Datasets:

math-ai

/

AutoMathText

like 150

Follow

math-ai 20

https://arxiv.org/abs/1510.08765

Representation for the Gauss-Laplace Transmutation

Under certain conditions, a symmetric unimodal continuous random variable $\xi$ can be represented as a scale mixture of the standard Normal distribution $Z$, i.e., $\xi = \sqrt{W} Z$, where the mixing distribution $W$ is independent of $Z.$ It is well known that if the mixing distribution is inverse Gamma, then $\xi$ is student's $t$ distribution. However, it is less well known that if the mixing distribution is Gamma, then $\xi$ is a Laplace distribution. Several existing proofs of the latter result rely on complex calculus and change of variables in integrals. We offer two simple and intuitive proofs based on representation and moment generating functions.

\section{Main Result to Prove} Let $\text{Expo}(1)$ denote the standard Exponential distribution with mean one, $\text{Laplace}$ the standard Laplace distribution with mean zero and variance two, $N(\mu,\sigma^2)$ the Normal distribution with mean $\mu$ and variance $\sigma^2.$ If two random variables $A$ and $B$ have the same distribution, then we write $A\sim B.$ The Gauss--Laplace transmutation states that $$ V\sim 2\text{Expo}(1),\quad L\mid V\sim N(0, V) \Longrightarrow L\sim \text{Laplace}, $$ or equivalently, if $\text{Expo}\sim \text{Expo}(1)$ is independent of $Z\sim N(0,1)$, then $$ L = \sqrt{2 \text{Expo}} Z \sim \text{Laplace}. $$ Interestingly, the Laplace and Normal distributions are Pierre-Simon Laplace's first and second law of errors \citep{wilson1923first}, which are tied together by a scaling factor $\sqrt{2 \text{Expo}}.$ In high dimensional statistics, this result is crucial to efficiently simulate posterior distribution of the Bayesian Lasso \citep{park2008bayesian}, which imposes Laplace priors on the regression coefficients. Some proofs of the Gauss--Laplace transmutation exist in the literature \citep{andrews1974scale, west1987scale}, which rely on advanced theory of complex calculus. In this note, we offer two simple proofs based on representation and moment generating functions (MGFs). \section{Review of Some Basic Representations} Let $Z\sim N(0,1), \text{Expo}\sim \text{Expo}(1), X^2_k\sim \chi^2_k$, and $L\sim \text{Laplace}$. We use these symbols for Normal, Exponential, Chi-Squared, and Laplace random variables from now on. We also use $S$ for a random sign taking values $\pm 1$ with probabilities $1/2.$ Let IID denote ``independently and identically distributed.'' The following representations are useful. \begin{enumerate} [(a)] \item\label{rep:a} If $Z_1, Z_2$ are IID $N(0,1)$, then $ 2\text{Expo}\sim X^2_{2}\sim Z_1^2+Z_2^2$. \begin{proof} Let $\text{Gamma}(a)$ be a Gamma random variable with shape parameter $a$ and rate parameter one. We have $2\text{Expo}\sim 2 \text{Gamma}(1) \sim X^2_2$ by the relationship between Gamma and Chi-Squared random variables, and $2\text{Expo}\sim Z_1^2 + Z_2^2$ by the definition the Chi-Squared random variable. \end{proof} \item\label{rep:b} If $Z_1, Z_2$ are IID $N(0,1)$, then $Z_1Z_2\sim (Z_1^2-Z_2^2)/2.$ \begin{proof} Because the bivariate Normal distributions satisfy $(Z_1,Z_2)\sim (Z_1+Z_2, Z_1-Z_2)/\sqrt{2}$, we have $Z_1Z_2\sim (Z_1+Z_2)( Z_1-Z_2)/2 = (Z_1^2-Z_2^2)/2.$ \end{proof} \item\label{rep:c} If $\text{Expo}_1, \text{Expo}_2$ are IID $\text{Expo}(1)$ independent of $S$, then $L\sim S \text{Expo}_1 \sim \text{Expo}_1 - \text{Expo}_2.$ \begin{proof} Obviously, $L\sim S \text{Expo}_1$, which justifies the other name, double exponential, of the Laplace distribution. By symmetry and the memoryless property, we have $$ \text{Expo}_1 - \text{Expo}_2\sim S \{ \max(\text{Expo}_1, \text{Expo}_2) - \min(\text{Expo}_1, \text{Expo}_2) \} \sim S \text{Expo}_1. $$ \end{proof} \end{enumerate} \section{A Proof Based on Representation} \begin{proof} We let $Z, Z_1,Z_2,Z_3,Z_4$ be IID $N(0,1)$, and $\text{Expo}_1,\text{Expo}_2$ be IID $\text{Expo}(1).$ Using representation (\ref{rep:a}), we have $$ L = \sqrt{2 \text{Expo}} Z\sim \sqrt{ Z_1^2 + Z_2^2 } Z. $$ By conditioning on $(Z_1,Z_2)$, we have $$ Z_1Z_3 + Z_2Z_4 \sim N(0, Z_1^2 + Z_2^2) \sim \sqrt{ Z_1^2 + Z_2^2 } Z, $$ which implies $ L\sim Z_1Z_3 + Z_2Z_4. $ Using representation (\ref{rep:b}), we have $$ L\sim (Z_1^2-Z_3^2)/2 + (Z_2^2-Z_4^2)/2 = (Z_1^2+Z_2^2)/2 - (Z_3^2+Z_4^2)/2 . $$ Using representation (\ref{rep:a}) again, we have $ L\sim \text{Expo}_1 - \text{Expo}_2. $ Therefore, the final result follows from representation (\ref{rep:c}). \end{proof} \section{A Proof Based on Moment Generating Functions} The MGF of $Z$ is $M_N(t) = E(e^{tZ}) = e^{t^2/2}$, and the MGF of $\text{Expo}$ is $M_E(t) = E(e^{t\text{Expo}}) = 1/(1-t)$ for $ t <1.$ We can prove the result by calculating the MGFs directly. \begin{proof} Using the known MGFs above and two conditional arguments, we can verify that the MGF of $\sqrt{2 \text{Expo}} Z $ is $$ E( e^{t\sqrt{2 \text{Expo}} Z } ) =E\{ E( e^{t\sqrt{2 \text{Expo}} Z } \mid \text{Expo}) \} = E\{ M_N(t\sqrt{2 \text{Expo}}) \} = E\{ e^{t^2 \text{Expo}} \} = M_E(t^2) = {1\over 1-t^2} , $$ and the MGF of $L\sim \text{Laplace}\sim S \text{Expo}$ is $$ E(e^{tS\text{Expo}}) = E\{ E(e^{tS\text{Expo}} \mid S) \} = E\{ M_E(tS) \} = E\left ( 1 \over 1-tS \right) = \frac{1}{2}\left( \frac{1}{1-t} + \frac{1}{1+t} \right) = {1\over 1-t^2} . $$ \end{proof} \section{Discussion} Because of representation (\ref{rep:a}), we have $\sqrt{2 \text{Expo}} Z \sim S \sqrt{X^2_1 X^2_2}$ and $L\sim S\text{Expo}\sim S X_2^2/2$. Therefore, the Gauss--Laplace transmutation reduces to $4 X^2_1 X^2_2 \sim (X_2^2)^2$. In fact, a general result holds: $4 X^2_{k} X^2_{k+1} \sim (X^2_{2k})^2.$ This is the stochastic analogue of Legendre's duplication formula for Gamma functions as discussed in \citet{gordon1989bounds}, \citet{gordon1994stochastic} and the Supplementary Materials. The proof by representation also establishes the result $ \sqrt{X_2^2} Z\sim Z_1Z_3 + Z_2Z_4$. In fact, a general result holds: $\sqrt{X_n^2} Z \sim \sum_{i=1}^n Z_{1i} Z_{2i}$, where $Z_{1i}, Z_{2i}\ (i=1,\ldots,n)$ are IID $N(0,1).$ This result is related to the Bartlett decomposition of the Wishart matrix \citep{anderson2003introduction} as commented in the Supplementary Materials. As a byproduct of our proof by representation, we have shown that $ Z_1Z_3 + Z_2Z_4$ follows a Laplace distribution. Because of symmetry $Z_2\sim -Z_2$, the determinant of a two by two matrix with IID $N(0,1)$ entries, $Z_1Z_3 - Z_2Z_4$, also follows a Laplace distribution. This result has been established early in the literature \citep{nyquist1954distribution, nicholson1958distribution}, with a proof similar to our representation by \citet{mantel1966light}, and a proof based on characteristic functions by \citet{mantel1973characteristic}. In a series of exchanges in {\it The American Statistician}, these proofs were revived by \citet{missiakoulis1985distribution}, \citet{farebrother1986pitman} and \citet{mantel1987laplace}. Professor Christian P. Robert posted online a proof based on directly calculating the density function. He generously allowed us to include his proof, which is in the Supplementary Materials. In a blog article about Gauss--Laplace transmutation, he was enthusiastic about proofs without complicated integrals or advanced theory, which motivated our note here. \bibliographystyle{apalike}

https://arxiv.org/abs/1406.1561

Formal Verification of Medina's Sequence of Polynomials for Approximating Arctangent

The verification of many algorithms for calculating transcendental functions is based on polynomial approximations to these functions, often Taylor series approximations. However, computing and verifying approximations to the arctangent function are very challenging problems, in large part because the Taylor series converges very slowly to arctangent-a 57th-degree polynomial is needed to get three decimal places for arctan(0.95). Medina proposed a series of polynomials that approximate arctangent with far faster convergence-a 7th-degree polynomial is all that is needed to get three decimal places for arctan(0.95). We present in this paper a proof in ACL2(r) of the correctness and convergence rate of this sequence of polynomials. The proof is particularly beautiful, in that it uses many results from real analysis. Some of these necessary results were proven in prior work, but some were proven as part of this effort.

\section{Introduction} \label{intro} In this paper, we describe a formalization in ACL2(r) of a polynomial approximation to arctangent. The obvious approach to approximating a transcendental function is to use a general approximation scheme, such as the Taylor Series. However, the Taylor Series for arctangent converges very slowly: \begin{equation} \label{taylor-arctan} \arctan(x) = x - \frac{x^3}{3} + \frac{x^5}{5} - \dots = \sum_{k=0}^{\infty}{\frac{(-1)^k}{2k+1}x^{2k+1}} \end{equation} As Equation~\ref{taylor-arctan} shows, the denominators are growing at the rate of $O(n)$, not $O(n!)$ as is the case for the Taylor series of sine, cosine, or $e^x$. Consequently, the n$^\text{th}$ terms in the series decrease much more slowly, and the convergence rate is disastrous. The long-term goal of this research project is to formally model the x86 instructions that compute trigonometric, logarithmic, and exponential functions~\cite{Rus:transcendentals}. So it is of practical importance to use a polynomial approximation that converges more quickly to arctangent. A recent result of Medina's provides such an approximation~\cite{Med:arctan}, and this paper describes a formalization of that result in ACL2(r). The paper is organized as follows. In Section~\ref{arctan}, we describe how the arctangent function can be introduced in ACL2(r). Section~\ref{polys} presents a necessary detour into the basic calculus of polynomials, including the rules for integrating and differentiating polynomials. Section~\ref{medina} deals with Medina's polynomial approximation. Finally, Section~\ref{conclusion} presents some concluding remarks on the use of ACL2(r) for this project. \section{The Arctangent in ACL2(r)} \label{arctan} \subsection{Introducing Arctangent} We begin this discussion by introducing the arctangent function into ACL2(r). From the perspective of ACL2(r), the exponential function $e^x$ is the most fundamental of the transcendental functions. It is defined as a power series over the complex plane, and the trigonometric functions sine and cosine are introduced in terms of $e^x$. The tangent function itself is introduced as the quotient of sine and cosine. ACL2(r) allows the definition of inverse functions, such as arctangent~\cite{GaCo:inverses}. In order to introduce the inverse function for $f(x)$, it is necessary to prove certain obligations (which correspond to constraints in a hidden \texttt{encapsulate}): \begin{itemize} \item $f:D\rightarrow R$ is defined on interval $D$, and its range is the interval $R$. \item $f$ is 1-to-1 over the domain $D$. \item $f$ is continuous over $D$. \item If $y\in R$, there are $x_1\in D$ and $x_2\in D$ such that $f(x_1) \le y \le f(x_2)$. \end{itemize} The challenge, then, is to prove that tangent has these properties, in order to introduce its inverse, arctangent. By convention, we chose the relevant domain of tangent to be $(-\pi/2, \pi/2)$, and the range of tangent over this domain is the entire number line $\mathbb{R}$. Next we show that tangent is 1-to-1 on the domain $(-\pi/2, \pi/2)$. We do this with a little calculus. If we can show that the derivative of tangent is positive on $(-\pi/2, \pi/2)$, then it must, necessarily, be increasing over this range. Moreover, if tangent is differentiable on $(-\pi/2, \pi/2)$, it must also be continuous on that range. Thus, the derivative of tangent provides two of the needed proof obligations. Tangent is defined in ACL2(r) as $\tan(x) \equiv \frac{\sin(x)}{\cos(x)},$ so its derivative follows from the product and quotient rules and the derivatives of sine and cosine~\cite{GaCo:chain-rule,ReGa:automatic-differentiator}. the major complication is proving that $\cos(x)$ is non-zero for $x\in (-\pi/2, \pi/2)$. This was actually proven earlier, in part to define the constant $\pi$ in ACL2(r) as (twice) the first positive zero of cosine~\cite{Gam:dissertation}! It should be noted that the result of this effort is that \begin{align} \frac{d(\frac{\sin(x)}{\cos(x)})}{dx} & = \frac{\sin(x) [(-1)(-\sin(x))]}{\cos^2(x)} + \cos(x) \frac{1}{\cos(x)} \\ &= \frac{\sin^2(x)}{\cos^2(x)} + 1 \end{align} It takes (proving and) using the trigonometric identity $\tan^2(x) + 1 = \sec^2(x)$ to reduce this expression to the familiar $\tan'(x) = \sec^2(x)$. As mentioned previously, now that the derivative is known, it follows directly that tangent is continuous on the desired interval. To show that tangent is 1-to-1 on the interval, we use the fact that the derivative $\sec^2(x)$ is positive on $(-\pi/2, \pi/2)$. We found it surprising that it was not already proven in ACL2(r) that a positive $f'$ guarantees increasing $f$. We formalized this small result using the Mean Value Theorem (MVT). If there are $x_1$ and $x_2$ such that $x_1 > x_2$ but $f(x_1) \le f(x_2)$, then by the MVT there is a point $c$ such that $x_1 < c < x_2$ and $f'(c) = \frac{f(x_2)-f(x_1)}{x_2-x_1} \le 0$. Since $f'$ is positive, no such point $c$ exists, hence no such $x_1$ and $x_2$ can be found. The final proof obligation is that for any $y\in\mathbb{R}$, we can find $x_1$ and $x_2$ in $(-\pi/2, \pi/2)$ such that $\tan(x_1) \le y \le \tan(x_2)$. This turned out to be a significant challenge, which we tackled in parts. For the first part, suppose $0 \le y \le 1$. Then $\tan(0) \le y \le \tan(\pi/4)$, since $\tan(0) = 0$, $\tan(\pi/4) = 1$, and tangent is an increasing function. So setting $x_1=0$ and $x_2=\pi/4$ will work. Before tackling the second part, we find an important lower bound on $\tan(x)$ whenever $\pi/4 \le x < \pi/2$. The lower bound is easily found since $\tan(x) = \sin(x)/\cos(x)$, sine is increasing on $[0, \pi/2]$, and $\sin(\pi/4) = 1/\sqrt{2}$, so $\tan(y) \ge 1/(\sqrt{2}\cos(x))$ when $\pi/4 \le y < \pi/2$. For the second part, suppose that $y>1$. The lower bound on tangent above can be turned into a range on arctangent as follows. Since $y > 1$, it follows that $1/(\sqrt{2}y) \in (0, 1)$. In turn, this means that $\arccos(y) \in (0, \pi/2)$. Actually, since cosine is decreasing on $(0, \pi/2)$, and $\cos(\pi/4)=1/\sqrt{2}$, $\arccos(y)$ is further restricted to $(\pi/4, \pi/2)$. So for $y > 1$, it follows that $\tan(0) \le y \le \tan(\arccos(1/(\sqrt{2}y)))$, so setting $x_1=0$ and $x_2=\arccos(1/(\sqrt{2}y))$ will work. The third and final part, when $y < 0$, can be derived from the results above by observing that $\tan(-y) = -\tan(y)$, so it is sufficient to find the bound for $\arctan(-y)$ and swap signs. At this point, the proof obligations for inverse functions are fulfilled, so we can introduce arctangent using \texttt{definv}. \subsection{The Derivative of Arctangent} The next step is to define the derivative of arctangent. The derivative of inverse functions was proven in~\cite{GaCo:chain-rule} and is given by \begin{equation} \frac{d(f^{-1}(y))}{dy} = \frac{1}{f'(f^{-1}(y))}. \end{equation} This formula is valid only when $f'$ is never infinitesimally small in the range of $y$. In the previous section, we showed that the derivative of tangent is $\sec^2(x) = 1/\cos^2(x)$. This function achieves its minimum when cosine achieves its maximum magnitude, i.e., when $\cos(x) = \pm 1$. Consequently, $\tan'(x) \ge 1$, so it is never infinitesimally small. That means \begin{equation} \frac{d(\tan^{-1}(y))}{dy} = \frac{1}{\sec^2(\arctan(y))} = \frac{1}{\tan^2(\arctan(y)) + 1} = \frac{1}{y^2+1}. \end{equation} The Fundamental Theorem of Calculus (FTC) was first proved in ACL2(r) in~\cite{Kau:ftc}, and we recently redid that proof to make the final statement of the FTC more direct. Using this result, it follows that \begin{equation} \int_{a}^{b}{\frac{dx}{1+x^2}} = \arctan(b) - \arctan(a). \end{equation} This result will play a key role in Section~\ref{medina}. \section{Polynomial Calculus} \label{polys} \subsection{The Derivative and Integral of $x^n$} We now turn our attention to the derivative and integral of the function $x^n$. Because this is really a binary function, of both $x$ and $n$, it illustrates the difficulties of working with the non-standard definition of derivative. For example, a direct way of proving that $\frac{d(x^n)}{dx} = n \cdot x^{n-1}$ is by using induction, invoking the product rule during the inductive step. The problem is that the non-standard definition of differentiability requires that, $small(\epsilon) \Rightarrow \frac{(x+\epsilon)^n - x^n}{\epsilon} \approx n \cdot x^{n-1}$. This is a non-classical formula, so it cannot be proved using functional instantiation with a pseudo-lambda expression, e.g., $f(x) \rightarrow (\lambda (x) x^n)$. That is part of the motivation behind proving in ACL2(r) that the $\epsilon$-$\delta$ definition of derivative is equivalent to the non-standard definition used in ACL2(r)~\cite{CoGa:equivalences}. Indeed, using the $\epsilon$-$\delta$ definition of derivative, it is possible to prove the derivative of $x^n$ by induction. However, there are still potential pitfalls. In particular, the key lemma in the inductive step requires the use of the product rule, $(f \times g)' = f \times g' + f' \times g$. But the proof obligations of the functional instantiation include the theorem $\frac{d(x^{n-1})}{dx} = (n-1) \cdot x^{n-2}$. This is part of the induction hypothesis, but injecting hypotheses into proof obligations of functional instantiation is a difficult problem. So we opted for a slightly more general approach. There are two different ways of writing $x^n$ in ACL2(r): \begin{itemize} \item \texttt(expt x n) \item \texttt(raise x n) \end{itemize} The \texttt{expt} function is identical to its counterpart in ACL2, so it is defined by induction on \texttt{n} (which must be an integer, not necessarily a natural number). The \texttt{raise} function is defined using $x^n = e^{n \ln(x)}$. For integer exponents $n$, these two definitions are known to be equal. The idea, then, is to use the derivative of $e^{n \ln(x)}$ to find the derivative of $x^n$. Previously, we had shown that the derivative of $e^x$ is precisely $e^x$~\cite{ReGa:automatic-differentiator}. With the use of the Chain Rule~\cite{GaCo:chain-rule} and the derivative of $\ln(x)$~\cite{ReGa:automatic-differentiator}, this means that \begin{align} \frac{d(x^n)}{dx} &= \frac{d(e^{n\ln(x)})}{dx} \\ &= n \frac{1}{x} e^{n\ln(x)} \\ &= n \frac{1}{x} x^n \\ &= n x^{n-1}. \end{align} However, this derivation makes several hidden assumptions that need to be addressed. The first problem is that the derivative of $\ln(x)$ is only known for $x>0$. (While the function $\ln(x)$ is defined for all non-zero complex numbers, derivatives in ACL2(r) are restricted to real-valued functions of real numbers.) So for positive values of $x$, this argument does hold, and we proved that \begin{equation} x>0 \Rightarrow \frac{d(x^n)}{dx} = n x^{n-1}. \end{equation} When $x<0$, $e^{n\ln(x)}$ isn't even necessarily defined over the reals, e.g., $(-1)^{\frac{1}{2}} = e^{\frac{1}{2} \ln(-1)} = i \not\in \mathbb{R}$. However, we can restrict $n$ to range over the integers, and then $x^n$ is defined even for negative $n$. Our approach was to show that whenever $x<0$, \begin{align} x^n &= e^{n \ln(x)} \\ &= e^{n \ln(-|x|)} \\ &= e^{n (\ln(|x|) + i \pi)} \\ &= e^{n \ln(|x|) + i \pi n} \\ &= e^{n \ln(|x|)} e^{i \pi n} \\ &= e^{n \ln(|x|)} (-1)^n \end{align} In the last step, $(-1)^n$ can be represented using either \texttt{raise} or \texttt{expt}, since $n$ is restricted to the integers. This means that $(-1)^n$ is equal to $1$ when $n$ is even and $-1$ when $n$ is odd, and these cases can be considered separately. At this point, the derivative of $x^n$ can be reduced to the case where $x>0$, since $|x|>0$. This shows that \begin{equation} x<0 \wedge n \in \mathbb{Z} \Rightarrow \frac{d(x^n)}{dx} = n x^{n-1}. \end{equation} That leaves the case when $x=0$. Again, we restrict ourselves to the case of integer $n$, because it is possible for $\epsilon$ to be infinitessimally close to $0$ yet still be negative. Moreover, $n$ cannot be negative, because in that case $0^n$ is undefined. When $n=0$, $x^n=1$, so the derivative of $x^n$ is 0, which is equal to $n x^{n-1} = 0\cdot x^{-1} = 0$. Note: This uses the fact that $1/0 = 0$ according to the axioms of ACL2. When $n>0$, $0^n = 0$ and $|\epsilon^n| \le |\epsilon|$ for $|\epsilon| < 1$. If $n=1$, then $\epsilon^n=\epsilon$, and the derivative of $x^n$ is just 1, and since $0^0=1$, this is exactly the same as $n x^{n-1} = 1 \cdot 0^{0} = 1$. When $n>1$, for infinitesimal $\epsilon$, $\epsilon^n \approx 0 = n 0^{n-1} = n \cdot 0$. So we have shown that \begin{equation} x=0 \wedge n \in \mathbb{N} \Rightarrow \frac{d(x^n)}{dx} = n x^{n-1}. \end{equation} Combining these results, we have that \begin{equation} \label{deriv-expt} \left[ (x>0) \vee (x<0 \wedge n \in \mathbb{Z}) \vee (x=0 \wedge n \in \mathbb{N}) \right] \Rightarrow \frac{d(x^n)}{dx} = n x^{n-1}. \end{equation} It is interesting that so many hypotheses are needed for this result, which is taken for granted in calculus. However, the assumption there is that the result holds only when all expressions in the theorem are defined. This is a powerful assumption that hides hypotheses. Before proceeding, we would like to make the following observation. Many of the theorems require hypotheses such as $n \in \mathbb{Z}$. Since $n$ is not one of the parameters of the function $f$ that is being functionally instantiated, these arguments have to be ``infected'' when using functional instantiation. One of the traditional approaches is to use a pseudo-lambda term with a condition and a default value, as in the following: \begin{lstlisting} :functional-instance useful-theorem (f (lambda (x) (if (not (integerp n)) 0 (expt x n)))) \end{lstlisting} However, since many such functions need to be instantiated, it is not always obvious how to define the ``unintended domain'' cases so that the constraints of all the combined functions hold. So we found it more productive to move these hypotheses into the definitions, as in the following: \begin{lstlisting} (defun raise-to-int (x n) (raise (realfix x) (ifix n))) \end{lstlisting} Then we proved the required theorems about the ``fixed'' functions, and only later raised the hypotheses to the statements as in Equation~\ref{deriv-expt}. Once the derivative of $x^n$ is known, it is a simple matter to invoke the FTC to find the integral of $x^n$: \begin{equation} \left[ (x>0) \vee (x<0 \wedge n \in \mathbb{Z}) \vee (x=0 \wedge n \in \mathbb{N}) \right] \Rightarrow \int_{a}^{b}{x^n dx} = \frac{b^{n+1}}{n+1} - \frac{a^{n+1}}{n+1}. \end{equation} \subsection{The Derivative and Integral of Polynomials} It is now time to extend the results in the previous section to polynomials. The first challenge is to capture the notion of polynomials in ACL2(r), and we chose to use the characterization described in~\cite{GaCo:cantor-trio}. Polynomials are encoded as lists of coefficients, with the first coefficient being the constant term, and subsequent coefficients corresponding to higher powers of $x$. For example, the polynomial $3 + x^2$ is encoded as the list \texttt{(3 0 1)}. The function \texttt{eval-polynomial} evaluates a polynomial at a point, and what we have to show is that its derivative is also a polynomial. That particular function used the following recursive scheme: \begin{equation} evalpoly(cons(c,rest), x) = c + x\cdot evalpoly(rest, x) \end{equation} It is an easy challenge to define an alternative execution based on a scheme that uses $x^n$: \begin{equation} evalpoly(cons(c,rest), x, n) = c\cdot x^n + evalpoly(rest, x, n+1) \end{equation} Once these two functions are proved equivalent, the results from the previous section can be used directly. So the first step is to define the list of coefficients of the derivative of a polynomial. This is easily done, e.g., as in the following definition: \begin{lstlisting} (defun derivative-polynomial-aux (poly n) (if (and (real-polynomial-p poly) (natp n) (consp poly)) (if (< 0 n) (cons (* n (car poly)) (derivative-polynomial-aux (cdr poly) (1+ n))) (derivative-polynomial-aux (cdr poly) (1+ n))) nil)) \end{lstlisting} The proof that this polynomial is the derivative of the original polynomial can proceed by induction. Recall that one of the complications described in the previous section is the difficulty of pushing the inductive hypothesis into the proof obligations of a functional instantiation. However, the key lemma that is required in this case is that $(f+g)'(x) = f'(x)+g'(x)$. The proof of this lemma is easy enough that it can be carried out as part of the induction. The trick is to do induction such that $\langle poly, n, \epsilon \rangle \rightarrow \langle cdr(poly), n+1, \epsilon/2 \rangle$. As before, once the derivative of polynomials is established, it is easy to invoke the FTC in order to introduce the integral of polynomials. We defined a function similar to \texttt{derivative-polynomial-aux} that computes the coefficient of the integral. \section{Medina's Result} \label{medina} Now that all preliminaries have been dealt with, we can formalize Medina's main result. In order to make arctangent more tractable, Medina first reduces the domain of arctangent to $[0,1]$. He can do this by using the following lemmas: \begin{gather} x > 1 \Rightarrow \arctan(x) = \frac{\pi}{2} - \arctan\left(\frac{1}{x}\right) \label{medina-lemma-1}\\ x < 0 \Rightarrow \arctan(x) = -\arctan(-x) \label{medina-lemma-2} \end{gather} The proof of Equation~\ref{medina-lemma-1} follows by proving that the tangent of both sides is equal, and then using the uniqueness of inverse functions (in the appropriate domain). Equation~\ref{medina-lemma-2} follows even more directly using the same approach. Incidentally, neither of these lemmas requires the given hypothesis. Now that these lemmas are proved, we can restrict $x$ to the range $x\in[0,1]$. Medina defines the following sequence of polynomials: \begin{align} p_1(x) &= 4 - 4x^2 + 5x^4 - 4x^5 + x^6 \\ p_m(x) &= x^4(1-x)^4 p_{m-1}(x) + (-4)^{m-1} p_1(x) \end{align} The first step is to find a more direct way of writing $p_m$. For $m\ge 2$, the polynomial can be written as follows: \begin{equation} \label{medina-alternative} p_m(x) = \frac{x^{4m}(1-x)^{4m} + (-4)^m}{1+x^2}. \end{equation} This is not obviously a polynomial, but $1+x^2$ is actually a factor of the numerator. But since the structure is not clearly that of a polynomial, we introduced the functions $p_m$ explicitly, instead of using \texttt{eval-polynomial}. The proof of Equation~\ref{medina-alternative} is quite involved, although it requires only induction on $m$ and elementary algebra. The difficulty comes from the necessary algebraic manipulations. We next focus on the term $x(1-x)=x-x^2$ when $x\in[0,1]$. The derivative of this polynomial is $1-2x$, and this is zero when $x=1/2$. In prior work, we had proved the Extreme Value Theorem that says the derivative is zero when the function achieves a maximum or minimum~\cite{Gam:dissertation}. Unfortunately, that is not the lemma that is required here. Instead, what is needed is to show that when the derivative is zero \emph{and some other conditions hold}, the function is at a maximum. The ``other conditions'' can vary, but we chose to formalize the First-Derivative Test. That is, if the derivative is positive for all $x<a$, zero at $a$, and negative for all $x>a$, then $f$ achieves a maximum at $a$. More precisely, the variable $x$ is restricted to range over some interval $I$ containing $a$, not over all reals---although in this case, that would have been sufficient. Since $x(1-x)$ achieves a maximum at $1/2$, we have that $x(1-x) \le 1/4$ for all $x\in[0,1]$. Moreover, since $x(1-x)\ge 0$ when $x\in[0,1]$, it follows that \begin{equation} x^{4m}(1-x)^{4m}\le \left(\frac{1}{4}\right)^{4m}. \end{equation} Now, $1+x^2 \ge 1$, so we have also shown that \begin{equation} \frac{x^{4m}(1-x)^{4m}}{1+x^2} \le \left(\frac{1}{4}\right)^{4m}. \end{equation} Taking the integral of both sides shows the following: \begin{align} \int_{0}^{x}\frac{t^{4m}(1-t)^{4m}}{1+t^2} dt &\le \int_{0}^{x} \left(\frac{1}{4}\right)^{4m} dt \\ &=\left(\frac{1}{4}\right)^{4m} x \\ &\le \left(\frac{1}{4}\right)^{4m} \label{medina-bound} \end{align} Note that the last step follows only because $x\in[0,1]$. We now return to Equation~\ref{medina-alternative}, which we reproduce below: \begin{equation} p_m(x) = \frac{x^{4m}(1-x)^{4m} + (-4)^m}{1+x^2}. \end{equation} This can be rewritten as follows: \begin{align} \frac{x^{4m}(1-x)^{4m}}{1+x^2} &= p_m(x) + \frac{(-4)^m}{1+x^2} \\ &= p_m(x) - \frac{(-1)^{m+1}4^m}{1+x^2}. \end{align} Notice that the left-hand side is non-negative for $x\in [0,1]$, so the right-hand side must be non-negative as well. We will use that observation in the next step, but first we take integrals of both sides and use Inequality~\ref{medina-bound}: \begin{align} p_m(t) - \frac{(-1)^{m+1}4^m}{1+t^2} &= \frac{t^{4m}(1-t)^{4m}}{1+t^2} \\ \int_{0}^{x} p_m(t) - \frac{(-1)^{m+1}4^m}{1+t^2} dt &= \int_{0}^{x} \frac{t^{4m}(1-t)^{4m}}{1+t^2} dt \le \left(\frac{1}{4}\right)^{4m} \end{align} The next step is to divide the last equation by $(-1)^{m+1}4^m$. This can change the direction of the inequality, but since both terms are positive (as discussed above), the magnitude of absolute values is preserved. This results in the following: \begin{equation} \left| \int_{0}^{x} \frac{p_m(t)}{(-1)^{m+1}4^m} - \frac{1}{1+t^2} dt \right| \le \left|\frac{1}{4}\right|^{5m} \end{equation} Now, we use the derivative of arctangent to integrate the second term in the integral. \begin{align} \left| \int_{0}^{x} \frac{p_m(t)}{(-1)^{m+1}4^m} - \frac{1}{1+t^2} dt \right| &\le \left|\frac{1}{4}\right|^{5m} \\ \left| \int_{0}^{x} \frac{p_m(t)}{(-1)^{m+1}4^m} dt - \int_{0}^{x}\frac{1}{1+t^2} dt \right| &\le \left|\frac{1}{4}\right|^{5m} \\ \left| \int_{0}^{x} \frac{p_m(t)}{(-1)^{m+1}4^m} dt - \arctan(x) \right| &\le \left|\frac{1}{4}\right|^{5m} \end{align} All that is left is to define the polynomial approximation: \begin{equation} h_m(x) \equiv \int_{0}^{x} \frac{p_m(t)}{(-1)^{m+1}4^m} dt. \end{equation} The previous results show that $h_m(x)$ is a good approximation to arctangent. In particular, \begin{equation} \left| h_m(x) - \arctan(x) \right| \le \left|\frac{1}{4}\right|^{5m}. \end{equation} The $1/4^{5m}$ term on the right-hand side shows that the convergence is quite good. As before, it is not at all obvious that $h_m(x)$ is actually a polynomial. But this does follow because $p_m(x)$ is a polynomial, the other term inside the integral is a constant, and the integral of a polynomial is also a polynomial. It would be more satisfying, however, to have an expression for $h_m(x)$ that is an actual list of coefficients. Medina does derive a closed form for $p_m$, and hence for $h_m$, and we have formalized that proof in ACL2(r). The details of that proof involve mostly tedious algebra, so we do not present them here. \section{Conclusion} \label{conclusion} This paper formalized a result of Medina's which defined a polynomial approximation to arctangent that converges quickly. The proof made heavy use of results from prior work formalizing real analysis, such as the FTC, the MVT, composition rules for derivatives, etc. In addition, a handful of results were missing and were proved as part of this effort, such as the First Derivative Test. In some ways, the result is an obvious candidate for ACL2(r), as opposed to ACL2, since the final theorem uses the transcendental function arctangent: \begin{equation} \label{medina-eqn-1} \left| h_m(x) - \arctan(x) \right| \le \left|\frac{1}{4}\right|^{5m}. \end{equation} However, one can envision a way of proving this result in ACL2, and this is not unreasonable, since ACL2 has been used in the past to prove the correctness of hardware approximations of functions that do not technically exist in ACL2, such as the square root function. The key step is to start with an approximation of the given function, and then show that some other (e.g., faster) approximation is also close. For instance, instead of using arctangent, we could start with the Taylor approximation in Equation~\ref{taylor-arctan}. In particular, the polynomial $T_n(x)$ could be defined as the Taylor approximation of order $n$. This could lead to a theorem such as the following: \begin{equation} \label{medina-taylor-eqn-1} \left| h_m(x) - T_n(x) \right| \le \left|\frac{1}{4}\right|^{5m}. \end{equation} The problem is that it is not obvious how to compare $h_m$ and $T_n$. Certainly, the theorem will not hold when $n=m$. After all, $h_m$ should converge to arctangent much more quickly than $T_n$! Moreover, a recent discussion in the ACL2 mailing list has brought attention to the fact that proving that two different series converge to the same value can be very difficult in ACL2. The solution suggested by the experts in the mailing list is to show that each of the two series converges to some function, and that the functions the series converge to are the same. But such a strategy could not be carried out in this case, since arctangent is provably not in ACL2. E.g., $\arctan(1) = \pi/4$ is not a number in ACL2, since it is irrational. So we believe that it is necessary to have support for the reals in order to reason about results such as Inequality~\ref{medina-eqn-1} and even Inequality~\ref{medina-taylor-eqn-1}, and we are delighted that enough of real analysis has been formalized in ACL2(r) that the formalization effort was mostly focused on the results specific to the problem at hand, and (with the exception of the First Derivative Test) not on more fundamental results. \bibliographystyle{eptcs}

https://arxiv.org/abs/1505.07941

Solutions of equations over finite fields: enumeration via bijections

We present a simple proof of the well-known fact concerning the number of solutions of diagonal equations over finite fields. In a similar manner, we give an alternative proof of the recent result on generalizations of Carlitz equations. In both cases, the use of character sums is avoided by using an elementary combinatorial argument.

\section{Introduction.} Let $\mathbb F_q$ be a finite field of $q$ elements and let $\mathbb F_q^*=\mathbb F_q\setminus\{0\}$. We consider equations of the form $f_1(x_1,\dots,x_n)=f_2(x_1,\dots,x_n)$, where $f_1$ and $f_2$ are polynomials with coefficients in $\mathbb F_q$. Explicit formulas for the number of solutions $(x_1,\dots,x_n)\in\mathbb F_q^n$ are known only for certain classes of equations (see \cite{BEW,LN,MP}). The traditional approach is to express the number of solutions in terms of character sums and to apply known results about these sums. In this note, we present an alternative approach, based on the observation that for some families of equations the number of solutions does not depend on the coefficients, that is, the same formula is valid for the whole family. This suggests that one should try to establish a bijective correspondence between the sets of solutions of equations belonging to a given family and to use this correspondence to obtain an explicit formula for the number of solutions. In Sections~\ref{s2} and \ref{s3}, we illustrate these ideas. We denote by $N[f_1(x_1,\dots,x_n)=f_2(x_1,\dots,x_n)]$ the number of solutions of the equation $f_1(x_1,\dots,x_n)=f_2(x_1,\dots,x_n)$ in $\mathbb F_q^n$ and by $N^*[f_1(x_1,\dots,x_n)=f_2(x_1,\dots,x_n)]$ the number of such solutions with ${x_1\cdots x_n\ne 0}$. Unless explicitly stated otherwise, we will assume that equations have coefficients in $\mathbb F_q^*$, exponents occurring in them are positive integers and $n\ge 2$. We write $|\mathcal{S}|$ for the number of elements of a finite set~$\mathcal{S}$. \section{Diagonal equations.}\label{s2} We consider an equation of the type $$ a_1^{}x_1^{m_1}+\dots+a_n^{}x_n^{m_n}=0. $$ Such an equation is called a diagonal equation. As $x_j$ runs through all elements of $\mathbb F_q$, $x_j^{m_j}$ runs through the same elements as $x_j^{\gcd(m_j,q-1)}$ does with the same multiplicity. Therefore, $$ N[a_1^{}x_1^{m_1}+\dots+a_n^{}x_n^{m_n}=0]=N[a_1^{}x_1^{d_1}+\dots+a_n^{}x_n^{d_n}=0], $$ where $d_j=\gcd(m_j,q-1)$. In 1971, Joly~\cite{joly} obtained the following result. \begin{theorem} \label{t1} Assume that $\gcd(d_j,d_1\cdots d_n/d_j)=1$ for some $j$. Then $$ N[a_1^{}x_1^{d_1}+\dots+a_n^{}x_n^{d_n}=0]=q^{n-1}. $$ \end{theorem} In his proof, Joly used the expression for the number of solutions in terms of Gauss sums. We give an alternative proof based on the ideas mentioned in the introduction. \begin{proof} Without loss of generality we may assume that $\gcd(d_1,d_2\cdots d_n)=1$. For $c\in\mathbb F_q$, let $\mathcal{S}_c$ denote the set of solutions $(x_1,\dots,x_n)\in\mathbb F_q^n$ of the equation $$ a_1^{}cx_1^{d_1}+a_2^{}x_2^{d_2}+\dots+a_n^{}x_n^{d_n}=0 $$ and $\mathcal{S}'_c$ denote the set of such solutions with $x_1\ne 0$. Clearly, $\mathcal{S}_{c_1}\setminus\mathcal{S}'_{c_1}=\mathcal{S}_{c_2}\setminus\mathcal{S}'_{c_2}$ for any $c_1,c_2\in\mathbb F_q$. In particular, we have $|\mathcal{S}_1\setminus\mathcal{S}'_1|=|\mathcal{S}_0\setminus\mathcal{S}'_0|$, or, equivalently, \begin{equation} \label{eq1} |\mathcal{S}_1|=|\mathcal{S}'_1|+|\mathcal{S}_0|-|\mathcal{S}'_0|. \end{equation} Observe that for any $n$-tuple $(x_1,\dots,x_n)\in\mathbb F_q$ with $x_1\ne 0$ there exists a unique $c\in\mathbb F_q$ such that $(x_1,\dots,x_n)\in\mathcal{S}'_c$. In other words, we have $$ \bigvee_{c\in\mathbb F_q}\mathcal{S}'_c=\mathbb F_q^*\times\underbrace{\mathbb F_q\times\dots\times\mathbb F_q}_{n-1}. $$ This yields \begin{equation} \label{eq2} \sum_{c\in\mathbb F_q}|\mathcal{S}'_c|=q^{n-1}(q-1). \end{equation} Since $\gcd(d_1,d_2\cdots d_n)=1$, we can choose a positive integer $t$ with $t\equiv 0\!\pmod{d_1}$ and $t\equiv 1\pmod{d_2\cdots d_n}$. It is easy to see that for a fixed $c\in\mathbb F_q^*$ the map $$ (x_1,x_2,\dots,x_n) \longmapsto (c^{t/d_1}x_1,c^{(t-1)/d_2}x_2,\dots,c^{(t-1)/d_n}x_n) $$ is a bijection between $\mathcal{S}'_c$ and $\mathcal{S}'_1$. This implies that $|\mathcal{S}'_c|=|\mathcal{S}'_1|$ for any $c\in\mathbb F_q^*$, so that \eqref{eq2} can be rewritten as $|\mathcal{S}'_0|+(q-1)|\mathcal{S}'_1|=q^{n-1}(q-1)$, or, equivalently, \begin{equation} \label{eq3} |\mathcal{S}'_1|=q^{n-1}-\frac 1{q-1}\,|\mathcal{S}'_0|. \end{equation} Finally, note that \begin{align*} |\mathcal{S}_0|&=qN[a_2^{}x_2^{d_2}+\dots+a_n^{}x_n^{d_n}=0],\\ |\mathcal{S}'_0|&=(q-1)N[a_2^{}x_2^{d_2}+\dots+a_n^{}x_n^{d_n}=0], \end{align*} and thus \begin{equation} \label{eq4} |\mathcal{S}_0|=\frac q{q-1}\,|\mathcal{S}'_0|. \end{equation} Combining \eqref{eq1}, \eqref{eq3} and \eqref{eq4}, we find that $$ N[a_1^{}x_1^{d_1}+\dots+a_n^{}x_n^{d_n}=0]=|\mathcal{S}_1|=q^{n-1}, $$ as desired. \end{proof} The inclusion-exclusion principle of combinatorics yields the following corollary, which will be useful in the sequel. \begin{corollary} \label{c1} If $d_1,\dots,d_n$ are pairwise coprime then $$ N^*[a_1^{}x_1^{d_1}+\dots+a_n^{}x_n^{d_n}=0]=\frac{(q-1)^n+(-1)^n(q-1)}q. $$ \end{corollary} \section{Generalizations of Carlitz equations.}\label{s3} Carlitz~\cite{carlitz1} studied equations of the form $$ (x_1+\dots+x_n)^2=bx_1\cdots x_n $$ over finite fields of odd characteristic. In particular, he proved that $$ N[(x_1+x_2+x_3)^2=bx_1x_2x_3]=q^2+1. $$ In \cite{baoulina}, the above result of Carlitz was generalized as follows: $$ N[(x_1^{m_1}+\dots+x_n^{m_n})^k=bx_1^{}\cdots x_n^{}]=q^{n-1}+(-1)^{n-1}, $$ provided that $\gcd(\sum_{j=1}^n (M/m_j)-kM,(q-1)/D)=1$ and $d_1,\dots,d_n$ are pairwise coprime, where $d_j=\gcd(m_j,q-1)$, $j=1,\dots,n$, $M=\text{lcm}[m_1,\dots,m_n]$ and $D=\text{lcm}[d_1,\dots,d_n]$ (here $q$ is not necessarily odd). See the references in~\cite{baoulina} for earlier results in this direction. A further generalization of the Carlitz equation was recently considered in~\cite{PZC}. By using the augmented degree matrix and Gauss sums, the authors established the following: if $\gcd(\sum_{j=1}^n (k_jm_1\cdots m_n/m_j)-km_1\cdots m_n,q-1)=1$, then $$ N[(a_1^{}x_1^{m_1}+\dots+a_n^{}x_n^{m_n})^k=bx_1^{k_1}\cdots x_n^{k_n}]=q^{n-1}+(-1)^{n-1}. $$ Let us remark that, with the notation as above, we have $$ \sum_{j=1}^n \frac{k_jm_1\cdots m_n}{m_j}-km_1\cdots m_n=\frac{m_1\cdots m_n}M\,\biggl(\sum_{j=1}^n \frac{k_jM}{m_j}-kM\biggr). $$ Therefore the condition $\gcd(\sum_{j=1}^n (k_jm_1\cdots m_n/m_j)-km_1\cdots m_n,q-1)=1$ is equivalent to the following two conditions: $$ \gcd\biggl(\sum_{j=1}^n \frac{k_jM}{m_j}-kM,q-1\biggr)=1,\qquad \gcd\biggl(\frac{m_1\cdots m_n}M,q-1\biggr)=1. $$ It is readily seen that $M\cdot\gcd(m_i,m_j)\mid m_1\cdots m_n$. Since $\gcd(d_i,d_j)\mid(q-1)$ and $\gcd(d_i,d_j)\mid\gcd(m_i,m_j)$, we conclude that $\gcd(m_1\cdots m_n/M,q-1)=1$ if and only if $d_1,\dots,d_n$ are pairwise coprime. Summarizing, we reformulate the main result of~\cite{PZC} in the following way: \begin{theorem} \label{t2} Assume that $\gcd(\sum_{j=1}^n (k_jM/m_j)-kM,q-1)=1$ and $d_1,\dots,d_n$ are pairwise coprime. Then $$ N[(a_1^{}x_1^{m_1}+\dots+a_n^{}x_n^{m_n})^k=bx_1^{k_1}\cdots x_n^{k_n}]=q^{n-1}+(-1)^{n-1}. $$ \end{theorem} Our proof of Theorem~\ref{t2} uses an argument similar to the one used in the proof of Theorem~\ref{t1}. \begin{proof} For $c\in\mathbb F_q$, let $\mathcal{S}_c$ denote the set of solutions $(x_1,\dots,x_n)\in\mathbb F_q^n$ of the equation $$ (a_1^{}x_1^{m_1}+\dots+a_n^{}x_n^{m_n})^k=bcx_1^{k_1}\cdots x_n^{k_n} $$ and $\mathcal{S}^*_c$ denote the set of such solutions with $x_1\cdots x_n\ne 0$. Proceeding by the same type of argument as in the proof of Theorem~\ref{t1}, we find that \begin{equation} \label{eq5} |\mathcal{S}_1|=|\mathcal{S}^*_1|+|\mathcal{S}_0|-|\mathcal{S}^*_0| \end{equation} and \begin{equation} \label{eq6} \sum_{c\in\mathbb F_q}|\mathcal{S}^*_c|=(q-1)^n. \end{equation} Let $g$ be a generator of the cyclic group $\mathbb F_q^*$. Then $g^{\sum_{j=1}^n (k_jM/m_j)-kM}$ is also a generator of $\mathbb F_q^*$, in view of the fact that $\gcd(\sum_{j=1}^n (k_jM/m_j)-kM,q-1)=1$. For a fixed $c\in\mathbb F_q^*$, there is an integer $t$ such that $c=g^{t(\sum_{j=1}^n (k_jM/m_j)-kM)}$. It is readily seen that the map $$ (x_1,\dots,x_n) \longmapsto (g^{tM/m_1}x_1,\dots,g^{tM/m_n}x_n) $$ is a bijection between $\mathcal{S}^*_1$ and $\mathcal{S}^*_c$, and, exactly as in the proof of Theorem~\ref{t1}, we deduce from \eqref{eq6} that $$ |\mathcal{S}^*_1|=(q-1)^{n-1}-\frac 1{q-1}\,|\mathcal{S}^*_0|. $$ Substituting this into \eqref{eq5} and recalling that \begin{align*} |\mathcal{S}_0|&=N[a_1^{}x_1^{d_1}+\dots+a_n^{}x_n^{d_n}=0]=q^{n-1},\\ |\mathcal{S}^*_0|&=N^*[a_1^{}x_1^{d_1}+\dots+a_n^{}x_n^{d_n}=0]=\frac{(q-1)^n+(-1)^n(q-1)}q, \end{align*} by Theorem~\ref{t1} and Corollary~\ref{c1}, we obtain the result. \end{proof} Note that we used the condition that $d_1,\dots,d_n$ are pairwise coprime only for determining the number of solutions of the corresponding diagonal equation. By relaxing this condition and proceeding exactly as above, we establish the following result. \begin{theorem} \label{t3} Assume that $\gcd(\sum_{j=1}^n (k_jM/m_j)-kM,q-1)=1$. Then \begin{multline*} N[(a_1^{}x_1^{m_1}+\dots+a_n^{}x_n^{m_n})^k=bx_1^{k_1}\cdots x_n^{k_n}]=(q-1)^{n-1}\\ +N[a_1^{}x_1^{d_1}+\dots+a_n^{}x_n^{d_n}=0]-\frac q{q-1}N^*[a_1^{}x_1^{d_1}+\dots+a_n^{}x_n^{d_n}=0]. \end{multline*} \end{theorem} Theorem~\ref{t3} allows us to determine $N[(a_1^{}x_1^{m_1}+\dots+a_n^{}x_n^{m_n})^k=bx_1^{k_1}\cdots x_n^{k_n}]$ explicitly in certain cases when the number of solutions of the corresponding diagonal equation is known. In particular, using \cite[Theorem~2]{SY} and the inclusion-exclusion principle, we can generalize Theorem~\ref{t2} as follows (for a similar result in the case $a_1=\dots=a_n=1$, $k_1=\dots=k_n=1$ we refer to \cite{baoulina}). \begin{theorem} \label{t4} Assume that $\gcd(\sum_{j=1}^n (k_jM/m_j)-kM,q-1)=1$, $d_1,\dots,d_t$ are odd, $d_{t+1},\dots,d_n$ are even, $d_1,\dots,d_t,d_{t+1}/2,\dots,d_n/2$ are pairwise coprime, ${0\le t\le n}$. Then \begin{align*} N[(a_1^{}x_1^{m_1}&+\dots+a_n^{}x_n^{m_n})^k=bx_1^{k_1}\cdots x_n^{k_n}]=q^{n-1}+(-1)^{n-1}\\ &+(-1)^{n-1}\sum_{j=1}^{[(n-t)/2]}\eta((-1)^j)\sigma_{2j}(\eta(a_{t+1}),\dots,\eta(a_n))q^j\\ &+\begin{cases} \eta((-1)^{n/2}a_1\cdots a_n)q^{(n-2)/2}(q-1)&\text{if $t=0$ and $n$ is even,}\\ 0&\text{otherwise,} \end{cases} \end{align*} where $\sigma_{2j}(z_1,\dots,z_{n-t})$ are the elementary symmetric polynomials and $\eta$ is the quadratic character on $\mathbb F_q$ ($\eta(x)=+1,-1,0$ according as $x$ is a square, a non-square or zero in $\mathbb F_q$). \end{theorem} \section{Concluding remarks.} The method used in the proof of Theorem~\ref{t2} can be applied to a more general class of equations, namely, $$ f(x_1,\dots,x_n)=bx_1^{k_1}\cdots x_n^{k_n}, $$ where $f\in\mathbb F_q[x_1,\dots,x_n]$ is a polynomial for which there exist positive integers $r,r_1,\dots,r_n$ such that $f(c^{r_1}x_1,\dots,c^{r_n}x_n)=c^rf(x_1,\dots,x_n)$ for every $c\in\mathbb F_q$ (such a polynomial is sometimes called a quasi-homogeneous polynomial). If, in addition, $\gcd(\sum_{j=1}^n k_jr_j-r,q-1)=1$, then we have the expression \begin{multline*} N[f(x_1,\dots,x_n)=bx_1^{k_1}\cdots x_n^{k_n}]=(q-1)^{n-1}\\ +N[f(x_1,\dots,x_n)=0]-\frac q{q-1}N^*[f(x_1,\dots,x_n)=0], \end{multline*} and so the number of solutions of $f(x_1,\dots,x_n)=bx_1^{k_1}\cdots x_n^{k_n}$ is easily evaluated once $N[f(x_1,\dots,x_n)=0]$ and $N^*[f(x_1,\dots,x_n)=0]$ are known.

https://arxiv.org/abs/2210.17225

An isoperimetric problem with two distinct solutions

In this paper we prove that among all convex domains of the plane with two axis of symmetry, the maximizer of the first non trivial Neumann eigenvalue $\mu_1$ with perimeter constraint is achieved by the square and the equilateral triangle. Part of the result follows from a new general bound on $\mu_1$ involving the minimal width over the area. Our main result partially answers to a question addressed in 2009 by R. S. Laugesen, I. Polterovich, and B. A. Siudeja.

\section{Introduction} In this paper, we are interested in the shape optimization of the second Neumann eigenvalue of the Laplacian, denoted by $\mu_1$, see Section \ref{section0} for the Notation. We underline that the shape optimization of the first Neumann eigenvalue $\mu_0$ is not interesting, since $\mu_0(\Omega)=0$ for every shape $\Omega$. A well-known result, first proved by Szeg\H{o} (for plane simply-connected domains) and then generalized by Weinberger, says that the ball maximizes $\mu_1(\Omega)$ with volume constraint. The Dirichlet analogue is the Faber-Krahn inequality, which says that the ball minimizes the first Dirichlet eigenvalue of the Laplacian $\lambda_1(\Omega)$ under volume constraint. But less is known for the perimeter constraint. Concerning the first Dirichlet eigenvalue, it is easy to see that the ball is again the minimizer with fixed perimeter. Indeed, by successively using the isoperimetric inequality and then Faber-Krahn inequality, we get (here $\Omega$ is a bounded open set in $\mathbb{R}^N$ and $\omega_N$ is the volume of the unit ball $B_1$) \begin{eqnarray} P(\Omega)^{\frac{2}{N-1}}\lambda_1(\Omega) &\geq& \left(N\omega_N^{\frac{1}{N}}\right)^{\frac{2}{N-1}} |\Omega|^{\frac{2}{N}} \lambda_1(\Omega) \notag \\ &\geq& \left(N\omega_N^{\frac{1}{N}}\right)^{\frac{2}{N-1}} |B_1|^{\frac{2}{N}} \lambda_1(B_1)= P(B_1)^{\frac{2}{N-1}} \lambda_1(B_1),\notag \end{eqnarray} with equality for the ball in each of the above inequalities. For the Neumann eigenvalue $\mu_1(\Omega)$, the maximizer with perimeter constraint does not exist in general (see Proposition \ref{nonexistence1}). On the other hand, it is easy to see that a maximizer exists among convex sets (see Proposition \ref{existence}). However, up to our knowledge, even in dimension 2, the convex maximizer is not known. The question was actually explicitly addressed in an Oberwolfach meeting in 2009 by R. Laugesen and I. Polterovich, \cite[Problem 9.2]{LS}, see also \cite[Open Problem 6.66]{LS2}, from which we can conjecture the following. \begin{conj}\label{conj1} For all planar convex domain we have \begin{eqnarray} P^2(\Omega)\mu_1(\Omega)\leq 16\pi^2. \label{ineqConj} \end{eqnarray} The maximum being achieved by squares, and equilateral triangles. \end{conj} The eigenvalue $\mu_1(\Omega)$ cannot be computed explicitly in general. However, this can be done for simple domains, such as equilateral triangles, rectangles, and disks (see, e.g., \cite[Section 1.2.5]{livre_vert} and \cite{LS}). In particular, the equality in \eqref{ineqConj} for squares and equilateral triangles follows by a direct computation. Moreover, in \cite{LS} it was proved that among all triangles, the equilateral triangle is the maximizer. For the unit disk $\mathbb D$ (actually, for any disk, by scale invariance of the functional under study) we have $P^2(\mathbb{D})\mu_1(\mathbb{D})=4\pi^2(j'_{1,1})^2 \simeq (4\times 3,39)\pi^2$ which is far from achieving the bound $16\pi^2$. This gap is large enough to imply that all regular polygons with a number of sides $N\geq 5$ will satisfy \eqref{ineqConj}, by comparison with a ball (see Proposition \ref{polygon}). Let us also mention that the conjecture is supported by numerical simulations, performed by Beniamin Bogosel (private communication). \\ In this paper we partially prove the conjecture. Our main result is the following. \begin{theorem}\label{main} For all planar convex domains with two axis of symmetry, $$ P^2(\Omega)\mu_1(\Omega)\leq 16\pi^2. $$ The maximum is achieved by squares and equilateral triangles and only by them. \end{theorem} One of the key ingredients in the proof of Theorem \ref{main} is a new geometrical bound on $\mu_1(\Omega)$, valid for any bounded Lipschitz open set $\Omega \subset \mathbb{R}^2$ (see also Lemma \ref{boundmu1} for a general statement on any set, not necessarily Lipschitz). Besides the application to our problem, the result is interesting in itself. \begin{lemma}\label{joli} For all planar bounded (Lipschitz) open set $\Omega$ we have \begin{equation}\label{jolieq} \mu_1(\Omega)\leq \pi^2\frac{w(\Omega)^2}{|\Omega|^2} \end{equation} where $|\Omega|$ is the area of $\Omega$ and $w(\Omega)$ is the minimal width of $\Omega$. The inequality becomes an equality for all and only rectangles. \end{lemma} The proof of this lemma is quite short and it is given in Section \ref{sectionI} (together with a generalization in higher dimension, see Lemma \ref{boundmu1} and Lemma \ref{boundmu1N}). It allows us to conclude that Conjecture \ref{conj1} holds true for all planar convex domains with two orthogonal axis of symmetry. Actually, from Lemma \ref{joli} one can state the following immediate corollary. \begin{corollary} \label{corr} Any planar convex domain $\Omega$ satisfying \begin{eqnarray} P(\Omega) w(\Omega)\leq 4 |\Omega|, \label{inequG} \end{eqnarray} also satisfies inequality \eqref{ineqConj}. \end{corollary} As a matter of fact, any convex domain with two orthogonal axis of symmetry automatically satisfies \eqref{inequG} as it is proved in Proposition \ref{orthogonal}, which therefore answers to the conjecture for the case of convex domains with two orthogonal axis of symmetry, with a quite short proof. Moreover, among convex domains with two orthogonal axis of symmetry, only squares satisfy \eqref{jolieq} and \eqref{inequG}, and thus \eqref{ineqConj}, with an equality sign. It is worth noticing that there exist other convex domains satisfying \eqref{inequG}, and thus Conjecture \ref{conj1}, without having necessarily two axis of symmetry: this is the case, e.g., of parallelograms (see Proposition~\ref{parallelogram}) and centrally symmetric convex domains (see Remark \ref{centrally}). Our study of parallelograms contains and generalizes the result by Raiko \cite{raiko}, with a completely different proof. In order to obtain the main result (Theorem \ref{main}) in the generic case, we consider a domain $\Omega$ with two axis of symmetry with a given angle $\theta \in (0,2\pi)$. We rapidly arrive to the conclusion that it suffices to treat angles of the form $\theta=\pi/N$ with $N\geq 3$ (see Proposition \ref{axes}). Notice that when $\theta=\pi/4$, or more generically $\theta=\pi/N$ with $N=2^k$, $k\geq 2$, then the domain has two orthogonal axis of symmetry, and the argument above applies. The case of $\theta=\pi/N$ with $N\geq 5$ is treated in Section \ref{sec-small}: for such domains, the comparison with the disk, which in general gives a too rough estimate, gives the desired bound with the strict inequality sign. The last case that we need to consider is that of an angle $\theta=\pi/3$, i.e., when $\Omega$ admits the same axis of symmetry as the equilateral triangle. This case is much more involved and it is treated in Section \ref{sectionIII}. We have to introduce a parameter $a\in [0, 1/3]$, which, roughly speaking, describes how far the domain is from an equilateral triangle. The complicated situation occurs for small values of the parameter, $a\in [0,1/4]$, corresponding to shapes not far from the equilateral triangle. Here the strategy is to use a convex combination of the eigenfunctions of two suitable equilateral triangles as a test function for $\mu_1(\Omega)$. After a series of hard computations and estimates, we come to the conclusion that nothing but a finite number of configurations have to be tested, which is finally done by computer. The most technical parts are postponed to the Appendix. On the other hand, for the remaining intermediate values of the parameter, namely $a\in [1/4, 1/3]$, we can easily conclude by using Szeg\H{o}-Weinberger's inequality together with a nice reverse isoperimetric inequality (involving the perimeter and the area). This concludes the proof of Theorem~\ref{main}. \subsection*{Acknowledgements} The authors want to thank Beniamin Bogosel, Lorenzo Cavallina, Andrea Colesanti, Kei Funano, Richard Laugesen, and Shigeru Sakaguchi for the stimulating discussions. This work was partially supported by the project \emph{Shape Optimization (SHAPO)} n. ANR-18-CE40-0013 financed by the French \emph{Agence Nationale de la Recherche (ANR)}, and by the Lorraine and Tohoku Universities through a joint research project. IL is member of the Italian research group \emph{Gruppo Nazionale per l'Analisi Matematica, la Probabilit\`{a} e le loro Applicazioni (GNAMPA)} of the \emph{Istituto Nazionale di Alta Matematica (INdAM)}. \section{Notation and useful bounds}\label{section0} We start this section by fixing the notation that will be used throughout the paper. Then we give some known bounds of $P(\Omega)$ and $\mu_1(\Omega)$ coming from the literature that will be used later. As a direct application of one of these bounds (Szeg\H{o}-Weinberger's inequality) we give a short proof of the validity of Conjecture \ref{conj1} for regular polygons. \subsection*{Notation.} We denote by $Vect[\xi_1, \ldots, \xi_{m}]$, $m\in \mathbb N$, the vector space of $\mathbb R^N$ generated by the vectors $\xi_1, \ldots, \xi_m \in \mathbb R^N$. The orthogonal complement of a subspace $V$ of $\mathbb R^N$ is denoted by $V^\perp$. Given three points $A,B,C$ in the plane, we denote by $\overline{AB}$ the length of the segment joining $A$ and $B$, by $(AB)$ the line passing through $A$ and $B$ (when they do not coincide), and by $\widehat{ABC}$ the angle of vectors between $\overset{\longrightarrow}{AB}$ and $\overset{\longrightarrow}{BC}$. Given a set $\Omega$, we denote its area by $|\Omega|$, its perimeter (always defined when $\Omega$ is convex) by $P(\Omega)$, its diameter by $D(\Omega)$, and its convex hull by $\mathrm{conv}(\Omega)$. We denote by $\mathcal{H}^1$ the one dimensional Hausdorff measure, $\mathbb{S}^1$ the unit circle, and $\mathbb{D}$ the unit disk. The minimal width of the convex set $\Omega \subset \mathbb{R}^2$ (or more generally of any bounded open set) is defined by $$ w(\Omega):=\min_{\nu \in \mathbb{S}^1} \mathcal{H}^1(p_\nu(\Omega)),$$ where $p_\nu : \mathbb{R}^2 \to \mathbb{R} \nu$ is the orthogonal projection onto the vectorial line $\mathbb{R} \nu$ oriented by the vector $\nu$. If $\Omega$ is a bounded Lipschitz domain (as for instance a convex domain), the spectrum of the Neumann Laplacian is discrete, and the first Neumann eigenvalue is always zero. We will denote by $\mu_1(\Omega)$ the second eigenvalue, defined by \begin{equation}\label{mu1} \mu_1(\Omega)=\min_{u \in H^1(\Omega) \; : \; \int_{\Omega} u \;\mathrm{d} x=0} \frac{\int_{\Omega} |\nabla u|^2 \; \mathrm{d} x}{\int_{\Omega} u^2 \;\mathrm{d} x}. \end{equation} \subsection*{Some useful bounds.} The first inequality that we mention is the following: for all convex domains (with non empty interior) \begin{eqnarray} P(\Omega)> 2D(\Omega). \label{inclusion} \end{eqnarray} The second inequality says that among all domains the ball maximizes $\mu_1$ with fixed volume. More precisely (see \cite{W56} or \cite[7.1.2]{livre_vert}), for any Lipschitz domain $\Omega \subset \mathbb{R}^2$ we have \begin{eqnarray} |\Omega|\mu_1(\Omega)\leq \pi \mu_1(\mathbb{D})=\pi(j'_{1,1})^2, \quad\quad \text{ (Szeg\H{o}-Weinberger)} \label{SzegoW} \end{eqnarray} where $j'_{1,1}$ is the first zero of the derivative of the Bessel function $J_1$. In particular, $(j'_{1,1})^2\simeq 3.39$ (see \cite[page 11]{livre_vert}). The third inequality is a classical lower bound for $\mu_1$, valid for any convex domain $\Omega\subset \mathbb{R}^2$ (see \cite{PW}), \begin{eqnarray} D(\Omega)^2 \mu_1(\Omega)\geq \pi^2. \quad\quad \text{ (Payne-Weinberger)} \label{PW} \end{eqnarray} The last inequality that we mention gives a counterpart of the above inequality, with an upper bound. It was explicitly stated by Ba\~{n}uelos and Burdzy in \cite[Corollary 2.1]{BB}, but actually it follows from a more general result by Cheng \cite{cheng}: for any convex domain $\Omega \subset \mathbb{R}^2$, \begin{eqnarray} D(\Omega)^2 \mu_1(\Omega)\leq 4j_{0,1}^2. \quad\quad \text{ (Cheng, Ba\~{n}uelos-Burdzy)} \label{cheng} \end{eqnarray} Here $j_{0,1}$ is the first zero of the Bessel function $J_0$, and it holds $j_{0,1}\simeq 2.405$ (see \cite[page 11]{livre_vert}). \subsection*{The case of regular polygons}\label{polygons} As any regular polygon admits at least two axis of symmetry, we can treat them as a direct consequence of our main result (Theorem~\ref{main}). However, Szeg\H{o}-Weinberger's inequality allows us to give an independent and shorter proof. \begin{proposition} \label{polygon}For any regular polygon $\Omega \subset \mathbb{R}^2$ inequality \eqref{ineqConj} holds true. \end{proposition} \begin{proof} Let $\Omega$ be a regular polygon with $N$ sides. We denote by $A$ its area and $P$ its perimeter. Then a simple computation shows that $$ P=2N\sin\left(\frac{\pi}{N}\right), \quad A=N\sin\left(\frac{\pi}{N}\right)\cos\left(\frac{\pi}{N}\right). $$ Next, by Szeg\H{o}-Weinberger's inequality \eqref{SzegoW}, we know that $$ A\mu_1(\Omega)\leq \pi \mu_1(\mathbb{D}). $$ It follows that \eqref{ineqConj} will be true if $P^2\mu_1(\mathbb{D})\leq 16 A \pi$ or, equivalently, if $$ \mu_1(\mathbb{D})4 N^2 \sin^2\left(\frac{\pi}{N}\right) \leq 16N\sin\left(\frac{\pi}{N}\right)\cos\left(\frac{\pi}{N}\right)\pi. $$ Since $\cos\left(\frac{\pi}{N}\right)$ never vanishes for $N\geq 3$, we can transform the relation into $$ N \tan\left(\frac{\pi}{N}\right) \leq \frac{4\pi}{\mu_1(\mathbb{D})} . $$ Now $\mu_1(\mathbb{D})\simeq 3.39$ thus $\frac{4\pi}{\mu_1(\mathbb{D})}>3.706$. On the other hand, the function $x\mapsto x \tan\left(\frac{\pi}{x}\right)$ is decreasing on $[5,+\infty[$ and $$ 5 \tan\left(\frac{\pi}{5}\right)<3.633. $$ This means that \eqref{ineqConj} is satisfied for every polygon with $N\geq 5$. For $N=3,4$, as we already know, the desired inequality is an equality. \end{proof} \section{Existence and non existence results}\label{sectionexis} In this section we prove the existence of a convex maximizer for the quantity $P^2(\Omega)\mu_1(\Omega)$. Then we stress the non existence for the analogue (again among convex sets) minimizing problem and for the maximizing problem without convexity constraint. \begin{proposition}\label{existence} The problem $$\max\{P^2(\Omega)\mu_1(\Omega) \quad : \quad \Omega \subset \mathbb{R}^2 \text{, convex }\},$$ admits a solution. \end{proposition} \begin{proof}Let $\{\Omega_n\}_{n\in \mathbb N}$ be a maximizing sequence (of plane convex domains). By the scale invariance, we can assume that all these domains have diameter equal to 1, so that they are all contained into a fixed ball. By Blaschke selection theorem, only two situations can occur: \begin{enumerate} \item either there exists a convex set $\Omega^*$ with non empty interior such that $\Omega_n$ (or a subsequence) converges to $\Omega^*$ for the complementary Hausdorff distance, as $n\to \infty$. In that case, as the quantities $P(\Omega_n)$ and $\mu_1(\Omega_n)$ are continuous for the Hausdorff convergence in the class of convex domains, the set $\Omega^*$ will be the maximizer (for the continuity of $P(\Omega_n)$ see \cite[Lemma 2.3.5]{BcB} and for the continuity of $\mu_1(\Omega_n)$ see \cite{Henrot-Pierre} or \cite{henrot-ftouhi}); \item or the sequence $\Omega_n$ converges (for the complementary Hausdorff distance) to a segment of diameter 1, as $n\to \infty$. But in that case, the perimeter of $\Omega_n$ converges to $2 = 2D(\Omega_n)$, thus $$\lim_n P^2(\Omega_n )\mu_1 (\Omega_n ) = 4\lim_n D^2(\Omega_n )\mu_1(\Omega_n ) \leq 16 (j_{0,1})^2$$ by Cheng inequality \eqref{cheng}, thus we would have $${\rm sup}P^2(\Omega)\mu_1(\Omega) \leq 16 (j_{0,1})^2 < 16\pi^2$$ that is impossible. \end{enumerate} The proof is therefore completed. \end{proof} Regarding to the minimization of $P^2(\Omega) \mu_1(\Omega)$, we notice that a minimizer does not exist as stated in the following proposition. \begin{proposition} For all planar convex domains we have \begin{eqnarray} P^2(\Omega) \mu_1(\Omega) {>} 4 \pi^2. \label{lowerB} \end{eqnarray} The lower bound is optimal and it is reached asymptotically by a sequence of rectangles shrinking to a segment. \end{proposition} \begin{proof} Payne-Weinberger's inequality \eqref{PW} says that $$D^2(\Omega)\mu_1(\Omega)\geq \pi^2,$$ where $D(\Omega)$ is the diameter of $\Omega$. On the other hand we can use \eqref{inclusion} which, for the recall, says that for any planar convex domain we always have $P(\Omega)> 2 D(\Omega)$. These two inequalities give \eqref{lowerB}. Now by considering a sequence of rectangles $\Omega_\varepsilon:=[0,1]\times [0,\varepsilon]$, $0<\varepsilon <<1$, that shrink to the segment $[0,1]\times\{0\}$ as $\varepsilon \to 0$, we observe that $$P^2(\Omega_\varepsilon)\mu_1(\Omega_\varepsilon)=4(1+\varepsilon)^2 \pi^2\to 4\pi^2.$$ Here we have used the exact value of $\mu_1$ for rectangles of the form $[0,L]\times [0,l]$ with $L\geq l$, which reads $\mu_1([0,L]\times[0,l])= \frac{\pi^2}{L^2}$ (see, e.g., \cite[Proposition 1.2.13]{livre_vert}). \end{proof} We now show that if we relax the convexity constraint, then the maximizing problem has no solution. \begin{proposition} \label{nonexistence1}The following equality holds: $$\sup\left\{P^2(\Omega)\mu_1(\Omega) \quad : \quad \Omega \subset \mathbb{R}^2 \right\}=+\infty.$$ \end{proposition} \begin{proof} Let $a>0$ be a small parameter (that will be taken infinitesimal at the end of the proof) and let $Q_a=[0,a]\times [0,a]$ be a square of size $a$. We will construct a sequence of sets $\Omega_n$, $n\in \mathbb N$, such that for every $n\in \mathbb N$ the perimeter satisfies $P(\Omega_n)=1$, whereas, in the limit as $n\to \infty$, $\Omega_n \to Q_a$ with respect to the Hausdorff distance. For that purpose, for a given $n$ we construct a piecewise linear affine function $f_n:[0,a]\to \mathbb{R}$ as follows. We prescribe its value on a uniform partition of $[0,a]$ and then we extend it in an affine way: {\renewcommand{\arraystretch}{1.5} $$ \left\{ \begin{array}{lc} f_n(a\frac{k}{2n})=0 & \text{ for all } k=0,\dots,n-1, \\ f_n(a(\frac{k}{n}+\frac{1}{2n})) = b \frac{a}{2n} & \text{ for all } k=0,\dots,n-1, \\ \text{ linear affine otherwise,} \end{array} \right. $$ } where the parameter $b$ is $$b:=\sqrt{\frac{1}{16 a^2}-1}.$$ The graph of $f_n$ is the union of $n$ isosceles triangles with base of length $a/n$ and legs of length $$\sqrt{\left(\frac{a}{2n}\right)^2+\left(\frac{ba}{2n}\right)^2}=\frac{1}{8n}.$$ In particular, the graph of $f_n$ over $[0,a]$ has length exactly $1/4$, and does not depend on $n$. Moreover, since $\|f_n\|_\infty = ab/(2n)\to 0$, it follows that $f_n \to 0$ uniformly in $[0,a]$, as $n\to \infty$. Finally, we notice that $f_n$ is uniformly Lipschitz with constant $b$. Then we reproduce the graph of $f_n$ over each side of the square $Q_a$, see Figure \ref{Figure1} below: this procedure allows to construct (the boundary of) a sequence of domains $\Omega_n$, uniformly Lipschitz, converging to $Q_a$ with respect to the Hausdorff distance. \begin{figure}[h!] \begin{center} \includegraphics[height=7truecm]{dessin1.pdf} \caption{\textit{The construction of the sequence of sets $\Omega_n$.}}\label{Figure1} \end{center} \end{figure} By applying \cite{henrot-ftouhi}, (see also \cite[Section 2.3]{livre_vert}) we know that $\mu_1(\Omega_n) \to \mu_1(Q_a)=\frac{\pi^2}{a^2}$. On the other hand $P(\Omega_n)=1$ for all $n$, so that $P^2(\Omega_n)\mu_1(\Omega_n)\to \frac{\pi}{a^2}$. Consequently, $$\sup\left\{P^2(\Omega)\mu_1(\Omega) \quad : \quad \Omega \subset \mathbb{R}^2 \right\} \geq \frac{\pi^2}{a^2},$$ and we conclude by finally letting $a\to 0$ in the above inequality. \end{proof} \begin{remark} Let us observe that there exist domains with infinite perimeter but with a finite Neumann eigenvalue. This is, in particular, the case of some fractal domains as the Koch snowflake. Obviously, this provides another proof of Proposition \ref{nonexistence1}. \end{remark} \section{A general bound on $\mu_1$ and the case of orthogonal axis of symmetry}\label{sectionI} This section is devoted to a new upper bound of $\mu_1$, which is valid for generic planar domains (not necessarily convex) and which involves two geometric functionals: the area and the minimal width. In the convex setting, this allows us to show that the conjecture holds true for a broad class of domains: those for which the perimeter satisfies a suitable upper bound that matches the new upper bound of $\mu_1$. In the second part of the section, we provide some examples of such domains, proving in particular the conjecture \eqref{ineqConj} for convex sets with two orthogonal axis of symmetry. \begin{lemma}\label{boundmu1} For every (non-empty) open and bounded set $\Omega\subset \mathbb{R}^2$, we have \begin{equation}\label{inf} \inf_{u \in H^1(\Omega) \; : \; \int_{\Omega} u \mathrm{d} x=0} \frac{\int_{\Omega} |\nabla u|^2 \; \mathrm{d} x}{\int_{\Omega} u^2 \;\mathrm{d} x} \leq \pi^2\frac{w^2(\Omega)}{|\Omega|^2}. \end{equation} In particular, if $\mu_1(\Omega)$ is well defined, then $\mu_1(\Omega)\leq \pi^2w^2(\Omega)/|\Omega|^2$.\\ Equality occurs in this inequality only for rectangles. \end{lemma} \begin{proof} Let $\Omega\subset \mathbb R^2$ be a (non-empty) bounded open set. Throughout the proof, let $A:=|\Omega|$ and $w:=w(\Omega)$. Without loss of generality, we may assume that the coordinate system is chosen so that the minimal width of $\Omega$ is achieved in the vertical direction. In other words $\Omega$ is contained into the horizontal strip $S:= \mathbb{R} \times [-w/2 , w/2]$. In $S$ we consider the rectangle $$ R:=[-L/2 , L/2]\times [-w/2 , w/2], \quad \hbox{with }L:=A/w, $$ having area $|R|=A$, and we define the function $$ u(x,y):= \left\{ \begin{array}{lll} \sin(\pi x /L) & \text{ in } R\\ +1 & \text{ in } S \setminus R, \ x\geq L/2 \\ -1 & \text{ in } S \setminus R,\ x\leq -L/2. \end{array} \right. $$ Since $\Omega\subset \subset S$, the function $u$ is well defined in $\Omega$, but it does not have (in general) zero average in it. For $t\in \mathbb R$, let us denote by $\Omega_{t}$ the translation $\Omega_{t}:=\Omega + (t,0)$, which is still contained into the strip $S$. The map $t \mapsto \int_{\Omega_{t}} u $ is continuous, moreover, by construction of $u$ it is negative for $t<0$ large enough in modulus, and it is positive for $t>0$ large enough. Therefore, there exists $x_0\in \mathbb R$ such that $\int_{\Omega_{x_0}} u= 0 $. Since all the functionals appearing in \eqref{inf} are invariant under translation, it is enough to prove the statement for $\Omega_{x_0}$. The choice of $x_0$ allows us to obtain $$ \inf_{f \in H^1(\Omega_{x_0}) \; : \; \int_{\Omega_{x_0}} f \mathrm{d} x=0} \frac{\int_{\Omega_{x_0}} |\nabla f|^2 \;\mathrm{d} x}{\int_{\Omega_{x_0}} f^2 \;\mathrm{d} x} \leq \frac{\int_{\Omega_{x_0}} |\nabla u|^2 \;\mathrm{d} x}{\int_{\Omega_{x_0}} u^2 \;\mathrm{d} x}. $$ Let us bound from above the numerator and from below the denominator of the Rayleigh quotient: \begin{eqnarray} \int_{\Omega_{x_0}} |\nabla u|^2 \;\mathrm{d} x & =& \int_{\Omega_{x_0}\cap R} |\nabla u|^2 \;\mathrm{d} x \leq \int_R |\nabla u|^2\;\mathrm{d} x; \notag \\ \int_{\Omega_{x_0}} u^2 \;\mathrm{d} x & =& \int_{\Omega_{x_0}\cap R} u^2 \;\mathrm{d} x + |\Omega_{x_0} \setminus R| \notag \\ &= & \int_R u^2 \;\mathrm{d} x - \int_{R\setminus \Omega_{x_0}} u^2 \;\mathrm{d} x + |\Omega_{x_0} \setminus R| \notag \\ &\geq & \int_{R} u^2 \;\mathrm{d} x - |R\setminus \Omega_{x_0}| + |\Omega_{x_0} \setminus R| = \int_{R} u^2 \;\mathrm{d} x. \notag \end{eqnarray} \begin{figure}[h] \begin{center} \includegraphics[height=7truecm]{figure_2_NNew.png} \caption{\textit{The competitor $u$ for the domain $\Omega$.}}\label{Figure2} \end{center} \end{figure} In the lower bound of the denominator we have used $|u|\leq 1$ (true by construction) and $|\Omega_{x_0} \setminus R|= |R\setminus \Omega_{x_0}|$ (true since $|\Omega_{x_0}|=A=|R|$). A direct computation gives $$ \int_R |\nabla u|^2\;\mathrm{d} x = \frac{\pi^2 w}{2L}, \quad \int_{R} u^2 \;\mathrm{d} x = \frac{w L}{2}, $$ so that, recalling the choice $L=A/w$, $$ \inf_{f \in H^1(\Omega_{x_0}) \; : \; \int_{\Omega_{x_0}} f \mathrm{d} x=0} \frac{\int_{\Omega} |\nabla f|^2 \;\mathrm{d} x}{\int_{\Omega_{x_0}} f^2 \;\mathrm{d} x} \leq \frac{\pi^2}{L^2}= \frac{\pi^2w^2}{A^2}, $$ which concludes the proof of \eqref{inf}. The equality case comes immediately from the study of the two chains of inequalities. \end{proof} In the present paper we deal with planar shapes, however Lemma \ref{boundmu1} can be generalized in higher dimension, as we show in Lemma \ref{boundmu1N}. To state this result, we need some notation. Let $\Omega$ be any bounded set in $\mathbb{R}^N$. We successively define, for $k=1,2,\ldots N-1$: \begin{itemize} \item $w_1$ the minimal width of $\Omega$ and $\xi_1$ the unit vector defining the direction of this minimal width, \item $w_2$ the minimal width of the orthogonal projection of $\Omega$ on the space $Vect[\xi_1]^\perp$ and $\xi_2$ the unit vector defining the direction of this minimal width, \item $\vdots$ \item $w_k$ the minimal width of the orthogonal projection of $\Omega$ on the vector space $Vect[\xi_1,\xi_2,\ldots,\xi_{k-1}]^\perp$ and $\xi_k$ the unit vector defining the direction of this minimal width. \end{itemize} Then we can state: \begin{lemma}\label{boundmu1N} For every open, non-empty and bounded set $\Omega\subset \mathbb{R}^N$, we have \begin{equation}\label{inf2b} \inf_{u \in H^1(\Omega) \; : \; \int_{\Omega} u \; \mathrm{d} x=0} \frac{\int_{\Omega} |\nabla u|^2 \; \mathrm{d} x}{\int_{\Omega} u^2 \;\mathrm{d} x} \leq \pi^2\frac{\prod_{k=1}^{N-1}w_k^2}{|\Omega|^2}. \end{equation} In particular, if $\mu_1(\Omega)$ is well defined, then $\mu_1(\Omega)\leq \pi^2\prod_{k=1}^{N-1}w_k^2/|\Omega|^2$. \end{lemma} The proof of this Lemma follows exactly the same line as in the two-dimensional case by introducing the cuboid $$K=\left[-\frac{L}{2}, \frac{L}{2}\right]\times \left[-\frac{w_{N-1}}{2},\frac{w_{N-1}}{2}\right]\times \ldots \times \left[-\frac{w_1}{2}, \frac{w_1}{2}\right]$$ with $L$ chosen such that $L \prod_{k=1}^{N-1} w_k = |\Omega|$ and $u$ the same test function depending only on the first variable $x_1$. The proof is left to the reader. \medskip A direct consequence of Lemma \ref{boundmu1} is the immediate Corollary \ref{corr} stated in the Introduction, which says that the convex sets for which \begin{eqnarray} P(\Omega) w(\Omega)\leq 4 |\Omega| \label{inequG2} \end{eqnarray} satisfy the conjecture \eqref{ineqConj}. In what follows we provide some examples. We start with parallelograms: the next proposition generalizes the result contained in \cite{raiko}, where a similar statement has been proved only for some parallelograms of particular type. \begin{proposition}\label{parallelogram} All parallelograms satisfy \eqref{inequG2}. In particular all parallelograms satisfy \eqref{ineqConj}. \end{proposition} \begin{proof} Let $\Omega$ be a parallelogram with area $A$, minimal width $w$, and perimeter $P$. Let $L$ be largest side and $\ell$ the smaller side i.e. $\ell\leq L$. Let $\theta$ be the smallest angle, as in Figure \ref{Figure3}. \begin{figure}[h] \begin{center} \includegraphics[height=4truecm]{PL.png}\\ \caption{\textit{A parallelogram.}}\label{Figure3} \end{center} \end{figure} It is easy to see that $$ P=2\ell +2 L, \quad w \leq \ell \sin(\theta), \quad A=\sin(\theta)\ell L. $$ These inequalities imply \eqref{inequG2}: $$ Pw\leq (2\ell +2L) \sin(\theta)\ell\leq 4 L \sin(\theta)\ell = 4A. $$ The validity of \eqref{ineqConj} follows by Corollary \ref{corr}. \end{proof} Another family of domains satisfying $P(\Omega)w(\Omega) \leq 4 |\Omega|$ is that of convex sets with two orthogonal axis of symmetry. \begin{proposition}\label{orthogonal} All planar convex domains with two orthogonal axis of symmetry satisfy \eqref{inequG2}. In particular all planar convex domains with two orthogonal axis of symmetry satisfy inequality \eqref{ineqConj}; moreover, the equality sign is satisfied only for squares. \end{proposition} \begin{proof} In virtue of Corollary \ref{corr}, to get \eqref{ineqConj} we only need to prove \eqref{inequG2}. By approximation of $\Omega$ with polygons, we may directly assume that $\Omega$ is a polygon. We also assume that the two axis of symmetry are $\{x=0\}$ and $\{y=0\}$. We consider the part of $\Omega$ in the first quadrant $\{x\geq 0, \ y\geq 0\}$: by connecting the origin with the vertexes of the polygon, we divide $\Omega$ into triangles. Let $N$ be the number of such triangles (in the first quadrant). For $i=1, \ldots, N$, the $i$-th triangle has one side on $\partial \Omega$, whose length is denoted by $a_i$. The corresponding height is an orthogonal segment passing through the origin, whose length is denoted by $h_i$. According to this notation, we have $$ P(\Omega) = 4 \sum_{i=1}^N a_i, \quad |\Omega|= 2 \sum_{i=1}^N a_i h_i. $$ The key point of the proof relies on the following elementary observation about the minimal width (see Fig. \ref{Figure4}): \begin{equation}\label{cs} \forall i , \quad \quad 2 h_i\geq w(\Omega). \end{equation} \begin{figure}[h] \begin{center} \includegraphics[height=6truecm]{dessin4.pdf} \caption{\textit{Polygons with two orthogonal axis of symmetry.}}\label{Figure4} \end{center} \end{figure} Thus we infer that $$ |\Omega| \geq \sum_{i=1}^N a_i w(\Omega) =\frac14 w(\Omega) P(\Omega), $$ that is, \eqref{inequG2}. This concludes the proof of \eqref{ineqConj}. Let us consider the equality cases: take a planar convex domain $\Omega$ with two orthogonal axis of symmetry, such that $P^2(\Omega) \mu_1(\Omega)=16 \pi^2$. Then, in view of \eqref{inf} and \eqref{ineqConj}, it must satisfy $$ \mu_1(\Omega) = \pi^2 \frac{w^2(\Omega)}{|\Omega|}, \quad P(\Omega)w(\Omega) = 4 |\Omega|. $$ The former, thanks to Lemma \ref{boundmu1}, implies that $\Omega$ is a rectangle, say $\Omega=[0,L]\times [0,\ell]$, $\ell\leq L$. For rectangles, the second equality is satisfied if and only if $$ 2(L+\ell) \ell = 4 L \ell, $$ namely if and only if $L=\ell$ and $\Omega$ is a square. This concludes the proof.\end{proof} \begin{remark}\label{centrally}The key point in the proof of Proposition \ref{orthogonal} is the inequality \eqref{cs}, which relates the distance of two supporting lines to the height of the triangles. The inequality holds true more in general for centrally symmetric convex sets, implying the validity of the conjecture \eqref{ineqConj} for these domains. This fact has been pointed out by K. Funano and A. Colesanti, while discussing on the problem. \end{remark} \section{The case of axis of symmetry forming a small angle}\label{sec-small} This section contains two results: first we show that we can restrict ourselves to the case in which the angle between two axis of symmetry is of the form $\pi/N$, $N\geq 3$; then we analyse the case of small angles of the form $\pi/N$ with $N\geq 5$, proving the validity of \eqref{ineqConj} in this framework. Note that the case $\pi/4$ has been treated in the previous section, while the case $\pi/3$ will be the object of the next section. \begin{proposition}\label{axes}Let $\Omega\subset \mathbb{R}^2$ be any planar convex domain which admits two axis of symmetry with respective angle $\theta\in (0,2\pi)$. Then: \begin{enumerate} \item either $\theta = \alpha 2\pi$ with $\alpha \not \in \mathbb{Q}$ and $\Omega$ is a disk; \item or $\theta = \alpha 2\pi$ with $\alpha \in \mathbb{Q}$ and $\Omega$ admits two (possibly other) axis of symmetry, with respective angle $\frac{\pi}{N}$ with $N\geq 3$. \end{enumerate} \end{proposition} \begin{proof} The proof is divided into three steps. In the following, $\Omega$ is a planar convex set with two axis of symmetry $L_1$ and $L_2$, forming and angle $\theta\in (0,2\pi)$. \emph{Step 1: invariance under rotation.} Since the composition of two symmetries produces the rotation of twice the angle, we infer the following two properties of $\Omega$: \begin{itemize} \item[i)] it is invariant under rotation of angle $2\theta$, denoted by $R_{2\theta}$; \item[ii)] if $L$ is an axis of symmetry, then $R_{2\theta}(L)$ is an axis of symmetry, too. \end{itemize} By repeatedly applying (ii) to $L_1$ and $L_2$, we infer that $R_{2\theta}^k(L_1)$ and $R_{2\theta}^k(L_2)$, $k\in \mathbb N^*$, are axis of symmetry, forming with $L_1$ an angle of $2k\theta$ and $\theta + 2k\theta$, respectively. Therefore, for every $m\in \mathbb N^*$, there exist an axis of symmetry (either a rotation of $L_1$ or a rotation of $L_2$) forming an angle of $m\theta$ with $L_1$. In view of (i) we infer the set $\Omega$ is invariant under all the rotations $R_{2m \theta}$, $m\in \mathbb N^*$. \emph{Step 2: case of $\theta$ non rational multiple of $2\pi$.} Assume that $\theta=\alpha2\pi$, for some $\alpha \in \mathbb R \setminus \mathbb Q$. Then the set $$ \{k\alpha2\pi + 2j\pi \quad : \quad k\in \mathbb{Z} \text{ and } j \in \mathbb{Z}\} $$ is dense in $\mathbb{R}$ (because a subgroup of $\mathbb{R}$ is either discrete or dense, and this one can not be discrete). In this case we deduce that $\Omega$ is invariant under a dense set of rotations, therefore $\Omega$ is a disk, which proves (1). \emph{Step 3: case of $\theta$ rational multiple of $2\pi$.} Let now $\theta = \alpha 2 \pi$ with $\alpha \in \mathbb Q$, say $\alpha=\frac{p}{q}$ with $p$ and $q$ in $\mathbb Z$. Thanks to Bezout Lemma, we can find two integers $a,b\in \mathbb{Z}$ such that $$ap+bq=1.$$ Then we can write $$a\theta=a \alpha 2 \pi = a\frac{p}{q}2\pi=\frac{1}{q}2\pi-b2\pi,$$ namely there exists a multiple of $\theta$ which is equal to $\frac{2\pi}{q}$ mod $2\pi$. By Step~1, this implies that there exist two axis of symmetry with respective angle $\tilde{\theta}:=2\pi/q$, $q\in \mathbb N$. Let $L'$ and $L''$ denote the two axis. Without loss of generality, we may take $L'$ horizontal. If $q$ is even, (2) is proved with $N=q/2$. Let us assume that $q$ is odd, of the form $q=2h+1$, for some $h\in \mathbb N^*$. If $h$ is even, we apply (ii) $h/2$ times to the axis $L'$: the rotated axis $R_{2\tilde{\theta}}^{h/2}(L')$ is again an axis of symmetry, forming with $L'$ two supplementary angles: $$ \frac{h}{2} (2 \tilde{\theta})= \frac{2 h}{2h+1} \pi \quad \hbox{and} \quad \pi - \frac{h}{2} (2 \tilde{\theta}) = \frac{\pi}{2h+1} = \frac{\pi}{q}. $$ The second angle concludes the proof of (2) with $N=q$. The remaining case is when $\tilde{\theta}=2\pi/q$ with $q=2h+1$ and $h$ odd. Here we apply (ii) $(h-1)/2$ times to $L''$: the rotated axis $R_{2\tilde{\theta}}^{(h-1)/2}(L'')$ is again an axis of symmetry, forming with $L'$ two supplementary angles: $$ \tilde{\theta} + \frac{h-1}{2}(2\tilde{\theta}) = \frac{2h}{2h+1} \pi \quad \hbox{and} \quad \pi - \left[\tilde{\theta} + \frac{h-1}{2}(2\tilde{\theta}) \right]= \frac{\pi}{q}, $$ giving (2) with $N=q$. This concludes the proof. \end{proof} Let us now prove the main result when the angle between the two axis of symmetry is small. In view of the previous proposition, this covers almost all the possible cases. \begin{proposition}\label{sym5} Let $\Omega$ be a planar convex domain with two axis of symmetry with angle $\theta= \frac{\pi}{N}$ with $N\geq 5$. Then the inequality \eqref{ineqConj} holds true strictly. \end{proposition} \begin{proof} The strategy of the proof is to compare $\Omega$ with a disk by using the Szeg\H{o}-Weinberger's inequality \eqref{SzegoW}: \begin{eqnarray} \mu_1(\Omega)\leq \frac{\pi(j'_{1,1})^2}{|\Omega|}. \label{SSe} \end{eqnarray} In virtue of \eqref{SSe}, the proposition will be proved if we are able to establish $$\frac{P^2(\Omega)}{|\Omega|}\leq \frac{16\pi}{(j'_{1,1})^2}.$$ Since $(j'_{1,1})^2<3.5$ and $16\pi > 50$, it is enough to prove that \begin{equation}\label{enough} \frac{P^2(\Omega)}{|\Omega|}<14.7. \end{equation} For that purpose we solve some kind of reverse isoperimetric inequality by maximizing the ratio $P^2(\Omega)/|\Omega|$ among convex sets lying in the sector defined by the two axis of symmetry. More precisely, we will bound the perimeter from above and the area from below. Let us assume without loss of generality that the two axis cross at the origin $O$, that the first axis of symmetry is the line $\{x=0\}$, and that the point $A=(0,1)$ belongs to $\partial \Omega$. By assumption, the second axis of symmetry forms the angle $\theta$ with $\{x=0\}$. We denote by $B$ the point on $\partial \Omega$ that belongs to this second axis, and by $\alpha$ the angle $\widehat{OAB}$ as in Fig. \ref{Figure5}. \begin{figure}[h] \begin{center} \includegraphics{dessin5.pdf} \caption{\textit{Proof of Proposition \ref{sym5}.}}\label{Figure5} \end{center} \end{figure} Notice that, by convexity, $B$ must lie between the two points $B_1$ and $B_2$ as in Fig. \ref{Figure5}. which provides the bound $$\frac{\pi}{2}-\theta\leq \alpha \leq \frac{\pi}{2}.$$ We denote by $S$ the positive sector delimited by the half-lines $(OA)$ and $(OB)$, and we consider the restricted sets $$\Omega_0:=\Omega\cap S \quad \text{ and } \quad \Gamma_0:= \partial \Omega \cap S.$$ By symmetry we know that $$|\Omega|=2N |\Omega_0| \quad \text{ and } \quad P(\Omega)= 2 N \mathcal{H}^1(\Gamma_0).$$ The remaining part of the proof is divided into two steps, in which we provide a lower bound for the area and an upper bound for the perimeter, respectively. \emph{Step 1: lower bound for the area.} To estimate $|\Omega_0|$ we will use the fact that, by convexity, it has bigger area than the triangle $OAB$, which turns out to be equal to $\frac{1}{2}x_B$, being $x_B$ the $x$-coordinate of the point $B$. The equation of the line $(AB)$ is $$ \cos(\alpha) x +\sin(\alpha) y- \sin(\alpha)=0, $$ and $B$ also lies on the line $(OB)$ whose equation is $y=\frac{x}{\tan(\theta)}$. We deduce that $$x_B=\frac{\sin(\alpha)\sin(\theta)}{\sin(\theta+\alpha)} \quad \quad \text{ and } \quad \quad y_B=\frac{\sin(\alpha)\cos(\theta)}{\sin(\theta+\alpha)}.$$ Therefore, \begin{eqnarray} |\Omega|= N|\Omega_0|\geq N x_B = N \frac{\sin(\alpha)\sin(\theta)}{\sin(\theta+\alpha)} . \label{lowerbound} \end{eqnarray} \emph{Step 2: upper bound for the perimeter.} Notice that by symmetry, the perpendicular to the line $(OB)$ passing through $B$ is the tangent line to $\partial \Omega$ at $B$. Similarly, the line perpendicular to $\{x=0\}$ through $A$ is the tangent to $\partial \Omega$ at $A$. These two tangents meet at a point $H$ and, by convexity, $\partial \Omega$ stays below the two. This implies that we can estimate the length of $\Gamma_0$ as $\mathcal{H}^1(\Gamma_0)\leq \overline{AH}+\overline{HB}$, see also Fig. \ref{Figure6}. \begin{figure}[h] \begin{center} \includegraphics{dessin6.pdf} \caption{\textit{Proof of Proposition \ref{sym5}.}}\label{Figure6} \end{center} \end{figure} Let us determine the $x$-coordinate $x_H$ of $H$ (the $y$-coordinate is by construction 1). The equation of the line $(HB)$ is given by $$\tan(\theta)(x-x_B)+(y-y_B)=0.$$ Therefore $$x_H=\frac{y_B-1}{\tan(\theta)}+x_B=\frac{\sin(\alpha)}{\sin(\theta)\sin(\theta+\alpha)}-\frac{\cos(\theta)}{\sin(\theta)} =-\cot(\alpha +\theta),$$ and we deduce that $$\overline{AH} = x_H=-\cot(\alpha+\theta).$$ Next, \begin{eqnarray} \overline{BH}^2&=&(x_H-x_B)^2+(1-y_B)^2 \notag \\ &=&\frac{1}{\sin^2(\alpha+\theta)}\Big(\cos(\alpha+\theta)+\sin(\alpha)\sin(\theta)\Big)^2+\Big(\sin(\alpha+\theta)-\sin(\alpha)\cos(\theta)\Big)^2 \notag \\ &=& \frac{1}{\sin^2(\alpha+\theta)}\cos^2(\alpha) \notag \end{eqnarray} from which we deduce $$\overline{BH}=\frac{\cos(\alpha)}{\sin(\theta+\alpha)}.$$ This means that \begin{eqnarray} \mathcal{H}^1(\Gamma_0)\leq \frac{\cos(\alpha)-\cos(\alpha+\theta)}{\sin(\alpha+\theta)}=\frac{2\sin(\frac{\theta}{2})\sin(\alpha+\frac{\theta}{2})}{\sin(\alpha+\theta)}, \label{upperbound} \end{eqnarray} and $P(\Omega)={2}N\mathcal{H}^1(\Gamma_0)$. By combining \eqref{lowerbound} and \eqref{upperbound} we get \begin{eqnarray} \frac{P^2(\Omega)}{|\Omega|} \leq N\frac{{16}\sin^2(\frac{\theta}{2})\sin^2(\alpha+\frac{\theta}{2})}{\sin(\alpha+\theta)\sin(\alpha)\sin(\theta)}={8}N\tan\left(\frac{\theta}{2}\right)\frac{\sin^2(\alpha+\frac{\theta}{2})}{\sin(\alpha+\theta)\sin(\alpha)}. \notag\label{maximiz} \end{eqnarray} Now we maximize in the variable $\alpha \in [\frac{\pi}{2}-\theta , \frac{\pi}{2}]$ the function $$g:\alpha \mapsto \frac{\sin^2(\alpha+\frac{\theta}{2})}{\sin(\alpha+\theta)\sin(\alpha)}.$$ For this purpose we notice that $g$ can be rewritten in the following form $$g(\alpha)=\frac{1-\cos(2\alpha+\theta)}{\cos(\theta)-\cos(2\alpha+\theta)},$$ from which we easily see that $g$ is non increasing on $[ \frac{\pi}{2}-\theta, \frac{\pi}{2}-\frac{\theta}{2}]$ and non decreasing on $[ \frac{\pi}{2}-\frac{\theta}{2}, \frac{\pi}{2}]$. This implies that \begin{eqnarray} \max_{\alpha \in[\frac{\pi}{2}-\theta , \frac{\pi}{2}] }g(\alpha)= g\left(\frac{\pi}{2}\right)=g\left(\frac{\pi}{2}-\theta\right)=\frac{\cos^2(\theta/2)}{\cos(\theta)}, \label{claim} \end{eqnarray} and we deduce that \begin{eqnarray} \frac{P^2(\Omega)}{|\Omega|} \leq {8}N \tan\left(\theta/2\right) \frac{\cos^2(\theta/2)}{\cos(\theta)}={4}N\tan(\theta)={4}N\tan({\pi/N}). \label{boundD} \end{eqnarray} Since $4N\tan\theta$ equals the ratio $P^2(\Omega)/|\Omega|$ for both triangles $OAB_1$ and $OAB_2$ (see Figure \ref{Figure5}) we have actually proved that these two triangles solve the reverse isoperimetric problem in the sector $S$. Next, we study the function $$h:t\mapsto {4}t\tan({\pi/t}).$$ It is easily seen that $h(t)$ is decreasing in $[{2},+\infty[$. Indeed, a direct computation reveals that $h'(t)$ has the same sign as \begin{eqnarray} \sin\left(\frac{\pi}{t}\right)\cos\left(\frac{\pi}{t}\right)-\frac{\pi}{t}. \label{hprime} \end{eqnarray} Then for $t\geq {2}$ we have $0\leq \cos(\frac{\pi}{t})\leq 1$ and $0\leq \sin(\frac{\pi}{t}) \leq \frac{\pi}{t}$ from which we deduce that the expression in \eqref{hprime} is non negative, so that $h'(t)\leq 0$ for $t\geq {2}$. In particular, for every $N\geq 5$, $h(N)\leq h(5)$, thus \eqref{boundD} gives $$ \frac{P^2(\Omega)}{|\Omega|} \leq h(5) = 20 \tan(\pi/5) < 14.54. $$ This concludes the proof thanks to \eqref{enough}. \end{proof} \section{The case of the axis of symmetry of the equilateral triangle} \label{sectionIII} In this section we analyze the maximization over the convex sets with the same axis of symmetry of the equilateral triangle. \begin{proposition}\label{prop-3sym} Let $\Omega$ be a planar convex domain with two axis of symmetry with angle $\pi /3$. Then the inequality \eqref{ineqConj} holds true. The equality occurs for all and only equilateral triangles. \end{proposition} \begin{proof} The proof is detailed in the next pages and occupies the whole section. Let us summarize it here: in Lemma \ref{Ca2} we provide a representation result, by dividing the class of domains under study into a 1-parameter family $\mathcal C_a$, with $a$ varying in the interval $[0,1/3]$, being $a=0$ associated to all and only the equilateral triangles. Then, in \S \ref{ss1} we prove the statement for $a\in [0,1/4]$. Finally, in \S \ref{ss2} we treat the case $a\in [1/4, 1/3]$. \end{proof} \subsection*{The family $\mathcal C_a$} We denote by $T$ the equilateral triangle with side 1 and vertexes at $(\pm 1/2,0)$ and $(0,\sqrt{3}/2)$. For $a\in [0,1/2]$ we denote by $\hat{T}_a$ the equilateral triangle bounded by the lines $$ y=\frac{\sqrt{3}}{2}(1-a), \quad y= \pm \sqrt{3}x - \sqrt{3} \left(\frac12 - a\right). $$ These triangles share the same axis of symmetry of $T$, but have the horizontal side on the top instead of the bottom. For the extremal values of $a$, namely $a=0, 1/2$, $\hat{T}_a$ is either circumscribed to $T$ or inscribed into it. We denote by $\widehat{T}$ the inscribed equilateral triangle $\hat{T}_{1/2}$. For the intermediate values of $a$, namely $0<a<1/2$, $\hat{T}_a$ crosses $T$. We define the intersection as \begin{equation}\label{Oa} \Omega_a:= \hat{T}_a \cap T \end{equation} that is, the hexagon with vertexes $$ \left(\pm \frac{a}{2}; \frac{\sqrt{3}}{2}(1-a)\right), \quad \left(\pm \left(\frac12 -a\right); 0\right), \quad \left(\pm \frac12 (1-a); \frac{\sqrt{3}}{2}a\right). $$ By connecting the midpoints of the sides $\Omega_a$, we obtain an hexagon, denoted by $H_a$, with vertexes $$ \left(0; \frac{\sqrt{3}}{2}(1-a)\right), \quad \left( \pm \frac14, \frac{\sqrt{3}}{4}\right), \quad \left(\pm \left(\frac12 -\frac34 a\right); \frac{\sqrt{3}}{4} a\right), \quad \left(0,0\right). $$ These domains are represented in Fig. \ref{fig-domains}. \begin{figure}[h] \begin{center} {\includegraphics[height=3truecm] {fig-TThat.pdf}}\quad {\includegraphics[height=3truecm] {fig-Oa.pdf}}\quad {\includegraphics[height=3truecm] {fig-Ha.pdf}} \end{center} \caption{\textit{The equilateral triangles $T$ and $\hat{T}_{1/2}$, and the hexagons $\Omega_a$ and $H_a$ for $a= 0.15$.}}\label{fig-domains} \end{figure} We are now in a position to introduce the subfamily $\mathcal C_a$. \begin{definition} Let $a\in [0,1/2]$. We say that $\Omega\in \mathcal C_a$ if the following conditions hold: \begin{itemize} \item[i)] $\Omega$ is convex and has the same axis of symmetry of $T$; \item[ii)] $H_a \subset \Omega\subset \Omega_a$. \end{itemize} \end{definition} A priori, $a\in [0,1/2]$. However, as we prove in the next lemma, the interval $[0,1/3]$ is enough to describe all the domains under study. \begin{lemma}\label{Ca2} For every convex set $\Omega$ with the same axis of symmetry of the equilateral triangle, there exists $a\in {[0,1/3]}$ such that, up to a rigid motion and a homothety, $\Omega$ belongs to $\mathcal C_a$. \end{lemma} \begin{proof} Let $\Omega$ be as in the assumption. Throughout the proof, with a slight abuse of notation, we will use the same letter for the class of shapes that are obtained by a rigid motion or a homothecy by $\Omega$. Without loss of generality, we may assume that $\Omega$ has the same axis of symmetry of $T$, it lies into the half-plane $y\geq 0$, and it is tangent to the horizontal line $y=0$, as it occurs for $T$. By combining the symmetry assumption on $\Omega$ with the (sharp) bound $y\geq 0$ of its points, we infer that $\Omega$ is contained into $T$ and touches the three segments of $\partial T$ at the midpoints. In particular, by convexity, the triangle $\hat{T}_{1/2}$ is contained into $\Omega$. Let now $r$ be the horizontal line tangent to $\Omega$ from above. In view of the inclusions $\hat{T}_{1/2}\subset \Omega \subset T$, we infer that $r$ is of the form $y=(1-a)\sqrt{3}/2$ for some $a\in [0,1/2]$. Arguing as above, we infer that $\Omega$ is contained into the equilateral triangle $\hat{T}_a$ and touches the three segments of $\partial \hat{T}_a$ at the midpoints. All in all, we infer that $\Omega$ is contained into $T\cap \hat{T}_a=\Omega_a$ and that it touches $\partial \Omega_a$ at the midpoints of the boundary segments, which by definition are the vertexes of $H_a$. By convexity, $H_a\subset \Omega$. In other words $\Omega\in \mathcal C_a$. In order to conclude the proof, we need to show that $a$ can be taken in $[0,1/3]$ instead of $[0,1/2]$. To this aim we use a very simple trick, which, roughly speaking, consists in changing the role of $T$ and $\hat{T}$. The class $\mathcal C_a$ is constructed starting from the hexagon $\Omega_a$, which is the intersection between $T$ and $\hat{T}_a$. The parameter $a$ is nothing but the length of the portions of $\partial \hat{T}_a$ included into $T$. The same hexagon can be described in terms of the length, say $b$, of the portions of $\partial T$ included into $\hat{T}_a$. In other words, changing the role of $T$ and $\hat{T}_a$, we may write $\Omega_a$ as an $\Omega_b$, for $b=1-2a$. If a shape $\Omega$ belongs to the class $\mathcal C_a$ associated to $T$, then the rescaled shape $(1/(2-3a))\Omega$ belongs to $\mathcal C_{(1-2a)/(2-3a)}$ associated to the triangle $(1/(2-3a))\hat{T}_a$. The triangle $(1/(2-3a))\hat{T}_a$ is equilateral and has sides of length 1. For $a\in [1/3, 1/2]$, the parameter $(1-2a)/(2-3a)$ runs from $0$ to $1/3$. This concludes the proof. \end{proof} \subsection*{{Description of shapes in $\mathcal C_a$}} In this paragraph we give some definitions and properties to represent the symmetric shapes of $\mathcal C_a$. We start with a useful formula to make integrations over symmetric sets. Let $\ell_0$ denote the vertical line $x=0$, $\ell_1$ the line $-x + \sqrt{3} y = 1/2$, and $\ell_2$ the line $x + \sqrt{3} y = 1/2$. Let $\sigma_1$ and $\sigma_2$ denote the symmetries with respect to $\ell_1$ and $\ell_2$, respectively: they are given by the transformations \begin{equation}\label{sigmai} \sigma_1: \begin{cases} x' = \frac12 x + \frac{\sqrt{3}}{2} y -\frac14 \\ y' = \frac{\sqrt{3}}{2} x - \frac12 y + \frac{\sqrt{3}}{4} \end{cases}\quad \sigma_2: \begin{cases} x' = \frac12 x - \frac{\sqrt{3}}{2} y +\frac14 \\ y' = -\frac{\sqrt{3}}{2} x - \frac12 y + \frac{\sqrt{3}}{4}. \end{cases} \end{equation} Any set $E$ with the axis of symmetry $\ell_0, \ell_1, \ell_2$ can be written as the disjoint union of three components: $\tilde{E} \cup \sigma_1(\tilde{E}) \cup \sigma_2(\tilde{E})$, for some set $\Tilde{E}$ symmetric with respect to the vertical axis. As a consequence, we get for every function $f$ \begin{equation}\label{trick} \int_{E} f = \int_{\tilde{E} \cup \sigma_1(\tilde{E}) \cup \sigma_2(\tilde{E})} f = \int_{\tilde{E}} (f + f\circ \sigma_1 + f\circ \sigma_2). \end{equation} This formula applies in particular to the elements of $\mathcal C_a$. \medskip Let us exploit again the three axis of symmetry to represent the elements of $\mathcal C_a$. We start from $\Omega_a$: it is the union of the hexagon $H_a$ and 6 copies (obtained by reflection) of the triangle $ABC$, with vertexes \begin{equation}\label{ABC} A=\left(0, \frac{\sqrt{3}}{2}(1-a) \right),\quad B=\left(\frac{a}{2}, \frac{\sqrt{3}}{2}(1-a) \right), \quad C=\left(\frac14, \frac{\sqrt{3}}{4} \right). \end{equation} Any other element $\Omega \in \mathcal C_a$ is the union of $H_a$ and 6 copies (reflections) of the convex set \begin{equation}\label{omega} \omega:=\Omega \cap ABC. \end{equation} Of course $\omega=\emptyset$ if $\Omega= H_a$. The perimeter of $\Omega$ is nothing but \begin{equation}\label{Pomega} P(\Omega) = 6 [P(\omega) - \overline{AC}]. \end{equation} Going back to \eqref{trick}, we deduce an alternative formula of integration in the case of $f$ even in $x$, i.e., $f(-x,y)=f(x,y)$: \begin{equation}\label{trick2} \int_{\Omega} f = \int_{H_a} f + 2 \int_\omega (f + f\circ \sigma_1 + f \circ \sigma_2). \end{equation} \subsection{The case $a\in [0,1/4]$}\label{ss1} As announced in the Introduction, this is the more delicate case to handle. The difficulty comes from calculations; on the other hand, the strategy is simple to present. Given $\Omega\in \mathcal C_a$, we look for three objects: an approximating convex set from outside $\Omega_1 \supset \Omega$, an approximating convex set from inside $\Omega_2 \subset \Omega$, and a function $v\in C^\infty(\mathbb R^2)$ with zero average on $\Omega$. By construction, we have \begin{equation}\label{firstub} P^2(\Omega)\mu_1(\Omega) \leq P^2(\Omega_1) \frac{\int_{\Omega_1} |\nabla v|^2}{\int_{\Omega_2} v^2}. \end{equation} The goal is to choose the three objects in a smart way, making this upper bound less than or equal to $16\pi^2$. \medskip We start with the choice of the test function. We take \begin{equation}\label{vk} v_k:=(1-k)u_1+k\hat{u}_1, \end{equation} with $k\in [0,1]$, which will be chosen later, with $u_1$ one of the first eigenfunctions of $T$ associated to $\mu_1(T)$: \begin{equation}\label{u1} u_1(x,y)= \sin\left(\frac43 \pi x \right) + 2 \cos\left(\frac{2}{\sqrt{3}}\pi y \right) \sin \left( \frac23 \pi x\right), \end{equation} and with $\hat{u}_1$ one of the first eigenfunctions of $\widehat{T}$ associated to $\mu_1(\hat{T})$: \begin{equation}\label{u1hat} \hat{u}_1(x,y)=\sin\left(\frac{8}{3} \pi x\right) - 2\cos\left(\frac{4}{\sqrt{3}} \pi y\right) \sin\left(\frac{4}{3} \pi x\right). \end{equation} Note that $u_1(-x,y)=-u_1(x,y)$ and $\hat{u}_1(-x,y)=-\hat{u}_1(x,y)$, therefore $v_k$ is odd too in the $x$ variable and has zero average in $\Omega$, which is symmetric with respect to the $y$ axis. This ensures that $v_k$ is an admissible test function for $\mu_1(\Omega)$. Let us now pass to the construction of $\Omega_1$ and $\Omega_2$. Let $\omega$ be the set associated to $\Omega$ according to \eqref{omega}, namely $\omega=\Omega \cap ABC$. Take now the line parallel to $AC$ tangent to $\omega$. This line intersects the sides $AB$ and $BC$ in two points (possibly coinciding): we call $Q_1$ the point on the segment $AB$ and $Q_2$ the point on the side $BC$. We introduce a parameter to describe these points: $Q_2$ is of the form $$ Q_2(c)=\left( \frac{c}{2}; \frac{\sqrt{3}}{2} (1-c)\right), \quad \hbox{for some } c\in [a,1/2]. $$ Note that $B=Q_2(a)$ and $C=Q_2(1/2)$. Using the expression of $Q_2$ we can deduce the coordinates of $Q_1$ as functions of $c$. Let us now consider one of the tangent points of $Q_1(c)Q_2(c)$ to $\omega$: it is of the form $$ Q(s)=(1-s)Q_1(c) + s Q_2(c), \quad s\in [0,1]. $$ \begin{figure}[h] \begin{center} {\includegraphics[height=5truecm] {figure88.png}} \caption{The parameters $c$ and $s$}\label{figuretets1} \end{center} \end{figure} These definitions are represented in Figure \ref{figuretets1}. By convexity $\omega$ contains the triangle $AQ(s)C$ and is contained into the trapezoid $AQ_1(c)Q_2(c)C$. In particular, we may choose as $\Omega_1$ the union of $H_a$ and 6 copies (rotations) of the trapezoid $AQ_1(c)Q_2(c)C$, and as $\Omega_2$ the union of $H_a$ and 6 copies (rotations) of the triangle $AQ(s) C$. With this choice of $\Omega_1$, using the perimeter formula \eqref{Pomega}, we obtain $$ P(\Omega) \leq P(\Omega_1) = 6 [ \overline{AQ_1(c)} + \overline{Q_1(c)Q_2(c)} + \overline{Q_2(c)C} ]. $$ By applying the integration formula \eqref{trick2} to $f=|\nabla v_k|^2$ and then $f=|v_k|^2$, satisfying both $f(-x,y)=f(x,y)$, we get \begin{align*} & \int_{\Omega_1} |\nabla v_k|^2 = \int_{H_a} |\nabla v_k|^2 + 2 \int_{AQ_1(c)Q_2(c)C} \left[|\nabla v_k|^2 + |\nabla v_k|^2\circ \sigma_1 + |\nabla v_k|^2\circ \sigma_2\right], \\ & \int_{\Omega_2} v_k^2 = \int_{H_a} v^2 + 2 \int_{AQ(s)C}\left[ |v_k|^2 + |v_k|^2\circ \sigma_1 + |v_k|^2\circ \sigma_2\right]. \end{align*} By combining these results, setting for brevity \begin{align} V & :=|\nabla v_k|^2 + |\nabla v_k|^2\circ \sigma_1 + |\nabla v_k|^2\circ \sigma_2, \label{defV0} \\ U & := |v_k|^2 + |v_k|^2\circ \sigma_1 + |v_k|^2\circ \sigma_2,\label{defU0} \end{align} we can rewrite the bound \eqref{firstub} as follows: \begin{equation}\label{estimate} \begin{split} &{P^2(\Omega) \mu_1(\Omega)} \\ &{\leq \left[6\, \left( \overline{AQ_1(c)} + \overline{Q_1(c)Q_2(c)} + \overline{Q_2(c)C} \right)\right] ^2 \frac{\int_{H_a} |\nabla v_k|^2 + 2 \int_{AQ_1(c)Q_2(c)C} V}{ \int_{H_a} |v_k|^2 + 2 \int_{AQ(s)C}U}.} \end{split} \end{equation} \medskip We explain now our strategy to prove that the right-hand side of \eqref{estimate} is less than or equal to $16\pi^2$: {\bf 1st step.} We prove that the integral $\int_{AQ(s)C}U$ appearing in the denominator is decreasing in $s$, at least when $0\leq k\leq 1/6$. Therefore, to find the upper bound, we can choose $s=1$, namely $Q(s)=Q_2(c)$. More precisely, denoting by $\widetilde{P}(a,c):=6\, \big(\overline{AQ_1(c)} + \overline{Q_1(c) Q_2(c) } + \overline{Q_2(c) C} \big)$, $N(a,c):=\int_{H_a} |\nabla v_k|^2 + 2 \int_{AQ_1(c)Q_2(c)C} V$, and $D(a,c):=\int_{H_a} v_k^2 + 2 \int_{AQ_2(c)C}U$, we obtain $$ P^2(\Omega)\mu_1(\Omega) \leq \left(\widetilde{P}(a,c)\right)^2 \frac{N(a,c)}{D(a,c)}. $$ Note that the upper bound is less than or equal to $16 \pi^2$ if and only if \begin{equation}\label{defF} F(a,c):=\left(\widetilde{P}(a,c)\right)^2 N(a,c) -16 \pi^2 D(a,c) \leq 0. \end{equation} \smallskip {\bf 2nd step.} By precise estimates, we prove that $F(a,c) \leq 0$ for all $a\leq 1/60$ (and for all $c$) or for all $c\leq 0.16$ (and for all $a$) if we choose $k=0$, namely as a test function $v_0=u_1$. \smallskip {\bf 3rd step.} For the remaining cases, $a\in [1/60,1/4]$, we provide a more numerical proof. First we compute exactly all the integrals occurring in $N(a,c)$ and $D(a,c)$. Then, by estimating the partial derivatives $\frac{\partial F}{\partial a}$ and $\frac{\partial F}{\partial c}$, we are led to compute $F(a,c)$ only in a finite number of points, namely 530\,000 points in the sector $a\in [1/60,1/4]$, $c\in [\max(a,0.16),1/2]$. For that purpose, we use successively as test functions $v_k$ for $k=0$, $k=0.06$ and $k=0.12$. We prove that way the desired inequality $F(a,c)<0$. The most technical parts of this program are postponed to the Appendix. \subsection*{1st step: getting rid of the parameter $s$.} We start with some preliminary results on the functions $U,V$ and on their integrals over $T$ and $\Omega_a$. In view of definition \eqref{vk}, the functions $V$ and $U$ introduced in \eqref{defV0}-\eqref{defU0} are of the form \begin{eqnarray} &V=(1-k)^2 V_1 +k^2 V_2 +2k(1-k) V_3,\label{defVbis} \\ &U=(1-k)^2 U_1 +k^2 U_2 +2k(1-k) U_3,\label{defUbis} \end{eqnarray} with \begin{equation}\label{defVall} \begin{array}{l} V_1(x,y):= |\nabla u_1|^2 (x,y) + |\nabla u_1|^2( \sigma_1(x,y)) + |\nabla u_1|^2(\sigma_2(x,y)), \\ V_2(x,y):= |\nabla \hat{u}_1|^2 (x,y) + |\nabla \hat{u}_1|^2( \sigma_1(x,y)) + |\nabla \hat{u}_1|^2(\sigma_2(x,y)), \\ V_3(x,y):= \nabla u_1.\nabla \hat{u}_1 (x,y) + \nabla u_1.\nabla \hat{u}_1( \sigma_1(x,y)) + \nabla u_1.\nabla \hat{u}_1(\sigma_2(x,y)), \end{array} \end{equation} and \begin{equation}\label{defUall} \begin{array}{l} U_1(x,y):=u_1^2 (x,y) + u_1^2( \sigma_1(x,y)) + u_1^2(\sigma_2(x,y)) , \\ U_2(x,y):=\hat{u}_1^2 (x,y) + \hat{u}_1^2( \sigma_1(x,y)) + \hat{u}_1^2(\sigma_2(x,y)) , \\ U_3(x,y):=u_1 \hat{u}_1 (x,y) + u_1 \hat{u}_1( \sigma_1(x,y)) + u_1 \hat{u}_1(\sigma_2(x,y)) .\\ \end{array} \end{equation} \begin{lemma} \label{lemma8.2} Let $V_i$ and $U_i$, $i=1,2,3$ be the functions defined in \eqref{defVall}-\eqref{defUall}, with $\sigma_1$ and $\sigma_2$ the symmetries introduced in \eqref{sigmai}. Let us introduce the following functions: \begin{equation}\label{fipsieta} \begin{array}{l} \varphi(x,y)=\cos \frac{4\pi y}{\sqrt{3}} - 2 \cos \frac{2\pi y}{\sqrt{3}} \cos 2\pi x, \\ \psi(x,y)=\cos 4\pi x - 2\cos \frac{6\pi y}{\sqrt{3}} \cos 2\pi x, \\ \eta(x,y)=\cos \frac{8\pi y}{\sqrt{3}} + 2 \cos \frac{4\pi y}{\sqrt{3}} \cos 4\pi x. \end{array} \end{equation} Then \begin{eqnarray*} V_1(x,y) &=& \frac{8\pi^2}{3} \left( 3 - \varphi(x,y)\right),\label{V1}\\ V_2(x,y) &=& \frac{32\pi^2}{3} \left( 3 - \eta(x,y)\right),\label{V2}\\ V_3(x,y) &=& \frac{16\pi^2}{3} \left( \psi(x,y) - \varphi(x,y)\right),\label{V3}\\ U_1(x,y) &=& 3 \left( \frac32 + \varphi(x,y)\right), \label{U1}\\ U_2(x,y) &=& 3 \left( \frac32 + \eta(x,y)\right), \label{U2}\\ U_3(x,y) &=& -3 \left( \frac12 \psi(x,y) + \varphi(x,y)\right). \label{U3}\\ \end{eqnarray*} \end{lemma} \begin{proof} The proof is straightforward. For example, for $u_1$ using the expressions of the symmetries $\sigma_1, \sigma_2$ given in \eqref{sigmai}, we get \begin{equation}\label{u1sig1} u_1(\sigma_1(x,y))=-u_1(x+1,y)\quad u_1(\sigma_2(x,y))=-u_1(x-1,y) . \end{equation} Now since $u_1^2$ can be written $$u_1^2(x,y)=\frac{1}{2}-\frac{1}{2} \cos \frac83 \pi x +2 \cos \frac{2\pi y}{\sqrt{3}} \big(\cos \frac23 \pi x - \cos 2\pi x \big)+ \big(1+\cos \frac{4\pi y}{\sqrt{3}} \big)\big(1-\cos \frac43 \pi x \big)$$ the formula for $U_1$ follows thanks to the property $\cos(a)+\cos(a+2\pi/3)+\cos(a+4\pi/3)=0$ that holds for any $a$.\\ The same argument is used for all the other functions. \end{proof} The first application of the previous lemma is the monotonicity of the integrand $U$, which allows us to get rid of the parameter $s$. \begin{lemma}\label{lemmas} Let $U$ be the function defined in \eqref{defUbis}, namely $U=(1-k)^2 U_1 +k^2 U_2 +2k(1-k) U_3$ with $U_i$ introduced in \eqref{defUall}. Then, for all $k\in [0,1/6]$ its derivatives satisfy $$\frac{\partial U}{\partial x} \leq 0 \qquad \frac{\partial U}{\partial y} \geq 0.$$ Consequently, the function $s\mapsto \int_{AQ(s)C} U(x,y)\mathrm{d} x\mathrm{d} y$ is decreasing and, for an upper bound of the Rayleigh quotient, we can choose $s=1$, namely $Q(s)=Q_2(c)$. \end{lemma} The proof is postponed to the Appendix. We conclude the paragraph with some useful formulas, which are deduced by combining the integration formula \eqref{trick} together with the representation result of Lemma \ref{lemma8.2}. \begin{lemma}\label{lem-F12} \begin{align*} \int_{T\setminus \Omega_a} u_1^2 \mathrm{d}x\mathrm{d}y & = \frac{9 \sqrt{3}a^2}{8} + \frac{3\sqrt{3}}{8\pi^2}(1-\cos 2\pi a + 2\pi a \sin 2\pi a)=:F_1(a), \\ \int_{T\setminus \Omega_a} |\nabla u_1|^2 \mathrm{d}x\mathrm{d}y &= 2\sqrt{3} \pi^2 a^2 - \frac{\sqrt{3}}{3}(1-\cos 2\pi a + 2\pi a \sin 2\pi a)=:F_2(a),\\ \int_{T\setminus \Omega_a} \hat{u}_1^2 \mathrm{d}x\mathrm{d}y & = \frac{9 \sqrt{3}a^2}{8} + \frac{3\sqrt{3}}{32\pi^2}(1-\cos 4\pi a + 4\pi a \sin 4\pi a)=:F_3(a),\\ \int_{T\setminus \Omega_a} |\nabla \hat{u}_1|^2 \mathrm{d}x\mathrm{d}y &= 8\sqrt{3} \pi^2 a^2 - \frac{\sqrt{3}}{3}(1-\cos 4\pi a + 4\pi a \sin 4\pi a)=:F_4(a),\\ \int_{T\setminus \Omega_a} u_1 \hat{u}_1 \mathrm{d}x\mathrm{d}y & = \frac{3\sqrt{3}}{16\pi^2}(\cos 2\pi a +\cos 4\pi a -2 -4\pi a \sin 2\pi a)=:F_5(a), \\ \int_{T\setminus \Omega_a} \nabla u_1 .\nabla \hat{u}_1 \mathrm{d}x\mathrm{d}y & =\frac{2\sqrt{3}}{3}(2\cos 2\pi a -1 -\cos 4\pi a - 2\pi a \sin 2\pi a)=:F_6(a). \end{align*} \end{lemma} \begin{proof} Let $T_a$ be the triangle located at the north of $T$, namely the triangle with vertexes $\left(\pm \frac{a}{2}; \frac{\sqrt{3}}{2}(1-a)\right)$ and $(0,\frac{\sqrt{3}}{2})$. According to the symmetry property \eqref{trick}, we have for example $$\int_{T\setminus \Omega_a} u_1^2\mathrm{d} x\mathrm{d} y = \int_{T_a} U_1(x,y)\mathrm{d} x\mathrm{d} y, \quad \mbox{and} \quad \int_{T\setminus \Omega_a} |\nabla u_1|^2\mathrm{d} x\mathrm{d} y = \int_{T_a} V_1(x,y)\mathrm{d} x\mathrm{d} y, $$ and Lemma \ref{lem-F12} follows from a straightforward computation of these integrals on the triangle $T_a$ using Lemma \ref{lemma8.2}. The computation is the same for the other functions. \end{proof} \begin{remark}\label{remT} Without making use of the reflections, we can integrate over the whole $T$ the following quantities: \begin{equation}\label{Ray-u1} \int_T |\nabla u_1|^2 \mathrm{d}x\mathrm{d}y = \frac{2\pi^2}{\sqrt{3}}, \quad \int_T u_1^2 \mathrm{d}x\mathrm{d}y = \frac{3\sqrt{3}}{8}, \end{equation} \begin{equation}\label{Ray-u1hat} \int_T |\nabla \hat{u}_1|^2 \mathrm{d}x\mathrm{d}y = \frac{8\pi^2}{\sqrt{3}}, \quad \int_T \hat{u}_1^2 \mathrm{d}x\mathrm{d}y = \frac{3\sqrt{3}}{8}, \end{equation} \begin{equation}\label{mixte} \int_T \nabla u_1 \cdot \nabla \hat{u}_1\mathrm{d}x\mathrm{d}y=0 , \quad \int_T u_1\hat{u}_1\mathrm{d}x\mathrm{d}y=0. \end{equation} Notice that the former allow us to recover $\mu_1(T)={\int_T |\nabla u_1|^2}/{\int_T u_1^2} = 16 \pi^2 /9$. \end{remark} \subsection*{2nd step: small values of $a$.} We want to prove that, choosing the test function $u_1$, the quantity $F(a,t)$ defined in \eqref{defF} is negative for small values of $a$, $a \neq 0$ (whereas it vanishes for $a=0$). We are going to prove \begin{lemma}\label{smalla} The function $F$ defined in \eqref{defF} satisfies $$F(a,c) \leq 0\quad \mbox{ when $a\in [0,1/60]$ (for any $c$) or when $c\in [0,0.16]$ (for any $a$)}.$$ Equality occurs only for $a=0$, namely for the equilateral triangle. \end{lemma} The proof of the Lemma is postponed to the Appendix. \subsection*{3rd step: computer assisted proof} In view of Lemma \ref{smalla}, it remains us to prove the inequality $F(a,c) < 0$ in the zone \begin{equation}\label{defZ} \mathcal{Z}:=\{(a,c)\ :\ 1/60 \leq a \leq 1/4,\, \max(a,0.16) \leq c \leq 1/2\}. \end{equation} \begin{figure}[h!] \includegraphics[scale=0.25]{zone1.jpg} \caption{The zone $1/60 \leq a \leq 1/4, \max(a,0.16) \leq c \leq 1/2$ and the three sub-zones $\mathcal Z_{I}$, $\mathcal Z_{II}$, $\mathcal Z_{III}$.}\label{figZ} \end{figure} To this aim, we divide $\mathcal Z$ into three subregions, according to the value of $a$, see also Fig. \ref{figZ}: \begin{equation}\label{defZi} \begin{array}{l} \mathcal Z_I := \mathcal Z \cap \{a\in [1/60, 0.06]\}, \\ \mathcal Z_{II} := \mathcal Z \cap \{a\in [0.06, 0.12]\}, \\ \mathcal Z_{III} := \mathcal Z \cap \{a\in [0.12, 0.25]\}. \end{array} \end{equation} In each regime, we choose different values for the parameter $k$: $$ \begin{cases} k=0 \quad & \hbox{in }\mathcal Z_I \\ k=0.06 \quad & \hbox{in }\mathcal Z_{II} \\ k=0.12 \quad & \hbox{in }\mathcal Z_{III}. \end{cases} $$ We recall that $k$ appears in the definition \eqref{vk} of the test function $v_k$, and thus in the definition \eqref{defF} of $F$. In the next lemma we bound the modulus of the partial derivatives of $F$ (with respect to $a$ and $c$) in each of the three regimes. \begin{lemma}\label{estimategrad} Let $F(a,c)$ be the function introduced in \eqref{defF} and let $\mathcal Z_{I}$, $\mathcal Z_{II}$, $\mathcal Z_{III}$ be the regions defined in \eqref{defZi}. Then $$ \left| \frac{\partial F}{\partial a} (a,c) \right| \leq \begin{cases} 819.6011 \quad & \hbox{in }\mathcal Z_{I} \\ 1048.9639 \quad & \hbox{in }\mathcal Z_{II} \\ 1353.8951 \quad & \hbox{in }\mathcal Z_{III} \end{cases} , \quad \left| \frac{\partial F}{\partial c}(a,c) \right| \leq \begin{cases} 84.4817 \quad & \hbox{in }\mathcal Z_{I} \\ 170.9884 \quad & \hbox{in }\mathcal Z_{II} \\ 352.1112 \quad & \hbox{in }\mathcal Z_{III} \end{cases}. $$ \end{lemma} The proof of the Lemma is postponed to the Appendix. Let us now explain the computer assisted part. We start by considering the rectangular regions $\mathcal Z_I$ and $\mathcal Z_{II}$, which are of the form $[a_{min}, a_{max}]\times[c_{min}, c_{max}]$. Let for brevity $\mathcal R$ denote one of these rectangles. Given $n_a, n_c\in \mathbb N$ and setting $$ \delta_a=\frac{a_{max}-a_{min}}{n_a}, \ \delta_c=\frac{c_{max}-c_{min}}{n_c}, $$ the family of points $$ \{(a_i,c_j)\}_{i=0,\ldots, n_a\,,j=0, \ldots, n_c}, \quad a_i= a_{min} + i\delta_a, \ c_j= c_{min} + j \delta_c, $$ forms a grid in $\mathcal R$ and defines a uniform partition of $\mathcal R$ into $n_a \cdot n_c$ rectangles. Given a pair $(a,c)\in \mathcal R$, there exist $i,j$ such that $|a-a_i| \leq \delta_a/2$, $|c-c_j|\leq \delta_c/2$. Therefore, we deduce that for a pair $(a,c)\in \mathcal R$ and the associated grid point $(a_i,c_j)\in \mathcal R$, there holds \begin{equation}\label{boundF} F(a,c) \leq F(a_i, c_j) + \|\partial_a F\|_\infty \delta_a/2 + \|\partial_c F\|_\infty \delta_c/2. \end{equation} Denoting by $\mathcal E$ the remainder $$ \mathcal E:= \|\partial_a F\|_\infty \delta_a/2 + \|\partial_c F\|_\infty \delta_c/2, $$ we deduce from \eqref{boundF} that for every $(a,c) \in \mathcal R$, \begin{equation}\label{boundF2} F(a,c) \leq \max_{(a_i, c_j )\in \mathcal R}F(a_i, c_j) + \mathcal E. \end{equation} The same conclusion holds true for the region $\mathcal Z_{III}$, but the grid has to be slightly modified: we start from the grid associated to the larger rectangle $[a_{min}, a_{max}]\times[c_{min}, c_{max}]$; then we refine the partition of the rectangles which cross the bisector line $c=a$, by adding the midpoints of the boundary segments and the intersection point of the diagonals; finally, we remove the points $(a_i, c_j)$ that fall outside $\mathcal Z_{III}$, namely the ones for which $c_j < \max(a_i, 0.16)$. The second operation is needed in order to guarantee that every point of $\mathcal Z_{III}$ stays at distance at most $\delta_a/2$ from one of the $a_i$s and at most $\delta_c/2$ from one of the $c_j$s. We now bound from above the two terms appearing in the right-hand side of \eqref{boundF2}, by choosing different values of $n_a$, $n_c$, thus of $\delta_a$ and $\delta_c$: in $\mathcal Z_{I}$ we take $(n_a,n_c)=(155,150)$, in $\mathcal Z_{II}$ we take $(n_a,n_c)=(160,155)$, and in the rectangle containing $\mathcal Z_{III}$ we take $(n_a, n_c)=(730,735)$. This corresponds to a non-uniform grid of $\mathcal Z$ of about 530\,000 points (we recall that in $Z_{III}$ we neglect the points in the half plane $c<a$). Using the upper bounds found in Lemma \ref{estimategrad}, we get $$ \mathcal E \leq \begin{cases} 0.21032 \quad & \hbox{in }\mathcal Z_{I} \\ 0.38422 \quad & \hbox{in }\mathcal Z_{II} \\ 0.202 \quad & \hbox{in }\mathcal Z_{III} \end{cases} $$ The numerical computation in the 530\,000 grid points gives $$ F(a_i,c_j) \leq \begin{cases} -0.21184 \quad & \hbox{in }\mathcal Z_{I} \\ -0.39006 \quad & \hbox{in }\mathcal Z_{II} \\ -0.20324 \quad & \hbox{in }\mathcal Z_{III}. \end{cases} $$ Using these estimates in \eqref{boundF2}, we conclude that $$ F(a,c) \leq \begin{cases} -0.00152 \quad & \hbox{in }\mathcal Z_{I} \\ -0.00584 \quad & \hbox{in }\mathcal Z_{II} \\ -0.00124 \quad & \hbox{in }\mathcal Z_{III} \end{cases} $$ namely $F$ is negative everywhere. This concludes the proof. \begin{remark} We underline that the numerical computation of $F$ is performed on its explicit expression as a function of $k$, $a$, and $c$, involving only trigonometric functions and square roots. Indeed, the integrals appearing in $N$ and $D$ can be computed as linear combinations of integrals over the whole $T$ of $u_1^2, \hat{u}_1^2, |\nabla u_1|^2, |\nabla \hat{u}_1|^2$ (for the explicit expressions, see Remark \ref{remT}), integrals over $T\setminus \Omega_a$ of $u_1^2$, $\hat{u}_1^2$, $u_1 \hat{u_1}$, $|\nabla u_1|^2$, $|\nabla \hat{u}_1|^2$, $\nabla u_1\cdot \nabla \hat{u_1}$ (see Lemma \ref{lem-F12}), and integrals over the triangles $T_N=Q_1(c)BQ_2(c)$ and $T_D=ABQ_2(c)$ of $\varphi, \psi, \eta$ (see Remark \ref{remInt} in the Appendix). \end{remark} \subsection{The case $a\in [1/4, 1/3]$}\label{ss2} This case is ruled out by the Szeg\H{o}-Weinberger inequality, already mentioned in \eqref{SzegoW}. This inequality states that the disk maximizes $\mu_1$ among sets of given area, in particular \begin{equation}\label{ratio} P^2(\Omega) \mu_1(\Omega) = \frac{P^2(\Omega)}{|\Omega|}\mu_1(\Omega)|\Omega|\leq \frac{P^2(\Omega)}{|\Omega|} \pi (j'_{1,1})^2, \end{equation} where $j'_{1,1}$ is the first zero of the derivative of the first Bessel function $J_1$. In order to conclude, we need to solve, for every $a\in [1/4, 1/3]$, $$ \max_{\Omega \in \mathcal C_a} \frac{P^2(\Omega)}{|\Omega|}. $$ This shape optimization problem is of isoperimetric type, since one can fix the perimeter of the admissible shapes. Contrary to the classical isoperimetric problem, here the area functional is minimized. In this sense, we call such a problem a reverse isoperimetric problem. The study of this kind of functional is interesting in itself, therefore we state a more general result. We also point out that the statement is valid for $a$ in a wider range. \begin{proposition}[A reverse isoperimetric problem]\label{prop-reverse} Let $a \in [0,1/2]$. Let $\Psi,\Phi:\mathbb [0,+\infty[ \to \mathbb [0,+\infty[$ be two continuous functions with $\Psi$ increasing. Let $\omega$ be a convex subset of the triangle $ABC$ defined in \eqref{ABC} whose boundary is the union of the segment $AC$ with a curve $\Gamma$. Then the maximum \begin{equation}\label{form} \max_{\omega} \left\{\Psi(\mathcal H^1(\Gamma))\Phi(|\omega|)\right\} \end{equation} can be searched in the family of triangles with vertexes $A$, $B^*$ and $C$, with $B^*$ belonging either to $AB$ (for $a\leq 1/3$) or to $BC$ (for $a\geq 1/3$). \end{proposition} \begin{proof} Let us denote for by $T_{ABC}$ the triangle $ABC$. We first notice that the maximum exists: since by construction the area and the perimeter of admissible configurations are bounded, then the supremum of $\Psi(\mathcal H^1(\Gamma))\Phi(|\omega|)$ is finite. The existence of an optimal configuration follows by combining the compactness of the admissible sets with respect to the complementary Hausdorff convergence, the continuity of perimeter and area in the class, and the continuity of $\Psi$ and $\Phi$ in $\mathbb R^+$. Without loss of generality, we may search for maximizers in the dense subclass of convex polygons inside $T_{ABC}$, having $A$ and $C$ among their vertexes: a maximizer among polygons is also a maximizer among convex sets. The proof is divided into three steps. \smallskip \noindent \textit{Step 1.} Arguing by contradiction, it is easy to see that no vertex of an optimal shape is in the interior of $T_{ABC}$: let $Q_1, Q_2, Q_3$ be three vertexes with $Q_2$ in the interior of $T_{ABC}$. We now perform a {\it parallel chord movement}: consider a small perturbation of the set obtained by sliding the point $Q_2$ on the line parallel to $Q_1Q_3$ passing through $Q_2$: since $Q_2$ is in the interior, the deformation can be done in both directions, either towards $Q_1$ or towards $Q_3$. In both cases, convexity and area are preserved; but, at least in one of the two directions, the perimeter increases, giving the contradiction. \smallskip \noindent \textit{Step 2.} The previous step implies that the maximizers have at most 4 vertexes: $A$, $C$, one point $Q_1$ in $AB$ and one point $Q_2$ in $BC$. Let us exclude the case in which $Q_1$ and $Q_2$ are (both) in the interior of the segments $AB$ and $BC$, respectively. They are of the form $Q_1= A + t e_1$, for some $t\in ]0,a/2[$, and $Q_2= C + s \nu$, for some $s\in ]0,1/2-a[$, where $e_1=(1,0)$ and $\nu=(-1/2, \sqrt{3}/2)$ are the unit vectors in the direction of the segments $AB$ (from $A$ to $B$) and $CB$ (from $C$ to $B$), respectively. Let us perturb the configuration keeping the area constant as follows: we replace $Q_1$ and $Q_2$ with \begin{equation} Q_1^\varepsilon:= A + (t+\varepsilon) e_1, \quad Q_2^\varepsilon:=C + (s+\delta(\varepsilon))\nu, \end{equation} for $\varepsilon\in \mathbb R$ small and for a suitable $\delta(\varepsilon)$. Imposing that the corresponding shape $\omega_\varepsilon$ has the same area, we deduce the value of $\delta(\varepsilon)$: $$ \frac{\mathrm{d}}{\mathrm{d}\varepsilon}|\omega_\varepsilon|_{\lfloor_{\varepsilon=0}} = 0 \quad \Leftrightarrow \quad \frac{\mathrm{d}}{\mathrm{d}\varepsilon} \left[\left( \frac{a}{2}-t-\varepsilon\right)\left(\frac12 - a -s - \delta(\varepsilon) \right)\right]_{\lfloor_{\varepsilon=0}} =0, $$ so that $$ \delta(\varepsilon)=-2x\varepsilon + o(\varepsilon), \quad \hbox{with}\quad x:= \frac{\sigma}{\tau}, \quad \hbox{being}\quad \sigma := \frac14 - \frac{a}{2} - \frac{s}{2}, \quad \tau:= \frac{a}{2}-t. $$ Notice that $\tau \in ]0,a/2[$, $\sigma \in ]0,1/4 - a/2[$, and $x \in ]0, +\infty[$. As for the perimeter term, namely the length of the curve $\Gamma_\varepsilon$, the following expression holds: $$ \mathcal H^1(\Gamma_\varepsilon) = t + \varepsilon + s + \delta(\varepsilon) + \sqrt{ \left(\tau + \sigma - \varepsilon - \frac12 \delta(\varepsilon) \right)^2 + \frac{3}{4}\left( 2\sigma - \delta(\varepsilon)\right)^2}. $$ Let us now compute the derivative of the perimeter with respect to $\varepsilon$, avaluated at $\varepsilon=0$: \begin{align*} \frac{\mathrm{d}}{\mathrm{d}\varepsilon} \mathcal H^1(\Gamma_\varepsilon)_{\lfloor_{\varepsilon=0}} & = 1 + \delta'(0) + \frac{ \left[2 (\tau + \sigma) (-1-\delta'(0)/2) + 3 \sigma (-\delta'(0)) \right] }{2\sqrt{ (\tau + \sigma)^2 + 3 \sigma^2}} \\ & = 1 - 2 x + \frac{(2x+1) (2x-1)}{\sqrt{ 1 + 4 x^2 + 2x}} \\ & = (2x-1) \left[ \frac{2x+1}{\sqrt{ 1 + 4 x^2 + 2x}} -1 \right]. \end{align*} Recalling that $x\in ]0, +\infty[$, we notice that the derivative above vanishes for $x=1/2$, it is negative in $]0,1/2[$ and positive in $]1/2, +\infty[$. In particular, as a function of $x$, the perimeter does not have any maximum in the interior of $\mathbb R^+$. This allows us to conclude that none of these configurations is optimal: indeed, for fixed values of $t$ and $s$, the choice of the sign of $\varepsilon$ makes the corresponding $x$ increase of decrease, and at least for one of these cases, the perimeter increases. \smallskip \noindent \textit{Step 3.} In view of the previous steps, we infer that, among polygons, maximizers have to be searched in a particular class of triangles: the ones with two vertexes in $A$ and $C$, and the third one either between $A$ and $B$ or between $B$ and $C$. Using the notation introduced above (Step 2), we describe the former type of point with a parameter $t\in [0,a/2]$ and denote it by $Q_1(t)$, and the latter with $s\in [0,1/2-a]$ denoted by $Q_2(s)$. Here we allow $t$ and $s$ to take the extremal values, corresponding to triangles degenerating to the segment $AC$ or coinciding with the whole triangle $T_{ABC}$. In particular, all the possible areas of subsets of $T_{ABC}$ are attained by both families of triangles. The maximization problem among polygons becomes scalar and a maximizer exists. In this step we compare the pairs of triangles $AQ_1(t)C$ and $AQ_2(s)C$ having the same area. Writing $Q_1(t)$ in coordinates $$ Q_1(t)= \left(t, \frac{\sqrt{3}}{2}(1-a) \right), $$ and recalling the definition \eqref{ABC} of $A$ and $C$, we infer that the perimeter term $\mathcal H^1(\Gamma)$ of $AQ_1(t)C$ is $$ p_1(t) := t + \sqrt{\left(\frac14 - t\right)^2 + \frac34 \left(\frac12 - a \right)^2}. $$ Similarly, writing $Q_2(s)$ in coordinates $$ Q_2(s)= \left(\frac14 - \frac{s}2, \frac{\sqrt{3}}{2} \left( \frac12 + s \right)\right), $$ we deduce that the perimeter term of $AQ_2(s)C$ reads $$ p_2(s) := s + \sqrt{ \left(\frac14 - \frac{s}{2}\right)^2 + \frac{3}{4} \left( s-\frac12 + a\right)^2}. $$ Imposing that the areas coincide, we infer that, given $t$, $s$ has to be chosen as $s(t)= t (1-2a)/a$. We underline that, given $a$, this map is a bijection from the interval $[0,a/2]$, in which $t$ varies, and in interval $[0,1/2-a]$, in which $s$ varies. Thus we are led to study the sign of $p_1(t)-p_2(t(1-2a)/a)$: this function depends on two variables, $a$ varying in $]0,1/2[$ (for the extremal cases there is nothing to prove) and $t$ varying in $[0,a/2]$. In order to get rid of the dependence on $a$ of the domain of definition of $t$, let us introduce the variable $y:=2t/a\in [0,1]$. According to this definition, $t=ay/2$ and $s(t)=y (1-2a)/2$. A direct computation shows that the terms of $p_1-p_2$ can be rearranged as follows: \begin{equation}\label{p1p2} p_1(ay/2)-p_2(y(1-2a)/2)=\frac14[\varphi(y) - \tilde{\varphi}(y)], \end{equation} with $$ \varphi(y):= 2(3a-1) y + \sqrt{\left(1 - 2 ay\right)^2 + 3 \left(1 - 2a \right)^2}, \quad$$ $$ \tilde{\varphi}(y):= \sqrt{ \left(1 - y(1-2a) \right)^2 + 3 (1- y)^2(1-2a)^2}. $$ The function $\tilde\varphi$ is clearly non negative. Let us prove that the function $\varphi$ is non negative, too. If $a\geq 1/3$ this is clearly true. If $a\leq 1/3$, then it is easy to see that $\varphi$ is decreasing in $y$, so that $$ \varphi(y) \geq \varphi(1) = 2 (3a-1) + 2(1-2a) = 2a \geq 0. $$ Therefore, we are allowed to compare the squares of these functions: $\varphi^2(y) \geq \tilde{\varphi}^2(y)$ if and only if $$ 4 y (1-3a)\left[2 (1-a(1+y)) - \sqrt{\left(1 - 2 ay\right)^2 + 3 \left(1 - 2a \right)^2} \right] \geq 0. $$ The function in square brackets is always positive, since for every $y\in [0,1]$, $a\in [0,1/2]$, there holds $ (1-a(1+y)) \geq 0$ and $$[2 (1-a(1+y))]^2 - \left[\left(1 - 2 ay\right)^2 + 3 \left(1 - 2a \right)^2 \right] = 4a(1-2a)(1-y)\geq 0. $$ Going back to \eqref{p1p2}, we infer that $p_1(t)>p_2(s(t))$ for $a<1/3$, $p_1(t) \equiv p_2(s(t))$ for $a=1/3$, and $p_1(t)<p_2(s(t))$ in the remaining case, $a>1/3$. For every fixed $a$ we have found an optimal triangle among the convex admissible shapes with the same area. The proof is concluded. \end{proof} \begin{lemma}\label{lem-SW} For every $\Omega \in \mathcal C_a$, $a\in [1/4, 1/3]$, there holds \begin{equation}\label{ratio2} \frac{P^2(\Omega)}{|\Omega|}\leq \frac{P^2(H_a)}{|H_a|}. \end{equation} \end{lemma} \begin{proof} Let $\omega$ be the intersection of $\Omega$ with the triangle $ABC$ defined in \eqref{ABC}. Its boundary is the union of the segment $AC$ and a curve, that we denote here by $\Gamma$. By construction, $\Omega$ is the union of $H_a$ and 6 copies (reflections) of $\omega$, so that $|\Omega|= |H_a| + 6 |\omega|$ and $P(\Omega)= 6 \mathcal H^1(\Gamma)$. All in all, we may write \begin{equation}\label{ratio3} \frac{P^2(\Omega)}{|\Omega|} = \frac{36 (\mathcal H^1(\Gamma))^2}{|H_a| + 6|\omega|}. \end{equation} The right-hand side is of the form \eqref{form} and satisfies the assumptions of Proposition \ref{prop-reverse}. Therefore, recalling that $a\leq 1/3$, we deduce that the optimal $\Omega$ is a hexagon, associated to an $\omega$ belonging to the 1-parameter family of triangles $AB(t)C$, being $B(t)$ the point of $AB$ at distance $t$ from $A$, $t\in [0,a/2]$. Such triangle has area $|\omega | = \sqrt{3}(1-2a)t/8$ and the length of $\Gamma$ is $$ p(t):=\overline{AB(t)} + \overline{B(t)C} = t + \sqrt{\left( \frac{1}{4} - t \right)^2 + \frac34 \left(\frac12 -a\right)^2}. $$ Using $|H_a| =\sqrt{3}(2 - 3a)/8$, we deduce that \begin{equation}\label{maxt} \frac{P^2(\Omega)}{|\Omega|} \leq 36 \frac{8}{\sqrt{3}} \max_{t\in [0,a/2]} \frac{p(t)^2}{(2 - 3a) + 6 (1-2a) t }. \end{equation} Let us prove that the function in the right-hand side is convex. To this aim, we introduce the auxiliary functions $$ D(t):=(2-3a) + 6 (1-2a) t , \quad T(t):= \frac14 - t, \quad R(t):= \sqrt{T^2(t) + \frac34 \left(\frac12 -a\right)^2}. $$ According to this notation $p(t)=t + R(t)$ and the function in the right-hand side of \eqref{maxt} is $p^2(t)/D(t)$. Exploiting the fact that $$ p'(t)=\frac{R(t) - T(t)}{R(t)} = \frac{p(t)-1/4}{R(t)}, \quad p''(t)= p'(t) \frac{R(t)+T(t)}{R^2(t)}, \quad D''(t)=0, $$ we get \begin{align} \frac{\mathrm{d}^2}{\mathrm{d}t^2} \frac{p^2(t)}{D(t)} \notag & = \frac{2}{D^3} \left[(p')^2D^2 + p p''D^2 - 2 p p' D' D + (D')^2p^2 \right]\notag \\ & \geq \frac{2}{D^2} \left[(p')^2D + p p''D - 2 p p' D' \right]\notag \\ &= \frac{2p'}{D^2} \left[\frac{p-1/4}{R} D + pD \frac{(R+T)}{R^2} - 2 p D' \right]\notag \\ &= \frac{2p'}{D^2} \left[2p\left(\frac{D}{R} - D' \right) - D \left(\frac{1}{4R} - \frac{pT}{R^2}\right) \right]. \label{stima} \end{align} Setting $\psi(t):= 2p'(t)/D^2(t)$ and $$\phi_1(t):=2\frac{p(t)}{R(t)}\left[D(t) - D'(t)R(t) \right] , \quad \phi_2(t):= \frac{D(t)}{4 R^2(t)}\left[R(t) - 4 p(t)T(t)\right], $$ the last term in \eqref{stima} is $\psi(\phi_1-\phi_2)$. It is immediate to check that $\psi\geq 0$: this is a consequence of $R(t)\geq |T(t)|>T(t)$, giving $p'(t)\geq 0$. In particular, we deduce that $$ \frac{\mathrm{d}^2}{\mathrm{d}t^2} \frac{p^2(t)}{D(t)} \geq \psi(t) \left[ \min_{[0,a/2]} \phi_1 - \max_{[0,a/2]} \phi_2\right]. $$ The two functions $\phi_1$ and $\phi_2$ are both increasing, since \begin{align*} \phi_1'(t) & = \frac{2}{R^3} \left[ D (R^2 + t T) + D' R^2 (t+T) \right] \geq 0, \\ \phi_2'(t) &= \frac{p'}{R^3} \left[ t (D' R^2 + 2TD) + Dp R \right] \geq 0. \end{align*} This implies that \begin{align*} \min_{[0,a/2]} \phi_1 & = \phi_1(0)=2\Big(2-3a-3(1-2a)\sqrt{1-3a+3a^2}\Big). \\ \max_{[0,a/2]} \phi_2 & = \phi_2(a/2)=\frac{a(1-3a^2)}{(1-2a)}. \end{align*} Evaluating $\min {\phi_1} - \max \phi_2$ for $a$ in the range $[1/4, 1/3]$, we infer that the second order derivative of $p^2/D$ is positive, thus the function is convex. In particular, its maximum is either at $0$ or at $a/2$. A direct computation shows that $$ \frac{p^2(0)}{D(0)} - \frac{p^2(a/2)}{D(a/2)} = \frac{ 1-3a+3a^2}{4(2-3a)} - \frac{(1-a)^2}{8(1-3a^2)} = \frac{a (1-2a) (1-3a)^2}{8 (1-3a^2) (2-3a)} >0. $$ The maximizer $t=0$ corresponds to $\omega=\emptyset$ in the ratio \eqref{ratio3}, associated to $H_a$. This concludes the proof. \end{proof} Let us now maximize the right-hand side of \eqref{ratio2}: recalling that $$ P(H_a)=3 \sqrt{1 - 3 a + 3a^2}, \quad |H_a| = \frac{\sqrt{3}}{8}(2 - 3a), $$ and exploiting the decreasing monotonicity of the function $$ [1/4, 1/3] \ni x\mapsto \frac{1-3x+3x^2}{2-3x}, $$ we infer that $P^2(H_a)/|H_a|\leq P^2(H_{1/4})/|H_{1/4}|$. Thus \eqref{ratio} and \eqref{ratio2} give \begin{equation}\label{case2} P^2(\Omega) \mu_1(\Omega) \leq \pi (j'_{1,1})^2 \frac{P^2(H_{1/4})}{|H_{1/4}|} = \pi (j'_{1,1})^2 \frac{42 \sqrt{3}}{5} < 50 \pi, \end{equation} where in the last inequality we have used the fact that $(j'_{1,1})^2<3.4$. The last term is below $16\pi^2$, and this gives the proof of Proposition \ref{prop-3sym} in the range $a\in[1/4,1/3]$.

https://arxiv.org/abs/1801.07529

Connections between rank and dimension for subspaces of bilinear forms

Let $K$ be a field and let $V$ be a vector space of dimension $n$ over $K$. Let $M$ be a subspace of bilinear forms defined on $V\times V$. Let $r$ be the number of different non-zero ranks that occur among the elements of $M$. Our aim is to obtain an upper bound for $\dim M$ in terms of $r$ and $n$ under various hypotheses. As a sample of what we prove, we mention the following. Suppose that $m$ is the largest integer that occurs as the rank of an element of $M$. Then if $m\leq \lceil n/2\rceil$ and $|K|\geq m+1$, we have $\dim M\leq rn$. The case $r=1$ corresponds to a constant rank space and it is conjectured that $\dim M\leq n$ when $M$ is a constant rank $m$ space and $|K|\geq m+1$. We prove that the dimension bound for a constant rank $m$ space $M$ holds provided $|K|\geq m+1$ and either $K$ is finite or $K$ has characteristic different from 2 and $M$ consists of symmetric forms. In general, we show that if $M$ is a constant rank $m$ subspace and $|K|\geq m+1$, then $\dim M\leq \max\,(n,2m-1)$. We also provide more detailed results about constant rank subspaces over finite fields, especially subspaces of alternating or symmetric bilinear forms.

\section{Introduction} \noindent Let $K$ be a field and let $V$ be a vector space of finite dimension $n$ over $K$. We let $V^\times$ denote the subset of non-zero elements of $V$, and use similar notation for the subset of non-zero elements in any vector space. Let $\Bil(V)$ denote the $K$-vector space of all bilinear forms defined on $V\times V$. Let $\Alt(V)$ denote the subspace of $\Bil(V)$ consisting of alternating bilinear forms and $\Symm(V)$ the subspace of symmetric bilinear forms. Let $f$ be an element of $\Bil(V)$. Let $\rad_L f$ denote the left radical of $f$ and $\rad_R f$ denote the right radical of $f$. It is known that $\dim \rad_L f=\dim \rad_R f$ and the number $n-\dim \rad_L f$ is called the rank of $f$, which we denote by $\rank f$. In the case that $f$ is alternating or symmetric, $\rad_L f=\rad_R f$ and we call the common subspace the radical of $f$, denoted by $\rad f$. \begin{definition} \label{rank_definition} Let $\M$ be a non-zero subspace of $\Bil(V)$. We let $\rank(\M)$ denote the set of different integers that occur as the ranks of the non-zero elements of $\M$. If $\M$ is the zero subspace, we set $\rank(\M)=0$. \end{definition} Thus $\rank(\M)=\{ \rank f: f\in \M^\times\}$, where we only include the different ranks that occur. Clearly, we have $\rank(\N)\leq \rank(\M)$ when $\N$ is a subspace of $\M$. One purpose of this paper is to obtain an upper bound for $\dim \M$ in terms of $|\rank(\M)|$ and $n$. The results we obtain vary according to the nature of the forms and certain hypotheses we make, and may be subdivided into three types, described as follows. Let $\M$ be a non-zero subspace of $\Bil(V)$, let $m$ be the largest integer in $\rank(\M)$ and let $r=|\rank(\M)|$. Then if $m\leq \lceil n/2\rceil$ and $|K|\geq m+1$, we have \[ \dim \M\leq rn. \] This bound is optimal in non-trivial cases, but we do not know if the restriction on the size of $m$ is essential. See Theorem \ref{bilinear_several_ranks_bound}. Suppose next that $\M$, as above, is a subspace of $\Alt(V)$. Then if $m\leq \lfloor n/2\rfloor$ and $|K|\geq m+1$, we have \[ \dim \M\leq rn-\frac{r(r+1)}{2}. \] Examples show that this bound is also optimal in non-trivial ways, but the restriction on the size of $m$ is essential. See Theorem \ref{alternating_several_ranks_bound}. Finally, suppose that $\M$ is a subspace of $\Symm(V)$. Then if $K$ has characteristic different from 2 and $|K|\geq n$, we have \[ \dim \mathcal{M}\leq rn-\frac{r(r-1)}{2}. \] The hypothesis on the characteristic of $K$ is essential. This bound is optimal in some cases, but not so in general. For small values of $r$ and specific fields, examples show that the bound is reasonably precise. See Theorem \ref{symmetric_general_dimension_bound}. Our original motivation for undertaking investigations of this nature is as follows. Suppose that $|\rank(\M)|=1$, and let $m$ be the rank of all non-zero elements of $\M$. We say that $\M$ is a constant rank $m$ subspace of $\Bil(V)$. It is conjectured that the dimension of a constant rank $m$ subspace is at most $n$, provided that $|K|\geq m+1$. For many fields, including all finite fields, there are constant rank $m$ subspaces of dimension $n$, when $1\leq m\leq n$, and thus the conjectured upper bound is optimal if it is proved to hold. This dimension bound is trivial to prove for all fields if $m=n$ and thus interest is concentrated on the case when $m<n$. The dimension bound is a consequence of Theorem 7 for a constant rank $m$ subspace of $\Symm(V)$ when $K$ has characteristic different from 2 and $|K|\geq m+1$. We also prove the upper bound for finite fields of size at least $m+1$ (Theorem \ref{finite_constant_rank}). We already proved an equivalent theorem in \cite{G1}. We supplement this dimension bound with the following additional information when the maximum dimension $n$ occurs. Let $\M$ be an $n$-dimensional constant rank $m$ subspace of $\Bil(V)$ and let $K=\mathbb{F}_q$, where $q\geq m+1$. Then if $n\geq 2m+1$, all elements of $\M^\times$ either have the same left radical or the same right radical. See Theorem \ref{equality_of_left_radicals}. Some condition on the size of $m$ relative to $n$ is necessary for the truth of this theorem. Concerning the general case of the constant rank dimension bound, we prove the following results. Let $\M$ be a constant rank $m$ subspace of $\Bil(V)$. Then if $|K|\geq m+1$, we have $\dim \M\leq \max\,(n,2m-1)$. See Theorem \ref{bilinear_constant_rank_bound}. If we assume that $\M\leq \Alt(V)$, this bound can be improved to $\dim \M\leq \max\,(n-1,2m-1)$. See Theorem \ref{alternating_constant_rank_bound}. We have remarked above that if $\M$ is a constant rank $m$ subspace of $\Bil(V)$, we have $\dim \M\leq n$ if we work over a sufficiently large finite field, and this upper bound is optimal. For constant rank subspaces of $\Alt(V)$ and $\Symm(V)$ in the finite field case, there is reason to suppose that the upper bound of $n$ for the dimension can often be improved. To illustrate this point, let $\M$ be a constant rank $m$ subspace of $\Alt(V)$ and $K=\mathbb{F}_q$, where $q\geq m+1$. Then if $4\leq m\leq \lfloor n/2\rfloor$, we have $\dim \M\leq n-2$ (Theorem \ref{new_alternating_dimension_bound}). Similarly, let $\M$ be a constant rank $m$ subspace of $\Symm(V)$ and $K=\mathbb{F}_q$, where $q$ is odd and at least $m+1$. Then if $m\leq 2n/3$, we have $\dim \M\leq n-1$ (Theorem \ref{the_main_dimension_bound}). Additional results can also be obtained for certain constant rank subspaces of $\Alt(V)$ in the finite field case. By way of example, let $\M$ be an $n$-dimensional constant rank $m$ subspace of $\Alt(V)$ and let $K=\mathbb{F}_q$. Then if $q\geq m+1$, $n-m$ divides $n$. This follows from the fact that the different subspaces of dimension $n-m$ of $V$ that occur as the radicals of the elements of $\M^\times$ form a spread of $V$ (Theorem \ref{spread_theorem}). We also show that such $n$-dimensional constant rank $m$ subspaces of $\Alt(V)$ exist if $n$ is odd and $n-m$ divides $n$. \section{Basic theorems for studying bilinear forms} \noindent The following is fundamental to all the results we obtain in this paper. \begin{theorem} \label{left_radical_right_radical} Let $\M$ be a subspace of $\Bil(V)$ and let $m$ be the largest integer in $\rank(\M)$. Let $f$ be an element of $\M$ with $\rank f=m$. Let $u$, $w$ be arbitrary elements of $\rad_L f$, $\rad_R f$, respectively. Then if $|K|\geq m+1$, we have \[ g(u,w)=0 \] for all elements $g$ of $\M$. \end{theorem} \begin{proof} We set $U=\rad_L f$ and $W=\rad_R f$. These are both subspaces of dimension $n-m$. It follows that there is an automorphism, $\sigma$, say, of $V$ with $\sigma(U)=W$. For each element $g$ of $\M$, we define $g_\sigma$ in $\Bil(V)$ by setting \[ g_\sigma(x,y)=g(x, \sigma y) \] for all $x$ and $y$ in $V$. Clearly, since $\sigma$ is an automorphism of $V$, $g_\sigma$ has the same rank as $g$. Thus $f_\sigma$ has rank $m$ and its left radical is $U$. The right radical of $f_\sigma$ consists of those elements $y$ such that $\sigma y\in W$. Thus $y\in \sigma^{-1}(W)=U$. We see therefore that $U$ is both the left and right radical of $f_\sigma$. Let $U'$ be a complement for $U$ in $V$. With respect to a basis of $V$ consisting of bases of $U$ and $U'$, we can take the matrix of $f_\sigma$ to be \[ C=\left( \begin{array} {cc} 0&0\\ 0&A \end{array} \right), \] where $A$ is an invertible $m\times m$ matrix. Given $g$ in $\M$, let the matrix of $g_\sigma$ with respect to the same basis be \[ D=\left( \begin{array} {cc} A_1&A_2\\ A_3&A_4 \end{array} \right), \] where $A_1$ is an $(n-m)\times (n-m)$ matrix, $A_4$ is an $m\times m$ matrix, and $A_2$, $A_3$ are matrices of the appropriate compatible sizes. When we follow the proof of Theorem 1 of \cite{G2}, we find that $A_1=0$ if $|K|\geq m+1$. Now the fact that $A_1=0$ means that \[ g_\sigma(x,y)=0 \] for all $x$ and $y$ in $U$. It follows that \[ g(x,\sigma y)=0 \] for all $x$ and $y$ in $U$. However, since $\sigma(U)=W$, we obtain \[ g(u,w)=0 \] for all $u$ in $U$ and all $w$ in $W$. \end{proof} Next, we introduce a simple idea which is useful for induction arguments that establish a relationship between $\dim \M$ and $|\rank(\M)|$. Let $V^*$ denote the dual space of $V$ and let $u$ be any vector in $V$. We define a linear transformation $\epsilon_u:\M\to V^*$ by setting \[ \epsilon_u(f)(v)=f(u,v) \] for all $f\in \M$ and all $v\in V$. Likewise, we define a linear transformation $\eta_u:\M\to V^*$ by setting \[ \eta_u(f)(v)=f(v,u) \] for all $f\in \M$ and all $v\in V$. Let $\M_u^L$ denote the kernel of $\epsilon_u$. This is the subspace of $\M$ consisting of all those forms $f$ such that $u\in \rad_L f$. Similarly, let $\M_u^R$ denote the kernel of $\eta_u$. This is the subspace of $\M$ consisting of all those forms $f$ such that $u\in \rad_R f$. When $\M$ is a subspace of either $\Alt(V)$ or $\Symm(V)$, clearly $\M_u^L=\M_u^R$. In this case, we write $\M_u$ in place of $\M_u^L$. The following estimate for $\dim \M_u^L$ and $\dim \M_u^R$ follows from the fact that $\dim V^*=n$. See, for example, Lemma 2 of \cite{G4}. \begin{lemma} \label{M_u_dimension_bound} For each element $u$ of $V$, we have \[ \dim \M_u^L\geq \dim \M-n, \quad \dim \M_u^R\geq \dim \M-n. \] \end{lemma} The following more precise estimate for both $\dim \M_u^L$ and $\dim \M_u^R$ is also important throughout this paper. Its proof relies on Theorem \ref{left_radical_right_radical}. \begin{lemma} \label{improved_M_u_dimension_bound} Let $\M$ be a non-zero subspace of $\Bil(V)$ and let $m$ be the largest integer in $\rank(\M)$. Let $u$ be an element of $V$. Suppose that $\M_u^L$ contains an element of rank $m$. Then if $|K|\geq m+1$, we have \[ \dim \M_u^L\geq \dim \M-m. \] Furthermore, if $\dim \M_u^L=\dim \M-m$, then all elements of rank $m$ in $\M_u^L$ have the same right radical. Similarly, if $\M_u^R$ contains an element of rank $m$ and if $|K|\geq m+1$, \[ \dim \M_u^R\geq \dim \M-m. \] The equality $\dim \M_u^R=\dim \M-m$ implies that all elements of rank $m$ in $\M_u^R$ have the same left radical. \end{lemma} \begin{proof} Suppose that $\M_u^L$ contains an element $f$, say, of rank $m$. Then provided that $|K|\geq m+1$, Theorem \ref{left_radical_right_radical} implies that as $u\in \rad_L f$, \[ g(u,w)=0 \] for all $w$ in $\rad_R f$ and all $g$ in $\M$. We deduce that $\epsilon_u(\M)$ is contained in the annihilator of $\rad_R f$ in $V^*$. Now since $\rad_R f$ has dimension $n-m$, its annihilator in $V^*$ has dimension $m$. This establishes that $\dim \epsilon_u(\M)\leq m$. It is also clear that if $\dim \epsilon_u(\M)=m$, $\epsilon_u(\M)$ is precisely the annihilator of $\rad_R g$ for each element $g$ of rank $m$ in $\M_u^L$, and hence these right radicals are all the same. When we recall that \[ \dim \M=\dim \epsilon_u(\M)+\dim \M_u^L, \] we obtain the desired inequality $\dim \M_u^L\geq \dim \M-m$. An identical proof serves to estimate $\dim \M_u^R$ under the stated hypotheses and to identify left radicals when equality holds. \end{proof} \section{Constant rank subspaces in finite field case} \noindent The following lemma is our main tool to investigate constant rank subspaces of bilinear forms over finite fields. \begin{lemma} \label{constant_rank_dimension_lemma} Let $\M$ be a $d$-dimensional constant rank $m$ subspace of $\Bil(V)$ and let $K=\mathbb{F}_q$. Then we have \[ (q^d-1)(q^{n-m}-1)=\sum_{u\neq 0} (q^{d(u)}-1), \] where the sum extends over all non-zero vectors $u$ and $d(u)=\dim \M_u^L$. \end{lemma} \begin{proof} Let $\Omega$ be the set of pairs $(f,u)$, where $f$ is an element of $\M^\times$, and $u$ is an element of $(\rad_L f)^\times$. We first evaluate $|\Omega|$ by fixing $f$ and counting those $u\neq 0$ in $\rad_L f$. We obtain \[ |\Omega|=(q^d-1)(q^{n-m}-1). \] Next we evaluate $|\Omega|$ by fixing a non-zero $u$ and counting those $f$ with $u\in \rad_L f$. The required $f$ are the non-zero elements of $\M_u^L$. Thus, summing over all non-zero $u$, we derive the equality \[ |\Omega|=\sum_{u\neq 0} (q^{d(u)}-1). \] This proves what we want. \end{proof} This lemma enables us to prove that a constant rank subspace of $\Bil(V)$ has dimension at most $n$, provided that we work over a sufficiently large finite field. \begin{theorem} \label{finite_constant_rank} Let $\M$ be a constant rank $m$ subspace of $\Bil(V)$ and let $K=\mathbb{F}_q$, where $q\geq m+1$. Then we have $\dim \M\leq n$. (When $m=n$, the result is true for all fields $K$.) \end{theorem} \begin{proof} Suppose if possible that $\dim \M>n$. Then we may as well assume that $\dim \M=n+1$ and proceed to derive a contradiction. Given $u\in V$, we set $d(u)=\dim \M_u^L$. Since we are assuming that $\dim \M>n$, Lemma \ref{M_u_dimension_bound} implies that $d(u)\geq 1$. We deduce from Lemma \ref{improved_M_u_dimension_bound} that, since $\M$ is a constant rank $m$ subspace, \[ d(u)\geq n+1-m. \] Lemma \ref{constant_rank_dimension_lemma} shows that \[ \sum_{u\neq 0} (q^{d(u)}-1)=(q^{n+1}-1)(q^{n-m}-1). \] On expanding each side above, we obtain \[ (\sum_{u\neq 0} q^{d(u)})-q^n=q^{2n+1-m}-q^{n+1}-q^{n-m}. \] The highest power of $q$ dividing the right hand side is $q^{n-m}$. On the other hand, as $d(u)\geq n+1-m$, the highest power of $q$ dividing the left hand side is at least $q^{n+1-m}$. This is a contradiction, and we deduce that $\dim \M\leq n$. \end{proof} We gave a similar proof of an equivalent theorem in \cite{G1} but have included this proof to illustrate the ideas of bilinear form theory. Our next objective is to investigate what happens when we have the equality $\dim \M=n$ in Theorem \ref{finite_constant_rank}. We begin by defining two relevant subspaces of $V$. \begin{definition} \label{V^L_M_definition} Let $\M$ be a subspace of $\Bil(V)$. We set \[ V(\M)^L=\{ v \in V: \M_v^L\neq 0\} \mbox{ and } V(\M)^R=\{v \in V: \M_v^R\neq 0\}. \] \end{definition} \begin{lemma} \label{V(M)_is_a_subspace} Let $\M$ be a constant rank $m$ subspace of $\Bil(V)$, with $\dim \M\geq 2m+1$. Then, if $|K|\geq m+1$, $V(\M)^L$ and $V(\M)^R$ are both subspaces of $V$. \end{lemma} \begin{proof} Let $u$, $w$ be elements of $V(\M)^L$. We wish to show that $\M_{u+w}^L\neq 0$. Then, since $V(\M)^L$ is clearly closed under scalar multiplication, it will follow that $V(\M)^L$ is a subspace. We claim that if we can show that $\M_u^L\cap \M_w^L\neq 0$, this will establish that $\M_{u+w}^L\neq 0$. For suppose that $g$ is a non-zero element of $\M_u^L\cap \M_w^L$. Then $u$ and $w$ are contained in $\rad_L g$ and hence $u+w\in \rad_L g$. This implies that $\M_{u+w}^L\neq 0$, as required. We turn therefore to proving that $\M_u^L\cap \M_w^L\neq 0$. Lemma \ref{improved_M_u_dimension_bound} shows that both $\M_u^L$ and $\M_w^L$ have dimension at least $\dim \M-m$. In addition, we have the inequality \[ \dim(\M_u^L+\M_w^L)\leq \dim \M \] and hence \[ \dim \M_u^L+\dim \M_w^L -\dim (\M_u^L\cap \M_w^L)\leq \dim \M. \] Given the earlier inequalities for $\dim \M_u^L$ and $\dim \M_w^L$, we deduce that \[ 2(\dim \M-m)-\dim (\M_u^L\cap \M_w^L) \leq \dim \M \] and hence \[ \dim (\M_u^L\cap \M_w^L)\geq \dim \M-2m. \] Since we are assuming that $\dim \M\geq 2m+1$, we see that $\dim (\M_u^L\cap \M_w^L)\geq 1$, and this completes the proof that $V(\M)^L$ is a subspace. The proof that $V(\M)^R$ is also a subspace is identical, since the same inequalities hold. \end{proof} \begin{lemma} \label{left_orthogonal_right_orthogonal} Let $\M$ be a constant rank $m$ subspace of $\Bil(V)$, with $\dim \M\geq 2m+1$. Then, if $|K|\geq m+1$, we have \[ f(u,w)=0 \] for all $u\in V(\M)^L$, all $w\in V(\M)^R$, and all $f\in \M$. \end{lemma} \begin{proof} Let $u$ and $w$ be elements in $V(\M)^L$, $V(\M)^R$, respectively. Then we have $\dim \M_u^L\geq \dim \M-m$, $\dim \M_w^R\geq \dim \M-m$ by Lemma \ref{improved_M_u_dimension_bound}. The dimension argument used in Lemma \ref{V(M)_is_a_subspace} shows that $\M_u^L\cap \M_w^R\neq 0$, since we are assuming that $\dim \M\geq 2m+1$. Let $g$ be a non-zero element in $\M_u^L\cap \M_w^R$. Then $u$ is in $\rad_L g$, $w$ is in $\rad_R g$ and hence Theorem \ref{left_radical_right_radical} implies that \[ f(u,w)=0 \] for all $f$ in $\M$. This proves the lemma. \end{proof} \begin{lemma} \label{dimension_estimate_in_terms_of_V} Let $\M$ be a constant rank $m$ subspace of $\Bil(V)$, with $\dim \M\geq 2m+1$. Then, if $|K|\geq m+1$, we have \[ \dim \M_u^L\geq \dim \M-n+\dim V(\M)^R \] for all $u\in V(\M)^L$. \end{lemma} \begin{proof} Lemma \ref{left_orthogonal_right_orthogonal} implies that $\epsilon_u(\M)$ annihilates $V(\M)^R$. Thus $\epsilon_u(\M)$ is contained in the annihilator of $V(\M)^R$ in $V^*$. We deduce that \[ \dim \epsilon_u(\M)\leq n-\dim V(\M)^R. \] Since $\dim \epsilon_u(\M)=\dim \M-\dim \M_u$, the inequality follows. \end{proof} It is clear that if $\M$ is a subspace of $\Bil(V)$, $V(\M)^L$ is the union of the left radicals of the elements of $\M$, and similarly $V(\M)^R$ is the union of the right radicals. Thus, if $\M$ is a constant rank $m$ subspace of $\Bil(V)$ and if $V(\M)^L$ is a subspace of $V$, certainly $\dim V(\M)^L\geq n-m$, and $\dim V(\M)^L= n-m$ if and only if all elements of $\M^\times$ have the same left radical. Similarly, if $V(\M)^R$ is a subspace, its dimension is at least $n-m$ and it equals $n-m$ if and only if all elements of $\M^\times$ have the same right radical. We proceed now to prove a theorem about the equality of left radicals or of right radicals for constant rank $m$ subspaces of maximum dimension $n$ provided we assume that $n\geq 2m+1$ and we work over sufficiently large finite fields. \begin{theorem} \label{equality_of_left_radicals} Let $\M$ be an $n$-dimensional constant rank $m$ subspace of $\Bil(V)$ and let $K=\mathbb{F}_q$, where $q\geq m+1$. Then if $n\geq 2m+1$, the elements of $\M^\times$ either all have the same left radical or they have the same right radical. \end{theorem} \begin{proof} Let us assume that our assertion above about the left and right radicals is not true. Then we have the inequalities $\dim V(\M)^L\geq n-m+1$, $\dim V(\M)^R\geq n-m+1$ by the discussion after the proof of Lemma \ref{dimension_estimate_in_terms_of_V}, and we will show that these lead to a contradiction. Lemma \ref{constant_rank_dimension_lemma} shows that \[ \sum_{u\neq 0} (q^{d(u)}-1)=(q^{n}-1)(q^{n-m}-1). \] We are interested only in those $u$ for which $d(u)>0$ and these are the elements of $V(\M)^L$. Let us put $\dim V(\M)^L=d$. Then \[ \sum_{u\neq 0} (q^{d(u)}-1)=(\sum_{u\neq 0} q^{d(u)})-q^d+1. \] On expanding and rearranging, we obtain \[ (\sum_{u\neq 0}q^{d(u)})-q^d=q^{2n-m}-q^{n}-q^{n-m}. \] We are assuming that $d\geq n-m+1$ and also that $\dim V(\M)^R\geq n-m+1$. We then have $d(u)\geq n-m+1$ by Lemma \ref{dimension_estimate_in_terms_of_V}. Thus the power of $q$ dividing the left hand side above is at least $q^{n-m+1}$. However, the power of $q$ dividing the right hand side is exactly $q^{n-m}$. We have reached a contradiction, and thus either the left radicals are all equal, or the right radicals are all equal. \end{proof} We note that we cannot weaken the hypothesis that $\dim \M=n$ in Theorem \ref{equality_of_left_radicals}, since, for example, there exists a constant rank 2 subspace of $\Symm(V)$ of dimension $n-1$ in which all linearly independent elements have different radicals. Similarly, some restriction on the size of $m$ compared with $n$ is needed, since, as we shall see in Section 7, there exist $n$-dimensional constant rank $m$ subspaces of $\Alt(V)$, for which Theorem \ref{equality_of_left_radicals} cannot possibly be true. The simplest example occurs when $n=3$. $\Alt(V)$ is a three-dimensional constant rank 2 subspace, in which linearly independent forms have different one-dimensional radicals. \section{Dimension bounds for subspaces of symmetric forms} \noindent We begin this section on symmetric forms with an application of Theorem \ref{left_radical_right_radical}. \begin{lemma} \label{symmetric_no_elements_of_maximum_rank} Let $K$ be a field of characteristic different from $2$ and let $\M$ be a non-zero subspace of $\Symm(V)$. Let $m$ be the largest integer in $\rank(\M)$. Then if $|K|\geq m+1$, there exists an element $u$, say, in $V$ such that $\M_u$ contains no element of rank $m$. \end{lemma} \begin{proof} There exists an element $h\in \M$ and $u\in V$ with $h(u,u)\neq 0$, since $K$ has characteristic different from 2 and $\M\neq 0$ consists of symmetric bilinear forms. Suppose now that $\M_u$ contains an element $f$, say, of rank $m$. Then $u\in \rad f$ and hence Theorem \ref{left_radical_right_radical} implies that, if we assume that $|K|\geq m+1$, each element $g$ of $\M$ satisfies \[ g(u,u)=0. \] This contradicts our earlier statement about the existence of $h$ in $\M$ with $h(u,u)\neq 0$. We deduce that $\M_u$ contains no element of rank $m$, as required. \end{proof} We have enough information to prove our constant rank dimension bound for subspaces of $\Symm(V)$. \begin{theorem} \label{symmetric_constant_rank_bound} Let $\M$ be a constant rank $m$ subspace of $\Symm(V)$. Then if $K$ has characteristic different from $2$ and $|K|\geq m+1$, we have $\dim \M\leq n$. (When $m=n$, the bound $\dim \M\leq n$ holds for all fields $K$ without exception, as noted in Theorem \ref{finite_constant_rank}.) \end{theorem} \begin{proof} We suppose that $|K|\geq m+1$. Clearly, by the constant rank hypothesis, Lemma \ref{symmetric_no_elements_of_maximum_rank} implies that there is some $u$ in $V$ such that $\M_u=0$. Lemma \ref{M_u_dimension_bound} implies then that $\dim \M\leq n$, as required. \end{proof} We would like now to show that, in contrast to Theorem \ref{symmetric_constant_rank_bound}, given positive integers $m$ and $n$, with $2\leq m\leq n$, and making certain assumptions about $K$, $\Symm(V)$ contains a constant rank $m$ subspace $\M$ of dimension $m$ that is maximal with respect to the property of being constant rank $m$. In other words, $\M$ is not contained in any larger constant rank $m$ subspace of $\Symm(V)$. Thus for example, under fairly weak hypotheses on $K$, there are two-dimensional maximal constant rank 2 subspaces of $\Symm(V)$. The construction is based on simple concepts of field theory. Suppose that $K$ has a separable extension $L$, say, of degree $m\geq 2$, but is otherwise arbitrary. We consider $L$ as a vector space of dimension $m$ over $K$. Let $\ensuremath\mathrm{Tr}:L\to K$ denote the trace form. For each element $z$ of $L$, we define an element $f_z$ of $\Symm(L)$ by setting \[ f_z(x,y)=\ensuremath\mathrm{Tr}(z(xy)) \] for all $x$ and $y$ in $L$. Let $\N$ be the subspace of $\Symm(L)$ consisting of all the $f_z$. We have $\dim \N=m$, and each element of $\N^\times$ has rank $m$, since $\ensuremath\mathrm{Tr}$ is non-zero under the hypothesis of separability. We note the following property of $\N$ whose (omitted) proof depends on the fact that $\ensuremath\mathrm{Tr}$ is non-zero. \begin{lemma} \label{non_vanishing_property} Let $\N$ be the subspace of $\Symm(L)$ described above. Let $x$ and $y$ be elements of $L$ such that \[ f_z(x,y)=0 \] for all $f_z$ in $\N$. Then $x=0$ or $y=0$. \end{lemma} We can now construct our example. \begin{theorem} \label{maximal_example} Suppose that $K$ has a separable extension of degree $m\geq 2$. Then if $m\leq n=\dim V$ and $|K|\geq m+1$, $\Symm(V)$ contains a constant rank $m$ subspace $\M$ of dimension $m$ that is contained in no larger constant rank $m$ subspace of $\Symm(V)$. Thus, $\M$ is maximal with respect to containment in constant rank $m$ subspaces. \end{theorem} \begin{proof} Let $U$ be a subspace of $V$ with $\dim U=m$, and let $W$ be a complement for $U$ in $V$. We may identify $U$ with $L$ as a vector space of dimension $m$ over $K$, and hence may define a constant rank $m$ subspace of $\Symm(U)$ of dimension $m$. We may then extend this subspace to a subspace $\M$ of $\Symm(V)$ by setting the extensions to have radical $W$ in all non-zero cases. We claim that the subspace $\M$ thus constructed is maximal in $\Symm(V)$ with respect to being constant rank $m$. For suppose that $\M$ is contained in a larger constant rank $m$ subspace, $\M_1$, say, of $\Symm(V)$. Let $f$ be an element of $\M_1$ not contained in $\M$. We aim to show that $\rad f=W$. Now since $|K|\geq m+1$, Theorem \ref{left_radical_right_radical} implies that $\rad f$ is totally isotropic for $\M$. Let $v$ be any element of $\rad f$ and set $v=u+w$, where $u\in U$ and $w\in W$. Let $g$ be any element of $\M$. Then we have $g(v,v)=0$ and hence \[ g(u+w,u+w)=g(u,u)=0, \] since $w\in \rad g$. Thus $g(u,u)=0$ for all $g\in \M$. Lemma \ref{non_vanishing_property} implies that $u=0$. We see therefore that $\rad f\leq W$ and hence $\rad f=W$, by consideration of dimensions. We have thus proved that all elements of $\M_1^\times$ have radical $W$. It is permissible then to identify $\M_1$ with a constant rank $m$ subspace of $\Symm(V/W)$. Since $V/W$ has dimension $m$, it follows that $\dim \M_1\leq m$. This is a contradiction, and our claim that $\M$ is maximal is established. \end{proof} When we work over the field of real numbers, we can show that there are one-dimensional constant rank $m$ subspaces that are maximal for any positive integer value of $m$. Our next theorem provides the details. \begin{theorem} \label{real_example} Let $K$ be the field of real numbers and let $m$ be a positive integer with $m\leq n$. Let $\M$ be a constant rank $m$ subspace of $\Symm(V)$ that contains a non-zero positive semidefinite element. Then $\dim \M=1$. \end{theorem} \begin{proof} Let $f\neq 0$ be a positive semidefinite element in $\M$ and let $U$ be a complement to $\rad f$ in $V$. Then, $f$ is positive definite on $U\times U$. Now let $g$ be any non-zero element in $\M$ and let $R=\rad g$. Theorem \ref{left_radical_right_radical} implies that $f$ is zero on $R\times R$. Hence $f$ is also zero on $(R+\rad f)\times (R+\rad f)$. Since $f$ is positive definite on $U\times U$, and zero on $(R+\rad f)\times (R+\rad f)$, we have $U\cap (R+\rad f)=0$. It follows that $R=\rad f$ and thus all non-zero elements of $\M$ have the same radical, which we may take to be 0. We may thus assume that all non-zero elements have maximum rank $n$. Suppose if possible that $\dim \M\geq 2$. Let $g$ be an element of $\M$ linearly independent of $f$. By a well known theorem of linear algebra over the real numbers, there is a basis of $V$ that is orthonormal with respect to $f$ and orthogonal with respect to $g$. But then it follows easily that there is a linear combination of $f$ and $g$ that has non-zero radical, contradicting our statement in the paragraph above. We deduce that $\dim \M=1$, as required. \end{proof} \begin{corollary} \label{one_dimensional_example} Let $K$ be the field of real numbers and let $m$ be a positive integer with $m\leq n$. Then there exists a one-dimensional maximal constant rank $m$ subspace of $\Symm(V)$. \end{corollary} We return now to the theme of bounding $\dim \M$ in terms of $|\rank(\M)|$. We investigate subspaces of $\Symm(V)$ and employ the idea underlying the proof of Theorem \ref{symmetric_constant_rank_bound}. \begin{theorem} \label{symmetric_general_dimension_bound} Let $\mathcal{M}$ be a subspace of $\Symm(V)$ and let $r=|\rank(\M)|$. Then if $K$ has characteristic different from $2$ and $|K|\geq n$, we have \[ \dim \mathcal{M}\leq rn-\frac{r(r-1)}{2}. \] \end{theorem} \begin{proof} We proceed by induction on $r$. The result is trivially true when $r=0$, so we can therefore assume that $r\geq 1$. Let $m$ be the largest integer in $\rank(\M)$. Lemma \ref{symmetric_no_elements_of_maximum_rank} implies that there exists some element $u$ in $V$ such that $\M_u$ contains no element of rank $m$. Thus, $|\rank(\M_u)|\leq r-1$. Let $U$ be the one-dimensional subspace of $V$ spanned by $u$ and let $U'$ be a complement of $U$ in $V$. Since $U$ is in the radical of each element of $\M_u$, we may identify $\M_u$ with a subspace of $\Symm(U')$. Since $\dim U'=n-1$, and $|\rank(\M_u)|\leq r-1$, we have by induction that \[ \dim \M_u\leq (r-1)(n-1)-\frac{(r-1)(r-2)}{2}. \] Lemma \ref{M_u_dimension_bound} yields that $\dim \M\leq \dim \M_u+n$ and we obtain the bound \[ \dim \M\leq (r-1)(n-1)-\frac{(r-1)(r-2)}{2}+n=rn-\frac{r(r-1)}{2}, \] as required. \end{proof} \bigskip \noindent{\bf Example 1.} Let $r$ be a positive integer such that $r\leq n/2$. Let $\M'$ be the subspace of all $n\times n$ symmetric matrices of the form \[ \left( \begin{array} {cc} 0&A\\ A^T&0 \end{array} \right), \] where $A$ runs over all $r\times (n-r)$ matrices with entries in $K$, and $A^T$ denotes the transpose of $A$. Such a matrix has rank $2s$, where $s$ is the rank of $A$ (and hence $0\leq s\leq r$). Thus $|\rank(\M')|=r$ and $\dim \M'=r(n-r)$. $\M'$ defines a subspace $\M$, say, of $\Symm(V)$, with $\dim \M=\dim \M'=r(n-r)$ and $|\rank(\M)|=|\rank(\M')|=r$. The bound for $\dim \M$ given by our theorem differs from the exact dimension by $r(r+1)/2$. Thus, for small values of $r$, we have a reasonably accurate dimension bound. \bigskip \noindent {\bf Example 2.} It is straightforward to show that a real symmetric matrix of trace 0 cannot have rank one. It follows that if $V$ is a vector space of dimension $n$ over the field of real numbers, $\Symm(V)$ contains a subspace $\M$ of dimension $n(n+1)/2-1$ in which $|\rank(\M)|=n-1$. This shows that Theorem \ref{symmetric_general_dimension_bound} is precise in the case $|\rank(\M)|=n-1$ for subfields of the real numbers. \section{Constant rank subspaces of symmetric forms over finite fields} \noindent Theorem \ref{symmetric_constant_rank_bound} shows that the dimension of a constant rank subspace of $\Symm(V)$ is at most $n$, provided that the underlying field is sufficiently large and has characteristic different from 2. A lack of specific examples or of construction processes suggests that this upper bound is rarely obtained. This section is devoted to improving Theorem \ref{symmetric_constant_rank_bound}, although our definitive results are currently restricted to finite fields, as ultimately we employ counting techniques to complete our arguments. We begin with definitions of concepts which were implicit in the previous section. \begin{definition} \label{totally_isotropic_subspace} Let $\M$ be a subspace of $\Symm(V)$. We say that a subspace $U$ of $V$ is totally isotropic for $\M$ if \[ f(u,w)=0 \] for all $u$ and $w$ in $U$, and all $f$ in $\M$. \end{definition} \begin{definition} \label{isotropic_points} Let $\M$ be a subspace of $\Symm(V)$. We set \[ I(\M)= \{ w\in V: f(w,w)=0 \mbox{ for all } f\in \M\}. \] We also set $I(\M)^\times$ to be the subset of non-zero elements in $I(\M)$. \end{definition} Clearly, a non-zero vector is in $I(\M)^\times$ if and only if the one-dimensional subspace it spans is totally isotropic for $\M$. More generally, a subspace of $V$ that is totally isotropic for $\M$ is contained in $I(\M)$. We consider the converse next. \begin{lemma} \label{totally_isotropic_in_subspace} Let $U$ be a subspace of $V$ that is contained in $I(\M)$. Then $U$ is totally isotropic for $\M$ provided that $K$ has characteristic different from $2$. \end{lemma} \begin{proof} Let $u$ and $w$ be elements of $U$ and $f$ be an element of $\M$. Then since $U$ is contained in $I(\M)$, we have \[ f(u,u)=f(w,w)=f(u+w,u+w)=0 \] and hence $2f(u,w)=0$, using the symmetry of $f$. We thus have $f(u,w)=0$ when $K$ has characteristic different from 2, and $U$ is totally isotropic for $\M$. \end{proof} We remark that in this section, the key point of one argument involves the case that $I(\M)$ is itself a subspace of $V$. It is in this case that we have resorted to counting techniques to resolve our problems. \begin{definition} \label{A_u_definition} Let $\M$ be a subspace of $\Symm(V)$ and let $u$ be an element of $V$. We let $\A_u$ denote the subspace of $V$ annihilated by the subspace $\epsilon_u(\M)$ of $V^*$. Thus \[ \A_u=\{ w\in V: f(u,w)=0 \mbox{ for all }f\in \M\}. \] \end{definition} The lemma that follows is important for our analysis of constant rank subspaces of $\Symm(V)$ of dimension $n$ (assuming they exist). \begin{lemma} \label{key_lemma_isotropic} Let $\M$ be an $n$-dimensional constant rank $m$ subspace of $\Symm(V)$. Then, given a vector $u$ in $V^\times$, the subspace $\A_u$ is totally isotropic for $\M$ provided that $|K|\geq m+1$ and $K$ has characteristic different from $2$. \end{lemma} \begin{proof} Let $w$ be an element of $\A_u^\times$. We aim to show that $w\in I(\M)^\times$. Now since $\epsilon_u(\M)$ vanishes on $w$, the symmetry of our forms implies that $\epsilon_w(\M)$ vanishes on $u$. Then, since $u\neq 0$, $\epsilon_w(\M)\neq V^*$. Now we have the general equality \[ \dim \epsilon_w(\M)+\dim \M_w=\dim \M=n \] and since we know from above that $\dim \epsilon_w(\M)<n$, we deduce that $\M_w\neq 0$. It follows that there is some element $f$ of $\M^\times$ with $w\in \rad f$. But $\M$ is a constant rank $m$ subspace of $\Symm(V)$, and since we are assuming that $|K|\geq m+1$, Theorem \ref{left_radical_right_radical} implies that $\rad f$ is totally isotropic for $\M$. Thus $w\in I(\M)^\times$ and hence $\A_u$ is contained in $I(\M)$. Finally, applying the assumption that $K$ has characteristic different from 2, Lemma \ref{totally_isotropic_in_subspace} implies that $\A_u$ is totally isotropic for $\M$. \end{proof} The next lemma is elementary. \begin{lemma} \label{A_w_property} Let $\M$ be a subspace of $\Symm(V)$. Let $W$ be a subspace of $V$ that is totally isotropic for $\M$. Then $W$ is contained in $\A_w$ for each element $w$ of $W$. \end{lemma} \begin{proof} Since $W$ is totally isotropic for $\M$, \[ f(w,u)=0 \] for all $u$ in $W$ and all $f$ in $\M$. This implies that $W\leq \A_w$, as required. \end{proof} We proceed to determine $\dim \A_u$. \begin{lemma} \label{dimension_of_A_u} Let $\M$ be an $n$-dimensional constant rank $m$ subspace of $\Symm(V)$ and let $u$ be an element of $V^\times$. Then $\dim \A_u=\dim \M_u$. Suppose furthermore that $|K|\geq m+1$ and $K$ has characteristic different from $2$. Then $\A_u$ is non-zero if and only if $u\in I(\M)^\times$. \end{lemma} \begin{proof} We have \[ \dim \epsilon_u(\M)=\dim \M-\dim \M_u=n-\dim \M_u. \] Duality theory shows that the subspace of $V$ annihilated by $\epsilon_u(\M)$ has dimension $n-\dim \epsilon_u(\M)$ and this number equals $\dim \M_u$ by our equality above. Thus, $\dim \A_u=\dim \M_u$. Suppose that $K$ has the properties described above and $\A_u$ is non-zero. Then $\A_u$ is totally isotropic for $\M$, by Lemma \ref{key_lemma_isotropic}. Thus $u\in I(\M)^\times$. Conversely, suppose that $u\in I(\M)^\times$. Then $u\in \A_u$ by Lemma \ref{A_w_property} and hence $\A_u$ is non-zero. \end{proof} The next in our sequence of lemmas enables us to partition $I(\M)$. \begin{lemma} \label{partition_theorem} Let $\M$ be an $n$-dimensional constant rank $m$ subspace of $\Symm(V)$. Suppose that $|K|\geq m+1$ and $K$ has characteristic different from $2$. Let $u$ be an element of $I(\M)^\times$. Then we have $\A_w=\A_u$ for all elements $w$ of $\A_u^\times $. \end{lemma} \begin{proof} Let $w$ be an element of $\A_u^\times $. Since $\A_u$ is totally isotropic for $\M$ by Lemma \ref{key_lemma_isotropic}, $\A_u\leq \A_w$ by Lemma \ref{A_w_property}. Furthermore, since $\epsilon_u(\M)$ annihilates $w$, $\epsilon_w(\M)$ annihilates $u$ by symmetry and thus $u\in \A_w$. Since $\A_w$ is also totally isotropic for $\M$, $\A_w\leq \A_u$, again by Lemma \ref{A_w_property}. Thus $\A_w=\A_u$, as stated. \end{proof} \begin{corollary} \label{statement_of_partition_property} Let $\M$ be an $n$-dimensional constant rank $m$ subspace of $\Symm(V)$. Suppose that $|K|\geq m+1$ and $K$ has characteristic different from $2$. Then $I(\M)$ is the union of the subspaces $\A_u$, where $u$ runs through the elements of $I(\M)^\times$. Two subspaces $\A_u$ and $\A_w$ are either identical or $\A_u\cap \A_w=0$. \end{corollary} \begin{proof} Let $u$ be an element $I(\M)^\times$. We have seen that $\A_u$ is non-zero and contained in $I(\M)$ by Lemma \ref{key_lemma_isotropic}. Since $u\in \A_u$, $I(\M)$ is the union of subspaces of type $\A_u$. Consider now a second subspace $\A_w$. Suppose $v$ is a non-zero vector in $\A_u\cap \A_w$. Then $v\in \A_u$ and thus $\A_v=\A_u$ by Lemma \ref{partition_theorem}. Likewise, $v\in \A_w$ and thus $\A_v=\A_w$. Therefore, $\A_u=\A_w$. \end{proof} We shall assume for the rest of this section that the underlying field $K$ is finite of odd characteristic. Corollary \ref{statement_of_partition_property} implies that in this case $I(\M)$ is the union of a finite number, $r$, say, of subspaces $\A_w$ which intersect trivially pairwise. Thus $I(\M)^\times$ is a disjoint union of $r$ subsets $\A_i^\times$, where $\A_i=\A_{u_i}$, $1\leq i\leq r$. If $\A_u$ is one of the $\A_i$, we have $\dim \A_u=\dim \M_u=d(u)$, say, where $d(u)>0$, by Lemma \ref{dimension_of_A_u}. Thus if we set $d_i=d(u_i)$, we have \[ |I(\M)^\times|=\sum_{i=1}^r (q^{d_i}-1). \] Retaining this notation, we have the following result. \begin{theorem} \label{key_counting_theorem} Let $\M$ be an $n$-dimensional constant rank $m$ subspace of $\Symm(V)$ and $K=\mathbb{F}_q$, where $q$ is odd and at least $m+1$. Then if \[ I(\M)^\times=\bigcup _{i=1}^r \A_i^\times, \] where $|\A_i^\times|=q^{d_i}-1$, we have \[ \sum_{i=1}^r (q^{d_i}-1)^2=(q^n-1)(q^{n-m}-1). \] \end{theorem} \begin{proof} Lemma \ref{constant_rank_dimension_lemma} shows that \[ \sum_{u\neq 0} (q^{d(u)}-1)=(q^n-1)(q^{n-m}-1), \] where $d(u)=\dim \M_u$ and we need only to sum over the elements $u$ of $I(\M)^\times$. We evaluate the sum on the left by counting over the subsets $\A_i^\times$ which partition $I(\M)^\times$. The subset contains $\A_i^\times$ contains $q^{d_i}-1$ elements $u$, for each of which $d(u)=d_i$. Thus the contribution of the elements of $\A_i^\times$ to the sum is $(q^{d_i}-1)^2$. Summing over all $i$, we obtain \[ \sum_{i=1}^r (q^{d_i}-1)^2=(q^n-1)(q^{n-m}-1), \] as required. \end{proof} The next result shows that $I(\M)$ cannot be a subspace in the circumstances of Theorem \ref{key_counting_theorem}. \begin{lemma} \label{r=1_case} We cannot have $r=1$ in Theorem \ref{key_counting_theorem}. \end{lemma} \begin{proof} Suppose if possible that $r=1$ in Theorem \ref{key_counting_theorem}. Then setting $d=d_1$, we obtain \[ (q^d-1)^2=(q^n-1)(q^{n-m}-1). \] A simple substitution in the formula above shows that $d$ cannot equal $n-m$. Thus $d>n-m$. Expansion of the formula above yields \[ q^{2d}-2q^d=q^{2n-m}-q^n-q^{n-m}. \] Since we are assuming that $q$ is odd, the power of $q$ dividing the left hand side is $q^d$, and the power of $q$ dividing the right hand side is $q^{n-m}$. This implies that $d=n-m$, a case we have already eliminated. Consequently, $r=1$ is impossible. \end{proof} We turn to the proof of the main theorem of this section. \begin{theorem} \label{the_main_dimension_bound} Let $\M$ be a constant rank $m$ subspace of $\Symm(V)$ and $K=\mathbb{F}_q$, where $q$ is odd and at least $m+1$. Then if $m\leq 2n/3$, we have $\dim \M<n$. \end{theorem} \begin{proof} Suppose by way of contradiction that $\dim \M=n$. We may then apply Theorem \ref{key_counting_theorem}. Let the distinct dimensions $d_i$ that occur be $e_1$, \dots, $e_t$, where \[ n-m\leq e_1<\ldots <e_t. \] Suppose that dimension $e_i$ occurs with multiplicity $c_i$ (so that $r=c_1+ \cdots +c_t$). Then we have \[ \sum_{i=1}^t c_i(q^{e_i}-1)^2=(q^n-1)(q^{n-m}-1), \quad |I(\M)^\times|=\sum_{i=1}^t c_i(q^{e_i}-1). \] Let us first show that we cannot have $t=1$ and $e_1=n-m$. For if we take $t=1$ and $e_1=n-m$, we obtain \[ c_1(q^{n-m}-1)^2=(q^n-1)(q^{n-m}-1), \quad |I(\M)^\times|=c_1(q^{n-m}-1). \] These equations imply that $|I(\M)^\times|=q^n-1$, a clear contradiction since it means that every vector in $V$ is isotropic for $\M$. Thus our statement is established. Expanding the first equation involving $(q^n-1)(q^{n-m}-1)$, we obtain \[ \sum_{i=1}^t c_i(q^{2e_i}-2q^{e_i}+1)=q^{2n-m}-q^n-q^{n-m}+1. \] Now as $e_i\geq n-m$ for all $i$, $q^{n-m}$ divides $q^{2e_i}$, $q^{e_i}$, $q^{2n-m}$ and $q^n$. It follows that $q^{n-m}$ divides \[ (\sum_{i=1}^t c_i)-1=r-1. \] Since we have shown that $r>1$ in Lemma \ref{r=1_case}, we have then \[ (\sum_{i=1}^t c_i)=\lambda q^{n-m}+1 \] for some positive integer $\lambda$. Now as we already know that all $e_i$ satisfy $e_i\geq n-m$, and moreover that not all $e_i$ equal $n-m$, we deduce that \[ |I(\M)^\times|>(\lambda q^{n-m}+1)(q^{n-m}-1) \] and thus $|I(\M)^\times|>q^{2(n-m)}-1$. Now we can certainly assume that $m$ is even, since if $m$ is odd, a constant rank $m$ subspace of $\Symm(V)$ has dimension at most $m$, and thus our theorem is trivially true. See, for example, Corollary 3 of \cite{DGS}. We therefore set $m=2k$, where $k$ is a positive integer. Then since $\dim \M=n$, Theorem 5 of \cite{DGS} implies that \[ |I(\M)^\times|=(A-B)q^{-k}, \] where $A$ is the number of elements of Witt index $k$ in $\M$, $B$ is the number of elements of Witt index $k-1$ in $\M$, and $A+B=q^n-1$. Thus, since $|I(\M)^\times|$ is an integer, we have $|I(\M)^\times|\leq q^{n-k}-1$. If we now compare this upper bound with our earlier lower bound for $|I(\M)^\times|$, we obtain \[ q^{2n-2m}-1<q^{n-k}-1 \] and deduce that $2n-2m<n-k=n-m/2$. This yields $n<3m/2$, and contradicts our original hypothesis that $n\geq 3m/2$. We therefore conclude that $\dim \M<n$, as required. \end{proof} \section{Dimension bounds for subspaces of alternating forms} \noindent We begin this section by proving a sharpened version of Lemma \ref{M_u_dimension_bound} for subspaces of $\Alt(V)$. \begin{lemma} \label{M_u_alternating_bound} Let $\M$ be a subspace of $\Alt(V)$ and let $u$ be an element of $V$. Then we have $\dim \M_u\geq \dim \M-(n-1)$. \end{lemma} \begin{proof} The elements of $\M$ are alternating by hypothesis and hence $\epsilon_u(\M)$ vanishes on $u$. It follows that $\epsilon_u(\M)\neq V^*$. Thus $\dim \epsilon_u(\M)\leq n-1$ and the inequality for $\dim \M_u$ is a consequence of the rank-nullity theorem. \end{proof} The next result enables us to establish dimension bounds for subspaces $\M$ of $\Alt(V)$ by induction on $|\rank(\M)|$. \begin{theorem} \label{alternating_dimension_bound} Let $\M$ be a non-zero subspace of $\Alt(V)$ and let $m$ be the largest integer in $\rank(\M)$. Suppose that $|K|\geq m+1$ and for each element $w$ of $V$, $\M_w$ contains an element of rank $m$. Then there exist elements $u$ and $v$ in $V$ such that $\M_u\cap \M_v$ contains no elements of rank $m$ and the inequality \[ \dim \M\leq 2m-1+\dim (\M_u\cap \M_v) \] holds. Furthermore, if $U$ is the two-dimensional subspace of $V$ spanned by $u$ and $v$, and $W$ is a complement of $U$ in $V$, we may identify $\M_u\cap \M_v$ with a subspace of $\Alt(W)$. \end{theorem} \begin{proof} Since $\M$ is non-zero, there must exist $u$ and $v$ in $V$ such that $g(u,v)\neq 0$ for some $g\in \M$. We note that $u$ and $v$ are necessarily linear independent. We will now show that $\M_u\cap \M_v$ contains no elements of rank $m$. For suppose that some form $f$ of rank $m$ is in $\M_u\cap \M_v$. Then $u$, $v$ are in $\rad f$ and hence as $|K|\geq m+1$, Theorem \ref{left_radical_right_radical} implies that $h(u,v)=0$ for all $h\in\M$. This contradicts our earlier assertion about the existence of $g$ in $\M$ with $g(u,v)\neq 0$. We deduce that $\M_u\cap \M_v$ contains no elements of rank $m$, as asserted. We note that each element $\phi$, say, of the subspace $\M_u+\M_v$ of $\M$ satisfies $\phi(u,v)=0$, since $u$ is in the radical of each element of $\M_u$ and $v$ is in the radical of each element of $\M_v$. In view of our earlier statement about the existence of $g$ in $\M$ with $g(u,v)\neq 0$, we see that $\M_u+\M_v\neq \M$ and hence $\dim(\M_u+\M_v)<\dim \M$. It follows that \[ \dim \M_u+\dim \M_v<\dim \M+\dim(\M_u\cap \M_v). \] Now since $\M_u$ and $\M_v$ both contain elements of rank $m$, Lemma \ref{improved_M_u_dimension_bound} shows that \[ \dim \M-m\leq \dim \M_u, \quad \dim \M-m\leq \dim \M_v. \] Thus we have \[ 2(\dim \M-m)<\dim \M+\dim(\M_u\cap \M_v) \] and consequently the inequality \[ \dim \M\leq 2m-1+\dim (\M_u\cap \M_v) \] is valid. Finally, $U$ is contained in the radical of each element of $\M_u\cap \M_v$. It follows that if $W$ is a complement of $U$ in $V$, each element of $\M_u\cap \M_v$ is determined by its restriction to $W\times W$ and thus we may identify $\M_u\cap \M_v$ with a subspace of $\Alt(W)$. \end{proof} \begin{theorem} \label{alternating_constant_rank_bound} Let $\M$ be a non-zero constant rank $m$ subspace of $\Alt(V)$. Then if $|K|\geq m+1$, we have $\dim \M\leq \max\,(n-1, 2m-1)$. \end{theorem} \begin{proof} Suppose that for some $w$ in $V$, $\M_w=0$. Then it follows from Lemma \ref{M_u_alternating_bound} that $\dim \M\leq n-1$, and there is nothing more to prove. We may therefore assume that $\M_w\neq 0$ for all $w\in V$. It follows from Theorem \ref{alternating_dimension_bound} that there exist elements $u$ and $v$ in $V$ such that $\M_u\cap \M_v$ contains no elements of rank $m$. In view of the constant rank $m$ hypothesis, this implies that $\M_u\cap \M_v=0$ and hence \[ \dim \M\leq 2m-1, \] by Theorem \ref{alternating_dimension_bound}, as asserted. \end{proof} Our next step is to prove a version of Theorem \ref{alternating_constant_rank_bound} for subspaces $\M$ of $\Alt(V)$ with $|\rank(\M)|>1$. \begin{theorem} \label{alternating_several_ranks_bound} Let $\M$ be a non-zero subspace of $\Alt(V)$, let $m$ be the largest integer in $\rank(\M)$ and let $r=|\rank(\M)|$. Then if $m\leq \lfloor n/2\rfloor$ and $|K|\geq m+1$, we have \[ \dim \M\leq rn-\frac{r(r+1)}{2}. \] \end{theorem} \begin{proof} We proceed by induction on $r$ and may assume that $r\geq 1$. Suppose that there is an element $w$ of $V$ such that $\M_w$ contains no element of rank $m$. Let $U$ be the one-dimensional subspace of $V$ spanned by $w$ and let $U'$ be a complement of $U$ in $V$. Since $U$ is contained in the radical of each element of $\M_w$, we may identify $\M_w$ with a subspace of $\Alt(U')$. We also have $|\rank(\M_w)|\leq r-1$, since $\M_w$ contains no element of rank $m$. Let $m'$ be the largest integer in $\rank(\M_w)$. We have $m'\leq m-2$, since $m'<m$ and each element of $\Alt(V)$ has even rank. It follows then by a simple inspection that $m'\leq \lfloor (n-1)/2\rfloor$. By induction, we have \[ \dim \M_w\leq (r-1)(n-1)-\frac{(r-1)r}{2}. \] Then since $\dim \M\leq \dim \M_w+n-1$, by Lemma \ref{M_u_alternating_bound}, we obtain that \[ \dim \M\leq (r-1)(n-1)-\frac{(r-1)r}{2}+n-1= rn-\frac{r(r+1)}{2}. \] We have thus completed the induction step in this case. We can now assume that for each element $w$ of $V$, $\M_w$ contains an element of rank $m$. It follows from Theorem \ref{alternating_dimension_bound} that there exist elements $u$ and $v$ in $V$ such that $\M_u\cap \M_v$ contains no elements of rank $m$ and the inequality \[ \dim \M\leq 2m-1+\dim (\M_u\cap \M_v) \] holds. Moreover, we may identify $\M_u\cap \M_v$ with a subspace of $\Alt(W)$, where $W$ is a subspace of $V$ of dimension $n-2$. We clearly have $|\rank(\M_u\cap \M_v)|\leq r-1$. Moreover, if $m''$ is the largest integer in $\rank(\M_u\cap \M_v)$, we have $m''\leq m-2$, as above. Then it is easy to verify that $m''\leq \lfloor (n-2)/2\rfloor$. Thus, by induction \[ \dim (\M_u\cap \M_v)\leq (r-1)(n-2)-\frac{(r-1)r}{2} \] and hence by the inequality in the paragraph above, \[ \dim \M\leq (r-1)(n-2)-\frac{(r-1)r}{2}+2m-1. \] Our hypothesis that $m\leq \lfloor n/2\rfloor$ implies that $2m-1\leq n-1$. We thus have \[ \dim \M\leq (r-1)(n-2)-\frac{(r-1)r}{2}+n-1=rn-\frac{(r^2+3r-2)}{2}. \] It is straightforward to verify that $r^2+3r-2\geq r(r+1)$ when $r\geq 1$. Thus \[ \dim \M\leq rn-\frac{r(r+1)}{2} \] holds for $r\geq 1$ and we have completed the induction step in this second case. Therefore, the dimension bound holds for all $r$, as required. \end{proof} We remark that it easy to see that this dimension bound is optimal in certain cases. On the other hand, the bound does not hold without some restriction on $m$. For example, suppose that $n=2k+1$ is odd and $K$ is a finite field. Then it is possible to construct a subspace $\M$ of $\Alt(V)$ such that $\dim \M=(k-s+1)n$ and $\rank \M=\{2s, 2s+2, \ldots, 2k\}$, where the integer $s$ satisfies $1\leq s\leq k$. Note that when $s=1$, $\M$ is $\Alt(V)$. \section{Constant rank subspaces of alternating forms over finite fields} \noindent The general results we have obtained in Section 3 can be used to deduce reasonably precise information about $n$-dimensional constant rank subspaces of $\Alt(V)$ when we work over sufficiently large finite fields. Our first theorem gives the basic details. \begin{theorem} \label{spread_theorem} Let $\M$ be an $n$-dimensional constant rank $m$ subspace of $\Alt(V)$ and let $K=\mathbb{F}_q$, where $q\geq m+1$. Let $R_1$, \dots, $R_t$ be the different subspaces of dimension $n-m$ in $V$ that occur as the radicals of the elements of $\M^\times$. Then these subspaces form a spread of $V$. As a consequence, $t=(q^n-1)/(q^{n-m}-1)$ and hence $n-m$ divides $n$. Furthermore, if $\M_i$ is the subspace of $\M$ consisting of those elements of $\M$ whose radical contains $R_i$, $1\leq i\leq t$, these $t$ subspaces form a spread of $\M$. \end{theorem} \begin{proof} Let $u$ be any element of $V^\times$. Since $\M$ consists of alternating bilinear forms and $\dim \M=n$, Lemma \ref{M_u_alternating_bound} implies that $\M_u$ in non-zero. Thus, since $\M$ is a constant rank $m$ subspace and $q\geq m+1$, Lemma \ref{improved_M_u_dimension_bound} implies that $d(u)=\dim \M_u\geq n-m$. Lemma \ref{constant_rank_dimension_lemma} then yields that \[ \sum_{u\neq 0}(q^{d(u)}-1)=(q^n-1)(q^{n-m}-1). \] Since $d(u)\geq n-m$, we see that the left hand side above is at least $(q^n-1)(q^{n-m}-1)$, and can only equal $(q^n-1)(q^{n-m}-1)$ if $d(u)=n-m$ for all $u\neq 0$. We deduce that $d(u)=n-m$ for all non-zero $u$. It now follows from Lemma \ref{improved_M_u_dimension_bound} that each element of $\M_u^\times $ has the same radical, $R$, say, where $R$ depends on $u$. Let $w$ be an element of $V$ not in $R$ and let $S$ be the common radical of the elements in $\M_w^\times$. Note that since $w\in S$, $R\neq S$. We claim next that $\M_u\cap \M_w=0$ and $R\cap S=0$. For suppose that $f$ is a non-zero element of $\M_u\cap \M_w$. Then $u\in \rad f=R$ and $w\in \rad f=R$. This is clearly absurd, and we deduce that $\M_u\cap \M_w=0$, as claimed. Similarly, suppose that $v$ is a non-zero element of $R\cap S$. If $g$ is any element of $\M_u^\times$, $v\in \rad g=R$. But equally, if $h$ is any element of $\M_w^\times$, $v\in \rad h=S$. But we also have $g\in \M_v$ and $h\in \M_v$. This implies that $g$ and $h$ have the same radical, as all elements of $\M_v$ have the same radical. This gives the contradiction that $R=S$. We deduce that $R\cap S=0$. It follows that if $R_1$, \dots, $R_t$ are the different subspaces of dimension $n-m$ in $V$ that occur as the radicals of the elements of $\M^\times$, these subspaces form a spread of $V$. We deduce that $t=(q^n-1)/(q^{n-m}-1)$ and hence $n-m$ divides $n$. Likewise, if $\M_i$ is the subspace of $\M$ consisting of those elements of $\M$ whose radical contains $R_i$, $1\leq i\leq t$, these $t$ subspaces form a spread of $\M$. \end{proof} We turn now to considering to what extent the converse of Theorem \ref{spread_theorem} holds: in other words, if $n-m$ divides $n$ (and $m$ is even), is there a constant rank $m$ subspace of $\Alt(V)$ of dimension $n$? As we shall see, the answer is no in general, but there are many cases for which the converse is true. Our starting point is the following construction. Suppose that $U$ is a vector space of odd dimension $k>1$ over an arbitrary finite field. Then there is a $k$-dimensional constant rank $k-1$ subspace, $\N$, say, of $\Alt(U)$. The spread associated with $\N$ is the trivial one, consisting of all one-dimensional subspaces of $U$. See, for example, Theorem 7 of \cite{DGo}. If we take the field to be $\mathbb{F}_{q^t}$, and consider $U$ and $\N$ as vector spaces of dimension $n=kt$ over $\mathbb{F}_{q}$, the trace map from $\mathbb{F}_{q^t}$ to $\mathbb{F}_{q}$ enables us derive a constant rank $m=(k-1)t$ subspace $\M$, say, of $\Alt(V)$, where $\dim \M=\dim V=n=kt$, and $\M$ and $V$ are vector spaces over $\mathbb{F}_{q}$. We have therefore the following result. \begin{theorem} \label{constant_rank_construction} Let $K$ be a finite field and let $n\geq 3$ be a positive integer that is not a power of $2$. Let $k>1$ be an odd divisor of $n$ and let $m=n(k-1)/k$. Then $\Alt(V)$ contains an $n$-dimensional constant rank $m$ subspace. In particular, if $n$ is odd, this construction holds for any divisor of $n$ greater than $1$. \end{theorem} We mention the following non-existence criterion for constant rank subspaces. Suppose that $n=4k$, where $k$ is a positive integer, and $|K|\geq 2k+1$. Then $\Alt(V)$ does not contain a constant rank $2k$ subspace of dimension $n$. This is an immediate consequence of Theorem \ref{alternating_constant_rank_bound}. The remainder of this section is devoted to improving Theorem \ref{alternating_constant_rank_bound} in the case of finite fields. Initially, however, we work over general fields. \begin{definition} \label{R(M)_definition} Let $\M$ be a subspace of $\Alt(V)$. We set \[ V(\M)=\{v\in V: \M_v\neq 0\}. \] \end{definition} Since we are dealing with alternating bilinear forms, $V(\M)$ is identical with the subspaces $V(\M)^L$ and $V(\M)^R$ introduced in Definition \ref{V^L_M_definition}. Lemmas \ref{V(M)_is_a_subspace} and \ref{left_orthogonal_right_orthogonal} show that if $\M$ is a constant rank $m$ subspace of $\Alt(V)$, then $V(\M)$ is a subspace of $V$ that is totally isotropic for $\M$ provided that $\dim \M\geq 2m+1$ and $|K|\geq m+1$. We will show below that an improved result holds if $K$ is a finite field. We begin our analysis by looking at an extreme case. \begin{lemma} \label{gow_quinlan_lemma} Let $\M$ be a constant rank $m$ subspace of $\Alt(V)$ and $K=\mathbb{F}_q$. Suppose that $V(\M)=\rad f$ for some $f\in \M^\times$. Then $\dim \M\leq m/2$. \end{lemma} \begin{proof} We set $R=\rad f$. Our hypothesis implies that each element of $\M^\times$ has radical $R$. We may then identify $\M$ with a constant rank $m$ subspace of $\Alt(V/R)$. Since $\dim (V/R)=m$, we have $\dim \M\leq m/2$ by Lemma 3 of \cite{GQ}. \end{proof} \begin{lemma} \label{R(M)_is_a_subspace} Let $\M$ be a constant rank $m$ subspace of $\Alt(V)$ and $K=\mathbb{F}_q$. Suppose that $q\geq m+1$, $m>2$ and $\dim \M\geq 2m-1$. Then $V(\M)$ is a subspace of $V$ that is totally isotropic for $\M$. \end{lemma} \begin{proof} Let $u$ be an element of $V(\M)$. Lemma \ref{improved_M_u_dimension_bound} shows that $\dim \M_u\geq \dim \M-m$ and moreover, if $\dim \M_u=\dim \M-m$, all elements of $\M_u^\times$ have the same radical. Suppose then that we have $\dim \M_u=\dim \M-m$. Lemma \ref{gow_quinlan_lemma} applied to $\M_u$ implies that $\dim \M_u\leq m/2$ and this in turn implies that $\dim \M\leq 3m/2$. But we are already assuming that $\dim \M\geq 2m-1$ and hence $2m-1\leq 3m/2$. This last inequality implies that $m\leq 2$, which possibility is excluded by hypothesis. We have thus established that $\dim \M_u\geq \dim \M-m+1$ when $u\in V(\M)$. We intend to use this inequality to show that $V(\M)$ is a subspace. It is clear that $V(\M)$ is closed under scalar multiplication. Given elements $u$ and $v$ in $V(\M)$ it suffices then to show that $u+v\in V(\M)$. As in the proof of Lemma \ref{V(M)_is_a_subspace}, this follows if we can show that $\M_u\cap \M_v\neq 0$. We have \[ \dim(\M_u+\M_v)=\dim \M_u+\dim \M_v-\dim (\M_u\cap \M_v)\leq \dim \M. \] Since we have established that $\dim \M_u, \dim \M_v\geq \dim \M-m+1$, we obtain \[ 2(\dim \M-m+1)-\dim (\M_u\cap \M_v)\leq \dim \M. \] Given that we are assuming that $\dim \M\geq 2m-1$, this inequality leads to \[ \dim (\M_u\cap \M_v)\geq \dim \M-2m+2\geq 1 \] and we have thus proved that the intersection is non-zero, as required. Thus $V(\M)$ is a subspace. Finally, we need to show that $V(\M)$ is totally isotropic for $\M$. To achieve this, we must prove that $g(u,v)=0$ for all $u$ and $v$ in $V(\M)$, and all $g$ in $\M$. Now since we have proved that $\M_u\cap \M_v\neq 0$, $u$ and $v$ are both contained in $\rad f$, where $f$ is any non-zero element of $\M_u\cap \M_v$. Since $\rad f$ is totally isotropic for $\M$ by Theorem \ref{left_radical_right_radical}, we see that $g(u,v)=0$ for $g$ in $\M$. This establishes the total isotropy of $V(\M)$. \end{proof} \begin{lemma} \label{dimension_of_R(M)} Let $\M$ be a constant rank $m=2k$ subspace of $\Alt(V)$ and $K=\mathbb{F}_q$. Suppose that $V(\M)$ is a subspace of $V$ and not all elements of $\M^\times$ have the same radical. Then if $\dim \M> n-k$, we have $\dim V(\M)\geq n-m+2$. \end{lemma} \begin{proof} Since $V(\M)$ contains all radicals of elements of $\M^\times$, and we are assuming that not all these radicals are the same, $\dim V(\M)\geq n-m+1$. Suppose now by way of contradiction that $\dim V(\M)=n-m+1$. Since $V(\M)$ contains the radical of each element of $\M^\times$, and each radical has codimension 1 in $V(\M)$, the number of different radicals is at most $(q^{n-m+1}-1)/(q-1)$. On the other hand, let $S$ be the radical of some element of $\M^\times$, and let \[ \M_S=\{ g\in \M: S\leq \rad g\}. \] $\M_S$ is a subspace of $\M$ and Lemma \ref{gow_quinlan_lemma} implies that $\dim \M_S\leq m/2$. If $S'=\rad h$, where $h\in\M^\times$, and $S\neq S'$, we clearly have $\M_S\cap \M_{S'}=0$. It follows that the subspaces $\M_S$ form a partition of $\M$. Thus if $\dim \M=d$, the number of different radicals is at least $(q^d-1)/(q^k-1)$, where $m=2k$. We deduce that the inequality \[ \frac{q^d-1}{q^k-1}\leq \frac{q^{n-m+1}-1}{q-1} \] holds. This leads to the inequality \[ q^{d+1}\leq q^d+q^{n-k+1}-q^k-q^{n-2k+1}+q. \] It is clear, however, that this inequality cannot hold if $d>n-k$. \end{proof} \begin{lemma} \label{R(M)_inequality} Let $\M$ be a constant rank $m$ subspace of $\Alt(V)$ and $K=\mathbb{F}_q$. Suppose that $q\geq m+1$, $m>2$ and $\dim \M\geq 2m-1$. Then we have \[ \dim V(\M)\leq \dim \M_u+n-\dim \M \] for each element $u$ of $V(\M)$. \end{lemma} \begin{proof} Lemma \ref{R(M)_is_a_subspace} shows that $V(\M)$ is a subspace that is totally isotropic for $\M$. The rest of the proof follows from the argument used to prove Lemma \ref{dimension_estimate_in_terms_of_V}. \end{proof} \begin{theorem} \label{new_alternating_dimension_bound} Let $\M$ be a constant rank $m$ subspace of $\Alt(V)$, where $4\leq m\leq \lfloor n/2\rfloor$ and $K=\mathbb{F}_q$. Then if $q\geq m+1$, we have $\dim \M\leq n-2$. \end{theorem} \begin{proof} Suppose if possible that $\dim \M=n-1$. Then since we are assuming that $m\leq \lfloor n/2\rfloor$, we have $\dim \M\geq 2m-1$ and hence Lemma \ref{R(M)_is_a_subspace} implies that $V(\M)$ is a subspace of $V$ that is totally isotropic for $\M$. We assert furthermore that the elements of $\M^\times$ do not all have the same radical. This follows from Lemma \ref{gow_quinlan_lemma}, since $\dim \M>m/2$ in our circumstances. Now since we are assuming that $\dim \M=n-1$, we certainly have $\dim \M>n-m/2$ and hence $\dim V(\M)\geq n-m+2$ by Lemma \ref{dimension_of_R(M)}. Lemma \ref{R(M)_inequality} in turn implies that \[ \dim \M_u\geq n-m+1 \] for all $u$ in $V(\M)$. Lemma \ref{constant_rank_dimension_lemma} shows that we have the equality \[ \sum_{u\neq 0} (q^{d(u)}-1)=(q^{n-1}-1)(q^{n-m}-1), \] where $d(u)=\dim \M_u$. The sum clearly takes place over the non-zero elements $u$ of $V(\M)$. Thus, if we set $d=\dim V(\M)$, we have \[ (\sum_u q^{d(u)})-q^d+1=(q^{n-1}-1)(q^{n-m}-1). \] Consequently, \[ (\sum_u q^{d(u)})-q^d=q^{2n-m-1}-q^{n-1}-q^{n-m}. \] Since we have shown above that $d\geq n-m+2$ and $d(u)\geq n-m+1$, we see that the power of $q$ dividing the left hand side of the sum above is at least $q^{n-m+1}$, whereas the power of $q$ dividing the right hand side is exactly $q^{n-m}$. This is a contradiction, and we deduce that $\dim \M\leq n-2$. \end{proof} We note that the assumption that $m\geq 4$ is essential to this theorem, since for any field $K$, there is a constant rank 2 subspace of $\Alt(V)$ of dimension $n-1$. On the other hand, if $K$ is finite, $\Alt(V)$ contains a constant rank 4 subspace of dimension $n-2$ whenever $n$ is even and at least 4. \section{Dimension bounds for subspaces of bilinear forms} \noindent In this final section, we establish analogues of earlier results which apply to bilinear forms in general. \begin{theorem} \label{bilinear_dimension_bound} Let $\M$ be a non-zero subspace of $\Bil(V)$ that is not contained in $\Alt(V)$ and let $m$ be the largest integer in $\rank(\M)$. Suppose that $|K|\geq m+1$. Suppose also that for each element $w$ of $V$, both $\M_w^L$ and $\M_w^R$ contain elements of rank $m$. Then there exists an element $u$ in $V$ such that $\M_u^L\cap \M_u^R$ contains no elements of rank $m$ and the inequality \[ \dim \M\leq 2m-1+\dim (\M_u^L\cap \M_u^R) \] holds. Furthermore, if $U$ is the one-dimensional subspace of $V$ spanned by $u$ and $W$ is a complement of $U$ in $V$, we may identify $\M_u^L\cap \M_u^R$ with a subspace of $\Bil(W)$. \end{theorem} \begin{proof} Since $\M$ is not contained in $\Alt(V)$, there exists $u$ in $V$ such that $g(u,u)\neq 0$ for some $g\in \M$. We will now show that $\M_u^L\cap \M_u^R$ contains no elements of rank $m$. For suppose that some form $f$ of rank $m$ is in $\M_u^L\cap \M_u^R$. Then $u$ is in both $\rad_L f$ and $\rad_R f$, and hence as $|K|\geq m+1$, Theorem \ref{left_radical_right_radical} implies that $h(u,u)=0$ for all $h\in\M$. This contradicts our earlier assertion about the existence of $g$ in $\M$ with $g(u,u)\neq 0$. We deduce that $\M_u^L\cap \M_u^R$ contains no elements of rank $m$, as asserted. We note that each element $\phi$, say, of the subspace $\M_u^L+\M_u^R$ of $\M$ satisfies $\phi(u,u)=0$, since $u$ is in the left radical of each element of $\M_u^L$ and also in the right radical of each element of $\M_u^R$. In view of our earlier statement about the existence of $g$ in $\M$ with $g(u,u)\neq 0$, we see that $\M_u^L+\M_u^R\neq \M$ and hence $\dim(\M_u^L+\M_u^R)<\dim \M$. It follows that \[ \dim \M_u^L+\dim \M_u^R<\dim \M+\dim(\M_u^L\cap \M_u^R). \] Now since $\M_u^L$ and $\M_u^R$ both contain elements of rank $m$, Lemma \ref{improved_M_u_dimension_bound} shows that \[ \dim \M-m\leq \dim \M_u^L, \quad \dim \M-m\leq \dim \M_u^R. \] Thus we have \[ 2(\dim \M-m)<\dim \M+\dim(\M_u^L\cap \M_u^R) \] and consequently the inequality \[ \dim \M<2m+\dim (\M_u^L\cap \M_u^R) \] is valid. We therefore have $\dim \M\leq 2m-1+\dim (\M_u^L\cap \M_u^R)$. Finally, $U$ is contained in the left and right radical of each element of $\M_u^L\cap \M_u^R$. It follows that if $W$ is a complement of $U$ in $V$, each element of $\M_u^L\cap \M_u^R$ is determined by its restriction to $W\times W$ and thus we may identify $\M_u^L\cap \M_u^R$ with a subspace of $\Bil(W)$. \end{proof} \begin{theorem} \label{bilinear_constant_rank_bound} Let $\M$ be a non-zero constant rank $m$ subspace of $\Bil(V)$. Then if $|K|\geq m+1$, we have $\dim \M\leq \max\,(n,2m-1)$. \end{theorem} \begin{proof} Suppose that for some $w$ in $V$, $\M_w^L=0$ or $\M_w^R=0$ . Then it follows from Lemma \ref{M_u_dimension_bound} that $\dim \M\leq n$, and there is nothing more to prove. We may therefore assume that $\M_w^L\neq 0$ and $\M_w^R\neq 0$ for all $w\in V$. The required dimension bound certainly follows from Theorem \ref{alternating_constant_rank_bound} if $\M$ is a subspace of $\Alt(V)$ and thus we may additionally assume that $\M$ is not contained in $\Alt(V)$. It follows then from Theorem \ref{bilinear_dimension_bound} that there exists $u$ in $V$ such that $\M_u^L\cap \M_u^R$ contains no elements of rank $m$ and $\dim \M\leq 2m-1 +\dim (\M_u^L\cap \M_u^R)$. In view of the constant rank $m$ hypothesis, this implies that $\M_u^L\cap \M_u^R=0$ and hence \[ \dim \M\leq 2m-1, \] as asserted. \end{proof} Following the strategy of previous sections, we may now prove an upper bound for the dimension of a subspace $\M$ of $\Bil(V)$ in terms of $|\rank(\M)|$. \begin{theorem} \label{bilinear_several_ranks_bound} Let $\M$ be a non-zero subspace of $\Bil(V)$, let $m$ be the largest integer in $\rank(\M)$ and let $r=|\rank(\M)|$. Then if $m\leq \lceil n/2\rceil$ and $|K|\geq m+1$, we have \[ \dim \M\leq rn. \] \end{theorem} \begin{proof} We proceed by induction on $r$. If $\M\leq \Alt(V)$, we already have the inequality \[ \dim \M\leq rn-\frac{r(r+1)}{2} \] from Theorem \ref{alternating_several_ranks_bound} and our proposed inequality is clearly a consequence of this sharper inequality. We may therefore assume that $\M$ is not contained in $\Alt(V)$. Suppose next that for some $w$ in $V$, $\M_w^L$ contains no element of rank $m$. Then we certainly have $\dim \M_w^L\leq (r-1)n$ by induction and Lemma \ref{M_u_dimension_bound} yields $\dim \M\leq rn$, as required. An identical argument applies if $\M_w^R$ contains no element of rank $m$. We may thus assume that for all $w\in V$, both $\M_w^L$ and $\M_w^R$ contain elements of rank $m$. Theorem \ref{bilinear_dimension_bound} shows that in this case there exists an element $u$ in $V$ such that $\M_u^L\cap \M_u^R$ contains no elements of rank $m$ and the inequality \[ \dim \M\leq 2m-1+\dim (\M_u^L\cap \M_u^R) \] holds. We certainly have $|\rank(\M_u^L\cap \M_u^R)|\leq r-1$ and we may identify $\M_u^L\cap \M_u^R$ with a subspace of $\Bil(W)$, where $W$ is a subspace of $V$ of dimension $n-1$. Now if $m'$ is the largest integer in $\rank(\M_u^L\cap \M_u^R)$, we have $m'\leq m-1$ and it is easy to verify that, since $m\leq \lceil n/2\rceil$, the inequality $m'\leq \lceil (n-1)/2\rceil$ also holds. It follows by induction that \[ \dim (\M_u^L\cap \M_u^R)\leq (r-1)(n-1). \] Furthermore, since $m\leq \lceil n/2\rceil$, we have $2m-1\leq n$. Thus, our two displayed inequalities above lead to \[ \dim \M\leq n+(r-1)(n-1)=rn-(r-1)<rn. \] This inequality completes the induction step and establishes the theorem. \end{proof} Theorem \ref{bilinear_several_ranks_bound} is optimal in non-trivial cases, as we shall now explain. Suppose that $K$ has a separable extension $L$, say, of finite degree $s>1$. Let $W$ be a vector space of dimension $m$ over $L$. Given any integer $r$ satisfying $1\leq r\leq m$, $\Bil(W)$ contains a subspace $\N$, say, of dimension $rm$ in which all elements have rank at most $r$. We may realize $\N$ by taking all $m\times m$ matrices with entries in $L$ whose first $r$ columns are arbitrary and whose last $m-r$ columns are all zero. Let $V$ denote $W$ considered as a vector space of dimension $n=ms$ over $K$. Given $f$ in $\N$, we define $F$ in $\Bil(V)$ by setting $F(u,v)=\ensuremath\mathrm{Tr}(f(u,v))$ for all $u$ and $v$ in $V$, where $\ensuremath\mathrm{Tr}:L\to K$ is the trace form. We may verify that, since $\ensuremath\mathrm{Tr}$ is non-zero under the separability hypothesis, $\rank F=s\rank f$. The set of all elements $F$, as $f$ runs over $\N$, is then a subspace $\M$, say, of $\Bil(V)$ of dimension $rms=rn$, in which $|\rank(\M)|=r$. The elements of $\M^\times$ have rank $ts$, where $1\leq t\leq r$. This example shows that the dimension bound in Theorem \ref{bilinear_several_ranks_bound} is optimal in this case. It is possible to provide more detail about the structure of a subspace whose dimension equals the upper bound given in Theorem \ref{bilinear_several_ranks_bound}, as we shall now demonstrate. \begin{theorem} \label{subspace_meeting_upper_bound} Let $\M$ be a non-zero subspace of $\Bil(V)$, let $m$ be the largest integer in $\rank(\M)$ and let $r=|\rank(\M)|$. Suppose that $\dim \M=rn$, $m\leq \lceil n/2\rceil$ and $|K|\geq m+1$. Then for each integer $s$ satisfying $1\leq s\leq r$, there exists a subspace $\M_s$, say, of $\M$ such that $\dim \M_s=sn$ and $\rank(\M_s)$ consists of the $s$ smallest integers in $\rank(\M)$. \end{theorem} \begin{proof} We first note that $\M$ is not contained in $\Alt(V)$, since otherwise we have $\dim \M<rn$ by Theorem \ref{alternating_several_ranks_bound}. We claim that there is some $w\in V$ such that either $\M_w^L$ or $\M_w^R$ contains no element of rank $m$. For if this is not the case, Theorem \ref{bilinear_dimension_bound} shows that we can find $u\neq 0$ such that $\M_u^L\cap \M_u^R$ contains no element of rank $m$ and \[ \dim(\M_u^L\cap \M_u^R)\geq \dim \M-(2m-1). \] Now $|\rank(\M_u^L\cap \M_u^R)|\leq r-1$, since $\M_u^L\cap \M_u^R$ contains no element of rank $m$, and hence we have $\dim (\M_u^L\cap \M_u^R)\leq (r-1)n$ by Theorem \ref{bilinear_several_ranks_bound}. It follows that \[ (r-1)n\geq rn-(2m-1) \] and hence $n\leq 2m-1$. This inequality is incompatible with the hypothesis that $m\leq \lceil n/2\rceil$ and hence establishes our claim above. We may as well assume therefore that there is some $w$ such that $\M_w^L$ contains no element of rank $m$ and thus $|\rank(\M_w^L)|\leq r-1$. Lemma \ref{M_u_dimension_bound} implies that $\dim \M_w^L\geq (r-1)n$, whereas Theorem \ref{bilinear_several_ranks_bound} implies that $\dim \M_w^L\leq (r-1)n$. It follows that $\dim \M_w^L=(r-1)n$ and $|\rank(\M_w^L)|= r-1$. We now complete the proof by reverse induction on $r$. \end{proof} It would be valuable to know if Theorem \ref{bilinear_several_ranks_bound} holds in cases where the hypothesis $m\leq \lceil n/2\rceil$ is weakened. We describe below a simple example of a positive answer. \begin{theorem} \label{two_ranks_bound} Suppose that $n$ is even and $\M$ is a subspace of $\Bil(V)$ such that $\rank \M=\{ n/2, n\}$. Then if $|K|\geq n/2+1$, we have $\dim \M\leq 2n$. \end{theorem} \begin{proof} Let $u$ be any element of $V^\times$. $\M_u^L$ cannot contain any elements of rank $n$ and hence must be a constant rank $n/2$ subspace of $\Bil(V)$. Theorem \ref{bilinear_constant_rank_bound} implies that $\dim \M_u^L\leq n$ and then Lemma \ref{M_u_dimension_bound} implies that $\dim \M\leq 2n$, as desired. \end{proof} The discussion before this theorem shows that the bound of Theorem \ref{two_ranks_bound} can be optimal in non-trivial ways.

https://arxiv.org/abs/1607.06600

High girth hypergraphs with unavoidable monochromatic or rainbow edges

A classical result of Erdős and Hajnal claims that for any integers $k, r, g \geq 2$ there is an $r$-uniform hypergraph of girth at least $g$ with chromatic number at least $k$. This implies that there are sparse hypergraphs such that in any coloring of their vertices with at most $k-1$ colors there is a monochromatic hyperedge. We show that for any integers $r, g\geq 2$ there is an $r$-uniform hypergraph of girth at least $g$ such that in any coloring of its vertices there is either a monochromatic or a rainbow (totally multicolored) edge. We give a probabilistic and a deterministic proof of this result.

\section{Introduction} A classical result of Erd\H{o}s and Hajnal \cite{erd66}, Corollary 13.4, claims that for any integers $k, r, g \geq 2$ there is an $r$-uniform hypergraph of girth at least $g$ with chromatic number at least $k$. This implies that there are sparse hypergraphs such that in any coloring of their vertices with at most $k-1$ colors there is a monochromatic hyperedge. The original proof was probabilistic. Other probabilistic constructions were given by Ne\v{s}et\v{r}il and R\"odl \cite{nese78}, Duffus et al. \cite{duffus}, Kostochka and R\"odl \cite{kost10}, and, in case of graphs only, by Erd\H{o}s \cite{erd59}. Several explicit constructions were found later, see Lov\'asz \cite{lov68}, Erd\H{o}s and Lov\'asz \cite{erd75}, Ne\v{s}et\v{r}il and R\"odl \cite{nese79}, Duffus et al. \cite{duffus}, Alon et al. \cite{alon}, K\v{r}\'i\v{z} \cite{kriz89}, Kostochka and Ne\v{s}et\v{r}il \cite{kost99}. Ne\v{s}et\v{r}il \cite{nese13} as well as Raigorodskii and Shabanov \cite{raig11} gave surveys on the topic. Some interesting generalizations were treated by Feder and Vardi \cite{feder98}, Kun \cite{kun}, M\"uller \cite{muller75} , \cite{muller79}, as well as by Ne\v{s}et\v{r}il \cite{nese13}. When the number of colors used on the vertices of a hypergraph is not restricted, the monochromatic hyperedges could easily be avoided by simply using a lot of different colors. Then, however, so-called rainbow (totally multicolored) hyperedges could appear. The notion of a proper coloring when both rainbow and monochromatic hyperedges are forbidden was introduced by Voloshin in a concept called bihypergraphs, \cite{V}, see also Karrer \cite{karrer}. Here, we show that there are sparse hypergraphs in which monochromatic or rainbow hyperedges are unavoidable. {\it A cycle} of length $g$ in a hypergraph is a subhypergraph consisting of $g\geq 2$ distinct hyperedges $E_0, \ldots, E_{g-1}$ and containing distinct vertices $x_0, \ldots, x_{g-1}$, such that $x_i \in E_i\cap E_{i+1}$, $i=0, \ldots, g-1$, addition of indices modulo $g$. The {\it girth} of a hypergraph is the length of a shortest cycle if such exists, and infinity otherwise. Next is our main result. \begin{theorem}\label{main} For any integers $r, g\geq 2$ there is an $r$-uniform hypergraph of girth at least $g$ such that in any coloring of its vertices there is either a monochromatic or a rainbow (totally multicolored) edge. \end{theorem} We shall give a probabilistic proof and an explicit construction of a desired hypergraph. Our proofs are inspired by amalgamation and probabilistic techniques of Ne\v{s}et\v{r}il and R\"odl. To shorten the presentation, we shall say that a hypergraph is {\it rm-unavoidable} if any coloring of its vertices has either a rainbow or a monochromatic edge. We give an explicit construction and use it to prove the main theorem in Section \ref{explicit}. The probabilistic proof is given in Section \ref{probabilistic}. The proofs of a few standard results we use are presented in Appendix. \section{Explicit Construction of rm-unavoidable Hypergraphs}\label{explicit} The goal of this section is to construct, for each $r\geq 2$ and $g\geq 2$, an rm-unavoidable hypergraph, that we shall call $H(r,g)$, of uniformity $r$ and girth $g$. The three main concepts we use are amalgamation, special partite hypergraphs forcing rainbow edges, and so-called complete partite factors. All of these notions are defined for partite hypergraphs. A hypergraph is {\it $a$-partite} if its vertex set can be partitioned in at most $a$ parts such that each hyperedge contains at most one vertex from each part. We shall first define a part-rainbow-forced hypergraph as a hypergraph having some special coloring properties and give an explicit construction of such a hypergraph $PR(r,g)$. Then we incorporate this hypergraph into a more involved construction of an rm-unavoidable hypergraph $H(r,g)$. Both of these constructions use amalgamation.\\ \begin{figure}[htb] \includegraphics[scale=0.8]{Amalgamation2_newformat.pdf} \caption{Amalgamation of $F$ and $H$ along the $4^{{\rm th}}$ part. Here $F$ is a $3$-uniform cycle on $3$ edges, $H$ is $5$-uniform, $5$-partite with $4$ edges. The resulting graph is $5$-partite, $5$-uniform, with curves indicating hyperedges and colors indicating distinct copies of $H$, corresponding to the edges of $F$.} \label{amalgamation} \end{figure} \begin{description} \item[Amalgamation] Given an $a$-partite hypergraph $H$ with the $i^{{\rm th}}$ part of size $r_i$ and given an $r_i$-uniform hypergraph $F=(V, \mathcal{E})$, {\it an amalgamation} of $H$ and $F$ along the $i ^{{\rm th}}$ part, denoted by $H\star_i F$ is an $a$-partite hypergraph obtained by taking $|\mathcal{E}|$ vertex-disjoint copies of $H$ and identifying the $i$th part of each such copy with a hyperedge of $F$ such that distinct copies get identified with distinct hyperedges. Moreover, the $j^{\rm th}$ part of $H\star_i F$ is a pairwise disjoint union of the $j^{{\rm th}}$ parts from the copies of $H$, for $j\in \{1, \ldots, a\} \setminus \{i\}$, see Figure \ref{amalgamation}. We shall sometimes say that $H\star_i F$ is obtained by amalgamating copies of $H$ along the part $i$ using $F$.\\ \item[Part-rainbow-forced hypergraph] A vertex coloring of an $a$-partite hypergraph with parts $X_1, \ldots, X_a$ that assign $|X_i|$ colors to part $i$, $i=1, \ldots, a$ is called {\it part rainbow}. We say that an $a$-partite hypergraph is {\it part-rainbow-forced} if in any part-rainbow coloring there is a rainbow edge.\\ \begin{figure}[htb] \includegraphics[scale=0.8]{Partite-F-factor.pdf} \caption{An example of a complete $4$-partite $F$-factor, where $F$ is a $3$-partite $3$-uniform hypergraph with two edges.}\label{partite-factor} \end{figure} \item[Partite factor] Let $F$ be an $r$-uniform $r$-partite hypergraph. A {\it complete $a$-partite $F$-factor} is an $a$-partite $r$-uniform hypergraph $G$ that is a union of pairwise vertex-disjoint copies $F_1, \ldots, F_{\binom{a}{r}}$ of $F$, such that each part of $F_i$ is contained in some part of $G$, $i = 1, \ldots, \binom{a}{r}$ and such that the union of any $r$ parts of $G$ contains the vertex set of $F_i$, for some $i=1, \ldots, \binom{a}{r}$, see Figure \ref{partite-factor}.\\ \begin{figure}[htb] \includegraphics[scale=1.0]{Extension_figure_entwurf5.pdf} \caption{Extension of an $r$-partite $r$-uniform hypergraph $H_r$ to an $(r+1)$-partite $(r+1)$-uniform hypergraph $\widetilde{H}_{r}$.} \label{extension} \end{figure} \item[Construction of a hypergraph \textbf{\textit{PR(r,g)}}] ~~ Let $r, g \geq 2$, $g\geq 2$ be fixed. Let $g\geq 2$, let $PR(2, g)$ be a bipartite graph on vertices $x,y,z$ and edges $xy$, $yz$. Assume now that $PR(r,g)$ has been constructed and it is an $r$-uniform, $r$-partite hypergraph. Let $F' $ be an $\ell$-uniform hypergraph of girth at least $g$ and minimum degree $\ell (r+1)$, where $\ell = |E(H_r)|$. We show the existence of $F'$ in Appendix. For an $r$-uniform $r$-partite hypergraph $H$, let $\widetilde{H}$ be an $(r+1)$-partite $(r+1)$-uniform hypergraph that is obtained from $H$ by expanding each of its edges by a vertex in a new, $(r+1)^{{\rm st} }$ part such that each edge is extended by an own vertex, i.e., the size of the $(r+1)^{{\rm st} }$ part is equal to the number of edges in $H$, see Figure \ref{extension}. Let $PR(r+1, g) = \widetilde{PR(r, g)} \star _{r+1} F'$, i.e., it is an amalgamation of copies of $\widetilde{PR(r,g)}$ along the $(r+1)^{{\rm st}}$ part using $F'$, see Figure \ref{part-rainbow}. \end{description} ~\\ \begin{figure}[htb] \includegraphics[scale=0.85]{Cycle_newformat.pdf} \caption{Illustration of a part-rainbow-forced $(r+1)$-uniform hypergraph and a cycle of length $3$ in the amalgamated hypergraph $F'$. The bold hyperedges form a cycle of length $11$ in the resulted hypergraph.}\label{part-rainbow} \end{figure} \begin{lemma}\label{Lem2} For any integers $r, g\geq 2$, $PR(r,g)$ is a part-rainbow-forced $r$-uniform hypergraph of girth $g$. \end{lemma} \begin{proof} By construction, $PR(r,g)$ is an $r$-uniform $r$-partite hypergraph, $r\geq 2$. We shall prove by induction on $r$ that $PR(r,g)$ is part-rainbow-forced hypergraph of girth at least $g$. When $r=2$, we see that a part-rainbow coloring assigns distinct colors to $x$ and $z$. Thus, no matter how $y$ is colored, $xy$ or $yz$ is rainbow. Moreover this graph is acyclic, so it has infinite girth. Assume that $PR(r,g)$ is part-rainbow-forced hypergraph of girth at least $g$. Let's prove that $H_{r+1} = PR(r+1,g)$ is also part-rainbow-forced hypergraph of girth at least $g$. Let $H_r= PR(r, g)$. Recall that $H_{r+1}$ is an amalgamation of copies $\widetilde{H}_r^1$, $\widetilde{H}_r^1$, \ldots, $\widetilde{H}_r^{e'}$ of $\widetilde{H_r}$ along the $(r+1)^{{\rm st}}$ part using $F'$, where $F' $ is an $\ell$-uniform hypergraph of girth at least $g$, minimum degree $\ell (r+1)$, $\ell = |E(H_r)|$, and $e' = |E(F')|$. Recall further, that $\widetilde{H}_r^i$ is obtained by an extension operation tilde from $H_r^i$, a copy of $H_r$. First we shall verify that any part-rainbow coloring $c$ of $H_{r+1}$ results in a rainbow edge. For any $i=1, \ldots, e'$, consider a restriction of $c$ to the vertex set of $H_r^i$. Since it is a copy of $H_r = PR(r,g)$, it is again part-rainbow, so there is a rainbow edge $E'_i$ in that copy. Let $ E'_i\cup \{v_i\}$ be a corresponding uniquely defined edge of $\widetilde{H}^i_{r}$. The vertices $v_1, \ldots, v_{e'}$ are vertices of $F'$. Since the minimum degree of $F'$ is at least $\ell (r+1)$, then $e'=|E(F')|\geq |V(F')| \ell(r+1) / \ell = |V(F')| (r+1)$. Thus there are at least $r+1$ repeated vertices in the list $v_1, \ldots, v_{e'}$, i.e., w.l.o.g. $v=v_1= \ldots = v_{r+1}$. Thus $v$ extends rainbow edges $E'_1, E_2', \ldots, E_{r+1}'$ in $H_r^1, H_r^2, \ldots, H_r^{r+1}$. We claim that at least one of the extended edges $E'_1\cup \{v\}, E_2'\cup \{v\}, \ldots, E_{r+1}'\cup\{v\}$ is rainbow. Assume not, then $c(v)$ is present in each of $E'_1, E_2', \ldots, E_{r+1}'$. However, there are at most $r$ vertices of each given color in the first $r$ parts. Since $E'_1, E_2', \ldots, E_{r+1}'$ are pairwise disjoint, we have a contradiction. To see that the girth of $H_{r+1}$ is at least $g$, consider a cycle $C$ in $H_{r+1}$, see bold edges in Figure \ref{part-rainbow}. If the edges of $C$ come from one copy of $\widetilde{H}_r$, then the length of $C$ is at least $g$ as the girth of $\widetilde{H}_r$ is the same as girth of $H_r$. If the edges of $C$ come from at least two distinct copies of $\widetilde{H}_r$, then $C$ is a union of hyperpaths $P_0, P_1, \ldots, P_{m-1}$ from different copies of $\widetilde{H}_r$, such that the consecutive paths share a vertex in the last $(r+1)^{\rm st}$ part, i.e., $V(P_i)\cap V(P_{i+1}) = \{u_i\}$, $u_0, \ldots, u_{m-1}$ are distinct vertices from $V_{r+1}$, addition modulo $m$. Thus $u_{i}$ and $u_{i+1}$ belong to the same copy of $\widetilde{H}_r$ and thus the same edge of $F'$, $i=0, \ldots, m-1$, addition modulo $m$. We see that these edges of $F'$ form a cycle in $F'$ of length at most the length of $C$. On the other hand, we know that any cycle in $F'$ has length at least $g$, implying that $C$ has length at least $g$. This concludes the proof that $PR(r+1, g)$ is part-rainbow-forced of girth at least~$g$. \end{proof} Now we construct an rm-unavoidable hypergraph $H(r,g)$ of uniformty $r$ and girth at least~$g$.\\ \begin{description} \item[Construction of a hypergraph \textbf{\textit{H(r, g)}}] ~~ For $g=2$ and any $r\geq 2$, let $H(r, 2)$ be a complete $r$-uniform hypergraph on $(r-1)^2+1$ vertices. Assume that for any $r\geq 2$, $H(r,g-1)$ has been constructed. Let $F= PR(r, g)$ be as given in the previous construction. Let $a= (r-1)^2 +r$ and let $\mathcal{M}_1$ be a complete $a$-partite $F$-factor. For any partite hypergraph $G$, let $|G|_i$ denote the size of the $i^{{\rm th}}$ part of $G$.\\ Let $\mathcal{M}_2 = \mathcal{M}_1 \star_1 \mathcal{H}_1$, where $\mathcal{H}_1= H(|\mathcal{M}_1|_1, g-1)$. Let $\mathcal{M}_3= \mathcal{M}_2 \star_2 \mathcal{H}_2$, where $\mathcal{H}_2=H(|\mathcal{M}_2|_2, g-1)$. In general, let $\mathcal{M}_{j+1}= \mathcal{M}_j \star_j \mathcal{H}_j$, where $\mathcal{H}_j=H(|\mathcal{M}_j|_j,g-1)$. We see that the $j^{\rm th}$ part of $\mathcal{M}_{j+1}$ corresponds to the vertex set of $\mathcal{H}_j$. Let $H(r,g)=\mathcal{M}_{a+1}$. \end{description} \noindent Now, we shall prove that this construction gives an rm-unavoidable hypergraph that is $r$-uniform and has girth $g$. This will give a proof of Theorem \ref{main}. \begin{proof}[Proof of Theorem \ref{main}] We shall show that $H(r, g)$ is an rm-unavoidable hypergraph of girth at least $g$, by induction on $g$. When $g=2$, $H(r, 2)$ is a compete $r$-uniform hypergraph on $(r-1)^2 +1$ edges. It has girth $2$ and in any vertex coloring there are either $r$ vertices of the same color, forming a monochromatic edge, or $r$ vertices of distinct colors, forming a rainbow edge. Assume that for any $r\geq 2$, $H(r, g-1)$ is an rm-unavoidable hypergraph of girth at least $g-1$. Consider $H(r, g) = \mathcal{M} = \mathcal{M}_{a+1}$ given in the construction. Let $c$ be a vertex coloring of $ \mathcal{M} $. Consider the $a^{\rm th}$ part of $\mathcal{M}=\mathcal{M}_{a+1}$. This part corresponds to the vertex set of $\mathcal{H}_a = H(|\mathcal{M}_a|_a, g-1)$, an rm-unavoidable hypergraph. Thus, there is a monochromatic or rainbow subset $X_a$ in the $a^{\rm th}$ part of $\mathcal{M}$ of size equal to the uniformity of $\mathcal{H}_a$, i.e., of size $|\mathcal{M}_a|_a$. Since $X_{a}\in \mathcal{E}(\mathcal{H}_{a})$, $X_a$ is the $a^{\rm th}$ part of a copy of $\mathcal{M}_a$. Consider $(a-1)^{\rm st}$ part of this copy of $\mathcal{M}_a$. Similarly to the above, there is a monochromatic or rainbow subset $X_{a-1}$ of this part of size equal to the uniformity of $\mathcal{H}_{a-1}= H(|\mathcal{M}_{a-1}|_{a-1}, g-1)$, i.e., of size $|\mathcal{M}_{a-1}|_{a-1}$. Since $X_{a-1}\in \mathcal{E}(\mathcal{H}_{a-1})$, $X_{a-1}$ is the $(a-1)^{\rm st}$ part of a copy of $\mathcal{M}_{a-1}$ such that the $a^{\rm th}$ part of this copy is a subset of $X_a$. Continuing in this manner we see that there is a monochromatic or a rainbow subset $X_{j}$ of $j^{\rm th}$ part of $\mathcal{M}_{j+1}$ of size equal to the uniformity of $\mathcal{H}_{j}$, i.e., of size $|\mathcal{M}_{j}|_{j}$. We have that $X_{j}$ is the $j^{\rm th}$ part of a copy of $\mathcal{M}_{j}$ such that the $(j+t)^{\rm th}$ part of this copy is a subset of $X_{j+t}$, $j+t \in \{j+1, j+2, \ldots, a\}$. Thus $X_1, X_2, \ldots, X_a$ form parts of an $a$-uniform sub-hypergraph of $\mathcal{M}$ containing a copy of $\mathcal{M}_1$. Recall that $\mathcal{M}_1$ is a complete $a$-partite $F$-factor. Each of these parts is monochromatic or rainbow. Since $a= (r-1)^2 + r$, there are either at least $r$ parts that are rainbow or at least $(r-1)^2+1$ parts that are monochromatic. If there are $r$ rainbow parts, the copy of $F$ on these parts contains a rainbow edge as $F$ is part-rainbow-forced. So, assume that there are at least $(r-1)^2+1$ monochromatic parts. If there are $r$ of those that are of the same color, any edge in a copy of $F$ on these parts is monochromatic. Otherwise there are at most $(r-1)$ parts of each given color, so there are $r$ monochromatic parts of distinct colors. These $r$ parts in turn contain an edge of $F$, and since an edge has at most one vertex from each part, this edge is rainbow.\\ Now, we verify that the girth of $\mathcal{M}$ is at least $g$ by an argument similar to one of Lemma~\ref{Lem2}. To do that, we shall prove by induction on $j$, that $\mathcal{M}_j$ has girth at least $g$, $j=1, \ldots, a$. Since $\mathcal{M}_1$ is a complete $a$-partite $F$ factor, it has girth equal to the girth of $F$, that is at least $g$. Assume that $\mathcal{M}_j$ has girth at least $g$. Let's prove that $\mathcal{M}_{j+1}$ has girth at least $g$. Recall that $\mathcal{M}_{j+1} = \mathcal{M}_j \star_j \mathcal{H}_j$, i.e., $\mathcal{M}_{j+1}$ is obtained by amalgamating copies of $\mathcal{M}_j$ along $\mathcal{H}_j = H(|\mathcal{M}_j|_j, g-1)$. Let $X$ be the $j^{\rm th}$ part of $\mathcal{M}_{j+1}$, i.e., the vertex set of $\mathcal{H}_j$. Consider a shortest cycle $C$ in $\mathcal{M}_{j+1}$. If $C$ is a subgraph of one of these copies of $\mathcal{M}_j$, then by induction $C$ has length at least $g$. If the edges of $C$ come from at least two distinct copies of $\mathcal{M}_j$, then $C$ is an edge-disjoint union of hyperpaths $P_0, P_1, \ldots, P_{m-1}$, each with at least $2$ edges, from different copies of $\mathcal{M}_j$, such that the consecutive paths share a vertex in $X$, i.e., $V(P_i)\cap V(P_{i+1}) = \{u_i\}$, $i=0, \ldots, m-1$, and $u_0, \ldots, u_{m-1}$ are distinct vertices from $X$, addition modulo $m$. Thus $u_{i}$ and $u_{i+1}$ belong to the same copy of $\mathcal{M}_j$ and thus correspond to the vertices from the same edge of $\mathcal{H}_j$, $i=0, \ldots, m-1$, addition modulo $m$. We see that these edges of $\mathcal{H}_j$ form a cycle in $\mathcal{H}_j$ of length at most half the length of $C$. On the other hand, we know that any cycle in $\mathcal{H}_j$ has length at least $g-1$, implying that $C$ has length at least $2(g-1)\geq g$. This concludes the proof of Theorem \ref{main} using an explicit construction. \end{proof} \section{Proof of Theorem \ref{main} - Probabilistic Construction}\label{probabilistic} This proof is just a slight generalization of the probabilistic construction for high-girth, high-chromatic-number hypergraphs by Ne\v{s}et\v{r}il and R\"odl. Let an $\ell$-cycle be a cycle of length $\ell$. Let $r, g$ be fixed, put $R = (r-1)^2+1$ and consider an $R$-uniform hypergraph $\mathcal{H}= \mathcal{H}(n,R,g) =(X, \mathcal{E})$ with $n$ vertices, girth at least $ g$, and with $|\mathcal{E}|=\lceil n^{1+\frac{1}{g}}\rceil$. Such a graph exists, if $n$ is large enough by Lemma \ref{girth-edges}, see Appendix. Let's order the hyperedges of $\mathcal{H}$ as $E_1, E_2, \ldots, E_m$. Let $\mathcal{M}_n$ be the family of all sequences $(E_1',\ldots, E_m')$ such that $|E_i'|=r$ and $E_i'\subseteq E_i$, $i=1, \ldots, m$. For a given sequence $Q \in \mathcal{M}_n$, let $\mathcal{H}_Q$ be a hypergraph whose hyperedges are elements of $Q$. We say that a coloring of $X$ is {\it good} for $Q$ if there are no monochromatic and no rainbow edges under this coloring of $\mathcal{H}_Q$. We say that $Q$ is colorable if there is a coloring of $X$ that is good for $Q$. We shall count the number of colorable sequences and shall show that it is strictly less than the number of all sequences in $\mathcal{M}_n$. This will imply that there is a non-colorable sequence corresponding to an rm-unavoidable hypergraph. Each hypergraph $\mathcal{H}_Q$, $Q\in \mathcal{M}_n$ has girth at least $ g$ since $\mathcal{H}$ has this property. In addition $|\mathcal{M}_n|\ge a^{n^{1+\frac{1}{g}}}$, where $a = \binom{R}{r}$, since there are $a$ ways to choose an $r$-element subset from an edge of $\mathcal{H}$ and $m\geq n^{1+\frac{1}{g}}$. Now we consider a coloring of $X$ with arbitrary number of colors. Each edge $E$ of $\mathcal{H}$ is colored with at least $r$ or less than $r$ colors. If $E$ is colored with less than $r$ colors, there are $r$ vertices in $E$ of the same color since $E$ has $R = (r-1)^{2}+1$ elements and $\frac{R}{(r-1)}>(r-1)$. If $E$ is colored with at least $r$ colors, there are $r$ vertices with pairwise distinct colors. Thus each edge $E$ of $\mathcal{H}$ contains a "bad" subset that is either monochromatic or rainbow, and only at most $\binom{|E|}{r} -1 = \binom{R}{r}-1 = a-1$ of all $r$-element subsets of $E$ could be "good". Therefore each coloring $c$ of $X$ is good for at most $(a-1)^{\lceil n^{1+\frac{1}{g}}\rceil} \leq (a-1)^{ 1+n^{1+\frac{1}{g} }} $ members of $\mathcal{M}_n$. Since the total number of colors in $X$ is at most $n$ in any coloring, it is enough to consider colorings with colors ${1,\dots,n}$. Since there are $n^n$ colorings with $n$ colors we have that \begin{eqnarray*} |\{Q \in \mathcal{M}_n| ~Q \text{ is colorable} \}| &=&|\bigcup_{c:X\rightarrow [n]}{\bigcup_{Q \in \mathcal{M}_n} \{Q |~ c \text{ is good for } Q\}}| \\ &\le& \sum_{c:X\rightarrow [n]}{|\bigcup_{Q \in \mathcal{M}_n} \{Q |~ c \text{ is good for } Q\}|} \\ &\le& n^n \cdot (a-1)^{1+ n^{1+\frac{1}{g}}}. \end{eqnarray*} Next we shall show that $n^n \cdot (a-1)^{1+ n^{1+\frac{1}{g}}}< a^{ n^{1+\frac{1}{g}}}$ for all sufficiently large $n$. Indeed, $n^n (a-1)^{1+ n^{1+\frac{1}{g}}}< a^{ n^{1+\frac{1}{g}}} ~\Leftrightarrow~ n \ln(n) + \ln(a-1) < n^{1+\frac{1}{g}} \ln \left( {\frac{a}{a-1}} \right).$ The last inequality holds since $\ln({\frac{a}{a-1}})>0$. Therefore the number of colorable members from $\mathcal{M}_n$ is less than the total number of members in $\mathcal{M}_n$ and thus there is an non-colorable $Q\in \mathcal{M}_n$ that gives $\mathcal{H}_Q$, an $r$-uniform hypergraph of girth at least $g$ that is rm-unavoidable. \qed \bibliographystyle{plain}

https://arxiv.org/abs/1402.2087

Connected Colourings of Complete Graphs and Hypergraphs

Gallai's colouring theorem states that if the edges of a complete graph are 3-coloured, with each colour class forming a connected (spanning) subgraph, then there is a triangle that has all 3 colours. What happens for more colours: if we $k$-colour the edges of the complete graph, with each colour class connected, how many of the $\binom{k}{3}$ triples of colours must appear as triangles?In this note we show that the `obvious' conjecture, namely that there are always at least $\binom{k-1}{2}$ triples, is not correct. We determine the minimum asymptotically. This answers a question of Johnson. We also give some results about the analogous problem for hypergraphs, and we make a conjecture that we believe is the `right' generalisation of Gallai's theorem to hypergraphs.

\section{Introduction} Gallai's colouring theorem (see~\cite{gallai} or~\cite{maffray}) states that if we 3-colour the edges of $K_n$, the complete graph on $n$ vertices, in such a way that each colour class forms a connected spanning subgraph, then there exists a triangle that is \emph{multicoloured}, meaning that no two of its edges have the same colour. \\\\ What happens if we have 4 colours? Let us call a colouring of $K_n$ \emph{connected} if each colour class forms a connected spanning subgraph. So suppose that we have a connected 4-colouring of $K_n$: of the 4 possible triples of colours, how many must appear as the colour set of a multicoloured triangle? It is easy to see that we must have at least 3 triples. Indeed, if no triangle is coloured as 123 or 124 then, viewing the 4-colouring as a 3-colouring with colours 1, 2 and `3 or 4', we would contradict Gallai's theorem. And it is also immediate that we cannot guarantee all 4 triples (at least if $n$ is large): just take colour classes 1, 2 and 3 to be paths that are `completely unrelated' (i.e., the union of them does not contain a triangle), and let colour class 4 be everything else. This does not have any triangle with colours 123. \\\\ Johnson~\cite{johnson} asked: what happens if we have more colours? So suppose that we have a connected $k$-colouring of $K_n$. What is the least number of triples that must appear as the colour sets of multicoloured triangles (perhaps for $n$ large)? There is an obvious guess, namely that we repeat the above: so we let $k-1$ of the colour classes be paths, which are completely unrelated, and the other colour class be everything else. This gives $\binom{k-1}{2}$ triples. Is this the right answer? \\\\ Surprisingly, it turns out that one can do significantly better than this. In Section 2, we give a simple construction to show that the true answer is about $\frac{1}{3}k^2$. \\\\ In Section 3, we turn our attention to the corresponding question for hypergraphs. We concentrate on the 3-uniform case. Perhaps the first attempt to find an analogue of Gallai's theorem would be to ask: if we 4-colour the set of all 3-sets from an $n$-set, in such a way that each colour class is connected (in some sense or other), must there be a 4-set that is multicoloured (i.e. whose 3-sets receive all 4 colours)? There are several different ways to define `connected', but it turns out, as we will see, that even for the strongest notion of connectedness the answer is that we need not have such a 4-set. However, if we return to 3-colourings, and ask for a 4-set whose 3-sets receive all 3 colours, then we do not know what happens. We make various related conjectures, about this case and the $r$-uniform case. \\\\ We remark that Gallai's theorem has been the starting point for a considerable amount of work. For example, Ball, Pultr, and Vojt\v{e}chovsk\'{y}~\cite{ball} considered a special class of Gallai graphs, those where each triangle spans precisely two colours, and Gy\'{a}rf\'{a}s, Sark\"{o}zy, Seb\H{o} and Selkow~\cite{gyarfas} considered Ramsey-type results for Gallai colourings. See also~\cite{fujita,gurvich,gyarfas2,gyarfas3} for related results. \\\\ We write $[k]=\{1,2,\ldots,k\}$. In a $k$-colouring, we usually use colours from $[k]$. We also often refer to `different multicoloured triangles' for multicoloured triangles having different colour sets. \section{Multicoloured triangles in coloured complete graphs} In this section, we consider $f(k)$, the minimum number of triples that can appear as the colour sets of multicoloured triangles in a connected $k$-colouring of $K_n$, for any $n$. (We remark in passing that one might also ask for the minimum provided $n$ is sufficiently large - but in fact, as we will see later in the section, this is the same notion.) \\\\ We start with an easy lower bound of $f(k)$: any connected $k$-colouring of $K_n$ must contain at least $\frac{k(k-2)}{3}$ different multicoloured triangles. This is a consequence of Gallai's theorem and the following simple lemma. \begin{lemma} \label{setlemma} Let $\mathcal{A}$ be a family of subsets of size $3$ of $[k]$ such that whenever we partition $[k]$ into three non-empty subsets, $[k] = R_1 \cup R_2 \cup R_3$, there exists an $A \in \mathcal{A}$ with $A\cap R_i \neq \emptyset$ for $i=1,2,3$. Then $|\mathcal{A}|\geq \frac{k(k-2)}{3}$. \end{lemma} \begin{proof} We show that each element of $[k]$ is in at least $k-2$ sets of $\mathcal{A}$ (whence $|\mathcal{A}|\geq \frac{k(k-2)}{3}$ by double counting). If we fix an element $i\in[k]$ and consider the graph where the edges are induced by the sets containing $i$, then by the condition in the lemma, it is easy to see that this is a connected graph on $k-1$ vertices and so must have at least $k-2$ edges. \end{proof} \noindent For an alternative proof, note that, partitioning $[k]$ into $\{1\} \cup \{2\} \cup \{3,\ldots,k\}$, there must be a set $A_1$ in $\mathcal{A}$ containing $\{1,2\}$ and wlog $A_1 = \{1,2,3\}$. Then partitioning $[k]$ into $\{1\} \cup \{2,3\} \cup \{4,\ldots,k\}$, there must be another set $A_2$ in $\mathcal{A}$ containing $\{1,2\text{ or }3\}$ and wlog $A_2 = \{1,2\text{ or }3,4\}$. Continuing to partition $[k]$ into $\{1\} \cup \{2,3,4\} \cup \{5,\ldots,k\},\{1\} \cup \{2,3,4,5\} \cup \{6,\ldots,k\}, \ldots, \{1\} \cup \{2,\ldots,k-1\} \cup \{k\}$, we can see that there are at least $k-2$ sets in $\mathcal{A}$ containing $1$. \begin{corollary} \label{lowerbound} $f(k)\geq \frac{k(k-2)}{3}$. \end{corollary} \begin{proof} Suppose now that we have a connected $k$-colouring of $K_n$. The subgraph spanned by colours in $R$ is connected for any subset $R$ of $[k]$. If we partition $[k]$ into three non-empty subsets $R_1 \cup R_2 \cup R_3$, Gallai's theorem says that there must exist a multicoloured triangle with colour set intersecting $R_1$, $R_2$ and $R_3$. The family of colour sets of multicoloured triangles now satisfies the condition in Lemma~\ref{setlemma} and hence has size at least $\frac{k(k-2)}{3}$. \end{proof} \noindent We remark that, in the proof of Lemma~\ref{setlemma}, we only considered partitions with a singleton as a class. One might hope to improve this to get a better lower bound on $f(k)$, but the bound in Lemma~\ref{setlemma} is in fact best possible by an inductive construction shown by Diao, Liu, Rautenbach, and Zhao~\cite{diao}. (See the remark after the next result for an explicit construction.) \\\\ From the above lemma and the paths colouring discussed in the Introduction, we have $\frac{k(k-2)}{3} \leq f(k) \leq \frac{(k-1)(k-2)}{2}$. For the case $k=5$, this gives $f(5)=5$ or $6$, and it is natural to believe that the paths colouring would be the best, suggesting $f(5)=6$. But surprisingly, this is not the case. And in fact this paths colouring is not right in general, not even asymptotically. Indeed, we will give another colouring to improve the upper bound of $f(5)$ and in general $f(k)$. \\\\ To be able to have a connected $5$-colouring of $K_n$, we need each subgraph to have at least $n-1$ edges, implying that the minimal complete graph to have a connected $5$-colouring is $K_{10}$, with each colour class forming a tree. However, by going up to $K_{11}$, we are able to find a colouring with more symmetry, which turns out to give fewer multicoloured triangles. This is the case $k=5$ of the following result. \begin{prop} \label{primecolouring} Let $n=2k+1$ be prime. Then there is a connected $k$-colouring of $K_n$ with precisely $\frac{k(k-2)}{3}$ multicoloured triangles. \end{prop} \begin{proof} Let $V(K_n)=\{0,1,2,\ldots,n-1\}$. We can partition the edge set of $K_n$ into $k$ disjoint spanning cycles $C_i$, $i=1,2,\ldots,k$, where $E(V_i)=\{\{ai,(a+1)i\}:a=0,1,2,\ldots,n-1\}$. Here, we use multiplication and addition mod $n$. We now colour each $C_i$ with a different colour. This colouring is definitely connected as each colour class spans a cycle. It is also not hard to check that each colour is in precisely $k-2$ different multicoloured triangles. Hence the size of the family of colour sets of multicoloured triangles is exactly $\frac{k(k-2)}{3}$. \end{proof} \noindent We remark that for the case when $2k+1$ is prime, the family of colour sets of multicoloured triangles in the above colouring provides an explicit (non-inductive) construction attaining the bound in Lemma~\ref{setlemma}. \\\\ The colouring in Proposition~\ref{primecolouring} works for $n=2k+1$ - what about colourings for other values of $n$? For a smaller value of $n$, we note that the minimal complete graph to have a connected $k$-colouring is $K_{2k}$. So we can take the coloured $K_{2k+1}$ in the Lemma~\ref{primecolouring} and delete a vertex from it. Very fortunately, each colour class stays connected. For larger values of $n$, the following simple lemma shows that the above colouring is in fact enough to attain the lower bound of $f(k)$, for each $n\geq 2k$. \begin{lemma} \label{extendlemma} Suppose that there is a connected $k$-colouring of $K_m$ with $l$ different multicoloured triangles. Then, for any $n\geq m$, there is a connected $k$-colouring of $K_n$ with $l$ different multicoloured triangles. \end{lemma} \begin{proof} Let $c'$ be the above colouring of $K_m$. Partition the vertices of $K_n$ into $m$ non-empty vertex classes, $V_1 \cup V_2 \cup \ldots \cup V_m$. For $u_i \in V_i$ and $v_j \in V_j$, we define a colouring $c$ for $K_n$ as follows. \begin{equation*} c(u_iv_j) = \begin{cases} c'(ij) & \text{ if } i\neq j,\\ c'(12) & \text{ if } i = j. \end{cases} \end{equation*} \noindent It is easy to see that $c$ is a connected $k$-colouring of $K_n$ and any multicoloured triangle must have all three vertices from distinct vertex classes. Hence the family of coloured sets of multicoloured triangles of $c$ is exactly the same as the family of colour sets of multicoloured triangles of $c'$. \end{proof} \noindent Combining Proposition~\ref{primecolouring}, Lemma~\ref{extendlemma} and the discussion after Proposition~\ref{primecolouring}, when $2k+1$ is prime we have a connected $k$-colouring of $K_n$ for any $n\geq 2k$ with exactly $\frac{k(k-2)}{3}$ different multicoloured triangles. Together with the lower bound on $f(k)$, this gives the following corollary. \begin{corollary} \label{exact} $f(k)=\frac{k(k-2)}{3}$ when $2k+1$ is a prime.\qed \end{corollary} \noindent When $2k+1$ is not prime, we do not know an explicit connected $k$-colouring attaining the lower bound. Instead, we give an inductive colouring where the number of different multicoloured triangles is close to the lower bound in Corollary~\ref{lowerbound}. \\\\ The following technical lemma states that if a $k$-coloured complete graph satisfies certain conditions, we can extend this colouring to a larger complete graph by adding an extra colour without creating too many new multicoloured triangles. Indeed, only the minimum number (cf. Lemma~\ref{setlemma}) of multicoloured triangles will be created, that is, $k-1$ of them involving this new colour. \begin{lemma} \label{inductivecolouring} Let $c$ be a connected $k$-colouring of $K_n$ with the following properties. \begin{itemize} \item There are exactly $l$ different multicoloured triangles. \item There are exactly $k-2$ different multicoloured triangles using colour $k$. \item The subgraph spanned by colour $k$ is a cycle. \item The edges $v_i v_{i+2}$ have the same colour for all $i\in[n]$. (The subscripts are taken mod $n$, so $v_{n+1}=v_1$ and $v_{n+2}=v_2$.) \end{itemize} Then, there exists a connected $(k+1)$-colouring $c'$ of $K_{2n}$ with the following properties. \begin{itemize} \item There are exactly $l+k-1$ different multicoloured triangles. \item There are exactly $k-1$ different multicoloured triangles using colour $k+1$. \item The subgraph spanned by colour $k+1$ is a cycle. \item The edges $v_i' v_{i+2}'$ have the same colour for all $i\in[2n]$. (The subscripts are taken mod $2n$, so $v_{2n+1}=v_1$ and $v_{2n+2}=v_2$.) \end{itemize} \end{lemma} \begin{proof} Suppose $V(K_n)=\{v_1,v_2,\ldots,v_n\}$ and the subgraph spanned by colour $k$ has edges $v_1v_2,v_2v_3,\ldots,v_nv_1$. \\\\ Let $V(K_{2n})=\{x_1,x_2,\ldots,x_n,y_1,y_2,\ldots,y_n\}$. We define $c'$ on $K_{2n}$ as follows. \begin{eqnarray*} c'(x_ix_j) & = & c(v_iv_j),\\ c'(y_iy_j) & = & c(v_iv_j),\\ c'(x_iy_j) & = & \begin{cases} c(v_iv_j) &\text{ if } j\notin\{i,i+1\},\\ k+1 &\text{ otherwise}. \end{cases} \end{eqnarray*} Here we use addition mod $n$, so $x_{n+1}=x_1$ and $y_{n+1}=y_1$. \\\\ For each $i\in[k]$, the subgraph spanned by colour $i$ in $c'$ is two copies of the subgraph spanned by colour $i$ in $c$ with at least one edge joining them and so connected in $K_{2n}$. The subgraph spanned by colour $k+1$ is just a spanning cycle of $K_{2n}$ and so also connected. Hence, $c'$ is a connected $k+1$-colouring of $K_{2n}$. \\\\ The number of multicoloured triangles not using colour $k+1$ is exactly $l$. The number of multicoloured triangles using colour $k+1$ but not colour $k$ is the same as the number of multicoloured triangles using colour $k$ in $c$, that is $k-2$. And finally, there is only one multicoloured triangle using both colours $k$ and $k+1$. In total, there are $l+k-2+1=l+k-1$ different multicoloured triangles in $c'$ and the number of different multicoloured triangles using colour $k+1$ is precisely $k-1$, proving the lemma. \end{proof} \noindent From Corollary~\ref{exact}, we know the exact values of $f(k)$ for infinitely many $k$. Applying Lemma~\ref{inductivecolouring} to the explicit colourings in Lemma~\ref{primecolouring}, we have good upper bounds for $f(k)$ for all $k$'s between consecutive primes. Finally, to obtain the limit of $\frac{f(k)}{k^2}$, we need to know the gaps between consecutive primes. It is known (see e.g. \cite{hoheisel, baker}) that there exists a constant $\alpha <1$ such that $p_{n+1}-p_n<p_n^\alpha$ for sufficiently large $n$, where $p_n$ is the $n$th prime. This determines $f(k)$ asymptotically. \begin{thm} $f(k) = \frac{k^2}{3}\big(1+o(1)\big)$. \qed \end{thm} \noindent We have shown that $f(k)=\frac{k(k-2)}{3}$ for infinitely many $k$'s, but what is the exact value of $f(k)$ in general? We believe that a colouring attaining the lower bound in Corollary~\ref{lowerbound} always exists, but we have been unable to prove this. \begin{conjecture} $f(k)=\Big\lceil \frac{k(k-2)}{3} \Big\rceil$ for all $k\geq 3$. \qed \end{conjecture} \section{Multicoloured 4-sets in coloured complete 3-graphs} In this section, we wish to find analogues of these results for hypergraphs. We will focus on the case of 3-uniform hypergraphs (or 3-graphs for short). \\\\ An analogue of Gallai's theorem for 3-graphs would be the following statement. Suppose we connectedly (in some sense of connectedness) $4$-colour the edges of the complete 3-graph on $n$ vertices, $K_n^{(3)}$, then must there exist a multicoloured 4-set (that is, a $K_4^{(3)}$ with all its edges having different colours)? \\\\ The notion of connectedness in hypergraphs can be generalised in a natural way from the connectedness of 2-graphs. If we view connectedness as a `1-set property', then this would just be \emph{pointwise connectedness} (although some authors call this `connectedness', see e.g.~\cite{duchet}), that is to say a 3-graph is pointwise connected when there is a path between every pair of vertices, where a \emph{path} is a sequence of intersecting 3-edges. We say a colouring of $K_n^{(3)}$ is a \emph{pointwise connected colouring} if the subgraph spanned by each of the colours is pointwise connected on $n$ vertices. \\\\ It is easy to see that if we take a `cycles' colouring, analogous to the paths colouring from the Introduction, where we take colour classes 1, 2, and 3 to be completely unrelated spanning cycles, and class 4 to be everything else, then this does not contain a multicoloured 4-set. For example, let $n$ be prime and let $V(K_n^{(3)}) = \{0,1,2,\ldots,n-1\}$. We partition the edge set of $K_n^{(3)}$, $E(K_n^{(3)})$ into $\mathcal{A} \cup \mathcal{B} \cup \mathcal{C} \cup \mathcal{D}$, where \begin{eqnarray*} \mathcal{A} & = & \{012,123,\ldots,(n-2)(n-1)0,(n-1)01\},\\ \mathcal{B} & = & \{024,246,\ldots,(n-4)(n-2)0,(n-2)02\},\\ \mathcal{C} & = & \{036,369,\ldots,(n-6)(n-3)0,(n-3)03\},\\ \mathcal{D} & = & E(K_n^{(3)}) \setminus (\mathcal{A} \cup \mathcal{B} \cup \mathcal{C}). \end{eqnarray*} If we colour the edges in each of these sets differently, then each colour spans a pointwise connected subgraph. It is also easy to check that there is no multicoloured 4-set. \\\\ Note that the above example can be generalised to a $k$-colouring of the complete 3-graph in the obvious way. This is to say, there is a pointwise connected $k$-colouring of $K_n^{(3)}$ such that it contains no multicoloured 4-set. \\\\ What if we view connectedness as a 2-set property instead? That is to say, a 3-graph is connected when there is a \emph{strong path}, that is, a path where each of the intersection sizes is precisely two, between every pair of 2-sets. (Note that this is a stronger notion than being a covering, where we say a 3-graph is a \emph{covering} if every 2-set is in some edge. In fact, it is the strongest possible notion of connectness for 3-uniform hypergraphs, apart from topological notions such as spanning a disc.) Formally, and from now onwards, we say a 3-graph $H$ is \emph{connected} if for any $\{u,v\},\{u',v'\}$ in $V(H)^{(2)}$ there is a strong path $P=\{E_1,E_2,\ldots,E_k\}$ in $H$ such that $\{u,v\}\subset E_1$ and $\{u',v'\}\subset E_k$. And similarly, we say a coloured $K_n^{(3)}$ is \emph{connected} if the subgraph spanned by each of the colours is connected on the $n$ vertices. \\\\ With this notion of connectedness for 3-graphs, one might hope to have a direct analogue of Gallai's theorem. However, it turns out that the analogous statement is again false. We will first focus on general $k$-colourings, and will comment on the particular case of $k=4$ afterwards. \\\\ The idea is to inductively blow up a coloured complete 3-graph that contains no multicoloured 4-set and add a new colour to it without creating any multicoloured 4-set. \begin{thm} \label{norainbow} Let $k\geq 1$. Then there is a connected $k$-colouring of $K_n^{(3)}$, for some sufficiently large $n$, with no multicoloured 4-set. \end{thm} \begin{proof} The case $k=1$ is trivial. Suppose $c$ is a connected $k$-colouring of $K_n^{(3)}$ with no multicoloured 4-set. We show that we can $(k+1)$-colour $K_{n^2}^{(3)}$ such that it is connected and does not contain any multicoloured 4-set. \\\\ Let $V\big(K_{n^2}^{(3)}\big) = V_1 \cup V_2 \cup \ldots \cup V_n$, where $V_i=\{v_{ij}:1\leq j \leq n\}$. We define the $(k+1)$-colouring $c'$ as follows. \begin{equation*} c'(v_{ix}v_{jy}v_{lz}) = \begin{cases} c(ijl) & \text{ if $i,j,l$ all distinct},\\ c(xyz) & \text{ if $i,j,l$ not all distinct and $x,y,z$ all distinct},\\ k+1 & \text { otherwise}. \end{cases} \end{equation*} We claim that $c'$ is a connected colouring of $K_{n^2}^{(3)}$. We need to check that the subgraph spanned by colour $s \in [k+1]$, $H_s$ is connected. We shall check that for every pair of 2-sets, $\{v_{ix},v_{jy}\},\{v_{pz},v_{qt}\}$, there is always a strong path in $H_s$ between them. We will do the case when $s\in [k]$. The case $s=k+1$ is similar and hence is left for the reader. \\\\ If all the four vertices are from different blocks or they are all from the same block, it is clear that there is such a path, induced from colouring $c$. Suppose now that they are from three different blocks. There are two cases for this, that is, when $i=j, p\neq q,i\notin \{p,q\}$ and when $i=p,j\neq q,i\notin \{j,q\}$. For the former case, there must be an edge of colour $s$, $E=\{v_{ix},v_{iy},v_{ru}\}$ with $r\notin \{i,p,q\}$ and with the path between $\{v_{ix},v_{ru}\}$ and $\{v_{pz},v_{qt}\}$, induced from colouring $c$, we have the required path. For the latter case, since there is a path of colour $s$ in the colouring $c$ between $\{i,j\}$ and $\{i,q\}$, this induces a path in $H_s$ joining $\{v_{ix},v_{jy}\}$ and $\{v_{iz},v_{qt}\}$. The case when the four vertices are in two different blocks is similar. Hence, $c'$ is indeed a connected colouring. \\\\ Now, we claim that $c'$ does not span a multicoloured 4-set. Let $\{v_{ix},v_{jy},v_{pz},v_{qt}\}$ be a 4-set. If $i,j,p,q$ or $x,y,z,t$ are all distinct, then the colour of the 4-set is the same as a 4-set induced by $c$ on $K_n^{(3)}$, which is not multicoloured. Suppose now that they are in three different blocks, that is, $i=j, p\neq q,i\notin \{p,q\}$, then $c'(v_{ix}v_{pz}v_{qt})=c'(v_{jy}v_{pz}v_{qt})=c(ipq)$, hence not multicoloured. If they are from two different blocks, there are two cases to consider, that is, when $j=p=q, x=y$ and when $i=j,p=q,x=z$. For the former case, we have $c'(v_{ix}v_{pz}v_{qt})=c'(v_{jy}v_{pz}v_{qt})=c(xzt)$, hence not multicoloured. For the latter case, we have $c'(v_{ix}v_{jy}v_{qt})=c'(v_{jy}v_{pz}v_{qt})=c(xyt)$, also not multicoloured. \\\\ We have now exhibited a $(k+1)$-colouring of $K_{n^2}^{(3)}$ such that it is connected and contains no multicoloured 4-set. This completes the proof of the theorem. \end{proof} \noindent The theorem above says that we can connectedly 4-colour the complete 3-graph to avoid any multicoloured 4-set In how small a complete 3-graph can this be done? For example, the above colouring requires $n$, the number of vertices, to be about $3^8=6561$. \\\\ We now show that one may take $n=17$, by giving an explicit connected 4-colouring of $K_{17}^{(3)}$ with no multicoloured 4-set. We suspect that the value of 17 is optimal. \begin{prop} There is a connected 4-colouring of $K_{17}^{(3)}$ with no multicoloured 4-set. \end{prop} \begin{proof} We would like to have a very symmetric colouring, and indeed we will have that any two of our colour classes are isomorphic 3-graphs. Let the vertices of $K_{17}^{(3)}$ be $\{v_0,v_1,\ldots,v_{16}\}$. We define the \emph{distance} of two vertices, $v_i,v_j$ to be $\min\{|i-j|,17-|i-j|\}$. For each edge $v_iv_jv_k$, its `type' is a 3-tuple consisting the three distances of the three pairs of vertices. For example, we say the edge $v_1v_2v_4$ is of type $(1,2,3)$ (or simply type 123 in short). \\\\ All edges of a given type will receive the same colour. Note that there are 8 special types of edges with a repeated distance, namely $\text{type }112, \text{type }224, \ldots, \text{type } 881$. So each colour class should contain 2 of those and 4 other types of edges. \\\\ We are now ready to give a 4-colouring without multicoloured 4-set. Let $\mathcal{C}$ be a set of types of edges, namely $\mathcal{C}=\{112,336,145,235,347,458\}$. For a positive integer $k$, we write $k\mathcal{C}=\{k\times C: C\in \mathcal{C}\}$, where $k\times (a,b,c) = (ka\pmod{17}, kb\pmod{17}, kc\pmod{17})$. (Here, we view $x$ as the same as $17-x$.) \\\\ One can check that $\mathcal{C} \cup 2\mathcal{C} \cup 4\mathcal{C} \cup 8\mathcal{C}$ partitions the types of edge in $K_{17}^{(3)}$. Now we can colour each of the edges of $K_{17}^{(3)}$ by one of four different colours depending on which set its type lies in. \\\\ To check this colouring is indeed connected on $K_{17}^{(3)}$, we can check that in the subgraph spannned by each colour, there is a strong path from $\{v_0,v_1\}$ to every other pair of vertices. For example, from $\{v_0,v_1\}$ to $\{v_5,v_9\}$, we have the path $\{v_0v_1v_2, v_0v_2v_5,v_2v_5v_9\}$ in the subgraph spanned by the colour in correspondence to $\mathcal{C}$. Note that we only need to check for the case $\mathcal{C}$, as the four subgraphs spanned by the four colours are isomorphic. The rest of the cases are similar. \\\\ Suppose now that there is a multicoloured 4-set and one of the edges are from the special types. We may assume that this 4-set is $\{v_0,v_1,v_2,v_x\}$. It is enough to consider the cases when $3\le x\le 9$, and in each of these cases the 4-set is not multicoloured. So a multicoloured 4-set cannot have any special type edge. Suppose now that one of the edges is of type 145; again we may assume that the 4-set is $\{v_0,v_1,v_5,v_x\}$. For each value of $x$, we again claim that the 4-set is not multicoloured. For example, when $x=6$, the edge $v_0v_1v_5$ and the edge $v_1v_5v_6$ have the same colour, and hence not multicoloured. All the remaining cases are similar, and so there is no multicoloured 4-set in this colouring. \end{proof} \noindent From the above, it seems that there is no direct analogue of Gallai's theorem in 3-uniform hypergraphs. But perhaps this is because a multicoloured 4-set is too much to ask for, and maybe we should look for a 3-coloured 4-set instead? \\\\ In each of the colourings of $K_n^{(3)}$ without any multicoloured 4-set we had, there are many 4-sets that have three different edge colours. We say such 4-sets are \emph{tricoloured}. On the other hand, any non-trivial colouring of $K_n^{(3)}$ using at least two colours contains a 4-set that has at least two different edge colours. \\\\ So it is natural to ask: given some connectedness condition on the $k$-colouring of $K_n^{(3)}$, must it always contain a tricoloured 4-set? From the colourings we have on $K_n^{(3)}$ that avoid multicoloured 4-sets, one might hope that, for any connectedness condition we apply, such a colouring must contain a tricoloured 4-set. \\\\ Surprisingly, this is not entirely correct. Indeed, suppose we weaken the condition of connectedness of 3-graphs we had before by only requiring the presence of a path (and not a strong path) between every pair of 2-sets - note that this is exactly the condition of being a covering, as defined earlier. We now give a covering $k$-colouring of $K_n^{(3)}$ (again, this means that every colour class is a covering) without any tricoloured 4-set. This colouring is very similar to the one in Theorem~\ref{norainbow}, but rather easier as we have a weaker notion of connectedness. \begin{lemma} \label{notricoloured} Let $k\geq 1$. Then there is a covering $k$-colouring of $K_n^{(3)}$, for some sufficiently large $n$, with no tricoloured 4-set. \end{lemma} \begin{proof} The case $k=1$ is trivial. Suppose $c$ is a covering $k$-colouring of $K_n^{(3)}$ with no tricoloured 4-set. We want to $(k+1)$-colour $K_{n^2}^{(3)}$ such that it is a covering and does not contain any tricoloured 4-set. \\\\ Let $V\big(K_{n^2}^{(3)}\big) = V_1 \cup V_2 \cup \ldots \cup V_n$, where $V_i=\{v_{ij}:1\leq j \leq n\}$. We define the $(k+1)$-colouring $c'$ as follows. \begin{equation*} c'(v_{ix}v_{jy}v_{lz}) = \begin{cases} c(ijl) & \text{ if $i,j,l$ all distinct},\\ c(xyz) & \text{ if $i=j=l$},\\ k+1 & \text { otherwise}. \end{cases} \end{equation*} As in the proof of Theorem~\ref{norainbow}, it is not hard to check that $c'$ is in fact a covering $(k+1)$-colouring of $K_{n^2}^{(3)}$ without any tricoloured 4-set. \end{proof} \noindent Despite the above colouring with no tricoloured 4-set, we still believe that every connected $k$-coloured $K_n^{(3)}$ must contain an tricoloured 4-set. This is our conjectured extension of Gallai's theorem. \begin{conjecture} For all sufficiently large $n$, every connected $3$-colouring of $K_n^{(3)}$ must contain a tricoloured 4-set. \qed \end{conjecture} \section{Further remarks and questions} We remarked after Proposition~\ref{primecolouring} that of course $K_{2r}$ is the minimal complete graph to have a connected $k$-colouring, because a connected 2-graph on $n$ vertices must have at least $n-1$ edges. In order to determine the minimal complete 3-graph having a connected $k$-colouring, we need to know the minimal number of edges of a connected 3-graph on $n$ vertices. We have the following simple result. \begin{lemma} Let $H_n$ be a connected 3-graph on $n$ vertices. Then $\big|E(H_n)\big| \geq \big\lfloor \frac{1}{2}\binom{n}{2} \big\rfloor$. Moreover, this bound can be obtained. \end{lemma} \begin{proof} To show the lower bound, we construct a connected 2-graph $G_n$ on $\binom{n}{2}$ vertices from $H_n$. Let the vertex set of $G_n$ indexed by the 2-sets of vetices of $H_n$. For each edge $v_iv_jv_k$ in $H_n$, we add three edges $(v_iv_j)(v_iv_k)$, $(v_iv_j)(v_jv_k)$ and $(v_iv_k)(v_jv_k)$ to $G_n$. By the connectedness of $H_n$, we can see that $G_n$ is connected. In fact, if we delete one of the three edges added to $G_n$ from each edge $v_iv_jv_k$ in $H_n$, $G_n$ remains connected. \\\\ By construction, $G_n$ has $2\big|E(H_n)\big|$ edges and together with the fact that $G_n$ being connected implies that it has at least $\binom{n}{2} - 1$ edges, implying $H_n$ must have at least $\big\lfloor \frac{1}{2}\binom{n}{2} \big\rfloor$ edges. \\\\ For the upper bound, we show by inductive constructions that there is a connected 3-graph on $n$ vertices with $\big\lfloor \frac{1}{2}\binom{n}{2} \big\rfloor$ edges. \\\\ We first deal with the case when $n$ is even. Given $H_n$ with $V(H_n)=\{x_1,\ldots,x_k,y_1,\ldots,y_k\}$, we construct $H_{n+4}$ as follows. \begin{eqnarray*} V(H_{n+4})&:=&V(H_n) \cup \{a,b,c,d\}, \\ E(H_{n+4})&:=&E(H_n) \cup \{ax_iy_i:1\leq i\leq k\} \cup \{bx_iy_i:1\leq i\leq k-1\} \cup \\ & & \{cx_iy_i:1\leq i\leq k\} \cup \{dx_iy_i:1\leq i\leq k\} \cup \{abx_k,abc,acd,bdy_k\}. \end{eqnarray*} \noindent It is not hard to check that $H_{n+4}$ is connected if $H_n$ is connected. We need two base cases, that is, when $n=2,4$. For $n=2$, we can simply take $H_2$ to be the empty 3-graph on two vertices and for $n=4$, we can take $H_4$ to be the complete 3-graph on four vertices taking away an edge. Now $|E(H_{n+4})| = |E(H_n)|+ 2n +3 = \big\lfloor \frac{1}{2}\binom{n}{2} \big\rfloor + 2n+3 = \big\lfloor \frac{1}{2}\binom{n+4}{2} \big\rfloor$. \\\\ We can now construct a connected 3-graph on $n+1$ vertices from one on $n$ vertices, with $n$ being even. Given $H_n$ with $V(H_n)=\{x_1,\ldots,x_k,y_1,\ldots,y_k\}$, we construct $H_{n+1}$ as follows. \begin{eqnarray*} V(H_{n+1})&:=&V(H_n) \cup \{a\}, \\ E(H_{n+1})&:=&E(H_n) \cup \{ax_iy_i\}:1\leq i \leq k \}. \end{eqnarray*} \noindent It is straightforward to check that $H_{n+1}$ is indeed connected and $|E(H_{n+1})| = |E(H_n)|+ \frac{n}{2} = \big\lfloor \frac{1}{2}\binom{n}{2} \big\rfloor + \frac{n}{2} = \big\lfloor \frac{1}{2}\binom{n+1}{2} \big\rfloor$. \end{proof} \noindent In Section 3, we tried to extend Gallai's theorem to hypergraphs. Returning to graphs, we could also ask, what about a multicoloured $K_d$ in a connectedly $k$-coloured $K_n$, for any $d>3$? The exact same paths colouring we had in the Introduction shows that there exists a connectedly $k$-coloured $K_n$ without any multicoloured $K_d$. But another question would be, how many colours must some $K_d$ have in a connected $k$-colouring of $K_n$? For example, if we have a connected $6$-colouring of $K_n$, then there must exist a $K_4$ that spans at least four colours - this is a simple consequence of Gallai's theorem plus the fact that every vertex is incident with edges of all colours. In the other direction, we can take five disjoint paths on $n$ vertices such that the union of them contains no cycles of length at most 4 and give the paths colouring (as in the Introduction) to deduce that every $K_4$ spans at most four colours. \begin{prop} Let $3\leq d \leq k$. Then there is a $K_d$ that spans at least $d$ colours in any connectedly $k$-coloured $K_n$. Moreover, for all sufficiently large $n$, there exists a connectedly $k$-coloured $K_n$ with no $K_d$ spanning more than $d$ colours. \end{prop} \begin{proof} As above, the first statement is a simple consequence from Gallai's theorem plus the fact that every vertex is incident with edges of all colours. \\\\ The latter statement is trivially true for $d=k$. For $d<k$, we can take $k-1$ disjoint paths on $n$ vertices such that the union of them contains no cycles of length at most $d$ and give the paths colouring as the one mentioned in the introduction, that is, colour each of the spanning paths by a different colour and the rest of the edges by another colour, say green. Suppose there is a $K_d$ that spans $d+1$ colours, then there are at least $d$ non-green edges on these $d$ vertices, which implies that there is a cycle of length at most $d$ from the union of these paths, contradicting the assumption. \end{proof} \noindent Until now we have focused on graphs and 3-uniform hypergraphs, but it is natural to seek extensions to the case of general $r$-uniform hypergraphs. As before, we say that an $r$-graph is \emph{connected} if there is a strong path between every pair of $(r-1)$-sets. Here, a strong path is a sequence of $r$-edges where each consecutive pair of $r$-edges has intersection size precisely $r-1$. Again, we say a coloured $K_n^{(r)}$ is \emph{connected} if each colour class spans a connected subgraphs. It appears that the interesting case is still for 3 colours. \begin{conjecture} For all sufficiently large $n$, if we connectedly $3$-colour the edges of the complete $r$-graph on $n$ vertices, then there must exist an $(r+1)$-set that uses all three colours. \qed \end{conjecture} \noindent A slightly weaker notion would be to use covering, where we say an $r$-graph is a \emph{covering} if every $(r-1)$-set is in some $r$-edge. We say a colouring of the complete $r$-graph is \emph{covering} if each colour class spans a covering. \\\\ Unfortunately, as with 3-graphs (Lemma~\ref{notricoloured}), it is again not true that every weakly connected 3-colouring of a complete 4-graph contains a 5-set that uses all three colours. \begin{lemma} For all sufficiently large $n$, there is a covering 3-colouring of $K_n^{(4)}$ with no 5-set that uses all three colours. \end{lemma} \begin{proof} Suppose $c$ is a covering red/blue colouring of $K_n^{(4)}$ and $d$ is a covering blue/green colouring of $K_n^{(4)}$. \\\\ Let $V\big(K_{n^2}^{(3)}\big) = V_1 \cup V_2 \cup \ldots \cup V_n$, where $V_i=\{v_{ij}:1\leq j \leq n\}$. We can view this as the blow-up of $K_n^{(4)}$ of colouring $d$ with $n$ copies of $K_n^{(4)}$ of colouring $c$. There are three other different types of 4-edges to be coloured. Formally, we define the 3-colouring $c'$ as follows. \begin{equation*} c'(v_{ix}v_{jy}v_{pz}v_{qt}) = \begin{cases} d(ijpq) & \text{ if $i,j,p,q$ all distinct},\\ c(xyzt) & \text{ if $i=j=p=q$},\\ red & \text{ if $\big|\{i,j,p,q\}\big|=3$},\\ blue & \text{ if $i=j,p=q,i\ne p$},\\ green & \text { if $i=j=p,q\ne i$}. \end{cases} \end{equation*} It is now straightforward to check that $c'$ is in fact a covering 3-colouring of $K_{n^2}^{(4)}$ without any $K_5^{(4)}$ that uses all three colours. \end{proof} \noindent It seems that the above inductive colouring works because we are lucky to have exactly three colours, namely one to colour each of the three extra types of 4-edges to maintain the connectivity of the blow-up $K_{n^2}^{(4)}$. In fact, we do not see how to generalise this to greater values of $r$, even when we are allowed to use more colours. \\\\ Finally, returning to Theorem~\ref{norainbow}, it would be interesting to know what happens if the notion of connectedness is strengthened to some topological notion of connectedness (to do with the simplicial complex formed by the triples in each colour class): this is an idea of Thomass\'{e}~\cite{thomasse}.

https://arxiv.org/abs/1202.3065

Partially ample line bundles on toric varieties

In this note we study properties of partially ample line bundles on simplicial projective toric varieties. We prove that the cone of q-ample line bundles is a union of rational polyhedral cones, and calculate these cones in examples. We prove a restriction theorem for big q-ample line bundles, and deduce that q-ampleness of the anticanonical bundle is not invariant under flips. Finally we prove a Kodaira-type vanishing theorem for q-ample line bundles.

\section{$q$-ample line bundles} In the 1950s Serre gave a cohomological characterisation of ample line bundles: a line bundle is ample if and only if some sufficiently high power of it kills cohomology of any coherent sheaf in degrees above zero. This characterisation suggests the following generalisation of ampleness, introduced by Sommese \cite{Sommese}. (Note that Sommese's definition requires that some power of the line bundle be globally generated, but we drop that hypothesis here.) \begin{definition} \label{def-naive} Let $X$ be a projective variety. A line bundle $L$ on $X$ is called {\it $q$-ample} (for some integer $q \geq 0$) if for any coherent sheaf $F$ on $X$, there exists a natural number $n_0$ (depending on $F$) such that \begin{align*} H^i(X, L^n \otimes F) &=0 \text{ for all } i>q \text{ and } n \geq n_0. \end{align*} \end{definition} Any line bundle on a variety of dimension $n$ is $n$-ample; by Serre, $0$-ample is the same as ample. At first sight the $q$-ample condition seems hard to check since it involves tensoring with an arbitrary coherent sheaf. Totaro \cite[Definition 6.1]{Totaro} introduced the related notion of {\it $q$-T-ampleness}, which has the key advantage that it is defined by the vanishing of finitely many cohomology groups: \begin{definition} Let $X$ be a projective variety of dimesion $n$, and fix a Koszul-ample line bundle $\mathcal{O}_X(1)$ on $X$. A line bundle $L$ on $X$ is called {\it $q$-T-ample} if there exists a natural number $N$ such that \begin{align*} H^{q+1}(X,L^N(-n-1))=H^{q+2}(X,L^N(-n-2))=\cdots =H^n(X,L^N(-2n+q))=0. \end{align*} \end{definition} For the definitition of Koszul-ample see \cite{Totaro}; the precise meaning will not play an important role here. For our purposes it will be sufficient to know that: \begin{compactenum} [i)] \item some sufficiently high power of any ample line bundle is Koszul-ample \cite{Backelin}, \item the definition of $q$-T-ampleness is independent of the choice of Koszul-ample line bundle \cite[Corollary 6.2, Theorem 6.3]{Totaro}. \end{compactenum} The significance of this definition is that the two notions of $q$-ampleness coincide in characteristic zero \cite[Theorem 6.3]{Totaro}: \begin{theorem}[Totaro] \label{thm-qtample} Let $X$ be a projective variety over a field of characteristic zero. Then line bundle $L$ on $X$ is $q$-ample if and only if it is $q$-T-ample. \end{theorem} So in charactersistic zero we can check $q$-ampleness by looking at the cohomology of a finite set of line bundles. Following \cite[Section~8]{Totaro}, we define an $\text{\bf R}$-divisor to be $q$-ample if it is numerically equivalent to a sum $cD + A$ with $D$ a $q$-ample divisor, $c$ a positive rational number, and $A$ an ample $\text{\bf R}$-divisor. The $q$-ample cone $Amp_q(X)$ is the cone of classes of $q$-ample $\text{\bf R}$-divisors in $N^1(X)$. (Here $N^1(X)$ denotes the real vector space $\left( \mathrm{Div}(X)/ \equiv \right) \otimes \text{\bf R}$, where $\mathrm{Div}(X)$ is the group of all Cartier divisors on $X$, and $\equiv$ denotes numerical equivalence.) We emphasise that $Amp_q(X)$ is not a convex cone in general: typically the tensor product of two $q$-ample line bundles is only $2q$-ample, not $q$-ample. We will use Theorem~\ref{thm-qtample} to show that the $q$-ample cone of a simplicial (equivalently, $\text{\bf Q}$-factorial) projective toric variety in characteristic zero is the interior of a finite union of rational polyhedral cones. Roughly speaking, each vanishing condition in the definition of $q$-T-ample defines a polyhedral region, and these regions combine to give the $q$-ample cone. For later reference we state Totaro's result \cite[Theorem 8.3]{Totaro} (building on work of Demailly--Peternell--Schneider) that $q$-ampleness is an open condition in the space of line bundles up to numerical equivalence. \begin{theorem}[Demailly--Peternell--Schneider, Totaro] \label{thm-open} Let $X$ be a projective variety over a field of characteristic zero. Then the $q$-ample cone $Amp_q(X)$ is open in $N^1(X)$. \end{theorem} \section{Cohomology of line bundles on toric varieties} \label{sect-2} In this section we outline the description \cite{Broomhead} of cohomology of line bundles on toric varieties in terms of constructible sheaves on the polytope. We work over an algebraically closed field $\mathbf{k}$ of characteristic zero. Let $X = X(\Delta)$ be a simplicial projective $n$-dimensional toric variety, corresponding to some complete fan $\Delta$ in a lattice $N \cong \text{\bf Z}^n$. We denote the primitive generators of the rays of $\Delta$ by $\{v_i \in N \mid i \in I \}$. There is a one-to-one correspondence between prime torus-invariant divisors and rays of $\Delta$ \cite[Chapter 3]{Fulton}. We denote these divisors by $\{E_i \mid i \in I\}$ and the free group generated by them by $\text{\bf Z}^{I}$. The dual space $\text{\bf Z}_I$ is generated by the dual basis $\{e_i \mid i \in I\}$. Let $ M := \operatorname{Hom}(N, \text{\bf Z}) \cong \text{\bf Z}^n$ be the dual lattice to $N$, with pairing $ \langle \cdot , \cdot \rangle $. We have the following commutative diagram with exact rows \begin{equation} \label{eqn:seseq} \xymatrix{ 0 \ar[r] & M \ar@{=}[d] \ar[r] & \mathrm{Div}_T(X) \ar@{^{(}->}[d] \ar[r] & \Pic(X) \ar@{^{(}->}[d] \ar[r] & 0 \\ 0 \ar[r] & M \ar[r] & \text{\bf Z}^{I} \ar[r]^{[-]} & Cl(X) \ar[r] & 0 \\ } \end{equation} where $\mathrm{Div}_T(X)$ is the group of torus-invariant Cartier divisors, and $Cl(X)$ is the class group. Applying the functor $-\otimes_{\text{\bf Z}} \text{\bf R}$ and using the fact that simplicial toric varieties are $\text{\bf Q}$-factorial, we obtain the following exact sequence \begin{equation} \label{eqn:seseq2} \xymatrix{ 0 \ar[r] & M_\text{\bf R} \ar[r] & \text{\bf R}^{I} \ar[r]^{[-]} & N^1(X) \ar[r] & 0 \\ } \end{equation} where $N^1(X)$ denotes $\Pic(X) \otimes \text{\bf R}$. We denote the polytope of the toric variety $X$ by $P_X$. To fix notation, let us recall how this polytope is constructed. For an ample divisor $A = \sum_{i \in I} a_i E_i$. on $X$, there is an associated $n$-dimensional polytope in $M_\text{\bf R}$ given by \begin{equation*} P_A := \{u \in M_\text{\bf R} \mid \langle u , v_i \rangle \geq -a_i \; \forall i \in I\}. \end{equation*} This has $(n-1)$-dimensional faces \[ F_i:= \{u \in M_\text{\bf R} \mid \langle u , v_i \rangle = -a_i \text{ and } \; \langle u , v_j \rangle \geq -a_j \; \forall j \in I\backslash \{i\} \} \] indexed by the rays of $\Delta$. The combinatorial structure of this polytope is independent of the choice of ample divisor $A$ \cite[p. 72]{Fulton}, so we define $P_X$ to be the polytope $P_A$ for any ample divisor $A$. ~\newline If $D$ is any torus-invariant Cartier divisor on $X$, its cohomology $H^p(X,\mathcal{O}(D))$ splits into a direct sum of weight spaces indexed by the lattice $M$ \cite[Section 3.5, Proposition]{Fulton}: \begin{equation} \label{dsd} H^p(X,\mathcal{O}(D)) \cong \bigoplus_{m \in M}H^p(X,\mathcal{O}(D))_m. \end{equation} There are several known descriptions of these weight spaces, for example Demazure's description \cite{Demazure} in terms of local cohomology groups. Here we will use the description from \cite{Broomhead} in terms of cohomology of a constructible sheaf on the polytope $P_X$. More precisely, if $D = \sum_{i \in I} d_i E_i$ is the torus-invariant divisor, then for each $m \in M$ we define a closed subset of $P_X$ by \begin{align*} Z(m,D):= \bigcup_{ \{i \in I \mid \langle m , v_i \rangle < -d_i \} } F_{i} \quad = \quad \bigcup_{ \{i \in I \mid \langle m +D , e_i \rangle < 0 \} } \! \! F_{i} \, , \end{align*} a union of closed maximal dimensional faces of $P_X$, and denote by $W(m,D):= P_X \setminus Z(m,D)$ the complementary open subset of $P_X$. Let $j:W(m,D)\hookrightarrow P_X $ be the inclusion of $W(m,D)$ into $P_X$ and $ \text{\bf k}_{W}$ be the constant sheaf on $W$ with values in the ground field. Then \cite[Theorem~1.1]{Broomhead} states that for all $p \geq 0$ there are canonical isomorphisms \begin{equation} \label{eqn:pieces} H^p(X,\mathcal{O}(D))_m \cong H^p(P_X,j_! \text{\bf k}_{W(m,D)}). \end{equation} This leads to a practical method of computation as follows. There is a short exact sequence of constructible sheaves \begin{equation*} \label{seseq3} 0 \lra{} {j}_!{j}^*\text{\bf k}_{P_X} \lra{} \text{\bf k}_{P_X} \lra{} {i}_*{i}^*\text{\bf k}_{P_X} \lra{} 0 \end{equation*} where $i$ is the inclusion of $Z(m,D)$ into $P_X$, which induces a long exact sequence of cohomology \begin{equation*} \label{long} \dots \lra{} H^{p-1}(Z(m,D), \text{\bf k}) \lra{} H^{p}(P_X,j_! \text{\bf k}_{W(m,D)}) \lra{} H^{p}(P_X, \text{\bf k}) \lra{} \dots \end{equation*} Using this together with Equation~(\ref{eqn:pieces}) and observing that $P_X$ is contractible, we conclude that for all $p \geq 0$ we have \begin{equation}\label{reduced} H^p(X,\mathcal{O}(D))_m = \widetilde{H}^{p-1}(Z(m,D), \text{\bf k}) \end{equation} where $\widetilde{H}$ denotes reduced cohomology. (We use the convention that $\widetilde{H}^{-1}(Z, \text{\bf k})=0$ except when $Z = \emptyset$, and $\widetilde{H}^{-1}(\emptyset, \text{\bf k})=\text{\bf k}$.) In order to compute cohomology groups it is therefore important to understand the sets $Z(m,D)$. For each subset $\alpha \subseteq I$, we define the set \[ Z_\alpha := \bigcup_{ i \in \alpha } F_{i} \] and note that each $Z(m,D)$ is of this form. Given $\alpha \subseteq I$ we define a corresponding orthant in $\text{\bf R}^I$ as follows \[ O_\alpha := \{ D=\sum_{i \in I} d_i E_i \in \text{\bf R}^I \mid d_i < 0 \; \forall i\in \alpha, \; d_i \geq 0 \; \forall i\in I\backslash \alpha \}. \] With these definitions the following lemma is immediate: \begin{lemma}\label{orthlemm} $Z(m,D) = Z_\alpha$ if and only if $D+m \in O_\alpha$. \end{lemma} \section{The $q$-ample cone of a toric variety} In this section we will prove our main theorem. First we collect a few basic facts about rational cones that we will need for the proof. A subset of a vector space is a (closed, not necessarily strongly) convex polyhedral cone if and only if it is the intersection of a finite number of closed half-spaces. We will refer to the intersection of a finite number of half-spaces, some of which are open, as a partially open convex polyhedral cone. \begin{lemma}\label{orthanttocone} The image $[O_\alpha]$ of an orthant $O_\alpha$ in $N^1(X)$ is an intersection of a finite number of rational half-spaces (some of which may be open). Its closure $\overline{[O_\alpha]}$ is a convex rational polyhedral cone of top dimension. \hfill $\square$ \end{lemma} \begin{lemma}\label{complementcones} Let $C_1, \dots ,C_t$ be convex rational polyhedral cones (some of which may be partially open). The complement of their union is also a finite union of convex rational polyhedral cones (some of which may be partially open). \end{lemma} \begin{proof} For a single cone $C_1$, we write this as an intersection of half spaces. The set of (rational) hyperplanes bounding these half spaces split the complement of $C_1$ up into a finite number of convex rational polyhedral cones (some of which are partially open). The lemma then follows for a finite union $\bigcup_{j=1}^t C_j$ using De Morgan's law and the fact that the intersection of two rational polyhedral cones is again a rational polyhedral cone. \end{proof} Now we can prove our main theorem. \begin{theorem} \label{thm-main} The $q$-ample cone of a simplicial (equivalently, $\text{\bf Q}$-factorial) projective toric variety $X$ over an algebraically closed field of characteristic zero is the interior of a union of finitely many rational polyhedral cones. \end{theorem} \begin{proof} We denote by $n$ the dimension of $X$ and let $\mathcal{L} := \mathcal{O}(A) \in \Pic(X)$ be a Koszul-ample line bundle. We start by proving that the set \[S = \left\{ \mathcal{O}(D) \in \Pic(X) \mid H^i(X,\mathcal{O}(kD+(q-i-n)A) ) =0, \; \forall i>q, \; \forall k \gg 0 \right\}. \] is the set of classes of all $q$-ample Cartier divisors on $X$. In one direction, the elements of $S$ satisfy a condition which obviously implies $q$-$T$-ampleness, and are therefore classes of $q$-ample Cartier divisors by Theorem~\ref{thm-qtample} In the other direction, suppose that $D$ is a $q$-ample Cartier divisor. Then for any given coherent sheaf $E$ on $X$, there exists a positive integer $N$ (depending on $E$) such that $H^i(kD+E)=0$ for all $k \geq N$, in all degrees $i>q$. Applying this with $E$ chosen to be $(-n-1)A, (-n-2)A,\dots ,(-2n+q)A$, we get natural numbers $N_1, N_2,\dots ,N_{n-q}$ such that for each $j$, and each $k \geq N_j$, we have $H^i(kD+(q-j-n)A)=0$ for all $i>q$. Now set $N=\mathrm{max }\left\{N_j \right\}$: then for each $j=q+1,\dots ,n$ and each $k \geq N$ we have $H^i(kD+(q-j-n)A)=0$ for all $i>q$. In particular, choosing $i=j$, we see that the class of $D$ lies in $S$. Therefore, the $q$-ample cone $Amp_q(X)$ is the cone of $\text{\bf R}$-divisor classes of the form $[cD +A^\prime]$ where $[D] \in S$, $c$ is a positive rational number and $A^\prime$ is an ample $\text{\bf R}$-divisor. It remains to prove that this is the interior of a finite union of rational polyhedral cones. Using the weight space decomposition (\ref{dsd}) and Equation~(\ref{reduced}), we can describe the set $S$ defined above as \[ S= \left\{ \mathcal{O}(D) \in \Pic(X) \mid \widetilde{H}^{j}(Z(m,kD+(q-j-n-1)A), \text{\bf k} ) =0, \; \forall m\in M,\; \forall j \geq q, \; \forall k \gg 0 \right\} .\] Let $J_i:=\{ \alpha \subseteq I \mid \widetilde{H}^{i}(Z_\alpha, \text{\bf k} ) \neq 0 \}$ be the set indexing orthants which would contribute non-zero terms to the $i$-th cohomology, and let \[O_{J_i} := \bigcup_{ \alpha \in J_i } O_\alpha\] be the union of these orthants. Then using Lemma~\ref{orthlemm} we have \[S = \left\{ \mathcal{O}(D) \in \Pic(X) \mid (kD+(q-i-n-1)A+M) \cap O_{J_i} = \emptyset,\; \forall i \geq q, \; \forall k \gg 0 \right\}. \] Furthermore, using the short exact sequence~(\ref{eqn:seseq}) it is clear that for any divisor $E$ and subset $U \subseteq \text{\bf Z}^I$ we have \[ (E+M) \cap U = \emptyset \text{ if and only if } [E] \notin [U] \] and so \begin{equation}\label{eqn:amplecone} S = \left\{ \mathcal{O}(D) \in \Pic(X) \mid [kD+(q-i-n-1)A] \notin [O_{J_i}] ,\; \forall i \geq q, \; \forall k \gg 0 \right\}. \end{equation} Let $$K:= \bigcup_{i \geq q}[O_{J_i}]$$ which, by Lemma~\ref{orthanttocone}, is a union of (partially open) rational polyhedral cones in $N^1(X)$ and denote by $K^c$ its complement. By Lemma~\ref{complementcones} the set $K^c$ is also a union of (partially open) rational polyhedral cones in $N^1(X)$. We will show that $Amp_q = (K^c)^\circ$, which proves the theorem. First we will show that $Amp_q$ contains the cone $(K^c)^\circ$. Choose any point $[D^\prime]$ in $(K^c)^\circ$. Since this set is open, we can write $[D^\prime]$ in the form $[D + A^\prime]$ where $[D]$ is a rational point in $(K^c)^\circ$ and $A^\prime$ is some ample $\text{\bf R}$-divisor. There exists $\varepsilon>0$ such that the $\varepsilon$-ball $B_\varepsilon([D])$ around $[D]$ is contained in $K^c$. Since $K^c$ is closed under multiplication by a positive scalar, the cone $\Sigma$ generated by $B_\varepsilon([D])$ is also contained in $K^c$, and since $B_\varepsilon([D])$ is convex the cone $\Sigma$ is closed under addition. Let \[k_0> \operatorname{max}_{q \leq i < n} \|[D+(q-i-n-1)A]\|/\varepsilon\] so for every $q \leq i < n$ we have \[[D]+ \frac{1}{k_0}[D+(q-i-n-1)A] \in \Sigma\] and therefore \[(k_0+1)[D]+(q-i-n-1)[A] \in \Sigma.\] Then, since $\Sigma$ is closed under addition, for $k>k_0$ we have \begin{align*}[kD +(q-i-n-1)A] &= (k-(k_0+1))[D] + ((k_0+1)[D]+(q-i-n-1)[A]) \\ &\in \Sigma \subset K^c. \end{align*} In particular, $[kD+(q-i-n-1)A] \notin [O_{J_i}] $ for all $ i \geq q$ and for all $k\gg 0$. Therefore $D$ is $q$-ample and it follows that $(K^c)^\circ \subseteq Amp_q$. Next we will show that $Amp_q$ is contained in the cone $\overline{K^c}$, the closure of the complement of $K$. To see this, suppose for the sake of contradiction that $Amp_q \cap K^\circ \neq \emptyset$. Since $Amp_q$ and $K^\circ$ are both open and the rational points are dense, there is a rational point $[D] \in Amp_q \cap K^\circ$. Choose $\varepsilon_0 > 0$ such that $[D] + \varepsilon [A] \in K^\circ$ for all $\epsilon$ with $|\varepsilon|<\varepsilon_0$. In fact since $K$ is by definition a finite union of convex sets, for all $\varepsilon$ sufficiently close to zero, the point $[D] + \varepsilon [A]$ must lie in a given one of these sets, so there exists $i \geq q$ and $\widetilde{\varepsilon}_0 >0$ such that $[D] + \varepsilon [A] \in [O_{J_i}]$ for all $\varepsilon$ with $0<|\varepsilon|<\widetilde{\varepsilon}_0$. Therefore for all $k\gg 0$, $[D + \frac{(q-i-n-1)}{k} A] \in [O_{J_i}]$ and, multiplying through by $k$, it follows that $[D]\notin Amp_q$. This is a contradiction, so we conclude that $Amp_q \subseteq \overline{K^c}$. Putting the last two paragraphs together, we have \[(K^c)^\circ \subseteq Amp_q \subseteq \overline{K^c}. \] Using Theorem~\ref{thm-open} which says that $Amp_q$ is open in $N^1(X)$, this gives \[\overline{K}^c = (K^c)^\circ \subseteq Amp_q \subseteq \overline{K^c}^\circ = \overline{K^\circ}^c. \] Because $K$ is the union of top dimensional convex cones, every point in the closure is the limit point of a sequence of points in the interior (using the convexity and the fact that each cone has non-empty interior) so $\overline{K} = \overline{K^\circ}.$ Therefore we obtain the following descriptions of $Amp_q$: \begin{equation}\label{eqn:descriptions} Amp_q = \overline{K}^c = (K^c)^\circ = \overline{K^c}^\circ = \overline{K^\circ}^c. \end{equation} The theorem follows since $\overline{K^c}$ is also a union of finitely many rational polyhedral cones by Lemma~\ref{complementcones}. \end{proof} \section{Example} To illustrate the theorem, we will calculate the $q$-ample cone of the blowup of $\text{\bf P}^n$ in a single point, for all $n \geq 2$ and all interesting values of $q$. Let $X$ be the blowup of $\text{\bf P}^n$ in a point, which we may choose to be the intersection of $n$ of the $n+1$ torus-invariant divisors in $\text{\bf P}^n$. Then $X$ has $n+2$ torus-invariant divisors and $\Pic(X)$ has two generators given by $H$, the pullback to $X$ of the hyperplane class on $\text{\bf P}^n$, and $E$, the class of the exceptional divisor. The short exact sequence from diagram~(\ref{eqn:seseq}) is \[ \begin{CD} 0 @>>> M @>{\left( \begin{smallmatrix} {\bf I}_{n \times n} \\ -1 \, \dots \, -1 \\ 1 \; \dots \; \; 1 \end{smallmatrix} \right)}>> \text{\bf Z}^{n+2} @>{\left( \begin{smallmatrix} \begin{smallmatrix} 1 & \dots & 1 \\ -1 & \dots & -1 \end{smallmatrix} & {\bf I}_{2 \times 2} \end{smallmatrix} \right)}>> \text{\bf Z}^2 @>>> 0 \end{CD} \] The polytope $P_X$ is obtained by ``chopping off'' one vertex of the $n$-simplex. We label the top-dimensional faces of this polytope by $F_1, \ldots, F_{n+1}, F_{n+2}$, ordered so that $F_{n+1}$ and $F_{n+2}$ are the two disjoint faces. We need to understand the reduced homology groups of the subspaces $Z(m,D) \subset P_X$ defined in Section \ref{sect-2}. \begin{lemma} Let $Y$ be a union of closed top-dimensional faces of the polytope $P_X$. Then its reduced homology groups are \begin{align*} \tilde{H}_k(Y) = \begin{cases} \text{\bf k} & \text{ if } Y=\partial P_X, \, k=n-1 \\ \text{\bf k} & \text{ if } Y = \partial P_X \setminus \left\{ F_{n+1} \cup F_{n+2} \right\}, \, k=n-2 \\ \text{\bf k} & \text{ if } Y = F_{n+1} \cup F_{n+2}, \, k=0 \\ \text{\bf k} & \text{ if } Y = \emptyset, \, k=-1 \\ 0 & \text{ otherwise. } \end{cases} \end{align*} \end{lemma} \begin{proof} The proof comes from considering the long exact sequence of reduced homology associated to a sequence $A \hookrightarrow B \rightarrow B/A$, where $A$ is a (reasonable) closed subspace of a topological space $B$. Applying this with $B=Y$, a union of top-dimensional faces of $P_X$, and $A=Y \cap F_{n+2}$ we reduce the problem to calculating the reduced homology of either a union of faces of a simplex or the disjoint union of a point with a union of faces of a simplex. Using the fact that the union of any proper subset of faces of a simplex has no reduced homology, the result follows. \end{proof} This lemma immediately identifies the index sets $J_\alpha$ which give nonzero contributions to cohomology of a line bundle, as described in Section \ref{sect-2}. We have \begin{align*} J_0 &= \left\{ \alpha \subseteq I =\left\{1,\ldots,n+2\right\} | \tilde{H}^0(Z_\alpha,\text{\bf k}) \neq 0 \right\} = \left\{ \left\{n+1,n+2 \right\} \right\} \\ J_1 &= \cdots = J_{n-3} = \emptyset \\ J_{n-2} &= \left\{ \alpha \subseteq I | \tilde{H}^{n-2}(Z_\alpha,\text{\bf k}) \neq 0 \right\} = \left\{ \left\{1,\ldots,n \right\} \right\} \\ J_{n-1} &= \left\{ \alpha \subseteq I | \tilde{H}^{n-1}(Z_\alpha,\text{\bf k}) \neq 0 \right\} = \left\{ \left\{1,\ldots,n+2 \right\}\right\}. \\ \end{align*} The corresponding orthants in $\text{\bf Z}^I \cong \text{\bf Z}^{n+2}$ are then \begin{align*} O_{J_0} &= \left\{ \left(d_1,\ldots,d_{n+2} \right) \in \text{\bf Z}^{n+2} \ | \ d_{n+1} <0, d_{n+2} <0, d_i \geq 0 \text{ for all } i=1,\ldots,n \right\} \\ O_{J_1} &= \cdots = O_{J_{n-3}} = \emptyset \\ O_{J_{n-2}} &= \left\{ \left(d_1,\ldots,d_{n+2} \right) \in \text{\bf Z}^{n+2} \ | \ d_{n+1} \geq 0, \, d_{n+2} \geq 0, \, d_i<0 \text{ for all } i=1,\ldots,n \right\} \\ O_{J_{n-1}} &= \left\{ \left(d_1,\ldots,d_{n+2} \right) \in \text{\bf Z}^{n+2} \ | \ d_i<0 \text{ for all } i=1,\ldots,n+2 \right\} \\ \end{align*} The images of these orthants under the map $\text{\bf Z}^I \rightarrow \Pic(X)$ then look as in Figure \ref{Fig1}. In the proof of Theorem \ref{thm-main} we showed that the $q$-ample cone is given by $Amp_q=(K^c)^o$, where $K=\bigcup_{i \geq q}[O_{J_i}]$. So we can compute the $q$-ample cone for each $q$ by forming appropriate unions of the regions $\overline{[O_{J_i}]}$ and taking the complement. The results are shown in Figure \ref{Fig2}. \begin{figure} \centering \begin{tikzpicture} \draw[thin][draw=white][pattern=north east lines][pattern color=gray] (-2,0) -- (0,0) -- (2,-2) -- (-2,-2) --cycle; \draw[<->][very thick] (-2.1,0) -- (0,0) -- (2.05,-2.05); \draw[dotted] (-2,0) -- (2,0) [shift={(0.3,0)}] node{\tiny{$H$}}; \draw[dotted] (0,-2) -- (0,2) [shift={(0,0.3)}] node{\tiny{$E$}}; \draw (-1,-1) node{$\overline{[O_{J_0}]}$}; \draw [shift={(5,0)}][thin][draw=white][pattern=north east lines][pattern color=gray] (-2,2) -- (0,0) -- (2,0) -- (2,2) --cycle; \draw [shift={(5,0)}][<->][very thick] (-2.1,2.1) -- (0,0) -- (2.1,0); \draw [shift={(5,0)}][dotted] (-2,0) -- (2,0) [shift={(0.3,0)}] node{\tiny{$H$}}; \draw [shift={(5,0)}][dotted] (0,-2) -- (0,2) [shift={(0,0.3)}] node{\tiny{$E$}}; \draw [shift={(5,0)}] (1,1) node{$\overline{[O_{J_{n-2}}]}$}; \draw [shift={(2.5,-5)}][thin][draw=white][pattern=north east lines][pattern color=gray] (-2,2) -- (0,0) -- (0,-2) -- (-2,-2) --cycle; \draw [shift={(2.5,-5)}][<->][very thick] (-2.1,2.1) -- (0,0) -- (0,-2.1); \draw [shift={(2.5,-5)}][dotted] (-2,0) -- (2,0) [shift={(0.3,0)}] node{\tiny{$H$}}; \draw [shift={(2.5,-5)}][dotted] (0,-2) -- (0,2) [shift={(0,0.3)}] node{\tiny{$E$}}; \draw [shift={(2.5,-5)}] (-1,-1) node{$\overline{[O_{J_{n-1}}]}$}; \end{tikzpicture} \caption{Orthants of cohomology vanishing} \label{Fig1} \end{figure} \begin{figure} \centering \begin{tikzpicture} \draw[thin][draw=white][pattern=north east lines][pattern color=gray] (2,0) -- (0,0) -- (2,-2) --cycle; \draw[dotted][<->][very thick] (2.1,0) -- (0,0) -- (2.05,-2.05); \draw[dotted] (-2,0) -- (2,0) [shift={(0.3,0)}] node{\tiny{$H$}}; \draw[dotted] (0,-2) -- (0,2) [shift={(0,0.3)}] node{\tiny{$E$}}; \draw (2.5,-1) node{$Amp_0$}; \draw [shift={(5,0)}][thin][draw=white][pattern=north east lines][pattern color=gray] (2,0) -- (0,0) -- (0,-2) --cycle; \draw [shift={(5,0)}][<->][very thick][dotted] (2.1,0) -- (0,0) -- (0,-2.1); \draw [shift={(5,0)}][dotted] (-2,0) -- (2,0) [shift={(0.3,0)}] node{\tiny{$H$}}; \draw [shift={(5,0)}][dotted] (0,-2) -- (0,2) [shift={(0,0.3)}] node{\tiny{$E$}}; \draw [shift={(5,0)}] (2.5,-1.5) node{$Amp_q \, (1 \leq q \leq n-2)$}; \draw [shift={(2.5,-5)}][thin][draw=white][pattern=north east lines][pattern color=gray] (-2,2) -- (0,0) -- (0,-2) -- (2,-2) --(2,2) --cycle; \draw [shift={(2.5,-5)}][<->][very thick][dotted] (-2.1,2.1) -- (0,0) -- (0,-2.1); \draw [shift={(2.5,-5)}][dotted] (-2,0) -- (2,0) [shift={(0.2,0)}] node{\tiny{$H$}}; \draw [shift={(2.5,-5)}][dotted] (0,-2) -- (0,2) [shift={(0,0.3)}] node{\tiny{$E$}}; \draw [shift={(2.5,-5)}] (1,1) node{$Amp_{n-1}$}; \end{tikzpicture} \caption{The $q$-ample cones} \label{Fig2} \end{figure} We remark that Totaro \cite[Theorem 9.1]{Totaro} showed that $Amp_{n-1}$ is the complement of the negative of the closed effective cone, for any projective variety in characteristic zero. For $X$ the blowup of $\text{\bf P}^n$ in a point, it is easy to show that the effective cone is closed, spanned by $E$ and $H-E$, and the complement of the negative of that cone is indeed the cone $Amp_{n-1}$ in Figure \ref{Fig2}. \section{Asymptotic cohomological functions and q-ampleness} Totaro \cite[Section 10]{Totaro} asked whether the $q$-ample cone could be characterised by the vanishing of K\"uronya's asymptotic cohomological functions in degrees above $q$. (For the ample cone, this was proved by de Fernex--K\"uronya--Lazarsfeld \cite{dFKL}.) \begin{theorem} Let $D$ be a Cartier divisor on a projective toric variety. Then $D$ is $q$-ample if and only if $\widehat{h}^i$ vanishes identically in an open neighbourhood of $[D]$ in $N^1(X)$ for all $i>q$. \end{theorem} \begin{proof} First assume $D$ is $q$-ample. K\"uronya \cite{Kur} showed that the functions $\widehat{h}^i$ are homogeneous of degree equal to the dimension of $X$, and continuous in the Euclidean topology on $N^1(X)$, so to prove the claim it suffices to show that $\widehat{h}^i$ vanishes for all Cartier divisors spanning rays which pass sufficiently close to $[D]$. By openness of the $q$-ample cone, there exists an open neighbourhood $U$ of $[D]$ in $N^1(X)$ which also lies in the $q$-ample cone. Any Cartier divisor $D^\prime$ lying on a ray which intersects $U$ is $q$-ample, so applying Definition \ref{def-naive} with $F=\mathcal{O}_X$ we get $H^i(X, \mathcal{O}(nD^\prime))=0$ for $n$ sufficiently large and for $i>q$. Therefore $\widehat{h}^i(D^\prime)=0$, as required. To prove the converse, suppose $D$ is not $q$-ample. Using the description of the $q$-ample cone in the proof of Theorem~\ref{thm-main}, we have $[D]\in \overline{K}$. Therefore given any open neighbourhood of $[D]$ in $N^1(X)$, it intersects the interior of $K$. In fact it follows from the definition of $K$ that there exists some $i \geq q$ and $\alpha \in J_i$ such that the neighbourhood intersects the interior of $[O_{\alpha}]$. It will therefore be sufficient to show that for any Cartier divisor $D'$ with $[D^\prime] \in [O_{\alpha}]^\circ$ the function $\widehat{h}^{i+1}(D^\prime) \neq 0$. We choose a representative divisor $D^\prime$ of the class $[D^\prime]$ which lies in the interior of the orthant $O_{\alpha}$. If $U$ denotes the closed unit hypercube in $M$ we fix a large enough integer $k$ such that $ D^\prime + \frac{1}{k}U \subset O_{\alpha}$. It follows that for each positive integer $j$ we have $j k D^\prime + j U \subset O_{\alpha}$ and so for each $m \in jU$ the $m$th piece of the cohomology $H^{i+1}(X, \mathcal{O}(j k D^\prime))_m\neq 0$. Therefore $h^{i+1}(X, \mathcal{O}(j k D^\prime)) \geq j^n$ and it follows that $\widehat{h}^{i+1}(D^\prime) \neq 0$. \end{proof}

https://arxiv.org/abs/1906.10465

Probabilistic Error Analysis for Inner Products

Probabilistic models are proposed for bounding the forward error in the numerically computed inner product (dot product, scalar product) between of two real $n$-vectors. We derive probabilistic perturbation bounds, as well as probabilistic roundoff error bounds for the sequential accumulation of the inner product. These bounds are non-asymptotic, explicit, and make minimal assumptions on perturbations and roundoffs.The perturbations are represented as independent, bounded, zero-mean random variables, and the probabilistic perturbation bound is based on Azuma's inequality. The roundoffs are also represented as bounded, zero-mean random variables. The first probabilistic bound assumes that the roundoffs are independent, while the second one does not. For the latter, we construct a Martingale that mirrors the sequential order of computations.Numerical experiments confirm that our bounds are more informative, often by several orders of magnitude, than traditional deterministic bounds -- even for small vector dimensions~$n$ and very stringent success probabilities. In particular the probabilistic roundoff error bounds are functions of $\sqrt{n}$ rather than~$n$, thus giving a quantitative confirmation of Wilkinson's intuition. The paper concludes with a critical assessment of the probabilistic approach.

\section*{Acknowledgements} We thank Jack Dongarra, Nick Higham, and Clever Moler for helpful discussions. \section{Introduction} Probabilistic approaches towards roundoff analysis have been applied to: matrix inversion by von Neumann \& Goldstine \cite{vNG47} and Tienari \cite{Tie70}; matrix addition and multiplication, and Runge Kutta methods by Hull \& Swenson \cite{HS66}; solution of ordinary differential equations by Henrici \cite{Hen63}; Gaussian elimination by Barlow \& Bareiss \cite{BB80,BB85b,BB85a}; convolution and FFT by Calvetti \cite{Cal91a,Cal91b,Cal92}; solution of eigenvalue problems by Chatelin \& Brunet \cite{BBC88,BC86,CB90}; LU decomposition and linear system solution by Babu\v{s}ka \& S\"{o}derlind \cite{BS18} and Higham and Mary \cite{HM18}. Yet, the futility of probabilistic roundoff error analysis has also been pointed out \cite[page 2]{HS66}, \cite[Page 17]{Kah96}, since roundoffs apparently do not behave like random variables. Nevertheless, we present probabilistic perturbation and roundoff error bounds for the forward error in the numerically computed inner product\footnote{The superscript~$T$ denotes the transpose, and for relative bounds we assume $\vx^T\vy\neq 0$.}, \begin{eqnarray*} \vx^T\vy = x_1 y_1 +\cdots + x_ny_n, \end{eqnarray*} between two real $n$-vectors \begin{eqnarray*} \vx=\begin{pmatrix}x_1 \\\vdots \\ x_n\end{pmatrix}\in\rn \qquad \mathrm{and}\qquad \vy=\begin{pmatrix}y_1\\ \vdots \\ y_n\end{pmatrix}\in\rn. \end{eqnarray*} \paragraph{Contributions} The idea is to represent perturbations and roundoffs as random variables, express the total forward error as a sum of "local" forward errors, and then apply a concentration inequality to the sum. In contrast to some of the previous work, the roundoffs are not required to obey a particular probability distribution. We "motivate" the particular form of each probabilistic bound with a corresponding deterministic bound, and interpret the various random variables in terms of particular forward errors. Our probabilistic approach is most closely related to that of Higham and Mary~\cite{HM18} who derive backward error bounds. In contrast, our forward error bounds lead to new condition numbers (Sections \ref{s_perturb} and~\ref{s_prob1}), and they are tighter because they avoid a union bound for the probabilities. Our bounds are also simple, intuitive, and easy to interpret, with a clear relationship between failure probability and relative error. Compared to \cite[Theorem 3.1]{HM18}, our Corollary~\ref{c_4} is tighter and does not assume independence of roundoffs. \paragraph{Overview} To facilitate the introduction of the probabilistic approach, we start as simple as possible, with probabilistic perturbation bounds (Section~\ref{s_perturb}). The perturbations are represented as independent, bounded, zero-mean random variables; and the forward error is bounded by Azuma's inequaility. This is followed by probabilistic roundoff error bounds for the sequential accumulation of inner products (Section~\ref{s_roundoff1}). The roundoffs are represented as independent, bounded, zero-mean random variables; and the forward error is, again, bounded by Azuma's inequaility. However, numerical experiments (Section~\ref{s_numexp}) illustrate that for non-negative vectors of large dimension, the probabilistic expression stops being an upper bound. By way of an explanation, Henrici ends his 1963 paper \cite[page 11]{Hen63} with: \begin{quote} The crucial hypothesis for the above statistical theories is the hypothesis of independence of local errors. While this assumption seems to yield realistic results in many cases, some situations are known, [...], where local errors definitely cannot be considered to be independent. To elucidate the conditions under which local errors act like independent variables would seem to be a fascinating if difficult problem. \end{quote} As a consequence, and in contrast to \cite{HM18}, we relinquish the independence assumption and derive a general probabilistic roundoff error bound (Section~\ref{s_roundoff2}). The roundoffs are represented as bounded, zero-mean random variables; and the forward error is bounded by an Azuma-Hoeffding Martingale. In particular, we present a quantitative confirmation of Wilkinson's intuition \cite[Section 1.33]{Wilk94book} that the roundoff error in $n$ operations is proportional to $\sqrt{n}\>u$ rather than $n\>u$. The paper ends with a critical analysis of the probabilistic approach, and a long list of future work (Section~\ref{s_conc}). \section{Perturbation bounds}\label{s_perturb} To calibrate the roundoff error bounds and set the stage for the probabilistic approach, we start off with perturbation bounds: first, deterministic bounds that generalize the traditional bound and motivate the probabilistic bound (Section~\ref{s_perturb1}), and then the probabilistic bound (Section~\ref{s_perturb2}). We use the Hadamard product \begin{eqnarray*} \vx\circ\vy \equiv \begin{pmatrix} x_1y_1 & \cdots & x_ny_n\end{pmatrix}^T \end{eqnarray*} to compactly express componentwise relative perturbations as \begin{eqnarray*} \vhx=\begin{pmatrix}(1+\delta_1)\, x_1\\ \vdots \\ (1+\delta_n)\, x_n\end{pmatrix} =\vx+\vdelta\circ \vx,\qquad \vhy=\begin{pmatrix}(1+\theta_1)\,y_1\\ \vdots \\ (1+\theta_n)\,y_n\end{pmatrix} = \vy+\vtheta\circ\vy, \end{eqnarray*} where $|\delta_k|,|\theta_k|\leq u$, $1\leq k\leq n$, for some $u>0$, and the perturbation vectors are \begin{eqnarray*} \vdelta \equiv \begin{pmatrix} \delta_1 & \cdots &\delta_n\end{pmatrix}^T,\qquad \vtheta\equiv \begin{pmatrix} \theta_1 & \cdots & \theta_n\end{pmatrix}^T. \end{eqnarray*} \subsection{Deterministic perturbation bound}\label{s_perturb1} We generalize the traditional perturbation bound to a whole class of bounds, and single out a specific bound to motivate the probabilistic bound in Section~\ref{s_perturb2}. \begin{theorem}\label{t_2} If $\frac{1}{p}+\tfrac{1}{q}=1$, then the relative forward error in the perturbed inner product is bounded by \begin{eqnarray*} \left|\frac{\vhx^T\vhy-\vx^T\vy}{\vx^T\vy}\right| \leq \frac{\|\vx\circ\vy\|_p}{|\vx^T\vy|}\>\|\vdelta+\vtheta+\vdelta\circ\vtheta\|_q. \end{eqnarray*} \end{theorem} \begin{proof} From associativity, distributivity and the fact that all quantities are real follows \begin{eqnarray*} \vhx^T\vhy-\vx^T\vy&=&(\vdelta\circ \vx)^T\vy+\vx^T(\vtheta\circ \vy)+(\vdelta\circ\vx)^T(\vtheta\circ\vy)\\ &=& \sum_{k=1}^n{x_ky_k\>\left(\delta_k+\theta_k+\delta_k\theta_k\right)} = (\vx\circ\vy)^T\left(\vdelta+\vtheta+\vdelta\circ\vtheta\right). \end{eqnarray*} The H\"{o}lder inequality implies \begin{eqnarray*} \left|(\vx\circ\vy)^T\left(\vdelta+\vtheta+\vdelta\circ\vtheta\right)\right| \leq \|\vx\circ\vy\|_p\>\|\vdelta+\vtheta+\vdelta\circ\vtheta\|_q. \end{eqnarray*} $\quad$ \end{proof} Below is a specialization of Theorem~\ref{t_2} to popular $p$-norms. \begin{corollary}\label{c_2} Theorem~\ref{t_2} implies the following bounds. \begin{enumerate} \item Traditional bound $(p=1)$ \begin{eqnarray*} \left|\frac{\vhx^T\vhy-\vx^T\vy}{\vx^T\vy}\right| \leq \frac{\|\vx\circ\vy\|_1}{|\vx^T\vy|}\>\|\vdelta+\vtheta+\vdelta\circ\vtheta\|_{\infty}\leq \frac{|\vx|^T|\vy|}{|\vx^T\vy|}\>u(2+u). \end{eqnarray*} \item Same amplifier as in Theorem~\ref{t_1} $(p=2)$ \begin{eqnarray}\label{e_pb2} \left|\frac{\vhx^T\vhy-\vx^T\vy}{\vx^T\vy}\right| &\leq& \frac{\|\vx\circ\vy\|_2}{|\vx^T\vy|}\>\|\delta+\vtheta+\vdelta\circ\vtheta\|_2\\ &\leq & \sqrt{n}\>\frac{\|\vx\circ\vy\|_2}{|\vx^T\vy|}\>u(2+u).\nonumber \end{eqnarray} \item Smallest amplifier $(p=\infty)$ \begin{eqnarray*} \left|\frac{\vhx^T\vhy-\vx^T\vy}{\vx^T\vy}\right| \leq \frac{\|\vx\circ\vy\|_{\infty}}{|\vx^T\vy|}\>\|\vdelta+\vtheta+\vdelta\circ\vtheta\|_1\leq n\>\frac{\|\vx\circ\vy\|_{\infty}}{|\vx^T\vy|}\> u(2+u). \end{eqnarray*} \end{enumerate} \end{corollary} \begin{proof} The traditional bound follows from \begin{eqnarray*} \|\vx\circ\vy\|_1=\sum_{k=1}^n{|x_ky_k|}=\sum_{k=1}^n{|x_k|\,|y_k|}=|\vx|^T|\vy|. \end{eqnarray*} $\quad$ \end{proof} The numerical experiments in Section~\ref{s_perturbamp} suggest that the three bounds tend to differ by at most an order of magnitude or so, with the traditional bound being the tightest. \subsection{Probabilistic perturbation bound}\label{s_perturb2} We derive a probabilistic bound corresponding to the deterministic bound (\ref{e_pb2}), and then compare the two bounds. The basis for the probabilistic bounds is a concentration inequality, which bounds the deviation of a sum from its mean in terms of the deviations of the individual summands from their means. \begin{lemma}[Azuma's inequality, Theorem 5.3 in \cite{ChungLu2006}]\label{l_1} Let $Z\equiv Z_1+\cdots + Z_n$ be a sum of independent random variables $Z_1, \ldots, Z_n$ with \begin{eqnarray*} |Z_k-\E[Z_k]|\leq c_k, \qquad 1\leq k\leq n. \end{eqnarray*} Then for any $0<\delta<1$, with probability at least $1-\delta$, \begin{eqnarray*} \left|Z - \E[Z]\right|\leq \sqrt{\sum_{k=1}^n{c_k^2}}\>\sqrt{2\ln{(2/\delta)}}. \end{eqnarray*} \end{lemma} \begin{proof} In \cite[Theorem 5.3]{ChungLu2006} set \begin{eqnarray*} \delta\equiv \Pr\left[ |Z-\E[Z]| \geq t\right] \leq 2\exp\left(-\frac{t^2}{2\sum_{k=1}^n{c_k^2}}\right). \end{eqnarray*} and solve for $t$ in terms of $\delta$. If $|Z-\E[Z]| \geq t$ holds with probability at most~$\delta$, then the complementary event $|Z-\E[Z]| \leq t$ holds with probability at least~$1-\delta$. \end{proof} Thus, if each summand $Z_k$ is close to its mean $\E[Z_k]$, then with high probability, the sum $Z$ is also close to its mean $\E[Z]$. In the probabilistic perturbation bound below, the perturbations $\delta_k$ and $\theta_k$ are represented as independent, bounded, zero-mean random variables. \begin{theorem}\label{t_1} Let the perturbations $\delta_k,\theta_k$ be independent random variables with $\E[\delta_k]=\E[\theta_k]=0$ and $|\delta_k|,|\theta_k|\leq u$, $1\leq k\leq n$. Then for any $0<\delta<1$, with probability at least $1-\delta$, the relative forward error in the perturbed inner product is bounded by \begin{eqnarray*} \left|\frac{\vhx^T\vhy-\vx^T\vy}{\vx^T\vy}\right| &\leq& \frac{\|\vx\circ\vy\|_2}{|\vx^T\vy|}\>\sqrt{2\,\ln{(2/\delta)}}\> u(2+u)\\ &=& \frac{\sqrt{\sum_{k=1}^n{|x_ky_k|^2}}}{|\vx^T\vy|}\>\sqrt{2\,\ln{(2/\delta)}}\> u(2+u). \end{eqnarray*} \end{theorem} \begin{proof} Write the total forward error \begin{eqnarray*} Z\equiv \vhx^T\vhy-\vx^T\vy=Z_1 + \cdots + Z_n \end{eqnarray*} as a sum of independent random variables, where each summand represents a "local" forward error, \begin{eqnarray*} Z_k\equiv x_ky_k\, \left((1+\delta_k)(1+\theta_k)-1\right)=x_ky_k\,\left(\delta_k+\theta_k+\delta_k\theta_k\right), \qquad 1\leq k\leq n. \end{eqnarray*} From the linearity of the mean and $\delta_k$, $\theta_k$ being independent random variables with $\E[\delta_k]=\E[\theta_k]=0$ follows \begin{eqnarray*} \E[Z_k]=x_ky_k\>\left(\E[\delta_k] +\E[\theta_k] +\E[\delta_k]\E[\theta_k]\right)= 0, \qquad 1\leq k\leq n. \end{eqnarray*} The boundedness of $\delta_k$ and $\theta_k$ implies that the deviation of $Z_k$ from its mean $\E[Z_k]=0$ equals \begin{eqnarray*}\label{e_H1} \left|Z_k-\E[Z_jk\right| = |Z_k|= |x_ky_k|\,|\delta_k+\theta_k+\delta_k\theta_k|\leq c_k\equiv |x_ky_k|\,\tau, \qquad 1\leq k\leq n, \end{eqnarray*} where $\tau\equiv 2u+u^2=u(2+u)$. Therefore, the conditions of Lemma~\ref{l_1} are satisfied, and we have \begin{eqnarray*} \sum_{k=1}^n{c_k^2}=\sum_{k=1}^n{|x_ky_k|^2}\,\tau^2=\|\vx\circ\vy\|_2^2\,\tau^2. \end{eqnarray*} The linearity of the expected value implies \begin{eqnarray*} \E[\vhx^T\vhy-\vx^T\vy]=\E[Z]=\E[Z_1]+\cdots+\E[Z_n]= 0. \end{eqnarray*} Apply Lemma~\ref{l_1} to conclude that for any $0<\delta<1$, with probability at least $1-\delta$, \begin{eqnarray*} \left|\vhx^T\vhy-\vx^T\vy\right| =\left|Z-\E[Z]\right| \leq \|\vx\circ\vy\|_2\>\sqrt{2\,\ln{(2/\delta)}}\> \tau. \end{eqnarray*} At last divide both sides of the inequality by the constant $|\vx^T\vy|$. \end{proof} \begin{remark}[Comparsion]\label{r_1} The probabilistic bound in Theorem~\ref{t_1} is by a factor of $\sqrt{n}$ tighter than the deterministic bound (\ref{e_pb2}) in Corollary~\ref{c_2}. The probabilistic bound in Theorem~\ref{t_1} holds with probability at least $1-\delta$, \begin{eqnarray*} \left|\frac{\vhx^T\vhy-\vx^T\vy}{\vx^T\vy}\right| \leq \frac{\|\vx\circ\vy\|_2}{|\vx^T\vy|} \>\sqrt{2\,\ln{(2/\delta)}}\> u(2+u), \end{eqnarray*} while the deterministic bound~(\ref{e_pb2}) equals \begin{eqnarray*} \left|\frac{\vhx^T\vhy-\vx^T\vy}{\vx^T\vy}\right| \leq \frac{\|\vx\circ\vy\|_2}{|\vx^T\vy|} \>\sqrt{n}\>u(2+u). \end{eqnarray*} The two bounds differ in the factors $\sqrt{2\,\ln{(2/\delta)}}$ versus $\sqrt{n}$, which implies: \begin{enumerate} \item The deterministic bound depends explicitly on the dimension~$n$, while the probabilistic bound does not. \item The probabilistic bound is tighter than the deterministic bound for $n>2 \ln{(2/\delta)}$. Specifically, with a tiny failure probability of $\delta=10^{-16}$, the probabilistic bound is tighter for $n>76$, and $\sqrt{2\,\ln{(2/\delta)}}\leq 9$. \end{enumerate} The numerical experiments in Section~\ref{s_perturbcomp} illustrate that the probabilistic bound tends to be at least two orders of magnitude tighter than the deterministic bound. \end{remark} \begin{example}\label{ex_1} We illustrate the behaviour of the amplifier \begin{eqnarray*} \kappa_2\equiv \|\vx\circ\vy\|_2/|\vx^T\vy| \end{eqnarray*} in the probabilistic bound in Theorem~\ref{t_1} with three very special cases. \begin{enumerate} \item \textit{No cancellation:}\\ If all $x_ky_k$ have the same sign, then $\kappa_2^2= \tfrac{\sum_{k=1}^n{|x_ky_k|^2}}{\left(\sum_{k=1}^n{|x_ky_k|}\right)^2}\leq 1$, so that \begin{eqnarray*} \left|\frac{\vhx^T\vhy-\vx^T\vy}{\vx^T\vy}\right| \leq \sqrt{2\,\ln{(2/\delta)}}\> u(2+u). \end{eqnarray*} If also $x_ky_k=w\neq 0$ for $1\leq k\leq n$, then $\kappa_2^2=\tfrac{nw^2}{(nw)^2}=\tfrac{1}{n}$, so that $\kappa_2$ decreases with increasing dimension~$n$, \begin{eqnarray*} \left|\frac{\vhx^T\vhy-\vx^T\vy}{\vx^T\vy}\right| \leq \sqrt{\frac{2\,\ln{(2/\delta)}}{n}}\> u(2+u). \end{eqnarray*} \item \textit{Severe cancellation:}\\ If $x_ky_k=(-1)^k w$ for $1\leq k\leq n$, $w\neq 0$, and $n$ is odd, then $\kappa_2^2=\tfrac{nw^2}{w^2}=n$, so that $\kappa_2$ increases with increasing dimension~$n$, \begin{eqnarray*} \left|\frac{\vhx^T\vhy-\vx^T\vy}{\vx^T\vy}\right| \leq \sqrt{n}\>\sqrt{2\,\ln{(2/\delta)}}\,u(2+u). \end{eqnarray*} \end{enumerate} \end{example} \medskip \section{Probabilistic roundoff error bound, assuming independence of roundoff}\label{s_roundoff1} After presenting the model for independent roundoffs (Section~\ref{s_model1}), we derive a motivating deterministic bound (Section~\ref{s_motdet1}), followed by the probabilistic bound (Section~\ref{s_prob1}). \subsection{Roundoff error model}\label{s_model1} We assume that the elements of $\vx$ and $\vy$ are floating point numbers, and can be stored exactly. The inner product is computed via recursive summation \cite[Section 4.1]{Higham2002}, by accumulating partial sums sequentially from left to right, \begin{eqnarray*} z_1=x_1y_1, \qquad z_{k+1} =\sum_{j=1}^{k+1}{x_jy_j}, \qquad 1\leq k\leq n-1. \end{eqnarray*} The roundoff error model in Table~\ref{tab_fpt} corresponds to \cite[(3.1) and (3.2)]{Higham2002}. \begin{table}[tbhp] \caption{Traditional roundoff error model (guard digits, no fused multiply-add)}\label{tab_fpt} \begin{center} \begin{tabular}{|| l | l |}\hline Floating point arithmetic$\qquad\qquad\quad$ & Exact computation \\ \hline\hline $\hz_1 = x_1y_1\,(1+\theta_1)$ & $z_1=x_1y_1$ \\ $\hz_{k+1} = \left(\hz_{k}+x_{k+1}y_{k+1}\,(1+\theta_{k+1})\right)(1+\delta_{k+1})$ & $z_{k+1} = z_{k}+x_{k+1}y_{k+1}$ \\ \hline $\hz_{n}=\fl(\vx^T\vy)$ & $z_{n}=\vx^T\vy$\\ \hline \end{tabular} \end{center} \end{table} For $0<u< 1$ and $k\geq 1$, we use the abbreviation \begin{eqnarray}\label{e_gamma} \gamma_k &\equiv& (1+u)^k-1 =ku+ \mathcal{O}(u^2). \end{eqnarray} If $ku<1$ then \cite[Lemma 3.1]{Higham2002} \begin{eqnarray*} \gamma_k\leq \frac{ku}{1-ku}. \end{eqnarray*} \subsection{A motivating deterministic bound}\label{s_motdet1} First we unravel the expressions for the computed partial sums, and then bound the sums in terms of inputs and the roundoffs. \begin{lemma} The partial sums in Table~\ref{tab_fpt} are equal to \begin{eqnarray*} \hz_1 &= & x_1 y_1\,(1+\theta_1)\\ \hz_k & =& x_1y_1\,(1+\theta_1)\,\prod_{\ell=2}^k{(1+\delta_{\ell})}+ \sum_{j=2}^k{x_jy_j\,(1+\theta_j)\,\prod_{\ell=j}^k{(1+\delta_{\ell})}}, \qquad 2\leq k\leq n. \end{eqnarray*} If $|\delta_k|, |\theta_k|\leq u$, $1\leq k\leq n$, then the partial sums are bounded by \begin{eqnarray*} |\hz_1| & \leq & |x_1y_1|\,(1+u)\\ |\hz_k| & \leq & |x_1y_1|\, (1+u)^k + \sum_{j=2}^k{|x_jy_j|\,(1+u)^{k-j+2}}, \qquad 2\leq k\leq n. \end{eqnarray*} \end{lemma} \begin{lemma}\label{l_fpt2} The total forward error for the computed inner product $\hz_n=\fl(\vx^T\vy)$ in Table~\ref{tab_fpt} is expressed as a sum of "local forward errors", \begin{eqnarray*} \fl(\vx^T\vy)-\vx^T\vy=\hz_n-z_n=Z_1+\cdots + Z_n, \end{eqnarray*} with a local forward error for each summand, \begin{eqnarray*} Z_1 &\equiv & x_1y_1\,\left((1+\theta_1)\,\prod_{\ell=2}^n{(1+\delta_{\ell})}-1\right)\\ Z_k & \equiv & x_ky_k\,\left((1+\theta_k)\,\prod_{\ell=k}^n{(1+\delta_{\ell})}-1\right), \qquad 2\leq k\leq n \end{eqnarray*} If $|\delta_k|, |\theta_k|\leq u$, $1\leq k\leq n$, and $\gamma_k$ as in (\ref{e_gamma}), then \begin{eqnarray*} |Z_1| &\leq & c_1\equiv |x_1y_1|\,\gamma_n\\ |Z_k| &\leq & c_k\equiv |x_ky_k|\, \gamma_{n-k+2}, \qquad 2\leq k\leq n. \end{eqnarray*} \end{lemma} \begin{proof} This is analogous to \cite[Lemma 3.1]{Higham2002}. \end{proof} Now we can bound the total forward error. \begin{theorem} Let the roundoffs satisfy $|\delta_k|, |\theta_k|\leq u$, $1\leq k\leq n$, with $\gamma_k$ as in~(\ref{e_gamma}). Then the forward error of the computed inner product $\hz_n=\fl(\vx^T\vy)$ in Table~\ref{tab_fpt} is bounded by \begin{eqnarray*} \left|\fl(\vx^T\vy)-\vx^T\vy\right| =\left|\hz_{n}-z_{n}\right| \leq \sum_{k=1}^{n}{c_k} =|x_1y_1|\,\gamma_n +\sum_{k=2}^n{|x_ky_k|\,\gamma_{n-k+2}}. \end{eqnarray*} \end{theorem} \begin{proof} Applying the triangle inequality to the total forward error in Lemma~\ref{l_fpt2} gives \begin{eqnarray*} \left|\hz_{n}-z_{n}\right|\leq \sum_{k=1}^n{|Z_k|} \leq \sum_{k=1}^n{c_k}. \end{eqnarray*} $\quad$ \end{proof} The first consequence is the traditional forward error bound \cite[Section 3.1]{Higham2002}. \begin{corollary}[Traditional bound]\label{c_trad} Let the roundoffs satisfy $|\delta_k|, |\theta_k|\leq u$, $1\leq k\leq n$, with $\gamma_k$ as in~(\ref{e_gamma}). Then the relative forward error of the computed inner product $\hz_n=\fl(\vx^T\vy)$ in Table~\ref{tab_fpt} is bounded by \begin{eqnarray*} \left|\frac{\fl(\vx^T\vy)-\vx^T\vy}{|\vx^T\vy|}\right| &\leq& \frac{|\vx|^T|\vy|}{|\vx^T\vy|}\,\gamma_n. \end{eqnarray*} \end{corollary} \begin{proof} Define the vectors \begin{eqnarray*} \vv\equiv \begin{pmatrix} |x_1y_1| & \cdots &|x_ny_n|\end{pmatrix}^T, \qquad \vg \equiv \begin{pmatrix} \gamma_n & \gamma_n & \gamma_{n-1}& \cdots & \gamma_2\end{pmatrix}^T, \end{eqnarray*} and apply the H\"{o}lder inequality to \begin{eqnarray*} \sum_{k=1}^n{c_k} =\vv^T\vg\leq \|\vv\|_1\,\|\vg\|_{\infty}=\sum_{k=1}^n{|x_ky_k|}\>\gamma_n=|\vx|^T|\vy|\>\gamma_n. \end{eqnarray*} \end{proof} The second consequence is the motivation for the probabilistic bound to follow. \begin{corollary}[Deterministic version of Theorem~\ref{t_04}]\label{c_04} Let the roundoffs satisfy $|\delta_k|, |\theta_k|\leq u$, $1\leq k\leq n$, with $\gamma_k$ as in (\ref{e_gamma}). Then the relative forward error of the computed inner product $\hz_n=\fl(\vx^T\vy)$ in Table~\ref{tab_fpt} is bounded by \begin{eqnarray*} \left|\frac{\fl(\vx^T\vy)-\vx^T\vy}{|\vx^T\vy|}\right| &\leq& \frac{\sqrt{\sum_{k=1}^n{c_k^2}}}{|\vx^T\vy|}\>\sqrt{n} \end{eqnarray*} where $c_1\equiv |x_1y_1|\,\gamma_n$, and $c_k\equiv |x_ky_k|\, \gamma_{n-k+2}$, $2\leq k\leq n$. \end{corollary} \begin{proof} Define the non-negative vector $\ \vc\equiv \begin{pmatrix} c_1 & \cdots & c_n\end{pmatrix}^T$ and use the relation between vector norms \begin{eqnarray*} \sum_{k=1}^n{c_k}=\|\vc\|_1\leq \|\vc\|_2\>\sqrt{n} =\sqrt{\sum_{k=1}^n{c_k^2}}\>\sqrt{n}. \end{eqnarray*} $\quad$ \end{proof} \subsection{Probabilistic forward error bound}\label{s_prob1} Since the roundoffs are independent, bounded zero-mean random variables, we can use Azuma's inequality in Lemma~\ref{l_1}. \begin{theorem}\label{t_04} Let the roundoffs $\delta_k, \theta_k$ be independent random variables with $\E[\delta_k]=\E[\theta_k]=0$ and $|\delta_k|, |\theta_k|\leq u$, $1\leq k\leq n$, and let $\gamma_k$ as in (\ref{e_gamma}). Then for any $0<\delta<1$, with probability at least $1-\delta$, the relative forward error in the computed inner product $\hz_n=\fl(\vx^T\vy)$ in Table~\ref{tab_fpt} is bounded by \begin{eqnarray*} \left|\frac{\fl(\vx^T\vy)-\vx^T\vy}{\vx^T\vy}\right|=\left|\frac{\hz_{n}-z_{n}}{z_{n}}\right| \leq \frac{\sqrt{\sum_{k=1}^{n}{c_k^2}}}{|\vx^T\vy|}\>\sqrt{2\,\ln{(2/\delta)}}, \end{eqnarray*} where $c_1\equiv |x_1y_1|\,\gamma_n$, and $c_k\equiv |x_ky_k|\, \gamma_{n-k+2}$, $2\leq k\leq n$. \end{theorem} \begin{proof} Since the roundoffs are independent random variables, so is the total forward error in Lemma~\ref{l_fpt2}, \begin{eqnarray*} Z\equiv Z_1+\cdots + Z_n=\fl(\vx^T\vy)-\vx^T\vy. \end{eqnarray*} The random variables \begin{eqnarray*} Z_1 &\equiv & x_1y_1\,\left((1+\theta_1)\,\prod_{\ell=2}^n{(1+\delta_{\ell})}-1\right)\\ Z_k & \equiv & x_ky_k\,\left((1+\theta_k)\,\prod_{\ell=k}^n{(1+\delta_{\ell})}-1\right), \qquad 2\leq k\leq n, \end{eqnarray*} represent the local forward errors and have zero mean, $\E[Z_k]=0$. By linearity, the total forward error has zero mean as well, \begin{eqnarray*} \E[Z]=\E\left[Z_1+\cdots + Z_n\right]=\E[Z_1]+\cdots +\E[Z_n]=0. \end{eqnarray*} The deviations of the local errors from their means are bounded by \begin{eqnarray*} |Z_k-\E[Z_k]|=|Z_k|\leq c_k, \qquad 1\leq k\leq n, \end{eqnarray*} with $c_k$ as in Lemma~\ref{l_fpt2}. Thus we can apply Lemma~\ref{l_1} to $Z$, and then divide both sides by the constant $|\vx^T\vy|$. \end{proof} \begin{remark}[Comparison]\label{r_3} The probabilistic bound in Theorem~\ref{t_04} tends to be tighter than the corresponding deterministic bound in Corollary~\ref{c_04}. The probabilistic bound in Theorem~\ref{t_04} holds with probability at least $1-\delta$, \begin{eqnarray*} \left|\frac{\vhx^T\vhy-\vx^T\vy}{\vx^T\vy}\right| \leq \frac{\sqrt{\sum_{k=1}^n{c_k^2}}}{|\vx^T\vy|} \>\sqrt{2\,\ln{(2/\delta)}}, \end{eqnarray*} while the deterministic bound in Corollary~\ref{c_04} equals \begin{eqnarray*} \left|\frac{\vhx^T\vhy-\vx^T\vy}{\vx^T\vy}\right| \leq \frac{\sqrt{\sum_{k=1}^n{c_k^2}}}{|\vx^T\vy|} \>\sqrt{n}, \end{eqnarray*} where $c_1\equiv |x_1y_1|\,\gamma_n$, and $c_k\equiv |x_ky_k|\, \gamma_{n-k+2}$, $2\leq k\leq n$, with $\gamma_k$ as in (\ref{e_gamma}). As in Remark~\ref{r_1}, the two bounds differ in the factors $\sqrt{2\,\ln{(2/\delta)}}$ versus $\sqrt{n}$, which implies: \begin{enumerate} \item The deterministic bound depends explicitly on the dimension~$n$, while the probabilistic bound does not. \item The probabilistic bound is tighter than the deterministic bound for $n>2 \ln{(2/\delta)}$. Specifically, with a tiny failure probability of $\delta=10^{-16}$, the probabilistic bound is tighter for $n>76$, and $\sqrt{2\,\ln{(2/\delta)}}\leq 9$. \end{enumerate} The numerical experiments in Section~\ref{s_robi1} illustrate that the probabilistic expression can be as much as two orders of magnitude tighter then the deterministic bound, but stops being an upper bound for non-negative vectors of large dimension. \end{remark} \section{General probabilistic roundoff error bound}\label{s_roundoff2} In contrast to the previous section, we make no assumptions on the independence of roundoffs. After presenting the roundoff error model (Section~\ref{s_remodel2}), we derive a motivating deterministic bound (Section~\ref{s_motdet2}), and then present the probabilistic bound (Section~\ref{s_prob2}), followed by two upper bounds that take a simpler form (Section~\ref{s_uppersimpler}). \subsection{Roundoff error model}\label{s_remodel2} As in Section~\ref{s_model1}, we assume that the elements of $\vx$ and $\vy$ are floating point numbers, and can be stored exactly. Our model in Table~\ref{tab_fp} differs from the traditional model in Table~\ref{tab_fpt} only in the book keeping. It distinguishes each step that introduces a roundoff, and explicitly separates additions~($+$) from multiplications~($*$). There are $n$ multiplications and $n-1$ additions, so $2n-1$ distinct roundoffs. The model in Table~\ref{tab_fp} is designed to do without additional intermediate factors like $x_ky_k(1+\delta_{2k-2})$, and is expressed solely in terms of partial sums. Since we assume a guard digit model without fused multiply-add, the roundoff for addition can be recorded in a subsequent step. The very first partial sum incurs no addition, so we allocate the roundoff to the second partial sum for easier indexing. \begin{table}[tbhp] \caption{Our roundoff error model (guard digits, no fused multiply-add)}\label{tab_fp} \begin{center} \begin{tabular}{|c || l | l |}\hline $\quad$Operation$\quad$ & Floating point arithmetic$\qquad\qquad\quad$ & Exact computation \\ \hline\hline $*$ & $\hs_1 = x_1y_1$ & $s_1=x_1y_1$ \\ & $\hs_2 = \hs_1\,(1+\delta_1 )$ & $s_2=s_1$ \\ $*$ & $\hs_{2k-1} = \hs_{2k-2}+x_{k}y_{k}\,(1+\delta_{2k-2})$ & $s_{2k-1} = s_{2k-2}+x_ky_k$ \\ $+$ & $\hs_{2k} = \hs_{2k-1}\,(1+\delta_{2k-1})$ & $s_{2k} =s _{2k-1}$ \\ \hline Output & $\hs_{2n}=\fl(\vx^T\vy)$ & $s_{2n}=\vx^T\vy$\\ \hline \end{tabular} \end{center} \end{table} \subsection{A motivating deterministic bound}\label{s_motdet2} First we bound the computed partial sums in terms of the inputs, and the unit roundoff~$u$. \begin{lemma}\label{l_shat} Let the roundoffs satisfy $|\delta_k|\leq u$, $1\leq k\leq 2n-1$. Then the partial sums computed in Table~\ref{tab_fp} are bounded by \begin{eqnarray*} |\hs_{2k-1}|&\leq & |x_1y_1|\>(1+u)^{k-1} +|x_2y_2|\>(1+u)^{k-1}+\cdots +|x_{k}y_{k}|\> (1+u)\\ &=& |x_1y_1|\>(1+u)^{k-1}+\sum_{j=2}^k{|x_jy_j|\>(1+u)^{k-j+1}}, \qquad 1\leq k\leq n, \end{eqnarray*} and \begin{eqnarray*} |\hs_{2k}| &\leq & |x_1y_1|\,(1+u)^k + |x_2y_2|\>(1+u)^k+\cdots + |x_ky_k|\,(1+u)^2\\ &=& |x_1y_1|\>(1+u)^k+ \sum_{j=2}^{k}{|x_jy_j|\>(1+u)^{k-j+2}}, \qquad 1\leq k\leq n. \end{eqnarray*} \end{lemma} \begin{proof} The proof is by induction, starting with the basis for $k=1$, \begin{eqnarray*} |\hs_1|&=& |x_1y_1|=|x_1y_1|\>(1+u)^0,\\ |\hs_2| &=& |\hs_1\>(1+\delta_1)| \leq |x_1y_1|\>(1+u). \end{eqnarray*} Assuming, as the hypothesis, that the statement of the lemma is correct, the induction step gives for $1\leq k\leq n-1$, \begin{eqnarray*} |\hs_{2k+1}| &=& |\hs_{2k} +x_{k+1}y_{k+1}\>(1+\delta_{2k})|\leq |\hs_{2k}| +|x_{k+1}y_{k+1}|\>(1+u)\\ &\leq & |x_1y_1|\,(1+u)^{k} +\sum_{j=2}^k{|x_jy_j|\>(1+u)^{k-j+2}} +|x_{k+1}y_{k+1}|\> (1+u)\\ &=& |x_1y_1|\,(1+u)^{k} +\sum_{j=2}^k{|x_jy_j|\>(1+u)^{(k+1)-j+1}} +|x_{k+1}y_{k+1}|\> (1+u)\\ &=& |x_1y_1|\,(1+u)^{k} +\sum_{j=2}^{k+1}{|x_jy_j|\>(1+u)^{(k+1)-j+1}}, \end{eqnarray*} and for $1\leq k\leq n-1$, \begin{eqnarray*} |\hs_{2k+2}|&=& |\hs_{2k+1}\,(1+\delta_{2k+1})|\leq |\hs_{2k+1}|\,(1+u)\\ &=& |x_1y_1|\,(1+u)^{k+1} +\sum_{j=2}^{k+1}{|x_jy_j|\>(1+u)^{(k+1)-j+2}}. \end{eqnarray*} $\quad$ \end{proof} The \textit{total} forward error is \begin{eqnarray}\label{e_tfe2} Z_{2n} \equiv \hs_{2n}-s_{2n} =\fl(\vx^T\vy)-\vx^T\vy, \end{eqnarray} while the \textit{partial sum forward errors} are \begin{eqnarray*} Z_k \equiv \hs_k-s_k, \qquad 1\leq k\leq 2n, \end{eqnarray*} where $Z_1= 0$. We use these partial sum errors to distinguish the newly arrived roundoff from the previous roundoffs. Then we establish a recursion for the partial sum errors $Z_k$, and bound the difference between two successive partial sum errors $Z_k$ and $Z_{k-1}$ by the "incremental error" $c_k\>u$. This incremental error $c_k\>u$ captures the most recent roundoff introduced when moving from $Z_{k-1}$ to $Z_k$. \begin{lemma}\label{l_lfe} The forward errors for the partial sums in Table~\ref{tab_fp} satisfy the recursions \begin{eqnarray*} Z_{2k} &=& Z_{2k-1} +\hs_{2k-1}\> \delta_{2k-1}, \qquad 1\leq k\leq n,\\ Z_{2k-1} &=& Z_{2k-2} +x_ky_k\> \delta_{2k-2}, \qquad 2\leq k\leq n. \end{eqnarray*} If $|\delta_k|\leq u$, $1\leq k\leq 2n-1$, then \begin{eqnarray*} |Z_{2k}-Z_{2k-1}| \leq c_{2k-1}\>u, \qquad 1\leq k\leq n, \end{eqnarray*} where \begin{eqnarray*} |\hs_{2k-1}|\leq c_{2k-1}\equiv |x_1y_1|\>(1+u)^{k-1}+\sum_{j=2}^{k}{|x_jy_j|\>(1+u)^{k-j+1}}, \end{eqnarray*} and for $2\leq k\leq n$, \begin{eqnarray*} |Z_{2k-1}-Z_{2k-2}| \leq c_{2k-2}\>u, \qquad \text{where}\quad c_{2k-2} \equiv |x_{k}y_{k}|. \end{eqnarray*} \end{lemma} \begin{proof} The proof is by induction, following the recursions in Table~\ref{tab_fp}. Since $Z_1=0$, the induction starts one step later than the one in Lemma~\ref{l_lfe}, and the induction basis is \begin{eqnarray*} Z_2 & =& \hs_2 -s_2 = \hs_1(1+\delta_1) -s_1 = Z_1 + \hs_1\>\delta_1,\\ Z_3 &=& \hs_3-s_3 = \hs_2+x_2y_2\>(1+\delta_2)-\left(s_2+x_2y_2\right) =Z_2+x_2y_2\>\delta_2. \end{eqnarray*} Assuming, as the hypothesis, that the statement of the lemma is correct, the induction step gives for $1\leq k\leq n-1$, \begin{eqnarray*} Z_{2k+2}&=&\hs_{2k+2} -s_{2k+2} = \hs_{2k+1}\>(1+\delta_{2k+1}) -s_{2k+1} = Z_{2k+1} + \hs_{2k+1}\>\delta_{2k+1}, \end{eqnarray*} and for $2\leq k\leq n-1$, \begin{eqnarray*} Z_{2k+1} &= &\hs_{2k+1}-s_{2k+1} = \hs_{2k}+x_{k+1}y_{k+1}\>(1+\delta_{2k}) -(s_{2k}+x_{k+1}y_{k+1})\\ & = & Z_{2k} + x_{k+1}y_{k+1}\>\delta_{2k}. \end{eqnarray*} Lemma~\ref{l_shat} and the above recursions imply the bounds \begin{eqnarray*} |Z_2-Z_1|& =& |\hs_1\>\delta_1|\leq |\hs_1|\>u\leq c_1\>u \qquad \text{where} \quad c_1=|x_1y_1|,\\ |Z_3-Z_2|& =& |x_2y_2\>\delta_2| \leq |x_2y_2|\>u \qquad \text{where}\quad c_2=|x_2y_2|. \end{eqnarray*} In general, \begin{eqnarray*} |Z_{2k}-Z_{2k-1}| &=& |\hs_{2k-1}\>\delta_{2k-1}| \leq |\hs_{2k-1}|\> u\leq c_{2k-1}\> u, \qquad 2\leq k\leq n, \end{eqnarray*} where $c_{2k-1}=|x_1y_1|\>(1+u)^{k-1}+\sum_{j=2}^{k}{|x_jy_j|\>(1+u)^{k-j+1}}$, and \begin{eqnarray*} |Z_{2k+1}-Z_{2k}| &=& |x_{k+1}y_{k+1}\>\delta_{2k}|\leq |x_{k+1}y_{k+1}|\>u\leq c_{2k}\>u, \qquad 2\leq k\leq n-1, \end{eqnarray*} where $c_{2k}=|x_{k+1}y_{k+1}|$. \end{proof} \begin{theorem}[Deterministic version of Theorem~\ref{t_4}]\label{t_det4} Let the roundoffs satisfy $|\delta_k|\leq u$, $1\leq k\leq 2n-1$. Then the relative forward error of the computed inner product $\hs_{2n}=\fl(\vx^T\vy)$ in Table~\ref{tab_fp} is bounded by \begin{eqnarray*} \left|\frac{\fl(\vx^T\vy)-\vx^T\vy}{\vx^T\vy}\right| =\left|\frac{\hs_{2n}-s_{2n}}{s_{2n}}\right| \leq \sqrt{2n-1}\>\frac{\sqrt{\sum_{k=1}^{2n-1}{c_k^2}}}{|\vx^T\vy|}\> u, \end{eqnarray*} where \begin{eqnarray*} c_{2k-1} &=& |x_1y_1|\>(1+u)^{k-1}+\sum_{j=2}^{k}{|x_jy_j|(1+u)^{k-j+1}}, \qquad 1\leq k\leq n\\ c_{2k-2} &= &|x_{k}y_{k}| ,\qquad 2\leq k\leq n. \end{eqnarray*} \end{theorem} \begin{proof} Represent the total error (\ref{e_tfe2}) as a telescoping sum of incremental errors \begin{eqnarray*} \fl(\vx^T\vy)-\vx^T\vy=Z_{2n}=(Z_{2n}-Z_{2n-1})+(Z_{2n-1}-Z_{2n-2})+ \cdots +(Z_2-Z_1), \end{eqnarray*} where $Z_1=0$. With the expressions for $c_k$ from Lemma~\ref{l_lfe}, \begin{eqnarray*} |Z_{2n}| &\leq & \underbrace{|Z_{2n}-Z_{2n-1}|}_{\leq \,c_{2n-1} \,u} + \underbrace{|Z_{2n-1}-Z_{2n-2}|}_{\leq \,c_{2n-2}\, u}+ \cdots + \underbrace{|Z_2-Z_1|}_{\leq\, c_1\, u} \leq \sum_{k=1}^{2n-1}{c_k}\, u. \end{eqnarray*} As in the proof of Corollary~\ref{c_04}, the relation between the vector one- and two-norms implies \begin{eqnarray*} \sum_{k=1}^{2n-1}{c_k} \leq \sqrt{2n-1} \> \sqrt{\sum_{k=1}^{2n-1}{c_k^2}}. \end{eqnarray*} \end{proof} \subsection{Probabilistic forward error bound}\label{s_prob2} We derive a probabilistic bound based on an Azuma Martingale, which does not require independence of roundoffs, and then compare the probabilistic and deterministic bounds. \begin{definition}[Martingale, Definition 12.1 in \cite{MitzUp2005}]\label{d_4} A sequence of random variables $Z_1, Z_2\ldots$ is a Martingale with respect to a sequence $\delta_1, \delta_2, \ldots $ if for $k\geq 1$ \begin{enumerate} \item $Z_k$ is a function of $\delta_1, \ldots, \delta_{k-1}$, \item $\E[|Z_k|]<\infty$, \item $\E\left[Z_{k+1} |\delta_1, \ldots, \delta_{k-1}\right] = Z_k$. \end{enumerate} \end{definition} \medskip The version of the Martingale below is tailored to our context. \begin{lemma}[Azuma-Hoeffding Martingale, Theorem 12.4 in \cite{MitzUp2005}]\label{l_4} Let $Z_1, \ldots, Z_{2n}$ be a Martingale with \begin{eqnarray*} |Z_{k}-Z_{k-1}|\leq c_{k-1}, \qquad 2\leq k\leq 2n. \end{eqnarray*} Then for any $0<\delta<1$ with probability at least $1-\delta$, \begin{eqnarray*} \left|Z_{2n}-Z_1\right|\leq \sqrt{\sum_{k=1}^{2n-1}{c_k^2}}\>\sqrt{2\ln{(2/\delta)}}. \end{eqnarray*} \end{lemma} \begin{proof} In \cite[Theorem 12.4]{MitzUp2005}, set \begin{eqnarray*} \delta \equiv \Pr\left[ |Z_m-Z_0| \geq t\right] \leq 2\exp\left(-\frac{t^2}{2\sum_{k=1}^m{c_k^2}}\right). \end{eqnarray*} and $m=2n-1$, and then solve for $t$ in terms of $\delta$. If $|Z-\E[Z]| \geq t$ holds with probability at most~$\delta$, then the complementary event $|Z-\E[Z]| \leq t$ holds with probability at least~$1-\delta$. \end{proof} Again, the roundoffs are represented as bounded, zero-mean random variables, but now they are not required to be independent. The following bound resembles the one in Theorem~\ref{t_2}, but contains more summands. \begin{theorem}\label{t_4} Let the roundoffs $\delta_k$ be random variables with $\E[\delta_k]=0$ and $|\delta_k|\leq u$, $1\leq k\leq 2n-1$. Then for any $0<\delta<1$, with probability at least $1-\delta$, the relative forward error of the computed inner product $\hs_{2n}=\fl(\vx^T\vy)$ in Table~\ref{tab_fp} is bounded by \begin{eqnarray*} \left|\frac{\fl(\vx^T\vy)-\vx^T\vy}{\vx^T\vy}\right|=\left|\frac{\hs_{2n}-s_{2n}}{s_{2n}}\right| \leq \frac{\sqrt{\sum_{k=1}^{2n-1}{c_k^2}}}{|\vx^T\vy|}\>\sqrt{2\,\ln{(2/\delta)}} \> u, \end{eqnarray*} where \begin{eqnarray*} c_{2k-1} &= & |x_1y_1|\>(1+u)^{k-1}+\sum_{j=2}^{k}{|x_jy_j|\,(1+u)^{k-j+1}}, \qquad 1\leq k\leq n,\\ c_{2k-2} & = & |x_ky_k|, \qquad 2\leq k\leq n. \end{eqnarray*} \end{theorem} \begin{proof} Since $Z_1=0$, Table~\ref{tab_fp} implies for the total forward error (\ref{e_tfe2}) that \begin{eqnarray*} |\fl(\vx^T\vy)-\vx^T\vy|=|\hs_{2n}-s_{2n}|=|Z_{2n}|=|Z_{2n}-Z_1|. \end{eqnarray*} To apply Lemma~\ref{l_4}, we show that the partial sum forward errors $Z_1, Z_2\ldots, Z_{2n}$ form a Martingale with respect to the roundoffs $\delta_1,\ldots,\delta_{2n-1}$. To this end, we need to check the conditions in Definition~\ref{d_4} and Lemma~\ref{l_4}. \begin{enumerate} \item The recursions in Lemma~\ref{l_lfe} show that $Z_k$ is a function of the roundoffs $\delta_1,\ldots, \delta_{k-1}$, $2\leq k\leq 2n$. \item The expectation of $|Z_k|$ is finite because $|Z_k|$ is a finite sum of bounded summands, and the roundoffs have zero mean. \item Lemma~\ref{l_lfe} implies $Z_2=Z_1+x_1y_1\>\delta_1$ where $Z_1=0$. The linearity of expectation and zero-mean property of the roundoffs implies \begin{eqnarray*} \E[Z_2]=\E\left[Z_1+x_1y_1\>\delta_1\right]=Z_1+ x_1y_1\>\E[\delta_1]=Z_1. \end{eqnarray*} More generally, item~1 implies that $Z_{2k-2}$ depends on $\delta_1,\ldots, \delta_{2k-3}$, $2\leq k\leq n$. Conditioning on all of these roundoffs removes the randomness and produces a fixed value, \begin{eqnarray*} \E[Z_{2k-2}\, | \delta_1,\ldots,\delta_{2k-3}]=Z_{2k-2}, \qquad 2\leq k\leq n, \end{eqnarray*} Combine the above with the zero-mean property of the roundoffs \begin{eqnarray*} \E[\delta_{2k-2}\,|\delta_1,\ldots,\delta_{2k-3}]=\E[\delta_{2k-2}]=0 \end{eqnarray*} and Lemma~\ref{l_lfe} to conclude \begin{eqnarray*} \E[Z_{2k-1}\,|\delta_1,\ldots, \delta_{2k-3}]&=&\E[Z_{2k-2}+x_ky_k\>\delta_{2k-2}\,|\delta_1,\ldots, \delta_{2k-3}] \\ &=& Z_{2k-2}+x_ky_k \E[\delta_{2k-2}]=Z_{2k-2}, \qquad 2\leq k\leq n. \end{eqnarray*} Now consider the remaining recursions $Z_{2k}=Z_{2k-1} +\hs_{2k-1}\>\delta_{2k-1}$, $1\leq k\leq n$. Item~1 and Table~\ref{tab_fp} show that $Z_{2k-1}$ and $\hs_{2k-1}$ depend only on the roundoffs $\delta_1,\ldots,\delta_{2k-2}$. Conditioning $Z_{2k-1}$ and $\hs_{2k-1}$ on all of these roundoffs removes the randomness and produces fixed values, \begin{eqnarray*} \E[Z_{2k-1}\, | \delta_1,\ldots,\delta_{2k-2}] &=& Z_{2k-1}, \qquad 1\leq k\leq n,\\ \E[\hs_{2k-1}\, | \delta_1,\ldots,\delta_{2k-2}]&=& \hs_{2k-1}, \qquad 1\leq k\leq n. \end{eqnarray*} Arguing as above shows \begin{eqnarray*} \E[Z_{2k}\,|\delta_1,\ldots, \delta_{2k-2}]&=&\E[Z_{2k-1}+\hs_{2k-1}\delta_{2k-1}\,|\delta_1,\ldots, \delta_{2k-2}] \\ &=&Z_{2k-1}+\hs_{2k-1} \E[\delta_{2k-1}]=Z_{2k-1}, \qquad 1\leq k\leq n. \end{eqnarray*} Thus, $Z_1, Z_2, \ldots, Z_{2n}$ form a Martingale with respect to $\delta_1,\ldots, \delta_{2n-1}$. \item Lemma~\ref{l_lfe} implies \begin{eqnarray*} |Z_{2k}-Z_{2k-1}| \leq c_{2k-1}\>u, \qquad 1\leq k\leq n, \end{eqnarray*} where \begin{eqnarray*} |\hs_{2k-1}|\leq c_{2k-1}\equiv |x_1y_1|\>(1+u)^{k-1}+\sum_{j=2}^{k}{|x_jy_j|\>(1+u)^{k-j+1}}, \end{eqnarray*} and for $2\leq k\leq n$, \begin{eqnarray*} |Z_{2k-1}-Z_{2k-2}| \leq c_{2k-2}\>u, \qquad \text{where}\quad c_{2k-2} \equiv |x_{k}y_{k}|. \end{eqnarray*} \end{enumerate} Thus, the conditions for Lemma~\ref{l_lfe} are satisfied, and we can use it to bound $|Z_{2n}-Z_1|$ with the above $c_k$ from Lemma~\ref{l_lfe}. \end{proof} \begin{remark}[Comparison]\label{r_4} The probabilistic bound in Theorem~\ref{t_4} tends to be tighter than the deterministic bound in Theorem~\ref{t_det4}. The probabilistic bound in Theorem~\ref{t_4} holds with probability at least $1-\delta$, \begin{eqnarray*} \left|\frac{\fl(\vx^T\vy)-\vx^T\vy}{\vx^T\vy}\right|=\left|\frac{\hs_{2n}-s_{2n}}{s_{2n}}\right| \leq \frac{\sqrt{\sum_{k=1}^{2n-1}{c_k^2}}}{|\vx^T\vy|}\>\sqrt{2\,\ln{(2/\delta)}} \> u, \end{eqnarray*} while the deterministic bound in Theorem~\ref{t_det4} is \begin{eqnarray*} \left|\frac{\fl(\vx^T\vy)-\vx^T\vy}{\vx^T\vy}\right| =\left|\frac{\hs_{2n}-s_{2n}}{s_{2n}}\right| \leq \frac{\sqrt{\sum_{k=1}^{2n-1}{c_k^2}}}{|\vx^T\vy|}\> \sqrt{2n-1}\>u, \end{eqnarray*} where $c_{2k-1}=|x_{k}y_{k}|$ and $c_{2k} = \sum_{j=1}^{k}{|x_jy_j|(1+u)^{k-j+1}}$, $1\leq k\leq n$. The two bounds differ in the factors $\sqrt{2\,\ln{(2/\delta)}}$ versus $\sqrt{2n-1}$, which implies: \begin{enumerate} \item The deterministic bound increases with the dimension~$n$, while the probabilistic bound does not. \item The probabilistic bound is tighter for $n>\ln{(2/\delta)}+\tfrac{1}{2}$. Specifically, with a tiny failure probability of $\delta=10^{-16}$, the probabilistic bound is tighter for $n\geq 39$, and $\sqrt{2\,\ln{(2/\delta)}}\leq 9$. \end{enumerate} \end{remark} \subsection{Simpler forward error bounds}\label{s_uppersimpler} We derive two upper bounds for Theorem \ref{t_det4} and~\ref{t_4} that have a simpler form, and then conform Wilkinson's intuition \cite[Section 1.33]{Wilk94book}. The first bound is more compact than Theorem~\ref{t_4}, and makes use of abbreviations for the leading subvectors of $|\vx|\circ |\vy|$, and vectors containing powers of $1+u$. \begin{corollary}[Compact upper bound]\label{c_4a} Define the $k$-vectors \begin{eqnarray*} (\vx\circ\vy)_k \equiv \begin{pmatrix} |x_1y_1| \\ |x_2y_2|\\\vdots \\ |x_ky_k|\end{pmatrix}, \qquad \vu_k\equiv \begin{pmatrix}(1+u)^{k-1} \\ (1+u)^{k-1}\\ \vdots \\ 1+u\end{pmatrix}, \qquad 2\leq k\leq n. \end{eqnarray*} If $\tfrac{1}{p}+\tfrac{1}{q}=1$, then in Theorems \ref{t_det4} and~\ref{t_4} we have \begin{eqnarray*} \sum_{k=1}^{2n-1}{c_k^2}\leq \|\vx\circ \vy\|_2^2+\sum_{k=2}^{n}{\|(\vx\circ\vy)_k\|_p^2\, \|\vu_k\|_q^2}. \end{eqnarray*} \end{corollary} \begin{proof} Partition \begin{eqnarray*} \sum_{k=1}^{2n-1}{c_k^2} & = & \sum_{k=1}^n{c_{2k-1}^2}+\sum_{k=2}^{n}{c_{2k-2}^2} = \sum_{k=2}^n{c_{2k-1}^2}+ c_1^2+\sum_{k=2}^n{c_{2k-2}^2}. \end{eqnarray*} From $c_1=|x_1y_1|$ and $c_{2k-2}=|x_ky_k|$, $2\leq k\leq n$, follows \begin{eqnarray*} c_1^2+\sum_{k=2}^n{c_{2k-2}^2} = \sum_{k=1}^n{|x_ky_k|^2}=\|\vx\circ\vy\|_2^2. \end{eqnarray*} Thus $\sum_{k=1}^{2n-1}{c_k^2} = \|\vx\circ \vy\|_2^2+\sum_{k=2}^n{c_{2k-1}^2}$. In the remaining sum, apply H\"{o}lder's inequality to each summand, \begin{eqnarray*} c_{2k-1}&=& |x_1y_1|\>(1+u)^{k-1}+\sum_{j=2}^{k}{|x_jy_j|\,(1+u)^{k-j+1}}\\ &=&(\vx\circ\vy)_{k}^T\vu_{k} \leq \|(\vx \circ \vy)_{k}\|_p\,\|\vu_{k}\|_q, \qquad 2\leq k\leq n. \end{eqnarray*} $\quad$ \end{proof} The second bound, below, takes a much simpler form. \begin{corollary}[Simplest upper bound for Theorem~\ref{t_4}]\label{c_4} Let the roundoffs $\delta_k$ be random variables with $\E[\delta_k]=0$ and $|\delta_k|\leq u$, $1\leq k\leq 2n$; and let $\gamma_k$ as in (\ref{e_gamma}). Then for any $0<\delta<1$, with probability at least $1-\delta$, the relative forward error of the computed inner product $\hs_{2n}=\fl(\vx^T\vy)$ in Table~\ref{tab_fp} is bounded by \begin{eqnarray}\label{e_re1} \left|\frac{\fl(\vx^T\vy)-\vx^T\vy}{\vx^T\vy}\right| \leq \frac{|\vx|^T|\vy|}{|\vx^T\vy|}\>\sqrt{2\ln{(2/\delta)}} \>\sqrt{\frac{u\>\gamma_{2n}}{2}}. \end{eqnarray} \end{corollary} \begin{proof} In Corollary~\ref{c_4a}, choose $p=1$ and $q=\infty$, so that \begin{eqnarray*} \|(\vx\circ\vy)_k\|_1\|\vu_k\|_{\infty} \leq \|\vx\circ\vy\|_1 (1+u)^{k-1}, \qquad 2\leq k\leq n. \end{eqnarray*} The relation between vector norms implies $\|\vx\circ \vy\|_2\leq \|\vx\circ\vy\|_1$. Insert the preceding two inequalities into Corollary~\ref{c_4a}, \begin{eqnarray*} \sum_{k=1}^{2n-1}{c_k^2}\leq \|\vx\circ\vy\|_2^2+\sum_{k=2}^{n}{\|(\vx\circ\vy)_k\|_1^2\,\|\vu_k\|_{\infty}^2} \leq \|\vx\circ\vy\|_1^2\left(1+\sum_{k=2}^{n}{(1+u)^{2(k-1)}}\right). \end{eqnarray*} The second factor is a geometric sum, \begin{eqnarray*} 1+\sum_{k=1}^{n-1}{(1+u)^{2k}}=\sum_{k=0}^{n-1}{(1+u)^{2k}}=\frac{(1+u)^{2n}-1}{(1+u)^2-1}=\frac{\gamma_{2n}}{u^2+2u}. \end{eqnarray*} Combining the preceding inequalities gives \begin{eqnarray*} \sqrt{\sum_{k=1}^{2n-1}{c_k^2}}\leq \|\vx\circ\vy\|_1\,\sqrt{\frac{\gamma_{2n}}{u^2+2u}} \leq \|\vx\circ\vy\|_1\,\sqrt{\frac{\gamma_{2n}\,u}{2}}. \end{eqnarray*} At last substitute this into Theorem~\ref{t_4}. \end{proof} \begin{remark}[Comparison with traditional bound]\label{r_2} We quantify and confirm Wilkinson's intuition \cite[Section 1.33]{Wilk94book}, by illustrating that the probabilistic bounds in Theorem~\ref{t_4}, and Corollaries \ref{c_4a} and~\ref{c_4} are proportional to $\sqrt{n}\>u$, while the traditional bound in Corollary~\ref{c_trad} is proportional to $n\> u$. Let $\gamma_k=(1+u)^k-1$, $k\geq 1$, be as in (\ref{e_gamma}). The probabilistic bound in Corollary~\ref{c_4} holds with probability at least $1-\delta$, \begin{eqnarray*} \left|\frac{\fl(\vx^T\vy)-\vx^T\vy}{\vx^T\vy}\right| \leq \frac{|\vx|^T|\vy|}{|\vx^T\vy|}\>\sqrt{2\ln{(2/\delta)}} \>\sqrt{\frac{u\>\gamma_{2n}}{2}}, \end{eqnarray*} while the deterministic bound in Corollary~\ref{c_trad} equals \begin{eqnarray*} \left|\frac{\fl(\vx^T\vy)-\vx^T\vy}{|\vx^T\vy|}\right| &\leq& \frac{|\vx|^T|\vy|}{|\vx^T\vy|}\,\gamma_n. \end{eqnarray*} For large $n$, the bounds behave asymptotically like their first order terms, \begin{eqnarray*} \gamma_n\approx n\>u, \qquad \sqrt{\frac{u\>\gamma_{2n}}{2}}\approx \sqrt{n} \> u. \end{eqnarray*} For small $n$ with $2n\>u<1$, one can bound \cite[Lemma 3.1]{Higham2002}, \begin{eqnarray*} \gamma_n\leq \frac{nu}{1-nu}, \qquad \sqrt{\frac{u\>\gamma_{2n}}{2}} \leq \frac{\sqrt{n}\> u}{\sqrt{1-2n\>u}} \end{eqnarray*} Thus, the probabilistic bound is proportional to $\sqrt{n}\>u$. Furthermore, $\gamma_n>\sqrt{u\>\gamma_{2n}/2}$ for $n\geq 2$. With a failure probability of $\delta=10^{-16}$, the probabilistic bound is tighter than the deterministic bound for $n>80$. \end{remark} \section{Numerical experiments}\label{s_numexp} After describing the setup for the experiments (Section~\ref{s_setup}), we present experiments for the perturbation bounds (Section~\ref{s_eperturb}), the roundoff error bounds assuming independence (Section~\ref{s_robi1}), and the general roundoff error bounds (Section~\ref{s_rog}). \subsection{Experimental Setup}\label{s_setup} We use a tiny failure probability of $\delta=10^{-16}$, which gives a probabilistic factor of $\sqrt{2\,\ln{(2/\delta)}}\leq 8.7$. Two types of vectors $\vx$ and $\vy$ of dimension up to $n=10^8$ will be considered: \begin{itemize} \item The elements of $\vx$ and $\vy$ can have different signs. Specifically, $x_j$ and $y_j$ are iid\footnote{independent identically distributed} standard normal random variables with mean~0 and variance~1, and $\vx$ and $\vy$ are generated with the Matlab commands \begin{quote} \texttt{x = \texttt{single(rand(n, 1))}, y = \texttt{single(rand(n, 1))}} \end{quote} \item The elements of $\vx$ and $\vy$ all have the same sign. Specifically, $x_j$ and $y_j$ are absolute values of iid standard normal random variables, and $\vx$ and $\vy$ are generated with the Matlab commands \begin{quote} \texttt{x = \texttt{single(abs(rand(n, 1)))}, y = \texttt{single(abs(rand(n, 1)))}} \end{quote} \end{itemize} The exact inner products $\vx^T\vy$ are represented by the double precision computation \texttt{dot(double(x), double(y))} with unit roundoff $2^{-53}\approx 1.11\cdot 10^{-16}$. Bounds are computed in double precision. Computations were performed in Matlab R2017a, on a 3.1GHz Intel Core i7 processor. \subsection{Experiments for the perturbation bounds}\label{s_eperturb} We illustrate the perturbation bounds in Section~\ref{s_perturb}. Here the vectors $\vx$ and $\vy$ are perturbed, while the computations are exact. We select single precision perturbations $\delta_j$ and $\theta_j$ that are uniformly distributed in $[-u, u]$, where $u=2^{-24}\approx 5.96\cdot 10^{-8}$ is the single precision roundoff, and generate the perturbation vectors $\vdelta$ and $\vtheta$ each with the Matlab command \begin{quote} \texttt{u * (2 * double(single(rand(n, 1))) - ones(n, 1))}. \end{quote} The inner product of the perturbed vectors $\vhx^T\vhy$ is represented by the double precision computation \texttt{dot(double(xh), double(yh))}. \subsubsection{Amplifiers in Corollary~\ref{c_2}}\label{s_perturbamp} We compare the amplifiers of $u(2+u)$ in the upper bounds of Corollary~\ref{c_2}, listed again below, \begin{eqnarray}\label{e_amp} \kappa_1&\equiv& \frac{\|\vx\circ\vy\|_1}{|\vx^T\vy|}=\frac{|\vx|^T|\vy|}{|\vx^T\vy|}=\frac{\sum_{j=1}^n{|x_jy_j|}}{|\vx^T\vy|}\\ \kappa_2 &\equiv& \sqrt{n}\>\frac{\|\vx\circ\vy\|_2}{|\vx^T\vy|}=\sqrt{n}\>\frac{\sqrt{\sum_{j=1}^{n}{|x_jy_j|^2}}}{|\vx^T\vy|}\nonumber\\ \kappa_{\infty} &\equiv& n\>\frac{\|\vx\circ\vy\|_{\infty}}{|\vx^T\vy|}=n\>\frac{\max_{1\leq j\leq n}{|x_jy_j|}}{|\vx^T\vy|}.\nonumber \end{eqnarray} Figure~\ref{f_amp} illustrates that, among the three amplifiers in (\ref{e_amp}), the traditional $\kappa_1$ tends to be the lowest. It also illustrates that amplification of roundoff can be orders of magnitude larger for vector elements with different signs, compared to vectors where all elements have the same sign. \begin{figure}[h] \begin{center} \includegraphics[width = 2.5in]{Fig_Corollary22_diff.eps} \includegraphics[width = 2.5in]{Fig_Corollary22_same.eps} \end{center} \caption{Comparison of amplifiers in (\ref{e_amp}): $\kappa_1$ (blue), $\kappa_2$ (red), and $\kappa_{\infty}$ (green) versus vector dimensions $1\leq n\leq 10^8$ in steps of $10^6$. Vertical axis starts at 1 and ends at $10^8$. Left panel: Elements can have different signs. Right panel: All elements have the same sign.} \label{f_amp} \end{figure} \subsubsection{Probabilistic perturbation bound in Theorem~\ref{t_1} and Remark~\ref{r_1}}\label{s_perturbcomp} This experiment follows up on Remark~\ref{r_1}, where we compare the probabilistic bound from Theorem~\ref{t_1} to the corresponding deterministic bound from Corollary~\ref{c_2}. \begin{itemize} \item Deterministic bound \begin{eqnarray} \left|\frac{\vhx^T\vhy-\vx^T\vy}{\vx^T\vy}\right| & \leq & \frac{\|\vx\circ\vy\|_2}{|\vx^T\vy|}\>\sqrt{n}\>u(2+u)\label{e_eperturb1}. \end{eqnarray} \item Probabilistic bound holding with probability at least $1-\delta$, \begin{eqnarray} \left|\frac{\vhx^T\vhy-\vx^T\vy}{\vx^T\vy}\right| &\leq & \frac{\|\vx\circ\vy\|_2}{|\vx^T\vy|}\>\sqrt{2\,\ln{(2/\delta)}}\> u(2+u)\label{e_eperturb2}. \end{eqnarray} \end{itemize} Figure~\ref{f_pert} illustrates that the probabilistic bound~(\ref{e_eperturb2}) tends to be at least two orders orders of magnitude tighter than the deterministic bound~(\ref{e_eperturb1}). \begin{figure}[h] \begin{center} \includegraphics[width = 2.5in]{Fig_Theorem24_diff.eps} \includegraphics[width = 2.5in]{Fig_Theorem24_same.eps} \end{center} \caption{Comparison of probabilistic bound (red \ref{e_eperturb2}) with deterministic bound (blue \ref{e_eperturb1}), and relative error (green) versus vector dimensions $1\leq n\leq 10^8$ in steps of $10^6$. Vertical axis starts at $10^{-14}$ and ends at 1. Left panel: Elements can have different signs. Right panel: All elements have the same sign.} \label{f_pert} \end{figure} \subsection{Experiments for the roundoff error bounds based on independent roundoff}\label{s_robi1} We illustrate the roundoff error bounds in Section~\ref{s_roundoff1}. The inner products $\fl(\vx^T\vy)$ are computed in single precision with unit roundoff, in a loop that explicitly stores the products $x_ky_k$ before adding them to the partial sum, so as to bypass the fused multiply-add. Specifically, we compare the probabilistic bound in Theorem~\ref{t_04} with the corresponding deterministic bound in Corollary~\ref{c_04}. \begin{itemize} \item Deterministic bound \begin{eqnarray}\label{e_robi1} \left|\frac{\fl(\vx^T\vy)-\vx^T\vy}{|\vx^T\vy|}\right| &\leq& \frac{\sqrt{\sum_{k=1}^n{c_k^2}}}{|\vx^T\vy|}\>\sqrt{n} \end{eqnarray} \item Probabilistic bound holding with probability at least $1-\delta$, \begin{eqnarray}\label{e_robi2} \left|\frac{\fl(\vx^T\vy)-\vx^T\vy}{\vx^T\vy}\right|\leq \frac{\sqrt{\sum_{k=1}^{n}{c_k^2}}}{|\vx^T\vy|}\>\sqrt{2\,\ln{(2/\delta)}}, \end{eqnarray} \end{itemize} where $c_1\equiv |x_1y_1|\>\gamma_n$, and $c_k\equiv |x_ky_k|\> \gamma_{n-k+2}$, $2\leq k\leq n$, and $\gamma_k=(1+u)^k-1$ as in~(\ref{e_gamma}). Figure~\ref{f_robi1} illustrates that the probabilistic result (\ref{e_robi2}) tends to be two orders of magnitude tighter than the deterministic bound (\ref{e_robi1}) for vectors whose elements can have different signs. However, (\ref{e_robi2}) stops being a bound for vectors of large dimension all of whose elements have the same sign. \begin{figure}[h] \begin{center} \includegraphics[width = 2.5in]{Fig_Theorem36_diff.eps} \includegraphics[width = 2.5in]{Fig_Theorem36_same.eps} \end{center} \caption{Comparison of probabilistic bound (red \ref{e_robi2}) with deterministic bound (blue \ref{e_robi1}), and relative error (green) versus vector dimensions $1\leq n\leq 10^8$ in steps of $10^6$. Vertical axis starts at $10^{-8}$ and ends at $10^8$. Left panel: Elements can have different signs. Right panel: All elements have the same sign.} \label{f_robi1} \end{figure} Figure~\ref{f_robi1smalllarge} zooms in on the left panel in Figure~\ref{f_robi1} and illustrates that (\ref{e_robi2}) remains an upper bound for vector dimensions up to about $n=10^6$. The fact that it ceases to be an upper bound for $n>10^6$ does not appear to be a numerical issue, as nothing changes when the products $|x_ky_k|$ are sorted in increasing or in decreasing order of magnitude. \begin{figure}[h] \begin{center} \includegraphics[width = 2.5in]{Fig_Theorem36_samesmall.eps} \includegraphics[width = 2.5in]{Fig_Theorem36_samelarge.eps} \end{center} \caption{Comparison of probabilistic bound (red \ref{e_robi2}) with deterministic bound (blue \ref{e_robi1}), and relative error (green) versus vector dimensions when all elements have the same sign. Vertical axis starts at $10^{-8}$ and ends at $1$. Left panel: Small dimensions $1\leq n \leq 10^6$ in steps of $10^4$. Right panel: Large dimensions $10^6\leq n\leq 10^8$ in steps of $10^6$.} \label{f_robi1smalllarge} \end{figure} \subsection{Experiments for the general roundoff error bounds}\label{s_rog} We illustrate the roundoff error bounds in Section~\ref{s_roundoff2}. As in the previous section, the inner products $\fl(\vx^T\vy)$ are computed in single precision with unit roundoff, in a loop that explicitly stores the products $x_ky_k$ before adding them to the partial sum, so as to bypass the fused multiply-add. This experiment follows up on Remark~\ref{r_2}, where we compare the probabilistic bound in Corollary~\ref{c_4} to the corresponding deterministic bound in Corollary~\ref{c_trad}. \begin{itemize} \item Traditional bound \begin{eqnarray}\label{e_rog1} \left|\frac{\fl(\vx^T\vy)-\vx^T\vy}{\vx^T\vy}\right| \leq \frac{|\vx|^T|\vy|}{|\vx^T\vy|}\> \gamma_n, \end{eqnarray} \item Probabilistic bound \begin{eqnarray}\label{e_rog2} \left|\frac{\fl(\vx^T\vy)-\vx^T\vy}{\vx^T\vy}\right| \leq \frac{|\vx|^T|\vy|}{|\vx^T\vy|}\> \sqrt{\ln{(2/\delta)}}\> \sqrt{\frac{u\>\gamma_{2n}}{2}}, \end{eqnarray} \end{itemize} where $\gamma_k=(1+u)^k-1$ as in~(\ref{e_gamma}). Figure~\ref{f_rog1} illustrates that the probabilistic result (\ref{e_rog2}) tends to be at least two orders of magnitude tighter than the deterministic bound (\ref{e_rog1}) for vectors whose elements can have different signs. However, unfortunately, (\ref{e_rog2}) stops being a bound for vectors of large dimension all of whose elements have the same sign. \begin{figure}[h] \begin{center} \includegraphics[width = 2.5in]{Fig_Corollary48_diff.eps} \includegraphics[width = 2.5in]{Fig_Corollary48_same.eps} \end{center} \caption{Comparison of probabilistic bound (red \ref{e_rog2}) with deterministic bound (blue \ref{e_rog1}), and relative error (green) versus vector dimensions $1\leq n\leq 10^8$ in steps of $10^6$. Vertical axis starts at $10^{-8}$ and ends at $10^8$. Left panel: Elements can have different signs. Right panel: All elements have the same sign.} \label{f_rog1} \end{figure} \section{Conclusions, and future work}\label{s_conc} We presented derivations and numerical experiments for probabilistic perturbation and roundoff error bounds for the sequentially accumulated inner product of two real $n$-vectors, assuming a guard digit model and no fused multiply-add. The probabilistic bounds are tighter than the corresponding deterministic bounds, often by several orders of magnitude. \paragraph{Issues} However, for vectors of dimension $n\geq 10^7$ and a tiny failure probability of $\delta=10^{-16}$, the probabilistic results are not entirely satisfactory: On the one hand, they are still too pessimistic for vectors whose elements have different signs, while on the other hand they stops being upper bounds for vectors all of whose elements have the same sign --regardless of whether roundoffs are assumed to be independent or not. The latter phenomenon does not appear to be a numerical artifact. A simple fix would be to adjust the failure probability, making it even more stringent when elements can differ in sign, while relaxing it when all elements have the same sign. However, this does not get to the heart of the problem. Should the failure probability be explicitly and systematically tied to the dimension~$n$? This would be inconsistent with concentration inequalities, which do not explicitly depend on the number of summands. Alternatively, should one not model roundoffs as zero-mean random variables, but instead introduce a bias, possibly dimension-dependent, for vectors with structure, such as those where all elements have the same sign, see also \cite[section 4.2]{HM18}. \begin{comment} \paragraph{Future work} A few of the obvious next steps for inner products include: \begin{itemize} \item Different arithmetic models, e.g. fused multiply-add and no guard digits. \item Checking the informativeness of the bounds for a range of precisions, including half and quadruple. \item Different summation algorithms, such as tree summation, and approaches that exploit parallelism \item Modeling roundoffs as biased random variables in the context when vectors have structure, such as non-negative elements, geometrically decreasing elements, or sparsity, see also \cite[section 4.2]{HM18}. \item In the context of data analysis, relinquishing the focus on the number of correct digits in favour of checking the magnitude or even only the sign for correctness. \end{itemize} \end{comment}

https://arxiv.org/abs/2206.01608

Structure-Preserving Model Order Reduction for Index One Port-Hamiltonian Descriptor Systems

We develop optimization-based structure-preserving model order reduction (MOR) methods for port-Hamiltonian (pH) descriptor systems of differentiation index one. Descriptor systems in pH form permit energy-based modeling and intuitive coupling of physical systems across different physical domains, scales, and accuracies. This makes pH models well-suited building-blocks for component-wise modeling of large system networks. In this context, it is often necessary to preserve the pH structure during MOR. We discuss current projection-based and structure-preserving MOR algorithms for pH systems and present a new optimization-based framework for that task. The benefits of our method include a simplified treatment of algebraic constraints and often a higher accuracy of the resulting reduced-order model, which is demonstrated by several numerical examples.

\subsection{PH-preserving MOR techniques}% Traditional methods which directly retain the pH form in the reduction process are based on Galerkin projections, see, e.g., \cite{Antoulas2020, MehU22} for surveys. The original state $x(t)$ is approximated by ${x(t)\approx V x_r}$, where the columns of ${V\in\R^{n\times r}}$ form a basis for a suitably chosen subspace of dimension $r$. For instance, in the pH-ODE case with $\pE= I_n$, the ROM coefficient matrices are computed by \begin{equation*} \begin{array}{lll} \rJ=U^\T\pJ U, & \rG = U^\T\pG, & \rQ = V^\T\pQ V, \\ \rR=U^\T\pR U, & \rP = U^\T\pP, & \rN=\pN, \quad \rS=\pS, \end{array} \end{equation*} where ${U = QV(V^\T QV)^{-1}}$, which clearly enforces structure preservation. In \cite{Gugercin2012} an adaptation (IRKA-PH) of the well-known iterative rational Krylov algorithm (IRKA) was proposed that iteratively updates $V$ to fulfill a subset of $\cH_2$ optimality conditions via tangential interpolation of the original transfer function. While IRKA does not ensure stability (or even passivity) of the ROM a priori, it leads to (locally) $\cH_2$ optimal models upon convergence. IRKA-PH, on the other hand, preserves the pH structure and thus produces passive ROMs which, however, generally only fulfill a subset of the $\cH_2$ optimality conditions. Consequently, no $\cH_2$ optimality is achieved in general. The matrix $V$ can also be chosen in order to approximate the Dirac structure of the original model, resulting in the effort- and flow-constraint method \cite{Polyuga2012}. The extension of IRKA-PH and Dirac structure approximation to pH-DAEs is addressed in \cite{HauMM19,BeaGM21}. \subsection{Passivity-preserving MOR techniques}% Positive real balanced truncation (PRBT) for ODEs is a well-studied MOR method, see \cite{GugA04} and the references therein for a survey. The method is based on computing a minimal solution $X_{\min}$ of \eqref{eq:KYP_ineq} (in the sense of the Loewner ordering in symmetric matrices) and a minimal solution $Y_{\min}$ of the \emph{dual KYP inequality} \begin{equation*} \mat{-AY - YA^\T & B-YC^\T \\ B^\T-C Y & D+D^\T} \geq 0. \end{equation*} The solutions $X_{\min}$ and $Y_{\min}$ are then used to transform the system to a positive-real balanced realization where the transformed minimal solution of the KYP inequalities $\hat{X}_{\min}$ and $\hat{Y}_{\min}$ are equal and diagonal, i.\,e., $\hat{X}_{\min} = \hat{Y}_{\min} = \text{diag}(\eta_1,\ldots,\eta_n)$. Then the states corresponding to the small positive real characteristic values $\eta_i$ can be truncated. Finally, PRBT admits an a priori error bound in the gap metric that is derived in \cite{GuiO14}. Another passivity-preserving MOR technique is achieved via spectral factorization \cite{BreU21}. It initially requires a factorization \begin{equation}\label{eq:specfac} \mathcal{W}(X)=\mat{L&M}^\T\mat{L&M} \end{equation} with $L\in\R^{k\times n}$, $M\in\R^{k\times m}$ and where $\text{rank}\mat{L&M}$ is as small as possible. If the pair $(A,B)$ is \emph{stabilizable} and $X = X_{\min}$ is the minimal solution of $\mathcal{W}(X) \ge 0$, then the resulting spectral factor system $(A,B,L,M)$ may be reduced via traditional (unstructured) MOR techniques, such as IRKA or balanced truncation (BT), and the passive ROM is obtained from the reduced spectral factor. For all passivity-preserving reduction methods, the passive ROM may eventually be transformed back to a pH representation by applying Theorem \ref{thm:KYP}. \section{Introduction} \file{elsarticle.cls} is a thoroughly re-written document class for formatting \LaTeX{} submissions to Elsevier journals. The class uses the environments and commands defined in \LaTeX{} kernel without any change in the signature so that clashes with other contributed \LaTeX{} packages such as \file{hyperref.sty}, \file{preview-latex.sty}, etc., will be minimal. \file{elsarticle.cls} is primarily built upon the default \file{article.cls}. This class depends on the following packages for its proper functioning: \begin{enumerate} \item \file{natbib.sty} for citation processing; \item \file{geometry.sty} for margin settings; \item \file{fleqn.clo} for left aligned equations; \item \file{graphicx.sty} for graphics inclusion; \item \file{txfonts.sty} optional font package, if the document is to be formatted with Times and compatible math fonts; \item \file{hyperref.sty} optional packages if hyperlinking is required in the document; \item \file{endfloat.sty} optional packages if floats to be placed at end of the PDF. \end{enumerate} All the above packages (except some optional packages) are part of any standard \LaTeX{} installation. Therefore, the users need not be bothered about downloading any extra packages. Furthermore, users are free to make use of \textsc{ams} math packages such as \file{amsmath.sty}, \file{amsthm.sty}, \file{amssymb.sty}, \file{amsfonts.sty}, etc., if they want to. All these packages work in tandem with \file{elsarticle.cls} without any problems. \section{Major Differences} Following are the major differences between \file{elsarticle.cls} and its predecessor package, \file{elsart.cls}: \begin{enumerate}[\textbullet] \item \file{elsarticle.cls} is built upon \file{article.cls} while \file{elsart.cls} is not. \file{elsart.cls} redefines many of the commands in the \LaTeX{} classes/kernel, which can possibly cause surprising clashes with other contributed \LaTeX{} packages; \item provides preprint document formatting by default, and optionally formats the document as per the final style of models $1+$, $3+$ and $5+$ of Elsevier journals; \item some easier ways for formatting \verb+list+ and \verb+theorem+ environments are provided while people can still use \file{amsthm.sty} package; \item \file{natbib.sty} is the main citation processing package which can comprehensively handle all kinds of citations and works perfectly with \file{hyperref.sty} in combination with \file{hypernat.sty}; \item long title pages are processed correctly in preprint and final formats. \end{enumerate} \section{Installation} The package is available at author resources page at Elsevier (\url{http://www.elsevier.com/locate/latex}). It can also be found in any of the nodes of the Comprehensive \TeX{} Archive Network (\textsc{ctan}), one of the primary nodes being \url{http://tug.ctan.org/tex-archive/macros/latex/contrib/elsarticle/}. Please download the \file{elsarticle.dtx} which is a composite class with documentation and \file{elsarticle.ins} which is the \LaTeX{} installer file. When we compile the \file{elsarticle.ins} with \LaTeX{} it provides the class file, \file{elsarticle.cls} by stripping off all the documentation from the \verb+*.dtx+ file. The class may be moved or copied to a place, usually, \verb+$TEXMF/tex/latex/elsevier/+, or a folder which will be read by \LaTeX{} during document compilation. The \TeX{} file database needs updation after moving/copying class file. Usually, we use commands like \verb+mktexlsr+ or \verb+texhash+ depending upon the distribution and operating system. \section{Usage}\label{sec:usage} The class should be loaded with the command: \begin{vquote} \documentclass[<options>]{elsarticle} \end{vquote} \noindent where the \verb+options+ can be the following: \begin{description} \item [{\tt\color{verbcolor} preprint}] default option which format the document for submission to Elsevier journals. \item [{\tt\color{verbcolor} review}] similar to the \verb+preprint+ option, but increases the baselineskip to facilitate easier review process. \item [{\tt\color{verbcolor} 1p}] formats the article to the look and feel of the final format of model 1+ journals. This is always single column style. \item [{\tt\color{verbcolor} 3p}] formats the article to the look and feel of the final format of model 3+ journals. If the journal is a two column model, use \verb+twocolumn+ option in combination. \item [{\tt\color{verbcolor} 5p}] formats for model 5+ journals. This is always of two column style. \item [{\tt\color{verbcolor} authoryear}] author-year citation style of \file{natbib.sty}. If you want to add extra options of \file{natbib.sty}, you may use the options as comma delimited strings as arguments to \verb+\biboptions+ command. An example would be: \end{description} \begin{vquote} \biboptions{longnamesfirst,angle,semicolon} \end{vquote} \begin{description} \item [{\tt\color{verbcolor} number}] numbered citation style. Extra options can be loaded with\linebreak \verb+\biboptions+ command. \item [{\tt\color{verbcolor} sort\&compress}] sorts and compresses the numbered citations. For example, citation [1,2,3] will become [1--3]. \item [{\tt\color{verbcolor} longtitle}] if front matter is unusually long, use this option to split the title page across pages with the correct placement of title and author footnotes in the first page. \item [{\tt\color{verbcolor} times}] loads \file{txfonts.sty}, if available in the system to use Times and compatible math fonts. \item [{\tt\color{verbcolor} reversenotenum}] Use alphabets as author--affiliation linking labels and use numbers for author footnotes. By default, numbers will be used as author--affiliation linking labels and alphabets for author footnotes. \item [{\tt\color{verbcolor} lefttitle}] To move title and author/affiliation block to flushleft. \verb+centertitle+ is the default option which produces center alignment. \item [{\tt\color{verbcolor} endfloat}] To place all floats at the end of the document. \item [{\tt\color{verbcolor} nonatbib}] To unload natbib.sty. \item [{\tt\color{verbcolor} doubleblind}] To hide author name, affiliation, email address etc. for double blind refereeing purpose. \item[] All options of \file{article.cls} can be used with this document class. \item[] The default options loaded are \verb+a4paper+, \verb+10pt+, \verb+oneside+, \verb+onecolumn+ and \verb+preprint+. \end{description} \section{Frontmatter} There are two types of frontmatter coding: \begin{enumerate}[(1)] \item each author is connected to an affiliation with a footnote marker; hence all authors are grouped together and affiliations follow; \pagebreak \item authors of same affiliations are grouped together and the relevant affiliation follows this group. \end{enumerate} An example of coding the first type is provided below. \begin{vquote} \title{This is a specimen title\tnoteref{t1,t2}} \tnotetext[t1]{This document is the results of the research project funded by the National Science Foundation.} \tnotetext[t2]{The second title footnote which is a longer text matter to fill through the whole text width and overflow into another line in the footnotes area of the first page.} \end{vquote} \begin{vquote} \author[1]{Jos Migchielsen\corref{cor1}% \fnref{fn1}} \ead{[email protected]} \author[2]{CV Radhakrishnan\fnref{fn2}} \ead{[email protected]} \author[3]{CV Rajagopal\fnref{fn1,fn3}} \ead[url]{www.stmdocs.in} \cortext[cor1]{Corresponding author} \fntext[fn1]{This is the first author footnote.} \fntext[fn2]{Another author footnote, this is a very long footnote and it should be a really long footnote. But this footnote is not yet sufficiently long enough to make two lines of footnote text.} \fntext[fn3]{Yet another author footnote.} \affiliation[1]{organization={Elsevier B.V.}, addressline={Radarweg 29}, postcode={1043 NX}, city={Amsterdam}, country={The Netherlands}} \affiliation[2]{organization={Sayahna Foundation}, addressline={JWRA 34, Jagathy}, city={Trivandrum} postcode={695014}, country={India}} \end{vquote} \begin{vquote} \affiliation[3]{organization={STM Document Engineering Pvt Ltd.}, addressline={Mepukada, Malayinkil}, city={Trivandrum} postcode={695571}, country={India}} \end{vquote} The output of the above \TeX{} source is given in Clips~\ref{clip1} and \ref{clip2}. The header portion or title area is given in Clip~\ref{clip1} and the footer area is given in Clip~\ref{clip2}. \deforange{blue!70} \src{Header of the title page.} \includeclip{1}{130 612 477 707}{1psingleauthorgroup.pdf \deforange{orange} \deforange{blue!70} \src{Footer of the title page.} \includeclip{1}{93 135 499 255}{1pseperateaug.pdf \deforange{orange} Most of the commands such as \verb+\title+, \verb+\author+, \verb+\affiliation+ are self explanatory. Various components are linked to each other by a label--reference mechanism; for instance, title footnote is linked to the title with a footnote mark generated by referring to the \verb+\label+ string of the \verb=\tnotetext=. We have used similar commands such as \verb=\tnoteref= (to link title note to title); \verb=\corref= (to link corresponding author text to corresponding author); \verb=\fnref= (to link footnote text to the relevant author names). \TeX{} needs two compilations to resolve the footnote marks in the preamble part. Given below are the syntax of various note marks and note texts. \begin{vquote} \tnoteref{<label(s)>} \corref{<label(s)>} \fnref{<label(s)>} \tnotetext[<label>]{<title note text>} \cortext[<label>]{<corresponding author note text>} \fntext[<label>]{<author footnote text>} \end{vquote} \noindent where \verb=<label(s)>= can be either one or more comma delimited label strings. The optional arguments to the \verb=\author= command holds the ref label(s) of the address(es) to which the author is affiliated while each \verb=\affiliation= command can have an optional argument of a label. In the same manner, \verb=\tnotetext=, \verb=\fntext=, \verb=\cortext= will have optional arguments as their respective labels and note text as their mandatory argument. The following example code provides the markup of the second type of author-affiliation. \begin{vquote} \author{Jos Migchielsen\corref{cor1}% \fnref{fn1}} \ead{[email protected]} \affiliation[1]{organization={Elsevier B.V.}, addressline={Radarweg 29}, postcode={1043 NX}, city={Amsterdam}, country={The Netherlands}} \author{CV Radhakrishnan\fnref{fn2}} \ead{[email protected]} \affiliation[2]{organization={Sayahna Foundation}, addressline={JWRA 34, Jagathy}, city={Trivandrum} postcode={695014}, country={India}} \author{CV Rajagopal\fnref{fn1,fn3}} \ead[url]{www.stmdocs.in} \affiliation[3]{organization={STM Document Engineering Pvt Ltd.}, addressline={Mepukada, Malayinkil}, city={Trivandrum} postcode={695571}, country={India}} \end{vquote} \vspace*{-.5pc} \begin{vquote} \cortext[cor1]{Corresponding author} \fntext[fn1]{This is the first author footnote.} \fntext[fn2]{Another author footnote, this is a very long footnote and it should be a really long footnote. But this footnote is not yet sufficiently long enough to make two lines of footnote text.} \end{vquote} The output of the above \TeX{} source is given in Clip~\ref{clip3}. \deforange{blue!70} \src{Header of the title page..} \includeclip{1}{119 563 468 709}{1pseperateaug.pdf \deforange{orange} Clip~\ref{clip4} shows the output after giving \verb+doubleblind+ class option. \deforange{blue!70} \src{Double blind article} \includeclip{1}{124 567 477 670}{elstest-1pdoubleblind.pdf \deforange{orange} \vspace*{-.5pc} The frontmatter part has further environments such as abstracts and keywords. These can be marked up in the following manner: \begin{vquote} \begin{abstract} In this work we demonstrate the formation of a new type of polariton on the interface between a .... \end{abstract} \end{vquote} \vspace*{-.5pc} \begin{vquote} \begin{keyword} quadruple exiton \sep polariton \sep WGM \end{keyword} \end{vquote} \noindent Each keyword shall be separated by a \verb+\sep+ command. \textsc{msc} classifications shall be provided in the keyword environment with the commands \verb+\MSC+. \verb+\MSC+ accepts an optional argument to accommodate future revisions. eg., \verb=\MSC[2008]=. The default is 2000.\looseness=-1 \subsection{New page} Sometimes you may need to give a page-break and start a new page after title, author or abstract. Following commands can be used for this purpose. \begin{vquote} \newpageafter{title} \newpageafter{author} \newpageafter{abstract} \end{vquote} \begin{itemize} \leftskip-2pc \item [] {\tt\color{verbcolor} \verb+\newpageafter{title}+} typeset the title alone on one page. \item [] {\tt\color{verbcolor} \verb+\newpageafter{author}+} typeset the title and author details on one page. \item [] {\tt\color{verbcolor} \verb+\newpageafter{abstract}+} typeset the title, author details and abstract \& keywords one one page. \end{itemize} \section{Floats} {Figures} may be included using the command, \verb+\includegraphics+ in combination with or without its several options to further control graphic. \verb+\includegraphics+ is provided by \file{graphic[s,x].sty} which is part of any standard \LaTeX{} distribution. \file{graphicx.sty} is loaded by default. \LaTeX{} accepts figures in the postscript format while pdf\LaTeX{} accepts \file{*.pdf}, \file{*.mps} (metapost), \file{*.jpg} and \file{*.png} formats. pdf\LaTeX{} does not accept graphic files in the postscript format. The \verb+table+ environment is handy for marking up tabular material. If users want to use \file{multirow.sty}, \file{array.sty}, etc., to fine control/enhance the tables, they are welcome to load any package of their choice and \file{elsarticle.cls} will work in combination with all loaded packages. \section[Theorem and ...]{Theorem and theorem like environments} \file{elsarticle.cls} provides a few shortcuts to format theorems and theorem-like environments with ease. In all commands the options that are used with the \verb+\newtheorem+ command will work exactly in the same manner. \file{elsarticle.cls} provides three commands to format theorem or theorem-like environments: \begin{vquote} \newtheorem{thm}{Theorem} \newtheorem{lem}[thm]{Lemma} \newdefinition{rmk}{Remark} \newproof{pf}{Proof} \newproof{pot}{Proof of Theorem \ref{thm2}} \end{vquote} The \verb+\newtheorem+ command formats a theorem in \LaTeX's default style with italicized font, bold font for theorem heading and theorem number at the right hand side of the theorem heading. It also optionally accepts an argument which will be printed as an extra heading in parentheses. \begin{vquote} \begin{thm} For system (8), consensus can be achieved with $\|T_{\omega z}$ ... \begin{eqnarray}\label{10} .... \end{eqnarray} \end{thm} \end{vquote} Clip~\ref{clip5} will show you how some text enclosed between the above code\goodbreak \noindent looks like: \vspace*{6pt} \deforange{blue!70} \src{{\ttfamily\color{verbcolor}\expandafter\@gobble\string\\ newtheorem}} \includeclip{2}{1 1 453 120}{jfigs.pdf} \deforange{orange} The \verb+\newdefinition+ command is the same in all respects as its\linebreak \verb+\newtheorem+ counterpart except that the font shape is roman instead of italic. Both \verb+\newdefinition+ and \verb+\newtheorem+ commands automatically define counters for the environments defined. \vspace*{6pt} \deforange{blue!70} \src{{\ttfamily\color{verbcolor}\expandafter\@gobble\string\\ newdefinition}} \includeclip{1}{1 1 453 105}{jfigs.pdf} \deforange{orange} The \verb+\newproof+ command defines proof environments with upright font shape. No counters are defined. \vspace*{6pt} \deforange{blue!70} \src{{\ttfamily\color{verbcolor}\expandafter\@gobble\string\\ newproof}} \includeclip{3}{1 1 453 65}{jfigs.pdf} \deforange{orange} Users can also make use of \verb+amsthm.sty+ which will override all the default definitions described above. \section[Enumerated ...]{Enumerated and Itemized Lists} \file{elsarticle.cls} provides an extended list processing macros which makes the usage a bit more user friendly than the default \LaTeX{} list macros. With an optional argument to the \verb+\begin{enumerate}+ command, you can change the list counter type and its attributes. \begin{vquote} \begin{enumerate}[1.] \item The enumerate environment starts with an optional argument `1.', so that the item counter will be suffixed by a period. \item You can use `a)' for alphabetical counter and '(i)' for roman counter. \begin{enumerate}[a)] \item Another level of list with alphabetical counter. \item One more item before we start another. \end{vquote} \deforange{blue!70} \src{List -- Enumerate} \includeclip{4}{1 1 453 185}{jfigs.pdf} \deforange{orange} Further, the enhanced list environment allows one to prefix a string like `step' to all the item numbers. \begin{vquote} \begin{enumerate}[Step 1.] \item This is the first step of the example list. \item Obviously this is the second step. \item The final step to wind up this example. \end{enumerate} \end{vquote} \deforange{blue!70} \src{List -- enhanced} \includeclip{5}{1 1 313 83}{jfigs.pdf} \deforange{orange} \section{Cross-references} In electronic publications, articles may be internally hyperlinked. Hyperlinks are generated from proper cross-references in the article. For example, the words \textcolor{black!80}{Fig.~1} will never be more than simple text, whereas the proper cross-reference \verb+\ref{tiger}+ may be turned into a hyperlink to the figure itself: \textcolor{blue}{Fig.~1}. In the same way, the words \textcolor{blue}{Ref.~[1]} will fail to turn into a hyperlink; the proper cross-reference is \verb+\cite{Knuth96}+. Cross-referencing is possible in \LaTeX{} for sections, subsections, formulae, figures, tables, and literature references. \section[Mathematical ...]{Mathematical symbols and formulae} Many physical/mathematical sciences authors require more mathematical symbols than the few that are provided in standard \LaTeX. A useful package for additional symbols is the \file{amssymb} package, developed by the American Mathematical Society. This package includes such oft-used symbols as $\lesssim$ (\verb+\lesssim+), $\gtrsim$ (\verb+\gtrsim+) or $\hbar$ (\verb+\hbar+). Note that your \TeX{} system should have the \file{msam} and \file{msbm} fonts installed. If you need only a few symbols, such as $\Box$ (\verb+\Box+), you might try the package \file{latexsym}. Another point which would require authors' attention is the breaking up of long equations. When you use \file{elsarticle.cls} for formatting your submissions in the \verb+preprint+ mode, the document is formatted in single column style with a text width of 384pt or 5.3in. When this document is formatted for final print and if the journal happens to be a double column journal, the text width will be reduced to 224pt at for 3+ double column and 5+ journals respectively. All the nifty fine-tuning in equation breaking done by the author goes to waste in such cases. Therefore, authors are requested to check this problem by typesetting their submissions in final format as well just to see if their equations are broken at appropriate places, by changing appropriate options in the document class loading command, which is explained in section~\ref{sec:usage}, \nameref{sec:usage}. This allows authors to fix any equation breaking problem before submission for publication. \file{elsarticle.cls} supports formatting the author submission in different types of final format. This is further discussed in section \ref{sec:final}, \nameref{sec:final}. \enlargethispage*{\baselineskip} \subsection*{Displayed equations and double column journals} Many Elsevier journals print their text in two columns. Since the preprint layout uses a larger line width than such columns, the formulae are too wide for the line width in print. Here is an example of an equation (see equation 6) which is perfect in a single column preprint format: In normal course, articles are prepared and submitted in single column format even if the final printed article will come in a double column format journal. Here the problem is that when the article is typeset by the typesetters for paginating and fit within the single column width, they have to break the lengthy equations and align them properly. Even if most of the tasks in preparing your proof is automated, the equation breaking and aligning requires manual judgement, hence this task is manual. When there comes a manual operation that area is error prone. Author needs to check that equation pretty well. However if authors themselves break the equation to the single column width typesetters need not want to touch these area and the proof authors get will be without any errors. \setlength\Sep{6pt} \src{See equation (6)} \deforange{blue!70} \includeclip{4}{105 500 500 700}{1psingleauthorgroup.pdf} \deforange{orange} \noindent When this document is typeset for publication in a model 3+ journal with double columns, the equation will overlap the second column text matter if the equation is not broken at the appropriate location. \vspace*{6pt} \deforange{blue!70} \src{See equation (6) overprints into second column} \includeclip{3}{59 421 532 635}{elstest-3pd.pdf} \deforange{orange} \vspace*{6pt} \noindent The typesetter will try to break the equation which need not necessarily be to the liking of the author or as it happens, typesetter's break point may be semantically incorrect. Therefore, authors may check their submissions for the incidence of such long equations and break the equations at the correct places so that the final typeset copy will be as they wish. \section{Bibliography} Three bibliographic style files (\verb+*.bst+) are provided --- \file{elsarticle-num.bst}, \file{elsarticle-num-names.bst} and \file{elsarticle-harv.bst} --- the first one can be used for the numbered scheme, second one for numbered with new options of \file{natbib.sty}. The third one is for the author year scheme. In \LaTeX{} literature, references are listed in the \verb+thebibliography+ environment. Each reference is a \verb+\bibitem+ and each \verb+\bibitem+ is identified by a label, by which it can be cited in the text: \verb+\bibitem[Elson et al.(1996)]{ESG96}+ is cited as \verb+\citet{ESG96}+. \noindent In connection with cross-referencing and possible future hyperlinking it is not a good idea to collect more that one literature item in one \verb+\bibitem+. The so-called Harvard or author-year style of referencing is enabled by the \LaTeX{} package \file{natbib}. With this package the literature can be cited as follows: \begin{enumerate}[\textbullet] \item Parenthetical: \verb+\citep{WB96}+ produces (Wettig \& Brown, 1996). \item Textual: \verb+\citet{ESG96}+ produces Elson et al. (1996). \item An affix and part of a reference: \verb+\citep[e.g.][Ch. 2]{Gea97}+ produces (e.g. Governato et al., 1997, Ch. 2). \end{enumerate} In the numbered scheme of citation, \verb+\cite{<label>}+ is used, since \verb+\citep+ or \verb+\citet+ has no relevance in the numbered scheme. \file{natbib} package is loaded by \file{elsarticle} with \verb+numbers+ as default option. You can change this to author-year or harvard scheme by adding option \verb+authoryear+ in the class loading command. If you want to use more options of the \file{natbib} package, you can do so with the \verb+\biboptions+ command, which is described in the section \ref{sec:usage}, \nameref{sec:usage}. For details of various options of the \file{natbib} package, please take a look at the \file{natbib} documentation, which is part of any standard \LaTeX{} installation. In addition to the above standard \verb+.bst+ files, there are 10 journal-specific \verb+.bst+ files also available. Instruction for using these \verb+.bst+ files can be found at \href{http://support.stmdocs.in/wiki/index.php?title=Model-wise_bibliographic_style_files} {http://support.stmdocs.in} \section[Graphical ...]{Graphical abstract and highlights} A template for adding graphical abstract and highlights are available now. This will appear as the first two pages of the PDF before the article content begins. \pagebreak Please refer below to see how to code them. \begin{vquote} .... .... \end{abstract} \begin{graphicalabstract} \end{graphicalabstract} \begin{highlights} \item Research highlight 1 \item Research highlight 2 \end{highlights} \begin{keyword} .... .... \end{vquote} \section{Final print}\label{sec:final} The authors can format their submission to the page size and margins of their preferred journal. \file{elsarticle} provides four class options for the same. But it does not mean that using these options you can emulate the exact page layout of the final print copy. \lmrgn=3em \begin{description} \item [\texttt{1p}:] $1+$ journals with a text area of 384pt $\times$ 562pt or 13.5cm $\times$ 19.75cm or 5.3in $\times$ 7.78in, single column style only. \item [\texttt{3p}:] $3+$ journals with a text area of 468pt $\times$ 622pt or 16.45cm $\times$ 21.9cm or 6.5in $\times$ 8.6in, single column style. \item [\texttt{twocolumn}:] should be used along with 3p option if the journal is $3+$ with the same text area as above, but double column style. \item [\texttt{5p}:] $5+$ with text area of 522pt $\times$ 682pt or 18.35cm $\times$ 24cm or 7.22in $\times$ 9.45in, double column style only. \end{description} Following pages have the clippings of different parts of the title page of different journal models typeset in final format. Model $1+$ and $3+$ will have the same look and feel in the typeset copy when presented in this document. That is also the case with the double column $3+$ and $5+$ journal article pages. The only difference will be wider text width of higher models. Here are the specimen single and double column journal pages. \begin{comment} \begin{center} \hypertarget{bsc}{} \hyperlink{sc}{ {\bf [Specimen single column article -- Click here]} } \hypertarget{bsc}{} \hyperlink{dc}{ {\bf [Specimen double column article -- Click here]} } \end{center} \end{comment} \vspace*{-.5pc} \enlargethispage*{\baselineskip} \src{}\hypertarget{sc}{} \deforange{blue!70} \hyperlink{bsc}{\includeclip{1}{88 120 514 724}{elstest-1p.pdf}} \deforange{orange} \src{}\hypertarget{dc}{} \deforange{blue!70} \hyperlink{bsc}{\includeclip{1}{27 61 562 758}{elstest-5p.pdf}} \deforange{orange} ~\hfill $\Box$ \end{document} \section{} \label{} \section{} \label{} \section{} \label{} \section{Introduction}% \label{sec:introduction} \input{Introduction} \section{Preliminaries}% \label{sec:preliminaries} \input{Prelims} \section{A new optimization-based approach}% \label{sec:our_method} \input{pauls_approach} \input{tims_approach} \section{Numerical examples}% \label{sec:numerical_examples} \input{Experimental_setup.tex} \input{Results.tex} \section{Conclusion}% \label{sec:conclusion} We have presented two optimization-based methods for structure-preserving MOR of pH-DAEs. These make it possible to compute accurate ROMs with respect to either the $\hinf$ or the $\htwo$ norm. The main benefits compared to state-of-the-art methods are the simplified treatment of the algebraic equations, which can be incorporated into the parameterized ROM in a structure-preserving way without increasing its state dimension. Furthermore, our methods are data-driven, such that no transformations to the FOM system matrices are required. Nonetheless, we have shown, how transformations can be applied in order to obtain an accurate estimate of the feedthrough, which is essential in the $\htwo$ case. Finally, our numerical experiments show that the optimization-based methods often lead to a higher accuracy (especially in the $\hinf$ norm). We are currently investigating the application of our method to higher index pH-DAEs. These may have improper transfer functions, which are not currently supported in our parameterization. \bibliographystyle{elsarticle-num} \subsection{$\hinf$ approximation}% \label{sub:ourmethod_hinf} The algorithm presented in \cite{Schwerdtner2020} to obtain a good $\hinf$ approximation of a pH-FOM with transfer function $\pHtf$ using a parametrized low-order system with transfer function $\pHtfr( \cdot, \theta)$ is based on minimizing the objective function \begin{align} \begin{split} \pushleft{\loss(\theta;\pHtf,\gamma,\mathcal{S}) :=} \\ \pushright{\frac{1}{\gamma}\sum\limits_{s_i\in \mathcal{S}} \sum\limits_{j=1}^{m}{\left({\left[\sigma_j \left(\pHtf(s_i)-\pHtfr(s_i,\theta)\right)-\gamma\right]}_+\right)}^2} \end{split} \label{eq:loss} \end{align} with respect to $\theta$, where \begin{align*} {[ \cdot ]}_+: \R \rightarrow [0,\infty), \quad x \mapsto \begin{cases} x & \text{if } x\ge 0,\\ 0 & \text{if } x<0, \end{cases} \end{align*} for decreasing values of $\gamma > 0$. Here $\mathcal{S} \subset \ri \R$ is a set of sample points, at which the original and reduced transfer functions are evaluated and $\sigma_j(\cdot)$ denotes the $j$-th singular value of its matrix argument. The justification for using $\loss$ as a surrogate for the $\hinf$ error is that $\loss(\cdot;H,\gamma,\mathcal{S})$ attains its global minimum (at zero), when all singular values of the error transfer function at all sample points $s_i$ are below the threshold~$\gamma$. Therefore, if the sample points are chosen appropriately, then $\loss(\theta;\pHtf,\gamma,\mathcal{S}) = 0$ is an indication for ${{\|\pHtf - \pHtf_r(\cdot,\theta)\|}_{\cH_\infty} \le \gamma}$. In \cite{Schwerdtner2021}, an adaptive sampling procedure is introduced which ensures that the sample points are appropriately distributed along the imaginary axis based on the given FOM and status of the optimization. In this work, we use Algorithm~\ref{alg:bisection} to determine an $\hinf$ approximation via a bisection procedure, which determines the minimal value for $\gamma$ (up to a relative tolerance $\epsilon_1$) at which a minimization of $\loss$ with respect to $\theta$ terminates at zero. The tolerance $\epsilon_2$ that is used in line $6$ of the algorithm is the maximum value of $\loss$, which is still numerically interpreted as zero, such that $\gamma$ is reduced in the subsequent bisection step. The sample points are updated after each bisection step because the adaptive sampling update rule introduced in~\cite{Schwerdtner2021} depends on the current value of $\gamma$. \begin{algorithm}[tbh] \LinesNumbered \SetAlgoLined \DontPrintSemicolon \SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output} \Input{FOM transfer function $\pHtf\in\mathcal{RH}_{\infty}^{\enu \times \enu}$, initial ROM transfer function $\pHtfr( \cdot, \theta_0) \in \rhinf^{\enu \times \enu}$ with parameter $\theta_0 \in \R^{n_\theta}$, initial sample point set $\mathcal{S} \subset \mathrm{i}\R$, upper bound $\gamma_{\rm u} > 0$, bisection tolerance $\varepsilon_1 > 0$, termination tolerance $\varepsilon_2 > 0$ } \Output{Reduced pH-ODE of order $r$} Set $j:=0$ and $\gamma_{\rm l}:=0$.\; \While{$(\gamma_{\rm u}-\gamma_{\rm l})/(\gamma_{\rm u}+\gamma_{\rm l}) > \varepsilon_1$}{ Set $\gamma:=(\gamma_{\rm u}+\gamma_{\rm l})/2$.\; Update the sample set $\mathcal{S}$ using~\cite[Alg.~3.1]{Schwerdtner2021}. \; Solve the minimization problem $\alpha := \min_{\theta\in \R^{n_\theta}}\loss(\theta;\pHtf,\gamma,\mathcal{S})$ with minimizer $\theta_{j+1} \in \R^{n_\theta}$, initialized at $\theta_j$. \; \eIf{$\alpha > \varepsilon_2$}{ Set $\gamma_{\rm l}:=\gamma$.\; }{ Set $\gamma_{\rm u}:=\gamma$.\; } Set $j:=j+1$. } Construct the ROM with $\theta_j$ as in Lemma~\ref{lem:PHParam}. \caption{SOBMOR-$\hinf$} \label{alg:bisection} \end{algorithm} The benefits of using this approach instead of directly minimizing the $\hinf$ norm were discussed in detail in~\cite[Remark~3.3]{Schwerdtner2020}. The main reasons for using~\eqref{eq:loss} instead of the $\hinf$ norm are the differentiability of~$\loss$ with respect to~$\theta$, the local convergence of the method, and the prohibitive computational costs as well as reliability issues of the $\hinf$ norm computation (for the large-scale error system) inside an optimization loop. \subsection{$\htwo$ approximation}% \label{sub:ourmethod_htwo} In order to obtain a finite $\htwo$ error, the polynomial parts of the FOM and the ROM transfer function must be equal. When using projection-based methods on ODE models this \emph{feedthrough matching} is automatic. In the DAE case, this is typically obtained by preserving the algebraic part, i.\,e., by including the null-space of the $E$-matrix in the projection matrices. For systems with multiple algebraic constraints this is undesirable, and a reduction of the subsystem corresponding to the algebraic constraints may be necessary, see, for instance, \cite{MehS05}. Another remedy (used for pH-DAEs with index one in \cite{BeaGM21}) is to compute the polynomial part $\pHtf_{\rm p}(s) \equiv \mathcal{D}_0$ of the FOM before the reduction and include it in the feedthrough terms $\rS$ and $\rN$ of the ROM, which we will use here as well. The direct computation of $\mathcal{D}_0$ (see Section \ref{sec:preliminaries}) may, however, require transformations of the FOM. Alternatively, $\mathcal{D}_0$ can also be estimated by sampling the transfer function of the FOM at sufficiently large $s\in\C$ in an iterative manner as proposed in \cite{Schwerdtner2020a}. We then decompose it in its symmetric and skew-symmetric part, respectively, i.\,e., \begin{align} S_0 &:= \frac{1}{2}\left(\mathcal{D}_0^\T+\mathcal{D}_0\right), \label{eq:S0} \\ N_0 &:= \frac{1}{2}\left(\mathcal{D}_0^\T-\mathcal{D}_0\right). \label{eq:N0} \end{align} If the ROM is parameterized as in Lemma~\ref{lem:PHParam} then the $\cH_2$ error ${\Vert \pHtf - \pHtfr(\cdot,\theta)\Vert}_{\cH_2}$ is only well-defined if we have that \begin{align}\label{eq:SN_match} \rS(\theta) &= S_0,\quad \rN(\theta) = N_0, \end{align} since otherwise $\pHtf-\pHtfr(\cdot,\theta) \notin \mathcal{RH}_{2}^{m \times m}$. Consequently, we first have to fix all parameters in $\theta_N$ such that ${\rN(\theta) = N_0}$. Since we indirectly parameterize $\rS(\theta)$ via $\theta_W$, we first analyze which parameters in $\theta_W$ have an impact on $\rS(\theta)$. For this, consider a separation of ${\theta_W \in \R^{n_W}}$ into ${\theta_W =: \mat{\theta_{W_1}^{\T}, \theta_{W_2}^{\T}}^\T}$, where ${\theta_{W_1} \in \R^{n_W - m(m+1)/2}}$ and ${\theta_{W_2} \in \R^{m(m+1)/2}}$. Then we can decompose \begin{align} \vtu(\theta_W) = \mat{\Xi_1 & \Xi_2 \\ 0 & \Xi_3}, \end{align} where the matrices $\Xi_1, \,\Xi_2$ depend only on $\theta_{W_1}$, and $\Xi_3$ depends only on~$\theta_{W_2}$. Consequently, ${\rS(\theta) = \Xi_3\Xi_3^\T}$ only depends on~$\theta_{W_2}$ and we can set $\theta_{W_2}$ such that ${\rS(\theta) = S_0}$. The remaining parameters $\theta_{W_1}$ may still be subject to optimization and it holds that ${W(\theta) \geq 0}$ for all ${\theta_{W_1} \in \R^{n_W - m(m+1)/2}}$. Consequently, for minimizing the $\cH_2$ error, the parameter vector which is subject to optimization reduces to $\theta := \begin{bmatrix} \theta_J^\T, \theta_{W_1}^\T, \theta_Q^\T, \theta_{\pG}^\T \end{bmatrix}^\T.$ \begin{rem} Note that $S_0 \geq 0$ always holds since the implicit pH-ODE~\eqref{eq:FOM_decpl} has the same transfer function $\pHtf$ as the original pH-DAE and is therefore positive real, i.\,e., we have that \begin{equation*} \Phi(\mathsf{i}\omega) := \pHtf(\mathsf{i}\omega) + \pHtf(\mathsf{i}\omega)^{\mathsf{H}} \geq 0, \end{equation*} for all $\omega \in \R$ and consequently, $\lim_{\omega\rightarrow\infty}\Phi(\mathsf{i}\omega) = 2S_0\geq 0$. \end{rem} Now we can formulate the $\cH_2$ optimization problem in the pole-residue framework originally proposed in \cite{Beattie2009a} for unstructured LTI systems and extended to pH-ODE systems in \cite{Moser2020}. Assume that $(\rJ(\theta)-\rR(\theta))\rQ(\theta)$ is diagonalizable and consider the spectral decomposition \begin{equation} \label{eq:red_evp} (\rJ(\theta)-\rR(\theta))\rQ(\theta) Z(\theta) = Z(\theta)\Lambda(\theta), \end{equation} where $\Lambda(\theta) = \text{diag}(\reig_1(\theta),\,...\,,\reig_r(\theta))$ and $Z(\theta)$ contains the right eigenvectors as columns. If the eigenvalues $\reig_i(\theta)~\in~\C$, $i=1,\,\ldots,\,r$ are simple, then the transfer function $\pHtfr(\cdot,\theta)$ may be represented by the partial fraction expansion ${\pHtfr(s,\theta) = \sum_{i=1}^{r} \frac{\rl_i(\theta)\rr_i(\theta)^\T}{s-\reig_i(\theta)}+\rS(\theta)-\rN(\theta)}$, where ${\rl_i(\theta),\,\rr_i(\theta)\in \C^m}$ with \begin{align*} \rl_i(\theta) &= (\rG(\theta)+\rP(\theta))^\T\rQ(\theta) Z(\theta) e_i, \\ \rr_i(\theta) &= (\rG(\theta)-\rP(\theta))^\T Z(\theta)^{-\T}e_i, \end{align*} and where $e_i$ denotes the $i$-th standard basis vector of $\R^r$. Assuming that $\eqref{eq:SN_match}$ holds, we have that \begin{equation} \begin{aligned} {\Vert \pHtf-\pHtfr(\cdot,\theta) \Vert}_{\cH_2}^2 = &{\Vert \pHtf_{\rm sp} \Vert}_{\cH_2}^2 \\ & - 2\sum_{i=1}^{r}\rl_i(\theta)^\T \pHtf_{\rm sp}(-\reig_i(\theta))\rr_i(\theta) \\ & + \sum_{j,k=1}^{r} \frac{\rl_j(\theta)^\T\rl_k(\theta)\rr_k(\theta)^\T\rr_j(\theta)}{-\reig_j(\theta)-\reig_k(\theta)}, \end{aligned} \end{equation} as shown in \cite[Theorem 2.1]{Beattie2009a}. Since ${\Vert \pHtf_{\rm sp} \Vert}_{\cH_2}^2$ does not depend on $\theta$, it can be neglected in the optimization. Consequently, we define the objective functional \begin{align*} \losst(\theta;\pHtf) &:={\Vert \pHtf-\pHtfr(\cdot,\theta) \Vert}_{\cH_2}^2 - {\Vert \pHtf_{\rm sp} \Vert}_{\cH_2}^2 \\ &:= \hat{\mathcal{F}}\Big(\big[\rl_1(\theta)^\T,\ldots ,\rl_r(\theta)^\T,\rr_1(\theta)^\T,\ldots\\ & \quad\quad\quad\quad \quad\rr_r(\theta)^\T,\reig_1(\theta)^\T,\ldots,\reig_r(\theta)^\T\big]^\T\Big) \\ &= (\hat{\mathcal{F}} \circ q)(\theta), \end{align*} where \begin{align*} q(\theta) &:= [\rl_1(\theta)^\T,\ldots,\rl_r(\theta)^\T,\rr_1(\theta)^\T,\ldots \\ & \quad\quad\quad\quad\rr_r(\theta)^\T,\reig_1(\theta),\ldots,\reig_r(\theta)]^\T \in \C^{n_q}. \end{align*} This functional can be evaluated efficiently because it only requires the solution of the reduced-order eigenvalue problem in \eqref{eq:red_evp} as well as $r$ evaluations of $\pHtf_{\rm sp}$ at $-\reig_i(\theta)$. The eigenvalues $\reig_i(\theta)$ and rank-one residues $\rl_i(\theta)\rr_i(\theta)^\T$ are functions of the parameter vector $\theta$. If $\bar{\theta} \in \R^{n_\theta}$ is chosen such that all eigenvalues are simple, then $\mathcal{F}$ is differentiable in a neighborhood of $\bar{\theta}$. Its derivative is obtained by applying the chain rule, i.\,e., with the differentiation operator $\text{D}$ we obtain \begin{equation*} \text{D}\losst(\bar{\theta}) = \left(\nabla \losst(\bar{\theta})\right)^\T = \text{D}\hat{\mathcal{F}}(q(\bar{\theta}))\cdot \text{D}q(\bar{\theta}), \end{equation*} with \begin{multline*} \text{D}\hat{\losst}(q(\bar{\theta})) = \left[\text{D}_{b_1}\hat{\losst}(q(\bar{\theta})),\ldots,\text{D}_{b_r}\hat{\losst}(q(\bar{\theta})), \ldots \right. \\ \left. \text{D}_{c_1}\hat{\losst}(q(\bar{\theta})),\ldots, \text{D}_{c_r}\hat{\losst}(q(\bar{\theta})), \ldots\right. \\ \left.\text{D}_{\lambda_1}\hat{\losst}(q(\bar{\theta})),\ldots,\text{D}_{\lambda_r}\hat{\losst}(q(\bar{\theta}))\right]\in\C^{1\times n_q}, \end{multline*} and \begin{align*} \text{D}q(\bar{\theta}) &= \left[\text{D}_{\theta_1}{q(\bar{\theta})},\ldots,\text{D}_{\theta_{n_\theta}}{q(\bar{\theta})}\right]\in\C^{n_q\times n_{\theta}}. \end{align*} For all $i=1,\,\ldots,\,r$ it holds that \begin{align*} \text{D}_{\rr_i} \hat{\losst}(q(\bar{\theta})) &= 2\rl_i(\bar{\theta})^\T\big(\pHtfr(-\reig_i(\bar{\theta}))-\pHtf(-\reig_i(\bar{\theta}))\big), \\ \text{D}_{\rl_i} \hat{\losst}(q(\bar{\theta})) &= 2\rr_i(\bar{\theta})^\T\big(\pHtfr(-\reig_i(\bar{\theta}))-\pHtf(-\reig_i(\bar{\theta}))\big)^\T, \\ \text{D}_{\lambda_i} \hat{\losst}(q(\bar{\theta})) &= -2\rl_i(\bar{\theta})^\T\big(\pHtfr^{\prime}(-\reig_i(\bar{\theta}))-\pHtf^{\prime}(-\reig_i(\bar{\theta}))\big)\rr_i(\bar{\theta}), \end{align*} and we refer to~\cite{Moser2020} for the differentiation of $\text{D} q$. \begin{rem} The partial derivatives in $\mathrm{D}q(\bar{\theta})$ may be computed efficiently with block-wise expressions. For instance, the derivative $\mathrm{D}_{\theta_G}\rl_i(\bar{\theta}) \in \C^{m \times r\cdot m}$ can be computed as \begin{align*} \text{D}_{\theta_G}\rl_i(\bar{\theta}) &= \mat{z_i(\bar{\theta})^\T\rQ(\bar{\theta}) & 0 & 0 \\ 0 & \ddots & 0 \\ 0 & 0 & z_i(\bar{\theta})^\T\rQ(\bar{\theta})} \\ &= I_m \otimes z_i(\bar{\theta})^\T\rQ(\bar{\theta}), \end{align*} where $z_i(\bar{\theta})\in\C^r$ denotes the $i$-th column in $Z(\bar{\theta})$. This is also the case for more complex derivatives that involve the differentiation of the eigenvalue problem in \eqref{eq:red_evp}. \end{rem} Here, we highlight some important advantages of the pole-residue framework compared to recently proposed methods that are formulated in the Lyapunov framework (see~\cite{Sato2018,Jiang2019}), in particular for pH-DAEs. These methods require the solution of large-scale Lyapunov equations for the evaluation of $\losst$ and its gradient. Currently no structure-preserving Lyapunov-based methods exist for pH-DAEs; see \cite{SchM22_ppt} for new Lyapunov-based formulations of pH-DAEs. If the strictly proper part of the transfer function can be easily decoupled from the constant polynomial part, then for pH-DAEs with index one, the existing methods for pH-ODEs may be applied to this part. However, if the splitting into the strictly proper and polynomial part has first to be computed via a factorization method, then the sparsity patterns of the original pH-DAE may be lost which complicates the repetitive solution of Lyapunov equations for these systems in the large-scale setting. We highlight that the pole-residue framework only requires evaluations of $\pHtf_{\rm sp}$. Since we have that \begin{equation} \pHtf_{\rm sp}(s) = \pHtf(s) - (S_0-N_0), \end{equation} for all $s\in\C$, we may work directly with the sparse matrices of the original pH-DAE and do not require the solution of large-scale Lyapunov equations. \begin{algorithm}[tbh] \LinesNumbered \SetAlgoLined \DontPrintSemicolon \SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output} \Input{FOM transfer function $\pHtf\in\mathcal{RH}_{\infty}^{\enu \times \enu}$, reduced order $r\in\mathbb{N}$. } \Output{Reduced pH-ODE of order $r$} Compute $S_0,\,N_0$ as in \eqref{eq:S0}--\eqref{eq:N0}.\; Initialize $\theta_0$ s.t. $\rS(\theta_0)=S_0$, $\rN(\theta_0)=N_0$.\; Solve \vspace{-12pt}\begin{align*} \theta_{\rm fin} &= \argmin\limits_{\substack{\theta\in \R^{n_\theta}}} \losst(\theta; \pHtf) \\ &\text{s.t. } \rS(\theta)=S_0,\, \rN(\theta)=N_0. \end{align*}\vspace{-18pt} \; Construct the ROM with $\theta_{\rm fin}$ as in Lemma \ref{lem:PHParam}.\; \caption{PROPT-$\mathcal{H}_2$} \label{alg:H2OPT} \end{algorithm} Since the $\cH_2$ optimization problem is non-convex, the choice of the initial parameter vector $\theta_0$ will generally impact the fidelity of the final ROM obtained by Algorithm~\ref{alg:H2OPT}. Simple initialization strategies are, for instance, choosing $\theta_0$ randomly or using IRKA-PH (see~\cite{Moser2020, Sato2018}), which generally converges very quickly. Here, we propose another approach that may use \emph{unstructured} ROMs for initialization which is based on the following parameterization. \begin{lem}\label{lem:pHInit} Let $(\widetilde{A},\widetilde{B},\widetilde{C},\widetilde{D})$ be a ROM of state-space dimension $r$ such that $\widetilde{D} = S_0-N_0$ and such that $\widetilde{A}$ has all its eigenvalues in the open left half of the complex plane. Let $\theta_G \in\R^{r\cdot m}$ and $\theta_K \in\R^{r\cdot p}$ be two parameter vectors and define the matrix-valued functions \begin{align*} \iG(\theta_G) &:= \vtf_m(\theta_G), \\ \rK(\theta_K) &:= \vtf_q(\theta_K). \end{align*} Let $\iQ(\theta_K)>0$ solve the Lyapunov equation \begin{equation} \widetilde{A}^\T\iQ(\theta_K)+\iQ(\theta_K)\widetilde{A}+\rK(\theta_K)\rK(\theta_K)^\T = 0, \end{equation} and define \begin{align*} \iJ(\theta_K)&=\frac{1}{2}\left(\widetilde{A} \iQ(\theta_K)^{-1}-\iQ(\theta_K)^{-1}\widetilde{A}^\T\right), \\ \iR(\theta_K)&=-\frac{1}{2}\left(\widetilde{A} \iQ(\theta_K)^{-1}+\iQ(\theta_K)^{-1}\widetilde{A}^\T\right). \end{align*} Then the parametric system \begin{align} \label{eq:pHParamInit} \pHsysr(\theta_G,\theta_K): \begin{cases} \!\begin{aligned} \dot x_r(t)= &\big(\iJ(\theta_K)-\iR(\theta_K))\iQ(\theta_K\big)\rx(t) \\ &\quad \quad + \iG(\theta_G) u(t), \end{aligned}\\ \!\begin{aligned} \ry(t)= &\iG(\theta_G)^\T \iQ(\theta_K) \rx(t)\\ &\quad \quad+ (S_0-N_0) u(t) \end{aligned}\\ \end{cases} \end{align} is a pH-ODE system with \begin{equation*} \big(\iJ(\theta_K)-\iR(\theta_K)\big)\iQ(\theta_K) = \widetilde{A}. \end{equation*} \end{lem} Let $\widetilde{H}$ denote the transfer function of the (possibly unstructured) ROM $\big(\widetilde{A},\widetilde{B},\widetilde{C},\widetilde{D}\big)$ with ${\widetilde{\pHtf}(s)=\sum\limits_{i=1}^{r} \frac{\widetilde{\rl}_i\widetilde{\rr}_i^\T}{s-\widetilde{\reig}_i}+S_0-N_0}$ and ${\widetilde{\rl}_i,\,\widetilde{\rr}_i\in \C^m}$. Based on the parameterization in Lemma \ref{lem:pHInit}, we can then compute an initial pH model by minimizing the weighted sum of squared errors between the residuals in the Frobenius norm, i.\,e., \begin{equation*} \losst_0(\theta_G,\theta_K) := \sum_{i=1}^{r} \frac{1}{|\widetilde{\reig}_i|} \left\Vert \widetilde{\rl}_i\widetilde{\rr}_i^\T - \rl_i(\theta_G,\theta_K)\rr_i(\theta_G,\theta_K)^\T\right\Vert_{\rm F}^2, \end{equation*} where \begin{align*} \rl_i(\theta_G,\theta_K) &= \iG(\theta_G)^\T\iQ(\theta_K) \widetilde{Z} e_i, \\ \rr_i(\theta_G,\theta_K) &= \iG(\theta_G)^\T \widetilde{Z}^{-\T}e_i, \end{align*} for $i=1,\,\ldots,\,r$ and $\widetilde{Z}$ is, again under a diagonalizability assumption, obtained from the spectral decomposition \begin{equation*} \widetilde{A} \widetilde{Z} = \widetilde{Z}\widetilde{\Lambda}, \end{equation*} with $\widetilde{\Lambda} = \text{diag}(\widetilde{\reig}_1,\,...\,,\widetilde{\reig}_r)$. Note that the computation of the gradient of $\losst_0$ is very simple, since it does not involve a differentiation of the eigenvalues or eigenvectors. While the partial gradients of $\rl_i(\cdot)$ and $\rr_i(\cdot)$ with respect to $\theta_G$ are straightforward, the partial gradients of $\iQ(\cdot)$ with respect to the $l$-th entry in $\theta_K$ is the solution of the (reduced-order) Lyapunov equation \begin{multline*} \widetilde{A}^\T \frac{\partial\iQ(\theta_K)}{\partial \theta_{K,l}} + \frac{\partial\iQ(\theta_K)}{\partial \theta_{K,l}}\widetilde{A} \\ + \vtf_q(e_l)\rK(\theta_K)^\T+\rK(\theta_K)\vtf_q(e_l)^\T= 0, \end{multline*} where $e_l$ denotes the $l$-th standard basis vector of $\R^{r\cdot p}$. As the number of optimization parameters is reduced to $r(p+m)$, this initialization generally converges very quickly. In combination with Algorithm~\ref{alg:H2OPT}, this enables a \emph{two-step} approach with a more restrictive (yet simpler) pre-optimization of only the residuals and a subsequent (more complex) optimization of all system matrices. \begin{rem} Note that the sample-based SOBMOR method can be tuned to compute a ROM with small $\htwo$ error as well. Instead of using $\loss$ in conjunction with the bisection method outlined in Algorithm~\ref{alg:bisection}, the integral in~\eqref{eq:htwonorm} can be approximated by means of an adaptive quadrature rule (see \cite[Algorithm~1]{Gonnet2010} for a template method). In this way, it is possible to compute the $\htwo$ error and its gradient with respect to the free ROM parameters in terms of the error transfer function at specific sample points. This makes it possible to use the same optimization techniques as in~\cite{Schwerdtner2020}. In particular, \cite[Theorem~3.1]{Schwerdtner2020} for the gradient computation can be reused. We denote this method by SOBMOR-$\htwo$. Further details regarding the implementation of the adaptive integration are provided in the Appendix. \end{rem}

https://arxiv.org/abs/1906.09830

A story of balls, randomness and PDEs

Several differential equations usually appearing in mathematical physics are solved through a power series expansion, which reduces in solving difference equations. In this paper a probability problem is presented whose solution follows a completely reversed but systematic approach. Hence, this work is about illustrating how complex probability problems could be tackled with the more powerful techniques of a better studied and well understood field, that of differential equations. The problem is defined as follows: Inside a box containing r red and w white balls random removals occur. The balls are removed successively according to the three following rules. Rule I: If a white ball is chosen it is immediately discarded. If a red ball is chosen, it is placed back into the box and a new ball is randomly chosen. The second ball is then removed irrespective of the color. Rule II: Once one ball is removed, the game continues from Rule I. Rule III: The game ends once all the red balls are removed. The question posed is the determination of the probability that k white balls remain where k = 0, 1, 2, ..., w. Ending the game once all the white balls are removed, a second question is the determination of the probability that k red balls remain where k = 0, 1, 2, ..., r. While inductive solutions are possible, the current approach demonstrates a different and algorithmic route. In particular, the law of total probability yields a recursion that is transformed into a linear inhomogeneous 2D PDE, with suitable boundary conditions. The PDE solutions, which are found analytically, provide the generating functionals of the required probabilities as a function of r, w and k. Using the functionals, the probability formulas for any r, w and k are finally obtained in a closed form. Reproducing existing results of the literature this method is quite generic and adaptable to a large class of problems.

\section*{Disclaimer} \begin{document} \title{{\bf A story of balls, randomness and PDEs \\[1.5cm] }} \author{ \vspace{-1.2in} {\bf Anastasios Taliotis\thanks{e-mail: [email protected]} \thanks{Address: JPMorgan Chase, 25 Bank Street, Canary Wharf, London E14 5JP, UK.} \thanks{Disclaimer: \hspace{-0.04cm}The current \hspace{-0.04cm}paper \hspace{-0.04cm}consists of a personal \hspace{-0.04cm}work curried out by A. \hspace{-0.04cm}Taliotis, \hspace{-0.04cm}and it expresses \hspace{-0.02cm}his \hspace{-0.02cm}own personal views on the particular topic. In particular, this work is not, in any way, endorsed by or related with JPMorgan Chase (the firm) or its employees or stakeholders or any interests or entities represented by the firm.} \thanks{Keywords: Probability, recursions, generating functional, PDEs, Laurent expansion, residues, Python code.}} } \providecommand{\keywords}[1]{\textbf{\textit{Keywords:}} #1} \vspace{-3in} \maketitle \thispagestyle{empty} \begin{abstract}\hspace{-0.6cm} Several differential equations usually appearing in mathematical physics are solved through a power series expansion using the Frobenius method, which reduces in solving difference equations (recursions). In this paper, a probability problem is presented whose solution follows a completely reversed but systematic approach. Hence, this work is about illustrating how complex probability problems involving difference equations could be tackled with the more powerful techniques of a better studied and well understood field, that of differential equations. The problem is defined as follows: Inside a box containing $r$ red and $w$ white balls \hspace{-0.02cm}random\hspace{-0.02cm} removals \hspace{-0.04cm}occur. \hspace{-0.04cm}The balls are removed \hspace{-0.04cm}successively \hspace{-0.04cm}according to \hspace{-0.55cm}the three following rules. Rule I: If a white ball is chosen it is immediately discarded from the box. If a red ball is chosen, it is placed back into the box and a new ball is randomly chosen. The second ball is then removed irrespective of the color. Rule II: Once one ball is removed, the game continues from Rule I. Rule III: The game ends once all the red balls are removed. The first question posed is the determination of the probability that $k$ white balls remain where $k=0,\, 1,\, 2,\, ...,\, w$. Ending the game once all the white balls are removed, a second question is the determination of the probability that k red balls remain where $k=0,\, 1,\, 2,\, ...,\, r$. While inductive solutions are possible, the\hspace{-0.04cm} current approach demonstrates a different and algorithmic route. In particular, the law of total probability yields a recursive equation that is transformed into a linear PDE in two dimensions with inhomogeneous source terms, and suitable boundary conditions that depend on $k$. The PDE solutions, which are found analytically, provide up to a known rescaling factor, the (two) generating functionals of the required probabilities as a function of $r$, $w$ and $k$. Using then the derived functionals, the required probability formulas for any $r$, $w$ and $k$ are finally obtained in a closed form; the probability distributions turn out to be linear combinations of hypergeometric functions of type $_3F_2$. Reproducing existing results of the academic literature, which are special but less involved cases of the current completely generic solution, this method is quite generic and adaptable to a large class of problems. \end{abstract} \thispagestyle{empty} \newpage \tableofcontents \setcounter{page}{1} \section{Introduction} During interviews for quantitative researchers in investment banks, hedge funds, asset management firms, fin-tech companies etc, the candidates are often requested to solve various probability problems. The current problem is motivated by such interviews and is a more complicated version of these kind of questions. \vspace{0.2cm} \hspace{-0.6cm}{\bf Problem 1} \vspace{0.2cm} \hspace{-0.7cm} {\it Inside a box containing $r$ red and $w$ white balls random removals occur. The balls are removed one by one according to the three following rules. Rule I: If a white ball is chosen it is immediately discarded from the box. If a red ball is chosen, it is placed back into the box and a new ball is randomly chosen. The second ball is then removed irrespective of the color. Rule II: Once a ball is removed, the game continues from Rule I. Rule III: The game ends once all the red balls are removed. The question posed is the determination of the probability that $k$ white balls remain where $k=0,\, 1,\, 2,\, ...,\, w$. Changing rule Rule III into Rule IV: The game ends once all the white balls are removed. The second question posed is the determination of the probability that $k$ red balls remain where $k=0,\, 1,\, 2,\, ...,\, r$.} \vspace{0.2cm} This paper in not only about providing the solution to Problem 1, as, according to appendix \ref{Gminus}, a combinatorial-inductive solution\footnote{We are particularly grateful to D. Christofides for providing us with the solution through such an approach.} is (also) possible. This paper is rather about illustrating how difficult problems involving difference equations usually appearing in probability questions, could be tackled systematically with the more powerful techniques of a better studied and well understood field, that of differential equations. As it is known, differential equations appearing in mathematical physics, such as Bessel functions, Hypergeometric functions, Hermite polynomials, to name a few, are solved perturbatively using the Frobenius method. During the course of the solution, recursive relations that connect higher order coefficients with lower order ones appear; that is is difference equations. In this problem, we work in the reversed order following an algorithmic recipe explained below and summarized in five steps in section \ref{diss}. That is starting from a probability problem, which is reduced to a linear difference equation in two variables, we deduce an inhomogeneous linear partial differential equation (PDE) with suitable boundary conditions that we solve analytically. The solution, up to a known constant, provides the probability generating functional of the problem at hand. Expanding then the functional into a Laurent series, we explicitly obtain the probability solutions in closed form, which, as we will compute through our step by step systematic approach, are given by equations (\ref{finallypk}) and (\ref{finallypk4}) for the Rule III and the Rule IV respectively, and are reproduced below. In particular, the probability $p^{(k)}(r,w)$, where $k$ is the number of the white or red remaining balls, and $r$ and $w$ are the red and the white initial balls inside the box respectively, is given by \begin{equation*} \boxed{ \begin{aligned} p_{III}^{(k)}(r,w) &=k!\frac{ r!(r+w+1)}{(r+k+1)!}\,\, \frac{rw!(r+w-k-1)!}{(r+w)!(w-k)!} \mbox{, Rule III, }\\ p_{IV}^{(k)}(r,w) &=(2k+1)\frac{ r!(r+w+k)!}{(r+k+1)!(w+r)!}\,\, \frac{wr!(r+w-k-1)!}{(r+w)!(r-k)!}\mbox{, Rule IV. } \end{aligned} } \end{equation*} Satisfying the normalization conditions \begin{equation*} \boxed{ \begin{aligned} \sum_{k=0}^wp_{III}^{(k)}(r,w) =1,\,\,\, \sum_{k=0}^rp_{IV}^{(k)}(r,w) =1, \end{aligned} } \end{equation*} the two probability formulas consist of the main result of this paper. The proposed approach combines several beautiful branches of mathematics: probability theory, differential equations, special functions and peripherally, through analytic continuation, residues and Laurent expansions of functions, complex analysis. Most of the steps between equations are either done explicitly or explained in detail. Therefore, being self contained, this paper can be served as an introductory set of lecture notes for advanced undergraduate and graduate students with mathematical background that are interested in applications of the aforementioned fields of mathematics. For a quick but yet a detailed exposition of the reader to the main ideas of this work, it suffices to restrict to sections \ref{2.1}, \ref{pdeRes} and \ref{special}. \vspace{0.1in} This work is organized as follows. \vspace{0.1in} In section \ref{toy}, using $z$-transformations in two dimensions, we solve an easier problem, Problem 2, that involves algebraic but not differential equations. This serves as an introduction to Problem 1 and to the subsequent sections, and clarifies the formalism and the notation. \vspace{0.1in} In section \ref{pde} we return to the original problem where using the law of total probability and $z$-transformations, we eventually obtain the required PDE (of Problem 1). The boundary conditions are discussed in the subsequent sections. \vspace{0.1in} Section \ref{special} deals with Problem 3, which is a special case of Problem 1, and whose result is known in the community \cite{MuL}. Hence, we cross-check our method against known cases. \vspace{0.1in} Returning to Problem 1, Rule III, section \ref{genPDE} discusses the boundary conditions and derives the generic solution of the PDE for any remaining white balls $k=0,\,1,\,2\,...,\,w$. \vspace{0.1in} Using the solutions of the PDE from section \ref{genPDE}, the first part of section \ref{result} derives the explicit probability formulas for the cases $k=0,\,1,\,2$ (Rule III). Given these solutions and by observing the pattern (there exists an even more systematic approach), the remaining section derives the generic probabilities $p_{III}^{(k)}(r,w) $ as a function of the initial balls $r$ and $w$ and of the remaining white balls $k$, which consists the one of the two main results of this paper. The probabilities $p_{III}^{(k)}(r,w) $ are subsequently being investigated. \vspace{0.1in} Armed with the experience from earlier sections and in particular from section \ref{result}, section \ref{result4} outlines and eventually provides the second main result of this paper, namely the formula of $p_{IV}^{(k)}(r,w) $ for Problem 1, Rule IV. This section is much shorter than sections \ref{genPDE} and \ref{result}. \vspace{0.1in} The last section, section \ref{diss}, concludes by discussing the advantages of our approach. \vspace{0.1in} Appendices \ref{A}, \ref{B}, \ref{C} and \ref{D} are reserved for longer calculations and for proving several useful identities. Appendix \ref{Z} discusses alternative boundary conditions for Problem 1, Rule III. Appendix \ref{Gminus} provides the solution to Problem 1, Rule III through a combinatorial-inducive approach. On the other hand, appendix \ref{E} documents a different and hard question posed in \cite{OaPe}, which serves as another reference for these kind of problems. Lastly, appendix \ref{F} provides a python code that simulates Problem 1 and provides the results for a few cases contrasting them against the two analytically obtained formulas. \vspace{0.1in} \underline{Notation}: \vspace{0.1in} We use lower case latin letters for the remaining balls inside the box. In particular, we use the letter $r$ for the red balls and the letter $w$ for the white balls that the box contains at the beginning of the games. The letter $k$ is used for the remaining white or red balls (for Rule III and for Rule IV respectively) after the game ends. We use other lower case latin letters except from $k$, $r$ and $w$ in order to sum over other indices that may appear along the way. Finally, during the $z-$transformations, the letters $z$ and $u$ are associated with the red and the white balls respectively. We define the Kronecker delta by \begin{align} \delta_j^i = \mbox{$1$ if $i=j$ $(i,\,j\,=\,\mathbb{Z})$ and $0$ otherwise}. \end{align} The following definitions apply for Problem 1, Rule III (and for Problem 2). \begin{subequations} \label{pa2e} \begin{align} \label{pa} p_{III}^{(k)}(r,w) \equiv & \mbox{ The probability to remain with $k$ white balls} \\ \nonumber &\mbox{ starting with r red and w white balls} \\ \nonumber \equiv &\lim_{r'\to\ r} \left(\lim_{w'\to\ w}\left(\lim_{k'\to\ k} \left(p_{III}^{(k')}(r',w') \right)\right)\right), \\ \label{pb} p_{III}^{(k)}(0,w) \equiv &\,\delta_w^k,\\ \label{pc} p_{III}^{(0)}(0,0) \equiv &1\,,\\ \label{pbb} p_{III}^{(k)}(r\geq0,0) \equiv & \, \delta_0^k,\\% \lim_{r'\to\ r} \left(\lim_{w\to\ 0} \left(p^{(k)}(r',w) \right)\right), \,(r,k)\neq(0,0),\\ \label{pe} p_{III}^{(k)}(r,w<k) \equiv & \, 0. \end{align} \end{subequations} Equation (\ref{pc}) is a special case of (\ref{pbb}) but we make it explicit. What it should be emphasized is that equations (\ref{pb})-(\ref{pe}) should be consistent with equation (\ref{pa}); this can be checked explicitly once the formula for $p_{III}^{(k)}(r,w)$ is obtained. In particular, as we will see in section \ref{genPDE}, either equation (\ref{pb}) or equation (\ref{pbb}), under particular modifications that we will derive in what follows, should be used as the boundary conditions for the continuum version of the problem. The motivation behind the specification of the ordering of the limits in equation (\ref{pa}) is because we would like to write compact expressions for the final probability formulas by giving a meaning to subtle expressions such as $w/(w+r)$ or $r/(w+r)$ when $w=r=0$. Another example would be $(r+w-3)!/(r-1)!$ for $r=0$ and $w=2$ and so forth. The particular ordering of the limits implies that for the base case $r=0$ and $w=0$ (the game starts without any red balls, i.e. the game ends before it begins), the probability $p_{III}^{(k)}(r=0,w=0)$ is, by our limit convention, equal to one when $k=0$ and zero otherwise. To rephrase, the ordering of the limits essentially defines equations (\ref{pb}) and (\ref{pbb}) and hence, according to the discussion of the previous paragraph, such an ordering defines the boundary conditions of the problem, which are necessary for the uniqueness of the solution. We will see explicitly the necessity of the orderings of the limits and the applications of (\ref{pb})-(\ref{pe}) in what follows. For Problem 1, Rule IV, the roles of $r$ and $w$ are interchanged. In particular, we have \begin{subequations} \label{pa2e4} \begin{align} \label{pa4} p_{IV}^{(k)}(r,w) \equiv & \mbox{ The probability to remain with $k$ red balls} \\ \nonumber &\mbox{ starting with r red and w white balls} \\ \nonumber \equiv &\lim_{w'\to\ w} \left(\lim_{r'\to\ r}\left(\lim_{k'\to\ k} \left(p_{IV}^{(k')}(r',w') \right)\right)\right), \\ \label{pb4} p_{IV}^{(k)}(r,0) \equiv &\,\delta_r^k,\\ \label{pc4} p_{IV}^{(0)}(0,0) \equiv &1\,,\\ \label{pbb4} p_{IV}^{(k)}(0,w\geq0) \equiv & \, \delta_0^k,\\% \lim_{r'\to\ r} \left(\lim_{w\to\ 0} \left(p^{(k)}(r',w) \right)\right), \,(r,k)\neq(0,0),\\ \label{pe4} p_{IV}^{(k)}(r<k,w) \equiv & \, 0. \end{align} \end{subequations} Analogous comments apply here as the comments below equation (\ref{pa2e}). \section{A toy model}\label{toy} The following problem is given. \vspace{0.2cm} \hspace{-0.6cm} {\bf Problem 2} \vspace{0.2cm} \hspace{-0.7cm} {\it Inside a box containing $r$ red and $w$ white balls random removals occur. The balls are removed randomly one by one. The game ends once all the red balls are removed. The question posed is the determination of the probability that $k$ white balls remain where $k=0,\, 1,\, 2,\, ..., w$.} \vspace{0.2cm} \subsection{Solution through $z$-transformations}\label{2.1} \vspace{0.2cm} Using the law of total probability by conditioning on the first removal, the following recursive relation is obtained \begin{align}\label{rec1} p^{(k)}(r,w) = \frac{r}{r+w}p^{(k)}(r-1,w) + \frac{w}{r+w}p^{(k)}(r,w-1). \end{align} We now make the educated ansatz \begin{align}\label{ans1} p^{(k)}(r,w) = \frac{r!w!}{(r+w)!}f^{(k)}(r,w) \end{align} obtaining the much simpler equation for $f^{(k)}$ \begin{align}\label{rec2} f^{(k)}(r,w) = f^{(k)}(r-1,w) + f^{(k)}(r,w-1). \end{align} This ansatz is motivated because we want to factor out $\frac{r!w!}{(r+w)!}$, which provides the probability to remain with $k=w$ white balls; that is if we keep removing only red balls (equivalently only white balls). In the next step, we multiply both sides of (\ref{rec2}) by $z^{-r}$ and $u^{-w}$ and sum over $r$ and $w$ from one to infinity. Shifting some indices around and adding and subtracting suitable terms, we eventually obtain \begin{align}\label{sum1} &\sum_{r=0}^{\infty} \sum_{w=0} ^{\infty }f^{(k)}_{r,w} z^{-r} u^{-w} - \sum_{w=0} ^{\infty }f^{(k)}_{0,w} u^{-w} - \sum_{r=0} ^{\infty }f^{(k)}_{r,0} z^{-r} + f_{0,0} =\\ \nonumber & \frac{1}{z} \left( \sum_{r=0}^{\infty} \sum_{w=0} ^{\infty }f^{(k)}_{r,w} z^{-r} u^{-w} - \sum_{r=0} ^{\infty }f^{(k)}_{r,0} z^{-r} \right) + \frac{1}{u} \left( \sum_{r=0}^{\infty} \sum_{w=0} ^{\infty }f^{(k)}_{r,w} z^{-r} u^{-w} - \sum_{w=0} ^{\infty }f^{(k)}_{0,w} u^{-w} \right) \end{align} where we simplified the notation on $f^{(k)}$ setting $f^{(k)}(r,w)=f^{(k)}_{r,w}$. Defining the generating functional $Y^{(k)}$ by \begin{align}\label{Yfdef} Y^{(k)}(z,u) = \sum_{r=0}^{\infty} \sum_{w=0} ^{\infty }f^{(k)}_{r,w} z^{-r} u^{-w} \end{align} equation (\ref{sum1}) yields \begin{align}\label{Y1} Y^{(k)}(z,u) = \frac{1}{zu-z-u} \left( u(z-1) \sum_{r=0} ^{\infty }f^{(k)}_{r,0} z^{-r} + z(u-1) \sum_{w=0} ^{\infty }f^{(k)}_{0,w} u^{-w} - z uf^{(k)}_{0,0} \right). \end{align} \vspace{0.2cm} \underline{ Case $k>0$} \vspace{0.2cm} In this case, most of the terms in the right hand side of (\ref{Y1}) are zero. In particular, $f^{(k>0)}_{r,0}=0 \, \forall \,r \geq 0$ because the probability to remain with a positive number of white balls $k$ starting without any white balls at all must be zero. Another way to see it is through equation (\ref{pbb}). Also in the view of (\ref{pb}) all $f^{(k)}_{0,w}$'s are zero except from the term $f^{(k)}_{0,k}$, which is equal to one. Hence, \begin{align}\label{Y1k} Y^{(k>0)}(z,u) = \frac{z(u-1)}{zu-z-u} \times u^{-k} \mbox{ (because $f^{(k>0)}_{r,0}=0\, \,\forall \,r $ and $f^{(k)}_{0,w}=\delta^k_w$).} \end{align} The $f^{(k)}_{r,w}$'s are then obtained through a Laurent expansion of (\ref{Y1k}) in inverse powers of $z$ and $u$. In fact, the expansion can be derived from the expansion of $z(u-1)/(zu-z-u)$ followed by a shifting of the $w$ index of $u$ by $k$ due to the overall $u^{-k}$ factor in (\ref{Y1k}). Hence, \begin{align} \label{expand1} \frac{z(u-1)}{zu-z-u} & = \left(1-\frac{1}{u} \right)\frac{1}{1-\frac{1}{z}-\frac{1}{u}} = \left(1-\frac{1}{u} \right) \sum_{r=0}^{\infty} \sum_{w=0} ^{\infty } \frac{(r+w)!}{r!w!}\frac{1}{z^r}\frac{1}{u^w} \\ \nonumber &= \sum_{r=0}^{\infty} \sum_{w=0} ^{\infty } \frac{(r+w)!}{r!w!}\frac{1}{z^r}\frac{1}{u^w} - \sum_{r=0}^{\infty} \sum_{w=1} ^{\infty } \frac{(r+w-1)!}{r!(w-1)!}\frac{1}{z^r}\frac{1}{u^w}\\ \nonumber &= \sum_{r=0}^{\infty} \sum_{w=0} ^{\infty } \left(\frac{(r+w)!}{r!w!} - \frac{(r+w-1)!}{r!(w-1)!}\right) \frac{1}{z^r}\frac{1}{u^w} \left(\frac{\mbox{(i) assuming } (-1)!=\Gamma(0)=\infty,} {\mbox{\hspace{-0.8in}(ii) considering (\ref{pa}).} }\right)\\ \nonumber & = \sum_{r=0}^{\infty} \sum_{w=0} ^{\infty } \frac{r(r+w-1)!}{r!w!} \frac{1}{z^r}\frac{1}{u^w} \end{align} where in the third line we have extended the summation from $w=0$ because the denominator of the second term in the bracket vanishes assuming the factorial is analytically continued to a Gamma function such that $n!=\Gamma(n+1)$. Also the term $w=r=0$ in the sum of the last equality should be understood, in the view of (\ref{pb}) and (\ref{pc}), as taking the limit $w\rightarrow 0$ first followed by $r\rightarrow 0$ yielding a unit coefficient. Generally, the previous arguments employ the ordering of the limits defined in (\ref{pa}) (see also the discussion below (\ref{pa2e})). Combining now (\ref{expand1}) with (\ref{Y1k}) we obtain \begin{align}\label{Y1kfinal} Y^{(k)}(z,u) = \sum_{r=0}^{\infty} \sum_{w=k} ^{\infty } \frac{r(r+w-k-1)!}{r!(w-k)!} \frac{1}{z^r}\frac{1}{u^w}= \sum_{r=0}^{\infty} \sum_{w=0} ^{\infty } \frac{r(r+w-k-1)!}{r!(w-k)!} \frac{1}{z^r}\frac{1}{u^w} \end{align} where the in the second equality we have again extended the sum from $w=0$ in the view of the fact that $\Gamma(-n)=\infty,$ when $n\,=\,0,\,1,\,2,\,...\,.$ From (\ref{Y1kfinal}) and (\ref{Yfdef}) we identify the coefficients $f^{(k)}_{r,w}$. Using then (\ref{ans1}) we obtain the desired probability formula \begin{align} \label{p1old} p^{(k)}(r,w) = \frac{r (r+w-k-1)!w!}{(r+w)!(w-k)!},\, \forall\, r,\,w\geq 0,\, k>0. \end{align} \vspace{0.2cm} \underline{ Case $k= 0$} \vspace{0.2cm} In this case, the non-zero terms in the right hand side of (\ref{Y1}) are the $f^{(k=0)}_{r,0} \,\forall \,r \geq 0$ including the $f^{(k=0)}_{0,0}$ term. In fact, all of these terms are equal to one in the view of (\ref{pbb}); they represent the probability to remain with $k=0$ white balls when we start without any white balls at all and this should happen with probability equal to one. The fact that $f^{(k=0)}_{r,0}=\delta_0^0=1$ has already taken into account the rescaling pre-factor of (\ref{ans1}), which also reduces to one when $w=0$. Resumming then the right hand side of (\ref{Y1}) and simplifying yields \begin{align}\label{Y10} Y^{(k=0)}(z,u) = \frac{z(u-1)}{zu-z-u} \mbox{ (because $f^{(k=0)}_{r,0}=1\, \,\forall \,r $ and $f^{(k=0)}_{0,w}=\delta^0_w$).} \end{align} Comparing the right hand side of (\ref{Y10}) with that of (\ref{Y1k}) we realize that the probabilities for $k=0$ can be obtained from (\ref{p1old}) by taking the limit $k \to 0$. To conclude, the general solution is given by \begin{align} \label{p1} \boxed{p^{(k)}(r,w) = \frac{r (r+w-k-1)!w!}{(r+w)!(w-k)!}},\, \forall k,\,r,\,w \geq 0,\, k\leq w, \end{align} and it satisfies the normalization condition \begin{align} \label{normp2} \sum_{k=0}^wp^{(k)}(r,w) =1. \end{align} One may check that formula (\ref{p1}) satisfies (\ref{rec1}) and has the right behavior in several limiting cases. Indeed, $p^{(k)}(r,w<k)=0$ where we use the fact that the factorial of negative integer has a first order pole (tends to infinity). Also, in the view of (\ref{pa}), the probability formula yields $p^{(k)}(0,w)=\delta_w^k$. \subsection{A combinatorial derivation} In section \ref{2.1} we followed a rather complicated route in order to illustrate part of the formalism of Problem 1. However, Problem 2 may also be solved using simple combinatorics, which we present below for comparison. We begin by defining the events $A$ and $B$ where \begin{subequations} \begin{align} A\,\,= &\mbox{ the event to draw a red ball given that currently $k$ white and one red balls}\\ \nonumber & \mbox{ remained,}\\ B\,\,= &\mbox{ the event to draw $r-1$ red and $w-k$ white balls out of a population }\\ \nonumber &\mbox{ of $r$ red and $w$ white balls.} \end{align} \end{subequations} The following two key facts are emphasized: (i) The complement of the event $A$, $A^{\intercal}$, and the event to remain with $k$ white balls starting with $r$ reds and $w$ whites, are mutually exclusive. (ii) The event $B$ does not impose any constrains on the possible orderings where the $r-1$ red and the $w-k$ white balls could be chosen. With these two key facts at hand we apply the law of total probability by conditioning on the events $A$ and $A^{\intercal}$. Defining $P(A)$ and $P(B)$ as the probabilities for the occurrence of the events $A$ and $B$ respectively we obtain \begin{align}\label{p1comb} p^{(k)}(r,w)& = p^{(k)}(r,w|A^{\intercal})P(A^{\intercal}) + p^{(k)}(r,w|A)P(A) = 0 + p^{(k)}(r,w|A)P(A) \\ \nonumber & = p(B) \frac{1}{1+k} = \frac{\binom{r}{r-1} \binom{w}{w-k}}{\binom{r+w}{r+w-k-1}} \frac{1}{1+k} = \frac{r (r+w-k-1)!w!}{(r+w)!(w-k)!}, \end{align} which is identical to (\ref{p1}) and this completes the derivation. It is interesting to note that the probability of the event $B$, according to the second equality of the second line of (\ref{p1comb}), essentially follows a type $_2F_1$ hypergeometric distribution\footnote{We remind the reader that ``In probability theory and statistics, the hypergeometric distribution is a discrete probability distribution that describes the probability of $k$ successes (random draws for which the object drawn has a specified feature) in $n$ draws, without replacement, from a finite population of size $N$ that contains exactly $K$ objects with that feature, wherein each draw is either a success or a failure. If a random variable $X$ follows the hypergeometric distribution then its probability mass function is given by $p_X(k) = \frac{\binom{K}{k} \binom{N-K}{n-k} }{\binom{N}{n} }$" (Wikipedia).}. In problem $1$, as we will see further below in sections \ref{genPDE}, \ref{result} and \ref{result4}, the hypergeometric function is a key object. For later comparison with the results of Problem 1, we provide the probability generating functional of Problem 2 for fixed $r$ and $w$, which is given by \begin{align}\label{Gp2} G(r,w;z) \equiv \sum_{w=0}^kp^{(k)}(r,w)z^k = \frac{r}{r+w} \,_2F_1\left(1,-w,1-r-w ;z\right). \end{align} It is not a coincidence that equation (\ref{Gp2}) involves a type $_2F_1$ hypergeometric function. \section{Setting up the problem: From the law of total probability to the required PDE}\label{pde} We now return to Problem 1. Working as in previous section, we will derive a PDE rather than an algebraic equation for the analogous $Y^{{(k)}}$ we saw earlier. Using the law of total probability by conditioning on the first removal, we obtain \begin{align}\label{rec3} p^{(k)}(r,w) = \frac{r^2}{(r+w)^2}p^{(k)}(r-1,w) + \frac{w^2+2 r w}{(r+w)^2}p^{(k)}(r,w-1) \end{align} where the weighting probability of $p^{(k)}(r-1,w)$ is complement to the one of $p^{(k)}(r,w-1)$, and can be understood as $w/(w+r)+ rw/(r+w)^2$: the probability to remove a white ball equals the probability to either chose a white ball directly or to choose a red followed by a white ball. \subsection{The PDE of the generating functional with rescaling}\label{pdeRes} In this case too, we make an educated ansatz on (\ref{rec3}) as follows \begin{align}\label{ans2} p^{(k)}(r,w) = \left( \frac{r!w!}{(r+w)!}\right)^2 f^{(k)}(r,w) \end{align} obtaining the much simpler equation\footnote{Note added after the completion of this work: According to (\ref{psum}), the ansatz $p^{(k)}(r,w) = \frac{(r!)^2 k! w!}{((r+w)!)^2} f^{(k)}(r,w)$ yields an equation of the form (\ref{grec}), which is an even simpler equation for $ f^{(k)}(r,w)$ than equation (\ref{rec4}).} for $f^{(k)}$ \begin{align}\label{rec4} wf^{(k)}(r,w) = wf^{(k)}(r-1,w) + (2r+w)f^{(k)}(r,w-1). \end{align} This ansatz, as in Problem 2, is motivated because the expression $ (r!w!)^2/ ((r+w)!)^2$ provides the probability to remain with $k=w$ white balls; that is if we keep removing only red balls. Working then as in Problem 1 by multiplying both sides of (\ref{rec4}) by $z^{-r}$ and $u^{-w}$ and summing over $r$ and $w$ from one to infinity, after some algebra, we eventually obtain \begin{align}\label{sum2} \frac{1}{z}\sum_{r=0}^{\infty} \sum_{w=0} ^{\infty }wf^{(k)}_{r,w} z^{-r} u^{-w} &+\frac{1}{u}\sum_{r=0}^{\infty} \sum_{w=0} ^{\infty }(2r+w+1)f^{(k)}_{r,w} z^{-r} u^{-w} -\sum_{r=0}^{\infty} \sum_{w=0} ^{\infty }wf^{(k)}_{r,w} z^{-r} u^{-w} \\ \nonumber &= \frac{1}{u} \sum_{w=0} ^{\infty }(w+1)f^{(k)}_{0,w} u^{-w} -\sum_{w=0} ^{\infty }wf^{(k)}_{0,w} u^{-w}. \end{align} Defining $Y^{(k)}(z,u)$ as in (\ref{Yfdef}) and making the substitutions \begin{align}\label{partials} r\rightarrow -z\partial_z, \,\,\,\,\, w\rightarrow -u\partial_u \end{align} on the $r$ and the $w$ factors that multiply the quantity $f^{(k)}_{r,w} z^{-r} u^{-w}$, equation (\ref{sum2}) yields \begin{equation}\label{pdeG} \boxed{ \begin{aligned} -\left(\frac{u}{z}-u+1\right)\partial_uY^{(k)}(z,u)&-2\frac{z}{u}\partial_zY^{(k)}(z,u)+\frac{1}{u}Y^{(k)}(z,u) \\ &= \frac{1}{u} \sum_{w=0} ^{\infty }(w+1)f^{(k)}_{0,w} u^{-w} -\sum_{w=0} ^{\infty }wf^{(k)}_{0,w} u^{-w}. \end{aligned} } \end{equation} In what follows, we proceed using standard techniques of PDEs. In particular, we proceed using the fact that a first order PDE in two dimensions is equivalent to a system of two first order ODEs. In order to see that we define and solve the first ODE \begin{align}\label{dudz} \frac{du}{dz} = \frac{\frac{u}{z}-u+1}{2\frac{z}{u}} \implies u = \frac{z}{1+z+x_2\sqrt{z}} \end{align} where $x_2$ is the integration constant. The next step is to solve for $x_2=x_2(z,u)$ and make the change of variables \begin{subequations}\label{x12} \begin{align} \label{x1} x_1&= z,\\ \label{x2} x_2 &= \frac{z - u - z u}{\sqrt{z} u} = -\sqrt{z}\left(1-\frac{1}{u} + \frac{1}{z} \right) \end{align} \end{subequations} where we note that as $z,\,u\to \infty$ we have that $x_1\to \infty$ and $x_2 \to -\sqrt{z} \to -\infty$. In the final step, we apply the transformations (\ref{x12}) and the differential equation (\ref{pdeG}) becomes \begin{equation}\label{odeG} \boxed{ \hspace{-0.18cm}\begin{aligned} \frac{1+x_1+ x_2\sqrt{x_1}}{x_1} &\left (Y^{(k)}(x_1,x_2) -2x_1\partial_{x_1}Y^{(k)}(x_1,x_2) \right)=\\ \frac{1}{u} \sum_{w=0} ^{\infty }(w+1)f^{(k)}_{0,w} \left( u(x_1,x_2)\right) ^{-w} &-\sum_{w=0} ^{\infty }wf^{(k)}_{0,w} \left( u(x_1,x_2)\right)^{-w}, \,\,\,\, u=\frac{x_1}{1+x_1+x_2\sqrt{x_1}},\,\, z=x_1 \end{aligned} } \end{equation} where, as expected, the PDE reduces to a first order ODE. This is the second equation in the equivalent set of the two first order ODEs mentioned earlier. In the following sections, we will study (\ref{odeG}) in three different versions depending on the right hand side. The first version, section \ref{special}, corresponds to known results in the literature (Problem 3), the second version, sections \ref{genPDE} and \ref{result}, corresponds to the original problem, Problem 1, Rule III, while the third version corresponds to section \ref{result4}, Problem 1, Rule IV. \subsection{The PDE of the generating functional without rescaling}\label{pdeNoRes} For completeness, we present the PDE that corresponds to the initial recursion (\ref{rec3}) rather to (\ref{rec4}), and, which provides the actual probability generating functional. Working as before and defining \begin{align}\label{Yt} \tilde{Y}^{(k)}(z,u) = \sum_{r,w=0} ^{\infty } p^{(k)}(r,w)\frac{1}{z^r}\frac{1}{u^w}, \end{align} the PDE satisfied by $\tilde{Y}^{(k)}(z,u)$ is given by \begin{align}\label{pde21} \left(\frac{1}{z}+\frac{1}{u} \right) \tilde{Y}^{(k)} (z,u) &-\frac{1}{u} \left(u+z(2+u) \right) \partial_z \tilde{Y}^{(k)}(z,u) -\left(1+u \right) \partial_u\tilde{Y}^{(k)} (z,u)\\ \nonumber &-z\left(z-1 \right) \partial^2_{zz}\tilde{Y}^{(k)}(z,u) -2z\left(u-1 \right) \partial^2_{zu}\tilde{Y}^{(k)}(z,u) -u\left(u-1 \right) \partial^2_{uu}\tilde{Y}^{(k)} (z,u)\\ \nonumber &=\frac{1}{z}\sum_{r=0}^{\infty} (r+1)^2p^{(k)}(r,0)\frac{1}{z^r} +\frac{1}{u}\sum_{w=0}^{\infty} (w+1)^2p^{(k)}(0,w)\frac{1}{u^w}\\ \nonumber &-\sum_{r=0}^{\infty} r^2p^{(k)}(r,0)\frac{1}{z^r} -\sum_{w=0}^{\infty} w^2p^{(k)}(0,w)\frac{1}{u^w}. \end{align} We now have two cases to deal with. We start with Rule III for which conditions (\ref{pb}) and (\ref{pbb}) yield $p_{III}^{(k)}(0,w) =\delta_w^k$ and $p_{III}^{(k)}(r,0) =\delta_0^k$ respectively. Given this and using the fact that \begin{align} \frac{1}{z}\sum_{r=0}^{\infty} (r+1)^2p_{III}^{(k)}(r,0)\frac{1}{z^r} -\sum_{r=0}^{\infty} r^2p_{III}^{(k)}(r,0)\frac{1}{z^r} = \left(\frac{1}{z}\sum_{r=0}^{\infty} (r+1)^2\frac{1}{z^r} -\sum_{r=0}^{\infty} r^2\frac{1}{z^r} \right) \delta_0^k=0, \end{align} we conclude that for the Rule III case, equation (\ref{pde21}) reduces to \begin{subequations}\label{pde22} \begin{empheq}[box=\widefbox]{align} \label{pde3} D_{z,u} \tilde{Y}_{III}^{(k)} (z,u)& = (k+1)^2\frac{1}{u^{1+k}} - k^2 \frac{1}{u^k}, \mbox{ for Rule III,} \\ \label{difO} D_{z,u} &\equiv \left(\frac{1}{z}+\frac{1}{u} \right) -\frac{1}{u} \left(u+z(2+u) \right) \partial_z -\left(1+u \right) \partial_u \\ \notag &-z\left(z-1 \right) \partial^2_{zz} -2z\left(u-1 \right) \partial^2_{zu} -u\left(u-1 \right) \partial^2_{uu}. \end{empheq} \end{subequations} The second case is about Rule IV. Using an analogous approach as in the case of Rule III, it is eventually found that in the case of Rule IV, equation (\ref{pde21}) reduces to \begin{equation}\label{pde224} \boxed{ \begin{aligned} D_{z,u} \tilde{Y}_{IV}^{(k)} (z,u)& = (k+1)^2\frac{1}{z^{1+k}} - k^2 \frac{1}{z^k} , \mbox{ for Rule IV} \end{aligned} } \end{equation} where $D_{z,u}$ is the (same) differential operator defined in (\ref{difO}). We observe that the resulting PDEs are $2^{nd}$ order linear PDEs with mixed derivatives, which are generally harder equations to solve. \section{Reproducing known results from the literature}\label{special} The following known problem \cite{ MuL} is given. \vspace{0.2cm} \hspace{-0.6cm} {\bf Problem 3} \vspace{0.2cm} {\it Inside a box containing $r$ red and $w$ white balls random removals occur. The balls are removed one by one according to the three following rules. Rule I: If a white ball is chosen it is immediately discarded from the box. If a red ball is chosen, it is placed back into the box and a new ball is randomly chosen. The second ball is then removed irrespective of the color. Rule II: Once a ball is removed, the game continues from Rule I. Rule III: The game ends once all balls, except from one ball that remains, are removed. The question posed is the determination of the probability that the remaining ball is white.} \vspace{0.2cm} \hspace{-0.6cm} {\bf Solution} \vspace{0.2cm} All the steps that lead to (\ref{odeG}) are identical for this problem as well and hence, (\ref{odeG}) is the starting point. Given Rule III, which in particular, it implies that even at the event where all reds are removed and where $k>1$ white balls remain, we are still allowed to keep removing white balls (with probability one) until we end up with a single white ball. Thus, we now drop the superscript $k=1$ from $Y^{(k)}$ and $f^{(k)}$ as, according to Rule III, we will keep removing balls until one remains. Therefore, $f_{0,0}=0$ and $f_{0,w}=1, \,\,\forall \,w>0$ \footnote{In particular, equation (\ref{pb}) ceases to apply.} and hence, the right hand side of (\ref{odeG}) becomes $-1/u=-(1+x_1+x_2\sqrt{x_1})/x_1$ (see (\ref{dudz}) and (\ref{x12})). Hence, (\ref{odeG}) simplifies to \begin{subequations} \label{odespecialTemp} \begin{empheq}{align} Y(x_1,x_2) -2x_1\partial_{x_1}Y(x_1,x_2) & = -1, \\ \lim_{x_2 \to -\frac{1}{\sqrt{x_1}}-\sqrt{x_1}} Y(x_1,x_2) & = 0 \end{empheq} \end{subequations} where the boundary condition implies that when no white balls exist ($u\to\infty $), which is equivalent to $x_2 \to -\frac{1}{\sqrt{x_1}}-\sqrt{x_1}$ (see (\ref{x2})), the probabilities and hence, the generating functional to remain with one white ball at the end, should vanish. The general solution to (\ref{odespecialTemp}) is given by \begin{subequations} \label{odespecial} \begin{align} Y(x_1,x_2)& = -1+C(x_2)\sqrt{x_1}, \\ \lim_{x_2 \to -\frac{1}{\sqrt{x_1}}-\sqrt{x_1}} Y(x_1,x_2) & = 0 \end{align} \end{subequations} where $C(x_2)$ is an arbitrary integration constant to be specified by the boundary condition, which in particular, it implies \begin{align}\label{C2special} C(x_2)& = \frac{1}{\sqrt {\frac{1}{2} \left( -2+x_2^2-x_2\sqrt{4-x_2^2} \right) }} \mbox{, (using $x_2<0)$}\\ \nonumber &=\frac{1}{\sqrt {\frac{1}{2} \left( -2+x_2^2+x_2^2\sqrt{1-\frac{4}{x_2^2}} \right) }} \mbox{, (simplifying)} \\ \nonumber &=\frac{x_2}{2} \left(\sqrt{1-\frac{4}{x_2^2}} -1\right), \end{align} where the simplification from the second to the third line uses the following \begin{align}\label{idt} \mbox{if } \sqrt{x}+\frac{1}{\sqrt{x}}=-t \mbox{ then } x = \frac{1}{2} \left( -2+t^2 \pm t^2\sqrt{1-\frac{4}{t^2}} \right) = \left( \frac{t}{2}\left(1\pm \sqrt{1-\frac{4}{t^2}} \right)\right)^2. \end{align} The last simplification is very convenient when expanding in Laurent series. Combining now (\ref{odespecial}) with (\ref{C2special}) we arrive at the desired generating functional \begin{align}\label{gfs} \boxed{ Y(x_1(z,u),x_2(z,u)) = -1+\frac{x_2(z,u)}{2} \left(\sqrt{1-\frac{4}{x_2^2(z,u)}} -1\right)\sqrt{x_1(z,u)} } \end{align} where $x_1$ and $x_2$, according to (\ref{x12}), are functions of $z$ and $u$. It is straightforward to check that (\ref{gfs}) satisfies the set of equations in (\ref{odespecialTemp}). The final step is to use (\ref{gfs}) and (\ref{x12}) and Laurent expand in inverse powers of $z$ and of $u$ noting from (\ref{x12}) that as $z,\,u \to \infty$ we have $x_1 \to \infty$ and $\,x_2 \to -\infty$. In order to expand we make use of the following identities \begin{subequations}\label{exp} \begin{align} \label{expa} \frac{x_2}{2} \left(\sqrt{1-\frac{4}{x_2^2}} -1\right)&= - \sum_{i=0}^{\infty} \frac{1}{1+i} \frac{\Gamma(1+2i)}{\Gamma^2(1+i)} \left( \frac{1}{x_2}\right)^{1+2i},\\ \label{expb} \left(\frac{1}{x_2}\right)^l &= \left(-\sqrt{z}\left(1-\frac{1}{u} + \frac{1}{z} \right)\right)^{-l} \\ \nonumber &= (-1)^l z^{-\frac{l}{2}} \sum_{n=0}^{\infty} \sum_{m=0}^{\infty} \frac{(-1)^m\Gamma(1-l)}{\Gamma(1-l-n-m)\Gamma(1+n)\Gamma(1+m)} \left(\frac{1}{z}\right)^{n} \left(\frac{1}{u}\right)^{m} \end{align} \end{subequations} where we use Gamma functions instead of factorials for reasons that will become evident in what follows. Using the expansions (\ref{exp}), the generating functional (\ref{gfs}) expands as \begin{align}\label{Ysexp1} Y(z,u) = -1 +\sum_{i=0}^{\infty} \sum_{n=0}^{\infty} \sum_{w=0}^{\infty} &\Bigg[ \frac{1}{1+i} \frac{\Gamma(1+2i)}{\Gamma^2(1+i)} \times \\ \nonumber &\frac{(-1)^{w+2i+2}\Gamma(-2i)}{\Gamma(-2i-n-w)\Gamma(1+n)\Gamma(1+w)} \left(\frac{1}{z}\right)^{n+i} \left(\frac{1}{u}\right)^{w}\Bigg]\\ \nonumber = \sum_{\substack{i,n,w= 0\\(i,n,m)\neq (0,0,0)}}^{\infty} &\Bigg[\frac{1}{1+i} \frac{\Gamma(1+2i)}{\Gamma^2(1+i)} \times \\ \nonumber &\frac{(-1)^{w+2i+2}\Gamma(-2i)}{\Gamma(-2i-n-w)\Gamma(1+n)\Gamma(1+w)} \left(\frac{1}{z}\right)^{n+i} \left(\frac{1}{u}\right)^{w} \Bigg]. \end{align} Changing the dummy indices $i$ and $n$ by setting $i = r-n$ with $r=0,\,1,\,2,\,...$ and (a new) $n=0,\,1,\,2,\,...,\,r$ equation (\ref{Ysexp1}) becomes \begin{align} \label{fb1} Y(z,u) = \sum_{\substack{r,w= 0\\(r,w)\neq (0,0)}}^{\infty} \left(\frac{1}{z}\right)^{r} \left(\frac{1}{u}\right)^{w} \sum_{n=0}^{r} &\frac{1}{1+r-n} \frac{\Gamma(1+2r-2n)}{\Gamma^2(1+r-n)} \frac{1}{\Gamma(1+n)\Gamma(1+w)} \times \\ \nonumber &(-1)^{w}\frac{\Gamma(2n-2r)}{\Gamma(-2r+n-w)}. \end{align} Our approach is to use (\ref{fb1}) and compute the summation term over $n$, which, according to (\ref{Yfdef}), is by definition equal to $f_{r,w}$ (with the $k$ index dropped). With $f_{r,w}$ at hand, using (\ref{ans2}) we will eventually obtain the required probability formula. A comment is in order: both, the numerator and the denominator of the last fraction of the Gamma functions in (\ref{fb1}) diverge because $n$, $w$ and $r$ are positive integers with $r \geq n$. This is why we have replaced the factorials with Gamma functions as an analytic continuation of the factorials. These infinities are fictitious and are a consequence of interchanging the summation orderings. Similar infinities will appear in sections \ref{result} and \ref{result4}. Hence, in order to proceed we need the behavior of the Gamma functions near their poles; the corresponding expansions are documented in appendix \ref{A}. As we will show, the end result, when taking the limits carefully, will be finite and will reproduce exactly the correct coefficients such that the recursion (\ref{rec3}) and the boundary condition $p(r,0)=0$ are both fulfilled, which is what matters at the end. Using (\ref{ga}) in order to cancel the poles of the Gamma functions\footnote{We only consider the poles resulting from $r\to \mathbb{Z}_{\geq 0}$ because we ignore poles resulting from the dummy summation index $n\to \mathbb{Z}_{\geq 0}$. Another way to see it is to only consider the pole of the variable $-(r-n) \leq 0$. Furthermore, we ignore the pole in $\Gamma(-2(r-n)-n-w)$ coming from $w\to \mathbb{Z}_{\geq 0}$ (and from $n\to \mathbb{Z}_{\geq 0}$) because it is not compensated by a similar $w$ (or $n$) pole in the numerator and hence, it does not contribute a finite part.} the last fraction of the Gamma's in (\ref{fb1}) is exchanged according to \begin{align} (-1)^{w}\frac{\Gamma(2n-2r)}{\Gamma(-2r+n-w)}=(-1)^{w}\frac{\Gamma(-2(r-n))}{\Gamma(-2(r-n)-n-w)} \to (-1)^{n}\frac{\Gamma(2r+w-n+1)}{\Gamma(1+2r-2n)}. \end{align} Simplifying then the $\Gamma(1+2r-2n)$ that appears in the numerator and in the denominator of the resulting step, keeping track of the correct overall sign as $(-1)^n$, and changing the dummy index $n$ setting $n\to r-n$, transforms the summand of (\ref{fb1}) into $(-1)^{r+n} \Gamma(1+r+w+n)/\left((1+n)\Gamma(1+w) \Gamma(1+r-n) \Gamma^2(1+n) \right)$ where as before the (new) $n=0,\,1,\,...,\,r$. Performing then the summation of the resulting expression over $n$ using the summation identity \begin{align} \sum_{n=0}^r (-1)^{r+n}\frac{ \Gamma(1+r+w+n)} { (1+n)\Gamma(1+w) \Gamma(1+r-n) \Gamma^2(1+n) } =\left(\frac{(r+w)!}{r! w!}\right)^2 \frac{w}{(r+1)(r+w)} \end{align} proved in Appendix \ref{C}, equation (\ref{fb1}) eventually becomes\footnote{In Appendix \ref{B}, we will briefly provide a second derivation of (\ref{ffinal2}).} \begin{align}\label{ffinal2} Y(z,u) = \sum_{\substack{r,w= 0\\(r,w)\neq (0,0)}}^{\infty} \left(\frac{1}{z}\right)^{r} \left(\frac{1}{u}\right)^{w}\left(\frac{(r+w)!}{r! w!}\right)^2 \frac{w}{(r+1)(r+w)}. \end{align} The last step is to use (\ref{Yfdef}) in order to identify the $f_{r,w}$ coefficient from (\ref{ffinal2}) and multiply it by the overall ansatz factor from (\ref{ans2}) to obtain \begin{subequations}\label{solp2} \begin{empheq}[box=\widefbox]{align} p(0,0)& = 0, \\ p(r,w)& = \frac{w}{(w+r)(1+r)},\,\, \forall\, r,\,w \,\geq0,\,(r,w)\neq(0,0). \end{empheq} \end{subequations} Equation (\ref{solp2}) is the main result of the section and reproduces the result of \cite{ MuL} obtained through a completely different (inductive) approach. \section{Deriving the generic generating functional for any number of remaining white balls $k$, Rule III}\label{genPDE} Having solved Problem 3, which uses all the ingredients of our proposed approach, we gained confidence in order to attack Problem 1. In the following two sections, we simplify the notation by dropping the index $_{III}$ from $Y_{III}^{(k)}(x_1,x_2)$, $f^{(k)}_{III;r,w}$ and $p_{III}^{(k)}(r,w)$; we will restore it at the very end on the final formula for $p_{III}^{(k)}(r,w)$. Starting from (\ref{odeG}), we write the differential equation and the boundary conditions for Problem 1, Rule III. Noting that $f^{(k)}_{0,w} =\delta^k_w$, because $f^{(k)}_{0,w}$ denotes the probability to remain with $k$ white balls if we start with no red balls and with $w$ white balls (see (\ref{pb})), we obtain \begin{subequations} \label{odep1} \begin{empheq}[box=\widefbox]{align} \label{odep1a} \frac{1+x_1+ x_2\sqrt{x_1}}{x_1} \Big(Y^{(k)}(x_1,x_2)& -2x_1\partial_{x_1}Y^{(k)}(x_1,x_2) \Big) \\ \nonumber &= (1+k)\frac{1}{u^{1+k}(x_1,x_2)}-k\frac{1}{u^{k}(x_1,x_2)},\\ \label{odep1b} \lim_{z\to = \infty}Y^{(k)}(x_1,x_2) &= \frac{1}{u^k},\,\,z=z(x_1, x_2),\, u=u(x_1,x_2) \end{empheq} \end{subequations} where we have indicated explicitly that $z$ and $u$ are functions of $x_1$ and of $x_2$ (see (\ref{x12})). The boundary condition $z\to \infty$ (game starts without any red balls) is a boundary case where the game ends before it (even) begins. Then, the probability to remain with $k$ white balls should be equal to one if we start with $k$ white balls and zero otherwise \footnote{The argument takes into account the ansatz factor in (\ref{ans2}), which also reduces to one, when $r=0$.}. The boundary condition in the generating functional representation, equation (\ref{odep1b}), is a manifestation of equation (\ref{pb}), which is the boundary condition in the probability representation. The solution of (\ref{odep1}) that satisfies the boundary condition turns out to be \begin{subequations} \label{solk} \begin{empheq}[box=\widefbox]{align} \label{solka} & \hspace{-0.0in}Y^{(0)}(x_1,x_2) = 1 -\sqrt{x_1} \frac{2x_2+\sqrt{2\left(x_2^2 \left(1+\sqrt{1-\frac{4}{x_2^2}}\,\right)-2\right)}}{x_2^2\left(1+\sqrt{1-\frac{4}{x_2^2}} \,\right)-4} \\ \nonumber & \hspace{0.8in}=1 + \sqrt{x_1} \frac{x_2\left(\sqrt{1-\frac{4}{x_2^2}} -1\,\right)}{x_2^2\left(1+\sqrt{1-\frac{4}{x_2^2}} \,\right)-4}, \\ \label{solkb} &\hspace{-0.2in}Y^{(k>0)}(x_1,x_2) = \frac{2^{-2(1+k)}x_1^{-k}}{x_2^2\left(1-\frac{4}{x_2^2} \right)} \left(2+x_2 \sqrt{x_1}\left(1-\sqrt{1-\frac{4}{x_2^2}}\,\right) \right)^{1+2k} \times \\ \nonumber & \hspace{-0.35in} \Bigg \{ \hspace{-0.1in} - \hspace{-0.05in} \frac{2+x_2\sqrt{x_1}}{1+x_1+x_2\sqrt{x_1}} \hspace{-0.06in} \left(\hspace{-0.05in} 2\hspace{-0.03in} +\hspace{-0.03in} x_2 \sqrt{x_1}\left(1\hspace{-0.05in} + \hspace{-0.05in} \sqrt{1-\frac{4}{x_2^2}}\,\right) \hspace{-0.05in} \right) \hspace{-0.05in} \Bigg(\hspace{-0.05in} 1\hspace{-0.03in} + \hspace{-0.03in} \frac{2x_2\sqrt{x_1}\sqrt{1-\frac{4}{x_2^2}}}{2+x_2 \sqrt{x_1}\left(1-\sqrt{1-\frac{4}{x_2^2}}\,\right)} \Bigg)^k \\ \nonumber &\hspace{-0.3in} \hspace{0.25in}+ 2\frac{(1+k)x_2^2-2 }{1+2k} \,_2F_1\left(-1-2k,-k;-2k;-\frac{2x_2\sqrt{x_1}\sqrt{1-\frac{4}{x_2^2}}}{2+x_2 \sqrt{x_1}\left(1-\sqrt{1-\frac{4}{x_2^2}}\,\right)} \right) \Bigg \} \end{empheq} \end{subequations} where in (\ref{solka}) we simplified using (\ref{idt}), and where $_2F_1$ in (\ref{solkb}) is a hypergeometric function It is a straightforward but also a tedious process to show that (\ref{solk}) satisfies the differential equation (\ref{odep1}) and a computer program such as ``Mathematica" is recommended. The fact that $Y^{(0)}$ fulfills the boundary condition can be verified by taking first the limit $\lim {x_2 \to -\sqrt{z}}(1-\frac{1}{u})$ followed by the limit $\lim {z \to \infty}$. These limits are a consequence of (\ref{x12}) and of the boundary condition (\ref{odep1b}). In order to see that the boundary condition for $Y^{(k>0)}$ is also satisfied we take the same limits as the limits we took for $Y^{(0)}$. Then, the overall multiplicative factor of (\ref{solkb}) decays as $-(u-1)^{-3-2k}u^2/2\times1/z^{1+k}$. Expanding the first term of the curly bracket (not the one involving the $_2F_1$) to leading order in $z$ shows that this quantity grows as $ -2(u-1)^{2k+2}/u^{k+1} \times z^{k+1}$. Next, we move to the second term in the curly bracket, that is the hypergeometric function whose argument grows as $-(1-u)^2/u \times z$. Taking into account that if $k$ is a positive integer, we find that the $_2F_1\left(-1-2k,-k;-2k;-x\right)$ becomes a terminating polynomial of degree $k$, which grows as $(2k+1)/(k+1)x^{k}$ as $x \to \infty$. Thus, the whole term involving the $_2F_1$ grows as $2(1-u)^{2k+2}/u^{k+2} \times z^{k+1}$. Therefore, the curly bracket grows as $-2(u-1)^{2k+3}/u^{k+2} \times z^{k+1}$ and given the overall pre-factor $-(u-1)^{-3-2k}u^2/2\times1/z^{1+k}$ it is concluded that (\ref{odep1b}) converges to $1/u^k$ as should. We have thus just shown that the boundary condition is fulfilled $\forall \,k>0$. In the following section we will finalize the solution of Problem 1, Rule III. Using (\ref{solk}), we will compute explicitly the probabilities $p^{(k)}(r,w)$ for the first few values of $k$ and from there we will derive the formula for any $k$. It is also interesting to note that the alternative boundary conditions \begin{align}\label{x2bc} \lim_{x_2\to -\sqrt{x_1}-\frac{1}{\sqrt{x_1}}}Y^{(k)}(x_1,x_2) &= \delta_0^k \frac{x_1}{x_1-1} \end{align} yield the same exact solutions provided by equations (\ref{odep1}). In particular, the boundary conditions defined by (\ref{x2bc}) are the continuum version of the discrete boundary conditions provided by equation (\ref{pbb}). These boundary conditions imply that as $u\to \infty$ (game starts without any white balls), which due to (\ref{x12}) is equivalent to $x_2\to -\sqrt{x_1}-\frac{1}{\sqrt{x_1}}$, the probability to remain with $k>0$ white balls is zero. If, on the other hand, $k=0$ then the probability to remain with no white balls starting without any white balls, must be equal to one independently on the initial number of red balls \footnote{The argument takes into account the ansatz factor in (\ref{ans2}), which also reduces to one, when $w=0$.}. Indeed, expanding the expression $x_1/(x_1-1)$ in inverse powers of $x_1$ yields unit coefficients \footnote{We remind the reader that $x_1=z$ and that the variable $z$ corresponds to the red balls.}. We reserve the details, which show that the boundary conditions (\ref{x2bc}) are satisfied by (\ref{odep1}), for appendix \ref{Z}. In case the reader wonders about the uniqueness of the problem, the answer is that the two boundary conditions are equivalent because they define the same exact problem. In particular, one may choose to evolve the initial data starting from the surface $z \to \infty$ or to evolve the initial data starting from the surface $u \to \infty$, as long as the chosen boundary surface is known. In this problem, it happens that we know both boundaries. In what follows, things will become clearer and the simultaneous fulfillment of both boundary conditions will be checked once explicit formulas are obtained. \section{The general probability formula for Rule III}\label{result} Having developed all the necessary machinery we are now in the position to answer the question of Problem 1, Rule III. The strategy we will apply is as follows. We first consider (\ref{solk}) and take the limits $k \to 0,\,1\,,\,2$. Then, noting from (\ref{x12}) that as $z,\,u \to \infty$ we have $x_1\to \infty$ and $x_2\to -\infty$, we Laurent expand the solutions at infinity, that is in inverse powers of $x_1=z$ and $x_2$. Then, we re-expand $x_2$ in inverse powers of $z$ and $u$ in an analogous way as we worked in Problem 3, section \ref{special}. The final step is to identify the coefficients of $1/z^r$ and $1/u^w$ as the $f_{r,w}^{(k)}$'s of (\ref{Yfdef}) and from there, using (\ref{ans2}), we will obtain the $p^{(k)}(r,w)$. Using then the explicit expressions for $p^{(0)}(r,w)$, $p^{(1)}(r,w)$ and $p^{(2)}(r,w)$ and by studying the pattern, we will make an ansatz for the generic solution $p^{(k)}(r,w)$ for any $k$. The last step is to check that the ansatz fulfills the initial recursion we begun with, namely equation (\ref{rec3}), and the condition $p^{(k)}(0,w) = \delta_w^k$. It is also noted that in principle we could had reached the same answer without making any anstaz at all. That is through an even more systematic approach; by Laurent expanding the general PDE solutions (\ref{solk}) for arbitrary $k$. \subsection{The case $k=0$}\label{sk0} Despite this case can be obtained using the strategy outlined in the introduction of the current section, there is a way around it using the results of section \ref{special}. In particular, the probability of Problem 1, Rule III for $k=0$ is equal to the complement of the probability of Problem 3, which is the probability that the last ball is red; that it no white balls remain at the end of the game. Thus, using (\ref{solp2}), we find \begin{subequations}\label{solp10temp} \begin{empheq}{align p^{(0)}(0,0)& = 1, \\ p^{(0)}(r,w)& = r\frac{1 + r + w}{(w+r)(1+r)},\,\, (r,w)\neq(0,0). \end{empheq} \end{subequations} In the view of (\ref{pa}) and (\ref{pc}), equation (\ref{solp10temp}) can be written compactly as \begin{align}\label{solp10} \boxed{ p_{III}^{(0)}(r,w) = r\frac{1 + r + w}{(w+r)(1+r)},\,\, \forall \,r,\,w \geq 0. } \end{align} As a cross check of (\ref{solp10}), we substitute (\ref{x12}) into (\ref{solka}) and expand the result at $z=\infty$ and at $u=\infty$ up to $4^{th}$ order to obtain \begin{align} Y^{(0)}(z,u) = \left(1+O\left( \frac{1}{u^5}\right) \right) &+ \left(1+ \frac{3}{u}+\frac{6}{u^2}+ \frac{10}{u^3}+\frac{15}{u^4}+O\left( \frac{1}{u^5}\right) \right) \frac{1}{z}\\ \nonumber\ &+ \left(1+ \frac{8}{u}+\frac{30}{u^2}+ \frac{80}{u^3}+\frac{175}{u^4}+O\left( \frac{1}{u^5}\right) \right) \frac{1}{z^2}\\ \nonumber\ &+ \left(1+ \frac{15}{u}+\frac{90}{u^2}+ \frac{350}{u^3}+\frac{1050}{u^4}+O\left( \frac{1}{u^5}\right) \right) \frac{1}{z^3}\\ \nonumber\ &+ \left(1+ \frac{24}{u}+\frac{210}{u^2}+ \frac{1120}{u^3}+\frac{4410}{u^4}+O\left( \frac{1}{u^5}\right) \right) \frac{1}{z^4}+O\left( \frac{1}{z^5}\right). \end{align} We then compare previous expansion with the quantity \begin{align} Y^{(0)}(z,u) = \sum_{r=0}^{4} \sum_{w=0} ^{4 } f^{(0)}_{r,w} z^{-r} u^{-w} = 1+\mathop{\sum_{r=0}^{4}\sum_{w=0}^{4}}_{(r,w)\neq(0,0)} \left(\frac{(r+w)!}{r! w!}\right)^2 r\frac{1 + r + w}{(w+r)(1+r)} z^{-r} u^{-w} \end{align} observing an exact matching order to order in $z$ and in $u$. It is also notable that both boundary conditions (\ref{odep1b}) and (\ref{x2bc}) are satisfied. Indeed, checking the condition (\ref{odep1b}) is obvious. On the other hand, as $u\to \infty$, the expansion reduces to $1+1/z+1/z^2+1/z^3+....=z/(z-1)=x_1/(x_1-1)$ in accordance with the (alternative) boundary condition in (\ref{x2bc}). Thus, as discussed below (\ref{x2bc}), we can see explicitly the equivalence of the two boundary conditions. Equation (\ref{solp10}) will also be obtained as the $k\to 0$ limit of the general solution $p^{(k)}(r,w)$. \subsection{The case $k=1$} \label{sk1} Taking the limit in (\ref{solkb}) as $k \to 1$ yields \begin{align}\label{k1} Y^{(1)}(x_1,x_2) &=\left(1 +\frac{2}{3x_1} + \frac{x_2}{\sqrt{x_1}} \right) \\ \nonumber &+\frac{x_2\sqrt{x_1}}{6}\left(3-\sqrt{1-\frac{4}{x_2^2}}\,\, \right)-\frac{x_2^3\sqrt{x_1}}{6}\left(1-\sqrt{1-\frac{4}{x_2^2}}\,\, \right). \end{align} Using (\ref{x12}), the first line of (\ref{k1}) expands as \begin{align}\label{3z} 1 +\frac{2}{3x_1} + \frac{x_2}{\sqrt{x_1}} = \frac{1}{u}-\frac{1}{3 z} \end{align} while the second line, after some algebra, expressed as an expansion in $x_1$ and $x_2$ yields \begin{align}\label{intk1} \frac{x_2\sqrt{x_1}}{6}\left(3-\sqrt{1-\frac{4}{x_2^2}}\,\, \right)&-\frac{x_2^3\sqrt{x_1}}{6}\left(1-\sqrt{1-\frac{4}{x_2^2}}\,\, \right) \\ \nonumber & = -\sqrt{x_1}\sum_{i=2}^{\infty} \frac{(i-1)\Gamma(2i-1)}{i(1+i)\Gamma^2(i)} \frac{1}{x_2^{2i-1}}. \end{align} From here and on we work analogously to equations (\ref{exp})-(\ref{fb1}). In particular\footnote{The explanations that follow assume an identical indexing in $i$, $n$ $r$ and $w$ as the indexing of section \ref{special}.}, we expand $ \frac{1}{x_2^{2i-1}}$ using (\ref{expb}), exchange the index $i$ with $i=r-n+1$ where $r=1,\,2,\,...$ and $n=0,\,1,\,2,\,...\,r-1$ as before to finally obtain \begin{align}\label{prevk11} Y^{(1)}(z,u)= \sum_{r=1}^{\infty} \sum_{w=0}^{\infty} \left(\frac{1}{z}\right)^{r} \left(\frac{1}{u}\right)^{w} \Bigg \{ \sum_{n=0}^{r-1} &\frac{(r-n)}{(1+r-n)(2+r-n)} \frac{\Gamma(1+2(r-n))}{\Gamma^2(1+r-n)} \frac{(-1)^{w}}{\Gamma(1+n)\Gamma(1+w)} \\ \nonumber &\times \frac{\Gamma(2(n-r))}{\Gamma(-2r+n-w)} \Bigg \}+ \frac{1}{u}-\frac{1}{3z}. \end{align} We note that for $r\geq1$, $n<r$ and $w \geq0$ the numerator and the denominator in the last fraction of the Gammas in (\ref{prevk11}) diverges. Using then (\ref{gb}), which allows us to replace the last fraction of the Gamma functions with $(-1)^{n+w}\Gamma(1+2r-n+w)/\Gamma(1+2(r-n))$ \footnote{We only consider the poles resulting from $r\to \mathbb{Z}_{\geq 0}$ because we ignore poles resulting from the dummy summation index $n\to \mathbb{Z}_{\geq 0}$. Furthermore, we ignore the pole in $\Gamma(-2r+n-w)$ coming from $w\to \mathbb{Z}_{\geq 0}$ because it is not compensated by a similar $w$ pole in the numerator and hence, it does not contribute a finite part.}, and equation (\ref{Yfdef}), which defines the $f^{(k)}_{r,w}$ coefficients, we obtain \begin{align} \label{prevk12} f^{(1)}_{r,w} &= \sum_{\substack{n= 0\\r \geq1}}^{r-1} \frac{(-1)^n(r-n)}{(1+r-n)(2+r-n)} \frac{1}{\Gamma^2(1+r-n)} \frac{\Gamma(1+2r-n+w)}{\Gamma(1+n)\Gamma(1+w)} +\left(\frac{1}{u}-\frac{1}{3z} \right) \Bigg|_{r,w} \\ \nonumber\ &=\sum_{\substack{n= 1\\r \geq1}}^{r} \frac{(-1)^{r-n} n}{(1+n)(2+n)} \frac{1}{\Gamma^2(1+n)} \frac{\Gamma(1 + r + w + n)}{\Gamma(1+r-n)\Gamma(1+w)} +\left(\frac{1}{u}-\frac{1}{3z} \right) \Bigg|_{r,w}, \,\,r\geq 1, \,\,w\geq 0 \end{align} where in the second equality we changed the dummy index according to $n\to r-n$ and where $ |_{r,w}$ denotes the coefficient of $\frac{1}{z^r}\frac{1}{u^w}$ in the preceding bracket. Equation (\ref{prevk12}) can be partitioned in four cases according to \begin{subequations} \label{prevk13} \begin{align} \label{prevk13a} f^{(1)}_{1,0} & = \frac{1}{3}+\left( \frac{1}{u}-\frac{1}{3z} \right) \Bigg|_{1,0}=0, \\ \label{prevk13d} f^{(1)}_{0,1} & =\left( \frac{1}{u}-\frac{1}{3z} \right) \Bigg|_{0,1}=1,\\ \label{prevk13b} f^{(1)}_{r,w} &= \frac{(r+w-2)!(1+r+w)!}{(r-1)!(2+r)!(w-1)! w!}, \mbox{ $r\geq1,\,w\geq0$ and $(r,w)\neq(1,0)$}\\ f^{(1)}_{r,w} &=0 \mbox{ otherwise} \end{align} \end{subequations} where (\ref{prevk13a}) is obtained from (\ref{prevk12}) for $(r,w)=(1,0)$, and it has exactly the form we need in order to cancel the undesired $1/(3z)$ term coming from (\ref{prevk11}). Equation (\ref{prevk13b}) is obtained by performing the summation using the identity \begin{align}\label{bio2} \frac{(r+2)!}{(r+w+1)!}\sum_{n=1}^{r}\frac{(-1)^n n}{(1+n)(2+n)} \frac{(r+w+n)!}{(r-n)! (n!)^2} &= \sum_{n=1}^{r} (-1)^n \binom{r+2}{n+2} \binom{r+w+n}{n-1}\\ \nonumber\ &= (-1)^r \frac{(r+w-2)!}{(r-1)!(w-1)!}, \end{align} which can be proved working along the lines of equations (\ref{C1})-(\ref{CendProof}) with minor modifications. We note that in the limit $(r,w)\to(0,1)$, the right hand side of equation (\ref{prevk13b}) tends to one\footnote{The limit $r=0,\,w=1$ in (\ref{prevk13b}) must be taken according to equation (\ref{pa2e}) (see also comments that follow (\ref{pa2e})). In particular, when taking the first limit $w \to 1$, both of the $(r-1)!$ terms that appear in the numerator and the denominator of (\ref{prevk13b}) cancel out yielding $f^{(1)}_{0,1}=1$.}, and hence, (in theory) it provides the $1/u$ term (see boundary conditions, (\ref{odep1b})); in reality, the $1/u$ term is actually coming from $\left(\frac{1}{u}-\frac{1}{3z} \right) \big|_{r=0,w=1}$. Thus, equation (\ref{prevk13d}) can be absorbed in (\ref{prevk13b}) in the view of the limits ordering of (\ref{pa}), and by extending the range of validity suitably. Collecting all the terms of (\ref{k1}) using (\ref{prevk13}) and the observations we just made above, we finally obtain \begin{align}\label{k1exp} Y^{(1)}(z,u) = \sum_{r,w=0}^{\infty} f^{(k)}_{r,w}\left(\frac{1}{z}\right)^{r} \left(\frac{1}{u}\right)^{w}= \sum_{r=0}^{\infty} \sum_{w=0}^{\infty} \frac{(r+w-2)!(1+r+w)!}{(r-1)!(2+r)!(w-1)! w!} \left(\frac{1}{z}\right)^{r} \left(\frac{1}{u}\right)^{w} \end{align} where we extend the summations on $r$ and on $w$ according to $r,\,w \geq0$ in the view of the fact that $1/(r-1)!$ and $1/(w-1)!$ tend separately to zero as $r\to0$ and $w\to0$ respectively. As noted earlier, when $(r,w)=(0,1)$ we get $f^{(k)}_{0,1}=\delta_1^k$ (see (\ref{pa2e}) and the discussion below (\ref{pa2e}) in order to see how these limits should be taken). The reason we expect $f^{(1)}_{0,1}=1$ is because this coefficient encodes the $1/u$ term (see (\ref{odep1b}) and (\ref{3z})) and hence, equation (\ref{k1exp}) provides the right coefficient with value equal to one, which is precisely the probability to remain with one white ball if we start with one white ball and no red balls. To rephrase, if we start with no red balls $(r=0)$ and $w \neq1$ white balls the probability to remain with $k=1$ white balls is zero, otherwise if $w=1$ the probability is equal to one, and this is precisely the meaning of $f^{(k)}_{0,1}=\delta_1^k$. The last step is to use (\ref{Yfdef}) in order to identify $f^{(k)}_{r,w}$ from (\ref{k1exp}), multiply by the coefficient of (\ref{ans2}) and do the necessary simplifications to obtain \begin{align}\label{solp11} \boxed{ p_{III}^{(1)}(r,w) = \frac{rw(1 + r + w)}{(1+r)(2+r)(w+r-1)(w+r)},\,\,\forall \, r,\,w \geq0, } \end{align} which also includes the cases $w=0,\,\, r\neq1$ and $(r,w)=(0,1)$ in the view of (\ref{pbb}) and (\ref{pb}) respectively. One may verify that equation (\ref{solp11}) satisfies the initial recursion we begun with, namely equation (\ref{rec3}) and the boundary condition (\ref{pb}), and in fact, all the equations (\ref{pb})-(\ref{pe}). Comparing (\ref{solp10}) with (\ref{solp11}) we already start to see a pattern forming. In the next section, once the $k=2$ case is computed, the pattern will become obvious. \subsection{The case $k=2$}\label{sk2} Taking the limit in (\ref{solkb}) as $k \to 2$ yields \begin{align}\label{k2} Y^{(2)}(x_1,x_2) &=\left(1 +\frac{3}{5x_1^2} +\frac{3x_2}{2x_1^{\frac{3}{2}}} + \frac{4+3x_2^2}{3x_1}+ \frac{2x_2}{\sqrt{x_1}} \right) \\ \nonumber &+\sqrt{x_1}x_2 \left(\frac{1}{2}-\frac{2}{15}\sqrt{1-\frac{4}{x_2^2}}\,\, \right) +x_2^3\sqrt{x_1}\left(-\frac{1}{3}+\frac{7}{30}\sqrt{1-\frac{4}{x_2^2}}\,\, \right)\\ \nonumber &+\frac{\sqrt{x_1}x_2^5}{20}\left( 1-\sqrt{1-\frac{4}{x_2^2}}\,\, \right). \end{align} Before computing the general $f^{(k)}_{r,w}$ term, we expand (\ref{k2}) up to $4^{th}$ order in $1/z$ and $1/u$ as a way to cross-check our calculations along the way. The result is \begin{align}\label{k2exp} Y^{(2)}(z,u)& = \left(\frac{1}{u^2}+O\left( \frac{1}{u^5}\right) \right) + \left(\frac{1}{u^2}+ \frac{5}{3u^3}+\frac{5}{2u^4}+O\left( \frac{1}{u^5}\right) \right) \frac{1}{z}\\ \nonumber\ &+ \left(\frac{1}{u^2}+ \frac{4}{u^3}+\frac{21}{2u^4}+O\left( \frac{1}{u^5}\right) \right) \frac{1}{z^2} + \left(\frac{1}{u^2}+ \frac{7}{u^3}+\frac{28}{u^4}+O\left( \frac{1}{u^5}\right) \right) \frac{1}{z^3}\\ \nonumber\ &+ \left(\frac{1}{u^2}+ \frac{32}{3u^3}+\frac{60}{u^4}+O\left( \frac{1}{u^5}\right) \right) \frac{1}{z^4}+O\left( \frac{1}{z^5}\right). \end{align} As expected, according to the boundary condition (\ref{odep1b}), the lowest order in $1/u$ is $1/u^2$, which basically says that unless we start with at least two white balls, the probability to end up with two white balls is zero. More specifically, we know that if $(r,w)=(0,2)$ we should get a probability equal to one, which is precisely equal to the coefficient of the $1/z^01/u^2$ term in the expansion. We also note that as $u\to \infty$, $Y^{(2)}(z,u) \to 0$ in agreement with the alternative boundary condition (\ref{x2bc}). Returning now to equation (\ref{k2}) and expanding it yields \begin{align}\label{y21} Y^{(2)}(z,u) = \frac{1}{u^2}-\frac{1}{6z}-\frac{1}{2uz}+\frac{1}{10z^2} -2\sqrt{x_1(z,u)}\sum_{i=0}^{\infty}\frac{4^i i \Gamma(\frac{1}{2}+i)}{\sqrt{\pi}\Gamma(4+i)}\frac{1}{x_2^{1+2i}(z,u)} \end{align} where the sum corresponds to the last two lines of (\ref{k2}). We note that we already have the desired $1/u^2$ term. Working then analogously to sections \ref{special} and \ref{sk1}, and using the Gamma function identities \begin{subequations} \label{idsk2} \begin{align} \Gamma \left(\frac{1}{2}+x\right) = 2^{1-2x}\sqrt{\pi} \frac{\Gamma(2x)}{\Gamma(x)},\\ \pi \csc \left(2\pi x\right) = - 2x \Gamma(-2x) \Gamma(2x), \end{align} \end{subequations} and expanding $x_2$ as in (\ref{expb}) we find that (\ref{y21}) yields \begin{align} \label{sumk21} Y^{(2)}(z,u) &= \frac{1}{u^2}-\frac{1}{6z}-\frac{1}{2uz}+\frac{1}{10z^2} -\sum_{r,w=0}^{\infty} \left(\frac{1}{z}\right)^{r} \left(\frac{1}{u}\right)^{w} \Bigg \{2 (-1)^w \\ \nonumber &\times \sum_{n=0}^r \frac{ \pi \csc(2\pi(r-n))}{ \Gamma(1+w) \Gamma(r-n) \Gamma(4+r-n) \Gamma(1+n) \Gamma(-2r-w+n)} \Bigg \}. \end{align} We note that for any $w$, when $r=0$, and hence $n=0$, the summation term over $n$ becomes trivial and also it yields a zero result. The reason is because the numerator has only a single pole in $r$ coming from $\csc(2\pi (r-n)) $ while the denominator has a second order pole in $r$, one coming from $\Gamma(r-n)$ and one coming from $\Gamma(-2r-w+n)$. Hence, the summation over $r$ should begin from $r=1$. As it turns out the summation over $n$ is, up to overall Gamma function factors, a regularized hypergeometric function with integer coefficients that reduces to a hypergeometric function $_2F_1$. In particular, (\ref{sumk21}) yields \begin{align} \label{sumk22} Y^{(2)}(z,u) = & \frac{1}{u^2}-\frac{1}{6z}-\frac{1}{2uz}+\frac{1}{10z^2} \\ \nonumber &-\sum_{r=1,w=0}^{\infty} \left(\frac{1}{z}\right)^{r} \left(\frac{1}{u}\right)^{w} \Bigg \{ 2 (-1)^w \frac{ \pi \csc(2\pi r) \,_2F_1\left( -3-r,1-r;-2r-w;1\right)}{ \Gamma(1+w) \Gamma(4+r) \Gamma(r) \Gamma(-2r-w)} \Bigg \ \\ \nonumber =& \frac{1}{u^2}-\frac{1}{6z}-\frac{1}{2zu}+\frac{1}{10z^2} \\ \nonumber &+\sum_{r=1,w=0}^{\infty} \left(\frac{1}{z}\right)^{r} \left(\frac{1}{u}\right)^{w} \Bigg \{2 \frac{\Gamma(1+2r+w) \,_2F_1\left( -3-r,1-r;-2r-w;1\right)}{ \Gamma(1+w) \Gamma(4+r) \Gamma(r) } \Bigg \} \end{align} where in the second equality we used equation (\ref{gb}) on $\Gamma(-2r-w)$ and the fact that as $r\to j \in \mathbb{Z}$ then $\pi \csc(2\pi r) \approx \frac{1}{2(r-j)}$. These two facts allow the cancellation of the poles in the ratio $\pi \csc(2\pi r)/\Gamma(-2r-w)$ and its replacement with $-\Gamma(2r+w+1)(-1)^w$ \footnote{We only consider the pole resulting from $r\to \mathbb{Z}_{\geq 0}$ and we ignore the pole in $\Gamma(-2r-w)$ coming from $w\to \mathbb{Z}_{\geq 0}$ because this pole is not compensated by a similar $w$ pole in the numerator and hence, it does not contribute a finite part.}. Given also that $_2F_1\left( -4,0;-2;x\right)=1$, $_2F_1\left( -4,0;-3;x\right)=1$, $_2F_1\left( -5,-1;-4;x\right)=1-5/4 x$ and $_2F_1\left( -5,-1;-5;x\right)=1- x$, the curly bracket in (\ref{sumk22}) yields the following terms \begin{subequations} \label{cancellterms} \begin{align} \frac{1}{6z} &\mbox{ for $(r,w)=(1,0)$,}\\ \frac{1}{2zu} &\mbox{ for $(r,w)=(1,1)$,}\\ -\frac{1}{10z^2} &\mbox{ for $(r,w)=(2,0)$,}\\ 0 &\mbox{ for $(r,w)=(2,1)$,}\\ 0 &\mbox{ for $r=0$ and $w \geq 0$},\\%0 &\mbox{ for $r=0$ and $w > 2$} \label{w2r2} 0 &\mbox{ for $w<2$ and $r > 2$} \end{align} \end{subequations} Equation (\ref{w2r2}) is not obvious but it will become soon due to equation (\ref{casesk2b}), which implies that for $w=\{0,1 \}$ and $r>2$, the $\Gamma(w-1)$ term in the denominator diverges without any compensating factor coming from the numerator; thus the whole term is equal to zero. It is noted that the first three sub-equations of (\ref{cancellterms}) have precisely the right form in order to cancel the unnecessary terms from (\ref{sumk22}) (see (\ref{k2exp})), which thus, in the view of (\ref{Yfdef}), yields \begin{subequations} \begin{align} f^{(2)}_{r,w} \left(\frac{1}{z}\right)^{r} \left(\frac{1}{u}\right)^{w} & = \frac{1}{u^2}, \,\, (r,w)=(0,2),\\ f^{(2)}_{r,w} \left(\frac{1}{z}\right)^{r} \left(\frac{1}{u}\right)^{w} &=2 \left(\frac{1}{z}\right)^{r} \left(\frac{1}{u}\right)^{w} \frac{\Gamma(1+2r+w) }{ \Gamma(1+w) \Gamma(4+r) \Gamma(r) } \\ \nonumber &\times \,_2F_1\left( -3-r,1-r;-2r-w;1\right), \,\,\, r \geq 1,\,\, w\geq 2,\\ f^{(2)}_{r,w} \left(\frac{1}{z}\right)^{r} \left(\frac{1}{u}\right)^{w} &= 0 \mbox{ otherwise}. \end{align} \end{subequations} Next, we use the Gauss theorem on the hypergeometric function $_2F_1$, which states that \begin{align} _2F_1(a,b;c;1) = \frac{\Gamma(c) \Gamma(c-a-b)}{ \Gamma(c-a) \Gamma(c-b)},\,\, c-a-b>0. \end{align} Strictly speaking in our case $c-a-b = 2 -w\leq 0$ for $w\geq2$ and hence, the Gamma functions, of both, the numerator and the denominator, will diverge. Hence, we take the limits of the Gamma functions carefully by taking care of the poles that appear using (\ref{gpoles})\footnote{While taking the required limits, we cross-check by matching the $f^{(2)}_{r,w}$'s with the coefficients of (\ref{k2exp}).}. The end result is finite and is given by \begin{subequations} \label{casesk2} \begin{align} f^{(2)}_{r,w} & = 1, \,\, (r,w)=(0,2),\\ \label{casesk2b} f^{(2)}_{r,w} & = 2 \frac{\Gamma(r+w-2) \Gamma(r+w+2)}{ \Gamma(w-1) \Gamma(w+1) \Gamma(4+r) \Gamma(r) }, \,\,\, r \geq 1,\,\, w\geq 2.\\ f^{(2)}_{r,w} &= 0 \mbox{ otherwise}. \end{align} \end{subequations} Equation (\ref{casesk2}) reproduces the $\frac{1}{z^r} \frac{1}{u^w}$ coefficients of (\ref{k2exp}), and this provides confidence that the equation is correct; it can be written compactly as \begin{align} \label{f2comp} f^{(2)}_{r,w} & = 2 \frac{\Gamma(r+w-2) \Gamma(r+w+2)}{ \Gamma(w-1) \Gamma(w+1) \Gamma(4+r) \Gamma(r) }, \,\, \forall\, r,\,w \geq0 \end{align} where we have extended the range of $r$ and of $w$ by observing that: (i) If $r=0$ and $w=2$, taking the limits in (\ref{f2comp}) according to (\ref{pa}), yields $f^{(2)}_{0,2}=1$ in agreement with (\ref{pb}). (ii) Else if $r=0$ and $w>2$ we have one $r$ pole in the denominator coming from $\Gamma(r)$ that is not compensated by a similar pole in the numerator and hence, the result is zero. (iii) Else if $r=0$ and $w=0,\,1$, despite the ratio $\Gamma(r+w-2)/\Gamma(r)$ is finite, the $1/\Gamma(w-1)$ term tends to zero and hence, $f^{(2)}_{0,1}=f^{(2)}_{0,2}=0$. (iv) Else, if $w=0,\,1$ and $r \geq 1$ the result is also zero, including the cases $(r,w)=\{(1,0),\,(1,1)\}$, in the view of the orderings of the limits of (\ref{pa}) and the fact that the $1/\Gamma(w-1)$ in (\ref{f2comp}) goes to zero as $w\to0,\,1$. Given that we are dealing with $k=2$ and $w=0,\,1$, we see that case (iv) is consistent with equation (\ref{pe}) as should. The last step is to multiply (\ref{f2comp}) by the coefficient of (\ref{ans2}) and do the necessary simplifications to obtain \begin{align}\label{solp12} \boxed{ p_{III}^{(2)}(r,w) =2 \frac{rw(w-1)(1 + r + w)}{(1+r)(2+r)(r+3)(w+r-2)(w+r-1)(w+r)},\,\, r,w \geq 0. } \end{align} It can be checked that equation (\ref{solp12}) satisfies the initial recursion (\ref{rec3}) and the boundary condition (\ref{pb}), and in fact, all the equations (\ref{pb})-(\ref{pe}). This completes the derivation of $ p_{III}^{(2)}(r,w)$. In the following section, we will provide the general solution $ p_{III}^{(k)}(r,w)$ $\forall k$. \subsection{The general solution} Having computed $p^{(k)}(r,w)$ for $k=0,\,1$ and $2$, we are now in position to attempt for a general ansatz solution. Observing equations (\ref{solp10}), (\ref{solp11}) and (\ref{solp12}) one may guess that the solution should have the form \begin{align}\label{finallypka} p_{III}^{(k)}(r,w) &\sim r \frac{w(w-1)(w-2)...(w-k+1)}{(r+1)(r+2)...(r+k+1)} \frac{1+r+w}{(r+w)(r+w-1)....(r+w-k)}\\ \nonumber\ &\sim r\frac{r! w!}{(r+k+1)!(w-k)!} \frac{(1+r+w)(r+w-k-1)!}{(r+w)!} \end{align} up to an overall constant that does not depend on $r$ and $w$. The constant is then specified by the requirement that $p^{(k)}(0,k)=1$ in the view of the boundary condition (\ref{pb}) from where we infer that the solution must be \begin{align}\label{finallypk} \boxed{ p_{III}^{(k)}(r,w) =\frac{k! r!(r+w+1)}{(r+k+1)!}\,\, \frac{rw!(r+w-k-1)!}{(r+w)!(w-k)!}\,\, \forall r,\,w,\,k\geq0,\, k\leq w. } \end{align} The final step is to verify that (\ref{finallypk}) is the required probability solution. Indeed, it is a matter of straightforward algebra to show that (\ref{finallypk}) satisfies the initial recursion we begun with, namely equation (\ref{rec3}), and all the equations (\ref{pb})-(\ref{pe}). We do the following two cross-checks: (i) We first note that the solution reproduces the cases $k=0,\,1,\,2$, equations (\ref{solp10}), (\ref{solp11}) and (\ref{solp12}). (ii) Also, as it is shown in Appendix \ref{D}, the probability formula is normalized and hence, it satisfies \begin{align}\label{normalp} \sum_{k=0}^w p_{III}^{(k)}(r,w) =1, \end{align} which basically says that when the game ends, we will surely end up with a number of white balls between zero and the initial number $w$. We note that the normalization equation (\ref{normalp}) comes out automatically and this key fact serves as another cross check of the correctness of (\ref{finallypk}). At this stage, the solution is considered as complete. Two important observations can be made: (a) It is notable that (\ref{finallypk}) is, up to the rescaling factor $\frac{k! r!(r+w+1)}{(r+k+1)!}$, the same as the solution of Problem 2, equation (\ref{p1}). This implies that this overall factor, in a sense, encodes the additional complication of adding red balls back into the box. (ii) One can also show that the (true) generating probability functional $\tilde{Y}^{(k)}(z,u)= \tilde{Y}^{(k)}_{III}(z,u)$ defined in (\ref{Yt}) satisfies the PDE (\ref{pde22}) with $p^{(k)}(r,w)=p_{III}^{(k)}(r,w)$ given by the probability formula (\ref{finallypk}). We also compute the probability generating functional for fixed $r$ and $w$. It is is given by \begin{align}\label{Gp1III} G_{III}(r,w;z) \equiv \sum_{w=0}^kp_{III}^{(k)}(r,w)z^k = \frac{1+r+w}{1+r} \frac{r}{r+w} \,_3F_2\left(1,1,-w;2+r,1-r-w ;z\right), \end{align} which involves a $\,_3F_2$ hypergeometric function rather than a $\,_2F_1$, which was the case for Problem 2 (see (\ref{Gp2})). \subsubsection{Maximal probability}\label{max3} In this section we investigate the maxima of $p_{III}^{(k)}(r,w)$ as a function of $k$ for fixed values of $r$ and of $w$. We start by constraining $k$ solving the two inequalities \begin{subequations} \begin{align} \label{inq2a} p_{III}^{(k)}(r,w) \geq p_{III}^{(k+1)}(r,w), \\ \label{inq2b} p_{III}^{(k)}(r,w) \geq p_{III}^{(k-1)}(r,w), \end{align} \end{subequations} which provide the set of $k$'s for which $p_{III}^{(k)}(r,w)$ has local maxima (if any). The inequalities (\ref{inq2a}) and (\ref{inq2b}) imply \begin{subequations} \begin{align} k& \leq \frac{1}{2} (1+r)(r+w)-1, \\ k & \geq \frac{1}{2} (1+r)(r+w) \end{align} \end{subequations} respectively. Evidently the system of inequalities has no solution. This hinds that, for $r>0$, the maximal probability is at the boundary cases $p_{III}^{(0)}(r>0,w)$ or $p_{III}^{(w)}(r>0,w)$. In fact, we guess that the $p_{III}^{(0)}(r>0,w)$ should be the required maximal probability. In order to show the claim, we take the difference $p_{III}^{(k)}(r,w)-p_{III}^{(1+k)}(r,w)$ and simplify to obtain \begin{align} p_{III}^{(k)}(r,w)& - p^{(1+k)}(r,w) \\ \nonumber &= |C|r\Gamma(r+w-k-1) (-2 - 2 k + (1+r)(r+w)), \, r \geq 1,\,k\leq w-1 \end{align} where $|C|$ is a positive constant of ratios of factorials. The difference $p_{III}^{(k)}(r,w) - p_{III}^{(1+k)}(r,w)$ is minimum when $k=w-1$, and given that $r\geq1$, it implies that the difference is always positive and hence, $p_{III}^{(k)}(r,w)>p_{III}^{(1+k)}(r,w)$. Therefore, when $r>0$ we find that $p_{III}^{(k)}(r,w)$ is a monotonically decreasing function of $k$. On the other hand, we know that for $r=0$, $p_{III}^{(k)}(0,w) =\delta_w^k$. Thus, we have just proved \begin{subequations} \label{maxpk} \begin{empheq}[box=\widefbox]{align} \label{maxpk3a} &sup\left(p_{III}^{(k)}(r=0,w)|k=0,\,1,\,...\,w \right)=w,\\ \label{maxpk3b} &sup\left(p_{III}^{(k)}(r\geq1,w)|k=0,\,1,\,...\,w \right)=0. \end{empheq} \end{subequations} To conclude, the maximal probability occurs for: (i) $k=w$ if $r=0$ and is given by $p_{III}^{(k=w)}(0,w)=1$. (ii) Else, the maximal probability occurs for $k=0$ and the corresponding probability $p_{III}^{(k=0)}(r,w)$ is given by equation (\ref{solp10}). \subsubsection{Limiting cases} It is also interesting to investigate the behavior of (\ref{finallypk}) in the limits $r\to \infty$ for $w$ fixed and any $k$, and for $w \to \infty$ for $r$ and $k$ fixed. For this purpose we use the more convenient equation (\ref{finallypka}) from where it is deduced that \begin{subequations} \label{limcases} \begin{align} \label{limcasesr} p_{III}^{(k)}(r,w) &= k!\frac{1}{r^{2k} } \left( \delta_0^k + \prod_{i=0}^{k-1} (w-i) \right)+ O\left(\frac{1}{r^{2k+1}} \right),\,\, r \to \infty, \, r \gg w,\\ \label{limcasesw} p_{III}^{(k)}(r,w) &=k! \frac{r \,r!}{(r+k+1)! }+ O\left(\frac{1}{w} \right),\,\, w \to \infty, \, w \gg r,\, k, \end{align} \end{subequations} where $\prod_{i=0}^{k-1} $ for $k=0$ is defined to be zero. The asymptotic expansion (\ref{limcasesr}) says that in the limit $r \to \infty$ with $w$ kept fixed, the probability $p_{III}^{(k>0)}(r,w)$ decays as $\frac{1}{r^{2k}}$. Moreover, if $k=0$, the probability $p_{III}^{(k=0)}(r,w)$ tends to one, which means that when the red balls are much more than the white balls, the game will (most likely) end without any white balls. Both of these two asymptotic results behave as expected. On the other hand, the asymptotic expansion (\ref{limcasesw}) in the limit $w \to \infty$ with $r$ and $k$ kept fixed yields less expected results. In particular, the probability $p_{III}^{(k)}(r,w)$ (to leading order in $w$) becomes independent on the initial number of the white balls $w$. \section{The general probability formula for Rule IV}\label{result4} In this section we solve Problem 1, Rule IV. The ideas are similar as those of Problem 1, Rule III and hence, the derivations are sketchy and much shorter compared to those of sections \ref{genPDE} and \ref{result}. In particular, the proofs between several steps are similar or even identical to the aforementioned sections and therefore, they are omitted. \subsection{The PDE, the boundary conditions and the solution, Rule IV} The starting point is the PDE (\ref{odeG}) where the coefficients in the sums in the right hand side of the equation are all equal to $f_{0,w}^{(k)}=\delta_0^k$ (see (\ref{pbb4})). The (physical) reason is because $f_{0,w}^{(k)}$ is (proportional to) the probability to remain with $k$ red balls if we start without any and this should yield probability equal to one if $k=0$ and equal to zero otherwise. Summing then the two series over $w$ we find that the result is zero $\forall k$, including $k=0$. Hence, the PDE reduces to \begin{subequations}\label{pdebc4} \begin{align}\label{pde4} Y^{(k)}(x_1,x_2) -2x_1\partial_{x_1}Y^{(k)}(x_1,x_2) & =0,\\ \label{bc4} \lim_{x_2 \to -\sqrt{x_1}-\frac{1}{\sqrt{x_1}}} Y^{(k)}(x_1,x_2) & = \frac{1}{x_1^k} \end{align} \end{subequations} where the boundary condition (\ref{bc4}) is explained as follows. The condition $x_2 \to -\sqrt{x_1}-\frac{1}{\sqrt{x_1}}$, according to (\ref{x12}), is equivalent to the condition that all the white balls are out of the box ($u \to \infty$). Then, in this case, the only solution for $k$ red balls to remain is if inside the box exist exactly $k$ red balls. In other words, the lowest order term in the Laurent expansion of $Y^{(k)}(x_1,x_2) $ should be $\sim 1/z^k$. The solution then to (\ref{pdebc4}) is given by\footnote{Analogously to the Rule III case, we could had imposed the alternative boundary conditions $\lim_{x_1 \to \infty} Y^{(k)}(x_1,x_2) =\frac{u}{u-1}\delta_0^k$ and still obtain the same $Y^{(k)}(x_1,x_2)$ given by equation (\ref{sol4}).} \begin{align}\label{sol4} Y^{(k)}(x_1(z,u),x_2(z,u)) = \sqrt{x_1} \left( \frac{x_2}{2} \left( \sqrt{1-\frac{4}{x_2^2}} - 1\right) \right)^{2k+1}. \end{align} \subsection{The probabilities for the first few cases, Rule IV} In this section we Laurent expand (\ref{sol4}) for $k=0,\,\,1,$ and $2$ and obtain explicit formulas for $p_{IV}^{(k)}(r,w)$. In particular, we first expand in inverse powers of $x_2$ and then re-expand in inverse powers of $z$ and of $u$ identifying the coefficients with the $f^{(k)}_{r,w}$'s (see (\ref{ans2})). \subsubsection{The case $k=0$} In fact, the solution to this case is already found in section (\ref{special}) and is (almost) given by equation (\ref{solp2}). The precise answer is \begin{align}\label{solp04} \boxed{ p_{IV}^{(0)}(r,w) = w\frac{1}{1+r} \, \frac{1}{w+r}, \forall \, r,\,w \geq 0} \end{align} where we extend the applicability of the formula in order to include $(r,w)=(0,0)$ in the view of (\ref{pc4}) and the discussion below equation (\ref{pa2e4}). One can check that equation (\ref{solp04}) satisfies the recursion (\ref{rec3}), and also fulfills the boundary condition (\ref{pb4}). \subsubsection{The case $k=1$} It can be shown that (\ref{sol4}) for $k=1$ expands as \begin{align} Y^{(1)}(x_1(z,u),x_2(z,u)) &= \sqrt{x_1} \left( \frac{x_2}{2} \left( \sqrt{1-\frac{4}{x_2^2}} - 1\right) \right)^{3} \\ \notag &= -3 \sqrt{x_1} \sum_{i=0}^{\infty} \frac{i \,\Gamma(2i+1)}{(i+1)(i+2)\Gamma^2(i+1)} \frac{1}{x_2^{2i+1}}. \end{align} The next step is to follow exactly the same steps that led from equation (\ref{intk1}) to equation (\ref{k1exp}) to eventually obtain \begin{align}\label{solp14} \boxed{ p_{IV}^{(1)}(r,w) =3 w \frac{ r }{(r+1)(r+2)}\, \frac{ (r+w+1)}{(r+w-1)(w+r)}, \forall \, r,\,w \geq 0.} \end{align} One can check that equation (\ref{solp14}) satisfies the recursion (\ref{rec3}), and also fulfills the boundary condition (\ref{pb4}). \subsubsection{The case $k=2$} It can be shown that (\ref{sol4}) for $k=2$ expands as \begin{align} Y^{(2)}(x_1(z,u),x_2(z,u)) &= \sqrt{x_1} \left( \frac{x_2}{2} \left( \sqrt{1-\frac{4}{x_2^2}} - 1\right) \right)^{5} \\ \notag &= -\frac{5}{2} \sqrt{x_1} \sum_{i=0}^{\infty} \frac{(i-1)\, i\, 2^{2i+1}\Gamma \left(i+\frac{1}{2}\right)}{\sqrt{\pi} \Gamma(i+4)}\frac{1}{x_2^{2i+1}}. \end{align} The next step is to work as in the $k=1$ case by following exactly the same steps that led from equation (\ref{intk1}) to equation (\ref{k1exp}) to eventually obtain \begin{align}\label{solp24} \boxed{ p_{IV}^{(2)}(r,w) =5 w \frac{(r-1)r }{(r+1)(r+2)(r+3)} \, \frac{(r+w+1)(r+w+2)}{(r+w-2)(r+w-1)(w+r)}, \forall \, r,\,w \geq 0.} \end{align} One can check that equation (\ref{solp24}) satisfies the recursion (\ref{rec3}), and also fulfills the boundary condition (\ref{pb4}). \subsection{The general solution} Equations (\ref{solp04}), (\ref{solp14}) and (\ref{solp24}) motivate the following ansatz for the general solution for any k \begin{align}\label{finallypka4} p_{IV}^{(k)}(r,w) = (2k+1)w \frac{(r-k+1)...(r-1)r }{(r+1)(r+2)...(r+k+1)} \, \frac{(r+w+1)(r+w+2)...(r+w+k)}{(r+w-k)...(r+w-1)(w+r)}, \end{align} which can be re-written as \begin{align}\label{finallypk4} \boxed{p_{IV}^{(k)}(r,w) =(2k+1)\frac{ r!(r+w+k)!}{(r+k+1)!(w+r)!}\,\, \frac{wr!(r+w-k-1)!}{(r+w)!(r-k)!},\,\, \forall\, r,\,w,\,k\geq0,\, k\leq w.} \end{align} One may check that (\ref{finallypk4}) satisfies both, the recursive equation (\ref{rec3}) and the boundary condition (\ref{pb4}), and in fact, all the consistency-check equations (\ref{pb4})-(\ref{pe4}). As a cross check, we verified that equation (\ref{finallypk4}) reproduces (\ref{solp04}), (\ref{solp14}) and (\ref{solp24}) and most importantly, it satisfies the normalization condition \vspace{-0.2in} \begin{align}\label{normalp4} \sum_{k=0}^r p_{IV}^{(k)}(r,w) = 1. \end{align} We note that the normalization equation (\ref{normalp4}) comes out automatically. This key fact serves as another cross check of (\ref{finallypk4}). At this stage, the solution is considered as complete. Two important observations can be made: (a) It is notable that (\ref{finallypk4}) is, up to the rescaling factor $\frac{(2k+1) r!(r+w+k)!}{(r+k+1)!(w+r)!}$, the same as the solution of Problem 2, equation (\ref{p1}) with the (expected) reflection $ r \leftrightarrow w$. This implies that this overall factor, in a sense, encodes the additional complication of adding red balls back into the box. (ii) One can also show that the (true) generating probability functional $\tilde{Y}^{(k)}(z,u) = \tilde{Y}^{(k)}_{IV}(z,u)$ defined in (\ref{Yt}) satisfies the PDE (\ref{pde224}) with $p^{(k)}(r,w)=p_{IV}^{(k)}(r,w)$ given by the probability formula (\ref{finallypk4}). We conclude the section by providing the probability generating functional \vspace{-0.1in} \begin{align}\label{Gp1IV} \hspace{-0.03cm} \hspace{-0.1cm}G_{IV}(r,w;\hspace{-0.06cm}z)& \equiv \sum_{r=0}^kp_{IV}^{(k)}(r,w)z^k = \frac{w}{(1+r)(r+w)} \Big[ \,_3F_2\left(1,-r,1+r+w;2+r,1-r-w ;z\right)\\ \notag &+ 2r \frac{1+r+w}{(2+r)(r+w-1)} z \,_3F_2\left(2,1-r,2+r+w;3+r,2-r-w ;z\right)\Big] \end{align} for fixed $r$ and $w$. The generating functional involves a linear combination of $\,_3F_2$ hypergeo- metric functions rather than a single $\,_3F_2$, which was the case for Problem 1, Rule III (see (\ref{Gp1III})). \subsubsection{Maximal probability} This section investigates the set of $k$'s for which $p_{IV}^{(k)}(r,w)$ is maximized. For the boundary case $w=0$ we have $p_{IV}^{(k)}(r,0)=\delta_r^k$ (see (\ref{pb4})) and hence, the probability is maximized for $k=r$ with maximal value $p_{IV}^{(r)}(r,0)=1$. The other boundary case is when $r=0$ in which case $p_{IV}^{(k)}(0,w)=\delta_0^k$ (see (\ref{pbb4})) and hence, the probability is maximized for $k=0$ with maximal value $p_{IV}^{(0)}(0,0)=1$. Otherwise, if $w>0$ and $r>0$, we work analogously to section \ref{max3} and we find that the set of $k$'s that maximize (locally) the $p_{IV}^{(k)}(r,w)$ is specified by the two inequalities \begin{subequations} \begin{align} k&\geq \sqrt{\frac{(1+r)(r+w)}{2w-1}}-1, \,r,\,w \geq1, \\ k & \leq \sqrt{\frac{(1+r)(r+w)}{2w-1}},\,r,\,w \geq1. \end{align} \end{subequations} We observe that the inequalities are linear in $k$, which means that (modulus degeneracies) there exists only one unique maximum. Before investigating the maxima, we define $k_0$ by \begin{align} \label{k0} k_0 \equiv \sqrt{\frac{(1+r)(r+w)}{2w-1}}. \end{align} Given that the investigation here is for $r \geq1$ and $w \geq1$, it is deduced that $k_0>1$. We now want to constrain $k_0$ from above. As it can be shown, there are two cases: Case 1. If $w=1$, then $k_0=r+1$ and hence, the maximum occurs for $k=r$. Case 2. Else if $w>1$, then $k_0<r+1$. Case 2 is partitioned into two sub-cases: Case 2a. If $k_0 \in \mathbb{Z}_+$ then there are two maxima that correspond to $k=k_0-1$ and to $k=k_0$, which occur with equal probabilities. Case 2b. Otherwise if $k_0 \notin \mathbb{Z}_+$ then there exists a unique maximum that corresponds to the integer between $k_0 -1$ and $k_0$; that is for $k= \left \lfloor{k_0}\right \rfloor $ (the integer part of $k_0$). This is always the case for $r=1$ and $w>1$; the maximum occurs for $k=1$ because $k_0 \notin \mathbb{Z}_+$ with $k_0 \in (1,2)$. Collecting all cases together, we have just proved \begin{subequations} \label{maxpk4} \begin{empheq}[box=\widefbox]{align} \label{maxpk4a} &sup\left(p_{IV}^{(k)}(r,w=0)|k=0,\,1,\,...\,r \right)=r,\\ \label{maxpk4b} &sup\left(p_{IV}^{(k)}(r=0,w)|k=0 \right)=0,\\ \label{maxpk4c} &sup\left(p_{IV}^{(k)}(r \geq 1,w=1)|k=0,\,1,\,...\,r \right)=r,\\ \label{maxpk4d} &sup\left(p_{IV}^{(k)}(r\geq1,w \geq2)|k=0,\,1,\,...\,r \right)=\{k_0-1,k_0 \}, k_0\in \mathbb{Z}_+,\\ \label{maxpk4e} &sup\left(p_{IV}^{(k)}(r \geq1,w \geq2)|k=0,\,1,\,...\,r \right)= \left \lfloor{k_0}\right \rfloor, \, k_0 \notin \mathbb{Z}_+. \end{empheq} \end{subequations} This ends our investigation of the maxima of $p_{IV}^{(k)}(r ,w )$. \subsubsection{Limiting cases} In this section we investigate the behavior of (\ref{finallypk4}) in the limits $w\to \infty$ for $r$ fixed and any $k$, and for $r \to \infty$ for $w$ and $k$ fixed. For this purpose we use the more convenient equation (\ref{finallypka4}) from where it is deduced that \begin{subequations} \label{limcases4} \begin{align} \label{limcasesr4} p_{IV}^{(k)}(r,w) &= (2k+1) \frac{w}{r^2}+ O\left(\frac{1}{r^{3}} \right)= (2k+1)p_{IV}^{(0)}(r,w) + O\left(\frac{1}{r^{3}} \right),\,\, r \to \infty, \, r \gg w,\,k,\\ \label{limcasesw4} p_{IV}^{(k)}(r,w) &=(2k+1)\frac{(r!)^2}{(r-k)!(r+k+1)!}+ O\left(\frac{1}{w} \right),\,\, w \to \infty, \, w \gg r. \end{align} \end{subequations} The asymptotic expansion (\ref{limcasesr4}) says that in the limit $r \to \infty$ with $w$ and $k$ kept fixed, the probability $p_{IV}^{(k)}(r,w)$ decays as $\frac{1}{r^{2}}$ independently on $k$ \footnote{Modulus the $2k+1$ overall factor. More precisely, in this limit, the probabilities grow linearly in $k$ according to $(2k+1)p_{IV}^{(0)}(r,w)$; they grow as integer multiples of a fundamental quantity, the probability $p_{IV}^{(0)}(r,w)$.} contrary to the case of Rule III, equation (\ref{limcasesr}). Moreover, the asymptotic expansion (\ref{limcasesw4}) in the limit $w \to \infty$ with $r$ kept fixed is also interesting. In particular, the probability $p_{IV}^{(k)}(r,w)$ (to leading order in $w$) becomes independent on the initial number of the white balls $w$ analogously to the behavior of the Rule III case, equation (\ref{limcasesw}). In particular, if $r=1$, then $p_{IV}^{(0)}(1,w)\approx p_{IV}^{(1)}(1,w) \approx \frac{1}{2}$ for $w \gg 1$. \section{Discussion}\label{diss} Wow!! That has been a long but joyful journey for such a rather simple problem. One could think that our approach amounts in shooting a bug with a bazooka! While such a view might be right, in this work we have seen a concrete example of how a recursion equation arising from a probability problem, without a systematic method to approach, was transformed into a differential equation that was solved using standard PDE methods. Hence, we have seen a concrete example of how difference equations, such as those arising from discrete random processes, and for which the tools in the literature are less developed, could be transformed into differential equations, where the literature is rich and well studied. In particular, the approach we followed has been algorithmic and it may be summarized by the following recipe: \vspace{0.2in} \hspace{0.4cm} {\bf Step 1.} Usage of the total probability law in order to derive a difference equation (rec- \hspace{2.1cm} ursion)\hspace{-0.04cm} in\hspace{-0.04cm} one \hspace{-0.04cm}or \hspace{-0.04cm}more \hspace{-0.04cm}variables.\hspace{-0.04cm} Impose \hspace{-0.04cm}suitable \hspace{-0.04cm}discrete \hspace{-0.04cm}boundary \hspace{-0.04cm}conditions. \vspace{0.2in} \hspace{0.4cm} {\bf Step 2.} Apply z-transformations in order to transform the recursion into a differential \hspace{2.1cm} equation for the probability generating functional. \vspace{0.2in} \hspace{0.4cm} {\bf Step 3.} Translate the discrete level boundary conditions of Step 1 into continuum \hspace{2.1cm} boundary conditions at the (P)DE level. \vspace{0.2in} \hspace{0.4cm} {\bf Step 4.} Solve the (P)DE enforcing the boundary conditions of step 3. \vspace{0.2in} \hspace{0.4cm} {\bf Step 5.} Laurent expand the solutions taking care of any fictitious infinities that may \hspace{2.1cm}appear by taking suitable limits. Hence, obtain the required probabilities. \vspace{0.2in} This recipe is adaptable to large classes of probability problems, and generally for problems involving difference equations. For instance, adapting the ideas of Problem 1, Rule III, to those of Rule IV, whose results have been obtained fast and straightforwardly, has been effortless. Retracing the steps (for Problem 1, Rule III), we started from a probability question and derived a recursion, equation (\ref{rec3}), with two variables using the law of total probability. Then, through a 2D $z-$transformation (\ref{Yfdef}), we eventually derived a first order 2D PDE (\ref{pdeG}) with suitable source terms and boundary conditions (\ref{odep1}) whose solution, up to a factor (see (\ref{ans2})), provides the generating probability functional of the problem at hand for any number of remaining white balls $k$. Given that a first order 2D PDE is equivalent to a set of two first order ODEs, the PDE, through suitable coordinate transformations (see (\ref{x12})), was reduced into two decoupled first order ODEs, equations (\ref{dudz}) and (\ref{pdeG}). This step, in a sense, decouples the 2D problem into two 1D problems of suitable scaling variables, which are functions of the initial ones. While the two scaling variables are decoupled at the DE level, they are yet coupled in a non-trivial way through the boundary conditions (i.e. see (\ref{solk})). Solving then the differential equation, and hence, obtaining the general generating functional (\ref{solk}), we performed a closed form Laurent expansion. Using the formula from the expansion, which provides the coefficients $f^{(k)}(r,w)$ (see (\ref{Yfdef}) and (\ref{ans2})), we were eventually able to find the general probability formula, equation (\ref{finallypk}). Finally, we checked that the probability formula satisfies the initial recursion equation we begun with, and its boundary conditions (\ref{pb}), thus completing the solution of the problem. An analogous approach has been followed in Problem 1, Rule IV. It is notable how a relatively simple probability problem to state and to understand, has involved such a heavy computational machinery. The most challenging step has been to Laurent expand the generating functional and to deal with the fictitious infinities that appeared on the way due to the interchanging of the summation orderings. These infinities showed up as poles of Gamma functions; this part has been considerably more involved than solving the PDE itself. We have also seen that the probability generating functionals involve hypergeometric functions (see (\ref{Gp1III})) or linear combinations of hypergeometric functions (see (\ref{Gp1IV})) of type $_3F_2$. These results generalize the simpler version of the problem, which does not require placing balls back into the box, and, which involves a $_2F_1$ distribution instead (see (\ref{Gp2})). Hence, in a sense, the additional complication of adding balls back into the box is captured by extending the $_2F_1$ type of the probability generating functional into a $_3F_2$ type. As a bonus, we found the power series solution of the rather complicated family of PDEs given by equation (\ref{pde22}), which satisfy the same boundary conditions $\tilde{Y}_{III}^{(k)}(z \to \infty,u) =1/u^k$ as the boundary conditions satisfied by $Y_{III}^{(k)}(z,u) $ (see (\ref{odep1b})). The solution is given by $\tilde{Y}_{III}^{(k)}(z,u) = \sum_{r,w=0}^{\infty}p_{III}^{(k)}(r,w) \frac{1}{z^r} \frac{1}{u^w} $. Likewise, for Rule IV, $\tilde{Y}_{IV}^{(k)}(z,u) = \sum_{r,w=0}^{\infty}p_{IV}^{(k)}(r,w) \frac{1}{z^r} \frac{1}{u^w} $ is the solution of the PDE (\ref{pde224}), which satisfy the same boundary conditions $\tilde{Y}_{IV}^{(k)}(z,u \to \infty,u) =1/z^k$ as the boundary conditions satisfied by $Y_{IV}^{(k)}(z,u) $ (see (\ref{bc4}) and (\ref{x12})). It would be interesting to find simple probability problems, such as the ones presented in this work, whose differential equation representation is the same as that of known problems from other areas of mathematics or physics or even from finance. That would provide a sort of duality between a physically meaningful and involved problem, and a rather simple probability problem such as a box of balls of various colors. \section*{Acknowledgments} I would like to thank A. Kryftis for keep challenging me and in particular, A. Anastasiou (LSE and Un. of Cyprus) for reading the manuscript and for providing insightful comments, and D. Christofides (UCLAN, Cyprus) for providing the solution of appendix \ref{Gminus}, and also S. Agapiou (Un. of Cyprus), S. Hormann (Graz Un. of Technology, Austria), T. Bruss (Universite Libre de Bruxelles), and E. Mossel and S. Sheffield (MIT) for referring me to useful sources. I also thank my friends at Vincent House (London), and especially G. Vogiatzi for encouraging me to publish this work, and my colleagues at JPMorgan Chase, and in particular, S. El Hamoui, A. Eriksson, M. Green, J. Lorenzen and S. Mcgarvie for engaging in stimulating discussions\footnote{Disclaimer: \hspace{-0.04cm}The current \hspace{-0.04cm}paper \hspace{-0.04cm}consists of a personal \hspace{-0.04cm}work \hspace{-0.04cm}curried out by \hspace{-0.04cm}A. \hspace{-0.04cm}Taliotis, \hspace{-0.04cm}and it expresses \hspace{-0.02cm}his \hspace{-0.02cm}own personal \hspace{-0.04cm}views on the \hspace{-0.04cm}particular topic. \hspace{-0.04cm}In particular, \hspace{-0.04cm}this \hspace{-0.04cm}work is not, in \hspace{-0.04cm}any way, \hspace{-0.04cm}endorsed \hspace{-0.04cm}by \hspace{-0.04cm}or \hspace{-0.04cm}related with \hspace{-0.04cm}JPMorgan \hspace{-0.04cm}Chase (the firm) or its employees or stakeholders or any interests or entities represented by the firm.}. A warm thank you to Staxto for ``being here'', during the long nights after the JPMC office hours, while preparing this write-up. I would like to express my deep gratitude to my teacher and good friend Y. Kovchegov from The Ohio State University for patiently showing me, among other, how to carry on long and tedious calculations and how to, in a quantum field theoretical language, ``renormalize" (deal with) infinities, and eventually obtain finite and meaningful results. Lastly, I would like to thank Nikolas and Leo-Anastasis for invading into my routine teaching me the beauty in life, and for all the time I have taken away from them.

https://arxiv.org/abs/1711.03065

An Application of Mosaic Diagrams to the Visualization of Set Relationships

We present an application of mosaic diagrams to the visualisation of set relations. Venn and Euler diagrams are the best known visual representations of sets and their relationships (intersections, containment or subsets, exclusion or disjointness). In recent years, alternative forms of visualisation have been proposed. Among them, linear diagrams have been shown to compare favourably to Venn and Euler diagrams, in supporting non-interactive assessment of set relationships. Recent studies that compared several variants of linear diagrams have demonstrated that users perform best at tasks involving identification of intersections, disjointness and subsets when using a horizontally drawn linear diagram with thin lines representing sets, and employing vertical lines as guide lines. The essential visual task the user needs to perform in order to interpret this kind of diagram is vertical alignment of parallel lines and detection of overlaps. Space-filling mosaic diagrams which support this same visual task have been used in other applications, such as the visualization of schedules of activities, where they have been shown to be superior to linear Gantt charts. In this paper, we present an application of mosaic diagrams for visualization of set relationships, and compare it to linear diagrams in terms of accuracy, time-to-answer, and subjective ratings of perceived task difficulty. The study participants exhibited similar performance on both visualisations, suggesting that mosaic diagrams are a good alternative to Venn and Euler diagrams, and that the choice between linear diagrams and mosaics may be solely guided by visual design considerations.

\section{Introduction} \label{sec:introduction} The study of sets and their relationships is fundamental to the disciplines of mathematics, logic, and computer science. Visual representations of relationships among sets --- intersection, containment, and exclusion (disjoint sets) --- have been used for centuries. However, the development of interactive visualizations and tools in recent years has gained a new impetus due to the wide range of applications that these tools find in a variety of areas, including the analysis of healthcare and population data, representation of relationships in social networks, and the study of consumer purchasing patterns, to name a few. Visual representation of sets and their relationships is most commonly done through Venn and Euler diagrams \citep{bib:Baron1969}. These types of diagrams, however, have well known limitations \citep{bib:Rodgers2014,bib:Gottfried2014,bib:Gottfried2015}. They generally do not scale well beyond a small number of sets, and present usability problems. Automatic drawing of Venn and Euler diagrams is also problematic \citep{bib:RicheAndDwyer2010,bib:FlowerEtal2014,bib:Simonetto2016}. In response to these limitations, alternative set visualization techniques have been proposed. Linear diagrams \citep{bib:Gottfried2014}, which are of particular interest here, have been shown to compare favourably to Venn and Euler diagrams in terms of task completion time and the number of errors made by users \citep{bib:ChapmanEtal2014}, for example in tasks involving syllogistic reasoning \citep{bib:SatoAndMineshima2012}. In a recent paper \cite{bib:RodgersEtal2015} compared several versions of linear diagrams produced by varying essential properties of their corresponding {\em retinal} and {\em planar} variables \citep{bib:Bertin67Semiologie}. Their study concluded that users perform best at tasks involving identification of intersections, disjointness (exclusion) and subsets (inclusion) among sets when using a horizontally drawn linear chart with thin lines representing sets, and vertical guide lines for aiding the detection of alignment across the vertical axis. The essential visual task the user needs to perform in order to interpret this kind of linear diagrams is a Vernier acuity task, which basically requires vertical alignment of the beginning or end of a horizontal line with those of other lines above or below. In this paper, we present a study comparing linear diagrams with a space-filling alternative visualization based on mosaic diagrams, \citep{bib:LuzMasoodian2007} for representing set relationships as examined by \cite{bib:RodgersEtal2015}. The proposed mosaic diagrams (as shown below) employ a space-filling algorithm whereby intersections are denoted by shared areas (represented in different colours), subset relations are denoted by area containment, and exclusions are represented as uniformly (i.e. single) coloured areas. The primary motivation for this study was the fact that mosaic diagrams have previously been used in other applications (e.g. visualization of task schedules) where they have been shown to be superior to linear-style diagrams such as Gantt charts \citep{bib:Gantt19}. The current study therefore investigates whether this superiority of mosaic over linear diagrams also holds true in the case of set visualization. To the best of our knowledge, this study is the first to compare mosaic and linear diagrams in set visualization tasks. As such, it aims specifically at comparing static representations of set relations, and using two representations both based on linear structure, albeit employing different instantiations of retinal and planar variables (see further discussion below), rather than providing an exhaustive comparison of set visualisation methods based on disparate principles, or replicating previous comparisons. Thus it does not compare mosaic or linear diagrams to other static representations such as Euler diagrams, Venn diagrams or their modern variants described below, as comparisons between linear diagrams and Euler and Venn diagrams have been reported elsewhere \citep{bib:Gottfried2015,bib:ChapmanEtal2014}. In particular, as regards variants such as Bubble Sets \citep{bib:CollinsEtal2009,bib:RicheAndDwyer2010} and LineSets \citep{bib:AlperEtal2011}, these techniques are unlike linear (and mosaic) diagrams, which display only abstract set relations, in that they ``require the existence of embedded items'', as pointed out by \cite{bib:RodgersEtal2015}. Similarly, this study does not compare mosaics to the many interactive set visualization systems proposed in the burgeoning literature on this topic. The reader is referred to these works and to the literature review below for comparisons of interactive systems in terms of their design features \citep{bib:YalcinEtal2016} and task taxonomies \citep{bib:AlsallakhEtal2015}. While empirical studies of interactive versions of mosaic (and indeed linear diagrams) are of great practical interest for future work, comparisons of this kind lie beyond our scope here. This paper contributes to the information visualization literature by providing an analysis of time-on-task, accuracy and subjective difficulty ratings for each of these two linearly-structured visualizations, with essentially comparable forms of set representation. In addition to its empirical findings, the paper also discusses the relative advantages and disadvantages of mosaic and linear diagrams in terms of their design, including their potential uses as compact overviews of sets, and their ability to represent other properties (e.g. cardinality) of sets beyond the basic set relationships investigated in the current study. \section{Background} \subsection{Set visualizations} \label{sec:lit-review} Set visualization is a common and increasingly important task. Not surprisingly, a wide range of set visualization techniques have been proposed over the years. \cite{bib:AlsallakhEtal2015,bib:AlsallakhEtal2014} provide a comprehensive review of set visualizations in their state-of-the-art report. They classify set visualizations into six categories: \begin{enumerate} \item \emph{Euler and Venn diagrams:} As mentioned, these visualizations are the most common representations of sets, and a large number of variations have been designed to improve them. For surveys, see \cite{bib:Rodgers2014} and \cite{bib:RuskeyAndWeston2005}. \item \emph{Overlays:} These techniques present set memberships as secondary information over other visualizations (e.g. spatial, or temporal) which provide the context for analysis. These include the popular LineSets \citep{bib:AlperEtal2011}, Bubble Sets \citep{bib:CollinsEtal2009}, Kelp diagrams \citep{bib:MeulemansEtal2013}, and TimeSets \citep{bib:NguyenEtal2016}. \item \emph{Node-link diagrams:} These techniques represent relationships between sets and their members as edges of bipartite graphs whose nodes are the sets and elements. Node-link diagrams are considered to be easy to understand, and allow visual encoding of further information in representation of the nodes (i.e. each element or set). Node-link visualizations can also be combined with other representations such as matrix-based (e.g. OnSet \citep{bib:Sadana2014}), or aggregation-based (e.g. Radial Sets \citep{bib:AlsallakhEtal2013}) representations. \item \emph{Matrix-based techniques:} These visualizations use the matrix representation to show sets or set members as elements of matrices. Examples of this type of visualizations include UpSet \citep{bib:LexEtal2014}, and OnSet \citep{bib:Sadana2014}. \item \emph{Aggregation-based techniques:} Unlike some of the above mentioned techniques, aggregation-based visualizations do not aim to represent the relationships between individual elements of the sets involved. Instead, set elements are aggregated into their respective sets, and only relationships between those sets are represented. As such, aggregation-based techniques are more suitable for representing relationships between sets with large number of elements, where it would be impractical to show all the relationships between those elements. Examples of these techniques include AggreSet \citep{bib:YalcinEtal2016}, Radial Sets \citep{bib:AlsallakhEtal2013}, and PowerSets \citep{bib:AlsallakhAndRen:2017}. \item \emph{Other techniques:} There are also a range of other set visualization techniques, such as Scatter plots (e.g. scatter view and cluster view\citep{bib:AlsallakhEtal2015}), which represent set relationships using other visual methods than those described in the above categories. These include techniques such as bargrams, which resemble linear diagrams in some aspects but incorporate other extensions, such as set-valued attributes \citep{bib:WittenburgMaliziaEtAl12vsatp}, and can be categorised as frequency-based. \end{enumerate} More specifically, linear diagrams \citep{bib:Gottfried2014} fall into the category of aggregation-based techniques. Linear diagrams have been shown to be more effective than region-based representation such as Euler and Venn diagrams \citep{bib:ChapmanEtal2014}, which tend to be more cluttered due to overlapping, coincident, and tangentially touching contours, as demonstrated in an empirical study \citep{bib:Gottfried2015}. As mentioned earlier, \cite{bib:RodgersEtal2015} have also conducted a series of studies which compared the effectiveness of linear diagrams against Euler and Venn diagrams, as well as different variations of linear diagrams themselves, for preforming tasks requiring visualization of set relationships. These studies have shown that linear diagrams are superior to Euler and Venn diagrams for identification of set intersections, containment, and exclusions. They have also led to a number of visual design principles for creating more effective linear diagrams. These include: a) \emph{the use of a minimal number of line segments}, b) \emph{the use of guide lines where line overlaps start and end}, and c) \emph{the use of lines that are thin as opposed to thick bars} \citep{bib:RodgersEtal2015}. The effectiveness of these principles was demonstrated through a final study \citep{bib:RodgersEtal2015}, which we utilize in our own study, presented in this paper. \subsection{Mosaic diagrams} \label{sec:mosaic} Mosaic diagrams were originally proposed by Luz and Masoodian \cite{bib:LuzMasoodianAVI04,bib:LuzMasoodian2007} as an alternative to conventional timelines for visualization of temporal streams of media --- in their case, recorded during multimedia meetings. As shown in Figure~\ref{fig:timeline-mosaic-diagrams}, unlike timeline visualization which reserves horizontal rows for each data stream (Figure~\ref{fig:timeline-mosaic-diagrams}a), the mosaic visualization uses a pre-specified vertical space proportionally between only those streams which occur at that specific point in time (Figure~\ref{fig:timeline-mosaic-diagrams}b). \begin{figure}[!t] \centering \includegraphics[width=0.45\linewidth]{timeline.pdf}~a) \includegraphics[width=0.45\linewidth]{mosaic.pdf}~b) \caption{Visualization of 8 media streams (4 voice and 4 text) using a) timeline, and b) mosaic diagrams (from \citep{bib:LuzMasoodian2007}).} \label{fig:timeline-mosaic-diagrams} \end{figure} The mosaic visualization has also been used for representation of event schedules\cite{bib:LuzMasoodian2011}, in a manner similar to standard Gantt charts. A study \citep{bib:LuzMasoodian2011} comparing static Gantt charts and mosaic diagrams has shown that mosaic diagrams match Gantt charts, in terms of speed and accuracy, for all types of tasks requiring detection of relationships between schedule events (e.g. durations and overlap of events). Due to the similarity between Gantt charts and linear diagrams, we decided to investigate the use of mosaic diagrams as a potential alternative to linear diagrams for visualization of set relationships. In this form, mosaic diagrams are employed as an aggregation-based set visualization technique. Figure~\ref{fig:3sets} provides an example of the use of mosaic diagrams (\ref{fig:3sets}c) to represent set relationships, in comparison to Euler (\ref{fig:3sets}a) and linear (\ref{fig:3sets}b) diagrams. In this example, three sets of people are interested in books, technology, and cars. As can be seen, some people are interested only in books, some only in cars, some only in books and technology, and some in all the three categories. Furthermore, everyone who is interested in technology is also interested in books. \begin{figure}[tbp] \centering a)\includegraphics[width=0.3\linewidth]{3sets-euler.pdf} b)\includegraphics[width=0.3\linewidth]{3sets-linear.pdf} c)\includegraphics[width=0.3\linewidth]{3sets-mosaic.pdf} \caption{Relationships between three example sets, shown using a) Euler, b) linear, and c) mosaic diagrams.} \label{fig:3sets} \end{figure} \subsection{Visual Variables and Perceptual Tasks} \label{sec:visual_variables} Both mosaic and linear diagrams are in essence linearly-structured on a two-dimensional plane. In terms of Bertin's graphic sign system \citep{bib:Bertin81g,bib:Bertin67Semiologie} size and planar position can be used to convey association. However, while for linear diagrams these two variables would in principle suffice to communicate the relevant set relations (intersection, disjointness and subset), mosaics cannot avail of the alignment between horizontal bars and set labels the way linear diagrams do. Therefore, mosaics need to employ a further variable to distinguish the different signs for individual sets. As there are typically many sets to label, and since colour is generally recommended for label encoding \citep{bib:Ware12}, the colour hue attribute was chosen as the differentiating sign in mosaics. It should also be noted that \cite{bib:RodgersEtal2015} also considered colour as a variable in their evaluation of linear diagrams, but their results showed no significant differences in performance between colour-coded and monochrome diagrams. The use of colour places some constraints on mosaic diagrams. Notably, it limits the number of sets that can be encoded to the number of colours that can be reliably distinguished from each other if colour continuity issues are to be avoided. A study by \cite{bib:Healey96c} places this limit at 10 distinct hues. In order to maximise contrast in the mosaic one should not choose a colour that lies in the convex hull (in a uniform colour space) of the colours already in use. Thus a suitable set of colours might be, for instance, the edges of a convex hull in the CIEluv space \citep{bib:Ware12}. The use of high-saturation colours would also help improving discrimination of mosaic areas, as would the addition of thin, high luminance contrast boundaries to the different tiles. As will be discussed below, in the study reported here, we limited the use of colours to those colours used in the experiments of \cite{bib:RodgersEtal2015} in order to reduce the possibility of introducing confounds in the conditions we compared. In terms of perceptual tasks, viewers rely on their ability to verify the alignment of lines accurately in interpretation of linear and mosaics diagrams. As such, both types of diagrams benefit from (and to some extent depend on) the hyperacuity characteristic of the human visual perception \citep{bib:Westheimer09h}. This allows viewers to perform alignment tasks, as well as comparing length of lines, very effectively, even in small diagrams. Unfortunately however, performance on such tasks is known to degrade significantly if the lines to be compared are placed too far apart in the visual space, or when that space is crowded by intervening lines \citep{bib:LeviKleinAitsebaomo85v}. Furthermore, comparisons also become more challenging in the absence of contrast between the lines and their surrounding visual context (i.e. the background visual space) \citep{bib:Westheimer09h,bib:SayimWestheimerHerzog08c}. These factors have indeed contributed to, and demonstrated through empirical studies, to suggestions made by \cite{bib:RodgersEtal2015} for generating the most effective visual variants of linear diagrams for visualization of set relationships, as discussed previously. Therefore, we speculated that mosaic diagrams may be more effective than linear diagrams for Vernier acuity tasks due to their space-filling characteristic. This would make visual tasks such as identifying set relationships easier in mosaic diagrams, where background visual space is often filled using the colour(s) associated with set(s) of interest, unless of course when there are no relationships between sets which is much less likely in such visualizations. This space-filling characteristic also allows spaces associated with sets of interest to join one another not only horizontally, but more importantly vertically; making it easier to perform vertical alignment tasks. It should however be pointed out that, as is often the case in visualizations, there is a trade off in adding this space-filling visual element. In this case, space-filling creates shapes of different colours, which in turn can reduce detection of continuity of lines. Although continuity is important, and according to Gestalt laws should be preserved \citep{bib:Ware12}, mosaic relies on another powerful Gestalt principle, namely closure \citep{bib:Ware12}, to allow easier detection of individual sets by creating uniquely coloured shapes for each set. Finally, as a side note, it should be mentioned here that although another aggregation-based set visualization technique, called Mosaic Plots, has previously been proposed \citep{bib:HartiganAndKleiner1981,bib:Hofmann2000}, this technique is rather different from the use of mosaic diagrams as demonstrated here. Mosaic Plots are a combination of Spine Plots and bar charts, designed to allow representation of relationships between groups of sets --- e.g. two gender sets, and five age group sets for accident victims, as discussed by \cite{bib:Hofmann2000} --- rather than direct representations of relationships between individual sets as is the case of the mosaic diagrams investigated here. \section{Evaluation} In order to compare the effectiveness of mosaic and linear diagrams for visualization of set relationships, we adopted the same set of tasks used on the multiple comparisons of linear diagram variants carried out by \cite{bib:RodgersEtal2015}. As in that study, the diagrams used in our study were derived from the Twitter graph dataset available through the SNAP project \citep{bib:LeskovecKrevl14sd}. The variant of linear diagrams used in our comparisons was the variant found to be the most effective \citep{bib:RodgersEtal2015}. This variant uses: a) \emph{heuristically minimized number of segments}, and b) \emph{thin horizontal lines} for representing sets. These lines are distinguished from each other though the use of colour, and placed on a grid of guide lines meant to facilitate visual alignment (see the linear diagram shown in Figure~\ref{fig:tutorial-screenshot}, for instance\footnote{All the content used in the evaluation is available at \url{http://removed}}). In order to standardize the labelling in the linear diagrams with respect to mosaic diagrams for experimental comparison, the same legends were used in both diagram types. These legends preserve the line ordering of the original linear diagrams. The mosaic diagrams that were generated each corresponded to the linear diagram used in the final experiment of \cite{bib:RodgersEtal2015}, except that we standardized the number of sets to six in all tasks. We replicated the linear diagrams manually, and used a version of the freely-available Chronos software \citep{bib:LuzMasoodian2011 to produce the corresponding mosaic diagrams. All images were produced in PNG format, using the same size, colour combination, and resolution used by Rodgers et al. for their linear diagrams. Identical settings were employed in the production of the corresponding mosaic diagrams. \subsection{Methodology} \label{sec:methodology} Unlike \cite{bib:RodgersEtal2015}, who employed a between-subject design and collected their data through crowd-sourcing, we used a within-subject design, administered through a bespoke Java application, and recruited our participants locally by personal invitation in each of our respective universities. This alternative experimental set-up was adopted in order to enable us to recruit a smaller number of more suitable participants, and exercise better validation and control over experimental conditions and measurements. The choice of a within-subject (repeated measures) design was made because it allows each participant to experience each of the alternative visualizations under test (i.e. mosaic and linear) repeatedly, thus mitigating the effects of any potential inter-participant variations, and allows a smaller number of participants usually to reveal the relevant differences, should such differences exist. Well known shortcomings of this kind of repeated measures design were also addressed. Specifically carry-over effects were mitigated by alternation of the two conditions, as well as replications with the opposite alternation ordering (see Table~\ref{tab:questions-sets}), and practice effects were accounted for by the ordering of tasks from easy to difficult, again in alternation. Furthermore, the use of a specially designed application for the study enabled us to obtain precise answer timings, as well as collecting subjective task difficulty ratings. Answer time and ratings allowed us to compare the alternative visualizations in more detail, for instance in terms of the difficulties perceived by participants when performing similar tasks using each of the visualizations. This is in addition to the measures used by Rodgers et al. In this experiment we considered three factors, with the following possible levels: \begin{itemize} \item 2 visualization types: (L)inear vs. (M)osaic \item 3 task types: (I)ntersection, (S)ubset, and (D)isjunction \item 2 levels of difficulty: \begin{itemize} \item (E)asy: where the task involves identifying subsets, sets that intersect with, or sets that are disjoint from a set $X$, \item (H)ard: where the task is to identify subsets or sets that intersect with $X \cup Y$, or sets that are disjoint from $X \cap Y$. \end{itemize} \end{itemize} In order to make our study comparable to that of Rodgers et al., we adopted the same combinations used by them for two of these factors, namely task types and difficulty levels. Each participant was requested to answer 12 ($2 \times 3 \times 2$) task questions: 6 questions against different mosaic diagrams (MEI, MES, MED, MHI, MHS, MHD), and 6 questions against different linear diagrams (LEI, LES, LED, LHI, LHS, LHD). Each diagram used in the study depicted a collection of 6 sets and their relationships. Each question referred to a different collection of sets. These 6-set collections were drawn from a larger collection of 24 possible sets. The number of pairwise set relations (intersections, disjointness and subsets) for all sets used in this experiment, along with their respective mosaic and linear diagrams are shown in Table~\ref{tab:task-questions-images}. On average, taken in pairs, these sets contain 8.4 (SD=3.6) intersection, 6.1 (3.5) disjointness, and 1.6 (1.5) subset relations. The numbers of elements in these sets were left unspecified, as we were only interested in assessing abstract set relations, which are immediately supported by linear diagrams and their mosaic equivalents. However, see the discussion section for an example of how mosaics could support visualisation of proportional cardinality relations through a simple modification. Although irrelevant to this study, exact cardinality and composition of the sets used can be retrieved from the SNAP project website.\footnote{https://snap.stanford.edu/} As mentioned, the task questions were presented in alternation (a mosaic diagram following a linear diagram or vice-versa). In order to mitigate potential order effects, we distributed the questions so that a task was never followed by another task of the same type. Participants were assigned automatically by the system to one of the task question sets shown on Table~\ref{tab:questions-sets}, so as to ensure a balanced set of answers. Thus, for instance, on the first series, LEI (an Easy Inclusion task, with sets represented as a Linear diagram) is followed by a different type of task (an Easy Disjointness task) with sets represented as a Mosaic diagram (MED). The presentation sequences also contain no consecutive presentation of the same type of tasks (I, D, S). As regards difficulty level, we kept a fixed ordering whereby easier questions preceded harder questions, as mentioned earlier. Since this ordering is consistent across the two visualization types (i.e. experiment conditions), task difficulty should not affect the comparisons made between the two conditions. The results reported later in this paper showed that our labelling of tasks according to difficulty level conformed to the participants' levels of performance and subjective perceptions of difficulty. \begin{table}[!t] \caption{The two replications of task questions in terms of the sets and diagrams used in the study. L=linear diagram, M=mosaic, E=easy question, H=hard question, I=intersection, D=disjointness, S=subset.} \label{tab:questions-sets} \begin{tabular}{@{ }c@{\hspace{.95ex}}c@{\hspace{.95ex}}c@{\hspace{.95ex}}c@{\hspace{.95ex}}c@{\hspace{.95ex}}c@{\hspace{.95ex}}c@{\hspace{.95ex}}c@{\hspace{.95ex}}c@{\hspace{.95ex}}c@{\hspace{.95ex}}c@{\hspace{.95ex}}c@{}} \hline \\[-2ex] \multicolumn{6}{l}{\small Task set 1:}\\ \cline{1-3}\\[-2ex] 1 & 6 & 2 & 4 & 3 & 5 & 7 & 12 & 8 & 10 & 9 & 11 \\ LEI & MED & LES & MEI & LED & MES & LHI & MHD & LHS & MHI & LHD & MHS\\ \hline\\[-2ex] \multicolumn{6}{l}{\small Task set 2:}\\ \cline{1-3}\\[-2ex] 1 & 6 & 2 & 4 & 3 & 5 & 7 & 12 & 8 & 10 & 9 & 11 \\ MEI & LED & MES & LEI & MED & LES & MHI & LHD & MHS & LHI & MHD & LHS\\ \hline \end{tabular} \end{table} The participants were instructed to answer the questions as accurately and as quickly as possible. We measured time (T) and accuracy (A) as the main dependent variables. Once the participants answered each question, they were presented with a task difficulty rating for that question, which they were asked to complete. Ratings were entered on a Likert scale, ranging from 1 (very easy) to 7 (very difficult). Participants were informed that the time taken to enter the ratings was not recorded (i.e. it was not added to their answer times). A short text containing an explanation of how to interpret both mosaic and linear diagrams, including visual examples, was presented to each participant at the start of the study sessions. This was followed by the participants completing a 6-question tutorial in which task questions similar to those asked during the actual study were presented in the same manner as in the actual study. This tutorial set of questions was, of course, based on a different collection of sets than the one used in the study. After answering each of the tutorial task questions, participants were given the correct answer, along with a brief explanation of the answer. Figure~\ref{fig:tutorial-screenshot} shows a screen-shot of one of the tutorial task questions, after it has been completed, along with the difficulty rating, yet to be submitted. \begin{figure}[ht] \centering \includegraphics[width=.6\linewidth]{tutorial-screenshot.pdf} \caption{A screen-shot of one of the tutorial questions, with the completed answer and difficulty rating.} \label{fig:tutorial-screenshot} \end{figure} After finishing the tutorial, the participants were directed to the actual study. The study component functioned slightly differently from the tutorial session, in that the correct answers were not presented to the participants after they completed the test questions. \begin{table}[h!] \centering \caption{Task questions used in the study, with all the given choices shown in brackets, and answers in \emph{italics}.} \label{tab:task-questions}\vspace{1ex} \begin{tabular}{llp{.8\linewidth}} \hline No. & Type & Question\\ \hline 1 & EI & Tick the check boxes where some of the people are also interested in Books.\\ & & (\emph{Android}, \emph{Cars}, \emph{Media}, \emph{News}, \emph{Stars}, None of the above) \\ 2 & ES & Tick the check boxes where all of the people are also interested in Hifi.\\ & & (\emph{Android}, \emph{Books}, Cars, \emph{Design}, \emph{Media}, None of the above)\\ 3 & ED & Tick the check boxes where none of the people are also interested in Economics.\\ & & (\emph{Cars}, Food, \emph{Music}, \emph{Stars}, \emph{Travel}, None of the above)\\ 4 & EI & Tick the check boxes where some of the people are also interested in Games.\\ & & (Computers, \emph{Design}, \emph{Food}, \emph{Programming}, Travel, None of the above)\\ 5 & ES & Tick the check boxes where all of the people are also interested in Web.\\ & & (\emph{Hifi}, iPhone, News, Relaxation, Travel, None of the above)\\ 6 & ED & Tick the check boxes where none of the people are also interested in Programming.\\ & & (Camping, \emph{Food}, \emph{Journalism}, Stars, Web, None of the above)\\ 7 & HI & Tick the check boxes where some of the people are also interested in either Computers \\ & & or Economics. (Games, \emph{Journalism}, \emph{News}, \emph{Relaxation}, None of the above)\\ 8 & HS & Tick the check boxes where all of the people are also interested in either Economics or Web.\\ & & (Books, Computers, Internet, Media, \emph{None of the above})\\ 9 & HD & Tick the check boxes where none of the people are also interested in both Cars and Travel.\\ & & (Design, \emph{Health}, Media, Relaxation, None of the above)\\ 10 & HI & Tick the check boxes where some of the people are also interested in either College or Relaxation.\\ & & (\emph{Android}, Design, \emph{Internet}, \emph{Stars}, None of the above)\\ 11 & HS & Tick the check boxes where all of the people are also interested in either Design or Economics.\\ & & (Food, \emph{Internet}, Relaxation, \emph{Technology}, None of the above)\\ 12 & HD & Tick the check boxes where none of the people are also interested in both Books and Food.\\ & & (Camping, \emph{Economics}, Hifi, News, None of the above)\\ \hline \end{tabular} \end{table} \begin{table}[h!t] \centering \caption{Alternative linear and mosaic visualization images used for each task question. The numbers of non-empty, pairwise intersection (I), disjointness (D) and subset relations (S) are shown on the right.} \label{tab:task-questions-images}\vspace{1ex} \includegraphics[width=.8\linewidth]{table3} \end{table} \subsection{Task Questions} Table~\ref{tab:task-questions} presents the task questions used in this study, along with the choices given for each question (please note that the sets belonging to the correct answers are shown in italics). The selected questions covered all types and difficulty levels enumerated previously. Words representing quantifiers and logical relations (some, all, none, both, either/or) were highlighted in the questions, so as to draw attention to the set relations being assessed. We realise that the wording of the questions is complicated, and somewhat unnatural. However, given the difficulty in devising natural-sounding questions about abstract relations, and in order to facilitate comparison between our results and those of \cite{bib:RodgersEtal2015}, we chose to replicate the wording used in their experiment. Table~\ref{tab:task-questions-images} provides a small version of the Linear and Mosaic visualization images which were used alternatively for each question, and of course were counter-balanced. Figure~\ref{fig:session-screenshot-mei}.a shows a screen-shot of the Easy Intersection question presented using the Mosaic visualization (i.e. MEI) during the actual study session using Task set 2 (see Table~\ref{tab:questions-sets}). Figure~\ref{fig:session-screenshot-mei}.b, on the other hand, shows a screen-shot of the Easy Disjointness question presented using the Linear visualization (i.e. LED), also using Task set 2. \begin{figure}[htbp] \centering a)\includegraphics[width=.45\linewidth]{session-screenshot-mei.pdf}~b)\includegraphics[width=.45\linewidth]{session-screenshot-led.pdf} \caption{Sample screen shots of questionnaire system, showing completed answer and difficulty rating: a) question 1 (MEI) and b) question 6 (LED), both presented during the experiment using Task set 2. } \label{fig:session-screenshot-mei} \end{figure} \subsection{Participants} We initially conducted a power analysis to determine the number of participants needed in order to detect differences in user performance at the significance level $p < 0.05$. Assuming that interesting performance differences induced by the use of mosaic or linear diagrams would have relatively large effect sizes, say, $\eta^2$ slightly above 0.138 \citep{bib:Cohen88spanb} and aiming for 70\% power ($1-\beta$), we estimated that around 18 participants would be sufficient for this study. However, we recruited 26 participants in order ensure the availability of sufficient data. Two of these participants experienced technical difficulties during the experiment, and their answers were excluded from the analysis. This left us with a total of 24 participants who completed all the task questions. Of these, 18 were male and 6 female, and their age groups were distributed as follows: 20-29 (8), 30-39 (6), 40-49 (5), 50-59 (5). As regards their occupations, 10 were academics, 9 students, and 5 had other occupations. Ten participants (41.6\%) wore glasses, and none of the participants were colour blind. Once again, due to within-subject design of our study, these variations in participants attributes are likely to have little impact on the results of our study. \subsection{Results} The answers to the task questions were collated into a single data file containing all the 288 (24x12) answers, and analysed using the R language. We started by comparing the accuracy scores of mosaic and linear diagrams overall, and followed this up by comparing them according to task type (i.e. tasks involving visual detection of intersections, disjointness, and subsets, respectively). Analysis of accuracy figures are of special interest here, since accuracy analysis formed the basis for performance comparison in similar experiments \citep{bib:RodgersEtal2015}. Pearson's $\chi^2$ test revealed no differences in either overall or task specific comparisons. The results are summarized in Table~\ref{tab:accuracy}. Remarkably, the overall accuracy for mosaic diagrams was almost exactly the same as the accuracy for linear diagrams. When broken down by task types, we see a trend (but no statistical significance at $p<0.05$) for better performance of mosaic on tasks based on the detection of intersections (questions labelled {\em EI} and {\em HI} in tables~\ref{tab:task-questions} and \ref{tab:task-questions-images}), no difference on disjointness tasks (questions {\em ED} and {\em HD} in tables~\ref{tab:task-questions} and \ref{tab:task-questions-images}), and an advantage for linear diagrams in detection of subsets (questions {\em ES} and {\em HS} in tables~\ref{tab:task-questions} and \ref{tab:task-questions-images}). \begin{table}[htb] \centering \caption{Comparison of accuracy scores in task questions based on linear and mosaic diagrams. The figures represent the percentage of correct answers out of the total number of answers given.}\vspace{1ex} \label{tab:accuracy} \begin{tabular}{lccccc} \hline Task & Linear & Mosaic & $\chi^2$ & $p<$ & df \\ \hline Intersection & 70.8\% & 73.0\% & 0.00 & 1.00 & 1\\ Disjointness & 77.0\% & 77.0\% & 0.00 & 1.00 & 1\\ Subset & 79.1\% & 75.0\% & 0.05 & 0.80 & 1\\ All & 75.6\% & 75.0\% & 0.00 & 1.00 & 1\\ \hline \end{tabular} \end{table} Given these results, we further investigated accuracy by comparing the different types of tasks grouped according to their difficulty levels, that is, \emph{easy} ({\em EI}, {\em ED} and {\em ES}) versus \emph{hard} ({\em HI}, {\em HD} and {\em HS}). In these comparisons, we employed McNemar's test, as each group consisted of paired data. Once again the accuracy scores were rather similar, with no statistically significant differences shown (see Table~\ref{tab:accuracybytype}). However, there appears to be a tendency for greater accuracy on the easier tasks for linear diagrams (84.7\% versus 77.8\%, $p<0.40$), and conversely greater accuracy for mosaic on harder tasks (72.3\% versus 66.7\%, $p<0.47$). \begin{table}[ht] \centering \caption{Comparison of accuracy scores in task questions (Intersection, Disjointness, Subset) grouped according to difficulty level (Easy, Hard) for mosaic and linear diagrams. The $\chi^2$ values are computed according to McNemar's method.} \label{tab:accuracybytype}\vspace{1ex} \begin{tabular}{lccccc} \hline Question & Linear & Mosaic & $\chi^2$ & $p<$ & df \\ \hline EI & 79.2\% & 79.2\% & 0.00 & 1.00 & 1\\ ED & 87.5\% & 79.2\% & 0.12 & 0.72 & 1\\ ES & 87.5\% & 75.0\% & 0.57 & 0.45 & 1\\ HI & 62.5\% & 66.7\% & 0.00 & 1.00 & 1\\ HD & 66.7\% & 75.0\% & 0.17 & 0.68 & 1\\ HS & 70.8\% & 75.0\% & 0.00 & 1.00 & 1\\ \hline \end{tabular} \end{table} We then measured the participants' performance in terms of the time taken to answer each task question (excluding the time taken to rate task difficulty). The distributions of answer times are summarized on the box plots of figures~\ref{fig:times-easy} and \ref{fig:times-hard}, for \emph{easy} and \emph{hard} questions respectively. Overall mosaic users took on average 54s ($SD=27.2$) to answer a question, while linear diagram users took 49s ($SD=27.2$). \begin{figure} \centering \begin{minipage}[b]{.472\linewidth} \centering \includegraphics[width=.8\linewidth]{times-easy} \caption{Time to answer {\em easy} questions using linear and mosaic diagrams.} \label{fig:times-easy} \end{minipage}\hspace{2.2em}\begin{minipage}[b]{.472\linewidth} \centering \includegraphics[width=.8\linewidth]{times-hard} \caption{Time to answer {\em hard} questions using linear and mosaic diagrams.} \label{fig:times-hard} \end{minipage} \end{figure} Repeated measures analysis of variance (ANOVA) showed no significant effects for the two visualization types ($F(1,276)=3.3$, $p=0.07$) or task question types ($F(2,276)=0.33$, $p=0.72$). No significant interactions between these variables were found either. The only significant difference found was between \emph{easy} and \emph{hard} tasks ($F(1,276)=31.9$, $p<0.05$, adjusted), which simply validated our experiment design choices for task question difficulties. Nevertheless, Figure~\ref{fig:times-easy} shows a trend for users of linear diagrams to take slightly less time on the \emph{easy} tasks. This difference does not persist however in the \emph{hard} tasks (Figure~\ref{fig:times-hard}), reversing, in fact, for the subset type tasks (last tasks). While further investigation is necessary to clarify this reversal in performance, we hypothesize that it is due to the fact that at the beginning of the experiment linear diagrams are likely to be more familiar to users (perhaps as a consequence of previous exposure to similar diagrams, such as Gantt charts) than mosaic diagrams. As users gain familiarity with the mosaic representation, their performance improves. Finally, we compared the participants' subjective ratings for task difficulty. Figures~\ref{fig:likert-easy} and \ref{fig:likert-hard} show summaries of responses for the two difficulty levels (\emph{easy} and \emph{hard} respectively), grouped by the three task types and two diagram types. The ratings are again similar, but less consistent. The median rating is 3 for both mosaic and linear diagrams. The Kruskal-Wallis test showed no statistically significant difference ($\chi^2 = 2.68$, df = 1, $p = 0.10 $). Despite their subjectivity, the ratings generally correlate to time on task (Pearson's $\rho(2.91,142)=0.24$, $p < 0.01$, for mosaic diagrams, and $\rho(3.3,142)=0.30$, $p < 0.01$, for linear diagrams) lending additional support to the hypothesis that performance on mosaic diagrams tended to improve more than performance on linear diagrams over time. \begin{figure}[ht] \centering \includegraphics[width=.55\linewidth]{likert-easy} \caption{Ratings for task difficult, with respect to {\em easy} tasks ({\em EI, ED, ES}).} \label{fig:likert-easy} \end{figure} \begin{figure}[ht] \centering \includegraphics[width=.55\linewidth]{likert-hard} \caption{Ratings for task difficult, with respect to {\em hard} tasks ({\em HI, HD, HS}).} \label{fig:likert-hard} \end{figure} \subsection{Discussion} The study presented here has shown that ordinary mosaic diagrams are comparable in their effectiveness to the most effective linear diagrams that follow previously proposed visual design principles, as discussed earlier \citep{bib:RodgersEtal2015}. However, the superiority of temporal mosaics over temporal linear diagrams (Gantt charts) in the context of task schedulling \citep{bib:LuzMasoodian2011}, which we hypothesized would translate to the set comparison tasks, was not observed in the current study. While it is not entirely clear why accuracy and answer times were so similar for both diagrams, one could speculate about contributing factors. One such factor may be the kinds of tasks the user is asked to perform in each case. Even though the basic visual tasks are roughly similar (detection of gaps and overlaps), in schedule visualization the user is also asked to assess interval length and position on the timeline (start and end times), which therefore provides a structuring element which facilitates interpretation and might benefit mosaic, where these characteristics are represented more prominently. The complexity and level of abstraction of the questions asked in the present set relations task is likely to be another contributing factor. The questions in this task are rather more abstract, and as we have pointed out, their textual formulation has to balance naturalness with the need to avoid ambiguity, resulting in wordings that are sometimes rather difficult to interpret. This is likely to have played a role in levelling down user performance across the two conditions. There are, however, certain advantages to mosaic diagrams, which although not tested in the current study, are likely to positively influence their effectiveness. For instance, the space-filling property of mosaic diagrams preserves the overview of overlaps and exclusions even if the diagram is dramatically reduced in size. Linear diagrams, on the other hand, rely on the position of labels to identify relations (colour being, as we noted before, a redundant attribute). As these diagrams are scaled down, the user's ability to align vertically is greatly diminished, since the horizontally aligned labels would be impossible to preserve in miniatures, leaving the otherwise redundant colour attribute as the only means of identifying individual sets. In miniature linear diagrams, as in normal-sized ones, empty spaces will dominate the image, hindering the perception of vertical alignment of horizontal lines. Compare, for example, \includegraphics[height=1.7ex]{minimosaic} to \includegraphics[height=1.7ex]{minilinear-wide}. Such miniatures could be useful, for instance, in small-multiples diagrams \citep{bib:Tufte90}, or in ``mini-charts'' like sparklines \citep{bib:Tuft01} when presented along with tabular data. In addition, mosaics highlight overlaps by facilitating visual alignment tasks, because the edges of adjacent areas to be aligned stand out clearly. In linear diagrams, on the other hand, comparisons of set relationships can become increasingly more challenging as more sets are included, thus leading to increasing vertical distances between sets and including more distracting line segments between sets that are placed vertically far apart. As mentioned previously, this kind of line ``crowding'' is known to impair user performance in alignment tasks \citep{bib:SayimWestheimerHerzog08c,bib:LeviKleinAitsebaomo85v}. Furthermore, linear diagrams do not generally represent other set properties such as their cardinality, and while it has been suggested \citep{bib:RodgersEtal2015} that visual properties including line size (e.g. length or width), colour, and texture could be used to show set cardinality, it is acknowledged that their effectiveness has not been demonstrated. It could be argued that the use of line length for representing cardinality is potentially feasible, while changing line width may be less effective, given that it has been shown that thin lines are more effective than thick lines. Similarly, although colour and texture visual properties can be used for representing categorical variables, they are not very useful for representing ordinal variables (e.g. relative cardinality of different sets) \citep{bib:Mackinlay86Automating}. Mosaic diagrams, on the other hand, have been designed, and shown \citep{bib:LuzMasoodian2011} to facilitate comparisons of relative sizes (e.g. duration of task schedules). Figure~\ref{fig:3sets-mosaic-cardinal} provides a simple example of how comparison of set cardinalities could be supported by proportionally varying the length of mosaic segments representing set relationships in proportion to the cardinalities of the sets being compared. Figure~\ref{fig:3sets-mosaic-cardinal}a shows only the relationships between the three sets (Books, Technology, Cars) without conveying any information about their cardinalities. Figure~\ref{fig:3sets-mosaic-cardinal}b, however, makes comparisons of the proportional cardinalities of the three sets relatively easy. For instance, it is clear that half of the people interested in cars are also interested in both books and technology, while the other half are not. Similarly, half the people interested in books are also interested in technology, while the other half are not. Also, it can be seen that Books is the largest set, followed by Technology, and Cars. \begin{figure}[htbp] \centering a)\includegraphics[width=0.45\linewidth]{3sets-mosaic.pdf}~~ b)\includegraphics[width=0.45\linewidth]{3sets-mosaic-cardinal.pdf} \caption{Relationships between three example sets, shown using mosaic diagrams a) without, and b) with cardinality comparisons.} \label{fig:3sets-mosaic-cardinal} \end{figure} Although interactive visualization techniques are not discussed here, mosaic diagrams have been shown to lend themselves well to the incorporation of interactive elements (e.g. selection, brushing, zoom, etc.) in comparison to linear-style visualizations such as Gantt charts \citep{bib:LuzMasoodian2010}. It should be noted, however, that as with any visualization, the use of the colour hue attribute to encode data values places some restrictions on the visualization for viewers who suffer from colour-blindness. This is also true for mosaic diagrams. One possible solution in such cases is to use another colour attribute, such as tonal variations (i.e. value), or perhaps texture instead of hue variations. \section{Conclusions} In this paper, we have proposed the use of mosaic diagrams as an aggregation-based technique for visualization of set relationships. This is a novel use of mosaic diagrams, which have previously been shown to be very effective for visualization of temporal data such as multimedia streams, and task schedules. Although mosaics failed to yield performance improvements in comparison to linear diagrams for set visualization tasks, as we had expected based on reported results from a different task (schedule visualization) which compared similar diagrams, the potential value of mosaic diagrams for representing set relationships is supported by the fact that mosaic produced similar results as the most effective visual form of linear diagrams, as previously studied by \cite{bib:RodgersEtal2015}. Finally we have discussed a number of cases where mosaic diagrams are likely to be particularly suitable for visualization of set relationships. These include cases where visual space is limited and/or needs to be used more efficiently, cases where a larger number of sets need to be represented, or cases where other set properties such as their cardinalities also need to be presented. These, and other interactive properties of mosaic diagrams, still need to be further investigated within this particular task domain. We aim to carry out this work in the near future. \bibliographystyle{plainnat}

https://arxiv.org/abs/1410.5112

The Best Mixing Time for Random Walks on Trees

We characterize the extremal structures for mixing walks on trees that start from the most advantageous vertex. Let $G=(V,E)$ be a tree with stationary distribution $\pi$. For a vertex $v \in V$, let $H(v,\pi)$ denote the expected length of an optimal stopping rule from $v$ to $\pi$. The \emph{best mixing time} for $G$ is $\min_{v \in V} H(v,\pi)$. We show that among all trees with $|V|=n$, the best mixing time is minimized uniquely by the star. For even $n$, the best mixing time is maximized by the uniquely path. Surprising, for odd $n$, the best mixing time is maximized uniquely by a path of length $n-1$ with a single leaf adjacent to one central vertex.

\section{Introduction} We resolve the following extremal question for random walks on trees: what tree structure minimizes/maximizes the expected length of an optimal stopping rule to the stationary distribution $\pi$, given that we start at the most advantageous vertex? Naturally, the star $S_n$ is the minimizing tree structure, but the maximization problem has an unexpected twist. The path $P_{n}$ is the maximizing structure when $n$ is even, but for odd $n$ the maximizing structure is the near-path $Y_n$ consisting of a path on $n-1$ vertices with a single leaf adjacent to one of the two central vertices. We refer to this graph $Y_n$ as the \emph{wishbone}. This choice of name is suggested by the layout of $Y_n$ in Figure \ref{fig:menagerie}(c) below. A \emph{random walk} on an undirected graph $G=(V,E)$ consists of a sequence of vertices $(w_0, w_1, \ldots , w_n, \ldots)$ such that $\mbox{Pr}[w_{t+1} = w \mid w_t = v]$ is $1/ \deg(v)$ if $(v,w) \in E$ and $0$ otherwise. The \emph{hitting time} $H(v,w)$ from vertex $v$ to vertex $w$ is the expected number of steps before a random walk started at $v$ visits $w$ for the first time. We also define $H(v,v)=0$, while $\mbox{Ret}(v)$ denotes the expected number of steps before a random walk started at $v$ first returns to $v$. When $G$ is not bipartite, the distribution of $w_t$ converges to the \emph{stationary distribution} $\pi$, where $\pi_v = \deg(v) /2|E|.$ Inconveniently, we do not have convergence for bipartite graphs (including trees), but we can rectify this at the cost of doubling the expected length of any walk. Indeed the distribution $w_t$ converges to $\pi$ when we follow a \emph{lazy random walk} in which we remain at the current state with probability $1/2$ at each time step. A \emph{mixing measure} of a graph $G$ captures the rate of convergence to the stationary distribution $\pi$. Let $\sigma=\sigma_0$ denote our initial distribution and $\sigma_t$ denote the distribution for the $t$-th step of a random walk. For a fixed constant $0 < \epsilon < 1$, the (approximate) mixing time of $G$ is given by $T_{\rm mix}(\epsilon) = \max_{\sigma} \min \{T \geq 0 : \, \| \sigma_t - \pi \| < \epsilon \mbox{ for all } t \geq T \},$ where the maximum is taken over all possible starting distributions and $\| \cdot \|$ is the given metric. This mixing time depends upon the choice of the parameter $\epsilon$. Lov\'asz and Winkler \cite{lovasz+winkler} studied a class of parameterless mixing measures by using more sophisticated \emph{stopping rules} to drive the random walk to a desired distribution. Suppose that we are given a \emph{starting distribution} $\sigma$ and a \emph{target distribution} $\tau$. A \emph{$(\sigma,\tau)$ stopping rule} halts a random walk whose initial state is drawn from $\sigma$ so that the final state is governed by the distribution $\tau$. The \emph{access time} $H(\sigma, \tau)$ denotes the minimum expected length for a such a $(\sigma,\tau)$ stopping rule to halt. We say that a stopping rule is \emph{optimal} if it achieves this minimum expected length. Using access times, we have three natural mixing measures, the \emph{mixing time}, the \emph{reset time}, and the \emph{best mixing time} given respectively by \[ T_{\rm mix} = \max_{v \in V} H(v, \pi), \quad \mbox{and} \quad T_{\rm reset} = \sum_{v \in V} \pi_v H(v, \pi), \quad \mbox{and} \quad T_{\rm bestmix} = \min_{v \in V} H(v, \pi). \] These are the worst-case, average-case and best-case mixing measures. These quantities are called \emph{exact mixing measures}, as opposed to approximate mixing measures like $T_{\rm mix}(\epsilon)$ above. See \cite{ALW} for a taxonomy that compares exact and approximate mixing measures for Markov chains. We also note that exact stopping rules converge to $\pi$ on a bipartite graph, even for non-lazy walks. Indeed, stopping rules can employ randomness in deciding when to stop, which has the same periodicity-breaking effect as using a lazy walk. It is natural to wonder which graph structures lead to slow or rapid mixing, and the study of extremal graphs for random walks is well-established, cf. \cite{aldous+fill,bright, feige}. A natural line of inquiry is to restrict our attention to trees. The extremal tree structures on $n$ vertices for $T_{\rm mix}$ and $T_{\rm forget}$ were characterized in \cite{beveridge+wang}. Not surprisingly, the unique minimal structure is the star $S_n=K_{1,n-1}$ and the unique maximal structure is the path. \begin{theorem}[\cite{beveridge+wang}] \label{thm:mix-reset} If $G$ is a tree on $n \geq 3$ vertices then \[ \frac{3}{2} \leq T_{\rm mix} \leq \frac{2n^2 - 4n + 3}{6} \] and \[ 1 \leq T_{\rm reset} \leq \left\{ \begin{array}{cc} \frac{1}{4}(n^2 - 2n +2) & \mbox{if $n$ is even}, \\ \frac{1}{4}(n-1)^2 & \mbox{if $n$ is odd}. \\ \end{array} \right. \] In each instance, the lower bound is achieved uniquely by the star $S_n$ and the upper bound is achieved uniquely by the path $P_n$. \end{theorem} We add to these extremal tree characterizations by studying the best mixing time $T_{\rm bestmix}$. \begin{theorem} \label{thm:bestmix} Let $G$ be a tree on $n \geq 3$ vertices. Then \[ 1/2 \leq T_{\rm bestmix}(G) \leq \left\{ \begin{array}{rl} \frac{1}{12}(n^2 +4n -6) & \mbox{if $n$ is even}, \\ \frac{1}{12}(n^2+4n-15) & \mbox{if $n$ is odd}. \\ \end{array} \right. \] The star $S_n$ is the unique minimizing tree. For $n$ even, the path $P_n$ is the unique maximizing tree. For $n$ odd, the wishbone $Y_n$ is the unique maximizing tree. \end{theorem} Below, we assume that $n \geq 4$ since there is a unique tree for $n=3$ (and the formula is correct in that case). The appearance of the wishbone is quite unexpected. At the end of the next section, we discuss the characteristics of $Y_n$ that lead to its maximization of the best mixing time for odd $n$. The proof of Theorem \ref{thm:bestmix} is more involved than the proof of Theorem \ref{thm:mix-reset} in \cite{beveridge+wang}. The reason is that a minor change to the tree can abruptly shift the location of the initial vertex achieving $T_{\rm{bestmix}}$, whereas the behavior of $T_{\rm mix}$ and $T_{\rm reset}$ are more stable under minor structural alterations. Here is an overview of our proof. First, we replace the given tree with a caterpillar having a larger best mixing time. We then follow a prescribed algorithm, carefully moving one or two leaves at a time via a \emph{tree surgery} $S_i$. These surgeries monotonically increase $T_{\rm bestmix}$, until we arrive at an even path or an odd wishbone. Figure \ref{fig:examples} shows two sequences of surgeries, the first resulting in $P_{10}$ and the second resulting in $Y_{11}$. The surgeries and the guiding algorithm are described in Section \ref{sec:proof}. \begin{figure}[t] \begin{center} \begin{tikzpicture}[scale=.5, every node/.style={font=\scriptsize}, decoration={ markings, mark=at position 1 with {\arrow[scale=1.25,black]{latex}}; } ] \begin{scope} \draw (1,0) -- (7,0); \draw (2,0) -- (2,-1); \draw (5,0) -- (4.67,-1); \draw (5,0) -- (5.33,-1); \foreach \i in {1,2,3,5,6,7} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=black!20] (2,-1) circle (4pt); \draw[fill=black!40] (4.67,-1) circle (4pt); \draw[fill=black!60] (5.33,-1) circle (4pt); \draw[fill=white] (4,0) circle (6pt); \draw[fill=white] (4,0) circle (4pt); \draw[bend left=20, postaction=decorate] (8, -.5) to (8,-2.5); \node at (8.75,-1.5) {$S_1$}; \node at (2,-1.6) {$a$}; \node at (4.67,-1.5) {$b$}; \node at (5.33,-1.6) {$c$}; \node at (7,-.5) {$d$}; \end{scope} \begin{scope}[shift={(0,-3)}] \draw (1,0) -- (6,0); \draw (2,0) -- (2,-1); \draw (4,0) -- (4,-1); \draw (5,0) -- (4.67,-1); \draw (5,0) -- (5.33,-1); \foreach \i in {1,2,5,6} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=black!20] (2,-1) circle (4pt); \draw[fill=black] (4,-1) circle (4pt); \draw[fill=black!40] (4.67,-1) circle (4pt); \draw[fill=black!60] (5.33,-1) circle (4pt); \draw[fill=white] (3,0) circle (4pt); \draw[fill=white] (4,0) circle (6pt); \draw[fill=white] (4,0) circle (4pt); \draw[bend left=20, postaction=decorate] (8, -.5) to (8,-2.5); \node at (8.75,-1.5) {$S_4$}; \node at (2,-1.6) {$a$}; \node at (4.67,-1.5) {$b$}; \node at (5.33,-1.6) {$c$}; \node at (4,-1.5) {$d$}; \end{scope} \begin{scope}[shift={(0,-6)}] \draw (1,0) -- (7,0); \draw (2,0) -- (2,-1); \draw (4,0) -- (3.67,-1); \draw (4,0) -- (4.33,-1); \foreach \i in {1,2,5,6} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=black!20] (2,-1) circle (4pt); \draw[fill=black] (3.67,-1) circle (4pt); \draw[fill=black!40] (4.33,-1) circle (4pt); \draw[fill=black!60] (7,0) circle (4pt); \draw[fill=white] (3,0) circle (4pt); \draw[fill=white] (4,0) circle (6pt); \draw[fill=white] (4,0) circle (4pt); \draw[bend left=20, postaction=decorate] (8, -.5) to (8,-2.5); \node at (8.75,-1.5) {$S_6$}; \node at (2,-1.6) {$a$}; \node at (4.33,-1.5) {$b$}; \node at (7,-.6) {$c$}; \node at (3.67,-1.5) {$d$}; \end{scope} \begin{scope}[shift={(0,-9)}] \draw (0,0) -- (7,0); \draw (4,0) -- (3.67,-1); \draw (4,0) -- (4.33,-1); \foreach \i in {1,2,5,6} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=black!20] (0,0) circle (4pt); \draw[fill=black] (3.67,-1) circle (4pt); \draw[fill=black!40] (4.33,-1) circle (4pt); \draw[fill=black!60] (7,0) circle (4pt); \draw[fill=white] (3,0) circle (6pt); \draw[fill=white] (3,0) circle (4pt); \draw[fill=white] (4,0) circle (6pt); \draw[fill=white] (4,0) circle (4pt); \draw[bend left=20, postaction=decorate] (8, -.5) to (8,-2.5); \node at (8.85,-1.5) {$S_{11}$}; \node at (0,-.6) {$a$}; \node at (4.33,-1.5) {$b$}; \node at (7,-.6) {$c$}; \node at (3.67,-1.5) {$d$}; \end{scope} \begin{scope}[shift={(0,-12)}] \draw (-1,0) -- (8,0); \foreach \i in {1,2,5,6} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=black!20] (0,0) circle (4pt); \draw[fill=black] (-1,0) circle (4pt); \draw[fill=black!40] (8,0) circle (4pt); \draw[fill=black!60] (7,0) circle (4pt); \draw[fill=white] (4,0) circle (6pt); \draw[fill=white] (3,0) circle (4pt); \draw[fill=white] (4,0) circle (6pt); \draw[fill=white] (4,0) circle (4pt); \node at (0,-.6) {$a$};background \node at (8,-.5) {$b$}; \node at (7,-.6) {$c$}; \node at (-1,-.5) {$d$}; \end{scope} \begin{scope}[shift={(14,0)}] \draw (1,0) -- (7,0); \draw (3,0) -- (2.75,-1); \draw (3,0) -- (3.25,-1); \draw (4,0) -- (4,-1); \draw (5,0) -- (5,-1); \foreach \i in {1,2,5,6,7} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=black!20] (2.75,-1) circle (4pt); \draw[fill=black!40] (3.25,-1) circle (4pt); \draw[fill=black!60] (4,-1) circle (4pt); \draw[fill=black!80] (5,-1) circle (4pt); \draw[fill=white] (3,0) circle (4pt); \draw[fill=white] (4,0) circle (6pt); \draw[fill=white] (4,0) circle (4pt); \draw[bend left=20, postaction=decorate] (8, -.5) to (8,-2.5); \node at (8.75,-1.5) {$S_8$}; \node at (2.75,-1.6) {$a$}; \node at (3.25,-1.5) {$b$}; \node at (4,-1.6) {$c$}; \node at (5,-1.5) {$d$}; \end{scope} \begin{scope}[shift={(14,-3)}] \draw (1,0) -- (7,0); \draw (3,0) -- (2.75,-1); \draw (3,0) -- (3.25,-1); \draw (4,0) -- (3.75,-1); \draw (4,0) -- (4.25,-1); \foreach \i in {1,2,3,5,6,7} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=black!20] (2.75,-1) circle (4pt); \draw[fill=black!40] (3.25,-1) circle (4pt); \draw[fill=black!60] (3.75,-1) circle (4pt); \draw[fill=black!80] (4.25,-1) circle (4pt); \draw[fill=white] (3,0) circle (4pt); \draw[fill=white] (4,0) circle (6pt); \draw[fill=white] (4,0) circle (4pt); \draw[bend left=20, postaction=decorate] (8, -.5) to (8,-2.5); \node at (8.75,-1.5) {$S_9$}; \node at (2.75,-1.6) {$a$}; \node at (3.25,-1.5) {$b$}; \node at (3.75,-1.6) {$c$}; \node at (4.25,-1.5) {$d$}; \end{scope} \begin{scope}[shift={(14,-6)}] \draw (0,0) -- (7,0); \draw (3,0) -- (3,-1); \draw (4,0) -- (3.75,-1); \draw (4,0) -- (4.25,-1); \foreach \i in {1,2,3,5,6,7} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=black!20] (0,0) circle (4pt); \draw[fill=black!40] (3,-1) circle (4pt); \draw[fill=black!60] (3.75,-1) circle (4pt); \draw[fill=black!80] (4.25,-1) circle (4pt); \draw[fill=white] (3,0) circle (6pt); \draw[fill=white] (3,0) circle (4pt); \draw[fill=white] (4,0) circle (6pt); \draw[fill=white] (4,0) circle (4pt); \draw[bend left=20, postaction=decorate] (8, -.5) to (8,-2.5); \node at (8.9,-1.5) {$S_{10}$}; \node at (0,-.6) {$a$}; \node at (3,-1.5) {$b$}; \node at (3.75,-1.6) {$c$}; \node at (4.25,-1.5) {$d$}; \end{scope} \begin{scope}[shift={(14,-9)}] \draw (-1,0) -- (8,0); \draw (4,0) -- (4,-1); \foreach \i in {1,2,3,5,6,7} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=black!20] (0,0) circle (4pt); \draw[fill=black!40] (-1,0) circle (4pt); \draw[fill=black!60] (4,-1) circle (4pt); \draw[fill=black!80] (8,-0) circle (4pt); \draw[fill=white] (3,0) circle (6pt); \draw[fill=white] (3,0) circle (4pt); \draw[fill=white] (4,0) circle (6pt); \draw[fill=white] (4,0) circle (4pt); \node at (0,-.6) {$a$}; \node at (-1,-.5) {$b$}; \node at (4,-1.6) {$c$}; \node at (8,-.5) {$d$}; \end{scope}background \end{tikzpicture} \end{center} \caption{Two different trees that are converted to $P_{10}$ and $Y_{11}$, respectively, using the algorithm and the tree surgeries $S_i$ defined in Section \ref{sec:proof}. The value of $T_{\rm bestmix}$ monotonically increases throughout. The white vertices are the foci, with $T_{\rm bestmix}$ achieved by the circled vertex. Certain vertices are labeled by $a,b,c,d$ so they can be tracked across steps.} \label{fig:examples} \end{figure} \section{Preliminaries} For an introduction to the theory of exact stopping rules, see \cite{lovasz+winkler}. To prove Theorem \ref{thm:bestmix}, we need the formula for access times from singleton distributions to the stationary distribution. Therefore, we limit ourselves to describing results from \cite{lovasz+winkler} needed to calculate these values. \subsection{Pessimal vertices} \label{sec:pessimal} A \emph{$v$-pessimal vertex} $v'$ is a vertex that satisfies $H(v',v)=\max_{w\in V}{H(w,v)}.$ Note that pessimal vertices are not necessarily unique; for example, every leaf of the star $S_n$ is pessimal for the central vertex. We use $\psq{v}$ to denote a pessimal vertex for $\p{v}$, and we employ the notation $$\mathrm{Pess}(v) = H(\p{v},v) = \max_{w \in V} H(w,v),$$ so that we can refer to the pessimal hitting time to $v$ in a manner agnostic of the choice of pessimal vertex $\p{v}$. This allows us to define the lightweight notation \begin{equation} \label{eqn:pessdiff} \Delta\mathrm{Pess}(v)=\mathrm{Pess}_{\widetilde{G}}(v)-\mathrm{Pess}_G(v) \end{equation} which shields us from the fact that the $v$-pessimal vertices in $G$ and $\widetilde{G}$ may be distinct. \subsection{Calculating mixing times on trees} Next we address the calculation of $H(v, \pi)$, the expected length of an optimal stopping rule from the singleton distribution on $v$ to the stationary distribution $\pi$. We refer to $H(v,\pi)$ as the \emph{mixing time from $v$}. The following result holds for any graph $G$. \begin{theorem}[\cite{lovasz+winkler}] \label{thm:mix} The expected length of an optimal mixing rule starting from vertex $v$ is \begin{equation} \label{eqn:mix} H(v, \pi) = H(\p{v}, v) - \sum_{u \in V} \pi_u H(u,v) = \mathrm{Pess}(v) - \sum_{j \in V} \pi_u H(u,v). \end{equation} \end{theorem} The very useful formula \eqref{eqn:mix} allows us to calculate the access time $H(v, \pi)$ via a linear combination of vertex-to-vertex hitting times. In other words, if we know all the pairwise hitting times, it is easy to calculate the access time $H(v, \pi)$ for each starting vertex $v \in V$. For a given graph, we can determine every hitting time $H(u,v)$ by solving a system of linear equations. In the special case of trees, we have an explicit formula in terms of the distances and degrees of the graph. Variations on this formula appear in the literature (cf. \cite{beveridge, beveridge+wang, dtw, gw} and Chapter 5 of \cite{aldous+fill}). This paper builds on the results in \cite{beveridge,beveridge+wang}, so we opt for that formulation. From here forward, we assume that $G=(V,E)$ is a tree on $n \geq 4$ vertices. We start with a well-known result about the hitting time between adjacent vertices in trees. If $uv \in E$, then removing this edge breaks $G$ into two disjoint trees $G_u$ and $G_v$ where $u \in V(G_u)$ and $v \in V(G_v)$. We define $V_{u:v} = V(G_u)$ and $V_{v:u} = V(G_v)$. We think of $V_{u:v}$ as the set of vertices that are closer to $u$ than to $v$. For these adjacent vertices, let $F$ denote the induced tree on $V_{u:v} \cup \{ v \}$. We have \begin{equation} \label{eq:adjhtime} H(u,v)= \mbox{Ret}_{F} (v) - 1 = \sum_{w\in V_{u:v}}{\deg(w)} = 2|V_{u:v}|-1, \end{equation} where the first equality holds by the well-known equality $\mbox{Ret}(v) = 2|E|/\deg(v)$. Equation \eqref{eq:adjhtime} encodes a very useful property of trees: the hitting time from a vertex $u$ to an adjacent vertex $v$ only depends on $|V_{u:v}|$, independent of the particular structure of the tree. This equation can be used to reveal that when $G=P_n$ is the path ($v_1,v_2,\dots,v_n$), the hitting times are \begin{equation} \label{eq:htimepath2} H_{P_n}(v_i,v_j)= \left\{ \begin{array}{rl} (i-1)^2-(j-1)^2 & i\leq j,\\ (n-j)^2-(n-i)^2 & i>j. \end{array} \right. \end{equation} For example, the first of these formulas is calculated by repeatedly using equation \eqref{eq:adjhtime} to evaluate $H(v_i, v_j) = H(v_i, v_{i+1}) + H(v_{i+1}, v_{i+2}) + \cdots + H(v_{j-1}, v_{j})$. A similar argument produces a formula for the hitting times of an arbitrary tree, but first we need some additional notation. Let $u,v,w \in V$. Define $d(u,v)$ to be the distance between these two vertices and define \begin{displaymath} \ell(u,v;w)=\frac{1}{2}(d(u,w)+d(v,w)-d(u,v)) \end{displaymath} to be the length of the intersection of the $(u,w)$-path and the $(v,w)$-path. \begin{lemma}[\cite{beveridge}] \label{thm:htime} For any pair of vertices $v_i, v_j$, we have \begin{equation} \label{eq:htime} H(v_i,v_j)=\sum_{v\in V}{\ell(v_i,v;v_j)\deg(v)}. \end{equation} \end{lemma} Theorem \ref{thm:mix} and Lemma \ref{thm:htime} are the fundamental tools in our proof of Theorem \ref{thm:bestmix}. Together they provide a simple way to calculate state-to-state hitting times and mixing times. \subsection{The foci of a tree} The center and the barycenter are two well-established notions of centrality. The center of a tree $G$ is the vertex (or two adjacent vertices) that achieves $\min_{u \in V} \max_{v} d(v,u)$. The center does not appear to have any central properties with respect to random walks on trees. The barycenter is the vertex (or two adjacent vertices) that achieves $\min_{u \in V} \sum_{v \in V} d(v,u)$. Proposition 1 of \cite{beveridge} shows that that the barycenter is the ``average'' center for random walks on trees: this vertex also achieves $\min_{u} \sum_{v \in V} \pi_v H(v,u)$. A third type of centrality for trees was introduced in \cite{beveridge}: the \emph{focus} is the ``extremal'' center for random walks. \begin{definition} \label{def:focus} {\bf{(\cite{beveridge})}} A vertex $v$ achieving $\mathrm{Pess}(v) = \min_{w \in V} \mathrm{Pess}(w)$ is called a \emph{primary focus} of $G$. If all $v$-pessimal vertices are contained in a single subtree of $G-v$, then the unique $v$-neighbor $u$ in that subtree is also a focus of $G$. If $\mathrm{Pess}(u)=\mathrm{Pess}(v)$, then $u$ is also a primary focus. Otherwise $u$ is a \emph{secondary focus}. \end{definition} Every tree has either one focus, in which case it is \emph{focal}, or two adjacent foci, in which case it is \emph{bifocal}. We have the following two theorems, which relate the foci to mixing walks on $G$. \begin{theorem}[\cite{beveridge}] \label{thm:tifocus} If $G$ is focal with focus $u$, then for all $v$ $ H(v, \pi) = H(v,u) + H(u,\pi). $ If $G$ is bifocal with foci $u$ and $w$, then for $v \in V_{u:w}$, $ H(v, \pi) = H(v,u) + H(u,\pi) $ and for $v \in V_{w:u}$, $ H(v, \pi) = H(v,w) + H(w,\pi). $ \end{theorem} \begin{theorem}[\cite{beveridge}] \label{thm:tbmvertex} If $H(w,\pi) = T_{\rm{bestmix}}(G)$ then $w$ is a focus of $G$. For bifocal $G$ with foci $u$ and $v$, if $H(\p{v},u) < H(\p{u},v)$ then $v$ is the unique vertex achieving $T_{\rm{bestmix}}(G)$. If $H(\p{v},u) > H(\p{u},v)$ then $u$ is the unique vertex achieving $T_{\rm{bestmix}}(G)$. If $H(\p{v},u) = H(\p{u},v)$ then both vertices achieve $T_{\rm{bestmix}}$. \end{theorem} If $H(w,\pi) = T_{\rm{bestmix}}(G)$ then we say that $w$ is a \emph{best mix focus}. Note that a tree can have one or two best mix foci. In the former case, the best mix focus could be either the primary focus or the secondary focus: it depends on the relative sizes of $H(\p{v},u)$ and $H(\p{u},v)$. This unusual criterion is what makes handling the best mixing time so delicate. A small change to the tree structure can move the location of the best mix focus, even when the foci do not change. \subsection{The best mixing time for stars, paths and odd wishbones} \label{sec:bestmix-extreme} As our first application of these theorems and lemmas, let us calculate the best mixing time for the star, the path (both even and odd $n$) and the wishbone (for odd $n$). This will justify the expressions that appear in Theorem \ref{thm:mix}. Clearly, the central vertex $v$ of the star $S_n=K_{1,n-1}$ is its unique focus. Let $w \neq v$ be any other vertex. We have $\pi_v = 1/2$ and $\pi_w = 1/2(n-1)$. Furthermore, $H(w,v) = \mathrm{Pess}(v) = 1$ and obviously $H(v,v)=0$. Therefore \[ T_{\rm{bestmix}} (S_n) = \mathrm{Pess}(v) - \sum_{w \in V} \pi_w H(w,v) = 1 - \frac{1}{2(n-1)} (n-1) = \frac{1}{2}. \] In fact, it is easy to see that $S_n$ is the unique tree on $n$ vertices that achieves this value. Let $G \neq S_n$ and let $v$ be a best mix focus. Then $\pi_v = \deg(v)/2(n-1)< 1/2$ because the star is the unique tree with a vertex of degree $n-1$. An optimal mixing rule started at $v$ must exit with probability at least $1 - \pi_v > 1/2$, otherwise the ending distribution will weight vertex $v$ too heavily. This proves that the star is the unique minimizing structure for Theorem \ref{thm:bestmix}. Next, we calculate $T_{\rm{bestmix}}(P_n)$. Label the vertices as $(v_1, v_2, \ldots , v_n)$. We first consider odd $n=2r+1$. By symmetry, the unique focus is $v_{r+1}$. By Theorem \ref{thm:mix} and equation \eqref{eq:htimepath2}, \begin{eqnarray*} T_{\rm bestmix}(P_{2r+1}) &=& \mathrm{Pess}(v_{r+1}) - \sum_{i=1}^{2r+1} \pi_i H(v_i, v_{r+1}) \, = \, \mathrm{Pess}(v_{r+1}) - 2 \sum_{i=1}^{r} \pi_i H(v_i, v_{r+1}) \\ &=& r^2 - \frac{1}{2r} r^2 - \frac{1}{r} \sum_{i=2}^r \left( r^2 - (i-1)^2 \right) \,=\, \frac{2r^2+1}{6} \, = \, \frac{n^2-2n+3}{12}, \end{eqnarray*} where the second equality follows from the symmetry of the odd path. The calculation for the even path is a bit tougher. Setting $n=2r$, we have two best mix foci $v_r, v_{r+1}$ by symmetry. We will take $v_{r+1}$ as the best mix focus in our calculation. We have \begin{eqnarray*} T_{\rm bestmix}(P_{2r}) &=& \mathrm{Pess}(v_{r+1}) - \sum_{i=1}^{r} \pi_i H(v_i, v_{r+1}) - \sum_{i=r+2}^{2r} \pi_i H(v_i, v_{r+1}) \\ &=& \mathrm{Pess}(v_{r+1}) - 2\sum_{i=1}^{r-1} \pi_i H(v_i, v_{r}) - \sum_{i=1}^{r} \pi_i H(v_r, v_{r+1}) \\ &=& r^2 - \frac{2}{4r-2} (r-1)^2 - \frac{4}{4r-2} \sum_{i=2}^{r-1} \left( (r-1)^2 - (i-1)^2 \right) - \frac{1}{2} (2r-1) \\ &=& \frac{2r^2 + 4r-3}{6} \, = \, \frac{n^2 +4n - 6}{12} \end{eqnarray*} where we use the symmetry of the path and $H(v_i, v_{r+1}) = H(v_i, v_r) + H(v_r,v_{r+1})$ for $1 \leq i < r$ in the third equality. Finally, we determine $T_{\rm{bestmix}}(Y_{2r+1})$. Let the spine be $(w_1, w_2, \ldots, w_{2r})$ with a leaf $x$ adjacent to $w_{r+1}$. Arguing similarly to the even path, we have \begin{eqnarray*} T_{\rm{bestmix}}(Y_{2r+1}) &=& \mathrm{Pess}(w_{r+1}) - \sum_{i=1}^{r} \pi_i H(w_i, w_{r+1}) - \sum_{i=r+2}^{2r} H(w_i, w_{r+1}) - \pi_x H(x,w_{r+1}) \\ &=& \mathrm{Pess}(w_{r+1}) - 2\sum_{i=1}^{r} \pi_i H(w_i, w_r) - \sum_{i=1}^{r} H(v_r, w_{r+1}) - \pi_x H(x,w_{r+1}) \\ &=& r^2 - \frac{1}{2r} (r-1)^2 - \frac{1}{r} \sum_{i=2}^{r-1} ((r-1)^2 - (i-1)^2) - \frac{2r-1}{4r} (2r-1) - \frac{1}{4r} \\ &=& \frac{n^2 + 4n - 15}{12}. \end{eqnarray*} These calculations show that for odd $n=2r+1 \geq 5$, we have $T_{\rm{bestmix}}(Y_n) = T_{\rm{bestmix}}(P_n)+ (n-3)/2$. Furthermore, we can identify precisely what leads to the wishbone's advantage in equation \eqref{eqn:mix}. The pessimal hitting times to the best mix foci of $Y_{2r+1}$ and $P_{2r+1}$ are both equal to $r^2.$ So the average hitting time to the best mix focus makes the difference: this quantity is smaller for $Y_n$ We will see below that this effect can be attributed to the fact that $P_{2n+1}$ has a single focus. This special balance to the tree actually makes it \emph{easier} to mix from the focus. With these calculations in hand, we have proven the lower bound of Theorem \ref{thm:bestmix}, and calculated the two expressions that appear in the upper bound. The remainder of the paper proves that these upper bounds are correct. \section{Caterpillars and Tree Surgeries} In this section, we show that for a fixed $n \geq 4$, the tree on $n$ vertices that maximizes the best mixing time is a caterpillar. Next, we describe our basic techniques, called \emph{tree surgeries}, for making incremental changes to a caterpillars. We introduce some notation and useful lemmas for showing that the best mixing time monotonically increases after each tree surgery. \subsection{Tree to Caterpillar} We begin with a lemma showing that $T_{\rm bestmix}$ is maximized by a caterpillar. Recall that a \emph{caterpillar} is a tree such that every vertex is distance at most one from a fixed central path $W = \{ w_1, w_2, \ldots , w_t \}$, called the \emph{spine}. We employ the following notation for talking about sections of the caterpillar. For $2 \leq i \leq t-1$, let $U_i \subset V \backslash W$ denote the set of pendant leaves adjacent to $w_i$. Note that the spinal leaves are not in $V \backslash W$, so $w_1 \notin U_2$ and $w_t \notin U_{t-1}$. Finally, we also define $V_k = U_k \cup \{w_k \}$. See Figure \ref{fig:treetocat} for an example of $w_i$ and $U_i \subset V_i$. \begin{lemma} \label{lem:cat} Let $G$ be a tree on $n$ vertices that is not a caterpillar. There exists a caterpillar $\widetilde{G}$ on $n$ vertices such that $T_{\rm{bestmix}}(\widetilde{G}) > T_{\rm{bestmix}}(G)$. \end{lemma} \begin{proof} We give a simple construction of the caterpillar $\widetilde{G}$; an example is shown in Figure \ref{fig:treetocat}. \begin{figure}[t] \begin{center} \begin{tikzpicture}[scale=.6, every node/.style={font=\small}] \begin{scope} \draw (0,0) -- (9,0); \foreach \i in {0, ..., 9} { \draw[fill] (\i,0) circle (4pt); } \draw (4,-2.54) -- (3.5,-1.77) -- (3,-2.54); \draw (3.5,-1.77) -- (3,-1) -- (2.5,-1.77); \draw (3,-1) -- (3,0); \draw[fill=black!25] (3,-1) circle (4pt); \draw[fill=black!25] (2.5,-1.77) circle (4pt); \draw[fill=black!25] (3.5,-1.77) circle (4pt); \draw[fill=black!25] (3,-2.54) circle (4pt); \draw[fill=black!25] (4,-2.54) circle (4pt); \draw (5,-2) -- (5,0); \draw[fill=black!50] (5,-1) circle (4pt); \draw[fill=black!50] (5,-2) circle (4pt); \draw (7,-1) -- (7,0); \draw (7.5,-1.77) -- (7,-1) -- (6.5,-1.77); \draw[fill=black!75] (7,-1) circle (4pt); \draw[fill=black!75] (6.5,-1.77) circle (4pt); \draw[fill=black!75] (7.5,-1.77) circle (4pt); \node[left] at (5,-1) {$v$}; \node[above] at (-.5,0) {$\p{v}$}; \node[above] at (9.5,0) {$\psq{v}$}; \draw[gray] (7, -1.6) ellipse (.8 and .9); \draw[gray] (7, -1.1) ellipse (.95 and 1.75); \draw node at (7, -2.1) {$U_i$}; \draw node at (7, .3) {$w_i$}; \draw node at (8.33,-1.25 ) {$V_i$}; \node at (.5,-1.5) {$G$}; \end{scope} \node[single arrow, fill=gray, style={font=\tiny}] at (11.5,0) {\phantom{xxx}}; \begin{scope}[shift={(14,0)}] \draw (0,0) -- (9,0); \foreach \i in {0, ..., 9} { \draw[fill] (\i,0) circle (4pt); } \draw (2.6,-1) -- (3,0) -- (3.4,-1); \draw (2.2,-1) -- (3,0) -- (3.8,-1); \draw (3,-1) -- (3,0); \draw[fill=black!25] (3,-1) circle (4pt); \draw[fill=black!25] (2.6,-1) circle (4pt); \draw[fill=black!25] (3.4,-1) circle (4pt); \draw[fill=black!25] (2.2,-1) circle (4pt); \draw[fill=black!25] (3.8,-1) circle (4pt); \draw (4.8,-1) -- (5,0) -- (5.2,-1); \draw[fill=black!50] (4.8,-1) circle (4pt); \draw[fill=black!50] (5.2,-1) circle (4pt); \draw (7,-1) -- (7,0); \draw (7.4,-1) -- (7,0) -- (6.6,-1); \draw[fill=black!75] (7,-1) circle (4pt); \draw[fill=black!75] (7.4,-1) circle (4pt); \draw[fill=black!75] (6.6,-1) circle (4pt); \node[below] at (4.6,-1) {$v$}; \node[above] at (-.5,0) {$\p{v}$}; \node[above] at (9.5,0) {$\psq{v}$}; \draw[gray] (7, -1.25) ellipse (.75 and .7); \draw[gray] (7, -.85) ellipse (.9 and 1.5); \draw node at (7, -1.6) {$U_i$}; \draw node at (7, .3) {$w_i$}; \draw node at (8.25,-1 ) {$V_i$}; \node at (.5,-1.5) {$\widetilde{G}$}; \end{scope} \end{tikzpicture} \caption{Transforming a tree $G$ into a caterpillar $\widetilde{G}$ with $T_{\rm bestmix}(G) < T_{\rm bestmix}(\widetilde{G})$. Every vertex in $U_i$ becomes adjacent to $w_i$.} \label{fig:treetocat} \end{center} \end{figure} Choose any $v\in V(G)$. Recall from Section \ref{sec:pessimal} that $\p{v}$ is a $v$-pessimal vertex, and that $\psq{v}$ is a $\p{v}$-pessimal vertex. Let $ W=\{\p{v}=w_1,w_2,\dots,w_t=\psq{v}\} $ be the vertices on the unique $(\p{v},\psq{v})$ path $P \subset G$. These vertices will be the spine of the caterpillar $\widetilde{G}$. (Note that $v$ is not necessarily a member of $W$.) We replace the subforest attached to each spine vertex $w_i$ with a set of leaves adjacent to $w_i$. Define $\widetilde{G}$ to be the caterpillar with vertex set $V(G)$ and edge set $ E(\widetilde{G})= E(P) \cup \{uw_i\mid u\in U_i, 1<i<t\}. $ The foci of $\widetilde{G}$ are contained in $W$ because leaves cannot be foci when $n \geq 3$. Without loss of generality, let $w_r$ be the best mix focus of $\widetilde{G}$ and let $\p{w_r} = w_1$. If $G$ is bifocal, the other focus must be $w_{r-1}$ and $\p{w_{r-1}} = w_t$. For $1\leq i\leq t-1$, the quantity $$\sum_{v \in V_{w_i:w_{i+1}}} \deg_G(v) = 2 |V_{w_i:w_{i+1}}| - 1 = \sum_{v \in V_{w_i:w_{i+1}}} \deg_{\widetilde{G}}(v) $$ does not change. By equation \eqref{eq:adjhtime} we have \begin{equation} \label{eqn:nochange} H_G(w_i, w_j) = H_{\widetilde{G}}(w_i,w_j) \end{equation} for all $1 \leq i,j \leq t$. Considering any non-spine vertex $v \in U_i$, we have $$ H_{\widetilde{G}}(v, w_r) - H_G(v,w_r) = 1 + H_{\widetilde{G}}(w_i, w_r) - \big( H_{G}(v, w_i) + H_{G}(w_i, w_r) \big) = 1 - H_G(v,w_i) \leq 0. $$ We conclude that $w_1$ is $v_r$-pessimal for both $G$ and $\widetilde{G}$. Likewise, $w_t$ is $w_{r-1}$-pessimal for both $G$ and $\widetilde{G}$. By equation \eqref{eqn:nochange} and Theorem \ref{thm:tbmvertex}, the vertex $w_r$ is the best mix focus of $\widetilde{G}$ and $\mathrm{Pess}_G(w_r) = \mathrm{Pess}_{\widetilde{G}}(w_r)$. Finally, Theorem \ref{thm:mix} yields \begin{eqnarray*} \lefteqn{ T_{\rm bestmix}(\widetilde{G}) - T_{\rm bestmix}(G) \, = \, H_{\widetilde{G}}(w_r, \widetilde{\pi}) - H_{G}(w_r, \pi) } \\ &=& \sum_{i=1}^{t} \sum_{v \in V_i} \left( \pi_{v} (H_G (v, w_i)+ H_G (w_i, w_r)) - \widetilde{\pi}_{v} (H_{\widetilde{G}} (v, w_i) + H_{\widetilde{G}} (w_i, w_r)) \right) \\ &=& \sum_{i=1}^{t} \sum_{v \in V_i} \left( \pi_{v} H_G (v, w_i) - \widetilde{\pi}_{v} H_{\widetilde{G}} (v, w_i) \right) \, = \, \sum_{i=2}^{t-1} \sum_{v \in U_i} \left( \pi_{v} H_G (v, w_i) - \widetilde{\pi}_{v} H_{\widetilde{G}} (v, w_i) \right) \, \geq \, 0 \end{eqnarray*} because each $ \sum_{v \in V_i} \pi_{v} H_G (v, w_i) - \widetilde{\pi}_{v} H_{\widetilde{G}} (v, w_i) \geq 0$, with equality holding if and only if every $v \in U_i$ is already adjacent to $w_i$. Since $G$ was not a caterpillar, there must be at least one exceptional $v$ in some $U_i$, so this inequality is strict. \end{proof} \subsection{Leaf Transplants and Tree Surgeries} For the remainder of the paper, $G$ is a caterpillar on $n$ vertices with spine $W= \{ w_1, w_2, \ldots , w_t \}$. We choose to view $w_1$ as the leftmost vertex and $w_t$ as the rightmost vertex. For $v \in V$, we define the \emph{left pessimal hitting time} $L(v) = H(w_1,v)$ and the \emph{right pessimal hitting time} $R(v) = H(w_t,v)$. Of course, $\mathrm{Pess}(v) = \max \{ L(v), R(v) \}$. When $G$ is bifocal, we always label the foci as $w_{r-1}$ and $w_r$ where $w_r$ achieves $H(w_r,\pi)=T_{\rm bestmix}(G)$. In this case, we view $\{ w_1, \ldots , w_{r-2} \}$ as the \emph{left spine} and $\{ w_{r+1}, \ldots , w_{t} \}$ as the \emph{right spine}. The foci $\{ w_{r-1}, w_r \}$ are considered the \emph{central spine}. Note that this arrangement guarantees that $\p{w_r}=w_1$ and $\p{w_{r-1}}=w_t$. We collect some basic results about caterpillar foci in the next lemma. \begin{lemma} \label{lemma:focuscat} Let $G$ be a caterpillar with spine $W= \{ w_1, w_2, \ldots , w_t \}$. Then \begin{enumerate} \item[(a)] The vertex $w_r$ is the unique focus of $G$ if and only if $L(w_r)=R(w_r)$. \item[(b)] The foci of $G$ are $w_{r-1}, w_r$ if and only if $L(w_r)>R(w_r)$ and $R(w_{r-1})>L(w_{r-1}).$ \item[(c)] If $G$ has foci $w_{r-1}, w_r$ then vertex $w_r$ is a best mix focus if and only if $L(w_{r-1})\leq R(w_{r}).$ \item[(d)] If $L(w_r) > R(w_r)$ and $L(w_{r-1})\leq R(w_{r})$ then $w_r$ is a best mix focus and $w_{r-1}$ is also a focus of $G$. \end{enumerate} \end{lemma} \begin{proof} Parts (a) and (b) are a direct consequence of Definition \ref{def:focus}. Part (c) is restatement of Theorem \ref{thm:tbmvertex} for caterpillars. For part (d), we observe that $L(w_{r-1})\leq R(w_r) < R(w_{r-1})$, so that $w_{r-1}$ and $w_r$ satisfy (b) and (c). \end{proof} When $U_i \neq 0$, we define the \emph{leaf transplant} $\tr{i}{j}$ as the relocation of $x \in U_i$ so that this leaf is now adjacent to $w_j$ where $1 \leq j \leq t$. Usually, we have $2 \leq j \leq t-1$, so that the leaf $x$ becomes an element of $U_j$. Our most common operation will be an \emph{elementary leaf transplant} where $j=i \pm 1$. Occasionally, we will take $j \in \{1, t\}$, so that the leaf transplant actually increases the length of the spine. Finally, we need one special transplant that reduces the length of the spine by relocating the leaf $w_t$ to become adjacent to the best mix focus $w_r$: we use $\str{t}{r}$ to denote this rare spinal leaf transplant. The more general term \emph{tree surgery} denotes either a single leaf transplant (either $\tr{i}{j}$ or $\str{t}{r}$), or a pair of leaf transplants $\tr{i_1}{j_1} \wedge \tr{i_2}{j_2}$ performed simultaneously. We use $\mathcal{S}(G)$ to denote the caterpillar that results from applying tree surgery $\mathcal{S}$ to caterpillar $G$. Our algorithm uses eleven different tree surgeries. They are enumerated in Tables \ref{table:phase1}, \ref{table:phase2} and \ref{table:phase3} below. The remainder of this section is devoted to proving some fundamental results about the effects of tree surgeries on the best mixing time. Let $G=(V,E)$ be a caterpillar and let $\widetilde{G}=\mathcal{S}(G) = (V, \widetilde{E})$ be the caterpillar obtained by applying tree surgery $\mathcal{S}$, with stationary distribution $\widetilde{\pi} = \pi_{\widetilde{G}}$. We introduce notation for the comparison of various caterpillar measurements. Just as with the pessimal notation introduced in Section \ref{sec:pessimal}, we are motivated to shield ourselves from the particular locations of important vertices (like foci and pessimal vertices) which may be different in $G$ and $\widetilde{G}$. As with equation \eqref{eqn:pessdiff}, we introduce $\Delta$-notation to compactly represent the changes in various quantities due to a tree surgery. For vertices $v, v_1, v_2 \in V$, we define: $$ \begin{array}{ccc} \begin{array}{rcl} \Delta \deg(v) &=& \deg_{\widetilde{G}}(v) - \deg_G(v), \\ \Delta \pi(v) &=& \widetilde{\pi}(v) - \pi(v) , \\ \Delta H(v_1, v_2) &=& H_{\widetilde{G}}(v_1, v_2) - H_G(v_1, v_2) , \\ \Delta H(\pi, v) &=& H_{\widetilde{G}}(\widetilde{\pi}, v) - H_G(\pi, v), \\ \end{array} &\quad& \begin{array}{rcl} \Delta \mathrm{Pess}(v) &=& \mathrm{Pess}_{\widetilde{G}}(v) - \mathrm{Pess}_G(v), \\ \Delta L(v) &=& L_{\widetilde{G}}(v) - L_G(v), \\ \Delta R(v) &=& R_{\widetilde{G}}(v) - R_G(v), \\ \Delta T_{\rm bestmix} &=& T_{\rm bestmix}(\widetilde{G}) - T_{\rm bestmix}(G). \\ \end{array} \end{array} $$ Let $G$ be a caterpillar and let $\widetilde{G}=\mathcal{S}(G)$ be the result of a leaf transplant $\tr{i}{j}$. We give formulas for $\Delta H(v,w_r)$ for use in later sections. We only consider leaf transplants $\tr{i}{j}$ where either $1 \leq i,j \leq r$ or where $r+1 \leq i,j \leq t$ (so that we never transplant leaves from the left spine to the right spine, or vice versa). The formulas below cover two qualitatively different cases for $\tr{i}{j}$. When $2 \leq j \leq t-1$, the spine length is unaffected. However, when $j \in \{ 1, t \}$, the spine length increases. The formulas below still hold, though the arguments are slightly different. All four formulas below follow quickly from equation \eqref{eq:adjhtime}, and we leave these short calculations to the reader. Figure \ref{fig:transplants} shows the effect of two example transplants on hitting times. The arguments for $j=i \pm 1$ are straight forward, and the other cases follow inductively. \begin{figure}[t] \begin{center} \begin{tikzpicture}[scale=.7, every node/.style={font=\scriptsize}] \begin{scope} \draw (1,0) -- (9,0); \draw (2,0) -- (2,-1); \draw (8,0) -- (8,-1); \draw (8,0) -- (8.5,-1); \draw (8,0) -- (7.5,-1); \foreach \i in {1,2,3,4,7,8,9} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=black!30] (2,-1) circle (4pt); \draw[fill=black!30] (7.5,-1) circle (4pt); \draw[fill=black!30] (8,-1) circle (4pt); \draw[fill=black!30] (8.5,-1) circle (4pt); \draw[fill=white] (5,0) circle (4pt); \draw[fill=white] (6,0) circle (6pt); \draw[fill=white] (6,0) circle (4pt); \node[above] at (1.5,0) {1}; \node[above] at (2.5,0) {5}; \node[above] at (3.5,0) {7}; \node[above] at (4.5,0) {9}; \node[above] at (5.5,0) {11}; \node[above] at (6.5,0) {13}; \node[above] at (7.5,0) {15}; \node[above] at (8.5,0) {23}; \node[below] at (1.5,0) {23}; \node[below] at (2.5,0) {19}; \node[below] at (3.5,0) {17}; \node[below] at (4.5,0) {15}; \node[below] at (5.5,0) {13}; \node[below] at (6.5,0) {11}; \node[below] at (7.5,0) {9}; \node[below] at (8.5,0) {1}; \end{scope} \begin{scope}[shift={(11,0)}] \draw (1,0) -- (10,0); \draw (5,0) -- (5,-1); \draw (8,0) -- (8.33,-1); \draw (8,0) -- (7.67,-1); \foreach \i in {1,2,3,4,7,8,9} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=black!30] (5,-1) circle (4pt); \draw[fill=black!30] (7.67,-1) circle (4pt); \draw[fill=black!30] (8.33,-1) circle (4pt); \draw[fill=black!30] (10,0) circle (4pt); \draw[fill=white] (5,0) circle (4pt); \draw[fill=white] (6,0) circle (6pt); \draw[fill=white] (6,0) circle (4pt); \node[above] at (1.5,0) {1}; \node[above] at (2.5,0) {3}; \node[above] at (3.5,0) {5}; \node[above] at (4.5,0) {7}; \node[above] at (5.5,0) {11}; \node[above] at (6.5,0) {13}; \node[above] at (7.5,0) {15}; \node[above] at (8.5,0) {21}; \node[above] at (9.5,0) {23}; \node[below] at (1.5,0) {23}; \node[below] at (2.5,0) {21}; \node[below] at (3.5,0) {19}; \node[below] at (4.5,0) {17}; \node[below] at (5.5,0) {13}; \node[below] at (6.5,0) {11}; \node[below] at (7.5,0) {9}; \node[below] at (8.5,0) {3}; \node[below] at (9.5,0) {1}; \end{scope} \end{tikzpicture} \end{center} \caption{The effect of transplants on hitting times. The right tree results from applying $\tr{2}{5} \wedge \tr{8}{9}$ on the left tree. Each spine edge $(w_i, w_{i+1})$ is labeled by $H(w_i,w_{i+1})$ above and $H(w_{i+1},w_i)$ below. Other hitting times are linear combinations of those shown.} \label{fig:transplants} \end{figure} Suppose that the leaf transplant $\tr{i}{j}$ moves $x$ from $U_i$ to $U_j$. For $v \in V_k$, the value of $\Delta H (v,w_r)$ depends on the relative locations of $i,j$ and $k$: \begin{align} \label{eq:leftleft2} \mbox{If } 1\leq j<i\leq r \mbox{ then } & \Delta H(v,w_r)= \left\{ \begin{array}{cl} H_G(w_j,w_i)+2(i-j) & \quad v=x,\\ 2(i-j) & \quad k\leq j,\\ 2(i-k) & \quad j<k<i,\\ 0 & \quad k \leq i \leq r \mbox{ and } v \neq x. \end{array} \right. \\ \label{eq:leftright2} \mbox{If } 1\leq i<j\leq r \mbox{ then } & \Delta H(v,w_r)= \left\{ \begin{array}{cl} -H_G(w_i,w_j) & \quad v=x,\\ -2(j-i) & \quad k\leq i \mbox{ and } v \neq x,\\ -2(j-k) & \quad i<k<j,\\ 0 & \quad j \leq k \leq r. \end{array} \right. \\ \label{eq:rightleft2} \mbox{If } r \leq j<i\leq t \mbox{ then } & \Delta H(v,w_r)= \left\{ \begin{array}{cl} -H_G(w_i,w_j) & \quad v=x,\\ -2(i-j) & \quad i \leq k \mbox{ and } v \neq x,\\ -2(i-k) & \quad j < k < i,\\ 0 & \quad r < k \leq j. \end{array} \right. \\ \label{eq:rightright2} \mbox{If } r \leq i<j\leq t \mbox{ then } & \Delta H(v,w_r)= \left\{ \begin{array}{cl} H_G(w_j,w_i)+2 & \quad v=x,\\ 2(j-i) & \quad j \leq k,\\ 2(j-k) & \quad i < k < j,\\ 0 & \quad r < k \leq i \mbox{ and } v \neq x. \end{array} \right. \end{align} We get similar equations for hitting times to the second focus $w_{r-1}$, with the appropriate change to the bounds on the indices $i,j$. This completes our list of useful hitting time changes due to common leaf transplants. Next, we turn our attention to tracking the change in the best mixing time after a tree surgery. The lemmas that follow will be used to analyze all eleven surgeries used in our algorithm. If $w_r$ is the best mix focus of $G$ and $w_s$ is the best mix focus of $\widetilde{G}$, then \begin{equation} \label{eq:tbmchange} \Delta T_{\rm{bestmix}} = \mathrm{Pess}_{\widetilde{G}}(w_s) - \mathrm{Pess}_G(w_r) - \big( H_{\widetilde{G}}(\widetilde{\pi}, w_s) - H_G(\pi, w_r) \big). \end{equation} As we alter our caterpillar, we strive to maintain $w_r$ as the best mix focus and the leftmost spinal vertex as the $w_r$-pessimal vertex. In this case, equation \eqref{eq:tbmchange} simplifies to $ \Delta T_{\rm{bestmix}} = \Delta \mathrm{Pess}(w_r) - \Delta H(\pi, w_r) = \Delta L(w_r) - \Delta H(\pi, w_r). $ This can be rephrased as the simple criterion: \begin{equation} \label{eq:tbmchange2} \Delta T_{\rm{bestmix}} \geq 0 \quad \Longleftrightarrow \quad \Delta\mathrm{Pess}(w_r) \geq \Delta H(\pi,w_r). \end{equation} The next two results track the effect of a surgery on the location of the foci. \begin{lemma} \label{lem:criteria} Let $G$ be a caterpillar. Let $\mathcal{S}$ be a tree surgery and let $\widetilde{G}=\mathcal{S}(G)$. If \begin{eqnarray} \label{eq:lchange} L_G(w_r)-R_G(w_r) &>& \DeltaR(w_r)-\DeltaL(w_r) \\ \mbox{and} \quad \label{eq:newtbmfoc} R_G(w_r)-L_G(w_{r-1}) &\geq& \DeltaL(w_{r-1})-\DeltaR(w_r), \end{eqnarray} then $\widetilde{G}=\mathcal{S}(G)$ has best mix focus $w_{r}$ and $w_{r-1}$ is also a focus. \end{lemma} \begin{proof} These criteria are equivalent to the conditions of Lemma \ref{lemma:focuscat} (d) for $\widetilde{G}$. \end{proof} \begin{corollary} \label{cor:criteria} Let $G$ be a bifocal caterpillar with focus $w_{r-1}$ and best mix focus $w_r$. Let $\mathcal{S}$ be a tree surgery and let $\widetilde{G}=\mathcal{S}(G)$. If $\DeltaR(w_r)-\DeltaL(w_r) \leq 0$ and $\DeltaL(w_{r-1})-\DeltaR(w_r)\leq 0$ then $\widetilde{G}=\mathcal{S}(G)$ also has focus $w_{r-1}$ and best mix focus $w_r$. \end{corollary} \begin{proof} This follows directly from Lemma \ref{lem:criteria} and the inequalities $L(w_r)-R(w_r)\geq 1$ and $R(w_r)-L(w_{r-1})\geq 0$. \end{proof} Lemma \ref{lem:criteria} and Corollary \ref{cor:criteria} are key tools of our methodology. We use them repeatedly to verify the foci and best mix focus of our caterpillar after a surgery has been applied. Once we know that $w_r$ is still the best mix focus, we then verify that equation \eqref{eq:tbmchange2} holds. Next, we give a final test for $\Delta T_{\rm bestmix} \geq 0$. This lemma will be used very frequently in subsequent sections. \begin{lemma} \label{lem:slack} Let $G$ be a caterpillar and let $\mathcal{S}$ be a tree surgery such that $w_r$ is a best mix focus of $G$ and $\widetilde{G}$. Let $A=\{v\in V\mid \Delta H(v,w_r)>\Delta\mathrm{Pess}(w_r)\}$ and let $C = \{ v \in \overline{A} \mid \Delta(v) > 0 \}$. If there exists $C \subset B \subset \overline{A}$ for which \begin{align} \label{eqn:slack} &\sum_{v\in A}{ \Big(\deg_{\widetilde{G}}(v)(\Delta H(v,w_r)-\Delta\mathrm{Pess}(w_r))+\Delta\deg_G(v)H_G(v,w_r) \Big)} \nonumber \\ \leq&\sum_{u\in B}{ \Big( \deg_{\widetilde{G}}(u)(\Delta\mathrm{Pess}(w_r)-\Delta H(u,w_r))-\Delta\deg_G(u)H_G(u,w_r) \Big)} \end{align} then $\Delta T_{\rm{bestmix}} \geq 0$. \end{lemma} Before proving this lemma, we make a few comments its use in later sections. First, the set $A$ will always be small: it will only contain one or both of the leaves moved during the surgery. Second, if the lemma holds for $B$, then it holds for any superset of $B$, including $\overline{A}$. However, there is no need to calculate the right hand side for $\overline{A}$ when a small subset will do. For our earlier surgeries, the set $B$ will be small, typically consisting of a handful of spine vertices near the transplant locations. In later surgeries, $B$ will consist of half or all of the spine. \begin{proof} We decompose $\Delta H(\pi,w_r)$ as follows: \begin{align} \Delta H(\pi,w_r) &=\sum_{v\in V}{\big(\widetilde{\pi}(v)H_{\widetilde{G}}(v,w_r)-\pi(v)H_G(v,w_r)\big)}\nonumber \\ &=\sum_{v\in V}\big(\widetilde{\pi}(v)(\Delta H(v,w_r) + H_G(v,w_r))-\pi(v)H_G(v,w_r)\big)\nonumber \\ &=\sum_{v\in V}{\big(\widetilde{\pi}(v)\Delta H(v,w_r)+\Delta\pi(v)H_G(v,w_r)\big)}. \nonumber \end{align} Let $g(v) = \deg_{\widetilde{G}}(v)(\Delta\mathrm{Pess}(w_r) - \Delta H(v,w_r)) - \Delta\deg_G(v)H_G(v,w_r)$. We can rewrite \begin{align*} \Delta\mathrm{Pess}(w_r)-\Delta H(\pi,w_r) & = \Delta\mathrm{Pess}(w_r)-\sum_{v\in V}{\left(\widetilde{\pi}(v)\Delta H(v,w_r)+\Delta\pi(v)H_G(v,w_r)\right)}\\ &= \frac{1}{2|E|} \left( \sum_{v\in A} g(v) + \sum_{u\in \overline{A}} g(u) \right) \, \geq \,\frac{1}{2|E|} \left( \sum_{v\in A} g(v) + \sum_{u\in B} g(u) \right). \end{align*} The final inequality holds because $g(u) \geq 0$ for every $u \in \overline{A} \backslash B \subset \overline{A} \backslash C$. Therefore, if $ - \sum_{v\in A} g(v) \leq \sum_{u\in B} g(u)$ then $ \Delta\mathrm{Pess}(w_r)-\Delta H(\pi,w_r) \geq 0$. \end{proof} We have an immediate corollary for the special case where $A$ is empty. \begin{corollary} \label{cor:slack} Given a tree $G$ and any tree surgery $\mathcal{S}$ such that $w_r$ is the best mix focus of $G$ and $\widetilde{G}$, if $ \Delta H(v,w_r)\leq\Delta\mathrm{Pess}(w_r), $ for all $v\in V$, then $\Delta T_{\rm{bestmix}} \geq 0$. \hfill$\Box$ \end{corollary} Finally, we note that if the $w_r$-pessimal vertices in $G$ and $\widetilde{G}$ are the leftmost spinal vertices in these graphs, then we can replace $\Delta\mathrm{Pess}(w_r)$ with $\DeltaL(w_r)$ in Lemma \ref{lem:slack} and Corollary \ref{cor:slack}. \section{Proof of the upper bound in Theorem \ref{thm:bestmix}} \label{sec:proof} In Section \ref{sec:bestmix-extreme}, we proved the lower bound of Theorem \ref{thm:bestmix}, and calculated the best mix time for the even path and the odd wishbone. In this section, we prove the upper bound of Theorem \ref{thm:bestmix}, leaving the details to the lemmas in the accompanying subsections. Two examples of the algorithm in practice are shown in Figure \ref{fig:examples} above. \begin{proofof}{Theorem \ref{thm:bestmix}} Given a tree $G$ that is not an even path or an odd wishbone, we must show that it does not maximize the best mixing time. We may assume that $G$ is not a star since that is clearly not the maximizing structure. Starting from our tree $G$, we let $G_0$ be the caterpillar constructed by Lemma \ref{lem:cat}, so that $T_{\rm bestmix}(G) \leq T_{\rm bestmix}(G_0)$. Next, we apply a sequence of tree surgeries to produce a sequence of caterpillars $G_0,G_1, G_2, \ldots, G_m$ such that $T_{\rm bestmix}(G_{i-1}) \leq T_{\rm bestmix}(G_i)$, where the final caterpillar $G_m$ is either $P_n$ or $Y_n$. This occurs in an algorithmic manner, divided into three phases. \begin{figure}[t] \begin{center} \begin{tikzpicture}[scale=.66] \node at (0,-2.5) {(c)}; \draw (-1.5,1) -- (0,0) -- (2.25, 1.5); \draw (0,0) -- (0,-1); \draw[fill] (-1.5,1) circle (3pt); \draw[fill] (-.75,.5) circle (3pt); \draw[fill] (0,0) circle (3pt); \draw[fill] (.75,.5) circle (3pt); \draw[fill] (1.5,1) circle (3pt); \draw[fill] (2.25,1.5) circle (3pt); \draw[fill] (0,-1) circle (3pt); \begin{scope}[shift={(-5,-.5)}] \node at (.75,-2) {(b)}; \draw (0,3) -- (0,0) -- (1.5,0) -- (1.5,2); \draw (0,-1) -- (0,0); \draw (.33,-1) -- (0,0); \draw (-.33,-1) -- (0,0); \draw (1.5,-1) -- (1.5,0); \draw (1.83,-1) -- (1.5,0); \draw (2.16,-1) -- (1.5,0); \draw (1.17,-1) -- (1.5,0); \draw[fill] (0,3) circle (3pt); \draw[fill] (0,2) circle (3pt); \draw[fill] (0,1) circle (3pt); \draw[fill] (0,0) circle (3pt); \draw[fill] (1.5,0) circle (3pt); \draw[fill] (1.5,1) circle (3pt); \draw[fill] (1.5,2) circle (3pt); .3 \draw[fill] (0,-1) circle (3pt); \draw[fill] (.33,-1) circle (3pt); \draw[fill] (-.33,-1) circle (3pt); \draw[fill] (1.5,-1) circle (3pt); \draw[fill] (1.83,-1) circle (3pt); \draw[fill] (2.16,-1) circle (3pt); \draw[fill] (1.17,-1) circle (3pt); \end{scope} \begin{scope}[shift={(-12,0)}] \node at (.5,-2.5) {(a)}; \draw (-3,0) -- (0,0) -- (1,0) -- (5,0); \draw (0,-1) -- (0,0); \draw (.33,-1) -- (0,0); \draw (-.33,-1) -- (0,0); \draw (1.2,-1) -- (1,0); \draw (.8,-1) -- (1,0); \draw (3,1) -- (3,0); \draw (-2,1) -- (-2,0); \draw[fill] (-3,0) circle (3pt); \draw[fill] (-2,0) circle (3pt); \draw[fill] (-1,0) circle (3pt); \draw[fill] (0,0) circle (3pt); \draw[fill] (1,0) circle (3pt); \draw[fill] (2,0) circle (3pt); \draw[fill] (3,0) circle (3pt); \draw[fill] (4,0) circle (3pt); \draw[fill] (5,0) circle (3pt); \draw[fill] (-2,1) circle (3pt); \draw[fill] (3,1) circle (3pt); \draw[fill] (0,-1) circle (3pt); \draw[fill] (.33,-1) circle (3pt); \draw[fill] (-.33,-1) circle (3pt); \draw[fill] (1.2,-1) circle (3pt); \draw[fill] (.8,-1) circle (3pt); \end{scope} \end{tikzpicture} \end{center} \caption{Examples of the milestone graphs obtained at the end of each phase of our process: (a) a seesaw, (b) a twin broom, and (c) the wishbone $Y_7$.} \label{fig:menagerie} \end{figure} In order to define these phases, we define two special subfamilies of caterpillars which mark milestones in our graph sequence. Examples of these graphs and a wishbone are shown in Figure \ref{fig:menagerie}. A \emph{twin broom} is a bifocal caterpillar whose only non-spine leaves are adjacent to the foci $w_{r-1}, w_r$. In other words, $V= W \cup U_{r-1} \cup U_r$. A \emph{seesaw} is a twin broom that has at most one additional leaf on each side of the spine. In other words, a seesaw satisfies $| \cup_{i=2}^{r-2} U_i| \leq 1$ and $| \cup_{i=r+1}^{t-1} U_i| \leq 1$. Figure \ref{fig:examples} includes examples of each. In that figure, the output of $S_4$ and the initial tree on 11 vertices are both seesaws. Meanwhile, the outputs of $S_6$ and $S_8$ are both twin brooms (as are all graphs that follow them). We start with a caterpillar $G$ on $n$ vertices. In Phase One, we convert the caterpillar into a seesaw. Phase Two converts the seesaw into a twin broom. Phase Three converts the twin broom into one of $P_{n}$ (when $n$ is even) or $Y_n$ (when $n$ is odd). Lemmas \ref{lemma:cat-to-seesaw}, \ref{lemma:seesaw-to-twinbroom}, and \ref{lemma:twinbroom-to-end} below show that the best mixing time monotonically increases in every phase of this process. This proves Theorem \ref{thm:bestmix}. \end{proofof} Our caterpillar transformation consists of incremental steps that move one or two leaves at a time. This allows us to monitor the delicate balance maintained by the best mix focus. In particular, there may be a critical step at which we change the focus that attains the best mixing time. This happens in one of two ways. Usually, we make a small change that keeps the current best mix focus, but also causes a neighbor to also become a best mix focus. The other change is more abrupt: Surgery $S_5$ below changes the location of the unique best mix focus to a neighbor of the current one. In $S_5$, the prescribed caterpillar structure during that surgery makes this crucial transition manageable. The remainder of this paper is devoted to proving Lemmas \ref{lemma:cat-to-seesaw}, \ref{lemma:seesaw-to-twinbroom}, and \ref{lemma:twinbroom-to-end}. \subsection{Phase One: Caterpillar to Seesaw} In this subsection, we prove that Phase One is successful: we can transform any caterpillar into a seesaw while also increasing the best mixing time. Let $G$ be a caterpillar with spine $W=\{w_1, w_2, \ldots, w_t \}$ where $w_r$ is the best mix focus and $\p{w_r}=w_1$. \begin{lemma} \label{lemma:cat-to-seesaw} Let $G$ be a caterpillar $G$ on $n$ vertices. Phase One creates a seesaw $\widetilde{G}$ such that $T_{\rm{bestmix}}(\widetilde{G})\geqT_{\rm{bestmix}}(G)$. \end{lemma} \begin{proof} If $G$ is already a seesaw then $\widetilde{G}=G$. Table \ref{table:phase1} shows the five surgery types employed during Phase One. We defer the proofs that $\Delta T_{\rm bestmix} \geq 0$ for each of these surgeries to the lemmas that follow. Figure \ref{fig:phase1} shows the workflow for Phase One. First, if the caterpillar has a unique focus then we use $\mathcal{S}_1$ to create a caterpillar with two foci. From here forward, the caterpillar will remain bifocal. Let $w_{r-1}, w_r$ be the foci of a bifocal caterpillar with $w_r$ achieving $T_{\rm bestmix}$. A leaf $x \in V \backslash W$ is \emph{good} when $x \in U_{2} \cup U_{r-1} \cup U_{r} \cup U_{t-1}$. All other leaves in $V \backslash W$ are \emph{bad}. First, we repeatedly use $\mathcal{S}_2$ to move pairs of bad leaves on the same side of the spine (one towards the end and the other towards the center). This loop terminates whether there is at most one bad leaf on each side of the spine. At this point, we use $S_3$ to extend the left spine and to transplant a leaf to $U_r$. This requires there are at least two left leaves (which can only happen when at least one is in $U_2$ since we are done with $\mathcal{S}_2$). After extending the spine, the previously good leaves at $U_2$ become bad. This throws us back into the $\mathcal{S}_2$ loop. We exit the $S_3$ loop when there is at most one left leaf. At this point, we deal with the right leaves. When $L(w_r) > R(w_r) + 1$, we apply $S_4$, the right-hand surgery analogous to $S_3$. However, if $L(w_r)= R(w_r)+1$ then applying $S_4$ would create a focal caterpillar, which we choose to avoid. Instead, we apply $S_5$, which transplants a single right leaf to the end of the right spine. Applying either $S_4$ or $S_5$ might create bad right leaves, which puts us back into the $S_2$ loop. Surgeries $S_3, S_4, S_5$ reduce the number of non-spinal vertices, so Phase One must terminate. Ultimately, we create a caterpillar with at most one left leaf and at most one right leaf, while $U_{r-1} \cup U_{r}$ may contain many leaves. This is a seesaw graph, as desired. \end{proof} \begin{table}[ht] \begin{center} \begin{tabular}{|c|p{2.5in}|@{}c@{}|} \hline Surgery & Initial Conditions & Illustration\\ \hline $\mathcal{S}_1$ & \small $G$ is focal, so $L(w_r) = R(w_r)$. The transplant depends on whether $U_{t-1} = \emptyset$. & \begin{tabular}{c} \begin{tikzpicture}[scale=0.5, decoration={ markings, mark=at position 1 with {\arrow[scale=1.25,black]{latex}}; } ] \begin{scope}[shift={(0,-2)}] \draw[gray!70] (2,0) -- (2,-1); \draw[gray!70, fill=gray!70] (2,-1) circle (4pt); \draw[dashed] (0,0) -- (5,0); \draw (0,0) -- (.5,0); \draw (1.5,0) -- (2.5,0); \draw (3.5,0) -- (5,0); \foreach \i in {0,4,5} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=white] (2,0) circle (6pt); \draw[fill=white] (2,0) circle (4pt); \draw[bend left=20, postaction=decorate] (4.75, -.25) to (2.25,-1); \end{scope} \node at (-1.5,-1.75) {\small{or}}; \begin{scope} \draw[white] (0,0.5) circle (1pt); \draw[gray!70] (2,0) -- (2,-1); \draw[gray!70, fill=gray!70] (2,-1) circle (4pt); \draw[dashed] (0,0) -- (5,0); \draw (0,0) -- (.5,0); \draw (1.5,0) -- (2.5,0); \draw (3.5,0) -- (5,0); \draw[gray!70] (2,0) -- (2,-1); \draw[gray!70, fill=gray!70] (2,-1) circle (4pt); \foreach \i in {0,4,5} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=white] (2,0) circle (6pt); \draw[fill=white] (2,0) circle (4pt); \draw (4,0) -- (4,-1); \draw[fill] (4,-1) circle (4pt); \draw[bend left=20, postaction=decorate] (3.75, -1.25) to (2.25,-1.25); \end{scope} \end{tikzpicture} \end{tabular} \\ \hline $\mathcal{S}_2$ & \small $G$ has two (or more) bad left leaves or has two (or more) bad right leaves; \newline $R(w_r) \geq L(w_{r-1})$. & \begin{tabular}{c} \begin{tikzpicture}[scale=0.5, decoration={ markings, mark=at position 1 with {\arrow[scale=1.25,black]{latex}}; } ] \draw[color=white] (0, .5) circle (1pt); \draw[gray!70] (6,0) -- (6,-1); \draw[gray!70, fill=gray!70] (6,-1) circle (4pt); \draw(5,0) -- (5,-1); \draw[fill] (5,-1) circle (4pt); \draw[gray!70] (2,0) -- (2,-1); \draw[gray!70, fill=gray!70] (2,-1) circle (4pt); \draw(3,0) -- (3,-1); \draw[fill] (3,-1) circle (4pt); \draw[dashed] (0,0) -- (11,0); \draw (0,0) -- (.5,0); \draw (1.5,0) -- (3.5,0); \draw (4.5,0) -- (6.5,0); \draw (7.5,0) -- (9.5,0); \draw (10.5,0) -- (11,0); \foreach \i in {0,2,3,5,6,11} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=white] (8,0) circle (4pt); \draw[fill=white] (9,0) circle (6pt); \draw[fill=white] (9,0) circle (4pt); \draw[bend left=15, postaction=decorate] (2.9, -1.25) to (2.1,-1.25); \draw[bend right=15, postaction=decorate] (5.1, -1.25) to (5.9,-1.25); \node at (3.3,-1.25) {\scriptsize $x$}; \node at (4.6,-1.25) {\scriptsize $y$}; \end{tikzpicture} \end{tabular} \\ \hline $\mathcal{S}_3$ & \small $G$ has $x \in U_2$ and another left leaf $y$; \newline $R(w_r) > L(w_{r-1})$ & \begin{tabular}{c} \begin{tikzpicture}[scale=0.5, decoration={ markings, mark=at position 1 with {\arrow[scale=1.25,black]{latex}}; } ] \draw[color=white] (2, .5) circle (1pt); \draw[gray!70] (6,0) -- (6,-1); \draw[gray!70, fill=gray!70] (6,-1) circle (4pt); \draw(5,0) -- (5,-1); \draw[fill] (5,-1) circle (4pt); \draw[gray!70] (1,0) -- (2,0); \draw[gray!70, fill=gray!70] (1,0) circle (4pt); \draw(3,0) -- (3,-1); \draw[fill] (3,-1) circle (4pt); \draw[dashed] (2,0) -- (11,0); \draw (2,0) -- (3.5,0); \draw (4.5,0) -- (6.5,0); \draw (7.5,0) -- (9.5,0); \draw (10.5,0) -- (11,0); \foreach \i in {2,3,5,6,11} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=white] (8,0) circle (4pt); \draw[fill=white] (9,0) circle (6pt); \draw[fill=white] (9,0) circle (4pt); \draw[bend left=20, postaction=decorate] (2.75, -1.) to (1.2,-.25); \draw[bend right=15, postaction=decorate] (5.1, -1.25) to (5.9,-1.25); \node at (3.35,-1.25) {\scriptsize $x$}; \node at (4.65,-1.25) {\scriptsize $y$}; \end{tikzpicture} \end{tabular} \\ \hline $\mathcal{S}_4$ & \small $G$ has $y\in U_{t-1}$ and another right leaf $x$; \newline $R(w_r) \geq L(w_{r-1})$ and $L(w_r)>R(w_r)+1$ & \begin{tabular}{c} \begin{tikzpicture}[scale=0.5, decoration={ markings, mark=at position 1 with {\arrow[scale=1.25,black]{latex}}; } ] \draw[color=white] (-2, .5) circle (1pt); \draw[gray!70] (-6,0) -- (-6,-1); \draw[gray!70, fill=gray!70] (-6,-1) circle (4pt); \draw(-5,0) -- (-5,-1); \draw[fill] (-5,-1) circle (4pt); \draw[gray!70] (-1,0) -- (-2,0); \draw[gray!70, fill=gray!70] (-1,0) circle (4pt); \draw(-3,0) -- (-3,-1); \draw[fill] (-3,-1) circle (4pt); \draw[dashed] (-2,0) -- (-11,0); \draw (-2,0) -- (-3.5,0); \draw (-4.5,0) -- (-6.5,0); \draw (-7.5,0) -- (-9.5,0); \draw (-10.5,0) -- (-11,0); \foreach \i in {2,3,5,6,11} { \draw[fill] (-\i,0) circle (4pt); } \draw[fill=white] (-9,0) circle (4pt); \draw[fill=white] (-8,0) circle (6pt); \draw[fill=white] (-8,0) circle (4pt); \draw[bend right=20, postaction=decorate] (-2.75, -1.) to (-1.2,-.25); \draw[bend left=15, postaction=decorate] (-5.1, -1.25) to (-5.9,-1.25); \node at (-3.35,-1.25) {\scriptsize $y$}; \node at (-4.65,-1.25) {\scriptsize $x$}; \end{tikzpicture} \end{tabular} \\ \hline $\mathcal{S}_5$ & \small $G$ has $y\in U_{t-1}$ and another right leaf $x$; \newline $R(w_r) \geq L(w_{r-1})$ and $L(w_r)=R(w_r)+1$ & \begin{tabular}{c} \begin{tikzpicture}[scale=0.5, decoration={ markings, mark=at position 1 with {\arrow[scale=1.25,black]{latex}}; } ] bifocal \draw[color=white] (-2, .5) circle (1pt); \draw(-5,0) -- (-5,-1); \draw[fill] (-5,-1) circle (4pt); \draw[gray!70] (-1,0) -- (-2,0); \draw[gray!70, fill=gray!70] (-1,0) circle (4pt); \draw(-3,0) -- (-3,-1); \draw[fill] (-3,-1) circle (4pt); \draw[dashed] (-2,0) -- (-10,0); \draw (-2,0) -- (-3.5,0); \draw (-4.5,0) -- (-5.5,0); \draw (-6.5,0) -- (-8.5,0); \draw (-9.5,0) -- (-10,0); \foreach \i in {2,3,5,10} { \draw[fill] (-\i,0) circle (4pt); } \draw[fill=white] (-8,0) circle (4pt); \draw[fill=white] (-7,0) circle (6pt); \draw[fill=white] (-7,0) circle (4pt); \draw[bend right=20, postaction=decorate] (-2.75, -1.) to (-1.2,-.25); \node at (-3.35,-1.25) {\scriptsize $y$}; \node at (-4.65,-1.25) {\scriptsize $x$}; \end{tikzpicture} \end{tabular} \\ \hline \end{tabular} \hspace{-1in}\caption{The Phase One tree surgeries. Except for $\mathcal{S}_1$, the caterpillar is bifocal with best mix focus $w_r$ and another focus $w_{r-1}$. White vertices are foci and the circled vertices are best mix foci.} \label{table:phase1} \end{center} \end{table} \begin{figure}[t] \begin{center} \includegraphics[width=5in]{flowchart_phase_I} \end{center} \caption{Phase One of the algorithm, turning a caterpillar into a seesaw.} \label{fig:phase1} \end{figure} Next, we prove that surgeries $S_1, S_2, S_3, S_4, S_5$ each result in $\Delta T_{\rm bestmix} \geq 0$. If we start with a focal caterpillar $G$, we use surgery $\mathcal{S}_1$ to create a bifocal caterpillar $\widetilde{G} = \mathcal{S}_1(G)$. Depending on the structure of $G$, we use one of two leaf transplants: $\tr{{t-1}}{r}$ or $\str{t}{r}$. We note that $\str{t}{r}$ is the only transplant that removes a leaf from the spine. \begin{lemma} \label{lem:S1} Let $G$ be a caterpillar with a single focus $w_r$ and with the spine indexed such that $d(w_1,w_r)\geq d(w_t,w_r)$. If $U_{t-1} \neq \emptyset$, then let $S_1 = \tr{{t-1}}{r}$. If $U_{t-1} = \emptyset$ then let $\mathcal{S}_1 = \str{t}{r}$. The caterpillar $\widetilde{G} = \mathcal{S}_1(G)$ is bifocal and $\Delta T_{\rm{bestmix}} > 0$. \end{lemma} \begin{proof} We use Lemma \ref{lem:criteria} for our proof. The spine is indexed so that $r-1 \geq t-r$ and the unique focus means that $L_G(w_r) = R_G(w_r).$ For both surgeries under consideration, $\Delta R(w_r) < 0$ and $\Delta L(w_r) =0$, so equation \eqref{eq:lchange} is satisfied. Next, we observe that $R(w_r) - L(w_{r-1}) = L(w_r) - L(w_{r-1}) = H(w_{r-1}, w_r) = \sum_{v \in V_{w_{r-1}:w_r}} \deg(v) \geq 2(r-1)-1$ by equation \eqref{eq:adjhtime}, so it suffices to show that $\Delta L(w_{r-1}) - \Delta R(w_r) = - \Delta R(w_r) \leq 2(r-1) -1$ to verify equation \eqref{eq:newtbmfoc}. There are two cases. First, suppose that $U_{t-1} \neq \emptyset$, so that $\mathcal{S}_1=\tr{t-1}{r}$ does not alter the spine. By equation \eqref{eq:rightright2}, we have $ -\Delta R(w_r) = 2(t-1-r) < 2(r-1)-1.$ Next, suppose that $U_{t-1} = \emptyset$, so that $\mathcal{S}_1=\str{t}{r}$ does alter the spine. We have \begin{eqnarray*} -\Delta R(w_r) &=& -H_G(w_t,w_r) + H_{\widetilde{G}}(w_{t-1},w_r) \,= \, -1 - \Delta H(w_{t-1},w_r) \\ &=& -1+2(t-1-r) \, = \, 2(t-r)-3 \, < \, 2(r-1)-1. \end{eqnarray*} In either case, the conditions of Lemma \ref{lem:criteria} are satisfied, so the foci do not change. The result now follows from Corollary \ref{cor:slack} since $\Delta L(w_r)=0 \geq \Delta H(v,w_r)$ for all $v \in V$. \end{proof} Next, we discuss surgery $\mathcal{S}_2$ which transplants a pair of bad left-hand leaves, moving one toward $U_2$ and one toward $U_{r-1}$, as shown in Table \ref{table:phase2}. We repeat $\mathcal{S}_2$ until the left side of the caterpillar has at most one bad leaf. \begin{lemma} \label{lem:S2} Let $G$ be a bifocal caterpillar with leaves $x\in U_i$ and $y\in U_j$ where either $2 < i \leq j < r-1$ or $r<i\leq j <t-1$. If $\mathcal{S}_2=\tr{i}{i-1}\wedge\tr{j}{j+1}$, then $\Delta T_{\rm{bestmix}} > 0$. \end{lemma} \begin{proof} We consider the left-hand spine case $2<i\leq j<r-1$. The right-hand spine proof is analogous, switching the roles of $i$ and $j$. Suppose that there are bad leaves $x \in U_i$ and $y \in U_j$. First, we observe that $\Delta L (w_r) = 0 = \Delta L(w_{r-1})$: these net hitting time changes are $2-2=0$ as per equations \eqref{eq:leftleft2} and \eqref{eq:leftright2}. The right hand spine is unaffected, so $\Delta R (w_r) = 0 = \Delta R(w_{r-1})$. Corollary \ref{cor:criteria} guarantees that $\widetilde{G}$ also has foci $w_{r-1}$ and best mix focus $w_r$. Having established that the foci do not change, we show that $T_{\rm{bestmix}}$ increases. We must show that $ \DeltaL(w_r)=\Delta H(w_1,w_r)>\Delta H(\pi,w_r). $ We only argue the case $i<j$, as the case $i=j$ is a straight-forward adaptation. Equations \eqref{eq:leftleft2} and \eqref{eq:leftright2} give $$ \begin{array}{rclcrcll} \Delta H(x,w_r) &=& H_G(w_{i-1},w_i), & & \Delta H(v,w_r)&=& -2 &\mbox{ for } v \in (V_i \cup \cdots \cup V_j) \backslash \{x, y \}, \\ \Delta H(y,v_r) &=& 2-H_G(w_j,w_{j+1}), && \Delta H(v,w_r) &=&0 & \mbox{ for all other } v. \end{array} $$ The only degree changes are $\Delta \deg(w_i) = -1 = \Delta \deg(w_j)$ and $\Delta \deg(w_{i-1}) = 1 = \Delta \deg(w_{j+1})$. (When $i=j$, the only negative degree change is $\Delta \deg(w_i)=-2$.) We use Lemma \ref{lem:slack} with $A= \{x \}$ to show that $T_{\rm{bestmix}}$ does not decrease. The left side of inequality \eqref{eqn:slack} simplifies to $\Delta H(x,w_r) = H_G(w_{i-1},w_i)$. It remains to show that this value is a lower bound for the right side of \eqref{eqn:slack}. Taking $B= \{w_{i-1}, w_i, w_j, w_{j+1}, y \}$, we obtain \begin{align*} \MoveEqLeft -H_G(w_{i-1},w_r) + \left(2\deg_{\widetilde{G}}(w_i)+H_G(w_i,w_r) \right)+ \left(2\deg_{\widetilde{G}}(w_j)+H_G(w_j,w_r)\right)\\ &\qquad -H_G(w_{j+1},w_r)-(2-H_G(w_j,w_{j+1}))\\ &= 2(\deg_G(w_i)+\deg_G(w_j)-2)-2+H_G(w_i,w_r) -H_G(w_{i-1},w_r) \\ &\qquad +H_G(w_j,w_r) -H_G(w_{j+1},w_r)+H_G(w_j,w_{j+1})\\ &> 2H_G(w_j,w_{j+1})-H_G(w_{i-1},w_i) >H_G(w_{i-1},w_i), \end{align*} where the last inequality uses equation \eqref{eq:adjhtime} twice to justify $H_G(w_{i-1},w_i)<H_G(w_j,w_{j+1})$. Thus the condition of Lemma \ref{lem:slack} is satisfied, so $\Delta T_{\rm{bestmix}} > 0$. \end{proof} Once there is at most one bad left leaf, we increase the spine length, starting with the left side. Surgery $\mathcal{S}_3$ requires at least two left leaves, one of which must be in $U_2$. We also require $R(w_r)>L(w_{r-1})$, meaning that $w_r$ is the unique best mix focus. If this is not the case (that is, $R(w_r) = L(w_{r-1})$), then we will take $w_{r-1}$ to be the best mix focus, reverse the labeling of the spine, and then apply $\mathcal{S}_4$ below. \begin{lemma} \label{lem:S3} Let $G$ be a bifocal caterpillar with $R(w_r) > L(w_{r-1})$ and with distinct vertices $x\in U_2$ and $y \in U_i$ where $2\leq i\leq r-2$. If $\mathcal{S}_3=\tr{2}{1}\wedge\tr{i}{i+1}$, then $\Delta T_{\rm{bestmix}} \geq 0$. \end{lemma} \begin{proof} First, we use Lemma \ref{lem:criteria} to show that the foci do not change. This surgery extends the left-hand side of the spine. The left-pessimal hitting times to $w_{r-1}$ and $w_r$ increase by $\Delta L(w_r) = \Delta L(w_{r-1}) = 3-2=1$ (using equation \eqref{eq:adjhtime} for the effect of $\tr{2}{1}$ and equation \eqref{eq:leftleft2} for $\tr{i}{i+1}$). The right-pessimal hitting times do not change, $\Delta R(w_r) = \Delta(w_{r-1})=0$. Therefore, equation \eqref{eq:lchange} is satisfied. By assumption, $R( w_r) - L(w_{r-1}) \geq 1 = \DeltaL(w_{r-1})-\DeltaR(w_r)$ so inequality \eqref{eq:newtbmfoc} holds. We have, $\Delta H(v,w_r)\leq\DeltaL(w_r)$ for all $v\in V$ holds for all $v \in V$ since $H(x,w_r)$ is the only hitting time to $w_r$ that increases; all others decrease or are constant. By Corollary \ref{cor:slack}, we have $\Delta T_{\rm{bestmix}} \geq 0$. \end{proof} Surgery $\mathcal{S}_4$ is the right-hand version of $\mathcal{S}_3$. However, if $ L(w_r)=R(w_r)+1 $ then applying $\mathcal{S}_4$ would lead to $ L_{\widetilde{G}}(w_r)=R_{\widetilde{G}}(w_r) $ which indicates that $\widetilde{G}$ has one focus by Lemma \ref{lemma:focuscat}(a). We choose to avoid this situation, So we require that $L(w_r) > R(w_r)+1$ and handle the case $L(w_r)=R(w_r)+1$ with $S_5$ below. \begin{lemma} \label{lem:S4} Let $G$ be a bifocal caterpillar with $ L(w_r)>R(w_r)+1 $ that contains distinct vertices $y \in U_{t-1}$ and $x \in U_i$ where $r < i < t$. If $\mathcal{S}_4=\tr{t-1}{t-2}\wedge\tr{i}{i-1}$, then $\Delta T_{\rm{bestmix}} > 0$. \end{lemma} \begin{proof} This surgery extends the spine on the right-hand side. Analgous to the previous proof, this time the left-pessimal hitting times have $\Delta L(w_{r-1})= 0 = \Delta L(w_r)$ and the right-pessimal hitting times have $\Delta R(w_{r-1}) = 1 = \Delta R(w_r)$. Inequality \eqref{eq:lchange} holds because $L(w_r) - R(w_r) > 1 = \DeltaR(w_r)-\DeltaL(w_r),$ and inequality \eqref{eq:newtbmfoc} holds because $\DeltaL(w_{r-1})-\DeltaR(w_r)<0$. By Lemma \ref{lem:criteria}, $\widetilde{G}$ has focus $w_{r-1}$ and best mix focus $w_r$. Next, we show that $\Delta T_{\rm bestmix} > 0$ using Lemma \ref{lem:slack} with $A=\{ y \}$. We have $\Delta L(w_r) < \Delta H(y, w_r) = \Delta R(w_r)$, while $\Delta H(v,w_r) \leq 0 = \Delta L(w_r)$ for $v \neq y$. The left hand side of inequality \eqref{eqn:slack} equals $\Delta H(y,w_r) = 1$. We take $ B= \{ w_{i-1}, w_i, w_{t-1}, w_t \}$ because the contribution from each vertex in $\overline{A} \backslash B$ is positive. Using equation \eqref{eq:rightright2}, this sum is \begin{align*} \MoveEqLeft -H_G(w_{i-1},w_r) + \left( 2\deg_{\widetilde{G}}(w_i)+H_G(w_i,w_{r}) \right) + \left(2\deg_{\widetilde{G}}(w_{t-1})+H_G(w_{t-1},w_r) \right) \\ &\qquad + \left(-2\deg_{\widetilde{G}}(w_t) -H_G(w_t,w_r) \right)\\ &= H_G(w_i,w_{i-1}) -H_G(w_t, w_{t-1}) + 2 \left( \deg_{\widetilde{G}}(w_i) + \deg_{\widetilde{G}}(w_{t-1}) -\deg_{\widetilde{G}}(w_t) \right) \\ & > H_G(w_i,w_{i-1}) > 1. \end{align*} By Lemma \ref{lem:slack}, $\Delta T_{\rm bestmix} > 0$. \end{proof} We now consider surgery $\mathcal{S}_5$, which is only used when $L(w_r)=R(w_r)+1$. This is one of the crucial moments in our algorithm: a minor change threatens the balance described in Lemma \ref{lemma:focuscat}. In fact, surgery $\mathcal{S}_5=\tr{t-1}{t}$ is the first surgery that shifts the location of the foci. The resulting graph is bifocal with best mix focus $w_r$, but the second focus moves from $w_{r-1}$ to $w_{r+1}$. \begin{lemma} \label{lem:S5} Let $G$ be a bifocal caterpillar with $y\in U_{t-1}$ such that $ L(w_r)= R(w_r)+1. $ If $\mathcal{S}_5=\tr{t-1}{t}$, then $\Delta T_{\rm{bestmix}} > 0$. \end{lemma} We note that during Phase One, we will also have a second right leaf (otherwise we are done with the right spine). However, our proof does not require the existence of such a leaf. \begin{proof} Let $\widetilde{G} = \mathcal{S}_5(G)$ and let $y \in V_{t-1}$ be the transplanted vertex. The changes in the left and right hitting times to the foci are $\DeltaL(w_{r-1})=0= \DeltaL(w_{r})$ and $\DeltaR(w_{r-1})= 3 = \DeltaR(w_r).$ Crucially, inequality \eqref{eq:lchange} is not satisfied. We now verify that $w_r$ is the best mix focus of $\widetilde{G}$ and that $w_{r+1}$ is the other focus. We use Lemma \ref{lemma:focuscat} (d), replacing $r$ with $r+1$ and swapping left for right. In other words we must show that $R(w_r) > L(w_r)$ and $R(w_{r+1}) \leq L(w_r)$. Observe that $ R_{\widetilde{G}}(w_r) =R_G(w_r)+3 >L_G(w_r) =L_{\widetilde{G}}(w_r), $ and $$ R_{\widetilde{G}}(w_{r+1}) =R_G(w_{r+1})+3 < R_G(w_{r+1}) + H_{\widetilde{G}} ( w_{r+1}, w_r) = R_G(w_r) < L_G(w_r) =L_{\widetilde{G}}(w_r). $$ Next, we show that $T_{\rm{bestmix}}(\widetilde{G})>T_{\rm{bestmix}}(G)$ using Lemma \ref{lem:slack}. Note that the surgery shifts the foci, so $w_1$ is the $G$-pessimal vertex for $w_r$, while $y$ is the $G'$-pessimal vertex for $w_r$. We find that $\Delta\mathrm{Pess}(w_r)=2$ because $$ \mathrm{Pess}_{\widetilde{G}}(w_r) = H_{\widetilde{G}}(y,w_r) = 3+H_G(w_t,w_r) =2+H_G(w_1,w_r) = 2 + \mathrm{Pess}_G(w_r). $$ We will use Lemma \ref{lem:slack} with $A=\{y\}$. The left side of equation \eqref{eqn:slack} is $ \deg_{\widetilde{G}}(y)(\Delta H(y,w_r)-2)=3-2=1. $ We take $B = \{y, w_{t-1},w_t\}$ since $v \notin B$ means that $\Delta\deg(v)=0$ and $ \Delta H(v,w_r)\leq 2=\Delta\mathrm{Pess}(w_r).$ The right hand side of equation \eqref{eqn:slack} is $$ \left( 2\deg_{\widetilde{G}}(w_{t-1})+H_G(w_{t-1},w_r) \right) -H_G(w_t,w_r) = 2\deg_{\widetilde{G}}(w_{t-1})-H_G(w_t,w_{t-1}) > 1, $$ so by Lemma \ref{lem:slack}, $\Delta T_{\rm{bestmix}} > 0$. \end{proof} \subsection{Phase Two: Seesaw to Twin Broom} In this section, we discuss Phase Two, which inputs a seesaw and outputs a twin broom. \begin{table}[ht] \begin{center} \begin{tabular}{|c|p{2.75in}|@{}c@{}|} \hline Surgery & Initial Conditions & Illustration\\ \hline $\mathcal{S}_6$ & \small $G$ has exactly one left leaf and $2r-1 \leq t$ & \begin{tabular}{c} \begin{tikzpicture}[scale=0.5, decoration={ markings, mark=at position 1 with {\arrow[scale=1.25,black]{latex}}; } ] \draw[color=white] (2, .5) circle (1pt); \draw[gray!70] (1,0) -- (2,0); \draw[gray!70, fill=gray!70] (1,0) circle (4pt); \draw(4,0) -- (4,-1); \draw[fill] (4,-1) circle (4pt); \draw[dashed] (2,0) -- (9,0); \draw (2,0) -- (2.5,0); \draw (3.5,0) -- (5.5,0); \draw (5.5,0) -- (7.5,0); \draw (8.5,0) -- (9,0); \foreach \i in {2,4,9} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=white] (6,0) circle (4pt); \draw[fill=white] (7,0) circle (6pt); \draw[fill=white] (7,0) circle (4pt); \draw[bend left=20, postaction=decorate] (3.75, -1.1) to (1.2,-.25); \node at (4.35,-1.25) {\scriptsize $x$}; \end{tikzpicture} \end{tabular} \\ \hline $\mathcal{S}_7$ & \small $2r-2 \geq t$; $G$ has exactly one left leaf and one right leaf & \begin{tabular}{c} \begin{tikzpicture}[scale=0.5, decoration={ markings, mark=at position 1 with {\arrow[scale=1.25,black]{latex}}; } ] \draw[color=white] (2, .5) circle (1pt); \draw[gray!70] (11,0) -- (12,0); \draw[gray!70, fill=gray!70] (12,0) circle (4pt); \draw(9,0) -- (9,-1); \draw[fill] (9,-1) circle (4pt); \draw[gray!70] (1,0) -- (2,0); \draw[gray!70, fill=gray!70] (1,0) circle (4pt); \draw(4,0) -- (4,-1); \draw[fill] (4,-1) circle (4pt); \draw[dashed] (2,0) -- (11,0); \draw (2,0) -- (2.5,0); \draw (3.5,0) -- (4.5,0); \draw (5.5,0) -- (7.5,0); \draw (8.5,0) -- (9.5,0); \draw (10.5,0) -- (11,0); \foreach \i in {2,4,9,11} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=white] (6,0) circle (4pt); \draw[fill=white] (7,0) circle (6pt); \draw[fill=white] (7,0) circle (4pt); \draw[bend left=20, postaction=decorate] (3.75, -1.1) to (1.2,-.25); \draw[bend right=20, postaction=decorate] (9.25, -1.1) to (11.8,-.25); \node at (4.35,-1.25) {\scriptsize $x$}; \node at (8.65,-1.25) {\scriptsize $y$}; \end{tikzpicture} \end{tabular} \\ \hline $\mathcal{S}_8$ & \small $G$ has exactly one right leaf and no left leaves & \begin{tabular}{c} \begin{tikzpicture}[scale=0.5, decoration={ markings, mark=at position 1 with {\arrow[scale=1.25,black]{latex}}; } ] \draw[color=white] (2, .5) circle (1pt); \draw[gray!70] (3,0) -- (3,-1); \draw[gray!70, fill=gray!70] (3,-1) circle (4pt); \draw[dashed] (0,0) -- (7,0); \draw (0,0) -- (.5,0); \draw (1.5,0) -- (3.5,0); \draw (4.5,0) -- (5.5,0); \draw (6.5,0) -- (7,0); \foreach \i in {0,5,7} { \draw[fill] (\i,0) circle (4pt); } \draw (5,0) -- (5,-1); \draw[fill] (5,-1) circle (4pt); \node at (5.35,-1.25) {\scriptsize $x$}; \draw[bend left=20, postaction=decorate] (4.75, -1.25) to (3.25,-1.25); \draw[fill=white] (2,0) circle (4pt); \draw[fill=white] (3,0) circle (6pt); \draw[fill=white] (3,0) circle (4pt); \end{tikzpicture} \end{tabular} \\ \hline \end{tabular} \hspace{-1in}\caption{The Phase Two tree surgeries. The tree $G$ is a bifocal seesaw.} \label{table:phase2} \end{center} \end{table} \begin{figure}[t] \begin{center} \includegraphics[width=4.5in]{flowchart_phase_II} \end{center} \caption{Phase Two of the algorithm, turning a seesaw into a twin broom.} \label{fig:phase2} \end{figure} \begin{lemma} \label{lemma:seesaw-to-twinbroom} Let $G$ be a seesaw on $n$ vertices. Phase Two creates a twin broom $\widetilde{G}$ such that $T_{\rm{bestmix}}(\widetilde{G})\geqT_{\rm{bestmix}}(G)$. \end{lemma} \begin{proof} If $G$ is already a twin broom then $\widetilde{G}=G$. The three surgery types $\mathcal{S}_6, \mathcal{S}_7, \mathcal{S}_8$ employed in Phase Two are shown in Table \ref{table:phase2} and the workflow is shown in Figure \ref{fig:phase2}. The lemmas that follow show that $\Delta T_{\rm bestmix} \geq 0$ for all three surgeries. Suppose that $G$ has a left leaf. If $2r-1 \leq t$ then we use $\mathcal{S}_6$ to transplant the left leaf onto the end of the left spine. If $2r-2 \geq t$, or equivalently $r-2 \geq t-r$, then there must also be a right leaf because $w_r$ is the best mix focus (see Lemma \ref{lemma:focuscat}(c)). In this case, we use $S_7$ to simultaneously transplant both of these leaves to the spine. Next, if $G$ still has a right leaf (perhaps we just applied $\mathcal{S}_6$), then we apply $\mathcal{S}_8$ to move this leaf to the end of the right spine. The resulting graph is a twin broom with best mix focus $w_r$ and second focus $w_{r-1}, w_r$. \end{proof} First, we prove that $\mathcal{S}_6$ moves the final left leaf to the spine while also increasing $T_{\rm bestmix}$. \begin{lemma} \label{lem:S6} Let $G$ be a seesaw with $2r-1 \leq t$ and exactly one left leaf $x\in U_i$, where $2 \leq i \leq r-2$. If $\mathcal{S}_6=\tr{i}{1}$ then $\Delta T_{\rm{bestmix}} > 0$. \end{lemma} \begin{proof} The surgery extends the left-hand spine, so $L(w_{r-1}) = H_{\widetilde{G}} (x,w_{r-1}) = H_{\widetilde{G}} (x,w_1) + H_{\widetilde{G}} (w_1, w_{r-1}) = 1 + H_G(w_1 + w_{r-1}) + 2(i-1)$ by equation \eqref{eq:leftleft2}. Therefore $\DeltaL(w_{r-1})=2i-1$, and likewise $\DeltaL(w_r)=2i-1,$ while $\DeltaR(w_{r-1})=0=\DeltaR(w_r).$ We show that the foci have not changed using Lemma \ref{lemma:focuscat} (d). We have $L_{\widetilde{G}}(w_r) > L_G(w_r) \geq R_G(w_r) = R_{\widetilde{G}}(w_r)$. Equation \eqref{eq:htimepath2} and the assumption $2r-1 \leq t$ yield $ L_{\widetilde{G}}(w_{r-1}) = (r-1)^2 \leq (t-r)^2 \leq R_{\widetilde{G}}(w_r). $ Therefore, $\widetilde{G}$ has focus $w_{r-1}$ and best mix focus $w_r$. Next, we use Lemma \ref{lem:slack} with $A=\{x\}$. The left-hand side of inequality \eqref{eqn:slack} is $$ \Delta H(x,w_r)-\DeltaL(w_r) =(i^2-1)-(2i-1)=i^2-2i. $$ Let $B=\{w_{r-1},w_r,\dots,w_{t-1}\}$. For all $v \in B$, we have $\deg_{\widetilde{G}}(v) \geq 2$ and $\Delta H(v,w_r)=0=\Delta\deg(v)$. The right-hand side of inequality \eqref{eqn:slack} is $$ \sum_{v\in B}{2(2i-1)} =2(2i-1)(t-r) \geq 2(2i-1)(r-2) > i(i-2 )=i^2-2i. $$ Lemma \ref{lem:slack} gives $\Delta T_{\rm{bestmix}} > 0$. \end{proof} We employ surgergy $S_7$ is the particular case where $2r-2 \geq t$ and $G$ has both a left leaf and a right leaf. This surgery removes both leaves simultaneously. \begin{lemma} Let $G$ be a seesaw with exactly one left leaf $x\in U_i$ and exactly one right leaf $y\in U_j$, such that $2r-2 \geq t$. If $\mathcal{S}_7=\tr{i}{1}\wedge\tr{j}{t}$, then $\Delta T_{\rm{bestmix}} > 0$. \end{lemma} \begin{proof} First, we show that in fact, $2r-2 = t$, meaning that the left and right spines are the same length. If $r-2 > t-r$ then $L(w_{r-1}) = H(w_1, w_{r-1}) > (r-2)^2 \geq (t-r+1)^2 > H(w_t, w_r) = R(w_r)$, which contradicts Lemma \ref{lemma:focuscat}(c) because $w_r$ is a best mix focus. Next, we observe that $i-1 \geq t-j$, meaning that $i$ is further from the left endpoint than $j$ is from the right endpoint. Indeed, when the left and right spines are the same length, this is necessary for $L(w_{r-1}) \leq R(w_r)$. By equations \eqref{eq:leftleft2} and \eqref{eq:rightright2}, we have $\DeltaL(w_{r-1})=2i-1 = \DeltaL(w_r)$ and $\DeltaR(w_{r-1})=2(t-j)+1=\DeltaR(w_r)$. After the surgery, the left and right spines are equal length and leaf-free, so the conditions of Lemma \ref{lemma:focuscat}(d) hold for $\widetilde{G}$. In fact, both $w_{r-1}, w_r$ are now best mix foci because $L_{\widetilde{G}}(w_{r-1}) = R_{\widetilde{G}}(w_r)$. To prove that $T_{\rm{bestmix}}$ increases, we use Lemma \ref{lem:slack}. Taking $A=\{x,y\}$, the left hand side of inequality \eqref{eqn:slack} is $$ \big( i^2-1-(2i-1) \big) + \big((t+1-j)^2-1-(2i-1) \big) \leq 2 \big( i^2-1-(2i-1) \big) = 2i^2 - 4i. $$ For the right hand side of inequality \eqref{eqn:slack}, we take $B=W=\{ w_1, w_2, \ldots , w_t \}$, in other words, we ignore any central leaves. This right hand side is at least \begin{align*} \MoveEqLeft 2t \, \Delta\mathrm{Pess}(w_r) -2 \sum_{k=1}^t \Delta H(w_i, w_r) \\ & \qquad -H_G(w_1, w_r) + H_G(w_i, w_r) + H_G(w_j, w_r) - H_G(w_t,w_r) \\ % & = 2t (2i-1) -4 \sum_{k=1}^{i-1} (i-k) -4 \sum_{k=j+1}^{t} (k-j) - (i-1)^2- (t-j)^2 \\ % & \geq 2t (2i-1) - 4 i(i-1) - 2 ( i-1)^2 \, =\, 4(r-1)(2i-1) - 6i^2 + 8i -2 \\ & > 4(i-1)(2i-1) - 6i^2 + 8i -2 \, =\, 2i^2 - 4i +2. \end{align*} We have satisfied the conditions of Lemma \ref{lem:slack}, so $\Delta T_{\rm{bestmix}} > 0$. \end{proof} Finally, we consider surgery $\mathcal{S}_8$, which moves the final right leaf to the spine. \begin{lemma} Let $G$ be a seesaw with exactly one right leaf $x \in U_i$, where $r+1 \leq i \leq t-1$ and no left leaves. If $\mathcal{S}_8=\tr{i}{r}$, then $\Delta T_{\rm{bestmix}} > 0$. \end{lemma} \begin{proof} First, we use Lemma \ref{lem:criteria}, to show that $\widetilde{G}$ has focus $w_{r-1}$ and best mix focus $w_r$. By equation \eqref{eq:rightleft2}, $\DeltaL(w_{r-1})=0 =\DeltaL(w_r)=\Delta H(w_1,w_r)$ and $\DeltaR(w_{r-1})=-2(i-r)=\DeltaR(w_r)$. Inequality \eqref{eq:lchange} is clearly satisfied. Next, we verify inequality \eqref{eq:newtbmfoc}. By equation \eqref{eq:htimepath2}, we have $ R_G(w_r)-L_G(w_{r-1})=(t-r)^2+2(i-r)-(r-2)^2 $ and we claim that $(t-r)^2 - (r-2)^2 >0$. Indeed, since $w_r$ is a best mix focus, Lemma \ref{lem:criteria} (c) ensures that $L(w_{r-1}) \leq R(w_r)$, or in other words $(r-2)^2 \leq (t-r)^2 + 2(r-i) < (t-r+1)^2.$ Since $r-2$ and $t-r$ are both integers, we must have $r-2 \leq t-r$. This means that $R_G(w_r)-L_G(w_{r-1}) \geq 2(i-r) = \DeltaL(w_{r-1})-\DeltaR(w_r).$ This confirms that the foci do not change. Finally, for all $v\in V$, we have $\Delta H(v,w_r)\leq 0=L(w_r)$. By Corollary \ref{cor:slack}, $\Delta T_{\rm{bestmix}} > 0$. \end{proof} \subsection{Phase Three: Twin Broom to Path or Wishbone} Phase Three converts a twin broom into either $P_n$ or $Y_n$. The three surgeries are shown in Table \ref{table:phase3} and the workflow is shown in Figure \ref{fig:phase3}. \begin{lemma} \label{lemma:twinbroom-to-end} Let $G$ be a twin broom on $n$ vertices. If $n$ is even then Phase Three turns $G$ into the path $P_n$. If $n$ is odd, then Phase Three turns $G$ into the wishbone $Y_n$. Moreover, if $n$ is even then $T_{\rm bestmix}(G) \leq T_{\rm bestmix}(P_n)$ and if $n$ is odd then $T_{\rm bestmix}(G) \leq T_{\rm bestmix}(Y_n)$. Furthmore, equality holds if and only if $G=P_n$ for $n$ even, and $G=Y_n$ for $n$ odd. \end{lemma} \begin{proof} The Phase Three surgeries $\mathcal{S}_9, \mathcal{S}_{10}, \mathcal{S}_{11}$ are shown in Table \ref{table:phase3} and the workflow is shown in Figure \ref{fig:phase3}. The lemmas that follow show that $\Delta T_{\rm bestmix} > 0$ for all three surgeries. Suppose that $r-2 < t-r$, or equivalently $t < 2r-2$. Since $G$ is bifocal, there must be enough leaves in $U_{r-1}$ so that $(r-1)^2 + 2|U_{r-1}| = L(w_r) > R(w_r) = (t-r)^2$. We apply $\mathcal{S}_9$ until $r-2 = t-r$. At this point (and henceforth), both $w_{r-1}$ and $w_r$ are best mix foci by Lemma \ref{lemma:focuscat} (d). We may assume that $|U_{r-1}| \leq |U_{r}|$. If both $U_{r-1}$ and $U_r$ are nonempty, we use $\mathcal{S}_{10}$ to simultaneously extend the left and right spine by taking one vertex from each of these sets. We repeat this until $U_{r-1} = \emptyset$. Next, if there are multiple leaves remaining in $U_r$, we use $\mathcal{S}_{11}$ to extend both ends of the spine. We repeat $\mathcal{S}_{11}$ until there are 0 or 1 leaves left in $U_r$. At this point, we either have a path $P_n$ or a wishbone $Y_n$. We started this phase with a twin broom, which is bifocal by definition. Therefore, if $n$ is even we have constructed a path and if $n$ is odd we have constructed a wishbone. \end{proof} \begin{table}[ht] \begin{center} \begin{tabular}{|c|p{2.5in}|@{}c@{}|} \hline Surgery & Initial Conditions & Illustration \\ \hline $\mathcal{S}_9$ & \small $r-2<t-r$, which forces $U_{r-1} \neq 0$ & \begin{tabular}{c} \begin{tikzpicture}[scale=0.5, decoration={ markings, mark=at position 1 with {\arrow[scale=1.25,black]{latex}}; } ] \draw[color=white] (2, .5) circle (1pt); \draw[gray!70] (1,0) -- (2,0); \draw[gray!70, fill=gray!70] (1,0) circle (4pt); \draw(4,0) -- (4,-1); \draw[fill] (4,-1) circle (4pt); \draw[dashed] (2,0) -- (7,0); \draw (2,0) -- (2.5,0); \draw (3.5,0) -- (5.5,0); \draw (6.5,0) -- (7,0); \foreach \i in {2,7} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=white] (4,0) circle (4pt); \draw[fill=white] (5,0) circle (6pt); \draw[fill=white] (5,0) circle (4pt); \draw[bend left=20, postaction=decorate] (3.75, -1.1) to (1.2,-.25); \node at (4,-1.5) {\scriptsize $x$}; \end{tikzpicture} \end{tabular} \\ \hline $\mathcal{S}_{10}$ & \small $r-2=t-r$, and there are leaves $x\in U_{r-1}$ and $y\in U_r$ & \begin{tabular}{c} \begin{tikzpicture}[scale=0.5, decoration={ markings, mark=at position 1 with {\arrow[scale=1.25,black]{latex}}; } ] \draw[color=white] (2, .5) circle (1pt); \draw[gray!70] (7,0) -- (8,0); \draw[gray!70, fill=gray!70] (8,0) circle (4pt); \draw(5,0) -- (5,-1); \draw[fill] (5,-1) circle (4pt); \draw[gray!70] (1,0) -- (2,0); \draw[gray!70, fill=gray!70] (1,0) circle (4pt); \draw(4,0) -- (4,-1); \draw[fill] (4,-1) circle (4pt); \draw[dashed] (2,0) -- (7,0); \draw (2,0) -- (2.5,0); \draw (3.5,0) -- (5.5,0); \draw (6.5,0) -- (7,0); \foreach \i in {2,7} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=white] (4,0) circle (4pt); \draw[fill=white] (5,0) circle (6pt); \draw[fill=white] (5,0) circle (4pt); \draw[bend left=20, postaction=decorate] (3.75, -1.1) to (1.2,-.25); \draw[bend right=20, postaction=decorate] (5.25, -1.1) to (7.8,-.25); \node at (4,-1.5) {\scriptsize $x$}; \node at (5,-1.5) {\scriptsize $y$}; \end{tikzpicture} \end{tabular} \\ \hline $\mathcal{S}_{11}$ & \small $r-2=t-r$, $U_{r-1} = \emptyset$ and there are leaves $x,y\in U_r$ & \begin{tabular}{c} \begin{tikzpicture}[scale=0.5, decoration={ markings, mark=at position 1 with {\arrow[scale=1.25,black]{latex}}; } ] \draw[color=white] (2, .5) circle (1pt); \draw[gray!70] (7,0) -- (8,0); \draw[gray!70, fill=gray!70] (8,0) circle (4pt); \draw(5,0) -- (5.33,-1); \draw[fill] (5.33,-1) circle (4pt); \draw[gray!70] (1,0) -- (2,0); \draw[gray!70, fill=gray!70] (1,0) circle (4pt); \draw(5,0) -- (4.67,-1); \draw[fill] (4.67,-1) circle (4pt); \draw[dashed] (2,0) -- (7,0); \draw (2,0) -- (2.5,0); \draw (3.5,0) -- (5.5,0); \draw (6.5,0) -- (7,0); \foreach \i in {2,7} { \draw[fill] (\i,0) circle (4pt); } \draw[fill=white] (4,0) circle (4pt); \draw[fill=white] (5,0) circle (6pt); \draw[fill=white] (5,0) circle (4pt); \draw[bend left=20, postaction=decorate] (4.40, -1.1) to (1.2,-.25); \draw[bend right=20, postaction=decorate] (5.6, -1.1) to (7.8,-.25); \node at (4.6,-1.5) {\scriptsize $x$}; \node at (5.4,-1.5) {\scriptsize $y$}; \end{tikzpicture} \end{tabular} \\ \hline \end{tabular} \end{center} \caption{The Phase Three tree surgeries. The graph $G$ is a twin broom.} \label{table:phase3} \end{table} \begin{figure}[t] \begin{center} \includegraphics[width=5.5in]{flowchart_phase_III} \end{center} \caption{Phase three of the algorithm, turning a twin broom into a path or a wishbone.} \label{fig:phase3} \end{figure} We first consider $\mathcal{S}_9$, which we apply to equalize the lengths of the left and right spine. \begin{lemma} \label{lem:S9} Let $G$ be a twin broom with leaf $x\in U_{r-1}$ such that $2r-1 \leq t.$ If $\mathcal{S}_9=\tr{{r-1}}{1}$, then $\Delta T_{\rm bestmix} > 0$. \end{lemma} \begin{proof} First, it is routine to show that the simultaneous inequalities $2r-1 \leq t$ and $L_G(w_r) > R_G(w_r)$ require that $\deg(w_{r-1}) \geq 4$. By equation \eqref{eq:leftleft2}, we have $\DeltaL(w_{r-1})=2r-3=\DeltaL(w_r)$ and $\DeltaR(w_{r-1})=0=\DeltaR(w_r).$ We use the two criteria of Lemma \ref{lemma:focuscat} (d) to show that the foci roles do not change. Clearly $L_{\widetilde{G}}(w_r) > L_G(w_r) \geq R_G (w_r) = R_{\widetilde{G}}(w_r)$. Since $U_i=\emptyset$ for all $i\notin\{r-1,r\}$, we can use equation \eqref{eq:htimepath2} to calculate $L_{\widetilde{G}}(w_{r-1})=(r-1)^2\leq (t-r)^2=R_{\widetilde{G}}(w_r).$ Next, we show that the best mixing time increases using Lemma \ref{lem:slack} with $A= \{ x \}$. The left-hand side of inequality \eqref{eqn:slack} is $$ \Delta H(x,w_r) - \Delta\mathrm{Pess} (w_r) =( (r-1)^2-1) - (2r-3) = r^2 -4r+3. $$ For the right hand side, we take $B=W$, ignoring any other central leaves. We obtain \begin{align*} \MoveEqLeft (2t-1) \, \Delta\mathrm{Pess} (w_r) -2 \sum_{i=1}^{r-1} \Delta H(w_i,w_r) - H_G(w_1, w_r) + H_G(w_{r-1},w_r) \\ & = (2t-1) (2r-3) - 4 \sum_{i=1}^{r-1} (r-1-k) - (r-2)^2 \\ & \geq 2(2r-2) (2r-3) -3 r^2 +10 r -8 \, = \, 5r^2 -10r +4, \end{align*} which is clearly larger than the left hand side, so $\Delta T_{\rm bestmix} > 0$. \end{proof} Next, we discuss transplanting one leaf from each of $U_{r-1}$ and $U_{r}$. \begin{lemma} \label{lem:S10} Let $G$ be a twin broom where $t=2r-2$ and with vertices $x\in U_{r-1}$ and $y\in U_r$. If $\mathcal{S}_{10}=\tr{{r-1}}{1}\wedge\tr{r}{t}$, then $\Delta T_{\rm bestmix} > 0$. \end{lemma} \begin{proof} The right-hand and left-hand spines have equal lengths before and after this surgery. Corollary \ref{cor:criteria} holds because each of $\DeltaL(w_{r-1})$, $\DeltaL(w_{r})$, $\DeltaR(w_{r-1})$, and $\DeltaR(w_{r})$ are equal to $2(r-2)-1= 2r-3$. So $\widetilde{G}$ has focus $w_{r-1}$ and best mix focus $w_r$. We now show that the best mixing time increases. Setting $A=\{x,y\}$, the left hand side of inequality \eqref{eqn:slack} is $$\Delta H(x,w_r)+\Delta H(y,w_r)-2(2r-3)= 2 \big((r-1)^2-1 \big) - 2(2r-3) = 2r^2 -8r + 6. $$ where we use equations \eqref{eq:leftleft2} and \eqref{eq:rightright2}. As for the right-hand side, we take $B=W$, disregarding the remaining central leaves. This right hand side is at least \begin{align*} \MoveEqLeft 2t \, \Delta\mathrm{Pess}(w_r) - 2 \sum_{k=1}^t \Delta H(w_k,w_r) -H_G(w_1, w_r) +H(w_{r-1}, w_r) - H_G(w_t, w_r) \\ & = 2(2r-2) (2r-3) -2 \cdot 2 \sum_{k=1}^{r-1}2 (r-1-k) - 2 (r-2)^2 \, = \, 2r^2 -4. \end{align*} By Lemma \ref{lem:slack}, $\Delta T_{\rm bestmix} > 0$. \end{proof} Finally, we discuss transplanting pairs of leaves from $U_r$ to the ends of the spine. \begin{lemma} \label{lem:S10} Let $G$ be a twin broom with $t = 2r-2$ where $U_{r-1} = \emptyset$ and with distinct vertices $x,y \in U_r$. If $\mathcal{S}_{11}=\tr{r}{1}\wedge\tr{r}{t}$, then $\Delta T_{\rm bestmix} > 0$. \end{lemma} \begin{proof} This proof is similar to the previous one: the left and right spine lengths are equal before and after the surgery. We have $\DeltaL(w_{r})=2r-1$ while $\DeltaL(w_{r-1})=\DeltaR(w_{r-1})=\Delta R(w_{r})=2r-3,$ where the value for $\Delta R(w_{r-1})$ takes into account the removal of two leaves from $U_r$. We have $ \DeltaL(w_{r-1})-\DeltaR(w_r)=0$ and $\DeltaR(w_r)-\DeltaL(w_r)=-2,$ so Corollary \ref{cor:criteria} ensures that $\widetilde{G}$ has focus $w_{r-1}$ and best mix focus $w_r$. Setting $A= \{ x, y \}$, the left hand side of inequality \eqref{eqn:slack} is $$ (r^2 -1) + ((r-1)^2 -1) - 2(2r-1) = 2r^2 -2r + 1. $$ As for the right-hand side of inequality \eqref{eqn:slack}, we take $B=W$ and obtain \begin{align*} \MoveEqLeft 2t \, \Delta\mathrm{Pess}(w_r) - 2\sum_{k=1}^t \Delta H(w_k, w_r) -H_G(w_1, w_r) - H_G(w_t, w_r) \\ &= 2(2r-2) (2r-1) - 2\sum_{k=1}^{r-1} 2(r-k) - 2 \sum_{k=r+1}^{2r-2} 2(k-r) - (r-1)^2 - (r-2)^2 \\ &=2r^2+2r-5. \end{align*} Thus Lemma \ref{lem:slack} ensures that $\Delta T_{\rm bestmix} > 0$. \end{proof} \section{Conclusion} We have characterized the tree structures on $n$ vertices that minimize and maximize $T_{\rm bestmix} = \min_{v \in V} H(v,\pi)$. The star $S_n$ is the unique minimizing structure, but the maximization problem depends on the parity of $n$. For even $n$, the maximizing structure is the path $P_n$, and for odd $n$, it is the wishbone $Y_n$. It is a bit strange that the odd path is not the maximizing structure for $T_{\rm bestmix}$. But all is not lost: we believe that $P_n$ is the maximizing structure for a slightly different quantity. For any graph $G$, the \emph{forget distribution} $\mu$ is the distribution achieving $\max_{v \in V} H(v,\mu) = \min_{\tau} \max_{v \in V} H(v, \tau)$. Lov\'asz and Winkler \cite{lovasz+winkler-forget} shows that $\mu$ is unique, and they give a general formula. For a tree $G$, the forget distribution is concentrated on its foci \cite{beveridge}. When $G$ is focal, $\mu$ is a singleton distribution on the unique focus. When $G$ is bifocal, $\mu$ is given by $$ \mu_u = \frac{H(v',v) - H(u',v)}{2|E|} \quad \mbox{and} \quad \mu_v = \frac{H(u',u) - H(v',u)}{2|E|} $$ where $u,v$ are the foci of the tree. Instead of $T_{\rm bestmix}=\min_w H(w,\pi)$, we could instead consider the similar quantity $H(\mu, \pi)$. It is easy to see that $S_n$ minimizes $H(\mu, \pi)$ among all trees on $n$ vertices. We conjecture that $P_n$ maximizes this quantity for both even and odd $n$. Indeed, letting $G=P_n$ and $\widetilde{G}=Y_n$, calculations show that for odd $n$, we have $H_{P_n}(\mu,\pi) = H_{Y_n}(\widetilde{\mu}, \widetilde{\pi}) + (2n-3)/(2n-2)$, and for even $n$, we have $H_{P_n}(\mu,\pi) = H_{Y_n}(\widetilde{\mu}, \widetilde{\pi}) + 1/(2n-2)$. Our tree surgery methods should be a fruitful line of attack, though tracking the changes in $H(\mu, \pi)$ will require a new set of lemmas. We leave this problem for future work. \bibliographystyle{plain}

https://arxiv.org/abs/1407.4377

Recovery-Based Error Estimators for Diffusion Problems: Explicit Formulas

We introduced and analyzed robust recovery-based a posteriori error estimators for various lower order finite element approximations to interface problems in [9, 10], where the recoveries of the flux and/or gradient are implicit (i.e., requiring solutions of global problems with mass matrices). In this paper, we develop fully explicit recovery-based error estimators for lower order conforming, mixed, and non- conforming finite element approximations to diffusion problems with full coefficient tensor. When the diffusion coefficient is piecewise constant scalar and its distribution is local quasi-monotone, it is shown theoretically that the estimators developed in this paper are robust with respect to the size of jumps. Numerical experiments are also performed to support the theoretical results.

\section{Introduction}\label{intro} \setcounter{equation}{0} A posteriori error estimation for finite element methods has been extensively studied for the past three decades (see, e.g., books by Verf\"urth \cite{Ver:96, Ver:13}, Ainsworth and Oden \cite{AiOd:00}, Babu\v{s}ka and Strouboulis \cite{BaSt:01}, and references therein). The widely adapted estimator is probably the Zienkiewicz-Zhu (ZZ) recovery-based error estimator \cite{ZiZh:87, ZiZh:92} due to its easy implementation, generality, and ability to produce quite accurate estimations. By first recovering a gradient in the conforming $C^0$ linear vector finite element space from the numerical gradient, the ZZ estimator is defined as the $L^2$ norm of the difference between the recovered and the numerical gradients. Despite popularity of the ZZ estimator, it is also well known that the ZZ estimator over-refines regions where there are no error, and hence, they fail to reduce the global error. This is shown by Ovall in \cite{Ova:06b} through some interesting and realistic examples. Such a failure is simply caused by using continuous functions (recovered gradient/flux) to approximate discontinuous functions (true gradient/flux) in the recovery procedure. By recovering flux and/or gradient in the respective $H({\rm div};\O)$ and $H({\rm curl}; \O)$ conforming finite element spaces, in \cite{CaZh:09, CaZh:10a}, we developed and studied robust recovery-based implicit and explicit error estimators for various lowest order finite element approximations to the interface problems. The implicit error estimator requires solution of a global $L^2$ minimization problem, and the explicit error estimator uses a simple edge average. The explicit recovery introduced in \cite{CaZh:09, CaZh:10a} is limited to the Raviart-Thomas ($RT$) \cite{BrFo:91} and the first type of N\'ed\'elec ($N\!E$) \cite{Ned:80} elements of the lowest order for the respective flux and gradient recoveries. This simple averaging approach may not be extended to the Brezzi-Douglas-Marini ($B\!D\!M$) \cite{BrFo:91} and the second type of N\'ed\'elec \cite{Ned:86} ($N\!D$) elements of the lowest order and to the diffusion problem with full coefficient tensor. The purpose of this paper is first to introduce a general approach for constructing explicit recovery of the flux/gradient for various lower order finite element approximations to the diffusion problem with the full coefficient tensor. The approach, similar to \cite{CaZh:11}, is to localize the implicit recovery through a partition of the unity. For various lower order elements, we are able to reduce the local patch problem to the edge/face patch which contains at most two elements. Hence, by solving a local minimization problem on this two-element patch, we explicitly recover the flux/gradient. We then define the corresponding estimators and establish their reliability and efficiency. When the diffusion coefficient is piecewise constant and its distribution is local quasi-monotone, we are able to show theoretically that these estimators are robust with respect to the size of jumps. For a benchmark test problem, whose coefficient is not local quasi-monotone, numerical results also show the robustness of the estimators. For the conforming finite element approximation to the interface problem, robust error estimators have been studied by Bernardi and Verf\"urth \cite{BeVe:00} and Petzoldt \cite{Pet:02} for the residual-based estimator, Luce and Wohlmuth \cite{LuWo:04} for an equilibrated estimator on a dual mesh, and by us \cite{CaZh:09} for the recovery-based error estimator. Ainsworth in \cite{Ain:05, Ain:07} studied robust error estimators for nonconforming and mixed methods, respectively. Robust error estimators for locally conserved methods were studied by Kim \cite{Kim:07}. Recently, we studied robust recovery-based estimators for lowest order nonconforming, mixed, and discontinuous Galerkin methods (see \cite{CaZh:10a, CaYeZh:11}) via the $L^2$ recovery and for higher-order conforming elements in \cite{CaZh:10b} via a weighted $H({\rm div})$ recovery. Robust equilibrated residual error estimator are constructed by us in \cite{CaZh:11}. For interface problems with flux jumps, we studied robust residual- and recovery-based error estimators in \cite{CaZh:10c}. The paper is organized as follows. Section 2 describes the diffusion problem and its variational forms. Conforming, mixed, and nonconforming finite element methods are presented in Section 3. Section 4 introduces the explicit recoveries of the flux/gradient for those finite element approximations. The corresponding a posteriori error estimators are introduced in Section 5 and their reliability and efficiency bounds are established in Section 6. Finally, Section 7 provides numerical results for a benchmark test problem. \section{Diffusion Problem and Variational Form}\label{problems} \setcounter{equation}{0} Let $\O$ be a bounded polygonal domain in $\Re^2$, with boundary $\partial \O = \bar{\Gamma}_{_D} \cup\bar{ \Gamma}_{_N}$, $\Gamma_{_D}\cap \Gamma_{_N} = \emptyset$, and $\mbox{measure}\,(\Gamma_D)\not= 0$, and let ${\bf n}$ be the outward unit vector normal to the boundary. Consider diffusion equation \begin{equation}\label{pde} -\nabla\cdot (A(x) \nabla u) = f \quad\mbox{in} \quad \O\\ \end{equation} with boundary conditions \begin{equation}\label{bc} -A\nabla u \cdot {\bf n} = g_{_N} \quad\mbox{on} \quad \Gamma_{_N} \quad \mbox{and}\quad u = g_{_D} \quad\mbox{on}\quad \Gamma_{_D}. \end{equation} For simplicity of presentation, assume that $f \in L^2(\O)$, that $g_{_D}$ and $g_{_N}$ are piecewise affine functions and constants, respectively, and that $A$ is a symmetric, positive definite piecewise constant matrix. Here and thereafter, we use standard notations and definitions for the Sobolev spaces. Let \[ H^1_{g,D}(\O)=\{v\in H^1(\O)\,|\, v=g_{_D}\mbox{ on }\Gamma_D\} \quad\mbox{and}\quad H^1_D(\O)=H^1_{0,D}(\O) \] Then the corresponding variational problem is to find $u \in H^1_{g,D}(\O)$ such that \begin{equation} \label{vp} a(u,\,v)\equiv (A\nabla u, \nabla v) = (f, v) - ( g_{_N}, v )_{\Gamma_{_N}}\equiv f(v) \quad \forall \; v\in H^1_D(\O), \end{equation} where $(\cdot, \cdot)_{\omega}$ is the $L^2$ inner product on the domain $\o$. The subscript $\omega$ is omitted when $\o=\O$. In two dimensions, for $\mbox{\boldmath$\tau$}= (\tau_1,\,\tau_2)^t$, define the divergence and curl operators by \[ \nab\cdot \mbox{\boldmath$\tau$}:= \displaystyle\frac{\partial\tau_1}{\partial x_1} + \displaystyle\frac{\partial\tau_2}{\partial x_2} \quad\mbox{and}\quad {\nabla \times} \mbox{\boldmath$\tau$} := \displaystyle\frac{\partial\tau_2}{\partial x_1} - \displaystyle\frac{\partial\tau_1}{\partial x_2}, \] respectively. For a scalar-valued function $v$, define the operator $\nabla^{\perp}$ by $$ \nabla^{\perp} v: = (\displaystyle\frac{\partial v}{\partial x_2},\,- \displaystyle\frac{\partial v}{\partial x_1})^t. $$ We shall use the following Hilbert spaces \begin{eqnarray*} && H({\rm div};\O)=\{\mbox{\boldmath$\tau$}\in L^2(\O)^2 |\,\nabla\cdot\mbox{\boldmath$\tau$}\in L^2(\O)\}\\[2mm] \mbox{ and} && H({\rm curl};\O)=\{\mbox{\boldmath$\tau$}\in L^2(\O)^2 |\,{\nabla \times} \mbox{\boldmath$\tau$} \in L^2(\O)\} \end{eqnarray*} equipped with the norms \[ \|\mbox{\boldmath$\tau$}\|_{H({\rm div};\,\O)}=\left(\|\mbox{\boldmath$\tau$}\|^2_{0,\O}+\|\nabla\cdot\mbox{\boldmath$\tau$}\|^2_{0,\O} \right)^\frac12 \quad\mbox{and}\quad \|\mbox{\boldmath$\tau$}\|_{H({\rm curl};\,\O)}=\left(\|\mbox{\boldmath$\tau$}\|^2_{0,\O}+\|{\nabla \times}\mbox{\boldmath$\tau$}\|^2_{0,\O} \right)^\frac12, \] respectively. Let \begin{eqnarray*} H_{g,N}({\rm div};\O)\! &=&\{\mbox{\boldmath$\tau$}\in H({\rm div};\O) |\,\mbox{\boldmath$\tau$}\cdot {\bf n}|_{\Gamma_{_N}}= g_{_N}\},\quad H_N({\rm div};\O)\!=H_{0,N}({\rm div};\O),\\[2mm] \mbox{ and }\,\, H_D({\rm curl};\O)\! &=&\{\mbox{\boldmath$\tau$}\in H({\rm curl};\O) |\,\mbox{\boldmath$\tau$}\cdot {\bf t}\big|_{\Gamma_{_D}}=0\}, \end{eqnarray*} where ${\bf n} = ( n_1,n_2)^t$ and ${\bf t} =(t_1,t_2)^t = (-n_2,n_1)^t$ are the unit vectors outward normal to and tangent to the boundary $\partial\O$, respectively. Define the flux by \[ \mbox{\boldmath$\sigma$} = -A(x)\nabla u \quad\mbox{in }\,\O, \] then the mixed variational formulation is to find $(\mbox{\boldmath$\sigma$},\,u)\in H_{g,N}({\rm div};\O)\times L^2(\O)$ such that \begin{equation}\label{mixed} \left\{\begin{array}{lclll} (A^{-1}\mbox{\boldmath$\sigma$},\,\mbox{\boldmath$\tau$})-(\nab\cdot \mbox{\boldmath$\tau$},\, u) &=& -(\mbox{\boldmath$\tau$} \cdot {\bf n}, g_{_D})_{\Gamma_{_D}} \quad & \forall\,\, \mbox{\boldmath$\tau$} \in H_N({\rm div};\O),\\[2mm] (\nab\cdot \mbox{\boldmath$\sigma$}, \,v) &=& (f,\,v)&\forall \,\, v\in L^2(\O). \end{array}\right. \end{equation} \section{Finite Element Approximation}\label{FEMs} \setcounter{equation}{0} \subsection{Finite Element Spaces} For simplicity, consider only triangular elements. Let ${\cal T}=\{K\}$ be a regular triangulation of the domain $\O$, and denote by $h_{_K}$ the diameter of the element $K$. We assume that $A$ is piecewise constant matrix on the mesh ${\cal T}$. Denote the set of all nodes of the triangulation by $ {\cal N} := {\cal N}_{_I}\cup{\cal N}_{_D}\cup{\cal N}_{_N}, $ where ${\cal N}_{_I}$ is the set of all interior nodes and ${\cal N}_{_D}$ and ${\cal N}_{_N}$ are the sets of all boundary nodes belonging to the respective $\overline{\Gamma}_D$ and $\Gamma_N$. Denote the set of all edges of the triangulation by $ {\cal E} := {\cal E}_{_I}\cup{\cal E}_{_D}\cup{\cal E}_{_N}, $ where ${\cal E}_{_I}$ is the set of all interior element edges and ${\cal E}_{_D}$ and ${\cal E}_{_N}$ are the sets of all boundary edges belonging to the respective $\Gamma_D$ and $\Gamma_N$. For each $F \in {\cal E}$, denote by ${\bf n}_{_F}= (n_{1,{_F}},n_{2,{_F}})^t$ a unit vector normal to $F$; then ${\bf t}_{_F} = -(n_{2,{_F}},n_{1,{_F}})^t$ is a unit vector tangent to $F$. Let $K_{_F}^-$ and $K_{_F}^+$ be two elements sharing the common edge $F$ such that the unit outward normal vector of $K_{_F}^-$ coincide with ${\bf n}_{_F}$. When $F\in {\cal E}_{_D} \cup {\cal E}_{_N}$, ${\bf n}_{_F} $ is the unit outward vector normal to $\partial\O$ and denote by $K_{_F}^-$ the element having the edge $F$. For interior edges $F\in {\cal E}_{_I}$, the selection of ${\bf n}_{_F}$ is arbitrary but globally fixed. For a function $v$ defined on $K_{_F}^-\cup K_{_F}^+$, denote its traces on $F$ by $v|_{_F}^-$ and $v|_{_F}^+$, respectively. The jump over the edge $F$ is denoted by $$ \jump{v}_{_F} := \left\{ \begin{array}{llll} v|_{_F}^- - v|_{_F}^+ & F\in {\cal E}_{_I}, \\ v|_{_F}^- & F\in{\cal E}_{_D} \cup {\cal E}_{_N}. \end{array} \right. $$ (When there is no ambiguity, the subscript or superscript $F$ in the designation of jump and other places will be dropped.) For each $K\in{\cal T}$, let $P_k(K)$ be the space of polynomials of degree $k$. Denote the linear conforming and nonconforming (Crouzeix-Raviart) finite element spaces \cite{Cia:78, GiRa:86} associated with the triangulation ${\cal T}$ by \[\begin{array}{ll} &S = \{v\in H^1(\O)\,\big|\,v|_K\in P_1(K)\quad\forall\,\,K\in{\cal T}\}\\[2mm] and & S^{nc}= \{v\in L^2(\O)\,\big|\,v|_K\in P_1(K)\,\,\forall\,\,K\in{\cal T}, \mbox{ and } v \,\mbox{is continuous at}\,m_{_F} \;\forall\, F \in {\cal E}_{_I}\}, \end{array}\] respectively. Let \[ \begin{array}{ll} S_{g,{_D}}= \{v\in S |\,v=g_{_D} \,\,\mbox{on}\, \, \Gamma_{_D}\}, & \quad S^{nc}_{g,{_D}}=\{v\in S^{nc}\,|\, v(m_{_F}) =g_{_D}(m_{_F}) \,\,\forall\,\, F\in {\cal E}_{_D}\},\\[2mm] S_{{_D}}= \{v\in S |\,v=0 \,\,\mbox{on}\, \, \Gamma_{_D}\}, & \quad S^{nc}_{{_D}}=\{v\in S^{nc}\,|\, v(m_{_F}) =0 \,\,\forall\,\, F\in {\cal E}_{_D}\}. \end{array} \] The $H({\rm div};\,\Omega)$ conforming Raviart-Thomas (RT) and Brezzi-Douglas-Marini (BDM) spaces \cite{BrFo:91} of the lowest order are defined by \begin{eqnarray* && RT=\{\mbox{\boldmath$\tau$}\in H({\rm div};\,\Omega)\big|\, \mbox{\boldmath$\tau$}|_K\in RT(K)\,\,\forall\,K\in{\cal T}\}\\[2mm] \mbox{and} && B\!D\!M=\{\mbox{\boldmath$\tau$}\in H({\rm div};\,\Omega)\big|\, \mbox{\boldmath$\tau$}|_K\in BDM(K)\,\,\,\,\forall\,\,K\in{\cal T}\}, \end{eqnarray*} respectively, where $RT(K)=P_0(K)^2 +(x_1,x_2)^t\,P_0(K)$ and $B\!D\!M(K)=P_1(K)^2$. The $H({\rm curl}\,;\,\Omega)$-conforming first \cite{Ned:80} and second \cite{Ned:86} types of N\'ed\'elec spaces of the lowest order are defined by \begin{eqnarray*} && N\!E\!=\{\mbox{\boldmath$\tau$}\in H({\rm curl};\Omega)\big|\, \mbox{\boldmath$\tau$}|_K\in N\!E(K)\,\forall\,K\in{\cal T}\} \\[2mm] \mbox{and} && N\!D=\{\mbox{\boldmath$\tau$}\in H({\rm curl};\Omega)\big|\, \mbox{\boldmath$\tau$}|_K\in N\!D(K)\,\,\forall\,K\in{\cal T}\}, \end{eqnarray*} respectively, where $N\!E(K)\!=\!P_0(K)^2+(x_2,-x_1)^tP_0(K)$ and $N\!D(K) =P_1(K)^2$. For convenience, denote $RT(K)$ and $B\!D\!M(K)$ by ${\cal V}(K)$, $RT$ and $B\!D\!M$ by ${\cal V}$, $N\!E(K)$ and $N\!D(K)$ by ${\cal W}(K)$, and $N\!E$ and $N\!D$ by ${\cal W}$. Also, let $$ P_0 =\{v\in L^2(\O)\,\big|\, v|_K \in P_0(K) \,\, \forall\,\, K\in {\cal T}\}. $$ Definitions and properties of bases for the $RT$, $BDM$, $N\!E$, and $N\!D$ spaces on an element $K$ are presented in Appendix A. Finally, we define the discrete gradient, divergence, and curl operators by $$ (\nabla_h v)|_K := \nabla(v|_K), \quad (\nabla_h\cdot \mbox{\boldmath$\tau$})|_K := \nab\cdot(\mbox{\boldmath$\tau$}|_K), \quad\mbox{and}\quad (\nabla_h\times \mbox{\boldmath$\tau$})|_K := {\nabla \times}(\mbox{\boldmath$\tau$}|_K) $$ for all $K\in {\cal T}$, respectively. \subsection{Finite Element Approximation} The conforming finite element method is to seek $u_{c} \in S_{g,{_D}}$ such that \begin{eqnarray} \label{problem_c} ( A\nabla u_{c},\, \nabla v) &=& (f,v) \qquad \,\forall\, v\in S_{{_D}}, \end{eqnarray} the mixed finite element method is to seek $(\mbox{\boldmath$\sigma$}_{m},u_m) \in \left({\cal V}\cap H_{g,{_N}}({\rm div};\,\O)\right)\times P_0$ such that \begin{equation}\label{problem_mixed} \left\{\begin{array}{lclll} (A^{-1}\mbox{\boldmath$\sigma$}_m,\,\mbox{\boldmath$\tau$})-(\nab\cdot \mbox{\boldmath$\tau$},\, u_m)&=& -(\mbox{\boldmath$\tau$}\cdot{\bf n}, g_{_D})_{\Gamma_{_D}} \quad & \forall\,\, \mbox{\boldmath$\tau$} \in {\cal V} \cap H_N({\rm div};\O),\\[2mm] (\nab\cdot \mbox{\boldmath$\sigma$}_m,\, v) &=& (f,\,v)&\forall \,\, v\in P_0, \end{array}\right. \end{equation} and the nonconforming finite element method is to find $u_{nc} \in S^{nc}_{g,{_D}}$ such that \begin{eqnarray} \label{problem_nc} ( A\nabla_h u_{nc},\, \nabla_h v) &=& (f,v) \qquad \,\forall\, v\in S^{nc}_{{_D}}. \end{eqnarray} \section{Explicit Flux and Gradient Recoveries} \setcounter{equation}{0} In \cite{CaZh:09, CaZh:10a, CaYeZh:11}, we studied flux and/or gradient recoveries for various lower order finite element approximations to the diffusion problem. A unique feature of those recoveries is that the recovered quantities are in proper finite element spaces. However, those recoveries require solutions of global problems with mass matrices. In this section, we introduce explicit recovery procedures. This is done by first decomposing the error of the flux/gradient through a partition of the unity and then approximating the flux/gradient error by local patch problems. The partition of the unity is based on nodal basis functions of the non-conforming linear element, and hence the local patch problems contain at most two elements. \subsection{Explicit Flux Recovery for Conforming Method} Let $u_c$ be the conforming linear finite element approximation defined in (\ref{problem_c}). Denote by \[ \hat{\mbox{\boldmath$\sigma$}}_c=-A\nabla u_c, \quad e_c = u-u_c, \quad\mbox{and}\quad {\bf E}_c=\mbox{\boldmath$\sigma$}-\hat{\mbox{\boldmath$\sigma$}}_c = -A\nabla e_c, \] the numerical flux, the solution error, and the flux error, respectively. In this section, we introduce an explicit flux recovery procedure. This will be done through approximating the error flux ${\bf E}_c$ by local patch problems. To this end, let $\phi_{{_F}}^{nc}({\bf x}) \in S^{nc}$ be the nodal basis function of the linear nonconforming element associated with the edge $F \in {\cal E}$. Denote by \[\o_{{_F}} = \mbox{supp}(\phi_{{_F}}^{nc}({\bf x})) \] the support of $\phi_{{_F}}^{nc}$, which contains either two or one triangles for the respective interior or boundary edges. Denote the collection of triangles in $\o_{{_F}}$ by \[ {\cal T}_{_F} = \{ K\in {\cal T}: \o_{_F} \cap K \neq \emptyset \}. \] Let ${\cal E}_{b,{_F}}$ be the collection of the boundary edges of $\o_{_F}$ that does not contain the edge $F$. Then the collection of edges of triangles in $ {\cal T}_{_F}$ is given by $$ {\cal E}_{_F} = \{ E \in {\cal E} : E \cap \overline{\o}_{_F} \neq \emptyset\} = \{F\} \cup {\cal E}_{b,{_F}} \quad \forall\,\, F \in {\cal E}. $$ It is also easy to check that \begin{equation}\label{non-basisf} \phi_{_F}^{nc}({\bf x}) \equiv 1\quad\mbox{on }\,\, F \quad\mbox{and}\quad \int_{E} \phi_{_F}^{nc}\,ds =0 \quad\forall\,\, E \in {\cal E}_{b,{_F}}. \end{equation} The set of functions $\{\phi_{{_F}}^{nc}\}_{F\in{\cal E}}$ forms a partition of the unity in $\O$: \[ \sum_{F\in {\cal E}} \phi_{_F}^{nc}({\bf x}) \equiv 1 \quad \forall\, {\bf x}\in \O, \] which leads to the following decomposition of the error flux: \ {\bf E}_c= \sum_{F\in {\cal E}} \big(\phi_{{_F}}^{nc}\,{\bf E}_c\big) = \sum_{F\in {\cal E}} \big(-\phi_{{_F}}^{nc} A\nabla e_c\big). \ On edge $F\in {\cal E}_{_I}\cup{\cal E}_{_N}$, denote the normal components of the numerical flux by \begin{equation}\label{n-flux} \hat{\sigma}^+_{c,{_F}}=\big(\hat{\mbox{\boldmath$\sigma$}}_c|_{K^+_{_F}} \cdot{\bf n}_{_F}\big)|_{{_F}} \quad\mbox{and}\quad \hat{\sigma}^-_{c,{_F}}=\big(\hat{\mbox{\boldmath$\sigma$}}_c|_{K^-_{_F}} \cdot{\bf n}_{_F}\big)|_{{_F}} \end{equation} and the jump of the numerical flux by \ j^c_{f,{_F}} \equiv\jump{\hat{\mbox{\boldmath$\sigma$}}_c\cdot {\bf n}_{_F}}_{_F} = \left\{\begin{array}{lll} \hat{\sigma}^-_{c,{_F}}- \hat{\sigma}^+_{c,{_F}}, & \forall\,\, F\in {\cal E}_{_I},\\[3mm] \hat{\sigma}^-_{c,{_F}}-g_{_N}, &\forall\,\, F\in {\cal E}_{_N}. \end{array} \right. \ By the first equality in (\ref{non-basisf}) and the continuity of the normal component of the true flux, it is easy to see that the jump of the normal component of the local error flux $\phi_{{_F}}^{nc} \,{\bf E}_c$ on edge $F\in {\cal E}_{_I}\cup{\cal E}_{_N}$ satisfies \begin{equation}\label{c-jump} \jump{\phi_{_F}^{nc}\,{\bf E}_c \cdot {\bf n}_{_F}}_{_F} =- j^c_{f,{_F}} . \end{equation} Note that \[ \phi_{_F}^{nc}\,{\bf E}_c \cdot {\bf n}_{_F} =\left(-\phi_{_F}^{nc}A \nabla e_c\right) \cdot {\bf n}_{_E} \approx 0 \quad\mbox{on }\,\, E \in {\cal E}_{b, {_F}}. \] Therefore, we introduce the following approximation to the local error flux $\phi_{_F}^{nc}\,{\bf E}_c$ on the local patch $\o_{_F}$: \begin{itemize} \item[(1)] for every $F\in {\cal E}_{_D}$, set \begin{equation}\label{rt-D} \bsigma^{\Delta}_{c,{_F}} =0 \quad \mbox{on }\,\, K^-_{_F}; \end{equation} \item[(2)] for every edge $F\in {\cal E}_{_I}\cup{\cal E}_{_N}$, find $\bsigma^{\Delta}_{c,{_F}} \in {\cal V}^c_{_{-1, F}}$ such that \begin{equation} \label{c-mini} \| A^{-1/2}\bsigma^{\Delta}_{c, {_F}} \|_{0,\o_{_F}} = \min_{\mbox{\boldmath$\tau$} \in {\cal V}^c_{_{-1,F}}} \|A^{-1/2}\mbox{\boldmath$\tau$} \|_{0,\o_{_F}}, \end{equation} where ${\cal V}^c_{_{-1, F}}$ with ${\cal V}= RT$ or $BDM$ is a local finite element space defined by \[ {\cal V}^c_{_{-1, F}} = \{ \mbox{\boldmath$\tau$}\in\! L^2(\omega_{_F}\!) \big|\, \mbox{\boldmath$\tau$}|_{_K}\!\!\in\!{\cal V}(\!K)\,\forall\, K\in {\cal T}_{_F}, \, \jump{\mbox{\boldmath$\tau$}\cdot{\bf n}_{_F}}_{_F}\!\! =\! -j^c_{f,{_F}},\, \, \mbox{\boldmath$\tau$}|_{_E}\cdot{\bf n}_{_E}\! = 0\, \forall \,E \in\! {\cal E}_{b, {_F}}\!\}. \] \end{itemize} With the approximations defined in (\ref{rt-D}) and (\ref{c-mini}), the global approximation to the error flux is then defined by \begin{equation}\label{app-err-f} \bsigma^{\Delta}_c = \sum_{F\in{\cal E}} \bsigma^{\Delta}_{c,{_F}} = \sum_{F\in{\cal E}_{_I}} \bsigma^{\Delta}_{c,{_F}}+\sum_{F\in{\cal E}_{_N}} \bsigma^{\Delta}_{c,{_F}}. \end{equation} This yields the following recovered flux for the conforming linear element: \begin{equation}\label{app-f} \mbox{\boldmath$\sigma$}_c = \bsigma^{\Delta}_c +\hat{\mbox{\boldmath$\sigma$}}_c\in H({\rm div};\O). \end{equation} The fact that $\mbox{\boldmath$\sigma$}_c \in H({\rm div};\O)$ follows from (\ref{c-jump}) that \[ \jump{\mbox{\boldmath$\sigma$}_c\cdot{\bf n}_{_F}}_{_F} = \jump{(\bsigma^{\Delta}_c+ \hat{\mbox{\boldmath$\sigma$}}_c)\cdot{\bf n}_{_F}}_{_F} =- j^c_{f,{_F}}+ \jump{\hat{\mbox{\boldmath$\sigma$}}_c\cdot{\bf n}_{_F}}_{_F}=0 \quad\mbox{on }\,\, F\in{\cal E}_{_I}. \] \subsubsection{Solution of (\ref{c-mini})} The recovered flux defined in (\ref{app-f}) requires solutions of the local problems defined in (\ref{c-mini}), which are constrained minimization problems. This section studies solutions of (\ref{c-mini}). To this end, let $\mbox{\boldmath$\phi$}^{rt}_{_F}$ be the local $RT$ basis function given in (\ref{rt}) in Appendix A, define the global $RT$ basis function associated with the edge $F$ by \[ \mbox{\boldmath$\psi$}^{rt}_{{_F}} = \left\{\begin{array}{llll} \mbox{\boldmath$\phi$}^{rt}_{_F}|_{K_{_F}^-}, & {\bf x} \in K_{_F}^-, \\[4mm] -\mbox{\boldmath$\phi$}^{rt}_{_F}|_{K_{_F}^+}, & {\bf x} \in K_{_F}^+, \\[4mm] 0, & {\bf x} \not\in \o_{_F}, \end{array} \right. \forall\,\, F\in{\cal E}_{_I} \,\mbox{ and }\, \mbox{\boldmath$\psi$}^{rt}_{{_F}} = \left\{\begin{array}{llll} \mbox{\boldmath$\phi$}^{rt}_{_F}|_{K_{_F}^-}, & {\bf x} \in K_{_F}^-, \\[4mm] 0, & {\bf x} \not\in \o_{_F}, \end{array} \right. \forall\,\, F\in {\cal E}_{_D}\cup{\cal E}_{_N} . \] for any $F\in{\cal E}_{_I}$ and by \[ \mbox{\boldmath$\psi$}^{rt}_{{_F}} = \left\{\begin{array}{llll} \mbox{\boldmath$\phi$}^{rt}_{_F}|_{K_{_F}^-}& {\bf x} \in K_{_F}^-, \\[4mm] 0& {\bf x} \not\in \o_{_F} \end{array} \right. \] for any $F\in{\cal E}_{_D}\cup{\cal E}_{_N}$. Denote by $\mbox{\boldmath$\psi$}^{rt,-}_{{_F}}$ and $\mbox{\boldmath$\psi$}^{rt,+}_{{_F}}$ the restriction of $\mbox{\boldmath$\psi$}^{rt}_{{_F}}$ on $K_{_F}^-$ and $K_{_F}^+$, respectively. To solve (\ref{c-mini}), set \begin{equation}\label{sigmajN} \mbox{\boldmath$\sigma$}^c_{j,{_F}}=-j^c_{f,{_F}} \mbox{\boldmath$\psi$}_{_F}^{rt,-} \quad\mbox{on }\,\, K_{_F}^- \end{equation} for any $F\in{\cal E}_{_N}$ and set \begin{equation}\label{sigmajI} \mbox{\boldmath$\sigma$}^c_{j,{_F}} = \left\{\begin{array}{lll} -j^c_{f,{_F}} \mbox{\boldmath$\psi$}_{_F}^{rt,-}, & \mbox{on} & K_{_F}^-, \\[4mm] 0, & \mbox{on} & K_{_F}^+ \end{array}\right. \end{equation} for any $F\in {\cal E}_{_I}$. By (\ref{proprt}), it is easy to check that for $F\in{\cal E}_{_I}\cup{\cal E}_{_N}$ \[ \jump{\mbox{\boldmath$\sigma$}^c_{j,{_F}} \cdot{\bf n}_{_F}}_{_F}= - j^c_{f,{_F}} \quad\mbox{and}\quad \mbox{\boldmath$\sigma$}^c_{j,{_F}} \cdot{\bf n}_{_E}=0\quad\mbox{on}\,\,E\in{\cal E}_{b,{_F}}. \] Hence, for any Neumman boundary edge $F\in {\cal E}_{_N}$, we have \begin{equation}\label{rt-N} \bsigma^{\Delta}_{c,{_F}} =\mbox{\boldmath$\sigma$}^c_{j,{_F}} \quad\mbox{on }\,\, K_{_F}^-, \end{equation} and for any interior edge $F\in{\cal E}_{_I}$, we have \[ \bsigma^{\Delta}_{c,{_F}} - \mbox{\boldmath$\sigma$}^c_{j,{_F}} \in H({\rm div};\o_{_F}) \quad\mbox{and}\quad \left(\bsigma^{\Delta}_{c,{_F}} - \mbox{\boldmath$\sigma$}^c_{j,{_F}}\right)|_{_E} \cdot {\bf n}_{_E}=0\quad\forall\,\, E\in {\cal E}_{b,{_F}}. \] Let \[ \tilde{\bsigma}_{c,{_F}} = \bsigma^{\Delta}_{c,{_F}} - \mbox{\boldmath$\sigma$}^c_{j,{_F}}, \quad H_0({\rm div};\o_{_F})=\{ \mbox{\boldmath$\tau$}\in H({\rm div};\o_{_F})\, |\,\mbox{\boldmath$\tau$}\cdot{\bf n}|_{\partial \o_{_F}} = 0\}, \] and \[ {\cal V}^c_{_{F}} = \{ \mbox{\boldmath$\tau$}\in H_0({\rm div};\o_{_F}) \,|\, \mbox{\boldmath$\tau$}|_{_K}\in {\cal V}(K)\,\,\forall\,\, K\in {\cal T}_{_F} \}, \] then the minimization problem in (\ref{c-mini}) for $F\in {\cal E}_{_I}$ is equivalent to finding $\tilde{\bsigma}_{c,{_F}} \in {\cal V}^c_{_{F}}$ such that \ \| A^{-1/2}\left(\tilde{\bsigma}_{c,{_F}} +\mbox{\boldmath$\sigma$}^c_{j,{_F}}\right) \|_{0,\o_{_F}} = \min_{\mbox{\boldmath$\tau$} \in {\cal V}^c_{_{F}}} \|A^{-1/2}\left(\mbox{\boldmath$\tau$}+\mbox{\boldmath$\sigma$}^c_{j,{_F}}\right) \|_{0,\o_{_F}}. \ The corresponding variational formulation is to find $\tilde{\bsigma}_{c,{_F}} \in {\cal V}^c_{_{F}}$ such that \begin{equation} \label{c-var} \left(A^{-1}\tilde{\bsigma}_{c,{_F}} ,\,\mbox{\boldmath$\tau$}\right)_{0,\o_{_F}} = -\left(A^{-1}\mbox{\boldmath$\sigma$}^c_{j,{_F}} ,\,\mbox{\boldmath$\tau$}\right)_{0,\o_{_F}} \quad\forall\,\,\mbox{\boldmath$\tau$}\in {\cal V}^c_{_{F}}. \end{equation} Note that for any interior edge (\ref{c-var}) has either one (RT) or two (BDM) unknowns. Their explicit formulas will be introduced in the subsequent section. \subsubsection{Explicit Formula for Flux Recovery} This section derives explicit formulas for the solution of (\ref{c-var}) and, hence, for the $RT$ and $B\!D\!M$ recoveries. First, we consider the $RT$ recovery. Since $\tilde{\bsigma}_{c,rt,{_F}} \in RT^c_{_{F}}\subset H_0({\rm div};\o_{_F})$, we have \[ \tilde{\bsigma}_{c,rt,{_F}} =\bsigma^{\Delta}_{c,rt,{_F}} - \mbox{\boldmath$\sigma$}^c_{j,{_F}} = a_{rt,{_F}} j^c_{f,{_F}} \mbox{\boldmath$\psi$}_{{_F}}^{rt} \quad\mbox{on}\,\,\omega_{_F} \] for all $F\in {\cal E}_{_I}$, which, together with (\ref{c-var}), yields \[ a_{rt,{_F}} = \displaystyle\frac{ \beta_{rt,{_F}}^- } { \beta_{rt,{_F}}^- +\beta_{rt,{_F}}^+ } \quad\mbox{with}\quad \beta_{rt,{_F}}^\pm = \left(A^{-1} \mbox{\boldmath$\psi$}_{{_F}}^{rt},\,\mbox{\boldmath$\psi$}_{{_F}}^{rt}\right)_{K_{_F}^\pm}. \] Hence, for any interior edge $F\in{\cal E}_{_I}$, we have \begin{equation}\label{rt-I} \bsigma^{\Delta}_{c,rt,{_F}} = \tilde{\bsigma}_{c,rt,{_F}} + \mbox{\boldmath$\sigma$}^c_{j,{_F}} = \left\{ \begin{array}{lll} -\left(1-a_{rt,{_F}}\right) j^c_{f,{_F}} \mbox{\boldmath$\psi$}_{{_F}}^{rt,-}, & \mbox{on} & K_{_F}^-, \\[3mm] a_{rt,{_F}} j^c_{f,{_F}} \mbox{\boldmath$\psi$}_{{_F}}^{rt,+} , & \mbox{on} & K_{_F}^+. \end{array} \right. \end{equation} Combining with (\ref{app-err-f}) and (\ref{rt-N}), the global approximation to the error flux is given by \begin{equation}\label{app-err-f-rt} \bsigma^{\Delta}_{c,rt} = \sum_{F\in{\cal E}_{_I}} \bsigma^{\Delta}_{c,rt,{_F} -\sum_{F\in{\cal E}_{_N}} j^c_{f,{_F}} \mbox{\boldmath$\psi$}_{_F}^{rt,-} \end{equation} with $\bsigma^{\Delta}_{c,rt,{_F}}$ defined in (\ref{rt-I}). Since the numerical flux is a piecewise constant vector, it has the following local representation on each element $K\in{\cal T}$ (see Lemma 4.4 of \cite{CaZh:09}): \[ \hat{\mbox{\boldmath$\sigma$}}_c|_{_K} = \sum_{F\in \partial K} \left( \hat{\mbox{\boldmath$\sigma$}}_c|_{_F}\cdot {\bf n}_{_K} \right) \mbox{\boldmath$\phi$}_{{_F}}^{rt}, \] where ${\bf n}_{_K}$ is the unit outward vector normal to $\partial K$. Globally, for any interior edge $F\in{\cal E}_{_I}$, we have \begin{equation}\label{n-flux-rep} \hat{\mbox{\boldmath$\sigma$}}^c_{{_F}} = \left\{ \begin{array}{lll} \hat{\sigma}^-_{c,{_F}} \mbox{\boldmath$\psi$}_{{_F}}^{rt,-}, & \mbox{on} & K_{_F}^-, \\[5mm] \hat{\sigma}^+_{c,{_F}} \mbox{\boldmath$\psi$}_{{_F}}^{rt,+}, & \mbox{on} & K_{_F}^+. \end{array} \right. \end{equation} Now, by (\ref{app-f}) and (\ref{app-err-f-rt}), the explicit formula for the recovered flux using the RT element is then \begin{equation}\label{rt-explicit} \mbox{\boldmath$\sigma$}_{c}^{rt} = \bsigma^{\Delta}_{c,rt}+\hat{\mbox{\boldmath$\sigma$}}_c = \sum_{F\in{\cal E}} \sigma_{_{c,F}}^{rt} \mbox{\boldmath$\psi$}_{_F}^{rt}, \end{equation} where the nodal value (i.e., the normal component of $\mbox{\boldmath$\sigma$}_{c}^{rt}$ on the edge $F$), $\sigma_{_{c,F}}^{rt}$, is given by \begin{equation}\label{rt-coef} \sigma_{_{c,F}}^{rt} = \left\{\begin{array}{llll} a_{rt,{_F}} \hat{\sigma}^-_{c,{_F}} +\left(1-a_{rt,{_F}}\right) \hat{\sigma}^+_{c,{_F}}, & F\in{\cal E}_{_I}, \\[2mm] g_{_N}, & F\in{\cal E}_{_N},\\[2mm] \hat{\sigma}^-_{c,{_F}}, & F\in{\cal E}_{_D} \end{array} \right. \end{equation} with the nodal values of the numerical fluxes, $\hat{\sigma}^+_{c,{_F}}$ and $\hat{\sigma}^-_{c,{_F}}$, defined in (\ref{n-flux}). Note that for any interior edge $F\in{\cal E}_{_I}$, the nodal value of the recovered flux is an average of the numerical fluxes. For interface problems, the recovered flux in (\ref{rt-explicit}) and the resulting estimator are similar to those introduced and analyzed in \cite{CaZh:09}. To this end, let $A|_{_K} = \alpha_{_K} I$ for any $K\in{\cal T}$, where $\alpha_{_K}$ and $I$ are constant and the identity matrix, respectively. Let \[ \alpha_{_F}^-=\alpha_{K^-_{_F}} \quad{and}\quad\alpha_{_F}^+=\alpha_{K^+_{_F}}, \] then \[ \beta_{_F}^- = \displaystyle\frac{1}{\alpha_{_F}^-} \left(\mbox{\boldmath$\psi$}_{_F}^{rt},\, \mbox{\boldmath$\psi$}_{_F}^{rt}\right)_{K_{_F}^-} \quad\mbox{and}\quad \beta_{_F}^+ = \displaystyle\frac{1}{\alpha_{_F}^+} \left(\mbox{\boldmath$\psi$}_{_F}^{rt}, \,\mbox{\boldmath$\psi$}_{_F}^{rt}\right)_{K_{_F}^+}. \] For a regular triangulation, the ratio of $\left(\mbox{\boldmath$\psi$}_{_F}^{rt},\, \mbox{\boldmath$\psi$}_{_F}^{rt}\right)_{K_{_F}^-} $ and $\left(\mbox{\boldmath$\psi$}_{_F}^{rt},\, \mbox{\boldmath$\psi$}_{_F}^{rt}\right)_{K_{_F}^+} $ are bounded above and below. Thus \begin{equation}\label{c-equi} a_{rt,{_F}}=\displaystyle\frac{\beta_{{_F}}^-}{\beta_{{_F}}^- + \beta_{{_F}}^+} \approx \displaystyle\frac{\alpha_{_F}^+}{\alpha_{_F}^- + \alpha_{_F}^+} \quad \mbox{and}\quad 1-a_{rt,{_F}}=\displaystyle\frac{\beta_{{_F}}^+}{\beta_{{_F}}^- + \beta_{{_F}}^+} \approx \displaystyle\frac{\alpha_{_F}^-}{\alpha_{_F}^- + \alpha_{_F}^+}. \end{equation} (Here and thereafter, we will use $x \approx y$ to mean that there exist two positive constants $C_1$ and $C_2$ independent of the mesh size such that $C_1 x \leq y\leq C_2 x$.) (\ref{c-equi}) indicates that the weights in the nodal values of the recovered flux may be replaced by $\displaystyle\frac{\alpha_{_F}^+}{\alpha_{_F}^- + \alpha_{_F}^+}$ and $\displaystyle\frac{\alpha_{_F}^-}{\alpha_{_F}^- + \alpha_{_F}^+}$, respectively. Next, we consider the $B\!D\!M$ recovery. For edge $F\in{\cal E}$, let ${\bf s}_{_F}$ and ${\bf e}_{_F}$ be endpoints of $F$ such that ${\bf e}_{_F}-{\bf s}_{_F} =h_{_F}{\bf t}_{_F}$. Let $\mbox{\boldmath$\phi$}^{bdm}_{_{s, F}}$ and $\mbox{\boldmath$\phi$}^{bdm}_{_{e, F}}$ be the two local $B\!D\!M$ basis functions associated withe vertices ${\bf s}_{_F}$ and ${\bf e}_{_F}$, respectively. For $i=\{s,e\}$, define the global $B\!D\!M$ basis functions associated with the edge $F$ by \[ \mbox{\boldmath$\psi$}^{bdm}_{_{i, F}} \!=\! \left\{\!\!\begin{array}{llll} \mbox{\boldmath$\phi$}^{bdm}_{_{i, F}}|_{K_{_F}^-}, \!&\! {\bf x} \in K_{_F}^-, \\[3mm] -\mbox{\boldmath$\phi$}^{bdm}_{_{i,F}}|_{K_{_F}^+}, \!&\! {\bf x} \in K_{_F}^+, \\[3mm] 0, \!&\! {\bf x} \not\in \o_{_F}, \end{array} \right. \!\forall F\!\in\!{\cal E}_{_I} \mbox{ and } \mbox{\boldmath$\psi$}^{bdm}_{_{i, F}}\! = \!\left\{\!\!\begin{array}{llll} \mbox{\boldmath$\phi$}^{bdm}_{_{i, F}}|_{K_{_F}^-}, \!&\! {\bf x} \in K_{_F}^-, \\[4mm] 0, \!&\! {\bf x} \not\in \o_{_F}, \end{array} \right.\! \forall F\!\in \!{\cal E}_{_D}\!\!\cup{\cal E}_{_N}. \] Denote by $\mbox{\boldmath$\psi$}^{bdm,-}_{_{i, F}}$ and $\mbox{\boldmath$\psi$}^{bdm,+}_{_{i, F}}$ the restriction of $\mbox{\boldmath$\psi$}^{bdm}_{_{i, F}}$ on $K_{_F}^-$ and $K_{_F}^+$, respectively. Again, since $\tilde{\bsigma}_{c,bdm,{_F}}=\bsigma^{\Delta}_{c,bdm,{_F}} - \mbox{\boldmath$\sigma$}^c_{j,{_F}}\in B\!D\!M^c_{_{F}}\subset H_0({\rm div};\o_{_F})$, we have \[ \tilde{\bsigma}_{c,bdm,{_F} = a_{bdm,{_F}} j^c_{f,{_F}} \mbox{\boldmath$\psi$}_{s,{_F}}^{bdm} + b_{bdm,{_F}} j^c_{f,{_F}} \mbox{\boldmath$\psi$}_{e,{_F}}^{bdm} \quad \forall\,\, F\in {\cal E}_{_I}. \] For $i,\, j \in\{ s,\,e\}$, let \ \beta_{ij,{_F}}^\pm = \left(A^{-1} \mbox{\boldmath$\psi$}_{i,{_F}}^{bdm},\, \mbox{\boldmath$\psi$}_{j,{_F}}^{bdm}\right)_{K_{_F}^\pm} \quad\mbox{and}\quad \beta_{ij,{_F}} = \beta_{ij,{_F}}^- +\beta_{ij,{_F}}^+ . \ Solving (\ref{c-var}) yields \begin{equation}\label{BDM-a} a_{bdm,{_F}}= \displaystyle\frac{ (\beta_{ss,{_F}}^- + \beta_{se,{_F}}^-)\beta_{ee,{_F}} -(\beta_{se,{_F}}^- +\beta_{ee,{_F}}^- )\beta_{se,{_F}} } {\beta_{ss,{_F}}\beta_{ee,{_F}}- \beta_{se,{_F}}^2} \end{equation} and \begin{equation} \label{BDM-b} b_{bdm,{_F}}= \displaystyle\frac{ (\beta_{ss,{_F}}^- + \beta_{se,{_F}}^-)\beta_{ss,{_F}} -(\beta_{se,{_F}}^- +\beta_{ee,{_F}}^- )\beta_{se,{_F}} } {\beta_{ss,{_F}}\beta_{ee,{_F}}- \beta_{se,{_F}}^2} \end{equation} Since $\mbox{\boldmath$\psi$}_{{_F}}^{rt,\pm}=\mbox{\boldmath$\psi$}_{s,{_F}}^{bdm,\pm}+\mbox{\boldmath$\psi$}_{e,{_F}}^{bdm,\pm}$, we have \begin{equation}\label{bdm-I} \bsigma^{\Delta}_{c,bdm,{_F}}=\tilde{\bsigma}_{c,bdm,{_F}}+\mbox{\boldmath$\sigma$}^c_{j,{_F}} =\hat{a}_{bdm,{_F}} j^c_{f,{_F}} \mbox{\boldmath$\psi$}_{s,{_F}}^{bdm} + \hat{b}_{bdm,{_F}} j^c_{f,{_F}} \mbox{\boldmath$\psi$}_{e,{_F}}^{bdm} \quad \forall\,\, F\in {\cal E}_{_I} \end{equation} with \ \hat{a}_{bdm,{_F}} = \left\{ \begin{array}{llll} 1- a_{bdm,{_F}} , & \mbox{on}&K_{{_F}}^-, \\[2mm] a_{bdm,{_F}}, & \mbox{on} & K_{{_F}}^+, \end{array} \right. \quad\mbox{and}\quad \hat{b}_{bdm,{_F}} = \left\{ \begin{array}{llll} 1- b_{bdm,{_F}} , & \mbox{on}&K_{{_F}}^-, \\[2mm] b_{bdm,{_F}}, & \mbox{on} & K_{{_F}}^+. \end{array} \right. \ Thus, by (\ref{app-err-f}) and (\ref{rt-N}) the error flux is given by \begin{equation}\label{app-err-f-bdm} \bsigma^{\Delta}_{c,bdm} =\sum_{F\in{\cal E}_{_I}} \bsigma^{\Delta}_{c,bdm,{_F} -\sum_{F\in{\cal E}_{_N}} j^c_{f,{_F}} (\mbox{\boldmath$\psi$}_{s,{_F}}^{bdm,-} + \mbox{\boldmath$\psi$}_{e,{_F}}^{bdm,-}). \end{equation} with $\bsigma^{\Delta}_{c,bdm,{_F}}$ defined in (\ref{bdm-I}). Now, by (\ref{app-f}) and (\ref{n-flux-rep}), the explicit formula for the recovered flux using the $B\!D\!M$ element is then \begin{equation}\label{bdm-explicit} \mbox{\boldmath$\sigma$}_{c}^{bdm} = \bsigma^{\Delta}_{c,bdm}+\hat{\mbox{\boldmath$\sigma$}}_c = \sum_{F\in{\cal E}} \sigma_{c,s,{_F}}^{bdm} \mbox{\boldmath$\psi$}_{s,{_F}}^{bdm} + \sum_{F\in{\cal E}} \sigma_{c,e,{_F}}^{bdm} \mbox{\boldmath$\psi$}_{e,{_F}}^{bdm}, \end{equation} where \ \sigma_{c,s,{_F}}^{bdm} = \left\{ \begin{array}{lllll} a_{bdm,{_F}} \hat{\sigma}^-_{c,{_F}} +(1-a_{bdm,{_F}}) \hat{\sigma}^+_{c,{_F}}, & F\in{\cal E}_{_I},\\[2mm] g_{_N}, &F\in {\cal E}_{_N},\\[2mm] \hat{\sigma}^-_{c,{_F}}, & F\in{\cal E}_{_D}. \end{array} \right. \] and \[ \sigma_{c,e,{_F}}^{bdm} = \left\{ \begin{array}{lllll} b_{bdm,{_F}} \hat{\sigma}^-_{c,{_F}} +(1-b_{bdm,{_F}}) \hat{\sigma}^+_{c,{_F}}, & F\in{\cal E}_{_I},\\[2mm] g_{_N}, &F\in {\cal E}_{_N},\\[2mm] \hat{\sigma}^-_{c,{_F}}, & F\in{\cal E}_{_D}. \end{array} \right. \] Note that, for any interior edge $F\in{\cal E}_{_I}$, the coefficients of the recovered flux are again weighted averages of the numerical fluxes. For interface problems $A|_{_K} = \alpha_{_K} I$ with a regular triangulation, by a careful calculation, we can show that \begin{equation}\label{c-equi-bdm} a _{bdm,{_F}} \approx \displaystyle\frac{\alpha_{_F}^+}{\alpha_{_F}^- + \alpha_{_F}^+} \approx b _{bdm,{_F}} \quad \mbox{and}\quad 1-a_{bdm,{_F}} \approx \displaystyle\frac{\alpha_{_F}^-}{\alpha_{_F}^- + \alpha_{_F}^+} \approx 1-b_{bdm,{_F}}. \end{equation} \subsection{Explicit Gradient Recovery for Mixed Method} This section introduces an explicit gradient recovery based on the mixed finite element approximation in (\ref{problem_mixed}). Since derivation is similar to that in the previous section, we briefly describe the recovery procedure and present an explicit formula of the recovered gradient. Let $(\mbox{\boldmath$\sigma$}_m, u_m)$ be the solution of (\ref{problem_mixed}). Denote by \[ \hat{\bf{\rho}}_m= -A^{-1}\mbox{\boldmath$\sigma$}_m \quad\mbox{and}\quad {\bf E}_m=\mbox{\boldmath$\sigma$}-\mbox{\boldmath$\sigma$}_m=-A\left(\nabla u -\hat{\bf{\rho}}_m\right) \] the numerical gradient and the flux error, respectively. Then the gradient error is given by \begin{equation}\label{g-error} \nabla u -\hat{\bf{\rho}}_m=-A^{-1}{\bf E}_m. \end{equation} Denote the tangential components of the numerical gradient on edge $F\in{\cal E}$ by \begin{equation}\label{g-t} \hat{\rho}^+_{m, {_F}}=\left(\hat{\bf{\rho}}_m|_{K^+_{_F}} \cdot{\bf t}_{_F}\right)|_{{_F}} \quad\mbox{and}\quad \hat{\rho}^-_{m, {_F}}=\left(\hat{\bf{\rho}}_m|_{K^-_{_F}} \cdot{\bf t}_{_F}\right)|_{{_F}} \end{equation} and the edge jump of the numerical gradient on edge $F\in {\cal E}_{_I}\cup{\cal E}_{_D}$ by \ j^m_{g,{_F}} \equiv \jump{ \hat{\bf{\rho}}_m\cdot {\bf t}_{_F}}_{_F} =\left\{ \begin{array}{lll} \hat{\rho}^-_{m, {_F}}-\hat{\rho}^+_{m, {_F}}, & \forall \,\,F\in{\cal E}_{_I}, \\[2mm] \hat{\rho}^-_{m, {_F}}-\nabla g_{_D}\!\!\cdot{\bf t}_{_F} , &\forall \,\,F\in{\cal E}_{_D}. \end{array} \right. \ By the continuity of the tangential components of the true gradient, the edge jump of the tangential component of the local error gradient is given by \begin{equation}\label{g-jump} \jump{-\phi_{_F}^{nc} A^{-1} {\bf E}_m \cdot {\bf t}_{_F}}_{_F} =-j^m_{g,{_F}}. \end{equation} Since $\hat{\rho}^+_{m, {_F}}$ and $\hat{\rho}^-_{m, {_F}}$ are affine functions defined on $F\in{\cal E}$, the $N\!D$ element is needed for their approximations. To this end, as in the $B\!D\!M$ case, for edge $F\in{\cal E}$, let ${\bf s}_{_F}$ and ${\bf e}_{_F}$ be endpoints of $F$ such that ${\bf e}_{_F}-{\bf s}_{_F} =h_{_F}{\bf t}_{_F}$, let $\mbox{\boldmath$\phi$}^{nd}_{i, {_F}}$ ($i=s,\,e$) be the local $N\!D$ basis functions given in (\ref{nd}) in Appendix A, and define the global $N\!D$ basis functions associated with edge $F$ by \[ \mbox{\boldmath$\psi$}^{nd}_{i, {_F}} =\! \left\{\!\!\begin{array}{llll} \mbox{\boldmath$\phi$}^{nd}_{i, {_F}}|_{K_{_F}^-},& {\bf x} \in K_{_F}^-, \\[4mm] -\mbox{\boldmath$\phi$}^{nd}_{i,{_F}}|_{K_{_F}^+}, & {\bf x} \in K_{_F}^+, \\[4mm] 0,& {\bf x} \not\in \o_{_F}, \end{array} \right. \forall\,F\!\in{\cal E}_{_I} \mbox{ and } \mbox{\boldmath$\psi$}^{nd}_{i, {_F}} =\! \left\{\!\!\begin{array}{llll} \mbox{\boldmath$\phi$}^{nd}_{i, {_F}}|_{K_{_F}^-},& {\bf x} \in K_{_F}^-, \\[4mm] 0,& {\bf x} \not\in \o_{_F}, \end{array} \right. \forall\,F\! \in{\cal E}_{_D}\!\!\cup{\cal E}_{_N}. \] Denote by $\mbox{\boldmath$\psi$}^{nd,-}_{i, {_F}}$ and $\mbox{\boldmath$\psi$}^{nd,+}_{i, {_F}}$ the restriction of $\mbox{\boldmath$\psi$}^{nd}_{i, {_F}}$ on $K_{_F}^-$ and $K_{_F}^+$, respectively. Denote the tangential components of the numerical gradient, $\hat{\rho}^+_{m, {_F}}$ and $\hat{\rho}^-_{m, {_F}}$, at the endpoints by \[ d^\pm_{s,{_F}} = \hat{\rho}^\pm_{m,{_F}}({\bf s}_{_F}) \quad\mbox{and}\quad d^\pm_{e,{_F}} = \hat{\rho}^\pm_{m,{_F}}({\bf e}_{_F}), \] respectively. Then the numerical gradient has the following representation in local $N\!D$ bases $$ \hat{\bf{\rho}}_m = d^-_{s,{_F}}\mbox{\boldmath$\psi$}_{s,{_F}}^{nd,-} - d^-_{e, {_F}} \mbox{\boldmath$\psi$}_{e,{_F}}^{nd,-},\,\, \mbox{ on } \,\, K_{_F}^-, \quad\mbox{for}\quad F\in {\cal E}_{_D} $$ and $$ \hat{\bf{\rho}}_m = \left\{ \begin{array}{llll} d^-_{s,{_F}} \mbox{\boldmath$\psi$}_{s,{_F}}^{nd,-} - d^-_{e,{_F}} \mbox{\boldmath$\psi$}_{e,{_F}}^{nd,-}, & \mbox{on} & K_{_F}^-, \\[4mm] d^+_{s,{_F}} \mbox{\boldmath$\psi$}_{s,{_F}}^{nd,+} - d^+_{e,{_F}} \mbox{\boldmath$\psi$}_{e,{_F}}^{nd,+}, & \mbox{on} & K_{_F}^+, \end{array} \right. \quad\mbox{for}\quad F\in {\cal E}_{_I}. $$ Let \[ c_{s,{_F}} =\left\{\begin{array}{ll} d^-_{s,{_F}} - d^+_{s,{_F}}, & \forall \,\, F\in{\cal E}_{_I},\\[4mm] d^-_{s,{_F}} , & \forall \,\, F\in{\cal E}_{_D} \end{array}\right. \quad \mbox{and }\,\, c_{e,{_F}} = \left\{\begin{array}{ll} d^-_{e,{_F}} - d^+_{e,{_F}}, & \forall \,\, F\in{\cal E}_{_I}, \\[4mm] d^-_{e,{_F}}, & \forall \,\, F\in{\cal E}_{_D}. \end{array}\right. \] A simple calculation leads to \begin{equation}\label{jump-g} j^m_{g,{_F}} = \left\{ \begin{array}{lll} c_{s,{_F}} (\mbox{\boldmath$\psi$}^{nd}_{s,{_F}}\cdot{\bf t}_{_F}) - c_{e,{_F}} (\mbox{\boldmath$\psi$}^{nd}_{e,{_F}}\cdot{\bf t}_{_F}), & \forall \,\,F\in{\cal E}_{_I}, \\[4mm] c_{s,{_F}} (\mbox{\boldmath$\psi$}^{nd}_{s,{_F}}\cdot{\bf t}_{_F}) - c_{e,{_F}} (\mbox{\boldmath$\psi$}^{nd}_{e,{_F}}\cdot{\bf t}_{_F}) -\nabla g_{_D}\!\!\cdot{\bf t}_{_F} , &\forall \,\,F\in{\cal E}_{_D}. \end{array} \right. \end{equation} Let \begin{equation}\label{rhojD} \bf{\rho}^m_{j,{_F}} = - c_{s,{_F}} \mbox{\boldmath$\psi$}_{s,{_F}}^{nd,-} + c_{e,{_F}} \mbox{\boldmath$\psi$}_{e,{_F}}^{nd,-}, \quad\mbox{on }\,\, K_{_F}^- \end{equation} for $F\in{\cal E}_{_D}$, and let \begin{equation} \label{rhojI} \bf{\rho}^m_{j,{_F}} = \left\{ \begin{array}{lll} - c_{s,{_F}} \mbox{\boldmath$\psi$}_{s,{_F}}^{nd,-} + c_{e,{_F}} \mbox{\boldmath$\psi$}_{e,{_F}}^{nd,-}, & \mbox{on} & K_{_F}^-, \\[4mm] 0, & \mbox{on} & K_{_F}^+ \end{array} \right. \end{equation} for $F\in{\cal E}_{_I}$. By the properties of the $N\!D$ basis functions in (\ref{propnd0}) and (\ref{propnd}), it is easy to check that \[ \jump{\bf{\rho}^m_{j,{_F}}\cdot{\bf t}_{_F}}_{_F}=-j^m_{g,{_F}} \quad\mbox{and}\quad \bf{\rho}^m_{j,{_F}} \cdot{\bf t}_{_E}=0\quad\mbox{on}\,\,E\in{\cal E}_{b,{_F}} \] for $F\in {\cal E}_{_I}\cup{\cal E}_{_D}$. Let \[ H_0(\mbox{curl};\o_{_F})=\{ \mbox{\boldmath$\tau$}\in H(\mbox{curl};\o_{_F})\, |\,\mbox{\boldmath$\tau$}\cdot{\bf t} |_{\partial \o_{_F}} = 0\}, \] and let \[ N\!D^m_{{_F}} = \{ \mbox{\boldmath$\tau$}\in H_0(\mbox{curl};\o_{_F}) \,|\, \mbox{\boldmath$\tau$}|_{_K}\in N\!D(K)\,\,\forall\,\, K\in {\cal T}_{_F}\}. \] In a similar fashion as that of the previous section, by (\ref{g-jump}), we introduce the following approximation to the error gradient: \begin{equation}\label{app-err-g} \brho^{\Delta}_{m,nd}=\sum_{F\in{\cal E}_{_D}} \brho^{\Delta}_{m,nd,{_F}}+\sum_{F\in{\cal E}_{_I}} \brho^{\Delta}_{m,nd,{_F}}, \end{equation} where \begin{equation}\label{app-err-g-F} \brho^{\Delta}_{m,nd,{_F}}=\left\{\begin{array}{ll} \bf{\rho}^m_{j,{_F}}, & F\in {\cal E}_{_D},\\[2mm] \tilde{\bf{\rho}}_{m,{_F}} + \bf{\rho}^m_{j,{_F}}, & F\in {\cal E}_{_I}. \end{array}\right. \end{equation} Here, $\tilde{\bf{\rho}}_{m,{_F}}\in N\!D^m_{{_F}}$ is the solution of the following minimization problem: \begin{equation}\label{mix-mini} \|A^{1/2}\left(\tilde{\bf{\rho}}_{m,{_F}}+\bf{\rho}^m_{j,{_F}}\right)\|_{0,\o_{_F}} =\min_{\mbox{\boldmath$\tau$}\in N\!D^m_{{_F}}} \|A^{1/2}\left(\mbox{\boldmath$\tau$}+\bf{\rho}^m_{j,{_F}}\right)\|_{0,\o_{_F}}. \end{equation} Let $$ \gamma_{ij,{_F}}^- = (A \mbox{\boldmath$\psi$}_{i,{_F}}^{nd}, \mbox{\boldmath$\psi$}_{j,{_F}}^{nd})_{K_{_F}^-},\quad \gamma_{ij,{_F}}^+ = (A \mbox{\boldmath$\psi$}_{i,{_F}}^{nd}, \mbox{\boldmath$\psi$}_{j,{_F}}^{nd})_{K_{_F}^+}, \quad\mbox{and}\quad \gamma_{ij,{_F}} = \gamma_{ij,{_F}}^- +\gamma_{ij,{_F}}^+ $$ for $i,\, j \in\{ s,\,e\}$. Solving (\ref{mix-mini}) leads to $$ \tilde{\bf{\rho}}_{m,{_F}} = \rho_{s,{_F}} \mbox{\boldmath$\psi$}_{s,{_F}}^{nd} + \rho_{e,{_F}} \mbox{\boldmath$\psi$}_{e,{_F}}^{nd} $$ with coefficients given by \begin{eqnarray*} \rho_{s,{_F}} &=&\displaystyle\frac{\left(c_{s,{_F}}\gamma_{ss,{_F}}^- -c_{e,{_F}}\gamma_{se,{_F}}^- \right)\gamma_{ee,{_F}} - \left(c_{s,{_F}}\gamma_{se,{_F}}^- -c_{e,{_F}}\gamma_{ee,{_F}}^- \right)\gamma_{se,{_F}} }{\gamma_{ss,{_F}}\gamma_{ee,{_F}}- \gamma_{se,{_F}}^2}\\ &=& \displaystyle\frac{c_{s,{_F}} \left(\gamma_{ss,{_F}}^-\gamma_{ee,{_F}} - \gamma_{se,{_F}}^- \gamma_{se,{_F}} \right) - c_{e,{_F}} \left(\gamma_{se,{_F}}^- \gamma_{ee,{_F}} -\gamma_{ee,{_F}}^- \gamma_{se,{_F}}\right) }{\gamma_{ss,{_F}}\gamma_{ee,{_F}}- \gamma_{se,{_F}}^2} \end{eqnarray*} and \begin{eqnarray*} \rho_{e,{_F}} &=&\displaystyle\frac{\left(c_{s,{_F}}\gamma_{se,{_F}}^- -c_{e,{_F}}\gamma_{ee,{_F}}^- \right)\gamma_{ss,{_F}} - \left(c_{s,{_F}}\gamma_{ss,{_F}}^- -c_{e,{_F}}\gamma_{se,{_F}}^- \right)\gamma_{se,{_F}} }{\gamma_{ss,{_F}}\gamma_{ee,{_F}}- \gamma_{se,{_F}}^2}\\ &=& \displaystyle\frac{c_{s,{_F}} \left(\gamma_{se,{_F}}^-\gamma_{ss,{_F}} - \gamma_{ss,{_F}}^- \gamma_{se,{_F}} \right) - c_{e,{_F}} \left(\gamma_{ee,{_F}}^- \gamma_{ss,{_F}} -\gamma_{se,{_F}}^- \gamma_{se,{_F}}\right) }{\gamma_{ss,{_F}}\gamma_{ee,{_F}}- \gamma_{se,{_F}}^2}. \end{eqnarray*} Hence, we have \begin{equation}\label{app-err-g-I} \brho^{\Delta}_{m,nd,{_F}}=\tilde{\bf{\rho}}_{m,{_F}}+\bf{\rho}^m_{j,{_F}} = \left\{ \begin{array}{lllll} (\rho_{s,{_F}} - c_{s,{_F}}) \mbox{\boldmath$\psi$}_{s,{_F}}^{nd} + (\rho_{e,{_F}} +c_{e,{_F}}) \mbox{\boldmath$\psi$}_{e,{_F}}^{nd},&\mbox{on}& K_{_F}^- ,\\[4mm] \rho_{s,{_F}} \mbox{\boldmath$\psi$}_{s,{_F}}^{nd} + \rho_{e,{_F}} \mbox{\boldmath$\psi$}_{e,{_F}}^{nd},&\mbox{on}& K_{_F}^+ \end{array}\right. \end{equation} for interior edge $F\in{\cal E}_{_I}$. Now, the explicit formula for the recovered gradient using $N\!D$ element is then \begin{equation} \bf{\rho}^{nd}_m =\brho^{\Delta}_{m,nd}+\hat{\bf{\rho}}_m = \sum_{F\in {\cal E}} a^{nd}_{m,{_F}}\mbox{\boldmath$\psi$}^{nd}_{s,{_F}}+ \sum_{F\in {\cal E}} b^{nd}_{m,{_F}}\mbox{\boldmath$\psi$}^{nd}_{e,{_F}}, \end{equation} where the coefficients are given by $$ a^{nd}_{m,{_F}} = \left\{\begin{array}{lllll} \rho_{s,{_F}} +d_{s,{_F}}^+, & F\in {\cal E}_{_I}, \\[2mm] d^-_{s,{_F}},& F\in{\cal E}_{{_N}},\\[2mm] \nabla g_{{_D}}\!\!\cdot{\bf t}_{_F},& F\in{\cal E}_{{_D}} \end{array} \right. \quad \mbox{and} \quad b^{nd}_{m,{_F}} = \left\{\begin{array}{lllll} \rho_{e,{_F}} -d_{e,{_F}}^+, & F\in {\cal E}_{_I}, \\[2mm] -d^-_{e,{_F}},& F\in{\cal E}_{{_N}},\\[2mm] -\nabla g_{{_D}}\!\!\cdot{\bf t}_{_F},& F\in{\cal E}_{{_D}}. \end{array} \right. $$ Notice that $$ \rho_{s,{_F}} +d_{s,{_F}}^+ = \ell_{s,{_F}} d^-_{s,{_F}} + (1-\ell_{s,{_F}})d^+_{s,{_F}} - \displaystyle\frac{ c_{e,{_F}} \left(\gamma_{se,{_F}}^- \gamma^+_{ee,{_F}} -\gamma_{ee,{_F}}^- \gamma^+_{se,{_F}}\right) }{\gamma_{ss,{_F}}\gamma_{ee,{_F}}- \gamma_{se,{_F}}^2} $$ with $$ \ell_{s,{_F}} = \displaystyle\frac{ \left(\gamma_{ss,{_F}}^-\gamma_{ee,{_F}} - \gamma_{se,{_F}}^- \gamma_{se,{_F}} \right) } {\gamma_{ss,{_F}}\gamma_{ee,{_F}}- \gamma_{se,{_F}}^2}, $$ and $$ \rho_{e,{_F}} -d_{e,{_F}}^+ = - \ell_{e,{_F}} d^-_{e,{_F}} - (1-\ell_{e,{_F}})d^+_{e,{_F}} + \displaystyle\frac{c_{s,{_F}} \left(\gamma_{se,{_F}}^-\gamma^-_{ss,{_F}} - \gamma_{ss,{_F}}^- \gamma^+_{se,{_F}} \right) }{\gamma_{ss,{_F}}\gamma_{ee,{_F}}- \gamma_{se,{_F}}^2}. $$ with $$ \ell_{e,{_F}} = \displaystyle\frac{ \left(\gamma_{ee,{_F}}^-\gamma_{ss,{_F}} - \gamma_{se,{_F}}^- \gamma_{se,{_F}} \right) } {\gamma_{ss,{_F}}\gamma_{ee,{_F}}- \gamma_{se,{_F}}^2}. $$ Note that, for any interior edge $F\in{\cal E}_{_I}$, the coefficients of the recovered gradient are weighted averages of the numerical gradients plus some high order terms. For interface problems $A|_{_K} = \alpha_{_K} I$ with a regular triangulation, by a careful calculation, we can show that \begin{equation}\label{c-equi-bdm} \ell _{s,{_F}} \approx \displaystyle\frac{\alpha_{_F}^-}{\alpha_{_F}^- + \alpha_{_F}^+} \approx \ell _{e,{_F}} \quad \mbox{and}\quad 1-\ell_{s,{_F}} \approx \displaystyle\frac{\alpha_{_F}^+}{\alpha_{_F}^- + \alpha_{_F}^+} \approx 1-\ell_{e,{_F}}. \end{equation} \subsection{Explicit Flux and Gradient Recoveries for Nonconforming Method} Let $u_{nc}$ be the solution of (\ref{problem_nc}). Denote by \[ \hat{\bf{\rho}}_{nc}= \nabla_h u_{nc} \quad\mbox{and}\quad \hat{\mbox{\boldmath$\sigma$}}_{nc}= -A^{-1}\nabla_h u_{nc}=-A^{-1} \hat{\bf{\rho}}_{nc} \] the numerical gradient and the numerical flux, respectively. This section introduces explicit formulas of the recovered flux $\mbox{\boldmath$\sigma$}_{nc}\in H({\rm div}, \O)$ and the recovered gradient $\bf{\rho}_{nc}\in H({\rm curl},\O)$ based on $\hat{\mbox{\boldmath$\sigma$}}_{nc}$ and $ \hat{\bf{\rho}}_{nc}$. Again, derivations are similar to those in the previous sections and, hence, descriptions in this section are brief. Denote the solution error, the flux error, and the gradient error by \[ e_{nc} = u-u_{nc}, \quad {\bf E}_{nc} = \mbox{\boldmath$\sigma$}-\hat{\mbox{\boldmath$\sigma$}}_{nc} = -A\nabla_h e_{nc}, \mbox{ and } \nabla u -\nabla_h u_{nc}=\nabla_h e_{nc}= -A^{-1}{\bf E}_{nc}, \] respectively. Denote the normal components of the numerical flux on edge $F\in{\cal E}$ by \begin{equation}\label{n-flux-nc} \hat{\sigma}^+_{nc, {_F}}=\left(\hat{\mbox{\boldmath$\sigma$}}_{nc}|_{K^+_{_F}} \cdot{\bf n}_{_F}\right)|_{{_F}} \quad\mbox{and}\quad \hat{\sigma}^-_{nc, {_F}}=\left(\hat{\mbox{\boldmath$\sigma$}}_{nc}|_{K^-_{_F}} \cdot{\bf n}_{_F}\right)|_{{_F}} \end{equation} and the edge jump of the numerical flux by \[ j^{nc}_{f,{_F}} \equiv \jump{\hat{\mbox{\boldmath$\sigma$}}_{nc}\cdot{\bf n}_{_F}}_{_F} =\left\{ \begin{array}{lll} \hat{\sigma}^-_{nc, {_F}}-\hat{\sigma}^+_{nc, {_F}}, & \forall \,\,F\in{\cal E}_{_I}, \\[2mm] \hat{\sigma}^-_{nc, {_F}}-g_{_N} , &\forall \,\,F\in{\cal E}_{_N}. \end{array} \right. \] Denote the tangential components of the numerical gradient on edge $F\in{\cal E}$ by \begin{equation}\label{g-t-nc} \hat{\rho}^+_{nc, {_F}}=\left(\hat{\bf{\rho}}_{nc}|_{K^+_{_F}}\cdot{\bf t}_{_F}\right)|_{{_F}} \quad\mbox{and}\quad \hat{\rho}^-_{nc, {_F}}=\left(\hat{\bf{\rho}}_{nc}|_{K^-_{_F}}\cdot{\bf t}_{_F}\right)|_{{_F}} \end{equation} and the edge jump of the numerical gradient by \[ j^{nc}_{g,{_F}} \equiv \jump{\hat{\bf{\rho}}_{nc}\cdot{\bf t}_{_F}}_{_F} =\left\{ \begin{array}{lll} \hat{\rho}^-_{nc, {_F}}-\hat{\rho}^+_{nc, {_F}}, & \forall \,\,F\in{\cal E}_{_I}, \\[2mm] \hat{\rho}^-_{nc, {_F}}-\nabla g_{_D}\cdot{\bf t}_{_F}, &\forall \,\,F\in{\cal E}_{_D}. \end{array} \right. \] By the continuity of the true flux and true gradient, we have \begin{equation}\label{jump-nc} \jump{\phi_{_F}^{nc} {\bf E}_{nc} \cdot {\bf n}_{_F}}_{_F} =-j^{nc}_{f,{_F}} \,\,\mbox{ and }\,\, \jump{-\phi_{_F}^{nc} A^{-1} {\bf E}_{nc} \cdot {\bf t}_{_F}}_{_F} =-j^{nc}_{g,{_F}}. \end{equation} \subsubsection{ Explicit Formula for Flux Recovery} In a similar fashion as in Section~4.1, the approximation to the error flux using the $RT$ element is given by \begin{equation}\label{sdncrtf} \bsigma^{\Delta}_{nc,rt} = \sum_{F\in{\cal E}_{_I}}\bsigma^{\Delta}_{nc,rt,{_F}} + \sum_{F\in{\cal E}_{_N}}\bsigma^{\Delta}_{nc,rt,{_F}} = \sum_{F\in{\cal E}_{_I}}\bsigma^{\Delta}_{nc,rt,{_F}} -\sum_{F\in{\cal E}_{_N}}j^{nc}_{f,{_F}} \mbox{\boldmath$\psi$}_{_F}^{rt,-} \end{equation} with \begin{equation}\label{app-err-f-nc-F} \bsigma^{\Delta}_{nc,rt,{_F}} = \left\{ \begin{array}{lll} - \left(1-a_{rt,{_F}}\right) j^{nc}_{f,{_F}} \mbox{\boldmath$\psi$}_{{_F}}^{rt,-}, & \mbox{on} & K_{_F}^-, \\[3mm] a_{rt,{_F}} j^{nc}_{f,{_F}} \mbox{\boldmath$\psi$}_{{_F}}^{rt,+}, & \mbox{on} & K_{_F}^+, \end{array} \right. \end{equation} where $a_{rt,{_F}}$ is defined in Section 4.1.2. Now, the explicit flux recovery using the $RT$ element is given by \begin{equation}\label{rt-explicit-nc} \mbox{\boldmath$\sigma$}_{nc}^{rt} = \bsigma^{\Delta}_{nc,rt}+\hat{\mbox{\boldmath$\sigma$}}_c = \sum_{F\in{\cal E}} \sigma_{_{nc,F}}^{rt} \mbox{\boldmath$\psi$}_{_F}^{rt} \in H(\mbox{div},\O), \end{equation} where the nodal value $\sigma_{_{nc,F}}^{rt}$ is given by \begin{equation}\label{rt-coef-nc} \sigma_{_{nc,F}}^{rt} = \left\{\begin{array}{llll} a_{rt,{_F}} \hat{\sigma}^-_{nc,{_F}} +(1-a_{rt,{_F}})\hat{\sigma}^+_{nc,{_F}}, & F\in{\cal E}_{_I}, \\[2mm] g_{_N}, & F\in{\cal E}_{_N},\\[2mm] \hat{\sigma}^-_{nc,{_F}}, & F\in{\cal E}_{_D}. \end{array} \right. \end{equation} Using the $B\!D\!M$ element, the approximation to the error flux is given by \begin{eqnarray}\nonumber \bsigma^{\Delta}_{nc,bdm} &=& \sum_{F\in{\cal E}_{_I}}\bsigma^{\Delta}_{nc,bdm,{_F}} + \sum_{F\in{\cal E}_{_N}}\bsigma^{\Delta}_{nc,bdm,{_F}}\\[2mm]\label{bdm-Ib} &=&\sum_{F\in{\cal E}_{_I}} j^{nc}_{f,{_F}}\left(a_{bdm,{_F}} \mbox{\boldmath$\psi$}_{s,{_F}}^{bdm} + b_{bdm,{_F}} \mbox{\boldmath$\psi$}_{e,{_F}}^{bdm}\right) -\sum_{F\in{\cal E}_{_N}} j^{nc}_{f,{_F}} (\mbox{\boldmath$\psi$}_{s,{_F}}^{bdm,-}+\mbox{\boldmath$\psi$}_{e,{_F}}^{bdm,-}), \end{eqnarray} where $a_{bdm,{_F}}$ and $b_{bdm,{_F}}$ are defined in Section 4.1.2. Now, the explicit flux recovery using the $B\!D\!M$ element is given by \begin{equation}\label{bdm-explicit-nc} \mbox{\boldmath$\sigma$}_{nc}^{bdm} = \sum_{F\in{\cal E}} \sigma_{nc,s,{_F}}^{bdm} \mbox{\boldmath$\psi$}_{s,{_F}}^{bdm} +\sum_{F\in{\cal E}} \sigma_{nc,e,{_F}}^{bdm} \mbox{\boldmath$\psi$}_{e,{_F}}^{bdm} \in H(\mbox{div},\O) \end{equation} where where $\sigma_{nc,s,{_F}}^{bdm}$ and $\sigma_{nc,e,{_F}}^{bdm}$ is similar to $\sigma_{c,s,{_F}}^{bdm}$ and $\sigma_{c,e,{_F}}^{bdm}$ defined in Section 4.1.2, i.e., \ \sigma_{nc,s,{_F}}^{bdm} = \left\{ \begin{array}{lllll} a_{bdm,{_F}} \hat{\sigma}^-_{nc,{_F}} +(1-a_{bdm,{_F}}) \hat{\sigma}^+_{nc,{_F}}, & F\in{\cal E}_{_I},\\[2mm] g_{_N}, &F\in {\cal E}_{_N},\\[2mm] \hat{\sigma}^-_{nc,{_F}}, & F\in{\cal E}_{_D}. \end{array} \right. \] and \[ \sigma_{nc,e,{_F}}^{bdm} = \left\{ \begin{array}{lllll} b_{bdm,{_F}} \hat{\sigma}^-_{nc,{_F}} +(1-b_{bdm,{_F}}) \hat{\sigma}^+_{nc,{_F}}, & F\in{\cal E}_{_I},\\[2mm] g_{_N}, &F\in {\cal E}_{_N},\\[2mm] \hat{\sigma}^-_{nc,{_F}}, & F\in{\cal E}_{_D}. \end{array} \right. \] \subsubsection{Explicit Formula for Gradient Recovery} Let $\mbox{\boldmath$\phi$}^{ne}_{_F}$ be the local $N\!E$ basis function given in Appendix A, define the global $N\!E$ basis function associated with the edge $F$ by \ \mbox{\boldmath$\psi$}^{ne}_{F} =\! \left\{\!\!\begin{array}{llll} \mbox{\boldmath$\phi$}^{ne}_{_F}|_{K_{_F}^-},& {\bf x} \in K_{_F}^-, \\[2mm] -\mbox{\boldmath$\phi$}^{ne}_{_F}|_{K_{_F}^+}, & {\bf x} \in K_{_F}^+, \\[2mm] 0,& {\bf x} \not\in \o_{_F}, \end{array} \right. \forall F\in{\cal E}_{_I} \mbox{ and } \mbox{\boldmath$\psi$}^{ne}_{F} =\! \left\{\!\!\begin{array}{llll} \mbox{\boldmath$\phi$}^{ne}_{_F}|_{K_{_F}^-},& {\bf x} \in K_{_F}^-, \\[2mm] 0,& {\bf x} \not\in \o_{_F}, \end{array} \right. \forall F\in{\cal E}_{_D}\cup{\cal E}_{_N}. \ Denote by $\mbox{\boldmath$\psi$}^{ne,-}_{F}$ and $\mbox{\boldmath$\psi$}^{ne,+}_{F}$ the restriction of $\mbox{\boldmath$\psi$}^{ne}_{F}$ on $K_{_F}^-$ and $K_{_F}^+$, respectively. Let \[ a_{ne,{_F}} = \displaystyle\frac{ \beta_{ne,{_F}}^- } { \beta_{ne,{_F}}^- +\beta_{ne,{_F}}^+ } \quad\mbox{with}\quad \beta_{ne,{_F}}^\pm = \left(A^{-1} \mbox{\boldmath$\psi$}_{{_F}}^{ne},\,\mbox{\boldmath$\psi$}_{{_F}}^{ne}\right)_{K_{_F}^\pm}. \] Then the approximation to the gradient error is \begin{equation}\label{app-err-g-nc} \brho^{\Delta}_{{nc},ne} =\sum_{F\in{\cal E}_{_I}}\brho^{\Delta}_{{nc},ne,{_F}} +\sum_{F\in{\cal E}_{_D}}\brho^{\Delta}_{{nc},ne,{_F}} = \sum_{F\in{\cal E}_{_I}}\brho^{\Delta}_{{nc},ne,{_F}} +\sum_{F\in{\cal E}_{_D}}j^{nc}_{g,{_F}} \mbox{\boldmath$\psi$}_{_F}^{ne,-} \end{equation} with \begin{equation}\label{local-rho-nc} \brho^{\Delta}_{{nc},ne,{_F}} = \left\{ \begin{array}{lll} -\left(1-a_{ne,{_F}}\right) j^{nc}_{g,{_F}} \mbox{\boldmath$\psi$}_{{_F}}^{ne,-}, & \mbox{on} & K_{_F}^-, \\[3mm] a_{ne,{_F}} j^{nc}_{g,{_F}} \mbox{\boldmath$\psi$}_{{_F}}^{ne,+}, & \mbox{on} & K_{_F}^+. \end{array} \right. \end{equation} Now, the explicit gradient recovery using the $N\!E$ element is given by \begin{equation}\label{ne-explicit-nc} \bf{\rho}_{nc,{_F}}^{ne} = \brho^{\Delta}_{nc,ne} +\hat{\bf{\rho}}_{nc} = \sum_{F\in{\cal E}} \rho_{nc,{_F}}^{ne} \mbox{\boldmath$\psi$}_{_F}^{ne} \in H(\mbox{curl},\O), \end{equation} where the nodal value $\rho_{{nc,F}}^{ne}$ is given by \begin{equation}\label{ne-coef-nc} \rho_{nc,{_F}}^{ne} = \left\{ \begin{array}{lll} a_{ne,{_F}}\hat{\rho}^-_{nc,F} +\left(1- a_{ne,{_F}}\right)\hat{\rho}^+_{nc,F} & F\in{\cal E}_{_I}, \\[3mm] \nabla g_{_D}\!\!\cdot {\bf t}_{_F}, & F\in{\cal E}_{_D}, \\[3mm] \hat{\rho}^-_{nc,F}, & F\in{\cal E}_{_N}. \end{array} \right. \end{equation} Next, we describe the recovered gradient using the $N\!D$ element. Let \[ a_{_F}^{nc}= \displaystyle\frac{ (\gamma_{ss,{_F}}^- - \gamma_{se,{_F}}^-) \gamma_{ee,{_F}} -(\gamma_{se,{_F}}^- -\gamma_{ee,{_F}}^-) \gamma_{se,{_F}} } {\gamma_{ss,{_F}}\gamma_{ee,{_F}}- \gamma_{se,{_F}}^2} >0 , \quad \mbox{and} \] \[ b_{_F}^{nc}=\displaystyle\frac{ (\gamma_{se,{_F}}^- -\gamma_{ee,{_F}}^-) \gamma_{ss,{_F}} -(\gamma_{ss,{_F}}^- - \gamma_{se,{_F}}^-) \gamma_{se,{_F}} } {\gamma_{ss,{_F}}\gamma_{ee,{_F}}- \gamma_{se,{_F}}^2} <0 \] with $ \gamma_{ij,{_F}}^\pm$ and $ \gamma_{ij,{_F}}$, ($i,j \in \{s,e\}$) defined in Section 4.2. Similar to the gradient recovery using the $N\!D$ element for the mixed method, the approximation to the error gradient is \begin{equation}\label{app-err-g-nd-nc} \brho^{\Delta}_{nc, nd} = \sum_{F\in{\cal E}_{_I}} \left(\hat{a}^{nc}_{{_F}} j^{nc}_{g,{_F}} \mbox{\boldmath$\psi$}_{s,{_F}}^{nd} + \hat{b}_{_F}^{nc} j^{nc}_{g,{_F}} \mbox{\boldmath$\psi$}_{e,{_F}}^{nd}\right) -\sum_{F\in{\cal E}_{_D}} j^{nc}_{g,{_F}} (\mbox{\boldmath$\psi$}^{nd,-}_{s,{_F}}-\mbox{\boldmath$\psi$}^{nd,-}_{e,{_F}}). \end{equation} where the coefficients $\hat{a}^{nc}_{{_F}}$ and $\hat{b}^{nc}_{{_F}}$ are given by $$ \hat{a}^{nc}_{{_F}} = \left\{ \begin{array}{llll} 1-a_{_F}^{nc}, & \mbox{on}&K_{{_F}}^-, \\[2mm] -a_{_F}^{nc}, & \mbox{on}& K_{{_F}}^+, \end{array} \right. \quad\mbox{and}\quad \hat{b}^{nc}_{{_F}} = \left\{ \begin{array}{llll} 1+b_{_F}^{nc}, & \mbox{on}&K_{{_F}}^-, \\[2mm] b_{_F}^{nc}, & \mbox{on}& K_{{_F}}^+, \end{array} \right. $$ Now, the recovered gradient using the $N\!D$ element is given by \begin{equation} \bf{\rho}_{nc}^{nd} = \brho^{\Delta}_{nc, nd} + \hat{\bf{\rho}}_{nc}^{nd} = \sum_{F\in{\cal E}}\rho_{nc,s,{_F}}^{nd} \mbox{\boldmath$\psi$}_{s,{_F}}^{nd} +\sum_{F\in{\cal E}}\rho_{nc,s,{_F}}^{nd} j^{nc}_{g,{_F}} \mbox{\boldmath$\psi$}_{e,{_F}}^{nd}, \end{equation} where the coefficients of $\mbox{\boldmath$\psi$}_{s,{_F}}^{nd}$ and $\mbox{\boldmath$\psi$}_{e,{_F}}^{nd}$ are given by \begin{equation} \hat{\rho}^{nd}_{nc, s,{_F}} = \left\{ \begin{array}{lllll} a_{_F}^{nc} \hat{\rho}^{-}_{nc,{_F}}+ \left(1-a_{_F}^{nc}\right) \hat{\rho}^{+}_{nc,{_F}}, & F\in {\cal E}_{_I},\\[2mm] \nabla g_{_D} \!\!\cdot{\bf t}_{_F},& F\in {\cal E}_{_D},\\[2mm] j^{nc}_{g,{_F}}, & F\in {\cal E}_{_N}, \end{array} \right. \end{equation} and \begin{equation} \hat{\rho}^{nd}_{nc, e,{_F}} = \left\{ \begin{array}{lllll} b_{_F}^{nc} \hat{\rho}^{-}_{nc,{_F}}+ \left(1+b_{_F}^{nc}\right) \hat{\rho}^{+}_{nc,{_F}}, & F\in {\cal E}_{_I},\\[2mm] -\nabla g_{_D} \!\!\cdot{\bf t}_{_F},& F\in {\cal E}_{_D},\\[2mm] -j^{nc}_{g,{_F}}, & F\in {\cal E}_{_N}. \end{array} \right. \end{equation} \section{Explicit A Posteriori Error Estimators}\label{estimators-a} \setcounter{equation}{0} With the explicit recoveries of the flux and gradient introduced in Section~4 for various finite element approximations, this section describes the corresponding recovery-based a posteriori error estimators. For the conforming linear element, we study two estimators using the respective $RT$ and $BDM$ recoveries. The global $RT$ a posteriori error estimator is given by \[ \eta^{rt}_c = \|A^{-1/2}\left(\mbox{\boldmath$\sigma$}_c^{rt} + A \nabla u_c\right)\|_{0,\O} = \|A^{-1/2} \bsigma^{\Delta}_{c,rt} \|_{0,\O}, \] and the $RT$ local error indicators on element $K\in{\cal T}$ and on edge $F\in{\cal E}$ are given by \[ \eta^{rt}_{c,{_K}} = \|A^{-1/2} \bsigma^{\Delta}_{c,rt} \|_{0,K} \quad\mbox{ and }\quad \eta^{rt}_{c,{_F}} = \|A^{-1/2} \bsigma^{\Delta}_{c,rt,{_F}} \|_{0,\o_{_F}}, \] respectively, where $\bsigma^{\Delta}_{c,rt}$ is given in (\ref{app-err-f-rt}) and $\bsigma^{\Delta}_{c,rt,{_F}}$ in (\ref{rt-N}) and (\ref{rt-I}). The global $B\!D\!M$ a posteriori error estimator is given by \[ \eta^{bdm}_c = \|A^{-1/2} \left(\mbox{\boldmath$\sigma$}_c^{bdm} + A\nabla u_c\right)\|_{0,\O} = \|A^{-1/2}\bsigma^{\Delta}_{c,bdm} \|_{0,\O}, \] and the $BDM$ local error indicators on element $K\in{\cal T}$ and on edge $F\in{\cal E}$ are given by \[ \eta^{bdm}_{c,{_K}} = \|A^{-1/2} \bsigma^{\Delta}_{c,bdm} \|_{0,K} \quad\mbox{ and }\quad \eta^{bdm}_{c,{_F}} = \|A^{-1/2} \bsigma^{\Delta}_{c,bdm,{_F}} \|_{0,\o_{_F}}, \] respectively, where $\bsigma^{\Delta}_{c,bdm}$ is defined in (\ref{app-err-f-bdm}) and $\bsigma^{\Delta}_{c,bdm,{_F}}$ in (\ref{rt-N}) and (\ref{bdm-I}). For the lowest-order mixed element, we study one estimator based on the explicit $N\!D$ recovery. The local error indicators on element $K\in{\cal T}$ and on edge $F\in{\cal E}$ are defined by \[ \eta^{nd}_{m,{_K}} = \|A^{1/2} \brho^{\Delta}_{m,nd} \|_{0,K} \quad\mbox{ and }\quad \eta^{nd}_{m,{_F}} = \|A^{1/2} \brho^{\Delta}_{m,nd,{_F}} \|_{0,\o_{_F}}, \] respectively, where $\brho^{\Delta}_{m,nd}$ is defined in (\ref{app-err-g}) and $\brho^{\Delta}_{m,nd,{_F}}$ in (\ref{app-err-g-F}) and (\ref{app-err-g-I}). The global error estimator is then defined by \[ \eta^{nd}_m = \|A^{-1/2}\left(\mbox{\boldmath$\sigma$}_m + A \bf{\rho}_m^{nd}\right)\|_{0,\O} = \|A^{1/2} \brho^{\Delta}_{m,nd} \|_{0,\O}. \] For the nonconforming linear element, again we introduce two estimators based on the $RT$-$N\!E$ and $B\!D\!M$-$N\!D$ recoveries. Let $c_1,\,\,c_2\in (0,1)$ be parameters to be determined such that $c_1+c_2=1$ (e.g, $c_1=c_2 =1/2$). The global $RT$-$N\!E$ error estimator is defined by \[ \eta^{rh}_{nc} = \left(c_1 \|A^{-1/2} \bsigma^{\Delta}_{nc,rt} \|_{0,\O}^2+c_2 \|A^{1/2} \brho^{\Delta}_{nc,ne} \|_{0,\O}^2\right)^{1/2}, \] and the $RT$-$W\!H$ local error indicators on element $K\in{\cal T}$ and on edge $F\in{\cal E}$ are defined respectively by \begin{eqnarray*} \eta^{rh}_{nc,{_K}} &=& \left(\!c_1 \|A^{-1/2} \bsigma^{\Delta}_{nc,rt} \|_{0,K}^2\! +c_2 \|A^{1/2} \brho^{\Delta}_{nc,ne} \|_{0,K}^2\!\right)^{1/2}\\[2mm] \mbox{and} \quad \eta^{rh}_{nc, {_F}} &=& \left(c_1 \|A^{-1/2} \!\bsigma^{\Delta}_{nc,rt,{_F}} \|_{0,\o_{_F}}^2 +c_2 \|A^{1/2} \brho^{\Delta}_{nc,ne,{_F}} \|_{0,\o_{_F}}^2\right)^{1/2}, \end{eqnarray*} where $\bsigma^{\Delta}_{nc,rt}$, $\bsigma^{\Delta}_{nc,rt, {_F}}$, $\brho^{\Delta}_{nc,ne}$ and $\brho^{\Delta}_{nc,ne,{_F}}$ are defined in (\ref{sdncrtf}), (\ref{app-err-f-nc-F}), (\ref{app-err-g-nc}), and (\ref{local-rho-nc}), respectively. Similarly, The global $B\!D\!M$-$N\!D$ error estimator is defined by \[ \eta^{rh}_{nc} = \left(c_1 \|A^{-1/2} \bsigma^{\Delta}_{nc,bdm} \|_{0,\O}^2+c_2\|A^{1/2} \brho^{\Delta}_{nc,nd} \|_{0,\O}^2\right)^{1/2}, \] and the local $B\!D\!M$-$N\!D$ error indicators on element $K\in{\cal T}$ and on edge $F\in{\cal E}$ are defined respectively by \begin{eqnarray*} \eta^{rh}_{nc,{_K}} &=& \left(\!c_1 \|A^{-1/2} \bsigma^{\Delta}_{nc,bdm} \|_{0,K}^2\! +c_2 \|A^{1/2} \brho^{\Delta}_{nc,nd} \|_{0,K}^2\!\right)^{1/2}\\[2mm] \mbox{and} \quad \eta^{rh}_{nc, {_F}} &=& \left(c_1 \|A^{-1/2} \bsigma^{\Delta}_{nc,bdm,{_F}} \|_{0,\o_{_F}}^2 +c_2 \|A^{1/2} \brho^{\Delta}_{nc,nd,{_F}} \|_{0,\o_{_F}}^2\right)^{1/2}, \end{eqnarray*} where $\bsigma^{\Delta}_{nc,bdm}$ and $\brho^{\Delta}_{nc,nd}$ are defined in (\ref{bdm-Ib}) and (\ref{app-err-g-nd-nc}), respectively. \section{Efficiency and Reliability}\label{estimators-b} \setcounter{equation}{0} This section establishes efficiency and reliability bounds of the estimators defined in Section 5 for interface problems (i.e, $A = \alpha\, I$ and $\alpha(x)$ is a piecewise constant with respect to the triangulation ${\cal T}$.). In order to show that the reliability constant are independent of the jump of $\alpha$, as usual, we assume that the distribution of the coefficients $\alpha_{_K}$ for all $K\in {\cal T}$ is locally quasi-monotone \cite{Pet:02}, which is slightly weaker than Hypothesis 2.7 in \cite{BeVe:00}. For convenience of readers, we restate it here. Let $\o_z$ be the union of all elements having $z$ as a vertex. For any $z\in{\cal N}$, let \[ \hat{\omega}_z=\{K\in\omega_z \,:\, \alpha_{_K} = \max_{{_K}'\in\omega_z} \alpha_{{_K}'}\}. \] \begin{defn}\label{defnquasimonotone} Given a vertex $z \in {\cal N}$, the distribution of the coefficients $\alpha_{_K}$, $K\in\omega_z$, is said to be {\em quasi-monotone} with respect to the vertex $z$ if there exists a subset $\tilde{\o}_{{_K},z,qm}$ of $\omega_z$ such that the union of elements in $\tilde{\o}_{{_K},z,qm}$ is a Lipschitz domain and that \begin{itemize} \item if $z\in{\cal N}\backslash{\cal N}_{_D}$, then $\{K\}\cup \hat{\o}_z \subset \tilde{\o}_{{_K},z,qm}$ and $\alpha_{_K}\leq \alpha_{{_K}'} \; \forall {_K}' \in \tilde{\o}_{{_K},z,qm}$; \item if $z\in{\cal N}_{_D}$, then $K\in \tilde{\o}_{{_K},z,qm}$, $\partial\tilde{\omega}_{{_K},z,qm}\cap\Gamma_D \neq \emptyset$, and $\alpha_{_K}\leq \alpha_{{_K}'} \; \forall {_K}' \in \tilde{\o}_{{_K},z,qm}$. \end{itemize} The distribution of the coefficients $\alpha_{_K}$, $K\in{\cal T}$, is said to be locally {\em quasi-monotone} if it is quasi-monotone with respect to every vertex $z\in{\cal N}$. \end{defn} Let $f_{_{\cal T}}$ be the $L^2$ projection of $f$ onto the space of piecewise constant defined on elements of ${\cal T}$, let \begin{eqnarray* H_f &=& \left(\sum_{K\in{\cal T}} H_{f,K}^2\right)^{1/2} \quad\mbox{with}\quad H_{f,K} = \displaystyle\frac{h_K}{\sqrt{\alpha_K}}\, \|f-f_{_{\cal T}}\|_{0,K} \quad\forall\;K\in{\cal T}, \end{eqnarray*} and let \ \hat{H}_f=\left(\sum_{z\in {\cal N}\cap ({\cal S}\cup\Gamma_D)} \sum_{K\subset\omega_z} \displaystyle\frac{h^2_K}{\alpha_K}\,\|f\|^2_{0,K} +\sum_{z\in {\cal N}\setminus ({\cal S}\cup\Gamma_D)} \sum_{K\subset\omega_z} \displaystyle\frac{h^2_K}{\alpha_K}\, \|f-\Xint-_{\omega_z}f\,dx\|^2_{0,K}\right)^{1/2} \] where $\Xint-_{\omega_z}f\,dx=\int_{\hat{\o}_z} f \psi_z\,dx\big/ \int_{\hat{\o}_z} \psi_z\,dx$ is a weighted average of $f$ over $\hat{\o}_z$ and $\psi_z$ is a linear nodal basis function at $z\in{\cal N}$. \begin{rem} For various lower order finite element approximations, the second term in $\hat{H}_f$ is of higher order than $\eta_{_{\cal E}}$ {\em (}defined below in {\em (\ref{edgeestimator-c}))} for $f\in L^2(\O)$ and so is the first term for $f\in L^p(\O)$ with $p > 2$ {\em (}see {\em \cite{CaVe:99}}{\em )}. \end{rem} \subsection{Conforming Elements} \begin{thm} Assume that the distribution of the coefficients are quasi-monotone. Then the error estimators $\eta^{rt}_c$ and $\eta^{bdm}_c$ satisfy the global reliability bound: \begin{equation}\label{rel-rt \|\alpha^{1/2} \nabla e_c\|_{0,\O} \leq C(\eta^{rt}_c + \hat{H}_f ) \quad\mbox{and}\quad \|\alpha^{1/2} \nabla e_c\|_{0,\O} \leq C(\eta^{bdm}_c + \hat{H}_f), \ee where the constants above are independent of $\alpha$ and the mesh size. \end{thm} \begin{proof} Inequalities (\ref{rel-rt}) may be established in a similar fashion as those in \cite{CaZh:09, CaZh:10c}. \end{proof} To prove the efficiency bound, consider the edge error estimator and indicator of the residual type: \begin{equation}\label{edgeestimator-c} \eta_{c,{\cal E}} := \left( \sum_{F\in{\cal E}_{_I} \cup{\cal E}_{_N}} \eta_{c,{_F}}^2 \right)^{1/2} \quad\mbox{with}\quad \eta_{c,{_F}} = \left\{\begin{array}{lll} h_{_F} j^c_{f,{_F}}\big/ \sqrt{\alpha_{_F}^+ + \alpha_{_F}^-}, & F\in{\cal E}_{_I},\\[4mm] h_{_F} j^c_{f,{_F}}\big/ \sqrt{ \alpha_{_F}^-}, & F\in {\cal E}_{_N}. \end{array} \right. \end{equation} Without assumptions on the distribution of the coefficient $\alpha$, it was proved by Petzoldt (see equation (5.7) in \cite{Pet:02}) that there exists a constant $C>0$ independent of $\alpha$ and the mesh size such that \begin{equation} \label{edgeestimator} \eta_{c,{_F}}^2 \leq C\left( \|\alpha^{-1/2} \nabla e_c\|_{\o_{_F}}^2 + \sum_{K\in{\cal T}_{_F}} H_{f,{_K}}^2 \right). \end{equation} Let ${\cal T}_{_K} = \{T\in{\cal T}: T \mbox{ and } K \mbox{ share at least one edge} \}$. \begin{thm}\label{thm:eff-c} The local indicators $\eta^{rt}_{c,F}$, $\eta^{rt}_{c,K}$, $\eta^{bdm}_{c,F}$, and $\eta^{bdm}_{c,K}$ defined in {\em Section 5} are efficient, i.e., there exists a constant $C>0$ independent of $\alpha$ and the mesh size such that \begin{eqnarray} \label{c-eff-F} \eta^{bdm}_{c,{_F}} \leq \eta^{rt}_{c,{_F}} &\leq& C \|\alpha^{1/2} \nabla e_c\|_{0,\o_{_F}} +C\left( \sum_{K\in{\cal T}_{_F}} H_{f,{_K}}^2\right)^{1/2}\\[2mm] \label{c-eff-K1} \mbox{and}\quad \eta^{bdm}_{c,{_K}}, \,\, \eta^{rt}_{c,{_K}} & \leq& C \|\alpha^{1/2} \nabla e_c\|_{0,\o_{_K}}+ C\left( \sum_{T\in{\cal T}_{_K}} H_{f,{_T}}^2\right)^{1/2} \end{eqnarray} \end{thm} \begin{proof} Without loss of generality, we establish the efficiency bounds only for interior edges. The first inequality of (\ref{c-eff-F}) is a direct consequence of the minimization problem in (\ref{c-mini}) and the fact that $RT^c_{-1,F}\subset BDM^c_{-1,F}$. To prove the second inequality of (\ref{c-eff-F}), we assume that the triangulation is regular. By the equivalence in (\ref{c-equi}) and the fact that $\|\mbox{\boldmath$\psi$}^{rt}_{_F}\|_{0,\o_{_F}}\leq C\,h^2_{_F}$, we have \begin{eqnarray*} \left(\eta^{rt}_{c,{_F}} \right)^2 &=& \|\alpha_{{_F}^-}^{-1/2}\bsigma^{\Delta}_{c,rt,{_F}}\|_{0,K^-_{_F}}^2 + \|\alpha_{{_F}^+}^{-1/2}\bsigma^{\Delta}_{c,rt,{_F}}\|_{0,K^+_{_F}}^2 \\[2mm] &\leq& C\left( \displaystyle\frac{1}{\alpha_{_F}^-} \left(\displaystyle\frac{\alpha_{_F}^-}{\alpha_{_F}^- + \alpha_{_F}^+} \right)^2 + \displaystyle\frac{1}{\alpha_{_F}^+} \left(\displaystyle\frac{\alpha_{_F}^+}{\alpha_{_F}^- + \alpha_{_F}^+} \right)^2 \right) \left(j_{f,{_F}}^c h_{_F}\right)^2 = C\, \eta_{c,{_F}}^2, \end{eqnarray*} which, combining with (\ref{edgeestimator}), implies the second inequality of (\ref{c-eff-F}). It is easy to see that \[ \left( \eta^{rt}_{c,{_K}} \right)^2 \leq \sum_{F\in{\cal E}_{_K}} \left( \eta^{rt}_{c,{_F}} \right)^2 \quad\mbox{and}\quad \left( \eta^{bdm}_{c,{_K}} \right)^2 \leq \sum_{F\in{\cal E}_{_K}} \left( \eta^{bdm}_{c,{_F}} \right)^2. \] Now, (\ref{c-eff-K1}) follows from (\ref{c-eff-F}). This completes the proof of the theorem. \end{proof} \subsection{Mixed Elements} \begin{thm} Assume that the distribution of the coefficient $\alpha$ is quasi-monotone. Then the error estimator $\eta^{nd}_{m}$ satisfies the following global reliability bound: \begin{equation} \|\alpha^{-1/2}{\bf E}_m\|_{0,\O} \leq C (\eta_m^{nd}+H_f + G_{\nabla_h\times(\alpha^{-1}\mbox{\boldmath$\sigma$}_m)}), \end{equation} where $G_{\nabla_h\times(\alpha^{-1}\mbox{\boldmath$\sigma$}_m)}$ is a higher order term if $\nabla_h\times(\alpha^{-1}\mbox{\boldmath$\sigma$}_m) \in L^p(\O)$ with $p>2$. Moreover, if ${\cal V} = RT$, then \begin{equation} \|\alpha^{-1/2}{\bf E}_m\|_{0,\O} \leq C (\eta_m^{nd}+H_f). \end{equation} \end{thm} \begin{proof} Let $\hat{\eta}_m$ be the implicit recovery-based estimator introduced in \cite{CaZh:10a}, i.e., \[ \hat{\eta}_{m} = \min_{\mbox{\boldmath$\tau$} \in N\!D} \|\alpha^{1/2}\mbox{\boldmath$\tau$} + \alpha^{-1/2}\mbox{\boldmath$\sigma$}_m\|_{0,\O}. \] It is obvious that $\hat{\eta}_{m} \leq \eta^{nd}_{m}$. Now, the theorem is a direct consequence of Theorem 6.2 of \cite{CaZh:10a}. \end{proof} The efficiency of the $\eta^{nd}_{m}$ may be established by a direct calculation similar to the proof of Theorem \ref{thm:eff-c}. However, the calculation is quite complicated in this case. We will prove it through the following Helmholtz decomposition (see, e.g., \cite{GiRa:86}) of the error flux ${\bf E}_m$: there exist $\xi_m\in H_{_D}^1(\O)$ and $\zeta_m \in H^1_N(\O)\equiv \{v\in H^1(\O)\big| \,v = 0\mbox{ on }\Gamma_N\}$ such that \begin{eqnarray}\label{H-decom} && {\bf E}_m = \alpha \nabla \xi_m+\nabla^{\perp} \zeta_m\\[2mm]\nonumbe \mbox{and && \|\alpha^{-1/2}{\bf E}_m\|^2_{0,\O} =\|\alpha^{1/2}\nabla \xi_m\|^2_{0,\O} + \|\alpha^{-1/2} \nabla^{\perp} \zeta_m\|^2_{0,\O}. \end{eqnarray} \begin{thm} The local indicators $\eta^{nd}_{m,{_F}}$ and $\eta^{nd}_{m,{_K}}$ and the global error estimator $\eta^{nd}_m$ are efficient, i.e., there exists a constant $C>0$ independent of $\alpha$ and the mesh size such that \begin{eqnarray} \label{m-eff} \eta^{nd}_{m,{_F}} &\leq& C \|\alpha^{-1/2} \nabla^{\perp}\zeta_m\|_{0,\o_{_F}}, \quad \eta^{nd}_{m,{_K}} \leq C \|\alpha^{-1/2} \nabla^{\perp}\zeta_m\|_{0,\o_{_K}},\\[2mm] \label{m-eff-G} \mbox{and}\quad \eta^{nd}_{m} &\leq & C \|\alpha^{-1/2} \nabla^{\perp}\zeta_m\|_{0,\O} \leq C \|\alpha^{1/2} {\bf E}_m\|_{0,\O}. \end{eqnarray} \end{thm} \begin{proof} Without loss of generality, we establish the efficiency bounds only for interior edges. Let $\eta_{m,{_F}}$ and $\eta_{m}$ be the respective edge indicator and estimator defined in \cite{CaZh:10a}, where \begin{equation}\label{6.11a} \eta_{m,{_F}}^2=\displaystyle\frac{\alpha^-_{_F}+\alpha^+_{_F}}{2} h_{_F} \int_{_F} \large| j^m_{g,{_F}}\large|^2\,ds. \end{equation} It is proved in Proposition 6.6 of \cite{CaZh:10a} that \begin{equation}\label{mixedge} \eta_{m,{_F}} \leq \|\alpha^{-1/2} \nabla^{\perp} \zeta_m\|_{0,\omega_{_F}} \;\mbox{ and }\; \eta_{m}\leq C\|\alpha^{-1/2} \nabla^{\perp} \zeta_m\|_{0,\Omega} \leq C\|\alpha^{1/2}{\bf E}_m\|_{0,\Omega}. \end{equation} Since $\|\mbox{\boldmath$\psi$}^{nd}_{i,F}\|_K\approx C\,h_{_F}$ for $i=s,e$, it follows from (\ref{mix-mini}) with $\mbox{\boldmath$\tau$}={\bf 0}$, (\ref{rhojI}), and the triangle inequality that \begin{eqnarray*} \eta^{nd}_{m,{_F}} & =& \|\alpha^{1/2} \brho^{\Delta}_{m,nd,{_F}}\|_{0,\omega_{_F}} \leq \|\alpha^{1/2} \bf{\rho}^m_{j,{_F}}\|_{0,\omega_{_F}} \\[2mm] &=& \sqrt{\alpha^-_{_F}} \left(|c_{s,{_F}}|\, \|\mbox{\boldmath$\psi$}^{nd,-}_{s,{_F}}\|_{0,K_{_F}^-} +|c_{e,{_F}}|\, \|\mbox{\boldmath$\psi$}^{nd,-}_{e,F}\|_{0,K_{_F}^-}\right)\\[2mm] &\leq &C\, h_{_F}\sqrt{\alpha^-_{_F}}\left(|c_{s,{_F}}|+|c_{e,{_F}}|\right) \end{eqnarray*} Note that \[ c_{s,{_F}}= j^m_{g,{_F}}({\bf s}_{_F}) \quad\mbox{and}\quad c_{e,{_F}}= j^m_{g,{_F}}({\bf e}_{_F}) \] and that $j^m_{g,{_F}}$ is an affine function on $F$, it is then easy to check that there exists a constant $C>0$ independent of $\alpha$ and $h_{_F}$ such that \ |c_{s,{_F}}|+|c_{e,{_F}}| \leq C\,h_{_F}^{-1/2} \left(\int_{_F} \large| j^m_{g,{_F}}\large|^2\,ds\right)^{1/2}. \ By using the above two inequalities, we have \[ \eta^{nd}_{m,{_F}} \leq C\, h_{_F}^{1/2}\sqrt{\alpha^-_{_F}} \left(\int_{_F} \large| j^m_{g,{_F}}\large|^2\,ds\right)^{1/2} \leq C\, \eta_{m,{_F}}, \] which, together with (\ref{mixedge}), implies the validity of the first inequality in (\ref{m-eff}). Now, the second inequality in (\ref{m-eff}) and (\ref{m-eff-G}) are straightforward from the definitions and (\ref{mixedge}). \end{proof} \subsection{Nonconforming Elements} \begin{thm} Assume that the distribution of the coefficient $\alpha$ is quasi-monotone. Then the error estimators $\eta^{rh}_{nc}$ and $\eta^{bd}_{nc}$ satisfy the global reliability bounds: \begin{eqnarray}\label{nc-rel-1} && \|\alpha^{1/2} \nabla_h e_{nc}\|_{0,\O} \leq C\left(\eta^{bd}_{nc} +H_f\right)\\[2mm] \label{nc-rel-2} \mbox{and} && \|\alpha^{1/2} \nabla_h e_{nc}\|_{0,\O} \leq C\left(\eta^{rh}_{nc} +H_f\right). \end{eqnarray} \end{thm} \begin{proof} Let $\hat{\eta}_{nc}$ be the implicit recovery-based estimator introduced in \cite{CaZh:10a}: \[ \hat{\eta}_{nc}^2= c\, \hat{\eta}_{nc,1}^2+ (1-c)\hat{\eta}_{nc,2}^2 \] with $c\in (0,1)$ being a parameter to be determined, where \[ \hat{\eta}_{nc,1}=\min_{\mbox{\boldmath$\tau$} \in B\!D\!M} \|\alpha^{-1/2}\mbox{\boldmath$\tau$} + \alpha^{1/2}\nabla_h u_{nc}\|_{0,\O} \quad\mbox{and}\quad \hat{\eta}_{nc,2}=\min_{\mbox{\boldmath$\tau$} \in N\!D} \|\alpha^{1/2}(\mbox{\boldmath$\tau$} - \nabla_h u_{nc})\|_{0,\O}. \] It is obvious that $\hat{\eta}_{nc,1}\leq \eta^{bdm}_{nc} \leq \eta^{rt}_{nc}$ and that $\hat{\eta}_{nc,2}\leq \eta^{nd}_{nc} \leq \eta^{ne}_{nc}$. Now, (\ref{nc-rel-1}) and (\ref{nc-rel-2}) follow from Theorem 6.4 of \cite{CaZh:10a}. \end{proof} To prove the efficiency of the explicit error estimators, consider the weighted edge error estimator introduced in \cite{CaZh:10a}: \[ \eta_{nc,{\cal E}} := \left( \sum_{F\in{\cal E}} \eta_{nc,{_F}}^2 \right)^{1/2} \mbox{with}\quad \eta_{nc,{_F}}^2 = \left\{ \begin{array}{lll} \displaystyle\frac{2h_{_F}^2}{\alpha_K^+ + \alpha_K^-} \left(j^{nc}_{f,{_F}}\right)^2 + \displaystyle\frac{h_{_F}^2\alpha_K^+ \alpha_K^-}{\alpha_K^+ + \alpha_K^-} \left(j^{nc}_{g,{_F}}\right)^2, & F\in{\cal E}_{_I}, \\[4mm] \displaystyle\frac{h_{_F}^2}{ \alpha_K^-} (j^{nc}_{f,{_F}})^2, & F\in{\cal E}_{_N}, \\[4mm] h_{_F}^2 \alpha_K^- \left(j^{nc}_{g,{_F}}\right)^2, & F\in{\cal E}_{_D}. \end{array} \right. \] \begin{lem}\label{lem-nc1} There exist a positive constant $C$ independent of $\alpha$ and the mesh size such that \begin{equation}\label{edge-nc} \eta^{rt}_{nc,{_F}}\leq C\eta_{nc,{_F}} \quad \mbox{and}\quad \eta^{ne}_{nc,{_F}}\leq C\eta_{nc,{_F}} \end{equation} \end{lem} \begin{proof} Without loss of generality, we prove the validity of the lemma only for interior edges. Assume that the triangulation is regular, then $\|\mbox{\boldmath$\phi$}^{rt}_{_F}\|_{0,K}\leq C\,h_{_F}$. It follows from the definition of $\eta^{rt}_{nc,{_F}}$, (\ref{sdncrtf}), and the equivalence (\ref{c-equi}) that \begin{eqnarray*} \eta^{rt}_{nc,1,{_F}} &=& \|\alpha^{-1/2} \bsigma^{\Delta}_{nc,rt,{_F}}\|_{0,\o_{_F}} =\left( \|\alpha_{F^-}^{-1/2}\bsigma^{\Delta}_{nc,rt,{_F}}\|_{0,K^-_{_F}}^2 + \|\alpha_{F^+}^{-1/2}\bsigma^{\Delta}_{nc,rt,{_F}}\|_{0,K^+_{_F}}^2\right)^{1/2} \\[2mm] &\leq& C\, h_{_F} j_{f,{_F}}^{nc} \left( \displaystyle\frac{1}{\alpha_{_F}^-} \left(\displaystyle\frac{\alpha_{_F}^-}{\alpha_{_F}^- + \alpha_{_F}^+} \right)^2 + \displaystyle\frac{1}{\alpha_{_F}^+} \left(\displaystyle\frac{\alpha_{_F}^+}{\alpha_{_F}^- + \alpha_{_F}^+} \right)^2 \right)^{1/2} \\[2mm] &=& C\, \displaystyle\frac{h_{_F} j_{f,{_F}}^{nc}}{\sqrt{\alpha_{_F}^- + \alpha_{_F}^+}} \leq C \eta_{nc,{_F}}, \end{eqnarray*} which implies the first inequality in (\ref{edge-nc}). To prove the second inequality in (\ref{edge-nc}), for any $F\in{\cal E}_{_I}$, introduce \[ \bf{\rho}^{nc}_{j,{_F}}=\left\{\begin{array}{ll} j^{nc}_{g,{_F}} h_{_F}\mbox{\boldmath$\psi$}^{ne,-}_{_F}, & \mbox{on }\, K^-_{_F},\\[3mm] 0, & \mbox{on }\, K^+_{_F}. \end{array}\right. \] Without loss of generality, we assume that $\alpha_{_F}^- \leq \alpha_{_F}^+ $. (Otherwise, $\bf{\rho}^{nc}_{j,{_F}}$ may be redefined by exchanging $K^-_{_F}$ and $K^+_{_F}$.) Since $j^{nc}_{g,{_F}}$ is a constant on $F$ and $\|\mbox{\boldmath$\psi$}^{ne}_{_F}\|_K \approx C\,h_{_F}$, by the definitions of $\brho^{\Delta}_{nc,ne,{_F}}$ in (\ref{local-rho-nc}), we have \begin{eqnarray*} \eta^{ne}_{nc,{_F}} &=& \|\alpha^{1/2}\brho^{\Delta}_{nc,ne,{_F}}\|_{0,\o_{_F}} \leq \|\alpha^{1/2}\bf{\rho}^{nc}_{j,{_F}}\|_{0,\o_{_F}} = \sqrt{\alpha_{_F}^-}\, \|\bf{\rho}^{nc}_{j,{_F}}\|_{0,K_{_F}^-}\\[2mm] &= & \sqrt{\alpha_{_F}^-}\, j^{nc}_{g,{_F}} \|\mbox{\boldmath$\psi$}^{ne}_{_F}\|_{0,K_{_F}^-} \leq C \sqrt{\alpha_{_F}^-}\, j^{nc}_{g,{_F}} h_{_F} \\[2mm] & \leq & C \left(\displaystyle\frac{\alpha_K^+ \alpha_K^-}{\alpha_K^+ + \alpha_K^-} \right)^{1/2} j^{nc}_{g,{_F}} h_{_F} \leq C \eta_{nc,{_F}}. \end{eqnarray*} This completes the proof of the second inequality in (\ref{edge-nc}) and, hence, the lemma. \end{proof} \begin{thm} The local indicators $\eta^{rh}_{nc,{_F}}$, $\eta^{rh}_{nc,{_K}}$, $\eta^{bd}_{nc,{_F}}$, and $\eta^{bh}_{nc,{_K}}$ are efficient, i.e., there exists a constant $C>0$ independent of $\alpha$ and the mesh size such that \begin{equation} \label{nc-eff-F} \eta^{bd}_{nc,{_F}} \leq \eta^{rh}_{nc,{_F}} \leq C\, \|\alpha^{1/2} \nabla_h e_{nc}\|_{0,\o_{_F}} +C\left( \sum_{K\in{\cal T}_{_F}} H_{f,{_K}}^2\right)^{1/2} \end{equation} and that \begin{equation}\label{nc-eff-K} \eta^{rh}_{nc,{_K}},\,\, \eta^{bd}_{nc,{_K}} \leq C\, \|\alpha^{1/2} \nabla_h e_{nc}\|_{0,\o_{_K}}+ C\left( \sum_{T\in{\cal T}_{_K}} H_{f,{_T}}^2\right)^{1/2}. \end{equation} \end{thm} \begin{proof} Let \[ {\cal V}^{nc}_{-1,{_F}} = \{ \mbox{\boldmath$\tau$}\in\! L^2(\omega_{_F}\!) \big|\, \mbox{\boldmath$\tau$}|_{_K}\!\!\in\!{\cal V}(K)\,\,\forall\, K\in {\cal T}_{_F}, \, \jump{\mbox{\boldmath$\tau$}\cdot{\bf n}_{_F}}_{_F}\!\! =\! -j^{nc}_{f,{_F}},\, \, \mbox{\boldmath$\tau$}\cdot{\bf n}_{_E}\! = 0\,\, \mbox{on}\, \,E \in\! {\cal E}_{b, {_F}}\!\} \] with ${\cal V}=RT\,\mbox{or}\, B\!D\!M$ and let \[ {\cal W}^{nc}_{-1,{_F}} = \{ \mbox{\boldmath$\tau$}\in\! L^2(\omega_{_F}\!) \big|\, \mbox{\boldmath$\tau$}|_{_K}\!\!\in\!{\cal W}(K)\,\,\forall\, K\in {\cal T}_{_F}, \, \jump{\mbox{\boldmath$\tau$}\cdot{\bf t}_{_F}}_{_F}\!\! =\! -j^{nc}_{g,{_F}},\, \, \mbox{\boldmath$\tau$}\cdot{\bf t}_{_E}\! = 0\,\, \mbox{on} \,\,E \in\! {\cal E}_{b, {_F}}\!\}. \] with ${\cal W}=N\!E\,\mbox{or}\, N\!D$. Similar to Section 4.1, the approximation error fluxes $\bsigma^{\Delta}_{nc,v,{_F}}$ with $v= rt\mbox{ or } bdm$ and the approximation error gradients $\brho^{\Delta}_{nc,w,{_F}}$ with $w=ne\mbox{ or }nd$ are then the solutions of the minimization problems: \begin{eqnarray}\label{mini-flux-nc-1} && \|A^{-1/2} \bsigma^{\Delta}_{nc, v,{_F}} \|_{0,\o_{_F}} =\min_{\mbox{\boldmath$\tau$}\in{\cal V}^{nc}_{-1,{_F}}}\|A^{-1/2} \mbox{\boldmath$\tau$}\|_{0,\o_{_F}} \\[2mm] \label{mini-flux-nc-2} \,\mbox{ and } && \|A^{1/2} \brho^{\Delta}_{nc,w,{_F}} \|_{0,\o_{_F}} =\min_{\mbox{\boldmath$\tau$}\in{\cal W}^{nc}_{-1,{_F}}}\|A^{1/2} \mbox{\boldmath$\tau$}\|_{0,\o_{_F}}, \end{eqnarray} respectively. Since $RT^{nc}_{-1,F}\subset B\!D\!M^{nc}_{-1,F}$ and $N\!E^{nc}_{-1,F}\subset N\!D^{nc}_{-1,F}$, the first inequality in (\ref{nc-eff-F}) follows from their definitions. The second inequality in (\ref{nc-eff-F}) is from the minimization problems in (\ref{mini-flux-nc-1}) and (\ref{mini-flux-nc-2}), Lemma \ref{lem-nc1}, and Theorem 6.8 of \cite{CaZh:10a}. The bounds in (\ref{nc-eff-K}) are straightforward from their definitions and inequality (\ref{nc-eff-F}). \end{proof} \section{Numerical Experiments} \setcounter{equation}{0} In this section, we report some numerical results for an interface problem with intersecting interfaces used by many authors, e.g., \cite{Kim:07,CaZh:09,CaZh:10a,CaZh:10b}, which is considered as a benchmark test problem. For simplicity, we only test the conforming element with explicit RT recovery. Other cases behave similarly. Let $\O=(-1,1)^2$ and \[ u(r,\theta)=r^{\gamma}\mu(\theta) \] in the polar coordinates at the origin with $\mu(\theta)$ being a smooth function of $\theta$ \cite{CaZh:09}. The function $u(r,\theta)$ satisfies the interface equation with $A= \alpha I$, $\Gamma_N=\emptyset$, $f=0$, and \[ \alpha(x)=\left\{\begin{array}{ll} R & \quad\mbox{in }\, (0,1)^2\cup (-1,0)^2,\\[2mm] 1 & \quad\mbox{in }\,\O\setminus ([0,1]^2\cup [-1, 0]^2). \end{array}\right. \] The $\gamma$ depends on the size of the jump. In our test problem, $\gamma=0.1$ is chosen and is corresponding to $R\approx 161.4476387975881$. Note that the solution $u(r,\theta)$ is only in $H^{1+\gamma-\epsilon}(\O)$ for any $\epsilon>0$ and, hence, it is very singular for small $\gamma$ at the origin. This suggests that refinement is centered around the origin. \begin{figure}[!hts] \hfill \begin{minipage}[!hbp]{0.48\linewidth} \centering \includegraphics[width=0.99\textwidth,angle=0]{mesh_exp} \caption{mesh generated by $\eta$}% \end{minipage}% \quad \begin{minipage}[!htbp]{0.48\linewidth} \centering \includegraphics[width=0.99\textwidth,angle=0]{error_exp} \caption{error and estimator $\eta$}% \end{minipage}% \hfill \end{figure} Mesh generated by $\eta_c^{rt}$ is shown in Figure 1. The refinement is centered at origin. Similar meshes for this test problem generated by other error estimators can be found in \cite{CaZh:09,CaZh:10a,CaZh:11}. The comparison of the error and the $\eta_c^{rt}$ is shown in Figure 2. The effectivity index is close to $1$. Moreover, the slope of the log(dof)- log(relative error) for $\eta_c^{rt}$ is $-1/2$, which indicates the optimal decay of the error with respect to the number of unknowns.

https://arxiv.org/abs/1202.1331

Integer Subsets with High Volume and Low Perimeter

We consider a certain variation of the 'isoperimetric problem' adopted for subsets of nonnegative integers. More specifically, we explore the sequence P(n) as described in OEIS A186053. We provide the first exact formulas for P(n) including multiple recursive relations involving auxiliary functions as well as concise and satisfying representations and even quasi-explicit formulas. We also discuss some of the intricate fractal-like symmetry of the sequence as well as the development of algorithms for computing P(n). We conclude with open questions for further research.Note this is a more developed, but more concise version of a previous arXiv paperarXiv:1107.2954by the name "Sets with High Volume and Low Perimeter".

\section{Introduction}\label{section introduction} One of the most widely-known classical geometry problems is the so-called \textit{isoperimetric problem}, one equivalent variation of which is: \begin{quote} If a figure in the plane has area $A$, what is the smallest possible value for its perimeter? \end{quote} In the Euclidean plane, the optimal configuration is a circle, implying that any figure with area $A$ has perimeter at least $2 \sqrt{A\pi}$, and this lower bound may be obtained if and only if the figure is a circle. \paragraph*{}In 2011, Miller et al.~\cite{Miller} extended the isoperimetric problem in a new direction, in which integer subsets took the role of geometric figures. For any integer subset $A$, they defined its \textit{volume} as the sum over all its elements, and they defined its \textit{perimeter} as the sum of all elements $x \in A$ such that $\{x-1, x+1\} \not \subset A$. Thus, the volume can be thought of as the sum of all the elements of $A$, and the perimeter can be thought of as the sum of all the elements on the ``boundary" of $A$ (that is to say, the elements of $A$ whose successor and predecessor are not both in $A$). \paragraph*{}The main focus of \cite{Miller} was to examine the relationships between a set's perimeter and its volume. More specifically, the authors wanted to answer the corresponding ``isoperimetric question"\footnote{They focused on this question in particular because it turns out that all of the related extremal questions are trivial.}: \begin{quote} If a subset of $\{0, 1, \ldots \}$ has volume $n$, what is the smallest possible value for its perimeter? \end{quote} Adopting their notation, we will let $P(n)$ denote this value through the duration of this paper\footnote{This is sequence A186053 in OEIS.}. \paragraph*{}Because their work is so recent, Miller et al. are the only ones who have published on this variation of the isoperimetric problem or on the function $P(n)$. Their work was to provide bounds for $P(n)$, by which they were able to determine its asymptotic behavior. Specifically, their main result was \begin{theorem}\label{theirResult} \emph{(Miller et al., 2011)} Let $P(n)$ be as defined. Then $P(n) \thicksim \sqrt{2}n^{1/2}$. Moreover, for all $n \geq 1$, \begin{equation}\label{theirBounds} \sqrt{2}n^{1/2} - 1/2 < P(n) < \sqrt{2} n^{1/2} + (2 n^{1/4} + 8) \log _2 \log _2 n + 58. \end{equation} \end{theorem} \paragraph*{}Their proof of the lower bound will be reproduced in following sections. However, their proof of the upper bound was found by a construction argument, which we will not reproduce here since we will analytically derive a tighter bound in \textbf{Theorem \ref{myBounds}}. \paragraph*{}Beyond the inequalities in \eqref{theirBounds} provided by Miller et al., nothing else has been published on $P(n)$ except for some values for small $n$. It should be noted that \cite{Miller} provides very good bounds on a related function, in which the sets of interest are allowed to have both negative as well as positive elements. However, this result was also obtained by a construction argument, and it is not relevant to this paper. \subsection*{Outline of Results} In this paper, we focus on improving the few results known on $P(n)$, including deriving multiple exact formulas and developing an understanding of its interesting long-term behavior. Many of these results are stated in terms of an intimately related function, $Q(n)$, which may be briefly defined as \[ Q(n) := \min_{A \subseteq \{0, 1, \ldots\}} \Big \{ per(A^{c}) : vol(A) = n \Big \}. \] Since it proves to be so closely related to $P(n)$, we also provide results on $Q(n)$ throughout the paper. \paragraph*{}We begin in \secref{\ref{section preliminary}} with several prelimary lemmas including those used in \cite{Miller}. Then in \secref{\ref{section first recurrences}}, we define auxilary functions, with which we combinatorially derive several recursive formulas for $P(n)$. We then introduce the function $Q(n)$ and derive similar recursive formulas for it as well. \paragraph*{}In \secref{\ref{section second recurrences}}, we relate the functions $P(n)$ and $Q(n)$ by providing yet more recurrence relations for both of them, from which we see that each function completely determines the other. With this in place, we move on to \secref{\ref{section analysis of recurrences}}, in which we use these recurrences to determine several analytic results for $P(n)$ and $Q(n)$, including upper and lower bounds and derivations of their asymptotic behavior. \paragraph*{}Our work then culminates in \secref{\ref{section good recurrences}}, in which we state and prove the strongest results of the paper. By appealing to our analytic bounds on $P(n)$ and $Q(n)$, we show that for all sufficiently large values of $n$, the recurrences of \secref{\ref{section second recurrences}} admit certain drastic simplifications. By then combining this result with rigorous computer calculations, we arrive at the main theorem of the paper\footnote{More adequate introductions of the functions $f$ and $g$ are given in \secref{\ref{section analysis of recurrences}}.}: \begin{reptheorem}{bestResult} Let $P(n)$ and $Q(n)$ be as given. Then if $n \geq 0$ is not one of the $177$ known counterexamples tabulated in \textbf{Table \ref{table:exceptions}} of the appendix (in particular, for all $n > 149,894$), we have \begin{eqnarray*} P(n) &=& f(n) + Q(g(n)) \qquad \qquad \text{and}\\ Q(n) &=& 1 + f(n) + P(g(n)), \end{eqnarray*} where the functions $f(n)$ and $g(n)$, given by \[ f(n) := \left \lfloor \sqrt{2n} + 1/2 \right \rfloor = \left [ \sqrt{2n}\right ], \qquad \qquad \text{and} \qquad \qquad g(n) := \dfrac{f(n) [ f(n)+1]}{2} - n, \] are the smallest nonnegative integers satisfying $[1+2+3+ \cdots + f(n)] - g(n) = n$. \end{reptheorem} With this, we derive several other satisfying and revealing reccurence relations and quasi-explicit representations for $P(n)$ and $Q(n)$. We also breifly demonstate and discuss the intricate fractal-like symmetry of the graphs of these functions. We then conclude in \secref{\ref{section conclusion}} by noting applications in the design of algorithms related to this problem and with some open questions for future research. \paragraph*{}For an earlier version of this paper with somewhat more detailed proofs and expositions, see \cite{arXiv}. \section{Definitions and Notation}\label{section definitions} For the reader's possible convenience, a brief list of definitions used throughout the paper is given here. In each definition, $A$ is assumed to be a subset of $\{0, 1, 2, \ldots\}$, and $n$ and $k$ are assumed to be nonnegative integers. \begin{itemize} \item The \textit{boundary} of $A$, $\partial A$, is $\partial A := \{z \in A : \{ z-1, z+1\} \not \subseteq A\}.$ In words, it is the set of elements of $A$ whose successor or predecessor is not in $A$. \item The \textit{volume} and \textit{perimeter} of $A$ are defined as \[ vol(A) := \sum_{z \in A} z, \qquad \text{and} \qquad per(A) := \sum_{z \in \partial A} z, \] respectively. (For convention, the volume and perimeter of the empty set is 0.) \item $P(n) := \min_{A \subseteq \{0, 1, \ldots\}} \Big \{ per(A) : vol(A) = n \Big \}.$ \item The \textit{complement} of $A$ is $A^{c} := \{0, 1, \ldots \} \setminus A = \{z \in \{0, 1, \ldots \} : z \notin A\}.$ \item $Q(n) := \min_{A \subseteq \{0, 1, \ldots\}} \Big \{ per(A^{c}) : vol(A) = n \Big \}.$ \item The helper functions $p(n;k)$ and $q(n;k)$ are defined as \[ p(n;k) := \min_{A \subseteq \{0, 1, \ldots , k\}} \Big \{per(A) : vol(A) = n \Big \}, \quad \qquad q(n;k) := \min_{A \subseteq \{0, 1, \ldots , k\}} \Big \{per(A^{c}) : vol(A) = n \Big \}. \] \item The special helper function $\sigma (n;k)$ is $\sigma (n;k) := \min_{A \subseteq \{0, 1, \ldots, k\}} \Big \{per(A^c) : vol(A) = n, \quad \text{and} \quad k \in A \Big \}.$ \item The functions $f(n)$ and $g(n)$ are given by \[ f(n) = \left [\sqrt{2n} \right], \qquad \text{and} \qquad g(n) = \dfrac{f(n) [f(n) + 1]}{2} - n = \dfrac{\left [\sqrt{2n} \ \right] ^2 + \left [\sqrt{2n}\ \right]}{2} - n, \] where $[x]$ denotes the nearest integer function. In \textbf{Proposition \ref{fReps}}, we show that $f(n)$ and $g(n)$ are also the smallest nonnegative integers satisfying $[1+2+\cdots + f(n)] - g(n) = n$. \item For all $N$ (and particularly, for $N = 149,894$), we define $\phi (n; N) = \phi (n):=\min_{i \geq 0} \{g^{i} (n) \leq N\}$. \end{itemize} \section{Preliminary Results}\label{section preliminary} The following lemma is used throughout \cite{Miller} and is essential in proving their lower bound on $P(n)$. \begin{lemma}\label{basicInequality} \emph{(Miller et al., 2011)} Assume $A$ is a finite nonempty subset of $\{0, 1, \ldots \}$, and let $m$ denote its maximum element. Then \[ m \leq per(A) \leq vol(A) \leq \dfrac{m(m+1)}{2}. \] \end{lemma} Using this lemma, the following lower bound is immediately attained. \begin{proposition}\label{lowerBoundForP} Assume $A \subseteq \{0, 1, \ldots \}$ is finite. Then we have \[ \sqrt{2vol(A)} -1/2 \leq \dfrac{-1 + \sqrt{1+8 vol(A)}}{2} \leq per(A). \] Moreover, for any positive integer $n$, this implies \[ \sqrt{2} n^{1/2} - 1/2 \leq P(n). \] \end{proposition} As stated before, except for the previously mentioned constructive upper bound on $P(n)$, these two results are all that has been published about $P(n)$. The remainder of the paper is devoted to new results. \subsection*{Miscellaneous Lemmas} \begin{lemma} Let $A \neq \emptyset$ be a finite subset of $\{0, 1, \ldots \}$, and let $m$ denote its maximum element. Then \[ m+1 \leq per(A^c) \] with equality if and only if $\{1, \ldots, m\} \subseteq A$. \end{lemma} \begin{proof} Let $A$ be as given. Then $m \in A$, but we know $m+1 \notin A$. Therefore, $m+1 \in \partial A^c$ implying that $m+1 \leq per(A^c)$. Now since $m+1 \in \partial A^c$, we know that $m+1 = per(A^c)$ if and only if $\partial A^c$ is equal to either $\{m+1\}$ or $\{0, m+1\}$. But this happens if and only if $\{1, 2, \ldots , m\} \subseteq A$, as desired. \end{proof} \begin{proposition}\label{lowerBoundForQ} Assume $A \subseteq \{0, 1, \ldots \}$ is finite. Then we have \[ \sqrt{2vol(A)} + 1/2 \leq \dfrac{-1 + \sqrt{1+8 vol(A)}}{2} + 1 \leq per(A^c). \] Moreover, for any positive integer $n$, this implies \[ \sqrt{2} n^{1/2} + 1/2 \leq Q(n). \] \end{proposition} \begin{proof} This follows from the previous lemma in the same way as \textbf{Proposition \ref{lowerBoundForP}}. \end{proof} \section{Recurrence Relations using Auxilary Functions} \label{section first recurrences} We now derive our first set of recurrence relations for $P(n)$ and $Q(n)$. Although the relations derived in \secref{\ref{section second recurrences}} are actually more revealing, the relations presented here follow naturally, and they motivate the introduction of important auxilary functions. Moreover, because of their convenient structure, these relations are used extensively in the design of algorithms for computing values, as we breifly discuss in \secref{\ref{section conclusion}}. \subsection*{First Recurrence for $P(n)$} As is often the case in analyzing discrete functions, we may obtain an exact recurrence relation for $P(n)$ in terms of a related auxillary function. In our case, recall that $P(n)$ is the minimum perimeter among all subsets of $\{0, 1, \ldots \}$ having volume $n$. This suggests defining an auxilary function, $p (n;k)$, as \[ p (n; k) = \min_{A \subseteq \{0, 1, \ldots , k\}} \Big \{ per(A) : vol(A) = n \Big \}. \] Then for all $n\geq 0$, we have \[ P(n) = \min_{k \in \{0, 1, \ldots \}} \Bigg \{ \min_{A \subseteq \{0, 1, \ldots , k\}} \Big \{ per(A) : vol(A) = n \Big \} \Bigg \} = \min_{k \in \{0, 1, \ldots \}} \Bigg \{ p (n;k) \Bigg \}. \] \paragraph*{}From its definition, it is clear that for all fixed $n$, the function $p (n; k)$ is monotonically decreasing with $k$. Moreover, for all $K \geq n$, we have $p (n; K) = p (n; n)$ since any subset of $\{0, 1, \ldots \}$ having volume $n$ must necessarily be a subset of $\{0, 1, 2, \ldots , n\}$. Therefore the above equation simplifies to \begin{equation}\label{PRecHelp} P(n) = \min_{k \in \{0, 1, \ldots \}} \Bigg \{ p (n;k) \Bigg \} = \lim _{k \to \infty} p (n; k) = p (n; n). \end{equation} Thus, we now seek a recurrence for $p (n; k)$, which will provide us with $P(n)$ by calculating $p (n; n)$. \paragraph*{}For notational convenience, let $S(n; k)$ denote the set of all subsets of $\{0, 1, \ldots , k \}$ having volume $n$. Then to obtain our desired recurrence for $p (n;k)$, we will consider the following paritition of $S(n;k)$ \[ S(n;k) = \bigcup _{l=0} ^{k+1} \Big \{A \in S(n;k) : \{l, \ldots , k\} \subseteq A \quad \text{and} \quad l-1 \notin A \Big \}. \] From this partition, it follows that \begin{equation}\label{pHelpRec1} p (n;k) = \min_{l \in \{0, 1, \ldots , k+1\}} \Bigg \{ \min_{A \in S(n;k)} \Big\{ per(A) : \{l, \ldots , k\} \subseteq A \quad \text{and} \quad l-1 \notin A \Big \} \Bigg \}. \end{equation} \paragraph*{}Now let $0 \leq l \leq k+1$ be fixed. Then we have \begin{eqnarray*} & & \min_{A \in S(n;k)} \Big\{ per(A) : \{l, \ldots , k\} \subseteq A \quad \text{and} \quad l-1 \notin A \Big \}\\ & & \qquad = \min_{B \subseteq \{0,1, \ldots, l-2\}} \Big\{ per(B \cup \{l, l+1, \ldots, k\}) : vol(B \cup \{l, l+1, \ldots, k\})=n \Big \}\\ & & \qquad = \begin{cases}\displaystyle \min_{B \in \{0,1, \ldots, k-1\}} \Big\{ per(B) : vol(B)=n \Big \} \quad &\text{if $l = k+1$},\\ \displaystyle \min_{B \in \{0,1, \ldots, k-2\}} \Big\{ k + per(B) : vol(B)=n-k \Big \} \quad &\text{if $l = k$},\\ \displaystyle \min_{B \in \{0,1, \ldots, l-2\}} \Big\{k + l + per(B) : vol(B)=n - \left [k(k+1)/2 - l(l-1)/2 \right] \Big \} \quad &\text{if $0 \leq l < k$},\end{cases}\\ & & \qquad = \begin{cases}p(n;k-1) \quad &\text{if $l = k+1$},\\ k+ p (n-k; k-2) \quad &\text{if $l = k$},\\ k + l + p \big (n - [k(k+1)-l(l-1)]/2 ; l-2 \big ) \quad &\text{if $0 \leq l < k$}.\end{cases} \end{eqnarray*} Therefore, by substituting into \eqref{pHelpRec1}, we are able to obtain the recurrence \begin{eqnarray} p (n;k) &=& \min \Bigg \{ p(n;k-1), k + p(n-k;k-2),\nonumber\\ & & \qquad \qquad k+\min_{l \in \{0, \ldots, k-1\}} \Big\{l + p \big (n - [k(k+1)-l(l-1)]/2 ; l-2 \big ) \Big \} \Bigg \}\label{pHelpRec}, \end{eqnarray} which is valid for all $n \geq 1$ and for all $k \geq 1$. Moreover, as boundary conditions, which are clear from its definition, we have that $p (n;k)$ satisfies \[ p(n;k) = \begin{cases}0 \qquad &\text{if $n=0$,}\\ \infty \qquad &\text{if $n < 0$ or $k \leq 0 < n$.}\end{cases} \] Thus, this recurrence for $p (n;k)$ gives the following compact recursive representation for $P(n)$ for all $n \geq 0$: \begin{equation} P(n) = \min \Big \{p(n; n-1), n \Big \}. \end{equation} \subsection*{Introduction of $Q (n)$ and Derivation of First Recurrences} Because of its intimate connections with the function $P(n)$ that will be explored in subsequent sections, we now introduce the function $Q (n)$, which is defined as \[ Q (n) = \min_{A \subseteq \{0, 1, \ldots\}} \Big \{ per(A^c) : vol(A)=n \Big \}. \] The difference between this function and the function $P(n)$ is subtle, and based on how similarly the two functions are defined, one would expect their behavior to be very close. As we will see, this is indeed the case, and the connections between $P(n)$ and $Q(n)$ are actually of fundamental importance. However, it is important for the reader to keep in mind the difference in how these functions are defined. \paragraph*{}As with the function $P(n)$, we define the auxilary function $q (n;k)$ as \[ q (n;k) = \min_{A \subseteq \{0, 1, \ldots, k\}} \Big \{ per(A^c): vol(A)=n \Big \}, \] and just as before, for all $n\geq 0$, we have that \begin{equation}\label{QRecHelp1} Q (n) = q (n;n). \end{equation} \paragraph*{}Because of the difference between how the functions $P(n)$ and $Q (n)$ are defined, we now need to define a special auxilary function, $\sigma (n;k)$, in order to obtain a compact recurrence for $q (n)$. This function is defined by \[ \sigma (n;k) = \min_{A \subseteq \{0, 1, \ldots, k\}} \Big \{ per(A^c): vol(A)=n \quad \text{and} \quad k \in A \Big \}. \] Note the similarities between $\sigma (n;k)$ and $q (n;k)$. In fact, it is clear that for all $n \geq 1$ and $k \geq 0$, we have \begin{equation}\label{qRec} q (n; k) = \min_{l \in \{1, 2, \ldots , k\}} \Big \{ \sigma (n;l) \Big \}. \end{equation} Using this equation and \eqref{QRecHelp1}, we obtain that for all $n \geq 1$ \begin{equation}\label{QRecHelp2} Q (n) = \min_{l \in \{1, 2, \ldots , n\}} \Big \{ \sigma (n;l) \Big \}, \end{equation} with $Q (0) = 0$. \paragraph*{}Just as was the case for $P(n)$, in order to obtain a useful recurrence relation for $Q(n)$, it now only remains to find a recurrence for $\sigma (n;k)$. As before, we accomplish this by a simple partition yielding \[ \sigma (n;k) = k+1 + \min \Big \{\sigma (n-k; k-1) - k, \sigma (n-k; k-2), k-1 + q (n-k; k-3) \Big \}, \] which we obtain by partitioning the subsets of interest into the three groups (I) sets containing $k-1$, (II) sets containing $k-2$ but not $k-1$, and (III) sets containing neither $k-2$ nor $k-1$. \paragraph*{}At this point, we need to note that some care must be given to the interpretation of the above equation, which depends on how we define $\sigma (0;0)$. However, if we note and state as a boundary condition that $\sigma(n,n) = 2n$ for all $n \geq 1$, then these concerns are effectively removed. \paragraph*{}We then have a recurrence relation for $\sigma$. As boundary conditions for $\sigma (n; k)$, we have \[ \sigma (n;k) = \begin{cases}0 \qquad &\text{if $n=k=0$,}\\ 2n \qquad &\text{if $n=k \geq 1$,}\\ \infty \qquad &\text{if $n < 0$ or if $k \in \{0, 1\}$ and $n > k$,}\\ \infty \qquad &\text{if $0 \leq k > n \geq 0$.}\end{cases} \] Then for all $n \geq 2$, and $2 \leq k < n$, we have \[ \sigma (n;k) = k+1 + \min \Big \{k-1 + q (n-k; t-3), \sigma (n-k; k-2), \sigma (n-k; k-1) - k \Big \}. \] Thus, by using \eqref{QRecHelp2} we have a recurrence for $Q(n)$ as well. \section{More Direct Recurrence Relations}\label{section second recurrences} Now by making use of different partitions of the sets of interest, we derive the following recurrence relations, from which we see the first connections between the functions $P(n)$ and $Q(n)$. \subsection*{Recurrence for $P(n)$ involving $q (n;k)$ and $\sigma(n;k)$} We may calculate $P(n)$ by a ``more direct" recurrence relation, which is found by partitioning all sets of volume $n$ first according to their maximum element, $m$, and then according to the largest integer smaller than $m$ not contained in each set. \paragraph*{}Let $A$ be a set of volume $n$, let $m$ be its maximum element, and let $l$ be the largest element of $\{-1, 0, \ldots, m\}$ not contained in $A$. Then $A$ may be written uniquely as $A = \{0, 1, 2, \ldots, m\} \setminus B$ for some set $B \subseteq \{0, 1, \ldots , l\}$, where the volume of $B$ is equal to $(1 + 2 + \cdots + m)-n$ and $l \in B$. If $l=m-1$, then $per(A) = per(B^c)$. Else, we have $per(A) = m + per(B^c)$. \paragraph*{}From this observation, we obtain that for all $n \geq 2$ \begin{equation}\label{PGoodRec} P(n) = \min_{m \geq 1} \Big \{m + q ([1 + 2 + \cdots + m] - n; m-2), \sigma ([1 + 2 + \cdots + m] - n; m-1) \Big \}, \end{equation} where $q (n;k)$ and $\sigma (n; k)$ are defined as earlier. \subsection*{Recurrence for $Q(n)$ involving $p (n;k)$} As before, we also have a simple recurrence that can be used to calculate $Q (n)$ ``more directly". Let $A$ be a set of volume $n$ and maximum element $m$. Then the set $A$ may be written uniquely in the form $A = \{0, 1, 2, \ldots, m\} \setminus B$ for some set $B \subseteq \{0, 1, \ldots , m-1\}$, where the volume of $B$ is equal to $(1 + 2 + \cdots + m) - n$. Now we know that for all such sets $A$ and $B$, we have $per(A^c) = per(B) + (m+1)$. \paragraph*{}This observation leads to the simple and beautiful recurrence that for all $n\geq 2$, \begin{equation}\label{QGoodRec} Q (n) = 1 + \min_{m \geq 1} \Big \{m + p ([1 + 2 + \cdots + m] - n; m-1) \Big \}, \end{equation} where $p (n; k)$ is as defined earlier. \section{Analysis of Recurrences}\label{section analysis of recurrences} Although equations \eqref{PGoodRec} and \eqref{QGoodRec} appear somewhat intractible (and they offer little or no computational advantage over the first recurrences of \secref{\ref{section first recurrences}}), they turn out to be crucial in understanding the behavior of $P(n)$ (and of $Q(n)$ as well). In \secref{\ref{section good recurrences}}, we are able to greatly simplify these recurrence, but in order to do so, we must first derive some analytic bounds on $P(n)$ and $Q(n)$. \subsection*{Relevant Lemmas and Notions} \begin{lemma} Let $n$ be a positive integer. Then there exist unique nonnegative integers $f(n)$ and $g(n)$ satisfying \[ n = [0 + 1 + \cdots + f(n)] - g(n), \] where $0 \leq g(n) < f(n)$. Moreover, $f(n)$ and $g(n)$ are given by\footnote{We will use these explicit functional representations for $f(n)$ and $g(n)$ so that $f(0) = g(0) = 0$ is well-defined.} \[ f(n) = \left \lceil \dfrac{-1 + \sqrt{1+8n}}{2} \right \rceil, \qquad \text{and} \qquad g(n) = \dfrac{f(n) [ f(n) + 1]}{2} - n. \] \end{lemma} \paragraph*{}Having defined these functions, we may now restate previous lemmas involving $P(n)$ and $Q(n)$ in these terms. The most important result we will use combines \textbf{Propositions \ref{lowerBoundForP}} and \textbf{\ref{lowerBoundForQ}} as follows: \begin{corollary} \label{crudeLowerBounds} Restating earlier results in new notation, for all $n \geq 1$, we have that \[ P(n) \geq f(n), \qquad \text{and} \qquad Q(n) \geq f(n) + 1. \] \end{corollary} \paragraph*{}Finally, before moving on, we must present two more results on the functions $f(x)$ and $g(x)$. \begin{proposition}\label{fReps} Let $f(n) = \left \lceil \dfrac{-1+\sqrt{1+8n}}{2} \right \rceil$ as before. Then for all integers $n \geq 0$, we have \[ f(n) = \left \lceil \dfrac{-1+\sqrt{1+8n}}{2} \right \rceil = \left \lceil \sqrt{2n} -1/2 \right \rceil = \left[ \sqrt{2n} \right], \] where $[x]$ is the nearest integer function. \end{proposition} \begin{proof} It suffices to show the first part of the stated equation holds, and the fact that $\sqrt{2n}$ is never a half-integer will complete the proof. Now by way of contradiction, suppose that the first two representations are not equal. Then this would imply that there exist integers $p \in \mathbb{Z}$ and $n \in \{0, 1, \ldots \}$ such that \[ \sqrt{2n} -1/2 \leq p < \dfrac{\sqrt{1+8n} - 1}{2}, \] which implies $8n \leq (2p +1)^2 < 8n+1.$ But since $n$ and $p$ are integers, this forces $8n = (2p +1)^2$, which taken modulo 2 yields a contradiction. \end{proof} \begin{proposition}\label{gBound} Let $f(n)$ and $g(n)$ be defined as before. Then for all integers $L \geq 0$ and $n \geq 0$, we have \[ g^{L} (n) \leq 2 \cdot (n/2) ^{1/2^{L}}. \] \end{proposition} \begin{proof} The proof is by induction on $L$. If $L=0$, then the claim is trivially true, which establishes the base case. Now suppose the claim holds for $L = m$. Then for all $n \geq 0$, we have \[ g(n) \leq f(n) - 1 < \sqrt{2n} - 1/2 < \sqrt{2n}, \] which implies $g^{m+1}(n) = g( g^{m} (n) ) < \sqrt{2 \cdot g^{m} (n)}.$ Then using the induction hypothesis and that the square root function is increasing completes the proof. \end{proof} \subsection*{Upper Bounds and Asymptotics for $P(n)$ and $Q (n)$} Using the recurrences of \secref{\ref{section second recurrences}}, we now obtain simple upper bounds on $P(n)$ and $Q(n)$, which taken with the last few lemmas, yield good absolute upper bounds in terms of $n$. \begin{theorem}\label{goodUpperBounds} Let $f(n)$ and $g(n)$ be defined as before. Then for all $n \geq 0$, we have the bounds \begin{eqnarray*} P(n) &\leq& f(n) + Q(g(n)), \qquad \text{and}\\ Q(n) &\leq& 1 + f(n) + P(g(n)). \end{eqnarray*} \end{theorem} \begin{proof} For $n=0$ and $n=1$, the two inequalities hold. Then for all $n \geq 2$, we may appeal to \eqref{PGoodRec} to obtain \begin{eqnarray*} P(n) &=& \min_{m \geq 1} \Big \{m + q ([1 + 2 + \cdots + m] - n; m-2), \sigma ([1 + 2 + \cdots + m] - n; m-1) \Big \}\\ &\leq & f(n) + \min \Big \{q (g(n); f(n)-2), \sigma (g(n); f(n)-1) \Big \} = f(n) + q(g(n); f(n)-1) = f(n) + Q(g(n)), \end{eqnarray*} and the corresponding inequality for $Q(n)$ is proven analogously. \end{proof} \begin{corollary} For all nonnegative integers $n$ and $L$, we have that \begin{eqnarray*} P(n) &\leq& L + P( g^{2L} (n)) + \sum_{i=0} ^{2L-1} f(g^{i}(n)), \qquad \text{and}\\ Q (n) &\leq & L + Q( g^{2L} (n)) + \sum_{i=0} ^{2L-1} f(g^{i}(n)), \end{eqnarray*} where $g^i (n)$ is the $i$-fold composition of $g$ evaluated at $n$, and by convention we take $g^0 (n) = n$. \end{corollary} \begin{theorem}\label{myBounds} Let $P(n)$ and $Q(n)$ be as given. Then $P(n) \sim Q(n) \sim \sqrt{2} n^{1/2}$. Moreover, for all $n > 2$, \begin{eqnarray*} \sqrt{2} n^{1/2} - 1/2 < &P(n)& \leq \sqrt{2}n^{1/2} + (2^{3/4} \cdot n^{1/4} + 1)[\log_2 ( \log_2 (n/2)) - 1] + 7, \qquad \text{and}\\ \sqrt{2} n^{1/2} + 1/2 < &Q(n)& \leq \sqrt{2}n^{1/2} + (2^{3/4} \cdot n^{1/4} + 1)[\log_2 ( \log_2 (n/2)) - 1] + 7. \end{eqnarray*} \end{theorem} \begin{proof} The lower bounds in the asserted inequalities have already been proven. To prove the upper bounds, we merely combine the results in the last corollary with the past few bounds on $f(n)$ and $g(n)$. More specifically, assuming $n > 2$, we know from \textbf{Proposition \ref{gBound}} that if $L \geq (\log_2 ( \log_2 (n/2)) - 1)/2$, then \[ g^{2L}(n) \leq 2 \cdot (n/2) ^{1/2^{(\log_2 ( \log_2 (n/2)) - 1)}} = \cdots = 8. \] By considering values of $P(n)$ and $Q(n)$ for $n \leq 8$, we see that $g^{2L}(n) \leq 8$ implies $P(g^{2L}(n)) \leq 7$ and $Q(g^{2L}(n)) \leq 7$. Now by the last corollary and the past few lemmas, we have \begin{eqnarray*} P(n) & \leq & L + P(g^{2L}(n)) + \sum_{i=0} ^{2L-1} f(g^{i}(n)) \leq L + P(g^{2L}(n)) + \sum_{i=0} ^{2L-1} \sqrt{2 g^{i}(n)} + 1/2\\ & \leq & 2L + P(g^{2L}(n)) + \sum_{i=0} ^{2L-1} \sqrt{4 \cdot (n/2) ^{1/2^{i}}} \leq 2L + P(g^{2L}(n)) + \sqrt{2n} + 2 \sum_{i=1} ^{2L-1} \sqrt{(n/2) ^{1/2^{i}}}\\ & \leq & 2L + P(g^{2L}(n)) + \sqrt{2n} + 4L(n/2) ^{1/4}. \end{eqnarray*} Then taking $L = (\log_2 ( \log_2 (n/2)) - 1)/2$ proves the bound. The inequality for $Q(n)$ is proven analogously. \end{proof} Note that these bounds on $P(n)$ are slightly better than those of \cite{Miller} stated in \textbf{Theorem \ref{theirResult}}. Also note that the upper bound on the summation is very crude. However, these bounds are sufficient for our purposes. \section{Obtaining \textit{Good} Recurrences for $P(n)$ and $Q(n)$} \label{section good recurrences} Although the bounds in \textbf{Theorem \ref{myBounds}} are rather good, they reveal nothing about the actual fluctuations of $P(n)$ and $Q(n)$. And although we have already obtained multiple recurrence relations for finding exact values, these relations all involve auxilary helper functions, multiple variables, and unweildy minimum functions. In this section, we combine our analytic bounds and combinatorial results to obtain surprisingly simple and satisfying recurrence relations for $P(n)$ and $Q(n)$ and even quasi-explicit formulae. \subsection*{New Lower Bounds on $P(n)$ and $Q(n)$} \begin{lemma}\label{infinityBound} Let $n$ and $k$ be positive integers with $k < f(n)$. Then $p(n;k),$ $q(n;k)$, and $\sigma(n;k)$ are all infinite. \end{lemma} \begin{proof} This follows from the fact that if $k < f(n)$, there are no subsets of $\{0, 1, \ldots, k\}$ with volume $n$. \end{proof} \begin{lemma}\label{helperBound} Let $n$ and $m$ be positive integers with $m > f(n)$. Then we have \begin{eqnarray*} m + p ([1 + 2 + \cdots + m] - n; m-1) &\geq& f(n) + \sqrt{2 (g(n) + f(n)+1)} + 1/2 \qquad \qquad \text{and}\\ m + q ([1 + 2+\cdots +m] - n; m-2) &\geq& f(n) + \sqrt{2 (g(n) + f(n)+1)} + 3/2. \end{eqnarray*} \end{lemma} \begin{proof} Consider the following chain of inequalities, which uses the simple lower bound in \textbf{Theorem \ref{myBounds}} \begin{eqnarray*} p ([1 + 2+\cdots +m] - n; m-1) &\geq& P ([1 + 2+\cdots +m] - n) \geq \sqrt{2 ([1 + 2+\cdots +m] - n)} - 1/2\\ &\geq& \sqrt{2 (g(n) + [f(n)+1] + [f(n)+2] + \cdots + m)} - 1/2\\ &\geq& \sqrt{2 (g(n) + f(n)+1)} - 1/2. \end{eqnarray*} Adding $m \geq f(n) +1$ to both sides proves the first inequality, and the second is proven in the same way. \end{proof} \begin{lemma}\label{piBound} Let $n$ and $m$ be positive integers with $m \geq f(n)$. Then we have \[ \sigma ([1 + 2 + \cdots + m] - n; m-1) \geq 2 f(n)-2. \] \end{lemma} \begin{proof} We may assume $f(n) \geq 2$, or the claim is trivially true. Let $A \subseteq \{0, 1, \ldots, m-1\}$ be such that $vol(A)=[1+2+\cdots + m] -n$ and $m-1 \in A$. By way of contradiction, suppose that $per(A^c) < 2f(n) -2$. \paragraph*{}If $m \geq 2f(n) - 2$, then since $m-1 \in \partial A$, this would imply that $per(A^c) \geq m \geq 2f(n) -2$. Therefore, we may assume that $m \leq 2f(n) - 3$. Now since $m \geq f(n)$, the volume of $A$ may be written as \[ vol(A) [1 + 2 + \cdots + m] - n = g(n) + [(f(n) + 1) + (f(n) + 2) + \cdots + m] < f(n) + [f(n) + 1] + \cdots + m, \] and because $m \leq 2f(n) -3 = [f(n) - 2] + [f(n) - 1]$, we also have \[ vol(A) < [f(n)] + [f(n) + 1] + \cdots + [m - 1] + [f(n)-2] + [f(n) -1] = \sum_{i=f(n)-2}^{m-1} i. \] \paragraph*{}From this, we know that there is at least one element of $\{f(n)-2, f(n)-1, \ldots , m-2\}$ that is not contained in $A$, because otherwise the volume of $A$ would be too large. Let $l \in A^c$ be the largest integer satisfying $f(n)-2 \leq l \leq m-2$. Then since $m-1 \in A$, we know that $l \in \partial A^c$, which implies \[ per(A^c) \geq l + m \geq f(n) -2 + m \geq f(n) - 2 + f(n) = 2f(n) -2. \] But this contradicts the assumption that $per(A^c) < 2f(n) -2$, thus completing the proof. \end{proof} \paragraph*{}With these lemmas, we are now able to prove the following lower bounds. \begin{theorem}\label{goodLowerBounds} Let $P(n)$ and $Q(n)$ be as given. Then for all $n \geq 2$, we have \begin{eqnarray*} P(n) &\geq & f(n) + \min \Big \{Q (g(n)), \sqrt{2 (g(n) + f(n)+1)} + 3/2, f(n)-2 \Big \} \qquad \qquad \text{and}\\ Q(n) &\geq & 1 + f(n) + \min \Big \{P (g(n)), \sqrt{2 (g(n) + f(n)+1)} + 1/2 \Big \}. \end{eqnarray*} \end{theorem} \begin{proof} Starting with \eqref{PGoodRec} and applying \textbf{Lemmas \ref{infinityBound}, \ref{helperBound},} and \textbf{\ref{piBound}}, we obtain \begin{eqnarray*} P(n) &=& \min_{m > f(n)} \Big \{f(n) + q(g(n); f(n)-2), m + q ([1 + 2+\cdots +m] - n; m-2),\\ & & \qquad \qquad \sigma(g(n); f(n)-1), \sigma ([1 + 2+\cdots +m] - n; m-1) \Big \}\\ &\geq& f(n) + \min \Big \{Q (g(n)), \sqrt{2 (g(n) + f(n)+1)} + 3/2, f(n)-2 \Big \}. \end{eqnarray*} The second inequality is proven analogously by starting with \eqref{QGoodRec}. \end{proof} \ignore{ Similarly, we also may obtain the following lower bound on $Q(n)$. \begin{proposition}\label{goodLowerBoundForQ} Let $n \geq 2$. Then $Q(n)$ satisfies \[ Q(n) \geq 1 + f(n) + \min \Big \{P (g(n)), \sqrt{2 (g(n) + f(n)+1)} + 1/2 \Big \}. \] \end{proposition} \begin{proof}Just as before, starting from \eqref{QGoodRec}, for all $n \geq 2$ we have \begin{eqnarray*} Q (n) &=& 1 + \min_{m \geq 1} \Big \{m + p ([1 + 2 + \cdots + m] - n; m-1) \Big \}\\ &=& 1 + \min_{m \geq f(n)} \Big \{m + p ([1 + 2 + \cdots + m] - n; m-1) \Big \}\\ &\geq & 1 + \min_{m > f(n)} \Big \{f(n) + p (g(n); f(n)-1), m + p ([1 + 2 + \cdots + m] - n; m-1) \Big \}\\ &\geq & 1 + f(n) + \min_{m > f(n)} \Big \{P(g(n)), \sqrt{2 (g(n) + f(n)+1)} + 1/2 \Big \}, \end{eqnarray*} as desired. \end{proof} } \subsection*{Squeezing an Equation from Inequalities (Eventually)} At this point, we have simple upper bounds on $P(n)$ and $Q(n)$ provided by \textbf{Theorem \ref{goodUpperBounds}} and nearly simple lower bounds from \textbf{Theorem \ref{goodLowerBounds}}, which are complicated by the ``min" operators. Suppose we could show that \textit{eventually} $P(g(n))$ and $Q(g(n))$ happen to be the smallest terms in each minimum. Then our lower bounds would simplify drastically and our lower and upper bounds would squeeze together, yielding a simple pair of mutually recursive equations that would hold for all sufficiently large $n$. \paragraph*{}As it turns out, we can in fact prove this claim, which is the content of the following proposition: \begin{proposition}\label{eventuallyHappens} Let $P(n)$ and $Q(n)$ be as given. Then there exists an $N \in \mathbb{Z}$ such that for all $n \geq N$ \begin{eqnarray*} P(g(n)) &=& \min \Big \{P (g(n)), \sqrt{2 (g(n) + f(n)+1)} + 1/2 \Big \} \qquad \qquad \text{and}\\ Q(g(n)) &=& \min \Big \{Q (g(n)), \sqrt{2 (g(n) + f(n)+1)} + 3/2, f(n)-2 \Big \}. \end{eqnarray*} Moreover, these claims hold if we take $N$ to be $2,500,000$. \end{proposition} \begin{proof} We will first prove there is such an $N \in \mathbb{Z}$. Then we will discuss why we may take $N$ to be $2,500,000$. \paragraph*{}We need to show that eventually $P(g(n)) \leq \sqrt{2 (g(n) + f(n)+1)} + 1/2$. From \textbf{Theorem \ref{myBounds}}, we know \[ P(r) \leq \sqrt{2r} + o(\sqrt{r}). \] Therefore, there exists a constant $G$ such that for all $r \geq G$, we have \[ P(r) \leq \sqrt{2r} + o(\sqrt{r}) \leq \sqrt{4r}. \] From this, it follows that for all $n$, if $g(n) \geq G$, then we have \[ P(g(n)) \leq \sqrt{4g(n)} \leq \sqrt{2(g(n) + f(n) + 1)} + 1/2. \] \paragraph*{}Let $M$ be the maximum value taken by $P(k)$ for $0 \leq k \leq G$, and let $n \geq M^2 (M^2+1)/2$ be arbitrary. Now if $g(n) \geq G$, then we know the claim holds. Therefore, we can assume $g(n) < G$. But if this is the case, then we know $P(g(n)) \leq M$, which implies \[ P(g(n)) \leq M \leq \sqrt{f(n)} \leq \sqrt{2(g(n) + f(n)+1)}+1/2. \] \paragraph*{}Therefore, for all $n \geq M^2 (M^2+1)/2 =:N_P$, the first equation holds. In the same way, we may find a constant $N_Q$ after which the second inequality holds. Thus, taking $N := \max \{N_P, N_Q\}$ proves the existence of such an integer $N$. \paragraph*{}Now proving that we may in fact take $N$ to be $2,500,000$, follows from somewhat lengthy but routine refinements of the previous argument. In the above notation, the main idea is to first obtain any analytic upper bound on $G$. This upper bound on $G$ is then refined by using computer calculated data to compare $P(r)$ with $\sqrt{4r}$ to make $G$ as small as possible. Using this technique for both $N_P$ and $N_Q$ then proves the claim. \end{proof} With this proposition, we are able to prove our main result. \begin{theorem}\label{bestResult} Let $P(n)$ and $Q(n)$ be as given. Then if $n \geq 0$ is not one of the $177$ known counterexamples tabulated in \textbf{Table \ref{table:exceptions}} of the appendix (in particular, for all $n > 149,894$), we have \begin{eqnarray*} P(n) &=& f(n) + Q(g(n)) \qquad \qquad \text{and}\\ Q(n) &=& 1 + f(n) + P(g(n)), \end{eqnarray*} where as before, the functions $f(n)$ and $g(n)$, given by \[ f(n) := \left \lfloor \sqrt{2n} + 1/2 \right \rfloor = \left [ \sqrt{2n}\right ], \qquad \qquad \text{and} \qquad \qquad g(n) := \dfrac{f(n) [ f(n)+1]}{2} - n, \] are also the smallest nonnegative integers satisfying $[1+2+3+ \cdots + f(n)] - g(n) = n$. \end{theorem} \begin{proof} If $n \geq 2,500,000$, then the result follows by using the previous proposition to simplify the lower bounds of \textbf{Theorem \ref{goodLowerBounds}} and comparing these to the upper bounds in \textbf{Theorem \ref{goodUpperBounds}}. \paragraph*{}On the other hand, if $0 \leq n < 2,500,000$, then the result holds by performing an exhaustive computer seach for counterexamples\footnote{A brief discussion of the algorithms used for this search is provided in \secref{\ref{section conclusion}}. Code is available on request.}. There are only $177$ counterexamples in this range, as tabulated in \textbf{Table \ref{table:exceptions}} of the appendix. In particular, if $n > 149,894$, then the claim holds since $149,894$ is the largest counterexample. \end{proof} \subsection*{Corollaries and Remarks} There are many interesting implications of \textbf{Theorem \ref{bestResult}}; from this result, many things can be discovered about the behavior of $P(n)$ and $Q(n)$, and the intimate connection between these two functions is made evident. Although these results can be formulated simply as algebraic statements about the recurrence relations, the corresponding geometric statements about the graphs of these functions is perhaps more enlightening. \begin{figure}[htb] \centering \includegraphics[width=\textwidth]{P_graph.pdf} \caption{Graph of $P(n)$ (\emph{red}) and $P(n) - f(n) = P(n) - [\sqrt{2n}]$ (\emph{brown})} \label{fig:p_and_p_minus_f} \end{figure} \paragraph*{}Examining \textbf{Figures \ref{fig:p_and_p_minus_f}} and \textbf{\ref{fig:q_and_q_minus_f_minus_1}} suggests several apparent patterns of the graphs of these functions. For example, we see that the graphs $P(n)$ and $Q(n)$ are each ``drifting" upwards by a translation of $f(n)$. After compensating for this drift, the patterns in the graphs become more apparent. \begin{figure}[htb] \centering \includegraphics[width=\textwidth]{Q_graph.pdf} \caption{Graph of $Q(n)$ (blue) and $Q(n) - f(n) - 1 = Q(n) - [\sqrt{2n}] - 1$ (green)} \label{fig:q_and_q_minus_f_minus_1} \end{figure} \paragraph*{}Now the curves $P(n) - f(n)$ and $Q(n) - f(n) - 1$ (shown in brown and green respectively) appear to be almost ``periodic" in a sense, with zeroes at $0, 1, 3, 6, 10, \ldots$. This apparent behavior is even more pronounced when the values of these functions are laid out in the following triangular array \[ \begin{tabular}{cccccc} \multicolumn{5}{c}{\text{$\{a_{n}\}_{n=0} ^{\infty}$}}\\ & & & & $a_0$\\ & & & & $a_1$\\ & & & $a_2$ & $a_3$\\ & & $a_4$ & $a_5$ & $a_6$\\ & $a_7$ & $a_8$ & $a_9$ & $a_{10}$\\ $a_{11}$ & $a_{12}$ & $a_{13}$ & $a_{14}$ & $a_{15}$\\ \vdots & \vdots & \vdots & \vdots & \vdots\\ \end{tabular}, \qquad \text{which yields for example} \qquad \begin{tabular}{cccccc} \multicolumn{5}{c}{\text{$\{(f(n), g(n))\}_{n=0} ^{\infty}$}}\\ & & & & $(0,0)$\\ & & & & $(1,0)$\\ & & & $(2,1)$ & $(2,0)$\\ & & $(3,2)$ & $(3,1)$ & $(3,0)$\\ & $(4,3)$ & $(4,2)$ & $(4,1)$ & $(4,0)$\\ $(5,4)$ & $(5,3)$ & $(5,2)$ & $(5,1)$ & $(5,0)$\\ \vdots & \vdots & \vdots & \vdots & \vdots\\ \end{tabular}. \] \paragraph*{}Then arranging values in this triangular manner, we have \[ \begin{tabular}{cccccccccc} \multicolumn{10}{c}{\text{$\{P(n) - f(n)\}_{n=0} ^{\infty}$}}\\ & & & & & & & & & 0\\ & & & & & & & & & 0\\ & & & & & & & & 0 & 0\\ & & & & & & & 1 & 2 & 0\\ & & & & & & 2 & 3 & 2 & 0\\ & & & & & 3 & 3 & 4 & 2 & 0\\ & & & & 4 & 5 & 3 & 4 & 2 & 0\\ & & & 4 & 5 & 6 & 3 & 4 & 2 & 0\\ & & 6 & 4 & 5 & 6 & 3 & 4 & 2 & 0\\ & 7 & 7 & 4 & 5 & 6 & 3 & 4 & 2 & 0\\ 6 & 7 & 7 & 4 & 5 & 6 & 3 & 4 & 2 & 0 \end{tabular}\qquad \qquad \qquad \begin{tabular}{cccccccccc} \multicolumn{10}{c}{\text{$\{Q(n) - f(n) - 1\}_{n=0} ^{\infty}$}}\\ & & & & & & & & & -1\\ & & & & & & & & & 0\\ & & & & & & & & 1 & 0\\ & & & & & & & 2 & 1 & 0\\ & & & & & & 2 & 2 & 1 & 0\\ & & & & & 4 & 2 & 2 & 1 & 0\\ & & & & 5 & 4 & 2 & 2 & 1 & 0\\ & & & 3 & 5 & 4 & 2 & 2 & 1 & 0\\ & & 6 & 3 & 5 & 4 & 2 & 2 & 1 & 0\\ & 7 & 6 & 3 & 5 & 4 & 2 & 2 & 1 & 0\\ 6 & 7 & 6 & 3 & 5 & 4 & 2 & 2 & 1 & 0 \end{tabular}. \] Then it appears that the rows (read from right to left) of $\{P(n) - f(n)\}$ `approach' $0, 2, 4, 3, 6, 5, 4, 7, 7, 6, \ldots$, and the rows of $\{Q(n) - f(n) - 1\}$ `approach' $0, 1, 2, 2, 4, 5, 3, 6, 7, 6, \ldots$. Moreover, these two sequences seem to be just $\{Q(n)\}$ and $\{P(n)\}$, respectively. In fact, this follows as our first corollary of \textbf{Theorem \ref{bestResult}}: \begin{corollary} Let $\{P(n) - f(n)\}_{n=0} ^{\infty}$ and $\{Q(n) - f(n) - 1\}_{n=0} ^{\infty}$ be arranged in the triangular manner previously discussed. Then unless $n$ is one of the 177 counterexamples in \textbf{Table \ref{table:exceptions}} of the appendix, reading the rows of $\{P(n) - f(n)\}$ from to right to left exactly agrees with $Q(t)$, and reading the rows of $\{Q(n) - f(n) -1\}$ exactly agrees with $P(t)$. \end{corollary} \begin{proof} This follows immediately from \textbf{Theorem \ref{bestResult}} by how the triangular array was constructed. \end{proof} \paragraph*{}Formulating this as a geometric statement is to say that except for 177 particular points, each ``lump" in the graphs of $P(n) - f(n)$ and $Q(n) - f(n) - 1$ is simply a reflection of a partial copy of $Q(n)$ or $P(n)$, respectively. Thus, the graph of $P(n)$ eventually consists solely of ``shifted" and reflected partial copies of $Q(n)$, and similarly the graph of $Q(n)$ eventually consists solely of ``shifted" and reflected partial copies of $P(n)$. This mutual similarity of the two functions also induces self-similarity as shown in the following results. \begin{corollary} If $g(n) < f(n) - 1$, and if $n$ and $n-f(n)$ are not one of the 177 values in \textbf{Table \ref{table:exceptions}}, \begin{eqnarray*} P(n) &=& 1 + P(n-f(n)) \qquad \qquad \text{and}\\ Q(n) &=& 1 + Q(n-f(n)). \end{eqnarray*} \end{corollary} \begin{proof} This follows from \textbf{Theorem \ref{bestResult}} and the fact that if $g(n) \neq f(n)-1$, then $g(n) = g(n-f(n))$. \end{proof} This corollary is the statement that with a finite number of exceptions, unless $n$ is one of the values at the far left of a row, then the value for $n$ in the triangle for $\{P(n)\}_{n=0} ^{\infty}$ (or in $\{Q(n)\}_{n=0} ^{\infty}$) is simply one more than the value directly above that entry in the triangle. \begin{corollary} If $n$ and $g(n)$ are not one of the 177 values listed in \textbf{Table \ref{table:exceptions}} of the appendix (and in particular, if $g(n) > 149,894$), then we have \begin{eqnarray*} P(n) &=& 1 + f(n) + f(g(n)) + P(g^2(n)) \qquad \qquad \text{and}\\ Q(n) &=& 1 + f(n) + f(g(n)) + Q(g^2(n)). \end{eqnarray*} \end{corollary} \begin{proof} This follows immediately by applying \textbf{Theorem \ref{bestResult}} twice. \end{proof} \paragraph*{}This last recurrence is readily `solved' yielding the following quasi-explicit equations. \begin{proposition}\label{almostExplicit} For all $n \geq 0$, let $\phi(n;149,894) = \phi(n)$ denote the smallest nonnegative integer satisfying $g^{\phi(n)}(n) \leq 149,894$. Then for all $n \geq 0$, we have \begin{eqnarray*} P(n) &=& \begin{cases} P(g^{\phi(n)}(n)) + \sum_{i=1}^{\phi(n)} f(g^{i-1}(n)) + \phi(n)/2 \qquad &\text{if $\phi(n)$ is even}\\ Q(g^{\phi(n)}(n)) + \sum_{i=1}^{\phi(n)} f(g^{i-1}(n)) +[\phi(n)-1]/2 \qquad &\text{if $\phi(n)$ is odd,}\end{cases} \qquad \qquad \text{and}\\ Q(n) &=& \begin{cases} Q(g^{\phi(n)}(n)) + \sum_{i=1}^{\phi(n)} f(g^{i-1}(n)) + \phi(n)/2 \qquad &\text{if $\phi(n)$ is even}\\ P(g^{\phi(n)}(n)) + \sum_{i=1}^{\phi(n)} f(g^{i-1}(n)) +[\phi(n)+1]/2 \qquad &\text{if $\phi(n)$ is odd.}\end{cases} \end{eqnarray*} \end{proposition} \begin{proof} This follows easily from the previous corollary. Although the function $\phi(n)$ is much too elusive for most honest mathematicians to call these equations truly ``explicit", they ought not be considered recursive. This is because even though $P$ and $Q$ are referenced on the right-hand side, their arguments are bounded; therefore, by appealing to \textbf{Table \ref{table:exceptions}}, those terms are effectively known. \end{proof} This gives rise to the following, perhaps surprising fact: \begin{corollary}\label{QPDiff} Let $P(n)$ and $Q(n)$ be as given. Then for all $n \geq 0$, we have \[ -1 \leq Q(n) - P(n) \leq 2. \] \end{corollary} \begin{proof} For all $n \geq 0$, we can appeal to \textbf{Proposition \ref{almostExplicit}} to obtain that \[ Q(n) - P(n) = \begin{cases} Q(g^{\phi(n)}(n)) - P(g^{\phi(n)}(n)), \qquad &\text{if $\phi(n)$ is even},\\ P(g^{\phi(n)}(n)) - Q(g^{\phi(n)}(n)) + 1, \qquad &\text{if $\phi(n)$ is odd.}\end{cases} \] Moreover, for our purposes, we can assume that $\phi(n)$ is one of the 177 counterexamples tabulated in \textbf{Table \ref{table:exceptions}} or else we could continue to appeal to \textbf{Theorem \ref{bestResult}} until this is the case. But looking at a table of these 177 values, we see that if $k$ is one of those exceptions, then $0 \leq Q(k) - P(k) \leq 2$, which completes the proof. \end{proof} \section{Conclusion} \label{section conclusion} We conclude by discussing applications for computing $P(n)$ and $Q(n)$ and by listing some open questions. \subsection*{``Sufficiently Large" and Computer Algorithms} In \textbf{Proposition \ref{eventuallyHappens}}, we state results that hold for all sufficiently large values of $n$ (in particular, for all $n \geq 2,500,000$). We then use this result to prove \textbf{Theorem \ref{bestResult}}, and we use a computer aided search to completely classify all counterexamples, which brings up a brief discussion of algorithms. \paragraph*{}The most na\"ive approach to compute $P(n)$ would be simply to list all sets of volume $n$ and find which has the smallest perimeter. This would require roughly $\bigO{2^n}$ time and $\bigO{n}$ memory, which is much too slow for large $n$, and a different approach is needed. \paragraph*{}Using the recurrence relations in \secref{\ref{section first recurrences}}, dynamic programming enables us to design algorithms for computing $P(n)$ and $Q(n)$ taking $\bigO{n^2 f(n)} = \bigO{n^{2.5}}$ time and using $\bigO{n^2}$ memory. We can reduce this memory requirement to roughly $\bigO{n}$ by employing a custom data structure, which benefits from the fact that for fixed $n$, functions such as $p(n;k)$ seem to take very few distinct values. Using these algorithms, the author was able to check all values of $P(n)$ and $Q(n)$ for $n \leq 3,500,000$, which is more than enough to obtain the results of \textbf{Theorem \ref{bestResult}}. \paragraph*{}Now that we have proven the recurrences in \textbf{Theorem \ref{bestResult}} and \textbf{Proposition \ref{almostExplicit}}, we may use these to compute $P(n)$ or $Q(n)$ in $\bigO{\Phi(n)} \leq \bigO{\log_{2} \log_{2} (n/2)}$ time using no additional memory. Moreover, we can compute a list of $P(0), P(1), \ldots , P(n)$ [or $Q(0), Q(1), \ldots , Q(n)$] in $\bigO{n}$ time using the required $\bigO{n}$ memory. \paragraph*{}Thus, one can now simply use \textbf{Theorem \ref{bestResult}} and the 177 values in \textbf{Table \ref{table:exceptions}} to compute $P(n)$ and $Q(n)$ extremely quickly, and $P(n)$ and $Q(n)$ can be tabulated essentially as far out as desired. The author is more than willing to provide anyone interested with code and calculated results. \subsection*{Open Questions} There are several possible areas of future research. Because the function $P(n)$ was first introduced so recently, this paper serves as a comprehensive overview of all that is known. \begin{itemize} \item[--] Little is known about the behavior of the functions $p(n;k)$, $q(n;k)$, and $\sigma(n;k)$. \item[--] It appears that for any fixed $n \leq 100,000$ the function $p(n;k)$ takes at most two finite values as $k$ varies. This may be interesting and might be proveable by focusing on \textbf{Proposition \ref{eventuallyHappens}}. \item[--] Very little or nothing whatsoever is known about $\phi(n;N)$ from \textbf{Proposition \ref{almostExplicit}}. \item[--] Characterizing sets for which $P(n)$ is obtained may be interesting. It seems likely that the partitions used and the code developed in this paper would help with that. Moreover, the result of \textbf{Theorem \ref{bestResult}} seems likely to help with this. \item[--] Providing more direct (i.e., less analytic) proofs for these results would likely be quite enlightening. \item[--] There seems to be no pattern or unifying properties for the 177 counterexamples tabulated in \textbf{Table \ref{table:exceptions}}. Alternate proofs of the main results may shed light on these seemingly sporadic values. \end{itemize}

https://arxiv.org/abs/2205.12246

Localized versions of extremal problems

We generalize several classical theorems in extremal combinatorics by replacing a global constraint with an inequality which holds for all objects in a given class. In particular we obtain generalizations of Turán's theorem, the Erdős-Gallai theorem, the LYM-inequality, the Erdős-Ko-Rado theorem and the Erdős-Szekeres theorem on sequences.

\section{Introduction} In this paper we will consider a systematic way to generalize some classical problems in extremal combinatorics. In an extremal problem one typically seeks to maximize the number of some object among a class of finite structures (graphs, hypergraphs, sequences etc.) avoiding some other object. We seek to reformulate such problems in the following more general setting. Rather than considering structures of some class forbidding a given object, we define a weight function which holds for arbitrary structures from that class. We then prove a bound on this weight function that immediately implies the original extremal theorem. We begin by discussing the case of graphs. For a graph $G$, we denote the set of vertices and edges of $G$ by $V(G)$ and $E(G)$, respectively. A clique with $r$ vertices is denoted by $K_r$. For a given graph $F$, the Tur\'an number ${\rm ex}(n,F)$ is the maximum number of edges possible in an $n$-vertex graph which does not contain $F$ as a subgraph. Tur\'an's theorem asserts that for all $r$ we have ${\rm ex}(n,K_{r+1}) \le \frac{(r-1)n^2}{2r}$, and equality holds when $r$ divides $n$ and we take a complete $r$-partite graph with classes of equal size. Now we attempt to generalize this result by considering an arbitrary $n$-vertex graph $G$ and looking at the behavior of the edges. Since for any $r$ we wish to recover the bound of $\frac{(r-1)n^2}{2r}$ in the case when no edge is in a copy of $K_{r+1}$, we define a weight on the edges based on the size of the cliques in which that edge occurs. To this end, let \[ c(e) = \max\{r:\mbox{$e$ occurs in a subgraph of $G$ isomorphic to $K_r$}\}. \] It is then natural to assign each edge a weight $\slfrac{1}{\frac{(c(e)-1)n^2}{2c(e)}}$ with denominator matching the corresponding extremal bound. We wish to prove that the sum of these weights is at most $1$. Moving the $n^2/2$ factor to the other side of the inequality we arrive at the following natural assertion. \begin{theorem} \label{localizedTuran} Let $G$ be any $n$-vertex graph, then \[ \sum_{e \in E(G)} \frac{c(e)}{c(e)-1} \le \frac{n^2}{2}, \] and equality holds if and only if $G$ is a multipartite graph with classes of equal size. \end{theorem} We prove Theorem~\ref{localizedTuran} in Section~\ref{graphs} using an inductive argument. Theorem~\ref{localizedTuran} was very recently proved independently by Domagoj Brada\u{c}~\cite{bradac} using the method of Motzkin and Straus~\cite{motzkin1965maxima}. Brada\u{c}~\cite{bradac} communicates that the statement of Theorem~\ref{localizedTuran} was proposed by Balogh and Lidick\'y and discussed at a conference in Oberwolfach in 2022. Moreover he mentions that stability results in the case of $K_5$-free graphs have been obtained by Balogh and Lidick\'y. We will refer to Theorem~\ref{localizedTuran} as a localized version of Tur\'an's theorem since the weight assigned to each edge depends only on the structures which it participates in. The bound from Tur\'an's theorem for any $r$ is easily recovered by noting that in a $K_{r+1}$-free graph each edge has weight at most $\frac{r}{r-1}$. Next we turn our attention to the case of paths. Erd\H{o}s and Gallai~\cite{gallai1959maximal} proved that an $n$-vertex graph without a copy of a path of length $k$ contains at most $\frac{(k-1)n}{2}$ edges. Let $P_k$ denote the path of length $k$ (that is, with $k$ edges). Again we introduce a function on the edges \[ p(e) = \max\{k:\mbox{$e$ occurs in a subgraph of $G$ isomorphic to $P_k$}\}. \] Then we have the following localized version of the theorem of Erd\H{o}s and Gallai. \begin{theorem} \label{localizedErdosGallai} Let $G$ be any $n$-vertex graph, then \[ \sum_{e\in E(G)} \frac{1}{p(e)} \le \frac{n}{2}, \] and equality holds if an only if every connected component of $G$ is a clique. \end{theorem} Again, the original theorem of Erd\H{o}s and Gallai is easily recovered by taking a $P_k$-free graph~$G$ and using that $p(e) \le k-1$. Also one can rather simply give a similar result for stars. For any graph $G$, let $s(e)$ be the maximum number of edges in a star containing $e$. \begin{proposition} \label{stars} Let $G$ be any $n$-vertex graph, then \[ \sum_{e\in E(G)} \frac{1}{s(e)} \le \frac{n}{2}, \] and equality holds if and only if each connected component of $G$ is regular. \end{proposition} Abstractly we are looking for theorems of the following form. Suppose we have an infinite sequence of graphs $F_1,F_2,\dots$ with the property that for all $i$, $F_i$ is a subgraph of $F_{i+1}$. Let $G$ be an arbitrary $n$-vertex graph, and for each edge $e \in E(G)$ set \[f(e) = \max\{i:\mbox{$e$ is an edge of a subgraph of $G$ isomorphic to $F_i$}\}.\] Then we would like to know when it holds that \begin{equation} \label{weighted} \sum_{e\in E(G)} \frac{1}{{\rm ex}(n,F_{f(e)+1})} \le 1. \end{equation} A bound of this form simultaneously generalizes each of the corresponding results for extremal numbers~${\rm ex}(n,F_i)$. Indeed, if $G$ does does not contain a particular $F=F_i$ as a subgraph, then by the monotonicity of the extremal function ${\rm ex}(n,F)$ with respect to taking subgraphs, we have \[ \frac{\abs{E(G)}}{{\rm ex}(n,F)} \le \sum_{e\in E(G)} \frac{1}{{\rm ex}(n,F_{f(e)+1})} \le 1, \] and consequently $\abs{E(G)} \le {\rm ex}(n,F)$. A stronger corollary of a bound of the form~\eqref{weighted} is the following. Suppose for an infinite sequence of graphs $F_1\subset F_2\subset\cdots$ we have a bound of the form~\eqref{weighted}. Fix $F=F_i$ for some $i$, and partition the edge set $E(G)$ into two parts $S$ and $T$ where $S = \{e\in E(G): \mbox{$e$ is an edge of a subgraph of $G$ isomorphic to $F$}\}$ and $T = E(G) \setminus S$, then \begin{equation} \label{borgtype} \frac{\abs{S}}{{\rm ex}(n,F)} + \frac{\abs{T}}{\binom{n}{2}} \le 1. \end{equation} The bound~\eqref{borgtype} follows by replacing ${\rm ex}(n,F_j)$ with ${\rm ex}(n,F_i)$ for all $j<i$ and replacing ${\rm ex}(n,F_j)$ with $\binom{n}{2}$ for all $j>i$. Note that Theorems~\ref{localizedTuran} and~\ref{localizedErdosGallai} are not precisely of the form~\eqref{weighted} since the corresponding extremal bounds which we generalize are only tight in certain divisibility cases. It would be interesting to extend these results so that the inequality is sharp in all divisibility cases. Note that inequalities of the form~\eqref{weighted} do not hold for any possible graph sequence. Indeed, if $F_1$ is a triangle, $F_2$ is a bow tie and $G$ is a balanced complete bipartite graph plus an edge, then~\eqref{weighted} is violated. In extremal set theory bounds of the form~\eqref{borgtype} sometimes have had applications for `cross'-versions~\cite{hilton1977intersection} (also termed `multicolor'-versions~\cite{bollobas2004multicoloured,keevash2004multicolour}) of various problems. We now turn our attention to finding localized versions of some theorems in extremal set theory. First we recall the classical LYM-inequality~\cite{bollobas1965generalized,lubell1966short,meshalkin1963generalization,yamamoto1954logarithmic} which generalizes Sperner's theorem~\cite{sperner1928satz} about antichains in the Boolean lattice. Let $\mathcal{A} \subset 2^{[n]}$ be a family of sets forming an antichain with respect to the subset relation, then we have \begin{equation} \label{LYM} \sum_{A\in\mathcal{A}} \frac{1}{\binom{n}{\abs{A}}}\le 1. \end{equation} Daykin and Frankl proved the following natural extension of the LYM-inequality. Suppose that $\mathcal{A}$ is any family of subsets of $[n]$ and let $\mathcal{S} \subset \mathcal{A}$ be the collection of those sets in $\mathcal{A}$ incomparable to every set in $\mathcal{A}$. Write $\mathcal{T} = \mathcal{A} \setminus \mathcal{S}$, then \begin{equation}\label{daykin} \sum_{A\in\mathcal{S}} \frac{1}{\binom{n}{\abs{A}}} + \frac{\abs{\mathcal{T}}}{2^n} \le 1. \end{equation} The inequality~\eqref{daykin} can be used to directly bound the sum of the sizes of a family of cross-Sperner families. A $(k+1)$-chain in $2^{[n]}$ is a collection of sets $A_1,A_2,\dots,A_{k+1}$ such that for all $i$ we have $A_i \subset A_{i+1}$. Another natural extension of the LYM-inequality is the following. Suppose $\mathcal{A} \subset 2^{[n]}$ contains no $(k+1)$-chain, then \begin{equation}\label{katona} \sum_{A\in\mathcal{A}} \frac{1}{\binom{n}{\abs{A}}}\le k. \end{equation} The inequality~\eqref{katona} can be proved by applying a Mirsky-type~\cite{mirsky1971dual} decomposition to the set family or directly by imitating Lubell's proof of the LYM-inequality~\cite{katona}. Our main result in this topic is a generalization of~\eqref{katona}. For an arbitrary $\mathcal{A} \subset 2^{[n]}$, let \[ c(A) = \max \{k:\mbox{$A$ participates in a $k$-chain consisting of sets from $\mathcal{A}$}\}. \] \begin{theorem}\label{localizedLYM} Let $\mathcal{A}\subset 2^{[n]}$ be an arbitrary family of sets, then \[ \sum_{A\in \mathcal{A}} \frac{1}{\binom{n}{\abs{A}}c(A)}\le 1. \] Moreover equality holds if and only if $\mathcal{A}$ consists of a union of complete levels. \end{theorem} A further extension of Theorem~\ref{localizedLYM} to the setting of posets is given in Section~\ref{posets}, Corollary~\ref{posetcA}. Observe that if $\mathcal{A}$ is a $(k+1)$-chain free family, then $c(A) \le k$ and we recover~\eqref{katona}. The inequality in Theorem~\ref{localizedLYM} and~\eqref{daykin} seem to be independent of each other. Next we consider intersecting set families. Recall that the Erd\H{o}s-Ko-Rado~\cite{erdos1961intersection} theorem asserts than if $r<n/2$, then a pairwise intersecting family consisting of sets of size $r$ has size at most~$\binom{n-1}{r-1}$. In order to find a shorter proof for a theorem of Hilton~\cite{hilton1977intersection}, Borg~\cite{borg2009short} proved the following generalization of the Erd\H{o}s-Ko-Rado theorem. Let $r<n/2$ and let $\mathcal{A}$ be an arbitrary family of $r$-element subsets of $[n]$. Let $\mathcal{S} = \{A\in\mathcal{A}:\mbox{for all $B\in\mathcal{A}$, $A$ and $B$ intersect}\}$ and $\mathcal{T}=\mathcal{A}\setminus \mathcal{S}$, then \begin{equation} \abs{\mathcal{S}} + \frac{r}{n}\abs{\mathcal{T}} \le \binom{n-1}{r-1}. \end{equation} We now give a localized version generalizing Borg's result. \begin{theorem} \label{m(A)} Let $\mathcal{A}$ be a collection of $r$-element subsets of $[n]$. Let $m$ be the function on $\mathcal{A}$ defined by setting $m(A)=n/r$, if $A$ is contained in some matching of $\floor{n/r}$ sets from $\mathcal{A}$, and setting $m(A)$ equal to the size of the largest matching in $\mathcal{A}$ containing $A$, if that matching has fewer than $\floor{n/r}$ sets, then \begin{displaymath} \sum_{A\in\mathcal{A}} \frac{1}{m(A)} \le \binom{n-1}{r-1}. \end{displaymath} \end{theorem} Note that there are at least three distinct constructions attaining equality in Theorem~\ref{m(A)}. In addition to a star or a full level we may take the construction consisting of all sets of size $r$ containing at least one of two fixed elements. Finally, we turn out attention to perfect graphs, posets and sequences. Recall that a perfect graph is one in which the chromatic number and clique number of all induced subgraphs are equal. In particular for a perfect graph $G$ we have $\omega(G)= \chi(G)$ and so $\alpha(G)\omega(G)=\alpha(G)\chi(G)\ge n$. In Section~\ref{posets} we prove the analogous localized version of this bound as a corollary of a more technical statement about functions on perfect graphs that may be of independent interest. Let $G$ be a perfect graph, and for each $v\in V(G)$ let $c(v)$ be the largest size of a clique containing~$v$, and let $i(v)$ be the largest size of an independent set containing $v$. \begin{theorem}\label{localizedperfect} For any $n$-vertex perfect graph $G$, we have \[ \sum_{v\in V(G)} \frac{1}{c(v)i(v)} \le 1. \] \end{theorem} As a consequence of Theorem~\ref{localizedperfect} we deduce a localized version of the classical theorem of Erd\H{o}s and Szekeres~\cite{erdos1935combinatorial} about sequences. \section{Proofs of localized versions of extremal graph theorems} \label{graphs} \begin{proof}[Proof of Theorem~\ref{localizedTuran}] We prove the claim by induction on $n$. The result holds trivially for $n=1$, so assume $n>1$ and let $G$ be an $n$-vertex graph with edge set $E$. If the graph contains no edges, then the bound follows trivially so assume $G$ has at least one edge. Let $k$ be the size of the largest clique in $G$, and let $C$ be a clique of size $k$. We split the edges $E$ into three parts: $E_{C}$, the edges within $C$; $E_{G\setminus C}$, the edges within $G\setminus C$; and $E_{S}$, the edges that connect $C$ and $G\setminus C$. We bound the contribution to the sum from each of these separately. \begin{enumerate} \item Since $C$ is a clique of maximum size in $G$, we know that $E_{C}$ contains $k(k-1)/2$ edges, all with $c(e)=k$. So we can see that \begin{equation*} \sum_{e \in E_{C}}\frac{c(e)}{c(e)-1} =\frac{k(k-1)}{2}\frac{k}{k-1} =\frac{k^2}{2}. \end{equation*} \item For all $v \in V(G \setminus C$), let $C_v=\{w \in C \mid \{v,w\}\in E\}$. Note that $C_v\cup\{v\}$ is itself a clique, and so we may conclude that $\abs{C_v}+1\le k$, and $c(e)\ge \abs{C_v}+1$ for all edges $e$ that connect $v$ to $C$. Thus, we obtain that \begin{equation*} \sum_{e \in E_{S}}\frac{c(e)}{c(e)-1} \le \sum_{v\in V(G \setminus C)}\abs{C_v}\frac{\abs{C_v}+1}{\abs{C_v}} \le \sum_{v\in V(G \setminus C)} k =(n-k)k. \end{equation*} \item Lastly, our induction hypothesis implies that \begin{equation*} \sum_{e \in E_{G \setminus C}}\frac{c(e)}{c(e)-1} \le \frac{(n-k)^2}{2}. \end{equation*} \end{enumerate} Combining all three of the estimates above we obtain that \begin{equation}\label{combine} \sum_{e \in E} \frac{c(e)}{c(e)-1} \le \frac{k^2}{2}+k(n-k)+\frac{(n-k)^2}{2} =\frac{n^2}{2}, \end{equation} as required. Now suppose we have equality in~\eqref{combine}, then all three of the estimates above must be tight. From the third estimate we obtain that $G\setminus C$ is a complete multipartite graph with $m$ equal size classes where $m\le k$. Moreover, for each edge $e$ in $E_{G\setminus C}$ we have that $m=c_{G\setminus C}(e) = c_G(e)$. By the second estimate each vertex from $G\setminus C$ must have exactly $k-1$ neighbors in $C$. Consequently if we take a vertex from each class in $G\setminus C$, we obtain that these vertices have a common neighborhood in $C$ of size at least $k-m$. Thus every edge in $E_{G\setminus C}$ is part of a $k$-clique in $G$ and so $m=c_{G}(e)=k$. Finally, any selection of $k$ vertices from distinct classes of $G\setminus C$ must have an empty common neighborhood in $C$ for otherwise we would have a larger clique. It follows that for every class in $G\setminus C$ there is exactly one vertex in $C$ to which all vertices in that class are nonadjacent, and this vertex is distinct for each class in $G\setminus C$. Adding each such vertex from $C$ to its respective class in $G\setminus C$, we see that $G$ is a balanced multipartite graph as required. \begin{comment} \begin{corollary} $\sum_{v \in V} d(v) \frac{c(v)}{c(v)-1} \le n^2$ \end{corollary} \begin{proof} \begin{align*} \sum_{v \in V} d(v) \frac{c(v)}{c(v)-1} &= \sum_{v \in V} \sum_{e \in E: v \in e} \frac{c(v)}{c(v)-1} \\ &= \sum_{e \in E} \sum_{v \in e} \frac{c(v)}{c(v)-1} \\ &= \sum_{uv=e \in E} \frac{c(u)}{c(u)-1} + \frac{c(v)}{c(v)-1} \\ &\le 2 \sum_{e \in E} \frac{c(e)}{c(e)-1} \\ &\le n^2. \end{align*} The first inequality follows from the fact that $v \in e$ implies $c(v) \ge c(e)$ and $x/(x-1)$ is a decreasing function for $x \ge 2$. The second inequality follows from Theorem \ref{thm:main}. \end{proof} \end{comment} \end{proof} Now we give the proof of the localized version of the Erd\H{o}s-Gallai theorem. \begin{proof}[Proof of Theorem~\ref{localizedErdosGallai}] We prove the claim by induction on $n$. The claim follows trivially for $n=1$, so assume $n>1$. Furthermore, assume that $G$ is connected; if $G$ is not connected, we may simply apply the induction hypothesis to each connected component in turn, and sum the results to arrive at our claim. Assume the longest path in $G$ has length $k$, and let $P$ be a path achieving this length, with endpoints $v$ and $w$. We use $R(v)$ to denote the edges that have $v$ as an endpoint; note that $R(v)$ cannot contain any edges to vertices outside of $P$, since otherwise $P$ is not maximal. We define $R(w)$ similarly, and note it has the same property. If there exists a circuit $C$ that includes all the vertices on $P$, then we can see that $P$ must include all the vertices in $G$. Otherwise, since $G$ is connected, we can find a path from $C$ to any vertex $u$ not lying on it; but the path that starts at $u$, goes to $C$, and then goes around $C$ is strictly longer that $P$, a contradiction. So in this case, we have that $k=n-1$. Furthermore, every edge in the graph either lies on the circuit $C$ or forms a chord in $C$. In either case, we can easily construct a path of length $n-1$ containing it. So we can see that \begin{equation}\label{all} \sum_{e \in E(G)} \frac{1}{p(e)} = \abs{E(G)}\frac{1}{n-1} \le \binom{n}{2}\frac{1}{n-1} = \frac{n}{2}, \end{equation} and we are done. Equality in the theorem implies~\eqref{all} is tight and so $G$ is a clique. Now we assume that there is no circuit $C$ that includes all of the vertices of $P$. Note that this means that $\{v,w\}\notin E(G)$. Furthermore, let $\{v',w'\}$ be any internal edge of $P$, where $v'$ is closer to $v$ along $P$ (and so $w'$ is closer to $w$ along $P$). Note that $P\setminus\{\{v',w'\}\}\cup\{\{v,w'\},\{v',w\}\}$ would be a circuit, so either $\{v,w'\}\notin E(G)$ or $\{v',w\}\notin E(G)$. Recalling that $R(v)$ and $R(w)$ contain no edges to vertices outside $P$, we may thus conclude that \begin{equation*} \abs{R(v)}+\abs{R(w)}\le 2 + (k-2) = k. \end{equation*} So without loss of generality, we assume that $\abs{R(v)}\le k/2$. Note that any $e\in R(v)$ has $p(e)=k$, since there is always some $e'\in P$ that can be replaced with $e$ to form a new path $P'$. Thus, if we apply our induction hypothesis to $G\setminus v$, we get that \begin{equation}\label{longinequality} \sum_{e \in E(G)}\frac{1}{p(e)} =\sum_{e \in R(v)}\frac{1}{p(e)}+\sum_{E(G)\setminus R(v)}\frac{1}{p(e)} \le \sum_{e \in R(v)}\frac{1}{p(e)}+\sum_{E(G)\setminus R(v)}\frac{1}{p_{G\setminus v}(e)} \le \frac{k}{2}\frac{1}{k}+\frac{n-1}{2} =\frac{n}{2}. \end{equation} Equality in the theorem statement implies that both inequalities in~\eqref{longinequality} hold with equality for all $e \in E(G)$. This means that for the edges $e$ in $E(G\setminus v)$ we have $p(e) = p_{G\setminus v}(e)$. Moreover by induction $G\setminus v$ must be a disjoint union of cliques. If $k=1$ the graph consists of just an edge and we are done, so assume $k>1$. Consider the edge $e$ incident to $v$ in $P$ and the next edge $f$ in the path. Since $f$ belongs to a connected component which is a clique in $G\setminus v$ and $f$ is adjacent to $e$, we can easily find a longer path in $G$ containing $f$ and so $p(e)>p_{G\setminus v}(e)$. Thus in the present case, equality cannot hold in~\eqref{longinequality}. \end{proof} We conclude this section with the proof of the analogous statement about stars. \begin{proof}[Proof of Proposition~\ref{stars}] For a vertex $v$ of $G$ set $w(v)$ equal to the sum of $\frac{1}{s(e)}$ across edges $e$ which are incident to $v$. Then for all $v$, \begin{equation} \label{stareq} w(v) \le \sum_{e:v\in e} \frac{1}{d(v)} = 1. \end{equation} Then we have \[ 2 \sum_{e\in E(G)} \frac{1}{s(e)} = \sum_v w(v) \le n, \] and the result follows. Suppose we have equality in~\eqref{stareq}, then for every $v$ every edge $e$ incident to $v$ must satisfy $s(e)=d(v)$. It is then immediate that each component is regular, as claimed. \end{proof} Given that we have localized versions of the extremal results of paths and stars it would be natural to investigate what happens in the case of an arbitrary sequence of trees: $T_1 \subset T_2 \subset\cdots$. However, such an investigation may be difficult since we only know the validity of the Erd\H{o}s-S\'os conjecture in certain cases. Nonetheless one could hope to prove localized versions conditional on the Erd\H{o}s-S\'os conjecture. \section{Proofs of localized versions of extremal hypergraph theorems} \label{hypergraphs} We begin by proving the localized version of the LYM-inequality. \begin{proof}[Proof of Theorem~\ref{localizedLYM}] Let $\mathcal{A} \subset 2^{[n]}$ be any family of sets. Call a chain of sets $\mathcal{C}$ maximal if it contains a set of every possible cardinality. We will double count pairs $(A,\mathcal{C})$ where $A \in \mathcal{A}$ and $\mathcal{C}$ is a maximal chain in $2^{[n]}$ using a weight function $w$. We define $w$ by \begin{displaymath} w(A,\mathcal{C}) = \begin{cases} \frac{1}{c(A)} & \mbox{if } A\in \mathcal{A} \mbox{ and } A \in \mathcal{C}, \\ 0 & \mbox{otherwise.} \end{cases} \end{displaymath} First fix a set $A \in \mathcal{A}$. There are $\abs{A}!(n-\abs{A})!$ maximal chains containing $A$ and so \begin{equation} \label{eq1} \sum_{A\in \mathcal{A}} \sum_{\mathcal{C}} w(A,\mathcal{C}) = \sum_{A\in \mathcal{A}}\frac{\abs{A}!(n-\abs{A})!}{c(A)}. \end{equation} Now, fix a maximal chain $\mathcal{C}$. Suppose $\mathcal{C}$ contains $m$ sets from $\mathcal{A}$; then we have $c(A) \ge m$ for all such sets, since these $m$ sets themself form a chain of length $m$. Thus, we have \begin{equation} \label{eq2} \sum_{\mathcal{C}} \sum_{A\in \mathcal{A}} w(A,\mathcal{C}) \le \sum_{\mathcal{C}} \frac{m}{m} = n!. \end{equation} Comparing $\eqref{eq1}$ and $\eqref{eq2}$ yields \begin{equation} \label{eq3} \sum_{A\in \mathcal{A}} \frac{1}{c(A) \binom{n}{\abs{A}}} \le 1, \end{equation} as desired. Suppose we have equality in $\eqref{eq3}$. Then the total weight on every chain must be $1$. Note that this trivially holds in the case where $\mathcal{A}$ equals a union of levels. We claim that this is the only case where equality holds. Suppose $\mathcal{A}$ is not a union of levels. Then there is at least one level of $2^{[n]}$ containing both a set from $\mathcal{A}$ and one not from $\mathcal{A}$. Moreover, there must be such a pair with symmetric difference $2$. Call these sets $A$ and $B$ where $A \in \mathcal{A}$ with $c(A)=m$ and $B \notin \mathcal{A}$. Since $\abs{A}=\abs{B}=t$ and $A$ and $B$ have symmetric difference $2$, it follows that the intersection, $A\cap B$, is of size $t-1$ and the union, $A\cup B$, is of size $t+1$. Consider any chain of the form $\mathcal{C}_1 = \{\varnothing,A_1,A_2,\dots,A_{t-2},A\cap B, A,A\cup B, A_{t+2},\dots,[n]\}$, then the total weight along $\mathcal{C}_1$ must be $1$ in the equality case and so $\mathcal{C}_1$ contains $m$ sets from $\mathcal{A}$ each with weight $1/m$. Now, consider the chain $\mathcal{C}_2 = \{\varnothing,A_1,A_2,\dots,A_{t-2},A\cap B, B,A\cup B, A_{t+2},\dots,[n]\}$. Since this chain contains the same sets as $\mathcal{C}_1$ except with $B \notin \mathcal{A}$ instead of $A \in \mathcal{A}$, the total weight of $\mathcal{C}_2$ is $1-1/m$, a contradiction. \end{proof} \ \begin{proof} We will use Katona's method of cyclic permutations~\cite{katona1972simple}. By a cyclic permutation, $\sigma$, we mean a cyclic ordering $a_1 < a_2 < \dots < a_n < a_1$ of $[n]$. A set $A \in \mathcal{A}$ is an interval in $\sigma$ if its elements are consecutive in $\sigma$. An interval starts at $a_i$ if $a_i$ is the smallest element in the order on $A$ induced by $\sigma$. We will double count pairs $(A,\sigma)$ where $A\in\mathcal{A}$ and $\sigma$ is a cyclic permutation of $[n]$ with the following weight function: \begin{displaymath} w(A,\sigma)= \begin{cases} \frac{1}{m(A)}, & \text{if } A\in\mathcal{A} \text{ and } A \text{ is an interval in } \sigma, \\ 0, &\text{otherwise.} \end{cases} \end{displaymath} Every set $A\in\mathcal{A}$ is an interval in exactly $r!(n-r)!$ cyclic permutations. Thus, we have \begin{displaymath} \sum_{A\in\mathcal{A}}\sum_{\sigma} w(A,\sigma) = \sum_{A\in\mathcal{A}} \frac{r!(n-r)!}{m(A)}. \end{displaymath} Now we switch the order of summation and fix a cyclic permutation $\sigma$. In Lemma~\ref{lemma:cyclemma}, whose proof is given below, we will show that the the total weight contributed by intervals along any cyclic permutation is at most $r$. It follows that \begin{displaymath} \sum_{\sigma}\sum_{A\in\mathcal{A}} w(A,\sigma) \le \sum_{\sigma}r = (n-1)!r. \end{displaymath} Dividing through by $(n-r)!r!$ gives the desired result. \end{proof} \begin{lemma} \label{lemma:cyclemma} Let $\sigma$ be a cyclic permutation and denote by $\mathcal{A}^\sigma$ the collection of those sets in $\mathcal{A}$ which are intervals in $\sigma$. Let $m(A)$ be defined as in the theorem, then \begin{displaymath} \sum_{A\in \mathcal{A}^\sigma} \frac{1}{m(A)} \le r. \end{displaymath} \end{lemma} \begin{proof} Let $n=tr+s$ where $t$ and $s$ are integers and $0 \le s < r$. Similarly, let $\abs{A^\sigma} = l r+k$ where $l$ and $k$ are integers and $0 \le k < r$. First, we show that every set in $\mathcal{A}^\sigma$ is contained in an $l$-matching. Notice that if an interval $A$ does not begin less than $r$ positions before or after an interval $B$, then $A$ and $B$ are disjoint. Consider the cyclic order attained by removing all of those elements from $\sigma$ at which no $A \in \mathcal{A}^\sigma$ begins. If two elements $a$ and $b$ are more than $r$ positions apart in this contracted order, then they must have been at least $r$ positions apart in the original order, and so the interval beginning at $a$ in $\sigma$ must be disjoint from the interval beginning at $b$. Now take an arbitrary element of the contracted order and move along the order $r$ elements at a time. After $l-1$ repetitions we are still a distance of at least $r$ from the starting element. The starting element, along with the elements encountered at each of the $l-1$ steps constitute an $l$-matching. Since our choice of starting element in the contracted order was arbitrary, it follows that every $A\in \mathcal{A}^\sigma$ is contained in an $l$-matching. Now, we distinguish two cases based on whether $l=t$ or $l<t$. If $l=t$, then since every $A\in\mathcal{A}^\sigma$ is contained in an $l$-matching we have, \begin{displaymath} \sum_{A\in\mathcal{A}^\sigma}\frac{1}{m(A)} = \frac{\abs{\mathcal{A}^\sigma}}{n/r} \le \frac{n}{n/r}=r. \end{displaymath} It remains to show the bound for the case $l<t$. Since $l<t$ there are at least $r-k$ elements of $\sigma$ at which no set in $\mathcal{A}^{\sigma}$ begins. Choose any $r-k$ such elements. We will refer to these elements as ``fake'' starting positions as we will imagine additional intervals beginning at them. The size of $\mathcal{A}^{\sigma}$ plus the number of fake starting positions is then $(l+1)r$. As before, we consider the contracted order containing only elements beginning sets in $\mathcal{A}^{\sigma}$ and the fake starting positions. Since $(l+1)r$ is a multiple of $r$, we may partition the elements of the resulting order into $r$ classes each containing $l+1$ elements such that in any class the distance between two elements is a multiple of $r$. There are only $r-k$ fake starting positions, so it follows that some $k$ of these classes consist of only elements corresponding to sets in $\mathcal{A}^{\sigma}$. For each set $A$ contained in one of the $k$ classes consisting only of sets in $\mathcal{A}^{\sigma}$, we know that $m(A) \ge l+1$ since $A$ is in a $l+1$-matching with the other sets in the class. It follows that the net weight contributed by such sets is at most \begin{equation} \label{eqn:partone} \frac{k(l+1)}{l+1}=k. \end{equation} Subtracting off the sets in $\mathcal{A}^{\sigma}$ which fill one of the classes and the fake starting positions leaves \begin{displaymath} r(l+1)-k(l+1)-(r-k)=l(r-k) \end{displaymath} sets. Since every set in $\mathcal{A}^{\sigma}$ is contained in matching of size at least $l$, we have that the weight contribution of the remaining sets is at most \begin{equation} \label{eqn:parttwo} \frac{l(r-k)}{l}=r-k. \end{equation} Thus, adding \eqref{eqn:partone} and \eqref{eqn:parttwo} we have that \begin{displaymath} \sum_{A\in\mathcal{A}^{\sigma}} \frac{1}{m(A)}\le k+(r-k)=r, \end{displaymath} as desired. \end{proof} \section{Results on perfect graphs, posets and sequences}\label{posets} \begin{theorem} \label{superlemma} Let $G$ be a perfect graph, and $f$ be a function on the set of pairs $(H, v)$ where $H$ is an induced subgraph of $G$ and $v\in V(H)$. Assume that $f$ satisfies $f(H,v) \le f(G[V(H)\cup\{w\}],v)$ for inducerd subgraphs $H$ of $G$ and $w\in V(G) \setminus V(H)$. Also assume that if $H$ is an independent set in $G$, then we have \begin{equation*} \sum_{v\in V(H)}\frac{1}{f(H,v)} \le 1. \end{equation*} For an induced subgraph $H$ of $G$, let $c(H, v)$ be the size of the largest clique in $H$ that contains $v$. Then, we have \begin{equation*} \sum_{v \in V(H)}\frac{1}{f(H, v)c(H, v)}\le 1. \end{equation*} \end{theorem} \begin{proof} We use induction on the size of the largest clique in $H$. If $H$ is an independent set, then we have $c(H, v)=1$ for all $v\in V(H)$ and the desired bound follows by assumption. Otherwise, let $k$ be the size of the largest clique in $H$, and assume we have the desired bound whenever the maximum clique has size strictly less than $k$. Define $B$ as the set of all vertices contained in a $k$-clique, that is $B=\{v\in V(H): c(H, v) = k\}$. Since $G$ is perfect, we may conclude that the chromatic number of $H$ is also $k$. Let $R$ be any $k$-coloring of $H$. We form the natural partition of $B$ based on $R$ as \begin{equation*} B_r=\{v \in B: v\text{ has color $r$ under }R\}. \end{equation*} Now, for any color $r\in R$, we know that every $k$-clique in $H$ must contain exactly one vertex with color $r$, and so the largest clique in $H \setminus B_r$ must have size exactly $k-1$. We apply induction to get that \begin{equation}\label{eq:pg_sumup} \sum_{v\in V(H\setminus B_r)}\frac{1}{f(H\setminus B_r, v)c(H \setminus B_r, v)} \le 1 \end{equation} for every $r\in R$. Now, for all $v \in H$ and all $r \in R$, we have by assumption that $f(H, v) \ge f(H \setminus B_r, v)$ and trivially that $c(H, v) \ge c(H\setminus B_r,v)$. Recall that $B_r$ contains exactly one vertex from each $k$-clique with vertices in $B$ (which is every $k$-clique in $H$). Thus, for any $v \in B \setminus B_r$ it is clear that $c(H\setminus B_r,v)=k-1$, and so for all $v \in B$ we have $c(H \setminus B_r, v) = \frac{k-1}{k}c(H, v)$. Summing Equation~\eqref{eq:pg_sumup} over all $r$ gives us \begin{align*} k &\ge \sum_{r \in R}\sum_{v\in V(H\setminus B_r)}\frac{1}{f(H\setminus B_r, v)c(H \setminus B_r, v)}\\ &=\sum_{r\in R}\sum_{v\in V(H\setminus B)}\frac{1}{f(H\setminus B_r, v)c(H \setminus B_r, v)}+\sum_{r\in R}\sum_{v\in V(G[B\setminus B_r])}\frac{1}{f(H\setminus B_r, v)c(H \setminus B_r, v)}\\ &\ge k \sum_{v\in V(H\setminus B)}\frac{1}{f(H, v)c(H, v)} + (k-1)\sum_{v\in V(G[B])}\frac{1}{f(H, v)\frac{k-1}{k}c(H, v)}\\ &= k \sum_{v\in V(H)}\frac{1}{f(H, v)c(H, v)}. \end{align*} Dividing both sides by $k$ yields the desired inequality. \end{proof} The following corollary is a natural generalization of Theorem~\ref{localizedLYM} to any poset satisfying the LYM-inequality. \begin{corollary} \label{posetcA} Let $P$ be a ranked poset with $N_i$ elements of rank $i$. For $x\in P$, let $\abs{x}$ denote the rank of $x$. Suppose that any antichain $S \subset P$ satisfies the LYM-type inequality: \begin{displaymath} \sum_{x\in S} \frac{1}{N_{\abs{x}}}\le 1. \end{displaymath} Then, given any set $T \subset P$ we have \begin{displaymath} \sum_{x \in T} \frac{1}{N_{\abs{x}} c(T,x)} \le 1. \end{displaymath} \end{corollary} \begin{proof} Let $G$ be the comparability graph of $P$. A set $S \subset P$ corresponds to an induced subgraph~$G_S$ of~$G$. Observe that $S$ is an antichain if and only if $G_S$ is an independent set. Moreover the chains in the poset $P$ are in one-to-one correspondence with the cliques in $G$. The assumption that an antichain $S$ in $P$ satisfies the LYM-inequality shows that the conditions of Theorem~\ref{superlemma} are satisfied when $f(G_S,x)$ is taken to be $N_{\abs{x}}$ for all $S$ and $x$. Then the conclusion of Corollary~\ref{posetcA} is immediate from the conclusion of Theorem~\ref{superlemma}. \end{proof} Now we deduce Theorem~\ref{localizedperfect} from Theorem~\ref{superlemma}. \begin{proof}[Proof of Theorem~\ref{localizedperfect}] Let $G$ be a perfect graph and for an induced subgraph $H$ and $x\in V(H)$ let $f(H,v)$ be the maximum size of an independent set in $H$ which contains $v$. It is easy to see $f$ satisfies the condition increasing condition. If $H$ is an independent set then $f(H,v)=\abs{V(H)}$ and the other condition of Theorem~\ref{superlemma} is satisfied. Then taking $G=H$, the conclusion of Theorem~\ref{superlemma} gives the required bound: \[ \sum_{v\in V(G)} \frac{1}{i(v)c(v)} = \sum_{v\in V(G)} \frac{1}{f(G,v)c(G,v)} \le 1.\qedhere \] \end{proof} Again by considering the comparability graph the following corollary is immediate from Theorem~\ref{localizedperfect}. For any poset $P$ and $x \in P$ let $C(x)$ denote the size of the largest chain $x$ belongs to and $A(x)$ the size of the largest antichain $x$ belongs to. \begin{corollary} \label{dilworth} Let $P$ be any poset we have, \begin{displaymath} \sum_{x\in P} \frac{1}{A(x)C(x)}\le 1. \end{displaymath} \end{corollary} By considering the natural poset defined on sequences of numbers $x_1,x_2,\dots,x_N$ where $x_i \prec x_j$ if $i<j$ and $x_i<x_j$ we obtain a localized version of the classical result of Erd\H{o}s and Szekeres~\cite{erdos1935combinatorial}. For a finite sequence of real numbers let $i(x)$ and $d(x)$ denote the longest increasing and decreasing subsequence containing $x$, respectively. Then we have the following. \begin{corollary} For any finite sequence $S$ of real numbers \[ \sum_{x\in S} \frac{1}{i(x)d(x)} \le 1. \] \end{corollary} \section{Acknowledgements} The second author would like to thank Rutger Campbell and Tuan Tran for insightful discussions about this topic and Nika Salia for providing the example with the triangle and bow tie in the introduction. The research of the second author was supported by NKFIH grant K135800.

https://arxiv.org/abs/0804.4464

Stabbing simplices by points and flats

The following result was proved by Barany in 1982: For every d >= 1 there exists c_d > 0 such that for every n-point set S in R^d there is a point p in R^d contained in at least c_d n^{d+1} - O(n^d) of the simplices spanned by S.We investigate the largest possible value of c_d. It was known that c_d <= 1/(2^d(d+1)!) (this estimate actually holds for every point set S). We construct sets showing that c_d <= (d+1)^{-(d+1)}, and we conjecture this estimate to be tight. The best known lower bound, due to Wagner, is c_d >= gamma_d := (d^2+1)/((d+1)!(d+1)^{d+1}); in his method, p can be chosen as any centerpoint of S. We construct n-point sets with a centerpoint that is contained in no more than gamma_d n^{d+1}+O(n^d) simplices spanned by S, thus showing that the approach using an arbitrary centerpoint cannot be further improved.We also prove that for every n-point set S in R^d there exists a (d-2)-flat that stabs at least c_{d,d-2} n^3 - O(n^2) of the triangles spanned by S, with c_{d,d-2}>=(1/24)(1- 1/(2d-1)^2). To this end, we establish an equipartition result of independent interest (generalizing planar results of Buck and Buck and of Ceder): Every mass distribution in R^d can be divided into 4d-2 equal parts by 2d-1 hyperplanes intersecting in a common (d-2)-flat.

\section{Introduction} Let $S$ be an $n$-point set in $\mathbb{R}^d$ in general position (no $d+1$ points lying on a common hyperplane). The points of $S$ span $n\choose d+1$ distinct $d$-dimensional simplices. The following interesting and useful result in discrete geometry (called the \emph{First Selection Lemma} in \cite{matou_book}), shows that at least a fixed fraction of these simplices have a point in common: \begin{theorem}[B\'ar\'any \cite{barany}]\label{t:1sl} For every $n$-point set $S$ in $\mathbb{R}^d$ in general position there exists a point $p\in \mathbb{R}^d$ that is contained in at least $c_d n^{d+1} - O(n^d)$ simplices spanned by $S$, where $c_d$ is a positive constant depending only on~$d$ (and the implicit constant in the $O(\,)$ notation may also depend on $d$). \end{theorem} In this paper we investigate the value of $c_d$. More precisely, from now on, let $c_d$ denote the supremum of the numbers such that the statement of Theorem~\ref{t:1sl} holds for all finite sets $S$ in $\mathbb{R}^d$. \paragraph{Lower bounds.} B\'ar\'any's proof yields \begin{equation*} c_d \ge {1\over d! (d+1)^{d+1}}. \end{equation*} Wagner \cite{wagner_thesis} improved this bound by roughly a factor of $d$, to \begin{equation}\label{eq_wagner_bd} c_d \ge {d^2+1 \over (d+1)! (d+1)^{d+1}}. \end{equation} For the special case $d=2$, Boros and F\"uredi \cite{boros_furedi_ptintriag} achieved the better lower bound of $c_2 = 1/27$ (also see Bukh \cite{bukh} for a simpler proof of this planar bound). \paragraph{Upper bounds.} The following result was proved by K\'arteszi \cite{karteszi} for $d=2$ (also see Moon \cite[p.~7]{moon} and Boros and F\"uredi \cite{bf_italian, boros_furedi_ptintriag}) and by B\'ar\'any \cite{barany} for general $d$: \begin{theorem}\label{thm_any_S} If $S$ is any $n$-point set in general position in $\mathbb{R}^d$, then no point $p\in \mathbb{R}^d$ is contained in more than \begin{equation*} \frac {1}{2^d (d+1)!}\, n^{d+1} + O(n^d) \end{equation*} $d$-simplices spanned by $S$. \end{theorem} It follows without difficulty from a result of Wendel (reproduced as Lemma~\ref{lemma_wendel} below) that this bound is asymptotically attained with high probability by points chosen uniformly at random from the unit sphere. Alternatively, as was kindly pointed out to us by Uli Wagner, the tightness also follows by considering the Gale transform of the polar of a cyclic polytope; see, e.g., Welzl \cite{Welzl-enteringleaving} for the relevant background. B\'ar\'any's bound implies that \begin{equation*} c_d \le {1\over 2^d(d+1)!}, \end{equation*} which, to our knowledge, was the best known upper bound on $c_d$ for all $d\ge 3$. For $d=2$, Boros and F\"uredi \cite{boros_furedi_ptintriag} claimed the upper bound $c_2 \le 1/27$ (which would be tight), but it turns out that the construction in their paper gives only $c_2 \le 1/27 + 1/729$ (see Appendix~\ref{app_BF_construction} of this paper). Our first result is an improved upper bound for $c_d$ for every $d$ (and the first ``non-trivial" one, in the sense that it refers to a specific construction): \begin{theorem}\label{momentcurvthm} For every fixed $d\ge 2$ and every $n$ there exists an $n$-point set $S\subset \mathbb{R}^d$ such that no point $p\in \mathbb{R}^d$ is contained in more than $(n/(d+1))^{d+1} + O(n^d)$ $d$-simplices spanned by $S$. Thus, \begin{equation}\label{eq_cd_upper_bd} c_d \le (d+1)^{-(d+1)}. \end{equation} Moreover, such an $S$ can be chosen in convex position. \end{theorem} In particular, the planar bound of $c_2 = 1/27$ is tight, after all. \subsection{The First Selection Lemma and centerpoints} If $S$ is an $n$-point set in $\mathbb{R}^d$ and $p\in \mathbb{R}^d$, we say that $p$ lies at \emph{depth $m$ with respect to $S$} if every halfspace that contains $p$ contains at least $m$ points of $S$. A classical result of Rado \cite{rado} states that there always exists a point at depth $n/(d+1)$. Such a point is called a \emph{centerpoint}. Wagner proved the bound (\ref{eq_wagner_bd}) by showing the following: \begin{theorem}[\cite{wagner_thesis}]\label{thm_wagner_alpha} If $S$ is an $n$-point set in $\mathbb{R}^d$ and $p\in\mathbb{R}^d$ is a point at depth $\alpha n$ with respect to $S$, then $p$ is contained in at least \begin{equation*} \bigl((d+1)\alpha^d-2d\alpha^{d+1}\bigr)\frac{n^{d+1}}{(d+1)!} - O(n^d) \end{equation*} $d$-simplices spanned by $S$.\footnote{This is obtained by setting $k=0$ in the lower bound for $f_k(\mu, \mathbf o)$ in the proof of Theorem~4.32 in \cite{wagner_thesis}.} \end{theorem} This, together with Rado's Centerpoint Theorem, immediately implies (\ref{eq_wagner_bd}). In this paper we show that Theorem~\ref{thm_wagner_alpha} cannot be improved: \begin{theorem}\label{wagnersharpthm} For every $\alpha$, $0 < \alpha \le 1/2$, and every $n$, there exists an $n$-point set $S$ in $\mathbb{R}^d$ such that the origin is at depth $\alpha n$ with respect to $S$ but is contained in only \begin{equation*} \bigl((d+1)\alpha^d - 2d\alpha^{d+1} \bigr)\frac{n^{d+1}}{(d+1)!} + O(n^d) \end{equation*} $d$-simplices spanned by $S$. \end{theorem} Thus, the approach of taking an arbitrary centerpoint cannot yield any lower bound better than (\ref{eq_wagner_bd}) for the First Selection Lemma. \subsection{Stabbing $(d-k)$-simplices by $k$-flats} The First Selection Lemma can be generalized as follows: If $S$ is an $n$-point set in $\mathbb{R}^d$ and $k$ is an integer, $0\le k < d$, then there exists a $k$-flat that intersects at least $c_{d,k} n^{d-k+1} - O(n^{d-k})$ of the $(d-k)$-simplices spanned by $S$, for some positive constants $c_{d,k}$ that depend only on $d$ and $k$. (By a \emph{$k$-flat} we mean a $k$-dimensional affine subspace of $\mathbb{R}^d$.) The problem is to determine the maximum values of the constants $c_{d,k}$. Trivially we have $c_{d,k} \ge c_{d-k}$: Simply project $S$ into an arbitrary generic subspace of dimension $d-k$, and then apply the First Selection Lemma. Here we derive a nontrivial lower bound for the case $k = d-2$ (this is the only case for which we could obtain good lower bounds): \begin{theorem}\label{thmsel} If $S$ is a $n$-point set in $\mathbb{R}^d$, then there is a $(d-2)$-flat $\ell$ that intersects at least $\frac1{24}(1-1/(2d-1)^2)n^3 - O(n^2)$ triangles spanned by~$S$. Thus, \begin{equation} c_{d,d-2} \ge \frac 1{24}\left(1-\frac 1{(2d-1)^2}\right). \label{eq_bd_cdd2} \end{equation} \end{theorem} For $d=2$ this is just the planar version of First Selection Lemma with the optimal constant of $1/27$. And as $d$ increases, the right-hand-side of (\ref{eq_bd_cdd2}) increases strictly with $d$, approaching $1/24$ as $d$ tends to infinity. Indeed, it is impossible to stab more than $n^3/24$ triangles for any $d$, since then projecting into a plane perpendicular to $\ell$ would result in a point stabbing more than $n^3/24$ triangles in the plane, contradicting Theorem~\ref{thm_any_S}. Theorem~\ref{thmsel} is a consequence of the following equipartition result, which is interesting in its own right. Given an integer $m \ge 2$, define an \emph{$m$-fan} as a set of $m$ hyperplanes in $\mathbb{R}^d$ that pass through a common $(d-2)$-flat. Then: \begin{theorem}\label{thm_part_4d_2_cont} For every probability measure $\mu$ on $\mathbb{R}^d$ that is absolutely continuous with respect to the Lebesgue measure, there exists a $(2d-1)$-fan that divides $\mu$ into $4d-2$ equal parts. \end{theorem} For $d=2$ this theorem specializes to a result of Ceder \cite{ceder} (also see Buck and Buck \cite{buckbuck} for a special case). We also show that $2d-1$ is the largest possible number of hyperplanes in Theorem~\ref{thm_part_4d_2_cont}: \begin{theorem}\label{thm_part_optimal} For every integer $m\ge 2d$ there exists an absolutely continuous probability measure $\mu$ on $\mathbb{R}^d$ that cannot be partitioned into $2m$ equal parts by an $m$-fan. \end{theorem} \section{The construction for Theorem \ref{momentcurvthm}} We now prove Theorem~\ref{momentcurvthm} by constructing a suitable point set $S$. Given real numbers $a, b > 1$, let $a \ll b$ mean that $f(a) < b$ for some fixed, sufficiently large function $f$ (concretely, we can take $f(x) = (d+1)! x^{d+1}$). Our point set is $S = \{ p_1, \ldots, p_n\}$, with \begin{equation*} p_i = (p_{i1}, p_{i2}, \ldots, p_{id}) \in (1,\infty)^d, \end{equation*} where the components $p_{ij}$ satisfy \begin{equation*} p_{ij} \ll p_{i'j'} \qquad \text{whenever} \qquad j < j', \quad \text{or}\quad j = j' \text{ and } i<i' \end{equation*} (so the ordering of the $p_{ij}$ is first by the coordinate index $j$ and then by the point index $i$). The idea of taking points separated by rapidly-increasing distances is borrowed from Boros and F\"uredi's planar construction \cite{boros_furedi_ptintriag}. However, their construction is more complicated, with points grouped into three clusters; see Appendix~\ref{app_BF_construction}. \begin{figure} \centerline{\includegraphics{planar_case}} \caption{\label{fig_planar_case}(\emph{a}) In the $x_1 x_2$-plane, the segment $p'_i p'_j$ has a smaller slope than the segment $p'_j p'_k$. (\emph{b}) In the planar case, the point $r$ must lie above-right of $q_0$, above-left of $q_1$, and below-left of $q_2$.} \end{figure} \begin{lemma} The set $S$ is in convex position. \end{lemma} \begin{proof} Let $p'_i = (p_{i1}, p_{i2})$ be the projection of point $p_i$ into the $x_1 x_2$-plane, for $1\le i\le n$. We claim that the points $p'_i$ lie on an $x_1$-monotone convex curve in the $x_1 x_2$-plane (which implies the lemma). To this end, we show that for every three points $p'_i$, $p'_j$, $p'_k$, with $i < j < k$, the segment $p'_i p'_j$ has a smaller slope than the segment $p'_j p'_k$; see Figure \ref{fig_planar_case}(\emph{a}). Indeed, this is the case if and only if \begin{equation}\label{eq_slope} (p_{k2} - p_{j2}) (p_{j1} - p_{i1}) > (p_{j2} - p_{i2}) (p_{k1} - p_{j1}). \end{equation} But (\ref{eq_slope}) will hold as long as the function $f$ in the definition of $\ll$ is chosen large enough. Specifically, if $f(x) \ge 4 x^2$, then the left-hand side of (\ref{eq_slope}) is at least \begin{equation*} {1\over 2} p_{k2} \cdot {1\over 2} p_{j1} \ge {1 \over 4}p_{k2} \ge p_{j2}^2, \end{equation*} which is larger than the right-hand side of (\ref{eq_slope}). \end{proof} Next, we want to show that no point $r = (r_1, \ldots, r_d)\in\mathbb{R}^d$ is contained in more than $(n/(d+1))^{d+1} + O(n^d)$ of the $d$-simplices spanned by $S$. We can assume that $p_{1j} \le r_j \le p_{nj}$ for each coordinate $1\le j\le d$, since otherwise, $r$ is not contained in any $d$-simplex spanned by $S$. For each coordinate $j=1,\ldots, d$, we discard from $S$ the last point $p_i$ with $p_{ij} \le r_j$ and the first point $p_i$ with $p_{ij} \ge r_j$. Let $S'$ be the resulting set. Since we have discarded at most $2d$ points, the number of $d$-simplices involving any of the discarded points is only $O(n^d)$. And now, for every $p_i\in S'$ and every $j$, we have either $r_j \ll p_{ij}$ or $r_j \gg p_{ij}$. Let $a=(a_1,\ldots,a_d)\in \mathbb{R}^d$ be a point; we define the \emph{type} of $a$ with respect to $r$ as $\max\{k: a_j> r_j \text{ for all } j = 1, 2, \ldots, k \}$. Note that the type of $a$ is an integer between $0$ and $d$ (it is $0$ if $a_1\le r_1$). Let $p_{i_0}, \ldots, p_{i_{d}} \in S'$ span a $d$-simplex containing $r$, with $i_0 < \cdots < i_{d}$. For convenience, we rename these points and their coordinates as \begin{equation*} q_\ell = (q_{\ell 1}, \ldots, q_{\ell d}), \qquad \text{for } \ell=0,1, \ldots, d. \end{equation*} The proof of Theorem~\ref{momentcurvthm} will be almost finished once we establish the following: \begin{lemma}\label{l:} For each $\ell=0,1,\ldots,d$, the point $q_\ell$ has type $\ell$ with respect to $r$. (See Figure~\ref{fig_planar_case}(b) for an illustration of the planar case.) \end{lemma} Indeed, assuming this lemma, the proof of Theorem~\ref{momentcurvthm} is concluded as follows. Given $r$, we partition the points of $S'$ into $d+1$ subsets $S'_0, \ldots, S'_d$ according to their type. Then, for a $d$-simplex spanned by $d+1$ points from $S'$ to contain $r$, each point must come from a different $S'_k$. The number of such simplices is thus at most $\prod_{k=0}^d |S'_k| \le (n/(d+1))^{d+1}$, by the arithmetic-geometric mean inequality. \begin{proof}[Proof of Lemma \ref{l:}.] We are going to derive the following relations: \begin{equation}\label{e:q<<} q_{(j-1)j} \ll r_j \ll q_{jj}\quad \text{for every } j=1, 2, \ldots d. \end{equation} Let us first check that they imply the lemma. To see that $q_\ell$ has type $\ell$, we need that $q_{\ell j}>r_j$ for $j\le\ell$ and, if $\ell<d$, also that $q_{\ell(\ell+1)}\le r_{\ell+1}$. The last inequality follows from (\ref{e:q<<}) with $j=\ell+1$. To derive $q_{\ell j}>r_j$, we use that the coordinates of $q_\ell$ are increasing since $q_\ell\in S$, and thus $q_{\ell j} \ge q_{jj}>r_j$. Now we start working on (\ref{e:q<<}). First we express the condition that $r$ lie in the simplex spanned by $q_0,\ldots,q_d$ using determinants. For each $\ell$, the points $r$ and $q_\ell$ must lie on the \emph{same} side of the hyperplane spanned by the points $q_m$, $m\neq \ell$. Thus, let $M$ be the $(d+1)\times(d+1)$ matrix consisting of rows $(1,q_0)$, $(1,q_1)$,\ldots, $(1,q_d)$. For $k = 0,1, \ldots, d$, let $M_k$ be the matrix obtained from $M$ by replacing the row $(1, q_k)$ by $(1, r)$. Then, for each $k$, $\det M_k$ must have the same sign as $\det M$. Next, we show that $\det M$ and each $\det M_k$ are ``dominated" by a single product of entries. Let $A$ be one of the matrices $M, M_0, M_1,\ldots, M_d$, and denote by $a_{\ell j}$ the entry in row $\ell$ and column $j$ of $A$, for $0\le \ell,j\le d$. We claim that if the function $f$ in the definition of $\ll$ is chosen sufficiently large, then there is a single product of the form $\sign(\sigma) \prod_\ell a_{\ell \sigma(\ell)}$, for some permutation $\sigma$, which is larger in absolute value than the sum of absolute values of all the other products in $\det A$. Indeed, let $a_{\ell_d d}$ be the largest entry in the last column of $A$. This is also the largest entry in the entire matrix. Then, if we take $f(x) \ge (d+1)!x^{d+1}$, any permutation product involving $a_{l_d d}$ is larger than $(d+1)!$ times any permutation product \emph{not} involving this entry. Thus, we choose $a_{\ell_d d}$ as the first term in our product, we remove row $\ell_d$ and column $d$ from $A$, and we continue in this fashion leftwards. We obtain a product $\prod_\ell a_{\ell \sigma(\ell)}$ which is larger than $(d+1)!$ times any other permutation product in $\det A$. Therefore, this product ``dominates" $\det A$ in the above sense, and so $\sign(\det A) = \sign(\sigma)$. In particular, these considerations for $A=M$ show that $\det M$ is dominated by the product \begin{equation*} q_{dd} q_{(d-1)(d-1)} \cdots q_{11}\cdot 1 \end{equation*} corresponding to the identity permutation. Therefore, $\det M >0$, and so we must have $\det M_k >0$ for all $k$. Now we are ready to prove (\ref{e:q<<}). First we suppose for contradiction that $r_j \gg q_{jj}$ for some $j=1,2,\ldots,d$. We take the largest such $j$; thus, $q_{kk}\gg r_k$ for $k>j$. Then $\det M_{j-1}$ is dominated by the product \begin{equation*} q_{dd}\cdots q_{(j+1)(j+1)} r_j q_{j(j-1)} q_{(j-2)(j-2)}\cdots q_{11}\cdot 1, \end{equation*} so the sign of $\det M_{j-1}$ is the sign of the permutation associated with this product. This is a permutation with exactly one inversion, so $\det M_{j-1}<0$, which is a contradiction. Next, we suppose for contradiction that $r_j \ll q_{(j-1)j}$ for some $j=1,2,\ldots,d$. Now we take the \emph{smallest} such $j$. We have already shown that $r_k \ll q_{kk}$ for all $k$. Therefore, $\det M_j$ is dominated by the product \begin{equation*} q_{dd}\cdots q_{(j+1)(j+1)} q_{(j-1)j} r_{j-1} q_{(j-2)(j-2)}\cdots q_{11}\cdot 1. \end{equation*} Again, this product corresponds to a permutation with exactly one inversion, so we have $\det M_j<0$, which is again a contradiction. \end{proof} \section{The construction for Theorem \ref{wagnersharpthm}} We now present the construction that proves Theorem~\ref{wagnersharpthm}. Let us call a set $Y\subseteq \mathbb{R}^d$ \emph{antisymmetric} if $Y\cap (-Y)=\emptyset$. We make use of the following result of Wendel: \begin{lemma}[\cite{wendel_sphere}]\label{lemma_wendel} Let $X = \{x_1, \ldots, x_{d+1} \}$ be a set of $d+1$ points in general position on the unit sphere $\S^{d-1}$ in $\mathbb{R}^d$. Then there are exactly two antisymmetric $(d+1)$-point subsets of $X\cup (-X)$ whose convex hull contains the origin; if one of them is $Y$, then the other one is $-Y$. \end{lemma} Let $\alpha\in(0, 1/2]$ be a parameter and let $n$ be given. For the moment assume for simplicity that $\alpha n$ is an integer. Let $A$ be a set of $\alpha n$ points on $\S^{d-1}$ and let $p$ be another point on $\S^{d-1}$ such that $A \cup \{p\}$ is in general position --- namely, such that the set $A\cup(-A)\cup\{p\}$ has $2|A|+1$ points and no hyperplane containing $p$ and a $(d-1)$-point antisymmetric subset of $A\cup(-A)$ passes through the origin. Let $P$ be a very small cluster of $(1-2\alpha) n$ points around $p$. Our set is $S = A \cup (-A) \cup P$. Note that $|S| = n$ as required. The origin clearly lies at depth $\alpha n$ with respect to $S$. Thus, Theorem~\ref{wagnersharpthm} reduces to the following lemma: \begin{lemma} The number of $(d+1)$-point subsets $B$ of $S$ such that $\conv B$ contains the origin is \begin{equation*} \left((d+1) \alpha^d - 2d\alpha^{d+1}\right){n\choose d+1} + O(n^d). \end{equation*} \end{lemma} \begin{proof} The number of $(d+1)$-point subsets of $S$ that contain a pair of antipodal points (one in $A$ and one in $-A$) is $O(n^d)$, and so it suffices to count the number of $B$ that are antisymmetric. The choice of $A$ and $P$ guarantees that if $B$ is antisymmetric and $|B\cap P|\ge 2$, then $0\not\in\conv B$. So we need to consider the cases $B\cap P=\emptyset$ and $|B\cap P|=1$. Let us set $\tilde B= \{x\in A\cup P: x\in B \mbox{ or } -x\in B\}$. For $B\cap P=\emptyset$ there are $\alpha n\choose d+1$ ways of choosing $\tilde B\subseteq A$, and for each of them we have two choices for $B$ by Lemma~\ref{lemma_wendel}. For $|B\cap P|=1$, we have $(1-2\alpha)n{\alpha n\choose d}$ choices for $\tilde B$, and each of them yields exactly one $B$ (Lemma~\ref{lemma_wendel} with $X=\tilde B$ shows that there are two $B\subset \tilde B\cup(-\tilde B)$ \ with $0\in\conv B$, and exactly one of these contains the point $p\in P\cap \tilde B$, while the other contains $-p$). Altogether the number of $B$'s is $2{\alpha n\choose d+1}+ (1-2\alpha)n{\alpha n\choose d}+O(n^d)$, and the lemma follows by algebraic manipulation. \iffalse ****************************** PROB ARGUMENT CHANGED TO COUNTING Let $Z$ be the event that the set $B$ contains a pair of antipodal points from $A \cup (-A)$. If $Z$ occurs then we certainly have $0\in \conv B$; however, $Z$ only occurs with probability $O(1/n)$. Thus, let us analyze the probability that $0 \in \conv B$ conditioned on $\overline Z$. Let $m = |B \cap P|$. We observe that if $B$ contains no pair of antipodal points (from $A\cup(-A)$), and $m\ge 2$, then $\conv B$ is sufficiently flat so that it avoids the origin (assuming that the cluster $P$ is chosen sufficiently small). Thus, \begin{equation*} \Pr[ 0 \in \conv B \mid \overline Z \wedge m\ge 2] = 0. \end{equation*} So it suffices to consider the case $m\le 1$. From Lemma~\ref{lemma_wendel} we have \begin{equation*} \Pr[ 0\in \conv B \mid \overline Z \wedge m= 0] = 2^{-d} \end{equation*} (choose $B$ by taking a random $(d+1)$-point sample from $A$ and then deciding randomly which points to reflect). Moreover, by the observation following Lemma~\ref{lemma_wendel} we also have \begin{equation*} \Pr[ 0\in \conv B \mid \overline Z \wedge m=1] = 2^{-d}. \end{equation*} Finally, we have \begin{align*} \Pr[\overline Z \wedge m\le 1] &= \left( 2^{d+1} {\alpha n \choose d+1} + 2^d {\alpha n \choose d} (1-2\alpha) n \right) \bigg/ {n\choose d+1}\\ &= (d+1)(2\alpha)^d - d(2\alpha)^{d+1} + O(1/n). \end{align*} Putting everything together, \begin{align*} \Pr[0\in\conv B] &= \Pr[Z] + \Pr[\overline Z \wedge m\le 1] \cdot \Pr[0\in \conv B \mid \overline Z \wedge m\le 1 ]\\ &= (d+1)\alpha^d - 2d\alpha^{d+1} + O(1/n). \qedhere \end{align*} \fi \end{proof} If $\alpha n$ is not an integer, then apply the above argument using $\alpha' = \lceil \alpha n \rceil/n$, and use the fact that $\alpha' - \alpha < 1/n$. \section{Partitioning measures by fans of hyperplanes} In this section we prove Theorem~\ref{thm_part_4d_2_cont}, which is the main ingredient in the proof of Theorem~\ref{thmsel}. We then prove Theorem~\ref{thm_part_optimal}, showing that Theorem~\ref{thm_part_4d_2_cont} is optimal. Recall that an \emph{$m$-fan} is a set of $m$ hyperplanes in $\mathbb{R}^d$ sharing a common $(d-2)$-flat. Let $\S^{d-1}$ denote the unit sphere in $\mathbb{R}^d$, and let $\Stief_{d,2} = \{ (v,w) \in (\S^{d-1})^2 : v\perp w \}$ denote the set of ordered pairs of orthonormal vectors in $\mathbb{R}^d$ (called the \emph{Stiefel manifold} of orthogonal $2$-frames). The proof of Theorem~\ref{thm_part_4d_2_cont} is based on the following topological result: \begin{lemma}\label{lemma_no_func} There exists no continuous function $g \colon \Stief_{d,2} \to \S^{2d-4}$ with the following property: For every $(v,w) \in \Stief_{d,2}$, if $g(v,w) = (a_1, \ldots, a_{2d-3})$ (with $a_1^2 + \ldots + a_{2d-3}^2 = 1$), then \begin{itemize} \item $g(-v,w) = (-a_1, \ldots, -a_{d-1}, a_d, \ldots, a_{2d-3})$, and \item $g(v,-w) = (a_1, \ldots, a_{d-1}, -a_d, \ldots, -a_{2d-3})$. \end{itemize} \end{lemma} \begin{proof} The lemma is a result on nonexistence of an \emph{equivariant map}. Let us briefly recall the basic setting; for more background we refer to \cite{Zivaljevic-topmeth}, \cite{Mat-top}. Let $G$ be a finite group. A \emph{$G$-space} is a topological space $X$ together with an \emph{action} of $G$ on $X$, which is a collection $(\varphi_g)_{g\in G}$ of homeomorphisms $\varphi_g\:X\to X$ whose composition agrees with the group operation in $G$; that is, $\varphi_e={\rm id}_X$ for the unit element $e\in G$ and $\varphi_g\circ\varphi_h=\varphi_{gh}$ for all $g,h\in G$. In our case, the relevant group is $G:=\Z_2\times \Z_2$ (the direct product of two cyclic groups of order $2$). We can write $G=\{e, g_1, g_2, g_1g_2\}$, where $g_1$ and $g_2$ are two generators of $G$; in order to specify an action of $G$, it is enough to give the homeomorphisms corresponding to $g_1$ and $g_2$. The lemma deals with two $G$-spaces: \begin{itemize} \item The Stiefel manifold $\Stief_{d,2}$ with the action $(\varphi_g)_{g\in G}$ of $G$ given by $\varphi_{g_1}(v,w)=(-v,w)$, $\varphi_{g_2}(v,w)=(v,-w)$. \item The sphere $\S^{2d-4}$ with the action $(\psi_g)_{g\in G}$, where $\psi_{g_1}$ flips the signs of the first $d-1$ coordinates and $\psi_{g_2}$ flips the signs of the remaining $d-2$ coordinates. \end{itemize} We want to prove that there is no \emph{equivariant map} $f\: \Stief_{d,2}\to \S^{2d-4}$, where an equivariant map is a continuous map that commutes with the actions of $G$, i.e., such that $f\circ \varphi_g=\psi_g\circ f$ for all $g\in G$. The ``usual'' elementary methods for showing nonexistence of equivariant maps, explained in \cite{Zivaljevic-topmeth}, \cite{Mat-top} and based on the Borsuk--Ulam theorem and its generalizations, cannot be applied here. We use the \emph{ideal-valued cohomological index} of Fadell and Husseini \cite{FadellHusseini1} (also see \cite{Zivaljevic-guide2}). This method assigns to every $G$-space $X$ the \emph{$G$-index} of $X$, denoted by ${\rm Ind}_G(X)$, which is an ideal in a certain ring $R_G$ (depending only on $G$). A key property is that whenever there is an equivariant map $f\:X\to Y$, where $X$ and $Y$ are $G$-spaces, we have ${\rm Ind}_G(Y)\subseteq {\rm Ind}_G(X)$. For the considered $G=\Z_2\times\Z_2$, $R_G$ is the ring $\Z_2[t_1,t_2]$ of polynomials in two variables with $\Z_2$ coefficients. The general definition of ${\rm Ind}_G(X)$, as well as its computation, are rather complicated, but fortunately, in our case we can use ready-made results from the literature. For the $G$-space $\S^{2d-4}$ with the $G$-action as above, the $G$-index is the principal ideal in $\Z_2[t_1,t_2]$ generated by $t_1^{d-1}t_2^{d-2}$ according to Corollary~2.12 in \cite{Zivaljevic-guide2}. On the other hand, Fadell \cite{Fadell} proved that the $G$-index of the $G$-space $\Stief_{d,2}$ with the described $G$-action does not contain the monomial $t_1^{d-1}t_2^{d-2}$ (also see \cite{Inoue} for a statement of this result and some applications of it). This shows that an equivariant map as in the lemma is indeed impossible. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm_part_4d_2_cont}.] We follow the ``configuration space/test map'' paradigm (see, e.g., \cite{Zivaljevic-topmeth}). We encode each ``candidate'' for the desired equipartition, which in our case is going to be a certain special fan of $4d-2$ half-hyperplanes sharing the boundary $(d-2)$-flat, by a point of $\Stief_{d,2}$. Then we define a continuous \emph{test map} that assigns to each candidate fan of half-hyperplanes a $(2d-3)$-tuple of real numbers, which measures how far the given candidate is from being a $(2d-1)$-fan of hyperplanes. Finally we will check that if there were no equipartition, the test map would yield an equivariant map $\Stief_{d,2} \to \S^{2d-4}$, which would contradict Lemma~\ref{lemma_no_func}. The details follow. For the proof we may assume that every nonempty open set has a positive $\mu$-measure. (Given an arbitrary $\mu$, we can consider the convolution $\mu*\gamma_\varepsilon$ of $\mu$ with a suitable probability measure $\gamma_\varepsilon$ whose density function is everywhere nonzero but for which all but at most $\varepsilon$ of the mass lies in a ball of radius $\varepsilon$ around $0$. The convolution has the required property and then, given an equipartition for each $\mu*\gamma_\varepsilon$ a limit argument, letting $\varepsilon\to0$, yields an equipartition for the original $\mu$. See the proof of \cite[Theorem 3.1.1]{Mat-top} for a similar limit argument.) Let $m = 2d-1$. Suppose we are given two orthonormal vectors $v, w \in \S^{d-1}$. Let $h$ be the unique hyperplane orthogonal to $v$ that splits $\mathbb{R}^d$ into two halfspaces of equal measure with respect to $\mu$. We say that the halfspace in the direction of $v$ is ``above" $h$, and the other halfspace is ``below" $h$. Let $\ell$ be a $(d-2)$-flat orthogonal to $w$ contained in $h$. Note that $\ell$ splits $h$ into two half-hyperplanes. We say that the half of $h$ in the direction of $w$ lies ``left" of $\ell$, and the other half of $h$ lies ``right" of $\ell$. Every half-hyperplane with boundary $\ell$ is uniquely determined by the angle it makes with the left half of $h$. \begin{figure} \centerline{\includegraphics{fig_half_hyperplanes}} \caption{\label{fig_half_hyperplanes}$2m$ half-hyperplanes coming out of $\ell$ that partition the measure $\mu$ into $2m$ equal parts. (Here $m=5$.)} \end{figure} Let $f_0, f_1, \ldots, f_{2m-1}$ be $2m$ half-hyperplanes coming out of $\ell$, listed in circular order, that split the measure $\mu$ into $2m$ equal parts, as follows: \begin{itemize} \item $f_0$ is the left half of $h$; \item $f_1, \ldots, f_{m-1}$ lie above $h$; \item $f_m$ is the right half of $h$; and \item $f_{m+1}, \ldots, f_{2m-1}$ lie below $h$. \end{itemize} See Figure~\ref{fig_half_hyperplanes}. For $i = 1, \ldots, m-1$, let $\alpha_i$ be the angle between $f_0$ and $f_i$, and let $\beta_i$ be the angle between $f_m$ and $f_{m+i}$. Let $\gamma_i = \alpha_i - \beta_i$. Note that $\gamma_i = 0$ means that $f_i$ and $f_{m+i}$ are aligned into a hyperplane. Translating $\ell$ within $h$ to the left causes the $\alpha_i$'s to increase and the $\beta_i$'s to decrease, while translating it to the right has the opposite effect. Therefore, there exists a unique position of $\ell$ for which $\sum \alpha_i = \sum \beta_i$, or equivalently, $\sum \gamma_i = 0$, and we fix $\ell$ there. In this way, we have defined each $\alpha_i$ and $\beta_i$ as a function of the given vectors $v$, $w$. Using the assumption that $\mu$ is absolutely continuous with respect to the Lebesgue measure and each open set has a positive $\mu$-measure, it is routine to verify the continuity of the $\alpha_i$ and $\beta_i$ as functions of $v$ and $w$. \begin{figure} \centerline{\includegraphics{fig_flip_v_w}} \caption{\label{fig_flip_v_w} The effect of changing the sign of $v$ (left) or $w$ (right).} \end{figure} Let us examine what happens when we change the sign of $v$ or $w$. We have: \begin{align*} \alpha_i(-v,w) &= \pi - \beta_{m-i}(v,w),& \beta_i(-v,w) &= \pi - \alpha_{m-i}(v,w),\\ \alpha_i(v,-w) &= \pi - \alpha_{m-i}(v,w),& \beta_i(v,-w) &= \pi -\beta_{m-i}(v,w). \end{align*} See Figure \ref{fig_flip_v_w}. Therefore, \begin{align*} \gamma_i(-v,w) &= \gamma_{m-i}(v,w),\\ \gamma_i(v,-w) &= -\gamma_{m-i}(v,w). \end{align*} Now we introduce a suitable change of coordinates in the target space so that the resulting map behaves as the map $g$ considered in Lemma~\ref{lemma_no_func}. Namely, we set \begin{align*} \lambda_i &= \gamma_i - \gamma_{m-i}, \qquad \mathrm{for\ } i = 1, \ldots, (m-1)/2;\\ \mu_i &= \gamma_i + \gamma_{m-i}, \qquad \mathrm{for\ } i = 2, \ldots, (m-1)/2. \end{align*} Note that \begin{align*} \lambda_i(-v,w) &= -\lambda_i(v,w), & \mu_i(-v,w) &= \mu_i(v,w),\\ \lambda_i(v,-w) &= \lambda_i(v,w),& \mu_i(v,-w) &= -\mu_i(v,w). \end{align*} We have $\gamma_i = 0$ for all $i$ if and only if $\lambda_i, \mu_i = 0$ for all $i$. (Recall that $\sum \gamma_i = 0$.) Now we define the ``test map'' $G \colon \Stief_{d,2} \to \mathbb{R}^{m-2}$ by \begin{equation*} G(v,w) = (\lambda_1, \ldots \lambda_{(m-1)/2}, \mu_2, \ldots, \mu_{(m-1)/2}). \end{equation*} Then, our desired equipartition of $\mu$ exists if and only if $G(v,w) = (0,\ldots,0)$ for some $(v,w)$. But $G$ is a continuous map such that flipping $v$ flips the first $(m-1)/2 = d-1$ coordinates of the image, while flipping $w$ flips the last $(m-3)/2 = d-2$ coordinates of the image. If we had $G \neq (0,\ldots, 0)$ for all $(v,w)$, the map $g \colon \Stief_{d,2} \to \S^{2d-4}$ given by $g(v,w) = G(v,w) / \| G(v,w) \|$ would contradict Lemma~\ref{lemma_no_func}. Therefore, the desired equipartition exists. \end{proof} We conclude this section by proving Theorem~\ref{thm_part_optimal}, which shows that Theorem~\ref{thm_part_4d_2_cont} is best possible, in the sense that an equipartition of a measure $\mu$ in $\mathbb{R}^d$ by a fan of $2d$ or more hyperplanes does not necessarily exist. The proof is based on the following lemma: \begin{lemma}\label{lemma_exists_gral_pos} Let $m>0$ be an integer and let $t\ge 2d+m-1$. Then there exists a $t$-point set $T \subset \mathbb{R}^d$ that cannot be covered by any $m$-fan in $\mathbb{R}^d$. \end{lemma} The basic idea, roughly speaking, is that an $m$-fan in $\mathbb{R}^d$ has $2d+m-2$ degrees of freedom, while each point in $T$ takes away one degree of freedom. Therefore, $T$ can be completely covered by an $m$-fan only if it is degenerate an appropriate sense. \begin{proof}[Proof of Lemma~\ref{lemma_exists_gral_pos}] For convenience we first prove the result in $\Prj^d$, the $d$-dimensional projective space, and then we show that the result also applies to $\mathbb{R}^d$. A set of $m$ hyperplanes in $\Prj^d$ share a common $(d-2)$-flat if and only if their dual points, when considered as vectors in $\mathbb{R}^{d+1}$, span a vector space of dimension at most $2$. Thus, define the projective variety \begin{equation*} V = \bigl\{ (p_1,\dotsc,p_m)\in (\Prj^d)^m : \rank(p_1,\dotsc,p_m)\le 2 \bigr\}, \end{equation*} where $\rank(p_1,\dotsc,p_m)$ denotes the dimension of the vector space spanned by $p_1,\dotsc,p_m$ as vectors in $\mathbb{R}^{d+1}$. The variety $V$ has dimension $\dim V=2d+m-2$. Given a point $p=p_0 : p_1 : \dotsb : p_d\in\Prj^d$, let $p^*=\{x_0 : \dotsb : x_d\in \Prj^d : \sum_i x_i p_i=0\}$ denote the hyperplane dual to $p$. For each $v=(p_1,\dotsc,p_m)\in V$, let $v^*=\bigcup_i p_i^* \subset \Prj^d$ be the variety which consists of the union of the hyperplanes dual to the points in $v$. Finally, define the moduli space \begin{equation*} C = \bigl\{ (v, q_1,\dotsc,q_t)\in V\times (\Prj^d)^t : q_1,\dotsc,q_t\in v^* \bigr\} \end{equation*} of all $t$-tuples of points lying on a fan $v^*$, for all $v\in V$. The dimension of $C$ is $\dim C = \dim V+t(d-1) = 2d+m-2 + t(d-1)$. Consider the projection map $\pi\colon V \times (\Prj^d)^t \to (\Prj^d)^t$. Then the projection $\pi(C)$ is the set of $t$-tuples of points in $\Prj^d$ that can be covered by an $m$-fan. By the Tarski--Seidenberg Theorem \cite{real_algref} $\pi(C)$ is a semialgebraic subset of $(\Prj^d)^t$. Since projection does not increase dimension, $\pi(C)$ is of dimension at most $2d+m-2+t(d-1)$, which by our choice of $t$ is smaller than $td = \dim (\Prj^d)^t$. Thus, there exists a $t$-point set $T$ in $\Prj^d$ that cannot be covered by an $m$-fan. Finally, a generic $\mathbb{R}^d$ inside $\Prj^d$ completely contains $T$, and the lemma follows. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm_part_optimal}] Given an integer $m\ge 2d$, let $t = 2m - 1$. We have $t\ge 2d+m-1$, so by the preceding lemma there exists a $t$-point set $T\subset \mathbb{R}^d$ that cannot be covered by any $m$-fan in $\mathbb{R}^d$. There must exist a positive radius $r$ such that, for every $m$-fan $F$ in $\mathbb{R}^d$, some point of $T$ lies at distance at least $r$ from the closest hyperplane in $F$. (Otherwise a limit argument would yield an $m$-fan that covers $T$.) Let $C_r(p)$ denote the ball of radius $r$ centered at $p$. Let $\mu$ be the uniform measure on $\bigcup_{p\in T} C_r(p)$. Then, in every partition of $\mathbb{R}^d$ into $2m$ parts by an $m$-fan, there exists a part that completely contains one of the balls $C_r(p)$. This part has measure at least $\frac{1}{t}>1/(2m)$, and so the partition is not an equipartition. \end{proof} \section{Stabbing many triangles in $\mathbb{R}^d$} In this section we prove Theorem~\ref{thmsel} by means of Theorem~\ref{thm_part_4d_2_cont}. The proof is an extension of the technique in \cite{bukh} for the case $d=2$. By a standard approach (see e.g.~Theorem~3.1.2 in \cite{Mat-top}), Theorem~\ref{thm_part_4d_2_cont} implies the following discrete version, which is what we actually use: \begin{corollary}\label{c_part_4d_2} Let $S$ be a set of $n$ points in $\mathbb{R}^d$. Then there exist $2d-1$ hyperplanes passing through a common $(d-2)$-flat that divide the space into $4d-2$ parts, each containing at most $n/(4d-2) + O(1)$ points of $S$. \end{corollary} We start with the following lemma: \begin{lemma}\label{lemma_2m_parts_plane} Let $\ell_0, \ldots, \ell_{m-1}$ be $m$ lines in the plane passing through a common point $x$, dividing the plane into $2m$ sectors. Let $P = \{p_0, \ldots, p_{2m-1}\}$ be $2m$ points, one from each sector, listed in circular order around $x$. Then, out of the $2m\choose 3$ triangles defined by $P$, at least $(m+1)m(m-1)/3$ contain $x$. (This minimum is achieved if $P\cup \{x\}$ is in general position, in particular if no two points of $P$ are collinear with $x$.) \end{lemma} \begin{proof} Let $p_i p_j$ be a directed segment joining two points of $P$, and let $d = (j-i) \bmod 2m$. If $0\le d \le m-1$, we call the segment $p_i p_j$ \emph{short}; if $d = m$, we call it \emph{medium}; and if $m+1 \le d \le 2m-1$, we call it \emph{long}. A triangle $p_i p_j p_k$, with $i<j<k$, can have either three short sides, or two short sides and one medium side, or two short sides and one long side. \begin{figure} \centerline{\includegraphics{fig_triangle_types}} \caption{\label{fig_triangle_types}(\emph{a}) A triangle with three short sides always contains $x$. (\emph{b}) The triangles with one medium side can be partitioned into pairs, such that at least one triangle from each pair contains $x$. (\emph{c}) A triangle with one long side never contains $x$.} \end{figure} It is easy to see that all the triangles with three short sides contain $x$, and none of the triangles with one long side contain $x$. Furthermore, the triangles with one medium side can be grouped into pairs, such that from each pair, at least one triangle contains $x$ (exactly one triangle if $P\cup \{x\}$ is in general position). See Figure~\ref{fig_triangle_types}. The number of triangles with three short sides is \begin{equation*} {2m\over 3}\bigl(1 + 2 + \cdots + (m-2) \bigr) = {m(m-1)(m-2) \over 3}; \end{equation*} and the number of triangles with one medium side is $2m(m-1)$. Thus, $P$ defines at least \begin{equation*} {m(m-1)(m-2) \over 3} + m(m-1) = {(m+1)m(m-1)\over 3} \end{equation*} triangles that contain $x$ (exactly these many if $P\cup \{x\}$ is in general position). \end{proof} \begin{corollary}\label{cor_2m_parts_d} Let $h_0, \ldots, h_{m-1}$ be $m$ hyperplanes in $\mathbb{R}^d$ that pass through a common $(d-2)$-flat $\ell$ and divide space into $2m$ parts. Let $P = \{ p_0, \ldots, p_{2m-1} \}$ be $2m$ points, one from each part. Then $P$ defines at least $(m+1)m(m-1)/3$ triangles that intersect $\ell$. \end{corollary} \begin{proof}[Proof of Theorem \ref{thmsel}.] Let $S$ be an $n$-point set in $\mathbb{R}^d$. By Corollary~\ref{c_part_4d_2} there exist $2d-1$ hyperplanes that pass through a common $(d-2)$-flat $\ell$ and partition $S$ into parts of size at most $n/(4d-2) + O(1)$ each. We show that $\ell$ is our desired $(d-2)$-flat. Each part has at least $n/(4d-2) - O(1)$ points, so there are at least $\left({n \over 4d-2} - O(1) \right)^{4d-2}$ ways to choose $4d-2$ points, one from each part. By Corollary~\ref{cor_2m_parts_d}, each such choice of points defines at least $2d(2d-1)(2d-2)/3$ triangles that intersect $\ell$. On the other hand, each such triangle is counted at most $\left({n \over 4d-2} + O(1) \right)^{4d-5}$ times. Thus the number of triangles intersected by $\ell$ is at least \begin{multline*} \left({n \over 4d-2} - O(1) \right)^{4d-2} \cdot {2d(2d-1)(2d-2) \over 3} \bigg/ \left({n \over 4d-2} + O(1) \right)^{4d-5}\\ = {d^2-d \over 6(2d-1)^2} n^3 - O(n^2). \qedhere \end{multline*} \end{proof} \section{Discussion} The main open problem is to determine the exact value of the constants $c_d$ of the First Selection Lemma for $d\ge 3$. There remains a multiplicative gap of roughly $(d-1)!$ between the current lower bound (\ref{eq_wagner_bd}) and our upper bound (given by Theorem~\ref{momentcurvthm}). We conjecture that Theorem~\ref{momentcurvthm} is tight, and that the correct constants are $c_d = (d+1)^{-(d+1)}$. We suspect that the construction in Theorem~\ref{momentcurvthm} also witnesses sharpness of Theorem~\ref{thmsel}. But, to our embarrassment, we have been unable to find even the line that stabs most triangles in this construction for $d = 3$. \subsection{A generalization of centerpoints} Rado's Centerpoint Theorem \cite{rado} implies that for every $n$-point set $S$ in $\mathbb{R}^d$ there exists a $(d-2)$-flat $\ell$ that lies at depth $n/3$ with respect to $S$, in the sense that every halfspace that contains $\ell$ contains at least $n/3$ points of $S$. (Simply project $S$ into an arbitrary plane.) But the $(d-2)$-flat $\ell$ of Corollary~\ref{c_part_4d_2} lies at depth $(d-1)n/(2d-1) - O(1)$ with respect to $S$. (Indeed, every halfspace that contains $S$ completely contains $2d-2$ of the $4d-2$ parts mentioned in Corollary~\ref{c_part_4d_2}.) We do not know whether this bound is tight. This suggests the following generalization of Rado's Theorem: If $S$ is an $n$-point set in $\mathbb{R}^d$ and $0\le k < d$, then there always exists a $k$-flat at depth $\delta_{d,k} n$ with respect to $S$---a ``center-$k$-flat"---for some constants $\delta_{d,k}$. The general question of determining these constants $\delta_{d,k}$ has not been explored, as far as we know. (The formula $\delta_{d,k} = (k+1)/(d+k+1)$ seems to fit all the currently known data.) \subsection{From the First Selection Lemma to the Second} The First Selection Lemma has been generalized by B\'ar\'any et al.~\cite{bfl}, in conjunction with Alon et al. \cite{abfk_secondsel}, and \v Zivaljevi\'c and Vre\'cica \cite{zv_color_tverberg}, to the following result, called the \emph{Second Selection Lemma} in \cite{matou_book}: If $S$ is an $n$-point set in $\mathbb{R}^d$ and $\mathcal F$ is a family of $m\le {n \choose d+1}$ $d$-simplices spanned by $S$, then there exists a point $p\in \mathbb{R}^d$ contained in at least \begin{equation}\label{eq_2nd_SL} c'_d \left({ m \over n^{d+1}}\right)^{s_d} n^{d+1} \end{equation} simplices of $\mathcal F$, for some constants $c'_d$ and $s_d$ that depend only on $d$. (Note that $m/n^{d+1} = O(1)$, so the smaller the constant $s_d$, the stronger the bound.) The Second Selection Lemma is an important ingredient in the derivation of non-trivial upper bounds for the number of \emph{$k$-sets} in $\mathbb{R}^d$ (see \cite[ch.~11]{matou_book} for the definition and details). The derivation proceeds by ``lifting" the lemma by one dimension, obtaining that if $\mathcal F$ is a family of $m$ $d$-simplices spanned by $n$ points in $\mathbb{R}^{d+1}$, then there exists a \emph{line} that stabs $\Omega{\left( \left({ m \over n^{d+1}}\right)^{s_d} n^{d+1} \right)}$ simplices of $\mathcal F$. Does this lifting step result in a loss of tightness? If we may make an analogy from the results of this paper, it seems that the answer is yes. (As we showed, $c_{2,0} = 1/27$ by Theorem~\ref{momentcurvthm}, whereas $c_{3,1} \ge 1/25$ by Theorem~\ref{thmsel}.) The current best bound for the Second Selection Lemma for $d = 2$ is $\Omega(m^3 / (n^6 \log^2 n))$, due to Eppstein \cite{eppstein, nivasch_sharir} (so $s_2$ can be taken arbitrarily close to $3$ in (\ref{eq_2nd_SL})). On the other hand, we know that if $\mathcal F$ is a set of $m$ triangles in $\mathbb{R}^3$ spanned by $n$ points, there exists a line (specifically, a line determined by two points of $S$) that stabs $\Omega(m^3 / n^6)$ triangles of $\mathcal F$ (see \cite{dey_edelsbrunner} and \cite{shakharphd} for two different proofs of this fact). It might turn out that this logarithmic gap between the two cases is an artifact of the current proofs, but we believe that the three-dimensional problem \emph{does} have a larger bound than the planar one. \subsection*{Acknowledgment} We would like to thank an anonymous referee for useful comments. \bibliographystyle{alpha}

https://arxiv.org/abs/2208.12656

Ramanujan's $q$-continued fractions

Ramanujan's $q$-continued fractions are a central part of Ramanujan's development of basic hypergeometric series. They appear in Chapter 16 of Part III and Chapter 32 of Part V of {\em Ramanujan's Notebooks} edited by Berndt, and in Volume I of Andrews and Berndt's {\em Ramanujan's Lost Notebook}. In these references the continued fractions as presented in the order in which they appear in Ramanujan's original notebooks. We summarize the work of several authors on this topic and re-organize Ramanujan's $q$-continued fractions.

\section{Introduction} Ramanujan considered the $q$-analogue of the well-known Fibonacci continued fraction \begin{equation*}\label{golden-mean-cfrac} \frac{1}{1}\genfrac{}{}{0pt}{}{}{+} \frac{1}{1}\genfrac{}{}{0pt}{}{}{+}\frac{1}{1}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} \end{equation*} given by \begin{subequations} \begin{equation}\label{RRcfrac1} \frac{1}{1}\genfrac{}{}{0pt}{}{}{+}\frac{q}{1}\genfrac{}{}{0pt}{}{}{+}\frac{q^2}{1}\genfrac{}{}{0pt}{}{}{+}\frac{q^3}{1}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots}, \end{equation} and showed that for $|q|<1$, this continued fraction can be written as a ratio of two very similar sums: \begin{equation} \frac {\displaystyle\sum_{k=0}^{\infty} \frac{q^{k^2+k}}{{(1-q)(1-q^2)\cdots (1-q^k)}}} {\displaystyle\sum_{k=0}^{\infty} \frac{q^{k^2}}{(1-q)(1-q^2)\cdots (1-q^k)}}. \label{rr-cfrac1} \end{equation} The continued fraction \eqref{RRcfrac1} is called the Rogers--Ramanujan continued fraction. In view of the Rogers--Ramanujan identities, this ratio can be expressed as a ratio of infinite products \begin{equation} \frac {(1-q)(1-q^6)(1-q^{11}) \cdots } {(1-q^2)(1-q^7)(1-q^{12}) \cdots } \times \frac{(1-q^4)(1-q^9) (1-q^{14}) \cdots} {(1-q^3) (1-q^8)(1-q^{13}) \cdots} .\label{entry16.38.iii} \end{equation} \end{subequations} It is clear that Ramanujan considered his work on this continued fraction as one of the highlights of his work. Consider the last two entries of Chapter 16 of Ramanujan's second notebook \cite{RamanujanNB}. Entry 38 concerns the continued fraction above and the Rogers--Ramanujan identities. Entry 39 computes special cases (see Berndt \cite[p.~77--86]{Berndt1991}). As a corollary, Ramanujan obtains \begin{subequations} \begin{align} \frac{1}{1}\genfrac{}{}{0pt}{}{}{+}\frac{e^{-2\pi}}{1}\genfrac{}{}{0pt}{}{}{+}\frac{e^{-4\pi}}{1} \genfrac{}{}{0pt}{}{}{+}\frac{e^{-6\pi}}{1}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} &=\Bigg( \sqrt{\frac{5+\sqrt{5}}{2} } - \frac{\sqrt{5}+1}{2} \Bigg) e^{2\pi/5}, \label{II.16.39ii-cor} \\ \intertext{and} \frac{1}{1}\genfrac{}{}{0pt}{}{}{-}\frac{e^{-\pi}}{1}\genfrac{}{}{0pt}{}{}{+}\frac{e^{-2\pi}}{1} \genfrac{}{}{0pt}{}{}{-}\frac{e^{-3\pi}}{1}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} &= \Bigg(\sqrt{\frac{5- \sqrt{5}}{2} } - \frac{\sqrt{5}-1}{2}\Bigg) e^{\pi/5} . \label{II.16.39i-cor} \end{align} \end{subequations} Askey \cite{Askey1989} suggested that Ramanujan's $q$-extension \eqref{RRcfrac1} was the primary motivation for Ramanujan to discover the Rogers--Ramanujan identities, and indeed for Ramanujan's development of basic hypergeometric series. In support of this hypothesis, we note that Ramanujan ended his development of basic hypergeometric series with \eqref{II.16.39ii-cor} and \eqref{II.16.39i-cor}. Thus $q$-continued fractions may be quite central to Ramanujan's thought process. Ramanujan himself included these two continued fractions in his first letter \cite[p.~29]{BR1995} to Hardy in 1913; this is additional evidence that these were results Ramanujan liked and thought likely to impress Hardy. And impress him, they did. About these, and another formula related to this continued fraction, Hardy \cite[p.~9]{Hardy1959} wrote \begin{quote} (they) defeated me completely; I had never seen anything in the least like them before. A single look at them is enough to show that they could only be written by a mathematician of the highest class. \end{quote} The objective of this article is to summarize the work of many mathematicians, whose work has made it easier to understand and organize Ramanujan's continued fractions. % The three formulas \eqref{RRcfrac1}--\eqref{entry16.38.iii} serve as a model for this organization. As it turns out, there are only five general results. In all of these a ratio of two series is expressed as a continued fraction. The remaining continued fractions are special cases of these. These general results are given in \S\S\ref{g-sec}, \ref{G-sec} and \ref{sec:entry11-12}. In some cases, the same ratio of two series is expressed as two or three continued fractions; thus, they are transformation formulas too. We note some special cases in \S\ref{sec:cf-trans}. There are some special cases where the ratio reduces to a single series, and many where they reduce to a ratio of infinite products (see \S\ref{sec:cf-prods}). When this happens, such as in the case of the Rogers--Ramanujan continued fraction, Ramanujan computes special values. In this article, we list all of Ramanujan's $q$-continued fractions. ( Formulas for the computation of special values have not been considered here.) Apart from this organization, we are interested in the ideas involved in the proofs. A variety of clever ideas are required to prove Ramanujan's $q$-continued fractions; we present a sample, giving details when possible, and references otherwise. We conclude in \S\ref{sec:proofs} with some remarks arising from our study, and pointers to recent work in this area. \subsection*{Notation} The {\em $q$-rising factorial} is defined as $\qrfac{a}{0} :=1$, $$\qrfac{a}{n} := (1-a)(1-aq)\cdots (1-aq^{n-1}) \text{ for } n=1, 2, \dots; $$ and, for $|q|<1$, $$\qrfac{a}{\infty} := \prod_{k=0}^\infty (1-aq^{k}).$$ In addition, we use the short-hand notation \begin{align*} \qrfac{a_1, a_2,\dots, a_r}{k} &:= \qrfac{a_1}{k} \qrfac{a_2}{k}\cdots \qrfac{a_r}{k}. \end{align*} We will require the $q$-binomial coefficient, defined as $$\genfrac{[}{]}{0pt}{}{n}{k}_q := \frac{\qrfac{q}{n}}{\qrfac{q}{k}\qrfac{q}{n-k}},$$ where $n\geq k$ are non-negative integers. When $n<k$, we take $\genfrac{[}{]}{0pt}{}{n}{k}_q=0.$ \subsection*{Ramanujan's entries} We record the notation used to refer to Ramanujan's entries from \cite{RamanujanNB, Ramanujan-LN}. An entry takes its number from either one of the five volumes of Ramanujan's Notebooks edited by Berndt, or the Lost Notebook edited by Andrews and Berndt. We have prefixed the volume number and/or the chapter to the entry. For example, Entry III.16.39 refers to Entry 39, of Chapter 16 in Part III of Berndt's five volumes on {\em Ramanujan's notebooks}; that is, it refers to \cite[Ch.~16, Entry 39]{Berndt1991}. We place an additional L when referring to the Lost Notebook. Thus, L.I.6.2.1 refers to Entry 6.2.1 in Part I of the {\em Ramanujan's lost notebook} edited by Andrews and Berndt, that is, \cite[Entry 6.2.1]{AB2005}. \section{A more general continued fraction}\label{g-sec} In his second letter to Hardy, Ramanujan~\cite[p.~57]{BR1995} mentions that Entry~III.16.29~(i) is ``a particular case of a theorem on the continued fraction $$\frac{1}{1}\genfrac{}{}{0pt}{}{}{+} \frac{ax}{1}\genfrac{}{}{0pt}{}{}{+}\frac{ax^2}{1}\genfrac{}{}{0pt}{}{}{+}\frac{ax^3}{1}\genfrac{}{}{0pt}{}{}{+}\frac{ax^4}{1}\genfrac{}{}{0pt}{}{}{+} \frac{ax^5}{1}\genfrac{}{}{0pt}{}{}{+} \genfrac{}{}{0pt}{}{}{\cdots} ,$$ which is a particular case of the continued fraction $$ \frac{1}{1}\genfrac{}{}{0pt}{}{}{+}\frac{a x}{1+bx}\genfrac{}{}{0pt}{}{}{+}\frac{a x^2}{1+bx^2}\genfrac{}{}{0pt}{}{}{+}\frac{a x^3}{1+bx^3}\genfrac{}{}{0pt}{}{}{+} \genfrac{}{}{0pt}{}{}{\cdots}, $$ which is a particular case of a general theorem on continued fractions.'' A theorem regarding this continued fraction involves the sum \begin{equation} g(b,\lambda):=\sum_{k=0}^{\infty}\frac{\lambda^k q^{k^2}}{\qrfac{q,-bq}{k}}.\label{g} \end{equation} For the expressions corresponding to \eqref{g} in the letter, Ramanujan had $x$ in place of $q$, and $\lambda$ in place of $a$. Ramanujan records the following continued fractions in the Lost Notebook \cite[p.~40]{Ramanujan-LN} \subsection*{L.I.6.3.1 (i), (ii), (iii)} Let $|q|<1$. Then \begin{subequations} \begin{align} \frac{g(b,\lambda q)}{g(b,\lambda)} & =\frac{1}{1}\genfrac{}{}{0pt}{}{}{+}\frac{\lambda q}{1}\genfrac{}{}{0pt}{}{}{+}\frac{bq+\lambda q^2 }{1}\genfrac{}{}{0pt}{}{}{+}\frac{\lambda q^3}{1}\genfrac{}{}{0pt}{}{}{+} \frac{bq^2+\lambda q^4}{1}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} \label{g-cfrac1} \\ &=\frac{1}{1}\genfrac{}{}{0pt}{}{}{+}\frac{\lambda q}{1+bq}\genfrac{}{}{0pt}{}{}{+}\frac{\lambda q^2}{1+bq^2}\genfrac{}{}{0pt}{}{}{+}\frac{\lambda q^3}{1+bq^3}\genfrac{}{}{0pt}{}{}{+} \frac{\lambda q^4}{1+bq^4}\genfrac{}{}{0pt}{}{}{+} \genfrac{}{}{0pt}{}{}{\cdots} \label{g-cfrac2}.\\ \intertext{Let $|q|<1$ and $|b|< 1$. Then} \frac{g(b,\lambda q)}{g(b,\lambda)} &= \frac{1}{1-b}\genfrac{}{}{0pt}{}{}{+}\frac{b+\lambda q}{1-b}\genfrac{}{}{0pt}{}{}{+}\frac{b+\lambda q^2}{1-b}\genfrac{}{}{0pt}{}{}{+}\frac{b+\lambda q^3}{1-b}\genfrac{}{}{0pt}{}{}{+} \frac{b+\lambda q^4}{1-b}\genfrac{}{}{0pt}{}{}{+} \genfrac{}{}{0pt}{}{}{\cdots}. \label{g-cfrac3} \end{align} \end{subequations} \begin{Remarks} A continued fraction of Touchard~\cite{Touchard1952} can be considered to be a special case of \eqref{g-cfrac3}, see Prodinger~\cite{Prodinger2012}. The first two continued fractions can be obtained as special cases of the general continued fractions in \S\ref{G-sec}. We prove the third. \end{Remarks} \begin{proof}[Proof of \eqref{g-cfrac3}] We present J.~Cigler's proof (sent by email, Sept.~2012), which illustrates an idea of Askey~\cite{Askey1989} regarding a discovery approach to the Rogers--Ramanujan identities beginning with \eqref{RRcfrac1} (see also \cite{GB2015}). Let $F(b,\lambda)$ denote the continued fraction \eqref{g-cfrac3}. Then $$F(b,\lambda)=\frac{1}{1-b+(b+\lambda q)F(b,\lambda q)}.$$ To change this non-linear equation into a linear one, we substitute $$F(b,\lambda)=\frac{g(b,\lambda q)}{g(b,\lambda )},$$ to obtain Ramanujan's recurrence relation given in Entry L.I.6.3.1(iv): \begin{equation*} g(b,\lambda) = (1-b)g(b,\lambda q) +(b+\lambda q) g(b,\lambda q^2). \end{equation*} Now assume that $$g(b,\lambda)=\sum_{n=0}^\infty a_n\lambda^n,$$ substitute in Ramanujan's recurrence relation, and compare coefficients of $\lambda^n$, to obtain $$a_n = (1-b)q^n a_n + bq^{2n} a_n +q^{2n-1}a_{n-1},$$ or $$ a_n = \frac{q^{2n-1}}{(1-q^n)(1+bq^n)} a_{n-1}. $$ Iterating this we obtain $$ a_n = \frac{q^{n^2}}{\pqrfac{q, -bq}{n}{q}} a_0.$$ Now taking $a_0=1$, we obtain the expression \eqref{g} for $g(b,\lambda)$ and a (formal) proof of \eqref{g-cfrac3}. \end{proof} The continued fraction \eqref{g-cfrac2} appeared earlier as Entry III.16.15. When $b=0$ this reduces to a relation for an extension of the Rogers--Ramanujan continued fraction. \subsection*{Cor. to III.16.15} If $|q|<1$, then \begin{equation*}\label{cor-entry15} \frac{\displaystyle\sum_{k=0}^{\infty} \frac{\lambda^k q^{k^2+k}}{\qrfac{q}{k}}} {\displaystyle\sum_{k=0}^{\infty} \frac{\lambda^k q^{k^2}}{\qrfac{q}{k}}} = \frac{1}{1}\genfrac{}{}{0pt}{}{}{+} \frac{\lambda q}{1}\genfrac{}{}{0pt}{}{}{+}\frac{\lambda q^2}{1}\genfrac{}{}{0pt}{}{}{+}\frac{\lambda q^3}{1} \genfrac{}{}{0pt}{}{}{+}\frac{\lambda q^4}{1}\genfrac{}{}{0pt}{}{}{+} \genfrac{}{}{0pt}{}{}{\cdots} . \end{equation*} When $\lambda=1$, this reduces to the Rogers--Ramanujan continued fraction. In the very next entry Ramanujan gives a formula for the convergents of the (reciprocal of) this continued fraction. \subsection*{III.16.16} \begin{equation*} \frac{\mu_n(0)}{\mu_n(1)} = 1+ \frac{\lambda q}{1}\genfrac{}{}{0pt}{}{}{+} \frac{\lambda q^2}{1}\genfrac{}{}{0pt}{}{}{+}\frac{\lambda q^3}{1} \genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} \genfrac{}{}{0pt}{}{}{+} \frac{\lambda q^n}{1} ,\label{entry16} \end{equation*} where $\mu_n(s)$ is defined as \begin{equation*}\label{mu1-finite} \mu_n(s):= \sum_{k=0}^{\infty} q^{k^2+sk}\lambda^k \genfrac{[}{]}{0pt}{}{n-k-s+1}{k}_{q} , \end{equation*} for $s=0, 1, 2, \dots.$ \begin{Remarks}\ In our proof we illustrate a method due to Euler~\cite{LE1788-616} (see \cite{GB2014, BI2021}), that works on all the general $q$-continued fractions mentioned in this article. Euler's approach is based on using the elementary identity \begin{equation*}\label{div-1step} \frac{N}{D}=1+\frac{N-D}{D} \end{equation*} to \lq divide' two series, both of whose first terms are $1$. \end{Remarks} \begin{proof} [Proof of III.16.16] Note that the sum $\mu_n(s)$ is finite; the index $k$ goes from $0$ to $ \lfloor \frac{n-s+1}{2}\rfloor$. We will show \begin{equation}\label{gn-recursion} \frac{\mu_n(s)}{\mu_n(s+1)}= 1+\frac{\lambda q^{s+1}}{{\cfrac{\mu_n(s+1)}{\mu_n(s+2)}}}, \end{equation} for $s=0,1, 2, 3, \dots, n-1$. Formula \eqref{entry16} follows by taking $s=0, 1, \dots, n-1$ in turn. Observe that the iteration stops when $s=n-1$, because \begin{equation* \mu_n(n)=1=\mu_n(n+1). \end{equation*} To prove \eqref{gn-recursion}, note that \begin{align* \frac{\mu_n(s)}{\mu_n(s+1)} &=1+\frac{\mu_n(s)-\mu_n(s+1)}{\mu_n(s+1)} \\ & = 1+\frac{1}{\mu_n(s+1)} \\ &\hspace{0.35cm}\times \displaystyle\sum_{k=0}^{\infty} \frac{q^{k^2+sk}\lambda^k\qrfac{q}{n-k-s}}{\qrfac{q}{k}\qrfac{q}{n-2k-s+1}} \left[(1-q^{n-k-s+1})- (1-q^{n-2k-s+1})q^k\right] \\ &= 1+ \frac{1}{\mu_n(s+1)} \displaystyle\sum_{k=1}^{\infty} \frac{q^{k^2+sk}\lambda^k}{\qrfac{q}{k-1}} \frac{\qrfac{q}{n-k-s}}{\qrfac{q}{n-2k-s+1}} \\ &= 1+\frac{\lambda q^{s+1}}{\mu_n(s+1)} \displaystyle\sum_{k=0}^{\infty} \frac{q^{k^2+(s+2)k)}\lambda^{k}}{\qrfac{q}{k}} \frac{\qrfac{q}{n-k-s-1}}{\qrfac{q}{n-2k-s-1}}, \end{align*} on shifting the sum. From this \eqref{gn-recursion} follows and the proof is complete. \end{proof} \subsection*{III.16.13 and Corollary 6.2.4 to L.I.6.2.1(Eisenstein)}If $|q|<1$, then \begin{equation*} \sum_{k=0}^{\infty} (-a)^k q^{\frac{k(k+1)}{2}}= \frac{1}{1}\genfrac{}{}{0pt}{}{}{+}\frac{aq}{1}\genfrac{}{}{0pt}{}{}{+}\frac{a(q^2 - q)}{1}\genfrac{}{}{0pt}{}{}{+}\frac{aq^3}{1}\genfrac{}{}{0pt}{}{}{+}\frac{a(q^4-q^2)}{1} \genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} ,\label{entry13} \end{equation*} \begin{Remarks} Eisenstein's continued fraction appears in \cite[pp.~35--39]{Eisenstein1975}. The case $a=1$ was known to Gauss in 1797, see \cite[p.~152]{AB2005}. See Folsom~\cite{Folsom2006} for a history, and an account of Eisenstein's continued fractions. \end{Remarks} \begin{proof}[Proof outline] The special case $\lambda=a$ and $b=-a$ of \eqref{g-cfrac1} can be written as Entry III.16.13, where to obtain the sum on the left hand side, we have to apply the transformation formula III.16.9 to the sums. \end{proof} Ramanujan (see \cite{Berndt1991}) noted formulas for the denominators of the convergents of this continued fraction. They are given by the following: \begin{align*} D_{2n} &= \sum_{k=0}^n a^{k} q^{nk} \genfrac{[}{]}{0pt}{}{n}{k}_q, \\ \intertext{and} D_{2n+1} &= \sum_{k=0}^n a^{k} q^{(n+1)k} \genfrac{[}{]}{0pt}{}{n}{k}_q . \end{align*} \section{Still more general}\label{G-sec} Some candidates for what Ramanujan may have meant by the phrase `general theorem on continued fractions' in his letter, concern the sum \begin{equation}\label{G} G(a,b,\lambda):=\sum_{k=0}^{\infty} \frac{\qrfac{-\lambda/a}{k} \, a^k\, q^{(k^2+k)/2}}{\qrfac{q, -bq}{k}}. \end{equation} \begin{subequations} However, only special cases of these theorems appear in Ramanujan's notebooks, which is the only record of his discoveries before he wrote the aforementioned letter to Hardy. In the lost notebook, Ramanujan gave the following continued fractions and derived numerous special cases. \subsection*{L.I.6.2.1 and L.I.6.4.1} Let $a$, $b$, $\lambda$, and $q$ be complex numbers, with $|q|<1$. Then \begin{align}\label{G-cfrac1} &\hspace{-1cm}\frac{G(aq, b, \lambda q)}{G(a, b, \lambda)}\cr &=\frac{1}{1}\genfrac{}{}{0pt}{}{}{+}\frac{aq+\lambda q}{1}\genfrac{}{}{0pt}{}{}{+} \frac{bq+\lambda q^2 }{1}\genfrac{}{}{0pt}{}{}{+}\frac{aq^2+\lambda q^3}{1}\genfrac{}{}{0pt}{}{}{+}\frac{bq^2+\lambda q^4}{1} \genfrac{}{}{0pt}{}{}{\cdots} \\ &= \frac{1}{1+aq}\genfrac{}{}{0pt}{}{}{+}\frac{\lambda q - abq^2}{1+aq^2+bq}\genfrac{}{}{0pt}{}{}{+}\frac{\lambda q^2 -abq^4 } {1+aq^3+bq^2}\genfrac{}{}{0pt}{}{}{+}\frac{\lambda q^3-abq^6}{1+aq^4+bq^3}\genfrac{}{}{0pt}{}{}{+} \genfrac{}{}{0pt}{}{}{\cdots} . \label{G-cfrac2} \end{align} \begin{Remarks} \ \begin{enumerate} \item The continued fraction in \eqref{G-cfrac2}, Entry L.I.6.4.1, appears a few pages after \eqref{G-cfrac1} in the Lost Notebook \cite{Ramanujan-LN}, as a transformation formula given by the second equality above. \item Note that if we take $a=0$ in \eqref{G-cfrac1} we obtain \eqref{g-cfrac1}; and taking $a=0$ in \eqref{G-cfrac2} we obtain \eqref{g-cfrac2}. But there is no general continued fraction in Ramanujan's notebooks for ${G(aq, b, \lambda q)}/{G(a, b, \lambda)}$ which reduces to \eqref{g-cfrac3}. A candidate for such a continued fraction is ~\cite[Theorem~6.4.1]{AB2005} \begin{multline}\label{G-cfrac3} \frac{G(aq, b, \lambda q)}{G(a, b, \lambda)} =\\ \frac{1}{1}\genfrac{}{}{0pt}{}{}{+}\frac{aq+\lambda q}{1-aq+bq}\genfrac{}{}{0pt}{}{}{+}\frac{aq+\lambda q^2} {1-aq+bq^2}\genfrac{}{}{0pt}{}{}{+} \frac{aq+ \lambda q^3}{1-aq+bq^3}\genfrac{}{}{0pt}{}{}{+} \genfrac{}{}{0pt}{}{}{\cdots}, \\ \text{ where } |aq|<1 . \end{multline} When $b=0$ and $a\mapsto b/q$, this reduces to \eqref{g-cfrac3}. This continued fraction appears in different forms in Hirschhorn \cite{MDH1980} and Bhargava and Adiga \cite{BA1984}. We need to appeal to Entry III.16.8 in order to match the two forms. \item In addition to \eqref{G-cfrac1}, \eqref{G-cfrac2}, and \eqref{G-cfrac3}, there are two more continued fractions for ${G(aq, b, \lambda q)}/{G(a, b, \lambda)}$ in \cite{GB2014} (see also \cite{LMS2017}). \item Another candidate for a `general theorem on continued fractions' is Andrews \cite[Theorem~6]{Andrews1968}. The most general continued fraction of this type is due to Andrews and Bowman~\cite{AndBow1995}; see also Gupta and Masson~\cite{GM1999}. \end{enumerate} \end{Remarks} \end{subequations} \begin{proof}[On proofs of \eqref{G-cfrac1}--\eqref{G-cfrac3}] Andrews \cite{andrews1979} included a proof of \eqref{G-cfrac1} in his introductory article reporting his discovery of the Lost Notebook. Bhargava and Adiga~\cite{BA1984} proved \eqref{G-cfrac2} and \eqref{G-cfrac3}. \end{proof} There are several special cases of these continued fractions noted by Ramanujan. In each of the following, we assume the conditions of L.I.6.2.1 (if applicable). \subsection*{L.I.6.2.3} \begin{equation*}\label{G-cfrac1-cor} \frac{G(a, b, 0)}{G(aq, b, 0)} =1+\frac{aq}{1}\genfrac{}{}{0pt}{}{}{+}\frac{bq }{1}\genfrac{}{}{0pt}{}{}{+}\frac{aq^2}{1}\genfrac{}{}{0pt}{}{}{+}\frac{bq^2}{1} \genfrac{}{}{0pt}{}{}{\cdots}. \end{equation*} \begin{proof} Set $\lambda =0$ in \eqref{G-cfrac1} and take reciprocals. \end{proof} \subsection*{Cor. 6.2.9 to L.I.6.2.1} \begin{equation*} \sum_{k=0}^\infty (-1)^k q^{3k^2+2k} \big(1+q^{2k+1}\big) = \frac{1}{1}\genfrac{}{}{0pt}{}{}{+}\frac{q^2-q}{1}\genfrac{}{}{0pt}{}{}{+}\frac{ q^4 - q^2}{1}\genfrac{}{}{0pt}{}{}{+}\frac{q^6-q^3}{1}\genfrac{}{}{0pt}{}{}{+} \frac{ q^8 - q^4}{1}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} . \label{LNIcor6.2.9} \end{equation*} \begin{proof} We take $q\mapsto q^2$ and $a=-1/q$, $b=-1$, and $\lambda =1$ in \eqref{G-cfrac1}. The sum side requires an iterate of Heine's transformation formula, and L.I.9.5.1, see \cite[p.~155]{AB2005}. Heine's formula is Entry III.16.6. \end{proof} \subsection*{Cor.~6.2.11 to Entry L.I.6.2.1} \begin{equation*} 1-\sum_{k=1}^\infty q^{k(3k-1)/2} \big(1-q^{k}\big) = \frac{2}{2}\genfrac{}{}{0pt}{}{}{+}\frac{q+q}{1}\genfrac{}{}{0pt}{}{}{+}\frac{ q^2 + q^3}{1}\genfrac{}{}{0pt}{}{}{+}\frac{q^3+q^5}{1}\genfrac{}{}{0pt}{}{}{+} \frac{ q^4 + q^7}{1}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} . \label{LNIcor6.2.11} \end{equation*} \begin{proof} Take $q\mapsto q^2$, $a= 1/q$, $b=1$, and $\lambda = 1/q,$ in \eqref{G-cfrac1}. The sum side requires elementary series manipulations, a special case of the $q$-binomial theorem (III.16.2), the second iterate of Heine's transformation (\cite[eq.~(III.3)]{GR90}) and Entry L.I.9.4.7. See \cite[p.~156]{AB2005} for the details. \end{proof} There are further special cases in the next two sections. \section{Transformations of continued fractions} \label{sec:cf-trans} In \S\S\ref{g-sec} and \ref{G-sec}, we saw that a ratio of two series can be written in multiple ways as continued fractions. In this section we note two transformation formulas recorded by Ramanujan, obtained by setting the same special case of two of these continued fractions equal to one another. In addition, there is one transformation which follows by taking the odd part of a continued fraction. \subsection*{L.I.6.5.1} Let $k\geq 0$, $\alpha = (1+\sqrt{1+4k} )/2$, and $\beta = (-1+\sqrt{1+4k} )/2$. Then for $|q|<1$, \begin{subequations} \begin{align} \frac{1}{1}\genfrac{}{}{0pt}{}{}{+} & \frac{k+q}{1}\genfrac{}{}{0pt}{}{}{+} \frac{k+q^2}{1}\genfrac{}{}{0pt}{}{}{+} \frac{k+q^3}{1}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} \label{LI.6.5.1a} \\ &= \frac{1}{\alpha}\genfrac{}{}{0pt}{}{}{+} \frac{q}{\alpha+\beta q}\genfrac{}{}{0pt}{}{}{+} \frac{q^2}{\alpha+\beta q^2}\genfrac{}{}{0pt}{}{}{+} \frac{q^3}{\alpha+\beta q^3} \genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} .\label{LI.6.5.1b} \end{align} \end{subequations} \begin{proof} Observe that $\alpha\beta=k$ and $\alpha-\beta =1$. To show this transformation formula we take a special case $\lambda = 1/\alpha^2$, $b=\beta/\alpha$ in \eqref{g-cfrac2} and \eqref{g-cfrac3} and equate the two resulting continued fractions. After substituting in \eqref{g-cfrac2}, we multiply and divide each fraction by $\alpha$, and simplify to obtain $\alpha$ times \eqref{LI.6.5.1b}. Similarly, we obtain $\alpha$ times \eqref{LI.6.5.1a} from \eqref{g-cfrac3}. \end{proof} \subsection*{L.I.6.5.2} For $|q|<1$, \begin{align*} \frac{1}{1}\genfrac{}{}{0pt}{}{}{+} & \frac{2+q}{1}\genfrac{}{}{0pt}{}{}{+} \frac{2+q^2}{1}\genfrac{}{}{0pt}{}{}{+} \frac{2+q^3}{1}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} \cr &= \frac{1}{2}\genfrac{}{}{0pt}{}{}{+} \frac{q}{2+ q}\genfrac{}{}{0pt}{}{}{+} \frac{q^2}{2+q^2}\genfrac{}{}{0pt}{}{}{+} \frac{q^3}{2+ q^3} \genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots}. \end{align*} \begin{proof} This is a special case of Entry L.I.6.5.1, where $k=2$, so that $\alpha=2$ and $\beta =1$. \end{proof} \subsection*{L.I.6.4.2} If $|q|<1$, then \begin{align* \frac{1}{a+c}\genfrac{}{}{0pt}{}{}{-} & \frac{ab}{a+b+cq}\genfrac{}{}{0pt}{}{}{-}\genfrac{}{}{0pt}{}{}{\cdots}\genfrac{}{}{0pt}{}{}{-} \frac{ab}{a+b+cq^n}\genfrac{}{}{0pt}{}{}{-}\genfrac{}{}{0pt}{}{}{\cdots} \\ &= \frac{1}{c-b+a}\genfrac{}{}{0pt}{}{}{+} \frac{bc}{c-b+a/q}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots}\genfrac{}{}{0pt}{}{}{+} \frac{bc}{c-b+a/q^n}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} \end{align*} \begin{proof} Take $\lambda=0$, $a\mapsto -b/aq$, $b\mapsto c/a$ in \eqref{G-cfrac3} and \eqref{G-cfrac2}, and compare the resulting continued fractions. The transformation formula is obtained after some simplification, see \cite[p.~161]{AB2005} for more details. \end{proof} \subsection*{L.I.6.4.3} For $q\neq 0$, $n$ a positive integer, we have \begin{align}\label{LI6.4.3} 1+\frac{a}{1}\genfrac{}{}{0pt}{}{}{+} & \frac{b}{q}\genfrac{}{}{0pt}{}{}{+}\frac{a}{1}\genfrac{}{}{0pt}{}{}{+} \frac{b}{q^2}\genfrac{}{}{0pt}{}{}{\cdots}\genfrac{}{}{0pt}{}{}{+} \frac{b}{q^n}\genfrac{}{}{0pt}{}{}{+}\frac{a}{1} \cr &= 1+a - \frac{ab}{a+b+q}\genfrac{}{}{0pt}{}{}{-} \frac{ab}{a+b+q^2}\genfrac{}{}{0pt}{}{}{-}\genfrac{}{}{0pt}{}{}{\cdots}\genfrac{}{}{0pt}{}{}{-} \frac{ab}{a+b+q^n}. \end{align} \begin{proof} The proof is elementary. The continued fraction on the left hand side of \eqref{LI6.4.3} is of the form $1+{N_{2n+1}}/{D_{2n+1}}$, where $$\frac{N_{2n+1}}{D_{2n+1}} = \frac{a_1}{b_1}\genfrac{}{}{0pt}{}{}{+}\frac{a_2}{b_2}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} \genfrac{}{}{0pt}{}{}{+} \frac{a_{2n+1}}{b_{2n+1}}.$$ We consider instead the continued fraction with convergents $N_k/D_k$. The continued fraction on the right hand side is obtained by adding $1$ to the continued fraction with convergents $N_1/D_1$, $N_3/D_3$, $\dots$, ${N_{2n+1}}/{D_{2n+1}}$. This is known as the odd part of the continued fraction. The details are as follows. An equivalent form of Ramanujan's Entry II.12.1 is that $N_k$ and $D_k$ are both determined by a three-term recurrence relation of the form \begin{align*} P_k &= b_kP_{k-1} +a_kP_{k-2}, \end{align*} for $k = 3, 4, \dots, 2n+1$, with initial values $N_1=a_1$, $D_1=b_1$, $N_2=a_1b_1$ and $D_2=a_2+b_1b_2$. It is easy to see that $N_{2k+1}$ and $D_{2k+1}$ (for $k=2, 3, \dots, n$) also satisfy the same recurrence relation with $a_k \mapsto - {a_{2k-1}a_{2k}b_{2k+1}}/{b_{2k-1}}$, and $b_k \mapsto a_{2k+1}+b_{2k}b_{2k+1}+{a_{2k}b_{2k+1}}/{b_{2k-1}}.$ Now we substitute the values of $a_k$ and $b_k$ from the left hand side of \eqref{LI6.4.3}, and find the corresponding recurrence for the odd part of the continued fraction. On adding $1$ to the corresponding continued fraction, we obtain the right hand side of \eqref{LI6.4.3}. \end{proof} \section{Continued fractions written as products}\label{sec:cf-prods} So far we have seen continued fractions written as a ratio of two series. Many of Ramanujan's $q$-continued fractions are ratios of infinite products. It turns out that all of these are obtained when the series in question can be written as infinite products. An example is the Rogers--Ramanujan continued fraction highlighted in the introduction, which has the three expressions given by formulas \eqref{RRcfrac1}, \eqref{rr-cfrac1} and \eqref{entry16.38.iii}. In this section, we list such special cases of the general continued fractions given so far. In all the formulas in this section, we assume $|q|<1$. \subsection*{Corollary~6.2.2 to L.I.6.2.1} For any complex number $a\neq -q^{2n+1}$, $n\ge 1$ \begin{equation} \frac{\pqrfac{-aq^2}{\infty}{q^2}}{(-aq; q^2)^{ 2}_{\infty}} = \frac{1}{1}\genfrac{}{}{0pt}{}{}{+}\frac{aq}{1}\genfrac{}{}{0pt}{}{}{+}\frac{ q + aq^2}{1}\genfrac{}{}{0pt}{}{}{+}\frac{aq^3}{1}\genfrac{}{}{0pt}{}{}{+} \frac{ q^2 + aq^4}{1}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} . \label{LNIcor6.2.2} \end{equation} \begin{proof} Take $a=0$, $b=1$ and $\lambda\mapsto a$ in \eqref{G-cfrac1}. This requires the $q$-binomial theorem (Entry III.16.2) to sum both the sums. \end{proof} The next two continued fractions are further special cases of this continued fraction. \subsection*{V.32.21; Corollary~6.2.1 to L.I.6.2.1} \begin{equation*} \frac{\pqrfac{-q^2}{\infty}{q^2}}{(-q; q^2)_{\infty}} = \frac{\pqrfac{q}{\infty}{q^2}}{(q^2; q^4)^{ 2}_{\infty}} = \frac{1}{1}\genfrac{}{}{0pt}{}{}{+}\frac{q}{1}\genfrac{}{}{0pt}{}{}{+}\frac{ q + q^2}{1}\genfrac{}{}{0pt}{}{}{+}\frac{q^3}{1}\genfrac{}{}{0pt}{}{}{+} \frac{ q^2 + q^4}{1}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} . \label{LNIcor6.2.1} \end{equation*} \begin{proof} Take $a = 1$ in \eqref{LNIcor6.2.2}. \end{proof} \subsection*{Corollary~6.2.10 to L.I.6.2.1} \begin{equation} \frac{\pqrfac{-q^3}{\infty}{q^4}}{(-q; q^4)_{\infty}} = \frac{1}{1}\genfrac{}{}{0pt}{}{}{+}\frac{q}{1}\genfrac{}{}{0pt}{}{}{+}\frac{ q^2 + q^3}{1}\genfrac{}{}{0pt}{}{}{+}\frac{q^5}{1}\genfrac{}{}{0pt}{}{}{+} \frac{ q^4 + q^7}{1}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} . \label{LNIcor6.2.10} \end{equation} \begin{proof} Take $q\mapsto q^2$ and $a= 1/q$ in \eqref{LNIcor6.2.2}. \end{proof} Another continued fraction for the LHS of \eqref{LNIcor6.2.10} is as follows. \subsection*{V.32.20} \begin{equation*} \frac{\pqrfac{q^3}{\infty}{q^4}}{(q; q^4)_{\infty}} = \frac{1}{1}\genfrac{}{}{0pt}{}{}{-}\frac{q}{1+q^2}\genfrac{}{}{0pt}{}{}{-}\frac{ q^3}{1+q^4}\genfrac{}{}{0pt}{}{}{-}\frac{q^5}{1+q^6}\genfrac{}{}{0pt}{}{}{-} \frac{q^7}{1+q^8}\genfrac{}{}{0pt}{}{}{-} \genfrac{}{}{0pt}{}{}{\cdots} . \end{equation*} \begin{proof} This is obtained from \eqref{g-cfrac2} by taking $q\mapsto q^2$, $\lambda = -1/q$, $b=1$. The sums are special case of the $q$-binomial theorem (the case $a\mapsto 0$ of Entry III.16.2, with $q\mapsto q^4$). \end{proof} Next, we have two more special cases of \eqref{G-cfrac1}. \subsection*{Cor.~6.2.7 to L.I.6.2.1; V.32.18} \begin{equation*} \frac{\pqrfac{q, q^5}{\infty}{q^6}}{(q^3; q^6)^2_{\infty}} = \frac{\pqrfac{q}{\infty}{q^2}}{(q^3; q^6)^3_{\infty}} = \frac{1}{1}\genfrac{}{}{0pt}{}{}{+}\frac{q+q^2}{1}\genfrac{}{}{0pt}{}{}{+}\frac{ q^2 + q^4}{1}\genfrac{}{}{0pt}{}{}{+}\frac{q^3+q^6}{1}\genfrac{}{}{0pt}{}{}{+} \frac{ q^4 + q^8}{1}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} . \label{LNIcor6.2.7} \end{equation*} \begin{Remark} This continued fraction is known as Ramanujan's cubic continued fraction. See Chan \cite{Chan1995} and \cite[Chapter 3]{AB2005}. \end{Remark} \begin{proof} Take $q\mapsto q^2$, and take $a=1/q$, $b=1$, and $\lambda =1$ in \eqref{G-cfrac1}. To obtain the product side, we require some identities of Slater \cite{Slater1951, Slater1952}. For the details, see \cite[p.~154]{AB2005}. \end{proof} \subsection*{V.32.22; V.32.23; Cor.~6.2.8 to L.I.6.2.1} \begin{equation*} \frac{\pqrfac{q, q^7}{\infty}{q^8}}{(q^3, q^5; q^8)_{\infty}} = \frac{1}{1}\genfrac{}{}{0pt}{}{}{+}\frac{q+q^2}{1}\genfrac{}{}{0pt}{}{}{+}\frac{ q^4}{1}\genfrac{}{}{0pt}{}{}{+}\frac{q^3+q^6}{1}\genfrac{}{}{0pt}{}{}{+} \frac{ q^8}{1}\genfrac{}{}{0pt}{}{}{+}\genfrac{}{}{0pt}{}{}{\cdots} . \label{LNIcor6.2.8} \end{equation*} \begin{proof} Take $q\mapsto q^2$, and take $a=1/q$, $b=0$, and $\lambda =1$ in \eqref{G-cfrac1}. To obtain the product side, we require some identities of Slater \cite{Slater1952}. See \cite[p.~154]{AB2005} for more details. This continued fraction is referred to as the Ramanujan--G{\" o}llnitz--Gordon continued fraction. \end{proof} \subsection*{V.32.19} \begin{equation} \frac{\pqrfac{q^2}{\infty}{q^3}}{\pqrfac{q}{\infty}{q^3}} = \frac{1}{1}\genfrac{}{}{0pt}{}{}{-}\frac{q}{1+q}\genfrac{}{}{0pt}{}{}{-}\frac{ q^3}{1+q^2}\genfrac{}{}{0pt}{}{}{-}\frac{q^5}{1+q^3}\genfrac{}{}{0pt}{}{}{-} \frac{ q^7}{1+q^4}\genfrac{}{}{0pt}{}{}{-}\genfrac{}{}{0pt}{}{}{\cdots} . \label{V32.19} \end{equation} \begin{Remark} The product side of \eqref{V32.19} and a related continued fraction appeared in an intriguing claim of Ramanujan concerning a continued fraction that converges to three different limits. This claim and related ideas have been explored by Andrews, Berndt, Sohn, Yee and Zaharescu~\cite{ABSYZ2003, ABSYZ2005} and Ismail and Stanton \cite{IS2006}. \end{Remark} \begin{proof} The continued fraction in this entry can be found from \eqref{G-cfrac2}. Let $\omega = e^{2\pi i /3}$. Then take $\lambda =0$, $a=-\omega/q$, $b=-\omega^2$, to obtain the continued fraction $$\frac{1}{1-\omega}\genfrac{}{}{0pt}{}{}{-}\frac{q}{1+q} \genfrac{}{}{0pt}{}{}{-}\frac{ q^3}{1+q^2}\genfrac{}{}{0pt}{}{}{-}\genfrac{}{}{0pt}{}{}{\cdots},$$ from which \eqref{V32.19} can be obtained by taking reciprocals and adding $\omega$. A proof along these lines can be found in Berndt~\cite[p.~46]{Berndt1998}. See also \cite{ABSYZ2003}. See the recurrence relation for the function $G_2(s)$ in \cite[p.~64]{GB2014} for the connection with \eqref{G-cfrac2}. \end{proof} Note that the proofs of the last three entries require results that are not explicitly stated in Ramanujan's notebooks. In fact, these are the only three entries in this article with this property. \section{Entries III.16.11 and III.16.12}\label{sec:entry11-12} The first two $q$-continued fractions in Chapter 16 of Ramanujan's second notebook are Entries III.16.11 and III.16.12. Entry III.16.11 is motivated by a continued fraction recorded in Chapter 12 of Ramanujan's second notebook (see also \cite[Chapter~14, Vol.~1]{RamanujanNB}). In Entry II.12.18, Ramanujan finds a continued fraction for $$\frac{(x+1)^n - (x-1)^n}{(x+1)^n +(x-1)^n}.$$ Ramanujan notes four corollaries: continued fractions for $\tan^{-1} x$, $\log\frac{1+x}{1-x}$, $\tan x$, $ \frac{e^x-1}{e^x+1}$, so it would be natural to attempt an extension of II.14.18. We can expand $(x+1)^n$ and $(x-1)^n$ using the binomial theorem. If $n$ is a non-negative integer, then the numerator becomes a polynomial with terms of the form $a_kx^k$, where $k=n-1, n-3, \dots$; and the denominator has terms with $k=n, n-2, n-4, \dots$. Now the $q$-analog of the binomial theorem, given in Entry III.16.2, is \begin{equation*}\label{III.16.2} F(a,b):= \sum_{k=0}^\infty \frac{\qrfac{b/a}{k}}{\qrfac{q}{k}}a^k = \frac{\qrfac{b}{\infty}}{\qrfac{a}{\infty}}. \end{equation*} If we consider $$\frac{F(a,b)-F(-a,-b)}{F(a,b)+F(-a,-b)}$$ we obtain \begin{align*} \frac{\displaystyle \sum_{k=0}^{\infty}\frac{\qrfac{b/a}{2k+1}} {\qrfac{q}{2k+1}}a^{2k+1}} {\displaystyle \sum_{k=0}^{\infty}\frac{\qrfac{b/a}{2k}} {\qrfac{q}{2k}}a^{2k}} &= \frac{{\qrfac{-a, b}{\infty}}-{\qrfac{a, -b}{\infty}}} {{\qrfac{-a,b}{\infty}}+\qrfac{a, -b}{\infty}}. \end{align*} The numerator of the left hand side contains the odd terms of the $q$-binomial sum, and the denominator contains the even terms. The right hand appears in the following entry. \subsection*{III.16.11} Let $|q|<1$ and $|a|<1$. Then we have \begin{multline*} \frac{{\qrfac{-a, b}{\infty}}-{\qrfac{a, -b}{\infty}}} {{\qrfac{-a, b}{\infty}}+\qrfac{a, -b}{\infty}} = \frac{a-b}{1-q}\genfrac{}{}{0pt}{}{}{+}\frac{(a-bq)(aq-b)}{1-q^3}\genfrac{}{}{0pt}{}{}{+} \cr \frac{q(a-bq^2)(aq^2-b)}{1-q^5}\genfrac{}{}{0pt}{}{}{+} \frac{q(a-bq^3)(aq^3-b)}{1-q^7}\genfrac{}{}{0pt}{}{}{+} \genfrac{}{}{0pt}{}{}{\cdots} . \end{multline*} \begin{proof For a proof, see Berndt~\cite{Berndt1991}, which relies on ideas of Jacobsen \cite{LJ1989} to clarify an earlier proof by Adiga, Berndt, Bhargava and Watson~\cite{ABBW1985}. \end{proof} Entry III.16.12 has a very similar continued fraction. \subsection*{III.16.12} Let $|q|<1,$ and $|ab|<1$. Then we have \begin{multline*} \frac{\pqrfac{a^2q^3, b^2q^3}{\infty}{q^4}} {\pqrfac{a^2q, b^2q}{\infty}{q^4}} = \frac{1}{1-ab}\genfrac{}{}{0pt}{}{}{+}\frac{(a-bq)(b-aq)}{(1-ab)(1+q^2)}\genfrac{}{}{0pt}{}{}{+} \cr \frac{(a-bq^3)(b-aq^3)}{(1-ab)(1+q^4)}\genfrac{}{}{0pt}{}{}{+} \frac{(a-bq^5)(b-aq^5)}{(1-ab)(1+q^6)}\genfrac{}{}{0pt}{}{}{+} \genfrac{}{}{0pt}{}{}{\cdots}. \end{multline*} \begin{Remarks}\ \begin{enumerate} \item Entry III.16.12 is a $q$-analogue of Entry II.12.25. To see this, one has to use III.16.1(ii), which contains the definition of the $q$-gamma function. \item The continued fraction in Entry III.16.12 converges for $|ab|>1$. It also converges for $|ab|<1$ and $|q|>1$. The expression for both of these can be found using III.16.12, see \cite{Berndt1991, BI2021}. \end{enumerate} \end{Remarks} \begin{proof} For two inter-related proofs, see the work of Ismail and the author \cite{BI2021}. One of these proofs uses Euler's approach; and the other, a standard method from the theory of orthogonal polynomials. \end{proof} This continued fraction also arises from a ratio of two sums. From the $q$-binomial theorem, it follows that for $|q|<1$ and $|a|<1$: \begin{align*} \frac{\pqrfac{a^2q^3, b^2q^3}{\infty}{q^4}} {\pqrfac{a^2q, b^2q}{\infty}{q^4}} &= \frac{\displaystyle \sum_{k=0}^\infty \frac{\pqrfac{(bq/a)^2}{k}{q^4} }{\pqrfac{q^4}{k}{q^4}} (a^2q)^k } {\displaystyle \sum_{k=0}^\infty \frac{\pqrfac{(b/aq)^2}{k}{q^4} }{\pqrfac{q^4}{k}{q^4}} (a^2q^3)^k }. \end{align*} \section{The proofs of Ramanujan's $q$-continued fractions}\label{sec:proofs} The first proofs of many of Ramanujan's general $q$-continued fractions were given by Andrews \cite{andrews1979, An1981} and Adiga, Berndt, Bhargava and Watson \cite{ABBW1985}; proofs are collated in \cite{AB2005} and \cite{Berndt1991}. Even earlier, Selberg~\cite{Selberg1936} obtained many Rogers--Ramanujan type continued fractions, where the continued fractions are expressed as infinite products. Reuter~\cite{Reuter2014} has surveyed proofs of Ramanujan's hypergeometric continued fractions. As we have seen, an eclectic collection of techniques is used to prove Ramanujan's $q$-continued fractions. As Berndt~\cite{Berndt2010} wrote in a similar context: \begin{quote} Methods for proving these continued fraction formulas are varied and at times ad hoc. Ramanujan evidently had a systematic procedure for proving these continued fraction formulas, but we don't know what it is. \end{quote} Nevertheless, when it comes to $q$-continued fractions, there are several interesting systematic approaches. As we have seen, all of Ramanujan's general continued fractions (in \S\S\ref{g-sec}, \ref{G-sec}, \ref{sec:entry11-12}), can be obtained as a ratio of two similar series; the remaining results are obtained as special cases. In fact, Ramanathan~\cite{KGR1987a} derived all of these general continued fractions formally by finding three-term recurrence relations from contiguous relations for $_2\phi_1$ series. There is, however, a gap in this formal approach to continued fractions. One does not know that the continued fraction converges; and, even if it does, whether it converges to the ratio of series in question. Jacobsen~\cite{LJ1989} has clarified such issues (see \cite{LW1992} for a comprehensive exposition). Another way convergence has been addressed is by associating the continued fraction with orthogonal polynomials, see Ismail and Stanton~\cite{IS2006}. However, we require the parameters to be real (rather than complex) numbers in this approach. Continued fractions are closely related to the study of orthogonal polynomials (see, for example, Askey and Ismail~\cite{AI1984}). Al--Salam and Ismail \cite{Al-I1983} have studied orthogonal polynomials related to Ramanujan's continued fractions, see also \cite{BI2022, IS1997}. A third approach to show convergence is to use a theorem of Pincherle \cite[p.~164]{JT1980}, see, for examples, Masson~\cite{Masson1989} and Gupta and Masson~\cite{GM1999}. With regard to questions of convergence of Ramanujan's continued fractions, see also \cite{ABSYZ2003, ABSYZ2005, GB2014, BM2004, BM2007, IS2006}. In addition to Ramanathan~\cite{KGR1987a}, Lorentzen (n{\' e}e Jacobsen)~\cite{LL2008} suggested another systematic approach to Ramanujan's (hypergeometric) continued fractions. Yet another approach, following Euler's elementary idea works on all of Ramanujan's general $q$-continued fractions, see~\cite{GB2014, BI2021}. Sokal~\cite{Sokal2022} has developed Euler's idea further and given several examples. Hirsch\-horn~\cite{MDH1972, MDH1974, MDH1980} proves many of Ramanujan's continued fractions by finding analogous formulas for the convergents of continued fractions (see also Menon~\cite{Menon1965}, and \cite{BI2022, BMW2006, Prodinger2018}.) Regarding transformation formulas for continued fractions, Andrews and Berndt \cite{AB2005} used the Bauer--Muir transformation~\cite[p.~76--80]{LW1992} to prove Entry L.I.6.5.1. Lee, Mc Laughlin and Sohn~\cite{LMS2017} have applied this idea to Ramanujan's transformation formulas. By contrast, \cite{GB2014} explains the transformation of continued fractions \eqref{G-cfrac2} by transforming the series appearing on the left-hand side of \eqref{G-cfrac1}, and then obtaining the corresponding continued fractions by Euler's method. As we have seen, taking special cases of Ramanujan's general continued fractions involve the application of other transformation and summation formulas. In particular, in \S\ref{sec:cf-prods} we saw that the Rogers--Ramanujan type identities are very useful in expressing continued fractions as infinite products. The first paper with continued fractions of this kind was written by Selberg~\cite{Selberg1936}. Further examples appear in Gu and Prodinger~\cite{GP2011}. This prolific set of ideas surrounding Ramanujan's continued fractions is only to be expected. After all, as Hardy~\cite[p.~XXX]{Ramanujan-CW} famously remarked \begin{quote} (Ramanujan's) mastery of continued fractions was, on the formal side at any rate, beyond that of any mathematician in the world... \end{quote} \subsection*{Acknowledgements} A shorter version of this article appears in the Encyclopedia of Srinivasa Ramanujan and his Mathematics. We thank Krishnaswami Alladi, George Andrews, Bruce Berndt, Peter Paule, Ole Warnaar, and Ae Ja Yee for their helpful comments.

https://arxiv.org/abs/2005.05563

A classification of one dimensional affine rank three graphs

The rank three subgroups of a one-dimensional affine group over a finite field were classified in 1978 by Foulser and Kallaher. Although one can use their results for a classification of corresponding rank three graphs, the author did not find such a classification in a literature. The goal of this note is to present such a classification. It turned out that graph classification is much simpler than the group one. More precisely, it is shown that the graphs in the title are either the Paley graphs or one of the graphs constructed by Van Lint and Schrijver or by Peisert. Our approach is based on elementary group theory and does not use the classification of rank three affine groups.

\section{Main result and its proof} \vspace*{10mm} Recall that a {\it rank three graph} is an undirected graph $\Gamma$ the automorphism group of which has rank three. So, the arc set of such a graph coincides with one of 2-orbits (orbitals) of its automorphism group. Let $\F_q$ be a finite field of order $q=p^r$ where $p$ is a prime and $r$ is a positive integer. A graph $\Gamma$ with point set $\F_q$ will be called a {\it one dimensional affine rank three graph} if $\aut{\Gamma}\cap A\Gamma L(1,q)$ is a rank three group. In other words, the arc set of $\Gamma$ coincides with one of 2-orbits of a rank three subgroup $G \leq A\Gamma L(1,q)$. A rank three subgroup $G \leq A\Gamma L(1,q)$ contains a normal abelian subgroup which coincides with the additive group $\F_q^+$ of the field. It is a semidirect product $G = \F_q^+ G_0$ where $G_0$ is the stablilizer of the zero $0\in\F_q$. The group $G_0$ is contained in $\Gamma L_1(q)\cong\F_q^*\rtimes\langle \phi\rangle$ where $\phi:x\mapsto x^p$ is the Frobenius automorphism. Note that $G$ has rank three if and only if $G_0$ has two orbits in its natural action on $\F_q^*$. Let us denote these orbits as $O_1$ and $O_2$. Then the non-reflexive 2-orbits (orbitals) of $G$ are Cayley graphs over $\F_q^+$ the connection sets of which are $O_1$ and $O_2$. As was mentioned before, all subgroups of $\Gamma L_1(q)$ with two orbits on $\F_q^*$ were classified in \cite{FK}. But the orbits of these subgroups were not described there. Since orbit partitions of different groups may coincide, it could happen that the number of such partitions is much less than the number of groups. So, the classification of orbit partitions might be easier than the one of groups. Surprisingly, this is indeed the case. In this note we provide a complete classification of the orbits of two orbit affine subgroups. Our approach is straightforward and is not based on the classification of affine rank three groups obtained in \cite{FK}. In this note the group $\Gamma L_1(q)$ is considered as a permutation group in its natural action on $\F_q^*$. It is generated by permutations $\phi$ and $\hat{\alpha},\alpha\in\F_q^*$ where $\hat{\alpha}$ stands for the mapping $x\mapsto x\alpha,x\in\F_q^*$. Note that the mapping $\alpha\mapsto\hat{\alpha}\in A\Gamma L_1(q)$ is a group monomorphism and its image $\widehat{\F_q^*}$ is a cyclic subgroup of $A\Gamma L_1(q)$ which acts regularly on $\F_q^*$. We start with two main examples of rank three graphs. The first example was discovered in \cite{LS} and will be referred to as the Van Lint - Schrijver graph. It is described in the statement below which in fact coincides with Lemma 1 in \cite{LS}. We give here the proof to make the text self-contained. \begin{prop}\label{200320b} Let $p$ and $\ell$ be two primes satisfying ${\rm ord}_\ell(p)=\ell-1$. Pick an arbitrary positive integer $k$ and set $q := p^{(\ell-1)k}$. Let $\omega$ be a generator of $\F_q^*$. Then the subgroup $G_0=\langle \hat{\omega}^\ell, \phi\rangle\leq \Gamma L_1(q)$ has two orbits on $\F_q^*$: $\langle \omega^\ell\rangle, \F_q^*\setminus \langle \omega^\ell\rangle$. \end{prop} \begin{proof} It follows from $q= \left(p^{\ell-1}\right)^k$ that $p^{\ell-1} - 1$ divides $q-1$. Since $\ell$ divides $p^{\ell-1} - 1$, we conclude that $\ell\,|\,q-1$. Thus $C:=\langle \omega^\ell\rangle$ is a subgroup of $\F_q^*$ of index $\ell$. The subgroup $\widehat{C}$ has $\ell$ orbits on $\F_q^*$, namely: $$C, C\omega,...,C\omega^{\ell-1}.$$ The permutation $\hat{\omega}^\ell=\widehat{\omega^\ell}$ fixes each of the orbits setwise while the action of $\phi$ on them is equivalent to the permutation of $i\mapsto pi,i\in\Z_\ell$. It follows from the assumption ${\rm ord}_\ell(p)=\ell-1$ that the permutation $i\mapsto pi,i\in\Z_\ell$ has two orbits on $\Z_\ell$, namely:$\{0\}$ and $\Z_\ell^*$. Therefore $\phi$ fixes the orbit $C=\langle\omega^\ell\rangle$ and permutes transitively the others. \end{proof} {\bf Remark.} The subgroup $G_0=\langle \widehat{\omega^\ell}, \phi\rangle$ mentioned in Proposition~\ref{200320b} has a prime index $\ell$ in $\Gamma L_1(q)$, and, therefore, is maximal. The graph $\cay{\F_q^+}{C}$ is a particular case of a generalized Paley graph \cite{KP}. Since $\cay{\F_q^+}{C}$ is a rank three graph, it is strongly regular. In the case of $\ell=2$ it coincides with the classical Paley graph, see e.g. \cite{J}. If $\ell > 2$, then the graph is a Latin square graph for odd values of $k$ and a Negative Latin Square graph if $k$ is even \cite{LS}. \begin{prop}\label{200320c} Let $p\equiv\,3({\rm mod}\ 4)$ be a prime and $r$ an even positive integer. Denote by $\omega$ a generator of $\F_q^*$ where $q=p^r$. Then the subgroup $G_0=\langle \widehat{\omega^4},\phi\widehat{\omega}\rangle$ has two orbits on $\F_q^*$: $C\cup C\omega$ and $C\omega^2\cup C\omega^3$ where $C:=\langle\omega^4\rangle$ \end{prop} \begin{proof} The subgroup $\widehat{C}=\langle\widehat{\omega^4}\rangle$ has $4$ orbits on $\F_q^*$: $C,C\omega,C\omega^2,C\omega^3$. The element $\phi\widehat{\omega}$ permutes these orbits in the following way: $C\leftrightarrow C\omega, C\omega^2\leftrightarrow C\omega^3$. Now the claim follows. \end{proof} \noindent{\bf Remark.} Note that the group $G$ in Proposition~\ref{200320c} has index $4$ in $\Gamma L_1(q)$. There are three ways to coarsen the partition $C,C\omega,C\omega^2,C\omega^3$ into a partition with two classes of equal size: $$ \{C\cup C\omega^2,C\omega\cup C\omega^3\}, \{C\cup C\omega,C\omega^2\cup C\omega^3\}, \{C\cup C\omega^3,C\omega^2\cup C\omega\}. $$ The first partition is an orbit partition of the subgroup $\langle \widehat{\omega^2},\phi\rangle$ and yields the Paley graph, the second and the third one are orbit partitions of the groups $G_0$ and $G_0^\phi$, respectively. The latter two partitions were explicitly constructed by Peisert in \cite{P}. All three partitions yield strongly regular graphs with Paley parameters. The theorem below proves that the examples described above are the only one dimensional affine rank three graphs. \begin{theorem}\label{280320a} Let $\Gamma$ be a one dimensional affine rank three graph with point set $\F_q$. Then, up to a complement, $\Gamma$ is either the Van Lint - Schrijver, the Paley or the Peisert graphs. \end{theorem} The group $\Gamma L_1(q)$ is a semidirect product $\F_q^*\rtimes\langle \phi\rangle$ which acts naturally on $\F_q^*$. So, we may consider a bit more general action, namely, the one of the semidirect product $X:=\Z_n\rtimes\langle\alpha\rangle,\alpha\in\aut{\Z_n}$ on $\Z_n$: $$u^{(a,\alpha^i)} = \alpha^i (u) + a.$$ Pick an arbitrary subgroup $Y\leq X$ which has exactly two orbits on $\Z_n$, say $O_1$ and $O_2$ (we do not exclude the case when one of them is a singleton). Our goal is to describe all partitions of $\Z_n$ appearing in this way. Then we apply the obtained classification to prove Theorem~\ref{280320a}. To formulate the key lemma we need one more definition. A {\it radical} $\rad(S)$ of a subset $S\subseteq \Z_n$ is defined as the subgroup of $\Z_n$ consisting of all $u\in\Z_n$ satisfying $S+u=S$, \cite{MP}. Note that for any partition $\Z_n=O_1\cup O_2$ we have that $\rad(O_1)=\rad(O_2)$. Note that every subgroup $N$ of $\Z_n$ is normal in $X$. Its orbits coincide with its cosets and form an imprimitivity system for $X$. \begin{lemma}\label{140320b} Let $Y\leq X$ be a subgroup having two orbits on $\Z_n$, say $O_1$ and $O_2$, $|O_1|\leq |O_2|$. Denote by $m$ the index of $\rad(O_1)$ in $\Z_n$. Then one of the following holds \begin{enumerate} \item $m$ is prime, $O_1=m\Z_n$ and $\alpha$ permutes transitively non-zero elements of the factor group $\Z_n/m\Z_n\cong\Z_m$. \item $m=4$, $|O_1| = |O_2|$, $O_1$ is a union of two cosets of $4\Z_n$, $\alpha(x+4\Z_n) = -x+4\Z_n$ and $\{O_1,O_2\} = \{4\Z_n\cup i+4\Z_n, 2+4\Z_n\cup -i+4\Z_n\}, i\in\{1,3\}$. \end{enumerate} \end{lemma} \begin{proof} Write $\rad(O_1) = m\Z_n$. If $m\Z_n$ is nontrivial, then we can replace the original action of $X$ on $\Z_n$ by the action on the cosets $\Z_n/m\Z_n\cong\Z_m$. The resulting permutation group will be $\overline{X}=\Z_{\overline{n}}\rtimes\langle\overline{\alpha}\rangle$ where $\overline{n}=m$ and $\overline{\alpha}$ is the automorphism of $\Z_m$ induced by $\alpha$ . The image $\overline{Y}$ has two orbits on $\Z_{\overline{n}}$, namely: $\overline{O_1}$ and $\overline{O_2}$ where $\overline{O_j}=O_j/m\Z_n$. Note that $\rad(\overline{O_j})$ is trivial now. This observation reduces the general case to the one with a trivial radical. Thus starting from now we assume that $\rad(O_1)$ is trivial, i.e. $m=n$. Pick an arbitrary prime divisor $p$ of $n$ and consider the subgroup $P:=\frac{n}{p}\Z_n,|P|=p$ . The cosets $i+P,i=0,...,n/p-1$ are blocks of $X$ and, therefore, are blocks of $Y$ too. Thus $(i+P)\cap O_j$ are blocks of the transitive action of $Y$ on $O_j,j=1,2$. If some coset $i+P$ is contained in some $O_j$, then, it follows from transitivity of $Y$ on $O_j$ that $O_j$ is a union of $P$-cosets, contrary to the assumption $\rad(O_j)=\{0\}$. Hence, each coset $i+P$ intersects every orbit $O_j$ non-trivially. Since $(i+P)\cap O_j$ is a block of $Y$ and $Y$ is transitive on $O_j$, the number $|(i+P)\cap O_j|$ depends only on $j$. Therefore $|O_j|=\frac{n}{p}|(i+P)\cap O_j|$, and, consequently, $\frac{n}{p}$ divides $|O_j|$. If $n$ has two distinct prime divisors, say $p$ and $q$, then $|O_j|$ is divisible by $\frac{n}{p}$ and $\frac{n}{q}$. Therefore $|O_j|$ is divisible by $\lcm(\frac{n}{p},\frac{n}{q})=n$, a contradiction. Thus $n=p^t$ for some prime $p$ and $|O_j| = p^{t-1} x_j$ with $x_1+x_2=p$ and $x_1\leq x_2$. Since $O_1$ has trivial radical, the group $Y\cap\Z_n$ is trivial. Therefore, $Y\cong Y\Z_n/\Z_n\hookrightarrow \langle \alpha\rangle$. In particular, $|Y|\leq |\langle\alpha\rangle|\leq\varphi(n) < n$. On the other hand $|Y|$ is divisible by $\lcm(|O_1|,|O_2|)=\frac{n}{p}\lcm(x_1,x_2)=\frac{n}{p}x_1x_2$. Hence $\frac{n}{p}x_1x_2 < n\implies x_1x_2 < p \implies x_1 =1, x_2=p-1$. Therefore $(p-1)\frac{n}{p} = \frac{n}{p}x_1x_2 \leq |Y| \leq\varphi(n)=\frac{n}{p}(p-1)\implies |Y|=\varphi(n)=\frac{n}{p}(p-1)\implies Y\cong\langle\alpha\rangle\cong\Z_n^*\implies X\cong\mathsf{Hol}(\Z_n)$. Let $y$ be a generator of $Y$. Since $O_1$ and $O_2$ are the only orbits of $Y=\langle y\rangle$, the permutation $y$ has two cycles in its cyclic decomposition. One of length $|O_1|$ and another one of length $|O_2|$. It follows from $|O_1|=p^{t-1},|O_2|=p^{t-1}(p-1)$ that $y^{p^{t-1}}$ acts trivially on $O_1$. Therefore $y^{p^{t-1}}$ has at least $p^{t-1}$ fixed points. If $p$ is odd, then $y^{p^{t-1}}$ is a non-trivial $p'$-element of $\mathsf{Hol}(\Z_n)$. Therefore it has only one fixed point, and, consequently, $t=1$. Assume now that $p=2$. Then, as it was shown before, $\Z_{2^t}^*$ is cyclic. This is possible only when $t=1$ or $t=2$. The rest follows easily. \end{proof} \noindent {\bf Proof of Theorem~\ref{280320a}.} As in Lemma~\ref{140320b} we assume that $|O_1|\leq|O_2|$. Assume that case (1) of the Lemma occurs. Then $O_1$ is a subgroup of $\F_q^*$ of prime index $\ell$ and $\phi$ permutes transitively $\ell-1$ cosets $O_1\omega,...,O_1\omega^{\ell-1}$. Therefore the induced action of $\phi$ on the cosets: $O_1\omega^i\mapsto O_1\omega^{ip},i\in\Z_\ell^*$ is a full cycle. This implies that ${\rm ord}_\ell(p)=\ell-1$. Together with $p^r\equiv 1({\rm mod}\,\ell)$ this implies $\ell-1\,|\,r$. Thus $O_1$ is the set described in Proposition~\ref{200320b}. Assume now that the second case of the Lemma~\ref{140320b} occurs. Then both $O_1$ and $O_2$ are unions of two cosets of an index $4$ subgroup $C\leq\F_q^*$ and $(C\omega)^\phi = C\omega^3$. The latter equality is equivalent to $C\omega^p = C\omega^3$, and, therefore, implies $p\equiv 3({\rm mod}\,4)$. Together with $p^r\equiv 1({\rm mod}\,4)$ this yields $2\,|\,r$. This yields us the examples described in Proposition~\ref{200320c}.\hfill $\square$ \vspace*{10mm} \centerline{\bf Acknowledgments} \vspace*{5mm} The author is very grateful to I. Ponomarenko who attracted the author's attention to the topic and read carefully the first draft. I am also thankful to A. Vasil'ev for careful reading and to M. Klin for useful discussions.

https://arxiv.org/abs/2010.10227

Reconstructibility of matroid polytopes

We specify what is meant for a polytope to be reconstructible from its graph or dual graph. And we introduce the problem of class reconstructibility, i.e., the face lattice of the polytope can be determined from the (dual) graph within a given class. We provide examples of cubical polytopes that are not reconstructible from their dual graphs. Furthermore, we show that matroid (base) polytopes are not reconstructible from their graphs and not class reconstructible from their dual graphs; our counterexamples include hypersimplices. Additionally, we prove that matroid polytopes are class reconstructible from their graphs, and we present a $O(n^3)$ algorithm that computes the vertices of a matroid polytope from its $n$-vertex graph. Moreover, our proof includes a characterisation of all matroids with isomorphic basis exchange graphs.

\section{Overview} \noindent We begin with a brief overview of how this manuscript is organised. It consists of three further sections. In Section~\ref{sec:intro-rec}, we give an introduction to the classical problem of reconstructing a polytope from its graph (Problem~\ref{prob:rec}), or more general, from its $k$-skeleton. We introduce a variation of the classical reconstruction in Problem~\ref{prob:class-rec}, which we call \emph{class reconstruction problem}. This new problem is inspired by a conjecture of Joswig, \cite[Conjecture 3.1]{Joswig:2000}, stating that a cubical polytope is reconstructible from its dual graph. We examine this conjecture in this section and show that the conjecture should be about class reconstructibility of cubical polytopes (Example~\ref{ex:dualcubical}). In the remaining of the paper, we focus on the reconstructibility of matroid (base) polytopes from their graphs. In Section~\ref{sec:BasisExchangeGraphs}, we exploit Maurer's characterisation \cite{Maurer:1973} of basis exchange graphs, i.e.\ the vertex-edge graphs of matroid polytopes, to obtain Theorem~\ref{thm:recmatroids}, which characterises matroids with isomorphic basis exchange graphs. This part does not use any polytope theory. In Section~\ref{sec:RecMatroidPolytopes}, we present Algorithm~\ref{algo:label}, which, given a graph on $n$ nodes, computes the bases of a matroid or the vertices of its matroid polytope, whenever the graph is the basis exchange graph of an matroid. Furthermore, Algorithm~\ref{algo:label} returns \emph{false} if the graph is not that of a matroid polytope. This algorithm will terminate after at most $O(n^3)$ steps. Using Theorem~\ref{thm:recmatroids}, we derive that matroid polytopes are class reconstructible from their graphs (Theorem~\ref{thm:rec-matroid-polys}) and that they are neither class reconstructible from their dual graphs (Corollary~\ref{cor:rec-dual-matroid-polys}) nor reconstructible from their graphs (Proposition~\ref{prop:counterexample}). \section{The problem of polytope reconstruction}\label{sec:intro-rec} \noindent A {\it $d$-polytope} is a $d$-dimensional (convex) polytope. The $k$-dimensional {\it skeleton} of a polytope~$P$ is the set of all its faces of dimension at most~$k$. The $1$-skeleton of $P$ is its {\it (vertex-edge) graph} $G(P)$. The {\it combinatorial structure} of a polytope is given by its {\it face lattice}, a partial ordering of the faces by inclusion. Two polytopes are \emph{combinatorially equivalent} if there is an inclusion preserving bijection between their face lattices. In this article, we do not distinguish between combinatorially equivalent polytopes. For a polytope $P$ that contains the origin in its relative interior, the \textit{polar polytope} $P^*$ is defined as \[ P^*=\SetOf{y\in \RR^d}{x\cdot y\le 1 \text{ for all $x$ in $P$}}. \] If $P$ does not contain the origin, we translate the polytope so that it does. Note that translating the polytope $P$ changes the geometry of $P^*$, but not the faces structure as long as the origin is in the relative interor of $P$. The face lattice of $P^*$ is the inclusion reversed face lattice of $P$. In particular, the vertices of $P^*$ correspond to the \emph{facets} of $P$, i.e.\ codimension-one faces of $P$. The \emph{dual graph} of a polytope $P$ is the graph of the polar polytope, or equivalently, the graph on the set of facets of $P$ where two facets are adjacent in the dual graph if they share a \emph{ridge}, a codimension-two face. The classical graph reconstruction problem can be stated as follows. \begin{problem}[Reconstruction problem]\label{prob:rec} Given a parameter $d$, and the (dual) graph of a $d$-polytope, determine the face lattice of the polytope. \end{problem} Problem~\ref{prob:rec} requires that there is exactly one $d$-polytope whose (dual) graph is the given graph. Furthermore, the reconstruction problem can plainly be extended to skeletons of a polytope; see Chapter 12 of Gr\"unbaum's book on convex polytopes \cite{Gruenbaum:2003} or Bayer's survey on the topic \cite{Bayer:2018}. It is well known that the graph of a polytope does not in general determine its dimension. For example, the $k$-skeleton of the cyclic $d$-polytope on $n$ vertices \cite[page~11]{Ziegler:1995} agrees with the $k$-skeleton of the $(n-1)$-simplex, for every natural numbers $n$, $d$ and $k$ satisfying $n>d\geq 2k+2 \geq 4$; see \cite[Chapter 12]{Gruenbaum:2003}; in particular, the graph of the cyclic $4$-polytope on $n>4$ vertices is the complete graph on $n$ vertices. Moreover, these examples show that there are polytopes of distinct dimension that share the same $\lfloor\tfrac{d}{2}-1\rfloor$-skeleton as the cyclic $d$-polytope on $n$ vertices. However, the $\lfloor \tfrac{d}{2}\rfloor$-skeleton of a $d$-polytope cannot be shared by a polytope of another dimension; see \cite[Thm.~ 12.2.1]{Gruenbaum:2003}. In this manuscript, we focus on reconstructions from graphs. Graph reconstruction means algorithmically that, given the dimension of a polytope and the graph or dual graph of the polytope, it can be decided whether a given induced subgraph is the graph of a facet Every $d$-polytope is reconstructible from its $(d-2)$-skeleton \cite[Theorem~12.3.1]{Gruenbaum:2003}, and there are combinatorially inequivalent $d$-polytopes with the same $(d-3)$-skeleton, e.g., a bipyramid over a $(d-1)$-simplex, and the pyramid over a bipyramid over a $(d-2)$-simplex. The graph however determines the combinatorial structure if the polytope has some additional properties. Blind and Mani \cite{BlindMani:1987}, and later Kalai \cite{Kalai:1988}, showed that every {\it simple $d$-polytope}, a polytope in which every vertex is incident to precisely $d$ edges, can be reconstructed from its graph. Other recent reconstruction results can be found in \cite{DoolittleEtAl:2017}, \cite{PinedaEtAl:2017}, and \cite[Section~19.5]{GoodmanORourkeToth:2018}. In contrast, there are simple $d$-polytopes that cannot be reconstructed from their dual graphs; equivalently, \emph{simplicial polytopes}, polytopes in which every facet is a simplex, that cannot be reconstructed from their graphs. This follows from the facts that the graphs of neighbourly $d$-polytopes on $n$ vertices are complete graphs and neighbourly polytopes with even dimension are simplicial. A $d$-polytope is called {\it neighbourly} if every $\lfloor d/2 \rfloor$-element subset of its vertices is the vertex set of a face. In particular, there are two combinatorially inequivalent neighbourly 4-polytopes on eight vertices, one of them being a cyclic polytope \cite[Theorem~7.2.4]{Gruenbaum:2003}. Polytopes in which every facet is a cube are called {\it cubical polytopes}. Cubical polytopes are not determined by their graphs and their dimensions. Sanyal and Ziegler \cite[Corollary~ 3.8]{SanZie10} constructed combinatorially inequivalent, cubical $d$-polytopes on $n$ vertices with the same $\lfloor d/2\rfloor$-skeleton as the $n$-cube; these are \textit{neighbourly cubical $d$-polytopes}. Joswig \cite[Conjecture~3.1]{Joswig:2000} conjectured that cubical polytopes can be reconstructed from their dual graphs. Our first contribution is a counterexample (Example~\ref{ex:dualcubical}) to Joswig's conjecture in the strict sense of Problem~\ref{prob:rec}. \begin{figure}[t!] \caption{Schlegel projections of the $4$-polytopes $Q$ and its polar $Q^*$ from Example~\ref{ex:dualcubical}. The polar polytope $Q^*$ has the same dual graph as the $4$-cube. } \label{fig:p-cube-dual-graph} \begin{subfigure}[t]{0.45\textwidth} \centering \subcaption{A Schlegel projection of $Q$.} \input{4polyEx1.tikz} \end{subfigure} \begin{subfigure}[t]{0.45\textwidth} \centering \subcaption{A Schlegel projection of $Q^*$.} \input{4polyEx1dual.tikz} \end{subfigure} \end{figure} \begin{example}\label{ex:dualcubical} Consider the $4$-polytope $Q$ formed by the eight vertices: \begin{align*} (-2, -2, -2, -2), && (-2, -2, -2, 4), && (-2, -2, 4, -2), && (-2, 4, -2, -2),\\ ( 4, 4, 4, -2), && ( 4, -2, -2, -2), && ( 4, -5, -2, 4), && ( 1,-11, -1, 12) \end{align*} The vertices are grouped such that nonadjacent vertices are on top of each other. The polytope $Q$ has the following twelve facets: \begin{align*} 2x_2 +4x_3+ x_4&\geq -14& x_1+2x_2+15x_3 &\geq -36\\ 3x_1-3x_2 -4x_3-4x_4&\geq -8 & 14x_2+9x_4&\geq -46\\ -2x_1+2x_2 -2x_3+ x_4&\geq -10& 3x_1+ x_2 &\geq -8\\ x_1- x_2 -x_3- x_4&\geq -2 & x_1 &\geq -2\\ -x_1-2x_2+ x_3-2x_4 &\geq -4 & x_3 &\geq -2\\ -2x_1-2x_2+2x_3- x_4 &\geq -6 & x_4 &\geq -2\\ \end{align*} It follows that the vertex-edge graph of $Q$ is isomorphic to the vertex-edge graph of a four dimensional cross-polytope, the polar of a $4$-cube. See additionally Figure~\ref{fig:p-cube-dual-graph}. As a consequence, the polar polytope $Q^*$ of $Q$ has a dual graph isomorphic to the dual graph of a $4$-cube. However, note that the polytope $Q^*$ has eight facets, each with six or seven vertices, and so $Q^*$ is noncubical. \end{example} We have just shown the following. \begin{proposition} Cubical polytopes are not reconstructible from their dual graphs and dimensions. \end{proposition} Example~\ref{ex:dualcubical} is our motivation for the following definition. We say that a polytope within a class of polytopes is \emph{class reconstructible} from its graph or dual graph if no other polytope in the class has the same graph or dual graph, respectively. \begin{problem}[Class reconstruction problem]\label{prob:class-rec} Given a class of polytopes, a parameter $d$, and the (dual) graph of a $d$-polytope in the class, determine the face lattice of the polytope. \end{problem} To the best of our knowledge, it is still open whether Joswig's conjecture is true in the sense of Problem~\ref{prob:class-rec}; that is, that cubical $d$-polytopes are class reconstructible from their dual graphs. Their are partial results in that direction. Joswig proved that capped cubical polytopes, also called stacked cubical polytopes, are reconstructible from their dual graphs and dimension within the class of cubical polytopes; see \cite[Theorem 3.7]{Joswig:2000}. For the rest of this paper, we focus on matroids and their polytopes. A \emph{matroid}~$M$ is a nonempty collection $\mathcal B$ of subsets of a finite set $E$, called \emph{bases}, that satisfy the symmetric bases exchange property: \begin{center} For every pair of bases $A,B\in\mathcal B$ and each element $a\in A-B$ there exist an element $b\in B$ such that $A-a+b$ and $B+a-b$ are contained in $\mathcal B$. \end{center} We used $\pm$ to denote inclusion and exclusion of elements from to a set. There are many ways to define matroids see \cite[Chapter~1]{Oxley:2011} and \cite[Appendix by Thomas Brylawski]{White:1986} for further cryptomorphisms. The finite set $E$ is the \emph{ground set} of the matroid and two matroids are isomorphic if there is a bijection between their ground sets that maps bases to bases. From now on, we may assume that the ground set $E$ of a matroid is the set $[n]=\{1,2,\ldots,n\}$ for $n=\size E$ whenever the ground set $E$ is not specified. For every basis $B\in \mathcal B$, the indicator vector $e_{B}\in \mathbb R^{n}$ is defined as $e_{B}=\sum_{i\in B}e_{i}$ where $e_{i}$ is the standard vector with a one in the $i$-coordinate. The {\it matroid (base) polytope} $P(M)$ associated with the matroid $M$ is the convex polytope \[P(M)\coloneqq \conv\SetOf{e_B}{B\in\mathcal B}.\] Matroid polytopes, their regular subdivisions and normal fans play a fundamental role in tropical geometry; see \cite{MaclaganSturmfels:2015} and \cite{Joswig:2020} for further details. As mentioned above, one goal of this paper is to prove that matroid polytopes are class reconstructible from their graphs. In contrast to Problem~\ref{prob:rec} and Problem~\ref{prob:class-rec} where the input graph is understood to be the graph of a polytope, we do not make this requirement. Furthermore, we do not even need the dimension $d$ as an input parameter. Given a graph, Algorithm~\ref{algo:label} in Section~\ref{sec:RecMatroidPolytopes} decides whether or not the graph is the graph of a matroid polytope. If it is, it returns a collection of bases of an matroid that correspond to the vertices of the unique matroid polytope with that graph; if is not, it returns \emph{false}. \section{Basis exchange Graphs of Matroids}\label{sec:BasisExchangeGraphs} \noindent It is convenient to define the graph $G(M)$ of a matroid polytope $P(M)$ without resorting to the realisation of the polytope. The graph $G(M)$ can be defined on the set of bases $\mathcal B$ of a matroid $M$. Two nodes $A,B\in\mathcal B$ are adjacent if and only if $\size(A-B)=1$. We refer to $G(M)$ as the \emph{labelled basis exchange graph}, whenever its nodes are labelled by the matroid bases and as \emph{abstract basis exchange graph} if the labelling is unknown. In this section, we determinate under which conditions it is possible to read off the matroid $M$ from the abstract basis exchange graph, i.e.\ finding a node labelling by basis of $M$. Section~\ref{sec:RecMatroidPolytopes} presents an algorithm that, given the abstract graph, computes these labels and therefore a matroid. Let $M$ be a matroid on $E$ with bases $\mathcal B$. The dual matroid of $M$ is a matroid $M^{*}$ on the same ground set with bases $\mathcal B^{*} =\{E-B:B\in \mathcal B\}$. A {\it loop} of a matroid is an element of the ground set that is contained in no basis and a {\it coloop} is an element of the ground set that is contained in every basis. Clearly, some matroids lead to isomorphic basis exchange graphs. \begin{lemma}\label{lem:samegraph} Two matroids $M$ and $M'$ have isomorphic basis exchange graphs if \begin{enumerate} \item \label{item:iso} $M'$ is isomorphic to $M$, \item \label{item:dual} $M$ is the dual of $M'$, \item \label{item:loops} $M'$ equals $M$ up to loops and coloops. \end{enumerate} \end{lemma} \begin{proof} \eqref{item:iso} If $M'$ and $M$ are isomorphic then there is a bijection $\varphi$ of the ground sets, which maps bases to bases. Obviously this map induces a bijection between the nodes of $G(M')$ and $G(M)$. This is edge preserving, as for any basis $A$ and $B$ of $M'$ we have $\varphi(A)-\varphi(B) = \varphi(A-B)$ is of the same size as $A-B$. \eqref{item:dual} The map $B\mapsto E-B$ on the bases of $M$ is a bijection to the bases of the dual matroid~$M^*$. Moreover, for each pair of bases $A$ and $B$ we have $(E-A) - (E-B) = B - A$ and $\size(B-A) = \size(A-B)$. Thus, this map is a graph isomorphism. \eqref{item:loops} Adding a loop to a matroid does not change the set of bases and hence preserves the basis exchange graph. The isomorphism of the basis exchange graphs for colops follows from \eqref{item:dual}. \end{proof} The next step is to exploit the structure of a basis exchange graph. We benefit from results of Maurer, whose main theorem in \cite{Maurer:1973} is a characterisation of basis exchange graphs. A building block in his theorem are neighbourhood subgraphs. Given a node $B$ in a graph $G$. Its \emph{neighbourhood subgraph} $N(B)$ consists of all nodes that are adjacent to $B$ and all edges of $G$ that connect two of these nodes. Neighbourhood subgraphs in a basis exchange graph $G(M)$ are line graphs. The \emph{line graph} $L(H)$ of a graph $H$ is the graph whose nodes are the edges of $H$ and two nodes share an edge whenever the corresponding edges in $H$ are adjacent. A graph is {\it bipartite} if its vertex set can be partitioned into two sets $X_{1}$ and $X_{2}$ with every edge connecting a vertex in $X_{1}$ to a vertex in $X_{2}$; we call the sets $X_{1}$ and $X_{2}$ {\it partite sets}. \begin{lemma}[{Link condition of basis exchange graphs. \cite[Lemma 1.8]{Maurer:1973}, \cite{ChalopinEtAl:2015}}] \label{lem:linegraph} For any node $B$ in a (labelled) basis exchange graph, the neighbourhood subgraph is the line graph of a (labelled) bipartite graph with partite sets $B$ and $E-B$. \end{lemma} The embedding of the bipartite graph, that we denote by $L^{-1}(B)$, into $N(B)$ is explicitly given by the map from the edges of the bipartite graph to the nodes of $N(B)$ \[ (b,e) \mapsto B-b+e \enspace . \] Clearly two vertices are connected by an edge in the neighbourhood subgraph if and only if the two corresponding edges of the bipartite graph are adjacent. The \emph{rank} of a matroid $M$, denoted $\rank M$, is the size of a basis in $\mathcal B$ and its \emph{corank} is the rank of the dual matroid of $M$. The following examples show that we can not read off the rank and corank from a neighbourhood subgraph of $G(M)$. Moreover, they show that we have to extend the list given in Lemma~\ref{lem:samegraph}. \begin{figure} \caption{Isomorphic basis exchange graphs of isomorphic matroids of different ranks. (a) The basis exchange graph of the matroid $M_1$ of Example~\ref{ex:mat1}, with the bipartite graph that corresponds to the neighbourhood subgraph $N(12)$ highlighted. (b) The basis exchange graph of the matroid $M_2$, with the bipartite graph of the subgraph $N(123)$ highlighted. \label{fig:example}} \input{example.tikz} \end{figure} \begin{example} \label{ex:mat1} Consider the matroid $M_1$ on the ground set $E=\{1,\ldots, 6\}$ with the nine bases $12$, $13$, $14$, $25$, $26$, $35$, $36$, $45$ and $46$; it has rank two and corank four. To improve readability, we omitted parentheses and colons to denote the bases. The neighbourhood subgraph of $12$ of the basis exchange graph contains the four nodes $13$, $14$, $25$, and $26$ and two nonadjacent edges. See Figure~\ref{fig:example} (a). As another matroid $M_2$, we take the rank $3$ matroid on the same ground set with the nine bases $123$, $126$, $136$, $234$, $235$, $246$, $256$, $346$ and $356$. The neighbourhood subgraph of $123$ in the basis exchange graph contains the four nodes $126$, $136$, $234$, and $235$. See Figure~\ref{fig:example} (b). Figure~\ref{fig:example} illustrates that the basis exchange graph of $M_1$ is isomorphic to the basis exchange graph of $M_2$. Moreover, Figure~\ref{fig:example} shows that the two bipartite graphs $L^{-1}(12)$ and $L^{-1}(123)$, which correspond to the isomorphic neighbourhood subgraphs $N(12)$ and $N(123)$ in Figure~\ref{fig:example}~(a) and Figure~\ref{fig:example}~(b), are isomorphic. \end{example} More generally, let $M_1$ be the matroid with bases $\mathcal B_1$ on the ground set $E_1$ and and $M_2$ be the matroid with bases $\mathcal B_2$ on $E_2$, where the ground sets $E_1$ and $E_2$ are disjoint. Let $E\coloneqq E_{1}\cup E_{2}$ and $\mathcal B\coloneqq \SetOf{B_{1}\cup B_{2}}{B_{1}\in \mathcal B_{1},B_{2}\in \mathcal B_{2}}$. The collection $\mathcal B$ is the set of bases of a matroid $M$ on the ground set $E$ called the \emph{direct sum} $M_{1}\oplus M_{2}$ of $M_{1}$ and $M_{2}$. We say that a matroid is \emph{connected} if it is not the direct sum of two matroids. If a matroid $M=M_1\oplus\dots\oplus M_k$ is the direct sum of connected matroids then are the ground sets $E_1,\ldots,E_k$ the \emph{connected components of $M$}. The matroid polytope $P(M_1\oplus M_2)$ is the \emph{product} $P(M_{1})\times P(M_{2})$ of the polytopes $P(M_{1})$ and $P(M_{2})$; that is for $M=M_1\oplus M_2$, \[ P(M)=P(M_{1})\times P(M_{2})\coloneqq\SetOf{(x_{1},x_{2})\in \mathbb R^{\size E_{1}+\size E_{2}}}{ x_{1}\in P(M_1) \text{ and } x_{2}\in P(M_2)}. \] The basis exchange graph $G(M)$ is the \emph{Cartesian product} $G(M_{1})\times G(M_{2})$ of $G(M_{1})$ and $G(M_{2})$; that is, $G(M)$ is the graph on $V(G(M_{1}))\times V(G(M_{2}))$ with edge set $V(G(M_{1}))\times E(G(M_{2}))\cup E(G(M_{1}))\times V(G(M_{2}))$; here $V(\cdot)$ and $E(\cdot)$ denote the node and edge set of a graph, respectively. The graphs $G(M_1)$ and $G(M_2)$ are \textit{factors} of the product $G(M)$. The following extends our list of matroids with isomorphic basis exchange graphs. \begin{lemma}\label{lem:disconnected} Let $M_1$, $M'_1$ and $M_2$, $M'_2$ be two pairs of matroids with isomorphic basis exchange graphs. Then the basis exchange graph of $M_1\oplus M_2$ and $M'_1\oplus M'_2$ are isomorphic as well. \end{lemma} \begin{proof} The basis exchange graph $G(M_1\oplus M_2)$ is the Cartesian product of $G(M_1)$ and $G(M_2)$ and hence this graph is isomorphic to $G(M_1')\times G(M_2')=G(M'_1\oplus M'_2)$. \end{proof} In the rest of this section we will prove that Lemma~\ref{lem:samegraph} and Lemma~\ref{lem:disconnected} cover all cases where matroids have isomorphic basis exchange graphs. Note that all those properties are reflected in every bipartite graph $L^{-1}(B)$ of an node $B$. More precisely, loops and coloops correspond to isolated nodes in the bipartite graph, duality to a switch of the two node sets, matroid isomorphisms to graph isomorphisms and direct sums of matroids to the disjoint union of bipartite graphs. The next statement summarises the uniqueness of the rank and corank for connected matroids. \begin{lemma}\label{lem:connected} Let $M$ be a matroid and let $B$ be a basis of $M$. The following statements are equivalent: \begin{enumerate} \item \label{it:Mcon} $M$ is connected, \item \label{it:Ncon} the neighbourhood subgraph $N(B)$ is connected, \item \label{it:Bcon} the bipartite graph $L^{-1}(B)$ is connected. \end{enumerate} Moreover, the (abstract) graph $L^{-1}(B)$ determines the rank and corank of $M$ uniquely up to duality of connected components. \end{lemma} \begin{proof} Clearly, an edge path in the bipartite graph of $N(B)$ is mapped to a node path of $N(B)$. This already shows the equivalence of \eqref{it:Ncon} and \eqref{it:Bcon}. The matroid $M$ is disconnected if and only if $M = M_1\oplus M_2$. Let $E_1$ and $E_2$ be the ground sets of those matroids. For any $e\in E_1\cap B$ is $B-e+f$ a basis only if $f\in E_1$, hence there is no edge from a node in $E_1$ to one in $E_2$ in the bipartite graph. This proves that \eqref{it:Bcon} implies \eqref{it:Mcon}. On the other hand, suppose the bipartite graph $L^{-1}(B)$ is disconnected and $S\subset E$ is the set of nodes of a connected component. Then $S$ is a proper subset of $E$ and $\rank(S) = \rank(S\cap B)$, otherwise there has to be an edge between a node in $S-B$ and a node in $B-S$. The same argument holds for the complement $E-S$. Thus $S$ is a separator, i.e.\ $\rank(S) + \rank(E-S) \geq \rank(E)$, and the matroid $M$ is disconnected. These arguments show that from \eqref{it:Mcon} follows \eqref{it:Bcon}. Any connected bipartite graph $L^{-1}(B)$ determines the number of nodes in the two colour classes uniquely up to there interchange. This sizes are exactly the rank and corank of the matroid $M$. \end{proof} Suppose we have an abstract basis exchange graph with a marked node and its neighbourhood subgraph. Any labelling and ordering of the two colour classes of the corresponding abstract bipartite graph leads to a unique labelling of the neighbourhood subgraph and the marked node. Now, we describe how we can extend this labelling to a proper labelling of all nodes in the basis exchange graph. The key here is again the characterisation of Maurer. A \emph{common neighbour subgraph} is a subgraph of $G$ that contains two nodes of $G$ of distance two and all there common neighbours as additional nodes together with their connecting edges. The following is Maurer's characterisation. \begin{theorem}[{\cite[Theorem 2.1]{Maurer:1973}}]\label{thm:Maurer} A connected graph $G$ is a basis exchange graph of a matroid if and only if \begin{enumerate} \item\label{it:Maurer1} it meets the interval condition, i.e.\ each common neighbour subgraph is a square, a pyramid, or a octahedron, \item\label{it:Maurer2} it meets the positioning condition, i.e.\ for every node $B$ and every induced subgraph which is a square with nodes $A$, $C$, $D$, $E$ in cyclic order holds \[\dist(B,A)+\dist(B,D)=\dist(B,C)+\dist(B,E)\enspace,\] \item\label{it:Maurer3} the neighbourhood subgraph $N(B)$ is the line graph of a bipartite graph for some $B$ \end{enumerate} \end{theorem} The label of a single nonlabelled node in each of the three graphs of (\ref{it:Maurer1}) is given by all appearing elements in the labels of its neighbours that do not occur in the label of the unique node of distance two together with the intersection of all these labels. We label our graph along the distance levels to our fixed node and its labelled neighbourhood. We conclude the following. \begin{corollary}\label{cor:extlabelling} Any valid labelling of a node and its neighbourhood subgraph in a basis exchange graph has a unique extension to the entire graph. \end{corollary} We sum up our results in the following theorem. \begin{theorem}\label{thm:recmatroids} An abstract basis exchange graph determines its matroid $M$ up to \begin{enumerate} \item the isomorphic type of $M$, \item duality of the connected components of $M$, \item and possible loops and coloops. \end{enumerate} \end{theorem} In the next section we present Algorithm~\ref{algo:label} which constructs of a valid labelling from a basis exchange graph. \section{Reconstruction of matroid polytopes}\label{sec:RecMatroidPolytopes} \noindent In this section, we prove that matroid polytopes are class reconstructible from their graphs. Furthermore, we show that they are neither class reconstructible from their dual graphs nor reconstructible from their graphs. We begin by presenting Algorithm~\ref{algo:label}, which computes a (valid) labelling from a basis exchange graph if it exists. \begin{algorithm}[htb] \dontprintsemicolon \Input{A connected graph $G=(V,E)$ with at least two nodes} \Output{A valid labelling $\ell$ or \texttt{False}} \smallskip Pick $v\in V$\; $H$ $\leftarrow \SetOf{w\in V}{(v,w)\in E}$\; \If{ $(H,E\cap (H\times H))$ is not a line graph of a bipartite graph}{ \Return \texttt{False}\; } $(B_1 \cup B_2, E_B)$ $\leftarrow$ The ``original'' graph of the line graph $(H,E\cap (H\times H))$\; $\varphi:H\to E_B$ $\leftarrow$ The bijection of the above\; Set $\ell(v) = B_1$\; \For{$w\in H$}{ $(e,f)$ $\leftarrow$ $\varphi(w)$\; Set $\ell(w) = B_1-e+f$ \; } \For{$j=2,\ldots,\size{B_1}$}{ $L$ $\leftarrow$ $\SetOf{u\in V}{\dist(u,v)=j}$\; \While{ $L$ is not empty}{ Pick $u\in L$ and remove $u$ from $L$\; Pick $\hat v\in\SetOf{x\in V}{\dist(x,v)=j-2 \text{ and } \dist(x,u)=2}$\; $T$ $\leftarrow$ $\SetOf{w\in V}{\dist(w,u)=\dist(w,\hat v)=1}$\; $W$ $\leftarrow$ $\SetOf{(x_1,x_2)\in T\times T}{x_1\neq x_2 \text{ and } (x_1,x_2)\not\in E}$\; \If{$W$ is empty}{\Return \texttt{False}\;} Pick $(w_1,w_2)\in W$\; Set $\ell(u)=(\ell(w_1)\cap\ell(w_2)) \cup (\ell(w_1)+\ell(w_2)-\ell(\hat v))$\; } } \For{$u,w\in V$}{ \If{$\size (\ell(u)-\ell(v)) \neq \dist(u,v)$ \Or $\dist(u,w)=2$ and the common neighbour subgraph is neither a square, pyramid or octahedron}{\Return \texttt{False}\;} } \Return $\ell$\; \caption{Label the basis exchange graph} \label{algo:label} \end{algorithm} \begin{proposition} Algorithm~\ref{algo:label} decides whether a graph with $n$ nodes is a basis exchange graph of a matroid and returns a valid labelling of a graph if it exits in $O(n^3)$ steps. \end{proposition} \begin{proof} The graph with exactly one node and no edges is a basis exchange graph of a matroid. To avoid the situation that there are no edges we restrict our self's to the situation where $G$ is a connected graph with $n>1$ nodes. Let us first prove that Algorithm~\ref{algo:label} stops with a correct result. Our analysis relies once more on Mauerer's characterisation; see Theorem~\ref{thm:Maurer}. The graph $G$ violates condition (\ref{it:Maurer3}) if the procedure terminates in line 4. The interval condition (\ref{it:Maurer1}) guarantees that there are two nodes $x_1$, $x_2$ in the common neighbour subgraph of $u$ and $\hat v$ that form a square subgraph whenever $G$ is a basis exchange graph. Hence, $(x_1,x_2)$ is a nonedge and $W$ is nonempty in that case. We conclude that the interval condition is violated if the procedure terminates in line 19. Clearly, the algorithm stops at line 24 when either a common neighbour subgraph violates (\ref{it:Maurer1}) or the labelling is incorrect. If the graph $G$ is a basis exchange graph, then the node $v$ and its neighbourhood have a valid labelling, and by Corollary~\ref{cor:extlabelling} the labelling has a unique extension that is given by the exchange property, which is reflected in line 21. Thus if the node labelling of $G$ is incorrect then $G$ is not a basis exchange graph of a matroid. Our next goal is to show that $G$ is the basis exchange graph of a matroid with bases $\smallSetOf{\ell(v)}{v\in V}$ whenever the algorithm terminates in line 25. The algorithm stops in line 24 or earlier whenever $G$ violets (\ref{it:Maurer1}), and in line 4 if (\ref{it:Maurer3}) is violated. Note that a labelling is valid whenever $\ell(v)-\ell(w)$ is of size one is equivalent to $(v,w)$ is an edge of $G$. in that case the labelling reflects distances via \[ \dist(u,v) = \size(\ell(v)-\ell(u)) \enspace. \] It is easy to see that valid labels of nodes in a square are of the form \[ I+a+b,\, I+a+c,\, I+b+d,\, I+c+d \] for some set $I$ and elements $a,b,c,d$. With the above formula for the distances we get that the algorithms stops at line 24 if (\ref{it:Maurer2}) is violated. We conclude that the output is correct. Our next goal is to complete the proof of correctness by showing that the algorithm terminates. Moreover, we show that the algorithm takes $O(n^3)$ steps in the worst case. For this analysis. We may assume that $G$ is given as an adjacency matrix. This matrix is of size $n^2$ whenever $G$ has $n$ nodes, which is also an upper bound for the number of edges. Thus the second line takes less than $n^2$ steps. In the third line, the algorithm verifies whether the graph $(H,E\cap(H\times H))$ is a line graph of a graph $(B,E_B)$. This graph has $\size H < n$ nodes and $m < \size E$ edges. Already in 1973, Roussopoulos \cite{Roussopoulos:1973} produced an $O(\max\{\size H, m\})$ algorithm that recognises a given line graph and reconstructs the original graph $(B,E_B)$. Hence line 3--5 take $O(n^2)$ time. Here we also need to verify whether $(B,E_B)$ is a bipartite graph; this can be done by traversing the graph, e.g. via depth-first search. One either finds a two-colouring of its nodes, in the case that the graph is bipartite, or a cycle of odd length, if the graph is not bipartite. The graph traversal takes at most $O(n^2)$ steps. Moreover, the two colour classes $B_1$, $B_2$ form a partition of the nodes $B$ of the original graph. We may assume that no node of $(B,E_B)$ is isolated and thus $\size B_1 \leq \size E_B = \size H < n$. Assigning labels to $v$ and all its neighbours can be done in $O(n^2)$ time. It is efficient to compute all pairs of nodes of distance two in advance, before using this piece of information in lines 15 and 23. There are many ways to do so, e.g., by applying Floyd and Warshall's algorithm, which runs in $O(n^3)$. Moreover, we store additionally the the distances from $v$ to any other node. The two nested loops in line 11 and 13 go around $\size B_1-1<n $ and $\size L \leq n$ times, respectively. That is, the inner loop, line 14--21, iterates at most $n^2$ times. Initialising the list $L$ in line 12 takes linear time, while taking $u$ and removing it is constant. Finding the node $\hat v$ in line 15 can be done in $O(n)$ by using the distance information we computed in advance. The set $T$, which is the intersection of two sets, can be computed in $O(n)$, e.g., by iterating once through the nodes and picking those which are adjacent to $u$ and $\hat v$. The graph $G$ is not a graph of a matroid if $T$ has more than four elements. Thus we may assume that the instructions in line 17 to 20 take constant time. Finally, computing the label $\ell(u)$ takes a linear amount of time. Thus the inner loop is linear in $n$, and so lines 11--21 run in $O(n^3)$. There are fewer than $n^2$ pairs of nodes of distance two. Hence this is an upper bound for the number of iterations of the loop in line 22. Computing the common neighbour subgraph is essentially the same as computing $T$ in line 16. This graph has at most six nodes, thus checking if the graph is a square, a pyramid, or an octahedron is constant in time. All this takes $O(n^3)$ time. Independent of that loop, one might iterate trough all pairs of nodes to verify that the labelling is correct, which can also be done in $O(n^3)$ time. We conclude that the algorithm stops after at most $O(n^3)$ steps, with the correct result. \end{proof} \begin{remark} Note that a (valid) labelling of a basis exchange graph gives a matroid; more precisely, the collection of all bases. In Section~\ref{sec:BasisExchangeGraphs}, we saw that this matroid is not unique. The matroids with the same graph are well-characterised by Theorem~\ref{thm:recmatroids}. Moreover, the indicator vectors of the labels or bases form the vertices of the matroid polytope. Thus we may say that Algorithm~\ref{algo:label} computes a polytope. It is remarkable that the dimension of the polytope is not an input parameter of the algorithm. \end{remark} \begin{theorem}\label{thm:rec-matroid-polys} The (abstract) vertex-edge graph of a matroid polytope determinates the dimension and combinatorial type of the polytope uniquely in the class of matroid polytopes. \end{theorem} \begin{proof} Given the abstract vertex-edge graph of a matroid polytope or the basis exchange graph of a matroid we are able to reconstruct the matroid up to the three points given in Theorem~\ref{thm:recmatroids}. Now we will show that isomorphic matroids, matroids with dual connected components and matroids with loops and coloops lead to matroid polytopes with the same combinatorial type. Clearly, an isomorphism of matroids translates into a coordinate permutation of the space where the matroid polytope lies. The matroid polytope of a connected matroid is mapped to the matroid polytope of the dual matroid via the coordinate transformation $x_i\mapsto 1-x_i$ and hence they have the same combinatorial type. The matroid polytope of a direct sum is the product of the matroid polytopes of all the summands. This implies that matroids with dual connected components have combinatorial isomorphic matroid polytopes. A loop in a matroid fixes a coordinate to be $0$, and a coloop to be $1$. In other words, in both cases the matroid polytope is embedded in a hyperplane. We conclude that the vertex-edge graph determines the combinatorial type of its matroid-polytope uniquely. \end{proof} Theorem~\ref{thm:rec-matroid-polys} proves that matroid polytopes are class reconstructible from their graphs, in the sense of Problem~\ref{prob:class-rec}, via Algorithm~\ref{algo:label}. However, matroid polytopes are neither reconstructible nor class reconstructible from their dual graphs. This is illustrated best by the matroid polytopes of uniform matroids. A \emph{uniform matroid} $U_{r,n}$ of rank $r$ on $n$ elements has $\tbinom{n}{r}$ bases. That is every $r$-set forms a basis. Its matroid polytope is the \emph{hypersimplex} \[ \Delta(r,n) \coloneqq P(U_{r,n}) = \SetOf{x\in[0,1]^n}{\sum_{i=1}^n x_i = r} \] If $M$ is another matroid of rank $r$ on $n$ elements, then its polytope $P(M)$ a subpolytope of $\Delta(r,n)$ and the graph $G(M)$ is a subgraph of the graph $G(U_{r,n})$; see \cite{GelfandEtAl:1987}. Now we are ready to provide some counterexamples. \begin{proposition} Let $k,r,n$ be integers such that $0 \leq k < r\leq n$ and $k < n-r$. The $k$-skeleton of the polar polytope of the hypersimplex $\Delta(r,n)$ is isomorphic to the $k$-skeleton of the $n$-dimensional cross-polytope, which is the polar of the $n$-cube. \end{proposition} \begin{proof} The hypersimplex $\Delta(r,n)$ is the intersection of the hyperplane $\sum x_i = r$ with the $n$-dimensional $0/1$-cube, and all its facets are induced by those of the cube. Now, let $F$ be a $m$-dimensional face of that cube. Without loss of generality, we may assume that \[F = F_1\cap \ldots \cap F_\ell \cap F_{\ell+1} \cap \ldots \cap F_m,\] where $F_i = \SetOf{x\in[0,1]^n}{x_i=1}$ for $i\leq\ell$ and $F_i = \SetOf{x\in[0,1]^n}{x_i=0}$ for $\ell < i\leq m$. The face $F$ is a $n-m$ dimensional $0/1$-cube, and $F$ intersected with $\sum x_i = r$ is isomorphic to the hypersimplex $\Delta(r-\ell,n-m)$. This is a $(n-m-1)$ dimensional polytope unless $r=\ell$ or $m+r=n+\ell$. In these two cases, the hypersimplex is a single point. Therefore, the intersection of the facets $F_i\cap \smallSetOf{x\in\RR^n}{\sum x_i=r}$ of the hypersimplex form the $(n-m-1)$-face $F\cap \smallSetOf{x\in\RR^n}{\sum x_i=r}$ whenever $\ell<r$ and $r-\ell<n-m$. Hence, for every $m<\max\{r,n-r\}$ is the intersection of $m$ facets of $\Delta(r,n)$ a $(n-m-1)$-face. The claim follows from the observation that every face of the hypersimplex is of the form above. \end{proof} \begin{corollary}\label{cor:rec-dual-matroid-polys} The dual graphs of the $(n-1)$-dimensional polytopes $\Delta(r,n)$ and $\Delta(r+1,n)$ are isomorphic whenever $ n-2 > r\geq 2$. \end{corollary} We proceed with a discussion on the reconstructibility of matroid polytopes as in Problem~\ref{prob:rec}. We begin this discussion with the following classical example that we mentioned in Section~\ref{sec:intro-rec}. \begin{example} Consider the cyclic polytope of dimension $d\geq 4$ and $n>d$ vertices. Its vertex-edge graph is the complete graph with $n$ nodes; see \cite[Chapter 12]{Gruenbaum:2003}. This graph is also the vertex-edge graph of an $(n-1)$-dimensional simplex, which is the hypersimplex $\Delta(1,n)$, and thus the matroid polytope of an uniform matroid on $n$ elements of rank $1$. \end{example} \begin{figure}[t!] \caption{Schlegel projections of the two $4$-polytopes $P(M)$ and $Q$ of Example~\ref{ex:counterexample}. On the right, the nonmatroidal facet of $Q$ is exposed. } \label{fig:counterexample} \begin{subfigure}[t]{0.45\textwidth} \centering \subcaption{A Schlegel projection of $P(M)$.} \input{matroidpoly.tikz} \end{subfigure} \begin{subfigure}[t]{0.45\textwidth} \centering \subcaption{A Schlegel projection of $Q$.} \input{counterexample.tikz} \end{subfigure} \end{figure} However, the next example illustrates that matroid polytopes are not reconstructible from their graphs even if if the dimension is given, i.e.\ they are not reconstructible in the sense of Problem~\ref{prob:rec}. \begin{example}\label{ex:counterexample} Consider the matroid polytope $P(M)$ of the rank $2$ matroid $M$ on $5$ elements with the nine bases \[ 12,\,13,\,14,\,15,\,23,\,24,\,25,\,34,\,35 \] That is the elements $4$ and $5$ are parallel. Let us denote by $v_{ij} = e_i+e_j$ the vertices of $P(M)$. The polytope $P(M)$ has nine facets, three of which are tetrahedra, three of which are Egyptian pyramids, two of which are octahedra, and a prism over an triangle. Then the later is the convex hull of the six points $v_{14}$, $v_{15}$, $v_{24}$, $v_{25}$, $v_{34}$ and $v_{35}$. A Schlegel projection of the polytope is shown in Figure~\ref{fig:counterexample}. Now consider the $4$-polytope $Q$ which is derived from $P(M)$ by replacing the vertex $v_{23}= (0,1,1,0,0)$ by the point $w = (-\tfrac{1}{2},\tfrac{1}{2},1,\tfrac{1}{2},\tfrac{1}{2})$, i.e.\ $Q$ is the convex hull of the points $v_{12}$, $v_{13}$, $v_{14}$, $v_{15}$, $v_{24}$, $v_{25}$, $v_{34}$, $v_{35}$ and $w$. The point $w$ is a vertex of $Q$ and lies in the affine hull of the six points that form the prism over an triangle of $P(M)$. These seven points form a nonmatroidal facet of $Q$. All together, the polytope $Q$ has nine vertices, and ten facets, three of which are tetrahedra, six of which are Egyptian pyramids, and one which in not a matroid polytope. A Schlegel projection is shown in Figure~\ref{fig:counterexample}. It is straight forward to check that the two polytopes $P(M)$ and $Q$ have isomorphic graphs. \end{example} The above example is minimal in the sense that matroid polytopes with eight or fewer vertices are reconstructible from their graphs and dimension. We end this article with a last example that shows that hypersimplices are not reconstructible from their graphs and with related open questions. \begin{example}\label{ex:counterexamplehypersimplex} Consider the hypersimplex $\Delta(2,5)$, which is a 4-dimensional polytope with ten vertices and ten facets. Five of those facets are tetrahedra and the other five are octahedra. Let $Q$ be be the polytope that is derived from $\Delta(2,5)$ by pushing the vertex $(1,1,0,0,0)$ in direction of the center, say to the position $w=(\tfrac{5}{8},\tfrac{5}{8},\tfrac{1}{4},\tfrac{1}{4},\tfrac{1}{4})$. Then three of the octahedral facets of $\Delta(2,5)$ break into two egyptian pyramids each. Thus the polytope $Q$ is the convex hull of ten points with $13$ facets, five tetrahedra, six pyramids and two octahedra. However, the graphs of $\Delta(2,5)$ and $Q$ are isomorphic. \end{example} From Example~\ref{ex:counterexamplehypersimplex}, we derive the following statement. \begin{proposition}\label{prop:counterexample} Hypersimplices, and hence matroid polytopes, are not reconstructible from their graphs and dimension. \end{proposition} We saw that matroid polytopes and hypersimplices are not reconstructible from their graphs, i.e.\ their $1$-skeleta. It would be interesting to find the answers to the following questions. \begin{question} What is the smallest $k$ such that the hypersimplex $\Delta(r,n)$ of dimension $n-1$ is reconstructible from its $k$-skeleton and dimension? \end{question} \begin{question} What is the smallest $k$ such that every $n$-dimensional matroid polytope (of rank $r$) is reconstructible from its $k$-skeleton and dimension? \end{question} \section*{Acknowledgements} \noindent The authors are grateful for Michael Joswig comments about the reconstruction of cubical polytopes. The second author also thanks Federation University Australia for the hospitality during a two week stay. \bibliographystyle{alpha}

https://arxiv.org/abs/1208.0975

Covering Numbers in Linear Algebra

We compute the minimal cardinality of a covering (resp. an irredundant covering) of a vector space over an arbitrary field by proper linear subspaces. Analogues for affine linear subspaces are also given.

\section{Linear coverings} \noindent Let $V$ be a vector space over a field $K$. A \textbf{linear covering} of $V$ is a collection $\{W_i\}_{i \in I}$ of proper $K$-subspaces such that $V = \bigcup_{i \in I} W_i$. A linear covering is \textbf{irredundant} if for all $J \subsetneq I$, $\bigcup_{i \in J} W_i \neq V$. Linear coverings exist if and only if $\dim V \geq 2$. \\ \\ The \textbf{linear covering number} $\LC(V)$ of a vector space $V$ of dimension at least $2$ is the least cardinality $\# I$ of a linear covering $\{W_i\}_{i \in I}$ of $V$. The \textbf{irredundant linear covering number} $\ILC(V)$ is the least cardinality of an irredundant linear covering of $V$. Thus $\LC(V) \leq \ILC(V)$. \\ \\ The main result of this note is a computation of $\LC(V)$ and $\ILC(V)$. \begin{mainthm} Let $V$ be a vector space over a field $K$, with $\dim V \geq 2$. \\ a) If $\dim V$ and $\#K$ are not both infinite, then $\LC(V) = \# K +1$. \\ b) If $\dim V$ and $\# K$ are both infinite, then $\LC(V) = \aleph_0$. \\ c) In all cases we have $\ILC(V) = \# K + 1$. \end{mainthm} \noindent Here is a counterintuitive consequence: the vector space $\R[t]$ of polynomials has a countably infinite linear covering -- indeed, for each $n \in \Z^+$, let $W_n$ be the subspace of polynomials of degree at most $n$. However any irredundant linear covering of $\R[t]$ has cardinality $\# \R + 1 = 2^{\aleph_0}$. Redundant coverings can be much more efficient! \\ \\ The fact that a finite-dimensional vector space over an infinite field cannot be a finite union of proper linear subspaces is part of the mathematical folkore: the problem and its solution appear many times in the literature. For instance problem 10707 in this \textsc{Monthly} is intermediate between this fact and our main result. The editorial comments given on page 951 of the December 2000 issue of the \textsc{Monthly} give references to variants of this fact dating back to $1959$. Like many pieces of folklore, there seems to a be mild stigma against putting it in standard texts; an exception is \cite[Thm. 1.2]{Roman}. \\ \indent There are two essentially different arguments that establish this fact. Upon examination, each of these yields a stronger result, recorded as Theorem \ref{THMA} and Theorem \ref{THMB} below. From these two results the Main Theorem follows easily. \\ \indent The first two parts of the Main Theorem have appeared in the literature before (but only very recently!): they were shown by A. Khare \cite{Khare1}. I found these results independently in the summer of 2008. The computation of the irredundant linear covering number appears to be new. The proof of the Main Theorem is given in $\S 2$. \\ \indent There is an analogous result for coverings of a vector space by affine linear subspaces. This is stated in $\S 3$; we then briefly discuss what modifications must be made in the proof of the Main Theorem to obtain this affine analogue. \section{Proof of the Main Theorem} \noindent First we prove three lemmas, all consequences or special cases of the Main Theorem. \begin{lemma}(Quotient Principle) \label{QUOTIENTPRINCIPLE} \label{LEMMA0} Let $V$ and $W$ be vector spaces over a field $K$ with $\dim V \geq \dim W \geq 2$. Then $\LC(V) \leq \LC(W)$ and $\ILC(V) \leq \ILC(W)$. \end{lemma} \begin{proof} By standard linear algebra, the hypothesis implies that there is a surjective linear map $q: V \rightarrow W$. If $\{W_i\}_{i \in I}$ is a linear covering of $W$, then the complete preimages $\{q^{-1}(W_i)\}_{i \in I}$ give a linear covering of $V$. The preimage of an irredundant covering is easily seen to be irredundant. \end{proof} \begin{lemma} \label{LEMMA1} For any field $K$, the unique linear covering of $K^2$ is the set of all lines through the origin, of cardinality $\# K + 1$. It is an irredundant covering. \end{lemma} \begin{proof} The set of lines through the origin is a linear covering of $K^2$. Moreover, any nonzero $v \in K^2$ lies on a unique line, so all lines are needed. The lines through the origin are $\{y = \alpha x \ | \ \alpha \in K\}$ and $x = 0$, so there are $\# K + 1$ of them. \end{proof} \begin{lemma} \label{LEMMA1.5} Let $V$ be a vector space over a field $K$, with $\dim V \geq 2$. Then there are at least $\# K + 1$ hyperplanes -- i.e., codimension-one linear subspaces -- in $V$. \end{lemma} \begin{proof} When $\dim V = 2$ there are exactly $\# K + 1$ hyperplanes $\{L_i\}$. In general, take a surjective linear map $q: V \rightarrow K^2$; then $\{q^{-1}(L_i)\}$ is a family of distinct hyperplanes in $V$. \end{proof} \noindent The exact number of hyperplanes in $V$ is of course known, but not needed here. \begin{thm} \label{THMA} Let $V$ be a finite-dimensional vector space over a field $K$, and let $\{W_i\}_{i \in I}$ be a linear covering of $V$. Then $\# I \geq \# K + 1$. \end{thm} \begin{proof} Since every proper subspace is contained in a hyperplane, it suffices to consider hyperplane coverings. We go by induction on $d$, the case $d = 2$ being Lemma \ref{LEMMA1}. Assume the result for $(d-1)$-dimensional spaces and, seeking a contradiction, that we have a linear covering $\{W_i\}_{i \in I}$ of $K^d$ with $\# I < \# K + 1$. By Lemma \ref{LEMMA1.5}, there is a hyperplane $W$ such that $W \neq W_i$ for any $i \in I$. Then $\{W_i \cap W\}_{i \in I}$ is a covering of $W \cong K^{d-1}$ by at most $\# I < \# K +1$ hyperplanes, giving a contradiction. \end{proof} \begin{thm} \label{THMB} Let $V$ be a vector space over a field $K$, and let $\{W_i\}_{i \in I}$ be an irredundant linear covering of $V$. Then $\# I \geq \# K + 1$. \end{thm} \begin{proof} Let $\{W_i\}_{i \in I}$ be an irredundant linear covering of $V$. Choose one of the subspaces in the covering, say $W_{\bullet}$. By irredundancy, there exists $u \in W_{\bullet} \setminus \bigcup_{i \neq \bullet} W_i$; certainly there exists $v \in V \setminus W_{\bullet}$. Consider the affine line $\ell = \{tu + v \ | \ t \ \in K\}$; evidently $\# \ell = \# K$. If $w = tu + v \in \ell \cap W_{\bullet}$, then $v = w - tu \in W_{\bullet}$, giving a contradiction. Further, if for any $i \neq \bullet$ we had $\# (\ell \cap W_i) \geq 2$, then we would have $\ell \subset W_i$ and thus also the $K$-span of $\ell$ is contained in $W_i$, so $u = (u+v) - v) \in W_i$, again giving a contradiction. It follows that $\# \ell = \# K \leq \# (I \setminus \{\bullet \})$. \end{proof} \noindent \emph{Proof of the Main Theorem}: \\ \\ Let $V$ be a $K$-vector space of dimension at least $2$. By Theorem \ref{THMB}, $\ILC(V) \geq \#K + 1$, and by Lemmas \ref{LEMMA0} and \ref{LEMMA1}, $\ILC(V) \leq \# K + 1$. So $\ILC(V) = \#K +1$, proving part c) of the Main Theorem. It remains to compute $\LC(V)$. \\ Case 1: Suppose $2 \leq \dim V < \aleph_0$. By Theorem \ref{THMA} we have $\LC(V) \geq \# K + 1$, whereas by Lemma 1 and Lemma 2 we have $\LC(V) \leq \LC(K^2) = \# K + 1$. \\ Case 2: Suppose $\dim V \geq \aleph_0$ and $K$ is finite. Then $\LC(V) \leq \LC(K^2) = \# K + 1 < \aleph_0$. Suppose that $V$ had a linear covering $\{W_i\}_{i=1}^n$ with $n < \# K + 1$. Then, since $n$ is finite, we may obtain an irredundant subcovering simply by removing redundant subspaces one at a time, until we get an irredundant covering by $m$ subspaces, with $m \leq n < \#K + 1$ subspaces, contradicting Theorem \ref{THMB}. \\ Case 3: Suppose $\dim V$ and $\# K$ are both infinite. Consider $W = \bigoplus_{i=1}^{\infty} K$, a vector space of dimension $\aleph_0$. For $n \in \Z^+$, put $W_n := \bigoplus_{i=1}^n K$. Then $\{W_n\}_{n=1}^{\infty}$ gives a covering of $W$ of cardinality $\aleph_0$. Since $\dim V \geq \dim W$, by Lemma 4 we have $\LC(V) \leq \aleph_0$. Thus it remains to show that $V$ does not admit a finite linear covering. But once again, if $V$ admitted a finite linear covering it would admit a finite irredundant linear covering, contradicting Theorem \ref{THMB}. \section{Affine Covering Numbers} \noindent An \textbf{affine covering} $\{A_i\}_{i \in I}$ of a vector space $V$ is a covering by translates of proper linear subspaces. An affine covering is irredundant if no proper subset gives a covering. Irredundant affine coverings exist if and only if $\dim V \geq 1$. The \textbf{affine covering number} $\AC(V)$ is the least cardinality of an affine covering, and similarly the \textbf{irredundant affine covering number} $\IAC(V)$ is the least cardinality of an irredundant affine covering. \begin{thm} Let $V$ be a vector space over a field $K$, with $\dim V \geq 1$. \\ a) If $\min(\dim V, \# K)$ is finite, then $\AC(V) = \# K$. \\ b) If $\dim V$ and $\# K$ are both infinite, then $\AC(V) = \aleph_0$. \\ c) We have $\IAC(V) = \# K$. \end{thm} \noindent The proof of the Main Theorem goes through with minor modifications. Lemma \ref{QUOTIENTPRINCIPLE} holds verbatim. The following self-evident result is the analogue of Lemma \ref{LEMMA1}. \begin{lemma} \label{LEMMA1AFFINE} For any field $K$, the unique affine covering of $K^1$ is the set of all points of $K$, of cardinality $\# K$. It is an irredundant covering. \end{lemma} \noindent Combining these two results we get the analogue of Lemma \ref{LEMMA1.5}, in which $\# K + 1$ is replaced by $\# K$. To prove the analogue of Theorem \ref{THMA}, note that for two codimension-one affine subspaces $W_1$, $W_2$ of a vector space $V$, $W_1 \cap W_2$ is either empty or is a codimension-one affine subspace in each $W_i$. In the proof of the analogue of Theorem \ref{THMB} we use the line $\ell = \{(1-t) u + tv \ | \ t \in K\}$. \\ \\ \textbf{Acknowledgments.} This work was partially supported by National Science Foundation grant DMS-0701771.

https://arxiv.org/abs/1301.1270

s-Overlap Cycles for Permutations

The goal of this paper is to solve Problem 481 from the list of research problems in the special issue of Discrete Mathematics dedicated to the Banff International Research Station workshop on "Generalizations of de Bruijn Cycles and Gray Codes" in 2004. Overlap cycles are generalizations of de Bruijn cycles and Gray codes that were introduced originally in 2010 by Godbole et al. In this paper we prove that s-overlap cycles for k-permutations of [n] exist for all k<n.

\section{} \begin{abstract} The goal of this paper is to solve Problem 481 from the list of research problems in the special issue of Discrete Mathematics dedicated to the Banff International Research Station workshop on ``Generalizations of de Bruijn Cycles and Gray Codes" in 2004. Overlap cycles are generalizations of de Bruijn cycles and Gray codes that were introduced originally in 2010 by Godbole et al. In this paper we prove that $s$-overlap cycles for $k$-permutations of $[n]$ exist for all $k<n$. \end{abstract} \section{Introduction} Many applications require a set of combinatorial objects to be ordered in a specific manner. One such ordering is an \textbf{$s$-overlap cycle}, or \textbf{$s$-ocycle}. An $s$-ocycle is an ordering of a set of objects $\mathcal{C}$, each of which has size $n$ and is represented as a string. The ordering requires that object $b = b_1b_2 \ldots b_n$ follow object $a = a_1a_2 \ldots a_n$ only if $a_{n-s+1}a_{n-s+2} \ldots a_n = b_1b_2 \ldots b_s$. Ocycles were introduced by Godbole, Knisley, and Norwood in 2010 \cite{Godbole}. For the reader familiar with universal cycles, we note that an $(n-1)$-ocycle is a universal cycle. Universal cycles, or ucycles, were introduced in 1992 by Chung, Diaconis, and Graham \cite{UC}, and have been studied extensively ever since, for example over block designs \cite{Dewar}. Universal cycles for permutations have long been an interesting research problem. While ucycles using the standard permutation representation are impossible, it has been shown that when $n+1$ symbols are used to represent permutations of $[n]$, ucycles are possible \cite{HurlPerms, JohnsonPerms}. Another alternative is to consider $k$-permutations, as done by Jackson \cite{JacksonPerms}. Jackson proved the existence of ucycles for $k$-permutations of $[n]$ with $3 \leq k < n$. As a natural extension to the problem of finding ucycles for permutations, we consider ocycles. For example, Figure \ref{Fig1} gives a 3-ocycle for permutations of $[5]$, which is the largest allowable overlap for permutations of $[5]$. Omitting repeated elements and writing it as a string, the corresponding 3-ocycle is: $$1234512341523 \cdots 34152345.$$ \begin{figure} $$\begin{array}{l} 12345, 34512, 51234, 23415, 41523, 52314, 31425, 42513, 51324, 32415, 41532, 53214, 21435, \\ 43512, 51243, 24315, 31524, 52413, 41325, 32514, 51423, 42315, 31542, 54213, 21345, 34521, \\ 52134, 13425, 42531, 53124, 12435, 43521, 52143, 14325, 32541, 54123, 12354, 35412, 41235, \\ 23514, 51432, 43215, 21534, 53412, 41253, 25314, 31452, 45213, 21354, 35421, 42135, 13524, \\ 52431, 43125, 12534, 53421, 42153, 15324, 32451, 45132, 13245, 24513, 51342, 34215, 21543, \\ 54312, 31245, 24531, 53142, 14235, 23541, 54132, 13254, 25413, 41352, 35214, 21453, 45312, \\ 31254, 25431, 43152, 15243, 24351, 35124, 12453, 45321, 32145, 14523, 52341, 34125, 12543, \\ 54321, 32154, 15432, 43251, 25134, 13452, 45231, 23145, 14532, 53241, 24153, 15342, 34251, \\ 25143, 14352, 35241, 24135, 13542, 54231, 23154, 15423, 42351, 35142, 14253, 25341, 34152, \\ 15234, 23451, 45123 . \end{array}$$ \caption{List of Permutations of $[5]$ in a 3-Ocycle}\label{Fig1} \end{figure} In studying ocycles, it is often useful to consider the \textbf{overlap graph} for the set of objects. This graph has objects represented as vertices. An edge from object $a_1a_2 \ldots a_n$ to object $b_1b_2 \ldots b_n$ exists if and only if $a_{n-s+1}a_{n-s+2} \ldots a_n = b_1b_2 \ldots b_s$. In this transition graph, we are looking for a Hamilton cycle, which will correspond to an $s$-ocycle. In \cite{ResProbs}, the following question was posed by Anant Godbole, which we are able to answer affirmatively in Corollary \ref{main}. \begin{Q}\label{481} \emph{(\cite{ResProbs}, Problem 481.)} Let $P(k,n,s)$ be the overlap graph for $k$-permutations of $[n] = \{1, 2, \ldots , n\}$. Is $P(k,n,s)$ hamiltonian whenever $k<n$? \end{Q} \section{Results} We begin with a lemma from our previous paper on universal cycles for weak orders. \begin{lemma}\label{ocmultiset} \emph{(\cite{UCWO}, Lemma 4.3)} Let $n,s \in \mathbb{Z}^+$ with gcd$(n,s)=1$ and $1 \leq s \leq n-2$, and let $M$ be a multiset of size $n$. Then there is an $s$-ocycle for all permutations of $M$. \end{lemma} We can actually improve this and do even better for $s < \frac{n}{2}$. \begin{lemma}\label{ocmultisetsmall} Let $n,s \in \mathbb{Z}^+$ with $1 \leq s < \frac{n}{2}$. Let $M$ be a multiset of size $n$. Define the set $A$ to be the set of all permutations of $M$. Then there is an $s$-ocycle for all permutations of $A$. \end{lemma} \begin{proof} Construct the transition graph with vertices as $s$-prefixes and $s$-suffixes of words in $A$, and edges representing the words themselves. Fix an arbitrary vertex $v^{s-}=v_1v_2 \ldots v_s$ as the minimum vertex. To prove the existence of an Euler tour, and thus prove the existence of an $s$-ocycle, we will show that from any vertex $w^{s-}=w_1w_2 \ldots w_s$, we can find a path to the minimum vertex. Compare $v^{s-}$ and $w^{s-}$, and consider the first position in which they differ, say index $i$. In other words, $v_i \neq w_i$, and for all $1 \leq j < i$ we have $v_j = w_j$. We will choose any string $w=w_1w_2 \ldots w_n \in A$ so that $w$ has $s$-prefix $w^{s-}$. We have two cases. \begin{enumerate} \item First, if $v_i \in \{w_{i+1}, w_{i+2}, \ldots , w_s\}$, then we use the following undirected path, in which we merely transpose letters $w_i$ and $v_i$ in $w^{s-}$. \begin{eqnarray*} w_1w_2 \ldots w_{i-1}w_iw_{i+1} \ldots v_i \ldots w_s & \rightarrow & w_{n-s+1}w_{n-s+2} \ldots w_n \\ & \leftarrow & w_1w_2 \ldots w_{i-1} v_i w_{i+1} \ldots w_i \ldots w_s \end{eqnarray*} \item Second, if $v_i \in \{w_{s+1}, w_{s+2}, \ldots , w_n\}$, then $v_i$ does not appear in vertex $w^{s-}$. Note that vertex $w^{s-}=w_1w_2 \ldots w_s$ is connected to any $s$-suffix, which consists of all $s$-permutations of the set $$B=M \setminus \{w_1, w_2, \ldots , w_s\}.$$ Thus $v_i \in B$, so we may choose an edge leaving $w^{s-}$ that leads to an $s$-suffix not containing $v_i$, i.e. $v_i$ does not appear in either vertex. In this case, we may simply replace $w_i$ with $v_i$ as shown in the following path. \begin{eqnarray*} w_1w_2 \ldots w_{i-1}w_iw_{i+1} \ldots w_s & \rightarrow & w_{n-s+1}w_{n-s+2} \ldots w_n \\ & \leftarrow & w_1 w_2 \ldots w_{i-1} v_i w_{i+1} \ldots w_s \end{eqnarray*} \end{enumerate} At this point, we are one step closer to the minimum vertex, as now the two vertices agree in the first $i$ positions. Repeating, we will eventually find a path to the minimum vertex. Thus, the graph is connected. Finally, since clearly any $s$-suffix of a string in $A$ is also an $s$-prefix, the graph is balanced and hence eulerian. \end{proof} The previous lemmas immediately give us the following partial solution when we consider $s$-ocycles on permutations. \begin{res}\label{betterMSet} Let $n,s \in \mathbb{Z}^+$ with $n \geq 2$. If either (1) $1 \leq s < \frac{n}{2}$, or (2) gcd$(s,n)=1$ with $\frac{n}{2} \leq s <n-1$, then there exists an $s$-ocycle on the set of permutations of $[n]$. \end{res} \begin{proof} This is merely a concise restatement of Lemmas \ref{ocmultiset} and \ref{ocmultisetsmall} with $M = \{1, 2, \ldots , n\}$. \end{proof} This result gives us an alternative modification for dealing with the problem of finding ucycles for permutations. Instead of increasing the alphabet size from $n$ to $n+1$ as done in \cite{HurlPerms, JohnsonPerms}, an alternative could be to consider the largest possible overlap, with $s=n-2$ being the best alternative to a ucycle. Note that this result also shows that under the given conditions, all permutations of an $[n]$-set are connected within a larger transition graph. We can extend the ocycle result to $k$-permutations of $[n]$ to partially answer Question \ref{481}. \begin{res}\label{kpermOC} Let $n,s,k \in \mathbb{Z}^+$ with $1\leq k<n$. If either (1) $1 \leq s < \frac{k}{2}$, or (2) gcd$(s,k)=1$ with $\frac{k}{2} \leq s < k-1$, then there exists an $s$-ocycle on the set of $k$-permutations of $[n]$. \end{res} \begin{proof} First, construct the transition graph with vertices of length $s$ ($s$-prefixes of $k$-permutations) and edges representing $k$-permutations. We allow an edge from vertex $u$ to vertex $v$ if and only if $u$ is an $s$-prefix and $v$ is an $s$-suffix for some $k$-permutation. We will show that this graph is balanced and weakly connected, and thus is eulerian. From an Euler tour, we can find an $s$-ocycle. Define the minimum $k$-permutation $v = 123 \ldots k$, and the minimum vertex, $v^{s-}$, in the transition graph to be the $s$-prefix of $v$. Let $w^{s-} = w_1 w_2 \ldots w_s$ be the prefix of an arbitrary $k$-permutation $w_1 w_2 \ldots w_k$. By Result \ref{betterMSet}, all permutations of a $k$-set are connected under our hypthoses, so we may assume that $w$ is ordered as $w_1 < w_2 < \cdots < w_k$. We will show that there is a path in the graph from $w^{s-}$ to $v^{s-}$. Define $D = \{w_1, w_2, \ldots , w_k\} \setminus \{1,2,3, \ldots , k\}$ and $\overline{D} = \{1, 2, 3, \ldots , k\} \setminus \{w_1, w_2, \ldots , w_k\}$. Note that if $D = \emptyset$, then $w$ is a permutation of $v$ and so by the comments following Result \ref{betterMSet}, we know that there exists a path. For $D \neq \emptyset$, we choose $d$ letters $a_1, a_2, \ldots , a_d \in \overline{D}$, where $d = \min\{k-s,|\overline{D}|\}$. If $d < k-s$, then we may also select letters $a_{d+1}, a_{d+2}, \ldots , a_{k-s} \in \{w_{s+1}, w_{s+2}, \ldots , w_k\}$. Now in our graph we follow the edge corresponding to the $k$-permutation $w_1w_2 \ldots w_s a_1 a_2 \ldots a_{k-s}$. By following this edge, we have found a $k$-permutation with more letters in common with $v$ than $w$ did. Since Result \ref{betterMSet} implies that all permutations of a $k$-set are connected in the transition graph, we may arrange this string in increasing order, and by repeating this procedure we will eventually find a $k$-permutation that is simply a permutation of $v$, at which point we are done. Since we have shown that the graph is connected, we need only show that the graph is balanced in order to prove that an Euler tour exists. However it is clear that the graph is balanced, as any prefix of a $k$-permutation is also a suffix of a $k$-permutation. \end{proof} We now prove our main result, which will provide a complete solution to Question \ref{481}. \begin{res}\label{allkpermOC} For all $n,s,k \in \mathbb{Z}^+$ with $1 \leq s < k < n$, there is an $s$-ocycle for $k$-permutations of $[n]$. \end{res} \begin{proof} We construct the standard transition graph $G$ where vertices of length $s$ correspond to $s$-prefixes of $k$-permutations of $[n]$, and edges correspond to $k$-permutations of $[n]$. If we can show that this graph is eulerian, then we have shown that there exists an $s$-ocycle for $k$-permutations of $[n]$. First, we note that since any $s$-prefix of a $k$-permutation is also an $s$-suffix of a $k$-permutation, the graph is balanced. All that remains is to show that the graph is connected. Define the minimum vertex $v^{s-} = 12 \ldots s$, and let $w^{s-} = w_1w_2 \ldots w_s$ be an arbitrary vertex in the graph, which we assume to be an $s$-prefix of the $k$-permutation $w = w_1 w_2 \ldots w_k$. We will frequently refer to \textbf{rotations} of a vertex. This is defined as following the edges of the cycle corresponding to rotations of a $k$-permutation that the vertex is an $s$-prefix for. We next compare $w^{s-}$ with $v^{s-}$. Let $i$ be the first index in which $w_i \neq v_i$ and let $g = \hbox{gcd}(s,k)$. If $g=1$, we are done by Result \ref{kpermOC}. Otherwise, we note that rotations of $w$ partition the string into blocks of length $g$. All addition in indices will be modulo $k$. We have two cases. \begin{description} \item[Case 1:] If $i \not \in \{w_{i+1}, w_{i+2}, \ldots , w_k\}$: Rotate $w$ so that $w_i$ is in the first block. This means that we are considering some vertex $$w_{i-t} w_{i-t+1} \ldots w_{i-t+s-1}$$ with $t \leq g-1$, or $i \in [i-t,i-t+g-1]$. Follow the edge out of this vertex that corresponds to a rotation of $w$. This takes us to the vertex $$w_{i-t+k-s} w_{i-t+k-s+1} \ldots w_{i-t+k-1}.$$ Next we follow the backwards edge corresponding to the $k$-permutation $$w_{i-t}w_{i-t+1} \ldots w_{i-1} (i) w_{i+1} \ldots w_{i-t+k-1}.$$ Now we are at the vertex $$w_{i-t}w_{i-t+1} \ldots w_{i-1} (i) w_{i+1} \ldots w_{i-t+s-1}.$$ Finally we follow rotations of this vertex to end at the vertex $$12 \ldots (i) w_{i+1} w_{i+2} \ldots s.$$ This vertex is closer to the minimum vertex since the first $i$ letters agree. Repeating this procedure, we will eventually arrive at the minimum vertex. \item[Case 2:] If $i \in \{w_{i+1}, w_{i+2}, \ldots , w_k\}$: Rotate $w$ so that $i$ is in the first block. We are now considering some vertex $$a_1a_2 \ldots a_s$$ with $i \in \{a_1, a_2, \ldots , a_g\}$, so assume that $i = a_j$. Note that $$|[n] \setminus \{w_1, w_2, \ldots , w_k\}| > 1,$$ so choose some $x$ in this set. We follow the edge from $a_1 a_2 \ldots a_s$ corresponding to a rotation of $w$ $$a_1 a_2 \ldots a_{j-1} (i) a_{j+1} \ldots a_k.$$ This takes us to the vertex $$a_{k-s+1} a_{k-s+2} \ldots a_k,$$ which does not contain the letter $i$. From this vertex we follow the backward edge $$a_1 a_2 \ldots a_{j-1} (x) a_{j+1} \ldots a_k.$$ Note that $i$ does not appear in this edge, so we can go to Case (1). \end{description} When we have finished, we will have arrived at the minimum vertex. Thus the graph is weakly connected, and so is eulerian. \end{proof} Finally, the following answer to Question \ref{481} is an obvious Corollary to Result \ref{allkpermOC}. \begin{corr}\label{main} The permutation overlap graph $P(k,n,s)$ is hamiltonian whenever $k<n$. \end{corr}

https://arxiv.org/abs/1401.6188

On the local convergence of the Douglas-Rachford algorithm

We discuss the Douglas-Rachford algorithm to solve the feasibility problem for two closed sets $A,B$ in $\mathbb{R}^d$. We prove its local convergence to a fixed point when $A,B$ are finite unions of convex sets. We also show that for more general nonconvex sets the scheme may fail to converge and start to cycle, and may then even fail to solve the feasibility problem.

\section{Introduction} The Douglas--Rachford iterative scheme, originally introduced in \cite{douglas} to solve nonlinear heat flow problems, aims to find a point $x^*$ in the intersection of two closed constraint sets $A,B$ in $\mathbb R^d$ or in Hilbert space. Using monotone operator theory, Lions and Mercier \cite{lions} showed that the scheme converges weakly for two closed convex sets $A,B$ in Hilbert space with non-empty intersection. A rather comprehensive analysis of the convex case is given in \cite{combettes}. Due to its success in applications, the Douglas--Rachford scheme is frequently used in the nonconvex setting despite the lack of a satisfactory convergence theory. Recently Hesse and Luke \cite{hesse} made progress by proving local convergence of the scheme for $B$ an affine subspace intersecting the set $A$ transversally, where $A$ is no longer convex, but satisfies a regularity hypothesis called superregularity. Numerical experiments in the nonconvex case indicate, however, that the Douglas--Rachford scheme should converge in much more general situations. Very frequently one observes that the iterates settle for convergence after a chaotic transitory phase; see \cite{aragon,veit} and the references therein. Here we prove local convergence of the Douglas--Rachford scheme when $A,B$ are finite unions of convex sets. Our result is complementary to \cite{hesse}, because no transversality hypothesis is required. This result is proved in section \ref{convergence}. We will also show that for nonconvex sets $A,B$ the Douglas--Rachford scheme may fail to converge and start to cycle without solving the feasibility problem. We show that this may even lead to continuous limiting cycles. These are more delicate to construct, because in that case the Douglas--Rachford sequence $x_{n+1}\in T(x_n)$ is bounded and satisfies $x_{n+1}-x_n\to 0$, but fails to converge. Our construction is given in section \ref{failure}. \section{Preparation} Given a closed subset $A$ of $\mathbb R^d$, the projection onto $A$ is the set-valued mapping $P_A$ defined as \[ P_A(x) = \big\{ a\in A: \|x-a\| = {\rm dist}(x,A) \big\}, \] where $\|\cdot\|$ is the Euclidean norm, and dist$(x,A)=\min\{\|x-a\|: a\in A\}$. The reflection of $x$ in $A$ is then the set-valued operator \[ R_A = 2P_A - I. \] Given two closed sets $A,B$ in $\mathbb R^d$, the Douglas--Rachford iterative scheme, starting at $x_0$, generates a sequence $x_n$ by the recursion \[ x_{n+1} \in T(x_n), \qquad T :=\textstyle \frac{1}{2}\left( R_BR_A+I \right). \] We call $T$ the Douglas--Rachford operator, or shortly, DR operator. Suppose $x^+\in T(x)$ is one step of the Douglas--Rachford scheme. Then $x^+$ is obtained as \[ x^+ = x+b-a, \] where $a \in P_A(x)$, $y=2a-x$, and $b\in P_B(y)$. We call $a$ the shadow of iterate $x$ on $A$, $b$ the reflected shadow of $x$ in $B$, both used to produce $x^+$. We write $x^+=T(x)$ if the DR-operator is single-valued, and similarly for projectors $P_A,P_B$ and reflectors $R_A,R_B$. The fixed point set of $T$ is defined as $\ensuremath{\operatorname{Fix}}(T)=\{x\in\mathbb R^d: x\in T(x)\}$. Note that if $x^*\in \ensuremath{\operatorname{Fix}}(T)$, and if $a^*\in A$, $b^*\in B$ are the shadow and reflected shadow of $x^*$ used to produce $x^*\in T(x^*)$, then $a^*=b^*\in A \cap B$, so every fixed point gives rise to a solution $a^*\in A \cap B$ of the feasibility problem. However, in the set-valued case, it may happen that $x^*\in \ensuremath{\operatorname{Fix}}(T)$ has other shadow-reflected shadow pairs $(\tilde{a},\tilde{b})$ leading away from $x^*$, i.e., where $\tilde{a} \not = \tilde{b}$, so that $\tilde{x}=x^*+\tilde{b}-\tilde{a}\in T(x^*)\setminus \{x^*\}$. We therefore introduce the set of strong fixed-points of $T$ as \[ \ensuremath{\operatorname{{\bold{Fix}}}}(T) = \big\{x\in \mathbb R^n: T(x)=\{x\}\big\}. \] Note that $A \cap B \subset \ensuremath{\operatorname{{\bold{Fix}}}}(T) \subset \ensuremath{\operatorname{Fix}}(T)$. If $T$ is single-valued, then $\ensuremath{\operatorname{Fix}}(T)=\ensuremath{\operatorname{{\bold{Fix}}}}(T)$. These concepts are linked to discrete dynamical system theory, where fixed points are steady states. We recall that a steady state $x^*$ is stable in the sense of Lyapunov if for every $\epsilon > 0$ there exists $\delta > 0$ such that every trajectory $x_{n+1}\in T(x_n)$ with starting point $x_0\in B(x^*,\delta)$ satisfies $x_n\in B(x^*,\epsilon)$ for all $n$. It is clear that $x^*\in \ensuremath{\operatorname{Fix}}(T)\setminus \ensuremath{\operatorname{{\bold{Fix}}}}(T)$ can never be stable, because $x_0=x^*$ produces trajectories going away from $x^*$. \section{Unions of convex sets} \label{convergence} In this section $A=\bigcup_{i\in I}A_i$ and $B=\bigcup_{j\in J} B_j$ are finite unions of closed convex sets, a case which is of interest in a number of practical applications like rank or sparsity optimization \cite{hesse2}, or even road design \cite{road}, where finite unions of linear or affine subspaces are used. For every $i\in I$ and $j\in J$ let $T_{ij}$ be the Douglas--Rachford operator associated with the sets $A_i,B_j$. By convexity of $A_i,B_j$, the operators $T_{ij}$ are single-valued, and $T(x)\subset \{T_{ij}(x): i\in I, j\in J\}$. Since $A,B$ are finite unions of convex sets, every DR step is realized as the DR step of one of the operators $T_{ij}$, and in that case, we say that this operator is active. To make this precise, we define the set of active indices at $x$ as \begin{eqnarray} \label{active} K(x) =\big\{(i,j)\in I\times J: P_{A_i}(x)\in P_A(x), P_{B_j}\left(R_{A_i}\left( x\right)\right)\in P_B\left( R_{A_i}(x) \right)\big\}. \end{eqnarray} Note that if $(i,j)\in K(x)$, then $T_{ij}(x)\in T(x)$. Conversely, for every $x^+\in T(x)$, there exists $(i,j)\in K(x)$ such that $x^+=T_{ij}(x)$. However, be aware that $T_{ij}(x)\in T(x)$ may be true without $(i,j)$ being active at $x$. \begin{theorem} \label{theorem1} {\bf (Local attractor).} Let $A=\bigcup_{i\in I}A_i$ and $B=\bigcup_{j\in J}B_j$ be finite unions of closed convex sets, and let $x^*\in \ensuremath{\operatorname{{\bold{Fix}}}}(T)$ be a strong fixed point. Then $x^*$ has a radius of attraction $R>0$ with the following property: For every $0 < \epsilon < R$, suppose a Douglas--Rachford trajectory $x_{n+1}\in T(x_n)$ enters the ball $B(x^*,\epsilon)$. Then it stays there and converges to some fixed point $\bar{x}\in \ensuremath{\operatorname{Fix}}(T)$. Moreover, every accumulation point of the shadow sequence $a_n\in P_A(x_n)$ is a solution of the feasibility problem. The radius of attraction can be computed as \begin{eqnarray} \label{radius} R = \sup\big\{ \epsilon > 0: K\left( B(x^*,\epsilon)\right)\subset K(x^*)\big\}. \end{eqnarray} \end{theorem} \begin{proof} 1) The fact that $x^*\in \ensuremath{\operatorname{{\bold{Fix}}}}(T)$ is a strong fixed point has the following consequence. Whenever $(i,j)\in K(x^*)$ is active, then $P_{A_i}\left( x^*\right) = P_{B_j}\left( R_{A_i}(x^*)\right)\in A \cap B$. Therefore, for every $(i,j)\in K(x^*)$, the operator $T_{ij}$ has $x^*$ as a fixed point. 2) We now show the following. There exists $\epsilon>0$ such that every $x\in B(x^*,\epsilon)$ has $K(x)\subset K(x^*)$. Let us consider the set $I(x)=\{i\in I: \mbox{there exists } j\in J \mbox{ such that } (i,j)\in K(x)\}$ of active indices $i\in I$ at $x$. Then by definition \begin{eqnarray} \label{neu} \delta_1: = \min\big\{{\rm dist}(x^*,A_i): i\not\in I(x^*)\big\} - {\rm dist}(x^*,A)>0. \end{eqnarray} Similarly, we have \begin{eqnarray} \label{veryneu} \quad \delta_2 := \min \big\{{\rm dist}\left(R_{A_i}(x^*),B_j \right)-{\rm dist}\left(R_{A_i}(x^*),B\right) : i\in I(x^*), (i,j)\not\in K(x^*)\big\}>0. \end{eqnarray} Choose $\epsilon > 0$ such that $2\epsilon < \min\{\delta_1,\delta_2\}$. We show that $\epsilon$ is as claimed. Pick $(i,j)\not\in K(x^*)$. We have to show that $(i,j)\not \in K(x)$. \ First consider the case where $i\in I\setminus I(x^*)$. We show that $i\in I\setminus I(x)$. Indeed, \begin{align*} {\rm dist}(x,A) &\leq \|x-x^*\| + {\rm dist}(x^*,A) \\ &\leq \epsilon + {\rm dist}(x^*,A_i)-\delta_1 \qquad\qquad \mbox{(using (\ref{neu}))}\\ &\leq 2\epsilon + {\rm dist}(x,A_i)-\delta_1 < {\rm dist}(x,A_i), \end{align*} showing that $i\not \in I(x)$. \ We now discuss the case where $i\in I(x^*)$, but $(i,j)\not\in K(x^*)$. That means $ {\rm dist}(R_{A_i}(x^*),B_j) - {\rm dist}(R_{A_i}(x^*),B)\geq \delta_2$. Therefore \begin{align*} {\rm dist}(R_{A_i}(x),B) &\leq \|R_{A_i}(x)-R_{A_i}(x^*)\| + {\rm dist}(R_{A_i}(x^*),B) \\ &\leq \epsilon + {\rm dist}(R_{A_i}(x^*),B_j)- \delta_2\qquad\qquad\qquad \mbox{(using (\ref{veryneu}))} \\ &\leq 2\epsilon + {\rm dist}(R_{A_i}(x),B_j) - \delta_2 < {\rm dist}(R_{A_i}(x),B_j), \end{align*} proving $(i,j)\not\in K(x)$. 3) As an immediate consequence of 2) we have the following: If $x\in B(x^*,\epsilon)$ and $x^+\in T(x)$ is realized as $x^+ = T_{ij}(x)$ for some active operator $T_{ij}$, that is, for some $(i,j)\in K(x)$, then this operator $T_{ij}$ has $x^*$ as a fixed point. Namely, by 2) $x$ satisfies $K(x)\subset K(x^*)$, hence $(i,j)\in K(x^*)$, and therefore $P_{A_i}(x^*) = P_{B_j}(R_{A_i}(x^*))$, which proves what we claimed. 5) We next show that as soon as a DR sequence $x_{n+1}\in T(x_n)$ enters the ball $B(x^*,\epsilon)$, then it stays there and converges. This can be seen as follows. Suppose the trajectory enters $B(x^*,\epsilon)$ at stage $n$. Then the active operator $T_{i_nj_n}$ used to produce $x_{n+1}=T_{i_nj_n}(x_n)\in T(x_n)$ has $x^*$ as a fixed point, because $(i_n,j_n)\in K(x_n)\subset K(x^*)$. Therefore, by \cite[Prop.~4.21]{bauschke_book}, this operator satisfies $\|x_{n+1}-x^*\| = \|T_{i_nj_n}(x_n)- x^*\| \leq \|x_n-x^*\|\leq \epsilon$. The conclusion is that from index $n$ onward the sequence $x_n$ stays in the ball $B(x^*,\epsilon)$, and all operators $T_{i_mj_m}$ used from here on have $x^*$ as a common fixed point. Now we invoke Elsner {\em et al.} \cite[Thm.~1]{elser}, who show that $x_n$ converges to a common fixed point $\bar{x}$ of the operators $T_{i_mj_m}$, $m\geq n$. But $\bar{x}$ is then also a fixed point of $T$, as follows from the continuity of the distance functions. One has $\bar{x}\in B(x^*,\epsilon)$, and moreover, if $a_n = P_{i_n}(x_n)\in A_{i_n}\subset A$, then every accumulation point of the sequence $a_n$ is a solution of the feasibility problem. Namely, if we consider $b_n = P_{B_j}\left( R_{A_i}(x_n)\right)\in B$, and if we take accumulation points $a^*$ of $a_n$ and $b^*$ of $b_n$, then $a^*\in P_A(x^*)$, $b^*\in P_B\left( 2a-x^*)\right)$, hence $a^*=b^*$, because $x^*$ is a strong fixed point. 6) To conclude let us now define the radius of attraction $R$ as in formula (\ref{radius}). In 1) -- 5) above we have shown that there exists $\epsilon>0$ such that $K\left( B(x^*,\epsilon) \right) \subset K(x^*)$, and that this inclusion alone already implies convergence of every trajectory entering $B(x^*,\epsilon)$. This means that the supremum in (\ref{radius}) is over a nonempty set, and that is all that we need. \end{proof} \begin{remark} {\bf (Stable steady state.)} The dynamic system interpretation of Theorem \ref{theorem1} is that a strong fixed point $x^*\in \ensuremath{\operatorname{{\bold{Fix}}}}(T)$ is a stable steady state of the Douglas--Rachford dynamical system $x^+\in T(x)$ when $A,B$ are unions of convex sets. Note also that we do not claim that $\bar{x}\in \ensuremath{\operatorname{Fix}}(T)$ is strong, nor do we claim that the iterates converge to $x^*$ itself. \end{remark} The following observation is also of the essence. \begin{remark} {\bf (Strong fixed point needed).} Theorem \ref{theorem1} is not true if $x^*\in \ensuremath{\operatorname{Fix}}(T)\setminus \ensuremath{\operatorname{{\bold{Fix}}}}(T)$, that is, if $x^*$ is not a strong fixed point. Indeed, let $A = \{-1,1\}$ and $B=\{-2,1\}$. Then $x^*=0$ is a fixed-point of $T$, but not a strong one. Now there exist arbitrarily small $\epsilon \in (0,1)$ such that trajectories starting in $B(0,\epsilon)=(-\epsilon,\epsilon)$ will not stay in that ball. Indeed, for $-\epsilon < x < 0$, $x^+$ will move away from $0$ and will not stay in $B(0,\epsilon)$, while for $0<x <\epsilon$, $x^+$ stays. \hfill $\square$ \end{remark} \begin{remark} Note that if $A,B$ are convex sets, then all $K(x)$ are identical singleton sets, so formula (\ref{radius}) gives $R=\infty$, which means the DR scheme converges globally. \end{remark} \begin{remark} Formula (\ref{radius}) allows to compute the radius of attraction of a strong fixed-point $x^*\in \ensuremath{\operatorname{{\bold{Fix}}}}(T)$ in certain cases. For illustration, consider $A = \{(x,0):x\in \mathbb R\} \cup \{(0,y): y\in \mathbb R\}$ a union of two lines and $B= \{(x,y): y = -\frac{y^*}{x^*}x + y^*\}$ a line, where $x^*>0$, $y^*>0$. Then $(0,y^*)$ and $(x^*,0)$ are the two only fixed points of $T$, both strong, and one easily finds $R(x^*,0)=x^*/\sqrt{2}$ and $R(0,y^*)=y^*/\sqrt{2}$. \end{remark} \begin{remark} {\bf (Asymptotic stability).} Let $x^*\in \ensuremath{\operatorname{{\bold{Fix}}}}(T)$ be a strong fixed point, and suppose there exists $\delta > 0$ such that $B(x^*,\delta )$ containing no further fixed point of $T$. Then it follows from Theorem \ref{theorem1} that we can find $0 < \epsilon \leq \delta$ such that every trajectory $x_{n+1}\in T(x_n)$ entering $B(x^*,\epsilon)$ converges to $x^*$. In the dynamical system terminology, $x^*$ is then asymptotically stable in the sense of Lyapunov. Note that this still fails for an isolated fixed point $x^*\in \ensuremath{\operatorname{Fix}}(T)\setminus \ensuremath{\operatorname{{\bold{Fix}}}}(T)$. \end{remark} \begin{theorem} \label{theorem2} {\bf (Local convergence).} Let $A=\bigcup_{i\in I}A_i$ and $B=\bigcup_{j\in J}B_j$ be finite unions of convex sets. Let $x_{n+1}\in T(x_n)$ be a bounded Douglas--Rachford sequence satisfying $x_{n+1}-x_n\to 0$. Then $x_n$ converges to a fixed-point $\bar{x}\in \ensuremath{\operatorname{Fix}}(T)$. Moreover, every accumulation point of the shadow sequence $a_n\in P_A(x_n)$ is a solution to the feasibility problem. \end{theorem} \begin{proof} 1) For every $n\in \mathbb N$ let us choose an active index pair $(i_n,j_n)\in K(x_n)$ such that $x_{n+1}=T_{i_nj_n}(x_n)$. Put $a_n=P_{A_{i_n}}(x_n)$ and $b_n=P_{B_{j_n}}\left( R_{A_{i_n}}(x_n) \right)=P_{B_{j_n}}(2a_n-x_n)$, so that $x_{n+1}=x_n+b_n-a_n$. Note that $a_n-b_n = x_n-x_{n+1}\to 0$ by hypothesis. 2) Let $x^*$ be any accumulation point of the sequence $x_n$. We define a subset $K_0(x^*)$ of the active set $K(x^*)$ as \[ K_0(x^*) = \big\{(i,j)\in K(x^*): P_{A_i}(x^*)=P_{B_j}\left( R_{A_i}(x^*) \right)\big\}. \] Note that every $T_{ij}$ with $(i,j)\in K_0(x^*)$ has $x^*$ as a fixed point. 3) We now claim that for every accumulation point $x^*$ of the sequence $x_n$ there exists $\epsilon > 0$ and an index $n_0$ such that for every $x_n$ with $n\geq n_0$ and $x_n\in B(x^*,\epsilon)$, we have $(i_n,j_n)\in K_0(x^*)$. To prove this, assume on the contrary that for every $\epsilon = \frac{1}{k}$ there exists $n_k$ such that $x_{n_k} \in B(x^*,\frac{1}{k})$, but $(i_{n_k},j_{n_k})\not \in K_0(x^*)$. Moreover, let $n_k < n_{k+1}\to \infty$. Then $x_{n_k}\to x^*$. Passing to another subsequence which we also denote by $x_{n_k}$, we may assume that $i_{n_k}=i$, $j_{n_k}=j$. Then $a_{n_k}=P_{A_i}(x_{n_k})\to a^*\in A$, $b_{n_k}=P_{B_j}\left( R_{A_i}(x_{n_k}) \right) \to b^*\in B$. Since $a_n-b_n\to 0$ by part 1), we deduce that $a^*=b^*\in A \cap B$. Since we also have $a_{n_k}=P_{A_i}(x_{n_k})\in P_A(x_{n_k})$ and $b_{n_k}\in P_B\left( R_{A_i}(x_{n_k})\right) \subset P_B\left( R_A(x_{n_k})\right)$, we get $a^*\in P_A(x^*)$ and $b^*\in P_B\left( R_A(x^*)\right)$, hence $(i,j)\in K(x^*)$. Since $a^*=b^*$, we have $(i,j)\in K_0(x^*)$. This contradiction proves the claim. 4) Since $x^*$ is an accumulation point of the sequence $x_n$, there exist infinitely many indices with $x_n\in B(x^*,\epsilon)$. Choose one with $n\geq n_0$. Then $x_{n+1}=T_{i_n j_n}(x_n)$, and by part 3) we have $(i_n,j_n)\in K_0(x^*)$. By part 2), $T_{i_nj_n}$ has therefore $x^*$ as a fixed point. Using \cite[Prop.~4.21]{bauschke_book}, we deduce $\|x_{n+1}-x^*\| = \|T_{i_nj_n}(x_n)-x^*\| \leq \|x_n-x^*\|\leq \epsilon$, hence $x_{n+1}$ stays in the ball $B(x^*,\epsilon)$. This means we can repeat the argument, showing that the entire sequence $x_m$, $m\geq n$, stays in $B(x^*,\epsilon)$. By part 3) the operators $T_{i_mj_m}$, $m\geq n$, have the common fixed point $x^*$, hence we conclude again using \cite[Thm.~1]{elser} that $x_m$ converges to some $\bar{x}$, which must be a fixed point of $T$. The second part of the statement follows now from $a_n-b_n\to 0$. \end{proof} \begin{remark} \label{example1} {\bf (Discrete limit cycle).} Let $A= \{(x,y): y=0\}\subset \mathbb R^2$ be the $x$-axis, fix $1\geq \eta > 0$, and put $B = \{(0,0), (7+\eta,\eta),(7,-\eta)\}$. When started at $x_1=(7,\eta)$, the method cycles between the four points $x_1$, $x_2=(7+\eta,0)$, $x_3=(7+\eta,-\eta)$, $x_4=(7,0)$. Note that $B$ is a finite union of bounded convex sets and $A$ is convex, the iterates $x_2$ and $x_4$ reach $A$, but the method fails to converge, and it also fails to solve the feasibility problem. \end{remark} \begin{remark} {\bf (Several shadows)}. Let $B$ be a circle in $\mathbb R^2$, and let $A$ consist in the union of two circles which touch $B$ from outside in $a_1^*,a_2^*$. Then the centre $x^*$ of $B$ is a fixed-point of the Douglas--Rachford operator $T$, and the two points $a_i^*\in A \cap B$ are both shadows of $x^*$. This shows that even in the case of convergence of $x_n$ we do not expect the shadows $a_n$ to converge. \end{remark} \begin{remark} {\bf (More than two sets.)} It is a standard procedure in applications to extend the Douglas--Rachford scheme to solve the feasibility problem for a finite number of constraint sets $C_1,\dots,C_m$ in $\mathbb R^d$. One defines $A$ to be the diagonal in $\mathbb R^d \times \dots \times \mathbb R^d$ ($m$ times), and chooses as $B=C_1 \times \dots \times C_m$ in the product space. Then if $\bigcap_{i=1}^m C_i \not= \emptyset$, the Douglas--Rachford algorithm in product space can be used to compute a point in this intersection. The interesting observation is that if each $C_i$ is a finite union of convex sets, then this remains true for the set $B$, hence our convergence theory applies. \end{remark} \section{Existence of a continuous limit cycle} \label{failure} In this section we construct two closed bounded sets $A,B$ such that the Douglas--Rachford iteration $x_{n+1}\in T(x_n)$ with $T=\frac{1}{2}\left( R_BR_A +I\right)$ fails to converge and produces a continuum of accumulation points $F \subset \ensuremath{\operatorname{Fix}}(T)$ forming a continuous limit cycle. We let $A$ be the cylinder mantle \[ A = \big\{ (\cos t,\sin t,h): 0 \leq t \leq 2\pi, 0 \leq h \leq 1\big\}, \] and $B$ a double spiral consisting of two logarithmic spirals in 3D winding down against the cylinder, one from inside, one from outside. That is, \[ B= \big\{ ((1\pm e^{-t})\cos t,(1 \pm e^{-t} )\sin t, e^{-t/2}): 0 \leq t\big\} \cup F, \] where $F = \{(\cos\alpha,\sin\alpha,0): \alpha \in [0,2\pi]\}$. Note that $A \cap B=F$. We will construct a Douglas--Rachford sequence $x_{n+1}=T(x_n)$, whose set of accumulation points is the entire set $F$. It will be useful to divide the spiral in its outer and inner part \[ B_\pm = \big\{ ((1\pm e^{-t})\cos t,(1 \pm e^{-t} )\sin t, e^{-t/2}): 0 \leq t\big\} \cup F \] so that $B= B_-\cup B_+$ and $B_-\cap B_+=F$ with these notations. \begin{theorem} \label{theorem3} {\bf (Continuous limit cycle).} Let $x_{n+1} = T(x_n)$ be any Douglas--Rachford sequence between $A$ and $B$ with starting point $x_1\in B_-\setminus F$. Then the sequence $x_n$ is bounded, satisfies, $x_{n+1}-x_n\to 0$, but fails to converge. Its set of accumulation points is $F=A\cap B \subset \ensuremath{\operatorname{{\bold{Fix}}}}(T)$. \end{theorem} \begin{proof} 1) For $t\geq 0$ let us introduce the notations \[ a(t)= \big(\cos t,\sin t, e^{-t/2}\big) \in A, \] and \[ b_{\pm}(t) = \big((1\pm e^{-t})\cos t,(1\pm e^{-t})\sin t,e^{-t/2}\big)\in B_{\pm}\setminus F. \] The set $\{a(t): t\geq 0\}\subset A$ is the shadow of the spiral on the cylinder mantle. Namely, it is clear that for $t > 0$, \begin{eqnarray} \label{fifth} P_A\left( b_+(t) \right) = P_A\left( b_-(t)\right) = a(t). \end{eqnarray} In particular, \begin{eqnarray} \label{fourth} \|b_\pm(t) - P_A\left( b_\pm(t) \right)\| = e^{-t}. \end{eqnarray} In consequence \[ R_A\left( b_+(t)\right) = b_-(t), \quad R_A\left( b_-(t)\right) = b_+(t). \] In words, the two branches $B_\pm$ of the double spiral are the reflections of each others in the cylinder mantle. 2) Let us now analyze the projection of $a(t)$ on the double spiral $B$. We consider the projections of $a(t)$ on each of the branches $B_\pm$. We start with the analysis of $b\in P_{B_+}\left( a(t) \right)$. We first claim that $b\not\in F$. Indeed, the point $v\in F$ closest to $a(t)$ is $v= (\cos t,\sin t,0)=P_F\left( a(t)\right)$, so $\|v-a(t)\| = e^{-t/2}$. But $\|b_+(t) - a(t)\| = e^{-t} < e^{-t/2}$, hence there are points on $B_+\setminus F$ closer to $a(t)$ than $v$. This shows that any projected point $b\in P_{B_+}(a(t))$ has to be of the form $b_+(\tau)$ for some $\tau \geq 0$. Now consider some such $b_+(\tau)\in P_{B_+}\left( a(t)\right)$, then \begin{eqnarray} \label{first} e^{-t} = \| a(t)-b_+(t) \| \geq \| a(t) - b_+(\tau) \| \geq \| a(\tau) - b_+(\tau)\| = e^{-\tau}, \end{eqnarray} which shows $\tau \geq t$. Here the second estimate follows from $a(\tau)=P_A\left( b_+(\tau)\right)$. 3) Let us further observe that $\tau > t$. Namely, if we had $\tau = t$, then we would have a fixed point pair for the method of alternating projections between $A$ and $B_+$ in the sense that $a(t) = P_A\left( b_+(t)\right)$, $b_+(t)\in P_{B_+}\left(a(t) \right)$. That would mean the distance squared $\tau \mapsto \frac{1}{2} \| a(t) - b_+(\tau)\|^2$ had a local minimum at $\tau = t$. But the derivative of this function at $\tau = t$ is $-e^{-2t} < 0$, so $\tau = t$ is impossible, and we deduce $\tau > t$. 4) Using $b_+(\tau)\in P_{B_+}\left( a(t)\right)$ and (\ref{first}), we find \[ e^{-t}\geq \|a(t)-b_+(\tau)\| \geq | e^{-t/2}-e^{-\tau/2}| = e^{-t/2}-e^{-\tau/2} =e^{-t} \left( e^{t/2}-e^{t-\tau/2} \right), \] which shows \[ 0 < e^{t/2} - e^{t-\tau/2} \leq 1. \] This can be re-arranged as \begin{eqnarray} \label{bound} 0 < 1 - e^{t/2-\tau/2} \leq e^{-t/2}. \end{eqnarray} In particular, for $t\to \infty$ we must have $\tau-t\to 0$. 5) Let us next show that the projection $b_+(\tau) = P_{B_+}\left( a(t)\right)$ is unique for $t$ sufficiently large. Indeed, suppose we find $t<\tau_1<\tau_2$ such that $b_+(\tau_1),b_+(\tau_2)\in P_{B_+}\left( a(t)\right)$. Then by (\ref{bound}) we have $t < \tau_1 < \tau_2 < t - 2 \log\left( 1-e^{-t/2} \right)$. Define the function $d_+(x) = \frac{1}{2} \|a(t)-b_+(x)\|^2$, then $d_+(t) = \frac{1}{2} e^{-2t}$, $d_+'(t) = -e^{-2t} < 0$. Since $\tau_1,\tau_2$ are local minima, we have $d_+'(\tau_1)=d_+'(\tau_2)=0$. But $d_+''(x) = \cos(t-x) + 2e^{-2x}+2e^{-x}\sin(t-x)+\frac{3}{2}e^{-x}-\frac{1}{4}e^{-x/2-t/2}$. In consequence, for $t$ large and $x$ moving in the interval $x\in \left(t,t - 2 \log\left( 1-e^{-t/2} \right)\right)$, we have $d_+''(x) \approx \cos(t-x)\approx 1$, so certainly $d_+''(x)>0$ for these $x$, and since the local minima $\tau_i$ are in that interval for $t$ large, $d_+'(\tau_2)=0$ is impossible. This proves $b_+(\tau)=P_{B_+}\left( a(t)\right)$ for $t$ sufficiently large. \centerline{ \includegraphics[width=16cm,height=8cm]{spirale.pdf} } \hspace*{-.3cm} 6) Let us now consider the point $b_-(\tau)\in B_-\setminus F$ on the inner spiral. We claim that $b_-(\tau)$ is closer to $a(t)$ than $b_+(\tau)$. Indeed, the set of points $w$ having equal distance to $b_-(\tau)$ and $b_+(\tau)$ is the tangent plane to the cylinder at the point $a(\tau)=\frac{1}{2}\left( b_-(\tau)+b_+(\tau) \right)$. But the cylinder lies in one of the half-spaces associated with this plane, namely the one containing $b_-(\tau)$, $\|b_-(\tau) - a(t)\| < \|b_+(\tau)-a(t)\|$. Since $b_+(\tau)$ is the nearest point to $a(t)$ in $B_+$, we deduce that $P_B\left( a(t) \right) \subset P_{B_-}\left( a(t)\right)$ for all $t$. In other words, projections from the shadow of the spiral onto the double spiral always go to the inner spiral. We could also use an analytic argument to prove this. Let $d_-(x)=\frac{1}{2}\|a(t)-b_-(x)\|^2$ and consider the function $f(x) =d_+(x)-d_-(x)$, then $f(x)=\frac{1}{2}e^{-x}\left( 1-\cos(x-t) \right)$, so $f\ge 0$, and $f=0$ for $x=t+2k\pi$. Since $t<\tau < t-2\log(1-e^{-t/2})\ll t + 2\pi$, this proves $f(\tau) > 0$. 7) Let $b\in P_B\left( a(t)\right) = P_{B_-}\left( a(t)\right)$ a projected point of $a(t)$ in the inner spiral. We know already that $b\not\in F$, hence $b = b_-(\sigma)$ for some $\sigma \geq 0$. Repeating the argument in part 2), it follows that $\sigma > t$. Indeed, like in (\ref{first}) we have \[ e^{-t} = \|a(t)-b_-(t) \| \geq \| a(t)-b_-(\sigma) \| \geq \|a(\sigma)-b_-(\sigma)\| = e^{-\sigma}, \] and the same argument as in part 2) shows $\sigma > t$. But then again \[ e^{-t} \geq \|a(t) - b_-(\sigma) \| \geq | e^{-t/2}-e^{-\sigma/2}| = e^{-t/2}-e^{-\sigma/2} = e^{-t} \left( e^{t/2}-e^{t-\sigma/2} \right), \] which shows \[ 0 < e^{t/2} - e^{t-\sigma/2} \leq 1. \] This can be re-arranged to \begin{eqnarray} \label{second} 0 < 1 - e^{t/2-\sigma/2} \leq e^{-t/2}. \end{eqnarray} In particular, for $t\to \infty$ we must have $\sigma -t\to 0$, and in particular $0<\sigma-t \ll 2\pi$. Therefore projected points $b_-(\sigma)\in P_B\left( a(t)\right)$ lie on the same tour of the spiral as $a(t),b_-(t)$, and one does not take shortcuts by jumping down a full turn of the spiral $B_-$ or more. Repeating the argument of 5), we also see that $b_-(\sigma)=P_{B_-}\left( a(t)\right)$ is unique for $t>0$ large enough. 8) Let us now generate our Douglas--Rachford sequence $x_n$, starting at $x_1=b_-(t_1)\in B_-$ with $t_1 > 0$, excluding $t_1=0$ for simplicity to have a unique projection on the cylinder mantle at the start. We get $a_1=P_A(x_1)= a(t_1)$, hence $R_A(x_1)= b_+(t_1)$. Since $b_+(t_1)\in B$, it is its own reflection in $B$, and we get $b_1 = b_+(t_1)$. Averaging then gives $x_2 = (b_1+x_1) /2= a_1 = a(t_1)$, which concludes the first DR-step. The second DR-step proceeds as follows. Since $x_2 = a(t_1)\in A$, it is its own reflection in $A$, so $a_2 = x_2 = a_1=R_A(x_1)$. Now let $b_2=P_B(R_A(x_2))$, then $b_2 = P_B\left( a(t_1) \right) = P_{B_-}\left( a(t_1) \right)$, so $b_2 = b_-(t_2)$ for some $t_2 > t_1$, where $t_2$ is for $t_1$ what $\sigma$ was for $t$ in part 7). So the reflected point is $2b_-(t_2)-a(t_2)$ and averaging then gives $x_3 = b_-(t_2)\in B_-$. Proceeding in this way, we generate a strictly increasing sequence $t_n$ such that \begin{eqnarray} \label{dr} x_{2k-1} = b_-(t_k) \in B_-, \quad x_{2k} = a(t_k)\in A. \end{eqnarray} Moreover, the shadow and reflected shadow are \[ P_A\left(x_{2k-1} \right) = a_{2k-1}=a(t_k), \quad P_B\left( R_A\left( x_{2k-1}\right)\right)=b_{2k-1}=b_+(t_k) \] respectively, \[ P_A\left( x_{2k} \right)= a_{2k}=a(t_k),\quad P_B\left( R_A\left( x_{2k}\right)\right)= b_{2k}=b_-\left( t_k\right). \] Furthermore, note that we generate a sequence of alternating projections between $A$ and $B_-$. Namely \begin{eqnarray} \label{map} b_0:= x_1 \stackrel{P_A}{\rightarrow} a_1=a_2 \stackrel{P_{B_-}}{\rightarrow}{ b_2} \stackrel{P_A}{\rightarrow} a_3=a_4 \stackrel{P_{B_-}}{\rightarrow}b_4 \dots \end{eqnarray} 9) We now argue that $t_n$ so constructed tends to $\infty$. Suppose on the contrary that $t_n < t_{n+1} \to t^* < \infty$. Then from the construction we see that we create a pair $a(t^*)\in A$, $b_-(t^*)\in B_-$ such that $a(t^*)\in P_{A}(b_-(t^*))$ and $b_-(t^*)\in P_{B_-}(a(t^*))$. Arguing as before, this would imply that $\tau \mapsto \frac{1}{2} \| a(t^*)-b_-(\tau)\|^2$ had a local minimum at $\tau = t^*$, which it does not because its derivative at $t^*$ is $-e^{-2t^*} < 0$. Hence $t^* < \infty$ is impossble, and we have $t_n\to \infty$. As a consequence, the statements about uniqueness of the operators and the estimate (\ref{first}) are now satisfied from some counter $n_0$ onward. 10) To conclude, observe that $x_n-x_{n+1} \to 0$ by (\ref{second}), and that the $a(t_n)$ are $2e^{-t_n}$-dense in the interval $[t_n,t_n+2\pi]$, because of \[ \|a(t_n)-a(t_{n+1})\| \leq \|a(t_n)-b_-(t_{n+1})\| + \|b_-(t_{n+1})-a(t_{n+1})\| \leq e^{-t_n} + e^{-t_{n+1}} \leq 2e^{-t_n}. \] Using (\ref{fourth}) and the fact that every $a(t)$ with $t\in [t_n,t_n+2\pi]$ is at distance $\leq e^{-t_n/2}$ to the set $F$, we deduce that every point in $F$ is an accumulation point of both the DR-sequence (\ref{dr}) and the MAP sequence (\ref{map}). \end{proof} \begin{remark} {\bf (Strong fixed points need not be stable).} The system-theoretic interpretation of this result is that a strong fixed-point $x^*\in F\subset \ensuremath{\operatorname{{\bold{Fix}}}}(T)$ need not be a stable steady state. This is in contrast with Theorem \ref{theorem1}, where this was shown to be true when $A,B$ are finite unions of convex sets. A second interpretation is that $F$ is a stable attractor for the dynamic system $x^+=T(x)$. \end{remark} \begin{remark} {\bf (Shadows need not converge).} We note that not only does the DR sequence $x_n$ fail to converge in Theorem \ref{theorem3}, also the sequences $a_n=P_A(x_n)\in A$, $b_n=P_B(R_A(x_n))\in B$ fail to converge and have the same continuum set of accumulation points $F$. Presently no example of failure of convergence of a local sequence $x_n$ is known where the shadow sequence $a_n$ converges to a single limit $a^*\in A \cap B$. It is clear that this could only happen when $F \subset \{x: \|x-a^*\|=\epsilon\}$ for some $\epsilon > 0$. \end{remark} \begin{corollary} {\bf (Continuous limit cycle for MAP).} Let $x_n$ be the Douglas--Rachford sequence constructed above, and let $a_n\in A$, $b_n\in B$ be the shadows associated with $x_n$. Then $b_{2n}, a_{2n+1}$ is a sequence of alternating projections between the cylinder $A$ and the inner spiral $B_-$. This sequence also fails to converge and has the same set of accumulation points $F=A\cap B$. \hfill$\square$ \end{corollary} \begin{corollary} {\bf (Limit cycle for MAP with one set convex).} Every sequence of alternating projections $a_n,b_n$ between the outer spiral $B_+$ and the solid cylinder {\rm conv}$(A)$ started at $b_1\in B_+\setminus F$ is bounded, satisfies $a_n-b_n\to 0$, but fails to converge and has the set $F = B_+ \cap {\rm conv}(A)$ as its set of accumulation points. \end{corollary} \begin{proof} In part 2) of the proof of Theorem \ref{theorem3} we analyzed this sequence, which is generated by the building blocks $a(t) \to b_+(\tau) \to a(\tau)$. \end{proof} \begin{remark} Here we have an example of a semi-algebraic convex set conv$(A)$, and the spiral $B_+$, which is the projection of a semi-analytic set in $\mathbb R^4$, where the MAP sequence fails to converge and leads to a continuous limit cycle. The first example with a continuous limit cycle appears in \cite{spiral}, but with more pathological sets $A,B$. The fact that $B_+$ is not subanalytic can be deduced from \cite[Cor.~7]{aude}. Currently we do not have an example where the DR-algorithm fails to converge and creates a continuous limit cycle with one of the sets convex. \end{remark} \section*{Acknowledgements} \noindent H.H.\ Bauschke was supported by the Natural Sciences and Engineering Research Council of Canada and the Canada Research Chair Program. D. Noll acknowledges hospitality of the University of British Columbia Kelowna and support by the Pacific Institute of the Mathematical Sciences during the preparation of this paper. The visualization was made possible by GeoGebra \cite{geogebra}.

https://arxiv.org/abs/2207.13455

Compactness and Symmetric Well Orders

We introduce and investigate a topological version of Stäckel's characterization of finite sets, with the goal of obtaining an interesting notion that characterizes or is a close variant of compactness. Define a $T_2$ topological space $(X, \tau)$ to be Stäckel-compact if there is some linear ordering $\prec$ on $X$ such that every non-empty $\tau$-closed set contains a $\prec$-least and a $\prec$-greatest element. We find that compact spaces are Stäckel-compact but not conversely, and Stäckel-compact spaces are countably compact. The equivalence of Stäckel-compactness with countable compactness remains open, but our main result is that this equivalence holds in scattered spaces of Cantor-Bendixson rank $< \omega_2$ under ZFC. Under V=L, the equivalence holds in all scattered spaces.

\section{Introduction and Summary of Results.} A familiar phenomenon in point-set topology is that a purely combinatorial set-theoretic condition that characterizes finiteness of sets will often have a corresponding analogue in topological spaces which characterizes compactness, or at least a variant of compactness (Tao~\cite{Tao} illustrates this with example properties; see also \cite{Engelking,Handbook}). The purpose of this article to is to introduce and investigate the topological analogue of a specific property due to St\"ackel~\cite{Stackel} that characterizes the finiteness of a set, namely the existence of some ordering on the set in which every non-empty subset has a smallest and a greatest element (a ``symmetric well-order''). Section~\ref{sec:formulation} defines the corresponding topological property which we call \emph{St\"ackel-compactness,} namely the existence of some ordering on the space such that every non-empty \emph{closed} subset has a smallest and a greatest element. Our problem is to study how close this notion is to ordinary compactness. In Section~\ref{sec:propstack}, we establish some basic properties of St\"ackel-compactness and find that it has similarities with other variants of compactness such as pseudocompactness and sequential compactness (although distinct from them). We observe that every compact Hausdorff space is St\"ackel-compact but not conversely, showing that the space \(\omega_1\) is St\"ackel-compact. Also, every St\"ackel-compact space is countably compact, and so in metric spaces, St\"ackel-compactness coincides with usual compactness. In Section~\ref{sec:ccompact}, we obtain our main result: \emph{In scattered Hausdorff spaces of Cantor-Bendixson-rank less than \(\omega_2\), St\"ackel-compactness coincides with countable compactness} (Theorem~\ref{theo:main}). Finally, in Section~\ref{sec:bzfc} we go beyond ZFC and combine our results with those from~\cite{mathoverflow} to remove the restriction on Cantor-Bendixson rank: Under V=L, St\"ackel-compactness coincides with countable compactness in arbitrary scattered Hausdorff spaces. We also list a few open questions. \section{Symmetric Well Orders and St\"ackel-compactness.}% \label{sec:formulation} \begin{definition} An order \(\prec\) on a set \(X\) will be called a \emph{symmetric well-order} if every non-empty subset of \(X\) contains both a least and a greatest element (or equivalently, if both \(\prec\) and the reverse order \(\succ\) well order \(X\)). \end{definition} \begin{proposition}[St\"ackel~\cite{Stackel}]\label{prop:stackfin} A set \(X\) is finite if and only if a symmetric well-order can be defined on \(X\). \end{proposition} This is easily proved in ZF without AC (see \cite{Suppes}, p.108 and p.149). \medskip We now define a topological version of symmetric well-ordering, and formulate our main notion, \emph{St\"ackel-compactness.} \begin{definition} Let \(X\) be a Hausdorff topological space. \begin{enumerate} \item An ordering \(\prec\) on \(X\) will be called a \emph{symmetric topological well-order} if every non-empty closed subset of \(X\) has both a least and a greatest element. \item \(X\) will be called \emph{St\"ackel-compact} if there exists some ordering \(\prec\) on \(X\) which is a symmetric topological well order. \end{enumerate} \end{definition} \emph{Caveat: In these definitions, the order \(\prec\) on \(X\) is not assumed to be related to the topology on \(X\) in any other way. In general, the order topology given by the order \(\prec\) will be unrelated to the original topology of \(X\).} \medskip The unit interval \([0,1]\) is St\"ackel-compact, as the usual ordering is itself a symmetric topological well order. More examples will appear below. In the remaining sections, we focus on our main problem: \medskip \noindent \textbf{Problem.\;} How close is St\"ackel-compactness to compactness? \medskip Throughout, we will restrict our attention to Hausdorff topological spaces, and will assume the Axiom of Choice, i.e., work in ZFC. (Notations and results used can be found in standard references in point-set topology and set theory, such as \cite{Engelking,Handbook,Jech}.) \section{Basic Properties of St\"ackel-compact spaces.}% \label{sec:propstack} St\"ackel-compactness shares these properties of ordinary compactness: \begin{proposition} A closed subspace of a St\"ackel-compact space is St\"ackel-compact. If a topology is St\"ackel-compact, then so is any weaker \(T_2\) topology. \end{proposition} \begin{proof} Immediate from the definition. \end{proof} \begin{proposition}\label{prop:compactimpliesSC} Every compact Hausdorff space \(X\) is St\"ackel-compact. \end{proposition} \begin{proof} \(X\) is homeomorphic to a closed subspace of \([0,1]^\mu\) for some ordinal \(\mu\). The lexicographic order on \([0,1]^\mu\) is readily verified to be a symmetric topological well order. Hence \([0,1]^\mu\), and so \(X\), is St\"ackel-compact. \end{proof} \begin{proposition} Every St\"ackel-compact space \(X\) is countably compact. \end{proposition} \begin{proof} Fix an ordering on \(X\) in which every non-empty closed set contains both a least and a greatest element. Since \(X\) is Hausdorff, it suffices to show that every infinite subset of \(X\) has a limit point. Let \(A\) be an infinite subset of \(X\). Then (e.g.\ by Proposition~\ref{prop:stackfin}) \(A\) contains a non-empty subset \(B\) which either has no least element or has no greatest element. \(B\) cannot be closed, since \(X\) is St\"ackel-compact. Hence \(B\) has a limit point, and so \(A\) has a limit point. \end{proof} \begin{corollary}\label{coro:metricequiv} Let \(X\) be a Hausdorff space which is either Lindel\"of or paracompact. Then \(X\) is St\"ackel-compact if and only if \(X\) is compact. In particular, in metrizable spaces St\"ackel-compactness coincides with compactness. \end{corollary} At this point, we have the following implications (in Hausdorff spaces): \[ \text{Compact } \implies \text{ St\"ackel-compact } \implies \text{Countably compact} \] We next ask: Can the first implication be reversed? Is every St\"ackel-compact space compact? It turns out that the answer is negative. \begin{theorem}\label{theo:omegaone} The space \(\omega_1\), consisting of all countable ordinals with the order topology, is St\"ackel-compact. Thus there are St\"ackel-compact spaces which are not compact. \end{theorem} \begin{proof} Let \(<\) denote the usual ordering on ordinals. Partition \(\omega_1\) into two stationary sets \(A\) and \(B\), and let \(\prec\) be the order on \(\omega_1\) in which ``\(A\) is followed by the reverse of \(B\)'', or more precisely by defining \(\alpha \prec \beta\) if and only if either \(\alpha \in A\) and \(\beta \in B\), or \(\alpha, \beta \in A\) and \(\alpha < \beta\), or \(\alpha, \beta \in B\) and \(\alpha > \beta\). Let \(F\) be a non-empty closed subset of \(\omega_1\). We will show that \(F\) contains both a \(\prec\)-least element and a \(\prec\)-greatest element. If \(F\) meets both \(A\) and \(B\) then \(F\) clearly has both a \(\prec\)-least element and a \(\prec\)-greatest element, so assume that either \(F \subseteq A\) or \(F \subseteq B\). Then \(F\) must be countable, since an uncountable closed set in \(\omega_1\) would meet both \(A\) and \(B\). Suppose that \(F \subseteq A\). Now \(F\) is a non-empty countable closed set in \(\omega_1\), and so \(F\) contains a \(<\)-least and a \(<\)-greatest ordinal under the usual ordering \(<\) of the ordinals. But on \(A\), the ordering \(\prec\) coincides with the usual ordering \(<\) of the ordinals, so \(F\) contains both a \(\prec\)-least element and a \(\prec\)-greatest element. (A similar argument applies if \(F \subseteq B\).) Thus \(\omega_1\) is St\"ackel-compact. \end{proof} A similar argument shows that the long line is St\"ackel-compact. Also, (by results below) a limit ordinal \(\lambda\) of uncountable cofinality is St\"ackel-compact if \(\lambda < \omega_2\), and under V=L all limit ordinals of uncountable cofinality are St\"ackel-compact. \medskip We do not know if the product of two St\"ackel-compact spaces is St\"ackel-compact. However, if one of the two spaces is additionally assumed to be compact, then the product will be St\"ackel-compact. \begin{proposition}\label{prop:prodcompsc} Let \(X\) and \(Y\) be Hausdorff spaces. If \(X\) is St\"ackel-compact and \(Y\) is compact, then \(X \times Y\) is St\"ackel-compact. \end{proposition} \begin{proof} By Proposition~\ref{prop:compactimpliesSC}, \(Y\) is also St\"ackel-compact. Therefore we can fix symmetric topological well orders on \(X\) and on \(Y\). Then the lexicographical order on \(X \times Y\) defined by \[ (u,v) \prec (x,y) \iff u \prec x \text{ in \(X\), or } u = x \text{ and \(v \prec y\) in \(Y\)} \] is a symmetric topological well order on \(X \times Y\). To see this, let \(C\) be a non-empty closed subset of \(X \times Y\). Then the projection \(P\) of \(C\) onto \(X\), \[ P := \{ x \in X \colon (x,y) \in C \text{ for some \(y \in Y\)} \} \] is also a non-empty closed subset of \(X\) since \(Y\) is compact. So \(P\) will have a least element, say \(x_0\), with respect to the symmetric topological well order on \(X\). Now the set \(Q := \{ y \in Y \colon (x_0, y) \in C \}\) is a non-empty closed subset of \(Y\) and so will have a least element, say \(y_0\), with respect to the symmetric topological well order on \(Y\). Then \((x_0, y_0)\) will be the least element of \(C\) under the lexicographic order. Similarly \(C\) will also have a largest element. \end{proof} \begin{corollary}% \label{coro:prodexample} The product space \(\omega_1 \times (\omega_1 + 1)\) is St\"ackel-compact. We thus have St\"ackel-compact Tychonov spaces which are not normal. (This answers a question of Rao~\cite{Rao}.) \end{corollary} The basic observations of this section indicate that St\"ackel-compactness behaves in ways similar (but not identical) to some other variants of compactness. Like pseudo\-compactness and countable compactness, St\"ackel-compact\-ness is a necessary but not sufficient condition for compactness in Hausdorff spaces, and in metric spaces it coincides with compactness. Unlike pseudocompactness, St\"ackel-compactness implies countable compactness. We next look at the question of reversal of this implication. \section{The Case of Countably Compact Spaces.} \label{sec:ccompact} We now ask if the second implication mentioned after Corollary~\ref{coro:metricequiv} can be reversed: \smallskip \noindent \textbf{Question.\;} Are all countably compact \(T_2\) spaces St\"ackel-compact? \smallskip We do not know the full answer to this question, but we will prove a partial result and show that certain types of countably compact spaces are St\"ackel compact, under certain restrictions. This is the main result of this article, Theorem~\ref{theo:main}. When we try to improve Theorem~\ref{theo:main} by relaxing its restrictive conditions, we get into set theoretical considerations involving additional hypotheses beyond the standard ZFC axioms (Section~\ref{sec:bzfc}). However, in this section all results are proved under ZFC. First, we set up some standard terminology and notation. \begin{definition} Let \(X\) be a topological space and \(E\) be a subset of \(X\). \begin{enumerate} \item \(\Lim E\) denotes the set of limit points of \(E\). \item \(E\) is \emph{perfect} if \(E\) is closed and dense-in-itself, that is, if \(\Lim E = E\). \item The space \(X\) is \emph{scattered} if no non-empty subset is perfect. \end{enumerate} \end{definition} We get the \emph{Cantor-Bendixson derivatives} of \(X\) by repeatedly applying the \(\Lim\) operation through all ordinals, taking intersections at limit stages: \begin{definition} Let \(X\) be a topological space. For each ordinal \(\alpha\) we define a subset \(X^{(\alpha)}\) of \(X\) by transfinite recursion as follows: \begin{align*} X^{(0)} &:= X,\\ X^{(\alpha+1)} &:= \Lim X^{(\alpha)},\\ X^{(\alpha)} &:= \bigcap_{\beta < \alpha} X^{(\beta)}\quad \text{if \(\alpha\) is a limit ordinal.} \end{align*} \(X^{(\alpha)}\) is called \emph{the \(\alpha\)-th Cantor-Bendixson derivative of \(X\)} (\(\alpha\) any ordinal). \end{definition} The Cantor-Bendixson derivatives \(X^{(\alpha)}\) are closed sets that decrease with \(\alpha\), i.e.\@ \(\alpha < \beta \implies X^{(\alpha)} \supseteq X^{(\beta)}\): \[ X = X^{(0)} \supseteq X^{(1)} \supseteq X^{(2)} \supseteq \;\cdots\; \supseteq X^{(\alpha)} \supseteq X^{(\alpha + 1)} \supseteq \;\cdots\; \] Also, there must be an ordinal \(\rho\) with \(X^{(\rho+1)} = X^{(\rho)}\) (since we can fix an ordinal \(\mu\) such that there is no injective mapping from \(\mu\) into the power set of \(X\) (Hartogs), and so the sets \(X^{(\alpha)}, \alpha < \mu\), cannot be all distinct). \begin{definition} For a topological space \(X\), the least ordinal \(\rho = \rho(X)\) such that \(X^{(\rho+1)} = X^{(\rho)}\) is called the \emph{Cantor-Bendixson rank} (or \emph{CB-rank}) of \(X\). \end{definition} Note that if \(\rho = \rho(X)\) is the Cantor-Bendixson rank of \(X\), then \(X\) is perfect if and only if \(\rho = 0\), and \(X\) is scattered if and only if \(X^{(\rho)} = \emptyset\). Also, if \(X\) is countably compact and scattered, then its CB-rank \(\rho = \rho(X)\) is either a successor ordinal or must have uncountable cofinality (\(\cf \rho > \omega\)). The simplest example of a countably compact scattered Hausdorff space with uncountable CB-rank is \(\omega_1\) (under the usual order topology), which we saw in Theorem~\ref{theo:omegaone} to be St\"ackel-compact. We may therefore try to somehow ``lift the proof'' of Theorem~\ref{theo:omegaone} to general countably compact scattered Hausdorff spaces. This is done in Theorem~\ref{theo:scattered} below under a ``reflection assumption''. \medskip We now state our main result. \begin{theorem}[ZFC]\label{theo:main} Let \(X\) be a scattered Hausdorff space with CB-rank less than \(\omega_2\). Then \(X\) is St\"ackel-compact if and only if \(X\) is countably compact. \end{theorem} Using Theorem~\ref{theo:main}, we get more examples of St\"ackel-compact spaces: \(X := \omega_1 \times \omega_1\), \(Y := (\omega_1 + 1) \times (\omega_1 + 1) \setminus \{ (\omega_1, \omega_1) \} \), and \(Z := \omega_1 \times (\omega_1 + 1)\). (In Corollary~\ref{coro:prodexample}, \(Z\) was shown to be St\"ackel-compact using Proposition~\ref{prop:prodcompsc}, but Proposition~\ref{prop:prodcompsc} does not help for \(X\) or \(Y\).) \medskip The rest of the section is for the proof Theorem~\ref{theo:main}. \begin{definition} Let \(A\) be a set and \(\alpha\) be an ordinal with \(\cf \alpha > \omega\). We say that: \begin{enumerate} \item \emph{\(A\) reflects at \(\alpha\)} if \(A \cap \alpha\) is stationary in \(\alpha\). \item \emph{\(A\) reflects everywhere on \(\rho\)} (where \(\rho\) is an arbitrary ordinal) if \(A\) reflects at \(\beta\) for every \(\beta \leq \rho\) with \(\cf \beta > \omega\). \end{enumerate} \end{definition} \begin{theorem}[ZFC]\label{theo:scattered} Let \(X\) be a countably compact scattered Hausdorff space with CB-rank \(\rho = \rho(X)\), and suppose that there exist disjoint sets \(A\) and \(B\) each of which reflects everywhere on \(\rho\). Then \(X\) is St\"ackel-compact. \end{theorem} \begin{proof} The proof improves upon the proof that \(\omega_1\) is St\"ackel-compact. By the given condition, we can partition \(\rho\) into two disjoint sets \(A, B\) such that each of \(A\) and \(B\) reflects everywhere on \(\rho\). Define, for each ordinal \(\alpha\): \[ Y_\alpha := X^{(\alpha)} \setminus X^{(\alpha + 1)}. \] Then \(\{ Y_\alpha \,\colon\; \alpha < \rho \}\) forms a partition of \(X\). Fix a well order of \(X\) such that \(Y_\alpha\) precedes \(Y_\beta\) in this order if \(\alpha < \beta < \rho\). Now define: \[ X_A := \bigcup_{\alpha \in A} Y_\alpha \qquad \text{ and } \qquad X_B := \bigcup_{\beta \in B} Y_\beta. \] Then \(X_A\) and \(X_B\) form a partition of \(X\). Now take the order on \(X\) in which \(X_A\) precedes \(X_B\), \(X_A\) is ordered by the above well-order, and \(X_B\) is ordered by the reverse of that well-order. We now show that under this order, every non-empty closed subset \(F\) of \(X\) has a least and a greatest element. Given a non-empty closed set \(F\) in \(X\), consider the two sets \(F \cap X_A\) and \(F \cap X_B\). If both of these sets are non-empty, then \(F\) will contain least and greatest elements (since \(X_A\) is well-ordered and \(X_B\) is reverse well-ordered by our new chosen order on \(X\)). So we may assume that one of the sets \(F \cap X_A\) and \(F \cap X_B\) is empty, and without loss of generality that \(F \cap X_B = \emptyset\), that is, \(F \subseteq X_A\). So \(F\) has a least element. We will show that \(F\) has a greatest element as well. Define: \[ C := \{ \alpha < \rho \,\colon\; F \cap Y_\alpha \neq \emptyset \}. \] Then \(C \subseteq A\) by our assumption \(F \subseteq X_A\), and \(C \neq \emptyset\) since \(F \neq \emptyset\). If \(C\) has a largest element \(\mu\), then \(F \cap Y_\mu\) must be non-empty finite by countable compactness, and so will have a largest element, which must then be the greatest element of \(F\). Hence it suffices to show that \(C\) has a largest element. Suppose (for contradiction) that \(C\) does not have a largest element. Then \(\sup C \in \Lim C \,\setminus\, C\). Let: \[ \mu := \min ( \Lim C \,\setminus\, C). \] Thus \(\mu\) is a limit ordinal \(\leq \rho\), and \(\cf \mu \geq \omega_1\) by countable compactness. Note that \(C \cap \mu\) is a closed unbounded set in \(\mu\). Now, since \(B\) reflects at \(\mu\), \(B \cap \mu\) is stationary in \(\mu\), and so \((C \cap \mu) \cap (B \cap \mu)\) must be non-empty. But this implies that \(C \cap B \neq \emptyset\) which is a contradiction since \(C \subseteq A\). \end{proof} Our goal now is to try to use the above theorem to show that if a scattered Hausdorff space is countably compact, then it is St\"ackel-compact. But, as mentioned earlier, we are unable to do this without additional set-theoretic hypothesis beyond ZFC (Section~\ref{sec:bzfc}). In ZFC alone, we can use the next theorem below along with Theorem~\ref{theo:scattered} to obtain the result for spaces with CB-rank \(< \omega_2\), giving us Theorem~\ref{theo:main}. \begin{theorem}[ZFC]% \label{theo:lessthanomega2} If \(\rho < \omega_2\), then there exist disjoint sets \(A\) and \(B\) each of which reflects everywhere on \(\rho\). \end{theorem} \begin{proof} The proof is by induction on \(\rho\). Suppose that \(\rho < \omega_2\) and that for every \(\xi < \rho\) there are disjoint \(A_\xi, B_\xi\) each of which reflects everywhere on \(\xi\). Without loss of generality we can assume that \(\rho\) is a limit ordinal, so there are two cases: \(\cf \rho = \omega\) and \(\cf \rho = \omega_1\). \medskip \textbf{Case 1}: \(\cf \rho = \omega\). We can then choose a countable sequence of ordinals \[ 0 = \rho_0 < \rho_1 < \dots < \rho_n < \rho_{n+1} < \dots \] such that \(\sup_n \rho_n = \rho\). By induction hypothesis, for each \(n \in \omega\) we can fix disjoint sets \(A_n, B_n\) such that both of them reflect everywhere on \(\rho_{n+1}\). Now let: \[ A := \bigcup_{n \in \omega} A_n \cap (\rho_{n+1} \setminus \rho_n) \qquad \text{ and } \qquad B := \bigcup_{n \in \omega} B_n \cap (\rho_{n+1} \setminus \rho_n). \] (For ordinals \(\alpha\) and \(\beta\), the set-difference \(\alpha \setminus \beta\) equals \(\{\xi \,\colon\; \beta \leq \xi < \alpha \}\).) Notice that \(A \cap B = \emptyset\). We show that both \(A\) and \(B\) reflect everywhere on \(\rho\). Suppose that \(\alpha \leq \rho\), with \(\cf \alpha \geq \omega_1\). Then \(0 < \alpha < \rho\) (since \(\cf \rho = \omega\) and \(\cf \alpha \geq \omega_1\)), and so there is \(n\) such that \(\rho_n < \alpha \leq \rho_{n+1}\). Now \(A_n\) reflects everywhere on \(\rho_{n+1}\), so \(A_n \cap \alpha\) is stationary in \(\alpha\), and therefore \(A_n \cap (\alpha \setminus \rho_n)\) is also stationary in \(\alpha\) (as \(\alpha \setminus \rho_n\) is closed unbounded in \(\alpha\)). But \[ A_n \cap (\alpha \setminus \rho_n) \subseteq A_n \cap (\rho_{n+1} \setminus \rho_n) \subseteq A, \] hence \(A \cap \alpha\) is stationary in \(\alpha\). Similarly, \(B \cap \alpha\) is stationary in \(\alpha\). Thus both \(A\) and \(B\) reflect everywhere on \(\rho\). \textbf{Case 2}: \(\cf \rho = \omega_1\). Fix \(E \subseteq \rho\) of order type \(\omega_1\) with \(\sup E = \rho\), and let \(L := \rho \,\cap\, \Lim E\). Then \(L\) is closed unbounded in \(\rho\) of order type \(\omega_1\) and each \(\lambda \in L\) is a limit ordinal of countable cofinality. Enumerate \(L\) increasingly as \(\seq{\lambda_\xi}_{\xi < \omega_1}\): \[ L = \{ \lambda_\xi \,\colon\; 0 \leq \xi < \omega_1 \}, \quad\text{with \(\lambda_\alpha < \lambda_\beta\) for all \(\alpha < \beta < \omega_1\)}. \] Now, for each \(\alpha < \omega_1\), we have \(\lambda_{\alpha+1} < \rho\), so by induction hypothesis we can find disjoint sets \(A_\alpha, B_\alpha\) such that both \(A_\alpha\) and \(B_\alpha\) reflect everywhere on \(\lambda_{\alpha + 1}\). Define: \[ A^* := \bigcup_{\alpha < \omega_1} \left[ A_\alpha \cap (\lambda_{\alpha + 1} \setminus (\lambda_{\alpha} + 1)) \right] = \bigcup_{\alpha < \omega_1} \left[ A_\alpha \cap \{\xi \,\colon\; \lambda_{\alpha} < \xi < \lambda_{\alpha + 1}\} \right], \] \[ B^* := \bigcup_{\alpha < \omega_1} \left[ B_\alpha \cap (\lambda_{\alpha + 1} \setminus (\lambda_{\alpha} + 1)) \right] = \bigcup_{\alpha < \omega_1} \left[ B_\alpha \cap \{\xi \,\colon\; \lambda_{\alpha} < \xi < \lambda_{\alpha + 1}\} \right]. \] Note that the sets \(A^*\), \(B^*\), and \(L\) are pairwise disjoint. As \(L\) is closed unbounded in \(\rho\), we can fix disjoint subsets \(C\) and \(D\) of \(L\) that are stationary in $\rho$. Finally, define: \[ A := A^* \cup C \qquad\text{and}\qquad B := B^* \cup D. \] Then \(A \cap B = \emptyset\). We show that both \(A\) and \(B\) reflect everywhere on \(\rho\). Suppose that \(\alpha \leq \rho\), with \(\cf \alpha > \omega\). If \(\alpha = \rho\), then \(C\), and so \(A\), is stationary in \(\rho = \alpha\). If \(\alpha < \rho\), then \(\lambda_{\xi} \leq \alpha < \lambda_{\xi + 1}\) for some \(\xi < \omega_1\). Since the limit ordinal \(\lambda_{\xi}\) has countable cofinality but \(\cf \alpha > \omega\), we get \(\lambda_{\xi} < \alpha\). Also, \(A_\xi\) reflects everywhere on \(\lambda_{\xi + 1}\), so \(A_\xi \cap \alpha\) is stationary in \(\alpha\), and therefore \(A_\xi \cap (\alpha \setminus (\lambda_{\xi}+1))\) is stationary in \(\alpha\) (as \(\alpha \setminus (\lambda_{\xi}+1)\) is closed unbounded in \(\alpha\)). But \[ A_\xi \cap (\alpha \setminus (\lambda_{\xi}+1)) \subseteq A_\xi \cap (\lambda_{\xi + 1} \setminus (\lambda_{\xi}+1)) \subseteq A^* \subseteq A, \] hence \(A \cap \alpha\) is stationary in \(\alpha\). Similarly, \(B \cap \alpha\) is stationary in \(\alpha\). Thus both \(A\) and \(B\) reflect everywhere on \(\rho\). \end{proof} Theorem~\ref{theo:main} now follows immediately from Theorem~\ref{theo:scattered} and Theorem~\ref{theo:lessthanomega2}. \medskip \section{Beyond ZFC.}% \label{sec:bzfc} This section uses extra set-theoretic hypotheses to improve Theorem~\ref{theo:main}. As usual (see Jech \cite{Jech}), V=L denotes G\"odel's Axiom of Constructibility (which states that every set is constructible), and ZFC\(+\)V=L denotes the axioms of ZFC augmented with this axiom. Under ZFC\(+\)V=L, St\"ackel-compactness coincides with countable compactness in all scattered Hausdorff spaces (Theorem~\ref{theo:veql} below). \smallskip For any well-ordered set \(X\), we will use the notation \(\ord(X)\) to denote the order-type of \(X\), that is, the unique ordinal isomorphic to \(X\). \smallskip Let \(\kappa\) be an uncountable cardinal. Recall Jensen's \emph{square principle} \(\square_\kappa\), a combinatorial set-theoretic hypothesis true under ZFC\(+\)V=L (see Jech \cite{Jech}). The square principle \(\square_\kappa\) is the following statement: \smallskip \noindent \(\square_\kappa\): There is a sequence \(\seq{C_\alpha \colon \alpha < \kappa^+, \text{\(\alpha\) a limit ordinal}}\) of sets, known as a \emph{\(\square_\kappa\)-sequence,} such that for all limit \(\alpha < \kappa^+\) we have: \begin{enumerate} \item \(C_\alpha\) is closed unbounded in \(\alpha\); \item \(\cf \alpha < \kappa\) implies \(|C_\alpha| < \kappa\); and \item \(\beta \in \Lim(C_\alpha)\) implies \(C_\beta = C_\alpha \cap \beta\). \end{enumerate} For such a \(\square_\kappa\)-sequence, we have \(\ord(C_\alpha) \leq \kappa\) for all limit \(\alpha < \kappa+\), \begin{proposition}\label{prop:omegatwo} Assume \(\square_{\omega_1}\). Then \(\omega_2\) has a pair of disjoint subsets which reflect everywhere on \(\omega_2\). \end{proposition} \begin{proof} By \(\square_{\omega_1}\), we can fix a sequence \(\seq{C_\alpha \colon \alpha < \omega_2, \text{\(\alpha\) a limit ordinal}}\) such that for all limit \(\alpha < \omega_2\): \begin{enumerate} \item \(C_\alpha\) is closed unbounded in \(\alpha\); \item \(\cf \alpha = \omega_1\) implies \(\ord(C_\alpha) = \omega_1\); and \item \(\beta \in \Lim(C_\alpha)\) implies \(C_\beta = C_\alpha \cap \beta\). \end{enumerate} Fix disjoint stationary subsets \(A_0, B_0\) of \(\omega_1\) consisting of limit ordinals. For each \(\mu < \omega_2\) with \(\cf \mu = \omega_1\), the set \(C_\mu\) is closed unbounded in \(\mu\) and of order type \(\omega_1\), so we can form ``isomorphic copies of \(A_0\) and \(B_0\) within \(C_\mu\)'' (under the unique order-isomorphism between \(\omega_1\) and \(C_\mu\)) by defining: \[ A_\mu := \{ \xi \in C_\mu \colon \ord(C_\mu \cap \xi) \in A_0 \} \;\text{ and }\; B_\mu := \{ \xi \in C_\mu \colon \ord(C_\mu \cap \xi) \in B_0 \}. \] Note that \(A_\mu \subseteq \Lim C_\mu\) and \(B_\mu \subseteq \Lim C_\mu\). Finally, ``put them all together'' by defining: \[ A := \bigcup_{\substack{\mu < \omega_2 \\ \cf \mu = \omega_1}} A_\mu \qquad\text{ and }\qquad B := \bigcup_{\substack{\mu < \omega_2 \\ \cf \mu = \omega_1}} B_\mu. \] Then \(A\) and \(B\) are disjoint, since if \(\xi \in A_\mu \cap B_\nu\) then \(\xi \in \Lim C_\mu \cap \Lim C_\nu\), so \(C_\mu \cap \xi = C_\xi = C_\nu \cap \xi\), so \(\ord(C_\mu \cap \xi) = \ord(C_\nu \cap \xi)\), which is a contradiction since \(\ord(C_\mu \cap \xi) \in A_0\) and \(\ord(C_\nu \cap \xi) \in B_0\), while \(A_0\) and \(B_0\) are disjoint. Now if \(\mu < \omega_2\) and \(\cf \mu = \omega_1\), then \(A_\mu\) and \(B_\mu\) are stationary in \(C_\mu\) and hence in \(\mu\), so \(A\) and \(B\) reflect at \(\mu\). Also, \(A\) and \(B\) reflect at \(\mu = \omega_2\) as well, since they are stationary in \(\omega_2\). So \(A\) and \(B\) are disjoint sets which reflect everywhere on \(\omega_2\). \end{proof} Combining the above proposition with Theorem~\ref{theo:scattered} we get: \begin{corollary}\label{coro:myfinal} Assuming \(\square_{\omega_1}\), every countably compact scattered Hausdorff space of CB-rank \(\omega_2\) is St\"ackel-compact. \end{corollary} Hamkins \cite{mathoverflow} shows that if the global square principle is assumed, then for every \(\rho \geq \omega_2\), there are disjoint sets which reflect everywhere on \(\rho\) (in fact, there exist disjoint proper classes \(A, B\) of ordinals such that \(A \cap \alpha\) and \(B \cap \alpha\) are stationary in \(\alpha\) for every ordinal \(\alpha\) with uncountable cofinality). Since the global square principle holds under ZFC\(+\)V=L, we can combine the result of Hamkins with Theorem~\ref{theo:scattered} to get our final conclusion. \begin{theorem}\label{theo:veql} Assume ZFC\(+\)V=L, and let \(X\) be a scattered Hausdorff space. Then \(X\) is St\"ackel-compact if and only if it is countably compact. \end{theorem} \emph{Remark.} Eisworth \cite{mathoverflow} notes that under large cardinals, disjoint pairs which reflect everywhere on \(\mu\) may not exist for \(\mu \geq \omega_2\) (although this does not necessarily imply the existence of countably compact spaces which are not St\"ackel-compact). \medskip Many questions about St\"ackel-compact spaces remain unanswered. Here are some examples. Note that St\"ackel-compact spaces are Hausdorff, by definition. \begin{enumerate} \item Are there countably compact Hausdorff spaces which are not St\"ackel-compact? \item Is the product of two St\"ackel-compact spaces St\"ackel-compact? \item Is every St\"ackel-compact space regular? \item Is the continuous image of a St\"ackel-compact space in a Hausdorff space necessarily St\"ackel-compact? \end{enumerate} \section{History, Acknowledgements, and Remarks.} The author initially obtained the results up to Corollary~\ref{coro:myfinal} and presented them in seminars in Detroit and Ann Arbor. Ioannis Souldatos then posted a MathOverflow question in~\cite{mathoverflow}. The answers there by Joel Hamkins and Todd Eisworth showed that the existence of disjoint sets which reflect everywhere is both consistent with and (modulo large cardinals) independent of ZFC. The question whether every St\"ackel-compact space is normal was asked by T.~S.~S.~R.~K.~Rao~\cite{Rao}.

https://arxiv.org/abs/2010.12265

Model Interpretability through the Lens of Computational Complexity

In spite of several claims stating that some models are more interpretable than others -- e.g., "linear models are more interpretable than deep neural networks" -- we still lack a principled notion of interpretability to formally compare among different classes of models. We make a step towards such a notion by studying whether folklore interpretability claims have a correlate in terms of computational complexity theory. We focus on local post-hoc explainability queries that, intuitively, attempt to answer why individual inputs are classified in a certain way by a given model. In a nutshell, we say that a class $\mathcal{C}_1$ of models is more interpretable than another class $\mathcal{C}_2$, if the computational complexity of answering post-hoc queries for models in $\mathcal{C}_2$ is higher than for those in $\mathcal{C}_1$. We prove that this notion provides a good theoretical counterpart to current beliefs on the interpretability of models; in particular, we show that under our definition and assuming standard complexity-theoretical assumptions (such as P$\neq$NP), both linear and tree-based models are strictly more interpretable than neural networks. Our complexity analysis, however, does not provide a clear-cut difference between linear and tree-based models, as we obtain different results depending on the particular post-hoc explanations considered. Finally, by applying a finer complexity analysis based on parameterized complexity, we are able to prove a theoretical result suggesting that shallow neural networks are more interpretable than deeper ones.

\subsection{The complexity of \textsc{MinimumChangeRequired}} \label{ssec:mcr} In what follows we determine the complexity of the \textsc{MinimumChangeRequired} problem for the three classes of models that we consider. \begin{restatable}{proposition}{kminchange} \label{prp:k-minchange} The \textsc{MinimumChangeRequired} query is (1) in $\text{\rm PTIME}$ for FBDDs, (2) in $\text{\rm PTIME}$ for perceptrons, and (3) $\text{\rm NP}$-complete for MLPs. \end{restatable} \begin{proofsketch} This query has been shown to be solvable in \text{\rm PTIME}~for \emph{ordered} binary decision diagrams (OBDDs, a restricted form of FBDDs) by Shih et al.~\cite[Theorem 6]{shih2018formal} (the query is called \emph{robusteness} in the work of Shih et al.~\cite{shih2018formal}). We show that the same proof applies to FBDDs. Recall that in an FBDD every internal node is labeled with a feature index in~$\{1,\ldots,n\}$. The main idea is to compute a quantity~$\mathrm{mcr}_u({\bm{x}}) \in \mathbb{N}\cup \{\infty\}$ for every node~$u$ of the FBDD~$\mathcal{M}$. This quantity represents the minimum number of features that we need to flip in~${\bm{x}}$ to modify the classification~$\mathcal{M}({\bm{x}})$ if we are only allowed to change features associated with the paths from~$u$ to some leaf in the FBDD. One can easily compute these values by processing the FBDD bottom-up. Then the minimum change required for~${\bm{x}}$ is the value~$\mathrm{mcr}_r({\bm{x}})$ where~$r$ is the root of~$\mathcal{M}$, and thus we simply return \textsc{Yes} if~$\mathrm{mcr}_r({\bm{x}})\leq k$, and~\textsc{No} otherwise. For the case of a perceptron~$\mathcal{M}=({\bm{w}},b)$ and of an instance~${\bm{x}}$, let us assume without loss of generality that~$\mathcal{M}({\bm{x}})=1$. We first define the \emph{importance}~$s(i) \in \mathbb{Q}$ of every input feature at position~$i$ as follows: if~$x_i =1$ then~$s(i) \coloneqq w_i$, and if~$x_i=0$ then~$s(i) \coloneqq -w_i$. Consider now the set~$S$ that contains the top~$k$ most important input features for which~$s(i) > 0$. We can easily show that it is enough to check whether flipping every feature in~$S$ changes the classification of~${\bm{x}}$, in which case we return~\textsc{Yes}, and return~\textsc{No} otherwise. Finally, \text{\rm NP}~membership of MCR for MLPs is clear: guess a partial instance~${\bm{y}}$ with~$\mathrm{d}({\bm{x}}, {\bm{y}}) \leq k$ and check in polynomial time that~$\mathcal{M}({\bm{x}}) \neq \mathcal{M}({\bm{y}})$. We prove hardness with a simple reduction from the~\textsc{VertexCover} problem for graphs, which is known to be \text{\rm NP}-complete. \end{proofsketch} Notice that this result immediately yields Theorems~\ref{theo:fbdd-mlp}, \ref{theo:perceptron-mlp}, and \ref{prop:perceptron-fbdd} for the case of MCR. \subsection{The complexity of \textsc{MinimumSufficientReason}} We now study the complexity of \textsc{MinimumSufficientReason}. The following result yields Theorems~\ref{theo:fbdd-mlp}, \ref{theo:perceptron-mlp}, and \ref{prop:perceptron-fbdd} for the case of MSR. \begin{restatable}{proposition}{ksuff} \label{prp:ksuff} The \textsc{MinimumSufficientReason} query is (1) $\text{\rm NP}$-complete for FBDDs (and hardness holds already for decision trees), (2) in~$\text{\rm PTIME}$ for perceptrons, and (3) $\Sigma_2^p$-complete for MLPs. \end{restatable} \begin{proofsketch} Membership of the problem in the respective classes is easy. We show \text{\rm NP}-completeness of the problem for FBDDs by a nontrivial reduction from the \text{\rm NP}-complete problem of determining whether a directed acyclic graph has a dominating set of size at most~$k$~\cite{King_2005}. For a perceptron~$\mathcal{M}=({\bm{w}},b)$ and an instance~${\bm{x}}$, assume without loss of generality that~$\mathcal{M}({\bm{x}})=1$. As in the proof of Proposition~\ref{prp:k-minchange}, we consider the importance of every component of ${\bm{x}}$, and prove that it is enough to check whether the $k$ most important features of ${\bm{x}}$ are a sufficient reason for it, in which case we return~\textsc{Yes}, and simply return~\textsc{No} otherwise. Finally, the~$\Sigma_2^p$-completeness for~MLPs is obtained again using a technical reduction from the problem called \textsc{Shortest Implicant Core}, defined and shown to be~$\Sigma_2^p$-complete by Umans~\cite{Umans2001}. \end{proofsketch} To refine our analysis, we also consider the natural problem of \emph{checking} if a given partial instance is a sufficient reason for an instance. \begin{center} \fbox{\begin{tabular}{rl} Problem: & \textsc{CheckSufficientReason (CSR)} \\ Input: & Model $\mathcal{M}$, instance~${\bm{x}}$ and a partial instance ${\bm{y}}$ \\ Output: & \textsc{Yes}, if ${\bm{y}}$ is a sufficient reason for~${\bm{x}}$ wrt. $\mathcal{M}$, and \textsc{No} otherwise \end{tabular}} \end{center} We obtain the following (easy) result. \begin{restatable}{proposition}{checksuff} \label{prp:checksuff} The query \textsc{CheckSufficientReason} is (1) in~$\text{\rm PTIME}$ for FBDDs, (2)~in~$\text{\rm PTIME}$ for perceptrons, and (3) $\text{\rm co-NP}$-complete for MLPs. \end{restatable} We note that this result for FBDDs already appears in~\cite{darwiche2002knowledge} (under the name of \emph{implicant check}). Interestingly, we observe that this new query maintains the comparisons in terms of c-interpretability, in the sense that~$\mathcal{C}_\text{FBDD}$ and~$\mathcal{C}_\text{Perceptron}$ are strictly more c-interpretable than $\mathcal{C}_\text{MLP}$ with respect to~CSR. \subsection{The complexity of \textsc{CountCompletions}} What follows is our main complexity result regarding the query \textsc{CountCompletions}, which yields Theorems~\ref{theo:fbdd-mlp}, \ref{theo:perceptron-mlp}, and \ref{prop:perceptron-fbdd} for the case of CC. \begin{restatable}{proposition}{counting} \label{prp:counting} The query \textsc{CountCompletions} is (1) in $\text{\rm PTIME}$ for FBDDs, (2) \text{$\#$}\text{\rm P}-complete for perceptrons, and (3) \text{$\#$}\text{\rm P}-complete for MLPs. \end{restatable} \begin{proofsketch} Claim (1) is a a well-known fact that is a direct consequence of the definition of FBDDs; indeed, we can easily compute by bottom-up induction of the FBDD a quantity representing for each node the number of positive completions of the sub-FBDD rooted at that node (e.g., see~\cite{darwiche2002knowledge,wegener2004bdds}). We prove (2) by showing a reduction from the \text{$\#$}\text{\rm P}-complete problem \textsc{$\#$Knapsack}, i.e., counting the number of solutions to a~$0/1$ knapsack input.\footnote{Recall that such an input consists of natural numbers (given in binary) $s_1,\ldots,s_n, k \in \mathbb{N}$, and a solution to it is a set $S \subseteq \{1,\ldots,n\}$ with $\sum_{i \in S} s_i \leq k$.} For the last claim, we show that~MLPs with ReLU activations can simulate arbitrary Boolean formulas, which allows us to directly conclude~(3) since counting the number of satisfying assignments of a Boolean formula is \text{$\#$}\text{\rm P}-complete. \end{proofsketch} \paragraph{Comparing perceptrons and MLPs.} Although the query \textsc{CountCompletions} is \text{$\#$}\text{\rm P}-complete for perceptrons, we can still show that the complexity goes down to \text{\rm PTIME}~if we assume the weights and biases to be integers given in unary; this is commonly called \emph{pseudo-polynomial time}. \begin{restatable}{proposition}{pseudo} \label{prp:pseudo} The query \textsc{CountCompletions} can be solved in pseudo-polynomial time for perceptrons (assuming the weights and biases to be integers given in unary). \end{restatable} \begin{proofsketch} This is proved by first reducing the problem to \textsc{$\#$Knapsack}, and then using a classical dynamic programming algorithm to solve \textsc{$\#$Knapsack} in pseudo-polynomial time. \end{proofsketch} This result establishes a difference between perceptrons and MLPs in terms of CC, as this query remains \text{$\#$}\text{\rm P}-complete for the latter even if weights and biases are given as integers in unary. Another difference is established by the fact that \textsc{CountCompletions} for perceptrons can be efficiently approximated, while this is not the case for MLPs. To present this idea, we briefly recall the notion of {\em fully polynomial randomized approximation scheme} (FPRAS~\cite{jerrum1986random}), which is heavily used to refine the analysis of the complexity of \text{$\#$}\text{\rm P}-hard problems. Intuitively, an FPRAS is a polynomial time algorithm that computes with high probability a $(1-\epsilon)$-multiplicative approximation of the exact solution, for $\epsilon > 0$, in polynomial time in the size of the input and in the parameter $1 / \epsilon$. We show: \begin{restatable}{proposition}{approx} \label{prp:approx} The problem \textsc{CountCompletions} restricted to perceptrons admits an FPRAS (and the use of randomness is not even needed in this case). This is not the case for MLPs, on the other hand, at least under standard complexity assumptions. \end{restatable} \subsection{Specific models} \label{subsec:models} \paragraph{Binary decision diagrams.} A \emph{binary decision diagram} (BDD~\cite{wegener2004bdds}) is a rooted directed acyclic graph~$\mathcal{M}$ with labels on edges and nodes, verifying: (i) each leaf is labeled with~$\true$ or with~$\false$; (ii) each internal node (a node that is not a leaf) is labeled with an element of~$\{1,\ldots,n\}$; and (iii) each internal node has an outgoing edge labeled~$1$ and another one labeled~$0$. Every instance ${\bm{x}}=(x_1,\ldots,x_n)\in \{0,1\}^n$ defines a unique path $\pi_{\bm{x}}$ from the root to a leaf in $\mathcal{M}$, which satisfies the following condition: for every non-leaf node~$u$ in~$\pi_{\bm{x}}$, if~$i$ is the label of~$u$, then the path~$\pi_{\bm{x}}$ goes through the edge that is labeled with~$x_i$. The instance ${\bm{x}}$ is positive, i.e.,~$\mathcal{M}({\bm{x}}) \coloneqq 1$, if the label of the leaf in the path $\pi_{\bm{x}}$ is~$\true$, and negative otherwise. The \emph{size}~$|\mathcal{M}|$ of~$\mathcal{M}$ is its number of edges. A binary decision diagram~$\mathcal{M}$ is \emph{free} (FBDD) if for every path from the root to a leaf, no two nodes on that path have the same label. A \emph{decision tree} is simply an FBDD whose underlying graph is a tree. \paragraph{Multilayer perceptron (MLP).} A multilayer perceptron $\mathcal{M}$ with $k$ layers is defined by a sequence of {\em weight} matrices ${\bm{W}}^{(1)},\ldots,{\bm{W}}^{(k)}$, {\em bias} vectors ${\bm{b}}^{(1)},\ldots,{\bm{b}}^{(k)}$, and {\em activation} functions~$f^{(1)},\ldots,f^{(k)}$. Given an instance ${\bm{x}}$, we inductively define \begin{equation}\label{eq:mlp} {\bm{h}}^{(i)} \coloneqq f^{(i)}({\bm{h}}^{(i-1)}{\bm{W}}^{(i)} + {\bm{b}}^{(i)}) \quad \quad \quad (i \in \{1,\dots,k\}), \end{equation} assuming that ${\bm{h}}^{(0)}\coloneqq {\bm{x}}$. The output of $\mathcal{M}$ on ${\bm{x}}$ is defined as $\mathcal{M}({\bm{x}}) := {\bm{h}}^{(k)}$. In this paper we assume all weights and biases to be rational numbers. That is, we assume that there exists a sequence of positive integers $d_0,d_1,\ldots,d_k$ such that ${\bm{W}}^{(i)}\in \mathbb{Q}^{d_{i-1}\times d_i}$ and ${\bm{b}}^{(i)}\in \mathbb{Q}^{d_i}$. The integer $d_0$ is called the \emph{input size} of $\mathcal{M}$, and $d_k$ the \emph{output size}. Given that we are interested in binary classifiers, we assume that $d_k=1$. We say that an~MLP as defined above has $(k-1)$ \emph{hidden layers}. The {\em size} of an~MLP $\mathcal{M}$, denoted by $|\mathcal{M}|$, is the total size of its weights and biases, in which the size of a rational number $\nicefrac{p}{q}$ is $\log_2(p)+\log_2(q)$ (with the convention that $\log_2(0)=1$). We focus on~MLPs in which all internal functions $f^{(1)},\dots,f^{(k-1)}$ are the ReLU function~$\ensuremath{\operatorname{relu}}(x)\coloneqq \max(0,x)$. Usually, MLP binary classifiers are trained using the \emph{sigmoid} as the output function~$f^{(k)}$. Nevertheless, when an MLP classifies an input (after training), it takes decisions by simply using the \emph{pre activations}, also called \emph{logits}. Based on this and on the fact that we only consider already trained~MLPs, we can assume without loss of generality that the output function~$f^{(k)}$ is the \emph{binary step} function, defined as~$\ensuremath{\operatorname{step}}(x)\coloneqq 0$ if~$x<0$, and~$\ensuremath{\operatorname{step}}(x)\coloneqq 1$ if~$x\geq 0$. \paragraph{Perceptron.} A perceptron is an MLP with no hidden layers (i.e., $k=1$). That is, a perceptron~$\mathcal{M}$ is defined by a pair~$({\bm{W}},{\bm{b}})$ such that~${\bm{W}}\in\mathbb{Q}^{d\times 1}$ and~${\bm{b}}\in \mathbb{Q}$, and the output is~$\mathcal{M}({\bm{x}})=\ensuremath{\operatorname{step}}({\bm{x}}{\bm{W}}+{\bm{b}})$. Because of its particular structure, a perceptron is usually defined as a pair~$({\bm{w}},b)$ with~${\bm{w}}$ a rational vector and~$b$ a rational number. The output of~$\mathcal{M}({\bm{x}})$ is then~$1$ if and only if $\langle {\bm{x}},{\bm{w}}\rangle + b\geq 0$, where~$\langle {\bm{x}},{\bm{w}}\rangle$ denotes the dot product between~${\bm{x}}$ and~${\bm{w}}$. \subsection{Specific queries} Given instances~${\bm{x}}$ and~${\bm{y}}$, we define~$\mathrm{d}({\bm{x}},{\bm{y}}) \coloneqq \sum_{i=1}^n|{\bm{x}}_i-{\bm{y}}_i|$ as the number of components in which~${\bm{x}}$ and~${\bm{y}}$ differ. We now formalize the minimum-change-required problem, which checks if the output of the model can be changed by flipping the value of at most~$k$ components in the input. \begin{center} \fbox{\begin{tabular}{rl} Problem: & \textsc{MinimumChangeRequired (MCR)} \\ Input: & Model $\mathcal{M}$, instance ${\bm{x}}$, and~$k \in \mathbb{N}$ \\ Output: & \textsc{Yes}, if there exists an instance ${\bm{y}}$ with $\mathrm{d}({\bm{x}},{\bm{y}}) \leq k$ \\ & and $\mathcal{M}({\bm{x}})\neq\mathcal{M}({\bm{y}})$, and \textsc{No} otherwise \end{tabular}} \end{center} Notice that, in the above definition, instead of ``finding'' the minimum change we state the problem as a \textsc{Yes}/\textsc{No} query (a decision problem) by adding an additional input~$k \in \mathbb{N}$ and then asking for a change of size at most~$k$. This is a standard way of stating a problem to analyze its complexity~\cite{arora2009computational}. Moreover, in our results, when we are able to solve the problem in \text{\rm PTIME}~then we can also output a minimum change, and it is clear that if the decision problem is hard then the optimization problem is also hard. Hence, we can indeed state our problems as decision problems without loss of generality. To introduce our next query, recall that a partial instance is a vector~${\bm{y}}=(y_1,\ldots,y_n) \in \{0,1,\bot\}^n$, and a completion of it is an instance~${\bm{x}}=(x_1,\ldots,x_n) \in \{0,1\}^n$ such that for every~$i$ where~$y_i \in \{0,1\}$ it holds that~$x_i = y_i$. That is, ${\bm{x}}$ coincides with ${\bm{y}}$ on all the components of~${\bm{y}}$ that are not $\bot$. Given an instance~${\bm{x}}$ and a model~$\mathcal{M}$, a \emph{sufficient reason for~${\bm{x}}$ with respect to~$\mathcal{M}$}~\cite{shih2018symbolic} is a partial instance~${\bm{y}}$, such that ${\bm{x}}$ is a completion of~${\bm{y}}$ and every possible completion~${\bm{x}}'$ of~${\bm{y}}$ satisfies~$\mathcal{M}({\bm{x}}')=\mathcal{M}({\bm{x}})$. That is, knowing the value of the components that are defined in~${\bm{y}}$ is enough to determine the output~$\mathcal{M}({\bm{x}})$. Observe that an instance~${\bm{x}}$ is always a sufficient reason for itself, and that~${\bm{x}}$ could have multiple (other) sufficient reasons. However, given an instance~${\bm{x}}$, the sufficient reasons of~${\bm{x}}$ that are most interesting are those having the least possible number of defined components; indeed, it is clear that the less defined components a sufficient reason has, the more information it provides about the decision of~$\mathcal{M}$ on~${\bm{x}}$. For a partial instance~${\bm{y}}$, let us write~$\|{\bm{y}}\|$ for its number of components that are not~$\bot$. The previous observations then motivate our next interpretability query. \begin{center} \fbox{\begin{tabular}{rl} Problem: & \textsc{MinimumSufficientReason (MSR)} \\ Input: & Model $\mathcal{M}$, instance ${\bm{x}}$, and~$k \in \mathbb{N}$ \\ Output: & \textsc{Yes}, if there exists a sufficient reason~${\bm{y}}$ for ${\bm{x}}$ wrt.~$\mathcal{M}$ with $\|{\bm{y}}\|\leq k$, \\ & and \textsc{No} otherwise \end{tabular}} \end{center} As for the case of MCR, notice that we have formalized this interpretability query as a decision problem. The last query that we will consider refers to counting the number of {positive completions} for a given partial instance. \begin{center} \fbox{\begin{tabular}{rl} Problem: & \textsc{CountCompletions (CC)} \\ Input: & Model $\mathcal{M}$, partial instance ${\bm{y}}$ \\ Output: & The number of completions ${\bm{x}}$ of ${\bm{y}}$ such that $\mathcal{M}({\bm{x}})=1$ \end{tabular}} \end{center} Intuitively, this query informs us on the proportion of inputs that are accepted by the model, given that some particular features have been fixed; or, equivalently, on the \emph{probability} that such an instance is accepted, assuming the other features to be uniformly and independently distributed. \subsection{Main interpretability theorems} We can now state our main theorems, which are illustrated in Figure~\ref{fig:main-results}. In all these theorems we use~$\mathcal{C}_\text{MLP}$ to denote the class of all models (functions from $\{0,1\}^n$ to $\{0,1\}$) that are defined by MLPs, and similarly for~$\mathcal{C}_\text{FBDD}$ and~$\mathcal{C}_\text{Perceptron}$. The proofs for all these results will follow as corollaries from the detailed complexity analysis that we present in Section~\ref{sec:complexity}. We start by stating a strong separation between FBDDs and~MLPs, which holds for all the queries presented above. \begin{theorem}\label{theo:fbdd-mlp} $\mathcal{C}_\text{\emph{FBDD}}$ is strictly more c-interpretable than $\mathcal{C}_\text{\emph{MLP}}$ with respect to \textsc{MCR}, \textsc{MSR}, and~\textsc{CC}. \end{theorem} For the comparison between perceptrons and MLPs, we can establish a strict separation for MCR and~MSR , but not for CC. In fact, CC has the same complexity for both classes of models, which means that none of these classes strictly ``dominates'' the other in terms of c-interpretability for~CC. \begin{theorem}\label{theo:perceptron-mlp} $\mathcal{C}_\text{\emph{Perceptron}}$ is strictly more c-interpretable than $\mathcal{C}_\text{\emph{MLP}}$ with respect to ~\textsc{MCR} and \textsc{MSR}. In turn, the problems $\textsc{CC}(\mathcal{C}_\text{\emph{Perceptron}})$ and $\textsc{CC}(\mathcal{C}_\text{\emph{MLP}})$ are both complete for the same complexity class. \end{theorem} The next result shows that, in terms of c-interpretability, the relationship between FBDDs and perceptrons is not clear, as each one of them is strictly more c-interpretable than the other for some explainability query. \begin{theorem}\label{prop:perceptron-fbdd} The problems~$\textsc{MCR}(\mathcal{C}_\text{\emph{FBDD}})$ and~$\textsc{MCR}(\mathcal{C}_\text{\emph{Perceptrons}})$ are both in~\text{\rm PTIME}. However,~$\mathcal{C}_\text{\emph{Perceptron}}$ is strictly more c-interpretable than~$\mathcal{C}_\text{\emph{FBDD}}$ with respect to~\textsc{MSR}, while~$\mathcal{C}_\text{\emph{FBDD}}$ is strictly more c-interpretable than~$\mathcal{C}_\text{\emph{Perceptron}}$ with respect to~\textsc{CC}. \end{theorem} We prove these results in the next section, where for each query~$Q$ and class of models~$\mathcal{C}$ we pinpoint the exact complexity of the problem~$Q(\mathcal{C})$. \begin{figure}[H] \input{figures/diagram-results} \caption{Illustration of the main interpretability results. Arrows depict that the pointed class of models is harder with respect to the query that labels the edge. We omit labels (or arrows) when a problem is complete for the same complexity class for two classes of models.} \label{fig:main-results} \end{figure} \section{Introduction} \label{sec:introduction} \input{introduction} \section{A framework to compare interpretability} \label{sec:framework} \input{framework} \section{Instantiating the framework and main results} \label{sec:instantiation} \input{instantiation} \section{The complexity of explainability queries} \label{sec:complexity} \input{complexity} \section{Parameterized results for MLPs in terms of number of layers} \label{sec:p-complexity} \input{p-complexity} \section{Discussion and concluding remarks} \label{sec:discussion} \input{discussion} \section{Broader impact} Although interpretability as a subject may have a broad practical impact, our results in this paper are mostly theoretic, so we think that this work does not present any foreseeable societal consequences. \begin{ack} Barceló and Pérez are funded by Fondecyt grant 1200967. \end{ack} \newpage \bibliographystyle{abbrv} \subsection{Hardness} \label{subsec:p-hardness} \input{p-hardness} \subsection{Membership} \label{subsec:p-membership} \input{p-membership} \section{Simulating Boolean formulas/circuits with MLPs} \label{sec:simulation} \input{simulations} \section{Proof of Proposition~\ref{prp:k-minchange}} \label{sec:proof-5} \input{proof-5} \section{Proof of Proposition~\ref{prp:ksuff}} \label{sec:proof-6} \input{proof-6} \section{Proof of Proposition~\ref{prp:checksuff}} \label{sec:proof-7} \input{proof-7} \section{Proof of Proposition~\ref{prp:counting}} \label{sec:proof-8} \input{proof-8} \section{Proof of Proposition~\ref{prp:pseudo}} \label{sec:proof-9} \input{proof-9} \section{Proof of Proposition~\ref{prp:approx}} \label{sec:proof-10} \input{proof-10} \section{Background in parameterized complexity} \label{sec:p-background} \input{p-background} \section{Proof of Proposition~\ref{prp:mlpt}} \label{sec:proof-11} \input{proof-11} \section{Proof of Proposition~\ref{prp:result-layers}} \label{sec:proof-12} \input{proof-12}

https://arxiv.org/abs/2211.01596

Probability bounds for $n$ random events under $(n-1)$-wise independence

A collection of $n$ random events is said to be $(n - 1)$-wise independent if any $n - 1$ events among them are mutually independent. We characterise all probability measures with respect to which $n$ random events are $(n - 1)$-wise independent. We provide sharp upper and lower bounds on the probability that at least $k$ out of $n$ events with given marginal probabilities occur over these probability measures. The bounds are shown to be computable in polynomial time.

\section{Introduction} \label{sec:intro} Let $\Omega = \{\omega_J : \, J \subseteq [n]\}$ be the sample space freely generated by $n$ random events $A_1, \dots, A_n$, so that $\{\omega_J\} = \bigcap_{j \in J} A_j \cap \bigcap_{j \notin J} \overline{A}_j =: \A^J$ for all subsets $J \subseteq [n]$. Here, as usual, for any integer $n \ge 0$, let $[n] := \{1, \dots, n\}$, let $\subseteq$ (resp.\ $\subset$) denote the subset (resp.\ proper subset) relation, let $\overline{A}$ denote the complement of $A$ and let $|J|$ denote the cardinality of a set $J$. With $\Sigma$ as the $\sigma$-algebra of all subsets of $\Omega$, the unique probability measure $P$ on the measurable space $(\Omega, \Sigma)$ with respect to which $A_1, \dots, A_n$ are mutually independent and whose marginal probabilities are $P(A_j) =: a_j$ for all $j \in [n]$ is given by: \begin{equation} \label{eq: mutualIndependenceSolution} P(\A^J) = \prod_{j \in J} a_j \times \prod_{j \notin J} (1 - a_j) =: \a^J, \quad \text{for all } J \subseteq [n]. \end{equation} We may relax the condition that the $n$ events are mutually independent to require only that every $(n - 1)$ events among $A_1, \dots, A_n$ are mutually independent. This weaker condition is sometimes known as \emph{$(n - 1)$-wise independence}. It is known that $(n - 1)$-wise independence does not imply that the $n$ events are mutually independent in general (see \cite{Wang,wangstoy} for counterexamples), although the converse is true. Comparisons of various notions of independence and dependence for random variables can be found in the literature \cite{Feller, Mukhopadhyay, Stoyanov, Wong}. The special case of Bernoulli random variables correspond to the setting of random events in this paper. Bernstein \cite[p.\ 126]{Feller} constructed his classic example of $n = 3$ pairwise independent events which are not mutually independent. The example in \cite{Wang} generalizes his construction to $n \ge 3$ random events that are $(n - 1)$-wise independent but not mutually independent. We discuss this construction next. \begin{example} [Construction of $(n-1)$-wise independent events] \label{ex:wang} Let $a_j = 1/2$ for all $j \in [n]$ where $A_1,\ldots,A_{n-1}$ are mutually independent events and suppose that the event $A_n$ occurs given that an even number of events among $A_1,\ldots,A_{n-1}$ occur. It can be verified that these $n$ events are $(n - 1)$-wise independent but not mutually independent. The induced probability measure is not the unique one with respect to which $A_1,\ldots,A_{n}$ are $(n - 1)$-wise independent but not mutually independent. For example, if we suppose instead that $A_n$ occurs given that an odd number of events among $A_1,\ldots,A_{n-1}$ occur, we obtain a distinct probability measure. \end{example} \subsection*{Overview} In Section \ref{sec:characterization_(n-1)independence}, we characterize all probability measures with respect to which $n$ random events are $(n - 1)$-wise independent (Theorem \ref{thm: characterisationOfProbabilityMeasures}). Now we briefly outline the steps leading to Theorem \ref{thm: characterisationOfProbabilityMeasures}. First, in Proposition \ref{prop: formalEquivalence}, we identify a system of $2^n - 1$ equations linear in $P(\A^I)$ for all $I \subseteq [n]$ that is satisfied if and only if $A_1, \dots, A_n$ are $(n - 1)$-wise independent. We then relax the condition that $P(\A^I)$ are nonnegative and show in Lemma \ref{lem:1} and Corollary \ref{cor: characterisationOfUnsignedMeasures} that the solutions are parameterized by a single real parameter. Reimposing the condition that $P(\A^I)$ are nonnegative forces this parameter to lie within a compact interval which is identified in \eqref{eq: solveSystemForPositivityOfMeasure}. Lemmas \ref{lem2}--\ref{lem: minimizeAtomicProbabilityForUpperBound} then explicitly identifies the endpoints of this interval in terms of the marginal probabilities. Theorem \ref{thm: characterisationOfProbabilityMeasures} then follows. Section \ref{sec:atleastksharpbounds} applies Theorem \ref{thm: characterisationOfProbabilityMeasures} to identify sharp bounds on the probability that at least $k$ out of $n$ events which are $(n - 1)$-wise independent occurs (Theorem \ref{thm:atleastksharpbounds}) while Theorem \ref{thm:complexity} shows that these bounds are computable in polynomial time. Examples \ref{ex:reductiontopairwise} and \ref{ex:reductiontobonferroni} in Section \ref{sec:examples} give cases when the newly derived bounds are instances of known universal bounds such as the classical Bonferroni bounds. Example \ref{ex:lovaszlocallemma} illustrates the connection of the results to the probabilistic method which provides conditions for the non-occurrence of ``bad" events when events are mostly independent. Example \ref{ex:comparisonmakarov} illustrates the usefulness of the bounds in providing robust estimates when mutual independence breaks down or when existing bounds are not sharp. \section{Characterization of $(n-1)$-wise independence}\label{sec:characterization_(n-1)independence} \begin{proposition} \label{prop: formalEquivalence} A collection of $n$ random events $A_1, \dots, A_n$ is $(n - 1)$-wise independent with $P(A_j) = a_j$ for all $j \in [n]$ if and only if: \begin{equation} \label{eq: n-1wiseIndependence} \sum_{I \supseteq J} P(\A^I) = \prod_{j \in J} a_j, \quad \text{for all } J \subset [n]. \end{equation} \end{proposition} \begin{proof} Since the $\A^I$'s are mutually exclusive, we have $\sum_{I \supseteq J} P(\A^I) = P(\bigcap_{j \in J} A_j) $ for any $J$. Thus \eqref{eq: n-1wiseIndependence} is equivalent to $$P(\bigcap_{j \in J} A_j) = \prod_{j \in J} a_j, \quad \text{for all } J \subset [n],$$ which is in turn equivalent to the $(n - 1)$-wise independence of $A_1, \dots, A_n$. \end{proof} The system of linear equations in \eqref{eq: n-1wiseIndependence} is inhomogeneous with \eqref{eq: mutualIndependenceSolution} as a particular solution. \begin{lemma} \label{lem:1} The general solution of the system of linear equations: \begin{equation} \label{eq: associatedHomogeousSystem} \sum_{I \supset J} P(\A^I) = 0, \quad \text{for all } J \subset [n], \end{equation} in the $2^n$ variables $\{P(\A^J)\}_{J \subseteq [n]}$ (not necessarily nonnegative), is given by $P(\A^J) = (-1)^{|J|} s$, where $s \in \mathbb{R}$ is a free parameter. \end{lemma} \begin{proof} The following inhomogeneous linear system has a unique solution, namely the mutually independent probability measure in \eqref{eq: mutualIndependenceSolution}: $$\sum_{I \supseteq J} P(\A^I) = \prod_{j \in J} a_j, \quad \text{for all } J \subseteq [n].$$ Hence the square coefficient matrix of its associated homogeneous linear system is invertible. Thus, the equations of the homogeneous subsystem \eqref{eq: associatedHomogeousSystem} are linearly independent. Therefore its solution space has dimension $1$, because there are $2^n - 1$ equations and $2^n$ variables. But, substituting $P(\A^I) = (-1)^{|I|} s$ into \eqref{eq: associatedHomogeousSystem}, where $s \in \mathbb{R}$ is a parameter, we have for any $J \subset [n]$: \begin{align*} \sum_{I \supset J} (-1)^{|I|} s &= s \sum_{q = |J|}^n \binom{n - |J|}{q - |J|}(-1)^q\\ &= s (-1)^{|J|} \sum_{q = |J|}^n \binom{n - |J|}{q - |J|}(-1)^{q - |J|} \notag \\ &= s(-1)^{|J|}(1 - 1)^{n - |J|} \\ &= 0, \end{align*} where we use the Binomial Theorem in the third equality and the fact that $n - |J| > 0$ for all $J \subset [n]$ in the last equality. Thus $P(\A^J) = (-1)^{|J|} s$. \end{proof} \noindent Recall that a measure $P$ is \emph{unitary} if $P(\Omega) = 1$. \begin{corollary} \label{cor: characterisationOfUnsignedMeasures} Every unitary measure $P$ (not necessarily nonnegative) on $(\Omega, \Sigma)$ with $P(A_j) = a_j$ for all $j \in [n]$ and with respect to which $A_1, \dots, A_n$ are $(n - 1)$-wise independent has the form: \begin{equation} \label{eq:characterisationOfN-1wiseUnsignedMeasures} P(\A^J) = \a^J + (-1)^{|J|} s, \quad \text{for all } J \subseteq [n], \end{equation} for some scalar parameter $s \in \mathbb{R}$. \end{corollary} \begin{proof} Add the particular solution $P(\A^J) = \a^J$ (when the collection of events $\{A_1, \dots, A_n\}$ is mutually independent) with the general solution of the associated homogenous linear system given in the previous lemma. \end{proof} \noindent To characterise when $P$ in \eqref{eq:characterisationOfN-1wiseUnsignedMeasures} is a valid probability measure, we have to ensure that the following $2^n$ nonnegativity conditions are satisfied: \begin{equation} \label{eq: systemOfLinearInequalitiesToEnsurePositivity} P(\A^J)=\a^J + (-1)^{|J|} s \ge 0, \quad \text{for all } J \subseteq [n]. \end{equation} Simplifying, this system gives: \begin{equation} \begin{cases} s \ge -\a^J, &\text{ for all }J \subseteq [n]:\, |J| \text{ is even}, \\ s \le \a^J, &\text{ for all } J \subseteq [n]:\, |J| \text{ is odd}. \end{cases} \end{equation} Therefore it is a valid probability measure for all values of $s$ that satisfy: \begin{equation} \label{eq: solveSystemForPositivityOfMeasure} - \min_{J \subseteq [n]:\, |J| \text{ is even}} \a^J \le s \le \min_{J \subseteq [n]:\, |J| \text{ is odd}} \a^J. \end{equation} \noindent From this point onwards, we make the following assumption. \begin{assumption} \label{ass:order} The events are ordered by nondecreasing value of their marginal probabilities, i.e. $a_1 \le \cdots \le a_n$. \end{assumption} \noindent The next lemma provides a lower bound on $\a^J$ for any set $J \subseteq [n]$, which will be used to establish the precise interval for the parameter $s$ in \eqref{eq: solveSystemForPositivityOfMeasure} and thus to identify all probability measures with respect to which $A_1, \dots, A_n$ are $(n - 1)$-wise independent. \begin{lemma} \label{lem2} $\a^J \ge \a^{\initialSubset{|J|}}$ for all $J \subseteq [n]$. \end{lemma} \begin{proof} Use the notation $J \succeq I$ to denote $\a^J \ge \a^I$. Say $|J| = \ell$. We need to show that $J \succeq \initialSubset{\ell} = \{1,\ldots,\ell\}$. If $J=[\ell]$, we are done. Otherwise, if $J \neq \initialSubset{\ell}$, then there is a smallest index $r \le \ell$ that is not in $J$. Indeed, if there did not exist such an $r$, then the smallest index not in $J$ is strictly greater than $\ell$, hence $J \supset [\ell]$, which contradicts $|J| = \ell$. Hence the smallest index not in $J$, which we denote by $k$, satisfies $k > \ell$. Let $J^\prime := J \cup \{r\} \setminus \{k\}$ be the set obtained by replacing $k$ with $r$ in $J$; hence $|J^\prime| = |J| = \ell$ have the same cardinality. Now, a common factor of $\a^J$ and $\a^{J^\prime}$ is $C := \prod_{j \in J \setminus \{k\}} a_j \times \prod_{j \notin J \cup \{r\}} (1 - a_j)$, so \begin{equation*} \a^{J} - \a^{J^\prime} = C \big((1 - a_{r}) a_k - a_{r}(1- a_k) \big) = C (a_k - a_{r}) \ge 0, \end{equation*} since $a_k \ge a_{r}$ because $k > r$. Thus \begin{equation*} J \succeq J^\prime. \end{equation*} Now, if $J^\prime=[\ell]$, we are done, else, repeating this procedure, we get a finite sequence of subsets, each of cardinality $\ell$ which terminates at $J^{\prime \dots \prime}=[\ell]$ after $|J \setminus [\ell]|$ replacements (since each iteration replaces exactly one element of $J \setminus [\ell]$ with one element of $[\ell] \setminus J$). \iffalse Now $|J| = |J^\prime|$ is invariant. However, since $r \le \ell$ is the smallest index not in $J$, so $\initialSubset{r - 1} \subset J$ but $\initialSubset{r} \subsetneqq J$. But $r \in J^\prime$ and $k > r$, so $\initialSubset{r} \subset J^\prime$. What this means is that the initial segment of indices has lengthened as we replace $J$ with $J^\prime$. Formally, $\max\{k = 1, \dots, n:\, \initialSubset{k} \subset J\} < \max\{k = 1, \dots, n:\, \initialSubset{k} \subset J^\prime\}$. Repeating this procedure, we get a sequence of subsets, all of cardinality $\ell$: \begin{equation*} J \succeq J^\prime \succeq \cdots \succeq \initialSubset{\ell}. \end{equation*} The reason why this sequence terminates is because the length of the initial segment is strictly increasing, but the total cardinality is (bounded by) $\ell$. Thus the last subset is $\initialSubset{\ell}$, the unique subset of cardinality $\ell$, whose initial segment is equal to itself. \fi \end{proof} \noindent We next define two integer invariants $p$ and $m$ of the ordered sequence of marginal probabilities. These invariants are used to formulate the sharp lower bound on $\a^J$ for odd $|J|$ and that for even $|J|$ respectively. Lemma \ref{lem2} is used to obtain these bounds. Associate to $a_1 \le \cdots \le a_n$ an integer $p \in \{0, 1, \dots, \lfloor (n - 1)/2 \rfloor\}$ defined as: \begin{multline} \label{eq: defineInvariantForInf} p \mbox{ is the largest integer such that } \\ a_2 + a_3 \le \cdots \le a_{2 p} + a_{2 p + 1} \le 1. \end{multline} \begin{lemma} \label{lem: minimizeAtomicProbabilityForLowerBound} If $|J|$ is odd, then: \begin{equation} \a^J \ge \a^{\initialSubset{2p + 1}}, \end{equation} where $p$ is defined in \eqref{eq: defineInvariantForInf}. \end{lemma} \begin{proof} Say $|J| = 2q + 1$. First use Lemma \ref{lem2} to get $\a^J \ge \a^{\initialSubset{2q + 1}}$. If $q = p$, we are done. Otherwise, either $q < p$ or $q > p$. \paragraph{Case 1: Suppose $q < p$} Then, we compare $\a^{\initialSubset{2q + 1}}$ with $\a^{\initialSubset{2q + 3}}$, which have $C := a_1 \cdots a_{2q + 1} (1 - a_{2q + 4}) \cdots (1 - a_n)$ as a common factor, hence \begin{align*} \a^{\initialSubset{2q + 1}} - \a^{\initialSubset{2q + 3}} &= C\big( (1 - a_{2q + 3})(1 - a_{2q + 2}) - a_{2q + 3} a_{2q + 2} \big) \\ &= C\big( 1 - a_{2q + 2} - a_{2q + 3} \big)\\ &\ge 0, \end{align*} since $a_{2q + 2} + a_{2q + 3} \le a_{2p } + a_{2p + 1} \le 1$. Thus $\a^{\initialSubset{2q + 1}} \ge \a^{\initialSubset{2q + 3}}$. Repeating this process, we get \begin{equation*} \a^{\initialSubset{2q + 1}} \ge \a^{\initialSubset{2q + 3}} \ge \cdots \ge \a^{\initialSubset{2p + 1}}. \end{equation*} \paragraph{Case 2: Suppose $q > p$} One can similarly show that $\a^{\initialSubset{2q + 1}} \ge a^{\initialSubset{2q - 1}}$ using $a_{2q} + a_{2q + 1} > 1$ because $p$ is the largest integer such that $a_{2p} + a_{2p + 1} \le 1$. Repeat the process to get $\a^{\initialSubset{2q + 1}} \ge \a^{\initialSubset{2q - 1}} \ge \cdots \ge \a^{\initialSubset{2p + 1}}$. \end{proof} \noindent Associate also to $a_1 \le \cdots \le a_n$ an integer $m \in \{0, 1, \dots, \lfloor n/2\rfloor\}$ defined as: \begin{multline} \label{eq: defineInvariantForSup} m \text{ is the largest integer such that } \\ a_1 + a_2 \le \cdots \le a_{2m - 1} + a_{2m} \le 1. \end{multline} The proof of the next lemma is similar to the previous lemma and we omit it. \begin{lemma} \label{lem: minimizeAtomicProbabilityForUpperBound} If $|J|$ is even, then: \begin{equation} \a^J \ge \a^{\initialSubset{2m}}, \end{equation} where $m$ is defined in \eqref{eq: defineInvariantForSup}. \end{lemma} \noindent This brings us to the following theorem. \begin{theorem} \label{thm: characterisationOfProbabilityMeasures} Let $p$ and $m$ be defined as in \eqref{eq: defineInvariantForInf} and \eqref{eq: defineInvariantForSup} respectively where Assumption \ref{ass:order} holds. Then every probability measure $P$ on $(\Omega, \Sigma)$ with $P(A_j) = a_j$ for all $j \in [n]$ and with respect to which $A_1, \dots, A_n$ are $(n - 1)$-wise independent has the form: $$ P(\A^J) = \a^J + (-1)^{|J|} s,\quad \mbox{for all} \;J \subseteq [n],$$ where $s$ is a scalar parameter satisfying: \begin{equation} \label{eq: rangeOfFreeParameterToGetPositivity} -\prod_{i=1}^{2m}a_i\prod_{i=2m+1}^{n}(1-a_i) \leq s \leq \prod_{i=1}^{2p+1}a_i\prod_{i=2p+2}^{n}(1-a_i). \end{equation} \end{theorem} \begin{proof} By Corollary \ref{cor: characterisationOfUnsignedMeasures}, the conditions that $P(A_j) = a_j$ for all $j \in [n]$ and $A_1, \dots, A_n$ are $(n - 1)$-wise independent with respect to $P$ entails that $ P(\A^J) = \a^J + (-1)^{|J|} s $ for $J \subseteq [n]$ for some $s \in \mathbb{R}$. In order for $P$ to be a valid probability measure, $s$ has to satisfy \eqref{eq: solveSystemForPositivityOfMeasure}. This gives: \begin{align} s &\in [-\min_{J \subseteq [n]: |J| \text{ is even}} \a^J , \min_{J \subseteq [n]: |J| \text{ is odd}} \a^J] \notag\\ &= [-\a^{\initialSubset{2m}}, \a^{\initialSubset{2p + 1}}],\notag \end{align} where the equality follows from Lemma \ref{lem: minimizeAtomicProbabilityForLowerBound} and Lemma \ref{lem: minimizeAtomicProbabilityForUpperBound}. \end{proof} \noindent Previous results in the literature are limited to the construction of specific counterexamples showing that $(n - 1)$-wise independence does not imply the mutual independence of $n$ events (see \cite{Wang,wangstoy}). Theorem \ref{thm: characterisationOfProbabilityMeasures} comprehensively characterizes all such counterexamples. Note that $s = 0$ corresponds to mutual independence. \begin{remark} Let us revisit the constructions given in Example \ref{ex:wang} in view of Theorem \ref{thm: characterisationOfProbabilityMeasures}. The probability measure where $A_n$ occurs given that an even number of events in $A_1, \dots, A_{n - 1}$ occur is given by $$ P(\A^J) = \frac{1}{2^n} - \frac{(-1)^{|J|}}{2^n}, \quad \text{for all } J \subseteq [n].$$ Since $s = - 1/2^n \in [-1/2^n, 1/2^n]$, it follows from Theorem \ref{thm: characterisationOfProbabilityMeasures} that the events are $(n - 1)$-wise independent. Note that the events are not mutually independent since $s \neq 0$. The other construction where $A_n$ occurs given that an odd number of events in $A_1, \dots, A_{n - 1}$ occur is the case of $s = 1/2^n$. \end{remark} \begin{proposition} \label{prop1} Either $m = p$ or $m = p + 1$. \end{proposition} \begin{proof} Let $k$ be the largest integer such that $a_i + a_{i + 1} \le 1$ for all $i \in [k]$. If $k$ is odd, then $k = 2m - 1$ and $k - 1 = 2 p$, hence $m = (k + 1)/2 = p + 1$. On the other hand, if $k$ is even, then $k = 2 p $ and $k - 1 = 2 m - 1$, hence $m = k/2 = p$. \end{proof} \section{Probability bounds on at least $k$ events occurring} \label{sec:atleastksharpbounds} \noindent In this section, we derive sharp bounds on the probability that at least $k$ out of $n$ events that are $(n - 1)$-wise independent occur. Bounds of this type under differing assumptions on the dependence structure of the random events have been studied (see \cite{rusch,boros1989}). Here we provide results for $(n-1)$-wise independence. From Theorem \ref{thm: characterisationOfProbabilityMeasures}, we can represent all such probabilities by: \begin{align*} P_{s}(n,k,\a) &:= \sum_{q \ge k} \sum_{J \subseteq [n]: \, |J|=q} P(\A^J)\\ & = \sum_{q \ge k}^n \sum_{J \subseteq [n]: \, |J|=q} \left(\a^J + (-1)^{|J|} s\right), \end{align*} where $s \in [- \a^{\initialSubset{2m}}, \a^{\initialSubset{2p + 1}}]$ . Then $P_0(n,k,\a)$ is the probability of occurrence of at least $k$ out of $n$ mutually independent events with the given marginal probabilities $\a$. We show that $P_{s}(n,k,\a)$ is linear in $s$. Recall the binomial coefficient given by $\binom{z}{m} = z(z - 1) \cdots (z - m + 1)/m!$ for integers $z \ge 0$ and $n$. \begin{lemma}\label{lem:atleastkevents} For any integer $k \ge 0$, \begin{align}\label{eq:atleastkintermsofalln} P_{s}(n,k,\a) = P_0(n,k,\a) +(-1)^k\dbinom{n-1}{k-1} s, \end{align} where $s \in [- \a^{\initialSubset{2m}}, \a^{\initialSubset{2p + 1}}]$. \end{lemma} \begin{proof} We have: \begin{align*} P_{s}(n,k,\a) &= P_0(n, k, \a) + s \sum_{q \ge k} (-1)^q \sum_{J \subseteq [n] : \, |J| = q} 1 \\ &= P_0(n, k, \a) + s \sum_{q \ge k} (-1)^q \binom{n}{q} \end{align*} The result then follows from the combinatorial identity $\sum_{q \ge k} (-1)^{q} \binom{n}{q} = (-1)^k\binom{n-1}{k-1} $. \end{proof} We next derive sharp upper and lower bounds on the probability that at least $k$ out of $n$ events occur under $(n-1)$-wise independence. \begin{theorem}\label{thm:atleastksharpbounds} Let $p$ and $m$ be defined as in \eqref{eq: defineInvariantForInf} and \eqref{eq: defineInvariantForSup} respectively where Assumption \ref{ass:order} holds. Then every probability measure $P$ on $(\Omega, \Sigma)$ with $P(A_j) = a_j$ for all $j \in [n]$ and with respect to which $A_1, \dots, A_n$ are $(n - 1)$-wise independent satisfies: \begin{enumerate}[label=\roman*),wide=0pt] \item For $k$ odd: \begin{align} P(n,k,\a) &\ge P_0(n,k,\a) - \binom{n-1}{k-1}\prod_{i=1}^{2p+1} a_i \prod_{i=2p+2}^{n} (1-a_i), \label{eq: lowerBoundForatleastkProbability2} \\ P(n,k,\a) &\le P_0(n,k,\a) + \binom{n-1}{k-1}\prod_{i=1}^{2m} a_i \prod_{i=2m+1}^{n} (1-a_i), \label{eq: upperBoundForatleastkProbability2} \end{align} \item For $k$ even: \begin{align} P(n,k,\a) &\ge P_0(n,k,\a) - \dbinom{n-1}{k-1}\prod_{i=1}^{2m} a_i \prod_{i=2m+1}^{n} (1-a_i), \label{eq: lowerBoundForatleastkProbability1} \\ P(n,k,\a) &\le P_0(n,k,\a) + \dbinom{n-1}{k-1}\prod_{i=1}^{2p+1} a_i \prod_{i=2p+2}^{n} (1-a_i). \label{eq: upperBoundForatleastkProbability1} \end{align} \end{enumerate} Moreover, all the bounds are sharp. The lower bound for odd $k$ in \eqref{eq: lowerBoundForatleastkProbability2} and the upper bound for even $k$ in \eqref{eq: upperBoundForatleastkProbability1} is uniquely achieved with $P = P_{\a^{\initialSubset{2p + 1}}}$. The upper bound for odd $k$ is \eqref{eq: upperBoundForatleastkProbability2} and the lower bound for even $k$ in \eqref{eq: lowerBoundForatleastkProbability1} is uniquely achieved with $P = P_{-\a^{\initialSubset{2m}}}$. \end{theorem} \begin{proof} The result is obtained from Lemma \ref{lem:atleastkevents} and optimally selecting $s$ in $[-\a^{\initialSubset{2m}},\a^{\initialSubset{2p + 1}}]$ from Theorem \ref{thm: characterisationOfProbabilityMeasures}. \end{proof} \noindent Probability bounds on the occurrence of at least $k$ out of $n$ events that are $\ell$-wise independent (i.e. every $\ell$ out of the $n$ events are mutually independent) have been studied for particular values of $\ell$. The case of $\ell = 1$ represents arbitrary dependence among the random events, for which the sharp upper bound is derived for $k = 1$ in \cite{boole} and for general $k$ in \cite{Ruger}. At the other extreme is mutual independence ($\ell = n$), where the said probability is unique. For $\ell = 2$, the sharp upper bound on the probability of the union ($k = 1$) of pairwise independent random events has been recently derived in \cite{ramanatarajan2021pairwise} and new bounds that are not necessarily sharp have been proposed for $k \ge 2$. Further, to the best of our knowledge, sharp bounds for other values of $\ell \in [3,n-1]$ have not been identified in the literature. Our results contribute to this line of work by finding sharp bounds for $l =n-1$. We next demonstrate, as an immediate implication, the computability of the sharp bounds in Theorem \ref{thm:atleastksharpbounds}. \begin{theorem} \label{thm:complexity} The sharp upper and lower bounds in Theorem \ref{thm:atleastksharpbounds} are computable in polynomial time. \end{theorem} \begin{proof} The value of $P_0(n,k,\a) $ is computable in polynomial time using dynamic programming. To see this, let $P_0(r,t,\a)$ denote the probability that at least $t$ events occur out of the first $r$ events where $r \geq t \geq 0$. Then the probabilities satisfy the recursive formula: \begin{equation*} P_0(r,t,\a) = P_0(r-1,t-1,\a)p_r + P_0(r-1,t,\a)(1-p_r), \end{equation*} where the boundary conditions are $P_0(r,0,\a) = 1$ for $r \geq 0$ and $P_0(r,t,\a) = 0$ for $t > r$ (see \cite{hong2013poissonbinomial}). The probability $P_0(n,k,\a) $ is hence computable in $O(n^2)$ time. Since the additional term in the formulas \eqref{eq: lowerBoundForatleastkProbability2}-\eqref{eq: upperBoundForatleastkProbability1} is efficiently computable using sorting, evaluating binomial coefficients and multiplication, all the bounds are computable in polynomial time; specifically $O(n^2)$ time. \end{proof} \noindent The sharp bounds for $k = 1$ (union of events) and $k = n$ (intersection of events) are detailed next. \begin{corollary} \label{unionintersect} Let $p$ and $m$ be defined as in \eqref{eq: defineInvariantForInf} and \eqref{eq: defineInvariantForSup} respectively where Assumption \ref{ass:order} holds. Then every probability measure $P$ on $(\Omega, \Sigma)$ with $P(A_j) = a_j$ for all $j \in [n]$ and with respect to which $A_1, \dots, A_n$ are $(n - 1)$-wise independent satisfies: \begin{enumerate}[label=\roman*),wide=0pt] \item For the union of events: \begin{align} P(\bigcup_{j = 1}^n A_j) &\ge 1 - \left(\prod_{i=1}^{2p+1}(1-a_i)+\prod_{i=1}^{2p+1}a_i\right)\prod_{i=2p+2}^{n}(1 - a_i), \label{eq: lowerBoundForUnionProbability} \\ P(\bigcup_{j = 1}^n A_j) &\le 1 - \left(\prod_{i=1}^{2m}(1-a_i)-\prod_{i=1}^{2m}a_i\right)\prod_{i=2m+1}^{n}(1 - a_i), \label{eq: upperBoundForUnionProbability} \end{align} \item For the intersection of an even number of events: \begin{align} P(\bigcap_{j = 1}^n A_j) &\ge \prod_{i=1}^{2m}a_i\left(\prod_{i=2m+1}^{n}a_i-\prod_{i=2m+1}^{n}(1-a_i)\right), \label{eq: lowerBoundForIntersectProbabilityeven} \\ P(\bigcap_{j = 1}^n A_j) &\le\prod_{i=1}^{2p+1}a_i\left(\prod_{i=2p+2}^{n}a_i+\prod_{i=2p+2}^{n}(1-a_i)\right), \label{eq: upperBoundForIntersectProbabilityeven} \end{align} \item For the intersection of an odd number of events: \begin{align} P(\bigcap_{j = 1}^n A_j) &\ge \prod_{i=1}^{2p+1}a_i\left(\prod_{i=2p+2}^{n}a_i-\prod_{i=2p+2}^{n}(1-a_i)\right), \label{eq: lowerBoundForIntersectProbabilityodd} \\ P(\bigcap_{j = 1}^n A_j) &\le \prod_{i=1}^{2m}a_i\left(\prod_{i=2m+1}^{n}a_i+\prod_{i=2m+1}^{n}(1-a_i)\right)\label{eq: upperBoundForIntersectProbabilityodd}. \end{align} \end{enumerate} Each of these bounds is sharp and is achieved by a unique probability measure $P(\A^J) = \a^J + (-1)^{|J|}s$, where either $s =-\a^{\initialSubset{2m}}$ or $s =\a^{\initialSubset{2p + 1}}$. \end{corollary} \begin{proof} With $k=1$, we have $P_0(n,1,\a)=1-\prod_{i=1}^n (1-a_i)$ and the result immediately follows from \eqref{eq: lowerBoundForatleastkProbability2} and \eqref{eq: upperBoundForatleastkProbability2} in Theorem \ref{thm:atleastksharpbounds}. With $k=n$, we have $P_0(n,n,\a)=\prod_{i=1}^n a_i$ and the result immediately follows from Theorem \ref{thm:atleastksharpbounds} depending on whether $k = n$ is even or odd. \end{proof} \section{Examples}\label{sec:examples} \noindent In this section, we discuss several examples to illustrate the connection of the newly proposed bounds with existing bounds and provide numerical evidence of the quality of the bounds. \begin{example} [Bounds for $n = 3$ pairwise independent events]\label{ex:reductiontopairwise} For $n=3$ events, $(n-1)$-wise independence is pairwise independence. In this case from \eqref{eq: defineInvariantForSup} and \eqref{eq: upperBoundForUnionProbability} where $a_1 \leq a_2 \leq a_3$, we obtain the sharp upper bound on the union as: \begin{equation} \label{eq: sharpUpperBoundForThreeEvents} P(\bigcup_{j = 1}^3 A_j) \le \min\left(a_1+a_2+a_3-a_3(a_1+a_2),1\right) \end{equation} Another proof of the sharpness was given in \cite{ramanatarajan2021pairwise}. Kounias \cite{kounias} showed that every probability measure satisfies $P(\bigcup_{j = 1}^3 A_j) \le \min\left(a_1+a_2+a_3-a_{31}+a_{32}),1\right)$, where $a_{ij} := P(A_i \cap A_j)$ for $i, j \in [3]$ denotes the bivariate joint probability. Therefore, \eqref{eq: sharpUpperBoundForThreeEvents} entails that the upper bound of Kounias is achieved by some probability measure with respect to which $A_1, A_2, A_3$ are pairwise independent. For the sharp lower bound, from \eqref{eq: defineInvariantForInf} and \eqref{eq: lowerBoundForUnionProbability}, we get: \begin{multline*} P(\bigcup_{j = 1}^3 A_j) \ge \max(a_2+a_3-a_2a_3, \\ a_1+a_2+a_3-a_1a_2-a_1a_3-a_2a_3), \end{multline*} Similarly, a corresponding universal lower bound of Kounias in terms of bivariate joint probabilities is therefore achievable under pairwise independence. Likewise for the intersection of three pairwise independent events, we can verify that the sharp bounds are given as: \begin{equation*} P(\bigcap_{j = 1}^3 A_j) \le \min(a_1a_2, (1 - a_1)(1 - a_2)(1 - a_3) + a_1 a_2 a_3), \end{equation*} and \begin{equation*} P(\bigcap_{j = 1}^3 A_j) \ge \max\left(a_1(a_2+a_3-1),0\right). \end{equation*} An alternative proof of the sharpness of the lower bound is given in \cite{ramanatarajan2021pairwise}. \end{example} \begin{example} [Bonferroni bounds]\label{ex:reductiontobonferroni} Suppose the sum of the two largest marginal probabilities satisfies $a_{n-1} + a_{n} \leq 1$ and $n$ is even. Then $m=n/2$ in \eqref{eq: defineInvariantForSup}, hence from \eqref{eq: upperBoundForUnionProbability}, we get the sharp upper bound on the union: \begin{align*} P(\bigcup_{j = 1}^n A_j) &\le 1 - \prod_{i=1}^{n}(1 - a_i) + \prod_{i=1}^{n}a_i \\ &= \sum_{k = 0}^{n - 2} (-1)^k \sum_{1 \le i_0 < \cdots < i_k \le n} a_{i_0} \cdots a_{i_k}. \end{align*} Bonferroni \cite{Bonferroni} showed that every probability measure satisfies \begin{equation} P(\bigcup_{j = 1}^n A_j) \le \sum_{k = 0}^{n - 2} (-1)^k \sum_{1 \le i_0 < \cdots < i_k \le n} a_{i_0 \cdots i_k}, \end{equation} where $a_{i_0 \cdots i_k} := P(A_{i_0} \cap \cdots \cap A_{i_k})$ for $i_0, \dots, i_k \in [n]$ is the joint probability. Therefore the Bonferroni upper bound is achieved by some probability measure with respect to which $A_1, \dots, A_n$ are $(n - 1)$-wise independent, in this case. Similarly if $a_{n - 1} + a_n \le 1$ and $n$ is odd, then $p = (n - 1)/2$ in \eqref{eq: defineInvariantForInf} and thus the sharp lower bound in \eqref{eq: lowerBoundForUnionProbability} becomes: \begin{align*} P(\bigcup_{j = 1}^n A_j) &\ge 1 - \prod_{i=1}^{n}(1 - a_i) - \prod_{i=1}^{n}a_i \\ &= \sum_{k = 0}^{n - 2} (-1)^k \sum_{1 \le i_0 < \cdots < i_k \le n} a_{i_0} \cdots a_{i_k}. \end{align*} Again, a corresponding lower bound of Bonferroni in terms of joint probabilities of up to $n - 1$ events is thus achievable under $(n - 1)$-wise independence. \end{example} The next example shows the connection of the bound to the probabilistic method which has proved to be very useful tool in combinatorics (see \cite{alon}). \begin{example} [Probabilistic method]\label{ex:lovaszlocallemma} Suppose there are $n$ random ``bad'' events, each of which occurs with probability $a_j$ for $j \in [n]$. When the events are mutually independent, the probability of no bad event occurring is strictly positive when the probability of each bad event is strictly less than 1 (namely $\max_j a_j < 1$). On the other hand, if the events can be arbitrarily dependent, from Boole's union bound \cite{boole}, the sum of the probabilities must be strictly less than 1 (namely $\sum_j a_j < 1$) to guarantee the same. The Lov\'{a}sz local lemma \cite{erdos} is a powerful tool that allows one to relax the assumption of mutual independence to weak dependence while allowing for the probability of each bad event to be fairly large and still guarantee that no bad event occurs with strictly positive probability. Specifically consider a graph $G$ on $n$ nodes where each node $i \in [n]$ is associated with an event $A_i$ and $A_i$ is independent of the collection of events $\{A_j: (i,j) \notin G\}$ for each $i \in [n]$. If $G$ has maximum degree $d$ and $\max_i a_i \leq 1/4d$, then the probability of no bad event occurring satisfies (see \cite{erdos,tetali}): \begin{align} P(\bigcap_{j = 1}^n \overline{A}_j) \ge \prod_{i=1}^{n}(1-2a_i) > 0.\label{eq:lovaszlower} \end{align} Computing the tightest lower bound in terms of the dependency graph is known to be NP-complete \cite{shearer}. More generally, in \cite{shearer} it was shown that for $d \geq 2$, $\max_i a_i < (d-1)^{d-1}/d^d$ and for $d = 1$, $\max_i a_i < 1/2$ guarantees that there is a strictly positive probability that no bad event occurs. For the specific case of $d=1$, we can compare our results with the lower bound as shown next (although the Lov\'{a}sz local lemma holds more generally for lesser independence with $d \ge2 $). When the events are $(n-1)$-wise independent, using \eqref{eq: upperBoundForUnionProbability}, the probability that none of the events occur is strictly positive if $a_n< 1$ and $a_1+a_2 < 1$. Indeed, then $(1 - a_1)(1 - a_2) > a_1 a_2$ and $(1 - a_{2k - 1})(1 - a_{2k}) \ge a_{2k - 1} a_{2k}$ for all $k \in \{2, \dots, m\}$. Hence $\prod_{i=1}^{2m}(1-a_i) = \prod_{k = 1}^{m}((1 - a_{2k - 1})(1 - a_{2k})) > \prod_{k = 1}^{m} (a_{2k - 1} a_{2k}) = \prod_{i=1}^{2m}a_i$ and from \eqref{eq: upperBoundForUnionProbability}: $$ P(\bigcap_{j = 1}^n \overline{A}_j) \ge \left(\prod_{i=1}^{2m}(1-a_i)-\prod_{i=1}^{2m}a_i\right)\prod_{i=2m+1}^{n}(1 - a_i) > 0. $$ \noindent When all the marginal probabilities $a_1=\ldots=a_n= a$, are identical, the condition $a_1 + a_2 < 1$ gives $a < 1/2$ which exactly corresponds to the condition identified in \cite{spencer,shearer} for $d = 1$. It is easy to verify that the lower bound on the probability of no bad event occurring in this case is given by $(1-a)^n-a^n$ for $n$ even and $(1-a)^n-a^{n-1}(1-a)$ for $n$ odd which is the sharp lower bound instance wise. In comparison, the lower bound identified above in \eqref{eq:lovaszlower} is $(1-2a)^n$. For example with $n = 6$ and $a = 0.1$, the sharp lower bound is $0.53144$ while the weaker lower bound is $0.262144$. In fact for $a = 1/2$, the first construction in Example \ref{ex:wang} has a zero probability that no bad event occurs since $A_n$ must occur when none of the events in $\{A_1,A_2,\ldots A_{n-1}\}$ occur. \end{example} \noindent We next provide a numerical example to illustrate the performance of the bounds in Theorem \ref{thm:atleastksharpbounds} and compare it with an existing bound. Specifically tail probability bounds on the sum of two random variables given their marginal distribution functions were derived by Makarov in \cite{makarov1982}. We adopt these closed-form bounds also known as ``standard" bounds in our context as follows. Given that $n$ random events $A_1,\ldots A_n$ with respective marginal probabilities $a_1 \le \cdots \le a_n$ are $(n - 1)$-wise independent, define two random variables as follows: $Y_1=\sum_{i=1}^{n-1} \mathbb{1}_{A_i},\;Y_2=\mathbb{1}_{A_n}$ where $ \mathbb{1}_{A}$ is the indicator function of event $A$ occurring. Here $ Y_1 \sim \operatorname{PoissonBinomial}(n-1,a_1,a_2,\ldots a_{n-1})$ is an integer random variable taking values in $[0,n-1]$ while $Y_2 \sim \operatorname{Bernoulli}(a_n)$. Let $F_1$ and $F_2$ be the resepective distribution functions of $Y_1$ and $Y_2$. Then the Makarov upper bound for the probability that the sum of $Y_1$ and $Y_2$ is at least an integer $k \in [n]$ is given from \cite{makarov1982,rusch} as follows: \begin{equation}\label{eq: makarovupper} \begin{array}{rll} P(Y_1+Y_2\ge k)& \leq & \min(2-(F_{1}\vee F_{2})^{-}(k),1), \end{array} \end{equation} where $(F_{1}\vee F_{2})^{-}(k)=\underset{u \in \mathbb{R}}{\max}(F_{1}(k-u)^{-} +F_{2}(u))$ is the left continuous version of the supremum convolution $F_{1}\vee F_{2}$. Since $Y_2$ is a Bernoulli random variable, it is sufficient to maximize over $u \in \{0,1\}$ and thus we have : \begin{align*} & (F_{1}\vee F_{2})^{-}(k) & =&\max(F_{1}(k-1) +a_n,\;F_{1}(k-2) +1). \end{align*} The Makarov lower bound can be similarly derived as \begin{equation}\label{eq: makarovlower} \begin{array}{rll} P(Y_1+Y_2\ge k) \ge\max(1-\min(F_{1}(k),F_{1}(k-1) +a_n),0). \end{array} \end{equation} We next illustrate through a numerical example that the Makarov bound is not sharp in general under $(n-1)$-wise independence since we lose out on using additional independence information available in our context. For example, our bounds assume that any $n-2$ events from the first $n-1$ events $A_1,\ldots,A_{n-1}$ along with the last event $A_n$ are mutually independent while the Makarov bounds do not assume so. \begin{example} [Numerical example] \label{ex:comparisonmakarov} Here we compute the exact probability for $n = 8$ with identical marginal probabilities $a_i = a \in \{0.1,0.2,0.3,0.4,0.5\}$ for different values of $k$ assuming mutual independence. In addition we compute the sharp lower and upper bounds with $7$-wise independence from Theorem \ref{thm:atleastksharpbounds} (here $p = 3$ and $m = 4$ for all considered values of $a$). We also provide the Makarov lower and upper bounds from (\ref{eq: makarovlower}) and (\ref{eq: makarovupper}) to highlight that if more information is known on the independence of the random variables, we can exploit it tightening the bounds. \begin{table}[H] \footnotesize \caption{$k=1$ to $k=4$ - For each value of $a$, the first row provides the Makarov lower bound from \eqref{eq: makarovlower}, the second row provides the sharp lower bound with $7$-wise independence, the third row provides the exact value with $8$ mutually independent events, the fourth row provides the sharp upper bound with $7$-wise independence and the fifth row provides the Makarov upper bound from \eqref{eq: makarovupper}} \label{tab:heteroall3vstight1} \begin{center} \scriptsize{\begin{tabular} {|l|c|c|l|l|l|l|l|l|l|} \hline \mbox{a} & $k = 1$ & $k = 2$ & $k = 3$ & $k = 4$ \\ \hline 0.1 & 4.6953e-01 & 8.6895e-02 & 5.0243e-03 & 4.3165e-04\\ 0.1 & 5.6953e-01 & 1.8690e-01 & 3.8090e-02 & 5.0240e-03\\ 0.1 & 5.6953e-01 & 1.8690e-01 & 3.8092e-02 & 5.0244e-03\\ 0.1 & 5.6953e-01 & 1.8690e-01 & 3.8092e-02 & 5.0275e-03\\ 0.1 & 1.0000e+00 & 5.6953e-01 & 1.8690e-01 & 3.8092e-02\\ \hline 0.2 & 6.3223e-01 & 2.9668e-01 & 5.6282e-02 & 1.0406e-02\\ 0.2 & 8.3222e-01 & 4.9667e-01 & 2.0287e-01 & 5.6192e-02\\ 0.2 & 8.3223e-01 & 4.9668e-01 & 2.0308e-01 & 5.6282e-02\\ 0.2 & 8.3223e-01 & 4.9676e-01 & 2.0314e-01 & 5.6640e-02\\ 0.2 & 1.0000e+00 & 8.3223e-01 & 4.9668e-01 & 2.0308e-01\\ \hline 0.3 & 7.4470e-01 & 4.4823e-01 & 1.9410e-01 & 5.7968e-02\\ 0.3 & 9.4220e-01 & 7.4424e-01 & 4.4501e-01 & 1.9181e-01\\ 0.3 & 9.4235e-01 & 7.4470e-01 & 4.4823e-01 & 1.9410e-01\\ 0.3 & 9.4242e-01 & 7.4577e-01 & 4.4960e-01 & 1.9946e-01\\ 0.3 & 1.0000e+00 & 9.4235e-01 & 7.4470e-01 & 4.4823e-01\\ \hline 0.4 & 8.9362e-01 & 6.8461e-01 & 4.0591e-01 & 1.7367e-01\\ 0.4 & 9.8222e-01 & 8.8904e-01 & 6.6396e-01 & 3.8298e-01\\ 0.4 & 9.8320e-01 & 8.9362e-01 & 6.8461e-01 & 4.0591e-01\\ 0.4 & 9.8386e-01 & 9.0051e-01 & 6.9837e-01 & 4.4032e-01\\ 0.4 & 1.0000e+00& 9.8320e-01 & 8.9362e-01 & 6.8461e-01\\ \hline 0.5 & 9.6484e-01 & 8.5547e-01 & 6.3672e-01 & 3.6328e-01\\ 0.5 & 9.9219e-01 & 9.3750e-01 & 7.7344e-01 & 5.0000e-01\\ 0.5 & 9.9609e-01 & 9.6484e-01 & 8.5547e-01 & 6.3672e-01\\ 0.5 & 1.0000e+00 & 9.9219e-01 & 9.3750e-01 & 7.7344e-01\\ 0.5 & 1.0000e+00 & 9.9610e-01 & 9.6484e-01 & 8.5547e-01\\ \hline \end{tabular}} \end{center} \end{table} \begin{table}[H] \footnotesize \caption{$k=5$ to $k=8$} \label{tab:heteroall3vstight2} \begin{center} \scriptsize{\begin{tabular}[t] {|l|c|c|l|l|l|l|l|l|l|} \hline \mbox{a} & $k = 5$ & $k = 6$ & $k = 7$ & $k = 8$ \\ \hline 0.1 & 2.3410e-05 & 7.3000e-07 & 9.9999e-09 & 0.0000e+00\\ 0.1 & 4.2850e-04 & 2.3200e-05 & 1.0000e-07 & 0.0000e+00\\ 0.1 & 4.3165e-04 & 2.3410e-05 & 7.3000e-07 & 1.0000e-08\\ 0.1 & 4.3200e-04 & 2.5300e-05 & 8.0000e-07 & 1.0000e-07\\ 0.1 & 5.0244e-03 & 4.3165e-04 & 2.3410e-05 & 7.3000e-07\\ \hline 0.2 & 1.2314e-03 & 8.4480e-05 & 2.5600e-06 & 0.0000e+00\\ 0.2 & 1.0048e-02 & 1.1776e-03 & 1.2800e-05 & 0.0000e+00\\ 0.2 & 1.0406e-02 & 1.2314e-03 & 8.4480e-05 & 2.5600e-06\\ 0.2 & 1.0496e-02 & 1.4464e-03 & 1.0240e-04 & 1.2800e-05\\ 0.2 & 5.6282e-02 & 1.0406e-02 & 1.2314e-03 & 8.4480e-05\\ \hline 0.3 & 1.1292e-02 & 1.2903e-03 & 6.5610e-05 & 0.0000e+00\\ 0.3 & 5.2610e-02 & 9.9144e-03 & 2.1870e-04 & 0.0000e+00\\ 0.3 & 5.7968e-02 & 1.1292e-02 & 1.2903e-03 & 6.5610e-05\\ 0.3 & 6.0264e-02 & 1.4507e-02 & 1.7496e-03 & 2.1870e-04\\ 0.3 & 1.9410e-01 & 5.7968e-02 & 1.1292e-02 & 1.2903e-03\\ \hline 0.4 & 4.9807e-02 & 8.5200e-03 & 6.5536e-04 & 0.0000e+00\\ 0.4 & 1.3926e-01 & 3.6045e-02 & 1.6384e-03 & 0.0000e+00\\ 0.4 & 1.7367e-01 & 4.9807e-02 & 8.5197e-03 & 6.5536e-04\\ 0.4 & 1.9661e-01 & 7.0451e-02 & 1.3107e-02 & 1.6384e-03\\ 0.4 & 4.0591e-01 & 1.7367e-01 & 4.9807e-02 & 8.5197e-03\\ \hline 0.5 & 1.4453e-01 & 3.5156e-02 & 3.9063e-03 & 0.0000e+00\\ 0.5 & 2.2656e-01 & 6.2500e-02 & 7.8125e-03 & 0.0000e+00\\ 0.5 & 3.6328e-01 & 1.4453e-01 & 3.5156e-02 & 3.9063e-03\\ 0.5 & 5.0000e-01 & 2.2656e-01 & 6.2500e-02 & 7.8125e-03\\ 0.5 & 6.3672e-01 & 3.6328e-01 & 1.4453e-01 & 3.5156e-02\\ \hline \end{tabular}} \end{center} \end{table} \noindent As it can be observed, the sharp bounds with $(n-1)$-wise independence clearly improve upon the Makarov bounds, especially as $k$ increases (for the same $a$) and $a$ decreases (for the same $k$), where the bounds can be a couple or more magnitude of orders apart. In other words, the sharp bounds especially provide value in the regime where the right tail probabilities are more constrained i.e. large $k$ and small $a$. Such bounds are useful in providing robust estimates of the probabilities when the assumption of mutual independence breaks down. \end{example} \subsection*{Acknowledgements} \noindent The research of the first and third authors was partly supported by MOE Academic Research Fund Tier 2 grant T2MOE1906, ``Enhancing Robustness 770 of Networks to Dependence via Optimization''. The authors would like to thank the Area Editor Henry Lam, the Associate Editor and the anonymous reviewer for valuable comments.

https://arxiv.org/abs/1801.04483

Waring's Theorem for Binary Powers

A natural number is a binary $k$'th power if its binary representation consists of $k$ consecutive identical blocks. We prove an analogue of Waring's theorem for sums of binary $k$'th powers. More precisely, we show that for each integer $k \geq 2$, there exists a positive integer $W(k)$ such that every sufficiently large multiple of $E_k := \gcd(2^k - 1, k)$ is the sum of at most $W(k)$ binary $k$'th powers. (The hypothesis of being a multiple of $E_k$ cannot be omitted, since we show that the $\gcd$ of the binary $k$'th powers is $E_k$.) Also, we explain how our results can be extended to arbitrary integer bases $b > 2$.

\section{Introduction} Let ${\mathbb{N}} = \{ 0,1,2,\ldots \}$ be the natural numbers and let $S \subseteq {\mathbb{N}}$. The principal problem of additive number theory is to determine whether every integer $N$ (resp., every sufficiently large integer $N$) can be represented as the sum of some {\it constant\/} number of elements of $S$, not necessarily distinct, where the constant does not depend on $N$. For a superb introduction to this topic, see \cite{N1}. Probably the most famous theorem of additive number theory is Lagrange's theorem from 1770: every natural number is the sum of four squares \cite{L}. Waring's problem (see, e.g., \cite{Sm,VW}), first stated by Edward Waring in 1770, is to determine $g(k)$ such that every natural number is the sum of $g(k)$ $k$'th powers. (A priori, it is not even clear that $g(k) < \infty$, but this was proven by Hilbert in 1909.) From Lagrange's theorem we know that $g(2) = 4$. For other results concerning sums of squares, see, e.g., \cite{G,MW}. If every natural number is the sum of $k$ elements of $S$, we say that $S$ forms a {\it basis} of order $k$. If every sufficiently large natural number is the sum of $k$ elements of $S$, we say that $S$ forms an {\it asymptotic basis} of order $k$. In this paper, we consider a variation on Waring's theorem, where the ordinary notion of integer power is replaced by a related notion inspired from formal language theory. Our main result is Theorem~\ref{main} below. We say that a natural number $N$ is a {\it base-$b$ $k$'th power} if its base-$b$ representation consists of $k$ consecutive identical blocks. For example, 3549 in base $2$ is $$ 1101 \, 1101 \, 1101 ,$$ so 3549 is a base-2 (or binary) cube. Throughout this paper, we consider only the {\it canonical} base-$b$ expansions (that is, those without leading zeros). The binary squares $$ 0,3,10,15,36,45,54,63,136,153,170,187,204,221,238,255,528,561,594,627,\ldots$$ form sequence \seqnum{A020330} in Sloane's {\it On-Line Encyclopedia of Integer Sequences} \cite{Sl}. The binary cubes $$ 0,7,42,63,292,365,438,511,2184,2457,2730,3003,3276,3549,3822,4095,16912,\ldots$$ form sequence \seqnum{A297405}. Notice that a number $N>0$ is a base-$b$ $k$'th power if and only if we can write $N = a \cdot c_k^b (n)$, where $$c_k^b (n) := \frac{b^{kn}-1}{b^n - 1} = 1 + b^n + \cdots + b^{(k-1)n}$$ for some $n \geq 1$ such that $b^{n-1} \leq a < b^n$. (The latter condition is needed to ensure that the base-$b$ $k$'th power is formed by the concatenation of blocks that begin with a nonzero digit.) Such a number consists of $k$ consecutive blocks of digits, each of length $n$. For example, $3549 = 13 \cdot c_3^2 (4)$. We define $${\mathcal{S}}_k^b := \left\{ n \geq 0 \ : \ n \text{ is a base-$b$ $k$'th power} \right\} = \left\{ a \cdot c_k^b (n) \ : \ n \geq 1, \ b^{n-1} \leq a < b^n \right\}.$$ The set ${\mathcal{S}}_k^b$ is an interesting and natural set to study because its counting function is $\Omega(N^{1/k})$, just like the ordinary $k$'th powers. It has also appeared in a number of recent papers (e.g., \cite{BLS}). However, there are two significant differences between the ordinary $k$'th powers and the base-$b$ $k$'th powers. The first difference is that $1$ is not a base-$b$ $k$'th power for $k > 1$. Thus, the base-$b$ $k$'th powers cannot, in general, form a basis of finite order, but only an asymptotic basis. A more significant difference is that the gcd of the ordinary $k$'th powers is always equal to $1$, while the gcd of the base-$b$ $k$'th powers may, in some cases, be greater than one. This is quantified in Section~\ref{gcd2}. Thus, it is not reasonable to expect that every sufficiently large natural number can be the sum of a fixed number of base-$b$ $k$'th powers; only those that are also a multiple of the $\gcd$ can be so represented. \section{The greatest common divisor of ${\mathcal{S}}_k^b$} \label{gcd2} \begin{theorem} For $k \geq 1$ define \begin{align*} A_k &= \gcd( {\mathcal{S}}_k^b) ,\\ B_k &= \gcd( c_k^b (1), c_k^b (2), \ldots) , \\ C_k &= \gcd( c_k^b (1), c_k^b (2), \ldots, c_k^b (k)) , \\ D_k &= \gcd( c_k^b (1), c_k^b (k)) , \\ E_k &= \gcd\!\left( \frac{b^k - 1}{b-1}, k\right) . \end{align*} Then $A_k = B_k = C_k = D_k = E_k$. \label{sanna} \end{theorem} \begin{proof} \noindent $A_k = B_k$: If $d$ divides $B_k$, then it clearly also divides all numbers of the form $a \cdot c_k^b (n)$ with $b^{n-1} \leq a < b^n$ and hence $A_k$. On the other hand if $d$ divides $A_k$, then it divides $c_k^b (1)$. Furthermore, $d$ divides $b^{n-1} \cdot c_k^b (n)$ and $(b^{n-1} + 1) c_k^b (n)$ (both of which are members of ${\mathcal{S}}_k^b$ provided $n \geq 2$). So it must divide their difference, which is just $c_k^b (n)$. So $d$ divides $B_k$. \medskip \noindent $B_k = C_k$: Note that $d$ divides $B_k$ if and only if it divides $c_k^b (1)$ and also $c_k^b (n) \bmod c_k^b (1)$ for all $n \geq 1$. Now it is well known that, for $b \geq 2$ and integers $n, k \geq 1$, we have $$b^n \equiv \modd{b^{n \bmod k}} {b^k - 1}.$$ Hence \begin{align*} c_k^b (n) &= 1 + b^n + \cdots + b^{(k-1)n} \equiv \modd{1 + b^{n \bmod k} + \cdots + b^{(k-1)n \bmod k}} {b^k - 1} \\ &\equiv \modd{1 + b^a + \cdots + b^{(k-1)a}} {b^k-1} \\ & \equiv \modd{1 + b^a + \cdots + b^{(k-1)a}} {c_k^b(1)} \\ & \equiv \modd{c_k^b (a)} {c_k^b (1)}, \end{align*} where $a = n \bmod k$. Thus any divisor of $C_k$ is also a divisor of $B_k$. The converse is clear. \medskip \noindent $D_k = E_k$: It suffices to observe that \begin{align*} c_k^b (k) &= 1 + b^k + \cdots + b^{(k-1)k} \\ &\equiv \modd{\overbrace{1 + 1 + \cdots + 1}^k} {b^k - 1} \\ &\equiv \modd{k} {b^k - 1} \\ & \equiv \modd{k} {\frac{b^k-1}{b-1}} \\ & \equiv \modd{k} {c_k^b (1)}. \end{align*} \medskip \noindent $B_k = E_k$: Every divisor of $B_k$ clearly divides $D_k$, and above we saw $D_k = E_k$. We now show that every prime divisor of $E_k$ divides $B_k$ to at least the same order, thus showing that every divisor of $E_k$ divides $B_k$. We need the following classic lemma, sometimes called the ``lifting-the-exponent'' or LTE lemma \cite{C}: \begin{lemma}\label{lem:LTE} If $p$ is a prime number and $c \neq 1$ is an integer such that $p \mid c - 1$, then \begin{equation*} \nu_p\!\left(\frac{c^n - 1}{c - 1}\right) \geq \nu_p(n) , \end{equation*} for all positive integers $n$, where $\nu_p (n)$ is the $p$-adic valuation of $n$ (the exponent of the highest power of $p$ dividing $n$). \end{lemma} Fix an integer $\ell \geq 1$ and let $p$ be a prime factor of $E_k$. On the one hand, if $p \mid b^\ell - 1$, then by Lemma~\ref{lem:LTE} we get that \begin{equation*} \nu_p\!\left(c_k^b(\ell)\right)=\nu_p\!\left(\frac{b^{k\ell} - 1}{b^\ell - 1}\right) \geq \nu_p(k) \geq \nu_p(E_k) , \end{equation*} since $E_k \mid k$. Hence $p^{\nu_p(E_k)} \mid c_k^b(\ell)$. On the other hand, if $p \nmid b^\ell - 1$, then $p^{\nu_p(E_k)}$ divides $c_k^b(\ell) = \frac{b^{k\ell} - 1}{b^\ell - 1}$ simply because $p^{\nu_p(E_k)}$ divides the numerator but does not divide the denominator. In both cases, we have that $p^{\nu_p(E_k)} \mid c_k^b(\ell)$, and since this is true for all prime divisors of $E_k$, we get that $E_k \mid c_k^b(\ell)$, as desired. \end{proof} \begin{remark} For $b = 2$, the sequence $E_k$ is sequence \seqnum{A014491} in Sloane's {\it Encyclopedia}. We make some additional remarks about the values of $E_k$ in Section~\ref{final}. \end{remark} In the remainder of the paper, for concreteness, we focus on the case $b = 2$. We set $c_k (n) := c_k^2 (n)$ and ${\mathcal{S}}_k := {\mathcal{S}}_k^2$. However, everything we say also applies more generally to bases $b > 2$, with one minor complication that is mentioned in Section~\ref{final}. \section{Waring's theorem for binary $k$'th powers: proof outline and tools} We now state the main result of this paper. \begin{theorem} Let $k \geq 1$ be an integer. Then there is a number $W(k) < \infty$ such that every sufficiently large multiple of $E_k = \gcd(2^k - 1,k)$ is representable as the sum of at most $W(k)$ binary $k$'th powers. \label{main} \end{theorem} \begin{remark} The fact that $W(2) \leq 4$ was proved in \cite{MNRS}. \end{remark} \begin{proof}[Proof sketch] Here is an outline of the proof. All of the mentioned constants depend only on $k$. Given a number $N$, a multiple of $E_k$, that we wish to represent as a sum of binary $k$'th powers, we first choose a suitable power of $2$, say $x = 2^n$, and think of $N$ as a degree-$k$ polynomial $p$ evaluated at $x$. For example, we can represent $N$ in base $2^n$; the ``digits'' of this representation then correspond to the coefficients of $p$. Similarly, the integers $c_k (n), c_k(n+1), \ldots, c_k (n+k-1)$ can also be viewed as polynomials in $x = 2^n$. By linear algebra, there is a unique way to rewrite $p$ as a linear combination of $c_k (n), c_k(n+1), \ldots, c_k (n+k-1)$, and this linear transformation can be represented by a matrix $M$ that depends only on $k$, and is independent of $n$. At first glance, such a linear combination would seem to provide a suitable representation of $N$ in terms of binary $k$'th powers, but there are three problems to overcome: \begin{enumerate}[(a)] \item the coefficients of $c_k (i)$, $n \leq i < n+k$, could be much too large; \item the coefficients could be too small or negative; \item the coefficients might not be integers. \end{enumerate} Issue (a) can be handled by choosing $n$ such that $2^n \approx N^{1/k}$. This guarantees that the resulting coefficients of the $c_k (n)$ are at most a constant factor larger than $2^n$. Using Lemma~\ref{split} below, the coefficients can be ``split'' into at most a constant number of coefficients lying in the desired range. Issue (b) is handled by not working with $N$, but rather with $Y := N - D$, where $D$ is a suitably chosen linear combination of $c_k (n), c_k (n+1), \ldots, c_k(n+k-1)$ with large positive integer coefficients. Any negative coefficients arising in the expression for $Y$ can now be offset by adding the large positive coefficients corresponding to $D$, giving us coefficients for the representation of $N$ that are positive and lie in a suitable range. Issue (c) is handled by rounding down the coefficients of the linear combination to the next lower integer. This gives us a representation, as a sum of binary $k$'th powers, for some smaller number $N' < N$, where the difference $N - N'$ is a sum of at most $k^2$ terms of the form $2^i/d$, where $d$ is the determinant of $M$. However, the base-$2$ representation of $1/d$ is, disregarding leading zeros, actually periodic with some period $p$. By choosing an appropriate small multiple of a binary $k$'th power corresponding to $k$ copies of this period, we can approximate each $2^i/d$, and hence $N - N'$, from below by some number $N''$ that is a sum of binary $k$'th powers. The remaining error term is $Q := N - N' - N''$, which turns out to be at most some constant depending on $k$. Since $N$ is a multiple of $E_k$ and $N'$ and $N''$ are sums of binary $k$'th powers, it follows that $Q$ is also a multiple of $E_k$. With care we can ensure that $Q$ is larger than the Frobenius number of the binary $k$'th powers, and hence $Q$ can be written as a sum of elements of ${\mathcal{S}}_k$. On the other hand, since $Q$ is a constant, at most a constant number of additional binary $k$'th powers are needed to represent it. This completes the sketch of our construction. It is carried out in more detail in the rest of the paper. \end{proof} \begin{remark} In what follows, we spend a small amount of time explaining that certain quantities are actually constants that depend only on $k$. By estimating these constants we could come up with an explicit bound on $W(k)$, but we have not done so. \end{remark} \subsection{Expressing multiples of $c_k (n)$ as a sum of binary $k$'th powers} As we have seen, a number of the form $a \cdot c_k (n)$ with $2^{n-1} \leq a < 2^n$ is a binary $k$'th power. But how about larger multiples of $c_k (n)$? The following lemma will be useful. \begin{lemma} Let $a \geq 2^{n-1}$. Then $a \cdot c_k (n)$ is the sum of at most $\lceil \frac{a}{2^n -1} \rceil$ binary $k$'th powers. \label{split} \end{lemma} \begin{proof} Clearly the claim is true for $2^{n-1} \leq a < 2^n$. Otherwise, define $b := \lceil \frac{a}{2^n -1} \rceil$ and $c := (2^n-1)b - a$, so that $0 \leq c < 2^n - 1$. Then $a = (b-2)(2^n-1) + d_1 + d_2$, where $d_1 = \lfloor (2^n - 1) - \frac{c}{2} \rfloor$ and $d_2 = \lceil (2^n - 1) - \frac{c}{2} \rceil$. A routine calculation now shows that $2^{n-1} \leq d_1 \leq d_2 < 2^n$, and so $a \cdot c_k (n)$ is the sum of $b$ binary $k$'th powers. \end{proof} \subsection{Change of basis and the Vandermonde matrix} \label{vander} In what follows, matrices and vectors are always indexed starting at $0$. Recall that a Vandermonde matrix $$ V(a_0, a_1, \ldots, a_{k-1})$$ is a $k\times k$ matrix where the entry in the $i$'th row and $j$'th column, for $0 \leq i, j < k$, is defined to be $a_i^j$. The matrix is invertible if and only if the $a_i$ are distinct. Recall that $c_k (n) = 1 + 2^n + 2^{2n} + \cdots + 2^{(k-1)n}$. For $k \geq 1$ and $n \geq 0$ we have \begin{equation} \left[ \begin{array}{c} c_k (n) \\ c_k (n+1) \\ \vdots \\ c_k(n+k-1) \end{array} \right] = M_k \left[ \begin{array}{c} 1 \\ 2^n \\ \vdots \\ 2^{(k-1)n} \end{array} \right], \label{kane} \end{equation} where $M_k = V(1, 2, 4, \ldots, 2^{k-1})$. For example, $$ M_4 = \left[ \begin{array}{cccc} 1 & 1 & 1 & 1 \\ 1 & 2 & 4 & 8 \\ 1 & 4 & 16 & 64 \\ 1 & 8 & 64 & 512 \end{array} \right].$$ Let a natural number $Y$ be represented as an ${\mathbb{N}}$-linear combination $$Y = a_0 + a_1 2^n + \cdots + a_{k-1} 2^{(k-1)n}.$$ Then, multiplying Eq.~\eqref{kane} on the left by $$ [b_0 \quad b_1 \quad \cdots \quad b_{k-1}] := [a_0 \quad a_1 \quad \cdots \quad a_{k-1}] M_k^{-1},$$ we get the following expression for $Y$ as a ${\mathbb{Q}}$-linear combination of binary $k$'th powers: $$ Y = b_0 c_k (n) + b_1 c_{k}(n+1) + \cdots + b_{k-1} c_k(n+k-1) .$$ It remains to estimate the size of the coefficients $b_i$, as well as the sizes of their denominators. The Vandermonde matrix is well studied (e.g., \cite[pp.~43, 105]{PS}). We recall one basic fact about it. \begin{lemma} The determinant of $V(a_0, a_1, \ldots, a_{k-1})$ is $$ \prod_{0 \leq i < j < k} (a_j - a_i) .$$ \label{van} \end{lemma} We now define $d_k$ to be the determinant of $M_k$, and $\ell_k$ to be the largest of the absolute values of the entries of $M_k^{-1}$. Note that, by Lemma~\ref{van}, $d_k$ is positive. Also, Laplace's formula tells us that $M_k^{-1} = M'_k d_k^{-1}$, where $M'_k$ is the adjugate (classical adjoint) $M'_k$ of $M_k$. Furthermore, since $M_k$ has integer entries, so does $M'_k$. \begin{proposition} We have $0 < d_k < 2^{k^3/3}$ for $k \geq 1$. \end{proposition} \begin{proof} By the formula of Lemma~\ref{van} we know that $$d_k = \prod_{0 \leq i < j< k} (2^j - 2^i) < \prod_{0 \leq i < j< k} 2^j = 2^{k^3/3 -k^2/2 + k/6} < 2^{k^3/3}$$ for $k \geq 1$. \end{proof} Our next result demonstrates that $\ell_k$, the absolute value of the largest entry in $M_k^{-1}$, is bounded above by a constant. \begin{proposition} We have $\ell_k < 34$. \end{proposition} \begin{proof} As is well known (see, e.g., \cite[Exercise 1.2.3.40]{K}, the $i$'th column in the inverse of the Vandermonde matrix $V(a_0, a_1, \ldots, a_{k-1})$ consists of the coefficients of the polynomial $$ p(x) := \frac{\prod_{{0 \leq j< k} \atop {j\not=i}}(x-a_i)}{\prod_{{0 \leq j< k} \atop {j\not=i}} (a_j - a_i)}.$$ We also observe that if $$(x-b_1)(x-b_2) \cdots (x-b_n) = x^n + c_{n-1} x^{n-1} + \cdots + c_1x + c_0,$$ is a polynomial with real roots, then the absolute value of every coefficient $c_i$ is bounded by $$ |c_0| + \cdots + |c_{n-1}| \leq \prod_{1 \leq i \leq n} (1 + |b_i|) .$$ Putting these two facts together, we see that all of the entries in the $i$'th column of $V(a_0, a_1, \ldots, a_{n-1})^{-1}$ are, in absolute value, bounded by $$ P_k (i) := \frac{\prod_{{0 \leq j< k} \atop {j\not=i}}(1+ |a_j|)}{\prod_{{0 \leq j< k} \atop {j\not=i}} |a_j - a_i|}.$$ Now let's specialize to $a_i = 2^i$. We get $$ P_k (i) := \frac{\prod_{{0 \leq j< k} \atop {j\not=i}}(2^j +1)}{\prod_{{0 \leq j< k} \atop {j\not=i}} |2^j - 2^i|} \\ \leq \frac{\prod_{0 \leq j< k}(2^j + 1)}{\prod_{{0 \leq j< k} \atop {j\not=i}} |2^j - 2^i |}. $$ To finish the proof of the upper bound, it remains to find a lower bound for the denominator $$ Q_k (i) := \prod_{{0 \leq j < k} \atop {j \not= i}} |2^j - 2^i| .$$ We claim, for $k \geq 2$, that \begin{equation} Q_k (0) \geq Q_k (1) \label{ineq1} \end{equation} and \begin{equation} Q_k (1) \leq Q_k (2) \leq \cdots \leq Q_k (k-1). \label{ineq2} \end{equation} To see \eqref{ineq1}, note that $Q_k (0) = \prod_{2 \leq j < k} (2^j - 1)$ and $Q_k (1) = \prod_{2 \leq j < k} (2^j - 2)$. On the other hand, by telescoping cancellation we see, for $1 \leq i \leq k - 2$, that $$ \frac{Q_k (i)}{Q_k (i+1)} = \frac{2^{k-1} - 2^i}{(2^{i+1} - 1)2^{k-2}} < \frac{2^{k-1}}{3 \cdot 2^{k-2}} = \frac23 ,$$ which proves \eqref{ineq2}. Hence $Q_k (i)$ is minimized at $i = 1$. Now \begin{align*} \ell_k &\leq \frac{\prod_{0 \leq j < k} (2^j + 1)}{Q_k (i)} \leq \frac{\prod_{0 \leq j < k} (2^j + 1)}{Q_k (1)} \\[1pt] & = \frac{\prod_{0 \leq j < k} (2^j + 1)}{\prod_{2 \leq j < k} (2^j - 2)} < 2 \cdot 3 \cdot \prod_{j \geq 2} \frac{2^j + 1}{2^j - 2} \doteq 33.023951743\cdots < 34. \end{align*} \end{proof} \begin{remark} The tightest upper bound seems to be $\ell_k < 5.194119929183\cdots$ for all $k$, but we did not prove this. \end{remark} \subsection{Expressing fractions of powers of $2$ as sums of binary $k$'th powers} \label{fractions} In everything that follows, $k$ is an integer greater than $1$. \begin{lemma} Let $f > 1$ be an odd integer. Define $e = \lfloor \log_2 f \rfloor$, so that $2^e < f < 2^{e+1}$. Let $m$ be the order of $2$ in the multiplicative group of integers modulo $f$. Then for all integers $j \geq 1$, the number $$ \left\lfloor \frac{2^{jm+e}}{f} \right\rfloor$$ is a binary $j$'th power, whose base-$2$ representation consists of $j$ repetitions of a block of size $m$. \label{jef5} \end{lemma} \begin{proof} Since $m$ is the multiplicative order of $2$ modulo $f$, we have $2^m - 1 = fq$ for some positive integer $q$. Then $$ \frac{1}{f} = \frac{q}{2^m - 1} = q \sum_{i \geq 1} 2^{-im}.$$ Multiplying by $2^{jm+e}$ and splitting the summation into two pieces, we see that \begin{align}\label{twojmeoverf} \frac{2^{jm+e}}{f} &= q \cdot 2^{jm+e} \sum_{i \geq 1} 2^{-im} \nonumber\\ &= q \cdot 2^{jm+e} \sum_{1 \leq i \leq j} 2^{-im} + q \cdot 2^{jm+e} \sum_{i > j} 2^{-im} \nonumber\\ &= q \cdot 2^e \cdot \frac{2^{jm}-1}{2^m - 1} + \frac{2^e}{f} . \end{align} Since $2^e < f$, the right-hand side of Eq.~(\ref{twojmeoverf}) is the sum of an integer and a number strictly between $0$ and $1$. It follows that $$ \left\lfloor \frac{2^{jm+e}}{f} \right\rfloor = q \cdot 2^e \cdot \frac{2^{jm}-1}{2^m - 1}.$$ It remains to see that $q \cdot 2^e$ is in the right range: we must have $2^{m-1} \leq q \cdot 2^e < 2^m$. To see this, note that $$ q \cdot 2^e = \frac{2^m -1}{f} \cdot 2^e \\ = \frac{2^{m+e}}{f} - \frac{2^e}{f} , $$ and, since $0 < 2^e / f < 1$, it follows that $$ \frac{2^{m+e}}{f} - 1 < q \cdot 2^e < \frac{2^{m+e}}{f} .$$ Rewriting gives $$ 2^{m-1} - 1 < 2^{m-1} \left( \frac{2^{e+1}}{f} \right) - 1 < q \cdot 2^e \leq \left( \frac{2^e}{f} \right) 2^m < 2^m,$$ or $2^{m-1} \leq q \cdot 2^e < 2^m$, as desired. \end{proof} \begin{lemma} Let $g$ be an integer with $g = 2^\ell \cdot f$, where $f \geq 1$ is odd. Then for all $n \geq kf + \ell + \log_2 f$, the number $ \lfloor \frac{2^n}{g} \rfloor$ can be written as the sum of at most $2^{kf-1}$ binary $k$'th powers and an integer $t$ with $0 \leq t \leq 2^{kf-1}$. \label{pows} \end{lemma} \begin{proof} There are two cases: (a) $f = 1$ or (b) $f > 1$. \medskip \noindent (a) $f = 1$: Using the division algorithm write $n - \ell = rk + i$ for $0 \leq i\leq k-1$. Since $n \geq \ell$ we have $$ \left\lfloor \frac{2^n}{g} \right\rfloor = \frac{2^n}{g} = 2^{n-\ell} = 2^{rk+i} = 2^i (2^{rk}- 1) + 2^i.$$ The base-$2$ representation of $2^{rk}-1$ is clearly a binary $k$'th power. Take $t = 2^i$. \medskip \noindent (b) $f > 1$: Let $e = \lfloor \log_2 f \rfloor$ and let $m$ be the order of $2$ in the multiplicative group of integers modulo $f$. Using the division algorithm, write $n - \ell - e = rkm + i$ for some $i$ with $0 \leq i \leq km - 1$. Note that since $n \geq kf + \ell + \log_2 f \geq km + \ell + e$ we have $r \geq 1$. Then \begin{align*} \frac{2^n}{g} &= \frac{2^{rkm+i+\ell+e}}{2^\ell \cdot f} = \frac{2^{rkm+i+e}}{f} \\ &= 2^i \cdot \frac{2^{rkm+e}}{f} = 2^i \left\lfloor \frac{2^{rkm+e}}{f} \right\rfloor + t, \end{align*} with $0 \leq t < 2^i$. Now take the floor of both sides and apply Lemma~\ref{jef5}. \end{proof} \subsection{The Frobenius number} Let $S$ be a set and $x$ be a real number. By $xS$ we mean the set $\{ xs \ : \ s \in S \}$. Let $S \subseteq {\mathbb{N}}$ with $\gcd(S) = 1$. The Frobenius number of $S$, written $F(S)$, is the largest integer that cannot be represented as a non-negative integer linear combination of elements of $S$. See, for example, \cite{RA}. As we have seen, $\gcd({\mathcal{S}}_k) = E_k = \gcd(k, 2^k - 1)$. Thus $\gcd(E_k^{-1} {\mathcal{S}}_k) = 1$. Define $F_k$ to be the Frobenius number of the set $E_k^{-1} {\mathcal{S}}_k$. In this section we give a weak upper bound for $F_k$. \begin{lemma} For $k \geq 2$ we have $F_k \leq 2^{k^2+k}$. \label{frobl} \end{lemma} \begin{proof} Consider $T = \{g_1, g_2, g_3 \}$ where $g_1 = 2^k - 1$, $g_2 = (2^k - 2) \frac{2^{k^2} - 1}{2^k -1}$, and $g_3 = (2^k - 1) \frac{2^{k^2} - 1}{2^k -1}$. We have $T \subseteq {\mathcal{S}}_k$. Let $d$ be the greatest common divisor of $T$. Then $d$ divides $g_3 - g_2 = \frac{2^{k^2} - 1}{2^k -1}$ and $g_1 = 2^k - 1$. So $d$ divides $D_k$. On the other hand, clearly, $A_k$ divides $d$, while from Theorem~\ref{sanna} we know that $A_k = D_k = E_k$. Hence, $d = E_k$. Clearly $F(E_k^{-1} {\mathcal{S}}_k) \leq F(E_k^{-1} T)$. Furthermore, since $g_1 \, | \, g_3$, it follows that $F(E_k^{-1} T) = F(\{ E_k^{-1} g_1, E_k^{-1} g_2 \})$. By a well-known result (see, e.g., \cite[Theorem 2.1.1, p.~31]{RA}), we have $F(\{a,b\}) = ab - a - b$, and the desired claim follows. \end{proof} \begin{remark} We compute explicitly that $F_2 = 17$, $F_3 = 723$, $F_4 = 52753$, $F_5 = 49790415$, and $F_6 = 126629$. \end{remark} \section{The complete proof} We are now ready to fill in the details of the proof of our main result, Theorem~\ref{main}. We recall the definitions of the following quantities that will figure in the proof: \begin{itemize} \item $c_k (n) = 1 + 2^n + \cdots + 2^{(k-1)n}$; \item $E_k = \gcd(k, 2^k - 1)$ is the greatest common divisor of the set ${\mathcal{S}}_k$ of binary $k$'th powers; \item $F_k$ is the Frobenius number of the set $E_k^{-1} {\mathcal{S}}_k$; \item $d_k$ is the determinant of the Vandermonde matrix $M_k = V(1,2,\ldots, 2^{k-1})$; \item $\ell_k$ is the largest of the absolute values of the entries of $M_k^{-1}$ \end{itemize} We will show that, for $k \geq 1$, there exists a constant $W(k)$ such that every integer $N > F_k E_k$ that is a multiple of $E_k = \gcd(k, 2^k - 1)$ can be written as the sum of $W(k)$ binary $k$'th powers. \begin{proof} The result is clear for $k = 1$, so let us assume $k \geq 2$ and that $N$ is a multiple of $E_k$. Define $Z = (F_k + 1) E_k$. In the proof there are several places where we need $N$ to be ``sufficiently large''; that is, greater than some constant $C > Z$ depending only on $k$; some are awkward to write explicitly, so we do not attempt to do so. Instead we just assume $N$ satisfies the requirement $N > C$. The cases $F_k E_k < N \leq C$ are then handled by writing $N$ as a sum of a constant number of elements of ${\mathcal{S}}_k$. Let $X := N - Z$. Let $c$ be a constant specified below, and let $n$ be the largest integer such that $2^n < c X^{1/k}$; we assume $N$ is sufficiently large so that $n \geq 1$. First we explain how to write $X = Y + D$, where \begin{enumerate}[(a)] \item $Y < c_k (n)$; and \item $D$ is an ${\mathbb{N}}$-linear combination of $c_k (n), \ldots, c_k(n+k-1)$ with all coefficients sufficiently large. \end{enumerate} To do so, define $Q = c_k (n) + \cdots + c_k (n+k-1)$, and $R = \lfloor X/Q \rfloor$. We have now obtained $RQ$ (a good approximation of $X$), which is an ${\mathbb{N}}$-linear combination of $c_k (n), \ldots, c_k(n+k-1)$ with every coefficient equal to $R$. Note that $0 \leq X-RQ < Q$. We now improve this approximation of $X$ using a greedy algorithm, as follows: from $x-RQ$ we remove as many copies as possible of $c_k(n + k-1)$, then as many copies as possible of $c_k (n+k-2)$, and so forth, down to $c_k (n)$. More precisely, for each index $i = k-1, k-2, \ldots, 0$ (in that order) set $$r_i = \left\lfloor \frac{X - RQ - \sum_{i < j < k} r_j c_k(n+j)}{c_k (n+i)} \right\rfloor ,$$ and then put $$D := RQ + r_0 c_k (n) + r_1 c_k (n+1) + \cdots + r_{k-1} c_k(n+k-1) .$$ By the way we chose the $r_i$, we have $0 \leq r_{k-1} < 2$ and $0 \leq r_i < c_k(n+i+1)/c_k(n+i) < 2^{k-1}$ for $0 \leq i \leq k-2$. Furthermore, $0 \leq y < c_k (n)$. Define $e_i = R + r_i$ for $0 \leq i < k$. Then $D = \sum_{0 \leq i < k} e_i c_k (n+i)$. Since $Y < c_k (n)$, we can express $Y$ in base $2^n$ as $Y = a_0 + a_1 2^n + \cdots + a_{k-1} 2^{(k-1)n}$, where each $a_i$ is an integer satisfying $0 \leq a_i < 2^n$. Apply the transformation discussed above in Section~\ref{vander}, obtaining the ${\mathbb{Q}}$-linear combination $$ Y = \sum_{0 \leq i < k} b_i c_k (n+i).$$ It follows that $X = \sum_{0 \leq i < k} (e_i+b_i) c_k (n+i)$. Furthermore, from Section~\ref{vander} we know that each $b_i$ is at most $k\ell_k \cdot 2^n$ in absolute value, and the denominator of each $b_i$ is at most $d_k$. Now we want to ensure that, for $0 \leq i < k$, it holds \begin{equation}\label{ebrightsize} 2^{n + i - 1} \leq e_i + b_i < c^\prime 2^n , \end{equation} where $c^\prime > 0$ is a constant depending only on $k$. We choose the constant $c$ mentioned above to get the bound (\ref{ebrightsize}). Pick $c > 0$ such that $c^{-k} = \left(2^{k-2} + k \ell_k + 1\right) 2^{k^2 - k + 1}$. Then we have \begin{align*} e_i + b_i &\geq R - k\ell_k \cdot 2^n > \frac{X}{Q} - k\ell_k \cdot 2^n - 1 > \frac{2^{kn} c^{-k}}{2 \cdot 2^{(k-1)(n+k) + 1}} - k\ell_k \cdot 2^n - 1 \\ &= 2^n \left(c^{-k} 2^{-k^2 + k - 1} - k\ell_k - 2^{-n}\right) > 2^{n + k - 2} \geq 2^{n + i - 1} , \end{align*} as desired. For the upper bound, recalling that our choice of $n$ implies that $c X^{1/k} \leq 2^{n+1}$, we have $$ e_i + b_i < R + 2^{k-1} + k\ell_k 2^n \leq \frac{X}{Q} + 2^{k-1} + k\ell_k 2^n < \frac{2^{k(n+1)}c^{-k}}{2^{(k-1)(n + k - 1)}} + 2^{k-1} + k\ell_k 2^n,$$ and a routine calculation shows that $e_i + b_i \leq c' 2^n$, where the $c^\prime$ depends only on $k$. The only problem left to resolve is that the $e_i + b_i$ need not be integers. Write $X = X_1 + X_2$, where $$X_1 := \sum_{0 \leq i < k} \lfloor e_i + b_i \rfloor c_k (n+i) .$$ Thanks to (\ref{ebrightsize}), we can use Lemma~\ref{split} to rewrite $X_1$ as a sum of a constant number of binary $k$'th powers. Then $$ X_2 = \sum_{0 \leq i < k} a_i c_k (n+i),$$ where $0 < a_i < 1$ is a rational number with denominator $d_k$. Writing $a_i = v_i / d_k$, we see \begin{align*} X_2 &= \sum_{0 \leq i < k} \frac{v_i}{d_k} \sum_{0 \leq j < k} 2^{(n+i)j} \\ &= \sum_{0 \leq i, j < k} v_i \cdot \frac{2^{(n+i)j}}{d_k} \\ &= \sum_{0 \leq i, j < k} v_i \left\lfloor \frac{2^{(n+i)j}}{d_k} \right\rfloor + X_3, \end{align*} where $0 \leq X_3 < d_k \cdot k^2$. By Lemma~\ref{pows} we know that, provided $(n+i)j > kd_k + 2 \log_2 d_k$, each term $\left\lfloor \frac{2^{(n+i)j}}{d_k} \right\rfloor$ is the sum of a constant number of binary $k$'th powers, plus an error term that is at most $2^{k d_k}$. Thus, provided $n$ (and hence $N$) are large enough, this will be true for all exponents except those corresponding to $j = 0$. Those exponents are not a problem, since for $j = 0$ we have $\left\lfloor \frac{2^{(n+i)j}}{d_k} \right\rfloor = 0$, because $d_k > 1$ for $k > 1$. It follows that $X_4 := X_2-X_3$ is the sum of a constant number of binary $k$'th powers. Putting this all together, we have expressed $$ N = X_1 + X_4 + X_3 + Z,$$ where $X_1$ and $X_4$ are both the sum of a constant number of binary $k$'th powers, and $X_3 + Z$ is bounded below by $(F_k+1)E_k$ and above by a constant. Now $N$, $X_1$, and $X_4$ are all multiples of $E_k$, so the ``error term'' $X_3 + Z$ must also be a multiple of $E_k$. Furthermore, the error term is larger than $E_k F_k$ and hence is representable as a sum of a constant number of binary $k$'th powers. \end{proof} \section{Final remarks} \label{final} Everything we have done in this paper is equally applicable to expansions in bases $b > 2$, with one minor complication: it may be that if $b$ is not a prime, then the base-$b$ expansion of $1/d$ might not be purely periodic (after removing any leading zeros), but only ultimately periodic. This adds a small complication in Section~\ref{fractions}. However, this case can easily be handled, and we leave the details to the reader. The bounds we obtained in this paper for $W(k)$ are very weak --- at least doubly exponential --- and can certainly be improved. We leave this as work for the future. For example, we have \begin{conjecture} Every natural number $> 147615$ is the sum of at most nine binary cubes. The total number of exceptions is 4921. \end{conjecture} \begin{remark} We have verified this claim up to $2^{27}$. \end{remark} There is another approach to Waring's theorem for binary powers that could potentially give much better bounds for $W(k)$. For sets $S, T \subseteq {\mathbb{N}}$ define the {\it sumset} $S + T$ as follows: $$ S + T = \lbrace s + t \ : \ s \in S, t \in T \rbrace.$$ We make the following conjecture: \begin{conjecture} Writing $C_n$ for the set $\{ a\cdot c_k^2 (n) \ : \ 2^{n-1} \leq a < 2^n \}$ of cardinality $2^{n-1}$ (i.e., the $kn$-bit binary $k$'th powers), for $n, k \geq 1$, all the elements in the sumset $$C_n + C_{n+1} + \cdots + C_{n+k-1},$$ are actually represented {\it uniquely\/} as a sum of $k$ elements, one chosen from each of the summands. \end{conjecture} If this conjecture were true --- we have proved it for $1 \leq k \leq 3$ --- it would prove that ${\mathcal{S}}_k$ has positive density, and hence, by a result of Nathanson \cite[Theorem 11.7, p.~366]{N2}, that it is an asymptotic additive basis. From this we could obtain better bounds on $W(k)$. \medskip In the light of our results, it seems natural to ask about the set ${\mathcal{T}}_1^b$ of positive integers $k$ such that $\gcd({\mathcal{S}}_k^b) = 1$. Indeed, we have that the elements of ${\mathcal{T}}_1^b$ are exactly the integers $k$ such that ${\mathcal{S}}_k^b$ forms an asymptotic additive basis for $\mathbb{N}$. It turn out that ${\mathcal{T}}_1^b$ has a natural density, and even more can be said: since $\left(\frac{b^k - 1}{b - 1}\right)_{k \geq 1}$ is a Lucas sequence, we can employ the same methods of \cite{ST} to prove the following result: \begin{theorem} For all integers $g \geq 1$, $b \geq 2$, the set ${\mathcal{T}}_g^b$ of positive integers $k$ such that $\gcd({\mathcal{S}}_k^b) = g$ has a natural density, given by \begin{equation*} \mathbf{d}({\mathcal{T}}_g^b) = \sum_{d \geq 1 \text{ coprime with } b} \frac{\mu(d)}{L_b(dg)}, \end{equation*} where $\mu$ is the M\"obius function and $L_b(x) := \lcm(x, \ord_x(b))$, where $\ord_x(b)$ is the multiplicative order of $b$, modulo $x$. In particular, the series converges absolutely. Furthermore, $\mathbf{d}({\mathcal{T}}_g^b) > 0$ if and only if ${\mathcal{T}}_g^b \neq \varnothing$ if and only if $g = \gcd\!\left(L_b(g), \frac{b^{L_b(g)} - 1}{b - 1}\right)$. \end{theorem} Also, employing the methods of \cite{LS}, the counting function of the set $\{g \geq 1 : {\mathcal{T}}_g^b \neq \varnothing\}$ can be shown to be $\gg x / \log x$ and at most $o(x)$, as $x \to +\infty$. Note only that, in doing so, where in \cite{LS} results of Cubre and Rouse~\cite{CubreRouse} on the density of the set of primes $p$ such that the rank of appearance of $p$ in the Fibonacci sequence is divisible by a fixed positive integer $m$ are used, one should instead use results on the density of the set of primes $p$ such that $\ord_p(b)$ is divisible by $m$ --- for example, those given by Wiertelak~\cite{Wie}. \section*{Acknowledgment} We are grateful to Igor Pak for introducing the first and third authors to each other. \bibliographystyle{amsplain}

https://arxiv.org/abs/math/0507472

A Result About the Density of Iterated Line Intersections in the Plane

Let $S$ be a finite set of points in the plane and let $\mathcal{T}(S)$ be the set of intersection points between pairs of lines passing through any two points in $S$. We characterize all configurations of points $S$ such that iteration of the above operation produces a dense set. We also discuss partial results on the characterization of those finite point-sets with rational coordinates that generate all of $\mathbb Q^2$ through iteration of $\mathcal{T}$.

\section{Introduction} Let $S$ be a set of points in the plane and let $L = \{L_{i}\}_{i \in I}$ be the set of lines between pairs of points in $S$. Consider the following operation on $S$: \begin{equation} \mathcal{T}(S) = \bigcup\limits_{i \ne j} {L_i \cap L_j } \subseteq \mathbb R^2. \end{equation} In other words, $\mathcal{T}(S)$ is the set of intersection points between pairs of distinct lines in $L$. If $S$ consists of $n$ collinear points (or no points at all), then the union above is empty; so to keep the notation consistent, we set $\mathcal{T}(S) = \emptyset$ for these cases. \begin{figure}[!htbp] \begin{center} \includegraphics[scale=0.45]{trapezoiditeration} \end{center} \caption{$\mathcal{T}(S)$ for a set of points $S$ that form a trapezoid.} \label{fig.trapezoid} \end{figure} As a simple example of the operation $\mathcal{T}$, let $S$ consist of four black points that are the vertices of a trapezoid as in Figure \ref{fig.trapezoid}. Then, $\mathcal{T}$$(S)$ consists of the original four points along with two additional ones shown in gray. It should be clear that for a set of points not all collinear, we have $S \subseteq \mathcal{T}(S)$. Moreover, $\mathcal{T}(S)$ is finite for finite sets $S$. We are interested here in the iterations, $\mathcal{T}^i(S)$, and specifically, the limiting behavior of such operations on arbitrary finite sets $S$. The study of such phenomenon naturally leads to the notion of the order of a set $S$, which we define below. As a matter of convention, we set $\mathcal{T}^0(S)$ = S. \begin{defn} Let $S$ be a set of points in $\mathbb R^2$. The \emph{order} of $S$ is the smallest positive integer $n$ such that $\mathcal{T}^n(S) = \mathcal{T}^{n-1}(S)$. If there is no such $n$, then the order of $S$ is defined to be $\infty$. \end{defn} For example, the order for a set of points forming the vertices of a square is 2. If the order of a set $S$ is 1, then we call $S$ \emph{fixed} under $\mathcal{T}$. A set $S$, therefore, has finite order if and only if $\mathcal{T}^n(S)$ is fixed for some nonnegative integer $n$. \begin{prob}\label{fixedprob} Describe the finite point-sets that have finite order. \end{prob} Before discussing the answer to this problem (in Section \ref{fixedsetssec}), we describe a nontrivial infinite point-set that has finite order. Let $S$ be the set of rational points on the unit circle, $x^2 + y^2 = 1$. For a given $P \in \mathbb Q^2$, choose two points $A$ and $B$ in $S$ such that $PA$ and $PB$ are not tangent to the unit circle. Then, if $C$ and $D$ are the points of intersection of $PA$ and $PB$ (respectively) with the circle, it turns out \cite[p. 249]{numtheory} that $C$ and $D$ are both rational. It follows that $P \in \mathcal{T}(S)$ for every $P \in \mathbb Q^2$, and thus \[\mathcal{T}^2(S) = \mathcal{T}(\mathbb Q^2) = \mathbb Q^2 = \mathcal{T}(S).\] Excluding the sets of finite order, it follows that iteration of $\mathcal{T}$ produces a strictly increasing chain of sets of points in the plane. In light of this observation, a natural question is whether we arrive at a dense set of points by such a procedure. In other words, is $\bigcup\nolimits_{i \geq 0} {\mathcal{T}^{i} \left( S \right)}$ dense in $\mathbb R^2$? A more difficult but related question is whether we get all of $\mathbb Q^2$ when $S$ consists of only rational points. We address both of these questions with a complete answer to the first in Section \ref{denseproof} and some partial results for the second in Section \ref{rationalsection}. \begin{thm}\label{densethm} Let $S$ be a finite set of points in the plane. Then, $S$ has infinite order if and only if $\bigcup\nolimits_{i \geq 0} {\mathcal{T}^{i} \left( S \right)}$ is dense in $\mathbb R^2$. \end{thm} The answer to Problem \ref{fixedprob} found in Corollary \ref{fixedthmcor} below, therefore, gives a complete characterization of when iterated line intersections are dense. \begin{cor}\label{fulldensechar} Let $S$ be a finite set of points in the plane. Then, $\bigcup\nolimits_{i \geq 0} {\mathcal{T}^{i} \left( S \right)}$ is dense in $\mathbb R^2$ if and only if $S$ is not one of the following sets: \begin{enumerate} \item The empty set. \item A set of collinear points. \item A set of collinear points with one additional noncollinear point. \item The vertices of a parallelogram. \item The vertices of a parallelogram and the intersection of its two diagonals. \end{enumerate} \end{cor} In the rational case, we conjecture a more exact result. \begin{conj}\label{rationalthm} Let $S$ be a finite set of points in the plane with rational coordinates. Then, $S$ has infinite order if and only if $\bigcup\nolimits_{i \geq 0} {\mathcal{T}^{i} \left( S \right)} = \mathbb Q^2.$ \end{conj} As a step in the direction of this conjecture, we offer the following; its proof can be found in Section \ref{rationalsection}. \begin{thm}\label{rationalparallelthm} Let $R, P, Q, T \in S$ be rational points in the plane with $RQ$ and $PT$ parallel and suppose that $RP$ is not parallel to $QT$. Then, $\bigcup\nolimits_{i \geq 0} {\mathcal{T}^{i} \left( S \right)} = \mathbb Q^2$. \end{thm} Though we were not motivated by any other particular work, we should remark that a similar question posed by Fejes-Toth (with circles replacing lines) was addressed by Bezdek and Pach in \cite{pach}, and related results can also be found in the papers \cite{barany, king}. Additionally, Theorem \ref{densethm} has also been discovered recently (independently) by Ismailescu and Radoicic \cite{otherproof}. \section{Finite Fixed Sets}\label{fixedsetssec} We begin by characterizing sets of finite order. Although one may deduce the main result of this section from Lemmas \ref{trianglelemma} and \ref{trianglelemma2} in Section \ref{denseproof}, the methods employed here are less cumbersome and might be of independent interest. We will need the following result from elementary geometry. \begin{thm}[The Sylvester-Gallai Theorem]\label{sylthm} For every set of $n$ noncollinear points in the plane, there exists a line that contains exactly two of the points. \end{thm} Although this fact seems intuitively obvious, its proof eluded even Sylvester, and it was only solved (in published form) some $50$ years after being posed by him \cite{solvedsylvester}. We refer the reader to \cite{bookproof} for more details. We are ready to approach Problem \ref{fixedprob}. \begin{thm}\label{fixedthm} A finite set $S$ fixed under $\mathcal{T}$ must be one of the following configurations: \begin{enumerate} \item The empty set. \item A set of collinear points with one additional noncollinear point. \item The vertices of a parallelogram and the intersection of its two diagonals. \end{enumerate} \end{thm} \begin{proof} Let $S$ be a set of $n$ noncollinear points in the plane that is fixed by $\mathcal{T}$. Using Theorem \ref{sylthm}, there exists a line intersecting $S$ in exactly two points $P$ and $Q$. By assumption, there is some other point $X$ not on this line, and we can choose $X$ so that its altitude from $PQ$ is largest. If all other points lie on the line $XP$ or if all of them lie on $XQ$, then we are in configuration (2) above. The remaining possibilities break up into two cases. \begin{figure}[!htbp] \begin{center} \includegraphics[scale=0.45]{fixedpointdiagram2} \end{center} \caption{Case 1 in the proof of Theorem \ref{fixedthm}} \label{fig.fixedproof1} \end{figure} \underline{Case 1}: There is a point $Y \in S$ not on $XP$ and not on $XQ$. We first claim that $Y$ must lie on the line through $X$ that is parallel to $PQ$. Indeed, any other position for $Y$ would give rise to an intersection between $XY$ and $PQ$ that is not $P$ or $Q$, contrary to our use of Theorem \ref{sylthm} and our assumption that $\mathcal{T}(S) = S$. Relabeling if necessary, Figure \ref{fig.fixedproof1} depicts the situation. Since $S$ is fixed, the intersection point, $Z$, of $XQ$ and $PY$ is in $S$. It follows that $XP$ and $YQ$ must be parallel (otherwise, if $W$ is the intersection point of $XP$ and $YQ$, then $ZW$ would intersect $PQ$). Finally, it is easy to see that there can be no other points in $S$ by our choice of $P$ and $Q$. \begin{figure}[!htbp] \begin{center} \includegraphics[scale=0.45]{fixedpointdiagramcase2} \end{center} \caption{Case 2 in the proof of Theorem \ref{fixedthm}} \label{fig.fixedproof2} \end{figure} \underline{Case 2}: Every point in $S$ lies on one of the lines $XP$ or $XQ$. If $S$ is not a configuration of type (2), then there are points $R,T \in S$ such that $R$ is on the line $XP$, $T$ is on the line $XQ$, and $R,T$ are not $X,P$, or $Q$. By the assumption on $X$ and the line $PQ$, only two configurations for $R$ and $T$ are possible; these are depicted in Figure \ref{fig.fixedproof2}. In both cases, two iterations of $\mathcal{T}$ give rise to a point in $S$ on the line $PQ$, a contradiction. Therefore, no fixed point-sets other than those of configuration (2) may take this form. This completes the proof. \end{proof} \begin{cor}\label{fixedthmcor} The finite point-sets with finite order are \begin{enumerate} \item The empty set. \item A set of collinear points. \item A set of collinear points with one additional noncollinear point. \item The vertices of a parallelogram. \item The vertices of a parallelogram and the intersection of its two diagonals. \end{enumerate} \end{cor} \begin{proof} Let $S$ be a finite set in $\mathbb R^2$ with order $n$. Applying Theorem \ref{fixedthm}, it follows that $R = \mathcal{T}^{n-1}(S)$ must be one of three types. When $R$ is empty, then $S$ is either itself empty or a set of collinear points. Similarly, a set $R$ of collinear points with one additional point can only be obtained from a set $S$ that is the same as $R$. Finally, when $R$ forms a parallelogram with the intersection of its diagonals, the set $S$ must either be $R$ or $R$ without its diagonal intersection. \end{proof} \section{The Density Theorem}\label{denseproof} Before proving Theorem \ref{densethm}, we record the following technical lemmas, the first of which provides a useful characterization of sets of infinite order. For ease of presentation, we say that a point is \emph{strictly contained} in a set $K$ if it is located in its interior. \begin{lem}\label{trianglelemma} Let $S$ be a finite set of infinite order. Then, there exists $n \in \mathbb N$ such that $\mathcal{T}^n(S)$ contains a subset of 4 points in which 3 of the points are noncollinear and the fourth point is strictly contained in the triangle determined by these 3 points. \end{lem} \begin{proof} We consider the number of vertices $v$ on the convex hull $H$ of $S$. When $v = 2$, the set $S$ cannot have infinite order. So suppose that $v = 3$. If there is a point of $S$ strictly contained inside $H$, then we are done. Otherwise, since $S$ has infinite order, there must be two points of $S$ on different edges of $H$. An iteration of $\mathcal{T}$ then produces our desired point. \begin{figure}[!htbp] \begin{center} \includegraphics[scale=0.45]{trianglemma} \end{center} \caption{Four vertices on the convex hull of $S$} \label{fig.trianglempic} \end{figure} Assume now that $H$ has exactly four vertices. If these vertices do not form a parallelogram, then one iteration of $\mathcal{T}$ gives us what we want (see Figure \ref{fig.trianglempic}). Otherwise, there is a point in $S$ which is not a vertex of $H$ and not the intersection of the diagonals of the quadrilateral determined by $H$. Again in this case, one iteration of $\mathcal{T}$ (giving us the intersection of the two diagonals of $H$) produces the desired result. Finally, if $v > 4$, then we proceed as follows. Pick two adjacent vertices $A$ and $B$. There must be two other vertices $C$ and $D$ such that the edges $AB$ and $CD$ are not parallel ($H$ has at least $5$ vertices and is convex). This reduces the problem to the case of $4$ vertices not forming a parallelogram (encountered above) and completes the proof of the lemma. \end{proof} Our next result allows one to produce a convergent, nested sequence of triangles. \begin{lem}\label{trianglelemma2} Let $A$, $B$, and $C$ be noncollinear points, and let $P$ be a point strictly inside $\triangle ABC$. Then, there exist triangles $\triangle A_nB_nC_n$ ($n = 1,2,\ldots$) strictly containing $P$ such that $\mathop {\lim }\limits_{n \to \infty } A_n = \mathop {\lim }\limits_{n \to \infty } B_n = \mathop {\lim }\limits_{n \to \infty } C_n = P$, and for each $n$, \[A_n,B_n,C_n \in \bigcup\limits_{j = 0}^\infty { \mathcal{T}^{(j)} \left(\{A,B,C,P\} \right )}.\] \end{lem} \begin{figure}[!htbp] \begin{center} \includegraphics[scale=0.45]{nestedtriangles} \end{center} \caption{Nested triangle iteration} \label{fig. nestedtriangles} \end{figure} \begin{proof} Given a triangle $\triangle ABC$ and a point $P$ strictly contained in it, we may construct the vertices of another triangle containing this point by intersecting the lines $AP$, $BP$, and $CP$ with the edges of $\triangle ABC$. Iterating this procedure produces a nested sequence of triangles strictly containing $P$ with vertices in $\bigcup\nolimits_{j \geq 0} {\mathcal{T}^{(j)} \left( \{A,B,C,P \}\right)}$ (see Figure \ref{fig. nestedtriangles}). This sequence contains two types of triangles; we label the odd iterates $\triangle D_nE_nF_n$, while even iterates are denoted by $\triangle A_nB_nC_n$. Here, the $A_n$ (resp. $B_n$, $C_n$) are labeled so that they are the ones on the line $AP$ (resp. $BP$, $CP$). We claim the vertices of the triangles $\triangle A_nB_nC_n$ all converge to $P$. \begin{figure}[!htbp] \begin{center} \includegraphics[scale=0.5]{nestedtriangles2} \end{center} \caption{Iterations decrease triangle areas} \label{fig. nestedtriangles2} \end{figure} To verify this assertion, it suffices to show that $|A_1P| < |AA_1|$, $|B_1P| < |BB_1|$, and $|C_1P| < |CC_1|$. Without loss of generality, we prove that $|A_1P| < |AA_1|$. Reducing further, we observe that it is enough to show that the area of $\triangle PD_1F_1$ is less than the area of $\triangle AD_1F_1$ (drop altitudes to $D_1F_1$ from $A$, $P$ and compare similar triangles). Next, draw the line $JK$ that is parallel to $D_1F_1$ and passes through $P$, and label the angles formed as in Figure \ref{fig. nestedtriangles2}. Since $F_1P$ and $AJ$ (resp. $D_1P$ and $AK$) intersect at $B$ (resp. $C$), it follows that $\alpha < \beta$ and $\gamma < \delta$. Therefore, when we form the triangle $\triangle QD_1F_1$ that is congruent to $\triangle PD_1F_1$, it must lie entirely inside $\triangle AD_1F_1$. This finishes the proof. \end{proof} \begin{lem}\label{densetrianglelem} Let $A$, $B$, $C$ be noncollinear points in the plane. If $K$ is a dense set of points in $\triangle ABC$, then $\mathcal{T}(K)$ is a dense set of points in the entire plane. \end{lem} \begin{proof} Let $P$ be a point in the plane, and let $Q_1, Q_2$ and $R_1, R_2$ be points strictly inside $\triangle ABC$ such that $Q_1Q_2$ and $R_1R_2$ intersect at $P$. Since $K$ is dense in $\triangle ABC$, there are a sequence of points $Q_{1n},Q_{2n} \in K$ and $R_{1n}, R_{2n} \in K$ that converge to $Q_1, Q_2$ and $R_1, R_2$, respectively. Since the intersection of two lines formed by four points is continuous in the four points (the intersection is a rational function in the coordinates of the four points), it follows that the intersections of $Q_{1n}Q_{2n}$ and $R_{1n}R_{2n}$ (which are in $\mathcal{T}(K)$) converge to $P$. This completes the proof. \end{proof} We are ready to prove Theorem \ref{densethm}. \begin{proof}[Proof of Theorem \ref{densethm}] The if-direction ($\Leftarrow$) in the theorem statement is immediate. Therefore, let $S$ be a finite set of infinite order. Using Lemma \ref{trianglelemma}, there exists $n \in \mathbb N$ such that $\mathcal{T}^n(S)$ contains a triangle of vertices and a fourth point strictly contained in the triangle determined by these 3 vertices. We claim that iteration of $\mathcal{T}$ on these 4 points produces a dense set of points in the triangle. The theorem then follows from Lemma \ref{densetrianglelem}. Let $A$, $B$, and $C$ be the vertices of the triangle strictly containing $P$. Suppose that $K = \bigcup\nolimits_{j \geq 0} {\mathcal{T}^{(j)} \left( A,B,C,P \right)}$ does not contain a dense set of points in $\triangle ABC$; we will derive a contradiction. Using Lemma \ref{trianglelemma2}, we can produce a sequence of triangles, $\triangle A_iB_iC_i$, with vertices in $K$ such these vertices converge to $P$. Let $h$ be so large that the circle centered at $P$ with radius equal to twice the largest side of $\triangle A_hB_hC_h$ is strictly contained in $\triangle ABC$. Since $K$ is not dense in $\triangle ABC$, it follows that $K$ cannot be dense in $\triangle A_hB_hC_h$ (again using Lemma \ref{densetrianglelem}). Let $\overline{K}$ be the closure of $K$ and set $W = \overline{K} \cap \triangle A_hB_hC_h$. Also, let Int$(\triangle A_hB_hC_h)$ denote the interior of $\triangle A_hB_hC_h$. Since $K$ is not dense in the triangle $\triangle A_hB_hC_h$, the (nonempty) open set Int$(\triangle A_hB_hC_h) \setminus W$ contains an open ball centered at some point $X$ inside $\triangle A_hB_hC_h$. Consider the set of all closed balls centered at $X$ that do not intersect $\overline{K}$, and let $r > 0$ denote the supremum over all radii of such balls. The closed ball $\overline{B}(X,r)$ of radius $r$ centered at $X$ must be strictly contained in $\triangle ABC$ since its interior cannot contain $A_h$, $B_h$, or $C_h$ (they are in $K$) and because of how we chose $h$. \begin{figure}[!htbp] \begin{center} \includegraphics[scale=0.45]{densityprooffig} \end{center} \caption{Obtaining a contradiction} \label{fig.balltangent} \end{figure} By construction of $\overline{B}(X,r)$, there exists a point $Y \in \overline{K}$ intersecting the boundary of $\overline{B}(X,r)$. Consider the lines $AY$, $CY$, and $BY$, and notice that they cannot all be tangent to the ball $\overline{B}(X,r)$ (there is only one tangent line through a point on a circle). Therefore, at least one of these lines through $Y$, say $AY$, must intersect the interior of $\overline{B}(X,r)$. Let $Z$ be the intersection of the line $AY$ with the boundary of $\overline{B}(X,r)$ (the point $Z$ need not be in $\overline{K}$). The situation is depicted in Figure \ref{fig.balltangent}. The dashed line through $Y$ is the line tangent to the boundary of $\overline{B}(X,r)$ at $Y$, while the dashed line through $Z$ is parallel to it. To continue, we observe the following straightforward fact that was discussed in the proof of Lemma \ref{densetrianglelem}: If $U,V,Q,R \in \overline{K}$ determine two nonparallel lines $UV$ and $QR$, then the intersection point of $UV$ and $QR$ is in $\overline{K}$. With this observation in mind, we may use Lemma \ref{trianglelemma2} to obtain vertices of triangles $\triangle A'_iB'_iC'_i$ in $\overline{K}$ that contain $Y$ and that also converge to $Y$. None of the vertices $A'_i$, $B'_i$, or $C'_i$ is in the interior of $\overline{B}(X,r)$ by our choice of $r$. Finally, we claim that for large enough $n$, the segment $YZ$ must intersect a side of $\triangle A'_n B'_n C'_n$ in the interior of $\overline{B}(X,r)$, a contradiction to our assumption on $r$. To see this, notice that for a large $n$, at least one of the vertices of $\triangle A'_n B'_n C'_n$ must lie between the two parallel lines (depicted in Figure \ref{fig.balltangent}) through $Y, Z$, while none of them will lie beneath the line through $Z$. It follows that an edge of $\triangle A'_n B'_n C'_n$ intersects the line $AY$ inside $\overline{B}(X,r)$. This contradiction completes the proof. \end{proof} \section{The Rational Case}\label{rationalsection} We now turn our attention to the case of rational points as in the statement of Conjecture \ref{rationalthm}. We note the following simple observation. \begin{lem}\label{6pointlemma} Suppose that $S = \{(0,0),(0,1),(0,2),(1,0),(1,1),(1,2)\}$ or that $S = \{(0,0),(0,1),(0,2),(1,0),(1,-1),(1,-2)\}$. Then, $\bigcup\nolimits_{i \geq 0} {\mathcal{T}^{i} \left( S \right)} = \mathbb Q^2.$ \end{lem} \begin{proof} Iteration of $\mathcal{T}$ on both sets above gives all of $\mathbb Z^2$, and it is easily verified that $\mathbb Z^2$ generates all of $\mathbb Q^2$. \end{proof} \begin{figure}[!htbp] \begin{center} \includegraphics[scale=0.45]{midpointlemma} \end{center} \caption{Midpoint Lemma \ref{midpointlemma}} \label{fig.midpoint} \end{figure} We next restrict our attention to a particular case involving a pair of parallel lines. We need the following fact from plane geometry. \begin{lem}\label{midpointlemma} Let $R, P, Q, T$ be points in the plane with $RQ$ and $PT$ parallel and suppose that $RP$ is not parallel to $QT$. Let $Y$ be the intersection of $RT$ and $PQ$ and set $X$ to be the intersection of $RP$ and $QT$. Then, $XY$ intersects $RQ$ and $PT$ in their midpoints $U$ and $V$, respectively. \end{lem} \begin{proof} Since $\bigtriangleup RUY$ and $\bigtriangleup TVY$ are similar triangles, we have $RU/TV = UY/VY$. The same reasoning gives us that $UY/VY = UQ/VP$. Examining the large triangles $\bigtriangleup XVT$ and $\bigtriangleup XPV$, it is also clear that $UQ/TV = XU/XV = RU/VP$. Therefore, \begin{equation*} UQ = TV \cdot \frac{RU}{VP} = TV^2 \cdot \frac{UQ}{VP^2}, \end{equation*} so that $TV = VP$. A similar computation shows that $RU = UQ$. \end{proof} We finally arrive at our main result in the rational case. It will be a consequence of Lemma \ref{midpointlemma}, and it is the closest we come to proving Conjecture \ref{rationalthm}. \begin{proof}[Proof of Theorem \ref{rationalparallelthm}] Since a (rational) translation does not change the problem, we may assume that $Q = (0,0)$. Moreover, it is easy to see that if $M \in GL_2(\mathbb Q)$, then \[M \cdot S = \left\{M \left[ {\begin{array}{*{20}c} a \\ b \\ \end{array} } \right] \ : \ (a,b) \in S\right\}\] gives rise to $\mathbb Q^2$ through iteration of $\mathcal{T}$ if and only if $S$ does. Suppose that $R = (a,b)$, $P = (c,d)$, and $T = (u,v)$ with $a,b,c,d,u,v \in \mathbb Q$. Since $RQ$ and $PT$ do not define the same line, it follows that $bu-av \neq 0$. Also, since $RQ$ and $PT$ are parallel, we have $bu-av = bc - ad$. Consider the following matrix: \begin{equation*} M = \frac{1} {{bu - av}}\left[ {\begin{array}{*{20}c} b & { - a} \\ { - v} & u \\ \end{array} } \right]. \end{equation*} A straightforward computation gives $M \cdot S = \left\{(0,0),(0,1),(1,0),\left(1,\frac{du-cv}{bu-av}\right)\right\}.$ Moreover, since $RP$ is not parallel to $QT$, it follows that $\frac{du-cv}{bu-av} \neq 1$. Next, set $\frac{r}{s} = \frac{du-cv}{bu-av}$ in which $r,s \in \mathbb Z$ and gcd$(r,s) = 1$. Suppose first that $r/s > 0$. By successively applying Lemma \ref{midpointlemma}, iteration of $\mathcal{T}$ on $M \cdot S$ produces the points: \begin{equation*} \left\{\left(0,\frac{l_1}{2^k}\right),\left(1,\frac{rl_2}{s2^k}\right) \ : \ l_1,l_2,k \in \mathbb N; \ 0 \leq l_1,l_2 \leq 2^k \right\}. \end{equation*} It follows that if we choose $k$ such that $2^{k-1} \geq \text{max}\{r,s\}$, we will have \begin{equation*} \left\{\left(0,0\right),\left(0,\frac{r}{2^k}\right),\left(0,\frac{2r}{2^k}\right), \left(1,0\right),\left(1,\frac{r}{2^k}\right),\left(1,\frac{2r}{2^k}\right)\right\} \subseteq \mathcal{T}^{k} \left( M \cdot S \right). \end{equation*} Therefore, letting $N = \left[ {\begin{array}{*{20}c} 1 & 0 \\ 0 & {\frac{{2^k }} {r}} \\ \end{array} } \right]$, we must have \begin{equation*} \left\{(0,0),(0,1),(0,2),(1,0),(1,1),(1,2)\right\} \subseteq N \cdot \mathcal{T}^{k} \left( M \cdot S \right). \end{equation*} An application of Lemma \ref{6pointlemma} now concludes the proof of this case. Finally, if $r/s < 0$, then the same examination as above reduces the situation to $S = \{(0,0),(0,1),(0,2),(1,0),(1,-1),(1,-2)\}$, also covered by Lemma \ref{6pointlemma}. \end{proof} \section{Acknowledgments} We would like to thank the anonymous referees for introducing us to the references \cite{barany, pach, otherproof, king} and for several suggestions that improved the exposition of this paper. Special thanks also go to Kelli Carlson for helping to simplify the proof of Lemma \ref{trianglelemma2} and for giving useful comments on preliminary versions of this paper.

https://arxiv.org/abs/1802.08934

The Archimedean limit of random sorting networks

A sorting network (also known as a reduced decomposition of the reverse permutation), is a shortest path from $12 \cdots n$ to $n \cdots 21$ in the Cayley graph of the symmetric group $S_n$ generated by adjacent transpositions. We prove that in a uniform random $n$-element sorting network $\sigma^n$, all particle trajectories are close to sine curves with high probability. We also find the weak limit of the time-$t$ permutation matrix measures of $\sigma^n$. As a corollary of these results, we show that if $S_n$ is embedded into $\mathbb{R}^n$ via the map $\tau \mapsto (\tau(1), \tau(2), \dots \tau(n))$, then with high probability, the path $\sigma^n$ is close to a great circle on a particular $(n-2)$-dimensional sphere in $\mathbb{R}^n$. These results prove conjectures of Angel, Holroyd, Romik, and Virag.

\section{Introduction} \label{S:intro} Let $\Gamma(S_n)$ be the Cayley graph of the symmetric group $S_n$ with generators given by the adjacent transpositions $\pi_i = (i, i + 1)$ for $i = 1, \mathellipsis, n -1$. A {\bf sorting network} is a shortest path from the identity permutation $\text{id}_n = 12 \cdots n$ to the reverse permutation $\text{rev}_n = n \cdots 21$ in $\Gamma(S_n)$. The length of any sorting network is $N = {n \choose 2}$. \medskip Sorting networks are known as {\bf reduced decompositions} of the reverse permutation, as any sorting network can equivalently be represented as a minimal length decomposition of the reverse permutation as a product of adjacent transpositions: $\text{rev}_n = \pi_{k_N} \mathellipsis \pi_{k_1}$. The combinatorics of sorting networks have been studied in detail under this name. There are connections between sorting networks and Schubert calculus, quasisymmetric functions, zonotopal tilings of polygons, and aspects of representation theory. For more background on these topics, see Stanley \cite{stanley1984number}; Bj\"orner and Brenti \cite{bjorner2006combinatorics}; Garsia \cite{garsia2002saga}; Tenner \cite{tenner2006reduced}; and Manivel \cite{manivel2001symmetric}. \begin{figure} \centering \includegraphics[scale=0.7]{Wiring-3.pdf} \caption{A ``wiring diagram" for a sorting network with $n = 4$. In this diagram, trajectories are drawn as continuous curves for clarity.} \label{fig:wiring} \end{figure} \medskip In computer science, sorting networks are viewed as $N$-step algorithms for sorting a list of $n$ numbers. At step $i$, the sorting network algorithm sorts the numbers at positions $k_i$ and $k_{i+1}$ into increasing order. This process sorts any list in $N$ steps. \medskip As in this algorithmic interpretation of sorting networks, we will think of the elements $\{1, \mathellipsis, n \}$ as particles being sorted in time (see Figure \ref{fig:wiring}). We use the notation $\sigma(x, t) = \pi_{k_\floor{t}} \mathellipsis \pi_{k_2}\pi_{k_1}(x)$ for the position of particle $x$ at time $t$. We call $(k_1, \dots, k_N)$ the \textbf{swap sequence} for $\sigma$. \subsection{Main limit theorems} \label{SS:main-thms} Angel, Holroyd, Romik, and Vir\'ag \cite{angel2007random} initiated the study of uniform random $n$-element sorting networks. They studied the limiting behaviour of the space-time swap distribution, rescaled particle trajectories, time-$t$ permutation matrices, and the Cayley graph path itself. \medskip They proved a law of large numbers for the space-time swap distribution, and based on strong numerical evidence, made conjectures about the limiting behaviour of the other three objects. In this paper, we prove these conjectures. \bigskip \noindent {\bf I. Rescaled particle trajectories.} \qquad For a sorting network $\sigma$, define the global trajectory $$ \sigma_G(x, t) = \frac{2\sigma(x, Nt)}n - 1. $$ The function $\sigma_G(x, \cdot)$ is the trajectory of particle $x$, with time rescaled so that the sorting process finishes at time $1$, and space rescaled so that the trajectory stays in the interval $[-1, 1]$. In \cite{angel2007random}, the authors conjectured that with high probability, all particle trajectories in a uniform random sorting network are close to sine curves (see Figure \ref{fig:sinecurves}). \medskip They proved that limiting trajectories are $1/2$-H\"older with H\"older constant $\sqrt{8}$. Dauvergne and Vir\'ag \cite{dauvergne1} improved upon this by showing that limiting trajectories are $\pi$-Lipschitz, and that trajectories of particles starting within $\epsilon$ of the edge are distance $2 \sqrt{\epsilon}$ away from a sine curve in uniform norm. \medskip Our first theorem proves the sine curve conjecture from \cite{angel2007random}. Here and throughout the paper we use the notation $\sigma^n$ for a uniform random $n$-element sorting network. \begin{customthm}{1}[Sine curve limit] \label{T:sine-curves} For each $n$ there exist random variables $\{(A^n_i, \Theta^n_i) \in [0, 1] \times [0, 2\pi]\}_{i \in \{1, \mathellipsis, n\}}$ such that for any $\epsilon > 0$, we have that $$ \mathbb{P} \left( \max_{i \in [1, n] } \max_{t \in [0, 1]} \card{ \sigma^n_G(i, t) - A_i^n \sin(\pi t + \Theta_i^n) } > \epsilon \right) \to 0 \qquad \;\text{as}\; \;\; n \to \infty. $$ \end{customthm} \begin{figure} \centering \includegraphics[scale=0.8]{RSN-sinecurves.pdf} \caption{This is a diagram of selected particle trajectories in a 2000 element sorting network. This image is taken from \cite{angel2007random}.} \label{fig:sinecurves} \end{figure} \bigskip \noindent {\bf II. Permutation Matrices.} \qquad For a uniform $n$-element sorting network $\sigma^n$, define $$ \rho^n_t= \frac{1}n \sum_{i=1}^n \delta \left( \sigma^n_G(i, 0), \sigma^n_G(i, t) \right). $$ Here $\delta(x, y)$ is a $\delta$-mass at the point $(x, y)$. The measure $\rho^n_t$ rescales the time-$t$ permutation matrix of $\sigma^n$, placing atoms of weight $1/n$ at the positions of the $1$'s. Define the {\bf Archimedean measure} $\mathfrak{Arch}$ on the square $[-1, 1]^2$ to be the measure with Lebesgue density $$ f(x, y) = \frac{1}{2\pi\sqrt{1 - x^2 - y^2}} $$ on the unit ball $B(0, 1)$, and $0$ elsewhere. This is simply the projected surface area measure of the $2$-sphere. Define the {\bf time-$t$ Archimedean measure} by $$ (X, X \cos (\pi t) + Y \sin (\pi t)) \stackrel{d}{=} \mathfrak{Arch}_{t}, \qquad \text{ where } \;\; (X, Y) \stackrel{d}{=} \mathfrak{Arch}. $$ Note that $\mathfrak{Arch} = \mathfrak{Arch}_{1/2}$. In \cite{angel2007random}, the authors conjectured that $\rho^n_t$ converges weakly to $\mathfrak{Arch}_{t}$ (see Figure \ref{fig:circles}). They proved that for any $t$, the support of $\rho^n_t$ lies in a particular octagon with high probability. In \cite{dauvergne1}, the authors showed that the support of any limit of $\rho^n_t$ lies within the elliptical support of $\mathfrak{Arch}_t$. \begin{figure} \centering \includegraphics[scale=0.9]{RSN-circles.pdf} \caption{This is a diagram of the measures $\{\rho^n_t : t \in \{0, 1/10, 2/10, \mathellipsis 1\} \}$ in a $500$-element sorting network. The blue octagons are the octagonal support bounds proved in \cite{angel2007random}, and the red ellipses are the support bounds proved in \cite{dauvergne1}. This figure is from \cite{angel2007random}.} \label{fig:circles} \end{figure} \medskip Our second theorem proves the weak convergence of the random measures $\rho^n_t$. We also show that the support of $\rho^n_t$ and $\mathfrak{Arch}_t$ are close. To state this theorem, recall that the Hausdorff distance between two sets $A, B \subset \mathbb{R}^2$ is $$ d_H(A, B) = \max \left\{\sup_{a \in A} \inf_{b \in B} d(a, b), \;\sup_{b \in B} \inf_{a \in A} d(a, b) \right\}. $$ \begin{customthm}{2}[Permutation matrix limit] \label{T:matrices} For any $t \in [0, 1]$, $ \rho^n_t \to \mathfrak{Arch}_t $ in probability as $n \to \infty$. That is, for any weakly open set $U$ in the space of probability measures on $[-1, 1]^2$ containing $\mathfrak{Arch}_t$, we have that $$ \mathbb{P}(\rho^n_t \in U) \to 1 \qquad \;\text{as}\; \quad n \to \infty. $$ Moreover, $$ d_H(\text{supp}(\rho^n_t), \text{supp}(\mathfrak{Arch}_t)) \to 0 \qquad \text{ in probability } \;\text{as}\; n \to \infty. $$ \end{customthm} Angel, Holroyd, Romik, and Vir\'ag \cite{angel2007random} also considered a permutation matrix evolution for $\sigma^n$. Let $j \in \{1, \mathellipsis, n\}$, and consider the random complex-valued function $$ Z^n_j(t) = e^{\pi i t} \left[\sigma^n_G(j, t) + i\sigma^n_G(j, t + 1/2) \right], \qquad t \in [0, 1/2]. $$ For a fixed $t$, $(Z^n_1(t), \mathellipsis, Z^n_n(t))$ is the set of points in the scaled permutation matrix for \[ \sigma^n(\cdot, t + 1/2)(\sigma^n)^{-1}(\cdot, t) \] after a counterclockwise rotation by $\pi t$ (see Figure \ref{fig:window}). Theorem \ref{T:sine-curves} guarantees that each of the paths $Z^n_j$ localize. \begin{customthm}{3}[Path Localization] \label{T:unif-rotation} Let $\sigma^n$ be a uniform random $n$-element sorting network. Then $$ \max_{j \in [1, n]} \max_{s, t \in [0, 1]} |Z^n_j(t) - Z^n_j(s)| \to 0 \qquad \text{in probability as } \;\; n \to \infty. $$ \end{customthm} \bigskip \noindent {\bf III. Great Circles.} \qquad We can embed the vertices of $\Gamma(S_n)$ into $\mathbb{R}^n$ by sending the permutation $\tau \in S_n$ to the point $\close{\tau} = (\tau(1), \mathellipsis, \tau(n)) \in \mathbb{R}^n$. For any $\tau \in S_n$, the point $\close{\tau}$ lies on the $(n-2)$-sphere $\mathbb{S}^{n-2} = \mathbb{L}_n \cap \mathbb{K}_n$, where \begin{align*} \mathbb{L}_n =& \left\{ (x_1, \mathellipsis x_n) \in \mathbb{R}^n : \sum_{i=1}^n x_i = \frac{n(n+1)}2 \right\} \qquad \;\text{and}\; \\ \mathbb{K}_n =& \left\{ (x_1, \mathellipsis x_n) \in \mathbb{R}^n : \sum_{i=1}^n x_i^2 = \frac{n(n+1)(2n+1)}6\right\}. \end{align*} If we also embed the edges of $\Gamma(S_n)$ as straight lines between the embedded vertices, we get an object called the {\bf permutahedron} (see Figure \ref{fig:permutahedron}). \medskip \begin{figure} \centering \includegraphics[scale=0.3]{Permutahedron.png} \caption{The permutahedron for $S_4$.} \label{fig:permutahedron} \end{figure} In \cite{angel2007random}, the authors conjectured that a uniform random sorting network is close to a great circle in $\mathbb{S}^{n-2}$ under this embedding. They showed that this conjecture implies the other global limiting results for uniform random sorting networks. Our strongest theorem proves this conjecture. \medskip For two functions $f, g: [0, 1] \to \mathbb{R}^n$, define the distance function $$ d_\infty(f, g) = \sup_{t \in [0, 1]} ||f(t) - g(t)||_\infty. $$ This is the uniform norm on functions, where the pointwise distance is the $L^\infty$ distance. Also, for an $n$-element sorting network $\sigma$, we define the embedded and time-rescaled path $ \bar{\sigma}:[0, 1] \to \mathbb{S}^{n-2} $ by letting $$ \bar{\sigma}(t) = (\sigma(1, Nt), \sigma(2, Nt), \dots, \sigma(n, Nt)). $$ \begin{customthm}{4}[Great circles] \label{T:geom-limit} Let $\bar{\sigma}^n$ be the embedding of $\sigma^n$ into $\mathbb{S}^{n-2}$. For every $n$, there exists a random path $C_n:[0, 1] \to \mathbb{S}^{n-2}$ such that $C_n$ is a constant-velocity parametrization of an arc of a great circle in $\mathbb{S}^{n-2}$ starting at $(1, \mathellipsis, n)$ and finishing at $(n, \mathellipsis, 1)$, and such that $$ \frac{d_\infty(\bar{\sigma}^n, C_n)}{n} \to 0 \qquad \text{in probability as} \;\; n \to \infty. $$ \end{customthm} Note that it is easy to find sorting networks that aren't close to a great circle in $\mathbb{S}^{n-2}$. For example, the ``bubble sort" sorting network given by the swap sequence $$ (1, 2, \mathellipsis n-1, 1, 2, \mathellipsis, n-2, \mathellipsis, 1, 2, 1) $$ is $d_\infty$-distance $n - 1 - o(1)$ from any great circle. \subsection{Random trajectory limits} To prove the main theorems of Section \ref{SS:main-thms}, we first analyze the limit of a random particle trajectory. This approach was first considered by Rahman, Vir\'ag, and Vizer \cite{rahman2016geometry}, and continued in \cite{dauvergne1}. \medskip Let $\mathcal{D}$ be the closure in the uniform norm $||\cdot||_u$ of the space of all possible sorting network trajectories $\sigma_G(x, \cdot):[0, 1] \to [-1, 1]$. The space $(\mathcal{D}, ||\cdot||_u)$ is a complete separable metric space. The only functions in $\mathcal{D}$ are continuous functions and the sorting network trajectories themselves. \medskip Let $Y_n \in \mathcal{D}$ be a uniform $n$-element sorting network trajectory. That is, if $\sigma^n$ is a uniform $n$-element sorting network, and $I_n$ is an independent uniform random variable on $\{1, \mathellipsis n\}$, then $$ Y_n = \sigma^n_G(I_n, \cdot). $$ We refer to $Y_n$ as the {\bf trajectory random variable} of $\sigma^n$. In \cite{dauvergne1}, Dauvergne and Vir\'ag proved that the sequence $\{Y_n\}_{n \in \mathbb{N}}$ is precompact in distribution, and that any subsequential limit is almost surely Lipschitz (we state their results precisely in Section \ref{S:local}). In this paper, we show that $\{Y_n\}_{n \in \mathbb{N}}$ converges in distribution, and identify its limit. \begin{customthm}{5}[The weak trajectory limit] \label{T:weak-limit} Let $(X, Z) \sim \mathfrak{Arch}$, and define the Archimedean path $\mathcal{A} \in \mathcal{D}$ by $\mathcal{A}(t) = X\cos(\pi t) + Z\sin(\pi t)$. Then $$ Y_n \stackrel{d}{\to} \mathcal{A} \qquad \;\text{as}\; \; n \to \infty. $$ \end{customthm} Theorem \ref{T:weak-limit} will be used in the proof of all our main theorems from Section \ref{SS:main-thms}. Most of the paper is devoted to its proof. We note here that we can equivalently write $$ \mathcal{A}(t) = \sqrt{1 - V^2}\sin(\pi t + 2\pi U), $$ where $V$ and $U$ are independent uniform random variables on $[0, 1]$. \bigskip \noindent \textbf{Random $m$-out-of-$n$ sorting networks.} \qquad We will also use Theorem \ref{T:weak-limit} to identify the limit of random $m$-out-of-$n$ subnetworks. This answers a question of Angel and Holroyd \cite{angel2010random}. This limit can also be by found by using the stronger great circle theorem (Theorem \ref{T:geom-limit}), as was noted in \cite{angel2010random}. \medskip Let $\sigma$ be an $n$-element sorting network. For $A \subset \{1, \dots, n\}$, let $\sigma|_A$ be the $|A|$-element sorting network given by restricting $\sigma$ to the set $A$. Specifically, for $i \in \{1, \dots, N\}$ define $\sigma^*_A(\cdot, i)$ be the relative ordering of the particles in $A$ in the permutation $\sigma(\cdot, i)$. This gives a sequence of $N$ permutations $\{\sigma^*_A(\cdot, i) \in S_{|A|} : i \in \{1, \dots N\} \}$. Removing duplicates gives the permutation sequence for sorting network $\sigma|_A$. For $m < n$, let the \textbf{random $m$-out-of-$n$ subnetwork $\tau^n_m$} be the restriction of $\sigma^n$ to a uniform $m$-element subset of $\{1, \mathellipsis n\}$, chosen independently from $\sigma^n$. \medskip Let $\{x_1, \mathellipsis x_n\}$ be a set of points in $\mathbb{R}^2$ in general position, and such that no two pairs of points determine parallel lines. Label the points in order of increasing $x$-coordinate. For all but finitely many angles $\theta$, listing the labels of the points $\{x_1, x_2, \mathellipsis x_n\}$ in increasing order of their horizontal projections after rotation by $\pi \theta$ gives a permutation $\tau_\theta$. The \textbf{geometric sorting network} associated to $\{x_1, \mathellipsis x_n\}$ is simply the sequence of permutations $\{\tau_\theta, \theta \in [0, 1] \}$ listed in order of increasing $\theta$. \begin{customthm}{6}[The subnetwork limit] \label{T:subnetwork} Let $\{X_1, X_2, \mathellipsis, X_m \}$ be random points in the unit ball $B(0, 1)$ sampled from the Archimedean distribution $\mathfrak{Arch}$, and let $\tau_m$ be the associated geometric sorting network. Then $$ \tau^n_m \stackrel{d}{\to} \tau_m \qquad \;\text{as}\; \; n \to \infty. $$ \end{customthm} \subsection{The local speed distribution} As a by-product of the proof of Theorem \ref{T:weak-limit}, we find the distribution of speeds in the local limit of random sorting networks. To state this result, we first give an informal description of this limit (a precise description is given in Section \ref{S:local}). The existence of this limit was established independently by Angel, Dauvergne, Holroyd, and Vir\'ag \cite{angel2017local}, and by Gorin and Rahman \cite{gorin2017}. Define $$ U_n (x, t) = \sigma^n(\floor{n/2} + x, nt) - \floor{n/2}. $$ Each path $U_n(x, \cdot)$ is a locally scaled particle trajectory. With an appropriate notion of convergence, we have that $$ U_n \stackrel{d}{\to} U, $$ where $U$ is a random function from $\mathds{Z} \times [0, \infty) \to \mathds{Z}$. $U$ is the local limit at the centre of the sorting network. We can also take a local limit centred at particle $\floor{\alpha n}$ for any $\alpha \in (0, 1)$. The result is the process $U$ with time rescaled by a semicircle factor $2\sqrt{\alpha(1 - \alpha)}$. \medskip In \cite{dauvergne1}, Dauvergne and Vir\'ag proved that particles in $U$ have asymptotic speeds. Specifically, they showed that for every $x \in \mathds{Z}$, the limit $$ S(x) = \lim_{t \to \infty} \frac{U(x, t) - U(x, 0)}{t} \qquad \text{exists }\text{almost surely}. $$ They showed that $S(x)$ has distribution $\mu$ independent of $x$. In this paper, we identify $\mu$. \begin{customthm}{7} \label{T:main-2} The measure $\mu$ is the arcsine distribution on $[-\pi, \pi]$ given by the Lebesgue density $$ f(x) = \frac{1}{\pi\sqrt{\pi^2 - x^2}}. $$ \end{customthm} \subsection*{Related work} In addition to the papers mentioned above, the local behaviour of random sorting networks has also been studied by Angel, Gorin, and Holroyd \cite{angel2012pattern}. The frequency of particular substrings in the swap sequence of a random sorting network has been analyzed by Reiner \cite{reiner2005note} and Tenner \cite{tenner2014expected}. Fulman and Stein \cite{fulman2014stein} have analyzed the distribution of the first swap in a random sorting network. \medskip All of this work relies on a bijection of Edelman and Greene \cite{edelman1987balanced} between sorting networks and Young tableaux of a particular shape. Little \cite{little2003combinatorial} found another bijection between these two sets, and Hamaker and Young \cite{HY} proved that these bijections coincide. \medskip Problems involving limits of sorting networks under different measures have been considered by Angel, Holroyd, and Romik \cite{angel2009oriented}, and also by Young \cite{young2014markov}. Uniform ``relaxed" random sorting networks have been analyzed by Kotowski and Vir\'ag \cite{kotowski2016limits} (see also \cite{rahman2016geometry}). \subsection*{Outline of the paper} Most of the paper is devoted to proving Theorem \ref{T:weak-limit}. The key to proving this theorem is the notion of particle flux across a curve $h$. Heuristically, particle flux measures the number of times that particles cross $h$ in a large-$n$ sorting network. It is defined in terms of the local speed distribution $\mu$. \medskip Particle flux is a useful quantity because any limit of $Y_n$ must minimize this quantity among curves $h$ with $h(0) = -h(1)$. This is due to the fact that in any sorting network, every pair of particles swaps \textit{exactly} once, whereas any particle must cross a line with $h(0) = -h(1)$ \textit{at least} once. \medskip In Section \ref{S:local}, we introduce the necessary background for the paper. In Section \ref{S:flux-intro}, we formally define particle flux and establish its properties. In Section \ref{S:unique}, we then use these properties to partially characterize minimal flux paths $h$ with $h(0) = -h(1)$. In Section \ref{S:path-speed}, we use this characterization to relate the local speed distribution $\mu$ to the derivative distribution for subsequential limits of $Y_n$. \medskip This allows us to derive an integral transform formula for $\mu$, which we do in Section \ref{S:integral-formula}. In Section \ref{S:transform}, we invert this integral transform in order to find $\mu$. This allows us to immediately prove Theorem \ref{T:weak-limit}. In this section, we also prove a slightly stronger version of Theorem \ref{T:weak-limit}, allowing us to establish Theorem \ref{T:subnetwork}. \medskip In the remainder of the paper, we use Theorem \ref{T:weak-limit} to establish our stronger limit theorems. In Section \ref{S:strong-limit}, we combine Theorem \ref{T:weak-limit} with bounds from \cite{angel2007random} to prove Theorems \ref{T:sine-curves}, \ref{T:matrices}, and \ref{T:unif-rotation}. Finally, in Section \ref{S:geom-limit}, we combine Theorem \ref{T:weak-limit} with Theorem \ref{T:sine-curves} to prove Theorem \ref{T:geom-limit}. \section{Preliminaries} \label{S:local} In this section, we introduce the necessary background about sorting networks. The most basic observation about uniform $n$-element sorting networks is that they exhibit a type of time-stationarity. This was first observed in \cite{angel2007random}. \begin{theorem} \label{T:time-stat} Let $(K^n_1, K^n_2, \mathellipsis K^n_N)$ be the swap sequence of a random $n$-element sorting network $\sigma^n$. Then $$ \left(K^n_1, K^n_2, \mathellipsis K^n_N\right) \stackrel{d}{=} \left(n - K^n_N, K^n_1, K^n_2, \mathellipsis K^n_{N - 1}\right). $$ \end{theorem} We will repeatedly use time-stationarity of sorting networks to reduce proofs to statements about the beginning of a sorting network. Using the Edelman-Greene bijection between sorting networks and Young tableaux, Angel, Holroyd, Romik, and Vir\'ag \cite{angel2007random} also found an explicit formula for the distribution of $K^n_1$. \begin{theorem} \label{T:dist-k1} Let $K^n_1$ be the location of the first swap of $\sigma^n$. For any $i \in \{1, \dots, n-1\}$, we have that $$ \mathbb{P}(K^n_1 = i) = \frac{1}N \frac{[3\cdot5\cdot7\cdots(2i - 1)][3\cdot 5 \cdots (2(n-i) - 1)]}{[2\cdot4\cdot6\cdots(2i - 2)][2\cdot 4 \cdots (2(n-i) - 2)]} \le \frac{3}n. $$ Moreover, if $\{i_n\}_{n \in \mathbb{N}}$ is any sequence of integers such that $i_n/n\to (\alpha + 1)/2$ for some $\alpha \in (-1, 1)$, then $$ n\mathbb{P}(K^n_1 = i_n) \to \frac{4\sqrt{1 - \alpha^2}}{\pi} \qquad \;\text{as}\; n \to \infty. $$ \end{theorem} \bigskip \noindent \textbf{The local limit.} \qquad Define a {\bf swap function} as a map $V:\mathbb{Z} \times [0, \infty) \to \mathds{Z}$ with the following properties: \smallskip \begin{enumerate}[nosep,label=(\roman*)] \item For each $x$, $V(x,\cdot)$ is cadlag with nearest neighbour jumps. \item For each $t$, $V(\cdot,t)$ is a bijection from $\mathds{Z}$ to $\mathds{Z}$. \item Define $V^{-1}(x, t)$ by $V(V^{-1}(x, t),t) = x$. Then for each $x$, $V^{-1}(x, \cdot)$ is a cadlag path with nearest neighbour jumps. \item For any time $t \in (0, \infty)$ and any $x \in \mathds{Z}$, $$ \lim_{s \to t^-} V^{-1}(x, s) = V^{-1}(x + 1, t) \qquad \text{if and only if} \qquad \lim_{s \to t^-} V^{-1}(x +1, s) = V^{-1}(x, t). $$ \end{enumerate} We think of a swap function as a collection of particle trajectories $\{V(x, \cdot) : x \in \mathds{Z}\}$. Condition (iv) guarantees that the only way that a particle at position $x$ can move up at time $t$ is if the particle at position $x+1$ moves down. That is, particles move by swapping with their neighbours. \medskip Let $\mathcal{A}$ be the space of swap functions endowed with the following topology. A sequence of swap functions $V_n \to V$ if each of the cadlag paths $V_n(x, \cdot) \to V(x, \cdot)$ and $V^{-1}_n(x, \cdot) \to V^{-1}(x, \cdot)$. Convergence of cadlag paths is convergence in the Skorokhod topology. We refer to a random swap function as a \textbf{swap process}. \medskip For a swap function $V$ and a time $t \in (0, \infty)$, define $$ V(\cdot, t, s) = V(V^{-1}(t, \cdot), t + s). $$ The function $V(\cdot, t, s)$ is the increment of $V$ from time $t$ to time $t + s$. \medskip Now for $i \in \{1, \dots n\}$, define $$ r_n(i) = \sqrt{1 - (2i/n - 1)^2}, $$ and consider the shifted, time-scaled swap process $$ U_{n}^{i}(x, s) = \sigma^n \left(i + x, \frac{ns}{r_n(i)} \right) - i. $$ To ensure that $U_{n}^{i}$ fits the definition of a swap process, we can extend it to a random function from $\mathds{Z} \times [0, \infty) \to \mathds{Z}$ by letting $U_{n}^{i}$ be constant after time $\frac{n-1}{2r_n(i)}$, and with the convention that $U_{n}^{i}(x, s)= x$ whenever $x \notin \{1 -i, \mathellipsis n - i\}$. In the swap processes $U_{n}^{i}$, all particles are labelled by their initial positions. The following is shown in \cite{angel2017local}, and also essentially in \cite{gorin2017}. \begin{theorem} \label{T:local} There exists a swap process $U$ such that the following holds. For any $\alpha \in (-1, 1)$, and any sequence of integers $\{i_n\}_{n \in \mathbb{N}}$ such that $i_n/n \to (\alpha + 1)/2$, we have that $$ U_{n}^{i_n} \stackrel{d}{\to} U \qquad \;\text{as}\; \; n \to \infty. $$ The swap process $U$ has the following properties: \begin{enumerate}[nosep,label=(\roman*)] \item $U$ is stationary and mixing of all orders with respect to the spatial shift $\tau U(x, t) = U(x + 1, t) - 1$. \item $U$ has stationary increments in time: for any $t \ge 0$, the process $U(\cdot, t, s)_{s\ge 0}$ has the same law as $U(\cdot,s)_{s\geq 0}$. \item $U$ is symmetric: $U(\cdot, \cdot) \stackrel{d}{=} - \; U(- \; \cdot, \cdot)$. \item For any $t \in [0, \infty)$, $\mathbb{P}($There exists $x \in \mathds{Z}$ such that $U(x, t) \neq \lim_{s \to t^-} U(x, t)) = 0$. \end{enumerate} \smallskip Moreover, for any sequence of times $\{t_n : n \in \mathbb{N}\}$ such that $\frac{n-1}{2r_n(i_n)} - t_n \to \infty$ as $n \to \infty$, $$ U^{i_n}_n (\cdot, t_n, \cdot) \stackrel{d}{\to} U \qquad \;\text{as}\; n \to \infty. $$ \end{theorem} Now, for a swap function $V$, let $W(V, t)$ be the number of times that the particles at locations $0$ and $1$ swap in the interval $[0, t]$. That is, $$ W(V, t) = \card {\left\{s \in (0, t] : \lim_{r \to t^-} V^{-1}(0, r) = V^{-1}(1, s) \right\}}. $$ As a by-product of the proof of convergence of $U_n^{i_n}$ to $U$, the authors of \cite{angel2017local} also found the expectation of $W(U, t)$. \begin{theorem}[Proposition 7.10, \cite{angel2017local}] \label{T:swaps} Let $\alpha \in (-1, 1)$, and let $\{i_n\}_{n \in \mathbb{N}}$ be any sequence of integers converging to $(\alpha + 1)/2.$ Then for any $t \in [0, \infty)$, we have that $$ \mathbb{E} W(U_n^{i_n}, t) \to \mathbb{E} W(U, t) = \frac{4t}{\pi}. $$ \end{theorem} Now let $Q(V, t)$ be the number of swaps that particle $0$ makes by time $t$ in a swap function $V$. That is, $$ Q(V, t) = \card {\Big\{ s \in (0, t] : \lim_{r \to s^-} V(0, r) \neq V(0, s) \Big\}}. $$ Dauvergne and Vir\'ag \cite{dauvergne1} used a stationarity argument to prove an analogous result to Theorem \ref{T:swaps} for $Q(u, t)$. \begin{theorem}[Lemma 3.2, \cite{dauvergne1}] \label{T:swaps-2} For any $t \in [0, \infty)$, we have that $$ \mathbb{E} Q(U, t) = \frac{8t}{\pi}. $$ \end{theorem} The fact that $U$ is stationary in both time and space implies that the point process of swaps of a given particle $x$ in $U$ is also stationary. This realization combined with the ergodic theorem allows us to conclude that all particles in $U$ have asymptotic speeds. This observation was used in \cite{dauvergne1} to prove results about the relationship between the local and global limit. We use their results as a starting point in our proofs. \begin{theorem}[Theorem 1.7, \cite{dauvergne1}] \label{T:local-speeds} For every $x \in \mathds{Z}$, the limit $$ S(x) = \lim_{t \to \infty} \frac{U(x, t) - U(x, 0)}{t} \qquad \text{exists } \text{almost surely}. $$ $S(x)$ is a symmetric random variable with distribution $\mu$ independent of $x$. The support of $\mu$ is contained in the interval $[-\pi, \pi]$. \end{theorem} We refer to $\mu$ as the {\bf local speed distribution}. Dauvergne and Vir\'ag \cite{dauvergne1} also used Theorem \ref{T:local-speeds} to find limiting swap rates in $U$. Define \begin{equation*} D_\mu^+(c) = \int(y - c)^+d \mu(y), \;\;\;\; D_\mu^-(c) = \int(y - c)^-d \mu(y) \;\;\;\;\;\text{and}\; \;\;\;\; D_\mu(c) = \int|y - c|d \mu(y). \end{equation*} \begin{theorem}[Theorem 1.8, \cite{dauvergne1}] \label{T:particle-swaps} Let $S(0)$ be the asymptotic speed of particle $0$ in $U$. For any $x \in \mathds{Z}$, $$ \frac{Q(U, t)}t \to D_\mu(S(0)) \qquad \text{ almost surely and in } L^1. $$ In particular, the random variables $Q(U, t)/t$ are uniformly integrable and $\mathbb{E} D_\mu(S(0)) = 8/\pi$. \end{theorem} We also need an analogous result regarding crossings of lines in the local limit. Let $L(t) = ct + d$ be a line of constant slope $c$. For a swap function $V$, define $$ C^+(V, L, t) = \card{\Big\{x \in \mathds{Z}: V(x, 0) \le L(0), V(x, t) > L(t)\Big\}}. $$ $C^+(V, L, t)$ is the total number of particles that are below $L$ at time $0$ and above $L$ at time $t$. We symmetrically define $C^-(V, L, t)$ as the total number of particles that are above $L$ at time $0$ and below $L$ at time $t$, and let $C(V, L, t) = C^-(V, L, t) + C^+(V, L, t)$. \begin{theorem}[Theorem 5.7, \cite{dauvergne1}] \label{T:line-rate} Let $L(t) = ct + d$. Then almost surely and in $L^1$, we have that \begin{align*} \lim_{t \to \infty} \frac{C^+(U, L, t)}t = D_\mu^+(c), \qquad \lim_{t \to \infty} \frac{C^-(U, L, t)}t = D_\mu^-(c), \qquad \lim_{t \to \infty} \frac{C(U, L, t)}t = D_\mu(c). \end{align*} \end{theorem} We also record here a few basic facts about the functions $D_\mu$ and $D_\mu^+$ that will be used throughout the paper. These properties can be proven using basic facts about integrals, and the fact that $\mu$ is symmetric and supported in $[-\pi, \pi]$. \begin{lemma} \label{L:convex} \begin{enumerate}[nosep, label=(\roman*)] \item Both $D_\mu$ and $D_\mu^+$ are convex, $1$-Lipschitz functions. \item For all $x$, we have that $D_\mu^+(x) \le D_\mu(x) \le |x| \vee \pi$. \item Suppose that $L(t) = at + b$ is tangent to either $D_\mu$ or $D_\mu^+$. Then $b \in [0, \pi]$. \item $D_\mu$ is a symmetric function, and hence minimized at $0$. \item $D_\mu^+(\pi) = 0$, and $D_\mu(\pm \pi) = \pi$. \end{enumerate} \end{lemma} \bigskip \noindent \textbf{Subsequential limits of $Y_n$.} \qquad Recall that $Y_n$ is the trajectory random variable of $\sigma^n$. We record here the main result of \cite{dauvergne1} regarding subsequential limits of $Y_n$. Here and throughout the paper, the phrase ``subsequential limit of $Y_n$'' always refers to a subsequential limit of $Y_n$ in distribution. \medskip We say that a path $y \in \mathcal{D}$ is $g(y)$-Lipschitz if $y$ is absolutely continuous and if for almost every $t$, $|y'(t)| \le g(y(t))$. \begin{theorem}[Theorem 1.4, \cite{dauvergne1}] \label{T:bounded-speed} \begin{enumerate}[nosep, label=(\roman*)] \item The sequence $\{Y_n\}$ is precompact in distribution. \item Suppose that $Y$ is a subsequential limit of $Y_n$ (in distribution). Then $$ \mathbb{P}\bigg(Y \text{ is } \pi\sqrt{1-y^2}\text{-Lipschitz},\;Y(0) = -Y(1) \bigg) = 1. $$ Moreover, $Y(t)$ is uniformly distributed on $[-1, 1]$ for every $t$. \end{enumerate} \end{theorem} In addition, we observe here that any subsequential limit $Y$ of $Y_n$ inherits certain symmetries from $\sigma^n$. \begin{prop} \label{P:Y-symmetries} Let $Y$ be any subsequential limit of $Y_n$. \smallskip \begin{enumerate}[nosep, label=(\roman*)] \item Define $Y_t \in \mathcal{D}$ by $$ Y_t(s) = \begin{cases} \;\;Y(s + t), \quad &s \le t.\\ \;\; -Y(s + t - 1), \quad &s > t. \end{cases} $$ For any $t \in [0, 1]$, we have that $Y_t \stackrel{d}{=} Y$. \item $Y \stackrel{d}{=} - Y$. \item Define $Z \in \mathcal{D}$ by $Z(t) = Y(1 - t)$. Then $Z \stackrel{d}{=} Y$. \end{enumerate} \end{prop} \begin{proof} Property (i) follows from time stationarity of random sorting networks (Theorem \ref{T:time-stat}). Properties (ii) and (iii) follow from the corresponding properties of the swap sequence $(K^n_1, \dots, K^n_N)$ of $\sigma^n$: \[ \left(K^n_1, K^n_2, \mathellipsis K^n_N\right) \stackrel{d}{=} \left(n - K^n_1, n - K^n_2, \mathellipsis, n - K^n_{N}\right) \stackrel{d}{=} \left(K^n_N, \mathellipsis, K^n_2, K^n_1\right). \qquad \qedhere \] \end{proof} \section{Particle flux for Lipschitz paths} \label{S:flux-intro} In this section, we introduce particle flux and prove that it measures the amount of particles that cross a line in a typical sorting network. Let $\text{\fontfamily{ppl}\selectfont Lip}$ be the set of Lipschitz paths from $[0,1] \to [-1, 1]$ (we will use this notation throughout the paper). Define the \textbf{local speed} of a function $h$ at time $t$ by $$ s(t) = \frac{d(\arcsin(h))}{dt} = \begin{cases} \frac{h'(t)}{\sqrt{1 - h^2(t)}}, \quad &|h(t)| < 1 \\ 0, \quad &|h(t)| = 1. \end{cases} $$ When $h \in \text{\fontfamily{ppl}\selectfont Lip}$, the local speed $s(t)$ exists for almost every time $t$. Recalling the definition of $D_\mu$ from the previous section, we then define the {\bf particle flux} of $h$ over a set $A$ by \begin{equation} \label{E:path-flux} J(h; A) = \frac{1}2 \int_A D_\mu(s(t)) \sqrt{1-h^2(t)}dt. \end{equation} We define $J(h) = J(h; [0, 1])$. Note that $J(h) < \infty$ for any Lipschitz function $h$. This follows from Lemma \ref{L:convex} (ii), which implies that $$ D_\mu(s(t)) \sqrt{1-h^2(t)} \le [\pi \vee |s(t)|] \sqrt{1-h^2(t)} \le \pi \vee |h'(t)|. $$ We will also consider {\bf positive particle flux} $J^+(h; A)$ and {\bf negative particle flux} $J^-(h; A)$ defined by $$ J^+(h; A) = \frac{1}2 \int_A D^+_\mu(s(t)) \sqrt{1-h^2(t)}dt, \qquad J^-(h; A) = \frac{1}2 \int_A D^-_\mu(s(t)) \sqrt{1-h^2(t)}dt. $$ Again, we let $J^+(h) = J^+(h; [0, 1])$ and $J^-(h) = J^-(h; [0, 1])$. \medskip We now connect flux to random sorting networks. If a random sorting network resembles the local limit in a local window around the global space-time position $(t, h(t))$, then by Theorem \ref{T:line-rate}, the number of distinct particles that cross $h$ in this window should be proportional to $$ D_\mu\left(\frac{h'(t)}{\sqrt{1 -h^2(t)}}\right) \sqrt{1-h^2(t)}. $$ The scaling factors of $\sqrt{1-h^2}$ come from the semicircle rescaling of time in the local limit away from the center. \medskip Therefore in a typical large-$n$ sorting network, where most local windows resemble the local limit, $J(h)$ should be proportional to the number of particles that cross the line $h$, counting multiple crossings for a given particle if and only if the crossings happen at globally distinguishable locations. \medskip The factor of $1/2$ is to account for the difference between the ${n \choose 2}^{-1}$ scaling in the global limit and the $n^{-1}$ scaling in the local limit. Similarly, $J^+(h)$ and $J^-(h)$ should be proportional to the number of upcrossings and downcrossings of the line $h$ in a large $n$ sorting network. \medskip Now let $\text{\fontfamily{ppl}\selectfont Lip}_r$ be the set of paths $h \in \text{\fontfamily{ppl}\selectfont Lip}$ with $h(0) = -h(1)$. If $h \in \text{\fontfamily{ppl}\selectfont Lip}_{r}$, then in any sorting network, every particle must cross $h$ at least once unless the particle starts at $h(0)$. Therefore $J(h)$ should be bounded below for such paths. Since any two particles in a sorting network cross each other exactly once, $J(h)$ should achieve this lower bound when $h$ is a trajectory limit. \medskip The next theorem makes rigorous this intuition behind the definition of particle flux. To state the theorem, for a random variable $Y \in \mathcal{D}$ and a path $h \in \text{\fontfamily{ppl}\selectfont Lip}$, we define $$ P^+_Y(h; [a, b]) = \mathbb{P} \Big( \exists s < t \in [a, b] \text{ such that } Y(s) < h(s) \;\text{and}\; Y(t) > h(t) \Big). $$ In other words, $P^+_Y(h; [a, b])$ is the probability that $Y$ upcrosses $h$ in the interval $[a, b]$. We similarly define the downcrossing probability $$ P^-_Y(h; [a, b]) = \mathbb{P} \Big( \exists s < t \in [a, b] \text{ such that } Y(s) > h(s) \;\text{and}\; Y(t) < h(t) \Big). $$ \begin{theorem} \label{T:particle-crossings} Let $Y$ be any subsequential limit of $Y_n$. \begin{enumerate}[label=(\roman*)] \item Let $h \in \text{\fontfamily{ppl}\selectfont Lip}$ and $[a, b] \subset [0, 1]$. Then $P^+_Y(h; [a, b]) \le J^+(h; \; [a, b])$ and $P^-_Y(h; [a, b]) \le J^-(h; [a, b])$. \item Let $h \in \text{\fontfamily{ppl}\selectfont Lip}_r$. Then $J(h) \ge 1$. \item $\mathbb{P}(J(Y) = 1) = 1$. \end{enumerate} \end{theorem} To prove this theorem, we first need two key lemmas about convergence to the local limit. Recall that $C^+(V, L, t)$ is the number of upcrossings of a line $L(s) = cs + d$ in the interval $[0, t]$ in a swap function $V$. Recall also the definition of $U_n^{i}$ from Section \ref{S:local}.\begin{lemma} \label{L:L1-swap-rate} Let $\alpha \in (-1, 1)$, and suppose that $\{i_n\}_{n \in \mathbb{N}}$ is a sequence of integers such that $i_n/n \to (1 + \alpha)/2$. Let $X$ be a uniform random variable on $[0, 1]$ that is independent of $U, U_n^{i_n}$, let $\{c_n\}_{n \in \mathbb{N}}$ be a sequence of real numbers converging to $c \in \mathbb{R}$, and let $\{d_n\}_{n \in \mathbb{N}}$ be a sequence of real numbers in $[0, 1]$. Define $$ L_n(s) = c_ns + d_n + X, \quad \text{and} \quad L(s) = cs + X. $$ Then for any time $t \in (0, \infty)$, we have that \begin{enumerate}[label=(\roman*)] \item $C^+(U_n^{i_n}, L_n, t) \stackrel{d}{\to} C^+(U, L, t)$ as $n \to \infty$. \item $\mathbb{E} C^+(U_n^{i_n}, L_n, t) \to \mathbb{E} C^+(U, L, t)$ as $n \to \infty$. \item $\mathbb{E} C^+(U_n^{i_n}, L_n, t) \le 3t + |c_n t| + 2$ for all $n$. \end{enumerate} \end{lemma} \begin{proof} First note that $d_n + X \Mod 1 \stackrel{d}{=} X$ for all $n$. Therefore by the stationarity of $U$ with respect to integer-valued spatial shifts (Theorem \ref{T:local} (i)), we have that \begin{equation} \label{E:dn-shift} C^+(U^{i_n}_n, L_n, t) \stackrel{d}{\to} C^+(U, L, t) \quad \text{if and only if} \quad C^+(U^{i_n}_n, c_ns + X, t) \stackrel{d}{\to} C^+(U, L, t). \end{equation} Now if $V_n \in \mathcal{A}$ is a sequence of swap functions converging to swap function $V$, then $$ C^+(V_n, c_ns + X, t) \to C^+(V, L, t) $$ unless $V$ either has a swap at time $t$, or $ct + X \in \mathds{Z}$. By Theorem \ref{T:local} (iv), for any time $t$, the event where $U$ has a swap at time $t$ has probability $0$. Moreover, the probability that $ct + X \in \mathds{Z}$ is also $0$. Therefore $$ C^+(U^{i_n}_n, c_ns + X, t) \stackrel{d}{\to} C^+(U, L, t), $$ and hence (i) follows by statement \eqref{E:dn-shift}. Now recall that $W(V, t)$ is the number of swaps at location $0$ in a swap function $V$ in the interval $[0, t]$. For any swap function $V$ and any line $H(s) = as+b$, we have that \begin{equation} \label{E:C-W} C^+(V, H, t) \le W(V, t) + |a t| + 2. \end{equation} To see why this is true, observe that every particle $x$ with $x \le 0$ and $V(x, t) > 1$ must move from position $0$ to position $1$ at least once in the interval $[0, t]$, therefore contributing to $W(V, t)$. Every particle $x$ that upcrosses $H$ in the interval $[0, t]$ fits this description, unless $$ at \le V(x, t) \le 1. $$ There are at most $|a t| + 2$ such values of $x$, proving \eqref{E:C-W}. \medskip Now again since $U$ has no swap at time $t$ almost surely (Theorem \ref{T:local} (iv)), $W(U_n^{i_n}, t) \stackrel{d}{\to} W(U, t)$. Also, by Theorem \ref{T:swaps}, we have that $$ \mathbb{E} W(U_n^{i_n}, t) \to \mathbb{E} W(U, t). $$ Hence, the random variables $W(U_n^{i_n}, t)$ are uniformly integrable (see Proposition 3.12, \cite{kallenberg2006foundations}). Therefore by \eqref{E:C-W} applied to the swap processes $U_n^{i_n}$ and the lines $L_n$, the random variables $C^+(U_n^{i_n}, L_n, t)$ are also uniformly integrable, and hence converge in expectation since they converge in distribution (again by Proposition 3.12, \cite{kallenberg2006foundations}). \medskip Finally, the bound on the probability distribution for the first swap location in a random sorting network (Theorem \ref{T:dist-k1}) and time stationarity (Theorem \ref{T:time-stat}) allows us to conclude that for any $n, i_n$ and $t$, that $$ \mathbb{E} W(U_n^{i_n}, t) \le 3t, $$ which in turn proves the bound on the expectation of $C^+(U_n^{i_n}, L_n, t)$ via \eqref{E:C-W}. \end{proof} This next lemma is similar to Lemma \ref{L:L1-swap-rate}, but deals with particle speeds rather than upcrossing rates. For a swap function $V$ we define $$ S(V, t) = \frac{V(0, t)}{t}. $$ This is the average speed of particle $0$ in the interval $[0, t]$. \begin{lemma} \label{L:L1-flux} Let $\{U_n^{i} : i \in \{i, \dots n\}, n \in \mathbb{N}\}$ be the array of locally scaled random sorting networks defined in Section \ref{S:local}, and let $U$ be the local limit. Let $I$ be a uniform random variable on $(-1, 1)$, independent of everything else. For each $n$, define $I_n = \ceil{n(I + 1)/2}$, let $$ R_n = \sqrt{1 - \left[\frac{2I_n}n - 1\right]^2}, \qquad \;\text{and}\; \qquad R = \sqrt{1 - I^2}. $$ Then the following statements hold. \begin{enumerate}[label=(\roman*)] \item For any $t \in (0, \infty)$, we have that $R_nD_\mu(S(U_n^{I_n}, R_nt)) \xrightarrow[n \to \infty]{d} RD_\mu(S(U, R t))$. \item For any $t \in (0, \infty)$, we have that $\mathbb{E} R_n D_\mu(S(U_n^{I_n}, R_nt)) \xrightarrow[n \to \infty]{} \mathbb{E} R D_\mu \left(S(U, R t) \right)$. \item $\mathbb{E} R D_\mu \left(S(U, R t) \right) \to 2$ as $t \to \infty$. \end{enumerate} \end{lemma} \begin{proof} Fix $t \in (0, \infty)$. If $V_n$ is a sequence of swap functions converging to a swap function $V$ with no swap at time $t$, then $D_\mu(S(V_n, t)) \to D_\mu(S(V, t))$. Now condition on $I$. This fixes $R_n, I_n,$ and $R$. For any fixed time $t \in (0, \infty)$, the process $U$ has no swap at time $t$ almost surely (Theorem \ref{T:local} (iv)). Therefore for any bounded continuous test function $f$, we have that $$ \mathbb{E} \bigg[f\Big(R_nD_\mu(S(U_n^{I_n}, R_nt))\Big) \; \Big | \; I \bigg] \to \mathbb{E} \bigg[f\Big(RD_\mu(S(U,Rt))\Big) \; \Big| \; I \bigg]. $$ Taking expectations proves the distributional convergence in (i). Now, by Lemma \ref{L:convex}, we have that $D_\mu(s) \le |s| + \pi$ for any $s \in \mathbb{R}$. Recalling that $Q(V, t)$ is the number of swaps made by particle $0$ in the interval $[0, t]$ in a swap process $V$, this implies that \begin{equation} \label{E:Q-bd} R_n\left(D_\mu(S_{I_n}(U_n^{I_n}, R_nt))\right)\le R_n(|S_{I_n}(U_n^{I_n}, R_nt)| + \pi) \le \frac{Q(U_n^{I_n}, R_nt)}{t} + \pi. \end{equation} Now, we similarly have that $Q(U_n^{I_n}, R_nt) \stackrel{d}{\to} Q(U, Rt)$, again since $U$ has no swap at time $t$ almost surely. Moreover, $$ \mathbb{E} Q(U_n^{I_n}, R_nt) = \frac{2\floor{nt}}n, $$ since this expectation is simply the expected number of swaps made by a random particle in a sorting network after $2nt$ steps. By Theorem \ref{T:swaps-2}, we also have that $$ \mathbb{E} Q(U, Rt) = \frac{4t}{\pi} \int_{-1}^1 \sqrt{1 - x^2}dx = 2t, $$ so $\mathbb{E} Q(U_n^{I_n}, R_nt) \to \mathbb{E} Q(U, Rt)$. Again, by Proposition 3.12 from \cite{kallenberg2006foundations}, this implies that the random variables $\{Q(U_n^{I_n}, R_nt)\}$ are uniformly integrable, and hence so are the random variables $R_n D_\mu(S(U_n^{I_n}, R_nt))$. Since these random variables converge in distribution to $R D_\mu(S(U, R t))$, they must also converge in expectation, proving (ii). \medskip Now we prove (iii). First, Theorem \ref{T:local-speeds} implies that \begin{equation} \label{E:D-mu-lim} R D_\mu(S(U, Rt)) \xrightarrow[t \to \infty]{} R D_\mu(S(0)) \quad \text{almost surely}, \end{equation} where $S(0)$ is the speed of particle $0$ in $U$. Analogously to \eqref{E:Q-bd}, we also have that $$ R D_\mu(S(U, R t)) \le \frac{Q(U, t)} {t} + \pi. $$ By Theorem \ref{T:particle-swaps}, the right hand side above is uniformly integrable, and hence so is the left hand side. Therefore since $RD_\mu(S(U, R t))$ has an almost sure limit by \eqref{E:D-mu-lim}, it also converges in expectation. Finally, Theorem \ref{T:particle-swaps} implies that \begin{equation*} \mathbb{E} R D_\mu(S(0)) = 2. \qedhere \end{equation*} \end{proof} To prove Theorem \ref{T:particle-crossings}, we also need two auxiliary results. The first is a simple lemma about uniform convergence of functions. This will be used in the proof of Theorem \ref{T:particle-crossings} (i). \begin{lemma} \label{L:mesh} Let $f_n:[0, 1] \to [-1, 1]$ be a sequence of functions converging uniformly to a continuous function $f$, and let $h:[0, 1] \to [-1, 1]$ be any continuous function. Let $$ \{\Pi_n = \{t_{n, 0} =0< t_{n, 1} < \dots < t_{n, m(n)}=1\}\}_{n \in \mathbb{N}} $$ be a sequence of partitions of $[0, 1]$ such that $$ \text{mesh}(\Pi_n) = \max_{i \in \{0, 1, \dots, m(n) - 1\}} |t_{n, i + 1} - t_{n, i}| \to 0 \qquad \;\text{as}\; n \to \infty. $$ Let $[a, b] \subset [0, 1]$. If there exist times $s < t \in [a, b]$ such that $f(s) < h(s)$ and $f(t) > h(t)$, then for all large enough $n$, there exists a time $t_{n, i} \in [a, b]$ such that $f_n(t_{n, i}) \le h(t_{n, i})$ and $f_n(t_{n, i + 1}) > h(t_{n, i + 1})$. \end{lemma} \begin{proof} By the continuity of $f$ and $h$, there exists an $\epsilon > 0$ and disjoint intervals $[s, s_+]$ and $[t_-, t]$ such that $f(r) < h(r) - \epsilon$ for all $r \in[s, s_+]$ and $f(r) > h(r) + \epsilon$ for all $r \in [t_-, t]$. Therefore by uniform convergence, for all large enough $n$ we have that $f_n(r) < h(r) - \epsilon/2$ for all $r \in[s, s_+]$ and $f(r) > h(r) + \epsilon/2$ for all $r \in [t_-, t]$. Now, for large enough $n$ we also have that $$ \text{mesh}(\Pi_n) < \min(s_+ - s, t - t_-). $$ Therefore for such $n$, there must exists $i_1 < i_2 \in \Pi_n \cap [a, b]$ such that $f_n(t_{n, i_1}) < h(t_{n, i_1})$ and $f_n(t_{n, i_2}) > h(t_{n, i_2})$. Thus for some $i \in \{i_1, \dots i_2 - 1\}$, we must have that $f_n(t_{n, i}) \le h(t_{n, i})$ and $f_n(t_{n, i + 1}) > h(t_{n, i + 1})$. \end{proof} To prove part (iii), we also need that $J(\cdot)$ is lower semicontinuous. \begin{prop} \label{P:lsc-E} Let $h_n \in \text{\fontfamily{ppl}\selectfont Lip}$ be a sequence converging uniformly to $h$. Then for any set $A \subset [0, 1]$, we have that $$ J(h; A) \le \liminf_{n \to \infty} J(h_n ; A). $$ \end{prop} \begin{proof} Note that $J(h; A)$ is of the form $\int_A G(h(t), h'(t))dt$, where $G$ is a continuous, positive function, such that for any fixed value of $a$, $G(a, \cdot)$ is convex. This follows from the convexity of $D_\mu$ (Lemma \ref{L:convex}). Functionals of this form are lower semicontinuous in the uniform norm by Theorem 1.6 from \cite{struwe1996variational}. \end{proof} \begin{proof}[Proof of Theorem \ref{T:particle-crossings}.] \noindent \textbf{Proof of (i):} \qquad Note first that by the symmetry of $Y$ (Proposition \ref{P:Y-symmetries}), that $$ \mathbb{P}^-_Y(h; [a, b]) = \mathbb{P}^+_{-Y}(-h; [a, b]) = \mathbb{P}^+_{Y}(-h; [a, b]) $$ for any path $h \in \text{\fontfamily{ppl}\selectfont Lip}$ and any interval $[a, b]$. Also, $J^+(-h) = J^-(h)$ by the symmetry of $\mu$ (Theorem \ref{T:local-speeds}). Therefore the assertion that $ \mathbb{P}^-_Y(h; [a, b]) \le J^-(h) $ is equivalent to the assertion that $\mathbb{P}^+_Y(-h; [a, b]) \le J^+(-h)$, and so to prove Theorem \ref{T:particle-crossings} (i), it suffices to prove that $\mathbb{P}^+_Y(h; [a, b]) \le J^+(h)$ for every path $h \in \text{\fontfamily{ppl}\selectfont Lip}$. \medskip We first prove this for $h \in \text{\fontfamily{ppl}\selectfont Lip}$ with range in the open interval $(-1, 1)$. Let $t \in \mathds{Z} \cap (0, \infty)$, and for $s \in [0, 1]$, define $$ s_{n, t} = \frac{2t}{n-1} \floor{\frac{(n-1)s}{2t}}, \quad \;\text{and}\; \quad s^+_{n, t} = \min\left(s_{n, t} + \frac{2t}{n-1}, 1\right). $$ Let $X$ be a uniform random variable on $[0, 1]$, independent of the sequence $\{Y_n\}$, and let $$ A^+_{n, t, s} = \left\{ Y_n(s_{n, t}) < h(s_{n, t}) + \frac{2X}n \;\; \;\text{and}\; \;\; Y_n(s^+_{n, t}) \ge h(s^+_{n, t}) + \frac{2X}n \right\}. $$ Now observe that when $s < b_{n, t}$, then $s_{n, t}^+ \le b_{n, t}$, and so $s^+_{n, t} = s_{n, t} + 2t/(n-1)$. Therefore for such $s$, time stationarity of random sorting networks (Theorem \ref{T:time-stat}) implies that $A^+_{n, t, s}$ has the same probability as the event $$ \left\{ Y_n(0) < h(s_{n, t}) + \frac{2X}n \;\; \;\text{and}\; \;\; Y_n\left(\frac{2t}{n-1}\right) \ge h(s^+_{n, t}) + \frac{2X}n \right\}. $$ Here have used that $t \in \mathds{Z}$ to apply time-stationarity. We can then express the upcrossing probability $\mathbb{P}(A^+_{n, t, s})$ in terms of the expected number of upcrossings in the local scaling. For $u \in [-1, 1]$, define $$ [u]_n = \floor{\frac{n(u + 1)}2}, \quad \;\; \{u\}_n = \frac{n(u + 1)}2 - \floor{\frac{n(u + 1)}2}, \quad \;\; g_n(u) = \sqrt{1 - \left(\frac{2[u]_n}n - 1\right)^2}. $$ Then for $s < b_{n, t}$, we have that $$ n \mathbb{P}(A^+_{n, t, s}) = \mathbb{E} C^+\left(U_n^{[h(s_{n, t})]_n}, L_{n, s}, g_n(h(s_{n, t})) t \right), $$ where $$ L_{n, s}(r) = \frac{n\left(h(s^+_{n, t}) - h(s_{n, t})\right)}{2tg_n(h(s_{n, t}))} r + \{h(s_{n, t})\}_n + X. $$ Also, let $$ L_{s}(r) = \frac{h'(s)}{\sqrt{1 -h^2(s)}}r + X. $$ Since $h$ is Lipschitz and hence differentiable almost everywhere, Lemma \ref{L:L1-swap-rate} (ii) implies that for almost every $s \in [0, b)$, that \begin{equation} \label{E:n-PA} \lim_{n \to \infty} n\mathbb{P}(A^+_{n, t, s}) = \lim_{n \to \infty} \mathbb{E} C^+\left(U_n^{[h(s_{n, t})]_n}, L_{n, s}, g_n(h(s_{n, t})) t \right) =\mathbb{E} C^+(U, L_s, \sqrt{1 - h^2(s)}t). \end{equation} Here we have used that the range of $h$ is in $(-1, 1)$ to ensure convergence. Moreover, there exists a constant $C$ such that \begin{equation} \label{E:finite-bd} n\mathbb{P}(A^+_{n, t, s}) \le Ct \end{equation} for all $n \in \mathbb{N}, s < b_{n, t}$. This follows from Lemma \ref{L:L1-swap-rate} (iii) and the fact that $h$ is Lipschitz. Now let $Z_{n, t}$ be the number of times $s$ of the form $s_{n, t}$ in the interval $[a_{n, t}, b_{n, t})$ such that the upcrossing event $A^+_{n, t, s}$ occurs. We have that $$ Z_{n, t} = \int_{a_{n, t}}^{b_{n, t}} \frac{(n-1)\mathbbm{1} (A^+_{n, t, s})}{2t} ds. $$ Therefore by the bounded convergence theorem, we have that \begin{equation} \label{E:cvg-ab-flux} \mathbb{E} Z_{n, t} = \int_{a_{n, t}}^{b_{n, t}} \frac{(n-1)\mathbb{P} (A^+_{n, t, s})}{2t} ds \xrightarrow[n \to \infty]{} \frac{1}{2} \int_{a}^b \frac{\mathbb{E} C^+(U, L_s, \sqrt{1 - h^2(s)}t)}t ds. \end{equation} Now, $\mathbb{E} Z_{n, t}$ bounds $\mathbb{P}(Z_{n, t} \ge 1)$. Therefore for any subsequential limit $Y$ of $Y_n$, Lemma \ref{L:mesh} applied to a subsequence $Y_{n_i} - 2X/n \to Y$ implies that $$ \lim_{n \to \infty} \mathbb{E} Z_{n, t} \ge \liminf_{i \to \infty} \mathbb{P}(Z_{n_i, t} \ge 1) \ge P^+_Y(h; [a, b]). $$ The integrand on the right hand side of \eqref{E:cvg-ab-flux} is bounded above by $C$ for all $t \in \mathds{Z} \cap (0, \infty)$ by \eqref{E:n-PA} and \eqref{E:finite-bd}. Therefore by Theorem \ref{T:line-rate} and the bounded convergence theorem, the right hand side of \eqref{E:cvg-ab-flux} converges to $$ \frac{1}{2} \int_{a}^b D^+_\mu\left(\frac{h'(s)}{\sqrt{1 - h^2(s)}}\right) \sqrt{1 - h^2(s)} ds = J^+(h; \; [a, b]) \qquad \;\text{as}\; \;\; t \to \infty. $$ This proves Theorem \ref{T:particle-crossings} (i) for $h \in \text{\fontfamily{ppl}\selectfont Lip}$ with range in $(-1, 1)$. Now we extend this to all $h \in \text{\fontfamily{ppl}\selectfont Lip}$. Define $h_m = h \vee(-1 + 1/m) \wedge (1 - 1/m)$, and let $$ A_m = \{t \in [0, 1] : h(t) \ge 1- 1/m \text{ or } h(t) \le -1 + 1/m\}. $$ Letting $\mathcal{L}$ be Lebesgue measure on $[0, 1]$, we have that $$ J^+(h_m ; A_m) = D^+_\mu(0)\sqrt{1 - (1- 1/m)^2}\mathcal{L}(A_m), \qquad \;\text{and}\; \qquad J^+(h_m ; A_m^c) = J(h ; A_m^c). $$ The flux $J^+(h_m ; A_m) \to 0$ as $m \to \infty$, so $$ \limsup_{m \to \infty} J^+(h_m; [a, b]) \le J^+(h; [a, b]). $$ Moreover, $h_m$ converges uniformly to $h$, so if $Y$ upcrosses $h$, then $Y$ will eventually upcross $h_m$. Therefore $$ \liminf_{m \to \infty} P^+_Y(h_m; [a, b]) \ge P^+_Y(h; [a, b]). $$ Putting these two inequalities together completes the proof. \medskip \noindent \textbf{Proof of (ii):} \qquad Let $h \in \text{\fontfamily{ppl}\selectfont Lip}_r$, and let $Y$ be a subsequential limit of $Y_n$. Therefore by Theorem \ref{T:particle-crossings} (i), \begin{align*} \mathbb{P} \Big(\exists r < t \in [0, 1] \text{ such that either } Y(r) < h(r), \;Y(t) > h(t),\text{ or } &Y(r) > h(r), \;Y(t) < h(t) \Big) \\ &\le J^+(h) + J^-(h) = J(h). \end{align*} The event above holds unless $Y(0) = h(0)$ since $Y(0) = - Y(1)$ almost surely by Theorem \ref{T:bounded-speed}. Since $Y(0)$ is uniformly distributed (Theorem \ref{T:bounded-speed}), $Y(0) \ne h(0)$ with probability one, so the left hand side above is equal to $1$. \medskip \noindent \textbf{Proof of (iii):} \qquad Let $Y$ be any subsequential limit of $Y_n$. Fix $t \in (0, \infty) \cap \mathds{Z}$, and define $$ t_{n, j} = \frac{2jt}{n-1} \quad \text{for } j \in \{0, \mathellipsis \floor{(n-1)/(2t)}\} \quad \text{ and let } t_{n, \floor{\frac{n-1}{2t}} + 1} = 1. $$ Define $Y_{n, t}$ so that $ Y_{n, t}\left(t_{n, j}\right) = Y_{n}\left(t_{n, j}\right)$ for $j \in \{0, \mathellipsis \floor{(n-1)/(2t)} + 1\}$, and so that $Y_{n, t}$ is linear at times in between. By time stationarity of random sorting networks (Theorem \ref{T:time-stat}), we can write \begin{equation} \label{E:JY-ineq} \mathbb{E} J(Y_{n, t}) = \sum_{j=0}^\floor{\frac{n-1}{2t}} \mathbb{E} J(Y_{n, t} ; \; [t_{n, j}, t_{n, j+1}] ) \le \left(\floor{\frac{n-1}{2t}} + 1\right) \mathbb{E} J(Y_{n, t} ; \; [0, t_{n, 1}]). \end{equation} Here we have used that $t \in \mathds{Z}$ to apply time stationarity of sorting networks. The final term in the sum above may be slightly smaller than the previous terms since the length of interval may be less than $2t/(n-1)$; this gives rise to the inequality. \medskip Now we have that \begin{align} \nonumber J(Y_{n, t} ; \; [0, t_{n, 1}]) &= \frac{1}{2} \int_0^{t_{n, 1}} \int \left|Y_{n, t}'(0)- \sqrt{1 - Y_{n, t}^2(r)}x\right|d\mu(x) dr \\ \label{E:frac-en} &= \frac{1}{2} \int_0^{t_{n, 1}} \int \left|Y_{n, t}'(0)- \sqrt{1 - Y_{n, t}^2(0)}x\right|d\mu(x) dr + \epsilon_n \end{align} for some error term $\epsilon_n$. Here we have used that $Y'_{n, t}(r) = Y'_{n, t}(0)$ for all $r \in [0, t_{n, 1}]$ by piecewise linearity of $Y'_{n, t}$. Now if $|Y_{n, t}'(0)| \ge \pi$, then since $\mu$ is symmetric and supported in $[-\pi, \pi]$ (Theorem \ref{T:local-speeds}), the two inner integrals are the same. Therefore $\epsilon_n = 0$ in this case. Also, when $ |Y_{n, t}'(0)| < \pi $ and $x \in [-\pi, \pi]$, the difference of the integrands is bounded above by $$ \left|\sqrt{1 - Y_{n, t}^2(0)}x - \sqrt{1 - Y_{n, t}^2(r)}x\right| \le |x|\sqrt{1 - \left(1 - \frac{2\pi t}{n-1}\right)^2} \le \pi\sqrt{\frac{4\pi t}{n-1}}. $$ Hence $|\epsilon_n| \le k(t/n)^{3/2}$ for some constant $k$. Now, letting $I_n = n(Y_{n, t}(0) + 1)/2$, and using the notation of Lemma \ref{L:L1-flux}, \eqref{E:frac-en} is equal to $$ \frac{t}{n-1}R_nD_\mu\left(S(U_n^{I_n}, R_n t)\right) + \epsilon_n. $$ Therefore by Lemma \ref{L:L1-flux} (ii) and the bound on $\epsilon_n$, \eqref{E:JY-ineq} implies that \begin{align*} \mathbb{E} J(Y_{n, t}) \le \left(\floor{\frac{n - 1}{2t}} + 1 \right)\left[\frac{t}{n-1} \mathbb{E} R_nD_\mu\left(S(U_n^{I_n}, R_n t)\right) + \mathbb{E} \epsilon_n \right]\xrightarrow[n \to \infty]{} \frac{1}2 \mathbb{E} R D_\mu \left(S(U, Rt) \right). \end{align*} Letting $t \to \infty$, Lemma \ref{L:L1-flux} (iii) then implies that \begin{equation} \label{E:flux-lim} \lim_{t \to \infty} \lim_{n \to \infty} \mathbb{E} J(Y_{n, t}) \le 1. \end{equation} Since subsequential limits of $Y_n$ are Lipschitz by Theorem \ref{T:bounded-speed}, $Y$ is a subsequential limit of $Y_n$ if and only if $Y$ is a subsequential limit of $Y_{n, t}$. Therefore, by Fatou's lemma and the lower semicontinuity of $J(\cdot)$ (Proposition \ref{P:lsc-E}), \begin{equation*} \lim_{n \to \infty} \mathbb{E} J(Y_{n, t}) \ge \mathbb{E} J(Y) \end{equation*} for all $t$, so $\mathbb{E} J(Y) \le 1$ by \eqref{E:flux-lim}. Moreover, $J(Y) \ge 1$ almost surely by Theorem \ref{T:particle-crossings} (i), so $J(Y) = 1$ almost surely. \end{proof} \section{Characterization of minimal flux paths} \label{S:unique} In this section, we show that any subsequential limit $Y$ of $Y_n$ is uniquely determined by its initial position, maximum absolute value, and whether it is initially increasing or decreasing. By Theorem \ref{T:particle-crossings} (iii), it is enough to characterize paths $h \in \text{\fontfamily{ppl}\selectfont Lip}_r$ with $J(h) = 1$. \medskip Let $\text{\fontfamily{ppl}\selectfont Lip}_0$ be the set of path $h \in \text{\fontfamily{ppl}\selectfont Lip}$ with $h(0) = h(1) = 0$. We can first recognize that to characterize minimal flux paths $h \in \text{\fontfamily{ppl}\selectfont Lip}_r$, it is enough to characterize minimal flux paths $h \in \text{\fontfamily{ppl}\selectfont Lip}_0$. This is a simple consequence of the following simple fact. \begin{lemma} \label{L:path-shift} Let $h \in \text{\fontfamily{ppl}\selectfont Lip}_r$ and $t_0 = \inf \{t : h(t) = 0\}$. Define the path \begin{equation*} h_0(t) = \begin{cases} &h(t + t_0), \qquad \quad \quad t \le 1- t_0 \\ &-h(t + t_0 - 1), \qquad t > 1 - t_0. \end{cases} \end{equation*} Then $J(h) = J(h_0)$. In particular, every path $h \in \text{\fontfamily{ppl}\selectfont Lip}_r$ with flux $J(h) = 1$ can be shifted to a path $h_0 \in \text{\fontfamily{ppl}\selectfont Lip}_r$ with $J(h_0) = r$. \end{lemma} We build up to the following characterization of minimal flux paths in $h \in \text{\fontfamily{ppl}\selectfont Lip}_0$. Define the maximum height $m(h)$ of a continuous path $h:[0, 1] \to [-1, 1]$ by $$ m(h) = \max \{ |h(t)| : t \in [0, 1] \}. $$ Recall also the definition of $g(y)$-Lipschitz from Section \ref{S:local}. \begin{theorem} \label{T:path-set} For every $m \in [0, 1]$, there exists exactly one $\pi\sqrt{1- h^2}$-Lipschitz path $h_m \in \text{\fontfamily{ppl}\selectfont Lip}_0$ such that $J_m(h_m) = 1$, $m(h_m) = m$, and $h_m \ge 0$. The paths $h_m$ satisfy the following properties. \smallskip \begin{enumerate}[nosep,label=(\roman*)] \item $h_m(1/2) = m$, and $h_m(t) = h_m(1 - t)$ for $t \in [0, 1/2]$. \item $h_m$ is strictly increasing on the interval $[0, 1/2]$. \item For any $\ell \in [0, m]$, we have that $h_{\ell}(s) \le h_{m}(s)$ for all $s \in [0, 1]$. \end{enumerate} \smallskip Now define $h_{-m} = -h_m$. If $h \in \text{\fontfamily{ppl}\selectfont Lip}_0$ is a $\pi\sqrt{1 -h^2}$-Lipschitz path with $J_m(h) = 1$ and $m(h) = m$, then either $h = h_m$ or $h = h_{-m}$. \end{theorem} Theorem \ref{T:path-set} will be proven as Proposition \ref{P:monotone-traj}, Proposition \ref{P:unique-path}, Corollary \ref{C:sym}, Lemma \ref{L:no-cross}, and Proposition \ref{P:exist-m}. We also state a analogue of Theorem \ref{T:path-set} for minimal flux paths $h \in \text{\fontfamily{ppl}\selectfont Lip}_r$. First, for a path $h \in \text{\fontfamily{ppl}\selectfont Lip}$, define $$ S(h) = \sup_{t \in [0, 1]} h(t) \qquad \;\text{and}\; \qquad I(h) = \inf_{t \in [0, 1]} h(t). $$ \begin{theorem} \label{T:unique-2} Fix $a \in [-1, 1]$ and $k \in [|a|, 1]$. Then the following hold: \begin{enumerate}[nosep,label=(\roman*)] \item There is exactly one $\pi\sqrt{1 - h^2}$-Lipschitz path $h_{a, k} \in \text{\fontfamily{ppl}\selectfont Lip}_r$ such that $h_{a, k}(0) = - h_{a, k}(1) = a$ and $m(h) = S(h) = k$. \item There is a exactly one path $\pi\sqrt{1 - h^2}$-Lipschitz path $h_{a, -k} \in \text{\fontfamily{ppl}\selectfont Lip}_r$ such that $h_{a, -k}(0) = - h_{a, -k}(1) = a$ and $-m(h) = I(h) = - k$. We have that $h_{a, k}(t) = -h_{a, k}(1 - t)$ for all $t$. \item If $a \neq 0$, there is a unique time $t \in (0, 1)$ such that $h_{a, k}(t) = 0$. Moreover, the path $$ g(s) = \begin{cases} &h(t + s), \qquad s \le 1 - t, \\ &-h(t + s - 1), \qquad s > 1 - t \end{cases} $$ is equal to $h_k$ if $a < 0$ and $h_{-k}$ if $a > 0$. \item For $k_1 \le k_2 \in [-1, -|a|] \cup [|a|, 1]$, we have that $h_{a, k_1}(t) \le h_{a, k_2}(t)$ for all $t \in [0, 1]$. \item $h_{a, a} = h_{-a, a}$. \end{enumerate} \end{theorem} All parts of this theorem follow from applying Lemma \ref{L:path-shift} to Theorem \ref{T:path-set}, except part (iv). This will be proven in Lemma \ref{L:no-cross}. \subsection{Basic bounds on $D_\mu$} \label{SS:basic-bounds} In order to work with the functional $J(\cdot)$, we first prove a few basic bounds on $D_\mu$. \begin{lemma} \label{L:max-heights} Let $Y$ be a subsequential limit of $Y_n$. For all $a \in [0, 1]$, we have that $$ \mathbb{P}(m(Y) > a) \leq \frac{D_\mu(0)\sqrt{1- a^2}}{2}. $$ \end{lemma} \begin{proof} Since $Y(0) = -Y(1)$ and $Y(0)$ is uniformly distributed (Theorem \ref{T:bounded-speed}), the left hand side of this inequality can be written as \begin{align*} &\mathbb{P}\bigg(Y(0) < a \;\text{and}\; Y(t) > a \text{ for some } t > 0, \;\text{or}\; Y(0) > - a \;\text{and}\; Y(t) < - a \text{ for some } t > 0 \bigg). \end{align*} By Theorem \ref{T:particle-crossings}, this is bounded above by $J^+(g_a) + J^-(g_{-a})$, where $g_a$ is the path of constant height $a$. Finally, \begin{equation*} J^+(g_a) + J^-(g_{-a}) = \frac{D_\mu(0)\sqrt{1- a^2}}2. \qedhere \end{equation*} \end{proof} \begin{lemma} \label{L:expect-2} $D_\mu(0) = 2.$ That is, if $X$ is a random variable with distribution $\mu$, then $\mathbb{E}|X| = 2$. \end{lemma} \begin{proof} Let $Y$ be a subsequential limit of $Y_n$, and let $\alpha = D_\mu(0)$. By Lemma \ref{L:max-heights}, we have that $$ 1 = \mathbb{P}(m(Y) > 0) \leq \frac{\alpha}2, $$ so $\alpha \geq 2$. The equality above follows since $m(Y) \ge |Y(0)|$ and $|Y(0)|$ is uniformly distributed on $[-1, 1]$ (Theorem \ref{T:bounded-speed}). Now, for every height $a$ such that $$ a > \sqrt{1 - \left(\frac{2}{\alpha}\right)^2}, $$ Lemma \ref{L:max-heights} guarantees that $m(Y) \le a$ with positive probability. Therefore by Theorem \ref{T:particle-crossings} (iii), for any $\epsilon > 0$ there is a $\pi\sqrt{1-y^2}$-Lipschitz path $h_\epsilon \in \text{\fontfamily{ppl}\selectfont Lip}_r$ with $J(h_\epsilon) = 1$ such that \begin{equation} \label{E:m-hep} m(h_\epsilon) \leq m_\epsilon := \sqrt{1 - \left(\frac{2}{\alpha} - \epsilon\right)^2}. \end{equation} Using that $D_\mu$ is minimized at $0$ (Lemma \ref{L:convex}), we have the bound \begin{equation} \label{E:flux-he} \begin{split} 1 = J(h_\epsilon) &\ge \frac{\alpha}2 \int_0^1\sqrt{1-h_\epsilon^2(t)}dt. \end{split} \end{equation} The above integrand is always bounded by $2/\alpha - \epsilon$. Also, since $h_\epsilon(0) = -h_\epsilon(1)$ and $h_\epsilon$ is $\pi$-Lipschitz, the amount of time that $h_\epsilon$ spends in the interval $[-m_\epsilon/2, m_\epsilon/2]$ is at least $m_\epsilon/(2\pi)$. Therefore $$ \frac{2}{\alpha} \ge \int_0^1\sqrt{1-h_\epsilon^2(t)}dt \ge \left(\frac{2}\alpha - \epsilon\right)\left(1 - \frac{m_\epsilon}{2\pi}\right) + \frac{m_\epsilon}{2\pi}\sqrt{1 - \frac{m_\epsilon^2}{4}}. $$ Letting $\epsilon \to 0$, we get an inequality in $\alpha$ which implies that $\alpha = 2$. \end{proof} \begin{lemma} \label{L:crude-deviation-bd} There is a sequence $u_n \to 0$ such that $$ D_\mu(u_n) \leq \frac{2}{\cos(u_n/2)}. $$ \end{lemma} \begin{proof} For any subsequential limit $Y$ of $Y_n$ and any $\epsilon > 0$, Lemmas \ref{L:max-heights} and \ref{L:expect-2} imply that $$ \mathbb{P}(m(Y) \le \epsilon) \geq 1 - \sqrt{1-\epsilon^2} > 0. $$ Also, $Y(0)$ is uniformly distributed by Theorem \ref{T:bounded-speed}, so $\mathbb{P}(m(Y) = 0) = 0$. Therefore by Theorem \ref{T:particle-crossings} (iii), we can find a sequence of positive numbers $\alpha_n \to 0$ and a sequence of paths $h_n \in \text{\fontfamily{ppl}\selectfont Lip}_r$ with $m(h_n) = \alpha_n$ and $J(h_n) = 1$. \medskip Let $s_n$ be the local speed of $h_n$. The total variation of each of the paths $h_n$ is at least $2\alpha_n$, and hence the average absolute local speed of each $h_n$ is at least $2 \arcsin(\alpha_n)$. The convexity and symmetry of $D_\mu$ (Lemma \ref{L:convex}) then gives the following bound. \begin{align*} J(h_n) &\geq \sqrt{1- \alpha_n^2}\int_0^1 D_\mu(s_n(t)) dt \\ &\geq \sqrt {1-\alpha_n^2} D_\mu(2 \arcsin(\alpha_n)). \end{align*} Letting $u_n = 2\arcsin(\alpha_n)$ and rearranging completes the proof. \end{proof} \subsection{Monotonicity and uniqueness for minimal flux paths} \label{SS:monotone} In this subsection, we prove that minimal flux paths $h \in \text{\fontfamily{ppl}\selectfont Lip}_0$ with a particular maximum height $m(h)$ are unique up to sign, and that they satisfy a particular monotonicity relation. \medskip We start with two simple lemmas. The first lemma shows that minimal flux paths minimize flux on every subinterval of $[0, 1]$. \begin{lemma} \label{L:min-path} Let $h \in \text{\fontfamily{ppl}\selectfont Lip}_r$ be a path with $J(h) = 1$. For any interval $[a, b] \subset [0, 1]$ and any path $g \in \text{\fontfamily{ppl}\selectfont Lip}$ with $g(a) = h(a)$ and $g(b) = h(b)$, we have that $$ J(h; [a, b]) \le J(g; [a, b]). $$ Moreover, if $f \in \text{\fontfamily{ppl}\selectfont Lip}_r$ is another path with $J(f) = 1$, $f(a) = h(a)$, and $f(b) = h(b)$, then we can form a new path $p \in \text{\fontfamily{ppl}\selectfont Lip}_r$ with $J(p) = 1$ by letting \begin{equation*} p(t) = \begin{cases} &f(t) \qquad t \leq t_1, t \ge t_2 \\ &h(t) \qquad t \in [t_1, t_2]. \end{cases} \end{equation*} \end{lemma} \begin{proof} If there is some $g$ with $J(h; [a, b]) > J(g; [a, b])$, then we can make a new path $p \in \text{\fontfamily{ppl}\selectfont Lip}_r$ which is equal to $g$ on $[a, b]$ and equal to $h$ on $[a, b]^c$. This path $p$ will have $J(p) < 1$, contradicting Theorem \ref{T:particle-crossings} (ii). The second part of the lemma is a consequence of the fact that \[ J(p) = J(p ; [a, b]) + J(p ; [a, b]^c). \qquad \qquad \qedhere \] \end{proof} The next lemma uses the bounds established in Section \ref{SS:basic-bounds} to eliminate plateaus in minimal flux paths. \begin{lemma} \label{L:no-plat} For any height $\alpha \in (0, 1)$ and any interval $[t_1, t_2] \subset [0, 1]$, we have that $$ \inf_{h \in \text{\fontfamily{ppl}\selectfont Lip}} \big\{J(h; [t_1, t_2]) : h(t_1) = \alpha, h(t_2) = \alpha \big\} < (t_2-t_1)\sqrt{1-\alpha^2}. $$ \end{lemma} \begin{proof} Without loss of generality, we may assume that $t_1 = 0$. Let $\{u_n\}_{n \in \mathbb{N}}$ be as in Lemma \ref{L:crude-deviation-bd}, and define a sequence of paths $h_n \in \text{\fontfamily{ppl}\selectfont Lip}$ by letting \begin{equation*} h_n(t) = \begin{cases} \;\sin (\arcsin(\alpha) + u_n t), \qquad &t \leq t_2/2 \\ \;h_n(t_2 - t), \qquad \qquad \quad \;\;\;\;\; &t \in [t_2/2, t_2] \\ \;\alpha, \qquad \qquad \qquad \quad \quad \; &t > t_2. \end{cases} \end{equation*} Lemma \ref{L:crude-deviation-bd} gives the following bound on the flux of $h_n$ on the interval $[0, t_2]$. \begin{align*} J(h_n; [0, t_2]) &\leq \frac{2}{\cos(u_n/2)} \int_0^{t_2/2} \cos(\arcsin(\alpha) + u_n t)dt \\ &= t_2 \sqrt{1-\alpha^2} - \frac{t_2^2 \alpha u_n}{4} + O(u_n^2). \end{align*} For small enough $u_n$, this calculation shows that $J(h_n) < t_2 \sqrt{1-\alpha^2}$. \end{proof} We can now prove that minimal flux paths $h \in \text{\fontfamily{ppl}\selectfont Lip}_0$ are unimodal. \begin{prop} \label{P:monotone-traj} Let $h \in \text{\fontfamily{ppl}\selectfont Lip}_0$ be such that $J(h) = 1$, and $m(h) \in (0, 1)$. Then $|h(1/2)| = m(h)$. Moreover, if $h(1/2) = m(h)$, then $h$ is strictly increasing on $[0, 1/2]$ and strictly decreasing on $[1/2, 1]$. If $-h(1/2) = m(h)$, then $h$ is strictly decreasing on $[0, 1/2]$ and strictly increasing on $[1/2, 1]$. \end{prop} \begin{proof} First consider the case where $h \ge 0$. Let $t_m$ be any time when $h(t_m) = m(h)$. Suppose that there exist times $s_1 < s_2 \in [0, t_m]$ such that $h(s_1) = h(s_2)$, and $h(t) \le h(s_1)$ on the interval $[s_1, s_2]$. Define a new path $$ r(t) = \begin{cases} h(t), \qquad &t \le s_1, t \ge t_m\\ h(t + (s_2 - s_1)), \qquad &t \in (s_1, t_m - (s_2 -s_1)]\\ m(h) \qquad &t \in [t_m -(s_2 - s_1), t_m]. \end{cases} $$ The path $r$ replaces the segment of $h$ in the interval $[s_1, s_2]$ with a plateau at height $t_m$ at a later time. This operation cannot increase flux since $D_\mu$ is minimized at $0$ (Lemma \ref{L:convex}), so $r$ must be a minimal flux path in $\text{\fontfamily{ppl}\selectfont Lip}_r$. However $$ J(r; [t_m -(s_2 - s_1), t_m]) = (s_2 - s_1)\sqrt{1 - m^2(h)}, $$ which contradicts Lemmas \ref{L:min-path} and \ref{L:no-plat}. Therefore there is no interval $[s_1, s_2] \subset [0, t_m]$ where $h(s_1) = h(s_2)$ and $h(t) \le h(s_1)$ for $t \in [s_1, s_2]$, so $h$ must be strictly increasing on $[0, t_m]$. Similarly, $h$ is strictly decreasing on $[t_m, 1]$. The point $t_m$ is the unique time when $h$ achieves its maximum. \medskip We now show that $t_m = 1/2$. Without loss of generality, assume that $t_m \le 1/2$. Define a new path $g(t) = h(1 - t)$. The path $g \in \text{\fontfamily{ppl}\selectfont Lip}_0$ also satisfies $J(g) = 1$ and $m(g) = m$. We have that $g(1/2) = h(1/2)$, and so by Lemma \ref{L:min-path} we can create a path $p \in \text{\fontfamily{ppl}\selectfont Lip}_r$ with $J(p) = 1$ and $m(p) = m$ by letting \begin{equation*} p(t) =\begin{cases} &h(t) \qquad t \leq 1/2\\ &g(t) \qquad t \geq 1/2. \end{cases} \end{equation*} The path $p(t)$ achieves its maximum height at both $t_m$ and $1- t_m$, so we must have that $t_m = 1 - t_m$, and hence $t_m = 1/2$. \medskip Now if $h$ is not a strictly non-negative path, then we can create a non-negative path $p(t) = |h(t)|$. By the symmetry of $\mu$ (Theorem \ref{T:local-speeds}), the path $p$ must again have minimal flux, so by the above argument $p$ is strictly increasing on $[0, 1/2]$ and strictly decreasing on $[1/2, 1]$. Therefore either $h = p$ or $h = -p$, completing the proof. \end{proof} We can now use Proposition \ref{P:monotone-traj} to prove uniqueness of minimal flux paths with a given maximum height. \begin{prop} \label{P:unique-path} For every $m \in [0, 1]$, there is at most one $\pi\sqrt{1- y^2}$-Lipschitz path $h \in \text{\fontfamily{ppl}\selectfont Lip}_0$ with $J(h) = 1$, $m(h) = m$, and $h \ge 0$. If such a path $h$ exists, then the only other $\pi\sqrt{1- y^2}$-Lipschitz path $g \in \text{\fontfamily{ppl}\selectfont Lip}_0$ with $J(g) = 1$ and $m(g) = m$ is $g = -h$. \end{prop} \begin{proof} First observe that the only $\pi\sqrt{1- y^2}$-Lipschitz paths in $\text{\fontfamily{ppl}\selectfont Lip}_0$ with $m(h) = 1$ are $h = \pm \sin(\pi t)$. Similarly, the only path $h \in \text{\fontfamily{ppl}\selectfont Lip}$ with $m(h) = 0$ is $h = 0$. This proves the proposition for $m \in \{0, 1\}$. Now we assume $m \in (0, 1)$. \medskip Suppose that $h, g \in \text{\fontfamily{ppl}\selectfont Lip}_0$ are two distinct non-negative paths with $J(h) = J(g) = 1$ and $m(h) = m(g) = m$. By Proposition \ref{P:monotone-traj}, $h(1/2) = g(1/2) = 1/2$. Without loss of generality, there exists a time $t_2 \in [0, 1/2)$ such that $g(t_2) < h(t_2)$. Let $t_1 < t_2$ be such that $h(t_1) = g(t_2)$. Define the shifted path \begin{equation*} g^*(t) = \begin{cases} \;\;g(t + (t_2 - t_1)), \qquad &t \leq 1 - (t_2 - t_1) \\ \;\;- g(t + (t_2 - t_1) - 1), \qquad& t \geq 1 - (t_2 - t_1). \end{cases} \end{equation*} By Proposition \ref{P:monotone-traj}, we have that $$ m = g^*(1/2 - (t_2 - t_1)) > h(1/2 - (t_2 - t_1)) \quad \;\text{and}\; \quad g^*(1/2) < h(1/2) = m. $$ Therefore there is some time $t_3 \in (1/2 - (t_2 - t_1), 1/2)$ such that $g^*(t_3) = h(t_3)$. Moreover, $g^* \in \text{\fontfamily{ppl}\selectfont Lip}_r$, $J(g^*) = 1$, and $g^*(t_1) = h(t_1)$. Therefore by Lemma \ref{L:min-path} we can create a path $p \in \text{\fontfamily{ppl}\selectfont Lip}_0$ with $J(p) = 1$ by letting \begin{equation*} p(t) = \begin{cases} \;\;h(t), \qquad& t \leq t_1 \;\text{or}\; t \geq t_3 \\ \;\;g^*(t), \qquad& t \in (t_1, t_3). \end{cases} \end{equation*} This new path $p$ is a non-negative minimal flux path in $\text{\fontfamily{ppl}\selectfont Lip}_0$ which does not uniquely achieve its maximum value at $1/2$, contradicting Proposition \ref{P:monotone-traj}. \medskip Now, if $g \in \text{\fontfamily{ppl}\selectfont Lip}_0$ is another minimal flux path with $J(g) = 1$ and $m(g) = m$, then $|g|$ is a non-negative minimal flux path in $\text{\fontfamily{ppl}\selectfont Lip}_0$, so $|g| = h$. Since $h(t) > 0$ for $t \in (0, 1)$ by Proposition \ref{P:monotone-traj}, either $g = h$ or $g = -h$. \end{proof} Proposition \ref{P:unique-path} gives us the following easy corollary. \begin{corollary} \label{C:sym} Any $\pi\sqrt{1-h^2}$-Lipschitz path $h \in \text{\fontfamily{ppl}\selectfont Lip}_0$ with $J(h) = 1$ has $h(t) = h(1-t)$ for all $t \in [0, 1]$. \end{corollary} \begin{proof} Without loss of generality, assume that $h \ge 0$. Both $h(t)$ and $h(1-t)$ are non-negative minimal flux paths satisfying all the conditions of Proposition \ref{P:unique-path} and with the same maximum height, so by that proposition, $h(t) = h(1-t)$ for all $t$. \end{proof} \begin{remark} \label{R:shift} Note that the uniqueness proofs in this section automatically give the uniqueness of the paths $h_{a, k}$ in Theorem \ref{T:unique-2}. The fact that uniqueness immediately carries over to shifted paths follows from the strict monotonicity in Proposition \ref{P:monotone-traj}. \end{remark} \subsection{Existence of minimal flux paths} \label{SS:unique} We now show that for every $m \in [0, 1]$, that there exists a minimal flux path in $\text{\fontfamily{ppl}\selectfont Lip}_0$ with maximum height $m$. This is a consequence of the following proposition, which shows that the inequality in Theorem \ref{T:particle-crossings} is an equality for minimal flux paths. Recall the definition of the upcrossing probability $P^+_Y(h; [a, b])$ and the downcrossing probability $P^-_Y(h; [a, b])$ from Section \ref{S:flux-intro}. \begin{prop} \label{P:flux-upcross} Let $h \in \text{\fontfamily{ppl}\selectfont Lip}_r$ be a path with $J(h) = 1$. Let $Y$ be a subsequential limit of $Y_n$. Then for any interval $[a, b]$, we have that $$ P^+_Y(h; [a, b]) = J^+(h; [a, b]) \quad \;\text{and}\; \quad P^-_Y(h; [a, b]) = J^-(h; [a, b]). $$ \end{prop} \begin{proof} We will only prove the first equality, as the second one follows by the symmetry of $Y$ and $\mu$ (Theorem \ref{T:local-speeds}, Proposition \ref{P:Y-symmetries}). Since $Y(0) = -Y(1)$ almost surely by Theorem \ref{T:bounded-speed}, we have that \begin{equation*} 1 = P^+_Y(h; [0, 1]) + P^-_Y(h; [0, 1])= J^+(h) +J^-(h). \end{equation*} Therefore by Theorem \ref{T:particle-crossings} (i), we have that $P^+_Y(h; [0, 1]) = J^+(h)$. Now we have that \begin{equation} \label{E:big-J+} \begin{split} P^+_Y(h; [0, 1])& \le P^+_Y(h; [0, a]) + P^+_Y(h; [a, b]) + P^+_Y(h; [b, 1])\\ &\le J^+(h; [a, b]^c) + P^+_Y(h; [a, b]) \\ &\le J^+(h; [a, b]^c) + J^+(h; [a, b]) = J^+(h). \end{split} \end{equation} To see the first inequality above, note that the union of the three events on the right hand side contains the event on the left hand side minus the set $\{Y(a) = h(a) \;\text{or}\; Y(b) = h(b)\}$. This set has probability $0$ since $Y(t)$ is uniformly distributed for all $t$ (Theorem \ref{T:bounded-speed}). The second and third inequalities follow from Theorem \ref{T:particle-crossings} (i). \medskip Since $P^+_Y(h; [0, 1]) = J^+(h)$, all the inequalities in \eqref{E:big-J+} must in fact be equalities, proving the lemma. \end{proof} To apply the above proposition in order to prove the existence of minimal flux paths at every height, we need three lemmas. The first lemma shows that any two minimal flux paths in $\text{\fontfamily{ppl}\selectfont Lip}_r$ cannot cross each other more than once. This lemma also proves part (iii) of Theorem \ref{T:path-set}. \begin{lemma} \label{L:no-cross} Suppose that $h, g \in \text{\fontfamily{ppl}\selectfont Lip}_r$ are non-negative paths with $J(h) = J(g) = 1$. Then there cannot exist times $t_0 < t_1 < t_2 \in [0, 1]$ such that $h(t_0) < g(t_0)$, $h(t_1) > g(t_1)$, and $h(t_2) = g(t_2)$. In particular, if $g(0) = h(0)$, then either $h \le g$ or $g \le h$. \end{lemma} \begin{proof} Suppose there exist times $t_0 < t_1 < t_2 \in [0, 1]$ such that $h(t_0) < g(t_0)$, $h(t_1) > g(t_1)$, and $h(t_2) = g(t_2)$. First, by possibly shifting the paths as in Lemma \ref{L:path-shift}, we may assume that $t_2 = 1$. Hence $g(1) = h(1) = -g(0) = -h(0)$. Now let $s \in (t_0, t_1)$ be such that $h(s) = g(s)$. Letting $a = g(0)$, Remark \ref{R:shift} implies that $h = h_{a, k_1}$ and $g = h_{a, k_2}$ for some $k_1, k_2 \in [-1, -|a|] \cup [|a|, 1]$. Here the paths $h_{a, k}$ are as in Theorem \ref{T:unique-2}. Without loss of generality, assume that $k_2 > k_1$ and that $k_2 \ge a$; the other cases follow by symmetric arguments. \medskip Define $p = \max(h, g)$. By Lemma \ref{L:min-path}, $J(p) = 1$. Moreover, $p(0) = -p(1) = a$, and since $k_2 > k_1$ and $k_2 \ge |a|$, we have that $m(p) = S(p) = k_2$. This contradicts the uniqueness of the path $h_{a, k_2}$ established in Remark \ref{R:shift}. \end{proof} The next lemma establishes a strong concavity property for minimal flux paths. \begin{lemma} \label{L:concave-paths} For every $k \in [0, 1]$, the path $\arcsin(h_k)$ is concave. In other words, the local speed $s_k$ of $h_k$ is a non-increasing function of $t$. \end{lemma} \begin{proof} By the symmetry of $h_k$ (Corollary \ref{C:sym}), it suffices to prove that $\arcsin(h_k)$ is concave on $[0, 1/2]$. Suppose that there exist times $t_1 < t_2 < t_3 \le 1/2$ such that $h(t_2) < p(t_2)$, where $p$ is the path of constant local speed $\close{s} \in [0, \pi]$ with $p(t_1) = h(t_1)$ and $p(t_2) = h(t_2)$. Let $$ t_4 = \sup \{t \in [t_1, t_2] : p(t) \le h(t)\} \qquad \;\text{and}\; \qquad t_5 = \inf \{t \in [t_2, t_3] : p(t) \le h(t)\}. $$ Then $h(t_4) = p(t_4)$ and $h(t_5) = p(t_5)$, and for every $t \in (t_4, t_5)$, we have that $h(t) < p(t)$. By Lemma \ref{L:convex}, we can find a line $L(s) = as + b$ with $b \ge 0$ such that $L(s) \leq D_\mu(s)$ for all $s$, and such that $L(\close{s}) = D_\mu(\close{s}).$ Therefore \begin{align*} J(h; [t_4, t_5]) & \geq \int_{t_4}^{t_5} (as(t) + b) \sqrt{1-h^2(t)} dt \\ &\ge \int_{t_4}^{t_5} ah'(t) dt + \int_{t_4}^{t_5} b \sqrt{1-p^2(t)} dt. \end{align*} The inequality in the second line follows since $p(t) > h(t) \ge 0$ for all $t \in (t_4, t_5)$. By the fundamental theorem of calculus, since $p(t_4) = h(t_4)$ and $p(t_5) = h(t_5)$, we have \begin{align*} \int_{t_4}^{t_5} ah'(t) dt + \int_{t_4}^{t_5} b \sqrt{1-p^2(t)} dt &= \int_{t_4}^{t_5} ap'(t) dt + \int_{t_4}^{t_5} b \sqrt{1-p^2(t)} dt \\ & = J(p; [t_4, t_5]). \end{align*} We can then create a non-negative $\pi\sqrt{1 - h^2}$-Lipschitz path $g \in \text{\fontfamily{ppl}\selectfont Lip}_0$ with $m(g) = k$ by letting $g(t) = p(t)$ for $t \in [t_4, t_5]$ and $g(t) = h(t)$ otherwise. $J(g) \le 1$, contradicting the uniqueness of $h_k$. \end{proof} The next lemma is a consequence of the symmetry and concavity of $Y$. \begin{lemma} \label{L:Y-sym-type} Let $Y$ be a subsequential limit of $Y_n$, let $t \in [0, 1)$, and let $[a, b] \subset [0, 1]$. Then $Y'(t)$ exists almost surely and \begin{enumerate}[label=(\roman*)] \item $\mathbb{P}(Y'(t) = 0) = 0.$ \item $\mathbb{P}(Y'(t) > 0 , \; Y(t) \in [a, b]) = \mathbb{P}(Y'(t) < 0 , \; Y(t) \in [a, b]) = 1/2.$ \end{enumerate} \end{lemma} \begin{proof} By time-stationarity of $Y$ (Proposition \ref{P:Y-symmetries} (i)) it suffices to consider $t = 0$. This also implies that $$ Y'(0) \stackrel{d}{=} Y'(U), $$ for an independent uniform random variable $U$ on $[0, 1]$. Since $Y$ is Lipschitz and hence almost everywhere differentiable, this proves that $Y'(0)$ exists almost surely. By the concavity and strict monotonicity of minimal flux paths (Lemma \ref{L:concave-paths} and Proposition \ref{P:monotone-traj}), we have that \begin{equation} \label{E:prob-Y-0} \mathbb{P} (Y'(U) = 0) = \mathbb{P}( m(Y) = |Y(U)|) = 0. \end{equation} Define $Z(t) = -Y(1- t)$. Since $Y(0) = -Y(1)$ (Theorem \ref{T:bounded-speed}), we have that $ Z(0) = Y(0) $ and $Z'(0) = -Y'(0)$. By Proposition \ref{P:Y-symmetries}, $Y \stackrel{d}{=} Z$, so $$ \mathbb{P}(Y'(0) > 0 , Y(0) \in [a, b]) = \mathbb{P}(Y'(0) < 0 , Y(0) \in [a, b]). $$ Putting this together with \eqref{E:prob-Y-0} completes the proof. \end{proof} We are now ready to prove the existence of minimal flux paths at every height. \begin{prop} \label{P:exist-m} For every $m \in [0, 1]$, there exists a $\pi\sqrt{1- h^2}$-Lipschitz path $h_m \in \text{\fontfamily{ppl}\selectfont Lip}_0$ such that $m(h_m) = m$, $J(h_m) = 1$, and $h \ge 0$. \end{prop} \begin{proof} First observe that we can set $h_0(t) = 0$ and $h_1(t) = \sin (\pi t)$. Both of these paths satisfy all of the above properties. For this, we use that $D_\mu(\pm \pi) = \pi$ and $D_\mu(0) = 2$ (Lemma \ref{L:convex} and Lemma \ref{L:expect-2}). \medskip Let $A \subset [0, 1]$ be the set of all $m$ such there is a $\pi\sqrt{1- h^2}$-Lipschitz path $h_m \in \text{\fontfamily{ppl}\selectfont Lip}_0$ with $h_m \ge 0$, $J(h_m) = 1$ and $m(h_m) = m$. Let $m \in \close{A}$, and let $m_k \in A$ be a sequence converging to $m$. \medskip By the Arzel\`a-Ascoli Theorem, there exists a $\pi\sqrt{1-h^2}$-Lipschitz subsequential limit $h_m \in \text{\fontfamily{ppl}\selectfont Lip}_0$ of $h_{m_k}$ with $m(h_m) = m$ and $h_m \ge 0$. By the lower semicontinuity of $J$ (Proposition \ref{P:lsc-E}), $J(h_m) = 1$. Therefore $A$ is closed. \medskip Now suppose that $A \ne [0, 1]$. Then there exists an interval $(\ell, m) \subset [0, 1]$ such that $m \in A$ and $(\ell, m) \cap A = \emptyset$. Let $k = (\ell + m)/2$, and choose $t_1 < t_2 \in [0, 1/2]$ so that $ h_m(t_1) = \ell $ and $h_m(t_2) = k.$ We will show that \begin{equation} \label{E:P=P} P^+_Y(h_m ; [t_1, t_2]) + P^-_Y(h_m ; [t_1, t_2]) = P^+_Y\left(k ; [t_1, t_2]\right) + P^-_Y\left(k; [t_1, t_2]\right). \end{equation} Here we use $k$ as shorthand for the line $h(t) = k$. \medskip Let $g \in \text{\fontfamily{ppl}\selectfont Lip}_r$ be any minimal flux $\pi\sqrt{1- y^2}$-Lipschitz path. By the monotonicity established in Proposition \ref{P:monotone-traj}, $g$ upcrosses $k$ at most once in the interval $[0, 1]$, and downcrosses $k$ at most once. Moreover, by Lemma \ref{L:no-cross}, $g$ crosses $h_m$ at most once. \medskip If $g(t_1) \notin [\ell, k]$, and $g(t_2) \ne k$, then the total number of crossings of the lines $h_m$ and $k$ is even. Hence, either $g$ crosses $h_m$ the same amount of times that $g$ crosses $k$, or $g$ first crosses $k$ going down, and then crosses $k$ going up. This second possibility cannot happen by the monotonicity of $g$ (Proposition \ref{P:monotone-traj}). \medskip If $g(t_1) \in (\ell, k)$ and $g'(t_1) < 0$, then again by the monotonicity properties of $g$, $g$ downcrosses $h_m$, does not upcross $h_m$, and neither upcrosses nor downcrosses $k$. \medskip If $g(t_1) \in (\ell, k)$, $g'(t_1) > 0$, and $g(t_2) \ne k$, then $m(g) \ge m$ since $(\ell, m) \cap A = \emptyset$. By the monotonicity of $g$ (Proposition \ref{P:monotone-traj}), $g(t^*) = m(g)$ for some $t^* > t_1$. If $g(t_2) < h(t_2)$, then since both $g$ and $h_m$ are minimal flux paths, the restrictions on crossings imposed by Lemma \ref{L:no-cross} imply that $t^* < t_2$. In this case, $g$ both upcrosses and downcrosses $k$ in the interval $[t_1, t_2]$, and downcrosses but does not upcross $h_m$. \medskip On the other hand, if $g(t_2) > h(t_2)$, then $g$ upcrosses $k$ in the interval $[t_1, t_2]$ but does not downcross $k$, and does not cross $h$ at all in this interval. \medskip Now let $Y$ be any subsequential limit of $Y_n$. By Theorem \ref{T:bounded-speed} and Lemma \ref{L:Y-sym-type} (i), we have that $$ \mathbb{P} \Big(Y(t_1) \notin \{\ell, k\}, Y(t_2) \ne k, Y'(t_1) \text{ exists and is not equal to $0$}\Big) = 1. $$ Therefore by the above analysis of minimal flux $\pi\sqrt{1- y^2}$-Lipschitz paths $g \in \text{\fontfamily{ppl}\selectfont Lip}_r$, we have that \begin{align*} P^+_Y(h_m ; [t_1, t_2]) - P^+_Y\left(k ; [t_1, t_2]\right) &= - \mathbb{P}(Y'(t_1) > 0, Y(t_1) \in [\ell, k]), \quad \;\text{and}\; \\ P^-_Y(h_m ; [t_1, t_2]) - P^-_Y\left(k ; [t_1, t_2]\right) &= \mathbb{P}(Y'(t_1) < 0, Y(t_1) \in [\ell, k]). \end{align*} Equation \eqref{E:P=P} then follows from the symmetry established in Lemma \ref{L:Y-sym-type} (ii). Now by Proposition \ref{P:flux-upcross}, the left hand side of \eqref{E:P=P} is equal to $J(h_m ; [t_1, t_2])$, and by Theorem \ref{T:particle-crossings} (i), the right hand side of \eqref{E:P=P} is bounded above by $J\left(k; [t_1, t_2] \right)$. However, since $D_\mu$ is minimized at $0$ by Lemma \ref{L:convex}, we can also easily calculate that \begin{align*} J(h_m ; [t_1, t_2]) &\ge \frac{D_\mu(0)}{2} \int_{t_1}^{t_2} \sqrt{1 - h^2_m(t)}dt \\ &> \frac{D_\mu(0)}2(t_2 - t_1)\sqrt{1 - {k^2}} = J\left(k; [t_1, t_2] \right). \end{align*} This is a contradiction, so $A$ must be the whole interval $[0,1]$. \end{proof} \section{The derivative distribution} \label{S:path-speed} Let $ \mathbb{P}_{a}(Y \in \cdot) $ be the regular conditional distribution of $Y$ given that $Y(0) = a$. In this section, we use the structure of minimal flux paths to find $\mathbb{P}_a(Y'(0) \in \cdot)$. \begin{prop} \label{P:local-global-deriv} Let $Y$ be any subsequential limit of $Y_n$. With probability one, $Y'(0)$ exists. Moreover, \begin{equation} \label{E:X-Y'} \mathbb{P}_{a} \left(\frac{Y'(0)}{\sqrt{1 - a^2}} \in \cdot\right) = \mu, \end{equation} for Lebesgue-a.e. $a \in (-1, 1)$. \end{prop} As an immediate corollary of Proposition \ref{P:local-global-deriv}, we will be able to find the distribution of the maximum height function $m(Y)$ for any subsequential limit $Y$ of $Y_n$. This will allow us to show that the sequence $\{Y_n\}$ has a unique limit point. \medskip Moreover, Proposition \ref{P:local-global-deriv} will be used later to prove an integral transform formula for $\mu$, which will allow us to determine $\mu$, and then $Y$. This is done in Section \ref{S:integral-formula}. \medskip To prove Proposition \ref{P:local-global-deriv}, we first show that minimal flux paths fill space. For this lemma, we use the notation of Theorem \ref{T:unique-2}. \begin{lemma} \label{L:closed-int} Let $a \in [0, 1]$, and let $$ M_a = \left\{h_{a, k} : k \in [-1, -a] \cup [a, 1] \right\}. $$ Then for any $t \in [0, 1]$, we have that $$ \{h(t) : h \in M_a\} = [\sin(-\pi t + \arcsin(a)), \sin(\pi t + \arcsin(a))]. $$ \end{lemma} \begin{proof} Let $K_{a, t} = \{h(t) : h \in M_a\}.$ $K_{a, t}$ contains the endpoints of the above interval, since $M_a$ contains the two extremal functions $h_{a, -1}(t) = \sin(-\pi t + \arcsin(a))$ and $h_{a, 1}(t) = \sin(\pi t + \arcsin(a))$. Moreover, the ordering on minimal flux paths (Theorem \ref{T:unique-2} (iv)) implies that $$ K_{a, t} \subset [\sin(-\pi t + \arcsin(a)), \sin(\pi t + \arcsin(a))]. $$ Now suppose that $c \in (\sin(-\pi t + \arcsin(a)), \sin(\pi t + \arcsin(a)))$. Let $$ A = \{h_{a, k} \in M_a : h_{a, k}(t) \le c\} \qquad \;\text{and}\; \qquad B = \{h_{a, k} \in M_a : h_{a, k}(t) > c\}. $$ Both of these sets are non-empty. By Theorem \ref{T:unique-2} (iv), $M_a$ is a linearly ordered subset of the set of $\pi \sqrt{1-h^2}$-Lipschitz paths in $\text{\fontfamily{ppl}\selectfont Lip}$ with the usual partial order on functions. Therefore the sets $A$ and $B$ have a supremum and infimum in $\text{\fontfamily{ppl}\selectfont Lip}$ by the Arzel\`a-Ascoli Theorem. By the lower semicontinuity of flux (Proposition \ref{P:lsc-E}), $\sup A$ and $\inf B$ lie in the set $M_a$. This implies that \begin{align*} \sup A &= h_{a, k_A}, \quad \text{ where } \quad k_A = \sup \{ k : h_{a, k} \in A\}, \quad \;\text{and}\; \\ \inf A &= h_{a, k_B}, \quad \text{ where } \quad k_B = \inf \{ k : h_{a, k} \in B\}. \end{align*} Since $A \cup B = M_a$, we either have that $k_A = k_B$, or else $k_A = -a$ and $k_B = a$. Since $h_{a, a} = h_{-a, a}$ by Theorem \ref{T:unique-2} (v), this implies that $h_{a, k_A}(t) = h_{a, k_B}(t) = c$, so $c \in K_{a, t}$. \end{proof} \begin{proof}[Proof of Proposition \ref{P:local-global-deriv}] First note that $Y'(0)$ exists almost surely by Lemma \ref{L:Y-sym-type}. We now show that for every $q \in (-\pi, \pi)$ and $b < c \in (-1, 1)$, that \begin{equation} \label{E:deriv-integrated} \int_{b}^c \mathbb{E}_{a} (Y'(0) - \sqrt{1 - a^2}q)^+ da = D_\mu^+(q) \int_b^c \sqrt{1-a^2}da. \end{equation} Here $\mathbb{E}_a$ is the expectation taken with respect to $\mathbb{P}_a$. Lemma \ref{L:closed-int} guarantees that there exists a time $t_0>0$ such that for any $r \in [0, t_0]$ and $a \in [b, c]$, there is a minimal flux $\pi\sqrt{1-g^2}$-Lipschitz path $g \in \text{\fontfamily{ppl}\selectfont Lip}_r$ with $g(0) = a$ and $g(r) = \sin(\arcsin(a) + qr)$. Call this path $g_{a, r}$. Letting $$ s_{a, r}(t) = \frac{g'_{a, r}(t)}{\sqrt{1 - g^2_{a, r}(t)}} $$ be the local speed of $g_{a, r}$, we have that \begin{align} \nonumber J^+(g_{a, r} ; [0, r]) &\ge \frac{1}2 \min_{t \in [0, r]} \sqrt{1 - g^2_{a, r}(t)} \int_0^t D^+_\mu(s_{a, r}(t)) dt \\ \label{E:J-lower} &\ge \frac{1}2 \min_{t \in [0, r]} {\sqrt{1 - g^2_{a, r}(t)}} t D^+_\mu(q). \end{align} Here the second inequality follows by Jensen's inequality, using that $D^+_\mu$ is convex (Lemma \ref{L:convex}) and that the average local speed of $g_{a, r}$ is $q$. Now letting $g_a(t) = \sin(\arcsin(a) + qt)$, Lemma \ref{L:min-path} implies that \begin{equation} \label{E:g-J-h} J^+(g_{a, r} ; [0,r]) \le J^+(g_a ; [0, r]) \le \frac{1}2 \max_{t \in [0, r]} {\sqrt{1 - g^2_{a}(t)}} r D^+_\mu(q). \end{equation} Combining this inequality with \eqref{E:J-lower} implies that $$ \lim_{r \to 0} \frac{J^+(g_{a, r} ; [0, r])}{r} = \frac{D_\mu^+(q)}2 \sqrt{1 - a^2}. $$ Moreover, \eqref{E:g-J-h} implies that $J^+(g_{a, r} ; [0, r]) \le rD^+_\mu(q)$ for all $a \in [b, c], r \in [0, t_0]$. Therefore by the bounded convergence theorem, we have that \begin{equation} \label{E:want-1} \lim_{t \to 0} \int_{b}^c \frac{J^+(g_{a, r} ; [0, r])}r da = \frac{D^+_\mu(q)}2\int_b^c \sqrt{1- a^2}da. \end{equation} Now recall that $P^+_Y(g_{a, r} ; [0, r])$ is the probability that $Y$ upcrosses $g_{a, r}$ in the interval $[0, r]$. Since minimal flux paths cross at most once (Lemma \ref{L:no-cross}), $$ \mathbb{P}(Y(0) < a, Y(r) > g_{a, r}(r))= P^+_Y(g_{a, r} ; [0, r]). $$ Therefore by Proposition \ref{P:flux-upcross}, we can write \begin{align} \nonumber \int_{b}^c \frac{J^+(g_{a, r} ; [0, r])}r da &= \int_{b}^c \frac{\mathbb{P}(Y(0) < a, Y(r) > g_{a, r}(r))}r da. \end{align} Then defining $P_x(a, r) = \mathbb{P}_x(Y(r) > g_{a, r}(r))/r$, the right hand side above is equal to \begin{align} \label{E:J-+-da} \int_{b}^c \frac{1}2 \int_{-1}^a P_x(a, r)dxda = \frac{1}2 \int_{-1}^c \int_{x \vee b}^c P_x(a, r) dadx. \end{align} Now define $S_{x, y} = \sin(\arcsin(x) + (\pi - q)y)$. Since $Y$ is $\pi\sqrt{1-h^2}$-Lipschitz almost surely (Theorem \ref{T:bounded-speed}), for almost every $x \in [-1, 1]$, we have that $$ \mathbb{P}_x(Y(t) > g_{a, r}(r)) = 0 $$ whenever $a > S_{x, r}.$ Therefore we can rewrite the right hand side of \eqref{E:J-+-da} as \begin{equation} \label{E:want-2} \begin{split} \frac{1}2 \left[\int_b^{c} \int_{x}^{S_{x, r}} P_x(a, r) da dx - \int_{S_{c, -r}}^c \int_c^{S_{x, r}} P_x(a, r) da dx + \int_{S_{b, -r}}^b \int_b^{S_{x, r}} P_x(a, r) da dx\right]. \end{split} \end{equation} In the second and third terms above, the integrand is of size $O(r^{-1})$, whereas the region of integration is of size $O(r^{2})$. Therefore these terms go to zero as $r \to 0$. Now letting $$ f(a, x, s, r) = \frac{\sin(\arcsin(a) + sr) - x}r, $$ and making the substitution $u = f(a, x, q, r)$, we can write the first term of \eqref{E:want-2} as \begin{equation} \label{E:b-c} \begin{split} \int_b^c \int_{f(x, x, q, r)}^{f(x, x, \pi, r)} \mathbb{P}_x\left(\frac{Y(r) - Y(0)}r > u \right) \frac{\cos(\arcsin(ru + x) - qr)}{\sqrt{1 - (ru + x)^2}}du dx. \end{split} \end{equation} Moreover, for every $x \in [-1, 1]$ such that $Y'(0)$ exists $\mathbb{P}_x$-almost surely, and for $u \in [q\sqrt{1 - x^2}, \pi\sqrt{1 - x^2}]$ for which $u$ is not an atom of the distribution $\mathbb{P}_x(Y'(0) \in \cdot)$, we have that $$ \mathbb{P}_x\left(\frac{Y(r) - Y(0)}r > u \right) \xrightarrow[r \to 0]{} \mathbb{P}_x(Y'(0) > u). $$ Therefore by the dominated convergence theorem, \eqref{E:b-c} converges to $$ \int_b^c \int_{q\sqrt{1 - x^2}}^{\pi\sqrt{1 - x^2}} \mathbb{P}_x(Y'(0) > u) du dx = \int_{b}^c \mathbb{E}_{x} (Y'(0) - \sqrt{1 - x^2}q)^+ dx \qquad \;\text{as}\; \quad r \to 0. $$ The equality above again uses that $|Y'(0)| \le \pi\sqrt{1- Y^2(0)}$ almost surely, since $Y$ is $\pi\sqrt{1- h^2}$-Lipschitz. Combining this with \eqref{E:want-1}, \eqref{E:J-+-da}, and \eqref{E:want-2} proves \eqref{E:deriv-integrated}. Now define \[ d(Y) = \frac{Y'(0)}{\sqrt{1 - Y^2(0)}}. \] By \eqref{E:deriv-integrated}, for almost every $a \in (-1, 1)$, for every $q \in \mathbb{Q} \cap (\pi, \pi)$, we have that $$ \sqrt{1 - a^2}\mathbb{E}_{a} \left(d(Y) - q\right)^+ = \mathbb{E}_{a} \left(Y'(0) - \sqrt{1 - a^2} q\right)^+ = \sqrt{1- a^2}D_\mu^+(q). $$ Therefore by continuity in $s$ of the functions $\mathbb{E}_{a} \left(d (Y) - s\right)^+$ and $D_\mu^+(s)$, we have that \begin{equation} \label{E:daY} \mathbb{E}_{a} \left(d (Y) - s\right)^+ = D_\mu^+(s) \qquad \text{ for all } s \in (-\pi, \pi). \end{equation} Now, $D_\mu(s)$ is Lipschitz and hence differentiable almost everywhere (Lemma \ref{L:convex}), so we can differentiate both sides of the above equation to get that for almost every $s \in (-\pi, \pi)$, that $$ \mathbb{P}_a(d(Y) > s) = \mu(s, \infty). $$ Finally, for almost every $a \in (-1, 1)$, we have that $$ \mathbb{P}_a(d(Y) \in [-\pi, \pi]) = \mu[\pi, \pi] = 1 $$ since $Y$ is almost surely $\pi\sqrt{1 - h^2}$-Lipschitz (Theorem \ref{T:bounded-speed}). Therefore $\mathbb{P}_a(d(Y) \in \cdot) = \mu$. \end{proof} \subsection{The max-height distribution and the uniqueness of $Y$} As an application of Proposition \ref{P:local-global-deriv}, we can find $\mathbb{P}(m(Y) > a)$ for all $a \in [0, 1]$. This allows us to deduce the uniqueness of $Y$. \begin{theorem} \label{T:max-height} Let $Y$ be any subsequential limit of $Y_n$. For all $a \in [0, 1]$, we have that $ \mathbb{P}(m(Y) > a) = \sqrt{1 - a^2}$. That is, $m(Y) \stackrel{d}{=} \sqrt{1- U^2}$ for a uniform random variable $U$ on $[0, 1]$. \end{theorem} \begin{proof} Let $V(Y)$ be the total variation of $Y$ on the interval $[0, 1]$. We have that \begin{align*} \mathbb{E} V(Y) &= \mathbb{E} \int_0^1 |Y'(t)|dt = \mathbb{E} \left(\frac{|Y'(0)|}{\sqrt{1-Y^2(0)}}\right) \sqrt{1 - Y^2(0)} = 2 \mathbb{E} \sqrt{1 - U^2}, \end{align*} where $U$ is a uniform random variable on $[0, 1]$. Here the second equality follows from time stationarity of $Y$ (Proposition \ref{P:Y-symmetries} (i)). The third equality follows by Proposition \ref{P:local-global-deriv}, the fact that the first moment of $\mu$ is $2$ (Lemma \ref{L:expect-2}), and the fact that $Y(0)$ is uniformly distributed (Theorem \ref{T:bounded-speed}). \medskip Moreover, the characterization of minimal flux paths (Theorem \ref{T:path-set}) implies that $ V(Y) = 2 \mathbb{E} m(Y), $ so $\mathbb{E} m(Y) = \mathbb{E} \sqrt{1 - U^2}$. Finally, the bound in Lemma \ref{L:max-heights} implies that $$ \mathbb{P}(m(Y) > a) \le \frac{D_\mu(0)}2 \sqrt{1 - a^2} = \sqrt{1- a^2} = \mathbb{P}(\sqrt{1 - U^2} > a), $$ so the random variable $m(Y)$ is stochastically dominated by $\sqrt{1 - U^2}$. Hence $m(Y) \stackrel{d}{=} \sqrt{1 - U^2}$. \end{proof} Now we can prove the existence of a unique limit $Y$ of $Y_n$. First extend the paths $h_m$ defined in Theorem \ref{T:path-set} to paths $h_m:[0, \infty) \to [-1, 1]$ by letting $h_m(t) = -h_m(t + 1)$ for all $t > 1$. \begin{theorem} \label{T:Y-unique} The sequence $Y_n$ has a distributional limit $Y$ given by $$ Y(t) = h_{\sqrt{1 - U^2}}(V + t). $$ Here $U$ is a uniform random variable on $[0, 1]$ and $V$ is uniform on $[0, 2]$, independent of $U$. \end{theorem} \begin{proof} By Theorem \ref{T:unique-2} and Theorem \ref{T:particle-crossings} (iii), any subsequential limit $Y$ of $Y_n$ is supported on the set of paths $$ \{h_m (\cdot + u) : m \in [0, 1], u \in [0, 2]\}, $$ so we can write $ Y(\cdot) = h_M(\cdot + V), $ for a pair of random variables $(M, V)$. By Theorem \ref{T:max-height}, $M = \sqrt{1 - U^2}$, for a uniform random variable $U$ on $[0, 1]$. By the time stationarity of $Y$ (Proposition \ref{P:Y-symmetries} (i)), for any $t \in [0, 2]$, we have that $$ h_M(\cdot + t + V) \stackrel{d}{=} h_M(\cdot + V), $$ which in turn implies that $(M, V) \stackrel{d}{=} (M, V + t \Mod 2 )$ for any $t \in [0, 2]$. This implies that $V$ has uniform distribution, and that $M$ is independent of $V$. \end{proof} \section{The integral transform formula} \label{S:integral-formula} Let $\mu_+$ be the pushforward of the local speed distribution $\mu$ under the map $x \mapsto |x|$. In this section we prove an integral transform formula for $\mu_+$. This transform formula allows us to identify $\mu$ as the arcsine distribution on $[-\pi, \pi]$. Once we know $\mu$, we can find the set of minimal flux paths $h_m$ introduced in Theorem \ref{T:path-set}, and then in turn identify $Y$. \medskip To find the integral transform formula, we will calculate $\mathbb{P}(m(Y) > k)$ in two different ways. We first give a heuristic explanation of how to do this when the local speed distribution $\mu$ has no atoms. By Theorem \ref{T:max-height}, we have that $$ \mathbb{P}(m(Y) > k) = \sqrt{1 - k^2}. $$ We can also calculate $\mathbb{P}(m(Y) > k)$ by integrating the marginal probabilities $\mathbb{P}_a(m(Y) > k)$. This gives that \begin{equation} \label{E:heur1} \sqrt{1 - k^2} = \frac{1}2 \int_{-1}^1 \mathbb{P}_a(m(Y) > k)da = 1 - k + \frac{1}2 \int_{-k}^k \mathbb{P}_a(m(Y) > k)da. \end{equation} Now we want to find an expression for $\mathbb{P}_a(m(Y) > k)$ when $|a| < k$. By the ordering on minimal flux paths, if $Y(0) = a$, then $m(Y) > k$ if and only if $|Y'(0)|$ is greater than some threshold value. Therefore for some $s_{a, k} \in \mathbb{R}$, we have that \begin{equation} \label{E:heur2} \mathbb{P}_a(m(Y) > k) = \mathbb{P}_a (|Y'(0)| > \sqrt{1 - a^2}s_{a, k}) = \mu_+(s_{a, k}, \infty). \end{equation} The final equality here follows from Proposition \ref{P:local-global-deriv}. To find $s_{a, k}$, we calculate $\mathbb{P}(m(Y) > k | \; m(Y) > a)$. If we let \[ T_a = \inf \{t \in [0, 1] : Y(t) = a\}, \] then we should have that $$ \mathbb{P}(m(Y) > k | \; m(Y) > a) = \mathbb{P}(|Y'(T_a)| > \sqrt{1 - a^2}s_{a, k}). $$ Now, the ``amount of time" that $Y$ spends at height $a$ is inversely proportional to its speed at that location. Therefore the distribution of $|Y'(T_a)|$ should be a size-biased version of the distribution of $Y'(0)$ given that $Y(0) = a$. Hence, $$ \mathbb{P}(|Y'(T_a)| > \sqrt{1 - a^2}s_{a, k}) = \hat{\mu}_+(s_{a, k}, \infty), $$ where $\hat{\mu}_+$ is the size-biased distribution of $\mu_+$ (we define this formally in the next paragraph). Finally, we can also calculate $\mathbb{P}(m(Y) > k | \; m(Y) > a)$ using Theorem \ref{T:max-height}. This gives that $$ \hat{\mu}_+(s_{a, k}, \infty) = \mathbb{P}(m(Y) > k | \; m(Y) > a) = \frac{\mathbb{P}(m(Y) > k)}{\mathbb{P}(m(Y) > a)} = \frac{\sqrt{1 - k^2}}{\sqrt{1 - a^2}}. $$ We can combine this with \eqref{E:heur2} and \eqref{E:heur1} to get an integral transform formula involving the function $r_{\mu_+}(x) = S(\hat{S}^{-1}(x))$, where $S$ and $\hat{S}$ are the survival functions of $\mu_+$ and $\hat{\mu}_+$, respectively. \medskip We now precisely define everything that is needed to state the integral transform formula. For a probability measure $\nu$ on $(0, \infty)$ with finite mean, define the {\bf size-biased distribution} $\hat{\nu}$ on $(0, \infty)$ by the Radon-Nikodym derivative formula $$ \frac{d\hat{\nu}}{d \nu}(x) = \frac{x}{\int_0^\infty x d \nu(x)}. $$ In order to define the integral transform when $\mu$ has atoms, we define the \textbf{extended survival function} $S:\mathbb{R} \times [0,1] \to [0, 1]$ of a probability measure $\nu$ by $$ S(x, q) = \nu(x, \infty) + (1 - q) \nu(x). $$ $S$ is a non-increasing, continuous function in the lexicographic ordering on $\mathbb{R} \times [0,1]$. $S$ can be thought of as the survival function in the lexicographic ordering for $\nu \times \mathcal{L}$, where $\mathcal{L}$ is uniform measure on $[0, 1]$. Now define the \textbf{ size-bias ratio function} $r_\nu(x):[0, 1] \to [0, 1]$ of a probability measure $\nu$ on $(0, \infty)$ with finite mean by \begin{equation*} r_\nu(x) = S(\hat{S}^{-1}(x)), \end{equation*} where $\hat{S}$ is the extended survival function of $\hat{\nu}$ and $S$ is the extended survival function of $\nu$. Here the inverse function is given by $$ \hat{S}^{-1}(x) = \sup \{(y, q) \in \mathbb{R} \times [0,1] : \hat{S}(y, q) = x \}, $$ where the supremum is taken with respect to the lexicographic ordering. When $\nu$ has no atoms, we can define $r_\nu$ in terms of the usual survival functions. \begin{prop}[The integral transform] \label{P:integral-equality} Let $\mu_+$ be the pushforward of the measure $\mu$ under the map $f(x) = |x|$. For every $k \in (0, 1)$, we have that \begin{equation} \label{E:integral-H} 1 - k + \int_0^k r_{\mu_+} \left(\frac{\sqrt{1 - k^2}}{\sqrt{1- x^2}}\right) dx = \sqrt{1-k^2}. \end{equation} \end{prop} To prove Proposition \ref{P:integral-equality}, we first establish that $\mu(0) = 0$. \begin{lemma} \label{L:X-atom} $\mu(0) = 0$. \end{lemma} \begin{proof} By Lemma \ref{L:Y-sym-type}, $\mathbb{P} \left(\frac{Y'(0)}{\sqrt{1 - Y^2(0)}} = 0 \right) = 0$. By Proposition \ref{P:local-global-deriv}, this is equal to $\mu(0)$. \end{proof} Now let $m < 1$, and let $h_m$ be as in Theorem \ref{T:path-set}. For any $a \in [0, m]$, define $$ s_{a, m} = \frac{1}{\sqrt{1 - a^2}} \liminf_{r \to 0} \frac{|h_m(t^* + r) - h_m(t^*)|}{r}, $$ where $t^*$ is any point where $h_m(t^*) = a$. Note that $s_{a, m}$ is independent of $t^*$ by the symmetry of minimal flux paths (Theorems \ref{T:path-set} (i)). We now prove an integral formula relating the speeds $s_{a, m}$ to the local speed distribution. \begin{lemma} \label{L:bias-unbias} Let $\nu$ be the law of $m(Y)$. For almost every $a \in [0, 1)$, for every $k \in (a, 1)$ there exists a constant $q_{a, k} \in [0, 1]$ such that \begin{equation} \begin{split} \mathbb{P}_a(m(Y) > k) &= \frac{1}{\sqrt{1 - a^2}} \int \mathbbm{1} \Big(s_{a, m} > s_{a, k} \;\; \;\text{or}\; \;\; s_{a, m} = s_{a, k}, \; m > k\Big) \frac{2}{s_{a, m}} d\nu(m) \\ \label{E:s-a-k} &= \mu_+(s_{a, k}, \infty) + q_{a, k}\mu_+(s_{a, k}). \end{split} \end{equation} Moreover, $\mathbb{P}_a(m(Y) > k)$ is a continuous function of $k \in (a, 1)$ for almost every $a \in [0, 1)$. \end{lemma} \begin{proof} Let $h_m$ be as in Theorem \ref{T:path-set}, and let $a < b \in [0, 1)$. We first compute the amount of time that $h_m$ spends in the interval $[a, b]$. Define $h_m^{-1}:[0, m] \to [0, 1/2]$ by $$ h_m^{-1}(x) = \inf \{t : h_m(t) = x \}. $$ By the strict monotonicity and symmetry of $h_m$ (Theorem \ref{T:path-set} (i), (ii)), we can write \begin{equation} \label{E:L-h} \mathcal{L}\{t: |h_m(t)| \in [a, b]\} = 2[h_m^{-1}(b) - h_m^{-1}(a)], \end{equation} where $\mathcal{L}$ is Lebesgue measure on $[0, 1]$. Now by the concavity of $h_m$ (Lemma \ref{L:concave-paths}), the inverse $h_m^{-1}(x)$ is almost everywhere differentiable with derivative $2/[s_{x, m} \sqrt{1 - x^2}]$. Therefore the left hand side of \eqref{E:L-h} is equal to $$ \int_a^{b} \frac{2}{s_{x, m} \sqrt{1 - x^2}}dx. $$ Now letting $U$ be a uniform random variable on $[0, 1]$ that is independent of $Y$, for any $a < b < k \in [0, 1)$, we have that \begin{equation*} \begin{split} \mathbb{P}\big(m(Y) > k \;\text{and}\; Y(0) \in [a, b]\big) &= \frac{1}2\mathbb{P}\big(m(Y) > k \;\text{and}\; |Y(U)| \in [a, b]\big) \\ &= \frac{1}2 \int_k^1 \mathcal{L}\{t: |h_m(t)| \in [a, b]\} d\nu(m) \\ &= \int_a^{b} \frac{1}{\sqrt{1 - x^2}} \int_k^1 \frac{1}{s_{x, m}} d\nu(m)dx. \end{split} \end{equation*} The first equality above follows by the time-stationarity and symmetry of $Y$ (Proposition \ref{P:Y-symmetries} (i) and (ii)). The second equality follows since $Y$ is supported on shifts of the minimal flux paths $h_m$ (Theorem \ref{T:unique-2}). This implies that for almost every pair $a < k \in [0, 1)$, that \begin{equation} \label{E:f-amid} \mathbb{P}_a(m(Y) > k) = \frac{1}{\sqrt{1 - a^2}} \int_k^1 \frac{2}{s_{a, m}} d\nu(m). \end{equation} Now by the ordering on minimal flux paths (Theorem \ref{T:unique-2}(iv)), we have that \begin{equation} \label{E:speed-mono} \text{if} \;\; s_{a, m(Y)} < s_{a, k}, \qquad \text{then } m(Y) < k. \end{equation} This allows us to rewrite \eqref{E:f-amid} to get that \begin{equation} \label{E:f-amid-2} \mathbb{P}_a(m(Y) > k) = \frac{1}{\sqrt{1 - a^2}} \int \mathbbm{1} \Big(s_{a, m} > s_{a, k} \;\; \;\text{or}\; \;\; s_{a, m} = s_{a, k}, m > k\Big) \frac{2}{s_{a, m}} d\nu(m) \end{equation} for almost every $a < k \in [0, 1)$. Moreover, for almost every $a \in [0, 1)$, we have that $$ |Y'(0)| = \sqrt{1 - a^2}s_{a, m(Y)} \qquad \mathbb{P}_a\text{-almost surely}. $$ This follows by time stationarity of $Y$, since $Y$ is almost everywhere differentiable. Also, $Y'(0)/\sqrt{1- a^2}$ has distribution $\mu$ for almost every $a$ (Proposition \ref{P:local-global-deriv}). Therefore \eqref{E:speed-mono} implies that for almost every $a \in [0, 1)$, for every $k \in (a, 1)$ there exists a constant $q_{a, k} \in [0, 1]$ such that \begin{equation} \label{E:Pa-mY} \mathbb{P}_a(m(Y) > k) = \mu_+(s_{a, k}, \infty) + q_{a, k}\mu_+(s_{a, k}). \end{equation} Now let $a$ be such that \eqref{E:Pa-mY} holds for every $k \in (a, 1)$ and \eqref{E:f-amid} and \eqref{E:f-amid-2} hold for almost every $k \in (a, 1)$. \medskip By concavity of minimal flux paths (Lemma \ref{L:concave-paths}), $s_{a, m} > 0$ whenever $m > a$. Therefore since $\nu$ has a Lebesgue density by Theorem \ref{T:max-height}, the right hand side of \eqref{E:f-amid} is continuous and non-increasing. Since both sides of \eqref{E:f-amid-2} are also non-increasing, and are equal to the right hand side of \eqref{E:f-amid} for almost every $k \in (a, 1)$, they must be equal for every $k \in (a, 1)$. Combining this with \eqref{E:Pa-mY} proves the lemma. \end{proof} \begin{lemma} \label{L:equal-measure} Let $\mathfrak{s}_a$ be the law of $\mathbb{P}(s_{a, m(Y)} \in \cdot \; | m(Y) > a)$, and define the measure $\bar{\mathfrak{s}}_a$ by the Radon-Nikodym formula $$ \frac{2}{s} d\mathfrak{s}_a(s) = d\bar{\mathfrak{s}}_a(s). $$ Then for almost every $a \in [0, 1)$, we have that $\bar{\mathfrak{s}}_a = \mu_+$. In particular, for such $a$, we have that \begin{equation} \label{E:q-ak} q_{a, k}\mu_+(s_{a, k}) = \frac{2}{s_{a, k}\sqrt{1 - a^2}} \int \mathbbm{1} \Big(s_{a, m} = s_{a, k}, m > k\Big)d\nu(m). \end{equation} \end{lemma} \begin{proof} Let $a$ be such that \eqref{E:s-a-k} holds for every $k \in (0, 1)$ and such that $\mathbb{P}_a(m(Y) > k)$ is continuous. Further assume that $$ \mathbb{P}_a\left( Y'(0) \text{ exists and is non-zero }, m(Y) < 1, J(Y) = 1 \right) = 1. $$ These conditions hold for almost every $a \in [0, 1)$ (Proposition \ref{P:local-global-deriv}, Lemma \ref{L:bias-unbias}). Noting that $\sqrt{1 - a^2} = \mathbb{P}(m(Y) > a)$ by Theorem \ref{T:max-height}, equation \eqref{E:s-a-k} implies that for every $k \in (a, 1)$, there exists a $p_{a, k} \in [0, 1]$ such that \begin{equation} \label{E:comp} \mathbb{P}_a(m(Y) > k) = \bar{\mathfrak{s}}_a(s_{a, k}, \infty) + p_{a, k}\bar{\mathfrak{s}}_a(s_{a, k}) = \mu_+(s_{a, k}, \infty) + q_{a, k}\mu_+(s_{a, k}). \end{equation} Now, since $Y'(0) \ne 0$, $\mathbb{P}_a$-almost surely, the concavity of minimal flux paths (Lemma \ref{L:concave-paths}) implies that $\mathbb{P}_a(m(Y) = a) = 0$. Moreover, since $m(Y) < 1$, $\mathbb{P}_a$-almost surely, we have that $\mathbb{P}_a(m(Y) \in (a, 1)) = 1$, and so \begin{equation} \label{E:mY} \lim_{k \to 1} \mathbb{P}_a(m(Y) > k) = 0 \quad \;\text{and}\; \quad \lim_{k \to a} \mathbb{P}_a(m(Y) > k) = 1. \end{equation} Therefore, \eqref{E:comp} and the continuity of $\mathbb{P}_a(m(Y) > k)$ implies that \begin{equation} \label{E:full} \mu_+ \left( s_{a, k} : k \in (0, 1) \right) = \bar{\mathfrak{s}}_a \left( s_{a, k} : k \in (0, 1) \right) = 1. \end{equation} Now fix $k \in (a, 1)$. Let $$ k^* = \inf \{\ell \in [k, 1] : s_{a, k} < s_{a, m} \;\; \text{ for all } \;\;m > \ell \}. $$ Equation \eqref{E:comp} and the continuity of $\mathbb{P}_a(m(Y) > k)$ then implies that $$ \mathbb{P}_a(m(Y) > k^*) = \bar{\mathfrak{s}}_a(s_{a, k}, \infty) = \mu_+(s_{a, k}, \infty). $$ Combining this with \eqref{E:full} proves that $\mu_+ = \bar{\mathfrak{s}}_a$. Equation \eqref{E:q-ak} follows by using that $\mu_+ = \bar{\mathfrak{s}}_a$ to simplify equation \eqref{E:s-a-k}. \end{proof} \begin{proof}[Proof of Proposition \ref{P:integral-equality}] Fix $a$ so that the conclusion of Lemma \ref{L:equal-measure} holds, and let $k > a$. By Theorem \ref{T:max-height} and the ordering on minimal flux paths, we can write \begin{align*} \nonumber\mathbb{P}(m(Y) > k | \; m(Y) > a) &= \frac{1}{\sqrt{1 - a^2}} \left[\int \mathbbm{1}(s_{a, m} > s_{a, k} ) d\nu(m) + \int \mathbbm{1}(s_{a, m} = s_{a, k}, m > k) d\nu(m) \right]. \end{align*} By Lemma \ref{L:equal-measure}, we can rewrite the first integral above in terms of $\mathfrak{s}_a$, and then in terms of $d\mu_+$. This gives that $$ \frac{1}{\sqrt{1 - a^2}}\int \mathbbm{1}(s_{a, m} > s_{a, k} ) d\nu(m) = \int \mathbbm{1}(s > s_{a, k}) d\mathfrak{s}_a(s) = \int \mathbbm{1}(s > s_{a, k})\frac{s}2d\mu_+(s). $$ We can also rewrite the second integral using \eqref{E:q-ak}. This implies that $$ \mathbb{P}(m(Y) > k | \; m(Y) > a) = \int \mathbbm{1}(s > s_{a, k})\frac{s}2d\mu_+(s) + \frac{s_{a, k}}2 q_{a, k} \mu_+(s_{a, k}). $$ By Lemma \ref{L:expect-2}, we can recognize $\frac{s}{2}$ as the Radon-Nikodym derivative of the size-biased random variable $\hat{\mu}_+$ with respect to $\mu_+$, proving that $$ \frac{\sqrt{1 - k^2}}{\sqrt{1 - a^2}} = \mathbb{P}(m(Y) > k | m(Y) > a) = \hat{\mu}(s_{a, k}, \infty) + q_{a, k}\hat{\mu}(s_{a, k}) $$ for almost every $a < k \in [0, 1)$. Now, Lemma \ref{L:bias-unbias} allow us to conclude that \begin{align*} \mathbb{P}_a(m(Y) > k) &= S(s_{a, k}, 1 - q_{a, k}) = S\left(\hat{S}^{-1}\left(\frac{\sqrt{1- k^2}}{\sqrt{1- a^2}}\right)\right) \end{align*} for almost every $a < k \in [0, 1)$. Combining this with the symmetry of $Y$ (Proposition \ref{P:Y-symmetries}) and Theorem \ref{T:max-height} implies that \[ \sqrt{1 - k^2} = \mathbb{P}(m(Y) > k) = \int_{0}^1 \mathbb{P}_a(m(Y) > k) da = 1 - k + \int_0^k r_{\mu_+}\left(\frac{\sqrt{1- k^2}}{\sqrt{1- a^2}}\right) da. \qedhere \] \end{proof} \section{The weak trajectory limit} \label{S:transform} In this section we show that the integral transform in Proposition \ref{P:integral-equality} determines the local speed distribution. We begin with two basic lemmas about size-bias ratio functions. \begin{lemma} \label{L:loc-lip} For any probability distribution $\nu$ on $(0, \infty)$ with finite first moment $m$, the size-bias ratio function $r_\nu$ is locally Lipschitz on $[0, 1)$ and continuous on $[0, 1]$. \end{lemma} \begin{proof} Let $\pi_1(x, y) = x.$ By calculus, when $x \in (0, 1)$ we have that $$ \partial_+ r_\nu(x) = \lim_{h \to 0+} \frac{m}{\pi_1(\hat{S}^{-1}(x + h))} \qquad \;\text{and}\; \qquad \partial_- r_\nu(x) = \lim_{h \to 0+} \frac{m}{\pi_1(\hat{S}^{-1}(x - h))}. $$ The first equation also holds as $x = 0$. As $\pi_1(\hat{S}^{-1}(y))$ is a decreasing function of $y$, and strictly positive for all $y \in [0, 1)$, this shows that $r_\nu$ is locally Lipschitz on $[0, 1)$. It is straightforward to check that $r_\nu$ is continuous at $1$. \end{proof} \begin{lemma} \label{L:rX-det} Suppose that $\nu_1$ and $\nu_2$ are probability measures on $(0, \infty)$ with the same first moment, such that $r_{\nu_1} = r_{\nu_2}$. Then $\nu_1 = \nu_2$. \end{lemma} \begin{proof} We prove the contrapositive. Let $\nu_1$ and $\nu_2$ be measures with the same first moment, and suppose that $\nu_1 \ne \nu_2$. Since $\nu_1$ and $\nu_2$ have the same first moment, $\hat{\nu}_1 \ne \hat{\nu}_2$, so for some value of $(y, q) \in (0, \infty) \times [0, 1]$, we have that $\hat{S}_{\nu_1} (y, q) \ne \hat{S}_{\nu_2} (y, q)$. Since the functions $\hat{S}_{\nu_1}(y, \cdot)$ and $\hat{S}_{\nu_2}(y, \cdot)$ are linear for any fixed value of $y$, this implies that $\hat{S}_{\nu_1} (y, r) \ne \hat{S}_{\nu_2} (y, r)$ for some $r \in \{0, 1\}$. Moreover, since $$ \hat{S}_{\nu_i}(y, 1) = \sup \{\hat{S}_{\nu_i}(z, 0) : z > y \}, \qquad i = 1, 2, $$ we can conclude that $\hat{S}_{\nu_1} (z, 0) \ne \hat{S}_{\nu_2} (z, 0)$ for some $z \in (0, \infty)$. Without loss of generality, assume that $a_1 := \hat{S}_{\nu_1} (z, 0) > a_2 :=\hat{S}_{\nu_2} (z, 0)$. Therefore using the notation of the previous lemma, letting $b = (a_1 + a_2)/2$, we have that $$ \lim_{h \to 0^+} \pi_1(\hat{S}_{\nu_1}^{-1}(b + h)) \ge z, \qquad \text{whereas} \qquad \lim_{h \to 0^+} \pi_1(\hat{S}_{\nu_1}^{-1}(b + h)) < z. $$ Since $b \in (0, 1)$, the derivative computation in the previous lemma combined with the fact that $\nu_1$ and $\nu_2$ have the same first moment implies that $r_{\nu_1} \ne r_{\nu_2}$. \end{proof} Now let $\mathcal{X}$ be the space of continuous functions from $[0, 1] \to \mathbb{R}$ that are locally Lipschitz on $[0, 1)$. Define an integral transform $H$ on $\mathcal{X}$ by \begin{equation*} H(r)(k) = \int_0^k r \left(\frac{\sqrt{1 - k^2}}{\sqrt{1- x^2}}\right) dx, \end{equation*} for $k \in [0, 1]$. By Lemma \ref{L:loc-lip}, any linear combination of size-bias ratio functions is in $\mathcal{X}$, so if the integral transform $H$ is injective on $\mathcal{X}$, then $H(r_{\mu_+})$ determines $r_{\mu_+}$. \begin{lemma} \label{L:injective-transform-lip} $H$ is injective on $\mathcal{X}$. \end{lemma} \begin{proof} Let $r \in \mathcal{X}, r \ne 0$. We will show that $H(r) \ne 0$. Without loss of generality, we may assume that $$ \max_{x \in [0, 1]} r(x) = \delta > 0. $$ Letting $u = \frac{\sqrt{1 - k^2}}{\sqrt{1- x^2}}$ and $y = \sqrt{1 -k^2}$, we have that \begin{equation*} H(r)(\sqrt{1-y^2}) = \int_y^1 r(u) \frac{y^2}{u^2\sqrt{u^2 - y^2}}du. \end{equation*} For $y \in (0, 1]$, we have that \begin{equation} \label{E:H-bar} \close{H}(r)(y) := \frac{H(r)(\sqrt{1 - y^2})}{y^2} = \int_y^1 \frac{r(u)}{u^2\sqrt{u^2 - y^2}}du. \end{equation} It suffices to show that $\close{H}(r)(y) \ne 0$ for some $y \in (0, 1]$. Observe that if $r(1) > 0$, then there would exist a $\gamma > 0$ such that $r(x) > 0$ for all $x > \gamma$, so $\close{H}(r)(\gamma) > 0$. Also, if $r(0) > 0$, then $\close{H}(r)(y) \to \infty$ as $y \to 0$. Therefore there must exist $y \in (0,1)$ be such that $r(y) = \delta$. Since $r$ is locally Lipschitz on $[0, 1)$, we can find $k, \epsilon > 0$ such that $r(x) \ge \delta - k \epsilon$ for all $x \in [y, y + \epsilon]$. Therefore \begin{align} \label{E:de-k-bd} \int_y^{y+\epsilon}\frac{r(u)}{u^2\sqrt{u^2 - y^2}}du \ge \frac{(\delta - k \epsilon)\sqrt{\epsilon^2 + 2y\epsilon}}{y^2(y + \epsilon)}. \end{align} Now consider the difference between $\close{H}(r)(y)$ and $\close{H}(r)(y + \epsilon)$. We have \begin{align} \nonumber \close{H}&(r)(y) - \close{H}(r)(y + \epsilon) \\ \nonumber &= \int_y^{y+\epsilon}\frac{r(u)}{u^2\sqrt{u^2 - y^2}}du + \int_{y+ \epsilon}^1 \frac{r(u)}{u^2\sqrt{u^2 - y^2}}du - \int_{y+ \epsilon}^1 \frac{r(u)}{u^2\sqrt{u^2 - (y+ \epsilon)^2}}du \\ \nonumber &\ge \int_y^{y+\epsilon}\frac{r(u)}{u^2\sqrt{u^2 - y^2}}du + \delta \int_{y + \epsilon}^1 \left[\frac{1}{u^2\sqrt{u^2 - y^2}} - \frac{1}{u^2\sqrt{u^2 - (y + \epsilon)^2}}\right]du \\ \label{E:int-alm} &\ge \frac{(\delta - k \epsilon) \sqrt{\epsilon^2 + 2y\epsilon}}{y^2(y + \epsilon)} + \delta \left[\frac{(y + \epsilon)^2\sqrt{1-y^2} -y^2\sqrt{1-(y + \epsilon)^2}- (y + \epsilon)\sqrt{\epsilon^2 + 2y\epsilon}}{(y + \epsilon)^2y^2}\right] . \end{align} In above calculation the first inequality comes from the fact that $r(x) \le \delta$ for all $x$, and the observation that $$ \frac{1}{u^2\sqrt{u^2 - y^2}} < \frac{1}{u^2\sqrt{u^2 - (y + \epsilon)^2}} $$ for all $u \in (y + \epsilon, 1)$. The second equality follows by integration and plugging in the bound in \eqref{E:de-k-bd}. Now expanding in $\epsilon$ about $\epsilon = 0$, we get that \eqref{E:int-alm} is equal to \begin{align*} \frac{(2-y^2)\delta\epsilon}{y^3\sqrt{1-y^2}} + O(\epsilon^{3/2}). \end{align*} This is strictly greater than 0 for small enough $\epsilon$, so $ \close{H}(r)(y) - \close{H}(r)(y + \epsilon) \ne 0 $ for such $\epsilon$. Hence $\close{H}(r)(x) \ne 0$ for some $x \in [0, 1)$. \end{proof} \begin{prop} \label{P:arcsin-en} Let $\mathfrak{arc}$ be the arcsine distribution on $[-\pi, \pi]$, and let $\mathfrak{arc}_+$ be the pushforward of $\mathfrak{arc}$ under the map $x \mapsto |x|$. Then for every $k \in [0, 1]$, we have that \begin{equation} \label{E:integral-H-2} 1 - k + \int_0^k r_{\mathfrak{arc}_+} \left(\frac{\sqrt{1 - k^2}}{\sqrt{1- x^2}}\right) dx = \sqrt{1-k^2}. \end{equation} \end{prop} \begin{proof} The distribution $\mathfrak{arc}$ has density $(\pi \sqrt{\pi^2 - x^2})^{-1}$ on $[-\pi, \pi]$. From this, we can calculate that \begin{align*} r_{\mathfrak{arc}_+}(y) = 1 - \frac{2}{\pi}\arcsin(\sqrt{1-y^2}). \end{align*} We now use the connection between the arcsine distribution and the Archimedean measure $\mathfrak{Arch}$ to help evaluate the integral in \eqref{E:integral-H-2}. Let $(X_1, X_2) \sim \mathfrak{Arch}$ and let $$ \mathcal{A}(t) = X_1 \cos (\pi t) + X_2 \sin(\pi t) $$ be the Archimedean path. Writing $\mathbb{P}(m(\mathcal{A}) > k) = \mathbb{P}(X_1^2 + X_2^2 > k^2)$ in both polar and Cartesian coordinates, we get that $$ \int_k^1 \frac{rdr}{\sqrt{1 - r^2}} = 4 \int_{0}^1 \int_{\sqrt{k^2 - y^2} \vee 0}^{\sqrt{1 - y^2}} \frac{dxdy}{2 \pi \sqrt{1 - x^2 - y^2}}. $$ The left hand side can easily be evaluated as $\sqrt{1 - k^2}$. The right hand side is equal to \begin{align*} (1 - k) + \int_{0}^k \frac{2}{\pi} \left(\frac{\pi}{2} - \arcsin\left(\frac{\sqrt{k^2 - y^2}}{\sqrt{1- y^2}}\right)\right)dy = (1- k) + \int_{0}^k r_\mathfrak{arc}\left(\frac{\sqrt{1 - k^2}}{\sqrt{1- y^2}}\right) dy. \qquad \qedhere \end{align*} \end{proof} We can now prove Theorem \ref{T:main-2}, and in turn use that to prove Theorem \ref{T:weak-limit}. \begin{proof}[Proof of Theorem \ref{T:main-2}] By Proposition \ref{P:integral-equality}, Lemma \ref{L:injective-transform-lip} and Proposition \ref{P:arcsin-en}, we have that $r_{\mu_+} = r_{\mathfrak{arc}_+}$. Moreover, the first moment of $\mathfrak{arc}_+$ is $2$. By Lemma \ref{L:expect-2}, this matches the first moment of $\mu_+$. Therefore by Lemma \ref{L:rX-det}, $\mu_+ = \mathfrak{arc}_+$. Finally, symmetry of both measures implies that $\mu = \mathfrak{arc}$. \end{proof} \begin{proof}[Proof of Theorem \ref{T:weak-limit}] Fix $m > 0$. Let $g_m(t) = m \sin (\pi t)$, and let $g_m^{-1}$ be the inverse of $g_m$ on the interval $[0, 1/2]$. Let $$ s_m(t) = \frac{g_m'(t)}{\sqrt{1 - g_m^2(t)}} $$ be the local speed of $g_m$. We can calculate that \begin{align} \nonumber J(g_m) = 2J(g_m; [0, 1/2]) &= \int_0^{1/2} D_\mu(s_m(t)) \sqrt{1 - g_m^2(t)}dt \\ \label{E:s-h} &= \int_0^m \frac{D_\mu(s_m(g_m^{-1}(x))}{s_m(g_m^{-1}(x))}dx. \end{align} Here we have made the substitution $x = g_m(t)$ to go from the first to the second line. Now using Theorem \ref{T:main-2}, we can calculate $$ D_\mu(c) = D_\mathfrak{arc}(c) = \frac{2}{\pi} \left(c \arcsin \left(\frac{c}{\pi}\right) + \sqrt{\pi^2 - c^2}\right). $$ From here we can use that $s_m(g_m^{-1}(x)) = \frac{\pi \sqrt{m^2 - x^2}}{\sqrt{1 - x^2}}$ to compute that \begin{equation*} J(g_m) = \int_0^m \frac{2}{\pi} \left[ \frac{\sqrt{1-m^2}}{\sqrt{m^2 - x^2}} + \arcsin\left(\frac{\sqrt{m^2 - x^2}}{\sqrt{1- x^2}}\right)\right]dx \end{equation*} The first part of this integral can be easily evaluated, and the second part can be evaluated by comparing with the integral in the final line of the proof of Proposition \ref{P:arcsin-en}. Putting this all together yields that $J(g_m) = 1$, as desired. Therefore by Theorem \ref{T:Y-unique}, we can write $$ Y(t) = \sqrt{1 - V^2}\sin(\pi t + 2 \pi U), $$ where $U$ and $V$ are independent uniform random variables on $[0, 1]$. This is the Archimedean path. \end{proof} \subsection{The empirical distribution of trajectories} \label{SS:slightly} By a compactness argument, we can immediately prove a slightly stronger version of Theorem \ref{T:weak-limit}. This will allow us to conclude Theorem \ref{T:subnetwork}. This theorem will also be necessary for establishing the stronger limits in Sections \ref{S:strong-limit} and \ref{S:geom-limit}. \medskip For a Polish space $S$, let $\mathcal{M}(S)$ be the space of probability measures on $S$ with the topology of weak convergence. Note that $\mathcal{M}(S)$ is itself a Polish space. Recall the notation $\sigma_G$ introduced in Section \ref{S:intro}. For a fixed sorting network $\sigma$, let $\nu_\sigma \in \mathcal{M}(D)$ be uniform measure on the set $\{\sigma_G(i, \cdot)\}_{i \in \{1, \dots, n\}}$. Now let $\Omega_n$ be the space of all $n$-element sorting networks, and define $$ \nu_n = \frac{1}{\card{\Omega_n}} \sum_{\sigma \in \Omega_n} \delta(\nu_\sigma). $$ For each $n$, $\nu_n \in \mathcal{M}(\mathcal{M}(\mathcal{D}))$. We now extend Theorem \ref{T:weak-limit} to give a limit theorem for the sequence $\nu_n$. \begin{theorem} \label{T:weak-limit'} Let $\nu_n \in \mathcal{M}(\mathcal{M}(\mathcal{D}))$ be as above, and let $\mathbb{P}_\mathcal{A}$ be the law of the Archimedean path $\mathcal{A}(t) = X_1\cos(\pi t) + X_2 \sin (\pi t)$, where $(X_1, X_2) \sim \mathfrak{Arch}$. Then $\nu_n \to \delta_{\mathbb{P}_\mathcal{A}}$. \end{theorem} Note that the law of $Y_n$ is the expectation of the random measure $\nu_n$. Therefore Theorem \ref{T:weak-limit} gives convergence in expectation for $\nu_n$, whereas Theorem \ref{T:weak-limit'} gives pointwise convergence. \begin{proof} By Remark 2.4 from \cite{dauvergne1}, the sequence $\nu_n$ is precompact. For any subsequential limit $\nu$ of $\nu_n$, Theorem \ref{T:weak-limit} implies that $\nu$-almost every $\rho$ is supported on curves of the form $a\sin(\pi t) + b \cos(\pi t)$. Hence if $Z$ is a random path with law $\rho$, then $$ Z(t) = X_1 \cos(\pi t) + X_2 \sin(\pi t) $$ for some random variables $(X_1, X_2)$. Moreover, $Z(t)$ is uniform for every $t$ since each measure $\nu_n$ is supported on the set $\{\nu_\sigma\}_{\sigma \in \Omega_n}$. Therefore $(X_1, X_2) \sim \mathfrak{Arch}$, so $Z$ is the Archimedean path, and hence $\nu = \delta_{\mathbb{P}_\mathcal{A}}$. \end{proof} \begin{proof}[Proof of Theorem \ref{T:subnetwork}] Fix $m$. With the notation of Theorem \ref{T:subnetwork}, we have that \begin{equation} \label{E:2-cvg} \tau^n_m \stackrel{d}{\to} \tau_m \quad \;\text{as}\; n \to \infty \quad \text{ if and only if } \quad \frac{1}{|\Omega_n|} \sum_{\sigma \in \Omega_n} \nu_n^m \to \delta_{\mathbb{P}_\mathcal{A}^m} \;\text{as}\; n \to \infty. \end{equation} in the space $\mathcal{M}(\mathcal{M}(\mathcal{D}^n))$. Here $\nu_n^m$ is the $m$-fold product measure $\nu_n \times \nu_n \times \dots \nu_n$. The reason for this is that is if $I_1, \dots, I_m$ are uniform random variables on $\{1, \dots n\}$, then $I_j \ne I_k$ for all $j, k \in \{1, \dots n\}$ with high probability. Hence to find the limit of an $m$-out-of-$n$ sorting network, it suffices to find the limit of $m$ independently chosen trajectories. Finally, the second convergence statement in \eqref{E:2-cvg} follows from Theorem \ref{T:weak-limit'}. \end{proof} \begin{section}{The strong sine curve limit} \label{S:strong-limit} Theorem \ref{T:weak-limit} shows that for any $\epsilon > 0$, with high probability $(1- \epsilon)n$ particle trajectories in a random sorting network are close to sine curves. In this section, we extend this result to all particle trajectories, thus proving Theorem \ref{T:sine-curves}. By combining Theorem \ref{T:sine-curves} and Theorem \ref{T:weak-limit}, we also prove Theorems \ref{T:matrices} and \ref{T:unif-rotation}. \medskip The idea behind the proof is as follows. By Theorem \ref{T:weak-limit}, we know that most trajectories in a typical large-$n$ sorting network are close to sine curves. Since any two trajectories must cross exactly once, this restricts the type of behaviour that the remaining trajectories can have. Specifically, this forces all other trajectories to be either sine curves themselves, or to spend a lot of time at the edge of the sorting network. We can eliminate this second case by using the ``octagon" bound from \cite{angel2007random}. To state this bound, let $A_{n, \gamma}$ be the event where \begin{align*} \left|\sigma^n_G(i, t) - \sigma^n_G(i, 0)\right| < 2\sqrt{2t - t^2} + \gamma \;\; &\;\text{and}\; \;\; \left|\sigma^n_G(i, t) - \sigma^n_G(i, 1)\right| < 2\sqrt{1 - t^2} + \gamma \\ &\text{for all } \;\; t \in [0, 1], i \in \{1, \mathellipsis n\}. \end{align*} \begin{theorem}[Octagon bound, \cite{angel2007random}] \label{T:octagon} For any $\gamma > 0$, we have that $$ \lim_{n \to \infty} \mathbb{P}(A_{n, \gamma}) = 1. $$ \end{theorem} Now define $$ L^n_{i}(\delta) = \mathcal{L} \{ t: |\sigma^n_G(i, t)| \ge 1 - \delta\}, $$ where $\mathcal{L}$ is Lebesgue measure on $[0, 1]$. This is the amount of time that particle $i$ spends within $\delta$ of the edge in the random sorting network $\sigma^n$. \begin{lemma} \label{L:edge-time} For every $\epsilon > 0$, we have that $$ \lim_{n \to \infty} \mathbb{P} \left( \max_{i \in [1, n] } L^n_i(\epsilon^2/16) > \epsilon \right) =0. $$ \end{lemma} \begin{proof} Fix $\epsilon > 0, n \in \mathbb{N}$, and let $i$ be such that $|\sigma_G(i, 0)| \ge 1 - \epsilon^2/16$. On the event $A_{n, \epsilon^4/64}$, a simple calculation shows that $$ -1 + \epsilon^2/16 < \sigma_G(i, t) < 1 - \epsilon^2/16 \qquad \text{ for all } t \in [\epsilon/2, 1 - \epsilon/2]. $$ Therefore by Theorem \ref{T:octagon}, we have that \begin{equation} \label{E:time-shift} \lim_{n \to \infty} \mathbb{P} \bigg( \max \left\{L^n_i(\epsilon^2/16) : \; i \in [1, n], \;|\sigma^n_G(i, 0)| \ge 1 - \epsilon^2/16 \right\} > \epsilon \bigg) = 0. \end{equation} By time stationarity of random sorting networks (Theorem \ref{T:time-stat}), for each $n$ the above probability is greater than or equal to $$ \epsilon \mathbb{P} \left( \max_{i \in [1, n] } L^n_i(\epsilon^2/16) > \epsilon \right). $$ Therefore \eqref{E:time-shift} implies the lemma. \end{proof} Recall that $\nu_\sigma$ is uniform measure on the trajectories of a sorting network $\sigma$, and that $\mathbb{P}_\mathcal{A}$ is the law of the Archimedean path. To prove Theorem \ref{T:sine-curves} from here, we must show that if $\nu_\sigma$ is close to $\mathbb{P}_\mathcal{A}$ in the weak topology, then for any particle $i$, either $\sigma_G(i, \cdot)$ spends a lot of time at the edge of the sorting network, or else $\sigma_G(i, \cdot)$ is close to a sine curve. \medskip Recall that $\mathcal{D}$ is the closure of all sorting network trajectories in the uniform norm. To metrize weak convergence on the space of probability measures on $\mathcal{D}$, we use the L\'evy-Prokhorov metric $d_{LP}$. For a set $A \subset \mathcal{D}$, define $$ A^\epsilon = \{ f \in \mathcal{D} : ||f - g||_u < \epsilon \;\; \text{for some} \; g \in A \}. $$ Here and throughout the remaining proofs, $||\cdot||_u$ is the uniform norm. For two probability measures $\nu_1, \nu_2$ on $\mathcal{D}$, define $$ d_{LP}(\nu_1, \nu_2) = \inf \big\{ \epsilon > 0 : \nu_1(A) \le \nu_2(A^\epsilon) + \epsilon \text{ for all Borel sets } A \subset \mathcal{D} \big\}. $$ We now prove two lemmas characterizing particle behaviour when $\nu_\sigma$ is close to ${\mathbb{P}_\mathcal{A}}$. The first gives conditions under which a particle spends time close to the edge of a sorting network. Let $a_{n, i} = 2i/n - 1$, and define $$ C_{n, i}(t) = \sin(\arcsin(a_{n, i}) + \pi t) \qquad \;\text{and}\; \qquad c_{n, i}(t) = \sin(\arcsin(a_{n, i}) - \pi t). $$ Note that these are simply the paths $h_{a_{n, i}, \pm 1}$ in Theorem \ref{T:unique-2}. Recalling the definition of the Archimedean measure $\mathfrak{Arch}$ from Section \ref{S:intro}, for $\epsilon \in (0, 1]$ define $$ L(\epsilon) =\mathfrak{Arch}( r \ge 1 - \epsilon, \theta \in [2\pi x, 2\pi (x + \epsilon)]) =\epsilon\sqrt{2\epsilon - \epsilon^2}, $$ where $(r, \theta)$ are polar coordinates. Note that $L(\epsilon)$ is independent of $x$ by the rotational symmetry of $\mathfrak{Arch}$. \begin{lemma} \label{L:edge-cond} Let $\gamma \in (0, 2]$ and $\epsilon \in (0, \gamma^2/100)$. Suppose that for a fixed $n$-element sorting network $\sigma$ and a particle $i \in \{1, \mathellipsis n \}$, that $d_{LP}(\nu_\sigma, {\mathbb{P}_\mathcal{A}}) < L(\epsilon)/2$, and either $$ \max_{t \in [0, 1]} [\sigma_G(i, t) - C_{n,i}(t)]\ge \gamma \quad \text{or} \quad \max_{t \in [0, 1]} [c_{n, i}(t) - \sigma_G(i, t)] \ge \gamma. $$ Then $$ \mathcal{L}\left\{ t: |\sigma_G(i, t)| \ge 1- 3\epsilon \right\} \ge \gamma/6. $$ \end{lemma} \begin{proof} Without loss of generality, we can assume that there exists an $s \in [0, \arccos(a_{n, i})/\pi]$ such that \begin{equation} \label{E:sig-ga} \sigma_G(i, s) - C_{n, i}(s) \ge \gamma. \end{equation} The other cases follow by symmetric arguments. Since $d_{LP}(\nu_\sigma, {\mathbb{P}_\mathcal{A}}) < L(\epsilon)/2$, for every $x \in [0, 1]$, there exists a particle $j(x)$ such that $$ ||\sigma_G(j(x), t) - r\sin(\pi t + 2 \pi (x + \alpha)) ||_u < L(\epsilon)/2 $$ for some $r \ge 1 - \epsilon$ and $\alpha \in [0, \epsilon]$. Now since $$ ||r\sin(\pi t + 2 \pi (x + \alpha)) - \sin(\pi t + 2 \pi x)||_u \le 2 \epsilon, $$ this implies that \begin{equation} \label{E:sig-jx} ||\sigma_G(j(x), t) - \sin(\pi t + 2 \pi x)||_u < L(\epsilon)/2 + 2 \epsilon \end{equation} for all $x \in [0, 1]$. Let $\delta = \gamma - \frac{L(\epsilon)}2 - 2 \epsilon$. Note that $\delta$ is positive. For all $\alpha \in [0, \delta]$, the inequality \eqref{E:sig-ga} implies that \begin{align} \label{E:cond-1} \sigma_G(i, s)- \sin(\arcsin(a_{n, i}) + \alpha + \pi s) > \gamma - \delta = L(\epsilon)/2 + 2 \epsilon. \end{align} Now define $j_{\alpha} = j(\arcsin(a_{n, i}) + \alpha)/(2\pi)).$ Combining \eqref{E:cond-1} with \eqref{E:sig-jx}, we have that \begin{align} \nonumber \sigma_G(i, s) > \sigma_G\left(j_\alpha, s\right). \end{align} Moreover, if $1 - \cos(\alpha) > L(\epsilon)/2 + 2 \epsilon$, then $$ \sigma_G(i, 0) = a_{n, i} < \sin(\arcsin(a_{n, i}) + \alpha) - L(\epsilon)/2 + 2 \epsilon, $$ and so $$ \sigma_G(i, 0) < \sigma_G\left(j_\alpha, 0\right). $$ Thus for every $\alpha \in [\arccos(1 - L(\epsilon)/2 - 2 \epsilon), \delta]$, the particles $i$ and $j_\alpha$ must cross during the interval $[0, s]$. Therefore \begin{equation} \label{E:sig-G} \sigma_G(i, t) > \sigma_G\left(j_\alpha, t\right)\qquad \text{ for all } t > s, \;\; \alpha \in [\arccos(1 - L(\epsilon)/2 - 2 \epsilon), \delta]. \end{equation} We now show that this forces the particle $i$ to spend a large amount of time close to the edge of the sorting network. For all $\alpha \in [0, \delta]$, there must be some time $t_\alpha \in [0, 1]$ such that $$ \sin(\pi t_\alpha +\arcsin(a_{n, i}) + \alpha) = 1. $$ The time $t_\alpha \notin [0, s]$ since for every $\alpha \in [0, \delta]$, we have that $$ ||\sin(\pi t +\arcsin(a_{n, i}) + \alpha) - C_{n, i}(t)||_u < \gamma, \quad \;\text{and}\; \;\; C_{n, i}(t) \le C_{n, i}(s) \le 1 - \gamma \text{ for all } t \in [0, s]. $$ We have used \eqref{E:sig-ga} to get the second statement above. Therefore for all $\alpha \in [\arccos(1 - L(\epsilon)/2 - 2 \epsilon), \delta]$, \eqref{E:sig-G} implies that $$ \sigma_G(i, t_\alpha) > \sigma(j_\alpha, t_\alpha) \ge 1 - L(\epsilon)/2 - 2\epsilon \ge 1 - 3\epsilon. $$ Using the fact that $\arccos(1 - x) \le 2\sqrt{x}$ for $x \in [0, \pi/2)$, we have that \[ \mathcal{L} \big\{t_\alpha : \alpha \in [\arccos(1 - L(\epsilon)/2 - 2 \epsilon), \delta] \big\} \ge \frac{\delta - 2\sqrt{L(\epsilon)/2 + 2 \epsilon}}{\pi} \ge \frac{\gamma - 3\epsilon - 2\sqrt{3\epsilon}}{\pi} \ge \frac{\gamma}{6}. \] Here the final bound follows from the fact that $\epsilon < \gamma^2/100$. \end{proof} The second lemma shows that if a curve $\sigma_G(i, \cdot)$ stays close to the region between the curves $c_{n, i}$ and $C_{n, i}$, then it must be close to a sine curve. \begin{lemma} \label{L:sine-close} Let $\sigma$ be an $n$-element sorting network, $i \in \{1, \mathellipsis n \}$, and $\gamma \in (0, 1)$. Suppose that $$ \max_{t \in [0, 1]} [\sigma_G(i, t) - C_{n,i}(t)] < \gamma \quad \text{and} \quad \max_{t \in [0, 1]} [c_{n, i}(t) - \sigma_G(i, t)] < \gamma, $$ and that $d_{LP}(\nu_\sigma, {\mathbb{P}_\mathcal{A}}) < \frac{\gamma^4}{128}$. Then there exist constants $a \in [0, 1]$ and $\theta \in [0, 2\pi]$ such that $$ ||\sigma_G(i, t) - a\sin(\pi t + \theta) ||_u < 2\gamma + 2/n. $$ \end{lemma} \begin{proof} Suppose first that for some $s \in [0,1]$, that \begin{equation} \label{E:sig-GG} \sigma_G(i, s) \in [c_{n, i}(s) + \gamma, C_{n, i}(s) - \gamma]. \end{equation} Observe that $s \in [\arcsin(\gamma)/\pi, 1- \arcsin(\gamma)/\pi]$, since otherwise the above interval is empty. By \eqref{E:sig-GG}, we can find a point $(x_0, y_0) \in B(0, 1 - \gamma)$ such that $$ x_0 = \sigma_G(i, 0) \qquad \;\text{and}\; \qquad x_0\cos(\pi s) + y_0 \sin(\pi s) = \sigma_G(i, s). $$ Since $s \le 1- \arcsin(\gamma)/\pi$, we can find another point $ (x_1, y_1) \in B((x_0, y_0), \gamma) \subset B(0, 1) $ such that \begin{equation} \label{E:x1-x0} x_1 > x_0 + \gamma^2/3 \quad \;\text{and}\; \quad x_1\cos(\pi s) + y_1\sin(\pi s) > x_0\cos(\pi s) + y_0\sin(\pi s) + \gamma^2/3. \end{equation} Now observe that $$ \gamma^4/64\le \inf \left\{ \mathfrak{Arch}(B(x, \gamma^2/6)) : x \in B(0, 1) \right\}. $$ Therefore since $d_{LP}(\nu_\sigma, \mathbb{P}_{\mathbb{P}_\mathcal{A}}) < \gamma^4/128$, there must be a point $(x_2, y_2) \in B(0, 1) \cap B((x_1, y_1),\gamma^2/6)$ and a particle $j \in \{1, \dots, n\}$ such that \begin{equation} \label{E:G-jj} \begin{split} ||\sigma_G(j, t) - (x_1\cos(\pi t) &+ y_1\sin(\pi t))||_u \\ &\le ||\sigma_G(j, t) - (x_2\cos(\pi t) + y_2\sin(\pi t))||_u + \gamma^2/6 < \gamma^2/3. \end{split} \end{equation} By \eqref{E:x1-x0}, this implies that $$ \sigma_G(j, 0) > \sigma_G(i, 0) \qquad \;\text{and}\; \qquad \sigma_G(j, s) > \sigma_G(i, s), $$ and hence $\sigma_G(j, t) > \sigma_G(i, t)$ for all $t \in [0, s]$. Combining this with \eqref{E:G-jj} and the fact that $(x_1, y_1) \in B((x_0, y_0), \gamma)$ implies that $$ x_0\sin(\pi t) + y_0 \cos(\pi t) - \sigma_G(i, t) \le \gamma^2/3 + \gamma < 2\gamma \qquad \text{ for all } t \in [0, s]. $$ Symmetric arguments give the same upper bound on this difference over the interval $[s, 1]$, and on the difference $\sigma_G(i, t) - x_0\sin(\pi t) + y_0 \cos(\pi t)$ over the interval $[0, 1]$. This implies that $$ ||\sigma_G(i, t) - [x_0\sin(\pi t) + y_0 \cos(\pi t)]||_u \le 2\gamma. $$ Now suppose that there does not exist an $s \in [0, 1]$ such that $\sigma_G(i, s) \in [c_{n, i}(s) + \gamma, C_{n, i}(s) - \gamma]$. Then either $$ ||\sigma_G(i, \cdot) - C_{n, i}(\cdot)||_u < 2\gamma + \frac{2}n \qquad \;\text{or}\; \qquad ||\sigma_G(i, \cdot) - c_{n, i}(\cdot)||_u < 2\gamma + \frac{2}n. $$ This can be seen by simply analyzing the paths $C_{n, i}$ and $c_{n, i}$, noting that the trajectory $\sigma_G(i, \cdot)$ can ``jump" distances of size $2/n$. \medskip As all the functions $C_{n, i}$, $c_{n, i}$, and $x_0\sin(\pi t) + y_0 \cos(\pi t)$ are of the form $a\sin(\pi t + \theta)$, for some $(a, \theta) \in [0, 1] \times[0, 2\pi]$, this proves the lemma. \end{proof} We can now put together all the ingredients to prove Theorem \ref{T:sine-curves}. \begin{proof} Fix $\gamma \in (0, 1)$, and recall from Section \ref{SS:slightly} that $\nu_n$ is uniform measure on the set $\{\nu_\sigma\}_{\sigma \in \Omega_n}$. For small enough $\epsilon > 0$, Lemma \ref{L:edge-time} and Theorem \ref{T:weak-limit'} imply that $$ \mathbb{P} \left ( \max_{i \in [1, n]} \mathcal{L}\left\{t: |\sigma^n_G(i, t)| \ge 1 -3\epsilon \right\} < \frac{\gamma}{6}, \;\; d_{LP}(\nu_n, \mathbb{P}_\mathcal{A}) < \frac{\gamma^4}{128} \wedge \frac{L(\epsilon)}2 \right) \to 1 \qquad \;\text{as}\; n \to \infty. $$ Combining Lemmas \ref{L:edge-cond} and \ref{L:sine-close}, this implies that there exist random variables $A_{n, i, \gamma} \in [0, 1]$ and $\Theta_{n, i, \gamma} \in [0, 2\pi]$ such that $$ P_{n, \gamma} := \mathbb{P} \left(\max_{i \in [1, n]} ||\sigma^n_G(i, t) - A_{n, i, \gamma} \sin(\pi t + \Theta_{n, i, \gamma})||_u < 2\gamma + \frac{1}n \right) \to 1 \qquad \;\text{as}\; n \to \infty. $$ As constructed, the random variables $A_{n, i, \gamma}$ and $\Theta_{n, i, \gamma}$ depend on $\gamma$. To remove this dependence, let $\gamma_n \to 0$ be a sequence such that $P_{n, \gamma_n} \to 1$ as $n \to \infty$. Let $A_{n, i} = A_{n, i, \gamma_n}$ and $\Theta_{n, i} = \Theta_{n, i, \gamma_n}$, and define \begin{equation} \label{E:Bnga} B_{n, \gamma} = \left\{ \max_{i \in [1, n]} ||\sigma^n_G(i, t)- A_{n, i} \sin(\pi t + \Theta_{n, i})||_u < \gamma \right\}, \end{equation} Then for any $\gamma > 0$, we have that $\mathbb{P}(B_{n, \gamma}) \to 1$ as $n \to \infty$. \end{proof} We can also prove Theorem \ref{T:matrices} and Theorem \ref{T:unif-rotation}. For these we use the notation $B_{n, \gamma}$ from \eqref{E:Bnga}. \begin{proof}[Proof of Theorem \ref{T:matrices}] Fix $t \in [0, 1]$. Since $\nu_{n} \to \delta_{\mathbb{P}_\mathcal{A}}$ (Theorem \ref{T:weak-limit'}), the random measure $\rho_t^n$ converges in probability to the law of $(X, X \cos (\pi t )+ Y \sin (\pi t))$, where $(X, Y) \sim \mathfrak{Arch}$. This law is simply $\mathfrak{Arch}_t$. \medskip Now on the event $B_{n, \gamma}$, the support of the measure $\rho_t^n$ is contained in the set $$ \text{supp}(\mathfrak{Arch}_t)^\gamma = \{ x \in [-1, 1]^2 : d(x, \text{supp}(\mathfrak{Arch}_t)) < \gamma \}. $$ Moreover, since $\mathfrak{Arch}_t$ has a Lebesgue density that is bounded below, the weak convergence of $\rho^n_t$ to $\rho$ implies that with high probability, $ \text{supp}(\mathfrak{Arch}_t) \subset \text{supp}(\rho_n^t)^\gamma $ for any $\gamma > 0$. Therefore since $\mathbb{P}(B_{n, \gamma}) \to 1$ as $n \to \infty$ by Theorem \ref{T:sine-curves}, we have that \[ \lim_{n \to \infty} d_H(\text{supp}(\rho^n_t), \text{supp}(\mathfrak{Arch}_t)) = 0 \qquad \text{ in probability. } \quad \qedhere \] \end{proof} \begin{proof}[Proof of Theorem \ref{T:unif-rotation}] First observe that for any $a \in [0, 1]$, $t \in [0, 1/2]$, and $\theta \in [0, 2\pi]$, that $$ e^{\pi i t} \left[a\sin(\pi t + \theta) + ia\sin(\pi t + \pi/2 + \theta) \right] = a \sin(\theta). $$ Therefore on the event $B_{n, \gamma}$, we have that $$ \max_{j \in [1, n]} \max_{s, t \in [0, 1]} |Z^n_j(t) - Z^n_j(s)| \le 2\gamma. $$ Since $\mathbb{P}(B_{n, \gamma}) \to 1$ as $n \to \infty$ for any $\gamma > 0$ by Theorem \ref{T:sine-curves}, this proves the theorem. \end{proof} \end{section} \section{The geometric limit} \label{S:geom-limit} In this section, we use Theorem \ref{T:sine-curves} and Theorem \ref{T:weak-limit} to prove Theorem \ref{T:geom-limit}. Recall from Section \ref{S:intro} that $d_\infty(f, g)$ is the uniform norm between two $\mathbb{R}^n$-valued functions $f$ and $g$, where the pointwise distance is the $L^\infty$-distance. Recall also that $\bar{\sigma}$ is the embedding of a sorting network $\sigma$ into the $(n-2)$-dimensional sphere $\mathbb{S}^{n-2} \subset \mathbb{R}^n$. \medskip By Theorem \ref{T:sine-curves} and a change of variables, there exist random vectors $\mathbf{X^n} = (X^n_1, \mathellipsis X^n_n)$ and $\mathbf{V^n} = (V^n_1, \mathellipsis V^n_n)$ such that $d_\infty(f_n, \close{\sigma}^n)/n \to 0$ in probability, where $$ f_n(t) = \mathbf{X^n} \cos(\pi t) + \mathbf{V^n} \sin(\pi t) + \mathbf{c}, \qquad \text{where} \qquad \mathbf{c} = \left( \frac{n + 1}2, \mathellipsis , \frac{n + 1}2\right). $$ Moreover, we can assume that $X^n_i = i - (n + 1)/2$ and that $V^n_i = \sigma^n(i, N/2) - (n + 1)/2$, as these changes only shift the curve $\mathbf{X^n} \cos(\pi t) + \mathbf{V^n} \sin(\pi t)$ by $d_\infty$-distance $o(n)$ in probability. \medskip The point $\mathbf{c}$ is the center of $\mathbb{S}^{n-2}$. It remains to show that we can shift the curve $f_n$ by $d_\infty$-distance $o(n)$ to obtain a great circle in $\mathbb{S}^{n-2}$. For this we need the following lemma. \begin{lemma} \label{L:dot-prod-small} Let the vectors $\mathbf{X^n}$ and $\mathbf{V^n}$ be as above. Then $$ \lim_{n \to \infty} \frac{\left\langle\mathbf{X^n} , \mathbf{V^n} \right\rangle}{n^3} = 0 \qquad \text{in probability.} $$ \end{lemma} \begin{proof} Let $I_n$ be a uniform random variable on $\{1, \dots, n\}$, independent of $\mathbf{X^n}$ and $\mathbf{V^n}$, and define $$ (\tilde{X}^n, \tilde{V}^n) = \left(\frac{2X^n_{I_n}}n, \frac{2V^n_{I_n} }n\right). $$ We have that $$ (\tilde{X}^n, \tilde{V}^n) \stackrel{d}{=} (Y_n(0) - 1/n, Y_n(1/2) - 1/n), $$ where $Y_n$ is the trajectory random variable of $\sigma^n$. Therefore by Theorem \ref{T:weak-limit}, $$ (\tilde{X}^n, \tilde{V}^n) \stackrel{d}{\to} (X, V) \sim \mathfrak{Arch}. $$ By the bounded convergence theorem, this implies that $\mathbb{E} \tilde{X}^n\tilde{V}^n \to \mathbb{E} XV$ in probability. Observing that $n^3 \mathbb{E} \tilde{X}^n\tilde{V}^n = 4\left\langle\mathbf{X^n} , \mathbf{V^n} \right\rangle$ and that $\mathbb{E} XV = 0$ completes the proof. \end{proof} \begin{proof}[Proof of Theorem \ref{T:geom-limit}.] Fix $n \in \mathbb{N}$. For ease of bounding errors, we assume that $n \ge 27$. Let $\sigma$ be a fixed $n$-element sorting network. Our goal is to perturb the vector $\mathbf{V^n}$ to a new vector $\mathbf{W^n}$ so that the path \begin{equation} \label{E:C-nep} C_{n}(t) := \mathbf{X^n} \cos(\pi t) + \mathbf{W^n} \sin(\pi t) + \mathbf{c} \end{equation} follows a great circle on $\mathbb{S}^{n-2}$, and so that $d_\infty(C_{n}, f_n)$ is small. In order to do this, we just need to find $\mathbf{W^n}$ such that $$ \sum_{i=1}^n W^n_i = 0, \;\left\langle\mathbf{X^n} , \mathbf{W^n} \right\rangle = 0, \quad \;\text{and}\; \quad ||W^n||_2^2 = (n^3 - n)/12. $$ The first of these conditions guarantees that $C_n$ lies in the correct hyperplane $\mathbb{L}_n \subset \mathbb{R}^n$, and the second and third conditions guarantee that $C_n$ traces out a great circle on the sphere $\mathbb{S}^{n-2}$ within that hyperplane. We first perturb $\mathbf{V^n}$ to a vector $\mathbf{Z^n}$ satisfying the first two properties. Define the random variable $$ A_n = \frac{\left|\left\langle\mathbf{X^n} , \mathbf{V^n} \right\rangle\right|}{n^3}. $$ Without loss of generality, assume that $\left\langle\mathbf{X^n} , \mathbf{V^n} \right\rangle > 0$. \medskip Choose any $i \ge 3(n + 1)/4$. Let $k > 0$, and decrease the value of $V^n_i$ by $kn$ and increase the value of $V^n_{n + 1 - i}$ by $kn$. Call this new vector $\mathbf{V^n_1}$. Note that $X^n_i \ge n/4$, and that $X^n_i = -X^n_{n + 1 - i}$. Therefore $\left\langle\mathbf{X^n} , \mathbf{V^n_1} \right\rangle \le\left\langle\mathbf{X^n} , \mathbf{V^n} \right\rangle - kn^2/2$. \medskip By iterating the above procedure to repeatedly lower the dot product $\left\langle\mathbf{X^n} , \mathbf{V^n} \right\rangle$, we can obtain a vector $\mathbf{K_n} = (K_1n, \mathellipsis K_nn)$ with the following properties (here is where we require that $n \ge 27$). \smallskip \begin{enumerate}[nosep,label=(\roman*)] \item For all $i$, $|K_i| \le 9A_n$, $- K_i = K_{n + 1 - i}$, and $k_i = 0$ if $i \in \left(\frac{n+1}4, \frac{3(n+1)}4\right)$. \item $\left\langle\mathbf{X^n} , \mathbf{V^n + K^n} \right\rangle = 0.$ \end{enumerate} \smallskip Let $\mathbf{Z^n} = \mathbf{V^n + K^n}$. Observe that $\sum_{i=1}^n Z^n_i = 0$ since both $\sum_{i=1}^n K_i = 0$ and $\sum_{i=1}^n V^n_i = 0$. The fact that $\sum_{i=1}^n V^n_i = 0$ follows since $\mathbf{V^n} + \mathbf{c}$ is a permutation of the vector $(1, 2, \mathellipsis n)$. For any $t$, we have that \begin{align*} ||f_n(t) - (\mathbf{X^n} \cos(\pi t) + \mathbf{Z^n} \sin(\pi t) + \mathbf{c})||_\infty = \max_{i \in [1, n]} |V^n_i \sin(\pi t) - Z^n_i \sin(\pi t)| \le 9 A_n n. \end{align*} We can now define $\mathbf{W^n} = M_n\mathbf{Z^n}$, where $M_n$ is a random constant chosen so that $||\mathbf{W^n}||^2_2 = (n^3 - n)/12$. Using the definition of $C_n$ in \eqref{E:C-nep}, for every $t \in [0, 1]$, we have that \begin{equation} \label{E:fC} \begin{split} ||f_n(t) - C_n(t)||_\infty &\le 9\epsilon n + \max_{i \in [1, n]} |W^n_i \sin(\pi t) - Z^n_i \sin(\pi t)| \\ &< 9 A_n n + |M_n - 1|\left[\frac{n + 1}2 + 9 A_n n\right]. \end{split} \end{equation} In the last inequality, we have used that $|Z^n_i - V^n_i| \le 9 A_n n$. The inequality \eqref{E:fC} implies that \begin{equation} \label{E:d-inf-comp} \frac{d_\infty(C_n, \close{\sigma}^n)}n \le 9 A_n + |M_n - 1|\left[\frac{n + 1}{2n} + 9 A_n \right] + \frac{d_\infty(f_n, \close{\sigma}^n)}n. \end{equation} Now again using that $|Z^n_i - V^n_i| \le 9 A_n n$, we have that \begin{align*} \big|\;||\mathbf{Z^n}||^2_2 - ||\mathbf{V^n}||^2_2 \; \big| \le \sum_{i=1}^n |Z^n_i - V^n_i||Z^n_i + V^n_i| \le 9A_n n^2(n + 1 + 9 A_n n). \end{align*} Therefore since $||\mathbf{V^n}||^2_2 = (n^3 - n)/12$, we have that $$ M_n \in \left[\sqrt{1 - 217A_n}, \sqrt{1 + 217A_n} \right] $$ whenever $A_n < 1 /217$. Lemma \ref{L:dot-prod-small} implies that $A_n \to 0$ as $n \to \infty$ in probability, and hence so does $M_n$. Therefore since $d_\infty(f_n, \close{\sigma}^n)/n \to 0$ in probability as $n \to \infty$, \eqref{E:d-inf-comp} implies that $d_\infty(C_n, \close{\sigma}^n)/n$ does as well. \end{proof} \section*{Acknowledgements} The author would like to thank B\'alint Vir\'ag for many fruitful discussions about the problem, and for many constructive comments about previous drafts. \bibliographystyle{alpha}

https://arxiv.org/abs/2011.02258

Concentration Inequalities for Statistical Inference

This paper gives a review of concentration inequalities which are widely employed in non-asymptotical analyses of mathematical statistics in a wide range of settings, from distribution-free to distribution-dependent, from sub-Gaussian to sub-exponential, sub-Gamma, and sub-Weibull random variables, and from the mean to the maximum concentration. This review provides results in these settings with some fresh new results. Given the increasing popularity of high-dimensional data and inference, results in the context of high-dimensional linear and Poisson regressions are also provided. We aim to illustrate the concentration inequalities with known constants and to improve existing bounds with sharper constants.

\section*{Acknowledgments} \author[Zhang H., Chen S.X.]{Huiming Zhang\affil{1}\comma\affil{3}\comma\affil{4}, Song Xi Chen\affil{1}\comma\affil{2}\comma\affil{3}\corrauth} \address{\affilnum{1}\ School of Mathematical Sciences, \affilnum{2}\ Guanghua School of Management, \affilnum{3}\ Center for Statistical Sciences, Peking University, Beijing, P.R. China;\\\affilnum{4} Department of Mathematics, Faculty of Science and Technology, University of Macau, Macau, China.} \emails{{\tt [email protected]} (Zhang H.), {\tt [email protected]} (Chen S.X.)} \begin{abstract} This paper gives a review of concentration inequalities which are widely employed in non-asymptotical analyses of mathematical statistics in a wide range of settings, from distribution-free to distribution-dependent, from sub-Gaussian to sub-exponential, sub-Gamma, and sub-Weibull random variables, and from the mean to the maximum concentration. This review provides results in these settings with some fresh new results. Given the increasing popularity of high-dimensional data and inference, results in the context of high-dimensional linear and Poisson regressions are also provided. We aim to illustrate the concentration inequalities with known constants and to improve existing bounds with sharper constants. \end{abstract} \ams{60F10, 60G50, 62E17} \keywords{constants-specified inequalities, sub-Weibull random variables, heavy-tailed distributions, high-dimensional estimation and testing, finite-sample theory, random matrices.} \maketitle \tableofcontents \section{Introduction} In probability theory and statistical inference, researchers often need to bound the probability of a difference between a random {quantity from its target}, usually the error bound of estimation. Concentration inequalities (CIs) are tools for attaining such bounds, and play important roles in deriving theoretical results for various inferential situations in statistics and probability. The recent developments in high-dimensional (HD) statistical inference, and statistical and machine learning have generated renewed interests in the CIs, as reflected in \cite{Koltchinskii11}, \cite{Vershynin18}, \cite{Wainwright19} and \cite{Fanj20}. As the CIs are diverse in their forms and the underlying distributional requirements, and are scattered around in references, there is an increasing need for a review which collects existing results together with some new results (sharper and constants-specified CIs) from the authors for researchers and graduate students working in statistics and probability. This motivates the writing of this review. {CIs enable us to obtain non-asymptotic results for estimating, constructing confidence intervals, and doing hypothesis testing with a high-probability guarantee. For example, the first-order optimized condition for HD linear regressions should be held with a high probability to guarantee the well-behavior of the estimator. The concentration inequality for error distributions is to ensure the concentration from first-order optimized conditions to the estimator.} {Our review focuses on four types of CIs: \begin{center} $P(Z_n > \mathrm{E}Z_n + t),\quad P(Z_n < \mathrm{E}Z_n - t), \quad P(|Z_n - \mathrm{E}Z_n| > t)$ and $ \mathrm{E}(\max\limits_{i=1,\cdots,n} |{X_i }|)$ \end{center} where $Z_n: = f({X_1}, \cdots ,{X_n})$ and ${X_1}, \cdots ,{X_n}$ are random variables.} We present two types of CIs: distribution-free and distribution-dependent. Distribution free CIs are free of distribution assumptions, while the distribution-dependent CIs are based on exponential moment conditions reflecting the tail property for the particular class of distributions. {Concentration phenomenons for a sum of sub-Weibull random variables will lead to a mixture of two tails: sub-Gaussian for small deviations and sub-Weibull for large deviations from the mean, and it is closely related to Strong Law of Large Numbers, Central Limit Theorem, and Law of the Iterative Logarithm. }We provide applications of the CIs to empirical processes and high-dimensional data settings. The latter includes the linear and Poisson regression with a diverging number of covariates. We organize the materials in the forms of Lemmas, Corollaries, Propositions, and Theorems. Lemmas and Corollaries are on existing results usually without proof except for a few fundamental ones. Propositions are also for existing results but with sharper or more precise constants {and sometimes come with proofs}. Theorems are for new results. This review contains 26 Lemmas, 21 Corollaries, 14 Propositions, and 4 Theorems. The review is organized as follows. Section 2 outlines distribution-free CIs. CIs for Sub-Gaussian, Sub-exponential, sub-Gamma, and sub-Weibull random variables are given in Section 3, 4, 5, and 6 respectively. Section 7 reports concentration for the maximal of random variables and suprema of empirical processes. Applications for high dimensional linear and Poisson regression are outlined in Section 8. {Section 9 discusses extensions to other settings. } \section{Distribution-free Concentration Bounds}\label{Distribution-free} The purpose here is to introduce distribution-free CIs. We first review Markov's, Chebysheff's and Chernoff's tail probability bounds that constitute fundamental inequalities for deriving most of the concentration bounds; see Chap. 1 of \cite{Durrett2019} or Appendix B in \cite{Giraud2014} for the proofs. \begin{lemma}[Markov's inequality]\label{Markov} Let $\varphi(x): \mathbb{R} \rightarrow \mathbb{R}^{+}$ be any non-decreasing positive function. For any real valued random variable (r.v.) $X,$ $ {P}(X \geq a) \leq \mathrm{E}[\varphi(X)]\frac{1}{\varphi(a)} ,~\forall~ a \in \mathbb{R}. $ \end{lemma} By letting $\varphi(x)=x^2$, the following Chebyshev's inequality is merely an application of Markov's inequality for $|X-\mathrm{E}X|$. \begin{lemma}[Chebyshev's Inequality] Let $X$ be a r.v. with expectation $\mathrm{E}X$ and variance $\operatorname{Var}X$. Then, for any $a \in \mathbb{R}^{+}:$ $ {P}(|X-\mathrm{E}X| \geq a) \leq \frac{\operatorname{Var}X}{a^{2}}. $ \end{lemma} The Chebyshev's inequality prescribes a polynomial rate of convergence depending on the {variance assumption}. Another application of Markov's inequality is the Chernoff's bound which is sharper by optimizing the upper bounds. \begin{lemma}[Chernoff's inequality] For a r.v. $X$ with $\mathrm{E}{e^{tX}}< \infty$, $P(X \ge a) \le {\inf}_{t > 0} \left\{ {{e^{ - ta}}\mathrm{E}{e^{tX}}} \right\}.$ \end{lemma} \begin{proof} Lemma \ref{Markov} with $\varphi(x)=e^{t x}$ implies $P(X \geq a) \leq {{e^{ - ta}}\mathrm{E}{e^{tX}}}$ and minimize $t$ on $t>0$. \end{proof} The Jensen's inequality and its truncated version [Lemma 14.6 in \cite{Buhlmann11}] are another powerful tool to derive useful inequalities by the convexity. \begin{lemma}[Jensen's inequality] For any convex function $\varphi: \mathbb{R}^{d} \rightarrow \mathbb{R}$ and any r.v. $X$ in $\mathbb{R}^{d},$ such that $\varphi(X)$ is integrable, we have $\varphi(\mathrm{E}X) \leq \mathrm{E}[\varphi(X)]$. \end{lemma} \begin{lemma}[Truncated Jensen's inequality]\label{lem:Truncated} Let $g(\cdot)$ be an increasing function on $[0, \infty),$ which is concave on $[c, \infty)$ for some $c \geq 0 .$ Then $\mathrm{E} g(|Z|) \leq g[\mathrm{E}|Z|+c {P}(|Z|<c)]$ for r.v. $Z$. \end{lemma} The Chebyshev's, Markov's, Chernoff's and Jensen's inequalities are also valid for conditional expectations (Chapter 4 of \cite{Durrett2019}). The Chernoff's bound typically lead to a tighter bound than Markov's inequality by optimization via an exponential $\varphi(x)$ function. A sharper bound for the sum of independent random variables(r.vs) was attempted in \cite{Hoeffding63}. The following is a slightly sharper bound from Theorem 1.2 in \cite{Bosq1998}. \begin{corollary}[Hoeffding's inequality]\label{lm:Hoeffding} Let ${X_1}, \cdots ,{X_n}$ be independent r.vs on $\mathbb{R}$ satisfying bound condition $ {a_i}\le {{X_i}} \le {b_i}~\text{for}~i = 1,2, \cdots ,n. $ Then for $t,u> 0$ \begin{enumerate} \item[\rm{(}a\rm{)}] \textbf{Hoeffding's lemma}: ${\rm{E}}e^{u \sum\limits_{i = 1}^n ({{X_i}}-{\rm{E}}{X_i})} \le e^{ \frac{u^2}{8}\sum\limits_{i = 1}^n {(b_i-a_i)^2} }~\text{and}~{\rm{E}} e^{u| \sum\limits_{i=1}^{n} ({{X_i}}-{\rm{E}}{X_i})|}\leq 2e^{\frac{u^{2}}{8} \sum\limits_{i=1}^{n} (b_i-a_i)^{2}}$; \item[\rm{(}b\rm{)}] \textbf{Hoeffding's inequality}: $P(|\sum_{i = 1}^n ({{X_i}}-{\rm{E}}{X_i}) | \ge t) \le 2e^{-{2 t^{2}}/{\sum_{i=1}^{n}\left(b_{i}-a_{i}\right)^{2}}}$. \end{enumerate} \end{corollary} Corollary \ref{lm:Hoeffding} has a sharper bound than the Markov's inequality or Chebyshev's inequality with the requirement of first or moment condition on $X$. Hoeffding's inequality has many applications in statistics as shown in the next example. \iffalse The exponential bounds have term 2 in $2e^{-{2 t^{2}}/{\sum_{i=1}^{n}\left(b_{i}-a_{i}\right)^{2}}}$ (and in subsequent sections) which is from separate bounds on the right tail probability $P(\sum_{i = 1}^n {{X_i}} \ge t)$ and the left tail $P(\sum_{i = 1}^n {{X_i}} \le -t)$. So it is enough to show the the exponential bound of the right tail, since by the symmetry of centralized r.vs, the right tail follows from the left tail inequality applied to $-\sum_{i = 1}^n {{X_i}}$. \fi \textbf{The proof of Hoeffding's lemma.} Without loss of generality, we assume $\mathrm{E}X_i=0$. This is from the fact that the concentration inequality is location shift-invariance. Since $f(x)=e^x$ is convex, for $u>0$, then $e^{u x} \leq \frac{b_{i}-x}{b_{i}-a_{i}} e^{u a_{i}}+\frac{x-a_{i}}{b_{i}-a_{i}} e^{u b_{i}},~a_{i} \leq x \leq b_{i}.$ Taking expectation, it gives by $\mathrm{E}X_i=0$ \begin{equation} \label{hoeffd1} \mathrm{E}e^{u X_{i}} \le \frac{b_{i}}{b_{i}-a_{i}} e^{u a_{i}}-\frac{a_{i}}{b_{i}-a_{i}} e^{u b_{i}} =\left[1-s+s e^{u\left(b_{i}-a_{i}\right)}\right] e^{-s u\left(b_{i}-a_{i}\right)} \triangleq e^{f(r)}, \end{equation} where $r=u(b_i-a_i), s=-a_i/(b_i-a_i)$ and $f(r)=-sr+\log(1-s+se^r).$ We can show that $f'(r) = - s + \frac{{s{e^r}}}{{1 - s + s{e^r}}},~~f''(r) = \frac{{(1 - s)s{e^r}}}{{{{(1 - s + s{e^r})}^2}}}\le \frac{{\rm{1}}}{{\rm{4}}}$ for all $r\ge 0$. Note that $f(0)=f'(0)=0$. Consider the Taylor's expansion of $f$, there exists $\xi\in[0,1]$ such that $f(r)=r^{2} f^{\prime \prime}(\xi r) / 2 \leq r^{2} / 8=u^{2}\left(b_{i}-a_{i}\right)^{2} / 8.$ Substitute it to \eqref{hoeffd1}, we get the Hoeffding's lemma. The last assertion of Lemma~\ref{lm:Hoeffding}(a) is by letting $Z=u\sum_{i=1}^{n} ({{X_i}}-{\rm{E}}{X_i})$, so that \begin{equation}\label{eq:abZ} {\rm{E}}e^{|Z|}={\rm{E}} e^{-Z} \cdot 1(Z \leq 0)+{\rm{E}}e^{Z} \cdot 1(Z>0)\leq 2e^{\frac{1}{8} u^{2} \sum_{i=1}^{n} (b_i-a_i)^{2}}. \end{equation} \textbf{The proof of Hoeffding's inequality.} Let $S_n =\sum_{i=1}^n X_i$ and $c_i=a_i-b_i$. For any $t,u> 0$, \begin{align}\label{eq:SN0} P(S_n -\mathrm{E}S_n \ge t) &= P(e^{u (S_n -\mathrm{E}S_n) } \ge e^{ut}) \le \mathop {\inf }\limits_{u > 0} e^{-ut} \prod_{i=1}^n \mathrm{E}e^{u (X_i - \mathrm{E}X_i)}~[\text{Chernoff's inequality}]\nonumber \\ [\text{Hoeffding's lemma}] & \le \mathop {\inf }\limits_{u > 0} e^{-ut} \prod{}_{i=1}^n e^{ u^2 c_i^2/8} = \mathop {\inf }\limits_{u > 0} e^{-ut + u^2 \sum_{i=1}^nc_i^2/8}=e^{ {-2t^{2}}/\sum_{i=1}^{n} c_{i}^{2}}. \end{align} The smallest bounded is attained at $u = 4t/\sum_{i=1}^nc_i^2$ and $P(-[S_n -\mathrm{E}S_n] \ge t) \le e^{ {-2t^{2}}/\sum_{i=1}^{n} c_{i}^{2}}$ similarly. Hence, the Hoeffding's inequality is verified via $$ P(|S_n -\mathrm{E}S_n| \ge t)\le P(S_n -\mathrm{E}S_n \ge t) + P(-[S_n -\mathrm{E}S_n] \ge t)\le 2e^{ {-2t^{2}}/\sum_{i=1}^{n} c_{i}^{2}}. $$ Corollary \ref{lm:Hoeffding} has a sharper bound than the Markov's inequality or Chebyshev's inequality with the requirement of first or moment condition on $X$. A second approach for proving Hoeffding's lemma is given in Lemma 1.8 of \cite{Rigollet19}. Hoeffding's inequality has many applications in statistics as shown in the next example. \begin{example}[Empirical distribution function, EDF]\label{eg:edf} Let $\{ {X_i}\} _{i = 1}^n \stackrel{\rm IID}{\sim} F(x)$ for a distribution $F$. Let $\mathbb{F}_{n}(x):={\frac 1n}\sum _{{i=1}}^{n}{\rm{1}}_{{\{X_{i}\leq x\}}}(x),~x\in {\mathbb{R}}$ be the empirical distribution. By Hoeffding's inequality ($a_i-b_i=1/n$), $P(|\mathbb{F}_{n}(x)-F(x)|>\varepsilon )\leq 2e^{-2n\varepsilon ^{2}},~\forall \varepsilon >0.$ \end{example} McDiarmid's inequality (also called bounded difference inequality, see \cite{McDiarmid89}) is a concentration inequality for a multivariate function of random sequence $\{X_i\}_{i=1}^n$, says $f(X_{1},...,X_{n})$. As a generalization of Hoeffding's inequality, it does not require any distribution assumptions about r.vs and the $f(X_{1},...,X_{n})$ may be dependent sum of r.vs. The only requirement is the bounded difference condition by replacing $X_{j}$ by $X_{j}^{'}$ meanwhile maintaining the others fixed in $f(X_{1},...,X_{n})$. \begin{lemma}[McDiarmid's inequality]\label{lm:bd} Suppose $X_{1},\cdots,X_{n}$ are independent r.vs all taking values in the set $A$, and assume $f:A^n\rightarrow\mathbb{R}$ satisfies the \emph{bounded difference condition} \begin{center} ${\sup}_{{x_{1},\cdots,x_{n},x_{k}^{'}\in A}}\vert f(x_{1},\cdots,x_{n})-f(x_{1},\cdots,x_{k-1},x_{k}^{'},x_{k+1},\cdots,x_{n})\vert\le c_{k}.$ \end{center} Then, $P\left( {\left| {f({X_1},\cdots,{X_n}) - {\rm{E}}\left\{f({X_1},\cdots,{X_n})\right\}} \right| \ge t} \right) \le 2e^{- 2{t^2}/\sum_{i = 1}^n {c_i^2}}~~\forall t>0.$ \end{lemma} One method of proof is by the martingale argument, which needs to check the Azuma-Hoeffding's inequality below, see Section 2.2.2 in \cite{Wainwright19}. Theorem 3.3.14 of \cite{Gine15} gives another proof based on the entropy method. \begin{lemma}[Azuma-Hoeffding's inequality]\label{le:RAH} Let $\{X_{n}\}_{n=0}^{\infty}$ be a sequence of martingale (or supermartingale), adapted to an increasing filtration $\{\mathcal{F}_{n}\}_{n=0}^{\infty}$. Suppose $\{X_{n}\}_{n=0}^{\infty}$ satisfies the \emph{bounded difference condition} $ a_{k}\le X_{k}-X_{k-1}\le b_{k},~~\rm{a.s.} $ for $k=1, \ldots, n$. Then, ${P}\left(\left|X_{n}-X_0\right|>t\right) \leq 2e^{ -2{t^2}/\sum_{i = 1}^n {(b_{k}-a_{k})^2}},~t\ge 0$. \end{lemma} Two typical examples with bounded differences function are the concentration for U-statistics (a dependent summation) and the integral error of the kernel density estimation. \begin{example}[U-statistics]\label{eg:u} Let $\{X_{i}\}_{i=1}^n$ be independent and identically distributed (IID) r.vs and $g: \mathbb{R}^{2} \rightarrow \mathbb{R}$ be the bounded and symmetric function. Define a \emph{U-statistic of order 2} as $U_n={\tiny \left(\begin{array}{l} n \\ 2 \end{array}\right)}^{-1} \sum_{i<j} g\left(X_{i}, X_{j}\right):=f(x_{1},...,x_{n}).$ Its bounded difference condition is \begin{align*} &~~~~\left|f\left(x_{1}, \ldots, x_{k-1}, x_{k}, x_{k+1}, \ldots x_{n}\right)-f\left(x_{1}, \ldots, x_{k-1}, x_{k}^{\prime}, x_{k+1}, \ldots x_{n}\right)\right| \\ &=\frac{1}{\tiny \left(\begin{array}{l} n \\ 2 \end{array}\right)}|\sum_{j=1, j \neq k}^{n}[g(x_{k}, x_{j})-g(x_{k}^{\prime}, x_{j})]| \leq \frac{2\cdot 2(n-1)\|g\|_{\infty}}{n(n-1)}=\frac{4\|g\|_{\infty}}{n}. \end{align*} So we have $ {P}(|U_n-\mathrm{E}U_n|>t) \leq 2 e^{-{n t^{2}}/{8\|g\|_{\infty}^{2}}} $. \end{example} \begin{example}[$L_{1}$-error in kernel density estimation] Let $\{X_{i}\}_{i=1}^n \stackrel{\rm IID}{\sim} F(x)$ with density function $f(x)$. Define the \emph{kernel density estimator} by $\hat { f }_{n,h} ( x ) = \frac { 1 } { n } \sum _ { i = 1 } ^ { n } \frac { 1 } { h } K \left( \frac { x - X _ { i } } { h } \right)$ where $K(\cdot)>1$ is the kernel function and $h>0$ is a smoothing parameter called the bandwidth. Usually, the kernel function $K(\cdot)$ is symmetric probability density and $h>0$ with $h \to 0$ and $nh \to \infty$. Define the $L_{1}$-error of $\hat { f }_{n,h} ( x ) $ by ${Z_n}=g({X}_{1}, \ldots,{X}_{{n}})=\int|\hat { f }_{n,h} ( x )-f({x})| {dx}. $ By $\int K(u) d u=1$, the McDiarmid's inequality with bound difference condition \begin{center} $\left|g\left({x}_{1}, \ldots, {x}_{{n}}\right)-g({x}_{1}, \ldots, {x}_{{i}}^{\prime}, \ldots, {x}_{{n}})\right|\leq \frac{1}{n } \int\left|K\left(\frac{x-x_{i}}{h}\right)-K\left(\frac{x-x_{i}^{\prime}}{h}\right)\right| d \left(\frac{x}{h}\right) \le \frac{2}{n}$ \end{center} gives $P(|Z_n-\mathrm{E}Z_n|\ge t)\leq 2 e^{-{2 t^{2}}/{n(\frac{2}{n})^{2}} }=2 e^{- n t^{2}/2 }$, which is free of the bandwidth. \end{example} \section{Sub-Gaussian Distributions}\label{sec-sg} \subsection{Motivations} In probability, there is a well-known inequality for bounding the Gaussian tail. If $X \sim N(0,1)$, \cite{Gordon1941} obtained for $x>0$ \begin{equation}\label{eq:Mills} \left( {\frac{1}{x} - \frac{1}{{{x^3}}}} \right)\cdot \frac{e^{ - {x^2}/2}}{{\sqrt {2\pi } }}< \left(\frac{x}{x^{2}+1} \right)\cdot\frac{e^{ - {x^2}/2}}{{\sqrt {2\pi } }} \le P(X \ge x) \le \frac{1}{x}\cdot \frac{e^{ - {x^2}/2}}{{\sqrt {2\pi }}}, \end{equation} which is called \emph{Mills's inequality}, relating to Mills's ratio \citep{Mills1926}. The upper bound in \eqref{eq:Mills} is mostly used to derive law of the iterated logarithm \citep{Durrett2019}. However, if $x$ tends to zero the upper bound goes to $+\infty$ which makes it meaningless. So the Mill's inequality is useful only for larger $x$. We need a better inequality. In fact, the upper bound in \eqref{eq:Mills} can be strengthened as in Lemma B.3 in \cite{Giraud2014}: $P(|X| \ge x) \le {e^{ - {x^2}/2}}.$ We refer it as the \emph{sharper Mill's inequality}. In statistics, people want to study a general class of error distributions (beyond Gaussian) whose \textit{moment generating function} (MGF): $\mathrm{E}e^{s X}$ have similar Gaussian properties with $s$ in specific subset of $\mathbb{R}$. To derive sharper Mill's inequality, it is natural to define the class of sub-Gaussian r.v. as follows. \begin{definition}[Sub-Gaussian distribution]\label{def:Sub-Gaussian} A r.v. $X \in \mathbb{R}$ with mean zero is sub-Gaussian with a \textit{variance proxy} $\sigma^{2}$ (denoted $X \sim \operatorname{subG}\left(\sigma^{2}\right)$) if its MGF satisfies $\mathrm{E}e^{s X}\le e^{\frac{\sigma^{2} s^{2}}{2}},~\forall s \in \mathbb{R}.$ \end{definition} With Definition \ref{def:Sub-Gaussian} and Chernoff's inequality, we will get the exponential decay of the tail as the alternative definition of sub-Gaussian: ${P}\left( X \ge t \right)\le \mathop {\inf }\limits_{s > 0} e^{-s t}\mathrm{E} e^{s X} \le \mathop {\inf }\limits_{s > 0} e^{-s t + \frac{\sigma^2s^2}{2}}\xlongequal{s={t}/{\sigma^2}} e^{- \frac{t^2}{2 \sigma^2}}$. This argument is called \emph{Cramer-Chernoff method}, and it is applied in proving Hoeffding's lemma for sum of independent variables. In general. let $Z_{1}, \ldots, Z_{n}$ be $n$ independent centralized r.vs, and suppose there exists a convex function $g(t)$ and a domain $D_0$ containing $\{0\}$ such that $\mathrm{E}e^{t \sum_{i=1}^{n} Z_{i}} \leq e^{n g(t)},~\forall t \in D_0 \subset \mathbb{R}.$ Denote $g^{*}(s)=\sup _{t \in D_0}\{t s-g(t)\}$ as the \textit{convex conjugate function} of $g,$ therefore the Chernoff's inequality implies \begin{center} $P(|\frac{1}{n} \sum_{i=1}^{n} Z_{i}|>s) \leq 2e^{-n g^{*}(s)},~\forall s>0$, \end{center} which has rich applications in high-dimensional statistics, machine learning, random matrix theory, and other fields on non-asymptotic results. Note that $\operatorname{subG}(\sigma^{2})$ denotes a class of distributions rather than a single distribution. Trivially, the Gaussian distribution is a special case of sub-Gaussian. \begin{example} [Normal distributions] Consider the normal r.v. $X\sim N(\mu, \sigma^{2})$. With the MGF of $X$: $\mathrm{E}e^{s X}:= e^{\frac{\sigma^{2} s^{2}}{2}},~\forall s \in \mathbb{R}$, it is sub-Gaussian with the variance proxy $\sigma^{2}=\operatorname{Var}(X)$. \end{example} \begin{example}[Bounded r.vs] By Hoeffding's lemma, $\mathrm{E}e^{s X} \le e^{ \frac{1}{8}{s^2}(b-a)^2}~\text{for}~s> 0$ for the centralized bounded variable $X \in [{a},{b}]$. So $X$ is essentially sub-Gaussian with variance proxy $\sigma^{2}=\frac{1}{4}{(b-a)^2}$. For Bernoulli variable $X \in \{0,1\}$, we have $X \sim \operatorname{subG}\left(\frac{1}{4}\right)$. \end{example} There are at least seven equivalent forms for sub-Gaussian as shown in the following. \begin{corollary}[Characterizations of sub-Gaussian]\label{Sub-gaussian distribution} Let $X$ be a r.v. in $\mathbb{R}$ with $\mathrm{E} X = 0$. Then, the following are equivalent for finite positive constants $\{K_i\}_{i=1}^7$. \begin{enumerate}[\rm{(}1\rm{)}] \item \label{p: sub-gaussian MGF} The MGF of $X$: $ \mathrm{E} e^{s X} \le e^{K_1^2 s^2}~\text{for all } s \in \mathbb{R} $; \item \label{p: sub-gaussian tail} The tail of $X$: $ P \{ |X| \ge t \} \le 2 e^{-t^2/K_2^2}~ \text{for all } t \ge 0; $ \item \label{p: sub-gaussian moments} The moments of $X$: $ (\mathrm{E} |X|^k)^{1/k} \le K_3 \sqrt{k}~\text{for all integer}~k \ge 1; $ \item \label{p: sub-gaussian MGF square finite} The exponential moment of $X^2$ \index{Moment Generating Function}: $ \mathrm{E} e^{X^2/K_4^2}\le 2; $ \item \label{p: sub-gaussian MGF square} The local MGF of $X^2$: $\mathrm{E} e^{\l^2 X^2} \le e^{K_5^2 \l^2}~\text{for all $\l$ in a local set} |\l| \le \frac{1}{K_5}. $ \item There is a constant $\sigma \geq 0$ such that $\mathrm{E} e^{{\lambda X^{2}}/{ K_6^2}} \le ({1-\lambda})^{-1/2}~\text { for all } \lambda \in[0,1)$. \item Union bound condition: $\exists c>0$ s.t. $\mathrm{E}\left[\max \left\{|X_{1}|, \ldots,|X_{n}|\right\}\right] \leq c \sqrt{\log n}$ for all $n \geq c$, where $\{X_i\}_{i=1}^n$ are IID copies of $X$. \end{enumerate} \end{corollary} \begin{remark} The $\mathrm{E} X = 0$ is for convenience as the zero mean is used in the proof of Corollary \ref{Sub-gaussian distribution}(1), see \cite{Vershynin2010} for the details and the proof of the equivalences (1)-(5). The equivalences (6) is given in Theorem 2.6 of \cite{Wainwright19} and the equivalences (7) is present in Page24 of \cite{Sen2018}. The moment condition for integers $k$ in (3) can be relaxed to \emph{even integers} $k$ by the symmetrization technique. By symmetry of $X$, let us consider a negative independent copy $-X^{\prime}$ which is independent of $X$ and has the same distribution as $X$. If (3) is true and $\mathrm{E}(-X^{\prime})=0$, from Jensen's inequality ${\rm{E}}{e^{\theta ( - {X^\prime })}} \ge {e^{\theta {\rm{E}}( - {X^\prime })}} = 1$ since $-{X^\prime }$ has zero mean. So we have by the independence of $ X^{\prime}$ and $X$: \begin{align*} {\rm{E}}{e^{\theta X}}& \le {\rm{E}}{e^{\theta X}}{\rm{E}}{e^{\theta ( - {X^\prime })}} = {\rm{E}}{e^{\theta (X - {X^\prime })}}= 1 + \sum\limits_{k = 1}^\infty {\frac{{{\theta ^{2k}}{\rm{E}}{{(X - {X^\prime })}^{2k}}}}{{(2k)!}}} \le 1 + \sum\limits_{k = 1}^\infty {\frac{{{\theta ^{2k}}{\rm{E(}}|X| + |{X^\prime }|{)^{2k}}}}{{(2k)!}}} \\ [\text{By}~\eqref{eq:Jensen}]& \le 1 + \sum\limits_{k = 1}^\infty {\frac{{{\theta ^{2k}}{{\rm{2}}^{2k}}{\rm{E}}|X{|^{2k}}}}{{(2k)!}}} < 1 + \sum\limits_{k = 1}^\infty {\frac{{{{(2\theta K_3\sqrt {2k} )}^{2k}}}}{{{k^k}k!}}} = 1 + \sum\limits_{k = 1}^\infty {\frac{{{{(8{\theta ^2}K_3^2)}^k}}}{{k!}}} = {e^{8{\theta ^2}K_3^2}}~\forall~\theta \in \mathbb{R} \end{align*} where the last inequality is due to ${(2k)!}>k^k\cdot k!$. \end{remark} \subsection{The variance proxy and sub-Gaussian norm} We show that the $\sigma^{2}$ in Definition \ref{def:Sub-Gaussian} is indeed the upper bounds of variance of $X$. The $\sigma^{2}$ not only characterizes the speed of decay in the sub-Gaussian tail probability, but also bounds the variance of $n^{-1 / 2} \sum_{i=1}^{n} X_{i}$. The $\operatorname{Var}X \le \sigma^{2} $ is because, by the sub-Gaussian MGF \begin{align}\label{eq:var} {\frac{\sigma^{2} s^{2}}{2}}+o(s^2)= e^{\frac{\sigma^{2} s^{2}}{2}} -1 \ge \mathrm{E}e^{s X}-1=s\mathrm{E} X +\frac{s^{2}}{2}\mathrm{E} X^2+\cdots= \frac{s^{2}}{2}\cdot\operatorname{Var}X+o\left(s^{2}\right) \end{align} (Dividing $s^{2}$ on both sides of \eqref{eq:var} and taking $s \rightarrow 0$). \begin{definition}[Sub-Gaussian norm]\label{def: sub-gaussian} For a sub-Gaussian r.v. $X$, the sub-Gaussian norm of $X$, denoted $ \|X\|_{\psi_2}$, is defined by: $\|X\|_{\psi_2} = \inf \{ t>0 :\; \mathrm{E} e^{X^2/t^2} \le 2 \}.$ \end{definition} From Corollary \ref{Sub-gaussian distribution}(4), $\|X\|_{\psi_2}$ is the smallest $K_4$. An alternative definition of the sub-Gaussian norm is $\|X\|_{\psi _2} := \sup_{p \ge 1} p^{-1/2} ({\rm{E}} |X|^p)^{1/p}$ \citep{Vershynin2010}. The definition for sub-Gaussian norm makes Corollary~\ref{Sub-gaussian distribution} easily presented. In fact, if $\mathrm{E} e^{X^2/\|X\|_{\psi_2}^2} \le 2$, \begin{equation} P(|X| \ge t)=P({e^{ {X^2}/{\|X\|_{\psi_2}^2}}} \ge {e^{ {t^2}/{\|X\|_{\psi_2}^2}}}) \le {{\mathrm{E}{e^{{X^2}/\|X\|_{\psi_2}^2}}}}/{{{e^{{t^2}/\|X\|_{\psi_2}^2}}}} \le 2{e^{ - {t^2}/{\|X\|_{\psi_2}^2}}}.\label{eq: sub-gaussian tail} \end{equation} \begin{example}[The sub-Gaussian norm of bounded r.vs.]\label{eg:sub-Gaussianboundedr.v.} Consider a r.v. $|X| \le M<\infty$. Set $\mathrm{E} e^{X^2/t^2}\le e^{M^2/t^2}\le 2 $ and $t\ge M/\sqrt{\log 2}$, By the definition of the sub-Gaussian norm, we have $\|X\|_{\psi_2}= M/\sqrt{\log 2}$. The equality holds if $|X| = M<\infty$. \end{example} \begin{example}[The sub-Gaussian norm of Gaussian r.vs.]\label{eg:sub-GaussianGaussianr.v.} For a $N(0,\sigma^{2})$ and $t > \sqrt 2 \sigma $, ${\rm{E}}{e^{{X^2}/{t^2}}}=\int {{e^{{x^2}/{t^2}}}\frac{e^{ - {{{{x^2}}/{2{\sigma ^2}}}}}}{ \sqrt{{{2\pi\sigma ^2}}}}} dx= \frac{t}{{{{({t^2} - 2{\sigma ^2}{\rm{)}}}^{1/2}}}}\int {{e^{{x^2}/{t^2}}}{{{\rm{(2}}\pi {\textstyle{{{\sigma ^2}{t^2}} \over {{t^2} - 2{\sigma ^2}}}}{\rm{)}}}^{ - 1/2}}} \exp \{ - \frac{{{x^2}}}{{2({\textstyle{{{\sigma ^2}{t^2}} \over {{t^2} - 2{\sigma ^2}}}})}}\} dx= \frac{{ t}}{{{{({t^2} - 2{\sigma ^2}{\rm{)}}}^{1/2}}}} \le 2 ~\Rightarrow~t \ge \sqrt {\frac{8}{3}} {\sigma }.$ By the definition, $\|X\|_{\psi_2}=\sqrt {\frac{8}{3}} \sigma > \sqrt {2} \sigma$. \end{example} However, the neat notation for defining sub-Gaussian norm sometime leads to unknown constants in the CIs as shown next. \begin{corollary}[Theorem 2.6.3, \cite{Vershynin18}]\label{coro:ghd} Let $\{ {X_i}\} _{i = 1}^n$ be independent mean-zero sub-Gaussian, $\forall~t \geq 0,~{P}\{|\frac{1}{n}\sum_{i=1}^{n} X_{i}| \geq t\} \leq 2 e^{-{C(n t)^{2}}/{\sum_{i=1}^{n}\|X_{i}\|_{\psi_{2}}^{2}}},~\forall~t \geq 0$ for a constant $C$. \end{corollary} The unknown constant $C$ makes the above CIs cannot be used in constructing confidence bands for $\mu$. To obtain more specific bounds (data dependent bounds as a statistics), we adopt the follow three propositions under sub-Gaussian. \begin{proposition}[Sub-Gaussian properties]\label{Sub-gaussianproperties} Let $X \sim \operatorname{subG}(\sigma^{2})$, then for any $t>0,$ \begin{enumerate}[\rm{(}a\rm{)}] \item \label{p: sub-gaussiantail} the tail satisfies $P(|X|>t) \leq 2e^{-{t^{2}}/{2 \sigma^{2}}} $; \item (a) implies that moments $\mathrm{E}|X|^{k} \leq(2 \sigma^{2})^{k / 2} k \Gamma(\frac{k}{2})$ and $(\mathrm{E}(|X|^{k}))^{1 / k} \leq \sigma e^{1 / e} \sqrt{k},~k \geq 2$; \item If \eqref{p: sub-gaussiantail} holds and $\mathrm{E} X = 0$, then $\mathrm{E}e^{s X}\leq e^{4 \sigma^{2} s^{2}}$ for any $s>0$; \item If $X \sim \operatorname{subG}\left(\sigma^{2}\right)$, then $\|X\|_{\psi_2} \le \frac{{2\sqrt 2}}{{\sqrt {\log 2} }}\sigma$; conversely, if $\|X\|_{\psi_2} = \sigma$ then $X \sim \operatorname{subG}(4\sigma^{2})$. \end{enumerate} \end{proposition} \begin{proof} The proofs of (a)-(c) are in Lemma 1.4 and 1.5 in \cite{Rigollet19}). The proofs of (a, b) is similar to Proposition \ref{pro-moment}(a, b) below. For (d), note that \begin{align}\label{eq:evenmoment} \mathrm{E} \exp \left(s^{2} X^{2}\right)=1+\sum_{k=1}^{\infty} \frac{s^{2 k} \mathrm{E}X^{2 k}}{k !}&\stackrel{\rm (b)}{\le} 1+\sum_{k=1}^{\infty} \frac{2s^{2 k} (2 \sigma^{2})^{k} k \Gamma(k)}{k !}=1+4s^{2}\sigma^{2}\sum_{k=0}^{\infty} (2s^{2}\sigma^{2})^{k}\nonumber\\ &\xlongequal{\forall~|2s^{2}\sigma^{2}|<1}1+ \frac{4s^{2}\sigma^{2}}{1-2s^{2}\sigma^{2}} \stackrel{\forall~|s| \le 1/({2}\sigma)}{\le} 1+ 8s^{2}\sigma^{2} \le e^{8 \sigma^{2} s^{2}}. \end{align} By \eqref{eq:evenmoment}, set $\mathrm{E} e^{s_0^{2} X^{2}}\le {e^{8{{s_0}^{\rm{2}}}\sigma^2}} \le {\rm{2}}$ for some $s_0$. Then $s_0^2 \le \frac{{\log 2}}{{8\sigma^2}}$ and $|s_0| \le \frac{{\sqrt {\log 2} }}{{2\sqrt 2\sigma}} \le \frac{1}{2\sigma}$. Put $|{s_0}| =\frac{{\sqrt {\log 2} }}{{2\sqrt 2\sigma}} $ and the sub-Gaussian norm gives $ {\rm{E}}{e^{ {X^2}/(\frac{{2\sqrt 2\sigma}}{{\sqrt {\log 2} }})^2}} \le {\rm{2}} \Rightarrow {\left\| X \right\|_{{\varphi _2}}} \le \frac{{2\sqrt 2\sigma}}{{\sqrt {\log 2} }}.$ Conversely, if $\|X\|_{\psi_2}= \sigma$ then \eqref{eq: sub-gaussian tail} gives $$P(|X| > t) \le 2{e^{ - {t^2}/{\sigma ^2}}}{\rm{ = }}2{e^{ - \frac{{{t^2}}}{{2{{(\sigma /\sqrt 2 )}^2}}}}}.$$ Then Proposition \ref{Sub-gaussianproperties}(c) concludes $\mathrm{E}e^{s X}\leq {e^{4{{(\sigma /\sqrt 2 )}^2}{s^2}}} = {e^{4{\sigma ^2}{s^2}/2}}$ for any $s>0$, and we have $X \sim \operatorname{subG}(4\sigma^{2})$. \end{proof} Let $\{ {Y_i}\} _{i = 1}^n$ be a sequence of \textit{exponential family} (EF) r.vs, its density \begin{equation}\label{eq:E-P} f(y_i;\theta_i ) = c(y_i)\exp \{y_i\theta_i - b(\theta_i )\} \end{equation} with ${\rm{E}}Y_i=\dot{b}(\theta_i)$ and ${\rm{Var}}Y_i=\ddot{b}(\theta_i)$. We next introduce the sub-Gaussian CIs for the non-random weighted sum of EF r.vs with compact parameter space, adapted from Lemma 6.1 in \cite{Rigollet12} with more specific constants. \begin{proposition}[Concentration for weighted E-F summation] \label{pro-moment} We assume \eqref{eq:E-P} and \begin{itemize} \item [\textbullet]\emph{(E.1)}: Uniformly bounded variances condition: there exist a compact set $\Omega$ and some constant $C_{b}$ such that $\mathop {\sup }_{\theta_i \in \Omega } \ddot b (\theta_i ) \le {C_{ b}^2}~\text{for all}~i$. \end{itemize} Let $\bm w := ({w_1}, \cdots ,{w_n})^T \in {\mathbb{R}^n}$ be a non-random vector and define $S_n^w = :\sum_{i = 1}^n {{w_i}{Y_i}} $. Then \begin{enumerate}[\rm{(}a\rm{)}] \item Closed under addition: $S_{n}^{w}-{\rm{E}}S_{n}^{w} \sim {\rm{subG}}({C_b^{2}\|\bm w\|_{2}^{2}})$ and $P\{|S_{n}^{w}-{\rm{E}}S_{n}^{w}|>t\}\leq2 e^{-{t^{2}}/(2C_b^{2}\|\bm w\|_{2}^{2})}$; \item Let $C_n:=S_{n}^{w}-{\rm{E}}S_{n}^{w}$ and $\Gamma (t):=\int_{0}^\infty x^{t-1} e^{-x} dx$ be the Gamma function. For all integer $k \ge 1$, we have moments bound: ${\rm{E}}|C_n|^k \le k{(2{C_b^2})^{k/2}}\Gamma (k/2)\left\|\bm w \right\|_2^k.$ \item The MGF of centralized $|C_n|^2$: ${\rm{E}}e^{s[|C_n|^2-{\rm{E}}|C_n|^2]} \le e^{{s^{\rm{2}}}{{({\rm{8}}\sqrt {\rm{2}} C_b^2\left\| w \right\|_2^2)}^2}/2},~\forall ~ | s | \le ({{\rm{8}}C_b^2\|\bm w \|_2^2})^{-1}.$ \item \emph{In this case, we do not assume \eqref{eq:E-P} and (E.1)}. Suppose $\{ Y_{i}-{\rm{E}}Y_{i}\} _{i = 1}^n $ are independent distributed as $\{\operatorname{subG}(\sigma_i^{2})\} _{i = 1}^n$ with ${C_b^2}=:\mathop {\max }_{1 \le i \le n} \sigma_i^{2}>0$, then $\rm{(a)-(c)}$ also hold. \end{enumerate} \end{proposition} \begin{proof} Based on the MGF and uniformly bounded variances condition, the proof of (a) can be found in Lemma 6.1 of \cite{Rigollet12}. In the proof of (c), we update the constant, and (d) is similar for the sub-Gaussian case. \item (b) The proof relies on expectation formula for positive r.v. (in terms of integral of tail probability) which transforms tail bound to moment bound. For any integer $k \ge 1$, \begin{align}\label{eq:EFmgf} \mathrm{E}\left| S_{n}^{w}-\mathrm{E} S_{n}^{w} \right|^k& = \int_{0}^\infty P( \left| S_{n}^{w}-\mathrm{E} S_{n}^{w} \right|^k \ge s ) ds\xlongequal{t =s^{1/k}} \int_{0}^\infty k t^{k-1}P\left( \left| S_{n}^{w}-\mathrm{E} S_{n}^{w} \right| \ge t \right) dt. \end{align} Applying tail bound in (a), we have by letting ${D_{k,C}} = k{(2{C_b^2})^{k/2}}\Gamma (k/2)$ \begin{align*} \mathrm{E}\left| S_{n}^{w}-\mathrm{E} S_{n}^{w} \right|^k &\le 2k \int_{0}^\infty t^{k-1} e^{-\frac{t^{2}}{2C_b^{2}\|\bm w\|_{2}^{2}}} dt \xlongequal{x ={t^{2}}/{2C_b^{2}\|\bm w\|_{2}^{2}}} k (2C_b^2)^{\frac{k}{2}} \|\bm w\|_{2}^{k} \int_{0}^\infty x^{{\frac{k}{2}}-1} e^{-x} dx={D_{k,C}}\left\| \bm w \right\|_2^k. \end{align*} \item (c) The proof will resort to $(\mathrm{E}|Z|)^{{k}} \leq \mathrm{E}|Z|^{k}$ and Jensen's inequality \begin{align}\label{eq:Jensen} {(\frac{{|a| + |b|}}{2})^k} \le \frac{1}{2}{|a|^k} + \frac{1}{2}{|b|^k},~\text{for integer}~k \ge 1. \end{align} From Taylor's expansion, \eqref{eq:Jensen} gives \begin{align*}\label{eq:EFmgf} {\rm{E}}e^{s[|C_n|^2-{\rm{E}}|C_n|^2]}& = 1 + \sum_{k=2}^\infty \frac{s^{k} \mathrm{E} [|C_n|^2-\mathrm{E}|C_n|^2]^{k}}{k!}\le 1 + \sum\limits_{k = 2}^\infty {\frac{{{s^k}{{\rm{2}}^{k - 1}}{\rm{E}}\{ {|C_n|^{2k} + {{\left( {{\rm{E}}|C_n|{^2}} \right)}^k}} \}}}{{k!}}}\\ [\text{By}~(\mathrm{E}|Z|)^{{k}} \leq \mathrm{E}(|Z|^{k})]~~&\le1 + \sum\limits_{k = 2}^\infty {\frac{{{s^k}{{\rm{2}}^{k - 1}}{\rm{E}}\left\{ {|C_n|^{2k}+ {\rm{E}}|C_n|^{2k}} \right\}}}{{k!}}}\le1 + \sum\limits_{k = 2}^\infty {\frac{{{s^k}{{\rm{2}}^k} \cdot 2k{{(2C_b^2\left\|\bm w \right\|_2^2)}^k}\Gamma (k)}}{{k!}}}, \end{align*} where the last inequality is by Proposition~\ref{pro-moment}(b). Then, under $| {4sC_b^2\left\| \bm w \right\|_2^2} | < {\rm{1}}$, we have \begin{align*} {\rm{E}}e^{s[|C_n|^2-{\rm{E}}|C_n|^2]}&= 1 + 2\sum\limits_{k = 2}^\infty {{{(4sC_b^2\left\|\bm w \right\|_2^2)}^k}} = 1 + \frac{{2{{(4sC_b^2\left\| \bm w \right\|_2^2)}^2}}}{{1 - 4|s|C_b^2\left\| \bm w \right\|_2^2}}\\ [| {4sC_b^2\| \bm w \|_2^2} | \le \textstyle{{\rm{1}}\over{{\rm{2}}}} \Leftrightarrow | s |~ \le \textstyle{{{\rm{1}}}\over{{{\rm{8}}C_b^2\|\bm w \|_2^2}}}]~~&\le 1 + \frac{{{s^{\rm{2}}}{{({\rm{8}}\sqrt {\rm{2}} C_b^2\left\| \bm w \right\|_2^2)}^2}}}{{\rm{2}}} \le {{\rm{e}}^{{s^{\rm{2}}}{{({\rm{8}}\sqrt {\rm{2}} C_b^2\left\| \bm w \right\|_2^2)}^2}/2}}. \end{align*} \item (d) It follows by defining ${C_b^2}=:\mathop {\max }\limits_{1 \le i \le n} \sigma_i^{2}>0$ as the common variance proxy for $\{ {Y_i}\} _{i = 1}^n$. For $i=1,2,\cdots,n$, we have: $\mathrm{E}e^{sw_i(Y_{i}-{\rm{E}}Y_{i})}\leq e^{\sigma^{2} s^{2}w_i^{2}/{2}}$, $\forall s \in \mathbb{R}$. \end{proof} Proposition~\ref{pro-moment}(a) yields the following results (The first result is in Corollary 1.7 of \cite{Rigollet19}). The second sub-Gaussian CI below specifies the unknown constant in Theorem 2.6.2 of \cite{Vershynin18}. \begin{proposition}\label{Sub-gaussianConcentration} Let $\{X_{i}\}_{i=1}^n$ be $n$ independent $\operatorname{subG}(\sigma_i^{2})$. Define ${\sigma ^2} = \mathop {\max }_{1 \le i \le n} \sigma _i^2$, \begin{center} $P( {| {\sum\limits_{i = 1}^n {{w_i}} {X_i}}| > t} ) \le 2e^{{ - {{{t^2}}}/({2{\sigma ^2}\|\bm w\|_2^2})}}$ and ${P}( {| {\sum\limits_{i = 1}^n {{w_i}} {X_i}}| > t} ) \le 2 e^{-{ t^{2}}/(8\sum_{i=1}^{n}\|{w_i}X_{i}\|_{\psi_{2}}^{2})},~\forall~t \geq 0$ \end{center} for any non-random vector $\bm w := ({w_1}, \cdots ,{w_n})^T$. \end{proposition} \begin{proof} To see the second CI, just use the Proposition \ref{Sub-gaussianproperties}(d) and the Proposition~\ref{pro-moment}(d), by noticing that if $\|X_i\|_{\psi_2} < \infty$ then ${w_i} X_i \sim \operatorname{subG}(4\|{w_i} X_i\|_{\psi_2}^2)$. \end{proof} \subsection{Randomly weighted sum of independent sub-Gaussian variables}\label{se;weights} In this part, we outline the sub-Gaussian type CIs for the randomly weighted sum of exponential family r.vs: $S_n^{W } = :\sum_{i = 1}^n {{W_i}{Y_i}}$ where $\{W_i\}_{i = 1}^n$ are called the multipliers (or random weights) which are independent from $\{Y_i\}_{i = 1}^n$. The normalized sum $\frac{1}{{\sqrt n }}(S_n^W - {\rm{E}}S_n^W)$ is also call \emph{multiplier empirical processes}, and it serves for the multiplier Bootstrap inference where the multipliers $\{W_i\}$ are r.vs independent from $\{Y_i\}_{i = 1}^n$, see Chapter 2.9 of \cite{van96}. To get sub-Gaussian concentration, some regularity conditions for the parameter space are required. \begin{itemize} \item [\textbullet](E.2): Let $\bm W := ({W_1}, \cdots ,{W_n})^T \in {\mathbb{R}^n}$ be a random vector with some bounded components, i.e. $| W_{i}| \leq w_i <\infty$ for a non-random vector $\bm w := ({w_1}, \cdots ,{w_n})^T \in {\mathbb{R}^n}$. \end{itemize} \begin{theorem}[Concentration inequalities for randomly weighted sum] \label{lem-weighted} Let $\{ {Y_i}\} _{i = 1}^n$ belong to the canonical exponential family \eqref{eq:E-P}, and let $\{ {W_i}\} _{i = 1}^n$ be independent of $\{ {Y_i}\} _{i = 1}^n$. Define the randomly weighted sum $S_n^{W } = :\sum_{i = 1}^n {{W_i}{Y_i}} $, then under \rm{(E.1)} and \rm{(E.2)} \[P( {|S_n^W - {\rm{E}}S_n^W| \ge t} ) \le 2e^{-{ t^{2}}/({2C_{ b}^{2}\|\bm w\|_{2}^{2}})}.\] \end{theorem} \begin{proof} Let $Y_i = \dot{b}(\theta_i)+ Z_i$, where $\{ {Z_i}\} _{i = 1}^n$ are centralized and independent E-F r.vs. From ${\rm{E}}Y_i=\dot{b}(\theta_i)$ and the identity \eqref{eq: identity} for a dominating measure $\mu(\cdot)$ \begin{align}\label{eq: identity} \smallint {dF_{Y_i}(y)} = 1 \Leftrightarrow \smallint {c({y}){e^{{y}{\theta _i}}}\mu (dy)} =e^{b({\theta _i})}. \end{align} Let ${\rm{E}}_{\cdot|\bm W}(\cdot):={\rm{E}}(\cdot |\bm W)$ and $s$ be in $(-\delta,\delta )$ (a neighbourhood of zero). Then, \begin{align*} {\rm{E}}_{\cdot|\bm W}[{\rm{e}}^{s{W_i}{Y_i}}] & = \int {\rm{e}}^{s{W_i}{Y_i}}dF_{Y_i|\bm W}(y)= \int {\rm{e}}^{s{W_i}{Y_i}}dF_{Y_i}(y)~~[\text{by}~\{W_i\}_{i = 1}^n \bot \{Y_i\}_{i = 1}^n] \\ & = \int {c(y)}e^{y{\theta _i} - b ({\theta _i})}e^{s{W_i}y}\mu(dy)\xlongequal{\eqref{eq: identity}} e^{ b ({\theta _i} + s{W_i}) - b ({\theta _i})}. \end{align*} It can be easily derived from (E.2) and Taylor's expansion, \begin{align}\label{eq:mgf} {\rm{E}}_{\cdot|\bm W}[{\rm{e}}^{s({W_i}{Y_i}-{\rm{E}}_{\cdot|\bm W}({W_i}{Y_i}))}] &= e^{ b ({\theta _i} + s{W_i}) - b ({\theta _i}) - \dot b ({\theta _i}){W_i}s}\xlongequal{\exists {\tilde \eta _{i}} \in [{\theta _i},{\theta _i} + s{W_i}]} e^{\frac{{{s^2}W_i^2}}{2}\ddot b (\tilde \eta _i)}\le e^{\frac{{{s^2}{C_{b}^2}W_i^2}}{2}}. \end{align} By the conditional independence for $\{{W_i}{Z_i}|\bm W\}$ and \eqref{eq:mgf}, it follows that when $s \in (-\delta,\delta )$ \begin{align}\label{eq:Cmgf} {{\rm{E}}_{ \cdot |\bm W}}[e^{s\sum_{i = 1}^n [ {W_i}{Z_i} - {\rm{E}}_{\cdot|\bm W}({W_i}{Z_i})]} ]&=\prod\limits_{i = 1}^n {{{\rm{E}}_{ \cdot |\bm W}}e^{ s[{W_i}{Z_i} - {\rm{E}}_{\cdot|\bm W}({W_i}{Z_i})]} } \le \prod_{i=1}^n e^{\frac{ s^2C_{b}^2W_i^2}{2}} \le 2e^{\frac{ s^{2} C_{b}^{2}\left\| \bm w \right\|_{2}^{2}}{2} }, \end{align} where the last inequality is from $\{| W_{i}| \leq w_i\}$ for a non-random vector $\bm w := ({w_1}, \cdots ,{w_n})^T$. By the conditional Markov's inequality and symmetry of $Z_i$, we have, as $s \in (-\delta,\delta )$ \begin{align}\label{eq:P2} P(|\sum{}_{i = 1}^n [ {W_i}{Z_i} - {\rm{E}}_{\cdot|\bm W}({W_i}{Z_i})]| \ge t|\bm W )&\le \mathop {\inf }\limits_{s > 0}[ {e^{-s t}}{ {{\rm{E}}_{ \cdot |\bm W}}e^{ s(\tilde S_{n}^{W}-{\rm{E}}_{\cdot|\bm W}\tilde S_{n}^{W})}}+{e^{-s t}}{ {{\rm{E}}_{ \cdot |\bm W}}e^{ -s(\tilde S_{n}^{W}-{\rm{E}}_{\cdot|\bm W}\tilde S_{n}^{W})}}]\nonumber\\ &\le 2\mathop {\inf }\limits_{s > 0}e^{\frac{ s^{2} C_{b}^{2}\left\| \bm w \right\|_{2}^{2}}{2} - s t }=2e^{-\frac{t^{2}}{2C_{b}^{2}\|\bm w\|_{2}^{2}}} \end{align} where the last equality is minimized by setting $s = t/(C_{b}^2\|\bm w\|_2^2)$. \end{proof} Note that Lemma 6.1 in \cite{Rigollet12} is about the concentration for the non-random weighted sum of exponential family r.vs. The assumption of compact parameter space for exponential family is vital for obtaining the sub-Gaussian type concentration. If we do not impose condition (E.2) and the assumption that $\{W_i\}_{i=1}^n$ and $\{Y_i\}_{i=1}^n$ are dependent, a counterexample for sub-Gaussian concentration is $W_i=Y_i$. Thus, $S_n^W$ is a quadratic form, and $S_n^W-{\rm{E}} S_n^W$ is sub-exponential by Lemma~\ref{lem: sub-exponential squared} below. If $\{W_i\}_{i=1}^n$ and $\{Y_i\}_{i=1}^n$ are dependent but $\{W_i\}_{i=1}^n$ are still bounded, another counterexample is $W_i=\rm{sign}(Y_i)$. Therefore, $S_n^{W } =\sum_{i = 1}^n {|Y_i|} $ is not zero-mean, and the concentration of $\sum_{i = 1}^n {|Y_i|} $ fails. \subsection{Concentration for Lipschitz functions of random vectors}\label{Lipschitz} In the analyses of high-dimensional statistics by empirical processes, researches often resort to the CIs of Lipschitz functions for either bounded or strongly log-concave random vectors \citep{Wainwright19}. \begin{lemma}[Theorem 2.26, \cite{Wainwright19}]\label{lem:caussiancon} Let $\bm N \sim N(\bm 0, \mathrm{I}_p)$. Let $f:\mathbb{R}^n \rightarrow \mathbb{R}$ be L-Lipschitz with respect to (w.r.t.) the Euclidean norm: $| f ( \bm a ) - f ( \bm b ) | \leq L \| \bm a - \bm b \| _ { 2 }$ for any $\bm a , \bm b \in \mathbb { R } ^ { n }$. Then, $P\left(|f(\bm N)-\mathrm{E} f(\bm N) |\geq t\right) \leq 2e^{-t^2 /(2L^2)},~\forall~t>0$. \end{lemma} A non-negative function $f(\bm x):\mathbb{R}^n \rightarrow \mathbb{R}$ is \emph{log-concave} if \begin{align}\label{eq:log-concave} \log f(\lambda \bm x+(1-\lambda) \bm y) \geq {\lambda}\log f(\bm x)+ ({1-\lambda})\log f(\bm y),~\forall~\lambda \in[0,1]~\text{and}~\bm x, \bm y \in \mathbb{R}^{n}. \end{align} A function $\psi(\bm x) : \mathbb{R}^{n} \rightarrow \mathbb{R}$ is $\gamma$-\emph{strongly concave} if there is some $\gamma>0$ s.t. \begin{center} $ \lambda \psi(\bm x)+(1-\lambda) \psi(\bm y)-\psi(\lambda \bm x+(1-\lambda) \bm y) \le \frac{\gamma}{2} \lambda(1-\lambda)\|\bm x-\bm y\|_{2}^{2},~\forall~\lambda \in[0,1]~\text{and}~\bm x, \bm y \in \mathbb{R}^{n}, $ \end{center} A continuous probability density $f(\bm x)$ and the corresponding r.v. are log-concave (or strongly log-concave) if $f(\bm x)$ is a log-concave function (or strongly log-concave function), see \cite{Saumard14} for a review of the log-concavity in statistics. \begin{lemma}[Theorem 3.16, \cite{Wainwright19}]\label{3.16} Let $\mathbb{P}$ be any $\gamma-$strongly log-concave distribution on $\mathbb{R}^n$ with parameter $\gamma>0$. Then for any function $f: \mathbb{R}^{n} \rightarrow \mathbb{R}$ that is $L$-Lipschitz w.r.t. the Euclidean norm, we have \begin{center} ${P}[f(X)-\mathrm{E}f(X) \geq t] \leq e^{-\frac{\gamma t^{2}}{4 L^{2}}}$ for $X \sim \mathbb{P}$ and $t\ge 0$. \end{center} \end{lemma} The standard Gaussian random vector is $1-$strongly log-concave distributed. However, Lemma~\ref{lem:caussiancon} has the sharper constant ${2L^{2}}$ than the Gaussian case of Lemma \ref{3.16} with constant ${4 L^{2}}$. Beyond Gaussian and strongly log-concave, it is possible to establish concentration for distributions involving bounded r.vs. A function $f(\bm x): \mathbb{R}^{n} \rightarrow \mathbb{R}$ is said to be \emph{separately convex} if, the univariate function $y_{k} \mapsto f\left(x_{1}, x_{2}, \ldots, x_{k-1}, y_{k}, x_{k+1}, \ldots, x_{n}\right)$ for each index $k \in\{1,2, \ldots, n\}, $ is convex for each fixed vector $\left(x_{1}, x_{2}, \ldots, x_{k-1}, x_{k+1}, \ldots, x_{n}\right) \in \mathbb{R}^{n-1} .$ \begin{lemma}[Theorem 3.4, \cite{Wainwright19}]\label{3.17} Let $\left\{X_{i}\right\}_{i=1}^{n}$ be independent r.vs, each supported on the interval $[a, b]$. Let $f: \mathbb{R}^{n} \rightarrow \mathbb{R}$ be separately convex, and $L$-Lipschitz w.r.t. the Euclidean norm. Then, $ {P}[f(X)- \mathrm{E}f(X) \ge t] \leq e^{-{t^{2}}/{4 L^{2}(b-a)^{2}}}. $ for $X \sim \mathbb{P}$ and $t\ge 0$. \end{lemma} \begin{example}[Order Statistics] From Lemma \ref{3.16} and Lemma \ref{3.17}, suppose that $\left\{X_{i}\right\}_{i=1}^{n}$ are independent r.vs which are $\gamma-$strongly log-concave distributed satisfying ${P}[f(X)-\mathrm{E}f(X) \geq t] \leq e^{-\frac{\gamma t^{2}}{4 L^{2}}}$ for any function $f: \mathbb{R}^{n} \rightarrow \mathbb{R}$ that is $L$-Lipschitz w.r.t. the Euclidean norm. Let $X_{(k)}$ be the $k$-th order statistic of $X_{1}, \ldots, X_{n}$, it can be shown that \begin{center} ${P}(|X_{(k)}-\mathrm{E}X_{(k)}|>\delta) \leq 2 e^{-\delta^{2} / 2}$ by checking $|X_{(k)}-Y_{(k)}| \leq\|X-Y\|_{2}$, i.e. $L=1$. \end{center} Indeed, we have \begin{center} $X_{(k)}-Y_{(k)} \le \left|X_{l}-Y_{l}\right| \leq\|X-Y\|_{2}$ for some $l \in \{1,2,\cdots,n\}$. \end{center} More results of the tail bounds for the order statistics of IID r.vs are reported in \cite{Boucheron12}. \end{example} \section{Sub-exponential Distributions}\label{sec-se} \subsection{Characterizations} The requirement in definition of sub-Gaussian r.v. $\mathrm{E}e^{s X}\le e^{\frac{\sigma^{2} s^{2}}{2}},~\forall s \in \mathbb{R}$ is too strong. We consider the MGF of exponential distributions: \begin{example} [MGF of exponential distributions]\label{eg:Exponential} Consider the exponential r.v. $X \sim {\rm{Exp}}(\mu)$ with $\mathrm{E}X=\mu>0$. The MGF of $X-\mu$ satisfies \begin{align}\label{eq:Exponentialmgf} {\rm{E}}{e^{s(X - \mu )}} = \frac{e^{ - s\mu }}{{1 - {\rm{s}}\mu } } = {({e^{ - s\mu/2 }}{\left( {1 - {\rm{s}}\mu } \right)^{ - 1/2}})^2} \le {e^{2{{(s\mu /2)}^2}}} < {e^{{s^2}(2\mu)^2/2}}, \forall ~ |s| \le (2\mu)^{-1} \end{align} where the second last inequality is by ${e^{-t}}/{\sqrt{1-2 t}} \leq e^{2 t^{2}}$ for $|t|\le 1/4$. (This is from the property of $f(t): = (1 - 2t){e^{4{t^2} + 2t}}$ with $f(0)=1$: (a). $f'(t) > 0,~0 < t < 1/4$; (b). $f(t) \ge 1, ~- 1/4 < t < 0$). \end{example} In \eqref{eq:Exponentialmgf}, the MGF of the exponential r.v. is divergent on $s=1/\mu$ and it cannot be bounded by a Gaussian MGF of $s$ in $\mathbb{R}$, and the exponential MGF is bounded by Gaussian MGF for $|s| \le \frac{1}{{2\mu }}$ via inequality \eqref{eq:Exponentialmgf}. Motivated by Example \ref{eg:Exponential}, the first definition of sub-exponential distribution \eqref{eq:sube1} below is exactly the locally sub-Gaussian property. \begin{definition}[Sub-exponential distributions]\label{def:Sub-exponential} A r.v. $X \in \mathbb{R}$ with $\mathrm{E} X = 0$ is sub-exponential with parameter $\lambda$ (denoted $X \sim \operatorname{subE}(\lambda))$ if its MGF satisfies \begin{equation}\label{eq:sube1} {\mathrm{{E}}}e^{s X} \leq e^{\frac{s^{2}\lambda ^{2}}{2}} \quad \text{for all } |s| <{1}/{\lambda}. \end{equation} In \cite{Wainwright19}, sub-exponential distributions are generally defined by two positive parameters $(\lambda, \alpha)$ (denoted $X \sim \operatorname{subE}(\lambda, \alpha)$): ${\mathrm{{E}}}e^{s X} \leq e^{\frac{s^{2}\lambda ^{2}}{2}}~\text { for all }|s|<{1}/{\alpha}.$ \end{definition} The $\lambda^2$ in \eqref{eq:sube1} is treated as a \emph{variance proxy} and ${\alpha}$ is seen as \emph{locally sub-Gaussian factor}, see Remark \ref{re:locally sub-Gaussian} later. Specifically, $\operatorname{subE}(\lambda)=\operatorname{subE}(\lambda, \lambda)$. Sub-Gaussian r.vs are sub-exponential by definition, but not vice verse. In Corollary \ref{Sub-gaussian distribution}, one equivalence of sub-Gaussian r.vs is that the survival function is bounded by the Gaussian-like survival function up to a constant. Similarly, the sub-exponential r.v. has a characterization that the survival function is bounded by that of a exponential distribution. Similar to sub-Gaussian characterizations, there are at least six equivalent forms for sub-exponential distributions which are useful for checking the sub-exponential distribution. \begin{corollary}[Characterizations of sub-exponential]\label{prop: sub-exponential properties} Let $X$ be a r.v. in $\mathbb{R}$ with $\mathrm{E} X = 0$. Then the following properties are equivalent, where $\{K_i\}_{i=1}^6$ are positive constants. \begin{enumerate}[\rm{(}1\rm{)}] \item \label{p: exponential tail} The tails of $X$ satisfy $ P \{ |X| \ge t \} \le 2 e^{-t/K_1}~\text{for all } t \ge 0; $ \item \label{p: exponential MGF} The MGF of $X$ satisfies $ \mathrm{E} e^{\l X} \le e^{K_2^2 \l^2}~\text{for all}~|\l| \le \frac{1}{K_2}; $ \item \label{p: exponential moments} The moments of $X$ satisfy $ (\mathrm{E} |X|^p)^{1/p} \le K_3 p~\text{for integer}~p \ge 1; $ \item \label{p: exponential MGF abs} The MGF of $|X|$ satisfies $ \mathrm{E} e^{\l |X|} \le e^{K_4 \l}~\text{for all } 0 \le \l \le \frac{1}{K_4}; $ \item \label{p: exponential MGF finite} The MGF of $|X|$ is bounded at some point: $\mathrm{E} e^{|X|/K_5} \le 2$; \item \label{p: exponential MGF finite2} Bounded MGF of $X$ in a compact set: $ \mathrm{E} e^{tX} < \infty,~\forall |t|< 1/K_6. $ \end{enumerate} \end{corollary} The zero mean is only used in the proof of \eqref{p: exponential MGF} of Corollary \ref{prop: sub-exponential properties}. The equivalence among \eqref{p: exponential tail}--\eqref{p: exponential MGF finite} is proved in \cite{Vershynin18} and that between \eqref{p: exponential MGF finite} and \eqref{p: exponential MGF finite2} can be found in Lemma 5 of \cite{Petrov1995}. The \eqref{p: exponential MGF finite2} is the called \emph{Cramer's condition} which is an essential characterization, it signifies that: \emph{All r.vs. are sub-exponential if their MGF exist in a neighborhood of zero.} \cite{Pistone1999} names the property \eqref{p: exponential MGF finite2} as the \emph{exponentially integrable} r.v. \begin{example}[Moment of exponential distributions]\label{eg:conExponential} The $P(X - \mu \ge t) = {e^{ - (t + \mu )/\mu }} \le {e^{ - t/\mu }}$ and the symmetry of $X - \mu$ implies $K_1=\mu$ in Corollary \ref{prop: sub-exponential properties}. Continue to Example \ref{eg:Exponential}, the ``$\le$'' in \eqref{eq:Exponentialmgf} implies $ {\rm{E}}{e^{s(X - \mu )}} \le {e^{{{(s\mu /{\sqrt 2})}^2}}}\le {e^{{{(2s\mu )}^2}}}, \forall ~ |s| <(2\mu)^{-1}$. So $ K_2=\mu/\sqrt 2 $ and $ K_6=2\mu$ in Corollary \ref{prop: sub-exponential properties}. Next, we evaluate the moment of $X$ for any $p \geq 1,$ $ \mathrm{E}|X|^{p}=\int_{0}^{\infty} x^{p} \cdot \mu^{-1} e^{-\mu^{-1} x} d x\xlongequal{y=\mu^{-1} x} {\mu^{p}} \int_{0}^{\infty} y^{p} e^{-y} d y={\Gamma(p+1)}{\mu^{p}}. $ By $\Gamma(p+1) \leq p^{p}$ for $p \geq 1,$ it gives: $(\mathrm{E}|X|^{p})^{1 / p}={(\Gamma(p+1))^{1 / p}}{\mu} \leq {p}\mu$. Via \eqref{eq:Minkowski} shows that $(\mathrm{E}|X-\mu|^{p})^{1 / p}\leq 2{p}\mu$ and thus $ K_3=2\mu$ in Corollary \ref{prop: sub-exponential properties}. Assume ${\mathrm{E}}X=0$, then by Stirling's approximation $p ! \geq (p / e)^{p}$ \begin{align}\label{eq:subE6} \mathrm{E}{e^{\lambda |X|}} = 1 + \sum\limits_{p = 2}^\infty {\frac{{{\lambda ^p}\mathrm{E}|X{|^p}}}{{p!}}} & \le 1 + \sum\limits_{p = 2}^\infty {\frac{{{{(\lambda {K_3}p)}^p}}}{{{{(p/e)}^p}}}} = 1 + \sum\limits_{p = 2}^\infty {{{(e{K_3}\lambda )}^p}} = 1 + \frac{{{{(e{K_3}\lambda )}^2}}}{{1 - e{K_3}\lambda }},~\forall~|e{K_3}\lambda | < 1\nonumber\\ (\text{Restrict}~e{K_3}\lambda \le 1/2) ~~& \le 1 + 2{(e{K_2}\lambda )^2} \le {e^{2{{(e{K_3}\lambda )}^2}}} \le {e^{e{K_3}\lambda }} \le{e^{2e{K_3}\lambda }},~\forall~\lambda \le 1/(2e{K_3}). \end{align} Thus $K_4=2e{K_3}=4e\mu$. That $\mathrm{E}{e^{\lambda |X|}} \le {e^{e{K_3}\lambda }}~\text{for}~0<\lambda \le 1/(2e{K_2})$ in \eqref{eq:subE6} implies $\mathrm{E}{e^{|X|/(2e{K_3})}} < {e^{1/2}} < 2$. Hence $K_5={K_3}$. \end{example} \begin{example} [Geometric distributions] {\color{black}{The \emph{geometric distribution} $X\sim \mathrm{Geo} (q) $ for r.v. $X$ is defined by: $P(X=k) ={(1 - q)} {q^{k-1}},(q \in (0,1),~k=1,2,\cdots)$. The mean and the variance of ${\rm{Geo}} (q) $ are ${(1 - q)}/{{q}}~\text{and}~{(1 - q)}/{{{{q}^2}}}$ respectively. Apply Lemma 4.3 in \cite{Hillar2013}, we get ${({\rm{E}}|X{|^k})^{1/k}} <{{-2k}}/{{ \log (1 - q)}}$. It follows from the Minkowski's inequality and Jensen's inequality $(\mathrm{E}|Z|)^{{k}} \leq \mathrm{E}|Z|^{k}$ for integer $k \ge 1$ that \begin{equation}\label{eq:Minkowski} (\mathrm{E}|X-\mathrm{E}X|^{k})^{1 / k} \le (\mathrm{E}|X|^{k})^{1 / k}+|\mathrm{E}X| \leq 2 (\mathrm{E}|X|^{k})^{1 / k}\le {{-4k}}/{{ \log (1 - q)}} \end{equation} and Corollary \ref{prop: sub-exponential properties}(3) implies the \textit{centralized $\mathrm{Geo} (q)$} is sub-exponential with $K_3={{-4}}/{{ \log (1 - q)}}$.}} \end{example} \begin{example}[Discrete Laplace r.vs] A r.v. $X \sim \mathrm{DL}(q)$ obeys the discrete Laplace distribution if $f_{q}(k)=\mathbb{P}(X=k)=\frac{1-q}{1+q} q^{|k|},~k \in \mathbb{Z}=\{0,\pm 1,\pm 2, \ldots\}$ with parameter $q \in(0,1).$ The discrete Laplace r.v. is the difference of two IID $\mathrm{Geo} (q) $. The geometric distribution is sub-exponential, thus Corollary \ref{sub-exponentialConcentration}(a) mentioned later implies that the discrete Laplace is also sub-exponential distributed. In differential privacy of network models, the noises are assumed following the discrete Laplace distribution, see \cite{Fan20} and references therein. \end{example} The next result shows that a sum of independent sub-exponential r.vs has two tails with difference convergence rate, which is slightly different from Hoeffding's inequality. Deviating from the mean, it tells us that the tail of the sum of sub-exponential r.vs behaves like a combination of a Gaussian tail and a exponential tail. \begin{corollary}[Concentration for weighted sub-exponential sums]\label{sub-exponentialConcentration} Let $\{ X_{i}\} _{i = 1}^n $ be independent $\{\operatorname{subE}(\lambda_i)\} _{i = 1}^n$ distributed with zero mean. Define $\lambda= \mathop {\max }_{1 \le i \le n} \lambda_i>0$ and the non-random vector $\bm w := ({w_1}, \cdots ,{w_n})^T \in {\mathbb{R}^n}$ with $w= \mathop {\max }_{1 \le i \le n} |w_i|>0$, we have \begin{enumerate}[\rm{(}a\rm{)}] \item Closed under addition: $\sum_{i = 1}^n {{w_i}{X_i}} \sim {\rm{subE}}(\| \bm w \|_2{\lambda} )$; \item $P( {| {\sum\limits_{i = 1}^n {{w_i}{X_i}} }| \ge t} ) \le 2e^{ { - \frac{1}{2}( {\frac{{{t^2}}}{{\left\|\bm w \right\|_2^2{\lambda ^2}}} \wedge \frac{t}{{w\lambda }}})} } = \left\{ {\begin{array}{*{20}{c}} {2{e^{ - {t^2}/2\left\|\bm w \right\|_2^2{\lambda ^2}}},~~~~0 \le t \le \left\|\bm w \right\|_2^2\lambda /w}\\ {2{e^{ - t/2w\lambda }},~~~~~~~t > \left\|\bm w \right\|_2^2\lambda /w} \end{array}} \right..$ \item \emph{Let $\{ X_{i}\} _{i = 1}^n $ be independent zero-mean $\{\operatorname{subE}(\lambda_i,\alpha_i )\} _{i = 1}^n$ distributed}. Define $\alpha:= \mathop {\max }_{1 \le i \le n} \alpha_i>0$, $\| \bm\lambda \|_2:= (\sum_{i = 1}^n {\lambda _i^2} )^{1/2}$ and $\bar \lambda : = (\frac{1}{n}\sum_{i = 1}^n {\lambda _i^2} )^{1/2}$. Then $\sum_{i = 1}^n {{X_i}} \sim {\rm{subE}}(\| \bm \lambda \|_2,\alpha)$ and \begin{align}\label{eq:subEE} P(|\frac{1}{n}\sum\limits_{i = 1}^n {{X_i}} | \ge t) \le 2{e^{ - \frac{1}{2}(\frac{{n{t^2}}}{{{{\bar \lambda }^2}}} \wedge \frac{{nt}}{\alpha })}}= \left\{ {\begin{array}{*{20}{c}} 2{e^{ - \frac{{n{t^2}}}{{2{{\bar \lambda }^2}}}}},~0 \le t \le \frac{{{{\bar \lambda }^2}}}{\alpha }\\ 2{e^{ - \frac{{nt}}{{2\alpha }}}},~~~~t > \frac{{{{\bar \lambda }^2}}}{\alpha } \end{array}} \right.,~\forall t \ge 0. \end{align} \end{enumerate} \end{corollary} \begin{remark}\label{re:locally sub-Gaussian} The $(\frac{{n{t^2}}}{{{{\bar \lambda }^2}}} \wedge \frac{{nt}}{\alpha })$ in \eqref{eq:subEE} reveals that the smaller $\alpha$ (locally sub-Gaussian factor) leads to sharper sub-exponential concentration. The sub-exponential concentration tends to the sub-Gaussian concentration with variance proxy ${{\bar \lambda }^2}$ when ${\alpha } \to 0$, which coincides the locally sub-Gaussian definition for sub-exponential distribution in Definition \ref{def:Sub-exponential}. \end{remark} \begin{proof} {\rm(a)} By definition of sub-exponential r.vs, ${\rm{E}}{e ^{s{w_i}{X_i}}} \le {e ^{{s^2}w_i^2\lambda _i^2/2}}~\forall~| s| \le \frac{{\rm{1}}}{{|w_i|{\lambda _i}}},\;i = 1,2, \cdots ,n$, and it implies ${\rm{E}}{e ^{s{w_i}{X_i}}} \le {e ^{{s^2}w_i^2\lambda _i^2/2}},~| s| \le \frac{{\rm{1}}}{{{w}{\lambda}}}$ for all $i$. By the independence among $\{ X_{i}\} _{i = 1}^n $, \begin{center} ${\rm{E}}\exp \{ s\sum\limits_{i = 1}^n {{w_i}{X_i}} \} = \prod\limits_{i = 1}^n {\rm{E}} {e^{s{w_i}{X_i}}} \le \exp \{ {s^2}\sum\limits_{i = 1}^n {w_i^2\lambda _i^2/2} \} \le {e^{{s^2}\left\| \bm w \right\|_2^2{\lambda ^2}/2}},~| s| \le \frac{{\rm{1}}}{{{w}{\lambda}}}.$ \end{center} {\rm(b)} The proof can be found in Theorem 1.13 of \cite{Rigollet19}. \noindent{\rm(c)}. The proof is similar to {\rm(b)}, see page 29 of \cite{Wainwright19}. \end{proof} Corollary \ref{sub-exponentialConcentration}(b) is due to Petrov, and it is also called Petrov's exponential inequalities, see \cite{Lin11}. Although Corollary \ref{sub-exponentialConcentration}(b,c) are non-asymptotically valid for any number of summands. Nevertheless, it also has asymptotical merit, which implies: \emph{Strong Law of Large Numbers} (SLNN), \emph{Central Limit Theorem} (CLT), and \emph{Law of the Iterated Logarithm} (LIL) for sub-exponential sums, as discussed below. \begin{enumerate}[\rm{(}1\rm{)}] \item \textbf{SLNN}. Let $w_i=1/n$. Consider the sample mean ${{\bar X}_n}=\frac{1}{n} \sum_{i=1}^n X_i$ for IID $\{\operatorname{subE}(\lambda_i)\} _{i = 1}^n$ data $\{X_i\}_{i=1}^n$ with population mean $\mu$, and we can use Corollary~\ref{sub-exponentialConcentration}(b) to prove that ${{\bar X}_n} \overset{a.s.}{\rightarrow}\mu$. We verify the Borel-Cantelli lemma by observing that $\sum_{n=1}^{\infty}P(|{{\bar X}_n}-\mu|>\varepsilon)\leq\sum_{n=1}^{\infty}2e^{ { - \frac{n}{2}( {\frac{{{\varepsilon^2}}}{{{\lambda ^2}}} \wedge \frac{\varepsilon}{\lambda }} )} }<\infty,$ which shows the strong convergence: ${{\bar X}_n} \overset{a.s.}{\rightarrow}\mu$. Corollary~\ref{sub-exponentialConcentration}(b) also implies the rate of convergence for sample mean for all $n$ with a high probability. It is easy to see that the sample mean ${{\bar X}_n}$ has the non-asymptotic error bounds by \begin{equation}\label{eq:subEclt} | {{\bar X}_n}-\mu| \le \sqrt {\frac{{2{\lambda ^{\rm{2}}}t}}{n}} \vee \frac{{2\lambda t}}{n}=\begin{cases} \sqrt {\frac{{2{\lambda ^{\rm{2}}}t}}{n}} , & n \ge 2t~(\text{slow global rate}) \\ \frac{{2\lambda t}}{n}, & n < 2t~(\text{fast local rate}) \end{cases} \end{equation} $\forall t> 0$ with the probability at least $1-2e^{-t}$. \item \textbf{CLT}. To study the convergence rate of CLT, we standardize the sum by letting $w_i=1/\sqrt n$ and apply Corollary \ref{sub-exponentialConcentration}(b) to $$ P (|\sqrt n {{\bar X}_n}| \ge t ) \le 2\exp \left\{ { - \frac{1}{2}\left( {\frac{{{t^2}}}{{{\lambda ^2}}} \wedge \frac{t}{{\lambda /\sqrt n }}} \right)} \right\}=\begin{cases} 2 e^{-ct^2/{\lambda ^2}}, & t \le \lambda \sqrt n; \\ 2 e^{-t \sqrt{n}/{\lambda }}, & t > \lambda \sqrt{n}. \end{cases} $$ The above deviation inequality is powerful as it indicates the phase transition about the tail behavior of $\sqrt n {{\bar X}_n}$: {\textbf{Small Deviation Regime}.} In the regime $t \le \lambda \sqrt n$, we have a sub-Gaussian tail bound with variance proxy ${\lambda ^2}$ as if the sum had the {\em normal distribution} with a constant variance. Note that the domain $t \le \lambda \sqrt n$ widens as $n$ increases and then the central limit theorem becomes more powerful. \textbf{Large Deviation Regime}. In the regime $t \ge \lambda \sqrt n$, the sum has a heavier tail. The sub-exponential tail bound is affected from \emph{the extreme variable among $\{\operatorname{subE}(\lambda_i)\} _{i = 1}^n$ with parameter $\lambda /\sqrt n$}. \item \textbf{LIL}. Let $w_i=1/n$ and $t = \frac{{R\sqrt {\log \log n} }}{{\sqrt n }} \le \left\|\bm w \right\|_2^2\lambda /w = \lambda $ for some positive constant $R$. Corollary \ref{sub-exponentialConcentration}(b) claims \begin{align*} P(|{{\bar X}_n}| \ge \frac{{R\sqrt {\log \log n} }}{{\sqrt n }})& \le 2{e^{ - {t^2}/2\left\|\bm w \right\|_2^2{\lambda ^2}}} = 2\exp \{ - \frac{n}{{2{\lambda ^2}}} \cdot \frac{{{R^2}\log \log n}}{n}\} \\ & = 2\exp \{ \log {(\log n)^{ - {R^2}/2{\lambda ^2}}}\} = {2}/{{{{(\log n)}^{{R^2}/2{\lambda ^2}}}}}. \end{align*} Therefore, with probability $1 - {2}/{{{{(\log n)}^{{R^2}/2{\lambda ^2}}}}}$, $|{{\bar X}_n}| \le \frac{{R\sqrt {\log \log n} }}{{\sqrt n }}.$ Although some researchers claims that LIL is useless, we clarify that there are still some meaningful applications of LIL, see \cite{Jamieson14} and \cite{Yang19} for the statistical and machine learning applications of the LIL. \end{enumerate} \subsection{Sub-exponential norm} Recall the Corollary \ref{prop: sub-exponential properties}(\ref{p: exponential MGF finite}): The absolute value of sub-exponential r.v. $|X|$ has a bound MGF at point $K_5^{-1}$: $\phi_{|X|}(K_5^{-1}):=\mathrm{E} e^{|X|/K_5} \le 2$. Similar to the definition of sub-Gaussian norm, we define the sub-exponential norm. \begin{definition}[sub-exponential norm]\label{def: sub-exponential} The sub-exponential norm of $X$ is defined as \begin{equation}\label{eq: psione} \|X\|_{\psi_1} = \inf \left\{ t>0 :\; \mathrm{E} \exp(|X|/t) \le 2 \right\}. \end{equation} \end{definition} An alternative definition of the sub-exponential norm is $\|X\|_{\psi_1} := \sup_{p \ge 1} p^{-1} ({\rm{E}} |X|^p)^{1/p}$ as in \cite{Vershynin2010}. The sub-exponential r.v. $X$ satisfies the equivalent properties in Corollary~\ref{prop: sub-exponential properties} (Characterizations of sub-exponential). Next, we present a useful lemma below which is to determine the sub-exponential parameter in the Definition \ref{def:Sub-exponential} by its MGF if we adopt Definition \ref{def: sub-exponential} of the sub-exponential norm. \begin{proposition}[Properties of sub-exponential norm]\label{prop:Psub-E} If $\mathrm{E} \exp(|X|/\|X\|_{{\psi _1}} ) \le 2$, then \begin{enumerate}[\rm{(}a\rm{)}] \item Tail bounds $P( |X| > t ) \le 2e^{-t/\|X\|_{{\psi _1}}}~\text{for all } t \ge 0$; \item Moment bounds $\mathrm{E} |X|^k \le 2{\|X\|_{{\psi _1}}^k}k! \quad \text{for all integer}~k \ge 1;$ \item If $\mathrm{E} X=0$, the MGF bounds ${\mathrm{{E}}}e^{s X} \leq e^{({\rm{2}} \left\| X \right\|_{{\psi _1}})^{2}s^{2}}~\text { for all }|{s}|<1/(2\left\| X \right\|_{{\psi _1}})$, i.e.$X \sim \operatorname{subE}(2\left\| X \right\|_{{\psi _1}})$. \end{enumerate} \end{proposition} \begin{proof} \rm{(a)}. To verified (a), using exponential Markov's inequality, we have $ P(|X| \geq t)={P}(e^{|X / \|X\|_{{\psi _1}}|} \geq e^{t / \|X\|_{{\psi _1}}}) \leq e^{-t / \|X\|_{{\psi _1}}|} \mathrm{E} e^{|X / \|X\|_{{\psi _1}}|} \leq 2 e^{-t / \|X\|_{{\psi _1}}} $ by Definition \ref{def: sub-exponential}. \noindent\rm{(b)}. Similar to the proof of Theorem~\ref{lem-moment} (b), we get from (a) \begin{align*} \mathrm{E}|X{|^k} &= \int_0^\infty P (|X| \ge t)k{t^{k - 1}}dt \le 2k\int_0^\infty {{e^{ - t/\|X\|_{{\psi _1}}}}} {t^{k - 1}}dt\nonumber\\ [\text{let}~s = {t/\|X\|_{{\psi _1}}}]~~&={{2k}}\int_0^\infty {{e^{ - s}}} {({s}\|X\|_{{\psi _1}})^{k - 1}}\|X\|_{{\psi _1}}ds= 2{\|X\|_{{\psi _1}}^k}{k}\Gamma( {{k-1}})= 2{\|X\|_{{\psi _1}}^k}k!. \end{align*} \rm{(c)}. Applying Taylor's expansion to MGF, we have \begin{align*} \mathrm{E} \exp (s X)&=1+\sum_{k=2}^{\infty} \frac{s^{k} \mathrm{E} X^{k}}{k !} \stackrel{\rm (b)}{\le} 1 + {\rm{2}}\sum\limits_{k = 2}^\infty {{{(s\left\| X \right\|_{{\psi _1}})}^k}}=1+\frac{2(s\left\| X \right\|_{{\psi _1}})^{2}}{1-{s\left\| X \right\|_{{\psi _1}}}},~(|{s\left\| X \right\|_{{\psi _1}}}|<1)\\ [\text{if}~|{s}|<1/(2\left\| X \right\|_{{\psi _1}})]~&\le 1+{4(s\left\| X \right\|_{{\psi _1}})^{2}} \leq e^{({\rm{2}} \left\| X \right\|_{{\psi _1}})^{2}s^{2}},~. \end{align*} Therefore, $X \sim \operatorname{subE}(2\left\| X \right\|_{{\psi _1}})$. \end{proof} Proposition \ref{prop:Psub-E}(c) implies the following user-friendly concentration inequality which contains all known constant. One should note that Theorem 2.8.1 of \cite{Vershynin18} includes an un-specific constant, which makes it is inefficacious when constructing non-asymptotic confident intervals for sub-exponential sample mean. \begin{proposition}[Concentration for r.v. with sub-exponential sum]\label{Sub-expConcentration} Let $\{ X_{i}\} _{i = 1}^n $ be zero mean independent sub-exponential distributed with $\|X_i\|_{\psi_1}\le \infty$. Then for every $t \ge 0$, \begin{center} $P( {| {\sum\limits_{i = 1}^n {{X_i}} }| \ge t} ) \le 2 \exp\{ { - \frac{1}{4}( {\frac{{{t^2}}}{\sum\nolimits_{i = 1}^n {2\|X_i\|_{\psi_1}^2}} } \wedge \frac{t}{\mathop {\max }\limits_{1 \le i \le n}\|X_i\|_{\psi_1}}}) \}.$ \end{center} \end{proposition} \begin{proof} If $\mathrm{E} \exp(|X|/\|X\|_{\psi_1} ) \le 2$, then $X \sim \operatorname{subE}(2\left\| X \right\|_{{\psi _1}})$ by using Proposition \ref{prop:Psub-E}(c). The result follows by employing Corollary \ref{sub-exponentialConcentration}(b). \end{proof} \cite{Gotze2019} mentions an explicitly calculation the sub-exponential norm with example of Poisson distributions. Therefore, it is convenient to apply Proposition~\ref{Sub-expConcentration} to get the concentration of sub-exponential summation. \begin{lemma} If $\|X\|_{\psi_{1}}$ exists, then $\|X\|_{\psi_{1}}=1 / \phi_{|X|}^{-1}(2)$ for the MGF $\phi_{X}(t):=\mathrm{E} e^{tX}$. \end{lemma} \begin{proof} Note that $\|\cdot\|_{\psi_{1}}$ is the smallest $t$ such that $\mathrm{E}e^{|X| / t}=\phi_{|X|}({t^{-1}})\le 2,$ so ${t^{-1}} \le \phi_{|X|}^{-1}(2)$ and $t \ge 1 / \phi_{|X|}^{-1}(2)$. By the definition of $\|\cdot\|_{\psi_{1}}$ again, we have $\|X\|_{\psi_{1}}=1 / \phi_{|X|}^{-1}(2)$. \end{proof} \begin{example}[The sub-exponential norm of bounded r.v.]\label{eg:boundedr.v.} Consider a r.v. $|X| \le M<\infty$. Set $\mathrm{E} e^{|X|/t}\le e^{M/t}\le 2 $ and $t\ge M/\log 2$, By the definition of $\|X\|_{\psi_{1}}$, we have $\|X\|_{\psi_{1}}= M/\log 2$. \end{example} \begin{example}[The sub-exponential norm of Poisson r.v.] Poisson r.v. $X$ has the probability mass function $P({X=k}) = \frac{{\lambda ^{k}}}{{k!}}{e^{ - k}},(k = 1,2, \cdots ,n;\lambda>0).$ We denote it as $X\sim {\rm{Poisson}}(\lambda)$. The MGF of the $ {\rm{Poisson}}(\lambda)$ is ${\phi _X}(t): = {e^{\lambda ({e^t} - 1)}}$. We have $\left\|X\right\|_{\psi_{1}}=[\log (\log (2) \lambda^{-1}+1)]^{-1}$, and the triangle inequality shows $\left\|X-\mathrm{E} X \right\|_{\psi_{1}}\le \left\|X\right\|_{\psi_{1}}+\left\|\mathrm{E} X \right\|_{\psi_{1}}= \left\|X\right\|_{\psi_{1}}+\frac{\lambda}{{ {\log 2} }} \le[\log (\log (2) \lambda^{-1}+1)]^{-1}+\frac{\lambda}{{ {\log 2} }} \propto \lambda$, where we use inequality $\left\|\mathrm{E} X \right\|_{\psi_{1}} = \frac{|\mathrm{E} X |}{{ {\log 2} }}$ by Example \ref{eg:boundedr.v.}. \end{example} Corollary \ref{Sub-expConcentration} is useful in the next subsection for the concentration for quadratic forms. \subsection{Concentration for quadratic forms and norm of random vectors}\label{quadratic} All concentration results in the above sections are about the mean. The inference for the variance and covariance in high-dimensional models is an important problem, see Section 6 of \cite{Wainwright19}. It is connected with squares of r.vs. The sample variance is a quadratic form (with shift term) of the data. The data are often postulated as sub-Gaussian. For the square of a sub-Gaussian r.v., it is natural to ask what is the behavior of the tail (or the exponential moment). The answer is sub-exponential by using \eqref{p: sub-gaussian MGF square} in Corollary \ref{Sub-gaussian distribution}. A simple example that the quadratic form of Gaussian is $\chi^2$ distributed, and the $\chi^2$-distribution of 2 degrees of freedom is exponentially distributed with mean 2. Let us look the $\chi^2$-concentration below: \begin{example}[Chi-squared r.vs] If $\{ X_{i}\} _{i = 1}^n\stackrel{\text {IID}}{\sim}N(0,1)$, then we say $Y_n:=\sum_{i=1}^n X_i^2$ follows $\chi^2$-distribution with $n$-degree of freedom, denoted as $Y_n \sim \chi^2(n)$. The density function is $ f(y)={\Gamma^{-1}(\frac{n}{2})}{(\frac{1}{2})^{\frac{n}{2}}}y^{\frac{n}{2}-1}e^{-\frac{y}{2}}\cdot 1(y>0)$. As $s<1/2$, the MGF of $X_i^2-1$ is \begin{center} $ \mathrm{E}e^{s(X_i^2-1)}=\frac{1}{\sqrt{2 \pi}} \int_{-\infty}^{+\infty} e^{s\left(x^{2}-1\right)} e^{-x^{2} / 2} d x =\frac{e^{-s}}{\sqrt{1-2 s}}\le e^{2s^{2}}= e^{(2s)^{2} / 2}, \quad \text { for all }|s|<\frac{1}{4} $ \end{center} where the second last inequality is due to $\frac{e^{-t}}{\sqrt{1-2 t}} \leq e^{2 t^{2}}$ for $|t|<1/4$. Then $X_i^2 \sim \operatorname{subE}(2,4)$. Applying Corollary \ref{sub-exponentialConcentration}(c), we have $Y_n \sim \operatorname{subE}(2\sqrt n ,4)$, therefore $P(|\frac{{{Y_n} - n}}{n}| \ge t) \le 2{e^{ - \frac{n}{8}({t^2} \wedge t)}}.$ \end{example} Similar sub-exponential results also hold for independent sum of square of sub-Gaussian r.vs. The following two lemmas in Page31 of \cite{Vershynin18} confirm this simple example to the general situation. \begin{lemma}[Square and product of sub-Gaussian are sub-exponential]\label{lem: sub-exponential squared} \rm{(a)}. A r.v. $X$ is sub-Gaussian if and only if $X^2$ is sub-exponential. Moreover, $\|X^2\|_{\psi_1} = \|X\|_{\psi_2}^2$; \rm{(b)}. Let $X$ and $Y$ be sub-Gaussian r.vs. Then $XY$ is sub-exponential and $\|XY\|_{\psi_1} \le \|X\|_{\psi_2} \, \|Y\|_{\psi_2}. $ \end{lemma} For Lemma \ref{lem: sub-exponential squared}(a), it follows from $\|X^2\|_{\psi_1} = \|X\|_{\psi_2}^2$ and Lemma \ref{prop:Psub-E} that Corollary \ref{Sub-expConcentration} coincides Proposition \ref{Sub-gaussianConcentration} as ${ {\max }_{1 \le i \le n}\|X_i\|_{\psi_1}}\to 0$, i.e. the sub-exponential r.v. degenerates to the sub-Gaussian r.v. the next proposition gives the accurately sub-exponential parameter for the square of sub-Gaussian r.v. in Definition \ref{def:Sub-exponential}, and it improves the constant in Lemma 1.12 of \cite{Rigollet19} (from $\operatorname{subE}\left(16 \sigma^{2}\right)$ to $\operatorname{subE}({{\rm{8}}\sqrt {\rm{2}} } \sigma^{2})$). \begin{proposition}\label{lem:quadraticsub-Gaussian} Let $X \sim \operatorname{subG}(\sigma^{2})$, then $Z:=X^{2}-\mathrm{E}X^{2} \sim \operatorname{subE}({{\rm{8}}\sqrt {\rm{2}} } \sigma^{2})$ or $\sim \operatorname{subE}({{\rm{8}}\sqrt {\rm{2}} } \sigma^{2}, 8\sigma^{2})$. \end{proposition} \begin{proof} The proof is immediately from Proposition \ref{pro-moment}(c) by letting $\bm w := (1, 0,\cdots ,0)^T $. \end{proof} In below, we deal with a sharper Hanson-Wright inequality in \cite{Bellec2019}. The Hanson-Wright (HW) inequality is a general concentration result for quadratic forms of sub-Gaussian r.vs, which was first studied in \cite{Hanson71}. Let $\mathbf{A}=(a_{ij})\in\mathbb{R}^{n\times n}$ be a real matrix and the ${\boldsymbol{\xi}}=(\xi_1,...,\xi_n)^T$ be a centered random vector with independent components. Define the \emph{Frobenius norm} (Hilbert-Schmidt norm) $\|\mathbf{A}\|_{\mathrm{F}}:=\sqrt{\operatorname{tr}\left(\mathbf{A}^{T} \mathbf{A} \right)}=\sqrt{\sum_{i, j} A_{i, j}^{2}}$ and the \emph{spectral norm} (operator norm) $\|\mathbf{A} \|_{2}:=\sup _{\|\bm u\|_{2} \leq 1}\|\mathbf{A}\bm u\|_{2}$. As an extension of $\chi^2$ r.vs, it is of interest to study the concentration behavior of ${\boldsymbol{\xi}}^T \mathbf{A} {\boldsymbol{\xi}} - \mathrm{E}[ {\boldsymbol{\xi}}^T \mathbf{A} {\boldsymbol{\xi}}]$. Under the setting above, Example 2.12 in \cite{Boucheron13} gives the Gaussian chaos concentration. \begin{corollary}[Gaussian chaos of order 2] \label{prop:gaussian-chaos} Let $\xi_1,...,\xi_n$ be zero-mean Gaussian with $\mathrm{E} \xi_i^2=\sigma_i^2$, Define $\mathbf{D}_\sigma = \mbox{diag}(\sigma_1,...,\sigma_n)$, then for any $x>0$ \begin{equation} P({\boldsymbol{\xi}}^T \mathbf{A} {\boldsymbol{\xi}} - \mathrm{E}[{\boldsymbol{\xi}}^T \mathbf{A} {\boldsymbol{\xi}} ] \ge 2 \hsnorm{\mathbf{D}_\sigma \mathbf{A} \mathbf{D}_\sigma} \sqrt{x} + 2 \opnorm{\mathbf{D}_\sigma \mathbf{A} \mathbf{D}_\sigma}x ) \le e^{-x}. \label{eq:gaussian-chaos} \end{equation} \end{corollary} The similar concentration phenomenon is also available for sub-Gaussian r.vs. which is named as the HW inequality. \cite{Rudelson13} gives a modern proof by the so-called \emph{decoupling argument} attributed to \cite{Bourgain1996}. \begin{corollary}[R-V's HW inequality] \label{prop:hanson} Let $n\ge 1$ and ${\boldsymbol{\xi}}:=(\xi_1,...,\xi_n)^T$ be an independent zero-mean sub-Gaussian r.vs with $\max_{i=1,...,n} \psinorm{\xi_i} \le K$ for $K>0$. Let $A$ be any $n\times n$ real matrix. Then there exists a constant $c>0$ such that \begin{equation} P( {\boldsymbol{\xi}}^T \mathbf{A} {\boldsymbol{\xi}} - \mathrm{E} [ {\boldsymbol{\xi}}^T \mathbf{A} {\boldsymbol{\xi}} ] > t) \le e^{ - c ( \frac{t^2}{K^4 \hsnorm{\mathbf{A}}^2}\wedge \frac{t}{K^2 \opnorm{\mathbf{A}}} )},~t \ge 0 \label{eq:hanson-t} \end{equation} Furthermore, for any $x > 0$, $P({\boldsymbol{\xi}}^T \mathbf{A} {\boldsymbol{\xi}} - \mathrm{E}[{\boldsymbol{\xi}}^T \mathbf{A} {\boldsymbol{\xi}} ] \le c K^2( \opnorm{\mathbf{A}}x + \hsnorm{\mathbf{A}} \sqrt{x}))\ge 1-e^{-x}$. \end{corollary} Intuitively, the term $K^2 \hsnorm{\mathbf{A}}$ is seen as the ``variance term''. When $\mathbf{A}$ is diagonal-free (i.e. the $\mathbf{A}$ matrix has zeros down its diagonal: $a_{ii}=0$ ), the r.v. ${\boldsymbol{\xi}}^T \mathbf{A} {\boldsymbol{\xi}}$ is zero-mean. \cite{Pollard2015} shortens the proof without unknown constant. \begin{corollary}[Diagonal-free Hanson-Wright inequality] Let $\xi_{1}, \ldots, \xi_{n}$ be independent, centered sub-Gaussian r.vs with $\max_{i=1,...,n} \psinorm{\xi_i} \le K<\infty$. Let $\mathbf{A}$ be an $n \times n$ matrix of real numbers with $a_{i i}=0$ for each $i$. Then ${P}({\boldsymbol{\xi}}^T \mathbf{A} {\boldsymbol{\xi}} \geq t)\leq e^{- (\frac{t^{2}}{64 K^{4}\|\mathbf{A}\|_{\mathrm{F}}}\wedge\frac{t}{8 \sqrt{2} K^{2}\|\mathbf{A}\|_{2}})}~ \text { for } t \geq 0.$ \end{corollary} Under assumptions on the moments of $\xi_1,...,\xi_n$ (do not need sub-Gaussian assumption), the next corollary provides a concentration inequality for quadratic forms of independent r.vs satisfying Bernstein's moment condition (discussed in the next subsection). \begin{corollary}[Quadratic forms concentration with moment conditions]\label{thm:hanson-moment} Assume that the r.v. ${\boldsymbol{\xi}}=(\xi_1,...,\xi_n)^T$ satisfies the condition on independent variables $\xi_1^2,...,\xi_n^2$: $\mathrm{E} |\xi_i|^{2p} \le \tfrac{1}{2} \; p! \; \sigma_i^2 \; \kappa^{2p-2}~\forall p \ge 1$ for some $\kappa>0$. Let $A$ be any $n\times n$ real matrix. Then for all $t\ge 0$, \begin{equation}\label{eq:hanson-moment-t} P({\boldsymbol{\xi}}^T \mathbf{A} {\boldsymbol{\xi}} - \mathrm{E}[{\boldsymbol{\xi}}^T \mathbf{A} {\boldsymbol{\xi}}] > t) \le e^{-(\frac{t^2}{192 \kappa^2 \hsnorm{\mathbf{A} \mathbf{D}_\sigma}^2}\wedge \frac{t}{ 256 \kappa^2 \opnorm{\mathbf{A}}})}, \end{equation} where $\mathbf{D}_\sigma := \mbox{diag}(\sigma_1,...,\sigma_n)$. Furthermore, with probability greater than $1-e^{-x}$, \begin{equation} {\boldsymbol{\xi}}^T \mathbf{A} {\boldsymbol{\xi}} - \mathrm{E}[{\boldsymbol{\xi}}^T \mathbf{A} {\boldsymbol{\xi}} ] \le 256 \kappa^2 \opnorm{\mathbf{A}} x + {8\sqrt{3}} \kappa \hsnorm{\mathbf{A} \mathbf{D}_\sigma} \sqrt{x},~\forall x \ge 0 . \label{eq:hanson-moment-x} \end{equation} \end{corollary} The bound in (\ref{eq:hanson-moment-t}) is exactly $ \exp( - \frac{t^2}{192 \kappa^2 \hsnorm{ \mathbf{A} \mathbf{D}_\sigma}^2} )$ if $t$ is small, and while the $\exp(- c \frac{t^2}{K^4 \hsnorm{\mathbf{A}}^2})$ in right hand side of the R-V's HW inequality \eqref{eq:hanson-t} has an unspecific constant $c>0$. {\color{black}{We finish this subsection with an exponential inequality for quadratic forms of a sub-Gaussian random vector. Consider the $n$-dimensional unit sphere ${S^{n - 1}} := \{ \boldsymbol{x} \in {\mathbb{R}^n}:{\left\| \boldsymbol{x} \right\|_2} = 1\}.$ Early in \cite{Fukuda1990}, a random vector $\boldsymbol{X}$ in $\mathbb{R}^n$ is called sub-Gaussian (sub-exponential) if the one-dimensional marginals $\langle {\boldsymbol{X},\boldsymbol{x}} \rangle $ are sub-Gaussian (sub-exponential) r.vs for all $\boldsymbol{x} \in \mathbb{R}^n$. Naturally, the sub-Gaussian (sub-exponential) norm of $\boldsymbol{X}$ is defined as $\|\boldsymbol{X}\|_{\psi _2} := \sup_{\boldsymbol{x} \in S^{n-1}} \|\langle {\boldsymbol{X},\boldsymbol{x}} \rangle \|_{\psi _2}$ ( $\|\boldsymbol{X}\|_{\psi _1} := \sup_{\boldsymbol{x} \in S^{n-1}} \|\langle {\boldsymbol{X},\boldsymbol{x}} \rangle \|_{\psi _1}$). For the sub-Gaussian, \cite{Fukuda1990}'s definition is equivalent to Chapter 6.3 of \cite{Wainwright19}, a random vector $\boldsymbol{X} \in \mathbb{R}^{d}$ with parameter $\sigma \in \mathbb{R}$ is sub-Gaussian (denote $\mathrm{subGV}(\sigma^2)$) so that: \begin{equation}\label{eq:SGV} \mathrm{E} e^{\lambda(\bm{v}, \boldsymbol{X}-\mathrm{E} \boldsymbol{X})} \leq e^{\lambda\sigma^{2}/{2}}, ~ \forall ~ \lambda \in \mathbb{R}^{n}~\text{and}~\bm{v} \in {S^{n - 1}}~\Leftrightarrow~\mathrm{E}e^{\boldsymbol{\alpha}^{T}(\boldsymbol{X}-\mathrm{E} \boldsymbol{X})} \le e^{\|\boldsymbol{\alpha}\|^{2} \sigma^{2} / 2},~\forall~\boldsymbol{\alpha} \in \mathbb{R}^{n} \end{equation} In \cite{Pan2020}, the subG random vector with parameter $v_{0} \ge 1$ is defined by ${P}\left(|\langle\boldsymbol{u}, \boldsymbol{X}\rangle| \ge v_{0}\|\boldsymbol{u}\|_{\boldsymbol{\Sigma}} \cdot t\right) \le 2 e^{-t^{2} / 2}$ for all $\boldsymbol{u} \in \mathbb{R}^{n}$ and $t \ge 0,$ where $\boldsymbol{\Sigma}=\mathrm{E}\left(\boldsymbol{X} \boldsymbol{X}^T\right)$ and $\|\boldsymbol{u}\|_{\mathbf{A}}=\left\|\mathbf{A}^{1 / 2} \boldsymbol{u}\right\|_{2}$ is the norm indexed by ${\mathbf{A}}$. For Definition \eqref{eq:SGV}, \cite{Hsu2012} obtains a tail bound for subG random vectors: \begin{corollary}[Tail inequality for quadratic forms of sub-Gaussian vectors]\label{thm:zhangt} Let $\bm{\Sigma}=\mathbf{A}^{T} \mathbf{A}$ for $p \times n$ matrix $\mathbf{A}$. Consider a sub-Gaussian random vector ${\boldsymbol{\xi}}=(\xi_1,...,\xi_n)^T\sim \mathrm{subGV}(\sigma^2)$ with independent components for $\boldsymbol{\mu}=\mathrm{E}{\boldsymbol{\xi}}$. Then, for any $t\ge 0$ \begin{center} ${P}\{\|\mathbf{A} {\boldsymbol{\xi}}\|^{2}>\sigma^{2}[\operatorname{tr}(\bm{\Sigma})+2 \operatorname{tr}(\bm{\Sigma}^{2} t)^{1 / 2}+2\|\bm{\Sigma}\|_2 t]+\operatorname{tr}(\bm{\Sigma} \boldsymbol{\mu} \boldsymbol{\mu}^{T})(1+2 \sqrt{\textstyle{{\|\bm{\Sigma}\|_2^{2}}\over{\operatorname{tr}(\bm{\Sigma}^{2})}} t})\} \leq e^{-t}.$ \end{center} \end{corollary} Conditioning on a divergence number of non-random covariates, an application of Corollary \ref{thm:zhangt} for the prediction error \eqref{olsR} in regressions with sub-Gaussian noise is given in Section \ref{se;lm}. The concentration bounds of sub-Gaussian random vectors depend on the parameter $\sigma$: the smaller $\sigma$, the tighter concentration bounds. Equation \eqref{eq:SGV} requires the distribution of subG random vectors to be isotropic, and the random vectors have an exponential tail, but the sub-Gaussian parameter $\sigma$ may be large, which leads to loose bounds for constructing confidence bands. To establish tighter bounds, \cite{Jin19} define a different and general class of sub-Gaussian distributions in $\mathbb{R}^{n}$, called norm-subGaussian random vectors as follows. \begin{definition}[Norm-subGaussian]\label{def:norm-subGaussian} A random vector $\boldsymbol{X} \in \mathbb{R}^{n}$ is norm-subG (denoted $\mathrm{nsubG}(\sigma^2)$), if $\exists \sigma$ so that: $ {P}(\|\boldsymbol{X}-\mathrm{E} \boldsymbol{X}\| \geq t) \leq 2 e^{-\frac{t^{2}}{2 \sigma^{2}}}, \quad \forall t \in \mathbb{R}^{+}. $ \end{definition} The Definition \ref{def:norm-subGaussian} only requires the tail probability estimate has sub-Gaussian tail under $l_2$-norm, avoiding the uniform condition in \eqref{eq:SGV}. If $\mathrm{E} \boldsymbol{X}=\bm 0$ and $2 e^{-t^{2} / 2\sigma^2} \ge {P}\left(\|\boldsymbol{X}\| \ge t\right)$ from $\mathrm{nsubG}(\sigma^2)$, we get ${P}\left(|\langle\boldsymbol{u}, \boldsymbol{X}\rangle| \ge t\right) \le {P}\left(\|\boldsymbol{X}\| \ge t\right) \le 2 e^{-t^{2} / 2\sigma^2},~\bm{u} \in {S^{n - 1}}$ by Cauchy's inequality. Thus, $\mathrm{nsubG}(\sigma^2)$ implies \eqref{eq:SGV}, and this verifies that the norm-subG is more general. \cite{Jin19} show that if $\boldsymbol{X} \in \mathbb{R}^{n}$ is $\mathrm{subGV}(\sigma^2/{n})$, then $\boldsymbol{X} \sim \mathrm{nsubG}(8\sigma^2)$. }} \section{Sub-Gamma Distributions and Bernstein's Inequality} \subsection{Sub-Gamma distributions} Comparing to the classical Chebyshev's inequality, Bernstein-type inequalities have more precise concentration, it originally is an extension of the Hoeffding's inequality with bounded assumption [see \cite{Bernstein1924}, \cite{Bennett1962}]. As mentioned by \cite{Pollard2015}, the proof of Hoeffding's inequality with endpoints of the interval $[a,b]$ in Lemma~\ref{lm:Hoeffding} (with $n=1$) crudely depends on the variance bound: \begin{equation}\label{eq:VarX} {\mathop{\rm Var}\nolimits} X = \mathrm{E}{(X - \mathrm{E} X)^2} \le \mathrm{E}[X - ({{b - a}})/{2}]^2 \le [({{b - a}})/{2}]^2~\text { if } a \leq X \leq b. \end{equation} If $X$ takes values near the endpoints of the interval $[a,b]$ with a small probability, it is guessed that the sharper concentration could be improved by adding variance condition. The following tail bound for the sum $S_n:=\sum_{i = 1}^n {{X_i}}$ needs extra variance information. \begin{corollary}[Bernstein's inequality with the bounded condition]\label{lm-Bernsteinbd} Let $X_{1},\ldots ,X_{n}$ be centralized independent variables such that $|X_{i}|\leq M$ a.s. for all $i$. Then, $\forall~t>0$ \begin{center} ${P} (|S_n|\geq t)\leq 2 e^{-{\frac {t^{2}/2}{\sum _{i=1}^{n}\mathrm{Var}X_{i}+Mt/3}}},~P\{|S_n|\geq(2t \sum_{i=1}^{n}\mathrm{Var}X_{i})^{1/2}+\frac{{Mt}}{3}\} \leq 2e^{-t}.$ \end{center} \end{corollary} The next example illustrates a sharp confidence interval for sample mean if we known that the variance is sufficient small. \begin{example}[Non-asymptotic confidence intervals] Let $\{X_i\}_{i=1}^n \stackrel{\rm{IID}}{\sim} X$ with the support $[-c,c]$ and the mean $\mu$. Hoeffding's and Bernstein's inequalities show for ${\bar X}:={n}^{-1} \sum{}_{i=1}^{n} X_{i}$ $$ \begin{array}{c} {P}(|\bar X-\mu | \le \sqrt{\frac{2 c^{2} \log (2 / \delta)}{n}}) \geq 1-\delta \quad \text { Hoeffding};\\ {P}(|\bar X-\mu | \le \frac{c}{3 n} \log (2 / \delta)+\sqrt{\frac{2(\operatorname{Var} X) \log (2 / \delta)}{n}}) \geq 1-\delta \quad \text { Bernstein}. \end{array} $$ For large $n$, the Bernstein's confidence interval is substantially shorter if ${X_i}$ has relatively small variance, i.e. ${\rm{Var}}{X} \ll {c^2}$ (the factor $\sqrt{\frac{\log (2 / \delta)}{n}}$ is a dominated term). The Hoeffding's confidence is shorter as ${\rm{Var}}{X}= {c^2}$ (This extreme case attains the upper bound ${\rm{Var}}{X} \le {c^2}$ in \eqref{eq:VarX} due to $b-a=2c$). But, for the case ${\rm{Var}}{X}< {c^2}$, if $n$ is sufficient small s.t. $\frac{c}{{3n}}\log (2/\delta ) \ge (c - \sqrt {\rm{Var}X} )\sqrt {\frac{{2\log (2/\delta )}}{n}} $, i.e. we need restrictions $n \le \frac{1}{{18}}{\left( {\frac{c}{{c - \sqrt {{\rm{Var}}X} }}} \right)^2}\log (\frac{2}{\delta })\ge 1$ to ensure Hoeffding's confidence interval is more accurate when $\delta \le 2\exp \{ - \frac{1}{{18}}{(\frac{{c - \sqrt {{\rm{Var}}X} }}{c})^2}\}$. \end{example} To prove Corollary \ref{lm-Bernsteinbd}, we need get the sharp bounds of the MGF of the single variable and then do aggregation for the summation. By the Taylor expansion, we have \begin{center} ${\rm{E}}{e^{s{X_i}}} = 1 + \sum\limits_{k = 2}^\infty {{s^k}} \frac{{{\rm{E}}{X_i^k}}}{{k!}} \le 1 + \sum\limits_{k = 2}^\infty {{s^k}} \frac{{{M^{k - 2}}{\rm{Var}}{X_i}}}{{k!}} \le 1 + {s^2}{\rm{Var}}{X_i}\sum\limits_{k = 2}^\infty {\frac{{{{(|s|M)}^{k - 2}}}}{{k!}}},~1 \leq i \leq n.$ \end{center} Applying the inequality $k ! / 2 \geq 3^{k-2}$ for any $k \geq 2$, it implies \begin{equation}\label{eq:GammaGMFbd} {\rm{E}}{e^{s{X_i}}} \le 1 + \frac{{{s^2}{\rm{Var}}{X_i}}}{2}\sum\limits_{k = 2}^\infty {{{\left( {\frac{{|s|M}}{3}} \right)}^{k - 2}}} = 1 + \frac{{{s^2}{\rm{Var}}{X_i}/2}}{{1 - |s|M/3}} \le \exp \left( {\frac{{{s^2}{\rm{Var}}{X_i}/2}}{{1 - |s|M/3}}} \right). \end{equation} The upper bounds of MGF essentially have the same form in comparison with Gamma distribution below whose MGF is bounded by \eqref{eq:GammaGMF} in following example. \begin{example}[Gamma r.vs]\label{ex:Gamma} The Gamma distribution with density $f(x)=\frac{x^{a-1} e^{-x / b}}{\Gamma(a) b^{a}},~x \geq 0$ is denote as $\Gamma(a,b)$. We have $\mathrm{{E}} X=a b$ and $\operatorname{Var}X=a b^{2}$ for $X\sim\Gamma(a,b)$. The Page28 of \cite{Boucheron13} illustrates that the log-MGF of a centered $\Gamma(a,b)$ is bounded by \begin{equation}\label{eq:GammaGMF} \log (\mathrm{{E}}{e^{s({X}-\mathrm{{E}} X)}})=a(-\log (1-sb)-sb) \leq s^{2}a b^{2}/[2(1- bs)], \quad \forall~0<s<b^{-1}. \end{equation} \end{example} Motivated by the MGF bounds in \eqref{eq:GammaGMF}, \cite{Boucheron13} defines the sub-Gamma r.v. based on the right tail and left tail with variance factor $v$ and scale factor $b$. \begin{definition}[Sub-Gamma r.v.]\label{def: sub-Gamma} A centralized r.v. $X$ is {\em sub-Gamma} with the \emph{variance factor} $\upsilon>0$ and the \emph{scale parameter} $c>0$ (denoted by $X \sim \mathrm{sub}\Gamma(\upsilon,c)$) if \begin{equation}\label{eq:sub-gamma} \log (\mathrm{{E}}{e^{s{X}}}) \leq {s^{2}}{\upsilon}/[2({1-c|s|})], \quad \forall~0<|s|<c^{-1}. \end{equation} \end{definition} If the restriction $0<|s|<b^{-1}$ is replaced by one side conditions $0<s<b^{-1}$ (or $0<-s<b^{-1}$), we call it \emph{sub-Gamma on the right tail} (or \emph{sub-Gamma on the left tail}), denoted as $\mathrm{sub}\Gamma_{+}(\upsilon,c)$ (or $\mathrm{sub}\Gamma_{-}(\upsilon,c)$). In Example \ref{ex:Gamma}, the Gamma r.v. $X \sim \mathrm{sub}\Gamma_{+}(a b^{2},b)$. The \eqref{eq:sub-gamma} is called \emph{two-sided Bernstein's condition}. \begin{example}[Sub-exponential r.vs] The sub-exponential distribution with positive support implies the sub-Gamma condition: $\log (\mathrm{{E}}{e^{s{X}}}) \le {\frac{s^{2}\lambda ^{2}}{2}} \le \frac{s^{2}\lambda ^{2}}{ {2(1 - \lambda|s|)}},~\forall~|s| <\frac{1}{\lambda}.$ This shows that $X \sim \operatorname{subE}(\lambda)$ implies $X \sim \mathrm{sub}\Gamma(\lambda^{2},{\lambda})$. \end{example} The sub-Gamma condition \eqref{eq:sub-gamma} leads to the useful tail bounds and moment bounds. \begin{lemma}[Sub-gamma properties, \cite{Boucheron13}]\label{sub-gammaproperties} If $X \sim \mathrm{sub}\Gamma(\upsilon,c)$, then \begin{equation}\label{eq:sub-gammasingle} P(|X|>t)\leq 2 e^{-\frac{\upsilon}{c^{2}} h\left(\frac{c t}{\upsilon}\right)}\leq 2e^{-\frac{t^2/ 2}{v+ct}}, \end{equation} where $h(u)=1+u-\sqrt{1+2|u|}$. Moreover, we have $ P\{|X|>\sqrt{2 v t}+c t\} \leq e^{-t}. $ \end{lemma} The tail bound in Lemma \ref{sub-gammaproperties} verifies that, the sub-Gamma variable has sub-Gaussian tail behavior with parameter $\upsilon$ for suitably small $t$, and it has exponential tail behavior for larger $t$. The proof is originated from \cite{Bennett1962}. \begin{proof} By Chernoff's inequality, $ P\left( {X - {\rm{E}}X \ge t} \right) \le \mathop {\inf }\limits_{s > 0} {e^{ - st}}{\rm{E}}{e^{s(X - {\rm{E}}X)}}$. It remains to bound $\log({e^{ - st}}{\rm{E}}{e^{s(X - {\rm{E}}X)}})$ by definition of sub-Gamma variable for all $0<|s|<c^{-1}$ \begin{center} $\mathop {\inf }\limits_{{c^{ - 1}} \ge s > 0} \log ({e^{ - st}}{\rm{E}}{e^{s(X - {\rm{E}}X)}}) \le \mathop {\inf }\limits_{{c^{ - 1}} \ge u > 0} \left( {\frac{{{u^2}}}{2}\frac{v}{{1 - cu}} - ut} \right) = - \frac{v}{{{c^2}}}h(\frac{{ct}}{v}) \le - \frac{{{t^2}/2}}{{v + ct}},$ \end{center} where the last inequality is from $h(u) = 1 + u - \sqrt {1 + 2|u|} \ge \frac{{{u^2}/2}}{{1 + u}}$. So we conclude \eqref{eq:sub-gammasingle}. \end{proof} \begin{proposition}[Concentration for sub-Gamma sum]\label{sub-GammaConcentration} Let $\{ X_{i}\} _{i = 1}^n $ be independent\\ $\{\mathrm{sub}\Gamma(\upsilon_i,c_i)\} _{i = 1}^n$ distributed with zero mean. Define $c= {\max }_{1 \le i \le n} c_i$, we have \begin{enumerate}[\rm{(}a\rm{)}] \item Closed under addition: $ S_n:=\sum_{i = 1}^n {{X_i}} \sim {\rm{sub}}\Gamma({\sum_{i = 1}^n\upsilon_i},c)$; \item For every $t \ge 0$: ${P} (|S_n|\geq t)\leq 2 \exp \left(-{\frac {t^{2}/2}{\sum _{i=1}^{n}\upsilon_i+ct}}\right)$ and $P\{|S_n|>(2t \sum_{i=1}^{n}\upsilon_i)^{1/2}+c t\} \leq 2 e^{-t}$; \item If $X \sim \mathrm{sub}\Gamma(\upsilon,c)$, the moments bounds satisfy for any integer $k\ge 1$: \begin{center} $\mathrm{E}{X^{k}} \le k{2^{k - 2}}[2{(\sqrt {2v} )^k}\Gamma (\frac{k}{2}) + c{(\sqrt {2v} )^{k - 1}}\Gamma (\frac{{k + 1}}{2}) + 3{c^k}\Gamma ({k})].$ \end{center} \item If $X \sim \mathrm{sub}\Gamma(\upsilon,c)$, the even moments bounds satisfy $\mathrm{E}{X^{2k}} \le k!{({8v} )^k}+(2k)!{({4c} )^{2k}},~k\ge 1.$ \item If $P\{|X|>(2t \upsilon)^{1/2}+c t\} \leq 2 e^{-t}$, then $X \sim \mathrm{sub}\Gamma(32(\upsilon+2c^2),8c)$. \end{enumerate} \end{proposition} \begin{proof} {\rm(a)} By definition of $\{\mathrm{sub}\Gamma(\upsilon_i,c_i)\} _{i = 1}^n$, we have $\log (\mathrm{{E}}{e^{s{X_i}}}) \leq \frac{s^{2}}{2} \frac{\upsilon_i}{1-c_i|s|},~\forall~0<|s|<c^{-1}$, from which and the independence among $\{ X_{i}\} _{i = 1}^n $, thus \begin{center} $\log (\mathrm{{E}}{e^{sS_n}}) \leq \frac{s^{2}}{2}\sum_{i = 1}^n \frac{\upsilon_i}{1-c_i|s|}\leq \frac{s^{2}}{2} \frac{\sum_{i = 1}^n\upsilon_i}{1-c|s|},\quad {\rm{for~all~}}0<|s|<c^{-1}.$ \end{center} \noindent{\rm(b)} Employing Proposition \ref{sub-gammaproperties}, we immediately obtain {\rm(b)} due to {\rm(a)}. \noindent{\rm(c)} Applying the integration form of the expectation formula, it yields \begin{align*} {\rm{E}}{X^k}&\le {\rm{E}}{|X|^k} = k\int_0^\infty {{x^{k - 1}}} P\{ |X| > x\} dx= k\int_0^\infty {{x^{k - 1}}} P\{ |X| > \sqrt {2vt} + ct\} (\frac{{\sqrt {2v} }}{{2\sqrt t }} + c)dt\\ & \le 2k\int_0^\infty {{{(\sqrt {2vt} + ct)}^{k - 1}}} (\frac{{\sqrt {2vt} + 2ct}}{{2t}}){e^{ - t}}dt= k\int_0^\infty {[{{(\sqrt {2vt} + ct)}^k} + ct{{(\sqrt {2vt} + ct)}^{k - 1}}} ]\frac{{{e^{ - t}}}}{t}dt. \end{align*} From {\rm(b)} and inequality \eqref{eq:Jensen}, \begin{align*} {\rm{E}}{X^k}& \le k\int_0^\infty {\left\{ {{2^{k - 1}}[{{(\sqrt {2vt} )}^k} + {{(ct)}^k}] + ct{2^{k - 2}}[{{(\sqrt {2vt} )}^{k - 1}} + {{(ct)}^{k - 1}}]} \right\}} \frac{{{e^{ - t}}}}{t}dt\\ & = k{2^{k - 2}}\int_0^\infty {[2{{(\sqrt {2v} )}^k}{t^{(k/2) - 1}} + c{{(\sqrt {2v} )}^{k - 1}}{t^{(k + 1)/2-1}} + 3{c^k}{t^{k - 1}}]} {e^{ - t}}dt\\ & = k{2^{k - 2}}[2{(\sqrt {2v} )^k}\Gamma (\frac{k}{2}) + c{(\sqrt {2v} )^{k - 1}}\Gamma (\frac{{k - 1}}{2}) + 3{c^k}(k - 1)!]. \end{align*} \noindent{\rm(d,e)} The proofs are in Theorem 2.3 of \cite{Boucheron13}. \end{proof} Having obtained Proposition \ref{sub-gammaproperties}(b), from the upper bound in \eqref{eq:GammaGMFbd}, we finish the proof of Proposition \ref{lm-Bernsteinbd} by treating $X_i \sim \mathrm{sub}\Gamma({\rm{Var}}{X_i}/2,M/3)$ for $i=1,2,\cdots,n$. \subsection{Bernstein's growth of moments condition} {In some settings, one can not assume the r.vs being bounded. } Bernstein's inequality for the sum of independent r.vs allows us to estimate the tail probability by a weaker version of an exponential condition on the growth of the $k$-moment without the boundedness. \begin{corollary}[Bernstein's inequality with the growth of moment condition]\label{lm-Bernsteingm} If the centred independent r.vs $X_1,\ldots, X_n$ satisfy \emph{the growth of moments condition} \begin{equation}\label{eq-Bernsteinmoments} {\rm{E}}{\left| {{X_i}} \right|^k} \le {2}^{-1}v_i^2{\kappa_i^{k - 2}}k!,~(i = 1,2, \cdots ,n),~\text{for all}~k\ge 2 \end{equation} where $\{\kappa_i\}_{i=1}^n, \{v_i\}_{i=1}^n$ are constants independent of $k$. Let ${\nu _n^2} = \sum_{i = 1}^n {v _i^2} $ (the fluctuation of sums) and $\kappa={\max }_{1 \le i \le n} \kappa_i$. Then, we have $X_{i}\sim \mathrm{sub}\Gamma(v_i,\kappa_i)$ and for $t>0$ \begin{equation}\label{eq-Bernstein} P\left( {\left| {{S_n}} \right| \ge t} \right) \le 2e^{ - \frac{{{t^2}}}{{2{\nu _n^2} + 2\kappa t}}},~~P( {\left| {{S_n}} \right| \ge \sqrt {2\nu _n^2t} + \kappa t}) \le 2{e^{ - t}}. \end{equation} \end{corollary} \begin{proof} Given that $\kappa_i|s|<1$ for all $i$, \eqref{eq-Bernsteinmoments} implies that $X_{i}\sim \mathrm{sub}\Gamma(v_i,\kappa_i)$ for $1 \leq i \leq n$ \begin{center} $ {\rm{E}} e^{s X_{i}} \leq 1+\frac{v_{i}^{2}}{2} \sum_{k=2}^{\infty}|s|^{k} \kappa_i^{k-2}=1 +\frac{s^{2} v_{i}^{2}}{2(1-|s| \kappa_i)} \leq e^{s^{2} v_{i}^{2} /(2-2 \kappa_i|s|)}. $ \end{center} The independence among $\{X_i\}_{i=1}^n$ and Proposition \ref{sub-GammaConcentration}(a,b) implies \eqref{eq-Bernstein}. \end{proof} The \eqref{eq-Bernsteinmoments} is also called \emph{Bernstein's moment condition}. Corollary~\ref{lm-Bernsteingm} slightly extends Lemma 2.2.11 in \cite{van96} for the case ${\kappa _i} \equiv \kappa$ (a fixed number). It should be noted that \eqref{eq-Bernsteinmoments} can be replaced by $\frac{1}{n} \sum_{i=1}^{n} {\rm{E}}{\left| {{X_i}} \right|^k} \le \frac{1}{2}v^2{\kappa^{k - 2}}k!, k=3,4, \ldots, \forall i$, where the $v^2$ is a variance-depending constant such that $\frac{1}{n} \sum_{i=1}^{n} {\rm{E}}{\left| {{X_i}} \right|^2} \le v^2$. Then \eqref{eq-Bernstein} still holds with $\nu _n^2=nv$, see Theorem 2.10 in \cite{Boucheron13}. \begin{example}[Normal r.v.]\label{ex:bounded} Applying the relation between MGF and moment, the $k$-th moment of $X \sim N(\mu,\sigma^{2})$ is ${\rm{E}}{ X ^{2k - 1}} = 0;~{\rm{E}}{\left| X \right|^{2k}} = {\sigma ^{2k}}(2k - 1)(2k - 3) \cdots 3 \cdot 1 \le {2}^{-1}(2{\sigma ^2}){\sigma ^{2k - 2}}(2k)!$, which satisfies \eqref{eq-Bernsteinmoments} with $v^2 = 2{\sigma ^2},\kappa = {\sigma ^2}$. \end{example} \subsection{Concentration of exponential family without compact space} Theory and statistical applications of natural exponential family \eqref{eq:E-P} have attracted renewed attention in the past years \citep{Lehmann06}. in Lasso penalized \emph{generalized linear models}(GLMs), the results of oracle inequalities lie on CIs of a quantity that can be represent as Karush-Kuhn-Tucker conditions (see \eqref{eq:kkt}) related to the centralized exponential family empirical process: $\sum_{i = 1}^n {{w_i}({Y_i}} - {\rm{E}}{Y_i}) $ for no-random weights $\{w_i\}_{i=1}^n$ depending on the fixed design. \cite{Kakade10} has studied the sub-exponential growth of the cumulants of an exponential family distribution and studied oracle inequalities of Lasso regularized GLMs, but the constant in their result is not specific. In this part, we obtain cental moments bounds with a specific constant, which gives the Bernstein's inequality for the general exponential family, and the proof is based on the Cauchy formula of higher-order derivatives for complex functions [Corollary 4.3 in \cite{Shakarchi10}]. \begin{lemma}[Cauchy's derivative inequalities]\label{lem:Cauchy} If $f$ is analytic in an open set that contains the closure of a disk $D$ centered at $z_0$ of radius $0<r<\infty$, then $ |f^{(n)}(z_{0})| \leq \frac{n !}{r^{n}}\sup _{z : |z-z_0|=r}|f(z)|. $ \end{lemma} \noindent \cite{Zhang14} adopts a similar approach for recovering the probability mass function (p.m.f.) from the characteristic function. It is well-known that exponential families on the natural parameter space, $\Theta :=\{ \theta \in \mathbb { R } ^ { k } :{e^{b({\theta})}} := \int_{} {c(y){e^{y\theta }}\mu (dy)}< \infty \}$ have finite analytic (standardized) moments and cumulants, see Lemma 3.3 in \cite{Kakade10}. The natural parameter space of an exponential family is convex, see \cite{Lehmann06}. A nice property in Lehmann's measure-theoretical statistical inference book is that: \begin{lemma}[Analytic property of MGF in the exponential family]\label{le:Analytic} The MGF $m_{\theta_i}(s):={{\rm{E}}_{\theta_i} }{e^{s{X_i} }}$ on $s\in \mathbb{C}$ of exponential family r.vs indexed by $\theta_i$, is analytic on $\Theta$ [see Theorem 2.7.1 in \cite{Lehmann06} or Theorem 2 in \cite{Pistone1999}]. \end{lemma} First, let us check the following lemma which is deduced by Cauchy's inequalities for the Taylor's series coefficients of a complex analytic function. \begin{proposition} The $s \mapsto \bar{m}_{\theta _i}(s):={{\rm{E}}_{\theta _i}}{e^{s|{Y_i} - \dot b({\theta _i}){\rm{|}}}}$ is analytic on the natural parameter space $\Theta$ with radius ${\rm{r}}(\Theta)$, and the $k$-th absolute central moment of $\{w_i{Y_i}\}_{i=1}^n$ is bounded by \begin{center} ${\rm{E}}_{\theta _i}{|{{w_i}({Y_i}} - {\rm{E}}{Y_i}) |^k}\le \frac{k!}{2}{({w}\sqrt {2C_{{\theta _i}}} )^2}({w}C_{{\theta _i}})^{k - 2},~k=2,3,\cdots$ \end{center} where $\{w_i\}_{i=1}^n$ are non-random with $w:= \mathop {\max }\limits_{1 \le i \le n} |w_i|>0$, and ${C_{\theta_i}}: = \mathop {\inf }\limits_{0<r \le {\rm{r}}(\Theta)} {r}^{-1}{{{{{{\rm{E}}_{{\theta _i}}}{e^{r|{X_i} - \dot b({\theta _i})|}}}}}} $. \end{proposition} \begin{proof} Let $s \in \mathbb{R}{\mathrm{i}}:=\{b{\mathrm{i}}:b \in \mathbb{R}\}$ be a given complex number on imaginary axis. \begin{align}\label{eq:barm} {{\bar m}_{\theta _i}}(s) &= {{\rm{E}}_{{\theta _i}}}( {{e^{s[{Y_i} - \dot b({\theta _i})]}}1\{ {Y_i} \ge \dot b({\theta _i})\} } ) + {{\rm{E}}_{{\theta _i}}}( {{e^{s[\dot b({\theta _i}) - {Y_i}]}}1\{ {Y_i} < \dot b({\theta _i})\} } )\nonumber\\ &= \smallint{}_{x \ge \dot b({\theta _i})} {c(x){e^{x({\theta _i} + s)}}{e^{ - \dot b({\theta _i})}}\mu (dx)} + \smallint{}_{x < \dot b({\theta _i})} {c(y){e^{x({\theta _i} - s)}}{e^{ - \dot b({\theta _i})}}\mu (dx)} \nonumber\\ &= {e^{ - \dot b({\theta _i})}}[ {\smallint{}_{x \ge \dot b({\theta _i})} {c(x){e^{x({\theta _i} + s)}}\mu (dx)} + \smallint{}_{x < \dot b({\theta _i})} {c(x){e^{x({\theta _i} - s)}}\mu (dx)} } ]. \end{align} The natural parameter space implies $\int {c(x){e^{x{\theta _i}}}\mu (dx)}$ is finite and analytic for ${\theta _i} \in \Theta$, so \begin{center} $\smallint{1\{ x \ge \dot b({\theta _i})\} c(x){e^{x({\theta _i} +s)}}\mu (dx)}$ and ${\int {1\{ x < \dot b({\theta _i})\} c(y){e^{x({\theta _i} - s)}}\mu (dx)} }$ \end{center} are finite and analytic for $s \in \mathbb{r}{\rm{i}}, {\theta _i}\in\Theta$. By Lemma \ref{le:Analytic}, $\bar{m}_{\theta _i}(s)$ in \eqref{eq:barm} is analytic on \begin{center} $D_{{\theta _i}}:=\left\lbrace s \in \mathbb{C}: \textrm{Re}({\theta _i}+s) \in \textrm{Int}(\Theta)~\textrm{and} ~\textrm{Re}({\theta _i}-s) \in \textrm{Int}(\Theta) \right\rbrace .$ \end{center} by using analytic continuation [i.e. the $\bar{m}_{\theta _i}(s)$ has an analytic continuation from ${{\bar m}_{\theta _i}}(s)$ on $s\in D_{{\theta _i}}$ to ${{\bar m}_{\theta _i}}(s)$ on $s \in \mathbb{C}$, see Corollary 4.9 in \cite{Shakarchi10}]. Since $0+{\theta _i}=\theta_i \in D_{{\theta _i}} \subset \textrm{Int}(\Theta)$, ${{\bar m}_{\theta _i}}(s)$ is analytic at the point 0 and hence the function is also analytic in a neighborhood of 0. By the analyticity of the functions $\{\bar{m}_{\theta _i}(s)\}_{{\theta _i}\in\Theta}$ on $s\in \textrm{Int}(\Theta)$, and Cauchy's derivative inequality with $z_0=0$, we have \begin{align}\label{eq:Cauchy} {\rm{E}}_{\theta _i}{| {{Y_i} - \dot b({\theta _i})} |^k} &= {{\bar m}_{\theta _i}}^{(k)}(0) \le k!{{{r^{-k}}}}\mathop {\sup }{}_{|s|= r} |{{\rm{E}}_{\theta _i}}{e^{s|{Y_i} - \dot b({\theta _i}){\rm{|}}}}|,~~{0 < r \le {\rm{r}}(\Theta)}. \end{align} Let $s = r(\cos \omega + {\rm{i}}\sin \omega ),\omega \in [0,2\pi ]$. Then, we get ${{\rm{E}}_{{\theta _i}}}{e^{s|{Y_i} - \dot b({\theta _i}){\rm{|}}}} = {{\rm{E}}_{{\theta _i}}}{e^{r(\cos \omega + {\rm{i}}\sin \omega )|{Y_i} - \dot b({\theta _i}){\rm{|}}}}$\\$ = {{\rm{E}}_{{\theta _i}}}[ {{e^{r\cos \omega |{Y_i} - \dot b({\theta _i})|}}{e^{{\rm{i}}r\sin \omega |{Y_i} - \dot b({\theta _i})|}}} ].$ Hence, \eqref{eq:Cauchy} gives \begin{align*} k!\frac{1}{{{r^k}}}\mathop {\sup }\limits_{|s| = r}| {{{\rm{E}}_{{\theta _i}}}{e^{s|{Y_i} - \dot b({\theta _i})|}}} |&\le k!\frac{1}{{{r^k}}}\mathop {\sup }\limits_{\omega \in [0,2\pi ]}{{{\rm{E}}_{{\theta _i}}}{e^{r\cos \omega |{Y_i} - \dot b({\theta _i})|}}} \le k!\frac{{{\rm{E}}_{{\theta _i}}}{e^{r|{Y_i} - \dot b({\theta _i})|}}}{{{r^k}}},\\ (\text{Due to}~{{\rm{E}}_{{\theta _i}}}{e^{r|{Y_i} - \dot b({\theta _i})|}} \ge 1)~~& = k!{{\rm{\{ }}\frac{{{{{\rm{[}}{{\rm{E}}_{{\theta _i}}}{e^{r|{Y_i} - \dot b({\theta _i})|}}]}^{1/k}}}}{r}\} ^k} \le k!{{\rm{\{ }} \frac{{{{{\rm{[}}{{\rm{E}}_{{\theta _i}}}{e^{r|{Y_i} - \dot b({\theta _i})|}}]}}}}{r}\} ^k}. \end{align*} From \eqref{eq:Cauchy}, it shows that by take infimum over ${0 < r \le{\rm{r}}(\Theta)}$, \begin{center} ${\rm{E}}_{\theta _i}{| {{Y_i} - \dot b({\theta _i})} |^k} \le k!{{\rm{\{ }}\mathop {\inf }\limits_{0 < r \le {\rm{r}}(\Theta) } {r^{-1}}{{{{{\rm{[}}{{\rm{E}}_{{\theta _i}}}{e^{r|{Y_i} - \dot b({\theta _i})|}}]}}}}\} ^k}\le k!C_{\theta_i}^k =\frac{k!}{2}{(\sqrt {2C_{{\theta _i}}} )^2}C_{{\theta _i}}^{k - 2}$ \end{center} where ${C_{\theta_i}}: = \mathop {\inf }\limits_{0 < r \le {\rm{r}}(\Theta)} {r^{-1}}{{{{{\rm{[}}{{\rm{E}}_{{\theta _i}}}{e^{r|{Y_i} - \dot b({\theta _i})|}}]}}}} $. Then for $\{{{w_i}({Y_i}} - {\rm{E}}{Y_i})\}_{i=1}^n$, we have $ {\rm{E}}_{\theta _i}{|{{w_i}({Y_i}} - {\rm{E}}{Y_i}) |^k}\le \frac{1}{2}k!{(\sqrt {2C_{{\theta _i}}} )^2}C_{{\theta _i}}^{k - 2}{w}^k=\frac{1}{2}k!{({w}\sqrt {2C_{{\theta _i}}} )^2}({w}C_{{\theta _i}})^{k - 2},~~k=2,3,\cdots.$ \end{proof} Therefore, ${w_i}X_{i}\sim \mathrm{sub}\Gamma(w\sqrt {2C_{{\theta _i}}},wC_{{\theta _i}})$ by Proposition \ref{lm-Bernsteingm} and we can apply the Bernstein's inequality with the growth of moments condition to get the following concentration of exponential family on a natural parameter space. \begin{theorem}[Concentration of exponential family] \label{lem-moment} Let $\{ {Y_i}\} _{i = 1}^n$ be a sequence of independent r.vs with their densities $\{f(y_i;\theta_i )\}_{i = 1}^n$ belong to canonical exponential family \eqref{eq:E-P} on the natural parameter space ${\theta _i} \in\Theta$. Given non-random weights $\{w_i\}_{i=1}^n$ with $w= \mathop {\max }_{1 \le i \le n} |w_i|>0$, then \begin{equation}\label{eq-BernsteinEP} P( {| \sum\limits_{i = 1}^n {{w_i}({Y_i} - {\rm{E}}{Y_i})}| \ge t} ) \le 2\exp( - \frac{{{t^2}}}{{4w^2 \sum_{i = 1}^n {C_{{\theta _i}}} + 2w\mathop {\max }\limits_{1 \le i \le n} C_{{\theta _i}} t}}). \end{equation} \end{theorem} Theorem \ref{lem-moment} has no compact space assumption. If we impose the compact space assumption {\rm{(E.1)}} in Proposition \ref{pro-moment}, it leads to the sub-Gaussian concentration as presented in Proposition \ref{lem-moment}. The constant $C_{{\theta _i}}$ in Theorem \ref{lem-moment} is hard to determine in general exponential family with infinite support. However, if the exponential family is Poisson, the $C_{{\theta _i}}$ can be obtained as an explicit form. \begin{theorem}[Concentration for weighted Poisson summation]\label{col:Poisson} Let $\{ {Y_i}\} _{i = 1}^n$ be independent $\{{\rm{Poisson}}(\lambda_i)\} _{i = 1}^n$ distributed. For non-random weights $\{w_i\}_{i=1}^n$ with $w= \mathop {\max }_{1 \le i \le n} |w_i|>0$, put $S_n^w:=\sum_{i = 1}^n {{w_i}({Y_i} - {\rm{E}}{Y_i})} $, then for all $t \ge 0$ \begin{equation}\label{eq:Poisson} {P} (|S_n^w|\geq t)\leq 2 \exp(-{\frac {t^{2}/2}{{{w^2}\sum_{i = i}^n\lambda_i}+ w t/3}}),~~P\{|S_n^w|> w [(2t \sum\limits_{i = 1}^n\lambda_i)^{1/2}+\frac{t}{3}]\} \leq e^{-t}. \end{equation} \end{theorem} \begin{proof} We evaluate the log-MGF of centered Poisson r.vs $\{Y_{i} - {\mathrm{{E}}}Y_{i}\}_{i=1}^n$ \begin{center} $\log {\mathrm{{E}}} {e^{s{w_i}({Y_{i}} - {\mathrm{{E}}}{Y_{i}})}} = - s{w_i}{\mathrm{{E}}}{Y_{i}} + \log \mathrm{E} {e^{s{w_i}{Y_{i}}}} = - {\lambda_{i}}s{w_i} + \log {e^{\lambda_i ({e^{s{w_i}}}- 1)}} ={\lambda_i ({e^{s{w_i}}}-s{w_i}- 1)}.$ \end{center} Note that, for $s$ in a small neighbourhood of zero, \begin{align}\label{eq:lessq} {\lambda_i ({e^{s{w_i}}}-s{w_i}- 1)}= \lambda_i\sum\limits_{k = 2}^\infty {\frac{{{{(s{w_i})}^k}}}{{k!}}} & \le \lambda_i\sum\limits_{k = 2}^\infty {\frac{{{{(\left| {sw} \right|)}^k}}}{{k!}}} = \frac{{\lambda_i{s^2}{w^2}}}{2}\sum\limits_{k = 2}^\infty {\frac{{{{ w }^{k - 2}}}}{{k(k - 1) \cdots 3}}} \nonumber\\ &\le \frac{\lambda_i{{s^2}{w^2}}}{2}\sum\limits_{k = 2}^\infty {{{\left( {\frac{{\left| {sw} \right|}}{3}} \right)}^{k - 2}}} = \frac{{{s^2}}}{2}\frac{{{w^2}\lambda_i}}{{1 - w\left| s \right|/3}}, \end{align} for $\left| s \right| \le 3/w$, which implies $w_i({Y_{i}} - {\mathrm{{E}}}{Y_{i}})\sim \mathrm{sub}\Gamma({{w^2}\lambda_i},w/3)$. By Proposition \ref{sub-GammaConcentration}(a), we have $ S_n^w \sim \mathrm{sub}\Gamma({{w^2}\sum_{k = i}^n\lambda_i},w/3).$ Then applying Proposition \ref{sub-GammaConcentration}(b), we get \eqref{eq:Poisson}. \end{proof} Before ending this section, we show a result for checking Bernstein's moment condition by the moment recurrence condition of log-concave distributions. \begin{definition}[Moment recurrence condition] \label{ex:log-concave} A r.v. $Z$ is called \emph{moment bounded} with parameter $L>0$ if it has recurrence condition $\mathrm{E} |Z|^p \le \; p \; L \cdot \mathrm{E}|Z|^{p-1}$ for any integer $p \ge 1$. \end{definition} By the recursion relation, Definition \ref{ex:log-concave} implies that any moment bounded r.v. $Z$ satisfies $\mathrm{E} |Z|^p \le p!L^P$. Hence, the tails of its moment bounded r.vs decay as the Bernstein's growth of moment condition. So the constant $C_{{\theta _i}}$ in Theorem \ref{lem-moment} is relatively easy to find. Lemmas 7.2, 7.3, 7.6 and 7.7 in \cite{Schudy11} showed that any \emph{log-concave continuous distribution}(see Section 3.4) and \emph{log-concave discrete distribution} $X$ with density $f$ is moment bounded with parameter $L \propto \mathrm{E}|X|$. \begin{example}[Log-concave continuous distributions, \cite{Bag2005}] Many continuous distributions, such as \emph{normal distribution}, \emph{exponential distribution}, \emph{uniform distribution over any convex set}, \emph{logistic distribution}, \emph{extreme value distribution}, \emph{chi-square distribution}, \emph{chi distribution}, \emph{hyperbolic secant distribution}, \emph{Laplace distribution}, \emph{Weibull distribution} (the shape parameter $\theta \ge 1$), \emph{Gamma distribution} (the shape parameter $a\ge 1$) and \emph{Beta distribution} (both shape parameters are $\ge 1$) have log-concave continuous densities. \end{example} Analogous to the log-concave continuous function in \eqref{eq:log-concave}, we can define log-concave sequence for the p.m.f. of discrete r.v., which also has Bernstein-type concentrations. \begin{definition}[Log-concave discrete distributions] \label{ex:log-concavede} A sequence $\{p_i\}_{i \in \mathbb{Z}}$ (or $\{p_i\}_{i \in \mathbb{N}}$) is said to be \emph{log-concave} if $p_{i+1}^{2} \geq$ $p_{i} p_{i+2}$ for all $i \in \mathbb{Z}$ (or $i \in \mathbb{N}$). An integer-valued r.v. $X$ is log-concave if its probability mass function (p.m.f.) $p_{i}:=P(X=i)$ is log-concave sequence. \end{definition} \begin{example}[Log-concave discrete distributions] \emph{Bernoulli and binomial distributions}, \emph{Poisson distribution}, \emph{geometric distribution}, and \emph{negative binomial distribution} (with number of success $>1$) and \emph{hypergeometric distribution} have log-concave integer-valued p.m.f., see \cite{johnson05}. \end{example} \section{Sub-Weibull distributions}\label{Sub-Weibull} \subsection{Sub-Weibull r.vs and $\psi_{\theta}$-norm} A r.v. is heavy-tailed if its distribution function fails to be bounded by a decreasing exponential function \citep{Foss2011}. We first give a simple example of the heavy-tailed distributions {arisen by} multiplying sub-Gaussian r.vs. The proof is motivated by Lemmas 2.7.7 of \cite{Vershynin18}. \begin{lemma}[The product of sub-Gaussians]\label{lem: dproduct sub-gaussian} Suppose $\{X^{(m)}\}_{m=1}^{d}$ are sub-Gaussian (may be dependent). Then $\prod\limits_{m = 1}^d {|{X^{(m)}}{|^{2/d}}} $ is sub-exponential and $ \|\prod\limits_{m = 1}^d {{{[{X^{(m)}}]}^{2/d}}} \|_{\psi_1} \le \prod\limits_{m = 1}^d\|{X^{(m)}}\|_{\psi_2}^{2/d}.$ \end{lemma} \begin{proof} By the definition of sub-Gaussian norm, ${\rm{E}}e^{ |{X^{(m)}}/{\| {{X^{(m)}}}\|_{{\psi _2}}}|^2} \le 2,~m = 1,2, \cdots ,d.$ Applying the elementary inequality $\prod_{m = 1}^d {{a_m}} \le \frac{1}{d}\sum_{m = 1}^d {a_m^d} $, we get by Jensen's inequality \begin{align}\label{eq: X Y sub-w} {\rm{E}}e^{ \prod\limits_{m = 1}^d [|{X^{(m)}}{|^{2/d}}/{{\| {{X^{(m)}}}\|}_{{\psi _2}}^{2/d}}] } \le {\rm{E}}e^{ \frac{1}{d}\sum\limits_{m = 1}^d {{{[{X^{(m)}}/{{\| {{X^{(m)}}}\|}_{{\psi _2}}}]}^2}} } \le \frac{1}{d}\sum\limits_{m = 1}^d {{\rm{E}}e^{{[{X^{(m)}}/{{\| {{X^{(m)}}} \|}_{{\psi _2}}}]}^2}} \le {\rm{2}}. \end{align} The proof is finished by the definition of the sub-exponential norm. \end{proof} In probability, Weibull r.vs are generated from the power of the exponential r.vs. \begin{example}[Weibull r.vs]\label{eq:Weibullb} The Weibull r.v. $X \in \mathbb{R}^+$ is defined by its survival function \begin{center} ${P}(X \ge x) = e^{-bx^{\theta}}~(x\ge 0)~\text{for the scale parameter}\,b>0~\text{and the shape parameter}~\theta>0.$ \end{center} \end{example} Sub-Weibull distribution is characterized by the right tail of the Weibull distribution and is a generalization of both sub-Gaussian and sub-exponential distributions. \begin{definition}[Sub-Weibull distributions] \label{def:subweibull} A r.v. $X$ satisfying $ {P}(|X| \ge x) \le ae^{-bx^{\theta}}$ for given $a, b, \theta>0$, is called a sub-Weibull r.v. with tail parameter $\theta$~(denoted by $X \sim \operatorname{subW}(\theta)$). \end{definition} A $\operatorname{subW}(\theta)$'s tail is no heavier than that of a Weibull r.v. with tail parameter $\theta$. It is emphasized that $X \sim \operatorname{subW}(\theta)$ r.vs with $\theta< 1$ belongs to heavy-tailed r.vs. Recently, the Weibull-like tail condition is also studied in high-dimensional statistics and random matrix theory [see \cite{Tao13}, \cite{Kuchibhotla18} and \cite{Wong17}]. \cite{Gotze2019} names $\operatorname{subW}(\theta)$ as $\theta$-sub-exponential r.v. There are 4 equivalent conditions to reveal the sub-Weibull tail condition which is useful in applications. \begin{corollary}[Characterizations of sub-Weibull condition]\label{th:subWeibull} Let $X$ be a r.v.. Then the following properties are equivalent. \begin{enumerate}[\rm{(}1\rm{)}] \item The tails of $X$ satisfy ${P}(|X| \ge x) \le e^{ - (x/ K_1)^{\theta}},~\text{for all } x \ge 0$. \item The moments of $X$ satisfy $ \|X\|_k:=(\mathrm{E}|X|^{k})^{1 / k} \leq K_{2} k^{1 /\theta}~\text{for all } k \ge 1 \wedge \theta$; \item The MGF of $|X|^{1/\theta}$ satisfies $ \mathrm{E}e^{ \lambda^{1/\theta} |X|^{1/\theta} }\le e^{\lambda^{1/\theta} K_3^{1/\theta} }$ for $|\lambda| \le \frac1{K_3}$; \item The MGF of $|X|^{1/\theta}$ is bounded at some point: ${\rm{E}}e^ {|X/K_4|^{1/\theta}} \le 2.$ \end{enumerate} \end{corollary} The proof can be founded in \cite{Wong17}, \cite{Vladimirova2019} by mimicking the proof of Proposition 2.5.2 in \cite{Vershynin18}. It follows from Corollary~\ref{th:subWeibull}(4) that $X$ is sub-Weibull with tail parameter $\theta$ if and only if $|X|^{1 / \theta}$ is sub-exponential. Let $\theta_{1}$ and $\theta_{2}$ ($0<\theta_{1} \leq \theta_{2}$) be two sub-Weibull parameters. Corollary \ref{th:subWeibull} implies $ \operatorname{subW}\left(\theta_{1}\right) \subset \operatorname{subW}\left(\theta_{2}\right). $ The following Orlicz-type norms play crucial roles in deriving tail and maximal inequality for sub-Weibull r.vs without the zero-mean assumption. \begin{definition}[Sub-Weibull norm or $\psi_{\theta}$-norm] Let $\psi_{{\theta}}(x)=e^{x^{{\theta}}}-1$. The sub-Weibull norm of $X$ for any $\theta>0$ is defined as $ \|X\|_{\psi_{\theta}}:=\inf \{C\in(0, \infty): ~ \mathrm Ee^{|X|^{\theta}/C^{\theta}}\leq 2\}. $ \end{definition} From Corollary \ref{th:subWeibull}(4), a second useful definition of sub-Weibull r.vs is the r.vs with finite $\psi_{\theta}$-norm. Sub-Weibull norm is a special case of the Orlicz norm \citep{Well17}. \begin{definition}[Orlicz norms]\label{def:IncOrliczNorm} Let $g:\,[0, \infty) \to [0, \infty)$ be a non-decreasing convex function with $g(0) = 0$. The ``$g$-Orlicz norm'' of a r.v. $X$ is $\|X\|_{g}:=\inf \{\eta>0: \mathrm{E}[g(|X| / \eta)] \leq 1\}.$ \end{definition} Let $g(x)=e^{x^{{\theta}}}-1$ and $ \mathrm{E}[g(|X| / \eta)] \leq 1$ implies $\mathrm E[\exp(|X|^{\theta}/ \eta^{\theta})]\leq 2$, which is the definition of sub-Weibull norm. Similar to sub-exponential, \cite{Zajkowski19}, \cite{Wong17}, \cite{Vladimirova2019} attained the following. \begin{corollary}[Properties of sub-Weibull norm]\label{prop:Psub-W} \text{If}~$\mathrm{E} e^{{|X / \|X\|_{\psi_{\theta}}|^{\theta}}} \le 2$, then \rm{(}a\rm{)}. $P (|X| > t) \le 2 e^{-(t / \|X\|_{\psi_{\theta}})^{\theta}}~\text{for all } t \ge 0$; \rm{(}b\rm{)}. Moment bounds: $\mathrm{E} |X|^k \le 2 \|X\|_{\psi_{\theta}}^{k} \Gamma(\frac{k}{\theta}+1)$. \end{corollary} \subsection{Concentrations for sub-Weibull summation} The Chernoff inequality tricks in the derivation of Corollary \ref{sub-exponentialConcentration} for sub-exponential concentration is not valid for sub-Weibull distributions, since the exponential moment equivalent conditions of sub-Weibull are on the absolute value $|X|$. However, Bernstein's moment condition is the exponential moment of the absolute value. An alternative method is given by \cite{Kuchibhotla18}, who defines the so-called Generalized Bernstein-Orlicz (GBO) norm. Fixed $\alpha > 0$ and $L \ge 0$, define a function $\Psi_{\theta, L}(\cdot)$ with its inverse function $ \Psi_{\theta, L}^{-1}(t) := \sqrt{\log (t+1)} + L[\log (t+1)]^{1/\theta}~\forall~t\ge 0.$ A promising development is that the following GBO norm helps us derive tail behaviors for sub-Weibull r.vs. \begin{definition}[Generalized Bernstein-Orlicz Norm]\label{def:GBOnorm} The \emph{generalized Bernstein-Orlicz} (GBO) norm of a r.v. $X$ is then given by: $\|X\|_{\Psi_{\theta, L}}:=\inf \{\eta>0: \mathrm{E}[{\Psi_{\theta, L}}(|X| / \eta)] \le 1\}.$ \end{definition} The monotone function $\Psi_{\theta, L}(\cdot)$ is motivated by the classical Bernstein's inequality for sub-exponential r.vs. Like the sub-Weibull norm properties Corollary \ref{prop:Psub-W}(a), the following proposition in \cite{Kuchibhotla18} allows us to get the concentration inequality for r.vs with finite GBO norms. \begin{corollary}[GBO norm concentration]\label{prop:GBO norm} For any r.v. $X$ with $\norm{X}_{\Psi_{\theta, L}}< \infty$, we have $ {P}(|X| \ge \norm{X}_{\Psi_{\theta, L}}\{\sqrt{t} + Lt^{1/\theta}\}) \le 2e^{-t},~\mbox{for all}~ t\ge 0.$ \end{corollary} From Corollary \ref{prop:GBO norm}, it is easy to derive the concentration inequality for a single sub-Weibull r.v. or even the sum of independent sub-Weibull r.vs. Theorem 3.1 in \cite{Kuchibhotla18} obtains an upper bound for the GBO norm of the summation. \begin{corollary}[Concentration for sub-Weibull summation]\label{thm:SumNewOrliczVex} If $\{X_i\}_{i=1}^n$ are independent centralized r.vs such that $\norm{X_i}_{\psi_{\theta}} < \infty$ for all $1\le i\le n$ and some $\theta > 0$, then for any weight vector $\bm w= (w_1, \ldots, w_n)\in\mathbb{R}^n$, we have $\|\sum_{i=1}^n w_iX_i\|_{\Psi_{\theta, L_n(\theta)}} \le 2eC(\theta)\norm{ \bm b}_2$ and \begin{equation}\label{eq:sWc} P(|\sum{}_{i = 1}^n {{w_i}{X_i}} | \ge 2eC(\theta ){\left\|\bm b \right\|_2}\{ \sqrt t + {L_n}(\theta ){t^{1/\theta }}\} ) \le 2{e^{ - t}} \end{equation} where $\bm b = (w_1\norm{X_1}_{\psi_{\theta}}, \ldots, w_n\norm{X_i}_{\psi_{\theta}})^T\in\mathbb{R}^n$, $L_n(\theta) := \frac{4^{1/\theta}}{\sqrt{2}\norm{\bm b}_2}\times\begin{cases}\norm{\bm b}_{\infty},&\mbox{if }\theta < 1,\\ {4e\norm{\bm b}_{\frac{\theta }{{{\rm{1}} - \theta }}}}/{C(\theta)},&\mbox{if }\theta \ge 1\end{cases}$ \begin{center} $ \text{and}~C(\theta) ~:=~ \max\{ \sqrt{2}, 2^{1/\theta}\} \times \begin{cases}\sqrt{8}e^3(2\pi)^{1/4}e^{1/24}(e^{2/e}/\theta)^{1/\theta},&\mbox{if }\theta < 1,\\ 4e + 2(\log 2)^{1/\theta},&\mbox{if }\theta \ge 1. \end{cases} $ \end{center} \end{corollary} The upper bound of sub-Weibull norm for summation provided by Corollary \ref{thm:SumNewOrliczVex} dependents on $\left\|X_{i}\right\|_{\psi_{\theta}}$ and the $\bm w$. \cite{Zhang20} gives a sharper version of Corollary \ref{thm:SumNewOrliczVex}. The $\theta=1$ is the phrase transition point, and reflect the fact that Weibull r.vs are log-convex for $\theta \leq 1$ and log-concave for $\theta \geq 1$. At last, we mention a generalized Hanson-Wright inequality for sub-Weibull r.vs in Proposition 1.5 of \cite{Gotze2019}. Let $\max _{i=1, \ldots, n}\|(a_{i j})_{j}\|_{2}:=\|A\|_{2 \rightarrow \infty}$ where $\|A\|_{p \rightarrow q}:=\sup \{\|A x\|_{q}:\|x\|_{p} \leq 1\} .$ \begin{corollary}[Concentration for the quadratic form of sub-Weibull r.vs.]\label{prop:Qsub-Weibull} Let $q \in \mathbb{N}, A=\left(a_{i j}\right)$ be a symmetric $n \times n$ matrix and let $\{X_i\}_{i=1}^n$ be independent and centered r.vs with $\left\|X_{i}\right\|_{\Psi_{2 / q}} \leq M$ and $\mathrm{E} X_{i}^{2}=\sigma_{i}^{2} .$ We have ${P}(|\sum_{i, j} a_{i j} X_{i} X_{j}-\sum_{i=1}^{n} \sigma_{i}^{2} a_{i i}| \geq t) \leq 2 e^{-\eta(A, q, t / M^{2})/C}$, $\forall t \ge 0$, where $\eta(A, q, t):=\min (\frac{t^{2}}{\|A\|_{\mathrm{F}}^{2}}, \frac{t}{\|A\|_{\mathrm{op}}},(\frac{t}{\max_{i=1, \ldots, n}\|(a_{i j})_{j}\|_{2}})^{\frac{2}{q+1}},(\frac{t}{\|A\|_{\infty}})^{\frac{1}{q}})$ and $C$ is a constant. \end{corollary} \section{Concentration for Extremes}\label{Extremes} The CIs presented so far only concern with linear combinations of independent r.vs or Lipschitz function of random vectors. In many statistics applications, we have to control the maximum of the $n$ r.vs when deriving the error bounds, while these r.vs may be arbitrarily dependent. This section is developed on advanced proof skills. So we present the proofs even for existing results, which are applications of CIs in a probability aspect. \subsection{Maximal inequalities} This section presents the maximal inequalities for r.vs. $\{X_i\}_{i=1}^n$ which may not be independent. In the theory of empirical process, it is of interest to bound $\mathrm{E}{{\max }_{1 \le i \le n} |X_i|}$ [Section 2.2, \cite{van96}]. If $\{X_i\}_{i=1}^n$ are arbitrary sequence of real-valued r.vs and have finite $r$-th moments ($r\ge1$), \cite{Aven85} gives a crude upper bounds for $\mathrm{E}{{\max }_{1 \le i \le n} {X_i}}$ by Jensen's inequality \begin{align}\label{le:crude} \mathrm{E}\{\mathop {\max }\limits_{1 \le i \le n}\left|X_{i}\right|^r\}^{1/r} \le \{\mathrm{E}\mathop {\max }\limits_{1 \le i \le n}\left|X_{i}\right|^r\}^{1/r} \le {\{ \sum\limits_{i = 1}^n{\rm{E}} {\left| {{X_i}} \right|^r}\} ^{1/r}} \le {n^{1/r}} \mathop {\max }\limits_{1 \le i \le n} {{\rm{(E}}{\left| {{X_i}} \right|^r})^{1/r}} \end{align} Page314 of \cite{Vaart1998} mentions a sharper version of \eqref{le:crude} without the proof. In below, we introduce the proof by the truncation technique. \begin{corollary}[Sharper maximal inequality]\label{le:sharpmax} Let $\{X_i\}_{i=1}^n$ be identically distributed but not necessarily independent and assume $\mathrm{E}(|X_1|^p) < \infty,(p \ge 1)$. Then, $ \mathrm{E}{\mathop {\max }_{1 \le i \le n} |X_i|}=o(n^{1 / p}). $ \end{corollary} \begin{proof} Let $M_{n}:={\mathop {\max }_{1 \le i \le n} |X_i|}$. For any $\epsilon>0,$ we truncate $M_{n}$ by ${\epsilon n^{1 / p}}$, $$ \begin{aligned} \mathrm{E}M_{n} &=\int_{0}^{\epsilon n^{1 / p}} {P}\left(M_{n}>t\right) d t+\int_{\epsilon n^{1 / p}}^{\infty}{P}\left(M_{n}>t\right) d t \leq \int_{0}^{\epsilon n^{1 / p}} 1 d t+\int_{_{\epsilon n^{1 / p}}}^{\infty} n {P}\left(|X_{1}|>t\right) d t \\ &=\epsilon n^{1 / p}+n^{1 / p} \int_{\epsilon n^{1 / p}}^{\infty} n^{(p-1) / p}{P}\left(|X_{1}|>t\right) d t\leq \epsilon n^{1 / p}+\frac{n^{1 / p}}{\epsilon^{p-1}} \int_{\epsilon n^{1 / p}}^{\infty} t^{p-1} {P}\left(|X_{1}|>t\right) d t. \end{aligned} $$ Thus, by dividing $n^{1 / p}$ we have $\frac{\mathrm{E}M_{n}}{n^{1 / p}} \leq \epsilon+\frac{1}{\epsilon^{p-1}} \int_{\epsilon n^{1 / p}}^{\infty} t^{p-1}{P}\left(|X_{1}|>t\right) d t= \epsilon+o(1)$, where we adopt the fact $\int_{\epsilon n^{1 / p}}^{\infty} t^{p-1}{P}\left(|X_{1}|>t\right) d t=o(1)$ from moment condition: $\mathrm{E}\left|X_{1}\right|^{p}<\infty$. Finally, it implies that ${\limsup}_{n \rightarrow \infty} \frac{\mathrm{E}M_{n}}{n^{1 / p}} \leq \epsilon$, which gives $\mathrm{E}M_{n}=o(n^{1 / p})$ by letting $\epsilon \to 0$. \end{proof} Corollary \ref{le:sharpmax} reveals that ${{\max }_{1 \le i \le n} |X_i|}$ diverges at rate slower than $n^{1 /r}$ under the $r$-th moment condition. As $r$ increases, it will slow down the divergence rate of the maxima. If we have arbitrary finite $r$-th moment conditions (such as Gaussian distribution), it means that the divergence rate of maxima is slower than any polynomial rate $n^{1 /r}$. This suggests that the rate may be logarithmic. With the sub-Gaussian assumptions, the logarithmic divergence rate is possible and the proof is based on controlling the expectation of the supremum of variables, from the argument in \cite{Pisier1983}. \begin{corollary}[Sub-Gaussian maximal inequality, \cite{Rigollet19}]\label{eq:Sub-Gaussianmaximal} Let $\{X_{i}\}_{i=1}^n$ be r.vs (without independence assumption) such that $X_{i} \sim \operatorname{subG}\left(\sigma^{2}\right).$ Then, (a) $\mathrm{E}[\max\limits _{1 \leq i \leq n} X_{i}] \leq \sigma \sqrt{2 \log n}$ and $ \mathrm{E}[\max\limits _{1 \leq i \leq n}\left|X_{i}\right|] \leq \sigma \sqrt{2 \log (2 n)}$. (b) ${P}(\max \limits_{1 \leq i \leq n} X_{i}>t) \leq n e^{-\frac{t^{2}}{2 \sigma^{2}}} ~ \text { and }~{P}(\max \limits_{1 \leq i \leq n}\left|X_{i}\right|>t) \leq 2 n e^{-\frac{t^{2}}{2 \sigma^{2}}}.$ \end{corollary} \begin{proof} (a) By the property of maximum, sub-Gaussian MGF and Jensen's inequality, \begin{align*} \mathrm{E} \max_{1 \le i \le n} X_i& =\mathop {\inf }\limits_{s > 0} s^{-1} \mathrm{E} \log e^{s\max\limits_{1 \le i \le n} X_i } \le \mathop {\inf }\limits_{s > 0} s^{-1} \log \mathrm{E} e^{s\max\limits_{1 \le i \le n} X_i}\le \mathop {\inf }\limits_{s > 0} s^{-1}\log\sum_{i=1}^n \mathrm{E} e^ {s X_i}\le \mathop {\inf }\limits_{s > 0} s^{-1} \log \sum_{i=1}^n e^{\frac{\sigma^2s^2}{2}}\\ &=\mathop {\inf }\limits_{s > 0} \left( {\frac{{\log n}}{s} + \frac{{{\sigma ^2}s}}{2}} \right) =\sigma \sqrt{2 \log n}~~[\text{Setting}~s = \sqrt {\frac{{{\rm{2}}\log n}}{{{\sigma ^2}}}}~\text{as the optimal bound.}] \end{align*} Let $Y_{2i-1} = X_i$ and $Y_{2i}=-X_i (1 \le i \le n)$. It gives $\mathrm{E} \max_{1 \le i \le n} |X_i| = \mathrm{E} \max_{1 \le i \le n} \max\{X_i, -X_i\} = \mathrm{E}\max_{1 \le i \le 2n} Y_i.$ The previous result for sample size $2n$ finishes the proof of the second part. (b) By Chernoff inequality and the sub-Gaussian MGF, we have ${P}( \max_{1 \le i \le n} X_i > t )\le \mathop {\inf }\limits_{s > 0} e^{-s t} \mathrm{E} e^{s \max\limits_{1 \le i \le n} X_i}\le \mathop {\inf }\limits_{s > 0} e^{-s t}\sum_{i=1}^n \mathrm{E} e^{s X_i} \le \mathop {\inf }\limits_{s > 0} n e^{-s t + \frac{\sigma^2s^2}{2}}\xlongequal{s={t}/{\sigma^2}} n e^{- \frac{t^2}{2 \sigma^2}}.$ For the second part, note that ${P}( \max_{1 \le i \le n} |X_i| > t ) = {P}( \max_{1 \le i \le n} X_i > t,\max_{1 \le i \le n} -X_i \ge t) \le 2 {P}( \max_{1 \le i \le n} X_i > t ). $ \end{proof} By a similar proof, Corollary \ref{eq:Sub-Gaussianmaximal} can be extended to other r.vs, such as sub-Gamma r.vs and r.vs characterized by sub-Weibull norm (or Orlicz norm) as presented before. \begin{corollary}[Concentration for maximum of sub-Gamma r.vs]\label{sub-GammaMAX} Let $\{ X_{i}\} _{i = 1}^n $ be independent zero-mean $\{\mathrm{sub}\Gamma(\upsilon_i,c_i)\} _{i = 1}^n$. Then, for $\max_{i=1,...,n}\upsilon_{i} =: \upsilon$ and $\max_{i=1,...,n}c_{i} =: c$, \begin{center} $\mathrm{E} (\max\limits_{i=1,\cdots,n}|X_i|) \leq [2\upsilon \log (2n)]^{1/2} + c \log (2n)$. \end{center} \end{corollary} \noindent(See Theorem 3.1.10 in \cite{Gine15} for the proof of Corollary \ref{sub-GammaMAX}.) In bellow, based on the sub-Weibull norm condition, a fundamental result due to \cite{Pisier1983} is given for obtaining the divergence rate of the maxima of sub-Weibull r.vs. \begin{corollary}[Maximal inequality for sub-Weibull r.vs]\label{pp-MaximalSW} For $\theta>0$, consider the sub-Weibull norm $\|X\|_{\psi_{\theta}}:=\inf_{C\in(0, \infty)}\{ ~ \mathrm E e^{|X|^{\theta}/C^{\theta}}\leq 2\}$ for $\psi_{\theta}(x)=e^{x^{\theta}}-1$. For any r.vs $\{X_{i}\}_{i=1}^n$, \begin{equation}\label{eq-MaximalSW} \mathrm{E}({\mathop {\max }\limits_{1 \le i \le n} |X_i|})\le \psi^{-1}_{\theta}(n)\mathop {\max }\limits_{1 \le i \le n}\|X_i\|_{\psi_{\theta}}=(\log (1+n))^{1 / \theta}\mathop {\max }\limits_{1 \le i \le n}\|X_i\|_{\psi_{\theta}}. \end{equation} If the function ${\psi_{\theta}}(x)$ is replaced by any non-decreasing convex function $g(x)$ with $g(0) = 0$ in the definition of Orlicz norm: $\|X\|_{g}:=\inf \{\eta>0: \mathrm{E}[g(|X| / \eta)] \le 1\}$, then \begin{center} $\mathrm{E}({\mathop {\max }\limits_{1 \le i \le n} |X_i|})\le g^{-1}(n)\mathop {\max }\limits_{1 \le i \le n}\|X_i\|_g$~~~~for finite $\mathop {\max }\limits_{1 \le i \le n}\|X_i\|_g$. \end{center} \end{corollary} \begin{proof} From Jensen's inequality, for $C\in(0, \infty)$ and $\psi_{\theta}(x)=e^{x^{\theta}}-1$ we get \begin{equation}\label{eq-Maximaljs} \psi_{\theta} [ {\mathrm{E}( { \max\limits_{1 \le i \le n} | {{X_i}}|/C} )} ] \le \mathrm{E}[ {\max\limits_{1 \le i \le n} \psi_{\theta}( {| {{X_i}} |/C} )} ] \le {\sum_{i = 1}^n \mathrm{E}{\psi_{\theta} ( {| {{X_i}} |/C} )} } \le n \end{equation} where the last inequality is by the definition of sub-Weibull norm: $\mathrm{E}{\psi_{\theta} ( {| {{X_i}} |/C} )}\le 1$. Let $C={\max }_{1 \le i \le n}\|X_i\|_{\psi_{\theta}}$. Applying the non-decreasing property of $\psi_{\theta}(x)$ (so does its inverse $\psi_{\theta}^{ - 1}(x)$), the \eqref{eq-Maximaljs} implies ${\rm{E}}( {{\max }_{1 \le i \le n}| {{X_i}}|/C} )\le \psi _\theta ^{ - 1}(n)$ by operating the map $\psi_{\theta}^{ - 1}$, and so we have \eqref{eq-MaximalSW}. The derivation of Orlicz norm case is the same. \end{proof} By Hoeffding's lemma, the following results on the maxima of the sum of independent r.vs, is useful for bounding empirical processes. \begin{corollary}[Maximal inequality for bounded r.vs, Lemma 14.14 in \cite{Buhlmann11}]\label{pp-Maximalbd} Let $\{X_{i}\}_{i=1}^n$ be independent r.vs on $\mathcal{X}$ and $\{f_{i}\}_{i=1}^n$ be real-valued functions on $\mathcal{X}$ which satisfy ${\rm{E}}f_{j}(X_{i})=0,~\vert f_{j}(X_{i})\vert \le a_{ij}$ for all $j=1,...,p$ and all $i=1,...,n$. Then \begin{center} ${\rm{E}}( \underset{1\le j\le p}{\max}| \sum\limits_{i=1}^{n}f_{j}(X_{i})|) \le [2\log(2p)]^{1 / 2}\underset{1\le j\le p}{\max}(\sum\limits_{i=1}^{n}a_{ij}^{2})^{1 / 2}.$ \end{center} \end{corollary} \begin{proof} Let ${V_j} = \sum_{i = 1}^n {{f_j}({X_i})} $. By Jensen's inequality and Hoeffding's lemma \begin{align*} {\rm{E}}\mathop {\max }\limits_{1 \le j \le p} |{V_j}| &= \frac{1}{\lambda }{\rm{E}}\log {{\rm{e}}^{\lambda \mathop {\max }\limits_{1 \le j \le p} |{V_j}|}} \le \frac{1}{\lambda }\log {\rm{E}}{{\rm{e}}^{\lambda \mathop {\max }\limits_{1 \le j \le p} |{V_j}|}}\le \frac{1}{\lambda }\log \sum\limits_{i = 1}^p{\rm{E}}{{\rm{e}}^{\lambda |{V_j}|}} \\ [\text{Corollary \ref{lm:Hoeffding}}]~&\le \frac{1}{\lambda }\log [\sum\limits_{j = 1}^p {2e^{ \frac{1}{2}{\lambda ^2}\sum\limits_{i = 1}^n {a_{ij}^2} } } ] \le \frac{1}{\lambda }\log {\rm{[2}}pe^{\frac{1}{2}{\lambda ^2}\mathop {\max }\limits_{1 \le j \le p} \sum\limits_{i = 1}^p {a_{ij}^2}}]= \frac{1}{\lambda }\log {\rm{(2}}p) + \frac{1}{2}\lambda \mathop {\max }\limits_{1 \le j \le p} \sum\limits_{i = 1}^n {a_{ij}^2} . \end{align*} Then ${\rm{E}}\mathop {\max }\limits_{1 \le j \le p} |{V_j}|\le \mathop {\inf }\limits_{\lambda > 0}\{\frac{1}{\lambda }\log {\rm{(2}}p) + \frac{1}{2}\lambda \mathop {\max }\limits_{1 \le j \le p} \sum\limits_{i = 1}^n {a_{ij}^2}\}=\sqrt{2\log(2p)}\underset{1\le j\le p}{\max}(\sum_{i=1}^{n}a_{ij}^{2})^{1 / 2}$. \end{proof} If Hoeffding's lemma for moment is replaced by Bernstein's moment conditions, then the maximal inequality for the sum of independent bounded r.v.s in Corollary \ref{pp-Maximalbd} can be extended to Bernstein's moment conditions. We give a modified version of Corollary 14.1 in \cite{Buhlmann11} based on truncated Jensen's inequality. \begin{proposition}[Maximal inequality with Bernstein's moment conditions]\label{pp-MaximalB} If $\{X_{ij}\}, j=1, \ldots, p$ are read-valued independent variables across $i=1, \ldots, n$. Assume $\mathrm{E}X_{ij}=0$ and Bernstein's moment conditions:$ \frac{1}{n} \sum_{i=1}^{n} {\rm{E}}{\left| {{X_{ij}}} \right|^k} \le \frac{1}{2}v^2{\kappa^{k - 2}}k!,~~ k=2,3, \ldots, \forall j,$ Then, \begin{center} ${\rm{E}}( \underset{1\le j\le p}{\max}| \frac{1}{n}\sum_{i=1}^{n}X_{ij}|^m) \le [{\textstyle{{\kappa \log (2p)} \over n}}+(v^2+1)\sqrt{\textstyle{{\log (2p)} \over n}}]^{m},~\forall~1\le m \leq 1+\log p~\text{and}~p \geq 2.$ \end{center} \end{proposition} \begin{proof} Let $M_{n,m}={\max}_{1\le j\le p}| \frac{1}{n}\sum_{i=1}^{n}X_{ij}|^m$. First, we show for any r.v. $X$ and all $m \geq 1$, \begin{align} \label{moment1} \mathrm{E}|X|^{m} \leq \log ^{m}(\mathrm{E} e^{|X|}-1+e^{m-1}). \end{align} The function $g(x)=\log ^{m}(x+1), x \geq 0$ is concave for all $x \geq e^{m-1}-1 .$ By the truncated Jensen's inequality in Lemma \ref{lem:Truncated} with $Z:= e^{|X|}-1,~c=e^{m-1}-1$, we have \begin{center} $ \begin{array}{c} \mathrm{E}|X|^{m}=\mathrm{E} \log ^{m}({e}^{|X|}-1+1) \leq \log ^{m}[\mathrm{E}({e}^{|X|}-1)+1+\left({e}^{m-1}-1\right)]=\log ^{m}[\mathrm{E}( e^{|X|}-1)+e^{m-1}]. \end{array} $ \end{center} Then for all $L,~m>0,$ \begin{align} \label{JensenP} (\frac{L}{n})^{-m}{\rm{E}}M_{n,m}& \le \log ^{m}[\operatorname{E}e^{ \underset{1\le j\le p}{\max}|\sum\limits_{i=1}^{n}{X_{ij}}/{L}|}-1+e^{m-1}]\le\log ^{m}[\sum_{j=1}^{p}\operatorname{E}(e^{ |\sum\limits_{i=1}^{n}{X_{ij}}/{L}|}-1)+e^{m-1}]. \end{align} Therefore, it is sufficient to bound $\operatorname{E}e^{ |\sum_{i=1}^{n}X_{ij}|/ L}$ uniformly in $j$. Second. To bound the MGF in \eqref{JensenP}, then we show that for any real-valued r.v. $X,$ \begin{align} \label{moment2} \mathrm{E}e^{X} \leq e^{\mathrm{E} e^{|X|}-1-\mathrm{E}|X|},~\text{with}~\mathrm{E} X=0. \end{align} Indeed, for any $c>0$, we have $ e^{X-c}-1 \leq \frac{e^{|X|}}{1+c}-1=\frac{{e}^{X}-1-X+X-c}{1+c} \leq \frac{e^{|X|}-1-| X|+X-c}{1+c}. $ Let $ c=\mathrm{E} e^{|X|}-1-\mathrm{E}|X|. $ Note that $\mathrm{E} X=0$, so $\mathrm{E} e^{X-c}-1 \leq \frac{\mathrm{E} e^{|X|}-1-\mathrm{E}|X|-c}{1+c}=0.$ Hence $\log(\mathrm{E}e^{X}) \leq c$. Using Taylor's expansion, the \eqref{moment2} and ${e}^{|x|} \leq {e}^{x}+{e}^{-x}$ give \begin{align} \label{moment3} \operatorname{E}e^{ | \sum_{i=1}^{n}X_{ij}|/ L}-1\le \operatorname{E}e^{ \sum_{i=1}^{n}X_{ij}/ L}&+\operatorname{E}e^{ -\sum_{i=1}^{n}X_{ij}/ L}-1\le 2e^{\sum_{i=1}^{n} \mathrm{E}(e^{|X_{ij}| / L}-1-|X_{ij}| / L)}-1 = 2e^{\sum\limits_{m=2}^{\infty} \frac{\sum_{i=1}^{n} \mathrm{E}|X_{ij}|^{m}}{L^{m} m !} }-1\nonumber\\ [\text{By moment conditions}]~ &\leq 2e^{\frac{nv^2}{ 2L^{2}} \sum_{m=2}^{\infty}(\kappa / L)^{m-2}}-1=2e^{\frac{nv^2}{ 2L^{2}(1-\kappa / L)}}-1. \end{align} Combining \eqref{moment3} and \eqref{JensenP}, we obtain for $L >\kappa=:L-\sqrt{{n}/{2 \log \left( p+e^{m-1}\right)}} $ \begin{align*} \label{JensenP1} &~~~~{\rm{E}}M_{n,m} \leq (\frac{L}{n})^{m} \log ^{m}[p(2e^{\frac{nv^2}{ 2L^{2}(1-\kappa / L)}}-1)+e^{m-1}] {\leq} (\frac{L}{n})^{m} \log ^{m}[(p+e^{m-1})e^{\frac{nv^2}{2\left(L^{2}-L\kappa\right)}}]\\ &=(\frac{L\log ( p+e^{m-1})}{n}+{\frac{v^2L}{2(L^{2}-L\kappa)}})^{m}=(\frac{\kappa\log ( p+e^{m-1})}{n}+\sqrt{\frac{ \log \left( p+e^{m-1}\right)}{2n}}+{\frac{v^2}{2(L-\kappa)}})^{m}\\ & \le [\frac{\kappa\log ( p+e^{m-1})}{n}+(v^2+1)\sqrt{\frac{ \log \left( p+e^{m-1}\right)}{2n}} ]^{m}\le [\frac{\kappa\log (2p)}{n}+(v^2+1)\sqrt{\frac{ \log \left( 2p\right)}{2n}}]^{m}. \end{align*} where the second and last inequality is by ${\frac{v^2}{L-\kappa}} = [\sqrt{\frac{n}{2 \log \left( p+e^{m-1}\right)}}]^{-1}{v^2}$ and $e^{m-1} \le p$. \end{proof} \subsection{Concentration for suprema of empirical processes} Let $\{X_{i}\}_{i=1}^n$ be a random sample from a measure {{${P}$}} on a measurable space $(\mathcal{X},\mathcal{A})$. The empirical distribution $\mathbb{P}_{n}:=n^{-1} \sum_{i=1}^{n} \delta_{X_{i}},$ where $\delta_{x}$ is the probability mass of 1 at $x$. Given a measurable function $f : \mathcal{X} \mapsto \mathbb{R}$, let $\mathbb{P}_{n} f:=\frac{1}{n} \sum_{i=1}^{n} f\left(X_{i}\right)$ be the expectation of $f$ under the empirical measure $\mathbb{P}_{n}$, and ${P} f:=\int f dP$ be the expectation under {{${P}$}}. The $\mathbb{P}_{n} f$ is called the \emph{empirical process} index by $n$. The study of the empirical processes begins with the uniform limit law of EDF in Example \ref{eg:edf}. The Glivenko-Cantelli theorem extends the LLN for EDF and gives uniform convergence: $\left\|\mathbb{F}_{n}-F\right\|_{\infty}=\sup _{t \in \mathbb{R}}\left|\mathbb{F}_{n}(t)-F(t)\right| \stackrel{\text { as }}{\rightarrow} 0.$ Moreover, a stronger result than Example \ref{eg:edf} is the Dvoretzky-Kiefer-Wolfowitz (DKW) inequality \citep{Dvoretzky56} \begin{equation}\label{eq:DKW} P({\sup{}_{x\in \mathbb{R} }}|\mathbb{F}_{n}(x)-F(x)|>\varepsilon )\leq 2e^{-2n\varepsilon ^{2}}\quad \forall \varepsilon >0. \end{equation} The DKW inequality is a uniform version of Hoeffding's inequality, which also strengthens the Glivenko-Cantelli theorem since \eqref{eq:DKW} implies Glivenko-Cantelli: $\|\mathbb{F}_{n}-F\|_{\infty }\xrightarrow{a.s.}0$ by Borel-Cantelli lemma: $X_n \xrightarrow{a.s.} 0 \Leftrightarrow \sum_{n=1}^{\infty}P(|X_n|\geq \epsilon) < \infty~\text{for any}~\varepsilon > 0.$ \cite{Dvoretzky56} proves $P({\sup_{x\in \mathbb{R} }}|\mathbb{F}_{n}(x)-F(x)|>\varepsilon )\leq Ce^{-2n\varepsilon ^{2}}$ with an unspecified constant $C$. \cite{Massart96} attains the sharper constant $C = 2$. In some statistical applications, given an estimator $\hat{\theta},$ and $f_{\hat{\theta}}(X_i)$ is a function of $X_i$ and ${\hat{\theta}}$. We want to study its asymptotic properties for sums of $f_{\hat{\theta}}(X_i)$ that changes with both $n$ and $\hat{\theta}$, \begin{center} $\frac{1}{n} \sum_{i=1}^{n}[f_{\hat{\theta}}\left(X_{i}\right)-\mathrm{E} f_{\hat{\theta}}(X_{i})],~(\text{a dependent sum}).$ \end{center} A possible route to attain results is via the suprema of the empirical process for all possible the `` true'' parameter $\theta_{0}$ on a set $K$: \begin{center} $ \frac{1}{n} \sum_{i=1}^{n}[f_{\hat{\theta}}\left(X_{i}\right)-\mathrm{E} f_{\hat{\theta}}(X_{i})]\le \sup_{\theta_{0} \in K} | \frac{1}{n} \sum_{i=1}^{n}[f_{\theta_{0}}\left(X_{i}\right)-\mathrm{E} f_{\theta_{0}}(X_{i})]|=: \sup_{{\theta_{0}} \in K} |(\mathbb{P}_{n}-P) f_{\theta_{0}}|. $ \end{center} Fortunately, the summation in the sup enjoys independence. So, the study of convergence rate suprema of empirical processes is important if we consider a functional class ${\cal F}$ instead of the set $K$ such that : ${\mathop {\sup }_{f \in {\cal F}} |(\mathbb{P}_{n}-P) f|}=\sup_{\theta_0 \in K} |(\mathbb{P}_{n}-P)f_{\theta_{0}}|.$ Let $(\mathcal{F},\|\cdot\|)$ be a normed space of real functions $f: \mathcal{X} \rightarrow \mathbb{R}$. For a probability measure $Q$, define the $L_{r}(Q)$-space with $L_{r}(Q)$-norm by $\|f\|_{L_{r}(Q)}=\left(\int|f|^{r} d Q\right)^{1 / r}$. Given two functions $l(\cdot)$ and $u(\cdot),$ the bracket $[l, u]$ is the set of all functions $f \in \mathcal{F}$ with $l(x) \leq f(x) \leq u(x),$ for all $x \in \mathcal{X} .$ An $\varepsilon$-bracket is a bracket $[l, u]$ with $\|l-u\|_{L_{r}(Q)}<\varepsilon$. The \emph{bracketing number} $N_{[~]}\left({\varepsilon}, \mathcal{F}, L_{r}(Q)\right)$ is minimum number of $\varepsilon$-brackets needed to cover $\mathcal{F}$, i.e. \begin{center} $N_{[~]}\left({\varepsilon}, \mathcal{F}, L_{r}(Q)\right)=\inf \{n: \exists l_{1}, u_{1}, \ldots, l_{n}, u_{n} \text { s.t. } \cup_{i=1}^{n}\left[l_{i}, u_{i}\right]=\mathcal{F} \text { and } \|l_n-u_n\|_{L_{r}(Q)} < \varepsilon \}.$ \end{center} The \emph{covering number} $N\left(\varepsilon, \mathcal{F}, L_{r}(Q)\right)$ is the minimal number of $L_{r}(Q)$-balls of radius $\varepsilon$ needed to cover the set $\mathcal{F}$. The uniform covering numbers is $\sup _{Q} N(\varepsilon\|F\|_{ L_{r}(Q)}, \mathcal{F}, L_{r}(Q))$, where the supremum is taken over all probability measures $Q$ for which $\|F\|_{ L_{r}(Q)}>0$. Two conditions to get the convergence of $\sup_{f \in {\cal F}} |(\mathbb{P}_{n}-P)f_{\theta}|$ are the finite \emph{bracketing number} condition with $L_{1}(P)$-norm in Theorem 19.4 of \cite{Vaart1998} (or finite uniform covering numbers in Theorem 19.13 of \cite{Vaart1998}). \begin{lemma}[Glivenko-Cantell class] \label{NA-GC} For every class $\mathcal{F}$ of measurable functions, if \begin{center} $N_{[~]}\left(\varepsilon, \mathcal{F}, L_{1}(P)\right)<\infty$~(or $\sup _{Q} N(\varepsilon\|F\|_{ L_{1}(Q)}, \mathcal{F}, L_{1}(Q))<\infty$ with $P^{*} F<\infty$) for every $\varepsilon>0$. \end{center} Then, $\mathcal{F}$ is $P$-Glivenko-Cantelli, i.e. ${ {\sup }_{f \in {\cal F}} |(\mathbb{P}_{n}- {P})f|}\stackrel{\text { as }}{\rightarrow} 0$. \end{lemma} \begin{example} [Empirical process with indicator functions] Let $\mathcal{F}$ be the collection of all indicator functions of the form $f_{t}(x)=1_{(-\infty, t]}(x),$ with $t$ ranging over $\mathbb{R}.$ Then, $\mathcal{F}$ is \text{P-Glivenko-Cantelli}, see Example 19.4 of \cite{Vaart1998}. \end{example} \begin{example} [Weighted empirical process with dependent weights]\label{eg:Weighted} Suppose we observe a sequence of IID observations $\{\left({X}_{i}, Y_{i}\right)\}_{i=1}^n$ drawn from a random pair $({X}, Y)$. Given some weighted functions $W(\cdot)$ and a bounded estimator $\hat t \in (0,\tau]$, we want to study the stochastic convergence of \textit{dependent weighted empirical processes} \begin{center} $|\frac{1}{n} \sum_{i=1}^{n} 1\left(Y_{i} \geq \hat t\right) W\left({X}_{i}\right)-\mu\left(\hat t ; W\right)|~~(\le \sup _{0 \leq t \leq \tau}|\frac{1}{n} \sum_{i=1}^{n} 1\left(Y_{i} \geq t\right) W\left({X}_{i}\right)-\mu\left(t ; W\right)|)$ \end{center} where $\mu\left(t ; W\right)=\mathrm{E}_{{X}, Y}\{1(Y \geq t)W\left({X}_{i}\right)\}<\infty$ and $W\left({X}_{i}\right) \leq U_{f}<\infty$ and $ \tau <\infty$. Consider the class of functions indexed by $t$, \begin{center} $\mathcal{F}=\left\{1(y \geq t) W\left(x\right)/ U_{f}: t \in[0, \tau], y \in \mathbb{R}, W\left(x\right)\leq U_{f}\right\}.$ \end{center} It is crucial to evaluate $N_{[~]}\left({\varepsilon}, \mathcal{F}, L_{1}(Q)\right)$. Given $\epsilon \in (0,1),$ let $t_{i}$ be the $i$-th $\lceil 1 / \varepsilon\rceil$ quantile of $Y,$ thus $ P\left(Y \leq t_{i}\right)=i \varepsilon,~i=1, \ldots,\left\lceil\frac{1}{\varepsilon}\right\rceil-1. $ Furthermore, take $t_{0}=0$ and $t_{\lceil 1 / \varepsilon\rceil}=+\infty .$ For $i=1,2, \cdots,\lceil 1 / \varepsilon\rceil,$ define brackets $\left[L_{i}, U_{i}\right]$ with \begin{center} $L_{i}(x, y)=1\left(y \geq t_{i}\right) \frac{W\left(x\right)}{U_{f}}, \quad U_{i}(x, y)=1\left(y>t_{i-1}\right) \frac{W\left(x\right)}{U_{f}}$ \end{center} such that $L_{i}(x, y) \leq 1(y \geq t) e^{f_{\theta}(x)} / U_{f} \leq U_{i}(x, y)$ as $t_{i-1}<t \leq t_{i} .$ The Jensen's inequality gives $\mathrm{E}|U_{i}-L_{i}|\le |E[\frac{W\left(X_i\right)}{U_{f}}\left\{1\left(Y \geq t_{i}\right)-1\left(Y>t_{i-1}\right)\right\}]| \le |P\left(t_{i-1}<Y \leq t_{i}\right)|=\varepsilon.$ Therefore, $N_{[~]}\left({\varepsilon}, \mathcal{F}, L_{1}(P)\right) \leq\lceil 1 / \varepsilon\rceil <\infty \text { for every } \varepsilon>0$. So the class $\mathcal{F}$ is P-Glivenko-Cantelli. \end{example} If the upper bounds of $N_{[~]}\left(\varepsilon, \mathcal{F}, L_{2}(P)\right)$ and $\sup{}_{Q} N\left(\varepsilon, \mathcal{F}, L_{2}(Q)\right)$ have polynomial rates w.r.t. $O(1/\varepsilon)$, the following tail bound estimate gives the convergence rate of suprema of empirical processes in Lemma~\ref{NA-GC} obtained by \cite{Talagrand94}. It extends DKW inequality to general empirical processes with the bounded function classes. \begin{lemma}[Sharper bounds for suprema of empirical processes]\label{tm:Talagrand} Consider a probability space $(\Omega, \Sigma, P),$ and consider $n$ IID r.vs $\{X_{i}\}_{i=1}^n$ valued in $\Omega,$ of law $P$. Let $\mathcal{F}$ be a class of measurable functions $f: \mathcal{X} \mapsto[0,1]$ that satisfies bracketing number conditions: $$N_{[~]}\left(\varepsilon, \mathcal{F}, L_{2}(P)\right) \leq ({K}/{\varepsilon})^{V}~(\text{or}~\sup{}_{Q} N\left(\varepsilon, \mathcal{F}, L_{2}(Q)\right) \leq({K}/{\varepsilon})^{V}),~\text { for every } 0<\varepsilon<K.$$ Then, for every $t>0$ \begin{center} $P(\sqrt n {\mathop {\sup }\limits_{f \in {\cal F}} |(\mathbb{P}_{n}- \mathbb{P})f|\ge t} ) \leq\left(\frac{D(K) t}{\sqrt{V}}\right)^{V} e^{-2 t^{2}}$ with a constant $D(K)$ depending on $K$ only. \end{center} \end{lemma} The explicit constant $D(K)$ can be found in \cite{Zhang06}, who studied the tail bounds for the suprema of the unbounded and non-IID empirical process. \cite{Kong14} derived the rate of convergence for the Lasso regularized Cox models by using sharper concentration inequality for the suprema of empirical processes in Example \ref{tm:Talagrand} related to the negative log-partial likelihood function. In Example \ref{eg:Weighted}, we have \[{\left\{ {{\rm{E}}|{U_i} - {L_i}{|^2}} \right\}^{1/2}} \le {\{ {E{[\frac{{W\left( {{X_i}} \right)}}{{{U_m}}}\left\{ {1\left( {Y \ge {t_i}} \right) - 1\left( {Y > {t_{i - 1}}} \right)} \right\}]^2}} \}^{1/2}} \le |P\left( {{t_{i - 1}} < Y \le {t_i}} \right)|^{1/2} = \sqrt \varepsilon,\] which implies $N_{[~]}\left(\sqrt{\varepsilon}, \mathcal{F}, L_{2}(P)\right) \leq\lceil 1 / \varepsilon\rceil \leq 2 / \varepsilon~\text { for every } \varepsilon>0$. Hence, $N_{[~]}\left({\varepsilon}, \mathcal{F}, L_{2}(P)\right) \leq 2 / \varepsilon^2$. By applying Lemma \ref{tm:Talagrand} with $V=2,~K=\sqrt 2$, we have \begin{center} $P({\sup }_{0 \le t \le \tau } |\frac{1}{{{U_f}\sqrt n }}\sum\limits_{i = 1}^n {[1} \left( {{Y_i} \ge t} \right)W\left( {{X_i}} \right) - \mu \left( {t;W} \right)]| \ge t) \le \frac{{{D^2}(\sqrt 2)}}{2}{t^2}{e^{ - 2{t^2}}}.$ \end{center} The next two results are the \emph{symmetrization theorem} and the \emph{contraction theorem}, which are fundamental tools to get sharper bounds for suprema of empirical processes. \begin{lemma}[Symmetrization theorem]\label{tm:Symmetrization} Let $\{\textit{\textbf{X}}_{i}\}_{i=1}^n$ be independent r.vs with values in some space $\mathcal{X}$ and $\mathcal{F}$ be a class of measurable real-valued functions on $\mathcal{X}$. Let $\{\epsilon_{i}\}_{i=1}^n$ be a Rademacher sequence with uniform distribution on $\{ - 1,1\}$, independent of $\{\textit{\textbf{X}}_{i}\}_{i=1}^n$ and $f\in \mathcal{F}$. If ${\rm{E}}|f(\textit{\textbf{X}}_{i})|< \infty~\forall~i$, then \begin{center} ${\rm{E}}\{{\sup}_{f \in \mathcal{F}}\Phi ( \sum_{i=1}^{n}[ f(\textit{\textbf{X}}_{i})-{\rm{E}}f(\textit{\textbf{X}}_{i})] )\}\le {\rm{E}}\{{\sup}_{f \in \mathcal{F}}\Phi[2\sum_{i=1}^{n} \epsilon_{i}f(\textit{\textbf{X}}_{i})]\}$ \end{center} for every nondecreasing, convex $\Phi(\cdot): \mathbb{R} \mapsto \mathbb{R}$ and class of measurable functions $\mathcal{F}$. \end{lemma} \begin{lemma}[Contraction theorem]\label{lm:Contraction} Let $x_{1},...,x_{n}$ be the non-random elements of $\mathcal{X}$ and $\varepsilon_{1},...,\varepsilon_{n}$ be Rademacher sequence. Consider $c$-Lipschitz functions $g_{i}$, i.e. $\left| {{g_i}(s) - {g_i}(t)} \right| \le c \left| {s - t} \right|,\forall s,t \in \mathbb{R}$. Then for any function $f$ and $h$ in $\mathcal{F}$, we have \begin{center} ${\rm{E}}_{\epsilon}[{\sup }_{f \in {\mathcal F}} | {\sum_{i = 1}^n {{\varepsilon_i}} [ {{g_i}\{f({x_i})\} - {g_i}\{h({x_i})\}} ]}|] \le 2c {\rm{E}}_{\epsilon}[{\sup }_{f \in {\mathcal F}} | {\sum_{i = 1}^n {{\varepsilon_i}\{f({x_i}) - h({x_i})\}} } |].$ \end{center} \end{lemma} A gentle introduction to suprema of empirical processes and its statistical applications are nicely presented in \cite{Sen2018}. {\color{black}{To further bound $\mathrm{E} \{ {\sup }_{f \in {\cal F}}\sqrt n|(\mathbb{P}_{n}- \mathbb{P})f|\}$ in Lemma \ref{tm:Symmetrization} with $\Phi(t)=|t|$, Theorem 3.5.4 in \cite{Gine15} gave a constants-specified upper bound for the expectation of suprema of unbounded empirical processes. \begin{lemma}[Moment bound for suprema of unbounded empirical processes]\label{lm:ContractionEntropy} Let $\mathcal{F}$ be a countable class of measurable functions with $0 \in \mathcal{F},$ and let $F$ be a strictly positive envelope for $\mathcal{F}$. Assume that $J(\mathcal{F}, F, t):=\int_{0}^{t} \sup _{Q} \sqrt{\log [2 N(\mathcal{F}, L_{2}(Q), \tau\|F\|_{L_{2}(Q)})]} d \tau<\infty$ for some (for all) $t>0$. Given $\mathcal{X}$-valued IID r.vs $\{X_{i}\}_{i=1}^n$ with law $P$ s.t. $P F^{2}<\infty$. Set $U=\max_{1 \leq i \leq n} F\left(X_{i}\right)$ and $\delta=\sup _{f \in \mathcal{F}} \sqrt{P f^{2}} /\|F\|_{L^{2}(P)} .$ Then, for $A_{1}=8 \sqrt{6} \text { and } A_{2}=2^{15} 3^{5 / 2}$, \begin{equation}\label{eq:EEntropy} \mathrm{E} \{\sup{}_{f \in {\cal F}}\sqrt n |(\mathbb{P}_{n}-{P})f|\} \leq A_{1}\|F\|_{L^{2}(P)} J(\mathcal{F}, F, \delta)\vee[{A_{2}\|U\|_{L^{2}(P)} J^{2}(\mathcal{F}, F, \delta)}/({\sqrt{n} \delta^{2}})]. \end{equation} \end{lemma} The bound in \eqref{eq:EEntropy} matches the non-asymptotically sub-exponential CLT in \eqref{eq:subEclt}, and it reveals that $\sup{}_{f \in {\cal F}}\sqrt n |(\mathbb{P}_{n}-{P})f|$ has the sub-exponential behaviour, although with a huge parameter (the constant $A_{2}=2^{15} 3^{5 / 2}$ is so large). Recently, Theorem 2 in \cite{Baraud2020} sharpened bound \eqref{eq:EEntropy} when $\mathcal{F}$ takes values in $[-1,1]$. {\color{black}{The \emph{uniform entropy integral} $J(\mathcal{F}, F, t)$ can be evaluated by VC dimension of $\mathcal{F}$, see Theorem 7.11 in \cite{Sen2018}. Applying \cite{Adamczak2008}'s tail inequalities for the summation $Z_n:=\sup _{f \in \mathcal{F}}n|(\mathbb{P}_{n}-{P})f|$}} with unbounded $\mathcal{F}$, they obtained following result. \begin{lemma}[Tail estimates for suprema of empirical processes under sub-Weibull norms]\label{lm:Adamczak} Let $\{X_{i}\}_{i=1}^n$ be independent $\mathcal{X}$-valued r.vs and let $\mathcal{F}$ be a countable class of measurable functions $f: \mathcal{X} \rightarrow \mathbb{R}$. For some $\alpha \in(0,1]$, assume $\|\sup _{f \in \mathcal{F}} | f\left(X_{i}\right)-\mathrm{E} f\left(X_{i}\right)| \|_{\psi_{\alpha}}<\infty$ for every $i$. Define $ \sigma_n^{2}=\sup _{f \in \mathcal{F}} \sum_{i=1}^{n} \mathrm{Var} f\left(X_{i}\right). $ For all $\eta \in (0,1)$ and $\delta>0,$ then there exists a constant $C_{\alpha, \eta, \delta}>0$ s.t. both ${P}(Z_n\geq(1+\eta) \mathrm{E} Z_n+t)$ and ${P}(Z_n\leq(1-\eta) \mathrm{E}Z_n-t)$ are bounded by {\color{black}{ \begin{center} $\delta_{n,t,\eta,\delta}(\sigma_n^{2},\alpha):=\exp (-\frac{t^{2}}{2(1+\delta) \sigma_n^{2}})+3 \exp (-(\frac{t}{C_{\alpha, \eta, \delta}\|\max\limits_{i} \sup _{f \in \mathcal{F}}|f\left(X_{i}\right)-\mathrm{E} f\left(X_{i}\right)|\|_{\psi_{\alpha}}})^{\alpha})$ for all $t \geq 0$. \end{center}}} \end{lemma} So, Lemmas \ref{lm:ContractionEntropy} and \ref{lm:Adamczak} give ${P}({n}^{-1}Z_n\geq(1+\eta){n}^{-1/2}[\text{Right hand side of }\eqref{eq:EEntropy}]+t)\le {P}(Z_n\geq(1+\eta){n}^{1/2}\mathrm{E}[{n}^{-1/2} Z_n]+nt)\le \delta_{n,nt,\eta,\delta}(\sigma_n^{2},\alpha)$. We have \begin{align*} P(\sup{}_{f \in {\cal F}}|(\mathbb{P}_{n}-{P})f|\le (1+\eta){n}^{-1/2}[\text{Right hand side of }\eqref{eq:EEntropy}]+t)\ge 1-\delta_{n,nt,\eta,\delta}(\sigma_n^{2},\alpha). \end{align*} The constant-unspecific version of Lemma \ref{lm:ContractionEntropy} (Lemma 19.36-19.38 in \cite{Vaart1998} or other versions) has wide applications in deriving the rate of convergence for kernel density estimations, M-estimators in high-dimensional and increasingly-dimensional regressions; see \cite{Gine15}, \cite{Buhlmann11}, \cite{Pan2020} and references therein. }} \section{Concentration for High-dimensional Statistics} With the emergence of high-dimensional (HD) data such as the gene expression data, there are renewed interests on the CIs. One aspect of the HD data is such that the number of variables $p$ can be comparable to or even greater than the sample size $n$. This section provides results in three commonly encountered settings: \textit{increasing-dimensional} ($p_n=o(n)<n$), \textit{large-dimensional} ($p_n=O(n)$) and \textit{high-dimensional} setting ($p_n \gg n,~p_n=e^{o(n)}$). \subsection{Linear models with diverging number of covariates}\label{se;lm} Suppose that we have an $n$-dimensional random vector $\emph{\textbf{Y}}$ which contains $n$ responses $\{Y_i\}_{i=1}^n$ to $p$ covariates ${\emph{\textbf{X}}_i} = ({x_{i1}}, \cdots ,{x_{ip}})^T$, respectively. The $n$ copies of ${\emph{\textbf{X}}_i}$ as row vectors make a $n \times p$ design matrix $\textbf{X} = ({\emph{\textbf{X}}_1}, \cdots ,{\emph{\textbf{X}}_n})^T$. The conditional expectation ${\mathrm{E}[Y_i|{\emph{\textbf{X}}_i}]}$ is linearly related to a coefficient vector $\boldsymbol{\beta}^{*}= {(\beta _1^*, \cdots ,\beta _p^*)^T}$ such that \begin{equation} \label{eq:LMs} {\bm Y} = {\mathbf X}\bm\beta^* + \bm \varepsilon \end{equation} where $\{\varepsilon _i\}_{i=1}^n$ in the error vector $\boldsymbol{\varepsilon} = {({\varepsilon _1}, \cdots ,{\varepsilon _n})^T}$ are IID with zero mean and finite variance $\sigma ^2$. The $\boldsymbol{\beta}^{*}$ needs to be estimated. This subsection only considers the case that $p$ is increasing but $p<n$. The ordinary least square (OLS) estimator is \iffalse For the case $n<p \to \infty$ , it is the HD statistical inferences which would be discussed in Section \ref{linear}, \ref{Poisson} and \ref{test}. \fi \begin{equation}\label{eq:ols} {\hat {\boldsymbol{\beta}}}_{LS} = \mathop {{\rm{argmin}}}{}_{{\boldsymbol{\beta}} \in {\mathbb{R}^p}} ||\textit{\textbf{Y}} - {\textbf{X}\boldsymbol{\beta}||_2^2}. \end{equation} Assume ${\rm{rank}}(\mathbf{X})=p$, which is not hard to meet since $p < n$, $\hat {\boldsymbol{\beta}}_{LS}=(\mathbf{X}^T\mathbf{X})^{-1}\mathbf{X}^T\textit{\textbf{Y}}$ is the unique solution of the \eqref{eq:ols}. The following result for the OLS estimator {is well known.} \begin{lemma} \label{tm:basic-ols} Under the assumptions on the linear models and the rank of $\textbf{X}$ is $p$, then\\ (i) Let $\mathbf{A}$ be a $p \times n$ matrix, then ${\rm E}\|\mathbf{A}\boldsymbol{\varepsilon}\|_2^2 = {\rm E}({\boldsymbol{\varepsilon}^T}{\mathbf{A}^T}\mathbf{A}\boldsymbol{\varepsilon}) = {\sigma ^2}{\rm{tr}}({\mathbf{A}^T}\mathbf{A})$. \noindent (ii)(\textbf{\rm The curse of dimensionality}) The \emph{mean square error} and the average \emph{in-sample $\ell_{2}$ risk} of the OLS estimator are ${\rm E}\|\hat {\boldsymbol{\beta}}_{LS} - {\boldsymbol{\beta} ^*}\|_2^2 = {\rm{tr}}((\mathbf{X}^T\mathbf{X})^{-1}){\sigma ^2}~\text{and}~\frac{1}{n}{\rm E}\|\mathbf{X}(\hat {\boldsymbol{\beta}}_{LS}- {\boldsymbol{\beta} ^*})\|_2^2 =\frac{p{\sigma ^2}}{n}.$ \end{lemma} \begin{remark}As $p,n \to \infty $ with $p < n$, part (ii) implies that the OLS estimator may had poor performance if $p/n \to c>0$. The average in-sample $\ell_{2}$-risk tends to zero if $p_{n}=o(n)$. \end{remark} Put $\hat {\boldsymbol{\beta}}:=\hat {\boldsymbol{\beta}}_{LS}$. Let $\{\lambda_{i}\left(\mathbf{X}^T\mathbf{X}\right) \}_{i=1}^k$ be the eigenvalue values of $\mathbf{X}^T\mathbf{X}$. Markov's inequality and Lemma \ref{tm:basic-ols} with $\mathbf{A}=\left(\mathbf{X}^T\mathbf{X}\right)^{-1} \mathbf{X}^T$ implies \begin{align*} P\{\|\hat{\bm\beta}-{\bm\beta}^*\|_2>t\} \le \frac{\sigma^{2} \operatorname{tr}[\left(\mathbf{X}^T\mathbf{X}\right)^{-1}]}{t^{2}}=\frac{\sigma^{2}}{t^{2}} \sum_{i=1}^{p} \frac{1}{\lambda_{i}\left(\mathbf{X}^T\mathbf{X}\right)} \leq \frac{\sigma^{2}}{t^{2}} \frac{p}{\lambda_{\min }\left(\mathbf{X}^T\mathbf{X}\right)}=:{\delta _n}, \end{align*} which implies that, with probability greater than $1-{\delta _n}$, \begin{align}\label{eq:ls1e} \|\hat{\bm\beta}-{\bm\beta}^*\|_2 \le \sigma \sqrt {{p}/{n}}\cdot [{\delta _n}{\lambda _{\min }}( {{\textstyle{1 \over n}}{{\bf{X}}^T}{\bf{X}}})]^{-1/2}. \end{align} Assume that $p:=p_n=o(n^r)$ as $n \rightarrow \infty,~p_n<n$. We specific two groups of regularity conditions and the value of $r$ such that the $l_2$ consistency ($\|\hat{\bm\beta}-{\bm\beta}^*\|_2\stackrel{p}{\longrightarrow}0$) is true. (1) By Lemma \ref{tm:basic-ols}, if $\frac{1}{n}\mathbf{X}^T\mathbf{X}$ is uniformly positive definite ($\exists~ c>0$ s.t. $\frac{1}{n}\mathbf{X}^T\mathbf{X}\succ c\mathbf{I}_p$) then \begin{center} $\frac{p{\sigma ^2}}{n}=\frac{1}{n}\mathrm{E}\|\mathbf{X}(\bm{\hat{\beta}}-\bm{\beta}^*)\|_2^2=\mathrm{E}[(\bm{\hat{\beta}}-\bm{\beta}^*)^T\frac{1}{n}\mathbf{X}^T\mathbf{X}(\bm{\hat{\beta}}-\bm{\beta}^*)]\ge c\mathrm{E}\|\hat{\bm\beta}-{\bm\beta}^*\|_2^2.$ \end{center} If $p=o(n)$, then $\mathrm{E}\|\hat{\bm\beta}-{\bm\beta}^*\|_2^2=o(1)$ which implies $\|\hat{\bm\beta}-{\bm\beta}^*\|_2=o_p(1)$. (2) From \eqref{eq:ls1e}, if $p = o({\lambda _{\min }}\left( {{\mathbf{X}^{T}}\mathbf{X}} \right))$, we have $\|\hat{\bm\beta}-{\bm\beta}\|_2=o_p(1)$. In this case, if we consider: ``$\frac{1}{n}\mathbf{X}^T\mathbf{X}$ is positive definite" in (1), it also leads to $p= o({\lambda _{\min }}\left( {{\mathbf{X}^{T}}\mathbf{X}} \right))=o(n)$. In \eqref{eq:LMs} with fixed design, suppose that the $\varepsilon_{1}, \ldots, \varepsilon_{n}$ are sub-Gaussian zero-mean noise for which there exists a $\sigma>0$ such that $\mathrm{E}e^{\sum_{i=1}^{n} \alpha_{i}\varepsilon_i} \le e^{\sigma^{2} \sum_{i=1}^{n} \alpha_{i}^{2}},~\forall \alpha_{1}, \ldots, \alpha_{n} \in \mathbb{R}$. Suppose that the Gram matrix $\bm{S}_n:=\frac{1}{n}\mathbf{X}^T\mathbf{X}$ is invertible. The \textit{excess in-sample prediction error} $R(\hat{\mbox{\boldmath $\beta$}})$ is the difference between the expected squared error for $\bm X_{i}^{T}\hat{\mbox{\boldmath $\beta$}}$ and for $\bm X_{i}^{T}\mbox{\boldmath $\beta$}^*:$ \begin{align}\label{olsR} R(\hat{\mbox{\boldmath $\beta$}}):&=\frac{1}{n}\{\mathrm{E}[ \sum_{i=1}^{n}(\bm X_{i}^{T} \boldsymbol{\beta}-Y_{i})^{2}]-\mathrm{E}[ \sum_{i=1}^{n}(X_{i}^{T} \mbox{\boldmath $\beta$}^*-Y_{i})^{2}]\}|_{\boldsymbol{\beta}=\boldsymbol{\hat\beta}}\nonumber\\ &=\frac{1}{n}\{\mathrm{E}[ \sum_{i=1}^{n}(\bm X_{i}^{T}{\mbox{\boldmath $\beta$}}-\bm X_{i}^{T}\mbox{\boldmath $\beta$}^*+\bm X_{i}^{T}\mbox{\boldmath $\beta$}^*-Y_{i})^{2}]|_{\boldsymbol{\beta}=\boldsymbol{\hat\beta}}-\mathrm{E}[ \sum_{i=1}^{n}(X_{i}^{T} \mbox{\boldmath $\beta$}^*-Y_{i})^{2}]\}\nonumber\\ &=\frac{1}{n}\|\mathbf{X}(\hat {\boldsymbol{\beta}} - {\boldsymbol{\beta} ^*})\|_2^2 +\frac{1}{n}\mathrm{E}[ \sum_{i=1}^{n}(\bm X_{i}^{T}{\mbox{\boldmath $\beta$}}-\bm X_{i}^{T}\mbox{\boldmath $\beta$}^*)\cdot \varepsilon_i]|_{\boldsymbol{\beta}=\boldsymbol{\hat\beta}}=\frac{1}{n}\|\mathbf{X}(\mathbf{X}^T\mathbf{X})^{-1}\mathbf{X}^T\bm\varepsilon\|_2^2 . \end{align} which is a quadratic form of sub-Gaussian vector. By Corollary \ref{thm:zhangt} with $\mathbf{A}:=\mathbf{X}(\mathbf{X}^T\mathbf{X})^{-1}\mathbf{X}^T/\sqrt{n}$, $\mbox{\boldmath $\xi$}:=\bm\varepsilon,~\bm \mu =\bm 0$ and $\bm{\Sigma}:=\mathbf{A}^{T} \mathbf{A}=\mathbf{X}(\mathbf{X}^T\mathbf{X})^{-1}\mathbf{X}^T/n$, \[\operatorname{tr}(\bm{\Sigma})=\operatorname{tr}((\mathbf{X}^T\mathbf{X})^{-1}\mathbf{X}^T\mathbf{X})/n=p/n,~~\operatorname{tr}\left(\bm{\Sigma}^{2} \right)=p/n^2,~~\|\bm{\Sigma}\|_2=1/n. \] where last identity is due to $\mathbf{X}(\mathbf{X}^T\mathbf{X})^{-1}\mathbf{X}^T$ being a projection matrix. Thus $P[R(\hat{\mbox{\boldmath $\beta$}})>\frac{\sigma^{2}(p+2 \sqrt{p t}+2 t)}{n}] \le e^{-t}$, i.e. with probability $1-e^{-t}$, \begin{center} $R(\hat{\mbox{\boldmath $\beta$}})\le \frac{\sigma^{2}(p+2 \sqrt{p t}+2 t)}{n}.$ \end{center} For Gaussian noise, $\mathrm{E}R(\hat{\mbox{\boldmath $\beta$}})=\frac{\sigma^{2} p}{n}$ in Lemma \ref{tm:basic-ols}, so $P\{R(\hat{\mbox{\boldmath $\beta$}})-\mathrm{E}R(\hat{\mbox{\boldmath $\beta$}})\le \frac{\sigma^{2}(2 \sqrt{p t}+2 t)}{n}\}\ge 1-e^{-t}$. \subsection{Non-asymptotic Bai-Yin theorem for random matrix}\label{Bai-Yin} Let $\mathbf{A}$ be a $p \times p$ Hermitian matrix with real eigenvalues: $\lambda_{\max}:=\lambda_1 \ge \cdots \ge \lambda_p=:\lambda_{\min}$. The \emph{empirical spectral distribution} (ESD) of $\mathbf{A}$ is \begin{center} $F_{\mathbf{A}}(x)=\frac{1}{p}\mathop{\sum_{j=1}^p {\rm{1}}(\lambda_j\leq x)}, $ \end{center} which resembles the EDF of IID samples. Let $\{{\mathbf{A}}_n\}_{n \geq 1}$ be a sequence of $p \times p$ Hermitian random matrices indexed by the sample size $n$, and $F_{{\mathbf{A}}_n}$ be the ESD of ${\mathbf{A}}_n$. A major interest in random matrix theory is to investigate the convergence of $F_{{\mathbf{A}}_n}$ as a sequence of distributions to a limit $F$. Recall the $F$ is said to be the limit of $\{F_{{\mathbf{A}}_n}\}_{n \geq 1}$ if $\mathop{\lim_{n\to \infty}F_{{\mathbf{A}}_n}}(x) =F(x)~\hbox{for all}~x \in C(F), $ where $C(F)$ is the set of continuous point set of $F(x)$. In multivariate statistics, it is of interest to study the \emph{sample covariance matrix} $\mathbf{S}_{n}:=\frac{1}{n}\mathbf{X}{\mathbf{X}^T}$ where the double array $\mathbf{X}=\left\{X_{i j}, i=1, \ldots, p ; j=1, \ldots, n\right\}$ contains zero-mean IID r.vs $\{X_{i j}\}$ with variance $\sigma^{2}.$ Suppose that the dimensions $n$ and $p$ grow to infinity while $p/n$ converges to a constant in $[0,1]$. \cite{Marcenko67} gives the limit behavior of {the ESD} of $\mathbf{S}_{n}$. \cite{Bai1993} obtained a strong version of the Mar{\v{c}}enko-Pastur law. \begin{corollary}[Bai-Yin theorem]\label{Bai1993} Let $\mathbf{X}$ be an $n \times p$ random matrix whose entries are independent copies of a r.v. with zero mean, unit variance, and finite fourth moment ($\mathrm{E} | X_{11}|^{4}<\infty$). As $n \rightarrow \infty, p \rightarrow \infty, p / n \rightarrow y \in(0,1)$, then \begin{center} $\mathop{\lim_{n \to \infty}} \lambda_{\min}(\mathbf{S}_{n}) =\sigma^2(1-\sqrt y)^2,~~~~ \mathop{\lim_{n \to \infty}} \lambda_{\max}(\mathbf{S}_{n}) =\sigma^2(1+\sqrt y)^2.$ \end{center} \end{corollary} Note that $\lambda_i(\mathbf{S}_{n})= \lambda_{i}(\mathbf{X}/{\sqrt n }) $ for all $i$, Bai-Yin's law asserts that if $\sigma^2=1$: $ \lambda_{\min}(\mathbf{X}/{\sqrt n }) =1 - \sqrt {{p}/{n}} + o(\sqrt {{p}/{n}} ), ~ \lambda_{\max}(\mathbf{X}/{\sqrt n }) =1 + \sqrt {{p}/{n}} + o(\sqrt {{p}/{n}} )~ \text{a.s.}. $ Theorem 4.6.1 in \cite{Vershynin18} studies the non-asymptotic upper and lower bounds of the extreme eigenvalues of $\mathbf{S}_{n}$ with independent sub-exponential entries, but the bounds contained un-specific constants. We give a constant-specified version: \begin{proposition}[Constants-specified non-asymptotic Bai-Yin theorem] \label{NA-BY} Let $\mathbf{X}$ be an $n \times p$ matrix whose rows $\bm{X}_i$ are independent sub-Gaussian random vectors in $\mathbb{R}^p$ with $\mathrm{Var}(\bm{X}_i)=\mathbf{I}_p$. Define $Z_i:=|\langle {\boldsymbol{X}_i,\boldsymbol{x}} \rangle|,~\forall~\boldsymbol{x} \in S^{n-1}$. Further assume that $\{ {Z_i^2-1}\} _{i = 1}^n$ are $\operatorname{subE}(\theta)$, then \begin{equation}\label{P1} {P} \{ \big\| {n}^{-1}\mathbf{X}^T\mathbf{X}-\mathbf{I}_p \big\|\leq 2c\theta\max \left( \delta , \delta ^ { 2 } \right) \} \geq 1-2 e^{-ct^2},~~t \ge 0 \end{equation} where $\delta =2c(\sqrt {{p}/{n}} + {t}/{{\sqrt n }})$ with $t = c\theta\max \left( \delta , \delta ^ { 2 } \right)$ and $c \geq {2n \log 9}/{p}$. Moreover, \begin{equation}\label{P2} {P}\left\{{1-t^2} \le \lambda_{\min}(\mathbf{S}_{n}) \le \lambda_{\max}(\mathbf{S}_{n}) \le {1+t^2}\right\}\geq 1 - 2 e^{-ct^2}. \end{equation} \end{proposition} Proposition \ref{NA-BY} does not require ${p}/{n} \rightarrow y \in(0,1)$ as in Corollary \ref{Bai1993}. \begin{proof} \textbf{Step1}. We introduce a \emph{counting measure} for measuring the complexity of a set in some space. The \emph{covering number} $\mathcal{N}(K,\varepsilon )$ is the smallest number of closed balls centered at $K$ with radii $\varepsilon$ whose union covers $K$. For some $\varepsilon \in [0,1)$, a subset $\mathcal{N}_{\varepsilon} \subset \mathbb{R}$ is an \emph{$\varepsilon$ -net} for $S^{n-1}$ if for all $\boldsymbol{x} \in S^{n-1},$ there is an $\boldsymbol{y} \in \mathcal{N}_{\varepsilon},$ such that $\|\boldsymbol{x}- \boldsymbol{y}\|<\varepsilon$. We use the following results in Lemma 5.2 and 5.4 of \cite{Vershynin2010}. \begin{lemma}[Covering numbers of the sphere]\label{net cardinality} $\mathcal{N}(S^{n-1},\varepsilon) \le ( 1 + \frac{2}{\varepsilon})^n $ for every $\varepsilon >0$. \end{lemma} \begin{lemma}[Computing the spectral norm on a net] \label{norm on net general} Let $\mathbf{B}$ be an $p \times p$ matrix. Then \begin{center} $ \big\|\mathbf{B}\big\|:= \max_{||\textbf{x}||_2 =1}\big\|\mathbf{B}\boldsymbol{x}\big\|_2 =\sup_{\boldsymbol{x} \in S^{p-1}} |\langle {\mathbf{B}\boldsymbol{x},\boldsymbol{x}} \rangle| \le (1 - 2\varepsilon)^{-1} \sup_{\boldsymbol{x} \in \mathcal{N}_{\varepsilon}} |\langle {\mathbf{B}\boldsymbol{x},\boldsymbol{x}} \rangle|. $ \end{center} \end{lemma} \noindent Lemma \ref{norm on net general} shows that $ \big\| \frac{1}{n}\mathbf{X}^T\mathbf{X}-\mathbf{I}_p \big\| \le 2 \max_{\boldsymbol{x} \in \mathcal{N}_{1/4}} \big| \frac{1}{n} \|\mathbf{X}\boldsymbol{x}\|_2^2 - 1 \big|. $ Indeed, note that $ \langle \frac{1}{n} \mathbf{X}^T\mathbf{X}\boldsymbol{x} - \boldsymbol{x}, \boldsymbol{x} \rangle = \langle \frac{1}{n} \mathbf{X}^T\mathbf{X}\boldsymbol{x}, \boldsymbol{x} \rangle - 1 = \frac{1}{n} \|\mathbf{X}\boldsymbol{x}\|_2^2 - 1. $ By setting $\varepsilon = 1/4$ in Lemma~\ref{norm on net general}, we get \begin{equation}\label{eq:SW} \big\| {n}^{-1}\mathbf{X}^T\mathbf{X}-\mathbf{I}_p \big\| \leq ( 1 - 2 \varepsilon ) ^ { - 1 } \sup {}_ {\boldsymbol{x} \in \mathcal { N } _ { \varepsilon } } | \langle {n}^{-1} \mathbf{X}^T\mathbf{X}\boldsymbol{x} -\boldsymbol{x}, \boldsymbol{x} \rangle | = 2 \sup{}_ { \boldsymbol{x} \in \mathcal { N } _ { 1 / 4 } } | {n}^{-1} \| \mathbf{X} \boldsymbol{x} \| _ { 2 } ^ { 2 } - 1 |. \end{equation} By \eqref{eq:SW}, we have \begin{align}\label{eq:NNN} &{P}\{ \big\| {n}^{-1}\mathbf{X}^T\mathbf{X}-\mathbf{I}_p \big\|\geq 2t \}\leq {P} \{ 2 \sup_ { \bm{x} \in \mathcal { N } _ { 1 / 4 } }| {n}^{-1} \| \mathbf{X} \boldsymbol{x} \| _ { 2 } ^ { 2 } - 1 | \geq 2t \} \leq \sum_{ x \in \mathcal { N } _ { 1 / 4 } } {P}\{ | {n}^{-1}\| \mathbf{X} \boldsymbol{x} \| _ { 2 } ^ { 2 } - 1 | \geq t\} \nonumber\\ & \leq \mathcal { N } ( S ^ { n - 1 } , {1}/{4} ) {P} \{ | {n}^{-1} \| \mathbf{X} \boldsymbol{x} \| _ { 2 } ^ { 2 } - 1 | \geq t\}\leq 9^n {P} \{ |{n}^{-1}\| \mathbf{X} \boldsymbol{x} \| _ { 2 } ^ { 2 } - 1 | \geq t\},~~\forall~{ x \in \mathcal { N } _ { 1 / 4 } }, \end{align} where the last inequality follows Lemma~\ref{net cardinality} with $\varepsilon = 1/4$. \textbf{Step2}. It is sufficient to bound ${P} \{ | \frac { 1 } { n } \| \mathbf{X} \boldsymbol{x} \| _ { 2 } ^ { 2 } - 1 | \geq t\} $. Let $Z_i:=|\langle {\boldsymbol{X}_i,\boldsymbol{x}} \rangle|,~\forall~\boldsymbol{x} \in S^{n-1}$. Observe that $\|\mathbf{X}\boldsymbol{x}\|_2^2 = \sum_{i=1}^n |\langle {\boldsymbol{X}_i,\boldsymbol{x}} \rangle|^2 =: \sum_{i=1}^n Z_i^2.$ Apply the sub-exponential concentration inequality in Corollary \ref{sub-exponentialConcentration}, $P({n}^{-1}| \| \mathbf{X} \boldsymbol{x} \| _ { 2 } ^ { 2 } - 1 || \ge t )=P( {n}^{-1}| {\sum_{i = 1}^n {(Z_i^2 - 1)} } | \ge t )\le 2e^{ { - \frac{n}{2}( {\frac{t^2}{\theta^2} \wedge \frac{t}{\theta}})} } .$ Specially, let $t = c\theta\max \left( \delta , \delta ^ { 2 } \right) =c\theta[ \delta \mathrm{I}_{\{\delta \leq 1\}} + \delta ^ { 2 }\mathrm{I}_{\{\delta > 1\}}]$ with $\delta :=2c( {{p}/{n}} +{t}/{{\sqrt n }})$. From \eqref{eq:NNN}, \begin{align*} &~~~~ {P}\{ \big\| n^{-1}\mathbf{X}^T\mathbf{X}-\mathbf{I}_p \big\|\geq 2t \}\leq 9^n \mathrm{P}\{|{n}^{-1}\| \mathbf{X} \boldsymbol{x} \| _ { 2 } ^ { 2 } - 1| \geq c \theta\max \left( \delta , \delta ^ { 2 } \right) \}\\ & \leq 2\cdot 9^ne^{[ - \frac{cn}{2} \min\{\delta^2 \mathrm{I}_{\{\delta \leq 1\}} + \delta ^ { 4 }\mathrm{I}_{\{\delta > 1\}}, \delta \mathrm{I}_{\{\delta \leq 1\}} + \delta ^ { 2 }\mathrm{I}_{\{\delta > 1\}}\}} = 2 \cdot 9^ne^{ -\frac{cn}{2} \delta ^2} =e^{ - \frac{c}{2} (\sqrt{p} + t)^2} \leq 2 \cdot 9^ne^ {- c( p + t ^ { 2 } ) /2} \end{align*} where the last inequality is obtained by using the inequality $( a + b ) ^ { 2 } \geq a^2 + b^2 $ for $a, b \geq 0$. For $c \geq {n \log 9/p}$, $2 \cdot 9^n e^{-c( p+t^2)} \leq 2 e^{-ct^2}$, which proves \eqref{P1} . \textbf{Step3}. To show \eqref{P2}, the $\max_{||\boldsymbol{x}||_2 =1}|\|\frac{1}{\sqrt n}\mathbf{X}\boldsymbol{x}\|_2^2-1|=\max_{||\boldsymbol{x}||_2 =1}\big\|(\frac{1}{n}\mathbf{X}^T\mathbf{X}-\mathbf{I}_p)\boldsymbol{x}\big\|_2^2=\big\| \frac{1}{n}\mathbf{X}^T\mathbf{X}-\mathbf{I}_p \big\|^2\leq t^2$ implies that ${1-t^2} \le \lambda_{\max}(\mathbf{S}_{n}) \le {1+t^2}.$ Similarly, for $\lambda_{\min}(\mathbf{S}_{n})$, \begin{center} $\min_{||\boldsymbol{x}||_2 =1}|\|\frac{1}{\sqrt n}\mathbf{X}\boldsymbol{x}\|_2^2-1|=\min_{||\boldsymbol{x}||_2 =1}\big\|(\frac{1}{n}\mathbf{X}^T\mathbf{X}-\mathbf{I}_p)\boldsymbol{x}\big\|_2^2 \le \max_{||\boldsymbol{x}||_2 =1}\big\|(\frac{1}{n}\mathbf{X}^T\mathbf{X}-\mathbf{I}_p)\boldsymbol{x}\big\|_2^2\leq t^2.$ \end{center} So $\lambda_{\min}(\mathbf{S}_{n}) \in [{1-t^2},{1+t^2}]$ and $\{\| \mathbf{X}^T\mathbf{X}-\mathbf{I}_p\|^2\leq t^2\}~\subset~\left\{{1-t^2} \le \lambda_{\min}(\mathbf{S}_{n}) \le \lambda_{\max}(\mathbf{S}_{n}) \le {1+t^2}\right\}.$ Then ${P}\{{1-t^2} \le \lambda_{\min}(\mathbf{S}_{n}) \le \lambda_{\max}(\mathbf{S}_{n}) \le {1+t^2}\}\geq {P} \{\big\| \frac{1}{n}\mathbf{X}^T\mathbf{X}-\mathbf{I}_p \big\|^2\leq t^2 \} \geq 1 - 2 e^{-ct^2}.$ \end{proof} \subsection{Oracle inequalities for penalized linear models}\label{linear} This section introduces the proofs of the error bounds from the perspective of Lasso penalized linear models with the $\ell_2$-loss function. When $p>n$, the OLS estimator is no longer available as $\frac{1}{n} \sum_{i = 1}^n {{\textit{\textbf{X}}_i}{\textit{\textbf{X}}_i^T}}$ is of invertible. A common way for obtaining a plausible estimator for the true parameter $\mbox{\boldmath $\beta$}^*$ is by adding penalized function to the square loss function. For $0< q\leq\infty$, we write $\| \mbox{\boldmath $\beta$}\|_{q}:=(\sum_{i=1}^{p}|\beta_{i}|^{q})^{1/q}$ as the $\ell_{q}$-norm for $\mbox{\boldmath $\beta$} \in \mathbb{R}^p$. If $q=\infty$, $\| \mbox{\boldmath $\beta$}\|_{\infty}:=\max_{i=1,...,p}|\beta_{i}|$; if $q=0$, $\| \mbox{\boldmath $\beta$}\|_{0}:=\sum_{i = 1}^p {{\rm{1(}}{\beta _i} \ne 0{\rm{)}}} $. There are two types \textit{statistical guarantees} of $\hat \mbox{\boldmath $\beta$}$ as mentioned in \cite{Bartlett12}. \begin{enumerate} \item \textbf{Persistence}: $\hat \mbox{\boldmath $\beta$}$ performs well on a new sample ${\bm X^*}\stackrel{d}{=}{\bm X}$ (equal in distribution), i.e. ${\rm{E}}\{[{\bm X^*}( \hat \mbox{\boldmath $\beta$}- {\mbox{\boldmath $\beta$} ^*} )]^2|{\bm X^*}\}\to 0$. \item \textbf{$\ell_q$-consistency} ($q\ge 1$): $\hat \beta$ approximates $\beta^{*}$, i.e. with high probability $\|\hat \mbox{\boldmath $\beta$}- {\mbox{\boldmath $\beta$} ^*}\|_q\to 0$. \end{enumerate} The persistence and $\ell_1$-consistency are respectively obtained by error bounds: \begin{center} $\|\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}\|_1 \le {O_p}(s{\lambda _n}),~~~~{\rm{E}}\{[{\bm X^*}( \hat \mbox{\boldmath $\beta$}- {\mbox{\boldmath $\beta$} ^*} )]^2|{\bm X^*}\} \le {O_p}(s{\lambda _n^2})$,~\text{(says \textit{oracle inequalities})} \end{center} where ${\lambda _n}\to 0$ is a tuning parameter and $s:=\|\mbox{\boldmath $\beta$}^*\|_{0}$. In the following, we focus on the $\ell_1$ estimation and prediction consistencies for the penalized linear models. Let $\lambda > 0$ be a tuning parameter, the \textit{Lasso estimator} \citep{Tibshirani1996} for Model \eqref{eq:LMs} is \begin{align}\label{eq:lasso} {\boldsymbol{\hat \beta}}_L={ \rm{argmin}}{}_{\boldsymbol{\beta} \in {\mathbb{R}^p}} \{ {\| {\bm{Y} - \textbf{X}\boldsymbol{\beta} } \|_2^2}/n + \lambda {\left\| \boldsymbol{\beta} \right\|_1}\}. \end{align} By sub-derivative techniques in convex optimizations, the Karush-Kuhn-Tucker (KKT) condition of Lasso optimization function is \begin{center} $ \left\{ \begin{aligned}\label{eq:kktl} 2{[ {{{\bm X}^T}({\bm Y} - {\mathbf X}{\hat \mbox{\boldmath $\beta$}}_L)} ]_j}/n = - {\lambda } {\rm{sign}}(\hat\beta_{Lj})\, \text{ if } \hat\beta_j\neq 0,\\ 2|{[ {{{\mathbf X}^T}({\bm Y} - {\mathbf X}{\hat \mbox{\boldmath $\beta$}} )}_L ]_j}|/n \le {\lambda } \qquad\text{ if } \hat\beta_{Lj}=0 \end{aligned} \right. $. \end{center} which implies ${\| {\frac{1}{n}{{\mathbf X}^T}({\bm Y} - {\mathbf X}{\hat \mbox{\boldmath $\beta$}})} \|_\infty } \le \frac{\lambda}{2} $. Another approach to get the Lasso-like sparse estimator is attained by Dantzig selector (DS) \begin{align}\label{eq:DS} {{\hat \mbox{\boldmath $\beta$} }_{DS}} = {\arg \min }_{\boldsymbol{\beta} \in {\mathbb{R}^p}} \{ {\left\| \mbox{\boldmath $\beta$} \right\|_1}:{\| {{{\mathbf X}^T}({\bm Y} - {\mathbf X}\mbox{\boldmath $\beta$} )} \|_\infty/n } \le {\lambda}/{2}\}. \end{align} see \cite{Candes2007}. Lasso and DS are capable of producing sparse estimates with only a few (hence sparse) nonzero coefficients among the $p$ coefficients of the covariates. The idea of Lasso and DS was presented in a geophysics literature \citep{Levy81}. By \eqref{eq:DS}, we get $\|\hat{\beta}_{DS}\|_{1} \le\|\hat{\beta}_{L}\|_{1}$, which signifies that the DS may be more sparse than the Lasso. It is well-known that $\bm{\Sigma}:=\frac{1}{n} \sum_{i = 1}^n {{\textit{\textbf{X}}_i}{\textit{\textbf{X}}_i^T}}$ is singular when $p>n$. To obtain oracle inequalities for the Lasso estimator with the minimax optimal rate \citep{Ye10}, the restricted eigenvalues proposed in \cite{Bickel09} is usually needed. Let $S(\boldsymbol{\beta}^{*}): = \{ j:{\beta_j^{*}} \ne 0,~\boldsymbol{\beta}^{*} = ({\beta_1^{*}}, \cdots,{\beta_p^{*}})^T\}$ and ${s} := \left| {S(\boldsymbol{\beta}^{*})} \right|$. For any vector $\bm{b} \bm \in \mathbb{R}^p$ and any index set $H \subset \{1,2,\cdots,p\}$, define the sub-vector indexed by $H$ as $\bm{b}_H = (\cdots ,{\tilde b_j}, \cdots )^T \in {\mathbb{R}^p}$ with ${\tilde b_j}= b_j$ if $j \in H$ and $\tilde b_j = 0$ if $j \notin H$. Define the \textit{conic set} for a sparse $\boldsymbol{\beta}^{*}$ with support $S(\boldsymbol{\beta}^{*})$: \begin{equation} \label{eq:compare3} {\rm{C}}(\eta ,S(\boldsymbol{\beta}^{*}))=\{ \bm{b} \in {\mathbb{R}^p}:{\| {{\bm{b}_{{S(\boldsymbol{\beta}^{*})^c}}}} \|_1} \le \eta {\| {{\bm{b}_{S(\boldsymbol{\beta}^{*})}}} \|_1}\},~\eta>0. \end{equation} Denote the \emph{restricted eigenvalue condition} (RE) as $RE(\eta ,S(\boldsymbol{\beta}^{*}),{\bm \Sigma}) = \mathop {\inf }\limits_{0 \ne {\bm{b}} \in {\rm{C}}(\eta ,S(\boldsymbol{\beta}^{*}))} \frac{{{{({{\bm{b}}^T}{\bm \Sigma} {\bm{b}})}^{1/2}}}}{{{\left\| \bm{b} \right\|_2}}} > 0$ for any $p\times p$ matrix ${\bm \Sigma}$. In the following, we present a modified version of Theorem 7.2 in \cite{Bickel09} from Lemma 2.5 of \cite{Li17} beyond Gaussian noise. \begin{proposition}[The rate of convergence of the Lasso]\label{thm:lassoo} Suppose that $\mathbf X$ is the fixed design matrix and the error sequence $\{ {\varepsilon _i}\} _{i = 1}^n \stackrel{\rm{IID}}{\sim} N(0, \sigma^2) $ or $\{ {\varepsilon _i/\sigma}\} _{i = 1}^n \stackrel{\rm{IID}} \sim 2-$strongly log-concave distribution satisfying Lemma \ref{3.16}. Let $\{\bm X_{ (j)}\}_{j=1}^p\in {\mathbb{R}^n}$ be column vectors of $\mathbf{X}$. We assume that $\frac{1}{n}\bm X_{(j)}^T{\bm X_{(j)}} = 1$. If $\lambda = A\sigma \sqrt {{{\log p}}/{n}} $ satisfies the KKT condition for $\bm\beta^*$, \begin{equation} \label{eqn:ineq1} \|\mathbf X^{T}(\bm Y - \mathbf X \bm\beta^*)/n\|_\infty \leq \lambda/2. \end{equation} \begin{enumerate}[\rm{(}1\rm{)}] \item Then the estimated error $\bm u := \bm{\hat \beta}_L - {\bm\beta ^*}$ satisfies $\|\bm u_{S(\boldsymbol{\beta}^{*})^c}\|_1 \leq 3\|\bm u_{S(\boldsymbol{\beta}^{*})}\|_1$, i.e. $\bm u \in {\rm{C}}(3 ,S(\boldsymbol{\beta}^{*}))$. \item Suppose that $\mathbf X$ satisfies the RE condition $\gamma:={\rm{RE}}(3, S(\boldsymbol{\beta}^{*}),\frac{1}{n} \sum_{i = 1}^n {{\textit{\textbf{X}}_i}{\textit{\textbf{X}}_i^T}})>0$. We have \emph{non-asymptotic oracle inequalities} with probability greater than $1-2p^{1-\frac{A^2}{8}}$: \begin{align} &(a).\|\bm{\hat \beta}_L - \bm\beta^*\|_1 \leq \frac{{3A\sigma }}{{{\gamma ^2}}}s\sqrt {\frac{{\log p}}{n}} ;~~(b).\|\bm{\hat \beta}_L- \bm\beta^*\|_2^2 \leq \frac{{9A{\sigma ^2}}}{{{\gamma ^2}}}\frac{{s\log p}}{n}; \label{thm:event1}\\ &(c).\frac{1}{n} \|\mathbf {X}(\bm{\hat \beta}_L- \bm\beta^*)\|_2^2 \leq \frac{{9A{\sigma}}}{{{\gamma}}}\frac{{s\log p}}{n},~A>2\sqrt{2}. \label{thm:event2} \end{align} \end{enumerate} \end{proposition} \begin{proof} The proof consists of 3 steps. : \emph{1. Checking $\bm{\hat \beta}_L - {\bm\beta ^*}$ be in cone set by using definition of Lasso and KKT conditions; 2. Verifying the high probability of the KKT condition; 3. Deriving the oracle inequalities from restricted eigenvalue condition with some elementary inequalities.} \textbf{Step1}: By the Lasso optimization \eqref{eq:lasso}, \begin{align} (2n)^{-1}\|\bm Y-\mathbf X\hat{\bm\beta}_L\|_2^2+\lambda\|\hat{\bm\beta}_L\|_1 \leq (2n)^{-1}\|\bm Y-\mathbf X \bm\beta^{*}\|_2^2+\lambda\|\bm\beta^{*}\|_1. \end{align} From $\frac{\|\bm Y-\mathbf X\hat{\bm\beta}_L\|_2^2}{2n}= \frac{1}{2n}\|\mathbf X\bm\beta^{*}+\bm\varepsilon-\mathbf X\hat{\bm\beta}_L\|_2^2=\frac{1}{2n}\|\mathbf X \bm\beta^{*}-\mathbf X\hat{\bm\beta}_L\|_2^2+\frac{\|\bm\varepsilon\|_2^2}{2n}-\frac{1}{n}\bm\varepsilon^{T} \mathbf X(\hat{\bm\beta}_L-\bm\beta^{*})$ and $\frac{1}{2n}\|\bm Y-\mathbf X\beta^{*}\|_2^2=\frac{\|\bm\varepsilon\|_2^2}{2n},$ thus $\frac{1}{2n}\|\mathbf X\bm\beta^{*}-\mathbf X\hat{\bm\beta}_L\|_2^2+\frac{\|\bm\varepsilon\|_2^2}{2n}-\frac{1}{n}\bm\varepsilon^{T}\mathbf X(\hat{\bm\beta}-\bm\beta^{*})+\lambda\|\hat{\bm\beta}_L\|_1 \leq \frac{\|\bm\varepsilon\|_2^2}{2n}+\lambda\|\bm\beta^{*}\|_1$. Then, \begin{align}\label{eq:compare0} (2n)^{-1}\|\mathbf X(\hat{\bm\beta}_L-\bm\beta^{*})\|_2^2+\lambda||\hat{\bm\beta}_L\|_1 \leq {n}^{-1}\bm\varepsilon^{T}\mathbf X(\hat{\bm\beta}_L-\bm\beta^{*})+\lambda\|\bm\beta^{*}\|_1. \end{align} The \eqref{eq:compare0} is usually called the \emph{basic inequality} in the proof of Lasso oracle inequalities. The first term in the left side of inequality \eqref{eq:compare0} is the empirical prediction error, while on the right side, $ \frac{1}{n}\bm\varepsilon^{T}\mathbf X(\hat{\bm\beta}-\bm\beta^{*}) $ is random and $\lambda||\bm\beta^{*}||_1$ is still fixed and unknown. For $\frac{1}{n}\bm\varepsilon^{T}\mathbf X(\hat{\bm\beta}-\bm\beta^{*})$, if we can get a sharper upper bound and it approaching 0 as $n \to \infty$, then we can achieve a sharper oracle inequality in below. By \eqref{eqn:ineq1}, \begin{align}\label{eq:compare} \frac{\|\mathbf X(\hat{\bm\beta}-\bm\beta^{*})\|_2^2}{2n}+\lambda\|\hat{\bm\beta}\|_1\leq \|\frac{1}{n}\bm\varepsilon^{T}\mathbf X\|_{\infty}\|\hat{\bm\beta}-\bm\beta^{*}\|_1+\lambda\|\bm\beta^{*}\|_1\leq \frac{\lambda}{2}\|\hat{\bm\beta}-\bm\beta^{*}\|_1+\lambda\|\bm\beta^{*}\|_1. \end{align} \noindent Let $S:=S(\boldsymbol{\beta}^{*})$ and notes that $||\hat{\bm\beta}_{S}||_1=||\bm\beta^{*}_{S}+(\hat{\bm\beta}_{S}-\bm\beta_{S}^{*})||_1\geq||\bm\beta^{*}_{S}||_1-||\hat{\bm\beta}_{S}-\bm\beta_{S}^{*}||_1$, then \begin{align}\label{eq:triangle} ||\hat{\bm\beta}||_1=||\hat{\bm\beta}_{S^{c}}||_1+||\hat{\bm\beta}_{S}||_1\geq ||\bm\beta^{*}_{S}||_1-||\hat{\bm\beta}_{S}-\bm\beta_{S}^{*}||_1+||\hat{\bm\beta}_{S^{c}}||_1. \end{align} From \eqref{eq:compare}, we get $\|\bm u_{S^c}\|_1\leq 3\|\bm u_{S}\|_1$ be checking \begin{align}\label{eq:triangle1} &0\leq {(2n)}^{-1}||\mathbf X(\hat{\bm\beta}-\bm\beta^{*})\|_2^2\leq{\lambda}\|\hat{\bm\beta}-\bm\beta^{*}||_1/2+\lambda\|\bm\beta^{*}\|_1-\lambda\|\hat{\bm\beta}\|_1\nonumber\\ &\leq \frac{\lambda}{2}\{\|\hat{\bm\beta}_{S}-\bm\beta^{*}_{S}\|_1+\|\hat{\bm\beta}_{S^{c}}||_1\}+\lambda\|\bm\beta^{*}_{S}\|_1-\lambda\{ \|\bm\beta^{*}_{S}\|_1-\|\hat{\bm\beta}_{S}-\bm\beta_{S}^{*}\|_1+\|\hat{\bm\beta}_{S^{c}}\|_1 \}~~[\text{By}~\eqref{eq:triangle}]\nonumber\\ &=\frac{3\lambda}{2}\|\hat{\bm\beta}_{S}-\bm\beta^{*}_{S}\|_1-\frac{\lambda}{2}\|\hat{\bm\beta}_{S^{c}}\|_1=:\frac{3\lambda}{2}\|\bm u_{S}\|_1-\frac{\lambda}{2}\|\bm u_{S^{c}}\|_1. \end{align} \textbf{Step2}: The Gaussian error vector $\boldsymbol{\varepsilon}$ enables us to get the Gaussian concentration around its mean, we can shows that \eqref{eqn:ineq1} occurs with a high probability. So next we need to check the Lipschitz condition in Lemma~\ref{lem:caussiancon}. Use Lemma~\ref{lem:caussiancon}, it implies that \begin{equation}\label{eq:kktlip} P({n}^{-1}| {\bm X_{(j)}^T(\bm Y - \mathbf X{\bm\beta ^*})| \ge t}) \le 2pe^{ - {{n{t^2}}}/{{2{\sigma ^2}}}},~\forall~j. \end{equation} under the presupposition $\|{\bm X_{(j)}}\|_2^2=\bm X_{(j)}^T{\bm X_{(j)}} = n$. The Lipschitz condition depends on the design matrix $ \mathbf X $. The different types of CIs require different assumptions on the design matrix (the random design is allowed if we adopt empirical process theory). In Lemma~\ref{lem:caussiancon}, put $f(\bm a): =\frac{1}{n}|\bm X_{(j)}^T(\sigma \bm a - \mathbf X{\bm\beta ^*})|$. Then, Cauchy's inequality implies, \begin{center} $f(\bm a) - f(\bm b) \le\frac{\sigma }{n}|\bm X_{(j)}^T{{(\bm b - \bm a)}}|\le \frac{\sigma }{{n }}\|{\bm X_{(j)}}\|_2 \cdot \|\bm b -\bm a\|_2=\frac{\sigma }{{\sqrt n }}\|\bm b -\bm a\|_2~\forall~j.$ \end{center} Hence, $f(\bm a)$ is ${\sigma }/{{\sqrt n }}$-Lipschitz. Recall $\lambda = A\sigma \sqrt {{{\log p}}/{n}} $. So \eqref{eq:kktlip} implies \begin{center} $P(\|\frac{1}{n}\mathbf X^{T}(\bm Y-\mathbf X \bm\beta^{*})\|_{\infty}\geq \frac{\lambda}{2})\le \sum_{j=1}^p P( { \frac{1}{n}| {\bm X_{(j)}^T(\bm Y - \mathbf X{\bm\beta ^*})}| \ge \frac{1}{2}A\sigma\sqrt {\frac{{{\log}p}}{n}}} )\leq 2 p^{1 -\frac{A^2}{8}}.$ \end{center} By Lemma \ref{3.16}, \eqref{eq:kktlip} is also held for $\{ {\varepsilon _i/\sigma}\} _{i = 1}^n\sim$ $2$-strongly log-concave distribution. \textbf{Step3}: Next we can start on the proof based on cone set condition \eqref{eq:compare3}. Since the $\mathbf X$ satisfies RE condition $\gamma:={\rm{RE}}(3, S,{n}^{-1}\sum_{i = 1}^n {{\textit{\textbf{X}}_i}{\textit{\textbf{X}}_i^T}})>0$, by \eqref{eq:compare3} we have \begin{center} $\gamma\|\bm u\|_2^2\leq\frac{1}{n}\|\mathbf X\bm u\|_2^2\stackrel{\eqref{eq:triangle1}}{\le}\lambda(3\|{\bm u}_{S}\|_1-\|{\bm u}_{S^c}\|_1)\leq 3\lambda\|{\bm u}_S\|_1\leq 3\lambda\sqrt{s}\|{\bm u}_{S}\|_2\leq 3\lambda\sqrt{s}\|{\bm u}\|_2,$ \end{center} where the second last inequality is by Cauchy's inequality. Therefore, \begin{center} $\|\hat{\mbox{\boldmath $\beta$}}_L-\bm\beta^{*}\|_2^2=:\|{\bm u}\|_2^2\leq\frac{9\lambda^2 s}{\gamma^2}=\frac{9A^2\sigma^2}{\gamma^2}\frac{s\log p}{n},~~\frac{\|\mathbf X(\hat{\bm\beta}_L-\bm\beta^{*})\|_2^2}{n}=:\frac{\|\mathbf X\bm u\|_2^2}{n}\leq\frac{9\lambda^2s}{\gamma}=\frac{9A^2\sigma^2}{\gamma}\frac{s \log p}{n}.$ \end{center} So $\|\hat{\bm\beta}_L-\bm\beta^{*}\|_1=:\|{\bm u}\|_1\leq\sqrt{s}\|{\bm u}\|_2\leq \frac{3\lambda s}{\gamma}=\frac{3A\sigma}{\gamma}s\sqrt{\frac{\log p}{n}}$ by Cauchy's inequality. \end{proof} According to \eqref{eq:ls1e}, the OLS with diverging number of covariates has the convergence rate $O(\sqrt {{p}/{n}} )$ under the minimal eigenvalue condition ${\lambda _{\min }}\left( {{\mathbf{X}^{T}}\mathbf{X}} \right)=O(n)$. In contrast, due to the sparse restriction and the RE condition in Proposition \ref{thm:lassoo}, the factor $\sqrt {\log p} $ is much more small that the factor $\sqrt {p} $ in the convergence rate \eqref{eq:ls1e}. Under the RE condition, Proposition \ref{thm:lassoo} reveals that Lasso is $\ell_2$-consistent if $\frac{{s\log p}}{n}\to 0$, and $s\sqrt {\frac{{\log p}}{n}} \to 0$ guarantees $\ell_1$-consistency. Theorem 7.1 in \cite{Bickel09} also gives oracle inequalities \eqref{thm:event1} and \eqref{thm:event2} for the DS estimator \eqref{eq:DS}. \subsection{High-dimensional Poisson regressions with random design}\label{Poisson} The Poisson regression \citep{McCullagh83} is a model for nonnegative integers response variables, i.e. ${Y_i} \stackrel{\rm IID}{\sim} \mathrm{Poisson}(\lambda_i),$ where $\log (\lambda _i) = {\bm X_i^T}\bm\beta$ for $i=1,\cdots,n$. We presume that the $\{\bm X_i\}_{i=1}^n$ are IID r.vs on some space $\mathcal{X}$, and we observe $n$ copies of $\{ ({Y_i},{\emph{\textbf{X}}_i})\}_{i = 1}^n \sim ({Y},{\emph{\textbf{X}}})\in \mathbb{R}\times \mathbb{R}^p$. The average negative log-likelihood of Poisson regressions is $\ell_n(\mbox{\boldmath $\beta$} ): = - \frac{1}{n}\sum_{i = 1}^n {[{Y_i}{\bm X_i^T}\bm\beta - {e^{{\bm X_i^T}\bm\beta}}]}$ and the Lasso penalized estimator is \begin{equation}\label{eq:enp} \boldsymbol{\hat \beta}:= \boldsymbol{\hat \beta} ({\lambda })=\mathop {\rm{argmin}}{}_{\boldsymbol{\beta} \in {\mathbb{R}^p}} \{\ell_n(\boldsymbol{\beta}) + \lambda {{\left\|\boldsymbol{\beta}\right\|}_1}\}~\text{with a turning parameter}~{\lambda } > 0. \end{equation} Lemma 4.2 in \cite{Buhlmann11} shows the first-order conditions for the optimization in \eqref{eq:enp}. \begin{lemma}[Necessary and sufficient condition]\label{lem:iff} Let $j \in \{ 1,2, \cdots ,p\} $ and ${\lambda } > 0$. Then, a necessary and sufficient condition for the Lasso estimates \eqref{eq:enp} is \begin{eqnarray}\label{eq:kkt} \left\{ \begin{aligned} {n^{-1}\sum{}_{i = 1}^n {{{{X_{ij}}({Y_i} - {e^{{\textit{\textbf{X}}_i^T} \boldsymbol{\hat \beta} }})}}} }= - {\lambda } {\rm{sign}}(\hat\beta_j) \quad\, \text{ if } \hat\beta_j\neq 0,\\ |{n^{-1}\sum{}_{i = 1}^n {{{{X_{ij}}({Y_i} - {e^{{\textit{\textbf{X}}_i^T} \boldsymbol{\hat \beta} }})}}} }| \le {\lambda} \qquad\quad\text{ if } \hat\beta_j=0. \end{aligned} \right. \end{eqnarray} \end{lemma} Let $ l(Y,{\emph{\textbf{X}}},{{\mbox{\boldmath $\beta$} }})=- Y{\bm{X}^T}\mbox{\boldmath $\beta$}+e^{{\textit{\textbf{X}}^T{\mbox{\boldmath $\beta$} }}}$ be the Poisson loss function. The true coefficient ${{{\mbox{\boldmath $\beta$} }} ^{*}}$ is the minimizer of the expected Poisson loss, i.e. \begin{equation}\label{eq:oracle} {\mbox{\boldmath $\beta$} } ^{*} = \argmin{}_{\boldsymbol{\beta} \in {{\mathbb{R}}^{p}}} { \mathrm{{E}}} l(Y,{\emph{\textbf{X}}},{{\mbox{\boldmath $\beta$} }}). \end{equation} The KKT condition of the $\ell _{1}$-penalized likelihood is for the estimated parameter. But, here we use the true parameter version of the KKT conditions: $|{\frac{1}{n}\sum_{i = 1}^n {{{{X_{ij}}({Y_i} - {\rm{E}}{Y_i})}}} }| \le {\lambda},~~j=1,...,p$ by replacing ${e^{{\textit{\textbf{X}}_i^T} \boldsymbol{\hat \beta} }}$ by ${\rm{E}}{Y_i}=e^{{\textit{\textbf{X}}_i^T} {\mbox{\boldmath $\beta$} } ^{*}}$ to approximate the estimated version \eqref{eq:kkt}. To motivate the next two propositions concerning high-probability events, let us consider the following notations and the decomposition of empirical process. The Poisson loss $l(\mbox{\boldmath $\beta$} ,\bm{X},Y) = {l_1}(\mbox{\boldmath $\beta$} ,\bm{X},Y) + {l_2}(\mbox{\boldmath $\beta$} ,\bm{X})$ is decomposed into two parts where ${l_1}(\mbox{\boldmath $\beta$} ): = {l_1}(\mbox{\boldmath $\beta$} ,\bm{X},Y) := - Y{\bm{X}^T}\mbox{\boldmath $\beta$} $ and ${l_2}(\mbox{\boldmath $\beta$}): = {l_2}(\mbox{\boldmath $\beta$},\textit{\textbf{X}}) :=e^{{\textit{\textbf{X}}^T}\mbox{\boldmath $\beta$} } $ is free of response. Let $\mathbb{P}l({{\mbox{\boldmath $\beta$} }}) := {\rm E}l(\mbox{\boldmath $\beta$},\bm{X},Y)$ be the expected {loss}. We are {interested in} the centralized empirical loss $\left( \mathbb{P}_{n}-\mathbb{P}\right) l(\mbox{\boldmath $\beta$})$ representing fluctuations between the expected and empirical losses. Note that \begin{equation}\label{eq:EPP} \left( \mathbb{P}_{n}-\mathbb{P}\right) l(\mbox{\boldmath $\beta$})=\left( \mathbb{P}_{n}-\mathbb{P}\right) l_{1}(\mbox{\boldmath $\beta$})+\left( \mathbb{P}_{n}-\mathbb{P}\right) l_{2}(\mbox{\boldmath $\beta$}), \end{equation} which is crucial in attaining the convergence rate of $\|\bm{\hat \beta} - \bm\beta^*\|_1$. Motivated from rate of convergence theorem [Theorem 3.2.5 of \cite{van96}] for M-estimation with functional parameter in some metric space, we study the upper bounds (or the rate) for the first and second part of the difference of the centralized empirical process between $\mbox{\boldmath $\beta$}^{*}$ and $\hat{\mbox{\boldmath $\beta$}}$: $(\mathbb{P}_{n}-\mathbb{P})( l_{m}(\mbox{\boldmath $\beta$}^{*})-l_{m}(\hat{\mbox{\boldmath $\beta$}}))$, for $m=1,2$. \begin{proposition}[Convergence rate of $(\mathbb{P}_{n}-\mathbb{P})( l_{1}(\mbox{\boldmath $\beta$}^{*})-l_{1}(\hat{\mbox{\boldmath $\beta$}}))$]\label{prop:upbound1} Suppose that \begin{equation}\label{eq:H1} \mathop {\sup }{}_{1 \le i \le \infty} {\| \textit{\textbf{X}}_i \|_\infty } \le L <\infty~\text{\rm a.s.}~\text{and}~\|\mbox{\boldmath $\beta$}^* \|_1 \le B. \end{equation} In the event of ${\cal A} := \bigcap_{j = 1}^p {\{ {| {\frac{1}{n}\sum_{i = 1}^n {{{{X_{ij}}({Y_i} - {\rm{E}}{Y_i})}}} } |\le \frac{{{\lambda }}}{4}} \}}, $ we have \begin{equation}\label{eq:L1} (\mathbb{P}_{n}-\mathbb{P})( l_{1}(\mbox{\boldmath $\beta$}^{*})-l_{1}(\hat{\mbox{\boldmath $\beta$}}))\le \frac{{{\lambda }}}{4} {\|{{\hat \mbox{\boldmath $\beta$} }} - \mbox{\boldmath $\beta$} ^*\|_1}. \end{equation} If $\lambda \ge \max \{ \frac{{16{A^2}L\log (2p)}}{{3n}},8AL{e^{LB/2}}\sqrt {\frac{{\log (2p)}}{n}} \} $ with $A>1$, we have $P(\mathcal{A}) \ge 1- {(2p)^{1 - {A^2}}}$. \end{proposition} \begin{proof} Note that, on the event ${\cal A}$ \begin{align*} (\mathbb{P}_{n}-\mathbb{P}) ( l_{1}(\mbox{\boldmath $\beta$}^{*})-l_{1}(\hat{\mbox{\boldmath $\beta$}})) &=\frac{{ - 1}}{n}\sum\limits_{i = 1}^n {({Y_i}} - {\rm{E}}{Y_i})\textit{\textbf{X}}_i^T({\mbox{\boldmath $\beta$} ^*} - \hat \mbox{\boldmath $\beta$} ) = \sum\limits_{j = 1}^p {({{\hat \beta }_j} - \beta _j^*)}{\frac{1}{n}\sum\limits_{i = 1}^n {{{{X_{ij}}({Y_i} - {\rm{E}}{Y_i})}}} } \\ & \le \sum\limits_{j = 1}^p {|{{\hat \beta }_j} - \beta _j^*|}\cdot |{\frac{1}{n}\sum\limits_{i = 1}^n {{{{X_{ij}}({Y_i} - {\rm{E}}{Y_i})}}} }| \stackrel{\mathcal{A}}{\le} \frac{{{\lambda }}}{4} {\|{{\hat \mbox{\boldmath $\beta$} }} - \mbox{\boldmath $\beta$} ^*\|_1}. \end{align*} Next, we show that $\mathcal{A}$ is a high probability event if the ${\lambda}$ is well chosen. For $j=1,...,p$ and $i=1,...,n$, $P({{\mathcal{A}}^c}) \le \sum\limits_{j = 1}^p P \{ {|{\frac{1}{n}\sum\limits_{i = 1}^n {{{{X_{ij}}({Y_i} - {\rm{E}}{Y_i})}}} } | > \frac{{{\lambda }}}{4}}\}.$ Given $\textbf{X}$, $\{S_{nj}(Y,X):={\frac{1}{n}{{{{X_{ij}}({Y_i} - {\rm{E}}{Y_i})}}} }\}_{i = 1}^n$ are conditional independent for each $j=1,...,p$. Thus Corollary~\ref{col:Poisson} with $w_i={X_{ij}}/n$ gives \begin{equation}\label{eq:majo-proba-A-bis} \begin{aligned} {P}(|S_{nj}(Y,X)|\geq t|\textbf{X})&\leq 2 \exp\{-{\frac {nt^{2}/2}{{\frac{1}{n}\sum\limits_{i = 1}^n e^{{\textit{\textbf{X}}_i^T} {\mbox{\boldmath $\beta$} } ^{*}}{\mathop {\max }\limits_{1 \le i \le n} X_{ij}^2}}+ {\mathop {\max }\limits_{1 \le i \le n} \frac{|X_{ij}|t}{3}}}}\} \le 2({e^{\frac{{ - n{t^2}}}{{4{L^2}{e^{LB}}}}}} \vee {e^{\frac{{ - 3nt}}{{4L}}}}) \end{aligned} \end{equation} where the last inequality is from ${{\mathop{\rm e}\nolimits} ^{ - \frac{a}{{b + c}}}} \le {{\mathop{\rm e}\nolimits} ^{\frac{{ - a}}{{2b}}}}\vee{{\mathop{\rm e}\nolimits} ^{\frac{{ - a}}{{2c}}}}$ for any positive numbers $a, b$ and $c$. Let $t=\frac{\lambda}{4}$. Assumptions \eqref{eq:H1} and \eqref{eq:majo-proba-A-bis} give for $j=1,...,p$ \begin{center} ${P}(|{\frac{1}{n}\sum\limits_{i = 1}^n {{{{X_{ij}}({Y_i} - {\rm{E}}{Y_i})}}} } |\geq \frac{\lambda}{4} )=\mathrm{E}{P}(|{\frac{1}{n}\sum\limits_{i = 1}^n {{{{X_{ij}}({Y_i} - {\rm{E}}{Y_i})}}} } |\geq \frac{\lambda}{4}|\textbf{X})\le 2\max \{ {e ^{\frac{{ - n\lambda^2}}{64L^2e^{LB}}}},{e ^{\frac{{ - 3n\lambda}}{16L}}}\},$ \end{center} which implies that $P({{\mathcal{A}}^c}) \le 2p\max \{ {e ^{\frac{{ - n\lambda^2}}{64L^2e^{LB}}}},{e ^{\frac{{ - 3n\lambda}}{16L}}}\} .$ Finally, if $\lambda \ge \max \{ \frac{{16{A^2}L\log (2p)}}{{3n}},8AL{e^{LB/2}}\sqrt {\frac{{\log (2p)}}{n}} \}~(A>1)$, so $P({{\mathcal{A}}^c}) \le {(2p)^{1 - {A^2}}}.$ \end{proof} Next, we provide a crucial lemma to bound ${{({\mathbb{P}_n} - \mathbb{P})\left( {{l_2}({\boldsymbol{\beta} ^*}) - {l_2}(\boldsymbol{\beta} )} \right)}}$. Let ${\nu _n}(\boldsymbol{\beta} ,{\boldsymbol{\beta} ^*}): = \frac{{({\mathbb{P}_n} - \mathbb{P})\left( {{l_2}({\boldsymbol{\beta} ^*}) - {l_2}(\boldsymbol{\beta} )} \right)}}{{ {\|{\boldsymbol{\beta}} - \boldsymbol{\beta} ^*\|_1} }}$ the normalized empirical process indexed by $\boldsymbol{\beta}$. Denote the $\ell_1$-ball by ${{\cal S}_{M}}(\mbox{\boldmath $\beta$} ^*):= \left\{ {\mbox{\boldmath $\beta$} \in {\mathbb{R}^p}:{\|{\mbox{\boldmath $\beta$} } - \mbox{\boldmath $\beta$} ^*\|_1} \le {M}<\infty} \right\}$, we define the \emph{local stochastic Lipschitz constant}: \begin{center} ${Z_M}(\mbox{\boldmath $\beta$}^*):={{\rm{sup}}}_{\boldsymbol{\beta}\in {{\cal S}_M}(\boldsymbol{\beta}^*)} |{\nu _n}(\mbox{\boldmath $\beta$} ,{\mbox{\boldmath $\beta$} ^*})|~\text{and a random event}~{\cal B} := \{ {Z_M}(\mbox{\boldmath $\beta$}^*)\le {{{{\lambda _1}}}}/{4}\}$. \end{center} It is easy to see $| {{\nu _n}( \hat \mbox{\boldmath $\beta$} ,{\mbox{\boldmath $\beta$} ^*})} | \le \mathop {\sup }_{{{\cal S}_M}(\boldsymbol{\beta}^*)} | {{\nu _n}( \hat \mbox{\boldmath $\beta$} ,{\mbox{\boldmath $\beta$} ^*})} | \le \frac{{{\lambda_1}}}{4}$, which gives $| {({\mathbb{P}_n} -\mathbb{P})({l_2}({\hat \mbox{\boldmath $\beta$} }) - {l_2}(\mbox{\boldmath $\beta$}^* ))} | \le \frac{{{\lambda_1}}}{4}{\|{ \hat \mbox{\boldmath $\beta$} } -\mbox{\boldmath $\beta$} ^*\|_1}$, provided that $\hat\mbox{\boldmath $\beta$} \in {{\cal S}_{M}}(\mbox{\boldmath $\beta$} ^*)$. Then we have follow result. \begin{proposition}[Convergence rate of $(\mathbb{P}_{n}-\mathbb{P})( l_{2}(\mbox{\boldmath $\beta$}^{*})-l_{2}(\hat{\mbox{\boldmath $\beta$}}))$]\label{lem:upbound2} Assume that there exists a large constant ${M}$ such that $\hat\mbox{\boldmath $\beta$}$ is in the $\ell_1$-ball ${{\cal S}_{M}}(\mbox{\boldmath $\beta$} ^*)$. Under assumption \eqref{eq:H1}, we have \begin{equation} P({Z_M}(\mbox{\boldmath $\beta$}^*) \ge {5AL e^{LB} }\sqrt {\frac{{\log 2p}}{n}} ) \le {(2p)^{ - {A^2}}}. \end{equation} If $\lambda \ge 20AL{e^{LB}}\sqrt{\frac{{2\log 2p}}{n}}$, we get $P\{| {({\mathbb{P}_n} -\mathbb{P})({l_2}({\hat \mbox{\boldmath $\beta$} }) - {l_2}(\mbox{\boldmath $\beta$}^* ))} | \le \frac{{{\lambda}}}{4}{\rm{(}} {\|{ \hat \mbox{\boldmath $\beta$} } -\mbox{\boldmath $\beta$} ^*\|_1} )\} \ge 1-{(2p)^{ - {A^2}}}.$ \end{proposition} \begin{proof} In the first step, we apply following McDiarmid's inequality to ${Z_M}(\mbox{\boldmath $\beta$}^*)$ by showing that ${Z_M}(\mbox{\boldmath $\beta$}^*)$ is fluctuated of no more than $\frac{{2e^{LB}}}{{n}}$. Let us check it. Put $\mathbb{P}_{n}:=\frac{1}{n}\sum_{j=1}^{n}1_{\textit{\textbf{X}}_{j},Y_{j}}$ and $\mathbb{P}_{n}^{'}:=\frac{1}{n}\sum_{j=1, j\neq i}^{n}1_{\textit{\textbf{X}}_{j},Y_{j}}+1_{\textit{\textbf{X}}_{i}^{'},Y_{i}^{'}},$ where $({\textit{\textbf{X}}_{i}^{'},Y_{i}^{'}})$ is the independent copy of $({\textit{\textbf{X}}_{j},Y_{j}})$. Let $\textit{\textbf{X}}_{i}^{T}\tilde{\mbox{\boldmath $\beta$}}_{i}$ ($\textit{\textbf{X}}{'}_{i}^{T}\tilde{\mbox{\boldmath $\beta$}}_{i}$) be an intermediate point between $\textit{\textbf{X}}_{i}^{T}\mbox{\boldmath $\beta$}$ ($\textit{\textbf{X}}{'}_{i}^{T}\mbox{\boldmath $\beta$}$) and $\textit{\textbf{X}}_{i}^{T}\mbox{\boldmath $\beta$}^*$ ($\textit{\textbf{X}}{'}_{i}^{T}\mbox{\boldmath $\beta$}^*$) from the Taylor's expansion of function $F(x):=e^{x}$. It deduces \begin{align*} & ~~~~\mathop {\sup }\limits_{\boldsymbol{ \beta} \in {S_M}}\frac{{|({\mathbb{P}_n} - \mathbb{P})({l_2}({\mbox{\boldmath $\beta$} ^*})- {l_2}(\mbox{\boldmath $\beta$} ))| }}{{\|{\mbox{\boldmath $\beta$} ^*} - \mbox{\boldmath $\beta$} \|_1 }}-\mathop {\sup }\limits_{\boldsymbol{ \beta} \in {S_M}}\frac{| ({\mathbb{P}{'}_n} - \mathbb{P})({l_2}({\mbox{\boldmath $\beta$} ^*}) - {l_2}(\hat \mbox{\boldmath $\beta$} ))| }{{\|{\mbox{\boldmath $\beta$} ^*} - \mbox{\boldmath $\beta$} \|_1 }} \\ & \le \mathop {\sup }\limits_{\boldsymbol{ \beta} \in {S_M}}\frac{|{{l_2}({\mbox{\boldmath $\beta$} ^*},{\textit{\textbf{X}}_i}) - {l_2}(\mbox{\boldmath $\beta$} ,{\textit{\textbf{X}}_i}) - {l_2}({\mbox{\boldmath $\beta$} ^*},{\textit{\textbf{X}}{'}_i}) + {l_2}(\mbox{\boldmath $\beta$} ,{\textit{\textbf{X}}{'}_i})}|}{{n\|{\mbox{\boldmath $\beta$} ^*} - \mbox{\boldmath $\beta$} \|_1}}\\ & \le \mathop {\sup }\limits_{\boldsymbol{ \beta} \in {S_M}}\frac{1}{n}{e^{\textit{\textbf{X}}_i^T\tilde \mbox{\boldmath $\beta$} }} \cdot \frac{{|\textit{\textbf{X}}_i^T{\mbox{\boldmath $\beta$} ^*} - \textit{\textbf{X}}_i^T \mbox{\boldmath $\beta$}| }}{{\|{\mbox{\boldmath $\beta$} ^*} - \mbox{\boldmath $\beta$} \|_1 }} +\mathop {\sup }\limits_{\boldsymbol{ \beta} \in {S_M}}\frac{1}{n}{e^{\textit{\textbf{X}}_i^T\tilde \mbox{\boldmath $\beta$} }} \cdot \frac{{|\textit{\textbf{X}}{'}_i^T{\mbox{\boldmath $\beta$} ^*} - \textit{\textbf{X}}{'}_i^T \mbox{\boldmath $\beta$} |}}{{\|{\mbox{\boldmath $\beta$} ^*} - \mbox{\boldmath $\beta$} \|_1 }}\le \mathop {\sup }\limits_{\boldsymbol{ \beta} \in {S_M}}\frac{{2Le^{LB}}}{n}\frac{{\|{\mbox{\boldmath $\beta$} ^*} - \mbox{\boldmath $\beta$} \|_1}}{{\|{\mbox{\boldmath $\beta$} ^*} - \mbox{\boldmath $\beta$} \|_1 }} = \frac{{2Le^{LB}}}{{n}}. \end{align*} where the first inequality stems from $ \left| {f(x)} \right| - \mathop {\sup }\limits_x \left| {g(x)} \right| \le \left| {f(x) - g(x)} \right|$ (and take suprema over $x$ again). Apply McDiarmid's inequality to ${Z_M}(\mbox{\boldmath $\beta$}^*)$, we have $P({Z_M}(\mbox{\boldmath $\beta$}^*) - {\rm{E}}{Z_M}(\mbox{\boldmath $\beta$}^*) \ge {\lambda}) \le e^{ - \frac{{{{n}}{{\lambda}^2}}}{2L^2e^{2LB}}} $. Let ${(2p)^{ - {A^2}}}=\exp \{ - \frac{{{{n}}{{\lambda}^2}}}{2L^2e^{2LB}}\}$, we get $ {\lambda} \ge ALe^{LB}\sqrt {\frac{{2\log (2p)}}{n}} $ for $A>0$, therefore \begin{equation} P({Z_M}(\mbox{\boldmath $\beta$}^*) - {\rm{E}}{Z_M}(\mbox{\boldmath $\beta$}^*) \ge {\lambda }) \le {(2p)^{ - {A^2}}}. \label{proba1-gpl} \end{equation} The next step is to estimate the sharper upper bounds of ${\rm{E}}{Z_M}(\mbox{\boldmath $\beta$}^*)$ by Lemma \ref{tm:Symmetrization} with $\Phi(t)=|t|$ and Lemma \ref{lm:Contraction}. Note that ${({\mathbb{P}_n} - \mathbb{P})\left\{{l_2}({\mbox{\boldmath $\beta$} ^*}) - {l_2}( \mbox{\boldmath $\beta$} )\right\}}={{\mathbb{P}_n}\left\{{l_2}({\mbox{\boldmath $\beta$} ^*}) - {l_2}(\mbox{\boldmath $\beta$} )\right\}}-{\rm{E}}{\left\{{l_2}({\mbox{\boldmath $\beta$}^*}) - {l_2}( \mbox{\boldmath $\beta$} )\right\}}$, by symmetrization theorem, the expected terms is canceled. To see contraction theorem, for ${Z_M}(\mbox{\boldmath $\beta$}^*) = \mathop {\sup }\limits_{\boldsymbol{ \beta} \in {S_M}} \left\{ \frac{1}{{n\|{\boldsymbol{ \beta} ^*} - \boldsymbol{\beta} \|_1 }}{|\sum\limits_{i = 1}^n (e^{{\textit{\textbf{X}}_i^T} {\mbox{\boldmath $\beta$} } ^{*}}-e^{{\textit{\textbf{X}}_i^T} {\mbox{\boldmath $\beta$} } })-n{\rm{E}}{[{l_2}({\mbox{\boldmath $\beta$} ^*}) - {l_2}( \mbox{\boldmath $\beta$} )]}|}\right\}$, it is required to check the Lipschitz property of $g_i$ in Lemma~\ref{lm:Contraction} with ${\mathcal F}=\mathbb{R}^p$. Let $f({x_i})= {{x_i^T} \boldsymbol{\beta}}/{\|{\boldsymbol{ \beta} ^*} - \boldsymbol{\beta} \|_1 },~h({x_i})= {{x_i^T} \boldsymbol{\beta}}^*/{\|{\boldsymbol{ \beta} ^*} - \boldsymbol{\beta} \|_1 }$ and ${g_i}(t) =\frac{{e^{{t{\|{\boldsymbol{ \beta} ^*} - \boldsymbol{\beta} \|_1 } }}}}{n\|{\boldsymbol{ \beta} ^*} - \boldsymbol{\beta} \|_1 }~(|t|\le LB/{\|{\boldsymbol{ \beta} ^*} - \boldsymbol{\beta} \|_1 })$. Then the function ${g_i}(t)$ here is $\frac{e^{LB}}{n }$-Lipschitz. In fact \begin{center} $\left| {{g_i}(s) - {g_i}(t)} \right| =\frac{e^{\tilde t}}{n } \cdot |s-t|\le \frac{e^{LB}}{n}|s-t|,~t,s\in [-LB/{\|{\boldsymbol{ \beta} ^*} - \boldsymbol{\beta} \|_1 },LB/{\|{\boldsymbol{ \beta} ^*} - \boldsymbol{\beta} \|_1 }]$ \end{center} where $\tilde t \in [-LB/{\|{\boldsymbol{ \beta} ^*} - \boldsymbol{\beta} \|_1 },LB/{\|{\boldsymbol{ \beta} ^*} - \boldsymbol{\beta} \|_1 }]$ is an intermediate point between $t$ and $s$ given by applying Lagrange mean value theorem. The symmetrization theorem and the contraction theorem imply \begin{align*} {\rm{E}}{Z_{M}(\mbox{\boldmath $\beta$}^*)} & \le \frac{4e^{LB} }{n} {\rm{E}} (\underset{\beta \in \mathcal{S}_{M}}{\sup}\left\lvert\sum_{i=1}^{n} \frac{\epsilon_{i}\textit{\textbf{X}}_{i}^{T}({\mbox{\boldmath $\beta$}^{*}}-\mbox{\boldmath $\beta$})}{{\|\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}\|_1 }} \right\rvert) \le \frac{4 e^{LB} }{n} {\rm{E}}(\underset{\beta \in \mathcal{S}_{M}}{\sup}\mathop {\max }\limits_{1 \le j \le p}|\sum_{i=1}^{n} {\epsilon_{i}\textit{X}_{ij}}|\cdot \frac{\|\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}\|_1 }{{\|\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}\|_1 }}) \\ & \le \frac{4 e^{LB} }{n}{\rm{E}}( \mathop {\max }\limits_{1 \le j \le p}|\sum_{i=1}^{n} {\epsilon_{i}\textit{X}_{ij}} |)= \frac{4 e^{LB} }{n}{\rm{E}}({\rm{E}}[ \mathop {\max }\limits_{1 \le j \le p}|\sum_{i=1}^{n} {\epsilon_{i}\textit{X}_{ij}} ||\textbf{X}]). \end{align*} From Corollary~\ref{pp-Maximalbd}, with ${\rm{E}}_{\epsilon}[{\epsilon_{i}\textit{X}_{ij}}|\textbf{X}]=0$ we get $\frac{4 e^{LB} }{n}{\rm{E}}({\rm{E}}[ \mathop {\max }\limits_{1 \le j \le p}|\sum_{i=1}^{n} {\epsilon_{i}\textit{X}_{ij}} ||\textbf{X}])\le \frac{4 e^{LB} }{n}\sqrt {2\log 2p} \cdot \sqrt {n{L^2}} ={4 e^{LB} }L\sqrt {\frac{{2\log 2p}}{n}} .$ Thus, for $A\ge 1$, \begin{equation}\label{proba2-gpl} {\rm{E}}{Z_M}(\mbox{\boldmath $\beta$}^*) \le{4 e^{LB} L}\sqrt {\frac{{2\log 2p}}{n}} \le {4AL e^{LB} }\sqrt {\frac{{2\log 2p}}{n}}. \end{equation} With ${\lambda} \ge ALe^{LB}\sqrt {\frac{{2\log (2p)}}{n}} $ and (\ref{proba2-gpl}), we conclude from (\ref{proba1-gpl}) that $P({Z_M}(\mbox{\boldmath $\beta$}^*) \ge {5AL e^{LB} }\sqrt {\frac{{\log 2p}}{n}} ) \le P({Z_M}(\mbox{\boldmath $\beta$}^*) \ge {\lambda} + {\rm{E}}{Z_M}(\mbox{\boldmath $\beta$}^*)) \le {(2p)^{ - {A^2}}}.$ Finally, we complete the proof of Proposition \ref{lem:upbound2} by letting $\frac{{{\lambda}}}{4} \ge {5AL e^{LB} }\sqrt {\frac{{2\log 2p}}{n}}$ and setting $\mbox{\boldmath $\beta$} = \hat \mbox{\boldmath $\beta$} \in {Z_M}(\mbox{\boldmath $\beta$}^*)$. \end{proof} Let $S:={S(\boldsymbol{\beta}^{*})}$ for $\boldsymbol{\beta}^{*}$ defined in \eqref{eq:oracle} and $s:=\vert S\vert$. To obtain sharp oracle inequalities for Lasso penalized Poisson regression, we consider the following regularity conditions: \begin{itemize} \item [\textbullet] (H.1): The covariate $\bm X$ is almost surely bounded $\parallel \bm X\parallel_{\infty} \le L$ a.s. for $L>0$; \item [\textbullet] (H.2): There exists a constant $B>0$ such that $\Vert \mbox{\boldmath $\beta$}^*\Vert_{1}\le B$; \item [\textbullet] (H.3): (Stabil Condition) For $\Sigma:=\mathrm{E}(\bm X \bm X^{T})$, there exsist a $k \in (0,1)$ such that \begin{center} $\delta ^{T}\Sigma \delta\geqslant k\sum_{j\in S} \delta_{j}^{2}$ for any $\delta \in C(c_{0},S):=\lbrace \delta \in \mathbb{R}^{p}: \sum_{j \in {S}^{c} }\vert \delta_{j}\vert\le c_{0}\sum_{j \in S }\vert \delta_{j}\vert \rbrace $. \end{center} \end{itemize} The \emph{Stabil Condition} (H.3) is denoted as $S(c_{0},S,k, \Sigma)$ which is a similar version of the RE condition in the Lasso linear models proposed in \cite{Bunea08}. Due to the random variance, Poisson regression is more complex than the linear model with the constant variance assumption. {Thus,} (H.1) and (H.2) are stronger {than those assumed }for the linear models. Based on the high-probability event $\mathcal{A}$ and $\mathcal{B}$, we have the oracle inequalities for estimation and prediction for Lasso estimator $\boldsymbol{\hat \beta}$ in \eqref{eq:enp} for the Poisson regressions. \begin{theorem}\label{le-gp-glm} Assume conditions $(H.1)-(H.3)$ hold. Let $\lambda$ be chosen such that \begin{equation}\label{lambda} \lambda \ge \max \{ \frac{{16{A^2}L\log (2p)}}{{3n}},8AL{e^{LB/2}}\sqrt {\frac{{\log (2p)}}{n}}, 20AL{e^{LB}}\sqrt {\frac{{2\log 2p}}{n}}\}~\text{for}~A>\sqrt{2}. \end{equation} Suppose that we have a new covariate vector $\bm X^*$ (as the test data) which is an independent copy of $ X$ (as the training data), and ${\rm{E^*}}$ represents the expectation w.r.t. $\bm X^*$ only, then \begin{center} $P({\rm{E^*}}{[{\bm X^{*}}( \hat \mbox{\boldmath $\beta$}- {\mbox{\boldmath $\beta$} ^*} )]^2}\le\dfrac{{12e^{10LB}}}{k} s\lambda^2)~\text{and}~P(\Vert \hat{\beta}-\beta^{*}\Vert_{1}\le \dfrac{4{e^{5LB}}}{k} s\lambda) \ge 1 - {(2p)^{1 - {A^2}}}-{(2p)^{ - {A^2}/2}}.$ \end{center} \end{theorem} The Theorem \ref{le-gp-glm} leads to the persistence and $\ell_1$-consistency if $\max\{s\lambda,s\lambda^2\}\to 0$. \begin{proof} The proof consists of three steps. The techniques are adapted from \cite{Zhang17}, \cite{Huang2020} and references therein. {\bf{Step1: Check $\hat{\mbox{\boldmath $\beta$}}-\mbox{\boldmath $\beta$}^{*} \in C(3,S)$}}. From the definition of the Lasso estimates $\hat{\mbox{\boldmath $\beta$}}$ (see \eqref{eq:enp}), \begin{equation}\label{eq:def} {\mathbb{P}_n}l(\hat\mbox{\boldmath $\beta$}) + {\lambda }||\hat\mbox{\boldmath $\beta$} ||_1 \le {\mathbb{P}_n}l(\mbox{\boldmath $\beta$}^{*}) + \lambda ||{\mbox{\boldmath $\beta$}^*}||_1. \end{equation} By adding $\mathbb{P}(l(\hat\mbox{\boldmath $\beta$})- l(\mbox{\boldmath $\beta$}^{*}))+\frac{\lambda}{2} ||\hat\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$}^*}||_1$ to both sides of (\ref{eq:def}), we have \begin{equation*}\label{eq:def-putting} \mathbb{P}(l(\hat\mbox{\boldmath $\beta$} )- l(\mbox{\boldmath $\beta$}^{*}))+ \dfrac{\lambda}{2}\|\hat\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$}^*}\|_1 \le ({\mathbb{P}_n} - \mathbb{P})(l({\mbox{\boldmath $\beta$}^*})-l(\hat\mbox{\boldmath $\beta$})) +\dfrac{\lambda}{2}\|\hat\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$}^*}\|_1 + \lambda (||{\mbox{\boldmath $\beta$}^*} \|_1- ||\hat\mbox{\boldmath $\beta$}\|_1), \end{equation*} which leads \begin{align}\label{eq:lambda1} \mathbb{P}(l(\hat\mbox{\boldmath $\beta$})- l(\mbox{\boldmath $\beta$}^{*}))+ \frac{\lambda}{2} ||\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}||_1 & \le ({\mathbb{P}_n} - \mathbb{P})(l({\mbox{\boldmath $\beta$} ^*})-l(\hat\mbox{\boldmath $\beta$} )) +\frac{\lambda}{2} ||\hat\mbox{\boldmath $\beta$} - \mbox{\boldmath $\beta$}^*||_1+ \lambda (\|{\mbox{\boldmath $\beta$} ^*} \|_1 - \|\hat\mbox{\boldmath $\beta$}\|_1) \nonumber\\ & \le {\lambda} \|\hat\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$}^*}\|_1+ \lambda (\|{\mbox{\boldmath $\beta$} ^*} \|_1 - \|\hat\mbox{\boldmath $\beta$}\|_1). \end{align} By the definition of $\beta^{*}$, $\mathbb{P}(l(\hat \mbox{\boldmath $\beta$})- l(\mbox{\boldmath $\beta$}^{*}))\ge 0$. The above inequality and the fact: $|{{\hat \beta }_j} - \beta _j^*|+|\beta _j^*| - |{{\hat \beta }_j}|=0$ for $j\notin S$ and $|{{\hat \beta }_j}| - |\beta _j^*| \le |{{\hat \beta }_j} - \beta _j^*|$ for $j\in S$ lead to \begin{align}\label{eq-WC1} {\lambda }\|\hat\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}\|_1/2 & \le \lambda \|\hat\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}\|_1+ \lambda (\|\mbox{\boldmath $\beta$} ^* \|_1 -\|\hat\mbox{\boldmath $\beta$}\|_1) \le 2 \lambda \|(\hat\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})_{S}\|_1. \end{align} Thus, $ \frac{\lambda }{2} ||(\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})_{S^c}||_1 \le 1.5 \lambda ||(\hat \mbox{\boldmath $\beta$}- {\mbox{\boldmath $\beta$}^*})_{S}||_1 $ and then $\hat{\mbox{\boldmath $\beta$}}-\mbox{\boldmath $\beta$}^{*} \in C(3,S)$. {\bf{Step2: Choosing ${\lambda }$}}. Since $\mathbb{P}(l(\hat\mbox{\boldmath $\beta$})- l(\mbox{\boldmath $\beta$}^{*}))\ge 0$, \eqref{eq:lambda1} implies \begin{equation}\label{eq:lambda2} \begin{aligned} {\lambda } \|\hat\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$}^*}\|_1/2 & \le \lambda\|\hat\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$}^*}\|_1 + \lambda (\|\mbox{\boldmath $\beta$}^* \|_1 - \|\hat\mbox{\boldmath $\beta$}\|_1)\\ & \le \lambda \|\hat\mbox{\boldmath $\beta$} \|_1 + \lambda \|{\mbox{\boldmath $\beta$} ^*}\|_1 + \lambda (\|{\mbox{\boldmath $\beta$} ^*} \|_1 - \|\hat\mbox{\boldmath $\beta$}\|_1) ={\rm{2}}\lambda\|{\mbox{\boldmath $\beta$}^*}\|_1 . \end{aligned} \end{equation} Thus (H.2) implies $||\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$}^*}|{|_1} \le {{{\rm{4}}B}} .$ After having shown Propositions \ref{prop:upbound1} and \ref{lem:upbound2}, we need the result on the high probability of the event $\mathcal{A}\bigcap\mathcal{B}$, whose proof is skipped. \begin{proposition}\label{prop:upbound3} Under the event $\mathcal{A}\bigcap\mathcal{B}$ with (H.1)-(H.3), we have $\hat\mbox{\boldmath $\beta$} \in {{\cal S}_{4B}}(\mbox{\boldmath $\beta$} ^*)$. And if $\lambda$ are chosen as \eqref{lambda}, then $P({\cal A} \cap {\cal B}) \ge 1 - {(2p)^{1 - {A^2}}}-{(2p)^{ - {A^2}/2}}.$ \end{proposition} {\bf{Step3: Error bounds from Stabil Condition}}. As ${\bm X^*}$ is an independent copy of $\bm{X}$, \begin{align*} &~~~~\mathbb{P}\{l(\hat \mbox{\boldmath $\beta$} ) - l({\mbox{\boldmath $\beta$} ^*})\}= {{\rm{E}}^{\rm{*}}}[{\rm{E}}\{l(\boldsymbol{\hat\beta}) - l({\mbox{\boldmath $\beta$} ^*})|{\bm X^{*}}\}]:={{\rm{E}}^{\rm{*}}}\{{\rm{E}} { {[ - Y{\textit{\textbf{X}}^{*T}}(\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}) +e^{{\textit{\textbf{X}}^{*T}} {\boldsymbol{\beta} }}-e^{{\textit{\textbf{X}}^T} \boldsymbol{\beta}^* }]|{\bm X^{*}}} }\}|_{\boldsymbol{\beta}=\boldsymbol{\hat\beta}}\\ &= { {{{\rm{E}}^{\rm{*}}}\{ {{\rm{E}}[ - Y|{\bm X^{*}}]{\bm X^{*T}}(\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}) +(e^{{\textit{\textbf{X}}^{*T}} {\boldsymbol{\hat\beta} }}-e^{{\textit{\textbf{X}}^T} \boldsymbol{\beta}^* })]|{\bm X^{*}} }\}}|_{\boldsymbol{\beta}=\boldsymbol{\hat\beta}}},~~({{\rm{E}}^{\rm{*}}}[Y|{\bm X^{*}}] ={{{e^{{\bm X^{*T}}{\boldsymbol{\beta}^*}}}}})\\ &= {{\rm{E}}^{\rm{*}}}{{\{ {-{{{e^{{\bm X^{*T}}{\boldsymbol{\beta}^*}}}}} + {{{e^{{\bm X^{*T}}{\boldsymbol{\beta}^*}}}}} + {2}^{-1}{{{e^{{\bm X^{*T}} \boldsymbol{\tilde\beta} }}{{[{\bm X^{*T}}(\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})]}^2}}}} \}} |_{\boldsymbol{\beta}=\boldsymbol{\hat\beta}}}={2}^{-1}{{\rm{E}}^{\rm{*}}}{{\{ {{e^{{\bm X^{*T}} \boldsymbol{\tilde\beta} }}{{[{\bm X^{*T}}(\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})]}^2}} \}} |_{\boldsymbol{\beta}=\boldsymbol{\hat\beta}}}, \end{align*} where ${{\textit{\textbf{X}}^{*T}} {\tilde\mbox{\boldmath $\beta$} } }=(1-t){\textit{\textbf{X}}^{*T}}{\mbox{\boldmath $\beta$}} ^*+t{\textit{\textbf{X}}^{*T}}{\hat\beta}$ is an intermediate point of ${\textit{\textbf{X}}^{*T}}$ and ${\textit{\textbf{X}}^{*T}}{\hat\beta}$ with $t\in [0,1]$. Note that $\Vert \mbox{\boldmath $\beta$}^*\Vert_{1}\le B$ by (H.1) and $||\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$}^*}||_1\le {{{\rm{4}}B}}$, (H.2) yields \begin{equation*}\label{eq:bd} |{{\textit{\textbf{X}}^{*T}} {\tilde\mbox{\boldmath $\beta$} } }| \le t|{\textit{\textbf{X}}^{*T}}{\hat\beta}-{\textit{\textbf{X}}^{*T}}{\mbox{\boldmath $\beta$}} ^*| + |{\textit{\textbf{X}}^{*T}}{\mbox{\boldmath $\beta$}} ^*| \le ||{\textit{\textbf{X}}^{*}}||_{\infty}\cdot ||\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$}^*}||_1+ |{\textit{\textbf{X}}^{*T}}{\mbox{\boldmath $\beta$}} ^*|\le 4LB+LB=5LB, \end{equation*} which implies for $c :={e^{-5LB}}/2$ \begin{align}\label{equation-Steinwart} \mathbb{P}\{l(\hat \beta ) - l({\beta ^*})\}\ge {\inf }_{\left| t \right| \le 5LB} {2}^{-1}{{\rm{E}}^{\rm{*}}}{{\{ {{e^{{\bm X^{*T}} \boldsymbol{\tilde\beta} }}{{[{\bm X^{*T}}(\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})]}^2}} \}} |_{\boldsymbol{\beta}=\boldsymbol{\hat\beta}}}=: c{\rm{E^*}}[{\bm X^{*T}}(\hat\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})]^2. \end{align} As ${\rm{E^*}}({\bm X^{*}}{\bm X^{*T}})=\boldsymbol{\Sigma}$, ${\rm{E^*}}{[{\bm X^{*}}( \hat \mbox{\boldmath $\beta$}- {\mbox{\boldmath $\beta$} ^*} )]^2} = {(\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})}\boldsymbol{\Sigma} (\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}).$ Having checked the cone condition $C(3,S)$, we apply the Stabil Condition \begin{equation}\label{eq-Stabil} c{(\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})}\boldsymbol{\Sigma} (\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}) \ge ck||(\hat\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})_{S}||_2^2 . \end{equation} From (\ref{eq:lambda1}), (\ref{eq-WC1}) and (\ref{equation-Steinwart}), we get \begin{equation}\label{eq-WC2} c{\rm{E^*}}{[{\bm X^{*}}( \hat \mbox{\boldmath $\beta$}- {\mbox{\boldmath $\beta$} ^*} )]^2} + \frac{\lambda }{2} ||\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}||_1 \le \mathbb{P}(l(\hat \mbox{\boldmath $\beta$} )- l(\mbox{\boldmath $\beta$}^{*}))+ \frac{\lambda }{2} ||\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}||_1 \le 2 \lambda ||(\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})_{S}||_1, \end{equation} which gives $ck||(\hat\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})_{S}||_2^2 +\frac{\lambda }{2} ||\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}||_1 \le 2 \lambda ||(\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})_{S}||_1$ by plugging \eqref{eq-Stabil} into \eqref{eq-WC2}. Then, employing Cauchy's inequality, we have \begin{align}\label{equation-CS} 2ck||(\hat\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})_{S}||_2^2 +\lambda ||\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}||_1 \le 4\lambda (s \cdot \|{{(\hat\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})}_S}\|_2^2)^{1/2}\le 4t{\lambda ^2}s + {\textstyle{1 \over t}}||(\hat\mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})_{S}||_2^2, \end{align} where the last inequality is from the elementary inequality $2xy \le tx^{2}+y^{2}/t$ for all $t>0$. Let us set $t = {(2ck)^{ - 1}}$ in (\ref{equation-CS}), thus $\|\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}\|_1 \le 4t\lambda s = \frac{{2\lambda s}}{{ck}}=\frac{{4e^{5LB}}}{k} s\lambda .$ To derive the oracle inequality of prediction error, from \eqref{eq-WC2}, we obtain \begin{center} $c{\rm{E^*}}{[{\bm X^{*}}( \hat \mbox{\boldmath $\beta$}- {\mbox{\boldmath $\beta$} ^*} )]^2} \le 1.5 \lambda ||(\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})_{S}||_1 \le 1.5 \lambda ||(\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})||_1$ \end{center} which implies ${\rm{E^*}}{[{\bm X^{*}}( \hat \mbox{\boldmath $\beta$}- {\mbox{\boldmath $\beta$} ^*} )]^2}\le 1.5 \lambda ||(\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*})||_1/c \le \frac{{3s\lambda^2}}{{c^2k}}=\frac{{12e^{10LB}}}{k} s\lambda^2$, where the last inequality is from $\|\hat \mbox{\boldmath $\beta$} - {\mbox{\boldmath $\beta$} ^*}\|_1 \le \frac{{4e^{5LB}}}{k} s\lambda.$ \end{proof} For general losses beyond linear models, the crucial techniques in the non-asymptotical analysis of increasing-dimensional and high-dimensional regressions, which are \textit{Bahadur representation's for the M-estimator} \citep{Kuchibhotla18D,Pan2020} and \textit{concentration for Lipschitz loss functions} \citep{Buhlmann11,Zhang17}, respectively. In large-dimensional regressions with $p/n \to c$, the \textit{theory of random matrix} \citep{Yao15}, \textit{leave-one-out analysis} \citep{Leil18,Karoui2013} and \textit{approximate message passing} \citep{Karoui2013,Donoho2016,Karoui2018} play important roles for obtaining asymptotical results. \section{Extensions} The review has been focused on the sum of independent r.vs in the Euclidean space. However, independence structure may not be suitable for some applications, for instance, econometrics, survival analysis, and graphical models. At the same time, the Euclidean valued r.vs may not be appropriate for functional data and image data. {In the following we point out results in settings not covered to broaden this review. } By CIs for the martingales, oracle inequalities have been proposed for Lasso penalized Cox models, see \cite{Huang13}. Some statistical models, such as the Ising model involving Markov's chains. \cite{Miasojedow2018} applied Hoeffding's inequality for Markov's chains to deal with this difficulty, see \cite{Fanj18} for a review. In time series analysis, \cite{Xie2018} studies the square-root Lasso method for HD linear models with $\alpha$, $\rho$, $\phi$-mixing or $m$-dependent errors. The Hoeffding's and Bernstein's CIs for weakly dependent summations can be found in \cite{Bosq1998}. Via sub-Weibull concentrations under $\beta$-mixing, non-asymptotic inequalities for estimation errors, and the prediction errors are obtained by \cite{Wong17} for the Lasso-regularized sparse VAR model with sub-Weibull innovations. U-Statistic is another dependent sum, and Example \ref{eg:u} provides a concentration result by McDiarmid's inequality. \cite{Borovskikh1996} introduces the concentration for the Banach-valued U-statistics. In non-parametric regressions, the corresponding score functions may be r.vs in Banach (or Hilbert) space; see the monographs \cite{Ledoux91}, \cite{Yurinsky95} for introductions. Exponential tail bounds for Banach- or Hilbert-valued r.vs are indispensable for deriving sharp oracle inequalities of the error bounds, see \cite{ZhangT05}, \cite{Lei20}. Recently, Banach-valued CIs are applied to conceive non-asymptotic hypothesis testing for non-parametric regressions, see \cite{Yang20}. To extend the empirical covariance matrices from finite to infinite dimension, the sample covariance operator is treated as a random element in Banach spaces. The concentrations of empirical covariance operator also have been raised attention in kernel principal components analysis, and functional data analysis, see \cite{Rosasco10}, \cite{Bunea15}. Testing hypotheses on the regression coefficients are a necessity in measuring the effects of covariates on the certain response variables. Scientists are interested in testing the significance of a large number of covariates simultaneously. From this backgrounds, \cite{Zhong11} proposed simultaneous tests for coefficients in HD linear models under the ``large $p$, small $n$'' situations by U-statistics motivated by \cite{Chen2010}. However, their HD tests are asymptotical without a non-asymptotic guarantee. Motivated by \cite{Arlot10}, \cite{ZhuBardic18} invents a new methodology for testing the linearity hypothesis in HD linear models, and the test they proposed does not impose any restriction of model sparsity. Based on the concentration of Lipschitz functions of Gaussian distributions or strongly log-concave distribution, \cite{Zhu18} developed a new concentration-based test in HD regressions. Recently, \cite{Wang20} studied non-asymptotical two-sample testing using \emph{Projected Wasserstein Distance}, via McDiarmid's inequality. {\color{black}{In future, it would be essential and practical to study the estimator for the sub-exponential, sub-Gaussian, sub-Weibull and GBO norms as the unknown parameters when constructing non-asymptotical and data-driven confidence intervals.}}\\ \noindent{\bf Acknowledgments}\\ ~~~\\ The authors thank Haoyu Wei, Xiang Li, Xiaoyu Lei, Qiuping Wang, Yanpeng Li, Chang Cui, Shengming Zhong, and Han Run for comments on the early versions of this paper. {\color{black}{The authors also thank two referees for helpful suggestions.}} This research is funded by National Natural Science Foundation of China Grants 92046021, 12071013, 12026607 and 71973005, and LMEQF at Peking University.

https://arxiv.org/abs/1405.2379

A Probabilistic Approach to Generalized Zeckendorf Decompositions

Generalized Zeckendorf decompositions are expansions of integers as sums of elements of solutions to recurrence relations. The simplest cases are base-$b$ expansions, and the standard Zeckendorf decomposition uses the Fibonacci sequence. The expansions are finite sequences of nonnegative integer coefficients (satisfying certain technical conditions to guarantee uniqueness of the decomposition) and which can be viewed as analogs of sequences of variable-length words made from some fixed alphabet. In this paper we present a new approach and construction for uniform measures on expansions, identifying them as the distribution of a Markov chain conditioned not to hit a set. This gives a unified approach that allows us to easily recover results on the expansions from analogous results for Markov chains, and in this paper we focus on laws of large numbers, central limit theorems for sums of digits, and statements on gaps (zeros) in expansions. We expect the approach to prove useful in other similar contexts.

\section{Introduction} There are many ways to generalize base $B$ expansions of integers. One interesting choice is the Zeckendorf decomposition. If we define the Fibonacci numbers $\{F_n\}$ by $F_1 = 1$, $F_2 = 2$ and $F_{n+2} = F_{n+1} + F_n$, then every integer can be written uniquely as a sum of non-adjacent Fibonacci numbers. This is known as Zeckendorf's Theorem \cite{Ze}. For integers $m \in [F_n, F_{n+1})$, using a continued fraction approach Lekkerkerker \cite{Lek} proved that the average number of summands is $n/(\varphi^2 + 1)$, with $\varphi = \frac{1+\sqrt{5}}2$ the golden mean. The precise probabilistic meaning of ``average" is the expectation with respect to the uniform measure on the decompositions of integers in $[F_n,F_{n+1})$, and then Zeckendorf's theorem provides an asymptotic statement on a certain statistic under the sequence of uniform probability measures on decompositions of length $n$, as $n\to\infty$. Analogues hold for more general recurrences, such as linear recurrences with non-negative coefficients \cite{Al,BCCSW,Day,GT,Ha,Ho,Ke,Len,MW1,MW2}, generalizations where additionally the summands are allowed to be signed \cite{DDKMU,MW1}, and $f$-decompositions (given a function $f:\mathbb{N} \to \mathbb{N}$, if $a_n$ is in the decomposition then we do not have $a_{n-1}, \dots, a_{n-f(n)}$ in the decomposition) \cite{DDKMMU}. The notion of a legal decomposition below generalizes the non-adjacency condition. \begin{defi} \label{def:lin_recur} Given a \textbf{length} $L\in \mathbb{N}$ and \textbf{coefficients} $c_1,\dots,c_L\in \ensuremath{\mathbb{Z}}_+$ with $c_1 C_L>0$, the corresponding \textbf{positive linear recursion} is a sequence $1=G_1,G_2,\dots\in \mathbb{N}$ satisfying: \begin{align} G_{n+1} &\ = \ c_1 G_n + c_2 G_{n-1} + \cdots + c_n G_{1}+1,~n=1,\dots, L.\nonumber\\ G_{n+1} &\ = \ \sum_{j=1}^L c_j G_{n+1 - j},~ n=L,L+1,\dots. \end{align} \end{defi} \begin{defi} \label{def:legal} Given a positive linear recursion with coefficients $c_1,\dots,c_L$, an integer $N$ has a \textbf{legal} decomposition of length $n\in \mathbb{N}$ if there exist $a_1\in \mathbb{N}, a_2,\dots,a_n \in \ensuremath{\mathbb{Z}}_+$, such that \be\label{eq:decomposition} N= \sum_{i=1}^{n} {a_i G_{n+1-i}},\ee and \begin{itemize} \item $n<L$ and $a_i=c_i$ for $1\le i \le n$; or \item there exists some $s\in\{1,\dots, L\}$ such that \be \label{eq:legalcondition2} \left. \begin{array}{l} a_1\ = \ c_1,\ a_2\ = \ c_2,\ \dots,\ a_{s-1}\ = \ c_{s-1}, \mbox{ and } a_s<c_s, \\ a_{s+1}, \dots, a_{s+\ell} = 0\mbox{ for some }\ell \ge 0,\\ \{b_i\}_{i=1}^{n-s-\ell}\mbox{ with }b_i = a_{s+\ell+i}, \mbox{ is either legal or empty.} \end{array}\right\} \ee \end{itemize} \end{defi} \begin{thm}[Generalized Zeckendorf Decomposition] \label{th:legal} Consider a positive linear recurrence with coefficients $c_1,\dots,c_L$. Then every $N\in\mathbb{N}$ has a unique legal decomposition. \end{thm} In the sequel we will fix a linear recurrence as in Definition \ref{def:lin_recur}. From Theorem \ref{th:legal} it follows that there's a one-to-one correspondence between the set of integers in $[G_n,G_{n+1})$ through \eqref{eq:decomposition}, where the integer $N$ is mapped to its legal decomposition $(a_1(N),\dots,a_n(N))$. Let $Q_n$ denote the uniform distribution on the legal decompositions of integers in $[G_n,G_{n+1})$, and with this identification it is natural to consider $N$ and $a_1(N),\dots,a_n(N)$ as random variables. In what follows, we denote expectation with respect to $Q_n$ by $E^{Q_n}$.\\ For $N \in [G_n,G_{n+1})$, \eqref{eq:decomposition} can be rewritten as \be\label{eq:kN_def} N \ = \ G_{i_1(N)} + G_{i_2(N)} + \cdots + G_{i_{k(N)}},\ee where $1\le i_1 \le \dots \le i_{k(N)}\le n$. The random variable $k(N)$ gives the number of summands in the generalized Zeckendorf decomposition, and was the main object of previous works. The first result was Lekkerkerker's theorem on the asymptotic expectation of $k(N)$ when $G_n= F_n$. Here is its generalization to our setting. \begin{thm}[Generalized Lekkerkerker's Theorem] \label{th:gen_lek} There exist constants $C_{{\rm Lek}} > 0$ and $d$ such that \be E^{Q_n} k(N) =C_{{\rm Lek}}n + d + o(1),\mbox{ as }n\to\infty.\ee \end{thm} Once the average number of summands has been determined, it is natural to investigate other and finer properties of the decompositions. Three natural questions concern the fluctuations in the number of summands $k(N)$ about the mean, the distribution of gaps $i_{j+1}(N) - i_j(N),~j=1,\dots,k(N)-1$ between adjacent summands, and the length of the longest gap in a decomposition. For positive linear recurrences as in Theorem \ref{th:legal}, the distribution of the number of summands converges to a Gaussian with computable mean and variance, both of order $n$. There is an extensive literature on these result. See \cite{DG,FGNPT,GTNP,LT,Ste1} for an analysis using techniques from ergodic theory and number theory, and \cite{KKMW,MW1,MW2} for proofs via a combinatorial perspective. As before, all these are statements on the asymptotic behavior of certain statistics of generalized Zeckendorf decompositions of integers in $[G_n,G_{n+1})$ under the uniform measure, as $n\to\infty$. \\ Results on the distribution of gaps between adjacent summands have recently been obtained by Beckwith, Bower, Gaudet, Insoft, Li, Miller and Tosteson \cite{BBGILMT, BILMT}. They show that the distribution of gaps larger than the recurrence length converges to that of a geometric random variable whose parameter is the largest eigenvalue of the characteristic polynomial of the recurrence relation. For gaps smaller than the recurrence relation closed forms exist for special recurrences, though with enough work explicit formulas can be derived for any given relation. They also determine the distribution of the longest gap, and prove the behavior is similar to that of the length of the longest run of heads in a sequence of tosses of a possibly biased coin. Their proofs are a mix of combinatorics and a careful analysis of polynomials associated with the recurrence relations. The details become involved as some of the associated polynomials depend on the interval $[G_n, G_{n+1})$ under consideration. \\ The main goal of this paper is to obtain a Markov-chain description of the sequence $Q_n$ rather than focusing on statements on asymptotic behavior of some explicit statistics. We will show that there exists a finite-state Markov chain such the distribution of the first $n$ steps, conditioned to avoid a certain fixed subset of states is equal to $Q_n$. This type of process is also known in the literature as the Doob transform or the $h$-transform of the Markov chain. We will show how to adapt classical results on finite-state Markov chains to our model, and will apply this to obtain a new proof for previously derived results mentioned above, as well as some new results. The focus in this paper is on laws of large numbers, central limit theorem and gap distribution (which could be viewed as renewal theory), but the message is that essentially any result on finite state Markov chains could be adapted to this new situation, and this includes large deviations, law of iterated logarithm, functional central limit theorem, etc. \\ The paper is organized as follows. In Section \ref{sec:prob} we describe the Markovian model, and how to obtain large-time asymptotics for our model from that of the underlying Markov chain. In Section \ref{sec:additive} we present the results additive functionals in a setting which includes our particular model, first by introducing the theoretical results in Section \label{sec:additive_theory} and then applying them to the generalized decompositions in Section \ref{sec:application_zeckendorf}. These results include sharp estimates on expectation, a law of large numbers and a central limit theorem. In Section \ref{sec:gaps} we then apply the results on additive functionals to the generalized Zeckendorf decompositions. Our treatment includes all additive functionals, and the number of summands, extensively studied in the past is just one example. As another example, consider the number of times a certain subset of digits (values attained by $a_s$) appears, and the number of $s$ for which $a_s<c_s$, or even (pushing the ideas further), the number of gaps of a certain size. In Section \ref{sec:gaps} we treat the gap distribution as a consequence of the regenerative structure of the underlying Markov chain, and the analogy with Bernoulli trials. \section{Probabilistic Approach}\label{sec:prob} We remind that throughout the discussion we assume that $L\in\mathbb{N}$ and the coefficients $c_1,\dots,c_L \in \ensuremath{\mathbb{Z}}_+$ satisfy $c_1 c_L>0$ as in Definition \ref{def:lin_recur}. \\ The main idea is to show that for a given $n\in \ensuremath{\mathbb{Z}}_+$, the uniform distribution on generalized Zeckendorf decompositions consisting of $n+1$ digits (that is, the $(n+1)$-th digit is non-vanishing and all higher digits are not present) coincides with the distribution of a certain conditioned Markov chain. This provides a unified framework for the model, which, in particular, gives rather easy access to many asymptotic results. We first define the Markov chain. Let $(X,Y) = \bigl((X_n,Y_n) : n\in\ensuremath{\mathbb{Z}}_+\bigr)$ be the two-dimensional process with $X_n \in \{0,\dots ,\max_i c_i\}$ and $Y_n \in \{1,\dots ,L\}$. The idea is that $X_0,X_1,\dots$ will be used to represent the coefficients $a_i$ in \eqref{eq:decomposition}, while $Y_0,Y_1,\dots$ will be used to keep track whether the $X_n$'s satisfy the condition \eqref{eq:legalcondition2}. This will be explained below, after we finish describing our construction. Let $P$ denote the distribution under which this is an IID process, $(X_0,Y_0)$ being uniformly distributed over $\{0,\dots,\max_i c_i\}\times \{1,\dots,L\}$. \begin{defi} \label{def:legal_real} Suppose $L\in \mathbb{N}$ and $c_1,\dots,c_L\in \ensuremath{\mathbb{Z}}_+,~c_1c_L>0$ are the coefficients of a linear recursion. We say that the realization $\bigl((X_0,Y_0),(X_1,Y_1), \dots\bigr)$ of the process $(X,Y)$ is \textbf{legal} with respect to the recursion if \begin{enumerate} \item $X_0>0$ and $Y_0=1$. \item There exists a random variable $J \in \ensuremath{\mathbb{Z}}_+$ such that $X_J>0$, $X_n =0$ and $Y_n=1$ for $n> J$. \item For all $n\in \mathbb{N}$, either \begin{enumerate} \item $X_n <c_{Y_n}$ and $Y_{n+1}=1$. \item $X_n= c_{Y_n}$ and $Y_n=Y_{n+1}+1$. \end{enumerate} \end{enumerate} \end{defi} Note that condition 3b and the assumption that $Y_n \in\{1,\dots, L\}$ for all $n$ implicitly mean that in a legal realization $X_n = c_{Y_n}$ only if $Y_n < L$. \\ The main observation is the following. Given a legal realization and letting (compare to \eqref{eq:decomposition}) \be\label{eq:decomp2} N= \sum_{j=0}^n X_j G_{n-j+1}, \ee then $(X_0,\dots,X_n)$ is the legal decomposition of $N \in [G_{n+1},G_{n+2})$, according to Definition \ref{def:legal}. Let \begin{equation}\tau \ = \ \inf \{n \in \ensuremath{\mathbb{Z}}_+: \left ( (X_0,Y_0) , (X_1,Y_1), \dots,(X_n,Y_n)\right) \mbox{ does not extend to a legal realization}\}.\end{equation} With a slight abuse of notation, let $Q_n$ be the probability measure on the $\sigma$-algebra generated by $(X_0,Y_0),\dots,(X_n,Y_n)$ defined through \begin{equation} Q_n (B) \ = \ P( B | \tau >n).\end{equation} Since $P$ is uniform, $Q_n$ is uniform over all finite realizations $(X_0,Y_0),\dots,(X_n,Y_n)$ that extend to legal realizations. Any such finite realization corresponds to a unique Zeckendorf decomposition of length $n+1$ given in \eqref{eq:decomp2}. Conversely, every integer with Zeckendorf decomposition of length $n+1$ corresponds to a unique finite realization $(X_0,Y_0),\dots,(X_n,Y_n)$ extending to a legal realization. Therefore $Q_n$ could be identified with the uniform distribution on generalized Zeckendorf decompositions of length $n+1$. \\ We now define an auxiliary process that allows us to introduce ideas on conditioned Markov chains. The reason for doing that is the following: $\tau$ is not a hitting or even stopping time for $(X,Y)$, as in order to determine whether $\tau=n$, it is evident from Definition \ref{def:legal_real} 3b. that on certain circumstances the value of $Y_{n+1}$ is needed. Therefore, the probabilistic analysis of Markov chains through stopping times, and which is key to our approach, cannot be applied. To fix this, let $Z_n = (X_n,Y_n,Y_{n+1})$, and let $Z=(Z_n:n\in\ensuremath{\mathbb{Z}}_+)$. Below we will write $Z_n (1)$ for $X_n$, $Z_n(2)$ for $Y_n$ and $Z_n(3)$ for $Y_{n+1}$. It is easy to see that $\tau$ is a hitting time for $Z$. Specifically, letting \begin{align} {\mathcal L} &\ = \ \{ (x,j,j'): (x < c_j \mbox{ and }j'=1) \mbox{ or } (j<L \mbox { and } x=c_j \mbox{ and } j'=j+1) \};\nonumber\\ \label{eq:L_0} {\mathcal L}_0&\ =\ {\mathcal L} \cap \{ (x,1,j'):x>0\}, \end{align} then \begin{equation}\twocase{\tau \ = \ }{0}{if $Z_0 \not \in {\mathcal L}_0$}{\inf \{n: Z_n \not \in {\mathcal L}\}}{otherwise.}\end{equation} Under $P$, $Z$ is a Markov chain. We abuse notation and denote its transition function by $P$ as well. Since the measure $P$ is uniform, it immediately follows that the restriction $P_{\mathcal L}$ of the transition function $P$ to ${\mathcal L}\times {\mathcal L}$ is an irreducible and aperiodic substochastic matrix. From the Perron-Frobenius theorem we know that $P_{\mathcal L}$ possesses a Perron root $\lambda_c \in (0,1)$ and corresponding left and right eigenfunctions, $\nu_c$ and $\varphi_c$, respectively, whose entries are strictly positive. We normalize them so that $\varphi_c$ and $\nu_c \varphi_c$ are probability measures. Let $Q$ be a stochastic transition function on ${\mathcal L}\times{\mathcal L}$ defined as follows: \begin{equation} \label{eq:Q} Q(z,z') =\frac{1}{\lambda_c\varphi_c(z) } P_{\mathcal L}(z,z') \varphi_c(z'). \end{equation} Observe that $Q$ inherits irreducibility and being aperiodic from $P_{\mathcal L}$. As a result, $Q$ is ergodic, and we denote its unique stationary distribution by $\pi^Q$. Recall that from the definition of a stationary distribution, $\pi^Q Q = \pi^Q$, if $\pi^Q$ is considered as a row vector, and it immediately follows that \begin{equation} \label{eq:piQ} \pi^Q(z) \ = \ \nu_c (z) \varphi_c(z). \end{equation} We also define the marginal of the first coordinate $\pi^Q_1$ by letting \begin{equation} \pi^Q_1 (x) \ = \ \sum_{b,b'} \pi^Q (x,b,b').\end{equation} Next we fix some notation. We write $P_\mu$ for the distribution of the Markov chain $Z$ under $P$ with initial distribution $\mu$, and $E^P_\mu$ for the corresponding expectation. When $\mu$ is a point mass $\delta_z$, we abbreviate and use $z$ as a subscript instead of the notationally correct but cumbersome $\delta_z$. We also define the analogous expressions with $Q$ instead of $P$. Finally, if $\mu$ is a probability distribution on ${\mathcal L}$, we write $E^P_{\mu}$ for the expectation associated with the process $Z$ and initial distribution $\mu$, and $E_\mu$ for expectation with respect to the probability $\mu$, equal to the expectation of the marginal $Z_0$ under $P_\mu$. The following result identifies the uniform distribution $Q^n$ with the distribution of the Markov chain $Z$ under $Q$. \begin{theorem} \label{th:structure} Let $f=f (Z_0,\dots,Z_n)$ be a complex-valued random variable. Then \begin{equation} E^{Q_n}( f )\ = \ \frac{ E_{\tilde \varphi_c}^Q \left ( \frac{f}{\varphi_c(Z_n)}\right) }{E_{\tilde \varphi_c}^Q\left ( \frac{1}{\varphi_c(Z_n)}\right)},\end{equation} where $\tilde\varphi_c$ is the probability measure given by $\varphi_c$ conditioned on ${\mathcal L}_0$ in \eqref{eq:L_0}. \end{theorem} \begin{proof}[Proof of Theorem \ref{th:structure}] Observe that if $z_0\in {\mathcal L}_0$ and $z_1,\dots,z_n \in {\mathcal L}$, then \begin{align} P_{z_0}( \prod_{j=0}^{n} \{ Z_j = z_j\} , \tau> n) & \ = \ \prod_{j=0}^{n-1} P(z_j,z_{j+1})\nonumber\\ &\ = \ \lambda_c^n \prod_{j=0}^{n-1} \varphi_c(z_j)Q(z_j,z_{j+1}) \frac{1}{\varphi_c(z_{j+1})} \nonumber\\ & \ = \ \lambda_c^n \varphi_c (z_0) Q _{z_0}( \prod_{j=0}^{n} \{ Z_j =z_j\} ) \frac{1}{\varphi_c (z_{n})}, \end{align} and otherwise $P_{z_0}( \prod_{j=0}^{n} \{ Z_j = z_j\} , \tau> n)=0$. In particular, if $f=f(Z_0,\dots,Z_n)$ is a complex valued random variable, then \begin{align} E^P ( f , \tau > n ) &\ = \ \sum_{z_0 \in {\mathcal L}_0} E^P(f, \tau>n, Z_0=z_0) \ = \ \sum_{z_0 \in {\mathcal L}_0} E^P ({\bf 1}_{\{Z_0=z_0\} } f(z_0,\dots,Z_n),\tau >n)\nonumber\\ & \ = \ \sum_{z_0\in {\mathcal L}_0} P(Z_0=z_0) E_{z_0}^P( f (Z_0,\dots,Z_n), \tau>n) \nonumber\\ & \ = \ \lambda_c^n \sum_{z_0\in {\mathcal L}_0} P(Z_0=z_0) \varphi_c (z_0) E^Q_{z_0} \left ( \frac{f }{\varphi_c(Z_n)}\right). \end{align} Since $P$ is uniform, it follows that $P(Z_0=z_0)$ is constant on ${\mathcal L}_0$, and the result follows. \end{proof} Next we consider limits. The following provides sufficient conditions under which $Q_n$ expectations and expectations with respect to $Q$ are asymptotically equivalent. \begin{prop} \label{lem:asymp} Suppose that for $n\in\ensuremath{\mathbb{Z}}_+$, $f_n (Z_0,\dots,Z_n)$ is a complex-valued random variable, and $(j_n:n\in\ensuremath{\mathbb{Z}}_+)$ is a subsequence of $\ensuremath{\mathbb{Z}}_+$ such that \begin{enumerate} \item $\min (j_n, n -j_n) \to \infty$, \item $E_{\tilde \varphi_c} ^Q |f_n - f_{j_n}| \to 0$. \end{enumerate} Then \begin{equation} | E^{Q_n} f_n - E^Q_{\tilde\varphi_c} f_n | \ = \ o(1) \max (|E^Q_{\tilde\varphi_c} (f_n)|,1) .\end{equation} \end{prop} \begin{proof} Because of condition (2), we have \begin{equation} E^{Q_n} \left ( f_n\right) \ = \ \frac {E_{\tilde \varphi_c}^Q \left ( \frac{f_{j_n} }{\varphi_c (Z_n)}\right)}{E_{\tilde \varphi_c}^Q\left ( \frac{1}{\varphi_c(Z_n)}\right)} + o(1).\end{equation} Then, by the Markov property, \begin{equation} E_{\tilde \varphi_c}^Q \left ( \frac{f_{j_n} }{\varphi_c (Z_n)}\right) \ = \ E_{\tilde\varphi_c}^Q \left ( f_{j_n} E_{Z_{j_n}} \left ( \frac{1}{\varphi_c (Z_{n-j_n})}\right)\right).\end{equation} The ergodicity of $Z$ under $Q$ and the fact that $n-j_n\to\infty$ guarantee that $ E_{Z_{j_n}}^Q\left ( \frac{1}{\varphi_c (Z_{n-j_n})}\right)=E_{\pi^Q} \frac{1}{\varphi_c} +o(1)=\|\nu_c\|_1+o(1)$. Thus \begin{eqnarray} E^{Q_n} (f_n) & \ = \ & \frac{ (\|\nu_c\|_1 + o(1)) E_{\tilde \varphi_c}^Q ( f_{j_n})}{\|\nu_c\|_1 + o(1)}+o(1) \nonumber\\ & \ = \ & (1 + o(1)) E_{\tilde \varphi_c}^Q (f_{j_n} ) + o(1)\ = \ (1 + o(1))E_{\tilde \varphi_c}^Q (f_n) + o(1).\end{eqnarray} \end{proof} For applications, it would be useful to know more about $Q$. It turns out that the underlying structure is determined by the matrix $C$, which we now describe. Let $C$ be the $L\times L$ matrix given by $C=(C_{i,j})$, $C_{i,1} = c_i$ and $C_{i,i+1} = 1$, and all other entries equal to $0$: \be \label{eq:C} C\ =\ \left ( \begin{array}{ccccc} c_1 & 1 & 0 & \cdots \\ c_2 & 0 & 1 & 0 & \cdots \\ \vdots & 0 & \cdots \\ c_{L-1} & 0 & \dots & & 1\\ c_L & 0 & \dots & & 0 \end{array} \right) \ee Let $\lambda_C$ denote the Perron eigenvalue of $C$, $\varphi_C$ a corresponding positive right eigenvector and $\nu_C$ a corresponding left eigenvector. A straightforward computation gives the following. \begin{lemma} Let $C$ be as in \eqref{eq:C}. Then \begin{enumerate} \item The characteristic polynomial of $C$ is $ \lambda^L - \sum_{j=1}^L c_j \lambda^{L-j}$. \item Up to multiplicative constants: $\nu_C(b) = \lambda_C^{-b}$ and $\varphi_C(b') = \lambda_C ^{b'} - \sum_{j=1}^{b'-1}c_j \lambda_C^{b'-j} $. \end{enumerate} \end{lemma} With this lemma we obtain a description of $Q$. \begin{prop} Let $C$ be as in \eqref{eq:C}, and let $\lambda_c,\nu_c,\varphi_c$, respectively, be the Perron eigenvalue, and corresponding left and right eigenvalues for $P_{\mathcal L}$, the restriction of the transition function $P$ to ${\mathcal L}$, normalized so that $\varphi_c$ and $\nu_c\varphi_c$ are probability distributions. Then \label{pr:quantities} \begin{enumerate} \item $\lambda_c =\frac{ \lambda_C}{(\max c_i +1)L}$, \item There exist positive constants $K_1,K_2$ such that $\varphi_c (a,b,b') =K_1 \varphi_C (b')$ and $\nu_c(a,b,b') = K_2\nu_C(b)$. In particular, $ \pi^Q (a,b,b') = K_1 K_2 \nu_C(b) \varphi_C(b')$, and $K_1 K_2 = \frac{1}{\lambda_C \sum_{b=1}^L \nu_C(b) \varphi_C(b)}$. \item $ Q((a,b,b'),(a',b',b'') ) = \frac{\varphi_C (b'')}{ \lambda_C\varphi_C (b')}$ for allowed transitions and is $0$ otherwise. \\ Furthermore, allowed transitions satisfy either of the following: \begin{enumerate} \item $b''= 1$ and then the probability of the transition is $\frac{\varphi_C(1)}{\lambda_C\varphi_C(b')}$; \item $b''=b'+1$ and then the probability of the transition is $1- \frac{ \varphi_C(1) c_{b'}}{\lambda_C \varphi_C (b')}$. \end{enumerate} \end{enumerate} \end{prop} \begin{example} \label{ex:zeck_quant} For the standard Zeckendorf decomposition, we have \begin{enumerate} \item $ C = \left ( \begin{array}{cc} 1 & 1 \\ 1 & 0 \end{array} \right)$. In particular, \begin{enumerate} \item The characteristic polynomial is $\lambda^2-\lambda -1$, and $\lambda_C= \phi$, where $\phi$ is the golden ratio $\phi = \frac{1+\sqrt{5}}{2}$. \item $\nu_C (b) = \phi^{-b}$, and $\varphi_C(b') = \phi^{2-b'}$. \end{enumerate} \item ${\mathcal L} = \{(0,1,1),(0,2,1), (1,1,2)\}$. Identifying these states as $1$, $2$ and $3$ in the order written, then \begin{enumerate} \item $ Q = \left ( \begin{array}{ccc} \frac{1}{\phi} & 0 & 1- \frac{1}{\phi} \\ \frac{1}{\phi} & 0 & 1- \frac{1}{\phi} \\ 0 & 1 & 0 \end{array} \right).$ \item $\pi^Q(0,1,1) = \frac{\phi}{2+\phi},~\pi^Q(0,2,1) = \frac{1}{2+\phi},\pi^Q(1,1,2) = \frac{1}{2+\phi}$, and \\ $\pi^Q_1(0) = \frac{1+\phi}{2+\phi},~\pi^Q_1(1) = \frac{1}{2+\phi}.$ \item $\varphi_c = \frac{1}{2\phi+1} \left (\phi,\phi,1 \right)^t$. \item $\nu_c = \frac{1}{\phi+2}\left (2\phi+1,\phi+1,2\phi+1\right)^t$. \end{enumerate} \end{enumerate} \end{example} \begin{proof}[Proof of Proposition \ref{pr:quantities}]~\\ 1. The first part is a straightforward calculation. \\ 2. Observe that for the row of $P$ corresponding to transition from $(a,b,b')$, we have exactly $|{\mathcal S}_1|\times |{\mathcal S}_2|=(\max_i c_i +1)L$ allowed sites to transition to, and due to the choice of uniform distribution, all are of equal probability. As $P$ is stochastic, its nonzero entries are equal to $\gamma = \frac{1}{(\max c_i +1)L }$. We first study the restriction $P_{\mathcal L}$ of $P$ to ${\mathcal L}\times {\mathcal L}$. Recall that the elements of $\mathcal L$ are of the form $(x,k,1)$, where $x < c_k$ or $(c_k,k,k+1)$ where $k=1,\dots,L-1$. For each $(a,b,b')\in {\mathcal L}$, $P_{\mathcal L}$ has a corresponding row, listing all transitions from $(a,b,b')$. We will count the number of such non-zero entries according to the value of $b'$. If $b'\in \{1,\dots, L-1\}$ then there are $1+c_{b'}$ transitions: one to the site $(c_{b'},{b'},b'+1)$ and $c_{b'}$ to $(x,b',1)$ where $x \in \{0,\dots, c_{b'}-1\}$. If $b' = L$ then there are only $c_L$ allowed transitions, all of which are of the second kind. We define a function $\varphi$ on ${\mathcal L}$ by letting $\varphi (a,b,b') = \varphi_C (b')$. Fix $(a,b,b')\in A$. If $b' < L$, then according to the allowed transitions listed above, we have \begin{equation} P_{\mathcal L} \varphi (a,b,b') \ = \ \gamma ( \varphi_C (b'+1) + c_{b'} \varphi_C (1)) \ = \ \gamma (C \varphi_C )(b') \ = \ \gamma \lambda_C \varphi(a,b,b').\end{equation} Similarly, if $b'=L$, then $P_{\mathcal L} \varphi(a,b,L) = \gamma c_L \varphi_C (1) = \gamma \lambda_C \varphi(a,b,L)$. Thus $\gamma \lambda_C=\lambda_c$, the Perron root for $P_{\mathcal L}$, and $\varphi$ is a corresponding positive eigenvector. Next we want to find the corresponding left-eigenvector for $P_{\mathcal L}$. To do that, let $D$ be the transpose of $C$, and let $\nu_C$ be a Perron eigenvector. Define $\nu_c (a,b,b') := \nu_C (b)$. If $b \in \{2,\dots, L\}$, then there is exactly one allowed transition to it, that is from $(c_{b-1},b-1,b)$. As a result, $\nu_c P_{\mathcal L} (a,b,b') = \gamma \nu_c (c_{b-1},b-1,b) = \gamma (D \nu_C)(b) =\gamma \lambda_C \nu_c (a,b,b')$. Next, if $b=1$, then the allowed transitions are from $(x,k,1)$ where $k=1,\dots, L$ and $x \in \{0,\dots,c_k-1\}$. We obtain $ \nu_c P_{\mathcal L} (a,1,b') = \gamma \sum_{k=1}^L c_k \nu_C (k) = \gamma (D \nu_C)(1) =\gamma \lambda_C \nu_c (a,1,b')$.\\ The formula for $\pi^Q$ follows directly from \eqref{eq:piQ} and the preceding identities, while the formula for $K_1K_2$ follows from the calculation below. \begin{align} \sum_{a,b,b'} \pi^Q(a,b,b') & \ = \ \sum_{a,b} \pi^Q (a,b,1) + \sum_{a,b} \pi^Q( a,b,b+1) \nonumber\\ & \ = \ K_1K_2 \left( \sum_{b=1}^{L} c_b \nu_C (b) \varphi_C(1)+ \sum_{b=1}^{L-1} \nu_C(b) \varphi_C(b+1)\right)\nonumber\\ & \ = \ K_1K_2 \sum_{b=1}^{L} \nu_C (b) \left ( c_b \varphi_C(1) + \varphi_C(b+1)\right) \nonumber\\ & \ = \ K_1 K_2 \lambda_C \sum_{b=1}^L \nu_C(b) \varphi_C (b). \end{align} \noindent 3. This follows from \eqref{eq:Q} and parts 1. and 2. \end{proof} \section{Additive functionals} \label{sec:additive} \subsection{General Theory} \label{sec:additive_theory} In this section we will study some theoretical aspects of large-time behavior of additive functionals of an ergodic finite-state Markov chain, under a change of measure which generalizes the way $Q_n$ was obtained from $Q$. The assumptions in this section are the following: \begin{defi} \label{def:additive} Let $Z=(Z_n:n\in\ensuremath{\mathbb{Z}}_+)$ be an irreducible and aperiodic Markov chain on the finite state space ${\mathcal L}$ with transition function $Q$. Let $\varphi:{\mathcal L} \to (0,\infty)$ be a positive function, and let $\mu$ be a probability distribution on ${\mathcal L}$. For every $n\in\ensuremath{\mathbb{Z}}_+$, let $Q_n$ be a probability measure on $\sigma (Z_0,\dots,Z_n)$ given by \be Q_n ( A )\ =\ \frac{E^Q_{\mu} \left ( \frac{{\bf 1}_A}{\varphi(Z_n)}\right)}{E^Q_{\mu} \frac{1}{\varphi(Z_n)}},\ \ \ A \in \sigma (Z_0,\dots,Z_n).\ee \end{defi} We will consider the behavior of additive functionals of the form $S_n = \sum_{j=0}^n g (Z_j)$ where $g :{\mathcal L} \to \ensuremath{\mathbb{C}}$ under $Q_n$ as $n\to\infty$. In the context of generalized Zeckendorf decompositions, an example for an additive functional is the number of, say, nonzero digits in the decomposition. In the next section, we show that gaps in the decomposition can be viewed as additive functionals of some Markov chain, so we can treat them with the same tools. \\ We need to fix some notation. Functions on ${\mathcal L}$ will interchangeably be viewed as column vectors. As an example, if $g$ is such a function then $Qg$ is to be identified as the function or, equivalently the column $f$ vector given by $f(z) = \sum_{z' \in {\mathcal L}} Q(z,z') g(z)$. We will write $hg$ for the product of such two functions, namely $hg$ is the function given by $(hg)(z) = h(z) g(z),~z\in {\mathcal L} $. In addition, $h (Qg)$ means the product of the function $h$ and the function $Qg$, not their scalar product. \\ Let $\pi^Q$ denote the stationary distribution for $Q$. Recall that $I-Q$ is invertible on the $Q$-invariant subspace of $V$, where $V =\{g:E_{\pi^Q} g(z)=0\}$. We denote this inverse by $Q^{\#}$, and extend it to all functions by letting $Q^{\#} {{\bf 1}} =0$. This is the only choice that guarantees that $Q$ and $Q^\#$ commute, and $Q^\#$ is known as the group inverse of $Q$. It is well-known that \begin{equation} \label{eq:Qsharp} \sum_{j=0}^\infty E_z^{Q} (g(Z_j) -E_{\pi^Q} g)\ = \ (Q^\# g) (z). \end{equation} Our first result is the following. \begin{theorem} \label{th:additive_functionals} Let $g:{\mathcal L} \to \ensuremath{\mathbb{C}}$. Let $\tilde g= g - E_{\pi^Q} g$, and $\tilde S_n = \sum_{j=0}^{n} \tilde g(Z_j) $. Then \begin{eqnarray} \label{eq:lekkerkerker2} E^{Q_n}\tilde S_n &\ = \ & E_{\mu} (Q^\# g) + \frac{ E_{\pi^Q} \tilde g (Q^\# \frac{1}{\varphi})}{E_{\pi^Q} \frac{1}{\varphi} } + o(1) \end{eqnarray} \begin{eqnarray} \label{eq:sec_mom} E^Q_{\pi^Q} \tilde S_n^2 &\ = \ & (n+1)E_{\pi^Q} \left (\tilde g ((2Q^\# - I) \tilde g)\right)+o(1)\mbox{ and } E^{Q_n} \tilde S_n^2\ =\ (1+o(1))E^Q_{\pi^Q} \tilde S_n^2.\ \ \end{eqnarray} \end{theorem} \begin{proof} We will first prove \eqref{eq:lekkerkerker2}. From Theorem \ref{th:structure} with $f= \tilde S_n$, and the Markov property, we have that \begin{align} E^Q_{\mu} \frac{1}{\varphi(Z_n)}\times E^{Q_n} \tilde S_n & \ = \ \sum_{j=0}^n E^Q_{\mu} \tilde g (Z_j) E_{Z_j} \frac{1}{\varphi(Z_{n-j})}\nonumber\\ & \ = \ \usb{(I)}{\sum_{j=0}^n E^Q_{\mu} \tilde g (Z_j)\left ( E_{Z_j}^Q \frac{1}{\varphi(Z_{n-j})} - E_{\pi^Q} \frac{1}{\varphi}\right)}\nonumber\\ &\ \ \ \ \ \ \ + E_{\pi^Q} \frac{1}{\varphi}\usb{(II)}{ \sum_{j=0}^n E^Q_{\mu} \tilde g (Z_j) }. \end{align} By \eqref{eq:Qsharp}, $(II)\to E_{\mu} (Q^\# \tilde g)=E_{\mu} Q^\# g$, because $Q^\#$ maps constant function to $0$. In order to estimate $(I)$, we recall that from the exponential ergodicity of irreducible finite state Markov chains, there exists $\rho \in(0,1)$ and $c_1>0$, such that for every function $h$ and $k\in \ensuremath{\mathbb{Z}}_+$, \begin{equation} \label{eq:exp_ergo} \sup_{z} | E_z^Q h (Z_k) - E_{\pi^Q} h|\ \le\ c_1 \|h\|_\infty \rho^k. \end{equation} Letting $h(z) = \frac{1}{\varphi(z)}-E_{\pi^Q} \frac{1}{\varphi}$, we have that $E_{\pi^Q} h =0$. This allows us torewrite $(I)$ as $ \sum_{j=0}^n E_{\mu} \tilde g (Z_j) E_{Z_j} h (Z_{n-j})$. In order to estimate this sum, we break it into two parts. First \begin{equation}\left|\sum_{j=0}^{\lfloor n/2 \rfloor } E_{\mu}^Q \tilde g (Z_j) E_{Z_j} h (Z_{n-j}) \right| \ \le \ \|\tilde g\|_\infty \sum_{j=0}^{\lfloor n/2 \rfloor } \sup_{z} |E_{z} h (Z_{n-j}) | \ \le \ c_1\|\tilde g\|_\infty \|h\|_\infty \rho^{n/2} n/2\to 0,\end{equation} where the last inequality follows from \eqref{eq:exp_ergo}. Next, let $h_{k}(z) = \tilde g(z) E_z h (Z_{k})$. Then \begin{equation} \sum_{j={\lfloor n/2 \rfloor +1}}^n E_{\mu}^Q \tilde g(Z_j) E_{Z_j} \frac{1}{\varphi(Z_{n-j})} \ = \ \sum_{j={\lfloor n/2 \rfloor +1}}^n E_{\mu}^Q h_{n-j} (Z_j) .\end{equation} Applying \eqref{eq:exp_ergo} to each of the functions $h_{k}$, and observing that $\|h_k\|_\infty \le \|\tilde g\|_\infty \|h\|_\infty$, it follows that for $j\ge \lfloor n/2 \rfloor +1$, \begin{equation}|E_{\mu}^Q h_k (Z_j) - E_{\pi^Q} h_k |\ \le \ c_1 \|\tilde g\|_\infty\|h\|_\infty\rho^{n/2}.\end{equation} Also, since $\pi^Q$ is the stationary distribution for $Q$, we have that $ E_{\pi^Q} h_k = E_{\pi^Q}^Q h_k(Z_j)$, and as a result \begin{equation}\sum_{j={\lfloor n/2 \rfloor +1}}^n \left ( E_{\mu}^Q h_{n-j} (Z_j) - E_{\pi^Q}^Q h_{n-j} (Z_j) \right)\ \le \ c_1 \|\tilde g\|_\infty\|h\|_\infty \rho^{n/2}n/2 \to 0.\end{equation} In addition, $E_{\pi^Q}^Q h_{n-j} (Z_j) = E_{\pi^Q} \tilde g(Z_0) E_{Z_0} h (Z_{n-j})$, and therefore \begin{equation}\sum_{j=\lfloor n/2 \rfloor +1}^n E_{\pi^Q}^Q h_{n-j} (Z_j) \ = \ \sum_{k=0}^{n- \lfloor n/2 \rfloor -1} E_{\pi^Q}^Q \tilde g (Z_0) E_{Z_0} h (Z_k).\end{equation} Since by our choice $E_{\pi^Q} h = 0$, it follows from \eqref{eq:Qsharp} that the righthand side is equal to $E_{\pi^Q} \tilde g (Q^\# h)+o(1)$. As a result, $ (I) = E_{\pi^Q} \tilde g (Q^\# \frac{1}{\varphi})+ o(1)$, completing the proof of \eqref{eq:lekkerkerker2}. \\ We turn to proving \eqref{eq:sec_mom}. We first prove the first equality. \begin{align} E_{\pi^Q}^Q \left ( \tilde S_n^2 \right)&\ = \ \sum_{j=0}^n E_{\pi^Q}^Q \tilde g ^2 (Z_j) + 2 \sum_{0\le j < k\le n} E_{\pi^Q} \tilde g(X_j) \tilde g (X_k)\nonumber\\ & \ = \ (n+1)E_{\pi^Q}^Q \tilde g ^2 + 2 \sum_{0\le j < k\le n} E_{\pi^Q}^Q \tilde g(X_0) E_{X_0}^Q \tilde g (X_{k-j})\nonumber\\ & \ = \ -(n+1)E_{\pi^Q} \tilde g^2 + 2 \sum_{j=0}^n \sum_{k=j}^n E_{\pi^Q}^Q \tilde g(X_0) E_{X_0}^Q \tilde g (X_{k-j}) \nonumber\\ & \ = \ - (n+1) E_{\pi^Q} \tilde g ^2 + 2 \sum_{j=0}^n E_{\pi^Q}^Q \tilde g(X_0)\left ( \sum_{k=0}^{n-j} E_{X_0}^Q \tilde g (X_k)\right) \nonumber\\ & \ = \ -(n+1) E_{\pi^Q} \tilde g^2 + 2 \sum_{j=0}^n E_{\pi^Q} \tilde g Q^\# \tilde g - 2 \usb{(*)}{\sum_{j=0}^n E_{\pi^Q}^Q \left ( \tilde g (X_0) \sum_{k > n-j} E^Q_{X_0} \tilde g (X_k)\right)}\nonumber\\ & \ = \ (n+1) E_{\pi^Q} \tilde g (2Q^\#-I) \tilde g + (*). \end{align} Observe that by exponential ergodicity, \eqref{eq:exp_ergo}, $|E_z^Q \tilde g (X_k) |\le c_1 \|\tilde g\|_\infty \rho^k$, uniformly over $z$, and so \begin{equation}|(*)| \ \le c_1 \|\tilde g\|_\infty^2 \sum_{j=0}^n \frac{\rho^{n-j+1}}{1-\rho}\le c_1 \|\tilde g\|_\infty^2 \frac{1}{(1-\rho)^2}=O(1).\end{equation} This completes the proof of the first equality in \eqref{eq:sec_mom}. It remains to the asymptotic equivalence of $E^Q_{Q_n} \tilde S_n^2$ and $E^Q_{\mu} \tilde S_n^2$. This, again, follows from the exponential ergodicity, as we now explain. We have \be \tilde S_n^2\ = \ \tilde S_m^2 + 2 \tilde S_m (\tilde S_n -\tilde S_m) + (\tilde S_n -\tilde S_m)^2.\ee From the Markov property and exponential ergodicity \eqref{eq:exp_ergo}, it follows that \be |E_{\mu} (S_n -S_m)^2 -E_{\pi^Q} \tilde S_{n-m}^2 |\ \le\ c_1 \|\tilde g\|_\infty n^2 \rho^{m}.\ee Choose $m = c \ln n$ for $c=4/\ln (1/\rho) $. It follows that righthand side tends to $0$ as $n\to\infty$. In particular, $E_{\mu} (S_n -S_m)^2 \le c_2 n$. Next, observe that $ E_{\mu} \tilde S_m^2 \le \|\tilde g\|_\infty^2 m^2$, and by Cauchy-Schwarz, $|E_{\mu} \tilde S_m (\tilde S_n -\tilde S_m)| \le\sqrt{ E_{\mu} \tilde S_m^2 }\sqrt{E_{\mu} (\tilde S_n -\tilde S_m)^2}\le c_3 m \sqrt{n}$. In summary, for all $n$ large enough, \be |E_{\mu}\left ( \tilde S_m^2 + 2 \tilde S_m (\tilde S_n -\tilde S_m)\right) |\ \le\ c_4 (\ln n)^2 \sqrt{n}\le c_4 n^{3/4}.\ee In particular, \be |E_{\mu} \tilde S_n^2 - E_{\pi^Q}^Q \tilde S_{n-m}^2 |\ \le\ c_4 n^{3/4},\ee so that \be E_{\mu} S_n^2 \ =\ (1 + o(1)) n E_{\pi^Q} \tilde g (2Q^\#-I) \tilde g,\ee and the claim is proved. \end{proof} We turn to laws of large numbers and central limit theorems for additive functionals. \begin{theorem}~ Under the same assumptions of Theorem \ref{th:additive_functionals} we have: \label{th:LLN_CLT} \begin{enumerate} \item Weak Law of Large Numbers: For $\epsilon>0$, $\lim_{n\to\infty} Q_n \left (|\frac{ \tilde S_n }{n+1} |> \epsilon \right)=0$. \item Central Limit Theorem: $ Q_n \left( \frac{\tilde S_n}{\sqrt{n+1}} \le x~\right) \Rightarrow P( Y \le x)$ where $Y\sim N(0,\sigma^2)$, and $\sigma^2 = E_{\pi^Q} \tilde g ((2Q^\# -I)\tilde g)$. \end{enumerate} \end{theorem} \begin{proof}[Proof of Theorem \ref{th:LLN_CLT}] The Weak Law of Large Numbers follows from Chebychev's inequality and the asymptotic estimate for $E^{Q_n} \tilde S_n^2$ given in Theorem \ref{th:additive_functionals}: \be Q_n\left (|\frac{ \tilde S_n}{n+1} |\ > \ \epsilon\right)\ \le\ \frac{ E^{Q_n} \tilde S_n^2 }{(n+1)^2 \epsilon^2} \ = \ \frac{E_{\pi^Q} \tilde g (2Q^\# -I) \tilde g }{(n+1)\epsilon^2}\to 0,\mbox{ as } n\to\infty.\ee We now prove the Central Limit Theorem. To do this we apply Proposition \ref{lem:asymp} with $j_n = n- \lfloor \ln n\rfloor $ and \begin{equation} f_n \ = \ \exp\left ( \frac{i \theta}{\sqrt{n+1}} \tilde S_n \right).\end{equation} Observe that the choice of $j_n$ guarantees that condition 1. in the proposition holds. Next, \begin{equation} E^Q_z |f_n - f_{j_n} | \ \le \ E_z^Q | 1- E_{Z_{j_n}}^Q e^{\frac{ i \theta}{\sqrt{n+1}} \tilde S_{n-j_n} }|\ \le \ \max_{z} \left ( |1- E_z^Q\cos ( \frac{\theta \tilde S_{n-j_n}}{\sqrt{n+1}}) | + |E_z^Q \sin (\frac{\theta \tilde S_{n-j_n}} {\sqrt{n+1}} )|\right).\end{equation} Since $|S_{n-j_n}| =O( \ln n)$, it follows from bounded convergence that $ \sup_{z} E^Q_z |f_n - f_{j_n} | \to 0$, and so condition 2. holds. Finally, we recall from the Central Limit Theorem for additive functionals of finite state Markov chains (e.g. \cite{maxwell_woodroofe},\cite[Theorem 5]{benari_neumann}, that \begin{equation} E_{\mu}^Q (f_n) \to e^{-\frac{ \sigma^2}{2}},\end{equation} where $\sigma^2 = \lim_{n\to\infty} \frac{1}{n+1} E_{\pi^Q}^Q \left ( \tilde S_n ^2 \right)$. The result now follows from Theorem \ref{th:additive_functionals}. \end{proof} \subsection{Application to Zeckendorf} \label{sec:application_zeckendorf} In this section we show how the results obtained in Section \ref{sec:additive_theory} apply to generalized Zeckendorf decompositions. In particular we will show that the generalized Lekkerkerker's theorem (Theorem \ref{th:gen_lek}) and the corresponding Central Limit Theorem are specials cases to Theorem \ref{th:additive_functionals}-1 and Theorem \ref{th:LLN_CLT}-2. We will also carry out explicit computations for the standard Zeckendorf decomposition, where all quantities are easily computable. \\ In order to apply the results in the context of generalized Zeckendorf decomposition, in Definition \ref{def:additive} we identify ${\mathcal L}$, $Z$ and $Q$ in the definition as the same quantities defined in Section \ref{sec:prob}, and also set $\varphi=\varphi_c$, and $\mu= \tilde\varphi_c$, where $\varphi_c$ and $\tilde \varphi_c$ are as in Section \ref{sec:prob}. With these choices, the measure $Q_n$ of Definition \ref{def:additive} coincides with $Q_n$ of Section \ref{sec:prob}. \\ Recall $k(N)$, the number of nonzero summands in the generalized Lekkerker decomposition of $N$, defined in \eqref{eq:kN_def}. Let $g:{\mathcal L} \to \{0,1\}$ be defined as $g(x,j,j')=1$ if and only if $x>0$. Then if $N \in [G_{n+1},G_{n+2})$, the from \eqref{eq:decomp2} have that that $k(N) =S_n$, where $S_N$ is the additive functional $S_n = \sum_{j=0}^n g (Z_j)$. Observe that $E_{\pi^Q} g = 1- \pi_1(0)$, and so $\tilde g= g - 1 + \pi_1(0)$. Furthermore, since $\pi^Q(z) = \varphi_c (z) \nu_c(z)$, it follows that $E_{\pi^Q} \frac{1}{\varphi_c} = \|\varphi\|_1$. The following therefore follow immediately from Theorem \ref{th:additive_functionals} and Theorem \ref{th:LLN_CLT}. \begin{cor} For generalized Zeckendorf decomposition: \label{cor:lekkerkerker} \begin{enumerate} \item Generalized Lekkerkerker's Theorem (Th. \ref{th:gen_lek}): \be E^{Q_n} k(N)\ =\ C_{\rm Lek} (n+1) + d\ee where \be C_{\rm Lek} \ = \ 1-\pi_1(0),~d \ = \ E_{\tilde \varphi_c} Q^\#(1-\delta)+\frac{ E_{\pi^Q} (1-\delta) (Q^\# \frac{1}{\varphi_c})}{\|\nu_c\|_1}, \ee \item Variance: \be E^{Q_n} (k(N) - C_{\rm Lek} (n+1) )^2\ =\ (1+o(1)) (n+1) \sigma^2\ee where \be \sigma^2\ =\ E_{\pi^Q} \tilde g((2Q^\#- I ) \tilde g).\ee \end{enumerate} \end{cor} \begin{cor} For generalized Zeckendorf decomposition: \begin{enumerate} \item Law of Large Numbers: \be Q_n\left( \left | k(N) - C_{\rm Lek}(n+1) \right | > n \epsilon\right)\ \to\ 0.\ee \item Central Limit Theorem: \be Q_n\left( \frac{k(N) - C_{\rm Lek}(n+1) }{\sqrt{n+1}} \in \cdot\right)\ \to\ N(0,\sigma^2) \ee where $\sigma^2$ is as in Corollary \ref{cor:lekkerkerker} \end{enumerate} \end{cor} In the remainder of the section we compute all constants above for the standard Zeckendorf decomposition. First we need to compute $Q^\#$. \begin{example} For the standard Zeckendorf decomposition, \begin{equation} Q^{\#} \ = \ \frac 15 \left ( \begin{array} {ccc} 5-\phi & \phi -4 & -1 \\ -\phi & \phi+1 & -1 \\ 1-3\phi & 2\phi -2 & \phi +1 \end{array}\right).\end{equation} \end{example} To prove the identity, recall the expressions for $Q$ and $\pi^Q$ computed in Example \ref{ex:zeck_quant}. Let $A=I-Q$, and let $v_1 = (0,1,-1)^t$, $v_2=(1,0,-\phi)^t$, and $v_3 = (1,1,1)^t$. Then $E_{\pi^Q} v_1 = E_{\pi^Q} v_2=0$. Since $v_1$ and $v_2$ are linearly independent, it follows that they span the $A$-invariant space $V =\{v: E_{\pi^Q} v=0\}$. In addition $A v_3=0$. Letting $q= 1- \frac{1}{\lambda_C} = \frac{1}{\lambda_C^2}$, a straightforward calculation shows that $A v_1 = q v_2 + (1+q) v_1$, and $A v_2 = v_2$. Thus $v_1 = q v_2 + (1+q) Q^\# v_1$, $Q^\# v_2 =v_2$ and $Q^\# v_3 = 0$. These determine $Q^\#$. \\ Also, from Example \ref{ex:zeck_quant} we have that $\pi_1(0) = \frac{\phi+1}{\phi+2}$, $\tilde \varphi_c$ is a point mass, and $\|\nu_c\|_1 = \frac{5\phi+3}{\phi+2}$. In addition, $\pi^Q = \frac{1}{\phi+2} \left(\phi ,1,1\right)^t$, and $\varphi_c = \frac{1}{2\phi+1} (\phi,\phi,1)^t$. Since also $g= (0,0,1)^t$, we have $Q^\# g =\frac 15 \left (1-3\phi,2\phi-2,\phi+1 \right )^t$, and $Q^\# \frac{1}{\varphi_c} = \frac{1}{5(\phi-1)}\left(-1, -1, \phi+1\right)^t$. As a result, we have the following. \begin{example} For the standard Zeckendorf: \begin{equation} C_{\rm Lek} \ = \ \frac{1}{\phi+2} \ = \ \frac{5 -\sqrt{5}}{10},\ ~d\ =\ \frac 35.\end{equation} \end{example} We finally compute $\sigma^2$. Clearly, $\tilde g=(0,0,1)^t - \frac{1}{2+\phi} (1,1,1)^t=\frac{1}{2+\phi} ( -1,-1,1+\phi)^t$. It therefore follows that $\tilde g Q^\# \tilde g= \tilde g Q^\# (0,0,1)^t = \frac 15 \left ( \frac{1}{\phi+2} ,\frac{1}{2+\phi} ,(1-\frac{1}{2+\phi}) (\phi+1) \right)^t$, and so the expectation is equal to \begin{equation} \label{eq:first} 2E_{\pi^Q}\tilde gQ^\# \tilde g \ = \ \frac 25 \left ( \frac{1+\phi+(1+\phi)^2}{(\phi+2)^2}\right)\ = \ \frac{2(\phi+2)}{25}. \end{equation} Since $\tilde g^2 = (\frac{1}{(\phi+2)^2}, \frac{1}{(\phi+2)^2}, \frac{(\phi+1)^2}{(\phi+2)^2})^t= \frac {1}{5 (1+\phi)}( 1,1,(1+\phi)^2)^t$, it follows that \begin{equation} \label{eq:second} E_{\pi^Q} \tilde g^2 \ = \ \frac{1}{5(1+\phi)} \frac{( \phi +1)+(\phi+1)^2}{\phi+2} \ = \ \frac 15. \end{equation} We therefore have \begin{example} For the standard Zeckendorf decomposition: $\sigma^2 = \frac{2\phi-1}{25} = \frac{\sqrt{5}}{25}$. \end{example} \section{Gaps in Zeckendorf Decomposition} \label{sec:gaps} \subsection{Gap Distribution} In this section we consider the asymptotic distribution of gaps between non-zero terms in the generalized Zeckendorf decomposition. This will be an application of our results on additive functionals from the previous section. We will first prove a statement on an ``average" gap distribution, Theorem \ref{th:gap_dist}, and we will later prove convergence of empirical gap measures in probability, Theorem \ref{th:weaklim_prob}. Let us first define the notion of a gap. We work under the same assumptions and notation as in Section \ref{sec:prob}. Suppose that $N\in \mathbb{N}$ admits a legal decomposition \eqref{eq:decomp2} with $X_0>0$. Note that $X_j$ counts the repetitions of $G_{n-j+1}$, and if repeating more than $1$ times, we can view this as $X_j-1$ gaps of size zero. If if $X_j>0$, than we have a gap of size $1$ or larger, the size of the gap equal to $\min\{k \ge 1: X_{j+k} >0\}$. Let $N_n (k)$ denote the number of gaps of size $k$ in the first $n$ digits, and let $N_n= \sum_k N_n(k)$. We define the gap distribution $\mu_n$ as a probability measure on $\ensuremath{\mathbb{Z}}_+$ given by \begin{equation}\mu_n (k) \ = \ \frac{ E^{Q_n} N_n (k)}{E^{Q_n} N_n}.\end{equation} To state the next theorem, let \be \nu(k) \ =\ \lambda_C ^{-(k-1)} (1-\lambda_C^{-1}) \ee denote the probability density of a Geometric random variable with parameter $\lambda_C^{-1}$. We have \begin{theorem} Let $ H_1 = \{(0,b,1) \in{\mathcal L}\}$ and $H_2 = \{(0,b+1,b+2)\in {\mathcal L}: c_b>0,c_{b+1} =0\}$. For $z=(0,b+1,b+2) \in H_2$ we let \begin{align} r(b) &\ = \ \max\{j : c_{b+j} = 0\},~\nonumber\\ \rho(b) &\ = \ Q ( (0,b+r(b),b+r(b)+1),(0,b+r(b)+1,1))= \frac{\varphi_C (1)}{\lambda_C \varphi_C(b+r(b)+1)}\mbox{, and }\\ h (b,k) &\ = \ \begin{cases} 0 & k < r (b)+1 \\ 1- \rho(b) & k = r(b)+1 \\ \rho (b) \lambda_C^{-(k-r(b)-2)}(1-\lambda_C^{-1}) & k> r(b)+1.\end{cases} \end{align} Then \label{th:gap_dist} \begin{enumerate} \item $\lim_{n\to\infty} \frac{1}{n} E^{Q_n} N_n =M_{\pi^Q_1}$. \item \be \lim_{n\to\infty} \mu_n(k)\ =\ \begin{cases} 1- \frac{1-\pi^Q_1(0)}{M_{\pi^Q_1}}& k= 0 \\ \frac{ 1- \pi^Q_1(0) - \pi^Q(H_1) ( 1- \lambda_C^{-1}) -\sum_{z\in H_2} \pi^Q (z) (1- \rho(z(2)))}{M_{\pi^Q_1}}& k=1. \end{cases}\ee \item For $k\ge 2$, \begin{align} \nonumber \label{eq:zzz}\lim_{n\to\infty} \mu_n (k) &\ = \ \frac{ \pi^Q (H_1) \nu(k-1)}{M_{\pi^Q_1}}.\nonumber\\ &\ \ \ \ \ \ +\ \frac {\sum_{z \in H_2} \pi^Q ( z ) \left ( h (z(2)-1,k) -\rho(z(2))\nu (k-1)\right)}{M_{\pi^Q_1}}. \end{align} \end{enumerate} \end{theorem} Since $\sum_{k\ge 2} \nu(k-1)=\sum_{k\ge 2} h(b,k) =1$, it follows that the limit $\lim_{n\to\infty} \mu_n(\cdot )$ is a probability measure, which we denote by $\mu_\infty$. A simple argument shows that a stronger result holds. For $n\in\mathbb{N}$, define the empirical gap distribution $\hat \mu_n$ as a random measure on $\ensuremath{\mathbb{Z}}_+$, defined by \be \hat \mu_n (A)\ =\ \frac{ \sum_{k \in A} N_n(k) }{\max (N_n,1)}.\ee We therefore have the following. \begin{theorem} \label{th:weaklim_prob} For any $A\subset \ensuremath{\mathbb{Z}}_+$ and $\epsilon>0$, \be \lim_{n\to\infty} Q_n (\left |\hat \mu_n (A) - \mu_\infty (A)\right|>\epsilon)\ =\ 0.\ee \end{theorem} We comment that the expression for the limit in Theorem \ref{th:gap_dist} is much simpler when $c_j>0$ for all $j=1,\dots,L$. In this case $H_2=\emptyset$. For the standard Zeckendorf, we have the following. \begin{example} For the standard Zeckendorf decomposition, $M_{\pi^Q_1} = \pi^Q_1 (1) = \frac{1}{\phi+2}$ and $\lambda_C = \phi$. Therefore \begin{enumerate} \item $\lim_{n\to\infty} \frac{1}{n} E^{Q_n} N_n =\frac{1}{\phi+2}$. \item $\lim_{n\to\infty} \mu_n(k)\ =\ \begin{cases} 0 & k= 0,1 \\ \phi^{-k} & k\ge 2. \end{cases}$ \end{enumerate} \end{example} When some of the coefficients are zero, then some gaps of size $\ge 2$ are forced by the recurrence relation, and taking this into account is the source of the lengthy expression in the theorem. \begin{example} Consider the recurrence relation with $L=4$, $c_1=1,c_2=c_3=0,c_4=2$. Then $\lambda_C $ is the largest (real) root of $\lambda^3 (\lambda-1) = 2$, $\lambda_C \thickapprox 1.5437$. We have \be h(k)\ =\ \begin{cases} 0 & k <3 \\ \frac 12 & k=3 \\ \frac 12 \lambda_C^{-(k-4)} (1-\lambda_C^{-1}) & k \ge 4\end{cases}\ee and \be \lim_{n\to\infty} \mu_n (k)\ =\ \begin{cases} 0 & k= 0 \\ 2-\frac{ \lambda_C^2+1}{3\lambda_C} & k=1 \\ \frac{(\lambda_C-1)^2 - \lambda_C}{3\lambda_C}\nu (k-1) + \frac{ 2\lambda_C-1}{3\lambda_C} h(k) & k\ge 2. \end{cases} \ee \end{example} In this example, \be {\mathcal L}\ =\ \left\{z^1 = (0,1,1),z^2=(1,1,2),z^3=(0,2,3),z^4=(0,3,4),z^5=(0,4,1),z^6=(1,4,1)\right\}.\ee There are no gaps of size $0$ as the coefficients immediately show. Gaps of size $1$ only appear in the form $(1,4,1)$ followed by $(1,1,2)$. Larger gaps can be formed as follows. \begin{itemize} \item Gaps of size $k\ge 2$ through a sequence of the form $(1,4,1), (0,1,1),\dots,(1,1,2)$, with $(0,1,1)$ repeated $k-1$ times. \item Gaps of size $k\ge 3$ through a sequence beginning with $(1,1,2),(0,2,3),(0,3,4)$, followed by $(1,4,1)$ if length is $3$, or by $k-3$ repetitions of $(0,1,1)$ followed by $(1,1,2)$ otherwise. \end{itemize} The larger gaps of the second type are forced by the recurrence, in the sense that the condition $c_2=c_3=0$ implies $Q( (1,1,2),(0,2,3)) = Q((0,2,3),(0,3,4))=1$, and so every time the sequence hits the state $ (1,1,2)$, a gap of minimal size $3$ occurs. Let us see how this is reflected in the formula. $H_1 = \{(0,1,1),(0,4,1) \}$ and $H_2=\{(0,2,3)\}$. There's only one element in $H^2$ and therefore we omit the reference to $b$ in the functions $r,\rho,h$. So $r=2$, $\rho = Q( (0,3,4),(0,4,1)) =\frac 12$, and the expression for $h$ follows. \\ We now compute $\pi^Q$. Let $p=\pi^Q(z^2)$. Since $Q(z^2,z^3)=Q(z^3,z^4)=1$, we have that $p=\pi^Q(z^3)=\pi^Q(z^4)$. Next, $Q(z^4,z^5)=Q(z^4,z^6)=\frac 12$, and so $\pi^Q (z^5)=\pi^Q (z^6) = p/2$. We also observe that \be \pi^Q (z^1)\ =\ \pi^Q (z^1) \lambda_C^{-1}+ \pi^Q (z^5) Q (z^5,z^1) + \pi^Q(z^6) Q (z^6,z^1)\ee Therefore, $\pi^Q(z^1)= \frac{p}{\lambda_C-1}$. Now we have $1 = \frac{p}{\lambda_C-1} + 4p $, so altogether, $p=\frac{\lambda_C-1}{4\lambda_C-3}$, and the expression for the limit of $\mu_n$ follow after some algebra. \begin{proof}[Proof of Theorem \ref{th:gap_dist}] For a real number $x$, let $x_+= \max (x,0)$. We begin with gaps of size $0$: \begin{equation}E^Q_z N_n(0) \ = \ \sum_{j=0}^{n-1} (Z_j(1)-1)_+.\end{equation} The ergodicity of $Z$ under $Q$ implies that \begin{equation} \label{eq:N0} \lim_{n\to\infty} \frac{E^Q_z N_n(0)}{n} \ = \ \sum_{z=(x,j,j')} \pi^Q (z)(x-1)_+=M_{\pi^Q_1}-1+\pi^Q_1(0). \end{equation} Before moving to gaps of larger size, we consider the total number of jumps. We have \begin{align} \nonumber \frac 1n E_z^Q \sum_{k \ge 1} N_n (k) &\ = \ \nonumber \frac 1n E^Q_z \sum_{j=0}^{n-1} {\bf 1}_{\{Z_j(0)>0\}}\\ \label{eq:N12} & \underset{n\to\infty} { \to} 1-\pi^Q_1(0), \end{align} and so from \eqref{eq:N0}, \eqref{eq:N12} \begin{equation} \label{eq:N} \lim_{n\to\infty} \frac 1n E^Q_z N_n \ = \ M_{\pi^Q_1}. \end{equation} We move to calculation of gaps of size $\ge 2$. We will treat gaps of size $1$ last. Let $k\ge 2$. Then \begin{align} \frac 1n E^Q_z N_n(k) &\ = \ \frac 1n E^Q_z \sum_{j=0}^{n-k}{\bf 1}_{\{Z_{j}(1)> 0\}}(\prod_{\ell=1}^{k-1} {\bf 1}_{\{Z_{j+\ell}(1)=0\}} ) {\bf 1}_{\{Z_{j+k}(1) > 0\}}. \end{align} Let $B=\{(0,b,b')\in {\mathcal L}\}$. It therefore follows from the Markov property and ergodicity that \be \lim_{n\to\infty} \frac1n E^Q_z N_n(k)\ =\ \sum_{z^0 \in A} \pi^Q (z^0) f_B(z^0)\ee where for $D\subset {\mathcal L}$ we have \be f_D(z^0)\ =\ \left( \sum_{z^1\in D, \dots,z^{k-1}\in B } \prod_{\ell=1}^{k-1} Q(z^{\ell-1},z^\ell) \right) Q(z^{k-1},A). \ee Letting \begin{align} B_0 &\ = \ \{(0,1,1)\}\nonumber\\ B_1 &\ = \ \{(0,b+1,1)\in {\mathcal L}: b\ge 1,~c_b >0\},\nonumber\\ B_2 &\ = \ \{ (0,b+1,b+2)\in {\mathcal L}: b\ge1,~c_b >0, c_{b+1} =0\}, \mbox{ and}\nonumber\\ B_3 &\ = \ \{(0,b+1,1)\in {\mathcal L}: b \ge 1,~c_b=0,c_{b+1} >0\}, \end{align} we can write \be \sum_{z^0 \in A} \pi^Q (z^0) f_B(z^0)\ =\ \sum_{m=0}^3 \sum_{z^0 \in A} \pi^Q (z^0) f_{B_m} (z^0).\ee Note that $\cup_{m=0}^3 B_m = \{(0,b,b') \in {\mathcal L}: c_{b-1} \ne 0 \mbox{ or } c_{b'} \ne 0\}$, and so this union does not necessarily contain all elements $(0,b,b') \in {\mathcal L}$. However, it does contain all such elements which are accessible from $A$ in one step (and more, whenever $B_3$ is not empty). \\ We now simplify the expression, beginning with the sum over $B_1$. It is important to observe that $B_1$ is the subset of states in $B$ accessible in one step only from $A$, In addition, if $z^1 \in B_1$, then it immediately follows that $z^2=\dots = z^{k-1} = (0,1,1)$, and that allowed transition to $(0,1,1)$ always have probability $\lambda_C^{-1}$. As a result, we have that \be f_{B_1} (z^0)\ =\ Q(z^0, z^1) \lambda_C^{-(k-2)} (1-\lambda_C^{-1}),\ee and thus \be \sum_{z^0 \in A} \pi^Q (z^0) f_{B_1}(z^0)\ =\ \pi^Q (B_1) \nu (k-1) .\ee Next we consider the sum over $B_0$, namely $z^1 = (0,1,1)$. Clearly: \begin{align} \sum_{z^0 \in A} \pi^Q (z^0) f_{(0,1,1)} (z^0) &\ = \ \sum_{z^0\in {\mathcal L}} \pi^Q (z^0) f_{(0,1,1)} (z^0) - \sum_{z^0 \in B }\pi^Q (z^0) f_{(0,1,1)} (z^0). \end{align} Since $(0,1,1)$ is accessible in one step either from $A$ or from states in $z \in B_0 \cup B_1\cup B_3$ and for all such $z$, $Q(z,(0,1,1)) = \lambda_C^{-1}$, it follows that \be \sum_{z^0 \in A} \pi^Q (z^0) f_{(0,1,1)} (z^0)\ = \ \left ( \pi^Q ( (0,1,1) ) -\pi^Q(B_0 \cup B_1 \cup B_3 )\lambda_C^{-1} \right) \nu (k-1).\ee Hence, \be \sum_{z^0 \in A} \pi^Q (z^0) f_{B_0 \cup B_1} (z^0) \ =\ \pi^Q (B_0 \cup B_1\cup B_3 ) (1-\lambda_C^{-1}) \nu (k-1) - \pi^Q (B_3) \nu (k-1).\ee We now consider $z^1 \in B_2$. Suppose then that $z^0 \in A$ and $z^1 \in B_2$ and $Q(z^0,z^1)>0$. Since $z^1 = (0,b+1,b+2)$, it follows that $z^0=(c_b,b,b+1)$ and $c_b >0$. Now if $c_{b+2} =0$, then the only allowed transition form $z^1$ is to $z^2 = (0,b+2,b+3)$. Let $r=r(b)$ and $\rho =\rho(b)$ as defined in the statement of the theorem. Then $z^j=(0,b+j,b+j+1)$ for all $j=1,\dots r$, and we conclude that $Q(z^j,z^{j+1}) =1$ for $j=0,\dots, r$. We continue according the the following two cases. \\ ~\\ \noindent {\bf 1. $r> k-1$.} In this case $Q^k (z^0,A)=0$.\\ \noindent{\bf 2. $r \le k-1$.} Then either \begin{itemize} \item $r=k-1$, in which case $Q^k(z^0,A) = Q( (0,b+r,b+r+1),A)=1-\rho$; or \item $1< r\le k-2$, in which case $z^{r+1} = (0,b+r+1,1)$ and $z^{r+l}= (0,1,1)$ for all $2\le l \le k-1-r$. In particular, since $Q ( (0,1,1) , A) = Q ( (0,b+r+1,1),A)= 1-\lambda_C^{-1}$, we have that \be Q^k(z^0,A) =Q ( (0,b+r,b+r+1), (0,b+r+1,1))\nu (k-r-1).\ee \end{itemize} The only allowed transitions from $(0,b+r,b+r+1)$ to $(x,b+r+1,1)$ are to $x=0,\dots,c_{b+r+1}-1$, all with equal transition probability. Since there are exactly $c_{b+r+1}-\delta_L(b+r+1)$ possible values for $x$, exactly one of which is with $x=0$, letting $\rho (b) = \frac{1}{c_b - \delta_L (b)}$, we have \be Q^k (z^0,A)\ =\ \begin{cases} 0 & k < r (z)+1 \\ 1- \rho (b+r+1) & k = r(z)+1 \\ \rho (b+r+1) \nu (k-r-1) & k> r(z)+1.\end{cases} \ee Summarizing the two cases, we conclude that \be \sum_{z^0 \in A} \pi^Q (z^0) f_{B_2} (z^0)\ =\ \sum_{z^1 \in B_2} h (z^1(2)-1,k).\ee Next, when $z^0\in A$, and $z^1\in B_3$, then $Q(z^0,z^1)=0$. Thus, we have proved \be \sum_{m=0}^3 \pi^Q (z^0) f_{B_m} (z^0)\ = \ \left( (1-\lambda_C^{-1}) \pi^Q (H_1) - \pi^Q (B_3) \right) \nu (k-1) + \sum_{z^0=(0,b+1,b+2) \in B_2} \pi^Q(z^0) h(b-1,k).\ee Let $z' \in B_3$. Then there exists a unique $z^1=(0,b+1,b+2) \in B_2$ such that $z^1 = (0,b,b+1),~z^2 = (0,b+2,b+3),\dots, z^{r(b)} =(0,b+r(b),b+r(b)+1)$ and $z^{r(b)+1}=z'$. Since $Q(z^k,z^{k+1})=1$ for $k=1,\dots,r(b)-1$, it easily follows that $\pi^Q (z') = \pi^Q (z^r) \rho (b)=\pi^Q (z^{r-1}) \rho (b)=\dots = \pi^Q (z^1) \rho(b)$. This shows that $\pi^Q(B_3) = \sum_{\{z^1 =(0,b+1,b+2) \in B_2\} }\pi^Q (z^1) \rho(b)$. Plugging this into the formula above, and noting that $H_1$ in the theorem is $B_0 \cup B_1\cup B_3$ and $H_2$ in the theorem is $B_2$, we obtain \begin{eqnarray}\label{eq:place_holder} \lim_{n\to\infty} \frac 1n E^Q_z N_n(k)&\ =\ & \sum_{m=0}^2 \pi^Q (z^0) f_{B_m} (z^0)\nonumber\\ &\ = \ & (1-\lambda_C^{-1}) \pi^Q (H_1)\nu(k-1) \nonumber\\ & & \ \ \ +\ \sum_{z^0=(0,b+1,b+2) \in H_2} \pi^Q ( z^0) \left ( h(b-1,k) -\rho(b) \nu(k-1))\right). \end{eqnarray} We turn to gaps of size $1$: \begin{align} \frac 1n E^Q_z N_n(1) &\ = \ \frac 1n E^Q_z \sum_{j=0}^{n-1} {\bf 1}_{\{Z_j(1) >0\}}{\bf 1}_{\{Z_{j+1} (1) >0\}}\nonumber\\ &\ = \ \frac 1n \sum_{j=0}^{n-1} E_z^Q{\bf 1}_{\{Z_j(1) > 0\}}E_{Z_j}^Q {\bf 1}_{\{Z_1(1)> 0\}}, \end{align} where in the second line we applied the Markov property. Let \be A \ = \ \{(x,b,b')\in {\mathcal L}: x> 0\}.\ee Ergodicity of $Z$ under $Q$ then gives \begin{align} \label{eq:nice1} \lim_{n\to\infty} \frac 1n E^Q_z N_n(1)\ =\ \sum_{z\in A } \pi^Q(z) Q(z,A) = \pi^Q (A) - \sum_{z \in A^c} \pi^Q (z) Q (z,A). \end{align} Given $z=(0,b,b') \in A^c$, exactly one of the following holds. \begin{itemize} \item $c_b=0$, $b'=b+1$, $c_{b+1}=0$, and then $Q(z,A)=0$. \item $c_b=0$, $b'=b+1$, $c_{b+1}>0$. From the argument in the paragraph above \eqref{eq:place_holder}, and since $Q(z,A) = 1-Q(z,A^c)$ we obtain that \bea \sum_{\{z=(0,b,b+1)\in {\mathcal L}: c_b =0,c_{b+1} =1\}} \pi^Q (z) Q(z,A) &\ =\ & \pi^Q(B_2) - \pi^Q (B_3) \nonumber\\ &=& \sum_{z^0=(0,b+1,b+2) \in H_2} \pi^Q (z^0) (1- \rho(b)).\ \ \ \eea \item $c_b >0$ and then $b'=1$, equivalently, $z \in H_1$, in which case $Q(z,A) = 1-Q(z,(0,1,1))=1-\lambda_C^{-1}$. \end{itemize} Summarizing, \begin{align} \label{eq:nice2} \lim_{n\to\infty} \frac 1n E^Q_z N_n(1) = 1-\pi^Q_1(0) - (1-\lambda_C^{-1})\pi^Q_1(H_1)-\left ( \sum_{z^0=(0,b+1,b+2) \in H_2} \pi^Q (z^0) (1- \rho(b))\right). \end{align} To finish the proof, we need to show that the results continue to hold when considering the measure $Q^n$ instead of $Q$. However, by the Markov property, the expectation under $Q^n$ of $N_n$, and $N_n(k)$ are equal to the expectations of corresponding additive functionals. Therefore it follows from Theorem \ref{th:additive_functionals} that the expectations of $N_n(k)$ and $N_n$ under $Q_n$ are asymptotical equivalent to their expectations with respect to $Q_{\varphi_c}$. The theorem now follows.\end{proof} \begin{proof}[Proof of Theorem \ref{th:weaklim_prob}] We have \bea \{ |\hat \mu_n (A) - \mu_\infty (A) |>\epsilon \} & \ \subset \ & \cup_{k\in A} \{ \left | \hat \mu_n (k) - \mu_\infty (k) \right| > \epsilon\}\nonumber\\ &\ = \ & \cup_{k\in A} \{ \left | N_n (k) -\mu_\infty (k) N_n \right|> \epsilon N_n \}\cup \{N_n =0\}.\eea Since $Q_n (N_n=0) =Q_n (Z_0>0,Z_1=\dots =Z_n=0)\to 0$, we can ignore the event $\{N_n=0\}$. Now for every fixed $k\in A$, we have \be \{ \left | N_n (k) -\mu_\infty (k) N_n \right|> \epsilon N_n \}\subset \{ |N_n(k) - \mu_\infty (k) E_{\pi^Q}^Q N_n | > \epsilon/2 \} \cup \{ | N_n - E_{\pi^Q}^Q N_n| > \epsilon/2\}.\ee Next observe that both $N_n$ and $N_n(k)$ are additive functionals for the process $Z^k = (Z^k_n:k\in\ensuremath{\mathbb{Z}}_+)$, where $Z^k_n =(Z_n,Z_{n+1},\dots,Z_{n+k})$, and so we can consider $N_n(k)$ and $N_n$ as additive functionals of $Z^k$. Letting $ \varphi' _c(z^0,z^1,\dots,z^k)= \varphi_c (z^0)$, and $\tilde \varphi_c'(z^0,\dots,z^k)$, the distribution of $Z_0,\dots,Z_k$ under $Q_{\tilde\varphi_c}$, then if as in Definition \ref{def:additive} we define \be Q'_{n,k} (A)\ =\ \frac{E_{\tilde \varphi_c'}^Q \left (\frac{{\bf 1}_{A}}{\varphi_c'(Z_n^k)}\right)}{E^Q_{\tilde \varphi_c'}\left ( \frac{1}{\varphi_c'(Z_n^k)}\right)},\ee it follows that the restriction of $Q'_{n,k}$ to events generated by $Z_0,\dots,Z_n$ coincides with $Q_n$. In particular, the distribution of the additive functionals $N_n$ and $N_n(k)$ for $Z^k$ under $Q_{n,k'}$ coincides with their distribution under $Q_n$. From the variance estimate \eqref{eq:sec_mom} in Theorem \ref{th:additive_functionals} applied to these additive functionals under $Q_{n,k}'$, we conclude that \be Q_n ( \{ |N_n(k) - \mu_\infty (k) E_{\pi^Q}^Q N_n | > \epsilon/2 \} ) =O(n^{-1})\mbox{ and }Q_n (\{ | N_n - E_{\pi^Q}^Q N_n| > \epsilon/2\}) = O (n^{-1}).\ee Therefore if $A$ is finite, we obtain that \be \label{eq:weak_conv} \lim_{n\to\infty} Q_n ( |\hat \mu_n (A) - \mu_\infty (A) |>\epsilon )=0.\ee Now if $A$ is infinite, letting $A_M = A \cap \{0,\dots,M\}$, we observe that \begin{align} \nonumber | \hat \mu_n (A) - \mu_\infty (A) |&\ = \ |\hat \mu_n (A_M) - \mu_\infty (A_M) | + \hat \mu_n (\{M+1,\dots\}) + \mu_\infty (\{M+1,\dots \})\nonumber\\ &\ \le\ |\hat \mu_n (A_M) - \mu_\infty (A_M) | + \hat\mu (\{M+1,\dots\})+ \mu_\infty (\{M+1,\dots\}). \end{align} Fix $\epsilon$, and let $M$ be such that $\mu_{\infty}(\{M+1,\dots\})< \epsilon$. Thus for $n$ large enough, \be \{ | \hat \mu_n (A) - \mu_\infty (A) |>5\epsilon \}\subset \{ |\hat \mu_n (A_M) - \mu_\infty (A_M) | > 2\epsilon\}\cup \{ \hat \mu_n (\{M+1,\dots\}) >2\epsilon\}.\ee The measure of the first event on the right-hand side tends to $0$ as $n\to\infty$ by \eqref{eq:weak_conv}. As for the second event, it is equal to the event $\{\hat \mu_n (\{0,\dots,M\}) <1-2\epsilon\}$. However, since, again by \eqref{eq:weak_conv} $\mu ( |\hat \mu_n (\{0,\dots,M\})- \mu_\infty (\{M+1,\dots\})|> \epsilon/2)$ tends to $0$, it follows that $Q_n ( \{ \hat \mu_n (\{0,\dots,M\}> 1-3\epsilon /2)$ tends to $1$. But this event is $\{ \hat \mu_n (\{M+1,\dots,\})<3 \epsilon /2\}$, and so $Q_n ( \hat \mu_n (\{M+1,\dots,\}) > 2\epsilon) $ tends to $0$ as well. The result now follows.\end{proof} \subsection{Maximal Gap} Next we consider the maximal gap $M_n$, defined as \begin{equation} M_n \ = \ \sup\{k\in \ensuremath{\mathbb{Z}}_+: N_n(k) >0\}.\end{equation} Although we can prove the results at the same level of generality as in the previous section, we prefer to keep the expressions cleaner and simpler, and will assume throughout this section that $c_1,\dots,c_L >0$. \\ Our analysis is based on a renewal structure we now describe. We refer to the gaps of length $k\ge 2$ as ``long gaps", and denote the lengths of the long gaps, indexed by order of appearance, by $(R_j:j\in\mathbb{N})$. Observe that any long gap is followed by a (possibly empty) sequence of length-$0$ and length-$1$ gaps, independent of $k$, and is then followed again by an independent long gap. The number of the small gaps is bounded above by $(L-1)+\sum_i (c_i-1)=(\sum_i c_i)-1$, as the first summand bounds the number of length-$1$ gaps and the second summand bounds the number of length $0$-gaps. Let $T_m$ denote the first time exactly $m$ long gaps are completed, $ m(n) = \sup\{m : T_m \le n\}$. Observe that a long gap is completed whenever the digit zero is followed by a nonzero digit. Therefore \begin{equation}m(n) \ = \ \sum_{j=0}^{n-1} {\bf 1}_{0}(Z_j(1)){\bf 1}_{\{Z_{j+1}(1)> 0\}}.\end{equation} From the Markov property, \begin{equation} E^Q m(n) \ = \ \sum_{j=0}^{n-1} E^Q \left ( {\bf 1}_{0} (Z_j(1))Q_{Z_j}(Z_1(1)> 0)\right),\end{equation} Letting $A= \{z= (x,b,b') \in{\mathcal L}: x>0\}$, and repeating a similar computation as in the proof of the case $k=1$ in Theorem \ref{th:gap_dist}, it follows that \begin{align} \lim_{n\to\infty} \frac{1}{n} E^Q m(n) \ &\ = \ \ \sum_{z\in A^c} \pi^Q(z)Q(z,A)\ = \pi^Q(A^c) - \sum_{z \in A} \pi^Q (z) Q(z,A) \nonumber\\ &\ = \ (1-\pi^Q(0))- (1- \pi^Q(0))+(1-\lambda_C^{-1}) \pi^Q(0),\end{align} where the last equality follows from \eqref{eq:nice1} and \eqref{eq:nice2}. Also, by the renewal theorem \cite[Theorem 2.4.6]{durrett} \begin{equation} \label{eq:renewal} \lim_{n\to\infty} \frac{m(n)}{n} \ = \ \alpha,~Q\mbox{-a.s.}, \end{equation} where $\alpha = 1 / E_{\rho}^Q T_1$ and $\rho$ is the uniform distribution on $c_1$ elements: $(x,1,1),~1<x<c_1$ and $(c_1,1,2)$. The limit above also holds in $L^1(Q)$, as $m(n)\le n$. Consequently \begin{equation}\alpha \ = \ \pi^Q_1(0)\left(1-\frac{1}{\lambda_C}\right).\end{equation} To state our result we need to introduce some additional assumption. We say that a sequence $(n_k:k\in\mathbb{N})$ of natural numbers tending to $\infty$ satisfies the {\it spacing condition} with respect to $\alpha$ and $q$ if \begin{equation}\label{eq:spacingcondition}\liminf_{k\to\infty} \inf_{z\in \ensuremath{\mathbb{Z}}_+} \left| \frac{ \ln (n_k \alpha)}{\ln \frac1q} -z\right|\ >\ 0.\end{equation} Roughly speaking, this means that $n_k\alpha$ is eventually uniformly far from integer powers of $1/q$ in some normalized sense. \begin{theorem} \label{th:maxgap} Assume $c_1c_2\cdots c_L>0$. Then for every $k\in \ensuremath{\mathbb{Z}}$, \begin{equation} \lim_{n\to\infty} Q_n\left(M_n \ \le \ \left\lfloor \frac{\ln n \pi_1(0)(1- \frac{1}{\lambda_C})}{\ln \lambda_C}\right\rfloor +k\right) \ = \ e^{-\lambda_C^{-(k-2)}},\end{equation} when the limit is taken along any sequence satisfying the spacing condition \eqref{eq:spacingcondition} with respect to $\alpha = \pi_1(0)(1- \frac 1 {\lambda_C})$ and $q=\frac1{\lambda_C}$. \end{theorem} \begin{example} For the standard Zeckendorf decomposition, $\lambda_C = \phi$ and $\pi_1(0)= \frac{\phi+1}{\phi+2}$. This gives: \begin{equation}\lim_{n\to\infty} Q_n\left(M_n \le \left\lfloor \frac{\ln n- \ln (\phi+2) }{\ln \phi }\right\rfloor +k\right) \ = \ e^{-\phi^{-(k-2)}}.\end{equation} \end{example} \begin{proof}[Proof of Theorem \ref{th:maxgap}] To prove the theorem, we need to recall some facts on the maximum of negative geometric random variables. Let $\G$ be a negative geometric random variable with parameter $p\in (0,1)$. That is, for $k \in \ensuremath{\mathbb{Z}}_+$, $P( \G \ge k) = q^k$ where $q=1-p$. Let $G$ be negative geometric with parameter $p$. That is, $\G$ takes values in $\ensuremath{\mathbb{Z}}_+$, and $P(\G\ge k) = q^k$, where $q=1-p$. We denote this distribution by ${\rm Geom}^-(p)$. Let $(\G_k:k\in\mathbb{N})$ be IID ${\rm Geom}^-(p)$-distributed random variables, and let $M^\G_m = \max_{k\le m} \G_k$. Then $ P( M^\G_m \le j) = (1-q^j)^m$. For each $m\in \mathbb{N}$, let $\delta_m$ be chosen so that $\frac{\ln m \delta_m}{\ln 1/q} = \left\lfloor \frac{\ln m}{\ln 1/q}\right\rfloor$. Observe then that $\delta_m \in (q,1]$. From this we obtain that for any $k\in\ensuremath{\mathbb{Z}}$, \begin{equation} \label{eq:max_geom} P\left(M_m^\G \le \left\lfloor \frac{\ln m}{\ln 1/q}\right\rfloor + k\right) \ = \ \left(1- \frac{q^k}{m\delta_m}\right)^m\underset{m\to\infty}{\to} e^{-q^k}. \end{equation} We return to the proof. Fix some sequence satisfying the spacing condition. Abusing notation, we will refer to a generic element in the sequence as $n$. Observe that if we choose $\G_j=R_j-2$, then $(\G_j:j\in \mathbb{N})$ is an IID sequence of ${\rm Geom}^-(p)$ random variables with $p=1-\lambda_C^{-1}$. In particular, for every $m$ \begin{equation} M_{T_m} \ = \ M^\G_m +2.\end{equation} Clearly $T_{m(n)} \le n$, but also by the law of large numbers and \eqref{eq:renewal} \begin{equation} \frac{ T_{m(n)}}{ n }\ = \ \frac{ T_{m(n)}}{m(n)} \times \frac{ m(n) }{n} \underset{n\to\infty}{\to} 1,~Q\mbox{-a.s.}\end{equation} From \eqref{eq:renewal} we can find $\epsilon_n>0$ with $\lim_{n\to\infty} \epsilon_n = 0$ and satisfying \begin{equation}Q\left( \frac{m(n)}{n} \in [1-\epsilon_n,1+\epsilon_n]\alpha\right) \ \underset{n\to\infty} {\to}\ 1.\end{equation} Observe then that \begin{eqnarray} Q\left( M_n \le \left\lfloor \frac{\ln n \alpha}{\ln \frac1q}\right\rfloor + k\right) &\ \ge\ & Q\left(M_n \le \left\lfloor \frac{\ln n \alpha}{\ln \frac1q}\right\rfloor + k, 0< m(n) \le (1+\epsilon_n) n \alpha\right)\nonumber\\ &\ge & Q\left(M^\G_{\left\lfloor(1+\epsilon_n)n \alpha \right\rfloor} \le \left\lfloor \frac{\ln n \alpha}{\ln \frac1q}\right\rfloor + k-2\right) - Q (m(n) \nonumber\\ &>& (1+\epsilon_n) n \alpha)-Q(m(n)=0). \end{eqnarray} The last two terms on the righthand side tend to $0$. In addition, since $\ln (n(1+\epsilon_n) \alpha) - \ln (n \alpha)\underset{n\to\infty} {\to} 0$, it follows from the spacing condition that for all $n$ large enough,$ \left\lfloor \frac{ \left\lfloor(1+\epsilon_n)n \alpha \right\rfloor}{\ln \frac 1q}\right\rfloor= \left\lfloor \frac{\ln n \alpha}{\ln \frac1q}\right\rfloor$. It then follows from \eqref{eq:max_geom} that \begin{equation}\liminf_{n\to\infty} Q\left(M_n \le \left\lfloor \frac{\ln n \alpha}{\ln \frac1q}\right\rfloor + k\right)\ \ge\ e^{-q^{k-2}}.\end{equation} We turn to the upper bound. \begin{eqnarray} Q\left( M_n \le \left\lfloor \frac{\ln n \alpha}{\ln \frac1q}\right\rfloor + k\right) &\ \le\ & Q\left(M_n \le \left\lfloor \frac{\ln n \alpha}{\ln \frac1q}\right\rfloor + k-2, m(n) \ge (1-\epsilon_n) n \alpha\right)\nonumber\\ & & \ \ \ + \ Q\left(m(n) < (1-\epsilon_n)n \alpha\right)\nonumber\\ & \le & Q\left( M^\G_{\lceil (1-\epsilon_n)n \alpha\rceil} \le \left\lfloor \frac{\ln n \alpha}{\ln \frac1q}\right\rfloor + k\right) +o(1). \end{eqnarray} The same argument as before shows that for $n$ large enough, $\left\lfloor \frac{\ln \lceil(1-\epsilon_n)n \alpha \rceil }{\ln \frac 1q}\right\rfloor =\left\lfloor\frac{\ln n \alpha}{\ln \frac1q}\right\rfloor$, and so \begin{equation} \limsup_{n\to\infty} Q\left( M_n \le \left\lfloor \frac{\ln n \alpha}{\ln \frac1q}\right\rfloor + k\right)\ \le \ e^{-q^{k-2}}.\end{equation} Summarizing, \begin{equation} \label{eq:QM_lim} \lim_{n\to\infty} Q\left(M_n \le \left\lfloor \frac{\ln n \alpha}{\ln \frac1q}\right\rfloor + k\right)\ =\ e^{-q^{k-2}}. \end{equation} It remains to convert the result to $Q_n$. Let $A_n =\{M_n \ge \left\lfloor \ln \ln n\right\rfloor \}$. Then $Q(A_n) \underset{n\to\infty}{\to} 1$. Let $b_n = \left\lfloor \ln \ln n \right\rfloor$. Then as $n-b_n = n (1 +o(1))$, we conclude that the sequence $n- b_n$ also satisfies the spacing condition. Furthermore, for sufficiently large $n$, $\left\lfloor \frac{\ln (n -b_n) \alpha } {\ln 1/q} \right\rfloor = \left\lfloor \frac{\ln n \alpha } {\ln 1/q} \right\rfloor$. Thus, from \eqref{eq:QM_lim} \begin{equation} \lim_{n\to\infty} Q\left( M_{n- b_n} \le \left\lfloor \frac{\ln n \alpha}{\ln \frac 1q} \right\rfloor + k\right) \ = \ e^{-q^{k-2}}.\end{equation} Letting $B_n = \{M_{n- b_n} \le \left\lfloor \frac{\ln n \alpha}{\ln \frac 1q} \right\rfloor + k\}$, it follows from the Markov property and the ergodicity of $Z$ that \begin{align} E^Q \left ( {\bf 1}_{B_n} \frac{1}{\varphi_c(Z_n)}\right) &\ = \ E^Q \left ( {\bf 1}_{B_n} E_{X_{n-b_n}} \frac{1}{\varphi_c(X_{b_n})} \right) \nonumber\\ \label{eq:ergo_pass} &\ = \ E^Q \left ( {\bf 1}_{B_n} E_{\pi^Q} \frac{1}{\varphi_c}\right) + o(1)=Q(B_n)+o(1). \end{align} Now \begin{align} Q\left( M_n\le \left\lfloor \frac{\ln n \alpha}{\ln \frac 1q} \right\rfloor + k, \frac{1}{\varphi_c(X_n)}\right) &\ \le\ Q\left( {\bf 1}_{B_n}, E^Q_{X_{n-b_n}} \frac{1}{\varphi_c(X_{b_n})}\right) \nonumber\\ & \ = \ Q(B_n) E_{\pi^Q} \frac{1}{\varphi_c}+ o(1), \end{align} and so \begin{equation}\limsup_{n\to\infty} Q_n\left(M_n \le \left\lfloor \frac{ \ln n\alpha}{\ln 1/q}\right\rfloor+k\right) \ \le\ e^{-q^{k-2}}.\end{equation} We turn to the lower bound. Observe that $M_n > M_{n-b_n}$ only if one of the last $b_n+1$ long gaps among the first $m(n)$ is maximal. Fix $c>0$, then for all $n$ large enough, depending on $c$ and on the event $\{M_n > c \ln n\}$, those maximal gap among the last $b_n+1$ must begin before $n-b_n$ (because otherwise it will have length at most $b_n < c \ln n$) and end after $n-b_n$ (otherwise already included in $M_{n-b_n}$). That is, \begin{equation}\{M_n > M_{n-b_n} \} \cap \{M_n > c \ln n\}\subset \{ \max_{j=1,\dots, m(n-b_n)+1} \G_j \ = \ \G_{m(n-b_n)+1}\}.\end{equation} Denote the event on the right-hand side by $C_n$. We have that \begin{align} Q(C_n) & \ \le\ Q(C_n, m(n) \in ( (1-\epsilon) n \alpha, (1+\epsilon)n \alpha)) + o(1)\nonumber\\ &\ \le\ 2 \epsilon n \alpha \times \frac{1}{(1-\epsilon)n \alpha}+o(1)\underset{n\to\infty} { \to} \frac{2\epsilon}{1-\epsilon}. \end{align} Since $\epsilon$ is arbitrary, we conclude that $Q(C_n) \underset{n\to\infty}{\to}0$. Hence \begin{align} E^Q\left( M_n > \left\lfloor\frac{\ln n \alpha}{\ln \frac 1q}\right\rfloor + k,\frac{1}{\varphi_c(Z_n)}\right) &\ \le\ Q\left(M_{n-b_n} > \left\lfloor \frac{ \ln n \alpha}{\ln \frac 1q}\right\rfloor + k, C_n^c,\frac{1}{\varphi_c(X_n)}\right)+Q(C_n)\nonumber\\ &\ \le\ Q\left(M_{n-b_n} > \left\lfloor \frac{ \ln (n-b_n) \alpha}{\ln \frac 1q}\right\rfloor + k,\frac{1}{\varphi_c(Z_n)}\right)+o(1). \end{align} The remainder of the proof is identical to the argument presented in \eqref{eq:ergo_pass}, with the obvious changes. This gives the lower bound \begin{equation} \liminf_{n\to\infty}Q^n\left(M_n \le \frac{\left\lfloor \ln n \alpha\right\rfloor}{\ln \frac 1q} + k\right)\ \ge\ e^{-q^{k-2}},\end{equation} thus completing the proof.\end{proof} \begin{comment} \section{Renewal Approach to Gaps} \label{sec:renewal} The structure of the gaps corresponds to a classical probabilistic model, known as discrete-time renewal. We now describe the ingredients. Let $N_n(k)$ denote the number of gaps of size $k$ up to time $n$. Our analysis of $N_n(k)$ rests on the observation that for fixed $k$, the function $n\to N_n(k)$ is a renewal process. To see this we need some preparation. Let $J_0= \inf\{n\in \ensuremath{\mathbb{Z}}_+: X_n =0\}$, and continue inductively by letting $T_{n+1} = \inf\{n > J_n: X_n \ne 0\mbox{ and } X_{n-1}=X_{n-2} = \dots =X_{n-(k-1)}=0\}$ and $J_{n+1} = \inf\{n> T_{n+1}: X_n = 0\}$. Observe that the first gap of size $k$ ends at time $T_1$, the second gap ends at time $T_2$, etc. Furthermore, the random variables $\{J_{i+1}-J_i:i\in \ensuremath{\mathbb{Z}}_+\}$ are IID. Let $R_n = \sup\{ i: J_i \le n\}$. Then $(R_n:n\in\ensuremath{\mathbb{Z}}_+)$ is the renewal process corresponding to the IID sequence $(J_{i+1}-J_i:i\in\ensuremath{\mathbb{Z}}_+)$. Clearly, $N_n(k)-1 \le R_n \le N_n(k)$. We now consider the asymptotic behavior of $R_n$. \begin{lemma} \label{lem:R_n} Let $\rho>0$. Then as $n\to\infty$, $ E e^{\rho R_n} \sim \frac{(1-E z_0^{J_1})E z_0^{J_1}}{(1-z_0) E J_1 z_0^{J_1}} z_0^{-n}$ where $z_0\in (0,1)$ is the unique solution in $(0,1)$ to the equation $ E z^{J_1}=e^{-\rho}$. \end{lemma} We prove this through the $z$-transform defined as follows: \begin{equation} \phi(\rho, z) \ = \ \sum_{n=0}^\infty E e^{\rho R_n} z^n.\end{equation} Since $R_n \le n$, $\phi(\rho,\cdot)$ is analytic in some neighborhood of the origin. Furthermore, \begin{equation} E e^{\rho R_n} \ = \ \sum_{m=0}^\infty e^{\rho m} P(R_n = m) \ = \ \sum_{m=0}^\infty e^{\rho m}( P( J_m \le n ) - P(J_{m+1} \le n)).\end{equation} Therefore \begin{equation} \label{eq:phi_defined} \phi(\rho,z) \ = \ \sum_{m=0}^\infty e^{\rho m } \left (\sum_{n=0}^\infty z^n \left ( P(J_{m} \le n) - P(J_{m+1} \le n) \right)\right). \end{equation} Since $J_{m+1}=J_m + (J_{m+1}-J_m)$, we have that \begin{align} \sum_{n=0}^\infty z^n P(J_{m+1}\le n ) &\ = \ \sum_{n=0}^\infty \sum_{k=0}^n z^{n-k} P(J_m \le n-k)z^k P(J_1 = k) \nonumber\\ &\ = \ \sum_{k=0}^\infty z^k P(J_1=k) z^{k} \sum_{l=0}^\infty z^l P(J_m \le l)\nonumber\\ &\ = \ E z^{J_1} \sum_{n=0}^\infty z^n P(J_m \le n). \end{align} By induction \begin{equation} \sum_{n=0}^\infty z^n P(J_m \le n) \ = \ (E z^{J_1} )^{m-1} H(z),~m\ge 1,\end{equation} where $H(z) = \sum_{n=0}^\infty z^n P(J_1 \le n)= \frac{1}{1-z} - \sum_{n=0}^\infty z^n P(J_1 > n)$. Plugging this back into \eqref{eq:phi_defined} we obtain \begin{align} \phi(\rho,z) &\ = \ \frac{1}{1-z} - H(z) + \nonumber\\ &\ \ \ \ \ \ +\ H(z) \sum_{m=1}^\infty e^{\rho m} (E z^{J_1})^{m-1}(1- E z^{J_1}) \nonumber\\ &\ = \ \umessclean{F_1(z)}{\sum_{n=0}^\infty z^n P(J_1 > n)}+ \umessclean{F_2(\rho,z)}{ \frac{H(z) e^{\rho}(1-E z^{J_1})}{1-e^\rho E z^{J_1}}}. \end{align} As is easy to see, $J_1$ has exponential tails, and as a result the radius of convergence of $F_1$ is strictly greater than $1$. Next, $H$ is analytic on the open unit disk. In particular, the only poles of $\phi(\rho,\cdot)$ in the open unit disk are the solutions to $Ez^{J_1} = e^{-\rho}$. There exists a unique real solution $z_0\in (0,1)$. Call this solution $z_0$. It is clear that there does not exist any other solution $z$ with $|z|\le z_0$, because $|Ez^{J_1}|\le E|z^{J_1}|= E |z|^{J_1} \le E z_0^{J_1}$, and the first inequality is an equality if and only if $z^{J_1}$ has a constant argument a.s. We further observe that $\frac{1-e^{\rho} E z^{J_1}}{z-z_0}$ is analytic (except for the removable singularity at $z=z_0$), and is non-vanishing in some neighborhood of $z_0$, because its limit as $z\to z_0$ is $-e^\rho E J_1 z_0^{J_1-1}$, which is strictly negative. Therefore $(z-z_0)F_2 (\rho,z)$ is analytic on the disk $|z|<z_0(1+\epsilon)$ for some $\epsilon>0$, and we can write \begin{align} \phi (\rho,z) &\ = \ -\umessclean{c_{-1}}{\frac{ H(z_0)(1-E z_0^{J_1})}{ E J_1 z_0^{J_1-1}}}(z-z_0)^{-1} + F_3(\rho,z)\nonumber\\ & \ = \ \sum_{n=0}^{\infty} c_{-1} z_0^{-1-n} z^n + F_3(\rho,z). \end{align} where $F_3(\rho,z)$ is analytic on the disk $|z|<(1+\epsilon) z_0$. In particular, the radius of convergence of $\sum_{n=0}^\infty a_n z^n = \sum_{n=0}^\infty (E e^{\rho R_n} -c_{-1} z_0^{-n-1}) z^n$ is bounded below by $(1+\epsilon)z_0$, and this implies that $\lim_{n\to\infty}|a_n| / (c_{-1} z_0^{-n-1})=0$, which proves the lemma. \end{comment}

https://arxiv.org/abs/1310.1423

On lattice sums and Wigner limits

Wigner limits are given formally as the difference between a lattice sum, associated to a positive definite quadratic form, and a corresponding multiple integral. To define these limits, which arose in work of Wigner on the energy of static electron lattices, in a mathematically rigorous way one commonly truncates the lattice sum and the corresponding integral and takes the limit along expanding hypercubes or other regular geometric shapes. We generalize the known mathematically rigorous two and three dimensional results regarding Wigner limits, as laid down in [Analysis of certain lattice sums, D. Borwein, J. M. Borwein, and R. Shail, 1989], to integer lattices of arbitrary dimension. In doing so, we also resolve a problem posed in Chapter 7 of [Lattice Sums: Then and Now, J. M. Borwein, L. Glasser, R. McPhedran, J. G. Wan, and I. J. Zucker, 2013].For the sake of clarity, we begin by considering the simpler case of cubic lattice sums first, before treating the case of arbitrary quadratic forms. We also consider limits taken along expanding hyperballs with respect to general norms, and connect with classical topics such as Gauss's circle problem. An appendix is included to recall certain properties of Epstein zeta functions that are either used in the paper or serve to provide perspective.

\section{Introduction}\label{sec:intro} Throughout this paper, $Q(x)=Q (x_1, \ldots, x_d)$ is a positive definite quadratic form in $d$ variables with real coefficients and determinant $\Delta > 0$. As proposed in {\cite[Chapter 7]{latticesums}}, we shall examine the behaviour of \[ \sigma_N (s) := \alpha_N (s) - \beta_N (s) \] as $N \rightarrow \infty$, where $\alpha_N$ and $\beta_N$ are given by \begin{eqnarray} \alpha_N (s) & := & \sum_{n_1 = - N}^N \cdots \sum_{n_d = - N}^N \frac{1}{Q (n_1, \ldots, n_d)^s}, \label{eq:alpha}\\ \beta_N (s) & := & \int_{- N - 1 / 2}^{N + 1 / 2} \cdots \int_{- N - 1 / 2}^{N + 1 / 2} \frac{\mathrm{d} x_1 \cdots \mathrm{d} x_d}{Q (x_1, \ldots, x_d)^s} . \label{eq:beta} \end{eqnarray} As usual, the summation in \eqref{eq:alpha} is understood to avoid the term corresponding to $(n_1, \ldots, n_d) = (0, \ldots, 0)$. If $\op{Re} s > d / 2$, then $\alpha_N (s)$ converges to the Epstein zeta function $\alpha (s) = Z_Q (s)$ as $N \rightarrow \infty$. A few basic properties of $Z_Q$ are recollected in Section \ref{sec:basics}. On the other hand, each integral $\beta_N (s)$ is only defined for $\op{Re} s < d / 2$. \textit{A priori} it is therefore unclear, for any $s$, whether the \emph{Wigner limit} $\sigma (s) := \lim_{N \rightarrow \infty} \sigma_N (s)$ should exist. In the sequel, we will write $\sigma_Q (s)$ when we wish to emphasize the dependence on the quadratic form $Q$. For more on the physical background, which motivates the interest in the limit $\sigma (s)$, we refer to Section \ref{sec:phys} below. In the case $d = 2$, it was shown in {\cite[Theorem 1]{latticesums-bbs}} that the limit $\sigma (s)$ exists in the strip $0 < \op{Re} s < 1$ and that it coincides therein with the analytic continuation of $\alpha (s)$. Further, in the case $d = 3$ with $Q (x) = x_1^2 + x_2^2 + x_3^2$, it was shown in {\cite[Theorem 3]{latticesums-bbs}} that the limit $\sigma (s)$ exists for $1 / 2 < \op{Re} s < 3 / 2$ as well as for $s = 1 / 2$. However, it was determined that $\sigma (1 / 2) - \pi / 6 = \lim_{\varepsilon \rightarrow 0^+} \sigma (1 / 2 + \varepsilon)$. In other words, the limit $\sigma (s)$ exhibits a jump discontinuity at $s = 1 / 2$. It is therefore natural to ask in what senses the phenomenon, observed for the cubic lattice when $d = 3$, extends both to higher dimensions and to more general quadratic forms. We largely resolve the following problem which is a refinement of one posed in the recent book {\cite[Chapter 7]{latticesums}}. \begin{problem}[Convergence] \label{prob}For dimension $d >1$, consider $\sigma_N$ as above. \begin{quote}Show that the limit $\sigma(s) := \lim_{N \rightarrow \infty} \sigma_N (s)$ exists in the strip $d / 2 - 1 < \op{Re} s < d / 2$. Does the limit exist for $s = d / 2 - 1$? If so, is the limit discontinuous at $s = d / 2 - 1$, and can the height of the jump discontinuity be evaluated?\end{quote} \end{problem} In Proposition \ref{prop:sigmastrip}, we show that the limit indeed exists in the strip suggested in Problem \ref{prob}. In the case of $Q (x) := x_1^2 + \cdots + x_d^2$, we then show in Theorem \ref{thm:jump} that $\sigma(s)$ also converges for $s = d / 2 - 1$. As in the case $d = 3$, we find that $\sigma(s)$ has a jump discontinuity, which we evaluate in closed form. In Theorem \ref{thm:jumpx} we extend this result less explicitly to arbitrary positive definite quadratic forms $Q$. \subsection{Motivation and physical background}\label{sec:phys} As described in {\cite[Chapter 7]{latticesums}}: {\small \begin{quote} In 1934 Wigner introduced the concept of an electron gas bathed in a compensating background of positive charge as a model for a metal. He suggested that under certain circumstances the electrons would arrange themselves in a lattice, and that the body-centred lattice would be the most stable of the three common cubic structures. Fuchs (1935) appears to have confirmed this in a calculation on copper relying on physical properties of copper. The evaluation of the energy of the three cubic electron lattices under precise conditions was carried out by Coldwell-Horsefall and Maradudin (1960) and became the standard form for calculating the energy of static electron lattices. In this model electrons are assumed to be negative point charges located on their lattice sites and surrounded by an equal amount of positive charge uniformly distributed over a cube centered at the lattice point. \end{quote}} In three dimensions, this leads precisely to the problem enunciated in the previous section. That is, Wigner, when $d=3$, $s=1/2$ and $Q(x)=x_1^2+x_2^2+x_3^2$, proposed considering, after appropriate renormalization, the entity \begin{equation} \sigma(s):= \sum_{n_1 = - \infty}^\infty \cdots \sum_{n_d = - \infty}^\infty \frac{1}{Q (n_1, \ldots, n_d)^s}- \int_{- \infty}^{\infty} \cdots \int_{-\infty}^{\infty} \frac{\mathrm{d} x_1 \cdots \mathrm{d} x_d}{Q (x_1, \ldots, x_d)^s} . \label{eq:wigner} \end{equation} As a physicist Wigner found it largely untroubling that in \eqref{eq:wigner} the object of study $\sigma(s)$ never makes unambiguous sense. Nor even does it point the way to formalize its mathematical content. The concept is thus both natural physically and puzzling mathematically for the reasons given above of the non-convergence of the integral whenever the sum converges. The best-behaved case is that of two dimensions, which is also physically meaningful if used to consider planar lamina. In \cite[\S3]{latticesums-bbsz} a `meta-principle' was presented justifying the evaluation of $\sigma$ as the analytic continuation of $\alpha=Z_Q$. This was followed by a discussion and analysis of various important hexagonal and diamond, cubic and triangular lattices in two and three space \cite[\S4]{latticesums-bbsz}. In particular, the values of $\alpha$ obtained agreed with values in the physical literature whenever they were known. It was this work which led to the analysis in \cite{latticesums-bbs}. The entirety of \cite[Chapter 7]{latticesums} is dedicated to the analysis of such `electron sums' in two and three dimensions. While we make no direct claim for the physical relevance of the analysis with $d>3$, the delicacy of the mathematical resolution of Problem \ref{prob} is certainly informative even just for general forms in three dimensions. \subsection{Structure of the paper} The remainder of the paper is organized as follows. In Section \ref{sec:basics}, we establish some basic properties of $\alpha_N$ and $\beta_N$. Then, in Section \ref{sec:wigconv}, we establish convergence in the strip for a general quadratic form (Proposition \ref{prop:sigmastrip}). Next, in Section \ref{sec:wigjump}, we consider convergence on the boundary of the strip. In particular, we explicitly evaluate the jump discontinuity in the cubic case (Theorem \ref{thm:jump}). In the non-cubic case the same phenomenon is established, though the corresponding evaluation of the jump is less explicit (Theorem \ref{thm:jumpx}). In Section \ref{sec:alt}, we consider other limiting procedures which replace limits over expanding cubes by more general convex bodies. The paper concludes with a brief accounting of the underlying theory of cubic lattice sums in Appendix \ref{sec:cubiclatticesums}. For more details the reader is referred to \cite{latticesums} and the other cited works. \section{Basic analytic properties}\label{sec:basics} Any quadratic form $Q(x)=Q (x_1, \ldots, x_d)$ can be expressed as \begin{equation}\label{eq:Q} Q (x) = Q_A (x) := x^T A x = \sum_{1 \leq i, j \leq d} a_{i j} x_i x_j, \end{equation} for a matrix $A = (a_{i j})_{1 \leq i, j \leq d}$ which is symmetric (that is, $a_{i j} = a_{j i}$ for all $1 \leq i, j \leq d$). If $Q$ is positive definite, then $A$ is a positive definite matrix of determinant $\Delta = \det(A) > 0$. A basic property of a positive definite matrix $A$, given in most linear algebra texts, is that it can be decomposed as $A=L^TL$, where $L$ is a non-singular matrix. This property is used implicitly when making coordinate transformations as in \eqref{eq:intQQ} and in the proof of Lemma \ref{lem:intQAB} below. As indicated in the introduction, the limit of $\alpha_N (s)$ is the Epstein zeta function \begin{equation} \alpha(s) := Z_Q (s) := \sum_{n_1, \ldots, n_d}' \frac{1}{Q (n_1, n_2, \ldots, n_d)^s} \label{eq:def:epsteinzeta} . \end{equation} Standard arguments show that $Z_Q (s)$ is an analytic function in the domain $\op{Re} s > d / 2$. In fact, see {\cite{epstein-zeta}} or {\cite[Chapter 2 or 8]{berndtknopp-hecke}}, the Epstein zeta function $Z_Q (s)$ has a meromorphic continuation to the entire complex plane and satisfies the functional equation \begin{equation} \frac{Z_Q (s) \Gamma (s)}{\pi^s} = \frac{1}{\sqrt{\Delta}} \frac{Z_{Q^{- 1}} (d / 2 - s) \Gamma (d / 2 - s)}{\pi^{d / 2 - s}}, \label{eq:epsteinfunc} \end{equation} where $ Q (x) = x^T A x$ and $Q^{-1} (x) = x^T A^{-1} x.$ Moreover, the only pole of $Z_Q (s)$ occurs at $s = d / 2$, is simple, and has residue \begin{equation} \op{res}_{d / 2} Z_Q (s) = \frac{1}{\sqrt{\Delta}} \frac{\pi^{d / 2}}{\Gamma (d / 2)} . \label{eq:epsteinres} \end{equation} A particularly important special case of Epstein zeta functions is that of cubic lattice sums, which correspond to the choice $Q (x) = x_1^2 + \cdots + x_d^2$. In Appendix \ref{sec:cubiclatticesums}, we recall some of their basic properties, which provide further context for the questions discussed herein. These remarks made, it is natural to begin our investigation of Problem \ref{prob} by discussing some related properties of the limit of $\beta_N (s)$. In the sequel, we use the notation $\|x\|_{\infty} := \max (|x_1 |, \ldots, |x_d |)$ for vectors $x = (x_1, \ldots, x_d) \in \mathbb{R}^d$. \begin{proposition} \label{prop:betares}Let $Q$ be a $d$-dimensional positive definite quadratic form of determinant $\Delta > 0$. The integral $\beta_N (s)$ extends meromorphically to the entire complex plane with a single pole at $s = d / 2$, which is simple and has residue \begin{equation} \op{res}_{d / 2} \beta_N (s) = - \frac{1}{\sqrt{\Delta}} \frac{\pi^{d / 2}}{\Gamma (d / 2)} = - \op{res}_{d / 2} \alpha (s) . \label{eq:betaNres} \end{equation} \end{proposition} \begin{proof} The integral $\beta_N (s)$, as defined in \eqref{eq:beta}, is analytic for $\op{Re} s < d / 2$. On the other hand, we easily see that the difference \begin{equation} \beta_N (s) - \int_{Q (x) \leq N} \frac{1}{Q (x)^s} \mathrm{d} x \label{eq:betadiff} \end{equation} is an entire function in $s$. The latter integral can be evaluated in closed form. Indeed, for $\op{Re} s < d / 2$, \begin{eqnarray}\label{eq:intQQ} \int_{Q (x) \leq N} \frac{1}{Q (x)^s} \mathrm{d} x & = & \frac{1}{\sqrt{\Delta}} \int_{\|x\|_2^2 \leq N} \frac{1}{\|x\|^{2 s}_2} \mathrm{d} x\nonumber\\ & = & \frac{1}{\sqrt{\Delta}} \op{vol} (\mathbb{S}^{d - 1}) \int_0^{\sqrt{N}} r^{d - 1 - 2 s} \mathrm{d} r\nonumber\\ & = & \frac{1}{\sqrt{\Delta}} \frac{N^{d / 2 - s}}{d / 2 - s} \frac{\pi^{d / 2}}{\Gamma (d / 2)} . \end{eqnarray} In light of \eqref{eq:betadiff}, this shows that $\beta_N (s)$ has an analytic continuation to the full complex plane with a single pole at $s = d / 2$, which is simple and has residue as claimed in \eqref{eq:betaNres}. The second equality in \eqref{eq:betaNres} follows from \eqref{eq:epsteinres}. \end{proof} We note that the fact that the residue of $\beta_N (s)$ does not depend on $N$ reflects that, for any $N, M > 0$, the differences $\beta_N (s) - \beta_M (s)$ are, as in \eqref{eq:betadiff}, entire functions. We further record that the proof of Proposition \ref{prop:betares} is related to the following observation. For any reasonable function $F_s : \mathbb{R}^d \rightarrow \mathbb{R}$ such that $F_s (\lambda x) = | \lambda |^{- 2 s} F_s (x)$, \begin{eqnarray} \int_{\|x\|_{\infty} \leq 1} F_s (x) \mathrm{d} \lambda_d & = & \int_0^1 \int_{\|x\|_{\infty} = t} F_s (x) \mathrm{d} \lambda_{d - 1} \mathrm{d} t \nonumber\\ & = & \int_0^1 t^{d - 1 - 2 s} \int_{\|x\|_{\infty} = 1} F_s (x) \mathrm{d} \lambda_{d - 1} \mathrm{d} t \nonumber\\ & = & \frac{1}{d - 2 s} \int_{\|x\|_{\infty} = 1} F_s (x) \mathrm{d} \lambda_{d - 1}, \label{eq:intFs} \end{eqnarray} where $\lambda_d$ denotes the $d$-dimensional Lebesgue measure and $\lambda_{d - 1}$ the induced $(d - 1)$-dimensional surface measure (that is, $\mathrm{d} \lambda_{d - 1} = \mathrm{d} x_1 \cdots \mathrm{d} x_{j - 1} \mathrm{d} x_{j + 1} \cdots \mathrm{d} x_d$ on the part of the domain where $x_j$ is constant). \begin{remark} Since the notation used in (\ref{eq:intFs}) is rather terse, let us, for instance, spell out the crucial first equality. By separating the variable of maximal absolute value and then interchanging summation and integration, \begin{eqnarray*} \int_{\|x\|_{\infty} \leq 1} F_s (x) \mathrm{d} \lambda_d & = & 2 \sum_{j = 1}^d \int_0^1 \left( \int_{[- x_j, x_j]^{d - 1}} F_s (x) \mathrm{d} x_1 \cdots \mathrm{d} x_{j - 1} \mathrm{d} x_{j + 1} \cdots \mathrm{d} x_d \right) \mathrm{d} x_j\\ & = & \int_0^1 \left( \sum_{j = 1}^d \int_{\substack{\|x\|= |t| \\ x_j = \pm t}} F_s (x) \mathrm{d} x_1 \cdots \mathrm{d} x_{j - 1} \mathrm{d} x_{j + 1} \cdots \mathrm{d} x_d \right) \mathrm{d} t\\ & = & \int_0^1 \int_{\|x\|_{\infty} = t} F_s (x) \mathrm{d} \lambda_{d - 1} \mathrm{d} t. \end{eqnarray*} We note that the relation between first and final integral also holds with $\| \cdot \|_{\infty}$ replaced by $\| \cdot \|_2$ (in which case $\lambda_{d - 1}$ would now refer to the surface measure on the Euclidean sphere $\{x \in \mathbb{R}^d : \|x\|_2 = t\}$). \hspace*{\fill}$\Diamond$\medskip \end{remark} Based on \eqref{eq:intFs}, we obtain the following consequence of Proposition \ref{prop:betares}, which will be important for our purposes later on. \begin{lemma} \label{lem:intQdi}Let $Q$ be a $d$-dimensional positive definite quadratic form of determinant $\Delta > 0$. Then we have \begin{equation} \int_{\|x\|_{\infty} = 1} \frac{1}{Q (x)^{d / 2}} \mathrm{d} \lambda_{d - 1} = \frac{2}{\sqrt{\Delta}} \frac{\pi^{d / 2}}{\Gamma (d / 2)} . \label{eq:intQdi} \end{equation} \end{lemma} \begin{proof} From the arguments in the proof of Proposition \ref{prop:betares}, we know that \[ \int_{\|x\|_{\infty} \leq 1} \frac{1}{Q (x)^s} \mathrm{d} \lambda_d \] is a meromorphic function with a simple pole at $s = d / 2$. The computation in \eqref{eq:intFs} shows that \[ \int_{\|x\|_{\infty} = 1} \frac{1}{Q (x)^{d / 2}} \mathrm{d} \lambda_{d - 1} = - 2 \op{res}_{d / 2} \int_{\|x\|_{\infty} \leq 1} \frac{1}{Q (x)^s} \mathrm{d} \lambda_d = \frac{2}{\sqrt{\Delta}} \frac{\pi^{d / 2}}{\Gamma (d / 2)}, \] with the last equality following in analogy with Proposition \ref{prop:betares}. \end{proof} \begin{example}[Generalized $\arctan(1)$] The special case $Q (x) := x_1^2 + \cdots + x_d^2$ results in the integral evaluation \begin{equation} \int_{[- 1, 1]^{d - 1}} \frac{1}{(1 + x_1^2 + \cdots x_{d - 1}^2)^{d / 2}} \mathrm{d} x = \frac{1}{d} \frac{\pi^{d / 2}}{\Gamma (d / 2)}, \label{eq:arctan1} \end{equation} which has been derived in {\cite[Section 5.7.3]{bb}} as a radially invariant generalization of $\arctan (1)$. Let us indicate an alternative direct derivation of \eqref{eq:arctan1}. To this end, recall that the gamma function is characterized by \[ \frac{\Gamma (s)}{A^s} = \int_0^{\infty} t^{s - 1} e^{- A t} \mathrm{d} t. \] Applying this integral representation, which is valid for $\op{Re} s > 0$, with $A = 1 + x_1^2 + \cdots x_{d - 1}^2$, we find \begin{eqnarray*} & & \int_{[- 1, 1]^{d - 1}} \frac{1}{(1 + x_1^2 + \cdots x_{d - 1}^2)^s} \mathrm{d} x\\ & = & \frac{1}{\Gamma (s)} \int_{[- 1, 1]^{d - 1}} \int_0^{\infty} t^{s - 1} e^{- (1 + x_1^2 + \cdots x_{d - 1}) t} \mathrm{d} t \mathrm{d} x\\ & = & \frac{1}{\Gamma (s)} \int_0^{\infty} t^{s - 1} e^{- t} \int_{[- 1, 1]^{d - 1}} e^{- (x_1^2 + \cdots x_{d - 1}) t} \mathrm{d} x \mathrm{d} t\\ & = & \frac{1}{\Gamma (s)} \int_0^{\infty} t^{s - 1} e^{- t} \left( \int_{- 1}^1 e^{- x^2 t} \mathrm{d} x \right)^{d - 1} \mathrm{d} t\\ & = & \frac{2}{\Gamma (s)} \int_0^{\infty} u^{2 s - 1} e^{- u^2} \left( \int_{- 1}^1 e^{- x^2 u^2} \mathrm{d} x \right)^{d - 1} \mathrm{d} u. \end{eqnarray*} In particular, for $s = d / 2$, \begin{eqnarray*} \int_{[- 1, 1]^{d - 1}} \frac{1}{(1 + x_1^2 + \cdots x_{d - 1}^2)^{d / 2}} \mathrm{d} x & = & \frac{2}{\Gamma (d / 2)} \int_0^{\infty} e^{- u^2} \left( u \int_{- 1}^1 e^{- x^2 u^2} \mathrm{d} x \right)^{d - 1} \mathrm{d} u. \end{eqnarray*} Define \[ f (u) = u \int_{- 1}^1 e^{- x^2 u^2} \mathrm{d} x, \] which, in terms of the error function, can be expressed as $f (u) = \sqrt{\pi} \op{erf} (u)$. We note that $f (u) \rightarrow \sqrt{\pi}$ as $u \rightarrow \infty$. Further, the derivative is simply \[ f' (u) = 2 e^{- u^2} . \] After the substitution $v = f (u)$, we thus find \[ \int_{[- 1, 1]^{d - 1}} \frac{1}{(1 + x_1^2 + \cdots x_{d - 1}^2)^{d / 2}} \mathrm{d} x = \frac{1}{\Gamma (d / 2)} \int_0^{\sqrt{\pi}} v^{d - 1} \mathrm{d} v = \frac{1}{d} \frac{\pi^{d / 2}}{\Gamma (d / 2)}, \] as claimed. \hspace*{\fill}$\Diamond$\medskip \end{example} As in \eqref{eq:Q}, we denote with $Q=Q_A$ the quadratic form $Q (x)$ on $\mathbb{R}^d$ associated with the symmetric matrix $A = (a_{i j})_{1 \leq i, j \leq d}$. We now record an extension of Lemma \ref{lem:intQdi}, which will prove useful for our purposes (the case $B = A$ in \eqref{eq:intQAB} reduces to \eqref{eq:intQdi}, and it is the case $B = A^2$ that will appear later). Recall that the \emph{trace} of a square matrix is given by $\op{tr}A = \sum_{j=1}^d a_{jj}$ and defines a Euclidean norm on the symmetric $d \times d$ matrices via $\langle A_1,A_2 \rangle = \op{tr} (A_1A_2)$. \begin{lemma} \label{lem:intQAB}For matrices $A, B \in \mathbb{R}^{d \times d}$, with $A$ positive definite, \begin{equation} \int_{\|x\|_{\infty} = 1} \frac{Q_B (x)}{Q_A (x)^{1 + d / 2}} \mathrm{d} \lambda_{d - 1} = \frac{\op{tr} (B A^{- 1})}{\sqrt{\det (A)}} \frac{\pi^{d / 2}}{\Gamma (1 + d / 2)} . \label{eq:intQAB} \end{equation} \end{lemma} \begin{proof} On decomposing $A$ as $A = L^T L$, we find \[ \int_{Q (x) \leq 1} \frac{Q_B (x)}{Q_A (x)^{s + 1}} \mathrm{d} x = \frac{1}{\det (L)} \int_{\|x\|_2^2 \leq 1} \frac{Q_C (x)}{\|x\|^{2 s + 2}_2} \mathrm{d} x, \] with $C := (L^{- 1})^T B L^{- 1}$. For the residue of the latter integral only the quadratic terms $C_{11} x_1^2 + \cdots + C_{d d} x_d^2$ in $Q_C (x)$ contribute; indeed, in the present case the contributions of the mixed terms $C_{i j} x_i x_j$, $i \neq j$, integrate to zero. Because of symmetry we thus obtain \begin{eqnarray*} \int_{Q (x) \leq 1} \frac{Q_B (x)}{Q_A (x)^{s + 1}} \mathrm{d} x & = & \frac{\op{tr} (C)}{\sqrt{\det (A)}} \int_{\|x\|_2^2 \leq 1} \frac{x_1^2}{\|x\|^{2 s + 2}_2} \mathrm{d} x\\ & = & \frac{\op{tr} (C)}{d \sqrt{\det (A)}} \int_{\|x\|_2^2 \leq 1} \frac{1}{\|x\|^{2 s}_2} \mathrm{d} x\\ & = & \frac{\op{tr} (C)}{d \sqrt{\det (A)}} \frac{1}{d / 2 - s} \frac{\pi^{d / 2}}{\Gamma (d / 2)}, \end{eqnarray*} with the final step as in \eqref{eq:intQQ}. Since the trace is commutative, \[ \op{tr} (C) = \op{tr} (L^{- T} B L^{- 1}) = \op{tr} (B L^{- 1} L^{- T}) = \op{tr} (B A^{- 1}). \] We conclude that, for any compact region $D \subset \mathbb{R}^d$ containing a neighborhood of the origin, \begin{equation} \op{res}_{d / 2} \int_D \frac{Q_B (x)}{Q_A (x)^{s + 1}} \mathrm{d} x = - \frac{\op{tr} (B A^{- 1})}{d \sqrt{\det (A)}} \frac{\pi^{d / 2}}{\Gamma (d / 2)} . \label{eq:resintQAB} \end{equation} In light of the computation \eqref{eq:intFs}, we arrive at \[ \int_{\|x\|_{\infty} = 1} \frac{Q_B (x)}{Q_A (x)^{1 + d / 2}} \mathrm{d} \lambda_{d - 1} = - 2 \op{res}_{d / 2} \int_{\|x\|_{\infty} \leq 1} \frac{Q_B (x)}{Q_A (x)^{s + 1}} \mathrm{d} x, \] which, together with \eqref{eq:resintQAB}, implies \eqref{eq:intQAB}. \end{proof} \section{Convergence of Wigner limits}\label{sec:wigconv} Our next goal is to show that $\sigma_N (s)$ indeed converges in the vertical strip suggested in Problem \ref{prob}. As discussed in \cite[Chapter 2 and 8]{latticesums}, convergence over such hyper-cubes is more stable than that over Euclidean balls and similar shapes. Other limit procedures are compared in Section \ref{sec:alt}. \begin{proposition}[Convergence in a strip] \label{prop:sigmastrip}Let $Q$ be an arbitrary positive definite quadratic form on $\mathbb{R}^d$. Then the limit $\sigma (s) := \lim_{N \rightarrow \infty} \sigma_N (s)$ exists in the strip $d / 2 - 1 < \op{Re} s < d / 2$ and coincides therein with the analytic continuation of $\alpha (s)$. \end{proposition} \begin{proof} For the first part of the claim, we follow the proof given in {\cite{latticesums-bbs}} for binary forms $Q$. Fix $\sigma > 0$ as well as $R > 0$ and set $\Omega := \{s : \op{Re} s > \sigma, \hspace{1em} |s| < R\}$. All order terms below are uniform with respect to $s$ in the bounded region $\Omega$. For $N \geq 1$ let \begin{eqnarray*} \delta_N (s) & := & \sigma_N (s) - \sigma_{N - 1} (s)\\ & = & \sum_{\|n\|_{\infty} = N} \int_{\|x\|_{\infty} \leq 1 / 2} \left[ \frac{1}{Q (n)^s} - \frac{1}{Q (n + x)^s} \right] \mathrm{d} x. \end{eqnarray*} Here and in the sequel, we let $f (x) := Q (n + x)^{- s}$ with $\|n\|_{\infty} = N$ and $\|x\|_{\infty} \leq 1 / 2$. Since we may assume $Q (x) = \sum_{i, j} a_{i j} x_i x_j$, with $a_{i j} = a_{j i}$, is positive definite, we have the estimate \begin{eqnarray*} f_{i j} (x) & = & \frac{4 s (s + 1)}{Q (n + x)^{s + 2}} \sum_k a_{i k} (n_k + x_k) \sum_{\ell} a_{j \ell} (n_{\ell} + x_{\ell}) - \frac{2 a_{i j} s}{Q (n + x)^{s + 1}}\\ & = & O (N^{- 2 \sigma - 2}). \end{eqnarray*} Here, the indices of $f$ indicate partial derivatives with respect to the $i$-th or $j$-th variable. We thus have \begin{equation} f (x) - f (0) = \sum_i x_i f_{i} (0) + O (N^{- 2 \sigma - 2}). \label{eq:fx0} \end{equation} Consequently, \begin{eqnarray*} & & \int_{\|x\|_{\infty} \leq 1 / 2} \left[ \frac{1}{Q (n)^s} - \frac{1}{Q (n + x)^s} \right] \mathrm{d} x\\ & = & - \int_{\|x\|_{\infty} \leq 1 / 2} \left[ \sum_i x_i f_{i} (0) \right] \mathrm{d} x + O (N^{- 2 \sigma - 2})\\ & = & O (N^{- 2 \sigma - 2}), \end{eqnarray*} because the final integral, being odd, vanishes. Hence, \[ \delta_N (s) = O (N^{d - 2 \sigma - 3}), \] and so, for all $s \in \Omega$, $| \delta_N (s) | \leq M N^{d - 2 \sigma - 3}$ for some $M$, which is independent of $N$ and $s$. Assume now that $\sigma > d / 2 - 1$. Since $\delta_N (s)$ is an entire function, the Weierstrass $M$-test shows that \[ \delta (s) := \sum_{N = 1}^{\infty} \delta_N (s) \] is an analytic function in $\Omega$. Since $R$ was arbitrary, $\delta (s)$ is in fact analytic in the half-plane $\op{Re} s > d / 2 - 1$. By construction, \begin{equation} \delta (s) = \lim_{N \rightarrow \infty} \left[ \sigma_N (s) - \sigma_0 (s) \right] = \lim_{N \rightarrow \infty} \left[ \sigma_N (s) + \beta_0 (s) \right] . \label{eq:delta1} \end{equation} It follows that the limit $\sigma (s)$ exists if, additionally, $\op{Re} s < d / 2$. \medskip For the second part of the claim, we begin with the simple observation that, for $\op{Re} s < d / 2$, \begin{eqnarray} \beta_N (s) & = & \int_{\|x\|_{\infty} \leq N + 1 / 2} \frac{1}{Q (x)^s} \mathrm{d} x \nonumber\\ & = & (2 N + 1)^{d - 2 s} \int_{\|x\|_{\infty} \leq 1 / 2} \frac{1}{Q (x)^s} \mathrm{d} x \nonumber\\ & = & (2 N + 1)^{d - 2 s} \beta_0 (s) . \label{eq:betaN0} \end{eqnarray} As shown in Proposition \ref{prop:betares}, both $\beta_N$ and $\beta_0$ have meromorphic extensions to the entire complex plane, and the relation \eqref{eq:betaN0} continues to hold. In particular, this shows that, for $\op{Re} s > d / 2$, the meromorphic continuation of $\beta_N$ satisfies \[ \lim_{N \rightarrow \infty} \beta_N (s) = \lim_{N \rightarrow \infty} (2 N + 1)^{d - 2 s} \beta_0 (s) = 0. \] Working from \eqref{eq:delta1}, we thus have, for $\op{Re} s > d / 2$, \begin{equation} \delta (s) = \lim_{N \rightarrow \infty} \left[ \alpha_N (s) - \beta_N (s) + \beta_0 (s) \right] = \alpha (s) + \beta_0 (s) . \label{eq:deltaab} \end{equation} On the other hand, we have shown via \eqref{eq:delta1} that, for $\op{Re} s < d / 2$, \begin{equation} \delta (s) = \sigma (s) + \beta_0 (s) . \label{eq:deltasb} \end{equation} Since both $\delta (s)$ and $\beta_0 (s)$ are meromorphic in the half-plane $\op{Re} s > d / 2 - 1$, comparing \eqref{eq:deltaab} and \eqref{eq:deltasb} proves that the analytic continuations of $\sigma (s)$ and $\alpha (s)$ agree. In particular, in the strip $d / 2 - 1 < \op{Re} s < d / 2$, the limit $\sigma (s)$, which was shown to exist, equals the analytic continuation of $\alpha (s)$. \end{proof} We note that Proposition \ref{prop:sigmastrip} agrees with the results known for $d = 2, 3$. In the $d = 2$ case, the limit $\sigma (s)$ exists for $0 < \op{Re} s < 1$, in accordance with {\cite[Theorem 1]{latticesums-bbs}}. In the $d = 3$ case, the limit $\sigma (s)$ exists for $1 / 2 < \op{Re} s < 3 / 2$, which is consistent with the special case of the cubic lattice discussed in {\cite[Theorem 3]{latticesums-bbs}}. \section{Jump discontinuities in Wigner limits}\label{sec:wigjump} In {\cite[Theorem 3]{latticesums-bbs}} it was shown that, in the case of the cubic lattice, the limit $\sigma (s)$ also exists for $s = 1 / 2$, but is discontinuous there. In fact, it was shown that \[\sigma (1 / 2) - \pi / 6 = \lim_{\varepsilon \rightarrow 0^+} \sigma (1 / 2 + \varepsilon).\] We now extend this result to cubic lattices in arbitrary dimensions, in which case we can and do evaluate the jump discontinuity in simple terms. We then show that an analogous result is true for arbitrary positive definite quadratic forms, though the proof is more technical and no simple closed-form expression for the jumps is given. \begin{remark}[$\sigma(0)$] \label{rk:sigma0}Note that, for trivial reasons, the limit $\sigma (0)$ always exists and is given by $\sigma (0) = - 1$, which agrees with the value $\alpha (0) = - 1$, obtained by analytic continuation from \eqref{eq:epsteinfunc} and \eqref{eq:epsteinres}. (The value $s = 0$ is missed in the statement of Theorem 3 in {\cite{latticesums-bbs}}.) \hspace*{\fill}$\Diamond$\medskip \end{remark} \begin{theorem}[Cubic jump discontinuity] \label{thm:jump}Let $Q (x) = x_1^2 + \cdots + x_d^2$. Then the corresponding limit $\sigma (s) := \lim_{N \rightarrow \infty} \sigma_N (s)$ exists in the strip $d / 2 - 1 < \op{Re} s < d / 2$ and for $s = d / 2 - 1$. In the strip, $\sigma (s)$ coincides with the analytic continuation of $\alpha (s)$. On the other hand, \[ \sigma (d / 2 - 1) - \frac{1}{6} \frac{\pi^{d / 2}}{\Gamma (d / 2 - 1)} = \alpha (d / 2 - 1) = \lim_{\varepsilon \rightarrow 0^+} \sigma (d / 2 - 1 + \varepsilon) . \] In particular, for $d \geq 3$, $\sigma (s)$ is discontinuous at $s = d / 2 - 1$. \end{theorem} \begin{proof} In light of Proposition \ref{prop:sigmastrip}, we only need to show the statement about the value of $\sigma (s)$ at $s = d / 2 - 1$. Let us adopt the notation used in Proposition \ref{prop:sigmastrip}, including, in particular, the definitions of $\delta_N$ and $f (x) := Q (n + x)^{- s}$ with $\|n\|_{\infty} = N$ and $\|x\|_{\infty} \leq 1 / 2$. Proceeding as for \eqref{eq:fx0}, we have that \[ f (x) - f (0) = \sum_i x_i f_{i} (0) + \frac{1}{2} \sum_{i, j} x_i x_j f_{i j} (0) + O (N^{- 2 \sigma - 3}). \] Since terms of odd order in the $x_i$ are eliminated in the subsequent integration, we focus on the terms $f_{i i}$. In the present case of the cubic lattice, \begin{eqnarray}\label{eq:sumfij0} \sum_i f_{i i} (0) & = & \sum_i \left[ \frac{4 s (s + 1)}{Q (n)^{s + 2}} n_i^2 - \frac{2 s}{Q (n)^{s + 1}} \right]\nonumber\\ & = & \frac{4 s (s + 1) - 2 d s}{Q (n)^{s + 1}}\nonumber\\ & = & \frac{2 s (2 s - (d - 2))}{Q (n)^{s + 1}} . \end{eqnarray} We thus find that \begin{eqnarray*} & & \int_{\|x\|_{\infty} \leq 1 / 2} \left[ \frac{1}{Q (n)^s} - \frac{1}{Q (n + x)^s} \right] \mathrm{d} x\\ & = & - \frac{1}{2} \int_{\|x\|_{\infty} \leq 1 / 2} \left[ \sum_i x_i^2 f_{i i} (0) \right] \mathrm{d} x + O (N^{- 2 \sigma - 3})\\ & = & - \frac{1}{24} \sum_i f_{ii} (0) + O (N^{- 2 \sigma - 3}),\end{eqnarray*} \noindent (on integrating term-by-term). Then, on appealing to \eqref{eq:sumfij0}, \begin{eqnarray*} \int_{\|x\|_{\infty} \leq 1 / 2} \left[ \frac{1}{Q (n)^s} - \frac{1}{Q (n + x)^s} \right] \mathrm{d} x& = & \frac{1}{12} \frac{s (d - 2 - 2 s)}{Q (n)^{s + 1}} + O (N^{- 2 \sigma - 3}) . \end{eqnarray*} Hence, \begin{eqnarray*} \delta_N (s) & = & \sum_{\|n\|_{\infty} = N} \int_{\|x\|_{\infty} \leq 1 / 2} \left[ \frac{1}{Q (n)^s} - \frac{1}{Q (n + x)^s} \right] \mathrm{d} x\\ & = & \frac{s (d - 2 - 2 s)}{12} \sum_{\|n\|_{\infty} = N} \frac{1}{Q (n)^{s + 1}} + O (N^{d - 2 \sigma - 4})\\ & = & \frac{s (d - 2 - 2 s)}{12 N^{2 s - d + 3}} \frac{1}{N^{d - 1}} \sum_{\|n\|_{\infty} = N} \frac{1}{Q (n / N)^{s + 1}} + O (N^{d - 2 \sigma - 4}) . \end{eqnarray*} We now note that \begin{equation*} \frac{1}{2 d N^{d - 1}} \sum_{\|n\|_{\infty} = N} \frac{1}{Q (n / N)^{s + 1}} = V_N(s) + O (N^{- 1}), \end{equation*} where \begin{equation*} V_N(s) := \frac{1}{N^{d - 1}} \sum_{- N \leq n_i < N} \frac{1}{(1 + (n_1 / N)^2 + \cdots + (n_{d - 1} / N)^2)^{s + 1}}. \end{equation*} We first show that $V_N (s)$ approaches the integral \[ V (s) := \int_{[- 1, 1]^{d - 1}} \frac{1}{(1 + x_1^2 + \cdots + x_{d - 1}^2)^{s + 1}} \mathrm{d} x. \] Indeed, this follows since, for $\op{Re} s \geq - 2$, \begin{eqnarray*} |V (s) - V_N (s) | & = & \left| \sum_{- N \leq n_i < N} \int_{n_i \leq N x_i \leq n_i + 1} \left[ \frac{1}{(1 + x_1^2 + \cdots + x_{d - 1}^2)^{s + 1}} \right. \right.\\ & & \left. \left. - \frac{1}{(1 + (n_1 / N)^2 + \cdots + (n_{d - 1} / N)^2)^{s + 1}} \right]\,\mathrm{d} x \right|\\ & \leq & \sum_{- N \leq n_i < N} \int_{n_i \leq N x_i \leq n_i + 1} (d - 1) \frac{2 |s + 1|}{N} \,\mathrm{d} x\\ & = & \frac{2^d (d - 1) |s + 1|}{N} . \end{eqnarray*} To bound the above integrand, we used that $|x^{\lambda} - y^{\lambda} | \leq | \lambda | |x - y|$ when $\op{Re} \lambda \leq 1$ and $x, y \geq 1$ (as follows from the mean value theorem). Combining these estimates, we can thus write \begin{equation} \delta_N (s) = \frac{d}{6} \frac{s (d - 2 - 2 s)}{N^{2 s - d + 3}} V (s) + W_N (s), \label{eq:deltanVW} \end{equation} where $W_N (s) = O (N^{d - 2 \sigma - 4})$. For $\sigma > d / 2 - 3 / 2$, the sum \[ W (s) := \sum_{N = 1}^{\infty} W_N (s) \] converges and, by the Weierstrass $M$-test, defines an analytic function. If, further, $\op{Re} s > d / 2 - 1$ then, from \eqref{eq:deltanVW}, the sum $\delta (s) := \sum_{N = 1}^{\infty} \delta_N (s)$ converges and we have \[ \delta (s) = \frac{d}{6} s (d - 2 - 2 s) \zeta (2 s - d + 3) V (s) + W (s) . \] In particular, since $\zeta (s)$ has a simple pole at $s = 1$ of residue $1$, we find \begin{eqnarray}\label{eq:lim}\lim_{\varepsilon \rightarrow 0^+} \delta (d / 2 - 1 + \varepsilon) = - \frac{d}{6} (d / 2 - 1) V (d / 2 - 1) + W (d / 2 - 1) . \end{eqnarray} On the other hand, it follows from \eqref{eq:deltanVW} that $\delta_N (d / 2 - 1) = W_N (d / 2 - 1)$. Hence, the defining series for $\delta (s)$ also converges when $s = d / 2 - 1$ and we obtain \begin{eqnarray}\label{eq:sigv} \delta (d / 2 - 1) = W (d / 2 - 1) . \end{eqnarray} Using the consequence \eqref{eq:arctan1} of Lemma \ref{lem:intQdi}, we have \[ \frac{d}{6} (d / 2 - 1) V (d / 2 - 1) = \frac{1}{6} \frac{\pi^{d / 2}}{\Gamma (d / 2 - 1)} . \] Since, by construction, $\delta (s) = \sigma (s) - \sigma_0 (s) = \sigma (s)$, on comparing \eqref{eq:lim} and \eqref{eq:sigv} we are done. \end{proof} \begin{example}[Explicit evaluations in even dimensions] In the case of cubic lattice sums and small even dimension, the value $\sigma (d / 2 - 1)$, at the jump discontinuity, can be given explicitly by combining Theorem \ref{thm:jump} and the closed forms for the corresponding Epstein zeta function, recalled in Example \ref{eg:cubicexact} below. Let $Q_d(x) = x_1^2+\ldots+x_d^2$. \begin{eqnarray*} \sigma_{Q_2} (0) & = & \alpha_{Q_2} (0) = - 1,\\ \sigma_{Q_4} (1) & = & \frac{\pi^2}{6} + \alpha_{Q_4} (1) = \frac{\pi^2}{6} - 8 \log 2,\\ \sigma_{Q_6} (2) & = & \frac{\pi^3}{6} + \alpha_{Q_6} (2) = \frac{\pi^3}{6} - \frac{\pi^2}{3} - 8 G,\\ \sigma_{Q_8} (3) & = & \frac{\pi^4}{12} + \alpha_{Q_8} (3) = \frac{\pi^4}{12} - 8 \zeta (3),\\ \sigma_{Q_{24}} (11) & = & \frac{\pi^{12}}{6 \cdot 10!} + \alpha_{Q_{24}} (11) = \frac{\pi^{12}}{6 \cdot 10!} - \frac{8}{691} \zeta (11) + \frac{271435}{5528} L_\Delta (11) . \end{eqnarray*} Here, $G = \sum_{n = 1}^{\infty} \chi_{- 4} (n) / n^2$ denotes Catalan's constant, and $L_\Delta$ is (the analytic continuation of) $L_{\Delta}(s) = \sum_{n=1}^\infty \tau(n) / n^s$ with $\tau(n)$ Ramanujan's $\tau$-function. A few properties of this remarkable function are commented on in Example \ref{eg:cubicexact}. We note that we have used the appropriate reflection formulas to simplify these evaluations. We note that the above values mix numbers of different `order', such as $\pi^4$ and $\zeta (3)$ which have order $4$ and $3$, respectively. This may be another argument to use $\alpha(d/2-1)$ as the `value' of the Wigner limit even when the limit $\sigma(d/2-1)$ itself converges. \hspace*{\fill}$\Diamond$\medskip \end{example} We now extend Theorem \ref{thm:jump} to arbitrary definite quadratic forms. For the most part, the proof is a natural extension of the proof of Theorem \ref{thm:jump}. For the convenience of the reader, we duplicate some parts, as well as the overall structure, of the previous proof. As in \eqref{eq:Q}, let $Q = Q_A$ be the positive definite quadratic form associated to the symmetric matrix $A$. Set also $B(s) := \op{tr} (A) A - 2 (s + 1) A^2$. Finally, define \begin{align}\label{eq:vq} V (s) := V_Q(s) := \int_{\|x\|_{\infty} = 1} \frac{Q_{B(s)} (x)}{Q_A (x)^{s + 2}} \mathrm{d} \lambda_{d - 1}, \end{align} with $\lambda_{d - 1}$ the induced $(d-1)$-dimensional measure as in \eqref{eq:intFs}. \begin{theorem}[General jump discontinuity] \label{thm:jumpx}Let $Q$ be an arbitrary positive definite quadratic form. Then the corresponding limit $\sigma (s) := \lim_{N \rightarrow \infty} \sigma_N (s)$ exists in the strip $d / 2 - 1 < \op{Re} s < d / 2$ and for $s = d / 2 - 1$. In the strip, $\sigma (s)$ coincides with the analytic continuation of $\alpha (s)$. On the other hand, \begin{equation}\label{eq:jumpx} \sigma (d / 2 - 1) + \frac{d / 2 - 1}{24} V'_Q (d / 2 - 1) = \alpha (d / 2 - 1) = \lim_{\varepsilon \rightarrow 0^+} \sigma (d / 2 - 1 + \varepsilon), \end{equation} with $V_Q$ as introduced in equation \eqref{eq:vq}. \end{theorem} \begin{proof} In light of Proposition \ref{prop:sigmastrip}, we only need to show the statement about the value of $\sigma (s)$ at $s = d / 2 - 1$. Let us adopt the notation used in Proposition \ref{prop:sigmastrip}, including, in particular, the definitions of $\delta_N$ and $f (x) := Q (n + x)^{- s}$ with $\|n\|_{\infty} = N$ and $\|x\|_{\infty} \leq 1 / 2$. Proceeding as for \eqref{eq:fx0}, we have that \[ f (x) - f (0) = \sum_i x_i f_i (0) + \frac{1}{2} \sum_{i, j} x_i x_j f_{i j} (0) + O (N^{- 2 \sigma - 3}). \] Since terms of odd order in the $x_i$ are eliminated in the subsequent integration, we focus on the terms $f_{i i} (0)$, which are given by \[ f_{i i} (0) = \frac{4 s (s + 1)}{Q (n)^{s + 2}} \left[ \sum_{k = 1}^d a_{i k} n_k \right]^2 - \frac{2 a_{i i} s}{Q (n)^{s + 1}} . \] Hence, equation \eqref{eq:sumfij0} generalizes to \begin{eqnarray} \sum_{i = 1}^d f_{i i} (0) & = & \frac{4 s (s + 1)}{Q (n)^{s + 2}} \sum_{1 \leq k, l \leq d} \left( \sum_{i = 1}^d a_{k i} a_{i l} \right) n_k n_l - 2 s \frac{\op{tr} (A)}{Q (n)^{s + 1}} \nonumber\\ & = & \frac{4 s (s + 1)}{Q (n)^{s + 2}} Q_{A^2} (n) - 2 s \frac{\op{tr} (A)}{Q (n)^{s + 1}} \nonumber\\ & = & \frac{2 s}{Q (n)^{s + 2}} \left[ 2 (s + 1) Q_{A^2} (n) - \op{tr} (A) Q (n) \right] . \label{eq:sumfij0x} \end{eqnarray} We thus find, on integrating term-by-term, that \begin{eqnarray*} & & \int_{\|x\|_{\infty} \leq 1 / 2} \left[ \frac{1}{Q (n)^s} - \frac{1}{Q (n + x)^s} \right] \mathrm{d} x\\ & = & - \frac{1}{2} \int_{\|x\|_{\infty} \leq 1 / 2} \left[ \sum_{i = 1}^d x_i^2 f_{i i} (0) \right] \mathrm{d} x + O (N^{- 2 \sigma - 3})\\ & = & - \frac{1}{24} \sum_{i = 1}^d f_{i i} (0) + O (N^{- 2 \sigma - 3})\\ & = & \frac{s}{12} \frac{\op{tr} (A) Q (n) - 2 (s + 1) Q_{A^2} (n)}{Q (n)^{s + 2}} + O (N^{- 2 \sigma - 3}) . \end{eqnarray*} In the final step, we appealed to \eqref{eq:sumfij0x}. Hence, \begin{eqnarray*} \delta_N (s) & = & \sum_{\|n\|_{\infty} = N} \int_{\|x\|_{\infty} \leq 1 / 2} \left[ \frac{1}{Q (n)^s} - \frac{1}{Q (n + x)^s} \right] \mathrm{d} x\\ & = & \frac{s}{12} \sum_{\|n\|_{\infty} = N} \frac{\op{tr} (A) Q (n) - 2 (s + 1) Q_{A^2} (n)}{Q (n)^{s + 2}} + O (N^{d - 2 \sigma - 4})\\ & = & \frac{s}{12 N^{2 s + 2}} \sum_{\|n\|_{\infty} = N} \frac{\op{tr} (A) Q (n / N) - 2 (s + 1) Q_{A^2} (n / N)}{Q (n / N)^{s + 2}} + O (N^{d - 2 \sigma - 4}) . \end{eqnarray*} Consider, as defined above, $B(s) = \op{tr} (A) A - 2 (s + 1) A^2$. As in the proof of Theorem \ref{thm:jump}, one obtains that, for $\op{Re} s \geq - 2$, \begin{align*} V (s) = \int_{\|x\|_{\infty} = 1} \frac{Q_{B(s)} (x)}{Q_A (x)^{s + 2}} \mathrm{d} \lambda_{d - 1} = \frac{1}{N^{d - 1}} \sum_{\|n\|_{\infty} = N} \frac{Q_{B(s)} (n / N)}{Q_A (n / N)^{s + 2}} + O (N^{- 1}), \end{align*} with $\lambda_{d - 1}$ as in \eqref{eq:intFs}. Combining these, we can thus write \begin{equation} \delta_N (s) = s \frac{V (s)}{12 N^{2 s - d + 3}} + W_N (s), \label{eq:deltanVWx} \end{equation} where $W_N (s) = O (N^{d - 2 \sigma - 4})$. For $\sigma > d / 2 - 3 / 2$, the sum \[ W (s) := \sum_{N = 1}^{\infty} W_N (s) \] converges and, by the Weierstrass $M$-test, defines an analytic function. If, further, $\op{Re} s > d / 2 - 1$ then, from \eqref{eq:deltanVWx}, the sum $\delta (s) := \sum_{N = 1}^{\infty} \delta_N (s)$ converges and we have \begin{equation} \delta (s) = \frac{s V (s)}{12} \zeta (2 s - d + 3) + W (s) . \label{eq:deltaVWx} \end{equation} Since \[ \op{tr} (B(s) A^{- 1}) = \op{tr} (\op{tr} (A) I - 2 (s + 1) A) = (d - 2 (s + 1)) \op{tr} (A), \] Lemma \ref{lem:intQAB} shows that \begin{equation} V (d / 2 - 1) = 0. \label{eq:V0} \end{equation} Using that $\zeta (s)$ has a simple pole at $s = 1$ of residue $1$, we thus deduce from \eqref{eq:deltaVWx} that \begin{equation}\label{eq:deltapx} \lim_{\varepsilon \rightarrow 0^+} \delta (d / 2 - 1 + \varepsilon) = \frac{d / 2 - 1}{24} V' (d / 2 - 1) + W (d / 2 - 1) . \end{equation} On the other hand, \eqref{eq:deltanVWx} together with \eqref{eq:V0} implies $\delta_N (d / 2 - 1) = W_N (d / 2 - 1)$. Hence, the defining series for $\delta (s)$ also converges when $s = d / 2 - 1$ and we obtain \[ \delta (d / 2 - 1) = W (d / 2 - 1) . \] The claim follows on comparison with \eqref{eq:deltapx}. \end{proof} \subsection{The behaviour of $V_Q'(d/2-1)$} We now examine the nature of $V_Q'(d/2-1)$ in somewhat more detail. Specifically, we are interested in the following question: \begin{problem}\label{prob:VQD} Let $d>1$. Are there positive definite quadratic forms $Q$ on $\mathbb{R}^d$ such that $V_Q'(d/2-1) = 0$? \end{problem} Recall that, in light of Theorem \ref{thm:jumpx}, if $V_Q'(d/2-1) \ne 0$ for a quadratic form $Q$ on $\mathbb{R}^d$, with $d>2$, then the corresponding Wigner limit $\sigma_Q(s)$ exhibits a jump discontinuity at $s=d/2-1$. In fact, in all the cases of $Q$, that we consider in this section, including the cubic lattice case, we find that $V_{Q}'(d/2-1) < 0$, which leads us to speculate whether this inequality holds in general. From the definition \eqref{eq:vq} we obtain that \begin{align}\label{eq:VD0} V_Q'(d/2-1) &= \int_{\|x\|_{\infty} = 1} \frac{-2 Q_{A^2}(x)} {Q_A (x)^{d/2 + 1}} \mathrm{d} \lambda_{d - 1} \nonumber\\ &- \int_{\|x\|_{\infty} = 1} \frac{\op{tr}(A) Q_A(x) - d Q_{A^2}(x)} {Q_A (x)^{d/2 + 1}} \log Q_A(x) \mathrm{d} \lambda_{d - 1} \\ &= -\frac{4\op{tr} (A)}{d\sqrt{\det (A)}} \frac{\pi^{d / 2}}{\Gamma (d / 2)} \nonumber\\ &- \int_{\|x\|_{\infty} = 1} \frac{\op{tr}(A) Q_A(x) - d Q_{A^2}(x)} {Q_A (x)^{d/2 + 1}} \log Q_A(x) \mathrm{d} \lambda_{d - 1}. \label{eq:VD} \end{align} The last equality is a useful consequence of Lemma \ref{lem:intQAB}. We also have that \begin{align}\label{eq:scale} V'_{\lambda Q} (d / 2 - 1) = \lambda^{- (d / 2 - 1)} V_Q' (d / 2 -1) . \end{align} Indeed, it follows directly from the definition (\ref{eq:vq}) that, for $\lambda > 0$, $V_{\lambda Q} (s) = \lambda^{- s} V_Q (s)$ and hence, by (\ref{eq:V0}), that \eqref{eq:scale} holds. The rescaling result \eqref{eq:scale} shows that scaling $Q$ does not change the sign of $V_Q' (d / 2 - 1)$. Also note that both integrals in (\ref{eq:VD0}) scale in the same way; that this is true for the integral involving the logarithm is equivalent to \[ \int_{\|x\|_{\infty} = 1} \frac{\op{tr} (A) Q_A (x) - d Q_{A^2} (x)}{Q_A (x)^{d / 2 + 1}} \mathrm{d} \lambda_{d - 1} = 0, \] which follows from Lemma \ref{lem:intQAB}. \begin{example}[Recovery of cubic jump] Let us demonstrate that Theorem \ref{thm:jumpx} reduces to Theorem \ref{thm:jump} in the cubic lattice case. In that case, $A = I$ and $\op{tr}(A) =d$, so that the integral in \eqref{eq:VD}, involving the logarithm, vanishes. Hence, \[ V' (d / 2 - 1) = - 4 \frac{\pi^{d / 2}}{\Gamma (d / 2)}, \] in agreement with the value given in Theorem \ref{thm:jump}. \hspace*{\fill}$\Diamond$\medskip \end{example} We now give a simple criterion that $V_Q' (d / 2 - 1) < 0$ for certain $Q = Q_A$. Suppose that there is some $\lambda>0$ such that, for all $x$ with $\|x\|_{\infty} = 1$, \begin{equation} 2 Q_{A^2} (x) \geq \left[ d Q_{A^2} (x) - \op{tr} (A) Q_A (x) \right] \, \log Q_{\lambda A} (x). \label{eq:Qineq} \end{equation} It then follows from \eqref{eq:VD0} that $V_Q' (d / 2 - 1) \leq 0$. To see that, in fact, $V_Q' (d / 2 - 1) < 0$, we note that \eqref{eq:Qineq} cannot be an equality for all $x$, because $d Q_{A^2} (x) - \op{tr} (A) Q_A (x)$ does not vanish identically unless $A$ is a multiple of the identity matrix (which corresponds to the cubic case, for which we know the explicit values from Theorem \ref{thm:jump}). In the non-cubic case, the right-hand side thus is a nonzero polynomial times the logarithm of a nonconstant polynomial, while the left-hand side is a polynomial. Since $\log Q_{\lambda A}(x) = \log\lambda + \log Q_A(x)$, a $\lambda>0$ satisfying \eqref{eq:Qineq} certainly exists if the sign of $d Q_{A^2} (x) - \op{tr} (A) Q_A (x)$ is constant for all $x$ with $\|x\|_{\infty} = 1$. We have thus proved the following result. \begin{proposition} \label{prop:Qineqc}Let $Q$ be a positive definite quadratic form on $\mathbb{R}^d$ such that \begin{equation} d \,Q_{A^2} (x) \leq \op{tr} (A) Q_A (x) \label{eq:Qineqc} \end{equation} for all $x$ with $\|x\|_{\infty} = 1$. Then $V_Q' (d / 2 - 1) < 0$. The same conclusion holds if `$\leq$' is replaced with `$\geq$' in (\ref{eq:Qineqc}). \end{proposition} \begin{example}[Some non-cubic lattices] Consider the case when $A$ is given by $A_p := I - p E$, where $E$ is the matrix with all entries equal to $1$. One easily checks that $A_p$ is positive definite if and only if $p < 1 / d$. Hence, we assume $p < 1 / d$. We further observe that \[ Q_{A_p} (x) = \|x\|_2^2 - p \left( \sum_{j = 1}^d x_j \right)^2, \] as well as $A_p^2 = A_{p (2 - d p)}$. Thus equipped, a brief calculation reveals that \[ d Q_{A_p^2} (x) - \op{tr} (A_p) Q_{A_p} (x) = p d \|x\|_2^2 - p \left[ 1 - (d - 1) p \right] \left( \sum_{j = 1}^d x_j \right)^2 . \] Notice that, by H\"older's inequality, \[ \left( \sum_{j = 1}^d x_j \right)^2 \leq \|x\|_1^2 \leq d \|x\|_{\infty} \|x\|_2^2 . \] Assume further that $p \geq 0$, so that $p \left[ 1 - (d - 1) p \right] > 0$. We then find that, for all $x$ with $\|x\|_{\infty} = 1$, \[ d Q_{A_p^2} (x) - \op{tr} (A_p) Q_{A_p} (x) \geq p^2 d (d - 1) \|x\|_2^2 \geq 0. \] By Proposition \ref{prop:Qineqc}, we have thus shown that $V_Q' (d / 2 - 1) < 0$, with $Q = Q_{A_p}$, for all $0 \leq p < 1 / d$. \hspace*{\fill}$\Diamond$\medskip \end{example} Continuing in this vein, we explicitly determine $V_Q'(d/2-1)$ for some very simple binary forms. \begin{example} To indicate the nature of the quantities $V_Q (s)$ and, in consequence, $V_Q' (d / 2 - 1)$, let us consider the very basic case of $Q (x_1, x_2) := a x_1^2 + b x_2^2$, with $a, b > 0$, (of course, the factor $d / 2 - 1$ in (\ref{eq:jumpx}) vanishes in this case, so the contribution of $V'_Q (d / 2 - 1)$ is not, in the end, brought to bear). We have \begin{eqnarray*} V_Q (s) & = & \int_{\|x\|_{\infty} = 1} \frac{(a b - (2 s + 1) a^2) x_1^2 + (a b - (2 s + 1) b^2) x_2^2}{(a x_1^2 + b x_2^2)^{s + 2}} \mathrm{d} \lambda_1\\ & = & 4 \int_0^1 \frac{(a b - (2 s + 1) a^2) x_1^2 + (a b - (2 s + 1) b^2)}{(a x_1^2 + b)^{s + 2}} \mathrm{d} x_1\\ & & + 4 \int_0^1 \frac{(a b - (2 s + 1) a^2) + (a b - (2 s + 1) b^2) x_2^2}{(a + b x_2^2)^{s + 2}} \mathrm{d} x_2. \end{eqnarray*} Using the basic integral \begin{equation} \int_0^1 \frac{1}{(a x^2 + b)^s} \mathrm{d} x = \pFq21{1 / 2, s}{3 / 2}{- a}, \label{eq:intab2F1} \end{equation} and some standard hypergeometric manipulations, we thus find \begin{equation} V_Q (s) = \frac{- 8 s}{(a + b)^s} \left[ \pFq21{1, 1 / 2 - s}{3 / 2}{- \frac{a}{b}} + \pFq21{1, 1 / 2 - s}{3 / 2}{ - \frac{b}{a}} \right] . \end{equation} The factor of $s$ in $V_Q (s)$, together with the elementary special case $s = 1$ of (\ref{eq:intab2F1}), allows us to conclude that \begin{equation*} V_Q' (0) = - 8 \left[ \sqrt{\frac{b}{a}} \arctan \sqrt{\frac{a}{b}} + \sqrt{\frac{a}{b}} \arctan \sqrt{\frac{b}{a}} \right] . \end{equation*} In particular, we observe that $V_Q' (d / 2 - 1) < 0$, though Proposition \ref{prop:Qineqc} does not apply in the present case. \hspace*{\fill}$\Diamond$\medskip \end{example} Sadly, as illustrated by this example, Proposition \ref{prop:Qineqc} is not always accessible. Indeed, it can fail quite comprehensively. \begin{example}[Some scaled cubic lattices] Consider the case when $A$ is given by $A_p := I + p D(a)$, where $D(a)=D(a_1,\ldots,a_d)$ is a diagonal matrix and, without loss, $p \geq 0$. The matrix $A_p$ is positive definite if and only if $p a_k+1 >0$ for all $1 \leq k \leq d$. Suppose that $\op{tr} (D(a))=0$, so that $\op{tr} (A_p)=d$. Also $A_p^2 = I+2pD(a) +p^2D(a_1^2,\ldots,a_k^2)$. Thence, \[ d Q_{A_p^2} (x) - \op{tr} (A_p) Q_{A_p} (x) = pd \sum_{k=1}^d a_k(1+pa_k)x_k^2, \] which must change signs on the sphere, since the $a_k$ vary in sign, and so Proposition \ref{prop:Qineqc} does not apply. \hspace*{\fill}$\Diamond$\medskip \end{example} We conclude this section with a comment on the behaviour of $\sigma(s)$ at the other side of the strip of convergence, that is, as $s \to d/2$. \begin{remark}[$\sigma(s)$ as $s\to d/2$]\label{rem:spole} From Proposition \ref{prop:betares} and the fact that $\alpha_N(s)$ is an entire function, we know that $\sigma_N(s)$ has a simple pole at $s=d/2$ with the same residue as $\alpha(s)$ (which, by Proposition \ref{prop:sigmastrip}, is the analytic continuation of the limit $\sigma(s)$). \hspace*{\fill}$\Diamond$\medskip \end{remark} \section{Alternative procedures for Wigner limits}\label{sec:alt} The limit $\sigma (s) = \lim_{N \rightarrow \infty} \left[ \alpha_N (s) - \beta_N (s) \right]$, considered in the preceding sections, is built from the sum $\alpha_N (s)$, which sums over the lattice points in the hypercube $\{x \in \mathbb{R}^d : \|x\|_{\infty} \leq N\}$. In this section, we will show that some of the previous discussion carries over to the case when the hypercubes get replaced by more general sets. For simplicity, we restrict to the case of general hyperballs $\{x \in \mathbb{R}^d : \|x\| \leq N\}$ where $\| \cdot \|$ is any norm in $\mathbb{R}^d$, and consider \begin{eqnarray} \widehat{\alpha}_N (s) & := & \sum_{0 < \|n\| \leq N} \frac{1}{Q (n)^s}, \label{eq:alphah}\\ \widehat{\beta}_N (s) &: = & \int_{\|x\| \leq N} \frac{1}{Q (x)^s} \mathrm{d} x, \label{eq:betah} \end{eqnarray} as well as $\widehat{\sigma}_N := \widehat{\alpha}_N - \widehat{\beta}_N$. Again, if $\op{Re} s > d / 2$, then $\widehat{\alpha}_N (s)$ converges to the Epstein zeta function $\alpha (s) = Z_Q (s)$ as $N \rightarrow \infty$. Of particular interest is the case $\| \cdot \| = \| \cdot \|_2$, in which the lattice sum extends over the usual Euclidean $d$-balls of radius $N$. This case was considered in {\cite[Theorem 2]{latticesums-bbs}} when $d = 2$ and it was shown that the limit $\widehat{\sigma} (s) := \lim_{N \rightarrow \infty} \widehat{\sigma}_N (s)$ exists in the strip $1 / 3 < \op{Re} s < 1$ and coincides therein with the analytic continuation of $\alpha (s)$. As we will see below, this strip can be extended on the left-hand side, though not below $1 / 4$. In contrast to Remark \ref{rk:sigma0}, we note that $\widehat{\sigma}_N (0)$ usually does not converge. We therefore let $\lambda$ be the infimum of all values $\ell \geq 0$ such that \begin{equation} \widehat{\sigma}_N (0) =\# \{n \in \mathbbm{Z}^d : \|n\| \leq N\} - \op{vol} \{x \in \mathbb{R}^d : \|x\| \leq N\} - 1 = O (N^{\ell}) . \label{eq:sigmah0} \end{equation} The determination of $\lambda$, especially for the $p$-norms $\| \cdot \|_p$, is a famous problem and in several cases still open. In particular, when \ $d = 2$ and $\| \cdot \| = \| \cdot \|_2$, this is Gauss's circle problem. For a recent survey, we refer to {\cite{ikn-lattice04}}. A number of results on the values of $\lambda$ are discussed in the proof of Corollary \ref{cor:sigmahstrip2} and the remarks thereafter. We also recall the well-known fact, due to Weierstrass, that the balls in the $p$-norm have volume \[ \op{vol} \{x \in \mathbb{R}^d : \|x\|_p \leq N\} = \frac{2^d \Gamma^d (1 + 1 / p)}{\Gamma (1 + d / p)} N^d . \] We prove the following analog of Proposition \ref{prop:sigmastrip}, which includes {\cite[Theorem 2]{latticesums-bbs}} as the special case $d = 2$ and $\| \cdot \| = \| \cdot \|_2$. \begin{proposition} \label{prop:sigmahstrip}Let $\| \cdot \|$ be a norm on $\mathbb{R}^d$, and assume that $\lambda$ is the infimum of all values $\ell \geq 0$ such that \eqref{eq:sigmah0} holds. Further, let $Q$ be a positive definite quadratic form. Then the limit $\widehat{\sigma} (s) := \lim_{N \rightarrow \infty} \widehat{\sigma}_N (s)$ exists in the strip $\max (d / 2 - 1, \lambda / 2) < \op{Re} s < d / 2$ and coincides therein with the analytic continuation of $\alpha (s)$. \end{proposition} \begin{proof} As before, we fix $\sigma > 0$ as well as $R > 0$ and set $\Omega = \{s : \op{Re} s > \sigma, \hspace{1em} |s| < R\}$. All order terms below are uniform with respect to $s$ in the bounded region $\Omega$. In order to proceed along the lines of Proposition \ref{prop:sigmastrip}, we introduce \[ \tilde{\beta}_N (s) := \sum_{\|n\| \leq N} \int_{\|x\|_{\infty} \leq 1 / 2} \frac{1}{Q (n + x)^s} \mathrm{d} x, \] and observe that, by \eqref{eq:sigmah0} and the fact that all norms on $\mathbb{R}^d$ are equivalent, \begin{equation} \widehat{\beta}_N (s) - \tilde{\beta}_N (s) = O (N^{- 2 s + \ell}) \label{eq:betaht} \end{equation} for all values $\ell > \lambda$. On the other hand, set $\tilde{\sigma}_N (s) := \widehat{\alpha}_N (s) - \tilde{\beta}_N (s)$ and let \begin{eqnarray*} \delta_N (s) & := & \tilde{\sigma}_N (s) - \tilde{\sigma}_{N - 1} (s)\\ & = & \sum_{N - 1 < \|n\| \leq N} \int_{\|x\|_{\infty} \leq 1 / 2} \left[ \frac{1}{Q (n)^s} - \frac{1}{Q (n + x)^s} \right] \mathrm{d} x.\\ & = & O (N^{- 2 \sigma - 2}) \sum_{N - 1 < \|n\| \leq N} 1\\ & = & O (N^{d - 2 \sigma - 3}), \end{eqnarray*} where the estimates follow as in the proof of Proposition \ref{prop:sigmastrip}. Again, we conclude that the series $\delta (s) := \sum_{N = 1}^{\infty} \delta_N (s)$ converges in the half-plane $\op{Re} s > d / 2 - 1$ and defines an analytic function therein. By construction, \begin{equation} \delta (s) = \lim_{N \rightarrow \infty} \left[ \tilde{\sigma}_N (s) + \tilde{\beta}_0 (s) \right] . \label{eq:delta1t} \end{equation} Since $\tilde{\beta}_0 (s)$ is analytic for $\op{Re} s < d / 2$, it follows that the limit $\tilde{\sigma} (s) := \lim_{N \rightarrow \infty} \tilde{\sigma}_N (s)$ exists in the strip $d / 2 - 1 < \op{Re} s < d / 2$. In combination with \eqref{eq:betaht}, this shows that the limit $\widehat{\sigma} (s)$ exists in the strip $\max (d / 2 - 1, \lambda / 2) < \op{Re} s < d / 2$ and equals $\tilde{\sigma} (s)$ therein. For the second part of the claim, we proceed as in the proof of Proposition \ref{prop:sigmastrip} and observe that, for $\op{Re} s < d / 2$, \begin{equation} \widehat{\beta}_N (s) = N^{d - 2 s} \int_{\|x\| \leq 1} \frac{1}{Q (x)^s} \mathrm{d} x = N^{d - 2 s} \widehat{\beta}_1 (s) . \label{eq:betahN1} \end{equation} We note that Proposition \ref{prop:betares}, with the same proof, also applies to $\widehat{\beta}_N$ in place of $\beta_N$. In particular, $\widehat{\beta}_N$ and $\widehat{\beta}_0$ have meromorphic continuations to the entire complex plane, and the relation induced by \eqref{eq:betahN1} continues to hold. For $\op{Re} s > d / 2$, \[ \lim_{N \rightarrow \infty} \widehat{\beta}_N (s) = \lim_{N \rightarrow \infty} N^{d - 2 s} \widehat{\beta}_1 (s) = 0. \] We therefore have, for $\op{Re} s > d / 2$, \begin{equation} \delta (s) = \lim_{N \rightarrow \infty} \left[ \widehat{\alpha}_N (s) - \widehat{\beta}_N (s) + \tilde{\beta}_0 (s) \right] = \alpha (s) + \tilde{\beta}_0 (s) . \label{eq:deltaabh} \end{equation} On the other hand, it follows from \eqref{eq:betaht} and \eqref{eq:delta1t} that, for $\op{Re} s < d / 2$, \begin{equation} \delta (s) = \widehat{\sigma} (s) + \tilde{\beta}_0 (s) . \label{eq:deltasbh} \end{equation} Since both $\delta (s)$ and $\tilde{\beta}_0 (s)$ are meromorphic in the half-plane $\op{Re} s > d / 2 - 1$, comparing \eqref{eq:deltaabh} and \eqref{eq:deltasbh} proves that the analytic continuations of $\widehat{\sigma} (s)$ and $\alpha (s)$ agree. \end{proof} \begin{corollary}[Four and higher dimensions] \label{cor:sigmahstrip2}Let $Q$ be a positive definite quadratic form on $\mathbb{R}^d$ for $d \geq 4$. Then the limit \[ \widehat{\sigma} (s) = \lim_{N \rightarrow \infty} \left[ \sum_{0 <\|n\|_2 \leq N} \frac{1}{Q (n)^s} - \int_{\|x\|_2 \leq N} \frac{1}{Q (x)^s} \mathrm{d} x \right] \] exists in the strip $d / 2 - 1 < \op{Re} s < d / 2$ and coincides therein with the analytic continuation of $\alpha (s)$. \end{corollary} \begin{proof} We recall, see {\cite{ikn-lattice04}}, the fact that, for all $d \geq 5$, \[ \# \{n \in \mathbbm{Z}^d : \|n\|_2 \leq N\} - \op{vol} \{x \in \mathbb{R}^d : \|x\|_2 \leq N\} = O (N^{d - 2}), \] while, for $d = 4$, the right-hand side needs to be replaced with, for instance, the rather classical $O (N^2 (\log N))$, or the improved $O (N^2 (\log N)^{2 / 3})$ shown in {\cite{walfisz59}}. In any case, we conclude that, for all $d \geq 4$, the infimum $\lambda_d$ of all values $\ell \geq 0$ such that \eqref{eq:sigmah0} holds, is $\lambda_d = d - 2$. The claim therefore follows from Proposition \ref{prop:sigmahstrip}. \end{proof} \begin{remark}[Two and three dimensions] Thorough reports on the current status of the cases $d = 2$ and $d = 3$, missing in Corollary \ref{cor:sigmahstrip2}, can be found in {\cite{ikn-lattice04}}. In the case $d = 2$, it was shown by Hardy as well as Landau that $\lambda_2 \geq 1 / 2$. While it is believed that in fact $\lambda_2 = 1 / 2$, the best currently known bound is $\lambda \leq 131 / 208 \approx 0.6298$, obtained in {\cite{huxley-gauss2}}. For $d = 3$, it is known that $\lambda_3 \geq 1$ and it is believed that $\lambda_3 = 1$, in which case the conclusion of Corollary \ref{cor:sigmahstrip2} would also hold for $d = 3$. The smallest currently fully proven upper bound is $\lambda_3 \leq 21 / 16 = 1.3125$ from {\cite{hb-gauss3}}. \hspace*{\fill}$\Diamond$\medskip \end{remark} \begin{remark}[more general $p$-norms] Let us briefly note some results and their consequences for more general $p$-norms, again referring to {\cite{ikn-lattice04}} for further details and missing cases. Let $d \geq 2$. For integers $p > d + 1$ it is known that \[ \# \{n \in \mathbbm{Z}^d : \|n\|_p \leq N\} - \op{vol} \{x \in \mathbb{R}^d : \|x\|_p \leq N\} = O \left( N^{(d - 1) (1 - 1 / p)} \right), \] and that the exponent in this estimate cannot be improved. This result was obtained in {\cite{randol-gaussp}} for even $p$, and in {\cite{kraetzel-gaussp}} for odd $p$. In light of Proposition \ref{prop:sigmahstrip}, we conclude that the limit \[ \widehat{\sigma} (s) = \lim_{N \rightarrow \infty} \left[ \sum_{0 <\|n\|_p \leq N} \frac{1}{Q (n)^s} - \int_{\|x\|_p \leq N} \frac{1}{Q (x)^s} \mathrm{d} x \right] \] exists in the strip $(d - 1) (1 - 1 / p) / 2 < \op{Re} s < d / 2$ and coincides therein with the analytic continuation of $\alpha (s)$. We note that this strip shrinks to $d / 2 - 1 / 2 < \op{Re} s < d / 2$ as $p \rightarrow \infty$. In particular, for $d = 2$, the physically interesting value $\widehat{\sigma} (1 / 2)$ always exists and equals $\alpha \left( 1 / 2 \right)$. \hspace*{\fill}$\Diamond$\medskip \end{remark}

https://arxiv.org/abs/physics/0605197

Optimal Data-Based Binning for Histograms

Histograms are convenient non-parametric density estimators, which continue to be used ubiquitously. Summary quantities estimated from histogram-based probability density models depend on the choice of the number of bins. We introduce a straightforward data-based method of determining the optimal number of bins in a uniform bin-width histogram. By assigning a multinomial likelihood and a non-informative prior, we derive the posterior probability for the number of bins in a piecewise-constant density model given the data. In addition, we estimate the mean and standard deviations of the resulting bin heights, examine the effects of small sample sizes and digitized data, and demonstrate the application to multi-dimensional histograms.

\section{Introduction} Histograms are used extensively as nonparametric density estimators both to visualize data and to obtain summary quantities, such as the entropy, of the underlying density. However in practice, the values of such summary quantities depend on the number of bins chosen for the histogram, which given the range of the data dictates the bin width. The idea is to choose a number of bins sufficiently large to capture the major features in the data while ignoring fine details due to `random sampling fluctuations'. Several rules of thumb exist for determining the number of bins, such as the belief that between 5-20 bins is usually adequate (for example, \texttt{Matlab} uses 10 bins as a default). \citet{Scott:1979} and \cite{Freedman&Diaconis:1981} derived formulas for the optimal bin width by minimizing the integrated mean squared error of the histogram model $h(x)$ of the true underlying density $f(x)$, \begin{equation} L(h(x), f(x)) = \int{\bigl(h(x)-f(x)\bigr)^2}. \end{equation} For $N$ data points, the optimal bin width $v$ goes as $\alpha N^{-1/3}$, where $\alpha$ is a constant that depends on the form of the underlying distribution. Assuming that the data are normally distributed with a sample variance $s$ gives $\alpha = 3.49s$ (Scott, 1979), and \begin{equation} v_{\mathrm{scott}} = 3.49s N^{-1/3}. \end{equation} Given a fixed range $R$ for the data, the number of bins $M$ then goes as \begin{equation} M_{\mathrm{scott}} = \lceil \frac{R}{3.49 s} N^{1/3}\rceil.\label{eq:scott} \end{equation} Freedman and Diaconis report similar results, however they suggest choosing $\alpha$ to be twice the interquartile range of the data. While these appear to be useful estimates for unimodal densities similar to a Gaussian distribution, they are known to be suboptimal for multimodal densities. This is because they were derived assuming particular characteristics of the underlying density. In particular, the result by Freedman and Diaconis is not valid for some densities, such as the uniform density, since it derives from the assumption that the density $f$ satisfies $\int{f'^2>0}$. \citet{Stone:1984} chooses to minimize $L(h,f) - \int{f^2}$ and obtains a rule where one chooses the bin width $v$ to minimize \begin{equation}\label{eq:stone} K(v,M) = \frac{1}{v}\biggl( \frac{2}{N-1} - \frac{N+1}{N-1} \sum_{m=1}^{M}{\pi_i^2} \biggr) \end{equation} where $M$ is the number of bins and $\pi_i$ are the bin probabilities. \citet{Rudemo:1982} obtains a similar rule by applying cross-validation techniques with a Kullback-Leibler risk function. We approach this problem from a different perspective. Since the underlying density is not known, it is not reasonable to use an optimization criterion that relies on the error between our density model and the true density. Instead, we consider the histogram to be a piecewise-constant model of the underlying probability density. Using Bayesian probability theory we derive a straightforward algorithm that computes the posterior probability of the number of bins for a given data set. This enables one to objectively select an optimal piecewise-constant model describing the density function from which the data were sampled. \section{The Piecewise-Constant Density Model} We begin by considering the histogram as a piecewise-constant model of the probability density function from which $N$ data points were sampled. This model has $M$ bins with each bin having width $v_k$, where $k$ is used to index the bins. We further assume that the bins have equal width $v = v_k$ for all $k$, and together they encompass an entire width $V = Mv$.\footnote{For a one-dimensional histogram, $v_k$ is the width of the $k^{th}$ bin. In the case of a multi-dimensional histogram, this will be a multi-dimensional volume.} Each bin has a ``height'' $h_k$, which is the constant probability density over the region of the bin. Integrating this constant probability density $h_k$ over the width of the bin $v_k$ leads to a total probability mass of $\pi_k = h_k v_k$ for the bin. This leads to the following piecewise-constant model $h(x)$ of the unknown probability density function $f(x)$ \begin{equation} h(x) = \sum_{k = 1}^{M}{h_k~\Pi(x_{k-1}, x, x_k)}, \end{equation} where $h_k$ is the probability density of the $k^{th}$ bin with edges defined by $x_{k-1}$ and $x_k$, and $\Pi(x_{k-1}, x, x_k)$ is the boxcar function where \begin{equation} \Pi(x_a, x, x_b) = \left\{ \begin{array}{rl} 0 & \mbox{if}~~x < x_a\\ 1 & \mbox{if}~~x_a \leq x < x_b\\ 0 & \mbox{if}~~x_b \leq x \end{array} \right. \end{equation} Our density model can be re-written in terms of the bin probabilities $\pi_k$ as \begin{equation} h(x) = \frac{M}{V} \sum_{k = 1}^{M}{\pi_k~\Pi(x_{k-1}, x, x_k)}. \end{equation} It is important to keep in mind that the piecewise-constant density model is not a histogram in the usual sense. We use the bin heights to represent the probability density and not the probability mass of the bin. Furthermore, we will show that the mean value of the probability mass of a bin is computed in a slightly different way. Given $M$ bins and the normalization condition that the integral of the probability density equals unity, we are left with $M-1$ bin probabilities: $\pi_1, \pi_2, \ldots, \pi_{M-1}$, each describing the probability that samples will be drawn from each of the $M$ bins. The normalization condition requires that $\pi_M = 1-\sum_{k=1}^{M-1}{\pi_k}$. For simplicity, we assume that the bin alignment is fixed so that extreme data points lie precisely at the center of the extreme bins of the histogram (that is, the smallest sample is at the center of the leftmost bin, and similarly for the largest sample). As we will show, this technique is easily extended to multi-dimensional densities of arbitrarily high dimension. \subsection{The Likelihood of the Piecewise-Constant Model} The likelihood function is a probability density that when multiplied by $dx$ describes probability that a datum point $d_n$ is found to have a value in the infinitesimal range between $x$ and $x + dx$. For $d_n$ to have a value of $x$ occurring in the $k^{th}$ bin, the likelihood is simply the probability density in the region defined by the bin \begin{equation} p(d_n | \pi_k, M, I) = h_k = \frac{\pi_k}{v_k} \end{equation} where $I$ represents our prior knowledge about the problem, which includes the range of the data and the bin alignment. For equal width bins, the likelihood density reduces to \begin{equation} p(d_n | \pi_k, M, I) = \frac{M}{V}\pi_k. \end{equation} For $N$ independently sampled data points, the joint likelihood is given by \begin{equation} \label{eq:likelihood} p(\underline{d} | \underline{\pi}, M, I) = \biggl(\frac{M}{V}\biggr)^N \pi_1^{n_1} \pi_2^{n_2} \ldots \pi_{M-1}^{n_{M-1}} \pi_{M}^{n_M} \end{equation} where $\underline{d} = \{d_1, d_2, \ldots, d_N\}$ and $\underline{\pi} = \{\pi_1, \pi_2, \ldots, \pi_{M-1}\}$. Equation (\ref{eq:likelihood}) is data-dependent and describes the likelihood that the hypothesized piecewise-constant model accounts for the data. Individuals who recognize this as having the form of the multinomial distribution may be tempted to include its familiar normalization factor. However, it is important to note that this likelihood function is properly normalized as is, which we now demonstrate. For a single datum point $d$, the likelihood that it will take the value $x$ is \begin{equation} p(d=x | \underline{\pi}, M, I) = \frac{1}{v} \sum_{k = 1}^{M}{\pi_k~\Pi(x_{k-1}, x, x_k)}, \end{equation} where we have written $v = \frac{V}{M}$. Multiplying the probability density by $dx$ to get the probability and integrating over all possible values of $x$ we have \begin{align} \int_{-\infty}^{\infty}{dx~p(d=x | \underline{\pi}, M, I)} & = \int_{-\infty}^{\infty}{dx~\frac{1}{v} \sum_{k = 1}^{M}{\pi_k~\Pi(x_{k-1}, x, x_k)}}\\ & = \frac{1}{v} \sum_{k = 1}^{M}{\int_{-\infty}^{\infty}{dx~\pi_k~\Pi(x_{k-1}, x, x_k)}}\\ & = \frac{1}{v} \sum_{k = 1}^{M}{\pi_k~v}\\ & = \sum_{k = 1}^{M}{\pi_k}\\ & = 1. \end{align} \subsection{The Prior Probabilities} For the prior probability of the number of bins, we assign a uniform density \begin{equation} \label{eq:prior-for-M} p(M | I) = \left\{ \begin{array}{rl} C^{-1} & \mbox{if}~~1 \leq M \leq C\\ 0 & \mbox{otherwise} \end{array} \right. \end{equation} where $C$ is the maximum number of bins to be considered. This could reasonably be set to the range of the data divided by smallest non-zero distance between any two data points. We assign a non-informative prior for the bin parameters $\pi_1, \pi_2, \ldots, \pi_{M-1}$, the possible values of which lie within a simplex defined by the corners of an $M$-dimensional hypercube with unit side lengths \begin{equation} \label{eq:prior-for-pi's} p(\underline{\pi} | M, I) = \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(\frac{1}{2}\bigr)^M} \biggl[\pi_1 \pi_2 \cdots \pi_{M-1} \biggl(1-\sum_{i=1}^{M-1}{\pi_i}\biggr)\biggr]^{-1/2}. \end{equation} Equation (\ref{eq:prior-for-pi's}) is the Jeffreys's prior for the multinomial likelihood (\ref{eq:likelihood}) \citep{Jeffreys:1961,Box&Tiao:1992,Berger&Bernardo:1992}, and has the advantage in that it is also the conjugate prior to the multinomial likelihood. \subsection{The Posterior Probability} Using Bayes' Theorem, the posterior probability of the histogram model is proportional to the product of the priors and the likelihood \begin{equation} p(\underline{\pi}, M | \underline{d}, I) \propto p(\underline{\pi} | I)~p(M | I)~p(\underline{d} | \underline{\pi}, M, I). \end{equation} Substituting (\ref{eq:likelihood}), (\ref{eq:prior-for-M}), and (\ref{eq:prior-for-pi's}) gives the joint posterior probability for the piecewise-constant density model \begin{equation} \label{eq:joint-posterior} p(\underline{\pi}, M | \underline{d}, I) \propto \biggl(\frac{M}{V}\biggr)^N \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(\frac{1}{2}\bigr)^M}~ \pi_1^{n_1-\frac{1}{2}} \pi_2^{n_2-\frac{1}{2}} \ldots \pi_{M-1}^{n_{M-1}-\frac{1}{2}} \biggl(1-\sum_{i=1}^{M-1}{\pi_i}\biggr)^{n_M-\frac{1}{2}}, \end{equation} where $p(M|I)$ is absorbed into the implicit proportionality constant with the understanding that we will only consider a reasonable range of bin numbers. The goal is to obtain the posterior probability for the number of bins $M$. To do this we integrate the joint posterior over all possible values of $\pi_1, \pi_2, \ldots, \pi_{M-1}$ in the simplex. The expression we desire is written as a series of nested integrals over the $M-1$ dimensional parameter space of bin probabilities \begin{align} \label{eq:nonsimplified-nested-integrals} p(M | \underline{d}, I) \propto \biggl(\frac{M}{V}\biggr)^N \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(\frac{1}{2}\bigr)^M}~ \int^{1}_{0} & {d\pi_1~\pi_1^{n_1-\frac{1}{2}}} \int^{1-\pi_1}_{0} {d\pi_2~\pi_2^{n_2-\frac{1}{2}}} \ldots\\ & \ldots \int^{(1-\sum_{i=1}^{M-2}{\pi_i})}_{0} {d\pi_{M-1}~\pi_{M-1}^{n_{M-1}-\frac{1}{2}} \biggl(1-\sum_{i=1}^{M-1}{\pi_i}\biggr)^{n_M-\frac{1}{2}}}.\notag \end{align} In order to write this more compactly, we first define \begin{eqnarray} a_1 & = & 1\\ a_2 & = & 1-\pi_1 \notag \\ a_3 & = & 1-\pi_1-\pi_2 \notag \\ \vdots \notag \\ a_{M-1} & = & 1-\sum_{k=1}^{M-2}{\pi_k} \notag \end{eqnarray} and note the recursion relation \begin{equation} a_k = a_{k-1} - \pi_{k-1}.\label{eqn:recursion} \end{equation} These definitions greatly simplify the sum in the last term as well as the limits of integration \begin{align} \label{eq:nestedints} p(M | \underline{d}, I) \propto \biggl(\frac{M}{V}\biggr)^N \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(\frac{1}{2}\bigr)^M}~ \int^{a_1}_{0} & {d\pi_1~\pi_1^{n_1-\frac{1}{2}}} \int^{a_2}_{0} {d\pi_2~\pi_2^{n_2-\frac{1}{2}}} \ldots\\ & \ldots \int^{a_{M-1}}_{0} {d\pi_{M-1}~\pi_{M-1}^{n_{M-1}-\frac{1}{2}} (a_{M-1}-\pi_{M-1})^{n_M-\frac{1}{2}}}.\notag \end{align} To solve the set of nested integrals in (\ref{eq:nonsimplified-nested-integrals}), consider the general integral \begin{equation} I_k = \int^{a_k}_{0} {d\pi_k~\pi_k^{n_k-\frac{1}{2}} (a_k-\pi_k)^{b_k}} \end{equation} where $b_k \in \mathbb{R}^{+}$ and $b_k > 1/2$. This integral can be re-written as \begin{equation} I_k = a_k^{b_k} \int^{a_k}_{0} {d\pi_k~\pi_k^{n_k-\frac{1}{2}} \biggl( 1 - \frac{\pi_k}{a_k} \biggr)^{b_k}}. \end{equation} Setting $\displaystyle u_k = \frac{\pi_k}{a_k}$ we have \begin{eqnarray} I_k & = & a_k^{b_k} \int^{1}_{0} {du~a_k^{n_k+\frac{1}{2}} u^{n_k-\frac{1}{2}} (1 - u)^{b_k}} \nonumber\\ & = & a_k^{b_k+n_k+\frac{1}{2}} \int^{1}_{0} {du~u^{n_k-\frac{1}{2}} (1 - u)^{b_k}}, \nonumber \\ & = & a_k^{b_k+n_k+\frac{1}{2}} B(n_k + \frac{1}{2}, b_k + 1)\label{eq:intsoln} \end{eqnarray} where $B(\cdot)$ is the Beta function with \begin{equation} \label{eq:beta-ito-gamma} B(n_k + \frac{1}{2}, b_k + 1) = \frac{\Gamma\biggl( n_k + \frac{1}{2} \biggr) \Gamma(b_k + 1)}{\Gamma\biggl( n_k + \frac{1}{2} + b_k + 1 \biggr)}. \end{equation} To solve all of the integrals we rewrite $a_k$ in (\ref{eq:intsoln}) using the recursion formula (\ref{eqn:recursion}) \begin{equation} I_k = (a_{k-1}-\pi_{k-1})^{b_k+n_k+\frac{1}{2}} B(n_k + \frac{1}{2}, b_k + 1). \end{equation} By defining \begin{eqnarray} \label{eq:b-recursion} b_{M-1} & = & n_M-\frac{1}{2}\\ b_{k-1} & = & b_k+n_k+\frac{1}{2}\notag \end{eqnarray} we find \begin{eqnarray} b_{1} & = & N-n_1+\frac{M}{2}-\frac{3}{2}. \end{eqnarray} Finally, integrating (\ref{eq:nestedints}) gives \begin{equation} p(M | \underline{d}, I) \propto \biggl(\frac{M}{V}\biggr)^N \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(\frac{1}{2}\bigr)^M}~ \prod_{k=1}^{M-1}{B(n_k + \frac{1}{2}, b_k + 1)}, \end{equation} which can be simplified further by expanding the Beta functions using (\ref{eq:beta-ito-gamma}) \begin{align} p(M | \underline{d}, I) \propto \biggl(\frac{M}{V}\biggr)^N \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(\frac{1}{2}\bigr)^M}~ \frac{\Gamma(n_1 + \frac{1}{2}) \Gamma(b_1 + 1)}{\Gamma(n_1 + \frac{1}{2} + b_1 + 1)} \cdot \frac{\Gamma(n_2 + \frac{1}{2}) \Gamma(b_2 + 1)}{\Gamma(n_2 + \frac{1}{2} + b_2 + 1)} \cdot \cdots & \nonumber \\ \cdots \cdot \frac{\Gamma(n_{M-1} + \frac{1}{2}) \Gamma(b_{M-1} + 1)}{\Gamma(n_{M-1} + \frac{1}{2} + b_{M-1} + 1)} & \end{align} Using the recursion relation (\ref{eq:b-recursion}) for the $b_k$, we see that the general term $\Gamma(b_k+1)$ in each numerator, except the last, cancels with the denominator in the following term. This leaves \begin{equation} p(M | \underline{d}, I) \propto \biggl(\frac{M}{V}\biggr)^N \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(\frac{1}{2}\bigr)^M}~ \frac{\prod_{k=1}^{M}{\Gamma(n_k+\frac{1}{2})}}{\Gamma(n_1+b_1+\frac{3}{2})}, \end{equation} where we have used (\ref{eq:b-recursion}) to observe that $\Gamma(b_{M-1}+1) = \Gamma(n_M+1/2)$. Last, again using the recursion relation in (\ref{eq:b-recursion}) we find that $b_1 = N-n_1+\frac{M}{2}-\frac{3}{2}$, which results in our marginal posterior probability \begin{equation} \label{eq:posterior-for-M} p(M | \underline{d}, I) \propto \biggl(\frac{M}{V}\biggr)^N \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(\frac{1}{2}\bigr)^M}~ \frac{\prod_{k=1}^{M}{\Gamma(n_k+\frac{1}{2})}}{\Gamma(N+\frac{M}{2})}. \end{equation} The normalization of this posterior probability density depends on the actual data used. For this reason, we will work with the unnormalized posterior, and shall refer to it as the relative posterior. In optimization problems, it is often easier to maximize the logarithm of the posterior \begin{align} \label{eq:log-posterior-for-M} \log{p(M | \underline{d}, I)} = N \log{M} + \log{\Gamma\biggl(\frac{M}{2}\biggr)} - M \log{\Gamma\biggl(\frac{1}{2}\biggr)} - \log{\Gamma\biggl(N+\frac{M}{2}\biggr)} + ~~~~ &\\ + \sum_{k=1}^{M}{\log{\Gamma\biggl(n_k+\frac{1}{2}\biggr)}} + K, & \nonumber \end{align} where $K$ represents the sum of the volume term and the logarithm of the implicit proportionality constant. The optimal number of bins $\hat{M}$ is found by identifying the mode of the logarithm of the marginal posterior \begin{equation} \hat{M} = \underset{M}{\textrm{arg max}} \{\log{p(M | \underline{d}, I)}\}. \end{equation} Such a result is reassuring, since it is independent of the order in which the bins are counted. Many software packages are equipped to quickly compute the log of the gamma function. However, for more basic implementations, the following definitions from \cite{Abramowitz&Stegun:1972} can be used for integer~$m$ . \begin{align} & \log{\Gamma(m)} = \sum_{k=1}^{m-1}{\log{k}}\\ & \log{\Gamma \biggl( m+\frac{1}{2} \biggr)} = \frac{1}{2}\log{\pi} - n\log{2} + \sum_{k=1}^{m}{\log{(2k-1)}} \end{align} Equation (\ref{eq:log-posterior-for-M}) allows one to easily identify the number of bins $M$ which optimize the posterior. We call this technique the \texttt{optBINS} algorithm and provide a \texttt{Matlab} code implementation in Appendix 1. \section{The Posterior Probability for the Bin Height} In order to obtain the posterior probability for the probability mass of a particular bin, we begin with the joint posterior (\ref{eq:joint-posterior}) and integrate over all the other bin probability masses. Since we can consider the bins in any order, the resulting expression is similar to the multiple nested integral in (\ref{eq:nonsimplified-nested-integrals}) except that the integral for one of the $M-1$ bins is not performed. Treating the number of bins as a given, we can use the product rule to get \begin{equation} p(\underline{\pi} | \underline{d}, M, I) = \frac{p(\underline{\pi}, M | \underline{d}, I)}{p(M | \underline{d}, I)} \end{equation} where the numerator is given by (\ref{eq:joint-posterior}) and the denominator by (\ref{eq:posterior-for-M}). Since the bins can be treated in any order, we derive the marginal posterior for the first bin and generalize the result for the $k^{th}$ bin. The marginal posterior is \begin{align} p(\pi_1 | \underline{d}, M, I) = \frac{(\frac{M}{V})^N \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(\frac{1}{2}\bigr)^M}}{p(M | \underline{d},I)} \pi_1^{n_1-\frac{1}{2}} & \int^{a_2}_{0} {d\pi_2~\pi_2^{n_2-\frac{1}{2}}} \int^{a_3}_{0} {d\pi_3~\pi_3^{n_3-\frac{1}{2}}} \ldots\\ & \ldots \int^{a_{M-1}}_{0} {d\pi_{M-1}~\pi_{M-1}^{n_{M-1}-\frac{1}{2}} (a_{M-1}-\pi_{M-1})^{n_M-\frac{1}{2}}}. \notag \end{align} Evaluating the integrals and substituting (\ref{eq:posterior-for-M}) into the denominator we get \begin{equation} p(\pi_1 | \underline{d}, M, I) = \frac{\prod_{k=2}^{M}{B(n_k+\frac{1}{2}, b_k+1)}}{\prod_{k=1}^{M}{B(n_k+\frac{1}{2}, b_k+1)}} \pi_1^{n_1-\frac{1}{2}} (1-\pi_1)^{b_1} \end{equation} Cancelling terms and explicitly writing $b_1$, the marginal posterior for $\pi_1$ is \begin{equation} p(\pi_1 | \underline{d}, M, I) = \frac{\Gamma(N+\frac{M}{2})}{\Gamma(n_1+\frac{1}{2}) \Gamma(N-n_1+\frac{M-1}{2})} \pi_1^{n_1-\frac{1}{2}} (1-\pi_1)^{N-n_1+\frac{M-3}{2}}, \end{equation} which can easily be verified to be normalized by integrating $\pi_1$ over its entire possible range from 0 to 1. Since the bins can be considered in any order, this is a general result for the $k^{th}$ bin \begin{equation} \label{eq:posterior-bin-height} p(\pi_k | \underline{d}, M, I) = \frac{\Gamma(N+\frac{M}{2})}{\Gamma(n_k+\frac{1}{2}) \Gamma(N-n_k+\frac{M-1}{2})} \pi_k^{n_k-\frac{1}{2}} (1-\pi_k)^{N-n_k+\frac{M-3}{2}}. \end{equation} The mean bin probability mass can be found from its expectation \begin{equation} \langle\pi_k\rangle = \int_0^1{d\pi_k~\pi_k~p(\pi_k | \underline{d}, M, I)}, \end{equation} which substituting (\ref{eq:posterior-bin-height}) gives \begin{equation} \langle\pi_k\rangle = \frac{\Gamma(N+\frac{M}{2})}{\Gamma(n_k+\frac{1}{2}) \Gamma(N-n_k+\frac{M-1}{2})} \int_0^1{d\pi_k~\pi_k^{n_k+\frac{1}{2}} (1-\pi_k)^{N-n_k+\frac{M-3}{2}}}. \end{equation} The integral again gives a Beta function, which when written in terms of Gamma functions is \begin{equation} \langle\pi_k\rangle = \frac{\Gamma(N+\frac{M}{2})}{\Gamma(n_k+\frac{1}{2}) \Gamma(N-n_k+\frac{M-1}{2})} \cdot \frac{\Gamma(n_k+\frac{3}{2}) \Gamma(N-n_k+\frac{M-1}{2})}{\Gamma(N+\frac{M}{2}+1)}. \end{equation} Using the fact that $\Gamma(x+1) = x\Gamma(x)$ and cancelling like terms, we find that \begin{equation} \label{eq:mean-of-bin-heights} \langle\pi_k\rangle = \frac{n_k+\frac{1}{2}}{N+\frac{M}{2}}. \end{equation} The mean probability density for bin $k$ (the bin height) is simply \begin{equation} \label{eq:mean-of-prob-density} \mu_k = \langle h_k \rangle = \frac{\langle \pi_k \rangle}{v_k} = \biggl(\frac{M}{V}\biggr) \biggl(\frac{n_k+\frac{1}{2}}{N+\frac{M}{2}}\biggr). \end{equation} It is an interesting result that bins with no counts still have a non-zero probability. This makes sense since no lack of evidence can ever prove conclusively that an event occurring in a given bin is impossible---just less probable. The Jeffrey's prior effectively places one-half of a datum point in each bin. The variance of the height of the $k^{th}$ bin is found similarly by \begin{equation} \sigma_k^2 = \biggl(\frac{M}{V}\biggr)^2 \bigl(\langle\pi_k^2\rangle - \langle\pi_k\rangle^2\bigr), \end{equation} which gives \begin{equation} \label{eq:variance-of-prob-density} \sigma_k^2 = \biggl(\frac{M}{V}\biggr)^2 \biggl(\frac{(n_k+\frac{1}{2})(N-n_k+\frac{M-1}{2})} {(N+\frac{M}{2}+1)(N+\frac{M}{2})^2}\biggr). \end{equation} Thus, given the optimal number of bins found by maximizing (\ref{eq:log-posterior-for-M}), the mean and variance of the bin heights are found from (\ref{eq:mean-of-prob-density}) and (\ref{eq:variance-of-prob-density}), which allow us to construct an explicit histogram model of the probability density and perform computations and error analysis. Note that in the case where there is one bin (\ref{eq:variance-of-prob-density}) gives a zero variance. \begin{figure} \centering \makebox{\includegraphics[scale=0.8]{figures/fig-1/fig-1-lo.eps}} \caption{\label{fig:01}To demonstrate the technique, 1000 samples were drawn from four different probability density functions (pdf) above. (A) The optimal histogram for 1000 samples drawn from a Gaussian density function is superimposed over a 100-bin histogram that shows the distribution of data samples. (B) The log posterior probability of the number of bins peaks at 14 bins for these 1000 data sampled from the Gaussian density. (C) Samples shown are from a 4-step piecewise-constant density function. The optimal binning describes this density accurately since (D) the log posterior peaks at four bins. (E) These data were sampled from a uniform density as verified by the log posterior probability (F). (G) shows a more complex example---three Gaussian peaks plus a uniform background. (H) The posterior, which peaks at 52 bins, demonstrates clearly that the data themselves support this detailed picture of the pdf.} \end{figure} \section{Results} \subsection{Demonstration using One-dimensional Histograms} In this section we demonstrate the utility of this method for determining the optimal number of bins in a histogram model of several different data sets. Here we consider 1000 data points sampled from four different probability density functions. The optimal histogram for 1000 data points drawn from a Gaussian distribution $\mathcal{N}(0,1)$ is shown in Figure 1A, where it is superimposed over a 100-bin histogram showing the density of the sampled points. Figure 1B shows that the logarithm of the posterior probability (\ref{eq:log-posterior-for-M}) peaks at 14 bins. Figure 1C shows the optimal binning for data sampled from a 4-step piecewise-constant density. The logarithm of the posterior (Figure 1D) peaks at 4 bins, which indicates that the algorithm can correctly detect the 4-step structure. In figures 1E and F, we see that samples drawn from a uniform density were best described by a single bin. This result is significant, since entropy estimates computed from these data would be biased if multiple bins were used. Last, we consider a density function that consists of a mixture of three sharply-peaked Gaussians with a uniform background (Figure 1G). The posterior peaks at 52 bins indicating that the data warrant a detailed model (Figure 1H). The spikes in the log posterior are due to the fact that the bin edges are fixed. The log posterior is large at values of $M$ where the bins happen to line up with the Gaussians, and small when they are misaligned. This last example demonstrates one of the weaknesses of the equal bin-width model, as many bins are needed to describe the uniform density between the three narrow peaks. \begin{figure} \centering \makebox{\includegraphics[scale=0.75]{figures/fig-2/Figure-2-lo.eps}} \caption{\label{fig:02} Four techniques are compared: \texttt{optBINS} (solid), AIC (dot-dash), Stone (dash), and Scott (dot). For each number of data points $N$, 50 separate sets of samples were drawn from a Gaussian distribution $\mathcal{N}(0,1)$. With each set of samples, each method was used to determine the optimal number of bins $M$ and the Integrated Square Error (ISE) between each of the 50 modelled density functions and the original Gaussian density. Mean and standard deviations of these results were computed for the 50 samples. (A) This pair of figures (top: inset, bottom: complete figure) shows the mean and standard deviation of the number of bins $M$ determined by each algorithm. For large numbers of data points $N>5000$, \texttt{optBINS} consistently chooses a smaller number of bins than the other techniques. (B) The figures on the right compare the ISE between the modelled density function and the true density function. The present algorithm \texttt{optBINS} consistently outperforms the other methods, despite the fact that Scott's Rule is derived to asymptotically optimize the ISE when the sampling distribution is a Gaussian.} \end{figure} \subsection{Comparison to Other Techniques} We now compare our results with those obtained using Scott's Rule, Stone's Rule, and the Akaike model selection criterion (\cite{Akaike:1974}). Since Freedman and Diaconis' (F\&D) method has the same functional form as Scott's Rule, the results using F\&D are not presented here. Their technique leads to a greater number of bins than Scott's Rule, which, as we will show, already prescribes more bins than are warranted by the data. Akaike's method, applied to histograms in \cite{Hartigan:1996}, balances the logarithm of the likelihood of the model against the number of model parameters. The number of bins is chosen to maximize \begin{equation}\label{eq:AIC} AIC(M) = \log p(d_n | \pi_k, M, I) - M. \end{equation} Since Scott's Rule was derived to be asymptotically optimal for Gaussian distributions, we limited our comparison to Gaussian-distributed data. We tested data sets with $12$ different numbers of samples including $N = \{200, 500, 1000, 2000, \cdots, 1000000\}$. For each $N$ we tested $50$ different histograms, each with $N$ samples drawn from a Gaussian distribution $\mathcal{N}(0,1)$, for a total of $700$ histograms in this analysis. The optimal number of bins was found using Scott's Rule (\ref{eq:scott}), Stone's Rule (\ref{eq:stone}), Akaike's AIC (\ref{eq:AIC}) and the present \texttt{optBINS} algorithm (\ref{eq:log-posterior-for-M}). The quality of fit was quantified using the Integrated Square Error (ISE), which is the criterion for which Scott's Rule is asymptotically optimized. This is \begin{equation} ISE = \int_{-\infty}^{\infty}{dx~\biggl(h(x,M) - \frac{1}{\sqrt{2 \pi}} \exp\bigl(-\frac{x^2}{2}\bigr)\biggr)^2}, \end{equation} where $h(x,M)$ is the piecewise-constant histogram model with $M$ bins. Figure \ref{fig:02}A shows the mean number of bins $M$ found using each of four methods for the values of the number of samples $N$ tested. The present algorithm \texttt{optBINS} tends to choose the least number of bins, and does so consistently for $N>5000$. More importantly, Figure \ref{fig:02}B shows the mean ISE between the modelled density function and the correct Gaussian distribution $\mathcal{N}(0,1)$. Since Scott's Rule was derived to asymptotically optimize the ISE, it may be surprising to see that \texttt{optBINS} outperforms each of these methods with respect to this measure. While Scott's Rule may be optimal as $N \rightarrow \infty$, \texttt{optBINS}, which is based on a Bayesian solution, is optimal for finite $N$. Furthermore if a log probability (or log-odds ratio) criterion were to be used, \texttt{optBINS} would be optimal by design. \begin{figure} \centering \makebox{\includegraphics[scale=0.75]{figures/fig-3/Figure-3-lo.eps}} \caption{\label{fig:small-samples}These figures demonstrate the behavior of the relative log posterior for small numbers of samples. (A) With only $N=2$ samples, the log posterior is maximum when the samples are in the same bin $M=1$. For $M>1$, the log posterior follows the function described in (\ref{eq:two-samples}) in the text. (B) The relative log posterior is slightly more complicated for $N=3$. For $M=1$ all three points lie in the same bin. As $M$ increases, two data points are in one bin and the remaining datum point is in another bin. The functional form is described by (\ref{eq:N=3,2|1}). Eventually, all three data points lie in separate bins and the relative log posterior is given by (\ref{eq:N=3,1|1|1}). (C) The situation is more complicated still for $N=5$ data points. As $M$ increases, a point is reached when, depending on the particular value of $M$, the points will be in separate bins. As $M$ changes value, two points may again fall into the same bin. This gives rise to this oscillation in the log posterior. Once all points are in separate bins, the behavior follows a well-defined functional form (\ref{eq:M>N}). (D) This plot shows the behavior for a large number of data points $N=200$. The log posterior now displays a more well-defined mode indicating that there is a well-defined optimal number of bins. As $M$ approaches 10000 to 100000 bins, one can see some of the oscillatory behavior demonstrated in the small $N$ cases.} \end{figure} \section{Effects of Small Sample Size} \subsection{Small Samples and Asymptotic Behavior} It is instructive to observe how this algorithm behaves in situations involving small sample sizes. We begin by considering the extreme case of two data points $N=2$. In the case of a single bin, $M=1$, the posterior probability reduces to \begin{eqnarray} p(M=1 | d_1, d_2, I) & \propto & M^N \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(\frac{1}{2}\bigr)^M}~ \frac{\prod_{k=1}^{M}{\Gamma(n_k+\frac{1}{2})}}{\Gamma(N+\frac{M}{2})} \nonumber \\ & \propto & 1^2 \frac{\Gamma\bigl(\frac{1}{2}\bigr)}{\Gamma\bigl(\frac{1}{2}\bigr)^1} \frac{\Gamma\bigl(2+\frac{1}{2}\bigr)}{\Gamma\bigl(2+\frac{1}{2}\bigr)} = 1, \end{eqnarray} so that the log posterior is zero. For $M>1$, the two data points lie in separate bins, resulting in \begin{eqnarray} \label{eq:two-samples} p(M | d_1, d_2, I) & \propto & M^N \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(\frac{1}{2}\bigr)^M}~ \frac{\prod_{k=1}^{M}{\Gamma(n_k+\frac{1}{2})}}{\Gamma(N+\frac{M}{2})} \nonumber \\ & \propto & M^2 \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(\frac{1}{2}\bigr)^M}~ \frac{\Gamma(1+\frac{1}{2})^2 \Gamma(\frac{1}{2})^{M-2}}{\Gamma(2+\frac{M}{2})}\nonumber \\ & \propto & M^2 \frac{\Gamma(\frac{3}{2})^2}{\Gamma\bigl(\frac{1}{2}\bigr)^2}~ \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma(2+\frac{M}{2})}\nonumber \\ & \propto & \frac{1}{2} \cdot \frac{M}{1+\frac{M}{2}}. \end{eqnarray} Figure \ref{fig:small-samples}A shows the log posterior which starts at zero for a single bin, drops to $\log(\frac{1}{2})$ for $M=2$ and then increases monotonically approaching zero in the limit as $M$ goes to infinity. The result is that a single bin is the most probable solution for two data points. For three data points in a single bin ($N=3$ and $M=1$), the posterior probability is one, resulting in a log posterior of zero. In the $M>1$ case where there are two data points in one bin and one datum point in another, the posterior probability is \begin{equation}\label{eq:N=3,2|1} p(M | d_1, d_2, d_3, I) \propto \frac{3}{4} \cdot \frac{M^2}{(2+\frac{M}{2})(1+\frac{M}{2})}, \end{equation} and for each point in a separate bin we have \begin{equation}\label{eq:N=3,1|1|1} p(M | d_1, d_2, d_3, I) \propto \frac{1}{4} \cdot \frac{M^2}{(2+\frac{M}{2})(1+\frac{M}{2})}. \end{equation} While the logarithm of the posterior in (\ref{eq:N=3,2|1}) can be greater than zero, as $M$ increases, the data points eventually fall into separate bins. This causes the posterior to change from (\ref{eq:N=3,2|1}) to (\ref{eq:N=3,1|1|1}) resulting in a dramatic decrease in the logarithm of the posterior, which then asymptotically increases to zero as $M \rightarrow \infty$. This behavior is shown in Figure \ref{fig:small-samples}B. The result is that either one or two bins will be optimal depending on the relative positions of the data points. More rich behavior can be seen in the case of $N=5$ data points. The results again (Figure \ref{fig:small-samples}C) depend on the relative positions of the data points with respect to one another. In this case the posterior probability switches between two types of behavior as the number of bins increase depending on whether the bin positions force two data points together in the same bin or separate them into two bins. The ultimate result is a ridiculous \emph{maximum a posteriori} solution of 57 bins. Clearly, for a small number of data points, the mode depends sensitively on the relative positions of the samples in a way that is not meaningful. In these cases there are too few data points to model a density function. With a larger number of samples, the posterior probability shows a well-defined mode indicating a well-determined optimal number of bins. In the general case of $M > N$ where each of the $N$ data points is in a separate bin, we have \begin{equation}\label{eq:M>N} p(M | \underline{d}, I) \propto \biggl(\frac{M}{2}\biggr)^N \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(N+\frac{M}{2}\bigr)}, \end{equation} which again results in a log posterior that asymptotically approaches zero as $M \rightarrow \infty$. Figure \ref{fig:small-samples}D demonstrates these two effects for $N=200$. This also can be compared to the log posterior for $1000$ Gaussian samples in Figure \ref{fig:01}B. \begin{figure}[t] \centering \makebox{\includegraphics[scale=0.3]{figures/fig-4/Figure-4-lo.eps}} \caption{\label{fig:sufficient}(A) An optimal density model ($M = 19$) for $N = 30$ data points sampled from a Gaussian distribution. The fact that the error bars on the bin probabilities are as large as the probabilities themselves indicates that this is a poor estimate. (B) The log posterior probability for the number of bins possesses no well-defined peak, and is instead reminiscent of noise. (C) This plot shows the standard deviation of the estimated number of bins $M$ for $1000$ data sets of $N$ points, ranging from $2$ to $200$, sampled from a Gaussian distribution. The standard deviation stabilizes around $\sigma_M = 2$ bins for $N > 150$ indicating the inherent level of uncertainty in the problem. This suggests that one requires at least $150$ data points to consistently perform such probability density estimates, and can perhaps get by with as few as $100$ data points in some cases.} \end{figure} \subsection{Sufficient Data} This investigation on the effects of small sample size raises the question as to how many data points are needed to estimate the probability density function. The general shape of a healthy log posterior reflects a sharp initial rise to a well-defined peak, and a gradual fall-off as the number of bins $M$ increases from one (eg. Fig. \ref{fig:01}B, Fig. \ref{fig:small-samples}D. With small sample sizes, however, one finds that the bin heights have large error bars (Figure \ref{fig:sufficient}A) so that $\mu_i \simeq \sigma_i$, and that the log posterior is multi-modal (Figure \ref{fig:sufficient}B) with no clear peak. We tested our algorithm on data sets with $199$ different sample sizes from $N = 2$ to $N = 200$. One thousand data sets were drawn from a Gaussian distribution for each value of $N$. The standard deviation of the number of bins obtained for these 1000 data sets at a given value if $N$ was used as an indicator of the stability of the solution. Figure \ref{fig:sufficient}C shows a plot of the standard deviation of the number of bins selected for the 1000 data sets at each value of $N$. As we found above, with two data points, the optimal solution is always one bin giving a standard deviation of zero. This increases dramatically as the number of data points increases, as we saw in our example with $N = 5$ and $M = 57$. This peaks around $N = 15$ and slowly decreases as $N$ increases further. The standard deviation of the number of bins decreased to $\sigma_M < 5$ for $N > 100$, and stabilized to $\sigma_M \simeq 2$ for $N > 150$. While 30 samples may be sufficient for estimating the mean and variance of a density function known to be Gaussian, it is clear that more samples are needed to reliably estimate the shape of an unknown density function. In the case where the data are described by a Gaussian, it would appear that at least $100$ samples, and preferentially $150$ samples, are required to accurately and consistently infer the shape of the density function. By examining the shape of the log posterior, one can easily determine whether one has sufficient data to estimate the density function. In the event that there are too few samples to perform such estimates, one can either incorporate additional prior information or collect more data. \section{Digitized Data} Due to the way that computers represent data, all data are essentially represented by integers \citep{Bayman&Broadhurst}. In some cases, the data samples have been intentionally rounded or truncated, often to save storage space or transmission time. It is well-known that any non-invertible transformation, such as rounding, destroys information. Here we investigate how severe losses of information due to rounding or truncation affects the \texttt{optBINS} algorithm. \begin{figure} \centering \makebox{\includegraphics[scale=0.3]{figures/fig-5/Figure-5-lo.eps}} \caption{\label{fig:rounded} $N=1000$ data points were sampled from a Gaussian distribution $\mathcal{N}(0,1)$. The top plots show (A) the estimated density function using optimal binning and (B) the relative log posterior, which exhibits a well-defined peak at $M=11$ bins. The bottom plots reflect the results using the same data set after it has been rounded with $\Delta x = 0.1$ to keep only the first decimal place. (C) There is no optimal binning as the algorithm identifies the discrete structure as being a more salient feature than the overall Gaussian shape of the density function. (D) The relative log posterior displays no well-defined peak, and in addition, for large numbers of $M$ displays a monotonically increase given by (\ref{eq:M>N}) that asymptotes to a positive value. This indicates that the data have been severely rounded.} \end{figure} When data are digitized via truncation or rounding, the digitization is performed so as to maintain a resolution that we will denote by $\Delta x$. That is, if the data set has values that range from 0 to 1, and we represent these numbers with an 8 bits, the minimum resolution we can maintain is $\Delta x = 1/2^8 = 1/256$. For a sufficiently large data set (in this example $N>256$) it will be impossible for every datum point to be in its own bin when the number of bins is greater than a critical number, $M > M_{\Delta x}$, where \begin{equation} M_{\Delta x} = \frac{V}{\Delta x}, \end{equation} and $V$ is the range of the data considered. Once $M > M_{\Delta x}$ the number of populated bins $P$ will remain unchanged since the bin width $w$ for $M > M_{\Delta x}$ will be smaller than the digitization resolution, $w < \Delta x$. For all bin numbers $M > M_{\Delta x}$, there will be $P$ populated bins with populations $n_1, n_2, \ldots, n_P$. This leads to a form for the marginal posterior probability for $M$ (\ref{eq:posterior-for-M}) that depends only on the number of instances of each discrete value that was recorded, $n_1, n_2, \ldots, n_P$. Since these values do not vary for $M > M_{\Delta x}$, the marginal posterior can be viewed solely as a function of $M$ \begin{equation} p(M | \underline{d}, I) \propto \biggl(\frac{M}{2}\biggr)^N \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(N+\frac{M}{2}\bigr)} \cdot 2^N \frac{\prod_{p=1}^{P}{\Gamma(n_p+\frac{1}{2})}}{\Gamma\bigl(\frac{1}{2}\bigr)^P}, \end{equation} where the product over $p$ is over populated bins only. Comparing this to (\ref{eq:M>N}), the function on the right-hand side asymptotically approaches a value greater than one---so that its logarithm increases asymptotically to a value greater than zero. As the number of bins $M$ increases, the point is reached where the data can not be further separated; call this point $M_{crit}$. In this situation, there are $n_p$ data points in the $p^{th}$ bin and the posterior probability can be written as \begin{equation}\label{eq:picket-fence-posterior} p(M | \underline{d}, I) \propto \biggl(\frac{M}{2}\biggr)^N \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(N+\frac{M}{2}\bigr)} \cdot \prod_{p = 1}^P {(2n_p-1)!!}, \end{equation} where $!!$ denotes the double factorial. For $M > M_{crit}$, as $M \rightarrow \infty$, the log posterior asymptotes to $\sum_{p = 1}^P {\log((2n_p-1)!!)}$, which can be further simplified to \begin{equation}\label{eq:picket-fence-log-asymptote} \sum_{p = 1}^P{\log((2n_p-1)!!)} = (P-N)\log(2) + \sum_{p = 1}^P{\sum_{s=n_p}^{2n_p-1}{\log s}}. \end{equation} To test for excessive rounding or truncation, the mode of $\log p(M | \underline{d}, I)$ for $M < M_{crit}$ should be compared to (\ref{eq:picket-fence-log-asymptote}) above. If the latter is larger, than the discrete nature of the data is a more significant feature than the general shape of the underlying probability density function. When this is the case, a reasonable histogram model of the density function can still be obtained by adding a uniformly-distributed random number, with a range defined by the resolution $\Delta x$, to each datum point \citep{Bayman&Broadhurst}. While this will produce the best histogram possible given the data, this will not recover the lost information. \begin{figure} \centering \makebox{\includegraphics[scale=0.85]{figures/fig-6/Figure-6-lo.eps}} \caption{\label{fig:06}10000 samples were drawn from a two-dimensional Gaussian density to demonstrate the optimization of a two-dimensional histogram. (A) The relative logarithm of the posterior probability is plotted as a function of the number of bins in each dimension. The normalization constant has been neglected in this plot, resulting in positive values of the log posterior. (B) This plot shows the relative log posterior as a contour plot. The optimal number of bins is found to be $12 \times 14$. (C) The optimal histogram for this data set. (D) The histogram determined using Stone's method has $27 \times 28$ bins. This histogram is clearly sub-optimal since it highlights random variations that are not representative of the density function from which the data were sampled.} \end{figure} \section{Multi-Dimensional Histograms} In this section, we demonstrate that our method can be extended naturally to multi-dimensional histograms. We begin by describing the method for a two-dimensional histogram. The constant-piecewise model $h(x,y)$ of the two-dimensional density function $f(x,y)$ is \begin{equation} h(x,y;M_x,M_y) = \frac{M}{V} \sum_{j = 1}^{M_x}{\sum_{k = 1}^{M_y}{\pi_{j,k}~\Pi(x_{j-1}, x, x_j)}\Pi(y_{k-1}, y, y_k)}, \end{equation} where $M = M_x M_y$, $V$ is the total area of the histogram, $j$ indexes the bin labels along $x$, and $k$ indexes them along $y$. Since the $\pi_{j,k}$ all sum to unity, we have $M-1$ bin probability density parameters as before, where $M$ is the total number of bins. The likelihood of obtaining a datum point $d_n$ from bin $(j,k)$ is still simply \begin{equation} p(d_n | \pi_{j,k}, M_x, M_y, I) = \frac{M}{V}\pi_{j,k}. \end{equation} The previous prior assignments result in the posterior probability \begin{equation} \label{eq:joint-posterior} p(\underline{\pi}, M_x, M_y | \underline{d}, I) \propto \biggl(\frac{M}{V}\biggr)^N \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(\frac{1}{2}\bigr)^M}~ \prod_{j = 1}^{M_x}{\prod_{k = 1}^{M_y}{\pi_{j,k}^{n_{j,k}-\frac{1}{2}}}}, \end{equation} where $\pi_{M_x,M_y}$ is $1$ minus the sum of all the other bin probabilities. The order of the bins in the marginalization does not matter, which gives a result similar in form to the one-dimensional case \begin{equation} \label{eq:posterior-for-MxMy} p(M_x, M_y | \underline{d}, I) \propto \biggl(\frac{M}{V}\biggr)^N \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(\frac{1}{2}\bigr)^M}~ \frac{\prod_{j=1}^{M_x}{\prod_{k=1}^{M_y}{\Gamma(n_{j,k}+\frac{1}{2})}}}{\Gamma(N+\frac{M}{2})}, \end{equation} where $M = M_x M_y$. For a D-dimensional histogram, the general result is \begin{equation} p(M_1, \cdots, M_D | \underline{d}, I) \propto \biggl(\frac{M}{V}\biggr)^N \frac{\Gamma\bigl(\frac{M}{2}\bigr)}{\Gamma\bigl(\frac{1}{2}\bigr)^M}~ \frac{\prod_{i_1=1}^{M_1}{\cdots\prod_{i_D=1}^{M_D}{\Gamma(n_{i_1,\ldots,i_D}+\frac{1}{2})}}}{\Gamma(N+\frac{M}{2})}, \end{equation} where $M_i$ is the number of bins along the $i^{th}$ dimension, $M$ is the total number of bins, $V$ is the $D$-dimensional volume of the histogram, and $n_{i_1,\ldots,i_D}$ indicates the number of counts in the bin indexed by the coordinates $(i_1, \ldots,i_D)$. Note that the result in (\ref{eq:log-posterior-for-M}) can be used directly for a multi-dimensional histogram simply by relabelling the multi-dimensional bins with a single index. Figure \ref{fig:06} demonstrates the procedure on a data set sampled from a two-dimensional Gaussian. In this example, 10000 samples were drawn from a two-dimensional Gaussian density. Figure \ref{fig:06}A shows the relative logarithm of the posterior probability plotted as a function of the number of bins in each dimension. The same surface is displayed as contour plot in Figure \ref{fig:06}B, where we find the optimal number of bins to be $12 \times 14$. Figure \ref{fig:06}C shows the optimal two-dimensional histogram model. Note that modelled density function is displayed in terms of the number of counts rather than the probability density, which can be easily computed using (\ref{eq:mean-of-prob-density}) with error bars computed using (\ref{eq:variance-of-prob-density}). In Figure \ref{fig:06}D, we show the histogram obtained using Stone's method, which results in a $27 \times 28$ array of bins. This is clearly a sub-optimal model since random sampling variations are easily visible. \section{Discussion} The optimal binning algorithm presented in this paper, \texttt{optBINS}, relies on finding the mode of the marginal posterior probability of the number of bins in a piecewise-constant density function. This posterior probability originates as a product of the likelihood of the density parameters given the data and the prior probability of those same parameter values. As the number of bins increases, the model can better fit the data, which leads to an increase in the likelihood function. However, by introducing additional bins, the prior probability, which must sum to unity, is now spread out over a parameter space of greater dimensionality resulting in a decrease of the prior probability. Thus a density model with more bins will result in a greater likelihood, but also in a smaller prior probability. Since the posterior is a product of these two functions, the maximum of the posterior probability occurs at a point where these two opposing factors are balanced. This interplay between the likelihood and the prior probability effectively implements Occam's razor by selecting the most simple model that best describes the data. The quality of this algorithm was demonstrated by testing it against several other popular bin selection techniques. The \texttt{optBINS} algorithm outperformed all techniques investigated in the sense that it consistently resulted in the minimum integrated square error between the estimated density model and the true density function. This was true even in the case of Scott's technique, which is designed to asymptotically minimize this error for Gaussian-distributed data. Our algorithm also can be readily applied to multi-dimensional data sets, which we demonstrated with a two-dimensional data set. In practice, we have been applying \texttt{optBINS} to three-dimensional data sets with comparable results. It should be noted that we are working with a piecewise-constant model of the density function, and \emph{not} a histogram. The distinction is subtle, but important. Given the full posterior probability for the model parameters and a selected number of bins, one can estimate the mean bin probabilities and their associated standard deviations. This is extremely useful in that it quantifies uncertainties in the density model, which can be used in subsequent calculations. In this paper, we demonstrated that with small numbers of data points the magnitude of the error bars on the bin heights is on the order of the bin heights themselves. Such a situation indicates that too few data exist to infer a density function. This can also be determined by examining the marginal posterior probability for the number of bins. In cases where there are too few data points, the posterior will not possess a well-defined mode. In our experiments with Gaussian-distributed data, we found that approximately 150 data points are needed to accurately estimate the density model when the functional form of the density is unknown. This approach has an additional advantage in that we can use the \texttt{optBINS} algorithm to identify data sets where the data have been excessively truncated. This will be explored further in future papers \citep{Knuth:2006}. As always, there is room for improvement. The \texttt{optBINS} algorithm, as currently implemented, performs a brute force search of the number of bins. This can be slow for large data sets that require large numbers of bins, or multi-dimensional data sets that have multiple bin dimensions. We have also made some simplifying assumptions in designing the algorithm. First, the endpoints of the density model are defined by the data, and are not allowed to vary during the analysis. Second, we use the marginal posterior to select the optimal number of bins and then use this value to estimate the mean bin heights. This neglects the fact that we actually possess uncertainty about the number of bins. Thus the uncertainty in the number of bins is not quantified. Last, equally-spaced bins can be very inefficient in describing multi-modal density functions (as in Fig. \ref{fig:01}G.) In such cases, variable bin-width models such as \texttt{Bayesian Blocks} \citep{Scargle:2005} or the Markov chain Monte Carlo implementations that we have experimented with may prove to be more useful. For most applications, \texttt{optBINS} efficiently delivers an appropriate number of bins that both maximizes the depiction of the shape of the density function while minimizing the appearance of random fluctuations. \section*{Appendix 1: Matlab code for the optimal number of uniform bins in a histogram} \begin{verbatim} function optM = optBINS(data,minM,maxM) if size(data)>2 | size(data,1)>1 error('data dimensions must be (1,N)'); end N = size(data,2); logp = zeros(1,maxM); for M = minM:maxM n = hist(data,M); part1 = N*log(M) + gammaln(M/2) - gammaln(N+M/2); part2 = - M*gammaln(1/2) + sum(gammaln(n+0.5)); logp(M) = part1 + part2; end [maximum, optM] = max(logp); return \end{verbatim}

https://arxiv.org/abs/2008.01542

On extremal eigenvalues of the graph Laplacian

Upper and lower estimates of eigenvalues of the Laplacian on a metric graph have been established in 2017 by G. Berkolaiko, J.B. Kennedy, P. Kurasov and D. Mugnolo. Both these estimates can be achieved at the same time only by highly degenerate eigenvalues which we call maximally degenerate. By comparison with the maximal eigenvalue multiplicity proved by I. Kac and V. Pivovarchik in 2011 we characterize the family of graphs exhibiting maximally degenerate eigenvalues which we call lasso trees, namely graphs constructed from trees by attaching lasso graphs to some of the vertices.

\section*{Overview} We are interested in the study of two bounds of the eigenvalues of the graph Laplacian proven in \cite{BKKM17} which depend on a few simple geometrical and topological properties of the graph, namely the total length $\mathcal{L}$, the number of Dirichlet and Neumann pendant vertices $\mathcal{D}$ and $\mathcal{N}$, and the first Betti number $\beta$. \begin{align} \label{lwb} \lambda_n \geq {{m}}_n &= \begin{cases} \frac{\pi^2}{\mathcal{L}^2} \frac{n^2}{4} &\text{if } n < \mathcal{N} + \beta, \\ \frac{\pi^2}{\mathcal{L}^2} \left( n - \frac{\mathcal{N} + \beta}{2}\right)^2 &\text{if } n \geq \mathcal{N} + \beta \end{cases} \qquad &n \geq 2 \\ \label{upb} \lambda_n \leq {{M}}_n &= \frac{\pi^2}{\mathcal{L}^2} \left( n - 2 + \mathcal{D} + \frac{\mathcal{N} + \beta}{2} + \beta \right)^2 &n \in \mathbb{N}, \end{align} If $\mathcal{D} \neq 0$ then the lower bound estimate holds for all $n \in \mathbb{N}$. Otherwise if $\mathcal{D} = 0$ then $\lambda_1 = 0$. In \cite{KuSe18} it was shown that (\ref{upb}) is attained by an infinite sequence of eigenvalues $\{\mu_{n_i} = {{M}}_{n_i} \}_{i \in \mathbb{N}}$ generated by a family of graphs with $\mathcal{D} = \mathcal{N} = 0$ and any $\beta \geq 2$, or $\beta = \mathcal{D} + \mathcal{N} = 1$, called respectively Windmill graphs, Neumann lasso graph and Dirichlet lasso graph. It can be observed that these graphs also provide examples of sequences of eigenvalues which exhibit the equality in (\ref{lwb}), $\{\nu_{n_j} = {{m}}_{n_j} \}_{j \in \mathbb{N}}$. Remarkably, the sequences $\mu_{n_i}$ and $\nu_{n_j}$ coincide. This is possible because the indices ${n_i}$ and ${n_j}$ are respectively the smallest and the largest of a sequence of degenerate eigenvalues with multiplicity $m$: \begin{equation}\label{eq:sharpdegenerate} i=j \iff \mu_{n_j} = \nu_{n_i} \iff n_j - n_i = m-1. \end{equation} We call {\bf lower sharp} and {\bf upper sharp} eigenvalues those which satisfy the equality in (\ref{lwb}) and (\ref{upb}) respectively, and in general we call {\bf (degenerate) sharp eigenvalues} the eigenvalues of multiplicity $m \geq 1$ ($m \geq 2$) with smallest index upper sharp and largest index lower sharp like those in equation (\ref{eq:sharpdegenerate}). The purpose of this work is to investigate which graphs exhibit sharp eigenvalues and discuss their properties. The text is organized into three sections: Introduction and notation, Properties of sharp eigenvalues, and Main results. We can summarize our findings as follows. In Proposition~\ref{prp:sharpmult}, by direct comparison of (\ref{lwb}) and (\ref{upb}), we obtain an upper bound for the maximal eigenvalue multiplicity $m_\mathcal{U} = m_\mathcal{U}(\mathcal{G}) = \mathcal{D} + \mathcal{N} + 2 \beta - 1$. Eigenvalues of multiplicity $m_\mathcal{U}$ are called {\bf maximally degenerate eigenvalues}. In Theorem~\ref{thm:characterization} we show that sharp eigenvalues are characterized by being maximally degenerate. This is used to show that sharp eigenvalues are preserved when multiple graphs are joined together at one of their Dirichlet pendant vertices (Lemma~\ref{lem:joinDirichlet}) or when a loop graph---with certain prescribed length---is attached to any Neumann pendant vertex (Lemma~\ref{lem:attachloop}). In the proof of the main result, Theorem~\ref{thm:main}, we show how the aforementioned Lemmata can be used to construct a graph with arbitrary $\mathcal{N}, \mathcal{D}, \beta$ which produce sequences of degenerate sharp eigenvalues. Graphs which can be constructed by recursive applications of Lemmata~\ref{lem:joinDirichlet} and~\ref{lem:attachloop} are trees where some of the pendant vertices have a loop graph attached, or, equivalently, trees decorated with some lasso graphs (also called tadpole or lollipop graphs), for this reason we call them {\bf lasso trees}. By comparing $m_\mathcal{U}$ with the maximal eigenvalue multiplicity $m_\mathcal{M}$ proved in \cite{KaPi11} we observe that $m_\mathcal{U}(\mathcal{G}) \geq m_\mathcal{M}(\mathcal{G})$ with the equality occurring if and only if $\mathcal{G}$ is a lasso tree. Finally we conclude in Theorem~\ref{thm:comparison} that sharp eigenvalues appear only in the spectra of lasso trees. \section{Introduction and notation} In this section we present the essentials about metric graphs used in the present text, for a more general introduction we refer to \cite{BeKu13} and \cite{Ku20}, see also \cite{KuSt02} and the recent \cite{Mu19}. \paragraph{Metric graphs.} Metric graphs are constructed as the quotient space of a set of distinct real intervals under an equivalence relation on the set of their endpoints. Let ${E} = \sqcup_{i} e_i,\, e_i := [x_{2i-1},x_{2i}]$ be the disjoint union of closed real intervals and let $\sim$ be an equivalence relation over the endpoints of ${E}$. The quotient space $\mathcal{G} = {E} / \sim$ is a metric graph, whose set of edges and vertices are ${E} = {E}(\mathcal{G})$ and, respectively, ${V} = {V}(\mathcal{G}) = \sqcup_{i} \{ x_{2i-1},x_{2i} \} \,/ \sim$. The metric and measure over $\mathcal{G}$ are inherited from the Euclidean metric and Lebesgue measure over the edges. Moreover we denote by $\mathcal{L} = \sum_{i} |x_{2i} - x_{2i-1}|$ the total length of $\mathcal{G}$. Since we are interested in metric graphs which are compact and with finite total length, we assume that $\mathcal{G}$ has a finite number of edges, $|{E}| < \infty$, each of them compact. We allow the presence of loops and multiple edges. \paragraph{Cycles and the first Betti number.} A cycle is a finite sequence of distinct edges $\{ e^{(j)} \}_{j=1}^n$ associated to a sequence of distinct vertices $\{ v^{(j)} \}_{j=1}^{n}$ such that \begin{itemize} \item $e^{(j)}$ is incident to $v^{(j)}$ and $v^{(j+1)}$ for $j=1,\dots,n-1$, \item $e^{(n)}$ is incident to $v^{(n)}$ and $v^{(1)}$. \end{itemize} A cycle of length $1$ is also called loop. A graph with just one edge which is also a loop is called loop graph. We denote by $\beta = \beta_1 = |{E}| - |{V}| + 1$, the first Betti number of $\mathcal{G}$; $\beta$ coincides with the circuit rank of $\mathcal{G}$, namely the least number of edges that need to be removed in order to turn $\mathcal{G}$ into a tree. \paragraph{Functions on metric graphs.} Let $\textit{L}_2(\mathcal{G}) := \bigoplus_{i} \textit{L}_2 [x_{2i-1},x_{2i}]$. The space $\textit{L}_2(\mathcal{G})$ equipped with the inner product $\langle f , g \rangle := \sum_{i} \int_{e_i} f \overline{g}\, dx$ is a well defined Hilbert space. If $f \in \textit{L}_2(\mathcal{G})$ is continuously differentiable over the edge $e_i = [x_{2i-1},x_{2i}]$, the oriented derivatives of $f$ at the endpoints of the interval $e_i$ are defined by $\partial f (x_{2i-1}) = f'(x_{2i-1})$ and $\partial f (x_{2i}) = -f'(x_{2i})$. The oriented derivatives are well defined for functions in the Sobolev space $H^2(\mathcal{G}) := \bigoplus_{i} H^2 [x_{2i-1},x_{2i}]$. \paragraph{Laplacian, vertex conditions, and eigenvalues.} Given a subset of the pendant vertices $D$, which we call Dirichlet vertices, we define the Laplacian operator $L = L(\mathcal{G},D) := -\frac{d^2}{dx^2}$ with domain $\mathfrak{D}(L) = \mathfrak{D}(\mathcal{G},D)$ as the set of functions $f \in H^2(\mathcal{G})$ subject to the following conditions: \begin{itemize} \item $f$ is continuous at the vertices, so $f \in \mathcal{C}(\mathcal{G})$ (continuity condition), \item the oriented derivatives of $f$ sum to zero at each non Dirichlet vertex: \\ $\sum_{x_j \in v} \partial f(x_j) = 0 ~~ \forall v \in {V} \setminus D$ (Kirchhoff condition), \item $f$ vanishes at the Dirichlet vertices, $f(v) = 0\,\forall v \in D$ (Dirichlet condition). \end{itemize} The continuity and Kirchhoff conditions together are called standard vertex conditions (in the literature sometimes called natural). Standard vertex conditions at pendant vertices $v \notin D$ read as Neumann conditions: $f'(v) = 0$; hence we call these vertices Neumann and denote their set by $N$. We denote by the corresponding calligraphic letter the cardinality of the two different type of pendant vertices $\mathcal{N} = |N|, \mathcal{D} = |D|$. Standard vertex conditions at a vertex $v$ of degree two read as continuity of both the function and its derivative. Let $\mathcal{G}$ have a pair of edges $e_i, e_j$ incident to a vertex $v$ of degree two and let $\mathcal{G}'$ be the graph where $e_i, e_j$ are replaced by a single edge with length equal to the sum of the lengths of $e_i, e_j$. Then $\mathcal{G}',\textit{L}_2(\mathcal{G}')$, and $\mathfrak{D}(\mathcal{G}',D)$ are, respectively, isomorphic to $\mathcal{G},\textit{L}_2(\mathcal{G})$, and $\mathfrak{D}(\mathcal{G},D)$. Hence, degree two vertices play no role in the study and can be freely removed whenever they occur in the construction of graphs. Under the above hypothesis $L$ is a self-adjoint unbounded positive operator whose spectrum is exclusively discrete with unique accumulation point at $+\infty$, \cite{BeKu13}. We denote the spectrum by $\sigma(L) = \{ \lambda_n \}_{n \in \mathbb{N}}$. Whenever we say that an indexed eigenvalue $\lambda_n$ has multiplicity $m$ we assume $n$ to be the smallest index of the degenerate eigenvalue, i.e. $\lambda_n = \dots = \lambda_{n+m-1}$. Moreover, unless differently stated, we shall assume $\mathcal{G}$ to be connected. Under this hypothesis follows that the ground state $\lambda_1$ is simple (see \cite{Ku19}) and $\lambda_1 = 0$ if and only if $D = \emptyset$. In \cite{BKKM17} the authors show that if $\mathcal{G}$ is not a loop graph then the eigenvalues of $L(\mathcal{G},D)$ satisfy the inequalities (\ref{lwb}), (\ref{upb}). \paragraph{Quantum Graphs.} The term quantum graph is used in general to refer to the triple $\Gamma = (\mathcal{G},-\frac{d^2}{dx^2} + q(x),\mathfrak{D}(\mathcal{G},D))$. In the present setup $q \equiv 0$, so a metric graph $\mathcal{G}$ together with a set of Dirichlet vertices $D$ suffice to determine a quantum graph $\Gamma$. Hence we use the generic term graph to refer to both metric and quantum graph whenever the set of Dirichlet vertices and the associated Laplacian are clear from the context. In particular we may speak of the eigenvalues of a graph meaning the eigenvalues of the associated Laplacian. In the study of the spectral estimates of quantum graphs several techniques have been developed, many of which put in relation modifications of the metric graph with the changes occurring in the spectrum. Several of these are discussed in \cite{BKKM19} under the name of \emph{surgery principles}. Lemmata \ref{lem:joinDirichlet} and \ref{lem:attachloop} make use of a particular case of two such principles. In line with the previous paragraph, any modification brought to a graph $\mathcal{G}$ and its set of Dirichlet vertices $D$ should be reflected in the associated Laplacian. \section{Properties of sharp eigenvalues} \subsection{Sharp and maximally degenerate eigenvalues} We start by making a general observation. \begin{observation}\label{obs:strictineq} The sequences $\{{{m}}_n\}_{n \in \mathbb{N}}$ and $\{{{M}}_n\}_{n \in \mathbb{N}}$ given by (\ref{lwb}) and (\ref{upb}) are strictly increasing. As a consequence of this, if $\lambda_n = {{m}}_n$ then $\lambda_n < {{m}}_{n+1} \leq \lambda_{n+1}$; hence \begin{itemize} \item if $\lambda_n$ is lower sharp then $\lambda_n < \lambda_{n+1}$. \end{itemize} \noindent Similarly $\lambda_n = {{M}}_n$ implies that \begin{itemize} \item if $\lambda_n$ is upper sharp then $\lambda_{n-1} < \lambda_n$. \end{itemize} \end{observation} \begin{proposition}\label{prp:sharpmult} The multiplicity of any eigenvalue of the graph Laplacian is at most $\mathcal{D} + \mathcal{N} + 2\beta - 1$. Eigenvalues with maximal multiplicity are called maximally degenerate eigenvalues. \end{proposition} \begin{proof} Consider $\lambda_n$ be a degenerate eigenvalue of multiplicity $m \geq 2$, so $n \geq 2$ by the simplicity of the ground state. By direct application of the inequalities (\ref{lwb},\ref{upb}) to $\lambda_n$ and $\lambda_{n+m-1}$ we have \begin{equation}\label{eq:ineq_lambda_mu} {{m}}_{n+m-1} \leq {{M}}_n. \end{equation} Assuming $n+m-1 \geq \mathcal{N} + \beta$, then (\ref{eq:ineq_lambda_mu}) implies $m \leq \mathcal{D} + \mathcal{N} + 2\beta - 1$. If instead we assume $n+m-1 < \mathcal{N} + \beta$ then (\ref{eq:ineq_lambda_mu}) implies $m \leq n + 2\mathcal{D} + \mathcal{N} + 3\beta - 3$ which combined with the assumption leads to $m < \mathcal{D} + \mathcal{N} + 2\beta - 1$. \end{proof} \begin{lemma}\label{lem:sharpmult} Let $\lambda_n$ be a degenerate eigenvalue of multiplicity $m \geq 2$. Then $\lambda_n$ is a sharp degenerate eigenvalue if and only if $\lambda_n$ is maximally degenerate $m = \mathcal{D} + \mathcal{N} + 2\beta - 1$. \end{lemma} \begin{proof} Because of the simplicity of the ground state, $n \geq 2$. Assume $n + m - 1 \geq \mathcal{N} + \beta$; hence by definition ${{m}}_{n + m - 1} = (\pi / \mathcal{L})^2 (n + m - 1 - (\mathcal{N} + \beta)/2)^2$ and \begin{align}\label{eq:lem_step1} {{M}}_n = {{m}}_{n + m - 1} &\iff n - 2 + \mathcal{D} + \frac{\mathcal{N} + \beta}{2} + \beta = n + m - 1 - \frac{\mathcal{N} + \beta}{2} \\ \nonumber &\iff m = \mathcal{D} + \mathcal{N} + 2\beta - 1.\\ \intertext{Assume instead $n+m-1 < \mathcal{N} + \beta$; hence ${{m}}_{n + m - 1} = (\pi / \mathcal{L})^2 (n + m - 1)^2 / 4$; we show that this is not compatible with the hypothesis. In fact} \label{eq:lem_step2} {{M}}_n = {{m}}_{n + m - 1} &\iff n - 2 + \mathcal{D} + \frac{\mathcal{N} + \beta}{2} + \beta = \frac{n + m - 1}{2} \\ &\iff m = n + 2\mathcal{D} + \mathcal{N} + 3\beta - 3, \nonumber \end{align} therefore $n+m-1 < \mathcal{N} + \beta$ reads $n + \mathcal{D} + \beta < 2$ which implies $n = 1$, a contradiction. \end{proof} The proof of Lemma~\ref{lem:sharpmult} suggests there might exist simple eigenvalues which are both upper sharp and lower sharp at the same time, which shall be called simple sharp eigenvalues. Consider first $n \geq 2$: if ${{M}}_n = {{m}}_n$ then either \begin{enumerate}[(i)] \item \label{case:1} $n < \mathcal{N} + \beta$, so by (\ref{eq:lem_step2}) $n=1$, which is excluded, \item \label{case:2} $n \geq \mathcal{N} + \beta$, so by (\ref{eq:lem_step1}) $\mathcal{D} + \mathcal{N} + 2\beta = 2$. \end{enumerate} \noindent If we consider $n = 1$ then either \begin{enumerate}[(i)] \setcounter{enumi}{2} \item \label{case:3} $\mathcal{D} = 0$ thus $\lambda_1 = 0 = {{M}}_1$, so $\mathcal{N} + 3\beta = 2$, \item \label{case:4} $\mathcal{D} \neq 0$ and $1 < \mathcal{N} + \beta$, so by (\ref{eq:lem_step2}) $2\mathcal{D} + \mathcal{N} + 3\beta = 3$, \item \label{case:5} $\mathcal{D} \neq 0$ and $1 \geq \mathcal{N} + \beta$, so by (\ref{eq:lem_step1}) $\mathcal{D} + \mathcal{N} + 2\beta = 2$. \end{enumerate} Case (\ref{case:2}) is satisfied by any of the following: \begin{itemize} \item $\beta = 1$ and $\mathcal{N} = \mathcal{D} = 0$, i.e. the loop graph which should be disregarded as it is an exceptional case for which neither (\ref{lwb}) nor (\ref{upb}) holds. \item $\beta = 0$ and $\mathcal{N} + \mathcal{D} = 2$, i.e. the interval with any admissible vertex conditions at its endpoints. \end{itemize} Case (\ref{case:3}) implies $\mathcal{N}=2, \beta = 0$, namely the Neumann-Neumann interval. Case (\ref{case:4}) does not have solutions. Case (\ref{case:5}) implies $\beta = 0$ and either $\mathcal{N} = 0, \mathcal{D} = 2$, or $\mathcal{N} = \mathcal{D} = 1$ hence the remaining two possible vertex conditions for the single interval. Therefore the single interval with any of the admissible vertex conditions provides the only three examples of graphs with simple sharp eigenvalues, and in particular the whole spectrum is composed only by simple sharp eigenvalues: \begin{itemize} \item $\mathcal{N}=2,\mathcal{D}=0$, the spectrum is $\lambda_n^{NN} = \frac{\pi^2}{\mathcal{L}^2}(n-1)^2$, \item $\mathcal{N}=1,\mathcal{D}=1$, the spectrum is $\lambda_n^{ND} = \frac{\pi^2}{\mathcal{L}^2}(n-\frac{1}{2})^2$, \item $\mathcal{N}=0,\mathcal{D}=2$, the spectrum is $\lambda_n^{DD} = \frac{\pi^2}{\mathcal{L}^2}n^2$. \end{itemize} \begin{proposition}\label{prp:sharpint} An eigenvalue is simple sharp if and only if the underlying graph is a single interval with any of the three possible combinations of vertex conditions listed above. \end{proposition} From now on we refer to as sharp eigenvalues the eigenvalues which are either simple or degenerate eigenvalues which are sharp. The above discussion can be summarized by the following statement: \begin{theorem}\label{thm:characterization} Let $\mathcal{G}$ be a graph which is not a cycle and let $\lambda$ be an eigenvalue of some Laplacian over $\mathcal{G}$. Then $\lambda$ is sharp if and only if it is maximally degenerate. \end{theorem} \subsection{Sharp eigenvalues and fully supported eigenspace} In this section we show a necessary property of the eigenfunctions associated to degenerate sharp eigenvalues. For its proof we need the following proposition about the regularity of the eigenvalues seen as functions dependent on the length of an edge of the graph. \begin{proposition}\label{prp:continuouslambda} For any fixed index $n \in \mathbb{N}$ and edge $e$ of length $\ell$, the function $\ell \mapsto \lambda_n(\ell)$ is continuous on $(0,+\infty)$. \end{proposition} \begin{proof} Consider the Courant-Fischer eigenvalues characterizations via the Rayleigh quotient \begin{equation}\label{eq:rayleigh} \lambda_n = \min_{\substack{X \subset \mathfrak{D}_q \\ \dim(X) = n}} \max_{u \in X} \frac{\norm{u'}_2^2}{\norm{u}_2^2} = \norm{\psi_n'}_2^2, \end{equation} where $\mathfrak{D}_q = \mathfrak{D}_q(\mathcal{G},D) := \{f \in \textit{H}^{\,1}(\mathcal{G}) \cap \mathcal{C}(\mathcal{G}) : f(v) = 0 \,\,\forall v \in D \}$ and $\psi_n$ is any normalized eigenfunction associated to $\lambda_n$. In order to prove the statement we show that $\rho \mapsto \lambda_n(\rho\ell)$ is continuous in $\rho = 1$. Let $\mathcal{G}_\rho$ be the modification of $\mathcal{G}$ where the edge $e$ is stretched by a factor $\rho$, i.e. $e$ is replaced by $\rho e$ and consequently $\ell$ replaced by $\rho \ell$. Let $X \subset \mathfrak{D}_q$ be any subset realizing the minimum in (\ref{eq:rayleigh}). Let $X_\rho \subset \mathfrak{D}_q(\mathcal{G}_\rho)$ be the space obtained from $X$ by stretching each function over the edge $e$, i.e. $f_\rho(x) = f(x)$ if $x \in \mathcal{G}_\rho \setminus \rho e$ and $f_\rho(x) = f(x/\rho)$ if $x \in \rho e$. From the Rayleigh quotient it follows that \begin{equation}\label{eq:rayleighbound2} \lambda_n(\rho \ell) \leq \max_{u_\rho \in X_\rho} \frac{\norm{u_\rho'}_2^2}{\norm{u_\rho}_2^2}. \end{equation} We compute \begin{equation} \begin{aligned} \norm{u_\rho'}_{L_2(\mathcal{G}_\rho)}^2 &= \int_{\mathcal{G}_\rho \setminus \rho e} (u_\rho')^2\,dx + \int_{\rho e} (u_\rho')^2\,dx \\ &=\int_{\mathcal{G} \setminus e} (u')^2\,dx + \int_{e} \left( \frac{1}{\rho} u' \right)^2 \rho \,dx \\ &=\int_{\mathcal{G}} (u')^2\,dx + \left( \frac{1}{\rho} - 1 \right) \int_{e} \left( u' \right)^2 \,dx \end{aligned} \end{equation} and similarly we also obtain \begin{eqnarray} \norm{u_\rho}_{L_2(\mathcal{G}_\rho)}^2 =\int_{\mathcal{G}} u^2\,dx + \left( \rho - 1 \right) \int_{e} u^2 \,dx. \end{eqnarray} Therefore, if $\rho \leq 1$ we have the following upper estimate \begin{equation} \begin{aligned} \max_{u_\rho \in X_\rho} \frac{\norm{{u_\rho}'}_2^2}{\norm{u_\rho}_2^2} &= \max_{u \in X} \frac{\norm{u'}_{L_2(\mathcal{G})}^2 + (\frac{1}{\rho} - 1)\norm{u'}_{L_2(e)}^2 }{\norm{u}_{L_2(\mathcal{G})}^2 + (\rho-1) \norm{u}_{L_2(e)}^2} \\ &\leq \frac{1}{\rho^2} \max_{u \in X} \frac{\norm{u'}_{L_2(\mathcal{G})}^2}{\norm{u}_{L_2(\mathcal{G})}^2} \\ &\leq \frac{1}{\rho^2} \lambda_n(\ell). \end{aligned} \end{equation} Moreover, from the monotonicity of the eigenvalues (see for example Corollary 3.12 in \cite{BKKM19}) we know that $\rho \leq 1 \Rightarrow \lambda_n(\ell) \leq \lambda_n(\rho \ell)$. Thus for $\rho \leq 1$ we have \begin{equation} \lambda_n(\ell) \leq \lambda_n(\rho \ell) \leq \frac{1}{\rho^2} \lambda_n(\ell). \end{equation} By changing $\ell$ with $\ell / \rho$ then we can deduce the more general inequality for any $\rho > 0$: \begin{equation} \min \left\{ 1, \rho^{-2} \right\} \lambda_n(\ell) \leq \lambda_n(\rho \ell) \leq \max \left\{ 1, \rho^{-2} \right\} \lambda_n(\ell), \end{equation} which shows the continuity in $\rho = 1$ of $\rho \mapsto \lambda_n(\rho\ell)$ and hence the claimed continuity of $\ell \mapsto \lambda_n(\ell)$ for $\ell \in (0,+\infty)$. \end{proof} The next lemma shows that both upper and lower sharp eigenvalues can be associated to eigenfunctions that do not identically vanish on any edge of the graph. \begin{lemma}\label{lem:prop} If $\lambda_n$ is either lower or upper sharp then for each edge $e$ there exists an eigenfunction associated to $\lambda_n$ which is not identically zero on $e$. \end{lemma} \begin{proof} Assume $\lambda_n = {{M}}_n$ and let $\ell(e) = \ell_0$. Making $e$ longer increases the total length of the graph and consequently decreases ${{M}}_n$, we write ${{M}}_n(\ell)$ to highlight the estimate dependence on the length of the edge $e$. In order not to violate (\ref{upb}), $\lambda_n$ must also decrease, at least as much as its upper bound. Assume $\lambda_n$ has multiplicity $m$; hence by Observation~\ref{obs:strictineq} $\lambda_{n-1} < \lambda_n = \dots = \lambda_{n+m-1} < \lambda_{n+m}$. By Proposition~\ref{prp:continuouslambda} all eigenvalues are continuous functions in the length $\ell(e)$, so there exists $\varepsilon > 0$ small such that $\forall \ell \in [\ell_0,\ell_0 + \varepsilon]$ \begin{equation}\label{eq:cond_and} \lambda_{n-1}(\ell) < \lambda_n(\ell) \quad \textit{and} \quad \lambda_{n+m-1}(\ell) < \lambda_{n+m}(\ell). \end{equation} Let $\{ \psi_j\}_{j=n}^{n+m-1}$ be a basis of the $m$-dimensional eigenspace associated to $\lambda_n$. If each $\psi_j$ is identically zero on $e$, then $\psi_j$ is still an eigenfunction after perturbing the length of $e$ over the interval $[\ell_0,\ell_0 + \varepsilon]$ and by the Rayleigh quotient it is associated to an eigenvalue equal to $\lambda_n(\ell_0)$ with the same multiplicity $m$. Because of (\ref{eq:cond_and}) the indices of the eigenvalues are preserved; hence \begin{equation} \lambda_j(\ell) \equiv \lambda_j(\ell_0) \quad \forall \ell \in [\ell_0,\ell_0 + \varepsilon]. \end{equation} This leads to the following contradiction \begin{equation} \begin{aligned} {{M}}_n(\ell_0 + \varepsilon) < {{M}}_n(\ell_0) &= \lambda_n(\ell_0) \\ &= \lambda_n(\ell_0 + \varepsilon) \leq {{M}}_n(\ell_0 + \varepsilon). \end{aligned} \end{equation} Thus there exists an eigenfunction not identically zero on $e$. \end{proof} We then have the next corollary. \begin{corollary}\label{cor:nonzeroeigf} If $\lambda$ is a sharp eigenvalue then there exists an eigenfunction associated to $\lambda$ which does not identically vanish on any edge of the graph. \end{corollary} \section{Main results} The main theorem is proven in a constructive manner and relies upon the next two lemmata, each of them providing an operation which preserves sharp eigenvalues. The first of them, Lemma~\ref{lem:joinDirichlet}, tells us that joining together graphs which share a sharp eigenvalue preserves not only the eigenvalue but also its sharpness. Let $\{\mathcal{G}_i\}_{i=1}^p$ be a finite set of graphs, each of them with $\mathcal{N}_i$ Neumann, $\mathcal{D}_i \neq 0$ Dirichlet pendant vertices and $\beta_i$ first Betti numbers respectively. For each $\mathcal{G}_i$ fix a Dirichlet vertex $v_i \in D_i$ (see for example the set of graphs on the left of figure \ref{fig:joinDirichlet}). Assume that the spectrum of each $\mathcal{G}_i$ contains the same eigenvalue $\lambda$, not necessarily with the same index $\lambda_{n_i}(\mathcal{G}_i) = \lambda \,\forall i$. Consider the graph $\mathcal{G}$ obtained by the disjoint union of all graphs $\bigsqcup_{i=1}^p \mathcal{G}_i$ with the vertices $v_i$ replaced by a single vertex endowed with standard vertex conditions as in Figure~\ref{fig:joinDirichlet}. Then $\lambda$ is still an eigenvalue of $\mathcal{G}$. We have then the following statement. \begin{lemma}\label{lem:joinDirichlet} If $\lambda_{n_i} = \lambda$ is a sharp eigenvalue of each $\mathcal{G}_i$, then $\lambda_n = \lambda, n = 2 - p + \sum_{i=1}^p n_i$ is also a sharp eigenvalue of $\mathcal{G}$. \end{lemma} \begin{figure}[ht] \centering \includegraphics[width=.8\textwidth]{figure1.pdf} \caption{Example of application of Lemma~\ref{lem:joinDirichlet} to three graphs (left) joined at one chosen Dirichlet vertex for each of them in order to obtain the graph on the right. Here and in the following figures the symbol $\circ$ stands for a Dirichlet pendant vertex and $\bullet$ for a Neumann pendant vertex.} \label{fig:joinDirichlet} \end{figure} The above lemma allows us to build trees with sharp eigenvalues: one starts by joining intervals into star graphs with at least one Dirichlet pendant and then the star graphs into a tree. This Lemma allows the construction of three graphs with any prescribed number of Dirichlet and Neumann pendant vertices which exhibit sharp eigenvalues. It remains to show that it is as well possible to prescribe the first Betti number and still be able to construct a graph with sharp eigenvalues. This is achieved by Lemma~\ref{lem:attachloop} which shows that sharp eigenvalues are preserved after attaching a cycle to a Neumann pendant. We have already mentioned that the loop graph is the only graph which does not satisfy the inequalities (\ref{lwb}~\ref{upb}), in particular we can notice the following: \begin{proposition} The spectrum of the loop graph $L_\mathcal{L}$ of length $\mathcal{L}$ is given by $\sigma(L_\mathcal{L}) = \{\lambda_1 = 0 \} \cup \left\{ \lambda_{2j} = \lambda_{2j+1} = \frac{\pi^2}{\mathcal{L}^2}(2j)^2 : j \in \mathbb{N}_{>0} \right\}$. The even eigenvalues of the loop graph exceeds the upper estimate (\ref{upb}) by a term $+\frac{1}{2}$ as follows: \begin{equation} \lambda_{2j}(L_\mathcal{L}) = \frac{\pi^2}{\mathcal{L}^2} \left( 2j - 2 + \frac{3}{2} + \frac{1}{2} \right)^2. \end{equation} The odd eigenvalues, excluded the first, differs from the lower estimate (\ref{lwb}) by a term $-\frac{1}{2}$ as follows: \begin{equation} \lambda_{2j+1}(L_\mathcal{L}) = \frac{\pi^2}{\mathcal{L}^2} \left( 2j + 1 - \frac{1}{2} - \frac{1}{2} \right)^2. \end{equation} \end{proposition} Therefore, given $\lambda > 0$ and $j \in \mathbb{N}$, the loop graph with length $\ell := 2 j \pi / \sqrt{\lambda}$ has the eigenvalue $\lambda_{2j} = \lambda$ with multiplicity $2$. Now consider $\mathcal{G}$ any graph with at least one Neumann pendant vertex $v$ with a certain eigenvalue $\lambda_n(\mathcal{G}) = \lambda$. Let $\mathcal{G}_v$ be the graph obtained by attaching the loop graph of length $\ell$ to the Neumann vertex $v$ with standard vertex conditions imposed there as in Figure~\ref{fig:attachloop}. We have the following statement: \begin{lemma}\label{lem:attachloop} If $\lambda_n = \lambda$ is a sharp eigenvalue of $\mathcal{G}$ then $\lambda_{n_v} = \lambda, n_v = n+2j-1$ is a sharp degenerate eigenvalue of $\mathcal{G}_v$. In particular, the multiplicity of $\lambda$ going from $\mathcal{G}$ to $\mathcal{G}_v$ increases by one. \end{lemma} \begin{figure}[ht] \centering \includegraphics[width=.8\textwidth]{figure2.pdf} \caption{Example of the construction considered in Lemma~\ref{lem:attachloop}. On the left the graph $\mathcal{G}$ and the loop graph, on the right the graph $\mathcal{G}_v$.} \label{fig:attachloop} \end{figure} Lemmata~\ref{lem:joinDirichlet} and~\ref{lem:attachloop} applied to a set of intervals and loop graphs provide the tools to derive the main result, Figure~\ref{fig:thm_main} shows an example of graph constructed following the proof of Theorem~\ref{thm:main}. \begin{theorem}\label{thm:main} Given $\mathcal{N},\mathcal{D}, \beta \in \mathbb{N} \cup \{0\}$ such that $\mathcal{N} + \mathcal{D} + \beta \geq 2$, there exists a graph with $\mathcal{N}$ Neumann, $\mathcal{D}$ Dirichlet pendant vertices respectively and first Betti number $\beta$ which exhibits an infinite sequence of sharp eigenvalues. \end{theorem} \begin{figure}[ht] \centering \includegraphics[width=.4\textwidth]{figure3.pdf} \caption{Example of graph constructed via Theorem~\ref{thm:main} with $\mathcal{D} = 4, \mathcal{N} = 2, \beta = 2$ with the lengths of the edges to scale.} \label{fig:thm_main} \end{figure} In \cite{KaPi11} the authors show that the maximal multiplicity of eigenvalues of the Schr\"odinger operator $-\frac{d^2}{dx^2} + q(x)$ with potential $q \in \textit{L}_1$ defined on a compact graph $\mathcal{G}$ is $m_\mathcal{M} = \mathcal{\beta} + \mathcal{P}^T - 1$, where $\mathcal{P}^T$ is the number of pendant vertices of the tree graph $T_\mathcal{G}$ obtained from $\mathcal{G}$ after contracting each cycle to a vertex. We observe that the contraction of any cycle may generate at most one new pendant vertex, thus $\mathcal{P}^T - (\mathcal{D} + \mathcal{N}) \leq \beta$. This means that $m_\mathcal{M} \leq m_\mathcal{U}$ with the equality occurring if and only if $T_\mathcal{G}$ has exactly $\beta$ pendant vertices more than $\mathcal{G}$, or equivalently $\mathcal{G}$ is a lasso tree. \begin{definition} A lasso tree is a compact metric graph where each cycle is a loop incident to a vertex of degree three. \end{definition} The previous observation together with Theorem \ref{thm:characterization} lead us to the following conclusion. \begin{figure}[ht] \centering \includegraphics[width=.5\textwidth]{figure4.pdf} \caption{Example of a lasso tree.} \label{fig:lasso_tree} \end{figure} \begin{theorem}\label{thm:comparison} If $\mathcal{G}$ is a metric graph with sharp eigenvalues, then $\mathcal{G}$ is a lasso tree. \end{theorem} \begin{remark} We point out that Lemmata \ref{lem:joinDirichlet} and \ref{lem:attachloop} can be applied recursively to construct lasso trees, with any possible topological structure, having sharp eigenvalues. \end{remark} \subsection{Proofs} \begin{proof}[Proof of Lemma~\ref{lem:joinDirichlet}] By Theorem~\ref{thm:characterization} each $\lambda_{n_i}$ is maximally degenerate; hence with multiplicity $m_i = \mathcal{D}_i + \mathcal{N}_i + 2\beta_i - 1$. The spectrum of the disjoint union of the graphs $\{\mathcal{G}_i\}$ is the disjoint union of their eigenvalues, therefore $\lambda$ is an eigenvalue of $\bigsqcup_{i=1}^p \mathcal{G}_i$ with multiplicity $\sum m_i$. Since $\lambda_{n_i - 1} < \lambda_{n_i}$ then there are $(\sum n_i) - p$ strictly smaller eigenvalues than $\lambda$, possibly zero. Hence the smallest index of $\lambda$ on $\bigsqcup_{i=1}^p \mathcal{G}_i$ is $1 + \sum (n_i - 1)$. Replacing the vertices $\{v_i\}$ by a single vertex $v$ endowed with standard vertex conditions is an operation which increases the dimension of the domain of the quadratic form associated to the Laplacian by one, thus it corresponds to a rank one perturbation of the operator which pushes all the eigenvalues down, but no further than one index, i.e. it interlaces the eigenvalues $\lambda_{j-1}( \bigsqcup_{i=1}^p \mathcal{G}_i) \leq \lambda_j( \mathcal{G}) \leq \lambda_{j}( \bigsqcup_{i=1}^p \mathcal{G}_i)\, \forall j \in \mathbb{N}$. This operation can be seen as the inverse of a particular case of Theorem 3.4 (2) in \cite{BKKM19}, see also Theorem 3.1.8 in \cite{BeKu13}. Since $\lambda$ has multiplicity $\sum m_i$ on $\bigsqcup_{i=1}^p \mathcal{G}_i$ after the change of vertex condition, $\lambda$ is still an eigenvalue on $\mathcal{G}$, with multiplicity at least $m = (\sum m_i) - 1$ and correspondingly with lowest index at most $n = 2 + \sum (n_i - 1) $. We shall now show that $n$ and $m$ are indeed exact. Notice that when going from $\bigsqcup_{i=1}^p \mathcal{G}_i$ to $\mathcal{G}$ we have that \begin{itemize} \item the number of Dirichlet pendant vertices is reduced by $p$,\\ $\mathcal{D} = (\sum_{i=1}^p \mathcal{D}_i) - p$; \item the number of Neumann pendant vertices is preserved,\\ $\mathcal{N} = \sum_{i=1}^p \mathcal{N}_i$; \item the first Betti number is preserved,\\ $\beta = \sum_{i=1}^p \beta_i$; \end{itemize} Therefore we compute that \begin{equation} \begin{aligned} m = \left( \sum m_i \right) - 1 &= \sum_{i=1}^n \left( \mathcal{D}_i + \mathcal{N}_i + 2\beta_i - 1 \right) - 1\\ &= \mathcal{D} + \mathcal{N} + 2\beta - 1 \end{aligned} \end{equation} which coincides with the maximal admissible multiplicity. Hence $\lambda$ must have precisely multiplicity $m$ and consequently lowest index $n$. By Theorem~\ref{thm:characterization} $\lambda_n$ must be a sharp eigenvalue. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:attachloop}.] Consider the spectrum of the disjoint union of $\mathcal{G}$ and $L_\ell$, which is the disjoint union of their spectra. Then $\lambda$ has now smallest index $n_v = (n-1) + (2j - 1) + 1$ and if $\lambda$ has multiplicity $m$ on $\mathcal{G}$ then its multiplicity on $\mathcal{G} \sqcup L_\ell$ is $m + 2$. The action of attaching the loop graph to $\mathcal{G}$ at the vertex $v \in \mathcal{G}$ is a rank one perturbation of the graph Laplacian which decreases the domain of its associated quadratic form and consequently pushes the eigenvalues up, but no further than the eigenvalue of next index; hence \begin{equation} \lambda = \lambda_{n + 2j - 1}(\mathcal{G} \sqcup L_\ell) \leq \lambda_{n + 2j - 1}(\mathcal{G}_v) \leq \lambda_{n+2j}(\mathcal{G} \sqcup L_\ell) = \lambda. \end{equation} We now show that after this operation the multiplicity of $\lambda$ is reduced by one, i.e. it is $m + 1$, and consequently the smallest index of $\lambda$ on $\mathcal{G}_v$ is still $n_v$. Let us parameterize the loop graph by the interval $[-\ell / 2,\ell / 2]$ with the zero placed in $v$. Any eigenfunction $\varphi$ on $\mathcal{G}$ can be extended to $\mathcal{G}_v$ by \begin{equation} \widetilde{\varphi}(x) := \begin{cases} \varphi(x) &\text{if }x \in \mathcal{G}, \\ \varphi(v)\cos(\sqrt{\lambda}x) &\text{if }x \in [-\ell / 2,+\ell / 2]. \end{cases} \end{equation} So all the eigenfunctions on $\mathcal{G}$ associated to $\lambda_n$ are embedded in $\mathcal{G}_v$. In addition the following eigenfunction $\widetilde{\varphi}$ from $L_\ell$ can be embedded in $\mathcal{G}_v$ \begin{equation} \widetilde{\varphi}(x) := \begin{cases} 0 &\text{if }x \in \mathcal{G}, \\ \sin(\sqrt{\lambda}x) &\text{if }x \in [-\ell / 2,+\ell / 2]. \end{cases} \end{equation} Hence going from $\mathcal{G} \sqcup L_\ell$ to $\mathcal{G}_v$ the multiplicity of $\lambda$ is reduced by one. Now notice that by Theorem~\ref{thm:characterization} the multiplicity of $\lambda$ on $\mathcal{G}$ is $m = \mathcal{D} + \mathcal{N} + 2 \beta - 1$, and hence \begin{equation} m + 1 = \mathcal{D} + (\mathcal{N}-1) + 2 (\beta+1) - 1. \end{equation} which is the maximal admissible eigenvalue multiplicity on $\mathcal{G}_v$ since this graph has one more cycle and one less Neumann pendant than $\mathcal{G}$. Again by Theorem~\ref{thm:characterization}, $\lambda$ must be a sharp degenerate eigenvalue of $\mathcal{G}_v$ with smallest index necessarily $n_v$. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:main}] Let $\mathcal{N}, \mathcal{D}$ and $\beta$ be given. In order to construct a graph with these corresponding numbers of Neumann pendants, Dirichlet pendants and first Betti number respectively, it is enough to consider \begin{itemize} \item $\mathcal{N} + \beta$ copies of Neumann-Dirichlet intervals $I^{ND}$ of length $\ell_N = \pi / 2$, \item $\mathcal{D}$ copies of Dirichlet-Dirichlet intervals $I^{DD}$ of length $\ell_D = \pi$, \item $\beta$ copies of loop graph $L$ of length $\ell_L = 2\pi$. \end{itemize} Notice that the above three graphs share the following sequence of eigenvalues: \begin{equation} \lambda^{ND}_{j} = \lambda^{DD}_{2j-1} = \lambda^{L}_{2(2j-1)} = (2j - 1)^2. \end{equation} Apply Lemma~\ref{lem:joinDirichlet} to all the above intervals, both Neumann-Dirichlet and Dirichlet-Dirichlet to deduce that $\{ \lambda_{n_j} = ( 2j + 1 )^2 \}_{j \in \mathbb{N}}$ is a sequence of sharp eigenvalues, each of multiplicity $\mathcal{N} + \beta + \mathcal{D} - 1$, where \begin{equation} \begin{aligned} n_j &= 2 - ((\mathcal{N} + \beta) + \mathcal{D}) + (\mathcal{N} + \beta) \cdot j + \mathcal{D} \cdot (2j - 1) \\ &= 2 - (\mathcal{N} + \beta) + (\mathcal{N} + \beta + 2 \mathcal{D}) j. \end{aligned} \end{equation} Notice that the length of the loop graph $L$ can be rewritten as \begin{equation} \ell_L = 2 \pi = \frac{\pi \cdot 2(2j - 1)}{\sqrt{\lambda^{L}_{2(2j-1)}}}. \end{equation} Therefore we can recursively apply Lemma~\ref{lem:attachloop} $\beta$ number of times and obtain the new sequence of sharp eigenvalues $\{ \lambda_{\widetilde{n}_j} = (2j - 1)^2 \}$, each of multiplicity $\mathcal{N} + \mathcal{D} + 2 \beta - 1$, which is maximal, where \begin{equation} \begin{aligned} \widetilde{n}_j &= n_j + 2(2j - 1)\beta - \beta \\ &= 2 - (\mathcal{N} + 4\beta) + (\mathcal{N} + 4\beta + 2 \mathcal{D}) j. \end{aligned} \end{equation} \end{proof} \begin{observation} The proof of Theorem~\ref{thm:main} with $\mathcal{N} = \mathcal{D} = 0$ recovers the family of Windmill graphs defined \cite{KuSe18}. \end{observation} \section*{Acknowledgement} The author wishes to thank Pavel Kurasov for valuable suggestions and guiding, and Jacob Muller.

https://arxiv.org/abs/1903.06737

Shortest paths in arbitrary plane domains

Let $\Omega$ be a connected open set in the plane and $\gamma: [0,1] \to \overline{\Omega}$ a path such that $\gamma((0,1)) \subset \Omega$. We show that the path $\gamma$ can be ``pulled tight'' to a unique shortest path which is homotopic to $\gamma$, via a homotopy $h$ with endpoints fixed whose intermediate paths $h_t$, for $t \in [0,1)$, satisfy $h_t((0,1)) \subset \Omega$. We prove this result even in the case when there is no path of finite Euclidean length homotopic to $\gamma$ under such a homotopy. For this purpose, we offer three other natural, equivalent notions of a ``shortest'' path. This work generalizes previous results for simply connected domains with simple closed curve boundaries.

\section{Introduction} \label{sec:introduction} Bourgin and Renz \cite{BourginRenz1989} proved that given a simply connected plane domain $\Omega$ with simple closed curve boundary, and given any two points $p,q \in \overline{\Omega}$, there exists a unique \emph{shortest} path in $\overline{\Omega}$ which is the uniform limit of paths which (except possibly for their endpoints $p$ and $q$) are contained in $\Omega$. If there is a rectifiable such curve (i.e.\ one with finite Euclidean length), then by shortest is meant the one with the smallest Euclidean length. If not, then by shortest is meant \emph{locally shortest}, which means that every subpath not containing the endpoints $p$ and $q$ is of finite Euclidean length and is shortest among all such paths joining the endpoints of the subpath. A related result for $1$-dimensional locally connected continua is proved in \cite{CannonConnerZastrow2002}. In \cref{thm:main1}, we extend the result of Bourgin and Renz to paths in multiply connected domains with arbitrary boundaries. Along the way, we characterize the concept of a shortest path in up to four different ways, outlined in the subsections below and stated in \cref{thm:main2}. To state these concepts and our results precisely, we must fix some terminology and notation regarding paths and homotopies. For notational convenience, we identify the plane $\mathbb{R}^2$ with the complex numbers $\mathbb{C}$. Fix a connected open set $\Omega \subset \mathbb{C}$ and points $p,q \in \overline{\Omega}$. We consider paths whose range is contained in $\Omega$, except that the endpoints of the path may belong to $\partial \Omega$. For brevity, we use the abbreviation ``e.p.e.''\ to mean ``except possibly at endpoints''. Formally, a \emph{path in $\Omega$ (e.p.e.)\ joining $p$ and $q$} is a continuous function $\gamma: [0,1] \to \mathbb{C}$ such that $\gamma(0) = p$, $\gamma(1) = q$, and $\gamma(s) \in \Omega$ for all $s \in (0,1)$. When $p,q \in \partial \Omega$, the existence of such a path is equivalent to the statement that $p$ and $q$ are \emph{accessible} from $\Omega$. We consider homotopies $h: [0,1] \times [0,1] \to \mathbb{C}$ such that for each $t \in [0,1)$, the path $h_t: [0,1] \to \mathbb{C}$ defined by $h_t(s) = h(s,t)$ is a path in $\Omega$ (e.p.e.)\ joining $p$ and $q$. Given a path $\gamma$ in $\Omega$ (e.p.e.)\ joining $p$ and $q$, let $\overline{[\gamma]}$ denote the set of all paths homotopic to $\gamma$ under such homotopies. Note that paths in $\overline{[\gamma]}$ may meet the boundary $\partial \Omega$ in more than just the endpoints. Let $[\gamma]$ denote the set of all paths in $\Omega$ (e.p.e.)\ which belong to $\overline{[\gamma]}$. Note that in spite of the notation, $\overline{[\gamma]}$ is not equal to the topological closure of $[\gamma]$ (e.g.\ in the function space). The main result of this paper is that if $\gamma$ is a path in $\Omega$ (e.p.e.)\ joining $p$ and $q$, then $\overline{[\gamma]}$ contains a unique \emph{shortest} path. By ``shortest'', we mean any one of the equivalent notions introduced next and collected in \cref{thm:main2} below. First, if there is a path in $\overline{[\gamma]}$ of finite Euclidean length, then we may consider the path in $\overline{[\gamma]}$ of smallest Euclidean length. Second, we adapt Bourgin \& Renz's condition of locally shortest as follows: \begin{defn*} \label{defn:locally shortest} Let $\gamma$ be a path in $\Omega$ (e.p.e.)\ joining $p$ and $q$, and let $\lambda \in \overline{[\gamma]}$. \begin{itemize} \item Given $0 < s_1 < s_2 < 1$, we define the class $\overline{[\lambda^\gamma_{[s_1,s_2]}]}$ as follows. Let $h: [0,1] \times [0,1] \to \overline{\Omega}$ be a homotopy such that $h_0 = \gamma$, $h_1 = \lambda$, and for each $t \in [0,1)$, $h_t$ is a path in $\Omega$ (e.p.e.)\ joining $p$ and $q$. We use the homotopy $h$ to ``pull off'' the path $\lambda {\upharpoonright}_{[s_1,s_2]}$ (e.p.e.)\ from $\partial \Omega$ as follows. Define the path $\lambda^\gamma_{[s_1,s_2]}: [s_1,s_2] \to \overline{\Omega}$ by \[ \lambda^\gamma_{[s_1,s_2]}(s) = h(s, 1 - (s-s_1)(s-s_2)) .\] See \cref{fig:locally shortest}. Though this path is defined in terms of the homotopy $h$, the class $\overline{[\lambda^\gamma_{[s_1,s_2]}]}$ does not depend on the choice of $h$ (see \cref{cor:locally shortest well-defined} below). \item A path $\lambda \in \overline{[\gamma]}$ is called \emph{locally shortest} if for any $0 < s_1 < s_2 < 1$, the path $\lambda {\upharpoonright}_{[s_1,s_2]}$ has finite Euclidean length, and this length is smallest among all paths in $\overline{[\lambda^\gamma_{[s_1,s_2]}]}$. \end{itemize} \end{defn*} \begin{figure} \begin{center} \includegraphics{LocallyShortest.pdf} \end{center} \caption{An illustration of the paths $\gamma$, $\lambda$, and $\lambda^\gamma_{[s_1,s_2]}$, for the definition of a locally shortest path.} \label{fig:locally shortest} \end{figure} Two further notions of ``shortest path'' are given in the next subsections. \subsection{Efficient paths} \label{sec:efficient} As above, fix a connected open set $\Omega \subset \mathbb{C}$ and points $p,q \in \overline{\Omega}$. \begin{defn*} \label{defn:efficient} Let $\gamma$ be a path in $\Omega$ (e.p.e.)\ joining $p$ and $q$. A path $\lambda \in \overline{[\gamma]}$ is called \emph{efficient} (in $\overline{[\gamma]}$) if the following property holds: \begin{quote} Given any $s_1,s_2 \in [0,1]$ with $s_1 < s_2$, let $\lambda'$ be the path defined by $\lambda'(s) = \lambda(s)$ for $s \notin [s_1,s_2]$, and $\lambda' {\upharpoonright}_{[s_1,s_2]}$ parameterizes the straight line segment $\overline{\lambda(s_1) \lambda(s_2)}$ or the constant path if $\lambda(s_1) = \lambda(s_2)$. If $\lambda' \in \overline{[\gamma]}$, then $\lambda = \lambda'$ (up to reparameterization). \end{quote} \end{defn*} The following is the first main result of this paper. The proof is in \cref{sec:proof main1}. \begin{thm} \label{thm:main1} Let $\Omega \subset \mathbb{C}$ be a connected open set, let $p,q \in \overline{\Omega}$, and let $\gamma$ be a path in $\Omega$ (except possibly at endpoints) joining $p$ and $q$. Then $\overline{[\gamma]}$ contains a unique (up to parameterization) efficient path. \end{thm} \subsection{Alternative notion of path length} \label{sec:len} Instead of the standard Euclidean path length, which can only distinguish between two paths if at least one of them is rectifiable, we can use an alternative notion of path length for which all paths have finite length, and which has other useful properties and many features in common with Euclidean length. A notion of length with such properties was first introduced in \cite{Morse1936} and further developed in \cite{Silverman1969}. A simlar notion is given in \cite{CannonConnerZastrow2002}, and the authors of the present paper modified and extended that notion in \cite{HOT2018-1}. Let $\mathsf{len}$ refer to either the path length function introduced in \cite{HOT2018-1} or in \cite{Morse1936}. The essential properties of $\mathsf{len}$ are: \begin{enumerate} \item $\mathsf{len}(\gamma)$ is finite (in fact $\mathsf{len}(\gamma) < 1$) for any path $\gamma$; \item $\mathsf{len}(\gamma) = 0$ if and only if $\gamma$ is constant; \item For any $p,q \in \mathbb{C}$, the straight line segment $\overline{pq}$ has smallest $\mathsf{len}$ length among all paths from $p$ to $q$. Moreover, if $\gamma$ is a path from $p$ to $q$ which deviates from the straight line segment $\overline{pq}$, or which is not monotone, then $\mathsf{len}(\gamma) > \mathsf{len}(\overline{pq})$; \item If $\Phi: \mathbb{C} \to \mathbb{C}$ is an isometry, then $\mathsf{len}(\Phi \circ \gamma) = \mathsf{len}(\gamma)$; \item Given $0 \leq c_1 < c_2 \leq 1$, $\mathsf{len}(\gamma {\upharpoonright}_{[c_1,c_2]}) \leq \mathsf{len}(\gamma)$. This inequality is strict unless $\gamma$ is constant outside of $[c_1,c_2]$; \item Given $c \in (0,1)$, $\mathsf{len}(\gamma) \leq \mathsf{len}(\gamma {\upharpoonright}_{[0,c]}) + \mathsf{len}(\gamma {\upharpoonright}_{[c,1]})$. This inequality is strict unless one of the subpaths is constant; \item $\mathsf{len}$ is a continuous function from the space of paths (with the uniform metric $d_{\mathrm{sup}}$) to $\mathbb{R}$. \end{enumerate} Our second main result is the following. The proof is in \cref{sec:proof main2}. \begin{thm} \label{thm:main2} Let $\Omega \subset \mathbb{C}$ be a connected open set, let $p,q \in \overline{\Omega}$, and let $\gamma$ be a path in $\Omega$ (except possibly at endpoints) joining $p$ and $q$. Then for a path $\lambda \in \overline{[\gamma]}$, the following are equivalent: \begin{enumerate} \item $\lambda$ is locally shortest; \item $\lambda$ is efficient; and \item $\lambda$ has smallest $\mathsf{len}$ length among all paths in $\overline{[\gamma]}$. \end{enumerate} Moreover, if $\overline{[\gamma]}$ contains a rectifiable path, then in addition to the above we have \begin{enumerate} \setcounter{enumi}{3} \item $\lambda$ has smallest Euclidean length among all paths in $\overline{[\gamma]}$. \end{enumerate} \end{thm} \subsection*{Ackowledgements} The authors would like to thank Professor Alexandre Eremenko for educating them on the history of analytic covering maps and the referee of an earlier version of this paper for several helpful comments which helped to clarify some of the arguments. \section{Preliminaries on bounded analytic covering maps} \label{sec:covering maps} Our arguments in Sections \ref{sec:proof main1} and \ref{sec:proof main2} make heavy use of the theory of complex analytic covering maps. In this section, we collect the results we will use later. Denote $\mathbb{D} = \{z \in \mathbb{C}: |z| < 1\}$. It is a standard classical result (see e.g.\ \cite{Ahlfors1973}) that for any connected open set $\Omega \subset \mathbb{C}$ whose complement contains at least two points, and for any $z_0 \in \Omega$, there is a complex analytic covering map $\varphi: \mathbb{D} \to \Omega$ such that $\varphi(0) = z_0$. Moreover, this covering map $\varphi$ is uniquely determined by the argument of $\varphi'(0)$. Many of the results below hold only for analytic covering maps $\varphi: \mathbb{D} \to \Omega$ to bounded sets $\Omega$. For the remainder of this subsection, let $\Omega \subset \mathbb{C}$ be a bounded connected open set, and let $\varphi: \mathbb{D} \to \Omega$ be an analytic covering map. We first state three classic results about the boundary behavior of $\varphi$. All of the subsequent results in this section will be derived from these and from standard covering map theory. \begin{thm}[Fatou \cite{Fatou06}, see e.g.\ {\cite[p.22]{Conway1995}}] \label{thm:Fatou} The radial limits $\lim_{r \to 1^-} \varphi(r\alpha)$ exist for all points $\alpha \in \partial \mathbb{D}$ except possibly for a set of linear measure zero in $\partial \mathbb{D}$. \end{thm} It can easily be seen that if $\alpha \in \partial \mathbb{D}$ is such that $\lim_{r \to 1^-} \varphi(r\alpha)$ exists, then the limit belongs to $\partial \Omega$. \begin{thm}[Riesz \cite{Riesz16, Riesz23}, see e.g.\ {\cite[p.22]{Conway1995}}] \label{thm:Riesz} For each $z \in \partial \Omega$, the set of points $\alpha \in \partial \mathbb{D}$ for which $\lim_{r\to 1^-} \varphi(r\alpha) = z$ has linear measure zero in $\partial \mathbb{D}$. \end{thm} \begin{thm}[Lindel\"{o}f \cite{Lindelof1915}, see e.g.\ {\cite[p.23]{Conway1995}}] \label{thm:Lindelof} Let $\widehat{\gamma}: [0,1] \to \overline{\mathbb{D}}$ be a path such that $\widehat{\gamma}([0,1)) \subset \mathbb{D}$ and $\widehat{\gamma}(1) = \alpha \in \partial \mathbb{D}$. Suppose that $\lim_{t \to 1^-} \varphi \circ \widehat{\gamma}(t)$ exists. Then the radial limit $\lim_{r\to 1^-} \varphi(r\alpha)$ exists and is equal to $\lim_{t \to 1^-} \varphi \circ \widehat{\gamma}(t)$. \end{thm} We define the \emph{extended covering map} \[ \overline{\varphi}: \mathbb{D} \cup \{\alpha \in \partial \mathbb{D}: \lim_{r \to 1^-} \varphi(r\alpha) \textrm{ exists}\} \to \overline{\Omega} \] by \[ \overline{\varphi}(\widehat{z}) = \begin{cases} \varphi(\widehat{z}) & \textrm{if } \widehat{z} \in \mathbb{D} \\ \lim_{r \to 1^-} \varphi(r\alpha) & \textrm{if } \widehat{z} = \alpha \in \partial \mathbb{D} . \end{cases} \] By \cref{thm:Fatou}, this function is defined on $\mathbb{D}$ plus a full measure subset of $\partial \mathbb{D}$. Note that this function $\overline{\varphi}$ is not necessarily continuous at points where it is defined in $\partial \mathbb{D}$. It is, however, continuous by definition along each radial segment from the center $0$ of $\mathbb{D}$ to any point $\alpha \in \partial \mathbb{D}$ where it is defined. In fact, more is true: if $\overline{\varphi}$ is defined at $\alpha$, then its restriction to any \emph{Stolz angle} at $\alpha$ is continuous (see e.g.\ \cite[p.23]{Conway1995}); however, we will not need this concept in this paper. As a general convention, we will put hats on symbols, as in $\widehat{z}$, $\widehat{A}$, or $\widehat{\gamma}$, to refer to points, subsets, or paths in $\overline{\mathbb{D}}$, and use symbols without hats to refer to points, subsets, or paths in $\overline{\Omega} \subset \mathbb{C}$, with the understanding that if both $z$ and $\widehat{z}$ appear in an argument, they are related by $z = \overline{\varphi}(\widehat{z})$ (and likewise for sets and paths). If $z = \overline{\varphi}(\widehat{z})$ for points $z \in \overline{\Omega}$ and $\widehat{z} \in \overline{\mathbb{D}}$ (respectively, $A = \overline{\varphi}(\widehat{A})$ for sets $A \subset \overline{\Omega}$ and $\widehat{A} \subset \overline{\mathbb{D}}$), we say $\widehat{z}$ is a \emph{lift} of $z$ (respectively, $\widehat{A}$ is a \emph{lift} of $A$). Similarly, if $\gamma = \overline{\varphi} \circ \widehat{\gamma}$ for paths $\gamma$ in $\overline{\Omega}$ and $\widehat{\gamma}$ in $\overline{\mathbb{D}}$, we say $\widehat{\gamma}$ is a \emph{lift} of $\gamma$. The next result about lifts of paths is very similar to classical results for covering maps. Since our extended map $\overline{\varphi}$ is not necessarily continuous at points in $\partial \mathbb{D}$, it is not a simple consequence of basic covering map theory. It is proved in \cite{HOT2018-2}, and we also include a proof here to convey the flavor of working with the extended map $\overline{\varphi}$. \begin{thm} \label{thm:lift} Let $p,q \in \overline{\Omega}$, let $z \in \Omega$, and let $\widehat{z} \in \mathbb{D}$ such that $\varphi(\widehat{z}) = z$. \begin{enumerate} \item Let $\gamma$ be a path in $\Omega$ (e.p.e.)\ joining $p$ and $q$, and suppose $\gamma(s_0) = z$ for some $s_0 \in [0,1]$. Then there exists a unique path $\widehat{\gamma}$ in $\mathbb{D}$ (e.p.e.)\ such that $\overline{\varphi} \circ \widehat{\gamma} = \gamma$ and $\widehat{\gamma}(s_0) = \widehat{z}$. \item Let $h: [0,1] \times [0,1] \to \mathbb{C}$ be a homotopy such that for each $t \in [0,1]$ the path $h_t(s) = h(s,t)$ is a path in $\Omega$ (e.p.e.)\ joining $p$ and $q$, and suppose $h(s_0,0) = z$ for some $s_0 \in [0,1]$. Then there exists a unique homotopy $\widehat{h}: [0,1] \times [0,1] \to \overline{\mathbb{D}}$ such that $\widehat{h}(s,t) \in \mathbb{D}$ and $\varphi \circ \widehat{h}(s,t) = h(s,t)$ for each $(s,t) \in (0,1) \times [0,1]$, $\widehat{h}(\{0\} \times [0,1])$ is a single point $\widehat{p}$ in $\overline{\mathbb{D}}$ with $\overline{\varphi}(\widehat{p}) = p$, $\widehat{h}(\{1\} \times [0,1])$ is a single point $\widehat{q}$ in $\overline{\mathbb{D}}$ with $\overline{\varphi}(\widehat{q}) = q$, and $\widehat{h}(s_0,0) = \widehat{z}$. \end{enumerate} \end{thm} \begin{proof} Observe that (1) follows from (2), by using the constant homotopy $h(s,t) = \gamma(s)$ for all $t \in [0,1]$. Thus it suffices to prove (2). Since $\varphi$ is a covering map and $h((0,1) \times [0,1]) \subset \Omega$, it follows from standard covering space theory that there is a unique homotopy $\widehat{h}: (0,1) \times [0,1] \to \mathbb{D}$ such that $\varphi \circ \widehat{h}(s,t) = h(s,t)$ for each $(s,t) \in (0,1) \times [0,1]$, and $\widehat{h}(s_0,0) = \widehat{z}$. It remains to prove that there are points $\widehat{p},\widehat{q} \in \mathbb{D}$ such that defining $\widehat{h}(\{0\} \times [0,1]) = \{\widehat{p}\}$ and $\widehat{h}(\{1\} \times [0,1]) = \{\widehat{q}\}$ makes $\widehat{h}$ into a continuous function from $[0,1] \times [0,1]$ to $\overline{\mathbb{D}}$. This is immediate if $p$ and $q$ belong to $\Omega$, so we assume $p,q \in \partial \Omega$. Observe that the set $\widehat{h}((0,\frac{1}{2}] \times [0,1])$ compactifies on a continuum $K \subset \partial \mathbb{D}$. We need to prove that $K$ is a single point $\{\widehat{p}\}$. Suppose for a contradication that $K$ contains more than one point. Then there exists by \cref{thm:Fatou} a set $E$ of positive measure in the interior of $K$ so that for each $\alpha \in E$, the radial limit $\lim_{r \to 1^-} \varphi(r\alpha)$ exists. Since the set $\widehat{h}((0,\frac{1}{2}] \times [0,1])$ compactifies on $K$ we can choose, for each $\alpha \in E$, a sequence $(s_n,t_n)$ in $(0,\frac{1}{2}] \times [0,1]$ such that $s_n \to 0$ and $\widehat{h}(s_n,t_n) = r_n \alpha$, with $r_n \to 1$. It follows that the radial limit $\lim_{r \to 1^-} \varphi(r\alpha) = p$ for each $\alpha \in E$, a contradiction with \cref{thm:Riesz}. Thus $K$ is a single point $\{\widehat{p}\}$, and so we can continuously extend $\widehat{h}$ to $\{0\} \times [0,1]$ by defining $\widehat{h}(\{0\} \times [0,1]) = \{\widehat{p}\}$. By \cref{thm:Lindelof} it follows that $\overline{\varphi}(\widehat{p}) = p$. Likewise, by considering $\widehat{h}([\frac{1}{2},1) \times [0,1])$ we obtain by the same argument a point $\widehat{q} \in \overline{\mathbb{D}}$ such that $\widehat{h}$ extends continuously to $\{1\} \times [0,1]$ by defining $\widehat{h}(\{1\} \times [0,1]) = \{\widehat{q}\}$, and $\overline{\varphi}(\widehat{q}) = q$. \end{proof} \cref{thm:lift}(2) implies that if $p,q \in \overline{\Omega}$, $\gamma$ is a path in $\Omega$ (e.p.e.)\ joining $p$ and $q$, $\widehat{\gamma}$ is a lift of $\gamma$ with endpoints $\widehat{p},\widehat{q} \in \overline{\mathbb{D}}$, and $\lambda \in [\gamma]$, then there exists a lift $\widehat{\lambda}$ of $\lambda$ with the same endpoints $\widehat{p},\widehat{q}$ as $\widehat{\gamma}$. We will prove a stronger statement in \cref{thm:class closure lift} below. The next result follows immediately from \cref{thm:lift}(1). \begin{cor} \label{cor:line lift} Let $L$ be an open arc in $\Omega$ whose closure is an arc with distinct endpoints in $\partial \Omega$. Let $\widehat{L}$ be a component of $\varphi^{-1}(L)$. Then $\widehat{L}$ is an open arc in $\mathbb{D}$ whose closure in $\overline{\mathbb{D}}$ is an arc with distinct endpoints in $\partial \mathbb{D}$. \end{cor} We next prove a result about the existence of small ``crosscuts'' in $\mathbb{D}$ straddling any point $\alpha \in \partial \mathbb{D}$ for which $\overline{\varphi}(\alpha) \in \partial \Omega$ is not isolated in $\partial \Omega$. For one-to-one analytic maps $\varphi: \mathbb{D} \to \mathbb{C}$, this result is standard; see e.g.\ \cite{Pommerenke1992}. We were unable to find a reference for the case of a bounded analytic covering map $\varphi$, so we include a proof for completeness. \begin{thm} \label{thm:crosscut} Let $\alpha \in \partial \mathbb{D}$ be such that the radial limit $\lim_{r \to 1^-} \varphi(r\alpha)$ exists, and let $p \in \partial \Omega$ be the limit. \begin{enumerate} \item If $p$ is not isolated in $\partial \Omega$, then for any sufficiently small simple closed curve $S$ in $\mathbb{C}$ containing $p$ in its interior, there is a component $\widehat{S}$ of $\varphi^{-1}(S)$ whose closure is an arc separating $\alpha$ from the center $0$ of $\mathbb{D}$ in $\overline{\mathbb{D}}$. \item If $p$ is isolated in $\partial \Omega$, then for any sufficiently small simple closed curve $S$ in $\Omega$ containing $p$ in its interior, there is a component $\widehat{S}$ of $\varphi^{-1}(S)$ whose closure is a circle whose intersection with $\partial \mathbb{D}$ is $\{\alpha\}$. \end{enumerate} Moreover, in both cases, the diameter of $\widehat{S}$ can be made arbitrarily small by choosing $S$ sufficiently small. \end{thm} \begin{proof} Let $S$ be small enough so that $\varphi(0)$ is not in the closed topological disk bounded by $S$. Let $r_0 < 1$ be close enough to $1$ so that $\varphi(r\alpha)$ is in the interior of $S$ for all $r \in [r_0,1)$. The closed (in $\mathbb{D}$) set $\varphi^{-1}(S)$ must separate $r_0 \alpha$ from $0$ in $\mathbb{D}$, since otherwise there would be a path from $0$ to $r_0 \alpha$ which would project to a path in $\Omega$ from $\varphi(0)$ to $\varphi(r_0 \alpha)$ without intersecting $S$, a contradiction. Therefore, there must be a component $\widehat{S}$ of $\varphi^{-1}(S)$ which separates $r_0 \alpha$ from $0$ in $\mathbb{D}$ (see e.g.\ \cite[p.438 \S 57 III Theorem 1]{Kuratowski1968}). For (1), suppose that $p$ is not isolated in $\partial \Omega$. Assume first that $S \cap \partial \Omega \neq \emptyset$. Let $C$ be the component of $S \cap \Omega$ containing $\varphi(\widehat{S})$. This $C$ is a path in $\Omega$ (e.p.e.)\ joining two points (not necessarily distinct) $a,b \in \partial \Omega$ with $a \neq p \neq b$. By \cref{thm:lift} and the fact that $C$ is an open arc in $\Omega$, we have that each component of $\varphi^{-1}(C)$ is an arc in $\mathbb{D}$ joining points $\widehat{a},\widehat{b} \in \partial \mathbb{D}$ at which the radial limits exist and are equal to $a$ and $b$, respectively; in particular, we have $\widehat{a} \neq \alpha \neq \widehat{b}$. It follows that the closure of $\widehat{S}$ is an arc with endpoints distinct from $\alpha$, which separates $\alpha$ from $0$ in $\overline{\mathbb{D}}$, as desired. Now choose $\varepsilon_1 > 0$ small enough so that $\varphi(0)$ is not in the closed disk $\overline{B}(p,\varepsilon_1)$ and such that $\partial B(p,\varepsilon_1) \cap \partial \Omega \neq \emptyset$. Assume that $S \subset B(p,\varepsilon_1)$. If $S \cap \partial \Omega \neq \emptyset$, then we are done by the previous paragraph; hence, suppose that $S \cap \partial \Omega = \emptyset$. Choose $\varepsilon_0 > 0$ small enough so that $B(p,\varepsilon_0)$ is contained in the interior of $S$, and such that $\partial B(p,\varepsilon_0) \cap \partial \Omega \neq \emptyset$. By the previous paragraph, there are components $\widehat{S}_0$ and $\widehat{S}_1$ of $\varphi^{-1}(\partial B(p,\varepsilon_0))$ and $\varphi^{-1}(\partial B(p,\varepsilon_1))$, respectively, which are arcs separating $\alpha$ from $0$. As above, there is a component $\widehat{S}$ of $\varphi^{-1}(S)$ which separates a tail end of the radial segment at $\alpha$ from $0$ in $\mathbb{D}$, and which separates $\widehat{S}_0$ from $\widehat{S}_1$ in $\mathbb{D}$ (i.e.\ lies between $\widehat{S}_0$ and $\widehat{S}_1$). This implies the endpoints of the closure of $\widehat{S}$ are distinct, hence the closure of $\widehat{S}$ is an arc, as desired. For (2), let $S$ be small enough so that $p$ is the only point of $\partial \Omega$ in the closed topological disk bounded by $S$. As above, we obtain a component $\widehat{S}$ of $\varphi^{-1}(S)$ which separates a tail end of the radial segment at $\alpha$ from $0$ in $\mathbb{D}$. Since $\varphi$ is a covering map, this $\widehat{S}$ must be an open arc in $\mathbb{D}$ whose two ends both accumulate on continua $K_1$ and $K_2$ in $\partial \mathbb{D}$. We first argue that $K_1$ and $K_2$ are single points. Suppose $\beta_1,\beta_2 \in K_1$ with $\beta_1 \neq \beta_2$. Then by \cref{thm:Fatou} there exists $\beta \in K_1$ between $\beta_1$ and $\beta_2$ where the radial limit $\lim_{r \to 1^-} \varphi(r\beta)$ exists. However, this radial segment meets $\widehat{S}$ arbitrarily close to $\beta$, hence $\lim_{r \to 1^-} \varphi(r\beta) \notin \partial \Omega$, a contradiction. Thus $K_1$ is a single point $\{\widehat{a}\}$. Similarly, $K_2$ is a single point $\{\widehat{b}\}$. If $\widehat{a} \neq \alpha$, by \cref{thm:Fatou} and \cref{thm:Riesz} we can find $\beta \in \partial \mathbb{D}$ between $\widehat{a}$ and $\alpha$ such that the radial limit $\lim_{r \to 1^-} \varphi(r\beta)$ exists and is different from $p$. The path $\overline{\varphi}(r\beta)$, $0 \leq r \leq 1$, is homotopic to one which does not enter the interior of $S$. By \cref{thm:lift}, we can lift this homotopy, to obtain a path from $0$ to $\beta$ in $\mathbb{D}$ which does not meet $\widehat{S}$. But this is a contradiction since $\widehat{S}$ separates $\beta$ from $0$. Therefore $\widehat{a} = \alpha$, and likewise $\widehat{b} = \alpha$. Thus the closure of $\widehat{S}$ is a circle meeting $\partial \mathbb{D}$ at $\alpha$ only, as desired. For the moreover part, we argue as in the proof of \cref{thm:lift} that if these components $\widehat{S}$ did not converge to $0$ in diameter as the diameter of $S$ is shrunk towards $0$, then they would accumulate on a non-degenerate continuum $K \subset \partial \mathbb{D}$, and we would obtain a contradiction by \cref{thm:Fatou} and \cref{thm:Riesz}. The details are left to the reader. \end{proof} \begin{lem} \label{lem:large lifts} Let $S$ be a straight line or round circle in $\mathbb{C}$, and let $\varepsilon > 0$. Then there are only finitely many lifts of components of $S \cap \Omega$ with diameter at least $\varepsilon$. \end{lem} \begin{proof} Suppose the claim is false, so that there exists $\varepsilon > 0$ and infinitely many lifts $\widehat{S}_n$, $n = 1,2,\ldots$, of components of $S \cap \Omega$ such that the diameter of $\widehat{S}_n$ is at least $\varepsilon$ for each $n$. These lifts accumulate on a non-degenerate continuum $K \subset \overline{\mathbb{D}}$. If $K \cap \mathbb{D} \neq \emptyset$, then let $\widehat{z} \in K \cap \mathbb{D}$ and let $\widehat{V}$ be a neighborhood of $\widehat{z}$ which maps one-to-one under $\varphi$ to a small round disk $V \subset \Omega$. Then $\widehat{V}$ meets infinitely many of the lifts $\widehat{S}_n$. On the other hand, because $V$ is a round disk contained in $\Omega$ and $S$ is a round circle or straight line, it follows that $V$ can only meet one component of $S \cap \Omega$. This is a contradiction since $\varphi$ is one-to-one on $\widehat{V}$. Suppose then that $K \subset \partial \mathbb{D}$. Then there exists by \cref{thm:Fatou} a set $E$ of positive measure in the interior of $K$ so that for each $\alpha \in E$, the radial limit $\lim_{r \to 1^-} \varphi(r\alpha)$ exists. If there is a single component $S'$ of $S \cap \Omega$ such that $\varphi(\widehat{S}_n) = S'$ for infinitely many $n$, then it is clear that the radial limit of $\varphi$ at each $\alpha \in E$ must belong to $\overline{S'} \cap \partial \Omega$, which contains at most two points. But this contradicts \cref{thm:Riesz}. Therefore we may assume that the components $\varphi(\widehat{S}_n)$ are all distinct, which means their diameters must converge to $0$. By passing to a subsequence if necessary, we may assume that the components $\varphi(\widehat{S}_n)$ converge to a single point $a \in S \cap \partial \Omega$. Then it is clear that the radial limit of $\varphi$ at each $\alpha \in E$ must equal $a$, again a contradiction by \cref{thm:Riesz}. \end{proof} Recall that the function $\overline{\varphi}$ is not necessarily continuous at points $\alpha \in \partial \mathbb{D}$ where it is defined. However, the next result shows that the restriction of $\overline{\varphi}$ to the region in between two lifted paths with the same endpoints is continuous. Given a continuum $X$ in $\mathbb{C}$, the \emph{topological hull} of $X$, denoted $\mathrm{Hull}(X)$, is the smallest simply connected continuum in $\mathbb{C}$ containing $X$. Equivalently, $\mathrm{Hull}(X)$ is equal to $\mathbb{C} \smallsetminus U$, where $U$ is the unbounded component of $\mathbb{C} \smallsetminus X$. \begin{lem} \label{lem:hull continuous} Suppose $\widehat{\lambda}: [0,1] \to \overline{\mathbb{D}}$ is a path such that $\overline{\varphi} \circ \widehat{\lambda}$ is a path in $\overline{\Omega}$ (i.e.\ is continuous). Let $X$ be the union of $\widehat{\lambda}([0,1])$ with the two radial segments from the center $0$ of $\mathbb{D}$ to $\widehat{p} = \widehat{\lambda}(0)$ and to $\widehat{q} = \widehat{\lambda}(1)$, and let $\Delta = \mathrm{Hull}(X)$. Then $\Delta$ is simply connected and locally connected, and the restriction $\overline{\varphi} {\upharpoonright}_\Delta$ is continuous on $\Delta$. \end{lem} \begin{proof} Note that $\Delta$ is simply connected by definition, and it is straightforward to see that the topological hull of any locally connected continuum is locally connected. Since $\varphi$ is continuous and $\overline{\varphi} = \varphi$ in $\mathbb{D}$, it remains to prove that the restriction of $\overline{\varphi}$ to $\Delta$ is continuous at each point of $\Delta \cap \partial \mathbb{D}$. Let $\alpha \in \Delta \cap \partial \mathbb{D}$, and suppose for a contradiction that the restriction of $\overline{\varphi}$ to $\Delta$ is not continuous at $\alpha$. Then there exists $\varepsilon > 0$ and a sequence of points $\langle \widehat{w}_n \rangle_{n=1}^\infty$ in $\Delta$ such that $\widehat{w}_n \to \alpha$, but $|\overline{\varphi}(\alpha) - \varphi(\widehat{w}_n)| \geq \varepsilon$ for all $n$. Note that since $\Delta \cap \partial \mathbb{D} \subset \widehat{\lambda}([0,1])$ and $\overline{\varphi} \circ \widehat{\lambda}$ is continuous, we have that the restriction of $\overline{\varphi}$ to $\Delta \cap \partial \mathbb{D}$ is continuous. Hence, we may assume that $\widehat{w}_n \in \Delta \cap \mathbb{D}$ for each $n$. Moreover, since the restriction of $\overline{\varphi}$ to the radial segment from $0$ to $\alpha$ is continuous, we may also assume that $\widehat{w}_n$ does not belong to this segment for all $n$. For each $n$, there is a lift $\widehat{C}_n$ of a component of $\Omega \cap \partial B(\overline{\varphi}(\alpha), \varepsilon)$ which separates $\widehat{w}_n$ from $\alpha$ in $\overline{\mathbb{D}}$, where $B(\overline{\varphi}(\alpha), \varepsilon)$ is the open disk centered at $\overline{\varphi}(\alpha)$ of radius $\varepsilon$. By \cref{thm:crosscut}, this lift $\widehat{C}_n$ is either an arc with endpoints in $\partial \mathbb{D}$, or, in the case that $\overline{\varphi}(\alpha)$ is the only point of $\partial \Omega$ in the closed disk $\overline{B}(\overline{\varphi}(\alpha), \varepsilon)$, a circle with one point on $\partial \mathbb{D}$. By passing to a subsequence if necessary, we may assume that all of the $\widehat{C}_n$ are distinct, hence they are pairwise disjoint. According to \cref{lem:large lifts}, the diameters of the lifts $\widehat{C}_n$ converge to $0$ as $n \to \infty$. Let $X$ be as in the statement of the lemma, and for each $n$ choose a point $\widehat{z}_n \in X \cap C_n$. Then $\widehat{z}_n \to \alpha$. Suppose first that there is a subsequence $\widehat{z}_{n_k}$ of $\widehat{z}_n$ such that for each $k$, $\widehat{z}_{n_k}$ belongs to the radial segment from $0$ to $\widehat{p}$. It follows that $\alpha = \widehat{p}$. Since $\overline{\varphi}$ is continuous on this radial segment, we have $\varphi(\widehat{z}_{n_k}) \to \overline{\varphi}(\alpha)$ as $k \to \infty$. But $\varphi(\widehat{z}_n) \in \partial B(\overline{\varphi}(\alpha), \varepsilon)$ for each $n$, so this is a contradiction. Likewise, we encounter a contradiction if infinitely many of the points $\widehat{z}_n$ belong to the radial segment from $0$ to $\widehat{q}$. Thus we may assume that all of the points $\widehat{z}_n$ belong to the set $\widehat{\lambda}([0,1])$. Let $s_n \in [0,1]$ such that $\widehat{z}_n = \widehat{\lambda}(s_n)$. By passing to a subsequence if necessary, we may assume that $s_n \to s_\infty$, and $\widehat{\lambda}(s_\infty) = \alpha$. Since $\overline{\varphi} \circ \widehat{\lambda}$ is continuous, it follows that $\overline{\varphi}(\widehat{z}_n) \to \overline{\varphi}(\alpha)$ as $n \to \infty$. But again this is a contradiction because $\varphi(\widehat{z}_n) \in \partial B(\overline{\varphi}(\alpha), \varepsilon)$ for each $n$. \end{proof} In the next result, we characterize paths in $\overline{[\gamma]}$ in terms of lifts. \begin{thm} \label{thm:class closure lift} Let $\gamma$ be a path in $\Omega$ (e.p.e.)\ joining $p$ and $q$, and let $\widehat{\gamma}$ be a lift of $\gamma$ with endpoints $\widehat{p},\widehat{q} \in \overline{\mathbb{D}}$. If $\lambda \in \overline{[\gamma]}$, then there exists a lift $\widehat{\lambda}$ of $\lambda$ (to $\overline{\mathbb{D}}$) with the same endpoints $\widehat{p},\widehat{q}$. Conversely, if $\lambda: [0,1] \to \overline{\Omega}$ is a path joining $p$ and $q$ which has a lift $\widehat{\lambda}: [0,1] \to \overline{\mathbb{D}}$ with the same endpoints $\widehat{p},\widehat{q}$, then $\lambda \in \overline{[\gamma]}$. \end{thm} \begin{proof} Let $\lambda \in \overline{[\gamma]}$, and let $h$ be a homotopy such that $h_0 = \gamma$, $h_1 = \lambda$, and $h_t$ is a path in $\Omega$ (e.p.e.)\ joining $p$ and $q$ for each $t \in [0,1)$. For each $s \in [0,1]$, consider the path $t \mapsto h_t(s)$. Apply \cref{thm:lift} to obtain a lift $\widehat{h}_t(s)$ such that $\widehat{h}_0(s) = \widehat{\gamma}(s)$. Define $\widehat{\lambda}: [0,1] \to \overline{\mathbb{D}}$ by $\widehat{\lambda}(s) = \widehat{h}_1(s)$. By \cref{thm:Lindelof}, we have $\overline{\varphi} \circ \widehat{\lambda}(s) = \lambda(s)$ for all $s \in [0,1]$, and $\widehat{\lambda}(0) = \widehat{p}$ and $\widehat{\lambda}(1) = \widehat{q}$. It remains to prove that $\widehat{\lambda}$ is continuous. Let $s \in [0,1]$. If $\widehat{\lambda}(s) \in \mathbb{D}$, then $\widehat{\lambda}$ is continuous at $s$ by standard covering space theory. Suppose for a contradiction that $\widehat{\lambda}(s) \in \partial \mathbb{D}$ and $\widehat{\lambda}$ is not continuous at $s$. We proceed with an argument similar to the one given for \cref{thm:lift}. The sets $\widehat{h}([s-\frac{1}{n},s+\frac{1}{n}] \times [1-\frac{1}{n},1)$, $n=1,2,\ldots$, accumulate on a non-degenerate continuum $K \subset \partial \mathbb{D}$, which contains, by \cref{thm:Fatou}, a set $E$ of positive measure such that the radial limit $\lim_{r \to 1^-} \varphi(r\alpha)$ exists for each $\alpha \in E$. We can choose, for each $\alpha \in E$, a sequence $(s_n,t_n)$ converging to $(s,1)$ such that $\widehat{h}(s_n,t_n) = r_n \alpha$, with $r_n \to 1$. It follows that the radial limit $\lim_{r \to 1^-} \varphi(r\alpha)$ is equal to \[ \lim_{n \to \infty} \varphi(r_n \alpha) = \lim_{n \to \infty} \varphi \circ \widehat{h}(s_n,t_n) = \lim_{n \to \infty} h(s_n,t_n) = \lambda(s) ,\] a contradiction with \cref{thm:Riesz}. Therefore, $\widehat{\lambda}$ is a continuous lift of $\lambda$. Conversely, suppose $\lambda: [0,1] \to \overline{\Omega}$ is a path joining $p$ and $q$ which has a lift $\widehat{\lambda}: [0,1] \to \overline{\mathbb{D}}$ with the same endpoints $\widehat{p},\widehat{q}$ as $\widehat{\gamma}$. Let $\widehat{c}$ be the path in $\mathbb{D}$ (e.p.e.)\ such that $\widehat{c}(0) = \widehat{p}$, $\widehat{c}(\frac{1}{2}) = 0$, $\widehat{c}(1) = \widehat{q}$, and $\widehat{c}$ linearly parameterizes the straight segments in between these points. Let $c = \overline{\varphi} \circ \widehat{c}$. Let $X = \widehat{\lambda}([0,1]) \cup \widehat{c}([0,1])$, and let $\Delta = \mathrm{Hull}(X)$. Since $\Delta$ is simply connected, it follows that there is a homotopy $\widehat{h}$ between $\widehat{c}$ and $\widehat{\lambda}$ within $\Delta$, such that $\widehat{h}_0 = \widehat{c}$, $\widehat{h}_1 = \widehat{\lambda}$, and $\widehat{h}_t$ is a path in $\mathbb{D}$ (e.p.e.)\ joining $\widehat{p}$ and $\widehat{q}$ for all $t \in [0,1)$. Since $\overline{\varphi}$ is continuous on $\Delta$ by \cref{lem:hull continuous}, the composition $\overline{\varphi} \circ \widehat{h}$ is a homotopy between $c$ and $\lambda$ which establishes that $\lambda \in \overline{[c]}$. By the same reasoning, we can show that $\gamma \in [c]$. Therefore, $\lambda \in \overline{[\gamma]}$. \end{proof} We now have the machinery in place to conclude that the class $\overline{[\lambda^\gamma_{[s_1,s_2]}]}$ described in the definition of a locally shortest path is well-defined. \begin{cor} \label{cor:locally shortest well-defined} Let $\gamma$ be a path in $\Omega$ (e.p.e.)\ joining $p$ and $q$, and let $\widehat{\gamma}$ be a lift of $\gamma$ with endpoints $\widehat{p},\widehat{q} \in \overline{\mathbb{D}}$. Let $\lambda \in \overline{[\gamma]}$, and let $\widehat{\lambda}$ be a lift of $\lambda$ (to $\overline{\mathbb{D}}$) with the same endpoints $\widehat{p},\widehat{q}$. Let $0 < s_1 < s_2 < 1$. A path $\rho$ belongs to $\overline{[\lambda^\gamma_{[s_1,s_2]}]}$ if and only if there is a lift $\widehat{\rho}$ of $\rho$ (to $\overline{\mathbb{D}}$) with endpoints $\widehat{\lambda}(s_1),\widehat{\lambda}(s_2)$. In particular, the definition of the class $\overline{[\lambda^\gamma_{[s_1,s_2]}]}$ is independent of the choice of homotopy $h$ between $\gamma$ and $\lambda$. \end{cor} \section{Proof of \cref{thm:main1}} \label{sec:proof main1} Let $\Omega \subset \mathbb{C}$ be a connected open set, $p,q \in \overline{\Omega}$, and let $\gamma$ be a path in $\Omega$ (e.p.e.)\ joining $p$ and $q$. We may assume that either $p \neq q$, or $\gamma$ is a non-trivial loop, so that $\overline{[\gamma]}$ contains no constant path, since otherwise this is obviously the unique efficient path. Let $D_\Omega$ be a large round disk in $\mathbb{C}$ which contains the entire path $\gamma([0,1])$. We may assume, without loss of generality, that $\Omega$ is contained in $D_\Omega$, hence in particular is a bounded subset of $\mathbb{C}$. Indeed, any efficient path in $\overline{[\gamma]}$ with respect to $\Omega$ also belongs to $\overline{[\gamma]}$ with respect to $\Omega \cap D_\Omega$, and vice versa. The same goes for the other notions of shortest path used in this paper. Hence, we assume $\Omega \subset D_\Omega$ for the remainder of this paper. Let $\varphi: \mathbb{D} \to \Omega$ be an analytic covering map, and let $\overline{\varphi}$ denote the extension of $\varphi$ to those points in $\partial \mathbb{D}$ where the radial limit is defined, as in \cref{sec:covering maps}. Choose any lift $\widehat{\gamma}: [0,1] \to \overline{\mathbb{D}}$ of $\gamma$ under $\overline{\varphi}$ (see \cref{thm:lift}(1)). So $\widehat{\gamma}$ is a path in $\mathbb{D}$ (e.p.e.)\ and $\overline{\varphi} \circ \widehat{\gamma} = \gamma$. Let $\widehat{p} = \widehat{\gamma}(0) \in \overline{\mathbb{D}}$ and $\widehat{q} = \widehat{\gamma}(1) \in \overline{\mathbb{D}}$. Since $\overline{[\gamma]}$ does not contain a constant path, we have $\widehat{p} \neq \widehat{q}$ (cf.\ \cref{thm:class closure lift}). In \cref{sec:sequence of approximations} and \cref{sec:convergence} below, we will establish the existence of an efficient path in $\overline{[\gamma]}$ via a recursive construction in which we repeatedly replace subpaths by straight line segments, when doing so does not change the homotopy class. We begin with some preliminary results in \cref{sec:lifts of lines}. \subsection{Lifts of lines} \label{sec:lifts of lines} Throughout this paper, when we use the word \emph{line} we mean straight line in $\mathbb{C}$. In this subsection, we consider lines $L$ in $\mathbb{C}$ which intersect $\Omega$, and lifts of closures of components of $L \cap \Omega$ under $\overline{\varphi}$. By abuse of terminology, any such lift will be called a \emph{lift of $L$}. For a given line $L$, $L \cap \Omega$ has at most countably many components, and each of these components has at most countably many lifts, each of which is, by \cref{cor:line lift}, an arc in $\overline{\mathbb{D}}$ whose (distinct) endpoints are in $\partial \mathbb{D}$, and which is otherwise contained in $\mathbb{D}$. Observe that if $\widehat{L}_1$ and $\widehat{L}_2$ are distinct lifts of lines, then $\widehat{L}_1 \cap \widehat{L}_2$ contains at most one point; moreover, if $\widehat{L}_1 \cap \widehat{L}_2 = \{\widehat{z}\}$ for some $\widehat{z} \in \mathbb{D}$, then $\widehat{L}_1$ and $\widehat{L}_2$ cross transversally at $\widehat{z}$. Our construction later in this section of an efficient path is based on the following reformulation of the definition of an efficient path. \begin{prop} \label{prop:efficient characterization} Let $\lambda \in \overline{[\gamma]}$, and let $\widehat{\lambda}$ be a lift of $\lambda$ (under $\overline{\varphi}$) joining $\widehat{p}$ to $\widehat{q}$. Then $\lambda$ is an efficient path in $\overline{[\gamma]}$ if and only if for any lift $\widehat{L}$ of a line intersecting $\Omega$, the set $\widehat{\lambda}^{-1}(\widehat{L})$ is connected (possibly empty). \end{prop} We next make a detailed study of lifts of lines, focusing on whether they separate $\widehat{p}$ from $\widehat{q}$ in $\overline{\mathbb{D}}$ or not. \begin{defn*} Let $L$ be a line which intersects $\Omega$ and which does not contain $p$ or $q$, and let $\widehat{L}$ be a lift of $L$. \begin{itemize} \item We call $\widehat{L}$ a \emph{separating lift} if it separates $\widehat{p}$ from $\widehat{q}$ in $\overline{\mathbb{D}}$. The component of $\mathbb{D} \smallsetminus \widehat{L}$ whose closure contains $\widehat{p}$ is called the \emph{$\widehat{p}$-side} of $\widehat{L}$, and the component of $\mathbb{D} \smallsetminus \widehat{L}$ whose closure contains $\widehat{q}$ is called the \emph{$\widehat{q}$-side} of $\widehat{L}$. \item We call $\widehat{L}$ a \emph{non-separating lift} if it does not separate $\widehat{p}$ from $\widehat{q}$ in $\overline{\mathbb{D}}$. The component of $\mathbb{D} \smallsetminus \widehat{L}$ whose closure does not contain $\widehat{p},\widehat{q}$ is called the \emph{shadow} of $\widehat{L}$, denoted $\mathrm{Sh}(\widehat{L})$. \end{itemize} \end{defn*} Given two lines $L_1,L_2$ which intersect $\Omega$, the \emph{distance} between $L_1$ and $L_2$ is the Hausdorff distance between $L_1 \cap \overline{D_\Omega}$ and $L_2 \cap \overline{D_\Omega}$ (recall that $D_\Omega$ is a fixed large disk in $\mathbb{C}$ containing $\Omega$); that is, the infimum of all $\delta > 0$ such that each point of $L_1 \cap \overline{D_\Omega}$ is within $\delta$ of a point in $L_2 \cap \overline{D_\Omega}$, and vice versa. \begin{defn*} \begin{itemize} \item Let $\widehat{V} \subset \mathbb{D}$ be an open set which maps one-to-one under $\varphi$ to an open set $V \subset \Omega$, and let $L$ be a line which intersects $V$. Let $\varepsilon > 0$, and assume $\varepsilon$ is small enough so that every line $L'$ which is within distance $\varepsilon$ of $L$ must also intersect $V$. The family of all lifts $\widehat{L}'$, of such lines $L'$, which intersect $\widehat{V}$ will be called a \emph{basic open set of lifts of lines}, and denoted $\widehat{\mathcal{N}}(\widehat{L},\widehat{V},\varepsilon)$. \item Let $\widehat{z} \in \mathbb{D}$. We say $\widehat{z}$ is a \emph{stable point} if there exists a basic open set of lifts of lines $\widehat{\mathcal{N}}(\widehat{L},\widehat{V},\varepsilon)$, each element of which is non-separating and contains $\widehat{z}$ in its shadow. \end{itemize} \end{defn*} \begin{lem} \label{lem:two lifts} A point $\widehat{z} \in \mathbb{D}$ is a stable point if and only if there are two intersecting non-separating lifts $\widehat{L}_1,\widehat{L}_2$ of distinct lines $L_1,L_2$ such that $\widehat{z}$ is in the shadow of both $\widehat{L}_1$ and $\widehat{L}_2$. Furthermore, whenever $\widehat{L}_1,\widehat{L}_2$ are intersecting non-separating lifts of distinct lines, the point of intersection of $\widehat{L}_1$ and $\widehat{L}_2$ is also a stable point. \end{lem} \begin{proof} That any stable point has this property is immediate, since any basic open set of lifts of lines clearly contains pairs of distinct intersecting lifts. Conversely, suppose $\widehat{L}_1,\widehat{L}_2$ are intersecting non-separating lifts of distinct lines $L_1 \supset \varphi(\widehat{L}_1), L_2 \supset \varphi(\widehat{L}_2)$, and let $\widehat{z}$ be any point in the shadow of both $\widehat{L}_1$ and $\widehat{L}_2$. Let $\widehat{w}$ be the point of intersection of $\widehat{L}_1$ and $\widehat{L}_2$, and let $\widehat{V}$ be a neighborhood of $\widehat{w}$ which maps one-to-one under $\varphi$ to a round disk $V$ centered at $w = \varphi(\widehat{w})$. \begin{figure} \begin{center} \includegraphics{StableCharacterization.pdf} \end{center} \caption{The configuration of lines described in the proof of \cref{lem:two lifts}.} \label{fig:two lifts} \end{figure} Consider a line $L$ not containing $w$ which intersects $L_1 \cap V$ and $L_2 \cap V$, and such that the lift $\widehat{L}$ of $L$ which intersects $\widehat{V}$ does not contain any point in the intersection of the shadows of $\widehat{L}_1$ and of $\widehat{L}_2$. Let $\varepsilon > 0$ be small enough so that any line $L'$ within distance $\varepsilon$ of $L$ has these same properties. Let $\widehat{L}'$ be an arbitrary element of the basic open set of lifts of lines $\widehat{\mathcal{N}}(\widehat{L},\widehat{V},\varepsilon)$; that is, $\widehat{L}'$ is the lift of a line $L'$ within distance $\varepsilon$ of $L$ such that $\widehat{L} \cap \widehat{V} \neq \emptyset$. Since $\widehat{L}'$ intersects $\widehat{L}_1$ and $\widehat{L}_2$ inside $\widehat{V}$, it cannot cross them again, and so one endpoint of $\widehat{L}'$ is in the closure of $\mathrm{Sh}(\widehat{L}_1) \smallsetminus \mathrm{Sh}(\widehat{L}_2)$ and the other is in the closure of $\mathrm{Sh}(\widehat{L}_2) \smallsetminus \mathrm{Sh}(\widehat{L}_1)$ (see \cref{fig:two lifts}). It follows that $\widehat{L}'$ is a non-separating lift and $\mathrm{Sh}(\widehat{L}')$ contains $\mathrm{Sh}(\widehat{L}_1) \cap \mathrm{Sh}(\widehat{L}_2)$ (in particular, it contains the point $\widehat{z}$), as well as the point $\widehat{w}$. Thus $\widehat{z}$ and $\widehat{w}$ are both stable points. \end{proof} \begin{lem} \label{lem:stable generic} The set of stable points is a dense open subset of $\mathbb{D}$. Moreover, except for a countable set of lines, every line $L$ has the property that for each lift $\widehat{L}$ of $L$, the set of stable points in $\widehat{L}$ is a dense open subset of $\widehat{L}$. \end{lem} \begin{proof} It follows immediately from \cref{lem:two lifts} that the set of stable points is open in $\mathbb{D}$. Now let $\widehat{z} \in \mathbb{D}$, and let $\widehat{V}$ be a neighborhood of $\widehat{z}$. Suppose $\widehat{z}$ is not a stable point. By shrinking $\widehat{V}$, we may assume that $\widehat{V}$ maps one-to-one under $\varphi$ to a disk $V$ centered at $z = \varphi(\widehat{z})$. For a given $\theta \in \mathbb{R}$, let $L(z,\theta)$ denote the straight line through $z$ making angle $\theta$ with the positive real axis. Let $\widehat{L}(\widehat{z},\theta)$ be the lift containing $\widehat{z}$ of the component of $L(z,\theta) \cap \Omega$ containing $z$. Denote the endpoints of $\widehat{L}(\widehat{z},\theta)$ by $e^+(\widehat{z},\theta)$ and $e^-(\widehat{z},\theta)$, where $e^+(\widehat{z},\theta)$ corresponds to following the line $L(z,\theta)$ to $\partial \Omega$ in the direction $\theta$ from $z$, and $e^-(\widehat{z},\theta)$ corresponds to following the line $L(z,\theta)$ to $\partial \Omega$ in the direction $\theta + \pi$ from $z$. As $\theta$ increases, the line $L(z,\theta)$ revolves about the point $z$. Correspondingly, the lift $\widehat{L}(\widehat{z},\theta)$ ``revolves'' about $\widehat{z}$. The endpoints of $\widehat{L}(\widehat{z},\theta)$ move monotonically in the circle $\partial \mathbb{D}$, but not necessarily continuously, as $\theta$ increases. Since $\widehat{z}$ is not stable, by \cref{lem:two lifts} there can be at most one non-separating lift of a line which contains $\widehat{z}$. It follows that if we consider the line $L(z,\theta)$ and increase $\theta$ to revolve the line about $z$, at some moment the endpoint $e^+(\widehat{z},\theta)$ must cross or ``jump over'' $\widehat{p}$, and at that same moment $e^-(\widehat{z},\theta)$ must cross or ``jump over'' $\widehat{q}$. That is, there exists $\theta_0$ such that for all $\alpha,\beta$ sufficiently close to $\theta_0$ with $\alpha < \theta_0 < \beta$, we have that $\widehat{p}$ is on the ``left'' side of the arc $\widehat{L}(\widehat{z},\alpha)$ (thinking of this arc as oriented from $e^-(\widehat{z},\alpha)$ to $e^+(\widehat{z},\alpha)$) and $\widehat{q}$ is on the ``right'' side, and $\widehat{p}$ is on the ``right'' side of $\widehat{L}(\widehat{z},\beta)$ and $\widehat{q}$ is on the ``left'' side (see \cref{fig:stable dense}). Let $\widehat{w}$ be any point in $\widehat{V} \smallsetminus \widehat{L}(\widehat{z},\theta_0)$ and let $w = \varphi(\widehat{w})$. Let $\alpha,\beta$ be as above and sufficiently close to $\theta_0$ so that the lines $L(w,\alpha)$ and $L(w,\beta)$ do not intersect either of the lines $L(z,\alpha)$ or $L(z,\beta)$ inside $\overline{D_\Omega}$ (recall that $D_\Omega$ is a fixed large disk in $\mathbb{C}$ containing $\Omega$). This means that the lifts $\widehat{L}(\widehat{w},\alpha)$ and $\widehat{L}(\widehat{w},\beta)$ do not cross either of the lifts $\widehat{L}(\widehat{z},\alpha)$ or $\widehat{L}(\widehat{z},\beta)$. Because of the locations of $\widehat{p}$ and $\widehat{q}$ with respect to the lines $\widehat{L}(\widehat{z},\alpha)$ and $\widehat{L}(\widehat{z},\beta)$, it follows that $\widehat{L}(\widehat{w},\alpha)$ and $\widehat{L}(\widehat{w},\beta)$ are both non-separating lifts (see \cref{fig:stable dense}). Hence, by \cref{lem:two lifts}, $\widehat{w}$ is a stable point. Thus, the set of stable points is dense in $\mathbb{D}$. \begin{figure} \begin{center} \includegraphics{StableDense.pdf} \end{center} \caption{The configuration of lines described in the proof of \cref{lem:stable generic}.} \label{fig:stable dense} \end{figure} For the second statement, we may apply the above argument at each $\widehat{z} \in \mathbb{D}$ to obtain a countable cover $\{\widehat{V}_i: i = 1,2,\ldots\}$ of $\mathbb{D}$ by open sets with the property that for each $i$ there is at most one lift of a line, $\widehat{L}_i$, such that every point of $\widehat{V}_i \smallsetminus \widehat{L}_i$ is stable. Consider the countable family of lines $\{\varphi(\widehat{L}_i): i = 1,2,\ldots\}$. Let $L$ be any line not in this family, let $\widehat{L}$ be any lift of $L$, and let $\widehat{z} \in \widehat{L}$. Choose $i$ so that $\widehat{z} \in \widehat{V}_i$. If $\widehat{z}$ is not stable, it must be the (unique) point of intersection of $\widehat{L}$ and $\widehat{L}_i$, hence every other point in $\widehat{L} \cap \widehat{V}_i$ is stable. Therefore the set of stable points in $\widehat{L}$ is a dense open subset of $\widehat{L}$. \end{proof} We remark that since in the proof of \cref{lem:stable generic} the point $\widehat{w}$ could be chosen on either side of $\widehat{L}(\widehat{z},\theta_0)$, it follows from \cref{prop:efficient characterization} and the proof of \cref{lem:stable generic} that if $\widehat{\gamma}_0$ is a lift of an efficient path in $\overline{[\gamma]}$ with endpoints $\widehat{p}$ and $\widehat{q}$, then $\widehat{\gamma}_0([0,1]) \cap \mathbb{D}$ coincides with the set of non-stable points. This observation will not be used in the proof of \cref{thm:main1}. \subsection{Sequence of approximations of the efficient path} \label{sec:sequence of approximations} Let $\mathcal{L}$ be a countable dense (in the sense of Hausdorff distanct) family of distinct straight lines which intersect $\Omega$ and do not contain $p$ or $q$, and such that for each lift $\widehat{L}$ of a line $L \in \mathcal{L}$ the set of stable points in $\widehat{L}$ is a dense open subset of $\widehat{L}$ (this is possible by \cref{lem:stable generic}). Let $\widehat{\mathcal{L}}$ denote the (countable) set of all lifts of lines in $\mathcal{L}$. We enumerate the elements of this set: $\widehat{\mathcal{L}} = {\langle \widehat{L}_i \rangle}_{i=1}^\infty$. We construct a sequence of paths $\gamma_i$, $i \geq 1$, by recursion. To begin, let $\gamma_1 = \gamma$ and $\widehat{\gamma}_1 = \widehat{\gamma}$. Having defined $\gamma_i$ and its lift $\widehat{\gamma}_i$, we define $\gamma_{i+1}$ and $\widehat{\gamma}_{i+1}$ as follows: \begin{list}{\textbullet}{\leftmargin=2em \itemindent=0em} \item If $\widehat{\gamma}_i^{-1}(\widehat{L}_i)$ has cardinality $\leq 1$, then put $\gamma_{i+1} = \gamma_i$ and $\widehat{\gamma}_{i+1} = \widehat{\gamma}_i$. Otherwise, let $s_1$ and $s_2$ be the smallest and largest (respectively) $s \in [0,1]$ such that $\widehat{\gamma}_i(s) \in \widehat{L}_i$. Let $\widehat{\gamma}_{i+1}$ be the path in $\mathbb{D}$ (e.p.e.)\ defined by $\widehat{\gamma}_{i+1}(s) = \widehat{\gamma}_i(s)$ for $s \notin [s_1,s_2]$, and $\widehat{\gamma}_{i+1} {\upharpoonright}_{[s_1,s_2]}$ parameterizes the subarc of $\widehat{L}_i$ with endpoints $\widehat{\gamma}_i(s_1)$ and $\widehat{\gamma}_i(s_2)$ (or $\widehat{\gamma}_{i+1} {\upharpoonright}_{[s_1,s_2]}$ is constantly equal to $\widehat{w}$ if $\widehat{\gamma}_i(s_1) = \widehat{\gamma}_i(s_2) = \widehat{w}$). Let $\gamma_{i+1} = \overline{\varphi} \circ \widehat{\gamma}_{i+1}$. \end{list} \begin{lem} \label{lem:connected intersection} Let $\widehat{L}$ be a lift of a line $L$. If $i$ is such that $\widehat{\gamma}_i^{-1}(\widehat{L}) \subset [0,1]$ is connected (respectively, empty), then $\widehat{\gamma}_j^{-1}(\widehat{L})$ is connected (respectively, empty) for all $j \geq i$. In particular, for all $j > i \geq 1$, $\widehat{\gamma}_j^{-1}(\widehat{L}_i)$ is connected (or empty). \end{lem} \begin{proof} Let $\widehat{L}$ be a lift of a line $L$, and fix $i$. We will argue that if $\widehat{\gamma}_i^{-1}(\widehat{L}) \subset [0,1]$ is connected (respectively, empty), then $\widehat{\gamma}_{i+1}^{-1}(\widehat{L})$ is connected (respectively, empty). The claim then follows by induction. To this end, given the definition of $\widehat{\gamma}_{i+1}$, clearly we may assume $\widehat{L} \neq \widehat{L}_i$, so that either $\widehat{L} \cap \widehat{L}_i \cap \mathbb{D} = \emptyset$ or $\widehat{L}$ and $\widehat{L}_i$ cross transversally in $\mathbb{D}$. Assume first that $\widehat{\gamma}_i([0,1]) \cap \widehat{L} = \emptyset$. This means in particular that $\widehat{p},\widehat{q} \notin \widehat{L}$, $\widehat{L}$ is a non-separating lift, and $\widehat{\gamma}_i([0,1])$ is disjoint from the shadow of $\widehat{L}$. If $\widehat{L}_i$ does not meet $\widehat{L}$ in $\mathbb{D}$, then clearly $\widehat{\gamma}_{i+1}([0,1]) \cap \widehat{L} = \emptyset$ as well. If $\widehat{L}_i$ meets $\widehat{L}$ in $\mathbb{D}$, then these two lifts cross transversally. If $\widehat{\gamma}_{i+1}([0,1])$ meets $\widehat{L}$, we must have that $\widehat{\gamma}_{i+1}([s_1,s_2]) \cap \widehat{L} \neq \emptyset$ where $s_1 = \min\{s \in [0,1]: \widehat{\gamma}_i(s) \in \widehat{L}_i\}$ and $s_2 = \max\{s \in [0,1]: \widehat{\gamma}_i(s) \in \widehat{L}_i\}$, since $\widehat{\gamma}_{i+1} = \widehat{\gamma}_i$ outside of $[s_1,s_2]$. Since $\widehat{\gamma}_{i+1}([s_1,s_2])$ parameterizes a subarc of $\widehat{L}_i$, which crosses $\widehat{L}$ transversally, it follows that either $\widehat{\gamma}_{i+1}(s_1)$ or $\widehat{\gamma}_{i+1}(s_2)$ is in the shadow of $\widehat{L}$. But $\widehat{\gamma}_{i+1}(s_1) = \widehat{\gamma}_i(s_1)$ and $\widehat{\gamma}_{i+1}(s_2) = \widehat{\gamma}_i(s_2)$, so this contradicts the assumption that $\widehat{\gamma}_i([0,1]) \cap \widehat{L} = \emptyset$. Thus $\widehat{\gamma}_{i+1}([0,1]) \cap \widehat{L} = \emptyset$, as desired. Now assume that $\widehat{\gamma}_i([0,1]) \cap \widehat{L} \neq \emptyset$, and $\widehat{\gamma}_i^{-1}(\widehat{L})$ is an interval $[a,b]$ (here we allow the possibility that $a = b$, which means $[a,b] = \{a\}$). Suppose $\widehat{\gamma}_{i+1}$ is obtained from $\widehat{\gamma}_i$ by replacing the subpath of $\widehat{\gamma}_i$ from $s_1$ to $s_2$ ($0 < s_1 < s_2 < 1$) with a segment of $\widehat{L}_i$. We leave it to the reader to confirm the following claims: \begin{itemize} \item If $s_2 \leq a$, then $\widehat{\gamma}_{i+1}^{-1}(\widehat{L}) = [a,b]$; \item If $s_1 \leq a \leq s_2 \leq b$, then $\widehat{\gamma}_{i+1}^{-1}(\widehat{L}) = [s_2,b]$; \item If $s_1 \leq a \leq b \leq s_2$, then $\widehat{\gamma}_{i+1}^{-1}(\widehat{L})$ is either empty (if the section $\widehat{\gamma}_{i+1}([s_1,s_2])$ of $\widehat{L}_i$ does not meet $\widehat{L}$), or consists of the single point where the section $\widehat{\gamma}_{i+1}([s_1,s_2])$ of $\widehat{L}_i$ crosses $\widehat{L}$; \item It is not possible that $a \leq s_1 < s_2 \leq b$, since $\widehat{L} \neq \widehat{L}_i$ (unless $\widehat{\gamma}_i$ is constant on $[s_1,s_2]$, in which case $\widehat{\gamma}_{i+1} = \widehat{\gamma}_i$); \item If $a \leq s_1 \leq b \leq s_2$, then $\widehat{\gamma}_{i+1}^{-1}(\widehat{L}) = [a,s_1]$; and \item If $b \leq s_1$, then $\widehat{\gamma}_{i+1}^{-1}(\widehat{L}) = [a,b]$. \end{itemize} In all cases, $\widehat{\gamma}_{i+1}^{-1}(\widehat{L})$ is connected (possibly empty), as desired. \end{proof} \subsection{Convergence to the efficient path} \label{sec:convergence} For each $n = 1,2,\ldots$, we choose sets $\mathcal{G}_n$ of lines in $\mathcal{L}$ with the following properties: \begin{itemize} \item $\mathcal{G}_n$ is a finite subset of $\mathcal{L}$ for each $n$; \item $\mathcal{G}_n \subset \mathcal{G}_{n+1}$ for each $n$; \item each component of $\Omega \smallsetminus \bigcup \mathcal{G}_n$ has diameter less than $\frac{1}{n}$; and \item if $G_1,G_2 \in \mathcal{G}_n$ are distinct lines and $\widehat{G}_1$ and $\widehat{G}_2$ are separating lifts of $G_1$ and $G_2$, respectively, then $\widehat{G}_1$ and $\widehat{G}_2$ are either disjoint or meet in a stable point. \end{itemize} Let $\widehat{\mathcal{G}}_n$ be the set of all lifts of lines in $\mathcal{G}_n$. By \cref{lem:large lifts} only finitely many of the elements of $\widehat{\mathcal{G}}_n$ intersect the original lifted path $\widehat{\gamma}_1$, and among them are those that are separating lifts. We denote the set of all separating lifts in $\widehat{\mathcal{G}}_n$ by $\widehat{\mathcal{G}}_n^s$. We define an order $<$ on $\widehat{\mathcal{G}}_n^s$ as follows. Let $\widehat{G}_1,\widehat{G}_2 \in \widehat{\mathcal{G}}_n^s$ be distinct separating lifts. First suppose $\widehat{G}_1 \cap \widehat{G}_2 = \emptyset$. Then $\widehat{G}_1 < \widehat{G}_2$ if $\widehat{G}_2$ is on the $\widehat{q}$-side of $\widehat{G}_1$; otherwise $\widehat{G}_2 < \widehat{G}_1$. Now suppose $\widehat{G}_1 \cap \widehat{G}_2 = \{\widehat{z}\}$ for some stable point $\widehat{z} \in \mathbb{D}$. Since $\widehat{z}$ is stable, it follows from density of the family of lines $\mathcal{L}$ that there exists $\widehat{W} \in \widehat{\mathcal{L}}$ such that $\widehat{W}$ is non-separating and contains $\widehat{z}$ in its shadow. Then $\widehat{G}_1 < \widehat{G}_2$ if $\widehat{G}_2 \smallsetminus \overline{\mathrm{Sh}(\widehat{W})}$ is on the $\widehat{q}$-side of $\widehat{G}_1$; otherwise $\widehat{G}_2 < \widehat{G}_1$. It is straightforward to see that this relation $<$ is a well-defined linear order on $\widehat{\mathcal{G}}_n^s$. For each $n = 1,2,\ldots$, choose a finite set $\widehat{\mathcal{W}}_n$ of lifts of lines from $\widehat{\mathcal{L}}$ such that for each pair of distinct intersecting separating lifts $\widehat{G}_1,\widehat{G}_2 \in \widehat{\mathcal{G}}_n^s$ there is a lift $\widehat{W} \in \widehat{\mathcal{W}}_n$ witnessing that $\widehat{G}_1 < \widehat{G}_2$ or that $\widehat{G}_2 < \widehat{G}_1$ (i.e.\ such that $\mathrm{Sh}(\widehat{W})$ contains $\widehat{G}_1 \cap \widehat{G}_2$). Let $i(n)$ be large enough so that $\widehat{\mathcal{W}}_n \subset \{\widehat{L}_i: i = 1,\ldots,i(n)-1\}$, and also all of the (finitely many) elements of $\widehat{\mathcal{G}}_n$ which intersect the original lifted path $\widehat{\gamma}_1$ are contained in $\{\widehat{L}_i: i = 1,\ldots,i(n)-1\}$. Take an arbitrary $n$, and let $\widehat{G}_1,\ldots,\widehat{G}_m$ enumerate the elements of $\widehat{\mathcal{G}}_n^s$, listed in $<$ order. For any $i \geq i(n)$, by \cref{lem:connected intersection} the lifted path $\widehat{\gamma}_i$ does not cross any of the non-separating lifts in $\widehat{\mathcal{G}}_n \cup \widehat{\mathcal{W}}_n$, and has connected intersection with all the lifts that it meets, in particular with the elements of $\widehat{\mathcal{G}}_n^s$. Thus there are $m+1$ components $\widehat{R}_1,\ldots,\widehat{R}_{m+1}$ of $\mathbb{D} \smallsetminus \bigcup (\widehat{\mathcal{G}}_n \cup \widehat{\mathcal{W}}_n)$ such that for any $i \geq i(n)$, the lifted path $\widehat{\gamma}_i$ starts at $\widehat{p}$ then runs in $\widehat{R}_1$, until it crosses $\widehat{G}_1$ and enters $\widehat{R}_2$, then crosses $\widehat{G}_2$ and enters $\widehat{R}_3$, and so on, until it crosses $\widehat{G}_m$ to enter $\widehat{R}_{m+1}$ and ends at $\widehat{q}$ (see \cref{fig:regions}). \begin{figure} \begin{center} \includegraphics{Regions.pdf} \end{center} \caption{An illustration depicting a possible configuration of the lifts of lines $\widehat{G}_k$ and regions $\widehat{R}_k$ in between, as described in \cref{sec:convergence}. The thick curve represents a path $\widehat{\gamma}_i$ (for $i \geq i(n)$), the dashed arcs represent elements of $\widehat{\mathcal{G}}_n$ (including the separating lifts $\widehat{G}_k$), and the dotted arcs represent elements of $\widehat{\mathcal{W}}_n$.} \label{fig:regions} \end{figure} For each of these regions $\widehat{R}_j$, $\varphi(\widehat{R}_j)$ is contained in a component of $\Omega \smallsetminus \bigcup \mathcal{G}_n$, hence has diameter less than $\frac{1}{n}$. Therefore, provided we parameterize the path $\gamma_i$ so that if $\widehat{\gamma}_{i(n)}(t) \in \widehat{R}_j$ then $\widehat{\gamma}_i(t) \in \widehat{R}_j$ for all $i \geq i(n)$, we have that $d_{\mathrm{sup}}(\gamma_{i_1},\gamma_{i_2}) < \frac{1}{n}$ for all $i_1,i_2 \geq i(n)$; in fact there is a homotopy between $\gamma_{i_1}$ and $\gamma_{i_2}$ which moves no point more than $\frac{1}{n}$, obtained by moving $\widehat{\gamma}_{i_1}(t)$ to $\widehat{\gamma}_{i_2}(t)$ within the closure of a region $\widehat{R}_j$ which contains $\widehat{\gamma}_{i(n)}(t)$. We remark that according to \cref{prop:efficient characterization}, if $\lambda \in \overline{[\gamma]}$ is any efficient path, it must follow this same pattern traversing through the (closures of the) regions $\widehat{R}_1,\ldots,\widehat{R}_{m+1}$ in the same monotone order. Therefore, any two such efficient paths, if parameterized appropriately, are within $\frac{1}{n}$ of each other. Since $n$ is arbitrary, this establishes that there can be at most one efficient path in $\overline{[\gamma]}$ (up to reparameterization). It remains to confirm that our construction above yields an efficient path in $\overline{[\gamma]}$. Since $n$ was arbitrary, we now have that $\langle \gamma_i \rangle_{i=1}^\infty$ is a Cauchy sequence of paths, hence it converges to a path $\gamma_\infty$ from $p$ to $q$. Moreover, by putting together the (smaller and smaller) homotopies between the paths in this sequence, we have that $\gamma_\infty \in \overline{[\gamma]}$. \begin{lem} \label{lem:reduced} $\gamma_\infty$ is an efficient path in $\overline{[\gamma]}$. \end{lem} \begin{proof} This follows from \cref{prop:efficient characterization}: if $L$ is any line intersecting $\Omega$ which has a lift $\widehat{L}$ such that $\widehat{\gamma}_\infty([0,1]) \cap \widehat{L}$ is not connected, then by density of the family $\mathcal{L}$ it is easy to see there must be a line $L' \in \mathcal{L}$ close to $L$, with a lift $\widehat{L}'$ close to $\widehat{L}$, such that $\widehat{\gamma}_\infty([0,1]) \cap \widehat{L}'$ is also not connected. But if $\widehat{L}' = \widehat{L}_i$ in the enumeration of $\widehat{\mathcal{L}}$, then for all $j > i$, $\widehat{\gamma}_j([0,1]) \cap \widehat{L}'$ is connected according to \cref{lem:connected intersection}. Since $\widehat{\gamma}_i \to \widehat{\gamma}_\infty$, the same must be true for $\widehat{\gamma}_\infty$, a contradiction. \end{proof} This completes the proof of \cref{thm:main1}. \section{Proof of \cref{thm:main2}} \label{sec:proof main2} Having established the existence of a unique efficient path $\lambda$ in $\overline{[\gamma]}$, we now prove \cref{thm:main2} using \cref{thm:main1} and its proof. In particular, we observe that in the construction of the efficient path in the proof of \cref{thm:main1}, we started with a path in $[\gamma]$ (such as $\gamma$ itself) and constructed a sequence of paths converging to the efficient path, where each path in the sequence is obtained from the previous one by replacing a subpath with a straight segment. Such a replacement will always decrease the $\mathsf{len}$ length of the path and also, when finite, the Euclidean length. It is clear that any path in $\overline{[\gamma]}$ of smallest $\mathsf{len}$ length (or Euclidean length if finite) must be efficient, because replacing a subpath with a straight segment decreases the length. Thus we have (3) $\Rightarrow$ (2) and (4) $\Rightarrow$ (2). For (2) $\Rightarrow$ (3), suppose for a contradiction that there is a path $\rho \in \overline{[\gamma]}$ with strictly smaller $\mathsf{len}$ length than the efficient path $\lambda \in \overline{[\gamma]}$. By continuity of the function $\mathsf{len}$, it follows that there exists $\rho' \in [\gamma]$ whose $\mathsf{len}$ length is also strictly smaller than $\lambda$. If we follow the construction in the proof of \cref{thm:main1} starting with $\rho'$ instead of $\gamma$, we would then obtain an efficient path in $\overline{[\gamma]}$ of $\mathsf{len}$ length strictly smaller than $\lambda$, contradicting the uniqueness of the efficient path. For (2) $\Rightarrow$ (4), suppose for a contradiction that there is a path $\rho \in \overline{[\gamma]}$ with finite Euclidean length $E$ which is strictly smaller than the Euclidean length of the efficient path $\lambda \in \overline{[\gamma]}$. This means there exists $\varepsilon > 0$ and values $0 = s_0 < s_1 < \cdots < s_N = 1$ such that $\sum_{i=1}^N |\lambda(s_i) - \lambda(s_{i-1})| > E + \varepsilon$. Let $\widehat{\gamma}$ be a lift of $\gamma$ with endpoints $\widehat{p},\widehat{q} \in \overline{\mathbb{D}}$, and let $\widehat{\lambda}$ and $\widehat{\rho}$ be lifts of $\lambda$ and $\rho$ with the same endpoints $\widehat{p},\widehat{q}$. We will replace sections of the path $\rho$ with straight line segments to obtain a new path $\rho' \in \overline{[\gamma]}$ which goes within $\frac{\varepsilon}{2N}$ of each of the points $\lambda(s_i)$, in order. To obtain $\rho'$, we proceed as follows. For each $i = 1,\ldots,N-1$, consider the point $\widehat{\lambda}(s_i) \in \overline{\mathbb{D}}$. \begin{itemize} \item If $\widehat{\lambda}(s_i) \in \mathbb{D}$, by \cref{lem:stable generic} we can choose two stable points $\widehat{w}_1,\widehat{w}_2$ in a small neighborhood $\widehat{V}$ of $\widehat{\lambda}(s_i)$ which projects one-to-one under $\varphi$ to a small disk centered at $\lambda(s_i)$ of radius less than $\frac{\varepsilon}{2N}$, so that $\widehat{w}_1$ and $\widehat{w}_2$ are on opposite sides of $\widehat{\lambda}([0,1]) \cap \widehat{V}$. Let $\widehat{L}_1$ and $\widehat{L}_2$ be non-separating lifts of lines such that $\widehat{w}_1 \in \mathrm{Sh}(\widehat{L}_1)$ and $\widehat{w}_2 \in \mathrm{Sh}(\widehat{L}_2)$. By \cref{prop:efficient characterization}, $\widehat{\lambda}$ does not cross $\widehat{L}_1$ or $\widehat{L}_2$. It follows that if we replace the section of $\widehat{\rho}$ between its first and last intersections with $\widehat{L}_1$ with an arc in $\widehat{L}_1$, and likewise for $\widehat{L}_2$, then the resultant path must meet $\widehat{V}$. \item If $\widehat{\gamma}(s_i) \in \partial \mathbb{D}$, choose an arc $\widehat{A}$ such that $\overline{\varphi}(\widehat{A})$ has diameter less than $\frac{\varepsilon}{2N}$, with endpoints $\widehat{\lambda}(s_i)$ and a stable point $\widehat{w} \in \mathbb{D}$. Let $\widehat{L}$ be a lift of a line such that $\widehat{w} \in \mathrm{Sh}(\widehat{L})$. By \cref{prop:efficient characterization}, $\widehat{\lambda}$ does not cross $\widehat{L}$. It follows that if we replace the section of $\widehat{\rho}$ between its first and last intersections with $\widehat{L}$ with an arc in $\widehat{L}$, then the resultant path must meet $\widehat{A}$. \end{itemize} After making the finitely many replacements of subpaths of $\widehat{\rho}$ with segments of lifts of lines as described above, the resultant path $\widehat{\rho}'$ is such that $\rho' = \overline{\varphi} \circ \widehat{\rho}'$ is continuous, and $\rho' \in \overline{[\gamma]}$ by \cref{thm:class closure lift}. Clearly the Euclidean length of $\rho'$ is not greater than $E$. On the other hand, by construction, there are values $0 = s_0' < s_1' < \cdots < s_N' = 1$ such that $|\rho'(s_i') - \lambda(s_i)| < \frac{\varepsilon}{2N}$ for each $i = 0,\ldots,N$. It follows that the Euclidean length of $\rho'$ is at least \begin{align*} \sum_{i=1}^N |\rho'(s_i') - \rho'(s_{i-1}')| &> \sum_{i=1}^N \left( |\lambda(s_i) - \lambda(s_{i-1})| - 2 \cdot \frac{\varepsilon}{2N} \right) \\ &= \left( \sum_{i=1}^N |\lambda(s_i) - \lambda(s_{i-1})| \right) - \varepsilon \\ &> (E + \varepsilon) - \varepsilon = E , \end{align*} a contradiction. Therefore, $\lambda$ has smallest Euclidean length among all paths in $\overline{[\gamma]}$. The proof that (1) $\Rightarrow$ (2) is straightforward and left to the reader. Finally, for (2) $\Rightarrow$ (1), let $0 < s_1 < s_2 < 1$ and consider the path $\lambda {\upharpoonright}_{[s_1,s_2]}$. We claim that there exist $s_1',s_2'$ such that $0 < s_1' \leq s_1 < s_2 \leq s_2' < 1$ and $\widehat{\lambda}(s_1')$ and $\widehat{\lambda}(s_2')$ can be joined by a path in $\overline{\mathbb{D}}$ which projects under $\overline{\varphi}$ to a path of finite Euclidean length. We rely on the fact that any two points $\widehat{w}_1,\widehat{w}_2 \in \mathbb{D}$ can be joined by a path in $\mathbb{D}$ which projects to a path of finite Euclidean length (e.g.\ a piecewise linear path). If $\widehat{\lambda}([0,s_1]) \cap \mathbb{D} \neq \emptyset$, then let $0 < s_1' \leq s_1$ be such that $\widehat{\lambda}(s_1') \in \mathbb{D}$, and let $\widehat{w}_1 = \widehat{\lambda}(s_1')$. If $\widehat{\lambda}([0,s_1]) \subset \partial \mathbb{D}$, we choose $s_1'$ and $\widehat{w}_1$ as follows. Note that in this case $\lambda(s_1)$ is not isolated in $\partial \Omega$, so we can apply \cref{thm:crosscut} to obtain a component $\widehat{C}_1$ of $\varphi^{-1}(\partial B(\lambda(s_1),\varepsilon))$, for some sufficiently small $\varepsilon > 0$, whose closure is an arc which separates $\widehat{\lambda}(s_1)$ from $\widehat{p}$ in $\overline{\mathbb{D}}$. In particular, there exists $0 < s_1' < s_1$ such that $\widehat{\lambda}(s_1')$ is in the closure of $\widehat{C}_1$. Let $\widehat{w}_1$ be any point in $\widehat{C}_1$. Thus, we have established that there exists $0 < s_1' \leq s_1$ and $\widehat{w}_1 \in \mathbb{D}$ such that either $\widehat{\lambda}(s_1') = \widehat{w}_1$ or $\widehat{\lambda}(s_1')$ can be joined to $\widehat{w}_1$ by a path in $\overline{\mathbb{D}}$ which projects to a circular arc. By the same reasoning, there exists $s_2 \leq s_2' < 1$ and $\widehat{w}_2 \in \mathbb{D}$ such that either $\widehat{\lambda}(s_2') = \widehat{w}_2$ or $\widehat{\lambda}(s_2')$ can be joined to $\widehat{w}_2$ by a path in $\overline{\mathbb{D}}$ which projects to a circular arc. It follows that $\widehat{\lambda}(s_1')$ can be joined to $\widehat{\lambda}(s_2')$ by a path in $\overline{\mathbb{D}}$ which projects to a path of finite Euclidean length. Therefore, by \cref{cor:locally shortest well-defined}, there exists a path of finite Euclidean length in $\overline{[\lambda^\gamma_{[s_1',s_2']}]}$. Since $\lambda$ is efficient, it follows from \cref{prop:efficient characterization} that $\lambda {\upharpoonright}_{[s_1',s_2']}$ is the efficient path in $\overline{[\lambda^\gamma_{[s_1',s_2']}]}$, hence has smallest Euclidean length in this class according to the implication (2) $\Rightarrow$ (4). In particular, $\lambda {\upharpoonright}_{[s_1',s_2']}$ has finite Euclidean length in light of the previous paragraph. This in turn implies that the subpath $\lambda {\upharpoonright}_{[s_1,s_2]}$ has finite Euclidean length. Again by \cref{prop:efficient characterization}, we have that $\lambda {\upharpoonright}_{[s_1,s_2]}$ is the efficient path in $\overline{[\lambda^\gamma_{[s_1,s_2]}]}$, hence has smallest Euclidean length in this class according to the implication (2) $\Rightarrow$ (4), as required. Therefore, $\lambda$ is locally shortest. \bibliographystyle{amsplain}

https://arxiv.org/abs/1308.4613

Random subtrees of complete graphs

We study the asymptotic behavior of four statistics associated with subtrees of complete graphs: the uniform probability $p_n$ that a random subtree is a spanning tree of $K_n$, the weighted probability $q_n$ (where the probability a subtree is chosen is proportional to the number of edges in the subtree) that a random subtree spans and the two expectations associated with these two probabilities. We find $p_n$ and $q_n$ both approach $e^{-e^{-1}}\approx .692$, while both expectations approach the size of a spanning tree, i.e., a random subtree of $K_n$ has approximately $n-1$ edges.

\section{Introduction} We are interested in the following two questions: \begin{center} \begin{itemize} \item [Q1.] What is the asymptotic probability that a random subtree of $K_n$ is a spanning tree? \item [Q2.] How many edges (asymptotically) does a random subtree of $K_n$ have? \end{itemize} \end{center} In answering both questions, we consider two different probability measures: a uniform random probability $p_n$, where each subtree has an equal probability of being selected, and a weighted probability $q_n$, where the probability a subtree is selected is proportional to its size (measured by the number of edges in the subtree). As expected, weighting subtrees by their size increases the chances of selecting a spanning tree, i.e., $p_n<q_n$. Table~\ref{globalprobdata} gives data for these values when $ n \leq 100$. \begin{table}[htdp] \begin{center} \begin{tabular}{c|ll} $n$ & $p_n$ & $q_n$ \\ \hline 10 & 0.617473 & 0.652736 \\ 20 & 0.657876 & 0.672725 \\ 30 & 0.669904 & 0.679294 \\ 40 & 0.675689 & 0.682552 \\ 50 & 0.67909 & 0.684497 \\ 60 & 0.681329 & 0.685789 \\ 70 & 0.682915 & 0.686711 \\ 80 & 0.684097 & 0.687401 \\ 90 & 0.685012 & 0.687936 \\ 100 & 0.685741 & 0.688365 \\ \end{tabular} \caption{Probabilities of selecting a spanning tree using uniform and weighted probabilities.} \label{globalprobdata} \end{center} \end{table} The (somewhat) surprising result is that $p_n$ and $q_n$ approach the same limit as $n \to \infty$. This is our first main result, completely answering Q1. \begin{thm}\label{T:main1} \begin{enumerate} \item Let $p_n$ be the probability of choosing a spanning tree among all subtrees of $K_n$ with uniform probability, i.e., the probability any subtree is selected is the same. Then $$\lim_{n \to \infty} p_n=e^{-e^{-1}}=0.692201\dots$$ \item Let $q_n$ be the probability of choosing a spanning tree among all subtrees of $K_n$ with weighted probability, i.e., the probability any subtree is selected is proportional to its number of edges. Then $$\lim_{n \to \infty} q_n=e^{-e^{-1}}=0.692201\dots$$ \end{enumerate} \end{thm} For the second question Q2, the expected number of edges of a random subtree of $K_n$ is $\sum pr(T) |E(T)|$, where $pr(T)$ is the probability a tree $T$ is selected, $E(T)$ is the edge set of $T$, and the sum is over all subtrees $T$ of $K_n$. Since we have two distinct probability functions $p_n$ and $q_n$, we obtain two distinct expected values. The relation between these two expected values is equivalent to a famous example from elementary probability: \begin{quote} All universities report ``average class size.'' However, this average depends on whether you first choose a class at random, or first select a student at random, and then ask that student to randomly select one of their classes. \end{quote} Our uniform expectation is exactly analogous to the first situation, and our weighted expectation is equivalent to the student weighted average. In this context, edges play the role of students and the subtrees are the classes. It is a straightforward exercise to show the student weighting always produces a larger expectation. This was first noticed by Feld and Grofman in 1977 in \cite{fg}. For our purposes, this result will show the weighted expectation is always greater than the uniform expectation. Since both of these expected values are obviously bounded above by $n-1$, the size of a spanning tree, we use a variation of {\it subtree density} first defined in \cite{jam1}. We divide our expected values by $n-1$, the number of edges in a spanning tree, to convert our expectations to densities. Letting $a_k$ equal the number of $k$-edge subtrees in $K_n$, this gives us two formulas for subtree density, one using uniform probability and one using weighted probability: $$\mbox{Uniform density: } \mu_p(n)=\frac{\sum_{k=1}^{n-1}ka_k}{(n-1)\sum_{k=1}^{n-1}a_k} \hskip.5in \mbox{Weighted density: } \mu_q(n)=\frac{\sum_{k=1}^{n-1}k^2a_k}{(n-1)\sum_{k=1}^{n-1}ka_k}$$ Table~\ref{globaldensities} gives data for these two densities when $n \leq 100$. \begin{table}[htdp] \begin{center} \begin{tabular}{c|ll} $n$ & $\mu_p(n)$ & $\mu_q(n)$ \\ \hline 10 & 0.945976 & 0.952436 \\ 20 & 0.977928 & 0.97912 \\ 30 & 0.986177 & 0.986661 \\ 40 & 0.989945 & 0.990205 \\ 50 & 0.9921 & 0.992263 \\ 60 & 0.993496 & 0.993607 \\ 70 & 0.994472 & 0.994553 \\ 80 & 0.995194 & 0.995255 \\ 90 & 0.995749 & 0.995797 \\ 100 & 0.996189 & 0.996228 \\ \end{tabular} \caption{Subtree densities using uniform and weighted probabilities.} \label{globaldensities} \end{center} \end{table}% Evidently, both of these densities approach 1. This is our second main result, and our answer to Q2. \begin{thm}\label{T:main2} \begin{enumerate} \item $\displaystyle{\lim_{n\to\infty}\mu_p(n) = 1,}$ \item $\displaystyle{\lim_{n\to\infty} \mu_q(n) = 1.}$ \end{enumerate} \end{thm} The fact that the probabilities and the densities do not depend on which probability measure we use is an indication of the dominance of the number of spanning trees in $K_n$ compared to the the number of non-spanning trees. Theorems~\ref{T:main1} and \ref{T:main2} are proven in Section~\ref{S:global}. The proofs follow from a rather detailed analysis of the growth rate of individual terms in the sums that are used to compute all of the statistics. But we emphasize that these proof techniques are completely elementary. Subtree densities have been studied before, but apparently only when the graph is itself a tree. Jamison introduced this concept in \cite{jam1} and studied its properties in \cite{jam2}. A more recent paper of Vince and Wang \cite{vw} characterizes extremal families of trees with the largest and smallest subtree densities, answering one of Jamison's questions. A recent survey of results connecting subtrees of trees with other invariants, including the Weiner index, appears in \cite{sw}. There are several interesting directions for future research in this area. We indicate some possible projects in Section~\ref{S:future}. \section{Global probabilities}\label{S:global} Our goal in this section is to provide proofs of Theorems~\ref{T:main1} and \ref{T:main2}. As usual, $K_n$ represents the complete graph on $n$ vertices. We fix notation for the subtree enumeration we will need. \begin{notation} Assume $n$ is fixed. We define $a_k, b_k, A$ and $B$ as follows. \begin{itemize} \item Let $a_k$ denote the number of $k$-edge subtrees in $K_n$. (We ignore subtrees of size 0, although setting $a_0=n$ will not change the asymptotic behavior of any of our statistics.) \item Let $\displaystyle{A=\sum_{k=1}^{n-1}a_k}$ be the total number of subtrees of all sizes in $K_n$. \item Let $b_k=ka_k$ denote the number of edges used by all of the $k$-edge subtrees in $K_n$. \item Let $\displaystyle{B=\sum_{k=1}^{n-1}b_k}$ be the sum of the sizes (number of edges) of all the subtrees of $K_n$. \end{itemize} \end{notation} It is immediate from Cayley's formula that $\displaystyle{ a_k= \binom{n}{k+1} (k+1)^{k-1}.}$ We can view $B$ as the sum of all the entries in a 0--1 edge--tree incidence matrix. The four statistics we study here, $p_n, q_n, \mu_p(n)$ and $\mu_q(n)$, can be computed using $A, B, a_k$ and $b_k$. We omit the straightforward proof of the next result. \begin{lem}\label{L:global} Let $p_n, q_n, \mu_p(n)$ and $\mu_q(n)$ be as given above. Then \begin{enumerate} \item $\displaystyle{p_n=\frac{a_{n-1}}{A}}$, \item $\displaystyle{q_n=\frac{b_{n-1}}{B}}$, \item $\displaystyle{\mu_p(n)=\frac{B}{(n-1)A}}$, \item $\displaystyle{\mu_q(n)=\frac{\sum_{k=1}^{n-1}kb_k}{(n-1)B}=\frac{\sum_{k=1}^{n-1}k^2a_k}{(n-1)B}}$. \end{enumerate} \end{lem} \begin{ex} We compute each of these statistics for the graph $K_4$. In this case, there are 6 subtrees of size one (the 6 edges of $K_4$), 12 subtrees of size two and 16 spanning trees (of size three). Then we find $\displaystyle{p_4=\frac{16}{34}=.471\dots}$, $\displaystyle{q_4=\frac{48}{78}=.615\dots}$, $\displaystyle{\mu_p(4)=\frac{78}{102}=.768\dots}$, and $\displaystyle{\mu_q(4)=\frac{198}{234}=.846\dots}$. \end{ex} The remainder of this section is devoted to proofs of Theorems~\ref{T:main1} and \ref{T:main2}. We first prove part (1) of Theorem~\ref{T:main2}, then use the bounds from Lemmas~\ref{lemmatop} and \ref{lemmabottom} to help prove both parts of Theorem~\ref{T:main1}. Lastly, we prove part (2) of Theorem~\ref{T:main2}. Thus, our immediate goal is to prove that the uniform density $\displaystyle{\mu_p(n)=\frac{B}{(n-1)A}}$ approaches 1 as $n \to \infty$. Our approach is as follows: We bound the numerator $B$ from below and the term $A$ in the denominator from above so that $\mu_p(n)$ is bounded below by a function that approaches 1 as $n \to \infty$. Lemma~\ref{lemmatop} establishes the lower bound for $B$ and Lemma~\ref{lemmabottom} establishes the upper bound for $A$, from which the result follows. \begin{lem} \label{lemmatop} Let $\displaystyle{B=\sum_{k=1}^{n-1}k\binom{n}{k+1} (k+1)^{k-1}}$, as above. Then \begin{equation} B > (n-1)n^{n-2} \left( \frac{n-3}{n-1}\right) e^{e^{-1}}. \label{topresult} \end{equation} \end{lem} \begin{proof} Recall $b_{n-1} = (n-1)n^{n-2}$ counts the total number of edges used in all the spanning trees of $K_n$. We examine the ratio $b_{i}/b_{i-1}$ in order to establish a lower bound for each $b_i$ in terms of $b_{n-1}$. \begin{equation} \frac{b_i}{b_{i-1}} = \frac{\binom{n}{i+1} (i+1)^{i-1} i}{\binom{n}{i} i^{i-2} (i-1)} = \frac{n-i}{i+1}\cdot \frac{i}{i-1}\cdot \frac{(i+1)^{i-1}}{i^{i-2}} = (n-i)\cdot\frac{i}{i-1} \cdot\left(\frac{i+1}{i}\right)^{i-2} \label{topratio} \end{equation} We use the fact that $e$ is the least upper bound for the sequence $\left\{\left(\frac{i+1}{i}\right)^{i-2}\right\}$ to rewrite \eqref{topratio} as the inequality $$ b_{i-1} \geq \frac{i-1}{i(n-i)e} \cdot b_i, $$ and this is valid for $i = 2, 3,\dots,n-1$. In general, for $i<n$, an inductive argument on $n-i$ establishes the following: \begin{equation} b_{i} \geq \frac{i}{(n-i-1)!(n-1)e^{n-i-1}} \cdot b_{n-1}. \label{topinequality} \end{equation} Then equation \eqref{topinequality} gives \begin{equation} B = \sum_{i=1}^{n-1} b_i \geq \frac{b_{n-1}}{n-1} \left((n-1) + \frac{n-2}{e} + \frac{n-3}{2e^2} + \frac{n-4}{6e^3} + \dots + \frac{1}{(n-2)!e^{n-2}}\right). \end{equation} We bound this sum below using standard techniques from calculus. Let $$h(k) = 1 + \frac{1}{e} + \frac{1}{2e^2} + \frac{1}{6e^3} + \dots + \frac{1}{k!e^{k}}.$$ Then \begin{equation} B \geq \frac{b_{n-1}}{n-1}(h(n-2) + h(n-3) + \dots + h(0)). \label{deriv} \end{equation} Now $\displaystyle{e^{x} = \sum_{i=0}^{k}\frac{x^i}{i!} + R_k(x)}$, where $\displaystyle{R_k(x) = \frac{e^y}{(k+1)!} x^{k+1}}$ for some $y \in (0,x)$. We are interested in this expression when $x=e^{-1}$. Then $R_k(x)$ is maximized when $x=y= e^{-1}$. So, using $e^{(e^{-1})} \approx 1.44 \ldots<2$, we have $$ R_k\left(e^{-1}\right) \leq \frac{e^{(e^{-1})}}{{(k+1)!} }(e^{-1})^{k+1} \leq \frac{2}{(k+1)!e^{k+1}}. $$ Now, using this upper bound on the error in the Maclaurin polynomial for $e^x$ at $x = e^{-1}$ gives $$ e^{e^{-1}} = h(k) + R_k\left(e^{-1}\right) \leq h(k) + \frac{2}{(k+1)!e^{k+1}}, $$ so $$ h(k) \geq e^{e^{-1}} - \frac{2}{(k+1)!e^{k+1}}. $$ Substituting into \eqref{deriv}, \begin{align*} B &\geq \frac{b_{n-1}}{n-1}\sum_{k=0}^{n-2} h(k) \geq \frac{b_{n-1}}{n-1} \left((n-1)e^{e^{-1}} - 2\sum_{k=0}^{n-2}\frac{1}{(k+1)!e^{k+1}}\right) \\ &> \frac{b_{n-1}}{n-1} \left((n-1)e^{e^{-1}} - 2\sum_{i=0}^\infty\frac{1}{i!e^i}\right) = \frac{b_{n-1}}{n-1} \left((n-1)e^{e^{-1}} - 2e^{e^{-1}}\right) = b_{n-1} \left( \frac{n-3}{n-1}\right)e^{e^{-1}}. \end{align*} \end{proof} We now give an upper bound for $A$, the total number of subtrees of $K_n$. \begin{lem} \label{lemmabottom} Let $\displaystyle{A=\sum_{k=1}^{n-1}a_k}$ be the total number of subtrees of all sizes in $K_n$, as above. Then, for every $\varepsilon >0$, there is a positive integer $r(\varepsilon) \in \mathbb{N}$ so that, for all $n>r(\varepsilon)$, \begin{equation} A <n^{n-2} \left(e^{(e-\varepsilon)^{-1}} + \frac{e}{r(\varepsilon)!}\right) \label{bottomresult} \end{equation} \end{lem} \begin{proof} Recall $a_{n-1}=n^{n-2}$ is the number of spanning trees in $K_n$. As in Lemma~\ref{lemmatop}, we examine ratios of consecutive terms, but this time we need to establish an upper bound for the $a_i$ in terms of $a_{n-1}$. Now \begin{equation} \frac{a_i}{a_{i-1}} = \frac{\binom{n}{i+1} (i+1)^{i-1}}{\binom{n}{i} i^{i-2}} = \frac{n-i}{i+1}\cdot \frac{(i+1)^{i-1}}{i^{i-2}} = (n-i) \left(\frac{i+1}{i}\right)^{i-2} \label{bottomratio} \end{equation} Let $\varepsilon > 0$. Since $\lim_{i\to\infty} \left(\frac{i+1}{i}\right)^{i-2} = e$, there exists a $k(\varepsilon)$ such that $$ a_{i-1} \leq \frac{a_i}{(n-i)(e-\varepsilon)} $$ for every $k(\varepsilon) < i \leq n-1$. As in the proof of Lemma~\ref{lemmatop}, an inductive argument can be used to show $$ a_i \leq \frac{a_{n-1}}{(n-i-1)!(e-\varepsilon)^{n-i-1}} $$ for all $i$ such that $k(\varepsilon) < i \leq n-1$. On the other hand, if $i\leq k(\varepsilon)$, then $$ a_{i-1}= \frac{a_i}{(n-i)\left(\frac{i+1}{i}\right)^{i-2}} \leq \frac{a_i}{n-i} \leq \frac{a_{n-1}}{(n-i)!} $$ where we have bounded $\left(\frac{i+1}{i}\right)^{i-2}$ below by 1 and the final inequality follows by a similar inductive argument. Therefore, \begin{equation*} A = \sum_{i=1}^{n-1}a_i \leq a_{n-1} (f(n,k(\varepsilon)) + g(n,k(\varepsilon))) \end{equation*} where \begin{equation*} f(n,k(\varepsilon)) = 1 + \frac{1}{e-\varepsilon} + \frac{1}{2!(e-\varepsilon)^2} + \dots + \frac{1}{(n-k(\varepsilon))!(e-\varepsilon)^{n-k(\varepsilon)}}, \end{equation*} corresponding to those terms where $i>k(\varepsilon)$, and \begin{equation*} g(n,k(\varepsilon)) = \frac{1}{(n-k(\varepsilon)+1)!} + \frac{1}{(n-k(\varepsilon)+2)!} + \dots + \frac{1}{(n-2)!} + \frac{1}{(n-1)!} \end{equation*} corresponds to the terms where $i\leq k(\varepsilon)$. Using the Maclaurin expansion for $e^x$ evaluated at $x = (e-\varepsilon)^{-1}$ gives an upper bound for $f(n,k(\varepsilon))$: \begin{equation*} f(n,k(\varepsilon)) = \sum_{i=0}^{n-k(\varepsilon)} \frac{1}{i!(e-\varepsilon)^i} < \sum_{i=0}^{\infty} \frac{1}{i!(e-\varepsilon)^i} = e^{(e-\varepsilon)^{-1}}. \end{equation*} For $g(n,k(\varepsilon))$, we have \begin{equation*} g(n,k(\varepsilon)) = \sum_{i = n-k(\varepsilon)+2}^{n} \frac{1}{(i-1)!} < \sum_{i=n-k(\varepsilon)+2}^{\infty} \frac{1}{(i-1)!} <\frac{e}{r(\varepsilon)!}, \end{equation*} where $r(\varepsilon) = n-k(\varepsilon) + 1$. Therefore, \begin{equation*} A \leq a_{n-1}(f(n,k(\varepsilon)) + g(n,k(\varepsilon))) < a_{n-1}\left(e^{(e-\varepsilon)^{-1}} + \frac{e}{r(\varepsilon)!}\right). \end{equation*} \end{proof} We can now prove part (1) of Theorem~\ref{T:main2}. \begin{proof} [Proof: Theorem~\ref{T:main2} (1)] Recall $b_{n-1} = (n-1)n^{n-2}$ and $a_{n-1} = n^{n-2}$, so \begin{equation*} \frac{1}{n-1} \cdot \frac{b_{n-1}}{a_{n-1}} = 1. \end{equation*} Therefore, \eqref{topresult} and \eqref{bottomresult} imply \begin{equation*} \mu_p(n) = \frac{1}{n-1} \cdot \frac{B}{A} > \frac{1}{n-1} \cdot \frac{b_{n-1}}{a_{n-1}} \cdot \frac{\left( \frac{n-3}{n-1} \right) e^{e^{-1}} }{e^{(e-\varepsilon)^{-1}} + \frac{e}{r(\varepsilon)!}} =\left( \frac{n-3}{n-1}\right) \cdot \frac{e^{e^{-1}} }{e^{(e-\varepsilon)^{-1}} + \frac{e}{r(\varepsilon)!}}. \end{equation*} Then $\displaystyle{\lim_{n\to\infty} \mu_p(n) = 1}$ as $n \to \infty$ since $\varepsilon$ can be chosen arbitrarily small and $r(\varepsilon)$ can be made arbitrarily large. \end{proof} We now prove Theorem~\ref{T:main1}. \begin{proof}[Proof: Theorem~\ref{T:main1}] \begin{enumerate} \item Recall $p_n=\frac{a_{n-1}}{A}$, where $a_{n-1}$ is the number of spanning trees in $K_n$ and $A$ is the total number of subtrees of $K_n$. Then the argument in Lemma~\ref{lemmabottom} can be modified to prove $$A \geq a_{n-1}\sum_{i=0}^{n-1}\frac{1}{i!e^i}.$$ This follows by bounding $\displaystyle{ \left(\frac{i+1}{i}\right)^{i-2}}$ below by $e$ in equation~\eqref{bottomratio} -- this is the same bound we needed in our proof of Lemma~\ref{lemmatop}. Then, bounding $\displaystyle{ \frac{1}{p_n}=\frac{A}{a_{n-1}} }$, we have $$e^{e^{-1}}-\varepsilon' \leq \sum_{1=0}^{n-1}\frac{1}{i!e^i} \leq \frac{A} {a_{n-1}}\leq \left(e^{(e-\varepsilon)^{-1}} + \frac{e}{r(\varepsilon)!}\right),$$ where $\varepsilon$ and $\varepsilon'$ can be made as small as we like, and $r(\varepsilon)$ can be made arbitrarily large. Hence $$p_n=\frac{a_{n-1}}{A} \to \frac{1}{e^{e^{-1}}} \approx 0.6922\dots$$ as $n \to \infty$. \item Note that $$q_n=\frac{(n-1)a_{n-1}}{B}=\frac{(n-1)A}{B}\cdot \frac{a_{n-1}}{A} = \frac{p_n}{\mu_p(n)}.$$ By Theorem~\ref{T:main2}(1), $\mu_p(n)\to 1$ and, by part (1) of this theorem, $p_n \to e^{-e^{-1}}$. The result now follows immediately. \end{enumerate} \end{proof} An interesting consequence of our proof of part (2) of Theorem~\ref{T:main1} is that $p=q\mu_p(n)$ for any graph $G$, so $p<q$. Thus, if $G=C_n$ is a cycle, then the (uniform) density is approximately $\frac12$, so we immediately get the weighted probability that a random subtree spans is approximately twice the probability for the uniform case (although both probabilities approach 0 as $n \to \infty$). We state this observation as a corollary. \begin{cor}\label{C:pnqn} Let $G$ be any connected graph, let $p(G)$ and $q(G)$ be the uniform and weighted probabilities (resp.) that a random subtree is spanning, and let $\mu(G)$ be the uniform subtree density. Then $p(G)=q(G)\mu(G).$ \end{cor} If $\mathcal{G}$ is an infinite family of graphs, then we can interpret Cor.~\ref{C:pnqn} asymptotically. In this case, the two probabilities $p$ and $q$ coincide (and are non-zero) in the limit if and only if the density approaches 1. We conclude this section with a very short proof of part (2) of Theorem~\ref{T:main2}. \begin{proof}[Proof: Theorem~\ref{T:main2} (2)] We have $\mu_p(n)<\mu_q(n)<1$ for all $n$ by a standard argument in probability (see \cite{fg}). Since $\mu_p(n) \to 1$ as $n \to \infty$, we are done. \end{proof} \section{Conjectures, extensions and open problems}\label{S:future} We believe the study of subtrees in arbitrary graphs is a fertile area for interesting research questions. We outline several ideas that should be worthy of future study. Many of these topics are addressed in \cite{cgmv}. \begin{enumerate} \item {\bf Local statistics.} Compute ``local'' versions of the four statistics given here. Fix an edge $e$ in $K_n$. Then it is straightforward to compute $p_n', q_n', \mu_{p'}(n)$ and $\mu_{q'}(n)$, where each of these is described below. \begin{itemize} \item $p_n'$ is the (uniform) probability that a random subtree {\it containing the edge $e$} is a spanning tree. This is given by $$p_n'=\frac{n^{n-3}}{\sum_{k=0}^{n-2}{n-2 \choose k}(k+2)^{k-1}}.$$ The number of spanning trees containing a given edge is $2n^{n-3}$ -- this is easy to show using the tree-edge incidence matrix. Incidence counts can also show the total number of subtrees containing the edge $e$ is given by $\displaystyle{\sum_{k=0}^{n-2}2{n-2 \choose k}(k+2)^{k-1}}$. This can also be derived by using the {\it hyperbinomial transform} of the sequence of 1's. \item $q_n'$ is the (weighted) probability that a random subtree {\it containing the edge $e$} is a spanning tree. This time, we get $$q_n'=\frac{(n-1)n^{n-3}}{\sum_{k=0}^{n-2}{n-2 \choose k}(k+2)^{k-1}}.$$ \item $ \mu_{p'}(n)$ is the local (uniform) density, so $$\mu_{p'}(n)=\frac{\sum_{k=0}^{n-2}(k+1){n-2 \choose k}(k+2)^{k-1}}{(n-1)\sum_{k=0}^{n-2}{n-2 \choose k}(k+2)^{k-1}}.$$ \item $ \mu_{q'}(n)$ is the local (weighted) density, which gives $$\mu_{q'}(n)=\frac{\sum_{k=0}^{n-2}(k+1)^2{n-2 \choose k}(k+2)^{k-1}}{(n-1)\sum_{k=0}^{n-2}(k+1){n-2 \choose k}(k+2)^{k-1}}.$$ \end{itemize} It is not difficult to prove the same limits that hold for the global versions of these statistics also hold for the local versions: $p'_n$ and $q'_n$ both approach $e^{-e^{-1}}$ and $ \mu_{p'}(n)$ and $\mu_{q'}(n)$ both approach 1 as $n \to \infty$. (All of these can be proven by arguments analogous to the global versions.) For $p'_n$, however, the connection to the global statistics is even stronger: $p_n'=q_n$ for all $n$, i.e., the global weighted probability exactly matches the local uniform probability of selecting a spanning tree. (This can be proven by a direct calculation, but it is immediate from the ``average class size'' formulation of the weighted probability.) Other statistics that might be of interest here include larger local versions of these four: suppose we consider only subtrees that contain a given pair of adjacent edges, or a given subtree with 3 edges. Will the limiting probabilities match the ones given here? \item {\bf Non-spanning subtrees.} Explore the probabilities that a random subtree of $K_n$ has exactly $k$ edges for specified $k<n-1$. The analysis given here can be used to show the uniform and weighted probabilities of choosing a subtree with $n-2$ edges (one less than a spanning tree) approaches $\displaystyle{\left(e^{-1-e^{-1}}\right) =0.254646\dots.}$ This follows by observing the ratio $a_{n-1}/a_{n-2} \to e$ as $n \to \infty$. (It is interesting that both sequences are decreasing here, in contrast to the sequences $p_n$ and $q_n$.) \item {\bf Other graphs.} Explore these statistics for other classes of graphs. For instance, when $G=K_{n,n}$ is a complete bipartite graph, we get limiting values for the probability (uniform or weighted) that a random subtree spans is $$e^{-2e^{-1}} = 0.478965\dots,$$ the square of the limiting value we obtained for the complete graph. Both the uniform and weighted densities approach 1 in this case (forced by Cor.~\ref{C:pnqn}). On the other hand, consider the theta graph $\theta_{n,n}$, formed adding an edge ``in the middle'' of a $2n-2$ cycle (see Fig.~\ref{F:theta}). Then the uniform density approaches $\frac23$ as $n \to \infty$, and both the uniform and weighted probabilities that a random subtree spans tend to 0 as $n$ tends to $\infty$ \cite{cgmv}. \begin{figure}[h] \centering \begin{tikzpicture}[scale=2.2, auto,swap] \foreach \pos/\name in {{(-0.5,0.866)/a}, {(0.5,0.866)/b}, {(1,0)/c}, {(0.5,-0.866)/d}, {(-0.5,-0.866)/e}, {(-1,0)/f}} \node[vertex] (\name) at \pos {}; \foreach \source/ \dest in {a/b, b/c, c/d, d/e, e/f, f/a, f/c} \path[edge] (\source) edge (\dest); \end{tikzpicture} \caption{$\theta_{4,4}$.} \label{F:theta} \end{figure} We conjecture that, for any given graph, the number of edges determines whether or not these limiting densities are 0. \begin{conj} If $|E|=O(n^2)$, then the probability (either uniform or weighted) of selecting a spanning tree is non-zero (in the limit). \end{conj} \begin{conj} If $|E|=O(n)$, then the probability (either uniform or weighted) of selecting a spanning tree is zero (in the limit). \end{conj} \item {\bf Optimal graphs.} Determine the ``best'' graph on $n$ vertices and $m$ edges. There are many possible interpretations for ``best'' here: for instance, we might investigate which graph maximizes $p(G)$ and which maximizes the density? Must the graph that maximizes one of these statistics maximize all of them? This is closely related to the work in \cite{vw}, where extremal classes of trees are determined for uniform density. \item {\bf Subtree polynomial.} We can define a polynomial to keep track of the number of subtrees of size $k$. If $G$ is any graph with $n$ vertices, let $a_k$ be the number of subtrees of size $k$. Then define a {\it subtree polynomial} by $$s_G(x) = \sum_{k=0}^{n-1} a_k x^k.$$ We can compute the subtree densities directly from this polynomial and its derivatives: $$\mu_p(G) = \frac{s_G'(1)}{(n-1)s_G(1)} \hskip.2in \mbox{and} \hskip.2in \mu_q(G) = \frac{s_G'(1)+s''_G(1)}{(n-1)s'_G(1)}$$ It would be worthwhile to study the roots and various properties of this polynomial. In particular, we conjecture that the coefficients of the polynomial are unimodal. \begin{conj} The coefficients of $s_G(x)$ are unimodal. \end{conj} As an example of an interesting infinite 2-parameter family of graphs, the coefficients of the $\theta$-graph $s_{\theta_{a,b}}(x)$ are unimodal (see Fig.~\ref{F:thetacoef}). \begin{figure}[h] \centering \includegraphics[width=.5\textwidth]{gscatter} \caption{Coefficients of $\theta_{24,48}$, with three coefficient regions highlighted} \label{F:thetacoef} \end{figure} The mode of the sequence of coefficients gives another measure of the subtree density. For the $\theta$-graph $\theta_{n,n}$, we can show the mode is approximately $\sqrt{2}n$ (in \cite{cgmv}). Stating the unmodality conjecture in terms of a polynomial has the advantage of potentially using the very well developed theory of polynomials. Surveys that address unmodality of coefficients in polynomials include Stanley's classic paper \cite{st} and a recent paper of Pemantle \cite{pe}. In particular, if all the roots of $s_G(x)$ are negative reals, then the coefficients are unimodal. (Such polynomials are called {\it stable}.) Although $s_G(x)$ is not stable in general, perhaps it is for some large class of graphs. It is not difficult to find different graphs (in fact, different trees) with the same subtree polynomial. \begin{conj} For any $n$, construct $n$-pairwise non-isomorphic graphs that all share the same subtree polynomial. \end{conj} This is expected by pigeonhole considerations -- there should be far more graphs on $n$ vertices than potential polynomials. \item {\bf Density monotonicity.} Does adding an edge always increase the density? This is certainly false for disconnected graphs -- simply add an edge to a small component; this will lower the overall density. But we conjecture this cannot happen if $G$ is connected: \begin{conj} Suppose $G$ is a connected graph, and $G+e$ is obtained from $G$ by adding an edge between two distinct vertices of $G$. Then $\mu(G)<\mu(G+e)$. \end{conj} One consequence of this conjecture would be that, starting with a tree, we can add edges one at a time to create a complete graph, increasing the density at each stage. This could be a useful tool in studying optimal families. \item {\bf Matroid generalizations.} Instead of using subtrees of $G$, we could use {\it subforests}. This has the advantage of being well behaved under deletion and contraction. In particular, the total number of subforests of a graph $G$ is an evaluation of its Tutte polynomial: $T_G(2,1)$. The number of spanning trees is also an evaluation of the Tutte polynomial: $T_G(1,1)$ (In fact, this is how Tutte defined his {\it dichromatic} polynomial originally.) All of the statistics studied here would then have direct analogues to the subtree problem. This entire approach would then generalize to matroids. In this context, subforests correspond to independent sets in the matroid, and spanning trees correspond to bases. It would be of interest to study basis probabilities and densities for the class of binary matroids, for example. \end{enumerate}

https://arxiv.org/abs/1801.10418

Extremal values of the Sackin tree balance index

Tree balance plays an important role in different research areas like theoretical computer science and mathematical phylogenetics. For example, it has long been known that under the Yule model, a pure birth process, imbalanced trees are more likely than balanced ones. Also, concerning ordered search trees, more balanced ones allow for more efficient data structuring than imbalanced ones. Therefore, different methods to measure the balance of trees were introduced. The Sackin index is one of the most frequently used measures for this purpose. In many contexts, statements about the minimal and maximal values of this index have been discussed, but formal proofs have only been provided for some of them, and only in the context of ordered binary (search) trees, not for general rooted trees. Moreover, while the number of trees with maximal Sackin index as well as the number of trees with minimal Sackin index when the number of leaves is a power of 2 are relatively easy to understand, the number of trees with minimal Sackin index for all other numbers of leaves has been completely unknown. In this manuscript, we extend the findings on trees with minimal and maximal Sackin indices from the literature on ordered trees and subsequently use our results to provide formulas to explicitly calculate the numbers of such trees. We also extend previous studies by analyzing the case when the underlying trees need not be binary. Finally, we use our results to contribute both to the phylogenetic as well as the computer scientific literature by using the new findings on Sackin minimal and maximal trees in order to derive formulas to calculate the number of both minimal and maximal phylogenetic trees as well as minimal and maximal ordered trees both in the binary and non-binary settings. All our results have been implemented in the Mathematica package SackinMinimizer, which has been made publicly available.

\section{Introduction} Rooted trees, and binary ones in particular, play a fundamental role in many sciences as they can be used as a basis for search algorithms as well as, amongst others, as a model for evolution. In many cases where these trees occur, probability distributions are not always uniform concerning the degree of tree balance -- for instance, the Yule model in phylogenetics, which is a pure birth process, has long been known to lead to more imbalanced trees. A simple example is depicted in Figure \ref{yule}, where it can be seen that if all leaves of a tree with three leaves are equally likely to give rise to a new leaf, then two of them lead to the same (`imbalanced') tree, whereas the other possible tree (the `balanced' one) occurs only once. So in order to understand such processes and their possible bias towards imbalanced (or, in other cases, balanced) trees, one has to be able to classify the degree of balance in more detail than just in a binary way (`balanced' versus `imbalanced'). Therefore, various balance indices were introduced and have been used over the years, e.g. \citep{blum,colless,cophenetic,sackin,steel2016}. One of the most frequently used and discussed such indices is the Sackin index \cite{sackin}. \begin{figure} \centering \includegraphics[scale=.3]{yule.eps} \caption{The Yule process splits one leaf at a time uniformly at random to form a so-called cherry. It can be easily seen that this leads to a tree shape bias already when there are $n=4$ leaves. This is due to the fact that the only rooted binary tree on three leaves has two leaves that give rise to the tree on the left, which is considered `imbalanced' (it is the so-called caterpillar tree $T^{cat}_4$ on 4 leaves), whereas only one leaf leads to the so-called fully balanced tree $T_2^{bal}$ of height 2. } \label{yule} \end{figure} This index has been observed to have some very nice properties -- for instance, it has been stated that its maximum is achieved by the caterpillar tree (the tree with only one `cherry', i.e. with only one internal node whose two descendants are both leaves) and that, whenever the number $n$ of leaves equals a power of 2, i.e. when $n=2^k$ for some $k$, the minimum is achieved by the so-called fully balanced tree of height $k$, i.e. the tree in which all leaves have distance $k$ to the root \cite{shao}. However, while these statements can be found in the literature, rigorous proofs of them are nowhere to be found. It is one aim of this manuscript to present proofs and thus to show that the Sackin index indeed has these properties. Note that these properties are desirable for a good tree balance index, as the caterpillar tree is normally perceived as very `imbalanced', whereas the fully balanced tree is normally referred to as very `balanced' (which explains its name). The idea of the Sackin index is to assign a small number to trees that are perceived as balanced and a high number to more imbalanced trees -- i.e. the higher the Sackin index, the more imbalanced the tree. So while some statements on the maximum of the Sackin index and its minimum in the special case $n=2^k$ can already be found in the literature, even if without proofs, little is known about trees with the minimum Sackin index for $n\neq 2^k$. Moreover, while it has long been known that the tree achieving this minimum in such cases need not be unique, the number of most balanced or most imbalanced trees has never been formally investigated. The main aim of this manuscript is to state and rigorously prove the following statements: \begin{enumerate} \item For all leaf numbers $n$, the caterpillar is the unique tree with maximal Sackin index. \item For $n \in \{2^m-1,2^m,2^m+1\}$ ($m \in \mathbb{N}$), the minimum of the Sackin index is achieved by a unique tree. If $n=2^m$, this tree is the fully balanced tree. \item If there is {\bf no} $m \in \mathbb{N}$ such that $n \in \{2^m-1,2^m,2^m+1\}$, then the minimum of the Sackin index is achieved by more than one tree, i.e. it is not unique. \end{enumerate} Moreover, we will present an algorithm which constructs {\em all} trees with minimal Sackin index. Last but not least, we use this algorithm to derive a recursive formula for the number of such trees, which leads to a sequence that has only now been submitted to the Online Encyclopedia of Integer Sequences, i.e. it has not occurred elsewhere in the literature before. \section{Preliminaries} Before we can start to discuss the Sackin tree balance index, we first need to introduce all concepts used in this manuscript. We start with trees: {\em Trees} are connected, acyclic graphs with node set $V$ and edge set $E$. We use $V^1$ in order to denote the set of {\em leaves} of a tree, i.e. the set of nodes of degree at most 1. All nodes $v$ that are not leaves, i.e. $v\in V\setminus V^1$, are called {\em internal nodes}. The set of internal nodes of a tree $T$ will be denoted by $\mathring{V}(T)$, or, whenever there is no ambiguity, simply by $\mathring{V}$. All trees $T$ in this manuscript are assumed to be {\em rooted} and {\em binary}, i.e. if they have an internal node at all, they have one root node $\rho$ of degree 2 and all other internal nodes have degree 3. The only rooted binary tree which does not have an internal node is the tree that consists of only one node and no edge -- in this special case, the only node is for technical reasons at the same time defined to be the root and the only leaf of the tree, so it is the only case where the root is not an internal node. Furthermore, for technical reasons all tree edges in this manuscript are implicitly assumed to be {\em directed} from the root to the leaves. Thus, for an edge $e=(u,v)$ of $T$, it makes sense to refer to $u$ as the {\em direct ancestor } or {\em parent } of $v$ (and $v$ as the {\em direct descendant} or {\em child} of $u$). More generally, when there is a directed path from $\rho$ to $v$ employing $u$, $u$ is called {\em ancestor} of $v$ (and $v$ descendant of $u$). Two leaves $v$ and $w$ are said to form a {\em cherry}, denoted by $[v,w]$, if $v$ and $w$ have the same parent, i.e. if there exists an internal node $u$ in $V$ such that $(u,v)$ and $(u,w)$ are edges in $E$. Note that every rooted binary tree with at least 2 leaves has at least one cherry. Let $T$ be a rooted binary tree with root $\rho$, and let $x \in V^1$ be a leaf of $T$. Then we denote by $\delta_x$ the {\em depth} of $x$ in $T$, which is the number of edges on the unique shortest path from $\rho$ to $x$. Then, the {\em height} of $T$ is defined as $h(T)=\max\limits_{x \in V^1} \delta_x$, i.e. as the maximum of these distances. Note that whenever a leaf $v$ has maximal depth, i.e. whenever $h(T)=\delta_v$, $v$ is element of a cherry. This is due to the fact that if the other direct descendant, say $w$, of the parent of $v$, say $u$, was not a leaf but an internal node, it would have descending nodes of a greater depth than $\delta_v=\delta_w$, which would contradict the maximality of $\delta_v$. This is why in this manuscript, instead of considering both $v$ and $w$ separately as leaves of maximal depth, we sometimes refer to a cherry $[v,w]$ as {\em cherry of maximal depth}. Moreover, recall that a rooted binary tree $T$ can be decomposed into its two maximal pending subtrees $T_a$ and $T_b$ rooted at the direct descendants $a$ and $b$ of $\rho$, and we denote this by $T=(T_a,T_b)$. Last but not least, we want to introduce two particular trees which play a crucial role in this manuscript, namely the so-called {\em caterpillar tree} $T^{cat}_n$ and the so-called {\em fully balanced tree} $T^{bal}_k$, respectively. $T^{cat}_n$ denotes the unique rooted binary tree with $n$ leaves that has only one cherry, while $T^{bal}_k$ denotes the unique tree with $n=2^k$ leaves in which all leaves have depth precisely $k$. Whenever there is no ambiguity concerning $n$ or $k$, we also write $T^{cat}$ and $T^{bal}$ instead of $T^{cat}_n$ and $T^{bal}_k$. $T^{cat}_4$ and $T^{bal}_2$ are depicted in the bottom row of Figure \ref{yule}. Note that without loss of generality, the unique rooted binary tree with only one leaf, which consists of only one node and no edges, is defined to be $T^{cat}_1$, and it is thus the only caterpillar tree which does not contain a cherry. This tree is at the same time equal to $T^{bal}_0$, i.e. the fully balanced tree of height 0. This technicality enables inductive proofs concerning $T^{cat} $ and $T^{bal}$ to start at $n=1$. We are now in a position to define the central concept of this manuscript, namely the Sackin index. As there are four different versions of this index to be found in the literature, we will define all of them first and subsequently investigate their respective relationships. Note, however, that we focus on the first two definitions in the present manuscript, which can be shown to be equivalent. Therefore, we will not refer to the other two definitions as Sackin index, but give them modified names instead. \begin{definition}\label{sackin} \cite{cophenetic,steel2016} The {\em Sackin index} of a rooted binary tree $T$ is defined as $\mathcal{S}(T)=\sum\limits_{u \in \mathring{V}(T)}n_u$, where $n_u$ denotes the number of leaves in the subtree of $T$ rooted at $u$. \end{definition} Note that in the following, whenever we have for two trees $T_1$ and $T_2$ that $\mathcal{S}(T_1) < \mathcal{S}(T_2)$, then $T_1$ is called {\em more balanced} than $T_2$. \begin{definition}\label{sackinalternative} \cite{cophenetic,steel2016} The {\em Sackin index} of a rooted binary tree $T$ is defined as $\bar{\mathcal{S}}(T)=\sum\limits_{x \in V^1(T)}\delta_x$. \end{definition} \begin{definition}\citep[(3.10)]{steel2016}\label{mike} The {\em Sackin index} of a rooted binary tree $T$ with root $\rho$ is defined as $\widetilde{\mathcal{S}}(T)=\sum\limits_{u \in V(T)\setminus\{\rho\}} n_u$, where $n_u$ denotes the number of leaves in the subtree of $T$ rooted at $u$. \end{definition} Note that in the original paper by Sackin \citep{sackin}, in fact no index is defined at all. Instead, a sequence $b$ of leaf depths is defined, which implies that Definition \ref{sackinalternative} is probably most closely related to what Sackin originally intended. However, we will show in Lemma \ref{equidefs} that the first three definitions are in fact equivalent, which does not hold for the following definition. \begin{definition}\label{noah}\cite{noah} The {\em normalized Sackin index} of a rooted binary tree $T$ with $n$ leaves is defined as $\widehat{\mathcal{S}}(T)=\frac{1}{n}\sum\limits_{x \in V^1(T)}\delta_x$, where $\delta_x$ denotes the depth of leaf $x$. \end{definition} We now state a first lemma, which has already been partially stated (albeit without proof) in the literature \cite{cophenetic}. \begin{lemma}\label{equidefs} Definitions \ref{sackin}, \ref{sackinalternative} and \ref{mike} are equivalent, i.e. for any rooted binary tree $T$, we have $\mathcal{S}(T)=\bar{\mathcal{S}}(T)=\widetilde{\mathcal{S}}(T)$. \end{lemma} \begin{proof} We first prove $\mathcal{S}(T)=\bar{\mathcal{S}}(T)$. Therefore, consider $\bar{\mathcal{S}}(T)=\sum\limits_{x \in V^1(T)}\delta_x$. Note that while $\delta_x$ by definition denotes the number of edges separating leaf $x$ from the root $\rho$ of $T$, this is equivalent to the number of internal nodes on the path from $\rho$ to $x$ (including $\rho$). This leads to: $$\bar{\mathcal{S}}(T)=\sum\limits_{x \in V^1(T)}\delta_x= \sum\limits_{x \in V^1(T)} |\{u \in \mathring{V}(T): \mbox{$u$ is an ancestor of $x$}\}|$$ $$ = |\{(x,u): \mbox{ $x \in V^1(T)$, $u \in \mathring{V}(T)$ and $u$ is an ancestor of $x$}\}|$$ $$ = |\{(x,u): \mbox{ $x \in V^1(T)$, $u \in \mathring{V}(T)$ and $x$ is a descendant of $u$}\}|$$ $$= \sum\limits_{u \in \mathring{V}(T)} |\{x \in V^1(T): \mbox{$x$ is a descendant of $u$}\}| =\sum\limits_{u \in \mathring{V}(T)} n_u=\mathcal{S}(T).$$ So now we have $\mathcal{S}(T)=\bar{\mathcal{S}}(T)$, and next we show that $\mathcal{S}(T)=\widetilde{\mathcal{S}}(T)$. We have $$\mathcal{S}(T) = \sum\limits_{u \in \mathring{V}(T)}n_u = \sum\limits_{u \in V(T)}n_u - \sum\limits_{x \in V^1(T)}n_x = \left(\sum\limits_{u \in V(T)}n_u \right)-n.$$ The latter equality is due to the fact that the only leaf belonging to a subtree rooted at a leaf is the leaf itself, so as we have $n$ leaves, this gives $n$ summands that contribute 1 to the sum. Now recall that $n_\rho = n$, so this leads to $$\mathcal{S}(T) =\left( \sum\limits_{u \in V(T)}n_u\right) -n_\rho = \sum\limits_{u \in V(T)\setminus\{\rho\}}n_u=\widetilde{\mathcal{S}}(T).$$ This completes the proof. \end{proof} So because Definitions \ref{sackin}, \ref{sackinalternative} and \ref{mike} are equivalent, we do not have to distinguish between them and use them interchangeably. In fact, we will focus on in this manuscript mainly on the first two definitions. The normalized Sackin index, however, is a modification of the Sackin index whenever trees with different numbers of leaves are considered, because the ranking induced by the normalized Sackin index can even reverse the ranking induced by the Sackin index. For instance, consider the two trees $T_1=T_{37}^{cat}$, i.e. the caterpillar tree with 37 leaves, and $T_2=T_{9}^{bal}$ the fully balanced tree with $2^9=512$ leaves. Then, it can easily be verified that we have $\mathcal{S}(T_1)=702 < 4608=\mathcal{S}(T_2)$, but $\widehat{\mathcal{S}}(T_1)\approx 18.97 > 9 =\widehat{\mathcal{S}}(T_2)$.\footnote{Note that these numbers can also be verified later on by using Propositions \ref{sackinCat} and \ref{sackinBal}.} In fact, this ranking modification is no artifact but the very purpose of the normalization: The effect that many leaves automatically may lead to more `imbalance' shall be eliminated. So in fact, $\mathcal{S}$ and $\widehat{\mathcal{S}}$ can be very different -- but only when different leaf numbers are considered! As long as $n$ is fixed, the induced rankings of the two indices are of course equivalent, and in this case, $\widehat{\mathcal{S}}$ is just $\mathcal{S}$ divided by the constant factor $n$. So when we discuss for instance the question how many trees with $n$ leaves exist that have maximal or minimal Sackin index, the answers for $\mathcal{S}$ and $\widehat{\mathcal{S}}$ will be the same. Therefore, as this is sufficient for the number of minima and maxima, we focus in this manuscript on Definitions \ref{sackin} and \ref{sackinalternative}. \section{Results} It is the main aim of this manuscript to fully characterize trees with minimal and maximal Sackin index, respectively, and to count such trees. We will start with the easier case, which is the maximum, before we consider the more involved and therefore more interesting case of the minimum. \subsection{Maximally imbalanced trees / Minimally balanced trees} \par In this section, we present an upper bound for $\mathcal{S}(T)$ and prove that this bound is tight as it is achieved by $T^{cat}$. Moreover, we will show that for all values of $n$, $T^{cat}_n$ is even the unique tree maximizing $\mathcal{S}$, i.e. the unique most imbalanced tree. This result has been stated in the literature before, e.g. in \cite{cophenetic}, but so far, a formal proof has not been stated anywhere. However, we start by establishing the upper bound on $\mathcal{S}(T)$ before we can proceed as explained. \begin{theorem}\label{maxsackin} Let $T$ be a rooted binary tree with $n$ leaves. Then, we have: $$\mathcal{S}(T)\leq \frac{n\cdot(n+1)}{2}-1.$$ \end{theorem} Before we can prove this theorem, we need one more lemma. \begin{lemma} \label{cherrydeleted}Let $T$ be a rooted binary tree with $n\geq 2$ leaves. Let $[u,v]$ be a cherry of maximal depth in $T$ with parent $w$, and let $\widetilde{T}$ be the tree derived by $T$ by deleting $u$ and $v$ as well as the edges $(w,u)$ and $(w,v)$. Then, for the Sackin indices of $T$ and $\widetilde{T}$ we have: $$ \mathcal{S}(T)=\mathcal{S}(\widetilde{T})+\delta_u+1.$$ \end{lemma} \begin{proof} By Definition \ref{sackinalternative} and Lemma \ref{equidefs}, we have $\mathcal{S}(T)=\sum\limits_{x \in V^1}\delta_x$. By construction of $\widetilde{T}$, as $u$ and $v$ have been deleted and $w$ is a leaf in $\widetilde{T}$ but was an internal node in $T$, we have $\mathcal{S}(\widetilde{T})= \mathcal{S}(T)-\delta_u-\delta_v+\delta_w$. Note that as $u$ and $v$ form a cherry, we have $\delta_u=\delta_v$, and as $w$ was the parent of $u$ and $v$ in $T$, we have $\delta_w=\delta_u-1$. Thus, altogether we get $\mathcal{S}(\widetilde{T})= \mathcal{S}(T)-2\delta_u+(\delta_u-1)=\mathcal{S}(T)-\delta_u-1.$ Rearranging the latter term yields the desired result. \end{proof} Now we can use Lemma \ref{cherrydeleted} to prove Theorem \ref{maxsackin}. \begin{proof}[Proof of Theorem \ref{maxsackin}] We prove the statement by induction on $n$. For $n=1$ there is only one tree $T$, which consists of only one leaf and has no internal nodes. Thus, by Definition \ref{sackin}, $\mathcal{S}(T)=0$ (as the sum in the definition of $\mathcal{S}(T)$ is empty). Moreover, $ 0=\frac{1\cdot2}{2}-1= \frac{n\cdot(n+1)}{2}-1$, which completes the base case of the induction. Now we assume that for all trees with $n$ leaves the claim is proven and consider a tree $T$ with $n+1$ leaves. It remains to show $\mathcal{S}(T)\leq \frac{(n+1)(n+2)}{2}-1$. We construct a tree $\widetilde{T}$ by taking a cherry $[u,v]$ of $T$ of maximal depth $\delta_u=\delta_v$ and deleting $u$ and $v$ as well as the edges leading from $u$ and $v$ to their direct ancestor, say $w$. Thus, $\widetilde{T}$ has $(n+1)-2+1=n$ leaves, because leaves $u$ and $v$ have been deleted, but $w$, which is an internal node in $T$, is a leaf in $\widetilde{T}$. By Lemma \ref{cherrydeleted}, we have $\mathcal{S}(T)=\mathcal{S}(\widetilde{T})+\delta_u+1$. We use the inductive hypothesis for $\widetilde{T}$ and derive: $$\mathcal{S}(T)=\mathcal{S}(\widetilde{T})+\delta_u+1\leq \frac{n(n+1)}{2}-1 + \delta_u +1= \frac{n(n+1)}{2} + \delta_u . $$ Note that for all leaves, including those of maximal depth, their depths are bounded by one less than the number of leaves (cf. Lemma \ref{maxdepth} in the appendix). We use this to conclude that $\delta_u \leq (n+1)-1=n$. Therefore, we have: $$\mathcal{S}(T)\leq \frac{n(n+1)}{2} + n =\frac{n^2+3n}{2} = \frac{n^2+3n}{2}+\frac{2}{2} -1=\frac{(n+1)(n+2)}{2}-1.$$ This completes the proof. \end{proof} We now consider the caterpillar tree $T^{cat}$ and show that it achieves the bound induced by Theorem \ref{maxsackin}, which implies that this bound is indeed tight. \begin{proposition}\label{sackinCat} Let $T^{cat}=T^{cat}_n$ be the caterpillar tree with $n$ leaves. Then, we have: $\mathcal{S}(T^{cat})=\frac{n\cdot(n+1)}{2}-1$, i.e. $T^{cat}$ takes on the upper bound provided by Theorem \ref{maxsackin}. \end{proposition} \begin{proof} If $n=1$, $T^{cat}$ consists only of a leaf and has no internal nodes, so the sum over all internal nodes in Definition \ref{sackin} is empty, which implies $\mathcal{S}(T^{cat})=0$. On the other hand, in this case we have $\frac{n\cdot(n+1)}{2}-1= \frac{1\cdot(1+1)}{2}-1=0=\mathcal{S}(T^{cat})$. This completes the case $n=1$. Now consider the case $n\geq 2$. By Definition \ref{sackinalternative}, we have $\mathcal{S}(T^{cat})=\sum\limits_{x \in V^1(T^{cat})}\delta_x$. But note that in $T^{cat}$ there is precisely one leaf of depth 1, one leaf of depth 2 and so forth. In the unique cherry of $T^{cat}$, both leaves have depth $n-1$. So in total, we have: $$\mathcal{S}(T^{cat})=\sum\limits_{x \in V^1(T^{cat})}\delta_x = 1+2+ \ldots + (n-1) + (n-1) = \left(\sum\limits_{i=1}^{n-1}i \right) +(n-1).$$ Using the Gaussian sum, this leads to: $$\mathcal{S}(T^{cat})= \frac{(n-1)n}{2} +(n-1)=\frac{n^2-n}{2} +\frac{2n-2}{2}= \frac{n^2+n-2}{2}=\frac{n(n+1)}{2}-1.$$ This completes the proof. \end{proof} So by Proposition \ref{sackinCat} we know that the caterpillar tree assumes the maximal Sackin index for all possible leaf numbers $n$. However, Proposition \ref{sackinCat} does not make a statement about whether there exist other trees with maximal Sackin index. We will now show that this is not the case, i.e. the caterpillar tree is the unique most imbalanced tree. \begin{theorem}\label{catisunique} Let $n \in \mathbb{N}_{\geq 1}$. Then, there is only one rooted binary tree $T$ with maximal Sackin index, i.e. with $\mathcal{S}(T)= \frac{n(n+1)}{2}-1$, and we have $T=T^{cat}$. Thus, the caterpillar tree is the unique tree maximizing the Sackin index. \end{theorem} Before we can prove this theorem, we need one more lemma. \begin{lemma}\label{optStandardDecomp} Let $T$ be a rooted binary tree with $n$ leaves and with maximal (or minimal) Sackin index for $n$, i.e. for all other trees $\widetilde{T}$ with $n$ leaves we have $\mathcal{S}(T) \geq \mathcal{S}(\widetilde{T})$ (or $\mathcal{S}(T) \leq \mathcal{S}(\widetilde{T})$, respectively). Let $T=(T_a,T_b)$ be the standard decomposition of $T$ into its two maximal pending subtrees $T_a$ and $T_b$, with $n_a$ and $n_b$ leaves, respectively. Then, $\mathcal{S}(T_a)$ and $\mathcal{S}(T_b)$ are maximal (minimal) for $n_a$ and $n_b$, respectively. \end{lemma} \begin{proof} We consider the case of maximality. The case of minimality can be shown analogously. Now, assume that $\mathcal{S}(T)$ is maximal. Using Definition \ref{sackin}, it is easy to see that $\mathcal{S}(T)=\mathcal{S}(T_a)+\mathcal{S}(T_b)+n$. Now assume that $\mathcal{S}(T_a)$ is not maximal, i.e. there is a tree $\widehat{T}$ with $n_a$ leaves such that $\mathcal{S}(\widehat{T})>\mathcal{S}(T_a)$. Then we can construct a tree $\widetilde{T}$ on $n$ leaves such that $\widetilde{T}=(\widehat{T},T_b)$, i.e. we replace $T_a$ in $T$ by $\widehat{T}$ to derive $\widetilde{T}$. Now for $\widetilde{T}$ we have by Definition \ref{sackin}: $\mathcal{S}(\widetilde{T}) = \mathcal{S}(\widehat{T}) + \mathcal{S}(T_b)+n > \mathcal{S}(T_a)+\mathcal{S}(T_b)+n=\mathcal{S}(T)$. This contradicts the maximality of $\mathcal{S}(T)$, which implies that the assumption was wrong. So $\mathcal{S}(T_a)$ has to be maximal, and analogously, $\mathcal{S}(T_b)$ has to be maximal, too. This completes the proof. \end{proof} We are now in a position to prove Theorem \ref{catisunique}. \begin{proof}[Proof of Theorem \ref{catisunique}] By Proposition \ref{sackinCat}, we have $\mathcal{S}(T^{cat})= \frac{n(n+1)}{2}-1$, which is maximal according to Theorem \ref{maxsackin}. Now assume there is another tree $T$ with $\mathcal{S}(T)= \frac{n(n+1)}{2}-1$. We prove by induction on $n$ that then $T$ must equal $T^{cat}$. For $n=1$, there is only one rooted binary tree, which is by definition a caterpillar, so there is nothing to show. This completes the base case of the induction. Next we assume that the statement is already proven for all leaf numbers up to $n-1$ and consider a rooted binary tree $T$ with $n$ leaves and $\mathcal{S}(T)= \frac{n(n+1)}{2}-1$. Note that without loss of generality, $n\geq 2$ (else we consider the base case of the induction again). As $n\geq 2$, we can consider the standard decomposition of $T$ into its two maximal pending subtrees $T_a$ and $T_b$ with $n_a$ and $n_b$ leaves, respectively. Note that $n=n_a+n_b$ and that we may assume without loss of generality that $n_a\geq n_b$. Moreover, recall that we have $\mathcal{S}(T)=\mathcal{S}(T_a)+\mathcal{S}(T_b)+n$ by Definition \ref{sackin}. By Lemma \ref{optStandardDecomp}, $\mathcal{S}(T_a)$ and $\mathcal{S}(T_b)$ are also maximal (as $\mathcal{S}(T)$ is maximal), and thus, by induction, $T_a$ and $T_b$ are both caterpillars. So we can conclude by Proposition \ref{sackinCat} that $\mathcal{S}(T_a)= \frac{n_a(n_a+1)}{2}-1$ and $\mathcal{S}(T_b)= \frac{n_b(n_b+1)}{2}-1$, which gives $$\mathcal{S}(T)=\mathcal{S}(T_a)+\mathcal{S}(T_b)+n = \frac{n_a(n_a+1)}{2}-1+\frac{n_b(n_b+1)}{2}-1+n. $$ Using $n_b=n-n_a$ and expanding all terms leads to \begin{equation}\label{Sfterm}\mathcal{S}(T)= \frac{1}{2}n^2-n\cdot n_a+n_a^2+\frac{3}{2}n-2.\end{equation} We now consider this term as a function of $n_a$ and analyze the values of this function more in-depth, i.e. we consider $f(n_a)=\frac{1}{2}n^2-n\cdot n_a+n_a^2+\frac{3}{2}n-2$. Note that as we have $n_a+n_b=n$ and as we assume without loss of generality that $n_a \geq n_b$, we have $n_a\geq \lceil\frac{n}{2}\rceil$ (and $n_b\leq \lfloor\frac{n}{2}\rfloor$). Moreover, note that $n_a\leq n-1$, because $n_b\geq 1$. We now show that for $n_a \in \left\{\lceil\frac{n}{2}\rceil,\ldots, n-1\right\}$, $f(n_a)$ is strictly monotonically increasing. Therefore, consider the first derivative of $f$: $f'(n_a)= -n+2n_a$, which equals 0 precisely if $n_a=\frac{n}{2}$ and is strictly larger than 0 for all $n_a>\frac{n}{2}$. In total, this implies that we have a unique minimum at $\frac{n}{2}$ and that $f$ strictly increases after $n_a$ passes this minimum. So indeed, $f$ is strictly monotonically increasing on $\left\{\lceil\frac{n}{2}\rceil,\ldots, n-1\right\}$, which implies that its unique maximum is assumed at $n-1$. This implies that $\mathcal{S}(T)$ is maximal if $n_a=n-1$, i.e. if $T_a$ is the caterpillar on $n-1$ leaves. Using $n=n_a+n_b$, this implies that $n_b=1$. So in total, $T_a$ is a caterpillar on $n-1$ leaves and $T_b$ consists of only one leaf. This implies that $T$ is a caterpillar on $n$ leaves, which completes the proof. \end{proof} \par In total, we conclude that the bound provided by Theorem \ref{maxsackin} is tight, as Proposition \ref{sackinCat} implies that for each $n$, the caterpillar reaches this bound, and by Theorem \ref{catisunique} we know that the caterpillar is unique with this property, i.e. the caterpillar is the unique tree with maximal Sackin index. So for all $n$, there is precisely one tree with maximal Sackin index. We will show in the following section that this is different for the minimal Sackin index, as it can be taken on by various trees (depending on $n$). We conclude this section by noting that the sequence $(a_n)_{n \in \mathbb{N}_{\geq 1}}$ with $a_n=\mathcal{S}(T^{cat}_n)=\frac{n(n+1)}{2}-1$ for $i \in \mathbb{N}\geq 1$, which starts with 2, 5, 9, 14, 20, 27, 35, 44, 54, 65, 77, 90, 104, 119, 135, 152, 170, 189, 209, $\ldots$, corresponds to sequence A000096 in the Online Encyclopedia of Integer Sequences OEIS \cite[Sequence A000096]{OEIS}, when the index is shifted by 1 (i.e. the $i$\textsuperscript{th} entry of the OEIS sequence corresponds to the $(i+1)$\textsuperscript{st} entry of our sequence). So this sequence has already occurred in other contexts, which might link the maximal Sackin index to other areas of research like the study of prime polyominoes or the traveling salesman polytope \cite[Sequence A000096]{OEIS}. \subsection{Maximally balanced trees} \par In this section, it is our aim to achieve the same results for trees with minimal Sackin index that the previous section stated for trees with maximal Sackin index. In particular, we want to find a tight bound for the minimal Sackin index and we want to characterize the trees that achieve it. However, it turns out that -- as opposed to the previous section -- the case of minimality is far more involved. In fact, depending on the number $n$ of leaves, the tree with the minimal Sackin index need not be unique, so counting these trees is more complicated. Note that while examples for the fact that the Sackin index can be minimized by more than one tree have been presented before, e.g. in \cite{cophenetic}, see Figure \ref{6taxabalanced}, it has so far not been investigated for which values of $n$ this happens and how many minima there are. It is the main aim of this section to present an algorithm that is able to systematically generate all trees with minimal Sackin index and that therefore also leads to a recursive formula for counting such trees. \begin{figure} \centering \includegraphics[scale=.4]{sackin6taxamin.eps} \caption{Two trees with $n=6$ leaves which both have Sackin index 16, which can be shown to be minimal for $n=6$. } \label{6taxabalanced} \end{figure} However, we start as in the previous section and state the first result, which provides a lower bound on the Sackin index of trees with $n$ leaves. \begin{theorem}\label{minsackin} Let $T$ be a rooted binary tree with $n$ leaves. Let $k=\lceil \log_2(n)\rceil$. Then, $\mathcal{S}(T) \geq -2^k+n(k+1)$. \end{theorem} \begin{proof} Note that as $k = \lceil \log_2(n)\rceil$, we have $n\leq 2^k$, and $k$ is the smallest integer with this property. In other words, we have $k=\min\{\widehat{k}: n\leq 2^{\widehat{k}}\}$. We now prove the theorem by induction on $n$. For $n=1$, we have $k=0$, as $1 \leq 2^0$. So the term on the right hand side equals $-2^k+n(k+1) = -2^0+1\cdot (0+1)=-1+1=0$. As $\mathcal{S}(T) \geq 0$ for all $T$ by the definition of the Sackin index, this condition is clearly fulfilled. This completes the base case of the induction. Let us now consider a tree $T$ with $n$ leaves. We assume that for all trees with $n-1$ leaves, the statement already holds, and we have to show that now $\mathcal{S}(T) \geq -2^k+n(k+1)$. We consider a cherry $[u,v]$ of $T$ of maximal depth $\delta_u=\delta_v$ and its parent node $w$. We consider tree $\widetilde{T}$ which we derive from $T$ as follows: We remove leaves $u$ and $v$ together with the edges $(w,u)$ and $(w,v)$. By Lemma \ref{cherrydeleted}, this leads to $\mathcal{S}(T) =\mathcal{S}(\widetilde{T})+ \delta_u +1$. Note that as $\widetilde{T}$ was derived from $T$ by deleting two leaves ($u$ and $v$) but creating one new one ($w$), for the number of leaves $\widetilde{n}$ of $\widetilde{T}$ we have $\widetilde{n}=n-1$. Let $\widetilde{k} = \lceil \log_2(\widetilde{n})\rceil=\min\{\widehat{k}: \widetilde{n}\leq 2^{\widehat{k}}\}$. Then there are two cases: either $\widetilde{k}=k$ or $\widetilde{k}=k-1$, as the deletion of one leaf might result in a smaller power of 2 being necessary to cover all leaves, but it cannot decrease by more than 1. \begin{enumerate} \item We first consider the case $\widetilde{k}=k-1$. This case implies that the deletion of one leaf leads to a smaller power of 2 necessary to cover $n-1$ than $n$. This is precisely the case if $n=2^{k-1}+1>2^{k-1}=2^{\widetilde{k}}$, because then $n-1=2^{k-1}=2^{\widetilde{k}}$, and in this case, the smallest $\widehat{k}$ such that $2^{\widehat{k}}$ is at least $n$ would be $(k-1)+1=k$. So now we have $\mathcal{S}(T) =\mathcal{S}(\widetilde{T})+ \delta_u +1$ and, by the inductive hypothesis, $\mathcal{S}(\widetilde{T}) \geq -2^{\widetilde{k}}+(n-1)(\widetilde{k}+1)= -2^{k-1}+2^{k-1} ((k-1)+1)=-2^{k-1}+2^{k-1}k= 2^{k-1}(k-1)$. Combining these two observations, we derive \begin{equation} \label{case2} \mathcal{S}(T) = \mathcal{S}(\widetilde{T})+ \delta_u +1 \geq 2^{k-1}(k-1) + \delta_u + 1.\end{equation} It remains to show that the latter term in Equation \ref{case2} is at least $-2^k+n(k+1)$. Note that $\delta_u$ is the maximal depth of $T$, because $u$ is an element of a cherry of maximal depth. It can easily be seen that as $n=2^{k-1}+1$, we have $\delta_u \geq k$ (see Lemma \ref{maxdepthleaf} in the appendix). So this leads to: \begin{equation}\label{deltabound} \delta_u \geq k = -2^k+2^{k-1}\left( k+1-(k-1)\right)+k\end{equation} Now we combine Equations (\ref{case2}) and (\ref{deltabound}) to derive the desired result: $$\mathcal{S}(T) \geq 2^{k-1}(k-1) -2^k+2^{k-1}\left( k+1-(k-1)\right)+k+1=-2^k+n(k+1).$$ This completes the proof of this case. \item Now we consider the case $\widetilde{k}=k$. In this case, we have $n \leq 2^k$ and $n-1 > 2^{k-1}$. This implies that $n>2^{k-1}+1$ and thus, again by Lemma \ref{maxdepthleaf}, we have $\delta_u \geq k$. Moreover, we still have $\mathcal{S}(T)=\mathcal{S}(\widetilde{T})+\delta_u+1$, and by the inductive hypothesis we have that $\mathcal{S}(\widetilde{T}) \geq -2^k+(n-1) \cdot k +(n-1)$. Combining all these observations leads to: \begin{equation} \label{case1} \mathcal{S}(T) =\mathcal{S}(\widetilde{T})+\delta_u+1 \geq -2^k+(n-1) \cdot k +(n-1) +k +1 =-2^k+n(k+1). \end{equation} This completes the proof. \end{enumerate} \end{proof} We now consider the so-called fully balanced tree $T^{bal}_k$ and show that its name is indeed justified in the sense that it achieves the bound induced by Theorem \ref{minsackin} whenever it is defined, i.e. whenever the number $n$ of leaves equals $2^k$. \begin{proposition}\label{sackinBal} Let $T^{bal}_k$ be the fully balanced tree of height $k$ with $n=2^k$ leaves. Then, we have: $S(T^{bal}_k)=k\cdot 2^k$. \end{proposition} \begin{proof} By Definition \ref{sackinalternative} and Lemma \ref{equidefs}, we have $S(T^{bal}_k)=\sum\limits_{x \in {V^1}(T)}\delta_x$. Note that $|V^1|=n=2^k$ and that each of the $2^k$ leaves in $T^{bal}_k$ has depth $k$. This leads to $S(T^{bal}_k)=k\cdot 2^k$, which completes the proof. \end{proof} \begin{corollary}\label{balmaxbal} Let $T^{bal}_k$ be the fully balanced tree of level $k$ with $n=2^k$ leaves. Then, $T^{bal}_k$ is maximally balanced, i.e. for all rooted binary trees $T$ with $n=2^k$ leaves we have: $\mathcal{S}(T)\geq S(T^{bal}_k)$. \end{corollary} \begin{proof} By Proposition \ref{sackinBal} we have $S(T^{bal}_k)=k\cdot 2^k$. Let $T$ be a rooted binary tree with $n=2^k$ leaves. By Theorem \ref{minsackin}, we then have: $$\mathcal{S}(T) \geq -2^k+n(k+1) = -2^k+\underbrace{2^k}_{=n}(k+1) = -2^k+2^k\cdot k+2^k = 2^k\cdot k = \mathcal{S}(T^{bal}).$$ This completes the proof. \end{proof} Corollary \ref{balmaxbal} shows that when $n=2^k$, the bound provided by Theorem \ref{minsackin} is tight, i.e. there is in fact a tree that obtains the lower bound of the Sackin index, but it does not consider any other values of $n$. We next want to show that the bound is in fact tight for all $n$, even if $n\neq 2^k$, and we will present an explicit construction to find such maximally balanced trees, i.e. trees with a minimal Sackin index. Therefore, we need the following Algorithm \ref{alg1}. The high-level idea of this algorithm is to start with a maximally balanced tree (and we know already by Corollary \ref{balmaxbal} that $T_k^{bal}$ has that property), possibly of too many leaves, and then deleting leaves, one at a time, in a way that keeps the balance maximal. In particular, we will show that this can be achieved by deleting the leaves of a cherry of maximum depth. \par \vspace{0.5cm} \begin{algorithm}[H]\label{alg1} \KwIn{number $n$ of leaves} \KwOut{tree $T$ with $n$ leaves and $\mathcal{S}(T)$ minimal. } Initialization: $k:=\lceil \log_2(n) \rceil$; $T:=T^{bal}_k$; $\widehat{n}:=2^k$\; \While{$\widehat{n} > n$ }{ Find a cherry $[u,v]$ of maximal depth in $T$ and call its parent node $w$\; Delete leaves $u$ and $v$ and edges $(w,u)$ and $(w,v)$ from $T$\; $\widehat{n}:=\widehat{n}-1$\; } \KwRet{$T$} \caption{Construction of maximally balanced rooted binary trees with $n$ leaves} \end{algorithm} \begin{theorem}\label{thm:alg} Let $n \in \mathbb{N}$ and let $k=\lceil \log_2(n) \rceil$. Then, Algorithm \ref{alg1} returns a tree $T$ with $\mathcal{S}(T)=-2^k+n(k+1)$. So Algorithm \ref{alg1} finds a tree with minimal $\mathcal{S}(T)$; i.e. a maximally balanced tree. \end{theorem} \begin{proof} Note that as before, $k=\lceil \log_2(n) \rceil$ implies $k=\min\{\widetilde{k}: n \leq 2^{\widetilde{k}}\}$. We now distinguish two cases. If $n=2^k$, the algorithm does not enter the while-loop, as $\widehat{n}=n$. Therefore, the algorithm, returns the tree from the initialization step, which is the fully balanced tree $T^{bal}_k$. We know by Proposition \ref{sackinBal} that $S(T^{bal}_k)=k\cdot 2^k$, which is minimal by Corollary \ref{balmaxbal}. Moreover, $S(T^{bal}_k)=k\cdot 2^k= -2^k+ (k\cdot 2^k) +2^k =-2^k+2^k(k+1)= -2^k+n(k+1).$ Here, the last equality is due to $n=2^k$. This completes the proof of this case. Now let us consider the case where $n<2^k$. Note that we additionally have $n > 2^{k-1}$, because otherwise $k$ would not be minimal with the property $n\leq 2^k$. We start with $T^{bal}_k$ employing $\widehat{n}=2^k$ leaves, and the while-loop reduces $\widehat{n}$ in each step by 1 and repeats that as long as $\widehat{n}>n$. So in total, this step is repeated $2^k-n$ times. In each step, the leaves of a cherry $[u,v]$ of maximal depth are deleted along with their respective pending edges. Let us call the tree before performing this step $T$ and the resulting tree after the cherry deletion $\widetilde{T}$. Note that the direct ancestor of $u$ and $v$ in $T$, say $w$, is a leaf in $\widetilde{T}$. So in total, $\widetilde{T}$ has one leaf less than $T$. So after repeating this step $2^k-n$ times, we end up with a tree with $2^k-(2^k-n)=n$ leaves. Moreover, as $[u,v]$ is a cherry of maximal depth, we know from Lemma \ref{cherrydeleted} that $\mathcal{S}(\widetilde{T})=\mathcal{S}(T)-\delta_u-1$. So each time we run through the while-loop, the Sackin index gets reduced by $(\delta_u+1)$. It is now crucial to note that the value of $\delta_u$ does not change throughout the procedure! This is because we start with $\widehat{n}=2^k$ and stop when $\widehat{n}=n>2^{k-1}$, and we always take a cherry of maximal depth. In the beginning, we have $\frac{2^k}{2}=2^{k-1}$ such cherries in $T^{bal}_k$ to choose from, and in each step, one of those cherries gets replaced by a leaf. So as long as one of the original cherries of $T^{bal}_k$ remains unchanged, the value of $\delta_u$ does not change. But if all of the original cherries were replaced by a leaf, we would end up with $T^{bal}_{k-1}$ and thus with $\widehat{n}=2^{k-1}<n$, which by definition of Algorithm \ref{alg1} cannot happen. So $\delta_u$ is the same each time we run the while-loop, and thus, in total, on the way from $T^{bal}_k$ to the final tree $T$, the Sackin index gets reduced by $(2^k-n)(\delta_u+1)$. So we have $\mathcal{S}(T)=S(T^{bal}_k)-(2^k-n)(\delta_u+1)$. Now note that the depth $\delta_u$ of a leaf of maximal depth in $T^{bal}_k$ is precisely $k$. In total, using Proposition \ref{sackinBal}, this leads to: $$\mathcal{S}(T)=\mathcal{S}(T^{bal}_k)-(2^k-n)(\delta_u+1) =k\cdot 2^k - (2^k-n)(k+1) =-2^k+n(k+1).$$ This completes the proof. \end{proof} \begin{remark} Note that Theorem \ref{thm:alg} proves that the bound provided by Theorem \ref{minsackin} is in fact tight for all $n$, and that Algorithm \ref{alg1} even provides a way of finding a maximally balanced tree for all $n$. The tight bound also implies that the sequence of minimal values of the Sackin index, starting at $n=1$, is $1,2,5,8,12,16,20,24,\ldots$ This corresponds to Sequence A003314 in the Online Encyclopedia of Integer Sequences OEIS \cite[Sequence A003314]{OEIS}, which is also often referred to as binary entropy function. \end{remark} Note that by Theorem \ref{thm:alg}, it is also clear why trees with minimal Sackin index are not necessarily unique with this property: In the while-loop, any cherry of maximum depth in a maximally balanced tree with $\widehat{n}$ leaves will provide a maximally balanced tree with $\widehat{n}-1$ leaves -- so the choice of the particular cherry of maximum depth might result in different trees. \begin{example} Consider the two trees $T_1$ and $T_2$ with $n=6$ leaves as depicted in Figure \ref{6taxabalanced}. These two trees both have Sackin index 16: $\mathcal{S}(T_1)=2+2+3+3+6=16=2+2+2+4+6=\mathcal{S}(T_2)$. Let $k=\lceil \log_2(6) \rceil=3$. Now it is easy to see that $\mathcal{S}(T_1)=\mathcal{S}(T_2)=16$ is indeed minimal, as for $n=6$ we have by Theorem \ref{minsackin} that $\mathcal{S}(T) \geq -2^k+n(k+1)=-2^3+6(3+1)=16$. And the reason why we have two maximally balanced trees with 6 leaves is due to the (unique) maximally balanced tree $T$ with 7 leaves (see Figure \ref{algfig}). In this tree we have three cherries of maximal depth. However, due to symmetry it does not make a difference if we choose either one of the two underbraced ones in the while-loop of Algorithm \ref{alg1}, because the resulting tree is in any case $T_1$. But if we choose the third cherry of maximal depth, the resulting tree is $T_2$. \end{example} \begin{figure} \centering \includegraphics[scale=.4]{algorithm1.eps} \caption{Illustration of Algorithm \ref{alg1}: For $n=5$, we consider $k=\lceil \log_2(5)\rceil =3 $ and thus start at $2^3=8$ leaves. Then, one by one, we delete a cherry of maximal depth until we reach $n=5$ leaves. In this case, when we go from 7 to 6, the particular choice of such a cherry can lead to different trees. \\Note that alternatively, instead of going from $2^k$ leaves down to $n$, one could start at $2^{k-1}$ leaves and go up to $n$, i.e. the arrows in the figure could be reversed. } \label{algfig} \end{figure} \begin{remark} It can be easily seen that instead of using the top-down principle as presented in Algorithm \ref{alg1}, it is equivalent to use a bottom-up principle. In particular, instead of constructing a maximally balanced tree with $n$ leaves according to the Sackin index by starting at $T_k^{bal}$ for $k= \lceil \log_2(n) \rceil$ and deleting cherries of maximal depth, one could instead start at $T_{k-1}^{bal}$ and replace leaves of minimal depth with a cherry. Figure \ref{algfig} illustrates this as all arrows could simply be reversed. However, as both algorithms can easily be shown to lead to equivalent results, we omit the proof for the second algorithm. \end{remark} We will now consider the set of trees of minimal Sackin index, i.e. the maximally balanced ones according to this index. As we will show, if the number of leaves is a power of 2, i.e. $n=2^k$, the maximally balanced tree is unique, namely $T_k^{bal}$. This has frequently been observed in the literature (cf. \citep{heard,shao}), but to the best of our knowledge, no formal proof has been presented so far. Moreover, no statement on the number of maximally balanced trees for leaf numbers that are not a power of 2 has been made so far. We will provide a recursive formula to calculate this number in the following. But before we can do that, we have to consider the easier case with $n=2^k$. \begin{theorem}\label{TkbalUNIQUE} Let $n=2^k$ for some $k\in \mathbb{N}$. Then, $T_k^{bal}$ is the {\bf unique} tree with minimal Sackin index, i.e. for any other tree $T$ with $n=2^k$ leaves we have $\mathcal{S}(T)>S(T_k^{bal})$. \end{theorem} \begin{proof} We know by Corollary \ref{balmaxbal} that $T_k^{bal}$ has minimum Sackin index, i.e. $S(T_k^{bal})=k\cdot2^k$. So all that remains to be shown is that $T_k^{bal}$ is unique with this property. We prove this statement by induction on $k$. For $k\leq 1$ there is nothing to show as then there is only one tree to consider, which gives therefore of course the unique minimum. Now we assume the statement holds for $k-1$ and we consider a tree $T$ with $n=2^k$ leaves and $\mathcal{S}(T)$ minimal, i.e. by Proposition \ref{sackinBal} and Corollary \ref{balmaxbal} we have $\mathcal{S}(T)=k\cdot 2^k$. We have to show that $T=T_k^{bal}$. Let $T=(T_a,T_b)$ be the standard decomposition of $T$ into its two maximal pending subtrees $T_a$ and $T_b$ with leaf numbers $n_a$ and $n_b$, respectively, such that $n=n_a+n_b$. Without loss of generality, we may assume $n_a \geq n_b$, i.e. $n_a \geq \frac{n}{2}$. Note that as $T$ has minimum Sackin index and as by definition we have $\mathcal{S}(T)= \mathcal{S}(T_a)+\mathcal{S}(T_b)+n$, by Lemma \ref{optStandardDecomp}, $\mathcal{S}(T_a)$ and $\mathcal{S}(T_b)$ must be minimal, too. We now distinguish two cases. \begin{enumerate} \item $n_a=n_b=\frac{n}{2}=2^{k-1}$. In this case, we know by the inductive hypothesis that $T_a$ and $T_b$ must equal $T_{k-1}^{bal}$, as this is the unique tree with minimum Sackin index and $ 2^{k-1}$ leaves. So $T=(T_a,T_b)$, and $T_a$ and $T_b$ are both equal to $T_{k-1}^{bal}$, which implies $T=T_k^{bal}$. This completes the proof of this case. \item $n_a > n_b$, i.e. $n_a > \frac{n}{2}=2^{k-1}$. Then, there exists an $s \in \mathbb{N}$, $s\geq 1$, such that $n_a=\frac{n}{2}+s=2^{k-1}+s$. Due to $2^k=n=n_a+n_b=2^{k-1}+s+n_b$ we conclude $2^k-2^{k-1}=s+n_b$ and thus $n_b=2^{k-1}-s$. As in the first case we have $\mathcal{S}(T)=\mathcal{S}(T_a)+\mathcal{S}(T_b)+n$ with $\mathcal{S}(T_a)$ and $\mathcal{S}(T_b)$ minimal. By Theorems \ref{minsackin} and \ref{thm:alg} we conclude that $\mathcal{S}(T_a)=-2^{k_a}+n_a\cdot(k_a+1)$ and $\mathcal{S}(T_b)=-2^{k_b}+n_b\cdot(k_b+1)$, where $k_a=\lceil \log_2(n_a) \rceil$ and $k_b=\lceil \log_2(n_b) \rceil$. Now, note that due to $n_a=2^{k-1}+s$ and $n_b=2^{k-1}-s$ for $s \geq 1$ we have: \begin{itemize} \item $k_a=k$ (as $n_a> 2^{k-1}$ and $n_a <n=2^k$) and \item $ 0 \leq k_b \leq k-1$ (as $n_b < 2^{k-1}$) and \item $s \leq 2^{k-1}-1$, because as $n_b \geq 1$ (otherwise, if $n<2$, we would be in the base case of the induction, so now we have $n\geq 2$ and therefore $n_b\geq 1$), we have $n_a \leq n-1$. This implies $\frac{n}{2}+s \leq n-1$, which yields $s \leq \frac{n}{2}-1$. Using $n=2^k$, we derive $s\leq 2^{k-1}-1$. \end{itemize} So, in total we get: $$\mathcal{S}(T)=\mathcal{S}(T_a)+\mathcal{S}(T_b)+n=-2^{k_a}+n_a(k_a+1)-2^{k_b}+n_b(k_b+1)+2^k.$$ Using $k_a=k$, $n_a=2^{k-1}+s$ and $n_b=2^{k-1}-s$, we get: $$\mathcal{S}(T)=-2^k +(2^{k-1}+s) \cdot (k+1)-2^{k_b}+(2^{k-1}-s)\cdot(k_b+1)+2^k$$ $$= 2^{k-1}(k+k_b)+s(k-k_b)+2^k-2^{k_b}. $$ Using $k_b =\lceil \log_2(n_b)\rceil= \lceil \log_2(2^{k-1}-s)\rceil$, this leads to: \begin{equation}\label{STbal}\mathcal{S}(T)= 2^{k-1}\cdot (k+ \lceil \log_2(2^{k-1}-s)\rceil) + s(k- \lceil \log_2(2^{k-1}-s)\rceil)+2^k-2^{ \lceil \log_2(2^{k-1}-s)\rceil}. \end{equation} We now consider the term $k_b=\lceil \log_2(2^{k-1}-s)\rceil$ in more detail. For a fixed value of $k$, this term depends only on $s$, so we will refer to it as $k_b(s)$. $k_b(s)$ can assume values ranging from $0=k-k$ (if $n_b=1$, i.e. if $T_b$ consists only of one leaf, which is the case when $s=2^{k-1}-1$) to $k-1$ if $s\in\{1,\ldots,2^{k-2}-1\}$ (note that as we are in the case where $n_a \neq n_b$, we already know that $s>0$). Moreover, as $k_b(s)=\lceil \log_2(2^{k-1}-s)\rceil$, we can conclude $s \geq 2^{k-1}-2^{k_b(s)}$. This implies: $$k_b(s)=k-a\mbox{ if } s\in \{2^{k-1}-2^{k-a},\ldots,2^{k-1}-2^{k-a-1}-1\},$$ where $a$ ranges from $1$ to $k$. The only exceptions are the cases $a=1$, where the possible choices for $s$ start at $1$, not at $2^{k-1}(2^0-1)=0$ (because we already know that $s>0$), and $a=k$, because $2^{k-(k+1)}(2^k-1)-1= 2^{k-1}-\frac{3}{2} \not\in \mathbb{N}$, so $s$ cannot assume this value (which is why in this case, the set of possible choices of $s$ contains only one element). Now we use these insights concerning $k_b(s)$ to find minimal values of $\mathcal{S}(T)$ using Equation (\ref{STbal}). First, we consider two trees $T_1$ and $T_2$ such that $s_1,s_2 \in \{2^{k-a}(2^{a-1}-1),\ldots,2^{k-(a+1)}(2^a-1)-1\}$ for some value of $a \in \{1,\ldots,k\}$. As both values $s_1$ and $s_2$ correspond to the same value of $a$, we have $k_b(s_1)=k_b(s_2)=k-a$ and therefore we get by Equation (\ref{STbal}): $$ \mathcal{S}(T_1)= 2^{k-1}\cdot (k+ (k-a)) + s_1(k- (k-a))+2^k-2^{ k-a}$$ $$=2^{k-1}\cdot (2k-a) + s_1\cdot a+2^k-2^{ k-a}.$$ Analogously, we get: $$ \mathcal{S}(T_2)= 2^{k-1}\cdot (k+ (k-a)) + s_2(k- (k-a))+2^k-2^{ k-a}$$ $$=2^{k-1}\cdot (2k-a) + s_2\cdot a+2^k-2^{ k-a}.$$ This implies that $ \mathcal{S}(T_1)< \mathcal{S}(T_2)$ if and only if $s_1<s_2$. So in order to minimize the Sackin index, in each possible line of Equation (\ref{kbcases}), i.e. for each possible value of $a$, the lower bound of the possible values for $s$ is the only candidate for a minimum. Now let us summarize what we have: If $a>1$, we know that $s=2^{k-a}(2^{a-1}-1)$ minimizes $\mathcal{S}$, and if $a=1$, $s=1$ minimizes $\mathcal{S}$. Next, we compare two trees $T_1$ and $T_2$ with different values $a_1$ and $a_2$, respectively, i.e. let $a_1$ and $a_2$ be in $\{2,\ldots,k\}$ such that $a_1<a_2$ (the case where one of the values is 1 will be considered later). Then, in order for $\mathcal{S}$ to be as small as possible, we have already seen that we must use the smallest possible values of $s_1$ and $s_2$, respectively, i.e. $s_1=2^{k-a_1}(2^{a_1-1}-1)$ and $s_2=2^{k-a_2}(2^{a_2-1}-1)$. Using this and considering again Equation (\ref{STbal}), we derive: $$ \mathcal{S}(T_1)= 2^{k-1}\cdot (k+ (k-a_1)) + s_1(k- (k-a_1))+2^k-2^{ k-a_1}$$ $$=2^{k}(k+1)-2^{k-a_1}\cdot (a_1+1),$$ $$ \mathcal{S}(T_2)= 2^{k}(k+1)\underbrace{-2^{k-a_2}}_{>-2^{k-a_1}}\cdot \underbrace{(a_2+1)}_{>(a_1+1)}> \mathcal{S}(T_1).$$ So if $a_1<a_2$ (for $a_1,a_2 \in \{2,\ldots,k\}$), we have $\mathcal{S}(T_1)<\mathcal{S}(T_2)$. This implies that for a minimal value of $\mathcal{S}$, $a$ has to be minimal. So the only candidate for $a$ from the set $\{2,\ldots,k\}$ is $a=2$. However, we have not investigated the case $a=1$ yet, which is different because in this case we are not allowed to choose $s=2^{k-1}(2^0-1)=0$ as we are in the case where $s\geq 1$. So what remains to be done is to compare the case $a=1$ and $s=1$ with the case $a=2$ and $s=2^{k-2}(2^1-1)=2^{k-2}$. Therefore, now let $T_1$ be such that $s_1=1$ and thus $a_1=1$, and let $T_2$ be such that $s_2=2^{k-2}$ and $a_2=2$. Then, we get by Equation (\ref{STbal}): $$ \mathcal{S}(T_1)= 2^{k-1}\cdot (k+ (k-a_1)) + s_1(k- (k-a_1))+2^k-2^{ k-a_1}= 2^{k}k + 1,$$ and analogously $$ \mathcal{S}(T_2)= 2^{k-1}\cdot (k+ (k-a_2)) + s_2(k- (k-a_2))+2^k-2^{ k-a_2}= 2^{k}k +2^{k-2}.$$ So as we are in the case where $k\geq 2$ (else consider the base case of the induction again), this leads to $ \mathcal{S}(T_2)\geq \mathcal{S}(T_1)=2^{k}k + 1$. So the minimal value of $\mathcal{S}$ is achieved by the choice of $s=1$ and thus $a=1$, i.e. $k_b=k-1$, but even for a tree fulfilling these conditions we have $\mathcal{S}(T)=2^{k}k + 1>2^kk=\mathcal{S}(T^{bal}_k)$. The latter equality holds due to Corollary \ref{balmaxbal}. So in total, we obtain $\mathcal{S}(T) >S(T_k^{bal})$, which contradicts the minimality of $\mathcal{S}(T)$. Therefore, the case $n_a\neq n_b$ leads to a contradiction, which is why it cannot occur. Thus we only have to consider the first case, for which we have already shown $T=T_k^{bal}$. This completes the proof. \end{enumerate} \end{proof} So now we know that in case of $n=2^k$, the minimum of the Sackin index is unique. But it has long been known that this need not be the case for other values of $n$. In fact, already for $n=6$, there are two minima, which are depicted in Figure \ref{6taxabalanced}. As stated before, this example is not new; it can for instance already be found in \cite{cophenetic}. However, so far the explicit number of Sackin minima for $n \neq 2^k$ has not been investigated. In order to derive such a formula, we first need the following theorem, which characterizes all Sackin minima. \begin{theorem}\label{SackinMinCharacterization} Let $T$ be a rooted binary tree with $n\geq 2$ leaves and $\mathcal{S}(T)$ minimal, i.e. there is no tree on $n$ leaves with a smaller Sackin index. Then, we have: $T$ has height $h_T=k$ with $k=\lceil \log_2(n) \rceil$, and $T$ has precisely $n-2^{k-1}$ cherries of depth $k$. \end{theorem} \begin{proof} We prove the statement by induction on $n$. Throughout this proof, for a rooted binary tree $T$, we denote by $c_T$ the number of cherries of maximal depth. Now for $n=2$, the only binary rooted tree $T$ consists only of a cherry, and this tree has $h_T=1 = \lceil \log_2(2) \rceil=k$, so the height statement holds. Moreover, $T$ has only one cherry, so $c_T=1=2-2^{1-1}=n-2^{k-1}$. This completes the base case. Next we assume that the statement is true for up to $n$ leaves and now consider a tree $T$ with $n+1$ leaves and with minimal Sackin index. First, in case $n+1= 2^{k}$ for some $k$, by Theorem \ref{TkbalUNIQUE} we know that $T=T_{k}^{bal}$. This immediately implies $h_T=\log_2(n+1)=k$ and $c_T=2^{k-1}=(n+1)-2^{k-1}.$ This completes the proof of the case where $n+1$ is a power of 2. So now we consider the case $2^{k-1}<n+1<2^k$ with $k=\lceil \log_2(n+1)\rceil$. Let $[u,v]$ be a cherry of maximal depth in $T$ with direct ancestor $w$. We remove $u$, $v$ as well as the edges $(w,u)$ and $(w,v)$ to get a tree $\widetilde{T}$ with $n$ leaves ($u$ and $v$ have been deleted, but $w$, which is in $T$ in internal node, is a leaf in $\widetilde{T}$). Now, again there are two cases: \begin{enumerate}\item $n=2^{k-1}$, which is precisely the case if $n+1 = 2^{k-1}+1$, or \item $n>2^{k-1}$.\end{enumerate} \begin{enumerate}\item We now consider the first (and easier) case: If $n=2^{k-1}$, there are again only two possibilities: \begin{enumerate}\item Either $\widetilde{T}$ has minimal Sackin index, in which case (by Theorem \ref{TkbalUNIQUE}) we know that $\widetilde{T}=T_{k-1}^{bal}$, or \item$\mathcal{S}(\widetilde{T})>S(T_{k-1}^{bal})$. \end{enumerate} \par \vspace{0.3cm} \begin{enumerate} \item \label{case1a} If $\widetilde{T}=T_{k-1}^{bal}$, we have $h_{\widetilde{T}}=k-1$, and as $\widetilde{T}$ was derived by $T$ by deleting one cherry of maximal depth, $T$ must look as depicted in Figure \ref{onecherry}. In particular, $h_T=(k-1)+1=k$, and $T$ has only one cherry of depth $k$ (namely $[u,v]$), as otherwise $h_{\widetilde{T}}$ would equal $h_T$. So $c_T=1=(2^{k-1}+1)-2^{k-1}=(n+1)-2^{k-1}$. This completes the proof for the case $\widetilde{T}=T_{k-1}^{bal}$. \begin{figure} \centering \includegraphics[scale=.4]{OneCherry.eps} \caption{Illustration of the scenario described in Case (\ref{case1a}). } \label{onecherry} \end{figure} \item Now assume that we have $\mathcal{S}(\widetilde{T})>\mathcal{S}(T_{k-1}^{bal})$, i.e. in summary we now have that $T$ with $n+1 = 2^{k-1}+1$ leaves has $\mathcal{S}(T)$ minimal, but $\widetilde{T}$ with $n=2^{k-1}$ leaves has a Sackin index strictly larger than that of $T_{k-1}^{bal}$, i.e. $\mathcal{S}(\widetilde{T})>S(T_{k-1}^{bal})= 2^{k-1}\cdot (k-1)$. The last equality is due to Proposition \ref{sackinBal}. Note that by construction of $\widetilde{T}$ and Lemma \ref{cherrydeleted}, we have $\mathcal{S}(\widetilde{T})=\mathcal{S}(T)-\delta_u-1$. Combining both statements on $\mathcal{S}(\widetilde{T})$, we conclude \begin{equation} \label{sackinofTildeT}\mathcal{S}(\widetilde{T})=\mathcal{S}(T)-\delta_u-1> 2^{k-1}\cdot (k-1)=S(T_{k-1}^{bal}). \end{equation} On the other hand, as $\mathcal{S}(T)$ is minimal by assumption, we know from Theorems \ref{minsackin} and \ref{thm:alg} that \begin{equation}\label{smallestsackin}\mathcal{S}(T)= -2^k+(n+1)(k+1)=-2^k+nk+n+k+1.\end{equation} Combining (\ref{sackinofTildeT}) with (\ref{smallestsackin}), we obtain: $$-2^k+nk+n+k+1-\delta_u-1>k\cdot 2^{k-1} -2^{k-1}.$$ Using $n=2^{k-1}$, we note that this holds precisely if $$-2^k+k\cdot 2^{k-1}+2^{k-1}+k+1-\delta_u-1>k\cdot 2^{k-1} -2^{k-1}.$$ Rearranging the inequality shows that this is fulfilled if and only if $$-2^k+2\cdot 2^{k-1}+k>\delta_u, $$ i.e. precisely if $k>\delta_u$. But recall that $u$ was part of a cherry of maximal depth, i.e. $u$ has maximal depth in $T$, and $T$ has $2^{k-1}+1>2^{k-1}$ leaves. Thus, it can easily be seen (cf. Lemma \ref{maxdepthleaf}) that $h_T\geq k$, which implies $\delta_u \geq k$. This contradicts $k>\delta_u$, which shows that this case cannot happen. So if $\widetilde{T}$ has $n=2^{k-1}$ leaves, $\widetilde{T}$ must equal $T_{k-1}^{bal}$, for which we have already proven the statement in (\ref{case1a}). This completes the proof of this case. \end{enumerate} \item Now what remains to be considered is the case $2^{k-1}<n<n+1<2^k$, which implies $k=\lceil \log_2(n)\rceil = \lceil \log_2(n+1)\rceil$. As above, we know that $\mathcal{S}(T)$ is minimal, i.e. by Theorems \ref{minsackin} and \ref{thm:alg} we have $\mathcal{S}(T)= -2^k+(n+1)(k+1)$, and again, by construction of $\widetilde{T}$, we have $\mathcal{S}(\widetilde{T})=\mathcal{S}(T)-\delta_u-1$. Again, we distinguish two cases: \begin{enumerate}\item $\mathcal{S}(\widetilde{T})$ is minimal for $n$, or \item there is another tree $\widehat{T}$ on $n$ leaves with $S(\widehat{T})<\mathcal{S}(\widetilde{T})$.\end{enumerate} \par\vspace{0.3cm} \begin{enumerate} \item \label{case2a} If $\mathcal{S}(\widetilde{T})$ is minimal for $n$, we know by Theorems \ref{minsackin} and \ref{thm:alg} that $\mathcal{S}(\widetilde{T})= -2^k+n(k+1).$ Combining this with $\mathcal{S}(\widetilde{T})=\mathcal{S}(T)-\delta_u-1$ yields \begin{equation}\label{lll}\mathcal{S}(T)=-2^k+nk+n+\delta_u+1.\end{equation} On the other hand, we also know that $\mathcal{S}(T)$ is minimal, so we have \begin{equation}\label{mmm}\mathcal{S}(T)=-2^k+(n+1)(k+1).\end{equation} Combining Equations (\ref{lll}) and (\ref{mmm}) leads to $-2^k+nk+n+\delta_u+1=-2^k+nk+n+k+1$, which immediately gives $\delta_u=k$. As $u$ was a leaf of maximal depth, we have $\delta_u=h_T$ and thus $h_T=k$. What remains to be shown is that $c_T=(n+1)-2^{k-1}$. But note that as $\mathcal{S}(\widetilde{T})$ is minimal and $\widetilde{T}$ has $n$ leaves, we know by induction that $c_{\widetilde{T}}= n-2^{k-1}$ and $h_{\widetilde{T}}=k$. We have already shown that $h_T=k$, too. So $T$ and $\widetilde{T}$ have the same height $k$, but $\widetilde{T}$ was constructed by deleting one cherry of maximal depth from $T$. So we have $c_T=c_{\widetilde{T}}+1= n-2^{k-1}+1=(n+1)-2^{k-1}$. This completes the proof of this case. \item On the other hand, if $\mathcal{S}(\widetilde{T})$ is not minimal, we know by Theorems \ref{minsackin} and \ref{thm:alg} that $\mathcal{S}(\widetilde{T})> -2^k+n(k+1).$ As above, we also know that $\mathcal{S}(T)=\mathcal{S}(\widetilde{T})+\delta_u+1$, and thus we conclude \begin{equation}\label{aaa}\mathcal{S}(T)>-2^k+n(k+1)+\delta_u+1.\end{equation} On the other hand, though, $\mathcal{S}(T)$ is minimal, so again by Theorems \ref{minsackin} and \ref{thm:alg} we have \begin{equation}\label{bbb}\mathcal{S}(T)=-2^k+(n+1)(k+1).\end{equation} Combining Inequality (\ref{aaa}) with Equation (\ref{bbb}), we conclude $$ -2^k+(n+1)(k+1) > -2^k+n(k+1)+\delta_u+1, $$ which holds precisely if $k>\delta_u$. However, as $u$ was a leaf of maximum depth in $T$, we have $h_T=\delta_u$ and thus $h_T<k$. But this is a contradiction, because as $T$ has more than $2^{k-1}$ leaves, it has at least height $k$, i.e. $h_T\geq k$ (cf. Lemma \ref{maxdepthleaf}). This implies that this case cannot happen, so in fact, $\mathcal{S}(\widetilde{T})$ has to be minimal and we must be in case (\ref{case2a}), for which we have already proven the statement. This completes the proof. \end{enumerate} \end{enumerate} \end{proof} Note that Theorem \ref{SackinMinCharacterization} characterizes trees with minimal Sackin index in terms of their height, which must be minimal, and in terms of the number of cherries of maximal depth, which must be maximal (both with respect to the number of leaves). So trees with minimal Sackin index can be easily identified with Theorem \ref{SackinMinCharacterization}. \begin{example} Consider again the two trees on $n=6$ taxa depicted in Figure \ref{6taxabalanced}, which -- as already stated above and as can be seen by exhaustive search through the space of all rooted binary trees with 6 leaves -- have minimal Sackin index for $n=6$. We have $k=\lceil \log_2(6)\rceil = 3$, and indeed, both trees have height 3 and 2 cherries of maximum depth 3, respectively, and thus $h_T=k=3$ and $c_T=n-2^{k-1}=6-2^{3-1}=6-4=2$. \end{example} Note that so far we have shown that all trees with minimal Sackin index have the height and cherry properties described in Theorem \ref{SackinMinCharacterization}, but as the previous example shows, we still have not shown that all trees with these properties are also automatically minimal (we had to hint at an exhaustive search in order to verify that the two trees in the example indeed were maximally balanced). But in fact, it turns out that the opposite is also true, i.e. if the height is minimal and if the number of cherries of maximal depth is maximal, then the tree under consideration has minimal Sackin index. Indeed, this must be true as we will show now that such a tree will be discovered by Algorithm \ref{alg1}, which by Theorems \ref{minsackin} and \ref{thm:alg} returns only trees with minimal Sackin index. \begin{theorem} \label{algrecovers} Let $T$ be a rooted binary tree with $n$ leaves and height $h_T=k=\lceil \log_2(n)\rceil$ and with $c_T=n-2^{k-1}$ cherries of maximal depth $k$. Then, $\mathcal{S}(T)$ is minimal, i.e. there is no tree $\widehat{T}$ with $n$ leaves and $S(\widehat{T})<\mathcal{S}(T)$. \end{theorem} \begin{proof} Let $T$ be as described in the theorem. We will first prove that $T$ can be discovered by Algorithm \ref{alg1}. Note that $2^{k-1}<n\leq 2^{k}$ for $k=\lceil \log_2(n)\rceil$. Recall that Algorithm \ref{alg1} starts with $T_k^{bal}$ and deletes one cherry of maximal depth at a time until $n$ leaves are reached. Moreover, as long as $n>2^{k-1}$ (which must be the case because otherwise Algorithm \ref{alg1} would start at $T_{k-1}^{bal}$ instead), we only delete at most $2^k-(2^{k-1}+1)=2^k-2^{k-1}-1=2^{k-1}(2-1)-1=2^{k-1}-1$ cherries\footnote{Note that this extreme case would correspond to $n=2^{k-1}+1$.} of the originally $2^{k-1} $ cherries of maximum depth in $T_k^{bal}$. So any tree on $n$ leaves with $2^{k-1}<n\leq 2^{k}$ that the algorithm recovers has height $k$ (as at least one cherry of maximum depth $k$ is kept throughout the algorithm!). Moreover, the number of cherries of maximum depth in a tree of $n$ leaves that are left by Algorithm \ref{alg1} is $c_T=2^{k-1}-(2^k-n)$, as we start with the $2^{k-1}$ cherries of $T_k^{bal}$ and make $2^k-n$ steps, in each of which we delete exactly one cherry of maximum depth. So we end up with $c_T=2^{k-1}-(2^k-n)=2^{k-1}-2^k+n=2^{k-1}(1-2)+n=n-2^{k-1}.$ So, in summary, all trees recovered by Algorithm \ref{alg1} have height $h_T=k$ and $c_T=n-2^{k-1}$. And, more importantly, as the particular choice of cherry of maximum depth that the algorithm deletes is arbitrary, {\em all} such trees can be recovered. So $T$ will be found by Algorithm \ref{alg1}, as all trees of height $k$ and with $c_T=n-2^{k-1}$ leaves of maximum depth can be reached by starting with $T_k^{bal}$ and deleting $2^{k-1}-(2^k-n)$ cherries of maximum depth. So $T$ can be recovered by Algorithm \ref{alg1}. But, on the other hand, we know by Theorem \ref{thm:alg} that any tree recovered by Algorithm \ref{alg1} has minimal Sackin index. This completes the proof. \end{proof} This immediately leads to the following corollary, which is a direct conclusion of Theorems \ref{SackinMinCharacterization} and \ref{algrecovers}. \begin{corollary}\label{optalg1} Let $T$ be a rooted binary tree. Then, $\mathcal{S}(T)$ is minimal if and only if $T$ can be recovered by Algorithm \ref{alg1}. \end{corollary} \begin{example} Consider again the two trees with $n=6$ taxa, which are depicted in Figure \ref{6taxabalanced}. As both trees can be recovered by Algorithm \ref{alg1} (c.f. Figure \ref{algfig}), both trees must by Theorem \ref{thm:alg} indeed have minimal Sackin index, and as they are the only two trees with six leaves that are recovered by Algorithm \ref{alg1}, they are by Corollary \ref{optalg1} indeed the only two trees with this property. So an exhaustive search through all trees on six taxa as proposed above is not necessary anymore -- instead, all trees with 6 leaves that can be recovered by Algorithm \ref{alg1} are maximally balanced, and all other trees cannot be. In this sense, Corollary \ref{optalg1} provides a complete characterization of trees with minimal Sackin index. \end{example} As the previous example shows, for some values of $n$ there is more than one tree with minimal Sackin index. However, without explicitly listing and enumerating all possible outputs of Algorithm \ref{alg1}, it is not easy to see why for six leaves there are two such trees and for, say, twelve leaves there are five such trees. Of course, by Corollary \ref{optalg1} it all has to do with the number of cherries of maximum depth we can choose from -- but that is not all. Some choices might lead to the same tree. For instance, consider $T_3^{bal}$ as depicted in Figure \ref{algfig} at the top. It can easily be checked that no matter which of the cherries we choose to delete in order to recover a tree with $n=7$ leaves and minimal Sackin index, we always end up with the same tree, namely with the tree depicted in the second line of the same figure. This is due to the symmetries of $T_3^{bal}$. However, when we go from $n=7$ down to $n=6$ and choose one of the underbraced cherries, we will end up with the same tree, namely with the one depicted in the third line of Figure \ref{algfig} on the left-hand side. But if we choose the third cherry of maximum depth in this tree, this will lead to another tree -- namely precisely the one depicted on the right-hand side of the same line in Figure \ref{algfig}. This is due to the fact that the two underbraced cherries are in some sense `symmetric' (in the sense that there is a graph isomorphism which exchanges them), while the other cherry clearly is not symmetric to the two first ones. However, it is not trivial to count such symmetries or isomorphisms. But luckily, it is quite easy to characterize the cases where the minimum is unique, as can be seen in the following corollary. \begin{corollary}\label{uniquesackinminima} Let $n\in \mathbb{N}$. Then, there is only one tree $T$ with minimum Sackin index if and only if there exists an $m \in \mathbb{N}$ such that $n\in \{2^m-1,2^m,2^m+1\}$. \end{corollary} \begin{proof} If $n=2^m$, then $m=\lceil\log_2(n)\rceil$, so $m=k$ in Theorem \ref{TkbalUNIQUE}, which implies that the minimum is indeed unique. If $n=2^m-1$, then again for $k=\lceil\log_2(n)\rceil$, we have $m=k$. So in this case, Algorithm \ref{alg1} would start with tree $T_k^{bal}$ and delete precisely one cherry. It can be easily seen that due to symmetry, all cherries lead to the same tree. So for $n=2^m-1$, we again have uniqueness of the most balanced tree. However, if $n=2^m+1$, for $k=\lceil\log_2(n)\rceil$ we have $k=m+1$. In this case, the algorithm would start at $T_{k+1}^{bal}$ and delete all but one cherries of maximal depth. Again, due to symmetry, it does not matter which of these cherries remains in the tree (the scenario is equivalent to the one depicted in Figure \ref{onecherry}). So whenever $n\in \{2^m-1,2^m,2^m+1\}$, Algorithm \ref{alg1} returns precisely one tree, which by Corollary \ref{optalg1} is the unique tree that minimizes the Sackin index in these cases. Now, assume that $n\notin \{2^m-1,2^m,2^m+1\}$ for any $m$. In particular, for $k=\lceil\log_2(n)\rceil$, this implies that $n \in \{2^{k-1}+2,\ldots,2^k-2\}$. We partition this set into two cases: either $n\in \{2^{k-1}+2,\ldots,2^{k}-2^{k-2}-1\}$ or $n \in \{2^{k}-2^{k-2},\ldots,2^k-2\}$. Recall that Algorithm \ref{alg1} starts with $T_k^{bal}$ with $2^k$ leaves that form $2^{k-1}$ cherries. Now in the first case, if $n\in \{2^{k-1}+2,\ldots,2^{k}-2^{k-2}-1\}$, at least $2^k-(2^k-2^{k-2}-1)=2^{k-2}+1$ cherries of $T_k^{bal}$ get replaced by single leaves in the course of the algorithm. This implies that at most $2^{k-1}-(2^{k-2}+1)=2^{k-2}-1$ cherries of maximal depth remain. So the remaining cherries can either all be together in one of the maximal pending subtrees of the tree, which we call $T_a$, or, for instance, one of them is in the other maximal pending subtree, which we call $T_b$. This already gives two options on how to choose a tree that can be recovered by Algorithm \ref{alg1}, which by Corollary \ref{optalg1} implies at least two optima. On the other hand, if $n \in \{2^{k}-2^{k-2},\ldots,2^k-2\}$, this implies that at most $2^k-(2^k-2^{k-2})=2^{k-2}$ cherries get replaced by single leaves. These replacements can either all happen in the same maximal pending subtree of the tree, which we then call $T_b$, or, for instance, all can happen in $T_b$ except for one which happens in the other maximal pending subtree, which we then call $T_a$. This again gives at least two options on how to choose a tree that can be recovered by Algorithm \ref{alg1}, which by Corollary \ref{optalg1} implies at least two optima. So in summary, the minimum is unique if and only if $n\in \{2^m-1,2^m,2^m+1\}$ for some $m \in \mathbb{N}$. This completes the proof. \end{proof} Note that Algorithm \ref{alg1} guarantees that the difference between the two maximal pending subtrees $T_a$ and $T_b$ of the standard decomposition of a minimum Sackin tree does not get too large (in terms of the number of leaves). We will quantify this in the following corollary, which is based on Corollary \ref{optalg1}. \begin{corollary} Let $T$ be a rooted binary tree with $n\in \mathbb{N}_{\geq 2}$ leaves. Moreover, let $T=(T_a,T_b)$ be the standard decomposition of $T$ into its two maximal pending subtrees, let $n_i$ denote the number of leaves in $T_i$ for $i\in \{a,b\}$, respectively, such that $n_a\geq n_b$. Let $k=\lceil \log_2 n\rceil$. Then, the following equivalence holds: $T$ has minimum Sackin index if and only if $ T_a$ and $T_b$ have minimum Sackin index and $n_a-n_b\leq \min\{n-2^{k-1},2^k-n\}.$ \end{corollary} \begin{proof} We first consider the case where $T$ has minimum Sackin index. It is clear that $T_a$ and $T_b$ then also have minimum Sackin index. Moreover, by considering Algorithm \ref{alg1}, one can easily see that the difference between $n_a$ and $n_b$ is maximal when you perform as many cherry deletions in $T_b$ as possible. If all cherry deletions happen in $T_b$, which implies $n\geq 2^k-2^{k-2}=3\cdot 2^{k-2}$, $T_a$ equals $T_{k-1}^{bal}$ and thus $n_a =2^{k-1}$ and thus $n_b=n-2^{k-1}$. So if $n\geq 2^k-2^{k-2}$, we have $n_a-n_b \leq 2^{k-1}-n+2^{k-1} = 2^k-n$. Moreover, as $n\geq 3\cdot 2^{k-2}$ also implies that $2^k-n \leq n-2^{k-1}$, this completes the proof for the case where all cherry deletions happen in $T_b$. However, if more than $2^{k-2}$ cherry deletions happen, $T_b$ equals $T_{k-2}^{bal}$ and thus we have $n_b=2^{k-2}$. Morever, in this case we have $n_a=n-2^{k-2}$, and $n < 3 \cdot 2^{k-2}$. Thus, on the one hand we have $n_a-n_b=n-2^{k-1}$and $n-2^{k-1} < 3 \cdot 2^{k-2} - 2^{k-1} = 2^{k-2}$. On the other hand, we have $2^k-n > 2^k-3 \cdot 2^{k-2}$. This implies $n_a-n_b=n-2^{k-1}<2^k-n$, which completes the first direction of the proof. Now let us consider $T=(T_a,T_b)$ such that $T_a$ and $T_b$ are Sackin minima and $n_a-n_b \leq \min\{n-2^{k-1},2^k-n\}$. We need to show that $T$ is also a Sackin minimum. First note that we have $n_a-n_b\leq 2^k-n$, which implies $n_a \leq 2^{k-1}$, and that we also have $n_a-n_b\leq n-2^{k-1}$, which implies $n_b\geq 2^{k-2}$. As $n=n_a+n_b$ and as $n_a\geq n_b$, we have $n_a \geq \frac{n}{2}$ and $n_b \leq \frac{n}{2}$. So altogether we get $\frac{n}{2} \leq n_a \leq 2^{k-1}$ and $2^{k-2} \leq n_b \leq \frac{n}{2}$. It is easy to see that these leaf numbers precisely correspond to the trees recovered by Algorithm \ref{alg1}. This completes the proof. \end{proof} As stated above, counting the symmetries that remain when cherries of maximal depth are deleted is not trivial, even if Corollary \ref{uniquesackinminima} already provides some insight into whether there is more than one tree that minimizes the Sackin index for a given value of $n$. So in order to conclude this section, we will show in the following that the number of trees with minimal Sackin index can be explicitly counted by a recursive formula. The proof of this formula exploits Corollary \ref{optalg1}. \begin{theorem} \label{recursion} Let $s(n)$ denote the number of binary rooted trees with $n$ leaves and with minimal Sackin index and let $k = \lceil \log_2(n)\rceil$. For any partition of $n$ into two integers $n_a$, $n_b$, i.e. $n=n_a+n_b$, we use $k_a$ and $k_b$ to denote $\lceil \log_2(n_a)\rceil$ and $\lceil \log_2(n_b)\rceil=\lceil \log_2(n-n_a)\rceil$, respectively. Moreover, let $$f(n)=\begin{cases}0 & \mbox{if $n$ is odd} \\ {s(\frac{n}{2})+1 \choose 2}& \mbox{else. }\end{cases}$$ Then, the following recursion holds: \begin{itemize} \item $s(1)=1$ \item $s(n)= \sum\limits_{\substack{(n_a,n_b): \\n_a+n_b=n,\\ n_a\geq \frac{n}{2},\\ k_a = k -1, \\ k_b = k -1, \\n_a \neq n_b }} s(n_a)\cdot s(n_b) +f(n)+s(n-2^{k-2})$ \end{itemize} \end{theorem} \begin{proof} First consider $n=1$. In this case, it is clear that there is only one rooted binary tree, namely the one consisting of only one node, which therefore has minimal Sackin index, which implies $s(1)=1$. Before we continue with the recursion, let us analyze Algorithm \ref{alg1} a little bit more in-depth. We know by Corollary \ref{optalg1} that if we want to count all trees with $n$ leaves with minimal Sackin index, we only need to count the ones that can be recovered by Algorithm \ref{alg1}. However, we have also already seen that all trees recovered by this algorithm have height $k= \lceil \log_2(n)\rceil$. Moreover, let $T$ be such a tree recovered by Algorithm \ref{alg1} and consider its standard decomposition $T=(T_a,T_b)$ with $n_a$ and $n_b$ leaves, respectively, such that $n=n_a+n_b$. As before, we can assume without loss of generality that $n_a\geq n_b$, i.e. that $n_a \geq \frac{n}{2}$. Then, Algorithm \ref{alg1} will assure that the larger of the two subtrees has height $k-1$. This is due to the fact that we start with $T_k^{bal}$, i.e. in the beginning both subtrees have height $k-1$, and then we delete at most all but one cherry of maximum depth (because otherwise the algorithm would have started at $T_{k-1}^{bal}$). So the larger of the two subtrees, $T_a$, always keeps its initial height, which is $k-1$, so we always have $k_a=k-1$. Therefore, by Corollary \ref{optalg1}, no trees with $k_a \neq k-1$ can have minimal Sackin index. Now consider $T_b$. For $T_k^{bal}$, both maximal pending subtrees have height $k-1$, but then some cherries of maximal depth are deleted. This way, it may happen that the height $k_b$ of $T_b$ is at some stage less than $k-1$ (given that the difference between $n$ and $2^k$ is large enough). However, note that we always have $k_b \geq k-2$ (and thus $n_b\geq 2^{k-2}$), because as soon as the height of $T_b$ is only $k_b=k-2$, there is no longer a cherry of maximum depth $k$ of $T$ to be found in $T_b$. So when the algorithm continues, it will never choose a cherry from $T_b$ again, because all cherries of maximum depth would now be in $T_a$. So in total, we have $k_b \in \{k-2,k-1\}$. So in summary, all trees $T$ recovered by Algorithm \ref{alg1}, and thus all trees with a minimal Sackin index, have the property that $k_a=k-1$ and $k_b \in \{k-2,k-1\}$ for $n=n_a+n_b$ and $n_a \geq \frac{n}{2}$. Trees with $n$ leaves which do not have these properties can by Corollary \ref{optalg1} thus not have minimal Sackin index. Moreover, we know that (as $\mathcal{S}(T)=\mathcal{S}(T_a)+\mathcal{S}(T_b)+n$ by definition) if $\mathcal{S}(T)$ is minimal, $\mathcal{S}(T_a)$ and $\mathcal{S}(T_b)$ must be minimal, too. Now we want to calculate $s(n)$. Therefore, we consider all integer partitions of $n$ into precisely 2 summands, i.e. $n=n_a+n_b$ for some $n_a$, $n_b \in \mathbb{N}$ such that $n_a \geq n_b$, i.e. $n_a\geq \frac{n}{2}$. Now we set $k:=\lceil \log_2(n)\rceil$, $k_a:=\lceil \log_2(n_a)\rceil$ and $k_b:=\lceil \log_2(n_b)\rceil$. Note that as $n_a \geq \frac{n}{2}$, $k_a \geq \lceil \log_2(\frac{n}{2})\rceil = \lceil \log_2(n)-\log_2(2)\rceil =\lceil \log_2(n)-1\rceil =\lceil \log_2(n)\rceil-1= k-1$, so $k_a \in \{k-1,k\}$. We distinguish the only three possible cases: \begin{enumerate} \item $k_a=k-1$ and $k_b=k-1$ and $n_a\neq n_b$, \item $k_a=k-1$ and $k_b=k-1$ and $n_a= n_b$, \item $k_a=k-1$ and $k_b=k-2$, \end{enumerate} \par\vspace{0.3cm} \begin{enumerate} \item Let $n_a$ be such that $k_a=k-1$ and $k_b=k-1$ and $n_a\neq n_b$. Any tree with $n$ leaves and minimum Sackin index and standard decomposition $T=(T_a,T_b)$ with $n=n_a+n_b$, $n_a \geq \frac{n}{2}$, $k_a=k-1$, $k_b=k-1$ and $n_a\neq n_b$ must have the property that $T_a$ and $T_b$ have minimum Sackin index, too, as $\mathcal{S}(T)=\mathcal{S}(T_a)+\mathcal{S}(T_b)+n$. This implies by Theorems \ref{minsackin} and \ref{thm:alg} that $\mathcal{S}(T_a)=-2^{k_a}+n_a(k_a+1)$ and $\mathcal{S}(T_b)=-2^{k_b}+n_b(k_b+1)$. Using this together with $k_a=k-1$ and $k_b=k-1$ and $n_b=n-n_a$, we derive: $$ \mathcal{S}(T) = \mathcal{S}(T_a)+\mathcal{S}(T_b)+n =-2^k+n(k+1).$$ So $\mathcal{S}(T)=-2^k+n(k+1)$, which we know must hold for any tree $T$ with $n$ leaves and minimum Sackin index (again due to Theorems \ref{minsackin} and \ref{thm:alg}). So {\em every } combination of $T_a$ and $T_b$ with the properties that $T_a$ and $T_b$ have minimum Sackin index and $n=n_a+n_b$, $k_a=k-1$, $k_b=k-1$ and $n_a\neq n_b$ leads to a valid tree $T=(T_a,T_b)$ on $n$ leaves, i.e. a tree which indeed has minimum Sackin index. So every such combination has to be considered. Thus, we sum over all those pairs $(n_a,n_b)$ and consider for each of them all possibilities to combine choices from the $s(n_a)$ trees with $n_a$ leaves with minimum Sackin index with the $s(n_b)$ trees with $n_b$ leaves and minimum Sackin index. There are $s(n_a)\cdot s(n_b)$ such combinations. This explains the first sum of the recursion. \item Now if $n$ is even, there is another summand: If $k_a=k-1$ and $k_b=k-1$ and $n_a= n_b$, we can proceed as in Case 1 but here, we have to consider all combinations of 2 trees from the $s(n_a)=s(n_b)$ trees with $n_a=n_b$ leaves and with minimum Sackin index, and there are ${s(n_a)-1+2 \choose 2} $ such combinations (unordered sampling of two trees from $s(n_a)$ trees with replacement, as the two trees may be equal and as their order does not matter), see for instance \cite[Equation (1.4.4)]{ProbTheory}) for more details. This explains the second summand in the recursion. \item Next, consider the case $k_a=k-1$ and $k_b=k-2$. Note that if $k_b=\lceil \log_2(n_b) \rceil= k-2$, we have $n_b\leq 2^{k-2}$. But as explained above, for a tree recovered by Algorithm \ref{alg1}, we always have $n_b\geq 2^{k-2}$, so now we must have $n_b=2^{k-2}$. By Theorem \ref{TkbalUNIQUE}, this implies that the only choice for $T_b$ with minimum Sackin index is $T_{k-2}^{bal}$. So for $T_b$, there is no alternative, but we can combine the only choice of $T_b$ with all possible choices of $T_a$, of which there are $s(n_a)=s(n-n_b)=s(n-2^{k-2})$ many. This explains the last summand in the above recursion. \end{enumerate} As we have considered all cases and added all contributions of each possible integer partition of $n$, there is nothing more to show. This completes the proof. \end{proof} \begin{remark} Starting at $n=1$ and continuing up to $n=32$, the sequence $s(n)$ of numbers of trees with $n$ leaves and with minimal Sackin index is 1, 1, 1, 1, 1, 2, 1, 1, 1, 3, 3, 5, 3, 3, 1, 1, 1, 4, 6, 14, 17, 27, 28, 35, 28, 27, 17, 14, 6, 4, 1, 1. We have calculated the values of $s(n)$ for up to $n=1024$. These data can be found online at \cite{NumberOfSackinMinima}. Note that this sequence is new to the Online Encyclopedia of Integer Sequences OEIS \cite[Sequence A299037]{OEIS}; it has been submitted in the scope of this manuscript. It had previously not been contained in the OEIS, i.e. this sequence has so far apparently not occurred in any other context. \end{remark} \section{Discussion} In this manuscript, we first proved some results which are maybe not surprising as they have been stated in the literature before, but as they still had to be proved, we delivered the mathematical arguments to back those results up. We also explicitly derived the sequences for the values $\mathcal{S}(T)$ (where $T$ is a tree with $n$ leaves) of the minimum and maximum Sackin indices for growing values of $n$, and showed that they are known sequences as they are already contained in OEIS. This might lead to future research as it connects the optimal Sackin index to other problems that can be found in the literature. Subsequently and more importantly, we provided two algorithms which can be used to explicitly find all trees with minimal Sackin index even in the complicated case where $n\neq 2^k$, about which previously little had been known. We then used these algorithms to characterize trees with a minimal Sackin index and to derive a recursion for the sequence of numbers of such trees, which is new to the OEIS. A question for future research could be to find a closed formula or a generating function describing this sequence. Another area of interest for future research might be the investigation of other well-known and frequently used balance indices like, for instance, the Colless index \cite{colless}. Moreover, the implications of our findings, i.e. of the number of extremal trees concerning the Sackin index, on evolutionary models and their induced probability distributions on the tree space are also of high interest. \section*{Acknowledgements} The author wishes to thank Lina Herbst and Kristina Wicke for very helpful discussions on the general topic and for comments concerning an earlier version of the manuscript. The author also wishes to thank Mike Steel for very helpful discussions and some suggestions, in particular concerning a shorter version of the first part of the proof of Lemma \ref{equidefs} and a shorter version of the proof of Proposition \ref{sackinCat}. \section*{Appendix} Here we state and prove some basic combinatorial and graph theoretic results that are helpful to understand the proofs in the main part of the manuscript. \begin{lemma} \label{h>k} Let $T$ be a rooted binary tree with $n$ leaves. Let $h=h(T)$ be the height of $T$, i.e. the number of edges on the longest path from the root to any of the leaves of $T$. Let $k \in \mathbb{N}$ be such that $n>2^k$. Then we have: $h> k$. \end{lemma} \begin{proof} Recall that in a rooted binary tree $T$ with node set $V$ and $n$ leaves, we have $|V|=2n-1$. So by assumption, we here have $|V|=2n-1> 2\cdot 2^k-1=2^{k+1}-1$, as we assume $n>2^k$. Now let $a_i$ denote the number of nodes in $V$ with level $i$, i.e. the number of nodes in $V$ whose shortest path to the root employs $i$ edges. Note that then, $a_0$ is 1, because the root $\rho$ itself is the only node with distance 0 to the root, and the node with maximal level has level $h$ -- it will be precisely one of the leaves whose depths defines the height of $T$. We now use the $a_i$ values to derive a second bound for $|V|$, namely by summing up over all possible levels: $ |V|=\sum\limits_{i=0}^{h}a_i.$ Note that it can be easily seen that as $T$ is binary, we have $a_i \leq 2^i$ for each $i=0,\ldots, h$, which leads to: $$|V|=\sum\limits_{i=0}^{h}a_i \leq \sum\limits_{i=0}^{h}2^i = 2^{h+1}-1.$$ The last equality uses the geometric series property. So, in total we conclude $2^{k+1}-1 < |V| \leq 2^{h+1}-1$. This implies $k < h$ and thus completes the proof. \end{proof} \begin{lemma}\label{maxdepthleaf} Let $T$ be a rooted binary tree with $n\geq2^{k-1}+1$ leaves for some $k \in \mathbb{N}_0$, and let $u$ be a leaf of maximal depth $\delta_u$ in $T$, i.e. $\delta_u=h(T)$. Then, $\delta_u=h(T) \geq k$. In particular, if $n=2^k$, we have $h(T) \geq k$. \end{lemma} \begin{proof} By Lemma \ref{h>k}, as $n\geq2^{k-1}+1> 2^{k-1}$, we conclude that $h(T)>k-1$ and thus $\delta_u>k-1$. As $\delta_u \in \mathbb{N}$, we conclude $\delta_u\geq k$. Moreover, if $n=2^k$, this implies $n\geq2^{k-1}+1$, so this completes the proof. \end{proof} \begin{lemma} \label{maxdepth} Let $T$ be a rooted binary tree with $n\geq 1$ leaves. Let $u$ be a leaf of maximal depth $\delta_u$ in $T$. Then, $\delta_u \leq n-1$. \end{lemma} \begin{proof} Considering that a rooted binary tree with $n$ leaves has $2n-1$ nodes in total and therefore $n-1$ internal nodes, the stated bound is obvious: The extreme case would be if a leaf had {\it all} internal nodes on its path to the root, which implies $\delta_u\leq n-1$ for all leaves $u$ of $T$. This completes the proof. \end{proof} \bibliographystyle{natbib}

https://arxiv.org/abs/0705.0998

The alternating sign matrix polytope

We define the alternating sign matrix polytope as the convex hull of nxn alternating sign matrices and prove its equivalent description in terms of inequalities. This is analogous to the well known result of Birkhoff and von Neumann that the convex hull of the permutation matrices equals the set of all nonnegative doubly stochastic matrices. We count the facets and vertices of the alternating sign matrix polytope and describe its projection to the permutohedron as well as give a complete characterization of its face lattice in terms of modified square ice configurations. Furthermore we prove that the dimension of any face can be easily determined from this characterization.

\section{Introduction and background} \label{sec:back} The Birkhoff polytope, which we will denote as $B_n$, has been extensively studied and generalized. It is defined as the convex hull of the $n \times n$ permutation matrices as vectors in $\mathbb{R}^{n^2}$. Many analogous polytopes have been studied which are subsets of $B_n$ (see e.g.~\cite{BRUALDI}). In contrast, we study a polytope containing $B_n$. We begin with the following definitions. \begin{definition} \label{def:asm} Alternating sign matrices (ASMs) are square matrices with the following properties: \begin{itemize} \item entries $\in \{0,1,-1\}$ \item the entries in each row and column sum to 1 \item nonzero entries in each row and column alternate in sign \end{itemize} \end{definition} \begin{figure}[htbp] \[ \left( \begin{array}{rrr} 1 & 0 & 0 \\ 0 & 1 & 0\\ 0 & 0 & 1 \end{array} \right) \left( \begin{array}{rrr} 1 & 0 & 0 \\ 0 & 0 & 1\\ 0 & 1 & 0 \end{array} \right) \left( \begin{array}{rrr} 0 & 1 & 0 \\ 1 & 0 & 0\\ 0 & 0 & 1 \end{array} \right) \left( \begin{array}{rrr} 0 & 1 & 0 \\ 1 & -1 & 1\\ 0 & 1 & 0 \end{array} \right) \] \[ \left( \begin{array}{rrr} 0 & 1 & 0 \\ 0 & 0 & 1\\ 1 & 0 & 0 \end{array} \right) \left( \begin{array}{rrr} 0 & 0 & 1 \\ 1 & 0 & 0\\ 0 & 1 & 0 \end{array} \right) \left( \begin{array}{rrr} 0 & 0 & 1 \\ 0 & 1 & 0\\ 1 & 0 & 0 \end{array} \right) \] \label{fig:n3asm} \caption{The $3\times 3$ ASMs} \end{figure} The total number of $n \times n$ alternating sign matrices is given by the expression \begin{equation} \label{eq:product} \displaystyle\prod_{j=0}^{n-1} \frac{(3j+1)!}{(n+j)!}. \end{equation} Mills, Robbins, and Rumsey conjectured this formula~\cite{MRRASMDPP}, and then over a decade later Doron Zeilberger proved it~\cite{ZEILASM}. Shortly thereafter, Kuperberg found a bijection between ASMs and the statistical physics model of square ice with domain wall boundary conditions (which is very similar to the \emph{simple flow grids} defined later in this paper), and gave a shorter proof using insights from physics~\cite{KUP_ASM_CONJ}. For a detailed exposition of the conjecture and proof of the enumeration of ASMs, see~\cite{BRESSOUDBOOK}. See Figure~\ref{fig:n3asm} for the seven $3\times 3$ ASMs. \begin{definition} The $n$th alternating sign matrix polytope, which we will denote as $ASM_n$, is the convex hull in $\mathbb{R}^{n^2}$ of the $n\times n$ alternating sign matrices. \end{definition} From Definition~\ref{def:asm} we see that permutation matrices are the alternating sign matrices whose entries are nonnegative. Thus $B_n$ is contained in $ASM_n$. The connection between permutation matrices and ASMs is much deeper than simply containment. There exists a partial ordering on alternating sign matrices that is a distributive lattice. This lattice contains as a subposet the Bruhat order on the symmetric group, and in fact, it is the smallest lattice that does so (i.e.\ it is the MacNeille completion of the Bruhat order)~\cite{TREILLIS}. Given this close relationship between permutations and ASMs it is natural to hope for theorems for $ASM_n$ analogous to those known for $B_n$. In this paper we find analogues for $ASM_n$ of the following theorems about the Birkhoff polytope (see the discussion in~\cite{BRUALDI} and~\cite{YEMELICHEV}). \begin{itemize} \item $B_n$ consists of the $n\times n$ nonnegative doubly stochastic matrices (square matrices with nonnegative real entries whose rows and columns sum to 1). \item The dimension of $B_n$ is $(n-1)^2$. \item $B_n$ has $n!$ vertices. \item $B_n$ has $n^2$ facets (for $n\ge3$) where each facet is made up of all nonnegative doubly stochastic matrices with a 0 in a specified entry. \item $B_n$ projects onto the permutohedron. \item There exists a nice characterization of its face lattice in terms of elementary bipartite graphs~\cite{BILLERA}. \end{itemize} As we shall see in Theorem~\ref{thm:asmchasing}, the row and column sums of every matrix in $ASM_n$ must equal 1. Thus the dimension of $ASM_n$ is $(n-1)^2$ because, just as for the Birkhoff polytope, the last entry in each row and column is determined to be precisely what is needed to make that row or column sum equal 1. In Section~\ref{sec:properties} we prove that $ASM_n$ has $4[(n-2)^2+1]$ facets and its vertices are the alternating sign matrices. We also prove analogous theorems about the inequality description of $ASM_n$ (Section~\ref{sec:asm}), the face lattice (Section~\ref{sec:facelattice}), and the projection to the permutohedron (Section~\ref{sec:properties}). See~\cite{ZIEGLER} for background and terminology on polytopes. The alternating sign matrix polytope was independently defined in~\cite{KNIGHT} in which the authors also study the integer points in the $r$th dilate of $ASM_n$ calling them \emph{higher spin alternating sign matrices}. \section{The inequality description of the ASM polytope} \label{sec:asm} The main theorem about the Birkhoff polytope is the theorem of Birkhoff~\cite{BIRKHOFFPOLY} and von~Neumann~\cite{VONNEUMANN} which says that the Birkhoff polytope can be described not only as the convex hull of the permutation matrices but equivalently as the set of all nonnegative doubly stochastic matrices (real square matrices with row and column sums equaling 1 whose entries are nonnegative). The inequality description of the alternating sign matrix polytope is similar to that of the Birkhoff polytope. It consists of the subset of doubly stochastic matrices (now allowed to have negative entries) whose partial sums in each row and column are between 0 and 1. The proof uses the idea of von Neumann's proof of the inequality description of the Birkhoff polytope~\cite{VONNEUMANN}. Note that in~\cite{KNIGHT} Behrend and Knight approach the equivalence of the convex hull definition and the inequality description of $ASM_n$ in the opposite manner, defining the alternating sign matrix polytope in terms of inequalities and then proving that the vertices are the alternating sign matrices. \begin{theorem} \label{thm:asmchasing} The convex hull of $n\times n$ alternating sign matrices consists of all $n\times n$ real matrices $X=\{x_{ij}\}$ such that: \begin{align} \label{eq:partialcolumnsum} 0 \le \sum_{i=1}^{i'} x_{ij}&\le 1 \hspace{.6in} \forall\mbox{ } 1\le i'\le n, 1\le j\le n.\\ \label{eq:partialrowsum} 0 \le \sum_{j=1}^{j'} x_{ij}&\le 1 \hspace{.6in} \forall\mbox{ } 1\le j'\le n, 1\le i\le n.\displaybreak\\ \label{eq:sumagain} \sum_{i=1}^n x_{ij}&= 1 \hspace{.6in} \forall\mbox{ }1\le j \le n.\\ \label{eq:sumagain2} \sum_{j=1}^n x_{ij}&= 1 \hspace{.6in} \forall\mbox{ } 1\le i \le n. \end{align} \end{theorem} \begin{proof} Call the subset of $\mathbb{R}^{n^2}$ given by the above inequalities $P(n)$. It is easy to check that the convex hull of the alternating sign matrices is contained in the set $P(n)$. It remains to show that any $X\in P(n)$ can be written as a convex combination of alternating sign matrices. Let $X\in P(n)$. Let $r_{ij}=\sum_{j'=1}^j x_{ij}$ and $c_{ij}=\sum_{i'=1}^i x_{ij}$. Thus the $r_{ij}$ are the row partial sums and the $c_{ij}$ are the column partial sums. It follows from (\ref{eq:partialcolumnsum}) and (\ref{eq:partialrowsum}) that $0\le r_{ij},c_{ij}\le1$ for all $1\le i,j\le n$. Also, from (\ref{eq:sumagain}) and (\ref{eq:sumagain2}) we see that $r_{in}=c_{nj}=1$ for all $1\le i,j\le n$. If we set $r_{i0}=c_{0j}=0$ we see that every entry $x_{ij}\in X$ satisfies $x_{ij}=r_{ij}-r_{i,j-1}=c_{ij}-c_{i-1,j}$. Thus \begin{equation} \label{eq:rcrel} r_{ij}+c_{i-1,j}=c_{ij}+r_{i,j-1}. \end{equation} Using von Neumann's terminology, we call a real number $\alpha$ \emph{inner} if $0<\alpha<1$. We construct a circuit in $X$ such that the partial sum between adjacent matrix entries in the circuit be an inner. So we rewrite the matrix $X$ with the partial sums between entries as shown below. \[ \left( \begin{array}{ccccccccccc} & \textcolor{blue}{c_{01}}&&\textcolor{blue}{c_{02}}&& &&\textcolor{blue}{c_{0,n-1}}&&\textcolor{blue}{c_{0n}}&\\ \textcolor{blue}{r_{10}}&x_{11}&\textcolor{blue}{r_{11}}&x_{12}&\textcolor{blue}{r_{12}}& &&x_{1,n-1}&\textcolor{blue}{r_{1,n-1}}&x_{1n}&\textcolor{blue}{r_{1n}}\\ & \textcolor{blue}{c_{11}}&&\textcolor{blue}{c_{12}}&&\ldots&&\textcolor{blue}{c_{1,n-1}}&&\textcolor{blue}{c_{1n}}&\\ \textcolor{blue}{r_{20}}&x_{21}&\textcolor{blue}{r_{21}}&x_{22}&\textcolor{blue}{r_{22}}& &&x_{2,n-1}&\textcolor{blue}{r_{2,n-1}}&x_{2n}&\textcolor{blue}{r_{2n}}\\ &&&&&&&&&&\\ &&\vdots&&&&&&\vdots&&\\ & \textcolor{blue}{c_{n-1,1}}&&\textcolor{blue}{c_{n-1,2}}&& &&\textcolor{blue}{c_{n-1,n-1}}&&\textcolor{blue}{c_{n-1,n}}&\\ \textcolor{blue}{r_{n0}}&x_{n1}&\textcolor{blue}{r_{n1}}&x_{n2}&\textcolor{blue}{r_{n2}}&\ldots&&x_{n,n-1}&\textcolor{blue}{r_{n,n-1}}&x_{nn}&\textcolor{blue}{r_{nn}}\\ & \textcolor{blue}{c_{n1}}&&\textcolor{blue}{c_{n2}}&&&&\textcolor{blue}{c_{n,n-1}}&&\textcolor{blue}{c_{nn}}& \end{array} \right) \] Begin at the vertex to the left or above any inner partial sum; if no such partial sum exists, then $X$ is an alternating sign matrix. Then there exists an adjacent inner partial sum by (\ref{eq:rcrel}). By repeated application of (\ref{eq:rcrel}) to each new inner partial sum, a path can then be formed by moving from entry to entry of $X$ along inner partial sums. Since $X$ is of finite size and all the boundary partial sums are 0 or 1 (i.e. non--inner), the path eventually reaches an entry in the same row or column as a previous entry yielding a circuit in $X$ whose partial sums are all inner. Using this circuit we can write $X$ as a convex combination of two matrices in $P(n)$, each with at least one more non--inner partial sum, in the following way. \begin{figure} \centering $\begin{array}{lcr} \left( \begin{array}{rrrrr} 0&.4&.5&.1&0\\ .4&-.4&.5&0&.5\\ .6&.4&-.3&-.1&.4\\ 0&.3&-.3&.9&.1\\ 0&.3&.6&.1&0 \end{array} \right) & \Rightarrow & \left( \begin{array}{ccccccccccc} & \textcolor{blue}{0}&&\textcolor{blue}{0}&&\textcolor{blue}{0}&&\textcolor{blue}{0}&&\textcolor{blue}{0}&\\ \textcolor{blue}{0}&0&\textcolor{blue}{0}&.4&\textcolor{red}{\textbf{.4}}&.5&\textcolor{red}{\textbf{.9}}&.1&\textcolor{blue}{1}&0&\textcolor{blue}{1}\\ & \textcolor{blue}{0}&&\textcolor{red}{\textbf{.4}}&&\textcolor{blue}{.5}&&\textcolor{red}{\textbf{.1}}&&\textcolor{blue}{0}&\\ \textcolor{blue}{0}&.4&\textcolor{red}{\textbf{.4}}&-.4&\textcolor{blue}{0}&.5&\textcolor{blue}{.5}&0&\textcolor{blue}{.5}&.5&\textcolor{blue}{1}\\ & \textcolor{red}{\textbf{.4}}&&\textcolor{blue}{0}&&\textcolor{blue}{1}&&\textcolor{red}{\textbf{.1}}&&\textcolor{blue}{.5}&\\ \textcolor{blue}{0}&.6&\textcolor{red}{\textbf{.6}}&.4&\textcolor{blue}{1}&-.3&\textcolor{red}{\textbf{.7}}&-.1&\textcolor{blue}{.6}&.4&\textcolor{blue}{1}\\ & \textcolor{blue}{1}&&\textcolor{red}{\textbf{.4}}&&\textcolor{red}{\textbf{.7}}&&\textcolor{blue}{0}&&\textcolor{blue}{.9}&\\ \textcolor{blue}{0}&0&\textcolor{blue}{0}&.3&\textcolor{red}{\textbf{.3}}&-.3&\textcolor{blue}{0}&.9&\textcolor{blue}{.9}&.1&\textcolor{blue}{1}\\ & \textcolor{blue}{1}&&\textcolor{blue}{.7}&&\textcolor{blue}{.4}&&\textcolor{blue}{.9}&&\textcolor{blue}{1}&\\ \textcolor{blue}{0}&0&\textcolor{blue}{0}&.3&\textcolor{blue}{.3}&.6&\textcolor{blue}{.9}&.1&\textcolor{blue}{1}&0&\textcolor{blue}{1}\\ & \textcolor{blue}{1}&&\textcolor{blue}{1}&&\textcolor{blue}{1}&&\textcolor{blue}{1}&&\textcolor{blue}{1}& \end{array} \right) \end{array}$ \caption[A circuit in a matrix in $P(n)$]{A matrix in $P(n)$ along with the matrix rewritten with the partial sums between the entries and a circuit of inner partial sums shown in boldface red} \label{fig:asmcircuit} \end{figure} Label the corner matrix entries in the circuit alternately ($+$) and ($-$). Define \[k'=\min(r_{ij},1-r_{i'j'},c_{i''j''},1-c_{i''',j'''})\] where $r_{ij}$, $r_{i'j'}$, $c_{i''j''}$, and $c_{i''',j'''}$ are taken over respectively the row partial sums to the right of a ($-$) corner along the circuit, the row partial sums to the right of a ($+$) corner, the column partial sums below a ($-$) corner, and the column partial sums below a ($+$) corner. Subtract $k'$ from the entries labeled ($-$) and add $k'$ to the entries labeled ($+$). Subtracting and adding $k'$ in this way preserves the row and column sums and keeps all the partial sums weakly between 0 and 1 (satisfying (\ref{eq:partialcolumnsum})--(\ref{eq:sumagain2})), so the result is another matrix $X'$ in $P(n)$ with at least one more non--inner partial sum than $X$. Now give opposite labels to the corners in the circuit in $X$ and subtract and add another constant $k''$ in a similar way to obtain another matrix $X''$ in $P(n)$ with at least one more non--inner partial sum than $X$. Then $X$ is a convex combination of $X'$ and $X''$, namely $X=\frac{k''}{k'+k''} X' + \frac{k'}{k'+k''}X''$. Therefore, by repeatedly applying this procedure, $X$ can be written as a convex combination of alternating sign matrices (i.e. matrices of $P(n)$ with no inner partial sums). \end{proof} \section{Properties of the ASM polytope} \label{sec:properties} Now that we can describe the alternating sign matrix polytope in terms of inequalities, let us use this inequality description to examine some of the properties of $ASM_n$, namely, its facets, its vertices, and its projection to the permutohedron. To make the proofs of the next two theorems more transparent, we introduce modified square ice configurations called \emph{simple flow grids} which will be used more extensively in Section~\ref{sec:facelattice}. Consider a directed graph with $n^2+4n$ vertices: $n^2$ `internal' vertices $(i,j)$ and $4n$ `boundary' vertices $(i,0)$, $(0,j)$, $(i,n+1)$, and $(n+1,j)$ where $i,j=1,\ldots,n$. These vertices are naturally depicted in a grid in which vertex $(i,j)$ appears in row $i$ and column $j$. Define the \emph{complete flow grid} $C_n$ to be the directed graph on these vertices with edge set $\{((i,j),(i,j\pm1)),((i,j),(i\pm1,j))\}$ for $i,j=1,\ldots,n$. So $C_n$ has directed edges pointing in both direction between neighboring internal vertices in the grid, and also directed edges from internal vertices to neighboring border vertices. \begin{definition} \label{definition:simpleflowgrid} A \emph{simple flow grid} of order $n$ is a subgraph of $C_n$ consisting of all the vertices of $C_n$ for which four edges are incident to each internal vertex: either four edges directed inward, four edges directed outward, or two horizontal edges pointing in the same direction and two vertical edges pointing in the same direction. \end{definition} \begin{proposition} \label{prop:sfgasmbij} There exists an explicit bijection between simple flow grids of order $n$ and $n\times n$ alternating sign matrices. \end{proposition} \begin{proof} Given an ASM $A$, we will define a corresponding directed graph $g(A)$ on the $n^2$ internal vertices and $4n$ boundary vertices arranged on a grid as described above. Let each entry $a_{ij}$ of $A$ correspond to the internal vertex $(i,j)$ of $g(A)$. For neighboring vertices $v$ and $w$ in $g(A)$ let there be a directed edge from $v$ to $w$ if the partial sum from the border of the matrix to the entry corresponding to $v$ in the direction pointing toward $w$ equals 1. By the definition of alternating sign matrices, there will be exactly one directed edge between each pair of neighboring internal vertices and also a directed edge from an internal vertex to each neighboring border vertex. Vertices of $g(A)$ corresponding to 1's are sources and vertices corresponding to $-1$'s are sinks. The directions of the rest of the edges in $g(A)$ are determined by the placement of the 1's and $-1$'s, in that there is a series of directed edges emanating from the 1's and continuing until they reach a sink or a border vertex. Thus $g(A)$ is a simple flow grid. Also, given a simple flow grid we can easily find the corresponding ASM by replacing all the sources with 1's and all the sinks with $-1$'s. Thus simple flow grids are in one-to-one correspondence with ASMs (see Figure~\ref{fig:5by5grid}). \end{proof} \begin{figure}[htbp] \centering \[ \includegraphics[scale=0.2]{5by5griddwbc.eps} \hspace{.7in} \includegraphics[scale=0.25]{matrix.eps} \] \caption{The simple flow grid---ASM correspondence} \label{fig:5by5grid} \end{figure} Simple flow grids are, in fact, almost the same as configurations of the six-vertex model of square ice with domain wall boundary conditions (see the discussion in~\cite{BRESSOUDBOOK}), the only difference being that the horizontal arrows point in the opposite direction. Recall that for $n\ge 3$ the Birkhoff polytope has $n^2$ facets (faces of dimension one less than the polytope itself). ($B_2=ASM_2$ is simply a line segment, so the number of facets equals the number of vertices which is 2.) Each facet of the Birkhoff polytope consists of all nonnegative doubly stochastic matrices with a zero in a fixed entry, that is, where one of the defining inequalities is made into an equality. The analogous theorem for $ASM_n$ is the following. \begin{theorem} \label{theorem:facets} $ASM_n$ has $4[(n-2)^2+1]$ facets, for $n\ge 3$. \end{theorem} \begin{proof} Note that the $4 n^2$ defining inequalities for $X\in ASM_n$ given in (\ref{eq:partialcolumnsum}) and (\ref{eq:partialrowsum}) can be restated as \[\sum_{i'=1}^{i} x_{i'j} \ge 0 \hspace{.5 in}\sum_{j'=1}^{j} x_{ij'} \ge 0\] \[\sum_{i'=i}^n x_{i'j} \ge 0 \hspace{.5 in}\sum_{j'=j}^n x_{ij'} \ge 0\] for $i,j=1,\ldots,n$. We have rewritten the statement that the row and column partial sums from the left or top must be less than or equal to $1$ as the row and column partial sums from the right and bottom must be greater than or equal to $0$. By counting these defining inequalities, one sees that there could be at most $4n^2$ facets, each determined by making one of the above inequalities an equality. It is left to determine how many of these equalities determine a face of dimension less than $(n-1)^2-1$. By symmetry we can determine the number of facets coming from the inequalities $\sum_{i'=1}^{i} x_{i'j} \ge 0$ for $i,j=1,\ldots n$ and then multiply by 4. Since the full row and column sums always equal $1$, the equalities such as $\sum_{i'=1}^n x_{i'j} = 0$ yield the empty face ($i=n$). Also, $\sum_{i'=1}^{n-1} x_{i'j}\ge 0$ is implied from the fact that $x_{nj}\ge0$ ($i=n-1$). The inequalities $x_{i'1}\ge 0$ for all $i'$, i.e. the entries in the first column are nonnegative, imply that $\sum_{i'=1}^i x_{i'1}\ge 0$ for $2\le i\le n-1$ ($j=1$), thus each of these sets is a face of dimension less than $(n-1)^2-1$, and similarly for $\sum_{i'=1}^i x_{i'n}\ge 0$ for $2\le i\le n-1$ ($j=n$) the partial sums of the last column. So we are left with the $(n-2)^2$ inequalities $\sum_{i'=1}^{i} x_{i'j} = 0$ for $i=1,\ldots n-2$ and $j=2,\ldots,n$ along with the inequality $x_{11}\ge 0$. For our symmetry argument to work, we do not include $x_{n1}\ge 0$ in our count since $x_{n1}\ge 0$ is also an inequality of the form $\sum_{j'=1}^{j} x_{ij'} \ge 0$. Thus $ASM_n$ has at most $4[(n-2)^2+1]$ facets, given explicitly by the $4(n-2)^2+4$ sets of all $X\in ASM_n$ which satisfy one of the following: \begin{equation} \label{eq:faceteq} \sum_{i'=1}^{i-1} x_{i'j}=0, \sum_{j'=1}^{j-1} x_{ij'}=0, \sum_{i'=i+1}^n x_{i'j}=0, \sum_{j'=j+1}^n x_{ij'}=0, i,j\in\{2,\ldots,n-1\}, \end{equation} \begin{equation} \label{eq:corners} x_{11}=0,\mbox{ } x_{1n}=0,\mbox{ } x_{n1}=0,\mbox{ or } x_{nn}=0. \end{equation} They are facets (not just faces) since each equality determines exactly one more entry of the matrix, decreasing the dimension by one. Recall that a directed edge in a simple flow grid $g(A)$ represents a location in the corresponding ASM $A$ where the partial sum equals 1, thus a directed edge missing from $g(A)$ represents a location in $A$ where the partial sum equals 0. Thus we can represent each of the $4(n-2)^2$ facets of (\ref{eq:faceteq}) as subgraphs of the complete flow grid $C_n$ from which a single directed edge has been removed: $((i\pm 1,j),(i,j))$ or $((i,j\pm 1),(i,j))$ with $i,j\in\{2,\ldots,n-1\}$. We can represent the facets of (\ref{eq:corners}) as subgraphs of $C_n$ from which two directed edges have been removed: $((1,1),(1,2))$ and $((1,1),(2,1))$, $((1,n),(1,n-1))$ and $((1,n),(2,n))$, $((n,1),(n-1,1))$ and $((n,1),(n,2))$, or $((n,n),(n,n-1))$ and $((n,n),(n-1,n))$. Now given any two facets $F_1$ and $F_2$, it is easy to exhibit a pair of ASMs $\{X_1,X_2\}$ such that $X_1$ lies on $F_1$ and not on $F_2$. Include the directed edge(s) corresponding to $F_2$ but not the directed edge(s) corresponding to $F_1$ in $g(X_1)$, then do the opposite for $X_2$. Thus each of the $4[(n-2)^2+1]$ equalities gives rise to a unique facet. \end{proof} \begin{corollary} For $n\ge 3$, the number of facets of $ASM_n$ on which an ASM $A$ lies is given by $2(n-1)(n-2)+(\mbox{number of corner 1's in $A$}).$ \end{corollary} \begin{proof} Each $0$ around the border of $A$ represents one facet. Thus the number of facets corresponding to border zeros of $A$ equals $4(n-1) - (\mbox{\# 1's around the border of $A$})$. Then there are $2(n-2)(n-3)$ facets represented by directed edges pointing in the opposite directions to the directed edges in the $(n-2)\times(n-2)$ interior array of $g(A)$. The sum of these numbers gives the above count. \end{proof} Even though $ASM_n$ is defined as the convex hull of the ASMs, it requires some proof that each ASM is actually an extreme point of $ASM_n$. \begin{theorem} The vertices of $ASM_n$ are the $n\times n$ alternating sign matrices. \end{theorem} \begin{proof} Fix an $n\times n$ ASM $A$. In order to show that $A$ is a vertex of $ASM_n$, we need to find a hyperplane with $A$ on one side and all the other ASMs on the other side. Then since $ASM_n$ is the convex hull of $n\times n$ ASMs, $A$ would necessarily be a vertex. Consider the simple flow grid corresponding to $A$. In any simple flow grid there are, by definition, $2n(n+1)$ directed edges, where for each entry of the corresponding ASM there is a directed edge whenever the partial sum in that direction up to that point equals 1. Since the total number of directed edges in a simple flow grid is fixed, $A$ is the only ASM with all of those partial sums equaling 1. Thus the hyperplane where the sum of those partial sums equals $2n(n+1)-\frac{1}{2}$ will have $A$ on one side and all the other ASMs on the other. Thus the $n\times n$ ASMs are the vertices of $ASM_n$. \end{proof} Another interesting property of the ASM polytope is its relationship to the permutohedron. For a vector $z = (z_1, z_2, \ldots, z_n)\in\mathbb{R}^n$ with distinct entries, define the permutohedron $P_z$ as the convex hull of all vectors obtained by permuting the entries of $z$. That is, \begin{equation} \label{eq:permut} P_z = \mbox{conv}\{(z_{\omega (1)}, z_{\omega (2)}, \ldots, z_{\omega (n)})\mbox{ }|\mbox{ } \omega \in S_n\}. \end{equation} Also, for such a vector $z$, let $\phi_z$ be the mapping from the set of $n\times n$ real matrices to $\mathbb{R}^n$ defined by \[\phi_z (X)= z X, \mbox{for any $n\times n$ real matrix $X$.}\] It is well known, and follows immediately from the definitions, that $P_z$ is the image of the Birkhoff polytope under the projection $\phi_z$. \begin{proposition} \label{prop:birkpermut} Let $B_n$ be the Birkhoff polytope and $z$ be a vector in $\mathbb{R}^n$ with distinct entries. Then \begin{equation} \phi_z(B_n)=P_z. \end{equation} \end{proposition} This result is one of many classical results about the Birkhoff polytope dating back to Hardy, Littlewood, and P\'olya~\cite{HARDYLP1}~\cite{HARDYLP2}. See~\cite{MIRSKY} for a nice summary of relevant results. The next theorem states that when the same projection map is applied to $ASM_n$, the image is the same permutohedron whenever $z$ is a decreasing vector. For the proof of this theorem we will need the concept of majorization~\cite{MAJORIZATION}. \begin{definition} \label{def:maj} Let $u$ and $v$ be vectors of length $n$. Then $u\preceq v$ (that is $u$ is \emph{majorized} by $v$) if \begin{equation} \begin{cases} \sum_{i=1}^k u_{[i]} \le \sum_{i=1}^k v_{[i]},&\mbox{for } 1\le k\le n-1\\ \sum_{i=1}^n u_i = \sum_{i=1}^n v_i&\end{cases} \end{equation} where the vector $(u_{[1]},u_{[2]},\ldots,u_{[n]})$ is obtained from $u$ by rearranging its components so that they are in decreasing order, and similarly for $v$. \end{definition} \begin{theorem} Let $z$ be a decreasing vector in $\mathbb{R}^n$ with distinct entries. Then \begin{equation} \phi_z(ASM_n)=P_z. \end{equation} \end{theorem} \begin{proof} It follows from Proposition~\ref{prop:birkpermut} and $B_n\subseteq ASM_n$ that $P_z\subseteq \phi_z(ASM_n)$. Thus it only remains to be shown that $\phi_z(ASM_n)\subseteq P_z$. Let $z$ be a decreasing $n$--vector (so that $z_i = z_{[i]}$) and $X = \{x_{ij}\}$ an $n\times n$ ASM. Then there is a proposition of Rado which states that for vectors $u$ and $v$ of length $n$, $u\preceq v$ if and only if $u$ lies in the convex hull of the $n!$ permutations of the entries of $v$~\cite{RADO}. Therefore the proof will be completed by showing $z X\preceq z$. By Definition~\ref{def:maj} we need to show \begin{align} \label{eq:proj1} \sum_{j=1}^k (z X)_{[j]}&\le \sum_{j=1}^k z_{j}, \hspace{.4in} 1\le k\le n-1\\ \label{eq:proj2} \sum_{j=1}^n (z X)_j&= \sum_{j=1}^n z_j \end{align} where the $j$th component $(z X)_j$ of $z X$ is given by $\sum_{i=1}^n z_i x_{ij}$. To verify (\ref{eq:proj2}) note that since $\sum_{j=1}^n x_{ij}=1$, \[\sum_{j=1}^n (z X)_j = \sum_{j=1}^n \sum_{i=1}^n z_i x_{ij} = \sum_{i=1}^n z_i \sum_{j=1}^n x_{ij} = \sum_{i=1}^n z_i.\] To prove (\ref{eq:proj1}) we will show that $\sum_{j\in J} (z X)_{j} \le \sum_{j=1}^{|J|} z_j$ given any $J\subseteq \{1,\ldots, n\}$, so that in particular $\sum_{j=1}^{|J|} (z X)_{[j]}\le \sum_{j=1}^{|J|} z_j$. We will need to verify the following: \begin{align} \label{eq:vim} \sum_{i=1}^m \sum_{j\in J} x_{ij}&\le \min(m,|J|) \hspace{.4in} \forall \mbox{ }m\in \{1,\ldots, n\}.\\ \label{eq:vik} \sum_{i=1}^n \sum_{j\in J} x_{ij}&= |J|. \end{align} To prove (\ref{eq:vim}) note that \[ \sum_{i=1}^m \sum_{j\in J} x_{ij} = \sum_{j\in J} \sum_{i=1}^m x_{ij} \le |J| \] since $\sum_{i=1}^m x_{ij}\le 1$. But also, since $\sum_{i=1}^m x_{ij}\ge 0$ and $\sum_{j=1}^n x_{ij} = 1$ we have that \[ \sum_{j\in J} \sum_{i=1}^m x_{ij} \le \sum_{j=1}^n \sum_{i=1}^m x_{ij} = \sum_{i=1}^m \sum_{j=1}^n x_{ij} = m. \] To prove (\ref{eq:vik}) observe, \[ \sum_{i=1}^n \sum_{j\in J} x_{ij} = \sum_{j\in J} \sum_{i=1}^n x_{ij} = \sum_{j\in J} 1 = |J| \] since the columns of $X$ sum to 1. Therefore using (\ref{eq:vim}) and~(\ref{eq:vik}) we see that \begin{align*} \sum_{i=1}^n z_i x_{ij}&= \sum_{j\in J} \sum_{i=1}^n z_i x_{ij} = \sum_{i=1}^n z_i \sum_{j\in J} x_{ij} = \sum_{k=1}^{n-1} (z_k-z_{k+1}) \sum_{i=1}^k \sum_{j\in J} x_{ij} + z_n \sum_{i=1}^n \sum_{j\in J} x_{ij}\\ &= \sum_{k=1}^{n-1} (z_k-z_{k+1}) \sum_{i=1}^k \sum_{j\in J} x_{ij} + z_n |J| && \text{by~(\ref{eq:vik})}\\ &= \sum_{k=1}^{|J|-1} (z_k-z_{k+1}) \sum_{i=1}^k \sum_{j\in J} x_{ij} + \sum_{k=|J|}^{n-1} (z_k-z_{k+1}) \sum_{i=1}^k \sum_{j\in J} x_{ij} + z_n |J|\\ &\le \sum_{k=1}^{|J|-1} (z_k-z_{k+1}) k + \sum_{k=|J|}^{n-1} (z_k-z_{k+1}) |J| + z_n |J| &&\text{by~(\ref{eq:vim})}\\ &\le \sum_{k=1}^{|J|-1} z_k - z_{|J|} (|J|-1) + (z_{|J|}-z_{n})|J| + z_n |J|\\ &\le \sum_{k=1}^{|J|} z_k. \end{align*} Thus $z X \preceq z$ and so $z X$ is contained in the convex hull of the permutations of $z$. Therefore $\phi_z(ASM_n)=P_z$. \end{proof} \section{The face lattice of the ASM polytope} \label{sec:facelattice} Another nice result about the Birkhoff polytope is the structure of its face lattice~\cite{BILLERA}. Associate to each permutation matrix $X$ a bipartite graph with vertices $u_1,u_2,\ldots,u_n$ and $v_1,v_2,\ldots,v_n$ where there is an edge connecting $u_i$ and $v_j$ if and only if there is a 1 in the $(i,j)$ position of $X$. Such a graph will be a perfect matching on the complete bipartite graph $K_{n,n}$. A graph $G$ is called elementary if every edge is a member of some perfect matching of~$G$. \begin{theorem}[Billera--Sarangarajan] \label{prop:elementary} The face lattice of the Birkhoff polytope is isomorphic to the lattice of elementary subgraphs of $K_{n,n}$ ordered by inclusion. \end{theorem} This lattice structure was first identified by Billera and Sarangarajan in~\cite{BILLERA} and~\cite{BILLERAFPSAC}, but the set of faces itself was first characterized and studied extensively by Brualdi and Gibson in~\cite{BRUALDIGIB1} and~\cite{BRUALDIGIB2} using certain 0-1 matrices which correspond trivially to elementary subgraphs of $K_{n,n}$. Other relevant results were also obtained by Balinski and Russakoff in~\cite{BALRUS1} and~\cite{BALRUS2}. A similar statement can be made about the face lattice of the ASM polytope using simple flow grids (see Definition~\ref{definition:simpleflowgrid}) in place of perfect matchings, the complete flow grid $C_n$ instead of the complete bipartite graph $K_{n,n}$, and \emph{elementary flow grids} in place of elementary graphs. \begin{definition} An \emph{elementary flow grid} $G$ is a subgraph of the complete flow grid $C_n$ such that the edge set of $G$ is the union of the edge sets of simple flow grids. \end{definition} Now for any face $F$ of $ASM_n$ define the grid corresponding to the face, $g(F)$, to be the union over all the vertices of $F$ of the simple flow grids corresponding to the vertices. That is, \[g(F) = \bigcup_{vertices\mbox{ }A\in F} g(A). \] Thus $g(F)$ is an elementary flow grid since its edge set is the union of the edge sets of simple flow grids. Now we wish to define the converse, that is, given an elementary flow grid $G$ we would like to know the corresponding face $f(G)$ of $ASM_n$. Define $f(G)$ to be the convex hull of the vertices of $ASM_n$ whose corresponding simple flow grids are contained in the elementary flow grid $G$. So let \[f(G) = \mbox{conv}\{\mbox{vertices }A\in ASM_n \mbox{ }|\mbox{ } g(A)\subseteq G\}.\] Recall that we can represent each of the facets of $ASM_n$ either as subgraphs of the complete flow grid $C_n$ from which one of the directed edges in the set $S=\{((i\pm 1,j),(i,j))$, $((i,j\pm 1),(i,j))$ $|$ $i,j\in\{2,\ldots,n-1\}\}$ has been removed or from which one of the pairs of directed edges in the set $T=\{\{((1,1)(1,2)),((1,1),(2,1))\}$, $\{((1,n),(1,n-1)),((1,n),(2,n))\}$, $\{((n,1),(n-1,1)),((n,1),(n,2))\}$, $\{((n,n),(n,n-1)),((n,n),(n-1,n))\}\}$ has been removed. Thus each of the directed edges in $S$ and the first of each pair of directed edges in $T$ that are not in $G$ represent facets that contain $f(G)$. Let the collection of these directed edges be called $\{e_1, e_2, \ldots, e_k\}$ and their corresponding facets $\{F_1, F_2, \ldots, F_k\}$. Let $I=\bigcap_{j=1}^k F_j$ be the intersection of these facets. Thus $I$ is a face of $ASM_n$ and $f(G)\subseteq I$. We wish to show that $f(G)$ equals $I$. So suppose $f(G)\subsetneq I$. Then since $I$ is a face of $ASM_n$ and $f(G)$ is defined as the convex hull of vertices of $ASM_n$ there exists an additional vertex $B\in I$ of $ASM_n$ such that $B\notin f(G)$. But $g(B)$ must be missing the directed edges $e_1,e_2,\ldots,e_k$ since $B\in I$, thus all the directed edges of $g(B)$ must be in $G$. Therefore $g(B)\subseteq G$ so that $B\in f(G)$ which is a contradiction. So $f(G)=I$. Thus $f(G)$ is a face of $ASM_n$ since it is the intersection of faces of $ASM_n$. It can easily be seen that $f(g(F))=F$ and $g(f(G))=G$. Also if $F_1$ and $F_2$ are faces of $ASM_n$ then $F_1\subseteq F_2$ if and only if $g(F_1)\subseteq g(F_2)$. Thus elementary flow grids are in bijection with the faces of $ASM_n$ (if we also regard the empty grid as an elementary flow grid). Elementary flow grids can be made into a lattice by inclusion, where the join is the union of the edge sets and the meet is the largest elementary flow grid made up of the directed edges from the intersection of the edges sets. This discussion yields the following theorem: \begin{theorem} \label{thm:facelattice} The face lattice of $ASM_n$ is isomorphic to the lattice of all $n\times n$ elementary flow grids (or equivalently all $n\times n$ square ice configurations with domain wall boundary conditions) ordered by inclusion. \end{theorem} The dimension of any face of $ASM_n$ can be determined by looking at $g(F)$ as in the following theorem. The characterization of edges of $ASM_n$ is analogous to the result for the Birkhoff polytope which states that the graphs representing edges of $B_n$ are the elementary subgraphs of $K_{n,n}$ which have exactly one cycle~\cite{BALRUS2}~\cite{BILLERA}~\cite{BRUALDIGIB2}. Given an elementary flow grid $G$, define a \emph{doubly directed region} as a collection of cells in $G$ completely bounded by double directed edges but containing no double directed edges in the interior (see Figure~\ref{fig:redbluecircuits}). Let $\alpha(G)$ denote the number of doubly directed regions in $G$. \begin{theorem} \label{thm:dim} The dimension of a face $F$ of $ASM_n$ is the number of doubly directed regions in the corresponding elementary flow grid $g(F)$. In particular, the edges of $ASM_n$ are represented by elementary flow grids containing exactly one cycle of double directed edges. \end{theorem} \begin{figure}[htp] \centering \includegraphics[scale=0.25]{redonlycircuitsdwbc.eps} \caption[An elementary flow grid containing 3 doubly directed regions]{An elementary flow grid containing 3 doubly directed regions which corresponds by Theorem~\ref{thm:dim} to a 3-dimensional face of $ASM_5$} \label{fig:redbluecircuits} \end{figure} \begin{proof} We proceed by induction on the dimension of the face of $ASM_n$. The simple flow grid corresponding to any ASM $A$ has no double directed edges, thus $\alpha(g(A))=0$. Now suppose for every $m$--dimensional face of $ASM_n$, the number of doubly directed regions of the elementary flow grid corresponding to the face equals $m$. Let $F$ be an $(m+1)$--dimensional face of $ASM_n$ and $F'$ an $m$--dimensional subface of $F$. We assume $\alpha(g(F')) = m$ and wish to show that $\alpha(g(F)) = m+1$. Now $g(F)$ is the elementary flow grid whose edge set is the union of the edge sets of $g(F')$ and $g(A)$ over all ASMs $A$ in $F-F'$. Every vertex in a simple flow grid must have even indegree and even outdegree. Therefore, if we wish to obtain $g(A)$ from $g(A')$, where $A'$ is an ASM in $F'$, by reversing some directed edges, the number of directed edges reversed at each vertex must be even. Thus taking the union of the directed edges of $g(A')$ with the directed edges of $g(A)$ forms one or more circuits of double directed edges, where at least one of the double directed edges is not in $g(F')$. Therefore $g(F)$ has at least one more doubly directed region than $g(F')$, so $\alpha(g(F))\ge m+1$. Then since $g(ASM_n)$ equals the complete flow grid $C_n$, we have that $\alpha(g(ASM_n)) =\alpha(C_n)= (n-1)^2 = \mbox{dim}(ASM_n)$. Therefore moving up the face lattice one rank increases the number of doubly directed regions by exactly one, so $\alpha(g(F)) = m+1$. \end{proof} See Figure~\ref{fig:redcircuit} for the elementary flow grid representing the edge in $ASM_5$ between \[ \left( \begin{array}{rrrrr} 0 & 1 & 0 & 0 & 0\\ 1 & -1 & 1 & 0 & 0\\ 0&0&0&1&0\\ 0&1&-1&0&1\\ 0&0&1&0&0 \end{array}\right) \mbox{ and } \left( \begin{array}{rrrrr} 0 & 1 & 0 & 0 & 0\\ 1 & -1 & 1 & 0 & 0\\ 0&1&0&0&0\\ 0&0&0&0&1\\ 0&0&0&1&0 \end{array} \right). \] \begin{figure}[htp] \centering \includegraphics[scale=0.25]{redcircuitdwbc.eps} \caption{The elementary flow grid representing an edge in $ASM_5$} \label{fig:redcircuit} \end{figure} \section{Acknowledgments} This work is based on research which is a part of the author's doctoral thesis at the University of Minnesota under the direction of Dennis Stanton. The author would like to thank Professor Stanton for the many helpful discussions and encouragement. \bibliographystyle{amsalpha}

https://arxiv.org/abs/2112.00064

Acute Tours in the Plane

We confirm the following conjecture of Fekete and Woeginger from 1997: for any sufficiently large even number $n$, every set of $n$ points in the plane can be connected by a spanning tour (Hamiltonian cycle) consisting of straight-line edges such that the angle between any two consecutive edges is at most $\pi/2$. Our proof is constructive and suggests a simple $O(n\log n)$-time algorithm for finding such a tour. The previous best-known upper bound on the angle is $2\pi/3$, and it is due to Dumitrescu, Pach and Tóth (2009).

\section{Introduction} The Euclidean traveling salesperson problem (TSP) is a well-studied and fundamental problem in combinatorial optimization and computational geometry. In this problem we are given a set of points in the plane and our goal is to find a shortest tour that visits all points. Motivated by applications in robotics and motion planning, in recent years there has been an increased interest in the study of tours with bounded angles at vertices, rather than bounded length of edges; see e.g. \cite{Aggarwal1999,Aichholzer2017,Dumitrescu12,Fekete1992,Fekete1997} and references therein. Bounded-angle structures (tours, paths, trees) are also desirable in the context of designing networks with directional antennas \cite{Aschner2017,Aschner2012,Carmi2011,Tran2017}. Bounded-angle tours (and paths), in particular, have received considerable attention following the PhD thesis of S. Fekete \cite{Fekete1992} and the seminal work of Fekete and Woeginger \cite{Fekete1997}. Consider a set $P$ of at least three points in the plane. A {\em spanning tour} is a directed Hamiltonian cycle on $P$ that is drawn with straight-line edges. When three consecutive vertices $p_i, p_{i+1},p_{i+2}$ of the tour are traversed in this order, the {\em rotation angle} at $p_{i+1}$ (denoted by $\angle p_ip_{i+1}p_{i+2}$) is the angle in $[0,\pi]$ that is determined by the segments $p_ip_{i+1}$ and $p_{i+1}p_{i+2}$. If all rotation angles in a tour are at most $\pi/2$ then it is called an {\em acute} tour. In 1997, Fekete and Woeginger \cite{Fekete1997} raised many challenging questions about bounded-angle tours and paths. In particular they conjectured that {\em for any sufficiently large even number $n$, every set of $n$ points in the plane admits an acute spanning tour $($a tour with rotation angles at most $\pi/2$$)$}. They stated the conjecture specifically for $n\geqslant 8$. The point set illustrated in Figure~\ref{lower-bound-fig}(a) (also described in \cite{Fekete1997}) shows that the upper bound $\pi/2$ is the best achievable. The conjecture does not hold if $n$ is allowed to be an odd number; for example if the $n$ points are on a line then in any spanning tour one of the rotation angles must be $\pi$. The conjecture also does not hold if $n$ is allowed to be small. For instance the 4-element point set consisting of the 3 vertices of an equilateral triangle with its center, must have a rotation angle $2\pi/3$ in any spanning tour. Also the 6-element point set of Figure~\ref{lower-bound-fig}(b) (also illustrated in \cite{Fekete1997} and \cite{Dumitrescu12}) must have a rotation angle of at least $2\pi/3-\epsilon$ in any spanning tour, for some arbitrary small constant $\epsilon$. \begin{figure}[htb] \centering \setlength{\tabcolsep}{0in} $\begin{tabular}{cc} \multicolumn{1}{m{.47\columnwidth}}{\centering\includegraphics[width=.26\columnwidth]{fig/lower-bound.pdf}} &\multicolumn{1}{m{.53\columnwidth}}{\centering\vspace{0pt}\includegraphics[width=.38\columnwidth]{fig/lower-bound2.pdf}} \\ (a) &(b) \end{tabular}$ \caption{(a) a general lower bound example, and (b) a lower bound example for $6$ points.} \label{lower-bound-fig} \end{figure} In 2009, Dumitrescu, Pach and T{\'{o}}th \cite{Dumitrescu12} took the first promising steps towards proving the conjecture. They confirmed the conjecture for points in convex position. For general point sets, they obtained the first partial result by showing that any point set (with even number of points) admits a spanning tour in which each rotation angle is at most $2\pi/3$. In this paper we prove the conjecture of Fekete and Woeginger for general point sets. \begin{theorem} \label{main-thr} Let $n\geqslant 20$ be an even integer. Then every set of $n$ points in the plane admits an acute spanning tour. Such a tour can be computed in linear time after finding an equitable partitioning of points with two orthogonal lines. \end{theorem} Due to our desire of having a short proof, we prove the conjecture for $n\geqslant 20$. Perhaps with some detailed case analysis one could extend the range of $n$ to a number smaller than $20$. \paragraph{Difficulties towards a proof.} Fekete and Woeginger \cite{Fekete1997} exhibited an arbitrary-large even-size point set for which an algorithm (or a proof technique), that always outputs the longest tour or includes the diameter in the solution, does not achieve an acute tour; the point set is similar to that of Figure~\ref{lower-bound-fig}(b) but has more than $6$ points. This somehow breaks the hope for finding an acute tour by using greedy techniques. Therefore, to prove the conjecture one might need to employ some nontrivial ideas. \subsection{Related problems} Another interesting conjecture of Fekete and Woeginger \cite{Fekete1997} is that any set of points in the plane admits a spanning path in which all rotation angles are at least $\pi/6$.\footnote{This bound is the best achievable as the three vertices of an equilateral triangle together with its center do not admit a path with rotation angles greater than $\pi/6$.} In 2008, B{\'{a}}r{\'{a}}ny, P{\'{o}}r, and Valtr \cite{Barany2009} obtained the first constant lower bound of $\pi/9$, thereby gave a partial answer to the conjecture. The full conjecture was then proved, although not yet written in a paper format, by J. Kyn\v{c}l \cite{Kyncl2019} (see also the note added in the proof of \cite{Barany2009}). Fekete and Woeginger \cite{Fekete1997} showed that any set of points in the plane admits an acute spanning path (where all intermediate rotation angles are at most $\pi/2$). Such a path can be obtained simply by starting from an arbitrary point and iteratively connecting the current point to its farthest among the remaining points. Notice that the resulting path always contains the diameter and by the difficulties mentioned above it cannot be completed to an acute tour. Carmi {et~al.}~\cite{Carmi2011} showed how to construct acute paths with shorter edges; again no guarantee to be completed to an acute tour. Aichholzer {et~al.}~\cite{Aichholzer2013} studied a similar problem with an additional constraint that the path should be {\em plane} (i.e., its edges do not cross each other). Among other results, they showed that any set of points in the plane in general position admits a plane spanning path with rotation angles at most $3\pi/4$. They also conjectured that this upper bound could be replaced by $\pi/2$. The bounded-angle minimum spanning tree (also known as $\alpha$-MST) is a related problem that asks for a Euclidean minimum spanning tree in which all edges incident to every vertex lie in a cone of angle at most $\alpha$. This problem is motivated by replacing omni-directional antennas---in a wireless network---with directional antennas which are more secure, require lower transmission ranges, and cause less interference; see e.g. \cite{Aschner2017,Aschner2012,Biniaz2020,Biniaz2022,Tran2017}. Another related problem (with an objective somewhat opposite to ours) is to minimize the total {\em turning angle} of the tour \cite{Aggarwal1999}.\footnote{The turning angle at a vertex $v$ is the change in the direction of motion at $v$ when traveling on the tour. It is essentially $\pi$ minus the rotation angle at $v$.} Similar problems also studied under {\em pseudo-convex} tours and paths (that make only right turns) \cite{Fekete1997} and {\em reflexivity} of a point set (the smallest number of reflex vertices in a simple polygonalization of the point set) \cite{Ackerman2009,Arkin2003}. The so-called {\em Tverberg cycle} is a cycle with straight-line edges such that the diametral disks\footnote{The diametral disk induced by an edge $pq$ is the disk that has $pq$ as its diameter.} induced by the edges have nonempty intersection. Recently, Pirahmad {et~al.}~\cite{Pirahmad2021} showed how to construct a spanning Tverberg cycle on any set of points in the plane. Although the constructed cycle has many acute angles, it is still far from being fully acute. \paragraph{Remark.}It is worth mentioning that having a tour with many acute angles, does not necessarily help in getting a fully acute tour because one can simply get a tour with at least $n-2$ acute angles by interconnecting the endpoints of acute paths obtained in \cite{Carmi2011,Fekete1997}. \section{Preliminaries for the proof} \label{preliminaries} A set of four points in the plane is called a {\em quadruple}. If the four points are in convex position then the quadruple is called {\em convex}, otherwise it is called {\em concave}; the quadruple in Figure~\ref{hook-fig}(a) is convex while the quadruples in Figures~\ref{hook-fig}(b) and \ref{hook-fig}(c) are concave. We refer to the interior point of a concave quadruple as its {\em center}. By connecting the center of a concave quadruple to its other three points we obtain three angles. If one of these angles is at most $\pi/2$ then the quadruple is called {\em concave-acute}, otherwise all the angles are larger than $\pi/2$ and the quadruple is called {\em concave-obtuse}; the quadruple in Figure~\ref{hook-fig}(b) is concave-acute while the one in Figure~\ref{hook-fig}(c) is concave-obtuse. A path, that is drawn by straight-line edges, is called {\em acute} if all the angles determined by its adjacent edges are at most $\pi/2$. For two directed paths $P_1$ and $P_2$, where $P_1$ ends at the same vertex at which $P_2$ starts, we denote their concatenation by $P_1\oplus P_2$. For two distinct points $p$ and $q$ in the plane, we say that $p$ is {\em to the left of} $q$ if the $x$-coordinate of $p$ is not larger than the $x$-coordinate of $q$. Analogously, we say that $p$ is {\em below} $q$ if the $y$-coordinate of $p$ is not larger than the $y$-coordinate of $q$. It is known that any set of $n$ points in the plane can be split into four parts of equal size using two orthogonal lines (see e.g. \cite{Roy2007} or \cite[Section 6.6]{Courant1979}); such two lines can be computed in $\Theta(n\log n)$ time \cite{Roy2007}. The following is a restatement of this result which is borrowed from \cite{Dumitrescu12}. \begin{lemma} \label{equitable-partition-lemma} Given a set $S$ of $n$ points in the plane $(n$ even$)$, one can always find two orthogonal lines $\ell_1$, $\ell_2$ and a partition $S = S_1 \cup S_2 \cup S_3 \cup S_4$ with $|S_1| = |S_3| = \lfloor \frac{n}{4}\rfloor$ and $|S_2| = |S4| = \lceil \frac{n}{4}\rceil$ such that $S_1$ and $S_3$ belong to two opposite closed quadrants determined by $\ell_1$ and $\ell_2$, and $S_2$ and $S_4$ belong to the other two opposite closed quadrants. \end{lemma} Our proof of Theorem~\ref{main-thr} shares some similarities with that of Dumitrescu {et~al.}~\cite{Dumitrescu12} (for points in convex position) in the sense that both proofs employ the equitable partitioning of Lemma~\ref{equitable-partition-lemma}. However, there are major differences between the two proofs mainly because simple structures, that appear in points in convex position, do not necessarily appear in general point sets. Therefore one needs to extract complex structures from general point sets and combine them to establish a proof. \section{Proof of Theorem~\ref{main-thr}} Throughout this section we assume that $n$ is an even integer. We show how to construct an acute tour on any set of $n\geqslant 20$ points in the plane, and thus proving Theorem~\ref{main-thr}. In Subsection~\ref{setup-section} we describe the setup for our construction, and then in Subsection~\ref{tour-section} we construct the tour. \subsection{The proof setup} \label{setup-section} \iffalse \begin{lemma} \label{convex-quadruple-lemma} Let $p,q,r,s$ be the points of a convex quadruple that appear in this order along the boundary of their convex hull. Then the 4-cycle $prsqp$ or the 4-cycle $prqsp$ is acute. \end{lemma} \begin{proof} Let $x$ denote the intersection point of the diagonals $pr$ and $qs$. If $\angle pxq \geqslant \pi/2$ then the cycle $prsqp$ is acute, otherwise (i.e. if $\angle pxs > \pi/2$) the cycle $prqsp$ is acute. \end{proof} \begin{lemma} \label{acute-quadruple-lemma} Let $\{p,q,r,s\}$ be an acute quadruple with center $s$, and let $\angle psq$ be its acute angle. Then the 4-cycle $psqrp$ is acute. \end{lemma} \begin{proof} Observe that $\angle sqr + \angle qrp + \angle rps = \angle psq \leqslant \pi/2$. Thus, the cycle $psqrp$ is acute. \end{proof} \fi Let $S$ be a set of $n\geqslant 20$ points in the plane. Let $\{S_1, S_2, S_3, S_4\}$ be an equitable partitioning of $S$ with two orthogonal lines $\ell_1$ and $\ell_2$ that satisfies the conditions of Lemma~\ref{equitable-partition-lemma}. After a suitable rotation and translation we may assume that $\ell_1$ and $\ell_2$ coincide with the $x$ and $y$ coordinate axes, respectively. Also, after a suitable relabeling we may assume that all points of $S_i$ belong the $i$th quadrant determined by the axes as depicted in Figure~\ref{hook-fig}(a). \begin{figure}[htb] \centering \setlength{\tabcolsep}{0in} $\begin{tabular}{ccc} \multicolumn{1}{m{.34\columnwidth}}{\centering\includegraphics[width=.28\columnwidth]{fig/convex-quadruple2.pdf}} &\multicolumn{1}{m{.33\columnwidth}}{\centering\vspace{0pt}\includegraphics[width=.28\columnwidth]{fig/acute-quadruple2.pdf}} &\multicolumn{1}{m{.33\columnwidth}}{\centering\vspace{0pt}\includegraphics[width=.28\columnwidth]{fig/obtuse-quadruple-revised.pdf}} \\ (a) Convex quadruple&(b) Concave-acute quadruple &(c) Concave-obtuse quadruple \end{tabular}$ \caption{Illustration of (a) Lemma~\ref{convex-acute-quadruple-lemma} where $P$ is convex and $\angle p_1xp_2\geqslant \pi/2$, (b) Lemma~\ref{convex-acute-quadruple-lemma} where $P$ is concave-acute and $\angle p_1p_2p_4\leqslant \pi/2$, and (c) Lemma~\ref{obtuse-quadruple-lemma} where all the three angles at $s$ are obtuse.} \label{hook-fig} \end{figure} Based on the above partitioning we introduce four types of quadruples. Let $P=\{p_1,p_2,p_3,p_4\}$ be a quadruple such that $p_i\in S_i$ for all $i=1,2,3,4$. We say that $P$ is {\em upward} if the path $p_2p_4p_3p_1$ (or equivalently $p_1p_3p_4p_2$) is acute, {\em downward} if the path $p_3p_1p_2p_4$ (or equivalently $p_4p_2p_1p_3$) is acute, {\em leftward} if the path $p_2p_4p_1p_3$ (or equivalently $p_3p_1p_4p_2$) is acute, and {\em rightward} if the path $p_1p_3p_2p_4$ (or equivalently $p_4p_2p_3p_1$) is acute. Such paths are referred to as ``hooks'' in \cite{Dumitrescu12}. The following lemmas and observation, although very simple, play important roles in our proof. \begin{lemma} \label{convex-acute-quadruple-lemma} Let $P=\{p_1,p_2,p_3,p_4\}$ be a quadruple such that $p_i\in S_i$ for all $i=1,2,3,4$. If $P$ is convex or concave-acute then it is upward and downward or it is leftward and rightward. \end{lemma} \begin{proof} First assume that $P$ is convex. Let $x$ denote the intersection point of the diagonals $p_1p_3$ and $p_2p_4$. If $\angle p_1xp_2 \geqslant \pi/2$ then the paths $p_2p_4p_3p_1$ and $p_3p_1p_2p_4$ are acute and thus $P$ is upward and downward; see Figure~\ref{hook-fig}(a). If $\angle p_1xp_4 > \pi/2$ then the paths $p_2p_4p_1p_3$ and $p_1p_3p_2p_4$ are acute and thus $P$ is leftward and rightward. Now assume that $P$ is concave-acute. Without loss of generality we assume that $p_2$ is the center of $P$. Observe that in this case $\angle p_1p_2p_3$ is obtuse. This and the fact that $P$ is concave-acute imply that one of $\angle p_1p_2p_4$ and $\angle p_3p_2p_4$ is acute. If $\angle p_1p_2p_4$ is acute as depicted in Figure~\ref{hook-fig}(b) then the paths $p_2p_4p_3p_1$ and $p_3p_1p_2p_4$ are acute and thus $P$ is upward and downward (observe that $\angle p_2p_1p_3 + \angle p_1p_3p_4 + \angle p_3p_4p_2 = \angle p_1p_2p_4 \leqslant \pi/2$). Analogously, if $\angle p_3p_2p_4$ is acute then the paths $p_2p_4p_1p_3$ and $p_1p_3p_2p_4$ are acute and thus $P$ is leftward and rightward. \end{proof} \begin{lemma} \label{obtuse-quadruple-lemma} Let $\{p,q,r,s\}$ be a concave-obtuse quadruple with center $s$. Then all angles $\angle pqs$, $\angle qps$, $\angle qrs$, $\angle rqs$, $\angle rps$, and $\angle prs$ are acute. \end{lemma} \begin{proof} See Figure~\ref{hook-fig}(c). In each of the triangles $\bigtriangleup spq$, $\bigtriangleup sqr$, and $\bigtriangleup srp$ the angle at $s$ is larger than $\pi/2$. Thus the other two angles are acute. \end{proof} \begin{lemma} \label{obtuse-quadruple-lemma2} Let $P=\{p_1,p_2,p_3,p_4\}$ be a quadruple such that $p_i\in S_i$ for all $i=1,2,3,4$. If $P$ is concave-obtuse then it is upward, downward, leftward, or rightward. \end{lemma} \begin{proof} Without loss of generality assume that $p_2$ is the center of $P$. See Figure~\ref{hook-fig}(c) where $p_2=s$. In the triangle $\bigtriangleup p_1p_3p_4$ the angle at $p_1$ or the angle at $p_3$ is acute. If the angle at $p_1$ is acute then the path $p_2p_4p_1p_3$ is acute and thus $P$ is leftward ($\angle p_2p_4p_1$ is acute by Lemma~\ref{obtuse-quadruple-lemma}). If the angle at $p_3$ is acute then the path $p_2p_4p_3p_1$ is acute and thus $P$ is upward ($\angle p_2p_4p_3$ is acute by Lemma~\ref{obtuse-quadruple-lemma}). \end{proof} \begin{observation} \label{opposite-quadrantobs} Let $p$, $q$, and $r$ be any three points in $S$ such that $q$ and $r$ lie in the quadrant that is opposite to the quadrant containing $p$. Then the angle $\angle qpr$ is acute. \end{observation} \subsection{The tour construction} \label{tour-section} In this section we show how to construct an acute tour on $S$ where $|S|\geqslant 20$. By Lemma~\ref{equitable-partition-lemma} each $S_i$ with $i\in\{1,2,3,4\}$ has at least $\lfloor 20/4\rfloor=5$ points. From each $S_i$ we select an arbitrary subset of 5 points, and then we partition (the total 20) selected points into 5 quadruples such that each quadruple contains exactly one point from each $S_i$. Let $\cal Q$ denote the set of these quadruples. For any quadruple $X$ in $\cal Q$ we denote the points of $X$ by $x_1,x_2,x_3,x_4$ where $x_i\in S_i$ for all $i=1,2,3,4$. Since $|{\cal Q}|\geqslant 5$, by the pigeonhole principle $\cal Q$ has three quadruples that are {\em vertical} (i.e. upward, downward, or both upward and downward) or three that are {\em horizontal} (i.e. leftward, rightward, or both leftward and rightward). Without loss of generality assume that $Q$ has three vertical quadruples. If two of these vertical quadruples are of opposite types, i.e. one upward and one downward, then we construct a tour as in case 1 below. Otherwise, the three quadruples are concave-obtuse and of the same type in which case we construct a tour as in case 2 below. Our constructions take linear time in both cases. \begin{wrapfigure}{r}{0.45\textwidth} \begin{center} \vspace{-22pt} \includegraphics[width=.42\textwidth]{fig/convex-proof3.pdf} \end{center} \vspace{-15pt} \caption{Illustration of Case 1.} \label{case-i} \vspace{-5pt} \end{wrapfigure} \paragraph{Case 1:} $\cal Q$ {\em contains two quadruples such that one is upward and the other is downward.} Let $P$ and $Q$ be such quadruples where $P$ is upward and $Q$ is downward. Since $P$ is upward, the path $p_1p_3p_4p_2$ is acute. Since $Q$ is downward, the path $q_4q_2q_1q_3$ is acute; see Figure~\ref{case-i}. Let $\overline{S_2S_4}$ be a polygonal path starting from $p_2$, ending in $q_4$, alternating between $S_2$ and $S_4$, and containing all points of $S_2\cup S_4$ except for $q_2$ and $p_4$. Let $\overline{S_3S_1}$ be a polygonal path starting from $q_3$, ending in $p_1$, alternating between $S_3$ and $S_1$, and containing all points of $S_3\cup S_1$ except for $p_3$ and $q_1$. Such polygonal paths exist because by Lemma~\ref{equitable-partition-lemma} we have $|S_2|=|S_4|$ and $|S_1|=|S_3|$. All intermediate angles of these two polygonal paths are acute by Observation~\ref{opposite-quadrantobs}. Then the tour $p_1p_3p_4p_2\oplus\overline{S_2S_4}\oplus q_4q_2q_1q_3\oplus\overline{S_3S_1}$ is acute, and it spans $S$. Notice that the angles at $p_1$, $p_2$, $q_3$ and $q_4$ are acute by Observation~\ref{opposite-quadrantobs}. \paragraph{Case 2:}$\cal Q$ {\em contains three concave-obtuse quadruples of the same type.} Let $P$, $Q$ and $R$ be such quadruples, and without loss of generality assume that they are upward. Thus, the paths $p_2p_4p_3p_1$ and $q_2q_4q_3q_1$ and $r_2r_4r_3r_1$ are acute. Since $P$, $Q$ and $R$ are concave-obtuse their centers should lie at endpoints of these paths (the centers cannot be interior vertices of acute paths). Thus the center of $P$ is either $p_1$ or $p_2$, the center of $Q$ is either $q_1$ or $q_2$, and the center of $R$ is either $r_1$ or $r_2$. This means that the centers lie in quadrants 1 and 2. By the pigeonhole principle, and after a suitable reflection, we may assume that at least two of the centers lie in quadrant 2. After a suitable relabeling assume that the centers of $P$ and $Q$ (i.e. $p_2$ and $q_2$) lie in quadrant 2. The center of $R$ lies either in quadrant 2 (i.e. it is $r_2$) or in quadrant 1 (i.e. it is $r_1$). After a suitable relabeling assume that $p_2$ lies below $q_2$, as in Figure~\ref{proof-fig2}. Now we build our tour as follows. First we connect $p_2$ to $p_1$ and $q_1$. The point $p_2$ is below $p_1$ because $p_2$ lies below the segment $p_1p_3$. The point $p_2$ is also below $q_1$ because $p_2$ is below $q_2$ which is in turn below $q_1$ (as $q_2$ lies below the segment $q_1q_3$). Thus $p_2$ is below both $p_1$ and $q_1$. Also notice that $p_2$ is to the left of both $p_1$ and $q_1$. Thus, the angle $\angle p_1p_2q_1$ is acute (imagine moving the origin to $p_2$, then both $p_1$ and $q_1$ would lie in the first quadrant). Then we connect $q_3$ to $q_1$ and $q_4$. The angle $\angle q_4q_3q_1$ is acute because $Q$ is upward (i.e. the path $q_2q_4q_3q_1$ is acute). The angle $\angle p_2q_1q_3$ is acute because both $p_2$ and $q_3$ lie below and to the left of $q_1$. Therefore, the path $p_1p_2q_1q_3q_4$ is acute; see Figure~\ref{proof-fig2}. In the rest of the construction we distinguish two subcases, depending on the center of $R$. \begin{figure}[htb] \centering \setlength{\tabcolsep}{0in} $\begin{tabular}{cc} \multicolumn{1}{m{.5\columnwidth}}{\centering\vspace{0pt}\includegraphics[width=.45\columnwidth]{fig/concave-proof3-revised.pdf}} &\multicolumn{1}{m{.5\columnwidth}}{\centering\vspace{0pt}\includegraphics[width=.45\columnwidth]{fig/concave-proof2-revised.pdf}} \\ (a) &(b) \end{tabular}$ \caption{Illustration of Case 2. Three concave-obtuse quadruples $P$, $Q$ and $R$ that are upward, and the centers of $P$ and $Q$ lie in quadrant 2. (a) Subcase 2.1 where the center of $R$ is in quadrant 1. (b) Subcase 2.2 where the center of $R$ is in quadrant 2.} \label{proof-fig2} \end{figure} \paragraph{Subcase 2.1:} {\em The center of $R$ is $r_1$.} This case is depicted in Figure~\ref{proof-fig2}(a). We connect $r_4$ to $r_2$ and $r_3$. The resulting path $r_2r_4r_3$ is acute (because $R$ is upward, i.e. the path $r_2r_4r_3r_1$ is acute). Let $\overline{S_4S_2}$ be a polygonal path starting from $q_4$, ending in $r_2$, alternating between $S_4$ and $S_2$, containing all points of $S_4\cup S_2$ except for $r_4,p_2$, and having $q_4q_2$ as its first edge. Let $\overline{S_3S_1}$ be a polygonal path starting from $r_3$, ending in $p_1$, alternating between $S_3$ and $S_1$, containing all points of $S_3\cup S_1$ except for $q_3,q_1$, and having $r_3r_1$ as its first edge and $p_3p_1$ as its last edge. All intermediate angles of these two paths are acute by Observation~\ref{opposite-quadrantobs}. By interconnecting the constructed paths we obtain the tour $p_1p_2q_1q_3q_4\oplus\overline{S_4S_2}\oplus r_2r_4r_3\oplus\overline{S_3S_1}$ which is acute, and it spans $S$. The angles at $p_1,r_3,q_4$ are acute by Lemma~\ref{obtuse-quadruple-lemma}, and the angle at $r_2$ is acute by Observation~\ref{opposite-quadrantobs}. \paragraph{Subcase 2.2:} {\em The center of $R$ is $r_2$.} This case is depicted in Figure~\ref{proof-fig2}(b). We connect $r_3$ to $r_4$ and $r_1$. The resulting path $r_4r_3r_1$ is acute (because $R$ is upward, i.e. the path $r_2r_4r_3r_1$ is acute). Let $\overline{S_4S_2S_4}$ be a polygonal path starting from $q_4$, ending in $r_4$, alternating between $S_4$ and $S_2$, containing all points of $S_4\cup S_2$ except for $p_2$, and having $q_4q_2$ as its first edge and $r_2r_4$ as its last edge. Let $\overline{S_1S_3S_1}$ be a polygonal path starting from $r_1$, ending in $p_1$, alternating between $S_1$ and $S_3$, containing all points of $S_1\cup S_3$ except for $q_1,q_3,r_3$, and having $p_3p_1$ as its last edge. Intermediate angles of these paths are acute by Observation~\ref{opposite-quadrantobs}. Thus $p_1p_2q_1q_3q_4\oplus\overline{S_4S_2S_4}\oplus r_4r_3r_1\oplus\overline{S_1S_3S_1}$ is an acute spanning tour. The angles at $q_4$, $r_4$, and $p_1$ are acute by Lemma~\ref{obtuse-quadruple-lemma}, and the angle at $r_1$ is acute by Observation~\ref{opposite-quadrantobs}. This finishes our proof of Theorem~\ref{main-thr}. \section{Concluding remarks} We showed how to construct an acute tour on any set of $n$ points in the plane, where $n$ is even and at least $20$. Our construction uses at most 12 points in each case (namely the points of quadruples $P$, $Q$ and $R$). One might be interested to extend the range of $n$ (to smaller even numbers) by taking advantage of the $8$ unused points, although this may require some case analysis. \paragraph{Acknowledgement.} I am very grateful to the anonymous SoCG 2022 reviewer who meticulously verified our proof, and provided valuable feedback that reduced the number of subcases to two (which was three in our original proof) and improved the bound on $n$ to $20$ (which was $36$ originally). \bibliographystyle{abbrv}

https://arxiv.org/abs/1502.07904

On commuting $U$-operators in Jordan algebras

Recently J.A.Anquela, T.Cortés, and H.Petersson proved that for elements $x, y$ in a non-degenerate Jordan algebra $J$, the relation $x \circ y = 0$ implies that the $U$-operators of $x$ and $y$ commute: $U_xU_y = U_yU_x$. We show that the result may be not true without the assumption on non-degeneracity of $J$. We give also a more simple proof of the mentioned result in the case of linear Jordan algebras, that is, when $char\, F\neq 2$.

\section{An Introduction} \hspace{\parindent} In a recent paper \cite{ACP} J.\,A.\,Anquela, T.\,Cort\'es, and H.\,Petersson have studied the following question for Jordan algebras: (1) does the relation $x\circ y = 0$ imply that the quadratic operators $U_x$ and $U_y$ commute? They proved that the answer is positive for non-degenerate Jordan algebras, and left open the question in the general case, not assuming nondegeneracy. \smallskip We show that the answer to question (1) is negative in general case. We give also a more simple proof of the result for linear non-degenerate Jordan algebras, that is, over a field $F$ of characteristic $\neq 2$. Unless otherwise stated, we will deal with associative and Jordan algebras over a field of arbitrary characteristic. \section{A counter-example} \hspace{\parindent} Let us recall some facts on Jordan algebras. We use as general references the books \cite{Jac, ZSSS, McC}, and the paper \cite{McC1}. \smallskip Consider the free special Jordan algebra $SJ[x,y,z]$ and the free associative algebra $F\langle x,y,z\rangle$ over a field $F$. Let $*$ be the involution of ${F\langle x,y,z\rangle}$ identical on the set $\{x,y,z\}$. Denote $\{u\}=u+u^*$ for $u\in {F\langle x,y,z\rangle}$, then $\{u\}\in{SJ[x,y,z]}$ \cite{Jac, ZSSS} (see also \cite{McC1} for the case of characteristic 2). Below $ab$ will denote the associative product in ${F\langle x,y,z\rangle}$, so that $a\circ b=ab+ba$ and $aU_b=bab$ are the corresponding linear and quadratic operations in ${SJ[x,y,z]}$. \smallskip For an ideal $I$ of $SJ[x,y,z]$, let $\hat I$ denote the ideal of $F\langle x,y,z\rangle$ generated by $I$. By Cohn's Lemma \cite[lemma 1.1]{Jac} (see also \cite[Corollary to Cohn's Criterion]{McC1}), the quotient algebra $J={SJ[x,y,z]}/I$ is special if and only if $I=\hat{I}\cap {SJ[x,y,z]}$. \begin{Lem}\label{z[Ux,Uy]} The following equality holds in ${SJ[x,y,z]}\subseteq {F\langle x,y,z\rangle}$: $$ z[U_x,U_y]=\{(x\circ y)zxy\}-zU_{x\circ y}. $$ \end{Lem} {\it Proof. } We have in ${F\langle x,y,z\rangle}$ \begin{eqnarray*} z[U_x,U_y]&=&yxzxy-xyzyx=(y\circ x)zxy-xyzxy-xyzyx=\\ &=&(y\circ x)zxy-xyz(x\circ y)=\{(x\circ y)zxy\}-(x\circ y)z(x\circ y). \end{eqnarray*} \hfill$\Box$ \begin{Th} Let $I$ denote the ideal of ${SJ[x,y,z]}$ generated by $x\circ y=xy+yx$ and $J={SJ[x,y,z]}/I$. Then for the images $\bar x,\bar y$ of the elements $x,y$ in $J$ we have $\bar x \circ \bar y=0$ but $[U_{\bar x},U_{\bar y}]\neq 0$. \end{Th} {\it Proof. } It suffices to show that $k=z[U_x,U_y]\notin I$. By lemma 1, $k=\{(x\circ y)zxy\}\ (mod\ I)$. Now, the arguments from the proof of \cite[theorem 1.2]{Jac}, show that $k\notin I$ when $F$ is a field of characteristic not 2 (see also \cite[exercise 1, page 12]{Jac}). \smallskip The result is also true in characteristic 2 for quadratic Jordan algebras. In this case, one needs certain modifications concerning the generation of ideals in quadratic case. The author is grateful to T.\, Cort\'es and J.\,A.\, Anquela who corrected the first ``naive'' author's proof and suggested the proper modifications which we give below. \smallskip We have to prove that $\{(x \circ y)zxy\} \not\in I$. By \cite[(1.9)]{NMcC}, the ideal $I$ is the outer hull of $F(x\circ y)+U_{x\circ y} \widehat{SJ[x,y,z]}$, where $\widehat J$ denotes the unital hull of $J$. Assume that there exists a Jordan polynomial $f(x, y, z, t) \in SJ[x, y, z, t]$ with all of its Jordan monomials containing the variable $t$, such that $\{(x \circ y)zxy\} = f(x, y, z, x \circ y)$. By degree considerations, $f = g + h$, where $g, h \in SJ[x, y, z, t]$, $g$ is multilinear, and $h(x, y, z, t)$ is a linear combination of $U_t z$ and $z\circ t^2$. On the other hand, arguing as in \cite[Theorem 1.2]{Jac}, $g \in SJ[x, y, z, t] \subseteq H(F\langle x, y, z, t\rangle, *)$, and because of degree considerations and the fact that $z$ occupies inside position in the associative monomials of $\{(x\circ y)zxy\}$, $g$ is a linear combination of $$ \{xzyt\}, \{xzty\}, \{tzxy\}, \{tzyx\}, \{yztx\}, \{yzxt\}, $$ and $h$ is a scalar multiple of $U_t z$. Hence $f$ has the form \begin{eqnarray*} f(x, y, z, t) &=& \alpha_1\{xzyt\} + \alpha_2\{xzty\} + \alpha_3\{tzxy\}\\ &+& \alpha_4\{tzyx\} + \alpha_5\{yztx\} + \alpha_6\{yzxt\}\\ &+& \alpha_7 tzt, \end{eqnarray*} and therefore \begin{eqnarray*} \{(x \circ y)zxy\} &=& \alpha_1\{xzy(x\circ y)\} + \alpha_2\{xz(x\circ y)y\} + \alpha_3\{(x\circ y)zxy\}\\ &+& \alpha_4\{(x\circ y)zyx\} + \alpha_5\{yz(x\circ y)x\} + \alpha_6\{yzx(x\circ y)\}\\ &+& \alpha_7 (x\circ y)z(x\circ y), \end{eqnarray*} Comparing coefficients as in \cite[Theorem 1.2]{Jac}, we get \begin{eqnarray*} \alpha_1 = \alpha_2 = \alpha_5 = \alpha_6 = 0,\\ \alpha_3 = \lambda + 1, \alpha_4 = \lambda, \alpha_7 = - 2\lambda, \end{eqnarray*} for some $\lambda\in F$. Going back to $f$, we get $$ f = (\lambda + 1)\{tzxy\} + \lambda\{tzyx\} - 2 \lambda tzt = \{tzxy\} + \lambda\{tz(x\circ y)\} - 2\lambda U_t z, $$ so that $\{tzxy\} \in SJ[x, y, z, t]$, which is a contradiction. In fact, the standard arguments with the Grassmann algebra do not work in characteristic 2, to prove that $\{tzxy\} \notin SJ[x, y, z, t]$, but one can check directly (or with aid of computer) that the space of symmetric multilinear elements in $F{\langle} x,y,z,t{\rangle}$ has dimension 12 while the similar space of Jordan elements has dimension 11. \hfill$\Box$ \section{The non-degenerate case} \hspace{\parindent} Here we will give another proof of the main result from \cite{ACP} that the answer to question (1) is positive for nondegenerate algebras, in the case of linear Jordan algebras (over a field $F$ of characteristic $\neq 2$). Let J be a linear Jordan algebra, $a\in J,\ R_a : x \mapsto xa$ be the operator of right multiplication on $a$, and $U_a = 2R^2_a- R_{a^2}$. As in \cite{ACP}, due to the McCrimmon-Zelmanov theorem \cite{MZ}, it suffices to consider Albert algebras. We will need only the fact that an Albert algebra $A$ is {\em cubic}, that is, for every $a\in A$, holds the identity \begin{eqnarray*} a^3=t(a)a^2-s(a)a+n(a), \end{eqnarray*} where $t(a), s(a), n(a)$ are linear, quadratic, and cubic forms on $A$, correspondingly \cite{Jac}. Linearizing the above identity on $a$, we get the identity \begin{eqnarray*} 2((ab)c+(ac)b+(bc)a)&=&2(t(a)bc+t(b)ac+t(c)ab)\\ &&-s(a,b)c-s(a,c)b-s(b,c)a+n(a,b,c), \end{eqnarray*} where $s(a,b)=s(a+b)-s(a)-s(b)$ and $n(a,b,c)=n(a+b+c)-n(a+b)-n(a+c)-n(b+c)+n(a)+n(b)+n(c)$ are bilinear and trilinear forms. In particular, we have \begin{eqnarray}\label{aab} a^2b+2(ab)a=t(b)a^2+2t(a)ab-s(a,b)a-s(a)b+\tfrac12n(a,a,b). \end{eqnarray} \begin{Lem}\label{lem[Ua,Ub]} Let $a,b\in J$ with $ab=0$. Then $[U_a,U_b]=[R_{a^2},R_{b^2}]$. \end{Lem} {\it Proof. } Linearizing the Jordan identity $[R_x,R_{x^2}]=0$, one obtains $$ [R_{a^2},R_b]=-2[R_{ab},R_a]=0, $$ and similarly $[R_a,R_{b^2}]=0$. Therefore, \begin{eqnarray*} [U_a,U_b]=[2R_a^2-R_{a^2},2R_b^2-R_{b^2}]=4[R_a^2,R_b^2]+[R_{a^2},R_{b^2}]. \end{eqnarray*} Furthermore, $[R_a^2,R_b^2]=[R_a,R_b^2R_a+R_aR_b^2]$. By the operator Jordan identity \cite[(1.$O_2$)]{Jac}, \begin{eqnarray*} R_b^2R_a+R_aR_b^2=-R_{(ba)b}+2R_{ab}R_b+R_{b^2}R_a=R_{b^2}R_a, \end{eqnarray*} therefore $[R_a^2,R_b^2]=[R_a,R_{b^2}R_a]=[R_a,R_{b^2}]R_a=0$, which proves the lemma. \hfill$\Box$ \begin{Th} Let $J$ be a cubic Jordan algebra and $a,b\in J$ with $ab=0$. Then $[U_a,U_b]=0$. \end{Th} {\it Proof. } For any $c\in J$ we have by Lemma 2 and by the linearization of the Jordan identity $(x,y,x^2)=0$ \begin{eqnarray*} c[U_a,U_b]=c[R_{a^2},R_{b^2}]=(a^2,c,b^2)=-2(a^2b,c,b). \end{eqnarray*} By (\ref{aab}), we have \begin{eqnarray*} (a^2b,c,b)&=&t(b)(a^2,c,b)-s(a)(b,c,b)-s(a,b)(a,c,b)\\ &=&-2t(b)(ab,c,a)-s(a,b)(a,c,b)=-s(a,b)(a,c,b). \end{eqnarray*} Substituting $c=a$, we get $(a^2b,a,b)=((a^2b)a)b=(a^2(ba))b=0$, which implies $0=s(a,b)(a,a,b)=s(a,b)(a^2b)$. Therefore, $s(a,b)=0$ or $a^2b=0$. In both cases this implies $c[U_a,U_b]=0$. \hfill$\Box$ \begin{Cor} In an Albert algebra $A$, the equality $ab=0$ implies $[U_a,U_b]=0$. \end{Cor} In connection with the counter-example above, we would like to formulate an open question. Let $f,g\in SJ[x,y,z]$ such that $g\in \widehat{(f)}$ but $g\not\in (f)$, where $(f)$ and $\widehat{(f)}$ are the ideals generated by $f$ in $SJ[x,y,z]$ and in $F{\langle} x,y,z{\rangle}$, respectively. Then the quotient algebra $SJ[x,y,z]/(f)$ is not special, due to Cohn's Lemma. It follows from the results of \cite{Z} that the quotient algebra $\widehat{(f)}/(f)$ is degenerated. The question we want to ask is the following: \smallskip {\em If $f=0$ in a nondegenerate Jordan algebra $J$, should also be $g=0$?} \smallskip Of course, there is a problem of writing $f$ and $g$ in an arbitrary Jordan algebra, we know only what they are in $SJ[x,y,z]$, but in the free Jordan algebra $J[x,y,z]$ they have many pre-images (up to $s$-identities), and one may choose pre-images for which the question has a negative answer. For example, the answer is probably negative for $f=x\circ y$ and $g=z[U_x,U_y]+G(x,y,z)$, where $G(x,y,z)$ is the Glennie $s$-identity \cite{Jac}. So we modify our question in the following way: {\it In the situation as above, is it true that there exists $g'\in J[x,y,z]$ such that $g - g'$ is an $s$-identity and $f=0$ implies $g'=0$ in non-degenerate Jordan algebras?} \section{Acknowledgements} The author acknowledges the support by FAPESP, Proc.\,2014/09310-5 and CNPq, Proc.\,303916/ 2014-1. He is grateful to professor Holger Petersson for useful comments and suggestions, and to professors Jos\'e \'Angel Anquela and Teresa Cort\'es for correction the proof of Theorem 1 in the case of characteristic 2. He thanks all of them for pointing out some misprints.

https://arxiv.org/abs/1512.04321

Cohomological characterizations of the complex projective space

In this survey, we discuss whether the complex projective space can be characterized by its integral cohomology ring among compact complex manifolds.

\section{Introduction} Our starting point is the following 1957 result from \cite{hk}. \begin{theo}[Hirzebruch--Kodaira, Yau] Any compact K\"ahler manifold which is homeomorphic to ${\mathbf{CP}}^n$ is biholomorphic to ${\mathbf{CP}}^n$. \end{theo} Actually, Hirzebruch--Kodaira proved this result under the additional assumptions that the K\"ahler manifold is {\em diffeomorphic} to ${\mathbf{CP}}^n$ and that, when $n$ is even, $c_1(T_X)$ is not $n+1$ times a negative generator of $H^2(X,{\mathbf Z})$. The first assumption was dropped (see \cite{mor}) when Novikov proved in \cite{nov} that Pontryagin classes are invariant under homeomorphisms; the second assumption was also dropped later thanks to work of Yau (\cite{yau}). If one does not assume that $X$ is K\"ahler, the conclusion still holds for $n\le 2$ (complex surfaces with even first Betti number are K\"ahler by \cite{buc,lam}), but nothing is known for $n\ge 3$; if the K\"ahler assumption can be dropped when $n=3$, the sphere ${\mathbf S}^6$ has no complex structure. Stronger characterizations were proved in dimensions $n\le 6$ by Fujita (\cite{fuj1}) and Libgober--Wood (\cite{lw}) assuming only that the K\"ahler manifold has the homotopy type of ${\mathbf{CP}}^n$. Looking carefully through their arguments, it is not too difficult to extract a proof of the following stronger result. \begin{theo} Let $n$ be an integer with $n\le 6$. Any compact K\"ahler manifold with the same integral cohomology ring as ${\mathbf{CP}}^n$ is \begin{itemize} \item either isomorphic to ${\mathbf{CP}}^n$; \item or a quotient of the unit balls ${\bf B}^4$ or ${\bf B}^6$. \end{itemize} \end{theo} No quotients of even dimensional unit balls ${\bf B}^{2m}$ with the same integral cohomology rings as ${\mathbf P}^{2m}$ are known (all known examples have torsion in $H^2$). It is therefore legitimate to ask the following question. \begin{ques*} Is any compact K\"ahler manifold with same integral cohomology ring as ${\mathbf{CP}}^n$ isomorphic to ${\mathbf{CP}}^n$? \end{ques*} The methods used in the proof of the theorem above are completely computational and it seems unlikely that they can be generalized to higher dimensions (we obtain only partial results in dimension 7 in Theorem \ref{th7}). Using geometrical arguments would perhaps be a good idea to make further progress. \subsection*{Acknowledgements} All the computations were done with the software {\tt Sage} (\cite{sage}). Many thanks to Pierre Guillot for his computations of the polynomials $t_n$ in Section \ref{seco}. \section{Preliminaries} From now on, we will write ${\mathbf P}^n$ instead of ${\mathbf{CP}}^n$ for the complex projective space of dimension $n$. \subsection{Hirzebruch--Riemann--Roch}\label{se1.1} Let $X$ be a projective complex manifold of dimension $n$. Following \cite{hir}, we set \begin{equation*} \chi^p(X):=\sum_{q=0}^n(-1)^qh^{p,q}(X)=\chi(X,\Omega_X^p) \end{equation*} and we define the $\chi_y$-genus \begin{equation*} \chi_y(X):=\sum_{p=0}^n \chi^p(X) y^p=\sum_{p, q=0}^n(-1)^qh^{p,q}(X)y^p \in {\mathbf Z}[y]. \end{equation*} For instance, $\chi_0(X)=\chi(X,\mathcal{O}_X )$ and $\chi_{-1}(X)=\chi_{\rm top}(X )$. One consequence of the Hirzebruch--Riemann--Roch theorem is that the coefficients $\chi^p(X)$ of the polynomial $\chi_y(X)$ can be expressed in terms of the Chern classes of $X $ (\cite[Section~IV.21.3,~(10)]{hir}) \begin{equation*} \chi^p(X)=T_n^p(c_1(X),\dots,c_n(X)) \quad\text{or}\quad \chi_y(X)=T_n(y;c_1(X),\dots,c_n(X)), \end{equation*} where $T_n(y;c_1,\dots,c_n):= \sum_{p=0}^nT_n^p(c_1,\dots,c_n)y^p$. These polynomials satisfy $T^p_n=(-1)^nT_n^{n-p}$ and they can be explicitly determined (\cite[Section~I.1.8,~(10)]{hir}). For example, the constant terms $T_n^0(c_1,\dots,c_n)$ (which are also $(-1)^n$ times the leading term) are the Todd polynomials $\td_n(c_1,\dots,c_n)$ (\cite[Section~I.1.7,~(10)]{hir}) and $T_n (-1;c_1,\dots,c_n)=c_n$, so that \begin{equation}\label{cn0} c_n(X)=\chi_{\rm top}(X) . \end{equation} Libgober--Wood also introduce the polynomials \begin{equation}\label{tn} t_n(z;c_1,\dots,c_n):=T_n(z-1;c_1,\dots,c_n) . \end{equation} They show (\cite[Lemma~2.2]{lw}) \begin{equation}\label{tnn} t_n(z;c_1,\dots,c_n) =c_n-\tfrac12nc_nz+\tfrac{1}{12}\bigl( \tfrac12 n(3n-5)c_n+c_1c_{n-1}\bigr)z^2 +\cdots \end{equation} and they compute these polynomials for $n\le 6$ (\cite[p.\ 145]{lw}). We extend their computations to all $n\le9$ in Section~\ref{seco}. \subsection{Compact K\"ahler manifolds with same Betti numbers as ${\mathbf P}^n$} Let $X$ be a compact K\"ahler manifold with the same Betti numbers as ${\mathbf P}^n$. Since $X$ is K\"ahler, one can compute the numbers $h^{p,q}(X)$ from its Betti numbers, and we see that $h^{p,q}(X)= h^{p,q}({\mathbf P}^n)=1$ if $p=q\in\{0,\dots,n\}$, and $h^{p,q}(X)=0$ otherwise. In particular, $X$ is projective (Kodaira). Setting $c_i(X):=c_i(T_X)\in H^{2i}(X,{\mathbf Z})$, we deduce from the Hirzebruch--Riemann--Roch theorem (Section~\ref{se1.1}) the equalities \begin{multline}\label{tnpn} \qquad t_n(z;c_1(X),\dots,c_n(X))=t_n(z;c_1({\mathbf P}^n),\dots,c_n({\mathbf P}^n)) \\{}=t_n\bigl(z;\tbinom{n+1}{1},\dots,\tbinom{n+1}{n}\bigr)=\sum_{i=0}^n\tbinom{n+1}{i+1}(-1)^iz^i.\qquad \end{multline} In particular, \eqref{tnn} implies \begin{equation}\label{cn} c_n(X)=c_n({\mathbf P}^n)=n+1\ ,\ c_1(X)c_{n-1}(X)=c_1({\mathbf P}^n)c_{n-1}({\mathbf P}^n)=\tfrac12n(n+1)^2. \end{equation} Assume now $c_1(X)<0$,\footnote{In the sense that the image of $c_1(X)$ in $H^2(X,{\mathbf R})$ is a negative multiple of the class of a K\"ahler metric.} so that $X$ is of general type. We have (\cite[Remark (iii)]{yau}) \begin{equation}\label{yauu} \begin{array}{c} \bigl(\tfrac{2(n+1)}{n}c_2(X)-c_1(X)^2\bigr)\cdot (-c_1(X))^{n-2}\ge 0 \\ \hbox{with equality if and only if $X$ is covered by the unit ball in ${\mathbf C}^n$,} \end{array} \end{equation} in which case, by the Hirzebruch proportionality principle, all the Chern numbers of $X$ are the same as those of ${\mathbf P}^n$ and $n$ is even. If on the other hand $c_1(X)>0$, so that $X$ is a Fano manifold, the group $\Pic(X)\simeq H^2(X,{\mathbf Z})$ is torsion-free (\cite[Proposition 2.1.2]{ip}) and we write $K_X =-c_1L$, where $L$ is an ample generator of $\Pic(X)$. We have (\cite{ko}) \begin{equation}\label{koo} c_1\le n+1 \quad \hbox{with equality if and only if $X\simeq {\mathbf P}^n$.} \end{equation} \subsection{Compact K\"ahler manifolds with same integral cohomology ring as ${\mathbf P}^n$} Assume now that $X$ has the same integral cohomology ring as ${\mathbf P}^n$. We have $\Pic(X)={\mathbf Z} L$, where $L$ is ample with $L^n=1$, and $\ell:=c_1(L)$ generates $H^2(X,{\mathbf Z}) $. We define integers $c_1,\dots,c_n$ by setting $c_i(X)=c_i\ell^i$ and we compute Euler characteristic s using the Hirzebruch--Riemann--Roch theorem (\cite[Theorem~20.3.2]{hir}) \begin{eqnarray} \chi(X,L^m)&=&\bigl[e^{m\ell}\cdot \Td(X)\bigr]_n = \Td_n(X)+\cdots +\frac{m^{n-2}}{(n-2)!}\Td_2(X) \label{holt}\\ &&\hskip55mm{}+\frac{m^{n-1}}{(n-1)!}\Td_1(X)+ \frac{m^n}{n!},\nonumber\\ \chi(X,T_X\otimes L^m)&=&\bigl[\ch(T_X)\cdot e^{m\ell}\cdot \Td(X)\bigr]_n. \label{tholt}\end{eqnarray} \begin{lemm}\label{l1} We have \begin{eqnarray} c_1-(n+1)&\equiv& 0\pmod2,\label{c1}\\ c_1^2+c_2-3nc_1+\tfrac12(n+1)(3n-2)&\equiv& 0\pmod{12}.\label{c1c2} \end{eqnarray} \end{lemm} \begin{proof} Since the polynomial in $m$ which appears in \eqref{holt} takes integral values at all integers $m$, it decomposes as $$\binom{m+n}{n}+a_1\binom{m+n-1}{n-1}+a_2\binom{m+n-2}{n-2}+\cdots$$ where $a_1,a_2,\dots$ are integers. The coefficient of $m^{n-1}$ is $$\frac1{n!}\sum_{i=1}^ni+\frac1{(n-1)!}a_1=\frac1{(n-1)!}\Td_1(X)=\frac1{(n-1)!}\frac12c_1,$$ hence $\frac{n+1}{2}+a_1=\frac12c_1$. This proves the congruence \eqref{c1}. The coefficient of $m^{n-2}$ is $$\frac1{n!}\sum_{1\le i<j\le n}ij+\frac1{(n-1)!}\sum_{i=1}^{n-1}ia_1+\frac1{(n-2)!}a_2=\frac1{(n-2)!}\Td_2(X)=\frac1{(n-2)!}\frac1{12}(c_1^2+c_2).$$ The first sum is $$\sum_{1\le i<j\le n}\! ij=\sum_{1\le j\le n}\! \frac{j(j-1)}{2}j=\frac12\Big[ \frac{n^2(n+1)^2}{4}-\frac{n(n+1)(2n+1)}{6}\Bigr]=\frac{n(n-1)(n+1)(3n+2)}{24}.$$ We obtain $$\frac1{12}(c_1^2+c_2)=\frac{(n+1)(3n+2)}{24}+\frac{n}{2}\Bigl( \frac12c_1-\frac{n+1}{2}\Bigr)+a_2 $$ and the congruence \eqref{c1c2} follows. \end{proof} \section{Surfaces} \begin{theo}[\cite{yau}] Any compact complex manifold with the same Betti numbers as ${\mathbf P}^2$ is \begin{itemize} \item either isomorphic to ${\mathbf P}^2$; \item or a quotient of the unit ball ${\bf B}^2$. \end{itemize} \end{theo} \begin{proof} A compact complex surface with even first Betti number is K\"ahler (\cite{buc,lam}). Equations \eqref{cn} then give $c_1(X)^2=9 $ and $c_2(X)=3$. If $c_1(X)>0$, the surface $X$ is isomorphic to ${\mathbf P}^2$ by \eqref{koo}. If $c_1(X)<0$, there is equality in \eqref{yauu} and $X$ is a quotient of ${\bf B}^2$. \end{proof} \begin{coro} Any compact complex manifold with the same integral cohomology groups as ${\mathbf P}^2$ is isomorphic to ${\mathbf P}^2$. \end{coro} \begin{proof} Compact quotients $X$ of the unit ball ${\bf B}^2$ are called fake projective planes. They are all classified and it was proved in \cite[Theorem~10.1]{py} that $H_1(X,{\mathbf Z})$ is always nonzero (and torsion). It follows that $H^2(X,{\mathbf Z})_{\rm tors}\simeq H_1(X,{\mathbf Z})_{\rm tors}$ is nonzero, so the integral cohomology groups of fake projective planes are different from those of ${\mathbf P}^2$. \end{proof} \section{Threefolds} In odd dimensions $2m-1$, it is definitely not enough to assume that the Betti numbers of $X$ and ${\mathbf P}^{2m-1}$ are the same: a smooth odd-dimensional quadric $X\subset {\mathbf P}^{2m}$ has this property, but, if $m\ge 2$, a positive generator $L$ of $H^2(X,{\mathbf Z})$ satisfies $L^{2m-1}= 2$, hence $X$ is not even homeomorphic to ${\mathbf P}^{2m-1}$. In dimension 3, there are two other examples of Fano threefolds with the same Betti numbers as ${\mathbf P}^3$ (this is equivalent in that case to $b_2=1$ and $h^{1,2}=0$): one with $L^3=5$ and one with $L^3=22$ (\cite[Table~12.2]{ip}). \begin{theo}[\cite{fuj1,lw}] Any compact K\"ahler manifold with the same integral cohomology ring as ${\mathbf P}^3$ is isomorphic to ${\mathbf P}^3$.\end{theo} \begin{proof} If $\ell$ is a positive generator of $H^2(X,{\mathbf Z})$, we write as before $c_i(X)=c_i\ell^i$. Equations \eqref{cn} give $c_3=4$ and $c_1 c_2 =24$. If $c_1 <0$, we get $c_2 <0$, but this contradicts \eqref{yauu}. Therefore, $c_1 $ is a positive divisor of $24$ which we can assume, by \eqref{koo}, to be 1, 2, 3, or 4. By Lemma \ref{l1}, $c_1$ is even, so we need only exclude $c_1=2$. In that case, $(X,L)$ is a so-called del Pezzo variety (coindex 2) with $L^3=1$. It is therefore isomorphic to a hypersurface of degree 6 in the weighted projective space ${\mathbf P}(3,2,1,1,1)$ (\cite{fuj}, \cite[Theorem 3.2.5]{ip}). But such a variety has $h^{1,2}=21$, so this is a contradiction. \end{proof} \section{Fourfolds} \begin{theo}[\cite{fuj1,lw}] Any compact K\"ahler manifold with the same integral cohomology ring as ${\mathbf P}^4$ is \begin{itemize} \item either isomorphic to ${\mathbf P}^4$; \item or a quotient of the unit ball ${\bf B}^4$. \end{itemize}\end{theo} Four examples of compact quotients $X$ of ${\bf B}^4$ with the same Betti numbers as ${\mathbf P}^4$ are known, but the groups $H_1(X,{\mathbf Z})$ are never zero (\cite[Theorem 4]{py2}) hence they do not have the same integral cohomology ring as ${\mathbf P}^4$. It is therefore possible that the second case of the theorem never occurs. \begin{proof} If $\ell$ is a positive generator of $H^2(X,{\mathbf Z})$, we write as before $c_i(X)=c_i\ell^i$. Equations \eqref{cn} give $c_4=5$ and $c_1 c_3 =50$. By Lemma \ref{l1}, $c_1$ is odd, so by \eqref{koo}, we are reduced to $c_1\in\{ \pm 1,\pm 5, -25\}$. Equation \eqref{cn} reads $c_4 =5 $. Equation \eqref{holt} with $m=0$ then gives (using the values for the Todd polynomials given in Section~\ref{seco}) $$1=\chi(X,\mathcal{O}_X)=\tfrac{1}{720}\bigl(-c_1^4 + 4c_1^2 c_2+ 3 c_2^2 + c_1 c_3- c_4 \bigr)=\tfrac{1}{720}\bigl( -c_1^4 + 4c_1^2 c_2+ 3 c_2^2 + 50-5 \bigr) $$ so $c_2$ is an integral root of the quadratic equation $$3c_2^2+4c_1^2c_2-c_1^4-675=0.$$ Its (reduced) discriminant $ 7c_1^4+2025 $ is therefore a square, which leaves only (among our possible values) the cases $c_1=\pm 5$, $c_2=10$. If $c_1=5$, the fourfold $X$ is isomorphic to ${\mathbf P}^4$ by \eqref{koo}. If $c_1=-5$, there is equality in \eqref{yauu}  and $X$ is a quotient of ${\bf B}^4$. \end{proof} \section{Fivefolds} \begin{theo} [\cite{fuj1,lw}] Any compact K\"ahler manifold with the same integral cohomology ring as ${\mathbf P}^5$ is isomorphic to ${\mathbf P}^5$. \end{theo} \begin{proof} If $\ell$ is a positive generator of $H^2(X,{\mathbf Z})$, we write as before $c_i(X)=c_i\ell^i$. Equations \eqref{cn} give $c_5=6$ and $c_1 c_4 =90$ and, by \eqref{c1}, $c_1$ is even. The relation $\chi(X,\mathcal{O}_X)=T^0_5(c_1,\dots,c_5)$ (Section~\ref{se1.1}) gives $$1=\chi(X,\mathcal{O}_X)=-\tfrac{1}{1440}\bigl(c_1^3 c_2 - 3 c_1 c_2^2 - c_1^2 c_3 + c_1 c_4\bigr)=-\tfrac{1}{1440}\bigl( c_1^3 c_2 - 3 c_1 c_2^2 - c_1^2 c_3 + 90 \bigr) $$ so $c_2$ is an integral root of the quadratic equation \begin{equation}\label{quad} 3c_1c_2^2-c_1^3c_2+c_1^2c_3-1530=0. \end{equation} Moreover, the congruence \eqref{c1c2} reads \begin{equation}\label{cong} c_1^2-3c_1+c_2+3\equiv 0\pmod{12}. \end{equation} If $c_1\equiv 0\pmod9$, we obtain, reducing \eqref{quad} modulo 27, the contradiction $1530\equiv 0\pmod{27}$. Using \eqref{koo}, it follows that the possible values for $c_1$ are $\pm 2, \pm 6 , - 10, - 30$. We rules these cases out one by one. \medskip \noindent{\bf Case $c_1=-30$.} Reducing \eqref{quad} modulo 25, we obtain $3\cdot (-30) c_2^2\equiv 1530\equiv 5\pmod{25}$, hence $1\equiv 3\cdot (-6) c_2^2\equiv 2 c_2^2 \pmod{5}$, which is impossible. \medskip \noindent{\bf Case $c_1=-10$.} Reducing \eqref{quad} modulo 25, we obtain $3\cdot (-10) c_2^2\equiv 1530\equiv 5\pmod{25}$, hence $ c_2^2 \equiv-1\pmod{5}$. We have $c_4=-9$. Equation \eqref{holt} reads, for $m=1$, \begin{eqnarray*} 720\chi(X, L)&=&720+ (-c_1^4 + 4c_1^2 c_2+ 3 c_2^2 + c_1 c_3- c_4) + 15c_1c_2+ 10(c_1^2+c_2 )+ 15 c_1+ 6\\ &\equiv& 9 +3\cdot (-1)+ 6\pmod5, \end{eqnarray*} which is absurd. \medskip \noindent{\bf Case $c_1=\pm 2$.} By \eqref{cong}, we can write $c_2=12d_2-1$. Assume $c_1=2$. Substituting $c_5=6$, $c_4=45$, and $c_3=(1530-6c_2^2+8c_2)/4$ (obtained from \eqref{quad}) in \eqref{tholt} with $m=-1$, we compute, following \cite{fuj1}, $$ 24\chi(X,T_X\otimes L^{-1})= -3 c_2^2 + 2c_2 + 731=-3 (12 d_2-1)^2 + 2(12 d_2-1) + 731 \equiv 6\pmod{24}, $$ which is absurd. When $c_1=-2$, we obtain similarly the contradiction $$ 24 \chi(X,T_X\otimes L)=3 c_2^2 - 2c_2 + 727\equiv 12\pmod{24}. $$ \medskip \noindent{\bf Case $c_1=-6$.} By \eqref{cong}, we can write $c_2=12 d_2+3$. We compute, using \eqref{quad} and \eqref{tholt} with $m=-1$ again, \begin{eqnarray*} 24 \chi(X,T_X\otimes L )&=& 45c_2^2 - \tfrac{520}{3}c_2 - 171= 45(12 d_2+3)^2 - \tfrac{520}{3}(12 d_2+3) - 171\\ &\equiv& 45\cdot 9-520\cdot 4 d_2-520- 171 \pmod{24}, \end{eqnarray*} but this is absurd since this last number is $\equiv 2 \pmod8$. \end{proof} \section{Sixfolds} \begin{theo}[\cite{lw}] Any compact K\"ahler manifold with the same integral cohomology ring as ${\mathbf P}^6$ is \begin{itemize} \item either isomorphic to ${\mathbf P}^6$; \item or a quotient of the unit ball ${\bf B}^6$. \end{itemize}\end{theo} \begin{proof} If $\ell$ is a positive generator of $H^2(X,{\mathbf Z})$, we write as before $c_i(X)=c_i\ell^i$. Equations \eqref{cn} gives $c_6=7$ and $c_1 c_5 =3\cdot 7^2=147$. From the fact that the polynomial $t_6(y;c_1,\dots,c_6)$ is the same for $X$ and ${\mathbf P}^6$, we obtain \begin{eqnarray*} 720\tbinom{7}{5}\hskip-2mm&=&\hskip-2mm - c_1^3 c_3 + 3 c_1 c_2 c_3 + c_1^2 c_4 - 3 c_3^2 + 3 c_2 c_4 + 69 c_1 c_5 + 186 c_6, \\ 60480\tbinom{7}{7}\hskip-2mm&=&\hskip-2mm 2 c_1^6 - 12 c_1^4 c_2 + 11 c_1^2 c_2^2 + 5 c_1^3 c_3 + 10 c_2^3 + 11 c_1 c_2 c_3 - 5 c_1^2 c_4 - c_3^2 - 9c_2 c_4 - 2 c_1 c_5 + 2c_6. \end{eqnarray*} Plugging in the values $c_6=7$ and $c_1 c_5 = 147$, we obtain (\cite[p.\ 150]{lw}) \begin{eqnarray} - c_1^3 c_3 + 3 c_1 c_2 c_3 + c_1^2 c_4 - 3 c_3^2 + 3 c_2 c_4 &=& 3675, \label{0}\\ 2 c_1^6 - 12 c_1^4 c_2 + 11 c_1^2 c_2^2 + 5 c_1^3 c_3 + 10 c_2^3 + 11 c_1 c_2 c_3 - 5 c_1^2 c_4 - c_3^2 - 9c_2 c_4 &=& 60760.\label{1} \end{eqnarray} Eliminating $c_4$ between these two equations, we see that $c_3$ is a solution of the quadratic equation \begin{equation}\label{quadeq} a_2c_3^2+a_1c_3+a_0=0, \end{equation} where \begin{eqnarray*} a_2&=&2(15+8c_1^2),\\ a_1&=& -4c_1c_2(15c_2+8c_1^2),\\ a_0&=&-30c_2^2-43c_1^2c_2^3+25c_1^4c_2^2+6c_1^6c_2+215355c_2-2c_1^8+79135c_1^2. \end{eqnarray*} In particular, $15c_2+8c_1^2$ divides $a_0$, hence also the remainder of the division of $1125a_0$ by $15c_2+8c_1^2$, which is $$R(c_1):=c_1^2(6758c_1^6-40186125). $$ Moreover, Equation \eqref{holt} gives that the polynomial \begin{eqnarray*} P(m)&=&m\frac{1}{1440}\bigl( -c_1c_4+c_1^2c_3+3c_1c_2^2 -c_1^3c_2 \bigr) + \frac{m^2}{2}\frac{1}{720}\Bigl(-c_4+c_1c_3+3c_2^2+4c_1^2c_2-c_1^4\Bigr)\\ && \hskip 1cm {}+ \frac{m^3}{6}\cdot\frac{1}{24}c_1c_2+ \frac{t^4}{24}\cdot\frac{1}{12}\Bigl(c_1^2+c_2\Bigr)+ \frac{m^5}{120}\cdot\frac{1}{2}c_1+ \frac{m^6}{720} \end{eqnarray*} takes integral values for all integers $m$. In particular, \begin{equation}\label{+} 720(P(1)+P(-1))=-c_4+c_1c_3+3c_2^2+4c_1^2c_2-c_1^4 + 5 (c_1^2+c_2 )+2 \equiv 0\pmod{720}. \end{equation} Finally, by \eqref{c1c2}, we have the congruence \begin{equation}\label{cog} c_1^2+c_2+6c_1 \equiv 4\pmod{12}. \end{equation} \medskip \noindent{\bf Case $ c_1\equiv 0\pmod3$. } We get $c_2 \equiv 1\pmod3$ (from \eqref{cog}), $ c_4 \equiv 1\pmod{3}$ (from \eqref{+}), and $ c_3 \equiv 0\pmod{3}$ (from \eqref{0}). We write $c_1=3d_1$, $c_2=3d_2+1$, and $c_4=3d_4+1$. We have $ -c_4 + 3 +5c_2 +2\equiv0\pmod{9}$ (from \eqref{+} again), \hbox{i.e.}, $ d_2+ d_4 \equiv 0 \pmod{3}$. Moreover, using \eqref{0} modulo 27, we obtain $9d_1^2(3d_4+1) +3(3d_2+1)(3d_4+1) \equiv 3\pmod{27} $ (from \eqref{0} again), \hbox{i.e.}, $d_4+d_2 + d_1^2 \equiv 0\pmod{3} $. This gives $d_1 \equiv 0\pmod{3}$, which is impossible, since $d_1$ is a power of $7$. \medskip The case $c_1=49$ being excluded by \eqref{koo}, there remains to consider the cases $c_1=-49$, $c_1=\pm1$, and $c_1=\pm7$. In all these cases, we have $c_2 \equiv -3\pmod{12}$ by \eqref{cog} and we write $c_2=12e_2-3$. We saw above that $R(c_1) $ must be divisible by $15c_2+8c_1^2=180e_2+8c_1^2-45$. \medskip \noindent{\bf Case $c_1=-49$.} The integer $d=15c_2+8c_1^2=180e_2-45+8\cdot 49^2$ is positive by \eqref{yauu} and divides $R(c_1)= 7^6\cdot 37\cdot 251\cdot 1559\cdot 131849$. A computer check gives us all the positive divisors $d$ of $R(c_1)$ for which $e_2$ is an integer; we then compute the discriminant $D$ of the quadratic equation \eqref{quadeq}. We find \begin{itemize} \item $d=37\cdot 1559$, for which $e_2=214$ and $D=2^7 \cdot 37 \cdot 1559 \cdot 7087681 \cdot 21780337$; \item $d=7\cdot 131849$, for which $e_2=5021$ and $D= 2^8 \cdot 7^3 \cdot 13 \cdot 193 \cdot 131849 \cdot 9694436995073$; \item $d=7^3\cdot 251\cdot 1559\cdot 131849 $, for which $e_2=98314662210$ and $D= 2^7 \cdot 7^8 \cdot 17 \cdot 251 \cdot 317 \cdot 1559 \cdot 131849 \cdot 165057229 \cdot 1203263426047496730660859$. \end{itemize} In each case, $D$ is not a perfect square hence the system of equations \eqref{0} and \eqref{1} has no integral solutions. \medskip \noindent{\bf Case $c_1=\pm1$.} We have $R(c_1)=-23 \cdot 1746929$ and there are no divisors of $R(c_1)$ for which $e_2$ is an integer. \medskip \noindent{\bf Case $c_1= 7$.} There is then equality in \eqref{koo}  and $X$ is isomorphic to ${\mathbf P}^6$. \medskip \noindent{\bf Case $c_1=-7$.} We have $R(c_1)=7^4 \cdot 101 \cdot 152533$. The only divisor of $R(c_1)$ for which $e_2$ is an integer is $7\cdot 101$, for which $e_2=2$ and $c_2=21$. There is then equality in \eqref{yauu}  and $X$ is a quotient of ${\bf B}^6$. \end{proof} \section{Sevenfolds} \begin{theo}\label{th7} Any compact K\"ahler manifold $X$ with the same integral cohomology ring as ${\mathbf P}^7$ is isomorphic to ${\mathbf P}^7$, unless $c_1(X)^7\in \{\pm 2^7,\pm 4^7\}$. \end{theo} \begin{proof} If $\ell$ is a positive generator of $H^2(X,{\mathbf Z})$, we write as before $c_i(X)=c_i\ell^i$. Equations \eqref{cn} give $c_7=8$ and $c_1 c_6 = 2^5\cdot 7=224$ and, by \eqref{c1}, $c_1$ is even. From the fact that the polynomial $t_7(y;c_1,\dots,c_7)$ is the same for $X$ and ${\mathbf P}^7$, we obtain, comparing the coefficients of $y^5$ and $y^7$ and plugging in the values $c_7=8$ and $c_1 c_6 = 224$, the equations \begin{eqnarray} 0 \hskip-2mm&=&\hskip-2mm c_1^3 c_4 - 3 c_1 c_2 c_4 - c_1^2 c_5 + 3 c_3 c_4 - 3 c_2 c_5 + 7728 , \label{75}\\ 0\hskip-2mm&=&\hskip-2mm -2 c_1^5 c_2 + 10 c_1^3 c_2^2 + 2 c_1^4 c_3 - 10 c_1 c_2^3 - 11 c_1^2 c_2 c_3 - 2 c_1^3 c_4 + c_1 c_3^2 + 9 c_1 c_2 c_4 \nonumber\\ &&\hskip35mm{} + 2 c_1^2 c_5 + 120512 \label{77}, \end{eqnarray} with $7728 =2^4 \cdot 3 \cdot 7 \cdot 23 $ and $ 120512=2^6 \cdot 7 \cdot 269$. By \eqref{c1c2}, we also have the congruence \begin{equation}\label{cog7} c_1^2+c_2+3c_1 \equiv 8\pmod{12}. \end{equation} We also compute, for all integers $m$, \begin{eqnarray} 60480\chi(X,L^m)&=&12 m^7 + 42 m^6 c_1 + 42 m^5 (c_1^2 + c_2) + 105 m^4 c_1 c_2\nonumber\\ &&{} + 2m^3(-7 c_1^4 + 28 c_1^2 c_2+ 21 c_2^2 + 7 c_1 c_3 - 7 c_4) \nonumber\\ &&{}+ m^2 (-21 c_1^3 c_2 + 63 c_1 c_2^2 + 21 c_1^2 c_3 - 21 c_1 c_4) \nonumber\\ &&{}+ m( 2 c_1^6 - 12 c_1^4 c_2 + 11 c_1^2 c_2^2 + 5 c_1^3 c_3 + 10 c_2^3 + 11 c_1 c_2 c_3 \label{lm} \\ &&\qquad{}- 5 c_1^2 c_4 - c_3^2 - 9 c_2 c_4 - 2 c_1 c_5 + 2 c_6) \nonumber\\ &&{}+60480,\nonumber\end{eqnarray} with $ 60480=2^6 \cdot 3^3 \cdot 5 \cdot 7$. \medskip \noindent{\bf Case $7\mid c_1 $.} We write $c_1=7d_1$. Equation \eqref{77} divided by 7 gives $$ - 10 d_1 c_2^3 + d_1 c_3^2 + 9 d_1 c_2 c_4 + 2^6\cdot 269 \equiv 0\pmod7, \text{ hence } d_1(3 c_2^3-c_3^2-2 c_2c_4 )\equiv 3 \pmod7,$$ and, from \eqref{lm} with $m=1$, we get $$24+2(10 c_2^3 - c_3^2 - 9 c_2 c_4 +2 c_6)\equiv 0\pmod7, \text{ hence }5+3 c_2^3 -c_3^2 - 2 c_2 c_4 +2 c_6 \equiv 0\pmod7. $$ Together, these two congruences imply $ 5 d_1+3 +2 d_1c_6 \equiv 0\pmod7$. Since $ d_1 c_6=224/7=32$, we finally get $ d_1\equiv 2\pmod7$. But $d_1\le 1$ by \eqref{koo} and $ d_1\mid 2^5$, and all these conditions are incompatible. \medskip Since $c_1$ must be even, we now would like to exclude the cases $c_1\in \{\pm 2,\pm 4,-8,-16,-32\}$. Unfortunately, playing around with congruences is not enough when $c_1\in \{\pm 2,\pm 4\}$ (even with \eqref{tholt}) and we were unable to exclude these cases. We therefore assume $c_1\in \{-8,-16,-32\}$ and write $c_1=4 d_1 $. \medskip Congruence \eqref{cog7} implies $ c_2 \equiv 0\pmod{4}$ and we write $c_2=4 d_2$. Equation \eqref{lm} with $m=1$, taken modulo 32 and divided by 2, gives \begin{eqnarray} 0&\equiv& - 2 d_1 c_3^2 - 8 d_1 d_2 c_4 + 8 d_1 c_3 - c_3^2 - 20 d_1 c_4 - 4 d_2 c_4 - 8 d_1 c_5 + 8 d_1 + 8 d_2 \nonumber \\ &&{}-14 c_4 + 2 c_6 + 12 \pmod{16}, \label{eq9}\end{eqnarray} hence $c_3$ is even; we write $c_3=2 d_3$. When $c_1=\pm 32$, we already get the contradiction $0\equiv 2^6\pmod{2^7}$ by reducing equation \eqref{77} modulo $2^7$. So we assume $(d_1,c_6)\in \{(-2,-28),(-4,-14)\}$ and we write $c_6=2 d_6$. We obtain from \eqref{eq9} that $c_4$ is even, we write $c_4=2 d_4$, and, after dividing by 4, we get \begin{eqnarray*} 0&\equiv& - 2 d_1 d_3^2 - d_3^2 - 2 d_1 d_4 - 2 d_2 d_4 - 2 d_1 c_5 + 2 d_1 + 2 d_2 - 7d_4 + d_6 + 3 \pmod{4} \end{eqnarray*} hence $ d_3\equiv d_4+ d_6+1\pmod2$. \medskip \noindent{\bf Case $ c_1=-16$. } We have $ d_3\equiv d_4 \pmod2$. Substituting the integers $d_2$, $d_3$, and $d_4$ into \eqref{77} and dividing by 64, we find $$160 d_2^3 - 10240 d_2^2 - 352 d_2 d_3 - d_3^2 - 18 d_2 d_4 + 131072 d_2 + 4096 d_3 + 256 d_4 + 8 c_5 + 1883=0,$$ hence $ d_3\equiv 1\pmod2$. Doing the same with \eqref{75}, we obtain, after dividing by 4, $$96 d_2 d_4 + 3 d_3 d_4 - 3 d_2 c_5 - 2048 d_4 - 64 c_5 + 1932=0,$$ hence $ d_2\equiv c_5\equiv 1\pmod2$. So we write $ d_2=2e_2+1$, $ d_3=2e_3+1$, $ d_4=2e_4+1$, $ c_5=2d_5+1$, and, substituting these variables into \eqref{lm} with $m=-2$, we obtain the contradiction $$756 \chi(X,L^{-2}) = -640 e_2^3 + 7424 e_2^2 + 704 e_2 e_3 + 2 e_3^2 + 36 e_2 e_4 + 80818 e_2 + 290 e_3 + 14 e_4 - 8 d_5 - 713529 $$ (the right side is odd, but the left side is even). \medskip \noindent{\bf Case $ c_1=-8$. } We have $ d_3\equiv d_4 +1\pmod2$. Substituting the integers $d_2$, $d_3$, and $d_4$ into \eqref{77} and dividing by 32, we find $$160 d_2^3 - 2560 d_2^2 - 176 d_2 d_3 - d_3^2 - 18 d_2 d_4 + 8192 d_2 + 512 d_3 + 64 d_4 + 4 c_5 + 3766=0,$$ hence $ d_3\equiv 0\pmod2$ and $ d_4\equiv 1\pmod2$. We write $ d_3=2e_3$ and $ d_4=2e_4+1$. Doing the same with \eqref{lm} with $m=-1$, we get \begin{eqnarray*} 60480\chi(X,L^{-1})&=& -50 d_2^3 + 1310 d_2^2 + 880 d_2 e_3 + 20 e_3^2 + 90 d_2 e_4+ 45 d_2 + 22670 d_2 \\ &&\hskip55mm {} - 160 e_3 - 20 e_4 - 80 c_5 - 469710, \end{eqnarray*} hence $d_2\equiv 0\pmod 2$. We write $ d_2=2e_2$ and, substituting into \eqref{77} and dividing by $64$, we obtain the contradiction $$ 640 e_2^3 - 5120 e_2^2 - 352 e_2 e_3 - 2 e_3^2 - 36 e_2 e_4 + 8174 e_2 + 512 e_3 + 64 e_4 + 2 c_5 + 1915 =0 .$$ This finishes the proof of the theorem. \end{proof} \section{Computations}\label{seco} \allowdisplaybreaks We list here the polynomials $t_n(z;c_1,\dots,c_n)$ (defined in \eqref{tn}) for $n\le 9$ (they were given for $n\le 6$ in \cite[p.\ 145]{lw}) \begin{eqnarray*} t_1\hskip-3mm&=&\hskip-3mm c_1-\tfrac{1}{2} c_1z\\ t_2\hskip-3mm&=&\hskip-3mm c_2 -c_2z+\tfrac{1}{12}( c_1^2 + c_2)z^2\\ t_3\hskip-3mm&=&\hskip-3mm c_3 -\tfrac{3}{2} c_3 z+\tfrac{1}{12}(c_1 c_2 + 6 c_3)z^2- \tfrac{1}{2 4} c_1 c_2 z^3\\ t_4\hskip-3mm&=&\hskip-3mm c_4 -2 c_4z+\tfrac{1}{12}(c_1 c_3 + 14 c_4)z^2+\tfrac{1}{12}(c_1 c_3 + 2 c_4)z^3+\tfrac{1}{720} ( -c_1^4 + 4c_1^2 c_2+ 3 c_2^2 + c_1 c_3- c_4)z^4\\ t_5\hskip-3mm&=&\hskip-3mm c_5 -\tfrac{5}{2} c_5z +\tfrac{1}{12} ( c_1 c_4 + 25 c_5)z^2 -\tfrac{1}{8}( c_1 c_4 +5 c_5)z^3 \\ && {}+\tfrac{1}{720} ( -c_1^3 c_2 +3 c_1 c_2^2 +c_1^2 c_3 + 29 c_1 c_4 +30 c_5)z^4+\tfrac{1}{1440} ( c_1^3 c_2 - 3 c_1 c_2^2 - c_1^2 c_3 + c_1 c_4)z^5\\ t_6\hskip-3mm&=&\hskip-3mm c_6 -3 c_6z +\tfrac{1}{12} ( c_1 c_5 + 39 c_6)z^2 -\tfrac{1}{6} ( c_1 c_5 +9 c_6)z^3\\ &&{} +\tfrac{1}{720} ( -c_1^3 c_3 +3 c_1 c_2 c_3 + c_1^2 c_4 - 3 c_3^2 +3 c_2 c_4 + 69 c_1 c_5 +186 c_6)z^4\\ &&{} +\tfrac{1}{720} (c_1^3 c_3 - 3c_1 c_2 c_3 - c_1^2 c_4 + 3 c_3^2 - 3 c_2 c_4 - 9 c_1 c_5 - 6 c_6)z^5\\ &&{} +\tfrac{1}{60480} ( 2c_1^6 - 12 c_1^4 c_2 + 11 c_1^2 c_2^2 +5 c_1^3 c_3 + c_2^3 + 11 c_1 c_2 c_3 - 5 c_1^2 c_4\\ &&\hskip15mm {}- c_3^2 - 9 c_2 c_4 - 2 c_1 c_5 + 2 c_6)z^6\\ t_7\hskip-3mm&=&\hskip-3mm c_7 -\tfrac{7}{2} c_7z +\tfrac{1}{12} (c_1 c_6 + 56 c_7)z^2 -\tfrac{5}{24}( c_1 c_6 + 14 c_7)z^3\\ &&{} +\tfrac{1}{720} (-c_1^3 c_4 + 3 c_1 c_2 c_4 + c_1^2 c_5- 3 c_3 c_4+3 c_2 c_5 +124 c_1 c_6+602 c_7)z^4\\ &&{} +\tfrac{1}{480} ( c_1^3 c_4 -3 c_1 c_2 c_4 - c_1^2 c_5 +3 c_3 c_4 - 3c_2 c_5 - 24 c_1 c_6 - 42 c_7)z^5\\ &&{} +\tfrac{1}{60480}(2 c_1^5 c_2 - 10 c_1^3 c_2^2 - 2c_1^4 c_3 + 10 c_1 c_2^3 + 11 c_1^2 c_2 c_3 - 40 c_1^3 c_4 - c_1 c_3^2 \\ &&\hskip15mm {}+ 117 c_1 c_2 c_4 + 40 c_1^2 c_5 - 126 c_3 c_4 + 126c_2 c_5 + 170 c_1 c_6 + 84 c_7)z^6\\ &&{} +\tfrac{1}{120960} ( - 2 c_1^5 c_2 + 10 c_1^3 c_2^2 + 2 c_1^4 c_3 -10 c_1 c_2^3 - 11 c_1^2 c_2 c_3- 2 c_1^3 c_4 \\ &&\hskip15mm {}+c_1 c_3^2+9c_1 c_2 c_4 +2c_1^2 c_5 - 2c_1 c_6)z^7\\ t_8\hskip-3mm&=&\hskip-3mm c_8 -4 c_8z +\tfrac{1}{12} (c_1 c_7 + 76 c_8)z^2 -\tfrac{1}{4} ( c_1 c_7 + 20 c_8)z^3\\ &&{} +\tfrac{1}{720} (-c_1^3 c_5 + 3 c_1 c_2 c_5+ c_1^2 c_6 - 3 c_3 c_5+3 c_2 c_6+194 c_1 c_7+ 1458 c_8)z^4\\ &&{} +\tfrac{1}{360}( c_1^3 c_5 - 3 c_1 c_2 c_5 - c_1^2 c_6 + 3 c_3 c_5 - 3c_2 c_6 - 44 c_1 c_7 - 138 c_8)z^5\\ &&{} +\tfrac{1}{6048 0}(2 c_1^5 c_3 -10 c_1^3 c_2 c_3 -2 c_1^4 c_4 + 10 c_1 c_2^2 c_3 +10 c_1^2 c_3^2 + c_1^2 c_2 c_4 - 96 c_1^3 c_5 - 10 c_2 c_3^2 \\ && {} + 10 c_2^2 c_4 - 11 c_1 c_3 c_4 + 295 c_1 c_2 c_5 + 96 c_1^2 c_6 - 20 c_4^2 - 264c_3 c_5 + 284 c_2 c_6 \\ &&\hskip10mm {}+1206 c_1 c_7 + 1524 c_8)z^6\\ &&{} +\tfrac{1}{60480} (-2 c_1^5 c_3 + 10 c_1^3 c_2 c_3 + 2 c_1^4 c_4 - 10 c_1 c_2^2 c_3 - 10 c_1^2 c_3^2 - c_1^2 c_2 c_4 + 12 c_1^3 c_5 + 10 c_2 c_3^2 \\ && {}- 10 c_2^2 c_4+ 11 c_1 c_3 c_4 - 43 c_1 c_2 c_5 - 12c_1^2 c_6 + 20 c_4^2 + 12 c_3 c_5 - 32 c_2 c_6\\ &&\hskip10mm {} -30 c_1 c_7 - 12 c_8)z^7\\ &&{} +\tfrac{1}{3628800}(-3 c_1^8 + 24 c_1^6 c_2 - 50 c_1^4 c_2^2 - 14 c_1^5 c_3 + 8 c_1^2 c_2^3 + 26 c_1^3 c_2 c_3 + 14 c_1^4 c_4 +21 c_2^4 \\ &&\hskip10mm {}+ 50 c_1 c_2^2 c_3 + 3 c_1^2 c_3^2 - 19 c_1^2 c_2 c_4 - 7 c_1^3 c_5 - 8 c_2 c_3^2 -34 c_2^2 c_4 - 13 c_1 c_3 c_4 \\ &&\hskip10mm {}- 16 c_1 c_2 c_5 + 7 c_1^2 c_6 +5 c_4^2 + 3 c_3 c_5 + 13 c_2 c_6 + 3 c_1 c_7 -3 c_8)z^8\\ t_9\hskip-3mm&=&\hskip-3mm c_9 -\tfrac{9}{2} c_9z +\tfrac{1}{12} ( c_1 c_8 + 99 c_9)z^2 +\tfrac{1}{24} (-7 c_1 c_8 -189 c_9)z^3\\ &&\hskip-3mm{}+\tfrac{1}{720} (- c_1^3 c_6 + 3 c_1 c_2 c_6 + c_1^2 c_7 - 3 c_3 c_6 + 3 c_2 c_7 + 279 c_1 c_8 + 2979c_9)z^4\\ &&\hskip-3mm{} +\tfrac{1}{2 88} ( c_1^3 c_6 - 3 c_1 c_2 c_6 - c_1^2 c_7 +3 c_3 c_6 - 3 c_2 c_7 -69 c_1 c_8 - 333 c_9)z^5\\ &&\hskip-3mm{} +\tfrac{1}{60480} (2 c_1^5 c_4 - 10 c_1^3 c_2 c_4 -2 c_1^4 c_5 + 10 c_1 c_2^2 c_4 + 10c_1^2 c_3 c_4 + c_1^2 c_2 c_5 - 173 c_1^3 c_6 - 10c_2 c_3 c_4 \\ &&\hskip1cm {} - 10c_1 c_4^2 + 10c_2^2 c_5 - c_1 c_3 c_5 + 526 c_1 c_2 c_6 \\ &&\hskip1cm {}+ 173 c_1^2 c_7 - 10 c_4 c_5 - 505 c_3 c_6 + 515 c_2 c_7 + 4041 c_1 c_8 + 9075c_9)z^6 \\ &&\hskip-3mm{}+\tfrac{1}{40320} (-2 c_1^5 c_4 + 10 c_1^3 c_2 c_4 + 2 c_1^4 c_5 - 10 c_1 c_2^2 c_4 - 10 c_1^2 c_3 c_4 - c_1^2 c_2 c_5 + 33 c_1^3 c_6 +10c_2 c_3 c_4 \\ &&\hskip1cm {} +10 c_1 c_4^2 -10 c_2^2 c_5 + c_1 c_3 c_5 - 106 c_1 c_2 c_6 - 33c_1^2 c_7 + 10 c_4 c_5 + 85 c_3 c_6 - 95 c_2 c_7 \\ &&\hskip1cm {}- 261 c_1 c_8 - 255 c_9)z^7\\ &&\hskip-3mm{} + \tfrac{1}{3628800}(-3 c_1^7 c_2 +21 c_1^5 c_2^2 +3c_1^6 c_3 - 42 c_1^3 c_2^3 - 29 c_1^4 c_2 c_3 \\ &&\hskip1cm {} + 57c_1^5 c_4 + 21c_1 c_2^4 +50 c_1^2 c_2^2 c_3 + 8 c_1^3 c_3^2 - 274c_1^3 c_2 c_4 - 57 c_1^4 c_5 - 8 c_1 c_2 c_3^2 \\ &&\hskip1cm {}+ 266 c_1 c_2^2 c_4 +287 c_1^2 c_3 c_4 + 14 c_1^2 c_2 c_5 - 153 c_1^3 c_6 - 300 c_2 c_3 c_4 - 295 c_1 c_4^2 \\ &&\hskip1cm {}+ 300 c_2^2 c_5- 27 c_1 c_3 c_5 + 673 c_1 c_2 c_6 + 153 c_1^2 c_7 - 300 c_4 c_5 - 30c_3 c_6 + 330 c_2 c_7 \\ &&\hskip1cm {}+ 267 c_1 c_8+90 c_9)z^8 \\ &&\hskip-3mm{}+\tfrac{1 }{7257600}(3c_1^7 c_2 - 21 c_1^5 c_2^2 - 3 c_1^6 c_3 + 42 c_1^3 c_2^3 + 29 c_1^4 c_2 c_3 +3 c_1^5 c_4 - 21c_1 c_2^4 \\ &&\hskip1cm {}- 50 c_1^2 c_2^2 c_3 - 8 c_1^3 c_3^2 - 26 c_1^3 c_2 c_4 - 3 c_1^4 c_5 +8c_1 c_2 c_3^2 + 34 c_1 c_2^2 c_4 \\ &&\hskip1cm {}+ 13c_1^2 c_3 c_4 +16 c_1^2 c_2 c_5 + 3 c_1^3 c_6 -5 c_1 c_4^2 - 3 c_1 c_3 c_5 -13 c_1 c_2 c_6 \\ &&\hskip1cm {}-3 c_1^2 c_7 + 3 c_1 c_8)z^9 \end{eqnarray*}

https://arxiv.org/abs/2201.01286

Catching Polygons

Consider an arrangement of $k$ lines intersecting the unit square. There is some minimum scaling factor so that any placement of a rectangle with aspect ratio $1 \times p$ with $p\geq 1$ must non-transversely intersect some portion of the arrangement or unit square. Assuming the lines of the arrangement are axis-aligned, we show the optimal arrangement depends on the aspect ratio of the rectangle. In particular, the optimal arrangement is either evenly spaced parallel lines or an evenly spaced grid of lines. We present the precise aspect ratios of rectangles for which each of the two nets is optimal.

\section{Discussion} \label{sec:discuss} In this work, we showed that for axis-aligned nets with $k$ lines and rectangular intruder with aspect ratio $1\times p$, the optimal net is evenly spaced parallel lines for $p\leq \frac{k+1}{\frac{k}{2}+1}$ and a grid of $\frac{k}{2}\times \frac{k}{2}$ evenly spaced lines for $p\geq \frac{k+1}{\frac{k}{2}+1}.$ As far as we know, the problem is still open for non axis-aligned nets and rectangular intruders. Our hope is that someone can show, that for any net, one can construct an axis-aligned net that does change the optimal scale factor. Non-rectangular intruders would also be interesting to explore. \section{Towards Axis-Aligned Net Optimality} \label{append:append-axis} In this appendix, we show that, for square intruder, evenly spaced axis-aligned vertical lines are generally a local optimum. This is a first step toward our conjecture that axis-aligned lines are a global optimum for rectangular intruder. \begin{lemma}\label{lem:local} Let $P$ be a intruder with aspect ration $p=1$, i.e., a square. Then evenly spaced axis-aligned vertical lines are a local optimum when the number of lines is $k >2$. \end{lemma} \begin{proof} % Denote the $k$ lines from left to right by $\ell_1, \ell_2, \ldots, \ell_k$. Assume, towards a contradiction, that, given some small $\epsilon > 0$, there are small shift and pivot values for each line that describe the translation of lines to an arrangement that results in a lower overall scaling factor $c$. Specifically, let these values be denoted $s_1, s_2, \ldots, s_k$ and $p_1, p_2, \ldots , p_k$, respectively, where $s_i, p_i \in [0, \epsilon]$, and at least one of these values $s_i$ or $p_i$ are nonzero. Notice that, regardless of the pivot value, shifting any line decreases the maximum intruder size for one neighboring face, but increases it for the other neighboring face, leading to a higher $c$-value overall. Thus, we must have $s_1 = s_2 = \ldots = s_k$ = 0. This means we must consider the effect of pivot values on unshifted lines. Notice that pivoting a single line that is adjacent to a vertical (un-pivoted) line will increase the maximum intruder size for the face between them. Since the lines $p_1$ and $p_k$ are always adjacent to the vertical edges of the bounding square, we must have $p_1 = p_k = 0$, otherwise the leftmost and rightmost faces would cause the arrangement to have a higher $c$-value. But then we must also have $p_2 = p_{k-1} = 0$, or else the second to leftmost and second to rightmost faces would cause the arrangement to have a higher $c$-value. Continuing this line of argument, we eventually see that $p_1 =p_2 = \ldots = p_k = 0$, contradicting our assumption that some $s_i$ or $p_i$ be nonzero. % \end{proof} \section{An odd number of lines} \label{append:odd} In this section, we prove that, when $k$ is odd, the net $\mathcal{N}(k,0)$ is optimal for $1\leq p\leq \frac{(k+1)\lfloor\frac{k}{2}\rfloor}{\lceil\frac{k}{2} \rceil^2}$ and the net $\mathcal{N}(\lceil\frac{k}{2}\rceil,\lfloor\frac{k}{2}\rfloor)$ is optimal for $p\geq \frac{(k+1)\lfloor\frac{k}{2}\rfloor}{\lceil\frac{k}{2} \rceil^2}.$ The net $\mathcal{N}(\lceil\frac{k}{2}\rceil,\lfloor\frac{k}{2}\rfloor)$ has $\frac{1}{\lceil\frac{k}{2}\rceil }\times \frac{1}{\lfloor\frac{k}{2}\rfloor }$ holes and the net $\mathcal{N}(k,0)$ has $1\times \frac{1}{k+1}$ holes. We consider the curves $\frac{1}{\lceil\frac{k}{2}\rceil} \mathcal{C}_{\frac{\lceil\frac{k}{2}\rceil}{\lfloor\frac{k}{2}\rfloor}}(p)$ and $\frac{1}{k+1}\mathcal{C}_{k+1}(p)$. See \figref{odd}. These curves intersect at $p=\frac{(k+1)\lfloor\frac{k}{2}\rfloor}{\lceil\frac{k}{2} \rceil^2}.$ We define the \emph{base curve} for $k$ odd to be \[ D_k(p) = \min\Big\{ \frac{1}{k+1}\mathcal{C}_{k+1}(p), \frac{1}{\lceil\frac{k}{2}\rceil} \mathcal{C}_{\frac{\lceil\frac{k}{2}\rceil}{\lfloor\frac{k}{2}\rfloor}}(p)\Big\}. \] \begin{figure}[htb] \centering \includegraphics[width=.4\textwidth]{odd-minEli} \caption{The minimum curve for $k$ odd, with $p_1 = \frac{(k+1)\lfloor\frac{k}{2}\rfloor}{\lceil\frac{k}{2} \rceil^2}$.} \label{fig:odd} \end{figure} \begin{theorem}[Regular is Optimal Odd]\label{thm:optimal-odd} For $k$ odd, the optimal axis aligned net for rectangular polygons is $\mathcal{N}(k,0)$ for aspect ratio $p\leq\frac{(k+1)\lfloor\frac{k}{2}\rfloor}{\lceil\frac{k}{2} \rceil^2}$ and $\mathcal{N}(\lceil\frac{k}{2}\rceil,\lfloor\frac{k}{2}\rfloor)$ for $p\geq\frac{(k+1)\lfloor\frac{k}{2}\rfloor}{\lceil\frac{k}{2} \rceil^2}$. \end{theorem} \begin{proof} Consider any axis aligned net over the square, let $v$ be the number of vertical lines and $h$ be the number of horizontal lines call the net $\mathcal{N}(v,h)$. Notice $v+h=k$. Recall evenly spaced lines give a smaller scale factor than irregularly spaced lines. If $v=\lceil\frac{k}{2}\rceil$ and $ \lfloor\frac{k}{2}\rfloor=n$ then we have a $\lceil\frac{k}{2}\rceil\times \lfloor\frac{k}{2}\rfloor$ net. If $n=0$ (or $v=0$), then we have $k$ parallel lines. The base curve is defined to be the minimum scale factor of $\mathcal{N}(k,0)$ and $\mathcal{N}(\lceil\frac{k}{2}\rceil,\lfloor\frac{k}{2}\rfloor)$ so the base curve has scale factor less than or equal to $\mathcal{N}(k,0).$ Consider any other $v$ and $h$. Without loss of generality assume $h<v$ we have $0<h< \lfloor\frac{k}{2}\rfloor<\lceil \frac{k}{2} \rceil< v<k$ There are $(v+1)(h+1)$ holes in the net and the average size of a hole is $\frac{1}{v+1}\times \frac{1}{h+1}$. There exits one hole at least as big as the average, that is, with width at least $\frac{1}{v+1}$ and height at least $\frac{1}{h+1}$. So we have a hole with aspect ratio $\frac{v+1}{h+1}$ scaled so the smaller side has length $\frac{1}{v+1}$. The maximum scale factor of a rectangle with aspect ratio $p$ that fits inside this hole is given by $\frac{1}{v+1}\mathcal{C}_{\frac{v+1}{h+1}}(p)$. We compare $\frac{1}{v+1}\mathcal{C}_{\frac{v+1}{h+1}}(p)$ to $D_k(p)$. Both $\frac{1}{v+1}\mathcal{C}_{\frac{v+1}{h+1}}(p)$ and $D_k(p)$ are piecewise functions, we directly examine all $p\geq 1$ to show $D_k(p)<\frac{1}{v+1}\mathcal{C}_{\frac{v+1}{h+1}}(p).$ For $1\leq p\leq \frac{(k+1)\lfloor\frac{k}{2}\rfloor}{\lceil\frac{k}{2} \rceil^2}$, we have $D_k(p)=\frac{1}{k+1}$ and $\frac{1}{v+1}\mathcal{C}_{\frac{v+1}{h+1}}(p)=\frac{1}{v+1}.$ since $v<k$, we have $\frac{1}{k+1}<\frac{1}{v+1}$. Then, for $\frac{(k+1)\lfloor\frac{k}{2}\rfloor}{\lceil\frac{k}{2} \rceil^2}\leq p\leq \frac{v+1}{h+1}$, $D_k(p)$ decreases and $\frac{1}{v+1}\mathcal{C}_{\frac{v+1}{h+1}}(p)=\frac{1}{k+1}$ is constant. For $\frac{v+1}{h+1}<p\leq w_{\frac{v+1}{h+1}}$, $D_k(p)=\min\Big\{ \frac{1}{k+1}\mathcal{C}_{k+1}(p), \frac{1}{\lceil\frac{k}{2}\rceil} \mathcal{C}_{\frac{\lceil\frac{k}{2}\rceil}{\lfloor\frac{k}{2}\rfloor}}(p)\Big\} \leq \frac{1}{\lceil\frac{k}{2}\rceil}\left(\frac{1}{p}\right)$ and $\frac{1}{v+1}\mathcal{C}_{\frac{v+1}{h+1}}(p)=\frac{1}{v+1}\left(\frac{v+1}{h+1}\right)\frac{1}{p}.$ Since $h\leq \lfloor\frac{k}{2}\rfloor$, we have $$\left(\frac{1}{\lceil\frac{k}{2}\rceil}\right)\left(\frac{1}{p}\right)<\frac{1}{h+1}\frac{1}{p}=\frac{1}{v+1}\left(\frac{v+1}{h+1}\right)\frac{1}{p}.$$ For $p\geq w_{\frac{v+1}{h+1}}$, we again solve the same constrained optimization problem as in the even case. The minimum occurs when $v=h$. This is a global minimum, so in the odd case the minimum occurs when me make $v$ as close to $h$ as possible. So, when $h< \lfloor\frac{k}{2}\rfloor<\lceil \frac{k}{2} \rceil< v$ we can fit the diagonal rectangle of $\mathcal{N}(\lceil\frac{k}{2}\rceil,\lfloor\frac{k}{2}\rfloor)$ inside the diagonal of the rectangle with dimensions $\frac{1}{v+1}\times\frac{1}{h+1}$ and the optimal scale factor must be larger. \end{proof} \section{Optimality Results} \label{sec:optimal} In this section, we show that for any axis-aligned net $\mathcal{N}$ with an even number of lines and any aspect ratio of the intruder $p$, the base curve has a scale factor that is less than or equal to the scale factor of $\mathcal{N}$. Let $\mathcal{N}(k,0)$ denote the net with $k$ evenly spaced parallel lines and let $\mathcal{N}(\frac{k}{2},\frac{k}{2})$ denote the net with $\frac{k}{2}$ evenly spaced vertical lines and $\frac{k}{2}$ evenly spaced horizontal lines. The rectangle aspect ratio where the base curve switches from $\mathcal{N}(k,0)$ to $\mathcal{N}(\frac{k}{2},\frac{k}{2})$ is $p=\frac{k+1}{\frac{k}{2}+1}$. We now state our main theorem: \begin{theorem}[Regular is Optimal]\label{thm:optimal} For $k$ even, the optimal axis aligned net for rectangular polygons is $\mathcal{N}(k,0)$ for aspect ratio $p\leq\frac{k+1}{\frac{k}{2}+1}$ and $\mathcal{N}(\frac{k}{2},\frac{k}{2})$ for aspect ratio $p\geq\frac{k+1}{\frac{k}{2}+1}$. \end{theorem} \begin{proof} First, notice that regularly spaced lines give a smaller scale factor than irregularly spaced lines. This is because a rectangular hole generated by regular spacing has height and width equal to the average. A rectangular hole generated by irregular spacing has a hole with height and width greater than or equal to the average. Consider any axis aligned net over the square, let $v$ be the number of vertical lines and $h$ be the number of horizontal lines call the net $\mathcal{N}(v,h)$. Notice $v+h=k$. If $v=n$ then we have a $\frac{k}{2}\times \frac{k}{2}$ net and if $n=0$ (or $v=0$), then we have $k$ parallel lines. The base curve is defined to be the minimum scale factor of $\mathcal{N}(\frac{k}{2},\frac{k}{2})$ and $\mathcal{N}(k,0)$. Thus, the scale factor of the base curve is less than or equal to either of these nets. Consider any other $v$ and $h$. Without loss of generality, assume $h<v$ we have $0<h<\frac{k}{2}<v<k.$ There are $(v+1)(h+1)$ holes in the net and the average size of a hole is $\frac{1}{v+1}\times \frac{1}{h+1}$. There exists one hole at least as big as the average, so, we have a hole at least as big as the hole with aspect ratio $\frac{v+1}{h+1}$ scaled so the smaller side has length $\frac{1}{v+1}$. The optimal scale factor of a rectangle with aspect ratio $p$ that fits inside this hole is given by $\frac{1}{v+1}\mathcal{C}_{\frac{v+1}{h+1}}(p)$. We compare $\frac{1}{v+1}\mathcal{C}_{\frac{v+1}{h+1}}(p)$ to $B_k(p)$. Both $\frac{1}{v+1}\mathcal{C}_{\frac{v+1}{h+1}}(p)$ and $B_k(p)$ are piecewise functions, we directly examine all $p\geq 1$ to show $B_k(p)\leq\frac{1}{v+1}\mathcal{C}_{\frac{v+1}{h+1}}(p)$, see \figref{optimal} for intuition. For $1\leq p\leq \frac{k+1}{\frac{k}{2}+1}$, we have $B_k(p)=\frac{1}{k+1}$ and $\frac{1}{v+1}\mathcal{C}_{\frac{v+1}{h+1}}(p)=\frac{1}{v+1}.$ since $v<k$, we have $\frac{1}{k+1}<\frac{1}{v+1}$. For $\frac{k+1}{\frac{k}{2}+1}\leq p\leq \frac{v+1}{h+1}$, $B_k(p)$ decreases and $\frac{1}{v+1}\mathcal{C}_{\frac{v+1}{h+1}}(p)=\frac{1}{k+1}$ is constant. For $\frac{v+1}{h+1}<p\leq w_{\frac{v+1}{h+1}}$, $B_k(p) =\min\bigg\{\frac{1}{k+1}\mathcal{C}_{k+1}(p), \frac{1}{\frac{k}{2}+1} \mathcal{C}_1(p) \bigg\} \leq \left(\frac{1}{\frac{k}{2}+1}\right)\left(\frac{1}{p}\right)$ and~$\frac{1}{v+1}\mathcal{C}_{\frac{v+1}{h+1}}(p)=\frac{1}{v+1}\left(\frac{v+1}{h+1}\right)\frac{1}{p}.$ Since $h\leq \frac{k}{2}$, we have $$\left(\frac{1}{\frac{k}{2}+1}\right)\left(\frac{1}{p}\right)<\frac{1}{h+1}\left(\frac{1}{p}\right)=\frac{1}{v+1}\left(\frac{v+1}{h+1}\right)\frac{1}{p}.$$ For $p\geq w_{\frac{v+1}{h+1}}$ the interior rectangle is placed diagonally. Let $c'$ be the scale factor value of the rectangle placed diagonally in $\mathcal{N}(\frac{k}{2},\frac{k}{2}).$ Consider the length of the rectangle, with shorter side equal to $c'$, placed diagonally in a rectangle with sides $\frac{1}{v+1}\times\frac{1}{h+1}.$ Let $a_1$ and $a_2$ be the legs of the small right triangle formed by $c'$, the squared length of this inscribed rectangle is $\ell^2(v,h,a_1,a_2)=(\frac{1}{v+1}-a_2)^2+(\frac{1}{h+1}-a_1)^2$. The minimum of this function along the constraints \eqnref{similar}, \eqnref{right}, and $v+h=k$ occurs when $v=h$. We omit the details, but this can be done with Lagrange multipliers. So, when $v\neq h$ we can fit the diagonal rectangle of $\mathcal{N}(\frac{k}{2},\frac{k}{2})$ inside the diagonal of the rectangle with dimensions $\frac{1}{v+1}\times\frac{1}{h+1}$ so the optimal scale factor must be~larger. \end{proof} \begin{figure}[htb] \centering \includegraphics[width=.36\textwidth]{optimalEli} \caption{The curves $B_k(p)$ and $ \frac{1}{v+1}\mathcal{C}_{\frac{v+1}{h+1}}(p)$. Here, $p_1 = \frac{k+1}{\frac{k}{2}+1}$, $p_2 = \frac{v+1}{h+1}$, $p_3 = \sqrt{2}+1$, and $p_4 = w_{\frac{v+1}{h+1}}$.} \label{fig:optimal} \end{figure} \section{Rectangular Intruders in Rectangular Nets} \label{sec:curve} When $P$ is a rectangle and the $k$ lines are axis-aligned, the problem of calculating the optimal scale factor $c$ reduces to finding the largest rectangle inscribed inside of another rectangle. Inscribing rectangles inside rectangles has been studied in \cite{Carver-1957,Dunkel-1920,wetzel-00}. In this section, we construct a curve that describes the optimal scale factor $c$ for rectangular intruders in rectangular nets for all aspect ratios of the inscribed rectangle. We fix the aspect ratio of the hole in the net to be $1\times n$ with $n\geq 1.$ Let the aspect ratio of the intruder be $1\times p$ with $p\geq 1.$ We express the optimal scale factor $c$ as a function of the aspect ratio $p$. We denote this curve by $\curve$. See \figref{curve} for an example. \begin{figure}[htb] \centering \includegraphics[width=.4\textwidth]{fig3Eli.png} \caption{The inscribing curve for some $n \in \R^{\geq 1}$. The $x$-axis is the aspect ration of the intruder and the $y$-axis is the optimal scale factor.} \label{fig:curve} \end{figure} The curve $\curve$ consists of three parts. For small $p\leq n$, placing the shorter side of the intruder parallel to the shorter side of the net is optimal and $c=1.$ For medium size $p$ values relative to $n$, the value of $c$ is limited by the height of the net and the longer side of the intruder is scaled to equal the longer side of the net. This gives $pc=n$ or $c=\frac{n}{p}$. \begin{wrapfigure}{l}{0.5\textwidth} \centering \raisebox{0pt}[\dimexpr\height-1\baselineskip\relax] {\includegraphics[width=.25\textwidth]{diagfishEli.png}} \caption{A diagonally inscribed rectangle.\label{fig:diagfish}} \vspace{-1.7cm} \end{wrapfigure} For large $p$ relative to $n$, placing the intruder diagonally is optimal. Using the variables indicated in \figref{diagfish}, we have the following three equations \begin{align} \centering &\frac{a_1}{a_2}=\frac{n-a_2}{1-a_1}\label{eqn:similar}\\ &a_1^2+a_2^2=c^2\label{eqn:right}\\ &(1-a_1)^2+(n-a_2)^2=(cp)^2\label{eqn:ceepee} \end{align} along with the natural constraints of the problem, $0<c,\ 0<a_1<1,\ 0<a_2<n$. \eqnref{similar} is due to all triangles being similar. \eqnref{right} and \eqnref{ceepee} are applications of Pythagorean's~theorem. Solving for $c$ gives $$c=\sqrt{\left(\frac{n-p}{1-p^2}\right)^2(p^2+1)-2\left(\frac{n-p}{1-p^2}\right)p+1}.$$ Therefore, we are able to give an explicit formula for $\curve$ for a given $n$, namely, \[ \curve = \begin{cases} 1 & 1\leq p\leq n \\ \frac{n}{p} & n<p\leq w_n \\ \sqrt{\left(\frac{n-p}{1-p^2}\right)^2(p^2+1)-2\left(\frac{n-p}{1-p^2}\right)p+1} & w_n<p \end{cases} \] where $w_n$ is the solution to $\sqrt{\left(\frac{n-p}{1-p^2}\right)^2(p^2+1)-2\left(\frac{n-p}{1-p^2}\right)p+1}=\frac{n}{p}$ for $1<n<p.$ The value $w_n$ represents the scale factor where the vertically inscribed rectangle has the same scale factor as the diagonally inscribed rectangle. When $k$ is even, the grid with $\frac{k}{2}$ horizontal lines and $\frac{k}{2}$ vertical lines has square holes with side length $\frac{1}{\frac{k}{2}+1}$. The arrangement with $k$ vertical and $0$ horizontal lines has rectangular holes with dimension $1\times \frac{1}{k+1}$. The curves $\frac{1}{\frac{k}{2}+1} \mathcal{C}_1(p)$ and $\frac{1}{k+1}\mathcal{C}_{k+1}(p)$ intersect at $p=\frac{k+1}{\frac{k}{2}+1}.$ For even values of $k$, we define the \emph{base curve} to be \[ B_k(p) =\min\bigg\{\frac{1}{k+1}\mathcal{C}_{k+1}(p), \frac{1}{\frac{k}{2}+1} \mathcal{C}_1(p) \bigg\} \] The case for odd values of $k$ is similar and included in \appendref{odd}. \section{Introduction} \label{sec:intro} During the open problem session of CCCG20, Joseph O'Rourke suggested the following problem. Consider an arrangement of $k$ lines, all of which intersect the unit square. For a fixed polygon $P$, there is some minimum scale factor $c > 0$ such that the scaled polygon $cP$ cannot be embedded in the unit square without intersecting any of the $k$ lines of the arrangement non-transversely (allowing translation and rotation). That is, the scaled polygon will `just touch' at least one line of the arrangement. Can we compute the minimum such $c$ over all possible arrangements of this type? Can we describe an arrangement that realizes this minimum? See \figref{intro} for an example. This problem can be described using an analogy to tripwire lasers. In this analogy, the polygon is an intruder in the unit square. The intruder can vary in size but always has the same shape. The goal is to minimize the size of the intruder that avoids the net of lasers. \begin{figure}[htb] \begin{minipage}[l]{0.67\textwidth} \includegraphics[scale=0.7]{introEli} \end{minipage}\hfill \begin{minipage}[c]{0.3\textwidth} \caption{(a) If $c$ is small, the polygon can avoid the lasers. (b) If $c$ is large, any rotation and translation of the polygon is caught in the lasers. (c) The minimum $c$ such that the polygon is caught in the lasers. } \label{fig:intro} \end{minipage} \end{figure} Here, we consider the special case where the lines are axis-aligned and $P$ is a rectangle (a lemma towards justifying the setting of axis-aligned lines is given in \appendref{append-axis}). We observe that, depending on the aspect ratio of the rectangle being considered, the optimal arrangement is either evenly spaced parallel lines or a grid of lines. See \figref{intro-2} for an example. O'Rourke asked, \emph{for what aspect ratios of rectangles is each of the two nets optimal?} In this work, we answer this question precisely. We are only aware of one work that directly considers this problem \cite{aghamolaei-capfwamn-20}, however the author has a different interpretation of the problem. Similar laser based localization problems are considered in \cite{ArkinD0GMPT20,boundary-lasers}. Using the results from \cite{ADS-95}, given any net and aspect ratio for the intruder one can compute the optimal scale factor. \begin{figure}[htb] \begin{minipage}[l]{0.55\textwidth} \includegraphics[scale=.5]{intro-2Eli} \end{minipage} \begin{minipage}{0.45\textwidth} \caption{(a) For a square the evenly spaced vertical lines are optimal but for a rectangle with a $(1\times 3)$ aspect ratio evenly spaced vertical lines is not optimal. (b) For a square the evenly spaced grid is not optimal but for a $(1\times 3)$ rectangle the evenly spaced grid is optimal.} \label{fig:intro-2} \end{minipage} \end{figure}

https://arxiv.org/abs/1506.09077

Chi-square Fitting When Overall Normalization is a Fit Parameter

The problem of fitting an event distribution when the total expected number of events is not fixed, keeps appearing in experimental studies. In a chi-square fit, if overall normalization is one of the parameters parameters to be fit, the fitted curve may be seriously low with respect to the data points, sometimes below all of them. This problem and the solution for it are well known within the statistics community, but, apparently, not well known among some of the physics community. The purpose of this note is didactic, to explain the cause of the problem and the easy and elegant solution. The solution is to use maximum likelihood instead of chi-square. The essential difference between the two approaches is that maximum likelihood uses the normalization of each term in the chi-square assuming it is a normal distribution, 1/sqrt(2 pi sigma-square). In addition, the normalization is applied to the theoretical expectation not to the data. In the present note we illustrate what goes wrong and how maximum likelihood fixes the problem in a very simple toy example which illustrates the problem clearly and is the appropriate physics model for event histograms. We then note how a simple modification to the chi-square method gives a result identical to the maximum likelihood method.

\section{Introduction} The problem of fitting an event distribution when the total expected number of events is not fixed, keeps appearing in experimental studies. Peelle's Pertinnent Puzzle (PPP) notes that in a $\chi^2$ fit, if overall normalization is one of the parameters parameters to be fit, the fitted curve may be seriously low with respect to the data points, sometimes below all of them. This puzzle was the subject of a NIM article by G. D'Agostini (NIMA 346 (1994) 306). This problem and the solution for it are well known within the statistics community, but, apparently, not well known among some of the physics community. The purpose of this note is didactic, to explain the cause of the problem and the easy and elegant solution. The solution is to use maximum likelihood (ML) instead of $\chi^2$. The essential difference between the two approaches is that ML uses the normalization of each term in the $\chi^2$ assuming it is a normal distribution, $1/\sqrt{2\pi\sigma^2}$. In addition, the normalization is applied to the theoretical expectation not to the data. In the present note we illustrate what goes wrong and how maximum likelihood fixes the problem in a very simple toy example which illustrates the problem clearly and is the appropriate physics model for event histograms. We then note how a simple modification to the $\chi^2$ method gives a result identical to the ML method. I will also discuss the models in G. d'Agostini's article (p. 309) and add one more. \section{Toy Model--$\chi^2$} Consider a simple data set with only two bins. Theory predicts that the expected value of $N$, the number of events in the bin should be the same for each bin, and that the bins are uncorrelated. Let $x_1$ and $x_2$ be the number of events experimentally found in the two bins. The variance ($\sigma^2$) is $N$ for each bin, ($\sigma=\sqrt{N}$). \begin{equation} \chi^2 = \frac{(N-x_1)^2}{\sigma^2} + \frac{(N-x_2)^2}{\sigma^2}. \end{equation} We want to find the minimum, $\frac{\partial\chi^2}{\partial N}= 0$. Call term 1, the derivative with respect to the numerators of the $\chi^2$. \begin{equation} \rm{Term\ 1} = 2\frac{(N-x_1 +N-x_2)}{N} = 2\big(1-\frac{x_1}{N}\big) +2\big(1-\frac{x_2}{N}\big). \end{equation} If we ignore the derivative of the denominator, then Term 1 = 0, is solved by $N = \frac{x_1+x_2}{2}$. Call this the naive solution. Call Term 2 the derivative with respect to the denominator of the $\chi^2$ \begin{equation} \rm {Term\ 2} = -\frac{(N-x_1)^2 +(N-x_2)^2}{N^2}. \end{equation} Term 2 is negative and O(1/N). The only way that Term 1 + Term 2 = 0 is for Term 1 to be positive. This means that the $\chi^2$ solution must have $N$ greater than the naive value. Although Term 1 is O(1), $x_1/N$ and $x_2/N$ are O(1/N). N is pulled up as the fit wants to make the fractional errors larger. (Had the normalization been put into the data not the theoretical value, the fitted curve would have been low.) \section{Toy Model--Maximum Likelihood} The likelihood ($\mathcal{L}$) is the probability density function for the two bins assuming each bin has a normal distribution. (This requires $N$ is not too small). \begin{equation} \mathcal{L} = \frac{1}{\sqrt{2\pi\sigma^2}} \frac{1}{\sqrt{2\pi\sigma^2}} e^{-(N-x_1)^2/(2\sigma^2)}e^{-(N-x_2)^2/(2\sigma^2)}. \end{equation} For $\sigma^2 = N$, the log of the likelihood is: \begin{equation} \ln\mathcal{L} = -\ln(2\pi)-\ln N -\chi^2/2. \end{equation} Let Term 3 be the derivative of the normalization. \begin{equation} \rm {Term\ 3 }= -\frac{1}{N}. \end{equation} The derivative of the $\ln\mathcal{L}$ is Term 3 $-$ (Term 1)/2 $-$ (Term 2)/2. \begin{equation} {\rm Term\ 3 - (Term\ 2)/2} = -\frac{1}{N} + \frac{(N-x_1)^2 +(N-x_2)^2}{2N^2} =\frac{ -2N +(N-x_1)^2 +(N-x_2)^2}{2N^2}. \end{equation} Since the expectation value $E(N-x_1)^2 = E(N-x_2)^2 = N$ , the expectation value of Term 3 - (Term 2)/2 =0. For fitted values a modification is needed. Assume that there is only one overall normalization factor and assume now that there are $n_b$ bins. The expectation value for a $\chi^2$ with $n_b$ bins and $n_f$ fitted parameters is $n_b -n_f$. This occurs because, after fitting, the multidimensional normal distribution loses $n_f$ variables. This means, for $n_b=2,\ n_f=1$, the value of Term 2 is $2\times 1/2 = 1$. The same loss in dimensions requires term 3, the normalization term of the multidimensional distribution to be multiplied by $(n_b - n_f)/n_b$ to match the change in $\chi^2$ since the fit has integrated over those variables. The change in expectation value occurs automatically in the fit, but the modification to Term 3 must be put in by hand. There is an easy general way to handle this problem. The problem arises because the error matrix is a function of normalization. When the simple $\chi^2$ method is applied, the derivative of the $\chi^2$ is in error because the change in the normalization of the particle density function is not taken into account. Including this term in the ML approach eliminates the problem. This leads to a simple approach using a modified $\chi^2$ analysis. Consider $n_b$ bins and $g$ fitting parameters $p_j$. Let $n_i(p_1,p_2,\cdots ,p_g)$ be the expected number of events in bin $i$. The distribution of experimental events in each bin is taken as approximately normal. The total number of events in the histogram is not fixed. Choose the set $n_i$ as the basis. The error matrix is diagonal in this basis. Ignoring the $2\pi$ constants: \begin{equation} \ln\mathcal{L} = \sum_{i=1}^{n_b} -\frac{\ln n_i}{2} -\frac{(x_i-n_i)^2}{2n_i}. \end{equation} \begin{equation} \frac{d\ln\mathcal{L}}{d n_i} = \frac{x_i-n_i}{n_i} +\frac{1}{2n_i}\big[\big( \frac{(x_i-n_i)^2}{n_i}\big)-1\big]. \end{equation} The expectation value for the term in square brackets is zero. Recall that the expectation refers to the average value over a number of repetitions of the experiment. It is $x_i$ that changes with each experiment not the theoretical expectation, $n_i$. The expectation value of the term in square brackets will remain zero even if it is multiplied by a complicated function of the $p_j$ fitting parameters. Ignoring this term leads to: \begin{equation} \frac{\partial \ln\mathcal{L}}{\partial p_j} = \sum_{i=1}^{n_b} \big(\frac{x_i-n_i}{n_i}\big)\frac{\partial n_i}{\partial p_j}. \end{equation} By expressing the $n_i$ as the appropriate functions of the $p_j$, the error matrix can be written in terms of the $p_j$. However, the derivative of the inverse error matrix does not appear in the transform of Equation 10. This result means that one can use a modified $\chi^2$ approach. Use the usual $\chi^2$, but, when derivatives are taken to find the $\chi^2$ minimum, omit the derivatives of the inverse error matrix. The result is identical to the result from ML. The modified $\chi^2$ method should be generally used in place of the regular $\chi^2$ method. In practice, since the differences are not precisely the expectation values for a given experiment, there is a small residual higher order effect, which causes no bias on the average. \section{Review of G. D'Agostini's models} The problem he discusses is a bit different than that treated in the toy model. He imagines that we have two measurements of the same physical quantity, but that there is a possible scale error $f$ and a best value $k$ of two measurements, $x_1$ and $x_2$ to be fit. The models presented by D'Agostini can be written in the form: \begin{equation} \chi^2_n= \frac{(fx_1-k)^2}{f^n\sigma_1^2 }+ \frac{(fx_2-k)^2}{f^n\sigma_2^2} +\frac{(f-1)^2}{\sigma_f^2} = \frac{(x_1 -k/f)^2}{f^{n-2}\sigma_1^2} + \frac{(x_2-k/f)^2}{f^{n-2}\sigma_2^2} + \frac{(f-1)^2}{\sigma_f^2}. \end{equation} He treats the cases n=2 (Model A) and n=0 (Model B). We will also discuss the case $n=-1$. D'Agostini finds that $n=2$ does not exhibit PPP, but $n=0$ does exhibit it. There are two errors in the method of D'Agostini, which we have already mentioned in the previous section. \begin{itemize} \item The use of the $\chi^2$ distribution incorrectly ignores the changes of normalization of the multidimensional density distribution as the normalization parameter is changed. \item The normalization parameter $N$ should be included in the theoretically expected value, not in the data value. The experimentally observed number of events is what it is. D'Agostini's $f = 1/N$. This has two effects. The first effect is that the normalization dependence of the error matrix is changed. The second effect is that the average of $N$ is not the same as the average of $1/N$. \end{itemize} First consider the ML solution. Using $N$ as normalization, \begin{equation} \chi^2 = \frac{(x_1-Nk)^2}{N^{2-n}\sigma_1^2} + \frac{(x_2-Nk)^2}{N^{2-n}\sigma_2^2} +\frac{(N-1)^2}{\sigma_N^2}. \end{equation} It is assumed here that $\sigma_N^2$ is a fixed number, rather than having $\sigma_f^2$ fixed. Let \begin{equation} \chi^{2*} =\chi^2 -\frac{(N-1)^2}{\sigma_N^2}. \end{equation} The derivative of the numerator of $\chi^2$ with respect to $N$ is: \begin{equation} \frac{2(Nk-x_1)}{N^{2-n}\sigma_1^2} + \frac{2(Nk-x_2)}{N^{2-n}\sigma_2^2} +\frac{2(N-1)}{\sigma^2_N}. \end{equation} The derivative of the denominator is: \begin{equation} \frac{n-2}{N}\chi^{2*}. \end{equation} For ML the $N$ dependent part of the normalization term is $(1/\sqrt{N^{2-n}})^2$. The log of this term is $-(2-n)\ln N$ and the derivative of the log with respect to $N$ is $(n-2)/N$. For ML then: \begin{equation} \frac{\partial {\rm ML}}{\partial N} = \frac{n-2}{N} - \frac{1}{2}\big( \frac{2(Nk-x_1)}{N^{2-n}\sigma_1^2} + \frac{2(Nk-x_2)}{N^{2-n}\sigma_2^2} +\frac{2(N-1)}{\sigma^2_N}+ \frac{(n-2)}{N}\chi^{2*}\big). \end{equation} Here, the expectation value of the $\chi^{2*}$ term is 1 after fitting and the normalization term is reduced to $(n-2)/(2N)$ to account for the loss of a degree of freedom. For any $n$, the ML normalization term cancels the expectation value of the denominator derivative. Next look at this using D'Agostini's calclulation. For any $n$ value, the derivative with respect to $k$ is: \begin{equation} \frac{\partial \chi^2_n}{\partial k} = \frac{2}{f^{n-1}}[\frac{(k/f-x_1)}{\sigma_1^2} + \frac{(k/f-x_2)}{\sigma_2^2}]=0. \end{equation} Hence, \begin{equation} k =f(\frac{x_1}{\sigma_1^2}+\frac{x_2}{\sigma_2^2})/(\frac{1}{\sigma_1^2}+\frac{1}{\sigma_2^2}), \end{equation} which is the expected result from combining two measurements of the same quantity, except for the factor $f$. Define the result for $f=1$ to be $\overline{x} $. \begin{equation} \overline{x} = (\frac{x_1}{\sigma_1^2}+\frac{x_2}{\sigma_2^2})/(\frac{1}{\sigma_1^2}+\frac{1}{\sigma_2^2}), \end{equation} Note that for $\frac{\partial \chi^2}{\partial f}$, the derivative of the numerators of the first two terms together (using $\frac{(fx_1-k)^2}{f^n\sigma_1^2 }+ \frac{(fx_2-k)^2}{f^n\sigma_2^2}) $ has been determined to be zero from the $\frac{\partial \chi^2}{\partial k}$ derivative. \subsection{$n=2$, Model A} Using the result from the derivative with respect to $k$, it is seen that for the derivative with respect to $f$, (using the 2nd expression in Equation 11 with $f^{n-2}=1$ in the denominator), the derivatives of the first two terms add to be zero from the result of the derivative with respect to $k$ seen in Equation 17, and then $f$ is forced to be 1. D'Agostini finds that this does not have a PPP problem as expected since the variance is independent of $f$. \subsection{$n=0$, Model B} \begin{equation} \chi^2_B = \frac{(fx_1-k)^2}{\sigma_1^2 }+ \frac{(fx_2-k^2)}{\sigma_2^2} +\frac{(f-1)^2}{\sigma_f^2} = \frac{(x_1 -k/f)^2}{\sigma_1^2} + \frac{(x_2-k/f)^2}{\sigma_2^2} + \frac{(f-1)^2}{\sigma_f^2}. \end{equation} \begin{equation} \frac{\partial \chi^2_B}{\partial k} = 2[\frac{(k -fx_1)}{\sigma_1^2} + 2[\frac{(k-fx_2)}{\sigma_2^2}]. \end{equation} Here, $f$ will not be one. Using the result from the partial derivative with respect to $k$, $\chi^2_B$ can be written: \begin{equation} \frac{\partial \chi^2_{B}}{\partial f} = 2f^2[\frac{(x_1-\overline{x})^2}{\sigma_1^2} +\frac{(x_2-\overline{x})^2}{\sigma_2^2}] +2\frac{(f-1)}{\sigma^2_f}. \end{equation} \begin{equation} \frac{1}{f} = \sigma_f^2[\frac{1}{\sigma_f^2} +\frac{(x_1-\overline{x})^2}{\sigma_1^2} +\frac{(x_2-\overline{x})^2}{\sigma_2^2}]. \end{equation} \begin{equation} f= 1/\big[1+\sigma_f^2\big(\frac{(x_1-\overline{x})^2}{\sigma_1^2} +\frac{(x_2-\overline{x})^2}{\sigma_2^2} \big)\big]. \end{equation} \begin{equation} x_1-\overline{x} = x_1 - (\frac{x_1}{\sigma_1^2} +\frac{x_2}{\sigma_2^2} ) /(\frac{1}{\sigma_1^2} + \frac{1}{\sigma_2^2}) =\frac{x_1-x_2}{\sigma_2^2(1/\sigma_1^2+1/\sigma_2^2)}. \end{equation} Similarly, $$x_2-\overline{x} = \frac{x_2-x_1}{\sigma_1^2(1/\sigma_1^2 +1/\sigma_2^2)}.$$ To find $f$, consider: \begin{equation} \frac{(x_1-\overline{x})^2}{\sigma_1^2} + \frac{(x_2-\overline{x})^2}{\sigma_2^2} = \frac{(x_1-x_2)^2}{\sigma_1^2\sigma_2^4(1/\sigma_1^2+\sigma_2^2)^2} + \frac{(x_1-x_2)^2}{\sigma_1^4\sigma_2^2(1/\sigma_1^2+\sigma_2^2)^2} =\frac{(x_1-x_2)^2}{\sigma_1^2+\sigma_2^2}. \end{equation} \begin{equation} f= \frac{1}{1+\sigma_f^2(x_1-x_2)^2/(\sigma_1^2+\sigma_2^2)}. \end{equation} $f$ is always less than one. This is the result obtained by D'Agostini. \subsection{$n=-1$, the Toy Model} Use the notation of D'Agostini. Again the first two terms of $\frac{\partial \chi^2_{n=-1}}{\partial f}$ are zero. \begin{equation} \frac{\partial \chi^2_{n=-1}}{\partial f} = \frac{1}{f}[\frac{(fx_1-k)^2}{f^{-1}\sigma_1^2} + \frac{(fx_2-k)^2}{f^{-1}\sigma_2^2}]+\frac{2(f-1)}{\sigma_f^2}. \end{equation} The expectation value of the first two terms is $\frac{2}{f}$. \begin{equation} \frac{\partial \chi^2_{n=-1}}{\partial f} \approx \frac{2}{f}+\frac{2(f-1)}{\sigma_f^2}. \end{equation} This will be far from $f=1$, unless $\sigma_f<<1$. However, the ML term is $\frac{1}{f}$. \begin{equation} \frac{\partial \ln\mathcal{L}_{n=-1}}{\partial f} = \frac{1}{f}-\frac{\chi^2_{n=-1}}{2}\approx \frac{1}{f}-\frac{1}{f} -\frac{(f-1)}{\sigma_f^2}. \end{equation} For the ML method, $f = 1$. \section{Summary} The PPP problem arises because the $\chi^2$ method incorrectly ignores the normalizations of the multidimensional probability density functions when the total expected number of events is not fixed. For an event histogram the maximum likelihood method is correct if: \begin{itemize} \item{} Errors are taken as the square root of the theory model; they are not to be taken as the square root of the number of events in the bin. \item{} The normalization factor is included with the theory model. \item{} The subtraction for noise is included with the theory model.The data is the number of events obtained experimentally. All corrections belong to the theory model. \end{itemize} This ML result is completely equivalent to a modified $\chi^2$ approach. Use the usual $\chi^2$, but, when derivatives are taken to find the $\chi^2$ minimum, omit the derivatives of the inverse error matrix. \end{document}

https://arxiv.org/abs/2103.14014

How does the chromatic number of a random graph vary?

How does the chromatic number of a graph chosen uniformly at random from all graphs on $n$ vertices behave? This quantity is a random variable, so one can ask (i) for upper and lower bounds on its typical values, and (ii) for bounds on how much it varies: what is the width (e.g., standard deviation) of its distribution?On (i) there has been considerable progress over the last 45 years; on (ii), which is our focus here, remarkably little. One would like both upper and lower bounds on the width of the distribution, and ideally a description of the (appropriately scaled) limiting distribution. There is a well known upper bound of Shamir and Spencer of order $\sqrt{n}$, improved slightly by Alon to $\sqrt{n}/\log n$, but no non-trivial lower bound was known until 2019, when the first author proved that the width is at least $n^{1/4-o(1)}$ for infinitely many $n$, answering a longstanding question of Bollobás.In this paper we have two main aims: first, we shall prove a much stronger lower bound on the width. We shall show unconditionally that, for some values of $n$, the width is at least $n^{1/2-o(1)}$, matching the upper bounds up to the error term. Moreover, conditional on a recently announced sharper explicit estimate for the chromatic number, we improve the lower bound to order $\sqrt{n} \log \log n /\log^3 n$, within a logarithmic factor of the upper bound.Secondly, we will describe a number of conjectures as to what the true behaviour of the variation in $\chi(G_{n,1/2})$ is, and why. The first form of this conjecture arises from recent work of Bollobás, Heckel, Morris, Panagiotou, Riordan and Smith. We will also give much more detailed conjectures, suggesting that the true width, for the worst case $n$, matches our lower bound up to a constant factor. These conjectures also predict a Gaussian limiting distribution.

\section{Introduction} Given a graph $G$, a \emph{colouring} of $G$ is an assignment of colours to the vertices of $G$ so that no two adjacent vertices are coloured the same. The smallest number of colours for which this is possible is called the \emph{chromatic number} of $G$, and is denoted by $\chi(G)$. This graph parameter plays a very important role in applications, in particular in assignment problems. Here, however, we focus on $\chi(G)$ from a theoretical point of view, simply as a natural and fundamental parameter of a graph. As with any important graph parameter, an interesting question is: what is its typical value, if we choose $G$ uniformly at random from all graphs on $n$ (labelled) vertices? Also, how much does the chromatic number fluctuate around this critical value? Given $n \in \N$ and $p \in [0,1]$, the \emph{binomial random graph} $\Gnp$ is the graph on $n$ labelled vertices where each possible edge is included independently with probability $p$, so a uniformly random graph on $n$ vertices is simply $G_{n,1/2}$. The question just described was raised (in the sparse setting) by Erd\H{o}s and R\'enyi~\cite{erdos1960evolution}, in one of their seminal papers which initiated the study of random graphs. Erd\H{o}s later posed this question for the dense case, see Bollob\'as~\cite{bollobas:concentrationfixed}. In this section we first outline the history of this problem, concentrating on the most relevant results. Then we state our new results. Finally, we present a number of conjectures as to the true behaviour of $\chi(G_{n,1/2})$, in various levels of detail. The basic conjecture is due to Bollob\'as, Morris, Panagiotou and Smith together with the present authors; the finer conjectures are new. \subsection{Past results and questions} In 1975, Grimmett and McDiarmid~\cite{grimmett1975colouring} found the likely order of magnitude of $\chi(G_{n,p})$ for $0<p<1$ constant. In a landmark contribution in 1987, Bollob\'as~\cite{bollobas1988chromatic} determined the asymptotic behaviour of $\chi(G_{n,p})$ in this case. In stating this result we follow a standard convention, writing $q$ for $1-p$ and $b$ for $1/q=1/(1-p)$ to make the formulae more compact. \begin{theorem}[\cite{bollobas1988chromatic}]\label{theorem:bollobas87} Let $0<p<1$ be constant, and let $b=1/(1-p)$. With high probability\footnote{As usual, we say that a sequence $(E_n)_{n \in \N}$ of events holds \emph{with high probability (whp)} if $\Pb(E_n) \rightarrow 1$ as $n \rightarrow \infty$.}, \[ \pushQED{\qed} \chi(G_{n,p}) \sim \frac{n}{2 \log_b n}. \qedhere \popQED \] \end{theorem} Formally, this means that for any constant $\eps>0$, with high probability $\chi(G_{n,p})$ is between $1-\eps$ and $1+\eps$ times the bound on the right-hand side. Theorem~\ref{theorem:bollobas87} has been sharpened several times~\cite{mcdiarmid1989method,panagiotou2009note,fountoulakis2010t}, most recently in~\cite{heckel2018chromatic}. \begin{theorem}[\cite{heckel2018chromatic}]\label{theorem:bounds} Fix $p \le 1-1/e^2$. Then, whp, \begin{equation}\label{eq:estimate} \chi(\Gnp)=\frac{n}{2\log_b n - 2 \log_b \log_b n-2\log_b 2}+o \left( \frac{n}{\log^2 n}\right) \end{equation} where $b=1/(1-p)$. \end{theorem} For constant $p >1-1/e^2$, there is a slightly more complicated expression which also determines $\chi(\Gnp)$ whp up to accuracy $o \left( \frac{n}{\log^2 n} \right)$~\cite{heckel2018chromatic}. {\L}uczak~\cite{luczak1991chromatic} extended Theorem~\ref{theorem:bollobas87} to the case $p \rightarrow 0$, giving a similar expression for $\chi(G_{n,p})$ whenever $p > C/n$ for some large enough constant $C$. All the results we have mentioned so far examine the likely \emph{value} of the chromatic number --- they give increasingly sharp upper or lower bounds for $\chi(G_{n,p})$ which hold with high probability. A separate line of enquiry asks for the \emph{concentration} of the chromatic number: even if we cannot pin down $\chi(\Gnp)$ exactly, can we say something about how much it varies? The starting point for these questions is the classic result of Shamir and Spencer from their 1987 paper~\cite{shamir1987sharp}, in which they pioneered the use of martingale concentration inequalities in probabilistic combinatorics, something which has now become a standard tool in the area. They proved that for any function $p=p(n)$, the chromatic number of $G_{n,p}$ takes one of at most about $\sqrt{n}$ consecutive values whp. \begin{theorem}[\cite{shamir1987sharp}]\label{theorem:shamirspencer} Let $p=p(n)\in (0,1)$ and $\omega(n) \rightarrow \infty$ be arbitrary functions. Then there is a sequence of intervals $[s_n, t_n]$ of length \[\ell_n := t_n - s_n \le \sqrt{n}\omega(n)\] such that, whp, \[ \pushQED{\qed} \chi(G_{n,p}) \in [s_n, t_n]. \qedhere \popQED \] \end{theorem} It is not hard to show that for certain extreme values of $p=p(n)$, Theorem~\ref{theorem:shamirspencer} is tight: Alon and Krivelevich~\cite{alon1997concentration} note that $\chi(G_{n,p})$ is not concentrated on fewer than $\Theta(\sqrt{n})$ values for $p=1-1/(10n)$. For the dense case, where $p$ is constant, Alon gave a slight improvement to intervals of length about $\frac{\sqrt{n}}{\log n}$ (\cite{alonspencer}, \S 7.9, Exercise 3; see also~\cite{scott2008concentration}). If $p$ tends to $0$ quickly enough, however, Theorem~\ref{theorem:shamirspencer} can be improved considerably. Shamir and Spencer~\cite{shamir1987sharp} showed that if $p=n^{-c}$ for $c \in (0, \frac 12)$, then $\chi(G_{n,p})$ is concentrated on at most about $n^{\frac 12 - c} \log n$ values. For $c > \frac 12$, they proved concentration on \emph{constantly} many values. {\L}uczak~\cite{luczak1991note} showed that if $c > \frac 5 6$, then $\chi(G_{n,p})$ is maximally concentrated: whp it takes one of at most two consecutive values. Finally, Alon and Krivelevich~\cite{alon1997concentration} proved two-point concentration whenever $p<n^{-c}$ with $c> \frac 12$ constant. It should be noted that none of these concentration results gives any information about the \emph{location} of the concentration intervals. In a breakthrough contribution, Achlioptas and Naor~\cite{achlioptas2005two} found two \emph{explicit} values for $\chi(G_{n,p})$ with $p=d/n$ where $d$ is constant. Later, Coja-Oghlan, Panagiotou and Steger~\cite{coja2008chromatic} extended this result to $p<n^{-\frac34 - \eps}$, giving three explicit values in this case. In view of strong results asserting sharp concentration of the chromatic number, starting in the late 1980s Bollob\'as raised, and he and Erd\H{o}s disseminated, the \emph{opposite} question: can we find any examples where the chromatic number of $G_{n,p}$ is \emph{not} very sharply concentrated? Of course there are cases where this is trivially true, such as when $p=1-1/(10n)$ as mentioned above. But what about interesting examples, and what about the most natural special case, $p=1/2$? It took quite a while for this question to appear in print. In an open problems appendix to the first edition of \emph{The Probabilistic Method}~\cite{alonspencerfirstedition}, Erd\H{o}s asked: can we prove that $\chi(\Gnh)$ is not concentrated on constantly many values? Bollob\'as reiterated this question in~\cite{bollobas:concentrationfixed}, asking for \emph{any} non-trivial results asserting a lack of concentration. The problem is also discussed in \cite{ alon1997concentration, bollobas:randomgraphs, chung1998erdos, glebov2015concentration, mcdiarmidsurvey}. The first result of this type was recently given by the first author in~\cite{heckel2019nonconcentration}: it turns out that, at least for some values of $n$, the chromatic number of $\Gnh$ is not concentrated on fewer than about $n^{\frac 14}$ values. \begin{theorem}[\cite{heckel2019nonconcentration}]\label{theorem:n14} For any constant $c< \frac 14$, there is no sequence of intervals of length $n^{c}$ which contain $\chi(\Gnh)$ with high probability. \qed \end{theorem} \subsection{Main results} In this paper, we improve the lower bound in Theorem~\ref{theorem:n14} to an almost optimal one, giving a lower bound on the concentration interval length which nearly matches the upper bound from Theorem~\ref{theorem:shamirspencer}. \begin{theorem}\label{theorem:n12} Fix $p \in (0,1)$ and $c < \frac 12 $. Then there is no sequence of intervals of length $n^{c}$ which contain $\chi(G_{n,p})$ whp. \end{theorem} It is clear from the form of the result that Theorem~\ref{theorem:n12} also holds if we replace $c$ with $\frac 12 - o(1)$ for some function $o(1)$ which tends to $0$ sufficiently slowly. Up to this vanishing term, the exponent matches the classic upper bound of $\sqrt{n}$ for the concentration interval length given by Shamir and Spencer, and Alon's improved upper bound of $\frac{\sqrt{n}}{\log n}$. Note that neither Theorem~\ref{theorem:n14} nor Theorem~\ref{theorem:n12} tells us anything about the concentration of the chromatic number of $G_{n,p}$ for any particular $n$, let alone every $n$. They only imply that whenever $[s_n, t_n]$ is a \emph{sequence} of intervals which contain $\chi(G_{n, p})$ whp, there must be at least \emph{one} long interval. Thus, these results do not rule out the unlikely scenario that the chromatic number of $\Gnp$ is spread out over about $\sqrt{n}$ values on some sparse subsequence of the integers, and is one-point concentrated everywhere else. We will prove a stronger result than Theorem~\ref{theorem:n12}, Theorem~\ref{theorem:nstar} below. To state this we introduce some notation, and review some classic results, concerning the independence number of $G_{n,p}$. A set of vertices is \emph{independent} in a graph $G$ if there are no edges of $G$ between them; the \emph{independence number} of $G$, denoted by $\alpha(G)$, is the maximum size of such a set in $G$. As before, let $q=1-p$ and $b=1/q$. For $p$ constant, $\alpha(G_{n,p})$ can be determined precisely as follows: let \begin{equation} \label{eq:a0def} \alpha_0 = \alpha_0(n) := 2 \log_b n - 2 \log_b \log_b n +2 \log_b \left( e/2 \right)+1; \end{equation} then Matula~\cite{matula1970complete,matula1972employee} and independently Bollob\'as and Erd\H os~\cite{erdoscliques} proved that \[\alpha(G_{n, p}) = \left\lfloor \alpha_0+o(1)\right \rfloor \text{ whp},\] pinning down $\alpha(G_{n, p})$ to at most two consecutive values. If we let \begin{equation}\label{alphadef} \alpha=\alpha(n):=\left \lfloor \alpha_0(n) \right \rfloor, \end{equation} then in fact for most $n$, whp $\alpha(G_{n, p}) = \alpha$. Given $t \ge 1$, we call an independent set of size $t$ a \emph{$t$-set}. Let $X_t$ count the number of $t$-sets in $G_{n,p}$, and let \begin{equation}\label{eq:mudef} \mu_t = \mu_t(n) := \E[X_t]=\binom{n}{t}q^{\binom{t}{2}}. \end{equation} If we interpret the formula above suitably for non-integer $t$, then $\alpha_0(n)$ is, to a good approximation, the value of $t$ at which $\mu_t=1$. In particular, unless $\alpha_0$ is very close to an integer, we expect many $\alpha$-sets and no $(\alpha+1)$-sets, so it is no surprise that $\alpha(\Gnp)=\alpha$ whp. With this notation, we can now state our next, more precise, result. \begin{theorem} \label{theorem:nstar} Fix $p \le 1-1/e^2$ and $\eps>0$, and let $[s_n, t_n]$ be a sequence of intervals such that $\Pr\bb{\chi(\Gnp)\in [s_n,t_n]}\ge 0.9$. Then, for each $n$ such that $\mu_{\alpha(n)}(n)< n^{1-\eps}$, there is an integer $n^* = (1+o(1)) n$ such that \[ t_{n^*} - s_{n^*} > C \frac{\sqrt{\mu_{\alpha(n^*)}(n^*)}}{ \log n^*}, \] where \[ C=C(p,\eps)=\frac{\eps \log b}{9} \] and, as usual, $b=1/(1-p)$. \end{theorem} Theorem~\ref{theorem:nstar} readily implies the case $p\le 1-1/e^2$ of Theorem~\ref{theorem:n12}: we simply pick a sequence of $n$ where $\mu_\alpha$ is close to $n$, which is certainly possible. (See, for example, Lemma 4 in~\cite{heckel2019nonconcentration}.) The case $p>1-1/e^2$ of Theorem~\ref{theorem:n12} will also follow easily from the proof of Theorem~\ref{theorem:nstar} (see the final part of \S\ref{ss:n12}). We have replaced the assumption that $\chi(\Gnp)$ is in a certain interval whp with a weaker concrete assumption, since this is what the proof allows. The specific constant $0.9$ is not optimized; see Remark~\ref{rem:tradeoff}. Theorem~\ref{theorem:nstar} still does not imply non-concentration for any particular $n$ --- this is a feature of the method --- but for every $n$ it will find some nearby $n^*$ where the concentration interval is long. In many cases we believe that the bound above is tight up to the constant factor, including the dependence on $\eps$; see Section~\ref{ss_conj}, and in particular Remark~\ref{rem:matches}. \subsubsection*{Even stronger bounds} Theorem~\ref{theorem:n12} implies that there are \emph{some} values of $n$ such that $\chi(G_{n, p})$ is not concentrated on fewer than $n^{\frac 12 - o(1)}$ values for some unspecified function $o(1)$. Can this be pushed any further towards Alon's upper bound of $\sqrt{n}/ \log n$? We focus on the case $p=\frac 12$. The main bottleneck is the form of the error term in the estimate \eqref{eq:estimate} in Theorem~\ref{theorem:bounds}, which we make essential use of in the proof of Theorem~\ref{theorem:n12}. Specifically, we use that we have an explicit estimate for $\chi(\Gnh)$, and that the derivative (w.r.t.\ $n$) of this estimate is sufficiently larger than $1 / \alpha(n)$; see Remark~\ref{rem:bottleneck} for how this affects the final bound. Konstantinos Panagiotou and the first author~\cite{HPbdd} recently announced a sharper explicit estimate for $\chi(\Gnh)$. To state this we need some definitions. \begin{definition}\label{def:tbdd} A vertex colouring of $G$ is \emph{$t$-bounded} if all colour classes have size at most $t$; the \emph{$t$-bounded chromatic number} of $G$, denoted $\chi_{t}(G)$, is the minimum number of colours in such a colouring. By an \emph{unordered ($t$-bounded) $k$-colouring} of a graph $G$, we mean a partition of $V(G)$ into $k$ independent sets (of size at most $t$). We may think of this as an equivalence class of $k$-colourings under permuting colours. Let $E_{n,k,t}$ denote the expected number of unordered $t$-bounded $k$-colourings of $\Gnh$. Then the \emph{$t$-bounded first moment threshold} of $\Gnh$ is defined to be \begin{equation}\label{ktdef} k_t(n) := \min\{k: E_{n,k,t} \ge 1 \}. \end{equation} Note that $E_{n,k,t}$ is increasing in $k$ for $k\le n$, so this definition makes sense. \end{definition} In~\cite{HPbdd} it is shown that if $a=a(n)$ is such that $n^{0.1}<\mu_a(n)<n^{1.9}$ (where $\mu_a(n)$ is defined in \eqref{eq:mudef}), then whp \begin{equation} \label{eq:announcedbounds} \chi_{a-1} (\Gnh) = k_{ a-1} + O( n^{0.99}). \end{equation} Unsurprisingly, when $\mu_\alpha$ is not too large, then $\chi_{\alpha-1}(\Gnh)$ and $\chi(\Gnh)$ are close, and then \eqref{eq:announcedbounds} (applied with $a=\alpha$) provides a good bound on the latter. For example, we trivially have that the expectation of the difference is at most $\mu_\alpha$, though we will need a much tighter bound (see Lemma~\ref{lem:as1}). Assuming (a special case of) \eqref{eq:announcedbounds}, we can prove a stronger lower bound on the non-concentration interval. \begin{theorem}\label{theorem:polylog} Suppose that \eqref{eq:announcedbounds} holds in the special case $\mu_a(n)=\Theta(n/\log^2 n)$. Then there is a constant $c>0$ so that for any sequence of intervals $[s_n, t_n]$ such that $\Pr\bb{\chi(\Gnh) \in [s_n, t_n]}\ge 0.9$, there is a sequence of integers $n^*$ such that \[ t_{n^*}-s_{n^*} \ge c\frac{\sqrt{n^*} \log\log n^*}{\log^3 n^*}. \] \end{theorem} \begin{remark}\label{rem:var} Theorem~\ref{theorem:polylog} immediately implies a corresponding lower bound on the variance of $Y_n=\chi(\Gnh)$: writing $w_n$ for $n^{1/2}\log\log n/\log^3 n$, if we take intervals $I_n$ of length $cw_n/2$ centered on the mean of $Y_n$, then there are infinitely many $n$ such that $\Pr(Y_n\notin I_n)> 0.1$, which implies $\Var(Y_n)> 0.1(cw_n/4)^2 = \Omega(w_n^2)$, so $\limsup \Var(Y_n)/w_n^2>0$. \end{remark} As we shall describe in the next section, we believe that the bound given by Theorem~\ref{theorem:polylog} is optimal up to the constant factor. \subsection{Conjectured behaviour}\label{ss_conj} The behaviour of the chromatic number of $\Gnp$ is closely linked to that of the number of large independent sets, and specifically to $X_\alpha$ and $X_{\alpha-1}$ (where $X_t$ is the number of independent $t$-sets), so we take a closer look at the distributions of these random variables. First consider $X_\alpha$. Let $\muexp=\muexp(n) = \log \mu_\alpha / \log n$, so that \begin{equation}\mu_\alpha = n^{\muexp}. \label{eq:mua} \end{equation} Standard calculations (see \S3.c in~\cite{mcdiarmid1989method}) give \begin{equation}\label{rhoalpha1} \muexp=\alpha_0 - \alpha + o(1) \in [-o(1), 1+o(1)]. \end{equation} Thus $\muexp$ behaves as shown in Figure~\ref{fig:alphaexponent}: when $\alpha_0$ is close to an integer, $\muexp$ is close to $0$. As we increase $n$, $\muexp$ increases to near $1$ (roughly linearly in $\log n$), until $\alpha_0$ gets close to the next integer. At this point $\alpha(n)$ increases by $1$ and $\muexp$ drops back to near $0$. \begin{figure}[tb] \begin{center} \begin{overpic}[width=0.9\textwidth]{alphaexponent.pdf} \put(96,-2.5){$\log n$} \put(-5,32){$\muexp(n)$} \put(-2,25.5){$1$} \put(-2,0){$0$} \end{overpic} \end{center} \caption{\label{fig:alphaexponent} The exponent $\muexp=\muexp(n)$ so that $\mu_\alpha=n^\muexp$. When $\alpha_0(n)$ is close to an integer, $\alpha(n)=\left \lfloor \alpha_0 \right \rfloor$ increases by $1$ and $\muexp$ drops from close to $1$ to close to $0$. Note that one can think of each line segment as graphing the expected number of $t$-sets of some particular size $t$ (or rather, the log of this divided by $\log n$). These lines extend above and below the strip shown in the figure, but when we are considering the largest independent set, we jump from one size to the next as $n$ increases.} \end{figure} As for $X_{\alpha-1}$, note that \begin{equation}\label{eq:mua-1} \mu_{\alpha-1} =\Theta\left(\frac{n}{ \log n} \mu_\alpha\right) = n^{1+\muexp+o(1)}. \end{equation} It turns out that both $X_\alpha$ and $X_{\alpha-1}$ are approximately Poisson for almost all $n$ (see Theorem 11.9 in~\cite{bollobas:randomgraphs}). In particular, $X_\alpha$ and $X_{\alpha-1}$ are not whp contained in any sequences of intervals shorter than $\sqrt{\mu_\alpha} = n^{\muexp/2}$ and $\sqrt{\mu_{\alpha-1}} = n^{(1+\muexp)/2+o(1)}$, respectively. \subsubsection{The Zigzag Conjecture}\label{ss:zigzag} We are now ready to state a conjecture on the correct length of the concentration interval made by Bollob\'as, Heckel, Panagiotou, Morris, Riordan and Smith. The conjecture states that the concentration interval length for $\chi(\Gnh)$ is essentially the maximum of two proposed lower bounds, one which comes from fluctuations in $X_\alpha$ and one which comes from fluctuations in $X_{\alpha-1}$, which we will describe below. We shall consider only the case $p=\frac12$, for a number of reasons. Firstly, this is the original question; secondly, this simplifies the formulae somewhat; finally, and most importantly, for some constant $p$ --- in particular when $p>1-1/e^2$ --- the chromatic number of $\Gnp$ behaves differently to the case $p=\frac12$ (see~\cite{heckel2018chromatic}), so its concentration may well behave differently too. The chromatic number of $\Gnh$ is closely linked to its independence number. Every colour class in a colouring is an independent set, and so for any graph $G$ on $n$ vertices, $\chi(G) \ge n / \alpha(G)$. In $\Gnh$, this simple bound for the value of the chromatic number is asymptotically correct: Bollob\'as' classic result implies that whp $\chi(\Gnh) \sim n / \alpha(\Gnh)$, and Theorem~\ref{theorem:bounds} states that whp, \[ \chi(\Gnh) = \frac{n}{\alpha_0 - 1- \frac{2}{\log 2}+o(1)} \approx \frac{n}{\alpha_0-3.89}. \] It is plausible that an optimal colouring of $\Gnh$ contains all or almost all $\alpha$-sets as colour classes. To see this heuristically, fix a number $k \approx \frac{n}{2 \log_2 n}$ of colours. Each essentially different colouring of the vertex set of $\Gnh$ with $k$ colours corresponds to a \emph{profile}, i.e., a sequence of sizes for the colour classes. Among all profiles, it turns out that the expected number of colourings\footnote{As before, we actually count partitions into independent sets (with a given profile), rather than colourings.} is maximised if all or almost all $\alpha$-sets are included as colour classes. (More precisely, the expectation is maximised by \emph{unrealizable} profiles containing even more $\alpha$-sets (order $n/\log n$). Although the expected number of colourings with such a profile is large, whp no such colouring exists, as there are not enough $\alpha$-sets.) We saw above that $X_\alpha$ is approximately Poisson with mean $n^\muexp$. In particular, $X_\alpha$ varies by about $\sqrt{\mu_\alpha} = n^{\muexp/2}$. If the number of available $\alpha$-sets for our colouring varies by $\sqrt{\mu_\alpha}$, intuitively the total number of colours we need should vary by at least about \begin{equation}\label{eq:firstlowerbound} \frac{\sqrt{\mu_\alpha}}{\log n} = \frac{n^{\muexp/2}}{\log n}. \end{equation} (Perhaps it is not immediately clear where the factor $\log n$ comes from. One heuristic way to see this is the following: if there are $n^{\muexp/2}$ fewer $\alpha$-sets, we can cover $n^{\muexp/2} \alpha$ fewer vertices with $\alpha$-sets and need to colour them in colour classes of size $\alpha-1$ or less. On average we colour with classes of size $\approx \alpha_0-3.89$. So each $\alpha$-set that we use covers $\Theta(1)$ extra vertices compared to a typical colour class, and hence saves $\Theta(1/\alpha_0)=\Theta(1/\log n)$ colours. This argument is an oversimplification; see \S\ref{ss_furtherconjs} for a detailed discussion.) The first part of the Zigzag Conjecture states that \eqref{eq:firstlowerbound} is indeed a lower bound for the concentration interval length of $\chi(G_{n,\frac 12})$ (see Figure~\ref{fig:conjecture}). \begin{figure}[tb] \begin{center} \begin{overpic}[width=0.9\textwidth]{conjecture.pdf} \put(96,-2.5){$\log n$} \put(-2,25.5){$\frac 12$} \put(-2,12.75){$\frac 14$} \put(-2,0){$0$} \end{overpic} \end{center} \caption{\label{fig:conjecture} Exponent of the concentration interval length (in $n$). The dashed line is the conjectured lower bound $\muexp/2$. The dotted line is the conjectured lower bound $(1-\muexp)/2$. The thicker `zigzag' line is the maximum of these two lower bounds. The Zigzag Conjecture proposes that the concentration interval length of $\chi(\Gnh)$ fluctuates between $n^{1/4+o(1)}$ and $n^{1/2+o(1)}$ along this line.} \end{figure} The second part is another conjectured lower bound which comes from the variations of $X_{\alpha-1}$, and is slightly trickier to understand. Again fix a number $k\sim \frac{n}{2\log_2 n}$ of colours, and consider the optimal colouring profile, that is, choose the number of colour classes of each possible size so that the expected number of such colourings is maximised. A reasonable guess is that $\chi(\Gnh)$ is close to the smallest $k$ such that, for the optimal colouring profile, the expected number of colourings is at least $1$; it can be shown that the expected total number of (equivalence classes under permuting colours of) colourings is then not much more than this. It turns out that the optimal profile contains $l = \Theta (n / \log n)$ colour classes of size $\alpha-1$, i.e., some constant proportion of colour classes have this size. We now make some extremely rough estimates on how much the expected number of $k$-colourings with this profile changes as $X_{\alpha-1}$ varies, at least in the highest order terms. Since we pick $l$ colour classes from the $X_{\alpha-1}$ available $(\alpha-1)$-sets, the expected number of $k$-colourings with the optimal profile should be roughly proportional to $\binom{X_{\alpha-1}}{l} \approx X_{\alpha-1}^l$. Of course, in reality, not every choice of $l$ colour classes is possible because not all $(\alpha-1)$-sets are disjoint, but the highest order term should match, or rather, they change in the same way as $X_{\alpha-1}$ varies. Consider $\Gnh$ conditioned on some typical values for $X_{\alpha-1}$ which are $r\approx\sqrt{\mu_{\alpha-1}}$ apart, first on $X_{\alpha-1}=m \approx \mu_{\alpha-1}$ and then on $X_{\alpha-1}=m-r$. In the second case, where we have $r$ fewer $(\alpha-1)$-sets, the expected number of $k$-colourings with optimum profile decreases by a factor of roughly \begin{equation} (m-r)^l / m^ l \approx \exp \left( -rl / m\right) = \exp\left( -\Theta \left( \frac{n}{\sqrt{\mu_{\alpha-1}}\log n} \right)\right). \label{eq:decrease} \end{equation} So how much does the chromatic number increase when $X_{\alpha-1}=m-r$ compared to the case $X_{\alpha-1}=m$? It can be shown (see Corollary~\ref{dnk}) that adding one colour increases the expected number of colourings by a factor of size $\exp \left(\Theta\left(\log^2 n \right) \right)$. So in order to make up for the decrease in the expectation in \eqref{eq:decrease}, we need to introduce order \[ \frac{n}{\sqrt{\mu_{\alpha-1}} \log^3 n} \] additional colours. By \eqref{eq:mua-1}, note that \begin{equation} \label{eq:secondlowerbound} \frac{n}{\sqrt{\mu_{\alpha-1}} \log^ 3 n} = \Theta \left( \frac{\sqrt{n}}{\sqrt{\mu_\alpha} \log^{5/2} n} \right) = \Theta \left( \frac{n^{(1-\muexp)/2}}{\log^{5/2} n} \right). \end{equation} The second part of the Zigzag Conjecture states that \eqref{eq:secondlowerbound} is another lower bound for the concentration interval length of $\chi(\Gnh)$ (see Figure~\ref{fig:conjecture}). Are counts of $\alpha$-sets and $(\alpha-1)$-sets the only significant sources of non-concentration of the chromatic number? A recently announced result by the first author and Konstantinos Panagiotou~\cite{HPbdd} strongly suggests this (at least for $p=\frac12$). Recall that the $t$-bounded chromatic number $\chi_{t}(G)$ is defined like the normal chromatic number except that we only allow colourings in which all colour classes have size at most $t$. The announced result is that the $(\alpha-2)$-bounded chromatic number of $G_{n, m}$ with $m= \frac 12\binom{n}{2}$ is $2$-point concentrated. In other words, once $\alpha$-sets and $(\alpha-1)$-sets are banned as colour classes, and the number of edges is fixed, the required number of colours is extremely narrowly concentrated. It is easy to see that, in $\Gnh$, the variation in the number of edges only has a very small effect on the chromatic number, accounting for fluctuations of order at most $\log n$; see \S3 of~\cite{heckel2019nonconcentration} for a simple coupling argument showing this. The full conjecture, therefore, states that the maximum of the lower bounds \eqref{eq:firstlowerbound} and \eqref{eq:secondlowerbound} is indeed the correct concentration interval length for $\chi(\Gnh)$ --- at least whenever $\muexp(n)$ is bounded away from $0$ and $1$. Ignoring terms of size $n^{o(1)}$, a simplified statement is the following. \begin{conjecture}[Zigzag Conjecture; Bollob\'as, Heckel, Morris, Panagiotou, Riordan and Smith]\label{conj:zz} Set $p=\frac 12$ and define $\muexp=\muexp(n)$ as in \eqref{eq:mua}. Let \begin{equation}\label{eq:ln} \lambda = \lambda(n) := \max\left(\frac{\muexp}{2}, \frac{1-\muexp}{2}\right). \end{equation} Then there is a sequence of intervals of length $ n^{\lambda+o(1)} $ which contains $\chi(G_{n,\frac 12})$ whp. However, for any fixed $\eps>0$ and any sequence $(I_n)_{n \in \N}$ of intervals of length $ n^{\lambda-\eps}$, we have \[ \Pb\left(\chi(\Gnh) \in I_n\right) =o(1). \] \end{conjecture} An analogous statement presumably holds for any constant $p\in (0,1-1/e^2]$, or perhaps $p\in (0,1-1/e^2)$. Conjecture~\ref{conj:zz} would imply that the concentration interval length of $\chi(\Gnh)$ fluctuates between $n^{\frac 14+o(1)}$ and $n^{\frac 12+o(1)}$ as shown in Figure~\ref{fig:conjecture}. Theorem~\ref{theorem:nstar} \emph{almost} proves the first lower bound \eqref{eq:firstlowerbound} coming from fluctuations in $X_\alpha$: we show that, for any integer $n$ with $\muexp(n)$ bounded away from $1$, there is another integer $n^*$ nearby such that \eqref{eq:firstlowerbound} holds. It is of course extremely unlikely that the width of the distribution of $\chi(\Gnh)$ is significantly different between $n$ and $n^*$, so our result presumably holds for all $n$, but we cannot prove this. \subsubsection{Further conjectures}\label{ss_furtherconjs} In this section we state a number of further conjectures refining Conjecture~\ref{conj:zz}. We will explain the intuition behind these conjectures in the appendix, \S\ref{s:intuition}. So far we have focussed on the width of the distribution as measured by concentration in an interval; here it will often be more convenient to work with the variance. Of course we expect these to be equivalent: if $Y_n:=\chi(\Gnh)$ then we \emph{expect} that $Y_n$ is concentrated on some sequence of intervals of length $\ell_n$ if and only if $\ell_n/\sigma_n\to\infty$, where $\sigma_n^2=\Var(Y_n)$. However, we do not know this, only the one-way implication that small variance implies tight concentration. We start with a conjecture on the worst case concentration width: we believe that, up to a constant factor, the lower bound given in Theorem \ref{theorem:polylog} is optimal. \begin{conjecture}\label{conj:wc} Let $p\in (0,1-1/e^2]$ be constant, let $Y_n=\chi(\Gnp)$, let $\sigma_n^2=\Var(Y_n)$, and set \begin{equation}\label{wndef} w_n := \frac{\sqrt{n}\log\log n}{\log^3 n}. \end{equation} Then \begin{equation}\label{eq:cwc} 0 < \limsup \frac{\sigma_n}{w_n} < \infty. \end{equation} Moreover, for any constant $c>0$ there is a constant $d>0$ such that along any sequence of integers $n$ with $\mu_{\alpha(n)}(n)\sim c n/\log^2n$ we have $\sigma_n\sim d w_n$. \end{conjecture} As noted in Remark~\ref{rem:var}, Theorem~\ref{theorem:polylog} implies the first inequality in \eqref{eq:cwc}, subject to \eqref{eq:announcedbounds}. We have a corresponding conjecture for the best case, although we are less confident of this, so we state only the basic form. \begin{conjecture}\label{conj:min} Let $p\in (0,1-1/e^2]$ be constant, let $Y_n=\chi(\Gnp)$, let $\sigma_n^2=\Var(Y_n)$, and set \begin{equation}\label{twdef} \widetilde{w}_n := \frac{n^{1/4}}{\log^{7/4} n}. \end{equation} Then \[ 0 < \liminf \frac{\sigma_n}{\widetilde{w}_n} < \infty. \] \end{conjecture} In fact, we believe that for most (probably all) $n$, the chromatic number is asymptotically normally distributed, with a variance that follows (a refined version of) the graph suggested by the Zigzag Conjecture. We are least confident about points close to the minima in this graph, which we call `bad'. Fix a constant $\delta>0$, and call $n$ `bad' if $n^{1/2-\delta}\le \mu_{\alpha(n)}(n)\le n^{1/2+\delta}$, and `good' otherwise. \begin{conjecture}\label{conj:normal} Let $p\in (0,1-1/e^2]$ be constant, and let $Y_n=\chi(\Gnp)$. There are functions $f(n)$ and $g(n)$ such that, at least for `good' $n$, \[ \frac{Y_n-f(n)}{g(n)} \dto N(0,1), \] where $N(0,1)$ is a standard Gaussian distribution. Moreover, $g(n)=n^{\lambda(n)+o(1)}$, where $\lambda(n)$ is defined in \eqref{eq:ln}. \end{conjecture} For good $n$, the dominant source of the variation should be (as described earlier) the variation in the number of independent sets of a certain size $a=\alpha-1$ or $a=\alpha$, depending on the parameters. Specifically, let \begin{equation}\label{andef} a(n) := \floor{\alpha_0(n)-1/2}, \end{equation} so, for good $n$, we have \[ n^{1/2+\delta+o(1)} \le \mu_{a(n)}(n) \le n^{3/2-\delta+o(1)}. \] One can alternatively take this to \emph{define} $a(n)$. For bad $n$, at least at a certain transition point that we don't identify precisely, it is not clear how to define $a(n)$. Indeed, in a certain range two sizes should contribute. However, the distribution should still be asymptotically normal, since a linear combination of two Gaussians is Gaussian. \begin{conjecture}\label{conj:whynormal} Let $p\in (0,1-1/e^2]$ be constant, let $Y_n=\chi(\Gnp)$, and let $Z_n$ be the number of independent sets of size $a(n)$ in $\Gnp$, where $a(n)$ is defined in \eqref{andef}. Then there are functions $f(n)$, $g(n)$ such that \[ \left( \frac{Y_n-f(n)}{g(n)}, \frac{Z_n-\E Z_n}{\sqrt{\Var(Z_n)}} \right) \dto (Z,Z), \] where $Z\sim N(0,1)$. \end{conjecture} In other words, knowing $f(n)$ and $g(n)$ (which we do not), the value of $Z_n$ is enough to predict $Y_n=\chi(\Gnp)$ up to an error that is $o(g(n))$, i.e., smaller order than the standard deviation. In fact, for good $n$, this $o(\cdot)$ term should be $n^{-\Omega(1)}$. We would expect the conclusion of Conjecture~\ref{conj:whynormal} to hold outside a much smaller `bad' set, perhaps only having to exclude $n$ such that $\mu_{\alpha(n)}(n)$ is $\Theta(h(n))$ for some $h(n)$ close to $n^{1/2}$. Finally, we believe that, except for `bad' $n$, we can describe the width of the distribution up to a constant factor, and in a significant fraction of cases, up to a $1+o(1)$ factor. Defining $a=a(n)$ as above, define $x=x(n)$ by \begin{equation}\label{xndef} \mu_a(n) = \frac{2xn}{a^2} = \Theta\left(\frac{xn}{\log^2 n}\right). \end{equation} The precise normalisation here is not so important; the second formula is the key one. \begin{conjecture}\label{conj:4cases} Define $a(n)$ and $x(n)$ as in \eqref{andef} and \eqref{xndef}. For good $n$, the function $g(n)$ in Conjecture~\ref{conj:normal}, or equivalently $\sigma_n=\sqrt{\Var(Y_n)}$, satisfies the following bounds, with $c_0=2/\log 2$. \noindent (i) if $x\to 0$, then \[ g(n) \sim \sqrt{\mu_a}\frac{\log\log n+\log(1/x)}{c_0 \log^2 n}, \] (ii) if $x=\Theta(1)$, then, defining $w_n$ as in \eqref{wndef}, \[ g(n)=\Theta\left(\sqrt{\mu_a}\frac{\log\log n}{\log^2 n}\right) = \Theta(w_n), \] (iii) if $x\to\infty$ with $x=n^{o(1)}$, then \[ g(n) \sim \frac{\sqrt{\mu_a}}{x}\cdot \frac{\log\log n+\log x}{c_0\log^2 n} \sim c_1 n^{1/2} \frac{\log\log n+\log x}{\sqrt{x} \log^3 n}, \] and (iv) if $x\ge (\log n)^C$ for some constant $C>0$, then \[ g(n) = \Theta\left(\frac{\sqrt{\mu_a}\log x}{x\log^2 n}\right) = \Theta\left(\frac{\sqrt{n}\log x}{\sqrt{x}\log^3 n}\right). \] \end{conjecture} Note that the four ranges above cover all good $n$, with some overlap between (iii) and (iv). (The formula for (iii) applies in case (iv) too, but simplifies to (iv) in that case.) We can give a single formula applicable in all cases, but it is not clear that this is informative -- the transition from case (i) to cases (iii)/(iv) is rather arbitrary, since in case (ii) we do not even have a guess as to what the implicit constant should be (as a function of $x$). Still, defining \[ g_0(n) = \sqrt{\mu_a} \frac{\log\log n+|\log x|}{c_0 (1+x)\log^2 n}, \] in all cases we conjecture that $g(n)=\Theta(g_0(n))$, with $\sim$ in cases (i) and (iii). \begin{remark} If $\log \mu_a / \log n$ is bounded away from $1/2$, $1$ and $3/2$, then the formulae in (i) and (iv) match our earlier heuristics \eqref{eq:firstlowerbound} and \eqref{eq:secondlowerbound} up to constant factors. Thus, cases (i) and (iv) of Conjecture~\ref{conj:4cases} refine the `zig' and `zag' parts of the Zigzag conjecture. Case (ii), and also case (iii), interpolate between these parts, describing the conjectured shape of the top of the zigzag curve. For the bottom, we haven't stated a very detailed conjecture, but extrapolating the formulae in (i) and (iv) suggests that when $\mu_{\alpha(n)}(n)=\Theta(\sqrt{n}/\log^{3/2}n)$ and $\mu_{\alpha(n)-1}(n)=\Theta(n^{3/2}/\log^{5/2}n)$ (which is within, and indeed in some sense the centre of, the `bad $n$' case), then the contributions from $(\alpha-1)$-sets and $\alpha$-sets to $g(n)$ should both be of order $n^{1/4}/\log^{7/4}n$, and this is how Conjecture~\ref{conj:min} arises. \end{remark} \begin{remark}\label{rem:matches} The agreement between the lower bound in Theorem~\ref{theorem:nstar} and case (i) of Conjecture~\ref{conj:4cases} is in some sense surprisingly strong. The formula for $t_{n^*}-s_{n^*}$ in the former matches $g(n^*)$ up to a constant factor, noting that $1/x$ is at least approximately $n^\eps$. Since we may let $\eps$ tend to zero at some rate, and the dependence on $\eps$ matches, this shows that $\sqrt{\Var(Y_n)}=\Omega(g_0(n))$ not for every $n$, but at least for some $n^*$ near any good $n$ with $\mu_a(n)\le n^{1-\gamma(n)}$, where $\gamma(n)$ is a function tending to zero at a rate that we have not determined. Similarly, it was quite a surprise to us (and not the case when we first formulated the conjectures) that we can prove a (conditional) lower bound (Theorem~\ref{theorem:polylog}) that (for a subsequence) matches the upper bound in Conjecture~\ref{conj:wc}. \end{remark} A completely satisfactory understanding of the asymptotic distribution of $\chi(\Gnh)$ would involve two further ingredients: we would like to know $g(n)$ (or $\sigma_n$) asymptotically, not just up to constant factors. It's quite possible that one could read out such a formula from our intuitive justification of the conjectures above (see \S\ref{s:intuition}), though of course we are nowhere near a proof. The second is that we would of course like to know $f(n)$ up to an additive error that is $o(g(n))$. This seems to be a much harder problem, for which we do not even have a conjecture. See the discussion in \S\ref{s:intuition}. \medskip The rest of the paper is organized as follows. First, in \S\ref{ss:outline}, we outline the general strategy of the proofs. In \S\ref{ss:framework} we state and prove a concrete `framework lemma' that formalizes this strategy, essentially giving a conditional result subject to two ingredients. In \S\ref{ss:coupling} we provide the first ingredient, a simple coupling lemma. The details of the other ingredient vary from case to case; after some preliminaries in \S\ref{ss:prelim} we provide these, and so prove Theorems~\ref{theorem:nstar} and~\ref{theorem:n12}, in \S\ref{ss:nstar} and \S\ref{ss:n12}, respectively. The (very much more involved) argument for Theorem~\ref{theorem:polylog} is given in \S\ref{s:polylog}, with the proof of the key lemmas in \S\ref{ss:polylog2} and \S\ref{ss:alphashift}. Finally, we discuss the intuition behind our more precise conjectures in \S\ref{s:intuition}. \section{Proofs} \subsection{Proof outline}\label{ss:outline} \begin{figure}[tb] \begin{center} \begin{overpic}[width=0.9\textwidth]{intervals2-blobs.pdf} \put(97,-2.5){$n$} \put(43,40.5){$f(n)$} \put(66,56.5){$f(n)\pm\Delta(n)$} \put(22,17){$s_n$} \put(22,22.5){$t_n$} \put(44.2,23){$s_{n'}$} \put(44.5,28.5){$t_{n'}$} \put(35.8,17){$\alpha r$} \put(52,22.4){$\le r$} \put(26.5,17.5){\makebox(21.5,5){\upbracefill}} \put(49.5,22.4){{\Large $\rbrace$}} \end{overpic} \end{center} \caption{\label{fig:intervals} Illustration of the basic strategy. We know $\chi(\Gnp)$ is concentrated in the (wide) grey band around a function $f(n)$ with slope more than $1/\alpha$. A coupling argument shows that for suitable $r$ it is likely that $\chi(G_{n',p})\le \chi(G_{n,p})+r$, where $n'=n+\alpha r$ (dotted lines with slope $1/\alpha$). If the concentration intervals $[s_n,t_n]$ are too short, a contradiction results.} \end{figure} Before turning to the details, we outline the method, which is in principle simple but involves significant calculation. Throughout we fix $0<p<1$. There will be two key ingredients. First, we take as an input a suitable result establishing whp \emph{concentration} of $\chi(\Gnp)$ on some explicit interval $f(n)\pm \Delta(n)$. Here the interval length $2\Delta(n)$ will be much larger than the scale on which we are aiming to establish non-concentration. It will be essential that, interpolating $f(n)$ to non-integer values, over the range of $n$ that we consider we have $\dfdn>1/\alpha$, where $\alpha=\alpha(n)=\floor{\alpha_0(n)}$ as before. (We will consider a range of values for $n$ such that $\alpha$ does not change in this `window'.) More specifically, we will suppose that \[ \dfdn \ge \frac{1}{\alpha} + \delta \] for some $\delta>0$. The second key ingredient is a simple coupling result, Lemma~\ref{lemma:coupling} below, and in particular its consequence, Corollary~\ref{corollary:coupling}, which states, slightly informally, that for $r$ not too large we may couple the random graphs $G_n=G_{n,p}$ and $G_{n'}=G_{n+\alpha r,p}$ so that with significant probability (say $> 0.4$, though we could write $> 0.99$ by changing the constants) we have $\chi(G_{n'})\le \chi(G_{n})+r$. Here we can take $r$ up to roughly $\sqrt{\mu_\alpha(n)}$, the standard deviation of the number $X_\alpha$ of $\alpha$-sets. The intuition behind this is, roughly speaking, that because the number of $\alpha$-sets varies by at least $r$, planting $r$ extra ones does not affect the distribution of our graph too much. Planting these sets in $G_{n'}$, $n'=n+\alpha r$, we can view the graph on the remaining vertices as $G_n$, giving the coupling. Suppose for the moment that $\chi(G_n)=\chi(G_{n,p})$ were in fact deterministic, equal to some function $f_0(n)$. Then the coupling just described would show that $f_0(n')\le f_0(n)+r$, i.e., (essentially) that the function $f_0(n)$ has slope $f_0'(n)$ at most $1/\alpha$. This would lead to a contradiction, considering a suitably large range of values of $n$. Indeed, by our first (concentration) assumption, $|f_0(n)-f(n)|\le \Delta(n)$. But (ignoring the variation of $\Delta(n)$ over the relevant window), a line $f_0(n)$ with slope $1/\alpha$ cannot stay this close to a curve $f(n)$ with slope at least $1/\alpha+\delta$ for more than roughly $2\Delta/\delta$ consecutive values of $n$. Of course, $\chi(G_n)$ is not deterministic. But in proving our result, we may assume that it is \emph{almost} deterministic: for each $n$, we may assume that $\chi(G_n)$ is concentrated on some interval $[s_n,t_n]$ of length $\ell_n$. With $n'=n+\alpha r$ for $r$ not too large, as before, our coupling implies that $s_{n'}\le t_n+r$, since it is reasonably likely that all inequalities in the chain $s_{n'}\le \chi(G_{n'})\le \chi(G_n)+r\le t_n+r$ hold. In turn, this gives \[ s_{n'}\le s_n + \ell_n + r. \] Can we still get a contradiction? Yes, if the numbers work out correctly. Defining $f_0(n)=s_n$, we see that between $n$ and $n'=n+\alpha r$ this function has slope \[ \frac{s_{n'}-s_n}{\alpha r} \le \frac{1}{\alpha}+\frac{\ell_n}{\alpha r}, \] so if $\ell_n/(\alpha r)\le \delta/2$, say, we will get a contradiction much as before. Hence there must be some $n$ such that $\ell_n > \alpha r\delta/2$. Note that to obtain a strong non-concentration result, we wish to take $r$ as large as possible. \subsection{The framework lemma}\label{ss:framework} In this subsection we formalize the outline above in the following lemma. We have replaced various $o(1)$ bounds here by concrete bounds for definiteness, though in the application we mostly start with $o(1)$ bounds and take $n$ large enough. In the application we will take $a=\alpha(n^-)=\alpha(n^+)$, the typical independence number of $\Gnp$. \begin{lemma}\label{lem:framework} Let $p$, $\delta$ and $\Delta$ be positive real numbers with $p<1$, and let $n^-<n^+$ and $a$ be positive integers. Let $I=[n^-,n^+]$. Suppose that the following hold. Firstly, there is some function $f(n)$ such that for each (integer) $n\in I$ we have \begin{equation}\label{fra1} \Pr\bb{ \chi(G_{n,p}) \in [f(n)-\Delta,f(n)+\Delta] } \ge 0.99. \end{equation} Secondly, for all (real) $n\in I$ we have \begin{equation}\label{fra2} \dfdn \ge \frac{1}{a} + \delta. \end{equation} Thirdly, for each $n\in I$ we have \begin{equation}\label{fra3} \Pr\bb{ \chi(G_{n,p}) \in [s_n,t_n] } \ge 0.9 \end{equation} for some integers $s_n$, $t_n$. Fourthly, there is an increasing integer-valued function $r(n)$ such that for each $n\in I$ we have a coupling of $G_{n,p}$ and $G_{n+ar(n),p}$ such that \begin{equation}\label{fra4} \Pr\bb{ \chi(G_{n+ar(n),p}) \le \chi(G_{n,p}) +r(n)} \ge 0.4. \end{equation} Finally, suppose also that \begin{equation}\label{fra5} n^+-n^- \ge 5\Delta/\delta, \text{\quad and\quad} n^+-n^-\ge 5ar(n^+). \end{equation} Then there is some integer $n\in I$ for which $t_n-s_n > \frac{a\delta r(n)}{2}$. \end{lemma} \begin{proof} We follow the plan described in the previous section, with the minor complication that we allow $r$ to vary with $n$. (This is not essential, but gives stronger results in some applications.) Throughout we write $G_n$ for $G_{n,p}$. Firstly, for $n\in I$ define $\ts_n=\max\{s_n,f(n)-\Delta\}$ and $\ttt_n=\min\{t_n,f(n)+\Delta\}$. Then by \eqref{fra1} and \eqref{fra3} we have \begin{equation}\label{tstt} \Pr\bb{ \chi(G_n) \in [\ts_n,\ttt_n] } \ge 0.89, \end{equation} and in particular this interval is non-empty, which implies that \begin{equation}\label{tsf} |\ts_n-f(n)| \le \Delta. \end{equation} Let us suppose for a contradiction that for every $n\in I$ we have \[ \ell_n := t_n-s_n \le \frac{a\delta r(n)}{2}, \] and note for later that $\ttt_n-\ts_n\le \ell_n$. Now, for any $n\in I$ such that $n'=n+ar(n)\in I$, by \eqref{fra4} and \eqref{tstt} (applied twice), with probability at least $0.4-2\times 0.11>0$ all three inequalities $\chi(G_{n'})\ge \ts_{n'}$, $\chi(G_n)\le \ttt_n$ and $\chi(G_{n'})\le \chi(G_n)+r(n)$ hold. Hence, with positive probability \[ \ts_{n'}\le \chi(G_{n'}) \le \chi(G_n) + r(n) \le \ttt_n+r(n), \] and in particular $\ts_{n'}\le \ttt_n+r(n)$. Since this is a deterministic statement, it always holds. Thus, recalling that $\ttt_n-\ts_n\le \ell_n$, we have \begin{equation}\label{tsnar} \ts_{n'} \le \ts_n + \ell_n +r(n) \le \ts_n + r(n) + \frac{a\delta r(n)}{2} = \ts_n + (n'-n)\left(\frac{1}{a}+\frac{\delta}{2}\right). \end{equation} Finally, define a sequence $(n_i)_{0\le i\le j}$ as follows: let $n_0=n^-$ and, given $n_i$, let $n_{i+1}=n_i+ar(n_i)$ unless this value exceeds $n^+$, in which case we set $j=i$ and stop. Note that by the stopping condition and the monotonicity of $r$, \begin{equation}\label{nstop} n_j > n^+ - a r(n_j) \ge n^+ - a r(n^+). \end{equation} Applying \eqref{tsnar} with $n=n_i$ (and so $n'=n_{i+1}$) for $0\le i<j$ and telescoping, we see that \[ \ts_{n_j} - \ts_{n_0} \le (n_j-n_0) \left(\tfrac{1}{a}+\tfrac{\delta}{2}\right). \] On the other hand, from \eqref{fra2}, $f(n_j)-f(n_0) \ge (n_j-n_0) \left(\tfrac{1}{a}+\delta\right)$, so writing $h(n)=f(n)-\ts_n$ we have \[ h(n_j) - h(n_0) \ge (n_j-n_0)\delta/2. \] From \eqref{nstop} we have $n_j-n_0> n^+-n^--ar(n^+)$. Hence, by \eqref{fra5}, we have $n_j-n_0 > (4/5)(n^+-n^-) \ge 4\Delta/\delta$. Thus $h(n_j)-h(n_0)> 2\Delta$, which contradicts \eqref{tsf}. \end{proof} \begin{remark}\label{rem:tradeoff} Since in some applications we obtain bounds that appear to be within a constant factor of the truth, let us comment briefly on the probability bounds in the lemma above. Let $\eps_1$ be the failure probability in \eqref{fra1}, $\eps_2$ that in \eqref{fra3}, and $\eps_3$ that in \eqref{fra4}, so $\eps_1=0.01$, $\eps_2=0.1$ and $\eps_3=0.6$ above. The argument works provided $2\eps_1+2\eps_2+\eps_3<1$. We wish to maximize $\eps_2$ (giving a stronger non-concentration conclusion). Usually $\eps_1\to 0$ as $n\to \infty$. There is a trade-off with $\eps_3$; the coupling lemma below allows us to reduce $\eps_3$ at the cost of reducing $r$. But in the end we will always need $\eps_2<1/2$. \end{remark} \subsection{The coupling argument}\label{ss:coupling} In this section we present the coupling lemma we shall use. We state it somewhat more generally than needed here; in the application we will take $a=\alpha(n)$ (the typical independence number of $G_{n,p}$). \begin{lemma} \label{lemma:coupling} Let $p \in (0,1)$ be constant, let $b=1/(1-p)$, and let $a=a(n)$ satisfy $ 1.01 \log_b n \le a \le 100 \log_b n$ and $\mu \le n^{1.99}$, where $\mu=\mu_a(n)=\binom{n}{a}(1-p)^{\binom{a}{2}}$. Then there is a coupling of the random graphs $G_n=G_{n,p}$ and $G_{n-a}=G_{n-a,p}$ with the property that \[ \Pr\bb{ \chi(G_n)\le \chi(G_{n-a})+1 } \ge 1- \frac{1+o(1)}{2\sqrt{\mu}}. \] \end{lemma} \begin{proof} Let $U$ be a uniform random subset of $V=[n]$ of size $a$. Given $U$, let $P_n$ be the random graph on $V$ with no edges inside $U$, in which each of the other $\binom{n}{2}-\binom{a}{2}$ possible edges is present independently with probability $p$. (Thus $P_n$ is $G_n$ with a random independent $a$-set `planted'.) From the definition, we may realise $G_{n-a}$ as $P_n[V\setminus U]$. Furthermore, since $U$ is an independent set in $P_n$, we have \[ \chi(P_n) \le \chi(G_{n-a}) +1. \] It remains only to show that we can couple the distributions of $P_n$ and $G_n$ to agree with sufficiently high probability. The key observation is that $P_n$ has the distribution of $G_n$ `size-biased' by the number $X_a$ of independent $a$-sets. To see this, let $H$ be any graph on $[n]$, let $q_H=\Pr(P_n=H)$, and let $X_a(H)$ be the number of independent $a$-sets in $H$. For $P_n=H$ to hold, our random set $U$ must be independent in $H$, which has probability $X_a(H)/\binom{n}{a}$. Given such a choice of $U$, exactly the right edges outside $U$ must be present. Hence \begin{equation}\label{qpH} q_H = X_a(H)\binom{n}{a}^{-1} p^{e(H)}(1-p)^{\binom{n}{2}-\binom{a}{2}-e(H)} = \frac{X_a(H)}{\mu} p^{e(H)}(1-p)^{e(H^\cc)} = \frac{X_a(H)}{\mu} p_H, \end{equation} where $p_H = \Pr(G_n=H)$. Let $\tau$ be the total variation distance between the distributions of $P_n$ and of $G_n$. Then \[ 2\tau := \sum_H |q_H-p_H| =\sum_H \left|\frac{X_a}{\mu}-1 \right|p_H = \E\left[\frac{|X_a-\mu|}{\mu}\right], \] where the expectation refers to the random graph $G_n$. Thus by Jensen's inequality (or Cauchy--Schwarz), \[ 4\tau^2 \le \mu^{-2}\E[ (X_a(G_n)-\mu)^2] = \mu^{-2}\Var[X_a(G_n)]. \] Writing $\Var[X_a(G_n)]$ as a sum (of covariances of indicator functions) over pairs $U_1$, $U_2$ of $a$-sets in $V$, the contribution from $U_1=U_2$ is at most $\mu$, while by a standard exercise the contribution from the remaining terms is $O(\mu^2a^4/n^2 + \mu a n(1-p)^{a-1})$, with the two terms corresponding to $U_1$ and $U_2$ intersecting in $2$ or $a-1$ vertices, respectively. Under our assumptions $\Var[X_a(G_n)]\sim \mu$, so $\tau \lesssim 1 / (2 \sqrt{\mu})$. Since $G_n$ and $P_n$ can be coupled to agree with probability $1-\tau$, this completes the proof. \end{proof} \begin{remark} It is perhaps interesting that the proof of our coupling lemma relies on a variance bound, i.e., an upper bound on how much $X_a(G_n)$ varies. In the end, we use the lemma to show, roughly speaking, that $\chi(G_n)$ varies \emph{at least} a certain amount, because $X_a(G_n)$ does. \end{remark} \begin{corollary} \label{corollary:coupling} Let $p \in (0,1)$ be constant, let $b=1/(1-p)$ and let $1\le a=a(n)\le n$ satisfy $ 1.02 \log_b n \le a \le 99 \log_b n$ and $\mu \le n^{1.98}$, where $\mu=\mu_a(n)=\binom{n}{a}(1-p)^{\binom{a}{2}}$. Let $r \le \sqrt{\mu}$ be an integer. Then if $n$ is large enough, there is a coupling of the random graphs $G_n=G_{n,p}$ and $G_{n+ar}=G_{n+ar,p}$ with the property that \[ \Pr\bb{ \chi(G_{n+ar})\le \chi(G_{n})+r } > 0.4. \] \end{corollary} \begin{proof} For $i= 0, \dots, r$, let $n_i=n+ai$, and let $\mu_i = \binom{n_i}{a} (1-p)^{\binom{a}{2}}$. In this notation, $n_0=n$, $n_r=n+ar$ and $\mu_0=\mu$. Since $r \le \sqrt{\mu} \le n^{0.99}$ and $a \le 99 \log_b n$, if $n$ is large enough, for all $0 \le i \le r$, \[n \le n_i \le n +n^{0.999}.\] In particular, $\log n_i = \log n + O(n^{-0.001})$, so $1.01 \log_b n_i \le a \le 100 \log_b n_i$ if $n$ is large enough. Furthermore, if $n$ is large enough, \[ \mu_i = \mu_0 \left(1+O\left( \frac{ra}{n} \right)\right)^a = \mu_0 \left(1+O\left( \frac{ra^2}{n} \right)\right) \le n^{1.99} \le n_i^{1.99}. \] So we may apply Lemma~\ref{lemma:coupling} to show that, for every $i\in \{1, \dots, r\}$, there is a coupling of the random graphs $G_{n_{i}}$ and $G_{n_{i-1}}$ such that \begin{equation}\label{eq:couplingi} \Pr\bb{ \chi(G_{n_{i}})\le \chi(G_{n_{i-1}})+1 } \ge 1-\frac{1+o(1)}{2\sqrt{\mu_{i}}} \ge 1- \frac{1+o(1)}{2\sqrt{\mu}} . \end{equation} The Gluing Lemma (which is trivial in this finite setting\footnote{Given couplings of $X$ and $Y$ and of $Y$ and $Z$, i.e., desired distributions for $(X,Y)$ and for $(Y,Z)$, construct $(X,Y,Z)$ by starting with $Y$ and, given the value of $Y$, taking the appropriate conditional distributions for $X$ and for $Z$ -- for example with conditional independence.}) implies that there is a joint coupling of the random graphs $G_{n_0}, \dots, G_{n_r}$ so that \eqref{eq:couplingi} holds for every $1 \le i \le r$. In this coupling, with probability at least \[ 1-(1+o(1))\frac{r}{2\sqrt{\mu}} \ge 1-\frac{1+o(1)}{2} > 0.4 \] we have $\chi(G_{n+ar})\le \chi(G_{n})+r$. \end{proof} \subsection{Preliminaries for Theorem~\ref{theorem:nstar}}\label{ss:prelim} Fix $p\le 1-1/e^2$, and let \begin{equation}\label{eq:deff} f(n)= f_p(n) := \frac{n}{2 \log_b n - 2 \log_b \log_b n -2 \log_b 2} \end{equation} be the estimate for $\chi(\Gnp)$ given in Theorem~\ref{theorem:bounds}. Note that \[ f(n) = \frac{n}{\alpha_0(n)-1 -\frac{2}{\log b}}, \] where $\alpha_0(n)$ was defined in \eqref{eq:a0def}. \begin{lemma}\label{lemma:derivative} Treating $\alpha_0$ and $f$ as functions of a continuous input $n$, we have \[\frac{\mathrm{d}f}{\mathrm{d}n} = \frac{1}{\alpha_0(n)}+\frac{1}{\alpha_0(n)^2} + O\left( \frac{1}{\log^3 n}\right).\] \end{lemma} \begin{proof} Elementary calculus! \end{proof} Let us note some simple properties of $\alpha_0(n)$, $\alpha(n)=\floor{\alpha_0(n)}$, and $\muexp(n)$, defined in \eqref{eq:mua}. Firstly, as noted in the introduction (see \eqref{rhoalpha1} and Figure~\ref{fig:alphaexponent}), \begin{equation}\label{rhoalpha} \muexp(n) = \alpha_0(n)-\alpha(n)+o(1). \end{equation} In other words, $\muexp$ is essentially the fractional part of $\alpha_0$. Secondly, it is immediate from the definition that $\alpha_0(n)$ is an increasing function of $n$ (for $n$ at least some constant $n_0$), and that \begin{equation}\label{alphagrow} n'\sim n \implies \alpha_0(n')=\alpha_0(n) + o(1). \end{equation} As outlined in Section~\ref{ss:framework}, we will want to compare $f'$ to $1/\alpha$. The following lemma is a convenient form of the statement, allowing us to conveniently consider all $n$ in a suitable range. \begin{lemma}\label{lemma:growthlowerbound} If $n \sim n'$, then \[ \left.\dfdn\right|_{n'} = \frac{1}{\alpha(n)}+\frac{1-\muexp(n)}{\alpha(n)^2} + o \left( \frac{1}{\log^2 n} \right) \] \end{lemma} \begin{proof} The case $n'=n$ is immediate from Lemma~\ref{lemma:derivative} and~\eqref{rhoalpha}. To see the result for $n'\sim n$, note that when we change $n$ by a factor of $1+o(1)$, from \eqref{alphagrow} the expression for $f'$ given in Lemma~\ref{lemma:derivative} changes by $o(1/\alpha_0^2)=o(1/\log^2 n)$. \end{proof} \subsection{Proof of Theorem~\ref{theorem:nstar}} \label{ss:nstar} \begin{proof}[Proof of Theorem~\ref{theorem:nstar}] Throughout we fix $p\le 1-1/e^2$ and $\eps>0$, and consider a positive integer (or rather a sequence) $n$ such that $\mu(n) := \mu_{\alpha(n)}(n) \le n^{1-\eps}$, or, equivalently, \begin{equation} \label{eq:defn1} \muexp(n) \le 1-\eps. \end{equation} We will find the required $n^*$ if $n$ is large enough. We will apply Lemma~\ref{lem:framework} with $n^-=n$. Thus from now on we write $n^-$ for our `input' value of $n$. We choose $\gamma=\gamma(n^-)$ tending to zero sufficiently slowly for various estimates below to hold, and will choose $n^+$ so that $n^+\le n^-(1+\gamma)$. Thus the desired condition $n^*\sim n^-$ will follow from the conclusion $n^*\in I=[n^-,n^+]$ of Lemma~\ref{lem:framework}. As noted above, it is immediate from the definition \eqref{eq:a0def} of $\alpha_0(n)$ that (i) $\alpha_0$ is an increasing function, and (ii) $n^+\sim n^-$ implies $\alpha_0(n^+)=\alpha_0(n^-)+o(1)$. From \eqref{rhoalpha}, the fractional part of $\alpha_0(n^-)$ is $\muexp(n^-)+o(1)$, which is at most $1-\eps/2$, say, if $n^-$ is large enough, which we assume from now on. Thus $\alpha(n^+)=\alpha(n^-)$. (In other words, the condition \eqref{eq:defn1} ensures that $\alpha_0(n)$ is not just about to pass through an integer value as we increase $n$ from $n^-$.) Let us write \[ a = \alpha(n^-), \] noting that in fact $\alpha(n)=a$ for all $n\in I$. Let $f(n)$ and $\Delta(n)$ be as in Theorem~\ref{theorem:bounds}. In particular, the error function $\Delta(n)$ is $o(n/\log^2n)$. We will take \begin{equation}\label{eq:Delta} \Delta=\max_{n\in I}\Delta(n) = o(n^-/\log^2n^-). \end{equation} By Lemma~\ref{lemma:growthlowerbound} and \eqref{eq:defn1}, if $n^-$ is large enough we have $f'(n)\ge 1/a+\delta$ for all $n \in I$, where \begin{equation}\label{eq:delta} \delta= \frac{\eps}{2a^2} \sim \frac{\eps}{8\log_b^2n^-}. \end{equation} So far, we have verified the first two conditions of Lemma~\ref{lem:framework}. For the third, by assumption we have $\Pr\bb{ \chi(G_{n,p}) \in [s_n,t_n] } \ge 0.9$, and our aim is to prove a lower bound on some $\ell_n:=t_n-s_n$. For \eqref{fra4}, we take $r(n)= \lfloor \sqrt{\mu(n)} \rfloor$ where $\mu(n)=\mu_a(n)=\binom{n}{a}(1-p)^{\binom{a}{2}}$ as usual. (Note that here $a=\alpha(n)$.) This is clearly an increasing function of $n$. Moreover, since $a\sim 2\log_b n$ the first condition of Corollary~\ref{corollary:coupling} holds with room to spare. For the second, for any $n\in I$ we have $\mu(n)=n^{\muexp(n)}$ by definition, and from \eqref{rhoalpha} and \eqref{alphagrow} we have $\muexp(n)=\muexp(n^-)+o(1)\le 1$, so $\mu(n)\le n\le n^{1.98}$. Hence Corollary~\ref{corollary:coupling} applies, establishing \eqref{fra4}. Finally, from \eqref{eq:Delta} and \eqref{eq:delta} we have $\Delta/\delta=o(n^-)$. Also, $r(n^+) \le \sqrt{\mu(n^+)}\le \sqrt{n^+}$ as above, so both lower bounds on $n^+-n^-$ in \eqref{fra5} are $o(n^-)$, and we can choose $n^+$ to satisfy these bounds as long as $\gamma(n)\to 0$ slowly enough. Thus, all conditions of Lemma~\ref{lem:framework} are met, and we conclude that there is some $n^*\in I$ such that \[ t_{n^*}-s_{n^*}\ge \frac{a\delta}{2} r(n^*) \sim \frac{\eps}{4\alpha(n^*)} \sqrt{\mu(n^*)} \sim \frac{\eps}{8\log_b n^*}\sqrt{\mu(n^*)} = \frac{\eps\log b}{8\log n^*}\sqrt{\mu(n^*)}. \] This establishes the conclusion of Theorem~\ref{theorem:nstar} if $n$ is large enough. \end{proof} \begin{remark}\label{rem:bottleneck} Let us comment briefly on how the error bound in Theorem~\ref{theorem:bounds} affects the final bounds we obtain. At first sight, it appears to play little role: the interval length we obtain depends on $\delta$ (the gradient difference) and $r=\sqrt{\mu(n)}$. However, via \eqref{fra5}, if $\Delta$ is large we need to consider a large range $I$ of possible values of $n$. This not only weakens the conclusion (finding $n^*$ far from $n$) but can cause a more serious problem: over the interval $I$ both $\mu(n)$ and $\delta(n)=f'(n)-1/\alpha(n)$ vary, so if our bound on $\Delta$ is too weak, we will not obtain a useful lower bound on $\delta$ and the argument will fail. Conversely, to obtain a final non-concentration length very close to $n^{1/2}$, we need to consider values of $n$ such that $\mu(n)$ is very close to $n$, which will only be true over a relatively short interval. So for this we need a better bound on $\Delta$. We revisit this in Section~\ref{s:polylog}. \end{remark} \subsection{Proof of Theorem~\ref{theorem:n12}}\label{ss:n12} Fix $c<1/2$, and suppose that $[s_n, t_n]$ is a sequence of intervals which contains $\chi(G_{n,p})$ whp, with interval lengths $\ell_n:=t_n-s_n$. We will show that there is an integer $n^*$ such that $\ell_{n^*} \ge \left(n^*\right)^c$, which suffices to prove Theorem~\ref{theorem:n12}. Let \begin{equation}\label{eq:epsdef} \eps = \frac{1-2c}{3} \in \left(0, \frac 13 \right). \end{equation} It is very easy to see that we can find an arbitrarily large integer $n$ such that \begin{equation}\label{rhochoice} \muexp(n) \in \left(1-2\eps, 1-\eps\right). \end{equation} (Recall from \eqref{rhoalpha} that $\muexp$ is essentially the fractional part of $\alpha_0(n)$, which increases smoothly with $n$.) By definition of $\muexp(n)$, this implies that \begin{equation} \mu(n) =n^{\muexp(n)} \in \left( n^{1-2\eps}, n^{1-\eps}\right). \label{eq:r1} \end{equation} We first consider the case $p \le 1-1/e^2$; the case $p>1-1/e^2$ will follow by some straightforward modifications which we describe at the end of the proof. By Theorem~\ref{theorem:nstar}, there is an integer $n^* \sim n$ such that \begin{equation}\label{eq:elln} \ell_{n^*} \ge C(\eps,p) \frac{\sqrt{\mu(n^*)}}{\log n^*} \end{equation} As $n^*\sim n$, it follows that $\alpha_0(n^*) = \alpha_0(n)+o(1)$, and so (by \eqref{rhoalpha} and \eqref{rhochoice}) $\alpha(n^*)=\alpha(n)$. Therefore, \[ \mu(n^*) \sim \mu(n) \left(n^*/n \right)^{\alpha(n)} = \mu(n)(1+o(1))^{O(\log n)} = \mu(n) n^{o(1)} \ge n^{1-2\eps+o(1)} = \left(n^*\right)^{1-2\eps+o(1)}. \] From \eqref{eq:elln} it follows that if $n$ is large enough, then \[ \ell_{n^*} \ge \left(n^*\right)^{\frac{1-2\eps+o(1)}{2}} > \left(n^*\right)^{\frac{1-3\eps}{2}} = \left(n^*\right)^{c}, \] as required. Now suppose that $p > 1-1/e^2$. So far in this paper, whenever we assumed $p\le 1-1/e^2$, it was only to be able to use the estimate for the chromatic number from Theorem \ref{theorem:bounds}. More specifically, we only used that we have some estimate $\chi(G_{n,p})=f(n)+o \left(\frac{n}{\log^2 n} \right)$ so that the derivative $f'(n)$ is sufficiently larger than $\tfrac 1\alpha$; namely that \begin{equation}\label{eq:whatweused} f'(n) \ge \frac{1}{\alpha(n)} + \frac{1-\muexp}{\alpha(n)^2}+o \left(\frac{1}{\log^2 n} \right). \end{equation} If $p>1-1/e^2$, \cite{heckel2018chromatic} gives a more complicated expression which also determines $\chi(G_{n,p})$ up to an error term of size $o \left( \frac {n} {\log^2n}\right)$. Fortunately, if $\muexp = \alpha_0 - \alpha +o(1)$ is close to $1$, this estimate takes a simple form which is given in the following lemma. \begin{lemma}\label{lem:estimatelargep} Fix $p> 1-1/e^2$, and let $u= \frac{2}{\log b}< 1$. For all $n$ such that $\alpha_0(n) - \alpha(n) \ge u$, whp \[ \chi(G_{n,p}) = \frac{n}{\alpha(n)-1} + o\left(\frac{n}{\log^ 2n}\right). \] \end{lemma} \begin{proof} By Theorem~1 in \cite{heckel2018chromatic}, whp \begin{equation}\label{eq:estimateplarge} \chi(G_{n,p}) = \frac{n}{\gamma(n) - x_0} + o\left(\frac{n}{\log^ 2n}\right), \end{equation} where $\gamma(n) = \alpha_0(n) -1-u$, and, letting $d = \gamma - \left \lfloor \gamma \right \rfloor$, $x_0$ is the smallest non-negative solution to \[ \varphi(x) := (1-d+x) \log (1-d+x) + (d-x)(1-d)/u \le 0. \] Suppose that $\alpha_0(n) - \alpha(n) \ge u$. Then $\left \lfloor \gamma \right \rfloor = \left \lfloor \alpha_0 - 1 - u \right \rfloor = \alpha - 1$ and $d = \gamma - \left \lfloor \gamma \right \rfloor = \alpha_0 - \alpha - u$. In particular, $u < 1- d$. Then for $0\le x \le d$, note that \begin{align*} \varphi'(x) &= \log (1-d+x) + 1 - (1-d)/u \le 1 - (1-d)/u < 0 \end{align*} As $\varphi(d)=0$, this implies that $d$ is the smallest nonnegative solution to $\varphi(x) \le 0$, and so $x_0=d$. By \eqref{eq:estimateplarge}, whp \begin{equation*} \chi(G_{n,p}) = \frac{n}{\gamma - d} + o\left(\frac{n}{\log^ 2n}\right) = \frac{n}{\left \lfloor \gamma \right \rfloor} + o\left(\frac{n}{\log^ 2n}\right) = \frac{n}{\alpha-1} + o\left(\frac{n}{\log^ 2n}\right). \end{equation*} \end{proof} Let $u = \tfrac 2 {\log b}$, and fix $\eps>0$. For $n$ large enough, if $\muexp > u+ \eps $ then $\alpha_0 - \alpha = \muexp+o(1) \ge u$ and Lemma~\ref{lem:estimatelargep} above applies. Let $f(n) = \frac{n}{\alpha(n)-1}$. If we only consider $n$ in an interval where $\alpha(n)$ is constant --- as we did the proof of Theorem~\ref{theorem:nstar} --- we have \[ f'(n) = \frac{1}{\alpha-1} = \frac{1}{\alpha} + \frac{1}{\alpha(\alpha-1)} > \frac{1}{\alpha} + \frac{1}{\alpha^2} \ge \frac{1}{\alpha}+\frac{1-\muexp}{\alpha^2}. \] Comparing this to \eqref{eq:whatweused}, all our conclusions from the case $p\le 1-1/e^2$ remain valid as long as $\muexp \in (u+\eps,1)$. To prove the statement of Theorem~\ref{theorem:n12}, we can assume $c$ is arbitrarily close to $\tfrac 12$, so by \eqref{eq:epsdef} we can make $\eps$ arbitrarily small. By \eqref{rhochoice}, we can assume that $\muexp \in (u+\eps,1)$. The rest of the proof of Theorem~\ref{theorem:n12} is unchanged from the case $p\le 1-1/e^2$. \qed \section{Proof of Theorem~\ref{theorem:polylog}}\label{s:polylog} In this section we prove our final result, Theorem~\ref{theorem:polylog}. Throughout, we fix $p=\tfrac 12$. We assume the announced result \eqref{eq:announcedbounds} of Panagiotou and the first author~\cite{HPbdd} in the special case $\mu_a(n) = \Theta (n / \log^2 n)$. The strategy is very similar to that we used for Theorem~\ref{theorem:nstar}, based on our Framework Lemma, Lemma~\ref{lem:framework}. Before turning to the details, let us outline roughly why we choose the parameters that we do, as motivation for the arguments that follow. We use the same coupling lemma as before which, in terms of the parameters of Lemma~\ref{lem:framework}, leads to choosing $r\approx \sqrt{\mu_\alpha(n)}$, where $\alpha=\alpha(n)$. We want $r$ to be as large as possible, so we try to choose $n$ so that $\mu_\alpha(n)$ is as large as possible, which turns out to be around $n/\log^2 n$. (For why this is optimal, see \S\ref{s:intuition}.) As we shall see below, in this range the difference $\delta$ between the slope of the chromatic number and $1/\alpha$ is quite small, of order $\Theta(\log\log n/\log^3 n)$. We will consider an interval of values of $n$ differing by at most a factor $1+x$ where $x\approx 1/\log n$, so that, over the range of $n$, $\mu_\alpha$, which is roughly proportional to $n^\alpha$, varies by a constant factor. This means that we need our $\Delta$ to be at most roughly $xn\delta\approx n \log\log n/\log^4 n$, to satisfy the first condition in \eqref{fra5}. The error bound from the concentration result \eqref{eq:announcedbounds} is much smaller than this. The trouble is that it applies to the $\beta$-bounded chromatic number $\chi_\beta$, where $\beta=\alpha-1$, not the chromatic number itself. However, it turns out that, by a first moment argument, we can bound $\chi$ from below by $k_\beta-O(\mu_\alpha(n)\log\log n/\log^2 n)$, where $k_\beta$ is the first moment threshold for $\beta$-bounded colourings; see Definition~\ref{def:tbdd}. Since \eqref{eq:announcedbounds} gives $\chi_\beta\le k_\beta+O(n^{0.99})$ and $\chi\le\chi_\beta$ by definition, we thus have that $\chi(\Gnh)$ is (just) close enough to $k_\beta$ for our argument to work. Throughout the section we consider (sometimes only integer, sometimes real) values of $n$ in a set $W\subset \R$ with the following property: $W$ is a disjoint union of intervals, on each of which $\alpha(n)$ is constant, where $\alpha(n)$ is defined in \eqref{eq:a0def} and \eqref{alphadef}. In short, $\alpha(n)$ is \emph{locally constant} on $W$, formally meaning that it has derivative zero. In the following arguments, there are two relevant ways that $n$ varies: within an interval, and between intervals. When we differentiate with respect to $n$, we are (by definition) working locally within an interval, and then $\alpha=\alpha(n)$ is constant. On the other hand, for asymptotics (such as the bound $\alpha(n)=O(\log n))$, the variation between intervals is relevant. Intuitively, one can think of $n$ as very large (so that various asymptotic estimates hold), and in the analysis, in particular the application of the framework lemma, it is only the variation within an interval that matters. So one should think of $\alpha(n)$ as a constant (derivative zero) that happens to be of logarithmic order. Formally, of course, there is no issue: $\frac{\dd \alpha}{\dd n}=0$ on the set $W$. The hardest part of the proof turns out to be understanding the behaviour of (a suitable approximation to) $k_\beta(n)$, where $\beta=\beta(n)=\alpha(n)-1$. The following lemma, proved in the next section, provides this. Note that we work almost all the time with $\alpha-1$ rather than $\alpha$, so to keep the formulae compact we write $\beta$ for $\alpha-1$. In fact, although we don't need it here, the same method works with no difficulty for $\alpha-2$ also. We prove the more general case since it may be useful elsewhere, but the reader may wish to simply consider $\beta=\alpha-1$. \begin{lemma}\label{lem:polylog} Define $\beta=\beta(n)=\alpha(n)-i$, where $\alpha(n)$ is defined in \eqref{alphadef} and $i\in \{1,2\}$ is constant, and let $W\subset \R$ consist of a disjoint union of intervals on each of which $\alpha(n)$ is constant. For \emph{integer} $n\in W$, define $k_\beta(n)$ as in \eqref{ktdef}. Then there is a real-valued function $k^*(n)=k^*_\beta(n)$, defined for all $n\in W$, with the following three properties: \[ k_\beta(n) = k^*(n) + O(\log^2 n) \text{ for all integer }n\in W, \] while for all real $n\in W$ we have \begin{equation}\label{noks} \frac{n}{k^*(n)} = \alpha(n) + \frac{\log(\mu_{\alpha(n)}(n))}{\log n-\log\log n} - \frac{2}{\log 2} - 1 + O(1/\log n), \end{equation} and \[ \left(\frac{\dd k^*(n)}{\dd n}\right)^{-1} = \frac{n}{k^*(n)} + \frac{2}{\log 2} + O(1/\log n). \] \end{lemma} The (somewhat lengthy) proof of Lemma~\ref{lem:polylog} is given in Section~\ref{ss:polylog2}. \begin{remark}\label{rem:noks} The formula \eqref{noks} may seem slightly mysterious; we make two observations. Firstly, $\alpha(n)$ here can be replaced by any integer $a(n)$ such that $a(n)=\alpha_0(n)+O(1)$, provided we use the same $a$ in both places. This follows from the fact that $\mu_a/\mu_{a-1}=\Theta(\log n/n)$ for $a=\alpha_0(n)+O(1)$. Secondly, a straightforward but rather tedious calculation shows that, for $a=\alpha_0(n)+O(1)$, we have \begin{equation}\label{noks0} a(n) + \frac{\log(\mu_{a(n)}(n))}{\log n-\log\log n} = \alpha_0(n) + \left(\frac{2}{\log 2}-\frac12\right)\frac{\log\log n}{\log n} + O(1/\log n). \end{equation} To make sense of this note that one can interpolate the definition of $\mu_a(n)$ to non-integer values of $a$ in a natural way. As noted above, the left-hand side is then (roughly) constant for $a$ near $\alpha_0$. Substituting in $a=\alpha_0$, we expect $\mu_{\alpha_0}$ to be close to $1$. This explains \eqref{noks0} apart from the $c\log\log n/\log n$ term. This term is only there because we have taken a simple definition of $\alpha_0$, rather than solve $\mu_{\alpha_0}=1$ very precisely; if we were to do so, we would simply have $\alpha_0(n)+O(1/\log n)$ here, but there would be minor additional complications in other formulae. Finally, we don't use this expression in \eqref{noks} because both in the proof and in the application, it is easier to work with $\alpha$ and $\mu_\alpha$ than with $\alpha_0$. \end{remark} Our next lemma, proved in \S\ref{ss:alphashift}, is the promised lower bound on $\chi(\Gnh)$ in terms of $k_\beta$. \begin{lemma}\label{lem:as1} Let $\eps>0$ be constant. Suppose that $\mu_{\alpha(n)}(n)=\Theta(n/\log^2 n)$. Then, whp, \[ \chi(\Gnh) \ge k_\beta(n) - (1+\eps)\mu_{\alpha(n)}(n)\frac{\log\log n}{c_0\log^2 n}, \] where $c_0=2/\log 2$, and $k_\beta(n)$ is defined in \eqref{ktdef}, with $\beta=\beta(n)=\alpha(n)-1$. \end{lemma} At this point we are ready to prove Theorem~\ref{theorem:polylog}, subject to the (in the first case lengthy) proofs of Lemmas~\ref{lem:polylog} and~\ref{lem:as1}, given in next two sections. \begin{proof}[Proof of Theorem~\ref{theorem:polylog}] We consider the set \begin{equation}\label{Wdef} W := \left\{ n\in \R\ :\ \frac{c_1n}{e\log^2 n} \le \mu_{\alpha(n)}(n) \le \frac{c_1n}{\log^2 n} \right\}, \end{equation} for a positive constant $c_1<\frac{1}{5c_0^2}$. The set $W$ is easily seen to be a disjoint union of intervals, one for each value of $\alpha(n)$. Our aim is to show the existence of at least one $n$ in each interval (apart perhaps from the first few) such that $\chi(\Gnh)$ is not too concentrated. First, we consider the length of a single interval $I=[n^-,n^+]\subset W$. With $a=\alpha(n)$ constant (as it is over $I$), $\mu_a(n)$ is proportional to $\binom{n}{a}$, which is asymptotically $n^a/a!$. It follows easily that $n^+=(1+\eps)n^-$ for some $\eps$ such that $(1+\eps)^a\sim e$. This gives $\eps\sim 1/a\sim 1/(c_0\log n^-)$, say. Thus \[ n^+-n^-\sim n^-/(c_0\log n^-). \] We will apply Lemma~\ref{lem:framework} to each interval, with $f(n)=k^*(n)$, where $k^*(n)$ is as in Lemma~\ref{lem:polylog}. Let $\beta=\beta(n)=\alpha(n)-1$, which is constant on each interval. We have $\chi_\beta(\Gnh) = k_\beta(n) + O(n^{0.99})$ whp by \eqref{eq:announcedbounds}.\footnote{This is the only place in the proof where we use \eqref{eq:announcedbounds}; the lemmas stated in this section do not rely on it.} By Lemma~\ref{lem:as1} and the definition of $W$, whp we have \[ \chi(\Gnh) \ge k_\beta(n) - (1+o(1))\mu_{\alpha(n)}(n)\frac{\log\log n}{c_0\log^2 n} \ge k_\beta(n) - (1+o(1)) \frac{c_1 n\log\log n}{c_0\log^4 n}. \] Thus, whp \[ k_\beta(n) - (1+o(1)) \frac{c_1 n\log\log n}{c_0\log^4 n} \le \chi(\Gnh) \le \chi_\beta(\Gnh) \le k_\beta(n)+O(n^{0.99}). \] From Lemma~\ref{lem:polylog} we have $k^*(n)-k_\beta(n)=O(\log^2 n)$ for integer $n\in W$, so it follows that for $n\in I$ we have $\chi(\Gnh)\in [f(n)-\Delta,f(n)+\Delta]$ whp, for some $\Delta$ satisfying \[ \Delta \sim \frac{c_1 n^-\log\log n^-}{c_0\log^4 n^-}. \] This establishes the first condition \eqref{fra1} of Lemma~\ref{lem:framework}.\footnote{The reader may wonder why we take $f(n)=k^*(n)$ rather than $f(n)=k_\beta(n)$. The reason is that we do not know precisely enough how the latter varies.} We set $a=\alpha(n)$, which is constant over the interval we are considering. For $n\in W$, by Lemma~\ref{lem:polylog} we have \begin{eqnarray*} \frac{1}{f'(n)} &=& \alpha(n) + \frac{\log(\mu_{\alpha(n)}(n))}{\log n-\log\log n} - 1 + O(1/\log n) \\ &=& a + \frac{\log n-2\log\log n+O(1)}{\log n-\log\log n} -1 + O(1/\log n) \\ &=& a - \frac{\log\log n}{\log n} + O(1/\log n). \end{eqnarray*} Since $a\sim c_0\log n$, it follows (using $(a-\eps)^{-1}=a^{-1}(1-\eps/a)^{-1} = a^{-1}+\eps a^{-2}+\cdots$) that \[ f'(n) = \frac{1}{a} + (1+o(1))\frac{\log\log n}{c_0^2\log^3 n}, \] so \eqref{fra2} holds for all $n\in I$ for some $\delta$ satisfying \[ \delta\sim \frac{\log\log n^-}{c_0^2\log^3 n^-}. \] As usual \eqref{fra3} is part of our assumption; we assume $\Gnh$ is concentrated like this and our aim is to give a lower bound on $t_n-s_n$ for some $n$. As before, condition \eqref{fra4} follows from our coupling result, Corollary~\ref{corollary:coupling}, taking $r(n)=\floor{\sqrt{\mu_a(n)}}$, say. Now $\Delta/\delta\sim c_0c_1n^-/\log n^-$, while $ar(n^+)$ is $O(\sqrt{n^-})$. By choice of $c_1$ we have $1/c_0>5c_0c_1$, so it follows that the inequalities in \eqref{fra5} hold for large enough $n$. Thus Lemma~\ref{lem:framework} implies that for some $n$ in each interval (except perhaps for the first $O(1)$), we have \[ t_n-s_n\ge \frac{a\delta r(n)}{2} \ge (1+o(1)) c_0\log n \frac{\log\log n}{c_0^2\log^3 n} \frac{\sqrt{c_1 n}}{2\sqrt{e}\log n} \sim c_2\frac{\sqrt{n}\log\log n}{\log^3 n} \] where $c_2=\sqrt{c_1}/(2c_0\sqrt{e})$, and we have replaced $n^-$ by $n$ since $n\sim n^-$. \end{proof} \subsection{Proof of Lemma~\ref{lem:polylog}}\label{ss:polylog2} In this section we prove Lemma~\ref{lem:polylog}. This will take some time. In principle, this is a matter of calculation, but it seems to require considerable work, and several tricks, to get the calculations to come out to the required accuracy. For the reader to refer back to later, we collect in Table~\ref{t1} some notation used in this and the next section. \begin{table}[htb] \begin{center} \begin{tabular}{|c|l|} \hline $E_{n,k,t}$ & Expected number of unordered $t$-bounded $k$-colourings. \\ $k_t(n)$ & Threshold where $E_{n,k,t}$ reaches $1$. \\ \hline $L_0(n,k,t)$ & approximation to $\log(E_{n,k,t})$ defined in \eqref{L0nktdef}\\ $\hL_0(n,k,t)$ & $\tfrac{1}{k}L_0(n,k,t)$ \\ $\tL_0(\rho,k,t)$ & defined by $\hL_0(n,k,t)=\tL_0(n/k,k,t)$ \\ \hline $k^*(n)$ & defined by solving $L_0=0$ (or $\hL_0=0$) \\ \hline \end{tabular} \end{center} \caption{The various functions involved in defining and approximating the $t$-bounded expectation threshold. While $L_0$ is the key approximation to $\log(E_{n,k,t})$, in different parts of the analysis it turns out to be much simpler to consider the transformed functions $\hL_0$ and $\tL_0$} \label{t1} \end{table} The $t$-bounded first moment threshold $k_t(n)$ is defined in terms of $E_{n,k,t}$, the expected number of $t$-bounded $k$-colourings of $\Gnh$. One key idea of the proof is to replace $E_{n,k,t}$ by a simpler quantity, and to define $k^*$ as the threshold for this simpler estimate to cross $1$. We will simplify in three simple steps, proved together in one lemma (Lemma~\ref{lemsimp} below): (i) we replace the expected number of colourings with a given profile (see below) by a simpler formula, (ii) we replace the sum over profiles by a maximum, and (iii) we replace the maximum over integer-valued profiles (a complicated set) by the maximum over a certain region in $\R^t$. To state and prove the lemma we need some notation. As before, let $\pi=(n_i)_{i=1}^{t}$ denote a (now $t$-bounded) profile, where $n_i$ represents the number of colour classes with $i$ vertices. Let $P_{n,k,t}$ denote the set of all profiles $\pi$ satisfying \begin{equation}\label{constraint} n_i\ge 0,\qquad \sum_{i=1}^t n_i=k \qquad \text{and}\qquad \sum_{i=1}^t i n_i = n. \end{equation} Thus $P_{n,k,t}$ consists of all profiles corresponding to $t$-bounded $k$-colourings. Extending to real values, given positive reals $k<n$ and a positive integer $t$, let \[ P^0_{n,k,t} = \bigl\{\ (n_i)_{i=1}^{t} \in \R^t : \text{\eqref{constraint} holds}\ \bigr\}. \] Two key quantities appearing in many places in our calculation will be \begin{equation}\label{didef} d_i := 2^{\binom{i}{2}}i! \end{equation} and \begin{equation}\label{L0nktdef} L_0(n,k,t):=\sup_{\pi \in P^0_{n,k,t}}\left\{ n\log n -n +k -\sum_{i=1}^t n_i\log(n_id_i) \right\}. \end{equation} As we now show, the latter is a good approximation to the log (hence `$L$') of $E_{n,k,t}$, the expected number of $t$-bounded $k$-colourings of $\Gnh$ (see Definition~\ref{def:tbdd}). \begin{lemma}\label{lemsimp} Suppose that $t=t(n)=O(\log n)$. For all (large enough) $n$ and for all $k$ with $1<n/k<t$ we have \[ \log(E_{n,k,t}) = L_0(n,k,t) + O(\log^4 n). \] \end{lemma} \begin{proof} For a given profile $\pi$, let $E_{\pi}$ be the expected number of unordered colourings with this profile, so by definition \[ E_{n,k,t} = \sum_{\pi\in P_{n,k,t}} E_{\pi}. \] Since the order of the parts does not matter, there are \[ \frac{1}{\prod_i n_i!} \frac{n!}{\prod_i i!^{n_i}} \] ways to partition $[n]$ into $k$ parts with $n_i$ of size $i$ for each $i$ (the second fraction is the relevant multinomial coefficient). Such a partition is indeed an unordered $k$-colouring if and only if there are no edges of $\Gnh$ within the parts. Hence \[ E_{\pi} = \frac{n!}{\prod_i n_i!} \frac{1}{\prod_i i!^{n_i}} 2^{-\sum_i n_i \binom{i}{2}} = \frac{n!}{\prod_i n_i!} \prod_i d_i^{-n_i}. \] Let \[ L_\pi := n\log n - \sum_i n_i \log(n_id_i) -n +k. \] Then using Stirling's formula it is easy to see that for any $\pi\in P_{n,k,t}$ we have \[ \log(E_{\pi}) = L_{\pi}+O(\log^2 n). \] Indeed, this follows by absorbing the (logarithm of) all $\sqrt{2\pi m}$ factors into the error term. There are at most $(n+1)^t=\exp(O(\log^2n))$ possible profiles, so $E^t_{n,k}$ is within this factor of $\max_\pi E_{\pi}$. Hence \begin{equation}\label{lnEb1} \log(E_{n,k,t}) = \max_{\pi\in P_{n,k,t}} L_\pi + O(\log^2 n). \end{equation} Now $P_{n,k,t}\subset P^0_{n,k,t}$, so the inequality $\max_{\pi\in P_{n,k,t}} L_\pi\le L_0(n,k,t)$ holds trivially. It remains to show the reverse inequality, up to a small error term. For this, let $\pi=(n_i)_{i=1}^t\in P^0_{n,k,t}$ be arbitrary. Our aim is to find a profile $\pi'\in P_{n,k,t}$ with $L_{\pi'}$ not too far from $L_\pi$. To do so, we modify $\pi$ in a series of small steps. Firstly, round each (non-integer) $n_i$ either up or down to the nearest integer, choosing whether to round up or down in such a way that after all such roundings $\sum n_i$ is unchanged. At this point, $\sum i n_i$ has changed by no more than $\sum_{i=1}^t i\le t^2$. We obtain $\pi'$ by making a number of further changes, each of which consists of altering the size of one class by $1$, i.e., decreasing some $n_i$ by $1$ and increasing either $n_{i-1}$ or $n_{i+1}$ by $1$; clearly we can fix the error in $\sum in_i$ by at most $t^2$ such changes. In total, we have made $O(t^2)$ small changes, each of which consists of altering a single value $n_i$ by at most $1$. Now each $d_i$ is at most $2^{t^2}t! = \exp(O(t^2))$. Also \[ \frac{\dd}{\dd n_i} n_i\log(n_id_i) = \log(n_id_i)+1, \] which is thus $O(\log n+t^2)=O(\log^2 n)$ for $1\le n_i\le n$. It is easy to check that $n_i\log(n_id_i)$ is $O(\log^2 n)$ for $0\le n_i\le 1$. It follows that each of the changes above (changing a single $n_i$ by at most $1$) changes $\sum n_i\log(n_id_i)$ by at most $O(\log^2 n)$. The remaining terms in $L_\pi$ are the same for $\pi'$ as for $\pi$, so we conclude that \[ |L_\pi-L_{\pi'}| = O(t^2 \log^2 n) = O(\log^4 n). \] Hence $L_0$ is within $O(\log^4 n)$ of the maximum over (integer) profiles $\pi$, which, combined with \eqref{lnEb1}, gives the result. \end{proof} At this point it will be convenient to rescale in two ways: we replace each $n_i$ by $p_i=n_i/k$, the fraction of colour-classes having size $k$ (at least, this is the interpretation when $n_i$ is an integer). We will also divide the logarithm we are considering by $k$. To formalize this, for $t$ a positive integer and $\rho$ a real number with $1<\rho<b$ define \[ \tP_{\rho,t} = \left\{\ (p_i)_{i=1}^t \in [0,1]^t\ :\ \sum_i p_i = 1\text{\quad and\quad} \sum_i i p_i = \rho\ \right\}. \] When $\rho=n/k$, this is exactly the set $P_{n,k,t}$ rescaled by replacing each $n_i$ by $p_i=n_i/k$. Note that $\tP_{\rho,t}$ is simply the set of probability distributions (or probability mass functions) on $[t]$ with expectation $\rho$. Let $t$ be a positive integer, and $\rho$ and $k$ positive reals with $1<\rho<t$. For $\bp=(p_i)_{i=1}^t \in \tP_{\rho,t}$, let \[ \tL(\rho,k,\bp) := \rho\log(\rho k)-\log k -\rho +1 -\sum_i p_i\log(p_id_i), \] and define \begin{equation}\label{tL0def} \tL_0(\rho,k,t) := \sup_{\bp \in \tP_{\rho,t}} \tL(\rho,k,\bp). \end{equation} \begin{lemma}\label{lscale} If $t$ is a positive integer and $n$ and $k$ are positive reals with $1<n/k<t$ then \[ L_0(n,k,t) = k \tL_0(\rho,k,t), \] where $\rho=n/k$. \end{lemma} \begin{proof} This is simply a matter of rescaling: for $\pi\in P^0_{n,k,t}$, letting $p_i=n_i/k$ we have \begin{multline*} \frac{L_\pi}{k} = \frac{n}{k}\log n -\sum_i \frac{n_i}{k}\log(n_id_i) -\frac{n}{k}+1 = \rho\log(\rho k)-\sum_i p_i\log(kp_id_i) - \rho + 1 \\ = \rho\log(\rho k)-\sum_i p_i\log(p_id_i) -\log k - \rho + 1, \end{multline*} since $\sum p_i=1$. The result follows from the bijection between $P^0_{n,k,t}$ and $\tP_{\rho,t}$ given by $p_i=n_i/k$. \end{proof} \begin{corollary}\label{cEnk} Suppose that $t=t(n)=O(\log n)$. For all (large enough) $n$ and for all $k$ with $1<k<n/t$ we have \[ \log(E_{n,k,t}) = k \tL_0(\rho,k,t) + O(\log^4 n), \] where $\rho=n/k$. \end{corollary} \begin{proof} Immediate from Lemmas~\ref{lemsimp} and~\ref{lscale}. \end{proof} Our next aim is to understand the location and value of the maximum of $\tL(\rho,k,\bp)$ over $\bp$. \begin{lemma}\label{lem:tLmax} Let $1<\rho<t$. Then there is a unique $\bp\in \tP_{\rho,t}$ maximizing $\tL(\rho,k,\bp)$, given by \begin{equation}\label{pixy} p_i = e^{x+iy} d_i^{-1} \end{equation} for $1\le i\le t$, where $x=x_t(\rho)$ and $y=y_t(\rho)$ satisfy \begin{equation}\label{xy1} \sum_{i=1}^t e^{x+iy} d_i^{-1} =1 \end{equation} and \begin{equation}\label{xy2} \sum_{i=1}^t i e^{x+iy} d_i^{-1} = \rho. \end{equation} Furthermore, \begin{equation}\label{tLxy} \tL_0(\rho,k,t) = \rho\log(\rho k)-\log k-\rho+1-x-\rho y. \end{equation} \end{lemma} \begin{proof} Throughout the proof $k$, $t$ and $\rho$ are fixed, and we are maximizing only over $\bp\in \tP_{\rho,t}$. Thus, the only term in $\tL(\rho,k,\bp)$ that varies is the term \[ f(\bp) = \sum_{i=1}^t -p_i\log(p_id_i). \] Note that $k$ does not appear in this expression. In contrast, $\rho$ appears implicitly via the constraint $\sum ip_i=\rho$. Hence the location of the maximum will depend on $\rho$ (and $t$) but not on~$k$. Now $-x\log(xd)$ is strictly concave as a function of $x$, so viewed as a function on $[0,1]^t$, $f(\bp)$ is a sum of concave functions and hence concave. It is thus concave also on the domain $\tP_{\rho,t}$. Thus $f(\bp)$, and hence $\tL(\rho,k,\bp)$, has a unique maximizer $\bp$. This maximizer lies in the interior of $\tP_{\rho,t}$, since the derivative of $-x\log(xd)$, namely $-\log(xd)-1$, approaches infinity as $x$ approaches~$0$.\footnote{To spell this out completely, suppose that at the maximum some $p_i=0$. To obtain a contradiction it suffices to find a direction that we can move within $\tP_{\rho,t}$ in which $p_i$ increases. Then for a small enough change in this direction, the increase in the term $-p_i\log(p_id_i)$ will outweigh the decrease in any other terms. Such a direction exists, because $\tP_{\rho,t}$ certainly contains a point $\bp'$ with $p_i'>0$, so we may choose the direction from $\bp$ to $\bp'$.} The second statement now follows easily by the method of Lagrange multipliers, viewing $f(\bp)$ as a function on $[0,1]^t$, which we wish to maximize subject to the constraints \begin{equation}\label{pconstr} \sum p_i=1\text{\quad and\quad} \sum ip_i=\rho. \end{equation} Indeed, we have \[ \frac{\partial f}{\partial p_i} = -\log(p_id_i)-1, \] so at the maximum there are $\lambda$ and $\mu$ such that \[ -\log(p_id_i)-1 = \lambda + \mu i \] for $1\le i\le t$. Rearranging and setting $y=-\mu$ and $x=-\lambda-1$ gives \eqref{pixy}. The relations \eqref{xy1} and \eqref{xy2} follow immediately from the constraints \eqref{pconstr}. Finally, to obtain \eqref{tLxy} we substitute \eqref{pixy} into the definition of $\tL_0$, noting that for this specific $\bp$ we have \[ \sum p_i\log(p_id_i) = \sum p_i(x+iy) = x + \rho y, \] again using \eqref{pconstr}. \end{proof} It is easy to see that, for a given integer $t$, \eqref{xy1} and \eqref{xy2} define $x$ and $y$ uniquely as functions of $\rho$ (where $1<\rho<t$), and furthermore that these functions $x(\rho)=x_t(\rho)$ and $y(\rho)=y_t(\rho)$ are (infinitely) differentiable. Indeed, dividing \eqref{xy2} by \eqref{xy1} gives \[ \frac{ \sum_{i=1}^t i e^{iy}d_i^{-1} } { \sum_{i=1}^t e^{iy}d_i^{-1} } = \rho. \] The left-hand side is strictly increasing and (infinitely) differentiable as a function of $y$, and tends to $1$ or to $t$ as $y$ tends to $-\infty$ or $+\infty$, respectively. Having solved this equation to determine $y_t(\rho)$, we may use \eqref{xy1}, say, to find $x_t(\rho)$. We next investigate the derivatives of $L_0$. \begin{lemma}\label{dLrk} For $t$ fixed the $2$-variable function $\tL_0(\rho,k,t)$ has partial derivatives \[ \frac{\partial}{\partial k} \tL_0(\rho,k,t) = \frac{\rho-1}{k} \text{\quad and\quad} \frac{\partial}{\partial\rho} \tL_0(\rho,k,t) = \log(\rho k)-y_t(\rho). \] \end{lemma} \begin{proof} We use \eqref{tLxy}, recalling that with $t$ fixed $x=x_t(\rho)$ and $y=y_t(\rho)$ depend only on $\rho$, not on $k$. The formula for the $k$-derivative is immediate (since then $x$, $y$ and $\rho$ are constants). For the $\rho$-derivative by elementary calculus we have \[ \frac{\partial}{\partial\rho} \tL_0(\rho,k,t) = \log(\rho k) - \frac{\dd}{\dd\rho}(x+\rho y) = \log(\rho k) - \frac{\dd x_t(\rho)}{\dd\rho} - \rho\frac{\dd y_t(\rho)}{\dd\rho} -y_t(\rho). \] At this point something miraculous-seeming happens: if we differentiate the constraint \eqref{xy1} with respect to $\rho$ we obtain \[ \sum_{i=1}^t \left( \frac{\dd x_t(\rho)}{\dd\rho} + i \frac{\dd y_t(\rho)}{\dd\rho} \right ) e^{x_t(\rho)+iy_t(\rho)}d_i^{-1} =0, \] which, using \eqref{xy1} and \eqref{xy2} simplifies to \[ \frac{\dd x_t(\rho)}{\dd\rho} + \rho \frac{\dd y_t(\rho)}{\dd\rho} =0. \] Combined with the formula above, this gives the result. \end{proof} So far, it was convenient to work in terms of $\rho=n/k$ and $k$ rather than $n$ and $k$, because certain key functions then depended only on $\rho$. However, in the end we wish to find a threshold $k^*$ as a function of $n$, so we now undo this change of variables. Noting/recalling that the definitions \eqref{tL0def} and \eqref{L0nktdef} of $\tL_0(\rho,k,t)$ and $L_0(n,k,t)$ do not require $n$ and $k$ to be integers, for $t$ a positive integer and $n$ and $k$ positive reals with $1<n/k<t$, define \begin{equation}\label{Lnkdef} \hL_0(n,k,t) := \tL_0(n/k,k,t), \end{equation} so, by Lemma~\ref{lscale}, \begin{equation}\label{L0hL0} L_0(n,k,t) = k \tL_0(n/k,k,t) = k \hL_0(n,k,t). \end{equation} \begin{lemma}\label{dLnk} For $t$ fixed the $2$-variable function $\hL_0(n,k,t)$ has partial derivatives \[ \frac{\partial}{\partial k} \hL_0(n,k,t) = -\frac{n}{k^2}(\log n-y_t(n/k)) + \frac{n}{k^2}-\frac{1}{k} \text{\quad and\quad} \frac{\partial}{\partial n} \hL_0(n,k,t) = \frac{\log n-y_t(n/k)}{k}. \] \end{lemma} \begin{proof} This is straightforward calculus: we have \[ \frac{\partial}{\partial k} \hL_0 = -\frac{n}{k^2} \frac{\partial}{\partial\rho} \tL_0 + \frac{\partial}{\partial k} \tL_0, \] and \[ \frac{\partial}{\partial n} \hL_0 = \frac{1}{k} \frac{\partial}{\partial\rho} \tL_0. \] The result thus follows from Lemma~\ref{dLrk}. \end{proof} Our next aim is to find the value of $y$; it turns out that a fairly crude bound is enough, and for this we can use a `soft' argument, rather than trying to exactly solve the constraints \eqref{xy1} and \eqref{xy2}. \begin{lemma}\label{lem:ybrho} Suppose that $\rho=t-\Theta(1)$ and $\rho\ge 2$.\footnote{This condition will be irrelevant in the end; $\rho$ and $t$ will be order $\log n$. It's needed only to rule out values of $\rho$ very close to $1$.} Then \[ y_t(\rho) = \log\left(t2^t\right) + O(1). \] \end{lemma} \begin{proof} Note that $y_t(\rho)$ is defined for any positive integer $t$ and any real $\rho$ with $1<\rho<t$. The statement is that if we restrict the parameter space to $(\rho,t)$ such that $\rho\ge 2$ and $c<t-\rho<C$ for some constants $C>c>0$, then the difference between $y_t(\rho)$ and $\log(t2^t)$ is bounded. Fix, for the moment, $\rho$ and $t$ with $1<\rho<t$, and let $y=y_t(\rho)$. Recall that $(p_i)_{i=1}^t$ with $p_i$ defined by \eqref{pixy} is a probability distribution on $[t]$ with mean $\rho$. For $2\le i\le t$, from \eqref{pixy} we have \begin{equation}\label{riform} r_i:= \frac{p_i}{p_{i-1}} = e^{y} \frac{d_{i-1}}{d_i} = \frac{e^y}{i 2^{i-1}}, \end{equation} recalling the definition \eqref{didef} of $d_i$. In particular, $(r_i)$ is a decreasing function of $i$, so the sequence $(p_i)$ is unimodal. Furthermore, for $i=t-O(1)$ we have \begin{equation}\label{riasymp} r_i = e^y \Theta\left(\frac{1}{t 2^t}\right), \end{equation} where the implicit constants do not depend on $t$ or $\rho$. We claim that, uniformly over $(\rho,t)$ with $t-\rho=\Theta(1)$, we have $r_t=r_t(\rho,t)=\Theta(1)$; then \eqref{riasymp} gives the result. To establish the claim suppose first (for a contradiction) that for fixed $c,C$ there exist $(\rho,t)$ with $c<t-\rho<C$ such that $r_t=r_t(\rho,t)$ is arbitrarily large. If $r_t\ge D$ then $r_i\ge D$ for all $2\le i\le t$, so (for large $D$) the sequence $p_i$ is rapidly increasing and the mean $\rho$ of this probability distribution is very close to $t$. We thus obtain a contradiction for some $D=D(c)$. Next suppose that, with $c<t-\rho<C$, we may choose parameters such that $r_t$ is arbitrarily small. Since $r_i=\Theta(r_t)$ for $i\ge t-2C$, say, $r_i$ is also small (say $<1/2$) for $i\ge t-2C$. Thus $p_i$ decreases rapidly on $[t-2C,t]$. If $t\ge 2C+1$ then it follows that the mean of this probability distribution is less than $t-C$, a contradiction. If $t\le 2C+1$ then we conclude that $p_i$ decreases rapidly on the whole domain $[1,t]$, which implies that the mean is less than $2$, again contradicting our assumptions. \end{proof} We also give a useful bound on $x+ty$, in a slightly more general form. \begin{lemma}\label{lem:xay} Suppose that $\rho\ge 2$, that $\rho=t-\Theta(1)$, and that $a=t+O(1)$ is a positive integer. Then \[ x_t(\rho)+a y_t(\rho) =\log(d_a)+O(1), \] where $d_a$ is defined in \eqref{didef}. \end{lemma} \begin{proof} We continue the argument in the proof of the previous lemma. As shown there, defining $r_i$ as in \eqref{riform}, we have $r_i=\Theta(r_t)=\Theta(1)$ for $i=t-O(1)$. Since $(p_i)$ is a probability distribution on $[t]$ with mean $\rho$, it follows that $p_t=\Theta(1)$. (Indeed, if $p_t=o(1)$ then we would have $p_i=o(1)$ for $i=t-O(1)$, contradicting that the mean $\rho$ is within $O(1)$ of $t$.) Now (purely as a notational convenience) extend the definition of $p_i$ to $i>t$ also, taking $p_i=e^{x+iy}d_i^{-1}$ as in \eqref{pixy}, with $x=x_t(\rho)$ and $y=y_t(\rho)$. Then \eqref{riform} holds for $i>t$ too, and (from this equation) we have $r_i=\Theta(r_t)$ for $i=t+O(1)$. Hence $p_a=\Theta(p_t)=\Theta(1)$. Taking logs, \[ \log(p_a) = x_t(\rho)+a y_t(\rho) -\log(d_a) = O(1), \] giving the result. \end{proof} We will be interested in the $\beta$-bounded chromatic number where $\beta=\alpha(n)-i=\alpha_0(n)+O(1)$, for $i=1$ (the important case for us) or $i=2$. It will turn out that the relevant values of $\rho$ (the average colour class size) are of the form $\beta-\Theta(1)$. The next corollary gives the value of $y$ in this key case. \begin{corollary}\label{cor_y} Suppose that $t=t(n)=\alpha_0(n)+O(1)$ is an integer. Uniformly over all $n$ and all real $\rho\ge 2$ such that $t-\rho=\Theta(1)$ we have \begin{equation}\label{yvalue} y_t(\rho) = 2\log n-\log\log n + O(1). \end{equation} \end{corollary} \begin{proof} We apply Lemma~\ref{lem:ybrho}, noting that, recalling \eqref{eq:a0def}, for $t=\alpha_0(n)+O(1)$ we have $2^t=\Theta(n^2/\log^2 n)$. \end{proof} Using this value of $y_t(\rho)$, and Lemma~\ref{lem:xay}, we can estimate $\hL_0$ (or $\tL_0$, which is the same function reparametrized). Recall that $\hL_0$ is defined by dividing $L_0$ (a good approximation to the logarithm of the expected number of $t$-bounded $k$-colourings) by $k$, so the $+O(1)$ error below corresponds in the end to a factor $\exp(O(k))=\exp(O(n/\log n))$. \begin{lemma}\label{L0approx} Suppose that $k<n$ are positive reals, and $t\le a$ are positive integers, such that $a,t=\alpha_0(n)+O(1)$ and $2\le n/k=t-\Theta(1)$. Then \begin{equation}\label{eq:L0approx} \hL_0(n,k,t) = \left(a-\rho -1-\frac{2}{\log 2}\right)(\log n-\log\log n)+\log(\mu_a(n))+O(1), \end{equation} where $\rho=n/k$ and, as usual, $\mu_a(n)=\binom{n}{a}2^{-\binom{a}{2}}$ is the expected number of independent $a$-sets in $\Gnh$. \end{lemma} \begin{proof} Let $\rho=n/k$, so by assumption $\rho\ge 2$ and $t-\rho=\Theta(1)$. By the formula \eqref{tLxy} from Lemma~\ref{lem:tLmax} we have \[ L:= \hL_0(n,k,t) = \tL_0(\rho,k,t) = \rho\log n -\log k -\rho +1 -(x+ay)+(a-\rho)y, \] where $x=x_t(\rho)$ and $y_t(\rho)$. By Lemma~\ref{lem:xay} we have \[ x+ay = \log(d_a)+O(1). \] Since \[ \mu_a(n) = \binom{n}{a}2^{-\binom{a}{2}} \sim \frac{n^a}{a!2^{\binom{a}{2}}} = \frac{n^a}{d_a}, \] we have $\log(d_a)=a\log n-\log(\mu_a(n))+o(1)$. Thus \begin{eqnarray} L &=& \rho\log n-\log k-\rho-a\log n+\log(\mu_a(n)) +(a-\rho)y+ O(1) \nonumber\\ &=& (a-\rho)(y-\log n) + \log(\mu_a(n))-\log k -\rho +O(1). \label{Larho} \end{eqnarray} Now by assumption \[ \rho=n/k=t-\Theta(1)=\alpha_0(n)+O(1) = 2\log_2 n -2\log_2\log n+O(1), \] since $\log_2n=\Theta(\log n)$. Thus \[ \rho = \frac{2}{\log 2}(\log n-\log \log n) +O(1). \] Also, crudely, $k=n/\rho=\Theta(n/\log n)$, so \[ \log k = \log n-\log\log n+O(1). \] Substituting the last two formulae into \eqref{Larho}, we have \[ L = (a-\rho)(y-\log n) + \log(\mu_a(n)) -\left(1+\frac{2}{\log 2}\right)(\log n-\log\log n) +O(1). \] Finally, note that $a-\rho=O(1)$ and that, from \eqref{yvalue}, $y=2\log n-\log\log n+O(1)$. Thus \[ L = \left(a-\rho -1-\frac{2}{\log 2}\right)(\log n-\log\log n)+\log(\mu_a(n))+O(1), \] as claimed. \end{proof} We can also use the value of $y$ from Corollary~\ref{cor_y} to give approximate bounds on the partial derivatives of $L_0$ and $\hL_0=L_0/k$. \begin{corollary}\label{dnk} Suppose that $t=t(n)=\alpha_0(n)+O(1)$ is an integer. Uniformly over all $k\le n/2$ such that $k=n/(t-\Theta(1))$ we have \[ \frac{\partial}{\partial k} \hL_0(n,k,t) = \Theta\left(\frac{\log^3 n}{n}\right) \text{,\qquad} \frac{\partial}{\partial n} \hL_0(n,k,t) = -\Theta\left(\frac{\log^2 n}{n}\right) \] and \[ \frac{\partial}{\partial k} L_0(n,k,t) = \frac{2}{\log 2}\log^2 n+O(\log n\log\log n). \] \end{corollary} \begin{proof} Note that the dependence of $t$ on $n$ is only relevant for the asymptotics; by definition of partial derivative, we hold $t$ constant when differentiating. (And in the end $t=\alpha(n)-1$ or $\alpha(n)-2$ will be locally constant.) The bounds on the partial derivatives of $\hL_0$ follow by substituting the value $y=2\log n+O(\log\log n)\sim 2\log n$ from \eqref{yvalue} into the conclusion of Lemma~\ref{dLnk}, noting that $n/k=\alpha_0(n)+O(1)=2\log_2 n+O(\log\log n)\sim 2\log_2 n$. For $L_0(n,k,t)=k\hL_0(n,k,t)$, calculating slightly more precisely, \begin{multline*} \frac{\partial}{\partial k} L_0 = \frac{\partial}{\partial k} (k\hL_0) = \hL_0 + k\frac{\partial}{\partial k}\hL_0 = \hL_0 + \frac{n}{k}(y_t(n/k)-\log n) + \frac{n}{k}-1 \\ = \hL_0 + \frac{n}{k}\log n +O(\log n\log\log n) = \hL_0 + \frac{2}{\log 2}\log^2 n +O(\log n\log\log n). \end{multline*} The result follows since $\hL_0(n,k,t)=O(\log n)$ by Lemma~\ref{L0approx}. \end{proof} For the rest of the section we consider a function $\beta(n)$ satisfying the following assumptions; the upper bound on $\beta$ is of no particular significance. \begin{assumption}\label{assump} The function $\beta$ is defined on a subset $W$ of $\R$ which is a union of intervals, and is constant on each interval. Furthermore, for some constant $\eps>0$ we have \[ \alpha_0(n)-1-\frac{2}{\log 2} + \eps \le \beta(n) \le \alpha_0(n)+100 \] for all large enough $n$. \end{assumption} Note in the assumptions of Lemma~\ref{lem:polylog}, we specified $\beta(n)=\alpha(n)-1$ or $\beta(n)=\alpha(n)-2$. These both satisfy Assumption~\ref{assump}, since $\beta\ge \alpha_0-3$ and $2/\log 2>2$. For $n\in W$ let \[ I_n = \left[\frac{n}{\beta-\eps/4}, \frac{n}{\alpha_0-100}\right]. \] Recall that $L_0(n,k,t)=k\hL_0(n,k,t)$, so one is zero if and only if the other is. \begin{lemma}\label{lem:k*} For each large enough (real) $n\in W$ there is a unique $k^*=k^*(n)\in I_n$ such that \begin{equation}\label{k*cond} \hL_0(n,k^*(n),\beta(n)) = 0 = L_0(n,k^*(n),\beta(n)). \end{equation} Furthermore, \[ \frac{n}{k^*(n)} = \alpha(n) -1-\frac{2}{\log 2} + \frac{\log(\mu_{\alpha(n)}(n))}{\log n-\log\log n}+O(1/\log n), \] and if $n$ is an integer then $k^*(n)-k_\beta(n)=O(\log^2 n)$. \end{lemma} \begin{proof} Keeping $n$ fixed, from Corollary~\ref{dnk}, if $n$ is large enough, then $\hL_0(n,k,\beta(n))$ is strictly increasing as a function of $k\in I_n$, with derivative $\Theta(\log^3 n/n)$. This implies uniqueness of $k^*(n)$ once we show existence. Define $k_0=k_0(n)$ by \[ \frac{n}{k_0} = \alpha(n)-1-\frac{2}{\log 2} + \frac{\log(\mu_{\alpha(n)}(n))}{\log n-\log\log n}. \] Then, recalling that $\mu_{\alpha(n)}(n)=n^{\alpha(n)-\alpha_0(n)+o(1)}$, we have $n/k_0=\alpha_0(n)-1-2/\log 2+o(1)\le \beta(n)-\eps/2$, so $k_0\in I_n$ with $\eps/4$ room to spare. By Lemma~\ref{L0approx} we have $\hL_0(n,k_0,\beta(n))=O(1)$; we chose $k_0$ so that the main term in \eqref{eq:L0approx} vanishes, leaving only the error term. Since, as a function of $k$, $\hL_0$ has derivative $\Theta(\log^3 n/n)$, it follows immediately that $k^*(n)$ exists, and that $k^*(n)-k_0(n)=O(n/\log^3 n)$. Since $k^*$ and $k_0$ are of order $n/\log n$, this translates to $n/k^*=n/k_0+O(1/\log n)$, proving the first statement. For the second statement, recall the bound \begin{equation}\label{close} \log(E_{n,k,\beta}) = k\hL_0(n,k,\beta) + O(\log^4 n) \end{equation} given by Corollary~\ref{cEnk} and \eqref{L0hL0}. Consider $k=k^*(n)+x$, where $x$ will be of larger order than $\log^2 n$ but not too large (say $o(n/\log^2 n)$). Then from the derivative bound, $\hL_0(n,k,\beta)=\Theta(x\log^3 n/n)$, so $k\hL_0(n,k,\beta)=\Theta(x\log^2 n)$. For $x$ of the magnitude indicated this quantity is $\omega(\log^4 n)$. Choosing such an $x$ so that $k$ is an integer, from \eqref{close} we conclude that $\log(E_{n,k,\beta})>0$, so $k_\beta(n)\le k=k^*(n)+x$. A similar argument with $x$ negative shows that $k_\beta(n)=k^*(n)+O(\log^2 n)$. \end{proof} \begin{lemma}\label{lem:k*diff} The function $k^*(n)$ is differentiable on $W$, and its derivative satisfies \[ \left(\frac{\dd k^*(n)}{\dd n}\right)^{-1} = \frac{n}{k^*(n)} + \frac{2}{\log 2} + O(1/\log n). \] \end{lemma} \begin{proof} The Implicit Function Theorem, applied to the continuously (in fact, infinitely) differentiable function $\hL_0(n,k,t)$ with $t$ fixed tells us that $k^*(n)$, defined by $\hL_0(n,k^*,\beta(n))=0$, is differentiable, and that its derivative is $-\frac{\partial\hL_0}{\partial n} / \frac{\partial\hL_0}{\partial k}$. Writing $\rho$ for $n/k$, by Lemma~\ref{dLnk} and Corollary~\ref{cor_y} the reciprocal of the derivative is thus \[ -\frac{\partial\hL_0}{\partial k} / \frac{\partial\hL_0}{\partial n} = \frac{n}{k}-\frac{\rho -1}{\log n-y_\beta(\rho)} = \rho + \frac{\rho+O(1)}{\log n-\log\log n+O(1)}. \] Now $\rho=\alpha_0(n)+O(1)=2\log_2 n-2\log_2\log_2 n+O(1)$, so the last fraction above is \[ \frac{2\log_2 n-2\log_2\log_2 n+O(1)}{\log n-\log\log n+O(1)} = \frac{2\log_2 n-2\log_2\log_2 n+O(1)}{(\log 2)(\log_2 n-\log_2\log_2 n)+O(1)} \] since $\log x=(\log 2)\log_2 x$, and hence $\log\log n=\log(\log_2 n)+O(1)=(\log 2)\log_2\log_2 n+O(1)$. The result follows. \end{proof} Together, Lemmas~\ref{lem:k*} and~\ref{lem:k*diff} imply Lemma~\ref{lem:polylog}, so the proof of that lemma, and hence of Theorem~\ref{theorem:polylog}, is complete. \subsection{Proof of Lemma~\ref{lem:as1}}\label{ss:alphashift} We shall prove the following sharper form of Lemma~\ref{lem:as1}, since it seems that the lower bound here is perhaps quite close to the truth (see the discussion in \S\ref{s:intuition}), so this might be useful elsewhere. \begin{lemma}\label{lem:alphashift} Suppose that $\log^5 n\le \mu_{\alpha(n)}(n)=O(n/\log^2 n)$. Then, whp, \[ \chi(\Gnh) \ge k^*(n) - (1+\eps)\frac{\mu\log\nu}{\alpha(\log n-\log\log n)}, \] where $\alpha=\alpha(n)$, $\mu=\mu(n)=\mu_{\alpha(n)}(n)$, $\nu=(n/\log n)/\mu$, $\eps=\eps(n)=O(1/\log\nu)\to 0$, and $k^*(n)=k_\beta^*(n)$ is defined in Lemma~\ref{lem:k*}. \end{lemma} Before giving the proof, we note that the result we need, Lemma~\ref{lem:as1}, follows. \begin{proof}[Proof of Lemma~\ref{lem:as1}] This is immediate from Lemma~\ref{lem:alphashift}, noting that by assumption $\nu(n)$ as defined there is $\Theta(\log n)$, so $\log\nu\sim\log\log n$, recalling that $\alpha(n)\sim c_0\log n$, and noting that by Lemma~\ref{lem:k*}, $k_\beta-k^*(n)=O(\log^2 n)$, which is much smaller than the error term we are aiming for. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:alphashift}] Let \[ \delta = C/\log \nu, \] where $C\ge 3$ is a constant that we will specify later. Let $k_0=\floor{k^*(n)-d}$, where \[ d= (1+5\delta)\frac{\mu\log\nu}{\alpha(\log n-\log\log n)}, \] so our aim is to show that whp $\chi(\Gnh)\ge k_0$. To do this, it suffices to show that whp $\Gnh$ has no proper $k_0$-colouring. Note for later that $k^*(n)=\Theta(n/\log n)$ while, recalling our assumptions on $\mu$, we have $d=O(\mu/\log n)=O(n/\log^3 n)$. Thus, crudely, $d=o(k^*/\log n)$ and it follows easily that \begin{equation}\label{nk0close} n/k_0 = n/k^*(n) +o(1). \end{equation} Let $\alpha=\alpha(n)$. By assumption, $\mu:=\mu_\alpha(n)=O(n/\log^2 n)$, so $\mu_{\alpha+1}(n)=O(\mu\log n/n)=O(1/\log n)\to 0$, and whp $\Gnh$ contains no independent sets of size $\alpha+1$. Thus it suffices to show that whp $\Gnh$ has no $\alpha$-bounded $k_0$-colouring. We will group the potential colourings (or, more precisely, partitions into independent sets), according to the number $m$ of $\alpha$-sets included. Let \[ m^+=\mu(1+\delta). \] Recalling that $X_\alpha$, the number of independent $\alpha$-sets, has mean $\mu$ and variance $O(\mu)$, we know from Chebyshev's inequality that whp $X_a\le m^+$. Thus it suffices to show that whp $\Gnh$ has no $\alpha$-bounded $k_0$-colouring using at most $m^+$ $\alpha$-sets. Let $C_m$ denote the number of partitions of $[n]=V(\Gnh)$ into exactly $k_0$ independent sets of which exactly $m$ have size $\alpha$ and none has size larger than $\alpha$. We \emph{claim} that, if $n$ is large enough, for each $m\le m^+$ we have \begin{equation}\label{ECM} \E[C_m] \le 1/n. \end{equation} Assuming this, then summing over the $m^++1=O(\mu)=o(n)$ values of $m$ and applying Markov's inequality, the proof is complete. Thus it suffices to prove \eqref{ECM}. Now a potential colouring/partition of the type counted by $C_m$ may be described as follows: we pick an unordered $m$-tuple of disjoint $\alpha$-vertex subsets of $[n]$, and then we pick a partition $P$ of the remaining $n-\alpha m$ vertices into $k_0-m$ parts of size at most $\alpha-1$. The partition gives a legal colouring if and only if the $m$ $\alpha$-sets are independent, and $P$ induces a legal colouring of the corresponding subgraph of $G$. Hence, \[ \E[C_m] = \frac{1}{m!} \binom{n}{\alpha} \binom{n-\alpha}{\alpha}\cdots\binom{n-(m-1)\alpha}{\alpha} 2^{-m\binom{\alpha}{2}} E_{n-\alpha m,k_0-m,\alpha-1}, \] where $E_{n',k',t}$ is the expected number of $t$-bounded unordered $k'$-colourings of $G_{n',1/2}$. Hence, bounding each binomial coefficient above by $\binom{n}{\alpha}$, we have \[ \E[C_m] \le \frac{\mu^m}{m!} E_{n-\alpha m,k_0-m,\alpha-1}. \] Taking logs, and using the standard bound $\mu^m/m!\le e^\mu$ (the former is one term in the expansion of the latter), we see that \[ \log\E[C_m] \le \mu + \log E_m \] where $E_m:=E_{n-\alpha m,k_0-m,\alpha-1}$. Fortunately, we have a good approximation for $\log E_m$. Recalling \eqref{nk0close} and noting from Lemma~\ref{lem:k*} that $n/k^*(n)\le \alpha(n)-2/\log 2<\alpha(n)-2$, we have $n/k_0<\alpha-1$ for $n$ large enough, and it follows that \[ \frac{n-\alpha m}{k_0-m}\le \frac{n}{k_0}< \alpha-1. \] Thus we can apply Lemma~\ref{lemsimp} to conclude that \[ \log E_m = L_0(n-\alpha m,k_0-m,\alpha-1) + O(\log^4 n), \] where $L_0$ is defined in \eqref{L0nktdef}. Unfortunately we do not have a direct formula for $L_0$ sufficiently accurate for our present purpose. Fortunately, however, we do have indirect bounds, expressed in terms of $k^*(n)$, defined in Lemma~\ref{lem:k*}. Note that we will consider a range of values $n'$ satisfying $n'\in I$, where \[ I = [n-\alpha m^+,n]. \] Since $\alpha m^+=O(\alpha\mu)=O(n/\log n)$, it follows easily that $\mu_\alpha(n')=\Theta(\mu)$ for all such $n'$. In particular, $\alpha(n')=\alpha$ does not vary over this range of $n'$, and it makes sense to consider $k^*(n')$ as in Lemma~\ref{lem:k*}, defined with $\beta=\alpha-1$, as a function of $n'$. By definition $L_0(n-\alpha m,k^*(n-\alpha m),\alpha-1)=0$ (see \eqref{k*cond}). From the last part of Corollary~\ref{dnk} we thus have \begin{equation}\label{L0k*} L_0(n-\alpha m,k_0-m,\alpha-1) \sim c_0\log^2 n(k_0-m-k^*(n-\alpha m)), \end{equation} where $c_0=2/\log 2$. Since $k_0$ is defined in terms of $k^*(n)$, the next step is to consider how $k^*(n')$ varies as $n'$ varies between $n$ and $n-\alpha m$. Now by Lemma~\ref{lem:polylog}, for $n'\in I$ we have \[ \left(\frac{\dd k^*(n')}{\dd n'}\right)^{-1} = \alpha+\frac{\log(\mu_{\alpha}(n'))}{\log n'-\log\log n'} - 1 + O(1/\log n') , \] recalling that $\alpha(n')=\alpha$ for all $n'\in I$. For $n'\in I$ we have $\log n'=\log n+o(1)$ and, as noted above, $\mu_\alpha(n')=\Theta(\mu)$. It follows that \[ \left(\frac{\dd k^*(n')}{\dd n'}\right)^{-1} = \alpha+\frac{\log\mu}{\log n-\log\log n} - 1 + O(1/\log n) = \alpha-\frac{\log\nu}{\log n-\log\log n} + O(1/\log n), \] recalling that $\nu=(n/\log n)/\mu$. Thus, \[ \left(\frac{\dd k^*(n')}{\dd n'}\right)^{-1} \ge \alpha-(1+\delta)\frac{\log\nu}{\log n-\log\log n}, \] provided the constant $C$ appearing in the definition of $\delta$ is chosen large enough. We now take the reciprocal. Using the expansion $(\alpha-x)^{-1}=\alpha^{-1}(1-x/\alpha)^{-1}= \alpha^{-1}+x\alpha^{-2}+\cdots$ we see that for $n'\in I$ we have \[ \frac{\dd k^*(n')}{\dd n'} \le \frac{1}{\alpha} + (1+2\delta)\frac{\log\nu}{\alpha^2(\log n-\log\log n)}. \] For any $m\le m^+$ this estimate applies for all $n'$ in the interval $(n-\alpha m,n)\subset I$, so it follows immediately that \[ k^*(n) - k^*(n-\alpha m) \le m + (1+2\delta)\frac{m\log\nu}{\alpha(\log n-\log\log n)}. \] Hence \begin{eqnarray*} k_0-m - k^*(n-\alpha m) &=& k_0-k^*(n)-m + (k^*(n)-k^*(n-\alpha m)) \\ &\le& -d - m + (k^*(n)-k^*(n-\alpha m)) \\ &\le& -d + (1+2\delta)\frac{m\log\nu}{\alpha(\log n-\log\log n)} \\ &\le& -d + (1+4\delta)\frac{\mu\log\nu}{\alpha(\log n-\log\log n)} \\ &\le& -\frac{\delta\mu\log\nu}{\alpha(\log n-\log\log n)} \\ &\le& -\frac{C \mu}{\alpha\log n}, \end{eqnarray*} where in the last three steps we used the fact that $m\le m^+=(1+\delta)\mu$, then the definition of $d$, and finally the definition of $\delta$. Hence, from \eqref{L0k*}, if $n$ is large enough \[ L_0(n-\alpha m,k_0-m,\alpha-1) \le -0.99c_0\log^2 n \frac{C \mu}{\alpha\log n} \le -0.98C \mu\le -2\mu, \] recalling that $C\ge 3$. Putting the pieces together, we have \[ \log\E[C_m] \le \mu -2\mu + O(\log^4 n) \sim -\mu, \] recalling that $\mu\ge \log^5 n$ by assumption. Thus, if $n$ is large enough, $\E[C_m]\le 1/n$ with plenty of room to spare, completing the proof of \eqref{ECM} and thus of the lemma. \end{proof} \section{Appendix: intuition behind conjectures}\label{s:intuition} In this section we motivate the more refined conjectures in \S\ref{ss_furtherconjs}. There are two basic starting points, both described previously, so we only recall them briefly. Firstly, the very first guess at the chromatic number is from the `expectation threshold', the least $k$ such that the expected number of partitions into $k$ independent sets is larger than $1$. In calculating this, since there are rather few profiles (a list specifying how many independent sets have each possible size) to consider, one can consider only the optimal profile. This intuition fails immediately when we look at independent sets of size $\alpha$: the naive `optimal profile' is `unachievable', because it would like us to use $\Theta(n/\log n)$ independent sets of size $\alpha$, but the actual number $X_\alpha$ will be close to $\mu_\alpha$ which (for most $n$) will be much smaller than this. So the first approximation is to consider $\alpha$-sets separately, expecting (since the naive optimum is to use many more than there are) that we will use as many as we can, and then considering the expectation threshold for colourings without $\alpha$-sets. This same `unachievability' phenomenon can also arise with $(\alpha-1)$-sets; again, the optimal profile would like to use $\Theta(n/\log n)$ of them. There are certainly enough present, but not necessarily enough \emph{disjoint} ones. Numerical calculations carried out by the first author suggest that this is an issue for $\mu_{\alpha-1}(n)$ up to around $n^{1+x_0}$ for some small positive constant $x_0$. As in \S\ref{ss_furtherconjs}, to avoid a discontinuity when $\alpha$ changes, from now on we work in terms of $a=a(n)$, chosen so that $\mu_a(n)$ is between $n^{1/2+\delta}$ and $n^{3/2-\delta}$ for some positive $\delta$. (We only consider the `good' $n$, for which such an $a$ exists.) Then $a=\alpha$ or $\alpha-1$. In the latter case $\mu_{\alpha}(n)$ is at most $n^{1/2-\delta}$. For us, the independent sets of size $\alpha=a+1$ can be ignored in this case: there may be enough of them to affect the chromatic number significantly, but the standard deviation of $X_\alpha$ is at most around $n^{1/4}$, which is smaller than any of our predictions for $g(n)$. Heuristically, we include all $\alpha$-sets in our colouring, but do not need to consider them any further. As outlined above, our main heuristic (we discuss another below) is as follows: to colour we choose as large as possible a collection $\cC$ of disjoint independent sets of size $a$. Then we assume that the rest of the graph can be coloured with colour classes of size $a-1$ as predicted by the relevant expectation threshold. Let us write $m$ for $X_a$, the number of independent sets of size $a$, which will typically be $\mu_a$ plus or minus order $\sqrt{\mu_a}$, recalling that the distribution of $X_a$ is approximately Poisson, and hence asymptotically Gaussian when $\mu_a\to\infty$. We write $t$ for the size of $\cC$. Somewhat informally, we need to understand: (I) roughly how big $t$ is, and (hence) roughly how much $t$ varies as $m$ varies, and (II) how much a given change in $t$ affects the $(a-1)$-bounded chromatic number of the remaining graph $G_{n',p}$, where $n'=n-ta$. Let us rescale by writing \[ m=\frac{2xn}{a^2} \text{\quad and\quad} t=\frac{2yn}{a^2}. \] Rather than consider the actual distribution of independent sets of size $a$, we work heuristically in the random hypergraph model $H_a(m)$, or rather the essentially equivalent variant where the $m$ hyperedges are chosen independently and uniformly from all $a$-sets. Since two $a$-sets intersect with probability $\sim a^2/n$, we see that on average one $a$-set intersects $\sim 2x$ others. Case 1: $x=o(1)$, i.e., $m=o(n/\log^2 n)$. Then almost all $a$-sets intersect no others, so we have $t\sim m$. Moreover, if we add an extra $a$-set, it is very likely to be disjoint from the current maximum matching, so (somewhat informally) \[ \frac{\dy}{\dx} = \frac{\dt}{\dm} \sim 1. \] Case 2: $x=\Theta(1)$. Here it is hard to say anything very precise, but it is nevertheless clear that $t=\Theta(m)$. (We still have a constant fraction of $a$-sets that intersect no others.) Certainly we expect that for some\footnote{One can probably describe $h$ in terms of the size of the largest independent set in a suitable random graph $G_{n,2x/n}$, but it is not clear that this adds much. In any case, we believe we understand the asymptotic behaviour as $x\to 0$ or $x\to\infty$ from cases 1 and 3.} well-behaved increasing function $h:(0,\infty)\to(0,\infty)$ we have $y\sim h(x)$, and hence \[ \frac{\dy}{\dx} \sim h'(x) = \Theta(1). \] Case 3: $x\rightarrow \infty$. This case is more difficult, but for our heuristic we assume that the maximum matching is at least approximately given by the first moment threshold in the random hypergraph, i.e., by solving \[ \binom{m}{t} \frac{(n)_{at}}{(n)_a^t} \approx 1, \] where $(n)_k$ is the falling factorial $n!/(n-k)!$, and the ratio above is the probability that $t$ randomly chosen $a$-sets are disjoint. In turn this gives \begin{equation}\label{eq:xy} \log x - \log y + 1 +o(1) = y(1+O(at/n)) = \Theta(y), \end{equation} and we arrive at \[ y = \Theta(\log x) \text{\quad and\quad} \frac{\dy}{\dx} = \Theta(1/x), \] with the implicit constants being $1+o(1)$ when $x=n^{o(1)}$. Let us now turn to (II), considering how the $(a-1)$-bounded chromatic number of the rest of the graph, which we treat simply as $G_{n',p}$, $n'=n-at$, varies as $t$, and hence $n'$, varies. Heuristically, we assume the actual number of colours needed will be essentially the relevant first moment threshold, or rather the approximation $k^*$ from Lemma~\ref{lem:polylog}. If there are $n'$ vertices left, then for each extra vertex covered by $a$-sets we expect to need $\tfrac{\dd k^*}{\dd n}|_{n=n'}$ fewer colours. We temporarily write $\gamma$ for the \emph{reciprocal} of this quantity. From Lemma~\ref{lem:polylog} and Remark~\ref{rem:noks} (which tells us that we can replace $\alpha(n)$ by $a$ in \eqref{noks}) we have \begin{eqnarray*} \gamma &=& a -1 + \frac{\log\mu_a(n')}{\log n'-\log\log n'} + O(1/\log n') \\ &=& a -1 + \frac{\log\mu_a(n')}{\log n-\log\log n} + O(1/\log n) \end{eqnarray*} since we'll always have $n'=\Theta(n)$. It is convenient to work in terms of $\mu_{a+1}=\Theta(\mu_a \log n/ n)$. Let $\mu' = \mu_{a+1}(n')$, then \begin{eqnarray*} \gamma &=& a + \frac{\log\mu'}{\log n-\log\log n} + O(1/\log n) \\ &=& a+(1+o(1)) \frac{\log\mu'}{\log n}, \end{eqnarray*} since we'll see later that $-\log\mu'$ is at minimum at least $\omega(1)$ (in fact at least order $\log\log n$). This gives \[ \frac{\dd k^*}{\dd n} = \gamma^{-1} = a^{-1} - (1+o(1))\frac{\log \mu'}{a^2\log n}. \] So each extra $a$-set in the matching should save $a$ times this many colours, minus the one used for the set itself, giving `benefit' (per $a$-set used) \[ B \sim \frac{-\log \mu'}{a\log n} \sim \frac{-\log \mu'}{c_0\log^2 n} \] where $c_0=2/\log 2$. Now \[ \mu_{a+1}(n) = \Theta(\mu_a(n)\log n/n) = \Theta(x/\log n). \] In all cases, writing $\approx$ for agreement up to constant factors, \[ \frac{\mu'}{\mu_{a+1}(n)} \sim (n'/n)^{a+1} \approx (n'/n)^a = (1-at/n)^a = (1-2y/a)^a. \] In cases 1 and 2, where $x$ and hence $y$ are $O(1)$, this is $\Theta(1)$ and hence irrelevant. In these cases we thus have $\mu'=\Theta(x/\log n)$, so $-\log \mu'\sim \log\log n + |\log x|$. Thus \[ B \sim \frac{\log\log n+|\log x|}{c_0\log^2 n}. \] In case 3, when $x$ grows but not too quickly, say $x=n^{o(1)}$, then $y=o(\log n)$ and hence, from \eqref{eq:xy}, $y\sim \log x$. Then \[ \frac{\mu'}{\mu_{a+1}(n)} \approx (1-2y/a)^a = \exp(-(2+o(1))\log x), \] so $\mu'$ is roughly $1/(x\log n)$, with asymptotic agreement in the logarithms. In this case we thus obtain \[ B \sim \frac{\log\log n+\log x}{c_0\log^2 n}. \] Finally, if $x$ is at least $n^{\Omega(1)}$, then cruder estimates give $\mu'=n^{-\Omega(1)}$, so $-\log\mu'=\Theta(\log x)$. In this case \[ B = \Theta\left(\frac{\log x}{\log^2n}\right). \] In all cases, multiplying $\sqrt{\mu_a}$, our estimate for how much the number $m$ of independent $a$-sets varies, by $\tfrac{\dy}{\dx}=\tfrac{\dt}{\dm}$, and then by $B$, gives our estimate for $g(n)$, the standard deviation of $\chi(\Gnp)$. \subsection{Complications} In this subsection we discuss a number of issues that arise when attempting to understand the behaviour of $\chi(\Gnp)$ even more precisely. First, we should note that in any attempt at \emph{proving} Conjecture~\ref{conj:4cases}, there are major problems with the heuristic above. The key one is that, having removed some collection of independent $a$-sets, the graph that remains certainly does not have the same distribution as $G_{n',p}$ for appropriate $n'$. But even at the intuitive level, there are additional complications. For one thing, the alert reader may have noticed that our heuristic above does not make sense in case 3 when $x$ is too large, in particular when $\mu_a\ge n^{1+x_0}$, the point up to which the naive optimum profile wants us to use more disjoint $a$-sets than can be found. Here we justify our prediction rather by the heuristic in \S\ref{ss:zigzag}. With $t$ fixed, then as $m=X_a$ varies, the number of ways of choosing $t$ (disjoint) $a$-sets varies, and this translates into variation in the chromatic number. Fortunately, for $\mu_a=n^{1+\Theta(1)}$ the two predictions agree within a constant factor, so we do not need to resolve exactly how they interact. This same effect arises in other cases, however. Suppose we have a strategy for partially colouring with $a$-sets where we use a slightly smaller than maximum matching, so there are $N\gg 1$ choices for this matching. Then we might expect to find a colouring if the expected number of $(a-1)$-bounded colourings of the remaining $n'$ vertices is roughly $1/N$. As noted earlier, from Corollary~\ref{dnk}, for given $n'$ we should expect the extra $N$ choices to lead to a reduction in $k^*$ of around $\log N/(c_0\log^2 n')=\Theta(\log N/\log^2 n)$. Considering the simpler case in which almost all $a$-sets are disjoint, we have $N\approx \binom{m}{t}$, so there is a large increase in the number of choices for leaving out the first few $a$-sets. Our calculations suggest that in this range we will leave out order $\Theta(\mu_a^2\log n/n)$ $a$-sets from a maximum matching. This will affect the chromatic number significantly, but we do not expect it to lead to a significant change in the variance of the chromatic number. A further issue is that in our case $x\to\infty$, the first moment threshold is not a terribly good estimate of the size of a maximum matching of $a$-sets. In the case where $x$ does not grow too quickly, a heuristic explanation is the following. Since two $a$-sets intersect with probability $\pi_0\sim a^2/n$, we expect $t$ $a$-sets to be disjoint with probability around $\exp(-\pi_0\binom{t}{2})$. However, there is some variability in the number $M$ of overlapping pairs of $a$-sets. This quantity, which is of order $M_0=\binom{m}{2}a^2/n$, varies by around $\Delta=\sqrt{M_0}$. If we condition on this number, then our new heuristic for the probability $t$ $a$-sets are disjoint is $\exp(-\pi\binom{t}{2})$ where $\pi=M/\binom{m}{2}=\pi_0(1+\Delta/M_0)$. This variation may well be significant, and it leads to a situation where the overall expectation of the number of $t$-matchings (collections of $t$ disjoint $a$-sets) is dominated by the contribution from the case where $M$ is atypically small. Hence the first moment will not be an accurate guide to the existence of a $t$-matching. We do not explore this further here since it does not seem to affect $g(n)$. However, this, and more complicated such effects, would (at least in some cases) alter $f(n)$ by a significant amount. Thus the problem of predicting, let alone proving, a `full result' $(\chi(\Gnp)-f(n))/\sqrt{g(n)} \dto N(0,1)$ seems extremely difficult. \bibliographystyle{plainnat}

https://arxiv.org/abs/0905.3765

Sequences of density zeta(k) - 1

At a social gathering of mathematicians, Herb Wilf noted that the numbers $\zeta(k) - 1$ sum to 1, and challenged the assembly to interpret the sequence as probabilities in some interesting number theoretic context. This short note provides one such interpretation.

\section{INTRODUCTION} At a recent post-seminar gathering, Herb Wilf casually mentioned to those of us assembled the fact that the quantities $\zeta(k)-1$ sum to 1, i.e., $$ \sum_{k=2}^{\infty} \left( \left( \sum_{i=1}^{\infty} \frac{1}{i^k} \right) - 1 \right) = 1 \, \text{.}$$ He then declared that when a sequence of nonzero positive numbers sums to 1, the entries of the sequence should well be interpretable as the probabilities of \emph{something}, and asked what it might be in this case. The challenge here is to provide events as interesting as the numbers $\zeta(k)-1$ themselves: rich with internal relations and easy to describe. One criterion for a really good answer to this challenge would be that the response supplies us with an intuitive feel for the relations among the quantities $\zeta(k)-1$. The context of the discussion was number-theoretic, so after some consideration I formalized the challenge thus: \begin{Problem} Partition $\mathbb{N}$ into sets $\{ {\cal{A}}_2, {\cal{A}}_3, \dots \}$ such that the asymptotic densities $\lim_{n \rightarrow \infty} \frac{1}{n} \# \left( {\cal{A}}_k \cap \{1,2,\dots , n \} \right) = \zeta(k)-1$. \end{Problem} This note provides one such partition. There are, it will be seen, others, but {\ae}sthetics guides the choice presented here: the partition can be described in terms of components that build on each other, and we can produce an algorithm that can quickly determine the ${\cal{A}}_k$ to which any number belongs. \section{THE CLASSES ${\cal{B}}_k$} The key both to constructing the classes ${\cal{A}}_k$ and to proving that they have the required densities is to break down the $\zeta(k)-1$ thus: \begin{equation}\label{ProbMatrix} \begin{matrix} \zeta(2) - 1 &= \frac{1}{2^2} &+ \frac{1}{3^2} &+ \frac{1}{4^2} &+ \frac{1}{5^2} &+ \dots \\ \zeta(3) - 1 &= \frac{1}{2^3} &+ \frac{1}{3^3} &+ \frac{1}{4^3} &+ \frac{1}{5^3} &+ \dots \\ \zeta(4) - 1 &= \frac{1}{2^4} &+ \frac{1}{3^4} &+ \frac{1}{4^4} &+ \frac{1}{5^4} &+ \dots \\ \zeta(5) - 1 &= \frac{1}{2^5} &+ \frac{1}{3^5} &+ \frac{1}{4^5} &+ \frac{1}{5^5} &+ \dots \\ \vdots & \vdots & \vdots & \vdots & \vdots & \\ 1 &= \frac{1}{2} &+ \frac{1}{6} &+ \frac{1}{12} & + \frac{1}{20} &+ \dots \end{matrix} \end{equation} \noindent where the entries of the last line are the summations of the (infinite) columns above. The strategy is now to construct classes each of which have asymptotic densities given by the individual entries of the table. As those entries are all unit fractions, our task is much simplified: ask for a set of integers of some peculiar irrational density and intuition would not necessarily leap to service; but ask for one-sixth of the integers, and any residue class mod 6 will do. So we begin in this section with that last row, by producing classes ${\cal{B}}_j$ of densities $\frac{1}{2}, \frac{1}{6}, \frac{1}{12}, \dots$, representing each column. In the next section we break each ${\cal{B}}_j$ apart into sections ${\cal{B}}_{j,k}$ of densities $\frac{1}{j^k}$, $2 \leq j,k < \infty $, to produce the individual entries of the table. The finishing stroke is then to collect across rows to form the ${\cal{A}}_k$. The constructions we use are often explained nicely by recasting them in terms of the \emph{factorial-base} representation of a number, its expansion as $x = a_1 \cdot 1! + a_2 \cdot 2! + \dots + a_n n!$, $0 \leq a_i \leq i$). This expansion is unique, as the student unfamiliar with this base system is encouraged to try proving on their own. For more on such expansions, see \cite{Barwell}. The class ${\cal{B}}_2$ must have asymptotic density $\frac{1}{2}$. Let us select it to consist of those $x \equiv 1$ mod 2. Modulo 6, the classes 1, 3, and 5 have been assigned, so for ${\cal{B}}_3$ we select the residue class 2 mod 6. Modulo 24, we require two residue classes in order that ${\cal{B}}_4$ should have density $\frac{1}{12}$. We select the first two available: those $x \equiv 4$ or 6 mod 24. Modulo 120, we require six residue classes in order that ${\cal{B}}_5$ should have density $\frac{1}{20}$, and the first six available are 10, 12, 16, 18, 22, and 24. Notice that we cannot use modulus 20 for ${\cal{B}}_5$: every residue class mod 20 has already had some subprogressions assigned. Instead, we will use the moduli $j!$. Since for $m<j$, residue classes modulo $m!$ are the same as $\frac{j!}{m!}$ equally spaced residue classes modulo $j!$, all previous assignments can be viewed as multiple assignments modulo $j!$. For each ${\cal{B}}_j$, we select the first $(j-2)!$ residue classes that have not yet been assigned. What are these in general? \begin{Lemma}\label{BAssignment} The assigned residue classes for ${\cal{B}}_m$ are $$\{ x \equiv \, 0! + \sum_{j=1}^{m-2} a_j \cdot j! \, \text{mod} \, m! \, , 1 \leq a_j \leq j \} \, \text{.}$$ \end{Lemma} \begin{proof} In the examples above, by the time we have selected component residue classes for ${\cal{B}}_j$ we have assigned membership in some ${\cal{B}}_i$ to the natural numbers $1, 2, \dots , (j-1)!$, as part of residue classes either modulo $j!$ or modulo divisors thereof (previous factorials). Let us proceed by induction, and suppose that this has been the case through ${\cal{B}}_{m-1}$: for each $b \leq m-1$, the numbers $1, 2, \dots (b-1)!$ have been assigned to some ${\cal{B}}_j$, $j \leq b$, as part of residue classes modulo $j!$. Given this induction hypothesis, what is the first residue class modulo $m!$ available for assignment to ${\cal{B}}_m$? None of 1 through $(m-2)!$, to begin with: each of those classes were assigned to ${\cal{B}}_{m-1}$ or some previous class. We do know that none of the next $(m-2)!$ residue classes mod $m!$, $(m-2)!+1$ through $(m-2)!+(m-2)!$, are assigned to ${\cal{B}}_{m-1}$. Examine this segment. Since $(m-2)! \equiv 0$ mod $(m-2)!$, our assumption tells us that the numbers $(m-2)!+1$ through $(m-2)!+(m-3)!$ must have been assigned to ${\cal{B}}_{m-2}$ or earlier classes, as residues modulo $(m-2)!$ or divisors thereof. However, none of $(m-2)!+(m-3)!+1$ through $(m-2)!+(m-3)!+(m-3)!$ were assigned to ${\cal{B}}_{m-2}$, and this segment is wholly contained within those numbers we already knew were not assigned to ${\cal{B}}_{m-1}$. At each stage of the analysis, we have an unbroken segment of assigned classes for the numbers $1, 2, \dots, (m-2)! + (m-3)! + \dots + (m-i)!$, and we know that the next $(m-i)!$ numbers are not assigned to ${\cal{B}}_i$ through ${\cal{B}}_{m-1}$. At the last step, we find that the last number assigned is $(m-2)! + (m-3)! + \dots + 2! + 1!$, an odd number in class ${\cal{B}}_2$. The very next number is in the segment we know has not been assigned to ${\cal{B}}_2$ or any previous class, so it is available. Thus the first available residue class modulo $m!$ to assign to ${\cal{B}}_m$ is $(m-2)! + (m-3)! + \dots + 2! + 1! + 0!$, a construction for which there are unfortunately multiple labels and notations in current usage. The relevant sequence ${1, 2, 4, 10, 34, 154, 874, \dots }$ is Sloane's A003422 (\cite{Sloane}) missing a leading zero term, and is there called the \emph{left factorial} and denoted $!n = 0! + \dots +(n-1)!$. We will use this notation, so that the first unassigned class modulo $m!$ is $!(m-1)$. Now let us determine the remainder of the assigned classes mod $m!$. At the next-to-last step of our analysis previously, assuming $m \geq 4$ so that such a step exists, we determined that $!(m-1) - 2$ must have been an assigned class modulo $3!$, that is, $!(m-1)-2 \equiv 2$ mod 6. Then $!(m-1)$ is not such a class and neither is $!(m-1)+2$: these are the two even classes mod 6 that were not assigned to ${\cal{B}}_3$. Nor can they be assigned to ${\cal{B}}_i$ for $i \geq 3$ as part of residue classes modulo multiples of $3!$. Thus, we can select these two as our first two residue classes modulo $m!$ for ${\cal{B}}_m$. If $m \geq 5$, we can ascend backward another step: there we find that $!(m-1)-4$ and $!(m-1)-6$ had to be the assigned residue classes mod $4!$. We already knew $!(m-1)$ and $!(m-1)+2$ were not in those classes, and here we find that they can be joined by $!(m-1)+6$ or $+12$, and $!(m-1)+2+6$ or $+12$. These are the six residue classes modulo 24 that were left after we assigned two of them for ${\cal{B}}_4$. At the $i$-th step of the ascent, we find that to all the entries we previously picked, we can add as many as $i-1$ multiples of $i!$. Thinking about it in the factorial-base representation, using $a_1 a_2 a_3 \dots$ to denote $a_1 \cdot 1! + a_2 \cdot 2! + a_3 \cdot 3! + \dots$, the assigned classes will be $02111\dots + 0 b_2 b_3 b_4 b_5 \dots$, with $0 \leq b_i \leq i-1$. And these are exactly the numbers claimed in the Lemma. \end{proof} \subsection{A Fast Membership Algorithm} Having assigned the whole numbers to their various ${\cal{B}}_m$, we would like a means of determining which ${\cal{B}}_m$ a given $x$ belongs to without constructing every ${\cal{B}}_m$ until we assign $x$. As it turns out, systematically selecting the first available classes at each step helps us write a computationally simple algorithm to do so. Write $x$ in the factorial-base representation, $x = x_1 \cdot 1! + x_2 \cdot 2! + x_3 \cdot 3! + \dots + x_j \cdot j!$, $0 \leq x_i \leq i$. We know that if the least positive residue of $x$ modulo any $n!$ is less than $(n-1)!$, $x$ must have been assigned to one of the ${\cal{B}}_{j \leq n}$, because from our proof above, the assigned residue classes for the ${\cal{B}}_{j \leq n}$ were an unbroken string from 1 to $(n-1)!$. A "carry" in factorial-base addition occurs when the sum of the coefficients on the two summands on $i!$ is at least $i+1$, since $(i+1)i! = 1 \cdot (i+1)!$. But the Lemma tells us that the residues assigned to ${\cal{B}}_m$ are of the form $02111\dots + 0 b_2 b_3 b_4 b_5 \dots$, with $0 \leq b_i \leq i-1$. So in order for a 0 to ever occur at the $i!$ place of the factorial-base expansion of a residue assigned to ${\cal{B}}_m$, when $1 < i \leq m-1$, we need a very specific summand. We need $b_2 = 1$, so that we carry from $2!$, leaving a 0. We then need $b_3 = 2$, so that the sum of coefficients on $3!$ is 1 from $!(m-1)$, plus 2 from $b_3$, plus a carried 1, making 4, so we carry 1 and leave a 0 again. This string has to continue up to the $i!$ term where we desire a 0. Thus we have one of two cases that will diagnose where $!(m-1)$ ends for a given $x$. We have that the least positive residue of $x$ modulo $m!$ is $\leq (m-1)!$. Possibly $x \equiv (m-1)!$ modulo $m!$ exactly, so the string of $x_i$ starts with a possibly empty string of 0s that terminates with a 1 in position $x_{m-1}$: we added the largest possible value for every $b_i$, and carried at every step of the addition. If we did not carry, then there is some smallest $m$ such that $x_{m-1} = 0$ but no previous $x_i = 0$, except for possibly an initial string that does \emph{not} terminate with a 1. Write out $x$ in the factorial-base representation, check for the first behavior, and if it doesn't happen find the first 0. The index where whichever of these occurs, occurs, is $m-1$. \section{THE CLASSES ${\cal{A}}_k$} Now that we have identified the classes ${\cal{B}}_m$, there remain the tasks of breaking them up into subclasses to give us the individual entries of Table \ref{ProbMatrix}, and collecting the corresponding parts "horizontally" to form the ${\cal{A}}_k$. Each of the columns' individual entries decrease in geometric progression. The class ${\cal{B}}_2$ has density $\frac{1}{2} = \frac{1}{2} \left( \frac{1}{2} + \frac{1}{4} + \frac{1}{8} + \dots \right) $. The class ${\cal{B}}_3$ has density $\frac{1}{6} = \frac{1}{6} \left( \frac{2}{3} + \frac{2}{9} + \frac{2}{27} + \dots \right) $. The class ${\cal{B}}_4$ has density $\frac{1}{12} = \frac{1}{12} \left( \frac{3}{4} + \frac{3}{16} + \dots \right) $, etc. These geometric series give us the subclasses we need. We assign the earliest fraction $\frac{m-1}{m}$ of ${\cal{B}}_m$ to the first subset, the next fraction $\frac{1}{m} \frac{(m-1)}{m}$ to the next set, and so forth. This can be done by a one-step digit test. For an example from familiar territory, to obtain $\frac{9}{10}$ of the integers, take those ending in 1 through 9. Then to obtain $\frac{9}{100}$ of the integers by choosing $\frac{9}{10}$ of the remaining tenth, take those that end in 0 but have one of 1 through 9 in the tens place. Since ${\cal{B}}_m$ is made up of residue classes modulo $m!$, rewrite $x$ as $x = x_0 + m! \cdot \left( b_0 m^0 + b_1 m^1 + b_2 m^2 + \dots \right) $, $0 \leq b_i \leq m-1$. If $n$ is the smallest number such that $b_n \neq m-1$ assign $x$ to the set ${\cal{B}}_{m,n+2}$. In table \ref{ProbMatrix}, $x$ is part of the set with asymptotic density given by the entry in column $m$, row $n+1$, which represents an entry in the sum for ${\cal{A}}_{n+2}$. Thus, for example, we break up ${\cal{B}}_2$ as follows: ${\cal{B}}_{2,2} = \{ x \equiv 1 \, \text{mod} \, 4(=2 \cdot 2!) \}$, with density $\frac{1}{4}$; ${\cal{B}}_{2,3} = \{ x \equiv 3 \, \text{mod} \, 8(=4 \cdot 2!) \}$, with density $\frac{1}{8}$; ${\cal{B}}_{2,4} = \{ x \equiv 7 \, \text{mod} \, 16(=4 \cdot 2!) \}$, etc. We break up ${\cal{B}}_3$ into fractions of size $\frac{2}{3}, \frac{2}{9}$, etc.: ${\cal{B}}_{3,2} = \{ x \equiv 2 \, \text{or} \, 8 \, \text{mod} \, 18(=3 \cdot 3!) \}$, ${\cal{B}}_{3,3} = \{ x \equiv 14 \, \text{or} \, 32 \, \text{mod} \, 54(=9 \cdot 3!) \}$, etc. We now sum up by collecting these subsets across rows: ${\cal{A}}_k = {\bigcup \atop m} {\cal{B}}_{m,k}$. Since the ${\cal{B}}_{m,k}$ are disjoint and the series of partial sums of their densities converges absolutely to $\zeta(k)-1$, ${\cal{A}}_k$ has exactly the required asymptotic density. The algorithm to determine which class ${\cal{A}}_k$ a given $x$ belongs to is thus: \begin{Algorithm} Write $x = a_1 \cdot 1! + a_2 \cdot 2! + a_3 \cdot 3! + \dots $, $0 \leq a_i \leq i$. Let $j$ be the smallest index such that $a_j \neq 0$. If $a_j=1$, then $m=j+1$. If $a_j \neq 1$, then $m$ is the smallest number such that $a_i \neq 0$ for $j \leq i < m-1$, and $a_{m-1} = 0$. Then $x \in {\cal{B}}_m$ . Rewrite $x$ as $x = x_0 + m! \cdot (b_0 m^0 + b_1 m^1 + b_2 m^2+ \dots )$, $0 \leq b_i < m$, $0 < x_0 < m!$. Let $n$ be the smallest index such that $b_n \neq m-1$. Then $k = n+2$, and $x \in {\cal{A}}_k$. \end{Algorithm} \section{DATA AND SPECULATION} The first few of the sets ${\cal{B}}_k$ and ${\cal{A}}_k$ are listed below. ${\cal{B}}_2 = \{ 1, 3, 5, 7, 9, 11, 13, 15, \dots \}$. ${\cal{B}}_3 = \{ 2, 8, 14, 20, 26, 32, \dots \}$. ${\cal{B}}_4 = \{ 4, 6, \quad 28, 30, \quad 52, 54, \quad 76, 78, \dots \}$. ${\cal{B}}_5 = \{ 10, 12, 16, 18, 22, 24, \qquad 130, 132, 136, 138, 142, 144, \dots \}$. ${\cal{B}}_6 = \{ 34, 36, 40, 42, 46, 48, 58, 60, 64, 66, 70, 72, \dots, 120, \dots \} $. ${\cal{A}}_2 = \{ 1,2,4,5,6,8,9,10,12,13,16,17,18,20,21,22,24,25,26, \dots \}$. ${\cal{A}}_3 = \{ 3,11,14,19,27,32,35,43,51,59,67,68,75,76,78,83,86, \dots \}$. ${\cal{A}}_4 = \{ 7,23,39,50,55,71,87,103,104,119,135,151,167,183,199, \dots \}$. ${\cal{A}}_5 = \{ 15,47,79,111,143,158,175,207,239,271,303,320,335, \dots \}$. More terms, for the series up to ${\cal{A}}_{10}$, have been submitted to Sloane's database \cite{Sloane} and should be available by the time of publication of this note. In terms of their component residue classes, the sets are: ${\cal{B}}_2 = \{ x \equiv 1 \, \text{mod} \, 2 \}$. ${\cal{B}}_3 = \{ x \equiv 2 \, \text{mod} \, 6 \}$. ${\cal{B}}_4 = \{ x \equiv 4, 6 \, \text{mod} \, 24 \}$. ${\cal{B}}_5 = \{ x \equiv 10, 12, 16, 18, 22, 24 \, \text{mod} \, 120 \}$. ${\cal{B}}_6 = \{ x \equiv 34, 36, 40, 42, 46, 48, 58, 60, 64, 66, 70, 72, \dots, 120 \, \text{mod} \, 720 \} $. ${\cal{A}}_2 = \{ x \equiv 1 \, \text{mod} \, 4; \, 2,8 \, \text{mod} \, 18; \, 4,6,28,30,52,54 \, \text{mod} \, 96 \dots \}$. ${\cal{A}}_3 = \{ x \equiv 3 \, \text{mod} \, 8; \, 14,32 \, \text{mod} \, 54; \, 76,78,172,174,268,270 \, \text{mod} \, 384 \dots \}$. ${\cal{A}}_4 = \{ x \equiv 7 \, \text{mod} \, 16; \, 50, 104 \, \text{mod} \, 162; \, 364, 366, \dots ,1134 \, \text{mod} \, 1536 \dots \}$. Some variations of this construction could be explored. Building ${\cal{B}}_j$, we had to make a series of choices. We chose the odd numbers to form ${\cal{B}}_2$, but the even numbers would have worked just as well. We chose the residue class $x \equiv 2$ mod 6 for the class ${\cal{B}}_3$, but could just as easily have chosen $x \equiv 4$ or 6. Choosing the first available residue class for each factorial modulus gave us classes ${\cal{B}}_m$ that could be determined with a short algorithm, providing a tidy answer to the original problem. However, other choices lead to answers with different features. For example, suppose that at each step we choose the \emph{last} available residue class modulo $j!$ to construct ${\cal{B}}_j$. Choose those $x \equiv 2$ mod 2 for ${\cal{B}}_2$. Of the three available odd residue classes mod 6, choose $x \equiv 5$ mod 6 for ${\cal{B}}_3$. Use the classes $x \equiv 19, 21$ mod 24 for ${\cal{B}}_4$. With such a decision procedure, some numbers are never assigned to any ${\cal{B}}_j$ at all! The "missed set" is ${1, 3, 7, 9, 13, 15, 25, 27, 31, 33, \dots = 1 + \sum a_i i!, 0 \leq a_i < i}$. All of the ${\cal{B}}_j$ will still have the same densities, summing to 1, so the missed set is "small" in that it is of asymptotic density 0. In fact, given any arithmetic progression $X$ mod $Y$, some subprogression will be assigned to a ${\cal{B}}_j$. (Show it!) On the other hand, at any given point the missed set may be rather large for some purposes: $1/n$ of the numbers smaller than $n!$ are permanently unassigned. These two assignment procedures are in some sense on opposite poles of an entire ensemble of possible procedures. An interested reader might burrow a layer deeper than we have: assign some straightforward process for describing and choosing assignment procedures to construct each ${\cal{B}}_j$ from residue classes mod $j!$ and examine the resulting ensemble of all possible constructions. Does it possess any striking structural features? An early stab at this problem involved the powerfree numbers. Squarefree numbers are those whose prime factors are all distinct: they have density $\frac{1}{\zeta(2)}$ in the whole numbers. Cubefree numbers have factorizations in which no prime factor is repeated more than twice, so squarefree numbers are also cubefree. Cubefree numbers are of density $\frac{1}{\zeta(3)}$, and those which are cubefree but not squarefree are of density $\frac{1}{\zeta(3)} - \frac{1}{\zeta(2)}$. Those that are 4th-powerfree but not cubefree are of density $\frac{1}{\zeta(4)} - \frac{1}{\zeta(3)}$, and so on. Every whole number is $(k+1)$st-powerfree but not $k$th-powerfree for some $k$, the largest exponent in its prime factorization. These densities seemed to suggest the possibility of defining a simple probability distribution on $\mathbb{N}$ that assigned a total probability of $\zeta(k) - 1$ to the event that a random integer variable would be $(k+1)$st-powerfree but not $k$th-powerfree. It is easy to provide such a distribution by brute force -- say, giving the $n$th number which is $(k+1)$st-powerfree but not $k$th-powerfree a probability of $(\zeta(k) - 1) \cdot \frac{1}{2^n}$.\footnote{Typo in this line corrected in preprint at kind communication from Michael Lugo.} But this does not seem to be a particularly illuminating illustration of any relations between the values $\zeta(k)$; indeed, such a construction works with any partitioning of the integers, and any set of probabilities instead of $\zeta(k)-1$. A distribution based on the factorization of $x$ would seem much more natural; can a simple one be produced? In closing, I would like to mention a personal recollection. In general, the density of ${\cal{B}}_j$ is $\frac{1}{j(j-1)}$. Seeing it written on the wall of my office, I was reminded of the first place I ever encountered Leibniz' summation of the triangular series: William Dunham's delightful popular-mathematics text, \emph{Journey Through Genius} \cite{JTG}. I read this book in high school, and it motivated in considerable part my decision to pursue mathematics in college. The present note gives me an opportunity to thank Mr. Dunham sincerely for the service.

https://arxiv.org/abs/2111.05425

Disjoint edges in geometric graphs

A geometric graph is a graph drawn in the plane so that its vertices and edges are represented by points in general position and straight line segments, respectively. A vertex of a geometric graph is called pointed if it lies outside of the convex hull of its neighbours. We show that for a geometric graph with $n$ vertices and $e$ edges there are at least $\frac{n}{2}\binom{2e/n}{3}$ pairs of disjoint edges provided that $2e\geq n$ and all the vertices of the graph are pointed. Besides, we prove that if any edge of a geometric graph with $n$ vertices is disjoint from at most $ m $ edges, then the number of edges of this graph does not exceed $n(\sqrt{1+8m}+3)/4$ provided that $n$ is sufficiently large.These two results are tight for an infinite family of graphs.

\section{Introduction} A \textit{geometric graph} \(G\) is a graph drawn in the plane by (possibly crossing) straight line segments, that is, its vertex set \(V(G)\) is a set of points in general position in the plane and its edge set \(E(G)\) is the set of straight line segments with endpoints belonging to \(V(G)\). One of the classical problems on geometric graphs is a question raised by Avital and Hanani~\cite{avital1966graphs}, Kupitz~\cite{Kupitz79}, Erd\H os and Perles. For positive integers \(k\) and \(n\), determine the smallest \(e_k(n)\) such that any geometric graph with \(n\) vertices and \(m>e_k(n)\) edges contains \(k+1\) pairwise disjoint edges. By results of Hopf and Pannwitz~\cite{hopf1934} and Erd\H os~\cite{Erdos46}, we know that \(e_1(n)=n\). The upper bound for \(e_2(n)\) was studied in papers of Alon and Erd\H os~\cite{AlonErdos89}, Goddard, Katchalski, and Kleitman~\cite{GoddardKatchalskiKleitman96}, M\'esz\'aros~\cite{meszaros98}. The current best upper bound, \(e_2(n)\leq \lceil 5n/2\rceil\), was proved by \v{C}ern\'{y}~\cite{Cerny05}. This bound is tight up to additive constant: Perles found an example showing that \(e_2(n)\geq \lfloor 5n/2 \rfloor -3\). Also, in~\cite{GoddardKatchalskiKleitman96} it was shown that \({7n/2-6\leq e_3(n)\leq 10n}\). For \(k\leq n/2\), Kupitz~\cite{Kupitz79} proved the lower bound \(e_k(n)\geq kn\), and later T\'oth and Valtr~\cite{TothValtr99} improved it: \({e_k(n)\geq 3(k-1)n/2-2k^2}\). Using Dilworth's theorem, Pach and T\"or\H{o}csik~\cite{PachTorocsik94} found a beautiful proof of the upper bound~\(e_k(n)\leq k^4 n\). Later, this bound was refined in~\cite{TothValtr99}, and the current best upper bound, \(e_k(n)\leq 256k^2n\), belongs to T\'oth~\cite{Toth00}; see also~Theorem~1.11 in~\cite{felsner2012geometric}. Another interesting result about disjoint edges of a \textit{convex graph} is due to Kupitz~\cite{kupitz1984pairs}. Recall that \textit{a convex graph} is a geometric graph whose vertices are in convex position. He proved that a convex graph on \(n\) vertices has no \(k+1\) pairwise disjoint edge, then its number of edges does not exceed \(kn\) provided \(n\geq 2k+1\). For further reading, we refer the interested readers to the survey of Pach~\cite{pach2013beginnings} on geometric and topological graphs. In the current paper, we study two questions about the local and global number of pairs of disjoint edges. For a geometric graph \( G \), let \(DJ(G)\) be the set of pairs of disjoint edges of geometric graphs. For an edge \(uv\in E(G)\), let \(DJ(uv)\) be the set of edges in \(G\) disjoint from \(uv\). Clearly, we have the equality \[ |DJ(G)| = \frac12 \sum_{uv\in E(G)} |DJ(uv)|. \] These are the main problems of the paper. \begin{problem} For integers \(n>0\) and \(m\geq0\), determine the greatest number \(e(n,m)\) such that a geometric graph \(G\) on \(n\) vertices has at most \(e(n,m)\) edges provided that \(|DJ(uv)|\leq m\) for any \(uv\in E(G)\). \end{problem} \begin{problem} For positive integers \(n\) and \(e\), determine the smallest number \(dj(n,e)\) such that for any geometric graph \(G \) with \(n\) vertices and \(e\) edges, we have \(|DJ(G)|\geq dj(n,e)\). \end{problem} In the current paper, we solve these problems for \textit{locally convex graphs}. To define these graphs, recall that the \textit{neighborhood} \(N(v)\) of a vertex \(v\in V(G)\) is the set of vertices adjacent to \( v \). A vertex is called \textit{convex} if it lies outside of the convex hull of its neighborhood. A locally convex graph is a geometric graph such that any of its vertices is convex. Clearly, any convex graph is locally convex as well. The goal of this paper is to prove the following two theorems. \begin{theorem}\label{theorem:main1} Let \(m\) be a non-negative integer and \(G\) be a geometric graph such that \( |DJ(uv)|\leq m\) for any edge \( uv\in E(G) \). Then \[ |E(G)|\leq \max\left(|V(G)|\left(\sqrt{1+8m}+3\right)/4, |V(G)| + 3m-1\right). \] \end{theorem} \begin{theorem}\label{theorem:main2} For a locally convex graph \( G \) with \( 2|E(G)|\geq |V(G)| \), we have \[ |DJ(G)| \geq \frac{|V(G)|}{2} \cdot \binom{d(G)}{3}, \] where \(d(G)=2|E(G)|/|V(G)|\) is the average degree of \(G\). \end{theorem} Note that Theorem~\ref{theorem:main1} is a strengthening of a result mentioned above. \begin{theorem}[Hopf and Pannwitz~\cite{hopf1934}, Erd\H{o}s~\cite{Erdos46}]\label{erdos} If every edge of a geometric graph \( G \) intersects all other edges of \(G\), then \(|E(G)|\leq |V(G)|\). \end{theorem} The rest of the paper is organized as follows. In Section~\ref{section:preliminaries}, we prove auxiliary lemmas. In Sections~\ref{section:proof1} and~\ref{section:proof2}, we prove Theorems~\ref{theorem:main1} and~\ref{theorem:main2}, respectively. In Section~\ref{section:discussion}, we show that these theorems are tight for an infinite family of graphs. Also, in this section, we discuss open problems related to strengthenings of our main results. \section{Preliminaries} \label{section:preliminaries} Throughout the next three sections, we denote by \( n \) and \(e\) the number of vertices and edges of a geometric graph \(G\), respectively, that is, \(n=|V(G)|\) and \(e=|E(G)|\). In this section, we assume that \( G \) is a locally convex graph such that every vertex has degree at least two. For distinct points \(x,y,z\in \mathbb R^2\) in general position, by the \textit{oriented angle} \( \angle xyz \) we mean the angle in \( (-\pi, \pi) \) starting at the ray \(yx\) and ending at the ray \(yz\), measured counterclockwise. By this definition, we have that \( \angle xyz=-\angle zyx\). For a vertex \( v\in V(G)\), choose \( x,y\in N(v)\) such that \begin{equation} \label{equation:max_angle} \angle xvy =\max\left\{\angle avb: a,b \in N(v)\right\}. \end{equation} Set \( r_v: = x \) and \( \ell_v:=y \). The edges \( v\ell_v \) and \( vr_v \) are called the \textit{leftmost edge} and the \textit{rightmost edge} of \( v \), respectively. For \( v \in V(G) \), let \( DJ_\ell(v) \) be the set of edges incident to one of the vertices of \( N(v) \) and disjoint from the leftmost edge \( v\ell_v \). Analogously, we denote by \(DJ_r(v) \) the set of edges incident to one of the vertices of \( N(v) \) and disjoint from the rightmost edge \( vr_v \). Clearly, \( DJ_\ell(v)\subseteq DJ(v\ell_v)\), \(DJ_r(v)\subseteq DJ(vr_v) \), and thus, we have \(|DJ(v\ell_v)| \geq |DJ_\ell(v)|\) and \( |DJ(vr_v)| \geq |DJ_r(v)|\). For a vertex \(v\in V(G)\), denote by \(\mathcal L(v)\) the set of edges \(\ell_vx\in E(G)\) such that the angle \(\angle x\ell_v v\) is positive. Analogously, denote by \(\mathcal R(v)\) the set of edges \(r_vv\in E(G)\) such that the angle \(\angle xr_vv\) is negative. Remark that if \(\ell_v x\in \mathcal L(v)\) or \(r_vx\in \mathcal R(v)\), then \(x\not\in N(v)\). \begin{lemma} \label{lemma:simple?} For any vertex \( v\in V(G) \), we have \begin{equation} \label{equation: first lemma} | DJ_\ell (v) | + | DJ_r (v) | \geq \sum_{w\in N(v)\setminus\{\ell_v, r_v\}} (\deg w -1) + |\mathcal L(v)| + |\mathcal R(v)|, \end{equation} where \( \deg w \) is the degree of vertex \(w\in V(G)\). \end{lemma} \begin{proof} To prove this lemma, we apply the so-called discharging method. Let us assign a charge to every edge \(wt\in E(G)\) as follows: \begin{itemize} \item[1.] If the vertex \(v\) is incident to \(wt\), then the charge of \(wt\) is 0. \item[2.] If \(wt\in \mathcal L(v)\cup \mathcal R(v)\), then the charge of \(wt\) is 1. \item[3.] If \(wt\not\in \mathcal L(v)\cup \mathcal R(v)\) and \(v\) is not incident to \(wt\), then the charge of \(wt\) is \(|\{w,t\}\cap N(v)\setminus\{\ell_v,r_v\}|\). \end{itemize} \begin{figure}[H] \label{figure1} \centering \begin{subfigure}[t]{0.3\textwidth} \centering \begin{tikzpicture}[line cap=round,line width = 1pt] \draw (0,0) node [below] {\( v \)} -- (-2,2) node [left] {\( \ell_v \)}; \draw (0,0) -- (2,2) node [right] {\( r_v \)}; \draw (0,0) -- (0,3) node [above] {\( w \)}; \draw (0,3) -- (2,1) node [below] {\( t \)}; \filldraw (0,0) circle (2pt); \filldraw (2,2) circle (2pt); \filldraw (-2,2) circle (2pt); \filldraw (0,3) circle (2pt); \filldraw (2,1) circle (2pt); \end{tikzpicture} \caption{} \label{figure1:a} \end{subfigure} \hfill \begin{subfigure}[t]{0.3\textwidth} \centering \begin{tikzpicture}[line cap=round,line width = 1pt \draw (0,0) node [below] {\( v \)} -- (-2,2) node [left] {\( \ell_v \)}; \draw (0,0) -- (2,2) node [right] {\( r_v \)}; \draw (0,0) -- (0,3) node [above] {\( w \)}; \draw (0,3) -- (2,2); \filldraw (0,0) circle (2pt); \filldraw (2,2) circle (2pt); \filldraw (-2,2) circle (2pt); \filldraw (0,3) circle (2pt); \end{tikzpicture} \caption{} \label{figure1:b} \end{subfigure} \hfill \begin{subfigure}[t]{0.3\textwidth} \centering \begin{tikzpicture}[line cap=round,line width = 1pt] \draw (0,0) node [below] {\( v \)} -- (-2,2) node [left] {\( \ell_v \)}; \draw (0,0) -- (2,2) node [right] {\( r_v \)}; \draw (0,0) -- (-1,3) node [above] {\( w \)}; \draw (0,0) -- (1,3) node [above] {\( t \)}; \draw (1,3) -- (-1,3); \filldraw (0,0) circle (2pt); \filldraw (2,2) circle (2pt); \filldraw (-2,2) circle (2pt); \filldraw (1,3) circle (2pt); \filldraw (-1,3) circle (2pt); \end{tikzpicture} \caption{} \label{figure1:c} \end{subfigure} \caption{An edge \(wt\) incident to a vertex of \(N(v)\setminus\{r_v,\ell_v\}\).} \end{figure} Remark that the charge of an edge is well-defined, and the sum of charges of all edges equals the right-hand side of the desired inequality. Hence it is enough to show that if an edge \(wt\in E(G)\) has charge 1 or 2, then it is disjoint from one or two of the edges \(vr_v\) and \(v\ell_v\), respectively. There are the following possible cases: \begin{itemize} \item[1.] If \(wt\in \mathcal R(v)\), then it has charge 1 and is disjoint from \(v\ell_v\). \item[2.] If \(wt \in \mathcal L(v)\), then it has charge 1 and is disjoint from \( vr_v\). \item[3.] If the edge \( wt \) connects a vertices in \(N(v)\setminus\{r_v,\ell_v\}\) with a vertex in \(V(G)\setminus (\{v\}\cup N(v)\setminus\{r_v,\ell_v\})\), then \(wt\) has charge 1 and is disjoint from at least one of the edges \(v\ell_v\) or \(vr_v\). See Figures~\ref{figure1:a} and~\ref{figure1:b}. \item[4.] If the edge \(wt\) connects two vertices from \(N(v)\setminus\{r_v,\ell_v\}\), then \(wt\) has charge 2 and lies inside of the angle \(r_vv\ell_v\), and thus, it is disjoint from the edges \(v\ell_v\) and \(vr_v\). See Figure~\ref{figure1:c}. \end{itemize} Since any other edge has charge 0, we are done. \end{proof} For \( v \in V(G) \), denote by \( \alpha_{\ell}(v) \) the number of vertices \( w\in N(v) \) with \( v=\ell_w \). Analogously, denote by \( \alpha_r(v) \) the number of vertices \( w\in N(v) \) with \( v=r_w \). Since any vertex has exactly one leftmost edge and exactly one rightmost edge, we get \[ \label{equation:sumofalpha} \sum_{ v\in V(G) } \alpha_\ell(v)=\sum_{v\in V(G)}\alpha_r(v)=n. \] Also, we use the standard equality \[ \label{equation:sumofdegrees} \sum_{w\in V(G)}\deg w=2e. \] \begin{corollary} \label{corollary:summation} We have \begin{gather} \sum_{v\in V(G)} \Big( | DJ_\ell (v) | + | DJ_r (v) | \Big) \geq \sum_{w\in V(G)} \deg^2 w - \\ \sum_{w\in V(G)}(\alpha_\ell(w) + \alpha_r(w))\deg w - 2(e -n) + \sum_{v\in V(G)} |\mathcal L(v)| + \sum_{v\in V(G)}|\mathcal R(v)| \end{gather} \end{corollary} \begin{proof} Summing up the inequality from Lemma~\ref{lemma:simple?} over all vertices of \(G\), we obtain that \begin{gather} \sum_{v\in V(G)} \Big( | DJ_\ell (v) | + | DJ_r (v) | \Big) \geq \sum_{v\in V(G)} \sum_{w\in N(v)} (\deg w - 1)- \\\sum_{w\in V(G)} (\alpha_\ell(w)+\alpha_r(w))(\deg w -1)+\sum_{v\in V(G)} |\mathcal L(v)| + \sum_{v\in V(G)}|\mathcal R(v)|. \end{gather} By~\eqref{equation:sumofalpha} and~\eqref{equation:sumofdegrees}, we are done. \end{proof} For a vertex \(v\), let \(\mathcal L'(v)\) be a subset of \(\mathcal L(v)\) consisting of edges \(\ell_vw\) such that \(\ell_w= \ell_v\). Analogously, let \(\mathcal R'(v)\) be a subset of \(\mathcal R(v)\) consisting of edges \(r_vw\) such that \(r_w= r_v\). \begin{lemma} \label{lemma:discharging!} For a vertex \(v\in V(G)\), the following equalities hold \[ \sum_{v\in V(G)}|\mathcal L'(v)| = \sum_{v\in V(G)}\frac{\alpha_\ell(v)(\alpha_\ell(v)-1)}{2} \text{\ \ and\ \ } \sum_{v\in V(G)} |\mathcal R'(v)| = \sum_{v\in V(G)} \frac{\alpha_r(v)(\alpha_r(v)-1)}{2}. \] \end{lemma} \begin{proof} Remark that if we prove \[ \label{equality1} \sum_{v\in V(G): \ell_v=w} |\mathcal L'(v)|=\frac{\alpha_\ell(w)(\alpha_\ell(w)-1)}{2} \] for any vertex \( w\in V(G) \), the first desired inequality follows. Analogously, one can show the second inequality. Since \eqref{equality1} is trivial if \(\alpha_\ell(w)=0 \), we may assume \(\alpha_\ell(w)=k>0\). Let \( u_1, \dots, u_k\in N(w) \) be distinct vertices with \(\ell_{u_i}=w\). Without loss of generality, assume that among \(wu_1, \dots, wu_k\), the edge \(wu_1\) is the leftmost edge, the edge \( wu_2 \) is the second leftmost edge, etc., that is, the angles \(\angle u_iwu_j\) for \( 1\leq i<j\leq k \) are positive. Therefore, \(|\mathcal L'(u_i)| = k-i\) for \(1\leq i\leq k\). Summing up all these equalities, we obtain~\eqref{equality1}. \end{proof} Denote by \( n_\ell\) the number of vertices \( v\in V(G) \) such that for every \(w\in N(v)\) we have \( v=\ell_w \). Analogously, denote by \( n_r \) the number of vertices \( v\in V(G) \) such that for every \( w\in N(v) \) we have \( v = r_w\). For a vertex \(v\in V(G)\), put \(\delta_\ell(v)=1\) if there is an edge \(\ell_vx\) such that \(\angle x\ell_v v\) is positive and \(\ell_x\ne \ell_v\), otherwise put \( \delta_\ell(v)=0 \). Analogously, we define \(\delta_r(v)\). Remark that \(\delta_\ell(v)=1\) if and only if the set \( \mathcal L(v)\setminus\mathcal L'(v)\) is not empty. Clearly, we have \[ \label{sum:ahaha} |\mathcal L(v)|\geq |\mathcal L'(v)|+\delta_\ell(v) \text{ and } |\mathcal R(v)|\geq |\mathcal R'(v)|+ \delta_r(v). \] \begin{corollary} \label{maincorollary} The following inequality holds \[ \sum_{v \in V(G)}\Big( |DJ_\ell(v)|+ |DJ_r(v)|\Big) \geq \frac{(2e-n)^2}{2(n - n_\ell)} + \frac{(2e-n)^2}{2(n - n_r)} - (2e - n)+\sum_{v\in V(G)}(\delta_\ell(v)+\delta_r(v)). \] \end{corollary} \begin{proof} By Corollary~\ref{corollary:summation}, Lemma~\ref{lemma:discharging!} and \eqref{sum:ahaha}, we obtain \begin{gather} \sum\limits_{v \in V(G)}\Big( |DJ_\ell(v)|+ |DJ_r(v)|\Big) \geq \sum_{w\in V(G)} \Big( \frac{\deg w^2}{2}-\alpha_\ell(w)\deg w + \frac{\alpha_\ell(w)(\alpha_\ell(w)-1)}{2}+ \\ \frac{\deg w^2}{2}-\alpha_r(w)\deg w + \frac{\alpha_r(w)(\alpha_r(w)-1)}{2}\Big)-2(e-n)+\sum_{v\in V(G)}\big(\delta_\ell(v)+\delta_r(v)\big)\label{lastsum} \end{gather} By~\eqref{equation:sumofalpha} and~\eqref{equation:sumofdegrees}, we have \begin{gather} \sum_{w\in V(G)} \Big( \frac{\deg w^2}{2}-\alpha_\ell(w)\deg w + \frac{\alpha_\ell(w)(\alpha_\ell(w)-1)}{2}\Big)=\\ \sum_{w\in V(G)}\frac{(\deg w - \alpha_\ell(w))^2}{2} -\frac{n}{2} \geq \frac{(2e-n)^2}{2(n-n_\ell)} - \frac{n}{2}, \end{gather} where to prove the last inequality, we use that the number of vanishing terms \(\deg w-\alpha_\ell(w)\) is equal to \(n_\ell\). Analogously, we show that \[ \sum_{w\in V(G)} \Big( \frac{\deg w^2}{2}-\alpha_r(w)\deg w + \frac{\alpha_r(w)(\alpha_r(w)-1)}{2}\Big)\geq \frac{(2e-n)^2}{2(n-n_r)} - \frac{n}{2}. \] Substituting these two inequalities in \eqref{lastsum}, we finish the proof of the corollary. \end{proof} \begin{corollary}\label{theorem1corollary} The following inequality holds \[ \sum\limits_{v \in V(G)}\Big( |DJ(v\ell_v)| + |DJ(vr_v)|\Big) \geq n\left(\frac{2e}{n}-1\right)\left(\frac{2e}{n}-2\right). \] \end{corollary} \begin{proof} By the definition of \( DJ_r(v) \) and \( DJ_\ell(v)\), we have \( |DJ(v\ell_v)| \geq |DJ_\ell(v)| \) and \( |DJ(vr_v)| \geq |DJ_r(v)|\), and therefore, \cref{maincorollary} yields \begin{gather} \sum\limits_{v \in V(G)}\Big( |DJ(v\ell_v)| + |DJ(vr_v)|\Big) \geq \frac{(2 e-n)^2}{n} - (2e - n)= n\left(\frac{2e}{n}-1\right)\left(\frac{2e}{n}-2\right), \end{gather} which finishes the proof. \end{proof} \section{Proof of \cref{theorem:main1}} \label{section:proof1} Consider two possible cases. \textit{Case 1.} Let \( G \) be a locally convex graph. Since \( (\sqrt{1+8m}+3)/4\geq 1\), we may assume that \( G \) does not contain vertices of degree 0 or 1; otherwise we can delete such a vertex and use induction on the number of vertices. Hence we can apply results of Section~\ref{section:preliminaries}. Combining \cref{theorem1corollary} together with \( |DJ(v\ell_v)| \leq m\) and \( |DJ(vr_v)| \leq m \), we obtain \[ 2mn \geq n\left(\frac{2e}{n}-1\right)\left(\frac{2e}{n}-2\right)=n\left(\left(\frac{2e}{n} -\frac32\right)^2 - \frac14\right), \] which finishes the proof of the first case. \textit{Case 2.} Let \( G \) have a non-convex vertex \( v \). Hence, there are vertices \( v_1, v_2, v_3 \in N(v) \) such that \( v \) lies in the convex hull of \( v_1, v_2,\) and \( v_3 \). Remark that any edge that is not incident to \( v \) is disjoint from at least one of the edges \( vv_1, vv_2, vv_3 \). Since each of the edges \( vv_1, vv_2, \) and \( vv_3\) is disjoint from at most \( m \) edges, we have that there are at most \( 3m + \deg v \) edges in \( G \). The inequality \( \deg v < n \) finishes the proof of the second case. \section{Proof of \cref{theorem:main2}} \label{section:proof2} The proof is by induction on \( e \). \textit{Base of induction.} Assume that \( e \leq 3n/2 \), and thus, \( d(G)=2e/n\leq 3\). If \(1\leq d(G) \leq 2\), the desired inequality trivially follows from \(\binom{d(G)}{3}\leq 0\). If \(2<d(G)\leq 3\), then consider a graph \( G' \) obtained from \( G \) by deleting one edge in every pair of disjoint edges. Clearly, \( |E(G')| \geq e - |DJ(G)| \) and edges in \( G' \) are pairwise intersecting. By \cref{erdos}, we have \( |E(G')| \leq n \). Thus, \[ |DJ(G)| \geq e - n = \frac{n(d(G)-2)}{2} \geq \frac{n}{2} \binom{d(G)}{3}. \] \textit{Induction step.} Assume that \( e > 3n/2 \), and thus, \(d(G)>3\). First, we show that we may assume that \( G \) does not contain vertices of degree 0 or~1. Indeed, let \[ F(G) := \frac{|V(G)|}{2} \binom{d(G)}{3} \] and suppose that \( G \) has a vertex of degree 0 or 1. Let \( G' \) be a graph obtained from \( G \) by removing this vertex. Then \(|DJ(G)| \geq |DJ(G')|\) and \begin{gather} F(G') \geq \frac{n-1}{2} \binom{\frac{2e-2}{n-1}}{3} = \frac{2e-2n}{12} \left(\frac{2e-2}{n-1}\right)\left(\frac{2e-n-1}{n-1}\right) \overset{(*)}{>} \\ \frac{2e-2n}{12} \left(\frac{2e-2}{n-1}\right)\left(\frac{2e-n-1}{n-1}\right)\left(\frac{e(n-1)}{(e-1)n}\right)\left(\frac{(2e-n)(n-1)}{(2e-n-1)n}\right)=\\ \frac{2e(2e-n)(2e-2n)}{12n^2} = F(G). \end{gather} Here in \( (*) \) we use that \[ \frac{e(n-1)}{(e-1)n} < 1 \text{ and } \frac{(2e-n)(n-1)}{(2e-n-1)n} < 1. \] Both of these inequalities follow from the fact that \( e > n > 1 \). Since \(d(G')>d(G)>3\), we may apply induction on the number of vertices and obtain the desired inequality. Therefore, without loss of generality we assume that all the vertices of \( G \) have degree at least 2, and thus, we can apply results of Section~\ref{section:preliminaries}. \medskip By \cref{maincorollary}, we may assume without loss of generality that \[ \label{equation of corollary} \sum\limits_{v \in V(G)}\ncrlsetcard{v} \geq \frac{(2 e-n)^2}{2(n - n_\ell) } -e +\frac{n}{2} + \sum_{v\in V(G)}\delta_\ell(v). \] Let \(G'\) be a graph obtained from \(G\) by deleting all leftmost edges and all vertices \(v\in V(G)\) with \(\alpha_\ell(v)=\deg v\), and thus, \(|V(G')|=n-n_\ell\) and \(|E(G')|=e-n+t_\ell\), where \(t_\ell\) is the number of double leftmost edges in \(G\). An edge \(uv\in E(G)\) is \textit{double leftmost} if \(u=\ell_v\) and \(v=\ell_u\). Before we apply the induction hypothesis for the graph \(G'\), let us prove a few auxiliary facts. \medskip First, notice that \begin{equation} \label{equation:djg} |DJ(G)| \geq |DJ(G')| + \sum\limits_{v \in V(G)} |DJ_\ell(v)| - |N_\ell(G)|, \end{equation} where \(N_\ell(G)\) is the set of pairs of leftmost edges \(v\ell_v, u\ell_u\in E(G)\) such that \(vu\) is an edge in \(G\). Remark that each such pair correspond to an edge in \(E(G')\) in the following way: for every edge \(uv \in E(G)\), the leftmost edges \(u\ell_u, v\ell_v\in E(G)\) are disjoint if and only if \(uv\in E(G')\). Therefore, \(|N_\ell(G)|=|E(G')|=e-n+t_\ell\). Second, let us show \[ \label{equation:the new inequality} \sum_{v\in V(G)}\delta_\ell(v)+n_\ell \geq 2t_\ell\geq n_\ell. \] Indeed, for any double leftmost edge \(uv\in E(G)\) we have \[ \delta_\ell(v)+\delta_\ell(u)+\beta_\ell(v)+\beta_\ell(u)= 2, \] where for every \(w\in V(G)\) put \(\beta_\ell(w)=1\) if \(\deg w = \alpha_\ell(w)\) and otherwise \(\beta_\ell(w)=0\). Summing up this equality over all double leftmost edges in \(G\) and using that a vertex is incident to at most one double leftmost edge, we easily obtain \eqref{equation:the new inequality}. Third, by~\eqref{equation:the new inequality}, we have \[ \label{equation:newavarage} d(G') = \frac{2(e-n+t_\ell)}{n - n_\ell}\geq \frac{2e-n}{n-n_\ell}-1\geq d(G)-2> 1, \] and thus, we can apply the induction hypothesis for \(G'\). \medskip By the induction hypothesis, \eqref{equation:djg}, \eqref{equation:newavarage}, and \eqref{equation of corollary}, we obtain \[ |DJ(G)|\geq \frac{(n - n_\ell)}{2} \cdot\binom{\frac{2e-n}{n-n_\ell} - 1}{3} + \frac{(2e-n)^2}{2(n - n_\ell)} -e+\frac{n}{2}+\sum_{v\in V(G)}\delta_\ell(v)- e + n - t_\ell. \] By~\eqref{equation:the new inequality}, we obtain \begin{gather} |DJ(G)|\geq \frac{n - n_\ell}{2} \cdot\binom{\frac{2e-n}{n-n_\ell} - 1}{3} + \frac{(2e-n)^2}{2(n - n_\ell) } - (2e-n) + \frac{n-n_\ell}{2} = \\ \frac{n-n_\ell}{2}\cdot \Bigg(\binom{\frac{2e-n}{n-n_\ell}-1}{3}+\left(\frac{2e-n}{n-n_\ell}-1\right)^2\Bigg)=\\ \frac{2e-n}{12}\cdot \Bigg(\left(\frac{2e-n}{n-n_\ell}\right)^2-1\Bigg)\geq \frac{2e-n}{12}\cdot \bigg(\left(\frac{2e-n}{n}\right)^2-1\bigg)=\frac{n}{2}\binom{\frac{2e}{n}}{3}, \end{gather} which finishes the proof of the theorem. \section{Discussion} \label{section:discussion} \subsection{Tightness of the theorems.} Let \( n, k \) be integers of different parity such that \( n-2 > k>2 \). Let \(G_{n,k}\) be a convex graph whose vertices are denoted \(x_1,\dots, x_n\) in cyclic order and edges are \(x_{i}x_{j}\) for \(j-i\equiv\frac{n-k-1}{2},\dots,\frac{n+k+1}{2} (\mathrm{mod}\ n)\); see Figure \ref{figure:example}. It is not difficult to verify that the number of edges of \(G_{n,k}\) is \(n(k+2)/2\). Moreover, \[ |DJ(G_{n,k})| = \frac{n}{2} \cdot \binom{k+2}{3} \text{ and } |DJ(uv)| \leq \frac{k(k+1)}{2} \text{ for any edge } uv\in E(G_{n,k}). \] Therefore, for the graph \(G_{n,k}\), the bounds in Theorems~\ref{theorem:main1} and~\ref{theorem:main2} are tight. \begin{figure}[h] \begin{subfigure}[b]{0.45\textwidth} \begin{tikzpicture}[line cap=round,line width = 1pt, x = 3.03085cm, y = 3.03085cm] \draw (1.0, 0.0) node [right]{\(x_1\)}; \draw (0.8412535328311812, 0.5406408174555976) node [right]{\( x_2 \)}; \draw (0.41541501300188644, 0.9096319953545183) node [right]{\( x_3 \)}; \draw (-0.142314838273285, 0.9898214418809328) node [right]{\( x_4 \)}; \draw (-0.654860733945285, 0.7557495743542583) node [above right]{\( x_5 \)}; \draw (-0.9594929736144974, 0.28173255684142967) node [above]{\( x_6 \)}; \draw (-0.9594929736144975, -0.2817325568414294) node [above left]{\( x_7 \)}; \draw (-0.6548607339452852, -0.7557495743542582) node [left]{\( x_8 \)}; \draw (-0.14231483827328523, -0.9898214418809327) node [left]{\( x_9 \)}; \draw (0.41541501300188605, -0.9096319953545186) node [right]{\(x_{10}\)}; \draw (0.8412535328311812, -0.5406408174555974) node [right]{\( x_{11}\)}; \filldraw (1.0, 0.0) circle (2pt); \filldraw (0.8412535328311812, 0.5406408174555976) circle (2pt); \filldraw (0.41541501300188644, 0.9096319953545183) circle (2pt); \filldraw (-0.142314838273285, 0.9898214418809328) circle (2pt); \filldraw (-0.654860733945285, 0.7557495743542583) circle (2pt); \filldraw (-0.9594929736144974, 0.28173255684142967) circle (2pt); \filldraw (-0.9594929736144975, -0.2817325568414294) circle (2pt); \filldraw (-0.6548607339452852, -0.7557495743542582) circle (2pt); \filldraw (-0.14231483827328523, -0.9898214418809327) circle (2pt); \filldraw (0.41541501300188605, -0.9096319953545186) circle (2pt); \filldraw (0.8412535328311812, -0.5406408174555974) circle (2pt); \draw (1.0, 0.0)--(-0.654860733945285, 0.7557495743542583); \draw (1.0, 0.0)--(-0.9594929736144974, 0.28173255684142967); \draw (1.0, 0.0)--(-0.9594929736144975, -0.2817325568414294); \draw (1.0, 0.0)--(-0.6548607339452852, -0.7557495743542582); \draw (0.8412535328311812, 0.5406408174555976)--(-0.9594929736144974, 0.28173255684142967); \draw (0.8412535328311812, 0.5406408174555976)--(-0.9594929736144975, -0.2817325568414294); \draw (0.8412535328311812, 0.5406408174555976)--(-0.6548607339452852, -0.7557495743542582); \draw (0.8412535328311812, 0.5406408174555976)--(-0.14231483827328523, -0.9898214418809327); \draw (0.41541501300188644, 0.9096319953545183)--(-0.9594929736144975, -0.2817325568414294); \draw (0.41541501300188644, 0.9096319953545183)--(-0.6548607339452852, -0.7557495743542582); \draw (0.41541501300188644, 0.9096319953545183)--(-0.14231483827328523, -0.9898214418809327); \draw (0.41541501300188644, 0.9096319953545183)--(0.41541501300188605, -0.9096319953545186); \draw (-0.142314838273285, 0.9898214418809328)--(-0.6548607339452852, -0.7557495743542582); \draw (-0.142314838273285, 0.9898214418809328)--(-0.14231483827328523, -0.9898214418809327); \draw (-0.142314838273285, 0.9898214418809328)--(0.41541501300188605, -0.9096319953545186); \draw (-0.142314838273285, 0.9898214418809328)--(0.8412535328311812, -0.5406408174555974); \draw (-0.654860733945285, 0.7557495743542583)--(1.0, 0.0); \draw (-0.654860733945285, 0.7557495743542583)--(-0.14231483827328523, -0.9898214418809327); \draw (-0.654860733945285, 0.7557495743542583)--(0.41541501300188605, -0.9096319953545186); \draw (-0.654860733945285, 0.7557495743542583)--(0.8412535328311812, -0.5406408174555974); \draw (-0.9594929736144974, 0.28173255684142967)--(1.0, 0.0); \draw (-0.9594929736144974, 0.28173255684142967)--(0.8412535328311812, 0.5406408174555976); \draw (-0.9594929736144974, 0.28173255684142967)--(0.41541501300188605, -0.9096319953545186); \draw (-0.9594929736144974, 0.28173255684142967)--(0.8412535328311812, -0.5406408174555974); \draw (-0.9594929736144975, -0.2817325568414294)--(1.0, 0.0); \draw (-0.9594929736144975, -0.2817325568414294)--(0.8412535328311812, 0.5406408174555976); \draw (-0.9594929736144975, -0.2817325568414294)--(0.41541501300188644, 0.9096319953545183); \draw (-0.9594929736144975, -0.2817325568414294)--(0.8412535328311812, -0.5406408174555974); \draw (-0.6548607339452852, -0.7557495743542582)--(1.0, 0.0); \draw (-0.6548607339452852, -0.7557495743542582)--(0.8412535328311812, 0.5406408174555976); \draw (-0.6548607339452852, -0.7557495743542582)--(0.41541501300188644, 0.9096319953545183); \draw (-0.6548607339452852, -0.7557495743542582)--(-0.142314838273285, 0.9898214418809328); \draw (-0.14231483827328523, -0.9898214418809327)--(0.8412535328311812, 0.5406408174555976); \draw (-0.14231483827328523, -0.9898214418809327)--(0.41541501300188644, 0.9096319953545183); \draw (-0.14231483827328523, -0.9898214418809327)--(-0.142314838273285, 0.9898214418809328); \draw (-0.14231483827328523, -0.9898214418809327)--(-0.654860733945285, 0.7557495743542583); \draw (0.41541501300188605, -0.9096319953545186)--(0.41541501300188644, 0.9096319953545183); \draw (0.41541501300188605, -0.9096319953545186)--(-0.142314838273285, 0.9898214418809328); \draw (0.41541501300188605, -0.9096319953545186)--(-0.654860733945285, 0.7557495743542583); \draw (0.41541501300188605, -0.9096319953545186)--(-0.9594929736144974, 0.28173255684142967); \draw (0.8412535328311812, -0.5406408174555974)--(-0.142314838273285, 0.9898214418809328); \draw (0.8412535328311812, -0.5406408174555974)--(-0.654860733945285, 0.7557495743542583); \draw (0.8412535328311812, -0.5406408174555974)--(-0.9594929736144974, 0.28173255684142967); \draw (0.8412535328311812, -0.5406408174555974)--(-0.9594929736144975, -0.2817325568414294); \end{tikzpicture} \subcaption{\( n = 11, k = 2\)} \end{subfigure} \begin{subfigure}[b]{0.45\textwidth} \begin{tikzpicture}[line cap=round,line width = 1pt, x=3cm,y=3cm] \draw (1.0, 0.0) node [right] {\( x_1 \)}; \draw (0.8660254038, 0.5) node [right] {\( x_2\)}; \draw (0.5, 0.8660254038) node [right] {\(x_3\)}; \draw (0.0, 1.0) node [right] {\( x_4 \)}; \draw (-0.5, 0.8660254038) node [above right] {\( x_5 \)}; \draw (-0.8660254038, 0.5) node [above right] {\( x_6 \)}; \draw (-1.0, 0.0) node [above] {\( x_7 \)}; \draw (-0.8660254038, -0.5) node [above left] {\( x_8 \)}; \draw (-0.5, -0.8660254038) node [above left] {\( x_9 \)}; \draw (0.0, -1.0) node [left] {\( x_{10} \)}; \draw (0.5, -0.8660254038) node [right] {\( x_{11} \)}; \draw (0.8660254038, -0.5) node [right] {\( x_{12} \)}; \filldraw (1.0, 0.0) circle (2pt); \filldraw (0.8660254038, 0.5) circle (2pt); \filldraw (0.5, 0.8660254038) circle (2pt); \filldraw (0.0, 1.0) circle (2pt); \filldraw (-0.5, 0.8660254038) circle (2pt); \filldraw (-0.8660254038, 0.5) circle (2pt); \filldraw (-1.0, 0.0) circle (2pt); \filldraw (-0.8660254038, -0.5) circle (2pt); \filldraw (-0.5, -0.8660254038) circle (2pt); \filldraw (0.0, -1.0) circle (2pt); \filldraw (0.5, -0.8660254038) circle (2pt); \filldraw (0.8660254038, -0.5) circle (2pt); \draw (1.0 , 0.0 ) -- ( -0.5 , 0.8660254038 ); \draw ( 1.0 , 0.0 ) -- ( -0.8660254038 , 0.5 ); \draw ( 1.0 , 0.0 ) -- ( -1.0 , 0.0 ); \draw ( 1.0 , 0.0 ) -- ( -0.8660254038 , -0.5 ); \draw ( 1.0 , 0.0 ) -- ( -0.5 , -0.8660254038 ); \draw ( 0.8660254038 , 0.5 ) -- ( -0.8660254038 , 0.5 ); \draw ( 0.8660254038 , 0.5 ) -- ( -1.0 , 0.0 ); \draw ( 0.8660254038 , 0.5 ) -- ( -0.8660254038 , -0.5 ); \draw ( 0.8660254038 , 0.5 ) -- ( -0.5 , -0.8660254038 ); \draw ( 0.8660254038 , 0.5 ) -- ( 0.0 , -1.0 ); \draw ( 0.5 , 0.8660254038 ) -- ( -1.0 , 0.0 ); \draw ( 0.5 , 0.8660254038 ) -- ( -0.8660254038 , -0.5 ); \draw ( 0.5 , 0.8660254038 ) -- ( -0.5 , -0.8660254038 ); \draw ( 0.5 , 0.8660254038 ) -- ( 0.0 , -1.0 ); \draw ( 0.5 , 0.8660254038 ) -- ( 0.5 , -0.8660254038 ); \draw ( 0.0 , 1.0 ) -- ( -0.8660254038 , -0.5 ); \draw ( 0.0 , 1.0 ) -- ( -0.5 , -0.8660254038 ); \draw ( 0.0 , 1.0 ) -- ( 0.0 , -1.0 ); \draw ( 0.0 , 1.0 ) -- ( 0.5 , -0.8660254038 ); \draw ( 0.0 , 1.0 ) -- ( 0.8660254038 , -0.5 ); \draw ( -0.5 , 0.8660254038 ) -- ( -0.5 , -0.8660254038 ); \draw ( -0.5 , 0.8660254038 ) -- ( 0.0 , -1.0 ); \draw ( -0.5 , 0.8660254038 ) -- ( 0.5 , -0.8660254038 ); \draw ( -0.5 , 0.8660254038 ) -- ( 0.8660254038 , -0.5 ); \draw ( -0.8660254038 , 0.5 ) -- ( 0.0 , -1.0 ); \draw ( -0.8660254038 , 0.5 ) -- ( 0.5 , -0.8660254038 ); \draw ( -0.8660254038 , 0.5 ) -- ( 0.8660254038 , -0.5 ); \draw ( -1.0 , 0.0 ) -- ( 0.5 , -0.8660254038 ); \draw ( -1.0 , 0.0 ) -- ( 0.8660254038 , -0.5 ); \draw ( -0.8660254038 , -0.5 ) -- ( 0.8660254038 , -0.5 ); \end{tikzpicture} \subcaption{\( n = 12, k = 3 \)} \end{subfigure} \caption{Graph \(G_{n,k}\).} \label{figure:example} \end{figure} \subsection{Open problems and conjectures.} Unfortunately, Theorem~\ref{theorem:main2} becomes wrong if one replaces the bound \[ \frac{n}{2}\binom{d(G)}{3}\quad \text{by}\quad \frac{1}{2}\sum_{v \in V(G)}\binom{\deg v}{3}. \] Indeed, consider a convex graph \(G\) on \(2n+2\) vertices consisting of two disjoint copies of a star \( S_n \)\footnote{A star \(S_k\) is a graph on \( k + 1 \) vertices with one vertex of degree \( k \) and \( k \) vertices of degree \( 1 \).}. For \( n\geq 9\), this graph \(G\) does not satisfy the desired inequality because of \[ \frac{1}{2}\sum_{v \in V(G)}\binom{\deg v}{3} = \binom{n}{3} > n^2= DJ(G). \] We believe that the following conjectures hold. \begin{conjecture} \label{conjecture:main} Let \(m\) be a non-negative integer and \(G\) be a geometric graph such that \( |DJ(uv)|\leq m\) for any edge \( uv\in E(G) \). Then \[ |E(G)|\leq \frac{\sqrt{1+8m}+3}{4}\cdot |V(G)|. \] \end{conjecture} \begin{conjecture} For any geometric graph \( G \) with \(2|E(G)|\geq |V(G)|\), we have \[ |DJ(G)| \geq \frac{n}{2} \cdot \binom{d(G)}{3}. \] \end{conjecture} At last, remark that Conjecture~\ref{conjecture:main} is proved if a geometric graph \( G\) has at least \(\frac{3(\sqrt{1+8m}-3)m}{2(m-1)}\) vertices. \bibliographystyle{amsalpha}

https://arxiv.org/abs/math/0511329

On the Inner Radius of Nodal Domains

Let M be a closed Riemannian manifold. We consider the inner radius of a nodal domain for a large eigenvalue \lambda. We give upper and lower bounds on the inner radius of the type C/\lambda^k. Our proof is based on a local behavior of eigenfunctions discovered by Donnelly and Fefferman, and a Poincaré type inequality proved by Maz'ya. Sharp lower bounds are known only in dimension two. We give an account of this case too.

\section{Introduction and Main Results} Let $(M, g)$ be a closed Riemannian manifold of dimension $n$. Let $\Delta$ be the Laplace--Beltrami operator on $M$. Let $0<\lambda_1\leq \lambda_2 \leq \ldots$ be the eigenvalues of $\Delta$. Let $\varphi_\lambda$ be an eigenfunction of $\Delta$ with eigenvalue $\lambda$. A \emph{nodal domain} is a connected component of $\{\varphi_\lambda\neq 0\}$. We are interested in the asymptotic geometry of the nodal domains. In particular, in this paper we consider the inner radius of nodal domains. Let $r_\lambda$ be the inner radius of the $\lambda$-nodal domain $U_\lambda$. Let $C_1, C_2, \ldots$ denote constants which depend only on $(M, g)$. We prove \begin{theorem} \label{thm:inrad} Let $M$ be a closed Riemannian manifold of dimension $n\geq 3$. Then $$ \frac{C_1}{\sqrt{\lambda}} \geq r_\lambda \geq \frac{C_2}{\lambda^{k(n)}(\log{\lambda})^{2n-4}}\, ,$$ where $k(n)=n^2-15n/8+1/4$. \end{theorem} In dimension two we have the following sharp bound \begin{theorem} \label{thm:inrad2} Let $\Sigma$ be a closed Riemannian surface. Then \begin{equation*} \frac{C_3}{\sqrt{\lambda}}\geq r_\lambda \geq \frac{C_4}{\sqrt{\lambda}} \, . \end{equation*} \end{theorem} \subsection{Upper Bound} We remark that the upper bound is more or less standard and has been used in the literature~(e.g.~\cite{don-fef-yau}). However, we explain it here also. We observe that $\lambda=\lambda_1(U_\lambda)$. This is true since the $\lambda$-eigenfunction does not vanish in $U_\lambda$ (\cite{chavel}, ch.~I.5). Therefore, the existence of the upper bound in Theorems~\ref{thm:inrad} and~\ref{thm:inrad2} follows from the following general upper bound on $\lambda_1$ of domains $\Omega\subseteq M$. \begin{theorem} \label{thm:inrad-gen} $$ \lambda_1(\Omega) \leq \frac{C_5}{\mathrm{inrad}(\Omega)^2}\, .$$ \end{theorem} The proof of this theorem is given in \S\ref{sec:gen-bounds}. \subsection{Lower Bound} For the lower bound on the inner radius in dimensions $\geq 3$, we give a proof in \S\ref{sec:lower3} which is based on a local behavior of eigenfunctions discovered by H.~Donnelly and C.~Fefferman (Theorem~\ref{thm:loc-cour}). The same proof gives in dimension two the bound $C/\sqrt{\lambda\log\lambda}$. In order to get rid of the factor $\sqrt{\log\lambda}$ in dimension two, we treat this case separately in \S\ref{subsec:lower2}. The proof for this case can basically be found in~\cite{ego-kon}, and we bring it here for the sake of clarity and completeness. For the dimension two case we also bring a new proof in \S\ref{sec:nazpolsod}. Moreover, this proof shows that a big inscribed ball can be taken to be with center at a maximal point of the eigenfunction in the nodal domain. This proof is due to F.~Nazarov, L.~Polterovich and M.~Sodin and is based on complex analytic methods. \subsection{A Short Background} Related to the problem discussed in this paper is the problem of estimating the $(n-1)$-Hausdorff measure $H_{n-1}(\lambda)$ of the nodal set, i.e.~the set where an eigenfunction vanishes. J.~Br\"uning and D.~Gromes proved in~\cite{brun-gromes} and~\cite{bruning} sharp lower estimates in dimension two. Namely, they showed $H_1(\lambda)\geq C\sqrt{\lambda}$. An estimate of the constant $C$ is given in~\cite{savo}. Later, S.~T.~Yau conjectured that in any dimension $C_1 \sqrt{\lambda}\geq H_{n-1}(\lambda)\geq C_2\sqrt{\lambda}$. This was proved in the case of analytic metrics by H.~Donnelly and C.~Fefferman in~\cite{don-fef-yau}. Regarding the inner radius of nodal domains, we would like to mention the recent work of B.~Xu \cite{xubin}, in which he obtains a sharp lower bound on the inner radius for at least two nodal domains, and the work of V.~Maz'ya and M.~Shubin \cite{maz-shub}, in which they give sharp bounds on the inner capacity radius of a nodal domain. \subsection{Acknowledgements} I am grateful to Leonid Polterovich for introducing me the problem and for fruitful discussions. I would like to thank Joseph Bernstein, Lavi Karp and Mikhail Sodin for enlightening discussions. I am thankful to Sven Gnutzmann for showing me the nice nodal domains pictures he generated with his computer program and for nice discussions. I owe my gratitude also to Moshe Marcus, Yehuda Pinchover and Itai Shafrir for explaining to me the subtleties of Sobolev spaces. I would like to thank Leonid Polterovich, Mikhail Sodin and F\"edor Nazarov for explaining their proof in dimension two to me, and for letting me publish it in \S\ref{sec:nazpolsod}. Special thanks are sent to Sagun Chanillo, Daniel Grieser, Mikhail Shubin, Bin Xu and the anonymous referee for their comments and corrections on the first manuscript of this paper. \section{The Lower Bound on the Inner Radius} \label{sec:lower} In this section we prove the existence of the lower bounds on the inner radius given in Theorems~\ref{thm:inrad} and~\ref{thm:inrad2}. \subsection{Lower Bound in Dimension $\geq$ 3} \label{sec:lower3} In this section we prove the existence of the lower bound in Theorem~\ref{thm:inrad}. The proof also gives a bound in the case where $\dim M = 2$, namely $r_\lambda\geq C/\sqrt{\lambda\log\lambda}$, but in the next section we treat this case separately to get rid of the $\sqrt{\log\lambda}$ factor. Let $\{\sigma_i\}$ be a finite cellulation of $M$ by cubes, such that for each $i$ we can put a Euclidean metric $e_i$ on $\sigma_i$, which satisfies $e_i/4 \leq g \leq 4e_i$. Let $r_{\lambda,i}$ be the inner radius of $U_{\lambda, i}=U_\lambda\cap \sigma_i$, and $r_{\lambda, i, e}$ be the Euclidean inner radius of $U_{\lambda, i}$. Notice that \begin{equation} r_{\lambda, i, e}\leq 2r_{\lambda,i} \leq 2r_\lambda\, . \end{equation} \noindent\underline{\scshape Step 1}. \begin{figure} $$\quad\quad\includegraphics{localcube.eps}$$ \caption{Proof of Lower Bound on Inner Radius} \label{fig:inrad} \end{figure} (See Fig.~\ref{fig:inrad}). We consider $\sigma_i$ as a compact cube in $\mathbb{R}^n$, with edges parallel to the axes directions. We cover $\sigma_i$ by non-overlapping small cubes with edges of size $4h$, where $r_{\lambda, i, e}<h<2r_{\lambda, i, e}$. Let $Q$ be a copy of one of these small cubes. Let $Q'$ be a concentric cube with parallel edges of size~$2h$. \noindent\underline{\scshape Step 2}. We note that each copy of $Q'$ contains a point $p\in\sigma_i\setminus U_\lambda$. Otherwise, we would have $r_{\lambda, i, e}\geq h$, which would contradict the definition of $h$. \noindent\underline{\scshape Step 3}. Denote by $\mathrm{hole}(p)$ the connected component of $Q\setminus (\sigma_i\cap U_\lambda)$ which contains~$p$. We claim \begin{equation} \label{eqn:loc-cour} \frac{\mathrm{Vol}_e(\mathrm{hole}(p))}{\mathrm{Vol}_e(Q)} \geq \frac{C_1}{\lambda^{\alpha(n)}(\log\lambda)^{4n}}, \end{equation} where $\alpha(n)=2n^2+n/4$, and $\mathrm{Vol}_e$ denotes the Euclidean volume. We will denote the right hand side term of~(\ref{eqn:loc-cour}) by~$\gamma(\lambda)$. Indeed, $\mathrm{hole}(p)$ is a connected component of $U_\lambda'\cap Q$ for some $\lambda$-nodal domain $U_\lambda'$. Hence, we can apply the following Local Courant's Nodal Domain Theorem. \begin{theorem}[\cite{don-fef, chan-mucken, lu}] \label{thm:loc-cour} Let $B\subseteq M$ be a fixed ball. Let $B'$ be a concentric ball of half the radius of $B$. Let $U_\lambda$ be a $\lambda$-nodal domain which intersects $B'$. Let $B_\lambda$ be a connected component of $B\cap U_\lambda$. Then \begin{equation} \label{eqn:df} \mathrm{Vol}(B_\lambda)/\mathrm{Vol}(B)\geq\frac{C_2}{\lambda^{\alpha(n)}(\log\lambda)^{4n}}, \end{equation} where $\alpha(n)=2n^2+n/4$. \end{theorem} We remark that in our case~(\ref{eqn:df}) is true also for the quotient of Euclidean volumes, since the Euclidean metric on $\sigma_i$ is comparable with the metric coming from $M$. \noindent\underline{\scshape Step 4}. We let $\tilde{\varphi}_{\lambda} =\chi(U_\lambda)\varphi_{\lambda}$, where $\chi(U_\lambda)$ is the characteristic function of $U_\lambda$, and similarly, $\tilde{\varphi}_{\lambda, i} = \chi(U_\lambda\cap \sigma_i) \varphi_\lambda$. Then we have the inequality \begin{equation} \label{ineq:loc_lambda} \int_{Q} |\tilde{\varphi}_{\lambda, i}|^2\,\mathrm{d}(\mathrm{vol}) \leq \beta(\lambda) h^2 \int_Q |\nabla\tilde{\varphi}_{\lambda,i}|^2\,\mathrm{d}(\mathrm{vol})\, , \end{equation} where $$\beta(\lambda) = \left\{\begin{array}{lcr} C_3\log(1/\gamma(\lambda)) &,& n=2\, , \\ C_4/(\gamma(\lambda))^{(n-2)/n} &,& n\geq 3\, . \end{array}\right.$$ \begin{proof} Observe that $\tilde{\varphi}_{\lambda, i}$ vanishes on $\mathrm{hole}(p)$. We will use the following Poincar\'{e} type inequality due to Maz'ya. We discuss it in \S\ref{subsec:poinca-cap}. A general version of this inequality with weights instead of Lebesgue measure is proved in~\cite{chan-wheed}. \begin{theorem} \label{thm:poinca-vol} Let $Q\subset \mathbb{R}^n$ be a cube whose edge is of length~$a$. Let $0<\gamma<1$. Then, $$\int_Q |u|^2\,\mathrm{d}(\mathrm{vol})\leq \beta a^2\int_Q |\nabla u|^2\,\mathrm{d}(\mathrm{vol})$$ for all Lipschitz functions $u$ on $Q$, which vanish on a set of measure $\geq\gamma a^n$, and where $$\beta = \left\{\begin{array}{lcr} C_{5}\log(1/\gamma) &,& n=2\, ,\\ C_{6}/\gamma^{(n-2)/n} &,& n\geq 3\, . \end{array}\right.$$ \end{theorem} From~(\ref{eqn:loc-cour}) and Theorem~\ref{thm:poinca-vol} applied to $\tilde{\varphi}_{\lambda, i}$, it follows \begin{equation*} \int_{Q} |\tilde{\varphi}_{\lambda,i}|^2\,\mathrm{d}(\mathrm{vol}_e) \leq \beta(\lambda) h^2 \int_Q |\nabla_e\tilde{\varphi}_{\lambda,i}|_e^2\,\mathrm{d}(\mathrm{vol}_e) \, . \end{equation*} Since the metric on $\sigma_i$ is comparable to the Euclidean metric, we have also inequality~(\ref{ineq:loc_lambda}). \end{proof} \noindent\underline{\scshape Step 5}. \begin{equation} \label{ineq:triangle} \int_{\sigma_i} |\tilde{\varphi}_\lambda|^2\,\mathrm{d}(\mathrm{vol}) \leq 16\beta(\lambda) r_\lambda^2 \int_{\sigma_i} |\nabla\tilde{\varphi}_\lambda|^2\,\mathrm{d}(\mathrm{vol}) \, . \end{equation} This is obtained by summing up inequalities~(\ref{ineq:loc_lambda}) over all cubes~$Q$ which cover $\sigma_i$, and recalling that $h< 2r_{\lambda, i, e}\leq 4r_\lambda$. \noindent\underline{\scshape Step 6}. We sum up~(\ref{ineq:triangle}) over all cubical cells $\sigma_i$ to obtain a global inequality. \begin{align} \label{ineq:global} \int_{U_\lambda}& |\varphi_\lambda|^2 \,\mathrm{d}(\mathrm{vol}) = \int_M |\tilde{\varphi}_\lambda|^2 \,\mathrm{d}(\mathrm{vol}) = \sum_i\int_{\sigma_i} |\tilde{\varphi}_\lambda|^2\,\mathrm{d}(\mathrm{vol}) \nonumber\\ & \leq 16\beta(\lambda) r_\lambda^2 \sum_i \int_{\sigma_i} |\nabla\tilde{\varphi}_\lambda|^2 \,\mathrm{d}(\mathrm{vol}) = 16\beta(\lambda) r_\lambda^2 \int_{M} |\nabla\tilde{\varphi}_\lambda|^2\,\mathrm{d}(\mathrm{vol})\nonumber\\ &= 16\beta(\lambda) r_\lambda^2 \int_{U_\lambda} |\nabla\varphi_\lambda|^2\,\mathrm{d}(\mathrm{vol}) \end{align} \noindent\underline{\scshape Step 7}. $$r_\lambda\geq\left\{ \begin{array}{lcr} C_{7}/\sqrt{\lambda\log\lambda} &, & n=2\, , \\ C_{8}/\lambda^{n^2-15n/8+1/4}(\log\lambda)^{2n-4}&, & n\geq 3\, . \end{array}\right.$$ Indeed, by~(\ref{ineq:global}) $$\lambda = \frac{\int_{U_\lambda} |\nabla\varphi_\lambda|^2\,\mathrm{d}(\mathrm{vol})} {\int_{U_\lambda} |\varphi_\lambda|^2 \,\mathrm{d}(\mathrm{vol})} \geq \frac{1}{16\beta(\lambda)r_\lambda^2}\, .$$ Thus, \begin{align*} r_\lambda &\geq \frac{1}{4\sqrt{\lambda\beta(\lambda)}}= \left\{\begin{array}{lcr} C_{9}/\sqrt{\lambda\log(1/\gamma(\lambda))} &,& n=2\, , \\ C_{10}\gamma(\lambda)^{(n-2)/2n}/\sqrt{\lambda} &,& n\geq 3\, . \end{array}\right. \\ & \geq \left\{\begin{array}{lcr} C_{11}/\sqrt{\lambda\log\lambda} &,& n=2\, , \\ C_{12}/\left(\lambda^{n^2-15n/8+1/4}(\log\lambda)^{2n-4}\right) &,& n\geq 3\, . \end{array}\right. \end{align*} \qed \subsection{Lower Bound in Dimension $=$ 2} \label{subsec:lower2} We prove the existence of the lower bound on the inner radius in Theorem~\ref{thm:inrad2}. The arguments below can basically be found in~\cite{ego-kon}, Chapter~7. We begin the proof of Theorem~\ref{thm:inrad2} with Step~1 and Step~2 of \S\ref{sec:lower3}. We proceed as follows. \noindent\underline{\scshape Step 3'}. If $\mathrm{hole}(p)$ does not touch $\partial Q$ $$\frac{\mathrm{Area}_e(\mathrm{hole}(p))}{\mathrm{Area}_e(Q)}\geq C_{1}\, ,$$ where $\mathrm{Area}_e$ denotes the Euclidean area. \begin{proof} We recall the Faber-Krahn inequality in $\mathbb{R}^n$. \begin{theorem} \label{thm:faber-krahn} Let $\Omega\subseteq\mathbb{R}^n$ be a bounded domain. Then $\lambda_1(\Omega) \geq C_{2}/\mathrm{Vol}(\Omega)^{2/n}$ \end{theorem} We apply Theorem~\ref{thm:faber-krahn} with $\Omega=\mathrm{hole}(p)$. We emphasize that $\lambda_1(\mathrm{hole}(p), g) \geq C_{3}\lambda_1(\mathrm{hole}(p), e)$, since the two metrics are comparable. Thus, we obtain $$\lambda = \lambda_1(\mathrm{hole}(p), g) \geq \frac{C_{4}}{\mathrm{Area}_e(\mathrm{hole}(p))}\, ,$$ or, written differently, $\mathrm{Area}_e(\mathrm{hole}(p))\geq C_{4}/\lambda$. On the other hand, $\mathrm{Area}_e(Q)=(4h)^2\leq 64 r_\lambda^2\leq 64C_5/\lambda$, where the last inequality is the upper bound on the inner radius in Theorem~\ref{thm:inrad2}. So take $C_{1}=C_{4}/(64C_5)$. \end{proof} \noindent\underline{{\scshape Step 4'} (part a)}. There exists an edge of $Q$, on which the orthogonal projection of $\mathrm{hole}(p)$ is of Euclidean size $\geq\gamma\cdot 4h$, where $0<\gamma<1$ is independent of $\lambda$. Let us denote by $|\mathrm{pr}(\mathrm{hole}(p))|$ the maximal size of the projections of $\mathrm{hole}(p)$ on one of the edges of $Q$. If $\mathrm{hole}(p)$ touches $\partial Q$, then $|\mathrm{pr}(\mathrm{hole}(p))|\geq 4h/4=h$, and we can take $\gamma=1/4$. Otherwise, by Step~3' \begin{align*} \label{ineq:pr} |\mathrm{pr}&(\mathrm{hole}(p))| \geq \sqrt{\mathrm{Area}_e(\mathrm{hole}(p))}\\ &\geq \sqrt{C_{1} (4h)^2} = 4\sqrt{C_{1}}h\, . \end{align*} So, we can take $\gamma=\sqrt{C_{1}}$. \noindent\underline{{\scshape Step 4'} (part b)}. \begin{equation} \label{eqn:locpoinca2} \int_Q |\tilde{\varphi}_{\lambda,i}|^2\,\mathrm{d} \mathrm{vol}_e \leq C_{6} h^2 \int_Q |\nabla\tilde{\varphi}_{\lambda,i}|^2 \,\mathrm{d} \mathrm{vol}_e . \end{equation} Notice that $\tilde{\varphi}_{\lambda, i}$ vanishes on $\mathrm{hole}(p)$. Hence, Step~4' (part a) permits us to apply the following Poincar\'{e} type inequality to $\tilde{\varphi}_{\lambda,i}$. Its proof is given in \S\ref{sec:poinca-2}. An inequality in the same spirit can be found in~\cite{leonsimon}. \begin{theorem}[\cite{ego-kon}, ch.~7] \label{thm:poinca-dim2} Let $Q\subseteq \mathbb{R}^2$ be a cube whose edge is of length $a$. Let $u$ be a Lipschitz function on $Q$ which vanishes on a curve whose projection on one of the edges is of size $\geq \gamma a$. Then $$\int_Q |u|^2\,\mathrm{d} x \leq C(\gamma) a^2 \int_Q |\nabla u|^2\,\mathrm{d} x\, .$$ \end{theorem} \noindent\underline{\scshape Steps 5'--7'}. To conclude we continue in the same way as in Steps~5--7 of \S\ref{sec:lower3}. \qed \section{A New Proof in Dimension Two} \label{sec:nazpolsod} This section is due to L.~Polterovich, M.~Sodin and F.~Nazarov. In dimension two we give a proof based on the harmonic measure and the fact due to Nadirashvili that an eigenfunction on the scale comparable to the wavelength is almost harmonic in a sense to be defined below. This proof also gives information about the location of a big ball inscribed in the nodal domain $U_\lambda$. Namely, we show that if $\phi_\lambda(x_0) = \max_{U_{\lambda}}|\phi_\lambda|$, then one can find a ball of radius $C/\sqrt{\lambda}$ centered at $x_0$ and inscribed in $U_\lambda$. Let $D_p\subseteq\Sigma$ be a metric disk centered at $p$. Let $f$ be a function defined on $D$. Let $\mathbb{D}$ denote the unit disk in $\mathbb{C}$. \begin{definition} We say that $f$ is \emph{$(K, \delta)$-quasi\-harmonic} if there exists a $K$-quasi\-conformal homeomorphism $h:D\to\mathbb{D}$, a harmonic function $u$ on $\mathbb{D}$, and a function $v$ on $\mathbb{D}$ with $1-\delta\leq v\leq 1$, such that \begin{equation} f= (v\cdot u)\circ h\, . \end{equation} \end{definition} \noindent\textbf{Remark}. We will assume without loss of generality that $h(p) =0$. \begin{theorem}[\cite{nadir, naz-pol-sod}] \label{thm:harmon} There exist $K, \varepsilon, \delta>0$ such that for every eigenvalue $\lambda$ and disk $D\subseteq\Sigma$ of radius $\leq\varepsilon/\sqrt{\lambda}$, $\varphi_\lambda|_{D}$ is $(K, \delta)$-quasi\-harmonic. \end{theorem} We now choose a preferred system of conformal coordinates on $(\Sigma, g)$. \begin{lemma} \label{lem:coor} There exist positive constants $q_+, q_-, \rho$ such that for each point $p\in M$, there exists a disk $D_{p,\rho}$ centered at $p$ of radius $\rho$, a conformal map $\Psi_p:\mathbb{D}\to D_{p, \rho}$ with $\Psi_p(0)=p$, and a positive function $q(z)$ on $\mathbb{D}$ such that $$\Psi_p^{*}(g) = q(z)|dz|^2, $$ with $q_-<q < q_+$. \end{lemma} Let us take a point $p$, where $|\varphi_\lambda|$ admits its maximum on $U_\lambda$. Let $R=\varepsilon/\sqrt{\lambda q_+}$. Let $D_{p, R\sqrt{q_+}}\subseteq D_{p, \rho}$ be a disk of radius $R\sqrt{q_+}$ centered at $p$. We now take the functions $u, v$ defined on $\mathbb{D}$ which correspond to $\varphi_\lambda|_{D_{p, R\sqrt{q_+}}}$ in Theorem~\ref{thm:harmon}. We observe that \begin{equation*} \varphi_\lambda(p)= u(0)v(0)\geq u(z)v(z)\geq u(z)(1-\delta) \, , \end{equation*} for all $z\in\mathbb{D}$. Hence \begin{equation} \label{ineq:max} u(0) \geq (1-\delta) \max_\mathbb{D} u \, . \end{equation} Now we apply the harmonic measure technic. Let $U_{\lambda}^0\subseteq \mathbb{D}$ be the connected component of $\{u>0\}$, which contains $0$. Let $E=\mathbb{D}\setminus U_{\lambda}^0$. Let $\omega$ be the harmonic measure of $E$ in $\mathbb{D}$. $\omega$ is a bounded harmonic function on $U_\lambda^0$, which tends to $1$ on $\partial U_\lambda^0\cap \mathrm{Int}(\mathbb{D})$ and to $0$ on the interior points of $\partial U_\lambda^0\cap \partial \mathbb{D}$. Let $r_0=\inf\{|z| :\, z\in E\}$. By the Beurling-Nevanlinna theorem (\cite{conf-inv}, sec.~3-3), \begin{equation} \label{ineq:be-ne-nad} \omega(0) \geq 1-C_1\sqrt{r_0}\, . \end{equation} By the majorization principle \begin{equation} \label{ineq:major-nad} u(0)/\max u \leq 1-\omega(0)\, . \end{equation} Combining inequalities~(\ref{ineq:max}),~(\ref{ineq:be-ne-nad}) and~(\ref{ineq:major-nad}) gives us \begin{equation} \label{ineq:r-c} r_0\geq C_2\, . \end{equation} In the final step we apply a distortion theorem proved by Mori for quasi\-conformal maps. Denote by $\mathbb{D}_r\subseteq \mathbb{C}$ the disk $\{|z| < r\}$. Observe that $$\Psi_p(\mathbb{D}_{R})\subseteq D_{p,R\sqrt{q_+}}\, .$$ Hence, we can compose $$\tilde{h} =h\circ \Psi_p :\mathbb{D}_{R}\to\mathbb{D}\, .$$ $\tilde{h}$ is a $K$-quasi\-conformal map. By Mori's Theorem (\cite{quasiconf}, Ch.~III.C) it is $\frac{1}{K}$-H\"{o}lder. Moreover, it satisfies an inequality \begin{equation} \label{ineq:mori} |\tilde{h}(z_1)-\tilde{h}(z_2)|\leq M\left(\frac{|z_1-z_2|}{R}\right)^{1/K}, \end{equation} with $M$ depending only on $K$. Inequalities~(\ref{ineq:r-c}) and~(\ref{ineq:mori}) imply that \begin{equation} \frac{\mathrm{dist}(p, \partial (U_\lambda\cap D_{p, R}))}{R}\geq \left(\frac{C_2}{M}\right)^K\sqrt{q_-}\, . \end{equation} Hence, \begin{equation} \mathrm{inrad}(U_\lambda)\geq \left(\frac{C_2}{M}\right)^K\sqrt{q_-} R = C_3/\sqrt{\lambda}\, , \end{equation} as desired. \section{A Review of Poincar\'{e} Type Inequalities} \label{sec:poinca} We give an overview of several Poincar\'e type inequalities. In particular, we prove Theorem~\ref{thm:poinca-vol} and Theorem~\ref{thm:poinca-dim2}. \subsection{Poincar\'e Inequality and Capacity} \label{subsec:poinca-cap} Theorem~\ref{thm:poinca-vol} is a direct corollary of the following two inequalities proved by Maz'ya. \begin{theorem}[\cite{mazya63}, \S 10.1.2 in \cite{mazya-book}] \label{thm:poinca-cap} Let $Q\subseteq\mathbb{R}^n$ be a cube whose edge is of length~$a$. Let $F\subseteq Q$. Then $$\int_{Q} |u|^2\,\mathrm{d}(\mathrm{vol})\leq \frac{C_1 a^n}{\mathrm{cap}(F, 2Q)}\int_Q |\nabla u|^2\,\mathrm{d}(\mathrm{vol})$$ for all Lipschitz functions $u$ on $Q$ which vanish on $F$. \end{theorem} A few remarks: \begin{itemize} \item[(a)] $2Q$ denotes a cube concentric with $Q$, with parallel edges of size twice as large. \item[(b)] If $\Omega\subseteq\mathbb{R}^n$ is an open set, and $\bar{F}\subseteq \Omega$, then $\mathrm{cap}(F, \Omega)$ denotes the $L^2$-capacity of $F$ in $\Omega$, namely $$\mathrm{cap}(F, \Omega) = \inf_{u\in\mathcal{F}}\left\{\int_\Omega |\nabla u|^2\,\mathrm{d} x\right\}, $$ where $\mathcal{F}=\{ u\in C^{\infty}(\Omega),\ u\equiv 1 \mbox{ on } F,\ \mathrm{supp} (u)\subseteq \Omega\}. $ \item[(c)] By Rademacher's Theorem~(\cite{ziemer}), a Lipschitz function is differentiable almost everywhere, and thus the right hand side has a meaning. \item[(d)] A generalization of the inequality to a body which is starlike with respect to a ball is proved in~\cite{maz-shub-disc}. \end{itemize} The next theorem is a capacity--volume inequality. \begin{theorem}[\S2.2.3 in \cite{mazya-book}] $$ \mathrm{cap} (F, \Omega) \geq \left\{\begin{array}{lcr} C_2/\log (\mathrm{Area}(\Omega)/\mathrm{Area}(F)) &,& n=2\, , \\ C_3 /(\mathrm{Vol}(F)^{-(n-2)/n}-\mathrm{Vol}(\Omega)^{-(n-2)/n}) &,& n\geq 3\, . \end{array}\right. $$ In particular, for $n\geq3$ we have \begin{equation*} \mathrm{cap}(F, \Omega) \geq C_3 \mathrm{Vol}(F)^{(n-2)/n} \, . \end{equation*} \end{theorem} \subsection{A Poincar\'e Inequality in Dimension Two} \label{sec:poinca-2} In this section we prove Theorem~\ref{thm:poinca-dim2}. The proof can be found in chapter~7 of~\cite{ego-kon}. We bring it here for the sake of clarity. \begin{proof} Let the coordinates be such that $Q=\{0\leq x_1, x_2 \leq a\}$. Let the given edge be $Q\cap \{x_1=0\}$, and let $\mathrm{pr}$ denote the projection from $Q$ onto this edge. Set $E = \mathrm{pr}^{-1}(\mathrm{pr}(\mathrm{hole}(p)))$. We claim \begin{equation} \label{ineq:E-Q} \int_{E} |u|^2\,\mathrm{d} x \leq a^{2} \int_Q |\nabla u|^2 \,\mathrm{d} x. \end{equation} Indeed, let $E_t := E\cap\{x_2 = t\}$. We recall the following Poincar\'e type inequality in dimension one whose proof is given below. \begin{lemma} \label{ineq:poin-1} \begin{equation} \label{eqn:poin-1} \int_a^b |u|^2\,\mathrm{d} x\leq |b-a|^2\int_a^b |u'|^2\,\mathrm{d} x \end{equation} for all Lipschitz functions $u$ on $[a, b]$ which vanish at a point of $[a,b]$. \end{lemma} By this lemma $$\int_{E_{t}} |u|^2\, \mathrm{d} x_1 \leq a^2 \int_{E_t} |\partial_1 u(x_1, t)|^2 \,\mathrm{d} x_1.$$ Integrating over $t\in\mathrm{pr}(\mathrm{hole}(p))$ gives us~(\ref{ineq:E-Q}). Next we show \begin{equation} \label{ineq:var2} \int_Q |u|^2 \, \mathrm{d} x \leq C_1 a^2 \int_Q |\nabla u|^2 \,\mathrm{d} x. \end{equation} By the mean value theorem $\exists t_0$ such that \begin{equation} \label{ineq-avg} \int_{E_{t_0}} |u|^2\, \mathrm{d} x_1 \leq \frac{1}{\gamma\cdot a} \int_E |u|^2\,\mathrm{d} x. \end{equation} In addition, we have \begin{eqnarray*} |u(x)|^2 &\leq& 2|u(x_1, t_0)|^2 +2 |u(x)-u(x_1, t_0)|^2 \\ &\leq& 2|u(x_1, t_0)|^2 +2 \left(\int_{t_0}^{x_2} |\partial_2 u(x_1, s)|\,\mathrm{d} s\right)^2 \\ &\leq& 2|u(x_1, t_0)|^2 +2\cdot a\int_{0}^{a} |\partial_2 u(x_1, s)|^2\,\mathrm{d} s. \end{eqnarray*} Integrating the last inequality over $Q$ gives us \begin{equation} \label{ineq-mid} \int_Q |u|^2\,\mathrm{d} x \leq 2\cdot a\int_{E_{t_0}} |u(x_1, t_0)|^2\,\mathrm{d} x_1 + 2a^2\int_Q |\partial_2 u|^2 \,\mathrm{d} x\, . \end{equation} Finally, we combine~(\ref{ineq:E-Q}),~(\ref{ineq-avg}) and~(\ref{ineq-mid}) to get~(\ref{ineq:var2}). \begin{align*} \int_Q & |u|^2\,\mathrm{d} x \leq 2 \cdot a\frac{1}{\gamma a}\int_E |u|^2\,\mathrm{d} x + 2\cdot a^2 \int_Q |\nabla u|^2\,\mathrm{d} x\\ & \leq C_1 a^2 \int_Q |\nabla u|^2 \,\mathrm{d} x\,. \end{align*} \end{proof} \subsection{A Poincar\'{e} Inequality in Dimension One} We prove Lemma~\ref{ineq:poin-1}. \begin{proof} By scaling, it is enough to prove~(\ref{eqn:poin-1}) for the segment $[0,1]$. Suppose $u(x_0)=0$. Since a Lipschitz function is absolutely continuous, we have $$|u(x)|^2 =\left|\int_{x_0}^x u'(t)\,\mathrm{d} t\right|^2\leq \int_0^1 \left|u'(t)\right|^2\,\mathrm{d} t.$$ We integrate over $[0,1]$ to get the desired inequality. \end{proof} \section{$\lambda_1$ and Inner Radius} \label{sec:gen-bounds} We prove Theorem~\ref{thm:inrad-gen}, which relates the inner radius to $\lambda_1$. \begin{proof} Let $\{V_i\}$ be a finite open cover of $M$, such that for each $i$ one can put a Euclidean metric $e_i$ on $V_i$, which satisfies $e_i/4\leq g\leq 4e_i$. Let $\alpha$ be the Lebesgue number of the covering. Let $r=\min(\mathrm{inrad}(\Omega), \alpha)$. Let $B\subseteq \Omega$ be a ball of radius $r$. We can assume that $B\subseteq V_1$. Let $B_e\subseteq B$ be a Euclidean ball of radius $r/2$. By monotonicity of $\lambda_1$, we know that $\lambda_1(\Omega, g)\leq \lambda_1(B, g)\leq \lambda_1(B_e, g)$, but since the Riemannian metric on $B_e$ is comparable to the Euclidean metric on it, it follows from the variational principle that $$\lambda_1(B_e, g) \leq C_1\lambda_1(B_e, e_1) = C_2/r^2\leq C_3/\mathrm{inrad}(\Omega)^2,$$ where in the last inequality we used the fact that $\mathrm{inrad}(\Omega)\leq C_4\alpha$. \end{proof} \noindent\textbf{Remark}. We would like to emphasize that in general there is no lower bound on $\lambda_1$ is terms of the inner radius. However, in dimension two, as pointed out to us by Daniel Grieser and Mikhail Shubin, there exists a lower bound on $\lambda_1$ in terms of the inner radius and the connectivity of $\Omega$. This was proved in~\cite{hayman} and~\cite{osserman}. For a more detailed account of the subject one can consult~\cite{lieblow}. \providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace} \providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR } \providecommand{\MRhref}[2]{% \href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2} } \providecommand{\href}[2]{#2}

https://arxiv.org/abs/2107.06233

Generalizations of the Yao-Yao partition theorem and the central transversal theorem

We generalize the Yao-Yao partition theorem by showing that for any smooth measure in $R^d$ there exist equipartitions using $(t+1)2^{d-1}$ convex regions such that every hyperplane misses the interior of at least $t$ regions. In addition, we present tight bounds on the smallest number of hyperplanes whose union contains the boundary of an equipartition of a measure into $n$ regions. We also present a simple proof of a Borsuk-Ulam type theorem for Stiefel manifolds that allows us to generalize the central transversal theorem and prove results bridging the Yao--Yao partition theorem and the central transversal theorem.

\section{Introduction} Mass partition problems study how one can split finite sets of points or measures in Euclidean spaces. They connect topological combinatorics and computational geometry \cites{matousek2003using, Zivaljevic:2017vi, roldan2021survey}. We say that a finite family $P$ of subsets of $\mathds{R}^d$ is a \textit{convex partition} of $\mathds{R}^d$ if the union of the sets is $\mathds{R}^d$, the interiors of the sets are pairwise disjoint, and each set is closed and convex. For a finite measure $\mu$ in $\mathds{R}^d$, we say that a convex partition of $\mathds{R}^d$ is an \textit{equipartition} of $\mu$ if each set in $P$ has the same $\mu$-measure. In 1985, Yao and Yao proved the following theorem, motivated by applications in geometric range queries \cite{yao1985general}. \begin{theorem}[Yao and Yao 1985]\label{thm:yaoyao} For any finite measure $\mu$ in $\mathds{R}^d$ that is absolutely continuous with respect to the Lebesgue measure, there exists a convex equipartition of $\mathds{R}^d$ into $2^d$ regions such that any hyperplane avoids the interior of at least one region. \end{theorem} For $d=2$ these partitions are made by two lines, but they become much more involved in higher dimensions. In the original proof, the measure has to be the integral of a twice differentiable positive density function. Lehec \cite{lehec2009yao} found an alternate proof that weakened the condition to every hyperplane having measure 0. In addition to the original applications by Yao and Yao, the discrete version of Theorem \ref{thm:yaoyao} has been applied to geometric Ramsey questions \cite{alon2005crossing}. The proof technique for the Yao--Yao partition theorem is quite different from standard results in mass partition problems. The question of how far the technique Yao and Yao used can be pushed has been relatively unexplored. In 2014, Rold\'an-Pensado and Sober\'on extended the theorem so that any hyperplane avoids two convex regions and proved a general upper bound in the case of avoiding $t$ convex regions in $\mathds{R}^2$ \cite{roldan2014extension}. They also provide improved asymptotic bounds when $d$ is fixed and $t$ tends to infinity. The following problem remains open. \begin{problem}\label{problem:generalizedyao} Let $t, d$ be positive integers. Find the smallest value $n$ such that for any finite measure $\mu$ in $\mathds{R}^d$ that is absolutely continuous with respect to the Lebesgue measure there exists a convex equipartition of $\mu$ into $n$ parts such that every hyperplane avoids the interior of at least $t$ regions. \end{problem} Currently, the best bound for fixed $t$ is the naive bound that comes from Theorem \ref{thm:yaoyao}. Take any Yao--Yao partition in $\mathds{R}^d$ and partition each cell using $t-1$ parallel hyperplanes. This gives us a convex equipartition such that any hyperplane avoids the interior of at least $t$ regions, so $n \leq t\cdot2^{d}$. In this paper, we improve the naive bound with the following theorem that generalizes Theorem \ref{thm:yaoyao} and one of the main results from \cite{roldan2014extension}. \begin{theorem}\label{thm:generalyao} Let $t,d$ be positive integers. For any finite measure $\mu$ in $\mathds{R}^d$, absolutely continuous with respect to the Lebesgue measure, there exists a convex partition of $\mathds{R}^d$ into $(t+1)2^{d-1}$ regions of equal $\mu$-measure such that every hyperplane avoids the interior of at least $t$ regions. \end{theorem} This bound is exact in the case of $d=1$ and matches the known bounds for $d=2$ and any $t$ and for $t=1,2$ and any $d$ \cite{roldan2014extension}. Problem \ref{problem:generalizedyao} focuses on hyperplane transversals, so a natural question to ask is for the smallest number of regions needed to avoid all affine subspace transversals of other dimensions. In this paper, we study the case of lines. Because a line can pass through a hyperplane at most once, it is sufficient to obtain partitions whose boundaries are contained in the union of few hyperplanes. Some existing mass partitions problems exhibit convex equipartitions with few hyperplanes containing all boundaries. One classic example is the Gr\"unbaum--Hadwiger--Ramos problem \cites{grunbaum1960partitions, Blagojevic:2018jc}, where the aim is to split simultaneously as many measures as possible into $2^k$ equal parts using $k$ hyperplanes. Another is a recent conjecture by Langerman \cites{barba2019sharing, Hubard:2019we}, which claims that \textit{any $dk$ measures can be simultaneously split into two equal parts by a chessboard coloring induced by $k$ hyperplanes}. The key difference with the problem we discuss here is that we don't require the hyperplanes to extend indefinitely. \begin{problem}\label{problem:smallestkforn} Given positive integers $n, d$, find the smallest integer $k$ such that any finite measure in $\mathds{R}^d$ absolutely continuous with respect to the Lebesgue measure can be partitioned into $n$ convex regions of equal measure whose boundaries are contained in the union of at most $k$ hyperplanes. \end{problem} Using the classic ham sandwich theorem in a recursive argument we obtain the following bounds. \begin{theorem}\label{thm:smallestkforn} Let $n,d$ be positive integers. The following bounds hold for Problem \ref{problem:smallestkforn} \[ \left \lceil \frac{n-1}{d} \right \rceil \leq k \leq \frac{n}{d} + (d-1)\log_2\left( \frac{n}{d}\right) + d-2. \] \end{theorem} For a fixed dimension, these bounds imply that $k = n/d +O(\log (n))$. One consequence is the following Yao--Yao type corollary. \begin{corollary}\label{coro-line-miss-many} Let $d$ be a fixed positive integer. For any positive integer $n$ and any measure $\mu$ in $\mathds{R}^d$ absolutely continuous with respect to the Lebesgue measure, there exists a convex partition of $\mathds{R}^d$ into $n$ regions of equal $\mu$-measure such that every line misses the interior of at least $(d-1)n/d - O(\log n)$ regions. \end{corollary} Using a generalization of the ham sandwich theorem for ``well-separated'' families proved by B\'ar\'any, Hubard, and Jer\'onimo \cite{barany2008slicing}, we can improve this bound even further in the cases of $d=2$ and $d=3$ \[ \left \lceil \frac{n-1}{d} \right \rceil \leq k \leq \frac{n}{d} + O(1) . \] This sharper bound mainly solves these two cases as the $O(1)$ term is bounded by $1/2$ and $13/3$ for $d=2$ and $d=3$, respectively. We present multiple constructions that exhibit each of these bounds, each with different geometric properties. Finally, we study the connection of Yao--Yao partitions with yet another generalization of the ham sandwich theorem. The following result, proved independently by Dolnikov \cite{dol1992generalization} and \v{Z}ivaljevi\'c and Vre\'cica \cite{zivaljevic1990extension}, is known as the central transversal theorem. \begin{theorem}[Central transversal theorem]\label{thm:ctt} Let $k, d$ be non-negative integers such that $k \leq d - 1$. For any set of k+1 absolutely continuous probability measures $\mu_{1}, \ldots, \mu_{k+1}$ there exists a $k$-dimensional affine space $V$ such that for any closed half-space $H$ with $V \subset H$ we have $\mu_{i}(H) \geq \frac{1}{d-k+1}$ for $i=1,\ldots, k+1$. \end{theorem} The known proofs of this result involve computing non-trivial topological invariants of some associated spaces, such as Stiefel--Whitney characteristic classes or the Fadell--Husseini index. We first show a new proof that uses a simple homotopy argument. This proof method also allows us to prove some new results, such as the following theorem. \begin{theorem}\label{thm:two-hyperplanes} Let $\mu_1, \ldots, \mu_{d}$ be $d$ probability measures in $\mathds{R}^d$. Then, we can find two hyperplanes $H_1, H_2$ such that \begin{itemize} \item $H_1 \cup H_2$ splits $\mu_1$ into four equal parts, \item each half-space $H$ that contains $H_1 \cap H_2$ satisfies $\mu_i(H) \ge \frac{1}{3}$ for $i=2,3,\ldots, d-1$, and \item $H_1$ splits $\mu_d$ into two equal parts. \end{itemize} \end{theorem} If we project onto the orthogonal complement of $H_1 \cap H_2$, the first equipartition is a Yao--Yao partition. The subspace $H_1 \cap H_2$ is also a central transversal for the next $d-1$ measures as in Theorem \ref{thm:ctt}. We describe similar results, such as Theorem \ref{thm:yao-transversal}, that interpolate between higher-dimensional Yao--Yao partitions and the central transversal theorem. \subsection{Structure of the paper} In Section \ref{sect:toolsnotation}, we begin by introducing notation that we use throughout the paper regarding Yao--Yao partitions and general properties of such partitions. Within this section, we also provide an overview of the proof technique used for the Yao--Yao partition theorem. We use these tools in Section \ref{sect:proofgeneralyao} to prove Theorem \ref{thm:generalyao}. In Section \ref{sect:linetransversal}, we prove our bounds for Problem \ref{problem:smallestkforn}. Section \ref{sect:stiefelyao} introduces results that connect Stiefel manifolds and Yao--Yao partitions. Lastly, we present remarks and open problems in Section \ref{sect:remarks}. \section{Tools and notation}\label{sect:toolsnotation} Let $\mu$ be a measure in $\mathds{R}^d$. We say that $\mu$ is \textit{absolutely continuous} if it is absolutely continuous with respect to the Lebesgue measure in $\mathds{R}^d$ and $\mu(A) \neq 0$ for every open $A$ subset of $\mathds{R}^d$. We say that $\mu$ is \textit{finite} if $\mu(\mathds{R}^d) < \infty$. Before we construct the convex partitions of $\mathds{R}^d$ for our main results, we define Yao--Yao partitions and provide an overview of the original proof, which contains tools we use to prove Theorem \ref{thm:generalyao}. A \textit{Yao--Yao partition} is a convex partition of a $\mathds{R}^d$ into $2^{d}$ regions, defined by a recursive process. A key property of these partitions is that every hyperplane misses the interior of at least one region. Yao--Yao partitions are determined by the choice of an ordered orthonormal basis $(u_1, u_2, \ldots, u_d)$. Each Yao--Yao partition $P$ has a point, denoted $C(P)$, which we call the \textit{center of the partition}. \begin{definition} Let $(u_1, u_2, \ldots, u_d)$ be an orthonormal basis of $\mathds{R}^d$. We say a hyperplane is \textit{horizontal} if it is orthogonal to $u_d$. For a horizontal hyperplane $H$, we define the open half-spaces induced by $H$ as $H_{+} = \{x + tu_{d} \mid x \in H, t > 0\}$ and $H_{-} = \{x + tu_{d} \mid x \in H, t < 0\}$. Similarly, for a measure $\mu$ we define $\mu_{+}$ as $\mu$ restricted to $H_{+}$ and $\mu_{-}$ as $\mu$ restricted to $H_{-}$. For a vector $v = (v_1, v_2, \ldots, v_d) \in S^{d-1}$ such that $v_d \neq 0$, the \textit{projection onto $H$ in the direction of $v$} is a mapping from $\mathds{R}^d$ to $H$ denoted by $p_{v}(\cdot)$. Denote the projection of $\mu_+$ and $\mu_-$ onto $H$ by $p_v(\mu_+)$ and $p_v(\mu_-)$, respectively. \end{definition} Let $p$ be a point and $(u_1, u_2, \ldots, u_d)$ be an orthonormal basis of $\mathds{R}^d$. A Yao--Yao partition $P$ of $\mathds{R}^d$ induced by the orthonormal basis $(u_1, u_2, \ldots, u_d)$ with center $p$ is a partition obtained in the following way. For $d=1$, it is the partition of $\mathds{R}^1$ into two infinite rays in opposite directions starting from $p$. For $d>1$, we take $H$ to be the horizontal hyperplane through $p$. Now, we construct two Yao--Yao partitions $P_+$ and $P_-$ of $H$ induced by the orthonormal basis $(u_1, \ldots, u_{d-1})$, each with center $p$ and a vector $v=(v_1, \ldots, v_d)$ such that $v_d \neq 0$. We take $P=\{p_v^{-1}({C})\cap H_+ : C \in P_+\}\cup \{p_v^{-1}({C})\cap H_- : C \in P_-\}$ as our final partition. Given the center and the orthonormal basis, a Yao--Yao partition is determined by a binary tree of height $d-1$ of projection vectors. Yao and Yao proved the following theorem. \begin{theorem}[Yao, Yao 1985] Let $d$ be a positive integer, $\mu$ be a finite absolutely continuous measure in $\mathds{R}^d$, and $(u_1, \ldots, u_d)$ be an orthonormal basis of $\mathds{R}^d$. There exists a unique Yao--Yao partition $P$ of $\mathds{R}^d$ induced by $(u_1, \ldots, u_d)$ such that \[ \mu(C) = \frac{\mu(\mathds{R}^d)}{2^d} \qquad \mbox{ for every }C \in P. \] \end{theorem} If the orthonormal basis is fixed, then the center depends only on the measure, so we denote it as $C(\mu)$. We call this partition the Yao--Yao equipartition of $\mu$ or the Yao--Yao partition of $\mu$ if there is no risk of confusion. Each cell of the partition is a cone with apex $C(\mu)$. In dimension $d=1$, the center is the point that bisects the measure. For dimensions $d \geq 2$, we bisect the measure with a horizontal hyperplane $H$ and find a projection vector $v$ such that $C(p_{v}(\mu_{+})) = C(p_{v}(\mu_{-}))$. Yao and Yao proved that there is a unique projection vector $v$ for which this happens. Moreover, as $\mu$ varies continuously, so does $C(\mu)$. The reason why $C(p_{v}(\mu_{+}))$ and $C(p_{v}(\mu_{-}))$ coincide for some projection vector is that otherwise we can construct a map from $B^{d-1}$, the $(d-1)$-dimensional unit ball, thought of as the set of points in $S^{d-1}$ with non-negative last coordinate to $S^{d-2}$ as \begin{align*} f:B^{d-1} & \to S^{d-2} \\ v &\mapsto \begin{cases} \frac{C(p_{v}(\mu_{+})) - C(p_{v}(\mu_{-}))}{||C(p_{v}(\mu_{+})) - C(p_{v}(\mu_{-}))||} & \mbox{if } v_d >0, \\ -v & \mbox{if } v_d = 0. \end{cases} \end{align*} Yao and Yao showed that this map is well defined and continuous if the top and bottom centers never coincide. This is a contradiction since it would be imply that $f|_{S^{d-2}}$ has degree zero instead of $(-1)^{d-1}$ (Yao and Yao instead finish the proof using the Borsuk--Ulam theorem). Additional geometric arguments show that there is only one projection vector up to scalar multiples that makes the top and bottom center coincide. To be precise, if the last coordinate of $v$ is positive and we choose a new projection vector $v' = v + h$ where $h$ is parallel to $H$ and not zero, then \begin{align*} \langle C(p_{v'}(\mu_{+})), h \rangle < \langle C(p_{v}(\mu_{+})), h \rangle \\ \langle C(p_{v'}(\mu_{-})), h \rangle > \langle C(p_{v}(\mu_{-})), h \rangle \end{align*} where $\langle \cdot, \cdot \rangle$ denotes the standard dot product. In other words, as the projection direction changes, the centers move in opposite directions. The first family of partitions we consider are based on Yao--Yao partitions. \begin{definition} Let $\mu$ be an absolutely continuous finite measure in $\mathds{R}^d$ and $(u_1, \ldots, u_d)$ be an orthonormal basis of $\mathds{R}^d$. For $\alpha, \beta > 0$ such that $\alpha + \beta = 1$ we define recursively an $(\alpha,\beta)$-partition of $\mu$ with $(\alpha, \beta)$-center in the following way. In dimension $d=1$, the $(\alpha, \beta)$-center $C(\mu)$ is the point such that the regions in the directions of $-u_1$ and $u_1$ have measure $\alpha(\mu(\mathds{R}))$ and $\beta(\mu(\mathds{R}))$, respectively. For dimensions $d \geq 2$, we first halve the measure with a horizontal hyperplane $H$ and find a projection vector $v$ such that $C(p_{v}(\mu_{+})) = C(p_{v}(\mu_{-}))$. In the previous equality, we are considering $C(p_{v}(\mu_{+}))$ the center of the $(\alpha, \beta)$-partition $P_+$ of $p_v(\mu_{+})$ with respect to the basis $(u_1, \ldots, u_{d-1})$ and $C(p_{v}(\mu_{-}))$ the center of the $(\alpha, \beta)$-partition $P_-$ of $p_v(\mu_{-})$ with respect to the basis $(u_1, \ldots, u_{d-1})$. Finally, we define the $(\alpha, \beta)$-partition as $P=\{p_v^{-1}({C})\cap H_+ : C \in P_+\}\cup \{p_v^{-1}({C})\cap H_- : C \in P_-\}$. \end{definition} Note that in the case $\alpha = \beta = \frac{1}{2}$, we have a Yao--Yao equipartition. The proof of existence and uniqueness of Yao--Yao partitions can be used verbatim to prove the existence and uniqueness of $(\alpha,\beta)$-partitions. Each $(\alpha, \beta)$-partition of a measure splits $\mathds{R}^d$ into $2^d$ convex parts. Given an $(\alpha, \beta)$-partition, every hyperplane misses the interior of at least one of the $2^d$ sections. The only change we make, relative to Yao and Yao's original construction, is in the first step of the construction. Each subsequent step uses a halving horizontal hyperplane. More general versions of these partitions, where the horizontal hyperplanes are not necessarily halving hyperplanes, were studied by Lehec \cite{lehec2009yao}. As we will show, $(\alpha, \beta)$-partitions have additional useful structural properties. For a finite absolutely continuous measure $\mu$ in $\mathds{R}^d$ and fixed $(\alpha, \beta)$, we introduce a lemma that describes the relationship between (classic) Yao--Yao equipartitions and $(\alpha, \beta)$-partitions of $\mu$. \begin{lemma}\label{lemma:structure1} Let $\alpha, \beta$ be positive real numbers whose sum is $1$ and let $(u_1, \ldots, u_d)$ be an orthonormal basis of $\mathds{R}^d$. Given a finite absolutely continuous measure $\mu$ in $\mathds{R}^d$, consider the $(\alpha, \beta)$-partition of $\mu$ induced by the basis $u_1, \ldots, u_d$. Then, we can split the regions of the partition into $2^{d-1}$ pairs $(A_i, B_i)$ so that \begin{itemize} \item $A_i \cup B_i$ is convex for each $i$, \item each region $A_i$ contains an infinite ray in the direction of $-u_1$, and \item each region $B_i$ contains an infinite ray in the direction of $u_1$. \end{itemize} Additionally, for each $i=1,\ldots,2^{d-1}$ the regions satisfy $\displaystyle \mu(A_i) = \frac{\alpha}{2^{d-1}} \mu(\mathds{R}^d)$ and $\displaystyle \mu(B_i) = \frac{\beta}{2^{d-1}} \mu(\mathds{R}^d)$. \end{lemma} \begin{proof} We proceed by induction on the dimension $d$. Clearly, the claim is true for $d=1$. Now we assume that the statement is true for $d-1$ and want to show that it is true for $d$. Project $\mu$ onto the horizontal halving hyperplane $H$ in the direction of the associated projection vector $v$ of the $(\alpha, \beta)$-partition of $\mu$. The projected partitions of $p_{v}(\mu_{+})$ and $p_{v}(\mu_{-})$ are both $(\alpha, \beta)$-partitions. Therefore, we can find pairings of regions in $H_{+}$ and $H_{-}$ that satisfy our requirements. Taking the pre-image of these partitions does not change convexity and the same rays in the directions $\pm u_1$ are contained. Therefore, these pairings work for the original $(\alpha, \beta)$-partition. Since $\mu_{+}$ and $\mu_{-}$ each have measure $\displaystyle \frac{1}{2} \mu(\mathds{R}^{d})$, we know $\displaystyle \mu(A_i) = \frac{\alpha}{2^{d-1}} \cdot \frac{1}{2} \mu(\mathds{R}^d)$ and $\displaystyle \mu(B_i) = \frac{\beta}{2^{d-1}} \cdot \frac{1}{2} \mu(\mathds{R}^d)$ for each $i$. Thus, the claim is true for all $d$. \end{proof} Throughout the rest of the paper we will use $A_i$, $B_i$ to denote the regions as described above. For each $i$, let $C_i = A_i \cup B_i$. We call $C_1, \ldots, C_{2^{d-1}}$ the \textit{frame} of the partition. Notice that the frame of the partition is a convex equipartition of $\mu$ into $2^{d-1}$ parts. We now show that the frame of the partition does not depend on the values of $\alpha, \beta$, as it is defined by the projection of $\mu$ onto the orthogonal complement of $u_1$. \begin{lemma}\label{lemma:structure2} Let $\alpha, \beta$ be positive real numbers whose sum is $1$ and let $u_1, \ldots, u_d$ be an orthonormal basis of $\mathds{R}^d$ for $d \geq 2$. Given a measure $\mu$ in $\mathds{R}^d$, let $A_1, \ldots, A_{2^{d-1}}, \\ B_1, \ldots, B_{2^{d-1}}$ be the $(\alpha, \beta)$-partition of $\mu$ induced by the basis $u_1, \ldots, u_d$. Let $V$ be the hyperplane orthogonal to $u_1$. Let $\pi: \mathds{R}^d \to V$ be the orthogonal projection onto $V$. We denote for each $1 \le i \le 2^{d-1}$ the sets $C_i = A_i \cup B_i$ and $D_i = \pi (C_i)$. Then, $D_1, \ldots, D_{2^{d-1}}$ is the Yao--Yao equipartition of $\pi (\mu)$ on $V$ induced by the basis $u_2, \ldots, u_{d}$. \end{lemma} \begin{proof} We proceed by induction on $d$. Clearly, the claim is true for $d=2$. Now we assume that the statement is true for $d-1$ and want to show that it is true for $d$. Let $R \subset V$ be the affine subspace of dimension $d-2$ that halves $\pi (\mu)$, notice that $R = \pi(H)$. We prove this lemma by using the equivalence of two projections onto $R$. Let $v$ be the associated projection vector of the $(\alpha, \beta)$-partition. By the inductive hypothesis, we know that the frame of the $(\alpha,\beta)$-partition of $p_v(\mu_+)$ projects onto the Yao--Yao partition of $\pi(p_v(\mu_+))$. Similarly, the frame of the $(\alpha,\beta)$-partition of $p_v(\mu_-)$ projects onto the Yao--Yao partition of $\pi(p_v(\mu_-))$. Therefore, the projection of the frame of the $(\alpha, \beta)$-partition of $\mu$ projects onto a Yao--Yao partition of $\pi(\mu)$ whose projection vector is $\pi(v)$. This is simply because $\pi \circ p_v = p_{\pi(v)} \circ \pi $. The resulting partition of $\pi(\mu)$ is clearly an equipartition, so we obtain the desired result. \end{proof} Now we show that the frame of a partition also remains constant if we restrict the measure to the $A_i$'s or to the $B_i$'s, which will be useful for our definition of multicenter partitions. \begin{corollary}\label{cor:restriction} With the notation of Lemma \ref{lemma:structure2} define $\displaystyle A = \bigcup_{i=1}^{2^{d-1}} A_i$ and $ \mu|_{A}$ the restriction of $\mu$ to $A$. Then, $D_1, \ldots, D_{2^{d-1}}$ is the Yao--Yao equipartition of $\pi(\mu|_{A})$ on $V$ induced by the basis $u_2, \ldots, u_{d}$. The lemma also holds if we replace $A$ by $\displaystyle B = \bigcup_{i=1}^{2^{d-1}} B_i$. \end{corollary} \begin{proof} By Lemma \ref{lemma:structure2}, we know that $D_1, \ldots, D_{2^{d-1}}$ is a Yao--Yao partition, and Lemma \ref{lemma:structure1} shows that $\mu(A_i) = \mu|_{A}(C_i)=\pi(\mu|_A)(D_i)$ has the same value for each $i$. As $\pi(\mu|_A)$ is a finite absolutely continuous measure in $V$, its Yao--Yao equipartition induced by $(u_2, \ldots, u_d)$ is unique. The proof is analogous for $B$ and $\mu|_B$. \end{proof} Given an $(\alpha, \beta)$-partition, each $C_i$ is split into two convex regions $A_i$ and $B_i$. The boundary between $A_i$ and $B_i$ must therefore be a hyperplane section. We denote the union of all these boundary pieces the \textit{wings} of the partition. As we vary $\alpha, \beta$, the frame remains constant while the center and the wings of the partition can change. Another consequence of the lemmas above is that, for a fixed absolutely continuous finite measure $\mu$ in $\mathds{R}^d$, there exists a line $\ell$ parallel to $u_1$ such that \begin{itemize} \item every $(\alpha, \beta)$-partition of $\mu$ has its center on $\ell$, \item the boundary of every part $A_i$ of an $(\alpha, \beta)$-partition of $\mu$ contains the infinite ray $\{c - \lambda u_1 : \lambda \ge 0\}$ where $c$ is the center of the $(\alpha, \beta)$-partition, and \item the boundary of every part $B_i$ of an $(\alpha, \beta)$-partition of $\mu$ contains the infinite ray $\{c + \lambda u_1 : \lambda \ge 0\}$ where $c$ is the center of the $(\alpha, \beta)$-partition. \end{itemize} We know by the arguments of Yao and Yao that every hyperplane $H$ must avoid the interior of at least one of the regions. Let left and right be the directions defined by $-u_1$ and $u_1$, respectively. If $H\cap \ell$ is a single point $p$, then it must avoid one of the regions $B_i$ if $p$ is at or left of the center $c$ and one of the regions $A_i$ if $p$ is at or right of the center $c$. The previous results can be improved to give us significant information about the $k$-skeletons of the cones forming and $(\alpha, \beta)$-partition. \begin{lemma}\label{lem:skeleton} Let $k < d$ be positive integers, $(u_1, \ldots, u_d)$ be an orthonormal basis of $\mathds{R}^d$ and $\mu$ be a finite absolutely continuous measure. For any $(\alpha, \beta)$-partition of $\mu$ induced by the basis $(u_1, \ldots, u_d)$, the union of the $k$-skeletons of the parts of the partition contains the translate of $\operatorname{span}\{u_1,\ldots, u_k\}$ that goes through the center of the partition. \end{lemma} \begin{proof} We proceed by induction on $k$. For $k=1$, the previous results imply the result. Now assume that $k>1$ and we know the result holds for $k-1$ and any $d \ge k$. With the notation of the previous lemmas, the frame of the $(\alpha, \beta)$-partition $P$ projects onto $V$ to the Yao--Yao partition $P'$ of $\pi (\mu)$ with basis $(u_2,\ldots, u_d)$. We know by induction that the union of the $(k-1)$-skeletons of the parts of $P'$ contain the translate of $\operatorname{span}\{u_2,\ldots, u_k\}$. When we take the inverse of $\pi(\cdot)$ we are extending these boundaries by $\pm u_1$ (one direction with the parts $A_i$ and the other with the parts $B_i$), which finishes the proof. \end{proof} If we consider the lemma above with $k=d-1$, we obtain the halving hyperplane at the core of the Yao--Yao construction. This way one can also argue this lemma by reducing the value of $k$ rather than increasing it. Since the location of the center $p$ on the line $\ell$ is important, we prove an additional technical lemma to improve Corollary \ref{cor:restriction}. We want to prove that a center for an $(\alpha',\beta')$-partition of $\mu|_{B}$ is further right than a center for $\mu$. \begin{lemma}\label{lem:technical-position} Let $d$ be a positive integer, $(u_1, \ldots, u_d)$ be an orthonormal basis of $\mathds{R}^d$, and $\alpha, \alpha', \beta, \beta'$ be positive real numbers such that $\alpha + \beta = \alpha' + \beta' = 1$. For an absolutely continuous finite measure $\mu$ in $\mathds{R}^d$, let $C(\mu)$ be the center of its $(\alpha, \beta)$-partition induced by the basis $(u_1, \ldots, u_d)$, and let $C(\mu|_{B})$ be the center of the $(\alpha',\beta')$-partition of $\mu|_B$. Then, $C(\mu|_{B})$ is farther right than $C(\mu)$ \end{lemma} \begin{proof} We proceed by induction on $d$. For $d=1$, the result is clear. Assume the result holds for $d-1$. Let $v$ be the projection vector of the $(\alpha, \beta)$-partition of $\mu$, and let $v'$ be the projection vector of the $(\alpha', \beta')$-partition of $\mu|_B$. We know by the previous lemmas that we can choose scalar multiples of $v$ and $v'$ so that $\pi(v) = \pi(v')$, as the frame of both partitions coincide. Consider the measures $p_v(\mu_+|_{B})$ and $p_v(\mu_-|_{B})$ on the halving hyperplane $H$. By the previous lemmas, their $(\alpha',\beta')$-partition shares the frame of $p_v(\mu)$, and their centers $C_1, C_2$ respectively must be to the right of $C(p_v(\mu))=C(\mu)$. If $C_1 = C_2$, this means that $v=v'$ and we are done. If $C_1 \neq C_2$, as we move $v$ to $v'$ the centers $C_1$ and $C_2$ move in opposite directions according to $u_1$ (formally, one of $\langle C_1, u_1\rangle, \langle C_2, u_1\rangle$ increases and the other decreases. This means that $C(\mu|_{B})$ must be between $C_1$ and $C_2$, and so it is to the right of $C(\mu)$. \end{proof} \begin{definition}\label{def:multicenter} We define a new family of partitions, called \textit{multicenter partitions} with $t$ centers for a measure $\mu$ in $\mathds{R}^d$ in the following way. \begin{itemize} \item For $t=1$, we take a classic Yao--Yao equipartition of $\mu$. \item For $t>1$, let $C_1, \ldots, C_{2^{d-1}}$ be the frame of all possible $(\alpha, \beta)$-partitions of $\mu$. We first construct a $(1/(t+1), t/(t+1))$-partition of $\mu$ and denote its regions by $A_1, \ldots, A_{2^{d-1}}, B_1, \ldots, B_{2^{d-1}}$. Let $P$ be a multicenter partition of $\mu|_{B}$ with $t-1$ centers. Our multicenter partition $Q$ with $t$ centers is defined as \[ Q = \{A_i : i =1,\ldots, 2^{d-1}\}\cup \{K \cap B_i : i = 1,\ldots, 2^{d-1}, K \subset C_i, K \in P\}. \] \end{itemize} We define the $t$ centers as the union of the center of the $(1/(t+1), t/(t+1))$-partition of $\mu$ and the $t-1$ centers of the multicenter partition $P$ we used in the construction. \end{definition} \begin{figure}[ht] \centering \includegraphics[scale=0.75]{Multicenter_Partition_R3.pdf} \caption{Example of a multicenter partition in $\mathds{R}^3$ with two centers.} \label{fig:R3convexpartition} \end{figure} Each part of a multicenter partition is contained in one of the parts of the frame of $\mu$. Each region of the frame is iteratively partitioned into $t+1$ convex regions by the wings of each $(\alpha, \beta)$-partition we took. A multicenter partition with $t$ centers will have exactly $(t+1)2^{d-1}$ convex parts of equal measure. An example is shown in Figure \ref{fig:R3convexpartition}. Notice that the subdivision of the right-most center doesn't extend past the wings induced by the first partition (induced by the left-most center). Now we can prove Theorem \ref{thm:generalyao}. \section{Proof of Theorem \ref{thm:generalyao}}\label{sect:proofgeneralyao} We want to show that for a multicenter partition with $t$ centers every hyperplane misses the interior of at least $t$ regions. Since a multicenter partition with $t$ centers has $(t+1)2^{d-1}$ regions, all of them convex, we would prove Theorem \ref{thm:generalyao}. In this section, we use a fixed orthonormal basis $(u_1, u_2, \ldots, u_d)$ and induct on $t$. \begin{theorem} Let $\mu$ be a finite absolutely continuous measure in $\mathds{R}^d$, and let $P$ be a multicenter partition of $\mu$ with $t$ centers. Every hyperplane misses the interior of at least $t$ regions. \end{theorem} \begin{proof} We proceed inductively. For $t=1$, this is the result of Yao and Yao. Assume the result holds for $t-1$. Let $Q$ be the multicenter partition, with parts labeled as in Definition \ref{def:multicenter}. Let $C(\mu)$ be the center of the $(1/(t+1),t/(t+1))$-partition of $\mu$, and let $\ell$ be the line with direction $u_1$ through $C(\mu)$. Let $L$ be a hyperplane. We first assume that $L \cap \ell$ is a single point. If $L \cap \ell$ is to the right or coincides with $C(\mu)$, then $L$ misses the interior of one of the regions $A_i$. By the induction hypothesis, $L$ misses the interior of at least $t-1$ regions of the form $B_{i'} \cap K$ for some $i'$. If $L \cap \ell$ is to the left of $C(\mu)$, then $L$ avoids the interior of some set $B_i$. This means that $L$ avoids all the $t$ parts of the partition of the form $B_i \cap K$. Any hyperplane $L$ for which $L \cap \ell$ is not a point contains a line parallel to $\ell$. Therefore, $\pi(L)$, the projection of $L$ onto the subspace orthogonal to $u_1$ is not surjective. Since the projection of the frame $C_1, \ldots, C_{2^{d-1}}$ is a Yao--Yao partition, this means that $L$ misses the interior of one of the $C_i$. With this, we have that $L$ misses the interior of the $t+1$ regions in the subdivision of that $C_i$. \end{proof} The proof above does not use Lemma \ref{lem:technical-position}. However, that lemma can help us get intuition regarding which regions of a multicenter partition are avoided. If $L \cap \ell$ is between the $j$-th and the $(j+1)$-th center of the partition, we can guarantee $L$ avoids $j$ ``left regions $A_i$'' in the recursive definition and $t-j$ right regions which at some point were of the form $B_i \cap K$ in the recursive definition. The union of the right regions avoided is convex. An example is shown in Figure \ref{sect:linetransversal}. In that figure, one of the avoided regions is shaded. We can see that the section right of the shaded region is also missed by the hyperplane, and their union is convex. \begin{figure}[ht] \centering \includegraphics[scale=0.75]{Avoidance_Multicenter_Partition_R3.pdf} \caption{Hyperplane transversal for a multicenter partition in $\mathds{R}^3$} \label{fig:R3avoidance} \end{figure} \section{Line transversal problem}\label{sect:linetransversal} We first introduce the notion of complexity for a partition. The \textit{complexity} of a convex partition refers to the smallest number of hyperplanes whose union contains all the boundaries between parts of the partition. If $P$ is a partition into $n$ parts with complexity $k$ then every line misses the interior of at least $n - (k+1)$ regions, so we care about minimizing the complexity to maximize the number of parts missed by any line. The partitions we use involve iterated partitions by successive hyperplanes. These are similar to those used recently for high-dimensional versions of the necklace splitting problem \cites{deLongueville:2006uo, Karasev:2016cn, Blagojevic:2018gt}. The key difference is that we wish each part of the partition to have the same size, as opposed to distribute the parts among a fixed number of participants. To obtain a general bound for the minimum complexity needed to find an equipartition of any measure, we use the ham sandwich theorem. This was first proved by Steinhaus, who attributed the result to Banach \cite{Steinhaus1938}. \begin{theorem}[Ham sandwich theorem]\label{thm:hamsandwich} Given $d$ finite measures in $\mathds{R}^d$, each absolutely continuous with respect to the Lebesgue measure, there exists a hyperplane that divides $\mathds{R}^d$ into two half-spaces of the same size with respect to each measure. \end{theorem} With this tool, we can introduce a lemma that allows us to save many hyperplanes as the dimension increases. \begin{lemma}\label{lem:d2xlemma} Let $d$ and $x$ be positive integers. For any absolutely continuous measure $\mu$ in $\mathds{R}^d$ there exists a convex equipartition into $d2^{x}$ sets of complexity $(d-1) + 2^{x} - 1$. \end{lemma} \begin{proof} First, we partition the measure into convex regions of equal measure using $d-1$ parallel hyperplanes. Now we can treat the $d$ regions as distinct measures. By the ham sandwich theorem, we can bisect all the regions using a hyperplane. We repeat this process recursively for the regions on the positive side of the bisecting hyperplane and for the negative side of the bisecting plane until we have regions of measure $\mu(\mathds{R}^d)/d2^{x}$. The total number of hyperplanes used is $2^x - 1$ aside from the initial $d-1$ hyperplanes. Hence, we use a total of $(d-1) + 2^{x} - 1$ to partition the measure into $d2^{x}$ regions of equal measure. \end{proof} Because we are using a single hyperplane to bisect multiple regions at once, we can save many hyperplanes and obtain the following bound. We assume we have a fixed orthonormal basis $\{u_1, u_2, \ldots, u_d\}$. We call hyperplane vertical if it is orthogonal to $u_1$. We are now ready to prove Theorem \ref{thm:smallestkforn}. \begin{proof}[Proof of Theorem \ref{thm:smallestkforn}] We first prove the upper bound. We can recursively define how we partition $\mu$ into $n$ regions of equal measure in the following way. Any positive integer $n$ can be expressed in the form $d2^{\alpha_1} + d2^{\alpha_2} + d2^{\alpha_3} + \ldots + \epsilon$ such that $\alpha_{j} > \alpha_{k}$ for $j < k$ and $\epsilon < d$. We can use a vertical hyperplane to partition $\mu$ such that the left side of the hyperplane has measure $(\mu(\mathds{R}^d)d2^{\alpha_1})/n$. Using another vertical hyperplane to partition the right region we repeat this process so that the section between the two hyperplanes has measure $(\mu(\mathds{R}^d)d2^{\alpha_2})/n$ and so on for each $\alpha_{i}$. We repeat this for at most ${\alpha_1} \leq \log_{2}(n/d)$ terms. Once the right-most side of the partition has measure less than $(\mu(\mathds{R}^d)d)/n$, we use $\epsilon - 1$ parallel hyperplanes to partition the region into convex regions of equal measure. In each region with measure $(\mu(\mathds{R}^d)d2^{\alpha_i})/n$, we can apply Lemma \ref{lem:d2xlemma} to partition that region into $d2^{\alpha_i}$ regions of equal measure using $(d-1) + 2^{\alpha_i} - 1$ hyperplanes. Therefore, we use at most $k \leq (d-1)\log_{2}(n/d) + (n/d - \epsilon/d) + \epsilon - 1$ hyperplanes and consequently $k = n/d + O(\log(n))$. We can use a measure concentrated around the moment curve to obtain a lower bound for the complexity of any convex partition of $\mathds{R}^d$. A single hyperplane can intersect a moment curve at no more than $d$ points and contribute at most $d+1$ convex regions in which each piece of the measure can lie. Each subsequent hyperplane contributes at most an additional $d$ regions. This implies that, for a convex equipartition of $\mu$ into $n$ regions with complexity $k$, we have $n \leq kd + 1$. Therefore the lower bound for the complexity is $(n-1)/d \leq k$. \end{proof} An intuitive way to look at the proof above is that every time we use use Lemma \ref{lem:d2xlemma}, we get $d-1$ hyperplanes above the $n/d$ ideal bound. This is done once for every $1$ in the binary representation of $\lfloor n/d \rfloor$. For $d=2$ and $d=3$ we can find much simpler partitions that yield the following sharper bound. \[ \left \lceil \frac{n-1}{d} \right \rceil \leq k \leq \frac{n}{d} + O(1) \] \smallskip The bounds largely solve the problem in those dimensions, as the term $O(1)$ is bounded by $1/2$ for $d=2$ and by $13/3$ for $d=3$. We prove this by using a generalization of the ham sandwich theorem for ``well-separated'' families proved by B\'ar\'any, Hubard, and Jer\'onimo \cite{barany2008slicing}. We say that a family $\mathcal{F}$ of subsets of $\mathds{R}^d$ is well-separated if for every $\mathcal{A}\subset\mathcal{F}$ we have that $\bigcup \mathcal{A}$ can be separated from $\bigcup (\mathcal{F}\setminus \mathcal{A})$ by a hyperplane. \begin{theorem}[Bárány, Hubard, Jerónimo 2008]\label{thm:bhj} Let $d$ be a positive integer. For all $i=1,\ldots, d$, let $\mu_{i}$ be a finite measure on $\mathds{R}^d$, absolutely continuous with respect to the Lebesgue measure, with support $K_{i}$ for all $i \in \{1, 2, \dots, d\}$. Assume the family $\mathcal{F} = \{K_{1},\dots,K_{d}\}$ is well-separated and let $\alpha = (\alpha_{1},\dots,\alpha_{d}) \in (0,1)^{d}$. Then there exists a half-space, $H$, such that $\mu_{i}(K_{i} \cap H) = \alpha_{i} \cdot \mu_{i}(K_{i})$, for every $i \in \{1, 2, \dots, d\}$. \end{theorem} B\'ar\'any et al also determined conditions that guarantee the uniqueness of the hyperplane above, but we do not require it for our proof. Using Theorem \ref{thm:bhj}, we introduce alternative partitions that yield better upper bounds than Theorem \ref{thm:smallestkforn} for $d=2$ and $d=3$. \subsection{d=2} Let $n = 2q + r$ for non-negative integers $q$ and $r < 2$. If $n < 2$, then $n=1$, so we use no hyperplanes. For $n \geq 2$, we first use $r$ vertical lines to partition the measure such that each of the $r$ left-most regions has measure $\mu(\mathds{R}^2)/n$. Then we bisect the rightmost region with a line. The two regions induced are well-separated, so we can apply Theorem \ref{thm:bhj} to the pair of regions. We can use $q-1$ lines to partition the pair such that the positive half-space of each line has measure $\mu(\mathds{R}^2)/q$ in each region of each pair. Therefore, for $n \geq 2$ the number of lines we use is \[ \frac{n}{2} + \frac{r}{2}. \] We refer to this partition as a B\'ar\'any, Hubard, and Jer\'onimo partition (see Figure \ref{fig:BHJdiagram}). Note that $r \leq 1$, so the complexity is bounded by $1/2$. The second partition we introduce does not use Theorem \ref{thm:bhj} but follows immediately from Theorem \ref{thm:generalyao}. If the number of regions $n$ is odd, we use a vertical line to partition the measure so the left side has measure $\mu(\mathds{R}^2)/n$ and then we take a multicenter equipartition for the right side of the line. In the case that $n$ is even, we can take a multicenter equipartition. The number of lines used in this partition is \[ k = \Bigl\lceil{\frac{n}{2}\Bigr\rceil}. \] We call these partitions \textit{modified multicenter partitions} (see Figure \ref{fig:multicenterR2diagram}). Note that this is the same bound proven by Rold\'an-Pensado and Sober\'on \cite{roldan2014extension}. The construction they provide is different because they use a rotating half-space to partition the measure (see Figure \ref{fig:RPSdiagram}). We refer to this partition as a Rold\'an-Pensado and Sober\'on partition. The differences between the partitions lies in their geometry. The modified multicenter partition iterates on the right regions, the Rold\'an-Pensado and Sober\'on partition uses rotating cuts, and the B\'ar\'any, Hubard, and Jer\'onimo partition preserves one cut in any direction and iteratively partitions the pair of regions. \begin{figure}[ht] \centering \begin{subfigure}[b]{0.3\textwidth} \centering \includegraphics[width=\textwidth, height=2.5cm]{MulticenterR2_Diagram.pdf} \caption{Modified multicenter partition in $\mathds{R}^2$} \label{fig:multicenterR2diagram} \end{subfigure} \hfill \begin{subfigure}[b]{0.3\textwidth} \centering \includegraphics[width=\textwidth, height=2.5cm]{RPS_Diagram.pdf} \caption{Rold\'an-Pensado and Sober\'on partition} \label{fig:RPSdiagram} \end{subfigure} \hfill \begin{subfigure}[b]{0.3\textwidth} \centering \includegraphics[width=\textwidth, height=2.5cm]{BHJ_Diagram.pdf} \caption{B\'ar\'any, Hubard, and Jer\'onimo partition} \label{fig:BHJdiagram} \end{subfigure} \caption{Each type of partition has its own geometry} \label{fig:d2diagrams} \end{figure} \subsection{d=3} In the case of $d=3$, there are two other constructions that use Theorem \ref{thm:bhj} to show the same bound. Let $n = 6q + r$ for non-negative integers $q$ and $r < 6$. If $r > 0$, we use $r-1$ we first partition the measure using $r$ vertical planes so that each of the $r$ left-most regions has measure $\mu(\mathds{R}^3)/n$. For the first approach, we use a partition result of Buck and Buck, that says that \textit{any finite measure in the plane can be split into six equal parts using three concurrent lines} \cite{buck1949equipartition}. Therefore, using a projection we can see that we can partition the rightmost region with three planes that share a common line of intersection such that each of the six resulting regions has equal measure. Pairing every other region into a group of three gives two triplets, each consisting of well-separated regions. Therefore, by Theorem \ref{thm:bhj} we can iteratively use $q-1$ planes to partition each triplet such that the resulting regions have measure exactly $\mu(\mathds{R}^3)/q$. In the second approach, we use two parallel planes to partition the right region into three regions of equal measure and then by the ham sandwich theorem we bisect all three regions using an additional plane. We group together the leftmost and rightmost region on the positive side of the bisecting plane and the center region on the negative side. The other triplet is formed by the remaining regions. In both cases for $n\geq6$, the number of planes we use is \[ \frac{n}{3} + \frac{2r}{3} + 1. \] \smallskip Note that $r \leq 5$, so the complexity is bounded by $(n+13)/3$. Therefore, tools such as Theorem \ref{thm:bhj} allow us to find partitions with smaller complexity and different geometric properties. \section{Stiefel manifolds and Yao--Yao partitions}\label{sect:stiefelyao} The Borsuk--Ulam theorem is the topological backbone of many mass partition results. The Yao--Yao theorem is one of such result, as the inductive step that allows us to find Yao--Yao partitions relies on the Borsuk--Ulam theorem. As one explores more elaborate mass partition theorems, we may either require more advanced topological machinery or tailor-made topological results similar to the Borsuk-Ulam theorem. In this section we first present a new proof a Borsuk-Ulam type theorem from Chan et al \cite{chan2020borsuk} where the domain is a Stiefel manifold of orthonormal $k$-frames in $\mathds{R}^d$ \[ V_k(\mathds{R}^d) = \{(v_1, \ldots, v_k): v_i \in S^{d-1} \mbox{ for }i=1,\ldots, k, \ \langle v_i, v_j\rangle = 0 \ \mbox{ for } i \neq j\}. \] The Stiefel manifold $V_k(\mathds{R}^d)$ is a manifold of dimension $(d-1) + (d-2) + \ldots + (d-k)$ with a free action of $(\mathds{Z}_2)^k = \{+1,-1\}^k$. Given an element $\lambda = (\lambda_1, \ldots, \lambda_k) \in (\mathds{Z}_2)^k$ and $v=(v_1, \ldots, v_k) \in V_k(\mathds{R}^d)$, we define \[ \lambda v := (\lambda_1 v_1, \ldots, \lambda_k v_k) \in V_k(\mathds{R}^d). \] We can also define an action of $(\mathds{Z}_2)^k$ on $R=\mathds{R}^{d-1} \times \mathds{R}^{d-2} \times \ldots \times \mathds{R}^{d-k}$ as the direct product of the action of $\mathds{Z}_2$ on each component. The only fixed point in $R$ is the zero vector. \begin{theorem}[Chan, Chen, Frick, Hull 2020 \cite{chan2020borsuk}]\label{thm:stiefel-BU} Let $k, d$ be positive integers such that $k \le d$. Let $f:V_k(\mathds{R}^d) \to \mathds{R}^{d-1} \times \mathds{R}^{d-2} \times \ldots \times \mathds{R}^{d-k}$ be a continuous and equivariant function with respect to the action of $(\mathds{Z}_2)^k$ on each space as defined above. Then, there exists $v \in V_k(\mathds{R}^d)$ such that $f(v) = 0$. \end{theorem} The case $k=1$ is one of the many equivalent forms of the Borsuk-Ulam theorem: \textit{every continuous odd map $f:S^{d-1}\to \mathds{R}^{d-1}$ has a zero.} The case $k=d-1$ is essentially a Borsuk-Ulam type theorem in which the domain is $O(d)$. The original proof of this theorem involves the computation a topological invariant constructed by the sum of the degrees of some associated continuous maps. It can also be proved using the Fadell--Husseini index (this proof is implicit in Fadell and Husseini's foundational paper \cite{Fadell:1988tm}). Similar results, where the domain is $O(d)$ or $SO(d)$, have been proven earlier, although they use slightly different group actions \cites{roldan2014extension, FHMRA19}. Since the dimension of the domain and the range is the same, the theorem above is also a consequence of Musin's Borsuk--Ulam type theorems for manifolds \cite{Mus12}*{Theorem 1}. We present a simple proof below for completeness. The proof below, as well as Musin's proof for this main theorem and the proofs for the results mentioned for $SO(d)$ all follow the scheme of Imre B\'ar\'any's proof of the Borsuk--Ulam theorem \cites{barany1980borsuk, matousek2003using}. \begin{proof} First, we construct a particular map $g:V_k(\mathds{R}^d) \to R$ that is continuous and equivariant. We denote the coordinates of each $v_i$ by $v_i = ((v_{i})_1, \ldots, (v_{i})_d)$. The function $g$ is defined by \begin{align*} g:V_k(\mathds{R}^d) &\to \mathds{R}^{d-1}\times \mathds{R}^{d-2} \times \ldots \times \mathds{R}^{d-k} \\ (v_1, \ldots, v_k) &\mapsto (x_1, \ldots, x_k) \end{align*} where \[ x_j = \begin{bmatrix} (v_j)_{j+1} \\ (v_j)_{j+2} \\ \vdots \\ (v_j)_{d} \end{bmatrix} \in \mathds{R}^{d-j}. \] A simple inductive argument shows that the only zeroes of this function are when for each $1\le j \le k$ we have $v_j \in \{e_j, -e_j\}$, where $(e_1,\ldots, e_d)$ is the canonical basis of $\mathds{R}^d$. In other words, there is a single $\left(\mathds{Z}_2 \right)^k$-orbit of zeros in $V_k(\mathds{R}^d)$, which is the set $\left(\mathds{Z}_2 \right)^k(e_1,\ldots, e_k)$. Now we construct a new map \begin{align*} T: V_k(\mathds{R}^d) \times [0,1] &\to \mathds{R}^{d-1}\times \mathds{R}^{d-2} \times \ldots \times \mathds{R}^{d-k} \\ (v,t) & \mapsto tf(v) + (1-t)g(v). \end{align*} Let $\varepsilon>0$ be a real number. There exists a smooth $(\mathds{Z}_2)^k$-equivariant map $T_{\varepsilon}: V_k(\mathds{R}^d) \times [0,1] \to \mathds{R}^{d-1}\times \mathds{R}^{d-2} \times \ldots \times \mathds{R}^{d-k}$ such that \begin{align*} T_{\varepsilon}(v,0) = T(v,0) = g(v) & \qquad \mbox{ for all }v\in V_k(\mathds{R}^d), \\ ||T_{\varepsilon}(v,t)-T(v,t)||< \varepsilon & \qquad \mbox{ for all }v\in V_k(\mathds{R}^d),\ t \in [0,1], \mbox{ and} \\ 0\mbox{ is a regular value of }T_{\varepsilon}. \end{align*} This follows from Thom's transversality theorem \citelist{\cite{thom1954quelques} \cite{guillemin2010differential}*{pp 68-69}}. Let us look at $T_{\varepsilon}^{-1}(0) \subset V_k(\mathds{R}^d)\times [0,1]$. This is a one-dimensional manifold with a free action of $\left( \mathds{Z}_2\right)^k$, whose connected components are diffeomorophic to circles or to intervals. The components with diffeomorphic to intervals must have their endpoints in $V_k(\mathds{R}^d) \times \{0,1\}$. As any continuous function from a closed interval to itself must have a fixed point, the group $(\mathds{Z}_2)^k$ acts freely on the set of intervals in $T^{-1}_{\varepsilon}(0)$. The set $T^{-1}_{\varepsilon}(0) \cap \left(V_k(\mathds{R}^d)\times \{0\}\right)$ has a single orbit of $\left( \mathds{Z}_2\right)^k$ by construction. Therefore, $T^{-1}_{\varepsilon}(0) \cap \left(V_k(\mathds{R}^d)\times \{1\}\right)$ must have an odd number of orbit of $\left( \mathds{Z}_2\right)^k$, which implies that $T^{-1}_{\varepsilon}(0) \cap \left(V_k(\mathds{R}^d)\times \{1\}\right)$ is not empty. As we make $\varepsilon \to 0$, the compactness of $V_k(\mathds{R}^d)$ implies that $f^{-1}(0)$ is not empty, as we wanted. \end{proof} A consequence of Theorem \ref{thm:stiefel-BU} is a generalization of the central transversal theorem mentioned in the introduction \cites{dol1992generalization, zivaljevic1990extension}. Given a finite probability measure $\mu$ in $\mathds{R}^d$ and an affine subspace $L$ of dimension $k$, we say that $L$ is a central $k$-transversal to $\mu$ if each half-space that contains $L$ has measure at least $1/(d-k+1)$ in $\mu$. The central transversal theorem says that \textit{for $0 \le \lambda \le d-1$, any $\lambda+1$ measures in $\mathds{R}^d$ have a common central $\lambda$-transversal}. The case $\lambda=d-1$ is the ham sandwich theorem and the case $\lambda=0$ is the centerpoint theorem. We obtain the following result. \begin{theorem}\label{thm:stiefel-transversal} Let $d$ be a positive integer, $\mu_1, \ldots, \mu_d$ be finite absolutely continuous measures in $\mathds{R}^d$, and $\lambda$ be an integer such that $0 \le \lambda \le d-1$. We can find affine subspaces $L_{\lambda} \subset L_{\lambda+1} \subset \ldots \subset L_{d-1}$ such that \begin{itemize} \item for each $\lambda \le i \le d-1$, the dimension of $L_i$ is $i$, \item the subspace $L_{\lambda}$ is a central $\lambda$-transversal to each of $\mu_1, \ldots, \mu_{\lambda+1}$, and \item for each $\lambda < i \le d-1$, the subspace $L_i$ is a central $i$-transversal to $\mu_{i+1}$. \end{itemize} \end{theorem} The second condition implies the central transversal theorem. The last condition can also be obtained from the central transversal theorem by a bootstrapping argument. We first find a common $\lambda$-transversal to the first $\lambda+1$ measures and then use the central transversal theorem to iteratively look for a direction to extend $L_n$ to $L_{n+1}$ for $\lambda \le n < d-1$. This version is slightly closer to the ham sandwich theorem than the central transversal theorem, as we always deal with $d$ measures. We use Theorem \ref{thm:stiefel-BU} to obtain a direct proof. \begin{proof} Let $k=d-\lambda$. We construct a function $f:V_k(\mathds{R}^d) \to R$ as in Theorem \ref{thm:stiefel-BU}. Let $(v_1, \ldots, v_k)$ be an element of $V_k(\mathds{R}^d)$. For $0 \le i \le \lambda=d-k$, let $\sigma_{i+1}$ be the orthogonal projection of $\mu_{i+1}$ onto $V_{k} = \operatorname{span}\{v_1, \ldots, v_k\}$. By the central point theorem, there is a centerpoint $p_{i+1}$ of $\sigma_{i+1}$ in $V_k$. The set of all possible centerpoints of $\sigma_i$ is a compact convex set, so we can choose $p_{i+1}$ to be the barycenter of this set. If $\Pi_k : \mathds{R}^d \to V_k$ is the orthogonal projection, the affine subspace $\Pi_k^{-1}(p_{i+1})$ is a central $\lambda$-transversal for $\mu_{i+1}$. For $\lambda < i \le d-1$, let $\sigma_{i+1}$ be the orthogonal projection of $\mu_{i+1}$ onto $V_{d-i}=\operatorname{span} \{ v_1, \ldots, v_{d-i}\}$. Let $p_{i+1}$ be the centerpoint of $\sigma_i$ in $V_{d-i}$ chosen as above. We denote by $\Pi_{d-i}: \mathds{R}^d \to V_{d-i}$ the orthogonal projection onto $V_{d-i}$. The affine subspace $\Pi^{-1}_{d-i}(p_{i+1})$ is a central $i$-transversal for $\mu_{i+1}$. Now we define \begin{align*} f:V_k(\mathds{R}^d) &\to \mathds{R}^{d-1} \times \mathds{R}^{d-2} \times \ldots \mathds{R}^{d-k} \\ (v_1, \ldots, v_k) &\mapsto (x_1, \ldots, x_k) \end{align*} where $x_j \in \mathds{R}^{d-j}$ is defined by \[ x_j = \begin{bmatrix} \langle v_j , p_2 - p_1 \rangle \\ \langle v_j , p_3 - p_1 \rangle \\ \vdots \\ \langle v_j , p_{d+1-j}-p_1\rangle \end{bmatrix}. \] This map is continuous and equivariant, so it must have a zero. If $(v_1, \ldots, v_k)$ is a zero of this map, let us show that the subspaces $L_i = \Pi_{d-i}^{-1}\left(\Pi_{d-i}(p_1)\right)$ for $\lambda \le i \le d-1$ satisfy the condition we want. We immediately get that the dimension of $L_i$ is equal to $i$ and that $L_\lambda \subset \ldots \subset L_{d-1}$. First, since $(v_1, \ldots, v_k)$ is an orthonormal frame and $p_{i+1} \in V_k$ for $i=0,\ldots, \lambda$, we know that $p_{i+1} = \sum_{j=1}^k \langle v_j, p_{i+1} \rangle v_j$. For $i \ge \lambda+1$, we have $p_{i+1} = \sum_{j=1}^{d-i}\langle v_j, p_{i+1} \rangle v_j$. If $x_j = 0$ for all $j$, this implies that $p_1 = p_i$ for $i=2,\ldots, \lambda+1$. Therefore, $L_{\lambda}$ is a central $\lambda$-transversal for each of $\mu_1, \ldots, \mu_{d+1}$. For $i>\lambda+1$ we have \[ \Pi_{d-i}(p_1) = \sum_{j=1}^{d-i}\langle v_j, p_1 \rangle v_j = \sum_{j=1}^{d-i}\langle v_j, p_{i+1} \rangle v_j = p_{i+1} \]. Therefore, $L_i$ is a central $i$-transversal to $\mu_{i+1}$. \end{proof} If we take $\lambda = 0$ we have the following corollary about full flags of subspaces. \begin{corollary}\label{cor:flag} Let $\mu_1, \ldots, \mu_d$ be finite measures in $\mathds{R}^d$. We can find affine subspaces $L_{0} \subset L_{1} \subset \ldots \subset L_{d-1}$ such that for each $0 \le i \le d-1$, the dimension of $L_i$ is $i$ and for each $0 \le i \le d-1$ the subspace $L_i$ is a central $i$-transversal to $\mu_{i+1}$. \end{corollary} The fact that our parametrization depends on $V_k(\mathds{R}^d)$ instead of a Grassman manifold means that in the proof above we may replace $p_i$ in the proof above by a point that depends continuously on the choice of $V_k(\mathds{R}^d)$, as long as the choice keeps our function equivariant. For example, we could take $p_i$ to be a Yao--Yao center of $\sigma_{i+1}$ on the corresponding subspace with the basis induced by $(v_1,\ldots, v_k)$. If we do this for several measures, even though the projections of the Yao--Yao centers will coincide we have no guarantee that the projections of the whole Yao--Yao partitions will coincide. We can now prove Theorem \ref{thm:two-hyperplanes}. \begin{proof}[Proof of Theorem \ref{thm:two-hyperplanes}] The proof is almost identical to the proof of Theorem \ref{thm:stiefel-transversal} with $k=2$ except for the definition of $p_1$. We define $p_1$ as the Yao--Yao center of $\Pi_2(\mu_1)$ induced by the basis $(v_2, v_1)$. Notice that a Yao--Yao partition in $\mathds{R}^2$ is an equipartition by two lines, one of which is parallel to $v_2$ if we choose the ordered basis as above. The inverse image under $\Pi_2$ of the two lines forming this Yao--Yao partition give us the hyperplanes we wanted. \end{proof} Let us compare Theorem \ref{thm:two-hyperplanes} with earlier results. For $d=3$ the central transversal theorem tells us that for any two measures we can find a common central line. A classic result of Hadwiger tells us that for any two measures in $\mathds{R}^3$ there are two planes that simultaneously split them into four equal parts \cite{Hadwiger1966} (a proof with new methods was recently found by Blagojevi\'c, Frick, Haase, and Ziegler \cite{Blagojevic:2018jc}*{Section 4}). The case $d=3$ of Theorem \ref{thm:two-hyperplanes} shows that for any two measure in $\mathds{R}^3$ we can find two planes that split the first measure into four equal parts and whose intersection is a central line for the second measure. We even have a degree of freedom since we can choose $\mu_3$ at will. Even the case $\mu_1 = \mu_2$ is non-trivial, as opposed to the previous two related results. If we are given more than $2d/3$ measures in $\mathds{R}^d$ it is possible that there is no pair of hyperplanes that split each of them into four equal parts \cite{Ramos:1996dm}. Therefore the corollary above is a sensible way of interpolating between central transversals and equipartitions by hyperplanes of many measures. It is not clear if we can make the equipartition part of Theorem \ref{thm:two-hyperplanes} be used for more measures. \begin{question} Is is true that for any $d-1$ measures $\mu_1, \ldots, \mu_{d-1}$ in $\mathds{R}^d$ there exist two hyperplanes $H_1, H_2$ such that \begin{itemize} \item $H_1 \cup H_2$ split each of $\mu_1, \mu_2$ into four equal parts and \item $H_1 \cap H_2$ is a central $(d-2)$ transversal to each of $\mu_3, \ldots, \mu_{d-1}$? \end{itemize} \end{question} We can also extend Corollary \ref{cor:flag} in a similar way but now using a full Yao--Yao partition. \begin{theorem}\label{thm:yao-transversal} Let $d$ be a positive integer and $\mu_1, \ldots, \mu_d$ be finite absolutely continuous measures in $\mathds{R}^d$. For an orthonormal basis $(u_1, \ldots, u_d)$ of $\mathds{R}^d$ let $C(\mu_1)$ be the center of the Yao--Yao partition of $\mu_1$ induced by the basis $(u_1, \ldots, u_d)$, and let $L_i$ be the translate of $\operatorname{span}\{u_1,\ldots, u_i\}$ through $C(\mu_1)$. Then, there exists a choice of an orthonormal basis such that for $i=1,\ldots,d-1$ we have that $L_i$ is a central $i$-transversal of $\mu_{i+1}$. \end{theorem} Due to Lemma \ref{lem:skeleton}, each $L_i$ is contained in the union of the $i$-skeletons of the parts of the Yao--Yao partition of $\mu_1$ we constructed, so these spaces appear naturally. \begin{proof} We follow verbatim the proof of Theorem \ref{thm:stiefel-transversal} with the only difference being that for each $v=(v_1, \ldots, v_d) \in V_d(\mathds{R}^d)$ we take $p_1$ to be the Yao--Yao center of $\mu_1$ according to the basis $(u_1, u_2, \ldots, u_d) = (v_d, v_{d-1}, \ldots, v_1)$. For $i=1,\ldots, d-1$, the point $p_{i+1}$ is still the centerpoint of the projection of $\mu_{i+1}$ onto $\operatorname{span}\{v_1,\ldots, v_{d-i}\}$. \end{proof} Just as the theorem above is closely related to Corollary \ref{cor:flag}, it is clear we can get an analogous extension of Theorem \ref{thm:stiefel-transversal} for any $\lambda$. We do this by choosing $p_1$ to be the Yao--Yao center of the projection of $\mu_1$ onto $\operatorname{span}\{v_1, \ldots, v_k\}$ induced by the basis $(v_k,\ldots, v_1)$. Any half-space that contains $L_{\lambda}$ would avoid the interior of one of the regions of the initial Yao--Yao partition constructed. \section{Additional remarks}\label{sect:remarks} Theorem \ref{thm:generalyao} gives an exact bound for $d=1$, but it is unclear whether this is the best that can be done for $d\geq2$. The only lower bound for Problem \ref{problem:generalizedyao} was proven by Rold\'an-Pensado and Sober\'on who showed that $\displaystyle n \geq 2^{\frac{d}{2} - 1}$ for $t=1$ \cite{roldan2014extension}. Besides this result, the following question remains open and relatively unexplored. \begin{problem}\label{problem:lowerboundn} Let $t, d$ be positive integers. Let $n$ be the smallest value such that for any finite measure $\mu$ in $\mathds{R}^d$ that is absolutely continuous with respect to the Lebesgue measure there exists a convex equipartition into $n$ parts such that every hyperplane avoids the interior of at least $t$ regions. What is the lower bound for $n$? \end{problem} In the case of $d=2$, we offer multiple constructions that match the current best upper bound. Looking at the complexity of a partition is not sufficient to give us a lower bound on $n$ in general. However, it allows us to bound $t$ in terms of $n$. \begin{problem}\label{problem:greatesttforn} Given $n, d$ find the largest $t$ such that for any finite measure $\mu$ in $\mathds{R}^d$ that is absolutely continuous with respect to the Lebesgue measure there exists a convex equipartition into $n$ parts such that every line misses the interior of at least $t$ parts. \end{problem} Let $k$ be the complexity for a convex partition of $\mathds{R}^d$. As we discussed in the previous sections, an upper bound on the complexity of an equipartition into $n$ parts gives us a lower bound on $t$ because $(n - (k + 1)) \leq t$. The following problem becomes relevant. \begin{problem}\label{problem:exactk} What is the smallest value of $k$ such that for every finite absolutely continuous measure $\mu$ in $\mathds{R}^d$ there exists a convex equipartition of $\mu$ with complexity at most $k$? \end{problem} Equipartitions impose greater geometric restrictions on possible partitions as Buck and Buck showed that there is no equipartition of $7$ regions of complexity $3$ \cite{buck1949equipartition}. Therefore, it seems that in order to prove an exact bound on $k$ the geometry of the equipartitions should be considered. It also seems that the original construction of Yao and Yao can be used to obtain more mass partition results. For example, consider the following proof of a special case of Theorem \ref{thm:two-hyperplanes} for $d=3$. \begin{lemma} Let $\mu_1, \mu_2$ be two finite absolutely continuous measures in $\mathds{R}^3$ whose supports can be separated by a plane $H$. Then, there exist two planes $H_1, H_2$ such that $H_1 \cup H_2$ splits $\mu_1$ into four equal parts and $H_1 \cap H_2$ is a central line for $\mu_2$. \end{lemma} Before showing the proof, notice the following property of the result above. Given a half-space $M$, if the bounding plane of $M$ hits $H_1 \cap H_2$ on the side of $H$ of $\mu_1$, then $\mu_2(M) \ge \mu_2(\mathds{R}^3)/3$. If it hits $H_1 \cap H_2$ on the side of $H$ of $\mu_2$, then $M$ contains one of the four regions that of the equipartition of $\mu_1$. \begin{proof} Assume without loss of generality that $H$ his a horizontal plane, and choose a basis $u_1, u_2$ of $H$ arbitrarily. For a vector $v \not\in \operatorname{span}\{u_1, u_2\}$, let $C_v(\mu_1)$ be the Yao--Yao center of $p_v(\mu_1)$ with respect to $u_1, u_2$. Let $C_v(\mu_2)$ be the centerpoint of $p_v(\mu_2)$. The exact same arguments of Yao and Yao show that, up to scalar multiplication, there exists a unique vector $v$ for which $C_v(\mu_1) = C_v(\mu_2)$ (the separation of the supports of $\mu_1, \mu_2$ by $H$ is needed for this). The Yao--Yao partition of $p_v(\mu_1)$ with respect to $u_1, u_2$ consists of two lines. If we extend these two lines by the direction $v$, we obtain $H_1$ and $H_2$. \end{proof} As a final remark, we compare Corollary \ref{coro-line-miss-many} to a similar problem regarding Yao--Yao type partitions for more measures. \begin{problem}[Problem 3.3.3 in \cite{roldan2021survey} for lines] Let $d$ be a positive integer. Find the smallest positive integer $n$ such that the following holds. For any $d-1$ absolutely continuous probability measures $\mu_1, \dots, \mu_{d-1}$ in $\mathds{R}^d$ there exists a convex partition $C_1, \dots C_n$ of $\mathds{R}^d$ such that \[ \mu_i (C_j) = \frac{1}{n} \qquad \text{for } i=1,\dots, d-1 \quad j= 1,\dots, n \] and every line misses the interior of at least one $C_j$. \end{problem} Corollary \ref{coro-line-miss-many} shows that the problem above is meaningful even if we have few measures. The problem above shows that we might expect results of this kind for up to $d-1$ measures simultaneously. Determining how the number of regions we are guaranteed to miss with a line decreases as we increase the number of measures in the equipartition is an interesting problem. \begin{bibdiv} \begin{biblist} \bib{alon2005crossing}{article}{ author={Alon, Noga}, author={Pach, J{\'a}nos}, author={Pinchasi, Rom}, author={Radoi{\v{c}}i{\'c}, Rado{\v{s}}}, author={Sharir, Micha}, title={Crossing patterns of semi-algebraic sets}, date={2005}, journal={Journal of Combinatorial Theory, Series A}, volume={111}, number={2}, pages={310\ndash 326}, } \bib{barany1980borsuk}{article}{ author={B{\'a}r{\'a}ny, Imre}, title={Borsuk's theorem through complementary pivoting}, date={1980}, journal={Mathematical Programming}, volume={18}, number={1}, pages={84\ndash 88}, } \bib{buck1949equipartition}{article}{ author={Buck, Robert~C}, author={Buck, Ellen~F}, title={Equipartition of convex sets}, date={1949}, journal={Mathematics Magazine}, volume={22}, number={4}, pages={195\ndash 198}, } \bib{Blagojevic:2018jc}{article}{ author={Blagojevi{\'c}, Pavle V.~M.}, author={Frick, Florian}, author={Haase, Albert}, author={Ziegler, G{\"u}nter~M.}, title={Topology of the {G}r{\"u}nbaum--{H}adwiger--{R}amos hyperplane mass partition problem}, date={2018}, journal={Transactions of the American Mathematical Society}, volume={370}, number={10}, pages={6795\ndash 6824}, } \bib{barany2008slicing}{article}{ author={B{\'a}r{\'a}ny, Imre}, author={Hubard, Alfredo}, author={Jer{\'o}nimo, Jes{\'u}s}, title={Slicing convex sets and measures by a hyperplane}, date={2008}, journal={Discrete \& Computational Geometry}, volume={39}, number={1-3}, pages={67\ndash 75}, } \bib{barba2019sharing}{article}{ author={Barba, Luis}, author={Pilz, Alexander}, author={Schnider, Patrick}, title={Sharing a pizza: bisecting masses with two cuts}, date={2019}, journal={arXiv preprint arXiv:1904.02502}, } \bib{Blagojevic:2018gt}{article}{ author={Blagojevi{\'c}, Pavle V.~M.}, author={Sober{\'o}n, Pablo}, title={Thieves can make sandwiches}, date={2018}, journal={Bulletin of the London Mathematical Society}, volume={50}, number={1}, pages={108\ndash 123}, } \bib{chan2020borsuk}{article}{ author={Chan, Yu~Hin}, author={Chen, Shujian}, author={Frick, Florian}, author={Hull, J~Tristan}, title={{Borsuk-Ulam theorems for products of spheres and Stiefel manifolds revisited}}, date={2020}, journal={Topological Methods in Nonlinear Analysis}, volume={55}, number={2}, pages={553\ndash 564}, } \bib{deLongueville:2006uo}{article}{ author={De~Longueville, Mark}, author={{\v Z}ivaljevi{\'c}, Rade~T.}, title={Splitting multidimensional necklaces}, date={2008}, journal={Advances in Mathematics}, volume={218}, number={3}, pages={926\ndash 939}, } \bib{dol1992generalization}{article}{ author={Dol'nikov, Vladimir~Leonidovich}, title={A generalization of the ham sandwich theorem}, date={1992}, journal={Mathematical Notes}, volume={52}, number={2}, pages={771\ndash 779}, } \bib{Fadell:1988tm}{article}{ author={Fadell, Edward}, author={Husseini, Sufian}, title={An ideal-valued cohomological index theory with applications to {B}orsuk--{U}lam and {B}ourgin--{Y}ang theorems}, date={1988}, journal={Ergodic Theory and Dynamical Systems}, volume={8}, pages={73\ndash 85}, } \bib{FHMRA19}{article}{ author={Fradelizi, Matthieu}, author={Hubard, Alfredo}, author={Meyer, Mathieu}, author={R{old{\'a}n-Pensado}, Edgardo}, author={Zvavitch, Artem}, title={{Equipartitions and Mahler volumes of symmetric convex bodies}}, date={2019}, journal={arXiv preprint arXiv:1904.10765}, } \bib{guillemin2010differential}{book}{ author={Guillemin, Victor}, author={Pollack, Alan}, title={Differential topology}, publisher={American Mathematical Society}, date={2010}, volume={370}, } \bib{grunbaum1960partitions}{article}{ author={Gr{\"u}nbaum, Branko}, author={others}, title={Partitions of mass-distributions and of convex bodies by hyperplanes.}, date={1960}, journal={Pacific Journal of Mathematics}, volume={10}, number={4}, pages={1257\ndash 1261}, } \bib{Hadwiger1966}{article}{ author={Hadwiger, Hugo}, title={Simultane {V}ierteilung zweier {K}{\"o}rper}, date={1966}, journal={Archiv der Mathematik}, volume={17}, number={3}, pages={274\ndash 278}, } \bib{Hubard:2019we}{article}{ author={Hubard, Alfredo}, author={Karasev, Roman~N.}, title={Bisecting measures with hyperplane arrangements}, date={2020}, journal={Mathematical Proceedings of the Cambridge Philosophical Society}, volume={169}, number={3}, pages={639\ndash 647}, } \bib{Karasev:2016cn}{article}{ author={Karasev, Roman~N.}, author={Rold{\'a}n-Pensado, Edgardo}, author={Sober{\'o}n, Pablo}, title={Measure partitions using hyperplanes with fixed directions}, date={2016}, journal={Israel journal of mathematics}, volume={212}, number={2}, pages={705\ndash 728}, } \bib{lehec2009yao}{article}{ author={Lehec, Joseph}, title={{On the Yao--Yao partition theorem}}, date={2009}, journal={Archiv der Mathematik}, volume={92}, number={4}, pages={366\ndash 376}, } \bib{matousek2003using}{book}{ author={Matou{\v s}ek, Ji{\v{r}}{\'\i}}, title={Using the {B}orsuk--{U}lam theorem: lectures on topological methods in combinatorics and geometry}, publisher={Springer}, date={2003}, } \bib{Mus12}{article}{ author={Musin, Oleg}, title={{Borsuk--Ulam type theorems for manifolds}}, date={2012}, journal={Proceedings of the American Mathematical Society}, volume={140}, number={7}, pages={2551\ndash 2560}, } \bib{Ramos:1996dm}{article}{ author={Ramos, Edgar~Arturo}, title={Equipartition of mass distributions by hyperplanes}, date={1996}, journal={Discrete {\&} Computational Geometry}, volume={15}, number={2}, pages={147\ndash 167}, } \bib{roldan2014extension}{article}{ author={R{old{\'a}n-Pensado}, Edgardo}, author={Sober{\'o}n, Pablo}, title={{An extension of a theorem of Yao and Yao}}, date={2014}, journal={Discrete \& Computational Geometry}, volume={51}, number={2}, pages={285\ndash 299}, } \bib{roldan2021survey}{article}{ author={R{old{\'a}n-Pensado}, Edgardo}, author={Sober{\'o}n, Pablo}, title={A survey of mass partitions}, date={2021}, journal={Bulletin of the American Mathematical Society}, note={Electronically published on February 24, 2021, DOI: https://doi.org/10.1090/bull/1725 (to appear in print).}, } \bib{Steinhaus1938}{article}{ author={Steinhaus, Hugo}, title={A note on the ham sandwich theorem}, date={1938}, journal={Mathesis Polska}, volume={9}, pages={26\ndash 28}, } \bib{thom1954quelques}{article}{ author={Thom, Ren{\'e}}, title={Quelques propri{\'e}t{\'e}s globales des vari{\'e}t{\'e}s diff{\'e}rentiables}, date={1954}, journal={Commentarii Mathematici Helvetici}, volume={28}, number={1}, pages={17\ndash 86}, } \bib{yao1985general}{inproceedings}{ author={Yao, Andrew~C}, author={Yao, F~Frances}, title={A general approach to d-dimensional geometric queries}, date={1985}, booktitle={{Proceedings of the seventeenth annual ACM symposium on Theory of Computing}}, pages={163\ndash 168}, } \bib{Zivaljevic:2017vi}{incollection}{ author={{\v{Z}}ivaljevi{\'c}, Rade~T.}, title={Topological methods in discrete geometry}, date={2017}, booktitle={Handbook of discrete and computational geometry, third edition}, publisher={CRC Press}, } \bib{zivaljevic1990extension}{article}{ author={{\v{Z}}ivaljevi{\'c}, Rade~T}, author={Vre{\'c}ica, Sini{\v{s}}a~T}, title={An extension of the ham sandwich theorem}, date={1990}, journal={Bulletin of the London Mathematical Society}, volume={22}, number={2}, pages={183\ndash 186}, } \end{biblist} \end{bibdiv} \end{document}

https://arxiv.org/abs/1601.02136

Spectra of general hypergraphs

Here, we show a method to reconstruct connectivity hypermatrices of a general hypergraph (without any self loop or multiple edge) using tensor. We also study the different spectral properties of these hypermatrices and find that these properties are similar for graphs and uniform hypergraphs. The representation of a connectivity hypermatrix that is proposed here can be very useful for the further development in spectral hypergraph theory.

\section{Introduction} Spectral graph theory has a long history behind its development. In spectral graph theory, we analyse the eigenvalues of a connectivity matrix which is uniquely defined on a graph. Many researchers have had a great interest to study the eigenvalues of different connectivity matrices, such as, adjacency matrix, Laplacian matrix, signless Laplacian matrix, normalized Laplacian matrix, etc. Now, a recent trend has been developed to explore spectral hypergraph theory. Unlike in a graph, an edge of a hypergraph can be constructed with more than two vertices, i.e., the edge set of a hypergraph is the subset of the power set of the vertex set of that hypergraph \cite{Voloshin}. Now, one of the main challenges is to uniquely represent a hypergraph by a connectivity hypermatrix or by a tensor, and vice versa. It is not trivial for a non-uniform hypergraph, where the cardinalities of the edges are not the same. Recently, the study of the spectrum of uniform hypergraph becomes popular. In a ($m$-) uniform hypergraph, each edge contains the same, ($m$), number of vertices. Thus an $m$-uniform hypergraph of order $n$ can be easily represented by an $m$ order $n$ dimensional connectivity hypermatrix (or tensor). In \cite{Cooper2012}, the results on the spectrum of adjacency matrix of a graph are extended for uniform hypergraphs by using characteristic polynomial. Spectral properties of adjacency uniform hypermatrix are deduced from matroids in \cite{Pearson2015}. In 1993, Fan Chung defined Laplacian of a uniform hypergraph by considering various homological aspects of hypergraphs and studied the eigenvalues of the same \cite{chung1993}. In \cite{Hu2013_2, Hu2014, Hu2015, Qi2014, Qi2014_2}, different spectral properties of Laplacian and signless Laplacian of a uniform hypergraph, defined by using tensor, have been studied. In 2015, Hu and Qi introduced the normalized Laplacian of a uniform hypergraph and analyzed its spectral properties \cite{Hu2013}. The important tool that has been used in spectral hypergraph theory is tensor. In 2005, Liqun Qi introduced the different eigenvalues of a real supersymmetric tensor \cite{Qi2005}. The various properties of the eigenvalues of a tensor have been studied in \cite{Chang2008, Chang2009, Li2013, Ng2009, Shao2013, shaoshan2013, Yang2010, YangYang20101}. But, still the challenge remains to come up with a mathematical framework to construct a connectivity hypermatrix for a non-uniform hypergraph, such that, based on this connectivity hypermatrix the spectral graph theory for a general hypergraph can be developed. Here, we propose a unique representation of a general hypergraph (without any self loop or multiple edge) by connectivity hypermatrices, such as, adjacency hypermatrix, Laplacian hypermatrix, signless Laplacian hypermatrix, normalized Laplacian hypermatrix and analyze the different spectral properties of these matrices. These properties are very similar with the same for graphs and uniform hypergraphs. Studying the spectrum of a uniform hypergraphs could be considered as a special case of the spectral graph theory of general hypergraphs. \section{Preliminary} Let $\mathbb{R}$ be the set of real numbers. We consider an $m$ order $n$ dimensional hypermatrix $A$ having $n^m$ elements from $\mathbb{R}$, where $$A=(a_{i_{1}, {i_{2}}, \dots, {i_{m}}}), a_{i_{1}, {i_{2}}, \dots, {i_{m}}} \in \mathbb{R} \text{ and } 1\leq i_{1}, i_{2}, \dots, i_{m} \leq n.$$ Let $x=(x_{1}, x_{2}, \dots, x_{n})\in \mathbb{R}^n$. If we write $x^{m}$ as an $m$ order $n$ dimension hypermatrix with $(i_1, i_2, \dots, i_m)$-th entry $x_{i_1} x_{i_2}\dots x_{i_m}$, then $Ax^{m-1}$, where the multiplication is taken as tensor contraction over all indices, is an $n$ tuple whose $i$-th component is $$ \sum_{i_{2}, i_{3}, \dots, i_{m}=1}^{n} a_{i i_{2}i_{3}\dots i_{m}} x_{i_{2}}x_{i_{3}} \dots x_{i_{m}} .$$ \begin{definition} Let $A$ be a nonzero hypermatrix. A pair $(\lambda, x) \in \mathbb{C} \times (\mathbb{C}^{n}\setminus \{0\})$ is called eigenvalue and eigenvector (or simply an eigenpair) if they satisfy the following equation $$Ax^{m-1}=\lambda x^{[m-1]}.$$ Here, $x^{[m]}$ is a vector with $i$-th entry $x^{m}_i$. We call $(\lambda, x)$ an $H$-eigenpair (i.e., $\lambda$ and $x$ are called $H$-eigenvalue and $H$-eigenvector, respectively) if they are both real. An $H$-eigenvalue $\lambda$ is called $H^+ (H^{++})$-eigenvalue if the corresponding eigenvector $x \in \mathbb{R}_+^n $ $(\mathbb{R}_{++}^n)$. \end{definition} \begin{definition} Let $A$ be a nonzero hypermatrix. A pair $(\lambda, x) \in \mathbb{C} \times (\mathbb{C}^{n}\setminus \{0\})$ is called an $E$-eigenpair (where $\lambda$ and $x$ are called $E$-eigenvalue and $E$-eigenvector, respectively) if they satisfy the following equations $$Ax^{m-1}= \lambda x,$$ $$ \sum_{i=1}^{n} x_{i}^2 =1.$$ We call $(\lambda, x)$ a $Z$-eigenpair if both of them are real. \end{definition} From the above definitions it is clear that, a constant multiplication of an eigenvector is also an eigenvector corresponding to an $H$-eigenvalue, but, this is not always true for $E$-eigenvalue and $Z$-eigenvalue. Now, we recall some results that are used in the next section. \begin{thm}[\cite{Qi2005}] The eigenvalues of $A$ lie in the union of $n$ disks in $\mathbb{C}$. These n disks have the diagonal elements of the supersymmetric tensor as their centers, and the sums of the absolute values of the off-diagonal elements as their radii. \end{thm} The above theorem helps us to bound the eigenvalues of a tensor. \begin{lemma}\label{co-spectral} Let $ A$ be an $m$ order and $n$ dimensional tensor and $D=diag(d_1,\dots ,d_n)$ be a positive diagonal matrix. Define a new tensor $${B}=A.D^{-(m-1)}.\overbrace{D \dots D}^{m-1} $$ with the entries $$B_{i_{1}i_{2} \dots i_{m}}=A_{i_{1}i_{2} \dots i_{m}}d_{i_1}^{-(m-1)}d_{i_{2}} \dots d_{i_{m}}.$$ Then $A$ and $B$ have the same $H$-eigenvalues. \end{lemma} \begin{proof} From the remarks of lemma (3.2) in \cite{Yang2010}. \end{proof} Some results of spectral graph theory\footnote{For different spectral properties of a graph see \cite{bapat2010, chung1997}} also hold for general hypergraphs. If $\lambda$ is any eigenvalue of an adjacency matrix of a graph $G$ with the maximal degree $\Delta$, then $\lambda\leqslant\Delta$. For a $k$-regular graph $k$ is the maximum eigenvalue with a constant eigenvector of the adjacency matrix of that graph. If $\lambda$ and $\mu$ are the eigenvalues of the adjacency matrices, represent the graphs $G$ and $H$, respectively, then $\lambda + \mu$ is also an eigenvalue of the same for $G \times H$, the cartesian product of $G$ and $H$. All the eigenvalues of a Laplacian matrix of a graph are nonnegative and a very rough upper bound of these eigenvalues is $2\Delta$, whereas, any eigenvalue of a normalized Laplacian matrix of a graph lies in the interval $[0, 2]$. Zero is always an eigenvalue for both, Laplacian and normalized Laplacian matrices, of a graph, with a constant eigenvector. If $\mathbb{A}$ and $\mathcal{L}$ are the normalized adjacency matrix and normalized Laplacian matrix, respectively, of a graph (such that $ \mathcal{L} = I - \mathbb{A}$) then the spectrum of $\mathbb{A}$, $\sigma(\mathbb{A})=1-\sigma(\mathcal{L})$. If $M$ is a connectivity matrix of a graph with $r$ connected components then $\sigma(M)=\sigma(M_1)\cup\sigma(M_2)\dots\cup\sigma(M_r)$, where $M_{i}$ is the same connectivity matrix corresponding to the component $i$. \section{Spectral properties of general hypergraphs} \begin{definition} A (general) hypergraph $G$ is a pair $G= (V, E)$ where $V$ is a set of elements called vertices, and $E$ is a set of non-empty subsets of $V$ called edges. Therefore, E is a subset of $\mathcal{P}(V) \setminus\{\emptyset\}$, where $\mathcal{P}(V)$ is the power set of $V$. \end{definition} \begin{exm}\label{Examp:Hypergraph} Let $G=(V, E)$, where $V=\{ 1, 2, 3, 4, 5 \}$ and $E= \big\{ \{ 1 \}, \{ 2, 3 \}, \{ 1, 4, 5 \} \big\}$. Here, $G$ is a hypergraph of $5$ vertices and $3$ edges. \end{exm} \subsection{Adjacency hypermatrix and eigenvalues} \begin{definition} Let $G=(V, E)$ be the hypergraph where $V= \{ v_{1}, v_{2}, \dots, v_{n} \}$ and $E = \{ e_{1}, e_{2}, \dots, e_{k} \}$. Let $m=max\{|e_{i}| : e_{i}\in E\}$ be the maximum cardinality of edges, $m.c.e(G)$, of $G$. Define the adjacency hypermatrix of $G$ as $$\mathcal{A}_G = (a_{i_{1}i_{2}\dots i_{m}}), \text{ } 1\leq i_{1}, i_{2}, \dots, i_{m} \leq n.$$ For all edges $e=\{ v_{l_{1}}, v_{l_{2}}, \dots, v_{l_{s}}\}\in E$ of cardinality $s \leq m$, $$a_{ p_{1}p_{2}\dots p_{m}}=\frac{s}{\alpha}, \text{ where } \alpha =\sum_{k_{1}, k_{2}, \dots, k_{s} \geq 1, \sum k_{i}=m} \frac{m!}{k_{1}! k_{2}!\dots k_{s}!}, $$ and $p_{1}, p_{2}, \dots, p_{m}$ chosen in all possible way from $\{l_{1}, l_{2}, \dots, l_{s}\}$ with at least once for each element of the set. The other positions of the hypermatrix are zero\footnote{For a similar construction on uniform multi-hypergraph see \cite{Pearson2014}.}. \end{definition} \begin{exm} Let $G=(V, E)$ be a hypergraph in example \ref{Examp:Hypergraph}. Here, the maximum cardinality of edges is $3$. The adjacency hypermatrix of $G$ is $\mathcal{A}_G = (a_{i_{1}i_{2}i_{3}})$, where $1\leq i_1, i_2, i_3 \leq 5$. Here, $a_{111}=1, a_{233}=a_{232}=a_{223}=a_{323}=a_{332}=a_{322}=\frac{1}{3}, a_{145}=a_{154}=a_{451}=a_{415}=a_{541}=a_{514}=\frac{1}{2}$, and the other elements of $\mathcal{A}_G$ are zero. \end{exm} \begin{definition} Let $G=(V, E)$ be a hypergraph. The degree, $d(v)$, of a vertex $v \in V$ is the number of edges consist of $v$. \end{definition} Let $G=(V, E)$ be a hypergraph, where $V=\{v_{1}, v_{2}, \dots, v_{n}\}$ and $E=\{e_{1}, e_{2}, \dots, e_{k}\}$. Then, the degree of a vertex $v_{i}$ is given by $$d(v_{i})= \sum_{i_{2}, i_{3}, \dots, i_{m}=1}^{n} a_{ii_{2}i_{3}\dots i_{m}}.$$ \begin{definition} A hypergraph is called $k$-\textit{regular} if every vertex has the same degree $k$. \end{definition} Now, we discuss some spectral properties of $\mathcal{A}_G$ of a hypergraph $G$. Some of these properties are very similar as in general graph (i.e.~for a $2$-uniform hypergraph). \begin{thm}\label{HEigUB} Let $\mu$ be an $H$-eigenvalue of $\mathcal{A}_G$. Then $|\mu|\leq\Delta$, where $\Delta$ is the maximum degree of $G$. \end{thm} \begin{proof} Let $G$ be a hypergraph with $n$ vertices and $m.c.e(G)=m$. Let $\mu$ be an $H$-eigenvalue of $\mathcal{A}_G=(a_{i_{1}i_{2}\dots i_{m}})$ with an eigenvector $x =( x_{1}, x_{2}, \dots, x_{n})$. Let $x_{p}=max \{ |x_{1}|, |x_{2}|, \dots, |x_{n}| \}$. Without loss of any generality we can assume that $x_{p}=1$. Now, $$|\mu|=|\mu x_{p}^{m-1}|=\bigg|\sum_{i_{2}, i_{3}, \dots, i_{m}=1}^{n} a_{pi_{2}i_{3}\dots i_{m}} x_{i_{2}}x_{i_{3}}\dots x_{i_{m}}\bigg|$$ $$ \leq \sum_{i_{2}, i_{3}, \dots, i_{m}=1}^{n} |a_{p i_{2}i_{3}\dots i_{m}}| |x_{p}|^{m-1} =d(v_{p})\leq \Delta.$$ \end{proof} Thus, for a $k$-regular hypergraph the theorem (\ref{HEigUB}) implies $|\mu| \leq k$. \begin{thm} Let $G=(V, E)$ be a $k$-regular hypergraph with $n$ vertices. Then, $\mathcal{A}_G=(a_{i_{1}i_{2}\dots i_{m}})$ has an $H$-eigenvalue $k$. \end{thm} \begin{proof} Since, $G$ is $k$-regular, then $d(v_{i})=k$ for all $v_{i}\in V$, $i\in\{1, 2, 3, ..., n\}$. Now, for a vector $x=(1, 1, 1, \dots, 1)\in \mathbb{R}^{n}$, we have $$\mathcal{A}_Gx^{m-1}= \sum_{i_{2}, i_{3}, \dots i_{m}=1}^{n} a_{ii_{2}i_{3}\dots i_{m}}=k.$$ Thus the proof. \end{proof} \begin{thm} Let $G=(V, E)$ be a $k$-regular hypergraph with $n$ vertices. Then, $\mathcal{A}_G=(a_{i_{1}i_{2}\dots i_{m}})$ has a $Z$-eigenvalue $k(\frac{1}{\sqrt{n}})^{m-2}$. \end{thm} \begin{proof} The vector $x=(\frac{1}{\sqrt{n}}, \frac{1}{\sqrt{n}}, \dots, \frac{1}{\sqrt{n}})\in \mathbb{R}^{n}$ satisfies the $Z$-eigenvalue equations for $\lambda =k(\frac{1}{\sqrt{n}})^{m-2}$. \end{proof} \begin{thm} Let $G$ be a hypergraph with $n$ vertices and maximum degree $\Delta$. Let $x=(x_{1}, x_{2}, \dots, x_{n})$ be a $Z$-eigenvector of $\mathcal{A}_G=(a_{i_{1}i_{2}\dots i_{m}})$ corresponding to an eigenvalue $\mu$. If $x_{p}=max \big\{|x_{1}|, |x_{2}|, \dots, |x_{n}| \big\}$, then $|\mu|\leq \frac{\Delta}{x_{p}}$. \end{thm} \begin{proof} The $Z$-eigenvalue equations of $\mathcal{A}_G$ for $\mu$ and $x$ are $Ax^{m-1}=\mu x$, and $\sum x_{i}^{2} =1$. Therefore, $|x_{i}|\leq 1$, for all $i=1, 2, 3, \dots, n$. Now, $$|\mu| |x_{j}|= \bigg|\sum_{i_{2}, i_{3}, \dots, i_{m}=1}^{n} a_{ji_{2}i_{3}\dots i_{m}} x_{i_{2}}x_{i_{3}}\dots x_{i_{m}} \bigg|, $$ which implies $|\mu| |x_{j}|\leq d(j)\leq \Delta, \forall j=1, 2, 3, \dots, n$. Therefore, $|\mu|\leq \frac{\Delta}{x_{p}}$. \end{proof} \begin{definition} A hypergraph $H=(V_{1}, E_{1})$ is said to be a spanning subhypergraph of a hypergraph $G=(V, E)$, if $V=V_{1}$ and $E_{1}\subseteq E$. \end{definition} \begin{thm} Let $G=(V, E)$ be hypergraph. Let $H=(V^{'}, E^{'})$ be a subhypergraph of $G$, such that, $ m.c.e(G)=m.c.e(H)$ be even. Then, $\mu _{max}(H)\leq \mu _{max}(G)$, where $\mu_{max} $ is the highest $Z$-eigenvalue of the corresponding adjacency hypermatrix. \end{thm} \begin{proof} Let $|V|=n$, $|V^{'}|=n^{'}$ ($\leq n$) and $m.c.e(G)=m.c.e(H)=m.$ Now, \begin{align*} \mu _{max}(H)& = max _{||x||=1} x^{t} \mathcal{A}_{H} x^{m-1} \text{ (by using lemma (3.1) in \cite{Li2013})}\\ & = max _{||x||=1} \bigg(\sum_{i_{1}, i_{2}, \dots, i_{m}=1}^{n^{'}} a^H_{i_{1}i_{2}...i_{m}} x_{i_{1}} x_{i_{2}}\dots x_{i_{m}}\bigg)\\ & = max _{||x||=1} \bigg(\sum_{i_{1}, i_{2}, \dots, i_{m}=1}^{n} a^H_{i_{1}i_{2}...i_{m}} x_{i_{1}} x_{i_{2}}\dots x_{i_{m}}\bigg), \text{ where } a^H_{i_{1}..i_{m}}=x_{i_r}=0 \text{ when } i_r >n^{'}\\ &\leq \bigg(\sum_{i_{1}, i_{2}, \dots, i_{m} =1}^{n} a^{G}_{i_{1}i_{2}\dots i_{m}} x_{i_{1}} x_{i_{2}}\dots x_{i_{m}}\bigg) \\ &\leq\mu _{max}(G), \end{align*} since each component of $x$ is nonnegative (by Perron-Frobenious theorem \cite{Chang2008}) and the number of edges of $G$ is greater than or equal to the number of edges of $H$. Hence the proof. \end{proof} \begin{definition} Let $G=(V, E)$ be a hypergraph with $V=\{v_{1}, v_{2}, \dots, v_{n}\}$, $E=\{e_{1}, e_{2}, \dots, e_{k}\}$, and $m.c.e(G)=m$. Let $x=(x_1, x_2, \dots, x_n)$ be a vector in $\mathbb{R}^n$ and $p\geq s-1$ be an integer. For an edge $e=\{ v_{l_{1}}, v_{l_{2}}, \dots, v_{l_{s}}\}$ and a vertex $v_{l_i}$, we define $$x^{e/v_{l_i}}_{p} := \sum x_{r_1}x_{r_2}\dots x_{r_p}, $$ where the sum is over $r_{1}, r_{2}, \dots, r_{p}$ chosen in all possible way from $\{l_{1}, l_{2}, \dots, l_{s}\}$, such that, all $l_j (j \ne i)$ occur at least once. Whereas, $$x^{e}_{p} := \sum x_{r_1}x_{r_2}\dots x_{r_p}, $$ where the sum is over $r_{1}, r_{2}, \dots, r_{p}$ chosen in all possible way from $\{l_{1}, l_{2}, \dots, l_{s}\}$ with at least once for each element of the set. The symmetric (adjacency) hypermatrix $\mathcal{A}_G$ of order $m$ and dimension $n$ uniquely defines a homogeneous polynomial of degree $m$ and in $n$ variables by $$F_{\mathcal{A}_G}(x)=\sum_{i_1, i_2, \dots, i_m =1}^{n} a_{i_1 i_2\dots i_m} x_{i_1}x_{i_2}\dots x_{i_m}.$$ \end{definition} We rewrite the above polynomial as: $$F_{\mathcal{A}_G}(x)= \sum_{e \in E} a_{e}^{G} x^{e}_{m}, $$ where $a_{e}^{G}=\frac{s}{\alpha}$, $\alpha =\sum_{k_{1}, k_{2}, \dots, k_{s} \geq 1, \sum k_{i}=m} \frac{m!}{k_{1}! k_{2}!\dots k_{s}!}$, and $s$ is the cardinality of the edge $e$. \begin{definition} Let $G$ and $H$ be two hypergraphs. The Cartesian product, $G \times H$, of $G$ and $H$ is defined by the vertex set $V(G \times H)=V(G) \times V(H)$ and the edge set $E(G \times H)= \big\{ \{v\} \times e : v \in V(G), e \in E(H)\big\} \bigcup \big\{ e \times \{v\} : e \in E(G), v \in V(H)\big\}.$ \end{definition} \begin{definition} Let $G$ be a hypergraph with the vertex set $V=\{v_{1}, v_{2}, \dots, v_{n}\}$ and $m.c.e(G)=m$. For an edge $e= \{ v_{l_1}, v_{l_2}, \dots, v_{l_s} \}$ and an integer $r \geq m$, the arrangement $(v_{p_1} v_{ p_2} \dots v_{p_r})$ (where $p_1, p_2, \dots, p_r$ are chosen in all possible way from $\{ l_{1}, l_{2}, \dots, l_{s} \}$ with at least once for each element of the set) represents the edge $e$ in order $r$. \end{definition} \begin{exm} Let $G=(V, E)$ where $V= \{1, 2, 3, 4, 5\}$ and $E= \big\{ \{1, 2, 3 \}, \{ 2, 3, 5 \}, \{ 1, 3, 4, 5\} \big\}$, then the arrangement $(12233)$ represents the edge $\{1, 2, 3 \}$ in order 5. $(12123)$ is also a representation of the edge $\{1, 2, 3 \}$ in order five, whereas, $(111123)$ represents the edge $\{1, 2, 3 \}$ in 6 order. \end{exm} Let $G=(V, E)$ be a hypergraph with $m.c.e(G) =m$ and $E_i= \{ e\in E : v_i \in e\}$. Now, the $H$-eigenvalue equation for $\mathcal{A}_G$ becomes $$\sum _{e\in E_i } a_{e}^{G} x^{e/v_i}_{m-1} =\lambda x_{i}^{(m-1)}, \text{ for all } i.$$ \begin{thm} Let $G$ and $H$ be two hypergraphs with $m.c.e(G) = m.c.e(H)$. If $\lambda$ and $\mu$ are $H$-eigenvalue for $G$ and $H$, respectively, then $\lambda +\mu$ is an $H$-eigenvalue for $G \times H$. \end{thm} \begin{proof} Let $n_1$ and $n_2$ be the number of vertices in $G$ and $ H$, respectively, and $m.c.e(G) = m.c.e(H)= m$. Let $(\lambda, \textbf{u})$ and $(\mu, \textbf{v})$ be $H$-eigenpairs of $\mathcal{A}_G$ and $\mathcal{A}_H$, respectively. Let $\textbf{w}\in\mathbb{C}^{n_{1}n_{2}}$ be a vector with the entries indexed by the pairs $(a, b)\in[n_1]\times[n_2]$, such that, $w(a, b)=u(a)v(b)$. Now, we show that $(\lambda+\mu, \textbf{w})$ is an $H$-eigenpair of $\mathcal{A}_{G\times H}$. \begin{align*} \sum_{e\in E_{(a, b)}}a_{e}^{G\times H}w^{e/{(a, b)}}_{m-1} &=\sum_{\substack{\{a\}\times e \in E_{(a, b)}\\ \mbox{ with } e\in E_b}}a_{e}^{G\times H} w^{\{a\}\times e/(a, b)}_{m-1} +\sum_{\substack {e \times \{b\} \in E_{(a, b)}\\ \mbox{ with } e \in E_a}} a_{e}^{G\times H}w^{e \times \{b\}/(a, b)}_{m-1}\\ &=\sum_{e \in H_b}a_{e}^{G\times H} u^{m-1}(a)v^{e/b}_{m-1}+\sum_{e \in G_a}a_{e}^{G\times H}u^{e/a}_{m-1}v^{m-1}(b)\\ &=u^{m-1}(a) \sum_{e \in H_b}a_{e}^{ H} v_{m-1}^{e/b}+ v^{m-1}(b) \sum_{e \in E_a}a_{e}^{G}u^{e/a}_{m-1}\\ &=u^{m-1}(a) \mu v^{m-1}(b)+v^{m-1}(b) \lambda u^{m-1}(a)\\ &=(\lambda+\mu)w^{m-1}(a, b). \end{align*} Hence the proof\footnote{For similar proof on uniform hypergraph see \cite{Cooper2012}.}. \end{proof} \begin{lemma} \label{largest z eigenvalue} Let $A$ and $B$ be two symmetric hypermatrices of order $m$ and dimension $n$, where $m$ is even. Then $\lambda_{max}(A+B)\leq\lambda_{max}(A)+\lambda_{max}(B)$, where $\lambda_{max}(A)$ denotes the largest $Z$-eigenvalue of $A$. \end{lemma} \begin{proof} \begin{align*} \lambda_{max}(A+B) & = max _{||x||=1} x^{t}(A+B)x^{m-1} \text{ (by using lemma (3.1) in \cite{Li2013})} \\ & \leq max _{||x||=1} x^{t}Ax^{m-1}+max _{||x||=1} x^{t}Bx^{m-1}\\ &= \lambda_{max}(A)+\lambda_{max}(B). \end{align*} \end{proof} Let $G=(V, E)$ be a hypergraph with the vertex set $V=\{v_{1}, v_{2}, \dots, v_{n}\}$ and $m=max\{|e_{i}| : e_{i}\in E\}$. We partition the edge set $E$ as, $E=E_{1} \cup E_{2} \cup \dots \cup E_{m}$, where $E_{i}$ contains all the edges of the cardinality $i$ and construct a hypergraph $G_{i}=(V, E_{i})$, for a nonempty $E_i$. \begin{definition} Define the adjacency hypermatrix of $G_{i}$ in $m$ $(> i)$ -order by an $n$ dimensional $m$ order hypermatrix $$\mathcal{A}_{G_{i}}^m = \big( (a_{G_{i}}^m)_{ p_{1}p_{2}\dots p_{m}} \big), \text { } 1 \leq p_{1}, p_{2}, \dots, p_{m} \leq n, $$ such that, for any $e=\{ v_{l_{1}}, v_{l_{2}}, \dots, v_{l_{i}} \} \in E_i$, $$(a_{G_{i}}^m)_{ p_{1}p_{2}\dots p_{m}}=\frac{i}{\alpha}, \text{ where } \alpha =\sum_{k_{1}, k_{2}, \dots, k_{i} \geq 1, \sum k_{j}=m} \frac{m!}{k_{1}! k_{2}!\dots k_{i}!}$$ and $p_{1}, p_{2}, \dots, p_{m}$ are chosen in all possible way from $\{l_{1}, l_{2}, \dots, l_{i}\}$ with at least once for each element of the set. The other positions of $\mathcal{A}_{G_{i}}^m$ are zero. \end{definition} Thus, we can represent a hypergraph $G$, with $m.c.e(G)=s$, in higher order $m > s$ by the hypermatrix $\mathcal{A}_{G}^m$. Clearly, all the eigenvalue equations show that the eigenvalues of $\mathcal{A}_{G}^{m_1}$ and $\mathcal{A}_{G}^{m_2}$ are not equal for $m_1 \ne m_2$. \begin{thm}\label{EdgePartitionTheo} Let $G=(V, E)$ be a hypergraph and $m.c.e(G)=m$ be even. Then $\lambda _{max}(\mathcal{A}_{G})\leq \sum _{i=1}^m\lambda _{max}(\mathcal{A}_{G_i}^m)$, where $\lambda_{max}({A})$ is the largest $Z$-eigenvalue of $A$. \end{thm} \begin{proof} Since $\mathcal{A}_{G}=\sum _{i=1}^{m} \mathcal{A}_{G_i}^m$, the proof follows from the lemma ($\ref{largest z eigenvalue}$). \end{proof} Moreover, the theorem (\ref{EdgePartitionTheo}) implies $\lambda _{max}(\mathcal{A}_{G})\leq \sum _{i=1}^m n_i \lambda _{max}(\mathcal{A}_{i}^m)$, where $n_{i}$ is the number of edges of cardinality $i$ and $\mathcal{A}_{i}^m$ is the adjacency hypermatrix in $m$-order of a hypergraph contains a single edge of cardinality $i$. \subsection{Laplacian hypermatrix and eigenvalues} \begin{definition} Let $G=(V, E)$ be a (general) hypergraph without any isolated vertex where $V= \{ v_{1}, v_{2}, \dots, v_{n} \}$ and $E = \{ e_{1}, e_{2}, \dots, e_{k} \}$. Let $m.c.e(G)=m$. We define the Laplacian hypermatrix, $L_{G}$, of $G=(V, E)$ as $L_{G}=D_{G}-\mathcal{A}_{G}=(l_{i_{1}i_{2}\dots i_{m}}), 1\leq i_{1}, i_{2}, \dots, i_{m} \leq n,$ where $D_{G}=(d_{i_{1}i_{2}\dots i_{m}})$ is the $m$ order $n$ dimensional diagonal hypermatrix with $d_{ii\dots i}=d(v_i)$ and others are zero. The signless Laplacian of $G$ is defined as $L_{G}=D_G+\mathcal{A}_{G}.$ \end{definition} Let $G=(V, E)$ be a hypergraph with $m.c.e(G)=m$. For any edge $e=\{ v_{l_{1}}, v_{l_{2}}, \dots, v_{l_{s}}\}$, we define a homogeneous polynomial of degree $m$ and in $n$ variables by $$ L(e)x^m = \sum_{j=1}^s x_{i_j}^m - \frac{s}{\alpha}x_m^e \text{ } (s\leq m). $$ \begin{proposition}\label{inequality} $ \sum_{j=1}^s x_{i_j}^m \geq \frac{s}{\alpha}x_m^e$ $(x_{i_j}\in \mathbb{R_{+}})$. \end{proposition} \begin{proof} $x_m^e$ is the sum of all possible terms, $x_{i_1}^{k_1} x_{i_2}^{k_2}\dots x_{i_s}^{k_s}$ (where $\sum k_i = m$ and $k_i \geq 1$) $$\text{ where } \alpha =\sum_{k_{1}, k_{2}, \dots, k_{s} \geq 1, \sum k_{i}=m} \frac{m!}{k_{1}! k_{2}!\dots k_{s}!}, $$with some natural coefficient. Now, by applying AM-GM inequality on $k_1 x_{i_1}^{m}, k_2 x_{i_2}^{m}, \dots, k_s x_{i_s}^{m}$ we get \begin{equation}\label{lemma:equation} \frac{1}{m} \sum_{j=1}^s k_j x_{i_j}^m \geq x_{i_1}^{k_1} x_{i_2}^{k_2}\dots x_{i_s}^{k_s}. \end{equation} If we apply (\ref{lemma:equation}) for each term of $x_m^e$ and take the sum, we get $$ \frac{\alpha}{s} \sum_{j=1}^s x_{i_j}^m \geq x_m^e. $$ \end{proof} Many properties of Laplacian and signless Laplacian tensors are discussed in \cite{Qi2014}. Now we show that some of the results in general graph are also true for non-uniform (general) hypergraph. Note that, here, $L$ is co-positive tensor since, $Lx^m=\sum _{e\in E}L(e)x^m \geq 0$ for all $x \in \mathbb{R}_{+}^n $. \begin{thm}\label{positivity} Let $G=(V, E)$ be a general hypergraph. Let $L=(l_{i_{1}i_{2}\dots i_{m}}) \text{ where } 1\leq i_{1}, i_{2}, \dots, i_{m} \leq n,$ be the Laplacian hypermatrix of $G$. Then $0\leq \lambda \leq 2\Delta,$ where $\lambda$ is an $H$-eigenvalue of $L$. \end{thm} \begin{proof} For a vector $y=(\frac{1}{n^{\frac{1}{m}}},\frac{1}{n^{\frac{1}{m}}},\dots, \frac{1}{n^{\frac{1}{m}}}), Ly^m=0$. Since $L$ is a co-positive tensor, thus $\textit{min}\{Lx^m :x \in \mathbb{R}_+^{n},\sum _{i=1}^{n} x_{i}^m=1\}=0$. Therefore $\lambda \geq 0$. Again using theorem ($6(a)$) of \cite{Qi2005} we have $$|\lambda -l_{ii\dots i}|\leq \sum_{\substack{i_{2}, i_{3}, \dots, i_{m}=1, \\ \delta _{i,i_2,\dots ,i_m}=0}}^{n} |l_{ii_{2}i_{3}\dots i_{m}}| = \Delta,$$ i.e., $|\lambda| \leq 2 \Delta.$ Thus $0\leq \lambda \leq 2\Delta$. \end{proof} \begin{thm} Let $G=(V, E)$ be a general hypergraph with $m.c.e(G) = m\geq 3$. Let $L$ be the Laplacian hypermatrix of $G$. Then \begin{enumerate}[(i)] \item $L$ has an $H$-eigenvalue 0 with eigenvector $(1, 1, \dots, 1)\in \mathbb{R}^n$ and an $Z$-eigenvalue 0 with eigenvector $x=(\frac{1}{\sqrt{n}}, \frac{1}{\sqrt{n}}, \dots, \frac{1}{\sqrt{n}})\in \mathbb{R}^{n}$. Moreover, $0$ is the unique $H^{++}$-eigenvalue of $L$. \item $\Delta$ is the largest $H^+$-eigenvalue of $L$. \item $(d(i), e^{(j)})$ is an $H$-eigenpair, where $e^{(j)}\in\mathbb{R}^{n}$ and $e^{(j)}_{i}=1$ if $i=j$, otherwise 0. \item For a nonzero $x\in \mathbb{R}^n$ $(d(v_i), x)$ is an eigenpair if $\sum_{e\in E_{i}}a_{G}^{e} x_{m-1}^{e/i}=0$. \end{enumerate} \end{thm} \begin{proof} \begin{enumerate}[(i)] \item It is easy to check that $0$ is an $H$-eigenvalue with the eigenvctor $(1,1,1,\dots, 1)\in \mathbb{R}^{n}$ and $0$ is an $Z$-eigenvalue with the eigenvector $x=(\frac{1}{\sqrt{n}}, \frac{1}{\sqrt{n}}, \dots, \frac{1}{\sqrt{n}})\in \mathbb{R}^{n}$. Let $x$ is an $H^{++}$-eigenvector of $L$ with eigenvalue $\lambda$. By theorem (\ref {positivity}), $\lambda \geq 0.$ Suppose $x_{j}=\underset{i}{min}\{x_i\}$. Therefore $x_j$ is positive. Now,$$\lambda x_j^{m-1}=d(v_j)x_j^{m-1}-\sum _{e \in E,j\in e,|e|=s}\frac{s}{\alpha}\sum _{\substack{e\equiv \{i,i_2,\dots ,i_m\} \\ \text {as set, } i,i_2,\dots ,i_m=1}}^n x_{i_2} x_{i_3} \dots x_{i_m},$$ which implies that $$\lambda=d(v_j)-\sum _{e \in E,j\in e,|e|=s}\frac{s}{\alpha}\sum _{\substack{e\equiv \{i,i_2,\dots ,i_m\} \\ \text {as set, } i,i_2,\dots ,i_m=1}}^n \frac{x_{i_2}}{x_j}\frac{x_{i_3}}{x_j}\dots \frac{x_{i_m}}{x_j}.$$ Thus, $\lambda \leq d(v_j)-d(v_{j})=0$. Hence $\lambda=0$. \item Suppose $\lambda$ is an $H^+$-eigenvalue with non-negative $H^+$-eigenvector, $x$ of $L$. Assume that $x_j>0$. Now, we have $$\lambda x_j^{m-1}=d(v_j)x_j^{m-1}-\sum _{e \in E,j\in e,|e|=s}\frac{s}{\alpha}\sum _{\substack{e\equiv \{i,i_2,\dots ,i_m\} \\ \text {as set, } i,i_2,\dots ,i_m=1}}^n x_{i_2} x_{i_3} \dots x_{i_m} \leq d(v_j)x_j^{m-1}.$$ Therefore $\lambda \leq d(v_j)\leq\Delta$. Thus, $\Delta$ is the largest $H^+$-eigenvalue of $L$. \item Proof is obvious. \item It is clear from the eigenvalue equation. \end{enumerate} \end{proof} Let $G=(V, E)$ be a general hypergraph and $m.c.e(G)=m$. The \textit {analytic connectivity}, $\alpha(G)$, of $G$ is defined as $\alpha(G)= \underset{j=1, \dots, n}{min} min\{Lx^m | x \in \mathbb{R}_{+}^n, \sum_{i=1}^{n}x_{i}^{m}=1, x_{j}=0\} $. \begin{thm} The general hypergraph $G=(V,E)$ with $m.c.e(G)\geq 3$ is connected if and only if $\alpha(G)>0.$ \end{thm} \begin{proof} Suppose $G=(V,E)$ is not connected. Let $G_1=(V_1,E_1)$ be a component of $G$. Then there exists $j \in V\setminus V_1$. Let $x=\frac{1}{|V_1|^{\frac{1}{m}}}\sum _{i \in V_1}e^{(i)}$. Then $x$ is a feasible point. Therefore $min\{Lx^m | x \in \mathbb{R}_{+}^n, \sum_{i=1}^{n}x_{i}^{m}=1, x_{j}=0\}=0$, which implies $\alpha (G)=0$. Let $\alpha(G)=0$. Thus there exists $j$ such that $min\{Lx^m | x \in \mathbb{R}_{+}^n, \sum_{i=1}^{n}x_{i}^{m}=1, x_{j}=0\}=0$. Suppose that $y$ is a minimizer of this minimization problem. Therefore $y_j=0$, $Ly^m=0$. By optimization theory, there exists a Lagrange multiplier $\mu$ such that for $i=1,2,\dots ,n$ and $i\neq j$, either, $y_i=0$ and \begin{equation} \frac{\partial}{\partial y_{i}}(Ly^m)\geq \mu \frac{\partial}{\partial y_i}(\sum_{i=1}^{n}y_{i}^{m}-1) \end{equation} or, $y_i>0$ and \begin{equation} \frac{\partial}{\partial y_{i}}(Ly^m)=\mu \frac{\partial}{\partial y_i}(\sum_{i=1}^{n}y_{i}^{m}-1). \end{equation} In (2) and (3) $y \in \mathbb{R}_{+}^n, \sum_{i=1}^{n}y_{i}^{m}=1, y_{j}=0$. Now, multiplying (2) and (3) by $y_i$ and summing them for $i=1, \dots,n$, we have $Ly^m=\mu (\sum_{i=1}^{n}y_{i}^{m})$. Thus $Ly^m=\mu$. Hence $\mu =0$. Therefore, for $i=1,2,\dots ,n$ and $i\neq j$, either $y_i=0$ or $\frac{\partial}{\partial y_i}(Ly^m)=0$. Hence, either $y_i=0$ or $d_i(y_i)^{m-1}-\sum_{i_{2}, i_{3}, \dots, i_{m}=1}^{n} a_{i i_{2}i_{3}\dots i_{m}} y_{i_{2}}y_{i_{3}} \dots y_{i_{m}}=0$. Let $y_k=max\{y_i:i=1,2,\dots ,n\}$. Hence, we have $$d_k-\sum_{i_{2}, i_{3}, \dots, i_{m}=1}^{n} a_{i i_{2}i_{3}\dots i_{m}}\frac{y_{i_2}}{y_k}\frac{y_{i_3}}{y_k}\dots \frac{y_{i_m}}{y_k}=0.$$ Again, we know that $$d(v_{k})= \sum_{i_{2}, i_{3}, \dots, i_{m}=1}^{n} a_{ii_{2}i_{3}\dots i_{m}}.$$ Therefore, $x_i=x_k$ as long as $i$ and $k$ belong to same edge. Thus, $x_i = x_k$ as long as $i$ and $k$ are in same component of $G$. Since $y_j=0$, we have, $j$ and $k$ are in different components of $G$. Hence, $G$ is not connected. This proves the theorem. \end{proof} \subsection{Normalized Laplacian hypermatrix and eigenvalues} Now, we define normalized Laplacian hypermatrix for a general hypergraph. For any graph, there are two ways to construct normalized Laplacian matrix (see \cite{Banerjee2008} and \cite{chung1997} for details)\footnote{These two matrices are similar, i.e., they have same eigenvalues.}. Motivated by these two similar constructions, here, we also define the normalized Laplacian hypermatrix in two different ways and show that they are cospectral. The first definition is similar to the normalized Laplacian matrix defined in \cite{Banerjee2008}. \begin{definition} Let $G=(V, E)$ be a general hypergraph without any isolated vertex where $V= \{ v_{1}, v_{2}, \dots, v_{n} \}$ and $E = \{ e_{1}, e_{2}, \dots, e_{k} \}$. Let $m.c.e(G)=m$. The normalized Laplacian hypermatrix $\mathcal{L} = (l_{i_1i_2\dots i_m})$, which is an $n$-dimensional $m$-th order hypermatrix, is defined as: for any edge $e=\{ v_{l_{1}}, v_{l_{2}}, \dots, v_{l_{s}}\}\in E$ of cardinality $s \leq m$, $$l_{ p_1p_{2}\dots p_{m}}=-\frac{s/\alpha}{d(v_{p_1})}, \text{ where } \alpha =\sum_{\substack{k_{1}, k_{2}, \dots, k_{s} \geq 1, \\ \sum k_{i}=m}} \frac{m!}{k_{1}! k_{2}!\dots k_{s}!}$$ and $p_1, p_{2}, \dots, p_{m}$ are chosen in all possible way from $\{l_{1}, l_{2}, \dots, l_{s}\}$, such that, all $l_j $ occur at least once. All the diagonal entries are 1 and the rest are zero. \end{definition} Clearly, the hypermatrix $\mathit{A}=\mathcal{I}-\mathcal{L}$, which is known as normalized adjacency hypermatrix, is a stochastic tensor, that is, $\mathit{A}$ is non-negative and $ \sum_{i_{2}, \dots, i_{m}=1}^{n} a_{i i_{2}\dots i_{m}}=1$, where $a_{i_1 i_{2}\dots i_{m}}$ is the $(i_1, i_2,\dots, i_m)$-th entry of $\mathit{A}$. The different properties of a stochastic tensor are discussed in \cite{YangYang20101} and which can be used to study the hypermatrices $\mathit{A}$ and $\mathcal{L}$. Now, we define the normalized Laplacian hypermatrix of a general hypergraph as it is defined for a graph in \cite{chung1997}. \begin{definition} Let $G=(V, E)$ be a general hypergraph without any isolated vertex, where $V= \{ v_{1}, v_{2}, \dots, v_{n} \}$ and $E = \{ e_{1}, e_{2}, \dots, e_{k} \}$. Let $m.c.e(G)=m$. The normalized Laplacian hypermatrix $\mathfrak{L}= (l_{i_1i_2\dots i_m})$, which is an $n$-dimension $m$-th order symmetric hypermatrix, is defined as: for any edge $e=\{ v_{l_{1}}, v_{l_{2}}, \dots, v_{l_{s}}\}\in E$ of cardinality $s \leq m$, $${l}_{ p_{1}p_{2}\dots p_{m}}=-\frac{s}{\alpha}\prod_{j=1}^m\frac{1}{\sqrt[m]{d(v_{p_{j}})}}, \text{ where } \alpha =\sum_{k_{1}, k_{2}, \dots, k_{s} \geq 1, \sum k_{i}=m} \frac{m!}{k_{1}! k_{2}!\dots k_{s}!}$$ and $p_{1}, p_{2}, \dots, p_{m}$ chosen in all possible way from $\{l_{1}, l_{2}, \dots, l_{s}\}$ with at least once for each element of the set. The diagonal entries of $\mathfrak{L}$ are $1$ and the rest of the positions are zero. \end{definition} \begin{thm} $\mathcal{L}$ and $\mathfrak{L}$ are co-spectral. \end{thm} \begin{proof} In the lemma (\ref{co-spectral}) choose a diagonal matrix $D=(d_{ij})_{n \times n}$ where $d_{ii}=(d(v_i))^{1/m}$. \end{proof} \begin{thm}\label{spectra} Let $G=(V, E)$ be a general hypergraph. Let $\mathcal{L}$, $\mathit{A}$ be the normalized Laplacian and normalized adjacency hypermatrices of $G$, respectively. If $G$ has at least one edge, then $\lambda \in \sigma(\mathit{A})$ if and only if $(1-\lambda)\in \sigma(\mathcal{L})$, otherwise, $\sigma(\mathit{A})=\sigma(\mathcal{L})={0}$, where $\sigma(\mathcal{L})$ denotes the spectrum of $\mathcal{L}$. \end{thm} \begin{proof} Since, $\mathcal{L}=\mathcal{I}-\mathit{A}$ and $\lambda$ is the eigenvalue of $\mathit{A}$ \textit{iff} $\det(\mathit{A}- \lambda \mathcal{I})=0,$ thus, $\det(\mathcal{L}- (1-\lambda) \mathcal{I})=0$ implies $(1-\lambda)\in \sigma(\mathcal{L})$. \end{proof} \begin{thm} Let $G=(V, E)$ be a general hypergraph. Let $\mathcal{L}=(l_{i_{1}i_{2}\dots i_{m}}) \text{ where } 1\leq i_{1}, i_{2}, \dots, i_{m} \leq n,$ and $A$ be the normalized Laplacian and normalized adjacency hypermatrices of $G$, respectively, then \begin{enumerate}[(i)] \item $\rho(A)=1$. \item $0\leq\lambda(\mathcal{L})\leq 2$. \item 1 is the largest $H^+$-eigenvalue of $\mathcal{L}$. \item $0$ is an eigenvalue of $\mathcal{L}$ with the eigenvector $(1, 1, \dots, 1)$ and $0$ is an $Z$-eigenvalue with eigenvector $x=(\frac{1}{\sqrt{n}}, \frac{1}{\sqrt{n}}, \dots, \frac{1}{\sqrt{n}})\in \mathbb{R}^{n}$. \item $0$ is the unique $H^{++}$-eigenvalue of $\mathcal{L}$ . \end{enumerate} \end{thm} \begin{proof} \begin{enumerate}[(i)] \item Since $A$ is stochastic tensor, it is obvious that the spectral radius of $A$ is 1. Moreover, $(1,1,\dots, 1)$ is an eigenvector with eigenvalue 1. \item We know that spectral radius of $A$ is 1 and $\mathcal{L}=\mathcal{I}-A$. By theorem (\ref{spectra}) $\lambda \in \sigma(\mathit{A})$ if and only if $(1-\lambda)\in \sigma(\mathcal{L})$. Since 1 is an eigenvalue of $A$, thus, $\lambda \geq 0$. Again, using theorem ($6(a)$) of \cite{Qi2005} we have $$|\lambda(\mathcal{L}) -1|\leq \sum_{\substack{i_{2}, i_{3}, \dots, i_{m}=1,\\ \delta _{i,i_2,\dots ,i_m}=0}}^{n} |l_{ii_{2}i_{3}\dots i_{m}}| = 1.$$ This implies $|\lambda(\mathcal{L})| \leq 2.$ Thus, we have $0\leq \lambda(\mathcal{L}) \leq 2$. \item Suppose that $\lambda$ is an $H^+$-eigenvalue with non-negative $H^+$-eigenvector, $x$ of $\mathcal{L}$. Assume that $x_j>0$. Now, we have $$\lambda x_j^{m-1}=x_j^{m-1}-\frac{1}{d(v_j)}\sum _{e \in E,j\in e,|e|=s}\frac{s}{\alpha}\sum _{\substack{e\equiv\{i,i_2,\dots ,i_m\} \\ \text {as set, } i,i_2,\dots ,i_m=1}}^n x_{i_2} x_{i_3} \dots x_{i_m}.$$ Hence, $\lambda x_j^{m-1}\leq x_j^{m-1}$ implies $\lambda \leq 1$. Thus, 1 is the largest $H^+$-eigenvalue of $\mathcal{L}$. \item It is easy to check that $0$ is an $H$-eigenvalue corresponding an eigenvector $(1,1,\dots, 1)\in \mathbb{R}^{n}$ and $0$ is an $Z$-eigenvalue with the eigenvector $x=(\frac{1}{\sqrt{n}}, \frac{1}{\sqrt{n}}, \dots, \frac{1}{\sqrt{n}})\in \mathbb{R}^{n}$. \item Let $x$ is an $H^{++}$-eigenvector of $\mathcal{L}$ with eigenvalue $\lambda$. From the part (ii) of this theorem we have $\lambda \geq 0.$ Suppose $x_{j}=\underset {i}{min}\{x_i\}$. Now,$$\lambda x_j^{m-1}=x_j^{m-1}-\frac{1}{d(v_j)}\sum _{e \in E,j\in e,|e|=s}\frac{s}{\alpha}\sum _{\substack{e\equiv\{i,i_2,\dots ,i_m\} \\ \text {as set} ,i,i_2,\dots ,i_m=1}}^n x_{i_2} x_{i_3} \dots x_{i_m},$$ which implies that $$\lambda=1-\frac{1}{d(v_j)}\sum _{e \in E,j\in e,|e|=s}\frac{s}{\alpha}\sum _{\substack{e\equiv\{i,i_2,\dots ,i_m\} \\ \text {as set, } i,i_2,\dots ,i_m=1}}^n \frac{x_{i_2}}{x_j}\frac{x_{i_3}}{x_j}\dots \frac{x_{i_m}}{x_j}.$$ Thus $\lambda \leq 1-1=0$. Hence $\lambda=0$. \end{enumerate} \end{proof} \begin{thm} Let $G=(V, E)$ be a general hypergraph and $m.c.e(G)=m$. Let $\mathcal{L}$ be the normalized Laplacian hypermatrix of $G$ of order m and dimension n. Let $m(\lambda)$ be the algebraic multiplicity of $\lambda \in \sigma(\mathcal{L})$, then $\sum_{\lambda \in \sigma(\mathcal{L})} m(\lambda)\lambda =n(m-1)^{n-1}.$ \end{thm} \begin{proof} Since, for any tensor $\mathcal{T} = (t_{i_1i_2\dots i_m}), t_{i_1i_2\dots i_m} \in \mathbb{C},\text{ } 1\leq i_{1}, i_{2}, \dots, i_{m} \leq n,$ $$\sum _{\lambda \in \sigma (\mathcal{T})}m(\lambda)\lambda=(m-1)^{(n-1)} \sum _{i=1}^{n} t_{ii\dots i} \text{ (see \cite{Hu2013_3})}.$$ Hence, we have $\sum_{\lambda \in \sigma(\mathcal{L})} m(\lambda)\lambda =n(m-1)^{n-1}.$ \end{proof} \begin{thm} Let $G=(V, E)$ be a general hypergraph and $A$ be any connectivity hypermatrix of $G$. If $G$ has $r\geq 1$ connected components, $G_{1}, G_{2}, \dots, G_{r}$, such that, $|V(G_{i})|=n_i>1$ and $m.c.e(G_i) = m.c.e(G)$ for each $i\in\{1, 2, \dots, r \}$. Then, as sets, $\sigma(A)=\sigma(A_{1})\cup \sigma(A_{2})\cup\dots \cup \sigma(A_{r}),$ where $A_i$ is the connectivity hypermatrix of $G_i$. \end{thm} \begin{proof} Using corollary (4.2) of \cite {shaoshan2013} we get $$\phi _{A}(\lambda)=\prod _{i=1}^{r}(\phi _{A}(\lambda))^{(m-1)^{n-n_i}},$$ where $\phi _{A}(\lambda) $ is the characteristic polynomial of the tensor $A$. Therefore, $\sigma(A)=\sigma(A_{1})\cup \sigma(A_{2})\cup\dots \cup \sigma(A_{r})$. \end{proof} \section{Discussion and conclusion} Here, we propose a mathematical framework to construct connectivity matrices for a general hypergraph and also study the eigenvalues of adjacency hypermatrix, Laplacian hypermatrix, normalized Laplacian hypermatrix. This connectivity hypermatrix reconstruction can be used for further development of spectral hypergraph theory in many aspects, but, this may not be quite useful to study dynamics on hypergraphs. \section*{Acknowledgements} The authors are thankful to Mithun Mukherjee and Swarnendu Datta for fruitful discussions. Financial support from Council of Scientific and Industrial Research, India, Grant no-09/921(0113)/2014-EMR-I is sincerely acknowledged by Bibhash Mondal.

https://arxiv.org/abs/1106.5844

Extremum problems for eigenvalues of discrete Laplace operators

The discrete Laplace operator on a triangulated polyhedral surface is related to geometric properties of the surface. This paper studies extremum problems for eigenvalues of the discrete Laplace operators. Among all triangles, an equilateral triangle has the maximal first positive eigenvalue. Among all cyclic quadrilateral, a square has the maximal first positive eigenvalue. Among all cyclic $n$-gons, a regular one has the minimal value of the sum of all nontrivial eigenvalues and the minimal value of the product of all nontrivial eigenvalues.

\section{Introduction} \subsection{} A polyhedral surface $S$ is a surface obtained by gluing Euclidean triangles. It is associated with a triangulation $T$. We assume that $T$ is simplicial. Suppose $(\Sigma,T)$ is a polyhedral surface so that $V,E,F$ are sets of all vertices, edges and triangles in $T.$ We identify vertices of $T$ with indices, edges of $T$ with pairs of indices and triangles of $T$ with triples of indices. This means $V=\{1,2,...|V|\}, E=\{ij\ |\ i,j\in V\}$ and $F=\{\triangle ijk\ |\ i,j,k\in V\}.$ A vector $(f_1,f_2,...,f_{|V|})^t$ indexed by the set of vertices $V$ defines a piecewise-linear function over $(S,T)$ by linear extension. The Dirichlet energy of a function $f$ on $S$ is $$E_S(f)=\frac12\int_S|\nabla f|^2 dA.$$ When $f$ is obtained by linear extension of $(f_1,f_2,...,f_{|V|})^t$, the Dirichlet energy of $f$ turns out to be $$E_{(S,T)}(f)=\frac14\sum_{ijk\in F}[\cot\alpha_{jk}^i(f_j-f_k)^2+\cot\alpha_{ki}^j(f_k-f_i)^2+\cot\alpha_{ij}^k(f_i-f_j)^2]$$ where the sum runs over all triangles of $T$ and for a triangle $ijk\in F$, $\alpha_{jk}^i, \alpha_{ki}^j, \alpha_{ij}^k$ are angles opposite to the edges $jk, ki, ij$ respectively. Collecting the terms in the sum above according to edges, we obtain \begin{align}\label{fm:cot} E_{(S,T)}(f)=\frac14\sum_{ij\in E}w_{ij}(f_i-f_j)^2 \end{align} where the sum runs over all edges of $T$ and $$ w_{ij}=\left \{ \begin{array}{lll} \frac12 (\cot\alpha_{ij}^k+\cot\alpha_{ij}^l) &\ \ \mbox{if $ij$ is shared by two triangles $ijk$ and $ijl$} \\ \frac12 (\cot\alpha_{ij}^k) &\ \ \mbox{if $ij$ is contained in one triangles $ijk$} \end{array} \right. $$ The Dirichlet energy of a piecewise linear function on a polyhedral surface was introduced and the formula (\ref{fm:cot}) was derived by R. J. Duffin \cite{D}, G. Dziuk \cite{Dz} and U. Pinkall \& K. Polthier \cite{PP} in different context. For application of the Dirichlet energy and formula (\ref{fm:cot}) in the characterization of Delaunay triangulations, see \cite{R, G, BS, CXGL}. For interesting application of the Dirichlet energy and formula (\ref{fm:cot}) in computer graphics, see the survey \cite{BKPAL}. \subsection{} The discrete Laplace operator $L$ can be introduced by rewriting the Dirichlet energy using notation of matrices $$E_{(S,T)}(f)=\frac12 (f_1,...,f_{|V|})L(f_1,...,f_{|V|})^t$$ where each entry of the matrix $L$ is given as $$ L_{ij}=\left \{ \begin{array}{lll} \sum_{ik\in E} w_{ik} &\ \ \mbox{if $i=j$} \\ -w_{ij} &\ \ \mbox{if $ij\in E$} \\ 0 &\ \ \mbox{otherwise} \end{array} \right. $$ By definition $L$ is positive semi-definite. Its eigenvalues are denoted by $$0=\lambda_0 \leq \lambda_1 \leq \lambda_2 \leq ... \leq \lambda_{|V|-1}.$$ The discrete Laplace operator and its eigenvalues are related to the geometric properties of the polyhedral surface $(S,T)$. For example, it is proved in \cite{CXGL} that among all triangulations, the Delaunay triangulation has the minimal eigenvalues. In \cite{GGLZ}, it is shown that a polyhedral metric on a surface is determined up to scaling by its discrete Laplace operator. \subsection{} In smooth case, the spectral geometry is to relate geometric properties of a Riemannian manifold to the spectra of the Laplace operator on the manifold. One of the interesting result is the following one due to G. P\'olya. For reference, for example, see \cite{H}, page 50. \begin{1}[P\'olya] The equilateral triangle has the least first eigenvalue among all triangles of given area. The square has the least first eigenvalue among all quadrilaterals of given area. \end{1} It is conjectured that, for $n\geq 5,$ the regular $n$-gon has the least first eigenvalue among all $n$-gons of given area. \subsection{} In this paper, similar results as P\'olya's theorem are obtained for the discrete Laplace operator. \begin{theorem}\label{thm:3} Among all triangles, an equilateral triangle has the maximal $\lambda_1$, the minimal $\lambda_2$ and the minimal $\lambda_1+\lambda_2$. \end{theorem} A cyclic polygon is a polygon whose vertices are on a common circle. By adding diagonals, a cyclic polygon is decomposed into a union of triangles. For each inner edge of any triangulation of a a cyclic polygon, the weight $w_{ij}$ is zero. Therefore the discrete Laplace operator is independent of the choice of a triangulation of a cyclic polygon. \begin{theorem}\label{thm:4} Among all cyclic quadrilaterals, a square has the maximal $\lambda_1$, the minimal $\lambda_1+\lambda_2+\lambda_3$, the minimal $\lambda_1\lambda_2+\lambda_2\lambda_3+\lambda_3\lambda_1$ and the minimal $\lambda_1\lambda_2\lambda_3.$ \end{theorem} \begin{theorem}\label{thm:n} For $n\geq 5,$ among all cyclic $n$-gons, a regular $n$-gon has the minimal $\sum_{i=1}^{n-1}\lambda_i$ and the minimal $\prod_{i=1}^{n-1}\lambda_i$. \end{theorem} \subsection{Plan of the paper} Theorem \ref{thm:3}, Theorem \ref{thm:4} and Theorem \ref{thm:n} are proved in section 2, section 3 and section 4 respectively. \section{Triangles} \subsection{} In this section we prove Theorem \ref{thm:3}. Let $\theta_1,\theta_2,\theta_3$ be the three angles of a triangle. Let $a_i:=\cot\theta_i$ for $i=1,2,3.$ The condition $\theta_1+\theta_2+\theta_3=\pi$ implies that \begin{align}\label{fm:3} a_1a_2+a_2a_3+a_3a_1=1. \end{align} The discrete Laplace operator is $$ L_3=\left( \begin{array}{cccc} a_1+a_3 & -a_1 &-a_3 \\ -a_1 & a_1+a_2 &-a_2 \\ -a_3 & -a_2 & a_2+a_3 \end{array} \right). $$ The characteristic polynomial of $L_3$ is \begin{align*} P_3(x)=\det(L_3-x I_3)&=-x^3+2(a_1+a_2+a_3)x^2-3(a_1a_2+a_2a_3+a_3a_1)x\\ &=-x^3+2(a_1+a_2+a_3)x^2-3x \end{align*} by the equation (\ref{fm:3}). The eigenvalues of $L_3$ are denoted by $0=\lambda_0\leq \lambda_1 \leq \lambda_2.$ \subsection{} Therefore $\lambda_1+\lambda_2=2(a_1+a_2+a_3).$ We claim that $a_1+a_2+a_3\geq \sqrt{3}$ and the equality holds if and only if $\theta_1=\theta_2=\theta_3=\frac\pi 3.$ Consider $f:=a_1(\theta_1)+a_2(\theta_2)+a_3(\theta_3)$ as a function defined on the domain $$\Omega_3:=\{(\theta_1,\theta_2,\theta_3)\ | \ \theta_1+\theta_2+\theta_3=\pi, \theta_i>0, i=1,2,3 \}.$$ To find the absolute minimum of $f$, we apply the method of Lagrange multiplier. Let $$F=a_1(\theta_1)+a_2(\theta_2)+a_3(\theta_3)+y(\theta_1+\theta_2+\theta_3-\pi).$$ Since $$\frac{d a_i(\theta_i)}{d \theta_i}=-\frac{1}{\sin^2\theta_i}=-(1+a_i^2),$$ we have \begin{align*} 0=\frac{\partial F}{\partial \theta_1}&=-(1+a_1^2)+y,\\ 0=\frac{\partial F}{\partial \theta_2}&=-(1+a_2^2)+y,\\ 0=\frac{\partial F}{\partial \theta_3}&=-(1+a_3^2)+y,\\ 0=\frac{\partial F}{\partial y}&=\theta_1+\theta_2+\theta_3-\pi. \end{align*} Therefore the function $f$ has the unique critical point $(\theta_1,\theta_2,\theta_3)=(\frac\pi 3, \frac\pi 3, \frac\pi 3).$ Next, we investigate the behavior of the function $f$ when the variable $(\theta_1,\theta_2,\theta_3)$ approaches the boundary of the domain $\Omega_3$. Let $(\theta_1(t),\theta_2(t), \theta_3(t)), t\in [0,\infty),$ be a path in the domain $\Omega_3$. Let $a_i(t)=\cot \theta_1(t)$ for $i=1,2,3.$ Without loss of generality, we assume $$\lim_{t\to \infty} (\theta_1(t),\theta_2(t), \theta_3(t))=(0, s_2, s_3)$$ where $s_2\geq 0, s_3\geq 0$ and $s_2+s_3=\pi.$ Then $\lim_{t\to \infty}a_1(t)=\infty$ and $a_2(t)+a_3(t)>0$ for $t\in[0,\infty)$. Hence $$\lim_{t\to \infty}(a_1(t)+a_2(t)+a_3(t))=\infty.$$ If $\theta_i+\theta_j<\pi,$ then $\cot\theta_i>\cot(\pi-\theta_j)=-\cot\theta_j$. Therefore $a_i+a_j>0.$ Hence $2f=(a_1+a_2)+(a_2+a_3)+(a_3+a_1)>0.$ Thus $f$ has the absolute minimum. But the absolute minimum can not be achieved at a point in the boundary of $\Omega_3$. It must be achieved at the unique critical point $(\frac\pi 3, \frac\pi 3, \frac\pi 3).$ This shows that $a_1+a_2+a_3\geq \sqrt{3}$ and the equality holds if and only if $\theta_1=\theta_2=\theta_3=\frac\pi 3.$ \subsection{} Since $$\lambda_2=a_1+a_2+a_3+\sqrt{(a_1+a_2+a_3)^2-3},$$ we have $\lambda_2\geq \sqrt{3}$ and the equality holds if and only if $\theta_1=\theta_2=\theta_3=\frac\pi3$. Since $\lambda_1\lambda_2=3,$ we have $\lambda_1\leq \sqrt{3}$ and the equality holds if and only if $\theta_1=\theta_2=\theta_3=\frac\pi3$. \section{quadrilaterals} \subsection{} The vertices of a cyclic quadrilateral decompose its circumcircle into four arcs. We assume the radius of the circumcircle is 1 and the lengths of the four arcs are $2\theta_1,2\theta_2,2\theta_3,2\theta_4.$ Let $a_i:=\cot\theta_i$ for $i=1,...,4.$ The condition $\theta_1+\theta_2+\theta_3+\theta_4=\pi$ implies \begin{align}\label{fm:4} a_1a_2a_3+a_1a_2a_4+a_1a_3a_4+a_2a_3a_4=a_1+a_2+a_3+a_4. \end{align} There are two ways to decompose a cyclic quadrilateral in to a union of two triangles. The two ways produce the same discrete Laplace operator: $$ L_4=\left( \begin{array}{cccc} a_1+a_4 & -a_1 &0 & -a_4 \\ -a_1 & a_1+a_2 &-a_2 & 0 \\ 0 &-a_2 &a_2+a_3 & -a_3 \\ -a_4 & 0 &-a_3 & a_3+a_4 \end{array} \right). $$ The characteristic polynomial of $L_4$ is \begin{align*} P_4(x)&=x^4-2(a_1+a_2+a_3+a_4)x^3\\ &\hspace{70pt}+(3(a_1a_2+a_2a_3+a_3a_4+a_4a_1)+4(a_1a_3+a_2a_4))x^2\\ &\hspace{140pt}-4(a_1a_2a_3+a_1a_2a_4+a_1a_3a_4+a_2a_3a_4)x\\ &=x^4-2(a_1+a_2+a_3+a_4)x^3\\ &\hspace{70pt}+(3(a_1a_2+a_2a_3+a_3a_4+a_4a_1)+4(a_1a_3+a_2a_4))x^2\\ & \hspace{206pt} -4(a_1+a_2+a_3+a_4)x, \end{align*} by the equation (\ref{fm:4}). \subsection{} By the similar argument in the case of triangles, we can show that $a_1+a_2+a_3+a_4$ has the unique critical point at $(\frac\pi4,\frac\pi4,\frac\pi4,\frac\pi4).$ And we have $2(a_1+a_2+a_3+a_4)=(a_1+a_2)+(a_1+a_3)+(a_3+a_4)+(a_4+a_1)>0.$ Next, we investigate the behavior of the function $a_1+a_2+a_3+a_4$ when the variable $(\theta_1,\theta_2,\theta_3, \theta_4)$ approaches the boundary of the domain $$\Omega_4=\{(\theta_1,\theta_2,\theta_3, \theta_4)\ | \ \theta_1+\theta_2+\theta_3+\theta_4=\pi, \theta_i>0, i=1,2,3, 4 \}.$$ Let $(\theta_1(t),\theta_2(t), \theta_3(t), \theta_4(t))$, $t\in [0,\infty),$ be a path in the domain $\Omega_4$. Let $a_i(t)=\cot \theta_1(t)$ for $i=1,2,3,4.$ Without loss of generality, we assume $$\lim_{t\to \infty} (\theta_1(t),\theta_2(t), \theta_3(t), \theta_4(t))=(0, s_2, s_3, s_4)$$ where $s_i\geq 0$ for $i=2,3,4$ and $s_2+s_3+s_4=\pi.$ And we can assume that $s_2<\frac\pi2$ and $s_3<\frac\pi2$. Then $\lim_{t\to \infty} a_1(t)=\infty$, $a_2(t)>0$ when $t$ is sufficiently large and $a_3(t)+a_4(t)>0$ for any $t\in [0,\infty)$. Hence $$\lim_{t\to \infty} (a_1(t)+a_2(t)+a_3(t)+a_4(t))=\infty$$. Therefore $a_1+a_2+a_3+a_4$ achieves its absolute minimum at the unique critical point $(\frac\pi4,\frac\pi4,\frac\pi4,\frac\pi4).$ Hence $a_1+a_2+a_3+a_4\geq 4$ and the equality holds if and only if $\theta_1=\theta_2=\theta_3=\theta_4=\frac\pi4$. Therefore $\lambda_1+\lambda_2+\lambda_3\geq 8$, $\lambda_1\lambda_2\lambda_3\geq 16$ and the equality holds if and only if $\theta_1=\theta_2=\theta_3=\theta_4=\frac\pi4$. \subsection{} To verify the statement about $\lambda_1\lambda_2+\lambda_2\lambda_3+\lambda_3\lambda_1$, by the formula of the characteristic polynomial $P_4(x)$, it is enough to show $$3(a_1a_2+a_2a_3+a_3a_4+a_4a_1)+4(a_1a_3+a_2a_4)\geq 20$$ and the equality holds if and only if $\theta_1=\theta_2=\theta_3=\theta_4=\frac\pi4$. In fact, consider $g:=3(a_1a_2+a_2a_3+a_3a_4+a_4a_1)+4(a_1a_3+a_2a_4)$ as a function defined on the domain $\Omega_4.$ To find the absolute minimum of $g,$ we apply the method of Lagrange multiplier. Let $$G=3(a_1a_2+a_2a_3+a_3a_4+a_4a_1)+4(a_1a_3+a_2a_4)+y(\theta_1+\theta_2+\theta_3+\theta_4-\pi).$$ Then \begin{align*} 0=\frac{\partial G}{\partial \theta_1}&=-(3a_2+3a_4+4a_3)(1+a_1^2)+y,\\ 0=\frac{\partial G}{\partial \theta_2}&=-(3a_1+3a_3+4a_4)(1+a_2^2)+y,\\ 0=\frac{\partial G}{\partial \theta_3}&=-(3a_2+3a_4+4a_1)(1+a_3^2)+y,\\ 0=\frac{\partial G}{\partial \theta_4}&=-(3a_3+3a_1+4a_2)(1+a_4^2)+y,\\ 0=\frac{\partial G}{\partial y}&=\theta_1+\theta_2+\theta_3+\theta_4-\pi. \end{align*} The first and the third equation above imply that $$(3a_2+3a_4+4a_3)(1+a_1^2)=(3a_2+3a_4+4a_1)(1+a_3^2)$$ which is equivalent to $$(a_1-a_3)(3a_1a_2+3a_1a_4+3a_2a_3+3a_3a_4+4a_1a_3-4)=0.$$ We claim that the second factor is positive, i.e., $3a_1a_2+3a_1a_4+3a_2a_3+3a_3a_4+4a_1a_3>4.$ In fact, since $\theta_1+\theta_2+\theta_3<\pi,$ then $\cot(\theta_1+\theta_2)>\cot(\pi-\theta_3).$ Then $$\frac{a_1a_2-1}{a_1+a_2}>-a_3$$ which is equivalent to \begin{align}\label{fm:ine} a_1a_2+a_2a_3+a_3a_1>1 \end{align} since $a_1+a_2>0.$ By the similar reason, $$a_1a_4+a_4a_3+a_3a_1>1.$$ At least one of $a_2$ and $a_4$ is positive. If $a_2>0$, then \begin{align*} &3a_1a_2+3a_1a_4+3a_2a_3+3a_3a_4+4a_1a_3\\ &\hspace{50pt}=3(a_1a_4+a_4a_3+a_3a_1)+(a_1a_2+a_2a_3+a_3a_1)+2(a_1+a_3)a_2\\ &\hspace{50pt}>3+1+0. \end{align*} If $a_4>0$, then \begin{align*} &3a_1a_2+3a_1a_4+3a_2a_3+3a_3a_4+4a_1a_3\\ &\hspace{50pt}=(a_1a_4+a_4a_3+a_3a_1)+3(a_1a_2+a_2a_3+a_3a_1)+2(a_1+a_3)a_4\\ &\hspace{50pt}>1+3+0. \end{align*} Thus the only possibility is $a_1=a_3$ which implies $\theta_1=\theta_3$. By the similar argument, $0=\frac{\partial G}{\partial \theta_2}$ and $0=\frac{\partial G}{\partial \theta_4}$ imply $\theta_2=\theta_4.$ Since $\theta_1+\theta_2+\theta_3+\theta_4=\pi,$ we have $\theta_1+\theta_2=\frac\pi2$ which implies $a_1a_2=1.$ Now $0=\frac{\partial G}{\partial \theta_1}$ and $0=\frac{\partial G}{\partial \theta_2}$ imply $$(3a_2+3a_4+4a_3)(1+a_1^2)=(3a_1+3a_3+4a_4)(1+a_2^2).$$ Since $a_1=a_3$ and $a_2=a_4$, we have $$(6a_2+4a_1)(1+a_1^2)=(6a_1+4a_2)(1+a_2^2).$$ Since $a_1a_2=1,$ we have $$(a_1-a_2)(a_1^2+a_2^2+a_1a_2+1)=0.$$ Since the second factor satisfies $$a_1^2+a_2^2+a_1a_2+1=\frac12(a_1^2+a_2^2)+\frac12(a_1+a_2)^2+1>0,$$ the only possibility is $a_1=a_2.$ Therefore the function $g=3(a_1a_2+a_2a_3+a_3a_4+a_4a_1)+4(a_1a_3+a_2a_4)$ has the unique critical point $(\frac\pi4,\frac\pi4,\frac\pi4,\frac\pi4).$ Next, we claim that $g>0$. Since at least three of $a_1,a_2,a_3,a_4$ are positive, without loss of generality, we may assume that $a_1>0,a_2>0,a_3>0.$ Let's write $$g=2(a_1a_2+a_2a_4+a_4a_1)+2(a_2a_3+a_3a_4+a_4a_2)+(a_2+a_4)a_1+(a_1+a_4)a_3+4a_1a_3.$$ Then each term of sum above is positive. At last, we investigate the behavior of $g$ when the variable approaches the boundary of the domain $\Omega_4$. Let $(\theta_1(t),\theta_2(t), \theta_3(t), \theta_4(t))$, $t\in [0,\infty),$ be a path in the domain $\Omega_4$. Let $a_i(t)=\cot \theta_1(t)$ for $i=1,2,3,4.$ Without loss of generality, we have $$\lim_{t\to \infty} (\theta_1(t),\theta_2(t), \theta_3(t), \theta_4(t))=(0, s_2, s_3, s_4)$$ where $s_i\geq 0$ for $i=2,3,4$ and $s_2+s_3+s_4=\pi.$ And we can assume that $s_2<\frac\pi2$ and $s_3<\frac\pi2$. Let's write \begin{align*} g&=2(a_1(t)a_2(t)+a_2(t)a_4(t)+a_4(t)a_1(t))\\ &+2(a_2(t)a_3(t)+a_3(t)a_4(t)+a_4(t)a_2(t))\\ &+(a_2(t)+a_4(t))a_1(t)+(a_1(t)+a_4(t))a_3(t)+4a_1(t)a_3(t). \end{align*} By the inequality (\ref{fm:ine}), $$a_1(t)a_2(t)+a_2(t)a_4(t)+a_4(t)a_1(t)>1$$ and $$a_2(t)a_3(t)+a_3(t)a_4(t)+a_4(t)a_2(t)>1$$ for any $t\in [0, \infty)$. Since $a_2(t)+a_4(t)>0$ when $t$ is sufficiently large, $\lim_{t\to \infty}(a_2(t)+a_4(t))a_1(t)=\infty.$ And $(a_1(t)+a_4(t))a_3(t)>0, 4a_1(t)a_3(t)>0$ when $t$ is sufficiently large. Hence $g$ approaches $\infty.$ Therefore $g$ has a lower bound and can not achieve its absolute minimum at a boundary point. It much achieve its absolute minimum at the unique critical point $(\frac\pi4,\frac\pi4,\frac\pi4,\frac\pi4).$ \subsection{} We verify the statement about $\lambda_1$ in this subsection. First, we verify that $\lambda_1\leq 2$ as follows. Let \begin{align*} Q(x):=\frac{P_4(x)}x&=x^3-2(a_1+a_2+a_3+a_4)x^2\\ & \hspace{65pt}+(3(a_1a_2+a_2a_3+a_3a_4+a_4a_1)+4(a_1a_3+a_2a_4))x\\ & \hspace{190pt} -4(a_1+a_2+a_3+a_4). \end{align*} We have $Q(0)=-4(a_1+a_2+a_3+a_4)\leq -16.$ If $Q(2)> 0,$ then the first root of $Q(x)$ is less that $2$, i.e., $\lambda_1< 2.$ If $Q(2)\leq 0$, we claim that $Q'(0)>0$ and $Q'(2)\leq 0.$ Once the two statements are established, $\lambda_1\leq \lambda_2\leq 2.$ In fact $Q'(0)=3(a_1a_2+a_2a_3+a_3a_4+a_4a_1)+4(a_1a_3+a_2a_4)\geq 20.$ To verify $Q'(2)\leq 2,$ we need to use the assumption $Q(2)\leq 2$. In fact $Q(2)\leq 2$ implies $$3(a_1a_2+a_2a_3+a_3a_4+a_4a_1)+4(a_1a_3+a_2a_4)\leq 6(a_1+a_2+a_3+a_4)-4.$$ Now \begin{align*} &Q'(2)\\ &=12-8(a_1+a_2+a_3+a_4)+3(a_1a_2+a_2a_3+a_3a_4+a_4a_1)+4(a_1a_3+a_2a_4)\\ &\leq 12-8(a_1+a_2+a_3+a_4)+6(a_1+a_2+a_3+a_4)-4\\ &=8-2(a_1+a_2+a_3+a_4)\\ &\leq 0, \end{align*} since $a_1+a_2+a_3+a_4\geq 4.$ Second, we verify that $\lambda_1=2$ if and only if $\theta_1=\theta_2=\theta_3=\theta_4=\frac\pi4$. Since $\lambda_1=2$ is the first root of $Q(x)$, we have $Q'(2)\geq 0.$ On the other hand, it is shown that $Q(2)\leq 0$ implies $Q'(2)\leq 0.$ Hence the only possibility is $Q'(2)=0.$ This requires that $a_1+a_2+a_3+a_4=4$. Therefore we must have $\theta_1=\theta_2=\theta_3=\theta_4=\frac\pi4$. \section{general cyclic polygons} \subsection{} Assume $n\geq 5.$ The vertices of a cyclic $n$-gon decompose its circumcircle into $n$ arcs. We assume the radius of the circumcircle is 1 and the lengths of the $n$ arcs are $2\theta_1,2\theta_2,...,2\theta_n.$ The discrete Laplace operator of a cyclic $n$-gon is independent of the choice of a triangulation. It is $$ L_n=\left( \begin{array}{ccccccc} a_1+a_n & -a_1 &0 &0 &... & -a_n \\ -a_1 & a_1+a_2 &-a_2 &0 &... & 0 \\ 0 &-a_2 &a_2+a_3 & -a_3 &... & 0 \\ 0 & 0 &-a_3 & a_3+a_4 &... &0 \\ 0 &0 &0 & -a_4 &... &0 \\ \ & \ &\ & \ & \ddots &\ \\ -a_n & 0 &0 &0 &... &a_{n-1}+a_n \end{array} \right). $$ The eigenvalues are $0=\lambda_0\leq \lambda_1 \leq ... \leq \lambda_{n-1}.$ \subsection{} We have $\sum_{i=1}^{n-1}\lambda_i=2\sum_{i=1}^na_i.$ By the similar argument in the case of triangles and cyclic quadrilaterals, we can show that $\sum_{i=1}^na_i$ has the unique critical point $(\theta_1,...,\theta_n)=(\frac\pi n,...,\frac\pi n)$. Since there is at most one non-positive number in $a_1,...,a_n$, without loss of generality, we may assume $a_1>0,...,a_{n-1}>0.$ Since $a_{n-1}+a_{n}>0,$ we have $\sum_{i=1}^na_i>0.$ We investigate the behavior of $\sum_{i=1}^na_i$ when the variable approaches the boundary of the domain $$\Omega_n=\{(\theta_1,...,\theta_n)\ | \ \theta_1+...+\theta_n=\pi, \theta_i>0, i=1,...,n \}.$$ Let $(\theta_1(t),\theta_2(t), \theta_3(t),\theta(t))$, $t\in [0,\infty),$ be a path in the domain $\Omega_4$. Let $a_i(t)=\cot \theta_1(t)$ for $i=1,2,3,4.$ Without loss of generality, we have $$\lim_{t\to \infty} (\theta_1(t),...,\theta_n(t))=(0, s_2,..., s_n)$$ where $s_i\geq 0$ for $i=2,...,n$ and $s_2+...+s_n=\pi.$ And we can assume that $s_2<\frac\pi2,...,s_{n-1}<\frac\pi2.$ Since $a_i(t)>0$ for $i=2,...,n-1$ and $a_{n-1}+a_n>0$ when $t$ is sufficiently large, $\lim_{t\to \infty}a_1=\infty$ implies that $\lim_{t\to \infty} \sum_{i=1}^na_i=\infty.$ Thus $\sum_{i=1}^na_i$ achieved the absolute minimum at $(\frac\pi n,...,\frac\pi n)$. \subsection{} In this subsection we verify the statement about $\prod_{i=1}^{n-1}\lambda_i.$ \begin{2} Let $M$ be an $n$ by $n$ matrix. If the sum of the entries of each row or each column of $M$ vanishes, all principle $n-1$ by $n-1$ submatrices of $M$ have the same determinant, and this value is equal to $\frac 1n$ times the product of all nonzero eigenvalues of $M$. \end{2} For the reference of the weighted matrix-tree Theorem, for example, see \cite{LW}, page 450, Problem 34A or \cite{DKM}, Theorem 1.2. In our case, according the weighted matrix-tree Theorem, to calculate $\prod_{i=1}^{n-1}\lambda_i$ of the matrix $L_n$, it is enough to calculate a particular principle $n-1$ by $n-1$ matrix. \begin{lemma} Let $N_n$ be the submatrix obtained by deleting the first row and first column of the matrix $L_n$. Then $$\det N_n=\sum_{i=1}^n a_1...\widehat{a_{i}}...a_n,$$ where $\widehat{a_{i}}$ means that $a_i$ is missing. \end{lemma} \begin{proof} We prove the statement by the mathematical induction. It holds for $n=4$ as we see in the formula of the characteristic polynomial $P_4(x)$. We assume it holds for $n\leq m-1.$ By the property of tridiagonal matrices, we have $$\det N_m=(a_{m-1}+a_m)\det N_{m-1}-a_{m-1}^2\det N_{m-2}.$$ Then by the assumption of the induction, \begin{align*} \det N_m&=(a_{m-1}+a_m)\sum_{i=1}^{m-1} a_1...\widehat{a_{i}}...a_{m-1}-a_{m-1}^2\sum_{i=1}^{m-2} a_1...\widehat{a_{i}}...a_{m-2}\\ &=a_{m-1}\sum_{i=1}^{m-1} a_1...\widehat{a_{i}}...a_{m-1}-a_{m-1}\sum_{i=1}^{m-2} a_1...\widehat{a_{i}}...a_{m-2}a_{m-1}\\ &\hspace{200pt}+a_m\sum_{i=1}^{m-1} a_1...\widehat{a_{i}}...a_{m-1}\\ &=a_1...a_{m-1}+a_m\sum_{i=1}^{m-1} a_1...\widehat{a_{i}}...a_{m-1}\\ &=\sum_{i=1}^m a_1...\widehat{a_{i}}...a_m.\\ \end{align*} \end{proof} In the following, we prove that $\sum_{i=1}^n a_1...\widehat{a_{i}}...a_n$ achieves its absolute minimum when $\theta_1=...=\theta_n=\frac\pi n$. It is enough to show that \begin{itemize} \item[a.] $\sum_{i=1}^n a_1...\widehat{a_{i}}...a_n$ has the unique critical point $(\frac\pi n,...,\frac\pi n)$; \item[b.] $\sum_{i=1}^n a_1...\widehat{a_{i}}...a_n>0$; \item[c.] $\sum_{i=1}^n a_1...\widehat{a_{i}}...a_n$ approaches $\infty$ as the variable approaches the boundary of the domain $\Omega_n$. \end{itemize} When $n=4,$ since $a_1a_2a_3+a_1a_2a_4+a_1a_3a_4+a_2a_3a_4=a_1+a_2+a_3+a_4,$ the three statements above are already shown to be true in section 3. We assume that the three statements above hold when $n\leq m-1.$ Let's check the three statements when $n=m.$ Consider the function $$H=\sum_{i=1}^m a_1...\widehat{a_{i}}...a_m-y(\theta_1+...\theta_m-\pi).$$ Then $0=\frac{\partial H}{\partial \theta_1}$ and $0=\frac{\partial H}{\partial \theta_2}$ imply that $$(a_3...a_m+a_2\sum_{i=3}^m a_3...\widehat{a_{i}}...a_m)(1+a_1^2)=(a_3...a_m+a_1\sum_{i=3}^m a_3...\widehat{a_{i}}...a_m)(1+a_2^2).$$ Since $a_1+a_2>0,$ it is equivalent to $$(a_1-a_2)(a_1+a_2)(a_3...a_m+\frac{a_1a_2-1}{a_1+a_2}\sum_{i=3}^m a_3...\widehat{a_{i}}...a_m)=0.$$ The third factor is $$\cot\theta_3...\cot\theta_m+\cot(\theta_1+\theta_2)\sum_{i=3}^m\cot\theta_3...\widehat{\cot\theta_i}...\cot\theta_m$$ which is written as $\sum_{i=1}^m \widetilde{a}_1...\widehat{\widetilde{a}_i}...\widetilde{a}_{m-1}$, where $\widetilde{a_1}=\cot(\theta_1+\theta_2), \widetilde{a}_i=\cot\theta_{i+1}$ for $i=2,...,m-1.$ This expression corresponds to a cyclic $(m-1)$-gon. By assumption of the induction, $\sum_{i=1}^m \widetilde{a}_1...\widehat{\widetilde{a}_i}...\widetilde{a}_{m-1}>0.$ Hence the only possibility is $a_1=a_2.$ By similar argument, we show that $a_i=a_j$ for any $i,j.$ Hence the function $\sum_{i=1}^m a_1...\widehat{a_i}...a_m$ has the unique critical point such that $\theta_i=\frac\pi m$ for any $i=1,...,m.$ Next, we claim that $\sum_{i=1}^m a_1...\widehat{a_i}...a_m>0.$ Without loss of generality, we assume that $a_1>0, a_2>0, ..., a_{m-1}>0.$ Now \begin{align*} &\sum_{i=1}^m a_1...\widehat{a_i}...a_m\\ &=a_1a_2...a_{m-2}(a_{m-1}+a_m)+\sum_{i=1}^{m-2} a_1...\widehat{a_i}...a_{m-2}(a_{m-1}a_m)\\ &=a_1a_2...a_{m-2}(a_{m-1}+a_m)+\sum_{i=1}^{m-2} a_1...\widehat{a_i}...a_{m-2}(a_{m-1}a_m-1)+\sum_{i=1}^{m-2} a_1...\widehat{a_i}...a_{m-2}\\ &=(a_{m-1}+a_m)(a_1a_2...a_{m-2}+\sum_{i=1}^{m-2} a_1...\widehat{a_i}...a_{m-2}\frac{a_{m-1}a_m-1}{a_{m-1}+a_m})+\sum_{i=1}^{m-2} a_1...\widehat{a_i}...a_{m-2}. \end{align*} Let $\widetilde{a}_{m-1}=\frac{a_{m-1}a_m-1}{a_{m-1}+a_m}=\cot(\theta_{m-1}+\theta_m).$ Then \begin{align*} &a_1a_2...a_{m-2}+\sum_{i=1}^{m-2} a_1...\widehat{a_i}...a_{m-2}\frac{a_{m-1}a_m-1}{a_{m-1}+a_m}\\ &=a_1a_2...a_{m-2}+\sum_{i=1}^{m-2} a_1...\widehat{a_i}...a_{m-2}\widetilde{a}_{m-1}.\\ \end{align*} Consider an cyclic $(m-1)$-gon with angles $\theta_1,...,\theta_{m-2}, \theta_{m-1}+\theta_m.$ By the assumption of induction, $$a_1a_2...a_{m-2}+\sum_{i=1}^{m-2} a_1...\widehat{a_i}...a_{m-2}\widetilde{a}_{m-1}>0.$$ Therefore $\sum_{i=1}^m a_1...\widehat{a_i}...a_m>0.$ At last, we investigate the behavior of the function $\sum_{i=1}^m a_1...\widehat{a_i}...a_m$ when the variable approaches the boundary of the domain $$\Omega_m=\{(\theta_1,...,\theta_m)\ | \ \theta_1+...+\theta_m=\pi, \theta_i>0, i=1,...,m \}.$$ Let $(\theta_1(t),..., \theta_m(t))$, $t\in [0,\infty),$ be a path in the domain $\Omega_m$. Let $a_i(t)=\cot \theta_1(t)$ for $i=1,...,m.$ Without loss of generality, we assume $$\lim_{t\to \infty}(\theta_1(t),...,\theta_m(t))=(0,s_2,...,s_m),$$ where $s_2\geq 0,...,s_m \geq 0$ and $s_2+...+s_m=\pi.$ And we can assume furthermore that $s_2<\frac\pi2,...,s_{m-1}<\frac\pi2.$ Thus $a_1(t)>0,...,a_{m-1}(t)>0$ when $t$ is sufficiently large. To simplify the notation, we denote $a_i(t)$ by $a_i$ in the follows. Now \begin{align*} &\sum_{i=1}^m a_1...\widehat{a_i}...a_m\\ &=(a_{m-1}+a_m)(a_1a_2...a_{m-2}+\sum_{i=1}^{m-2} a_1...\widehat{a_i}...a_{m-2}\frac{a_{m-1}a_m-1}{a_{m-1}+a_m})+\sum_{i=1}^{m-2} a_1...\widehat{a_i}...a_{m-2}\\ &=(a_{m-1}+a_m)(a_1a_2...a_{m-2}+\sum_{i=1}^{m-2} a_1...\widehat{a_i}...a_{m-2}\widetilde{a}_{m-1})+\sum_{i=1}^{m-2} a_1...\widehat{a_i}...a_{m-2}, \end{align*} where $\widetilde{a}_{m-1}=\frac{a_{m-1}a_m-1}{a_{m-1}+a_m}=\cot(\theta_{m-1}+\theta_m).$ By the assumption of induction, $$a_1a_2...a_{m-2}+\sum_{i=1}^{m-2} a_1...\widehat{a_i}...a_{m-2}\widetilde{a}_{m-1}>0$$ for any $t\in [0,\infty)$. Since $a_{m-1}+a_m>0$ for any $t\in [0,\infty)$ and $\sum_{i=1}^{m-2} a_1...\widehat{a_i}...a_{m-2}$ approaches $\infty,$ we see that $\sum_{i=1}^m a_1...\widehat{a_i}...a_m$ approaches $\infty.$

https://arxiv.org/abs/2111.03817

Cubes and Boxes have Rupert's passages in every direction

It is a $300$ year old counterintuitive observation of Prince Rupert of Rhine that in cube a straight tunnel can be cut, through which a second congruent cube can be passed. Hundred years later P. Nieuwland generalized Rupert's problem and asked for the largest aspect ratio so that a larger homothetic copy of the same body can be passed. We show that cubes and in fact all rectangular boxes have Rupert's passages in every direction, which is not parallel to the faces. In case of the cube it was assumed without proof that the solution of the Nieuwland's problem is a tunnel perpendicular to the largest square contained by the cube. We prove that this unwarranted assumption is correct not only for the cube, but also for all other rectangular boxes.

\section{Rupert's passages and Nieuwland's constants} It is commonly agreed that Prince Rupert of the Rhine (1619-1682) was the first person, who posed the following puzzle.: \medskip \noindent Rupert's passage problem: {\it Cut a hole in a cube, through which another cube of the same size shall be able to pass.} \begin{figure}[h] \begin{center} \includegraphics[scale=.6]{Rupertcube2.pdf} \caption{ The cube is Rupert. a) is a visual proof from 1816 \cite{swinden16}; b) is a proof without words from today.} \label{Ruperttunnel} \end{center} \end{figure} The affirmative answer of Rupert's passage problem is attributed to P. Nieuwland and was published in a paper of J.H. van Swinden \cite{swinden16} in 1816. It is common to say that {\it the cube is Rupert} (see the survey papers \cite{rickey05} and \cite{jerrard04}). On Figure~\ref{Ruperttunnel}b the four points marked by dots partition their edges in the ratio $1 : 3$. Repeated use of Pythagorean theorem reveals that these four points form a square of edge length $\frac{3\sqrt 2}4 = 1.06 \dots$ allowing to cut a suitable perpendicular passage \cite{gardner01}. Rupert's passage problem sounds paradoxial at first. After realizing that the problem is about finding two shadows so that one shadow fits inside the other, the problem becomes manageable and also triggers generalizations. Which solids are Rupert? If a solid is Rupert, what is the maximum size of the second body so that it still can be moved through the first one? The second question was asked for cubes by P. Nieuwland (1764-1794) almost a century after Rupert's question was posed. The later problem is referred to as {\it Nieuwland's passage problem} and the homothetic constant corresponding to the largest size (or rather the supremum of the possible sizes) is called {\it Nieuwland's constant}. Several survey papers popularized generalizations over the years, including papers of J.H. van Swinden \cite{swinden16} in 1816, of A. Ehrenfeucht \cite{ehrenfeucht64} in 1964, of D.J.E. Schrek \cite{schrek50} in 1950, of M. Gardner \cite{gardner01} in 2001 and of V.F. Rickey \cite{rickey05} in 2005. Over the years various solids were studied concerning Rupert's passage problem. Spheres and bodies of constant widths are obviously not Rupert. Positive answers were found for tetrahedron and for octahedron by Scriba \cite{scriba68} in 1968, for rectangular boxes by Jerrard and Wetzl \cite{jerrard04} in 2004, for universal stoppers by Jerrard and Wetzl in \cite{jerrard08} in 2008, for dodecahedron and for icosahedron by Jerrard, Wetzl and Yuan \cite{jerrard17} in 2017, for eight Archimedean solids by Chai, Yuan and Zamfirescu \cite{chai18} in 2018 and for $n$-cubes by Huber, Schultz and Wetzl \cite{huber18} in 2018. The most intriguing open question in this area is wether every convex polyhedron is Rupert (see \cite{jerrard17}). \section{New results concerning Rupert's passages} Neither Rupert nor Nieuwland, were precise about how they want to pass a cube through the other one. According to standards at their time, using words like 'passage' and 'tunnel' was enough to indicate that motion is expected to be a translation. There is another assumption which requires rigorous proof. We noticed this detail in a paper of Jerrard and Wetzel who wrote several interesting papers on Rupert's passages. \cite{jerrard04} appeared in this Monthly, and is about the solution of Nieuwland's type problem for passing rectangular boxes through a unit cube. In the introduction the authors state a comonly used unwarranted assumption \noindent \begin{quote} {\it Nieuwland's passage problem of finding the largest cube that can pass through a unit cube is equivalent to finding the largest square that fits in the unit cube, because once the largest square is located, the hole through the cube having that largest square as its cross section clearly provides the desired passage.} \end{quote} \noindent Indeed, once the largest homothetic square is located, the hole raised perpendicularly over this square clearly provides the desired passage. The other direction of the equivalency statement can not be assumed without proof. It is very unlikely that Rupert and Nieuwland wanted to restrict passages to tunnels raised perpendicularly over squares contained in the unit cube. Thus, before this equivalency issue is settled the constant found by Nieuwland or later by other authors are only lower bounds for the Nieuwland's constants of the cube and of other solids. The main goals of this paper were i) to show that cubes are Rupert in every direction and ii) to show equivalency of the later two problems for cubes. Along the way we noticed that essentially the same arguments, with some modifications, prove the analogous theorems for all rectangular boxes. We will prove \begin{theorem} \label{rectangle-in-projection} The interior of every hexagonal projection of a rectangular box with sides $a \leq b \leq c$ contains a rectangle with sides $a$ and $b$. Equivalently, rectangular boxes have Rupert's passages in every direction not parallel to faces. In particular, cubes have Rupert's passages in every direction not parallel to faces. \end{theorem} Using Theorem~\ref{rectangle-in-projection} we also prove \begin{theorem} \label{largestsquare} Let $B$ be a rectangular box with edge lengths $a \leq b \leq c$. If a homothetic box $\lambda B$ ($\lambda >0$) can be passed through $B$ by translation, then the interior of box $B$ contains rectangle with sides $\lambda a$ and $\lambda b$. With other words, the problem of finding Nieuwland constant of a given rectangular box $B$, is equivalent to finding the supremum of $\lambda$'s for which the interior of box $B$ contains a $\lambda$-homothetic copy of the smallest face of box $B$. \end{theorem} \section{Preliminary Lemmas} In this section we prove three lemmas: \begin{lemma} \label{cube-basics} Let $C$ be a unit cube in a general position in the $xyz$-coordinate space. Denote by $H$ the perpendicular projection of $C$ on the $xy$-plane (Figure~\ref{basic}). Then, \begin{enumerate} \item the projection $H$ is either a central symmetrical hexagon with obtuse angles or a rectangle, \item the area of $H$ is equal to the length of the perpendicular projection of $C$ on the $z$-axis, \item $p^2+q^2 +r^2=1$, where $p,q$ and $r$ are the third coordinates of the three non-parallel normal vectors of the faces of $C$. Similar equations hold for the second and for the first coordinates. \end{enumerate} \end{lemma} \begin{proof} [Proof of Lemma \ref{cube-basics}] Statement 1 follows from the visual geometric observation that in an infinite wedge bounded by two half planes with an acute angle, no solid cube can reach the edge of the wedge on its convex side. Indeed, if this would be possible, then a continuity argument would imply that the contact is possible also with one face of the cube on one of the bounding half planes, a contradiction. \begin{figure}[h] \begin{center} \includegraphics[scale=.5]{cubebasics.pdf} \caption{Basic properties of the projections of a cube.} \label{basic} \end{center} \end{figure} Statement 2 is a well known property of the cube. Figure~\ref{basic} becomes a proof without words, once it is noticed that the dot product of a unit edge vector and the vertical unit vector has two geometric meanings. On the one hand, it is equal to the area of the shadow of that face which is perpendicular to the unit edge vector. On the other hand, it is equal to the length of the perpendicular projection of the unit edge vector on the $z$-axis. Statement 3 must be also an old known property of the cube. Here we present a short proof via matrices. Let $\overrightarrow{P} , \overrightarrow{Q}$ and $\overrightarrow{R}$ be the three normal vectors with third coordinates p, q and r. Let $A$ be the matrix whose row vectors are these normal vectors. The linear transformation $A$ is an isometry so for all column vectors $\alpha , \beta \in \mathcal R^3$, $(A \alpha )^T \cdot (A\beta ) = \alpha^T \cdot \beta$. Equivalently, $(\alpha^T A^T) \cdot (A\beta ) = \alpha^T \cdot \beta$. Thus, $A^T A =I $, which in view of the row-column multiplication means that the column vectors of $A$ form an orthonormal basis. In particular, the third column vector is a unit vector, thus $p^2+q^2 +r^2=1$. \end{proof} Next, we will show Lemma \ref{squareatcorner}, which not only claims that the interior of every hexagonal shadow of a unit cube contains a unit square, but also explains, where such squares are located within a shadow. \begin{lemma} \label{squareatcorner} Let $H$ be a hexagonal projection of a unit cube (Figure~\ref{projections2}). Then, there are at least two pairs of opposite vertices of the hexagon $H$ so that each of these four (or six) vertices can share a vertex of a unit square in $B$. Moreover, each of these four (or six) squares are unique and have the following properties: \begin{itemize} \item The two vertices adjacent to the 'corner' vertex, lie on a pair of opposite sides of $H$. \item The fourth vertices of these squares belong to interior of $H$. \item Each of these squares can be moved in the interior of the hexagon $H$. \end{itemize} \end{lemma} \begin{remark} First we used the GeoGebra software to see how does the shadow of a cube change when a cube is rotated in $3$-space. The experiment made us believe that there might exists a unit square in every shadow, so that the sqaure shares a vertex with the projection of the cube. The conjectured special position allowed us to get the affirmative answer by carrying out a straightforward four page long computation, using two variables only. The simple proof bellow is different, it is adjusted to the box version and it reveals more geometric reasons. \end{remark} \begin{figure}[h] \begin{center} \includegraphics[scale=.5]{square-in-projection.pdf} \caption{Every projection of a unit cube contains a unit square.} \label{projections2} \end{center} \end{figure} \begin{proof}[Proof of Lemma \ref{squareatcorner}] Let $A$ be one of the vertices of the hexagonal projection $H$ of the cube. Let $V$ be that vertex of the cube whose vertical projection is $A$. Let us label the three unit edge vectors emanating from $V$ by $\overrightarrow{P}, \overrightarrow{Q}$ and $\overrightarrow{R}$. If $p,q$ and $r$ denote the absolute values of the $z$-coordinates of these vectors, then the lengths of the projections of these vectors are $\sqrt{1-p^2}, \sqrt{1-q^2}$ and $\sqrt{1-r^2}$. Let us orient the hexagon $H$ so that one pair of opposite sides is horizontal and the lengths of these sides equal to $\sqrt{1-p^2}$. Since all angles of $H$ are obtuse (Part 1 of Lemma~\ref{cube-basics}) we may assume the labelling of side lengths and the areas of the sub-parallelograms are as shown on Figure~\ref{projections2}. In order to do a little computation, we introduce an $xy$-coordinate system centered at $A$ so that the horizontal line is the $x$ coordinate axis. Let $B'$ and $C'$ be the adjacent vertices of $A$ in $H$. Choose points $B$ and $C$ on the horizontal sides of $H$ at distance $1$ from $A$. Let $B$ the point at distance $b$ from $B'$ and $B'$ at distance $b'$ from the $y$-axis. Distances $c$ and $c'$ are introduced analogously for points $C$ and $C'$. We have \noindent $(b+b')^2 = 1^2-(\frac r {\sqrt{1- p^2}})^2 = 1- \frac {r^2}{1-p^2} = \frac {q^2}{1-p^2}$, \noindent $(b')^2 = 1- q^2 - (\frac r {\sqrt{1- p^2}})^2 = \frac {q^2}{1-p^2} - q^2 = \frac {p^2 q^2}{1-p^2}$, \noindent $b = \frac q {\sqrt{1-p^2}} (1-p) = q \sqrt \frac {1-p}{1+p}$. \noindent Similar expression holds for $c$: $c = r \sqrt {\frac {1-p}{1+p}}$. \noindent Thus, we have the following three equivalent inequalities, \noindent $b+c < \sqrt {(1-p^2)} \Leftrightarrow (q+r) \sqrt {\frac {1-p}{1+p}} < \sqrt{1-p^2} \Leftrightarrow q+r < 1 +p$. Let $D$ be the vertex which completes $A,B,C$ to a square. Since both $H$ and $ABCD$ are centrally symmetric, $D$ lies on the horizontal line through the vertex of $H$ opposite to $A$. Now, the first and the last equivalent inequalities mean that $D$ is inside of $H$ if $q+r \leq 1 +p$. Since all $p, q, r \in (0,1)$, this holds, if $p$ is the largest or the second largest among $p,q,r$, which proves the first two properties listed in Lemma~\ref{squareatcorner}. A small shift followed by a small rotation of these unit square ensures that there is a unit square which lies in the interior of the hexagon. \end{proof} \begin{definition} An infinite vertical cylinder, which is raised over a horizontal rectangle will be called {\it rectangular tube}. \end{definition} We will need \begin{lemma} \label{squaretube} Every planar cross section of a rectangular tube contains a rectangle congruent to its base. \end{lemma} \begin{remark}This lemma is a special case of a very strong theorem of K\'os and T\"or\H ocsik \cite{koos90}. They proved that every convex disc covers its shadow. Here we present a new short proof for Lemma~\ref{squaretube}, so that this paper remains self contained. \end{remark} \begin{figure}[h] \begin{center} \includegraphics[scale=.8]{xy-setup.pdf} \caption{The plane of cross section folded over the horizontal plane.} \label{xy} \end{center} \end{figure} \begin{proof}[Proof of Lemma \ref{squaretube}] Let $R$ be a rectangle of sides $a$ and $b$, $a\leq b$. If the plane of the cross section is horizontal, then Lemma~\ref{squaretube} is obviously true. Otherwise, denote the angle between the plane of the tube's base $S$ and the plane of the cross section $P$ by $\alpha$ (Figure~\ref{xy}). $P$ is a parallelogram with opposite sides at least at distances $a$ and $b$ respectively. Introduce an $x,y$ coordinate system in the horizontal plane of $R$ so that the $x$ axis belongs also to the plane of $P$. Let us fold the plane of the cross section into the horizontal plane around the $x$-axis. It is clear that there is a congruent image $P^*$ of $P$ so that the linear map $(x,y) \rightarrow (x, \frac 1{\cos \alpha} y)$ takes the rectangle $R$ to parallelogram $P^*$. We denote the image of a point $p$ under this mapping by $p^*$. Let us label the vertices of rectangle $R$ by $1,2,3$ and $4$ in the order of their $y$ coordinates (Figure 1). It is easy to see that $\frac 1{\cos \alpha} > 1$ implies that $P^*$ has obtuse angles at vertices $2^*$ and $3^*$. Depending on how the diagonal $2^* 3^*$ splits the obtuse angle at $2^*$, we distinguish two cases: \noindent Case 1. Both sub-angles at $2^*$ are greater than equal to the corresponding sub-angles at $2$. In this case a centrally positioned copy of $R$ inside of $P^*$, so that one of the diagonal of $R$ lies on $2^* 3^*$, is contained in the parallelogram $P^*$. \noindent Case 2. One of the sub-angles, say $\angle 3^* 2^* 4^*$, is less than the corresponding sub-angle at $2$. In this case, the length of the side $2^* 4^*$ is greater than the length of side $24$, which is $a$. Since the distance between the sides $2^* 4^*$ and $1^* 3^*$ is at least $b$, the rectangle, which has a vertex at $2^*$ and has a side of length $a$ on the segment $2^* 4^*$ is contained in the parallelogram $P^*$. \end{proof} \section{Proof of Theorems} \begin{proof} [Proof of Theorem~\ref{rectangle-in-projection}] Let $B$ be a rectangular box with edge lengths $a \leq b \leq c$. Shorten the longest edges of $B$ to $b$. It is enough to show that the new box $B'$ contains a rectangle of dimensions $a,b$. Let hexagon $H$ be a shadow of $B'$. Two of the six sides of $H$ are projections of edges of length $a$, while the remaining four sides are projections of edges of length $b$. To depict this property we label the sides and their lengthes with $pr(a)$ or $pr(b)$ on Figure~\ref{projections}. We have two pairs of opposite vertices which are endpoints of $pr(a)$ sides. In view of Lemma~\ref{squareatcorner} at least one of these two pairs have vertices, say $A$ and $A'$, so that moving a cube with edge length $a$ to the corner of the pre-image of $A$, the projection of the cube contain a square whose vertex is $A$. Now it is easy to see that one can extend one pair of parallel sides of this square to $b$ so that the new rectangle is still contained in $H$, what we wanted. A small shift followed by a small rotation of this rectangle ensures that there is a unit square which lies in the interior of the hexagon. \end{proof} \begin{figure}[h] \begin{center} \includegraphics[scale=.6]{cube-application.pdf} \caption{Every projection of a box contains a copy of the smallest face. } \label{projections} \end{center} \end{figure} \begin{proof} [Proof of Theorem~\ref{largestsquare}] Assume that a homothetic box $\lambda B$ can be passed through a stationary box $B$ with sides $a \leq b \leq c$. Although Theorem~\ref{largestsquare} is obviously true if $\lambda \leq 1$, we present an argument which works for all $\lambda$. On Figure~\ref{largest}, which illustrates the proof, we choose a $\lambda$ much smaller than $1$ so that the figure becomes less crowded and the proof can be better understood. Let $H_1$ be the shadow of the stationary rectangular box $B$ and $H_{\lambda}$ be the shadow of the traveling homothetic rectangular box $\lambda B$. We proved already that the hexagon $H_{\lambda}$ contains a rectangle $\lambda R$ of sides $\lambda a$ and $\lambda b$. Since both $H_1$ and $\lambda R$ are central symmetric, the rectangle $\lambda R$ can be shifted in center position while remaining inside of the hexagon $H_1$. Now, choose two adjacent vertices, say $U$ and $V$, of this relocated rectangle. Since $U, V$ are inside of the shadow of $B$, we can choose two points $U^*$, $V^*$ inside of the box $B$ directly over $U$ and $V$. Finally, reflect $U^*$, $V^*$ through the center of the box $B$ two get a parallelogram $P$ in box $B$. $P$'s shadow is the rectangle of sides $a$ and $b$. According to Lemma~\ref{squaretube}, parallelogram $P$ contains a rectangle of sides $a$ and $b$. \end{proof} \begin{figure}[h] \begin{center} \includegraphics[scale=.8]{boxequivalencylemma.pdf} \caption{Proof of Theorem~\ref{largestsquare}.} \label{largest} \end{center} \end{figure}

https://arxiv.org/abs/2202.06718

Various New Inequalities for Beta Distributions

This note provides some new inequalities and approximations for beta distributions, including tail inequalities, exponential inequalities of Hoeffding and Bernstein type, Gaussian inequalities and approximations.

\section{Introduction} Beta distributions play an important role in statistics and probability theory (Gupta and Nadarajah, 2004), and they occur in various scientific fields (Skorski, 2021). A frequent obstacle in problems involving beta distributions is the lack of analytic expressions for their distribution function, the normalized incomplete beta function. Therefore one often resorts to inequalities and approximations, as, for example, in the proofs of Dimitriadis et al.\ (2022, Theorem~4.1) and D{\"u}mbgen and Wellner (2022, Lemma~S.8). This paper provides some new inequalities for the beta distribution $\mathrm{Beta}(a,b)$ with parameters $a, b > 0$, its distribution function $B_{a,b}$, survival function $\bar{B}_{a,b} = 1 - B_{a,b}$ and density function $\beta_{a,b}$ on $[0,1]$. The latter is given by \[ \beta_{a,b}(x) \ := \ B(a,b)^{-1} x^{a-1} (1 - x)^{b-1} , \quad x \in (0,1) , \] wheere $B(a,b) := \int_0^1 x^{a-1}(1 - x)^{b-1} \, \d x = \Gamma(a) \Gamma(b)/\Gamma(a+b)$, and $\Gamma(\cdot)$ denotes the gamma function. In Section~\ref{sec:Segura}, we refine the lower and upper bounds for $B_{a,b}$ and $\bar{B}_{a,b}$ by Segura~(2016) which are particularly accurate in the tails of $\mathrm{Beta}(a,b)$. As a by-product we obtain refinements of Segura's~(2014) bounds for the gamma distribution and survival functions. In Section~\ref{sec:Exponential} we present new exponential inequalities which are stronger than previously known inequalities of D{\"u}mbgen~(1998), Marchal and Arbel~(2017) and Skorski~(2021). Section~\ref{sec:Gaussian.tail.inequalities} presents inequalities for $B_{a,b}$ and $\bar{B}_{a,b}$ in terms of Gaussian distribution functions. Finally, Section~\ref{sec:Gaussian.approximation.Beta(a,a)} discusses the approximation of the symmetric distribution $\beta_{a,a}$ by Gaussian densities with mean $1/2$ in the spirit of D{\"u}mbgen et al.\ (2021). Most proofs are deferred to Section~\ref{sec:Proofs}. \section{Sharp tail inequalities} \label{sec:Segura} In what follows, let $p := a/(a+b)$, the mean of $\mathrm{Beta}(a,b)$. In a general setting including noncentral beta distributions, Segura (2016, inequalities (27), (29), (30)) uses extensions of l'Hopital's rule to derive inequalities for $B_{a,b}$ and $\bar{B}_{a,b}$. For symmetry reasons, we only consider $B_{a,b}$, because $\bar{B}_{a,b}(x) = B_{b,a}(1-x)$. We rephrase Segura's inequalities in terms of the ratio \[ Q_{a,b}(x) \ := \ \frac{B_{a,b}(x)}{x^a/[a B(a,b)]} . \] This is motivated by the fact that $\beta_{a,b}(x) = B(a,b)^{-1} x^{a-1} (1 + O(x))$ and thus $B_{a,b}(x) = x^a/[a B(a,b)] (1 + O(x))$ as $x \to 0$. The goal is to find upper and lower bounds for $Q_{a,b}(x)$ approaching $1$ as $x \to 0$. Now, for $x \in (0,1)$, \begin{equation} \label{ineq:Segura.left} (1 - x)^b (1 + c_{a,b} x) \ \le \ Q_{a,b}(x) \ \le \ \frac{(1 - x)^b}{(1 - x/p)^+} , \\ \end{equation} where $c_{a,b} := (a+b)/(a+1)$. Numerical examples reveal that these inequalities are rather accurate unless $x$ is close to or larger than $p$. Our first contribution is an improvement of Segura's bounds. In particular, the upper bound remains valid if $x/p$ is replaced with the strictly smaller term $\max\{c_{a,b},1\} x$. The results are stated in terms of the following auxiliary functions: \begin{align*} q_{a,b}^{(1)}(x) \ &:= \ \Bigl( 1 - \frac{ax}{a+1} \Bigr)^{b-1} , \\ q_{a,b}^{(2)}(x) \ &:= \ \frac{a(1-x)^{b-1} + 1}{a+1} - \frac{a(b-1)(b-2)x^2(1-x)^{(b-3)^+}}{2(a+1)(a+2)} , \\ q_{a,b}^{(3)}(x) \ &:= \ \frac{(1-x)^{b}}{(1 - c_{a,b}x)^+} . \end{align*} \begin{Theorem} \label{thm:Segura} For $x \in (0,1)$, \[ (1 - x)^b (1 + c_{a,b} x) \ < \ Q_{a,b}^L(x) \ \le \ Q_{a,b}^{}(x) \ \le \ Q_{a,b}^U(x) \ \le \ \frac{(1 - x)^b}{(1 - \max\{c_{a,b},1\}x)^+} , \] where \begin{align*} Q_{a,b}^L(x) \ &:= \ \begin{cases} q_{a,b}^{(1)}(x) & \text{if} \ \ b \not\in (1,2) , \\ q_{a,b}^{(2)}(x) & \text{if} \ \ b \in [1,2] , \end{cases} \\ Q_{a,b}^U(x) \ &:= \ \begin{cases} q_{a,b}^{(2)}(x) & \text{if} \ \ b \le 1 , \\ q_{a,b}^{(1)}(x) & \text{if} \ \ b \in [1,2] , \\ \min \bigl\{ q_{a,b}^{(2)}(x), q_{a,b}^{(3)}(x) \bigr\} & \text{if} \ \ b > 2 . \end{cases} \end{align*} \end{Theorem} Figure~\ref{fig:Segura} illustrates the bounds for $B_{a,b}$ resulting from \eqref{ineq:Segura.left} and Theorem~\ref{thm:Segura} in case of $(a,b) = (4,8), (2,0.5)$. \begin{figure} \centering \includegraphics[width=0.85\textwidth]{Segura_4_8} \includegraphics[width=0.85\textwidth]{Segura_2_p5} \caption{Inequalities for $B_{a,b}$ when $(a,b) = (4,8)$ (upper panel) and $(a,b) = (2,0.5)$ (lower panel). The green line shows $B_{a,b}$, the blue lines are Segura's (2016) bounds resulting from \eqref{ineq:Segura.left}, and the black lines are the bounds via Theorem~\ref{thm:Segura}. The vertical line indicates the mean $p$.} \label{fig:Segura} \end{figure} \begin{Remark} The new inequalites for $Q_{a,b}$ are equalities in the case of $b \in \{1,2\}$, because $q_{a,1}^{(1)}(x) = q_{a,1}^{(2)}(x) = Q_{a,1}(x) = 1$ and $q_{a,2}^{(1)}(x) = q_{a,2}^{(2)}(x) = Q_{a,2}(x) = 1 - ax/(a+1)$. Moreover, the upper bound is exact for $b = 3$, because $q_{a,3}^{(2)}(x) = Q_{a,3}(x) = 1 - 2ax/(a+1) + ax^2/(a+2)$. \end{Remark} \begin{Remark} Note that the ratio $Q_{a,b}$ as well as the bounds $Q_{a,b}^L, Q_{a,b}^U$ are equal to $1 - d_{a,b} x + O(x^2)$ as $x \to 0$, where $d_{a,b} = (b-1) a/(a+1)$. The lower bound in \eqref{ineq:Segura.left} has the same property, but the upper bound does not. \end{Remark} \paragraph{Gamma distributions.} There is a rich literature about inequalities for gamma distribution and survival functions, see, for instance, Qi and Mei~(1999), Neuman~(2013), Segura~(2014) and Pinelis~(2020). We just illustrate that our bounds in Theorem~\ref{thm:Segura} yield a connection to that literature. It is well-known that for a random variable $X_{a,b} \sim \mathrm{Beta}(a,b)$, the rescaled variable $b X_{a,b}$ converges in distribution to a gamma random variable with shape parameter $a$ and scale parameter $1$ as $b \to \infty$. Denoting the corresponding distribution and survival function with $G_a$ and $\bar{G}_a = 1 - G_a$, respectively, we have $G_a(x) = \lim_{b\to \infty} B_{a,b}(x/b)$, and one can deduce from Theorem~\ref{thm:Segura} the following bounds. \begin{Corollary} \label{cor:Segura} For $a, x > 0$, \begin{align*} \frac{x^a e^{-ax/(a+1)}}{a\Gamma(a)} \ \le \ G_a(x) \ &\le \ \frac{x^a}{a\Gamma(a)} \cdot \begin{cases} \displaystyle \frac{a e^{-x} + 1}{a + 1} - \frac{a x^2 e^{-x}}{2(a+1)(a+2)} , \\ \displaystyle \frac{e^{-x}}{(1 - x/(a+1))^+} , \end{cases} \\[1ex] \frac{(x + 1_{[a \not\in (1,2)]})^{a-1} e^{-x}}{\Gamma(a)} \ \le \bar{G}_a(x) \ &\le \ \begin{cases} \displaystyle \frac{(x + 1_{[a > 1]})^{a-1} e^{-x}}{\Gamma(a)} & \text{if} \ a \le 2 , \\[1.5ex] \displaystyle \frac{x^a e^{-x}}{\Gamma(a) (x - a+1)^+} & \text{if} \ a > 2 , \\[2ex] \displaystyle e^{-x}(x^2/2 + x + 1) & \text{if} \ a = 3 . \\ \end{cases} \end{align*} \end{Corollary} The lower bound for $G_a(x)$ is already known from Neuman~(2013, Theorem~4.1), and the upper bounds for $G_a(x)$ are a combination of Segura~(2014, Theorem~10, part~3) and a slight improvement of Neuman~(2013, Theorem~4.1). The lower bounds for $\bar{G}_a(x)$ are equalities if $a \in \{1,2\}$, and the upper bounds if $a \in \{1,2,3\}$. Our lower bound for $\bar{G}_a(x)$ extends the lower bound of Segura~(2014, Theorem~10, part~4) to $a < 1$, and it is stronger than the latter for $a > 2$. Our upper bound for $\bar{G}_a(x)$ extends the upper bound of Segura~(2014, Theorem~10, part~6) to $a < 1$, and it is stronger than the latter if $1 < a \le 2$. \section{Exponential inequalities} \label{sec:Exponential} Although the upper bounds in Theorem~\ref{thm:Segura} are numerically rather accurate in the tails, they can diverge to $\infty$ at $x = p$ as $a,b \to \infty$. Moreover, it is sometimes desirable to have bounds for $\log B_{a,b}(x)$ and $\log \bar{B}_{a,b}(x)$ in terms of simple, maybe rational, functions of $x$. Numerous exponential tail inequalities for $B_{a,b}$ and $\bar{B}_{a,b}$ have been derived already. We start with one particular result of D{\"u}mbgen (1998, Proposition~2.1). For $x \in [0,1]$ let \[ K(p,x) \ := \ p \log \Bigl( \frac{p}{x} \Bigr) + (1 - p) \log \Bigl( \frac{1 - p}{1 - x} \Bigr) \in [0,\infty] . \] This function $K(p,\cdot)$ is strictly convex with minimum $K(p,p) = 0$. For arbitrary $x \in [0,1]$, \begin{equation} \label{ineq:Duembgen} \frac{x^a (1-x)^b}{p^a (1-p)^b} = \exp \bigl( - (a+b) K(p,x) \bigr) \ \ge \ \begin{cases} B_{a,b}(x) & \text{if} \ x \le p , \\ \bar{B}_{a,b}(x) & \text{if} \ x \ge p . \end{cases} \end{equation} In case of $a \ge 1$ or $b \ge 1$, these inequalities can be improved as follows. \begin{Theorem} \label{thm:Beta.expo} Suppose that $a \ge 1$. Then for $p_r := (a-1)/(a+b-1) < p$ and $x \in [p_r,1]$, \[ \bar{B}_{a,b}(x) \ \begin{cases} \displaystyle \le \ \frac{x^{a-1} (1-x)^b}{p_r^{a-1} (1-p_r)^b} \ = \ \exp \bigl( - (a+b-1) K(p_r, x) \bigr) , \\[2ex] \displaystyle \ge \ \frac{x^{a-1}(1-x)^b}{b B(a,b)} . \end{cases} \] Suppose that $b \ge 1$. Then for $p_\ell := a/(a+b-1) > p$ and $x \in [0,p_\ell]$, \[ B_{a,b}(x) \ \begin{cases} \displaystyle \le \ \frac{x^a (1-x)^{b-1}}{p_\ell^a (1-p_\ell)^{b-1}} \ = \ \exp \bigl( - (a+b-1) K(p_\ell, x) \bigr) , \\[2ex] \displaystyle \ge \ \frac{x^a(1-x)^{b-1}}{a B(a,b)} . \end{cases} \] \end{Theorem} \begin{Remark} At first glance, the upper bounds in Theorem~\ref{thm:Beta.expo} seem to be weaker than the ones in \eqref{ineq:Duembgen}, at least in the tail regions, because the factor $a+b-1$ is strictly smaller than $a+b$. But elementary algebra reveals that in case of $a \ge 1$, \begin{align*} (&a+b-1) K(p_r,x) - (a+b) K(p,x) \\ &= \ \log \Bigl( \frac{x}{p} \Bigr) + (a+b-1) \log \Bigl( 1 + \frac{1}{a+b-1} \Bigr) - (a-1) \log \Bigl( 1 + \frac{1}{a-1} \Bigr) \\ &> \ 0 \quad \text{for} \ x \in [p,1) , \end{align*} because $h(y) := y \log(1 + 1/y)$ (with $h(0) := 0$) is strictly increasing in $y \ge 0$. Analogously, if $b \ge 1$, then \begin{align*} (&a+b-1) K(p_\ell,x) - (a+b) K(p,x) \\ &= \ \log \Bigl( \frac{1-x}{1-p} \Bigr) + (a+b-1) \log \Bigl( 1 + \frac{1}{a+b-1} \Bigr) - (b-1) \log \Bigl( 1 + \frac{1}{b-1} \Bigr) \\ &> \ 0 \quad \text{for} \ x \in (0,p] . \end{align*} Thus the bounds in Theorem~\ref{thm:Beta.expo} are strictly smaller than the bounds in \eqref{ineq:Duembgen}. This is illustrated in Figure~\ref{fig:Beta.expo} for $(a,b) = (4,8)$. \end{Remark} \begin{figure} \centering \includegraphics[width=0.85\textwidth]{Beta_expo_4_8} \caption{Exponential tail inequalites for $\mathrm{Beta}(a,b)$ when $(a,b) = (4,8)$. The green line shows $\bar{B}_{a,b}$, the black line is its upper bound from Theorem~\ref{thm:Beta.expo}, and the blue line is its upper bound from \eqref{ineq:Duembgen}. One also sees the distribution function $B_{a,b}$ and its bounds as dotted lines. The additional red line is the upper bound \eqref{eq:beta.expo.median} from Remark~\ref{rem:Median}.} \label{fig:Beta.expo} \end{figure} \begin{Remark} \label{rem:Median} The upper bound for $\bar{B}_{a,b}$ in Theorem~\ref{thm:Beta.expo} can be improved substantially if $1 \le a \le b$. Indeed, the proof of Theorem~\ref{thm:Beta.expo} shows that for arbitrary $0 < x_o \le x \le 1$, \[ \bar{B}_{a,b}(x) \ \le \ \frac{\bar{B}_{a,b}(x_o)}{x_o^{a-1}(1-x_o)^b} x^{a-1}(1-x)^b. \] Specifically, it is well-known that $\mathrm{Median}(\mathrm{Beta}(a,b)) \le p$, see Groeneveld and Meeden (1977), so $\bar{B}_{a,b}(p) \le 1/2$ and for $x \in [p,1]$, \begin{equation} \label{eq:beta.expo.median} \bar{B}_{a,b}(x) \ \le \ \frac{x^{a-1}(1-x)^b}{2p^{a-1}(1-p)^b} . \end{equation} The latter bound is strictly smaller than the upper bound of Theorem~\ref{thm:Beta.expo} (restricted to $x \in [p,1]$), provided that $2 p^{a-1} (1 - p)^b > p_r^{a-1} (1 - p_r)^b$, and this is equivalent to $h(a-1) > h(a+b-1) - \log(2)$ with the increasing function $h(y) = y \log(1 + 1/y)$, $y > 0$. Since $h(a+b-1) < \lim_{y\to \infty} h(y) = 1$, a sufficient condition is that $h(a-1) \ge 1 - \log(2)$, which is fulfilled for $a \ge 1.152$. \end{Remark} The inequalities in Theorem~\ref{thm:Beta.expo} imply Bernstein and Hoeffding type exponential inequalities. It follows from D\"umbgen and Wellner (2022, Lemma~S.12) and the well-known inequality $z(1-z) \le 1/4$ for $z \in \mathbb{R}$, that \begin{equation} \label{ineq:K} K(q,x) \ \ge \ \frac{(x - q)^2}{2 (2x/3 + q/3)(1 - 2x/3 - q/3)} \ \ge \ 2 (x - q)^2 \end{equation} for $q,x \in [0,1]$, where $K(0,x) := -\log(1 - x)$ and $K(1,x) := -\log(x)$. This leads to the following inequalities: \begin{Corollary} \label{cor:Beta.expo} If $a \ge 1$, then for $x \in [p_r,1]$, \begin{align*} \bar{B}_{a,b}(x) \ &\le \ \exp \Bigl( - \frac{(a+b-1) (x - p_r)^2} {2(2x/3 + p_r/3)(1 - 2x/3 - p_r/3)} \Bigr) \\ &\le \ \exp \bigl( - 2(a+b-1) (x-p_r)^2 \bigr) . \end{align*} If $b \ge 1$, then for $x \in [0,p_\ell]$, \begin{align*} B_{a,b}(x) \ &\le \ \exp \Bigl( - (a+b-1) \frac{(x - p_\ell)^2} {2(2x/3 + p_\ell/3)(1 - 2x/3 - p_\ell/3)} \Bigr) \\ &\le \ \exp \bigl( - 2(a+b-1) (x - p_\ell)^2 \bigr) . \end{align*} \end{Corollary} Further tail and concentration inequalities for the Beta distribution have been derived by Marchal and Arbel (2017) and Skorski (2021). Marchal and Arbel (2017) prove that $\mathrm{Beta}(a,b)$ is subgaussian with a variance parameter that is the solution of an equation involving hypergeometric functions. An analytic upper bound for the variance parameter is $(4(a+b+1))^{-1}$, which implies the tail inequalities \begin{align*} \exp\bigl(-2(a+b+1)(x-p)^2\bigr) \ \ge \ \begin{cases} B_{a,b}(x) & \text{if} \ x \le p , \\ \bar{B}_{a,b}(x) & \text{if} \ x \ge p. \end{cases} \end{align*} These bounds are weaker than the one-sided bounds in Corollary~\ref{cor:Beta.expo}. For the right tails, the difference \[ (a+b-1)(x-p_r)^2 - (a+b+1)(x-p)^2 \] is strictly concave in $x$ with value $b^2/[(a+b)^2(a+b-1)] > 0$ for $x \in \{p,1\}$. Analogously, for the left tails, the difference \[ (a+b-1)(x-p_{\ell})^2 - (a+b+1)(x-p)^2 \] is strictly concave in $x$ with value $a^2/[(a+b)^2(a+b-1)] > 0$ for $x \in \{0,p\}$. Skorski (2021) derives a Bernstein type inequality. With the parameters \[ v^2 \ := \ \frac{p(1 - p)}{a+b+1}, \quad c \ := \ \max \biggl( \frac{|1-2p|}{a+b+2}, \sqrt{\frac{p(1-p)}{a+b+2}}\biggr), \] he shows that for $X \sim \mathrm{Beta}(a,b)$ and $\varepsilon \geq 0$, \[ P( \pm (X - p) \ge \varepsilon) \ \le \ \exp \Bigl( - \frac{\varepsilon^2}{2(v^2 + c \varepsilon)} \Bigr) . \] The next result shows that our bounds imply a stronger version of these inequalities if $a,b \ge 1$. \begin{Corollary} \label{cor:Bernstein} Let $a,b \ge 1$. Then for $x \in [p,1]$, \[ \bar{B}_{a,b}(x) \ \le \ \exp \Bigl( - \frac{(a+b+1)(x - p)^2}{2p(1 - p) + (4/3)(1 - 2p)(x - p)} \Bigr) , \] and for $x \in [0,p]$, \[ B_{a,b}(x) \ \le \ \exp \Bigl( - \frac{(a+b+1)(x - p)^2}{2p(1 - p) + (4/3)(2p - 1) (p - x)} \Bigr) . \] \end{Corollary} With the notation of Skorski~(2021), our upper bounds read \[ P(\pm (X - p) \ge \varepsilon) \ \le \ \exp \Bigl( - \frac{\varepsilon^2}{2(v^2 \pm \tilde{c} \varepsilon)} \Bigr) \] with $v^2$ as before and $\tilde{c} = (2/3) (1 - 2p)/(a+b+1)$. In particular, \[ |\tilde{c}| \ = \ \frac{2(a+b+2)}{3(a+b+1)} \frac{|1 - 2p|}{a+b+2} \ \le \ \frac{2(a+b+2)}{3(a+b+1)} c . \] Since $a,b \ge 1$, the factor $2(a+b+2)/[3(a+b+1)]$ is at most $8/9$ and converges to $2/3$ as $a+b \to \infty$. \section{Gaussian tail inequalities} \label{sec:Gaussian.tail.inequalities} Now suppose that $a,b > 1$. With $p_o := (a-1)/(a+b-2) \in (0,1)$, the density $\beta_{a,b}$ may be written as \[ \log \beta_{a,b}(x) \ = \ \log \beta_{a,b}(p_o) - (a+b-2) K(p_o,x) , \] whereas the probability density $\phi_{p_o,\sigma}$ of $\mathcal{N}(p_o,\sigma^2)$ with $\sigma := (4(a+b-2))^{-1/2}$ satisfies \[ \log \phi_{p_o,\sigma}(x) \ = \ \log \phi_{p_o,\sigma}(p_o) - 2 (a+b-2) (x - p_o)^2 . \] In particular, $\rho := \log(\beta_{a,b}/\phi_{p_o,\sigma})$ satisfies \[ \rho'(x) \ = \ (a+b-2) (x - p_o) \bigl( 4 - 1/[x(1-x)] \bigr) , \] and since $x(1-x) \le 1/4$, $\rho(x)$ is monotone decreasing in $x \ge p_o$ and monotone increasing in $x \le p_o$, where $\beta_{a,b} := 0$ on $\mathbb{R} \setminus (0,1)$. Consequently, for $x \ge p_o$, \begin{align*} \bar{B}_{a,b}(x) \ &\le \ \frac{\bar{B}_{a,b}(x)}{\bar{B}_{a,b}(p_o)} \\ &= \ \frac{\int_x^\infty e^{\rho(t)} \phi_{p_o,\sigma}(t) \, \d t} {\int_{p_o}^x e^{\rho(t)} \phi_{p_o,\sigma}(t) \, \d t + \int_x^\infty e^{\rho(t)} \phi_{p_o,\sigma}(t) \, \d t} \\ &\le \ \frac{e^{\rho(x)} \int_x^\infty \phi_{p_o,\sigma}(t) \, \d t} {e^{\rho(x)} \int_{p_o}^x \phi_{p_o,\sigma}(t) \, \d t + e^{\rho(x)} \int_x^\infty \phi_{p_o, \sigma}(t) \, \d t} \\ &= \ \frac{\mathcal{N}(p_o, \sigma^2)([x,\infty))} {\mathcal{N}(p_o,\sigma^2)([p_o,\infty))} \\ &= \ 2 \Phi \bigl( - 2 \sqrt{a+b-2} (x - p_o) \bigr) . \end{align*} Analogous arguments apply to $B_{a,b}(x)$ for $x \le p_o$, and we obtain the following bounds. \begin{Lemma} \label{lem:Beta.Phi} For $a, b > 1$ and $p_o = (a-1)/(a+b-2)$, \begin{align*} \bar{B}_{a,b}(x) \ &\le \ 2 \Phi \bigl( - 2 \sqrt{a+b-2} (x - p_o) \bigr) \quad \text{for} \ \, x \ge p_o , \\ B_{a,b}(x) \ &\le \ 2 \Phi \bigl( 2 \sqrt{a+b-2} (x - p_o) \bigr) \quad \text{for} \ \, x \le p_o . \end{align*} \end{Lemma} \section{Gaussian approximation of $\mathrm{Beta}(a,a)$} \label{sec:Gaussian.approximation.Beta(a,a)} Inspired by D{\"u}mbgen et al.~(2021), we want to compare the densities $\beta_{a,a}$ with the density $\phi_{1/2,\sigma}$ of $\mathcal{N}(1/2, \sigma^2)$ for various choices of $\sigma > 0$, where $a > 1$. Precisely, we want to determine \[ R(\sigma) \ := \ \max_{x \in (0,1)} \frac{\beta_{a,a}}{\phi_{1/2,\sigma}}(x) , \] because for arbitrary Borel sets $S \subset \mathbb{R}$, \[ \mathrm{Beta}(a,a)(S) \ \le \ R(\sigma) \, \mathcal{N}(1/2,\sigma^2)(S) \] and \[ \bigl| \mathrm{Beta}(a,a)(S) - \mathcal{N}(1/2,\sigma^2)(S) \bigr| \ \le \ 1 - R(\sigma)^{-1} . \] Moreover, we want to find $\sigma > 0$ such that this quantity is minimal. To determine $R(\sigma)$, note first that for fixed $a$ and $\sigma$, \begin{align*} \log \frac{\beta_{a,a}}{\phi_{1/2,\sigma}}(x) \ &= \ \log \sqrt{2\pi\sigma^2} - \log B(a,a) + \frac{(x - 1/2)^2}{2\sigma^2} + (a-1) \log(x(1-x)) \\ &= \ \log \sqrt{2\pi\sigma^2} - \log B(a,a) + \frac{(x - 1/2)^2}{2\sigma^2} + (a-1) \log \bigl( 1/4 - (x - 1/2)^2 \bigr) \\ &= \ \mathrm{const}(a,\sigma) + \frac{y}{8\sigma^2} + (a-1) \log(1 - y) , \end{align*} where $y := (2x - 1)^2 \in [0,1)$. Since \[ \frac{d}{dy} \Bigl( \frac{y}{8\sigma^2} + (a-1) \log(1 - y) \Bigr) \ = \ \frac{1}{8\sigma^2} - \frac{a-1}{1-y} , \] the maximum of $\log(\beta_{a,a}/\phi_{1/2,\sigma})$ is attained at $x \in (0,1)$ such that $y = \bigl( 1 - 8\sigma^2(a-1) \bigr)^+$, and the resulting value of $\log R(\sigma)$ is \begin{align*} \log R(\sigma) \ = \ &\log \sqrt{2\pi} - \log B(a,a) + (a-1) \log(1/4) \\ &+ \ \log(\sigma^2)/2 + \bigl( (8\sigma^2)^{-1} - a + 1 \bigr)^+ + (a-1) \log \min \{ 8 \sigma^2 (a-1), 1 \} \\ = \ &\log \sqrt{2\pi} - \log B(a,a) - (2a-1/2) \log(2) \\ &+ \ \log(8\sigma^2)/2 + \bigl( (8\sigma^2)^{-1} - a + 1 \bigr)^+ + (a-1) \log \min \{ 8 \sigma^2 (a-1), 1 \} . \end{align*} This is strictly monotone increasing in $8\sigma^2 \ge (a-1)^{-1}$, so we restrict our attention to values $\sigma$ in $\bigl( 0,(8(a-1))^{-1/2} \bigr]$. Then, \begin{align} \label{eq:R.sigma.1} \log R(\sigma) \ = \ &\log \sqrt{2\pi} - \log B(a,a) - (2a-1/2) \log(2) \\ \nonumber &+ \ (8\sigma^2)^{-1} + (a - 1/2) \log(8\sigma^2) - a + 1 + (a-1) \log(a-1) . \end{align} Note also the Stirling type approximation \[ \log \Gamma(y) \ = \ \log \sqrt{2\pi} + (y - 1/2) \log(y) - y + r(y) , \] where $r(y) = 1/(12y) + O(y^{-2})$ as $y \to \infty$ (cf.\ D{\"u}mbgen et al.\ 2021, Lemma~10). Consequently, \begin{align*} \log\sqrt{2\pi} - \log B(a,a) \ &= \ \log\sqrt{2\pi} + \log \Gamma(2a) - 2 \log \Gamma(a) \\ &= \ (2a - 1/2) \log(2a) - 2(a - 1/2) \log(a) + r(2a) - 2 r(a) \\ &= \ (2a - 1/2) \log(2) + \log(a)/2 + \tilde{r}(a) , \end{align*} with $\tilde{r}(a) := \tilde{r}(2a) - 2 \tilde{r}(a)$. This leads to \begin{align} \label{eq:R.sigma.2} \log R(\sigma) \ = \ &\tilde{r}(a) + \log(a)/2 \\ \nonumber &+ \ (8\sigma^2)^{-1} + (a - 1/2) \log(8\sigma^2) - a + 1 + (a-1) \log(a-1) . \end{align} For the particular choice of $\sigma$, there are at least three possibilities:\\[0.5ex] \textbf{Moment matching.} \ A first candidate for $\sigma$ would be the standard deviation of $\mathrm{Beta}(a,a)$, \[ \sigma_1(a) \ := \ (8(a + 1/2))^{-1/2} . \] \textbf{Local density matching.} \ Since $\log \beta_{a,a}(x) - \log \beta_{a,a}(1/2)$ equals $- 4(a-1) (x - 1/2)^2$ plus a remainder of order $O((x - 1/2)^4)$ as $x \to 1/2$, another natural choice would be \[ \sigma_2(a) \ := \ (8(a-1))^{-1/2} . \] \textbf{Minimizing $R(\sigma)$.} \ Note that $\log R(\sigma) = \mathrm{const}(a) + (a - 1/2) \log(8\sigma^2) + (8\sigma^2)^{-1}$. Since \[ \frac{d}{dy} \bigl( (a - 1/2) \log(y) + y^{-1} \bigr) \ = \ \frac{a - 1/2}{y} - \frac{1}{y^2} \ = \ \frac{(a - 1/2)(y - (a - 1/2)^{-1})}{y^2} , \] the optimal value of $\sigma$ equals \[ \sigma_3(a) \ := \ (8 (a - 1/2))^{-1/2} . \] \paragraph{Numerical example.} Figure~\ref{fig:Beta.Gauss.A} shows for $a = 5$ the beta density $\beta_{a,a}$ and the Gaussian approximations $\phi_{1/2,\sigma}$, where $\sigma = \sigma_1(a), \sigma_2(a), \sigma_3(a)$. Figure~\ref{fig:Beta.Gauss.B} depicts the corresponding log-density ratios $\log(\beta_{a,a}/\phi_{1/2,\sigma})$. The values of $R(\sigma)$, rounded to four digits, are $R(\sigma_1(a)) = 1.1660$, $R(\sigma_2(a)) = 1.0905$ and $R(\sigma_3(a)) = 1.0582$. \begin{figure} \centering \includegraphics[width=0.85\textwidth]{Beta_Gauss_A} \caption{The density $\beta_{a,a}$ (black) for $a = 5$ and its Gaussian approximation $\phi_{1/2,\sigma}$ for $\sigma = (8(a+1/2))^{-1}$ (green), $\sigma = (8(a-1))^{-1/2}$ (red) and $\sigma = (8(a - 1/2))^{-1/2}$ (blue).} \label{fig:Beta.Gauss.A} \end{figure} \begin{figure} \centering \includegraphics[width=0.85\textwidth]{Beta_Gauss_B} \caption{The log-density ratios $\log(\beta_{a,a}/\phi_{1/2,\sigma})$ for $a = 5$, where $\sigma$ equals $\sigma_1(a)$ (green), $\sigma_2(a)$ (red) or $\sigma_3(a)$ (blue).} \label{fig:Beta.Gauss.B} \end{figure} Our specific values $\sigma_j(a)$ are of the type $\sigma = (8(a + \delta))^{-1/2}$ for some $\delta \ge -1$. The next lemma provides two important properties of the resulting value $\log R(\sigma)$. \begin{Lemma} \label{lem:R(sigma.delta)} Let $\sigma(a) := (8(a+\delta))^{-1/2}$ for $a > 1$ with a fixed number $\delta \ge -1$. Then $\log R(\sigma(a))$ is strictly decreasing in $a > 1$, and \[ \log R(\sigma) \ = \ \frac{\delta(\delta+1) + 3/4}{2a} + O(a^{-2}) . \] \end{Lemma} For our specific standard deviations $\sigma_j(a)$ we obtain the limits \[ \lim_{a\to \infty} \, a \log R(\sigma_j(a)) \ = \ \begin{cases} 3/4 & \text{if} \ j = 1 , \\ 3/8 & \text{if} \ j = 2 , \\ 1/4 & \text{if} \ j = 3 . \end{cases} \] \begin{Remark} Similarly as in Section~\ref{sec:Gaussian.tail.inequalities}, we may conclude that for arbitrary $x > 1/2$ and $\sigma = (8(a + \delta))^{-1/2}$, \[ \bar{B}_{a,a}(x) \ \le \ R(\sigma) \Phi \Bigl( - \frac{x - 1/2}{\sigma} \Bigr) \ \le \ \frac{R(\sigma)}{2} \exp \bigl( - 4 (a + \delta) (x - 1/2)^2 \bigr) . \] Even the latter bound is stronger than the bound $\exp \bigl( - 4 (a + 1/2)(x - 1/2)^2 \bigr)$ by Marchal and Arbel (2017), as soon as $\delta \ge 0.5$ and $R(\sigma) \le 2$. For $\delta = 0.5$, this is the case for $a \ge 1.4$, and for $\delta = 1$, we only need $a \ge 1.9$. \end{Remark} \section{Proofs} \label{sec:Proofs} \begin{proof}[\bf Proof of Theorem~\ref{thm:Segura}] Note that \[ Q_{a,b}(x) \ = \ \frac{a}{x^a} \int_0^x u^{a-1} (1 - u)^{b-1} \, \d u \ = \ a \int_0^1 w^{a-1} (1 - xw)^{b-1} \, \d w . \] Since $\d^2 (1 - xw)^{b-1} / \d w^2 = (b-1)(b-2) x^2 (1 - xw)^{b-3}$, the function $w \mapsto (1 - xw)^{b-1}$ is convex if $b \not\in (1,2)$ and concave if $b \in [1,2]$. Since $a \int_0^1 w^{a-1} \, \d w = 1$, it follows from Jensen's inequality that \[ \Bigl( a \int_0^1 w^{a-1} (1 - xw) \, \d w \Bigr)^{b-1} \ = \ q_{a,b}^{(1)}(x) \] is a lower bound for $Q_{a,b}$ if $b \not\in (1,2)$ and an upper bound if $b \in [1,2]$. To compare $Q_{a,b}$ with $q_{a,b}^{(2)}$, we use a well-known formula for linear interpolation of the function $w \mapsto (1 - xw)^{b-1}$ with second derivative $(b-1)(b-2) x^2 (1 - xw)^{b-3}$ on $[0,1]$, namely, \[ (1 - xw)^{b-1} \ = \ 1 - w + w(1 - x)^{b-1} - w(1-w) (b-1)(b-2) x^2 (1 - x\tilde{w})^{b-3}/2 \] for some $\tilde{w} = \tilde{w}(x,w) \in (0,1)$. Note that \[ (b-1)(b-2) (1 - x \tilde{w})^{b-3} \ \begin{cases} \ge \ (b-1)(b-2) & \text{if} \ b \le 1 , \\ \le \ (b-1)(b-2) & \text{if} \ b \in [1,2] , \\ \ge \ (b-1)(b-2) (1 - x)^{(b-3)^+} & \text{if} \ b \ge 2 . \end{cases} \] Hence, with $h_b(w) := 1 - w + w (1 - x)^{b-1} - w(1-w) (b-1)(b-2) x^2 (1 - x)^{(b-3)^+}/2$ we may conclude that \[ a \int_0^1 w^{a-1} h_b(w) \, \d w \ = \ q_{a,b}^{(2)}(x) \] is an upper bound for $Q_{a,b}(x)$ if $b \not\in (1,2)$ and a lower bound if $b \in [1,2]$. Concerning alternative bounds for $Q_{a,b}$, let $Q : [0,x_o] \to (0,\infty]$ be a continuous function for some $x_o \in (0,1]$. Viewing $Q$ as a bound of $Q_{a,b}$, $H(x) = x^a Q(x) /[a B(a,b)]$ is a bound for $B_{a,b}(x)$. If $Q$ is differentiable on $(0,x_o)$, then elementary calculus reveals that \[ H'(x) \ = \ \beta_{a,b}(x) J(x) \quad\text{with}\quad J(x) \ := \ \frac{Q(x) + Q'(x) x/a}{(1 - x)^{b-1}} . \] If we can show that $J \ge 1$ (or $J \le 1$) on $(0,x_o)$, we may conclude that $Q_{a,b} \le Q$ (or $Q_{a,b} \ge Q$) on $[0,x_o]$. For instance, let $Q(x) := (1 - x)^b (1 + cx)$ for some $c > 0$ and $x \in [0,1]$. Then one can show that $J \le 1$ on $[0,1]$, provided that $c \ge c_{a,b} = (a+b)/(a+1)$. This yields the lower bound for $Q_{a,b}$ in \eqref{ineq:Segura.left}. Now, let $Q(x) := (1 - x)^b /(1 - cx)$ for some $c > 0$ and $0 \le x \le x_o := \min\{c^{-1},1\}$. For $0 < x < x_o$, \[ J(x) \ = \ 1 + \frac{x}{a(1 - cx)} \Bigl( c(a+1) - (a+b) + (c-1) \frac{cx}{1 - cx} \Bigr) . \] If $c < 1$, then the infimum of $c(a+1) - (a+b) + (c-1) cx /(1 - cx)$ over all $x \in (0,x_o)$ equals $ca - (a+b) < 0$. If $c \ge 1$, that infimum equals $c(a+1) - (a+b) \ge 0$, provided that $c \ge (a+b)/(a+1)$. Consequently, $J \ge 1$ on $(0,x_o)$ if $c \ge \max\{c_\ell,1\}$, and this yields the upper bound in \eqref{ineq:Segura.left} as well as the upper bound $q_{a,b}^{(3)}(x)$ for $Q_{a,b}(x)$ in case of $b \ge 1$. It remains to verify the additional inequalities for $Q_{a,b}^L, Q_{a,b}^U$. Concerning the lower bound for $Q_{a,b}^L$, the inequality $q_{a,b}^{(1)}(x) > (1-x)^b (1 + c_{a,b}x)$ is equivalent to \[ \bigl( 1 - x / (a+1-ax) \bigr)^{-b} \ > \ 1 + bx/(a+1) - a(a+b) x^2 /(a+1)^2 . \] Indeed, by convexity of $(1 - \cdot)^{-b}$, the left-hand side is larger than $1 + b x/(a+1-ax) > 1 + bx/(a+1)$. Now let $b \ge 1$. For $b \in [1,2]$, $q_{a,b}^{(2)}(x) \ge (a(1-x)^{b-1} + 1)/(a+1)$, and the latter term is strictly larger than $(1-x)^b (1 + c_{a,b}x)$ if and only if \[ (1 - x)^{-(b-1)} \ > \ 1 + (b-1)x - (a+b) x^2 . \] Indeed, since $b-1 \ge 0$, the left-hand side is not smaller than $1 + (b-1)x$. Concerning the upper bound for $Q_{a,b}^U$, if $b \le 1$, then $c_{a,b} \le 1$ and $(1 - x)^{b-1} \ge 1$, so $q_{a,b}^{(2)}(x) \le (a(1 - x)^{b-1} + 1)/(a+1) \le (1 - x)^{b-1} = (1 - x)^b / (1 - \max\{c_{a,b},1\}x)^+$. If $1 < b \le 2$, then $c_{a,b} > 1$, and the inequality $q_{a,b}^{(1)}(x) < (1 - x)^b/(1 - c_{a,b} x)^+$ is equivalent to \[ (1 - (b-1) y)^+ (1 + y)^{b-1} \ < \ 1 \] with $y := (a+1)^{-1} x/(1-x) \in (0,\infty)$. By concavity of $(1 + \cdot)^{b-1}$, $(1 + y)^{b-1} \le 1 + (b-1)y$, whence $(1 - (b-1)y)^+(1 + y)^{b-1} \le (1 - (b-1)^2y^2)^+ < 1$. \end{proof} \begin{proof}[\bf Proof of Corollary~\ref{cor:Segura}] Recall Stirling's approximation $\Gamma(c) = \sqrt{2\pi} c^{c-1/2} e^{-c} (1 + o(1))$ as $c \to \infty$. This implies the following asymptotic expansions as $b \to \infty$: \[ \frac{1}{B(a,b)} \ = \ \frac{\Gamma(a+b)}{\Gamma(a) \Gamma(b)} \ = \ \frac{b^a (1 + a/b)^{a+b-1/2} e^{-a} (1 + o(1))}{\Gamma(a)} \ = \ \frac{b^a (1 + o(1))}{\Gamma(a)} . \] Consequently, \[ G_a(x) \ = \ \lim_{b\to \infty} B_{a,b}(x/b) \ = \ \lim_{b\to\infty} \frac{(x/b)^a}{a B(a,b)} Q_{a,b}(x/b) \ = \ \frac{x^a}{a\Gamma(a)} \lim_{b\to \infty} Q_{a,b}(x/b) . \] Bounding $Q_{a,b}(x/b)$ in terms of $q_{a,b}^{(\ell)}(x/b)$, $1 \le \ell \le 3$, as in Theorem~\ref{thm:Segura}, the asserted bounds for $G_a(x)$ follow immediately from the following limits: \begin{align*} q_{a,b}^{(1)}(x/b) \ &= \ \Bigl( 1 - \frac{ax}{(a+1) b} \Bigr)^{b-1} \ \to \ e^{-ax/(a+1)} , \\ q_{a,b}^{(2)}(x/b) \ &= \ \frac{a (1 - x/b)^{b-1} + 1}{a+1} - \frac{a(b - 1)(b - 2) x^2 (1-x/b)^{(b-3)^+}}{2b^2(a+1)(a+2)} \\ &\to \ \frac{ae^{-x} + 1}{a+1} - \frac{a x^2 e^{-x}}{2(a+1)(a+2)} , \\ q_{a,b}^{(3)}(x/b) \ &= \ \frac{(1 - x/b)^b}{\bigl( 1 - [(a+b)/b] x/(a+1) \bigr)^+} \ \to \ \frac{e^{-x}}{(1 - x/(a+1))^+} . \end{align*} As to $\bar{G}_a(x)$, we write $\bar{G}_a(x) = \lim_{b \to \infty} \bar{B}_{a,b}(x/b) = \lim_{b \to \infty} B_{b,a}(1 - x/b)$ and \[ B_{b,a}(1 - x/b) \ = \ \frac{(1 - x/b)^b}{b B(a,b)} Q_{b,a}(1 - x/b) \ = \ \frac{e^{-x} b^{a-1} (1 + o(1))}{\Gamma(a)} Q_{b,a}(1 - x/b) , \] so $\bar{G}_a(x)$ is $e^{-x}/\Gamma(a)$ times $\lim_{b \to \infty} b^{a-1} Q_{b,a}(1 - x/b)$. Bounding $Q_{b,1}(1-x/b)$ in terms of $q_{b,1}^{(\ell)}(1 - x/b)$, $1 \le \ell \le 3$, as in Theorem~\ref{thm:Segura}, the asserted bounds for $\bar{G}_a(x)$ follow immediately from the following limits: \begin{align*} b^{a-1} q_{b,a}^{(1)}(1 - x/b) \ &= \ b^{a-1} \Bigl( \frac{x+1}{b+1} \Bigr)^{a-1} \ \to \ (x + 1)^{a-1} , \\ b^{a-1} q_{b,a}^{(2)}(1 - x/b) \ &= \ \frac{b x^{a-1} + b^{a-1}}{b+1} - \frac{b^{a}(a-1)(a-2)(1-x/b)^{2}(x/b)^{(a-3)^+}}{2(b+1)(b+2)} \\ & \to \ \begin{cases} x^{a-1} & \text{if} \ a < 2 , \\ x+1 & \text{if} \ a = 2 , \\ \infty & \text{if} \ a > 2 , a \neq 3, \\ x^2+2x+2 & \text{if} \ a = 3, \end{cases} \\ b^{a-1} q_{b,a}^{(3)}(1 - x/b) \ &= \ x^a \Big/ \Bigl( \frac{(a+b) x - b (a-1)}{b+1} \Bigr)^+ \ \to \ \frac{x^a}{(x - a + 1)^+} . \end{align*}\\[-5ex] \end{proof} \begin{proof}[\bf Proof of Theorem~\ref{thm:Beta.expo}] Since $B_{a,b}(\cdot) = \bar{B}_{b,a}(1-\cdot)$ and $K(q,x) = K(1-q,1-x)$ for $q \in (0,1)$ and $x \in [0,1]$, it suffices to prove the result for $\bar{B}_{a,b}(x)$, $x \in [p_r,1]$. In case of $a = 1$, the asserted bounds are sharp, because $B(a,b) = 1/b$, $p_r = 0$ and $\bar{B}_{a,b}(x) = (1 - x)^b$. In case of $a > 1$, the ratio \[ Q(x) := \frac{\bar{B}_{a,b}(x)}{x^{a-1} (1 - x)^b} \ = \ \frac{B(a,b)^{-1}}{1 - x} \int_x^1 \Bigl( \frac{u}{x} \Bigr)^{a-1} \Bigl( \frac{1 - u}{1 - x} \Bigr)^{b-1} \, \d u \] is strictly decreasing in $x \in (0,1)$. Indeed, with $w(u) := (1 - u)/(1 - x) \in (0,1)$ for $u \in (x,1)$, we have $dw(u)/du = -1/(1-x)$, and $u = 1 - (1 - x) w(u)$, so \[ Q(x) \ = \ B(a,b)^{-1} \int_0^1 \Bigl( w + \frac{1 - w}{x} \Bigr)^{a-1} w^{b-1} \, \d w , \] which is strictly decreasing in $x \in (0,1)$ with limit $Q(1) = 1/[b B(a,b)]$. Consequently, for $0 < x_o \le x \le 1$, \[ Q(1) \ \le \ Q(x) \ \le \ Q(x_o) . \] Multiplying these inequalities with $x^{a-1} (1 - x)^b$ and setting $x_o = p_r$ yields the asserted bounds for $\bar{B}_{a,b}(x)$. \end{proof} \begin{proof}[\bf Proof of Corollary~\ref{cor:Bernstein}] For symmetry reasons, it suffices to derive the upper bound for $\bar{B}_{a,b}(x)$. It suffices to show that for $x \in [p,1]$, \[ (a+b-1) K(p_r,x) \ \ge \ \frac{(a+b+1)(x - p)^2}{2p(1-p) + (4/3)(1 - 2p)^+ (x - p)} . \] To simplify notation, we write $m := a+b$, $y := x - p \in [0,1-p]$ and $\delta := p - p_r = (1 - p)/(m-1)$. Then it follows from the first inequality in \eqref{ineq:K} that \[ (a+b-1) K(p_r,x) \ \ge \ \frac{(m-1)(y + \delta)^2}{2 (p + (2/3)y - \delta/3)(1 - p - (2/3)y + \delta/3)} , \] and we want to show that this is greater than or equal to \[ \frac{(m+1) y^2}{2 [p(1 - p) + (2/3)(1 - 2p) y]} . \] Note first that since $(m-1) \delta = (1 - p)$, \begin{align} \nonumber (m-1) (y + \delta)^2 \ &= \ (m-1) y^2 + 2(1 - p)y + (1 - p)\delta \\ \label{eq:enumerator} &= \ (m+1) y^2 + 2(1 - p - y)y + (1 - p) \delta \\ \nonumber &> \ (m+1) y^2 . \end{align} Furthermore, \begin{align} \nonumber (&p + (2/3)y - \delta/3)(1 - p - (2/3)y + \delta/3) \\ \label{eq:denominator} &= \ p(1 - p) + (2/3)(1 - 2p) y + (2p - 1) \delta/3 - (2y - \delta)^2/9 , \end{align} and in case of $0 < p \le 1/2$, the right hand side is not larger than $p(1-p) + (2/3)(1 - 2p) y$. This proves our assertion in case of $0 < p \le 1/2$ already, and it remains to treat the case $1/2 < p < 1$. To this end, we have to show that the ratio of \eqref{eq:enumerator} and $(m+1)y^2$ is not smaller than the ratio of \eqref{eq:denominator} and $p(1 - p) + (2/3)(1 - 2p) y$. This assertion is equivalent to the inequality \[ \frac{2(1 - p - y)y + (1 - p)\delta}{(m+1)y^2} \ \ge \ \frac{(2p-1) \delta/3 - (2y - \delta)^2/9}{p(1 - p) + (2/3)(1 - 2p) y} . \] With $\lambda := (m-1)^{-1}$ and $z:= y/(1 - p) \in [0,1]$, the latter inequality is equivalent to \[ \frac{2(1 - z)z + \lambda}{(\lambda^{-1}+2)z^2} \ \ge \ \frac{(2p-1) \lambda - (1-p)(2z - \lambda)^2/3}{3p + 2(1 - 2p)z} . \] Since the left-hand side is strictly positive and $3p + 2(1 - 2p)z \ge 3p + 2(1 - 2p) = 2 - p$, it suffices to show that \begin{equation} \label{eq:goal} \frac{(1 - p)(2z - \lambda)^2}{3(2 - p)} + \frac{2(1/z - 1) + \lambda/z^2}{1+2\lambda} \lambda \ \ \ge \ \frac{(2p-1)}{2-p} \lambda . \end{equation} If $z \le 2/3$, the second summand on the left-hand side is at least \[ \frac{2(3/2 - 1) + \lambda (3/2)^2}{1+2\lambda} \lambda \ = \ \frac{1 + (9/4)\lambda}{1 + 2\lambda} \lambda \ > \ \lambda \ > \ \frac{2p-1}{2-p} \lambda . \] Thus it suffices to consider $z \ge 2/3$. It follows from $1 \le b = (1 - p)m$ that $m \ge 1/(1-p)$, whence $\lambda \le (1-p)/p$. Thus $2z - \lambda \ge 2z - (1 - p)/p > 0$ and $(1 - p) \ge p \lambda$. Consequently, it suffices to verify that \begin{equation} \label{eq:goal2} \frac{p(2z - (1-p)/p)^2}{3(2 - p)} + \frac{2(1/z - 1) + \lambda/z^2}{(1+2\lambda)} \ \ \ge \ \frac{(2p-1)}{2-p} . \end{equation} The second summand on the left-hand side equals \[ \frac{1}{2z^2} \frac{4z(1-z) + 2\lambda}{1 + 2\lambda} \ = \ \frac{1}{2z^2} \Bigl(1 - \frac{1 - 4z(1-z)}{1 + 2\lambda} \Bigr) , \] an increasing function of $\lambda > 0$ for any fixed $z \in (0,1]$. Consequently, inequality \eqref{eq:goal2} would be a consequence of \[ \frac{p(2z - (1-p)/p)^2}{3(2 - p)} + 2(1/z - 1) \ \ \ge \ \frac{(2p-1)}{2-p} . \] Since $1/z - 1 = (1-z)/z \ge 1-z$, it even suffices to show that \begin{equation} \label{eq:goal3} (p/3)(2z - (1-p)/p)^2 + 2(2-p)(1 - z) \ \ \ge \ 2p-1 . \end{equation} The minimiser of the left-hand side, as a function of $z \in \mathbb{R}$, is given by \[ z_o \ = \ 2/p - 5/4 \ > \ 3/4 . \] If $p \le 8/9$, then $z_o \ge 1$, so it suffices to verify \eqref{eq:goal3} for $z = 1$. Indeed, \[ (p/3)(2 - (1-p)/p)^2) \ = \ (4/3)(2p-1) + (1-p)^2/(3p) \ > \ 2p-1 . \] If $8/9 \le p < 1$, then \eqref{eq:goal3} is equivalent to \[ 2p-1 \ \le \ (p/3)(2z_o - (1-p)/p)^2 + 2(2-p)(1 - z_o) \ \ = \ 10 - 5/p - (15/4) p . \] But this inequality is equivalent to \begin{equation} \label{eq:goal4} (23/4) p + 5/p \ \le \ 11 . \end{equation} The left-hand side is convex in $p$, so it suffices to verify it for $p = 8/9$ and $p = 1$. The left-hand side of \eqref{eq:goal4} equals $46/9 + 45/8 = 10 + 1/9 + 5/8 < 11$ if $p = 8/9$, and $10 + 3/4 < 11$ if $p = 1$. This concludes our proof of Corollary~\ref{cor:Bernstein}. \end{proof} \begin{proof}[\bf Proof of Lemma~\ref{lem:R(sigma.delta)}] At first we analyze $\tilde{r}(a)$. We use Binet's fomula $r(y) = \int_0^\infty e^{-yt} w(t) \, \d t$ with a certain function $w$ satisfying $12^{-1} e^{-t/12} < w(t) < 12^{-1}$, see D{\"u}mbgen et al.\ (2021, Lemma~10). Consequently, $2 r(a) - r(2a) = \int_0^\infty (2e^{-at} - e^{-2at}) w(t) \, \d t$, and since $2e^{-at} - e^{-2at} = e^{-at} (2 - e^{-at}) > 0$, we conclude that \[ 2 r(a) - r(2a) \ \begin{cases} \displaystyle < \ \frac{1}{12} \int_0^\infty (2e^{-at} - e^{-2at}) \, \d t \ = \ \frac{1}{8a} , \\[2ex] \displaystyle > \ \frac{1}{12} \int_0^\infty (2 e^{-(a+1/12)t} - e^{-(2a+1/12)t}) \, \d t \ = \ \frac{a+1/36}{8 (a + 1/12)(a + 1/24)} . \end{cases} \] In particular, as $a \to \infty$, \begin{equation} \label{eq:rtilde.asymptotic} \tilde{r}(a) \ = \ - \frac{1}{8a} + O(a^{-2}) . \end{equation} Moreover, \begin{equation} \label{ineq:rtilde.derivative} \frac{d}{da} \tilde{r}(a) \ = \ 2 \int_0^\infty t (e^{-at} - e^{-2at}) w(t) \, \d t \ < \ \frac{1}{6} \int_0^\infty t (e^{-at} - e^{-2at}) \, \d t \ = \ \frac{1}{8a^2} . \end{equation} Next we verify that $\log R(\sigma(a))$ is strictly decreasing in $a > 1$. It follows from representation \eqref{eq:R.sigma.2} that \begin{align} \label{eq:R.sigma.3} \log R(\sigma(a)) \ = \ &\tilde{r}(a) + \log(a)/2 \\ \nonumber &+ \ \delta - (a - 1/2) \log(a+\delta) + 1 + (a-1) \log(a-1) . \end{align} According to \eqref{ineq:rtilde.derivative}, the derivative of this with respect to $a$ is not greater than \[ \frac{1}{8a^2} + \frac{1}{2a} - \frac{a-1/2}{a+\delta} - \log(a + \delta) + \log(a-1) + 1 . \] For fixed $a > 1$, the derivative of this bound with respect to $\delta \ge 1$ equals $- (\delta + 1/2) /(a + \delta)^2$, so it is maximal for $\delta = -1/2$. This leads to \begin{align*} \frac{d}{da} \log R(\sigma(a)) \ &\le \ \frac{1}{8a^2} + \frac{1}{2a} + \log \Bigl( \frac{a-1}{a-1/2} \Bigr) \ = \ \frac{1}{8a^2} + \frac{1}{2a} + \log \Bigl( 1 - \frac{1}{2(a - 1/2)} \Bigr) \\ &< \ \frac{1}{8a^2} + \frac{1}{2a} + \log \Bigl( 1 - \frac{1}{2a} \Bigr) \ = \ - \sum_{k\ge 3} (2a)^{-k}/k \ < \ 0 . \end{align*} It remains to prove the expansion of $\log R(\sigma(a)))$ as $a \to \infty$. To this end, we rewrite \eqref{eq:R.sigma.3} as \[ \log R(\sigma(a)) \ = \ \tilde{r}(a) + \delta - (a-1/2) \log(1 + \delta/a) + (a-1) \log(1 - 1/a) . \] Since $\log(1 + y) = y + O(y^2) = y - y^2/2 + O(y^3)$ as $y \to 0$, \begin{align*} \delta - (a-1/2) \log(1 + \delta/a) \ &= \ \delta - \frac{(a-1/2)\delta}{a} + \frac{(a-1/2) \delta^2}{2a^2} + O(a^{-2}) \\ &= \ \frac{\delta(\delta+1)}{2a} + O(a^{-2}) , \\ 1 + (a-1) \log(1 - 1/a) \ &= \ 1 - \frac{a-1}{a} - \frac{a-1}{2a^2} + O(a^{-2}) \\ &= \ \frac{1}{2a} + O(a^{-2}) . \end{align*} Combining these expansions with \eqref{eq:rtilde.asymptotic} leads to the desired expansion of $\log R(\sigma(a))$. \end{proof} \paragraph{Acknowledgements.} We are grateful to Maciej Skorski and two anonymous referees for constructive comments.

https://arxiv.org/abs/1812.06347

Optimal Regular Expressions for Permutations

The permutation language $P_n$ consists of all words that are permutations of a fixed alphabet of size $n$. Using divide-and-conquer, we construct a regular expression $R_n$ that specifies $P_n$. We then give explicit bounds for the length of $R_n$, which we find to be $4^n n^{-(\lg n)/4+\Theta(1)}$, and use these bounds to show that $R_n$ has minimum size over all regular expressions specifying $P_n$.

\section{Introduction} Given a regular language $L$ defined in some way, it is a challenging problem to find good upper and lower bounds on the size of the smallest regular expression specifying $L$. (In this paper, by a regular expression, we always mean one using the operations of union, concatenation, and Kleene closure only.) Indeed, as a computational problem, it is known that determining the shortest regular expression corresponding to an NFA is PSPACE-hard~\cite{Meyer&Stockmeyer:1972}. Jiang and Ravikumar proved the analogous result for DFAs~\cite{Jiang&Ravikumar:1993}. For more recent results on inapproximability, see~\cite{Gramlich&Schnitger:2007}. For nontrivial families of languages, only a handful of results are already known. For example, Ellul et al.~\cite{Ellul&Krawetz&Shallit&Wang:2005} showed that the shortest regular expression for the language $\{ w \in \{ 0,1 \}^n \, : \, |w|_1 \text{ is even} \}$ is of length $\Omega(n^2)$. Here $|w|_1$ denotes the number of occurrences of the symbol $1$ in the word $w$. (A simple divide-and-conquer strategy provides a matching upper bound.) Chistikov et al.~\cite{Chistikov&Ivan&Lubiw&Shallit:2017} showed that the regular language $$ \{ ij \ : \ 1 \leq i < j \leq n \} $$ can be specified by a regular expression of size exactly $n (\lfloor \log_2 n \rfloor + 2) - 2^{\lfloor \log_2 n\rfloor + 1}$, and furthermore this bound is optimal. Mousavi~\cite{Mousavi:2017} developed a general program for computing lower bounds on regular expression size for the binomial languages $$B(n,k) = \{ w \in \{ 0, 1 \}^n \, : \, |w|_1 = k \}.$$ Let $n$ be a positive integer, and define $\Sigma_n = \{ 1,2,\ldots, n \}$. In this paper we study the finite language $P_n$ consisting of all permutations of $\Sigma_n$. Thus, for example, $$ P_3 = \{ 123,132,213,231,312,321 \}.$$ We are interested in regular expressions that specify $P_n$. In counting the length of regular expressions, we adopt the conventional measure of {\it alphabetic length} (see, for example, \cite{Ehrenfeucht&Zeiger:1976}): the length of a regular expression is the number of occurrences of symbols of the alphabet $\Sigma_n$. Thus, other symbols, such as parentheses and {\tt +}, are ignored. A brute-force solution, which consists of listing all the members of $P_n$ and separating them by the union symbol {\tt +}, evidently gives a regular expression for $P_n$ of alphabetic length $n \cdot n!$. This can be improved to $n!\sum_{0 \leq i < n} 1/i! \sim e\cdot n!$ by tail recursion, where $E(S)$ represents a regular expression for all permutations of the symbols of $S$: $$E (S) = \sum_{i \in S} i(E(S-\{ i \})); \quad E({i}) = i .$$ For example, for $P_4$ this gives\\ \centerline{ {\tt 1(2(34+43)+3(24+42)+4(23+32))+2(1(34+43)+3(14+41)+4(13+31))+}} \\ \centerline{ \quad \quad {\tt 3(1(24+42)+2(14+41)+4(12+21))+4(1(23+32)+2(13+31)+3(12+21))}.} Can we do better? Ellul et al.~\cite{Ellul&Krawetz&Shallit&Wang:2005} proved the following weak lower bound: every regular expression for $P_n$ has alphabetic length at least $2^{n-1}$. In this note we derive an upper bound through divide-and-conquer. We then show that the regular expression this strategy produces is, in fact, actually optimal. This improves the result from \cite{Ellul&Krawetz&Shallit&Wang:2005}. The language $P_n$ is of particular interest because its complement has short regular expressions, as shown in \cite{Ellul&Krawetz&Shallit&Wang:2005}. For other results concerning context-free grammars for $P_n$, see \cite{Ellul&Krawetz&Shallit&Wang:2005,Asveld:2006,Asveld:2008,Filmus:2011}. \section{Divide-and-conquer} \label{sec:divconq} Consider the following divide-and-conquer strategy. Let $S$ be an alphabet of cardinality $n$. We consider all subsets $T \subseteq S$ of cardinality $\lfloor n/2 \rfloor$. For each subset we recursively determine a regular expression for the permutations of $T$, a regular expression for the permutations of $S - T$, and concatenate them together. This gives \begin{equation} E(S) = \sum_{{T \subseteq S}\atop {|T| = \lfloor n/2 \rfloor}} (E(T))(E(S-T)); \quad E({i}) = i . \label{rec-defn} \end{equation} Finally, we define $R_n = E(\Sigma_n)$. Thus, for example, we get\\[.1in] \centerline{$R_4 =$ {\tt (12+21)(34+43)+(13+31)(24+42)+(23+32)(14+41)+}}\\ \centerline{{\tt (14+41)(23+32)+(24+42)(13+31)+(34+43)(12+21)}}\\ for $P_4$. The alphabetic length of the resulting regular expression $R_n$ for all permutations of $\Sigma_n$ is then $f(n)$, where $$ f(n) = \begin{cases} 1, & \text{if $n = 1$}; \\[.1in] {\dbinom{n} {\lfloor n/2 \rfloor}} \bigl( f (\lfloor n/2 \rfloor) + f(\lceil n/2 \rceil) \bigr), & \text{if $n > 1$.} \end{cases} $$ The first few values of $f(n)$ are given in the table below. \begin{table}[H] \begin{center} \begin{tabular}{|c|c|} \hline $n$ & $f(n)$\\ \hline 1 & 1 \\ 2 & 4 \\ 3 & 15 \\ 4 & 48 \\ 5 & 190 \\ 6 & 600 \\ 7 & 2205 \\ 8 & 6720 \\ 9 & 29988 \\ 10 & 95760 \\ \hline \end{tabular} \end{center} \end{table} \noindent It is sequence \seqnum{A320460} in the {\it On-Line Encyclopedia of Integer Sequences} \cite{Sloane:2018}. It seems hard to determine a simple closed-form expression for $f(n)$. Nevertheless, we can roughly estimate it as follows, at least when $n = 2^m$ is a power of $2$: \begin{align*} f(2^m) &= 2 {{2^m}\choose{2^{m-1}}} f(2^{m-1}) \\ &= 2^m {{2^m}\choose{2^{m-1}}} {{2^{m-1}}\choose{2^{m-2}}} \cdots {2 \choose 1} \\ &= 2^m {{ (2^m)!} \over { (2^{m-1})! \, (2^{m-2})! \, \cdots \, 2! \, 1!}} . \end{align*} Substituting the Stirling approximation $n! \sim \sqrt{2 \pi n} (n/e)^n$ and simplifying, we get that $f(2^m)$ is roughly equal to $$ 4^{2^m} e^{-1} \pi^{(1-m)/2} 2^{-(m^2-5m+6)/4}.$$ To make this precise, and make it work when $n$ is not a power of $2$, however, takes more work. The rest of the paper is organized as follows: in Section~\ref{sec:opt}, we prove that our regular expression is in fact optimal, assuming one result that is proven at the end of the paper. In Section~\ref{analy}, we establish some inequalities related to Stirling's formula. In Section~\ref{bounds}, we connect these inequalities to $f(n)$ and obtain the estimate mentioned in the abstract. Finally, in Section~\ref{sec:opt-choice} we use our obtained bounds on $f(n)$ to provide the missing piece in our optimality proof. \section{Optimality} \label{sec:opt} In order to show that our regular expression has minimum possible length, we use the following property of $f(n)$ that we prove in Section~\ref{sec:opt-choice}: \begin{restatable}{lemma}{optchoice} \label{lem:opt-choice} If $n\ge1$, then every integer $0<k<n$ satisfies $\binom nk(f(k)+f(n-k))\ge f(n)$. Equality occurs if and only if $k=\fl{n/2}$ or $k=\cl{n/2}$. \end{restatable} For $n\ge1$ and $1\le k\le n!$, define $\ell(n,k)$ to be the minimum alphabetic length of a regular expression specifying a subset of $P_n$, where the subset has cardinality at least $k$. \begin{lemma} \label{lem:main-opt} If $n\ge1$ and $1\le k\le n!$, then $\ell(n,k)/k \ge \ell(n,n!)/n! \ge f(n)/n!$. \end{lemma} \begin{proof} We prove this by induction over the lexicographical ordering of pairs $(n,k)$. This is easy for our base case $n=k=1$, as the best regular expression is a single character. We thus suppose $n\ge2$. Consider a regular expression for a subset of $P_n$ of cardinality at least $k\geq 1$ that has minimum alphabetic length. Clearly no such expression will involve $\epsilon$ or $\emptyset$. We now consider the possibilities for the last (outermost) operation in the regular expression. Clearly the only relevant possibilities are union and concatenation. If the last operation is a union, then it is the union of two subsets of $P_n$ of cardinalities $k_1,k_2\ge1$ where $k_1+k_2\ge k$, and $k_1,k_2<k$ by minimality. Then we get \begin{align*} \frac{\ell(n,k)}k &\ge \frac{\ell(n,k_1)+\ell(n,k_2)}{k_1+k_2} \\ &\ge \min\left\{\frac{\ell(n,k_1)}{k_1},\frac{\ell(n,k_2)}{k_2}\right\} \\ &\ge \frac{\ell(n,n!)}{n!}\ge\frac{f(n)}{n!}. \end{align*} If the last operation is a concatenation, then it is the concatenation of two regular expressions for subsets of $P_{n_1}$ and $P_{n_2}$ of cardinalities $k_1$ and $k_2$ respectively (possibly after changing alphabets) where $n_1+n_2=n$ and $k_1k_2\ge k$. By minimality, we have $n_1,n_2,k_1,k_2$ all positive, so $n_1,n_2<n$. We now obtain \begin{align*} \frac{\ell(n,k)}k &\ge \frac{\ell(n_1,k_1)+\ell(n_2,k_2)}{k_1k_2} \\ &\ge \frac{\ell(n_1,n_1!)+\ell(n_2,k_2)}{n_1!\ k_2} \\ &\ge \frac{\ell(n_1,n_1!)+\ell(n_2,n_2!)}{n_1!\ n_2!} \\ &\ge \frac{f(n_1)+f(n_2)}{n_1!\ n_2!} \\ &= \frac1{n!}\binom n{n_1}(f(n_1)+f(n-n_1)) \\ &\ge \frac1{n!}\binom n{\fl{n/2}}(f(\fl{n/2})+f(\cl{n/2})) & \text{(by Lemma~\ref{lem:opt-choice})} \\ &= \frac{f(n)}{n!}. \end{align*} In both cases, we get the desired inequalities for these choices of $n$ and $k$, completing our induction. \end{proof} \begin{theorem} Let $n\ge1$. Over all regular expressions for the permutation language $P_n$, the regular expression $R_n$ given by our divide-and-conquer strategy achieves the minimum alphabetic length. \end{theorem} \begin{proof} By our construction from Section~\ref{sec:divconq}, the regular expression $R_n$ specifies the entirety of $P_n$ and has alphabetic length $f(n)$. We thus get the upper bound $\ell(n,n!)\le f(n)$. By Lemma~\ref{lem:main-opt}, we have the matching lower bound $\ell(n,n!)\ge f(n)$. Thus, $R_n$ has minimum possible alphabetic length for a regular expression specifying $P_n$. \end{proof} \section{Analysis} In what follows we use $\ln$ to denote the natural logarithm, and $\lg$ to denote logarithms to the base $2$. \label{analy} Define $S:\mathbb{R}_{>0}\to\mathbb{R}_{>0}$ to be the usual Stirling approximation~\cite{Robbins:1955}: \begin{displaymath} S(x) = \sqrt{2\pi x}(x/e)^x . \end{displaymath} \begin{lemma} \label{lem:sa} For every $x\ge1$, we have the bounds \begin{displaymath} S\bigl(x+\frac12\bigr)^2 \le S(x)S(x+1) \le e^{1/(2x)}S\bigl(x+\frac12\bigr)^2 . \end{displaymath} \end{lemma} \begin{proof} The first two derivatives of $\ln S(x)$ are \begin{align*} \frac\mathop{}\!d{\mathop{}\!d x}\ln S(x) &= \frac1{2x}+\ln x \\[.1in] \frac{\mathop{}\!d^2}{\mathop{}\!d x^2} \ln S(x) &= -\frac1{2x^2} + \frac1x, \end{align*} and so we see that $\ln S(x)$ is convex (that is, its derivative is increasing) for all $x>1/2$. Thus, by Jensen's inequality~\cite{Jensen:1906} and exponentiating, we obtain the lower bound \begin{displaymath} S(x)S(x+1) \ge S \bigl(x+\frac12 \bigr)^2 \end{displaymath} for all $x\ge1$. Now, using the mean value theorem twice, we get that \begin{align} (\ln S(x)) + {1\over 2}\mu &\leq \ln S(x + {1 \over 2}) \label{ap1}\\ \ln S(x+1) &\leq (\ln S(x+{1 \over 2})) + {1 \over 2} M, \label{ap2} \end{align} where \begin{align*} \mu &= \inf_{z \in [x,x+{1 \over 2}]} (\ln S(z))' = {1 \over {2x}} + \ln x \\ M &= \sup_{z \in [x+{1 \over 2},x+1]} (\ln S(z))' = {1 \over {2(x+1)}} + \ln(x+1) . \end{align*} Adding the inequalities \eqref{ap1} and \eqref{ap2}, we get \begin{align*} (\ln S(x)) + (\ln S(x+1)) &\leq (2 \ln S(x+1/2)) - \mu/2 + M/2 \\ & \leq (2 \ln S(x+1/2)) + {1 \over 2}\ln({{x+1} \over x}) \\ & \leq (2 \ln S(x+1/2)) + {1 \over {2x}}. \end{align*} This gives us the inequality \begin{displaymath} S(x)S(x+1) \le e^{1/(2x)}S(x+1/2)^2 \end{displaymath} for all $x\ge1$. \end{proof} Next, for $\a\in\mathbb{R}$, define the function $g_\a:\mathbb{R}_{>0}\to\mathbb{R}_{>0}$ by \begin{displaymath} g_\a(x) = \frac{4^x}{x^{(\lg x)/4}}x^\a. \end{displaymath} Our goal is to show that $f$ can be approximated by $g_\a$ for some choice of $\a$. \begin{lemma} \label{lem:ga} Let $\a>0$. Then for every $x\ge4^\a$ we have \begin{displaymath} e^{-1/(2\sqrt x)}\frac52g_\a(x+\frac12) \le g_\a(x)+g_\a(x+1) \le e^{1/(2\sqrt x)}\frac52g_\a(x+\frac12) . \end{displaymath} \end{lemma} \begin{proof} We again compute the logarithmic derivative: \begin{displaymath} \frac\mathop{}\!d{\mathop{}\!d x}\ln g_\a(x) = \frac{\a-(\lg x)/2}x+\ln4 . \end{displaymath} For $x\ge4^\a$, this derivative is at most $\ln 4$, so by the mean value theorem, $$\ln g_\a(x+1)-\ln2\le\ln g_\a(x+\frac12)\le\ln g_\a(x)+\ln2 .$$ Exponentiating, we get \begin{equation} \frac12g_\a(x+1) \le g_\a(x+\frac12) \le 2g_\a(x). \label{gm1} \end{equation} Next, we note that the derivative of $\ln x$ exceeds that of $\sqrt x$ for $0<x<4$ and is less for $x>4$. So, since $\ln4<\sqrt4$, we have $\ln x<\sqrt x$ for all $x>0$. Hence \begin{align*} \frac\mathop{}\!d{\mathop{}\!d x}\ln g_\a(x) &= \frac\a x-\frac{\ln x}{2x\ln2}+\ln4 \\ &\ge \ln4-\frac{\sqrt x}{(2\ln2)x} \\ &\ge \ln4-1/\sqrt x \end{align*} for all $x>0$. Thus, by the mean value theorem and exponentiating again, we get \begin{equation} \frac12e^{1/(2\sqrt x)}g_\a(x+1)\ge g_\a(x+\frac12)\ge2e^{-1/(2\sqrt x)}g_\a(x). \label{gm2} \end{equation} We can now combine the inequalities \eqref{gm1} and \eqref{gm2} for $x\ge4^\a$ to get \begin{align*} e^{-1/(2\sqrt x)}\frac52g_\a(x+\frac12) &\le \frac12g_\a(x+\frac12)+2e^{-1/(2\sqrt x)}g_\a(x+\frac12) \\ &\le g_\a(x)+g_\a(x+1) \\ &\le 2e^{1/(2\sqrt x)}g_\a(x+\frac12)+\frac12g_\a(x+\frac12) \\ &\le e^{1/(2\sqrt x)}\frac52g_\a(x+\frac12), \end{align*} which gives us both desired bounds. \end{proof} We now show an identity relating $g_\a$ and $S$. \begin{lemma} \label{lem:gaS} Suppose $x>0$ and $\b>0$. If $\a=\lg\b+1/4-(\lg\pi)/2$, then \begin{displaymath} \b\frac{S(2x)}{S(x)^2}g_\a(x) = g_\a(2x). \end{displaymath} \end{lemma} \begin{proof} We have \begin{align*} \b\frac{S(2x)}{S(x)^2}g_\a(x) &= \b\frac{\sqrt{4\pi x}(2x/e)^{2x}}{(\sqrt{2\pi x}(x/e)^x)^2}\frac{4^x}{x^{\lg x/4}}x^\a \\ &= \b\frac{4^x}{\sqrt{\pi x}}\frac{4^x}{x^{\lg x/4}}x^\a \\ &= \frac{2^{\lg\b+1/4}}{\pi^{1/2}x^{1/4}x^{1/4}2^{1/4}}\frac{4^{2x}}{x^{\lg x/4}}x^\a \\ &= 2^{\lg\b+1/4-\lg\pi/2}\frac{4^{2x}}{2^{(1+\lg x)/4}x^{(1+\lg x)/4}}x^\a \\ &= \frac{4^{2x}}{(2x)^{(\lg2x)/4}}(2x)^\a \\ &= g_\a(2x). \qedhere \end{align*} \end{proof} \section{Bounds on \texorpdfstring{$f(n)$}{f(n)}} \label{bounds} In this section we obtain an estimate for $f(n)$, the size of the optimal regular expression for~$P_n$. \begin{theorem} \label{thm:fn-bounds} For all $n\ge1$ we have \begin{displaymath} 0.195\frac{4^n}{n^{(\lg n)/4}}n^{5/4-(\lg\pi)/2} \le f(n) \le {1\over4}\frac{4^n}{n^{(\lg n)/4}}n^{(\lg5)-3/4-(\lg\pi)/2} . \end{displaymath} Further, when $n$ is a power of two, we get the following upper bound, matching the general lower bound. \begin{displaymath} f(n) \le {1 \over 4}\frac{4^n}{n^{\lg n/4}}n^{5/4-(\lg\pi)/2} . \end{displaymath} \end{theorem} \begin{proof} Recall the Stirling approximation \begin{equation} e^{1/(12n+1)}S(n)\le n!\le e^{1/12n}S(n); \label{eqn:stirling} \end{equation} see \cite{Robbins:1955}. Now suppose that $f(n)\le r_ng_\a(n)$ and $f(n+1)\le r_{n+1}g_\a(n+1)$, where $n\ge\max\{1,4^\a\}$, for some non-decreasing function $r:\mathbb{N}\to\mathbb{R}_{>0}$. Then by combining Lemma~\ref{lem:sa}, Lemma~\ref{lem:ga}, and equation~\eqref{eqn:stirling}, we get \begin{align*} f(2n+1) &= \binom{2n+1}n(f(n)+f(n+1)) \\ &\le e^{\frac1{12(2n+1)}-\frac1{12n+1}-\frac1{12(n+1)+1}}\frac{S(2n+1)}{S(n)S(n+1)}(r_ng_\a(n)+r_{n+1}g_\a(n+1)) \\ &\le \frac52r_{n+1}e^{1/(2\sqrt n)}\frac{S(2n+1)}{S(n+1/2)^2}g_\a(n+1/2) \end{align*} and \begin{align*} f(2n) &= \binom{2n}n(f(n)+f(n)) \\ &\le 2r_ne^{\frac1{12(2n)}-2\frac1{12n+1}}\frac{S(2n)}{S(n)^2}g_\a(n) \\ &\le 2r_n\frac{S(2n)}{S(n)^2}g_\a(n). \end{align*} For the case where $n$ is a power of two only, we use $\b=2$ and $r_n=C$, we set $$\a=\lg\b+1/4-(\lg\pi)/2=5/4-(\lg\pi)/2,$$ so $\a>0$ and $4^\a<2$. Now Lemma~\ref{lem:gaS} gives us the identity $2\frac{S(2x)}{S(x)^2}g_\a(x)=g_\a(2x)$. Then by induction we have $f(n)\le Cg_\a(n)$ for all $n\ge1$, where $C$ is any constant that satisfies this bound for $n<4$. In particular, $C={1 \over 4}$ works, so we have $f(n)\le {1 \over 4}g_{5/4-\lg\pi/2}(n)$ for all $n\ge1$ that are powers of two. Next, for general $n$, we use $\b=5/2$ and $r_n=Ce^{-\frac{\sqrt5}{(4-\sqrt{10})\sqrt n}}$, we set $$\a=\lg\b+1/4-\lg\pi/2=\lg5-3/4-\lg\pi/2,$$ so $\a>0$ and $4^\a<4$. Now Lemma~\ref{lem:gaS} gives us the identity $\frac52\frac{S(2x)}{S(x)^2}g_\a(x)=g_\a(2x)$. For $n\ge4$, we get \begin{align*} r_{n+1}e^{1/(2\sqrt n)} &= Ce^{\frac1{2\sqrt n}-\frac{\sqrt5}{(4-\sqrt{10})\sqrt{n+1}}} \\ &\le C(e^{\frac1{\sqrt n}})^{\frac12-\frac{\sqrt5}{4-\sqrt{10}}\frac{\sqrt4}{\sqrt5}} \\ &= C(e^{\frac1{\sqrt n}})^{-\frac{\sqrt{10}}{2(4-\sqrt{10})}} \\ &= Ce^{-\frac{\sqrt5}{(4-\sqrt{10})\sqrt{2n}}} \\ &\le r_{2n+1}. \end{align*} Easily, we also get $2r_n\le\frac52r_{2n}$. Thus, by induction we have $f(n)\le r_ng_\a(n)$ for all $n\ge12$, where $C$ is chosen to make this work for $12\le n<24$. In particular, $C={1\over4}$ works again. Further, since we have $r_n<C$ for all $n\ge1$, we also have $f(n)\le{1\over4}g_{\lg5-3/4-\lg\pi/2}(n)$ for all $n\ge12$. Finally, we check manually that this last inequality holds for $1\le n<12$ too, and thus for all $n\ge1$. All that remains is the lower bound. We get similar recurrences, supposing $f(n)\ge r_ng_\a(n)$ and $f(n+1)\ge r_{n+1}g_\a(n+1)$, where $n\ge\max\{1,4^\a\}$ for some non-increasing $r:\mathbb{N}\to\mathbb{R}_{>0}$. Then by a similar argument as for the upper bounds, we have \begin{align*} f(2n+1) &= \binom{2n+1}n(f(n)+f(n+1)) \\ &\ge \frac52r_ne^{\frac1{12(2n+1)+1}-\frac1{12n}-\frac1{12(n+1)}}e^{-1/(2\sqrt n)}e^{-1/(2n)}\frac{S(2n+1)}{S(n+1/2)^2}g_\a(n+1/2) \\ &\ge \frac52r_ne^{-1/(2\sqrt n)-2/(3n)}\frac{S(2n+1)}{S(n+1/2)^2}g_\a(n+1/2) \end{align*} and \begin{align*} f(2n) &= \binom{2n}n(f(n)+f(n)) \\ &\ge 2r_ne^{\frac1{24n+1}-\frac2{12n}}\frac{S(2n)}{S(n)^2}g_\a(n) \\ &\ge 2r_ne^{-1/(6n)}\frac{S(2n)}{S(n)^2}g_\a(n). \end{align*} This time, we set $\b=2$ with $r_n=Ce^{1/3n}$ (indeed non-increasing), and $\a=5/4-\lg\pi/2$, noting $4^\a<4$. Now for $n=9$, we have $\ln\frac54\ge\frac29=\frac1{2\sqrt n}+\frac1{2n}$. Since the right-hand side of this inequality is non-increasing in $n$, the bound holds for all $n\ge9$. This implies \begin{align*} \frac52e^{-1/(2\sqrt n)-2/(3n)}r_n &\ge \frac52e^{-1/(6n)-\ln(5/4)}r_n \\ &= 2r_{2n} \\ &\ge 2r_{2n+1}. \end{align*} Further, $2r_ne^{-1/6n}=2r_{2n}$, so by induction we have $f(n)\ge r_ng_\a(n)$ for all $n\ge17$, where $C$ is chosen to satisfy this bound for $17\le n<34$. In particular, $C=0.195$ works. Since $r_n>C$ for all $n\ge17$, we also have $f(n)\ge0.195g_{5/4-\lg\pi/2}(n)$ for all $n\ge17$. Finally, we check manually that this works for all $1\le n<17$ too, and thus for all $n\ge1$. \end{proof} \section{Optimality revisited} \label{sec:opt-choice} We now give a simple lower bound on the growth of $f$. \begin{lemma} \label{lem:3fn} We have $f(n+1)\ge3f(n)$ for all $n \geq 1$. \end{lemma} \begin{proof} We prove this by induction on $n$. It is easy to verify the base case $f(2)=4\ge3=3f(1)$. Otherwise $n > 1$. Suppose the desired inequality holds for all smaller values of $n$. If $n\ge2$ is odd, then let $1\le m<n$ satisfy $2m+1=n$. Then \begin{align*} f(n+1) &= f(2m+2) \\ &= 2\binom{2m+2}{m+1}f(m+1) \\ &= 2\frac{2m+2}{m+1}\binom{2m+1}mf(m+1) \\ &= \binom{2m+1}m(f(m+1)+3f(m+1)) \\ &\ge \binom{2m+1}m(3f(m)+3f(m+1)) \\ &= 3f(2m+1) \\ &= 3f(n) . \end{align*} Otherwise, if $n\ge2$ is even, then let $1\le m<n$ satisfy $2m=n$. We note that $4m+2\ge 3m+3$, so $2\frac{2m+1}{m+1}\ge3$. We then have \begin{align*} f(n+1) &= f(2m+1) \\ &= \binom{2m+1}m(f(m)+f(m+1)) \\ &= \frac{2m+1}{m+1}\binom{2m}m(f(m)+f(m+1)) \\ &\ge \frac{2m+1}{m+1}\binom{2m}m(f(m)+3f(m)) \\ &= 2\frac{2m+1}{m+1}f(2m) \\ &\ge 3f(2m) \\ &= 3f(n) . \qedhere \end{align*} \end{proof} Armed with this inequality and the bounds given by Theorem~\ref{thm:fn-bounds} of Section~\ref{bounds}, we are ready to complete our proof of the optimality of $R_n$. We recall Lemma~\ref{lem:opt-choice}, which is what we have left to show: \optchoice* \begin{proof We easily check the cases $n<12$ by hand. Let $n\ge12$ be arbitrary. It suffices to consider the cases where $k\le\fl{n/2}$, as those where $k\ge\cl{n/2}$ are symmetric. Equality for the case $k=\fl{n/2}$ is given by the definition of $f(n)$. Suppose that $n/6\le k<\fl{n/2}$. Then $n > 9$, so $3n-3 > 2n+6$. Hence we get $$\frac{n/2-1/2}{n/2+3/2} > 2/3$$ and so \begin{equation} \frac{\fl{n/2}}{\cl{n/2}+1} > 2/3. \label{medk-1} \end{equation} Also, from $k \ge n/6$ we get $3k+3 \ge \cl{n/2}+2$, and so \begin{equation} \frac{k+1}{\cl{n/2}+2} \ge 1/3. \label{medk-2} \end{equation} Then \begin{align*} &\binom nk(f(k)+f(n-k)) \\ &\ge \binom nkf(n-k) \\ &\ge \binom nk \, 3^{\fl{n/2}-k} \, f(\cl{n/2}) & \text{(by Lemma~\ref{lem:3fn})} \\ &\ge 3^{\fl{n/2}-k} \, \binom nk \, \frac{f(\fl{n/2})+f(\cl{n/2})}2 \\[.1in] &= \frac{3^{\fl{n/2}-k}}2 \, \prod_{k\le j<\fl{n/2}}\frac{j+1}{n-j} \, \binom n{\fl{n/2}}\,\left(f(\fl{n/2})+f(\cl{n/2})\right) \\[.1in] &= \frac{3^{\fl{n/2}-k}}2 \, \frac{\prod_{k<j\le\fl{n/2}}\, j}{\prod_{\cl{n/2}<j\le n-k} \, j} \, f(n) \\[.1in] &= \frac{3^{\fl{n/2}-k}}2 \, \frac{\fl{n/2}}{\cl{n/2}+1} \, \frac{\prod_{k+1\le j\le\fl{n/2}-1}\, j}{\prod_{\cl{n/2}+2\le j\le n-k} \, j} \, f(n) \\[.1in] &> \frac{3^{\fl{n/2}-k}}2 \, (2/3) \, \prod_{1\le j\le\fl{n/2}-k-1}\frac{k+j}{\cl{n/2}+1+j} \, f(n) & \text{(by \eqref{medk-1})} \\ &\ge 3^{\fl{n/2}-k-1}\,\left(\frac{k+1}{\cl{n/2}+2}\right)^{\fl{n/2}-k-1}\,f(n) \\ &\ge 3^{\fl{n/2}-k-1}\,(1/3)^{\fl{n/2}-k-1}\,f(n) & \text{(by \eqref{medk-2})} \\ &= f(n). \end{align*} Next, suppose $1\le k<n/6$. Then $4k < n-n/3$ and so $4 < \frac{n-1}k$. Then if $k>1$, \begin{align*} \binom nk &= n\prod_{2\le j\le k}\frac{n-1-k+j}j \\ &\ge n\left(\frac{n-1}k\right)^{k-1} \\ &\ge n4^{k-1}. \end{align*} We note also that $\binom n1=n4^0$, so we have $\binom nk\ge n4^{k-1}$ for all $1\le k<n/6$. We note from our proof of Lemma~\ref{lem:ga} that the derivative of $\ln g_\a(x)$ is at most $\ln4$ for $x\ge4^\a$. In particular, for $\a=5/4-\lg\pi/2$, this derivative is at most $\ln4$ for all $x\ge2$, and so $4^kg_\a(n-k)\ge g_\a(n)$ here (as $n-k>5n/6\ge10$). Thus \begin{align*} & \binom nk(f(k)+f(n-k)) \\ &\ge \binom nkf(n-k) \\ &\ge n4^{k-1}\left(0.195g_{5/4-\lg\pi/2}(n-k)\right) & \text{(by Theorem~\ref{thm:fn-bounds})} \\ &\ge n4^{k-1}\left(0.195\frac{g_{5/4-\lg\pi/2}(n)}{4^k}\right) \\ &= 0.195n{1\over4}g_{5/4-\lg\pi/2}(n) \\ &= 0.195n{1\over4}n^{2-\lg5}g_{\lg5-3/4-\lg\pi/2}(n) \\ &\ge 0.195n^{3-\lg5}f(n) & \text{(by Theorem~\ref{thm:fn-bounds})} \\ &> f(n) & \text{(since $n\ge12>0.195^{-\frac1{3-\lg5}}$)} . & \qedhere \end{align*} \end{proof} \bibliographystyle{new}

https://arxiv.org/abs/1908.03649

An Extremal Problem for the Neighborhood Lights Out Game

Neighborhood Lights Out is a game played on graphs. Begin with a graph and a vertex labeling of the graph from the set $\{0,1,2,\dots, \ell-1\}$ for $\ell \in \mathbb{N}$. The game is played by toggling vertices: when a vertex is toggled, that vertex and each of its neighbors has its label increased by $1$ (modulo $\ell$). The game is won when every vertex has label 0. For any $n\in\mathbb{N}$ it is clear that one cannot win the game on $K_n$ unless the initial labeling assigns all vertices the same label. Given that the $K_n$ has the maximum number of edges of any simple graph on $n$ vertices it is natural to ask how many edges can be in a graph so that the Neighborhood Lights Out game is winnable regardless of the initial labeling. We find all such extremal graphs on $n$ vertices that have $\binom{n}{2} - c$ edges for $c\leq \lceil\frac{n}{2}\rceil +3$ and all those that have minimum degree $n-3$. The proofs of our results require us to introduce a new version of the Lights Out game that can be played given any square matrix.

\section{Introduction} The Lights Out game was originally created by Tiger Electronics. It has since been reimagined as a light-switching game on graphs. Several variations of the game have been developed (see, for example \cite{Craft/Miller/Pritikin:09} and \cite{paper14}), but all have some important elements in common. In each game, we begin with a graph $G$ and a labeling of $V(G)$ with labels in $\mathbb{Z}_\ell$ for some $\ell \ge 2$. The vertices can be toggled so as to change the labels of some of the vertices. Finally, there is some desired labeling (usually the labeling with all labels being 0, called the \emph{zero labeling} and denoted by $\textbf{0}$) that marks the end of the game. The most common variation of the Lights Out game is what we call the \emph{neighborhood Lights Out game}. This is a generalization of Sutner's $\sigma^+$-game (see \cite{Sutner:90}). Each time we toggle some $v \in V(G)$, the label of each vertex in the closed neighborhood of $v$, $N[v]$, is increased by 1 modulo $\ell$. The game is won when the zero labeling is achieved. This game was developed independently in \cite{paper11} and \cite{Arangala:12} and has been studied in \cite{Arangala/MacDonald/Wilson:14}, \cite{Arangala/MacDonald:14}, \cite{Hope:10}, \cite{paper13}, and \cite{Hope:19}. The original Lights Out game is the neighborhood Lights Out game on a grid graph with $\ell=2$ and has been studied in \cite{Amin/Slater:92}, \cite{Goldwasser/Klostermeyer:07}, and \cite{Sutner:90}. It is possible for a Lights Out game to be impossible to win. Much of the work on Lights Out games has centered on the conditions under which winning the game is possible. Winnability depends on the version of the game that is played, the graph on which the game is played, and on $\ell$. In our paper, we work with the neighborhood Lights Out game with labels in $\mathbb{Z}_\ell$ for arbitrary $\ell \ge 2$. For each $n \ge 2$ there exist many labelings of $K_n$ for which the neighborhood Lights Out game is impossible to win. In fact, any initial labeling in which not every vertex has the same label cannot be won. It is also true that $K_n$ has the most edges of any simple graph on $n$ vertices. It then makes sense to ask, given $n,\ell \ge 2$, what is the maximum size of a simple graph on $n$ vertices with labels in $\mathbb{Z}_\ell$ for which the neighborhood Lights Out game can be won for every possible initial labeling? We call this maximum size $\max(n,\ell)$. In addition, we seek to classify the winnable graphs of maximum size among all graphs on $n$ vertices with labels from $\mathbb{Z}_{\ell}$, which we call \emph{$(n,\ell)$-extremal graphs}. It appears that the complements of $(n,\ell)$-extremal graphs have the property that every non-pendant vertex is adjacent to a pendant vertex. As in \cite{Graf:14} we write $H\astrosun K_1$ for the graph in which, for each vertex $v$ of $H$, we add a new vertex adjacent to only $v$. We call such graphs \emph{pendant graphs}. In the case that a pendant graph is a tree or a forest, we use the terms \emph{pendant tree} or \emph{pendant forest}, respectively. Our main results are in Section~\ref{Extremal Graphs}, where we determine partial results on the classification of $(n,\ell)$-extremal graphs. In the case of $n$ odd, we show that all $(n,\ell)$-extremal graphs are complements of near perfect matchings. We also classify all $(n,\ell)$-extremal graphs when $n$ is even and $\ell$ is odd. In the remaining case we have the following conjecture. \begin{cnj}\label{conjecture} For $n,\ell$ even then \[\max(n,\ell) = \binom{n}{2} - \left(\frac{n}{2} + k\right)\] where $k$ is the smallest nonnegative integer such that $\gcd(n-2k-1,\ell)=1$. In each case the $(n,\ell)$-extremal graphs are precisely the complements of pendant graphs of order $n$ that have size $\binom{n}{2} - \left(\frac{n}{2} + k\right)$. \end{cnj} By proving that the complements of pendant graphs can be won no matter the initial labeling, we conclude that $\max(n,\ell)$ is at least the quantity given in Conjecture \ref{conjecture}. We also prove equality for $0\leq k \leq 3$ and in the family of all graphs that have minimum degree at least $n-3$. To determine winnability, we depend heavily on linear algebra methods similar to those in \cite{Anderson/Feil:98}, \cite{Arangala/MacDonald/Wilson:14}, \cite{Hope:10}, and \cite{paper11}. We discuss these methods in Section \ref{Linear Algebra}. Our techniques differ in that we introduce how to play Lights Out given any square matrix. These tools allow us to determine winnability in some dense graphs by considering winnability in a modified Lights Out game in their sparse complements, which we discuss further in Section \ref{Winnability}. Throughout the paper, we assume the vertex labels of any labeling are from $\mathbb{Z}_\ell$ for some $\ell \in \mathbb{N}$, and so any reference to $\ell$ refers to this set of labels. \section{Linear Algebra} \label{Linear Algebra} Winnability in the Lights Out game on graphs can be studied by determining a strategy for toggling the vertices. But it can also be determined using linear algebra. We proceed as in \cite{Anderson/Feil:98} and \cite{paper11}. Let $G$ be a graph with $V(G)=\{v_1,v_2, \ldots , v_n\}$, and let $N(G)=[N_{ij}]$ be the neighborhood matrix of $G$ (where $N_{ij}=1$ if and only if $v_i$ is adjacent to $v_j$ or $i=j$ and $N_{ij}=0$ otherwise). Define the vectors $\textbf{b}, \textbf{x} \in \mathbb{Z}_\ell^n$ so that $\textbf{b}[i]$ is the initial label of $v_i$, and $\textbf{x}[i]$ is the number of times $v_i$ is toggled. As explained in \cite[Lemma~3.1]{paper11}, $\textbf{x}$ represents a winning set of toggles if and only if it satisfies the matrix equation $N(G)\textbf{x}=-\textbf{b}$. In this linear algebra perspective we typically think of the initial labeling of the graph as a vector (as in $\textbf{b}$ above). When we determine winnability by playing the game we will typically think of the initial labeling as a function. As described above, the neighborhood Lights Out game can be played by knowing the neighborhood matrix and an initial labeling. However, we can also play a Lights Out game using any matrix. Let $M = [m_{ij}] \in M_n(\mathbb{Z}_\ell)$ (the set of $n\times n$ matrices with entries in $\mathbb{Z}_{\ell}$), and define the \emph{vertex set of $M$} as a set of $n$ elements $V(M) = \{ v_1,v_2,\ldots,v_n \}$. We then define the $M$-Lights Out game as follows. We label the elements of $V(M)$ with a vector $\textbf{b} \in \mathbb{Z}_\ell^n$, where each $v_i$ has label $\textbf{b}[i]$. We play the game by toggling elements of $V(M)$. Each time $v_j$ is toggled, we add $m_{ij}$ to the label of $v_i$ for all $1 \leq i \leq n$. As with the ordinary Lights Out game, we win the game when we achieve the labeling ${\bf 0}$. \begin{exa} Let $G$ be a graph. For $M=N(G)$, we get the neighborhood Lights Out game. If we let $M$ be the adjacency matrix $A(G)$, we get an analogue of the $\sigma$-game from Sutner (see \cite{Sutner:90}), where toggling a vertex $v$ increases the label of each vertex in the open neighborhood $N_G(v)$ of $v$ by 1 modulo $\ell$ and leaves the label of $v$ unchanged. We call this the \emph{adjacency Lights Out game}. \end{exa} Throughout, we shorten the names of the neighborhood Lights Out game and the adajacency Lights Out game to the $N(G)$-Lights Out game and the $A(G)$-Lights Out game, respectively. We shorten even further to the $N$-Lights Out game and the $A$-Lights Out game when the graph is clear. Though the adjacency matrix and the neighborhood matrix are both symmetric there is no requirement that $M$ be symmetric in the $M$-Lights Out game. Now we introduce some terminology related to whether a given $M$-Lights Out game can be won. \begin{Def} Let $M$ and $V=V(M)$ be as above. We call a labeling $\pi$ \emph{$M$-winnable} if the $M$-Lights Out game can be won with initial labeling $\pi$. We say that $V$ is \emph{$M$-always winnable}, or \emph{$M$-AW} for short, if all labelings of $V(M)$ are $M$-winnable. \end{Def} In the case that $V$ is the vertex set of a graph $G$, we refer to $G$ as being $M$-AW, with the understanding that $V(M)=V(G)$. In these cases, $M$ is often given by the neighborhood matrix or adjacency matrix. The following summarizes the connection between whether a given $M$-Lights Out game can be won and the linear algebraic properties of $M$. The proof follows from basic linear algebra. \begin{lem} \label{matrixLO} Let $M \in M_n(\mathbb{Z}_\ell)$ and $V(M)=\{v_1,v_2,\dots,v_n\}$. \begin{enumerate} \item \label{matrixLOsys} Let $\pi$ be a labeling of $V(M)$, and define $\textbf{b}[i] = \pi(v_i)$. Then $\pi$ is $M$-winnable with the toggles given by $\textbf{x}$ if and only if $M\textbf{x} = -\textbf{b}$. \item \label{matrixLOinv} The vertex set $V(M)$ is $M$-AW if and only if $M$ is invertible over $\mathbb{Z}_{\ell}$. \end{enumerate} \end{lem} In this paper, we focus on whether or not a graph $G$ is $N(G)$-AW, so we seek to determine whether or not a given neighborhood matrix is invertible. One straightforward way to apply linear algebra techniques is when two rows or columns of a matrix are identical. \begin{Def} Let $M \in M_n(\mathbb{Z}_\ell)$, and let $v,w \in V(M)$. We call $v$ and $w$ \emph{$M$-twins} if the rows or columns of $M$ represented by $v$ and $w$ are identical. \end{Def} In graph theory, two vertices $v$ and $w$ are twins provided that have the same open neighborhood excluding $v$ and $w$. Twin vertices that are adjacent in a graph result in identical rows and columns in the neighborhood matrix and thus are $N$-twins. Twin vertices that are not adjacent result in identical rows in the adjacency matrix and thus are $A$-twins. The following is immediate from considering the invertibility of the matrix. \begin{cor}\label{twins} Let $M \in M_n(\mathbb{Z}_\ell)$, and suppose there exist $M$-twins in $V(M)$. Then $V(M)$ is not $M$-AW. \end{cor} Note that $\mathbb{Z}_\ell$ is generally not a field, but we can still use the determinant of a matrix to determine its invertibility. In particular, a matrix is invertible if and only if its determinant is a unit \cite[Corollary~2.21]{Brown:93}. As in standard linear algebra, we can apply row operations to a matrix and leave the determinant unchanged or multiplied by a unit. In particular, the typical elementary row operations (multiplying a row by a unit in $\mathbb{Z}_{\ell}$, adding an integer multiple of one row to another, and switching two rows) have no effect on whether or not the determinant is a unit. We say that $M$ is \emph{row equivalent} to $M'$ if and only if $M$ can be turned into $M'$ by applying a sequence of elementary row operations. Since elementary row operations do not change whether or not the determinant is a unit, if $M, M' \in M_n(\mathbb{Z}_\ell)$ such that $M$ is row equivalent to $M'$ then $M$ is invertible if and only if $M'$ is invertible. Thus, if $M$ and $M'$ are row equivalent then a common vertex set $V$ is $M$-AW if and only if $V$ is $M'$-AW. Thus, we can determine whether or not a set $V$ is $M$-AW by applying some elementary row operations to $M$ to obtain $M'$, and then determining whether or not $V$ is $M'$-AW. We now apply this strategy to the neighborhood Lights Out game. Our general strategy is to use elementary row operations to transform $N(G)$ into a matrix whose Lights Out game is easy to play. Our first result using this technique will be for graphs that have a dominating vertex. Given graphs $G$ and $H$ we use $G\cup H$ to denote the disjoint union of the graphs. \begin{thm} \label{subsetjoindom} Let $G$ be a graph. Then $\overline{G \cup K_1}$ is $N$-AW if and only if $G$ is $A$-AW. \end{thm} \begin{proof} We have \[ N(\overline{G \cup K_1}) = \left[ \begin{array}{c|c} N(\overline{G}) & 1 \\ \hline 1 & 1 \\ \end{array} \right] \] where the last row and column represent $V(K_1)$. We multiply each row except the last by the unit $-1$ and then add to each of those rows the last row. This turns every $1$ of $N(\overline{G})$ into a $0$ and vice versa, resulting in the adjacency matrix of $G$. Thus, we get that $N(\overline{G \cup K_1})$ is row equivalent to \[ M = \left[ \begin{array}{c|c} A(G) & 0 \\ \hline 1 & 1 \\ \end{array} \right] .\] Thus it suffices to show that $\overline{G \cup K_1}$ is $M$-AW if and only if $G$ is $A$-AW. Note that the $M$-Lights Out game is played as the $A$-Lights Out game on $G$, each vertex toggled in $V(G)$ adds 1 to the label of the vertex $v \in V(K_1)$, and toggling $v$ increases its own label by 1 and has no other effect. First suppose that $G$ is $A$-AW, and let $\pi$ be a labeling of $\overline{G\cup K_1}$. Since $G$ is $A$-AW, we can toggle the vertices of $G$ in a way that wins the $A(G)$-Lights Out game for the labeling $\pi \mid_{V(G)}$. At this point, every vertex has label 0 except $v$. We then toggle $v$ until it has label 0. In the $M$-Lights Out game, toggling $v$ has no effect on labels of other vertices, so this wins the $M$-Lights Out game. Thus $\overline{G \cup K_1}$ is $M$-AW. Conversely, suppose that $G$ is not $A$-AW. We then give $V(G)$ a labeling that is not $A(G)$-winnable. In the $M$-Lights Out game the only vertices that affect the labels of $V(G)$ are the vertices in $V(G)$, so this is not a winnable labeling for the $M$-Lights Out game. Thus, $\overline{G \cup K_1}$ is not $M$-AW, which completes the proof. \end{proof} \section{Winnability in Dense Graphs} \label{Winnability} In proving Theorem~\ref{subsetjoindom}, we use elementary row operations to convert the neighborhood Lights Out game on a dense graph into something resembling the adjacency Lights Out game on a sparse graph. Since the extremal problem we are working on seeks dense, winnable graphs and playing the game on sparse graphs is typically easier, this technique works to our advantage. The next result allows us to make a graph denser by removing an edge from the complement graph when the complement graph is combined with $P_4$. \begin{thm} \label{joinp4} Let $G$ be a graph, $U \subseteq V(G)$ and $v$ be an end vertex of $P_4$. Let $H$ be the graph where $V(H)=V(G) \cup V(P_4)$ and $E(H) = E(G) \cup E(P_4) \cup \{ uv: u \in U\}$. Then $\overline{H}$ is $N$-AW if and only if $\overline{G \cup P_4}$ is $N$-AW. \end{thm} \begin{proof} Let $V=V(\overline{G \cup P_4})=V(\overline{H})$, and let $P_4$ in both $G \cup P_4$ and $H$ be given by $vv_2v_3v_4$. Note that $\overline{P_4}$ is the path given by $v_2v_4vv_3$. By \cite[Thm. 4.3]{paper11}, $P_4$ is $N$-AW for all $\ell$. It follows that in both $\overline{H}$ and $\overline{G \cup P_4}$, the subgraph induced by $\{ v,v_2,v_3,v_4\}$ is $N$-AW. Thus, we can toggle the vertices of $P_4$ in such a way that each vertex in $P_4$ has label zero. We first assume $\overline{H}$ is $N$-AW and show $\overline{G \cup P_4}$ is $N$-AW. To that end, we let $\pi: V \rightarrow \mathbb{Z}_\ell$ and show that $\pi$ is winnable on $\overline{G \cup P_4}$. As discussed above, we can assume that $\pi \mid_{V(P_4)} = 0$. Since $\overline{H}$ is $N$-AW, $\pi$ is winnable on $\overline{H}$. In this winning strategy, let each $w \in V(G)$ be toggled $x_w$ times, and let $v_2$ be toggled $x$ times. If we apply this strategy to $\overline{H}$ but refrain from toggling $v$, $v_3$, and $v_4$, this leaves $v_2$ and $v_4$ with label $x+\sum_{w \in V(G)} x_w$, $v$ with label $\sum_{w \in V(G)-U}x_w$, and $v_3$ with label $\sum_{w \in V(G)}x_w$. Since $v_4$ is the only remaining vertex adjacent to $v_2$, $v_4$ must be toggled $-x-\sum_{w \in V(G)}x_w$ times. This will leave both $v_2$ and $v_4$ with label zero. Since $v$ is the only remaining vertex adjacent to $v_4$, this means we do not toggle $v$ at all. Thus, $v_3$ (the only remaining untoggled vertex) must make its own label zero by being toggled $-\sum_{w \in V(G)}x_w$ times. This completes winning the game on $\overline{H}$. An important observation is that the vertices of $P_4$ are collectively toggled $-2\sum_{w \in V(G)}x_w$ times, and none of those toggles come from $v$. Since each of $v_2$, $v_3$, and $v_4$ is adjacent to every vertex in $V(\overline{G})$, this implies that toggling the vertices of $P_4$ adds $-2\sum_{w \in V(G)}x_w$ to the labeling of each vertex in $V(G)$. Looked at another way, if we only toggle the vertices in $V(G)$, this leaves each such vertex with label $2\sum_{w \in V(G)}x_w$. With the initial labeling $\pi$, we now apply the above toggling strategy to $V(G)$ in $\overline{G \cup P_4}$. By the above, each vertex in $V(G)$ has label $2\sum_{w \in V(G)}x_w$. Since $v$ and each of the $v_i$ are adjacent to all vertices in $V(G)$, it follows that toggling the vertices in $V(G)$ leaves $v$ and each $v_i$ with label $\sum_{w \in V(G)}x_w$. Each of $v_2$ and $v_3$ is now toggled $-\sum_{w \in V(G)}x_w$ times. This makes the label of $v$ and each $v_i$ zero. In addition, it adds $-2\sum_{w \in V(G)}x_i$ to the labels of $V(G)$, which gives each of them label zero as well. We proceed similarly for the converse. Assume $\overline{G \cup P_4}$ is $N$-AW, and let $\pi: V \rightarrow \mathbb{Z}_\ell$ be a labeling as above with $\pi \mid_{V(P_4)}=0$. We need to prove that $\pi$ is winnable on $\overline{H}$. As before, there is a winning toggling strategy for $\overline{G \cup P_4}$, where each $w \in V(G)$ is toggled $x_w'$ times, and $v_2$ is toggled $x'$ times. At this point, we determine the toggles for $v$ and each remaining $v_i$ as before, and it follows that the vertices are collectively toggled $-2\sum_{w \in V(G)}x_w'$ times. As before, this implies that toggling the vertices of $V(G)$ results in the label of each vertex in $V(G)$ being $2\sum_{w \in V(G)}x_w'$. Again, we apply the above toggling strategy just to the vertices of $V(G)$ in $\overline{H}$. This leaves each of $v_2$, $v_3$, and $v_4$ with label $\sum_{w \in V(G)}x_w'$ and $v$ with label $\sum_{w \in V(G)-U}x_w'$. We then win the game as follows: $v_2$ is toggled $-2\sum_{w \in U}x_w'-\sum_{w \in V(G)-U} x_w'$ times, $v_3$ is toggled $-\sum_{w \in V(G)}x_w'$ times, and $v_4$ is toggled $\sum_{w \in U}x_w'$ times. \end{proof} We can apply this result to complements of graphs that include components that are path graphs. For $k\in\mathbb{N}$ and $G$ a graph we use $kG$ to denote $k$ disjoint copies of $G$. \begin{cor} \label{pathrestrictions} Let $G$ be a graph of order $n$ that is $N$-AW. \begin{enumerate} \item \label{nopath3} No component of $\overline{G}$ can be $P_k$ such that $k$ is congruent to 3 mod 4. \item \label{onepath4} At most one component of $\overline{G}$ can be $P_k$ such that $k$ is congruent to 1 modulo 4. \item \label{extpathbound} If $\overline{G}$ is an $(n,\ell)$-extremal graph, then no component of $G$ is a path of order more than 4. \end{enumerate} \end{cor} \begin{proof} For (\ref{nopath3}), let $P$ be a component of $\overline{G}$ that is a path of order $4k+3$ with $k \in \mathbb{N} \cup \{0\}$. By Lemma~\ref{joinp4}, if we replace $P$ in $\overline{G}$ with $kP_4 \cup P_3$, the complement of the resulting graph is $N$-AW if and only if $G$ is. Thus, we can assume $P=P_3$. However, the end vertices of the $P_3$ component in $\overline{G}$ are $N(G)$-twins in $G$, so $G$ is not $N$-AW by Corollary \ref{twins}. For (\ref{onepath4}), we apply Lemma~\ref{joinp4} again. If we have more than one component of $\overline{G}$ is a path with order congruent to 1 modulo 4, we can assume that all such components are $P_1$. But the vertices of these components are all $N(G)$-twins, and so in order for $G$ to be $N$-AW, $\overline{G}$ can have at most one component be a path of order congruent to 1 modulo 4. Finally, (\ref{extpathbound}) follows from the fact that if we replace the component of $\overline{G}$ that is $P_k$ with $k > 4$ with $P_{k-4} \cup P_4$, the complement of the resulting graph will be $N$-AW with larger size, thus contradicting the assumption that $\overline{G}$ is $(n,\ell)$-extremal. \end{proof} Given a matrix $M$, let $\pi$ be a labeling of $V(M)$. For $U \subseteq V(M)$ and $r \in \mathbb{Z}_\ell$, we define the labeling $\pi_{U,r}: V(M) \rightarrow \mathbb{Z}_\ell$ as \[ \pi_{U,r}(v) =\begin{cases} \pi(v)+r & v \in U \\ \pi(v) & v \notin U \end{cases} . \] In the case $U=V(M)$, we write $\pi_{V(M),r}=\pi_r$. When we encounter these labelings in the proof of Theorem~\ref{pendantremove}, we are concerned not only if certain labelings are winnable, but also how many toggles can be used to win the game for these labelings. Recall that $\textbf{0}$ is the zero labeling, which assigns to every vertex a label of 0. \begin{Def} Let $M \in M_n(\mathbb{Z}_\ell)$, $r \in \mathbb{Z}_\ell$ and $U \subseteq V(M)$. We define the set of \emph{$U$-toggling numbers} $T_U^M(r) \subseteq \mathbb{Z}_\ell$ as follows. We say $t \in T_U^M(r)$ if the elements of $V(M)$ can be toggled to win the $M$-Lights Out game with initial labeling $\textbf{0}_{U,r}$ in such a way that the vertices in $U$ are collectively toggled $t$ times. \end{Def} Note that each number in $T_U^M(0)$ corresponds to a set of toggles that leaves the initial labeling unchanged. Such sets of toggles are called \emph{null toggles}. Null toggles function very similarly to null spaces of a linear transformation. For instance, there exist two sets of toggles with $t$ toggles and $t'$ toggles of the vertices of $U$, respectively, to have the same effect on the labels of $V(M)$ if and only if $t'=t+q$ for some $q \in T_U^M(0)$. In both the neighborhood and adjacency Lights Out games, winning a particular game is equivalent to winning the game on each individual connected component. This simplifies the computation of toggling numbers in these cases. Let $G$ be a graph with $U \subseteq V(G)$ and $M$ is $N(G)$ or $A(G)$. If $G_1,G_2, \ldots,G_c$ are the connected components of $G$, and $U_i = U \cap V(G_i)$, then $T_U^M(r) = \{ \sum_{i=1}^c t_i : t_i \in T_{U_i}^{M_i}(r) \}$, where $M_i$ is $N(G_i)$ or $A(G_i)$, respectively. Suppose we have two different sets of toggles and look at their effect individually on each vertex. For each $v \in V(M)$, suppose that the label of $v$ is increased by $r_v$ for the first set of toggles and is increased by $s_v$ for the second set of toggles. Then combining the two sets of toggles increases each $v \in V(M)$ by $r_v+s_v$. We use this observation to prove the following. \begin{lem} \label{basictoggling} Let $n\in\mathbb{N}$ and $M \in M_n(\mathbb{Z}_\ell)$, let $U \subseteq V(M)$, and let $r \in \mathbb{N}$ be minimal such that $T_U^M(r) \neq \emptyset$. Then $r \mid \ell$, and $T_U^M(s) \neq \emptyset$ if and only if $r \mid s$. \end{lem} \begin{proof} It is easy to show that $\{ s \in \mathbb{Z}_\ell: T_U^M(s) \ne \emptyset\}$ is an additive subgroup of $\mathbb{Z}_\ell$. The result follows easily. \end{proof} For graphs with a pendant vertex, it will be helpful to understand the relationship between winning the adjacency game on both the graph and a certain subgraph. \begin{lem} \label{noU} Let $G$ be a graph with a pendant vertex $p$. Let $v$ be the neighbor of $p$ in $G$, let $G'$ be the graph induced by $V(G)-\{p,v\}$, let $U=N_G(v)-\{p\}$. Finally, let $\pi$ be a labeling of $G$, and define the labeling $\pi'$ on $G'$ by \[ \pi'(w) = \begin{cases} \pi(w) - \pi(p) & w \in U \\ \pi(w) & \hbox{otherwise} \end{cases}. \] Then \begin{enumerate} \item \label{ptop'} $\pi'$ is $A(G')$-winnable with $t$ toggles from $V(G')-U$ (along with perhaps some toggles from $U$) if and only if $\pi$ is $A(G)$-winnable with $t-\pi(v)-\pi(p)$ toggles from $V(G)$. \item \label{togsame} If $s \in \mathbb{Z}_\ell$, then $T_{V(G)}^{A(G)}(s) = \{ t-2s: t \in T_{V(G')-U}^{A(G')}(s)\}$. \end{enumerate} \end{lem} \begin{proof} For (\ref{ptop'}), we first assume $\pi'$ is $A(G')$-winnable with $t$ toggles from $V(G')-U$. If we begin with the labeling $\pi$ on $G$, we begin by toggling the vertices as we would to win the adjacency game on $G'$ with labeling $\pi'$. When we do this, we subtract $\pi(w)$ from the label of each $w \in V(G')-U$ and subtract $\pi(w)-\pi(p)$ from each $w \in U$. This leaves each vertex in $V(G')-U$ with label 0 and each vertex in $U$ with label $\pi(p)$. If the vertices of $U$ get toggled $t_U$ times, it also leaves $v$ with label $\pi(v)+t_U$. Then $v$ is toggled $-\pi(p)$ times and $p$ is toggled $-\pi(v)-t_U$ times to win the game. The total number of toggles is $t+t_U-\pi(p)-\pi(v)-t_U = t-\pi(p)-\pi(v)$. If we assume $\pi$ is $A(G)$-winnable with $t-\pi(p)-\pi(v)$ toggles, note that since $v$ is the only neighbor to $p$ in $G$, $v$ must be toggled $-\pi(p)$ times to win the $A(G)$-Lights Out game with initial labeling $\pi$. This leaves each $w \in V(G')$ with label $\pi'(w)$. We then toggle the vertices of $G'$ as we do for winning the adjacency game on $G$ with initial labeling $\pi$. This will win the adjacency game on $G'$ with initial labeling $\pi'$. Note that if $t_U$ is the number of toggles among the vertices of $U$, then that leaves $v$ with label $\pi(v)+t_U$. This requires $p$ to be toggled $-\pi(v)-t_U$ times. If we let $t'$ be the number of toggles among vertices in $V(G')-U$, and if we total up the number of toggles altogether, we get $t-\pi(p)-\pi(v) = -\pi(p)+t'+t_U-\pi(v)-t_U = t'-\pi(p)-\pi(v)$. Thus, $t=t'$, which proves the result. Part (\ref{togsame}) follows from letting $\pi(w)=s$ for all $w \in V(G)$. \end{proof} \begin{thm} \label{pendantremove} Let $G$ be a graph with a pendant vertex $p$. Let $v$ be the neighbor of $p$ in $G$, and let $G'$ be the graph induced by $V(G)-\{p,v\}$. \begin{enumerate} \item \label{dompen} $G$ is $A(G)$-AW if and only if $G'$ is $A(G')$-AW. \item \label{penneighbor} Let $r \in \mathbb{N}$ be minimum such that $T_{V(G)}^{A(G)}(r) \ne \emptyset$, and let $t \in T_{V(G)}^{A(G)}(r)$. Then $\overline{G}$ is $N(\overline{G})$-AW if and only if \begin{enumerate} \item \label{allswinnable} For each labeling $\pi$ of $V(G)$, there is some $s \in \mathbb{Z}_\ell$ such that $\pi_s$ is $A(G)$-winnable. \item \label{nullshit} For each $z \in \mathbb{Z}_\ell$, there exists $q \in T_{V(G)}^{A(G)}(0)$ such that there is a solution to $(r+t)x \equiv z+q$ (mod $\ell$). \end{enumerate} \end{enumerate} \end{thm} \begin{proof} For (\ref{dompen}), we first assume $G$ is $A(G)$-AW. Let $G'$ have an arbitrary labeling. We extend this labeling to a labeling of $G$ by giving each of $p$ and $v$ a label of 0. This labeling of $G$ is winnable since $G$ is $A(G)$-AW, so we toggle the vertices of $G'$ as we would in a winning toggling of $G$. If not every label of $G'$ becomes 0, then we have to toggle $v$ to give $G$ a zero labeling. However, this leaves $p$ with a nonzero label. Since $v$ is the only neighbor of $p$, this implies that toggling $v$ makes the zero labeling on $G$ impossible. Thus, the toggles we did for $G'$ leaves all vertices in $G'$ with label 0, and so $G'$ is $A(G')$-AW. If we assume $G'$ is $A(G')$-AW and let $G$ have an arbitrary labeling, we first toggle $v$ so that $p$ has label 0. The resulting labeling restricted to $G'$ is winnable since $G'$ is $A(G')$-AW. We can then toggle the vertices of $G'$ so that all vertices of $G'$ have label 0. This leaves all vertices with label 0, except perhaps $v$ since $v$ is the only vertex not in $G'$ that is adjacent to a vertex in $G'$. We then toggle $p$ until $v$ has label 0, which wins that game. Thus, $G$ is $A(G)$-AW. For (\ref{penneighbor}), let $U=N_G(v)-\{p\}$. Then $N(\overline{G})$ looks like the following. \[ \begin{array}{c|c|c|c|c|} & V(G')-U & U & v & p \\ \hline V(G')-U & N(\overline{G'-U}) & * & 1 & 1 \\ \hline U & * & N(\overline{U}) & 0 & 1 \\ \hline v & 1 & 0 & 1 & 0 \\ \hline p & 1 & 1 & 0 & 1 \\ \hline \end{array} \] where $G'-U$ is the induced subgraph with vertex set $V(G')-U$ and the $*$ blocks are the entries that make the four top-left blocks $N(\overline{G'})$. We multiply each row except the last by the unit $-1$ and then add to each of those rows the last row to get \[ M=\begin{array}{c|c|c|c|c|} & V(G')-U & U & v & p \\ \hline V(G')-U & A(G'-U) & \overline{*} & -1 & 0 \\ \hline U & \overline{*} & A(U) & 0 & 0 \\ \hline v & 0 & 1 & -1 & 1 \\ \hline p & 1 & 1 & 0 & 1 \\ \hline \end{array} \] where the $\overline{*}$ blocks are obtained from $*$ by changing the 1 entries to 0 and the 0 entries to 1. This makes the top-left four blocks $A(G')$. So the $M$-Lights Out game is played as the $A(G')$-Lights Out game on $V(G')$; toggling any vertex in $V(G')$ adds 1 to the label of $p$; toggling any vertex in $U$ adds 1 to the label of $v$; toggling $v$ adds $-1$ to every vertex in $(V(G)-U) \cup \{v\}$; and toggling $p$ adds 1 to $v$ and $p$. We first assume $\overline{G}$ is $N(\overline{G})$-AW. Since $M$ is row equivalent to $N(\overline{G})$, $\overline{G}$ is $M$-AW. To prove (\ref{allswinnable}), consider the labeling giving $v$ and $p$ labels of 0, each $w \in U$ a label of $\pi(w)-\pi(p)$, and each $w \in V(G')-U$ a label of $\pi(w)$, which is $M$-winnable by assumption. If we toggle $v$ and $p$ as part of a winning toggling, we get the following labeling of $V(G')$. \[ \lambda(w) = \left\{ \begin{matrix} \pi(w)-\pi(p) & w \in U \\ \pi(w)+s & \hbox{otherwise} \end{matrix} \right. \] where $v$ is toggled $-s$ times. We claim that $\pi_s$ is $A(G)$-winnable. If we define $\pi_s'$ similarly as $\pi'$ in Lemma~\ref{noU}, then $\pi_s' = \lambda$, which we showed to be $A(G')$-winnable. By Lemma~\ref{noU}(\ref{ptop'}), $\pi_s$ is winnable. For (\ref{nullshit}), let $z \in \mathbb{Z}_\ell$, and consider the labeling where $p$ has label $-z$ and all other labels are 0. This labeling is $M$-winnable by assumption, so let $y_1$ be the number of times $v$ is toggled and $y_2$ be the number of times $p$ is toggled in order to win the $M$-Lights Out game with this labeling. This results in each vertex of $V(G')-U$ having label $-y_1$, each vertex of $U$ having label 0, $v$ having label $y_2-y_1$, and $p$ having label $-z+y_2$. At this point, we have only the vertices in $V(G')$ to toggle, which means the remaining toggles necessary to win the $M$-Lights Out game will also win the $A(G')$-Lights Out game with labeling $\textbf{0}_{V(G')-U,-y_1}$. Thus, $T_{V(G')-U}^{A(G')}(-y_1) \ne \emptyset$. By Lemma~\ref{noU}(\ref{togsame}), $T_{V(G)}^{A(G)}(-y_1) \ne \emptyset$, and so $-y_1=rx$ for some $x \in \mathbb{Z}$ by Lemma~\ref{basictoggling}. By assumption, $t \in T_{V(G)}^{A(G)}(r)$, and so $t=t'-2r$ for some $t' \in T_{V(G')-U}^{A(G')}(r)$ by Lemma~\ref{noU}(\ref{togsame}). Thus, there exists $t_U \in \mathbb{Z}$ such that we can collectively toggle the vertices of $U$ $t_U$ times and the vertices of $V(G')-U$ $t'$ times to win the $A(G')$-Lights Out Game with labeling $\textbf{0}_{V(G')-U,r}$. By repeating $x$ times the toggles we use for the labeling $\textbf{0}_{V(G')-U,r}$, we can toggle the vertices of $U$ and $V(G')-U$ $xt_U$ and $xt'$ times, respectively, to win the $A(G')$-Lights Out Game with labeling $\textbf{0}_{V(G')-U,rx}$. Since toggling the vertices of $G'$ to win the $M$-Lights Out game also must win the adjacency game on $G'$ with labeling $\textbf{0}_{V(G')-U,rx}$, we must toggle the vertices of $G'$ $xt_U+xt'+k$ for some $k \in T_{V(G')}^{A(G')}(0)$. If we let $k=q_1+q_2$, where $q_1$ is the number of toggles from $V(G')-U$ and $q_2$ is the number of toggles from $U$ in the null toggle, we have $q_1 \in T_{V(G')-U}^{A(G')}(0)$. Note that by negating all toggles in this null toggle, we still get a null toggle, and so $-q_1 \in T_{V(G')-U}^{A(G')}(0)$. This leaves all vertices in $V(G')$ with label 0, $v$ with label $y_2+xr+xt_U+q_2$ (by setting $-y_1=xr$), and $p$ with label $-z+y_2+xt_U+xt'+q_1+q_2$. All of the toggles have been accounted for, and so the labels of $v$ and $p$ must be 0. We then eliminate $y_2$ in the resulting system of equations to get $(r-t')x=-z+q_1$. Recall that $t=t'-2r$, and so $t'=t+2r$. This gives us $(-r-t)x=-z+q_1$, and so $(r+t)x=z-q_1$. Now let $q=-q_1$. As noted above, $q \in T_{V(G')-U}^{A(G')}(0)$. By Lemma~\ref{noU}(\ref{togsame}), $T_{V(G')-U}^{A(G')}(0)=T_{V(G)}^{A(G)}(0)$, and so $q \in T_{V(G)}^{A(G)}(0)$. Since $(r+t)x=z+q$, this proves (\ref{nullshit}). Now we assume that (\ref{allswinnable}) and (\ref{nullshit}) hold, and we prove that $\overline{G}$ is $N(\overline{G})$-AW. Since $M$ is row equivalent to $N(\overline{G})$, we need only prove that $\overline{G}$ is $M$-AW. Let $\pi$ be a labeling of $V(\overline{G})$. Consider the labeling $\lambda$ of $V(G)$ that is 0 on $p$ and $v$ and $\pi$ on $V(G')$. By (\ref{allswinnable}), $\lambda_s$ is $(A(G)$-winnable for some $s \in \mathbb{Z}_\ell$. If we define $\lambda_s'$ as in Lemma~\ref{noU}(\ref{ptop'}), we get $\lambda_s'=\pi_{V(G')-U,s} |_{V(G')}$. By Lemma~\ref{noU}(\ref{ptop'}), $\pi_{V(G')-U,s} |_{V(G')}$ is $A(G')$-winnable. Then $v$ can be toggled in the $M$-Lights Out game $-s$ times to obtain $\pi_{V(G')-U,s} |_{V(G')}$ on $V(G')$, and we can then toggle the vertices of $V(G')$ to give every vertex in $V(G')$ a label of 0. This leaves $v$ with label $a$ and $p$ with label $b$ for some $a,b \in \mathbb{Z}_\ell$. By Lemma~\ref{noU}(\ref{togsame}), $t=t'-2r$ for some $t' \in T_{V(G')-U}^{A(G')}(r)$. Thus, there exists $t_U \in \mathbb{Z}_\ell$ such that we can toggle the vertices of $V(G')-U$ $t'$ times and the vertices of $U$ $t_U$ times to win the $A(G')$-Lights Out game with labeling $\textbf{0}_{V(G')-U,r}$. Lemma~\ref{noU}(\ref{togsame}) implies that $T_{V(G)}^{A(G)}(0)=T_{V(G')-U}^{A(G')}(0)$, and so $q \in T_{V(G')-U}^{A(G')}(0)$. As we reasoned above, $-q \in T_{V(G')-U}^{A(G')}(0)$, and so there exists $q_U \in T_U^A(0)$ such that $q_U-q \in T_{V(G')}^{A(G')}(0)$. Now let $x$ be a solution to (\ref{nullshit}), where $z=a-b$. This gives us $(r-t')x=b-a-q$. If $v$ is toggled $-xr$ times and $p$ is toggled $-b-x(t_U+t')+(q-q_U)$ times, this leaves each vertex of $V(G')-U$ with label $xr$, each vertex of $U$ with label 0, $v$ with label $a-b+xr-x(t_U+t')+(q-q_U)$, and $p$ with label $-x(t_U+t')+(q-q_U)$. As we reasoned above, we can then win the $A(G')$-Lights Out game with labeling $\textbf{0}_{V(G')-U,xr}$ (and thus make the labels of $V(G')$ to be 0) by toggling the vertices of $U$ $xt_U$ times and the vertices of $V(G')-U$ a total of $xt'$ times. The vertices of $V(G')$ can be toggled in such a way that the vertices of $U$ are toggled $q_U$ times, the vertices of $V(G')-U$ are toggled $-q$ times, and these toggles collectively have no effect on the labels of $V(G')$. So we combine these to toggle the vertices of $U$ collectively $xt_U+q_U$ times and the vertices of $V(G')-U$ collectively $xt'-q$ times. This leaves the vertices of $V(G')$ with label 0, $p$ with label $(-x[t_U+t']+[q-q_U])+(xt_U+q_U+xt'-q)=0$, and $v$ with label \begin{align*} a-b+xr-x(t_U+t')+(q-q_U)+xt_U+q_U &= a-b+x(r-t')+q \\ &= a-b+(b-a-q)+q=0 \end{align*} This wins the game and shows that $\overline{G}$ is $N(\overline{G})$-AW. \end{proof} In Theorem~\ref{pendantremove}(\ref{penneighbor}), if $G$ is $A$-AW, then $\pi_s$ is automatically $A(G)$-winnable for all $s \in \mathbb{Z}_\ell$. Thus, $G$ satisfies Theorem~\ref{pendantremove}(\ref{allswinnable}) and makes $r=1$. Furthermore, $A(G)$ is invertible, so the only null toggle possible is where no buttons are pushed, making $T_{V(G)}^{A(G)}(0)=\{0\}$. This gives us the following. \begin{cor} \label{subsetjoinaw} Let $G$ be an $A$-AW graph with a pendant vertex. Let $t \in T_{V(G)}^{A(G)}(1)$. Then $\overline{G}$ is $N$-AW if and only if $\gcd(1+t,\ell)=1$. \end{cor} \begin{proof} Since $G$ is $A$-AW, part (\ref{allswinnable}) of Theorem~\ref{pendantremove} is automatically satisfied. Moreover, since $G$ is $A$-AW, $A$ is invertible, which implies that $T_{V(G)}^{A(G)}(0) = \{0\}$. The result then follows directly from Theorem~\ref{pendantremove}. \end{proof} Furthermore, for possible $(n,\ell)$-extremal graphs with a dominating vertex, Theorem~\ref{pendantremove}(\ref{dompen}) gives us a way to eliminate most graphs with pendant vertices. \begin{cor} \label{extdom} Let $G$ be a graph with a dominating vertex. If $\overline{G}$ has a pendant vertex that is not part of a component isomorphic to $P_2$, then $G$ is not $(n,\ell)$-extremal for any $n$ and $\ell$. \end{cor} \begin{proof} For contradiction, assume $G$ is $(n,\ell)$-extremal, that $\overline{G}$ has a pendant vertex $p$ with neighbor $v$, and that $p$ and $v$ are not the only vertices in their connected component of $\overline{G}$. Thus, $v$ has a neighbor other than $p$ in $\overline{G}$. Let $w$ be the dominating vertex in $G$, and let $G'$ be the subgraph of $\overline{G}$ induced by $V(\overline{G})-\{p,v,w\}$. If we remove the edges in $\overline{G}$ incident to $v$ but not $p$, we get the graph $H=G' \cup P_2 \cup P_1$. Note that $H$ has size smaller than $\overline{G}$, and so $\overline{H}$ has size greater than $G$. To contradict the assumption that $G$ is $(n,\ell)$-extremal, it then suffices to prove that $\overline{H}$ is $N$-AW. Since $G$ is $N$-AW, Theorem~\ref{subsetjoindom} implies that $\overline{G}-\{w\}$ is $A$-AW. By Theorem~\ref{pendantremove}(\ref{dompen}), this implies that $G'$ is $A$-AW. Since $P_2$ is $A$-AW for all $\ell$, it follows that $G' \cup P_2 = H-\{w\}$ is $A$-AW. By Theorem~\ref{subsetjoindom}, $\overline{H}$ is $N$-AW. This means $G$ is not $(n,\ell)$-extremal, a contradiction. \end{proof} One nice property of pendant graphs is that it is really easy to play the $A$-Lights Out game on them. This is demonstrated in the following result. Recall that $H\astrosun K_1$ denotes the graph in which, for each vertex $v$ of $H$, we add a new vertex adjacent to only $v$. \begin{lem} \label{pendwin} Let $H$ be a graph, and $G=H \astrosun K_1$. Then \begin{enumerate} \item \label{pendwinall} $G$ is $A$-AW for all $\ell \in \mathbb{N}$ \item \label{pendtoggen} If $G$ has size $m$ and order $n$, then $T_{V(G)}^{A(G)}(1)=\{ 2(m-n) \}$. \item \label{pendtreetog} If $G$ is a pendant forest with $c$ components, then $T_{V(G)}^{A(G)}(1)=\{ -2c \}$. \end{enumerate} \end{lem} \begin{proof} For (\ref{pendwinall}), we have the following algorithm for winning any $A$-Lights Out game on $G$. Toggle each vertex in $V(H)$ until its pendant neighbor has label 0. Then toggle each vertex not in $V(H)$ until its neighbor has label 0. This results in the zero labeling, which makes $G$ $A$-AW. For (\ref{pendtoggen}), we begin with the labeling $\textbf{0}_{1}$. Applying the above strategy, each vertex in $V(H)$ is toggled $-1$ times, giving a total of $-\frac{n}{2}$ toggles. Each vertex toggled also decreases by 1 the label of each adjacent vertex in $H$. Collectively, this decreases the labels of $V(H)$ by $2|E(H)| =2 \left( m- \frac{n}{2} \right)$. That means that when we toggle the pendant vertices, we must toggle $-1$ each for the initial label of 1 for each vertex in $H$ plus $\left( m- \frac{n}{2} \right)$ for the decrease in labels from toggling $V(H)$. In total, we get $-\frac{n}{2} - \frac{n}{2} + 2 \left( m- \frac{n}{2} \right) = 2(m-n)$, which completes the proof. Finally, (\ref{pendtreetog}) follows from the fact that for a forest, we have $m=n-c$. \end{proof} We can now determine the $N$-winnability of the complements of pendant graphs. Interestingly, the issue of whether or not a pendant graph is $N$-AW depends entirely on the size and order of the pendant graph. \begin{lem} \label{pendantN} Let $G$ be a pendant graph of size $m$ and order $n$. Then $\overline{G}$ is $N$-AW if and only if $\gcd(2[n-m]-1,\ell)=1$. Equivalently, if $G$ is a graph of even order $n$ and size $\binom{n}{2} - \left(\frac{n}{2} +k\right)$ such that $\overline{G}$ is a pendant graph then $G$ is $N$-AW if and only if $\gcd(n-2k-1,\ell)=1$. \end{lem} \begin{proof} By Lemma~\ref{pendwin}(\ref{pendwinall}), $G$ is $A$-AW, and so we can apply Corollary~\ref{subsetjoinaw}. By Lemma~\ref{pendwin}(\ref{pendtoggen}), $T_{V(G)}^{A(G)}(1) = \{ 2(m-n)\}$. By Corollary~\ref{subsetjoinaw}, $\overline{G}$ is $N$-AW if and only if $\gcd(2(m-n)+1,\ell)=1$. The second part follows from substituting $m=\frac{n}{2}+k$ to get $2(m-n)+1=-(n-2k-1)$. \end{proof} If $\overline{G}$ is a forest, then $n-m$ is the number of components of $\overline{G}$. This along with Lemma~\ref{pendantN} gives us the following. \begin{cor} \label{2c-1} Let $G$ be a graph such that the components of $\overline{G}$ are all pendant trees. If $c$ is the number of components of $\overline{G}$, then $G$ is $N$-AW if and only if $\gcd(2c-1,\ell)=1$. \end{cor} When we are classifying $(n,\ell)$-extremal graphs in Section \ref{Extremal Graphs}, it will be helpful to replace connected components of the complement of one graph with another graph without affecting the $N$-winnability of the original graph. The following guarantees that the conditions of Theorem~\ref{pendantremove}(\ref{penneighbor}) are unaffected by the replacement. \begin{cor} \label{extswitch} Let $G$ be a graph with a pendant vertex, and let $C$ be a connected component of $G$ that is $A$-AW. If there exists a graph $C'$ such that \begin{enumerate} \item $C'$ is $A$-AW. \item $T_{V(C')}^{A(C')}(1)=T_{V(C)}^{A(C)}(1)$ \item $C$ and $C'$ have the same order. \item $C'$ has smaller size than $C$. \end{enumerate} Then $\overline{G}$ is not $(n,\ell)$-extremal. \end{cor} \begin{proof} Let $G'$ be the graph identical to $G$ except that the component $C$ is replaced with $C'$. The winnability of the adjacency game is determined by the winnability of the adjacency game on each connected component of a given graph. Since both $C$ and $C'$ are $A$-AW, then given a labeling $\pi$ on $G$ (resp. a labeling $\pi'$ on $G'$), we have that $\pi_s$ is $A(G)$-winnable (resp. $\pi'_s$ is $A(G')$-winnable) if and only if $\pi_s$ restricted to $G-C$ (resp. $\pi'_s$ restricted to $G'-C'$) is $A(G-C)$-winnable (resp. $A(G'-C')$-winnable). Since $G-C=G'-C'$, this condition is identical for both $G$ and $G'$. Thus, either both of $G$ and $G'$ satisfy Theorem~\ref{pendantremove}(\ref{allswinnable}) or neither does. A similar argument gives us that either both of $G$ and $G'$ satisfy Theorem~\ref{pendantremove}(\ref{nullshit}) or neither does. Thus, $\overline{G}$ is $N$-AW if and only if $\overline{G'}$ is $N$-AW. Furthermore, since $C$ and $C'$ have the same order, so do $\overline{G}$ and $\overline{G'}$. Finally, since $C'$ has smaller size than $C$, $\overline{G'}$ has larger size than $\overline{G}$. Since $\overline{G'}$ and $\overline{G}$ have the same order and same winnability but $\overline{G'}$ has larger size, $\overline{G}$ cannot be $(n,\ell)$-extremal. \end{proof} \section{Extremal Graphs} \label{Extremal Graphs} Recall that $\max(n,\ell)$ is the maximum number of edges in an $N$-AW graph with $n$ vertices and a graph is $(n,\ell)$-extremal provided that it has order $n$, size $\max(n,\ell)$, and is $N$-AW. We begin with straightforward upper and lower bounds on $\max(n,\ell)$. \begin{prop} \label{bestcase} For any $n,\ell\in\mathbb{N}$ we have \[ \binom{n}{2} - (n-1) \leq \max(n,\ell) \leq \binom{n}{2} - \left\lfloor \dfrac{n}{2} \right\rfloor .\] \end{prop} \begin{proof} For the right inequality, if $|E(\overline{G})| < \left\lfloor \frac{n}{2} \right\rfloor$, at most $\left\lfloor \frac{n}{2} \right\rfloor-1$ edges are removed from $K_n$ to obtain $G$. Thus, at most $2\left(\left\lfloor \frac{n}{2} \right\rfloor -1\right)\le n-2$ vertices of $K_n$ can have their degrees reduced by one or more to obtain $G$. So, at least two vertices in $G$ are dominating vertices. Two dominating vertices are $N$-twins, and such a $G$ is not $N$-AW by Corollary \ref{twins}. On the other hand we know $\max(n,\ell)\geq \binom{n}{2} - (n-1)$ since the complement of any pendant tree is $N$-AW for all $\ell$ by Corollary \ref{2c-1}. \end{proof} To obtain the upper bound of Proposition~\ref{bestcase}, we need $G$ to be a perfect or near-perfect matching. Let $M_n$ be a perfect matching on $n$ vertices when $n$ is even and a near-perfect matching on $n$ vertices when $n$ is odd. The following two results show us when $M_n$ is $(n,\ell)$-extremal. \begin{prop}\label{prop:oddmatch} If $n$ is odd, then \[\max(n,\ell)=\binom{n}{2}-\left\lfloor\frac{n}{2}\right\rfloor\] for all $\ell\in\mathbb{N}$. Moreover, $\overline{M_n}$ is the unique $N$-AW of maximum size on $n$ vertices. \end{prop} \begin{proof} We have that $M_{n-1}$ is a pendant graph, specifically $\left(\frac{n-1}{2}K_1\right) \astrosun K_1$, so by Lemma \ref{pendwin}(\ref{pendwinall}), $M_{n-1}$ is $A$-AW for all $\ell$. By Theorem~\ref{subsetjoindom}, $\overline{M_n}$ is $N$-AW. So, $\max(n,\ell)\geq \binom{n}{2}-\left\lfloor\frac{n}{2}\right\rfloor$ when $n$ is odd. By Proposition \ref{bestcase}, $\max(n,\ell)=\binom{n}{2}-\left\lfloor\frac{n}{2}\right\rfloor$. Any graph $G\neq \overline{M_n}$ with $\binom{n}{2}-\left\lfloor\frac{n}{2}\right\rfloor$ edges must have at least two dominating vertices in $G$, which are $N$-twins. Thus, $\overline{M_n}$ is unique. \end{proof} However, when $n$ is even, not all complements of perfect matchings give us $(n,\ell)$-extremal graphs. \begin{prop} \label{evenmatch} If $n$ is even, then \begin{center} $\max(n,\ell)=\binom{n}{2}-\frac{n}{2}$ if and only if $\gcd(n-1,\ell)=1$. \end{center} If $n$ is even and $\gcd(n-1,\ell)=1$, then $\overline{M_n}$ is the unique $N$-AW graph of maximum size on $n$ vertices. \end{prop} \begin{proof} Each component of $M_n$ is a pendant tree. By Corollary \ref{2c-1} $\overline{M_n}$ is $N$-AW if and only if $\gcd \left( 2 \left( \frac{n}{2} \right)-1,\ell \right)=\gcd(n-1,\ell)=1$. That $\overline{M_n}$ is the unique $N$-AW graph of maximum size on $n$ vertices again follows from two dominating vertices being $N$-twins in any other case. \end{proof} It turns out that when $n$ is even finding $\max(n,\ell)$ is considerably more complicated, though we conjecture that almost all extremal graphs are complements of pendant graphs. In Proposition \ref{prop:triangle} we find the extremal graphs where $n$ is even and $\ell$ is odd. This is the only situation we have found in which an $(n,\ell)$-extremal graph is not the complement of a pendant graph. \begin{prop}\label{prop:triangle} Suppose that $n \ge 4$ is even. If $\ell$ is odd and $\gcd(n-1,\ell)\neq 1$ then $\max(n,\ell) = \binom{n}{2}-\left(\frac{n}{2}+1\right)$. In this case an $(n,\ell)$-extremal graph is $\overline{C_3 \cup \left( \frac{n-4}{2} \right) P_2 \cup K_1}$. \end{prop} \begin{proof} We first show that $H=\overline{C_3 \cup \left( \frac{n-4}{2} \right) P_2 \cup K_1}$, which has $\binom{n}{2}- \left( \frac{n}{2}+1 \right)$ edges, is $N$-AW. By Theorem~\ref{subsetjoindom}, we need only prove that $C_3 \cup \left( \frac{n-4}{2} \right) P_2$ is $A$-AW. Clearly, $P_2$ is $A$-AW for all $\ell \in \mathbb{N}$. We can see that $C_3$ is $A$-AW if and only if $\ell$ is odd by row reducing the adjacency matrix of $C_3$. By playing on each component, $C_3 \cup \left( \frac{n-4}{2} \right) P_2$ is $A$-AW, and so $H$ is $N$-AW. We now show that $H$ is $(n,\ell)$-extremal. Since $\gcd(n-1,\ell) \ne 1$, Proposition~\ref{evenmatch} implies that $M_n$ is not $(n,\ell)$-extremal. By the uniqueness of $M_n$, $\max(n,\ell) \leq \binom{n}{2}-\left(\frac{n}{2}+1\right)$. Since $H$ is $N$-AW and $|E(\overline{H})|=\binom{n}{2}-\left(\frac{n}{2}+1\right)$ $\max(n,\ell)=\binom{n}{2}-\frac{n}{2}+1$. \end{proof} Since $\overline{M_n}$ is the $(n,\ell)$-extremal graph in the case that $n$ is odd and $\overline{C_3 \cup \left( \frac{n-4}{2} \right) P_2 \cup K_1}$ is an $(n,\ell)$-extremal graph in the case that $n$ is even and $\ell$ is odd, from here on we consider only cases where $n$ and $\ell$ are both even. In this case we find the quantity given in Conjecture \ref{conjecture} is a lower bound. \begin{prop}\label{prop:lowerbound} If $n$ and $\ell$ are both even then \[ \max(n,\ell)\geq \binom{n}{2} - \left(\frac{n}{2} + k\right)\] where $k$ is the smallest nonnegative integer such that $\gcd(n-2k-1,\ell)=1$. \end{prop} \begin{proof} Suppose that $k$ is the smallest nonnegative integer such that $\gcd(n-2k-1,\ell)=1$. Let $G=kP_4\cup \left(\frac{n}{2} - 2k\right)P_2$. Then $G$ is a pendant graph and $\overline{G}$ has size $\binom{n}{2} - \left(\frac{n}{2} + k\right)$. So by Lemma \ref{pendantN} we have $\overline{G}$ is $N$-AW and the result follows. \end{proof} Note when $k= \frac{n}{2}-1$ we have $n-2k-1 = 1$ and so $\gcd(n-2k-1,\ell)=1$. This gives us the lower bound in Proposition \ref{bestcase}. In the following two subsections we find $\max(n,\ell)$ in two cases: finding all graphs with minimum degree $n-2$ or $n-3$ that are $(n,\ell)$-extremal for any $\ell$ in Section \ref{mindegree}, and finding all combinations of $n$ and $\ell$ such that the $(n,\ell)$-extremal graph has $\binom{n}{2}-\left(\frac{n}{2}+k\right)$ edges for $0\leq k\leq 3$ in Section \ref{smallk}. In both perspectives we are led to pendant graphs, which supports Conjecture \ref{conjecture}. \subsection{Extremal Graphs With a Given Minimum Degree }\label{mindegree} The minimum degree of an $(n,\ell)$-extremal graph can not be $n-1$ because $K_n$ is not $N$-AW. Moreover, if the minimum degree is $n-2$ then, to avoid twins, the complement graph must be $M_n$. So, Propositions \ref{prop:oddmatch} and \ref{evenmatch} tell us that if $G$ is an $(n,\ell)$-extremal graph with minimum degree $n-2$, then $G=\overline{M_n}$, which is the complement of a pendant graph when $n$ is even. Thus, in this section we find $\max(n,\ell)$ among all graphs with minimum degree $n-3$. Recall we can assume $n$ and $\ell$ are even. If $G$ has minimum degree $n-3$ the complement has maximum degree $2$. So the components of the complement graph are paths and cycles. We denote the cycle graph $C_k$ by $V(C_k) = \{ v_i : 1 \le i \le k \}$, $E(C_k) = \{ v_iv_{i+1}, v_kv_1 : 1 \le i \le k-1 \}$. Our approach to the $A$-Lights Out game on $C_k$ is similar to our approach to the $N$-Lights Out game in \cite{paper11}. We first reduce an arbitrary labeling to a canonical labeling, and then determine when these canonical labelings can be won. To that end, for $a,b \in \mathbb{Z}_\ell$ we define $\lambda_{a,b}$ to be the labeling where $v_1$ has label $a$, $v_2$ has label $b$, and the other vertices have label $0$. By a straightforward induction proof, given any initial labeling of $C_k$ in the $A$-Lights Out game, the vertices can be toggled to achieve the $\lambda_{a,b}$ labeling for some $a,b \in \mathbb{Z}_\ell$. These are our canonical labelings. The following lemma shows how we deal with the labelings $\lambda_{a,b}$ and $(\lambda_{a,b})_s$. \begin{lem} \label{labelingstuff} Let $\pi$ be a labeling of $V(C_k)$, and let $\ell$ be even. \begin{enumerate} \item \label{winlambda} The labeling $\lambda_{a,b}$ is $A$-winnable precisely in the following circumstances. \begin{itemize} \item When $n \equiv 0$ (mod 4) and $a=b=0$. \item When $n \equiv 1,3$ (mod 4) and $a$ and $b$ have the same parity. \item When $n \equiv 2$ (mod 4) and $a$ and $b$ are both even. \end{itemize} \item \label{strans} The labeling $(\lambda_{a,b})_s$ can be toggled in the $A$-Lights Out game to obtain the following labelings. \begin{itemize} \item When $n \equiv 0$ (mod 4), $\lambda_{a,b}$. \item When $n \equiv 1$ (mod 4), $\lambda_{a,b-s}$. \item When $n \equiv 2$ (mod 4), $\lambda_{a-s,b-s}$. \item When $n \equiv 3$ (mod 4), $\lambda_{a-s,b}$. \end{itemize} \end{enumerate} \end{lem} \begin{proof} For (\ref{winlambda}), let $t_i$ be the number of times we toggle $v_i$ By a straightforward induction proof, it follows that $\lambda_{a,b}$ is $A$-winnable if and only if $t_{n-1}+t_1=0$, $t_n+a+t_2=0$, and for $2 \le i \le n$ we have \begin{equation*} \label{tieq} t_i = \begin{cases} -t_2 & i \equiv 0 \hbox{ (mod 4)} \\ b+t_1 & i \equiv 1 \hbox{ (mod 4)} \\ t_2 & i \equiv 2 \hbox{ (mod 4)} \\ -b-t_1 & i \equiv 3 \hbox{ (mod 4)} \end{cases}. \end{equation*} Then (\ref{winlambda}) follows from using the equations above with $i=n-1$ and $i=n$. For (\ref{strans}), we begin with the labeling $(\lambda_{a,b})_s$, and then each $v_i$ with $2 \le i < 4\left\lfloor \frac{n}{4} \right\rfloor$ and $i \equiv 2,3$ (mod 4) is toggled $-s$ times. This results in the labeling where each $v_i$ with $1 \le i \le 4 \left\lfloor \frac{n}{4} \right\rfloor$ has label $\lambda_{a,b}(v_i)$ and each $v_i$ with $i>4\left\lfloor \frac{n}{4} \right\rfloor$ has label $(\lambda_{a,b})_s(v_i)$. This gives us the $n \equiv 0$ (mod 4) case, and the other cases follow from appropriately toggling some combination of $v_1$, $v_{n-1}$, and $v_n$. \end{proof} The next result helps us see how the presence of cycle components in a graph can affect how we apply Theorem~\ref{pendantremove}(\ref{penneighbor}). \begin{lem} \label{notswin} Let $G$ be a graph. \begin{enumerate} \item \label{notswineven} If $G$ has a connected component that is a cycle of even order, then $G$ has a labeling $\pi$ such that $\pi_s$ is not $A$-winnable for all $s \in \mathbb{Z}_\ell$. \item \label{notswinplural} If $G$ has two connected components that are cycles, then $G$ has a labeling $\pi$ such that $\pi_s$ is not $A$-winnable for all $s \in \mathbb{Z}_\ell$. \end{enumerate} \end{lem} \begin{proof} For (\ref{notswineven}), let $C$ be a cycle component of $G$ with even order, and define a labeling that is $\lambda_{1,0}$ on $C$ and arbitrary on the remaining vertices of $G$. Since a labeling is winnable on a graph if and only if it is winnable on each connected component, it suffices to prove that $(\lambda_{1,0})_s$ is not winnable on $C$ for all $s \in \mathbb{Z}_\ell$. If $C$ has order divisible by 4, then Lemma~\ref{labelingstuff}(\ref{strans}) implies that the vertices of $C$ can be toggled to achieve $\lambda_{1,0}$, which is not winnable by Lemma~\ref{labelingstuff}(\ref{winlambda}). If $C$ has order not divisible by 4, then by Lemma~\ref{labelingstuff}(\ref{strans}), the vertices can be toggled to achieve the labeling $\lambda_{1-s,-s}$. Since $1-s$ and $s$ can never both be even, Lemma~\ref{labelingstuff}(\ref{winlambda}) implies that $\lambda_{1-s,-s}$ is not winnable for all $s \in \mathbb{Z}_\ell$. In either case, $(\lambda_{1,0})_s$ is not winnable on $C$ for all $s \in \mathbb{Z}_\ell$, and so $(\lambda_{1,0})_s$ is not $A$-winnable for all $s \in \mathbb{Z}_\ell$. For (\ref{notswinplural}), let $C$ and $C'$ be two cycle components of $G$. By (\ref{notswineven}), we can assume each of $C$ and $C'$ has odd order. We claim that for any labeling $\pi$ that restricts to $\lambda_{1,0}$ on $C$ and $\lambda_{0,0}$ on $C'$, $\pi_s$ is not $A$-winnable for all $s \in \mathbb{Z}_\ell$. By Lemma~\ref{labelingstuff}(\ref{strans}), with initial labeling $\pi_s$, we can toggle the vertices of $G$ to obtain a labeling that restricts either to $\lambda_{1-s,0}$ or $\lambda_{1,-s}$ on $C$ and restricts either to $\lambda_{-s,0}$ or $\lambda_{0,-s}$ on $C'$. If $s$ is even, then $1-s$ and 0 as well as $1$ and $-s$ have opposite parity. If $s$ is odd, then $-s$ and 0 have opposite parity. In any case, Lemma~\ref{labelingstuff}(\ref{winlambda}) implies that $\pi_s$ is not winnable, and so $\pi$ is not $A$-winnable for all $s \in \mathbb{Z}_\ell$. \end{proof} Our next lemma helps us when we want to apply Theorem~\ref{subsetjoindom} to graphs with both a dominating vertex and a cycle component in its complement. \begin{lem} \label{cyclenotaw} If $\ell$ is even, then every cycle graph is not $A$-AW. \end{lem} \begin{proof} By Lemma~\ref{labelingstuff}(\ref{winlambda}), if $a,b \in \mathbb{Z}_\ell$ have opposite parity, then $\lambda_{a,b}$ is not $A$-winnable. The result follows. \end{proof} The following theorem gives us a connection between $(n,\ell)$-extremal graphs and pendant graphs, in support of Conjecture \ref{conjecture}. We use $\Delta(G)$ to denote the maximum degree of a graph $G$. \begin{thm}\label{thm:maxdegree2} Let $\ell$ be even, and let $G$ be an $(n,\ell)$-extremal graph of even order with $\Delta(\overline{G}) \le 2$. Then each connected component of $\overline{G}$ is either $P_2$ or $P_4$. \end{thm} \begin{proof} Suppose $\Delta(\overline{G})=1$. All dominating vertices in $G$ are $N$-twins so by Corollary \ref{twins} $G$ has at most $1$ dominating vertex. However, $G$ can not have only one dominating vertex since $G$ has even order. Thus, $G$ has no dominating vertices, and so $\overline{G}$ has no isolated vertices. It follows that each connected component of $G$ is $P_2$. In the case $\Delta(\overline{G})=2$, we first prove that $\overline{G}$ has at least one path component. If not, all connected components are cycles, and so $|E(\overline{G})|=|V(\overline{G})|$. However, note that any pendant tree of order $|V(G)|$ is $N$-AW for all $\ell$ by Corollary \ref{2c-1}. Since the pendant tree has size $|V(G)|-1$, this implies that $G$ is not $(n,\ell)$-extremal. Thus, $\overline{G}$ has at least one path component (possibly $P_1$). By Corollary~\ref{pathrestrictions}, no component of $\overline{G}$ is $P_k$ with $k \ge 5$ or $k=3$. Furthermore two components of $P_1$ in $\overline{G}$ would be $N$-twins in $G$, which is prohibited by Corollary \ref{twins}. If we have one component of $P_1$, Theorem~\ref{subsetjoindom} implies that all other connected components of $\overline{G}$ are $A$-AW. This excludes cycles by Lemma~\ref{cyclenotaw}. Since the remaining paths have even order, this would force $G$ to have odd order, which is a contradiction. So $G$ is $N$-AW and $\overline{G}$ has a pendant tree ($P_2$ or $P_4$) as a component. Thus, $\overline{G}$ has a pendant vertex, so we can use Theorem~\ref{pendantremove}(\ref{penneighbor}). This implies that $\overline{G}$ is $(A,\ell,s)$-winnable for some $s \in \mathbb{Z}_\ell$. However, Lemma~\ref{notswin} implies that this can not happen if either $\overline{G}$ has more than one cycle component or if $\overline{G}$ has a cycle component of even order. Moreover, if $\overline{G}$ has precisely one cycle component, and if that connected component has odd order, this implies that $G$ has odd order, which is a contradiction. Thus, $\overline{G}$ has no cycle components, and so each connected component is either $P_2$ or $P_4$, which completes the proof. \end{proof} Note that the $(n,\ell)$-extremal graphs given in Theorem \ref{thm:maxdegree2} are pendant graphs. By Lemma \ref{pendantN} $kP_4 \cup \frac{n-4k}{2}P_2$ is $N$-AW if and only if $\gcd(n-2k-1,\ell)=1$. This implies Conjecture \ref{conjecture} for the family of graphs which have minimum degree at least $n-3$. \subsection{Extremal Graphs with ${n}\choose{2}$$ - (\frac{n}{2}+k)$ edges}\label{smallk} In this section we prove Conjecture \ref{conjecture} for $0\leq k \leq 3$. We state Theorem \ref{thm:extremalpendant} in the language of that conjecture. \begin{thm}\label{thm:extremalpendant} For $n,\ell$ even and $0\leq k \leq 3$ \[\max(n,\ell)=\binom{n}{2} - \left(\frac{n}{2} + k\right)\] where $k$ is the smallest nonnegative integer such that $\gcd(n-2k-1,\ell)=1$. In each case the $(n,\ell)$-extremal graphs are precisely the complements of pendant graphs of order $n$ that have size $\binom{n}{2} - \left(\frac{n}{2} + k\right)$. \end{thm} We will prove this result using separate propositions for each $k$. When $k=0$, Proposition \ref{evenmatch} implies Theorem \ref{thm:extremalpendant}. The following lemma will help us for the cases $1\leq k\leq 3$. \begin{lem} \label{degreebound} Let $n \in \mathbb{N}$ be even, let $\ell\in\mathbb{N}$, and let $G$ be a $N$-AW graph with $|E(\overline{G})|=\frac{n}{2}+t$, where $t \ge 1$. Then $\Delta(\overline{G})\leq t+1$ where $\Delta(\overline{G})$ is the maximum degree of $\overline{G}$. \end{lem} \begin{proof} We let $v \in V(\overline{G})$ and show $\deg(v) \le t+1$, where $\deg(v)$ is the degree of $v$ in $\overline{G}$. Let $W=V(\overline{G})-N_{\overline{G}}[v]$, and note that $\deg(v)=|N_{\overline{G}}(v)|$. Then $|W|=n-\deg(v)-1$. In the graph $\overline{G}$, let $k$ be the number of edges incident only to vertices in $N_{\overline{G}}(v)$, let $r$ be the number of edges incident only to vertices in $W$, and let $s$ be the number of edges between a vertex in $N_{\overline{G}}(v)$ and a vertex in $W$. Since $|E(\overline{G})|=\frac{n}{2}+t$, we have $\frac{n}{2}+t = \deg(v)+k+r+s$, and so $k+r+s = \frac{n}{2}+t-\deg(v)$. Since $G$ is $N$-AW, it can not have any $N$-twins. Thus, no vertices in $N_{\overline{G}}(v)$ can be $N$-twins, so we can have at most one vertex in $N_{\overline{G}}(v)$ that is adjacent in $G$ to every vertex except $v$. In other words, there are at least $\deg(v)-1$ vertices in $W$ that are adjacent in $\overline{G}$ to vertices other than $v$. There can be at most two such vertices for each of the $k$ edges in $\overline{G}$ incident with two vertices in $N_{\overline{G}}(v)$, and at most one such vertex for each of the $s$ edges between vertices in $N_{\overline{G}}(v)$ and $W$. This means that there are at most $2k+s$ such vertices in $N_{\overline{G}}(v)$. It follows that $\deg(v)-1 \le 2k+s$, and so $\deg(v) \leq 2k+s+1$. In order to prevent any vertices in $W$ from becoming $N$-twins, we can have at most one vertex in $W$ that is adjacent to every vertex in $G$. In other words, there are at least $|W|-1 = n-\deg(v)-2$ vertices in $W$ with nonzero degree in $\overline{G}$. Similar reasoning as in the previous paragraph implies that there are at most $2r+s$ such vertices in $W$, and so $n-\deg(v)-2 \le 2r+s$. Thus, $\deg(v) \geq n-2r-s-2$. Since we have $n-2r-s-2 \leq \deg(v) \leq 2k+s+1$, it follows that $n-2r-s-2 \leq 2k+s+1$. This gives us $n-2k-2r-2s \leq 3$. Since the left side of the equation is even, this actually gives us $n-2k-2r-2s \leq 2$, and so $\frac{n}{2}-k-r-s \leq 1$. Rearranging this a bit gives us $k+r+s \geq \frac{n}{2}-1$. Now we use the fact $k+r+s=\frac{n}{2}+t-\deg(v)$ to get $\frac{n}{2}+t-\deg(v) \geq \frac{n}{2}-1$. Solving for $\deg(v)$ gives deg$(v) \leq t+1$. \end{proof} In the next proposition, we resolve the $k=1$ case of Theorem \ref{thm:extremalpendant}. \begin{prop}\label{prop:plus1} Suppose that $n$ and $\ell$ are even and $n\ge 4$. Then \begin{center} $\max(n,\ell)=\binom{n}{2} - \left( \frac{n}{2}+1 \right)$ if and only if $\gcd(n-1,\ell)\neq 1$ and $\gcd(n-3,\ell)=1$. \end{center} \noindent Moreover, the only $(n,\ell)$-extremal graph in this case is the complement of the unique pendant graph of order $n$ and size $\binom{n}{2}-\left(\frac{n}{2} + 1\right)$, which is $\overline{P_4\cup \left( \frac{n}{2}-2 \right) P_2}$. \end{prop} \begin{proof} Suppose $\gcd(n-1,\ell)\neq 1$ and $\gcd(n-3,\ell)=1$. Consider $H=P_4 \cup \left( \frac{n}{2} -2 \right) P_2$. Note that $H$ is a pendant graph with $n$ vertices and $\frac{n}{2} + 1$ edges. By Lemma \ref{pendantN}, $\overline{H}$ is $N$-AW if and only if $\gcd(n-3,\ell)=1$. Since $\gcd(n-1,\ell)\neq 1$ it follows from Proposition \ref{evenmatch} that $\max(n,\ell)=\binom{n}{2}-\left( \frac{n}{2} + 1 \right)$. Suppose $\max(n,\ell)=\binom{n}{2}-\left( \frac{n}{2}+1 \right)$. Then $\gcd(n-1,\ell)\neq 1$, since otherwise $\max(n,\ell) = \binom{n}{2}-\frac{n}{2}$ by Proposition \ref{evenmatch}. By Lemma \ref{degreebound}, if $G$ is $N$-AW with $|E(\overline{G})|=\frac{n}{2}+1$ then $\Delta(\overline{G})\leq 2$. So by Theorem \ref{thm:maxdegree2} each connected component of $\overline{G}$ is either $P_2$ or $P_4$. The only such graph with $\frac{n}{2}+1$ edges is $H$. Thus $\gcd(n-3,\ell)=1$. It is clear that $H$ is the only pendant graph of order $\frac{n}{2}+1$. \end{proof} In the next proposition, we resolve the $k=2$ case of Theorem \ref{thm:extremalpendant}. The proof considers the possible degree sequences of the complements of $(n,\ell)$-extremal graphs. To ease our explanation we introduce a notation. Let a \emph{$d$-vertex} refer to a vertex of degree $d$. A \emph{$d^+$-vertex} is a vertex of degree $d$ or more. \begin{lem}\label{lem:dvertices} Suppose $G$ is an $N$-AW graph. Then any $d$-vertex in $\overline{G}$ with $d\geq 2$ must have at least $d-1$ neighbors that are $2^+$-vertices. \end{lem} \begin{proof} Suppose that $v$ is a $d$-vertex in $\overline{G}$ with $d\geq 2$ and that $v$ has fewer than $d-1$ neighbors that are $2^+$ vertices. Then $v$ has two neighbors of degree $1$ in $\overline{G}$, which results in $G$ having $N$-twins. By Corollary \ref{twins}, $\overline{G}$ is not $N$-AW. \end{proof} \begin{prop}\label{prop:plus2} Let $n,\ell\in\mathbb{N}$ be even and $n\geq 6$. Then \begin{center} $\max(n,\ell)=\binom{n}{2}- \left( \frac{n}{2} + 2 \right)$ if and only if $\gcd(n-1,\ell)\neq 1$, $\gcd(n-3,\ell)\neq 1$, and $\gcd(n-5,\ell)=1$. \end{center} Moreover, the $(n,\ell)$-extremal graphs in this case are precisely the complements of pendant graphs of order $n$ and size $\frac{n}{2}+2$: $\overline{(P_3 \astrosun K_1)\cup \frac{n-6}{2} P_2}$ and $\overline{2P_4\cup \frac{n-8}{2} P_2}$ with the latter only possible when $n\geq 8$. \end{prop} \begin{proof} Suppose $\gcd(n-1,\ell)\neq 1$, $\gcd(n-3,\ell)\neq 1$ and $\gcd(n-5,\ell)=1$. By Proposition \ref{evenmatch} and Proposition \ref{prop:plus1} $\max(n,\ell)\leq \binom{n}{2}- \left( \frac{n}{2}+2 \right)$. Consider $H = (P_3 \astrosun K_1)\cup \frac{n-6}{2}P_2$ which has size $\frac{n}{2}+2$. By Lemma \ref{pendantN}, $\overline{H}$ is $N$-AW if and only if $\gcd(n-5,\ell)=1$. So, $\max(n,\ell)=\binom{n}{2}- \left( \frac{n}{2}+2 \right)$. Now suppose that $\max(n,\ell)=\binom{n}{2}- \left( \frac{n}{2}+2 \right)$. Then $\gcd(n-1,\ell)\neq 1$ and $\gcd(n-3,\ell)\neq 1$ by Propositions \ref{evenmatch} and \ref{prop:plus1}. We will describe all $G$ such that $G$ is $N$-AW and $E(\overline{G})=\frac{n}{2} + 2$ and show either that these graphs are not $(n,\ell)$-extremal or that they are $N$-AW if and only if $\gcd(n-5,\ell)=1$. Suppose that $G$ is $N$-AW with $E(\overline{G})=\frac{n}{2}+2$. The degree sum of $\overline{G}$ is $n+4$. By Lemma \ref{degreebound}, $\Delta(\overline{G}) \leq 3$. To avoid $N$-twins in $G$, $\overline{G}$ can have at most one $0$-vertex. Thus the only possible degree sequences for $\overline{G}$ are $d_0 = (3,3,1,1,\dots,1)$, $d_1=(3,2,2,1,1,\dots,1)$, $d_2 = (2,2,2,2,1,1,\dots,1)$, $d_3=(3,3,2,1,1,\dots,1,0)$, $d_4=(3,2,2,2,1,1,\dots,1,0)$, and $d_5 = (2,2,2,2,2,1,1,\dots,1,0)$. For a graph with degree sequence $d_0$ note that each of the 3-vertices must have at least two pendant neighbors. So by Lemma~\ref{lem:dvertices}, no graph with degree sequence $d_0$ is $N$-AW. If $\overline{G}$ has degree sequence $d_1$, Lemma \ref{lem:dvertices} implies that all $2^+$-vertices are in the same component. The other components of $\overline{G}$ must be a matching. So our options are $G' \cup \frac{n-4}{2} P_2$ where $G'$ is shown in Figure \ref{fig:G1} or $H=(P_3 \astrosun K_1)\cup \frac{n-6}{2}P_2$. Note $\overline{H}$ is $N$-AW if and only if $\gcd(n-5,\ell)=1$ as shown above. The graph $G'$ has order $4$ and size $4$. Moreover, given the initial labeling $\bf{0}_1$ we can achieve the $0$ labeling by toggling the vertices $b$ and $c$ each $-1$ times. Thus, $T_{V(G_1)}^A (1) = \{-2\}$. By Lemma \ref{matrixLO}, $G'$ is $A$-AW because the adjacency matrix is invertible. The graph $P_4 = P_2 \astrosun K_1$ has order $4$, size $3$, is $A$-AW by Lemma~\ref{pendwin}(\ref{pendwinall}), and, by Lemma~\ref{pendwin}(\ref{pendtoggen}), $T_{V(P_4)}^A(1) = \{-2\}$. Thus, $\overline{G'\cup \frac{n-4}{2} P_2}$ is not $(n,\ell)$-extremal by Corollary~\ref{extswitch}. \begin{figure} \begin{center} \begin{tikzpicture} \node at (0,-.75) (1) {a}; \node at (.75,0) (2) {b}; \node at (.75,-1.5) (3) {d}; \node at (1.5,0) (4) {c}; \draw (1) -- (2); \draw (1) -- (3); \draw (2) -- (3); \draw (2) -- (4); \end{tikzpicture} \caption{One option for the non-matching component of a graph with degree sequence $d_1$ in the proof of Proposition \ref{prop:plus2}.} \label{fig:G1} \end{center} \end{figure} If $\overline{G}$ has degree sequence $d_2$ then $\Delta(\overline{G})=2$ and so, by Theorem \ref{thm:maxdegree2}, each component is $P_2$ or $P_4$. This leaves just $2P_4 \cup \frac{n-8}{2}P_2$ which is a pendant graph and thus, by Lemma \ref{pendantN}, $N$-AW if and only if $\gcd(n-5,\ell)=1$. Suppose $\overline{G}$ has degree sequence $d_3$, $d_4$ or $d_5$. In these cases $\overline{G}$ has an isolated vertex so by Corollary~\ref{extdom}, any component with non-pendant vertices has no pendant vertices. Degree sequence $d_3$ is impossible because there are not enough $2^+$ vertices to be in a component with a $3$-vertex. If $\overline{G}$ has degree sequence $d_4$, this implies one of the components must have odd degree sum which is impossible. If $\overline{G}$ has degree sequence $d_5$ then $\Delta(\overline{G})= 2$. By Theorem \ref{thm:maxdegree2} if $G$ is $(n,\ell)$-extremal then each component of $\overline{G}$ is either $P_2$ or $P_4$. Since the number of $2$-vertices is odd no such graph exists. Thus if $\max(n,\ell)=\binom{n}{2} - \left( \frac{n}{2}+2 \right)$ then $\gcd(n-5,\ell)=1$. Moreover the unique $(n,\ell)$-extremal graphs are $\overline{(P_3 \astrosun K_1)\cup \frac{n-6}{2} P_2}$ and $\overline{2P_4\cup \frac{n-8}{2} P_2}$ which are the complements of the only pendant graphs of order $n$ and size $\frac{n}{2} + 2$. \end{proof} In the next proposition, we resolve the case $k=3$ of Theorem \ref{thm:extremalpendant}. We use the techniques of Proposition \ref{prop:plus2}, but must analyze more cases. \begin{prop}\label{prop:plus3} Let $n,\ell\in\mathbb{N}$ be even and $n\geq 8$. Then \begin{center} $\max(n,\ell)=\binom{n}{2}- \left( \frac{n}{2} + 3 \right)$ if and only if $\gcd(n-2k-1,\ell)\neq 1$ for $0\leq k\leq 2$, and $\gcd(n-7,\ell)=1$. \end{center} In this case the unique $(n,\ell)$-extremal examples are exactly those graphs whose complements are pendant graphs with $\frac{n}{2}+3$ edges: $(C_3 \astrosun K_1) \cup \frac{n-6}{2}P_2$, $(P_4 \astrosun K_1) \cup \frac{n-8}{2} P_2$, $(K_{1,3} \astrosun K_1) \cup \frac{n-8}{2}P_2$, $(P_3 \astrosun K_1) \cup P_4 \cup \frac{n-10}{2}P_2$, and $3P_4 \cup \frac{n-12}{2} P_2$. \end{prop} \begin{proof} Suppose $\gcd(n-2k-1,\ell)\neq 1$ for $0\leq k\leq 2$, and $\gcd(n-7,\ell)=1$. By Propositions \ref{evenmatch}, \ref{prop:plus1} and \ref{prop:plus2} we know $\max(n,\ell) \leq \binom{n}{2}- \left( \frac{n}{2}+3 \right)$. Consider $G=(P_4 \astrosun K_1)\cup \frac{n-8}{2}P_2$ which has $\frac{n}{2}+3$ edges. Since $\gcd(n-2(3)-1,\ell)=1$, $G$ is $(N,\ell)$-AW by Lemma \ref{pendantN}. So $\max(n,\ell) = \binom{n}{2}- \left( \frac{n}{2}+3 \right)$. Suppose that $\max(n,\ell)=\binom{n}{2} - \left( \frac{n}{2}+3 \right)$. Then $\gcd(\ell,2)\neq 1$ and $\gcd(n-2k-1,\ell)\neq 1$ for $0\leq k\leq 2$ by Propositions \ref{evenmatch}, \ref{prop:plus1} and \ref{prop:plus2}. We describe all $G$ such that $G$ is $(N,\ell)$-AW and $E(\overline{G}) = \frac{n}{2} + 3$ and show that these graphs are either not $(n,\ell)$-extremal or are $(N,\ell)$-AW if and only if $\gcd(n-7,\ell)=1$. Suppose that $G$ is $(N,\ell)$-AW with $E(\overline{G})= \frac{n}{2}+3$. The degree sum of $\overline{G}$ is $n+6$. By Lemma \ref{degreebound}, $\Delta(\overline{G}) \leq 4$. To avoid $N$-twins in $G$, $\overline{G}$ can have at most one $0$-vertex. We first consider the case when $\overline{G}$ has a $0$-vertex. By Corollary \ref{extdom} we know that $\overline{G}$ does not have a pendant vertex that is not part of a $P_2$ component. Thus the degree sequence of $\overline{G}$ has an even number of $1$-vertices. Since the total number of vertices is even and we have a $0$-vertex, we know the number of $2^+$-vertices will be odd. By considering all integer partitions of $7$ that when added to $(1,1,\dots,1,0)$ will satisfy having an even number of $2^+$ vertices and no $5^+$-vertex we get the following possible degree sequences: \begin{itemize} \item $d_0 = (4,4,2,1,1,\dots,1,0)$ \item $d_1 = (4,3,3,1,1,\dots, 1,0)$ \item $d_2 = (4,2,2,2,2,1,\dots,1,0)$ \item $d_3 = (3,3,2,2,2,1,\dots, 1,0)$ \item $d_4 = (2,2,2,2,2,2,2,1,\dots,1,0)$ \end{itemize} By Lemma \ref{lem:dvertices} degree sequences $d_0$ and $d_1$ can not have a realization that is $(N,\ell)$-AW for any $\ell$. In the case of $d_2$, by Corollary \ref{extdom} and Lemma \ref{lem:dvertices} all the $2^+$-vertices form a component with no $1$-vertices. The only realization of $(4,2,2,2,2)$ is the bowtie graph shown in Figure \ref{bowtie}. Let $H_{d_2}$ be the bowtie graph along with the required number of $P_2$ components and an isolated vertex. By Theorem \ref{subsetjoindom}, $\overline{H_{d_2}}$ is $(N,\ell)$-AW if and only if $H_{d_2}-P_1$ is $(A,\ell)$-AW. By row reducing the adjacency matrix, we find that the bowtie graph (and thus $H_{d_2}-P_1$) is $(A,\ell)$-AW if and only if $\ell$ is odd. Thus, $\overline{H_{d_2}}$ is $(N,\ell)$-AW if and only if $\ell$ is odd, and so the graph corresponding to $d_2$ is not $(n,\ell)$-extremal by Proposition \ref{prop:triangle}. \begin{figure} \begin{center} \begin{tikzpicture} \coordinate (1) at (0,0); \coordinate (2) at (.5,.5); \coordinate (3) at (.5,-.5); \coordinate (4) at (-.5, .5); \coordinate (5) at (-.5,-.5); \fill (1) circle (2pt); \fill (2) circle (2pt); \fill (3) circle (2pt); \fill (4) circle (2pt); \fill (5) circle (2pt); \draw (1) -- (2); \draw (1) -- (3); \draw (1) -- (4); \draw (1) -- (5); \draw (4) -- (5); \draw (2) -- (3); \coordinate (6) at (2,-.5); \coordinate (7) at (3,-.5); \coordinate (8) at (2,.25); \coordinate (9) at (3,.25); \coordinate (10) at (2.5, .75); \fill (6) circle (2pt); \fill (7) circle (2pt); \fill (8) circle (2pt); \fill (9) circle (2pt); \fill (10) circle (2pt); \draw (6) -- (7); \draw (7)-- (9); \draw (6) -- (8); \draw (8) -- (9); \draw (8) -- (10); \draw (9) -- (10); \end{tikzpicture} \end{center} \caption{The bowtie graph with degree sequence $(4,2,2,2,2)$ on the left, and the house graph with degree sequence $(3,3,2,2,2)$ on the right. Both graphs appear in the proof of Proposition \ref{prop:plus3}.} \label{bowtie} \end{figure} Again, by Corollary \ref{extdom} and Lemma \ref{lem:dvertices} we find that for the case of $d_3$, the $2^+$-vertices form a component with no $1$-vertices. Considering the cases in which the two $3$-vertices are adjacent and when they are not, we have that the only realizations of $(3,3,2,2,2)$ are the house graph (shown in Figure \ref{bowtie}) and $K_{2,3}$. Let $H_{d_3}$ be the house graph along with the required number of $P_2$ components and an isolated vertex. As in the previous paragraph, we use Theorem \ref{subsetjoindom} to determine the $N$-winnability of $\overline{H_{d_3}}$ by row reducing the adjacency matrix of the house graph, and we find $\overline{H_{d_3}}$ is never $(N,\ell)$-AW. Note that in $K_{2,3}$, the two $3$-vertices are $A$-twins and thus, by Theorem \ref{subsetjoindom}, the complement of this realization is never $(N,\ell)$-AW. In the case of $d_4$ we note that $\Delta(\overline{G})=2$. By Theorem \ref{thm:maxdegree2} if $G$ is $(n,\ell)$-extremal then each component of $\overline{G}$ is either $P_2$ or $P_4$. Since a realization of $d_4$ has a $P_1$ component, there is no such $(n,\ell)$-extremal graph. Therefore, given the hypotheses, there are no $(n,\ell)$-extremal graphs with $\binom{n}{2}-\left(\frac{n}{2}+3\right)$ edges and a dominating vertex. Now suppose there is no $0$-vertex in $\overline{G}$. To get a degree sum of $n+6$ we need to add integer partitions of $6$ to $(1,1,\dots,1)$. Considering all integer partitions of $6$ that have parts of size at most $3$ and adding these to $(1,1,\dots,1)$ we get the following possible degree sequences: \begin{itemize} \item $d_5 = (4,4,1,1,\dots,1)$ \item $d_6 = (4,3,2,1,\dots,1)$ \item $d_7 = (4,2,2,2,1,\dots,1)$ \item $d_8 = (3,3,3,1,\dots,1)$ \item $d_9 = (3,3,2,2,1,\dots,1)$ \item $d_{10} = (3,2,2,2,2,1,\dots,1)$ \item $d_{11}= (2,2,2,2,2,2,1,\dots,1)$ \end{itemize} We eliminate $d_5$ and $d_6$ using Lemma \ref{lem:dvertices}. In the case of $d_7$ all of the $2$-vertices must be adjacent to the $4$-vertex. Considering the possible adjacencies among the $2$-vertices the possible graphs are $H_{d_7} =K_{1,3}\astrosun K_1\cup \frac{n-4}{2}P_2$ and the graph $G_2\cup \frac{n-6}{2} P_2$ where $G_2$ is given in Appendix \ref{Appendix}. Since $H_{d_7}$ is a pendant graph we know $\overline{H_{d_7}}$ is $(N,\ell)$-AW if and only if $\gcd(n-7,\ell)=1$ by Lemma \ref{pendantN}. By Lemma \ref{matrixLO} and the fact that the adjacency matrix of $G_2$ is invertible we know $G_2$ is $(A,\ell)$-AW. From Appendix \ref{Appendix} $G_2$ has order $6$, size $6$, and $T_{G_2}^A(1) = -2$. However, $P_3\astrosun K_1$ is $(A,\ell)$-AW by Lemma \ref{pendwin}(\ref{pendwinall}), has order $6$, size $5$, and has $T_{V(P_3\astrosun K_1)}^A =-2$ by Lemma \ref{pendwin}(\ref{pendtreetog}). Thus by Corollary \ref{extswitch}, $\overline{G_2\cup \frac{n-6}{2} P_2}$ is not $(n,\ell)$-extremal. Consider degree sequence $d_8$. By Lemma \ref{lem:dvertices} all $3$-vertices must be adjacent to each other. Thus the only possible graph is $H_{d_8} = (C_3 \astrosun K_1)\cup \frac{n-6}{2} P_2$. Since $H_{d_8}$ is a pendant graph we know $\overline{H_{d_8}}$ is $(N,\ell)$-AW if and only if $\gcd(n-7,\ell)=1$ by Lemma \ref{pendantN}. For degree sequence $d_9$ again by Lemma \ref{lem:dvertices} all $2^+$-vertices must be in the same component. We generate all possible graphs with degree sequence $d_9$ by considering whether or not the two $3$-vertices are adjacent. If the two $3$-vertices are not adjacent (in the complement graph) then they each must be adjacent to both of the degree $2$ vertices, resulting in $G_{d_9}$ which is given Figure \ref{d9}. Since $\overline{G_{d_9} \cup \frac{n-6}{2} P_2}$ has $N$-twins ($v$ and $w$ in Figure \ref{d9}) this graph is not $(N,\ell)$-AW for any $\ell$ by Corollary \ref{twins}. \begin{figure} \begin{center} \begin{multicols}{2} \begin{tikzpicture} \coordinate (1) at (0,0); \coordinate (2) at (.5,.5); \coordinate (3) at (.5,-.5); \coordinate (4) at (1,0); \coordinate (5) at (1.5,0); \coordinate (6) at (-.5,0); \fill (1) circle (2pt); \fill (2) circle (2pt); \fill (3) circle (2pt); \fill (4) circle (2pt); \fill (5) circle (2pt); \fill (6) circle (2pt); \draw (1) -- (2); \draw (1) -- (3); \draw (2) -- (4); \draw (3) -- (4); \draw (4) -- (5); \draw (1) -- (6); \draw[above] (2) node {$v$}; \draw[below] (3) node {$w$}; \end{tikzpicture} \begin{tikzpicture} \coordinate (1) at (0,0); \coordinate (2) at (.5,.5); \coordinate (3) at (.5,-.5); \coordinate (4) at (1,0); \fill (1) circle (2pt); \fill (2) circle (2pt); \fill (3) circle (2pt); \fill (4) circle (2pt); \draw (1) -- (2); \draw (1) -- (3); \draw (2) -- (4); \draw (3) -- (4); \draw (1) -- (4); \draw[above] (2) node {$v$}; \draw[below] (3) node {$w$}; \end{tikzpicture} \end{multicols} \end{center} \caption{At left, the graph $G_{d_9}$ - the non-matching component of the only graph with degree sequence $d_9$ from Proposition \ref{prop:plus3} for which the two $3$-vertices are not adjacent. At right, the graph $G_{d_9}'$ - the non-matching component of only graph with degree sequence $d_9$ in which the two $3$-vertices are adjacent and have two neighbors in common in the proof of Proposition \ref{prop:plus3}.} \label{d9} \end{figure} Suppose the two $3$-vertices are adjacent. We consider cases based on their number of common neighbors. If there are no common neighbors then, to avoid twins, we get $(P_4\astrosun K_1) \cup \frac{n-8}{2} P_2$ or $G_3 \cup \frac{n-6}{2}P_2$ where $G_3$ is given in Appendix \ref{Appendix}. For the former, the complement is $(N,\ell)$-AW if and only if $\gcd(n-7,\ell)=1$ by Lemma \ref{pendantN}. For the latter we apply Corollary \ref{extswitch}. By Lemma \ref{matrixLO}, $G_3\cup \frac{n-6}{2} P_2$ is $(A,\ell)$-AW. By Appendix \ref{Appendix} graph $G_3$ has order $6$, size $6$, and $T_{G_3}^A(1) = -2$. However, $P_3\astrosun K_1$ is $(A,\ell)$-AW by Lemma \ref{pendwin}(\ref{pendwinall}), has order $6$, size $5$, and has $T_{V(P_3\astrosun K_1)}^A =-2$ by Lemma \ref{pendwin}(\ref{pendtreetog}). Thus by Corollary \ref{extswitch}, $\overline{G_3\cup \frac{n-6}{2} P_2}$ is not $(n,\ell)$-extremal. Now suppose the two degree $3$ vertices have one neighbor in common. Then we get the graph $G_4$ in Appendix \ref{Appendix}. We see $G_4$ has order $6$, size $6$, and has $T_{V(G_4)}^A(1) =\{-4\}$. Also $G_4$ is $(A,\ell)$-AW by Lemma \ref{matrixLO}. However, by Lemma \ref{pendwin}(\ref{pendtoggen}), $(P_2\cup P_1) \astrosun K_1 = P_4 \cup P_2$ has $T_{V(P_2\cup P_1)\astrosun K_1)}^A(1) = \{-4\}$. So by Corollary \ref{extswitch} $G_4$ is not $(n,\ell)$-extremal. Finally, suppose the two degree $3$ vertices have two neighbors in common. This results in $G_{d_9}'$ given on the right in Figure \ref{d9}. Since $\overline{G_{d_9}'\cup \frac{n-4}{2}}$ has $N$-twins ($v$ and $w$), it is not $(N,\ell)$-AW for any $\ell$ by Corollary \ref{twins}. Next consider degree sequence $d_{10}$. In this case it is not necessarily true that all $2^+$-vertices need to be in the same component. However, if the $2^+$ vertices form more than one component, it would have to be the case that one component had a $3$-vertex and two $2$-vertices by Lemma \ref{lem:dvertices}. In this case, the arguments in Proposition \ref{prop:plus2} for degree sequence $d_1$ apply. Now suppose all the $2^+$-vertices are in the same component. We consider cases based on the degrees of the vertices adjacent to the $3$-vertex. This could either be two $2$-vertices and one $1$-vertex or three $2$-vertices. Suppose there are two $2$-vertices and one $1$-vertex adjacent to the $3$-vertex. The two $2$-vertices each have an additional neighbor (not the $3$-vertex and not each other since then the $3$-vertex would be forced to have three neighbors of degree $2$). Call these additional neighbors $v$ and $w$. If one of these vertices is degree $1$ we end up with the graph $G_5$ in Appendix \ref{Appendix}. By Lemma \ref{pendwin}(\ref{pendtoggen}), $P_4 \cup P_4$ has $T_{V(P_4\cup P_4)}^A =-4$, and we find $G_5 \cup \frac{n-8}{2}P_2$ is not $(n,\ell)$-extremal by Corollary \ref{extswitch} . If $v$ and $w$ each have degree $2$ we consider the possibility that they are adjacent to each other and if they are not. This yields the graphs in Figure \ref{d9-2} and graph $G_6$ in Appendix \ref{Appendix}. The graph in Figure \ref{d9-2} has a labeling that is not $(A,\ell,s)$-winnable for all $s$, namely the labeling where one of the pendant vertices adjacent to a 2-vertex has label 1 and the remaining vertices have label 0. This would make $G$ not $(N,\ell)$-AW by Theorem~\ref{pendantremove}(\ref{penneighbor}). Since $P_3 \astrosun K_1$ has $T_{V(P_3\astrosun K_1)}^A =-2$ by Lemma \ref{pendwin}(\ref{pendtreetog}), $G_6 \cup \frac{n-6}{2}P_2$ is not $(n,\ell)$-extremal by Corollary \ref{extswitch}. Now suppose there are three $2$-vertices adjacent to the $3$-vertex. Considering whether two of the $2$-vertices are adjacent to each other or not we get graphs $G_7$ and $G_8$ in Appendix \ref{Appendix}. Since $T_{V(P_4 \cup 2P_2)}^A(1) = -6$ and $T_{V(P_4\cup P_4)}^A(1)=-4$ by Lemma \ref{pendwin}(\ref{pendtoggen}), we know neither $G_7 \cup \frac{n-6}{2}P_2$ nor $G_8 \cup \frac{n-8}{2}P_2$ is $(n,\ell)$-extremal by Corollary \ref{extswitch}. \begin{figure} \begin{center} \begin{tikzpicture} \coordinate (1) at (0,0); \coordinate (2) at (1,0); \coordinate (3) at (2,1); \coordinate (4) at (2,-1); \coordinate (5) at (3,1); \coordinate (6) at (3,-1); \coordinate (7) at (4,-1); \coordinate (8) at (4,1); \fill (1) circle (2pt); \fill (2) circle (2pt); \fill (3) circle (2pt); \fill (4) circle (2pt); \fill (5) circle (2pt); \fill (6) circle (2pt); \fill (7) circle (2pt); \fill (8) circle (2pt); \draw (1) -- (2); \draw (2) -- (3); \draw (2) -- (4); \draw (3) -- (5); \draw (4) -- (6); \draw (6) -- (7); \draw (5) -- (8); \end{tikzpicture} \end{center} \caption{The second possibility for a high degree component for degree sequence $d_{10}$ in which the degree $3$ vertex is adjacent to two degree $2$ vertices and one degree $1$ component.} \label{d9-2} \end{figure} If $\overline{G}$ has degree sequence $d_{11}$ then $\Delta(\overline{G})=2$. By Theorem \ref{thm:maxdegree2} if $G$ is $(n,\ell)$-extremal then each component of $\overline{G}$ is either $P_2$ or $P_4$. Thus, the graph must be $3P_4 \cup \frac{n-12}{2} P_2$, a pendant graph. This is $(N,\ell)$-AW if and only if $\gcd(n-7,\ell)=1$ by Lemma \ref{pendantN}. Therefore every possible graph that is $(N,\ell)$-AW and has $E(\overline{G}) = \frac{n}{2}+3$ is either not $(n,\ell)$-extremal or has $\gcd(n-7,\ell)=1$, as desired. \end{proof} \noindent\emph{Proof of Theorem \ref{thm:extremalpendant}}. The cases $0\leq k\leq 3$ are true by Propositions \ref{evenmatch}, \ref{prop:plus1}, \ref{prop:plus2}, \ref{prop:plus3}, respectively. \hfill $\qed$ \section{Open Problems} We close with three open problems related to our results. \bigskip \noindent (1) \emph{Does Theorem~\ref{thm:extremalpendant} hold for $k \ge 4$?} We made much progress on this result by considering the possible degree sequences. However, when $k=4$, there are $37$ partitions of $7$ and $8$. Even with the additional restriction of Lemma \ref{degreebound} there are $23$ different degree sequences to consider. Thus, we need an alternative method to solve the general problem. \bigskip \noindent (2) \emph{What are the graphs of maximum size that are $(N,\ell)$-AW for all $\ell$?} The best candidates we have found are complements of pendant trees, which have size $\bc{n}{2}-(n-1)$. They are all $(N,\ell)$-AW for all $\ell$, but it is not clear that they are $(n,\ell)$-extremal. \bigskip \noindent (3) \emph{What are the $(n,\ell)$-extremal graphs for other Lights Out games, such as the adjacency game?} \bibliographystyle{amsalpha}

https://arxiv.org/abs/2108.00448

Small order asymptotics for nonlinear fractional problems

We study the limiting behavior of solutions to boundary value nonlinear problems involving the fractional Laplacian of order $2s$ when the parameter $s$ tends to zero. In particular, we show that least-energy solutions converge (up to a subsequence) to a nontrivial nonnegative least-energy solution of a limiting problem in terms of the logarithmic Laplacian, i.e., the pseudodifferential operator with Fourier symbol $\ln(|\xi|^2)$. These results are motivated by some applications of nonlocal models where a small value for the parameter $s$ yields the optimal choice. Our approach is based on variational methods, uniform energy-derived estimates, and the use of a new logarithmic-type Sobolev inequality.

\section{Introduction} Nonlocal operators provide an important tool to model phenomena with anomalous diffusive behavior. Examples of this can be found in a wide variety of situations: in water waves, crystal dislocations, phase transitions, peridynamics, and finance, see \emph{e.g.} \cite{BV16,HB10,L04}. A particularly illustrative example of how the nonlocality can play a prominent role in a model comes from population dynamics \cite{PV18}, where a nonlinear (logistic-type) equation in terms of a fractional Laplacian of order $2s$ is used to describe the movement of a species. Here, a small order $s$ describes an almost static population (but capable of moving quickly long distances), whereas $s$ near $1$ relates to a very dynamic species which mostly moves short distances. Therefore, the parameter $s$ accounts for different dispersal strategies. Interestingly, in \cite{PV18}, it is shown that a very small order $s$ can be the best strategy for survival if the habitat is not too fragmented or not too hostile in average. Similarly, a small value for the exponent $s$ yields the optimal choice in other applications such as optimal control \cite{SV17}, approximation of fractional harmonic maps \cite{ABS21}, and fractional image denoising \cite{AB17}. This motivates the research of small order asymptotics of nonlinear fractional problems. Moreover, the mathematical structure that arises in this analysis is rich and interesting in itself from a theoretical point of view. In this paper, we develop a method to study the small order asymptotics of subcritical nonlinear problems. To make the main ideas in our arguments more transparent, we focus on a model problem with power-type nonlinearity, and we refer to Remark \ref{ext:rmk} for a discussion on extensions and generalizations. To be more precise, we consider the following fractional Dirichlet problem \begin{align}\label{subcritical:intro} (-\Delta)^s u_s = |u_s|^{p_s-2}u_s\quad \text{ in }\Omega,\qquad u_s=0\quad \text{ in }\mathbb{R}^N\backslash\Omega, \end{align} where $N\geq 1$, $s\in(0,\frac{1}{4})$, $p_s\in(2,2^*_s)$ is superlinear and subcritical, $2^*_s=\frac{2N}{N-2s}$ is the critical Sobolev exponent, and $\Omega\subset \mathbb{R}^N$ is a bounded open Lipschitz set. Here $(-\Delta)^s$ denotes the integral fractional Laplacian given by the hypersingular integral in the principal value sense \begin{align*} (-\Delta)^s u(x) = c_{N,s}\, p.v.\int_{\mathbb{R}^N}\frac{u(x)-u(y)}{|x-y|^{N+2s}}\ dy,\qquad c_{N,s}:= 4^s\pi^{-\frac{N}{2}}s(1-s)\frac{\Gamma(\tfrac{N}{2}+s)}{\Gamma(2-s)}. \end{align*} \medskip Our goal is to answer the following question: \begin{align}\label{q} \text{\emph{What is the limit of a solution $u_s$ of~\eqref{subcritical:intro} as $s\to 0^+$?} } \end{align} Note that (at least formally) both sides of equation \eqref{subcritical:intro} behave similarly in the limit, namely, $(-\Delta)^s u\to u$ and $|u|^{p_s-2}u\to u$ as $s\to 0^+$. Therefore, to answer question~\eqref{q}, it is necessary to consider the first order expansion (with respect to $s$) on both sides of equation~\eqref{subcritical:intro}. This leads us naturally to the logarithmic Laplacian $L_\Delta$, which was introduced in \cite{CW19} and has the following pointwise evaluation \begin{align*} L_\Delta u(x) = c_N\, p.v.\int_{B_1(x)}\frac{u(x)-u(y)}{|x-y|^{N}}\ dy -c_N\int_{\mathbb{R}^N\backslash B_1(x)}\frac{u(y)}{|x-y|^N}\ dy + \rho_N\, u(x), \end{align*} where $c_N$ and $\rho_N$ are explicit constants (see~\eqref{cN:def} and~\eqref{rho:def}) and $B_1(x)$ is the open ball in $\mathbb{R}^N$ of radius 1 centered at $x$. Furthermore, as shown in \cite{CW19}, $L_\Delta$ is a pseudodifferential operator with Fourier symbol $2\ln(|\xi|)$ and, for $u\in C^\infty_c(\mathbb{R}^N)$, \begin{align*} L_\Delta u = \lim_{s\to 0^+}\frac{d}{ds}(-\Delta)^s u. \end{align*} We give the following answer to question~\eqref{q}. We refer to Section \ref{not:sec} for the precise definitions of least-energy solutions and related notation. \begin{theorem}\label{main:thm}Let $N\geq 1$ and let $\Omega\subset \mathbb{R}^N$ be a bounded open Lipschitz set. For $s\in(0,\frac{1}{4})$, let $p_s:=p(s)$, where \begin{align}\label{p:hyp} p\in C^1([0,\tfrac{1}{4}]),\quad 2<p(s)<2^*_s:=\frac{2N}{N-2s}\ \ \text{ for $s\in\left(0,\tfrac{1}{4}\right)$,}\quad \text{ and } \quad p'(0)\not\in\left\{0,\tfrac{4}{N}\right\}. \end{align} Let $(s_k)_{k\in\mathbb{N}}\subset(0,\frac{1}{4})$ be such that $\lim_{k\to\infty}s_k=0$ and, for $k\in\mathbb{N},$ let $u_{s_k}\in\mathcal{H}^{s_k}_0(\Omega)$ be a least-energy solution of \begin{align}\label{equsk} (-\Delta)^{s_k} u_{s_k} = |u_{s_k}|^{p_{s_k}-2}u_{s_k}\quad \text{ in }\Omega,\qquad u_{s_k}=0\quad \text{ in }\mathbb{R}^N\backslash\Omega. \end{align} Then there is a least-energy solution $u_0\in \mathbb{H}(\Omega)\backslash\{0\}$ of \begin{align}\label{log:prob:intro} L_\Delta u_0 =\mu\ln(|u_0|)u_0\ \ \text{ in }\Omega,\qquad u_0=0\ \ \text{ in }\mathbb{R}^N\backslash \Omega, \qquad \mu :=p'(0) \end{align} such that, passing to a subsequence, \begin{align}\label{cL2} u_{s_k}\to u_0\quad \text{ in $L^2(\mathbb{R}^N)$ as $k\to\infty$.} \end{align} Moreover, \begin{align}\label{energies} \lim_{k\to\infty}\frac{1}{s_k}J_{s_k}(u_{s_k})=J_0(u_0)>0\qquad \text{ and }\qquad \lim_{k\to\infty}\|u_{s_k}\|_{s_k} = |u_0|_2. \end{align} \end{theorem} Here, $\mathcal{H}^{s_k}_0(\Omega)$ and $\mathbb{H}(\Omega)$ are suitable Hilbert spaces and $J_{s_k}$ and $J_0$ are the \emph{energy functionals} associated to~\eqref{equsk} and~\eqref{log:prob:intro} (see Section \ref{not:sec}). Note that~\eqref{p:hyp} includes the paradigmatic case \begin{align*} p_s=p(s):=\lambda\, 2^*_s+(1-\lambda)\, 2\qquad \text{ for some $\lambda\in(0,1)$.} \end{align*} The condition $p_s<2^*_s$ is needed to guarantee the existence of solutions (for example, if $p_s=2^*_s$, the solvability of \eqref{subcritical:intro} depends on the domain $\Omega$; see \emph{e.g.} \cite{HSS21} for some existence and nonexistence results for \eqref{subcritical:intro} in this case for any $s>0$). An interesting phenomenon is that, for~\eqref{log:prob:intro}, the ``logarithmic subcriticality" is reflected on the scalar factor $\mu$, where the fact that $\mu=p'(0)<\frac{4}{N}$ is crucial in our arguments to obtain compactness. See Remark \ref{open:rmk} for a discussion on the case $p'(0)\in\{0,\frac{4}{N}\}$. One of the main obstacles to show Theorem~\ref{main:thm} is to find a uniform bound in $\mathbb{H}(\Omega)$ for the sequence of solutions $(u_{s_k})_{k\in\mathbb{N}}$. To show this, we use an ``intermediate inequality" (see Lemma~\ref{prebd}) between the fractional Sobolev inequality and the logarithmic Sobolev inequality together with a series of delicate energy-derived uniform bounds. Furthermore, the existence of a least-energy solution of the limiting problem~\eqref{log:prob:intro} is also important. In particular, we show the following. \begin{theorem}\label{main:thm:2}Let $N\geq 1$ and let $\Omega\subset \mathbb{R}^N$ be a bounded open Lipschitz set. For every $\mu\in(0,\frac{4}{N})$, the problem \begin{align}\label{lambda:problem} L_\Delta u = \mu\ln(|u|)u\quad \text{ in }\Omega,\qquad u_0=0\quad \text{ in }\mathbb{R}^N\backslash \Omega, \end{align} has a least-energy solution $u\in \mathbb{H}(\Omega)\backslash\{0\}$ and \begin{align}\label{mountain:pass} J_0(u)=\inf_{\mathcal{N}_0} J_0 = \inf_{\sigma\in \mathcal T}\max_{t\in[0,1]}J_0(\sigma(t))>0, \end{align} where $\mathcal T:=\{\sigma\in C^0([0,1],\mathbb H(\Omega)): \sigma(0)=0, \sigma(1)\neq 0, J_0(\sigma(1)\leq 0)\}$. Furthermore, all least-energy solutions of \eqref{lambda:problem} do not change sign in $\Omega$. \end{theorem} Here $\mathcal{N}_0$ is the \emph{Nehari manifold} associated to~\eqref{lambda:problem} (see~\eqref{N0:def} below) and~\eqref{mountain:pass} shows that the problem~\eqref{lambda:problem} also has a mountain-pass structure. The Nehari manifold method is our main variational tool to show existence of solutions. The implementation of this approach, however, faces several difficulties in this setting. See Remark~\ref{diff:rmk} below for a discussion of some important differences between~\eqref{subcritical:intro} and~\eqref{lambda:problem}. The most relevant obstacle to show Theorem~\ref{main:thm:2} is that the convergence of energy-minimizing sequences cannot be guaranteed by compact Sobolev embeddings. In particular, for the Hilbert space $\mathbb{H}(\Omega)$ (defined in~\eqref{Hdef}), it is only known that \begin{align}\label{comp:em} \mathbb H(\Omega) \hookrightarrow L^2(\Omega) \quad \text{ is compact} \end{align} (see \cite[Theorem 2.1]{CdP18} or \cite[Corollary 2.3]{LW20}), which alone is not enough to articulate an existence proof. To compensate the loss of the Sobolev inequality, we use a series of logarithmic inequalities. Most prominently, the logarithmic Sobolev inequality \begin{align}\label{si:intro} \frac{2}{N}\int_\Omega \ln(u^{2})|u|^2 \leq\mathcal{E}_L(u,u)+\frac{2}{N}\left(\ln\left(\int_\Omega|u|^2\right)+a_N\right)\int_\Omega|u|^2,\qquad u\in \mathbb{H}(\Omega), \end{align} plays a crucial role. The above inequality is a reformulation of \cite[Theorem~3]{B95} and it appears as a first order expansion of the standard Sobolev inequality at $s=0$ (see the proof of Proposition~\ref{log:prop} below). As mentioned before, the notion of ``subcriticality'' of~\eqref{lambda:problem} is encoded in the assumption that $\mu<\frac{4}{N}$. Indeed, this fact, together with~\eqref{si:intro} and~\eqref{comp:em}, allows to obtain some compactness of energy minimizing sequences, see Proposition~\ref{prop:bd}. \medskip Theorem~\ref{main:thm} can be seen as a nonlinear analog of the results in \cite{FJW20}, where the asymptotic profile of $L^2$-normalized Dirichlet eigenfunctions is studied as $s\to 0^+$. In particular, in \cite{FJW20} it is shown that, if $(s_n)\subset(0,\frac{1}{4})$ is such that $\lim_{n\to 0}s_n=0$, then, up to a subsequence, \begin{align}\label{cLp} \text{$\varphi_{k,s_n}\to \varphi_{k,L}$ in $L^p(\Omega)$ as $n\to\infty$ for any $p\in[1,\infty)$,} \end{align} where $\varphi_{k,s_n}$ and $\varphi_{k,L}$ are the $k$-th eigenfunctions of $(-\Delta)^{s_n}$ and $L_\Delta$, respectively. Although the methods in \cite{FJW20} yield stronger results (compare, for instance, \eqref{cL2} and~\eqref{cLp}), they rely heavily on the linearity of the eigenvalue problem and therefore key arguments in \cite{FJW20} do not have a direct extension to the nonlinear setting. As far as we know, Theorem~\ref{main:thm:2} is the first result regarding existence of solutions to boundary value nonlinear problems involving the logarithmic Laplacian $L_\Delta$. Let us mention some previously known results regarding logarithmic operators. The Dirichlet linear problem for $L_\Delta$ has been studied in \cite{CW19}, which also contains a Faber-Krahn type inequality and maximum principles in weak and strong forms. Further studies on the spectral properties (including sharp upper bounds for the Riesz means and lower bounds for the first Dirichlet eigenvalue of $L_\Delta$) can be found in \cite{LW20}; see also \cite{CV20} for upper and lower bounds for the sum of the first $k$ eigenvalues of $L_\Delta$. The logarithmic Laplacian also arises in the geometric context of the $0$-fractional perimeter, which has been studied recently in \cite{CdLNP21}. In \cite{FKT20} an equation similar to~\eqref{lambda:problem} is considered with $\mu=\frac{4}{N}$ on the sphere and $L_\Delta$ is substituted with a conformally invariant logarithmic operator; in particular, the authors in \cite{FKT20} classify all nonnegative solutions of this equation, which are extremals of a Sobolev logarithmic inequality on the sphere presented in \cite[Theorem~3]{B95}. From the probabilistic point of view, we mention that geometric stable stochastic processes with a logarithmic operator as an infinitesimal generator are studied in \cite{KM13}. Logarithmic-type energies also appear in the recent paper \cite{DCKNP21}, where the asymptotic analysis for the $s$-fractional heat flow as $s\to 0^+$ (and as $s\to 1^-$) is examined. Let us also comment on similar results to Theorem~\ref{main:thm} whenever $s_k\not\to 0$. If $s_k\to 1^-$, then a study of the nonlocal-to-local transition can be found in \cite{FBS20} (in the more general context of the fractional $p$-Laplacian). See also \cite{BS19} and \cite{BS20} for convergence results in bounded and unbounded domains for subcritical Schrödinger equations. In the critical case, the only available result, as far as we know, is \cite{HSS21}, where convergence of solutions (in bounded and unbounded domains) is studied whenever $s_k\to s_0$ for any $s_0>0$ (including the higher order case $s_0>1$). In this setting, convergence of solutions (up to a subsequence) can be guaranteed in the $\mathcal{H}^{s_0-\delta}_0$-sense for any $\delta\in(0,s_0)$. Finally, if $v_s\in \mathcal{H}^s_0(\Omega)$ is a solution of the fractional Poisson problem $(-\Delta)^sv_s=f$ in some bounded open set $\Omega$, then much more is known about the mapping $s\mapsto v_s$. In particular, in \cite{JSW20}, the logarithmic Laplacian $L_\Delta$ is used to characterize the continuity, differentiability, and monotonicity properties of the solution mapping $s\mapsto v_s$ for $s\in[0,1)$, where the case $s=0$ was previously studied in \cite{CW19}. \medskip The paper is organized as follows. In Section \ref{not:sec} we present the relevant definitions for the study of least-energy solutions. In Section \ref{aux:sec} we derive some Taylor expansions for the quadratic forms, give a full proof of the logarithmic Sobolev inequality \eqref{si:intro}, and show that $J_0$ is a $C^1$ functional. The uniform bounds for least-energy solutions are obtained in Section \ref{U:sec}. Section \ref{Sec:MT2} is devoted to the proof of Theorem \ref{main:thm:2} and Section \ref{6:sec} contains the proof of our main result Theorem \ref{main:thm}. Finally, in Section \ref{c:rmks} we include some closing remarks. \section{Definitions and notations}\label{not:sec} For $p\in[1,\infty]$ we use $L^p(\Omega)$ to denote the standard Lebesgue spaces with the norms \begin{align*} |u|_p:=\left(\int_{\Omega}|u|^p\ dx\right)^{\frac{1}{p}}\quad\text{ for }p<\infty\qquad \text{ and }\qquad |u|_\infty:=\sup_{\Omega}|u|. \end{align*} The natural Hilbert space associated to~\eqref{subcritical:intro} is \begin{align}\label{Hs:def} \mathcal{H}^s_0(\Omega):=\{u\in H^s(\mathbb{R}^N)\::\: u=0\text{ in }\Omega\}, \end{align} where $H^s(\mathbb{R}^N)$ is the usual fractional Sobolev space. We say that $u_s\in \mathcal{H}^s_0(\Omega)$ is a (weak) solution of~\eqref{subcritical:intro} if \begin{align}\label{ws:intro} \mathcal{E}_s(u_s,\varphi)=\int_\Omega |u_s|^{p_s-2}u_s\varphi\ dx\qquad\text{ for all } \varphi\in \mathcal{H}^s_0(\Omega), \end{align} where \begin{align*} \mathcal{E}_s(u,v)&:=\left(\frac{c_{N,s}}{2} \int_{\mathbb{R}^N}\int_{\mathbb{R}^N}\frac{(u(x)-u(y))(v(x)-v(y))}{|x-y|^{N+2s}}\ dxdy \right)^\frac{1}{2} \end{align*} is a scalar product in the Hilbert space $\mathcal{H}^s_0(\Omega)$ with norm $\|u\|_s:=\mathcal{E}(u,u)^\frac{1}{2}$. The \emph{energy functional} associated to~\eqref{subcritical:intro} is given by \begin{align}\label{Js:def} J_s:\mathcal{H}^s_0(\Omega)\to\mathbb{R},\qquad J_s(u):=\frac{1}{2}\|u\|_s^2-I_s(u),\qquad I_s(u):=\frac{|u|_{p_s}^{p_s}}{p_s}=\frac{1}{p_s}\int_\Omega|u|^{p_s}\, dx. \end{align} Note that all nontrivial solutions of~\eqref{subcritical:intro} belong to the set \begin{align}\label{cNs:def} \mathcal{N}_s:=\left\{u\in \mathcal{H}^s_0(\Omega)\backslash\{0\}\::\: \|u\|_s^2 = |u|_{p_s}^{p_s}\right\}. \end{align} Then, a solution $u\in \mathcal{N}_s$ is a \emph{least-energy solution} of~\eqref{subcritical:intro} if \begin{align}\label{les} J_s(u)=\inf_{v\in \mathcal{N}_s}J_s(v),\quad \text{or, equivalently,}\quad \|u\|_s^2 =\inf_{v\in \mathcal{N}_s} \|v\|^2_s, \end{align} see Theorem~\ref{existence}. \medskip On the other hand, the natural Hilbert space for the problem~\eqref{lambda:problem} is \begin{align}\label{Hdef} \mathbb{H}(\Omega):=\left\{ u\in L^2(\mathbb{R}^N)\::\: \iint_{\substack{x,y\in\mathbb{R}^N\\|x-y|\leq 1}} \frac{|u(x)-u(y)|^2}{|x-y|^N}\ dx\,dy<\infty\text{ and }u=0\text{ in }\mathbb{R}^N\backslash \Omega \right\} \end{align} with the scalar product \begin{align}\label{cN:def} \mathcal{E}(u,v):=\frac{c_N}{2}\iint_{\substack{x,y\in\mathbb{R}^N\\|x-y|\leq 1}} \frac{(u(x)-u(y))(v(x)-v(y))}{|x-y|^N}\ dx\,dy,\qquad c_N:=\pi^{-\frac{N}{2}}\Gamma\left(\tfrac{N}{2}\right)>0, \end{align} and the norm $\|u\|:=\left(\mathcal{E}(u,u)\right)^\frac{1}{2}.$ The space of smooth functions with compact support in $\Omega$, denoted $C^\infty_c(\Omega)$, is dense in $\mathbb{H}(\Omega)$, see \cite[Theorem 3.1]{CW19}. The operator $L_\Delta$ has the following associated quadratic form \begin{align}\label{cE} \mathcal{E}_L(u,v):=\mathcal{E}(u,v)-c_N\iint_{\substack{x,y\in\mathbb{R}^N\\|x-y|\geq 1}}\frac{u(x)v(y)}{|x-y|^N}\ dxdy+\rho_N\int_{\mathbb{R}^N} uv\ dx, \end{align} where \begin{align}\label{rho:def} \rho_N:=2\ln 2+\psi(\tfrac{N}{2})+\gamma,\qquad \gamma:=-\Gamma'(1). \end{align} Here the constant $\gamma$ is known as the Euler-Mascheroni constant and $\psi:=\frac{\Gamma'}{\Gamma}$ is the digamma function. By \cite{CW19} (see also \cite{FJW20}), it holds that \begin{align}\label{fou} \mathcal{E}_L(u,u)=\int_{\mathbb{R}^N}2\ln|\xi||\widehat u(\xi)|^2\ d\xi\qquad \text{ for all }u\in C^\infty_c(\Omega), \end{align} where $\widehat u$ denotes the Fourier transform of $u$ given by \begin{align*} \widehat u(\xi)=\frac{1}{(2\pi)^\frac{N}{2}}\int_{\mathbb{R}^N}e^{-ix\cdot \xi}u(x)\ dx,\qquad \xi\in\mathbb{R}^N. \end{align*} We say that $u\in\mathbb H(\Omega)$ is a (weak) solution of~\eqref{lambda:problem} if \begin{align*} \mathcal{E}_L(u,v)=\mu\int_\Omega uv\ln|u|\ dx\qquad \text{ for all }v\in \mathbb{H}(\Omega). \end{align*} The \emph{energy functional} associated to~\eqref{lambda:problem} is given by \begin{align}\label{J0def} J_0:\mathbb{H}(\Omega)\to\mathbb{R},\qquad J_0(u):=\frac{1}{2}\mathcal{E}_L(u,u)-I(u),\qquad I(u):=\frac{\mu}{4}\int_\Omega u^2(\ln(u^2)-1)\ dx. \end{align} All nontrivial solutions of~\eqref{subcritical:intro} belong to the set \begin{align}\label{N0:def} \mathcal{N}_0:=\left\{u\in \mathbb{H}(\Omega)\backslash\{0\}\::\: \mathcal{E}_L(u,u) = \mu\int_\Omega u^2\ln|u|\ dx\right\}. \end{align} A solution $u\in \mathcal{N}_0$ is a \emph{least-energy solution} of~\eqref{lambda:problem} if \begin{align}\label{e2} J_0(u) =\inf_{v\in \mathcal{N}_0}J_0(v)\quad \text{ or, equivalently,}\quad |u|_2^2=\inf_{v\in \mathcal{N}_0}|v|_2^2. \end{align} \begin{remark}\label{diff:rmk} Let us point out some important differences between the study of \eqref{subcritical:intro} and \eqref{lambda:problem}. \begin{enumerate} \item The continuity and differentiability of $J_s$ is a direct consequence of Sobolev embeddings and Hölder's inequality. This is not the case for $J_0$. Note that Hölder's inequality cannot be used to bound $I$ (the nonlinear part of $J_0$). Therefore, it is not immediate that $J_0$ is of class $C^1$ (or even defined) in $\mathbb{H}(\Omega)$. We show in the next section that $J_0$ is in fact $C^1$ using a different argument. \item Note that $I_s$ defined in~\eqref{Js:def} is weakly continuous by the compactness of the Sobolev embedding $\mathcal{H}^s_0(\Omega)\hookrightarrow L^{p_s}(\Omega)$. This plays an important role to show the existence of solutions (see Theorem~\ref{existence}). But the embedding~\eqref{comp:em} is not enough to guarantee that $I$ is weakly continuous and this is one of the main technical obstacles to show Theorem~\ref{main:thm:2}. \item In~\eqref{N0:def}, note that $\mathcal{E}_L(u,u)$ and $\int_\Omega u^2\ln|u|\ dx$ could be negative or zero even if $u\neq 0$, whereas in~\eqref{cNs:def} the fact that $\|u\|_s$ and $|u|_{p_s}$ are positive for $u\neq 0$ provides the natural projection \begin{align*} \left(\frac{\|u\|_s^2}{|u|_{p_s}^{p_s}}\right)^\frac{1}{p_s-2}u\in\mathcal{N}_s\qquad \text{ for all }u\in\mathcal{H}_0^s(\Omega)\backslash\{0\}. \end{align*} We show in Lemma \ref{lem:A1A2} that the projection to $\mathcal{N}_0$ has to be given in terms of an exponential function. \item In~\eqref{les} the least-energy solution is characterized in terms of the norm in $\mathcal{H}^s_0(\Omega)$, whereas in~\eqref{e2} we only have information about the norm in $L^2(\Omega)$. This implies that an energy bound yields a weaker control on the minimizing sequences for $J_0$; therefore, additional new arguments are needed to improve these estimates, see Proposition \ref{prop:bd}. \end{enumerate} \end{remark} \section{Auxiliary results}\label{aux:sec} \subsection{Asymptotics for the best Sobolev constant} \begin{lemma}\label{lem:lim} If $a\neq 0$, then \begin{align}\label{lim0} \lim_{s\to 0^+}\left( 1+sa+o(s) \right)^{\frac{1}{s}}=e^{a}=\lim_{s\to 0^+}\left( 1+sa \right)^{\frac{1}{s}}. \end{align} \end{lemma} \begin{proof} We claim that \begin{align}\label{lim} \lim_{s\to 0^+}\left( 1+s(a+o(1)) \right)^{\frac{1}{s}}-\left( 1+sa \right)^{\frac{1}{s}}=0. \end{align} For $s\in(0,1)$, let $f:\mathbb{R}\to\mathbb{R}$ be given by $f(x)=(1+sx)^\frac{1}{s}$, then $f'(x)=(1+sx)^{\frac{1}{2}-1}$ and \begin{align*} f(x+h)-f(x) &= h\int_0^1 (1+s (x+\tau h))^{\frac{1}{s}-1}\ d\tau = h\int_0^1 (s\tau h + sx+1)^{\frac{1}{s}-1}\ d\tau; \end{align*} but, since $|h|<1$, $\tau<1$, and $s\in(0,1)$, \begin{align*} \left|\int_0^1 (s\tau h + sx+1)^{\frac{1}{s}-1}\ d\tau\right| &\leq \frac{|s(1 +|x|)+1|^{\frac{1}{s}}}{|s(1 +|x|)+1|}=e^{1+|x|}+o(1)<C\quad \text{as $s\to 0^+$,} \end{align*} where $C>0$ is independent of $s\in(0,1)$. Therefore, for any $x\in \mathbb{R}$, $s\in(0,1)$, and $|h|<1$, $|f(x+h)-f(x)| \leq Ch.$ This implies~\eqref{lim} and therefore~\eqref{lim0} holds. \end{proof} The next theorem is the fractional Sobolev inequality, see for example \cite[Theorem~1.1]{CT04}. \begin{theorem}[Fractional Sobolev inequality]\label{thm:sobolev} Let $N\geq 1$, $s\in(0,\frac{N}{2})$, and $2^*_s:=\frac{2N}{N-2s}$. Then \begin{align*} |u|^2_{2^*_s}\leq \kappa_{N,s}\|u\|^2_{s}\qquad \text{ for all $u\in H^s(\mathbb{R}^N),$} \end{align*} where \begin{equation}\label{eq:best_constant} \kappa_{N,s}=2^{-2s}\pi^{-s}\frac{\Gamma(\frac{N-2s}{2})}{\Gamma(\frac{N+2s}{2})} \left(\frac{\Gamma(N)}{\Gamma(\frac{N}{2})}\right)^{\frac{2s}{N}}. \end{equation} \end{theorem} Observe that the best Sobolev constant $\kappa_{N,s}$ is well behaved as $s\to 0^+$, in fact, $\lim_{s\to 0^+}\kappa_{N,s}=1$ and \begin{align} \lim_{s\to 0^+}\kappa_{N,s}^{\frac{1}{s}} &=\frac{1}{4\pi}\left(\frac{\Gamma(N)}{\Gamma(\frac{N}{2})}\right)^{\frac{2}{N}} \lim_{s\to 0^+}\left(1+s\partial_s\frac{\Gamma(\frac{N-2s}{2})}{\Gamma(\frac{N+2s}{2})}\Big|_{s=0}+o(s) \right)^\frac{1}{s} &=\frac{1}{4\pi}\left(\frac{\Gamma(N)}{\Gamma(\frac{N}{2})}\right)^{\frac{2}{N}}e^{-2\psi(\frac{N}{2})},\label{kappa:eq} \end{align} where we used Lemma~\ref{lem:lim}, Taylor's expansion, and the fact that \begin{align*} \partial_s\frac{\Gamma(\frac{N-2s}{2})}{\Gamma(\frac{N+2s}{2})}\Bigg|_{s=0} =-\frac{\Gamma(\frac{N-2s}{2})( \psi(\frac{N-2s}{2})+\psi(\frac{N+2s}{2}) )}{\Gamma(\frac{N+2s}{2})}\Bigg|_{s=0}=-2\psi(\tfrac{N}{2}), \end{align*} where $\psi=\frac{\Gamma'}{\Gamma}$ is the digamma function. \subsection{Bounds and expansions for \texorpdfstring{$\mathcal{E}_s$}{Es} and \texorpdfstring{$\mathcal{E}_L$}{EL}} \begin{lemma}\label{ln:bd} Let $\alpha$ and $\beta$ be such that $\beta>\alpha$, then \begin{align*} \ln(r^2)r^\alpha\leq \frac{2}{\beta-\alpha} r^\beta\qquad \text{ for all }r>1 \end{align*} \end{lemma} \begin{proof} We claim that $f(r)=\frac{2}{\beta-\alpha}r^{\beta-\alpha}-\ln(r^2)>0$ in $[1,\infty)$. Indeed, this follows since $f'(r) = 2r^{\beta-\alpha-1}-\frac{2}{r}=0$ if and only if $r=1$, which is the unique minimum of $f$ and $f(1)\geq 0$. \end{proof} \begin{lemma}\label{lem:5.5} Let $\Omega\subset \mathbb{R}^N$ be an open bounded set and let $u,v\in L^2(\Omega)$ be such that $v=u=0$ in $\mathbb{R}^N\backslash \Omega$, then \begin{align*} \iint_{\substack{x,y\in\mathbb{R}^N\\|x-y|\geq 1}}\frac{|u(x)v(y)|}{|x-y|^N}\ dxdy &\leq |u|_1|v|_1 \leq |\Omega| |u|_2|v|_2,\\ \iint_{\substack{x,y\in\mathbb{R}^N\\|x-y|\geq 1}}\frac{|u(x)u(y)|}{|x-y|^N}\ dxdy &\leq |u|_1^2 \leq |\Omega| |u|_2^2. \end{align*} \end{lemma} \begin{proof} Note that \begin{align*} \int_{\mathbb{R}^N}\int_{\mathbb{R}^N\backslash B_1(y)}\frac{|u(x)v(y)|}{|x-y|^N}\ dx\,dy &=\int_{\mathbb{R}^N}|v(y)|\int_{\mathbb{R}^N\backslash B_1(0)}\frac{|u(x+y)|}{|x|^N}\ dx\,dy\\ &\leq\int_{\mathbb{R}^N}|v(y)|\int_{\mathbb{R}^N}|u(x+y)|\ dx\,dy=|u|_1|v|_1. \end{align*} The result now follows from Hölder's inequality. \end{proof} \begin{lemma}\label{lem:cEbd} Let $u\in\mathcal{H}_0^s(\Omega)$ for some $s\in(0,1)$. Then $u\in \mathbb{H}(\Omega)$ and there are $C_1=C_1(N)>0$ and $C_2=C_2(\Omega)>0$ such that \begin{align*} |\mathcal{E}_L(u,u)|<C_1 |u|_1^2+\frac{1}{s}\|u\|^2_s\qquad\text{ and }\qquad \|u\|^2<C_2 |u|_2^2+\frac{1}{s}\|u\|^2_s. \end{align*} \end{lemma} \begin{proof} Let $u\in C^\infty_c(\Omega)$ and $C=\int_{B_1(0)} |\ln(|\xi|^2)|\ d\xi.$ Then, by~\eqref{fou} and Lemma~\ref{ln:bd} (with $\alpha=0$ and $\beta=2s$), \begin{align*} |\mathcal{E}_L(u,u)| &\leq \int_{\mathbb{R}^N} |\ln(|\xi|^2)||\widehat u(\xi)|^2\ d\xi\\ &\leq |\widehat u|_\infty^2\int_{B_1(0)} |\ln(|\xi|^2)|\ d\xi +\int_{\mathbb{R}^N\backslash B_1(0)} \ln(|\xi|^2)|\widehat u(\xi)|^2\ d\xi\\ &\leq C |u|_1^2+s^{-1}\mathcal{E}_s(u,u)=C |u|_1^2+s^{-1}\|u\|^2_s. \end{align*} Note that, by Lemma~\ref{lem:5.5}, there is $C'=C'(\Omega)>0$ such that $|\mathcal{E}_L(u,u)|\geq\|u\|^2-C'|u|_2^2$ and therefore \begin{align*} \|u\|^2\leq (C'+C|\Omega|)|u|_2^2+s^{-1}\|u\|^2_s. \end{align*} The result for $u\in \mathcal{H}^s_0(\Omega)$ follows by a standard density argument. \end{proof} \begin{lemma}\label{lem:sig:u} Let $s\in(0,\frac{1}{4})$, and let $u_s\in \mathcal{H}^s_0(\Omega)$ satisfy that \begin{align}\label{bd:hyp} \|u_s\|_s<C. \end{align} Then $u_s\in\mathcal{H}^\sigma_0(\Omega)$ for $\sigma\in(0,s)$ and there is $C'=C'(C,\Omega)>0$ such that \begin{align}\label{sigma:bd} |u_s|_2^2+\|u_s\|^2_\sigma<C'\qquad \text{ for all $\sigma\in(0,s)$.} \end{align} \end{lemma} \begin{proof} Note that \begin{align*} \|u_s\|^2_\sigma &=\int_{\mathbb{R}^N}|\xi|^{2\sigma}|\widehat u_s(\xi)|^{2}\ d\xi =\int_{\{|\xi|\leq 1\}}|\xi|^{2\sigma}|\widehat u_s(\xi)|^{2}\ d\xi +\int_{\{|\xi|>1\}}|\xi|^{2\sigma}|\widehat u_s(\xi)|^{2}\ d\xi\\ &\leq\int_{\mathbb{R}^N}|\widehat u_s(\xi)|^{2}\ d\xi +\int_{\mathbb{R}^N}|\xi|^{2s}|\widehat u_s(\xi)|^{2}\ d\xi=|u_s|_2^2+\|u_s\|_s^2<C^2+|u_s|_2^2. \end{align*} The claim now follows by Hölder's inequality, because \begin{align*} |u_s|_2^2\leq |\Omega|^\frac{p_s-2}{p_s} |u_s|_{p_s}^{2} =|\Omega|^\frac{p_s-2}{p_s} \|u_s\|_{s}^{\frac{4}{p_s}} \leq \max_{t\in[0,\frac{1}{4})}|\Omega|^\frac{p_t-2}{p_t} C^{\frac{4}{p_t}}. \end{align*} \end{proof} \begin{lemma}\label{Es:exp:lem} Let $0<\sigma<s<1$ and $u\in \mathcal{H}_0^s(\Omega)$. There is a constant $d_N>0$ depending only on $N$ such that \begin{align*} \left| \mathcal{E}_\sigma(u,u)-|u|_2^2-\sigma\mathcal{E}_L(u,u) \right| \leq d_N \frac{\sigma^2}{(s-\sigma)^2}\left( |u|_1^2+\mathcal{E}_s(u,u) \right). \end{align*} In particular, $\|u\|_\sigma\to |u|_2$ as $\sigma\to 0^+$. \end{lemma} \begin{proof} We argue as in \cite[Lemma 2.6]{FJW20}. For $\xi\in\mathbb{R}^N$, let $h(\sigma):=|\xi|^{2\sigma}$. Then $h'(\sigma)=|\xi|^{2\sigma}\ln(|\xi|^2)$, $h''(\sigma)=|\xi|^{2\sigma}(\ln(|\xi|^2))^2$, and \begin{align*} \left||\xi|^{2\sigma}-1-\sigma\ln(|\xi|^2)\right| &=|h(\sigma)-h(0)-\sigma h'(0)| =\left|\int_0^t h''(\tau)(\sigma-\tau)\ d\tau\right|\\ &\leq |\ln(|\xi|^2)|^2\int_0^\sigma |\xi|^{2\tau}|\sigma-\tau|\ d\tau =|\ln(|\xi|^2)|^2 \sigma^2 \int_0^1 |\xi|^{2\tau \sigma}|1-\tau|\ d\tau\\ &\leq \sigma^2|\ln(|\xi|^2)|^2 (\chi_{\{|\xi|\leq 1\}}+|\xi|^{2\sigma}\chi_{\{|\xi|> 1\}}), \end{align*} since $|\xi|^{2\tau \sigma}|1-\tau|\leq 1$ if $|\xi|\leq 1$ and $|\xi|^{2\tau \sigma}|1-\tau|<|\xi|^{2\sigma}$ if $|\xi|>1$ for $\tau\in(0,1)$. Note also that, for $|\xi|>1$ and $0<\sigma<s<1$, we have that $2\ln|\xi|<\frac{2}{s-\sigma}|\xi|^{s-\sigma}$, by Lemma~\ref{ln:bd}, and then \begin{align}\label{xi} |\ln(|\xi|^2)|^2|\xi|^{2\sigma}<\frac{4}{(s-\sigma)^2}|\xi|^{2s}\quad \text{ for }|\xi|>1. \end{align} Let $u\in C^\infty_c(\Omega)$ and $d_N:=4+(2\pi)^{-N}\int_{B_1(0)}|\ln(|\xi|^2)|^2\ d\xi$. Then, by~\eqref{xi}, \begin{align*} \Big| \mathcal{E}_\sigma(u,u)-|u|_2^2&-\sigma\mathcal{E}_L(u,u) \Big| =\left|\int_{\mathbb{R}^N}(|\xi|^{2\sigma}-1-\sigma\ln(|\xi|^2)|\widehat u(\xi)|^2\ d\xi \right|\\ &\leq \sigma^2 \int_{\mathbb{R}^N} |\ln(|\xi|^2)|^2 (\chi_{\{|\xi|\leq 1\}}+|\xi|^{2\sigma}\chi_{\{|\xi|> 1\}})|\widehat u(\xi)|^2\ d\xi\\ &\leq \sigma^2 \|\widehat u\|^2_\infty \int_{B_1(0)}|\ln(|\xi|^2)|^2\ d\xi + \sigma^2 \int_{\mathbb{R}^N\backslash B_1(0)}|\ln(|\xi|^2)|^2|\xi|^{2\sigma}|\widehat u(\xi)|^2\ d\xi\\ &\leq \sigma^2 (2\pi)^{-N}|u|_1^2\int_{B_1(0)}|\ln(|\xi|^2)|^2\ d\xi + \sigma^2 \frac{4}{(s-\sigma)^2} \int_{\mathbb{R}^N\backslash B_1(0)}|\xi|^{2s}|\widehat u(\xi)|^2\ d\xi\\ &\leq d_N \frac{\sigma^2}{(s-\sigma)^2}\left(|u|^2_1 + \mathcal{E}_s(u,u)\right). \end{align*} The result for $u\in \mathcal{H}^s_0(\Omega)$ follows by density. \end{proof} \subsection{Sharp logarithmic Sobolev inequality} Recall that \begin{align*} |u|_2:=\left(\int_{\mathbb{R}^N}|u|^2\right)^\frac{1}{2}\qquad \text{ for }u\in L^2(\Omega). \end{align*} Now we present a sharp logarithmic Sobolev inequality, which is one of our main tools to guarantee the compactness of the minimizing sequences of $J_0$ (see \eqref{J0def}). This result is shown in \cite[Theorem 3]{B95} for functions in the Schwarz space and using a different definition of the Fourier Transform (which influences the expression of the optimal constants). For completeness, we include here a slightly different proof and a statement which is more adequate for our purposes. \begin{proposition}\label{log:prop} For every $u\in \mathbb{H}(\Omega)$, \begin{align}\label{eq:ineq_becker} \frac{2}{N}\int_\Omega \ln(u^{2})u^2\ dx \leq\mathcal{E}_L(u,u)+\frac{2}{N}\ln(|u|_2^2)|u|_2^2+a_N|u|^2_2, \end{align} where $\psi$ is the digamma function and \begin{align}\label{A} a_N:=\frac{2}{N} \ln \left(\frac{\Gamma (N)}{\Gamma \left(\frac{N}{2}\right)}\right)-\ln (4 \pi )-2 \psi\left(\frac{N}{2}\right). \end{align} \end{proposition} \begin{proof} Let $u\in C^\infty_c(\mathbb{R}^N)$ and note that $|u|^{2^*_s}_{2^*_s} =|u|_2^2+s\frac{4}{N}\int_{\mathbb{R}^N} |u|^2\ln|u|\, dx+o(s)$ as $s\to 0^+$. Moreover, for $F\in C^1([0,1])$, \begin{align*} \partial_sF(s)^\frac{2}{2^*_s}=F(s)^{\frac{N-2 s}{N}} \left(\frac{(N-2 s) F'(s)}{N F(s)}-\frac{2 \ln (F(s))}{N}\right), \end{align*} and therefore, using $F(s)=|u|_{2^*_s}^{2^*_s}$, we obtain that $F(0)=|u|_2^2$ and \begin{align}\label{exp1} |u|_{2^*_s}^2 &=|u|_2^2+s\partial_t F(t)^\frac{2}{2^*_t}\Big|_{t=0}+o(s) =|u|_2^2+s\frac{4}{N}\left( \int_{\mathbb{R}^N} |u|^2\ln|u|\ dx-|u|_2^2 \ln |u|_2 \right) +o(s) \end{align} as $s\to 0^+$. Let $\kappa_{N,s}$ be given by~\eqref{eq:best_constant} and $a_N$ by~\eqref{A}, then $\kappa_{N,s}=1+s a_N+o(s)$ as $s\to 0^+$. Furthermore, by Lemma~\ref{Es:exp:lem}, $\|u\|_s^2=|u|^2_2+s\mathcal{E}_L(u,u)+o(s)$ as $s\to 0^+$ and therefore \begin{align}\label{exp2} \kappa_{N,s}\|u\|^{2}_s =|u|_2^2+s(a_N|u|_2^2+\mathcal{E}_L(u,u)) +o(s)\quad \text{ as }s\to 0^+. \end{align} By Theorem~\ref{thm:sobolev}, $|u|^2_{2^*_s}\leq\kappa_{N,s}\|u\|^{2}_s$, and, using~\eqref{exp1} and~\eqref{exp2}, \begin{align*} \frac{4}{N}\left( \int_{\mathbb{R}^N} |u|^2\ln|u|\ dx-|u|_2^2 \ln |u|_2 \right) \leq a_N|u|_2^2+\mathcal{E}_L(u,u). \end{align*} This yields the claim for $u\in C^\infty_c(\mathbb{R}^N)$ and the general statement for $u\in \mathbb{H}(\Omega)$ follows by density. Indeed, let $u\in\mathbb{H}(\Omega)$ and let $(u_n)_{n\in\mathbb{N}}\subset C^\infty_c(\Omega)$ be such that $u_n\to u$ in $\mathbb{H}(\Omega)$ ($C^\infty_c(\Omega)$ is dense in $\mathbb{H}(\Omega)$ by \cite[Theorem 3.1]{CW19}). By~\eqref{comp:em}, $u_n\to u$ in $L^2(\Omega)$, and therefore \begin{align}\label{Fa0} \frac{4}{N}\left( \int_{\mathbb{R}^N} |u_n|^2\ln|u_n| \ dx-|u|_2^2 \ln |u|_2 \right) \leq a_N|u|_2^2+\mathcal{E}_L(u,u)+o(1)\quad \text{as $n\to\infty$}. \end{align} It suffices to show that, up to a subsequence, \begin{align}\label{Fa} \int_{\mathbb{R}^N} |u|^2\ln|u| \ dx\leq \lim_{n\to \infty}\int_{\mathbb{R}^N}|u_n|^2 \ln|u_n|\ dx\qquad \text{ for all }n\in\mathbb{N}. \end{align} Using that $\Omega$ is bounded and the dominated convergence theorem, we have that, passing to a subsequence, \begin{align}\label{Fa1} \lim_{n\to\infty}\int_{\{|u_n|\leq 1\}} |u_n|^2\ln|u_n|\ dx =\int_{\{|u|\leq 1\}} |u|^2\ln|u|\ dx, \end{align} whereas, by Fatou's Lemma, we deduce that, passing to a subsequence, \begin{align}\label{Fa2} \int_{\{|u|\geq 1\}} |u|^2\ln|u|\ dx\leq \lim_{n\to\infty}\int_{\{|u_n|\geq 1\}} |u_n|^2\ln|u_n|\ dx. \end{align} Then~\eqref{Fa1} and~\eqref{Fa2} imply~\eqref{Fa}, and the logarithmic Sobolev inequality for general $u\in\mathbb{H}(\Omega)$ follows from~\eqref{Fa} and~\eqref{Fa0}. \end{proof} \subsection{Differentiability of the energy functional} Let $J_0$ and $I$ be given by~\eqref{J0def}. We show that $J_0$ is of class $C^1$ in $\mathbb{H}(\Omega)$. Recall that \begin{align}\label{cEuu} \mathcal{E}_L(u,u)=\|u\|^2-c_N\iint_{\substack{x,y\in\mathbb{R}^N\\|x-y|\geq 1}}\frac{u(x)u(y)}{|x-y|^N}\ dxdy+\rho_N|u|_2^2. \end{align} It is clear that $\|\cdot\|$ is differentiable and $|\cdot|_2$ is also differentiable by~\eqref{comp:em}. Moreover, \begin{align*} B(u,v):=\iint_{\substack{x,y\in\mathbb{R}^N\\|x-y|\geq 1}}\frac{u(x)v(y)}{|x-y|^N}\ dxdy \end{align*} is a bounded bilinear form (by Lemma~\ref{lem:5.5}). We show next the differentiability of $I$. \begin{lemma}\label{I:C1} Let $I$ be given by~\eqref{J0def}. Then $I$ is of class $C^1$ in $\mathbb{H}(\Omega)$ and $I'(u)v=\mu\int_\Omega uv\ln|u|$. In particular, $I'(u)\in\mathcal{L}(\mathbb{H}(\Omega),\mathbb{R})$ and $I':\mathbb{H}(\Omega)\to\mathcal{L}(\mathbb{H}(\Omega),\mathbb{R})$ is continuous. \end{lemma} \begin{proof} Let $x\in \Omega$, $u,v\in\mathbb{H}(\Omega)$, and $\delta\in(-1,1)\backslash\{ 0\}$. Let $h(t):=t^2(\ln(t^2)-1)$, $h'(t)=2t\ln|t|$. By the mean value theorem, there is $\tau\in[0,1]$ such that \begin{align*} Q_\delta(x):=\frac{h(u(x)+\delta v(x))-h(u(x))}{\delta} &=2(u(x)+\delta\tau v(x))v(x)\ln|u(x)+\tau \delta v(x)|. \end{align*} Assume first that $|u(x)+\tau \delta v(x)|\geq 1$, then \begin{align*} 1\leq |u(x)+\delta\tau v(x)|\leq |u(x)|+|v(x)|\leq 2\max\{|u(x)|,|v(x)|\} \end{align*} and, since $h'$ is monotone increasing in $(1,\infty)$, \begin{align*} |Q_\delta(x)| &\leq 4\max\{|u(x)|,|v(x)|\}|v(x)||\ln(2\max\{|u(x)|,|v(x)|\})|\\ &\leq 4\max\{|u(x)|,|v(x)|\}^2|\ln(2\max\{|u(x)|,|v(x)|\})|\\ &\leq |2u(x)|^2|\ln|2u(x)|| +|2v(x)|^2|\ln|2v(x)||=:M(x). \end{align*} Note that $M\in L^1(\Omega)$. Indeed, by~\eqref{comp:em}, Lemma~\ref{lem:5.5}, and Proposition~\ref{log:prop}, \begin{align*} \int_\Omega w^2|\ln |w|| &=\int_{\{|w|\geq 1\}} w^2\ln |w|-\int_{\{|w|<1\}} w^2\ln |w|=\int_{\Omega} w^2\ln |w|-2\int_{\{|w|<1\}} w^2\ln |w|\\ &\leq \frac{N}{2}\mathcal{E}_L(w,w)+\ln(|w|_2^2)|w|_2^2+\frac{N}{2}|a_N||w|_2^2+2|\Omega|\sup_{t\in(0,1)}t^2|\ln t| <\infty \end{align*} for all $w\in\mathbb{H}(\Omega)$. On the other hand, if $|u(x)+\tau \delta v(x)|<1$, then $|Q_\delta(x)|\leq 2e^{-1}$, because $\sup_{(0,1)}|h'|\leq 2e^{-1}$. Then $|Q_\delta|\leq M+2e^{-1}$ in $\Omega$ and, by the dominated convergence theorem, \begin{align*} I'(u)v=\lim_{\delta\to 0}\frac{I(u+\delta v)-I(u)}{\delta} =\mu\int_\Omega uv\ln|u|\ dx. \end{align*} Using similar arguments, one can show that, for $u,v,w\in\mathbb{H}(\Omega)$, \begin{align*} \lim_{\delta\to 0}|I'(u)(v+\delta w)-I'(u)v| &\leq \frac{\mu}{2}\lim_{\delta\to 0}\delta \int_\Omega |u w\ln(u^2)|=0,\\ \lim_{\delta\to 0}|I'(u+\delta w)v-I'(u)v| &\leq \frac{\mu}{2} \lim_{\delta\to 0}\int_\Omega |(u+\delta w)\ln((u+\delta w)^2)-u\ln(u^2)||v|=0, \end{align*} and therefore $I'(u)\in\mathcal{L}(\mathbb{H}(\Omega),\mathbb{R})$ and $I':\mathbb{H}(\Omega)\to\mathcal{L}(\mathbb{H}(\Omega),\mathbb{R})$ is continuous. \end{proof} \section{Uniform bounds for the elements in the Nehari manifold}\label{U:sec} In this section we show some uniform estimates for every $u\in\mathcal{N}_s$ with $s\in(0,\frac{1}{4})$. In particular, we show in Proposition~\ref{cor:Ebd} that all least-energy solutions of~\eqref{subcritical:intro} are uniformly bounded in $\mathbb{H}(\Omega)$. \medskip We begin with some auxiliary lemmas. \begin{lemma}\label{lem:Ebd} Let $p\in C^1([0,\tfrac{1}{4}])$ satisfy~\eqref{p:hyp}. There is a constant $c=c(p,\Omega)>0$ such that \begin{align}\label{bds} \|u\|_s>c\quad \text{for all $u\in\mathcal{N}_s$ and all $s\in(0,\tfrac{1}{4}).$} \end{align} \end{lemma} \begin{proof} Let $F_s:\mathcal{H}^s_0(\Omega)\backslash\{0\}\to \mathbb{R}$ be given by $F_s(u)=\|u\|_s^2 - |u|_{p_s}^{p_s}.$ By Hölder's inequality and Theorem~\ref{thm:sobolev}, \begin{align*} F_s(u)\geq \|u\|_s^2 - |\Omega|^\frac{2^*_s-p_s}{2^*_s}|u|_{2^*_s}^{p_s} \geq \|u\|_s^2 - |\Omega|^\frac{2^*_s-p_s}{2^*_s}\kappa_{N,s}^{\frac{p_s}{2}}\|u\|_{s}^{p_s} =\|u\|_s^2(1 - |\Omega|^\frac{2^*_s-p_s}{2^*_s}\kappa_{N,s}^{\frac{p_s}{2}}\|u\|_{s}^{p_s-2}). \end{align*} Let $g(t,s) :=1 - |\Omega|^\frac{2^*_s-p_s}{2^*_s}\kappa_{N,s}^{\frac{p_s}{2}}t^{p_s-2},$ where $\kappa_{N,s}$ is given in~\eqref{eq:best_constant}. Then \begin{align*} g(t,s)>0\quad \text{ if }\quad t<|\Omega|^\frac{2^*_s-p_s}{2^*_s(2-p_s)} \kappa_{N,s}^{\frac{p_s}{{2(2-p_s)}}}. \end{align*} Note that \begin{align*} \frac{2^*_s-p_s}{2^*_s(2-p_s)} =\frac{2^*_s-2}{2^*_s(2-p_s)}+\frac{1}{2^*_s} =-\frac{4}{(N-2s)2^*_s \int_0^1 p'(\tau s)\,d\tau}+\frac{1}{2^*_s} \to \frac{1}{2}-\frac{2}{Np'(0)}\quad \text{as $s\to 0$,} \end{align*} and therefore \begin{align*} \lim_{s\to 0}|\Omega|^\frac{2^*_s-p_s}{2^*_s(2-p_s)}= |\Omega|^{\frac{1}{2}-\frac{2}{Np'(0)}}>0. \end{align*} Furthermore, by~\eqref{kappa:eq}, \begin{align*} \lim_{s\to 0}\kappa_{N,s}^{\frac{p_s}{{2(2-p_s)}}}=\lim_{s\to 0} \left( \kappa_{N,s}^\frac{1}{s} \right)^{\frac{s p_s}{{2(2-p_s)}}} = \left(\frac{1}{4\pi}\left(\frac{\Gamma(N)}{\Gamma(\frac{N}{2})}\right)^{\frac{2}{N}}e^{-2\psi(\frac{N}{2})}\right)^{-\frac{1}{p'(0)}} >0. \end{align*} As a consequence, there is $c=c(p,\Omega)>0$ such that $F_s(u)>0$ if $\|u\|_s\in(0,c)$, and then $\|u\|_s>c$ for all $u\in \mathcal{N}_s$ and for all $s\in(0,\frac{1}{4})$, as claimed. \end{proof} \begin{lemma}\label{lem:t0} Let $s\in(0,\tfrac{1}{4})$, $\varphi\in C^\infty_c(\Omega)\backslash\{0\}$, and $p_s:=p(s)$, where $p$ satisfies~\eqref{p:hyp}. Let $t_\varphi^s$ be given by \begin{align}\label{ts:def} t_\varphi^s :=\left( \frac{\|\varphi\|_s^2}{|\varphi|_{p_s}^{p_s}}\right)^{\frac{1}{p_s-2}}. \end{align} Then $t^s_\varphi \varphi\in\mathcal{N}_s$ and \begin{align*} \lim_{s\to 0^+} t_\varphi^s = t_\varphi^0 :=\exp\left(\frac{\mathcal{E}_L(\varphi,\varphi)-p'(0)\int_\Omega\ln (|\varphi|)|\varphi|^2\ dx}{p'(0)|\varphi|_2^2}\right)>0. \end{align*} In particular, $\sup_{s\in(0,\frac{1}{4})}t_\varphi^s<\infty$. \end{lemma} \begin{proof} Let $s\in(0,\tfrac{1}{4})$ and $\varphi\in C^\infty_c(\Omega)\backslash\{0\}$. Then, by Lemma~\ref{Es:exp:lem}, $\|\varphi\|_s^2=|\varphi|_2^2+s\mathcal{E}_L(\varphi,\varphi)+o(s) $ as $s\to 0^+.$ On the other hand, $|\varphi|_{p_s}^{p_s}=|\varphi|^2_2 + s p'(0)\int_\Omega |\varphi|^2\ln |\varphi|\, dx+o(s).$ Let $A=|\varphi|_2^2$. Then, by Lemma~\ref{lem:lim}, \begin{align*} \lim_{s\searrow 0}t_\varphi^s &= \lim_{s\searrow 0}\left( \frac{1 + A^{-1}s\mathcal{E}_L(\varphi,\varphi) + o(s)} {1+s\frac{p'(0)}{A} \int_\Omega |\varphi|^2\ln |\varphi|\, dx+o(s)} \right)^{\frac{1}{p_s-2}}\\ &= \left(\frac{\lim_{s\searrow 0}(1 + A^{-1}s\mathcal{E}_L(\varphi,\varphi) + o(s))^\frac{1}{s}} {\lim_{s\searrow 0}( {1+s\frac{p'(0)}{A} \int_\Omega |\varphi|^2\ln |\varphi|\, dx+o(s)})^\frac{1}{s}} \right)^\frac{1}{p'(0)} \\ &= \left(\frac{ e^{A^{-1}\mathcal{E}_L(\varphi,\varphi)} }{ e^{p'(0)A^{-1} \int_\Omega |\varphi|^2\ln |\varphi|}} \right)^\frac{1}{p'(0)} =\exp\left(\frac{\mathcal{E}_L(\varphi,\varphi)-p'(0)\int_\Omega\ln (|\varphi|)|\varphi|^2}{p'(0)|\varphi|_2^2}\right)>0. \end{align*} This implies that the function $s\mapsto t^s_\varphi$ has a continuous extension to the compact set $[0,\tfrac{1}{4}]$, and therefore $\sup_{s\in(0,\frac{1}{4})}t_\varphi^s<\infty$, as claimed. \end{proof} \begin{remark} In Lemma~\ref{lem:A1A2} below we show that, in fact, $t_\varphi^0\varphi\in \mathcal{N}_0$ with $\mu=p'(0)$. \end{remark} The following result provides an ``intermediate" logarithmic-type Sobolev inequality. \begin{lemma}\label{prebd} Let $s\in(0,\tfrac{1}{4})$ and let $\varphi\in C^\infty_c(\Omega)$. It holds that \begin{align*} \int_0^1\frac{4N}{(N-2s\tau)^2}\int_\Omega|\varphi|^{2^*_{s\tau}}\ln|\varphi|\ dx\,d\tau\leq \int_0^1 k'(s\tau)\|\varphi\|_{s\tau}^{2^*_{s\tau}} +2k(s\tau)\int_{\mathbb{R}^N}|\xi|^{2s\tau}\ln|\xi| |\widehat \varphi(\xi)|^2\ d\xi\,d\tau, \end{align*} where $k(s):=(\kappa_{N,s})^{\frac{2^*_s}{2}}$ and with $\kappa_{N,s}$ as in Theorem~\ref{thm:sobolev}. Moreover, if \begin{align}\label{C:def} \|\varphi\|_s^2<C\qquad \text{ for some $C>0$,} \end{align} then there is $C_1=C_1(C,\Omega)>0$ such that \begin{align*} \int_0^1\int_{\{|\varphi|\geq 1\}}|\varphi|^{2^*_{s\tau}}\ln|\varphi|\ dx\,d\tau\leq C_1+\frac{N}{4}\int_0^1 k(s\tau)\int_{\mathbb{R}^N}|\xi|^{2s\tau}\ln(|\xi|^2) |\widehat \varphi(\xi)|^2\ d\xi\,d\tau. \end{align*} \end{lemma} \begin{proof} Let $s\in(0,\frac{1}{4})$, $k(s):=(\kappa_{N,s})^{\frac{2^*_s}{2}}$, and $G(s):=k(s)\|\varphi\|_s^{2^*_s}-|\varphi|_{2^*_s}^{2^*_s}.$ Then $k\in C^1([0,\tfrac{1}{4}])$ and $G\in C^1((0,\tfrac{1}{4}))$ with \begin{align*} G'(s)=k'(s)\|\varphi\|_s^{2^*_s} +2k(s)\int_{\mathbb{R}^N}|\xi|^{2s}\ln|\xi| |\widehat \varphi(\xi)|^2\ d\xi -\frac{4N}{(N-2s)^2}\int_\Omega|\varphi|^{2^*_s}\ln|\varphi|\, dx. \end{align*} Note that $\lim_{s\to 0^+}G(s)=0$ and, by Theorem~\ref{thm:sobolev}, $G(s)\geq 0$ for $s\in(0,\tfrac{1}{4})$. Then $\int_0^1 G'(s\tau)\ d\tau\geq 0$, namely, \begin{align} J&:=\int_0^1\frac{4N}{(N-2s\tau)^2}\int_\Omega|\varphi|^{2^*_{s\tau}}\ln|\varphi|\ dx\,d\tau\notag\\ &\leq \int_0^1 k'(s\tau)\|\varphi\|_{s\tau}^{2^*_{s\tau}} +2k(s\tau)\int_{\mathbb{R}^N}|\xi|^{2s\tau}\ln|\xi| |\widehat \varphi(\xi)|^2\ d\xi\,d\tau.\label{it:0} \end{align} By Lemma~\ref{lem:sig:u} and \eqref{C:def}, there is $M=M(C,\Omega)>0$ such that $\|\varphi\|^2_{s\tau}<M$ for all $\tau\in(0,1)$; but then, there is $C'=C'(C,\Omega)>0$ such that \begin{align}\label{it:1} \int_0^1|k'(s\tau)|\|\varphi\|_{s\tau}^{2^*_{s\tau}}\, d\tau<C'. \end{align} Moreover, \begin{align} J&=\int_0^1\frac{4N}{(N-2s\tau)^2}\int_{\Omega\cap \{|\varphi|\geq 1\}}|\varphi|^{2^*_{s\tau}}\ln|\varphi|\ dx\,d\tau+\int_0^1\frac{4N}{(N-2s\tau)^2}\int_{\Omega\cap\{|\varphi|<1\}}|\varphi|^{2^*_{s\tau}}\ln|\varphi|\ dx\,d\tau\notag\\ &\geq \frac{4}{N} \int_0^1\int_{\Omega\cap \{|\varphi|\geq 1\}}|\varphi|^{2^*_{s\tau}}\ln|\varphi|\ dx\,d\tau-|\Omega|c_1,\label{it:2} \end{align} where $c_1:=\sup_{t\in(0,1),\theta\in(0,\frac{1}{4})}\frac{4N}{(N-2\theta)^2} t^{2^*_{\theta}}|\ln|t||$. By~\eqref{it:0},~\eqref{it:1}, and~\eqref{it:2}, \begin{align*} \int_0^1\int_{\Omega\cap \{|\varphi|\geq 1\}}|\varphi|^{2^*_{s\tau}}\ln|\varphi|\ dx\,d\tau\leq \frac{N}{4}\left( |\Omega|c_1+C'+\int_0^1 2k(s\tau)\int_{\mathbb{R}^N}|\xi|^{2s\tau}\ln|\xi| |\widehat \varphi(\xi)|^2\ d\xi\,d\tau \right), \end{align*} and the claim follows. \end{proof} The next result shows that uniform bounds in $\mathcal{N}_s$ yield uniform bounds in $\mathbb{H}(\Omega)$. Note that Lemma~\ref{lem:cEbd} provides such a bound only for $s$ far away from zero. For $s$ near zero a finer analysis is required. \begin{lemma}\label{lem:vip} Let $C_0>0$, $s\in(0,\frac{1}{4})$, $p_s=p(s)$ with $p$ as in~\eqref{p:hyp}, and let $\varphi\in \mathcal{N}_s$ be such that $\|\varphi\|_s^2<C_0$. Then there is $C=C(C_0,p,\Omega)>0$ such that \begin{align*} \|\varphi\|^2=\mathcal{E}(\varphi,\varphi)<C. \end{align*} \end{lemma} \begin{proof} Assume first that $\varphi\in C^\infty_c(\Omega)\cap \mathcal{N}_s$. For $\tau\in(0,1)$ and $\sigma\in(0,\frac{1}{4})$, let \begin{align*} h_\sigma(\tau):=1-\frac{N}{4}(\kappa_{N,\sigma\tau})^{\frac{2^*_{\sigma\tau}}{2}}\sup_{(0,\sigma)}|p'|, \end{align*} where $\kappa_{N,s}$ is given in \eqref{kappa:eq}. Note that $p'(0)\in(0,\frac{4}{N})$ (by \eqref{p:hyp}) and $\kappa_{N,\sigma}\to 1$ as $\sigma\to 0^+$ (by \eqref{kappa:eq}). Therefore there is $s_0\in(0,\frac{1}{4})$ such that, if $s\in(0,s_0)$, then \begin{align}\label{s0} p'(s\tau)>0\quad \text{ for }\tau\in(0,1)\qquad \text{ and }\qquad \delta:=\min_{(0,1)}h_s>0. \end{align} For $s\in[s_0,\frac{1}{4})$, the claim follows from Lemmas \ref{lem:cEbd} and \ref{lem:sig:u}. To show the claim in $(0,s_0)$, let $s\in(0,s_0)$. Note that \begin{align} \mathcal{I}:=\frac{\|\varphi\|_s^2-|\varphi|_2^2}{s} =\int_0^1\int_{\mathbb{R}^N}|\xi|^{2s\tau}\ln(|\xi|^2)|\widehat \varphi(\xi)|^2\ d\xi \ d\tau.\label{Iphi} \end{align} On the other hand, using that $\varphi\in\mathcal{N}_s$ and \eqref{s0}, \begin{align*} \mathcal{I}&=\frac{|\varphi|_{p_s}^{p_s}-|\varphi|_2^2}{s} =\int_0^1 p'(s\tau) \int_\Omega |\varphi|^{p(s\tau)} \ln |\varphi|\ dx \ d\tau\\ &=\int_0^1 p'(s\tau) \int_{\{|\varphi|<1\}} |\varphi|^{p(s\tau)} \ln |\varphi|\ dx \ d\tau +\int_0^1 p'(s\tau) \int_{\{|\varphi|\geq 1\}} |\varphi|^{p(s\tau)} \ln |\varphi|\ dx \ d\tau\\ &\leq \sup_{(0,s)}|p'|\int_0^1 \int_{\{|\varphi|\geq 1\}} |\varphi|^{2^*_{s\tau}} \ln |\varphi|\ dx \ d\tau \end{align*} By Lemma~\ref{prebd}, there is $C_1=C_1(C_0,\Omega)>0$ such that \begin{align} \mathcal{I}\leq C_1+\sup_{(0,s)}|p'|\frac{N}{4}\int_0^1 (\kappa_{N,s\tau})^{\frac{2^*_{s\tau}}{2}}\int_{\mathbb{R}^N}|\xi|^{2s\tau}\ln(|\xi|^2) |\widehat \varphi(\xi)|^2\ d\xi\,d\tau.\label{Iphi2} \end{align} Furthermore, by~\eqref{Iphi} and~\eqref{Iphi2}, \begin{align}\label{it:3} \int_0^1\int_{\mathbb{R}^N}h_s(\tau)|\xi|^{2s\tau}\ln|\xi|^2 |\widehat \varphi(\xi)|^2\ d\xi\,d\tau\leq C_1. \end{align} Note that, if $C_2:=\int_{\{|\xi|\leq 1\}}|\ln|\xi|^2| \ d\xi$, \begin{align} \int_0^1\int_{\{|\xi|\leq 1\}}h_s(\tau)|\xi|^{2s\tau}\ln|\xi|^2 &|\widehat \varphi(\xi)|^2\ d\xi\,d\tau\notag\\ &=-\int_0^1\int_{\{|\xi|\leq 1\}}h_s(\tau)|\xi|^{2s\tau}|\ln|\xi|^2| |\widehat \varphi(\xi)|^2\ d\xi\,d\tau\notag\\ &\geq -\int_{\{|\xi|\leq 1\}}|\ln|\xi|^2| |\widehat \varphi(\xi)|^2\ d\xi\notag\\ &\geq-\delta\int_{\{|\xi|\leq 1\}}|\ln|\xi|^2| |\widehat \varphi(\xi)|^2\ d\xi -(1-\delta)|\widehat \varphi|_\infty^2\int_{\{|\xi|\leq 1\}}|\ln|\xi|^2| \ d\xi\notag\\ &\geq\delta\int_{\{|\xi|\leq 1\}}\ln|\xi|^2 |\widehat \varphi(\xi)|^2\ d\xi -C_2(1-\delta)|\varphi|_1^2 \label{it:4} \end{align} and, by \eqref{s0}, \begin{align}\label{it:5} \int_0^1\int_{\{|\xi|>1\}}h_s(\tau)|\xi|^{2s\tau}\ln|\xi|^2 |\widehat \varphi(\xi)|^2\ d\xi\,d\tau \geq \delta \int_{\{|\xi|>1\}}\ln|\xi|^2 |\widehat \varphi(\xi)|^2\ d\xi. \end{align} But then, by~\eqref{it:3},~\eqref{it:4}, and~\eqref{it:5}, \begin{align}\label{it:7} \mathcal{E}_L(\varphi,\varphi)=\int_{\mathbb{R}^N}\ln|\xi|^2 |\widehat \varphi(\xi)|^2\ d\xi\leq \frac{C_1}{\delta} +C_2\frac{1-\delta}{\delta}|\Omega||\varphi|_2^2. \end{align} By Lemma \ref{lem:sig:u}, there is $C_3=C_3(\Omega,C_0)>0$ such that \begin{align}\label{it:8} |\varphi|_2^2<C_3. \end{align} Moreover, by~\eqref{cEuu} and Lemma~\ref{lem:5.5}, \begin{align}\label{it:9} \mathcal{E}_L(\varphi,\varphi)\geq \|\varphi\|^2-C_4|\varphi|_2^2 \end{align} for some $C_4=C_4(\Omega)>0$. The claim now follows, for $\varphi\in C^\infty_c(\Omega)\cap \mathcal{N}_s$, from~\eqref{it:7},~\eqref{it:8}, and~\eqref{it:9}. The general case $\varphi\in\mathcal{N}_s$ follows from the density of $C^\infty_c(\Omega)$ in $\mathcal{H}^s_0(\Omega)$ and Lemma~\ref{lem:cEbd}. \end{proof} We are ready to show the main result in this section. \begin{proposition}\label{cor:Ebd} Let $s\in(0,\tfrac{1}{4})$, $p_s=p(s)$ with $p$ as in~\eqref{p:hyp}, and let $u_{s}\in\mathcal{N}_s$ be a least-energy solution of~\eqref{subcritical:intro}. There is $C=C(p,\Omega)>0$ such that \begin{align*} \|u_s\|^2=\mathcal{E}(u_{s},u_s)<C\qquad \text{ for all }s\in(0,\tfrac{1}{4}). \end{align*} \end{proposition} \begin{proof} Let $\varphi\in C^\infty_c(\Omega)\backslash \{0\}$. By~\eqref{les} and Lemma~\ref{lem:t0}, \begin{align}\label{C0} \|u_s\|_s^2=\inf_{v \in \mathcal{N}_s}\|v\|_s^2\leq (t^s_\varphi)^2\|\varphi\|_s^2\leq \sup_{s\in(0,\frac{1}{4})}(t^s_\varphi)^2\|\varphi\|_s^2=:C_0<\infty. \end{align} With this bound, the result follows from Lemma~\ref{lem:vip}. \end{proof} \section{Existence of a least-energy solution of the limiting problem}\label{Sec:MT2} In this section we show Theorem~\ref{main:thm:2}. In the following, we use that, by \cite[Theorem 1.4]{CW19}, \begin{align}\label{thm:eigen_ll} \lambda_1^L:=\min\{\mathcal{E}_{L}(u,u): u\in\mathbb H(\Omega), \ |u|_2=1\}\in\mathbb{R}. \end{align} Note that $\lambda_1^L$ can be zero o nonnegative. In fact, by \cite[Corollary 1.10]{CW19}, $\lambda_1^L\leq \ln(\lambda_1)$, where $\lambda_1$ is the first Dirichlet eigenvalue of $(-\Delta)$ in $\Omega$. \begin{lemma}\label{lem:bound_below_nehari} There are $c_0>0$ and $c_1>0$ such that $|u|_2\geq c_0$ and $\|u\|>c_1$ for every $u\in\mathcal N_0(\Omega)$. \end{lemma} \begin{proof} Let $\mu=\frac{4}{N}\lambda$ for some $\lambda\in(0,1)$. For $u\in\mathbb{H}(\Omega)$, let $ F_0(u):=\mathcal{E}_L(u,u)-\frac{2\lambda}{N}\int_{\Omega}u^2 \ln(u^2). $ Then, by~\eqref{eq:ineq_becker}, \begin{equation} F_0(u) \geq (1-\lambda) \mathcal{E}_L (u,u)-\left[2a_{N}+\frac{2}{N}\ln(|u|_2^2)\right]\lambda|u|_2^2 \qquad\textnormal{ for every } u\in\mathbb H(\Omega). \end{equation} Furthermore, using~\eqref{thm:eigen_ll}, \begin{equation} F_0(u) \geq \left[\frac{1-\lambda}{\lambda} \lambda_1^{L}-2a_{N}-\frac{2}{N}\ln(|u|_2^2)\right]\lambda|u|_2^2>0 \end{equation} if $|u|_2 < \exp\left({\frac{(1-\lambda)}{4\lambda}N\lambda_1^L-\frac{N}{2}a_{N}}\right)=: c_0$. Therefore $|u|_2\geq c_0$ for every $u\in\mathcal N_0(\Omega)$. Finally, by~\eqref{comp:em}, there is $C>0$ such that $c_0\leq|u|_2\leq C\|u\|$ for all $u\in\mathbb{H}(\Omega)$, and this ends the proof. \end{proof} \begin{lemma}\label{lem:A1A2} For $w\in \mathbb H\setminus\{0\}$, let \begin{align}\label{t0:def} t^0_w:=\exp\left(\frac{\mathcal{E}_L(w,w)-\mu\int_{\Omega}w^2\ln|w|}{\mu|w|_2^2}\right) \end{align} and let $\alpha_w(s):=J_0(sw)$. Then, $\alpha_w^\prime(s)>0$ for $0<s<t^0_w$ and $\alpha_w^\prime(s)<0$ for $s>t^0_w$. In particular, $s\mapsto J_0(sw)$ achieves its unique maximum at $s=t^0_w$ and $t^0_w w\in\mathcal{N}_0$. \end{lemma} \begin{proof} Note that \begin{align*} \alpha'_w(s) =\left(\mathcal{E}_L(w,w)-\mu\int_\Omega w^2\ln|sw|\right)s =\left(\mathcal{E}_L(w,w)-\mu|w|_2^2\ln|s|-\mu\int_\Omega w^2\ln|w|\right)s \end{align*} for $w\in\mathbb H(\Omega)$. The claim now follows by the definition of $s_w$ and a direct computation. \end{proof} \begin{proposition}\label{prop:bd} Let $(u_n)_{n\in\mathbb N}\subset \mathcal N_0$ be a sequence such that $\sup_{n\in\mathbb N} J_0(u_n)\leq C$ for some $C>0$. Then $(u_n)_{n\in\mathbb{N}}$ is bounded in $\mathbb H(\Omega)$ and, passing to a subsequence, there is $u\in\mathbb H(\Omega)\backslash\{0\}$ such that $u_n\rightharpoonup u$ weakly in $\mathbb H(\Omega)$ and $u_n\to u_0$ strongly in $L^2(\Omega)$ as $n\to\infty$. \end{proposition} \begin{proof} Note that $J_0(u_n)=\frac{\mu}{4}|u_n|_2^2$, and therefore $\sup_{n\in\mathbb{N}}|u_n|_2^2\leq \frac{4}{\mu}C=:C_1$. Moreover, by Proposition~\ref{log:prop}, \begin{align*} J_0(u_n)\geq \left(1-\frac{N\mu}{4}\right)\mathcal{E}_L(u_n,u_n)-\frac{\mu}{2}\ln(|u_n|_2^2)|u_n|_2^2-a_N\frac{N\mu}{4}|u_n|^2_2, \end{align*} which yields that $\sup_{n\in\mathbb{N}}\mathcal{E}_L(u_n,u_n) \leq C+\sup_{t\in[0,C_1]}\left(\frac{\mu}{2}|\ln(t)|t+|a_N|\frac{N\mu}{4}t\right)=:C_2 $. By Lemma~\ref{lem:5.5}, \begin{align*} C_2\geq \mathcal{E}_L(u_n,u_n)\geq \|u_n\|^2 - (|\Omega|+|\rho_N|)C_1, \end{align*} which implies that $\sup_{n\in \mathbb{N}}\|u_n\|<\infty$. Then, by~\eqref{comp:em}, passing to a subsequence, there is $u\in\mathbb{H}(\Omega)$ such that $u_n\rightharpoonup u$ in $\mathbb H(\Omega)$ and $u_n\to u$ in $L^2(\Omega)$. Finally, by Lemma~\ref{lem:bound_below_nehari}, there is $c_0>0$ such that $|u_0|_2=\lim_{n\to\infty}|u_n|_2\geq c_0>0$ and therefore $u_0\neq 0$. \end{proof} \begin{lemma}\label{lem:L2conv} Let $u_0\in L^2(\Omega)$ and let $(u_{k})_{k\in\mathbb{N}}\subset L^2(\Omega)$ be such that $u_k\to u_0$ in $L^2(\Omega)$ as $k\to\infty$. If $(\alpha_k)_{k\in\mathbb{N}}\subset[0,\tfrac{1}{2})$ is such that $\lim_{k\to\infty}\alpha_k=0$, then, passing to a subsequence, \begin{align*} \lim_{k\to\infty}\int_\Omega \ln(u_k^2)|u_k|^{\alpha_k}u_k\varphi \, dx =\int_\Omega \ln(u_0^2)u_0\varphi \, dx\qquad \text{ for all }\varphi\in C^\infty_c(\Omega). \end{align*} \end{lemma} \begin{proof} Note that $\sup_{t\in(0,1),\,k\in\mathbb{N}}|t|^{\alpha_k+1}|\ln(t^2)|<\infty.$ Then, since $\Omega$ is bounded, we can use the dominated convergence theorem to obtain that \begin{align} \lim_{k\to\infty}\int_{\{|u_{k}|\leq 1\}} \ln (u_{k}^2) |u_{k}|^{\alpha_k}u_{k}\varphi\ dx =\int_{\{|u_{0}|\leq 1\}} \ln (u_{0}^2)v_{0}\varphi\ dx.\label{bd0} \end{align} On the other hand, since $u_{k}\to u_0$ in $L^2(\Omega)$ as $k\to\infty$, there is a majorant $U\in L^2(\Omega)$ (see \emph{e.g.} \cite[Lemma A.1]{W96}) such that, passing to a subsequence, \begin{align*} |u_{k}|<U\qquad \text{in $\Omega$ for all $k\in\mathbb{N}$.} \end{align*} Moreover, on the set $\{|u_{k}|>1\}$, we can use Lemma~\ref{ln:bd} (with $\alpha=\frac{3}{2}<2=\beta$) to obtain that \begin{align*} \ln (u_{k}^2)|u_{k}|^{\alpha_k}u_{k}\varphi \leq\ln (u_{k}^2)|u_{k}|^{\frac{3}{2}}|\varphi|\leq 4|u_{k}|^2|\varphi|<4|U|^2\|\varphi\|_\infty\qquad \text{in $\{|u_k|>1\}$ for all }k\in\mathbb{N}. \end{align*} By dominated convergence, then \begin{align*} \lim_{k\to\infty}\int_{\{|u_{k}|>1\}} \ln (u_{k}^2) |u_{k}|^{\alpha_k}u_{k}\varphi\ dx =\int_{\{|u_{0}|>1\}} \ln (u_{0}^2)u_{0}\varphi\ dx, \end{align*} which, together with~\eqref{bd0}, yields the desired result. \end{proof} \begin{lemma}\label{mp:lem:ln} It holds that \begin{align}\label{mountain:pass:eq:lem} \inf_{\mathcal{N}_0} J_0 = \inf_{\sigma\in \mathcal T}\max_{t\in[0,1]}J_0(\sigma(t)), \end{align} where $\mathcal T:=\{\sigma\in C^0([0,1],\mathbb H(\Omega)): \sigma(0)=0, \sigma(1)\neq 0, J_0(\sigma(1)\leq 0)\}$. \end{lemma} \begin{proof} Let $v\in\mathcal{N}_0$, then \begin{align*} J_0(tv)&=t^2\mathcal{E}_L(v,v)-t^2\frac{\mu}{4}\int_\Omega v^2(\ln|(tv)^2|-1)\\ &=t^2\left(\mathcal{E}_L(v,v)+\frac{\mu}{4}(1-2\ln|t|)|v|_2^2-\frac{\mu}{4}\int_\Omega v^2\ln(v^2)\right). \end{align*} Therefore, there is $r_v>0$ such that $J_0(r_vv)<0$, and setting $\sigma_v(t):=tr_vv$ we have that $\sigma_v\in \mathcal T$. Note that $\max_{t\in[0,1]}J_0(\sigma_v(t))=J_0(v)$ (see Lemma~\ref{lem:A1A2}) and \begin{align}\label{mountain1} \inf_{\sigma\in\mathcal T}\max_{t\in[0,1]}J_0(\sigma(t))\leq \inf_{v\in \mathcal{N}_0}\max_{t\in[0,1]}J_0(\sigma_v(t)) =\inf_{v\in \mathcal{N}_0}J_0(v). \end{align} On the other hand, let $\kappa:\mathbb H(\Omega)\to \mathbb{R}$ be given by \begin{align*} \kappa(v):= \begin{cases} \exp\left(\frac{\mu\int_\Omega v^2\ln|v|\, dx-\mathcal{E}_L(v,v)}{|v|_2^2}\right), &\text{ if }v\neq0,\\ 0, &\text{ if }v=0. \end{cases} \end{align*} By Proposition~\ref{log:prop}, \begin{align*} \frac{\mu\int_\Omega v^2\ln|v|\, dx-\mathcal{E}_L(v,v)}{|v|_2^2} \leq \mu\ln(|v|_2)+\frac{N\mu}{4} a_N, \end{align*} and therefore $\kappa$ is continuous at $v=0$. Note that $\kappa(v)=1$ if and only if $v\in\mathcal{N}_0$. Furthermore, if $v\neq 0$ and $J_0(v)\leq 0$, then $\kappa(v)>1$. But then, for every $\sigma\in\mathcal T$, $\kappa(\sigma(0))=0$, $\kappa(\sigma(1))>1$, and then there is $t_0\in(0,1)$ such that $\kappa(\sigma(t_0))=1$, which implies that $\sigma(t_0)\in\mathcal{N}_0$. This yields that $\max_{t\in[0,1]}J_0(\sigma(t))\geq J_0(\sigma(t_0))\geq \inf_{\mathcal{N}_0}J_0$; but then $\inf_{\sigma\in\mathcal T}\max_{t\in[0,1]}J_0(\sigma(t))\geq \inf_{\mathcal{N}_0}J_0.$ This, together with~\eqref{mountain1}, implies~\eqref{mountain:pass:eq:lem} and ends the proof. \end{proof} We are ready to show Theorem~\ref{main:thm:2}. \begin{proof}[Proof of Theorem~\ref{main:thm:2}] Let $\Psi:\mathbb{H}(\Omega)\backslash\{0\}\to\mathbb{R}$ be given by $\Psi(u):=\mathcal{E}_L(u,u)-\frac{\mu}{2} \int_\Omega u^2 \ln(u^2) \ dx$. Then $\mathcal N_0=\Psi^{-1}(0)$ and $ \Psi'(u)u=2\mathcal{E}_L(u,u)-\mu \int_\Omega (\ln(u^2)+1)u^2 \ dx=-\mu|u|_2^2<0$ if $u\in\mathcal{N}_0$. In particular, $\mathcal{N}_0$ is a $C^1$-manifold. By Ekeland's variational principle \cite[Corollary 3.4]{E74}, there are $(u_n)_{n\in\mathbb{N}}\subset \mathcal{N}_0$ and $(\zeta_n)_{n\in\mathbb{N}}\subset \mathbb{R}$ such that \begin{align} 0\leq J_0(u_n)-\inf_{\mathcal{N}_0} J_0\leq \frac{1}{n^2} \quad \text{ and }\quad \left\| J_0'(u_n)-\zeta_n \Psi'(u_n) \right\|_{\mathcal{L}(\mathbb{H}(\Omega),\mathbb{R})}\leq \frac{1}{n}. \label{eke1} \end{align} In particular, $J_0(u_n)<\infty$ for all $n\in\mathbb{N}$ and \begin{align} o(1)&=\frac{1}{\|u_n\|}\left(J_0'(u_n)u_n-\zeta_n \Psi'(u_n)u_n\right)\notag\\ &=\frac{1}{\|u_n\|}\left( \mathcal{E}_L(u_n,u_n)-\frac{\mu}{2}\int_\Omega u_n^2 \ln(u_n^2) +\zeta_n \mu|u_n|_2^2\right) =\zeta_n \mu\frac{|u_n|_2^2}{\|u_n\|}\label{star} \end{align} as $n\to\infty$. By Proposition~\ref{prop:bd}, there is $C_2>0$ satisfying that \begin{align}\label{unbd} \|u_n\|<C_2\qquad \text{ for all }n\in\mathbb{N} \end{align} and there is $u_0\in \mathbb H(\Omega)\backslash\{0\}$ such that, passing to a subsequence, $u_n\rightharpoonup{u_0}$ weakly in $\mathbb H(\Omega)$ and $u_n\to{u_0}$ strongly in $L^2(\Omega)$ as $n\to\infty$. By \Cref{lem:bound_below_nehari}, there is $c_0>0$ such that $\frac{|u_n|_2^2}{\|u_n\|}>\frac{c_0}{C_2}$ for all $n\in\mathbb{N}$. Therefore~\eqref{star} implies that $\zeta_n\to 0$ as $n\to\infty.$ Moreover, by~\eqref{unbd}, there is $C_3>0$ such that \begin{align*} |\Psi'(u_n)v| = \left|2\mathcal{E}_L(u_n,v)-\mu \int_\Omega (\ln(u_n^2)+1)u_nv \ dx\right|<C_3 \end{align*} for all $v\in \mathbb{H}(\Omega)$ with $\|v\|=1$, where we used that $I'(u_n)$ is a bounded linear operator, by Lemma~\ref{I:C1}. As a consequence, by~\eqref{eke1}, \begin{align}\label{Jpz} \left\| J_0'(u_n) \right\|_{\mathcal{L}(\mathbb{H}(\Omega),\mathbb{R})}\leq \frac{1}{n} +\zeta_n \|\Psi'(u_n)\|_{\mathcal{L}(\mathbb{H}(\Omega),\mathbb{R})}\to 0\quad \text{ as }n\to\infty. \end{align} Then, by Lemma~\ref{lem:L2conv}, \begin{align*} 0 = \lim_{n\to\infty}J_0'(u_n)\varphi= \lim_{n\to\infty}\mathcal{E}_L(u_n,\varphi)-\frac{\mu}{2}\int_\Omega \ln(u_n^2)u_n \varphi\ dx =\mathcal{E}_L(u_0,\varphi)-\frac{\mu}{2}\int_\Omega \ln(u_0^2)u_0\varphi\ dx \end{align*} for all $\varphi\in C^\infty_c(\Omega)$. This implies that $u_0$ is a weak solution of~\eqref{lambda:problem}. Moreover, since $C^\infty_c(\Omega)$ is dense in $\mathbb{H}(\Omega)$, there is $(\varphi_n)_{n\in\mathbb{N}}\subset C^\infty_c(\Omega)$ such that $\varphi_n\to u_0$ in $\mathbb{H}(\Omega)$ as $n\to\infty$. Then, by Lemma~\ref{I:C1}, we have that \begin{align*} 0=\lim_{n\to \infty}\mathcal{E}_L(u_0,\varphi_n)-\frac{\mu}{2}\int_\Omega \ln(u_0^2)u_0\varphi_n\ dx =\mathcal{E}_L(u_0,u_0)-\frac{\mu}{2}\int_\Omega \ln(u_0^2)u_0^2\ dx. \end{align*} Therefore, $u_0\in\mathcal{N}_0$. Finally, using that $u_n,u_0\in\mathcal{N}_0$ and that $u_n\to u_0$ in $L^2(\Omega),$ \begin{align*} \inf_{\mathcal{N}_0}J_0&=\lim_{n\to\infty}J_0(u_n) =\lim_{n\to\infty}\mathcal{E}_L(u_n,u_n)-\frac{\mu}{4}\int_\Omega u_n^2(\ln(u_n^2)-1)\ dx\\ &=\lim_{n\to\infty}\frac{\mu}{4}\int_\Omega u_n^2\ dx= \frac{\mu}{4}\int_\Omega u_0^2\ dx=J_0(u_0). \end{align*} Note that~\eqref{mountain:pass} follows from Lemma \ref{mp:lem:ln}. Finally, let $u_0$ be any least-energy solution and we argue that $u_0$ does not change sign. By \cite[Lemma 3.3]{CW19}, we have that $|u_0|\in\mathbb H(\Omega)$ and \begin{align}\label{prop:comp_EL} \mathcal{E}_{L}(|u_0|,|u_0|)\leq \mathcal{E}_{L}(u_0,u_0). \end{align} Furthermore, the equality holds if and only if $u_0$ does not change sign. Let $t^0_{|u_0|}$ be given by~\eqref{t0:def} with $w=|u_0|$. Then $t_{|u_0|}^0|u_0|\in\mathcal{N}_0$ and, by~\eqref{prop:comp_EL} and because $u_0\in\mathcal{N}_0$, we have that $t_{|u_0|}^0\leq 1$. Therefore \begin{align*} J_0(u_0)=\inf_{\mathcal{N}_0}J_0\leq J_0(t_{|u_0|}^0|u_0|)=\frac{\mu}{4}(t_{|u_0|}^0)^2|u_0|_2^2\leq \frac{\mu}{4}|u_0|_2^2=J_0(u_0). \end{align*} This yields that $t_{|u_0|}^0=1$ and therefore~\eqref{prop:comp_EL} must hold with equality. This implies that $u_0$ does not change sign. \end{proof} \section{Convergence of solutions}\label{6:sec} In this section we show that least-energy solutions of~\eqref{subcritical:intro} converge in the $L^2$-sense, up to a subsequence, to a least-energy solution of~\eqref{log:prob:intro}. First, a standard use of the Nehari method yields the existence of least-energy solutions of~\eqref{subcritical:intro}. For completeness (and for comparison with Theorem~\ref{main:thm:2}) we include a short proof. \begin{theorem}\label{existence} Let $s\in(0,\tfrac{1}{4})$ and $p_s\in(2,2^*_s)$. There is a least-energy solution $u_s\in\mathcal{N}_s$ of~\eqref{subcritical:intro}, namely, $u_s$ satisfies~\eqref{ws:intro} and $J_s(u_s)=\inf_{\mathcal{N}_s}J_s.$ Furthermore, all least-energy solutions of~\eqref{subcritical:intro} are either positive or negative in $\Omega$. \end{theorem} \begin{proof} Let $\Psi:\mathcal{H}^s_0(\Omega)\backslash\{0\}\to\mathbb{R}$ be given by $\Psi(u):=\|u\|_s^2-|u|_{p_s}^{p_s}$. Then $\mathcal N_s=\Psi^{-1}(0)$ and $ \Psi'(u)u=2\|u_s\|^2_s-p_s|u_s|_{p_s}^{p_s}=(2-p_s)\|u_s\|_s^2<0$ if $u\in\mathcal{N}_s$. By Ekeland's variational principle \cite[Corollary 3.4]{E74} and Lemma~\ref{lem:Ebd}, there are $(u_n)_{n\in\mathbb{N}}\subset \mathcal{N}_s$, $(\zeta_n)_{n\in\mathbb{N}}\subset \mathbb{R}$, and $C>1$ such that \begin{align} C^{-1}\leq \|u_n\|_s\leq C,\qquad 0\leq J_s(u_n)-\inf_{\mathcal{N}_s} J_s\leq \frac{1}{n^2}, \qquad \left\| J_s'(u_n)-\zeta_n \Psi'(u_n) \right\|_{\mathcal{L}(\mathcal{H}_0^s(\Omega),\mathbb{R})}\leq \frac{1}{n} \label{eke1:s} \end{align} for all $n\in\mathbb{N}$. It follows that, $o(1)=\frac{1}{\|u_n\|_s}\left(J_s'(u_n)u_n-\zeta_n \Psi'(u_n)u_n\right) =\zeta_n (2-p_s)\|u_n\|_s$ as $n\to\infty$ and therefore $\zeta_n\to 0$ as $n\to\infty.$ Then $\left\|J_s'(u_n)\right\|_{\mathcal{L}(\mathcal{H}^s_0(\Omega),\mathbb{R})}\to 0$ as $n\to\infty$ and there is $u_s\in \mathcal{H}_0^s(\Omega)\backslash\{0\}$ such that, passing to a subsequence, $u_n\rightharpoonup{u_s}$ weakly in $\mathcal{H}^s_0(\Omega)$ and $u_n\to{u_s}$ strongly in $L^{p_s}(\Omega)$ as $n\to\infty$. With these facts, we conclude that $u_s$ is a weak solution of~\eqref{subcritical:intro}, $u_s\in\mathcal{N}_s$, and, up to a subsequence, \begin{align*} \inf_{\mathcal{N}_s}J_s&=\lim_{n\to\infty}J_s(u_n) =\left(\frac{1}{2}-\frac{1}{p_s}\right)\lim_{n\to\infty}\|u_n\|_s^2\geq \left(\frac{1}{2}-\frac{1}{p_s}\right)\|u_s\|_s^2= J_s(u_s)\geq \inf_{\mathcal{N}_s}J_s. \end{align*} Finally, let $u_s$ be a least-energy solution and let $t_{|u_s|}^s$ be given by~\eqref{ts:def}. Then $\||u_s|\|_s\leq \|u_s\|_s$, $t_{|u_s|}^s\leq 1$, and $t_{|u_s|}^s|u_s|\in\mathcal{N}_s$. Therefore, \begin{align*} J_s(u_s)\leq J_s(t_{|u_s|}^s|u_s|)=\left(\frac{1}{2}-\frac{1}{p_s}\right)(t_{|u_s|}^s)^{p_s}|u_s|_{p_s}^{p_s}\leq J_s(u_s), \end{align*} which yields that $t_{|u_s|}^s=1$ and $|u_s|$ is a nonnegative least-energy solution of~\eqref{subcritical:intro}. By the strong maximum principle (see \emph{e.g.} \cite[Proposition 3.3]{JF}) it follows that $|u_s|>0$ in $\Omega$ and therefore $u_s$ is either strictly positive or negative in $\Omega$. \end{proof} Next we recall some known properties. Let $u\in\mathbb H(\Omega)$ and $\varphi\in C^\infty_c(\Omega)$, then \begin{align}\label{ibyp} \mathcal{E}_L(u,\varphi)=\int_\Omega u L_\Delta \varphi\ dx, \end{align} see \cite[eq. (3.11)]{CW19}. Moreover, by \cite[Theorem 1.1]{CW19}, $L_\Delta \varphi\in L^p(\mathbb{R}^N)$ and \begin{align}\label{exp:thm} \lim_{s\to 0^+}\left|\frac{(-\Delta)^s \varphi- \varphi}{s}-L_\Delta \varphi\right|_p=0\qquad \text{ for all }0<p\leq \infty. \end{align} In particular, $(-\Delta)^s \varphi=\varphi+s L_\Delta \varphi+o(s)$ in $L^\infty(\mathbb{R}^N)$ as $s\to 0^+$. \medskip We are ready to show our main theorem. \begin{proof}[Proof of Theorem~\ref{main:thm}] Let $s_k\in(0,\frac{1}{4})$ be such that $s_k\to 0$ and let $(u_{s_k})_{k\in \mathbb{N}}$ be a sequence of least-energy solutions (the set of least-energy solutions is nonempty by Theorem~\ref{existence}). By~\eqref{C0} there is $C_0=C_0(\Omega,p)>0$ such that \begin{align}\label{C0thm} \|u_{s_k}\|_{s_k}<C_0\qquad \text{ for all }k\in\mathbb{N}. \end{align} Moreover, by Proposition~\ref{cor:Ebd}, the sequence $(u_{s_k})_{k\in \mathbb{N}}\subset \mathcal{N}_{s_k}$ is uniformly bounded in the Hilbert space $\mathbb H(\Omega)$. By the compact embedding of $\mathbb H(\Omega)$ into $L^2(\Omega)$, we obtain that, passing to a subsequence, there is $u_0\in \mathbb H(\Omega)$ such that \begin{align}\label{conv} u_{s_k}\rightharpoonup u_0\quad \text{ in }\mathbb H(\Omega),\qquad u_{s_k}\to u_0\quad \text{ in }L^2(\Omega)\text{ as }k\to\infty. \end{align} Observe that, if $f(s):=|t|^{p_s-2}t$, then $f'(s\tau)=p'(s\tau)\ln(|t|)|t|^{p(s\tau)-2}t$. Let $\varphi\in C^\infty_c(\Omega)$, then, by~\eqref{exp:thm}, \begin{align} \int_\Omega u_{s_k} (\varphi+{s_k} L_\Delta \varphi + o({s_k})) &=\int_\Omega u_{s_k} (-\Delta)^{s_k}\varphi= \int_\Omega |u_{s_k}|^{p_{s_k}-2}u_{s_k}\varphi\notag\\ &= \int_\Omega \left(u_{s_k}+{s_k}\int_0^1 p'(s_k\tau)\ln(|u_{s_k}|)|u_{s_k}|^{p(s_k\tau)-2}u_{s_k} \ d\tau\right)\varphi\ dx,\label{a:eq} \end{align} in $L^\infty(\Omega)$ as $k\to \infty$. Moreover, since $\int_\Omega u_{s_k} L_\Delta \varphi = \mathcal{E}_L (u_{s_k},\varphi)$, by~\eqref{ibyp}, then~\eqref{a:eq} implies that \begin{align} \mathcal{E}_L(u_{s_k},\varphi)+o(1) &=\int_\Omega u_{s_k} L_\Delta \varphi\ dx + o(1)\notag\\ &=\int_\Omega\int_0^1 p'(s_k\tau)\ln(|u_{s_k}|)|u_{s_k}|^{p(s_k\tau)-2}u_{s_k} \ d\tau\varphi\ dx,\label{rhs} \end{align} as $k\to \infty$ for all $\varphi\in C^\infty_c(\Omega)$. By Lemma~\ref{lem:L2conv}, passing to a subsequence, \begin{align*} \lim_{k\to\infty}\int_0^1 p'(s_k\tau) \int_{\Omega} \ln (|u_{s_k}|) |u_{s_k}|^{p(s_k\tau)-2}u_{s_k} \varphi\ dx\ d\tau =p'(0)\int_{\Omega} \ln (|u_{0}|)u_{0}\varphi\ dx. \end{align*} Therefore, letting $k\to \infty$ in~\eqref{rhs} we conclude that \begin{align}\label{ws} \mathcal{E}_L(u_0,\varphi) =p'(0) \int_\Omega \ln (|u_0|)u_0\varphi\qquad \text{ for all }\varphi\in C^\infty_c(\Omega). \end{align} Then, by density, $u_0$ is a weak solution of \begin{align*} L_\Delta u_0 = p'(0) \ln (|u_0|)u_0\quad \text{ in }\Omega. \end{align*} Let \begin{align*} \lambda_k=\frac{p(s_k)-2}{2^*_s-2}\in(0,1),\quad \alpha_k=(1-\lambda_k)2,\quad \beta_k=\lambda_k 2^*_{s_k},\quad r_k=\frac{1}{(1-\lambda_k)},\quad q_k=\frac{1}{\lambda_k}, \end{align*} and note that $\alpha_k+\beta_k=p(s_k)$, $\frac{1}{r_k}+\frac{1}{q_k}=1,$ and \begin{align*} \lim_{k\to\infty} \lambda_k = \frac{s_k\int_0^1 p'(s_k\tau)\ d\tau}{s_k \frac{4}{N-2s_k}} = p'(0)\frac{N}{4}\in(0,1), \end{align*} by~\eqref{p:hyp}. Then, by~\eqref{bds}, Theorem~\ref{thm:sobolev},~\eqref{C0thm}, and Hölder's inequality, there is $c=c(\Omega,p)>0$ and $C=C(\Omega,p)>0$ such that \begin{align*} c&< \|u_{s_k}\|_{s_k}^2 =\int_\Omega|u_{s_k}|^{p_{s_k}} =\int_\Omega|u_{s_k}|^{\alpha_k}|u_s|^{\beta_k}\\ &\leq \left(\int_\Omega|u_{s_k}|^{\alpha_k r_k}\right)^\frac{1}{r_k} \left(\int_\Omega|u_{s_k}|^{\beta_k q_k}\right)^\frac{1}{q_k}=|u_{s_k}|_2^{2(1-\lambda_k)} |u_{s_k}|_{2^*_{s_k}}^{2^*_{s_k}\lambda_k} \leq C|u_{s_k}|_2^{2(1-\lambda_k)}, \end{align*} for all $k\in\mathbb{N}$, and therefore \begin{align*} |u_0|_2=\lim_{k\to \infty}|u_{s_k}|_{2}\geq \left(\frac{c}{C}\right)^{\frac{1}{2(1-\frac{N}{4}p'(0))}}>0, \end{align*} This yields that $u_0\neq 0$ is a nontrivial weak solution and $u_0\in\mathcal{N}_0$. Next, we show that $u_0$ is a least-energy solution of the limiting problem, namely, that \begin{align*} \frac{p'(0)}{4}|u_0|_2^2=J_{0}(u_0)=\inf_{\mathcal{N}_0}J_{0}=:c_0>0, \end{align*} where $J_0(u)=\frac{1}{2}\mathcal{E}_L(u,u)-\frac{p'(0)}{4}\int_\Omega |u|^2(\ln(|u|^2)-1)\ dx$. Noting that $\lim_{k\to\infty}\frac{1}{s_k}\left(\frac{1}{2}-\frac{1}{p_{s_k}}\right)=\frac{p'(0)}{4}$, we have that \begin{align*} J_{s_k}(u_{s_k})=\left(\frac{1}{2}-\frac{1}{p_{s_k}}\right)\|u_{s_k}\|_{s_k}^2\quad \text{ and }\quad \lim_{k\to\infty}\frac{1}{s_k}J_{s_k}(u_{s_k})=\frac{p'(0)}{4}\lim_{k\to\infty}\|u_{s_k}\|_{s_k}^2. \end{align*} Let $c_k:=\frac{1}{s_k}\left(\frac{1}{2}-\frac{1}{p_{s_k}}\right)\|u_{s_k}\|_{s_k}^2$. Then, by~\eqref{C0thm}, there is $c^*\in\mathbb{R}$ such that, passing to a subsequence, $\lim_{k\to\infty}c_k=c^*$. We claim that $c^*=c_0$. By Fatou's Lemma, \begin{align}\label{3} c_0=\frac{p'(0)}{4}|u_0|_2^2 \leq \frac{p'(0)}{4}\liminf_{k\to\infty}\int_{\mathbb{R}^N}|\xi|^{2s_k}|\widehat u_{s_k}|^2\ d\xi =\liminf_{k\to\infty}c_k=c^*. \end{align} On the other hand, by Theorem~\ref{main:thm:2} with $\mu=p'(0)$, there is $v\in \mathcal{N}_0$ such that $J_0(v)=c_0$. Let $(v_n)_{n\in \mathbb{N}}\subset C^\infty_c(\Omega)\cap \mathcal{N}_0$ such that $v_n\to v$ as $n\to\infty$ in $\mathbb{H}(\Omega)$. By Lemma~\ref{lem:t0} and using that $v_n\in\mathcal{N}_0$, \begin{equation}\label{1} \lim_{k\to\infty}t^n_k= 1\quad \text{ for every }n\in\mathbb{N},\text{ where }t^n_k:=\left(\frac{\|v_n\|_{s_k}^2}{|v_n|_{p_{s_k}}^{p_{s_k}}}\right)^{\frac{1}{p_{s_k}-2}}. \end{equation} But then, using the minimality of $u_{s_k}$,~\eqref{1}, Lemma~\ref{Es:exp:lem}, and that $t^n_k v_n\in\mathcal{N}_s$, \begin{align*} c_*=\lim_{k\to\infty}c_{k} =\lim_{k\to\infty}\frac{1}{s_k}J_{s_k}(u_{s_k}) \leq \lim_{k\to\infty}\frac{1}{s_k}J_{s_k}(t^n_{k}v_n)=\lim_{k\to\infty}\frac{1}{s_k}\left(\frac{1}{2}-\frac{1}{p_{s_k}}\right)\|t_{k}^nv_n\|_{s_k}^2 =\frac{p'(0)}{4}|v_n|_2^2. \end{align*} Since $\lim_{n\to\infty}\frac{p'(0)}{4}|v_n|_2^2=\frac{p'(0)}{4}|v|_2^2=J_0(v)=c_0$, we have that $c_*\leq c_0$. Together with~\eqref{3}, we conclude that $c_*=c_0$. Finally, arguing as in~\eqref{3} and using~\eqref{conv}, \begin{align*} c_0\leq J_0(u_0)=\frac{p'(0)}{4}|u_0|_2^2 \leq \frac{p'(0)}{4}\liminf_{k\to\infty}\|u_{s_k}\|_{s_k}^2 =\lim_{k\to\infty}c_k=c_*=c_0, \end{align*} which implies that $J_0(u_0)=c_0$. \end{proof} \section{Closing remarks}\label{c:rmks} To finish this paper, we comment on the following. \begin{remark}\label{open:rmk}(On the extremal cases for $p'(0)$) The cases $p'(0)=0$ and $p'(0)=\frac{4}{N}$ are not covered by Theorem~\ref{main:thm} (note that the assumption $2<p(s)<2^*_s$ for $s\in(0,\frac{1}{4})$ implies that $p'(0)\in[0,\frac{4}{N}]$). For $p'(0)=0$, the characterization of the limiting problem (see the proof of Theorem \ref{main:thm}) requires a second order \textemdash or even higher, if $p''(0)=0$\textemdash expansion of the fractional Laplacian and of the power nonlinearity at $s=0$, and we do not pursue this here. On the other hand, if $p'(0)=\frac{4}{N}$ (which corresponds to the ``critical case"), then we cannot use the logarithmic Sobolev inequality to obtain estimates in the $\mathbb{H}(\Omega)$-norm (and gain compactness), see for example Lemma~\ref{lem:bound_below_nehari} and Proposition~\ref{prop:bd}, where the fact that $\lambda:=p'(0)\frac{N}{4}<1$ is crucial. For problem \eqref{subcritical:intro} with the critical Sobolev exponent $p_s=2^*_s$, the use of a suitable symmetric variational framework can yield the existence of solutions in some bounded (and unbounded) domains, see \emph{e.g.} \cite{HSS21} and the references therein. We conjecture that a similar approach can be used to study \eqref{lambda:problem} with $\mu=\frac{4}{N}$. \end{remark} \begin{remark}(On the positivity properties of solutions) In Theorem \ref{existence}, a strong maximum principle for $(-\Delta)^s$ is used to show that least-energy solutions of \eqref{subcritical:intro} are either strictly positive or strictly negative in $\Omega$. However, maximum principles for $L_\Delta$ are more delicate and do not hold in general (see \cite[Theorem 1.8]{CW19}). As a consequence, in Theorem \ref{main:thm:2}, we only show that the least-energy solutions are either nonnegative or nonpositive. \end{remark} \begin{remark}[Generalizations and extensions]\label{ext:rmk} Our main variational tool is the Nehari manifold method. This is a very flexible and versatile approach that can be applied to study a wide set of nonlinear problems. For instance, in \cite{SW10}, a generalized Nehari method is used to show the existence of a least-energy solution (\emph{a.k.a.} ground state) of the problem \begin{align}\label{L} -\Delta u -\lambda u= f(x,u)\quad \text{ in }\Omega,\qquad u=0\quad \text{ on }\partial \Omega, \end{align} where $\lambda\in\mathbb{R}$ and $f\in C(\Omega\times \mathbb{R},\mathbb{R})$ satisfies the following assumptions: \begin{enumerate} \item (Subcriticality) $|f(x,u)|\leq a(1+|u|^{q-1})$ for some $a>0$ and $2<q<\frac{2N}{N-2}=2^*_1$, $N\geq 3$. \item (Superlinearity) $f(x,u)=o(u)$ uniformly in $x$ as $u\to 0$ and $F(x,u)/u^2\to\infty$ uniformly in $x$ as $|u|\to\infty$, where $F(x,u):=\int_0^u f(x,s)\ ds$. \item (Monotonicity) $u\mapsto f(x,u)/|u|$ is strictly increasing on $(-\infty,0)$ and $(0,\infty)$. \end{enumerate} Note that $f$ is not assumed to be $C^1$. Furthermore, if $f$ is odd in $u$, then \eqref{L} has infinitely many solutions. Combining the approach from \cite{SW10} with the methods in this paper, it is possible to characterize the small order asymptotics of \eqref{L} when $-\Delta$ is substituted with $(-\Delta)^s$. In this case, the condition for subcriticality would be $|f_s(x,u)|\leq a(1+|u|^{q_s-1})$ for some $a>0$ and $2<q_s<\frac{2N}{N-2s}=2^*_s$, $s\in(0,\frac{N}{2}),$ $N\geq 1$. To analyze the small order limit as $s\to 0^+$, suitable (logarithmic-subcriticality) assumptions need to be imposed on the behavior of the map $s\mapsto f_s$ at $s=0$, as in \eqref{p:hyp}. \end{remark} \renewcommand{\abstractname}{Acknowledgements} \begin{abstract} \end{abstract} \vspace{-0.5cm} Víctor Hernández-Santamaría is supported by the program ``Estancias posdoctorales por M\'exico'' of CONACyT, Mexico. Alberto Saldaña is supported by UNAM-DGAPA-PAPIIT grant IA101721, Mexico. The authors thank Pierre Aime Feulefack, Sven Jarohs, and Tobias Weth for helpful discussions and Harbir Antil for sharing some relevant references.

https://arxiv.org/abs/2211.07689

Towards the Erdős-Gallai Cycle Decomposition Conjecture

In the 1960's, Erdős and Gallai conjectured that the edges of any $n$-vertex graph can be decomposed into $O(n)$ cycles and edges. We improve upon the previous best bound of $O(n\log\log n)$ cycles and edges due to Conlon, Fox and Sudakov, by showing an $n$-vertex graph can always be decomposed into $O(n\log^{*}n)$ cycles and edges, where $\log^{*}n$ is the iterated logarithm function.

\section{Introduction} When is it possible to decompose a graph into edge disjoint subgraphs with certain properties? Many classical problems in extremal combinatorics fall within this framework and its natural hypergraph generalisation, while decomposition problems have strong links to many other fields, including the design of experiments, coding theory, complexity theory and distributed computing (see, for example, \cite{network_decompositions1989,designs_applications1989,jukna_complexity2006}). The particular case where we seek to decompose a graph into cycles has a long history, dating back to the 18th century and Euler's result on the existence of Euler tours. As Veblen~\cite{veblen1,veblen2} observed for his algebraic approach to the Four-Colour Theorem, Euler's result immediately implies that any graph with even vertex degrees (i.e., any \emph{Eulerian} graph) has a decomposition into cycles. As it is immediate that any graph with a vertex of odd degree cannot be decomposed into cycles, this exactly characterises which graphs have cycle decompositions. Another very classical cycle decomposition result is due to Walecki \cite{lucas1883recreations} from 1892, who showed it is possible to decompose any complete graph with an odd number of vertices into Hamilton cycles. This gives a cycle decomposition into few cycles, indeed, into optimally few cycles. This raises a very natural question of whether every Eulerian graph has a cycle decomposition into few cycles? That only $O(n)$ cycles might be needed to decompose any $n$-vertex Eulerian graph is easily seen to be equivalent to the following classical conjecture of Erd\H{o}s and Gallai \cite{erdos1966representation} dating back to the 1960's, which is one of the major open problems on graph decompositions. \begin{conj}\label{conj:EG} Any $n$-vertex graph can be decomposed into $O(n)$ cycles and edges. \end{conj} While Conjecture~\ref{conj:EG} is equivalent to conjecturing that every $n$-vertex Eulerian graph can be decomposed into $O(n)$ cycles, as noted above, if they both hold then the optimal implicit constants in these conjectures seem likely to be different. For the Eulerian problem, Haj\'os conjectured that $\frac n2$ cycles should be sufficient \cite{lovasz1968covering} (see also~\cite{dean1986smallest,chung1980,bondyEG1990,fan2003covers,girao2021path}), while the best known lower bound for the number of cycles and edges required in Conjecture~\ref{conj:EG} is $(\frac{3}{2}-o(1))n$, as observed by Erd\H{o}s in 1983 \cite{erdHos1983some}, improving on a previous construction of Gallai \cite{erdos1966representation} (see Section~\ref{sec:final}). Since its formulation, the Erd\H{o}s-Gallai Conjecture has often been highlighted (see, for example,~\cite{bondyEG1990,pyber1996covering,pyberEG1991,dano2015,glock2016optimal,girao2021path,conlon2014cycle}), with Erd\H{o}s himself mentioning it in many of his open problem collections \cite{erdHos1983some,erdHossome1971,erdossolved1981,erdos1973problems}. Despite this attention, and a lot of work on related problems over the years, direct progress towards the Erd\H{o}s-Gallai Conjecture has only been made within the last decade. The previous related results, which we discuss first, are mostly on the analogous path decomposition problem and the covering version of the Erd\H{o}s-Gallai conjecture. \textbf{Path decompositions.} In the 1960's, Gallai~\cite{lovasz1968covering} posed the analogous path decomposition version of Conjecture~\ref{conj:EG}. In particular, he conjectured that any connected $n$-vertex graph can be decomposed into at most $\frac{n+1}{2}$ paths. Lov\'asz \cite{lovasz1968covering} in 1968 proved that any graph can be decomposed into at most $n-1$ paths. This follows easily from his complete solution to the problem of how many paths \emph{or} cycles one needs to decompose an $n$-vertex graph, to which the answer is $\floor{\frac n2}$. Currently the best general bound in the path decomposition problem is due independently to Dean and Kouider~\cite{dean2000gallai} and Yan~\cite{yan1998path}, who showed that any graph can be decomposed into at most $\lfloor \frac23 n\rfloor$ paths. Gallai's path decomposition conjecture is known to hold for quite a few special classes of graphs, with connected planar graphs being the most recent addition to the list. This latest result is due to Blanch\'e, Bonamy and Bonichon in \cite{blanche2021gallai}, where a more exhaustive list of partial results can also be found. \textbf{Covering problems.} Another interesting direction which has attracted a lot of attention is the covering version of Conjecture~\ref{conj:EG}, in which we do not insist that the cycles we find should be disjoint, only that together they contain all the edges of the host graph. In 1985, Pyber~\cite{pyber1985erdHos} proved the covering version of the Erd\H{o}s-Gallai conjecture, showing that the edges of any $n$-vertex graph can be covered with $n-1$ cycles and edges. The analogous covering version of Gallai's conjecture, raised by Chung \cite{chung1980} in 1980, has been settled first approximately by Pyber~\cite{pyber1996covering} in 1996 and then completely by Fan~\cite{fan2002subgraph} in 2002, who showed that the edges of any connected graph can be covered by $\ceil{ \frac{n}{2}}$ paths. The covering version of Haj\'os's conjecture was also solved by Fan~\cite{fan2003covers}, who showed that any $n$-vertex Eulerian graph can be covered by at most $\lfloor \frac{n-1}{2}\rfloor$ cycles, settling another conjecture of Chung. As with the other two covering results above, this bound is best possible. \textbf{Results on the Erd\H{o}s-Gallai conjecture.} In more recent years, the Erd\H{o}s-Gallai conjecture (along with more accurate results on the implicit bounds) has been shown to hold for two large specific classes of graphs -- random graphs and graphs with linear minimum degree. The conjecture was first established for a typical binomial random graph $G(n,p)$ (for any $p=p(n)$) by Conlon, Fox and Sudakov \cite{conlon2014cycle}. Kor{\'a}ndi, Krivelevich, and Sudakov \cite{dano2015} found the correct leading constant here, showing that $(\frac{1}{4}+\frac{p}{2}+o(1))n$ cycles and edges are typically sufficient to decompose $G(n,p)$. For constant edge probability $p$, Glock, K{\"u}hn, and Osthus~\cite{glock2016optimal} were even able to determine with high probability the exact minimum number of cycles and edges required to decompose a (quasi)random graph. On the other hand, the Erd\H{o}s-Gallai conjecture was first shown to hold for graphs with linear minimum degree again by Conlon, Fox and Sudakov \cite{conlon2014cycle}. Very recently, the asymptotically correct bound of $(\frac32+o(1)) n$ cycles and edges has been proved by Gir{\~a}o, Granet, K{\"u}hn, and Osthus~\cite{girao2021path} for large graphs with linear minimum degree. A fundamental challenge towards establishing the Erd\H{o}s-Gallai conjecture is its generality, and indeed these previous results make progress only by imposing a fairly strong constraint on the structure or randomness of the graph. For almost 50 years, the best known bound in the general case of the Erd\H{o}s-Gallai conjecture (as observed by Erd\H{o}s and Gallai) came from a simple argument involving the iterative removal of a longest cycle, which shows that an $n$-vertex graph can always be decomposed into $O(n\log n)$ cycles and edges. In 2014, Fox, Conlon and Sudakov~\cite{conlon2014cycle} made the first major breakthrough on this problem, showing that such a decomposition with only $O(n\log\log n)$ cycles and edges always exists. Here we will give the following improvement on this bound, where $\log^\star n$ is the iterated logarithm function. \begin{restatable}{theorem}{logstar} \label{thm:logstar} Any $n$-vertex graph can be decomposed into $O(n\log^\star n)$ cycles and edges. \end{restatable} Key to the decompositions used by Conlon, Fox and Sudakov~\cite{conlon2014cycle} was to show that \emph{a)} graphs $H$ with certain expansion properties can be decomposed into $O(|H|)$ cycles and few edges and \emph{b)} any $n$-vertex graph $G$ can be decomposed nicely into such `expanders' $H$ and a small number of leftover edges. Combined, this gives a decomposition of $G$ into $O(n)$ cycles and some leftover edges, and it can be shown that iterating this on the leftover edges while removing any particularly long cycles causes the average degree of the leftover edges to drop significantly each time, so that after $\log\log n$ iterations the decomposition given by~\cite{conlon2014cycle} is achieved. To prove Theorem~\ref{thm:logstar}, essentially we need the average degree of the leftover edges to drop much faster, and so at \emph{b)} we have to take a much weaker condition on the `expanders' $H$. Effectively we replace the strong expansion used in \cite{conlon2014cycle}, with a very weak \emph{sublinear} expansion (as introduced by Koml\'os and Szemer\'edi~\cite{K-Sz-1,K-Sz-2}), in particular using a \emph{robust} sublinear expansion where sets expand sublinearly despite the additional removal of a possibly-superlinear set of edges (see Section~\ref{sec:generalexpansionchat} for a discussion of these forms of expansion and their background). Using this much weaker form of expansion introduces a raft of issues when we decompose an expander into few cycles and edges (for \emph{a)} above), resulting in a very different approach to that used in \cite{conlon2014cycle}. In order to do this, we introduce a range of new tools, which we hope will find further applications. In particular, we would highlight a new approach to robust sublinear expansion (see Section~\ref{sec:expansion}) and the (surprisingly difficult) result that randomly sampling the vertices of an expander is likely to induce a subgraph with a (somewhat weaker) expansion property (see Lemma~\ref{lem:expandintorandom2}). Additional new tools include a similar result but while randomly sampling edges, the (almost) decomposition of any graph into robust sublinear expanders, and the finding of a sparse `connective skeleton' in expanders to connect vertex pairs with paths. These, and other tools, and how they come together to prove Theorem~\ref{thm:logstar}, are discussed in \Cref{subsec:sketch}. As discussed in Section~\ref{sec:generalexpansionchat}, sublinear expansion has been useful in many different settings in which our tools may also be useful (see Section~\ref{sec:final} for some examples). In particular, \emph{robust} sublinear expansion (specifically considering the deletion of superlinearly many edges) is a very recent concept and we hope our new perspective and tools will contribute to its development and use. Several of our intermediate results and tools might also ultimately prove useful towards proving the Erd\H{o}s-Gallai conjecture in full as they often decompose any $n$-vertex graph into $O(n)$ cycles and a graph with some other structure imposed. This is discussed further in our concluding remarks in \Cref{sec:final}. \section{Preliminaries}\label{sec:prelim} After we introduce our notation, we give a detailed sketch of our methods before outlining the rest of the paper. \subsection{Notation} Given a graph $G$ we will denote by $V(G)$ and $E(G)$ its vertex and edge set, respectively. Given a vertex $v\in V(G)$, we denote its degree by $d_G(v)$ and the set of its neighbours by $N_G(v)$. We write $\Delta(G)$ for the maximum degree of a vertex in a graph $G$. Given a subset of vertices $U \subseteq V(G)$ we denote by $N_G(U)$ the set of vertices in $V(G) \setminus U$ which have a neighbour in $U$. Given $U \subseteq V(G)$ we define $B^{i}_G(U)$ as the set of vertices at distance at most $i$ from a vertex of $U$ in the graph $G$, i.e.\ the ball of radius $i$ around $U$ in $G$, and write simply $B_G(U)=B^1_G(U)$. Given $V \subseteq V(G)$ we write $G[V]$ for the subgraph of $G$ induced by the vertex set $V$, and write $G \setminus V$ for $G[V(G) \setminus V]$. Given $F \subseteq E(G)$ we write $G-F$ for the subgraph of $G$ obtained by deleting all the edges in $F$. Given multiple (hyper)graphs $H_1,\ldots, H_t$ we write $H_1 \cup \ldots \cup H_t$ for the (hyper)graph with vertex set $\bigcup_{i\in [t]}V(H_i)$ and edge set $\bigcup_{i\in [t]}E(H_i)$. Given vertices $v$ and $u$, by a $vu$-path/walk we refer to a path/walk joining $v$ and $u$. We write $X \sim \text{Bin}(n,p)$ to mean that $X$ is a random variable distributed according to the binomial distribution with parameters $n$ and $p$. We denote by $\mathcal{G}(n,p)$ the binomial random graph defined as the graph with vertex set $[n]$ in which we sample every edge with probability $p$ independently from all other edges. We write $G \sim \mathcal{G}(n,p)$ to mean that $G$ is sampled according to $\mathcal{G}(n,p).$ All our logarithms have base two. For each $k\geq 1$, let $\log^{[k]}(n)=\underbrace{\log \log \ldots \log}_{k \text{ times}} n$, and let $\log^{[0]}n=n$. The iterated logarithm function $\log^{\star} n$ is the minimum number of times we need to apply the logarithm function to $n$ until it becomes at most one, that is, the least $k\geq 0$ such that $\log^{[k]}n\leq 1$. Throughout the paper we make no attempt to optimise constants and logarithmic factors; often we are wasteful to improve readability. With the same goal, we also omit floor and ceiling signs wherever they are not crucial. \subsection{Proof sketch}\label{subsec:sketch} Our methods to find an (edge) decomposition into cycles and edges is iterative, where a single iteration, applied to an $n$-vertex graph $G$ with average degree $d$, performs the following steps for some appropriately large constant $C$. \begin{itemize} \item Repeatedly remove any cycle of length at least $d$ and add it to the decomposition, giving at most $\frac n2$ new cycles. \item Decompose $G$ into edge disjoint subgraphs $R_i$ with a certain expansion property which, combined with the lack of long cycles, guarantees that $|R_i|= O(d \log^4 d)$. These subgraphs are almost vertex disjoint, so that $\sum_{i}|R_i|\leq 2n$. \item Decompose each of the subgraphs $R_i$ into cycles and edges, using in total $O(n)$ cycles and $O(n\log^Cd)$ edges. \end{itemize} This finds a decomposition of $G$ into $O(n)$ cycles and a subgraph consisting of leftover edges of average degree at most $O(\log^Cd)$. We now iterate by applying the same argument to our much sparser graph consisting of leftover edges. After at most $O(\log^\star d)$ iterations we will be left with a graph of constant average degree. We then simply make all its edges part of our decomposition, which together with the $O(n)$ cycles we found at each of the $O(\log^\star d)$ iterations gives our desired decomposition for Theorem~\ref{thm:logstar} -- in fact, using only $O(n\log^\star d)$ edges for any $n$-vertex graph with average degree $d$. The majority of our work lies in carrying out the final step in this iteration, where the key part of this step is to show the following intermediate result. \begin{restatable}{theorem}{decompcyclelinear} \label{thm:decompexander} There exists $C>0$ such that any $r$-vertex graph decomposes into $O(r)$ cycles and $O(r\log^{C}r)$ edges. \end{restatable} Note that this intermediate result is applied to graphs $R_i$, each of which has order at most $O(d \log^4 d)$, so that, as $\sum_{i}|R_i|\leq 2n$, we get in total $O(n)$ cycles and $O(\sum_{i}|R_i|\log^C|R_i|)=O(\sum_{i}|R_i|\log^Cd)=O(n\log^Cd)$ edges in total from the decomposition in the final step of the iteration. Though we use a slightly modified iteration argument, till now our approach has the same structure as the one taken by Conlon, Fox and Sudakov in \cite{conlon2014cycle}, where they prove a weaker version of Theorem~\ref{thm:decompexander} in which they allow, instead of $O(r\log^C r)$, up to $O(r^{2-1/10})$ edges in the decomposition. This much weaker bound leads them to iterate $O(\log\log d)$ times and thus use a decomposition using $O(n\log\log d)$ cycles and edges in total. The key difference is that we are able to replace the extremely strong expansion properties used in~\cite{conlon2014cycle} with a very weak form of expansion. To make this change successfully, we need to carefully develop the weak expansion property we use (which originates with Koml\'os and Szemer\'edi~\cite{K-Sz-1,K-Sz-2}) as well as solve the variety of problems caused by working with this weak expansion. For this development, the main insight is the new perspective we bring to robust sublinear expansion. That we can decompose an arbitrary graph into edges and expanders follows relatively naturally from the definition of robust sublinear expansion (see \Cref{sec:decomp-into-expanders} for more details on the expander partitioning lemma), and the same decomposition result allows us to carry out the second step of the iteration mentioned above. This leaves the difficult task of decomposing an expander graph into few cycles and edges, which we now discuss. \smallskip \noindent\textbf{Decomposing expanders into few cycles and edges.} For some constant $C>0$, we now assume that we wish to decompose an $n$-vertex graph $G$ with the following (slightly simplified) expansion condition (see Section~\ref{sec:expansion} for the full condition we use): for each $U\subseteq V(G)$ and $F\subseteq E(G)$ with $|U|\leq \frac23 n$ and $|F|\leq |U|\log^Cn$, we have \begin{equation}\label{expandprop} |N_{G-F}(U)|\geq \frac{1}{\log^2n}|U|. \end{equation} As the size of the neighbourhood guaranteed in \eqref{expandprop} is smaller than $|U|$, this type of expansion is known as \emph{sublinear expansion}. Roughly speaking, our main strategy is to set aside a `sparse connecting skeleton' $H\subseteq G$, before initially decomposing the edges of $G-H$ into $O(n)$ paths and cycles using a result of Lov\'asz~\cite{lovasz1968covering} stated in the introduction. We then use short paths from the connecting skeleton $H$ to connect up each initial path into a cycle, before simply taking each unused edge of $H$ as part of our decomposition. In order for this to produce a correct decomposition we need $H$ to be very sparse, in particular with at most $n\log^{O(1)}n$ edges. To aid the connection of the paths from the initial path/cycle decomposition, we need that each vertex does not appear too often as an endvertex of these paths. This we ensure by proving a simple, but crucial, corollary of Lov\'asz's result (see Corollary~\ref{cor:lovasz}), which will allow us to decompose $G-H$ into a collection $\mathcal{P}$ of $O(n)$ paths in which each vertex appears as an endvertex at most twice, so that the endvertices of the paths are well spread across the graph. More problematically, note that in order to get an actual cycle we need to connect the endvertices of each path $P\in \mathcal{P}$ using a path in $H$ which is internally vertex disjoint from $P$. To deal with this, we change this outline slightly as follows. We partition $V(G)=V_1\cup V_2\cup V_3$ by placing each vertex independently into a set $V_i$ uniformly at random, and show that $G$ contains sparse subgraphs $H_1,H_2,H_3$ (each with $n\log^{O(1)} n$ edges) with the following property for each $i\in [3]$, where a \emph{path through $V_i$} is one whose interior vertices are all in $V_i$. We place no restriction on the endvertices themselves, so in particular a single edge path is a path through any set since it contains no interior vertices. \begin{enumerate}[label = \textbf{P}] \item For any set $\mathcal{P}\subseteq \binom{V(G)}{2}$ such that each vertex appears in at most $2$ pairs in $\mathcal{P}$, there are edge disjoint paths $P_{xy}$, $\{x,y\}\in \mathcal{P}$, such that, for each $\{x,y\}\in \mathcal{P}$, $P_{xy}$ is an $xy$-path through $V_i$ in $H_i$ with length $O(\log^7n)$.\label{prop:sketch} \end{enumerate} We then split the edges of $G-H_1-H_2-H_3$ into three subgraphs $G_1,G_2,G_3$ in such a way that $V(G_i)= V_{i+1} \cup V_{i+2},$ for each $i\in [3]$, with indices taken modulo $3$. Applying our path decomposition corollary to each $G_i$ then gives a decomposition of all the edges outside of $H_1 \cup H_2 \cup H_3$ into paths whose endvertices are well spread across the graph. Note that the paths decomposing $G_i$ completely avoid $V_i$ so by using the property~\ref{prop:sketch} we can connect each of these paths into actual cycles using edges of $H_i$. In total we find $O(n)$ edge disjoint cycles which use all the edges of $G-H_1-H_2-H_3$ so the total number of uncovered edges, which all belong to $H_1\cup H_2\cup H_3$, is small. With such a weak expansion property as that at \eqref{expandprop}, whether we can do this is initially far from clear. Building up to this, we ask the following three questions. \vspace{-0.2cm} \begin{enumerate}[label = \roman{enumi})] \item Can we connect pairs of vertices with edge disjoint paths using the whole of $G$? \\I.e., does property \ref{prop:sketch} hold if $H_i=G$ and $V_i=V(G)$? \item If so, can we do this using only a random subset of vertices $V_i$ for the interior vertices of the paths? \\I.e., does property \ref{prop:sketch} hold if $H_i=G$? \item If so, can we do this using only a sparse subgraph $H_i$ of $G$? \\I.e., can we find a sparse $H_i \subseteq G$ so that property \ref{prop:sketch} holds for $H_i$? \end{enumerate} \vspace{-0.2cm} \smallskip \textbf{i) Finding edge disjoint paths in $G$.} The expansion condition on $G$ at \eqref{expandprop} is sufficient to imply that any pair of vertices in $H$ are connected by a path of length $O(\log^3n)$ by expanding the neighbourhoods around each of the two vertices until they become large enough that they must overlap. Moreover, the robustness of our condition at \eqref{expandprop} (i.e., that this expansion can avoid using an arbitrary, but not too large, set of edges $F$) allows us, with only a bit more work, to show that, for any collection $\mathcal{P}$ of pairs of vertices as in property \ref{prop:sketch}, we could find at least $\Theta(\log^4n)\cdot |\mathcal{P}|$ edge disjoint paths in $G$ which each connect \emph{some} vertex pair in $\mathcal{P}$ and have length $O(\log^3n)$. As this holds in fact for any subset $\mathcal{P}'\subseteq \mathcal{P}$ in place of $\mathcal{P}$, this allows us to use the Aharoni-Haxell hypergraph matching theorem (see Theorem~\ref{thm:hyperhall}) to select, for each pair $\{x,y\}\in \mathcal{P}$, an $xy$-path in $G$, so that all these paths are edge disjoint. In total this allows us to answer question i) positively. \textbf{ii) Connecting through the random vertex subset $V_i$.} With $i\in [3]$ and $V_i$ a random subset of $V$ with size approximately $\frac{n}3$ as chosen above, unfortunately it seems it does not follow easily that $H[V_i]$ likely satisfies a similar expansion property to~\eqref{expandprop}. Indeed, firstly, if the inequality \eqref{expandprop} is tight for a set $U$, then with probability $\exp(-\Theta(|U|/\log^2n))$ we have that $U$ has no neighbours selected into our random subset $V_i$, too high a probability to naively take a union bound to avoid this event over all sets $U$ with any fixed size. Secondly, the robustness condition we need in order to find multiple edge disjoint paths (to then apply the Aharoni-Haxell hypergraph matching theorem) requires us to avoid an arbitrary set of $|U|\log^{O(1)}n$ edges, and again there are too many choices to just take a union bound. The first problem here is the most difficult to overcome. The second problem can be overcome by splitting the edges of $G$ randomly into $t$ subgraphs $G_i$, $i\in [t]$, for some $t=\log^{O(1)}n$, and showing then that (with high probability) each of these has some (slightly weaker) expansion property (see Lemma~\ref{lem:partitionedgesintoexpanders}). When we look for edge disjoint collections of paths in each graph $G_i$ separately for the application of the Aharoni-Haxell hypergraph matching theorem, by finding $\frac 1t$ fraction of the required paths in each graph $G_i$ we need to avoid fewer edges by expanding in each $G_i$ separately rather than $G$. Thus, we have fewer sets of edges over which to take a union bound, solving the second problem above. Solving the first problem to get some weak expansion into $V_i$ is the crux of this paper, and is where our new perspective on robust sublinear expansion is critical. We will show that it follows from this new perspective that, for such sets $U$ and $F$, we have either \begin{enumerate} \item $N_{G-F}(U)$ is actually much larger than guaranteed by \eqref{expandprop}, or \item there is a set $U'\subseteq U$ which is much smaller than $U$ but whose neighbourhood alone contains at least $\frac{|U|}{\log^2n}$ vertices in $N_{G-F}(U)$. \end{enumerate} Given this, a natural approach is to take a union bound over all `well-expanding' sets $U$, meaning that they fall under condition 1.\ above, to guarantee a constant fraction of their neighbourhood gets sampled into our random set $V_i$. Since the subset $U'$ from condition 2.\ is well-expanding this will guarantee us that it expands inside $V_i$. We would now like to use the fact that $U'$ expands inside $V_i$ to conclude the same happens for the original set $U$ from condition 2.\ above containing $U'$. However, a major issue here is that in order to achieve this we would need a bound on $|(N_{G-F}(U')\setminus U)\cap V_i|$ to get the expansion for \emph{any} relevant set $U$ for which we use the well-expanding subset $U'$ -- having to bound a random variable depending on $U'$ and $U$ spoils our union bound approach over the smaller sets $U'$. For the first expansion of $U$ into $V_i$ this is avoidable (by only considering such sets disjoint from $V_i$), but we need to expand multiple times to reach most of the vertices in $V_i$ and, after the first expansion, avoiding expanding sets that contain vertices in $V_i$ is unavoidable. To get around this, when we identify the well-expanding set $U'$, we look at its successive neighbourhoods in $G-F$ and show that enough of these vertices are chosen to be in $V_i$ together with all the vertices along a path going back to $U'$ so that $U'$ expands via such short paths to reach more than half of the vertices of $V_i$. In an early draft of this work, this was shown by carefully analysing an intricate random process. Fortunately, however, we will instead give here a much easier proof by combining our new perspective on robust sublinear expansion with an adaptation of a clever application of the sprinkling method appearing in a very recent work of Tomon~\cite{tomon2022robust}. We defer a more detailed sketch for this part of the argument to \Cref{sec:connectinexpander}, in particular until after we have introduced in full our new perspective on robust sublinear expansion, which remains crucial for this new approach. In total, though, this will allow us to answer question ii) positively. I.e., property~\ref{prop:sketch} holds if we are allowed to use all the edges of $G$ to make connections through $V_i$. Let us also stress an important point, which already played a role at various points in the above arguments and that is that the paths $P_{xy}$ we find will always be short, namely of length $\log^{O(1)} n$. This again plays an important role in answering the next question, namely finding an appropriate sparse subgraph $H_i\subseteq G$ with the same property, which we turn to next. \textbf{iii) Finding sparse connecting skeletons.} Before we look for a subgraph $H_i\subseteq G$ with the property \ref{prop:sketch} and $n\log^{O(1)}n$ edges, can we even find any graph with these properties? A binomial random graph is a natural candidate for such a graph, and, indeed, if $H$ is a binomial random graph with vertex set $V(G)$ and edge probability $p=\omega\left(\frac{\log n}n\right)$ then it will have, with high probability, the property \ref{prop:sketch} if we replace $H_i$ with $H$ and choose $V_i$ to be any fixed set of linear size. (We prove this as Lemma~\ref{lem:template}, with a larger than optimal value of $p$ for simplicity.) We then use $H$ as a \emph{template} to construct the sparse expanding skeleton $H_i$. We first sample our large random subset of vertices $V_i$, and use our answer to question ii) to guarantee that property~\ref{prop:sketch} holds with high probability in a slightly stronger form where every vertex is allowed to appear in $O(\log^5n)$ pairs in $\mathcal{P}$. In particular, we will use it with $\mathcal{P}$ being the set of pairs of vertices making an edge of our template graph $H$, which we choose to be sparse and well-connected through $V_i$, as well as have maximum degree $O(\log^5n)$. For each edge $xy\in E(H)$, we find an $xy$-path $P_{xy}$ through $V_i$ with length $\log^{O(1)}n$ in $G$ so that all these paths are edge disjoint. We then let $H_i$ be the union of all these paths, noting that, as $H$ is sparse and the paths $P_{xy}$ are relatively short, $H_i$ is also relatively sparse. Then, given an arbitrary collection $\mathcal{P}$ of pairs to connect, we first find edge disjoint paths connecting them through $V_i$ in $H$, before replacing each edge $xy$ on one of these paths in $H$ with the corresponding path $P_{xy}$ through $V_i$. This creates a set of edge disjoint $xy$-\emph{walks} through $V_i$ in $H_i$ --- as each such walk contains an $xy$-path, we can find the paths required by property~\ref{prop:sketch}. Note that, when we do this for each $i\in [3]$, we need to ensure that the graphs $H_i$ we find are edge disjoint, but this is easy to do by reusing some of our previous work, splitting $G$ into a union of edge disjoint expanders $G_1,G_2,G_3$, before finding each subgraph $H_i$ in the respective subgraph $G_i$. \subsection{Organisation of the paper}\label{sec:organisation} In the rest of this section we will introduce some general preliminary results, including some concentration results in \Cref{sec:conc} and the Aharoni-Haxell hypergraph matching theorem in \Cref{subsec:AH-matching-thm}, before showing strongly expanding graphs (namely $\mathcal{G}(n,p)$) satisfy a certain strong connectivity property in \Cref{sec:template}. In \Cref{sec:expansion}, we introduce robust sublinear expansion and prove a number of useful properties of this type of expansion. In \Cref{sec:connectinexpander}, we establish that our weaker expansion implies a similar (though weaker) connectivity property as that used in \Cref{sec:template}. In \Cref{sec:logstarproof}, we use the machinery we developed to prove our main result, \Cref{thm:logstar}. Finally, in \Cref{sec:final}, we make some concluding remarks. \subsection{Concentration inequalities}\label{sec:conc} We will often use a basic version of Chernoff's inequality for the binomial random variable (see, for example, \cite{alon-spencer}). \begin{theorem}[Chernoff's bound]\label{chernoff} Let $n$ be an integer and $0\le \delta,p \le 1$. If $X \sim \bin(n,p),$ then, setting $\mu=\mathbb{E} X = np,$ we have $$\P(X>(1+\delta) \mu) \le e^{-\delta^2\mu/2},\quad\quad\quad \text{ and }\quad\quad\quad \P(X<(1-\delta) \mu) \le e^{-\delta^2\mu/3}.$$ \end{theorem} We will also make use of the following well-known martingale concentration result (see Chapter 7 of \cite{alon-spencer}). \begin{lemma}\label{lem:mcd} Suppose that $X:\prod_{i=1}^N\Omega_i\to \mathbb{R}$ is $k$-Lipschitz. Then, for each $t>0$, \[ \mathbb{P}(|X-\mathbb{E} X|>t)\leq 2\exp\left(\frac{-2t^2}{k^2N}\right). \] \end{lemma} \subsection{The Aharoni-Haxell hypergraph matching theorem and edge disjoint paths}\label{subsec:AH-matching-thm} We will use the following hypergraph version of Hall's theorem due to Aharoni and Haxell, which is an immediate consequence of Corollary~1.2 in \cite{aharoni2000hall} (noting that we can add new, unique, vertices to each edge in the theorem to make the hypergraphs $\ell$-uniform). A \emph{matching} in a hypergraph is a collection of pairwise vertex disjoint edges. \begin{theorem}\label{thm:hyperhall} Let $r\in \mathbb{N}$, and let $H_1,\ldots,H_r$ be a collection of hypergraphs with at most $\ell$ vertices in each edge. Suppose that, for each $I\subseteq [r]$, there is a matching in $\bigcup_{i\in I}H_i$ containing more than $\ell(|I|-1)$ edges. Then, there is an injective function $f:[r]\to \bigcup_{i\in [r]}E(H_i)$ such that $f(i)\in E(H_i)$ for each $i\in [r]$ and $\{f(i):i\in [r]\}$ is a matching of $r$ edges. \end{theorem} We will use \Cref{thm:hyperhall} to find edge disjoint paths between vertex pairs, to show that a graph is well-connected under the following definition, recalling that a path \emph{through} $V$ is a path with all its internal vertices in $V$. \begin{defn}\label{defn:path-connected} A graph $G$ is $(\ell,t)$-\emph{path connected} through a vertex subset $V \subseteq V(G)$ if, for any $\mathcal{P} \subseteq \binom{V(G)}{2}$ in which every vertex appears in at most $t$ pairs in $\mathcal{P}$, there are edge disjoint paths $P_{\{x,y\}}$, $\{x,y\}\in \mathcal{P}$, such that, for each $\{x,y\}\in \mathcal{P}$, $P_{\{x,y\}}$ is an $xy$-path through $V$ with length at most $\ell$. \end{defn} We denote by $\binom{V(G)}{2}$ the multiset of pairs of distinct vertices of $G$, so in particular the same pair may appear multiple times in the collection $\mathcal{P}$. Typically, $t$ will be a small constant and $\ell$ will be at most polylogarithmic in the number of vertices. Given a graph $G$, a vertex set $V\subseteq V(G)$ and a collection $\mathcal{P}\subseteq \binom{V(G)}{2}$, we translate the pair connectivity property into a hypergraph matching problem as follows. For some $\ell\in \mathbb{N}$, and each $\{x,y\}\in \mathcal{P}$, let $H_{\{x,y\}}$ be the hypergraph with vertex set $E(G)$ and add as an edge the set $E(P)$ for each $xy$-path $P$ in $G$ with interior vertices in $V$ and length at most $\ell$. If there is an injective function $f:\mathcal{P}\to \bigcup_{\{x,y\}\in \mathcal{P}}E(H_{\{x,y\}})$ such that $f(\{x,y\})\in E(H_{\{x,y\}})$ for each $\{x,y\}\in \mathcal{P}$ and $\{f(\{x,y\}):\{x,y\}\in \mathcal{P}\}$ is a matching, then, for each $\{x,y\}\in \mathcal{P}$, let $P_{\{x,y\}}$ be the path in $G$ with edge set $f(\{x,y\})$. By the definition of $H_{\{x,y\}}$, each path $P_{\{x,y\}}$ is an $xy$-path in $G$ with length at most $\ell$ and interior vertices in $V$, and, as $\{f(\{x,y\}):\{x,y\}\in \mathcal{P}\}$ is a matching, these paths are all edge disjoint. Therefore, in order to prove that $G$ is $(\ell,t)$-{path connected} through $V$, it suffices to take a general such collection $\mathcal{P}$, define the relevant hypergraphs $H_{\{x,y\}}$, and prove that the associated condition holds for an application of Theorem~\ref{thm:hyperhall}. \subsection{Existence of well-connected sparse graphs}\label{sec:template} Random graphs typically present a natural candidate for a sparse, well-connected, graph, and we use this to prove the existence of our template in Lemma~\ref{lem:template} (using the connectivity property in \Cref{defn:path-connected}). Similar properties have been studied before in random graphs for various applications (see e.g.\ \cite{broder1996efficient,haxell2001tree,glebov2013hamilton}), and our main lemma (Lemma~\ref{lem:template}) can be obtained as a (not quite immediate) corollary of Lemma 3.4.\ from \cite{montgomery2021spanning} combined with a multi-round exposure argument like the one we use below. We include a different proof of Lemma~\ref{lem:template} for completeness but also use this to introduce two intermediate results (\Cref{prop:connectonepair} and \Cref{lem:connectmanypairs}) that we later use in the same manner to prove our key technical result, \Cref{thm:pathconnect}. We first remind the reader that a path through a subset of vertices $V$ is a path whose internal vertices are all in $V$. Let us also introduce, given $U,V \subseteq V(G)$, the \emph{ball of radius $i$ around $U$ within $V$} which we will denote as $B^{i}_G(U,V)$, namely it is the set of vertices in $V$ which can be reached by a path through $V$ of length at most $i$ starting from a vertex in $U$. The starting vertex in $U$ is \emph{not} required to be in $V$ itself, as is usual with our definition of paths through a set. We do, however, only consider reachable vertices within $V$, so that $B^{i}_G(U,V) \subseteq V$. For \Cref{prop:connectonepair}, we take an expansion property of sets of size $t\in \mathbb{N}$ and use this to connect a pair of vertices from a set of $2t-1$ pairs (c.f.\ the collection $\mathcal{P}$ in \Cref{defn:path-connected}). \begin{prop}\label{prop:connectonepair} Let $1\leq \ell,t \leq n$. Let $G$ be an $n$-vertex graph and let $V\subseteq V(G)$ be of size $|V|\geq 4t-2$ such that, for every $U \subseteq V(G)$ with size $|U|=t$, we have $|B^\ell_{G}(U,V)|>\frac{|V|}2$. Let $x_1,\ldots,x_{2t-1},y_1,\ldots,y_{2t-1}$ be distinct vertices of $G$. Then, for some $j\in [2t-1]$, there is an $x_jy_j$-path in $G$ through $V$ with length at most $4\ell \log n$. \end{prop} \begin{proof} Let $I_x$ be the set of $i\in [2t-1]$ for which $|B^{2\ell\log n}_G(x_i,V)|\leq \frac{|V|}2$ and let $I_y$ be the set of $i\in [2t-1]$ for which $|B^{2\ell\log n}_G(y_i,V)|\leq \frac{|V|}2$. Note that the required path can be found if there exists some $j\in [2t-1]$ with $j\notin I_x$ and $j\notin I_y$, for then $B^{2\ell\log n}_{G}(x_j,V)$ and $B^{2\ell\log n}_{G}(y_j,V)$ each have size larger than $\frac{|V|}2$ and must therefore intersect. Thus, we can assume that there is no such $j$, and, therefore, without loss of generality, that $|I_x|\geq t$. Let $r\geq 0$ be the largest integer for which there is a set $X\subseteq \{x_i:i\in I_x\}$ for which $|X|\leq t\left(\frac23\right)^r$ and $|B^{(r+1)\ell}_G(X,V)|>\frac{|V|}2$, and let $X$ be any such set. Note that this is possible as $|I_x|\geq t$ and any subset of $\{x_i:i\in I_x\}$ with size $t$ satisfies these conditions for $r=0$ by the assumption of the proposition. Now, as $X\neq \emptyset$, we have that $t\left(\frac23\right)^r\geq 1$, so that $r\leq 2\log t\leq 2\log \frac{n}2$ (using $\left(\frac23\right)^2<\frac12$ and $n\geq |V|\geq 4t-2\geq 2t$) and thus $r< (2\log n)-1$. Thus, by the definition of $I_x$, $|B^{(r+1)\ell}_G(x_i,V)|\leq |B^{2\ell\log n}_G(x_i,V)|\leq \frac{|V|}2$ for each $i\in I_x$, and hence $|X|\geq 2$. This allows us to partition $X=X_0\cup X_1$ with $|X_0|,|X_1|\leq \frac23 |X|\leq t\left(\frac23\right)^{r+1}$. As \[ |B^{(r+1)\ell}_G(X_0,V)|+|B^{(r+1)\ell}_G(X_1,V)|\geq |B^{(r+1)\ell}_G(X_0\cup X_1,V)|=|B^{(r+1)\ell}_G(X,V)|>\frac{|V|}2\geq 2t-1, \] we can pick $j\in [2]$ such that $|B^{(r+1)\ell}_G(X_j,V)|\geq t$. Therefore, using the expansion of $t$-sets into $V$, we have $|B^{(r+2)\ell}_G(X_j,V)|>\frac{|V|}2$, contradicting the maximality of $r$ as $|X_j|\leq t\left(\frac23\right)^r$. \end{proof} We now take a stronger expansion property and use \Cref{prop:connectonepair} in combination with the Aharoni-Haxell hypergraph matching theorem (\Cref{thm:hyperhall}) to find edge disjoint paths connecting a set of vertex pairs (see also the discussion after \Cref{thm:hyperhall}), proving \Cref{lem:connectmanypairs}. \begin{lemma}\label{lem:connectmanypairs} Let $n\geq 2$ and $1\leq \ell,k \leq n$. Let $G$ be an $n$-vertex graph and let $V\subseteq V(G)$ so that $|V|\ge \frac n8+1$ and suppose, for each $U\subseteq V(G)$ and $F\subseteq E(G)$ with $U\neq\emptyset$ and $|F|\leq 2^9k|U|(\ell\log n)^2$, we have $|B^\ell_{G-F}(U,V)|>\frac{|V|}2$. Then, $G$ is $(4\ell\log n,k)$-path connected through $V$. \end{lemma} \begin{proof} To show that $G$ is $(4\ell\log n,k)$-path connected through $V$, let $\mathcal{P}\subseteq \binom{V(G)}{2}$ be an arbitrary collection of vertex pairs such that each vertex appears in at most $k$ pairs in $\mathcal{P}$. Let $r=|\mathcal{P}|$, and order the pairs in $\mathcal{P}$. For each $i\in [r]$, let $H_i$ be the hypergraph with vertex set $E(G)$ and edge set corresponding to the edge sets of paths through $V$ with length at most $4\ell\log n$ connecting the $i$-th pair of vertices in $\mathcal{P}$, noting that the size of any edge is at most $4\ell\log n.$ Now, for each $I\subseteq [r]$, let $M_I$ be a maximal matching in $\bigcup_{i\in I}H_i$. We will show that $|M_I|\geq 4\ell \log n \cdot |I|$ for each $I\subseteq [r]$. Towards a contradiction, suppose that, for some $I\subseteq [r]$, we have $|M_I|<4\ell \log n \cdot |I|$, noting that we must have $I\neq \emptyset$. Let $F$ be the set of edges of $G$ in any path corresponding to an edge in $M_I$, so that $F=\bigcup_{e\in M_I}V(e)$. Note that $|F|< 4\ell\log n \cdot 4\ell \log n \cdot |I| = (4\ell \log n)^2 |I|$. Let $I'$ be a maximal subset of $I$ such that no vertex appears in a pair in $I'$ more than once, so that $2k|I'|\geq |I|$. Note also that this ensures that $|I'|\le \frac n2$. Now, let $t=\ceil{|I'|/16}\geq 1$, so that $2t-1 \le 16t-15\le |I'|\le 16t$ and $4t-2 \le \ceil{\frac n8}+1\le|V|$. Note that $|F|< (4\ell\log n)^2|I|\leq k\frac{|I'|}2(8\ell\log n)^2\leq 8kt(8\ell\log n)^2$, and therefore any set $U\subseteq V(G)$ with size $t$ satisfies $|B^\ell_{G-F}(U,V)|>\frac{|V|}2$. Then, by Proposition~\ref{prop:connectonepair} applied to $G-F$, for some $j\in I'$ there is a path in $G-F$ between the $j$-th pair in $\mathcal{P}$ with interior vertices in $V$ and length at most $4\ell \log n$. Such a path corresponds to an edge of $H_j$ with no vertices in $F$, a contradiction to the maximality of $M_I$. Thus, we must have $|M_I|\ge 4\ell\log n \cdot |I|$. Therefore, by \Cref{thm:hyperhall}, there is a set of paths $P_i$, $i\in [r]$, in $G$ with $E(P_i)\in H_i$ for each $i\in [r]$, such that $E(P_i)$, $i\in [r]$, form edge disjoint sets. Thus, $G$ is $(4\ell\log n,k)$-path connected through $V$. \end{proof} We now prove the existence of our template graph, by showing an appropriate expansion condition is likely in a certain binomial random graph and applying \Cref{lem:connectmanypairs}. \begin{lemma}\label{lem:template} For any large enough $n$, there exists an $n$-vertex graph $G$ with $\Delta(G)\leq 2^{8}\log^5 n$ and a set $V\subseteq V(G)$ with $|V|=\frac n6$ such that $G$ is $\left(\frac14\log^2 n,2\right)$-path connected through $V$. \end{lemma} \begin{proof} Let $p=\frac{150\log^5 n}{n}$, let $V\subseteq [n]$ be a set of size $\frac n6$ and let $G \sim \mathcal{G}\left(n,p\right)$. As $p=\omega\left(\frac{\log n}{n}\right)$, a standard application of Chernoff's inequality (\Cref{chernoff}) shows that, with high probability, $\Delta(G)\leq \frac32pn= 225\log^5 n$. Therefore, it is sufficient to show that, with high probability, $G$ is $\left(\frac14\log^2 n,2\right)$-path connected through $V$. \begin{claim* With high probability, for each $U\subseteq V(G)$ and each $F\subseteq E(G)$ with $U\neq\emptyset$ and $|F|\leq 4|U|\log^4 n$ we have \begin{equation}\label{eqn:U} |B_{G-F}(U,V)| > \min\left\{2^{16}|U|,\frac {|V|}{2}\right\}. \end{equation} \end{claim*} \begin{proof} For large $n$, we will show for each $U\subseteq V(G)$ and $F\subseteq E(G)$ with $U\neq\emptyset$ and $|F|\leq 4|U|\log^4n$ that \eqref{eqn:U} holds with probability at least $1-2^{-10|U|\log^5n}$, so that \eqref{eqn:U} holds for all $U\subseteq V(G)$ and $F\subseteq E(G)$ with $|F|\leq 4|U|\log^4n$ with probability at least \[ 1-\sum_{u=1}^{n} \binom{n}{u}\sum_{f=0}^{4u\log^4n}\binom{\binom{n}{2}}{f} \cdot 2^{-10u\log^5 n}\geq 1-\sum_{u=1}^nn^u\cdot n^2\cdot n^{8u\log^4n}\cdot 2^{-10u\log^5 n}\geq 1-\sum_{u=1}^nn^{-2}= 1-n^{-1}, \] and the claim holds. Let then $U\subseteq V(G)$ with $u=|U|\geq 1$ and $F\subseteq E(G)$ with $f=|F|\leq 4u\log^4 n$. Note that if $|B_{G-F}(U,V)| \leq \nolinebreak \min\left\{2^{16}u,\frac {|V|}{2}\right\}$, then there is some set $X=B_{G-F}(U,V)$ with size at most $2^{16}u$ such that $|V\setminus X|\geq \frac{|V|}2$ and there are no edges between $U$ and $V\setminus X$ other than in $F$. The probability of such a set $X$ existing is at most \[ \sum_{i=0}^{2^{16}u}\binom{n}{i}\cdot (1-p)^{u\cdot \frac{|V|}{2}-|F|}\leq n^{2^{16}u+1}\cdot e^{-p(un/12-f)}\leq e^{-pun/15}\leq 2^{-10u\log^5n}, \] as required. \end{proof} With high probability then, we have that the conclusion of the claim holds for $G$. For any $U\subseteq V(G)$ and $F\subseteq E(G)$ with $U\neq\emptyset$ and $|F| \le 4|U| \log^4 n$, we have then by induction that, for each $i\geq 0$, \[ |B_{G-F}^i(U,V)| > \min\left\{2^{16i}|U|,\frac{|V|}2\right\}. \] Setting $i=\ell=\frac{\log n}{16}$, we thus have $|B^{\ell}_{G-F}(U) \cap V|> \frac{|V|}2.$ Thus, by \Cref{lem:connectmanypairs} with $\ell=\frac{\log n}{16}$ and $k=2$, we have that $G$ is $\left(\frac14\log^2 n,2\right)$-connected through $V$. \end{proof} \section{Robust sublinear expansion}\label{sec:expansion} In this section we will explain the expansion we use and its background, our new perspective on this type of expansion, and prove some key results using the expansion. Before doing so, for convenience we state the definition of expansion that we use, as follows. \begin{defn} \label{defn:robust-sublinear-expansion} An $n$-vertex graph $G$ is an $(\varepsilon,s)$-expander if, for every $U\subseteq V(G)$ and $F\subseteq E(G)$ with $1\le |U|\leq \frac23n$ and $|F|\leq s|U|$, we have \begin{equation} |N_{G-F}(U)|\geq \frac{\varepsilon|U|}{\log^2 n}.\label{eqn:expands} \end{equation} \end{defn} As the bound on the size of the neighbourhood guaranteed at \eqref{eqn:expands} is $o(|U|)$ as $n\to \infty$, we consider this to be \emph{sublinear} expansion. We often use Definition~\ref{defn:robust-sublinear-expansion} when $s$ is polylogarithmic in $n$, so that the set of edges $F$ may be of size $\omega(|U|)$ as $n\to\infty$, and call this \emph{robust} sublinear expansion. This is discussed in more detail with the relevant background in Section~\ref{sec:generalexpansionchat}. In Section~\ref{subsec:alternative}, we then introduce an alternative perspective of this expansion. In \Cref{sec:decomp-into-expanders}, we prove a lemma which almost decomposes an arbitrary graph into expanders. Finally, in \Cref{sec:edge-partition-of-expanders} we prove sublinear expanders can be (edge) partitioned into expanders (with only slightly weaker expansion parameters). \subsection{Expansion}\label{sec:generalexpansionchat} Classical graph expansion is an immensely powerful idea in graph theory and computer science that has seen a very wide variety of applications (see for example the survey \cite{expander-survey}). A typical such property in a graph $G$ says that $|N_G(U)|\geq \lambda|U|$ for any set $U\subseteq V(G)$ which is not too large, where $\lambda\geq 2$, though other notions have been considered instead of requiring a large $|N_G(U)|$, for example bounding the number of edges in $G$ between $U$ and $V(G)\setminus U$ (as indeed used by Conlon, Fox and Sudakov~\cite{conlon2014cycle}). Sublinear expansion is a weaker notion of this classical expansion introduced by Koml\'os and Szemer\'edi~\cite{K-Sz-1,K-Sz-2}, where we take a much smaller value of $\lambda$, but which is significant as every graph contains a sublinear expander $H$ with $\lambda=\Theta(1/\log^2|H|)$ (and even has a nice decomposition into sublinear expanders, as we will prove and use). Koml\'os and Szemer\'edi used sublinear expansion to find minors in sparse graphs, and more recently sublinear expansion has found a host of other applications (see, for example, \cite{shapira2015small,montgomery2015logarithmically,fernandez2022nested,fernandez2022build,chakraborti2021well,haslegrave2021crux,haslegrave2021ramsey,kim2017komlos,liu2017mader,liu2020solution,liu2020clique,letzter2023immersion}). Such sublinear expansion in a graph $G$ has some very weak \emph{robustness} properties, in that if $\lambda|U|/2$ vertices in $V(G)\setminus U$ are removed from the graph then the set $U$ will still expand (with the neighbourhood of $U$ still having at least $\lambda|U|/2$ vertices), and this property is used in many of the applications of sublinear expansion cited above. However, we will distinguish \emph{robust sublinear expansion} to be that where $U$ expands despite the removal of any set $F$ of at most $s|U|$ edges in $G$, where $s$ grows with $|G|$ so that this bound is superlinear in $|U|$, as in \Cref{defn:robust-sublinear-expansion}. Such robust sublinear expansion has recently been developed essentially independently by groups of different authors, appearing in some form in work by Haslegrave, Kim, and Liu~\cite{haslegrave2021extremal} and by Sudakov and Tomon~\cite{sudakov2022extremal} (with the parallel clearer in the expansion used in subsequent work by Jiang, Methuku and Yepremyan~\cite{jiang2021rainbow} and by Tomon~\cite{tomon2022robust}). Roughly speaking, the expansion we use, as given in Definition~\ref{defn:robust-sublinear-expansion}, is a slightly weaker version of that used by Haslegrave, Kim, and Liu~\cite{haslegrave2021extremal} (so that we can find an almost-decomposition into such expanders) and a stronger version than subsequent developments of the expansion used by Sudakov and Tomon~\cite{sudakov2022extremal} (which makes it more powerful when we use it). \subsection{An alternative notion of robustness}\label{subsec:alternative} In \Cref{defn:robust-sublinear-expansion}, we consider the expansion to be robust as sets expand despite an arbitrary removal of a small number of edges. This can be alternatively encoded by recording that every vertex subset $U$ either expands very well (by some factor greater than $1$) or that its `robust neighbourhood' of vertices with plenty of edges towards $U$ expands well, though perhaps sublinearly (as in Proposition~\ref{prop:expansion-red-blue} below). It will be convenient to define this robust neighbourhood for any parameter $d$ as $$N_{G,d}(U):=\{v \in V(G) \setminus U : |N_G(v) \cap U| \ge d\},$$ that is, the set of vertices in a graph $G$, outside of a subset of vertices $U$, which have degree at least $d$ towards $U$. \begin{prop}\label{prop:expansion-red-blue} Let $G$ be an $n$-vertex $(\varepsilon,s)$-expander, $U\subseteq V(G), |U|\le \frac23 n$ and $F$ a set of at most $s|U|/2$ edges. Then, for any $0<d \le s$, either \[ \text{\textbf{\emph{a)}} }\;\;|N_{G-F}(U)| \ge \frac{s|U|}{2d},\;\;\;\text{ or }\;\;\;\text{\textbf{\emph{b)}}}\;\; |N_{G-F,d}(U)|\ge \frac{\varepsilon|U|}{\log^{2} n}. \] \end{prop} \begin{proof} Suppose \textbf{a)} is not satisfied, so that $|N_{G-F}(U)|<\frac{s|U|}{2d}$. Let $X=N_{G-F}(U)\setminus N_{G-F,d}(U)$, so that $|X|<\frac{s|U|}{2d}$. Let $F'$ be the edges of $G-F$ between $U$ and $X$, so that $|F'| < |X|d\leq s|U|/2$, and hence $|F|+|F'|\leq s|U|$. Note that, by the definition of $F'$, we have $N_{G-F,d}(U)=N_{G-F-F'}(U)$. As $G$ is an $(\varepsilon,s)$-expander, we thus have \[ |N_{G-F,d}(U)|=|N_{G-F-F'}(U)|\geq \frac{\varepsilon|U|}{\log^{2} n}, \] and therefore \textbf{b)} holds, as required. \end{proof} The following proposition shows more structure can be found in both outcomes of the above proposition. Though a more general variant follows easily, the parameters are tailored for our intended application. \begin{prop}\label{prop:red-blue-expansion-robust-both} There is an $n_0$ such that the following holds for each $n\geq n_0$, $\varepsilon\ge2^{-9}$ and $s \ge 8\log^{13} n$. Let $G$ be an $n$-vertex $(\varepsilon,s)$-expander, let $U\subseteq V(G)$ have size $|U|\le \frac23 n$ and let $F$ be a set of at most ${s|U|}/4$ edges. Then, in $G-F$ we can find either \begin{enumerate}[label = \emph{\textbf{\alph{enumi})}}] \item $\frac{|U|}{\log^{7} n}$ vertex disjoint stars, each with $\log^{9} n$ leaves, centre in $U$ and all its leaves in $V(G)\setminus U$, or \item a bipartite subgraph $H$ with vertex classes $U$ and $X\subseteq V(G) \setminus U$ such that \begin{itemize} \item $|X| \ge \frac{\varepsilon|U|}{2\log^2 n}$ and \item every vertex in $X$ has degree at least $\log^{4} n$ in $H$ and every vertex of $U$ has degree at most $2\log^{9} n$ in $H.$ \end{itemize} \end{enumerate} \end{prop} \begin{proof} Take a maximal collection of vertex disjoint stars in $G-F$ with $\log^{9} n$ leaves and centre in $U$ and leaves outside of $U$. Let $C \subseteq U$ be the set of centres of these star and $L\subseteq V(G)\setminus U$ be the set consisting of all their leaves. Assuming \textbf{a)} does not hold, we can thus assume that $|C| \le \frac{|U|}{\log^{7} n}$, $|L| \le |U| \log^2 n$, and, by the maximality, that there is no vertex in $U\setminus C$ with at least $\log^9 n$ neighbours in $G-F$ in $V(G)\setminus (U\cup L)$. Thus, \begin{equation}\label{eqn:NGW} |N_{G-F}(U\setminus C)|\leq |C|+|L|+|U\setminus C|\cdot \log^9n\leq \frac{|U|}{\log^{7} n}+|U| \log^2 n+|U|\log^9n< 2|U|\log^9n. \end{equation} Let $d=\log^4n$ and $\Delta=2\log^{9} n$. We now construct the set $X\subseteq V(G)\setminus U$ and the bipartite subgraph $H$ through the following process, starting with $X_0=\emptyset$ and setting $H_0$ to be the graph with vertex set $U\cup X_0$ and no edges. Let $r=|V(G)\setminus U|$ and label the vertices of $V(G)\setminus U$ arbitrarily as $v_1,\ldots,v_r$. For each $i\geq 1$, if possible pick a star $S_i$ in $G-F$ with centre $v_i$ and $d$ leaves in $U$ such that $H_{i-1}\cup S_i$ has maximum degree at most $\Delta$, and let $H_i=H_{i-1}\cup S_i$ and $X_i=X_{i-1} \cup \{v_i\}$, while otherwise we set $H_i=H_{i-1}$ and $X_{i}=X_{i-1}$. Finally, let $H=H_r$ and $X=X_r=V(H_r)\setminus U$. We will show that \textbf{b)} holds for this choice of $H$ with bipartition $(U,X)$. Firstly, observe that $\Delta(H_i)\leq \Delta$ for each $i\in [r]$ by construction, and that every vertex $v_i$ in $X$ has degree exactly $d$ in $H$, so the second condition in \textbf{b)} holds. Thus, we only need to show that $|X| \ge \frac{\varepsilon|U|}{2\log^2 n}$ holds, which will follow as no vertex in $U\setminus C$ has $\frac{\Delta}2=\log^9n$ neighbours in $G-F$ in $X\setminus L$ due to the maximality of our family of disjoint stars defining $C$ and $L$. Indeed, let $U'$ be the set of vertices in $U \setminus C$ with degree exactly $\Delta$ in $H$. As each vertex in $U\setminus C$ has fewer than $\frac{\Delta}2$ neighbours in $G-F$ in $X\setminus L$, it must have at least $\frac{\Delta}2$ neighbours in $H$ in $X\cap L$. As each vertex in $X\cap L$ has $d$ neighbours in $H$, we have \[ |U'|\leq \frac{d|X \cap L|}{\Delta/2}\le \frac{2d|L|}{\Delta}\leq \frac{2d\cdot |U|\log^2n}{\Delta}=\frac{2\log^4 n\cdot |U|\log^2n}{2\log^9n}\le \frac{\varepsilon|U|}{8\log^2 n}, \] where the last inequality follows for sufficiently large $n$. Let $B=C \cup U'$, so that \[ |B|\leq \frac{|U|}{\log^{7} n}+\frac{\varepsilon|U|}{8\log^2 n}\le \frac{\varepsilon|U|}{4\log^2 n}, \] and, in particular, $|U\setminus B|\geq \frac{|U|}2$. Then, by \Cref{prop:expansion-red-blue} applied to $U\setminus B$ and $F$ with $d$, using that $|F| \le s|U|/4 \le s|U \setminus B|/2$, we have either $|N_{G-F}(U\setminus B)|\geq \frac{s|U\setminus B|}{2d}$ or $|N_{G-F,d}(U\setminus B)|\geq \frac{\varepsilon |U\setminus B|}{\log^2 n}$. As $$\frac{s|U\setminus B|}{2d}\geq \frac{s|U|}{4d}\geq 2|U|\log^9 n,$$ the former contradicts \eqref{eqn:NGW}, and therefore we must have that $|N_{G-F,d}(U\setminus B)|\geq \frac{\varepsilon |U\setminus B|}{\log^2n}$. Every vertex $v_i$ in $N_{G-F,d}(U\setminus B)$ has at least $d$ neighbours in $G-F$ in $U\setminus B$ which, not being in $B=U' \cup C$, by definition of $U'$ must all have degree strictly less than $\Delta$ in $H$. This implies $v_i \in X$, since we could add it together with some $d$ of these neighbours. Hence, we must have $N_{G-F,d}(U\setminus B)\subseteq X$, and \[ |X|\geq |N_{G-F,d}(U\setminus B)|\geq \frac{\varepsilon |U\setminus B|}{\log^2 n}\geq \frac{\varepsilon|U|}{2\log^2 n}, \] as required. \end{proof} \subsection{Almost decomposing an arbitrary graph into expanders}\label{sec:decomp-into-expanders} The following lemma almost decomposes a graph into robust sublinear expanders with, on average, very little overlap between their vertex sets. Setting $s=0$ in the below lemma (as we do in one application) obtains a full decomposition, although without any robustness. \begin{lemma}\label{splitting-into-expanders} Given an $n$-vertex graph $G$, a non-negative integer $s$ and $\varepsilon \le 2^{-5}$ we can delete up to $4sn \log n$ edges from $G$ so that the remaining edges may be partitioned into graphs $G_1,\ldots, G_r$ such that $\sum_{i=1}^r |G_i| \le 2n$ and each $G_i$ is an $(\varepsilon,s)$-expander. \end{lemma} \begin{proof} We prove this by induction on $n$, under the stronger condition that the graphs $G_1,\ldots,G_r$ in the partition satisfy $\sum_{i=1}^r |G_i| \le 2n-\frac{2n}{2+\log n}$. Since $2n-\frac{2n}{2+\log n}\ge n,$ and any $1$-vertex graph $G$ is trivially an $(\varepsilon,s)$-expander, the lemma holds for $n=1$ with $G_1=G$. Let us then assume $n \ge 2$ and that the claim holds for all graphs with at most $n-1$ vertices. Letting $G$ be an $n$-vertex graph, note that, as $2n-\frac{2n}{2+\log n}\ge n,$ if $G$ is an $(\varepsilon,s)$-expander then the trivial partition of $G_1=G$ demonstrates the claim holds for $G$. Thus, we can assume $G$ is not an $(\varepsilon,s)$-expander, and in particular that there exists a non-empty set of vertices $U \subseteq V(G)$ with $|U| \le \frac23n$ and a set $F$ of at most $s|U|$ edges such that $|N_{G-F}(U)| < \frac{\varepsilon |U|}{\log^2 n}.$ Let $G_1=G[U\cup N_{G-F}(U)]-F$ and let $G_2=G\setminus U-E(G_1)-F$, so that $G_1$ and $G_2$ form an edge partition of $G-F$ and, setting $n_1=|G_1|$ and $n_2=G_2$, we have \begin{equation}\label{eqn:n1n2} n_1+n_2=|G_1|+|G_2|=|G|+|N_{G-F}(U)| < n+\frac{\varepsilon |U|}{\log^2 n}\le n+\frac{\varepsilon n_1}{\log^2 n}. \end{equation} Now, $n_2 = n-|U|<n$ and \begin{equation}\label{eqn:n1} n_1\leq |U|+\frac{\varepsilon |U|}{\log^2 n}\leq \frac23n+\varepsilon n\leq \frac34n<n, \end{equation} so there exist sets $E_1\subseteq E(G_1)$ and $E_2\subseteq E(G_2)$ and partitions $G_{1,1},\ldots,G_{1,r_1}$ and $G_{2,1},\ldots,G_{2,r_2}$ of $G_1-E_1$ and $G_2-E_2$ into edge disjoint $(\varepsilon,s)$-expanders so that, for each $i\in [2]$, $|E_i|\leq 4sn_i\log n_i$, and \[ \sum_{j=1}^{r_i}|G_{i,j}|\leq 2n_i-\frac{2n_i}{2+\log n_i}. \] Therefore, we can remove $F\cup E_1\cup E_2$ from $G$ and decompose the remaining edges into $(\varepsilon,s)$-expanders $G_{1,1},\ldots,G_{1,r_1}$, $G_{2,1},\ldots,G_{2,r_2}$. We need then only check that $|F\cup E_1\cup E_2|\leq 4sn\log n$ and that the sum of the vertices of the expanders in this decomposition is at most $2n-\frac{2n}{2+\log n}$. Firstly, note that, from \eqref{eqn:n1}, we have $\log n_1\leq \log \frac34n<\log n-\frac 25$, so that \begin{align} \frac{1}{s}(|F|+|E_1|+|E_2|)&\leq |U|+4n_1\log n_1+4n_2\log n_2 \leq n_1+4n_1\left(\log n-\frac{2}{5}\right)+4n_2\log n \nonumber\\ & = 4(n_1+n_2)\log n -\frac{3}{5}n_1 \overset{\eqref{eqn:n1n2}}{\leq} 4\left(n+\frac{\varepsilon n_1}{\log^2 n}\right)\log n -\frac{3}{5}n_1 \nonumber\\ & \leq 4n\log n.\label{eqn:EEE} \end{align} Secondly, again as $ \log \frac34n<\log n-\frac 25$, we have \begin{equation}\label{eqn:n1logs} \frac{2n_1}{2+\log n_1} \overset{\eqref{eqn:n1}}{\ge} \frac{2n_1}{2+\log (3n/4)} \ge \frac{2n_1}{8/5+\log n}=\frac{2n_1}{2+\log n}+\frac{2n_1\cdot 2/5}{(8/5+\log n)(2+\log n)} > \frac{2n_1}{2+\log n}+\frac{n_1}{10\log^2 n}, \end{equation} so that \begin{align*} \sum_{i=1}^2\sum_{j=1}^{r_i}|G_{i,j}|&\leq 2n_1+2n_2-\frac{2n_1}{2+\log n_1}-\frac{2n_2}{2+\log n_2}\overset{\eqref{eqn:n1logs}}{<} (n_1+n_2)\left(2-\frac{2}{2+\log n}\right)-\frac{n_1}{10\log^2 n} \\ &\overset{\eqref{eqn:n1n2}}{\leq} \left(n+\frac{\varepsilon n_1}{\log^2 n}\right)\left(2-\frac{2}{2+\log n}\right)-\frac{n_1}{10\log^2 n}\leq 2n-\frac{2n}{2+\log n}. \end{align*} In combination with \eqref{eqn:EEE}, this shows that $G$ has the required decomposition, completing the inductive step and hence the proof. \end{proof} \subsection{Decomposing an expander into many expanders}\label{sec:edge-partition-of-expanders} The following lemma partitions the edges of an expander into a chosen number of expanders with the same vertex set (and a slightly weaker expansion condition). \begin{lemma}\label{lem:partitionedgesintoexpanders} Let $n,k,s\in \mathbb{N}$ and $0<\varepsilon\leq 1$. Suppose that $G$ is an $n$-vertex $(\varepsilon,s)$-expander and $s\geq 2^{12}\varepsilon^{-1} k^2\log^4n$. Then, there are edge disjoint graphs $G_1,\ldots,G_k$ such that $E(G)=\bigcup_{i\in [k]}E(G_i)$ and, for each $i\in [k]$, $G_i$ is an $\left(\frac{\varepsilon}4,\frac{\sqrt{s\varepsilon}}{ 8k \log n}\right)$-expander with vertex set $V(G)$. \end{lemma} \begin{proof} If $n=1$ then the claim is trivially true so let us assume that $n \ge 2$. Furthermore, observe that as $\varepsilon>0$, we must have $s\leq \delta(G)$ for otherwise we can remove all the neighbours of a vertex with minimum degree by removing at most $s$ edges, contradicting that $G$ is an $(\varepsilon,s)$-expander, so certainly $s\leq n$. Let $H$ be a random subgraph of $G$ with vertex set $V(G)$ which contains every edge independently with probability $\frac 1k$. Then, assign every edge of $G$ to one of the graphs $G_1,\ldots, G_k$ uniformly and independently at random, so that each $G_i$ is a random subgraph with the same distribution as $H$. Letting $s'=\frac{\sqrt{s\varepsilon}}{ 8k \log n}\geq 8\log n$, we will show that the probability $H$ is not an $\left(\frac{\varepsilon}4,s'\right)$-expander is strictly less than $\frac 1k$. Thus, by a union bound, the probability that each $G_i$ is an $\left(\frac{\varepsilon}4,s'\right)$-expander is strictly positive, so some decomposition as required by the lemma must exist. To show that $H$ is not an $\left(\frac{\varepsilon}4,s'\right)$-expander with probability less then $\frac 1k$, we will take a union bound over all subsets $U$ of $V(G)$ for the event that $U$ fails the conditions of $(\frac{\varepsilon}{4},s')$-expansion in $H$. For this, set $d=\sqrt{s/\varepsilon} \log n$ and note that $\frac{s}{d}=8ks'$, $s'=\frac{\varepsilon d}{8k \log^2 n}$ and $s\geq \sqrt{s}\cdot \sqrt{\varepsilon^{-1}\log^2n}=d\geq 64 k$. Let $U\subseteq V(G)$ and $u=|U|\leq \frac{2n}{3}$. By \Cref{prop:expansion-red-blue} with $d$, $U$ and $F=\emptyset$, we have either \textbf{a)} $|N_G(U)|\geq \frac{su}{2d}=4ks'u$ or \textbf{b)} $|N_{G,d}(U)|\geq \frac{\varepsilon u}{\log^2 n}$. If \textbf{a)} holds, then $|N_H(U)|$ is dominated by $\text{Bin}(4ks'u,1/k)$, so that $\P(|N_H(U)|\geq 2s'u)\geq 1-e^{-s'u/2}$ by a Chernoff bound (\Cref{chernoff}). Note that if $|N_H(U)|\geq 2s'u$ then for any $F\subseteq E(H)$ with $|F|\leq s'u$ we have $|N_{H-F}(U)|\geq s'u\geq \frac{\varepsilon u}{4\log^2n}$. Thus, when \textbf{a)} holds, the probability $U$ fails the $\left(\frac{\varepsilon}4,s'\right)$-expansion condition is at most $e^{-s'u/2}\leq e^{-4u\log n}$ as $s'\geq 8\log n$. If \textbf{b)} holds, then note first that the probability that any vertex $v\in N_{G,d}(U)$ is not in $N_{H,d/2k}(U)$ is at most $p:=\P\left(\text{Bin}\left(d,\frac1k\right)< \frac{d}{2k}\right)$, where we have once again by a Chernoff bound that \begin{equation}\label{eqn:p} p\leq e^{-d/8k}. \end{equation} Therefore, as $|N_{G,d}(U)|\geq \frac{\varepsilon u}{\log^2 n}$, we have \begin{equation}\label{eqn:NHU} \P\left(|N_{H,d/2k}(U)|< \frac{\varepsilon u}{2\log^2n}\right)\le \binom{\frac{\varepsilon u}{\log^2 n}}{\frac{\varepsilon u}{2\log^2 n}}\cdot p^{\frac{\varepsilon u}{2\log^2 n}} \le 2^{\frac{\varepsilon u}{\log^2 n}}\cdot p^{\frac{\varepsilon u}{2\log^2 n}}\overset{\eqref{eqn:p}}{\le} e^{\frac{-\varepsilon u d}{32k\log^2 n}}= e^{-s'u/4}\le e^{-2u\log n}. \end{equation} Note that, if $|N_{H,d/2k}(U)|\geq \frac{\varepsilon u}{2\log^2n}$, then, for any $F\subseteq E(H)$ with $|F|\leq s'u$, we have \[ |N_{H-F}(U)|\geq |N_{H,d/2k}(U)|-\frac{|F|}{d/2k}\geq \frac{\varepsilon u}{2\log^2n}-\frac{2ks'u}{d}=\frac{\varepsilon u}{2\log^2n}-\frac{su}{4d^2}= \frac{\varepsilon u}{4\log^2n}. \] Thus, \eqref{eqn:NHU} implies that the probability $U$ does not satisfy the $(\varepsilon/4,s')$-expansion condition is at most $e^{-2u\log n}$. Therefore, whichever of \textbf{a)} or \textbf{b)} holds, the probability $U$ does not satisfy the $(\varepsilon/4,s')$-expansion condition is at most $e^{-2u\log n}$. Hence, the probability that $H$ is not an $(\varepsilon/4,s')$-expander is at most $$ \sum_{u=1}^{2n/3}\binom{n}{u}e^{-2u\log n} \le \sum_{u=1}^{2n/3} n^u\cdot n^{-2u} \le n^{-1}+\sum_{u=2}^{2n/3} n^{-2}\leq \frac{2}{n}< \frac{1}{k},$$ as required, where in the last inequality we make use of the observation that $n\geq s\geq 2^{12}\varepsilon^{-1}k^2\log^4n>2k$. \end{proof} \section{Finding edge disjoint paths through random vertex sets}\label{sec:connectinexpander} We will now show that a robust sublinear expander is not only well-connected (in the sense of Definition~\ref{defn:path-connected}), but is likely to be well-connected through any large random vertex subset. That is, we prove the following result. \begin{theorem}\label{thm:pathconnect} Let $G$ be an $n$-vertex $(\varepsilon,s)$-expander with $1\ge \varepsilon \ge 2^{-7}$ and $s \ge \log^{135} n$. Let $V\subseteq V(G)$ be a random subset chosen by including each vertex independently at random with probability $\frac13$. Then, with high probability, $G$ is $(4\log^5 n, 2^8\log^5 n)$-path connected through $V$. \end{theorem} The challenge of proving Theorem~\ref{thm:pathconnect} is discussed in Section~\ref{subsec:sketch}, and in particular in the answer to question \textbf{ii)} there. Having since then proved Lemma~\ref{lem:connectmanypairs}, let us note that, for suitable polylogarithmic parameters $\bar{s}$ and $\ell$, by this lemma, to prove Theorem~\ref{thm:pathconnect} it is sufficient to show that, with high probability, $|B^\ell_{G-F}(U,V)|> \frac{|V|}2$ for each $U\subseteq V(G)$ and $F\subseteq E(G)$ with $|F|\leq \bar{s}|U|$. We show this property in three stages. Firstly, in Section~\ref{sec:expand1}, we show that for any set $U\subseteq V(G)$ with $|N_G(U)|\geq |U|\log^{24}n$ and any $F\subseteq E(G)$ with $|F|\leq |U|$ we have $|B^\ell_{G-F}(U,V)| > \frac{|V|}2$ holds with probability $1-\exp(-\Omega(|U|\log^{2}n))$. That is, we show the desired property holds for a \emph{well-expanding} set $U$ with high enough probability that we can take a union bound over all well-expanding sets $U$ (and any edge set $F$ with say $|F|\leq |U|$) to get that this property holds for all well-expanding sets with probability $1-o(1/n)$. As discussed in Section~\ref{subsec:sketch}, if the set $U$ does not expand well then we can not guarantee it has the property we want with high enough probability for a union bound over all sets $U$, so we only consider well-expanding sets $U$ here. Then, in Section~\ref{sec:expand2}, we show that every set $U$ in our expander contains a well-expanding set $U'$ which is not that much smaller than $U$ (see \Cref{prop:well-expanding-core}), indeed, we will find such a $U'$ satisfying $|U'|\geq s'|U|$, where $s'\geq 1/\log^{27}n$. Thus, for any edge set $F$ with $|F|\leq s'|U|\leq |U'|$ we have $|B^\ell_{G-F}(U,V)|\geq |B^\ell_{G-F}(U',V)| > \frac{|V|}2$. This is almost the condition we need to apply Lemma~\ref{lem:connectmanypairs}, however $s'$ is too small, namely by a polylogarithmic factor smaller than the value of $\bar{s}$ that we need. Therefore, in Section~\ref{subsec:well-connectivity-weak-expander}, we find the stronger expansion property we need by first splitting an $(\varepsilon,s)$-expander into polylogarithmically many edge disjoint expanders via \Cref{lem:partitionedgesintoexpanders}, before showing that with high probability each one of these has the above-mentioned weaker expansion property into $V$. Finally, we combine these properties to show that indeed $G$ has the desired stronger expansion property into $V$. This allows us to apply Lemma~\ref{lem:connectmanypairs}, completing the proof of Theorem~\ref{thm:pathconnect}. \subsection{Expansion of well-expanding sets into a random vertex set}\label{sec:expand1} In an $n$-vertex $(\varepsilon,s)$-expander $G$, given $U\subseteq V(G)$ such that $|N_{G}(U)|\geq |U|\log^{24}n$ and $F\subseteq E(G)$ with $|F|\leq s|U|/4$, when $V\subseteq V(G)$ is chosen by selecting each vertex independently at random with probability $\frac 13$, we wish to show that, with some large probability, we have $|B^\ell_{G-F}(U,V)|>\frac{|V|}2$, for some appropriate parameters $\varepsilon,s$ and $\ell$ (thus proving Lemma~\ref{lem:expandintorandom} below). To prove this we will adapt a `sprinkling' argument from a very recent work of Tomon~\cite{tomon2022robust}, thus avoiding a much more complex argument from initial versions of this work. To prove this we will reveal the vertices in $V$ in $\ell$ batches, using the so-called sprinkling method, by partitioning $V$ randomly into sets $V_1\cup\ldots\cup V_\ell$, weighted so that most of the vertices in $V$ are likely to be in $V_\ell$. A natural approach here would be to prove a likely bound on $B^i_{G-F}(U,V_1\cup\ldots \cup V_i)$ for each $i\in [\ell]$, resulting in a bound on $B^\ell_{G-F}(U,V_1\cup\ldots \cup V_\ell)=B^\ell_{G-F}(U,V)$, so let us emphasise that this is \emph{not} what we do. Instead, for $0\leq i\leq \ell$ we track the size of sets $B_i$ which are defined as the set of vertices $v \in V(G)$ which can be reached from $U$ by a path all of whose internal vertices are in $V_1 \cup \ldots \cup V_{i-1}$ (i.e., a path through this set) of length at most $i$. It is \emph{crucial} here that we do not insist $v$ belongs to either $U$ or $V_1\cup \ldots \cup V_i$. One can think of vertices in $B_i$ as having potential of making all their neighbours in $G-F$ reachable in the same way (so being in $B_{i+1}$) if they get sampled into our next random subset $V_{i}$. In particular, every vertex with a neighbour in $B_i$ which gets sampled into $V_{i}$ will join $B_{i+1}$. This combined with our notion of robust expansion (as discussed further below) allows us to show that it is likely that $B_{i+1}$ will increase in size compared to $B_i$ until for some $i\le \ell-1$ its size is at least $\frac23 n$. In particular, we will have $B_{\ell} \ge \frac23 n$. The final stage is slightly different, here since $B_{\ell}$ is independent of our final random set $V_{\ell}$ we will likely have almost $\frac23$ of the vertices of $V_\ell$ belonging to $B_{\ell}$. As our random sets are weighted heavily towards $V_\ell$, it is likely to contain more than $\frac34$ of the vertices of $V$ so that we will likely have $|V_\ell\cap B_\ell|>\frac{|V|}2$, so that, finally, we have \[ |B^\ell_{G-F}(U,V)|\geq |V_\ell\cap B_\ell|>\frac{|V|}2, \] as required. That the sets $B_i$, $1\leq i\leq \ell-1$, are very likely to increase notably in size will follow from our notion of robust expansion (as proved in the claim below). In particular, at step $i$, \Cref{prop:red-blue-expansion-robust-both} tells us that one of two cases \textbf{a)} or \textbf{b)} may occur. \textbf{a)} $B_i$ has many large vertex disjoint stars extending from $B_i$. In this case we use that, for each centre sampled into $V_i$, the (many) corresponding leaves are added to $B_{i+1}$. We will have that many more leaves are added for each successful centre than the sampling probability for $V_i$, so that this is a good increase in size. \textbf{b)} $B_i$ has a large robust neighbourhood whose vertices have many neighbours in $G-F$ in $B_i$. Each vertex in this robust neighbourhood is likely to have at least one of these neighbours in $B_i$ sampled into $V_i$, whereupon it will then be in $B_{i+1}$. (In fact, we need a slightly stronger property to hold so that the sampling of each vertex in $B_i$ does not have too strong an influence on the size of $B_{i+1}\setminus B_i$, which is why we use the subgraph $H$ provided by case \textbf{b)} of \Cref{prop:red-blue-expansion-robust-both}.) Thus, in either case $|B_i|$ is likely to increase. \begin{lemma}\label{lem:expandintorandom} Suppose that $G$ is an $n$-vertex $(\varepsilon,s)$-expander with $2^{-9}\leq \varepsilon \leq 1$ and $s \ge 8\log^{13} n$. Let $U\subseteq V(G)$ satisfy $|N_{G}(U)|\geq |U|\log^{24}n$ and let $F\subseteq E(G)$ satisfy $|F|\leq {|U|}$. Let $V\subseteq V(G)$ be a random subset chosen by including each vertex independently at random with probability $\frac13$. Then, with probability $1-e^{-\Omega\left(|U|\log^{2}n\right)}$, \begin{equation}\label{expand} |B^{\log^4n}_{G-F}(U, V)|>\frac{|V|}{2}. \end{equation} \end{lemma} \begin{proof} Let $\ell=\log^4n$, $q=\frac3{11}$ and let $p$ be such that $1-(1-p)^{\ell-1}(1-q)=\frac{1}{3}$, i.e., that $(1-p)^{\ell-1}=\frac{11}{12}$, so that \begin{equation}\label{eqn:p15} p\ge \frac{1}{15\log ^4 n}. \end{equation} Let $G$ be an $n$-vertex $(\varepsilon,s)$-expander, $U\subseteq V(G)$ with $|N_G(U)| \ge |U|\log^{24} n$ and $F\subseteq E(G)$ with $|F|\leq {|U|}$. Independently, for each $i\in \{1,\ldots,\ell\}$, let $V_i$ be a random subset of $V(G)$ with each vertex included independently at random with probability $p$ if $i\leq \ell-1$ and with probability $q$ if $i=\ell$. Set $V=V_1\cup\ldots \cup V_\ell$, and note that each vertex is included in $V$ independently at random with probability $\frac13$. Thus, we wish to show that, with probability at least $1-e^{-\Omega\left({|U|}{\log^{2}n}\right)}$ we have $|B^{\ell}_{G-F}(U , V)|>\frac{|V|}{2}$. For each $0\le i\leq \ell$, let $B_{i}$ be the set of vertices of $G$ which can be reached via a path in $G-F$ which starts in $U$ and has length at most $i$ and whose internal vertices (if there are any) are in $V_1\cup \ldots \cup V_{i-1}$. In particular, then, we have $B_0=U$ and $B_1=B_{G-F}(U)$. Observe also that $B_0\subseteq B_1\subseteq \ldots \subseteq B_{\ell}$. We emphasise that the vertices of $B_{i}$ do not have to themselves be inside $V_1\cup \ldots \cup V_{i-1}$, only the interior vertices of some path from $U$ to the vertex in $B_{i}$ are required to be inside $V_1\cup \ldots \cup V_{i-1}$. An important property of $B_{i}$ is that it is completely determined by the sets $U,V_1,\ldots, V_{i-1}$, so is in particular independent of $V_{i}$. Note also that any vertex in $N_{G-F}(B_i)$ with a neighbour in $B_i$ that gets sampled into $V_i$ belongs to $B_{i+1}$. These two observations will be the key behind why the sets $B_{i+1}$ will grow in size until they occupy most of the set $V(G)$. In particular, finally, observe that \begin{equation}\label{eqn:overkill4} B_\ell\cap V_\ell\subseteq B^{\ell}_{G-F}(U, V). \end{equation} We now show that indeed, for each $1\le i\le \ell-1$, that, unless $B_{i}$ is already very large, $B_{i+1}$ is likely to be larger than $B_i$, as follows. \begin{claim*} For each $1\le i\le \ell-1$, with probability $1-e^{-\Omega\left({|U|}{\log^{2}n}\right)}$, either $|B_i| \ge \frac23 n$, or $$|B_{i+1}\setminus B_i| \ge \frac{\varepsilon|B_i|}{2^6\log^{2} n}.$$ \end{claim*} \begin{proof} For each $v\in N_{G-F}(B_i)$, $v$ is in $B_{i+1}$ if at least one of its neighbours in $G-F$ in $B_i$ gets sampled into $V_{i}$. That is, \begin{equation}\label{eqn:overkill1} \{v\in N_{G-F}(B_i):N_{G-F}(v,B_i)\cap V_i\neq\emptyset\} \subseteq B_{i+1}\setminus B_i. \end{equation} We will show that, for any set $W\subseteq V(G)$ with $|W|\leq \frac23 n$ and $B_1\subseteq W$ \begin{equation}\label{eqn:overkill2} \P\left(|\{v\in N_{G-F}(W):N_{G-F}(v,W)\cap V_i\neq\emptyset\}|\geq \frac{\varepsilon|W|}{2^6\log^{2} n}\right)\ge 1-e^{-\Omega\left({|B_1|}/{\log^{22} n}\right)}. \end{equation} Thus, we will have for all $1\le i \le \ell -1$ \begin{align*} \P\left(|B_i| \ge \frac23 n \: \text{ or } \: |B_{i+1}\setminus B_i| \ge \frac{\varepsilon|B_i|}{2^6\log^{2} n}\right) &\overset{\textcolor{white}{\eqref{eqn:overkill1}}}{\geq} \P\left(|B_{i+1}\setminus B_i| \ge \frac{\varepsilon|B_i|}{2^6\log^{2} n}\: \big| \: |B_i|\leq \frac{2}{3}n\right) \\ &\overset{\eqref{eqn:overkill1}}{\geq} \P\left(|\{v\in N_{G-F}(B_i):N_{G-F}(v,B_i)\cap V_i\neq\emptyset\}|\ge \frac{\varepsilon|B_i|}{2^6\log^{2} n}\: \big| \: |B_i|\leq \frac{2}{3}n\right)\\ &\overset{\eqref{eqn:overkill2}}{\ge} 1-e^{-\Omega\left({|B_1|}/{\log^{22} n}\right)}\ge 1-e^{-\Omega\left({|U|}{\log^{2} n}\right)}, \end{align*} where in the last inequality we used $|B_1|=|B_{G-F}(U)|\ge |U|\log^{24} n-|F|\geq \frac12|U|\log^{24}n$. Let then $W\subseteq V(G)$ with $|W|\leq \frac23n$ and $B_1\subseteq W$. As $|W|\leq \frac23 n$, and $|F|\le |U|\leq |B_1|\leq |W|\le {s|W|}/4$ we can apply \Cref{prop:red-blue-expansion-robust-both} to $W$ and $F$ to show one of two cases \textbf{a)} or \textbf{b)} holds and we will show that \eqref{eqn:overkill2} holds in either case. \textbf{a)} Suppose $G-F$ contains $\frac{|W|}{\log^{7} n}$ vertex disjoint stars with $\log^{9} n$ leaves, with the centre in $W$ and all leaves in $N_{G-F}(W)$. Let $C\subseteq W$ be the set of centres of such a collection of stars, and note that \begin{equation}\label{eqn:overkill3} |\{v\in N_{G-F}(W):N_{G-F}(v,W)\cap V_i\neq \emptyset|\geq |C\cap V_i|\log^9n. \end{equation} By a Chernoff bound (\Cref{chernoff}) and \eqref{eqn:p15}, with probability at least $1-e^{-p|C|/8}= 1-e^{-\Omega\left({|W|}/{\log^{11}n}\right)}$, we have $|C\cap V_i|\geq \frac{p|C|}2\geq \frac{|W|}{2^6\log^{11}n}$. Thus, in combination with \eqref{eqn:overkill3}, we have that \eqref{eqn:overkill2} holds as $\varepsilon\leq 1$. \textbf{b)} Suppose instead that there is a bipartite subgraph $H\subseteq G-F$ with vertex classes $W$ and $X$ such that \begin{itemize} \item $|X| \ge \frac{\varepsilon|W|}{2\log^2 n}$ and \item every vertex in $X$ has degree at least $\log^{4} n$ in $H$ and every vertex of $U$ has degree at most $\Delta:=2\log^{9} n$ in $H.$ \end{itemize} For each $v\in X$, the probability that $v$ has no neighbours in $H$ in $V_i$ is at most \[ (1-p)^{\log^4n}\le e^{-p\log^4n}\overset{\eqref{eqn:p15}}{\leq} e^{1/15}\leq \frac{15}{16}. \] Let $Y$ be the random variable counting the number of vertices of $X$ having a neighbour in $V_{i}$ in $H$, so that $\mathbb{E} Y \ge \frac{|X|}{16}$. Observe also that $Y$ is $\Delta$-Lipschitz since for each $v\in W$ the event $\{v\in V_i\}$ affects $Y$ by at most $d_H(v)\leq \Delta$. Hence, by \Cref{lem:mcd} with $k=\Delta$, $t=\frac{|X|}{32}$ and $N=|W|$, we have $$\P\left(Y < \frac{|X|}{32}\right)\le \P\left(Y < \mathbb{E} Y - \frac{|X|}{32} \right)\le2\exp\left(-\frac{2^{-9}|X|^2}{\Delta^2|W|}\right)= e^{-\Omega\left({|W|}/{\log^{22} n}\right)}. $$ Each vertex in $X$ with a neighbour in $V_i$ in $H$ lies in $\{v\in N_{G-F}(W):N_{G-F}(v,W)\cap V_i\neq\emptyset\}$, so therefore, with probability at least $1-e^{-\Omega\left({|W|}/{\log^{22} n}\right)}$, we have $|\{v\in N_{G-F}(W):N_{G-F}(v,W)\cap V_i\neq\emptyset\}|\geq Y\geq \frac{|X|}{32}\geq \frac{\varepsilon|W|}{2^6\log^2 n}$ and thus \eqref{eqn:overkill2} holds as well in case \textbf{b)}, completing the proof. \end{proof} As $B_\ell$ and $V_\ell$ are independent and $q=\frac3{11}$ (so that $\frac{2q}{3}=\frac{2}{11}<\frac{4}{23}$), by Chernoff's bound (\Cref{chernoff}), we have that \[ \P\left(|B_{\ell}\cap V_\ell|\le\frac{4}{23}n \: \big| \: |B_{\ell}|\geq \frac{2}3n\right)\leq \P\left(\bin\left(\frac{2}3n,q\right)\leq \frac{4}{23}n\right)\le e^{-\Theta(n)}, \] and, similarly, as $\frac13<\frac8{23}$ we have $\P\left(|V|\geq \frac{8}{23}n\right)\le e^{-\Theta(n)}$. Thus, by the claim, we have in total that \begin{enumerate}[label=\roman*)] \item \label{itm1} for each $i\in [\ell-1]$, $|B_i| \ge \frac23 n$ or $|B_{i+1}\setminus B_i| \ge \frac{\varepsilon|B_i|}{2^6\log^{2} n}$, and \item $|B_\ell|<\frac23 n$ or $|B_\ell\cap V_\ell|>\frac{4}{23}n$, and \item \label{itm3} $|V|\leq \frac{8}{23}n$ \end{enumerate} with probability at least $$1-\log^4n\cdot e^{-\Omega\left({|U|}{\log^{2} n}\right)}-2^{-\Theta(n)}\ge 1-e^{-\Omega\left({|U|}{\log^{2} n}\right)}.$$ However, if \ref{itm1}--\ref{itm3} all hold, then, for each $i\in [\ell-1]$, we have $$|B_i| \ge \min\left\{\frac23 n,\left(1+\frac{\varepsilon}{2^6\log^2 n}\right)^{i}|U|\right\}\geq \min\left\{\frac23 n,\exp\left(\frac{\varepsilon i}{2^7\log^2 n}\right)\right\},$$ so that, setting $i=\ell=\log^4n$, we conclude $|B_{\ell}|\geq \frac23 n$, and hence, by ii) and iii), that $|B_\ell\cap V_\ell|>\frac{|V|}2$. Thus, by \eqref{eqn:overkill4}, we have that $|B^{\ell}_{G-F}(U,V)|>\frac{|V|}2$ with probability at least $1-e^{-\Omega\left({|U|}{\log^{2} n}\right)}$. \end{proof} \subsection{Expansion into a random vertex set}\label{sec:expand2} Having picked $V\subseteq V(G)$ with vertex probability $\frac13$ in an $n$-vertex $(\varepsilon,s)$-expander $G$, \Cref{lem:expandintorandom} tells us that for any \emph{fixed}, well-expanding subset of vertices $U$ and small set $F$ of edges we can reach more than one half of the vertices of $V$ by short paths through $V$ in $G-F$ with pretty high probability. We now want to use this to show that a similar property holds \emph{simultaneously} for \emph{all} vertex subsets $U$. As we cannot directly take a union bound over all subsets $U$, we first show that any vertex subset $U$ in an $(\varepsilon,s)$-expander contains a subset $U'\subseteq U$ which expands particularly well and which is not much smaller than $U$. This follows easily from the definition of expansion but is perhaps easier to immediately see why it is true from the perspective introduced in \Cref{prop:expansion-red-blue}. \begin{prop}\label{prop:well-expanding-core} Let $n\geq 2$, $0<\varepsilon\leq 1$ and $s \ge \log^{24} n$. Let $G$ be an $n$-vertex $(\varepsilon,s)$-expander and let $U\subseteq V(G)$ have size $|U|\le \frac{2}3n$. Then, there is a set $U'\subseteq U$ with $|N_{G}(U')|\geq |U'|\log^{24} n$ and $|U'|\geq \frac{\varepsilon|U|}{3\log^{26}n}$. \end{prop} \begin{proof} Let $U'\subseteq U$ be maximal subject to $|N_{G}(U')|\geq |U'|\log^{24}n$, noting this is possible as $U'=\emptyset$ satisfies these conditions. Suppose that $U\neq U'$, for otherwise $U$ satisfies the conditions itself. Then $|N_{G}(U')|< (|U'|+1)\log^{24}n+1$ or we could add an arbitrary vertex to $U'$ and contradict maximality. Similarly we know that, for every vertex $v\in U\setminus U'$, $v$ has at most $\log^{24}n$ neighbours outside of $U'\cup N_{G}(U')$, for otherwise $U'\cup \{v\}$ contradicts the maximality. Let $F$ be the set of edges between $U\setminus U'$ and $V(G)\setminus (U'\cup N_{G}(U'))$, so that $|F|\leq |U\setminus U'|\log^{24}n\leq s|U|$. Thus, we have by the definition of expansion that \[ \frac{\varepsilon|U|}{\log^2n} \le |N_{G-F}(U)| \le |N_{G}(U')| \le (|U'|+1)\log^{24}n+1, \] so that $|U'|\geq \frac{\varepsilon |U|}{3\log^{26}n}$, as required. \end{proof} We now show that we can ensure the conclusion of \Cref{lem:expandintorandom} holds for \emph{all} well-expanding sets simultaneously by taking a union bound, and then use Proposition~\ref{prop:well-expanding-core} to deduce an expansion property for all sets, as follows. \begin{lemma}\label{lem:expandintorandom2} Suppose that $G$ is an $n$-vertex $(\varepsilon,s)$-expander with $2^{-9}\leq \varepsilon\leq 1$ and $s \ge 2\log^{24} n$. Let $V\subseteq V(G)$ be a random subset chosen by including each vertex independently at random with probability $\frac13$. Then, with probability at least $1-o\left(1/n\right)$, for every $U\subseteq V(G)$ and every set $F\subseteq E(G)$ with $|F|\leq \frac{|U|}{\log^{27}n}$ \begin{equation}\label{eqn:B1UV} |B_{G-F}^{\log^4n}(U,V)|> \frac{|V|}{2}. \end{equation} \end{lemma} \begin{proof} Say a set $U'\subseteq V(G)$ \emph{expands well} in $G$ if $|N_{G}(U')| \ge |U'|\log^{24}{n}.$ Given a non-empty well-expanding set $U'\subseteq V(G)$ and a set of edges $F$ of size at most $|U'|$, \Cref{lem:expandintorandom} applied to $U'$ implies that \begin{equation}\label{eq:1} |B_{G-F}^{\log^4n}(U', V)|> \frac{|V|}2 \end{equation} fails with probability at most $e^{-\Omega(|U'|\log^2n)}. Now a union bound over all pairs $(U',F)$ such that $U'$ is a well-expanding set in $G$ and $F$ is a set of at most $|U'|$ edges tells us that \emph{some} such pair $(U',F)$ fails \eqref{eq:1} with probability at most \begin{align*} \sum_{(U',F)} e^{-\Omega(|U'|\log^2n)} &\le \sum_{u=1}^n\sum_{f=1}^{u} \binom{n}{u}\binom{n^2}{f} \cdot e^{-\Omega(u\log^2n)} \\ & \le \sum_{u=1}^n u\cdot n^{3u}\cdot e^{-\Omega(u\log^2n)} \le \sum_{u=1}^n e^{-\Omega(u\log^2n)}=o(1/n). \end{align*} Thus, with probability $1-o(1/n)$, we can assume that \eqref{eq:1} holds for every well-expanding set $U'$ and set $F\subseteq E(G)$ with $|F|\leq |U'|$. We will now show that this implies \eqref{eqn:B1UV} holds for all $U\subseteq V(G)$ and $F\subseteq E(G)$ with $|F|\leq \frac{|U|}{\log^{27}n}$, completing the proof. Let then $U\subseteq V(G)$ with $|U|\leq \frac23n$ and let $F\subseteq E(G)$ satisfy the (slightly weaker) condition $|F|\leq \frac{2 |U|}{\log^{27}n}$. Then, by \Cref{prop:well-expanding-core}, there is a set $U'\subseteq U$ which is well-expanding for which $|U'|\geq \frac{\varepsilon|U|}{3\log^{26}n}$. Noting that $|F|\leq |U'|$ (as we may assume $n$ is large with probability $1-o(1/n)$), we therefore have that \[ |B^{\log^4n}(U,V)|\geq |B^{\log^4n}(U',V)|> \frac{|V|}{2}. \] Finally, consider $U\subseteq V(G)$ with $|U|> \frac23n$ and let $F\subseteq E(G)$ satisfy $|F|\leq \frac{|U|}{\log^{27}n}$. Let $\bar{U}\subseteq U$ be an arbitrary subset with $\frac{n}2\leq |\bar{U}|\leq \frac23 n$, so that we have $|F|\leq \frac{2|\bar{U}|}{\log^{27}n}$, and hence, from what we have just shown, \[ |B^{\log^4n}(U,V)|\geq |B^{\log^4n}(\bar{U},V)|> \frac{|V|}{2}, \] as required. \end{proof} \subsection{Path connectedness through a random subset in expanders}\label{subsec:well-connectivity-weak-expander} We are now ready to prove Theorem~\ref{thm:pathconnect}. As discussed at the start of this section, we first split the edges of the graph $G$ into expanders, before applying Lemma~\ref{lem:expandintorandom2} to each of these, to get (with high probability), a strong enough expansion condition to apply \Cref{lem:connectmanypairs}. \begin{proof}[ of Theorem~\ref{thm:pathconnect}] To recap: we have an $n$-vertex $(\varepsilon,s)$-expander, $G$, with $2^{-7}\leq \varepsilon \leq 1$ and $s\geq \log^{135}n$, and a random subset $V\subseteq V(G)$ where each vertex is included independently at random with probability $\frac13$. To prove Theorem~\ref{thm:pathconnect}, we need to show that, with high probability, $G$ is $(4\log^5n,2^8\log^5n)$-path connected through $V$. Let $k=2^{17}\log^{42} n$, so that $s\geq 2^{12}\varepsilon^{-1}k^2\log^4n$, and let $s'=\frac{\sqrt{s\varepsilon}}{ 8k \log n} \ge 2\log ^{24} n$. Using Lemma~\ref{lem:partitionedgesintoexpanders}, take edge disjoint graphs $G_1,\ldots,G_k$ such that $E(G)=\bigcup_{i\in [k]}E(G_i)$ and, for each $i\in [k]$, $G_i$ is an $\left(\frac{\varepsilon}4,s'\right)$-expander. Then, by \Cref{lem:expandintorandom2} and a union bound over the $k$ graphs $G_i$, with high probability we can assume that, for each $i\in [k]$ and every $U\subseteq V(G_i)$ and $F\subseteq E(G_i)$ with $|F|\leq \frac{|U|}{\log^{27}n}$, \[ |B^{\log^4n}_{G_i-F}(U,V)|> \frac{|V|}{2}. \] Now, let $U\subseteq V(G)$ and $F\subseteq E(G)$ with $|F|\leq 2^{17}|U|\log^{15}n$. As the graphs $G_i$, $i\in [k]$, are edge disjoint, there must be some $i\in [k]$ with $|F\cap E(G_i)|\leq \frac{2^{17}|U|\log^{15}n}k=\frac{|U|}{\log^{27}n}$, and therefore \[ |B^{\log^4n}_{G-F}(U,V)|\geq |B^{\log^4n}_{G_i-F}(U,V)|> \frac{|V|}{2}. \] Thus, by \Cref{lem:connectmanypairs}, applied with $k=2^8\log^{5}n$ and $\ell=\log^4 n$ we conclude that $G$ is $(4\log^5n,2^8\log^5n)$-connected, as desired. \end{proof} \section{Cycle decompositions}\label{sec:logstarproof} In this section we will prove our main results, \Cref{thm:logstar,thm:decompexander}. Before doing this we need to put together a few final ingredients. In Section~\ref{sec:connectinexpander}, we established a very robust connectivity property of expanders. In \Cref{sec:skeletons} we will show that in an expander one can find a subgraph with few edges and yet (effectively) the same connectivity property through a random subset $V$, with high probability. We will refer to this subgraph as a \emph{skeleton} of our graph, divide the vertex set into three as $V(G)=V_1\cup V_2\cup V_3$, and find three matching skeletons. In \Cref{sec:lovasz} we show that any graph can be decomposed into few paths in such a way that no vertex is used as an endvertex many times. In \Cref{subsec:decompexpander} we combine these results to decompose any expander into linearly many cycles and a few leftover edges. As outlined in Section~\ref{subsec:sketch}, we achieve this by setting aside the skeletons, then decomposing the remainder of the graph into three sets of paths and finally using the connection properties of the skeletons to join the endvertices of these sets of paths, where the sets of paths are matched to the skeletons to ensure this creates edge disjoint cycles. These cycles decompose all the edges in the graph which are not in the skeleton and since the skeleton is chosen to be sparse this gives us the result. The final ingredient in the proof of \Cref{thm:decompexander}, given in \Cref{sec:decomp-general}, is to decompose an arbitrary graph into expanders and a few leftover edges via \Cref{lem:decompexander}, to each of which we can apply our expander decomposition result. All that will remain, then, is to iteratively apply \Cref{thm:decompexander} in \Cref{subsec:iteration}, while removing some additional cycles, to deduce \Cref{thm:logstar}. \subsection{Finding the skeletons}\label{sec:skeletons} To find the skeletons, we will use \Cref{thm:pathconnect} to embed a sparse well-connected `template' graph (from \Cref{lem:template}) with its edges replaced by relatively short edge disjoint paths, and show that the image of this embedding has the properties we need of a skeleton, as follows. \begin{lemma}\label{lem:sparseconnect} Let $G$ be an $n$-vertex graph which is an $(\varepsilon,s)$-expander with $2^{-7}\leq \varepsilon \leq 1$ and $s\geq \log^{135}n$. Let $V\subseteq V(G)$ be chosen by including each vertex independently at random with probability $\frac13$. Then, with high probability, there is a subgraph of $G$ with at most $2^{9}n\log^{10}n$ edges which is $(\log^7 n,2)$-path connected through $V$. \end{lemma} \begin{proof} By Theorem~\ref{thm:pathconnect} applied to $G$ and $V$, $G$ is with high probability $(4\log^5 n,2^8\log^5 n)$-path connected through $V$. Note that, by Chernoff's inequality (\Cref{chernoff}), we can in addition ensure with high probability that $|V|\geq \frac n6$, and since our goal is to show a statement with high probability we may assume that $n$ is large enough to apply \Cref{lem:template}. That is, by that lemma we may assume there is an auxiliary graph $H$ with vertex set $V(H)=V(G)$ which is $\left(\frac14\log^2 n,2\right)$-path connected through $V$, and such that $\Delta(H) \le2^8\log^5 n$. Note that $E(H)$ is a collection of pairs of vertices in $V(G)$ with the property that every vertex appears in at most $\Delta(H) \le 2^8\log^5 n$ pairs. Hence, since $G$ is $(4\log^5 n, 2^8\log^5 n)$-path connected through $V$, we can find for each $e\in E(H)$ a path $P_e$ through $V$ of length at most $4\log^5 n$, such that all the paths $P_e$, $e\in E(H)$, are edge disjoint. Let $G'$ have vertex set $E(G)$ and edge set $\bigcup_{e\in E(H)}E(P_e)$, noting that \[ |E(G')|\leq 4\log^5n \cdot |E(H)|\leq 2n\log^5n \cdot \Delta(H)\leq 2n\log^5n\cdot 2^8\log^5n=2^{9}n\log^{10}n. \] Therefore, we need only show that $G'$ is $(\log^7 n,2)$-path connected. For this, let $\mathcal{P}\subseteq \binom{V(G)}{2}$ be a family of pairs of vertices from $V(G)$ with each vertex appearing in at most $2$ different pairs. Since $H$ is $\left(\frac14\log^2 n,2\right)$-path connected through $V$ we can find edge disjoint paths in $H$ through $V$, each of length at most $\frac14\log^2 n$, through $V$, joining each pair in $\mathcal{P}$. If we now replace each edge $e$ of $H$ used by one of these paths with $P_e$ we obtain a collection of edge disjoint \emph{walks} in $G'$ of length at most $\log^7 n$, through $V$, joining each pair in $\mathcal{P}$. Replacing each of the walks with its shortest subwalk joining the same endvertices, we obtain paths which are edge disjoint, each have length at most $\log^7n$, and which connect the vertex pairs in $\mathcal{P}$. Thus, $G'$ is $(\log^7 n,2)$-path connected, as claimed. \end{proof} \subsection{Lov\'asz path covering with well-spread endvertices}\label{sec:lovasz} As quoted in the introduction, Lov\'asz proved the following classical decomposition result in 1968. \begin{theorem}[Lov\'asz~\cite{lovasz1968covering}]\label{thm:lovasz} Every $n$-vertex graph can be decomposed into at most $\frac n2$ paths and cycles. \end{theorem} Theorem~\ref{thm:lovasz} almost provides the path decompositions that we need. At the expense of using perhaps slightly more paths, we can ensure in addition that no vertex is used often as an endvertex of the paths in the decomposition through the following simple deduction. \begin{corollary}\label{cor:lovasz} Every $n$-vertex graph $G$ can be decomposed into paths so that each vertex of $G$ is an endvertex of at most two paths in the decomposition. \end{corollary} \begin{proof} Form a graph $G'$ from $G$ by adding a new vertex $v_0$ and an edge from $v_0$ to each vertex $v\in V(G)$ for which $d_G(v)$ is even. By Theorem~\ref{thm:lovasz}, there is a collection $\mathcal{C}$ of at most $\frac{n+1}2$ cycles and paths which decomposes $G'$. Note that each vertex $v\in V(G)$ has odd degree in $G'$, and therefore must be an endvertex of an odd number of paths in $\mathcal{C}$, and thus, in particular, must be an endvertex of some path in $\mathcal{C}$. As the paths in $\mathcal{C}$ together have at most $n+1$ endvertices, each vertex in $G$ is the endvertex of at most 1 path in $\mathcal{C}$, as otherwise there would need to be $n-1+3 > n+1$ endvertices. Note that, furthermore, as there must be at least $n$ endvertices together for the paths in $\mathcal{C}$, we must have that $\mathcal{C}$ contains at least $\frac n2$ paths. Thus, as $|\mathcal{C}|\leq \frac{n+1}2$, $\mathcal{C}$ is in fact a collection consisting only of paths, with no cycles. Now, for each path $P\in \mathcal{C}$, if $v_0\in V(P)$ then remove the vertex $v_0$ from $P$, and let $\mathcal{C'}$ be the collection of all the resulting paths. Then, $\mathcal{C}'$ is a decomposition of $G$ into paths. Furthermore, observe that each vertex $v\in V(G)$ is an endvertex of a path $P\in \mathcal{C}$ only if $v$ was an endvertex of some path in $\mathcal{C}$ which contained $P$ or if $P$ was created by removing the edge $vv_0$ from a path in $\mathcal{C}$. Thus, each vertex is an endvertex of at most two paths in $\mathcal{C'}$, so that $\mathcal{C}'$ decomposes $G$ as required. \end{proof} \subsection{Decomposing expanders}\label{subsec:decompexpander} In this section we will decompose an expander into linearly many cycles and a few leftover edges, as follows. \begin{lemma}\label{lem:decompexander} Any sufficiently large $n$-vertex graph $G$ which is an $(\varepsilon,s)$-expander with $2^{-5}\leq \varepsilon \leq 1$ and $s \ge \log^{273} n$ can be decomposed into at most $3n$ cycles and at most $2^{11}n\log^{10}n$ edges. \end{lemma} \begin{proof} Let $s'=\frac{\sqrt{s\varepsilon}}{24\log n}\ge \log^{135} n$. Using Lemma~\ref{lem:partitionedgesintoexpanders} with $k=3$, take edge disjoint $\left(\frac{\varepsilon}4,s'\right)$-expander subgraphs $G_1,G_2$ and $G_3$ of $G$, each with vertex set $V(G)$, so that $E(G)=E(G_1)\cup E(G_2)\cup E(G_3)$. Next, partition $V(G)=V_1\cup V_2\cup V_3$ by assigning each vertex to a set in the partition uniformly and independently at random. Using Lemma~\ref{lem:sparseconnect}, for each $i\in [3]$, find a subgraph $G_i'\subseteq G_i$ with at most $2^{9}n\log^{10}n$ edges which is $(\log^7 n,2)$-path connected through $V_i$. For each $i\in [3]$, let $H_i$ be the graph with vertex set $V(G)\setminus V_i$ whose edges are the edges of $G-G_1'-G_2'-G_3'$ lying within $V_{i+1}$ or between $V_{i+1}$ and $V_{i+2}$, with these indices taken appropriately modulo $3$, noting that these graphs partition $G-G_1'-G_2'-G_3'$. For each $i\in [3]$, then, apply Corollary~\ref{cor:lovasz} to decompose $H_i$ into a collection of paths $\mathcal{P}_i$ with the property that no vertex is an endvertex of more than two of the paths in $\mathcal{P}_i$, noting that, in particular, this implies that $|\mathcal{P}_i|\le n.$ Since $G_i'$ is $(\log^7 n,2)$-path connected through $V_i$ and the paths in $\mathcal{P}_i$ have no vertices in $V_i$, for each path $P\in\mathcal{P}_i$, we can find a path $Q_P$ through $V_i$ in $G_i'$ joining the endvertices of $P$, so that all these new paths are edge disjoint. As $V(H_i) \cap V_i =\emptyset$, for each $P\in \mathcal{P}_i$, $P\cup Q_P$ is a cycle, and furthermore, as $H_i$ and $G_i'$ are edge disjoint, the cycles form a collection, $\mathcal{C}_i$ say, of at most $n$ edge disjoint cycles whose edges contain all of the edges of $H_i$. As the subgraphs $H_i\cup G_i'$, $i\in [3]$, are edge disjoint, $\mathcal{C}_1\cup \mathcal{C}_2\cup \mathcal{C}_3$ is a collection of at most $3n$ cycles which contains every edge of $G$ except for, perhaps, some edges in $G_1'\cup G_2'\cup G_3'$. Thus, these cycles cover all but at most $2^{11}n\log^{10} n$ edges, giving us the desired decomposition. \end{proof} \subsection{Decomposing a general graph} \label{sec:decomp-general} We are now ready to prove \Cref{thm:decompexander}, which we do in the following more quantitative form for convenience. \begin{theorem}\label{thm:decompexander-explicit} Any $n$-vertex graph can be decomposed into at most $6n$ cycles and $O(n\log^{274}n)$ edges. \end{theorem} \begin{proof} Let $n_0\geq 2^{12}$ be sufficiently large so that the statement of \Cref{lem:decompexander} holds for every graph with at least $n_0$ vertices, and let $C=3n_0$. Let $s=\log^{273}n$ and $\varepsilon=2^{-5}$. Using Lemma~\ref{splitting-into-expanders}, decompose $G$ into subgraphs $G_1,\ldots, G_r$ and at most $4sn\log n$ edges so that $|G_1|+\ldots+|G_r| \le 2n$ and, for each $i\in [r]$, $G_i$ is an $(\varepsilon,s)$-expander. Let $I\subseteq [r]$ be the set of $i\in [r]$ with $|G_i|\geq n_0$, so that \[ \big|E(G)\setminus\big(\bigcup_{i\in I}E(G_i)\big)\big|\leq n_0\cdot \sum_{i=1}^r|G_i|+4s n \log n\leq 2n_0s\cdot n \log n. \] For each $i\in I$, using Lemma~\ref{lem:decompexander}, decompose $G_i$ into at most $3|G_i|$ cycles and $2^{11}|G_i|\log^{10}n$ edges, giving in total at most $3|G_1|+\ldots+3|G_r| \le 6n$ cycles and at most $2^{11}(|G_1|+\ldots+|G_r|)\log^{10}n\leq n_0s\cdot n \log n$ edges. In combination with the edges of $G-\bigcup_{i\in I}G_i$ this gives a decomposition of $G$ into at most $6n$ cycles and $3n_0sn\log n=Cn\log^{274}n$ edges. \end{proof} \subsection{Long cycles in expanders} A final ingredient we need before proving Theorem~\ref{thm:logstar} is to show that, after removing long cycles and adding them to a decomposition, any expander in the resulting graph must be quite small. That is to say, any expander contains a long cycle even if it has no robustness at all. This follows from a result of Krivelevich \cite{michael2019expanders} although since we do not need the full power or generality of that result, for completeness we include a short proof using the Depth First Search (DFS) algorithm (first used in this manner in \cite{dfs2012}), as follows. \begin{lemma}\label{lem:onelongcycle} Any $n$-vertex $(\varepsilon,0)$-expander, with $\varepsilon \ge 2^{-5}$ and $n\geq 2^{30}/\varepsilon^2$, contains a cycle of length $\Omega\left(\frac{n}{\log^4 n}\right)$. \end{lemma} \begin{proof} Let $G$ be our $(\varepsilon,0)$-expander. Observe first that the expansion condition guarantees $G$ is connected, since otherwise we could find a connected component of $G$ of size less than $\frac{|G|}2$, whose vertex set then does not expand at all. We now run the DFS algorithm on $G$ as follows. At any point during the process we have the set of unexplored vertices $U$, the path $P$ with active endvertex $t(P)$, and the set of processed vertices $R$. Picking an arbitrary vertex $r\in V(G)$, We start with $U=V(G)\setminus \{r\}, R=\emptyset$ and with $P$ being the path with vertex set $\{r\}$, and set $t(P)=r$. At each step, if there is a neighbour $v$ of $t(P)$ in $U$ we add it to $P$ with the edge $t(P)v$ and let $t(P)=v$. Otherwise, we move $t(P)$ from $P$ to $R$ and set its neighbour in $P$ as the new $t(P)$. In the above process, in each step we either move precisely one vertex from $U$ to $P$ or precisely one vertex from $P$ to $R$. Note also that at any point in the process there are no edges between $U$ and $R$ since a vertex is only moved to $R$ once it has no neighbours in $U$, and $U$ only ever has vertices removed from it. Finally, as $G$ is connected, note that the process finishes with all the vertices being in $R$ and $P$, and $U$ being empty. Thus, we start with $|U|=n-1$ and $|R|=0$ and finish with $|U|=0$ and $|R|=n$, at each step reducing $|U|$ by one or increasing $|R|$ by one. Therefore, at some point in the process we must have $|U|=|R|$. Since there are no edges between $U$ and $R$ we know that all the neighbours of $U$ must belong to $P$, which therefore has size $|P|=n-2|U|\geq |N_G(U)| \geq \frac{\varepsilon |U|}{\log^2n}$, where the last inequality follows by the expansion property applied to $U$, which we can do since $|U|=\frac{n-|P|}2\leq \frac{n}{2}$. This in turn implies that $|P|\geq \frac{\varepsilon n}{3\log^2 n}$ since either $|U| \ge \frac{n}{3}$ or $|P|=n-2|U| \ge \frac n3$. Now, let $X$, $Y$, $Z$ be sets of consecutive vertices of $P$ in that order which partition $V(P)$ so that $|X|,|Z|\geq \frac{|P|}3$ and $\frac{\varepsilon^2 n}{18\log^4n}\leq |Y|< \frac{\varepsilon^2 n}{9\log^4 n}$. If $X$ and $Z$ are connected by some path in $G\setminus Y$, then take a shortest path, $Q$ say, between $X$ and $Z$ in $G\setminus Y$ and note that, combined with the segment of $P$ between the endvertices of $Q$, this gives a cycle containing each vertex in $Y$, which thus has size $\Omega\left(\frac{n}{\log^4 n}\right)$. If $X$ and $Z$ are not connected by a path in $G\setminus Y$, then we can take a partition $V(G)\setminus Y=X'\cup Z'$ with no edges between $X'$ and $Z'$ in $G$, and $X\subseteq X'$ and $Z\subseteq Z'$. Without loss of generality, suppose that $|X|\leq \frac n2$. By the expansion condition we have $|N_G(X)|\geq \frac{\varepsilon |X|}{\log^2n} \geq \frac{\varepsilon |P|}{3\log^2n} \geq \frac{\varepsilon^2 n}{9\log^4 n}$, yet we also have $|N_G(X)|\leq |Y|< \frac{\varepsilon n}{9\log^4n}$, a contradiction. \end{proof} \subsection{Proof of \texorpdfstring{\Cref{thm:logstar}}{Theorem 2}} \label{subsec:iteration} To decompose a graph into cycles and edges and prove \Cref{thm:logstar}, we now repeatedly do the following: \begin{itemize} \item letting $d$ be the average degree of the graph consisting of the remaining edges, we remove maximally many edge disjoint cycles with length at least $d$, \item we then decompose the remaining edges exactly into expanders (using Theorem~\ref{splitting-into-expanders} with $s=0$) and show that these must be small subgraphs as they each have no cycle with length at least $d$ (using Lemma~\ref{lem:onelongcycle}), \item and finally we decompose each of these small subgraphs into cycles and edges using Theorem~\ref{thm:decompexander-explicit}. \end{itemize} Together this comprises the iterative step we use, which decomposes the $n$-vertex graph $G$ with average degree $d$ into $O(n)$ cycles and $n\log^{O(1)}d$ edges. We then iterate on the graph of the edges in this decomposition, noting that its average degree is much smaller than $d$. We state and prove the outcome of one iterative step as the following lemma, for convenience and its own interest. \begin{lemma}\label{lem:density} Any $n$-vertex graph $G$ with average degree $d\ge 2$ can be decomposed into $O(n)$ cycles and a subgraph with average degree $O(\log^{274} d)$. \end{lemma} \begin{proof} Let $\mathcal{C}$ be a maximal collection of edge disjoint cycles with length at least $d$ in $G$. As there are $\frac{nd}2$ edges in $G$, we have $|\mathcal{C}|\leq \frac n2$. Let $G'$ be $G$ with the edges of the cycles in $\mathcal{C}$ removed, so that $G'$ has no cycles with length at least $d$. Apply \Cref{splitting-into-expanders} with $s=0$ and $\varepsilon=2^{-5}$ to obtain a full decomposition (since $s=0$) of $G'$ into subgraphs $G_1,\ldots, G_k$, such that $|G_1|+\ldots+|G_k|\le 2n$ and, for each $i\in [k]$, $G_i$ is a $(2^{-5},0)$-expander. For each $i\in [k]$, as $G'$ and hence $G_i$ has no cycle with length at least $d$, we have by \Cref{lem:onelongcycle} that $|G_i| = O(d \log^4 d)$. Using \Cref{thm:decompexander-explicit}, we decompose each $G_i$ into at most $6|G_i|$ cycles and $O(|G_i| \log^{274} |G_i|)= O(|G_i| \log^{274} d)$ edges. Collecting these cycles and edges over all $i\in [k]$, and including the cycles in $\mathcal{C}$, we get a decomposition of $G$ into at most $$\frac n2+\sum_{i=1}^k 6|G_i| \le 13 n$$ cycles and $$O\left(\sum_{i=1}^k |G_i| \log^{274} d\right)=O(n \log^{274} d)$$ edges. Noting these edges form a subgraph with average degree $O(\log^{274} d)$ completes the proof. \end{proof} Finally, by iterating \Cref{lem:density}, we can prove \Cref{thm:logstar}, i.e., that any $n$-vertex graph can be decomposed into $O(n\log^*n)$ cycles and edges \begin{proof}[ of \Cref{thm:logstar}] Using Lemma~\ref{lem:density}, let $C\geq 1$ be large enough that any $n$-vertex graph with average degree $d\geq 2$ has a decomposition into at most $Cn$ cycles and a subgraph with average degree at most $C\log^{274}n$. Let $G_0$ be any $n$-vertex graph, and, for each $i\geq 0$, let $G_{i+1}$ be a graph with the fewest edges that can be formed by removing at most $Cn$ edge disjoint cycles from $G_{i}$. For each $i\geq 0$, let $d_i$ be the average degree of $G_i$, so that we have $d_{i+1}\leq C\log^{274}d_i$ for each $i\geq 0$ for which $d_i\geq 2$. Let $\ell$ be the largest integer such that the log function applied iteratively $\ell$ times to $n$ is still above $300C$, i.e., the largest integer such that $\log^{[\ell]}n\geq 300C$. Note that $\ell\leq \log^{*}n$ and $\log^{[\ell]}n< 2^{300C}$. We will show, for each $0 \le i\le \ell$, that $d_i\leq C(300\log^{[i]}n)^{274}$. Note that $d_0\leq n$, so that this easily holds with $i=0$. Then, assuming it is true for some $0 \le i \le \ell-1$ and that $d_i\geq 2$ (for otherwise $d_{i-1}\leq d_i\leq 2$), we have \begin{align} d_{i+1}&\leq C\log^{274}d_i\leq C(\log (C(300\log^{[i]}n)^{274}))^{274}\nonumber\\&= C(\log C+274\log300+274\log^{[i+1]}n)^{274} \leq C(300\log^{[i]}n)^{274},\label{eqn:Clogi} \end{align} where in the last inequality we used $26\log^{[i+1]}n \ge 26\log^{[\ell]} n\ge 26\cdot 300C \ge \log C + 274 \log 300$. Thus, in particular, we have that the average degree of $G_\ell$ is at most $C(300\log^{[\ell]}n)^{274}\leq C(300\cdot 2^{300C})^{274}=O(1)$. Furthermore, to get from $G_0$ to $G_\ell$ we have removed at most $C\ell\leq Cn\log^{\star}n$ cycles. Therefore, $G_0$ has a decomposition into $O(n\log^{*}n)$ cycles and $O(n)$ edges, which, as $G_0$ is an arbitrary $n$-vertex graph, completes the proof.\end{proof} \section{Concluding remarks}\label{sec:final} \textbf{Our results.} In this paper, we gave new bounds on two of the most central open problems on cycle decompositions -- the Erd\H{o}s-Gallai conjecture from 1966 that any $n$-vertex graph can be decomposed into $O(n)$ cycles and edges, and Haj\'os's conjecture from 1968 asserting that any $n$-vertex Eulerian graph can be decomposed into at most $\frac{n}{2}$ cycles, where the bound in Haj\'os's conjecture follows easily from Theorem~\ref{thm:logstar}. Indeed, given any $n$-vertex Eulerian graph, we can apply this result first to remove $O(n\log^{\star}n)$ cycles and leave only $O(n \log^{\star} n)$ remaining edges. The remaining edges then still form an Eulerian graph, which has a cycle decomposition by the observation of Veblen quoted in the introduction. As this cycle decomposition has fewer cycles than the number of these edges, we get, altogether, a decomposition of the original Eulerian graph into $O(n\log^{\star}n)$ cycles. \textbf{Lower bounds for the Erd\H{o}s-Gallai conjecture.} As noted in the introduction, Haj\'os's conjecture implies the Erd\H{o}s-Gallai conjecture, as any $n$-vertex graph can be decomposed into an Eulerian graph and at most $n-1$ edges --- for example by taking the union of a maximal collection of disjoint cycles, and the edges in the remaining acyclic subgraph. Therefore, if Haj\'os's conjecture holds, then any $n$-vertex graph would have a decomposition into at most $\frac{n}{2}$ cycles and at most $n-1$ edges, and thus at most $\frac32n$ cycles and edges. It is known that only $(\frac{3}{2}+o(1))n$ cycles and edges are needed to decompose any $n$-vertex graph with linear minimum degree (due to Gir{\~a}o, Granet, K{\"u}hn, and Osthus~\cite{girao2021path}), and also that $\frac{3}{2}$ is best possible here. This latter fact was observed by Erd\H{o}s~\cite{erdHos1983some} in 1983 who remarked that there are graphs requiring $(\frac{3}{2}-o(1))n$ cycles and edges, likely referring to the following generalisation of an example of Gallai (see \cite{erdos1966representation}). Take $k\in \mathbb{N}$ and consider the complete bipartite graph $G$ with disjoint vertex classes $A$ and $B$ with $|A|=2k+1$ and $|B|=n-2k-1$. Each vertex in $|B|$ has odd degree in $G$, and $G[B]$ contains no edges, so any decomposition of $G$ into cycles and edges must contain at least $|B|$ edges. As each cycle in $G$ has length at most $2|A|$, there is thus no decomposition of $G$ into fewer than $$|B|+\frac{|A||B|-|B|}{2|A|}=\left(\frac{3}{2}-\frac{1}{2|A|}\right)|B|=\left(\frac{3}{2}-\frac{1}{4k+2}-o(1)\right)n$$ cycles and edges. \textbf{Our tools.} Many of our tools decompose an $n$-vertex graph into $O(n)$ cycles and a `leftover' subgraph, and might prove useful towards settling the conjecture in full. For example, the proof of Lemma~\ref{lem:density} shows that any $n$-vertex graph with no cycles longer than $t$ decomposes into $O(n)$ cycles and $O(n\log^{274}t)$ edges. Furthermore, for any constant $k$, running our iteration argument using \Cref{lem:density} $k+1$ times shows that any $n$-vertex graph can be decomposed into $O(kn)$ cycles and $O(n\log^{[k]}n)$ edges (see \eqref{eqn:Clogi}). This latter result reduces the Erd\H{o}s-Gallai conjecture to the case of arbitrarily sparse graphs, although this seems only to focus on the most difficult case. The main bottleneck in our argument seems to be the number of edges left uncovered in the almost decomposition into robust expanders, i.e.\ when applying \Cref{splitting-into-expanders}. It appears hard to reduce the number of edges enough to make an improvement on Theorem~\ref{thm:logstar} while getting enough properties in the expansion (robust or otherwise) to aid any cycle decomposition. \textbf{Potential further applications.} In addition to their application towards the Erd\H{o}s-Gallai conjecture, we believe the tools developed here could be useful in other settings due to the variety of applications that have been found for sublinear expansion since its introduction in its original form by Koml\'os and Szemer\'edi, as well as the only very recent use of the robust sublinear expansion discussed in Section~\ref{sec:generalexpansionchat}. In fact, one can use some of our ideas, specifically a more precise version of \Cref{prop:expansion-red-blue}, to give a simpler proof of a result of Tomon \cite{tomon2022robust} on finding rainbow cycles in properly coloured graphs, and give a different proof on a result of Wang \cite{wang2022rainbow} (itself an improvement of a result of Tomon~\cite{tomon2022robust}). In addition, Letzter~\cite{shobro} has very recently adapted some of our methods in order to find separating paths systems. \medskip \textbf{Acknowledgments.} We would like to thank Benny Sudakov, Jacob Fox, Stefan Glock and Shoham Letzter for helpful comments which improved this paper. The first author would like to gratefully acknowledge the support of the Oswald Veblen Fund.

https://arxiv.org/abs/1311.6338

New Bounds for the Traveling Salesman Constant

Let $X_1, X_2, \dots, X_n$ be independent and uniformly distributed random variables in the unit square $[0,1]^2$ and let $L(X_1, \dots, X_n)$ be the length of the shortest traveling salesman path through these points. In 1959, Beardwood, Halton $\&$ Hammersley proved the existence of a universal constant $\beta$ such that $$ \lim_{n \rightarrow \infty}{n^{-1/2}L(X_1, \dots, X_n)} = \beta \qquad \mbox{almost surely.}$$ The best bounds for $\beta$ are still the ones originally established by Beardwood, Halton $\&$ Hammersley $0.625 \leq \beta \leq 0.922$. We slightly improve both upper and lower bounds.

\section{Introduction and statement of results} For given points $x_1, x_2, \dots, x_n \subset [0,1]^2$, let $L(x_1, \dots, x_n)$ denote the length of the shortest traveling salesman path through all this points. It was realized early (e.g. Fejes \cite{fejes} in 1940, Verblunsky \cite{verb} in 1951 and Few \cite{few} in 1955) that there are uniform estimates $$ L(x_1, \dots, x_n) \leq c_1\sqrt{n}+c_2$$ for some constants $c_1, c_2$. If the points are chosen at random, one would expect a universality phenomenon: finding the optimal path is in some sense 'equivalent' to finding the optimal path through the points in many small subset of the unit square and then patching these together: the problem is self-similar on a smaller scale and this should imply an averaging effect. That this is indeed the case constitutes one of the first limit theorems in combinatorial optimization \cite{beard}. \begin{thm}[Beardwood, Halton \& Hammersley, 1959] Let $X_1, X_2, \dots, X_n, \dots$ be i.i.d. uniformly distributed random variables in $[0,1]^2$. There exists a universal constant $\beta$ such that $$ \lim_{n \rightarrow \infty}{\frac{L(X_1, \dots, X_n)}{\sqrt{n}}} = \beta$$ with probability 1. \end{thm} The statement is by now classic and very well-known (see, for example, the textbooks of Applegate, Bixby, Chvatal \& Cook \cite{apple}, Finch \cite{finch}, Gutin \& Punnen \cite{gut}, Steele \cite{steeletext} or Venkatesh \cite{venk} or even a popular-science book \cite{cook}). It is relatively easy to deduce that if the points $X_i$ are random following an absolutely continuous probability distribution $f(x)$ on $\mathbb{R}^2$, then $$ \lim_{n \rightarrow \infty}{\frac{L(X_1, \dots, X_n)}{\sqrt{n}}} = \beta\int_{[0,1]^2}{f(x)^{ \frac{1}{2}}dx}.$$ The Beardwood-Halton-Hammersley limit law is true for various other problems (e.g. minimal spanning tree, Steiner trees, ...) with a constant depending on the functional: a unified approach to the theory is given by Steele's limit theorem \cite{steele}. Interestingly and despite considerable effort, the constant is not known in any of the aforementioned cases. In case of the traveling salesman, Beardwood, Halton \& Hammersley themselves proved that $$ 0.625 = \frac{5}{8} \leq \beta \leq \beta_{BHH} \sim 0.92116\dots,$$ where $$\beta_{BHH} = 2\int_{0}^{\infty}{\int_{0}^{\sqrt{3}}{\sqrt{z_1^2+z_2^2}\exp{\left(-\sqrt{3}z_1\right)} \left(1-\frac{z_2}{\sqrt{3}}\right)dz_2} dz_1}.$$ It should be noted that Beardwood, Halton \& Hammersley actually claim to prove the better result $\beta \leq 0.92037 \dots$ (a statement reiterated in many different books and papers), however, their computation relies on numerical integration and we believe this to be the origin of the error: for the convenience of the reader, we have quickly surveyed their argument (and the integral to be evaluated) below. Despite the relative fame of the Beardwood-Halton-Hammersley theorem, there has been no improvement in the constant over the years; a series of papers \cite{johnson, percus, val} carrying out numerical estimates with large data sets suggest $\beta \sim 0.712$. The purpose of this paper is to draw some attention to the problem, describe the existing original arguments and to improve them. \begin{theorem} We have $$ \frac{5}{8} + \frac{19}{5184} \leq \beta \leq \beta_{BHH} - \varepsilon_0$$ for some explicit $$\varepsilon_0 > \frac{9}{16}10^{-6}.$$ \end{theorem} We have an explicit representation of $\varepsilon_0$ as an integral in $\mathbb{R}^7$: a concentration of measure effect turns Monte-Carlo estimates into a highly stable method and suggests that actually $$\varepsilon_0 \sim 0.0148\dots,$$ however, we consider the underlying idea to be of greater interest than the actual numerical improvement -- in addition, certain natural generalizations of our method should be able to give at least $\beta \leq 0.891$ if one assumes that certain integrals in high dimensions can be evaluated (details are given below). While additional improvements of the upper bound may lead to integrals whose evaluations become nontrivial, the approach is conceptually clear: further improving the lower bound, however, seems more challenging and in need of new ideas. As of this moment, we know of no methodical approach how this could be accomplished. \section{Proof of the Upper Bound} \subsection{Reduction to Poisson processes.} The core of the proof is in the stochastic treatment of $n$ random points in $[0,1]^2$ locally on the scale $n^{-1/2}$. At this scale the law of small numbers (see e.g. \cite{sv}) implies that the process behaves essentially like a Poisson process with intensity $n$. This property was exploited by Beardwood, Halton \& Hammersley; using their result, we can replace the $n$ random points with a Poisson process with intensity $n$, which simplifies further computations (this argument was pointed out to me by J. Michael Steele). \begin{lemma} Let $\mathcal{P}_n$ denote a Poisson process with intensity $n$ on $[0,1]^2$. Then $$ \lim_{n \rightarrow \infty}{\frac{\mathbb{E} L(\mathcal{P}_{n})}{\sqrt{n}}} = \beta.$$ \end{lemma} The idea is rather simple: the number of points in a Poisson process (i.e. the Poisson distribution) has mean $n$ and variance $n$. This means that we usually expect $\left|\#\mathcal{P}_{n} - n\right| \sim \sqrt{n}$, which is rather small compared to $n$. The expected length of a traveling salesman path lies somewhere between $\sim \beta\sqrt{n-\sqrt{n}}$ and $\sim \beta\sqrt{n+\sqrt{n}}$, the difference of which is $\sim 1$ and thus of smaller order -- if we now assume the Beardwood, Halton \& Hammersley result, this implies convergence for all cases concentrated here. Additionally, to deal with the other cases, it suffices to show that is very unlikely to have an unusually large amount of points and that the uniform bound of Few suffices. We leave the details to the interested reader and remark that the inverse statement (i.e. that the result for the Poisson process implies the desired result is actually due to Beardwood, Halton \& Hammersley). \subsection{The original argument.} This section describes the original argument due to Beardwood-Halton-Hammersley. It should be noted that a very similar argument was already used few years earlier by Few \cite{few}. Let $X$ be a Poisson process with intensity $n$ in the unit square $[0,1]^2$. We look at the set $$ X^* = \left\{x \in X: \pi_2(x) \leq \frac{\sqrt{3}}{\sqrt{n}}\right\}, $$ where $\pi_2$ is the projection onto the second component. Instead of asking for a traveling salesman tour through all the points, we merely ask for one through this particular strip. The entire unit square is cut into stripes and within each strip a local path gets constructed: in the end they all get connected to yield a fully valid traveling salesman path. The simplest solution locally within a strip is to order the points in $X^*$ with respect to the first coordinate, i.e. order them in such a way that $$ \pi_1(x_1) < \pi_1(x_2) < \pi_1(x_3) < \dots $$ and then simply connected the points in that order.\\ \begin{figure}[h!] \begin{tikzpicture}[xscale = 1, yscale =1] \draw [thick] (0,0) --(10,0); \draw [thick](0,1.5) --(10,1.5); \draw [thick] (0,0) --(0,1.5); \draw [thick] (10,0) --(10,1.5); \draw [fill] (1,0.5) circle [radius=0.05]; \draw (1,0.5) --(1.3,1.15); \draw [fill] (1.3,1.15) circle [radius=0.05]; \draw (2.5,0.15) --(1.3,1.15); \draw [fill] (2.5,0.15) circle [radius=0.05]; \draw (2.5,0.15) --(3,1.25); \draw [fill] (3,1.25) circle [radius=0.05]; \draw (3,1.25) --(3.3,0.75); \draw [fill] (3.3,0.75) circle [radius=0.05]; \draw (4.8,1.35) --(3.3,0.75); \draw [fill] (4.8,1.35) circle [radius=0.05]; \draw (4.8,1.35) --(5.5,1); \draw [fill] (5.5,1) circle [radius=0.05]; \draw (6.3,0.75) --(5.5,1); \draw [fill] (6.3,0.75) circle [radius=0.05]; \draw (6.3,0.75) --(7,0.15); \draw [fill] (7,0.15) circle [radius=0.05]; \draw (8,1) --(7,0.15); \draw [fill] (8,1) circle [radius=0.05]; \end{tikzpicture} \caption{A strip containing some points.} \end{figure} \begin{thm}[Beardwood, Halton \& Hammersley, 1959] Let $X$ be a Poisson process with intensity $n$ in $[0,1]^2$ and let $F$ be the length of the path constructed in the way described above. Then $$ \lim_{n \rightarrow \infty}{\frac{\mathbb{E}F(X)}{\sqrt{n}}} = 0.92116\dots$$ \end{thm} \begin{proof}[Sketch of the proof.] We restrict the Poisson process with intensity $n$ to the strip $\pi_2(x) \leq \sqrt{3}/\sqrt{n}.$ Then the real random variables $$ \left\{\pi_1(x) : \pi_2(x) \leq \frac{\sqrt{3}}{\sqrt{n}}\right\}$$ are distributed following a Poisson process with intensity $\sqrt{3n}$ on $[0,1]$. Ordering the points with respect to increasing first coordinate will give $x-$coordinates whose consecutive differences are exponentially distributed $$|\pi_1(x_{i+1}) - \pi_1(x_i)| \sim \sqrt{3n}\exp{\left(-\sqrt{3n}z\right)},$$ while the $y-$coordinates are i.i.d. distributed following the uniform distribution on $[0, \sqrt{3}/\sqrt{n}]$. Therefore, the expected distance in joining one point to the next is given by $$ \mathbb{E}\|x_i - x_{i+1}\| = 2n\int_{0}^{\infty}{\int_{0}^{\sqrt{3/n}}{\sqrt{z_1^2+z_2^2}\exp{\left(-\sqrt{3n}z_1\right)} \left(1-\frac{\sqrt{n}z_2}{\sqrt{3}}\right)dz_2} dz_1}.$$ Substitution allows to rewrite the integral as $$ \mathbb{E}\|x_i - x_{i+1}\| = \left(2\int_{0}^{\infty}{\int_{0}^{\sqrt{3}}{\sqrt{z_1^2+z_2^2}\exp{\left(-\sqrt{3}z_1\right)} \left(1-\frac{z_2}{\sqrt{3}}\right)dz_2} dz_1}\right)\frac{1}{\sqrt{n}}.$$ Since we are actually joining all $n$ points, we have to jump from one strip to another $\sim \sqrt{n/3}$ times and each time the jump is of order $\sim 1/\sqrt{n}$; this implies that the contribution coming from these jumps is of order $\mathcal{O}(1)$ and the total expected length is simply given by $n$ times the expected length of a single jump, which gives $$ \left(2\int_{0}^{\infty}{\int_{0}^{\sqrt{3}}{\sqrt{z_1^2+z_2^2}\exp{\left(-\sqrt{3}z_1\right)} \left(1-\frac{z_2}{\sqrt{3}}\right)dz_2} dz_1}\right)\sqrt{n} \sim 0.92116\sqrt{n}.$$ \end{proof} The underlying 'layer'-method is easily extended to higher dimensions and variable densities, see a paper of Borovkov \cite{bor}. \subsection{Changing variables.} The argument contains all the necessary ingredients for our improved local construction: following the steps outlined above, we will study Poisson processes with intensity $n$ in the strip $$ \left\{(x,y) \in \mathbb{R}^2: 0 \leq x \leq 1 \wedge 0 \leq y \leq \frac{\sqrt{3}}{\sqrt{n}}\right\},$$ which, following the same variable transformation as above, turns into studying local properties of the Poisson process with intensity 1 in the infinite strip $ \left\{(x,y) \in \mathbb{R}^2: 0 \leq y \leq \sqrt{3}\right\}.$ We construct the Poisson distribution indirectly in the following way: since we are interested in the lengths of paths through a local number of points and the strip has a translation symmetry, we may assume the first point to be given by $p_1 = (0, y_1)$, where $y_1$ is uniformly distributed on $[0, \sqrt{3}]$. Adding now iteratively exponentially distributed random variables with parameter $\sqrt{3}$ to the first variables and replacing the second component by independent uniformly distributed random variables in $[0, \sqrt{3}]$ yields the Poisson process with intensity $1$ in the strip. \subsection{Counting zigzags} The key observation in our improvement is the following: the Beardwood-Halton-Hammersley method is locally quite a bad if we encounter what we will informally call a zig-zag structure in the points: 4 consecutive points with a small difference in the $x-$coordinate but a large difference in the $y-$coordinate. More precisely, we will say that 4 points $p_1, p_2, p_3, p_4$ (ordered such that their $x-$coordinates increase) form a zigzag if $$ \| p_1 - p_3\| + \|p_3 - p_2\| + \|p_2 - p_4\| \leq \| p_1 - p_2\| + \| p_2 - p_3\| + \| p_3 - p_4\|.$$ Given a zigzag, it is advantageous to locally change the structure of the path. \begin{figure}[h!] \begin{tikzpicture}[xscale = 2, yscale = 2] \draw [thick] (0,0) --(2,0); \draw [thick](0,1) --(2,1); \draw [thick] (0,0) --(0,1); \draw [thick] (2,0) --(2,1); \draw [fill] (0.5,0.2) circle [radius=0.03]; \draw [fill] (0.7,0.85) circle [radius=0.03]; \draw [fill] (1,0.1) circle [radius=0.03]; \draw [fill] (1.4,0.9) circle [radius=0.03]; \draw [thick] (0.5,0.2) --(0.7,0.85) --(1,0.1) --(1.4,0.9); \draw [thick] (3,0) --(5,0); \draw [thick](3,1) --(5,1); \draw [thick] (3,0) --(3,1); \draw [thick] (5,0) --(5,1); \draw [fill] (3.5,0.2) circle [radius=0.03]; \draw [fill] (3.7,0.85) circle [radius=0.03]; \draw [fill] (4,0.1) circle [radius=0.03]; \draw [fill] (4.4,0.9) circle [radius=0.03]; \draw [thick] (3.5, 0.2) -- (4,0.1) --(3.7,0.85) --(4.4,0.9); \end{tikzpicture} \caption{Changing a zigzag path into something more effective.} \end{figure} We introduce some notation. Let $x_2, x_3, x_4$ be i.i.d. variables distributed according to the exponential law $\sqrt{3}e^{-\sqrt{3}z}$ and let $y_1, y_2, y_3, y_4$ be i.i.d. random variables uniformly distributed in $[0,\sqrt{3}]$. We define four random points via \begin{align*} p_1 &= (0, y_1) \qquad p_2 =(x_2, y_2) \qquad p_3 = (x_2+x_3, y_3) \qquad p_4 = (x_2 + x_3 + x_4, y_4). \end{align*} We know from the previous section that for all $1 \leq i \leq 3$ $$ \mathbb{E}\| p_i - p_{i+1}\| = \left(2\int_{0}^{\infty}{\int_{0}^{\sqrt{3}}{\sqrt{z_1^2+z_2^2}\exp{\left(-\sqrt{3}z_1\right)} \left(1-\frac{z_2}{\sqrt{3}}\right)dz_2} dz_1}\right) \sim 0.92\dots$$ Given these four points, we introduce a stochastic event $(A)$. \begin{align*} \| p_1 - p_3\| + \|p_3 - p_2\| + \|p_2 - p_4\| &\leq \| p_1 - p_2\| + \|p_2 - p_3\| + \|p_3 - p_4\| \qquad (A) \end{align*} Furthermore, we will introduce the respective (random) difference \begin{align*} X &= (\| p_1 - p_2\| + \|p_2 - p_3\| + \|p_3 - p_4\|) - (\| p_1 - p_3\| + \|p_3 - p_2\| + \|p_2 - p_4\|) \end{align*} \begin{lemma} We have $$ \mathbb{E}\left(X \big| A \right) \mathbb{P}\left(A\right) \geq \frac{9}{4}10^{-6}.$$ \end{lemma} \begin{proof} Since we are only trying to show a positive lower bound, rough estimates suffice. We study the event $B$ defined as $$ \left(x_2 \leq \frac{\sqrt{3}}{9}\right) \wedge \left(x_3 \leq \frac{\sqrt{3}}{9}\right) \wedge \left(x_4 \leq \frac{\sqrt{3}}{9}\right) \wedge \left(\min(y_1, y_3) \geq \frac{8\sqrt{3}}{9}\right) \wedge \left(\max(y_2, y_4) \leq \frac{\sqrt{3}}{9}\right).$$ \begin{figure} \begin{tikzpicture}[scale = 3] \draw [thick] (0,0) --(0.7,0); \draw [thick] (0,0) --(0,1); \draw [thick] (0,1/9) --(0.7,1/9); \draw [thick] (0,2/9) --(0.7,2/9); \draw [thick] (0,3/9) --(0.7,3/9); \draw [thick] (0,4/9) --(0.7,4/9); \draw [thick] (0,5/9) --(0.7,5/9); \draw [thick] (0,6/9) --(0.7,6/9); \draw [thick] (0,7/9) --(0.7,7/9); \draw [thick] (0,8/9) --(0.7,8/9); \draw [thick] (0,9/9) --(0.7,9/9); \draw [thick] (1/9,0) --(1/9,1); \draw [thick] (2/9,0) --(2/9,1); \draw [thick] (3/9,0) --(3/9,1); \draw [thick] (4/9,0) --(4/9,1); \draw [fill] (0.05,0.97) circle [radius=0.02]; \draw [fill] (0.16,0.05) circle [radius=0.02]; \draw [fill] (0.27,0.92) circle [radius=0.02]; \draw [fill] (0.37,0.06) circle [radius=0.02]; \end{tikzpicture} \caption{An instance of the event $B$.} \end{figure} These variables are independent and all distributions are explicitely given: thus, for $2 \leq i \leq 4$, we have $$\mathbb{P}\left(x_i \leq \frac{\sqrt{3}}{9}\right) = \int_{0}^{\frac{\sqrt{3}}{9}}{\sqrt{3}e^{-\sqrt{3}z}dz} = 1-\frac{1}{e^{\frac{1}{3}}},$$ while $\mathbb{P}(\min(y_1, y_3) \geq \frac{8\sqrt{3}}{9}) = \mathbb{P} (\max(y_2, y_4) \leq \frac{\sqrt{3}}{9}) = 1/81$. $$\mathbb{P}(B) = \left(1-\frac{1}{e^{\frac{1}{3}}}\right)^3\left(\frac{1}{81}\right)^2 \geq 3 \cdot 10^{-6}.$$ At the same time, a simple computation yields that in the event $B$, we always have $$ X \geq \frac{3}{4}.$$ This implies that the event $B$ is a subset of the event $A$ and, trivially, Therefore $$ \mathbb{E}\left(X \big| A \right) \mathbb{P}\left(A\right) \geq \mathbb{E}\left(X \big| B \right) \mathbb{P}\left(B\right) \geq \frac{9}{4}10^{-6}.$$ \end{proof} \begin{proof}[Proof of the upper bound.] We follow the original idea of Beardwood, Halton \& Hammersley and partition the unit square into strips: since we are dealing with a Poisson process, the behavior within each strip is independent of that in all other strips; focusing on one strip, we are dealing with a Poisson process of intensity $n$. For any set of random points arising from the Poisson process, we order them with increasing $x-$coordinate $$\pi_1(x_1) \leq \pi_1(x_2) \leq \dots \leq \pi_1(x_k), $$ and consider the $4-$tuples $(x_1, x_2, x_3, x_4)$, $(x_5, x_6, x_7, x_8)$, and so on (with possibly up to 3 points left at the end of each strip). Whether or not any of these $4-$tuples contains a zigzag structure is an independent event: the computations in the previous section then imply that with probability at least $3 \cdot 10^{-6}$ a zigzag yielding a gain of at least $3/(4\sqrt{n})$ is present. There are are $n/4 - O(\sqrt{n})$ $4-$tuples to consider implying the gain in length to be at of order $(n/4-O(\sqrt{n}))(3/(4\sqrt{n}))(3 \cdot 10^{-6})$ and thus $$ \beta \leq \beta_{BHH} - \frac{9}{16}10^{-6}.$$ \end{proof} \textbf{Remark.} The gain in length was achieved by looking at $n/4-O(\sqrt{n})$ independent events: usual arguments would allow us to conclude that the predicted gain in length is actually tightly concentrated around its mean. This, however, is not necessary for our sort of argument: we already know that $\beta$ describes the limiting behavior almost surely: the expected length of any construction of deterministic paths is then necessarily an upper bound on $\beta$.\\ \textbf{Remark.} These problems exhibit a concentration of measure phenomenon implying the stability of Monte-Carlo estimates, which will then usually imply much stronger results. For comparison, we did ten samples of a million random points each, which suggests $$ \mathbb{P}\left(A\right) \sim 0.1418, \quad \mathbb{E}\left(X \big| A \right) \sim 0.4187 \qquad \mbox{and thus} \qquad \mathbb{E}\left(X \big| A \right) \mathbb{P}\left(A\right) \geq 0.059.$$ with a standard deviation of 0.0003 and 0.001, respectively. This would imply that indeed $$ \beta \leq 0.90632.$$ \subsection{Numerical estimates.} Our result was aimed towards the clearest presentation of the idea. Improvements of the idea are rather obvious, however, they require somewhat accurate bounds for certain finite-dimensional integrals. One particular generalization is as follows: one could study not merely zigzags but all 24 possible paths through six points leaving the first and the last point invariant; let us consider all 24 permutations over the symbols $\left\{2,3,4,5\right\}$ and denote the existence of an improved path as the stochastic event $(C)$ $$ \inf_{\pi \in S_4(\left\{2,3,4,5\right\})}{\|p_1 - p_{\pi(2)}\| + \sum_{i=2}^{4}{\|p_{\pi(i+1)} - p_{\pi(i)}\|} +\|p_{6}-p_{\pi(5)}\|} < \sum_{i=1}^{5}{\|p_{i+1}-p_{i}\|} \qquad \qquad (C)$$ and the respective improvement by $$ Z = \sum_{i=1}^{5}{\|p_{i+1}-p_{i}\|} - \inf_{\pi \in S_4(\left\{2,3,4,5\right\})}{\|p_1 - p_{\pi(2)}\| + \sum_{i=2}^{4}{\|p_{\pi(i+1)} - p_{\pi(i)}\|}}-\|p_{6}-p_{\pi(5)}\|.$$ Monte-Carlo methods (10 samples of 50000 sets of points each) suggest that $$ \mathbb{P}\left(C\right) \sim 0.3721 \qquad \mbox{and} \qquad \mathbb{E}\left(Z \big| C \right) \sim 0.4990$$ with a standard deviation of 0.02 and 0.004, respectively. These values would suggest $\beta \leq 0.8902$. As already hinted at in the introduction, there is a natural limit to these improvements: we study paths through random points with an additional restriction on their movement in one of the two dimensions, which corresponds to a different functional and this difference will be a great hindrance to further major improvements. \section{Proof of the lower Bound} \subsection{The original argument.} Proving an upper bound can (and has) been done by constructing an explicit path. Proving a lower bound has to pursue an entirely different strategy since we have very little idea what an optimal path could look like: we already know, however, that it is sufficient to prove lower bounds on the expected length of the traveling salesman path through points of a Poisson process with intensity $n$ in $[0,1]^2$. The only real basic information about paths at our disposal is that for every point there are two points to which that particular point is connected: suppose now that for every point, these two points are also the two closest points. \\ The second remark is that we may assume that the Poisson process is actually distributed with the intensity $n$ on all of $\mathbb{R}^2$: adding more points can only decrease the expected distance and allows us to disregard the behavior of the process close to the boundary of $[0,1]^2$. The following Lemma can be found in many basic books on probability theory. \begin{lemma} Let $P_n$ be a Poisson process on $\mathbb{R}^2$ with intensity $n$. Then for any fixed point $p \in \mathbb{R}^2$, the probability distribution of the distance between $p$ and the nearest point in $\mathcal{P}_n$ is given by $$ f(r) = 2\pi n r e^{-\pi n r^2}$$ and the distance to the second-nearest neighbour is given by $$ g(r) = 2\pi^2 n^2 r^3 e^{-\pi n r^2}.$$ \end{lemma} \begin{proof} We compute the probability of the closest point point lying at distance $(r, r+\varepsilon)$. This is precisely the case if there is \textit{no} point in a disk of radius $r$ around $p$ but at least 1 point in $(r, r+\varepsilon)$. It follows from the definition of the Poisson process that the probability of there being no point in the disk is given by $e^{-n \pi r^2}$ and thus $$ f(r) = \lim_{\varepsilon \rightarrow 0}{\frac{e^{-n \pi (r+\varepsilon)^2} - e^{-n \pi r^2}}{\varepsilon}} = 2r\pi n e^{-n \pi r^2}.$$ We compute the other expression in the same way: we require that there is precisely one point with distance at most $r$ and another point at distance $(r, r+\varepsilon)$. The probability of being precisely one point at distance at most $r$ is given by $ r^2\pi n e^{-r^2 \pi n}$ while the area of the annulus is simply $((r+\varepsilon)^2-r^2)\pi$ and the probability of one point being in there is $((r+\varepsilon)^2-r^2)\pi n e^{-((r+\varepsilon)^2-r^2)\pi n}.$ Altogether, we have $$ g(r) = (r^2\pi n e^{-r^2 \pi n}) \lim_{\varepsilon \rightarrow 0^+}{\frac{((r+\varepsilon)^2-r^2)\pi n e^{-((r+\varepsilon)^2-r^2)\pi n}}{\varepsilon}} = 2\pi^2 n^2 r^3 e^{-\pi n r^2}.$$ \end{proof} Standard calculations give that the distance $r$ to the nearest point has expectation $$ \int_0^{\infty}{rf(r)dr} = \frac{1}{2\sqrt{n}}$$ while the distance to the next-to-nearest point has expectation $$ \int_0^{\infty}{rg(r)dr} = \frac{3}{4\sqrt{n}}.$$ Given a traveling salesman path, every point is connected to two other points -- in the worst case, these are the nearest and the next-to-nearest point in all cases, yielding a lower bound of $$ \beta \geq \left(\frac{1}{2}+\frac{3}{4}\right)\frac{1}{2} = 0.625,$$ which is the original result of Beardwood, Halton \& Hammersley. \subsection{An improvement.} The previous argument assumed that it is always the worst case that occurs: every point is connected to its two closest neighbors. This, however, is not possible if we have the following constellation of points: a point $a$ with closest point $b$ at distance $r_1$ and its second-closest point $c$ at distance $r_2 > r_1$ and third-closest point $d$ at distance $r_3 > r_1 + 2r_2$. Our proof rests on an analysis of this situation. \begin{lemma} Let $P_n$ be a Poisson process on $\mathbb{R}^2$ with intensity $n$. Then for any fixed point $p \in \mathbb{R}^2$, the probability distribution of the distance between $p$ and the closest, second closest and third closest point is given by $$ h(r_1, r_2, r_3) = \begin{cases} e^{-n \pi r_3^2}(2n\pi)^3r_1 r_2 r_3 \qquad &\mbox{if~}r_1 < r_2 < r_3 \\ 0 \qquad &\mbox{otherwise.} \end{cases}$$ \end{lemma} \begin{proof} As before, we study the probability of precisely one point at distance $(r_1, r_1+\varepsilon)$ (event A), precisely one point in $(r_2, r_2+\varepsilon)$ (event B) and precisely one point in $(r_3, r_3+\varepsilon)$ (event C) and no points in between (event D). The probabilities for these events including their expansion up to first order in $\varepsilon$ are \begin{align*} \mathbb{P}(A) &= ((r_1+\varepsilon)^2-r_1^2)n\pi e^{(-(r_1+\varepsilon)^2+r_1^2)n\pi} = 2n\pi\varepsilon r_1 + O(\varepsilon^2)\\ \mathbb{P}(B) &= ((r_2+\varepsilon)^2-r_2^2)n\pi e^{(-(r_2+\varepsilon)^2+r_2^2)n\pi} = 2n\pi\varepsilon r_2 + O(\varepsilon^2)\\ \mathbb{P}(C) &= ((r_3+\varepsilon)^2-r_3^2)n\pi e^{(-(r_3+\varepsilon)^2+r_3^2)n\pi} = 2n\pi\varepsilon r_3 + O(\varepsilon^2)\\ \mathbb{P}(D) &= e^{-r_1^2\pi n}e^{(-r_2^2+(r_1+\varepsilon)^2)\pi n}e^{(-r_3^2+(r_2+\varepsilon)^2)\pi n} = e^{-n\pi r_3^2} + P(\varepsilon). \end{align*} This immediately implies the statement. \end{proof} \begin{proof}[Proof of the lower bound.] We start by showing that both the nearest as well as the next-to-nearest point of any element in $\left\{a,b,c\right\}$ also lies in the set. Let $x$ be some other point with $x \notin \left\{a,b,c\right\}$. Then $$ \| b - x\| \geq \|a - x\| - r_1 \geq r_3 - r_1 > 2r_2 \geq r_1 + r_2 \geq \|b-c\|$$ and therefore the second closest point from $b$ is $a$ or $c$. By the same token $$ \| c - x\| \geq \|a - x\| - r_2 \geq r_3 - r_2 \geq r_1 + r_2 \geq \|b-c\|$$ and therefore the second closest point from $c$ is $a$ or $b$. Cases of equality have probability 0 and can be ignored. \begin{figure}[h!] \begin{tikzpicture}[xscale = 2, yscale = 2] \draw [fill] (0.5,0.2) circle [radius=0.03]; \draw [fill] (1,0.8) circle [radius=0.03]; \draw [fill] (1.7,0.1) circle [radius=0.03]; \draw [fill] (4.5,0.5) circle [radius=0.03]; \draw [thick] (0.5,0.2) --(1,0.8) -- (1.7,0.1); \draw (1,0.9) node{a}; \draw (0.4,0.2) node{b}; \draw (1.8,0.1) node{c}; \draw (0.6,0.6) node{$r_1$}; \draw (1.5,0.5) node{$r_2$}; \draw (4.64,0.5) node{d}; \end{tikzpicture} \caption{$a$ and the three closest points of $a$.} \end{figure} If we simply connect every point to its two closest neighbours, we end up with a triangle where every point is connected to the two other points but no other point except those. This is clearly not possible for a traveling salesman path. Let us first compute the frequency of such an event. Using the Lemma, the probability of all of these distance relations being true for a fixed point $a$ is $$ \int_{0}^{\infty}{\int_{r_1}^{\infty}{\int_{r_1+2r_2}^{\infty}{e^{-n \pi r_3^2}(2n\pi)^3r_1 r_2 r_3 dr_3}dr_2}dr_1} = \frac{7}{324}.$$ There is a lack of independence: if it is true for $a$, it is likely to be true for $b$ and $c$ as well -- thus, we have only the trivial bound $$ \frac{1}{3}\frac{7}{324}n = \frac{7n}{972}$$ on the number of triples of points with this property. However, if the case occurs, then the algorithm connecting every point to its two nearest neighbours has an expected length which can be bounded from above by $$ r_1 + r_2 +2||a-c|| \leq 3(r_1 + r_2).$$ where the distance $\|a-c\|$ has to be counted twice because the algorithm cannot 'see' that it has created a triangle and counts the distance twice. \begin{figure}[h!] \begin{tikzpicture}[xscale = 0.8, yscale = 0.8] \draw [fill] (0,0) circle [radius=0.06]; \draw (-0.2,-0.1) node{a}; \draw [fill] (3,1) circle [radius=0.06]; \draw (3.2,0.8) node{b}; \draw [thick] (0,0) -- (3,1); \draw (1.6,0.8) node{$r_1$}; \draw [fill] (3,-1) circle [radius=0.06]; \draw (3.1,-0.8) node{c}; \draw [thick] (0,0) -- (3,-1); \draw (1.6,-0.8) node{$r_2$}; \draw [fill] (6,2) circle [radius=0.06]; \draw (6.2,1.8) node{d}; \draw [dashed] (3,1) -- (6,2); \draw [fill] (6,-2) circle [radius=0.06]; \draw (6.2,-1.9) node{e}; \draw [dashed] (3,-1) -- (6,-2); \end{tikzpicture} \caption{The best of the worst case.} \end{figure} In the case of three points isolated from the rest, there is one special case which is the easiest to connect to the remaining points: this is is when $b$ can be connected to a point $d$ having distance $r_3$ from $a$ and $c$ can be connected to a different point $e$ also at distance $r_3$ from $a$ and, additionally, $b$ lies on the line $\overline{ad}$ and $c$ lies on $\overline{ae}$. In this case, the required length is $$\|d-b\| + \|b-a\| + \|a-c\| + \|c-e\| \geq (r_3-r_1) + r_1 + r_2 + (r_3-r_2) = 2r_3.$$ This implies that whenever we are in this particular configuration, the actual path has to be at least a length $2r_3 - 3(r_1 + r_2)$ longer than what the greedy algorithm suggests. Note that $$ 2r_3 - 3(r_1 + r_2) \geq r_2-r_1 \geq 0$$ and that we always gain something at this point. In expectation, this is an average length of $$ \frac{324}{7}\int_{0}^{\infty}{\int_{r_1}^{\infty}{\int_{r_1+2r_2}^{\infty}{(2r_3-3r_1-3r_2)e^{-n \pi r_3^2}(2n\pi)^3r_1 r_2 r_3 dr_3}dr_2}dr_1} = \frac{57}{112}\frac{1}{\sqrt{n}}.$$ Altogether, this gives the lower bound $$ \beta \geq \left(\frac{1}{2}+\frac{3}{4}\right)\frac{1}{2} + \frac{7}{972}\frac{57}{112} = \frac{5}{8} + \frac{19}{5184}.$$ \end{proof} \textbf{Acknowledgments.} I am grateful to Steven Finch for comments on the history of the problem and indebted to J. Michael Steele for encouraging me to write this paper and a series of very valuable discussions. The author was supported by SFB 1060 of the DFG.

https://arxiv.org/abs/2010.08672

On Banzhaf and Shapley-Shubik Fixed Points and Divisor Voting Systems

The Banzhaf and Shapley-Shubik power indices were first introduced to measure the power of voters in a weighted voting system. Given a weighted voting system, the fixed point of such a system is found by continually reassigning each voter's weight with its power index until the system can no longer be changed by the operation. We characterize all fixed points under the Shapley-Shubik power index of the form $(a,b,\ldots,b)$ and give an algebraic equation which can verify in principle whether a point of this form is fixed for Banzhaf; we also generate Shapley-Shubik fixed classes of the form $(a,a,b,\ldots,b)$. We also investigate the indices of divisor voting systems of abundant numbers and prove that the Banzhaf and Shapley-Shubik indices differ for some cases.

\section{Introduction} In a weighted voting system, voters, or \textit{players}, have different amounts of the total votes, which are called \textit{weights}. A \textit{motion} is an agenda item that needs some amount of votes to be passed. This amount is called the \textit{quota}. A \textit{coalition} is a group of players, and the sum of all votes in a coalition is called the total voting power. A coalition is a \textit{winning coalition} if its total voting power meets or exceeds the quota. A player is called a \textit{critical player} in a winning coalition if the coalition would not be winning without the player. We examine the \textit{Banzhaf power index} \cite{Banzhaf} and the \textit{Shapley-Shubik power index} \cite{Shapley-Shubik}, which are two different methods of measuring a player's strength in a system. The Banzhaf power index of a player is the number of times that player is a critical player in all winning coalitions divided by the number of total times any player is a critical player. The Shapley-Shubik index looks at permutations of all players in a system, called \textit{sequential coalitions}. We sum each player's votes starting from the beginning of a sequential coalition, and see if the sum reaches the quota as we progress. The player whose votes first cause this sum to meet or exceed the quota is called a \textit{pivotal player}. The Shapley-Shubik power index of a player is the number of times that player is a pivotal player divided by the total number sequential coalitions. We present results in two different directions. The first is slightly more number-theoretic, while the second relies heavily on modelling with suitable equivalent algebraic equations. \medskip For an integer $n> 1$, the \textit{divisor voting system of $n$} is $[Q: n, d_k, \ldots, d_1, 1]$ where $d_1,\ldots,d_k$ are the divisors of $n$ distinct from $1$ and $n$; also $Q = (\sigma(n)+1)/2$ when $\sigma(n)$ is even and $Q= \sigma(n)/2$ when $\sigma(n)$ is odd -- so the quota is essentially a simple majority rule. Here $\sigma(n)$ denotes the sum of the divisors of $n$. In Section 2 we will show that for any positive integer $n$ where $\sigma(n)\geq 2n$, the Banzhaf power index and Shapley-Shubik index disagree on at least one divisor for the divisor voting system of $n$ for all $n$ such that $\sigma(n) = 2n+k$, $0\leq k \leq 5$. It is unlikely that such an approach would work in the general case and so it is more plausible to examine an inductive argument of the following sort -- given $n$, what can we say about $pn$ and $mn$, where $p$, $m$ are primes? We have some progress towards this approach in Section $2$. \medskip Now let us introduce the algebraic direction of our research in Section 3. Consider systems of the form $V = [1/2_s : a_1,\ldots, a_n]$ where $a_1,\ldots,a_n \geq 0$ are real numbers that sum to $1$ and all winning coalitions are those whose sum \textit{strictly exceeds} $1/2$. We can evaluate the Banzhaf and Shapley-Shubik powers of each player, thus obtaining tuples $BB(V)$ and $SS(V)$. We say that $V$ is a \textit{Banzhaf fixed point} if $V = BB(V)$ and a \textit{Shapley-Shubik fixed point} if $V = SS(V)$. A fixed point is \textit{primitive} if there is no player $i$ with $a_i = 0$. A fixed point is \textit{non-trivial} if not all players have the same power. In Section $3$ we derive algebraic equations whose solutions generate fixed points and in particular in Section $3.2$ we give new classes of Shapley-Shubik fixed points. We then categorize all fixed points under Shapley-Shubik and use this to show that there is only one fixed point for both Banzhaf and Shapley-Shubik. Note that in general the problem is difficult to model with an equation and so we will restrict to the (already far from trivial) tuples of one of the forms $(a,b,\ldots,b)$ and $(a,a,b,\ldots,b)$. \section{Divisor Voting Systems and Abundant Numbers} Our first main aim in this section is to prove the following. \begin{proposition} The Banzhaf power index and Shapley-Shubik index are different for at least one divisor for the divisor voting system of $n$ when $\sigma(n) = 2n+k$, $0\leq k \leq 5$. \end{proposition} An integer $n$ is called \textit{abundant} if $\sigma(n) > 2n$. We shall sometimes refer to the following auxilliary observation. \begin{proposition} \label{abundantdivisors} Any abundant number must have more than 5 divisors, and the only abundant number with 6 divisors is 20. \end{proposition} \begin{proof} Let $g: \mathbb{Z}\rightarrow\mathbb{R}$ be the function defined by $g(n) = \frac{n}{\phi{(n)}}$. By substituting Euler's product formula for $\phi{(n)}$, we have $g(n)={\displaystyle \prod_{p \mid n} \frac{p}{p-1}}$. Suppose $n$ is an odd abundant number. Then, since $n$ is abundant, $\frac{\sigma{(n)}}{n} > 2$. Since $g(n) > \frac{\sigma{(n)}}{n}$ when $n>1$, $g(n) > 2$. Now we will find the minimum number of divisors $n$ can have. The function $f(x)=\frac{x}{x-1}$ is a decreasing function over the interval $[2, \infty)$. Therefore, the two prime odd numbers for which $\frac{p}{p-1}$ is at its largest values are $3$ and $5$. However, $(\frac{3}{2})(\frac{5}{4}) < 2$. Therefore, $n$ must have at least 3 prime odd divisors to be abundant, or 8 total divisors. Suppose $n$ is an even abundant number. Since $n$ is abundant, $g(n) >2$. In this case, it is possible for $n$ to have only 2 prime divisors. These prime divisors must be 2 and 3, since $(\frac{2}{1})(\frac{3}{2}) > 2$. 2 and 3 are the only prime divisors for which this is possible, since $f(x) = \frac{x}{x-1}$ is decreasing for $x\geq 2$, and $(\frac{3}{2})(\frac{5}{4}) < 2$. 20 is the only abundant number with only 2 and 3 as prime divisors. Hence, 20 is the only abundant number with 6 divisors, and all other even abundant numbers must have at least 8 divisors. \end{proof} \subsection{Casework -- Proof of Proposition $2.1$} In this part $d$ is the number of divisors of $n$. We will prove Proposition $2.1$ by breaking it into cases based on the value of $\sigma{(n)}$. \subsubsection{$\sigma(n)=2n$} \label{subsubsection:sigma(n)=2n} \begin{proof} We will first find the Banzhaf power index of each divisor of $n$ based on the number of divisors $d$. We can calculate the number of times $n$ is a critical player in the divisor voting system of $n$. Consider the total number of possible combinations of the divisors that include $n$, $2^{d-1}$. Since the sum of all divisors excluding $n$ is also $n$, and the quota is $n+1$, the only combination that is not a winning coalition is the coalition that contains only $n$. Hence $n$ is a critical player $2^{d-1} -1$ times. Every divisor that is not $n$ is a critical player only when it is in the winning coalition consisting of $n$ and itself. So, the Banzhaf index for $n$ is $\frac{2^{d-1}-1}{2^{d-1}+(d-2)}$. The Banzhaf index for any other divisor $d_i$ of $n$ is $\frac{1}{2^{d-1}+(d-2)}$. We will now calculate the Shapley-Shubik power index for each divisor of $n$. A proper divisor $d_i$ of $n$ can be a pivotal player only once. It is a pivotal player in the sequential coalition where $n$ is the first player followed by $d_i$. Hence, the number of times $d_i$ is a pivotal player is $(d-2)!$. We can now find the number of times $n$ is a pivotal player. There are $d-1$ proper divisors of $n$ and $d!$ total pivotal players, so the number of times $n$ is a pivotal player is $d!-(d-2)!(d-1)$. So $n$ has a Shapley-Shubik index of $\frac{d!-(d-1)!}{d!}$ and every other divisor of $n$ has index $\frac{(d-2)!}{d!} = \frac{1}{d(d-1)}$. We now compare the Shapley-Shubik indices and the Banzhaf indices to show that they differ for at least one divisor of $n$. We can show that each proper divisor of $n$, $d_i$, has a Banzhaf index that differs from its Shapley-Shubik index by setting equal the their formulas we derived. So, we have \begin{equation*} \frac{1}{2^{d-1}+(d-2)}=\frac{1}{d(d-1)}. \end{equation*} The only integer solution to this equation is $d=2$ -- indeed, note that $2^{d-1} > d^2 - 2d + 2$ for $d\geq 6$ by induction, with the inductive step being $2^d = 2 \cdot 2^{d-1} > 2(d^2 - 2d + 2) > d^2 + 2 = (d+1)^2 - 2(d+1) + 2$ (the second inequality is equivalent to $d(d-4) + 2 > 0$). However, since $n$ is an abundant number, it must have more than six divisors by Proposition \ref{abundantdivisors}. Thus, for every perfect number, at least one of its divisors in its divisor voting system will have a Banzhaf index that differs from its Shapley-Shubik index. \end{proof} We followed a similar procedure to find and compare the Banzhaf and Shapley-Shubik power indices for each divisor in the divisor voting system of $n$ where $\sigma(n)=2n+k$. Note that we will mostly omit the derivations of the calculations, as well as reasons why two concrete expressions involving $d$ are equal apart from a few cases, since the ideas are largely the same. \subsubsection{$\sigma(n)=2n+1$} Since both the Banzhaf and Shapley-Shubik power indices of 1 are 0, we must compare the Banzhaf and Shapley-Shubik power index formulas for proper divisors $d_i$ that are not 1. Using the same method that used in 2.1.1, we can see that the formula for the Banzhaf index of each $d_i$ is $\frac{2}{2^{d-1}+2(d-2)}$. The formula for the Shapley-Shubik index of each $d_i$ is $\frac{2(d-2)!}{d!} = \frac{2}{d(d-1)}$. These two formulas are equal only when the number of divisors of $n$ is 2, 3, or 4, which is not possible since $n$ is an abundant number. Note that numbers $n$ where $\sigma(n)=2n+1$, are called \textit{quasiperfect numbers}, and it is unknown if any such numbers really exist \cite{quasiperfect}. \subsubsection{$\sigma(n)=2n+2$} Both the Banzhaf and Shapley-Shubik power indices of 1 do not equal 0. Therefore we can compare the Banzhaf and Shapley-Shubik indices for 1 to prove that 2.0.1 is true for this case. The formula for the Banzhaf index of 1 is $\frac{1}{2^{d-1}+3(d-2)-2}$ and the formula for the Shapley-Shubik index of 1 is $\frac{(d-2)!}{d!} = \frac{1}{d(d-1)}$. If these two expressions are equal to each other, then $d$ must be 4, but an abundant number must have at least 6 divisors. Hence when $\sigma(n)=2n+2$, the power indices of 1 in Shapley-Shubik and Banzhaf differ from one another. \subsubsection{$\sigma(n)=2n+3$} In this case, since both the Banzhaf and Shapley-Shubik power indices of 1 are 0, we must compare the power indices of divisors $d_i$ that are not 1, nor $n$. Then the Banzhaf index of each $d_i$ is $\frac{4}{2^{d-1}+4(d-2)-4}$ and the Shapley-Shubik index of $d_i$ is $\frac{4(d-3)!+2(d-2)!}{d!} = \frac{2}{(d-1)(d-2)}$. There are no integer solutions for when the two indices of $d_i$ are equal -- indeed, we can prove that $2^{d-1} > 2d^2 - 10d+16$ for $d\geq 6$ with induction. Suppose that the inequality is true for $d$. By manipulating this inequality, we have that $2^d = 2\cdot 2^{d-1} > 2(2d^2 - 10d+16) > 4d^2 - 20d - 32 = 2(d+1)^2 - 10(d+1)+16$ (the second inequality is equivalent to $2(d-4)(d-3) > 0$). Therefore, since the inequality is true for $d+1$, it is true for all $d \geq 6$. Hence, since an abundant number must have at least 6 divisors, there are no solutions for when the two indices are equal. \subsubsection{$\sigma(n)=2n+4$} We must consider two cases for when $n$ is even or odd. Suppose $n$ is even. Since the index of 1 is not 0, we can simply compare the Banzhaf and Shapley-Shubik indices for 1. The Banzhaf index of 1 is $\frac{1}{2^{d-1}+5(d-3)-1}$. The Shapley-Shubik index of 1 is $\frac{2(d-3)!}{d!} = \frac{2}{d(d-1)(d-2)}$. When we set these two expressions for the Banzhaf and Shapley-Shubik index equal and simplify, we have $d^3+2d-3d^2 = 2^d + 10d-32$. We will show that these expressions differ since $n$ is an abundant number and must have more than 5 divisors. Let $h: \mathbb{R}\rightarrow\mathbb{R}$ be the function defined by $h(x) = 2^x + 10x-32-(x^3+2x-3x^2)$. Thus $h(x) = 2^x-x^3 + 3x^2 + 8x-32$. By Proposition \ref{abundantdivisors}, an abundant number must have at least 6 divisors. The integers 6, 7, and 8 generate values of $h(x)$ that are not equal to 0. To show that $h(x) > 0$ for values $x\geq 9$, it is sufficient to show that $h(9) > 0$ and $h'(x) = 0$, which is true if $h^{(3)}(9) > 0$ and $h^{(4)}(x)>0$ for all $x\geq 9$. $h^{(3)}(x) = 2^xln^3(2)-6$, which is greater than 0 for all numbers greater than or equal to 9, and $h^{(4)}(x) = 2^xln^4(2)$, which is clearly greater than 0 for all values of $x$. Hence the two expressions for the Banzhaf and Shapley-Shubik index of 1 are never equal for an abundant number $n$. Now suppose $n$ is odd. Since the index of 1 is 0 for both Banzhaf and Shapley-Shubik indices, we must compare the formulas for the indices of proper divisors $d_i$ that are not 1. The formula for the Banzhaf index of each $d_i$ is $\frac{4}{2^{d-1}+4(d-2)-3}$, and the formula for the Shapley-Shubik index of $d_i$ is $\frac{2[(d-2)\cdot (d-2)!]+4(d-2)!}{d!} = \frac{2}{d-1}$. By simplifying $\frac{4}{2^{d-1}+4(d-2)-3} = \frac{2}{d-1}$ we have $2^{d-1}=-2d+9$. However, $2^{d-1} > -2d+9$ for all $d > 5$, since the graph $y=-2x+9$ intersects the x-axis only when $x<5$, and $y=2^{d-1}$ is an increasing function when $x\geq 0$. Therefore, since $n$ is an abundant number and must have at least $6$ divisors, there are no abundant numbers for which these expressions are equal. \subsubsection{$\sigma(n)=2n+5$} We again have two cases to consider for when $n$ is even or odd. Suppose $n$ is even. The Banzhaf and Shapley-Shubik indices of 1 are not 0, so we can simply compare these two indices for 1. The formula for the Banzhaf index of 1 is $\frac{1}{2^{d-1}+5(d-3)-2}$ and the formula for the Shapley-Shubik index of 1 is $\frac{2(d-3)!}{d!} = \frac{2}{d(d-1)(d-2)}$. Equating the expressions for the Banzhaf and Shapley-Shubik indices gives $2^d+10d-34 = d^3+2d-3d^2$. $n$ is an abundant number and thus must have at least 6 divisors by Proposition 2.2. Hence we will show that this equation has no integers solutions greater than 5 using the same method as in 2.1.5. Let $h: \mathbb{R}\rightarrow\mathbb{R}$ be the function defined by $h(x) = 2^x+10x-34 - (x^3+2x-3x^2)$. Hence $h(x) = 2^x-x^3+3x^2+8x-34$. We can manually verify that $h(6)$, $h(7)$, and $h(8)$ are not equal to 0. We will now show that if $x$ is greater than or equal to 9, $h(x)$ is greater than 0. We can demonstrate this by showing that $h^{(3)}(9)>0$ and $h^{(4)}(x) > 0$ for all $x\geq9$. $h^{(3)}(9) = 2^9\ln{(2)}^{3} - 6$, which is greater than 0 and $h^{(4)}(x) = 2^x\ln{(2)}^4$, which is also clearly greater than 0. Hence there are no integer solutions for when the Banzhaf index of 1 and the Shapley-Shubik index of 1 are equal. In the second case, $n$ is odd. Since the Banzhaf and Shapley-Shubik indices of 1 are both 0, we must compare the formulas for the indices of proper divisors $d_i$ where $d_i$ is not 1. The Banzhaf index of each $d_i$ is $\frac{4}{2^{d-1}+4(d-2)-6}$ and the formula for the Shapley-Shubik index of $d_i$ is $\frac{2[(d-2)\cdot(d-2)!]+4(d-2)!}{d!} = \frac{2}{d-1}$. By simplifying $\frac{4}{2^{d-1}+4(d-2)-6} = \frac{2}{d-1}$, we have $2^{d-1}=-2d+12$. Since $f(x)=2^{x-1}$ is an increasing function for $x\geq 0$ and $g(x)=-2x+12$ is a decreasing function that intersects the x-axis when $x=6$, $2^{d-1}>-2d+12$ for all $x\geq 6$. Hence there are no possible abundant numbers $n$ for when these two expressions are equal. \subsection{Forms $pn$ and $mn$} In order to extend the class of abundant numbers for which Proposition 2.1 is true, we can examine numbers of the form $pn$ and $mn$ where $p$ and $m$ are primes. \begin{conjecture} Let $n$ be an abundant number. If $p$ and $m$ are primes greater than $\sigma{(n)}$, then the divisors of the divisor voting systems of $pn$ and $mn$ have the same Shapley-Shubik power indices as well as the same Banzhaf power indices. \end{conjecture} We will prove a part of this conjecture, which is the following proposition. \begin{proposition} Let $n$ be an abundant number and let $p$ and $m$ be primes greater than $\sigma{(n)}+1$. Then the divisor voting systems of $pn$ and $mn$ have the same number of winning coalitions. \end{proposition} \begin{proof} Let the divisor voting system of $n$ be $\{1, d_1, d_2,\dots, d_k, n\}$, where each $d_i$ is a proper divisor of $n$. Since $p$ and $m$ are primes, the divisor voting systems of $pn$ and $mn$, $S_{pn}$ and $S_{mn}$ respectively, are the following: \begin{center} $S_{pn}=\{1, d_1,\dots, d_k, n, p, pd_1, pd_2, \dots, pd_k, pn\}$ \end{center} \begin{center} $S_{mn}=\{1, d_1,\dots, d_k, n, m, md_1, md_2, \dots, md_k, mn\}$. \end{center} Let $W_{pn}$ be the set of winning coalitions in $S_{pn}$ and let $W_{mn}$ be the set of all winning coalitions in $S_{mn}$. In order to prove that $W_{pn}$ and $W_{mn}$ have the same number of coalitions, we will prove that there exists a bijection between the two sets. Let $f\colon W_{mn} \rightarrow W_{pn}$ be the function defined by \begin{equation*} f(\{d_{j_1},d_{j_2},\dots, d_{j_l}, md_{c_1},md_{c_2},\dots, md_{c_s}\}) = \{d_{j_1},d_{j_2},\dots, d_{j_l}, pd_{c_1},pd_{c_2},\dots, pd_{c_s}\}. \end{equation*} Note that $f$ is injective, since if $$\{d_{j_1},d_{j_2},\dots, d_{j_l}, pd_{c_1},pd_{c_2},\dots, pd_{c_s}\} = \{d_{j_1}',d_{j_2}',\dots, d_{j_l}', pd_{c_1}',pd_{c_2}',\dots, pd_{c_s}'\},$$ then we have $$\{d_{j_1},d_{j_2},\dots, d_{j_l}, md_{c_1},md_{c_2},\dots, md_{c_s}\} = \{d_{j_1}',d_{j_2}',\dots, d_{j_l}', md_{c_1}',md_{c_2}',\dots, md_{c_s}'\}.$$ Now we prove that $f$ is surjective. Suppose we have a winning coalition $G$ in $S_{pn}$. Then we will prove that there exists a winning coalition $H$ in $S_{mn}$ such that $f$ maps $H$ to $G$. $G$ is of the form $[d_j, d_{j+1},\dots, d_{j+k}, pd_c,pd_{c+1},\dots, pd_{c+r}]$, where each $d_i$ is a divisor of $n$. Then, based on our construction of $f$, we know that $H$ should be of the form \begin{center} $[d_j, d_{j+1},\dots, d_{j+k}, md_c,md_{c+1},\dots, md_{c+r}]$. \end{center} We will now prove that $H$ is a winning coalition. Let $q_{pn}$ be the quota of $S_{pn}$. Then \begin{equation*} q_{pn}=\dfrac{p\sigma{(n)}+\sigma{(n)}}{2}+1. \end{equation*} Similarly, for the quota of $S_{mn}$ we have \begin{equation*} q_{mn}=\dfrac{m\sigma{(n)}+\sigma{(n)}}{2}+1. \end{equation*} Let $w_1$ be the total voting power of $G$. Since $G$ is a winning coalition, we have \begin{equation} w_1 \geq q_{pn} = \dfrac{p\sigma{(n)}+\sigma{(n)}}{2}+1. \end{equation} Let $$\beta=\sum_{n=1}^{k} d_{j+n}$$ and let $$\alpha =\sum_{n=1}^{r} d_{c+n}.$$ Then we can rewrite $w_1$ as $p\beta+\alpha$. By substituting $p\beta+\alpha$ for $w_1$ we have \begin{equation} p\beta+\alpha \geq \dfrac{p\sigma{(n)}+\sigma{(n)}}{2}+1. \end{equation} Hence \begin{equation} p \geq\dfrac{\sigma({n})+2-2\alpha}{2\beta-\sigma({n})}. \end{equation} The maximum value of the right-hand side of 2.3 occurs when the denominator is 1 and $\alpha$ is 0. Hence the right-hand side of 2.3 is at most $\sigma(n)+2$. Since $m > \sigma(n) +1$, we have \begin{equation} m \geq\dfrac{\sigma({n})+2-2\alpha}{2\beta-\sigma({n})} \end{equation} By rearranging 2.4, we have \begin{equation*} m\beta+\alpha \geq \dfrac{m\sigma{(n)}+\sigma{(n)}}{2}+1=q_{mn}. \end{equation*} The total voting power of $H$ is also $m\beta + \alpha$ due to its construction. Hence $H$ is a winning coalition and $f$ is surjective. Therefore, since there exists a bijection between $W_{mn}$ and $W_{pn}$, the divisor voting systems of $mn$ and $pn$ have the same number of winning coalitions. \qedhere \end{proof} \section{Fixed Points} \subsection{Banzhaf} The general Banzhaf power indices for an arbitrary distribution of weights is quite difficult to express, so we will only consider the weights of the form $(a, b, b, \dots , b)$ where there is one player with a vote of $a$ and $m$ players have a vote of $b$, and the weights sum to $1$. We refer to people with a vote of $a$ as type $A$ and people with a vote of $b$ as Type $B$. Then we have $a = 1 - mb$. Each winning coalition either contains the player of Type $A$, or it does not. \begin{itemize} \item If a winning coalition does not contain the player of Type $A$, then there must be at least $p$ players of type $B$, where $pb > \frac{1}{2}$. Since $p$ is an integer, this simplifies to $p \geq \lceil \frac{1}{2b} \rceil$. Since all players in our coalition have the same voting power, we must have $p = \lceil \frac{1}{2b} \rceil$ for any player to be a critical player. Then each player of type $B$ is a critical player exactly $\binom{m-1}{\lceil \frac{1}{2b} \rceil - 1}$ times because we can pick $\lceil \frac{1}{2b} \rceil - 1$ of the remaining $m-1$ players of type $B$ to be the other players on our coalition. \item Now suppose a winning coalition contains the player of Type $A$. Then if are $q$ players of type $B$, $q$ must satisfy $1 - mb + qb > \frac{1}{2}$ and $qb < \frac{1}{2}$. $q$ is also an integer, so this simplifies to $m - \lfloor \frac{1}{2b} \rfloor \leq q \leq \lfloor \frac{1}{2b} \rfloor$. Then the player of Type $A$ is critical $\binom{m}{m - \lfloor \frac{1}{2b} \rfloor} + \binom{m}{m - \lfloor \frac{1}{2b} \rfloor + 1} + \dots + \binom{m}{\lfloor \frac{1}{2b} \rfloor}$ times, since we pick any $q$ players of type $B$. Furthermore, players of type $B$ can also be critical when $q$ is a minimum, so players of Type $B$ are critical an additional $\binom{m-1}{m - \lfloor \frac{1}{2b} \rfloor - 1}$ times. \end{itemize} Therefore the player of type A has Banzhaf index \[ \frac{\binom{m}{m - \lfloor \frac{1}{2b} \rfloor} + \binom{m}{m - \lfloor \frac{1}{2b} \rfloor + 1} + \dots + \binom{m}{\lfloor \frac{1}{2b} \rfloor}}{m \left(\binom{m-1}{\lceil \frac{1}{2b} \rceil - 1} + \binom{m-1}{m - \lfloor \frac{1}{2b} \rfloor - 1} \right) + \left[\binom{m}{m - \lfloor \frac{1}{2b} \rfloor} + \binom{m}{m - \lfloor \frac{1}{2b} \rfloor + 1} + \dots + \binom{m}{\lfloor \frac{1}{2b} \rfloor} \right]} \] while the player of type B has Banzhaf index \[ \frac{\binom{m-1}{\lceil \frac{1}{2b} \rceil - 1} + \binom{m-1}{m - \lfloor \frac{1}{2b} \rfloor - 1} }{m \left(\binom{m-1}{\lceil \frac{1}{2b} \rceil - 1} + \binom{m-1}{m - \lfloor \frac{1}{2b} \rfloor - 1} \right) + \left[\binom{m}{m - \lfloor \frac{1}{2b} \rfloor} + \binom{m}{m - \lfloor \frac{1}{2b} \rfloor + 1} + \dots + \binom{m}{\lfloor \frac{1}{2b} \rfloor} \right]}. \] We need the first expression to be equal to $1-mb$ and the second expression to be equal to $b$ (but note that it is sufficient to investigate one of these, since by symmetry if one of them is true then the other one becomes automatically true). It is difficult to find classes of solutions solely based on this equation, so we will investigate fixed points of both Banzhaf and Shapley-Shubik indices instead, in Section $3.3$. \subsection{Shapley-Shubik} First let us examine the primitive non-trivial fixed points of $[1/2_s: a, b, b, \ldots, b]$, where the total sum of the votes is 1. We can now say that players of type $A$ have a weight of $a$ and players of type $B$ have a weight of $b$. When trying to compute the power of type $A$, suppose that there are $p$ players of type $B$ in the permutation before the person of type $A$. If the players of type $A$ is a critical player, then we know that $pb \leq \frac{1}{2}$ and $pb + a > \frac{1}{2}$. We know that $a = 1 - mb$, so we can substitute that into the second inequality to get $pb + (1 - mb) > \frac{1}{2}$. Solving these inequalities, we get $m - \frac{1}{2b} < p \leq \frac{1}{2b}$. We know that $p$ must be an integer, so assuming $\frac{1}{2b} \not \in \mathbb{Z}$, we can make the bounds stricter, which gives us $m - \lfloor \frac{1}{2b} \rfloor \leq p \leq \lfloor \frac{1}{2b} \rfloor$. To figure out how many different permutations correspond to each value of $p$, we can see that there must be $m!$ different permutations for how to arrange the players of type $B$, and then we just need to put the person of type $A$ in the appropriate location. Therefore, $m!(\lfloor \frac{1}{2b} \rfloor - (m - \lfloor \frac{1}{2b} \rfloor) + 1) = m!(2\lfloor \frac{1}{2b} \rfloor - m + 1)$ counts the number of permutations where type $A$ is the critical player, and $\frac{m!(2\lfloor \frac{1}{2b} \rfloor - m + 1)}{(m + 1)!} = \frac{2\lfloor \frac{1}{2b} \rfloor - m + 1}{m + 1}$ gives the power of the person of type $A$ if $\frac{1}{2b} \not \in \mathbb{Z}$. On the other hand, if $\frac{1}{2b} \in \mathbb{Z}$, then $m!\left(\frac{1}{2b} - (m - \frac{1}{2b})\right) = m!(\frac{1}{b} - m)$ represents the number of permutations such that the person of type $A$ is a critical player. This means that $\frac{1-mb}{b(m + 1)}$ is the power index of the person of type $A$. But note that this equals $1-mb$ if and only if $b= \frac{1}{m+1}$ in which case all players have the same power, i.e. we get a trivial fixed point! \begin{proposition}Let $C$ be a positive integer. If $m = 2k - 1$, then for large enough $k$ $($depending only on $C)$, $b = \frac{k - C}{2k^2- k}$ and $a = 1 - (2k - 1)b$ yield a non-trivial fixed point of the form $[1/2_s: a, b, \dots, b]$. Moreover, all non-trivial fixed points are of this form. \end{proposition} \begin{proof} If $b = \frac{k - C}{2k^2 - k}$, then $\frac{1}{2b} \neq \lfloor \frac{1}{2b} \rfloor$. We calculated the power of the person of type $A$ to be $\frac{2\lfloor \frac{1}{2b} \rfloor - m + 1}{m + 1}$, so $\frac{2\lfloor \frac{1}{2b} \rfloor - (2k - 1) + 1}{2k} = 1 - 2kb$ is necessary for this specific configuration to be a fixed point. If $b = \frac{k - C}{2k^2 - k}$, then $\left\lfloor \frac{1}{2 \cdot \frac{k - C}{2k^2 - k}} \right\rfloor = k + C - 1$ for large enough $k$. Substituting, we get \begin{align*} \frac{2\lfloor \frac{1}{2b} \rfloor - 2k + 2}{2k} &= \frac{2(k + C - 1) - 2k + 2}{2k} = \frac{C}{k} \\ &= \frac{C(2k - 1)}{k(2k - 1} = \frac{(2k^2 - k) - (k - C)(2k - 1)}{2k^2 - k} \\ &= 1 - (2k - 1)\frac{k - C}{2k^2 - k} \\ &= 1 - mb \end{align*} Therefore, the calculated value for type $A$'s weight is equal to $a$, and similarly the original value for type $B$'s weight is equal to $b$. All examples of this form must are indeed non-trivial fixed points. \qedhere To show that these classify all the fixed points of $[1/2_s: a, b, b, \dots]$, we note that $\lfloor \frac{1}{2b} \rfloor$ must be an integer, so we know there exists an integer $C$ such that $\lfloor \frac{1}{2b} \rfloor = k + C - 1$. Substituting this in and solving for $b$, we get $b = \frac{k - C}{2k^2 - k}$. We also know that $C > 0$ because if $C \leq 0$, then $\frac{k - C}{2k^2 - k} \geq \frac{1}{2k - 1}$, so this value of $b$ represents either a trivial case or an impossible one where $(2k - 1)b > 1$. Since there are no non-trivial fixed points in the other case, this classifies all the fixed points if $m = 2k - 1$ for $[1/2_s: a, b, \dots, b]$. \end{proof} \begin{proposition} Let $C$ be a positive integer. If $m = 2k$, then for large enough $k$, $b = \frac{k - C}{2k^2 + k}$ and $a = 1 - 2kb$ gives a non-trivial fixed point of the form $[1/2_s: a, b, \ldots, b]$. Moreover, all non-trivial fixed points are of this form. \end{proposition} \begin{proof} Since $\frac{1}{2b} \not \in \mathbb{Z}$, we know that $\frac{2\lfloor \frac{1}{2b} \rfloor - 2k + 1}{2k + 1} = 1 - 2kb$ is necessary for this value of $b$ to be a fixed point. If we plug in $b = \frac{k - C}{2k^2 + k}$ into $\lfloor \frac{1}{2b} \rfloor$, we get that this equals $k + C$ for large enough $k$. Substituting this in, we get \begin{align*} \frac{2\lfloor \frac{1}{2b} \rfloor - 2k + 1}{2k + 1} &= \frac{2C + 1}{2k + 1} \\ &= \frac{2kC + k}{2k^2 + k} \\ &= \frac{(2k^2 + k) - 2k(k - C)}{2k^2 + k} \\ &= 1 - 2kb \end{align*} Therefore, the calculated value for type $A$'s weight is equal to $a$, and similarly the original value for type $B$'s weight is equal to $b$. All examples of this form are hence non-trivial fixed points. \qedhere To show that these classify all the fixed points of $[1/2_s: a, b, b, \dots]$, we note that $\lfloor \frac{1}{2b} \rfloor$ must be an integer, so we know there exists an integer $C$ such that $\lfloor \frac{1}{2b} \rfloor = k + C$. Substituting this in and solving for $b$, we get $b = \frac{k - C}{2k^2 + k}$. We also know that $C > 0$ because if $C \leq 0$, then $\frac{k - C}{2k^2 + k} \geq \frac{1}{2k + 1}$, so this value of $b$ represents either a trivial case or an impossible one where $2kb > 1$. Since there are no non-trivial fixed points in the other case, this classifies all the fixed points if $m = 2k$ for $[1/2_s: a, b, b, \dots]$. \end{proof} We can now take a look at weighted voting systems in the form of $(a, a, b, \ldots, b)$, which we can analyze in a similar way to the previous parts. Again letting $m$ be the number of players of type $B$, we can then split into two cases. \begin{enumerate} \item $m = 2k$ In this case, there are two possibilities for a winning coalition in which the critical player is in type $A$. Either the other player of type $A$ is in the winning coalition or they are not. We can calculate how many permutations fit into one of the categories and then add them together to get the total number of permutations where a person of type $A$ is the critical player when trying to compute Shapley-Shubik. \begin{itemize} \item Only one player of type $A$ in the winning coalition. Let $p$ be the number of players of type $B$ in the winning coalition. We know that $pb \leq \frac{1}{2}$ and $pb + a > \frac{1}{2}$. Solving for $p$ and substituting $a = \frac{1 - 2kb}{2}$, we get the inequalities $p \leq \frac{1}{2b}$ and $p > k$. Since $p$ must be an integer, we can make stricter bounds by saying $k + 1 \leq p \leq \lfloor \frac{1}{2b} \rfloor$. Now, for a given value of $p$, we need to figure out how many permutations work with that value of $p$. Out of the $2k + 2$ people voting, the first $p + 1$ spots are part of the winning coalition in a given permutation, so the other player of type $A$ must be in the remaining $2k - p + 1$ locations. There are $(2k)!$ ways to arrange the rest of the players, so in total, there are $m!(2k - p + 1)$ permutations associated with a given value of $p$. We then need to sum this expression using the bounds from earlier to get the total number of permutations in this case to be $(2k)!\sum_{p = k + 1}^{\lfloor 1/(2b) \rfloor}(2k - p + 1) = \frac{(2k)!}{2}(\lfloor \frac{1}{2b} \rfloor - k)(3k - \lfloor \frac{1}{2b} \rfloor + 1)$. \item Both players of type $A$ are in the winning coalition. Again, let $p$ be the number of players of type $B$ in the winning coalition. We know that $pb + a \leq \frac{1}{2}$ and $pb + 2a > \frac{1}{2}$. Solving for $p$ and substituting $a = \frac{1 - 2kb}{2}$, we get the inequalities $p \leq k$ and $p > 2k - \frac{1}{2b}$. If we assume that $\frac{1}{2b} \not \in \mathbb{Z}$, then since $p$ must be an integer, we can write $2k - \lfloor \frac{1}{2b} \rfloor \leq p \leq k$. For each value of $p$, there are $p + 1$ places where we can place the other player of type $A$ and $(2k + 1)!$ ways to place the rest of the players, so there are $(2k)!\sum_{p = 2k - \lfloor \frac{1}{2b} \rfloor}^k (p + 1) = \frac{1}{2}\left(\lfloor \frac{1}{2b} \rfloor - k + 1\right)\left(3k - \lfloor \frac{1}{2b} \rfloor + 2\right)$. \end{itemize} Combining these two cases, along with the fact that there are $(2k + 2)!$ total permutations, we get that the power of a player of type $A$ must be $$\frac{\frac{1}{2}\left(\lfloor \frac{1}{2b} \rfloor - k\right)\left(3k - \lfloor \frac{1}{2b} \rfloor + 1\right) + \frac{1}{2}\left(\lfloor \frac{1}{2b} \rfloor - k + 1\right)\left(3k - \lfloor \frac{1}{2b} \rfloor + 2\right)}{(2k + 1)(2k + 2)},$$ and a value of $b$ is a fixed point if and only if this expression equals $\frac{1 - 2kb}{2}$. Using this equation, we can generate the following solutions for $b$. \begin{center} \begin{tabular}{c|l} $k = 1$ & $b = 1/3$ \\ \hline $k = 2$ & $b = 2/15, b = 1/5$ \\ \hline $k = 3$ & $b = 3/28, b = 1/7$ \\ \hline $k = 4$ & $b = 13/180, b = 4/45, b = 1/9$ \\ \hline $k = 5$ & $b = 7/110, b = 5/66, b = 1/11$ \end{tabular} \end{center} We can verify that $b = \frac{1}{2k + 1}$ and $b = \frac{k}{(k + 1)(2k + 1)}$ are classes of primitive non-trivial fixed points for large enough $k$ by showing that both sides equal each other when we substitute the values of $b$ in. \item $m = 2k + 1$ We again split into different cases depending on the number of players of type $A$ in the winning coalition. \begin{itemize} \item Only one player of type $A$ in the winning coalition. Let $p$ be the number of players of type $B$ in the winning coalition. We know that $pb \leq \frac{1}{2}$ and $pb + a > \frac{1}{2}$. Solving for $p$ and substituting $a = \frac{1 - (2k +1)b}{2}$, we get the inequalities $p \leq \frac{1}{2b}$ and $p > k$. Since $p$ must be an integer, we can make stricter bounds by saying $k + 1 \leq p \leq \lfloor \frac{1}{2b} \rfloor$. Now, for a given value of $p$, we need to figure out how many permutations work with that value of $p$. Out of the $2k + 3$ people voting, the first $p + 1$ spots are part of the winning coalition in a given permutation, so the other player of type $A$ must be in the remaining $2k - p + 2$ locations. There are $(2k + 1)!$ ways to arrange the rest of the players, so in total, there are $(2k + 1)!(2k - p + 2)$ permutations associated with a given value of $p$. We then need to sum this expression using the bounds from earlier to get the total number of permutations in this case to be $(2k + 1)!\sum_{p = k + 1}^{\lfloor 1/(2b) \rfloor}(2k - p + 2) = \frac{(2k + 1)!}{2}(\lfloor \frac{1}{2b} \rfloor - k)(3k - \lfloor \frac{1}{2b} \rfloor + 3)$. \item Both players of type $A$ are in the winning coalition. Again, let $p$ be the number of players of type $B$ in the winning coalition. We know that $pb + a \leq \frac{1}{2}$ and $pb + 2a > \frac{1}{2}$. Solving for $p$ and substituting $a = \frac{1 - (2k + 1)b}{2}$, we get the inequalities $p \leq k$ and $p > 2k + 1 - \frac{1}{2b}$. If we assume that $\frac{1}{2b} \not \in \mathbb{Z}$, then since $p$ must be an integer, we can write $2k + 1 - \lfloor \frac{1}{2b} \rfloor \leq p \leq k$. For each value of $p$, there are $p + 1$ places where we can place the other player of type $A$ and $(2k + 1)!$ ways to place the rest of the players, so there are $(2k + 1)!\sum_{p = 2k + 1 - \lfloor \frac{1}{2b} \rfloor}^k (p + 1) = \frac{1}{2}\left(\lfloor \frac{1}{2b} \rfloor - k\right)\left(3k - \lfloor \frac{1}{2b} \rfloor + 3\right)$. \end{itemize} Combining these two cases, along with the fact that there are $(2k + 3)!$ total permutations, we get that the power of a player of type $A$ must be $$\frac{\left(\lfloor \frac{1}{2b} \rfloor - k\right)\left(3k - \lfloor \frac{1}{2b} \rfloor + 3\right)}{(2k + 1)(2k + 2)},$$ and a value of $b$ is a fixed point if and only if this expression equals $\frac{1 - (2k + 1)b}{2}$. Using this equation, we can generate the following solutions for $b$. \begin{center} \begin{tabular}{c|l} $k = 1$ & $b = 2/15$ \\ \hline $k = 2$ & $b = 3/35, b = 11/105$ \\ \hline $k = 3$ & $b = 4/63, b = 11/126$ \\ \hline $k = 4$ & $b = 5/99, b = 31/495, b = 37/495$ \end{tabular} \end{center} \noindent And we can verify that $b = \frac{k + 1}{4(k + 1)^2 - 1}$ is a primitive fixed points for large enough $k$. \end{enumerate} \subsection{Both Banzhaf and Shapley-Shubik} To find primitive nontrivial fixed points for both power indices, we can test Shapley-Shubik fixed points to see if they are still fixed points for Banzhaf. Consider the Shapley-Shubik fixed points in $\textbf{Proposition 3.1}$. To calculate Banzhaf power indices for this class of fixed points, we see that each winning coalition either contains the player of type $A$ or it does not. \begin{itemize} \item Winning coalition does not contain player of type A: The winning coalition must contain at least $p$ players of type $B$, where $p(\frac{k-c}{2k^2-k}) > \frac{1}{2}$, or $p > \frac{2k^2-k}{2k-2c} = k + c - \frac{1}{2} + \frac{2c^2-c}{2k-2c}$. Since $p$ is an integer, we have $p \geq k + c$ for large enough $k$. In order for the coalition to have a critical player, we must have $p = k + c$, so each player of type $B$ is a critical player $\binom{2k-2}{k+c-1}$ times since we pick the other $p - 1$ players from the remaining $2k-2$ players of Type $B$. \item Winning coalition contains player of type A: The winning coalition must contain $q$ players of type $B$, where $q(\frac{k-c}{2k^2-k}) < \frac{1}{2}$ and $\frac{1}{k} + q(\frac{k-c}{2k^2-k}) > \frac{1}{2}$, which simplifies to $\frac{(k-2c)(2k-1)}{2k-2c} < q < \frac{2k^2-k}{2k-2c}$. The lower and upper bounds are equal to $k - c - \frac{1}{2} + \frac{c-2c^2}{2k-2c}$ and $k + c - \frac{1}{2} + \frac{2c^2-c}{2k-2c}$, respectively, so since $q$ is an integer, we have $k - c \leq q \leq k + c - 1$ for large enough $k$. Then the player of type $A$ is a critical player $\binom{2k-1}{k-c} + \binom{2k-1}{k-c+1} + \dots + \binom{2k-1}{k+c-1}$ times, where we choose $q$ players of type $B$ from the total $2k-1$. Furthermore, players of type $B$ are also critical players when $q = k-c$, so they are critical players $\binom{2k-2}{k-c-1}$ times. \end{itemize} Thus the Banzhaf index for a player of type $B$ is given by \[\dfrac{\binom{2k-2}{k+c-1} + \binom{2k-2}{k-c-1}}{(2k-1)\left( \binom{2k-2}{k+c-1} + \binom{2k-2}{k-c-1} \right) + \binom{2k-1}{k-c} + \binom{2k-1}{k-c+1} + \dots + \binom{2k-1}{k+c-1}}. \] We want this expression to equal $\frac{k-c}{2k^2-k}$, which we found to only hold true when $c=1$. \newline \\ Now consider the Shapley-Shubik fixed points in $\textbf{Proposition 3.2}$. To calculate Banzhaf power indices for this class of fixed points, we see that each winning coalition either contains the player of type $A$ or it does not. \begin{itemize} \item Winning coalition does not contain player of type A: The winning coalition must contain at least $p$ players of type $B$, where $p(\frac{k-c}{2k^2+k}) > \frac{1}{2}$, or $p > \frac{2k^2+k}{2k-2c} = k + c + \frac{1}{2} + \frac{2c^2+c}{2k-2c}$. Since $p$ is an integer, we have $p \geq k + c +1$ for large enough $k$. In order for the coalition to have a critical player, we must have $p = k + c+1$, so each player of type $B$ is a critical player $\binom{2k-1}{k+c}$ times since we pick the other $p - 1$ players from the remaining $2k-1$ players. \item Winning coalition contains player of type A: The winning coalition must contain $q$ players of type $B$, where $q(\frac{k-c}{2k^2+k}) < \frac{1}{2}$ and $\frac{1}{k} + q(\frac{k-c}{2k^2+k}) > \frac{1}{2}$, which simplifies to $\frac{(k-2c)(2k+1)}{2k-2c} < q < \frac{2k^2+k}{2k-2c}$. The lower and upper bounds are equal to $k - c + \frac{1}{2} - \frac{c+2c^2}{2k-2c}$ and $k + c + \frac{1}{2} + \frac{2c^2+c}{2k-2c}$, respectively, so since $q$ is an integer, we have $k - c +1\leq q \leq k + c $ for large enough $k$. Then the player of type $A$ is a critical player $\binom{2k}{k-c+1} + \binom{2k}{k-c+2} + \dots + \binom{2k}{k+c}$, where we choose $q$ players of type $B$ from the total $2k-1$. Furthermore, players of type $B$ are also critical players when $q = k-c+1$, so they are critical players $\binom{2k-1}{k-c}$ times. \end{itemize} Thus the Banzhaf index for a player of type $B$ is given by \[ \dfrac{\binom{2k-1}{k+c} + \binom{2k-1}{k-c}}{(2k)\left( \binom{2k-1}{k+c} + \binom{2k-1}{k-c} \right) + \binom{2k}{k-c+1} + \binom{2k}{k-c+2} + \dots + \binom{2k}{k+c}}. \] We want this expression to equal $\frac{k-c}{2k^2+k}$, but we found no integer solutions for $c$ for which this is true. \section{Future Research} To expand the class of abundant numbers for which Proposition 2.1 is true, it would be helpful to prove that if the Shapley-Shubik and Banzhaf power indices differ in the divisor voting system of an abundant number $n$, then they also differ in the divisor voting system of $2^{k}p$ for a prime $p$ and various values of $k$. It would also be useful to finish our proof that the players in the divisor voting system of $pn$ have equal Banzhaf indices as well as equal Shapley-Shubik indices as those in $mn$. By proving these statements, it would then be possible to prove Question $1.2.$ by inducting on the number of distinct prime divisors of an abundant number $n$. In further research concerning fixed points, we could try strengthening the algebraic approach for fixed points to find more classes (e.g. with three types of players). \section{Acknowledgements} We would like to thank Miroslav Marinov for his guidance, mentorship, and feedback on this research paper. We thank Joshua Zelinsky for proposing this research project and for his helpful suggestions. We would like to thank the PROMYS program, especially Professor Fried and Steve Huang for the opportunity to work on this project, and the Clay Mathematical Institute for their support of the returning student program. \section{Appendix} Here is the code that we used to compute the power indices. \lstinputlisting[language=Python]{code.py}

https://arxiv.org/abs/0805.2140

On the adjoint quotient of Chevalley groups over arbitrary base schemes

For a split semisimple Chevalley group scheme G with Lie algebra g over an arbitrary base scheme S, we consider the quotient of g by the adjoint action of G. We study in detail the structure of g over S. Given a maximal torus T with Lie algebra t and associated Weyl group W, we show that the Chevalley morphism t/W -> g/G is an isomorphism except for the group Sp_{2n} over a base with 2-torsion. In this case this morphism is only dominant and we compute it explicitly. We compute the adjoint quotient in some other classical cases, yielding examples where the formation of the quotient g -> g/G commutes, or does not commute, with base change on S.

\section{Introduction} \indent Let $G$ be a split semisimple Chevalley group scheme over a base scheme $S$ and let~$\fg$ be its Lie algebra. The quotient of $\fg$ by the adjoint action of $G$ in the category of schemes affine over~$S$, that is to say, the spectrum of the sheaf of $G$-invariant functions of $\fg$, is traditionally called the {\em adjoint quotient} of $\fg$ and denoted $\fg/G$. Let $T\subset G$ be a maximal torus and $\ft$ its Lie algebra. There is an induced action of the Weyl group $W=W_T$ on~$\ft$ and the inclusion $\ft\subset\fg$ induces a natural morphism $\pi:\ft/W\to \fg/G$. In this paper, we call it the {\em Chevalley morphism}. The situation where the base is the spectrum of an algebraically closed field whose characteristic does not divide the order of the Weyl group is well documented. In this case $\pi$ is an isomorphism, as proven by Springer and Steinberg~\cite{SS}. It is known also that the adjoint quotient is an affine space (see Chevalley \cite{Ch}, Veldkamp \cite{Ve}, Demazure \cite{De}). There are counter-examples to these statements when the characteristic divides the order of the Weyl group. Another difficulty comes from the fact that we are considering the quotient $\fg/\Ad(G)$ of the Lie algebra, and not $G/\Int(G)$, and at some point this derivation causes some trouble (Steinberg \cite[p.51]{St} was also lead to the same conclusion). In this paper, we turn our attention to the integral structure of the adjoint quotient and the Chevalley morphism, including the characteristics that divide the order of $W$. In other words we are interested in an arbitrary base scheme~$S$, and in the behaviour of the previous objects under base change $S'\to S$. It is not hard to extend the results from simple to semisimple groups, so for simplicity we restrict to simple Chevalley groups. Our main result (theorems~\ref{theo_chevalley_dominant} and \ref{theo_chevalley_iso}) is that in most cases the Chevalley morphism is an isomorphism, therefore reducing the calculation of $\fg/G$ to the calculation of a quotient by a finite group: \bigskip \noindent {\bf Theorem 1} {\em Let $G$ be a split simple Chevalley group scheme over a base scheme $S$. Then the Chevalley morphism $\pi:\ft/W\to\fg/G$ is schematically dominant, and is an isomorphism if $G$ is not isomorphic to $Sp_{2n}$, $n\ge 1$.} \bigskip Note that even when the base is a field, this improves the known results. Our proof follows a classical strategy. The main new imputs are: over a base field, a close analysis of the root systems and determination of the conditions of nonvanishing of the differentials of the roots (lemma~\ref{racine_non_primitive}), and over a general base, a careful control of the poles along the singular locus for the relative meromorphic functions involved in the proof. We treat separately the exceptional case (theorem~\ref{theo_sp2n}): \bigskip \noindent {\bf Theorem 2} {\em If $G=Sp_{2n}$ then the Chevalley morphism is an isomorphism \iff the base has no $2$-torsion. Moreover, over an open affine subscheme $\Spec(A)\subset S$, the ring of functions of $\fg/G$ is $$ A[c_2,c_4,\dots,c_{2n}] $$ where the functions $c_{2i}$ are the coefficients of the characteristic polynomial $c_{2i}$. The formation of the adjoint quotient commutes with arbitrary base change.} \bigskip We see that for $G=Sp_{2n}$, the formation of the adjoint quotient commutes with base change. If this was true for all split simple Chevalley groups, then we could deduce the main theorem~1 above from the case $S=\Spec(\bZ)$ which is significantly easier (see corollary~\ref{coro_factorial_ring}). Unfortunately it is not always so, and in order to see this, we study in detail the orthogonal groups in types $B$ and $D$. Our main result is (see after the theorem and subsection \ref{invariant_weyl} for the missing notations): \bigskip \noindent {\bf Theorem 3} {\em If $G=SO_{2n}$ or $G=SO_{2n+1}$ then over an open affine subscheme $\Spec(A)\subset S$, the ring of functions of $\fg/G$ is the following: \medskip {\rm (i)} if $G=SO_{2n}$: $A[c_2,c_4,\dots,c_{2n-2},\pf\,;\,x(\pi_1)^{\epsilon_1} \dots(\pi_{n-1})^{\epsilon_{n-1}}]$, where $x$ runs through a set of generators of the $2$-torsion ideal $A[2]\subset A$, and $\epsilon_i=0$ or $1$, not all $0$, \medskip {\rm (ii)} if $G=SO_{2n+1}$: $A[c_2,c_4,\dots,c_{2n},\,;\,x(\pi_1)^{\epsilon_1} \dots(\pi_{n})^{\epsilon_{n}}]$, where $x$ runs through a set of generators of $A[2]$ and $\epsilon_i=0$ or $1$, not all $0$.} \bigskip The functions that appear in the preceding theorem are the coefficients of the characteristic polynomial $c_{2i}$, the Pfaffian $\pf$ and some functions $\pi_i$ which we call the coefficients of the Pfaffian polynomial. The functions $c_{2i}$ and $\pf$ are invariant, but the functions $\pi_i$ are invariant only after multiplication by a $2$-torsion element. The definition of these objets needs some care, since it is not always the straightforward definition one would think of. Using the theorem above, we prove that the formation of the adjoint quotient for the orthogonal groups commutes with a base change $f:S'\to S$ if and only if $f^*S[2]=S'[2]$, where $S[2]$ is the closed subscheme defined by the ideal of $2$-torsion. This holds in particular if $2$ is invertible in ${\cal O}_S$, or if $2=0$ in ${\cal O}_S$, or if $S'\to S$ is flat. We prove also that if $S$ is noetherian and connected then the quotient is of finite type over~$S$, and is flat over $S$ if and only if $S[2]=S$ or $S[2]=\emptyset$. \bigskip We feel it useful to say that when we first decided to study the adjoint quotient over a base other than a field, we started with some examples among the classical Chevalley groups and considered their Lie algebras. To our surprise, already in the classical case we could not find concrete descriptions of them in the existing literature (for example the Lie algebra of $PSL_n$ over $\mathbb Z$). This lead to our study of the classical Lie algebras over arbitrary bases (subsection \ref{subsection_classical}). We also faced the problem of relating the Lie algebra of a group scheme and of any finite quotient of it (subsection \ref{subsection_quotient}); note that this subsection holds for any smooth group scheme, not necessarily affine over the base. Let us finally mention that spin groups over $\bZ$ have also been studied very recently in such a concrete way by Ikai \cite{Ik1}, \cite{Ik2}. \bigskip Here is the outline of the article. In the end of this section 1 we give our notations and prove a combinatorial lemma about root systems which is crucial in all the paper. In section \ref{section_Lie_algebras_of_Chev_gps} we give two dual exact sequences $$ 0\to \mathscr{L}ie(K)^\vee\to \mathscr{L}ie(G)^\vee\to \omega^1_{H/S}\to 0 $$ and $$ 0\to \mathscr{L}ie(G)\to \mathscr{L}ie(K)\to (\omega^1_{H/S})^\dag\to 0 $$ describing the relation between the Lie algebra of a smooth group scheme $G$ and the Lie algebra of a quotient $K:=G/H$ (see more precise assumptions in propositions \ref{exact_sequence_for_Lie} and \ref{prop_quotient_groupe_fini}). Then we specialize to Chevalley groups and their Lie algebras over $\bZ$. We describe their weight decomposition (subsection \ref{subsection_algebre_lie}), the intermediate quotients of $G\to G^{ad}$ and $\Lie(G)\to \Lie(G^{ad})$ (subsection \ref{diff_quotient_maps}) and we illustrate our results by describing the classical Chevalley Lie algebras (\ref{subsection_classical}). In section \ref{section_morphisme_de_Chev} we prove theorem~1 above. In the remaining sections \ref{case_S0_2n}, \ref{case_S0_2n+1} and \ref{case_Sp_2n} we treat the examples of theorems~2 and~3 above by computing explicitly the map $\ft/W\to \fg/G$ (see theorem~\ref{theo_so2n}, corollary~\ref{coro_so2n}, theorem~\ref{theo_so2n+1}, theorem~\ref{theo_sp2n}). \tableofcontents \subsection{General notations} \label{notations} \indent All rings are commutative with unit. If $A$ is a ring, we denote by $A[2]$ its $2$-torsion ideal, defined by $A[2]=\{a\in A,\,2a=0\}$. If $S$ is a scheme, we denote by $S[2]$ its closed subscheme defined by the $2$-torsion ideal sheaf. If $X$ is an affine scheme over $\Spec(A)$ we always denote by $A[X]$ its function ring. If $S$ is a scheme, $X$ is a scheme over $S$, and $T\to S$ is a base change morphism, we denote by $X\times_S T$ or simply $X_T$ the $T$-scheme obtained by base change. In all the article, we call relative Cartier divisor of $X$ over $S$ an effective Cartier divisor in $X$ which is flat over $S$. Finally, the linear dual of an ${\cal O}_S$-module ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}$ is denoted ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}^\vee:=\Hom_{{\cal O}_S}({\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I},{\cal O}_S)$. \subsection{Notations on group schemes} \indent Let $S$ be a scheme and let $G$ be a group scheme over $S$. We will use the following standard notation: $e_G:S\to G$ is the unit section of $G/S$, $\Omega^1_{G/S}$ is the sheaf of relative differential $1$-forms of $G/S$, and $\omega^1_{G/S}=e_G^*\Omega^1_{G/S}$. Recall that $\Omega^1_{G/S}=f^*\omega^1_{G/S}$ where $f:G\to S$ is the structure map, so that $\Omega^1_{G/S}$ is locally free over $G$ if and only if $\omega^1_{G/S}$ is locally free over $S$. We will write $\Lie(G/S)$ (or simply $\Lie(G)$) for the Lie algebra of $G/S$, and $\mathscr{L}ie(G/S)$ (or simply $\mathscr{L}ie(G)$) for the sheaf of sections of $\Lie(G/S)$. Note that $\Lie(G/S)$ is the vector bundle $\bV(\omega^1_{G/S})$, with the Grothendieck notation. Sometimes we shall also use gothic style letters for Lie algebras, like $\fg$, $\ft$, $\mathfrak{psl}} \def\fspin{\mathfrak{spin}$, $\mathfrak{so}} \def\fpso{\mathfrak{pso}$, etc. By Chevalley group scheme over a scheme $S$, we mean a deployable reductive group scheme over $S$, with the terminology of \cite[expos{\'e} XXII, d{\'e}finition 1.13]{sga}. By \cite[expos{\'e} XXIII, corollaire 5.3]{sga}, such a group is characterized up to isomorphism by its type (as defined in \cite[expos{\'e} XXII, d{\'e}finition 2.7]{sga}: this is essentially the root datum together with a module included in the weight lattice and containing the root lattice) and is equal to $G_S$, where $G$ is a Chevalley group scheme over the ring of integers. \subsection{Roots that are integer multiples of weights} The next lemma comes up at various places in the article. It has as a consequence the fact that the differential of a root can vanish along the Lie algebra of a maximal torus in a simple Chevalley group only in case this group is $Sp_{2n}$ (including $Sp_2 \simeq SL_2$) - see lemma~\ref{lemm_racine}. This will be crucial throughout the article: lemma~\ref{lemm_birational}, on which relies the proof of theorem~\ref{theo_chevalley_iso}, is again a consequence of this lemma, as well as the fact that a simply-connected Lie algebra is equal to its own derived algebra in all cases but $\mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n}$, see proposition~\ref{prop_commutateur}. \begin{lemm} \label{racine_non_primitive} Let $R$ be a simple reduced root system, $Q(R)$ the root lattice and $P(R)$ the weight lattice. Assume there exists $\alpha \in R,\lambda \in P(R),l \in \mathbb{N}$ such that $\alpha = l.\lambda$ and $l \geq 2$. Then $l=2$, and either $R$ is of type $A_1$, or $R$ is of type $C_n$ and $\alpha$ is a long root. \end{lemm} \begin{proo} Let us assume that $R$ is the root system defined in \cite[Planches I to IX]{bourbaki}. If $R$ is of type $A_1$, then the roots are $\alpha = \epsilon_1-\epsilon_2$ and $-\alpha$. Since $\epsilon_1 - (\epsilon_1+\epsilon_2)/2 = (\epsilon_1-\epsilon_2)/2$ is a weight, $\alpha$ is indeed twice a weight. Now let's assume that $R$ is of rank greater than 1. The hypothesis of the lemma implies that \begin{equation} \label{scalaire_l} \forall \beta \in R, \scal{\beta^\vee , \alpha} = l\scal{\beta^\vee , \lambda} \in l \mathbb Z. \end{equation} Let $\beta$ be a root. If $\alpha$ and $\beta$ have the same lentgh, by \cite[VI, no 1.3, Proposition 8]{bourbaki}, $\scal{\beta^\vee,\alpha} \in \{-1,0,1\}$. If moreover we know that $\scal{\beta^\vee,\alpha} \not = 0$, we see that (\ref{scalaire_l}) cannot hold. By \cite[VI, no 1, proposition 15, p. 154]{bourbaki}, we can assume that $\alpha$ is a simple root. This implies that in the Dynkin diagram of $R$, all edges containing the vertex corresponding to $\alpha$ must be multiple edges. This excludes all simply-laced root systems, as well as the root system of type $F_4$. Moreover, if $R$ is of type $B_n$ with $n \geq 3$, then $\alpha$ has to equal $\alpha_n$, but since $\scal{\alpha_{n-1}^\vee,\alpha_n} = -1$, we have a contradiction. If $R$ is of type $C_n$, then $\alpha$ has to equal $\alpha_n$ again. Since $\alpha_n = 2\epsilon_n$, we have indeed $\alpha \in l.P(R)$ with $l=2$. Since $B_2=C_2$, the last case to be settled is that of $G_2$. But in this case $Q(R) = P(R)$; since $R$ is reduced, it is not possible that a root be a multiple of a weight. \end{proo} \section{On the Lie algebra of Chevalley groups} \label{section_Lie_algebras_of_Chev_gps} \subsection{Lie algebras of quotients and coverings} \label{subsection_quotient} In this section, our aim is to relate the Lie algebra of a group $G$ and the Lie algebra of a quotient $G/H$. More precisely, we consider a scheme $S$, a flat $S$-group scheme $G$, and a subgroup scheme $H\subset G$ which is flat and of finite presentation over $S$. It follows from a theorem of Mickael Artin \cite[corollary 6.3]{artin} that the quotient fppf sheaf $K:=G/H$ is representable by an algebraic space over $S$. In the cases of interest to us, it will always be representable by a scheme. We let $\pi\colon G\to K$ denote the quotient morphism, and $e_K:=\pi\circ e_G$. Whether or not $H$ is normal, we write $\Lie(K)$ for the restriction of the tangent space along $e_K$ and $\mathscr{L}ie(K)$ for its sheaf of sections. \begin{prop} \label{exact_sequence_for_Lie} Assume that $G,H,K$ are as above. \begin{trivlist} \itemn{1} There is a canonical exact sequence of quasi-coherent ${\cal O}_S$-modules: $$ \omega^1_{K/S}\to \omega^1_{G/S}\to \omega^1_{H/S}\to 0\;, $$ where $\omega^1_{G/S}\to \omega^1_{H/S}$ is the natural map deduced from the inclusion $H \subset G$. \itemn{2} Assume furthermore that $G$ is smooth over $S$ and that there is schematically dominant morphism $i:U\to S$ such that $H\times_S U$ is smooth over $U$. Then, there is a canonical exact sequence of coherent ${\cal O}_S$-modules: $$ 0\to \mathscr{L}ie(K)^\vee\to \mathscr{L}ie(G)^\vee\to \omega^1_{H/S}\to 0 $$ and $\mathscr{L}ie(G)^\vee\to \omega^1_{H/S}$ is the natural map deduced from the inclusion $H \subset G$. \end{trivlist} \end{prop} Typically, in the applications, $U$ will be an open subscheme of $S$ or the spectrum of the local ring of a generic point. \begin{proo} (1) We have the fundamental exact sequence for differential $1$-forms: $$ \pi^*\Omega^1_{K/S}\to \Omega^1_{G/S}\to \Omega^1_{G/K}\to 0\;. $$ By right-exactness of the tensor product, the sequence remains exact after we pullback via $e_G$. The only thing left to prove is that there is a canonical isomorphism $e_G^*\Omega^1_{G/K}\simeq \omega^1_{H/S}$. In order to do so, we use the fact that $G\to K$ is an $H$-torsor, so that we have an isomorphism $t\colon H\times_S G \to G\times_K G$ given by $t(h,g)=(hg,g)$. We consider the fiber square: $$ \xymatrix{H\times_S G \ar[r]^t & G\times_K G \ar[r]^{\quad\pr_2} \ar[d]_{\pr_1} & G \ar[d]^\pi \\ & G \ar[r]^\pi & K} $$ Then, if we call $f\colon H\times_S G\to H$ the projection, we have the sequence of isomorphisms on $H\times_S G$: $$ t^*\pr_1^*\Omega^1_{G/K}\;\simeq\; t^*\Omega^1_{G\times_K G/G}\;\simeq\; \Omega^1_{H\times_S G/G}\;\simeq\; f^*\Omega^1_{H/S} $$ (the first and the third isomorphisms come from the invariance of the module of relative diffe\-rentials by base change, \cite{EGA},~IV.16.4.5). Pulling back along $e_H\times e_G$, we get the desired result. Moreover, following the identifications, we see that the map $\omega^1_{G/S}\to \omega^1_{H/S}$ is the same as the map induced by the inclusion $H \subset G$. \medskip \noindent (2) Since $G$ is smooth over $S$ and $G\to K$ is faithfully flat, then $K$ is also smooth over $S$. Hence $\cM=\omega^1_{K/S}$ and $\cN=\omega^1_{G/S}$ are locally free ${\cal O}_S$-modules of finite rank, so that $$ \cM\simeq \mathscr{L}ie(K)^\vee \quad \mbox{and} \quad \cN\simeq \mathscr{L}ie(G)^\vee \;. $$ It remains to check that $\cM\to\cN$ is injective. This will follow from the diagram $$ \xymatrix{i_*i^*\cM\ \ar@{^(->}[r] & i_*i^*\cN \\ \overset{}{\cM} \ar@{^(->}[u] \ar[r] & \overset{}{\cN} \ar[u]\\} $$ if we describe the injective morphisms therein. Since $i:U\to S$ is schematically dominant and $\cM$ is flat, we have an injective morphism $\cM\to \cM\otimes i_*{\cal O}_U$ and the target module is isomorphic to $i_*i^*\cM$ by the projection formula. Besides, the morphism $G\times_S U\to {G/H}\times_S U$ is smooth since $H\times_S U$ is smooth over $U$, so by the short exact sequence of $\Omega^1$'s for a smooth morphism, the morphism $i^*\cM\to i^*\cN$ is injective. By left exactness the morphism $i_*i^*\cM\to i_*i^*\cN$ is injective also. \end{proo} If $H$ is finite over $S$, like in the cases we have in mind, we can dualize the exact sequence of proposition~\ref{exact_sequence_for_Lie} thanks to a Pontryagin duality for certain torsion modules, which we now present. Let~$A$ be a commutative ring and let $Q$ be the total quotient ring of $A$, i.e. the localization with respect to the multiplicative set of nonzerodivisors (in fact we should better consider the module of global sections of the sheaf of total quotient rings on $\Spec(A)$, but in this informal discussion it does not matter). We wish to associate to any finitely presented torsion $A$-module $M$ a dual $M^\dag=\Hom_A(M,Q/A)$. For general $M$ this does not lead to nice properties such as biduality; for example if $A=k[x,y]$ is a polynomial ring in two variables and $M=A/(x,y)$ it is easy to see that $M^\dag=0$. In this example there is a presentation $A^2\to A\to M\to 0$ but one can see that there is no presentation $A^n\to A^m\to M\to 0$ with $n=m$. In fact, this is a consequence of our results below. Note that the fact that $M$ is torsion implies $n\ge m$, thus if we can find a presentation with $n=m$ it is natural to say that $M$ has {\em few relations}. By the structure theorem for modules over a principal ideal domain, all finite abelian groups have few relations, and from our point of view, this is the crucial property of finite abelian groups that makes Pontryagin duality work. These considerations explain the following definition. \begin{defi} \label{defi_TMFR} Let ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}$ be a coherent ${\cal O}_S$-module; denote by ${\cal K}} \def\cL{{\cal L}} \def\cM{{\cal M}} \def\cN{{\cal N}$ the sheaf of total quotient rings of ${\cal O}_S$. We say that ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}$ is a {\em torsion module with few relations} if ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}\otimes{\cal K}} \def\cL{{\cal L}} \def\cM{{\cal M}} \def\cN{{\cal N}=0$ and ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}$ is locally the cokernel of a morphism $({\cal O}_S)^n\to ({\cal O}_S)^n$ for some $n\ge 1$. \end{defi} We have the following easy characterization: \begin{prop} \label{characterization_of_TMFRs} Let $\varphi:\cE_1\to\cE_2$ be a morphism between locally free ${\cal O}_S$-modules of the same finite rank and let ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}=\coker(\varphi)$. Then ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}$ is a torsion module with few relations if and only if the sequence $0\to \cE_1\to\cE_2\to {\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}\to 0$ is exact, i.e. $\varphi$ is injective. \end{prop} \begin{proo} If $\varphi$ is injective, then locally over an open set where $\cE_1$ and $\cE_2$ are free, its determinant $\det(\varphi)\in{\cal O}_S$ is a nonzerodivisor. Therefore $\varphi\otimes\Id:\cE_1\otimes{\cal K}} \def\cL{{\cal L}} \def\cM{{\cal M}} \def\cN{{\cal N}\to\cE_2\otimes{\cal K}} \def\cL{{\cal L}} \def\cM{{\cal M}} \def\cN{{\cal N}$ is surjective, hence an isomorphism. It follows that ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}\otimes{\cal K}} \def\cL{{\cal L}} \def\cM{{\cal M}} \def\cN{{\cal N}=\coker(\varphi\otimes\Id)=0$. Conversely if ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}$ is a torsion module with few relations, then $\coker(\varphi\otimes\Id)={\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}\otimes{\cal K}} \def\cL{{\cal L}} \def\cM{{\cal M}} \def\cN{{\cal N}=0$ so that $\varphi\otimes\Id$ is an isomorphism. Since $\cE_1$ and $\cE_2$ are flat we have injections $$ \xymatrix{\cE_1\otimes{\cal K}} \def\cL{{\cal L}} \def\cM{{\cal M}} \def\cN{{\cal N}\ \ar@{^(->}[r] & \cE_2\otimes{\cal K}} \def\cL{{\cal L}} \def\cM{{\cal M}} \def\cN{{\cal N} \\ \overset{}{\cE_1} \ar@{^(->}[u] \ar[r] & \overset{}{\cE_2} \ar@{^(->}[u] \\} $$ and it follows that $\varphi$ is injective. \end{proo} \begin{defi} \label{PD} Given a coherent ${\cal O}_S$-module ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}$ we define its {\em Pontryagin dual} by $${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}^\dag=\cH om_{{\cal O}_S}({\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I},{\cal K}} \def\cL{{\cal L}} \def\cM{{\cal M}} \def\cN{{\cal N}/{\cal O}_S)\;.$$ \end{defi} As the following proposition proves, there is a satisfactory duality if we restrict to torsion modules with few relations. \begin{prop} \label{Pontryagin_duality} Let ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}$ be a torsion ${\cal O}_S$-module with few relations. Then: \begin{trivlist} \itemn{1} ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}^\dag$ is also a torsion ${\cal O}_S$-module with few relations, and the canonical morphism ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}\to {\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}^{\dag\dag}$ is an isomorphism. \itemn{2} For each exact sequence $0\to \cE_1\to\cE_2\to {\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}\to 0$ where $\cE_1$, $\cE_2$ are locally free ${\cal O}_S$-modules of the same finite rank, we have a canonical dual exact sequence $0\to \cE_2^\vee\to\cE_1^\vee\to {\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}^\dag\to 0$. \end{trivlist} \end{prop} \begin{proo} The assertions in point (1) are local over $S$ and therefore are easy consequences of point~(2). In order to prove point (2) we set $\cG=\coker(\cE_2^\vee\to\cE_1^\vee)$ and we construct a canonical nondegenerate pairing ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}\times\cG\to {\cal K}} \def\cL{{\cal L}} \def\cM{{\cal M}} \def\cN{{\cal N}/{\cal O}_S$, as follows. Since ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}$ is torsion and finitely generated, locally (over an open subset $U\subset S$) there is a nonzerodivisor $a\in{\cal O}_S$ such that $a\cE_2\subset\cE_1$. Given two sections $f\in\cE_2$ and $g\in\cE_1^\vee$ over $U$, we let $\langle f,g\rangle$ denote the class of $\frac{1}{a}g(af)\in {\cal K}} \def\cL{{\cal L}} \def\cM{{\cal M}} \def\cN{{\cal N}$ modulo ${\cal O}_S$. It is easy to check that this is independent of the choice of~$a$. If $f\in\cE_1$ or if $g\in\cE_2^\vee$ then $\langle f,g\rangle=0$ so there results a pairing ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}\times \cG\to {\cal K}} \def\cL{{\cal L}} \def\cM{{\cal M}} \def\cN{{\cal N}/{\cal O}_S$ and we will now check that it induces isomorphisms ${\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}\to\cG^\dag$ and $\cG\to{\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}^\dag$. By symmetry we will consider only $\sigma:\cG\to{\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}^\dag$. If $\langle \cdot,g\rangle$ is zero then we claim that $g\in \cE_1^\vee$ extends to a form on $\cE_2$. Indeed, for each $f\in\cE_2$ we have $\frac{1}{a}g(af)\in {\cal O}_S$ so that the definition $g(f):=\frac{1}{a}g(af)$ is unambiguous, since $a$ is a nonzerodivisor. It follows that $\sigma$ is injective. In order to check surjectivity we may assume that $S$ is the spectrum of a local ring, and in this case $\cE_1$, $\cE_2$ are trivial. Any $u:{\cal F}} \def\cG{{\cal G}} \def\cH{{\cal H}} \def\cI{{\cal I}\to {\cal K}} \def\cL{{\cal L}} \def\cM{{\cal M}} \def\cN{{\cal N}/{\cal O}_S$ factors through $\frac 1a{\cal O}_S/{\cal O}_S\subset {\cal K}} \def\cL{{\cal L}} \def\cM{{\cal M}} \def\cN{{\cal N}/{\cal O}_S$ and then induces a morphism $\cE_2\to {\cal O}_S/a{\cal O}_S$. Since $\cE_2$ is trivial this map lifts to $u':\cE_2\to{\cal O}_S$. Moreover if $x \in \cE_1$ then $u'(x) \in a {\cal O}_S$, so we can set $v(x) = \frac 1a u'(x)$; then it is easy to check that $v$ is a form $g$ on $\cE_1$ that gives rise to $u$. Hence $\sigma$ is surjective. \end{proo} If $S$ is a Dedekind scheme, that is to say a n\oe therian normal scheme of dimension~$1$, then all coherent torsion ${\cal O}_S$-modules are torsion modules with few relations (by the structure theorem for modules of finite type). However in general it is not so, as soon as $\dim(S)\ge 2$, and we saw a counter-example before definition \ref{defi_TMFR}. We are now able to dualize the sequence of Lie algebras~\ref{exact_sequence_for_Lie}~(2) either if $H$ is smooth or if it is finite. \begin{prop} \label{prop_quotient_groupe_fini} Let $G$ be a smooth $S$-group scheme and $H\subset G$ a subgroup scheme which is flat and of finite presentation over $S$. Let $K=G/H$ be the quotient. \begin{trivlist} \itemn{1} If $H$ is smooth over $S$, then we have an exact sequence of locally free Lie algebra ${\cal O}_S$-modules $$0\to\mathscr{L}ie(H)\to\mathscr{L}ie(G)\to\mathscr{L}ie(K)\to 0$$ and if furthermore $K$ is commutative, we have $$[\mathscr{L}ie(G),\mathscr{L}ie(G)]\subset \mathscr{L}ie(H)\;.$$ \itemn{2} If $H$ is finite over $S$ and there is a schematically dominant morphism $i:U\to S$ such that $H\times_S U$ is {\'e}tale over $U$, then there is a canonical exact sequence of coherent ${\cal O}_S$-modules: $$ 0\to \mathscr{L}ie(G)\to \mathscr{L}ie(K)\to (\omega^1_{H/S})^\dag\to 0 \;, $$ and if furthermore $H$ is commutative, then $$[\mathscr{L}ie(K),\mathscr{L}ie(K)]\subset \mathscr{L}ie(G)\;.$$ \end{trivlist} \end{prop} \begin{proo} (1) All the sheaves in the exact sequence~\ref{exact_sequence_for_Lie}~(2) are locally free, so dualization yields the asserted result. It is clear that the resulting sequence is an exact sequence of sheaves of Lie algebras, so $[\mathscr{L}ie(G),\mathscr{L}ie(G)]\subset \mathscr{L}ie(H)$ in case $K$ is commutative. \noindent (2) The exact sequence~\ref{exact_sequence_for_Lie}~(2) and proposition~\ref{characterization_of_TMFRs} imply that $\omega^1_{H/S}$ is a torsion module with few relations. We get the dual sequence from proposition~\ref{Pontryagin_duality}. Here $(\omega^1_{H/S})^\dag$ is not a Lie algebra, so it is a little more subtle to deduce that $[\mathscr{L}ie(K),\mathscr{L}ie(K)]\subset \mathscr{L}ie(G)$. The assertion is local on $S$ so we may assume that $H$ is embedded into an abelian scheme $A/S$ (that is to say a smooth proper group scheme over $S$ with geometrically connected fibers), by a theorem of Raynaud (\cite{BBM}, Theorem 3.1.1). Let $\pi:A\to B=A/H$ be the quotient abelian scheme, and let $G'=(G\times_S A)/H$ where $H$ acts by $h(g,a)=(hg,h^{-1}a)$. We have two exact sequences of smooth $S$-schemes: $$ 1\to G\to G'\stackrel{p}{\to} B\to 1 $$ and $$ 1\to A\to G'\to K\to 1 \;. $$ By smoothness we derive exact sequences of sheaves of Lie algebras $$ 0\to \mathscr{L}ie(G)\to \mathscr{L}ie(G')\stackrel{p}{\to} \mathscr{L}ie(B)\to 0 $$ and $$ 0\to \mathscr{L}ie(A)\stackrel{i}{\to} \mathscr{L}ie(G')\to \mathscr{L}ie(K)\to 0\;. $$ Combining these exact sequences we have an exact sequence $$ 0\to \mathscr{L}ie(G)\to\mathscr{L}ie(K)\to\mathscr{L}ie(B)/\pi(\mathscr{L}ie(A))\to 0 $$ where $\pi=p\circ i$. Here, the arrow $\mathscr{L}ie(K)\to\mathscr{L}ie(B)/\pi(\mathscr{L}ie(A))$ is induced by $p$ which is a morphism of Lie algebras. It follows immediately that $[\mathscr{L}ie(K),\mathscr{L}ie(K)]\subset \mathscr{L}ie(G)$. \end{proo} \subsection{Lie algebras of Chevalley group schemes} \label{subsection_algebre_lie} \indent Let $G$ be a split simple Chevalley group scheme over $\mathbb Z$, $T \subset G$ a split maximal torus over $\bZ$, and write as in \cite{sga} $T=D_\bZ(M)$, where $M$ is a free $\bZ$-module. Since $G$ is smooth, $\Lie(G)$ is a vector bundle and hence is determined by $\mathscr{L}ie(G)$. Since the base is affine, this is in turn determined by the free $\mathbb Z$-module $\mathscr{L}ie(G)(\mathbb Z)=\Lie(G)(\mathbb Z)$ together with its Lie bracket. \begin{prop} \label{prop_decomposition} There is a weight decomposition $$\Lie(G)(\mathbb Z) = \Lie(T)(\mathbb Z) \oplus \bigoplus_\alpha \Lie(G)(\mathbb Z)_\alpha$$ over the integers. Moreover, letting $Q(R)$ (resp. $P(R)$) denote the root (resp. weight) lattice, we have $Q(R) \subset M\subset P(R)$. \end{prop} \begin{proo} Since $G$ is a smooth split reductive group scheme over $\mathbb Z$, this essentially follows from \cite{sga}, Expos{\'e}~I, 4.7.3, as explained in \cite{sga}, Expos{\'e}~XIX, no 3. \end{proo} Now let $H$ be a subgroup of the center of $G$ and let $i_H$ denote the inclusion of the character group of $T/H$ in that of $T$. \begin{prop} Under the natural inclusions $$\Lie(G)(\mathbb Z) \subset \Lie(G)(\mathbb{Q}) = \Lie(G/H)(\mathbb{Q}) \supset \Lie(G/H)(\mathbb Z) \ ,$$ we have $\Lie(G)(\mathbb Z)_{i_H(\alpha)} = \Lie(G/H)(\mathbb Z)_\alpha$. \end{prop} \begin{proo} Since $H\subset T$, by proposition \ref{prop_quotient_groupe_fini}, there are injections $\Lie(G)(\mathbb Z) \subset \Lie(G/H)(\mathbb Z)$ and $\Lie(T)(\mathbb Z) \subset \Lie(T/H)(\mathbb Z)$, both of index $|H|$. All these maps are compatible with the injection in $\Lie(G)(\mathbb{Q})$. Thus for each $\alpha$, the inclusion $\Lie(G)(\mathbb Z)_{i_H(\alpha)} \subset \Lie(G/H)(\mathbb Z)_\alpha$ must be of index 1, proving the proposition. \end{proo} \begin{rema} \label{rema_serre} The Lie algebra over $\mathbb Z$ defined by generators and relations by Serre \cite{serre} is the simply-connected one, that is to say the Lie algebra of the simply-connected corresponding group scheme, because, with his notations, the generators $H_i$ are by definition the coroots. \end{rema} Recall that $\pi : G \rightarrow G/H$ denotes the quotient morphism. \begin{prop} \label{prop_commutateur} Assume $G$ is simply-connected. \begin{trivlist} \itemn{1} When $G$ is not $Sp_{2n}$, $n\ge 1$, we have $$[\Lie(G)(\mathbb Z),\Lie(G)(\mathbb Z)] = \Lie(G)(\mathbb Z)$$ and $[\Lie(G/H)(\mathbb Z),\Lie(G/H)(\mathbb Z)] = d\pi(\Lie(G)(\mathbb Z))$. \itemn{2} When $G = Sp_{2n}$, then $[\Lie(G)(\mathbb Z),\Lie(G)(\mathbb Z)]$ has index $2^{2n}$ in $\Lie(G)(\mathbb Z)$. \end{trivlist} \end{prop} \begin{proo} (1) Let $\mathfrak g = \Lie(G)(\mathbb Z)$ and choose a Cartan $\mathbb Z$-subalgebra $\fh \subset \mathfrak g$. Choose a basis of the roots, and denote by $\fu_+,\fu_- \subset \mathfrak g$ the direct sum of the positive (resp. negative) root-spaces. By corollary \ref{prop_decomposition}, we have $\mathfrak g = \mathfrak h \oplus \u_+ \oplus \u_-$. Since $G$ is neither $SL_2 (=Sp_2)$ nor $Sp_{2n}$, by lemma~\ref{racine_non_primitive}, no root is an integer multiple of a weight, and so $[\fh,\fu_\pm] = \fu_\pm$. Moreover, it follows from Serre's presentation of the simple Lie algebras in terms of the Cartan matrix (see remark~\ref{rema_serre}) that in this case $[\mathfrak g,\mathfrak g] \supset \mathfrak h$. In particular $$ \begin{array}{rcl} d\pi(\Lie(G)(\mathbb Z)) & = & d\pi([\Lie(G)(\mathbb Z),\Lie(G)(\mathbb Z)]) \\ & = & [d\pi(\Lie(G)(\mathbb Z)),d\pi(\Lie(G)(\mathbb Z))] \\ & \subset & [\Lie(G/H)(\mathbb Z) , \Lie(G/H)(\mathbb Z)] \ . \\ \end{array} $$ The reverse inclusion follows from proposition \ref{prop_quotient_groupe_fini}. \lpara \noindent (2) Assume that $G$ stabilises the form $\matdd 0{I_n}{-I_n}0$, where $I_n$ stands for the identity matrix. Then $$ \mathfrak g = \left \{ \matdd AB C{-{}^t A} : {}^t B = B,{}^t C = C \right \}. $$ If $A$ is an arbitrary matrix and $B$ is symmetric, then we have the equality $$\left [ \matdd A0 0{-{}^t A}, \matddr 0B 00 \right ] = \matdd 0{AB+B{}^t A} 00.$$ From this it follows that $[\mathfrak g,\mathfrak g] \subset \mathfrak g$ is the $\mathbb Z$-submodule of elements $\matdd AB C{-{}^t A}$ with $B$ and $C$ having even diagonal elements. Therefore, it is a submodule of index $2^{2n}$. \end{proo} \subsection{The differential of the quotient maps} \label{diff_quotient_maps} We will now describe the differentials of the quotient maps between Chevalley groups in the neighbourhood of a prime $p\in\Spec(\bZ)$. So we consider the base ring $R=\bZ_{(p)}$. Let $G$ be simply-connected and let $n$ be the order of the center of $G$. Assume moreover that the center of $G$ is the group of $n$-th roots of unity $\mu_{n}$ (this is the case if $G$ is not of type $D_{2l}$; for this particular case see subsection \ref{subsection_classical}). Write $n=p^km$ with $m$ prime to $p$, and $G_i:=G/\mu_{p^i}$. We have the successive quotients $$ G=G_0\to G_1\to G_2\to\dots\to G_k\to G^{ad} $$ and the corresponding sequence of Lie algebras $$ \Lie(G)=\Lie(G_0)\to \Lie(G_1)\to\Lie(G_2)\to\dots\to\Lie(G_k) \stackrel{\simeq}{\longrightarrow} \Lie(G^{ad}) \ . $$ On the generic fibre all these maps are isomorphisms. In order to study what happens on the closed fibre, we set $\fg=\Lie(G)(\mathbb{F}_p)$ and $\fg_i=\Lie(G_i)(\mathbb{F}_p)$, and we let $\mathfrak{z}_i$ resp. $\mathfrak{z}$ denote the center of $\fg_i$ resp. $\fg$. We start with a lemma: \begin{lemm} \label{lemm_centre} The center $\mathfrak{z}_i$ is isomorphic to the one-dimensional Lie algebra $\mathbb{F}_p$ if $i<k$, and the algebra $\fg_k$ has trivial center. \end{lemm} \begin{proo} Let $x\in\fg_i$ be a central element. According to the decomposition of proposition \ref{prop_decomposition}, we can write $ x = \sum x_\alpha + h$. The lemma is easily checked directly when $\fg_i=\fsl_2$ or $\fg_i=\mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n}$, so assume we are not in these cases. According to the following lemma~\ref{lemm_racine}, for any root $\beta$ we may assume that there exists $t \in \ft$ such that $d\beta (t) \not = 0$. We then have $0 = [ t , x ] = \sum d\alpha(t) x_\alpha$, from which it follows that $x_\beta = 0$. Thus $x=h \in \ft$. Now, let again $\beta$ be an arbitrary root and let $0 \not = y \in (\fg_k)_\beta$. We have $0 = [ x , y ] = d\beta(x) . y$, therefore $d\beta(x) = 0$. Since we can reverse the above argument, the center of $\fg_i$ consists of all the elements in $\ft$ along which all the roots vanish. With the notations of proposition \ref{prop_decomposition}, $\ft \simeq M^\vee \otimes \mathbb{F}_p$, and therefore $\mathfrak{z}_i \simeq \Hom(M/Q(R),\mathbb{F}_p)$. Since $M/Q(R) \subset P(R)/Q(R)$ and in our case $P(R)/Q(R)$ is principal, $M/Q(R)$ is also principal and $\mathfrak{z}_i$ can be at most 1-dimensional. Moreover, it is trivial \iff $Q(R)=M$, which means that $\fg_i$ is adjoint, or $i=k$. \end{proo} \begin{lemm} \label{lemm_racine} Assume that $\fg$ is neither isomorphic to $\fsl_2$ nor $\mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n}$, or that the characteristic of $k$ is not $2$. Then there exists a finite extension $K$ of $k$ and $t \in \ft \otimes K$ such that $\forall \alpha \in R$, $d\alpha(t) \not = 0$. \end{lemm} \begin{proo} Let $R$ denote the root system of $G$ and let $p$ denote the characteristic of $k$; we have by proposition \ref{prop_decomposition} $Q(R) \subset M \subset P(R)$. The linear functions on $\ft$ defined over $k$ are in bijection with $M \otimes k$; therefore a root $\alpha \in Q(R)$ will yield a vanishing function on $\ft$ \iff it is a $p$-multiple of some element in $M$. By lemma \ref{racine_non_primitive}, this can occur only if $p=2$, $M = P(R)$ (thus $G$ is simply-connected) and $G$ is of type $A_1$ or $C_r$. By assumption we are not in these cases. Taking a finite extension $K/k$ if needed, $\fg$ is not a union of a finite number of hyperplanes, so the lemma is proved. \end{proo} Let $\fg' := \fg / \mathfrak{z}$. We can now describe the maps $\Lie(G_i)\to\Lie(G_{i+1})$ on the closed fibre: \begin{prop} The Lie algebras $\fg_i$ are described as follows: \begin{trivlist} \itemn{1} for all $i$ with $0 < i < k$, we have an isomorphism of Lie algebras $\fg_i\simeq \fg' \oplus \mathbb{F}_p$. In particular, all these Lie algebras are isomorphic. \itemn{2} for $i=0$ we have a non-split exact sequence of Lie algebras $0 \rightarrow \mathbb{F}_p \rightarrow \fg_0 \rightarrow \fg' \rightarrow 0$. \itemn{3} for $i=k$ we have a non-split exact sequence of Lie algebras $0 \rightarrow \fg' \rightarrow \fg_k \rightarrow \mathbb{F}_p \rightarrow 0$. \end{trivlist} In these terms, the maps $\fg_i\to \fg_{i+1}$ are described as follows. The map $\fg_0\to \fg_1$ takes $\mathbb{F}_p$ to zero and maps onto $\fg'\subset \fg_1$, and for all $i$ with $0 < i < k$, the map $\fg_i\to \fg_{i+1}$ takes $\mathbb{F}_p$ to zero and maps $\fg'\subset \fg_i$ isomorphically onto $\fg'\subset \fg_{i+1}$. \end{prop} \begin{proo} Let $Z_i:=\ker(G_i\to G_{i+1})$. For all $i\le k-1$, we have $Z_i\simeq \mu_p$, and its Lie algebra is included in $\mathfrak{z}_i$. Because $Z_i\to G_{i+1}$ is trivial, the map $\fg_i\to \fg_{i+1}$ takes $\mathfrak{z}_i$ to $0$. By tensoring the result of proposition \ref{prop_quotient_groupe_fini} by $\mathbb{F}_p$, there is an exact sequence $$ \fg_i \to \fg_{i+1} \to \bZ/p\bZ \to 0\ , $$ from which it follows that $\fg_i/\mathfrak{z}_i$ is mapped isomorphically onto a codimension $1$ subalgebra of $\fg_{i+1,{\mathbb F}} \def\bG{{\mathbb G}} \def\bH{{\mathbb H}_p}$ denoted $\fg'_{i+1,{\mathbb F}} \def\bG{{\mathbb G}} \def\bH{{\mathbb H}_p}$. By lemma \ref{lemm_centre}, no $x \in \fg'_{i+1,{\mathbb F}} \def\bG{{\mathbb G}} \def\bH{{\mathbb H}_p}$ can be central in $\fg_{i+1}$ so that we have, for $0 < i < k$, $$\fg_{i,\mathbb{F}_p} = \fg_{i,\mathbb{F}_p}' \oplus \mathfrak{z}_{i,\mathbb{F}_p}$$ as vector spaces. Since $\mathfrak g_{i,\mathbb{F}_p}'$ is a Lie subalgebra, it is also an equality of Lie algebras. In particular all the Lie algebras $\fg_i$ for $0 < i < k$ are isomorphic. For $i=0$, we have an exact sequence of Lie algebras $0 \rightarrow \mathbb{F}_p \rightarrow \fg_{0,\mathbb{F}_p} \rightarrow \fg' \rightarrow 0$, but this sequence does not split (in fact if it did split, then we would have $[\fg_{0,\mathbb{F}_p} , \fg_{0,\mathbb{F}_p}] \subset \fg'$, contradicting proposition \ref{prop_commutateur}). For $i=k>0$, we have a sequence $0 \rightarrow \fg' \rightarrow \fg_{k,\mathbb{F}_p} \rightarrow \mathbb{F}_p \rightarrow 0$, which again cannot split because $\fg_{k,\mathbb{F}_p}$ has trivial center, by lemma \ref{lemm_centre}. \end{proo} \begin{rema} For $0<i<k$, consider the Lie algebras $\Lie(G_i)$, as schemes over $\Spec(\bZ_{(p)})$. They have isomorphic underlying vector bundles (namely the trivial vector bundle), and their generic fibres as well as their special fibres are isomorphic as Lie algebras. However, they need not be isomorphic. For example, if $G=SL_{p^k}$ then it is immediate from proposition~\ref{prop_commutateur} that $G_i=G/\mu_{p^i}$ uniquely determines $i$, because $[\Lie(G_i)(\bZ),\Lie(G_i)(\bZ)]=\Lie(G)(\bZ)$ and the quotient $\Lie(G_i)(\bZ)/\Lie(G)(\bZ)$ is a cyclic abelian group of order $p^i$. In fact the only difference between them comes from the definition of the Lie bracket on the total space. \end{rema} \subsection{Classical Lie algebras} \label{subsection_classical} We now give an explicit description of some of the classical Chevalley Lie algebras, in which the above sequence of Lie algebras will become very transparent. Let $M$ be a free $\mathbb Z$-module of rank $n$, and let $m$ be an integer dividing $n$. We define a $\mathbb Z$-Lie algebra $L(M|m)$ as follows: let $\Hom(M,\frac 1m M)$ denote the $\mathbb Z$-module of linear maps $f:M\otimes{\mathbb Q}} \def\bT{{\mathbb T}} \def\bV{{\mathbb V}\to M\otimes{\mathbb Q}} \def\bT{{\mathbb T}} \def\bV{{\mathbb V}$ such that $f(M)\subset \frac 1m M$. Any such map induces a map $\overline f:M/mM\to \frac 1m M/M$. Note that multiplication by $m$ induces a canonical isomorphism $\frac 1m M/M\simeq M/mM$ so that $\overline f$ may be seen as an endomorphism of the free $\bZ/m\bZ$-module $M/mM$. Finally, let $L(M|m)$ (resp. $S(M|m)$) denote the submodule of $\Hom(M,\frac 1m M)$ of elements $f$ such that $\overline f$ is a homothety (resp. a homothety with vanishing trace). These are obviously Lie subalgebras of $\End(M\otimes{\mathbb Q}} \def\bT{{\mathbb T}} \def\bV{{\mathbb V})$. \begin{prop} \label{prop_psln} Let $n,m$ be as above; then $\Lie(SL_n / \mu_m)(\mathbb Z) \simeq S(\mathbb Z^n|m)$. \end{prop} \begin{proo} Let $n,m$ be integers such that $m$ divides $n$. Let $\fsl_n$ denote the Lie algebra of $SL_n$, and let $\fsl_{n,m}$ denote the Lie algebra of the quotient $SL_n/\mu_m$. Obviously, we have $\fsl_n(\mathbb Z) \subset \fsl_n(\mathbb{Q})$, $\fsl_{n,m}(\mathbb Z) \subset \fsl_{n,m}(\mathbb{Q})$, and $\fsl_n(\mathbb{Q}) = \fsl_{n,m}(\mathbb{Q})$ is the usual Lie algebra over $\mathbb{Q}$ of traceless matrices. The exact sequence of proposition \ref{exact_sequence_for_Lie} translates in our case to $$ 0 \rightarrow \fsl_{n,m}(\mathbb Z)^\vee \rightarrow \fsl_n(\mathbb Z)^\vee \rightarrow \mathbb Z/m\mathbb Z \rightarrow 0. $$ Note that $\fsl_n(\mathbb Z)^\vee \simeq \mathfrak{gl}} \def\fsl{\mathfrak{sl}_n(\mathbb Z)^\vee / (trace)$, so that evaluation at $I$ modulo $m$ is well-defined and yields the last arrow. Let $f_{i,j} : \fsl_n(\mathbb{Q}) \rightarrow \mathbb{Q}$ be the linear form $M \mapsto M_{i,j}$. Thus $\fsl_{n,m}(\mathbb Z)^\vee \subset \fsl_n(\mathbb Z)^\vee$ is the set of linear forms $\sum \lambda_{i,j}f_{i,j}$ with $m|\sum \lambda_{i,i}$. Since $\fsl_{n,m}(\mathbb Z)$ is the dual in $\fsl_n(\mathbb{Q})$ of $\fsl_{n,m}(\mathbb Z)^\vee$, it follows applying $f_{i,j}$ $(i \not = j)$ (resp. $m.f_{i,i}$, $f_{i,i} - f_{j,j}$) that if $M \in \fsl_n(\mathbb{Q})$ belongs to $\fsl_{n,m}(\mathbb Z)$, then $M_{i,j} \in \mathbb Z$ (resp. $m.M_{i,i} \in \mathbb Z, M_{i,i}-M_{j,j} \in \mathbb Z$). Conversely, matrices satisfying these conditions are certainly in $\fsl_{n,m}(\mathbb Z)$. The proposition is therefore proved. \end{proo} Now, assuming that $n$ is even, we describe the four Lie algebras of type $D_n$: $\fspin_{2n}$, $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n}$, $\fpso_{2n}$, and a fourth one that we denote by $\mathfrak{pspin}_{2n}$. We need to introduce some notation. \begin{nota} Let $n$ be an integer and let us consider the following sublattices of $\mathbb{Q}^n$: \begin{itemize} \item $N_{z} = \bZ^n$. \item $N_{ad}$ is generated by $\mathbb Z^n$ and $\frac 12 (1,\ldots,1)$. \item If $l$ is even, $N_{ps}$ is the sublattice of $N_{ad}$ of elements $(x_i)$ with $\sum x_i$ divisible by 2. \item $N_{sc}$ is the sublattice of $\mathbb Z^n$ of elements $(x_i)$ with $\sum x_i$ divisible by 2. \item We denote by $L_*$ ($* \in \{z,ad,ps,sc \}$) the lattice of matrices with off-diagonal coefficients in $\bZ$ and with the diagonal in $N_*$. \end{itemize} \end{nota} \begin{prop} \begin{trivlist} \itemn{1} There is a natural identification of $\mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n}(\bZ)$ (resp. $\fpsp_{2n}(\bZ)$) with the Lie algebra of matrices of the form $\matdd ABC{-{}^t A}$ with ${}^t B=B,{}^t C=C$ $(n\times n)$-matrices with coefficients in $\bZ$ and $A \in \mathfrak{gl}} \def\fsl{\mathfrak{sl}_n(\bZ)$ (resp. $ A \in L(\bZ^n|2)$). \itemn{2} Assume $n$ is odd. Then there is a natural identification of $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n}(\bZ)$ (resp. $\fpso_{2n}(\bZ)$, $\fspin_{2n}(\bZ)$) with the Lie algebra of matrices of the form $\matdd ABC{-{}^t A}$ with ${}^t B = -B, {}^t C = -C$ and $A$ in $L_z$ (resp. $L_{ad}$, $L_{sc}$). \itemn{3} Assume $n$ is even. There is a natural identification of $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n}(\bZ)$ (resp. $\fpso_{2n}(\bZ)$, $\fspin_{2n}(\bZ)$, $\mathfrak{pspin}_{2n}(\bZ)$) with the Lie algebra of matrices of the form $\matdd ABC{-{}^t A}$ with ${}^t B = -B, {}^t C = -C$ and $A$ in $L_z$ (resp. $L_{ad}$, $L_{sc}$, $L_{ps}$). \end{trivlist} \end{prop} \begin{proo} This is a direct consequence of proposition \ref{prop_decomposition}. For example, let $\fg(\bZ)$ be a Lie algebra of type $D_n$ over $\bZ$. Proposition \ref{prop_decomposition} implies that all Lie algebras of type $D_n$ will differ only by their Cartan subalgebras. Therefore $\fg(\bZ)$ is in fact a set of matrices of the form $\matdd ABC{-{}^t A}$ with ${}^t B = -B, {}^t C = -C$, and the off-diagonal coefficients of $A$ in $\bZ$. Moreover, with the notations of proposition \ref{prop_decomposition}, $\fg(\bZ)$ corresponds to a module $M$ between the root lattice and the weight lattice, and the Cartan subalgebra (the subalgebra when $A$ is diagonal and $B=C=0$) identifies with the dual lattice of $M$. Therefore the description of the proposition follows from the description of the root and weight lattices in \cite{bourbaki}. \end{proo} Using this description or lemma \ref{lemm_centre} we can deduce the dimension of the center of $\fg(\mathbb{F}_2)$: \medskip $$ \begin{array}{lcccc} & \fspin_{2n}(\mathbb{F}_2) & \mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n}(\mathbb{F}_2) & \mathfrak{pspin}_{2n}(\mathbb{F}_2) & \fpso_{2n}(\mathbb{F}_2) \\ n \mbox{ even } & 2 & 1 & 1 & 0 \\ n \mbox{ odd } & 1 & 1 & - & 0 \\ \end{array} $$ \medskip For example, we give a description of the center of $\fspin_{2n}(\mathbb{F}_2)$. Let $C_1$ (resp. $C_2$) be the matrix of the form $\matdd ABC{-{}^t A}$ with $B=C=0$ and $A=I_n$ (the identity matrix, resp. $A=\matddr 2000$, the matrix with only one non-vanishing coefficient in the top-left corner, equal to 2). Note that $C_1,C_2 \in \fspin_{2n}(\bZ)$ (but $C_1,C_2$ are not divisible by 2 in $\fspin_{2n}(\bZ)$). For any matrix $B$ in $\fspin(\bZ) \subset \mathfrak{so}} \def\fpso{\mathfrak{pso}(\bZ)$, we obviously have $[C_1,B]=0$ and $2 | [C_2,B]$. From these remarks it follows that the classes of $C_1$ and $C_2$ modulo $2.\fspin(\bZ)$ generate over $\mathbb{F}_2$ the 2-dimensional center of $\fspin_{2n}(\mathbb{F}_2)$. \section{The Chevalley morphism} \label{section_morphisme_de_Chev} We start this section by some elementary results on regular elements in Lie algebras of algebraic groups (\ref{section_regular_elements}), to be used in~\ref{subsection_Chevalley_iso}, then we prove that the Chevalley morphism is unconditionnally schematically dominant (\ref{subsection_chevalley_dominant}) and finally we prove the the Chevalley morphism is an isomorphism unless $G=Sp_{2n}$ for some $n\ge 1$ (\ref{subsection_Chevalley_iso}). The case $G=Sp_{2n}$ will be treated later. \subsection{Regular elements} \label{section_regular_elements} Let $\fg$ be a restricted Lie algebra of dimension $d$ over a field. For each $x\in \fg$, we denote by $\chi(x)=t^d+c_1(x)t^{d-1}+\dots+c_{d-1}(x)t+c_d(x)$ the characteristic polynomial of $\ad x$ acting on~$\fg$. The {\em rank} of $\fg$ is the least integer $l$ such that $c_{d-l}\ne 0$, and we set $\delta:=c_{d-l}$. An element $x\in \fg$ is called {\em regular} if the nilspace $\fg_0(\ad x):=\ker (\ad x)^d$ of $\fg$ relative to $\ad x$ has minimal dimension~$l$. An element $x$ is regular if and only if $\delta(x)\ne 0$. Note that our definition of regular elements differs from that in \cite{Ve}, according to which the singular locus has codimension 3 in $\fg$. The Cartan subalgebras of minimal dimension are exactly the centralizers of regular elements. For these facts see \cite{Str}, pages 52-53. If, furthermore, $\fg$ is the Lie algebra of a smooth connected group $G$, then in fact the Cartan subalgebras are conjugate, and in particular they all have the same dimension. This is proven in \cite{Fa}, corollary 4.4. Also, in this case the coefficients of the characteristic polynomial $\chi$ are invariant functions for the adjoint action of $G$ on $\fg$. We will use the notation $\Sing(\fg)$ for the closed subscheme of singular elements, defined by the equation $\delta=0$, and $\Reg(\fg)$ for its complement, the open subscheme of regular elements. We have the corresponding subschemes $\Sing(\ft)$ and $\Reg(\ft)$ in $\ft$. In the relative situation, if $\fg$ is a Lie algebra of dimension $d$ over a scheme $S$, then the objects $\chi$, $\delta$, $\Reg(\fg)$, $\Sing(\fg)$ are defined by the same procedure as above. We recall our general convention that a {\em relative Cartier divisor} of some $S$-scheme $X$ is an effective Cartier divisor in $X$ which is flat over $S$. \begin{lemm} \label{lemm_sing_cartier} Let $G$ be a split simple Chevalley group over a scheme $S$, not isomorphic to $Sp_{2n}$, $n\ge 1$. Let $s:\Sing(\fg)\to S$ be the locus of singular elements. Then $s_*{\cal O}_{\Sing(\fg)}$ is a free ${\cal O}_S$-module, in particular $\Sing(\fg)$ is a relative Cartier divisor of $\fg$ over $S$. \end{lemm} \begin{proo} Since the objects involved have formation compatible with base change, it is enough to prove the lemma over $S=\Spec(\bZ)$. We have to prove that the ring $\bZ[\fg]/(\delta)$ is free as a $\bZ$-module. Since $\delta$ is homogeneous, this ring is graded. If we can prove that it is flat over $\bZ$, then its homogeneous components are flat also, and since they are finitely generated, they are free over $\bZ$, and the result follows. So it is enough to prove that $\bZ[\fg]/(\delta)$ is flat. By the corollary to theorem 22.6 in \cite{Ma}, it is enough to prove that the coefficients of $\delta$ generate the unit ideal, or in other words, that $\delta$ is a nonzero function modulo each prime $p$. So we may now assume that the base is a field $k$ of characteristic $p\ge 0$, and we may also assume that $k$ is algebraically closed. Let $\ft$ be the Lie algebra of a maximal torus $T$. By lemma~\ref{lemm_racine}, we can choose $t \in \ft_k$ such that $\forall \alpha \in R$, $d\alpha(t) \not = 0$. Then $\delta(t)$ is the product of the scalars $d\alpha(t)$, up to a sign. Hence it is nonzero. \end{proo} We continue with the split simple Chevalley group $G$ over $S$. In the sequel, products are understood to be fibred products over $S$. We now turn our attention to the morphism $G/T\times\ft\to \fg$. We use the same construction as in \cite[3.17]{SS}: note that the normalizer $N_G(T)$ acts on $G\times \ft$ by $n.(g,\tau) = (gn^{-1},\Ad(n).\tau)$ and this induces an action of $W=N_G(T)/T$ on $G/T \times \ft$. The morphism $G/T\times \ft\to \fg$ induced by the adjoint action is clearly $W$-invariant. \begin{lemm} \label{lemm_birational} Let $G$ be a split simple Chevalley group over a scheme $S$, not isomorphic to $Sp_{2n}$, $n\ge 1$. Then the map $G/T\times \ft\to \fg$ is schematically dominant. Its restriction $$ b:G/T\times \Reg(\ft)\to \Reg(\fg) $$ is a $W$-torsor and hence induces an isomorphism $(G/T\times \Reg(\ft))/W\to \Reg(\fg)$. \end{lemm} \begin{proo} Here again, the objects involved have formation compatible with base change, so it is enough to prove the lemma over $S=\Spec(\bZ)$. We will first prove that $b$ is a $W$-torsor. It is enough to prove that $b$ is surjective, \'etale, and that $W$ is simply transitive in the fibres. Indeed, if $b$ is \'etale then the action must also be free and $b$ induces an isomorphism $(G/T\times \Reg(\ft))/W\to \Reg(\fg)$. The map $c=\ad:G\times \Reg(\ft)\to \Reg(\fg)$ is surjective because if $x\in \Reg(\fg)$, then its centralizer $\mathfrak{z}(x)$ is a Cartan subalgebra, and since Cartan subalgebras are conjugate, there exists $g\in G$ such that $(\ad g)(\ft)=\mathfrak{z}(x)$. Thus there is $y\in\ft$ such that $(\ad g)(y)=x$ and clearly $y$ is regular. We now prove that $c$ is smooth. Since its source and its target are smooth over $S$, it is enough to prove that for all $s\in S$, the map $c_s$ is smooth. By homogeneity, it is enough to prove that the differential of $c_s$ at any point $(1,t)$ with $t \in \Reg(\ft)$ is surjective. Then $T_1G_k = \fg_k$ and the tangent map $\psi=dc : T_1G_k \times \ft_k \rightarrow \fg_k$ is given by $(x,\tau) \mapsto [x,t] + \tau$. Recall that $\fg_k = \bigoplus_{\alpha \in R} \fg_\alpha \oplus \ft_k$, where $[\tau,x] = d\alpha(\tau)x$ for all $x \in \fg_\alpha$. Again by lemma~\ref{lemm_racine}, we can choose $\tau \in \ft_k$ such that $\forall \alpha \in R$, $d\alpha(\tau) \not = 0$. Thus, we have $\psi(\fg_k \times \{0\}) = \bigoplus_{\alpha \in R} \fg_\alpha$. Since $\psi(\{0\} \times \ft_k) = \ft_k$, $\psi$ is a surjective linear map. It follows that $b$ is also surjective and smooth, hence \'etale, by dimension reasons. Finally, let $(g,x)$ and $(h,y)$ have the same image in $\fg$, for $x,y\in\ft$. This means that $(\ad w)(x)=y$, where $w=h^{-1}g$. Thus $(\ad w)(\mathfrak{z}(x))=\mathfrak{z}((\ad w)(x)=\mathfrak{z}(y)$, that is to say $(\ad w)(\ft)=\ft$ since $\mathfrak{z}(x)=\mathfrak{z}(y)=\ft$. By \cite[13.2,13.3]{humphreys}, $T$ is the only maximal torus with Lie algebra $\ft$, so it follows that $w$ normalizes $T$. Hence $w$ defines an element of the Weyl group $W$. Now we consider the map $G/T\times \ft\to \fg$. From the preceding discussion it is dominant in the fibres, and since $G/T\times\ft$ is flat over $S$, the map is itself schematically dominant by \cite{EGA}, th{\'e}or{\`e}me 11.10.9. This concludes the proof of the lemma. \end{proo} \subsection{The Chevalley morphism is always dominant} \label{subsection_chevalley_dominant} We now deal with the cases that are not covered by lemma~\ref{lemm_sing_cartier}. \begin{nota} Consider the following subalgebras of $\fsl_2$ and $\mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n}$: \begin{itemize} \item Let $\mathfrak{b}} \def\fg{\mathfrak{g}} \def\fh{\mathfrak{h} \subset \fsl_2$ be the subalgebra of upper-triangular matrices. \item Let $L$ denote the set of long roots of $\mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n}$; if $\alpha$ is a root, denote by $\mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n,\alpha}$ the corresponding root space. \item Let $\fh \subset \mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n}$ be the sum $\ft \oplus \bigoplus_{\alpha \in L} \mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n,\alpha}$. \end{itemize} \end{nota} \begin{lemm} \label{lemm_dominant_sur_fibres_sp} Let $k$ be a field. Then the maps $(SL_2)_k \times \mathfrak{b}} \def\fg{\mathfrak{g}} \def\fh{\mathfrak{h}_k \rightarrow (\fsl_2)_k$ and $(Sp_{2n})_k \times \fh_k \rightarrow (\mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n})_k$, given by restricting the adjoint action, are dominant. Moreover, $\fh$ is isomorphic, as a Lie algebra, to $\fsl_2^{\oplus n}$. \end{lemm} Combining the two statements of this lemma, it follows that in the proof of theorem~\ref{theo_chevalley_dominant} below, we will be able to replace the Lie subaglebra $\fh \simeq \fsl_2^{\oplus n}$ by a sum of the form $\mathfrak{b}} \def\fg{\mathfrak{g}} \def\fh{\mathfrak{h}^{\oplus n}$. \begin{proo} The result about $(\fsl_2)_k$ is an immediate consequence of the fact that, over an algebraically closed field, any matrix is conjugated to an upper-triangular matrix. To prove that $(Sp_{2n})_k \times \fh_k \rightarrow (\mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n})_k$ is dominant, we argue as in lemma \ref{lemm_birational}. Since all the short roots are not integer multiples of a weight, we can choose $t \in \ft_k$ such that for all short roots $\alpha$, we have $d\alpha(t) \not = 0$. Let $S$ denote the set of short roots of $\mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n}$. If $\psi$ denotes the differential at $(1,t)$ of the ajoint action, it follows that $\psi(\mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n} \times \{0\}) \supset \bigoplus_{\alpha \in S} \mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n,\alpha}$. Since $\fh_k = \ft_k \oplus \bigoplus_{\alpha \in L} \mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n,\alpha}$, it follows that $\psi$ is surjective and the restriction of the action is dominant. To prove that $\mathfrak h \simeq \fsl_2^{\oplus n}$, one can compute explicitly in the Lie algebra $\mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n}$. Assume that $\mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n}$ is defined by the matrix $\matdd 0I{-I}0$, where $I$ denotes the identity matrix of size~$n$. Then a matrix $\matdd ABCD$ belongs to $\mathfrak{sp}} \def\fpsp{\mathfrak{psp}_{2n}$ \iff $D = -{}^t A$ and $B$ and $C$ are symmetric matrices. Choosing the torus $\ft = \left \{ \matdd d00{-d} \right \}$ in $\fsl_{2n}$, and $\epsilon_i$ the coordinate forms on $\ft$, it is well-known and easy to check that the long roots are $\pm 2\epsilon_i$. It follows that $\fh = \left \{ \matdd d \delta \epsilon {-d} : d,\delta,\epsilon \mbox{ diagonal} \right \}$. Thus $\fh$ is isomorphic to $\fsl_2^{\oplus n}$. \end{proo} \begin{lemm} \label{lemm_injectivite_sl2} Let $S=\Spec(A)$ be an affine base scheme and $\mathfrak g = \fsl_2$. Then the restriction morphism $A[\mathfrak{b}} \def\fg{\mathfrak{g}} \def\fh{\mathfrak{h}]^T \rightarrow A[\ft]$ is injective. \end{lemm} \begin{proo} Let $f \in A[\mathfrak{b}} \def\fg{\mathfrak{g}} \def\fh{\mathfrak{h}]$. Writing a typical element in $\mathfrak{b}} \def\fg{\mathfrak{g}} \def\fh{\mathfrak{h}$ as $\matddr ab0a$, we identify $f$ with a polynomial in $a,b$. Since $$\matddr t00{t^{-1}}\matddr ab0a \matddr t00{t^{-1}} ^{-1} = \matdd a{t^2b}0a,$$ $f$ is $T$-invariant \iff $f(a,b) = f(a,t^2b) \in A[a,b,t]$. This means that $f$ does not depend on $b$, and we indeed have an injection $A[\mathfrak{b}} \def\fg{\mathfrak{g}} \def\fh{\mathfrak{h}]^T \rightarrow A[\ft]$. \end{proo} Using the two preceding lemmas, we can now prove the main result of this section: \begin{theo} \label{theo_chevalley_dominant} Let $S$ be a scheme and let $G$ be a split simple Chevalley group over~$S$. Then the Chevalley morphism $\pi:\ft/W\to\fg/G$ is schematically dominant. \end{theo} \begin{proo} First, let $S=\Spec(\bZ)$. Let us write $\fh = \mathfrak{b}} \def\fg{\mathfrak{g}} \def\fh{\mathfrak{h}$ in the case of $SL_2$, $\fh = \fsl_2^{\oplus n}$ in the case of $Sp_{2n}$, and $\fh = \ft$ in the other cases. We also write $H = T$ (resp. $H=SL_2^{\oplus n} , H=T$) in the case of $SL_2$ (resp. $Sp_{2n}$, the other cases). The adjoint action restricts to a map $\varphi:G \times \fh \to \fg$. If $G$ acts on itself by left translation, trivially on $\fh$, and by the adjoint action on $\fg$, then $\varphi$ is $G$-equivariant. Moreover, by lemmas \ref{lemm_birational} and \ref{lemm_dominant_sur_fibres_sp}, the restriction $\varphi_k$ of $\varphi$ to any fiber of $\Spec(\mathbb Z)$ is dominant. Since the schemes $G \times \fh$ and $\fg$ are flat over $\mathbb Z$, it follows from \cite[th{\'e}or{\`e}me 11.10.9]{EGA} that $\varphi$ is universally schematically dominant. Now we let $S$ be arbitrary, and prove that $\pi$ is schematically dominant. The question is local over $S$ so we may assume $S=\Spec(A)$ affine. On the function rings, the Chevalley morphism can be decomposed as two successive restriction morphisms $A[\fg]^G \to A[\fh]^H \to A[\ft]^W$. First we concentrate on $A[\fg]^G \to A[\fh]^H$. We have already proven that the map $\varphi^* : A[\fg] \rightarrow A[G] \otimes_A A[\fh]$ is injective. This map is $G$-equivariant, so if $f \in A[\fg]$ is $G$-invariant, we have $$ \varphi^*(f) \in (A[G] \otimes_A A[\fh])^G = A[G]^G \otimes_A A[\fh] = A \otimes_A A[\fh] = A[\fh] \ . $$ Therefore $\varphi^*(f) = 1 \otimes i^*(f)$ where $i:\fh \rightarrow \fg$ is the inclusion. Since $\varphi^* : A[\fg] \rightarrow A[G] \otimes_A A[\fh]$ is injective, it follows that $A[\fg]^G \rightarrow A[\fh]$ is injective. In case $G \not = Sp_{2l}$, we have $\ft = \fh$ so the proof of the theorem is complete. In case $G = SL_2$, lemma \ref{lemm_injectivite_sl2} shows the injectivity of the Chevalley morphism. Finally, let us consider the case of $Sp_{2n}$ with $n>1$. Since by lemma \ref{lemm_dominant_sur_fibres_sp}, $\fh \simeq \fsl_2^{\oplus n}$ and $H \simeq SL_2^n$, and since the theorem is proved for $SL_2$, we know that $A[\fh]^H \to A[\ft]$ is injective, and we can once again conclude. \end{proo} It is interesting to mention an easy consequence of this theorem~: if the base scheme is $\Spec(\bZ)$, then the Chevalley morphism is an isomorphism, see the corollary below. This will of course be a particular case of theorems~\ref{theo_chevalley_iso} and \ref{theo_sp2n}, however, while theorem~\ref{theo_chevalley_iso} needs some more work, we get the present result almost for free. This corollary is in fact worthwhile because it shows that if the formation of the adjoint quotient commuted with base change (which is not the case), then we would get the case of a general base scheme $S$ from the case $S=\Spec(\bZ)$ and hence everything would be finished right now. So here is this corollary: \begin{coro} \label{coro_factorial_ring} Assume that $S$ is the spectrum of a factorial ring with characteristic prime to the order of $W$. Then the Chevalley morphism $\pi:\ft/W\to\fg/G$ is an isomorphism. \end{coro} \begin{proo} Let $S=\Spec(A)$ and let $K$ be the fraction field of $A$. By theorem~\ref{theo_chevalley_dominant} it is enough to prove that the restriction morphism $\res:A[\fg]^G \to A[\ft]^W$ is surjective. Let $P$ be a $W$-invariant function on $\ft$. From the assumption on the characteristic of $K$, it follows that $K[\fg]^G \to K[\ft]^W$ is an isomorphism, so there is $Q\in K[\fg]^G$ such that $\res(Q)=P$. Since $A$ is factorial, we can write $Q=cQ_0$ where $Q_0$ is a primitive polynomial (i.e. the gcd of its coefficients is $1$) and $c\in K$ is the content of $Q$. If we write $c=r/s$ with $r$ and $s$ coprime in $A$, we claim that $s$ is a unit in $A$. For, otherwise, some prime $p\in A$ divides $s$. Then $\res(\bar{rQ_0})=\bar{sP}=0$ in $(A/p)[\ft]^W$ so $\res(\bar{Q_0})=0$ in $(A/p)[\ft]^W$, since $r$ and $s$ are coprime. By theorem~\ref{theo_chevalley_dominant} again, it follows that $\bar{Q_0}=0$ in $(A/p)[\fg]^G$, in contradiction with the fact that $Q_0$ is primitive. Hence $Q\in A[\fg]^G$ as was to be proved. \end{proo} \subsection{The Chevalley morphism is an isomorphism for $G\ne Sp_{2n}$} \label{subsection_Chevalley_iso} Before proving theorem~\ref{theo_chevalley_iso}, our main result, we state two lemmas which will ultimately allow us to show that a function vanishes on the locus of singular elements. \smallskip The first lemma is a general result about the ``indicator function'' of the locus of the base where a function has a zero along a fixed divisor~: \begin{lemm} \label{indicator_of_zeroes} Let $X\to S$ be a morphism of schemes and let $D\subset X$ be a relative effective Cartier divisor with structure morphism $p:D\to S$, such that $p_*{\cal O}_D$ is a locally free ${\cal O}_S$-module. Let $*$ be the one-point set, and given a global function $f$ on $X$, let $F$ be the functor on the category of $S$-schemes defined as follows~: for any $S$-scheme $T$, $$ F(T)=\left\{ \begin{array}{cl} * & \mbox{if $f_T$ has a zero along $D_T$}, \\ \emptyset & \mbox{otherwise.} \\ \end{array} \right. $$ Then $F$ is representable by a subscheme $S_0\subset S$ which is open and closed in $S$. \end{lemm} \begin{proo} We first prove that $F$ is represented by a closed subscheme $S_0\subset S$. The assertion is local on $S$ so we may assume $S$ affine and $p_*{\cal O}_D$ free. Let $f_i$ be the finitely many components of $f_{|D}$ on some basis of $\Gamma(D,{\cal O}_D)$ as a free $\Gamma(S,{\cal O}_S)$-module. Then, $f$ has a zero along $D$ if and only if all the $f_i$ vanish. Since $D$ is a relative Cartier divisor, the formation of these objects commutes with base change, so that the above description is functorial. Then obviously $F$ is represented by the closed subscheme of finite presentation $S_0\subset S$ defined by the ideal of $\Gamma(S,{\cal O}_S)$ generated by the coefficients $f_i$. It order to prove that $S_0\subset S$ is also open, we will use another description of $F$. Let $L=\Spec(\Sym({\cal O}_X(D))$ and $d:L\to{\mathbb A}} \def\bC{{\mathbb C}} \def\bD{{\mathbb D}^1_X$ be the morphism of line bundles over $X$ induced by the canonical morphism of invertible sheaves ${\cal O}_X\to{\cal O}_X(D)$. We can view $f$ as a section of the trivial line bundle, i.e. $f:X\to{\mathbb A}} \def\bC{{\mathbb C}} \def\bD{{\mathbb D}^1_X$. Then to say that $f$ has a zero along $D$ means that there exists a section $\varphi:X\to L$ of $L$ such that the diagram $$ \xymatrix{X \ar[r]^\varphi \ar[rd]_f & L \ar[d]^d \\ & {\mathbb A}} \def\bC{{\mathbb C}} \def\bD{{\mathbb D}^1_X \\} $$ commutes. Since $D$ is a relative Cartier divisor, such a $\varphi$ is unique if it exists. In other words, $F(T)$ is the set of sections $\varphi:X_T\to L_T$ such that $f_T=d_T\circ\varphi$, or again, if we set $Z:=d^{-1}(f(X))$, then $F(T)$ is the set of sections $\varphi:X_T\to L_T$ such that $\varphi(X_T)$ is a closed subscheme of $Z_T$. In the sequel, we will take the freedom to write simply $X$ for the divisor $\varphi(X)\subset L$, and $D$ for the pullback of $D$ via the structure morphism $L\to X$. Since $\varphi$ is determined by its image, we have described $F$ as an open subfunctor of a Hilbert functor of $Z/S$. To prove that $S_0\subset S$ is open, it is enough to prove that this is a smooth morphism. Since $S_0\subset S$ is of finite presentation, it is enough to prove that it is formally smooth at all points $s\in S_0$. To do this, we will show that the deformations of $\varphi_s$ are unobstructed. Let $k$ be the residue field of $s$, let $A$ be a local artinian ${\cal O}_S$-algebra with residue field $k$ and assume that $\varphi_s$ has been lifted to $\varphi_A$. We have to prove that for each nilthickening $A'\to A$, i.e. a surjective morphism with kernel $M$ annihilated by the maximal ideal of~$A'$, the space of obstructions to lifting $\varphi_A$ to $A'$ vanishes. By the theory of the Hilbert functor, the obstruction space is $\Ext^1_{Z_A}(\cI_A,({\cal O}_Z/\cI)\otimes_k M)$ where $\cI$ is the ideal sheaf of $X$ in $Z$; note that $({\cal O}_{Z_A}/\cI_A)\otimes_k M=({\cal O}_Z/\cI)\otimes_k M$. We will compute this group by using an explicit resolution of $\cI$. In the sequel, we write $X$ for $X_A$, $D$ for $D_A$, etc. We remark that $Z=X+D$ as a sum of Cartier divisors of $L$. This is not hard to see and we leave the details to the reader. Now let $a_D:{\cal O}_L\to {\cal O}_L(D)$ and $a_X:{\cal O}_L\to {\cal O}_L(X)$ be the canonical morphisms. Consider the sequence $$ \dots \stackrel{a_X}{\longrightarrow}{\cal O}_L(-2Z)\stackrel{a_D}{\longrightarrow}{\cal O}_L(-X-Z) \stackrel{a_X}{\longrightarrow}{\cal O}_L(-Z)\stackrel{a_D}{\longrightarrow}{\cal O}_L(-X) \stackrel{a_X}{\longrightarrow}{\cal O}_L \ . $$ Note that this is not even a complex, but we claim that after restricting to $Z$, we get a resolution of the ideal sheaf $\cI$~: $$ \dots \stackrel{a_X}{\longrightarrow}{\cal O}_L(-2Z)|_Z\stackrel{a_D}{\longrightarrow}{\cal O}_L(-X-Z)|_Z \stackrel{a_X}{\longrightarrow}{\cal O}_L(-Z)|_Z\stackrel{a_D}{\longrightarrow}{\cal O}_L(-X)|_Z \stackrel{a_X}{\longrightarrow}\cI\longrightarrow 0 \ . $$ Indeed the image of $a_X:{\cal O}_L(-X)|_Z\to{{\cal O}_L}|_Z={\cal O}_Z$ is $\cI$, and locally $X$ and $D$ have equations $t_X$ and $t_D$ in ${\cal O}_L$, and the sequence is the sequence of ${\cal O}_L/(t_Xt_D)$-modules which is alternatively multiplication by $t_X$ and $t_D$. From the fact that $t_X$ and $t_D$ are nonzerodivisors, the exactness of the sequence follows. So to compute $\Ext^1_Z(\cI,{\cal O}_Z/\cI\otimes_k M)$ we apply $\Hom_Z(\cdot,{\cal O}_Z/\cI\otimes_k M)$ and take the first cohomology group of the resulting complex. In fact $\Hom_Z(\cI,{\cal O}_Z/\cI\otimes_k M)=0$~; this is the tangent space to the functor $F$. Therefore $\Ext^1_Z(\cI,{\cal O}_Z/\cI\otimes_k M)$ is equal to $$ \ker\big(a_D:\Hom_Z({\cal O}_L(-X)|_Z,{\cal O}_Z/\cI\otimes_k M)\longrightarrow\Hom_Z({\cal O}_L(-Z)|_Z,{\cal O}_Z/\cI\otimes_k M)\big) \ . $$ Locally ${\cal O}_L(-X)|_Z\simeq {\cal O}_L(-Z)|_Z\simeq {\cal O}_Z$, and $a_D$ takes a map $\sigma:{\cal O}_Z\to{\cal O}_Z/(t_X)\otimes_k M$ to $t_D\sigma$. Since $D$ is a Cartier divisor in $X$, it follows that $a_D$ is injective. Consequently $$\Ext^1_Z(\cI,{\cal O}_Z/\cI\otimes_k M)=0 \ , $$ thus the functor $F$ is unobstructed, so $S_0$ is formally smooth at~$s$. \end{proo} The second lemma proves a statement which is used in \cite{SS} (proof of 3.17, point (2)). However we were not able to understand their proof, due to a vicious circle in the use of an argument from \cite{St}: \begin{lemm} \label{lemm_division} Let $k$ be a field and assume $S = \Spec(k)$. Let $f,g \in k[\fg]^G$ and assume $f_{|\ft} \ | \ g_{|\ft}$. Then $f \ | \ g$. \end{lemm} \begin{proo} We may assume that $k$ is algebraically closed. First assume that $f$ has no square factors. Let $x \in \fg$ be such that $f(x)=0$; it is then enough to show that $g(x)=0$. To this end, since $x$ belongs to a Borel subalgebra of $\fg$ (in fact, the Borel subalgebra are the maximal solvable algebras) and all the Borel subalgebras are conjugated, we may assume that $x = \tau + \eta$ with $\tau \in \ft$ and $\eta =\sum_{\alpha > 0} x_\alpha \in \bigoplus_{\alpha > 0} \fg_\alpha$. Let $X:{\mathbb G}_m \to T$ be a one-parameter-subgroup corresponding to a coweight $\omega^\vee$ with $\scal{\omega^\vee,\alpha} > 0$ for all positive roots $\alpha$. We therefore have $\Ad(X(t)).x = \tau + \sum_{\alpha > 0} t^{n_\alpha} x_\alpha$, with $\forall \alpha>0,n_\alpha > 0$. Therefore the closure of the $G$-orbit through $x$ contains $\tau$, and $f(x)=f(\tau)=0$. Thus $g(x)=g(\tau)=0$. Thus the lemma is proved in case $f$ is squarefree. Now let $f$ be arbitrary. Write $f=f_1f_2$ where $f_1$ is the product of the prime factors of $f$, with multiplicity $1$. So $f_1$ is squarefree, and since $G$ is connected and hence has no nontrivial characters, we see that $f_1$ is $G$-invariant. We have $(f_1)_{|\ft} \ | \ g_{|\ft}$ so $f_1 \ | \ g$~: let us write $g=f_1g_2$. By factoriality of $k[\fg]$ we have $(f_2)_{|\ft} \ | \ (g_2)_{|\ft}$, so the lemma follows by induction on the degree of $f$. \end{proo} We can now prove our main result. \begin{theo} \label{theo_chevalley_iso} Let $S$ be a scheme and let $G$ be a split simple Chevalley group over~$S$. Assume that $G$ is not isomorphic to $Sp_{2n}$, $n\ge 1$. Then the Chevalley morphism $\pi:\ft/W\to\fg/G$ is an isomorphism. \end{theo} \begin{proo} By theorem~\ref{theo_chevalley_dominant} it is enough to prove that the map on functions is surjective. Let $f$ be a $W$-invariant function on $\ft$ and let $f_1$ be the function on $G/T\times\ft$ defined by $f_1(g,x)=f(x)$. Since it is $W$-invariant, it induces a function on $(G/T\times\ft)/W$ which we denote by the letter $f_1$ again. By lemma~\ref{lemm_birational} the function $h:=f_1\circ b^{-1}$ is a $G$-invariant relative meromorphic function whose domain of definition contains $\Reg(\fg)$. We may write $h=k/\delta^m$ for some function $k$ not divisible by $\delta$, and some integer $m$. Since $k$ is $G$-invariant on a schematically dense open subset, it is $G$-invariant. Assume that $m\ge 1$. Let $s\in S$ be a point. Since a generic element of $\ft_s$ is regular, $h_s$ is defined as a rational function on $\ft_s$. By definition of $b$, we moreover have $(h_s)_{|\ft} = f_s$. It follows that we have $$ (k_s)_{|\ft} = f_s . (\delta_s)_{|\ft}^m. $$ Therefore the restriction of $\delta_s$ divides the restriction of $k_s$. Lemma~\ref{lemm_division} implies that $\delta_s$ divides $k_s$. Since this is true for all $s\in S$, and $p:\Sing(\fg)\to S$ is a relative Cartier divisor of $\fg/S$ with $p_*{\cal O}_D$ free over $S$ by lemma~\ref{lemm_sing_cartier}, then we can apply lemma~\ref{indicator_of_zeroes} to conclude that $\delta$ divides $k$. This is a contradiction with our assumptions, therefore $h$ is a regular function extending $f$ to a $G$-invariant function on $\fg$. \end{proo} In the remaining sections, we compute explicitly the ring of invariants in the case where $G$ is one of the groups $SO_{2n},SO_{2n+1}$ or $Sp_{2n}$. \section{The orthogonal group $SO_{2n}$} \label{case_S0_2n} In the case of the group $G=SO_{2n}$, the explicit computation will prove that the formation of the adjoint quotient for the Lie algebra does not commute with all base changes. In fact, we will be able to describe exactly when commutation holds. We will see also that over a base field, the quotient is always an affine space. \subsection{Definition of $SO_{2n}$} \label{subsection_def_SO_2n} \begin{noth} {\bf The orthogonal group.} The free $\bZ$-module of rank $2n$ is denoted by $E$; we think of it as the trivial vector bundle over $\Spec(\bZ)$. The standard quadratic form of $E$ is defined for $v=(x_1,y_1,\dots,x_n,y_n)$ by $$ q(v)=x_1y_1+\dots+x_ny_n \ . $$ It is nondegenerate in the sense that $\{q=0\}\subset \bP(E)$ is smooth over $\bZ$. The polarization of $q$ is $$ \langle v,v'\rangle=q(v+v')-q(v)-q(v')=x_1y'_1+x'_1y_1+\dots+x_ny'_n+x'_ny_n \ . $$ The orthogonal group $O_{2n}$ is the set of transformations $P\in GL_{2n}$ that preserve $q$, more precisely, the zero locus of the morphism $\Psi$ from $GL_{2n}$ to the vector space of quadratic forms defined by $\Psi(P)=q\circ P-q$. Thus the Lie algebra $\mathfrak{o}} \def\ft{\mathfrak{t}} \def\fu{\mathfrak{u}_{2n}$ is the subscheme of $\mathfrak{gl}} \def\fsl{\mathfrak{sl}_{2n}$ composed of matrices $M$ such that by $d\Psi_{\Id}(M)=0$ with $$ d\Psi_{\Id}(M)(v)=\langle v,Mv\rangle \ . $$ It is not hard to verify that $\mathfrak{o}} \def\ft{\mathfrak{t}} \def\fu{\mathfrak{u}_{2n}\subset \mathfrak{gl}} \def\fsl{\mathfrak{sl}_{2n}$ is a direct summand of the expected dimension, so that $O_{2n}$ is a smooth group scheme over $\bZ$. \begin{rema} \label{orth_matrix_preserves_polarization} Let us denote by $B$ the matrix of the polarization of $q$. Clearly, an orthogonal matrix $P$ preserves the polarization, and it follows that ${}^tPBP=B$. However, one checks easily that the subgroup $X\subset GL_{2n}$ defined by the equations ${}^tP BP=B$ is not flat over $\bZ$ because its function ring has $2$-torsion. In fact $O_{2n}$ is the biggest subscheme of $X$ which is flat over $\bZ$. Accordingly $\Lie(X)\subset\mathfrak{gl}} \def\fsl{\mathfrak{sl}_{2n}$ is defined by ${}^tMB+BM=0$, and $\mathfrak{o}} \def\ft{\mathfrak{t}} \def\fu{\mathfrak{u}_{2n}$ is the biggest $\bZ$-flat subscheme of $\Lie(X)$. \end{rema} \end{noth} \begin{noth} {\bf Dickson's invariant.} Over any field $k$, it is well-known that $O_{2n}\otimes k$ has two connected components. In odd characteristic, the determinant takes the value $1$ on one and $-1$ on the other. In characteristic $2$ the determinant does not help to separate the connected components. Instead one usually uses Dickson's invariant $D(P)$ defined, for an orthogonal matrix $P$, to be $0$ if and only if $P$ acts trivially on the even part of the center of the Clifford algebra. Equivalently, $D(P)=0$ if and only if $P$ is a product of an even number of reflections (there is just one exception; see \cite{Ta}, p. 160). Here is a more modern, base-ring-free way to consider the determinant and Dickson's invariant altogether: \begin{lemm} There is a unique element $\delta\in \bZ[O_{2n}]$ such that $\det=1+2\delta$. \end{lemm} \begin{proo} Since for any $P\in O_{2n}$, we have $\det(P) \in \{-1,1\}$, the function $\det-1$ vanishes on the fibre $O_{2n}\otimes {\mathbb F}} \def\bG{{\mathbb G}} \def\bH{{\mathbb H}_2$. Since $O_{2n}\otimes {\mathbb F}} \def\bG{{\mathbb G}} \def\bH{{\mathbb H}_2$ is reduced, $2$ divides $\det-1$, yielding the existence of $\delta$. It is unique because $O_{2n}$ is flat over $\bZ$, and in particular has no $2$-torsion. \end{proo} Let us introduce the $\bZ$-group scheme $\cG=\Spec(\bZ[u,\frac{1}{1+2u}])$ with unit $u=0$ and multiplication $u*v=u+v+2uv$. Its fibre at $2$ is isomorphic to the additive group while all other fibres are isomorphic to the multiplicative group. When one passes from $\det$ to $\delta$, the multiplicativity formula $\det(P_1P_2)=\det(P_1)\det(P_2)$ gives $\delta(P_1P_2)=\delta(P_1)+\delta(P_2)+2\delta(P_1)\delta(P_2)$. In other words, \begin{lemm} $\delta$ defines a morphism of groups $O_{2n}\to \cG$. \hfill $\square$ \end{lemm} The schematic image of $\delta$ is the subgroup of $\cG$ given by $u(u+1)=0$, isomorphic to the constant $\bZ$-group scheme $\zmod{2}$. \begin{defi} We define $SO_{2n}$ as the kernel of $\delta$. \end{defi} The group $SO_{2n}$ is smooth over $\bZ$ with connected fibres. The subgroup $T$ of diagonal matrices in $SO_{2n}$ is a maximal torus, we denote by $\ft$ its Lie algebra and by $\lambda_i$ its coordinate functions. Its normalizer $N$ is the subgroup of orthogonal monomial matrices. The Weyl group $W=N/T$ is the semi-direct product $(\zmod{2})^{n-1}\rtimes\mathfrak{S}_n$ where $\mathfrak{S}_n$ is the symmetric group on $n$ letters. It acts on $T$ as follows. The subgroup $(\zmod{2})^{n-1}$ is generated by the transformations $\varepsilon_{i,j}$ which take $\lambda_i$ and $\lambda_j$ to their opposite and leave all other $\lambda_k$ unchanged. The subgroup $\mathfrak{S}_n$ permutes the $\lambda_i$. The action of $W$ on $\ft$ has analogous expressions that are immediate to write down. \end{noth} \begin{noth} {\bf The Pfaffian.} Recall that there is a unique function on $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n}$, called the {\em pfaffian} and denoted $\pf$, such that $\det(M)=(-1)^n(\pf(M))^2$. (The sign $(-1)^n$ comes from the fact that in our context, the pfaffian is $\pf'(BM)$ where $\pf'$ is the usual pfaffian.) Furthermore the pfaffian is invariant for the adjoint action of $SO_{2n}$. \end{noth} \subsection{Invariants of the Weyl group} \label{invariant_weyl} \indent We denote by $\ft$ the $n$-dimensional affine space with coordinate functions $X_i$, and by $W$ the group generated by the permutations of the coordinates and the reflections $\varepsilon_{i,j}$ which map $X_i$ and $X_j$ to their opposite and leave the other coordinates invariant. We denote by $\sigma_k$ the complete elementary symmetric functions in $n$ variables. \begin{prop} \label{invariant_weyl_so2n} Let $A$ be a ring, then $A[\ft]^W$ is generated by $X_1\cdots X_n$, $\sigma_k(X_i^2)$, and $x\sigma_k(X_i)$, where $k<n$ and $x$ runs through the 2-torsion ideal of $A$. \end{prop} \begin{proo} Let $F$ be a function in $X_1,\dots,X_n$ which is invariant under the Weyl group. Let us say that a monomial is {\em good} if the exponents of its variables $X_i$ all have the same parity (this is either a monomial in the $X_i^2$, or $X_1\cdots X_n$ times a monomial in the $X_i^2$). We say that it is {\em bad} otherwise. We can write uniquely $F$ as the sum of its good part and its bad part: $$ F(X_1,\dots,X_n) =F_1(X_1^2,\dots,X_n^2,X_1\dots X_n)+F_2(X_1,\dots,X_n) \ . $$ The group $W$ respects this decomposition, hence $F$ being $W$-invariant, its good and bad parts also are. In particular they are $\mathfrak{S}_n$-invariant, so that $$ F(X_1,\dots,X_n) =G_1(\sigma_1(X_i^2),\dots,\sigma_{n-1}(X_i^2),X_1\dots X_n) +G_2(\sigma_1(X_i),\dots,\sigma_n(X_i)) \ . $$ Letting the $\varepsilon_{i,j}$ act, we see that all coefficients of $G_2$ must be $2$-torsion. The proposition is therefore proved. \end{proo} \subsection{Computation of the Chevalley morphism} \label{subsection_invariants_SO_even} \indent In this subsection we will describe explicitly the invariants of $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n}$ under $SO_{2n}$ that correspond to the Weyl group invariants under theorem~\ref{theo_chevalley_iso}, see theorem~\ref{theo_so2n} below. The Lie algebra $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n}$ has a universal matrix ${\sf M}$ whose most important attributes are its characteristic polynomial $\chi$ and its pfaffian $\pf=\pf({\sf M})$. In fact ${\sf M}$ and $\chi$ are the restrictions of the universal matrix of $\mathfrak{gl}} \def\fsl{\mathfrak{sl}_{2n}$ and its characteristic polynomial. From the equality ${}^t{\sf M} B+B{\sf M}=0$ (see \ref{orth_matrix_preserves_polarization}) it follows that $\chi$ is an even polynomial, that is to say $$ \chi(t)=\det(t\Id-{\sf M})=t^{2n}+c_2t^{2n-2}+\dots+c_{2n}. $$ The functions $c_{2i}$ are invariants of the adjoint action; note that $$c_{2n}=\det({\sf M})=(-1)^n(\pf({\sf M}))^2 \ .$$ There are some more invariants coming from characteristic $2$. Indeed, in this case the polar form is alternating, so homotheties are antisymmetric and we can define the {\em pfaffian characteristic polynomial} by $$\pi_{{\mathbb F}} \def\bG{{\mathbb G}} \def\bH{{\mathbb H}_2}(t)=\pf(t\Id-{\sf M}_{{\mathbb F}} \def\bG{{\mathbb G}} \def\bH{{\mathbb H}_2}) \ .$$ We have $\chi_{{\mathbb F}} \def\bG{{\mathbb G}} \def\bH{{\mathbb H}_2}(t)=(\pi_{{\mathbb F}} \def\bG{{\mathbb G}} \def\bH{{\mathbb H}_2}(t))^2$. Now let us consider one particular lift of $\pi_{{\mathbb F}} \def\bG{{\mathbb G}} \def\bH{{\mathbb H}_2}$ to $\bZ$: \begin{defi} \label{defi_coefs_pfaff_polynomial} Let $\sigma\colon{\mathbb F}} \def\bG{{\mathbb G}} \def\bH{{\mathbb H}_2\to \bZ$ be such that $\sigma(0)=0$ and $\sigma(1)=1$. The polynomial $\pi\in\bZ[\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n}][t]$ is defined as $\pi(t)=t^n+\pi_1t^{n-1}+\dots+\pi_{n-1}t+\pi_n$ where $\pi_n:=\pf({\sf M})$ and the other coefficients $\pi_i$ ($1\le i\le n-1$) are the lifts of the corresponding coefficients of $\pi_{{\mathbb F}} \def\bG{{\mathbb G}} \def\bH{{\mathbb H}_2}$ via $\sigma$. \end{defi} We note that for any ring $A$, the (images of the) elements $\pi_1,\dots,\pi_{n-1},\pi_n$ are algebraically independent over $A$, because they restrict on a maximal torus to the functions $\sigma_i(X_j)$, the symmetric functions in the coordinates, which are themselves algebraically independent over $A$. We defined the functions $\pi_i$ by arbitrary lifting, but we can make them somehow universal: \begin{prop} \label{universal_coefs_of_pfaffian} Let ${\cal O}$ be the ring $\bZ[X]/(2X)$ and denote by $\tau$ the image of $X$ in~${\cal O}$. Then $({\cal O},\tau)$ is universal among rings with a $2$-torsion element. Moreover, any monomial function $\tau(\pi_1)^{\alpha_1}\dots(\pi_{n-1})^{\alpha_{n-1}}$ on $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n,{\cal O}}$ is independent of the choice of the lifts $\pi_i$ and invariant under the adjoint action of $SO_{2n,{\cal O}}$. Finally, for each $i\in\{1,\dots,n\}$ we have $\tau(\pi_i)^2=\tau c_{2i}$. \end{prop} \begin{proo} The universality statement about $({\cal O},\tau)$ means that for any pair $(A,x)$ where $A$ is a ring and $x$ is a $2$-torsion element of $A$, there is a unique morphism $f:{\cal O}\to A$ such that $f(\tau)=x$. This is obvious. Since $2\tau=0$, it is clear also that $\tau(\pi_1)^{\alpha_1}\dots(\pi_{n-1})^{\alpha_{n-1}}$ is independent of the choice of $\pi_i$. The fact that this monomial is invariant comes from the invariance of the pfaffian characteristic polynomial in characteristic $2$. Finally the equalities $\tau(\pi_i)^2=\tau c_{2i}$ come from the equalities $(\pi_i)^2=c_{2i}$ in characteristic $2$. \end{proo} By proposition \ref{universal_coefs_of_pfaffian}, for any ring $A$ and any $x\in A[2]$, the quantity $x(\pi_1)^{\alpha_1}\dots(\pi_{n-1})^{\alpha_{n-1}}$ is a well-defined invariant function on $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n,A}$. \begin{theo} \label{theo_so2n} Let $A$ be a ring, $G=SO_{2n,A}$, $\fg = \mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n,A}$. The ring of invariants $A[\fg]^G$ is $$ A[c_2,c_4,\dots,c_{2n-2},\pf\,;\,x(\pi_1)^{\epsilon_1} \dots(\pi_{n-1})^{\epsilon_{n-1}}] $$ where $x$ runs through a set of generators of the $2$-torsion ideal $A[2]\subset A$, and $\epsilon_i=0$ or $1$, not all $0$. \end{theo} \begin{proo} By theorem~\ref{theo_chevalley_iso} and proposition \ref{invariant_weyl_so2n}, we have $A[\fg]^G=A[\sigma_k(X_i^2)\,;\,X_1\dots X_n\,;\,x\sigma_k(X_i)]$ where as before the $X_i$ are the coordinate functions on the torus. We now use the previous proposition. Since $c_{2k}$ restricts on the torus to $\pm \sigma_k(X_i^2)$, the pfaffian restricts to $X_1\dots X_n$, $x\pi_k$ restricts to $x\sigma_k(X_i)$, and since $x\pi_i^2 = xc_{2i}$, the theorem is proved. \end{proo} The behaviour of the ring of invariants is therefore controlled by the $2$-torsion. More precisely, for a scheme $S$ let $S[2]$ be the closed subscheme defined by the ideal of $2$-torsion. If $f:S'\to S$ is a morphism of schemes, we always have $S'[2]\subset f^*S[2]$. We have: \begin{coro} \label{coro_so2n} \begin{trivlist} \itemn{1} The formation of the quotient in the previous theorem commutes with a base change $f:S'\to S$ if and only if $f^*S[2]=S'[2]$. This holds in particular if $2$ is invertible in ${\cal O}_S$, or if $2=0$ in ${\cal O}_S$, or if $S'\to S$ is flat. \itemn{2} Assume that $S$ is noetherian and connected. Then the quotient is of finite type over~$S$, and is flat over $S$ if and only if $S[2]=S$ or $S[2]=\emptyset$. \end{trivlist} \end{coro} \begin{proo} First we recall some general facts on the formation of the ring of invariants for the action of an affine $S$-group scheme $G$ acting on an affine $S$-scheme $X$. The formation of $X/G=\Spec(({\cal O}_X)^G)$ commutes with flat base change, and in particular with open immersions. It follows that if $(S_i)$ is an open covering of $S$ then with obvious notation $X_i/G_i$ is an open set in $X/G$, and $X/G$ can be obtained by glueing the schemes $X_i/G_i$. Therefore if $(S'_{i,j})$ is an open covering of $S_i\times_S S'$ for all $i$, the formation of the quotient commutes with the base change $S'\to S$ if and only if for all $i,j$ the formation of the quotient $X_i/G_i$ commutes with the base change $S'_{i,j}\to S_i$. This reduces the proof to the case of a base change of affine schemes $S'=\Spec(A')\to S=\Spec(A)$. Call $B$ (resp. $B'$) the ring of invariants over $A$ (resp. $A'$). Observe that $B$ inherits a graduation from the graduation of the function algebra of $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n,A}$, and its only homogeneous elements of degree~$1$ are those of the form $x\pi_1$ with $x\in A[2]$. We proceed to prove (1) and (2). \noindent (1) The base change morphism $B\otimes_A A'\to B'$ is $A'[\underline c,\pf,x\underline\pi^\epsilon]\to A'[\underline c,\pf,x'\underline\pi^\epsilon]$ where $$\underline c=(c_2,c_4,\dots,c_{2n-2}) \ , \ x\underline\pi^\epsilon=x(\pi_1)^{\epsilon_1} \dots(\pi_{n-1})^{\epsilon_{n-1}} \mbox{ with } x\in A[2] $$ and $x'\underline\pi^\epsilon$ is the same quantity with $x'\in A'[2]$. This map is clearly injective. If it is surjective, then in particular for any $x'\in A'[2]$ we have $x'\pi_1\in A'[\underline c,\pf,x\underline\pi^\epsilon]$. Thus there is $a'\in A'$ and $x\in A[2]$ such that $x'\pi_1=a'x\pi_1$. Since $\pi_1$ is a nonzerodivisor we get $x'=a'x$, so $A'[2]$ is the image of $A[2]$. This is exactly the assertion that $f^*S[2]=S'[2]$. The converse is easy, as well as the particular cases stated in the lemma. \noindent (2) If $A$ is noetherian, $A[2]$ is finitely generated and then $B$ is of finite type over $A$. Now let $I:=A[2]$. If $I=0$ then $B$ is a polynomial ring, and this is also the case if $I=A$ because then $c_2,c_4,\dots,c_{2n-2}$ are polynomials in $\pf({\sf M}),\pi_1,\dots,\pi_{n-1}$. It remains to prove that if $B$ is flat over $A$ then $I=0$ or $I=A$. In this case the $2$-torsion ideal of $B$ is $IB$, as we see from tensoring by~$B$ the exact sequence $$ 0\to A/I\stackrel{\times 2}{\to} A\to A/2\to 0 \ . $$ So for any $y\in I$, we have $y\pi_1\in B[2]=IB$ hence we may write $y\pi_1=i_1b_1+\dots+i_rb_r$ with $i_k\in I$ and $b_k\in B$. Let $x_k\pi_1$ be the degree~$1$ component of $b_k$, then by taking the components of degree~$1$ and using the fact that $\pi_1$ is a nonzerodivisor, we find $y=i_1x_1+\dots+i_rx_r\in I^2$. Thus $I=I^2$, and if the spectrum of $A$ is connected, this implies $I=0$ or $I=A$. \end{proo} \section{The orthogonal group $SO_{2n+1}$} \label{case_S0_2n+1} For $G=SO_{2n+1}$, the computation of the quotient is a little more involved since using the natural representation of dimension $2n+1$ brings some trouble, as we explain below. We show which point of view on $SO_{2n+1}$ will lead to the definition of the correct invariants. Then, the results are essentially the same as for $G=SO_{2n}$. \subsection{Definition of $SO_{2n+1}$} \indent In this section, $E$ is the free $\bZ$-module $\bZ^{2n+1}$. Its standard quadratic form $q$ is $$ q(v)=x_1y_1+\dots+x_ny_n+z^2 $$ where $v=(x_1,y_1,\dots,x_n,y_n,z)$. It is nondegenerate, and its polarization is $$ \langle v,v'\rangle=q(v+v')-q(v)-q(v')=x_1y'_1+x'_1y_1+\dots+x_ny'_n+x'_ny_n+2zz' \ . $$ In contrast with the even dimensional case, in characteristic $2$ the polarization has a nonzero radical which is the line generated by the last basis vector of $E\otimes{\mathbb F}} \def\bG{{\mathbb G}} \def\bH{{\mathbb H}_2$. Now, let $\tilde E=\bZ^{2n+2}$ with canonical basis $(e_1,e'_1,\dots,e_{n+1},e'_{n+1})$ and standard quadratic form defined (as in \ref{subsection_def_SO_2n}) by $q(v)=x_1y_1+\dots+x_{n+1}y_{n+1}$ where $v=(x_1,y_1,\dots,x_{n+1},y_{n+1})$. We consider the isometric embedding $i\colon E\hookrightarrow \tilde E$ given by $$ i(x_1,y_1,\dots,x_n,y_n,z)=(x_1,y_1,\dots,x_n,y_n,z,z) \ . $$ Since $i$ is an isometry, it is harmless to use the same letter for $q$ and for $q_{|E}$. The orthogonal subspace of $E$ in $\tilde E$ is the free rank $1$ submodule generated by the vector $\varepsilon=e_{n+1}-e'_{n+1}$. Note that the group of transformations of $(\tilde E,q)$ is the group $O_{2n+2}$ as defined in \ref{subsection_def_SO_2n}. Then we define $SO_{2n+1}$ as a closed subgroup of $SO_{2n+2}$ by $$ SO_{2n+1}=\{\,P\in SO_{2n+2},\, P(\varepsilon)=\varepsilon \,\} \ . $$ Accordingly, its Lie algebra is $$ \mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n+1}=\{\,M\in \mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n+2},\, M(\varepsilon)=0 \,\} \ . $$ It is a simple exercise to verify that $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n+1}\subset \mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n+2}$ is a direct summand of the expected dimension, so that $SO_{2n+1}$ is a smooth group scheme over $\bZ$. \begin{rema} \label{another_presentation} Let $O(q|_E)$ be the group of linear transformations of $E$ that preserve~$q$, and $SO(q|_E)$ the kernel of the Dickson invariant. Since a matrix $P\in SO_{2n+1}$ preserves the line generated by $\varepsilon$, it preserves its orthogonal~$E$. This leads to a morphism $SO_{2n+1}\to SO(q|E)$. However, because of the existence of a one-dimensional radical in characteristic~$2$, one can see that the fibre $SO(q|E)\otimes{\mathbb F}} \def\bG{{\mathbb G}} \def\bH{{\mathbb H}_2$ is nonreduced and its reduced subscheme is the subgroup $H$ of transformations that act as the identity on the radical. Thus $SO_{2n+1}\to SO(q|E)$ is not an isomorphism. In fact, one may see that this map realizes $SO_{2n+1}$ as the dilatation of $SO(q|E)$ with center~$H$. Recall from \cite{BLR}, 3.2 that the dilatation is a map $\pi\colon SO_{2n+1}\to SO(q|E)$ which is universal for the properties: $SO_{2n+1}$ is $\bZ$-flat and its special fibre at $2$ is mapped into $H$. It can be checked that the dilatation is indeed smooth over $\bZ$ and is the Chevalley orthogonal group. In this formulation, the special orthogonal group is not naturally a group of matrices. This is why we used another presentation. \hfill $\square$ \end{rema} The subgroup $T\subset SO_{2n+2}$ of diagonal matrices fixing $\varepsilon$ is a maximal torus of $SO_{2n+1}$, we denote by $\ft$ its Lie algebra and by $\lambda_i$ its coordinate functions. Its normalizer $N$ is the subgroup of orthogonal monomial matrices fixing $\varepsilon$. The Weyl group $W=N/T$ is the semi-direct product $(\zmod{2})^n\rtimes\mathfrak{S}_n$. It acts on $T$ as follows. The subgroup $(\zmod{2})^{n-1}$ is generated by the transformations $\varepsilon_i$ which take $\lambda_i$ to its opposite and leave all other $\lambda_k$ unchanged. The subgroup $\mathfrak{S}_n$ permutes the $\lambda_i$. \subsection{Explicit computation of the Chevalley morphism} \label{subsection_invariants_SO_odd} \indent Let ${\sf M}$ be the universal matrix over $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n+1}$. Using the embedding of $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n+1}$ into $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n+2}$, we define invariants by restriction from those of $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n+2}$ defined in \ref{subsection_invariants_SO_even}. For example, let us view the universal matrix ${\sf M}$ as a matrix in $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n+2}$. Since ${\sf M}(\varepsilon)=0$, the determinant of~${\sf M}$ vanishes and hence its characteristic polynomial in dimension $2n+2$ is $$ t^{2n+2}+c_2t^{2n}+\dots+c_{2n}t^2 \ . $$ We {\em define} the {\em characteristic polynomial of ${\sf M}$} as $$ \chi(t)=t^{2n+1}+c_2t^{2n-1}+\dots+c_{2n}t \ . $$ Note that this is not the characteristic polynomial associated to an actual action on the natural representation of dimension $2n+1$. Using again the embedding in $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n+2}$, we see that in characteristic $2$ we have again a polynomial $\pi(t)$ defined uniquely by the identity $\chi_{{\mathbb F}} \def\bG{{\mathbb G}} \def\bH{{\mathbb H}_2}(t)=t(\pi_{{\mathbb F}} \def\bG{{\mathbb G}} \def\bH{{\mathbb H}_2}(t))^2$. By abuse, we call it again {\em pfaffian characteristic polynomial}. We may define lifts of its coefficients by the same process as in definition \ref{defi_coefs_pfaff_polynomial} and we obtain a polynomial $\pi(t)=t^n+\pi_1t^{n-1}+\dots+\pi_{n-1}t+\pi_n$ where $\pi_i\in\bZ[\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n}]$. As in subsection~\ref{subsection_invariants_SO_even}, for any ring $A$ the elements $\pi_1,\dots,\pi_{n-1},\pi_n$ are algebraically independent over $A$. In the same way as in subsection \ref{subsection_invariants_SO_even}, we prove: \begin{prop} Let $({\cal O},\tau)$ be the ring defined in proposition \ref{universal_coefs_of_pfaffian}. Then any monomial function $\tau(\pi_1)^{\alpha_1}\dots(\pi_n)^{\alpha_n}$ on $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n+1,{\cal O}}$ is independent of the choice of the lifts~$\pi_i$ and invariant under the adjoint action of $SO_{2n+1,{\cal O}}$. Also, for each $i\in\{1,\dots,n\}$ we have $\tau(\pi_i)^2=\tau c_{2i}$. \hfill $\square$ \end{prop} So for any ring $A$ and any $x\in A[2]$, the quantity $x(\pi_1)^{\alpha_1}\dots(\pi_n)^{\alpha_n}$ is a well-defined invariant function on $\mathfrak{so}} \def\fpso{\mathfrak{pso}_{2n+1,A}$. Exactly the same proof as the proof of \ref{theo_so2n} gives: \begin{theo} \label{theo_so2n+1} Let $A$ be a ring and $G=SO_{2n+1,A}$. Then the ring of functions of $\fg/G$ is $$ A[c_2,c_4,\dots,c_{2n}\,;\,x(\pi_1)^{\epsilon_1}\dots(\pi_n)^{\epsilon_n}] $$ where $x$ runs through a set of generators of the $2$-torsion ideal $A[2]\subset A$, and $\epsilon_i=0$ or $1$, not all $0$. \hfill $\square$ \end{theo} Finally, all the statements of corollary \ref{coro_so2n} hold also word for word for $G=SO_{2n+1}$. \section{The symplectic group $Sp_{2n}$} \label{case_Sp_2n} \indent The computation of the adjoint quotient and of the Chevalley morphism $\pi:\ft/W\to\fg/G$ for $Sp_{2n}$ requires the preliminary computation of the corresponding quantity for the group $SL_2$. We also found it interesting to deal with the case of $PSL_2$. \subsection{Preliminary cases: $SL_2$ and $PSL_2$} We denote by $\matddr abc{-a}$ the universal matrix of $\fsl_2$; therefore $A[a]$ is the ring of functions on $\ft$ over $A$. The same proof as that of proposition \ref{invariant_weyl_so2n} yields: \begin{fait} \label{fait_weyl_sl2} Let $A$ be a ring, then $A[\ft]^W$ is equal to $A[a^2] \oplus aA[2][a^2]$. \end{fait} \noindent We set $\det(a,b,c) = -a^2 - bc$. This fact and the next proposition show that we don't have $\ft/W \simeq \fg/G$: \begin{prop} \label{prop_sl2} Let $A$ be a ring, then $A[\fsl_2]^{SL_2} = A[\det]$. \end{prop} \begin{proo} The action of the diagonal matrix $\matddr u00{u^{-1}}$ on the coordinate functions reads: \begin{equation} \label{action_diagonale} \begin{array}{l} a \mapsto a\\ b \mapsto u^2b\\ c \mapsto u^{-2}c \end{array} \end{equation} Any invariant polynomial can therefore be written as a polynomial in $a$ and $bc$. On the other hand the action of the unipotent element $\matddr 1t01$ reads: \begin{equation} \label{action_unipotent} \begin{array}{l} a \mapsto a+tc\\ b \mapsto b-2ta-t^2c\\ c \mapsto c \end{array} \end{equation} Assume we have a homogeneous invariant of odd degree $2d+1$. Since it is a polynomial in $a$ and $bc$, it can be written as $af(a^2,bc)$, with $f$ homogeneous of degree $d$. We consider the identity $$ af(a^2,bc) = (a+tc)f((a+tc)^2,(b-2ta-t^2c)c), $$ and specialise to $a=0$. We get $tcf(t^2c^2,bc-t^2c^2)=0$ so $f(t^2c^2,bc-t^2c^2)=0$. Performing the invertible change of coordinates $d=b+t^2c$, we therefore get $f(t^2c^2,cd)=0=c^df(t^2c,d)$, from which it follows that $f=0$. \medskip Thus there are no invariants of odd degree and the image of the restriction morphism is included in $A[a^2]$. Since $\det$ is an invariant, this image is exactly $A[a^2]$, which implies the proposition. \end{proo} We pass to $PSL_2$. By proposition \ref{prop_psln} and its proof, the coordinate ring of $\mathfrak{psl}} \def\fspin{\mathfrak{spin}_2$ over $A$ is $A[\alpha , b , c]$, where $\alpha = 2a$. \begin{fait} \label{fait_weyl_psl2} Let $A$ be a ring, then $A[\ft]^W$ is equal to $A[\alpha^2] \oplus \alpha A[2][\alpha^2]$. \end{fait} \begin{prop} \label{prop_psl2} Let $A$ be a ring, then $A[\mathfrak{psl}} \def\fspin{\mathfrak{spin}_2]^{PSL_2} = A[4\det] + \alpha . A[2][4\det]$. \end{prop} \begin{proo} We know from theorem \ref{theo_chevalley_dominant} that $A[\mathfrak{psl}} \def\fspin{\mathfrak{spin}_2]^{PSL_2}$ injects into $A[\ft]^W = A[\alpha^2] \oplus \alpha A[2][\alpha^2]$. On the other hand, $4\det = -\alpha^2 - 4bc$ is certainly an invariant in the coordinate ring, as well as $x\alpha$, if $x \in A$ is a 2-torsion element, since by (\ref{action_unipotent}) $\alpha$ is mapped to $\alpha + 2tc$ under the action of the unipotent element $\matddr 1t01$. Thus the proposition is proved. \end{proo} \subsection{Explicit computation of the Chevalley morphism} \indent We denote by $\ft$ the $n$-dimensional affine space with coordinate functions $X_i$, and $W$ the group generated by the permutations of the coordinates and the reflections $\varepsilon_i$ which map $X_i$ to its opposite and leave the other coordinates invariant. Recall that $\sigma_k$ denotes the complete elementary symmetric functions in $n$ variables. The same proof as for proposition \ref{invariant_weyl_so2n} yields: \begin{prop} \label{invariant_weyl_sp2n} Let $A$ be a ring, then $A[\ft]^W$ is generated by $\sigma_k(X_i^2)$ and $x\sigma_k(X_i)$, where $k<n$ and $x$ runs through the 2-torsion ideal of $A$. \hfill $\square$ \end{prop} We denote by $E$ the natural representation of $G = Sp_{2n}$, of dimension $2n$. By definition, we therefore have a morphism $G \rightarrow GL(E)$, which also induces a morphism $\fg \rightarrow \mathfrak{gl}} \def\fsl{\mathfrak{sl}(E)$. Let ${\sf M}$ be the universal matrix over $\mathfrak{gl}} \def\fsl{\mathfrak{sl}(E)$, and let $\chi$ be its characteristic polynomial: $$ \chi(t)=\det(t\Id-{\sf M})=t^{2n}-c_1t^{2n-1}+c_2t^{n-2}+\dots+c_{2n} \ . $$ \begin{theo} \label{theo_sp2n} Let $A$ be a ring and $G=Sp_{2n,A}$. Then the morphism $\pi:\ft/W\to\fg/G$ is an isomorphism \iff $A$ has no 2-torsion. Moreover, the ring of functions of $\fg/G$ is $$ A[c_2,c_4,\dots,c_{2n}] \ . $$ The formation of the adjoint quotient $\fg\to\fg/G$ over a scheme $S$ commutes with any base change $S'\to S$. \end{theo} \begin{proo} Let $G=Sp_{2n,A}$ and $\fg = Lie(G)$. By theorem~\ref{theo_chevalley_dominant} and proposition~\ref{invariant_weyl_sp2n}, $A[\fg]^G$ is a subring of $A[\ft]^W=A[\sigma_k(X_i^2)\,;\,x.\sigma_k(X_i)]$. With the notations of lemma~\ref{lemm_dominant_sur_fibres_sp}, it is also a subring of the image of $A[\fh]^H$ in $A[\sigma_k(X_i^2)\,;\,x.\sigma_k(X_i)]$. The latter is $A[\sigma_k(X_i^2)]$ by proposition~\ref{prop_sl2}. Since $c_{2k}\in A[\fg]^G$ maps to $\pm \sigma_k(X_i^2)\in A[\ft]^W$, the theorem is proved. \end{proo}

https://arxiv.org/abs/1105.5778

Fejer-type inequalities

The aim of this paper is to present some new Fejer-type results for convex functions. Improvements of Young's inequality (the arithmetic-geometric mean inequality) and other applications to special means are pointed as well.

\section{Preliminaries} {We start from an important result related to the convex functions due to Ch. Hermite \cite{Her1883} and J. Hadamard \cite{Had1893} which asserts that for every continuous convex func}tion{\ \textit{\ }}$f:[a,b]\rightarrow \mathbb{R}${\textit{\ }}the following inequalities hold:{\ \begin{equation} f\left( \frac{a+b}{2}\right) \leq \frac{1}{b-a}\int_{a}^{b}f(x)\mathrm{d x\leq \frac{f(a)+f(b)}{2}. \end{equation } Fej\'{e}r \cite{Fej1906} established the following well-known weighted generalization: \begin{Prop} \label{prop_fej}If {\textit{\ }}$f:[a,b]\rightarrow \mathbb{R}${\textit{\ \ is }continuous and convex and if }$g:[a,b]\rightarrow \mathbb{R}_{+}$\textit \ }{\ is integrable and symmetric about }$\frac{a+b}{2}$ ( i.e. $g\left( x\right) =g\left( a+b-x\right) $)$,$ then{\ }the following inequalities hold \begin{equation} f\left( \frac{a+b}{2}\right) \int_{a}^{b}g(x)\mathrm{d}x\leq \int_{a}^{b}f(x)g\left( x\right) \mathrm{d}x\leq \frac{f(a)+f(b)}{2 \int_{a}^{b}g(x)\mathrm{d}x. \label{fej} \end{equation } \end{Prop} Before stating the results we recall some useful facts from the literature. S. S. Dragomir, P. Cerone and A. Sofo present in \cite{Dra2000a,Dra2000b} the following estimates of the precision in the Hermite-Hadamard inequality: \begin{Prop} \label{nic}Let $f:[a,b]\rightarrow \mathbb{R}$ be a twice differentiable function such that there exists real constants $m$ and $M$ so that $m\leq f^{\prime \prime }\leq M$. The \begin{equation} m\frac{(b-a)^{2}}{24}\leq \frac{1}{b-a}\int_{a}^{b}f(x)\mathrm{d}x-f\left( \frac{a+b}{2}\right) \leq M\frac{(b-a)^{2}}{24} \label{cpn_1} \end{equation \textit{\ and \begin{equation} m\frac{(b-a)^{2}}{12}\leq \frac{f(a)+f(b)}{2}-\frac{1}{b-a}\int_{a}^{b}f(x \mathrm{d}x\leq M\frac{(b-a)^{2}}{12}. \label{cpn_2} \end{equation \textit{\ } \end{Prop} These inequalities follow from the Hermite-Hadamard inequality, for the convex functions $f(x)-m\frac{x^{2}}{2}$ and $M\frac{x^{2}}{2}-f\left( x\right) $.\textit{\ } Motivated by the above results, the purpose of this paper is to discuss further inequalities of Fej\'{e}r type. \section{Fej\'{e}r type inequalities for convex functions} \begin{The} \label{th_2}Let $f:[a,b]\rightarrow \mathbb{R}$ be a twice differentiable function such that there exist real constants $m$ and $M$ so that $m\leq f^{\prime \prime }\leq M$. Then \begin{equation} m\frac{\lambda (1-\lambda )}{2}(a-b)^{2}\leq \lambda f(a)+(1-\lambda )f(b)-f(\lambda a+(1-\lambda )b)\leq M\frac{\lambda (1-\lambda )}{2 (a-b)^{2}, \label{ref_1} \end{equation for all $\lambda \in \lbrack 0,1].$ \end{The} \noindent \textit{Proof:} We consider the function $g:[0,1]\rightarrow \mathbb{R}$, defined by \begin{equation*} g(\lambda )=\lambda f(a)+(1-\lambda )f(b)-f(\lambda a+(1-\lambda )b)-m\frac \lambda (1-\lambda )}{2}(a-b)^{2}. \end{equation* Since \begin{equation*} g^{\prime \prime }(\lambda )=(a-b)^{2}[m-f^{\prime \prime }(\lambda a+(1-\lambda )b)]\leq 0, \end{equation* the function $g$ is concave. But $g(0)=g(1)=0$, which implies that \begin{equation*} 0=(1-\lambda )g(0)+\lambda g(1)\leq g((1-\lambda )\cdot 0+\lambda \cdot 1)=g(\lambda ), \end{equation* for all $\lambda \in \lbrack 0,1]$. Therefore, we obtain the first part of inequality (\ref{ref_1}).\newline To see that the later inequality holds, our next step is to take the convex function $h:[0,1]\rightarrow \mathbb{R}$, defined by \begin{equation*} h(\lambda )=\lambda f(a)+(1-\lambda )f(b)-f(\lambda a+(1-\lambda )b)-M\frac \lambda (1-\lambda )}{2}(a-b)^{2}. \end{equation* Since $h(0)=h(1)=0$, \begin{equation*} 0=(1-\lambda )h(0)+\lambda h(1)\geq h((1-\lambda )\cdot 0+\lambda \cdot 1)=h(\lambda ), \end{equation* for all $\lambda \in \lbrack 0,1]$. The assertion is now clear. \hfill \hbox{\rule{6pt}{6pt}} For a slight generalization and alternative proof of Theorem \ref{th_2} the reader is referred to \cite[Theorem 4.2]{FMM2011}.\medskip\ \begin{Rem} By integrating each term of the inequality (\ref{ref_1}) on $\left[ 0, \right] $ with respect to the variable $\lambda $ we recover the inequality \ref{cpn_2}). \end{Rem} \begin{Cor} Preserving the notation of Theorem \ref{th_2}, the following inequalities hold \begin{eqnarray} m\frac{(1-2\lambda )^{2}}{8}(a-b)^{2} &\leq &\frac{f(\lambda a+(1-\lambda )b)+f((1-\lambda )a+\lambda b)}{2}-f\left( \frac{a+b}{2}\right) \label{ref_3} \\ &\leq &M\frac{(1-2\lambda )^{2}}{8}(a-b)^{2} \end{eqnarray for all $\lambda \in \lbrack 0,1].$ \end{Cor} \textit{Proof:} According to Theorem \ref{th_2} for $\lambda =\frac{1}{2}$ we obtain the following result, previously established in \cite{AF2010}: \begin{equation} \frac{m}{8}(b-a)^{2}\leq \frac{f(a)+f(b)}{2}-f\left( \frac{a+b}{2}\right) \leq \frac{M}{8}(b-a)^{2}. \label{ref_2} \end{equation We consider the above inequality (\ref{ref_2}) replacing $a\rightarrow \lambda a+(1-\lambda )b$ and $b\rightarrow (1-\lambda )a+\lambda b$ (the hypothesis $m\leq f^{\prime \prime }\leq M$ is still working on the interval with these endpoints because it is contained by $[a,b]$) and we get the claimed result. \hfill \hbox{\rule{6pt}{6pt}} \begin{Rem} Notice that by integrating all terms of (\ref{ref_3}) on $\left[ 0,1\right] $ with respect to $\lambda $ we recover now the inequality (\ref{cpn_1}). \end{Rem} Next we give some estimates of the Fej\'{e}r inequalities (Proposition \re {prop_fej}): \begin{The} \label{fej_2}Let $f:[a,b]\rightarrow \mathbb{R}$ be a twice differentiable function such that there exist real constants $m$ and $M$ so that $m\leq f^{\prime \prime }\leq M$. Assume $g:[a,b]\rightarrow \mathbb{R}_{+}$\textit \ }{\ is integrable and symmetric about }$\frac{a+b}{2}.$ Then{\ }the following inequalities hold \begin{eqnarray} \frac{m}{2}\int_{a}^{b}(t-a)(b-t)g\left( t\right) \mathrm{d}t &\leq &\frac f\left( a\right) +f\left( b\right) }{2}\int_{a}^{b}g\left( t\right) \mathrm{ }t-\int_{a}^{b}f(t)g\left( t\right) \mathrm{d}t\ \label{1} \\ &\leq &\frac{M}{2}\int_{a}^{b}(t-a)(b-t)g\left( t\right) \mathrm{d}t\ \ \end{eqnarray an \begin{eqnarray} \frac{m}{8}\int_{a}^{b}(2t-a-b)^{2}g\left( t\right) \mathrm{d}t &\leq &\int_{a}^{b}f(t)g\left( t\right) \mathrm{d}t-f\left( \frac{a+b}{2}\right) \ \int_{a}^{b}g\left( t\right) \mathrm{d}t \label{2} \\ &\leq &\frac{M}{8}\int_{a}^{b}(2t-a-b)^{2}g\left( t\right) \mathrm{d}t.\ \ \end{eqnarray} \end{The} \textit{Proof:} We multiply (\ref{ref_1}) by $g\left( \lambda a+(1-\lambda )b\right) $ and integrate the result on $\left[ 0,1\right] $ with respect to the variable $\lambda $. Using the change of the variable $\lambda a+(1-\lambda )b=t$ we get \begin{eqnarray} &&\frac{m}{2}\int_{a}^{b}(t-a)(b-t)g\left( t\right) \mathrm{d}t \notag \\ &\leq &f\left( a\right) \int_{a}^{b}\frac{b-t}{b-a}g\left( t\right) \mathrm{ }t+f\left( b\right) \int_{a}^{b}\frac{t-a}{b-a}g\left( t\right) \mathrm{d t-\int_{a}^{b}f(t)g\left( t\right) \mathrm{d}t\ \notag \\ &\leq &\frac{M}{2}\int_{a}^{b}(t-a)(b-t)g\left( t\right) dt. \label{right_fejer_1} \end{eqnarray On the other hand, due to the symmetry property of $g$, for $t=a+b-x,$ we also have \begin{eqnarray} &&\frac{m}{2}\int_{a}^{b}(b-x)(x-a)g\left( x\right) dx \notag \\ &\leq &f\left( a\right) \int_{a}^{b}\frac{x-a}{b-a}g\left( x\right) dx+f\left( b\right) \int_{a}^{b}\frac{b-x}{b-a}g\left( x\right) dx-\int_{a}^{b}f(t)g\left( t\right) dt\ \notag \\ &\leq &\frac{M}{2}\int_{a}^{b}(b-x)(x-a)g\left( x\right) dx. \label{right_fejer_2} \end{eqnarray Summing (\ref{right_fejer_1}) and (\ref{right_fejer_2}) we find (\ref{1}). In order to prove the remaining inequalities (\ref{2})\textit{\ }we follow same steps as above, using (\ref{ref_3}) instead of (\ref{ref_1}). The computation is straightforward, taking into account the symmetry of $g$ (applied now as $g\left( \lambda a+(1-\lambda )b\right) =g\left( (1-\lambda )a+\lambda b\right) $). We omit the details. This completes the proof. \hfill \hbox{\rule{6pt}{6pt}} It is remarkable that (\ref{1}) agrees, having an extended form, with \cite pp.53, Exercise 4]{CPN2006}. \begin{Rem} If $g:\left[ a,b\right] \rightarrow \left[ 0,1\right] $ then the function h\left( x\right) =1-g\left( x\right) $ satisfies the same symmetry and positivity conditions and Theorem \ref{fej_2} also applies. That yields the following estimates of the precision in (\ref{1}): \begin{eqnarray} &&\left( b-a\right) \left( \frac{f\left( a\right) +f\left( b\right) }{2} \frac{1}{b-a}\int_{a}^{b}f(t)\mathrm{d}t\ -m\frac{(b-a)^{2}}{12}\right) \notag \\ &\geq &\frac{f\left( a\right) +f\left( b\right) }{2}\int_{a}^{b}g\left( t\right) \mathrm{d}t-\int_{a}^{b}f(t)g\left( t\right) \mathrm{d}t\ -\ \frac{ }{2}\int_{a}^{b}(t-a)(b-t)g\left( t\right) \mathrm{d}t\geq 0 \end{eqnarray an \begin{eqnarray} &&\left( b-a\right) \left( M\frac{(b-a)^{2}}{12}-\frac{f\left( a\right) +f\left( b\right) }{2}+\frac{1}{b-a}\int_{a}^{b}f(t)\mathrm{d}t\ \right) \notag \\ &\geq &\frac{M}{2}\int_{a}^{b}(t-a)(b-t)g\left( t\right) \mathrm{d}t-\frac f\left( a\right) +f\left( b\right) }{2}\int_{a}^{b}g\left( t\right) \mathrm{ }t+\int_{a}^{b}f(t)g\left( t\right) \mathrm{d}t\ \geq 0.\ \end{eqnarray By a similar technique one can estimate (\ref{2}). \end{Rem} \begin{Rem} For the particular case $g\left( x\right) =1$ if we apply Theorem \ref{fej_2} on the intervals $\left[ a,\frac{a+b}{2}\right] ,$ $\left[ \frac{a+b}{2}, \right] $ we get \begin{equation} \frac{m\left( b-a\right) ^{2}}{48}\leq \frac{1}{2}\left( \frac{f\left( a\right) +f\left( b\right) }{2}+f\left( \frac{a+b}{2}\right) \right) -\frac{ }{b-a}\int_{a}^{b}f(t)g\left( t\right) \mathrm{d}t\leq \frac{M\left( b-a\right) ^{2}}{48}\ \end{equation and \begin{equation} \frac{m\left( b-a\right) ^{2}}{96}\leq \frac{1}{b-a}\int_{a}^{b}f(t)g\left( t\right) \mathrm{d}t-\frac{1}{2}\left( f\left( \frac{3a+b}{4}\right) \ +f\left( \frac{a+3b}{4}\right) \right) \leq \frac{m\left( b-a\right) ^{2}}{9 }.\ \ \end{equation} \end{Rem} The following theorem gives new Fej\'{e}r-type inequalities. \begin{The} \label{th_3}Let $f:[a,b]\rightarrow \mathbb{R}$ be a continuous, convex function and $g:[a,b]\rightarrow \mathbb{R}_{+}$ be continuous. Then the following statements hold. 1) If $g$ is monotonically decreasing the \begin{equation} \frac{f\left( a\right) +f\left( b\right) }{2}\int_{a}^{b}g\left( t\right) \mathrm{d}t-\int_{a}^{b}f(t)g\left( t\right) \mathrm{d}t\ \geq \frac{f\left( a\right) +f\left( x\right) }{2}\int_{a}^{x}g\left( t\right) \mathrm{d t-\int_{a}^{x}f(t)g\left( t\right) \mathrm{d}t\ \geq 0; \label{fg_1} \end{equation} 2) If $g$ is monotonically increasing then \begin{equation} \int_{a}^{b}f(t)g\left( t\right) \mathrm{d}t\ -f\left( \frac{a+b}{2}\right) \int_{a}^{b}g\left( t\right) \mathrm{d}t\geq \int_{a}^{x}f(t)g\left( t\right) \mathrm{d}t\ -f\left( \frac{a+x}{2}\right) \int_{a}^{x}g\left( t\right) \mathrm{d}t\geq 0 \label{fg_2} \end{equation for all $x\in \left( a,b\right) .$ \end{The} \textit{Proof:} 1) We consider the function $h_{1}:[a,b]\rightarrow \mathbb{ }$, defined by \begin{equation*} h_{1}(x)=\frac{f\left( a\right) +f\left( x\right) }{2}\int_{a}^{x}g\left( t\right) \mathrm{d}t-\int_{a}^{x}f(t)g\left( t\right) \mathrm{d}t\ . \end{equation* Its first derivative i \begin{equation*} h_{1}^{\prime }(x)=\frac{f^{\prime }\left( x\right) }{2}\int_{a}^{x}g\left( t\right) \mathrm{d}t-\frac{f\left( x\right) -f\left( a\right) }{2}g\left( x\right) . \end{equation* Using the mean value theorems there exist $c_{1},c_{2}\in \left[ a,x\right] $ such tha \begin{equation*} h_{1}^{\prime }(x)=\left( \frac{f^{\prime }\left( x\right) }{2}g\left( c_{1}\right) -\frac{f^{\prime }\left( c_{2}\right) }{2}g\left( x\right) \right) \left( x-a\right) . \end{equation* Thus, by the convexity of $f$ and to the monotonicity of $g,$ we have f^{\prime }\left( x\right) \geq f^{\prime }\left( c_{2}\right) $ and g\left( c_{1}\right) \geq g\left( x\right) ,$ hence we conclude that $h_{1}$ is increasing on its domain and $h_{1}\left( b\right) \geq h_{1}\left( x\right) \geq h_{1}\left( a\right) =0.$ Thus we have (\ref{fg_1}), as asserted. 2) Similarly, we consider the function $h_{2}:[a,b]\rightarrow \mathbb{R}$, defined by \begin{equation*} h_{2}(x)=\int_{a}^{x}f(t)g\left( t\right) \mathrm{d}t\ -f\left( \frac{a+x}{2 \right) \int_{a}^{x}g\left( t\right) \mathrm{d}t \end{equation* and we compute its first derivative \begin{equation*} h_{2}^{\prime }(x)=\left( f(x)-f\left( \frac{a+x}{2}\right) \right) g\left( x\right) \ -\frac{1}{2}f^{\prime }\left( \frac{a+x}{2}\right) \int_{a}^{x}g\left( t\right) \mathrm{d}t. \end{equation* There exist $k_{1}\in \left[ \frac{a+x}{2},x\right] $ and $k_{2}\in \left[ a,x\right] $ such tha \begin{equation*} h_{2}^{\prime }(x)=\left( f^{\prime }\left( k_{1}\right) g\left( x\right) \ -f^{\prime }\left( \frac{a+x}{2}\right) g\left( k_{2}\right) \right) \frac x-a}{2}. \end{equation* Therefore, due to the monotonicity we have $f^{\prime }\left( k_{1}\right) \geq f^{\prime }\left( \frac{a+x}{2}\right) $ and $g\left( x\right) \geq g\left( k_{2}\right) ,$ which leads that $h_{2}$ is increasing on its domain and $h_{2}\left( b\right) \geq h_{2}\left( x\right) \geq h_{2}\left( a\right) =0.$ Thus the proof is completed. \hfill \hbox{\rule{6pt}{6pt}} The following direct consequence incorporates the classic statement of the Hermite-Hadamard inequality. \begin{Cor} \label{cor}Suppose $f:[a,b]\rightarrow \mathbb{R}$ is continuous and convex. The \begin{equation} \frac{f\left( a\right) +f\left( b\right) }{2}-\frac{1}{b-a}\int_{a}^{b}f(t \mathrm{d}t\ \geq \frac{x-a}{b-a}\left( \frac{f\left( a\right) +f\left( x\right) }{2}-\frac{1}{x-a}\int_{a}^{x}f(t)\mathrm{d}t\ \right) \geq 0 \label{c_1} \end{equation and \begin{equation} \frac{1}{b-a}\int_{a}^{b}f(t)\mathrm{d}t\ -f\left( \frac{a+b}{2}\right) \geq \frac{x-a}{b-a}\left( \frac{1}{x-a}\int_{a}^{x}f(t)\mathrm{d}t\ -f\left( \frac{a+x}{2}\right) \right) \geq 0 \label{c_2} \end{equation for all $x\in \left( a,b\right) .$ \end{Cor} \textit{Proof:} We apply Theorem \ref{th_3} with $g\left( x\right) =1,$ which satisfies both monotonicity conditions. \hfill \hbox{\rule{6pt}{6pt}} \begin{Rem} Under the same assumptions as in Proposition \ref{nic} when we apply Corollary \ref{cor} to the convex function $f(x)-m\frac{x^{2}}{2},$ we recover and improve the inequalities (\ref{cpn_1}) as follows \begin{eqnarray} &&\frac{1}{b-a}\int_{a}^{b}f(t)\mathrm{d}t-f\left( \frac{a+b}{2}\right) - \frac{(b-a)^{2}}{24} \notag \\ &\geq &\frac{x-a}{b-a}\left( \frac{1}{x-a}\int_{a}^{x}f(t)\mathrm{d t-f\left( \frac{a+x}{2}\right) -m\frac{(x-a)^{2}}{24}\right) \geq 0 \end{eqnarray an \begin{eqnarray} &&\frac{M}{8}(b-a)^{2}-\left( \frac{1}{b-a}\int_{a}^{b}f(t)\mathrm{d t-f\left( \frac{a+b}{2}\right) \right) \notag \\ &\geq &\frac{x-a}{b-a}\left[ \frac{M}{8}(x-a)^{2}-\left( \frac{1}{x-a \int_{a}^{x}f(t)\mathrm{d}t-f\left( \frac{a+x}{2}\right) \right) \right] \geq 0.\ \end{eqnarray \textit{\ Similarly if we use the convex function }$M\frac{x^{2}}{2}-f\left( x\right) $ we get improvements of (\ref{cpn_2}) which at this moment can easily be written by the interested reader. Obviously same steps could be followed from Theorem \ref{th_3}, improving that way Theorem \ref{fej_2}. \end{Rem} We end this section with the weighted statement of a known result concerning convex functions. In the light of Proposition \ref{prop_fej}, the following statement appears as a trivial generalization of a result due to Vasi\'{c} and Lackovi\'{c} \cite{vas76}, and Lupa\c{s} \cite{lup76} (cf. J. E. Pe\v{c}ari\'{c} et al. \cite[pp. 143]{pec92}) and we omit its proof. \begin{Prop} Let $p$ and $q$ be two positive numbers and $a_{1}\leq a\leq b\leq b_{1}.$ Let $g:[a,b]\rightarrow \mathbb{R}_{+}$\textit{\ }{\ be integrable and symmetric about }$A=\frac{pa+qb}{p+q}.$ Then the inequalitie \begin{equation} f\left( \frac{pa+qb}{p+q}\right) \int_{A-y}^{A+y}g\left( t\right) \mathrm{d t\leq \int_{A-y}^{A+y}\,f(x)\,g\left( x\right) \mathrm{d}x\leq \frac pf\left( a\right) +qf\left( b\right) }{p+q}\int_{A-y}^{A+y}g\left( t\right) \mathrm{d}t \end{equation hold for $y>0$ and all continuous convex functions $f:\left[ a_{1},b_{1 \right] \rightarrow \mathbb{R}$ if and only if \begin{equation*} y\leq \frac{b-a}{p+q}\min \left\{ p,q\right\} . \end{equation*} \end{Prop} \section{Application to special means} From the inequality (\ref{c_1}) applied to the convex function $t^{p}$, with $p\in \left( -\infty ,0\right) \cup \left[ 1,\infty \right) \setminus \left\{ -1\right\} $ we have$\ \begin{equation} \left( b-a\right) \left\{ \left[ A_{p}\left( a,b\right) \right] ^{p}-\left[ L_{p}\left( a,b\right) \right] ^{p}\right\} \geq \left( x-a\right) \left\{ \left[ A_{p}\left( a,x\right) \right] ^{p}-\left[ L_{p}\left( a,x\right) \right] ^{p}\right\} , \label{AL} \end{equation where $x\in \left[ a,b\right] .$ Here $A_{p}\left( a,b\right) =\left( \frac a^{p}+b^{p}}{2}\right) ^{1/p}$ is the power mean and $L_{p}\left( a,b\right) =\left( \frac{b^{p+1}-a^{p+1}}{\left( p+1\right) \left( b-a\right) }\right) ^{1/p}$ is the $p$-logarithmic mean. Also the limit case $p\rightarrow -1$ (or we may equivalently say the case of the convex function $1/t$) gives u \begin{equation*} \left( b-a\right) \left\{ \frac{1}{H\left( a,b\right) }-\frac{1}{L\left( a,b\right) }\right\} \geq \left( x-a\right) \left\{ \frac{1}{H\left( a,x\right) }-\frac{1}{L\left( a,x\right) }\right\} , \end{equation* where $H\left( a,b\right) =\frac{2ab}{a+b}$ is the harmonic mean and L\left( a,b\right) =\frac{b-a}{\log b-\log a}$ is the logarithmic mean. It is also useful to consider the inequality (\ref{c_1}) applied for the convex function $-\log t$, when we get \begin{equation*} \left[ \frac{A\left( a,b\right) }{I\left( a,b\right) }\right] ^{b-a}\geq \left[ \frac{A\left( a,x\right) }{I\left( a,x\right) }\right] ^{x-a}, \end{equation* for $a\neq b,$ $x\in \left[ a,b\right] ,$ where $A\left( a,b\right) =\frac a+b}{2}$ is the arithmetic mean and $I\left( a,b\right) =\frac{1}{e}$ \left( \frac{b^{b}}{a^{a}}\right) ^{\frac{1}{b-a}}$ is the identric mean. In the remainder, we focus on two immediate particular cases of Theorem \re {th_2} that help us to give improvements of the well known arithmetic-geometric mean inequality (also known as Young's inequality). \emph{1)} We apply the theorem to the function $f:[a,b]\rightarrow \mathbb{R \ (a>0)$ defined by $f(x)=-\log x$, which leads to \begin{equation} e^{\frac{\lambda (1-\lambda )(a-b)^{2}}{2b^{2}}}\leq \frac{\lambda a+(1-\lambda )b}{a^{\lambda }b^{1-\lambda }}\leq e^{\frac{\lambda (1-\lambda )(a-b)^{2}}{2a^{2}}}. \label{3} \end{equation Since $e^{\frac{\lambda (1-\lambda )(a-b)^{2}}{2b^{2}}}\geq 1$, we obtain a refinement of Young's inequality where $\lambda \in \lbrack 0,1].$ We also obtained a reverse inequality for Young's inequality. \emph{2)} Next, we apply the theorem to the function $f:[\log a,\log b]\rightarrow \mathbb{R}$, defined by $f(x)=\exp x$ and we arrive at \begin{equation} \frac{\lambda (1-\lambda )a}{2}\log ^{2}\left( \frac{a}{b}\right) \leq \lambda a+(1-\lambda )b-a^{\lambda }b^{1-\lambda }\leq \frac{\lambda (1-\lambda )b}{2}\log ^{2}\left( \frac{a}{b}\right) , \label{5} \end{equation where $a,b>0$ and $\lambda \in \lbrack 0,1]$. The inequality (\ref{5}) gives an improvement of Young's inequality. \section*{Acknowledgements} The first author was supported in part by the Romanian Ministry of Education, Research and Innovation through the PNII Idei project 842/2008. The second author was supported by CNCSIS Grant $420/2008.$

https://arxiv.org/abs/1612.06013

Sketch and Project: Randomized Iterative Methods for Linear Systems and Inverting Matrices

Probabilistic ideas and tools have recently begun to permeate into several fields where they had traditionally not played a major role, including fields such as numerical linear algebra and optimization. One of the key ways in which these ideas influence these fields is via the development and analysis of randomized algorithms for solving standard and new problems of these fields. Such methods are typically easier to analyze, and often lead to faster and/or more scalable and versatile methods in practice.This thesis explores the design and analysis of new randomized iterative methods for solving linear systems and inverting matrices. The methods are based on a novel sketch-and-project framework. By sketching we mean, to start with a difficult problem and then randomly generate a simple problem that contains all the solutions of the original problem. After sketching the problem, we calculate the next iterate by projecting our current iterate onto the solution space of the sketched problem.

\chapter[Introduction]{Introduction} \chaptermark{Introduction} \label{ch:introduction} { \epigraph{\emph{We can only see a short distance ahead, but we can see plenty there that needs to be done.}}{Alan Turing} \let\clearpage\relax \section{Introduction: What's to Come} } \label{secChOne:whats} This thesis explores the design and analysis of new randomized iterative methods for solving linear systems and inverting matrices. All the methods presented in this thesis are globally and linearly convergent. Consequently, the methods are well suited to quickly obtain approximate solutions. The methods are based on a novel sketch-and-project framework. By sketching we mean, to start with a difficult problem and then randomly generate a simple problem that contains all the solutions of the original problem. For instance, consider the linear system \begin{equation}\label{eq:linsysintro}Ax=b,\end{equation} where $A\in \mathbb{R}^{m\times n},x\in \mathbb{R}^n$ and $b\in \mathbb{R}^m$. We suppose throughout the thesis that there exists a solution $x^* \in \mathbb{R}^n$ to the linear system, that is, the linear system is \emph{consistent}. Let $S \in \mathbb{R}^{m\times q}$ be a random matrix with the same number of rows as $A$ but far fewer columns ($q \ll n$.) The resulting \emph{sketched} linear system is given by \[S^\top A x = S^\top b,\] which has a relatively small number of rows, and is thus easier to solve. There has been a concerted effort into designing the distribution of $S$ with the property that the solution set of the sketched linear system is close to the solution set of the original linear system with high probability, particularly so for solving linear systems that arise from the least-squares problem~\cite{Pilanci2014,Mahoney:2011,Drineas2011}. Determining an $S$ with such a property can be difficult and often depends on properties of $A$ that are expensive to compute. Here we take a different approach and use sketching combined with a projection process. We apply our sketch-and-project technique not only to solve linear systems, but also to find the projection of a vector onto the solution space of a linear system and to invert matrices (by sketching, for example, the inverse equation $AX=I$). Throughout the entirety of this thesis, we assume that we can access the system matrix $A$ through matrix-vector products. Thus every element $A_{ij}$ of the system matrix may not be explicitly available. The methods presented here are designed with this restriction in mind and are thus compatible with the setting where $A$ can only be accessed as an operator. In the remainder of this section we give a summary of each chapter of this thesis. The detailed proofs and careful deductions of any claims made here are left to the chapters. \paragraph{Chapter~\ref{ch:linear_systems}: Linear Systems with Sketch-and-Project.} In Chapter~\ref{ch:linear_systems} we develop an iterative process that gradually refines an approximate solution to~\eqref{eq:linsysintro} using a sequence of sketching matrices, as opposed to the one shot sketching method. To describe our method, let $x^k\in \mathbb{R}^n$ be our current estimate of the solution to~\eqref{eq:linsysintro}. We obtain an improved estimate by projecting $x^k$ onto the solution space of a sketched system \begin{equation} \label{C1eq:sketchproj} x^{k+1} = \arg \min_x \norm{x^k-x}_B^2 \quad \mbox{subject to} \quad S^\top Ax=S^\top b,\end{equation} where $S$ is drawn in each iteration independently from a pre-specified distribution, $B$ is a positive definite matrix and $\norm{x}_B^2 \eqdef \dotprod{Bx,x}.$ This iterative process has a closed form solution given by \begin{equation} \label{C1eq:xupdate}x^{k+1} = x^k - B^{-1}A^\top S(S^\top A B^{-1}A^\top S)^{\dagger}S^\top(Ax^k-b),\end{equation} where $\dagger$ denotes the (Moore-Penrose) pseudoinverse. Using the closed form expression for the update~\eqref{C1eq:xupdate} we show that the iterates converge if $A$ has full column rank and under mild assumptions on the distribution of $S$. In particular, the convergence analysis will depend heavily on the following random matrix \[Z \eqdef A^\top S(S^\top A B^{-1}A^\top S)^{\dagger}S^\top A,\] which governs the iterative process~\eqref{C1eq:xupdate}. Indeed, we will show that when $A$ has full column rank and for any $x_0 \in \mathbb{R}^n,$ the iterates~\eqref{C1eq:xupdate} converge to the unique solution $x^* \in \mathbb{R}^n$ of the linear system exponentially fast according to \begin{equation} \label{ch:one:Enormconv} \E {\norm{x^{k} -x^{*} }_B^2 } \leq \rho^k \;\cdot\; \norm{x^{0} - x^{*}}_B^2, \end{equation} where \[\rho \eqdef 1- \lambda_{\min}(B^{-1/2}\E{Z}B^{-1/2}).\] By $B^{-1/2}$ we denote the unique positive definite square root of $B^{-1}.$ Therefore $B^{-1/2}B^{-1/2} =B^{-1}.$ We use $\E{\cdot}$ to denote the expectation operator. For instance, $\E{Z}$ is the expected value of $Z$. Since $Z$ is a function of $S$, which is the only random component of $Z$, we have that $\E{Z}$ is an expectation taken over the distribution of $S$. Because of the importance that $Z$ plays in this thesis, later in this chapter in Section~\ref{C1sec:tools}, we prove several properties of $Z.$ In Section~\ref{C2sec:discrete} we design a discrete distribution for $S$ that yields easily interpretable convergence rates in terms of a scaled condition number. Furthermore, we show that by choosing $B$ and the distribution of $S$ appropriately we recover a comprehensive array of well known algorithms as special cases, such as the randomized Kaczmarz method~\cite{Strohmer2009} and randomized Coordinate Descent~\cite{Leventhal2010}, demonstrating the expressive power of the framework. Having established convergence rates for each method defined by $(B,S)$, we explore questions such as: \emph{what distribution of $S$ yields the optimal convergence rates}? We determine that the optimal distribution of $S$, chosen from a family of discrete distributions, is the solution to a particular semidefinite program. This result determines, for instance, the optimal distribution for selecting the rows of the linear system in the randomized Kaczmarz method. Furthermore, the optimized randomized Kaczmarz method is shown to converge significantly faster than the randomized Kaczmarz method using the standard distribution for $S$. Our framework also allows for $S$ to have a continuous distribution, and to give an insight into the possibilities, we present three methods based on a Gaussian sketching matrix $S$ and give convergence rates for each. We then conclude the chapter with further numeric experiments that compare the different methods presented throughout the chapter. \paragraph{Chapter~\ref{ch:SDA}: Stochastic Dual Ascent for Finding the Projection of a Vector onto a Linear System.} In Chapter~\ref{ch:SDA} we consider the more general problem of finding the projection of a given vector $c \in \mathbb{R}^n$ onto the solution space of a linear system, that is \begin{equation} \min_{x \in \mathbb{R}^n} \quad \tfrac{1}{2}\|x-c\|_B^2 \quad \text{subject to} \quad Ax=b. \label{C1eq:P} \end{equation} To solve this projection problem we develop a new randomized iterative algorithm: {\em stochastic dual ascent (SDA)}. The method is dual in nature, and iteratively solves the dual of the projection problem~\eqref{C1eq:P}. The dual problem is a non-strongly concave quadratic maximization problem without constraints given by \begin{equation}\label{C1eq:D} \max_{y \in \mathbb{R}^m} \quad (b-Ac)^\top y - \tfrac{1}{2}\|A^\top y\|_{B^{-1}}^2. \end{equation} Each iterate $y^{k+1} \in \mathbb{R}^m$ of the SDA method is carefully chosen from a random affine space that passes through the previous iterate \begin{equation}\label{C1eq:SDA-compact0} y^{k+1} \in y^k + \myRange{S}, \end{equation} where $S$ is a random matrix drawn independently from a fixed distribution. Specifically, $y^{k+1}$ is the point with least norm that maximizes the dual objective in~\eqref{C1eq:D} constrained within the random affine space~\eqref{C1eq:SDA-compact0}. By mapping our dual iterates~\eqref{C1eq:SDA-compact0} to primal iterates, we uncover that the SDA method is a dual version of the sketch-and-project method~\eqref{C1eq:xupdate}. We then proceed to strengthen our convergence results established in the previous chapter. First, we do away with the assumption that $A$ has full column rank and consider any nonzero matrix $A$ and consistent linear system. In this more general setting, we prove that for $x_0 \in c + \myRange{B^{-1}A^\top}$ the primal iterates converge to the solution of~\eqref{C1eq:P} exponentially fast in expectation according to~\eqref{ch:one:Enormconv} with a convergence rate of \[\rho = 1- \lambda_{\min}^+(B^{-1/2}A^THAB^{-1/2}),\] where \begin{equation}\label{eq:Hintro} H = \E{S(S^\top AB^{-1}A^\top S)^{\dagger}S^\top}, \end{equation} and $\lambda_{\min}^+(B^{-1/2}A^THAB^{-1/2})$ denotes the smallest nonzero eigenvalue of $B^{-1/2}A^THAB^{-1/2}$. The only condition for this convergence to hold is that $H$ be nonsingular. We completely characterize the discrete distributions of $S$ for which $H$ is nonsingular in Section~\ref{sec:Hnonsingular}. This shows, for instance, that the Kaczmarz method converges so long as the system matrix has no zero rows. Thus assuming that $A$ has full column rank, as is required in the convergence theorems in Chapter~\ref{ch:linear_systems}, is an unnecessary assumption for proving convergence of the sketch-and-project method. But the full column rank assumption makes the proofs of convergence simpler and thus, for pedagogic reasons, we have presented the proofs assuming that $A$ has full column rank earlier on in Chapter~\ref{ch:linear_systems}. We present further improvements in the convergence analysis and give a tight and insightful lower bound for the convergence rate that depends on the rank of $A.$ We also prove that the same rate of convergence $\rho$ governs the convergence of the dual function values, primal function values and the duality gap. Unlike traditional iterative methods, SDA converges under virtually no additional assumptions on the system (e.g., rank, diagonal dominance) beyond consistency. In fact, our lower bound improves as the rank of the system matrix drops. When our method specializes to a known algorithm, we either recover the best known rates, or improve upon them. Finally, we show that the framework can be applied to the distributed average consensus problem to obtain an array of new algorithms. The randomized gossip algorithm arises as a special case~\cite{Boyd2006,OlshevskyTsitsiklis2009}. \paragraph{Chapter~\ref{ch:inverse}: Randomized Matrix Inversion.} In Chapter~\ref{ch:inverse} we extend our method for solving linear systems to inverting matrices, and develop a family of methods with a specialized variant which maintains symmetry or positive definiteness of the iterates. The initial insight into our matrix inversion methods comes from the simple observation that for a nonsingular matrix $A \in \mathbb{R}^{n\times n}$ the inverse is the solution in $X$ to one of the \emph{inverse equations} \[AX =I \quad \mbox{or}\quad XA =I.\] Our method for inverting $A$ calculates the new iterate $X_{k+1} \in \mathbb{R}^{n\times n}$ by projecting the previous iterate $X_k \in \mathbb{R}^{n\times n}$ onto the solution space of a sketched version of one of the two inverse equations: either the \emph{row} sketched variant $S^\top AX=S^\top$ or the \emph{column} sketched variant $XAS = S,$ where $S \in \mathbb{R}^{n \times q}$ is a random matrix drawn from a fixed distribution at each iteration. For example, using the column sketched variant a new iterate is calculated by solving \begin{equation} \label{C1eq:sketchprojinve} X_{k+1} = \arg \min_{X\in \mathbb{R}^{n\times n}} \norm{X_k-X}_{F(B)}^2 \quad \mbox{subject to} \quad XAS=S,\end{equation} where the norm is the weighted Frobenius norm. When $A$ is symmetric, it can be advantageous to maintain symmetries in the iterates $X_k.$ We propose a sketch-and-project method that maintains symmetry in the iterates by imposing symmetry as a constraint in \begin{equation} \label{C1eq:sketchprojinvesym} X_{k+1} = \arg \min_{X\in \mathbb{R}^{n\times n}} \norm{X_k-X}_{F(B)}^2 \quad \mbox{subject to} \quad XAS=S, \quad X = X^\top.\end{equation} All the methods we present converge globally and linearly, with the same explicit rate of convergence $\rho= 1- \lambda_{\min}(B^{-1/2}\E{Z}B^{-1/2})$. In special cases, we obtain stochastic block variants of several quasi-Newton updates, including the Broyden-Fletcher-Goldfarb-Shanno (BFGS) update. Ours are the first stochastic quasi-Newton updates shown to converge to an inverse of a fixed matrix. Through a dual viewpoint we uncover a fundamental link between quasi-Newton updates and approximate inverse preconditioning methods, which results in a new interpretation of the quasi-Newton methods. For instance, the BFGS update is the solution in $X$ to \[X_{k+1} = \arg_X \min_{X\in \mathbb{R}^{n\times n}, Y \in \mathbb{R}^{n\times q}} \frac{1}{2}\norm{X -A^{-1}}_{F(A)}^2 \quad \mbox{subject to} \quad X = X_k + SY^\top +YS^\top,\] for a particular (deterministic) choice of $S.$ This shows that the BFGS udpate can be interpreted as a projection of $A^{-1}$ onto a space of symmetric matrices. With explicit convergence rates for each method characterized by the distribution of $S$, we again raise the question of selecting a distribution of $S$ that results in a method with a faster convergence rate. Though different from the setting of solving a linear system, the goals of finding the inverse of $A$ and of designing a distribution of $S$ that results in an improved convergence rate are in synchrony. In particular, for many choices of $B$, it will transpire that the covariance of $S$ should be an estimate of the inverse of $A.$ One way to interpret this result is that $S$ should be chosen so that it not only sketches/compresses the inverse equations $AX=I$ or $XA =I$, but also, it should improve the condition number of these equations. This reasoning leads us to develop an adaptive variant of a randomized block BFGS (AdaRBFGS), where the distribution of $S$ depends on $X_k$. By inverting several matrices from varied applications, we demonstrate that AdaRBFGS is highly competitive when compared with the well established Newton-Schulz~\cite{Schulz1933} and the approximate inverse preconditioning methods~\cite{Chow1998,Saad2003,Gould1998,Benzi1999}. In particular, on large-scale problems our method outperforms the standard methods by orders of magnitude. Since the inspiration behind AdaRBFGS method comes from the desire to design an {\em optimal adaptive} distribution for $S$ by examining the convergence rate, this work also highlights the importance of developing algorithms with explicit convergence rates. \paragraph{Organization of Thesis} Chapters~\ref{ch:linear_systems},~\ref{ch:SDA} and~\ref{ch:inverse} are largely based on the papers~\cite{Gower2015,Gower2015c} and~\cite{Gower2016}, respectively. Excluding preliminary results in linear algebra presented in Section~\ref{C1sec:tools}, each of these chapter is mostly self-contained, including the objective, contributions, definitions and notation. \section{Why Randomized Methods} Why use randomization in algorithmic design? We can answer this question, in practical terms, by measuring the advantages of using randomized algorithms as compared with existing deterministic methods. The advantages in using randomized methods include: often algorithms that are easier to analyze and implement, better convergence in terms of improved convergence rates or range of convergence (the randomized methods are almost always globally convergent), lower memory requirements, and more scalable and parallelizable methods in practice. To substantiate these claims, we now compare the sketch-and-project randomized methods~\eqref{C1eq:sketchproj} for solving the linear system~\eqref{eq:linsysintro} to stationary methods and the Krylov methods. Note that all the iterative methods mentioned here benefit from using a preconditioner, so much so, they are often only used in conjunction with a preconditioner. But to compare the methods on a equal footing, we assume no preconditioning strategy has been applied. \subsection{A Case Study Comparing to Stationary Methods} Iterative methods that fit the simple form \begin{equation}\label{eq:stationarym} x^{k+1} = G x^{k-1} + c, \end{equation} are known as Stationary Methods, where $G \in \mathbb{R}^{n\times n}$ is the \emph{iteration matrix} and $c \in \mathbb{R}^n$ is the \emph{bias term}. Here neither $G$ nor $c$ depend on the iteration count. Methods that fit the format~\eqref{eq:stationarym} include the Jacobi method, the Gauss-Seidel method, the Successive Overrelaxation (SOR) method and the Symmetric Successive Overrelaxation (SSOR) method~\cite{barrett1994,Saad1986}. The sketch-and-project methods presented in this thesis~\eqref{C1eq:xupdate} would fit the format~\eqref{eq:stationarym} if it were not for the fact that $G$ and $c$ in our methods are randomly generated, and thus, can differ from one iteration to the next. Despite this difference, the sketch-and-project methods and the stationary methods share many similarities. \paragraph{Low memory requirements.} Stationary methods and~\eqref{C1eq:xupdate} are both easy to implement and have low memory requirements. Often only the previous iterate $x^k$ needs to be stored to enable the calculation of the next iterate. \paragraph{Existence of convergence rates.} A stationary method only converges if the spectral radius of $G$ is less than one~\cite{Saad1986}, that is, if $\rho(G) <1.$ Often $G$ is constructed by splitting the system matrix $A$, e.g., for square systems the Jacobi method iteration matrix is $G = I-D^{-1}(A-D)$ where $D = \mbox{diag}(A_{11}, \ldots, A_{nn})$. Thus for $\rho(G)<1$ to hold one needs strong assumptions on the spectrum of $A.$ Again using the Jacobi method as an example, if $A$ is strictly diagonally dominant, then $\rho(G)<1$ holds. In contrast, the sketch-and-project methods converge for virtually any matrix $A$ with an explicit convergence rate, as we show in Chapter~\ref{ch:SDA}. \paragraph{Parallelizable.} Due to their simple recurrence relationship, the stationary methods lead to straight forward parallel and distributed variants, see Section 2.5 in~\cite{Bertsekas:1989}. Furthermore, variants of stationary method for solving nonlinear systems have also been adapted to parallel architectures~\cite{BertsekasT91,Robinson2015}. The sketch-and-project methods are similar in this aspect, in that, they are amenable to parallel implementations. For instance, the coordinate descent is member of the sketch-and-project family, and various distributed and parallel variants of coordinate descent have been designed to solve optimization problems~\cite{shotgun,Fercoq2013a,PCDM,TTD}. \subsection{A Case Study Comparing to Krylov Methods} The Krylov methods are a well studied and established class of iterative methods for solving linear systems. In fact, the sketch-and-project methods share a certain similarity to Krylov methods. This can be seen by using the following equivalent dual formulation of~\eqref{C1eq:sketchproj} given by \begin{equation} x^{k+1} \; = \; \arg\min_{x\in \mathbb{R}^n} \norm{x\phantom{^k}- x^{*}}_B^2 \quad \mbox{subject to} \quad x \in x^{k} + \myRange{B^{-1}A^\top S}. \label{ch:one:RF} \end{equation} We refer to~\eqref{ch:one:RF} as the \emph{constrain and approximate} viewpoint, because a new iterate $x^{k+1}$ is selected as the best possible \emph{approximation} to the solution $x^*$ \emph{constrained} to a randomly generated affine space. Later in Section~\ref{sec:sixviews} we prove that~\eqref{C1eq:sketchproj} and~\eqref{ch:one:RF} are equivalent. Formulation~\eqref{ch:one:RF} is similar to the framework often used to describe Krylov methods~\cite[Chapter 1]{Liesen2014}, which is \begin{equation} \label{eq:krylov} x^{k+1} = \arg\min_{x\in \mathbb{R}^n} \norm{x- x^{*}}_B^2 \quad \mbox{subject to} \quad x \in x^{0} + \mathcal{K}_{k+1}, \end{equation} where $\mathcal{K}_{k+1}\subset \mathbb{R}^n$ is a $(k+1)$--dimensional subspace. Note that the constraint $x \in x^{0}+\mathcal{K}_{k+1}$ is an affine space that passes through $x^{0}$, as opposed to passing through $x^{k}$ in the sketch-and-project formulation~\eqref{ch:one:RF}. The objective $\|x-x^{*}\|^2_B $ is a generalization of the residual, where $B=A^\top A$ is used to characterize minimal residual methods (MINRES and GMRES)~\cite{Paige1975,Saad1986} and $B=A$ is used to describe the Conjugate Gradients (CG) method~\cite{Hestenes1952}. Progress from one iteration to the next is guaranteed by using expanding nested search spaces at each iteration, that is, $\mathcal{K}_k \subset \mathcal{K}_{k+1}.$ \paragraph{Low memory requirements.} To make an efficient Krylov method, the problem~\eqref{eq:krylov} should not be solved from scratch, but rather, one should build upon the knowledge that the previous iterate $x^k$ is the minima restricted to $\mathcal{K}_k$, and calculate $x^{k+1}$ by updating $x^k$ in an inexpensive manner. This inexpensive update is typically achieved through a \emph{short recurrence} update, that is, an update applied to $x^k$ of the following form \[x^{k+1} = x^k + \sum_{i=1}^{s} p^{k+2-i},\] using a small number $s\in \mathbb{N}$ of vectors $p^{k+2-s},\ldots, p^{k+1}$. To arrive at a short recurrence, one needs to carefully design an orthonormal basis for the spaces $\mathcal{K}_k.$ Such a short recurrence does not always exist. By the Faber-Manteuffel Theorem~\cite{faber1984}, the sufficient and necessary conditions for a short recurrence in a Krylov method are that the system matrix is a nonsingular normal matrix with respect to the geometry defined by the $B$ matrix. \footnote{Though one cannot guarantee the existence of a short recurrence for a singular matrix in general, there do exist short recurrences for particular singular matrices that arise from deflation techniques~\cite{Gaul2014}.} In contrast, the sketch-and-project methods automatically have a short recurrence by design, as can be seen through~\eqref{C1eq:xupdate}, where $x^{k+1}$ is calculated by adding a single random vector to $x^k$ (a very short recurrence). Instead of using ``growing'' search spaces to guarantee convergence, the sketch-and-project methods are guaranteed to converge when the distribution of $S$ is such that the search space in~\eqref{ch:one:RF} ``covers'' the space of interest with a nonzero probability\footnote{As shown later in Chapter~\ref{ch:SDA}, this condition is captured in the requirement that $H$~\eqref{eq:Hintro} be nonsingular.}. Not only can we guarantee convergence on even singular matrices but, our lower bounds indicate that convergence improves for lower rank matrices. \paragraph{Existence of convergence analysis.} As for convergence rates, only certain instantiations of the Krylov methods, such as the CG, MINRES and GMRES methods have well understood convergence rates. These three aforementioned methods are only applicable when the system matrix $A$ is symmetric positive definite, symmetric and nonsingular, respectively. Again in contrast, the sketch-and-project methods converge for virtually any matrix $A$ with an explicit convergence rate, as we show in Chapter~\ref{ch:SDA}. \paragraph{Improved convergence rate.} As an example of how the rate of convergence of a Krylov method compares against the rate of convergence of a sketch-and-project type method, let us consider the setting where $A$ is positive definite. In this setting the CG method (a Krylov method) and the Coordinate Descent (\emph{CD}) method (a sketch-and-project method) are the methods of choice. The iteration complexity of the CG method is $O(\sqrt{\lambda_{\max}(A)/\lambda_{\min}(A)})$ while the iteration complexity of the CD method is $O(\Tr{A}/\lambda_{\min}(A))$ (proof in Section~\ref{C2sec:shs98ss}). Thus the CD method requires at least the square of the number of iterations that the CG method requires to reach the same relative accuracy (ignoring the expectation). But this iteration complexity needs to be counter balanced with the very low iteration cost of the CD method. That is, an iteration of the CD method costs $O(n)$ while an iteration of the CG method costs $O(n^2)$ (ignoring possible gains in efficiency due to sparsity for both methods). This is a significant difference in iteration cost when solving large dimensional linear systems. Strohmer and Vershynin~\cite{Strohmer2009} give examples of when this lower cost of the CD method\footnote{In the paper~\cite{Strohmer2009} the authors compare the iteration complexity and iteration cost of the randomized Kaczmarz method against the Conjugate Gradients method applied to the least-squares problem. But their comparisons hold verbatim for the CD method and the CG method applied to a positive definite linear system.} pays off for the higher iteration complexity. Furthermore, accelerated versions of the CD method~\cite{Lee2013,LiuWright-AccKacz-2016,Wright:ABCRRO,Fercoq2013a} improve the iteration complexity of the CD method to $O(\sqrt{\Tr{A}/\lambda_{\min}(A)})$ at the cost of only a constant increase in the iteration cost. Thus when taking iteration cost into account, the accelerated CD methods asymptomatically achieve a solution with higher accuracy than the the CG method for the same computational cost. We do not consider accelerated variants of the CD method, or the sketch-and-project method, in this thesis. For further insight into how accelerated sketch-and-project methods compare with the CG method, see ~\cite{LiuWright-AccKacz-2016} for a comparison to an accelerated randomized Kaczmarz method and~\cite{Lee2013} for a comparison to an accelerated CD method. \paragraph{Parallelizable.} Where the sketch-and-project methods truly diverge from Krylov methods is in how parallelizable they are. Though we do not address this in this thesis, methods based on~\eqref{C1eq:sketchproj} are easy to adapt to a parallel architecture. In contrast, adapting the Krylov methods towards parallel computing is challenging. This is, in part, because of the delicate orthogonality conditions that need to be enforced on the basis of the search spaces $\mathcal{K}_k$ in order to guarantee short recurrences and convergence. A few of the current strategies towards adapting the Krylov methods for a parallel implementations are the following. The conjugate gradients method can be paired with domain decomposition methods for solving linear systems that result from discretizing PDE's for an overall successful distributed method~\cite{DongarraHL2016}. But in this strategy the gains in parallelism come from the domain decomposition method and not from the conjugate gradient method \emph{per se}. Efforts towards designing a distributed or communication efficient variant of Krylov methods are focused on two fronts: exploring the parallelism in matrix-vector and vector-vector products performed in the the Krylov methods~\cite{Ballard2014}, and the so called $s$--step Krylov methods. The $s$--step methods, such as the $s$--step conjugate gradients methods~\cite{Chronopoulos1989}, re-order the computations of standard Krylov methods so that $s$ iterations can be performed simultaneously and in parallel. Though implementing fast and reliable $s$--step methods come with two challenges: achieving numerical stability is more challenging than traditional Krylov methods~\cite{Ballard2014} and correctly identifying the communication bottlenecks is difficult as it depends on the nonzero structure of the system matrix~\cite{Ballard2014}. The sketch and project methods, on the other hand, are relatively easy to adapt to a parallel setting. First, the methods~\eqref{C1eq:sketchproj} make use of shared memory parallelism through their \emph{block} variants. By block variants we refer to the setting when $S$ has more than one column. Through experiments and complexity analysis we see that block variants converge much fast than their ``single column'' counterparts. Furthermore, the main computational bottleneck, calculating $S^\top A$, can be completely amortized using multi-thread matrix-matrix products. This feature has been explored in implementing parallel coordinate descent methods~\cite{PCDM}. In a distributed computing setting, the independence of the sampling of the matrix $S$ allows for distributed implementations, which has already been explored again for coordinate descent methods~\cite{Richtarik2013a,NIPSdistributedSDCA,Maymounkov2010}. Thus there are a number of clear advantages in using the sketch-and-project methods, as compared with Krylov methods. This discussion has only been a small window into the advantages of randomized methods for solving linear systems. But randomized methods are having a profound impact on other related fields such as optimization methods, in particular, in minimizing partially separable functions, randomized methods are considered the state-of-the-art due, in part, to their fast convergence. Also in numerical linear algebra, with new randomized methods for solving least squares that, according to the experiments in~\cite{Maymounkov2010}, outperform the long-standing benchmark LAPACK. \section{Tools of the Trade} \label{C1sec:tools} In this section we present several lemmas concerning pseudoinverses and projections. Though these lemmas are elementary in nature we include them for completeness since they are called upon several times throughout the thesis. \subsection{Pseudoinverse}\label{C1subsec:pseudo} The pseudoinverse matrix was introduced by Moore~\cite{Moore1920} and Penrose~\cite{Penrose1955} in their pioneering work, though our exposition and definition follows that of~\cite{Desoer1963}. The following lemma is a standard result in linear algebra required in defining the pseudoinverse and projections. \begin{lemma}\label{lem:WGW}For any matrix $W$ and symmetric positive definite matrix $G$, \begin{equation}\label{eq:8ys98hs}\Null{W} = \Null{W^\top G W}\end{equation} and \begin{equation}\label{eq:8ys98h986ss}\myRange{W^\top } = \myRange{W^\top G W}.\end{equation} \end{lemma} \begin{proof} In order to establish \eqref{eq:8ys98hs}, it suffices to show the inclusion $\Null{W} \supseteq \Null{W^\top G W}$ since the reverse inclusion trivially holds. Letting $s\in \Null{W^\top G W}$, we see that $\|G^{1/2}Ws\|^2=0$, which implies $G^{1/2}Ws=0$. Therefore, $s\in \Null{W}$. Finally, \eqref{eq:8ys98h986ss} follows from \eqref{eq:8ys98hs} by taking orthogonal complements. Indeed, $\myRange{W^\top}$ is the orthogonal complement of $\Null{W}$ and $\myRange{W^\top G W}$ is the orthogonal complement of $\Null{W^\top G W}$. \end{proof} The pseudoinverse is a matrix that shares many properties with the inverse matrix. Given a real matrix $M \in \mathbb{R}^{m\times n},$ when the nullspace of $M$ contains a nonzero vector then $M$, seen as a linear transformation \[M: \mathbb{R}^n \mapsto \myRange{M},\] is not injective and thus $M$ has no inverse. But we can construct a \emph{pseudoinverse} of $M$. The pseudoinverse is constructed by considering a restriction of $M$ that is injective and invertible, and then extending this restriction. Specifically, consider the restriction \[M|_{\myRange{M^\top}}:\myRange{M^\top } \mapsto \myRange{M}. \] Note that this restriction is defined on the orthogonal complement of the nullspace of $M$, and thus removes the ``troublesome'' subspace that prevents $M$ from being invertible. Indeed, this restriction is invertible since $\myRange{MM^\top }=\myRange{M}$ and $\Null{MM^\top }=\Null{M^\top }$ by Lemma~\ref{lem:WGW}, thus the restriction is surjective and injective. The following extension of the inverse of this restriction is what we call the pseudoinverse, see Definition~\ref{def:pseudoinverse} \begin{definition} \label{def:pseudoinverse} Let $M\in \mathbb{R}^{m\times n}$ be any real matrix. $M^{\dagger}\in \mathbb{R}^{n\times m}$ is said to be the pseudoinverse if \begin{enumerate}[i)] \item $M^{\dagger} Mx = x $ for all $x \in \myRange{M^\top }.$ \item $M^{\dagger} x = 0$ for all $x \in \Null{M^\top }.$ \end{enumerate} \end{definition} Item $i)$ defines $M^{\dagger}$ on \[ \myRange{MM^\top } \overset{\eqref{eq:8ys98h986ss}}=\myRange{M},\] and item $ii)$ defines $M^{\dagger}$ on $\Null{M^\top }$. Thus the two items together define $M^{\dagger}$ uniquely over $\myRange{M}\oplus \Null{M^\top } = \mathbb{R}^m.$ For the original, and equivalent, definition of pseudoinverse see~\cite{Moore1920} and Penrose~\cite{Penrose1955}. Alternatively, for a definition of pseudoinverse based on the SVD decomposition see Section 5.2.2 in~\cite{Golub2013}. We now collect the properties of the pseudoinverse that we use through the thesis in a sequence of lemmas. \begin{lemma} \label{lem:pseudo} $ M M^\dagger M = M$ \end{lemma} \begin{proof} Let $z \in \mathbb{R}^n$ and consider the decomposition $z = y +x $ where $y \in \myRange{M^\top }$ and $x \in \Null{M}.$ By item i) of Definition~\ref{def:pseudoinverse} we have \[M M^\dagger M z = M y =M(y+x)=Mz.\] \end{proof} \begin{lemma} \label{lem:pseudosym} If $M$ is symmetric then $M^{\dagger}$ is symmetric. \end{lemma} \begin{proof} Let $z_1,z_2 \in \mathbb{R}^n$ and consider the decompositions $z_1 = M y_1 +x_1 $ and $z_2 = M y_2 +x_2$ where $y_1, y_2 \in \myRange{M}$ and $x_1,x_2 \in \Null{M},$ which by the symmetry of $M$ always exist. It follows that \begin{eqnarray*} z_1^\top (M^{\dagger})^\top z_2 &=& (M^{\dagger}z_1)^\top z_2\\ &\overset{\text{Definition~\ref{def:pseudoinverse} item ii)}}{=}& (M^{\dagger}My_1)^\top z_2\\ &\overset{\text{Definition~\ref{def:pseudoinverse} item i)}}{=}& y_1^\top z_2\\ & = & y_1^\top My_2. \end{eqnarray*} and \begin{eqnarray*} z_1^\top M^{\dagger} z_2 &\overset{\text{Definition~\ref{def:pseudoinverse} item ii)}}{=}& z_1^\top M^{\dagger}My_2\\ &\overset{\text{Definition~\ref{def:pseudoinverse} item i)}}{=}& z_1^\top y_2\\ & = & y_1^\top My_2, \end{eqnarray*} thus $(M^{\dagger})^\top = M^{\dagger}.$ \end{proof} \begin{lemma} \label{lem:pseudoposdef} If $M$ is symmetric positive semidefinite then $M^{\dagger}$ is symmetric positive semidefinite. \end{lemma} \begin{proof} For any $z \in \mathbb{R}^n$, consider the decomposition $z = M y +x $ where $y \in \myRange{M}$ and $x \in \Null{M},$ which by the symmetry of $M$ always exists. It follows that \[z^\top M^{\dagger} z \overset{\text{Definition~\ref{def:pseudoinverse} item ii)}}{=} y^\top M M^{\dagger}My \overset{\text{Definition~\ref{def:pseudoinverse} item i)}}{=} y^\top My \geq 0, \] which shows that $M^{\dagger}$ is positive semidefinite. The symmetry of $M^{\dagger}$ follows from Lemma~\ref{lem:pseudosym}. \end{proof} \begin{lemma}\label{lem:pseudoproj} The matrix $M^{\dagger}M$ projects orthogonally onto $ \myRange{M^\top }$ and along $\Null{M}.$ \end{lemma} \begin{proof} Consider the orthogonal decomposition $z = y +x$ where $y \in \myRange{M^\top }$ and $x \in \Null{M}.$ Then \[M^{\dagger}M z = M^{\dagger}M y = y,\] where we used item ii) then item i) of Definition~\ref{def:pseudoinverse}. The result now follows by observing that $\myRange{M^\top }$ and $\Null{M}$ are orthogonal complements.\end{proof} \begin{lemma} \label{lem:pseudoleastnorm} Consider the consistent linear system $Mx=d$ where $M,x$ and $d$ are of conforming dimensions. It follows that \begin{equation}\label{eq:pseudoleastnorm}M^{\dagger}d = \arg \min_x \norm{x}_2^2 \quad \mbox{subject to} \quad Mx=d. \end{equation} \end{lemma} \begin{proof} As $d \in \myRange{M}$ we have from Lemma~\ref{lem:pseudo} that $MM^{\dagger}d =d$. Using the change of variables $z = x-M^{\dagger}d$ in~\eqref{eq:pseudoleastnorm} gives \begin{equation}\label{eq:a3fa3fa} z^*\eqdef \arg\min_z \norm{z+M^{\dagger}d}_2^2, \quad \mbox{subject to} \quad z \in \Null{M}. \end{equation} By Lemma~\ref{lem:pseudoproj} we have that $M^{\dagger}d \in \myRange{M^\top } = \Null{M}^{\perp}.$ Consequently \[\norm{z+M^{\dagger}d}_2^2 = \norm{z}_2^2 + \norm{M^{\dagger}d}_2^2 \geq \norm{z}_2^2,\] thus the minimum $z^*$ of~\eqref{eq:a3fa3fa} is achieved at $z^*=0,$ from which it follows that the minimum of~\eqref{eq:pseudoleastnorm} is achieved at $x=z^*+ M^{\dagger}d = M^{\dagger}d.$ \end{proof} \subsection{Projection matrices} The following proposition is a variant of a standard result of linear algebra (which is often presented in the $B=I$ case). While the results are folklore and easy to establish, in the proof of our main theorems we need certain details which are hard to find in textbooks on linear algebra, and hence hard to refer to. For the benefit of the reader, we include the detailed statement and proof. \begin{proposition} [Decomposition and Projection]\label{prop:decomposition} Let $M \in \mathbb{R}^{m\times n}$ by a real matrix and $B \in \mathbb{R}^{n\times n}$ a symmetric positive definite matrix. Each $x\in \mathbb{R}^n$ can be decomposed in a unique way as $x = s(x) + t(x)$, where $s(x)\in \myRange{B^{-1}M^\top}$ and $t(x)\in \Null{M}$. Moreover, the decomposition can be computed explicitly as \begin{equation} \label{eq:98hs8hss}s(x) = \arg \min_{s} \left\{ \|x-s\|_B \;:\; s\in \myRange{B^{-1}M^\top} \right\}= B^{-1} Z_M x \end{equation} and \begin{equation} \label{eq:98hs8htt}t(x) = \arg \min_{t} \left\{ \|x-t\|_B \;:\; t\in \Null{M} \right\}= (I - B^{-1}Z_M) x,\end{equation} where \begin{equation}\label{eq:Z_M}Z_M\eqdef M^\top (MB^{-1}M^\top)^\dagger M.\end{equation} Hence, the matrix $B^{-1}Z_M$ is a projection in the $B$-norm onto $\myRange{B^{-1}M^\top}$, and $I-B^{-1}Z_M$ is a projection in the $B$-norm onto $\Null{M}$. Moreover, for all $x\in \mathbb{R}^n$ we have $\|x\|_B^2 = \|s(x)\|_B^2 + \|t(x)\|_B^2$, with \begin{equation}\label{eq:iuhiuhpp}\|t(x)\|_B^2 = \|(I-B^{-1}Z_{M})x\|_B^2 = x^\top (B-Z_{ M}) x\end{equation} and \begin{equation}\label{eq:iuhiuhppss}\|s(x)\|_B^2 = \|B^{-1}Z_{M} x\|_B^2 = x^\top Z_{M} x.\end{equation} Finally, \begin{equation}\label{eq:ugisug7sss}\Rank{M} = \Tr{B^{-1}Z_M}.\end{equation} \end{proposition} \begin{proof} Fix arbitrary $x\in \mathbb{R}^n$. We first establish existence of the decomposition. By Lemma~\ref{lem:WGW} applied to $W=M^\top$ and $G=B^{-1}$ we know that there exists $u$ such that $Mx = MB^{-1}M^\top u$. Now let $s = B^{-1}M^\top u$ and $t = x-s$. Clearly, $s\in \myRange{B^{-1}M^\top}$ and $t\in \Null{M}$. For uniqueness, consider two decompositions: $x = s_1+ t_1$ and $x=s_2 + t_2$. Let $u_1,u_2$ be vectors such that $s_i = B^{-1}M^\top u_i$, $i=1,2$. Then $MB^{-1}M^\top(u_1-u_2)=0$. Invoking Lemma~\ref{lem:WGW} again, we see that $u_1-u_2\in \Null{M^\top}$, whence $s_1 = B^{-1}M^\top u_1 = B^{-1}M^\top u_2 = s_2$. Therefore, $t_1 = x - s_1 = x-s_2 = t_2$, establishing uniqueness. Note that $s = B^{-1}M^\top y$, where $y$ is any solution of the optimization problem \[\min_y \tfrac{1}{2}\|x-B^{-1}M^\top y\|_B^2.\] The first order necessary and sufficient optimality conditions are $Mx = MB^{-1}M^\top y$. In particular, we may choose $y$ to be the least norm solution of this system, which by Lemma~\ref{lem:pseudoleastnorm} is given $y=(MB^{-1}M^\top)^\dagger Mx$, from which \eqref{eq:98hs8hss} follows. The variational formulation \eqref{eq:98hs8htt} can be established in a similar way, again via first order optimality conditions (note that the closed form formula \eqref{eq:98hs8htt} also directly follows from \eqref{eq:98hs8hss} and the fact that $t = x - s$). Next, since $x=s+t$ and $s^\top B t = 0$, \begin{equation}\label{eq:09u0hss} \|t\|_B^2=(t+s)^\top B t = x^\top B t \overset{\eqref{eq:98hs8htt}}{=} x^\top B (I-B^{-1}Z_{ M})x = x^\top (B - Z_{M}) x \end{equation} and \[ \|s\|_B^2 = \|x\|_B^2 - \|t\|_B^2 \overset{\eqref{eq:09u0hss}}{=} x^\top Z_M x.\] It only remains to establish \eqref{eq:ugisug7sss}. Since $B^{-1}Z_M$ projects onto $\myRange{B^{-1}M^\top}$ and since the trace of a projection is equal to the dimension of the space they project onto, we have $\Tr{B^{-1}Z_M} = \dim(\myRange{B^{-1}M^\top}) = \dim(\myRange{M^\top}) = \Rank{M}$ \end{proof} All the iterative methods presented in the thesis are based on projections. In particular, the projections that govern most of the iterative methods here are constructed from the matrix \begin{equation}\label{eq:Z-C1def} Z \eqdef Z_{S^\top A} \overset{\eqref{eq:Z_M}}{=} A^\top S(S^\top AB^{-1}A^\top S)^{\dagger}S^\top A. \end{equation} We now collect in the following lemma several properties pertaining to $Z$ that are repeatedly used throughout the thesis. \begin{lemma}\label{ch:one:lem:Z} The matrix $Z$ defined in~\eqref{eq:Z-C1def} is symmetric positive semidefinite. Furthermore $B^{-1}Z$ is a projection with respect to the $B$--norm such that \begin{equation}\label{eq:BZprojdef} \myRange{B^{-1} Zx} = \myRange{B^{-1}A^\top S} \quad \text{and} \quad \myRange{I-B^{-1}Z} = \Null{S^\top A}, \end{equation} and $B^{-1/2}ZB^{-1/2}$ is a projection with respect to the standard Euclidean geometry, consequently \begin{equation}\label{eq:B12ZB12proj} \norm{(I-B^{-1/2} Z B^{-1/2})x}_2^2 = \dotprod{(I-B^{-1/2} Z B^{-1/2})x,x}, \quad \forall x \in \mathbb{R}^n, \end{equation} and \begin{equation}\label{eq:B12ZB12trace} \Tr{B^{-1/2}ZB^{-1/2}} = \Rank{A^\top S}. \end{equation} \end{lemma} \begin{proof} First note that $(B^{-1/2}A^\top S)^\top B^{-1/2}A^\top S =S^\top AB^{-1}A^\top S$ is symmetric positive semidefinite, and consequently by Lemma~\ref{lem:pseudoposdef} the matrix $(S^\top AB^{-1}A^\top S)^{\dagger}$ is also symmetric positive semidefinite. Thus there exists $G$ such that $GG^\top =(S^\top AB^{-1}A^\top S)^{\dagger} $ and consequently $ (A^\top S G)( A^\top S G)^\top = Z$ which proves that $Z$ is symmetric positive semidefinite. By Lemma~\ref{prop:decomposition} (with $M=S^\top A$) we have that $B^{-1}Z$ is a projection, and~\eqref{eq:BZprojdef} follows by~\eqref{eq:98hs8hss} and~\eqref{eq:98hs8htt}. Again by Lemma~\ref{prop:decomposition} (with $M=S^\top AB^{-1/2}$ and $B=I$) we have that $B^{-1/2}ZB^{-1/2}$ projects orthogonally onto $\myRange{B^{-1/2}A^\top S}$, whence~\eqref{eq:B12ZB12proj} and~\eqref{eq:B12ZB12trace} follow from~\eqref{eq:iuhiuhpp} and~\eqref{eq:ugisug7sss}, respectively. \end{proof} \subsection{Random variables and the random matrix $S$} As explained in Section~\ref{secChOne:whats}, the methods proposed in this thesis depend on a random matrix $S\in \mathbb{R}^{m \times q}$. Consequently the iterates of these methods are random variables. Throughout the thesis we make little to no assumption on the distribution of $S$, and unless explicitly stated, the reader should assume that $S$ is a random matrix in the most general sense. Here we formalize what is a random variable and what is a random matrix in the most general sense, that is, in the probability measure sense. For the reader that is not familiar with probability spaces and measure theory, we suggest the book~\cite{williams1991probability} as quick an enjoyable introduction. To formalize the notion of a random variable we need the definition of a \emph{probability space.} A probability space $(\Omega,\mathcal{F},P)$ is defined by three objects: \begin{enumerate} \item The $\Omega$ is a given set known as the \emph{sample space}. It contains all the possible~\emph{outcomes} (elements). \item The $\mathcal{F}$ is a set of subsets of $\Omega.$ Specifically, it is a $\sigma$--algebra over $\Omega.$ It contains all the possible \emph{events} (subsets) we would like to consider. \item The $P$ is a function that maps from $\mathcal{F}$ to $[0, \, 1].$ That is, given an event $E \in \mathcal{F}$ it return the probability $P(E) \in [0, \, 1]$ of $E$ occurring. Moreover, $P$ is a probability measure and thus $P(\Omega) =1, P(\emptyset) =0$ and $P$ satisfies the countable additivity property~\cite{williams1991probability}. \end{enumerate} Often one is not interested in the probability space itself, but in functions over this probability space called random variables. Consider the map $r: \Omega \rightarrow \mathbb{R}.$ We say $r$ is a \emph{random variable} when \[ \{\omega \, : \, r(\omega)\leq a\} \subset \mathcal{F} \quad \quad \forall a \in \mathbb{R}.\] An equivalent statement is as follows: The function $r$ is a random variable if the inverse image of $r$ over the interval $(-\infty, a]$ is contained in $\mathcal{F}$ for every $a \in \mathbb{R}.$ A random matrix is simply a matrix valued map where each element is a random variable. That is, consider a map $S: \Omega \rightarrow \mathbb{R}^{m\times q}$ where $m,q\in \mathbb{N}.$ We say that $S$ is a \emph{random matrix} when each element of $S$ is a random variable. For brevity, and as is customary, we use $S \in \mathbb{R}^{m\times q}$ as a shorthand for $S(\omega) \in \mathbb{R}^{m\times q}$ for all $\omega \in \Omega.$ A \emph{random vector} is a random matrix that has only one column or one row. We now provide an example of a random matrix. Note that this example, and in fact all the examples in this thesis, are simple enough as to not require this formal probability measure context. \paragraph{Example:} Let $e_i \in \mathbb{R}^{m}$ be the $i$th coordinate vector. Let $S = e_i$ with probability $1/m$ for all $i \in \{1,\ldots,m\}.$ In other words, $P(S =e_i) = 1/m$ for all $i \in \{1,\ldots,m\}.$ We will now show that $S$ is a random matrix by constructing a suitable probability space. Let $\Omega = \{1,\ldots, m\}$, let $\mathcal{F} = 2^\Omega$ be the power set of $\Omega$ and let $P: \mathcal{F} \rightarrow [0, 1] $ be any probability measure that satisfies $P(\{i\}) = 1/m$ for $i=1,\ldots, m.$ Then the map defined by $S(i) = e_i$ is our desired random matrix. \subsection{Convergence of a random sequence} \label{sec:convrandseq} As the methods presented in the thesis depend on a random matrix $S$ the iterates of our methods are themselves random variables. To guarantee that the iterates converge to the desired solution we need to establish the convergence of a sequence of random variables. Throughout the thesis we use two notions of the convergence of random variables; the convergence of the norm of the expectation and the convergence of the expected norm, which we describe here. Consider a sequence of random matrices $(Y^k)_k$ on $\mathbb{R}^{m\times q}.$ Let $\dotprod{\cdot, \cdot}$ and $\norm{Y}^2 = \dotprod{Y,Y}$ be an inner product and induced norm, respectively. We say that the norm of the expectation of $(Y^k)_k$ converges to zero with rate $\rho \in [0,\, 1)$ if \begin{equation}\label{ch:one:eq:normexpconv} \norm{\E{Y^k}} \leq \rho^k \norm{Y^0}. \end{equation} Furthermore, from~\eqref{ch:one:eq:normexpconv} we see that $\E{Y^k} \rightarrow 0,$ and thus the sequence converges in expectation to zero. We say that the expected norm of $(Y^k)_k$ converges to zero with rate $\rho$ if \begin{equation}\label{ch:one:eq:expnormconv} \E{\norm{Y^k}^2} \leq \rho^k \norm{Y^0}^2. \end{equation} Note that the order of the expectation operator and the norm are now exchanged in relation to~\eqref{ch:one:eq:normexpconv}. The convergence of the expected norm implies the convergence of the norm of expectation, as we prove in Lemma~\ref{lem:convrandvar}. In this lemma we also show that the convergence~\eqref{ch:one:eq:expnormconv} implies \emph{convergence in probability}. We say that $Y^k$ converges in probability to zero if for every $\epsilon>0$ we have that \begin{equation} \label{eq:convinprob} \lim_{k\rightarrow \infty} \mathbb{P}\left(\norm{Y^k}^2 \geq \epsilon \norm{Y^0}^2 \right) =0. \end{equation} \begin{lemma} \label{lem:convrandvar} The convergence of the expected norm~\eqref{ch:one:eq:expnormconv} implies the convergence of the norm of the expectation~\eqref{ch:one:eq:normexpconv} as can be seen through the equality \begin{equation}\label{eq:convequality} \E{\big\| Y^k \big\|^2} = \big\|\E{Y^k}\big\|^2 + \E{\big\| Y^k - \E{Y^k}\big\|^2}.\end{equation} Furthermore, the convergence of the expected norm~\eqref{ch:one:eq:expnormconv} implies convergence in probability. \end{lemma} \begin{proof} Let $(Y^k)_k$ be a sequence of random vectors that converges to zero according to~\eqref{ch:one:eq:expnormconv}. The convergence of the norm of expectation follows from the equality \begin{align} \big\|\E{Y^k}\big\|^2 &= \big\|\E{Y^k}\big\|^2 + \E{\norm{Y^k}^2} - \E{\norm{Y^k}^2} \nonumber \\ &=\E{\norm{Y^k}^2} - \left(\E{\norm{Y^k}^2} -2\E{\dotprod{Y^k,\E{Y^k}}} + \big\|\E{Y^k}\big\|^2\right) \nonumber\\ & = \E{\big\|Y^k \big\|^2} - \E{\big\| Y^k - \E{Y^k}\big\|^2}. \label{C1eq:convcovar} \end{align} Indeed, since $\E{\big\| Y^k - \E{Y^k}\big\|^2} \geq 0$ we have that \[\big\|\E{Y^k}\big\|^2 \leq \E{\big\| Y^k\big\|^2} \overset{\eqref{ch:one:eq:expnormconv}}{\leq} \rho^k \norm{Y^0}^2. \] Consequently the norm of the expected error converges with rate $\sqrt{\rho}.$ Finally, let $\epsilon>0.$ Using Markov's inequality we have \begin{align}\label{eq:probconv} \mathbb{P}(\norm{Y^k}^2 \geq \epsilon \norm{Y^0}^2) &\leq \frac{\E{\norm{Y^k}^2}}{\epsilon \norm{Y^0}^2} \overset{\eqref{ch:one:eq:expnormconv}}{\leq} \frac{\rho^k}{\epsilon}. \end{align} Thus $ \mathbb{P}(\norm{Y^k}^2 \geq \epsilon \norm{Y^0}^2 ) \rightarrow 0$ as $k\rightarrow \infty.$ \end{proof} In every chapter that follows, we will present several convergence results of random sequences. In particular, in Chapter~\ref{ch:linear_systems} we prove the convergence of the expected norm~\eqref{ch:one:eq:expnormconv} and the convergence of the norm of the expectation~\eqref{ch:one:eq:normexpconv} of a sequence of random vectors $Y^k =x^k-x^*$. In Chapter~\ref{ch:inverse} we prove analogous convergence results of a sequence of random matrices $Y^k = X_k-A^{-1}.$ Thus Lemma~\ref{lem:convrandvar} is important as it sheds light on how these two types of convergence~\eqref{ch:one:eq:expnormconv} and~\eqref{ch:one:eq:normexpconv} are related. Convergence according to~\eqref{ch:one:eq:expnormconv} is also commonly referred to as \emph{linear convergence}. This is because, as explained in the next Section~\ref{secChone:itercompl}, the number of iterations required to reach a certain precision grows linearly and proportionally to $\left.1 \right/ (1-\rho).$ Another common synonym to linear convergence, that we sometimes use here, is to say that $Y^k$ converges \emph{exponentially fast} to zero. This is because the expected norm of $Y^k$ decreases according to the exponential function $\rho^k.$ Note that if $Y^k$ is a random vector defined on $\mathbb{R}^m$ with the standard Euclidean inner product then $\E{\left\| Y^k - \E{Y^k}\right\|_2^2} =\sum_{i=1}^m\E{ (Y^k_i - \E{Y^k_i})^2} = \sum_{i=1}^m \mathbf{Var}(Y^k_i)$, where $z_i^k$ denotes the $i$th element of $Y^k.$ Thus, in this case, the equality~\eqref{eq:convequality} also shows that if the norm of expectation converges and the variance of $Y^k_i$ converges to zero for $i=1,\ldots, m$, then the expected norm converges. \subsection{Iteration complexity} \label{secChone:itercompl} Both types of convergence~\eqref{ch:one:eq:normexpconv} and~\eqref{ch:one:eq:expnormconv} can be recast as iteration complexity bounds using the following lemma. With this lemma we can stipulate an lower bound on how many iterations are required to bring the sequence within an $\epsilon>0$ relative distance of its limit point. \begin{lemma} \label{lem:itercomplex} Consider the sequence $(\alpha_k)_k \in \mathbb{R}_+$ of positive scalars that converges to zero according to \begin{equation}\label{eq:alphaconv} \alpha_k \leq \rho^k \, \alpha_0,\end{equation} where $\rho \in [0, 1).$ For a given $1>\epsilon >0$ we have that \begin{equation} \label{C1eq:itercomplex}k\geq \frac{1}{1-\rho} \log\left(\frac{1}{\epsilon}\right) \quad \Rightarrow \quad \alpha_k \leq \epsilon\, \alpha_0. \end{equation} \end{lemma} \begin{proof} First note that if $\rho=0$ the result follows trivially. Assuming $\rho \in (0,\,1)$, rearranging~\eqref{eq:alphaconv} and applying the logarithm to both sides gives \begin{equation} \label{eq:logalphaconv} \log\left(\frac{\alpha_0}{\alpha_k}\right) \geq k \log\left(\frac{1}{\rho}\right). \end{equation} Now using that \begin{equation}\label{eq:logineq} \frac{1}{1-\rho} \log\left(\frac{1}{\rho}\right) \geq 1, \end{equation} for all $\rho \in (0,1)$ and assuming that \begin{equation}\label{eq:kiterassump} k\geq \frac{1}{1-\rho} \log\left(\frac{1}{\epsilon}\right) , \end{equation} we have that \begin{eqnarray*} \log\left(\frac{\alpha_0}{\alpha_k}\right) & \overset{\eqref{eq:logalphaconv}}{\geq}& k \log\left(\frac{1}{\rho}\right) \\ & \overset{\eqref{eq:kiterassump}}{\geq} & \frac{1}{1-\rho} \log\left(\frac{1}{\rho}\right) \log\left(\frac{1}{\epsilon}\right)\\ & \overset{\eqref{eq:logineq}}{\geq} & \log\left(\frac{1}{\epsilon}\right) \end{eqnarray*} Applying exponentials to the above inequality gives~\eqref{C1eq:itercomplex}. \end{proof} As an example of the use this lemma, consider the sequence of random vectors $(Y^k)_k$ for which the expected norm converges to zero according to~\eqref{ch:one:eq:expnormconv}. Then applying Lemma~\ref{lem:itercomplex} with $\alpha_k = \E{\norm{Y^k}^2}$ for a given $1>\epsilon>0$ states that \[k \geq \frac{1}{1-\rho} \log\left(\frac{1}{\epsilon}\right) \quad \Rightarrow \quad \E{\norm{Y^k}^2} \leq \epsilon\, \norm{Y^0}^2.\] To give further insight into the implications of the convergence of the expected norm, for a given $1>\epsilon >0$ consider the sequence $\alpha_0 =1/\epsilon$ and $\alpha^k = \mathbf{P}\left(\norm{Y^k}_2^2 \geq \epsilon \norm{Y^0}_2^2\right)$ for $k \geq 1.$ From~\eqref{eq:probconv} we know that this sequence converges according to $\alpha^k \leq \rho^k \alpha^0.$ We can now use Lemma~\ref{lem:itercomplex} to determine how many iterates are required so that $\norm{Y^k}_2^2 \leq \epsilon \norm{Y^0}_2^2 $ with high probability. Indeed, let $\delta \in (0,\, 1)$ then by Lemma~\ref{lem:itercomplex} we have that \[k \geq \frac{1}{1-\rho} \log\left(\frac{1}{\epsilon\delta}\right) \quad \Rightarrow \quad \mathbf{P}\left(\norm{Y^k}_2^2 \geq \epsilon \norm{Y^0}_2^2\right) \leq \delta.\] This shows that convergence of the expected norm is almost as good as linear convergence without the expectation, that is, one can guarantee the iterates are relatively close with a high probability at the cost of only an additional logarithmic growth in the number of iterates. \chapter[Randomized Iterative Methods for Solving Linear Systems]{Randomized Iterative Methods for Linear Systems} \chaptermark{Randomized Iterative Methods for Linear Systems} \label{ch:linear_systems} { \epigraph{\emph{Guid gear comes in sma’ bulk}. \\ Good things come in small sizes (like sketched linear systems!) }{Scottish proverb.} \let\clearpage\relax \section{Introduction} } The need to solve linear systems of equations is ubiquitous in essentially all quantitative areas of human endeavour, including industry and science. Linear systems are a central problem in numerical linear algebra, and play an important role in computer science, mathematical computing, optimization, signal processing, engineering, numerical analysis, computer vision, machine learning, and many other fields. For instance, in the field of large scale optimization, there is a growing interest in inexact and approximate Newton-type methods for ~\cite{Dembo1982,Eisenstat1994b,Bellavia1998,Zhao2010a,Wang2013,Gondzio2013}, which can benefit from fast subroutines for calculating approximate solutions of linear systems. In machine learning, applications arise for the problem of finding optimal configurations in Gaussian Markov Random Fields \cite{GMRFbook}, in graph-based semi-supervised learning and other graph-Laplacian problems \cite{Bengio+al-ssl-2006}, least-squares SVMs, Gaussian processes and more. In a large scale setting, direct methods can suffer from two shortcomings. First, direct methods often require direct access to individual elements of the system matrix and thus need the system matrix to be stored on RAM. But the dimensions and density of the problem at hand maybe such that the system matrix does not fit on RAM. Second, the complexity of direct methods is of order $O(n^3)$ which can be prohibitively slow when $n$ is large. While classical iterative methods are deterministic, recent breakthroughs suggest that randomization can play a powerful role in the design and analysis of efficient algorithms~\cite{Strohmer2009,Leventhal2010,Needell2010,Drineas2011,Zouzias2013a,Lee2013,Ma2015, Richtarik2015c} which are in many situations competitive or better than existing deterministic methods. In this chapter we develop the sketch-and-project family of randomized methods for solving linear systems that are well suited to quickly calculating approximate solutions. \subsection{Background and related work} The literature on solving linear systems via iterative methods is vast and has a long history~\cite{Kelley1995,Saad2003}. For instance, the Kaczmarz method, in which one cycles through the rows of the system and each iteration is formed by projecting the current point to the hyperplane formed by the active row, dates back to the 30's~\cite{Kaczmarz1937}. The Kaczmarz method is just one example of an array of row-action methods for linear systems (and also, more generally, feasibility and optimization problems) which were studied in the second half of the 20th century~\cite{rowaction1981}. Research into the Kaczmarz method was reignited in 2009 by Strohmer and Vershynin~\cite{Strohmer2009}, who gave a brief and elegant proof that a randomized variant thereof enjoys an exponential error decay (also know as ``linear convergence''). This has triggered much research into developing and analyzing randomized linear solvers. It should be mentioned at this point that the randomized Kaczmarz (RK) method arises as a special case (when one considers quadratic objective functions) of the stochastic gradient descent (SGD) method for {\em convex optimization} which can be traced back to the seminal work of Robbins and Monro's on stochastic approximation~\cite{RobbinsMonro:1951}. Subsequently, intensive research went into studying various extensions of the SGD method. However, to the best of our knowledge, no complexity results with exponential error decay were established prior to the aforementioned work of Strohmer and Vershynin~\cite{Strohmer2009}. This is the reason behind our choice of \cite{Strohmer2009} as the starting point of our discussion. Motivated by the results of Strohmer and Vershynin~\cite{Strohmer2009}, Leventhal and Lewis~\cite{Leventhal2010} utilize similar techniques to establish the first bounds for {\em randomized coordinate descent} methods for solving systems with positive definite matrices, and systems arising from least squares problems~\cite{Leventhal2010}. These bounds are similar to those for the RK method. This development was later picked up by the optimization and machine learning communities, and much progress has been made in generalizing these early results in countless ways to various structured convex optimization problems. For a brief up to date account of the development in this area, we refer the reader to \cite{Fercoq2013a,ALPHA} and the references therein. The RK method and its analysis have been further extended to the least-squares problem~\cite{Needell2010,Zouzias2013a} and the block setting~\cite{Needell2012,Needell2014}. In~\cite{Ma2015} the authors extend the randomized coordinate descent and the RK methods to the problem of solving underdetermined systems. The authors of~\cite{Ma2015,Ramdas2014} analyze side-by-side the randomized coordinate descent and RK method, for least-squares, using a convenient notation in order to point out their similarities. Our work takes the next step, by analyzing these, and many other methods, through a genuinely general analysis. Also in the spirit of unifying the analysis of different methods, in~\cite{Oswald2015} the authors provide a unified analysis of iterative Schwarz methods and Kaczmarz methods. The use of random Gaussian directions as search directions in zero-order (derivative-free) minimization algorithm was recently suggested~\cite{Nesterov2011}. Our Gaussian positive definite and Gaussian least-squares in Sections~\ref{C2sec:gaussA} and~\ref{C2sec:gaussATA}, respectively, are special cases of these zero order methods applied to linear systems. More recently, Gaussian directions have been combined with exact and inexact line-search into a single {\em random pursuit} framework~\cite{Stich2014a}, and further utilized within a randomized variable metric method~\cite{Stich2014,Stich2015}. \section{Contributions and Overview} Given a real matrix $A \in \mathbb{R}^{m \times n}$ and a real vector $b \in \mathbb{R}^m$, in this chapter we consider the linear system \begin{equation}\label{eq:Axb}Ax =b.\end{equation} We shall assume throughout that the system is {\em consistent}: there exists $x^*$ for which $Ax^*=b$. We now comment on the main contribution of this chapter. {\em 1. New method.} We develop a novel, fundamental, and surprisingly simple {\em randomized iterative method} for solving \eqref{eq:Axb}. {\em 2. Six equivalent formulations.} Our method allows for several seemingly different but nevertheless equivalent formulations. First, it can be seen as a {\em sketch-and-project} method, in which the system \eqref{eq:Axb} is replaced by its {\em random sketch}, and then the current iterate is projected onto the solution space of the sketched system. We can also view it as a {\em constrain-and-approximate} method, where we constrain the next iterate to live in a particular random affine space passing through the current iterate, and then pick the point from this subspace which best approximates the optimal solution. Third, the method can be seen as an iterative solution of a sequence of random (and simpler) linear equations. The method also allows for a simple geometrical interpretation: the new iterate is defined as the unique intersection of two random affine spaces which are orthogonal complements. The fifth viewpoint gives a closed form formula for the {\em random update} which needs to be applied to the current iterate in order to arrive at the new one. Finally, the method can be seen as a {\em random fixed point iteration.} {\em 3. Special cases.} These multiple viewpoints enrich our interpretation of the method, and enable us to draw previously unknown links between several existing algorithms. Our algorithm has two parameters, an $n\times n$ positive definite matrix $B$ defining geometry of the space, and a random matrix $S$. Through combinations of these two parameters, in special cases our method recovers several well known algorithms. For instance, we recover the randomized Kaczmarz method of Strohmer and Vershyinin~\cite{Strohmer2009}, randomized coordinate descent method of Leventhal and Lewis~\cite{Leventhal2010}, random pursuit~\cite{Nesterov2011,Stich2015,Stich2014,Stich2012} (with exact line search), and the stochastic Newton method recently proposed by Qu et al \cite{Qu2015}. However, our method is more general, and leads to i) various generalizations and improvements of the aforementioned methods (e.g., block setup, importance sampling), and ii) completely new methods. Randomness enters our framework in a very general form, which allows us to obtain a {\em Gaussian Kaczmarz method}, {\em Gaussian descent}, and more. {\em 4. Complexity: general results.} When $A$ has full column rank, our framework allows us to determine the complexity of these methods using a single analysis. Our main results are summarized in Table~\ref{ch:two:tab:complexity}, where $\{x^k\}$ are the iterates of our method, $Z$ is a random matrix dependent on the data matrix $A$, parameter matrix $B\in \mathbb{R}^{n\times n}$ and random parameter matrix $S \in \mathbb{R}^{m\times q}$, defined as \begin{equation}\label{eq:Z-first} Z \eqdef A^\top S(S^\top AB^{-1}A^\top S)^{\dagger}S^\top A, \end{equation} where $\dagger$ denotes the (Moore-Penrose) pseudoinverse. For the definition of pseudoinverse, see Section~\ref{C1subsec:pseudo}. Moreover, $\norm{x}_B \eqdef \sqrt{\dotprod{x,x}_B}$, where $\dotprod{x,y}_B \eqdef x^\top By$, for all $x,y \in \mathbb{R}^{n}$. As we shall see later, we will often consider setting $B=I$, $B=A$ (if $A$ is positive definite) or $B=A^\top A$ (if $A$ is of full column rank). In particular, we first show that the convergence rate $\rho$ is always bounded between zero and one. We also show that as soon as $\E{Z}$ is invertible (which can only happen if $A$ has full column rank, which then implies that $x^*$ is unique), we have $\rho<1$, and the method converges. Besides establishing a bound involving the {\em expected norm of the error} (see the last line of Table~\ref{ch:two:tab:complexity}), we also obtain bounds involving the {\em norm of the expected error} (second line of Table~\ref{ch:two:tab:complexity}). Studying the expected sequence of iterates directly is very fruitful, as it allows us to establish an {\em exact characterization} of the evolution of the expected iterates (see the first line of Table~\ref{ch:two:tab:complexity}) through a {\em linear fixed point iteration}. Both of these theorems on the convergence of the method can be recast as iteration complexity bounds by using Lemma~\ref{lem:itercomplex}. For instance from Theorem~\ref{ch:two:theo:normEconv} in Table~\ref{ch:two:tab:complexity} we observe that for a given $\epsilon >0$ we have that \begin{equation} \label{ch:two:eq:itercomplex}k \geq \frac{1}{1-\rho} \log\left(\frac{1}{\epsilon}\right) \quad \Rightarrow \quad \norm{\E{x^k-x^*}}_B \leq \epsilon \norm{x^0-x^*}_B. \end{equation} {\em 5. Complexity: special cases.} Besides these generic results, which hold without any major restriction on the sampling matrix $S$ (in particular, it can be either discrete or continuous), we give a specialized result applicable to discrete sampling matrices $S$ (see Theorem~\ref{theo:convsingleS}). In the special cases for which rates are known, our analysis recovers the existing rates. \begin{table} \centering \begin{tabular}{|c|c|} \hline & \\ $\E {x^{k+1} -x^{*}} = \left(I - B^{-1}\E{Z}\right) \E{x^{k} - x^{*}} $ & Theorem~\ref{ch:two:theo:normEconv}\\ & \\ $\norm{\E {x^{k+1} -x^{*}}}_B \leq \rho \; \cdot \; \norm{\E{x^{k} - x^{*}}}_B $ & Theorem~\ref{ch:two:theo:normEconv}\\ & \\ $ \E {\norm{x^{k+1} -x^{*} }_B^2 } \leq \rho \;\cdot\; \E{ \norm{x^{k} - x^{*}}_B^2}$ & Theorem~\ref{ch:two:theo:Enormconv}\\ & \\ \hline \end{tabular} \caption{Our main complexity results. The convergence rate is: $\rho = 1- \lambda_{\min}(B^{-1/2}\E{Z}B^{-1/2}).$} \label{ch:two:tab:complexity} \end{table} {\em 6. Extensions.} Our approach opens up many avenues for further development and research. For instance, it is possible to extend the results to the case when $A$ is not necessarily of full column rank, which we do in Chapter~\ref{ch:SDA}. Furthermore, as our results hold for a wide range of distributions, new and efficient variants of the general method can be designed for problems of specific structure by fine-tuning the stochasticity to the structure. Similar ideas can be applied to design randomized iterative algorithms for finding the inverse of a very large matrix, which is the focus of Chapter~\ref{ch:inverse}. \section{One Algorithm in Six Disguises} \label{sec:sixviews}Our method has {\em two parameters}: i) an $n\times n$ positive definite matrix $B$ which is used to define the $B$-inner product and the induced $B$-norm by \begin{equation} \label{eq:B-innerprod}\langle x, y\rangle_{B}\eqdef \langle Bx,y\rangle, \qquad \|x\|_B \eqdef \sqrt{ \langle x, x \rangle_B},\end{equation} where $\langle \cdot,\cdot \rangle$ is the standard Euclidean inner product, and ii) a random matrix $S\in \mathbb{R}^{m\times q}$, to be drawn in an i.i.d.\ fashion at each iteration. We stress that we do not restrict the number of columns of $S$; indeed, we even allow $q$ to vary (and hence, $q$ is a random variable). \subsection{Six viewpoints} Starting from $x^k\in\mathbb{R}^n$, our method draws a random matrix $S$ and uses it to generate a new point $x^{k+1}\in\mathbb{R}^n$. As proven at the end of this section, our iterative method can be formulated in {\em six seemingly different but equivalent ways:} \paragraph{1. Sketching Viewpoint: Sketch-and-Project.} $x^{k+1}$ is the nearest point to $x^k$ which solves a {\em sketched} version of the original linear system: \begin{equation} \boxed{\quad x^{k+1} \quad = \quad \arg\min_{x\in \mathbb{R}^n} \norm{x- x^{k}}_B^2 \quad \mbox{subject to} \quad S^\top Ax = S^\top b \quad} \label{ch:two:NF} \end{equation} This viewpoint arises very naturally. Indeed, since the original system \eqref{eq:Axb} is assumed to be complicated, we replace it by a simpler system---a {\em random sketch} of the original system \eqref{eq:Axb}---whose solution set $\{x \;|\; S^\top Ax = S^\top b\}$ contains all solutions of the original system. However, this system will typically have many solutions, so in order to define a method, we need a way to select one of them. The idea is to try to preserve as much of the information learned so far as possible, as condensed in the current point $x^k$. Hence, we pick the solution which is closest to $x^k$. \paragraph{2. Optimization Viewpoint: Constrain-and-Approximate.} $x^{k+1}$ is the best approximation of $x^*$ in a random space passing through $x^k$: \begin{equation}\boxed{\; x^{k+1} \; = \; \arg\min_{x\in \mathbb{R}^n} \norm{x\phantom{^k}- x^{*}}_B^2 \quad \mbox{subject to} \quad x = x^{k} + B^{-1}A^\top S y, \quad y \;\text{is free} \;} \label{ch:two:RF} \end{equation} The above step has the following interpretation. We choose a random affine space containing $x^k$, and constrain our method to choose the next iterate from this space. We then do as well as we can on this space; that is, we pick $x^{k+1}$ as the point which best approximates $x^*$. Note that $x^{k+1}$ does not depend on which solution $x^*$ is used in \eqref{ch:two:RF} (this can be best seen by considering the geometric viewpoint, discussed next). \begin{figure}[!h] \centering \includegraphics[width =7cm]{ortho_proof_standalone.pdf} \caption{\footnotesize The geometry of our algorithm. The next iterate, $x^{k+1}$, arises as the intersection of two random affine spaces: $x^k + \myRange{B^{-1}A^\top S}$ and $x^* + \Null{S^\top A}$ (see \eqref{eq:geometry}). The spaces are orthogonal complements of each other with respect to the $B$-inner product, and hence $x^{k+1}$ can equivalently be written as the projection, in the $B$-norm, of $x^k$ onto $x^* +\Null{S^\top A}$ (see \eqref{ch:two:NF}), or the projection of $x^*$ onto $x^k +\myRange{B^{-1}A^\top S}$ (see \eqref{ch:two:RF}). The intersection $x^{k+1}$ can also be expressed as the solution of a system of linear equations (see \eqref{eq:2systems}). Finally, the new error $x^{k+1}-x^*$ is the projection, with respect to the $B$-inner product, of the current error $x^k-x^*$ onto $\Null{S^\top A}$. This gives rise to a random fixed point formulation (see \eqref{eq:xZupdate}).} \label{ch:two:fig:proj} \end{figure} \paragraph{3. Geometric viewpoint: Random Intersect.} $x^{k+1}$ is the (unique) intersection of two affine spaces: \begin{equation} \label{eq:geometry} \boxed{ \quad \{x^{k+1}\} \quad =\quad \left( x^* + \Null{S^\top A}\right) \quad \bigcap \quad \left(x^k + \myRange{B^{-1}A^\top S}\right) \quad} \end{equation} First, note that the first affine space above does not depend on the choice of $x^*$ from the set of optimal solutions of \eqref{eq:Axb}. A basic result of linear algebra says that the nullspace of an arbitrary matrix is the orthogonal complement of the range space of its transpose. Hence, whenever we have $h\in \Null{S^\top A}$ and $y\in \mathbb{R}^q$, where $q$ is the number of rows of $S$, then $\dotprod{h, B^{-1}A^\top S y}_B =\dotprod{h, A^\top S y} = 0$. It follows that the two spaces in \eqref{eq:geometry} are orthogonal complements with respect to the $B$-inner product and as such, they intersect at a unique point (see Figure~\ref{ch:two:fig:proj}). \paragraph{4. Algebraic viewpoint: Random Linear Solve.} Note that $x^{k+1}$ is the (unique) solution (in $x$) of a linear system (with variables $x$ and $y$): \begin{equation}\label{eq:2systems} \boxed{\quad x^{k+1} \quad = \quad \text{solution of }\quad S^\top A x = S^\top b, \quad x = x^k + B^{-1}A^\top S y\quad } \end{equation} This system is clearly equivalent to \eqref{eq:geometry}, and can alternatively be written as: \begin{equation}\label{eq:view-algebraic}\begin{pmatrix}S^\top A & 0 \\ B & -A^\top S\end{pmatrix} \begin{pmatrix}x \\y\end{pmatrix} = \begin{pmatrix}S^\top b \\ B x^k\end{pmatrix}.\end{equation} Hence, our method reduces the solution of the (complicated) linear system \eqref{eq:Axb} into a sequence of (hopefully simpler) random systems of the form \eqref{eq:view-algebraic}. \paragraph{5. Algebraic viewpoint: Random Update.} By plugging the second equation in \eqref{eq:2systems} into the first, we eliminate $x$ and obtain the system $(S^\top A B^{-1}A^\top S) y = S^\top (b-Ax^k)$. Note that for all solutions $y$ of this system we must have $x^{k+1} = x^k + B^{-1}A^\top S y$. In particular, we can choose the solution $y=y^k$ of minimal Euclidean norm, which by Lemma~\ref{lem:pseudoleastnorm} is given by $y^k = (S^\top A B^{-1}A^\top S)^{\dagger}S^\top (b-Ax^k)$. This leads to an expression for $x^{k+1}$ with an explicit form of the {\em random update} which must be applied to $x^k$ in order to obtain $x^{k+1}$: \begin{equation}\label{eq:MP} \boxed{\quad x^{k+1} = x^k - B^{-1}A^\top S(S^\top A B^{-1}A^\top S)^{\dagger}S^\top (Ax^k-b) \quad }\end{equation} In some sense, this form is the standard: it is customary for iterative techniques to be written in the form $x^{k+1}=x^k + d^k$, which is precisely what \eqref{eq:MP} does. \paragraph{6. Analytic viewpoint: Random Fixed Point.} Note that iteration~\eqref{eq:MP} can be written as \begin{equation}\label{eq:xZupdate} \boxed{\quad x^{k+1} - x^* \quad =\quad (I- B^{-1}Z)(x^{k} -x^*)\quad} \end{equation} where $Z$ is defined in \eqref{eq:Z-first} and where we used the fact that $Ax^{*} =b$. Matrix $Z$ plays a central role in our analysis, and can be used to construct explicit projection matrices of the two projections depicted in Figure~\ref{ch:two:fig:proj}. The equivalence between these six viewpoints is formally captured in the next statement. \begin{theorem}[Equivalence]\label{thm:equivalence} The six viewpoints are equivalent: they all produce the same (unique) point $x^{k+1}$. \end{theorem} \begin{proof} The proof is simple, and follows directly from the above discussion. In particular, see the caption of Figure~\ref{ch:two:fig:proj}. \end{proof} \subsection{Projection matrices} The explicit projection matrices of the projections depicted in Figure~\ref{ch:two:fig:proj} can be constructed using the $Z$ matrix. Indeed, recall that $S$ is a $m\times q$ random matrix (with $q$ possibly being random), and that $A$ is an $m\times n$ matrix. Let us define the random quantity \begin{equation}\dd \eqdef \Rank{S^\top A}\end{equation} and notice that $d\leq \min\{q,n\}$, \begin{equation} \label{eq:dimproj} \dim \left(\myRange{B^{-1}A^\top S}\right) = \dd, \qquad \text{and}\qquad \dim\left(\Null{S^\top A}\right) = n-\dd. \end{equation} Recall that~\eqref{eq:BZprojdef} shows that $B^{-1}Z$ is a projection onto $\myRange{B^{-1}A^\top S}$ and along $\Null{S^\top A}.$ This sheds additional light on Figure~\ref{ch:two:fig:proj} as it gives explicit expressions for the associated projection matrices. This also shows that $I-B^{-1}Z$ is a projection and thus implies that $I-B^{-1}Z$ is a {\em contraction} with respect to the $B$-norm, which means that the random fixed point iteration~\eqref{eq:xZupdate} has only very little room not to work. While $I-B^{-1}Z$ is not a strict contraction, under some reasonably weak assumptions on $S$ it will be a strict contraction in expectation, which ensures convergence. We shall state these assumptions and develop the associated convergence theory for our method in Section~\ref{C2sec:convergence} and Section~\ref{C2sec:discrete}. \section{Special Cases: Examples} \label{C2sec:examples} In this section we briefly mention how by selecting the parameters $S$ and $B$ of our method we recover several existing methods. The list is by no means comprehensive and merely serves the purpose of an illustration of the flexibility of our algorithm. All the associated complexity results we present in this section, can be recovered from Theorem~\ref{theo:convsingleS}, presented later in Section~\ref{C2sec:discrete}. \subsection{The one step method} When $S$ is an $m\times m$ invertible matrix with probability one, then the system $S^\top Ax=S^\top b$ is equivalent to solving $Ax=b,$ thus the solution to~\eqref{ch:two:NF} must be $x^{k+1}=x^*$, independently of matrix $B.$ Our convergence theorems also predict this one step behaviour, since $\rho =0$ (see Table~\ref{ch:two:tab:complexity}). \subsection{Random vector sketch} When $S = s \in \mathbb{R}^{m}$ is restricted to being a random column vector, then from~\eqref{eq:MP} a step of our method is given by \begin{equation}\label{eq:vecsketch}x^{k+1} = x^{k} - \frac{s^\top (A x^{k}-b)}{ s^\top AB^{-1}A^\top s} B^{-1}A^\top s, \end{equation} if $A^\top s \neq 0$ and $x^{k+1} =x^k$ otherwise. This is because the pseudoinverse of a scalar $\alpha \in \mathbb{R}$ is given by \[\alpha^{\dagger} = \begin{cases} 1/\alpha & \mbox{if } \alpha \neq 0\\ 0 & \mbox{if } \alpha = 0. \end{cases} \] Next we describe several well known specializations of the random vector sketch and for brevity, we write the updates in the form of~\eqref{eq:vecsketch} and leave implicit that when the denominator is zero, no step is taken. \subsection{Randomized Kaczmarz} \label{ch:two:sec:RK} If we choose $S=e^i$ (unit coordinate vector in $\mathbb{R}^m$) and $B=I$ (the identity matrix), in view of \eqref{ch:two:NF} we obtain the method: \begin{equation} \label{eq:RKintro} x^{k+1} = \arg\min_{x\in \mathbb{R}^n} \norm{x- x^{k}}_2^2 \quad \mbox{ subject to } \quad A_{i:}x =b_{i}. \end{equation} Using~\eqref{eq:MP}, these iterations can be calculated with \begin{equation}\label{eq:RKiterate} \boxed{x^{k+1} = x^{k} - \frac{A_{i:} x^{k}-b_{i}}{\norm{A_{i:}}_2^2}(A_{i:})^\top } \end{equation} \paragraph{Complexity.} When $i$ is selected at random, this is the randomized Kaczmarz (\emph{RK}) method~\cite{Strohmer2009}. A specific non-uniform probability distribution for $S$ yields simple and easily interpretable (but not necessarily optimal) complexity bound. In particular, by selecting $i$ with probability proportional to the magnitude of row $i$ of $A$, that is $p_i = \norm{ A_{i:} }_2^2/\norm{A}_F^2$, it follows from Theorem~\ref{theo:convsingleS} that RK enjoys the following complexity bound: \begin{equation}\label{eq:RKconv}\E {\norm{x^{k} -x^{*} }_2^2 } \leq \left(1 - \frac{\lambda_{\min}\left(A^\top A \right)}{\norm{A}_F^2} \right)^k \norm{x^{0} - x^{*}}_2^2.\end{equation} This result was first established by Strohmer and Vershynin \cite{Strohmer2009}. We also provide new convergence results in Theorem~\ref{ch:two:theo:normEconv}, based on the convergence of the norm of the expected error. Theorem~\ref{ch:two:theo:normEconv} applied to the RK method gives \begin{equation}\label{eq:RKconv2}\norm{\E {x^{k} -x^{*} } }_2^2 \leq \left(1 - \frac{\lambda_{\min}\left(A^\top A \right)}{\norm{A}_F^2} \right)^{2k} \norm{x^{0} - x^{*}}_2^2.\end{equation} Now the convergence rate appears squared, which is a better rate, though, the expectation has moved inside the norm, which is a weaker form of convergence as proven in Lemma~\ref{lem:convrandvar}. Analogous results for the convergence of the norm of the expected error holds for all the methods we present, though we only illustrate this with the RK method. \paragraph{Re-interpretation as SGD with exact line search.} Using the ``Constrain and Approximate'' formulation~\eqref{ch:two:RF}, randomized Kaczmarz method can also be written as \[x^{k+1} =\arg \min_{x\in \mathbb{R}^n} \norm{x- x^*}_2^2 \quad \mbox{ subject to } \quad x = x^k + y(A_{i:})^\top , \quad y \in \mathbb{R}, \] with probability $p_i$. Writing the least squares function $f(x) = \tfrac{1}{2}\|Ax-b\|_2^2$ as \[f(x) = \sum_{i=1}^m p_i f_i(x), \qquad f_i(x) = \frac{1}{2 p_i}(A_{i:}x - b_i)^2,\] we see that the random vector $\nabla f_i(x) = \tfrac{1}{p_i}(A_{i:}x-b_i) (A_{i:})^\top $ is an unbiased estimator of the gradient of $f$ at $x$. That is, $\E{\nabla f_i(x)} = \nabla f(x)$. Notice that RK takes a step in the direction $-\nabla f_i(x)$. This is true even when $A_{i:}x-b_i = 0$, in which case, the RK does not take any step. Hence, RK takes a step in the direction of the negative stochastic gradient. This means that it is equivalent to the Stochastic Gradient Descent (SGD) method. However, the stepsize choice is very special: RK chooses the stepsize which leads to the point which is closest to $x^*$ in the Euclidean norm. Later in Section~\ref{subsec:RKvsRCA} in Chapter~\ref{ch:SDA} we give yet another interpretation of the RK method, namely, that the RK method is the equivalent to applying the randomized coordinate descent method to the dual of the least-norm problem. \subsection{Randomized Coordinate Descent: positive definite case} \label{C2sec:shs98ss} If $A$ is symmetric positive definite, then we can choose $B= A$ and $S = e^i$ in~\eqref{ch:two:NF}, which results in \begin{equation} \label{eq:CDpdintro} x^{k+1} \eqdef \arg\min_{x\in \mathbb{R}^n} \norm{x- x^{k}}_A^2 \quad \mbox{subject to} \quad (A_{i:})^\top x =b_{i}, \end{equation} where we used the symmetry of $A$ to get $(e^i)^\top A = A_{i:}=(A_{:i})^\top .$ The solution to the above, given by~\eqref{eq:MP}, is \begin{equation}\label{eq:09j0s9jsss} \boxed{x^{k+1} = x^{k} - \frac{(A_{i:})^\top x^{k}-b_i}{A_{ii}}e^{i}}\end{equation} \paragraph{Complexity.} When $i$ is chosen randomly, this is the \emph{Randomized CD} method (CD-pd). Applying Theorem~\ref{theo:convsingleS}, we see the probability distribution $p_i = A_{ii}/\Tr{A}$ results in a convergence with \begin{equation}\label{eq:CDposconv} \E {\norm{x^{k} -x^{*} }_{A}^2 } \leq \left(1- \frac{\lambda_{\min}\left(A \right)}{\Tr{A}}\right)^k \norm{x^{0} - x^{*}}_{A}^2.\end{equation} This result was first established by Leventhal and Lewis~\cite{Leventhal2010}. \paragraph{Interpretation.} Using the Constrain-and-Approximate formulation~\eqref{ch:two:RF}, this method can be interpreted as \begin{equation} \label{eq:CDpdRF} x^{k+1} =\arg \min \|x-x^*\|_A^2\quad \mbox{ subject to } \quad x = x^k + y e^i, \quad y \in \mathbb{R}, \end{equation} with probability $p_i$. Using the identity $Ax^* = b$, it is easy to check that the function $f(x)=\tfrac{1}{2}x^\top A x - b^\top x$ satisfies: $\|x-x^*\|_A^2 = 2f(x) + b^\top x^*$. Therefore, \eqref{eq:CDpdRF} is equivalent to \begin{equation} \label{eq:CDpdRFxx} x^{k+1} =\arg \min f(x) \quad \mbox{ subject to } \quad x = x^k + y e^i, \quad y\in \mathbb{R}. \end{equation} The iterates~\eqref{eq:09j0s9jsss} can also be written as \[x^{k+1} = x^k - \frac{1}{L_i}\nabla_{i} f(x^k) e^i,\] where $L_i = A_{ii}$ is the Lipschitz constant of the gradient of $f$ corresponding to coordinate $i$ and $\nabla_{i} f(x^k)$ is the $i$th partial derivative of $f$ at $x^k$. \subsection{Randomized block Kaczmarz} Our framework also extends to new block formulations of the randomized Kaczmarz method. Let $R$ be a random subset of $[m]$ and let $S =I_{:R}$ be a column concatenation of the columns of the $m\times m$ identity matrix $I$ indexed by $R$. Further, let $B=I$. Then \eqref{ch:two:NF} specializes to \[ x^{k+1} = \arg\min_{x\in \mathbb{R}^n} \norm{x- x^{k}}_2^2 \quad \mbox{subject to} \quad A_{R:}x = b_{R} . \] In view of~\eqref{eq:MP}, this can be equivalently written as \begin{equation}\label{eq:blockRK}\boxed{x^{k+1}=x^k - (A_{R:})^\top (A_{R:} (A_{R:})^\top )^{\dagger}(A_{R:}x^k - b_{R})} \end{equation} \paragraph{Complexity.} From Theorem~\ref{ch:two:theo:Enormconv} we obtain the following new complexity result: \[\E{\|x^{k}-x^*\|_2^2} \leq \left(1-\lambda_{\min}\left(\E{(A_{R:})^\top (A_{R:} (A_{R:})^\top )^{\dagger}A_{R:}}\right)\right)^k\|x^0-x^*\|^2_2.\] To obtain a more meaningful convergence rate, we would need to bound the smallest eigenvalue of $\E{(A_{R:})^\top (A_{R:} (A_{R:})^\top )^{\dagger}A_{R:}}.$ This has been done in~\cite{Needell2012,Needell2014} when the image of $R$ defines a row paving of $A$. Our framework paves the way for analysing the convergence of new block methods for a large set of possible random subsets $R,$ including, for example, overlapping partitions. \subsection{Randomized Newton: positive definite case} If $A$ is symmetric positive definite, then we can choose $B= A$ and $S = I_{:C}$, a column concatenation of the columns of $I$ indexed by $C$, which is a random subset of $\{1,\ldots, n\}$. In view of~\eqref{ch:two:NF}, this results in \begin{equation} \label{eq:CDpdblock} x^{k+1} \eqdef \arg\min_{x\in \mathbb{R}^n} \norm{x- x^{k}}_A^2 \quad \mbox{subject to} \quad (A_{:C})^\top x =b_{C}. \end{equation} In view of \eqref{eq:MP}, we can equivalently write the method as \begin{equation} \boxed{ \quad x^{k+1} \quad = \quad x^k - I_{:C} ((I_{:C})^\top A I_{:C})^{-1} (I_{:C})^\top (Ax^k - b) \quad }\label{eq:Method1} \end{equation} \paragraph{Complexity.} Clearly, iteration \eqref{eq:Method1} is well defined as long as $C$ is nonempty with probability 1. Such $C$ is referred to in~\cite{Qu2015} as a ``non-vacuous'' sampling. From Theorem \ref{ch:two:theo:Enormconv} we obtain the following convergence rate: \begin{align}\label{eq:Method1xxx} \E { \norm{x^{k} -x^{*}}_{A}^2 } & \leq \rho^k \|x^0 - x^*\|_A^2 \nonumber \\ &= \left(1- \lambda_{\min}\left( \E{ I_{:C} ((I_{:C})^\top A I_{:C})^{-1} (I_{:C})^\top A} \right)\right)^k \norm{x^{0} - x^{*}}_{A}^2. \end{align} The convergence rate of this particular method was first established and studied in \cite{Qu2015}. Moreover, it was shown in \cite{Qu2015} that $\rho<1$ if one additionally assumes that the probability that $i \in C$ is positive for each column $i\in \{1, \ldots, n\}$, i.e., that $C$ is a ``proper'' sampling. \paragraph{Interpretation.} Using formulation~\eqref{ch:two:RF}, and in view of the equivalence between $f(x)$ and $\|x-x^*\|_A^2$ discussed in Section~\ref{C2sec:shs98ss}, the Randomized Newton method can be equivalently written as \[x^{k+1} =\arg \min_{x\in \mathbb{R}^n} f(x) \quad \mbox{ subject to } \quad x = x^k + I_{:C}\, y, \quad y \in \mathbb{R}^{|C|}. \] The next iterate is determined by advancing from the previous iterate over a subset of coordinates such that $f$ is minimized. Hence, an exact line search is performed in a random $|C|$ dimensional subspace. Method~\eqref{eq:Method1} was first studied by Qu et al~\cite{Qu2015}, and referred therein as ``Method~1'', or {\em Randomized Newton Method}. The name comes from the observation that the method inverts random principal submatrices of $A$ and that in the special case when $C=\{1, \ldots, n\}$ with probability 1, it specializes to the Newton method (which in this case converges in a single step). The expression $\rho$ defining the convergence rate of this method is rather involved and it is not immediately obvious what is gained by performing a search in a higher dimensional subspace ($|C|>1$) rather than in the one-dimensional subspaces ($|C|=1$), as is standard in the optimization literature. Let us write $\rho = 1-\sigma_\tau$ in the case when the $C$ is chosen to be a subset of $\{1,\ldots, n\}$ of size $\tau$, uniformly at random. In view of Lemma~\ref{lem:itercomplex}, the method takes $\tilde{O}(1/\sigma_\tau)$ iterations to converge, where the tilde notation suppresses logarithmic terms. It was shown in \cite{Qu2015} that $1/\sigma_\tau \leq 1/(\tau \sigma_1)$. That is, one can expect to obtain at least {\em superlinear speedup} in $\tau$ --- this is what is gained by moving to blocks / higher dimensional subspaces. For further details and additional properties of the method we refer the reader to \cite{Qu2015}. \subsection{Randomized Coordinate Descent: least-squares version} By choosing $S=Ae^{i} =:A_{:i}$ as the $i$th column of $A$ and $B=A^\top A$, the resulting iterates~\eqref{ch:two:RF} are given by \begin{equation} \label{eq:CDLSintro} x^{k+1} = \arg\min_{x\in \mathbb{R}^n} \norm{Ax-b}_2^2 \quad \mbox{ subject to } \quad x = x^{k} + y \, e^{i}, \quad y \in \mathbb{R}. \end{equation} When $i$ is selected at random, this is the Randomized Coordinate Descent method (\emph{CD-LS}) applied to the least-squares problem: $\min_x \|Ax-b\|_2^2$. Using~\eqref{eq:MP}, these iterations can be calculated with \begin{equation}\label{eq:098sh98hs} \boxed{x^{k+1} = x^{k} - \frac{(A_{:i})^\top (A x^{k} -b)}{\norm{A_{:i}}_2^2} e^{i} } \end{equation} \paragraph{Complexity.} Applying Theorem~\ref{theo:convsingleS}, we see that by selecting $i$ with probability proportional to magnitude of column $i$ of $A$, that is $p_i = \norm{ A_{:i} }_2^2/\norm{A}_F^2$, results in a convergence with \begin{equation}\label{eq:CDLSconv} \E {\norm{x^{k} -x^{*} }_{A^\top A}^2 } \leq \rho^k \|x^0-x^*\|^2_{A^\top A} = \left(1 - \frac{\lambda_{\min}\left(A^\top A \right)}{\norm{A}_F^2} \right)^k \norm{x^{0} - x^{*}}_{A^\top A}^2.\end{equation} This result was first established by Leventhal and Lewis~\cite{Leventhal2010}. \paragraph{Interpretation.} Using the Constrain-and-Approximate formulation~\eqref{ch:two:RF}, the CD-LS method can be interpreted as \begin{equation} \label{eq:CDLSRF} x^{k+1} =\arg \min_{x\in \mathbb{R}^n} \norm{x- x^*}_{A^\top A}^2 \quad \mbox{ subject to } \quad x = x^k + y e^i , \quad y \in \mathbb{R}.\end{equation} The CD-LS method selects a coordinate to advance from the previous iterate $x^k$, then performs an exact minimization of the least squares function over this line. This is equivalent to applying coordinate descent to the least squares problem $\min_{x\in \mathbb{R}^n} f(x) \eqdef \tfrac{1}{2}\|Ax-b\|_2^2.$ The iterates~\eqref{eq:CDLSintro} can be written as \[x^{k+1} =x^k -\frac{1}{L_i}\nabla_i f(x^k) e^i,\] where $L_i \eqdef \norm{A_{:i}}_2^2$ is the Lipschitz constant of the gradient corresponding to coordinate $i$ and $\nabla_{i} f(x^k)$ is the $i$th partial derivative of $f$ at $x^k$. \section{Convergence: General Theory} \label{C2sec:convergence} We shall present two complexity theorems: we first study the convergence of $\norm{\E{x^{k}-x^*}}_B$ , and then move on to analysing the convergence of $\E{\norm{x^{k}-x^*}}_B$. Both theorems depend on the same convergence rate $\rho \in [0,\, 1]$, which we examine in the next section. In particular, we show that $\rho <1$ if and only if $A$ has full column rank. Thus the convergence results in this section only prove that the method convergence when $A$ has full column rank. Later in Chapter~\ref{ch:SDA} we extend these convergence results, and show that $A$ need not have full column rank. \subsection{The rate of convergence} All of our convergence theorems (see Table~\ref{ch:two:tab:complexity}) depend on the convergence rate \begin{equation}\label{eq:rho}\rho \eqdef 1 - \lambda_{\min}(B^{-1/2}\E{Z}B^{-1/2}).\end{equation} To show that the rate is meaningful, in Lemma~\ref{lem:rho1} we prove that $0\leq \rho \leq 1$. We also give an alternative expression and provide a meaningful lower bound for $\rho$. \begin{lemma}\label{lem:rho1} The quantity $\rho$ defined in~\eqref{eq:rho} satisfies: \begin{equation} \label{eq:rho_lower}0 \leq 1-\dfrac{\E{d}}{n}\leq \rho \leq 1,\end{equation} where $\dd=\Rank{S^\top A}$. Furthermore \begin{equation} \label{eq:rhoopnorm}\rho = \norm{I - B^{-1/2}\E{Z}B^{-1/2}}_2.\end{equation} \end{lemma} \begin{proof} Recall from Lemma~\ref{ch:one:lem:Z} that $B^{-1/2}ZB^{-1/2}$ is a projection, whence the spectrum of $B^{-1/2}ZB^{-1/2}$ is contained in $\{0,1\}$. Using this, combined with the fact that the mapping $A \mapsto \lambda_{\max}(A)$ is convex on the set of symmetric matrices and Jensen's inequality, we get \begin{equation}\label{eq:a89h93a8} \lambda_{\max} (B^{-1/2} \E{Z}B^{-1/2}) \leq\E{\lambda_{\max}(B^{-1/2}ZB^{-1/2})} \leq 1. \end{equation} The inequality $\lambda_{\min}(B^{-1/2}\E{Z}B^{-1/2})~\geq~0$ can be shown analogously using convexity of the mapping $A\mapsto -\lambda_{\min}(A)$. Thus, $ \lambda_{\min}(B^{-1/2}\E{Z}B^{-1/2}) \in [0, 1]$, which implies $0 \leq \rho \leq 1.$ We now refine the lower bound. As the trace of a matrix is equal to the sum of its eigenvalues, we have \begin{equation}\label{ch:one:eq:traceb1} \E{\Tr{B^{-1/2}ZB^{-1/2}}} =\Tr{\E{B^{-1/2}ZB^{-1/2}}} \geq n \, \lambda_{\min}(\E{B^{-1/2}ZB^{-1/2}}). \end{equation} From~\eqref{eq:B12ZB12trace} we that have $\Tr{B^{-1/2}ZB^{-1/2}} = \dd.$ Thus rewriting~\eqref{ch:one:eq:traceb1} gives $1-\E{\dd}/n \leq \rho.$ Finally, from the symmetry of $Z$ it follows that $(I-B^{-1/2}\E{Z}B^{-1/2})$ is symmetric, and consequently \begin{align*} \norm{(I-B^{-1/2}\E{Z}B^{-1/2})}_2 &= \lambda_{\max} (I-B^{-1/2}\E{Z}B^{-1/2})\\ &= \left(1-\lambda_{\min}(B^{-1/2}\E{Z}B^{-1/2})\right) = \rho.\nonumber \end{align*} \end{proof} The lower bound on $\rho$ in~\eqref{eq:rho_lower} has a natural interpretation which makes intuitive sense. We shall present it from the perspective of the Constrain-and-Approximate formulation~\eqref{ch:two:RF}. As the dimension ($\dd$) of the search space $B^{-1}A^\top S$ increases (see \eqref{eq:dimproj}), the lower bound on $\rho$ decreases, and a faster convergence is possible. For instance, when $S$ is restricted to being a random column vector, as it is in the RK~\eqref{eq:RKiterate}, CD-LS~\eqref{eq:098sh98hs} and CD-pd~\eqref{eq:CDposconv} methods, the convergence rate is bounded with $1 -1/n\leq \rho.$ Using Lemma~\ref{lem:itercomplex}, this translates into the simple iteration complexity bound of $k \geq n \log(1/\epsilon)$. On the other extreme, when the search space is large, then the lower bound is close to zero, allowing room for the method to be faster. We now characterize circumstances under which $\rho$ is strictly smaller than one. \begin{lemma}\label{lem:rho} If $\E{Z}$ is invertible, then $\rho <1$, $A$ has full column rank and $x^*$ is unique. \end{lemma} \begin{proof} Assume that $\E{Z}$ is invertible. First, this means that $B^{-1/2}\E{Z}B^{-1/2}$ is positive definite, which in view of \eqref{eq:rho} means that $\rho <1.$ If $A$ did not have full column rank, then there would be $0\neq x\in \mathbb{R}^n$ such that $Ax=0$. However, we then have $Zx = 0$ and also $\E{Z}x=0$, contradicting the assumption that $\E{Z}$ is invertible. Finally, since $A$ has full column rank, $x^*$ must be unique (recall that we assume throughout this chapter that the system $Ax=b$ is consistent). \end{proof} \subsection{Exact characterization and norm of expectation} We now state a theorem which exactly characterizes the evolution of the expected iterates through a linear fixed point iteration. As a consequence, we obtain a convergence result for the norm of the expected error. While we do not highlight this in the text, this theorem can be applied to all the particular instances of our general method we detail throughout this chapter. For any $M\in \mathbb{R}^{n\times n}$ let us define \begin{equation}\label{eq:specnorm} \|M\|_B \eqdef \max_{\norm{x}_B=1} \|Mx\|_B. \end{equation} \begin{theorem}[Norm of expectation]\label{ch:two:theo:normEconv} For every $x^*\in \mathbb{R}^n$ satisfying $Ax^*=b$ we have \begin{equation} \label{eq:Eerror} \E {x^{k+1} -x^{*}} = \left(I - B^{-1}\E{Z}\right) \E{x^{k} - x^{*}}. \end{equation} Moreover, the induced $B$-norm of the iteration matrix $I-B^{-1}\E{Z}$ is equal to $\rho$: \begin{equation}\label{eq:rhofixedpoint} \|I-B^{-1}\E{Z}\|_B = 1 - \lambda_{\min}(B^{-1/2}\E{Z}B^{-1/2}) = \rho.\end{equation} Therefore, \begin{equation} \label{ch:two:eq:normEconv} \norm{\E {x^{k} -x^{*}} }_B \leq\rho^{k} \norm{x^{0} - x^{*}}_B. \end{equation} \end{theorem} \begin{proof} Taking expectations conditioned on $x^k$ in~\eqref{eq:xZupdate}, we get \begin{equation}\label{eq:0suj9sj}\E{x^{k+1}-x^* \;|\; x^k} = (I-B^{-1}\E{Z})(x^k-x^*).\end{equation} Taking expectation again gives \begin{eqnarray*} \E{x^{k+1}-x^*} &=& \E{\E{x^{k+1}-x^* \;|\; x^k}} \\ &\overset{\eqref{eq:0suj9sj}}{=}& \E{(I-B^{-1}\E{Z})(x^k-x^*)} \\ &=& (I-B^{-1}\E{Z})\E{x^k-x^*},\end{eqnarray*} and thus~\eqref{eq:Eerror} holds. The equivalence~\eqref{eq:rhofixedpoint} follows by \begin{eqnarray*} \rho &\overset{\eqref{eq:rhoopnorm}}{=} & \norm{I - B^{-1/2}\E{Z}B^{-1/2}}_2 \\ &=& \max_{\norm{v}_2=1} \norm{(I - B^{-1/2}\E{Z}B^{-1/2})v}_2\\ &\overset{(v=B^{1/2}w)}{=}& \max_{\norm{B^{1/2}w}_2=1} \norm{B^{1/2}B^{-1/2}(I - B^{-1/2}\E{Z}B^{-1/2})B^{1/2}w}_2\\ &=&\max_{\norm{w}_B=1} \norm{(I - B^{-1}\E{Z})w}_B = \norm{I - B^{-1}\E{Z}}_B. \end{eqnarray*} For all $k$, define $r^{k} \eqdef B^{1/2}(x^{k}-x^*).$ Left multiplying~\eqref{eq:Eerror} by $B^{1/2}$ gives \[\E{r^{k+1}} = (I-B^{-1/2}\E{Z}B^{-1/2})\E{r^{k}}.\] Applying the norms to both sides we obtain the estimate \begin{equation}\label{eq:normEr} \norm{\E {r^{k+1}} }_2 \leq \norm{I-B^{-1/2}\E{Z}B^{-1/2}}_2 \, \norm{\E {r^k} }_2. \end{equation} The claim~\eqref{ch:two:eq:normEconv} now follows by observing~\eqref{eq:rhoopnorm}, that $\norm{r_k}_2 = \norm{x^{k}-x^*}_B$ and unrolling the recurrence in~\eqref{eq:normEr}. \end{proof} \subsection{Expectation of norm} We now turn to analysing the convergence of the expected norm of the error, for which we need the following technical lemma. \begin{theorem}[Expectation of norm]\label{ch:two:theo:Enormconv} If $\E{Z}$ is positive definite, then \begin{equation} \label{ch:two:eq:Enormconv} \E {\|x^{k} -x^{*} \|_B^2 } \leq \rho^k \norm{x^{0} - x^{*}}_B^2, \end{equation} where $\rho<1$ is given in~\eqref{eq:rho}. \end{theorem} \begin{proof} Let $r^k = B^{1/2}(x^k-x^*)B^{1/2}$. Taking expectation in~\eqref{eq:xZupdate} conditioned on $r^k$ gives \begin{eqnarray*} \E{\|r^{k+1}\|_2^2 \,\ | \,\ r^k} &\overset{\eqref{eq:xZupdate}}{=} & \E{\|(I-B^{-1/2}ZB^{-1/2})r^k\|_2^2 \,\ | \,\ r^k } \\ &\overset{\eqref{eq:B12ZB12proj}}{=}& \left< (I-B^{-1/2}\E{Z}B^{-1/2})r^k,r^k\right> \nonumber\\ & \leq & \norm{I-B^{-1/2}\E{Z}B^{-1/2}}_2 \, \|r^k\|_2^2. \end{eqnarray*} Using~\eqref{eq:rhoopnorm}, taking expectation again and unrolling the recurrence gives the result. \end{proof} The convergence rate $\rho$ of the expected norm of the error is ``worse'' than the $\rho^2$ rate of convergence of the norm of the expected error in Theorem~\ref{ch:two:theo:normEconv}. This should not be misconstrued as Theorem~\ref{ch:two:theo:normEconv} offering a ``better'' convergence rate than Theorem~\ref{ch:two:theo:Enormconv}, because, as explained in Lemma~\ref{lem:convrandvar}, convergence of the expected norm of the error is a stronger type of convergence. More importantly, the exponent is not of any crucial importance; clearly, an exponent of $2$ manifests itself only in halving the number of iterations (see Lemma~\ref{lem:itercomplex}). \section{Methods Based on Discrete Sampling}\label{C2sec:discrete} When $S$ has a discrete distribution, we can establish under reasonable assumptions when $\E{Z}$ is positive definite (Proposition~\ref{pro:Ediscrete}), we can optimize the convergence rate in terms of the chosen probability distribution, and finally, determine a probability distribution for which the convergence rate is expressed in terms of the scaled condition number (Theorem~\ref{theo:convsingleS}). \begin{assumption}\label{ass:complete} The random matrix $S$ has a discrete distribution. In particular, $S= S_i \in \mathbb{R}^{m \times q_i}$ with probability $p_i>0$, $\sum_{i=1}^r p_i =1$, where $S_i^\top A$ has full row rank and $q_i \in \mathbb{N},$ for $i=1,\ldots, r$. Furthermore $\mathbf{S} \eqdef [S_1, \ldots, S_r] \in \mathbb{R}^{m\times \sum_{i=1}^r q_i}$ is such that $A^\top \mathbf{S}$ has full row rank. \end{assumption} For simplicity, sampling $S$ satisfying the above assumption will be called a {\em complete discrete sampling}. We now give an example of such a sampling. If $A$ has full column rank and each row of $A$ is not strictly zero, $S =e^i$ with probability $p_i =1/n$, for $i =1,\ldots, n,$ then $\mathbf{S} =I$ and $S$ is a complete discrete sampling. In fact, from any basis of $\mathbb{R}^n$ we can construct a complete discrete sampling in an analogous way. When $S$ is a complete discrete sampling, then $S^\top A$ has full row rank and \[(S^\top A B^{-1}A^\top S)^{\dagger}=(S^\top A B^{-1}A^\top S)^{-1}.\] Therefore we replace the pseudoinverse in~\eqref{eq:MP} and~\eqref{eq:xZupdate} by the inverse. Furthermore, using a complete discrete sampling guarantees convergence of the resulting method. \begin{proposition}\label{pro:Ediscrete} If $S$ is a complete discrete sampling, $\E{Z}$ is positive definite. \end{proposition} \begin{proof} Let \begin{equation}\label{eq:defD} D \eqdef \mbox{diag}\left( \sqrt{p_1}((S_1)^\top A B^{-1}A^\top S_1)^{-1/2}, \ldots, \sqrt{p_r}((S_r)^\top A B^{-1}A^\top S_r)^{-1/2}\right) \end{equation} which is a block diagonal matrix, and is well defined and invertible as $S_i^\top A$ has full row rank for $i=1,\ldots, r$. Taking the expectation of $Z$~\eqref{eq:Z-first} gives \begin{align} \E{Z} &= \sum_{i=1}^r A^\top S_i (S_i^\top A B^{-1}A^\top S_i)^{-1}S_i^\top A p_i \nonumber \\ &= A^\top \left(\sum_{i=1}^r S_i \sqrt{p_i}(S_i^\top A B^{-1}A^\top S_i)^{-1/2} (S_i^\top A B^{-1}A^\top S_i)^{-1/2} \sqrt{p_i}S_i^\top \right) A \nonumber \\ &= \left( A^\top \mathbf{S} D\right) \left( D \mathbf{S}^\top A\right), \label{ch:one:eq:EZdiscrete} \end{align} which is positive definite because $A^\top \mathbf{S}$ has full row rank and $D$ is invertible. \end{proof} With $\E{Z}$ positive definite, we can apply the convergence Theorem~\ref{ch:two:theo:normEconv} and~\ref{ch:two:theo:Enormconv}, and the resulting method converges. \subsection{Optimal probabilities}\label{ch:two:sec:optprob} We can choose the discrete probability distribution that optimizes the convergence rate. For this, according to Theorems~\ref{ch:two:theo:Enormconv} and~\ref{ch:two:theo:normEconv} we need to find $p=(p_1,\dots,p_r)$ that maximizes the minimal eigenvalue of $B^{-1/2}\E{Z}B^{-1/2}$. Let $S$ be a complete discrete sampling and fix the sample matrices $S_1,\dots, S_r$. Let us denote $Z=Z(p)$ as a function of $p=(p_1,\dots,p_r)$. Then we can also think of the spectral radius as a function of $p$ where \[\rho(p) = 1 - \lambda_{\min}(B^{-1/2}\E{Z(p)}B^{-1/2}).\] If we let $\Delta_r =\left\{p = (p_1,\dots,p_r) \in \mathbb{R}^r \;:\; \sum_{i=1}^r p_i =1, \; p\geq 0\right\}$, the problem of minimizing the spectral radius (i.e., optimizing the convergence rate) can be written as \[\rho^* \quad \eqdef\quad \min_{p\in \Delta_r} \rho(p) \quad = \quad 1 - \max_{p\in \Delta_r} \lambda_{\min} (B^{-1/2}\E{Z(p)}B^{-1/2}).\] This can be cast as a convex optimization problem, by first re-writing \begin{align*} B^{-1/2}\E{Z(p)}B^{-1/2} &= \sum_{i=1}^r p_i \left(B^{-1/2}A^\top S_i (S_i^\top A B^{-1}A^\top S_i)^{-1}S_i^\top AB^{-1/2} \right)\\ &= \sum_{i=1}^r p_i \left(V_i (V_i^\top V_i)^{-1}V_i^\top \right), \end{align*} where $V_i = B^{-1/2}A^\top S_i.$ Thus \begin{equation}\label{eq:opt_sampling}\rho^* \quad = \quad 1 - \max_{p\in \Delta_r} \lambda_{\min} \left( \sum_{i=1}^r p_i V_i (V_i^\top V_i)^{-1} V_i^\top \right).\end{equation} To obtain $p$ that maximizes the smallest eigenvalue, we solve \begin{align} \max_{p,t} \,\, &\quad t \nonumber \\ \mbox{subject to}& \quad \sum_{i=1}^r p_i \left(V_i (V_i^\top V_i)^{-1}V_i^\top \right) \succeq t\cdot I, \label{eq:optconv}\\ & \quad p \in \Delta_r. \nonumber \end{align} Despite~\eqref{eq:optconv} being a convex semi-definite program{\footnote{When preparing a revision of the paper on which this chapter is based, we have learned about the existence of prior work~\cite{Dai2014} where the authors have also characterized the probability distribution that optimizes the convergences rate of the RK method as the solution to an SDP.}, which is apparently a harder problem than solving the original linear system, investing the time into solving~\eqref{eq:optconv} using a solver for convex conic programming such as \texttt{cvx}~\cite{cvx} can pay off, as we show in Section~\ref{C2sec:numopt}. Though for a practical method based on this, we would need to develop an approximate solution to~\eqref{eq:optconv} which can be efficiently calculated. \subsection{Convenient probabilities} \label{ch:two:sec:convprob} Next we develop a choice of probability distribution that yields a convergence rate that is easy to interpret. This result is an extension of Strohmer and Vershynin's~\cite{Strohmer2009} non-uniform probability distribution for the Kaczmarz method. Our extension includes a wide range of methods, such as the randomized Kaczmarz, randomized coordinate descent, as well as their block variants. However, it is more general, and covers many other possible particular algorithms, which arise by choosing a particular set of sample matrices $S_i$, for $i=1,\ldots, r.$ \begin{theorem} \label{theo:convsingleS} Let $S$ be a complete discrete sampling such that $S = S_i \in \mathbb{R}^{m}$ with probability \begin{equation}\label{ch:one:eq:convprob}p_i~=~\dfrac{\Tr{S_i^\top AB^{-1}A^\top S_i}}{\norm{B^{-1/2}A^\top \mathbf{S}}_F^2},\quad \mbox{for } \quad i=1,\ldots, { r}. \end{equation} Then the iterates~\eqref{eq:MP} satisfy \begin{equation} \label{eq:expnormcon} \E{\norm{x^{k} -x^{*} }_B^2} \leq \rho_c^k \,\norm{x^{0} - x^{*}}_B^2,\end{equation} where \begin{equation}\label{ch:one:eq:rhoconv} \rho_c = 1- \frac{\lambda_{\min}\left(\mathbf{S}^\top A B^{-1}A^\top \mathbf{S} \right)}{\norm{B^{-1/2}A^\top \mathbf{S}}_F^2}. \end{equation} \end{theorem} \begin{proof} Let $t_i = \Tr{S_i^\top A B^{-1}A^\top S_i}$, and with~\eqref{ch:one:eq:convprob} in~\eqref{eq:defD} we have \[ D^2 =\frac{1}{\norm{B^{-1/2}A^\top \mathbf{S}}_F^2}\mbox{diag}\left(t_1 (S_1^\top A B^{-1}A^\top S_1)^{-1}, \ldots, t_r (S_r^\top A B^{-1}A^\top S_r)^{-1}\right), \] thus \begin{equation}\label{eq:Dlambda} \lambda_{\min}(D^2) = \frac{1}{\norm{B^{-1/2}A^\top \mathbf{S}}_F^2}\min_i\left\{ \frac{t_i}{\lambda_{\max}(S_i^\top A B^{-1}A^\top S_i)} \right\} \geq \frac{1}{\norm{B^{-1/2}A^\top \mathbf{S}}_F^2}. \end{equation} Applying the above in~\eqref{ch:one:eq:EZdiscrete} gives \begin{align} \lambda_{\min}\left(B^{-1/2}\E{Z}B^{-1/2} \right) &= \lambda_{\min}\left(B^{-1/2} A^\top \mathbf{S} D^2 \mathbf{S}^\top A B^{-1/2}\right) \nonumber\\ &= \lambda_{\min}\left( \mathbf{S}^\top A B^{-1} A^\top \mathbf{S} D^2\right) \nonumber\\ & \geq \lambda_{\min}\left(\mathbf{S}^\top A B^{-1} A^\top \mathbf{S} \right)\lambda_{\min}(D^2) \label{eq:prodpdeig}\\ & \geq \frac{\lambda_{\min}\left(\mathbf{S}^\top A B^{-1} A^\top \mathbf{S} \right)}{\norm{B^{-1/2}A^\top \mathbf{S}}_F^2}, \nonumber \end{align} where in the first step we used the fact that for arbitrary matrices $B,C$ of appropriate sizes, $\lambda_{\min}(BC)=\lambda_{\min}(CB)$, and in the first inequality the fact that if $B,C \in \mathbb{R}^{n\times n}$ are positive definite, then $\lambda_{\min}(BC) \geq \lambda_{\min}(B)\lambda_{\min}(C)$. Finally \begin{equation}\label{eq:convrhobound} 1 - \lambda_{\min}\left(B^{-1/2}\E{Z}B^{-1/2} \right) \leq 1 - \frac{\lambda_{\min}\left(\mathbf{S}^\top A B^{-1} A^\top \mathbf{S} \right)}{\norm{B^{-1/2}A^\top \mathbf{S}}_F^2}.\end{equation} The result~\eqref{eq:expnormcon} follows by applying Theorem~\ref{ch:two:theo:Enormconv}. \end{proof} The convergence rate $\lambda_{\min}\left(\mathbf{S}^\top A B^{-1} A^\top \mathbf{S}\right)/ \norm{B^{-1/2}A^\top \mathbf{S}}_F^2$ is known as the scaled condition number, and naturally appears in other numerical schemes, such as matrix inversion~\cite{Edelman1992,Demmel1988}. When $S_i =s_i \in \mathbb{R}^n$ is a column vector then \[p_i~=~\left((s_i)^\top AB^{-1}A^\top s_i\right)/\norm{B^{-1/2}A^\top \mathbf{S}}_F^2,\] for $i=1,\ldots r.$ In this case, the bound~\eqref{eq:Dlambda} is an equality and $D^2$ is a scaled identity, so~\eqref{eq:prodpdeig} and consequently~\eqref{eq:convrhobound} are equalities. For block methods, it is different story, and there is much more slack in the inequality~\eqref{eq:convrhobound}. So much so, the convergence rate~\eqref{ch:one:eq:rhoconv} does not indicate any advantage of using a block method (contrary to numerical experiments). To see the advantage of a block method, we need to use the exact expression for $\lambda_{\min}(D^2)$ given in~\eqref{eq:Dlambda}. Though this results in a somewhat harder to interpret convergence rate, one could use a so called \emph{matrix paving} to explore this convergence rate, as was done for the block Kaczmarz method (see~\cite{Needell2014,Needell2012} for more details). By appropriately choosing $B$ and $S$, this theorem applied to RK method~\eqref{eq:RKintro}, the CD-LS method~\eqref{eq:CDLSintro} and the CD-pd method~\eqref{eq:CDpdintro}, yields the convergence results~\eqref{eq:RKconv},~\eqref{eq:CDLSconv} and~\eqref{eq:CDposconv}, respectively, for single column sampling or block methods alike. This theorem also suggests a preconditioning strategy, in that, a faster convergence rate will be attained if $\mathbf{S}$ is an approximate inverse of $B^{-1/2}A^\top .$ For instance, in the RK method where $B=I$, this suggests that an accelerated convergence can be attained if $S$ is a random sampling of the rows of a preconditioner (approximate inverse) of $A.$ \section{Methods Based on Gaussian Sampling} \label{C2sec:gauss} In this section we shall describe variants of our method in the case when $S$ is a Gaussian vector with mean $0\in \mathbb{R}^m$ and a positive definite covariance matrix $\Sigma\in \mathbb{R}^{m \times m}$. That is, $S =\zeta \sim N(0,\Sigma)$. This applied to~\eqref{eq:MP} results in iterations of the form \begin{equation}\label{eq:gaussupdate} \boxed{x^{k+1} = x^{k} - \frac{\zeta ^\top (A x^{k}-b)}{\zeta^\top AB^{-1}A^\top \zeta } B^{-1}A^\top \zeta} \end{equation} Unlike the discrete methods in Section~\ref{C2sec:examples}, to calculate an iteration of~\eqref{eq:gaussupdate} we need to compute the product of a matrix with a dense vector $\zeta$. This significantly raises the cost of an iteration. Though in our numeric tests in Section~\ref{C2sec:numerics}, the faster convergence of the Gaussian method often pays off for their high iteration cost. To analyze the complexity of the resulting method let $\xi \eqdef B^{-1/2}A^\top S,$ which is also Gaussian, distributed as $\xi\sim N(0, \Omega)$, where $\Omega\eqdef B^{-1/2}A^\top \Sigma A B^{-1/2}.$ In this section we assume {{\em $A$ has full column rank}, so that $\Omega$ is always positive definite. The complexity of the method can be established through \begin{eqnarray} \rho &=& 1 - \lambda_{\min}\left(\E{ B^{-1/2} ZB^{-1/2}}\right) = 1- \lambda_{\min}\left(\E{\frac{\xi\xi^\top }{\norm{\xi}^2_2}}\right). \end{eqnarray} We can simplify the above by using the lower bound \[ \E{\frac{\xi\xi^\top }{\norm{\xi}^2_2}} \succeq \frac{2}{\pi}\frac{\Omega}{\Tr{\Omega}},\] which is proven in Lemma~\ref{lem:gaussdiag} in the Appendix of this chapter. Thus \begin{equation}\label{eq:rhoboundgauss} 1- \frac{1}{n} \leq \rho \leq 1 -\frac{2}{\pi}\frac{\lambda_{\min}(\Omega)}{\Tr{\Omega}}, \end{equation} where we used the general lower bound in~\eqref{eq:rho_lower}. Lemma~\ref{lem:gaussdiag} also shows that $\E{\xi\xi^\top /\norm{\xi}^2_2}$ is positive definite, thus Theorem~\ref{ch:two:theo:Enormconv} guarantees that the expected norm of the error of all Gaussian methods converges exponentially to zero. This bound is tight upto a constant factor. Indeed, if $A=I=\Sigma$ then $\xi \sim N(0,I)$ and $\E{\xi\xi^\top /\norm{\xi}^2_2} = \tfrac{1}{n} I,$ which yields \[1-\dfrac{1}{n}\leq \rho \leq 1- \dfrac{2}{\pi }\cdot \dfrac{1}{n}.\] When $n=2$, then in Lemma~\ref{lem:2Dgausscov} of the Appendix of this chapter we prove that \[\E{\frac{\xi\xi^\top }{\norm{\xi}^2_2}} = \frac{\Omega^{1/2}}{\Tr{\Omega^{1/2}}},\] which yields a very favourable convergence rate. \subsection{Gaussian Kaczmarz}\label{C2sec:gaussI} Let $B=I$ and choose $\Sigma =I$ so that $S=\eta\sim N(0,I)$. Then~\eqref{eq:gaussupdate} has the form \begin{equation} \label{eq:gaussrkupdate} \boxed{ x^{k+1} = x^{k} - \frac{\eta^\top (Ax^{k}-b)}{\|A^\top \eta\|_2^2} A^\top \eta } \end{equation} which we call the \emph{Gaussian Kaczmarz} (GK) method, for it is the analogous method to the Randomized Kaczmarz method in the discrete setting. Using the formulation~\eqref{ch:two:RF}, for instance, the GK method can be interpreted as \[x^{k+1} =\arg \min_{x\in \mathbb{R}^n} \norm{x- x^{*}}^2 \quad \mbox{ subject to } \quad x = x^{k} + A^\top \eta y, \quad y\in \mathbb{R}. \] Thus at each iteration, a random normal Gaussian vector $\eta$ is drawn and a search direction is formed by $A^\top \eta.$ Then, starting from the previous iterate $x^{k}$, an exact line search is performed over this search direction so that the euclidean distance from the optimal is minimized. \subsection{Gaussian least-squares} \label{C2sec:gaussATA} Let $B=A^\top A$ and choose $S\sim N(0,\Sigma)$ with $\Sigma=AA^\top $. It will be convenient to write $S=A\eta$, where $\eta\sim N(0,I)$. Then method \eqref{eq:gaussupdate} then has the form \begin{equation} \label{eq:gausslsupdate} \boxed{x^{k+1} = x^{k} - \frac{\eta^\top A^\top (Ax^{k} -b)}{\|A\eta\|_2^2} \eta} \end{equation} which we call the \emph{Gauss-LS} method. This method has a natural interpretation through formulation~\eqref{ch:two:RF} as \[x^{k+1} =\arg \min_{x\in \mathbb{R}^n} \frac{1}{2}\|Ax-b\|_2^2 \quad \mbox{ subject to } \quad x = x^{k} + y\eta, \quad y \in \mathbb{R}. \] That is, starting from $x^{k}$, we take a step in a random (Gaussian) direction, then perform an exact line search over this direction that minimizes the least squares error. Thus the Gauss-LS method is the same as applying the Random Pursuit method~\cite{Stich2014} with exact line search to the Least-squares function. \subsection{Gaussian positive definite}\label{C2sec:gaussA} When $A$ is positive definite, we achieve an accelerated Gaussian method. Let $B =A$ and choose $S= \eta \sim N(0,I)$. Method~\eqref{eq:gaussupdate} then has the form \begin{equation}\label{eq:gausspd}\boxed{x^{k+1} = x^{k} - \frac{ \eta^\top (A x^{k}-b)}{\norm{\eta}_A^2} \eta} \end{equation} which we call the \emph{Gauss-pd} method. Using formulation~\eqref{ch:two:RF}, the method can be interpreted as \[x^{k+1} =\arg \min_{x\in \mathbb{R}^n} \left\{f(x) \eqdef \tfrac{1}{2} x^\top Ax -b^\top x\right\} \quad \mbox{ subject to } \quad x = x^{k} + y\eta, \quad y \in \mathbb{R}. \] That is, starting from $x^{k}$, we take a step in a random (Gaussian) direction, then perform an exact line search over this direction. Thus the Gauss-pd method is equivalent to applying the Random Pursuit method~\cite{Stich2014} with exact line search to $f(x).$ All the Gaussian methods can be extended to block versions. We illustrate this by designing a Block Gauss-pd method where $S \in \mathbb{R}^{n \times q}$ has i.i.d.\ Gaussian normal entries and $B=A.$ This results in the iterates \begin{equation} \label{eq:Bgausspd} x^{k+1} = x^{k} - S(S^\top AS)^{-1}S^\top (A x^{k}-b). \end{equation} \section{Numerical Experiments} \label{C2sec:numerics} We perform some preliminary numeric tests on consistent overdetermined linear systems and positive definite systems. Everything was coded and run in MATLAB R2014b. Let $\kappa_2 = \norm{A}\norm{A^{\dagger}}$ be the $2-$norm condition number. In comparing different methods for solving overdetermined systems, we use the relative residual $\norm{Ax^k-b}_2/\norm{b}_2,$ while for positive definite systems we use $\norm{x^k-x^*}_A/\norm{x^*}_A$ as a relative residual measure. We run each method until the relative residual is below $10^{-4} =0.01 \%$ or until $400$ seconds in time is exceeded. We test this relatively low precision of $0.01 \%$ because our applications of interest (e.g. ridge regression in machine learning) only require a low precision. Note that when $A$ has full column rank, the convergence of the relative residual implies the convergence of norm error $\norm{x-x^*}$. Indeed, this follows from \[\norm{Ax^k -b}_2^2 = \dotprod{A^\top A(x^k-x^*),x^k-x^*} \geq \lambda_{\min}(A^\top A) \norm{x^k-x^*}_2^2.\] Consequently bringing the relative residual $\norm{Ax^k-b}_2/\norm{b}_2$ below $0.01 \%$ implies that \[ \norm{x^k-x^*}_2 \leq \frac{\norm{b}_2}{\sqrt{\lambda_{\min}(A^\top A)}}10^{-4}.\] When the solution $x^*$ and the right hand $b$ were not supplied by the data, we generated them as follows. The solution was generated using $x^*=$\texttt{rand}$(n,1)$, that is, each entry of $x^*$ is selected uniformly at random from the interval $[0, 1]$. The right hand side $b$ was set to $b=Ax^*.$ As the starting point we used $x_0=0 \in \mathbb{R}^n$ in all experiments. In each figure we plot the relative residual in percentage on the vertical axis, starting with $100\%$. For the horizontal axis, we use either wall-clock time measured using the \texttt{tic-toc} MATLAB function or the total number of floating point operations (\emph{flops}). Specifically, we use an upper bound on the number of flops performed in each iteration as a proxy. For example, the number of flops required to compute the matrix-vector product $A v$ is bounded by $O(nnz(A))$ from above. This bound is tight when $v$ is dense. When $v$ is not dense, as is the case when $v= e_i,$ we use an appropriately tight upper bound, such as $O(nnz(A_{i:}))$ when $v=e_i$. In implementing the discrete sampling methods we used the convenient probability distributions~\eqref{ch:one:eq:convprob}. All tests were performed on a Desktop with 64bit quad-core Intel(R) Core(TM) i5-2400S CPU @2.50GHz with 6MB cache size with a Scientific Linux release 6.4 (Carbon) operating system. \subsection{Overdetermined linear systems} First we compare the methods Gauss-LS~\eqref{eq:gausslsupdate} , CD-LS~\eqref{eq:CDLSintro} , Gauss-Kaczmarz~\eqref{eq:gaussrkupdate} and RK~\eqref{eq:RKiterate} methods on synthetic linear systems generated with the matrix functions {\tt rand} and {\tt sprandn}, see Figure~\ref{fig:oversynth}. The {\tt rand}$(m,n)$ function returns a $m$-by-$n$ matrix where each entry is a random variable selected uniformly at random from the interval $[0, 1]$. The high iteration cost of the Gaussian methods resulted in poor performance on the dense problem generated using \texttt{rand} in Figure~\ref{fig:randch1}. In Figure~\ref{fig:sprandn} we compare the methods on a sparse linear system generated using the MATLAB sparse random matrix function \texttt{sprandn}($m,n$,{\tt density,rc}), where {\tt density} is the percentage of nonzero entries and {\tt rc} is the reciprocal of the condition number. The {\tt sprandn}($m,n$,{\tt density,rc}) returns a random $m$-by-$n$ sparse matrix with approximately {\tt density}$\,\times m \times n$ normally distributed nonzero entries. Please consult \url{http://uk.mathworks.com/help/matlab/ref/sprandn.html} for more details on {\tt sprandn}. On this sparse problem the Gaussian methods are more efficient, and converge at a similar rate in time to the discrete sampling methods. \begin{figure} \centering \begin{subfigure}[t]{0.7\textwidth} \includegraphics[width = 0.85\textwidth, height =4.5cm, trim= 32 270 47 285, clip ]{uniform-random1000X500.pdf} \caption{\texttt{rand}}\label{fig:randch1} \end{subfigure}% \hspace{0.05\textwidth} \begin{subfigure}[t]{0.7\textwidth} \includegraphics[width = 0.85\textwidth, height =4.5cm, trim= 32 270 47 285, clip ]{sprandn-1000-500-0-076206-0-0014142} \caption{\texttt{sprandn}} \label{fig:sprandn} \end{subfigure} \caption{ The performance of the Gauss-LS, CD-LS, Gauss-Kaczmarz and RK methods on synthetic MATLAB generated problems (a) \texttt{rand}$(n,m)$ with $(m;n)=(1000,500)$ (b) \texttt{sprandn}($m,n$,{\tt density,rc}) with $(m;n) =(1000,500)$, {\tt density}$=1/\log(nm)$ and {\tt rc}$= 1/\sqrt{mn}$. In both experiments dense solutions were generated with $x^*=$\texttt{rand}$(n,1)$ and $b=Ax^*.$ }\label{fig:oversynth} \end{figure} In Figure~\ref{fig:overMM} we test two overdetermined linear systems taken from the the Matrix Market collection~\cite{Boisvert1997}. The collection also provides the right-hand side of the linear system. Both of these systems are very well conditioned, but do not have full column rank, thus Theorem~\ref{ch:two:theo:Enormconv} does not apply. The four methods have a similar performance on Figure~\ref{fig:illc1033}, while the Gauss-LS and CD-LS method converge faster on~\ref{fig:well1033} as compared with the Gauss-Kaczmarz and Kaczmarz methods in terms of time taken. In terms of number of flops, the CD-LS and Kaczmarz method outperform the Gauss-LS and Gauss-Kaczmarz methods. \begin{figure} \centering \begin{subfigure}[t]{0.70\textwidth} \includegraphics[width = 0.85\textwidth, height =4.5cm, trim= 32 270 47 285, clip ]{illc1033.pdf} \caption{\texttt{illc1033}}\label{fig:illc1033} \end{subfigure}% \hspace{0.05\textwidth} \begin{subfigure}[t]{0.70\textwidth} \includegraphics[width = 0.85\textwidth, height =4.5cm, trim= 32 270 47 285, clip ]{well1033.pdf} \caption{\texttt{well1033} } \label{fig:well1033} \end{subfigure} \caption{The performance of the Gauss-LS, CD-LS, Gauss-Kaczmarz and RK methods on linear systems (a) \texttt{well1033} where $(m;n) = (1850,750)$, $nnz = 8758$ and $\kappa_2 = 1.8$ (b) \texttt{illc1033} where $(m;n) =(1033;320)$, $nnz= 4732$ and $\kappa_2 = 2.1$, from the Matrix Market~\cite{Boisvert1997}.}\label{fig:overMM} \end{figure} Finally, we test two problems, the \texttt{SUSY} problem and the \texttt{covtype.binary} problem, from the library of support vector machine problems LIBSVM~\cite{Chang2011}. These problems do not form consistent linear systems, thus only the Gauss-LS and CD-LS methods are applicable, see Figure~\ref{ch:one:fig:LIBSVM}. This is equivalent to applying the Gauss-pd and CD-pd to the least squares system $A^\top Ax=A^\top b,$ which is always consistent. \begin{figure} \centering \begin{subfigure}[t]{0.7\textwidth} \centering \includegraphics[width = 0.85\textwidth, height =4.5cm, trim= 28 270 47 285, clip ]{SUSY} \caption{\texttt{SUSY}} \end{subfigure}% \hspace{0.05\textwidth} \begin{subfigure}[t]{0.7\textwidth} \centering \includegraphics[width = 0.85\textwidth, height =4.5cm, trim= 30 270 47 285, clip ]{covtype-libsvm-binary} \caption{\texttt{covtype-libsvm-binary}} \end{subfigure} \caption{The performance of Gauss-LS and CD-LS methods on two LIBSVM test problems: (a) \texttt{SUSY}: $(m;n)=(5\times 10^6; 18)$ (b) \texttt{covtype.binary}: $(m;n)=(581,012; 54)$.} \label{ch:one:fig:LIBSVM} \end{figure} Despite the higher iteration cost of the Gaussian methods, their performance, in terms of the wall-clock time, is comparable to performance of the discrete methods when the system matrix is sparse. \subsection{Bound for Gaussian convergence} \label{C2sec:numgaussbound} Now we compare the error over the number iterations of the Gauss-LS method to theoretical rate of convergence given by the bound~\eqref{eq:rhoboundgauss}. For the Gauss-LS method~\eqref{eq:rhoboundgauss} becomes \[1-\frac{1}{n} \leq \rho \leq 1-\frac{2}{\pi}\lambda_{\min}\left( \frac{A^\top A}{\norm{A}_F^2}\right).\] In Figures~\ref{fig:overtheouniform} and~\ref{fig:overtheoliver} we compare the empirical and theoretical bound on a random Gaussian matrix and the \texttt{liver-disorders} problem~\cite{Chang2011}. Furthermore, we ran the Gauss-LS method 100 times and plot as a shaded region the outcomes within the 95\% and 5\% quantiles. These tests indicate that the bound is tight for well conditioned problems, such as Figure~\ref{fig:overtheouniform} in which the system matrix has a condition number equal to $1.94$. While in Figure~\ref{fig:overtheoliver} the system matrix has a condition number of $41.70$ and there is some much more slack between the empirical convergence and the theoretical bound. \begin{figure} \centering \begin{subfigure}[t]{0.47\textwidth} \includegraphics[width = \textwidth, trim= 110 280 100 285, clip ]{Gaussls-uniform-random500X50-bnbdist} \caption{\texttt{rand}$(n,m)$}\label{fig:overtheouniform} \end{subfigure}% \hspace{0.05\textwidth} \begin{subfigure}[t]{0.47\textwidth} \includegraphics[width = \textwidth, trim= 110 280 100 285, clip ]{Gaussls-liver-disorders-bnbdist.pdf} \caption{\texttt{liver-disorders}} \label{fig:overtheoliver} \end{subfigure} \caption{A comparison between the Gauss-LS method and the theoretical bound $\rho_{theo} \eqdef 1-\lambda_{\min}(A^\top A)/\norm{A}_F^2$ on (a) \texttt{rand}$(n,m)$ with $(m;n)=(500,50), \kappa_2 =1.94$ and a dense solution generated with $x^*=$\texttt{ rand}$(n,1)$ (b) \texttt{liver-disorders} with $(m;n) =(345,6)$ and $\kappa_2 = 41.70.$ }\label{fig:overtheo} \end{figure} \subsection{Positive definite} First we compare the two methods Gauss-pd~\eqref{eq:gausspd} and CD-pd~\eqref{eq:09j0s9jsss} on synthetic data in Figure~\ref{fig:2}. \begin{figure} \centering \begin{subfigure}[t]{0.47\textwidth} \includegraphics[width = \textwidth, trim= 40 310 60 310, clip ]{Hilbert-100-Block} \end{subfigure}% \hspace{0.05\textwidth} \begin{subfigure}[t]{0.47\textwidth} \includegraphics[width = \textwidth, trim= 40 310 60 310, clip ]{sprandsym-1000-0-072382-0-001-1} \end{subfigure} \caption{Synthetic MATLAB generated problem. The Gaussian methods are more efficient on sparse matrices. LEFT: The \texttt{Hilbert Matrix} with $n=100$ and condition number $\norm{A}\norm{A^{-1}} \approx 0.001133\times e^{349}$. RIGHT: Sparse random matrix $A=$ \texttt{sprandsym} ($n$, {\tt density}, {\tt rc}, {\tt type}) with $n=1000$, {\tt density}$ =1/\log(n^2)$ and ${\tt rc}= 1/n=0.001$. Dense solution generated with $x^{*}=$\texttt{rand}$(n,1).$}\label{fig:2} \end{figure} Using the MATLAB function {\tt hilbert} we generate the positive definite Hilbert matrix which has a very high condition number, see Figure~\ref{fig:2}(LEFT). Indeed, the $100\times 100$ Hilbert matrix we tested has a condition number of approximately $0.001133\times e^{349}$! Both methods converge slowly and, despite the dense system matrix, the Gauss-pd method has a similar performance to CD-pd. In Figure~\eqref{fig:2}(RIGHT) we compare the two methods on a system generated by the MATLAB function \texttt{sprandsym} ($m$, $n$, {\tt density}, {\tt rc}, {\tt type}), where {\tt density} is the percentage of nonzero entries, {\tt rc} is the reciprocal of the condition number and {\tt type=1} returns a positive definite matrix. The Gauss-pd and the CD-pd method have a similar performance in terms of wall clock time on this sparse problem. \begin{figure} \centering \begin{subfigure}[t]{0.475\textwidth} \centering \includegraphics[width = \textwidth, trim= 40 310 60 310, clip ]{aloi-ridge-Block} \caption{\texttt{aloi}} \end{subfigure}% \hspace{0.04\textwidth} \begin{subfigure}[t]{0.475\textwidth} \centering \includegraphics[width = \textwidth, trim= 40 310 60 310, clip ]{protein-ridge-Block} \caption{\texttt{protein}} \end{subfigure} \begin{subfigure}[t]{0.475\textwidth} \centering \includegraphics[width = \textwidth, trim= 40 310 60 310, clip ]{SUSY-ridge-Block} \caption{\texttt{SUSY}} \end{subfigure}% \hspace{0.04\textwidth} \begin{subfigure}[t]{0.475\textwidth} \centering \includegraphics[width = \textwidth, trim= 40 310 60 310, clip ]{covtype-libsvm-binary-ridge-Block} \caption{\texttt{covtype.binary}} \end{subfigure} \caption{The performance of Gauss-pd and CD-pd methods on four ridge regression problems: (a) \texttt{aloi}: $(m;n)=(108,000;128)$ (b) \texttt{protein}: $(m; n)=(17,766; 357)$ (c) \texttt{SUSY}: $(m;n)=(5\times 10^6; 18)$ (d) \texttt{covtype.binary}: $(m;n)=(581,012; 54)$.} \label{ch:one:fig:LIBSVMridge} \end{figure} To appraise the performance gain in using block variants, we perform tests using two block variants: the Randomized Newton method~\eqref{eq:CDpdblock}, which we will now refer to as the Block CD-pd method, and the Block Gauss-pd method~\eqref{eq:Bgausspd}. We set the block size to $q= \sqrt{n}$ in both methods. To solve the $q\times q$ system required in the block methods, we use MATLAB's built-in direct solver, sometimes referred to as ``back-slash''. Next we test the Newton system $\nabla^2 f(w_0) x = - \nabla f(w_0)$, arising from four ridge-regression problems of the form \begin{equation}\label{eq:ridgeMatrix} \min_{w\in \mathbb{R}^n}f(w)\eqdef \tfrac{1}{2} \norm{Aw-b}_2^2 + \tfrac{\lambda}{2} \norm{w}_2^2, \end{equation} using data from LIBSVM~\cite{Chang2011}. In particular, we set $w_0=0$ and use $\lambda =1$ as the regularization parameter, whence $\nabla f(w_0) = A^\top b$ and $\nabla^2 f(w_0) = A^\top A+ I$. In terms of wall clock time, the Gauss-pd method converged faster on all problems accept the \texttt{aloi} problem as compared with CD-pd. The two Block methods had a comparable performance on the \texttt{aloi} and the \texttt{protein} problem. The Block Gauss-pd method converged in one iteration on \texttt{covtype.binary} and was the fastest method on the \texttt{SUSY} problem. We now compare the methods on two positive definite matrices from the Matrix Market collection~\cite{Boisvert1997}, see Figure~\ref{fig:pdMM}. The right-hand side was not supplied by the data set, and thus we generated $b$ using {\tt rand(n,1)}. The Block CD-pd method converged much faster on both problems. The lower condition number ($\kappa_2=12$) of the \texttt{gr\_30\_30-rsa} problem resulted in fast convergence of all methods, see Figure~\ref{fig:gr_30_30-rsa}. While the high condition number ($\kappa_2=4.3 \cdot 10^4$) of the \texttt{bcsstk18} problem, resulted in a slow convergence for all methods, see Figure~\ref{fig:bcsstk18-rsa}. \begin{figure} \centering \begin{subfigure}[t]{0.47\textwidth} \includegraphics[width = \textwidth, trim= 40 310 60 310, clip ]{gr-30-30-rsa-Block} \caption{ \texttt{gr\_30\_30-rsa}}\label{fig:gr_30_30-rsa} \end{subfigure}% \hspace{0.05\textwidth} \begin{subfigure}[t]{0.47\textwidth} \includegraphics[width = \textwidth, trim= 40 310 60 310, clip ]{bcsstk18-rsa-block} \caption{\texttt{bcsstk18}} \label{fig:bcsstk18-rsa} \end{subfigure} \caption{The performance of the Gauss-pd, CD-pd and the Block CD-pd methods on two linear systems from the MatrixMarket (a) \texttt{gr\_30\_30-rsa} with $n = 900$, $nnz = 4322$ ({\tt density}$=0.53\%$) and $\kappa_2 =12.$ (b) \texttt{bcsstk18} with $n = 11948$, $nnz=80519$ ({\tt density}$=0.1\%$) and $\kappa_2 = 4.3 \cdot 10^{10} $.}\label{fig:pdMM} \end{figure} Despite the clear advantage of using a block variant, applying a block method that uses a direct solver can be infeasible on very ill-conditioned problems. As an example, applying the Block CD-pd to the Hilbert system, and using MATLAB back-slash solver to solve the inner $q\times q$ systems, resulted in large numerical inaccuracies, and ultimately, prevented the method from converging. This occurred because the submatrices of the Hilbert matrix are also very ill-conditioned. \subsection{Comparison between optimized and convenient probabilities}\label{C2sec:numopt} We compare the practical performance of using the convenient probabilities~\eqref{ch:one:eq:convprob} against using the optimized probabilities by solving~\eqref{eq:optconv}. We solved~\eqref{eq:optconv} using the disciplined convex programming solver \texttt{cvx}~\cite{cvx} for MATLAB. In Table~\ref{tab:CD-pd-opt} we compare the different convergence rates for the CD-pd method, where $\rho_c$ is the convenient convergence rate~\eqref{ch:one:eq:rhoconv}, $\rho^*$ the optimized convergence rate, $(1-1/n)$ is the lower bound, and in the final ``optimized time(s)'' column the time taken to compute $\rho^*$. In Figure~\ref{fig:CD-pd-opt}, we compare the empirical convergence of the CD-pd method when using the convenient probabilities~\eqref{ch:one:eq:convprob} and CD-pd-opt, the CD-pd method with the optimized probabilities. We tested the two methods on four ridge regression problems and a synthetic positive definite system which is the square of a uniform random matrix: $A~=~\bar{A}^\top \bar{A}$ where $\bar{A}=$\texttt{rand}$(50)$. We ran each method for $60$ seconds. In most cases using the optimized probabilities results in a much faster convergence, see Figures~\ref{fig:aloi-opt},~\ref{fig:liver-opt},~\ref{fig:mushrooms-opt} and~\ref{fig:uniform-opt}. In particular, the $7.401$ seconds spent calculating the optimal probabilities for \texttt{aloi} paid off with a convergence that was $55$ seconds faster. The \texttt{mushrooms} problem was insensitive to the choice of probabilities~\ref{fig:mushrooms-opt}. Finally despite $\rho^*$ being much less than $\rho_c$ on \texttt{covtype}, see Table~\ref{tab:CD-pd-opt}, using optimized probabilities resulted in an initially slower method, though CD-pd-opt eventually catches up as CD-pd stagnates, see Figure~\ref{fig:covtype-opt}. \begin{table} \centering \footnotesize \begin{tabular}{c|lll|c} \hline data set & $\rho_c$ &$\rho^* $ & $1-1/n$ & optimized time(s) \\ \hline \texttt{rand}(50,50) & $1-2\cdot 10^{-6}$ & $1-3.05\cdot 10^{-6}$ & $1-2.10^{-2}$ & 1.076\\ {\tt mushrooms-ridge} & $1-5.86\cdot 10^{-6}$ & $1-7.15\cdot 10^{-6}$ & $1-8.93\cdot 10^{-3}$ & 4.632\\ {\tt aloi-ridge} & $1-2.17\cdot 10^{-7}$ & $1-1.26\cdot 10^{-4}$ & $1-7.81\cdot 10^{-3}$ & 7.401\\ {\tt liver-disorders-ridge} & $1-5.16\cdot 10^{-4}$ & $1-8.25\cdot 10^{-3}$ & $1-1.67\cdot 10^{-1}$ & 0.413\\ {\tt covtype.binary-ridge} & $1-7.57\cdot 10^{-14}$ & $1-1.48\cdot 10^{-6}$ & $1-1.85\cdot 10^{-2}$ & 1.449\\ \end{tabular} \caption{Optimizing the convergence rate for CD-pd.} \label{tab:CD-pd-opt} \end{table} \begin{figure}[!h] \centering \begin{subfigure}[t]{0.45\textwidth} \centering \includegraphics[width = \textwidth, trim=100 280 100 290, clip ]{aloi-ridge-opt} \caption{\texttt{aloi}} \label{fig:aloi-opt} \end{subfigure}% \hspace{0.02\textwidth} \begin{subfigure}[t]{0.45\textwidth} \centering \includegraphics[width = \textwidth, trim=100 280 100 290, clip ]{covtype-libsvm-binary-ridge-opt} \caption{\texttt{covtype.libsvm.binary}} \label{fig:covtype-opt} \end{subfigure} \begin{subfigure}[t]{0.45\textwidth} \centering \includegraphics[width = \textwidth, trim=100 280 100 290, clip ]{liver-disorders-ridge-opt} \caption{\texttt{liver-disorders-ridge}} \label{fig:liver-opt} \end{subfigure} \begin{subfigure}[t]{0.45\textwidth} \centering \includegraphics[width = \textwidth, trim=100 280 110 270, clip ]{mushrooms-ridge-opt} \caption{\texttt{mushrooms-ridge}} \label{fig:mushrooms-opt} \end{subfigure} \begin{subfigure}[t]{0.45\textwidth} \centering \includegraphics[width = \textwidth, trim=100 280 100 290, clip ]{uniform-random-50X50-opt} \caption{$50X50$ synthetic positive definite} \label{fig:uniform-opt} \end{subfigure} \caption{The performance of CD-pd and optimized CD-pd methods on (a) \texttt{aloi}: $(m;n)=(108,000;128)$ (b) \texttt{covtype.binary}: $(m;n)=(581,012; 54)$ (c) \texttt{liver-disorders}: $(m;n)=(345,6)$ (c)\texttt{mushrooms}: $(m;n) = (8124,112)$ (d) $A~=~\bar{A}^\top \bar{A}$ where $\bar{A}=$\texttt{rand}$(50)$.} \label{fig:CD-pd-opt} \end{figure} In Table~\ref{tab:RK-pd-opt} we compare the different convergence rates for the RK method. In Figure~\ref{fig:kaczmacz-opt}, we then compare the empirical convergence of the RK method when using the convenient probabilities~\eqref{ch:one:eq:convprob} and RK-opt, the RK method with the optimized probabilities by solving~\eqref{eq:optconv}. The rates $\rho^*$ and $\rho_c$ for the \texttt{rand}(500,100) problem are similar, and accordingly, both the convenient and optimized variant converge at a similar rate in practice, see Figure~\ref{fig:kaczmacz-opt}b. While the difference in the rates $\rho^*$ and $\rho_c$ for the {\tt liver-disorders} is more pronounced, and in this case, the $0.83$ seconds invested in obtaining the optimized probability distribution paid off in practice, as the optimized method converged $1.25$ seconds before the RK method with the convenient probability distribution, see Figure~\ref{fig:kaczmacz-opt}a. \begin{table} \centering \footnotesize \begin{tabular}{c|lll|c} \hline data set & $\rho_c$ &$\rho^* $ & $1-1/n$ & optimized time(s) \\ \hline \texttt{rand}(500,100) & $1-3.37\cdot 10^{-3}$ & $1-4.27\cdot 10^{-3}$ & $1-1\cdot 10^{-2}$ & 33.121\\ {\tt liver-disorders} & $1-5.16\cdot 10^{-4}$ & $1-4.04\cdot 10^{-3}$ & $1-1.67 \cdot 10^{-1}$ & 0.8316\\ \end{tabular} \caption{Optimizing the convergence rate for randomized Kaczmarz.} \label{tab:RK-pd-opt} \end{table} \begin{figure}[!h] \centering \begin{subfigure}[t]{0.45\textwidth} \centering \includegraphics[width = \textwidth, trim=100 280 100 290, clip ]{liver-disorders-popt-k} \caption{\texttt{liver-disorders-popt-k}} \end{subfigure}% \hspace{0.02\textwidth} \begin{subfigure}[t]{0.45\textwidth} \centering \includegraphics[width = \textwidth, trim=100 280 100 290, clip ]{uniform-random500X100-popt-k} \caption{\texttt{rand}(500,100)} \end{subfigure} \caption{The performance of the Kaczmarz and optimized Kaczmarz methods on (a) \texttt{liver-disorders}: $(m;n)=(345,6)$ (b) \texttt{rand}(500,100)} \label{fig:kaczmacz-opt} \end{figure} We conclude from these tests that the choice of the probability distribution can greatly affect the performance of the method. Hence, it is worthwhile to develop approximate solutions to~\eqref{eq:opt_sampling}. \subsection{Conclusion of numeric experiments} We now summarize the findings of our numeric experiments. \begin{itemize} \item Consistently across our experiments, in terms of number of flops taken to reach a desired precision, the three discrete sampling methods CD-LS, CD-pd and Kaczmarz are the most efficient. That is, the Gaussian methods almost always require more flops to reach a solution with the same precision as their discrete sampling counterparts. This is due to the expensive matrix-vector product required by the Gaussian methods. \item In terms of wall-clock time, the Gaussian methods Guass-LS, Gauss-pd and Gauss Kaczmarz are competitive as compared to the discrete sampling methods. This occurred because MATLAB performs automatic multi-threading when calculating matrix-vector products, which was the bottleneck cost in the Gaussian methods. As our machine has four cores, this explains some of the difference observed when measuring performance in terms of number of flops and wall clock time. \item In terms of both time taken and flops, the block variants proved to be significantly more efficient. \item Using the optimized probabilities~\eqref{eq:optconv} for the discrete sampling methods RK and CD-pd can result in significant speed-ups, as compared to RK and CD-pd using the convenient probabilities~\eqref{ch:one:eq:convprob}. So much so, that the time spent solving the SDP~\eqref{eq:optconv} using {\tt cvx}~\cite{cvx} often paid off. We can draw two interesting conclusions from this: (1) using convenient probabilities~\eqref{ch:one:eq:convprob} is not the best choice, but simply, the choice that provides easily interpretable convergence rates, (2) it is worth further investigating the use of optimization probabilities, for instance, one should investigate it is possible to obtain affordable approximate solutions to~\eqref{eq:optconv}. \end{itemize} \section{Summary} \label{C2sec:conclusion} In this chapter we presented a unifying framework for the randomized Kaczmarz method, randomized Newton method, randomized coordinate descent method and random Gaussian pursuit. Not only can we recover these methods by selecting appropriately the parameters $S$ and $B$, but also, we can analyze them and their block variants through a single Theorem~\ref{ch:two:theo:Enormconv}. Furthermore, we obtain a new lower bound for all these methods in Theorem~\ref{ch:two:theo:normEconv}, and in the discrete case, recover all known convergence rates expressed in terms of the scaled condition number in Theorem~\ref{theo:convsingleS}. Theorem~\ref{theo:convsingleS} also suggests a preconditioning strategy. Developing preconditioning methods are important for reaching a higher precision solution on ill-conditioned problems. For as we have seen in the numerical experiments, the randomized methods struggle to bring the solution within $10^{-2}\%$ relative residual when the matrix is ill-conditioned. This is also a framework on which randomized methods for linear systems can be designed. As an example, we have designed a new block variant of RK, a new Gaussian Kaczmarz method and a new Gaussian block method for positive definite systems. Furthermore, the flexibility of our framework and the general convergence Theorems~\ref{ch:two:theo:Enormconv} and~\ref{ch:two:theo:normEconv} allows one to tailor the probability distribution of $S$ to a particular problem class. For instance, other continuous distributions such uniform, or other discrete distributions such Poisson might be more suited to a particular class of problems. Numeric tests reveal that the new Gaussian methods designed for overdetermined systems are competitive on sparse problems, as compared with the Kaczmarz and CD-LS methods. The Gauss-pd also proved competitive as compared with CD-pd on all tests. Though, when applicable, the combined efficiency of using a direct solver and an iterative procedure, such as in Block CD-pd method, proved the most efficient. The work opens up many possible future venues of research. Including investigating accelerated convergence rates through preconditioning strategies based on Theorem~\ref{theo:convsingleS} or by obtaining approximate optimized probability distributions~\eqref{eq:optconv}. \subsubsection*{Acknowledgments} I would like to thank Prof. Sandy Davie for useful discussions relating to Lemma~\ref{lem:2Dgausscov}, and Prof. Joel Tropp for help with formulating and proving Lemma~\ref{lem:gaussdiag}. \section{Appendix: A Bound on the Expected Gaussian Projection Matrix} We now bound the covariance of a random Gaussian vector projected onto the sphere. This bound is used to study the complexity of Gaussian methods in Section~\ref{C2sec:gauss}. \begin{lemma} \label{lem:gaussdiag} Let $D \in \mathbb{R}^{n\times n}$ be a positive definite diagonal matrix, $U \in \mathbb{R}^{n\times n}$ an orthogonal matrix and $\Omega =UDU^\top $. If $u \sim N(0,D)$ and $\xi \sim N(0,\Omega)$ then \begin{equation}\label{eq:gaussdiag}\E{\frac{\xi \xi^\top }{\xi^\top \xi}} =U\E{\frac{u u^\top }{u^\top u}}U^\top ,\end{equation} and \begin{equation}\label{eq:gaussupper} \E{\frac{\xi \xi^\top }{\xi^\top \xi}} \succeq \frac{2}{\pi}\frac{\Omega}{\Tr{\Omega}}. \end{equation} \end{lemma} \begin{proof} Let us write $S(\xi)$ for the random vector $\xi/\|\xi\|_2$ (if $\xi=0$, we set $S(\xi)=0$). Using this notation, we can write \[ \E{ \xi (\xi^\top \xi)^{-1} \xi^\top } = \E{S(\xi)(S(\xi))^\top } = \COV{S(\xi)},\] where the last identity follows since $\E{S(\xi)}=0$, which in turn holds as the Gaussian distribution is centrally symmetric. As $\xi = U u$, note that \[S(u) = \frac{U^\top \xi}{\|U^\top \xi\|_2} = \frac{U^\top \xi}{\|\xi\|_2} = U^\top S(\xi).\] Left multiplying both sides by $U$ we obtain $U S(u) = S(\xi)$, from which we obtain \[\COV{S(\xi)} = U \COV{ S(u) } U^\top , \] which is equivalent to~\eqref{eq:gaussdiag}. To prove\footnote{A version of Lemma~\ref{lem:gaussdiag} was conjectured in the original draft of the paper on which this chapter is based. Prof. Joel Tropp provided this formulation and the remainder of this proof.} \eqref{eq:gaussupper}, note first that $M \eqdef \E{u u^\top /u^\top u}$ is a diagonal matrix. One can verify this by direct calculation (informally, this holds because the entries of $u$ are independent and centrally symmetric). The $i$th diagonal entry is given by \[M_{ii} = \E{\frac{u_{i}^2}{\sum_{j=1}^n u_j^2}}.\] As the map $(x,y) \rightarrow x^2/y$ is convex on the positive orthant, we can apply Jensen's inequality, which gives \[\E{\frac{u_{i}^2}{\sum_{j=1}^n u_j^2}} \geq \frac{\left(\E{|u_{i}|}\right)^2}{\sum_{j=1}^n \E{u_j^2}} = \frac{2}{\pi}\frac{D_{ii}}{\Tr{D}},\] which concludes the proof. \end{proof} \section{Appendix: Expected Gaussian Projection Matrix in 2D} \begin{lemma} \label{lem:2Dgausscov} Let $\xi \sim N(0,\Omega)$ and $\Omega \in \mathbb{R}^{2\times 2}$ be a positive definite matrix, then \begin{equation}\label{eq:2Dgausscov}\E{\frac{\xi \xi^\top }{\xi^\top \xi}} = \frac{\Omega^{1/2}}{\Tr{\Omega^{1/2}}}.\end{equation} \end{lemma} \begin{proof} Let $\Sigma = U D U^\top $ and $u \sim N(0,D).$ Given~\eqref{eq:gaussdiag} it suffices to show that \begin{equation}\label{eq:h98hs98hs9ss}\COV{S(u)} = \frac{D^{1/2}}{\Tr{D^{1/2}}},\end{equation} which we will now prove. Let $\sigma_x^2$ and $\sigma_y^2$ be the two diagonal elements of $D.$ First, suppose that $\sigma_x = \sigma_y.$ Then $u = \sigma_x \eta$ where $\eta \sim N(0,I)$ and \[\E{\frac{u u^\top }{u^\top u}} = \frac{\sigma_x^2}{\sigma_x^2}\E{\frac{\eta \eta^\top }{\eta^\top \eta}} = \frac{1}{n}I = \frac{D^{1/2}}{\Tr{D^{1/2}}}. \] Now suppose that $\sigma_x \neq \sigma_y.$ We calculate the diagonal terms of the covariance matrix by integrating \[\E{\frac{u_1^2}{u_1^2+u_2^2}}=\frac{1}{2\pi \sigma_x\sigma_y}\int_{\mathbb{R}^2} \frac{x^2}{x^2+y^2} e^{-\frac{1}{2}\left( x^2/\sigma_x^2 +y^2/\sigma_y^2 \right) } dxdy. \] Using polar coordinates $x= R \cos(\theta)$ and $y =R \sin(\theta)$ we have \begin{equation}\int_{\mathbb{R}^2} \frac{x^2}{x^2+y^2} e^{-\frac{1}{2}\left( x^2/\sigma_x^2 +y^2/\sigma_y^2 \right) } dxdy = \int_0^{2\pi}\int_{0}^{\infty} R\cos^2(\theta) e^{-\frac{R^2}{2} C(\theta) } dRd\theta,\label{eq:ondstep} \end{equation} where $C(\theta) \eqdef \left( \cos(\theta)^2/\sigma_x^2 +\sin(\theta)^2/\sigma_y^2 \right). $ Note that \begin{equation}\label{eq:Rintfirst} \int_{0}^{\infty} R e^{-\frac{C(\theta)R^2}{2} } dR = \left.-\frac{1}{C(\theta)}e^{-\frac{C(\theta)R^2}{2}} \right|_{0}^{\infty} =\frac{1}{C(\theta)}. \end{equation} This applied in~\eqref{eq:ondstep} gives \begin{align*} \E{\frac{u_1^2}{u_1^2+u_2^2}} &=\frac{1}{2\pi \sigma_x \sigma_y} \int_0^{2\pi}\frac{\cos^2(\theta)}{ \cos(\theta)^2/\sigma_x^2 +\sin(\theta)^2/\sigma_y^2 }d\theta = \frac{b}{\pi} \int_0^{\pi}\frac{\cos^2(\theta)}{ \cos^2(\theta) +b^2\sin^2(\theta)}d\theta, \end{align*} where $b = \sigma_x/\sigma_y.$ Multiplying the numerator and denominator of the integrand by $\sec^4(x)$ gives the integral \[\E{\frac{u_1^2}{u_1^2+u_2^2}} = \frac{b}{\pi} \int_0^{\pi}\frac{\sec^2(\theta)}{ \sec(\theta)^2\left( 1 +b^2\tan^2(\theta) \right) }d\theta.\] Substituting $v=\tan(\theta)$ so that $v^2 +1=\sec^2(\theta)$, $dv =\sec^2(\theta)d \theta$ and using the partial fractions \[\frac{1}{(v^2+1)\left( 1 +b^2v^2 \right)} = \frac{1}{1-b^2}\left( \frac{1}{v^2+1} -\frac{b^2}{b^2v^2+1} \right), \] gives the integral \begin{align} \int \frac{dv}{ (v^2+1)\left( 1 +b^2 v^2 \right) } &= \frac{1}{1-b^2}\left(\arctan(v)-b\arctan(b v) \right) \nonumber\\ &= \frac{1}{1-b^2}\left(\theta-b\arctan(b \tan(\theta))\right). \label{eq:A121} \end{align} To apply the limits of integration, we must take care because of the singularity at $\theta=\pi/2$. For this, consider the limits \[\lim_{\theta \rightarrow (\pi/2)^-} \arctan(b\tan(\theta)) = \frac{\pi}{2}, \qquad \lim_{\theta \rightarrow (\pi/2)^+} \arctan(b\tan(\theta)) = -\frac{\pi}{2}.\] Using this to evaluate~\eqref{eq:A121} on the limits of the interval $[0,\, \pi/2]$ gives \[\lim_{t \rightarrow (\pi/2)^-}\left.\frac{1}{1-b^2}\left(\theta-b\arctan(b \tan(\theta))\right)\right|_0^{t}= \frac{1}{1-b^2}\frac{\pi}{2}(1-b) = \frac{\pi}{2(1+b)}.\] Applying a similar argument for calculating the limits from $\pi/2^+$ to $\pi$, we find \[ \E{\frac{u_1^2}{u_1^2+u_2^2}} =\frac{2b}{\pi} \frac{ \pi}{2(1+b)} =\frac{\sigma_x}{\sigma_y+\sigma_x}.\] Repeating the same steps with $x$ swapped for $y$ we obtain the other diagonal element, which concludes the proof of~\eqref{eq:h98hs98hs9ss}. \end{proof} \chapter[Stochastic Dual Ascent for Finding the Projection of a Vector onto a Linear System]{Stochastic Dual Ascent for Finding the Projection of a Vector onto a Linear System} \chaptermark{SDA for Finding the Projection of a Vector onto a Linear System} \label{ch:SDA} { \epigraph{\emph{Dyfal donc a dyr y garreg.}\\ Tapping~persistently~breaks~the~stone.}{Welsh proverb} \let\clearpage\relax \section{Introduction} } In this chapter we consider the more general problem of finding the projection of a given vector onto the solution space of a linear system. This projection problem includes the problem of determining the least norm solution of a linear system (when the given vector is the zero vector). To solve this projection problem, we develop a new randomized iterative algorithm---{\em stochastic dual ascent (SDA)}. The method is dual in nature: with the dual being a non-strongly concave quadratic maximization problem without constraints. By mapping our dual iterates to primal iterates, we uncover that the SDA method is a dual version of the sketch-and-project method~\eqref{C1eq:xupdate}. We then proceed to strengthen our convergence results established in Chapter~\ref{ch:linear_systems}. First, we do away with the assumption that the system matrix has full column rank that was required to establish convergence through Theorem~\ref{ch:two:theo:normEconv} and~\ref{ch:two:theo:Enormconv} and consider any matrix and consistent linear system. In this more general setting we show that the primal iterates still converge linearly with a convergence rate that is at least as small as the convergence rate established in Chapter~\ref{ch:linear_systems}. Furthermore we give a formula and a tighter lower bound for the convergence rate. We also prove that the same rate of convergence applies to dual function values, primal function values and the duality gap. Unlike traditional iterative methods, SDA converges under virtually no additional assumptions on the system (e.g., rank, diagonal dominance) beyond consistency. In fact, our lower bound improves as the rank of the system matrix drops. When our method specializes to a known algorithm, we either recover the best known rates, or improve upon them. Finally, we show that the framework can be applied to the distributed average consensus problem to obtain an array of new algorithms. The randomized gossip algorithm arises as a special case~\cite{Boyd2006,OlshevskyTsitsiklis2009}. \section{Contributions and Overview} \subsection{The problem:} \begin{equation}\label{eq:Axbx}Ax =b,\end{equation} where $A \in \mathbb{R}^{m \times n}$ and $b \in \mathbb{R}^m$. We shall only assume that the system is {\em consistent}, that is, that there exists $x^*$ for which $Ax^*=b$. Note that we make no assumptions on $n$ or $m$ and all the configurations $m<n,$ $m=n$ and $m>n$ are allowed. While we assume the existence of a solution, we do not assume uniqueness. In situations with multiple solutions, one is often interested in finding a solution with specific properties. For instance, in compressed sensing and sparse optimization, one is interested in finding the least $\ell_1$-norm, or the least $\ell_0$-norm (sparsest) solution. In this chapter we shall focus on the canonical problem of finding the solution of \eqref{eq:Axbx} closest, with respect to a Euclidean distance, to a given vector $c\in \mathbb{R}^n$: \begin{eqnarray} \text{minimize} \quad \ && P(x)\eqdef \tfrac{1}{2}\|x-c\|_B^2 \notag\\ \text{subject to} \quad \ && Ax=b \label{eq:P}\\ && x\in \mathbb{R}^n.\notag \end{eqnarray} where $B$ is an $n\times n$ symmetric positive definite matrix and $\|x\|_B \eqdef \sqrt{x^\top B x}$. By $x^*$ we denote the (necessarily) unique solution of \eqref{eq:P}. Of key importance in this chapter is the {\em dual problem}\footnote{Technically, this is both the Lagrangian and Fenchel dual of \eqref{eq:P}.} to \eqref{eq:P}, namely \begin{eqnarray}\label{eq:Dualfunc} \text{maximize}\quad \ && D(y)\eqdef (b-Ac)^\top y - \tfrac{1}{2}\|A^\top y\|_{B^{-1}}^2\\ \text{subject to} \quad \ && y \in \mathbb{R}^m. \notag \end{eqnarray} Due to the consistency assumption, strong duality holds and we have $P(x^*) = D(y^*)$, where $y^*$ is any dual optimal solution. \subsection{A new family of stochastic optimization algorithms} We propose to solve \eqref{eq:P} via a new method operating in the dual \eqref{eq:Dualfunc}, which we call {\em stochastic dual ascent} (SDA). The iterates of SDA are of the form \begin{equation}\label{eq:methoddual0} y^{k+1} = y^k + S \lambda^k,\end{equation} where $S$ is a random matrix with $m$ rows drawn in each iteration independently from a pre-specified distribution ${\cal D}$, which should be seen as a parameter of the method. In fact, by varying ${\cal D}$, SDA should be seen as a family of algorithms indexed by $\cal D$, the choice of which leads to specific algorithms in this family. By performing steps of the form \eqref{eq:methoddual0}, we are moving in the range space of the random matrix $S$. A key feature of SDA enabling us to prove strong convergence results despite the fact that the dual objective is in general not strongly concave is the way in which the ``stepsize'' parameter $\lambda^k$ is chosen: we choose $\lambda^k$ to be the {\em least-norm} vector for which $D(y^k + S\lambda)$ is maximized in $\lambda$. Plugging this $\lambda^k$ into~\eqref{eq:methoddual0}, we obtain the SDA method: \begin{equation}\label{eq:SDA-compact0} \boxed{\quad y^{k+1} = y^k + S \left(S^\top A B^{-1} A^\top S\right)^\dagger S^\top \left(b - A \left( c+B^{-1}A^\top y^k \right) \right) \quad } \end{equation} To the best of our knowledge, a randomized optimization algorithm with iterates of the {\em general} form \eqref{eq:methoddual0} was not considered nor analyzed before. In the special case when $S$ is chosen to be a random unit coordinate vector, SDA specializes to the {\em randomized coordinate descent method}, first analyzed by Leventhal and Lewis \cite{Leventhal2010}. In the special case when $S$ is chosen as a random column submatrix of the $m\times m$ identity matrix, SDA specializes to the {\em randomized Newton method} of Qu, Fercoq, Richt\'{a}rik and Tak\'{a}\v{c}~\cite{Qu2015}. With the dual iterates $\{y^k\}$ we associate a sequence of primal iterates $\{x^k\}$ as follows: \begin{equation} \label{eq:primaliterates0} x^k \eqdef c + B^{-1}A^\top y^k.\end{equation} In combination with \eqref{eq:SDA-compact0}, this yields the primal iterative process \begin{equation}\label{eq:SDA-primal0} \boxed{\quad x^{k+1} = x^k - B^{-1}A^\top S \left(S^\top A B^{-1} A^\top S\right)^\dagger S^\top \left(A x^k -b \right) \quad } \end{equation} Optimality conditions (see Section~\ref{subsec:OptCond}) imply that if $y^*$ is any dual optimal point, then $c+B^{-1}A^\top y^*$ is necessarily primal optimal and hence equal to $x^*$, the optimal solution of \eqref{eq:P}. Moreover, we have the following useful and insightful correspondence between the quality of the primal and dual iterates (see Proposition~\ref{lem:correspondence}): \begin{equation}\label{eq:iugs8gs}D(y^*) - D(y^k) = \tfrac{1}{2}\|x^k-x^*\|_B^2.\end{equation} Hence, {\em dual convergence in function values is equivalent to primal convergence in iterates.} This work belongs to a growing literature on randomized methods for various problems appearing in linear algebra, optimization and computer science. In particular, relevant methods include sketching algorithms, randomized Kaczmarz, stochastic gradient descent and their variants \cite{Strohmer2009,Needell2010,Drineas2011,hogwild,Zouzias2013a, Needell2012, Ramdas2014, SAG, Takac2013,Johnson2013, Richtarik2015c, S2GD, proxSVRG, SAGA, mS2GD, ZhaoZhang2015,Needell2014, Dai2014, NeedellWard2015, Ma2015, Gower2015,Oswald2015,LiuWright-AccKacz-2016} and randomized coordinate and subspace type methods and their variants \cite{Leventhal2010,Lin:2008:DCDM,ShalevTewari09,Nesterov2012,Wright:ABCRRO, shotgun,Richtarik2014a,Nesterov2011,PCDM,tao2012stochastic,Necoara:Parallel,ICD,Necoara:rcdm-coupled,Richtarik2013a,Hydra2,SDCA,Fercoq-paralleladaboost,Fercoq2013a, Lee2013,Qu2015b,SPCDM,ALPHA,ESO,SPDC,Qu2015,NIPSdistributedSDCA, Shalev-Shwartz2013b,APCG,Csiba2015,Gower2015}. \subsection{The main results} We now describe two complexity theorems which form the core theoretical contribution of this chapter. The results hold for a wide family of distributions ${\cal D}$, which we describe next. \paragraph{Weak assumption on ${\cal D}$.} In our analysis, we only impose a very weak assumption on $\cal D$. In particular, we only assume that the $m\times m$ matrix \begin{equation} \label{eq:H} H \eqdef \mathbf{E}_{S\sim {\cal D}} \left[ S\left(S^\top AB^{-1}A^\top S\right)^{\dagger}S^\top\right]\end{equation} is well defined and nonsingular.\footnote{Note that from Lemma~\ref{lem:pseudoposdef}, the pseudo pseudoinverse of a symmetric positive semidefinite matrix is again symmetric and positive semidefinite. As a result, if the expectation defining $H$ is finite, $H$ is also symmetric and positive semidefinite. Hence, we could equivalently assume that $H$ be positive definite.} Hence, we do not assume that $S$ is chosen from any particular random matrix ensemble. This makes it possible for practitioners to choose the best distribution specific to a particular application. We cast the first complexity result in terms of the primal iterates since solving \eqref{eq:P} is our main focus in this work. Let $\myRange{M}, \Rank{M}$ and $\lambda_{\min}^+(M)$ denote the range space, rank and the smallest nonzero eigenvalue of $M$, respectively. \begin{theorem}[\bf Convergence of primal iterates and of the residual] \label{theo:Enormerror} Assume that the matrix $H$, defined in \eqref{eq:H}, is nonsingular. Fix arbitrary $x^0\in \mathbb{R}^n$. The primal iterates $\{x^k\}$ produced by \eqref{eq:SDA-primal0} converge linearly in expectation to $x^* + t$, where $x^*$ is the optimal solution of the primal problem \eqref{eq:P}, and $t$ is the projection of $x^0-c$ onto $\Null{A}$: \begin{equation}\label{eq:def_of_t}t \eqdef \arg \min_{t'} \left\{ \|x^0-c - t'\|_B \;:\; t'\in \Null{A}\right\}.\end{equation} In particular, for all $k\geq 0$ we have \begin{eqnarray}\label{eq:Enormerror} \quad \text{Primal iterates:} \quad &&\E{\norm{x^{k} - x^{*}- t}_{B}^2}\leq \rho^k \cdot \norm{x^{0} - x^{*} - t}_{B}^2,\\ \label{eq:s98h09hsxxx} \text{Residual:} \quad && \E{\|A x^k-b\|_B} \leq \rho^{k/2} \|A\|_B \|x^0-x^*-t\|_B + \|At\|_B,\end{eqnarray} where $\|A\|_B \eqdef \max \{\|Ax\|_B \;:\; \|x\|_B\leq 1\}$ and \begin{equation}\label{ch:three:eq:rho} \rho \eqdef 1- \lambda_{\min}^+\left(B^{-1/2}A^\top H A B^{-1/2}\right). \end{equation} Furthermore, the convergence rate is bounded by \begin{equation} \label{eq:nubound} 1-\frac{\E{\Rank{S^\top A}}}{\Rank{A}}\leq \rho < 1. \end{equation} \end{theorem} As show shown in Section~\ref{ch:two:sec:RK}, if we let $S$ be a unit coordinate vector chosen at random, $B$ be the identity matrix and set $c=0$, then \eqref{eq:SDA-primal0} reduces to the {\em randomized Kaczmarz (RK)} method proposed and analyzed in a seminal work of Strohmer and Vershynin \cite{Strohmer2009}. Theorem~\ref{theo:Enormerror} implies that RK converges with an exponential rate so long as the system matrix has no zero rows (see Section~\ref{sec:discrete}). To the best of our knowledge, such a result was not previously established: current convergence results for RK assume that the system matrix is full rank~\cite{Ma2015, Ramdas2014}. Not only do we show that the RK method converges to the least-norm solution for any consistent system, but we do so through a single all encompassing theorem covering a wide family of algorithms. Likewise, convergence of block variants of RK has only been established for full column rank~\cite{Needell2012,Needell2014}. Block versions of RK can be obtained from our generic method by choosing $B=I$ and $c=0$, as before, but letting $S$ to be a random column submatrix of the identity matrix. Again, our general complexity bound holds under no assumptions on $A$, as long as one can find $S$ such that $H$ becomes nonsingular. The lower bound \eqref{eq:nubound} says that for a singular system matrix, the number of steps required by SDA to reach an expected accuracy is at best inversely proportional to the rank of $A$. If $A$ has row rank equal to one, for instance, then RK converges in one step (this is no surprise, given that RK projects onto the solution space of a single row, which in this case, is the solution space of the whole system). Our lower bound in this case becomes $0$, and hence is tight. While Theorem~\ref{theo:Enormerror} is cast in terms of the primal iterates, if we assume that $x^0 = c+ B^{-1}A^\top y^0$ for some $y^0\in \mathbb{R}^m$, then an equivalent dual characterization follows by combining \eqref{eq:primaliterates0} and \eqref{eq:iugs8gs}. In fact, in that case we can also establish the convergence of the primal function values and of the duality gap. {\em No such results were previously known. } \begin{theorem}[\bf Convergence of function values]\label{theo:2} Assume that the matrix $H$, defined in \eqref{eq:H}, is nonsingular. Fix arbitrary $y^0\in \mathbb{R}^m$ and let $\{y^k\}$ be the SDA iterates produced by \eqref{eq:SDA-compact0}. Further, let $\{x^k\}$ be the associated primal iterates, defined by \eqref{eq:primaliterates0}, $OPT \eqdef P(x^*)=D(y^*)$, \[U_0 \eqdef \tfrac{1}{2}\|x^0-x^*\|_B^2 \overset{\eqref{eq:iugs8gs}}{=} OPT -D(y^0),\] and let $\rho$ be as in Theorem~\ref{theo:Enormerror}. Then for all $k\geq 0$ we have the following complexity bounds: \begin{eqnarray} \quad \text{Dual suboptimality:} \quad &&\E{OPT - D(y^k)}\leq \rho^k U_0\label{eq:DUALSUBOPT} \\ \quad \text{Primal suboptimality:} \quad &&\E{P(x^k) - OPT}\leq \rho^k U_0 + 2 \rho^{k/2} \sqrt{OPT \times U_0} \label{eq:PRIMALSUBOPT} \\ \text{Duality gap:} \quad && \E{P(x^k) - D(y^k)}\leq 2\rho^k U_0 + 2 \rho^{k/2} \sqrt{OPT\times U_0} \label{eq:GAPSUBOPT} \end{eqnarray} \end{theorem} {\em In our analysis, no error bounds are necessary.} \subsection{Chapter outline} This chapter is structured as follows. Section~\ref{sec:SDA} describes the algorithm in detail, both in its dual and primal form, and establishes several useful identities. In Section~\ref{sec:discrete} we characterize discrete distributions for which our main assumption on $H$ is satisfied. We then specialize our method to several simple discrete distributions to better illustrate the results. We then show in Section~\ref{sec:gossip} how SDA can be applied to design new randomized gossip algorithms. We also show that our framework can recover some standard methods. Theorem~\ref{theo:Enormerror} is proved in Section~\ref{sec:proof} and Theorem~\ref{theo:2} is proved in Section~\ref{sec:proof2}. In Section~\ref{sec:experiments} we perform a simple experiment illustrating the convergence of the randomized Kaczmarz method on rank deficient linear systems. We then summarize in Section~\ref{sec:conclusion}. \section{Stochastic Dual Ascent} \label{sec:SDA} By {\em stochastic dual ascent} (SDA) we refer to a randomized optimization method for solving the dual problem \eqref{eq:Dualfunc} performing iterations of the form \begin{equation}\label{eq:methoddual} y^{k+1} = y^k + S \lambda^k,\end{equation} where $S$ is a random matrix with $m$ rows drawn in each iteration independently from a prespecified distribution. We shall not fix the number of columns of $S$; in fact, we even allow for the number of columns to be random. By performing steps of the form \eqref{eq:methoddual}, we are moving in the range space of the random matrix $S$, with $\lambda^k$ describing the precise linear combination of the columns used in computing the step. In particular, we shall choose $\lambda^k$ from the set \[Q^k \eqdef \arg \max_{\lambda} D(y^k + S\lambda) \overset{\eqref{eq:Dualfunc}}{=} \arg\max_{\lambda} \left\{ (b-Ac)^\top (y^k + S\lambda) - \tfrac{1}{2}\left\|A^\top (y^k + S \lambda)\right\|_{B^{-1}}^2\right\}.\] Since $D$ is bounded above (a consequence of weak duality), this set is nonempty. Since $D$ is a concave quadratic, $Q^k$ consists of all those vectors $\lambda $ for which the gradient of the mapping $\phi_k(\lambda): \lambda \mapsto D(y^k + S\lambda)$ vanishes. This leads to the observation that $Q^k$ is the set of solutions of a random linear system: \[Q^k = \left\{\lambda \in \mathbb{R}^m \;:\; \left(S^\top A B^{-1}A^\top S \right) \lambda = S^\top \left(b - Ac - A B^{-1}A^\top y^k \right) \right\}.\] If $S$ has a small number of columns, this is a small easy-to-solve system. A key feature of our method enabling us to prove exponential error decay despite the lack of strong concavity is the way in which we choose $\lambda^k$ from $Q^k$. In SDA, $\lambda^k$ is chosen to be the least-norm element of $Q^k$, \[\lambda^k \eqdef \arg\min_{\lambda \in Q^k} \|\lambda\|_2.\] Using Lemma~\ref{lem:pseudoleastnorm}, the least-norm solution to the above is given by \begin{equation}\label{eq:lambda_closed_form} \lambda^k = \left(S^\top A B^{-1} A^\top S\right)^\dagger S^\top \left(b - Ac - A B^{-1}A^\top y^k \right). \end{equation} Note that if $S$ has only a few columns, then~\eqref{eq:lambda_closed_form} requires projecting the origin onto a small linear system. The SDA algorithm is obtained by combining \eqref{eq:methoddual} with \eqref{eq:lambda_closed_form}. \begin{algorithm}[!h] \begin{algorithmic}[1] \State \textbf{parameter:} ${\cal D}$ = distribution over random matrices \State Choose $y^0 \in \mathbb{R}^m$ \Comment Initialization \For {$k = 0, 1, 2, \dots$} \State Sample an independent copy $S\sim {\cal D}$ \State $\lambda^k = \left(S^\top A B^{-1} A^\top S\right)^\dagger S^\top \left(b - A c - AB^{-1}A^\top y^k \right)$ \State $y^{k+1} = y^k + S \lambda^k$ \Comment Update the dual variable \EndFor \end{algorithmic} \caption{Stochastic Dual Ascent (SDA)} \label{alg:SDA} \end{algorithm} The method has one parameter: the distribution $\cal D$ from which the random matrices $S$ are drawn. Sometimes, one is interested in finding any solution of the system $Ax=b$, rather than the particular solution described by the primal problem \eqref{eq:P}. In such situations, $B$ and $c$ could also be seen as parameters. \subsection{Optimality conditions} \label{subsec:OptCond} For any $x$ for which $Ax=b$ and for any $y$ we have \[P(x) - D(y) \overset{\eqref{eq:P}+\eqref{eq:Dualfunc}}{=} \tfrac{1}{2}\|x-c\|_B^2 + \tfrac{1}{2}\|A^\top y\|_{B^{-1}}^2 + (c-x)^\top A^\top y \geq 0,\] where the inequality (weak duality) follows from the Fenchel-Young inequality\footnote{Let $U$ be a vector space equipped with an inner product $\langle \cdot, \cdot \rangle : U\times U \to \mathbb{R}$. Given a function $f:U \to \mathbb{R}$, its convex (or Fenchel) conjugate $f^*:U\to \mathbb{R}\cup \{+\infty\}$ is defined by $f^*(v) = \sup_{u \in U} \langle u, v \rangle - f(u)$. A direct consequence of this is the Fenchel-Young inequality, which asserts that $f(u) + f^*(v)\geq \langle u, v\rangle$ for all $u$ and $v$. The inequality in the main text follows by choosing $f(u)=\tfrac{1}{2}\|u\|_B^2$ (and hence $f^*(v)=\tfrac{1}{2}\|v\|^2_{B^{-1}}$), $u=x-c$ and $v=A^\top y$. If $f$ is differentiable, then equality holds if and only if $v=\nabla f(u)$. In our case, this condition is $x=c+B^{-1}A^\top y$. This, together with primal feasibility, gives the optimality conditions \eqref{eq:opt_cond}. For more details on Fenchel duality, see \cite{bookBorweinLewis2006}.}. As a result, we obtain the following necessary and sufficient optimality conditions, characterizing primal and dual optimal points. \begin{proposition} [Optimality conditions] \label{eq:prop_opt_cond}Vectors $x\in \mathbb{R}^n$ and $y\in\mathbb{R}^m$ are optimal for the primal \eqref{eq:P} and dual \eqref{eq:Dualfunc} problems respectively, if and only if they satisfy the following relation \begin{equation} \label{eq:opt_cond}Ax = b, \qquad x = c + B^{-1} A^\top y.\end{equation} \end{proposition} In view of this, it will be useful to define a linear mapping from $\mathbb{R}^m$ to $\mathbb{R}^n$ as follows: \begin{equation}\label{eq:98s98hs}x(y) = c + B^{-1}A^\top y.\end{equation} As an immediate corollary of Proposition~\ref{eq:prop_opt_cond} we observe that for any dual optimal $y^*$, the vector $x(y^*)$ must be primal optimal. Since the primal problem has a unique optimal solution, $x^*$, we must necessarily have \begin{equation}\label{eq:opt_primal}x^* =x(y^*) = c + B^{-1} A^\top y^*.\end{equation} Another immediate corollary of Proposition~\ref{eq:prop_opt_cond} is the following characterization of dual optimality: $y$ is dual optimal if and only if \begin{equation} \label{eq:98hs8h9sss}b - Ac = AB^{-1}A^\top y.\end{equation} Hence, the set of dual optimal solutions is ${\cal Y}^* = (AB^{-1}A^\top)^\dagger (b-Ac) + \Null{AB^{-1}A^\top}$. Since, $\Null{AB^{-1}A^\top} = \Null{A^\top}$ (see Lemma~\ref{lem:WGW}), we have \[{\cal Y}^* = \left(AB^{-1}A^\top\right)^\dagger (b-Ac) + \Null{A^\top}.\] Combining this with \eqref{eq:opt_primal}, we get \[x^* = c + B^{-1}A^\top \left(AB^{-1}A^\top \right)^\dagger(b-Ac).\] \begin{remark} [The dual is also a least-norm problem] Observe that: \begin{enumerate} \item The particular dual optimal point $y^* = (AB^{-1}A^\top)^\dagger (b-Ac)$ is the solution of the following optimization problem: \begin{equation} \label{eq:iusiuh7ss}\min \left\{ \tfrac{1}{2}\|y\|^2_2 \;:\; A B^{-1} A^\top y = b-Ac\right\}.\end{equation} Hence, this particular formulation of the dual problem has the same form as the primal problem: projection onto a linear system. \item If $A^\top A$ is positive definite (which can only happen if $A$ is of full column rank, which means that $Ax=b$ has a unique solution and hence the primal objective function does not matter), and we choose $B=A^\top A$, then the dual constraint \eqref{eq:iusiuh7ss} becomes \[A (A^\top A)^{-1}A^\top y = b - Ac.\] This constraint has a geometric interpretation: we are seeking a vector $y$ whose orthogonal projection onto the column space of $A$ is equal to $b-Ac$. Hence the reformulated dual problem \eqref{eq:iusiuh7ss} is asking us to find the vector $y$ with this property having the least norm. \end{enumerate} \end{remark} \subsection{Primal iterates associated with the dual iterates} With the sequence of dual iterates $\{y^k\}$ produced by SDA we can associate a sequence of primal iterates $\{x^k\}$ using the mapping \eqref{eq:98s98hs}: \begin{equation} \label{eq:primaliterates} x^k \eqdef x(y^k) = c + B^{-1}A^\top y^k.\end{equation} This leads to the following {\em primal version of the SDA method}. \begin{algorithm}[!h] \begin{algorithmic}[1] \State \textbf{parameter:} ${\cal D}$ = distribution over random matrices \State Choose $x^0 \in \mathbb{R}^n$ \Comment Initialization \For {$k = 0, 1, 2, \dots$} \State Sample an independent copy $S\sim {\cal D}$ \State $x^{k+1} = x^k - B^{-1}A^\top S \left(S^\top A B^{-1} A^\top S\right)^\dagger S^\top (A x^k -b )$ \Comment Update the primal variable \EndFor \end{algorithmic} \caption{Primal Version of Stochastic Dual Ascent (SDA-Primal)} \label{alg:SDA-Primal} \end{algorithm} \begin{remark} \label{lem:5shsuss} A couple of observations: \begin{enumerate} \item {\em Self-duality.} If $A$ is positive definite, $c=0$, and if we choose $B=A$, then in view of \eqref{eq:primaliterates} we have $x^k = y^k$ for all $k$, and hence Algorithms~\ref{alg:SDA} and \ref{alg:SDA-Primal} coincide. In this case, Algorithm~\ref{alg:SDA-Primal} can be described as {\em self-dual.} \item {\em Space of iterates.} A direct consequence of the correspondence between the dual and primal iterates \eqref{eq:primaliterates} is the following simple observation (a generalized version of this, which we prove later as Lemma~\ref{lem:error}, will be used in the proof of Theorem~\ref{theo:Enormerror}): Choose $y^0\in \mathbb{R}^m$ and let $x^0 = c + B^{-1}A^\top y^0$. Then the iterates $\{x^k\}$ of Algorithm~\ref{alg:SDA-Primal} are of the form $x^k = c + B^{-1} A^\top y^k$ for some $y^k\in \mathbb{R}^m$. \item {\em Starting point.} While we have defined the primal iterates of Algorithm~\ref{alg:SDA-Primal} via a linear transformation of the dual iterates---see \eqref{eq:primaliterates}---we {\em can}, in principle, choose $x^0$ arbitrarily, thus breaking the primal-dual connection which helped us to define the method. In particular, we can choose $x^0$ in such a way that there does not exist $y^0$ for which $x^0 = c + B^{-1}A^\top y^0$. As is clear from Theorem~\ref{theo:Enormerror}, in this case the iterates $\{x^k\}$ will not converge to $x^*$, but to $x^*+t$, where $t$ is the projection of $x^0-c$ onto the nullspace of $A$. \end{enumerate} \end{remark} It is now clear that the iterates of Algorithm~\ref{alg:SDA-Primal} are the same as the iterates defined by the Random Update viewpoint~\eqref{eq:MP} of the sketch-and-project method~\eqref{ch:two:NF} in Chapter~\ref{ch:linear_systems}. Thus we have uncovered a hidden dual nature of the sketch-and-project method. It is this dual relationship that allows us to formulate and prove convergence of the dual function values and duality gap proven in Theorem~\ref{theo:2}. In particular the duality gap gives a means to measure the distance from the optimality. This certificate of convergence is standard in optimization methods, but seldom appears in the numerical linear algebra literature. \subsection{Relating the quality of the dual and primal iterates} The following simple but insightful result (mentioned in the introduction) relates the ``quality'' of a dual vector $y$ with that of its primal counterpart, $x(y)$. It says that the dual suboptimality of $y$ in terms of function values is equal to the primal suboptimality of $x(y)$ in terms of distance. \begin{proposition}\label{lem:correspondence} Let $y^*$ be any dual optimal point and $y\in \mathbb{R}^m$. Then \[D(y^*) - D(y) = \tfrac{1}{2}\|x(y^*) - x(y)\|_B^2.\] \end{proposition} \begin{proof} Straightforward calculation shows that \begin{eqnarray*} D(y^*) -D(y) &\overset{\eqref{eq:Dualfunc}}{=}& (b-Ac)^\top (y^* - y) - \tfrac{1}{2}(y^*)^\top A B^{-1} A^\top y^* + \tfrac{1}{2}y^\top A B^{-1} A^\top y\\ &\overset{\eqref{eq:98hs8h9sss}}{=}&(y^*)^\top A B^{-1} A^\top (y^* - y) - \tfrac{1}{2}(y^*)^\top A B^{-1} A^\top y^* + \tfrac{1}{2}y^\top A B^{-1} A^\top y\\ &=& \tfrac{1}{2}(y-y^*)^\top AB^{-1} A^\top (y-y^*)\\ &\overset{\eqref{eq:98s98hs}}{=}& \tfrac{1}{2}\|x(y) - x(y^*)\|_B^2. \end{eqnarray*} \end{proof} Applying this result to the sequence $\{(x^k,y^k)\}$ of dual iterates produced by SDA and their corresponding primal images, as defined in \eqref{eq:primaliterates}, we get the identity: \[D(y^*)- D(y^k) = \tfrac{1}{2}\|x^k - x^*\|_B^2.\] Therefore, {\em dual convergence in function values $D(y^k)$ is equivalent to primal convergence in iterates $x^k$}. Furthermore, a direct computation leads to the following formula for the {\em duality gap}: \begin{equation}\label{eq:dualitygap09709709}P(x^k ) - D(y^k) \overset{\eqref{eq:primaliterates}}{=} (AB^{-1}A^\top y^k + Ac - b)^\top y^k = -(\nabla D(y^k) )^\top y^k.\end{equation} Note that computing the gap is significantly more expensive than the cost of a single iteration (in the interesting regime when the number of columns of $S$ is small). Hence, evaluation of the duality gap should generally be avoided. If it is necessary to be certain about the quality of a solution however, the above formula will be useful. The gap should then be computed from time to time only, so that this extra work does not significantly slow down the iterative process. \section{Discrete Distributions} \label{sec:discrete} Both the SDA algorithm and its primal counterpart are generic in the sense that the distribution $\cal D$ is not specified beyond assuming that the matrix $H$ defined in \eqref{eq:H} is well defined and nonsingular. In this section we shall first characterize finite discrete distributions for which $H$ is nonsingular. We then give a few examples of algorithms based on such distributions, and comment on our complexity results in more detail. \subsection{Nonsingularity of $H$ for finite discrete distributions} \label{sec:Hnonsingular} For simplicity, we shall focus on {\em finite discrete} distributions $\cal D$. That is, we set $S = S_i$ with probability $p_i>0$, where $S_1,\dots,S_r$ are fixed matrices (each with $m$ rows). The next theorem gives a necessary and sufficient condition for the matrix $H$ defined in \eqref{eq:H} to be nonsingular. \begin{theorem}\label{thm:H} Let $\cal{D}$ be a finite discrete distribution, as described above. Then $H$ is nonsingular if and only if \[\myRange{[S_1S_1^\top A, \cdots, S_r S_r^\top A]} = \mathbb{R}^m .\] \end{theorem} \begin{proof} Let $K_i = S_i^\top AB^{-1/2}$. In view of the identity $\left(K_i K_i^\top \right)^{\dagger} = (K_i^\dagger )^\top K_i^\dagger$, we can write \[H \overset{\eqref{eq:H}}{=} \sum_{i=1}^r H_i,\] where $H_i = p_i S_i (K_i^\dagger)^\top K_i^\dagger S_i^\top$. Since $H_i$ are symmetric positive semidefinite, so is $H$. Now, it is easy to check that $y^\top H_i y = 0$ if and only if $y \in \Null{H_i}$ (this holds for any symmetric positive semidefinite $H_i$). Hence, $y^\top H y = 0$ if and only if $y \in \cap_i \Null{H_i}$ and thus $H$ is positive definite if and only if \begin{equation} \label{eq:0h09sh0976}\bigcap_{i} \Null{H_i} = \{0\}.\end{equation} In view of Lemma~\ref{lem:WGW}, $\Null{H_i} = \Null{\sqrt{p_i}K_i^\dagger S_i^\top} = \Null{K_i^\dagger S_i^\top}$. Now, $y\in \Null{K_i^\dagger S_i^\top}$ if and only of $S_i^\top y \in \Null{K_i^\dagger} = \Null{K_i^\top} = \Null{A^\top S_i}$. Hence, $\Null{H_i} = \Null{A^\top S_i S_i^\top}$, which means that \eqref{eq:0h09sh0976} is equivalent to \\ $\Null{[S_1S_1^\top A, \cdots, S_r S_r^\top A]^\top} = \{0\}$ \end{proof} \bigskip We have the following corollary.\footnote{We can also prove the corollary directly as follows: The first assumption implies that $S_i^\top A B^{-1} A^\top S_i$ is invertible for all $i$ and that $V \eqdef \mbox{Diag}\left(p_i^{1/2}(S_i^\top A{B^{-1}}A^\top S_i)^{-1/2}\right)$ is nonsingular. It remains to note that \[ H \overset{\eqref{eq:H}}{=} \E{ S\left(S^\top AB^{-1}A^\top S\right)^{-1} S^\top} \\ = \sum_i p_i S_i\left(S_i^\top AB^{-1}A^\top S_i \right)^{-1} S_i^\top = \mathbf{S}V^2 \mathbf{S}^\top. \] } \begin{corollary}\label{cor:09hs09hs} Assume that $S_i^\top A$ has full row rank for all $i$ and that $\mathbf{S} \eqdef [S_1,\ldots, S_r]$ is of full row rank. Then $H$ is nonsingular. \end{corollary} We now give a few illustrative examples: \begin{enumerate} \item \emph{Coordinate vectors.} Let $S_i = e_i$ ($i^{\text{th}}$ unit coordinate vector) for $i=1,2,\dots,r=m$. In this case, $\mathbf{S} = [S_1,\dots,S_m]$ is the identity matrix in $\mathbb{R}^m$, and $S_i^\top A$ has full row rank for all $i$ as long as the rows of $A$ are all nonzero. By Corollary~\ref{cor:09hs09hs}, $H$ is positive definite. \item \emph{Submatrices of the identity matrix.} We can let $S$ be a random column submatrix of the $m\times m$ identity matrix $I$. There are $2^m-1$ such potential submatrices, and we choose $1\leq r \leq 2^m-1$. As long as we choose $S_1,\dots,S_r$ in such a way that each column of $I$ is represented in some matrix $S_i$, the matrix $\mathbf{S}$ will have full row rank. Furthermore, if $S_i^\top A$ has full row rank for all $i$, then by the above corollary, $H$ is nonsingular. Note that if the row rank of $A$ is $r$, then the matrices $S_i$ selected by the above process will necessarily have at most $r$ columns. \item \emph{Count sketch and Count-min sketch.} Many other ``sketching'' matrices $S$ can be employed within SDA, including the count sketch \cite{CountSketch2002} and the count-min sketch \cite{CountMinSketch2005}. In our context (recall that we sketch with the transpose of $S$), $S$ is a count-sketch matrix (resp. count-min sketch) if it is assembled from random columns of $[I,-I]$ (resp $I$), chosen uniformly with replacement, where $I$ is the $m\times m$ identity matrix. \end{enumerate} \subsection{Randomized Kaczmarz is the primal process associated with randomized coordinate ascent } \label{subsec:RKvsRCA} Let $B=I$ (the identity matrix). The primal problem then becomes \begin{eqnarray} \text{minimize} \quad \ && P(x)\eqdef \tfrac{1}{2}\|x-c\|_2^2 \notag\\ \text{subject to} \quad \ && Ax=b \notag\\ && x\in \mathbb{R}^n,\notag \end{eqnarray} and the dual problem is \begin{eqnarray}\notag \text{maximize}\quad \ && D(y)\eqdef (b-Ac)^\top y - \tfrac{1}{2}y^\top A A^\top y\\ \text{subject to} \quad \ && y \in \mathbb{R}^m. \notag \end{eqnarray} \paragraph{Dual iterates.} Let us choose $S=e^i$ (unit coordinate vector in $\mathbb{R}^m$) with probability $p_i>0$ (to be specified later). The SDA method (Algorithm~\ref{alg:SDA}) then takes the form \begin{equation}\label{eq:SDA-compact08986986098} \boxed{\quad y^{k+1} = y^k + \frac{b_i - A_{i}c - A_{i:} A^\top y^k }{\|A_{i:}\|_2^2}e_i \quad } \end{equation} This is the randomized coordinate ascent method applied to the dual problem. In the form popularized by Nesterov~\cite{Nesterov2012}, it takes the form \[y^{k+1} = y^k + \frac{e_i^\top \nabla D(y^k)}{L_i} e_i,\] where $e_i^\top \nabla D(y^k)$ is the $i$th partial derivative of $D$ at $y^k$ and $L_i>0$ is the Lipschitz constant of the $i$th partial derivative, i.e., constant for which the following inequality holds for all $\lambda\in \mathbb{R}$: \begin{equation}\label{eq:s97g98gs} | e_i^\top \nabla D(y + \lambda e_i) - e_i^\top \nabla D(y) | \leq L_i |\lambda|.\end{equation} It can be easily verified that \eqref{eq:s97g98gs} holds with $L_i=\|A_{i:}\|_2^2$ and that $e_i^\top \nabla D(y^k)=b_i - A_{i:}c - A_{i:} A^\top y^k $. \paragraph{Primal iterates.} The associated primal iterative process (Algorithm~\ref{alg:SDA-Primal}) takes the form \begin{equation}\label{eq:SDA-primal009s09us0098} \boxed{\quad x^{k+1} = x^k - \frac{A_{i:} x^k -b_i }{\|A_{i:}\|_2^2} A_{i:}^\top \quad } \end{equation} This is the randomized Kaczmarz method of Strohmer and Vershynin \cite{Strohmer2009}. \paragraph{The rate.} \bigskip Let us now compute the rate $\rho$ as defined in~\eqref{ch:three:eq:rho}. It will be convenient, but {\em not} optimal, to choose the probabilities according to Theorem~\ref{theo:convsingleS}, that is \begin{equation}\label{eq:089h08hs98xx}p_i = \frac{\norm{ A_{i:} }_2^2}{\norm{A}_F^2},\end{equation} where $\|\cdot\|_F$ denotes the Frobenius norm (we assume that $A$ does not contain any zero rows). Since \[H \overset{\eqref{eq:H}}{=} \E{S \left(S^\top A A^\top S \right)^{\dagger}S^\top } = \sum_{i=1}^m p_i \frac{e_i e_i^\top }{\|A_{i:}\|_2^2} \overset{\eqref{eq:089h08hs98xx}}{=} \frac{1}{\norm{A}_F^2}I,\] we have \begin{equation}\label{eq:98hs8h8ss}\rho = 1-\lambda_{\min}^+\left(A^\top H A \right) = 1-\frac{\lambda_{\min}^+\left(A^\top A\right)}{\norm{A}_F^2}. \end{equation} Furthermore, if $r= \Rank{A}$, then in view of \eqref{eq:nubound}, the rate is bounded as \[ 1- \frac{1}{r}\leq \rho <1. \] Assume that $A$ is of rank $r=1$ and let $A= uv^\top$. Then $A^\top A = (u^\top u) v v^\top$, and hence this matrix is also of rank 1. Therefore, $A^\top A$ has a single nonzero eigenvalue, which is equal to its trace. Therefore, $\lambda_{\min}^+(A^\top A) = \Tr{A^\top A} = \|A\|^2_F$ and hence $\rho =0$. Note that the rate $\rho$ reaches its lower bound and the method converges in one step. \paragraph{Remarks.} For randomized coordinate ascent applied to (non-strongly) concave quadratics, rate \eqref{eq:98hs8h8ss} has been established by Leventhal and Lewis \cite{Leventhal2010}. However, to the best of our knowledge, this is the first time this rate has also been established for the randomized Kaczmarz method. We do not only prove this, but show that this is because the iterates of the two methods are linked via a linear relationship. In the $c=0, B=I$ case, and for row-normalized matrix $A$, this linear relationship between the two methods was recently independently observed by Wright~\cite{Wright:CoorDescMethods-survey}. While all linear complexity results for RK we are aware of require full rank assumptions, there exist nonstandard variants of RK which do not require such assumptions, one example being the asynchronous parallel version of RK studied by Liu, Wright and Sridhar \cite{Wright:AsyncPRK}. Finally, no results of the type \eqref{eq:PRIMALSUBOPT} (primal suboptimality) and \eqref{eq:GAPSUBOPT} (duality gap) previously existed for these methods in the literature. \subsection{Randomized block Kaczmarz is the primal process associated with randomized Newton} Let $B=I$, so that we have the same pair of primal dual problems as in Section~\ref{subsec:RKvsRCA}. \paragraph{Dual iterates.} Let us now choose $S$ to be a random column submatrix of the $m\times m$ identity matrix $I$. That is, we choose a random subset $C\subset \{1,2,\dots,m\}$ and then let $S$ be the concatenation of columns $j\in C$ of $I$. We shall write $S=I_C$. Let $p_C$ be the probability that $S = I_C$. Assume that for each $j \in \{1,\dots,m\}$ there exists $C$ with $j \in C$ such that $p_C>0$. Such a random set is called {\em proper}~\cite{Qu2015}. The SDA method (Algorithm~\ref{alg:SDA}) then takes the form \begin{equation}\label{eq:SDA-compact08986986098BLOCK} \boxed{\quad y^{k+1} = y^k + I_C \lambda^k \quad } \end{equation} where $\lambda^k$ is chosen so that the dual objective is maximized (see \eqref{eq:lambda_closed_form}). This is a variant of the {\em randomized Newton method} studied in \cite{Qu2015}. By examining \eqref{eq:lambda_closed_form}, we see that this method works by ``inverting'' randomized submatrices of the ``Hessian'' $AA^\top$. Indeed, $\lambda^k$ is in each iteration computed by solving a system with the matrix $I_C^\top A A^\top I_C$. This is the random submatrix of $A A^\top$ corresponding to rows and columns in $C$. \paragraph{Primal iterates.} In view of the equivalence between Algorithm~\ref{alg:SDA-Primal} and the sketch-and-project method~\eqref{ch:two:NF}, the primal iterative process associated with the randomized Newton method has the form \begin{equation}\label{eq:SDA-primal009s09us0098BLOCK} \boxed{\quad x^{k+1} = \arg \min_{x} \left\{ \|x-x^k\| \;:\; I_C^\top Ax = I_C^\top b \right\} \quad } \end{equation} This method is a variant of the {\em randomized block Kaczmarz} method of Needell \cite{Needell2012}. The method proceeds by projecting the last iterate $x^k$ onto a subsystem of $Ax=b$ formed by equations indexed by the set $C$. \paragraph{The rate.} Provided that $H$ is nonsingular, the shared rate of the randomized Newton and randomized block Kaczmarz methods is \[\rho \overset{\eqref{ch:three:eq:rho} }{=} 1- \lambda_{\min}^+\left(A^\top \E{I_C\left(I_C^\top AA^\top I_C\right)^\dagger I_C^\top} A\right).\] Qu et al \cite{Qu2015} study the randomized Newton method for the problem of minimizing a smooth strongly convex function and prove linear convergence. In particular, they study the above rate in the case when $AA^\top$ is positive definite. Here we show that linear converges also holds for {\em weakly} convex quadratics (as long as $H$ is nonsingular). \subsection{Self-duality for positive definite $A$} If $A$ is positive definite, then we can choose $B=A$. As mentioned before, in this setting SDA is self-dual: $x^k=y^k$ for all $k$. The primal problem then becomes \begin{eqnarray} \text{minimize} \quad \ && P(x)\eqdef \tfrac{1}{2}x^\top A x \notag\\ \text{subject to} \quad \ && Ax=b \notag\\ && x\in \mathbb{R}^n.\notag \end{eqnarray} and the dual problem becomes \begin{eqnarray}\notag \text{maximize}\quad \ && D(y)\eqdef b^\top y - \tfrac{1}{2}y^\top A y\\ \text{subject to} \quad \ && y \in \mathbb{R}^m. \notag \end{eqnarray} Note that the primal objective function does not play any role in determining the solution; indeed, the feasible set contains a single point only: $A^{-1}b$. However, it does affect the iterative process. \paragraph{Primal and dual iterates.} As before, let us choose $S=e^i$ (unit coordinate vector in $\mathbb{R}^m$) with probability $p_i>0$, where the probabilities $p_i$ are arbitrary. Then both the primal and the dual iterates take the form \[ \boxed{\quad y^{k+1} = y^k - \frac{A_{i:} y^k - b_i}{A_{ii}}e_i \quad } \] This is the randomized coordinate ascent method applied to the dual problem. \paragraph{The rate.} If we choose $p_i = A_{ii}/\Tr{A}$, then \[H = \E{S\left(S^\top A S\right)^\dagger S^\top} = \frac{I}{\Tr{A}},\] whence \[\rho \overset{\eqref{ch:three:eq:rho} }{=} 1 - \lambda_{\min}^+ \left( A^{1/2}H A^{1/2}\right) = 1 - \frac{ \lambda_{\min}(A)}{\Tr{A}}.\] It is known that for this problem, randomized coordinate descent applied to the dual problem, with this choice of probabilities, converges with this rate\cite{Leventhal2010}. \section{Application: Randomized Gossip Algorithms} \label{sec:gossip} In this section we apply our method and results to the distributed consensus (averaging) problem. Let $(V,E)$ be a connected network with $|V|=n$ nodes and $|E|=m$ edges, where each edge is an unordered pair $\{i,j\} \in E$ of distinct nodes. Node $i \in V$ stores a private value $c_i\in \mathbb{R}$. The goal of the \emph{distributed consensus problem} is for the network to compute the average of these private values in a distributed fashion \cite{Boyd2006,OlshevskyTsitsiklis2009}. This means that the exchange of information can only occur along the edges of the network. The nodes may represent people in a social network, with edges representing friendship and private value representing certain private information, such as salary. The goal would be to compute the average salary via an iterative process where only friends are allowed to exchange information. The nodes may represent sensors in a wireless sensor network, with an edge between two sensors if they are close to each other so that they can communicate. Private values represent measurements of some quantity performed by the sensors, such as the temperature. The goal is for the network to compute the average temperature. \subsection{Consensus as a projection problem} We now show how one can model the consensus (averaging) problem in the form \eqref{eq:P}. Consider the projection problem \begin{eqnarray} \text{minimize} \quad \ && \tfrac{1}{2}\|x - c\|_2^2 \notag \\ \text{subject to} \quad \ & & x_1=x_2=\cdots = x_n, \label{eq:ohs09hud98yd} \end{eqnarray} and note that the optimal solution $x^*$ must necessarily satisfy \[x^*_i=\bar{c}\eqdef \frac{1}{n}\sum_{i=1}^n c_i,\] for all $i$. There are many ways in which the constraint forcing all coordinates of $x$ to be equal can be represented in the form of a linear system $Ax=b$. Here are some examples: \begin{enumerate} \item {\em Each node is equal to all its neighbours.} Let the equations of the system $Ax=b$ correspond to constraints \[x_i=x_j,\] for $\{i,j\}\in E$. That is, we are enforcing all pairs of vertices joined by an edge to have the same value. Each edge $e\in E$ can be written in two ways: $e = \{i,j\}$ and $e=\{j,i\}$, where $i,j$ are the incident vertices. In order to avoid duplicating constraints, for each edge $e\in E$ we use $e=(i,j)$ to denote an arbitrary but fixed order of its incident vertices $i,j$. We then let $A\in \mathbb{R}^{m\times n}$ and $b=0\in \mathbb{R}^m$, where \begin{equation}\label{eq:89hs87s8ys}(A_{e:})^\top = f_i - f_j,\end{equation} and where $e=(i,j)\in E$, $f_i$ (resp.\ $f_j$) is the $i^{\text{th}}$ (resp.\ $j^{\text{th}}$) unit coordinate vector in $\mathbb{R}^n$. Note that the constraint $x_i=x_j$ is represented only once in the linear system. Further, note that the matrix \begin{equation}\label{eq:Laplacian09709}L = A^\top A \end{equation} is the {\em Laplacian} matrix of the graph $(V,E)$: \[L_{ij} = \begin{cases} d_i & i=j\\ -1 & i\neq j,\,\, (i,j)\in E\\ 0 & \text{otherwise,} \end{cases}\] where $d_i$ is the degree of node $i$. \item {\em Each node is the average of its neighbours.} Let the equations of the system $Ax=b$ correspond to constraints \[x_i = \frac{1}{d_i}\sum_{j \in N(i)} x_j,\] for $i\in V$, where $N(i)\eqdef \left\{ j\in V \,\,: \,\, \{i,j\} \in E\right\}$ is the set of neighbours of node $i$ and $d_i \eqdef |N(i)|$ is the degree of node $i$. That is, we require that the values stored at each node are equal to the average of the values of its neighbours. This corresponds to the choice $b=0$ and \begin{equation}\label{eq:s8h98s78gd}(A_{i:})^\top = f_i - \frac{1}{d_i}\sum_{j \in N(i)}f_j.\end{equation} Note that $A\in \mathbb{R}^{n\times n}$. \item {\em Spanning subgraph.} Let $(V,E')$ be any connected subgraph of $(V,E)$. For instance, we can choose a spanning tree. We can now apply any of the 2 models above to this new graph and either require $x_i=x_j$ for all $\{i,j\}\in E'$, or require the value $x_i$ to be equal to the average of the values $x_j$ for all neighbours $j$ of $i$ in $(V,E')$. \end{enumerate} Clearly, the above list does not exhaust the ways in which the constraint $x_1=\dots=x_n$ can be modeled as a linear system. For instance, we could build the system from constraints such as $x_1 = x_2 + x_4 - x_3$, $x_1 = 5 x_2 - 4 x_7$ and so on. Different representations of the constraint $x_1=\cdots=x_n$, in combination with a choice of $\cal D$, will lead to a wide range of specific algorithms for the consensus problem \eqref{eq:ohs09hud98yd}. Some (but not all) of these algorithms will have the property that communication only happens along the edges of the network, and these are the ones we are interested in. The number of combinations is very vast. We will therefore only highlight two options, with the understanding that based on this, the interested reader can assemble other specific methods as needed. \subsection{Model 1: Each node is equal to its neighbours} Let $b=0$ and $A$ be as in~\eqref{eq:89hs87s8ys}. Let the distribution $\cal D$ be defined by setting $S=e_i$ with probability $p_i>0$, where $e_i$ is the $i^{\text{th}}$ unit coordinate vector in $\mathbb{R}^m$. We have $B=I$, which means that Algorithm~\ref{alg:SDA-Primal} is the randomized Kaczmarz (RK) method \eqref{eq:SDA-primal009s09us0098} and Algorithm~\ref{alg:SDA} is the randomized coordinate ascent method \eqref{eq:SDA-compact08986986098}. Let us take $y^0 = 0$ (which means that $x^0=c$), so that in Theorem~\ref{theo:Enormerror} we have $t=0$, and hence $x^k \to x^*$. The particular choice of the starting point $x^0=c$ in the primal process has a very tangible meaning: for all $i$, node $i$ initially knows value $c_i$. The primal iterative process will dictate how the local values are modified in an iterative fashion so that eventually all nodes contain the optimal value $x^*_i = \bar{c}$. \paragraph{Primal method.} In view of \eqref{eq:89hs87s8ys}, for each edge $e = (i,j)\in E$, we have $\|A_{e:}\|_2^2=2$ and $A_{e:}x^k = x^k_i - x^k_j$. Hence, if the edge $e$ is selected by the RK method, \eqref{eq:SDA-primal009s09us0098} takes the specific form \begin{equation}\label{eq:s98g98gsf66r6fs}\boxed{\quad x^{k+1} = x^k - \frac{x^k_i - x^k_j}{2} (f_i - f_j) \quad }\end{equation} From \eqref{eq:s98g98gsf66r6fs} we see that only the $i^{\text{th}}$ and $j^{\text{th}}$ coordinates of $x^k$ are updated, via \[x^{k+1}_i = x^k_i - \frac{x^k_i - x^k_j}{2} = \frac{x_i^k + x_j^k}{2}\] and \[x^{k+1}_j = x^k_j + \frac{x^k_i - x^k_j}{2} = \frac{x_i^k + x_j^k}{2}.\] Note that in each iteration of RK, a random edge is selected, and the nodes on this edge replace their local values by their average. This is a basic variant of the {\em randomized gossip} algorithm \cite{Boyd2006, ZouziasFreris2009}. \paragraph{Invariance.} Let $f$ be the vector of all ones in $\mathbb{R}^n$ and notice that from \eqref{eq:s98g98gsf66r6fs} we obtain $f^\top x^{k+1} = f^\top x^k$ for all $k$. This means that for all $k\geq 0$ we have the invariance property: \begin{equation}\label{eq:invariance}\sum_{i=1}^n x_i^k = \sum_{i=1}^n c_i.\end{equation} \paragraph{Insights from the dual perspective.} We can now bring new insight into the randomized gossip algorithm by considering the dual iterative process. The dual method \eqref{eq:SDA-compact08986986098} maintains weights $y^k$ associated with the edges of $E$ via the process: \[y^{k+1} = y^k - \frac{A_{e:} (c - A^\top y^k)}{2}e_e,\] where $e$ is a randomly selected edge. Hence, only the weight of a single edge is updated in each iteration. At optimality, we have $x^* =c + A^\top y^*$. That is, for each $i$ \[\delta_i\eqdef \bar{c} - c_i = x_i^* - c_i = (A^\top y^*)_i = \sum_{e\in E} A_{ei} y^*_e,\] where $\delta_i$ is the correction term which needs to be added to $c_i$ in order for node $i$ to contain the value $\bar{c}$. From the above we observe that these correction terms are maintained by the dual method as an inner product of the $i^{\text{th}}$ column of $A$ and $y^k$, with the optimal correction being $\delta_i = A_{:i}^\top y^*$. \paragraph{Rate.} Both Theorem~\ref{theo:Enormerror} and Theorem~\ref{theo:2} hold, and hence we automatically get several types of convergence for the randomized gossip method. In particular, to the best of our knowledge, no primal-dual type of convergence exist in the literature. Equation~\eqref{eq:dualitygap09709709} gives a stopping criterion certifying convergence via the duality gap, which is also new. In view of \eqref{eq:98hs8h8ss} and \eqref{eq:Laplacian09709}, and since $\|A\|_F^2 = 2m$, the convergence rate appearing in all these complexity results is given by \[\rho = 1 - \frac{\lambda_{\min}^+(L)}{2m},\] where $L$ is the Laplacian of $(V,E)$. While it is know that the Laplacian is singular, the rate depends on the smallest nonzero eigenvalue. This means that the number of iterations needed to output an $\epsilon$-solution in expectation scales as $O(\left(2m/\lambda_{\min}^+(L)\right)\log(1/\epsilon))$, i.e., linearly with the number of edges. \subsection{Model 2: Each node is equal to the average of its neighbours} Let $A$ be as in \eqref{eq:s8h98s78gd} and $b=0$. Let the distribution $\cal D$ be defined by setting $S=f_i$ with probability $p_i>0$, where $f_i$ is the $i^{\text{th}}$ unit coordinate vector in $\mathbb{R}^n$. Again, we have $B=I$, which means that Algorithm~\ref{alg:SDA-Primal} is the randomized Kaczmarz (RK) method \eqref{eq:SDA-primal009s09us0098} and Algorithm~\ref{alg:SDA} is the randomized coordinate ascent method \eqref{eq:SDA-compact08986986098}. As before, we choose $y^0=0$, whence $x^0=c$. \paragraph{Primal method.} Observe that $\|A_{i:}\|_2^2 = 1 + 1/d_i$. The RK method \eqref{eq:SDA-primal009s09us0098} applied to this formulation of the problem takes the form \begin{equation} \label{eq:sihiuhd098d7d} \boxed{\quad x^{k+1} = x^k - \frac{x^k_i - \frac{1}{d_i}\sum_{j\in N(i)}x^k_j }{1 + 1/d_i} \left(f_i - \frac{1}{d_i}\sum_{j\in N(i)}f_j \right) \quad } \end{equation} where $i$ is chosen at random. This means that only coordinates in $i\cup N(i)$ get updated in such an iteration, the others remain unchanged. For node $i$ (coordinate $i$), this update is \begin{equation}\label{eq:9g8g98gssdd} x^{k+1}_i = \frac{1}{d_i+1} \left( x_i^k + \sum_{j\in N(i)}x^k_j \right).\end{equation} That is, the updated value at node $i$ is the average of the values of its neighbours and the previous value at $i$. From \eqref{eq:sihiuhd098d7d} we see that the values at nodes $j\in N(i)$ get updated as follows: \begin{equation}\label{eq:98889ff} x^{k+1}_j = x_j^{k} + \frac{1}{d_i+1}\left(x_i^k - \frac{1}{d_i}\sum_{j'\in N(i)}x^k_{j'} \right).\end{equation} \paragraph{Invariance.} Let $f$ be the vector of all ones in $\mathbb{R}^n$ and notice that from \eqref{eq:sihiuhd098d7d} we obtain \[f^\top x^{k+1} = f^\top x^k -\frac{x^k_i - \frac{1}{d_i}\sum_{j\in N(i)}x^k_j }{1 + 1/d_i} \left(1 - \frac{d_i}{d_i} \right)=f^\top x^k, \] for all $k$. It follows that the method satisfies the invariance property \eqref{eq:invariance}. \paragraph{Rate.} The method converges with the rate $\rho$ given by \eqref{eq:98hs8h8ss}, where $A$ is given by \eqref{eq:s8h98s78gd}. If $(V,E)$ is a complete graph (i.e., $m=\tfrac{n(n-1)}{2}$), then $L = \tfrac{(n-1)^2}{n}A^\top A$ is the Laplacian. In this case, $\|A\|_F^2 = \Tr{A^\top A} = \tfrac{n}{(n-1)^2}\Tr{L}=\tfrac{n}{(n-1)^2}\sum_i d_i = \tfrac{n^2}{n-1}$ and hence \[\rho \overset{\eqref{eq:s8h98s78gd}}{=} 1 - \frac{\lambda_{\min}^+(A^\top A)}{\|A\|_F^2} = 1 - \frac{\tfrac{n}{(n-1)^2}\lambda_{\min}^+(L)}{\tfrac{n^2}{n-1}} = 1 - \frac{\lambda_{\min}^+(L)}{2m}.\] \section{Proof of Theorem~\ref{theo:Enormerror}} \label{sec:proof} In this section we prove Theorem~\ref{theo:Enormerror}. We proceed as follows: in Section~\ref{subsec:error} we characterize the space in which the iterates move, in Section~\ref{subsec:inequality} we establish a certain key technical inequality, in Section~\ref{subsec:convergence} we establish convergence of iterates, in Section~\ref{subsec:residual} we derive a rate for the residual and finally, and in Section~\ref{subsec:lower_bound} we establish the lower bound on the convergence rate. \subsection{An error lemma} \label{subsec:error} The following result describes the space in which the iterates move. It is an extension of the observation in item 2 of Remark~\ref{lem:5shsuss} to the case when $x^0$ is chosen arbitrarily. \begin{lemma} \label{lem:error} Let the assumptions of Theorem~\ref{theo:Enormerror} hold. For all $k\geq 0$ there exists $w^k\in \mathbb{R}^m$ such that $ x^k - x^* - t = B^{-1}A^\top w^k$. \end{lemma} \begin{proof} We proceed by induction. Since by definition, $t$ is the projection of $x^0-c$ onto $\Null{A}$ (see \eqref{eq:98hs8htt}), applying Proposition~\ref{prop:decomposition} we know that $x^0 -c = s+ t$, where $s = B^{-1}A^\top \hat{y}^0$ for some $\hat{y}^0\in \mathbb{R}^m$. Moreover, in view of \eqref{eq:opt_primal}, we know that $x^* =c + B^{-1}A^\top y^*$, where $y^*$ is any dual optimal solution. Hence, \[ x^0 - x^* - t = B^{-1}A^\top (\hat{y}^0-y^*).\] Assuming the relationship holds for $k$, we have \begin{eqnarray*} x^{k+1} - x^* - t &\overset{(\text{Alg}~\ref{alg:SDA-Primal})}{=} & \left[x^k - B^{-1}A^\top S (S^\top A B^{-1} A^\top S)^\dagger S^\top (A x^k -b ) \right] - x^*-t\\ &= & \left[x^* + t + B^{-1}A^\top w^k - B^{-1}A^\top S (S^\top A B^{-1} A^\top S)^\dagger S^\top (A x^k -b )\right] - x^* - t\\ &=& B^{-1}A^\top w^{k+1}, \end{eqnarray*} where $w^{k+1} = w^k - S (S^\top A B^{-1} A^\top S)^\dagger S^\top (A x^k -b )$. \end{proof} \subsection{A key inequality} \label{subsec:inequality} The following inequality is of key importance in the proof of the main theorem. \begin{lemma}\label{lem:WGWtight} Let $0\neq W \in \mathbb{R}^{m\times n}$ and $G \in \mathbb{R}^{m\times m}$ be symmetric positive definite. Then the matrix $W^\top G W$ has a positive eigenvalue, and the following inequality holds for all $y\in \mathbb{R}^m$: \begin{equation}\label{eq:WGWtight} y^\top WW^\top G WW^\top y \geq \lambda_{\min}^+(W^\top G W)\|W^\top y\|_2^2. \end{equation} Furthermore, this bound is tight. \end{lemma} \begin{proof} Fix arbitrary $y\in \mathbb{R}^m$. By Lemma~\ref{lem:WGW}, $W^\top y\in \myRange{W^\top G W}$. Since $W$ is nonzero the positive semidefinite matrix $W^\top G W$ is also nonzero, and hence it has a positive eigenvalue. Hence, $\lambda_{\min}^+(W^\top G W)$ is well defined. Let $\lambda_{\min}^+(W^\top G W) = \lambda_{1}\leq \cdots\leq \lambda_{\tau}$ be the positive eigenvalues of $W^\top GW$, with associated orthonormal eigenvectors $q_1,\dots,q_\tau$. We thus have \[W^\top G W = \sum_{i=1}^\tau \lambda_i q_i q_i^\top.\] It is easy to see that these eigenvectors span $\myRange{W^\top GW}$. Hence, we can write $W^\top y = \sum_{i=1}^{\tau} \alpha_i q_i$ and therefore \[y^\top WW^\top G W W^\top y = \sum_{i=1}^{\tau} \lambda_{i}\alpha_i^2 \geq \lambda_1\sum_{i=1}^{\tau} \alpha_i^2 = \lambda_1 \|W^\top y\|_2^2.\] Furthermore this bound is tight, as can be seen by selecting $y$ so that $W^\top y = q_1$. \end{proof} \subsection{Convergence of the iterates} \label{subsec:convergence} Subtracting $x^*+t$ from both sides of the update step of Algorithm~\ref{alg:SDA-Primal}, and letting \begin{equation} \label{ch:three:Z} Z\eqdef A^\top S (S^\top A B^{-1} A^\top S)^\dagger S^\top A,\end{equation} we obtain the identity \begin{equation} \label{eq:fixed0} x^{k+1} - (x^{*}+t) = (I-B^{-1}Z)(x^k-(x^{*}+t)), \end{equation} where we used that $t\in \Null{A}$. Left multiplying~\eqref{eq:fixed0} by $B^{1/2}$ we see that the residual \[r^k \eqdef B^{1/2}(x^{k+1} - (x^{*}+t)),\] satisfies the recurrence \begin{equation} \label{eq:fixedr} r^{k+1} = (I-B^{-1/2}ZB^{-1/2})r^k. \end{equation} In view of $\E{Z} = A^\top H A$, and taking norms and expectations (in $S$) on both sides of~\eqref{eq:fixedr} gives \begin{eqnarray} \E{\norm{r^{k+1}}_{2}^2\, | \, r^k} &=& \E{\norm{(I-{B^{-1/2}}ZB^{-1/2})r^k }^2_2} \nonumber \\ &\overset{\eqref{eq:B12ZB12proj}}{=}& \E{(r^k)^\top (I-{B^{-1/2}}ZB^{-1/2}) r^k} \nonumber\\ & =& \norm{r^k}_{2}^2- (r^k)^\top B^{-1/2} \E{Z} B^{-1/2}r^k \nonumber \\ &=& \norm{r^k}_{2}^2- (r^k)^\top B^{-1/2} A^\top H A B^{-1/2} r^k, \label{eq:theostep1} \end{eqnarray} In view of Lemma~\ref{lem:error}, let $w^k\in \mathbb{R}^{m}$ be such that $r^k = B^{1/2}(B^{-1}A^\top w^k) = B^{-1/2}A^\top w^k.$ Thus \begin{eqnarray} (r^k)^\top B^{-1/2}A^\top H A B^{-1/2}r^k \quad &=&(w^k)^\top AB^{-1}A^\top H A B^{-1}A^\top w^k\nonumber \\ &\overset{(\text{Lemma}~\ref{lem:WGWtight})}{\geq}& \lambda_{\min}^+(B^{-1/2}A^\top H A B^{-1/2})\cdot \|B^{-1/2}A^\top w^k\|^2_2 \nonumber\\ &=& (1- \rho) \cdot\norm{r^k}_2^2,\label{eq:theostep2} \end{eqnarray} where we applied Lemma~\ref{lem:WGWtight} with $W = AB^{-1/2}$ and $G = H$, so that $W^\top GW = B^{-1/2}A^\top H A B^{-1/2}.$ Substituting~\eqref{eq:theostep2} into~\eqref{eq:theostep1} gives $\E{\norm{r^{k+1}}_{2}^2\, | \, r^k} \leq \rho \cdot \norm{r^k}_2^2$. Using the tower property of expectations, we obtain the recurrence \[\E{\norm{r^{k+1}}_{2}^2} \leq \rho \cdot \E{\norm{r^k}_2^2}. \] To prove~\eqref{eq:Enormerror} it remains to unroll the recurrence. \subsection{Convergence of the residual} \label{subsec:residual} We now prove \eqref{eq:s98h09hsxxx}. Letting $V_k = \|x^k-x^*-t\|_B^2$, we have \begin{eqnarray*} \E{\|Ax^k - b\|_B} &=& \E{\|A(x^k-x^*-t) + At\|_B}\\ &\leq &\E{\|A(x^k-x^*-t) \|_B} + \|At\|_B\\ &\leq & \|A\|_B \E{\sqrt{V_k}} + \|At\|_B\\ &\leq &\|A\|_B \sqrt{\E{V_k}} + \|At\|_B\\ &\overset{\eqref{eq:Enormerror}}{\leq}&\|A\|_B \sqrt{\rho^k V_0 } + \|At\|_B, \end{eqnarray*} where in the step preceding the last one we have used Jensen's inequality. \subsection{Proof of the lower bound}\label{subsec:lower_bound} Now we prove~\eqref{eq:nubound}. Using Lemma~\ref{lem:WGW} with $G=H$ and $W=A B^{-1/2}$ gives \begin{align*} \myRange{B^{-1/2}A^\top HAB^{-1/2}}=\myRange{B^{-1/2}A^\top}, \end{align*} from which we deduce that \begin{eqnarray*}\Rank{A} &=& \dim\left(\myRange{A^\top}\right) \\ &=& \dim\left(\myRange{B^{-1/2}A^\top}\right)\\ &=& \dim\left(\myRange{B^{-1/2}A^\top HAB^{-1/2}}\right)\\ &=& \Rank{B^{-1/2} A^\top HA B^{-1/2}}.\end{eqnarray*} Hence, $\Rank{A}$ is equal to the number of nonzero eigenvalues of $B^{-1/2}A^\top HAB^{-1/2}$, from which we immediately obtain the bound \begin{eqnarray*} \Tr{B^{-1/2}A^\top HA B^{-1/2}} &\geq & \Rank{A }\, \lambda_{\min}^+(B^{-1/2}A^\top HAB^{-1/2}). \end{eqnarray*} To conclude the proof, note that $\E{Z} = A^\top H A$ where $Z$ is defined in~\eqref{ch:three:Z}. In order to obtain \eqref{eq:nubound}, it only remains to combine the above inequality with \[ \E{\Rank{S^\top A}} \overset{ \eqref{eq:B12ZB12trace}}{=} = \E{\Tr{B^{-1/2}Z B^{-1/2}}} \\ = \Tr{B^{-1/2}A^\top HAB^{-1/2}}. \] \section{Proof of Theorem~\ref{theo:2}} \label{sec:proof2} In this section we prove Theorem~\ref{theo:2}. We dedicate a subsection to each of the three complexity bounds. \subsection{Dual suboptimality} Since $x^0\in c+\myRange{B^{-1}A^\top}$, we have $t=0$ in Theorem~\ref{theo:Enormerror}, and hence \eqref{eq:Enormerror} says that \begin{equation}\label{eq:s98h98shs}\E{U_k} \leq \rho^k U_0.\end{equation} It remains to apply Proposition~\ref{lem:correspondence}, which says that $U_k = D(y^*)-D(y^k)$. \subsection{Primal suboptimality} Letting $U_k = \tfrac{1}{2}\|x^k- x^*\|_B^2$, we can write \begin{eqnarray} P(x^k) - OPT &=& \tfrac{1}{2}\|x^k - c\|_B^2 - \tfrac{1}{2}\|x^* - c\|_B^2\notag\\ &=& \tfrac{1}{2}\|x^k - x^* + x^* - c\|_B^2 - \tfrac{1}{2}\|x^* - c\|_B^2\notag\\ &=& \tfrac{1}{2}\|x^k- x^*\|_B^2 + (x^k-x^*)^\top B (x^*-c)\notag\\ &\leq & U_k + \|B^{1/2}(x^k-x^*)\|_2 \|B^{1/2}(x^*-c)\|_{2} \notag\\ &=& U_k +\|x^k-x^*\|_B \|x^*-c\|_B \notag \\ &= & U_k + 2 \sqrt{U_k} \sqrt{OPT}.\label{eq:iuhs89h98s6s} \end{eqnarray} By taking expectations on both sides of \eqref{eq:iuhs89h98s6s}, and using Jensen's inequality, we obtain \[\E{P(x^k)-OPT} \leq \E{U_k} + 2\sqrt{OPT} \sqrt{\E{U_k}} \overset{\eqref{eq:s98h98shs}}{\leq} \rho^k U_0 + 2 \rho^{k/2} \sqrt{OPT \times U_0},\] which establishes the bound on primal suboptimality \eqref{eq:PRIMALSUBOPT}. \subsection{Duality gap} Having established rates for primal and dual suboptimality, the rate for the duality gap follows easily: \begin{eqnarray*} \E{P(x^k) - D(y^k)} &=& \E{P(x^k) - OPT + OPT - D(y^k)}\\ &=& \E{P(x^k) - OPT} + \E{OPT - D(y^k)}\\ &\overset{\eqref{eq:DUALSUBOPT} + \eqref{eq:PRIMALSUBOPT}}{=}& 2 \rho^k U_0 + 2 \rho^{k/2} \sqrt{OPT \times U_0}. \end{eqnarray*} \section{Numerical Experiments: Randomized Kaczmarz Method with Rank-Deficient System} \label{sec:experiments} To illustrate some of the novel aspects of our theory, we perform numerical experiments with the Randomized Kaczmarz method~\eqref{eq:SDA-primal009s09us0098} (or equivalently the randomized coordinate ascent method applied to the dual problem~\eqref{eq:Dualfunc}) and compare the empirical convergence to the convergence predicted by our theory. We test several randomly generated rank-deficient systems and compare the evolution of the empirical primal error $\norm{x^k-x^*}_2^2/\norm{x^0-x^*}_2^2$ to the convergence dictated by the rate $\rho = 1-\lambda_{\min}^+\left(A^\top A\right)/\norm{A}_F^2$ given in~\eqref{eq:98hs8h8ss} and the lower bound $1-1/\Rank{A}\leq\rho$. From Figure~\ref{ch:three:fig:rand} we can see that the RK method converges despite the fact that the linear systems are rank deficient. While previous results do not guarantee that RK converges for rank-deficient matrices, our theory does as long as the system matrix has no zero rows. Furthermore, we observe in Figure~\ref{ch:three:fig:rand} that the lower the rank of the system matrix, the faster the convergence of the RK method, and moreover, the closer the empirical convergence is to the convergence dictated by the rate $\rho $ and lower bound on $\rho$. In particular, on the low rank system in Figure~\ref{ch:three:fig:randa}, the empirical convergence is very close to both the convergence dictated by $\rho$ and the lower bound. While on the full rank system in Figure~\ref{ch:three:fig:randd}, the convergence dictated by $\rho$ and the lower bound on $\rho$ are no longer an accurate estimate of the empirical convergence. \begin{figure}[!h] \centering \begin{subfigure}[t]{0.45\textwidth} \centering \includegraphics[width = \textwidth, trim= 80 270 80 280, clip ]{Karczmarz-uniform-random300X300_r_40-bnbdist} \caption{$\Rank{A}= 40$}\label{ch:three:fig:randa} \end{subfigure}% \begin{subfigure}[t]{0.45\textwidth} \centering \includegraphics[width = \textwidth, trim= 80 270 80 280, clip ]{Karczmarz-uniform-random300X300_r_80-bnbdist} \caption{$\Rank{A} = 80$}\label{ch:three:fig:randb} \end{subfigure}\\% \begin{subfigure}[t]{0.45\textwidth} \centering \includegraphics[width = \textwidth, trim= 80 270 80 280, clip ]{Karczmarz-uniform-random300X300_r_160-bnbdist} \caption{$\Rank{A} = 160$}\label{ch:three:fig:randc} \end{subfigure}% \begin{subfigure}[t]{0.45\textwidth} \centering \includegraphics[width = \textwidth, trim= 80 270 80 280, clip ]{Karczmarz-uniform-random300X300_r_300-bnbdist} \caption{$\Rank{A} = 300$}\label{ch:three:fig:randd} \end{subfigure}% \caption{Synthetic MATLAB generated problems. Rank deficient matrix $A~=~\sum_{i=1}^{\Rank{A}} \sigma_i u_i v_i^\top $ where $\sum_{i=1}^{300} \sigma_i u_i v_i^\top =$\texttt{rand}$(300,300)$ is an svd decomposition of a $300\times 300$ uniform random matrix. We repeat each experiment ten times. The blue shaded region is the $90\%$ percentile of relative error achieved in each iteration. }\label{ch:three:fig:rand} \end{figure} \section{Summary} \label{sec:conclusion} We have developed a versatile and powerful algorithmic framework for solving linear systems: {\em stochastic dual ascent (SDA)}. The SDA method finds the projection of a given point, in a fixed but arbitrary Euclidean norm, onto the solution space of the system. Our method is dual in nature, but can also be described in terms of primal iterates via a simple affine transformation of the dual variables. Viewed as a dual method, SDA belongs to a novel class of randomized optimization algorithms: it updates the current iterate by adding the product of a random matrix, drawn independently from a fixed distribution, and a vector. The update is chosen as the best point lying in the random subspace spanned by the columns of this random matrix. While SDA is the first method of this type, particular choices for the distribution of the random matrix lead to several known algorithms: randomized coordinate descent \cite{Leventhal2010} and randomized Kaczmarz \cite{Strohmer2009} correspond to a discrete distribution over the columns of the identity matrix, randomized Newton method \cite{Qu2015} corresponds to a discrete distribution over column submatrices of the identity matrix, and Gaussian descent~\cite{Stich2014} corresponds to the case when the random matrix is a Gaussian vector. We equip the method with several complexity results with the same rate of exponential decay in expectation (aka linear convergence) and establish a tight lower bound on the rate. In particular, we prove convergence of primal iterates, dual function values, primal function values, duality gap and of the residual. The method converges under very weak conditions beyond consistency of the linear system. In particular, no rank assumptions on the system matrix are needed. For instance, randomized Kaczmarz method converges linearly as long as the system matrix contains no zero rows. Further, we show that SDA can be applied to the distributed (average) consensus problem. We recover a standard randomized gossip algorithm as a special case, and show that its complexity is proportional to the number of edges in the graph and inversely proportional to the smallest nonzero eigenvalue of the graph Laplacian. Moreover, we illustrate how our framework can be used to obtain new randomized algorithms for the distributed consensus problem. Our framework extends to several other problems in optimization and numerical linear algebra. For instance, one can apply it to develop new stochastic algorithms for computing the inverse of a matrix and obtain state-of-the art performance for inverting matrices of huge sizes, which is the subject of the next chapter. \chapter{Randomized Matrix Inversion} \chaptermark{Randomized Matrix Inversion} \label{ch:inverse} \section{Introduction} Here we extend our randomized methods for solving linear systems to methods for inverting matrices. Though there exists applications where one needs the explicit inverse of a matrix\footnote{for instance when one needs to store a Schur complement or a projection matrix}, inverting a matrix is seldom required. In contrast, calculating the approximate inverse of a matrix finds many applications. Most notably, calculating an approximate inverse finds applications in preconditioning~\cite{Saad2003} and, if the approximate inverse is guaranteed to be positive definite, then an iterative scheme for inverting a matrix can be used to design variable metric optimization methods. The methods we propose here converge globally and linearly to the inverse matrix and thus are well suited for quickly calculating approximate inverse matrices. When only an approximate inverse is required, then iterative methods are the methods of choice, for they can terminate the iterative process when the desired accuracy is reached. This can be far more efficient than using an all-or-nothing direct method. Furthermore, iterative methods can make use of an initial estimate of the inverse when available. The driving motivation of this work is the need to develop algorithms capable of computing the approximate inverse of very large matrices, where standard techniques take an excessive amount of time or simply fail. In particular, using the sketch-and-project technique we develop a family of randomized/stochastic methods for inverting a matrix with specialized variants maintaining symmetry or positive definiteness of the iterates. All methods in the family converge globally (i.e., from any starting point) and linearly (i.e., the error decays exponentially). As special cases, we obtain stochastic block variants of several quasi-Newton updates, including bad Broyden (BB), good Broyden (GB), Powell-symmetric-Broyden (PSB), Davidon-Fletcher-Powell (DFP) and Broyden-Fletcher-Goldfarb-Shanno (BFGS). To the best of our knowledge, these are the first stochastic versions of quasi-Newton updates. Moreover, this is the first time that randomized quasi-Newton methods are shown to be iterative methods for inverting a matrix. We also offer a new interpretation of the quasi-Newton methods through a Lagrangian dual viewpoint. This new viewpoint uncovers a fundamental link between quasi-Newton updates and approximate inverse preconditioning. We develop an adaptive variant of randomized block BFGS, in which we modify the distribution underlying the stochasticity of the method throughout the iterative process to achieve faster convergence. Through extensive numerical experiments with matrices arising from several applications, we demonstrate that AdaRBFGS is highly competitive when compared with the well established Newton-Schulz and minimal residual methods. In particular, on large-scale problems our method outperforms the standard methods by orders of magnitude. The development of efficient methods for estimating the inverse of very large matrices is a much needed tool for preconditioning and variable metric methods in the advent of the big data era. \subsection{Chapter outline} In Section~\ref{sec:contributions} we summarize the main contributions of this chapter. In Section~\ref{sec:QN} we describe the quasi-Newton methods, which is the main inspiration of our methods. Subsequently, Section~\ref{sec:SIMI-nonsym} describes two algorithms, each corresponding to a variant of the inverse equation, for inverting general square matrices. We also provide insightful dual viewpoints for both methods. In Section~\ref{sec:SIMI-sym} we describe a method specialized to inverting symmetric matrices. Convergence in expectation is examined in Section~\ref{sec:conv}, were we consider two types of convergence: the convergence of i) the expected norm of the error, and the convergence of ii) the norm of the expected error. In Section~\ref{sec:discrete} we specialize our methods to discrete distributions, and comment on how one may construct a probability distribution leading to better complexity rates (i.e., importance sampling), and how to construct an adaptive probability distribution. In Section~\ref{sec:discretemethods} we detail several instantiations of our family of methods, and their resulting convergence rates. We show how via the choice of the parameters of the method, we obtain {\em stochastic block variants} of several well known quasi Newton methods. We also describe the simultaneous randomized Kaczmarz method here. Section~\ref{sec:AdaRBFGS} is dedicated to the development of an adaptive variant of our randomized BFGS method, AdaRBFS, for inverting positive definite matrices. Finally, in Section~\ref{sec:numerical} we show through numerical tests that AdaRBFGS significantly outperforms state-of-the-art iterative matrix inversion methods on large-scale matrices. \subsection{Notation} \label{subsec:notation} Let $I $ denote the $n\times n$ identity matrix. Let \[\dotprod{X,Y}_{F(B)} \eqdef \Tr{X^\top BYB},\] denote the weighted Frobenius inner product, where $X,Y \in \mathbb{R}^{n\times n}$ and $B \in \mathbb{R}^{n\times n}$ is a symmetric positive definite ``weight'' matrix. As the trace is invariant under cyclic permutations, a fact we use repeatedly throughout this chapter, we have \begin{equation}\label{eq:98y988ff} \norm{X}_{F(B)}^2 = \Tr{X^\top BX B} = \Tr{B^{1/2}X^\top B X B^{1/2}} = \norm{B^{1/2}XB^{1/2}}_{F}^2,\end{equation} where we have used the convention $F=F(I)$, since $\|\cdot\|_{F(I)}$ is the standard Frobenius norm. Let $\norm{\cdot}_{2}$ denote the induced operator norm for square matrices defined via \[\|Y\|_{2} \eqdef \max_{\norm{v}_2=1} \norm{Yv}_2.\] Finally, we define the weighted induced norm via \[\norm{Y}_{B}^* \eqdef \norm{B^{1/2}YB^{1/2}}_2. \] \subsection{Previous work}\label{sec:previous} A widely used iterative method for inverting matrices is the Newton-Schulz method~\cite{Schulz1933} introduced in 1933, and its variants which is still subject of ongoing research~\cite{Li2010}. The drawback of the Newton-Schulz methods is that they do not converge for any initial estimate. Instead, an initial estimate $X_0$ such that $\norm{I-X_0A}_2 <1$ is required to guarantee convergence. Though note that such a $X_0$ always exists\footnote{Take for example $X_0 = \alpha A^T$ with $0 < \alpha < 2/\norm{A}_2$.}. The Newton-Schulz method enjoys local quadratic convergence. As has been observed before~\cite{Pan1991}, and as we observe in our numerical experiments, the Newton-Schulz method can experience slow initial convergence before the asymptotic second order convergence rate sets in. This is contrast with the methods we present here that enjoy global linear convergence and a fast initial convergence. Bingham~\cite{Bingham1941} describes a method that uses the characteristic polynomial to recursively calculate the inverse, though it requires the calculating the coefficients of the polynomial when initiated, which is costly, and the method has fallen into disuse. Goldfarb~\cite{Goldfarb1972} uses Broyden's method~\cite{Broyden1965} for iteratively inverting matrices. Our methods include a stochastic variant of Broyden's method. The approximate inverse preconditioning (\emph{AIP}) methods~\cite{Chow1998,Saad2003,Gould1998,Benzi1999} calculate an approximate inverse by minimizing in $X \in \mathbb{R}^{n\times n}$ the residual $\norm{XA-I}_F$ (Frobenius norm). They accomplish this by applying a number of iterations of the steepest descent or minimal residual method. A considerable drawback of the AIP methods, is that the approximate inverses are not guaranteed to be positive definite nor symmetric, even when $A$ is both. A solution to the lack of symmetry is to ``symmetrize'' the estimate between iterations, but then it is difficult to guarantee the quality of the new symmetric estimate. Another solution is to calculate directly a factored form $LL^\top =X$ and minimize in $L$ the residual $\norm{L^\top AL-I}_F$. But now this residual is a non-convex function, and is thus difficult to minimize. A variant of our method naturally maintains symmetry of the iterates. \section{Contributions and Overview} \label{sec:contributions} In this section we describe the main contributions of this chapter. \subsection{New algorithms} \label{subsec:primal} We develop a novel and surprisingly simple family of stochastic algorithms for inverting matrices. The problem of finding the inverse of an $n\times n$ invertible matrix $A$ can be characterized as finding the solution to either one of the two {\em inverse equations}\footnote{One may use other equations uniquely defining the inverse, such as $AXA = A$, but we do not explore these in this thesis.} $AX=I$ or $XA=I.$ Our methods make use of randomized sketching~\cite{Pilanci2014,Gower2015,Pilanci2015,Pilanci2015a} to reduce the dimension of the inverse equations in an iterative fashion. To the best of our knowledge, these are the first stochastic algorithms for inverting a matrix with global complexity rates. In particular, our nonsymmetric method (Algorithm~\ref{alg:asym-row}) is based on the inverse equation $AX=I$, and performs the {\em sketch-and-project} iteration \begin{equation}\label{eq:98hs8s}X_{k+1} = \arg \min_{X\in \mathbb{R}^{n\times n} } \tfrac{1}{2}\norm{X - X_{k}}_{F(B)}^2 \quad \mbox{subject to } \quad S^\top AX =S^\top,\end{equation} where $S\in \mathbb{R}^{n\times q}$ is a random matrix drawn in an i.i.d.\ fashion from a fixed distribution $\cal{D}$, and $B \in \mathbb{R}^{n\times n}$ is symmetric positive definite. The distribution $\cal D$ and matrix $B$ are the parameters of the method. Note that if we choose $q\ll n$, the constraint in the projection problem \eqref{eq:98hs8s} will be of a much smaller dimension than the original inverse equation, and hence the iteration \eqref{eq:98hs8s} will become cheap. In an analogous way, we design a method based on the inverse equation $XA=I$ (Algorithm~\ref{alg:asym-col}). By adding the symmetry constraint $X=X^\top $ to~\eqref{eq:98hs8s}, we obtain Algorithm~\ref{alg:sym}---a specialized method for inverting symmetric matrices capable of maintaining symmetric iterates. \subsection{Dual formulation} Besides the {\em primal formulation} described in Section~\ref{subsec:primal}---\emph{sketch-and-project}---we also provide {\em dual formulations} of all three methods (Algorithms~\ref{alg:asym-row}, \ref{alg:asym-col} and \ref{alg:sym}). For instance, the dual formulation of \eqref{eq:98hs8s} is \begin{equation}\label{eq:98hs8sssds}X_{k+1} = \arg_X \min_{X\in \mathbb{R}^{n\times n}, Y\in \mathbb{R}^{n\times q}} \tfrac{1}{2}\norm{X_{k}-A^{-1}}_{F(B)}^2 \quad \mbox{subject to } \quad X = X_k + B^{-1}A^\top SY^\top .\end{equation} We call the dual formulation \emph{constrain-and-approximate} as one seeks to perform the best approximation of the inverse (with respect to the weighted Frobenius distance) while constraining the search to a random affine space of matrices passing through $X_k$. While the projection \eqref{eq:98hs8sssds} cannot be performed directly since $A^{-1}$ is not known, it can be performed indirectly via the equivalent primal formulation \eqref{eq:98hs8s}. \subsection{Quasi-Newton updates and approximate inverse preconditioning} As we will discuss in Section~\ref{sec:QN}, through the lens of the sketch-and-project formulation, Algorithm~\ref{alg:sym} can be seen as {\em randomized block extension of the quasi-Newton updates}~\cite{Broyden1965,Fletcher1960,Goldfarb1970,Shanno1971}. We distinguish here between quasi-Newton methods, which are algorithms used in optimization, and quasi-Newton updates, which are the {\em matrix-update} rules used in the quasi-Newton methods. Standard quasi-Newton updates work with $q=1$ (``block'' refers to the choice $q>1$) and $S$ chosen in a deterministic way, depending on the sequence of iterates of the underlying optimization problem. To the best of our knowledge, this is the first time stochastic versions of quasi-Newton updates were designed and analyzed. On the other hand, through the lens of the constrain-and-approximate formulation, our methods can be seen as {\em new variants of the approximate inverse preconditioning (\emph{AIP}) methods}~\cite{Chow1998,Saad2003,Gould1998,Benzi1999}. Moreover, the equivalence between these two formulations reveals deep connections between what were before seen as distinct fields: the quasi-Newton and AIP literature. Our work also provides several new insights for {\em deterministic} quasi-Newton updates. For instance, the {\em bad Broyden update}~\cite{Broyden1965,Griewank2012} is a particular best rank-1 update that minimizes the distance to the inverse of $A$ under the Frobenius norm. The {\em BFGS update}~\cite{Broyden1965,Fletcher1960,Goldfarb1970,Shanno1971} can be seen as a projection of $A^{-1}$ onto a space of rank-2 symmetric matrices. To the best of our knowledge, this has not been observed before. \subsection{Complexity} Our framework leads to global linear convergence (i.e., exponential decay) under very weak assumptions on $\cal D$. In particular, we provide an explicit convergence rate $\rho$ for the exponential decay of the norm of the expected error of the iterates (line~2 of Table~\ref{tab:complexity}) and the expected norm of the error (line~3 of Table~\ref{tab:complexity}), where $\rho$ is the same rate provided in Chapter~\ref{ch:linear_systems}, namely \begin{equation}\label{ch:four:eq:rho} \rho = 1- \lambda_{\min}(B^{-1/2}\E{Z}B^{-1/2}), \end{equation} where \[ Z \eqdef A^\top S(S^\top A B^{-1}A^\top S)^{-1}S A^\top .\] This sets our method apart from current methods for inverting matrices that lack global guarantees, such as Newton-Schulz, or the self-conditioning variants of the minimal residual method. \begin{table} \centering \begin{tabular}{|c|c|} \hline & \\ $\E {X_{k+1} -A^{-1}} = \left(I - B^{-1}\E{Z}\right) \E{X_{k+1} - A^{-1}} $ & Theorem 4.1\\ & \\ $\norm{\E {X_{k+1} -A^{-1}}}_{B}^* \leq \rho \; \cdot \; \norm{\E{X_{k+1} - A^{-1}}}_{B}^* $ & Theorem~\ref{theo:normEconv}\\ & \\ $ \E {\norm{X_{k+1} -A^{-1} }_{F(B)}^2 } \leq \rho \;\cdot\; \E{ \norm{X_{k+1} - A^{-1}}_{F(B)}^2}$ & Theorem~\ref{theo:Enormconv}\\ & \\ \hline \end{tabular} \caption{Our main complexity results.} \label{tab:complexity} \end{table} By optimizing an upper bound on~\eqref{ch:four:eq:rho}, we also obtain a new practical importance sampling. This should be contrasted with the optimized rate in Section~\ref{ch:two:sec:optprob} which results in a SDP, which is rarely practical to solve. \subsection{Adaptive randomized BFGS} We develop an additional highly efficient method---adaptive randomized BFGS (AdaRBFGS)---for calculating an approximate inverse of {\em positive definite matrices}. In extensive numeric tests in Section~\ref{sec:numerical} we show that the AdaRBFGS method greatly outperforms the Newton-Schulz and approximate inverse preconditioning methods at obtaining an approximate inverse (with a relative precision of $99\%$). Furthermore, the AdaRBFGS method preserves positive definiteness, a quality not present in previous methods. Therefore, AdaRBFGS can be used to precondition positive definite systems and to design new variable-metric optimization methods. Since the inspiration behind this method comes from the desire to design an {\em optimal adaptive} distribution for $S$ by examining the complexity rate $\rho$, this work also highlights the importance of developing algorithms with explicit convergence rates. \subsection{Extensions} This work opens up many possible avenues for extensions. For instance, new efficient methods could be achieved by experimenting and analyzing through our framework with different sophisticated sketching matrices $S$, such as the Walsh-Hadamard matrix~\cite{Lu2013,Pilanci2014}. Furthermore, our method produces low rank estimates of the inverse and can be adapted to calculate low rank estimates of any matrix. Our methods can be applied to singular matrices, in which case they converge to a particular pseudo-inverse. Our results can be used to push forward work into stochastic variable metric methods. Such as the work by Leventhal and Lewis~\cite{Leventhal2011}, where they present a randomized iterative method for estimating Hessian matrices that converge in expectation with known convergence rates for any initial estimate. Stich et al.\ \cite{Stich2015} use Leventhal and Lewis' method to design a stochastic variable metric method for black-box minimization, with explicit convergence rates, and promising numeric results. We leave these and other extensions to future work. \section{Randomization of Quasi-Newton Updates}\label{sec:QN} Our methods are inspired by, and in some cases can be considered to be, randomized block variants of the quasi-Newton updates. In this section we explain how our algorithms arise naturally from the quasi-Newton setting. Readers familiar with quasi-Newton methods may jump ahead to Section~\ref{subsec:iuh9898}. \subsection{Quasi-Newton methods} A problem of fundamental interest in optimization is the unconstrained minimization problem \begin{equation}\label{eq:opt}\min_{x\in \mathbb{R}^n} f(x),\end{equation} where $f:\mathbb{R}^n\to \mathbb{R}$ is a sufficiently smooth function. Quasi-Newton (QN) methods, first proposed by Davidon in 1959~\cite{Davidon1959}, are an extremely powerful and popular class of algorithms for solving this problem, especially in the regime of moderately large $n$. In each iteration of a QN method, one approximates the function locally around the current iterate $x_k$ by a quadratic of the form \begin{equation}\label{eq:98h98hff}f(x_k+s)\approx f(x_k) + (\nabla f(x_k))^\top s + \frac{1}{2}s^\top B_k s,\end{equation} where $B_k$ is a suitably chosen approximation of the Hessian: $B_k\approx \nabla^2 f(x_k)$. After this, a direction $s_k$ is computed by minimizing the quadratic approximation in $s$, obtaining \begin{equation}\label{eq:QNdirection}s_k = - B_k^{-1}\nabla f(x_k),\end{equation} if the matrix $B_k$ is invertible. The next iterate is then set to \[x_{k+1} = x_k+h_k, \quad h_k =\alpha_k s_k,\] for a suitable choice of stepsize $\alpha_k$, often chosen by a line-search procedure (i.e., by approximately minimizing $f(x_k+ \alpha s_k)$ in $\alpha$). Gradient descent arises as a special case of this process by choosing $B_k$ to be constant throughout the iterations. A popular choice is $B_k=L I$, where $I$ is the identity matrix and $L\in \mathbb{R}_+$ is the Lipschitz constant of the gradient of $f$. In such a case, the quadratic approximation \eqref{eq:98h98hff} is a global upper bound on $f(x_k+s)$, which means that $f(x_k+s_k)$ is guaranteed to be at least as good (i.e., smaller or equal) as $f(x_k)$, leading to guaranteed descent. Newton's method also arises as a special case: by choosing $B_k = \nabla^2 f(x_k)$. These two algorithms are extreme cases on the opposite end of a spectrum. Gradient descent benefits from a trivial update rule for $B_k$ and from cheap iterations due to the fact that no linear systems need to be solved. However, curvature information is largely ignored, which slows down the practical convergence of the method. Newton's method utilizes the full curvature information contained in the Hessian, but requires the computation of the Hessian in each step, which is expensive for large $n$. QN methods aim to find a sweet spot on the continuum between these two extremes. In particular, the QN methods choose $B_{k+1}$ to be a matrix for which the {\em secant equation} is satisfied: \begin{equation}\label{eq:secant}B_{k+1}(x_{k+1}-x_k) = \nabla f(x_{k+1})-\nabla f(x_k).\end{equation} The basic reasoning behind this requirement is the following: if $f$ is a convex quadratic then the Hessian matrix satisfies the secant equation for all pairs of vectors $x_{k+1}$ and $x_k$. If $f$ is not a quadratic, the reasoning is as follows. Using the fundamental theorem of calculus, we have that \[ \left(\int_{0}^1 \nabla^2 f(x_k + t h_k) \; dt\right) (x_{k+1}-x_k)= \nabla f(x_{k+1}) -\nabla f(x_k). \] By selecting $B_{k+1}$ that satisfies the secant equation, we are enforcing that $B_{k+1}$ mimics the action of the integrated Hessian along the line segment joining $x_k$ and $x_{k+1}$. Unless $n=1$, the secant equation \eqref{eq:secant} does not have a unique solution in $B_{k+1}$. All QN methods differ only in which particular solution is used. The formulas transforming $B_k$ to $B_{k+1}$ are called {\em QN updates}. Since these matrices are used to compute the direction $s_k$ via \eqref{eq:QNdirection}, it is often more reasonable to instead maintain a sequence of inverses $X_k = B_k^{-1}$. By multiplying both sides of \eqref{eq:secant} by $X_{k+1}$, we arrive at the {\em secant equation for the inverse:} \begin{equation}\label{eq:secant2}X_{k+1}(\nabla f(x_{k+1})-\nabla f(x_k)) = x_{k+1}-x_k.\end{equation} The most popular classes of QN updates choose $X_{k+1}$ as the closest matrix to $X_k$, in a suitable norm (usually a weighted Frobenius norm with various weight matrices), subject to the secant equation, often with an explicit symmetry constraint: \begin{equation}\label{eq:QN:project_form}X_{k+1} = \arg \min_{X\in \mathbb{R}^{n\times n}} \left\{ \|X-X_k\| \;:\; X y_k = h_k, \; X = X^\top \right\},\end{equation} where $y_k = \nabla f(x_{k+1}) - \nabla f(x_k)$, \subsection{Quasi-Newton updates} Consider now problem \eqref{eq:opt} with the quadratic objective \begin{equation} \label{eq:QN_quadratic}f(x) = \frac{1}{2}x^\top A x - b^\top x + c,\end{equation} where $A$ is an $n\times n$ symmetric positive definite matrix, $b\in \mathbb{R}^n$ and $c\in \mathbb{R}$. Granted, this is not a typical problem for which QN methods would be used by a practitioner. Indeed, the Hessian of $f$ does not change, and hence one {\em does not have to} track it. The problem can simply be solved by setting the gradient to zero, which leads to the system $Ax = b$, the solution being $x_*=A^{-1}b$. As solving a linear system is much simpler than computing the inverse $A^{-1}$, approximately tracking the (inverse) Hessian of $f$ along the path of the iterates $\{x_k\}$---the basic strategy of all QN methods---seems like too much effort for what is ultimately a much simpler problem. However, and this is one of the main insights of this work, instead of viewing QN methods as optimization algorithms, we can alternatively interpret them as iterative algorithms producing a sequence of matrices, $\{B_k\}$ or $\{X_k\}$, hopefully converging to some matrix of interest. In particular, one would hope that if a QN method is applied to the quadratic problem \eqref{eq:QN_quadratic}, with any symmetric positive definite initial guess $X_0$, then the sequence $\{X_k\}$ converges to $A^{-1}$. For $f$ given by \eqref{eq:QN_quadratic}, the QN updates of the minimum distance variety given by \eqref{eq:QN:project_form} take the form \begin{equation}\label{eq:QN:project_form2}X_{k+1} = \arg \min_{X\in \mathbb{R}^{n\times n}} \left\{ \|X-X_k\| \;:\; X A h_k = h_k , \; X = X^\top \right\}.\end{equation} \subsection{Randomized quasi-Newton updates}\label{subsec:iuh9898} While the motivation for our work comes from optimization, having arrived at the update \eqref{eq:QN:project_form2}, we can dispense of some of the implicit assumptions and propose and analyze a wider class of methods. In particular, in this chapter we analyze a large class of {\em randomized algorithms} of the type \eqref{eq:QN:project_form2}, where the vector $h_k$ is replaced by a random matrix $S$ and $A$ is \emph{any} invertible, and not necessarily symmetric or positive definite matrix. This constitutes a randomized block extension of the QN updates. \section{Inverting Nonsymmetric Matrices} \label{sec:SIMI-nonsym} In this chapter we are concerned with the development and complexity analysis of a family of stochastic algorithms for computing the inverse of a nonsingular matrix $A\in \mathbb{R}^{n\times n}$. The starting point in the development of our methods is the simple observation that the inverse $A^{-1}$ is the (unique) solution of a linear matrix equation, which we shall refer to as {\em inverse equation}: \begin{equation} \label{eq:inverse_equations} AX = I.\end{equation} Alternatively, one can use the inverse equation $XA = I$ instead. Since \eqref{eq:inverse_equations} is difficult to solve directly, our approach is to iteratively solve a small randomly relaxed version of \eqref{eq:inverse_equations}. That is, we choose a random matrix $S\in \mathbb{R}^{n \times q}$, with $q\ll n$, and instead solve the following {\em sketched inverse equation}: \begin{equation}\label{eq:inverse_eq_sketched}S^\top AX = S^\top .\end{equation} If we base the method on the second inverse equation, the sketched inverse equation $XA S = S$ should be used instead. Note that $A^{-1}$ satisfies \eqref{eq:inverse_eq_sketched}. If $q\ll n$, the sketched inverse equation is of a much smaller dimension than the original inverse equation, and hence easier to solve. However, the equation will no longer have a unique solution and in order to design an algorithm, we need a way of picking a particular solution. Our algorithm defines $X_{k+1}$ to be the solution that is closest to the current iterate $X_k$ in a weighted Frobenius norm. This is repeated in an iterative fashion, each time drawing $S$ independently from a fixed distribution $\cal D$. The distribution $\cal D$ and the matrix $B$ can be seen as parameters of our method. The flexibility of being able to adjust $\cal D$ and $B$ is important: by varying these parameters we obtain various specific instantiations of the generic method, with varying properties and convergence rates. This gives the practitioner the flexibility to adjust the method to the structure of $A$, to the computing environment and so on. \subsection{Projection viewpoint: sketch-and-project} The next iterate $X_{k+1}$ is the nearest point to $X_k$ that satisfies a \emph{sketched} version of the inverse equation: \begin{align} \boxed{ X_{k+1} = \arg \min_{X\in \mathbb{R}^{n\times n}} \frac{1}{2} \norm{X - X_{k}}_{F(B)}^2 \quad \mbox{subject to } \quad S^\top AX =S^\top } \label{eq:NF} \end{align} In the special case when $S=I$, the only such matrix is the inverse itself, and \eqref{eq:NF} is not helpful. However, if $S$ is ``simple'', \eqref{eq:NF} will be easy to compute and the hope is that through a sequence of such steps, where the matrices $S$ are sampled in an i.i.d. fashion from some distribution, $X_k$ will converge to $A^{-1}$. Alternatively, we can sketch the equation $XA=I$ and project onto $XAS=S$: \begin{equation}\label{eq:NFcols} \boxed{X_{k+1} = \arg \min_{X\in \mathbb{R}^{n\times n}} \frac{1}{2} \norm{X - X_{k}}_{F(B)}^2 \quad \mbox{subject to } \quad XAS =S}\end{equation} While the method~\eqref{eq:NF} sketches the rows of $A$, the method~\eqref{eq:NF} sketches the columns of $A.$ Thus we refer to~\eqref{eq:NF} as the row variant and to~\eqref{eq:NFcols} as the column variant. The two variants~\eqref{eq:NF} and~\eqref{eq:NFcols} both converge to the inverse of $A$, as will be established in Section~\ref{sec:conv}. If $A$ is singular, then the iterates of~\eqref{eq:NFcols} converge to the left inverse, while the iterates of~\eqref{eq:NF} converge to the right inverse, an observation we leave to future work. \subsection{Optimization viewpoint: constrain-and-approximate} The row sketch-and-project method can be cast in an apparently different yet equivalent viewpoint: \begin{align} \boxed{X_{k+1} = \arg_X \min_{X\in \mathbb{R}^{n\times n}, Y\in \mathbb{R}^{n\times q}} \frac{1}{2}\norm{X - A^{-1}}_{F(B)}^2 \quad \mbox{subject to } \quad X=X_{k} +B^{-1}A^\top SY^\top }\label{eq:RF} \end{align} In this viewpoint, at each iteration~\eqref{eq:RF}, we select a random affine space that passes through $X_k.$ After that, we select the point in this space that is as close as possible to the inverse. This random search space is special in that, independently of the input pair $(B,S)$ we can efficiently compute the projection of $A^{-1}$ onto this space, without knowing $A^{-1}$ explicitly. The column variant~\eqref{eq:NFcols} also has an equivalent constrain-and-approximate formulation: \begin{align} \boxed{X_{k+1} = \arg_X \min_{X\in \mathbb{R}^{n\times n}, Y\in \mathbb{R}^{n\times q}} \frac{1}{2}\norm{X - A^{-1}}_{F(B)}^2 \quad \mbox{subject to } \quad X=X_{k} +YS^\top A^\top B^{-1}}\label{eq:RFcols} \end{align} These two variants~\eqref{eq:RF} and~\eqref{eq:RFcols} can be viewed as new variants of the approximate inverse preconditioner (AIP) methods~\cite{Benzi1999,Gould1998,Kolotilina1993,Huckle2007}. The AIP methods are a class of methods for computing approximate inverses of $A$ by minimizing $\norm{XA-I}_F$ via iterative optimization algorithms. In particular, the AIP methods use variants of the steepest descent or a minimal residual method to minimize $\norm{XA-I}_F$. The idea behind the AIP methods is to minimize the distance of $X$ from $A^{-1}$ in some sense. Our variants do just that, but under a weighted Frobenius norm. Furthermore, our methods project onto a randomly generated affine space instead of employing steepest descent of a minimal residual method. \subsection{Equivalence} We now prove that~\eqref{eq:NF} and~\eqref{eq:NFcols} are equivalent to~\eqref{eq:RF} and~\eqref{eq:RFcols}, respectively, and give their explicit solution. \begin{theorem}\label{theo:NFRF} The viewpoints~\eqref{eq:NF} and~\eqref{eq:RF} are equivalent to~\eqref{eq:NFcols} and~\eqref{eq:RFcols}, respectively. Furthermore, if $S$ has full column rank, then the explicit solution to~\eqref{eq:NF} is \begin{equation} \boxed{ X_{k+1} = X_{k} +B^{-1}A^\top S(S^\top A B^{-1}A^\top S)^{-1}S^\top (I-AX_{k})} \label{eq:Xupdate} \end{equation} and the explicit solution to~\eqref{eq:NFcols} is \begin{equation} \boxed{X_{k+1} = X_{k} +(I-X_{k}A^\top )S(S^\top A^\top B^{-1}A S)^{-1}S^\top A^\top B^{-1}} \label{eq:Xupdatecols} \end{equation} \end{theorem} \begin{proof} We will prove all the claims for the row variant, that is, we prove that~\eqref{eq:NF} are~\eqref{eq:RF} equivalent and that their solution is given by~\eqref{eq:Xupdate}. The remaining claims, that~\eqref{eq:NFcols} are~\eqref{eq:RFcols} are equivalent and that their solution is given by~\eqref{eq:Xupdatecols}, follow with analogous arguments. It suffices to consider the case when $B=I$, as we can perform a change of variables to recover the solution for any $B$. Indeed, in view of \eqref{eq:98y988ff}, with the change of variables \begin{equation}\label{eq:varchangeW}\hat{X}\eqdef B^{1/2}X B^{1/2}, \quad \hat{X}_k\eqdef B^{1/2}X_kB^{1/2}, \quad \hat{A} \eqdef B^{-1/2}AB^{-1/2} \quad \mbox{and} \quad \hat{S} \eqdef B^{1/2}S, \end{equation} \eqref{eq:NF} becomes \begin{align} \min_{\hat{X} \in \mathbb{R}^{n\times n}} \frac{1}{2} \norm{\hat{X} - \hat{X}_{k}}_{F}^2 \quad \mbox{subject to } \quad \hat{S}^\top \hat{A}\hat{X} =\hat{S}^\top . \label{eq:NFbar} \end{align} Moreover, if we let $\hat{Y} = B^{1/2}Y$, then ~\eqref{eq:RF} becomes \begin{align} \min_{\hat{X} \in \mathbb{R}^{n\times n}, \hat{Y}\in \mathbb{R}^{n\times q}} \frac{1}{2} \norm{\hat{X} - \hat{A}^{-1}}_{F}^2 \quad \mbox{subject to } \quad \hat{X} =\hat{X}_k + \hat{A}^\top \hat{S} \hat{Y}^\top . \label{eq:RFbar} \end{align} By substituting the constraint in~\eqref{eq:RFbar} into the objective function, then differentiating to find the stationary point, we obtain that \begin{equation} \label{eq:Xupdatebar} \hat{X} = \hat{X}_k +\hat{A}^\top \hat{S}(\hat{S}^\top \hat{A}\hat{A}^\top \hat{S})^{-1}\hat{S }^\top (I-\hat{A}\hat{X}_k),\end{equation} is the solution to~\eqref{eq:RFbar}. After changing the variables back using~\eqref{eq:varchangeW}, the update~\eqref{eq:Xupdatebar} becomes~\eqref{eq:XZupdate}. Now we prove the equivalence of~\eqref{eq:NFbar} and~\eqref{eq:RFbar} using Lagrangian duality. The sketch-and-project viewpoint~\eqref{eq:NFbar} has a convex quadratic objective function with linear constraints, thus strong duality holds. Introducing Lagrangian multiplier $\hat{Y} \in \mathbb{R}^{n\times q}$, the Langrangian dual of~\eqref{eq:NFbar} is given by \begin{equation}\label{eq:lagdual} L(\hat{X},\hat{Y})= \frac{1}{2}\norm{\hat{X}-\hat{X}_k}_{F}^2 - \dotprod{{\hat{Y}}^\top ,\hat{S}^\top \hat{A}(\hat{X}-\hat{A}^{-1})}_{F}. \end{equation} Clearly \[\eqref{eq:NFbar} =\min_{X\in\mathbb{R}^{n\times n}} \max_{\hat{Y} \in\mathbb{R}^{n\times q}} L(\hat{X}, \hat{Y}). \] We will now prove that \[\eqref{eq:RFbar} =\max_{\hat{Y} \in\mathbb{R}^{n\times q}} \min_{X\in\mathbb{R}^{n\times n}} L(\hat{X}, \hat{Y}), \] thus proving that~\eqref{eq:NFbar} and~\eqref{eq:RFbar} are equivalent by strong duality. Differentiating the Lagrangian in $\hat{X}$ and setting to zero gives \begin{equation}\label{eq:primopt1} \hat{X} = \hat{X}_k +\hat{A}^\top \hat{S}{\hat{Y}}^\top . \end{equation} Substituting back into~\eqref{eq:lagdual} gives \begin{align*} L(\hat{X},\hat{Y}) &= \frac{1}{2}\norm{\hat{A}^\top \hat{S}\hat{Y}^\top }_{F}^2 - \dotprod{\hat{A}^\top \hat{S}{\hat{Y}}^\top , \hat{X}_k +\hat{A}^\top \hat{S}{\hat{Y}}^\top -\hat{A}^{-1}}_{F} \\ &= - \frac{1}{2}\norm{\hat{A}^\top \hat{S}{\hat{Y}}^\top }_{F}^2 - \dotprod{\hat{A}^\top \hat{S}{\hat{Y}}^\top , \hat{X}-\hat{A}^{-1} }_{F}. \end{align*} Adding $\pm \frac{1}{2}\norm{\hat{X}_k -\hat{A}^{-1}}_{F}^2$ to the above gives \[L(\hat{X},\hat{Y}) = -\frac{1}{2}\norm{\hat{A}^\top \hat{S}{\hat{Y}}^\top +\hat{X}_k-\hat{A}^{-1}}_{F}^2+\frac{1}{2}\norm{\hat{X}_k -\hat{A}^{-1}}_{F}^2.\] Finally, substituting~\eqref{eq:primopt1} into the above, minimizing in $\hat{X}$ then maximizing in $\hat{Y}$, and dispensing of the term $\frac{1}{2}\norm{\hat{X}_k -\hat{A}^{-1}}_{F}^2$ as it does not depend on $\hat{Y}$ nor $\hat{X}$, we have that the dual problem is \[ \max_{\hat{Y}}\min_{\hat{X}} L(\hat{X},\hat{Y}) =\min_{\hat{X},\hat{Y}} \frac{1}{2}\norm{\hat{X}-\hat{A}^{-1}}_{F}^2 \quad \mbox{subject to} \quad \hat{X} =\hat{X}_k+ \hat{A}^\top \hat{S}{\hat{Y}}^\top .\] It now remains to change variables using~\eqref{eq:varchangeW} and set $Y = B^{-1/2}\hat{Y}$ to obtain~\eqref{eq:RF}. \end{proof} Based on Theorem~\ref{theo:NFRF}, we can summarize the methods described in this section as Algorithm~\ref{alg:asym-row} and Algorithm~\ref{alg:asym-col}. \begin{algorithm}[!t] \begin{algorithmic}[1] \State \textbf{input:} invertible matrix $A \in \mathbb{R}^{n\times n}$ \State \textbf{parameters:} ${\cal D}$ = distribution over random matrices; positive definite matrix $B\in \mathbb{R}^{n\times n}$ \State \textbf{initialize:} arbitrary square matrix $X_0\in \mathbb{R}^{n\times n}$ \For {$k = 0, 1, 2, \dots$} \State Sample an independent copy $S\sim {\cal D}$ \State Compute $\Lambda = S(S^\top A B^{-1}A^\top S)^{-1}S^\top $ \State $X_{k+1} = X_{k} + B^{-1}A^\top \Lambda (I-AX_{k})$ \Comment This is equivalent to \eqref{eq:NF} and \eqref{eq:RF} \EndFor \State \textbf{output:} last iterate $X_k$ \end{algorithmic} \caption{Stochastic Iterative Matrix Inversion (SIMI) -- nonsymmetric row variant} \label{alg:asym-row} \end{algorithm} \begin{algorithm}[!h] \begin{algorithmic}[1] \State \textbf{input:} invertible matrix $A \in \mathbb{R}^{n\times n}$ \State \textbf{parameters:} ${\cal D}$ = distribution over random matrices; positive definite matrix $B\in \mathbb{R}^{n\times n}$ \State \textbf{initialize:} arbitrary square matrix $X_0\in \mathbb{R}^{n\times n}$ \For {$k = 0, 1, 2, \dots$} \State Sample an independent copy $S\sim {\cal D}$ \State Compute $\Lambda = S(S^\top A^\top B^{-1}A S)^{-1}S^\top $ \State $X_{k+1} = X_{k} + (I-X_{k} A^\top ) \Lambda A^\top B^{-1}$ \Comment This is equivalent to \eqref{eq:NFcols} and \eqref{eq:RFcols} \EndFor \State \textbf{output:} last iterate $X_k$ \end{algorithmic} \caption{Stochastic Iterative Matrix Inversion (SIMI) -- nonsymmetric column variant} \label{alg:asym-col} \end{algorithm} The explicit formulas~\eqref{eq:Xupdate} and~\eqref{eq:Xupdatecols} for \eqref{eq:NF} and~\eqref{eq:NFcols} allow us to efficiently implement these methods, and facilitate convergence analysis. In particular, we can now see that the convergence analysis of~\eqref{eq:Xupdatecols} will follow trivially from analyzing~\eqref{eq:Xupdate}. This is because~\eqref{eq:Xupdate} and~\eqref{eq:Xupdatecols} differ only in terms of a transposition. That is, transposing~\eqref{eq:Xupdatecols} gives \[ X_{k+1}^\top = X_{k}^\top +B^{-1}AS(S^\top A^\top B^{-1}A S)^{-1}S^\top (I-A^\top X_{k}^\top ),\] which is the solution to the row variant of the sketch-and-project viewpoint but where the equation $A^\top X^\top = I$ is sketched instead of $AX=I.$ Thus, since the weighted Frobenius norm is invariant under transposition, it suffices to study the convergence of~\eqref{eq:Xupdate}, then the convergence of~\eqref{eq:Xupdatecols} follows by simply swapping the role of $A$ for $A^\top .$ We collect this observation is the following remark. \begin{remark}\label{rem:alg2conv} The expression for the rate of convergence of Algorithm~\ref{alg:asym-col} is the same as the expression for the rate of convergence of Algorithm~\ref{alg:asym-row}, but with every occurrence of $A$ swapped for $A^\top .$ \end{remark} \subsection{Relation to multiple linear systems} \label{sec:relationlinear} Any iterative method for solving linear systems can be applied to the $n$ linear systems that define the inverse through $AX=I$ to obtain an approximate inverse. Though not all methods for solving linear systems can be applied to solve these $n$ linear systems simultaneously, that is calculating each column of $X$ simultaneously, which is necessary for an efficient matrix inversion method. The sketch-and-project methods we described in Chapters~\ref{ch:linear_systems} and~\ref{ch:SDA} can be easily and efficiently generalized to inverting a matrix, and the resulting method is equivalent to our row variant method~\eqref{eq:NF} and~\eqref{eq:RF}. To show this, we perform the change of variables $\hat{X}_k= X_kB^{1/2},$ $\hat{A} =B^{-1/2} A$ and $\hat{S} = B^{1/2}S$ then~\eqref{eq:NF} becomes \[\hat{X}_{k+1} \eqdef X_{k+1}B^{1/2} = \arg \min_{\hat{X} \in \mathbb{R}^{n\times n}} \frac{1}{2} \norm{B^{1/2}(\hat{X} - \hat{X}_{k})}_{F}^2 \quad \mbox{subject to } \quad \hat{S}^\top \hat{A}\hat{X}=\hat{S}^\top .\] The above is a separable problem and each column of $\hat{X}_{k+1}$ can be calculated separately. Let $\hat{x}_{k+1}^i$ be the $i$th column of $\hat{X}_{k+1}$ which can be calculated through \[\hat{x}_{k+1}^i = \arg \min_{\hat{x} \in \mathbb{R}^n}\frac{1}{2} \norm{B^{1/2}(\hat{x} - \hat{x}_{k}^i)}_{2}^2 \quad \mbox{subject to } \quad \hat{S}^\top \hat{A}\hat{x} =\hat{S}^\top e_i. \] The above is exactly an iteration of the sketch-and-project method~\eqref{ch:two:NF} applied to the system $\hat{A}\hat{x} = e_i.$ Thus the convergence results established in~\cite{Gower2015} carry over to our row variant~\eqref{eq:NF} and~\eqref{eq:RF}. In particular, the theory in~\cite{Gower2015} proves that the expected norm difference of each column of $B^{1/2}X_k$ converges to $B^{1/2}A^{-1}$ with rate $\rho$ as defined in~\eqref{ch:four:eq:rho}. This equivalence breaks down when we impose additional matrix properties through constraints, such as symmetry. \section{Inverting Symmetric Matrices} \label{sec:SIMI-sym} When $A$ is symmetric, it may be useful to maintain symmetry in the iterates, in which case the nonsymmetric methods---Algorithms~\ref{alg:asym-row} and \ref{alg:asym-col}---have an issue, as they do not guarantee that the iterates are symmetric. However, we can modify~\eqref{eq:NF} by adding a symmetry constraint. The resulting \emph{symmetric} method naturally maintains symmetry in the iterates. \subsection{Projection viewpoint: sketch-and-project} The new iterate $X_{k+1}$ is the result of projecting $X_k$ onto the space of matrices that satisfy a sketched inverse equation and that are also symmetric, that is \begin{align} \boxed{ X_{k+1} = \arg \min_{X\in \mathbb{R}^{n\times n}}\frac{1}{2} \norm{X - X_{k}}_{F(B)}^2 \quad \mbox{subject to }\quad S^\top AX =S^\top , \quad X = X^\top } \label{eq:NFsym} \end{align} See Figure~\ref{fig:proj} for an illustration of the symmetric update~\eqref{eq:NFsym}. This viewpoint can be seen as a randomized block version of the quasi-Newton methods~\cite{Goldfarb1970,Greenstadt1969}, as detailed in Section~\ref{sec:QN}. The flexibility in using a weighted norm is important for choosing a norm that better reflects the geometry of the problem. For instance, when $A$ is symmetric positive definite, it turns out that $B =A$ results in a good method. This added freedom of choosing an appropriate weighting matrix has proven very useful in the quasi-Newton literature, in particular, the highly successful BFGS method~\cite{Broyden1965,Fletcher1960,Goldfarb1970,Shanno1971} selects $B$ as an estimate of the Hessian matrix. \begin{figure} \centering \resizebox{0.6\textwidth}{!}{ \begin{tikzpicture}[>=triangle 45,font=\sffamily, ] \node (planes) at (0,0) {\includegraphics[width =10cm]{planes.jpg}}; \node (sym) at (3.5cm,-1cm) {$\{X \; : \; X= X^\top \}$}; \nodepoint{Prevp}{3.7cm}{-2.1cm}; \node[inner sep =0mm, outer sep =0mm, right = 0pt of Prevp] (Prev) {$X_{k}$}; \node (action) at (2.5cm,2.5cm) {$\left\{ X \; : \; S^\top AX= S^\top \right\}$}; \nodepoint{Nextp}{-3.1cm}{-2cm}; \node[inner sep =0mm, outer sep =0mm, left = 0pt of Nextp] (Next) {$X_{k+1}$}; \nodepoint{Invp}{-2.9cm}{-1cm}; \node[inner sep =0mm, outer sep =0mm, left = 0pt of Invp] (Inv) {$A^{-1}$}; \draw [semithick,->] (Prevp) -- (Nextp) node [midway,above=0.0cm] {Projection}; \draw (Nextp)++(0.05,0.3cm) -- ++ (0.3cm,0.0cm) -- ++ (-0.05cm, -0.3cm); \node[inner sep =0mm, outer sep =0mm] (orthodot) at (-3.1cm+0.17cm,-2cm+0.15cm) {$\cdot$}; \end{tikzpicture}} \caption{\footnotesize The new estimate $X_{k+1}$ is obtained by projecting $X_{k}$ onto the affine space formed by intersecting $\{X \; : \; X= X^\top \}$ and $\left\{ X \; :\; S^\top AX= S^\top \right\}$.} \label{fig:proj} \end{figure} \subsection{Optimization viewpoint: constrain-and-approximate} The viewpoint~\eqref{eq:NFsym} also has an interesting dual viewpoint: \begin{equation} \boxed{X_{k+1} = \arg_{X} \min_{X \mathbb{R}^{n\times n}, Y\in \mathbb{R}^{n\times q}} \frac{1}{2}\norm{X -A^{-1}}_{F(B)}^2 \quad \mbox{subject to} \quad X = X_k + \frac{1}{2}(YS^\top AB^{-1} + B^{-1}A^\top SY^\top )} \label{eq:RFsym} \end{equation} The minimum is taken over matrices $X\in \mathbb{R}^{n\times n}$ and $Y\in \mathbb{R}^{n\times q}$. The next iterate $X_{k+1}$ is the best approximation to $A^{-1}$ restricted to a random affine space of symmetric matrices. Furthermore, \eqref{eq:RFsym} is a symmetric equivalent of~\eqref{eq:RF}; that is, the constraint in~\eqref{eq:RFsym} is the result of projecting the constraint in~\eqref{eq:RF} onto the space of symmetric matrices. When $A$ is symmetric positive definite and we choose $B =A$ in~\eqref{eq:RF} and~\eqref{eq:RFcols}, then \[\norm{X - A^{-1}}_{F(A)}^2 = \Tr{ (X-A^{-1})A(X-A^{-1})A} = \norm{XA - I}_{F}^2.\] The above is exactly the objective function used in most approximate inverse preconditioners (AIP)~\cite{Benzi1999,Gould1998,Kolotilina1993,Huckle2007}. \subsection{Equivalence} We now prove that the two viewpoints~\eqref{eq:NFsym} and~\eqref{eq:RFsym} are equivalent, and show their explicit solution. \begin{theorem} If $A$ and $X_k$ are symmetric, then the viewpoints~\eqref{eq:NFsym} and~\eqref{eq:RFsym} are equivalent. That is, they define the same $X_{k+1}$. Furthermore, if $S$ has full column rank, then the explicit solution to ~\eqref{eq:NFsym} and~\eqref{eq:RFsym} is \begin{align}&\boxed{X_{k+1} =X_{k}-(X_{k}AS-S)\Lambda S^\top A B^{-1} +B^{-1}AS\Lambda(S^\top AX_{k}-S^\top )\left(AS\Lambda S^\top AB^{-1}-I\right)} \label{eq:Xupdatesym} \end{align} where $\Lambda \eqdef (S^\top AB^{-1}A S)^{-1}$. \end{theorem} \begin{proof} We first prove the equivalence of~\eqref{eq:NFsym} and~\eqref{eq:RFsym} using Lagrangian duality. It suffices to prove the claim for $B=I$ as we did in the proof of Theorem~\ref{theo:NFRF}, since using the change of variables~\eqref{eq:varchangeW} applied to~\eqref{eq:NFsym} we have that~\eqref{eq:NFsym} is equivalent to \begin{equation} \min_{\hat{X} \in \mathbb{R}^{n\times n}} \frac{1}{2} \norm{\hat{X} - \hat{X}_{k}}_{F}^2 \quad \mbox{subject to } \quad \hat{S}^\top \hat{A}\hat{X} =\hat{S}^\top , \quad \hat{X}= \hat{X}^\top . \label{eq:NFsymbar} \end{equation} Since~\eqref{eq:NFsym} has a convex quadratic objective with linear constraints, strong duality holds. Thus we will derive a dual formulation for~\eqref{eq:NFsymbar} then use the change of coordinates~\eqref{eq:varchangeW} to recover the solution to~\eqref{eq:NFsym}. Let $\hat{Y} \in \mathbb{R}^{n\times q}$ and $W \in \mathbb{R}^{n\times n}$ and consider the Lagrangian of~\eqref{eq:NFsymbar} which is \begin{equation}\label{eq:lagsym} L(\hat{X},\hat{Y}, W)= \frac{1}{2}\norm{\hat{X}-\hat{X}_k}_F^2 - \dotprod{\hat{Y}^\top ,\hat{S}^\top \hat{A}(\hat{X}-\hat{A}^{-1})}_{F} - \dotprod{W,\hat{X}-\hat{X}^\top }_F. \end{equation} Differentiating in $\hat{X}$ and setting to zero gives \begin{equation}\label{eq:optsym} \hat{X} = \hat{X}_k+ \hat{A}^\top \hat{S}\hat{Y}^\top +W-W^\top . \end{equation} Applying the symmetry constraint $X=X^\top $ gives \[W-W^\top = \frac{1}{2}\left(\hat{Y}\hat{S}^\top \hat{A} - \hat{A}^\top \hat{S}\hat{Y}^\top \right).\] Substituting the above into~\eqref{eq:optsym} gives \begin{equation}\label{eq:optsym2} \hat{X} = \hat{X}_k +\frac{1}{2}\left(\hat{Y}\hat{S}^\top \hat{A} + \hat{A}^\top \hat{S}\hat{Y}^\top \right). \end{equation} Now let $\Theta = \frac{1}{2}(\hat{Y}\hat{S}^\top \hat{A} + \hat{A}^\top \hat{S}\hat{Y}^\top )$ and note that, since the matrix $\Theta+\hat{X}_k -\hat{A}^{-1}$ is symmetric, we get \begin{equation}\label{eq:symmob} \dotprod{\hat{A}^\top \hat{S}\hat{Y}^\top ,\Theta+\hat{X}_k -\hat{A}^{-1}}_F = \dotprod{\Theta ,\Theta+\hat{X}_k -\hat{A}^{-1}}_F. \end{equation} Substituting~\eqref{eq:optsym2} into~\eqref{eq:lagsym} gives \begin{align} L(\hat{X},\hat{Y},W) &= \frac{1}{2}\norm{\Theta}_F^2 - \dotprod{\hat{A}^\top \hat{S}\hat{Y}^\top ,\Theta+\hat{X}_k -\hat{A}^{-1}}_F \nonumber \overset{\eqref{eq:symmob}}{=} \frac{1}{2}\norm{\Theta}_F^2 - \dotprod{\Theta,\Theta+\hat{X}_k -\hat{A}^{-1}}_F \nonumber \\ & = -\frac{1}{2}\norm{\Theta}_F^2 -\dotprod{\Theta,\hat{X}_k -\hat{A}^{-1}}_F . \label{eq:dualsym2} \end{align} Adding $\pm\frac{1}{2}\norm{\hat{X}_k-\hat{A}^{-1}}_F^2$ to~\eqref{eq:dualsym2} gives \[ L(\hat{X},\hat{Y},W) = -\frac{1}{2}\norm{\Theta +\hat{X}_k -\hat{A}^{-1}}_F^2+\frac{1}{2}\norm{\hat{X}_k-\hat{A}^{-1}}_F^2. \] Finally, using~\eqref{eq:optsym2} and maximizing over $\hat{Y}$ then minimizing over $X$ gives the dual problem \[\min_{\hat{X},\hat{Y}} \frac{1}{2}\norm{\hat{X} -\hat{A}^{-1}}_F^2 \quad \mbox{subject to} \quad \hat{X} = \hat{X}_k + \frac{1}{2}(\hat{Y} \hat{S}^\top \hat{A} + \hat{A}^\top \hat{S}\hat{Y}^\top ). \] It now remains to change variables according to~\eqref{eq:varchangeW} and set $Y = B^{-1/2}\hat{Y}.$ It was recently shown in~\cite[Section~2]{Gower2014c} and~\cite[Section~4]{Hennig2015}\footnote{To re-interpret methods for solving linear systems through Bayesian inference, Hennig constructs estimates of the inverse system matrix using the sampled action of a matrix taken during a linear solve~\cite{Hennig2015}.} that~\eqref{eq:Xupdatesym} is the solution to~\eqref{eq:NFsym}. But for completion, we now give a new simple proof. From~\eqref{eq:optsym2} we see that the solution is solely determined by $\hat{Y}\hat{S}^\top \hat{A}$, and thus we focus on obtaining this matrix. To simplify notation, let $\Gamma =\hat{A}^\top \hat{S}$ and let $\hat{Z} =\Gamma(\Gamma^\top \Gamma)^{\dagger}\Gamma^\top .$ As $\hat{Z}$ is a projection matrix we have that $\hat{Z}^2 = \hat{Z}$ and $(I-\hat{Z})\hat{Z}=0,$ two properties we will use repeatedly. Using the sketch constraint in~\eqref{eq:NFsymbar} we have \begin{equation} \label{eq:Gammasketch} \Gamma^\top \hat{X} =\hat{S}^\top \hat{A}\hat{X}=\hat{S}^\top = \Gamma^\top \hat{A}^{-1},\end{equation} therefore left multiplying~\eqref{eq:optsym2} by $\Gamma^\top $ gives \begin{equation} \label{eq:sketchdualconst} \Gamma^\top \hat{X} = \Gamma^\top \hat{X}_k +\frac{1}{2}\Gamma^\top \left(\hat{Y}\Gamma^\top + \Gamma\hat{Y}^\top \right) \overset{\eqref{eq:Gammasketch}}{=} \Gamma^\top \hat{A}^{-1}. \end{equation} Let $R = \hat{A}^{-1}-\hat{X}_k$ which is a symmetric matrix. Rearranging~\eqref{eq:sketchdualconst} gives \[ (\Gamma^\top \Gamma)\hat{Y}^\top =\Gamma^\top \left( 2R -\hat{Y}\Gamma^\top \right). \] The least norm solution of the above in term of $\hat{Y}^\top $ is given by \[\hat{Y}^\top =(\Gamma^\top \Gamma)^{\dagger}\Gamma^\top \left( 2R -\hat{Y}\Gamma^\top \right).\] Left multiplying the above by $\Gamma$ gives \begin{equation}\label{eq:LDproj} \Gamma\hat{Y}^\top = \Gamma(\Gamma^T\Gamma)^{\dagger}\Gamma^T\left( 2R -\hat{Y}\Gamma^\top \right) =\hat{Z}\left( 2R -\hat{Y}\Gamma^\top \right). \end{equation} This shows that $\Gamma\hat{Y}^\top $ is equal to a projection matrix times an unknown matrix that is \begin{equation}\label{eq:PsiZ}\Gamma\hat{Y}^\top = \hat{Z}\Psi = \hat{Z}\Psi \hat{Z} +\hat{Z}\Psi (I-\hat{Z}), \end{equation} where $\Psi \in \mathbb{R}^{n \times n}$ is the unknown matrix. Note that in~\eqref{eq:PsiZ} we have decomposed the rows of $\hat{Z}\Psi$ into orthogonal components. Substituting~\eqref{eq:PsiZ} into~\eqref{eq:LDproj} gives \begin{equation} \label{eq:ZPsiZ} \hat{Z}\Psi \hat{Z} +\hat{Z}\Psi (I-\hat{Z}) =\hat{Z}\left( 2R -\hat{Z}\Psi^\top \hat{Z} \right),\end{equation} where we used that $\hat{Z}(I-\hat{Z})=0.$ Right multiplying~\eqref{eq:ZPsiZ} by $\hat{Z}$ and re-arranging gives \begin{equation}\label{eq:lefthatZ} \hat{Z}(\Psi +\Psi^\top )\hat{Z} = 2\hat{Z}R\hat{Z}. \end{equation} Right multiplying~\eqref{eq:ZPsiZ} by $I-\hat{Z}$ and re-arranging gives \begin{equation}\label{eq:lefthatIZ} \hat{Z}\Psi (I-\hat{Z}) =2\hat{Z}R(I-\hat{Z}). \end{equation} Finally, inserting~\eqref{eq:PsiZ} into~\eqref{eq:optsym2} gives \begin{eqnarray*} \hat{X} &= & \hat{X}_k +\frac{1}{2}\left(\hat{Z}(\Psi+\Psi^\top ) \hat{Z} +(I-\hat{Z})\Psi^\top \hat{Z} +\hat{Z}\Psi (I-\hat{Z}) \right)\\ &\overset{\eqref{eq:lefthatZ}+\eqref{eq:lefthatIZ}}{=}&\hat{X}_k + \hat{Z}R\hat{Z} + \hat{Z}R(I-\hat{Z})+(I-\hat{Z})R\hat{Z}\\ &=& \hat{X}_k+ R\hat{Z} +\hat{Z}R(I-\hat{Z})\\ &=& \hat{X}_k+ (\hat{S}-\hat{X}_k\hat{A}\hat{S})\Lambda \hat{S}^\top \hat{A} +\hat{A}\hat{S}\Lambda (\hat{S}^\top -\hat{S}^\top \hat{A}\hat{X}_k)(I-\hat{A}\hat{S}\Lambda\hat{S}^\top \hat{A}), \end{eqnarray*} where we used that $\Lambda =(\hat{S}^\top \hat{A}\hat{A}^\top \hat{S})^{\dagger} = (S^\top AB^{-1}A^\top S)^{-1}.$ It now remains to use the change of variables~\eqref{eq:varchangeW} to obtain~\eqref{eq:Xupdatesym}. \end{proof} \begin{algorithm}[!h] \begin{algorithmic}[1] \State \textbf{input:} symmetric invertible matrix $A \in \mathbb{R}^{n\times n}$ \State \textbf{parameters:} ${\cal D}$ = distribution over random matrices; symmetric positive definite $B\in \mathbb{R}^{n\times n}$ \State \textbf{initialize:} symmetric matrix $X_0\in \mathbb{R}^{n\times n}$ \For {$k = 0, 1, 2, \dots$} \State Sample an independent copy $S\sim {\cal D}$ \State Compute $\Lambda = S(S^\top AB^{-1}A S)^{-1}S^\top $ \State Compute $\Theta = \Lambda A B^{-1}$ \State Compute $M_k = X_k A - I$ \State $X_{k+1} =X_{k}- M_k \Theta - (M_k \Theta)^\top + \Theta^\top (AX_{k}A-A) \Theta$ \Comment This is equivalent to \eqref{eq:NFsym} \& \eqref{eq:RFsym} \EndFor \State \textbf{output:} last iterate $X_k$ \end{algorithmic} \caption{Stochastic Iterative Matrix Inversion (SIMI) -- symmetric variant} \label{alg:sym} \end{algorithm} \section{Convergence} \label{sec:conv} We now analyze the convergence of the \emph{error}, $X_{k}-A^{-1}$, for iterates of Algorithms~\ref{alg:asym-row}, \ref{alg:asym-col} and~\ref{alg:sym}. For the sake of economy of space, we only analyze Algorithms~\ref{alg:asym-row} and \ref{alg:sym}. Convergence of Algorithm~\ref{alg:asym-col} follows from convergence of Algorithm~\ref{alg:asym-row} by observing Remark~\ref{rem:alg2conv}. The first analysis we present in Section~\ref{sec:normEconv} is concerned with the convergence of $\norm{\E{X_{k}-A^{-1}}}^2,$ that is, the {\em norm of the expected error}. We then analyze the convergence of $\E{\norm{X_{k}-A^{-1}}}^2,$ the {\em expected norm of the error}. The latter is a stronger type of convergence, as explained in Lemma~\ref{lem:convrandvar}. The convergence of Algorithms~\ref{alg:asym-row} and \ref{alg:sym} can be entirely characterized by studying the following random matrix \begin{equation}\label{eq:Z} Z \eqdef A^\top S(S^\top AB^{-1}A^\top S)^{-1}S^\top A. \end{equation} With this definition, the update step of Algorithm~\ref{alg:asym-row} can be re-written as a simple fixed point formula \begin{align} X_{k+1} -A^{-1}&= \left(I-B^{-1}Z\right)(X_{k}-A^{-1}). \label{eq:XZupdate} \end{align} We can also simplify the iterates of Algorithm~\ref{alg:sym} to \begin{align} X_{k+1}-A^{-1} &= \left(I-B^{-1}Z\right) (X_{k}-A^{-1})\left(I -ZB^{-1} \right). \label{eq:XZupdatesym} \end{align} Much like our convergence proofs in Section~\ref{C2sec:convergence}, the only stochastic component in our methods is contained in the matrix $Z$, and thus the convergence of the iterates will depend on the properties of $Z$ and its expected value $\E{Z}.$ In particular, recall from Lemma~\ref{ch:one:lem:Z} that $B^{-1}ZB^{-1}$ is an orthogonal projection. \subsection{Norm of the expected error} \label{sec:normEconv} We start by proving that the norm of the expected error of the iterates of Algorithm~\ref{alg:asym-row} and Algorithm~\ref{alg:sym} converges to zero. The following theorem is remarkable in that we do not need to make any assumptions on the distribution $S$, except that $S$ has full column rank. Rather, the theorem pinpoints that convergence depends solely on the spectrum of $I-B^{1/2}\E{Z}B^{1/2}.$ \begin{theorem} \label{theo:normEconv} Let $S$ be a random matrix which has full column rank with probability~$1$ (so that $Z$ is well defined). Then the iterates $X_{k+1}$ of Algorithm~\ref{alg:asym-row} satisfy \begin{equation} \label{eq:XXXX} \E{ X_{k+1} -A^{-1} } =(I-B^{-1}\E{Z}) \E{X_{k} - A^{-1}}. \end{equation} Let $X_0 \in \mathbb{R}^{n\times n}$. If $X_k$ is calculated in either one of these two ways \begin{enumerate} \item Applying $k$ iterations of Algorithm~\ref{alg:asym-row}, \item Applying $k$ iterations of Algorithm~\ref{alg:sym} (assuming $A$ and $X_0$ are symmetric), \end{enumerate} then $X_k$ converges to the inverse exponentially fast, according to \begin{equation}\label{eq:normexpconv} \norm{\E{X_{k}-A^{-1}}}_{B}^* \leq \rho^k \norm{X_{0}-A^{-1}}_{B}^*, \end{equation} where \begin{equation} \label{eq:rhoequiv}\rho \eqdef 1-\lambda_{\min}(B^{-1/2}\E{Z}B^{-1/2}).\end{equation} Moreover, we have the following lower and upper bounds on the convergence rate: \begin{equation}\label{eq:rholower} 0 \leq 1-\frac{\E{q}}{n} \leq \rho \leq 1. \end{equation} \end{theorem} \begin{proof} Let \begin{equation}\label{eq:oihsoi8dhyY8J} R_k \eqdef B^{1/2}R_k B^{1/2}\quad and \quad \hat{Z} \eqdef B^{-1/2}ZB^{-1/2},\end{equation} for all $k$. Thus $\hat{Z}$ is a projection matrix (see Lemma~\ref{ch:one:lem:Z}) and $\norm{R_k}_2 = \norm{X_{k}-A^{-1}}_B.$ Left and right multiplying~\eqref{eq:XZupdate} by $B^{1/2}$ gives \begin{equation}\label{eq:barRknext} R_{k+1} = (I -\hat{Z})R_k. \end{equation} Taking expectation with respect to $S$ in~\eqref{eq:barRknext} gives \begin{equation}\label{eq:EXinXk} \E{R_{k+1} \;| \; R_k} = (I - \E{\hat{Z}}) R_k. \end{equation} Taking full expectation in~\eqref{eq:barRknext} and using the tower rule gives \begin{eqnarray} \E{ R _{k+1}} &=& \E{\E{ R _{k+1} \;|\; R_{k}}} \nonumber \\ &\overset{\eqref{eq:EXinXk}}{=}& \E{(I -\E{\hat{Z}})R_{k}} \nonumber\\ &=& (I -\E{\hat{Z}})\E{R_{k}}. \label{eq:Efullasym} \end{eqnarray} Applying the norm in~\eqref{eq:Efullasym} gives \begin{align} \label{eq:normErecur} \norm{\E{ R_{k+1}}}_{2} &\leq \norm{I -\E{\hat{Z}}}_2 \norm{\E{ R_{k}}}_{2}. \end{align} Furthermore \begin{align} \norm{I -\E{\hat{Z}}}_2 &= \lambda_{\max}\left(I -\E{\hat{Z}}\right) \nonumber \\ &=1-\lambda_{\min}(\E{\hat{Z}}) \overset{\eqref{eq:rhoequiv}}{=} \rho, \label{eq:rhonorm} \end{align} where we used to symmetry of $(I-\E{\hat{Z}})$ when passing from the operator norm to the spectral radius. Note that the symmetry of $\E{\hat{Z}}$ derives from the symmetry of $\hat{Z}$. It now remains to unroll the recurrence in~\eqref{eq:normErecur} to get~\eqref{eq:normexpconv}. Now we analyze the iterates of Algorithm~\ref{alg:sym}. Left and right multiplying~\eqref{eq:XZupdatesym} by $B^{1/2}$ we have \begin{equation}\label{eq:barRevol} R_{k+1}= P(R_k)\eqdef \left(I-\hat{Z}\right) R_k\left(I -\hat{Z} \right). \end{equation} Defining $\bar{P}: R \mapsto \E{P(R) \, | \, R_k}$, taking expectation in \eqref{eq:barRevol} conditioned on $R_{k}$, gives \[ \E{R_{k+1}\; | \; R_k} = \bar{P}(R_{k}).\] As $\bar{P}$ is a linear operator, taking expectation again yields \begin{equation} \label{eq:barPZERk} \E{R_{k+1}} = \E{\bar{P}(R_{k} )} = \bar{P}(\E{R_{k} }). \end{equation} Let $||| \bar{P}|||_2 \eqdef \max_{\norm{R}_2=1} \norm{\bar{P}(R)}_2$ be the operator induced norm. Applying norm in~\eqref{eq:barPZERk} gives \begin{eqnarray} \norm{\E{X_{k+1}-A^{-1}}}_{B}^* &=& \norm{\E{R_{k+1}}}_2 \\ &\leq & ||| \bar{P}|||_2 \norm{\E{R_{k}}}_2 \nonumber\\ & = & ||| \bar{P}|||_2 \norm{\E{X_{k}-A^{-1}}}_{B}^*.\label{eq:normsplitbarP} \end{eqnarray} Clearly, $P$ is a \emph{positive linear map}, that is, it is linear and maps positive semi-definite matrices to positive semi-definite matrices. Thus, by Jensen's inequality, the map $\bar{P}$ is also a positive linear map. As every positive linear map attains its norm at the identity matrix (see Corollary 2.3.8 in~\cite{bhatia07}), we have that \begin{eqnarray*} ||| \bar{P}|||_2 &= & \norm{\bar{P}(I)}_2 \\ &\overset{\eqref{eq:barRevol}}{=} &\norm{\E{ \left(I-\hat{Z}\right) I\left(I -\hat{Z} \right)}}_2\\ &\overset{(\text{Lemma}~\ref{ch:one:lem:Z})}{=}&\norm{\E{ I-\hat{Z}}}_2 \overset{\eqref{eq:rhonorm}}{=} \rho. \end{eqnarray*} Inserting the above equivalence in~\eqref{eq:normsplitbarP}, unrolling the recurrence and using the substitution \eqref{eq:oihsoi8dhyY8J} gives~\eqref{eq:normexpconv}. Finally~\eqref{eq:rholower} follows immediately from Lemma~\ref{lem:rho1} as $A$ is invertible and $S$ has full column rank. \end{proof} If $\rho =1$, this theorem does not guarantee convergence. But when $\E{Z}$ is positive definite, as it will transpire in all practical variants of our method, some of which we describe in Section~\ref{sec:discretemethods}, the rate $\rho$ will be strictly less than one, and the norm of the expected error will converge to zero. \subsection{Expectation of the norm of the error} Now we consider the convergence of the expected norm of the error. \begin{theorem} \label{theo:Enormconv} Let $S$ be a random matrix that has full column rank with probability~$1$ and such that $\E{Z}$ is positive definite, where $Z$ is defined in~\eqref{eq:Z}. Let $X_0 \in \mathbb{R}^{n\times n}$. If $X_k$ is calculated in either one of these two ways \begin{enumerate} \item Applying $k$ iterations of Algorithm~\ref{alg:asym-row}, \item Applying $k$ iterations of Algorithm~\ref{alg:sym} (assuming both $A$ and $X_0$ are symmetric matrices), \end{enumerate} then $X_k$ converges to the inverse according to \begin{equation} \label{eq:Enormconv} \E{\norm{X_{k} -A^{-1} }_{F(B)}^2} \leq \rho^k \norm{X_{0} - A^{-1}}_{F(B)}^2. \end{equation} \end{theorem} \begin{proof} First consider Algorithm~\ref{alg:asym-row}, where $X_{k+1}$ is calculated by iteratively applying~\eqref{eq:XZupdate}. Using again the substitution~\eqref{eq:oihsoi8dhyY8J}, then from~\eqref{eq:XZupdate} we have \begin{equation}\label{eq:js9hf7HyT} R_{k+1} = \left(I-\hat{Z}\right) R_k. \end{equation} From this we obtain \begin{eqnarray} \norm{ R_{k+1}}_{F}^2 & \overset{\eqref{eq:js9hf7HyT}}{=} & \norm{\left(I-\hat{Z}\right) R_{k}}_{F}^2 \nonumber\\ & = &\Tr{\left(I-\hat{Z}\right)\left(I-\hat{Z}\right) R_k R_{k}^\top } \nonumber\\ &\overset{(\text{Lemma}~\ref{ch:one:lem:Z})}{=}& \Tr{\left(I-\hat{Z}\right) R_k R_{k}^\top } \label{eq:asymmapplylem}\\ &= & \norm{ R_k}_{F}^2 - \Tr{\hat{Z} R_{k} R_k^\top }. \nonumber \end{eqnarray} Taking expectations, conditioned on $ R_k$, we get \[\E{\norm{ R_{k+1}}_{F}^2 \, | \, R_k} =\norm{ R_k}_{F}^2 - \Tr{\E{\hat{Z}} R_{k} R_k^\top }.\] Using that $\Tr{\E{\hat{Z}} R_{k} R_k^\top } \geq \lambda_{\min}\left(\E{\hat{Z}}\right)\Tr{ R_k R_{k}^\top } $, which relies on the symmetry of $\E{\hat{Z}},$ we have that \begin{align*}\E{\norm{ R_{k+1}}_{F}^2 \,| \, R_k} &\leq \left(1-\lambda_{\min}\left(\E{\hat{Z}}\right) \right) \norm{ R_k}_{F}^2 =\rho\cdot \norm{ R_k}_{F}^2. \end{align*} In order to arrive at \eqref{eq:Enormconv}, it now remains to take full expectation, unroll the recurrence and use the substitution \eqref{eq:oihsoi8dhyY8J} together with $\norm{R_k}_{F}^2 \overset{\eqref{eq:98y988ff}}{=} \norm{X_k-A^{-1}}_{F(B)}^2.$ Now we assume that $A$ and $X_0$ are symmetric and $\{X_k\}$ are the iterates computed by Algorithm~\ref{alg:sym}. Left and right multiplying~\eqref{eq:XZupdatesym} by $B^{1/2}$ we have \begin{equation}\label{eq:barRevol2} R_{k+1}= \left(I-\hat{Z}\right) R_k\left(I -\hat{Z} \right). \end{equation} Taking norm we have \begin{eqnarray} \norm{ R_{k+1}}_{F}^2 & \overset{(\text{Lemma}~\ref{ch:one:lem:Z})}{=} & \Tr{ R_k\left(I -\hat{Z} \right) R_k\left(I -\hat{Z} \right)} \nonumber \\ &=& \Tr{ R_k R_k\left(I -\hat{Z} \right)} -\Tr{ R_k \hat{Z} R_k\left(I -\hat{Z} \right)} \nonumber\\ & \leq& \Tr{ R_k R_k\left(I -\hat{Z}\right) }, \label{eq:symEnorm1} \end{eqnarray} where in the last inequality we used that $I -\hat{Z}$ is an orthogonal projection and thus it is symmetric positive semi-definite, whence \[\Tr{ R_k \hat{Z} R_k\left(I -\hat{Z} \right)} = \Tr{\hat{Z}^{1/2} R_k\left(I -\hat{Z} \right)R_k \hat{Z}^{1/2}} \geq 0. \] The remainder of the proof follows similar steps as those we used in the first part of the proof from~\eqref{eq:asymmapplylem} onwards. \end{proof} Theorem~\ref{theo:Enormconv} establishes that for all three methods, the expected norm of the error converges exponentially fast to zero. Moreover, the convergence rate $\rho$ is the same that appeared in Theorem~\ref{theo:normEconv}, where we established the convergence of the norm of the expected error. Using Lemma~\ref{lem:itercomplex}, both of the convergence results in Theorems~\ref{theo:normEconv} and~\ref{theo:Enormconv} can be recast as iteration complexity bounds. For instance, for a given $0<\epsilon <1$, Theorem~\ref{theo:normEconv} combined with Lemma~\ref{lem:itercomplex} (with $\alpha_k =(\norm{\E{X_k-A^{-1}}}_{B}^*)^2$) gives \begin{equation} \label{eq:itercomplex}k \geq \left(\frac{1}{2}\right)\frac{1}{1-\rho} \log\left(\frac{1}{\epsilon}\right) \quad \Rightarrow \quad (\norm{\E{X_k-A^{-1}}}_{B}^*)^2 \leq \epsilon (\norm{X_0-A^{-1}}_{B}^*)^2. \end{equation} On the other hand, Theorem~\ref{theo:Enormconv} combined with Lemma~\ref{lem:itercomplex} (with $\alpha_k =\mathbf{E}[\norm{X_k-A^{-1}}_{F(B)}^2]$) gives \begin{equation} \label{eq:itercomplex2}k \geq \frac{1}{1-\rho} \log\left(\frac{1}{\epsilon}\right) \quad \Rightarrow \quad \E{\norm{X_k-A^{-1}}_{F(B)}^2} \leq \epsilon \norm{X_0-A^{-1}}_{F(B)}^2. \end{equation} To push the expected norm of the error below the $\epsilon$ tolerance~\eqref{eq:itercomplex2}, we require double the amount of iterates, as compared with bringing the norm of expected error below the same tolerance~\eqref{eq:itercomplex}. This is because in Theorem~\ref{theo:Enormconv} we determined that $\rho$ is the rate at which the expectation of the \emph{squared} norm error converges, while in Theorem~\ref{theo:normEconv} we determined that $\rho$ is the rate at which the norm, without the square, of the expected error converges. Though it takes double the number of iterations to decrease the expectation of the norm error, as proven in Lemma~\ref{lem:convrandvar}, the former is a stronger form of convergence. Thus, Theorem~\ref{theo:normEconv} does not give a stronger result than Theorem~\ref{theo:Enormconv}, but rather, these theorems give qualitatively different results and ultimately enrich our understanding of the iterative process. \section{Discrete Random Matrices} \label{sec:discrete} We now consider the case of a discrete random matrix $S$. We show that when $S$ is a \emph{complete discrete sampling}, then $\E{Z}$ is positive definite, and thus from Theorems~\ref{theo:normEconv} and~\ref{theo:Enormconv} together with Remark~\ref{rem:alg2conv}, Algorithms~\ref{alg:asym-row}, \ref{alg:asym-col} and~\ref{alg:sym} converge. \begin{definition}[Complete Discrete Sampling]\label{def:complete} The random matrix $S$ has a finite discrete distribution with $r$ outcomes. In particular, $S= S_i \in \mathbb{R}^{n \times q_i}$ with probability $p_i>0$ for $i=1,\ldots, r$, where $S_i$ is of full column rank. We say that $S$ is a complete discrete sampling when $\mathbf{S} \eqdef [S_1, \ldots, S_r] \in \mathbb{R}^{m\times \sum_{i=1}^r q_i}$ has full row rank. \end{definition} Since we consider $A$ to be invertible in this chapter, the above definition of complete discrete sampling is in synchrony with the definition presented in Chapter~\ref{ch:linear_systems}. As an example of a complete discrete sampling, let $S =e_i$ (the $i$th unit coordinate vector in $\mathbb{R}^n$) with probability $p_i =1/n$, for $i =1,\ldots, n.$ Then $\mathbf{S}$, as defined in Definition~\ref{def:complete}, is equal to the identity matrix: $\mathbf{S} =I$. Consequently, $S$ is a complete discrete sampling. In fact, from any basis of $\mathbb{R}^n$ we could construct a complete discrete sampling in an analogous way. Next we establish that when $S$ is discrete random matrix, that $S$ having a complete discrete distribution is a necessary and sufficient condition for $\E{Z}$ to be positive definite. \begin{proposition}\label{prop:Ediscrete} Let $S$ be a discrete random matrix with $r$ outcomes $S_r$ all of which have full column rank. The matrix $\E{Z}$ is positive definite if and only if $S$ is a complete discrete sampling. Furthermore \begin{equation}\E{Z} = A^\top \mathbf{S} D^2 \mathbf{S}^\top A \label{eq:EZdiscrete}, \end{equation} where \begin{equation} \label{eq:D} D~\eqdef~\mbox{Diag}\left( \sqrt{p_1}(S_1^\top AB^{-1}A^\top S_1)^{-1/2}, \ldots, \sqrt{p_r}(S_r^\top AB^{-1}A^\top S_r)^{-1/2}\right).\end{equation} \end{proposition} \begin{proof} The equation~\eqref{eq:EZdiscrete} was established in Proposition~\ref{pro:Ediscrete}. Since we assume that $S$ has full column rank with probability $1$, the matrix $D$ is well defined and nonsingular. Given that $\E{Z}$ is positive semi-definite, we need only show that $\Null{\E{Z}}$ contains only the zero vector if and only if $S$ is a complete discrete sampling. Let $v \in \Null{\E{Z}}$ and $v \neq 0,$ thus \[0=v^\top A^\top \mathbf{S} D^2 \mathbf{S}^\top Av = \norm{D\mathbf{S}^\top Av}_2^2, \] which shows that $\mathbf{S}^\top Av =0$ and thus $v \in \Null{\mathbf{S}^\top A}.$ As $A$ is nonsingular, it follows that $v=0$ if and only if $\mathbf{S}^\top $ has full column rank. \end{proof} With a closed form expression for $\E{Z}$ we can optimize $\rho$ over the possible distributions of $S$ to yield a better convergence rate. \subsection{Optimizing an Upper Bound on the Convergence Rate} \label{sec:discreteopt} So far we have proven two different types of convergence for Algorithms~\ref{alg:asym-row}, \ref{alg:asym-col} and \ref{alg:sym} in Theorems~\ref{theo:normEconv} and~\ref{theo:Enormconv}. Furthermore, both forms of convergence depend on the same convergence rate $\rho$ for which we have a closed form expression~\eqref{eq:rhoequiv}. The availability of a closed form expression for the convergence rate opens up the possibility of designing particular distributions for $S$ optimizing the rate. In Section~\ref{ch:two:sec:optprob} we showed that for a complete discrete sampling, computing the optimal probability distribution, assuming that the matrices $\{S_i\}_{i=1}^r$ are fixed, leads to a semi-definite program (SDP). Here we propose a more practical alternative: to optimize the following upper bound on the convergence rate: \[ \rho = 1- \lambda_{\min}(B^{-1/2}\E{Z} B^{-1/2}) \leq 1 - \frac{1}{\Tr{B^{1/2} (\E{Z})^{-1} B^{1/2}}} \eqdef \gamma.\] To emphasize the dependence of $\gamma$ and $Z$ on the probability distribution $p = (p_1,\ldots, p_r)\in \mathbb{R}^r$, let us denote \begin{equation}\label{eq:98h9s8h8s}\gamma(p) \eqdef 1 - \frac{1}{\Tr{B^{1/2} (\E{Z_p})^{-1} B^{1/2}}}, \end{equation} where we have added a subscript to $Z$ to indicate that it is a function of $p$. We now minimize $\gamma(p)$ over the probability simplex: \[\Delta_r \eqdef \left\{p = (p_1,\dots,p_r) \in \mathbb{R}^r \;:\; \sum_{i=1}^r p_i =1, \; p\geq 0\right\}.\] \begin{theorem} Let $S$ be a complete discrete sampling and let $\overline{S}_i\in \mathbb{R}^{n\times q_i}$, for $i=1,2,\dots,r$, be such that $\mathbf{S}^{-T} = [\overline{S}_1, \ldots, \overline{S}_r]$. Then \begin{equation}\label{eq:discoptrate} \min_{p\in \Delta_r}\gamma(p) \quad=\quad 1 -\frac{1}{\left(\sum_{i=1}^r \norm{B^{-1/2}A^\top S_i \overline{S}^\top _i A^{-T} B^{1/2}}_F\right)^2} . \end{equation} \end{theorem} \begin{proof} In view of \eqref{eq:98h9s8h8s}, minimizing $\gamma$ in $p$ is equivalent to minimizing $\Tr{B^{1/2} (\E{Z_p})^{-1} B^{1/2}}$ in $p$. Further, we have \begin{eqnarray} \Tr{B^{1/2} (\E{Z_p})^{-1}B^{1/2}} &\overset{\eqref{eq:EZdiscrete}}{=} & \Tr{B^{1/2} ( A^\top \mathbf{S} D^2 \mathbf{S}^\top A)^{-1}B^{1/2}}\\ &=& \Tr{B^{1/2} A^{-1} \mathbf{S}^{-T} D^{-2} \mathbf{S}^{-1} A^{-T} B^{1/2}} \nonumber \\ &\overset{\eqref{eq:D}}{=}&\sum_{i=1}^r \frac{1}{p_i}\Tr{B^{1/2} A^{-1} \overline{S}_i (S_i^\top AB^{-1}A^\top S_i) \overline{S}^\top _i A^{-T} B^{1/2}} \nonumber\\ &=&\sum_{i=1}^r \frac{1}{p_i}\norm{B^{-1/2}A^{-1} \overline{S}_i S_i^\top A B^{1/2}}_F^2. \label{eq:pifracsum} \end{eqnarray} Applying Lemma~\ref{lem:fracsum} in the Appendix, the optimal probabilities are given by \begin{equation} \label{eq:discoptprob} p_i = \frac{\norm{B^{-1/2}A^{-1}\overline{S}_i S_i^\top A B^{1/2}}_F}{\sum_{j=1}^r \norm{B^{-1/2}A^{-1}\overline{S}_j S_j^\top A B^{1/2}}_F}, \quad i=1,2,\dots,r \end{equation} Plugging this into~\eqref{eq:pifracsum} gives the result~\eqref{eq:discoptrate}. \end{proof} Observe that in general, the optimal probabilities~\eqref{eq:discoptprob} cannot be calculated, since the formula involves the inverse of $A$, which is not known. However, if $A$ is symmetric positive definite, we can choose $B^{-1}=A^2$, which eliminates this issue. If $A$ is not symmetric positive definite, or if we do not wish to choose $B^{-1}=A^2$, we can approach the formula \eqref{eq:discoptprob} as a recipe for a heuristic choice of the probabilities: we can use the iterates $\{X_{k}\}$ as a proxy for $A^{-1}$. With this setup, the resulting method is not guaranteed to converge by the theory developed in this thesis. However, in practice one would expect it to work well. We have not done extensive experiments to test this, and leave this to future research. To illustrate, let us consider a concrete simple example. Choose $B =I$ and $S_i =e_i$ (the unit coordinate vector in $\mathbb{R}^n$). We have $\mathbf{S} = [e_1,\dots,e_n] = I$, whence $\overline{S}_i =e_i$ for $i=1,\dots,r$. Plugging into \eqref{eq:discoptprob}, we obtain \[p_i = \frac{\norm{X_k e_i e_i^\top A}_F}{\sum_{j=1}^r \norm{X_k e_j e_j ^\top A}_F} = \frac{\norm{X_k e_i}_2\norm {e_i^\top A}_2}{\sum_{j=1}^r \norm{X_k e_j}_2\norm{ e_j ^\top A}_2}.\] \subsection{Adaptive Samplings} In Theorem~\ref{theo:convsingleS} we determined a discrete probability distribution $\mathbf{P}(S= S_i) =p_i$ that yields a convergence rate $\rho$ that is easy to interpret. We will now make use of this convenient probability distribution and pose the question: Having decided on the probabilities $p_1,\dots,p_r$, how should we choose the matrices $S_1,\ldots, S_r$ if we want $\rho $ to be as small as possible? To answer this question, first we observe that the convergence rate in Theorem~\ref{theo:convsingleS} is proportional to the scaled condition number defined by \begin{align}\label{eq:kappalower} \kappa_{2,F}(B^{-1/2}A^\top \mathbf{S}) &\eqdef \norm{(B^{-1/2}A^\top \mathbf{S})^{-1}}_2 \norm{B^{-1/2}A^\top \mathbf{S}}_F = \sqrt{ \frac{\Tr{\mathbf{S}^\top AB^{-1}A^\top \mathbf{S}}}{\lambda_{\min}\left(\mathbf{S}^\top AB^{-1}A^\top \mathbf{S} \right)} } \geq \sqrt{n}. \end{align} That is, if we select the probabilities \begin{equation} \label{eq:convprob} p_i = \left.\norm{B^{-1/2}A^\top S_i}_F^2\right/\norm{B^{-1/2}A^\top \mathbf{S}}_F^2, \end{equation} then Theorem~\ref{theo:convsingleS} combined with~\eqref{eq:kappalower} gives that the resulting convergence rate is \begin{equation} \rho = 1- \frac{\Tr{\mathbf{S}^\top AB^{-1}A^\top \mathbf{S}}}{\lambda_{\min}\left(\mathbf{S}^\top AB^{-1}A^\top \mathbf{S} \right)}= 1- \frac{1}{\kappa_{2,F}^2(B^{-1/2}A^\top \mathbf{S})}. \label{eq:rhoconv} \end{equation} Furthermore, following from Remark~\ref{rem:alg2conv}, we can determine a convergence rate for Algorithm~\ref{alg:asym-col} based on~\eqref{eq:rhoconv}. That is, by merely transposing each occurrence of $A$ we have that, by selecting $S_i$ with probability \begin{equation} \label{eq:convprob2} p_i = \left.\norm{B^{-1/2}A S_i}_F^2\right/\norm{B^{-1/2}A\mathbf{S}}_F^2, \end{equation} then Algorithm~\ref{alg:asym-col} converges at the rate \begin{equation}\label{eq:rhoconv2} \rho_2 = 1- \frac{1}{ \kappa_{2,F}^2(B^{-1/2}A\mathbf{S})}. \end{equation} Since these rates improve as the condition number $\kappa^2_{2,F}(B^{-1/2}A^\top \mathbf{S})$ (or $\kappa_{2,F}^2(B^{-1/2}A\mathbf{S})$ for Algorithm~\ref{alg:asym-col}) decreases, we should aim for matrices $S_1,\ldots, S_r$ that minimize the condition number. For instance, the lower bound in~\eqref{eq:kappalower} is reached for $\mathbf{S} = (B^{-1/2}A^\top )^{-1} = A^{-T}B^{1/2}$. While we do not know $A^{-1}$, we can use our best current approximation of it, $X_k$, in its place. This leads to a method which {\em adapts} the probability distribution governing $S$ throughout the iterative process. This observation inspires a very efficient modification of Algorithm~\ref{alg:sym}, which we call AdaRBFGS (Adaptive Randomized BFGS), and describe in Section~\ref{sec:AdaRBFGS}. Notice that, luckily and surprisingly, our twin goals of computing the inverse and optimizing the convergence rate via the above adaptive trick are compatible. Indeed, we wish to find $A^{-1}$, whose knowledge gives us the optimal rate. This should be contrasted with the SDP approach mentioned earlier in Section~\ref{ch:two:sec:optprob}: i) the SDP could potentially be harder than the inversion problem, and ii) having found the optimal probabilities $\{p_i\}$, we are still not guaranteed the optimal rate. Indeed, optimality is relative to the choice of the matrices $S_1,\dots,S_r$, which can be suboptimal. \begin{remark}[Adaptive sampling] The convergence rate~\eqref{eq:rhoconv2} suggests how one can select a sampling distribution for $S$ that would result in faster practical convergence. We now detail several practical choices for $B$ and indicate how to sample $S$. These suggestions require that the distribution of $S$ depends on the iterate $X_k$, and thus no longer fit into our framework. Nonetheless, we collect these suggestions here in the hope that others will wish to extend these ideas further, and as a demonstration of the utility of developing convergence rates. \begin{enumerate} \item If $B =I$, then Algorithm~\ref{alg:asym-row} converges at the rate $\rho = 1- 1/\kappa_{2,F}^2(A^\top \mathbf{S})$, and hence $S$ should be chosen so that $\mathbf{S}$ is a preconditioner of $A^\top $. For example $\mathbf{S}=X_{k}^\top ,$ that is, $S$ should be a sampling of the rows of $X_{k}$. \item If $B =I$, then Algorithm~\ref{alg:asym-col} converges at the rate $\rho = 1- 1/\kappa_{2,F}^2(A\mathbf{S})$, and hence $S$ should be chosen so that $\mathbf{S}$ is a preconditioner of $A$. For example $\mathbf{S}=X_{k}$; that is, $S$ should be a sampling of the columns of $X_{k}$. \item If $A $ is symmetric positive definite, we can choose $B=A$, in which case Algorithm~\ref{alg:sym} converges at the rate $\rho =1- 1/\kappa_{2,F}^2(A^{1/2}\mathbf{S}).$ This rate suggests that $S$ should be chosen so that $\mathbf{S}$ is an approximation of $A^{-1/2}.$ In Section~\ref{sec:AdaRBFGS} we develop this idea further, and design the AdaRBFGS algorithm. \item If $B=A^\top A$, then Algorithm~\ref{alg:asym-row} can be efficiently implemented with $S=AV$, where $V$ is a complete discrete sampling. Furthermore $\rho = 1- 1/\kappa_{2,F}^2(A\mathbf{V}),$ where $\mathbf{V} \eqdef [V_1,\ldots, V_r]$. This rate suggests that $V$ should be chosen so that $\mathbf{V}$ is a preconditioner of $A$. For example $\mathbf{V}=X_{k}$; that is, $V$ should be a sampling of the rows of $X_{k}$. \item If $B=AA^\top $, then Algorithm~\ref{alg:asym-col} can be efficiently implemented with $S=A^\top V$, where $V$ is a complete discrete sampling. From~\eqref{eq:rhoconv2}, the convergence rate of the resulting method is given by $1- 1/ \kappa_{2,F}^2(A^\top \mathbf{V}).$ This rate suggests that $V$ should be chosen so that $\mathbf{V}$ is a preconditioner of $A^\top $. For example, $\mathbf{V}=X_{k}^\top $; that is, $V$ should be a sampling of the columns of $X_{k}$. \item If $A$ is symmetric positive definite, we can choose $B = A^{-2}$, in which case Algorithm~\ref{alg:sym} can be efficiently implemented with $S =AV.$ Furthermore $\rho = 1- 1/\kappa_{2,F}^2(A\mathbf{V}).$ This rate suggests that $V$ should be chosen so that $\mathbf{V}$ is a preconditioner of $A$. For example $\mathbf{V}=X_{k},$ that is, $V$ should be a sampling of the rows or the columns of $X_{k}$. \end{enumerate} \end{remark} \section{Randomized Quasi-Newton Updates} \label{sec:discretemethods} Algorithms~\ref{alg:asym-row}, \ref{alg:asym-col} and \ref{alg:sym} are in fact families of algorithms indexed by the two parameters: i) positive definite matrix $B$ and ii) distribution $\cal D$ (from which we pick random matrices $S$). This allows us to design a myriad of specific methods by varying these parameters. Here we highlight some of these possibilities, focusing on complete discrete distributions for $S$ so that convergence of the iterates is guaranteed through Theorems~\ref{theo:normEconv} and~\ref{theo:Enormconv}. We also compute the convergence rate $\rho$ for these special methods for the convenient probability distribution given by~\eqref{eq:convprob} and~\eqref{eq:convprob2} (Theorem~\ref{theo:convsingleS}) so that the convergence rates~\eqref{eq:rhoconv} and~\eqref{eq:rhoconv2} depend on a scaled condition number which is easy to interpret. We will also make some connections to existing quasi-Newton and Approximate Inverse Preconditioning methods. Table~\ref{tbl:QN} provides a guide through this section. \begin{table}[!h] \begin{center} { \footnotesize \begin{tabular}{|c|c|c|c|c|c|} \hline $A$ & $B^{-1}$ & $S$ & Inverse Equation & Randomized Update & Section\\ \hline any & any & invertible & any & One Step & \ref{subsec:1step}\\ any & $ I$ & $e_i$ & $AX = I$ & Simultaneous Kaczmarz (SK) & \ref{subsec:09u09u09Kacz}\\ any & $I$ & vector & $XA = I$ & Bad Broyden (BB) & \ref{subsec:09u09u09}\\ sym. & $I$ & vector & $AX = I, X=X^\top $ & Powell-Symmetric-Broyden (PSB) & \ref{subsec:PSB} \\ any & $ I$ & vector & $XA^{-1} = I$ & Good Broyden (GB) & \ref{subsec:09u09u09good}\\ sym. & $A^{-1}-X_k$ & vector & $AX =I$ or $XA=I$ & Symmetric Rank 1 (SR1) & \ref{subsec:SR1}\\ s.p.d. & $A$ & vector & $XA^{-1} =I , X=X^\top $ & Davidon-Fletcher-Powell (DFP) & \ref{sec:RDFP} \\ s.p.d. & $A^{-1}$ & vector & $AX = I, X=X^\top $ & Broyden-Fletcher-Goldfarb-Shanno (BFGS) & \ref{sec:RBFGS} \\ any & $(A^\top A)^{-1}$ & vector & $AX=I$ & Column & \ref{sec:RCM}\\ \hline \end{tabular} } \end{center} \caption{Specific randomized updates for inverting matrices discussed in this section, obtained as special cases of our algorithms. First column: ``sym'' means ``symmetric'' and ``s.p.d.'' means ``symmetric positive definite''. Block versions of all these updates are obtained by choosing $S$ as a matrix with more than one column (i.e., not as a vector).}\label{tbl:QN} \end{table} \subsection{One Step Update}\label{subsec:1step} We have the freedom to select $S$ as almost any random matrix that has full column rank. This includes choosing $S$ to be a constant and invertible matrix, such as the identity matrix $I$, in which case $X_{1}$ must be equal to the inverse. Indeed, the sketch-and-project formulations of all our algorithms reveal that. For Algorithm~\ref{alg:asym-row}, for example, the sketched system is $S^\top AX = S^\top $, which is equivalent to $AX=I$, which has as its unique solution $X=A^{-1}$. Hence, $X_1 = A^{-1}$, and we have convergence in one iteration/step. Through inspection of the complexity rate, we see that $B^{-1/2}\E{Z}B^{-1/2} = I$ and $\rho = \lambda_{\min}(B^{-1/2}\E{Z}B^{-1/2}) =1$, thus this one step convergence is predicted in theory by Theorems~\ref{theo:normEconv} and~\ref{theo:Enormconv}. \subsection{Simultaneous randomized Kaczmarz update}\label{subsec:09u09u09Kacz} Perhaps the most natural choice for the weighting matrix $B$ is the identity $B=I.$ With this choice, Algorithm~\ref{alg:asym-row} is equivalent to applying the randomized Kaczmarz update simultaneously to the $n$ linear systems encoded in $AX=I$. To see this, note that the sketch-and-project viewpoint~\eqref{eq:NF} of Algorithm~\ref{alg:asym-row} is \begin{align} \label{eq:NFbadbroyadj} X_{k+1} = &\arg \min_{X\in \mathbb{R}^{n\times n}} \frac{1}{2}\norm{X - X_{k}}_{F}^2 \quad \mbox{subject to }\quad S^\top AX =S^\top , \end{align} which, by~\eqref{eq:Xupdate}, results in the explicit update \begin{equation} \label{eq:badbroyadj} X_{k+1} = X_{k}+ A^\top S(S^\top AA^\top S)^{-1}S^\top (I-AX_{k}). \end{equation} If $S$ is a random coordinate vector, then~\eqref{eq:NFbadbroyadj} is equivalent to projecting the $j$th column of $X_k$ onto the solution space of $A_{i:}x = \delta_{ij},$ which is exactly an iteration of the randomized Kaczmarz update applied to solving $Ax=e_j.$ In particular, if $S = e_i$ with probability $p_i = \norm{A_{i:}}_2^2/\norm{A}_F^2$ then according to~\eqref{eq:rhoconv}, the rate of convergence of update~\eqref{eq:badbroyadj} is given by \[\E{\norm{X_{k} -A^{-1} }_F^2} \leq \left(1- \frac{1}{\kappa_{2,F}^2(A)}\right)^k \norm{X_{0} -A^{-1} }_F^2 \] where we used that $\kappa_{2,F}(A) = \kappa_{2,F}(A^\top ).$ This is exactly the rate of convergence given by Strohmer and Vershynin in~\cite{Strohmer2009} for the randomized Kaczmarz method. \subsection{Randomized bad Broyden update}\label{subsec:09u09u09} The update~\eqref{eq:badbroyadj} can also be viewed as an adjoint form of the bad Broyden update~\cite{Broyden1965,Griewank2012}. To see this, if we use Algorithm~\ref{alg:asym-col} with $B=I$, then the iterative process is \begin{equation} \label{eq:badbroy} X_{k+1} = X_{k}+ (I-X_{k}A)S(S^\top A^\top AS)^{-1}S^\top A^\top . \end{equation} This update~\eqref{eq:badbroy} is a randomized block form of the {\em bad Broyden update}~\cite{Broyden1965,Griewank2012}. In the quasi-Newton setting, $S$ is not random, but rather the previous step direction $S = \delta \in \mathbb{R}^n$. Furthermore, if we rename $\gamma \eqdef AS\in \mathbb{R}^n$, then~\eqref{eq:badbroy} becomes \begin{equation}\label{eq:badbroyden} X_{k+1} = X_{k}+ \frac{\delta-X_k\gamma}{\norm{\gamma}_2^2}\gamma^\top , \end{equation} which is the standard way of writing the bad Broyden update~\cite{Griewank2012}. The update~\eqref{eq:badbroyadj} is an adjoint form of the bad Broyden in the sense that, if we transpose~\eqref{eq:badbroyadj}, then set $S = \delta$ and denote $\gamma = A^\top S$, we obtain the bad Broyden update, but applied to $X_k^\top $ instead. From the constrain-and-approximate viewpoint~\eqref{eq:RFcols} we give a new interpretation to the bad Broyden update, namely, the update~\eqref{eq:badbroyden} can be written as \[X_{k+1} = \arg_X \min_{X \in \mathbb{R}^{n\times n}, \; y\in \mathbb{R}^n} \frac{1}{2}\norm{X - A^{-1}}_{F}^2 \quad \mbox{subject to } \quad X=X_{k} +y \gamma^\top . \] Thus, {\em the bad Broyden update is the best rank-one update approximating the inverse.} We can determine the rate at which our randomized variant of the BB update~\eqref{eq:badbroy} converges by using~\eqref{eq:rhoconv2}. In particular, if $S =S_i$ with probability $p_i = \left.\norm{A S_i}_F^2\right/\norm{A\mathbf{S}}_F^2$, then~\eqref{eq:BFGSas} converges with the rate \[\E{\norm{X_{k} -A^{-1} }_F^2} \leq \left(1- \frac{1}{\kappa_{2,F}^2(A\mathbf{S})}\right)^k \norm{X_{0} -A^{-1} }_F^2. \] \subsection{Randomized Powell-Symmetric-Broyden update}\label{subsec:PSB} If $A$ is symmetric and we use Algorithm~\ref{alg:sym} with $B=I$, the iterates are given by \begin{align} X_{k+1} &= X_{k} +AS^\top (S^\top A^2S)^{-1}SA (X_{k}AS -S)\left((S^\top A^2S)^{-1} S^\top A-I \right) \nonumber \\ &-(X_{k}AS -S)(S^\top A^2S)^{-1} S^\top A, \label{eq:PSB} \end{align} which is a randomized block form of the Powell-Symmetric-Broyden update~\cite{Gower2014c}. If $S =S_i$ with probability $p_i = \norm{AS_i}_F^2/\norm{A\mathbf{S}}_F^2$, then according to~\eqref{eq:rhoconv}, the iterates~\eqref{eq:PSB} and~\eqref{eq:badbroyadj} converge according to \[\E{\norm{X_{k} -A^{-1} }_F^2} \leq \left(1 - \frac{1}{\kappa^2_{2,F}(A^\top \mathbf{S})}\right)^k \norm{X_{0} - A^{-1}}_F^2.\] \subsection{Randomized good Broyden update} \label{subsec:09u09u09good} Next we present a method that shares certain properties with Gaussian elimination and can be viewed as a randomized block variant of the good Broyden update~\cite{Broyden1965,Griewank2012}. This method requires the following adaptation of Algorithm~\ref{alg:asym-col}: instead of sketching the inverse equation, consider the update~\eqref{eq:guasssketch} that performs a column sketching of the equation $XA^{-1} = I$ by right multiplying with $Ae_i$, where $e_i$ is the $i$th coordinate vector. Projecting an iterate $X_k$ onto this sketched equation gives \begin{align}\label{eq:guasssketch} X_{k+1} = &\arg \min_{X \in \mathbb{R}^{n\times n}} \frac{1}{2} \norm{X - X_{k}}_{F}^2 \quad \mbox{subject to }\quad Xe_i =Ae_i. \end{align} The iterates defined by the above are given by \begin{equation}\label{eq:gaussinv}X_{k+1}=X_{k} +(A-X_k)e_ie_i^\top . \end{equation} Given that we are sketching and projecting onto the solution space of $XA^{-1}=I$, the iterates of this method converge to $A$. Therefore the inverse iterates $X^{-1}_k$ converge to $A^{-1}.$ We can efficiently compute the inverse iterates by using the Woodbury formula~\cite{Woodbury1950} which gives \begin{equation}\label{eq:gauss} X_{k+1}^{-1} = X_k^{-1}-\frac{(X_k^{-1}A-I)e_ie_i^\top X_k^{-1}}{e_i^\top X_k^{-1}Ae_i}. \end{equation} This update~\eqref{eq:gauss} behaves like Gaussian elimination in the sense that, if $i$ is selected in a cyclic fashion, that is $i=k$ on the $k$th iteration, then from~\eqref{eq:gaussinv} it is clear that \[ X_{k+1}e_i = Ae_i, \quad \mbox{thus} \quad X_{k+1}^{-1}Ae_i =e_i, \quad \mbox{for }i=1\ldots k.\] That is, on the $k$th iteration, the first $k$ columns of the matrix $X_{k+1}^{-1}A$ are equal to the first $k$ columns of the identity matrix. Consequently, $X_n = A$ and $X^{-1}_n = A^{-1}.$ If instead, we select $i$ uniformly at random, then we can adapt~\eqref{eq:rhoconv} by swapping each occurrence of $A^\top $ for $A^{-1}$ and observing that $S_i = Ae_i$ thus $\mathbf{S}=A$. Consequently the iterates~\eqref{eq:gaussinv} converge to $A$ at a rate of \[\rho = 1-\kappa_{2,F}^2 \left(A^{-1}A\right)= 1-\frac{1}{n}, \] and thus the lower bound~\eqref{eq:rholower} is achieved and $X_k$ converges to $A$ according to \[\E{\norm{X_{k} -A }_F^2} \leq \left(1-\frac{1}{n}\right)^k \norm{X_{0} -A}_F^2. \] Despite this favourable convergence rate, this does not say anything about how fast $X^{-1}_k$ converges to $A^{-1}$. Therefore~\eqref{eq:gauss} is not an efficient method for calculating an approximate inverse. If we replace $e_i$ by a \emph{step} direction $\delta_k \in \mathbb{R}^d$, then the update~\eqref{eq:gauss} is known as the {\em good Broyden} update~\cite{Broyden1965,Griewank2012}. \subsection{Approximate inverse preconditioning}\label{sec:AIP} When $A$ is symmetric positive definite, we can choose $B =A$, and Algorithm~\ref{alg:asym-row} is given by \begin{equation} \label{eq:BFGSas} X_{k+1} = X_{k} +S(S^\top AS)^{-1}S^\top (I-AX_{k}). \end{equation} The constrain-and-approximate viewpoint~\eqref{eq:RF} of this update is \[X_{k+1} = \arg_X \min_{X\in \mathbb{R}^{n\times n}, Y\in \mathbb{R}^{n\times q}} \frac{1}{2}\norm{A^{1/2}XA^{1/2} - I}_{F}^2 \quad \mbox{subject to } \quad X=X_{k} + S Y^\top .\] This viewpoint reveals that the update~\eqref{eq:BFGSas} is akin to the Approximate Inverse Preconditioning (\emph{AIP}) methods~\cite{Benzi1999,Gould1998,Kolotilina1993,Huckle2007}. We can determine the rate a which~\eqref{eq:BFGSas} converges using~\eqref{eq:rhoconv}. In particular, if $S =S_i$ with probability $p_i = \Tr{S_i^\top A S_i}/\Tr{\mathbf{S}^\top A\mathbf{S}}$, then~\eqref{eq:BFGSas} converges with rate \begin{equation} \rho \overset{\eqref{eq:rhoconv}}{=} 1- \frac{1}{\kappa_{2,F}^2(A^{1/2}\mathbf{S})} = 1- \frac{\lambda_{\min}(\mathbf{S}^\top A\mathbf{S})}{\Tr{\mathbf{S}^\top A\mathbf{S}}}, \label{eq:BBonv} \end{equation} and according to \[\E{\norm{A^{1/2}X_k A^{1/2} - I }_F^2} \leq \left(1- \frac{\lambda_{\min}(\mathbf{S}^\top A\mathbf{S})}{\Tr{\mathbf{S}^\top A\mathbf{S}}}\right)^k \norm{A^{1/2}X_0A^{1/2} - I}_F^2.\] This, as we will see in Section~\ref{sec:RBFGS}, is the same rate of convergence as a randomized variant of the BFGS method. \subsection{Randomized SR1} \label{subsec:SR1} The Symmetric Rank-1 (SR1) update~\cite{Davidon1968,Murtagh1969} does not explicitly fit into our framework, and nor does it fit into the traditional quasi-Newton framework, since it requires a $B$ that is not positive definite. Despite this, we present the update since it is still commonly used. Before choosing $B$, note that, though our theory requires that $B$ be positive definite, Algorithms~\ref{alg:asym-row} and~\ref{alg:asym-col} can be defined with a matrix $B^{-1}$ even if $B$ does not exist! For instance, when $A$ is symmetric and $B^{-1} = A^{-1}-X_k$ then from~\eqref{eq:Xupdate} or~\eqref{eq:Xupdatecols} we get \begin{equation} \label{eq:SR1} X_{k+1} = X_k +(I-AX_k)^\top S(S^\top (A-AX_kA)S)^{-1}S^\top (I-AX_k). \end{equation} This choice for $B$ presents problems, namely, the update~\eqref{eq:SR1} is not always well defined because it requires inverting $S^\top (A-AX_kA)S$ which is not necessarily invertible. To fix this, we should select the sketching matrix $S$ so that $ S^\top (A-AX_kA)S$ is invertible. But this in turn means that $S$ will depend on $X_k$ and most likely cannot be sampled in an i.i.d fashion. Alternatively, we can use the pseudoinverse of $S^\top (A-AX_kA)S$ in place of the inverse. Since $B$ is not positive definite, our convergence theory says nothing about this update. \subsection{Randomized DFP update}\label{sec:RDFP} If $A$ is symmetric positive definite then we can choose $B =A^{-1}.$ Furthermore, if we adapt the sketch-and-project formulation~\eqref{eq:NF} to sketch the equation $XA^{-1} = I$ by right multiplying by $AS,$ and additionally impose symmetry on the iterates, we arrive at the following update. \begin{align}\label{eq:DFPsketch} X_{k+1} = &\arg \min_{X \in \mathbb{R}^{n\times n}} \frac{1}{2} \norm{X - X_{k}}_{F(A^{-1})}^2 \quad \mbox{subject to }\quad XS =AS, \quad X =X^\top . \end{align} The solution to the above is given by\footnote{To arrive at this solution, one needs to swap the occurrences of $AS$ for $S$ and plug in $B=A^{-1}$ in~\eqref{eq:Xupdatesym}. This is because, by swapping $AS$ for $S$ in~\eqref{eq:DFPsketch} gives~\eqref{eq:NFsym}.} \begin{equation}\label{eq:DFPinv} X_{k+1}= AS (S^\top AS)^{-1} S^\top A +\left( I- AS (S^\top AS)^{-1} S^\top \right)X_k\left(I -S (S^\top AS)^{-1} S^\top A\right). \end{equation} Using the Woodbury formula~\cite{Woodbury1950}, we find that \begin{equation}\label{eq:DFP} X_{k+1}^{-1} = X_{k}^{-1} + AS (S^\top AS)^{-1} S^\top A - X_k^{-1} S \left( S^\top X_k^{-1} S \right)^{-1} S^\top X_k^{-1}.\end{equation} The update~\eqref{eq:DFP} is a randomized variant of the Davidon-Fletcher-Powell (DFP) update~\cite{Davidon1959,Fletcher1960}. We can adapt~\eqref{eq:rhoconv} to determine the rate at which $X_k$ converges to $A$ by swapping each occurrence of $A^\top $ for $A^{-1}$. Indeed, for example, let $S_i= Ae_i$ with probability $ p_i = \left. \lambda_{\min}(A) \right/ \Tr{A},$ then the iterates~\eqref{eq:gaussinv} converge to $A$ at a rate of \begin{equation}\label{eq:convDFP} \E{\norm{X_{k} -A }_{F(A^{-1})}^2} \leq \left(1-\frac{\lambda_{\min}(A)}{\Tr{A}}\right)^k \norm{X_{0} - A}_{F(A^{-1})}^2. \end{equation} Thus $X_k$ converges to $A$ at a favourable rate. But this does not indicate at what rate does $X_{k}^{-1}$ converge to $A^{-1}$. This is in contrast to the randomized BFGS, which produces iterates that converge to $A^{-1}$ at this same favourable rate, as we show in the next section. This sheds new light on why BFGS update performs better than the DFP update. \subsection{Randomized BFGS update}\label{sec:RBFGS} If $A$ is symmetric and positive definite, we can choose $B=A$ and apply Algorithm~\ref{alg:sym} to maintain symmetry of the iterates. The iterates are given by \begin{equation}\label{eq:qunac} X_{k+1} =S(S^\top AS)^{-1}S^\top + \left(I-S(S^\top AS)^{-1}S^\top A\right) X_{k} \left(I -AS(S^\top AS)^{-1}S^\top \right). \end{equation} This is a block variant, see~\cite{Gower2014c}, of the BFGS update~\cite{Broyden1965,Fletcher1960,Goldfarb1970,Shanno1971}. The constrain-and-approximate viewpoint gives a new interpretation to the Block BFGS update. That is, from~\eqref{eq:NFsym}, the iterates~\eqref{eq:qunac} can be equivalently defined by \[X_{k+1} = \arg_X \min_{X\in \mathbb{R}^{n\times n}, Y\in \mathbb{R}^{n\times q}} \frac{1}{2}\norm{XA -I}_{F}^2 \quad \mbox{subject to} \quad X = X_k + SY^\top +YS^\top .\] Thus the block BFGS update, and the standard BFGS update, can be seen as a method for calculating an approximate inverse subject to a particular symmetric affine space passing through $X_k.$ This is a completely new way of interpreting the BFGS update. If $p_i = \Tr{S_i^\top AS_i}/\Tr{\mathbf{S}A\mathbf{S}^\top }$, then according to ~\eqref{eq:rhoconv}, the update~\eqref{eq:qunac} converges according to \begin{equation}\label{eq:convqunac} \E{\norm{X_{k}A -I }_{F}^2} \leq \left(1 - \frac{1}{\kappa^2_{2,F}(A^{1/2}\mathbf{S})}\right)^k \norm{X_{0}A - I}_{F}^2. \end{equation} A remarkable property of the update~\eqref{eq:qunac} is that it preserves positive definiteness of $A$. Indeed, assume that $X_k$ is positive definite and let $v \in \mathbb{R}^n$ and $P \eqdef S(S^\top AS)^{-1}S^\top .$ Left and right multiplying~\eqref{eq:qunac} by $v^\top $ and $v$, respectively, gives \[v^\top X_{k+1}v = v^\top Pv+ v^\top \left(I-PA\right) X_{k} \left(I -AP\right)v \geq 0.\] Thus $v^\top X_{k+1}v =0$ implies that $Pv=0$ and $ \left(I -AP \right)v=0,$ which when combined gives $v =0.$ This proves that $X_{k+1}$ is positive definite. Thus the update~\eqref{eq:qunac} is particularly well suited for calculating the inverse of a positive definite matrices. In Section~\eqref{sec:AdaRBFGS}, we detail an update designed to improve the convergence rate in~\eqref{eq:convqunac}. The result is a method that is able to invert large scale positive definite matrices orders of magnitude faster than the state-of-the-art. \subsection{Randomized Column update}\label{sec:RCM} We now describe an update that has no connection to any previous updates, yet the convergence rate we determine~\eqref{eq:RRMconv} is favourable, and comparable to all the other updates we develop. For this update, we need to perform a linear transformation of the sampling matrices. For this, let $V$ be a complete discrete sampling where $V= V_i\in \mathbb{R}^{n\times q_i}$ with probability $p_i>0,$ for $i =1,\ldots,r.$ Let $\mathbf{V} =[V_1,\ldots, V_r].$ Let the sampling matrices be defined as $S_i=AV_i \in \mathbb{R}^{n \times q_i}$ for $i =1,\ldots,r$. As $A$ is nonsingular, and $\mathbf{S} = A \mathbf{V}$, then $S$ is a complete discrete sampling. With these choices and $B=A^\top A$, the sketch-and-project viewpoint~\eqref{eq:NF} is given by \begin{align} X_{k+1} =\arg \min_{X\in \mathbb{R}^{n\times n}} \frac{1}{2}\norm{X - X_{k}}_{F(A^\top A)}^2 \nonumber\quad \mbox{subject to } \quad V_i^\top A^\top AX = V_i^\top A^\top . \end{align} The solution to the above are the iterates of Algorithm~\ref{alg:asym-row}, which is given by \begin{equation}\label{eq:RRM} X_{k+1} = X_{k} +V_i(V_i^\top A^\top AV_i)^{-1}V_i^\top (A^\top -A^\top AX_{k}). \end{equation} From the constrain-and-approximate viewpoint~\eqref{eq:RF}, this can be written as \[X_{k+1} = \arg \min_{X\in \mathbb{R}^{n\times n}, Y\in \mathbb{R}^{n\times q}} \frac{1}{2}\norm{A(XA^\top -I)}_{F}^2 \quad \mbox{subject to } \quad X=X_{k} + V_i Y^\top.\] With these same parameter choices for $S$ and $B$, the iterates of Algorithm~\ref{alg:sym} are given by \begin{align} X_{k+1} &=X_{k} + V_i(V_i^\top A^2 V_i)^{-1}V_i^\top (AX_{k} -I)\left(A^2V_i(V_i^\top A^2V_i)^{-1}V_i^\top - I \right)\nonumber \\ & -(X_{k}A -I)AV_i(V_i^\top A^2 V_i)^{-1}V_i^\top . \label{eq:symROM} \end{align} If we choose $p_i = \norm{(AA^\top )^{-1/2}AA^\top V_i}_F^2 /\norm{(AA^\top )^{-1/2}AA^\top \mathbf{V}}_F^2 = \norm{A^\top V_i}_F^2/\norm{A^\top \mathbf{V}}_F^2,$ then according to ~\eqref{eq:rhoconv}, the iterates~\eqref{eq:RRM} and~\eqref{eq:symROM} converge exponentially in expectation to the inverse according to \begin{equation}\label{eq:RRMconv} \E{\norm{A(X_kA^\top -I) }_{F}^2} \leq \left(1 - \frac{1}{\kappa^2_{2,F}(A\mathbf{V})}\right)^k \norm{A(X_0A^\top -I)}_{F}^2. \end{equation} There also exists an analogous ``row'' variant of~\eqref{eq:RRM}, which arises by using Algorithm~\ref{alg:asym-col}, but we do not explore it here. \section{AdaRBFGS: Adaptive Randomized BFGS}\label{sec:AdaRBFGS} All the updates we have developed thus far use a sketching matrix $S$ that is sampled in an i.i.d.\ fashion from a fixed distribution $\cal D$ at each iteration. In this section we assume that $A$ is symmetric positive definite, and propose AdaRBFGS: a variant of the RBFGS update, discussed in Section~\ref{sec:RBFGS}, which {\em adaptively} changes the distribution $\cal D$ throughout the iterative process. Due to this change, Theorems~\ref{theo:normEconv} and~\ref{theo:Enormconv} are no longer applicable. Superior numerical efficiency of this update is verified through extensive numerical experiments in Section~\ref{sec:numerical}. \subsection{Motivation} We now motivate the design of this new update by examining the convergence rate~\eqref{eq:convqunac} of the RBFGS iterates~\eqref{eq:qunac}. Recall that in RBFGS we choose $B=A$ and $S=S_i$ with probability \begin{equation}\label{eq:ihsih0(0hI}p_i = \Tr{S_i^\top AS_i}/\Tr{\mathbf{S}A\mathbf{S}^\top }, \quad i=1,2,\dots,r,\end{equation} where $S$ is a complete discrete sampling and ${\bf S} = [S_1,\dots,S_r]$. The convergence rate is \[\rho = 1 - \frac{1}{\kappa_{2,F}^2(A^{1/2}\mathbf{S})} \overset{\eqref{eq:kappalower}}{=} 1 - \frac{\lambda_{\min}({\bf S}^\top A {\bf S})}{\Tr{{\bf S}^\top A {\bf S}}}.\] Consider now the question of choosing the matrix $\bf S$ in such a way that $\rho$ is as small as possible. Note that the optimal choice is any $\bf S$ such that \[{\bf S}^\top A {\bf S}= I.\] Indeed, then $\rho = 1-1/n$, and the lower bound~\eqref{eq:kappalower} is attained. For instance, the choice ${\bf S} = A^{-1/2}$ would be optimal. This means that in each iteration we would choose $S$ to be a random column (or random column submatrix) of $A^{-1/2}$. Clearly, this is not a feasible choice, as we do not know the inverse of $A$. In fact, it is $A^{-1}$ which we are trying to find! However, this leads to the following interesting observation: {\em the goals of finding the inverse of $A$ and of designing an optimal distribution $\cal D$ are in synchrony.} \subsection{The algorithm} While we do not know $A^{-1/2}$, we can use the information of the iterates $\{X_k\}$ themselves to construct a good {\em adaptive} sampling. Indeed, the iterates contain information about the inverse and hence we can use them to design a better sampling $S$. In order to do so, it will be useful to maintain a factored form of the iterates, \begin{equation}\label{eq:X_k_98987H}X_k = L_kL_k^\top ,\end{equation} where $L_k \in \mathbb{R}^{n\times n}$ is invertible. With this in place, let us choose $S$ to be a random column submatrix of $L_k$. In particular, let $C_1,C_2,\dots,C_r$ be nonempty subsets that form a partition of $\{1,2,\dots,n\}$, and at iteration $k$ choose \begin{equation}\label{eq:siug98sUJi}S = L_k I_{:C_i} \eqdef S_i, \end{equation} with probability $p_i$ given by \eqref{eq:ihsih0(0hI} for $i=1,2,\dots,r$. For simplicity, assume that $C_1=\{1,\dots,c_1\}$, $C_2 = \{c_1+1,\dots,c_2\}$ and so on, so that, by the definition of $\bf S$, we have \begin{equation}\label{eq:S_9898y8}{\bf S} = [S_1, \dots, S_r] = L_k.\end{equation} Note that now both $\bf S$ and $p_i$ depend on $k$. The method described above satisfies the following recurrence. \begin{theorem}\label{lem:9hs9hIKKK} Consider one step of the AdaRBFGS method described above. Then \begin{equation}\label{eq:98h98hJJJ}\E{\norm{X_{k+1} -A^{-1}}_{F(A)}^2\, | \, X_k} \leq \left(1-\frac{\lambda_{\min }(AX_k )}{\Tr{AX_k}}\right) \norm{X_k -A^{-1}}_{F(A)}^2. \end{equation} \end{theorem} \begin{proof} Using the same arguments as those in the proof of Theorem~\ref{theo:Enormconv}, we obtain \begin{equation}\label{eq:Ec} \E{\norm{X_{k+1} -A^{-1}}_{F(A)}^2\, | \, X_k} \leq \left(1-\lambda_{\min}\left(A^{-1/2}\E{Z \;|\; X_k}A^{-1/2}\right)\right) \norm{X_k -A^{-1}}_{F(A)}^2, \end{equation} where \begin{equation}\label{eq:Zxxx} Z \overset{\eqref{eq:Zxxx}}{=}AS_i(S_i^\top A S_i)^{-1}S_i^\top A. \end{equation} So, we only need to show that \[\lambda_{\min}\left(A^{-1/2}\E{Z \;|\; X_k}A^{-1/2}\right) \geq \frac{\lambda_{\min }(AX_k )}{\Tr{AX_k}}.\] Since $S$ is a complete discrete sampling, Proposition~\ref{prop:Ediscrete} applied to our setting says that \begin{equation}\label{eq:hYhGfR6}\E{Z \;|\; X_k} = A \mathbf{S} D^2 \mathbf{S}^\top A, \end{equation} where \begin{equation} \label{eq:Dxxx} D~\eqdef~\mbox{Diag}\left( \sqrt{p_1}(S_1^\top A S_1)^{-1/2}, \ldots, \sqrt{p_r}(S_r^\top A S_r)^{-1/2}\right).\end{equation} We now have \begin{eqnarray*} \lambda_{\min}\left(A^{-1/2}\E{Z \;|\; X_k}A^{-1/2}\right) &\overset{\eqref{eq:hYhGfR6}+\eqref{eq:S_9898y8}}{\geq}& \lambda_{\min}\left(A^{1/2}L_k L_k^\top A^{1/2}\right) \lambda_{\min}(D^2) \\ & \overset{\eqref{eq:X_k_98987H}}{=} & \frac{\lambda_{\min }(AX_k )}{\lambda_{\max}(D^{-2})}\\ &\overset{\eqref{eq:Dxxx}}{=}& \frac{\lambda_{\min }(AX_k )}{\max_{i} \lambda_{\max}(S_i^\top A S_i) / p_i}\\ &\geq & \frac{\lambda_{\min }(AX_k )}{\max_{i} \Tr{S_i^\top A S_i} / p_i}\\ &\overset{\eqref{eq:ihsih0(0hI} + \eqref{eq:S_9898y8}}{=} & \frac{\lambda_{\min }(AX_k )}{\Tr{A X_k}}, \end{eqnarray*} where in the second equality we have used the fact that the largest eigenvalue of a block diagonal matrix is equal to the maximum of the largest eigenvalues of the blocks. \end{proof} If $X_k$ converges to $A^{-1}$, then necessarily the one-step rate of AdaRBFGS proved in Theorem~\ref{lem:9hs9hIKKK} asymptotically reaches the lower bound \[\rho_k \eqdef 1 - \frac{\lambda_{\min}(AX_k)}{\Tr{A X_k}} \to 1-\frac{1}{n}.\] In other words, as long as this method works, the convergence rate gradually improves, and becomes asymptotically optimal and independent of the condition number. We leave a deeper analysis of this and other adaptive variants of the methods developed in this chapter to future work. \subsection{Implementation} To implement the AdaRBFGS update, we need to maintain the iterates $X_k$ in the factored form \eqref{eq:X_k_98987H}. Fortunately, a factored form of the update~\eqref{eq:qunac} was introduced in~\cite{Gratton2011}, which we shall now describe and adapt to our objective. Assuming that $X_k$ is symmetric positive definite such that $X_k = L_k L_k^\top $, we shall describe how to obtain a corresponding factorization of $X_{k+1}$. Letting $G_k = (S^\top L_k^{-T}L_k^{-1}S)^{1/2}$ and $R_k= (S^\top AS)^{-1/2}$, it can be verified through direct inspection~\cite{Gratton2011} that $X_{k+1} = L_{k+1} L_{k+1}^\top $, where \begin{equation}\label{eq:Chol}L_{k+1} = L_k +S R_k \left(G_k^{-1}S^\top L_k^{-T}-R_k ^\top S^\top A L_k\right). \end{equation} If we instead of \eqref{eq:siug98sUJi} consider the more general update $S = L_k \tilde{S}$, where $\tilde{S}$ is chosen in an i.i.d.\ fashion from some fixed distribution $\cal \tilde{D}$, then \begin{equation} L_{k+1} = L_k +L_k \tilde{S} R_k \left(({\tilde{S}}^\top \tilde{S})^{-1/2}\tilde{S}^\top -R_k^\top \tilde{S}^\top L_k^\top A L_k\right). \label{eq:Cholimprov} \end{equation} The above can now be implemented efficiently, see Algorithm~\ref{alg:AdaRBFGS}. \begin{algorithm}[!h] \begin{algorithmic}[1] \State \textbf{input:} symmetric positive definite matrix $A$ \State \textbf{parameter:} ${\cal \tilde{D}}$ = distribution over random matrices with $n$ rows \State \textbf{initialize:} pick invertible $L_0 \in \mathbb{R}^{n\times n}$ \For {$k = 0, 1, 2, \dots$} \State Sample an independent copy $\tilde{S} \sim {\cal \tilde{D}}$ \State Compute $S = L_k \tilde{S}$ \Comment $S$ is sampled adaptively, as it depends on $k$ \State Compute $R_k = (\tilde{S}^\top A \tilde{S})^{-1/2}$ \State $L_{k+1} = L_k +S R_k \left((\tilde{S}^\top \tilde{S})^{-1/2}\tilde{S}^\top -R_k^\top S^\top A L_k\right)$ \Comment Update the factor \EndFor \State \textbf{output:} $X_k = L_k L_k^\top $ \end{algorithmic} \caption{ Adaptive Randomized BFGS (AdaRBFGS)} \label{alg:AdaRBFGS} \end{algorithm} In Section~\ref{sec:numerical} we test two variants based on~\eqref{eq:Cholimprov}. The first is the \emph{AdaRBFGS\_gauss} update, in which the entries of $\tilde{S}$ are standard Gaussian. The second is \emph{AdaRBFGS\_cols}, where $\tilde{S} =I_{:C_i}$, as described above, and $|C_i| =q$ for all $i$ for some $q$. \section{Numerical Experiments} \label{sec:numerical} Given the demand for approximate inverses of positive definite matrices in preconditioning and in variable metric methods in optimization, and the author's own interest in the aforementioned applications, we restrict our test to inverting positive definite matrices. We test four iterative methods for inverting matrices. This rules out the all-or-nothing direct methods such as Gaussian elimination of LU based methods. For our tests we use two variants of Algorithm~\ref{alg:AdaRBFGS}: AdaRBFGS\_gauss, where $\tilde{S}\in\mathbb{R}^{n\times q}$ is a normal Gaussian matrix, and AdaRBFGS\_cols, where $\tilde{S}$ consists of a collection of $q$ distinct coordinate vectors in $\mathbb{R}^n$, selected uniformly at random. At each iteration the AdaRBFGS methods compute the inverse of a small matrix $S^\top AS$ of dimension $q\times q$. To invert this matrix we use MATLAB's inbuilt \texttt{inv} function, which uses $LU$ decomposition or Gaussian elimination, depending on the input. Either way, \texttt{inv} costs $O(q^3).$ For simplicity, we selected $q=\sqrt{n}$ in all our tests. We compare our method to two well established and competitive methods, the \emph{Newton-Schulz method}~\cite{Schulz1933} and the global self-conditioned Minimal Residual (\emph{MR}) method~\cite{Chow1998}. The Newton-Schulz method arises from applying the Newton-Raphson method to solve the equation $X^{-1}=A$, which gives \begin{equation}\label{eq:Newton-Schulz} X_{k+1} = 2X_{k} -X_{k} A X_{k}. \end{equation} The MR method was designed to calculate approximate inverses, and it does so by minimizing the norm of the residual along the preconditioned residual direction, that is \begin{equation} \label{eq:MRdef} \norm{I-AX_{k+1}}_F^2 = \min_{\alpha \in \mathbb{R}} \left\{\norm{I-AX}_F^2 \quad \mbox{subject to} \quad X = X_k +\alpha X_k(I-AX_k)\right\}, \end{equation} see~\cite[chapter 10.5]{Saad2003} for a didactic introduction to MR methods. The resulting iterates of the MR method are given by \begin{equation} \label{eq:MRiter} X_{k+1} =X_k +\frac{\Tr{R_k^\top AX_kR_k}}{\Tr{(AX_kR_k)^\top AX_kR_k}}X_kR_k, \end{equation} where $R_k = I-AX_k$. We perform two sets of tests. On the first set, we choose a different starting matrix for each method which is optimized, in some sense, for that method. We then compare the empirical convergence of each method, including the time taken to calculate $X_0$. In particular, the Newton-Schulz is only guaranteed to converge for an initial matrix $X_0$ such that $\rho(I-X_0A)< 1$. Indeed, the Newton-Schulz method did not converge in most of our experiments when $X_0$ was not carefully chosen according to this criteria. To remedy this, we choose $X_0 = 0.99 \cdot A^\top /\rho^2(A)$ for the Newton-Schulz method\footnote{During a revision of this thesis, it was kindly pointed out by the committee that setting $X_0 = \alpha A^\top$ with $\alpha = 1/\norm{A}_F^2$ would suffice to guarantee convergence and avoid the cost of calculating $\rho(A).$ Though we should note that the cost in calculating $\rho(A)$ was partly offset by using a C++ implementation, while all other methods were coded in native MATLAB. }, so that $\rho(I-X_0A)<1$ is satisfied. To compute $\rho(A)$ we used the inbuilt MATLAB function \texttt{normest} which is coded in C++. While for MR we followed the suggestion in~\cite{Saad2003} and used the projected identity for the initial matrix $X_0 = (\Tr{A}/\Tr{AA^\top }) \cdot I.$ For our AdaRBFGS methods we simply used $X_0 =I$, as this worked well in practice. In the second set of tests, which we relegate to the Appendix of this chapter, we compare the empirical convergence of the methods starting from the same matrix, namely the identity matrix $X_0 = I$. We run each method until the relative residual $\norm{I-AX_{k}}_{F}/\norm{I-AX_0}_{F}$ is below $10^{-2}.$ All experiments were performed and run in MATLAB R2014b. To appraise the performance of each method we plot the relative residual against time taken and against the number of floating point operations (\emph{flops}). \subsection{Experiment 1: synthetic matrices} First we compare the four methods on synthetic matrices generated using the \texttt{rand} function as follows: $A~=~\bar{A}^\top \bar{A}$ where $\bar{A}=$\texttt{rand}$(n)$. The resulting matrix $A$ is positive definite with high probability. To appraise the difference in performance of the methods as the dimension of the problem grows, we tested for $n=1000$, $2000$ and $5000.$ As the dimension grows, only the two variants of the AdaRBFGS method are able to reach the $10^{-2}$ desired tolerance in a reasonable amount time and number of flops (see Figure~\ref{fig:pdsynth}). \begin{figure} \centering \begin{subfigure}[t]{0.80\textwidth} \centering \includegraphics[width = \textwidth, trim= 40 300 50 300, clip ]{randn-1000} \caption{\texttt{rand} with $n=10^4$ }\label{fig:rand} \end{subfigure} \begin{subfigure}[t]{0.80\textwidth} \centering \includegraphics[width = \textwidth, trim= 40 300 50 300, clip ]{randn-2000} \caption{\texttt{rand} with $n=2 \cdot 10^4$ }\label{fig:rand2} \end{subfigure}\\ \begin{subfigure}[t]{0.80\textwidth} \centering \includegraphics[width = \textwidth, trim= 40 300 50 300, clip ]{randn-5000} \caption{\texttt{rand} with $n=5 \cdot 10^4$ }\label{fig:rand3} \end{subfigure}% \caption{Synthetic MATLAB generated problems. Uniform random matrix $A~=~\bar{A}^\top \bar{A}$ where $\bar{A}=$\texttt{rand}$(n).$ }\label{fig:pdsynth} \end{figure} \subsection{Experiment 2: LIBSVM matrices} Next we invert the Hessian matrix $\nabla^2 f(x)$ of four ridge-regression problems of the form \begin{equation}\label{eq:ridgeMatrix} \min_{x\in \mathbb{R}^n}f(x)\eqdef \frac{1}{2}\norm{Ax-b}_2^2 + \frac{\lambda}{2} \norm{x}_2^2,\quad \quad\nabla^2 f(x) = A^\top A+\lambda I, \end{equation} using data from LIBSVM~\cite{Chang2011}, see Figure~\ref{fig:LIBSVM}. We use $\lambda =1$ as the regularization parameter. On the two problems of smaller dimension, \texttt{aloi} and \texttt{protein}, the four methods have a similar performance, and encounter the inverse in a few seconds. On the two larger problems, \texttt{gisette-scale} and \texttt{real-sim}, the two variants of AdaRBFGS significantly outperform the MR and the Newton-Schulz method. \begin{figure} \centering \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim= 40 300 50 300, clip ]{aloi} \caption{\texttt{aloi}} \end{subfigure}\\ \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim=40 300 50 300, clip ]{protein} \caption{\texttt{protein}} \end{subfigure}\\ \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim=40 300 50 300, clip ]{gisette-scale} \caption{\texttt{gisette\_scale}} \end{subfigure}\\ \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim= 40 300 50 300, clip ]{real-sim} \caption{\texttt{real\_sim}} \end{subfigure} \caption{The performance of Newton-Schulz, MR, AdaRBFGS\_gauss and AdaRBFGS\_cols methods on the Hessian matrix of four LIBSVM test problems: (a) \texttt{aloi}: $(m;n)=(108,000;128)$ (b) \texttt{protein}: $(m; n)=(17,766; 357)$ (c) \texttt{gisette\_scale}: $(m;n)=(6000; 5000)$ (d) \texttt{real-sim}: $(m;n)=(72,309; 20,958)$.} \label{fig:LIBSVM} \end{figure} \subsection{Experiment 3: UF sparse matrices} For our final batch of tests, we calculate an approximate inverse of several sparse matrices from the Florida sparse matrix collection~\cite{Davis:2011}\footnote{One should never calculate an exact inverse of a sparse matrix in practice since the resulting matrix is dense. But given that the AdaBFGS method performs low rank updates, this allows us to \emph{implicitly} form an approximate inverse by storing the updates, and not explicitly forming the matrix. We explore this in the paper~\cite{GowerGold2016}}. We have selected six problems from six different applications, so that the set of matrices display a varied sparsity pattern and structure, see Figures~\ref{fig:UF1} and~\ref{fig:UF2}. On the matrix \texttt{Bates/Chem97ZtZ} of moderate size, the four methods perform well, with the Newton-Schulz method converging first in time and AdaRBFGS\_cols first in flops. While on the matrices of larger dimension, the two variants of AdaRBFGS converge much faster, often orders of magnitude before the MR and Newton-Schulz method reach the required precision. \begin{figure} \centering \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim=40 300 50 300, clip ]{Bates-Chem97ZtZ} \caption{\texttt{Bates/Chem97ZtZ}} \end{subfigure}\\ \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim= 40 300 50 300, clip ]{FIDAP-ex9} \caption{\texttt{FIDAP/ex9}} \end{subfigure} \\% \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim=40 300 50 300, clip ]{Nasa-nasa4704} \caption{\texttt{Nasa/nasa4704}} \end{subfigure} \\ \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim=40 300 50 300, clip ]{HB-bcsstk18} \caption{\texttt{HB/bcsstk18}} \end{subfigure}% \caption{The performance of Newton-Schulz, MR, AdaRBFGS\_gauss and AdaRBFGS\_cols on (a) \texttt{Bates-Chem97ZtZ}: $n= 2\,541$, (b) \texttt{FIDAP/ex9}: $n = 3,\,363 $, (c) \texttt{Nasa/nasa4704}: $n= 4\,,704$, (d) \texttt{HB/bcsstk18}: $n=11,\,948$. } \label{fig:UF1} \end{figure} \begin{figure} \centering \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim=40 300 50 300, clip ]{Pothen-bodyy4} \caption{\texttt{Pothen/bodyy4}} \end{subfigure} \\%\hspace{0.2\textwidth} \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim=40 300 50 300, clip ]{ND-nd6k} \caption{\texttt{ND/nd6k}} \end{subfigure}\\% \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim=40 300 50 300, clip ]{GHS-psdef-wathen100-M0-variable} \caption{\texttt{GHS\_psdef/wathen100}} \end{subfigure}% \caption{The performance of Newton-Schulz, MR, AdaRBFGS\_gauss and AdaRBFGS\_cols on (a) \texttt{Pothen/bodyy4}: $n =17,\,546$ (b) \texttt{ND/nd6k}: $n=18,\,000$ (c) \texttt{GHS\_psdef/wathen100}: $n = 30, \,401$. } \label{fig:UF2} \end{figure} The significant difference between the performance of the methods on large scale problems can be, in part, explained by their iteration cost. The iterates of the Newton-Schulz and MR method compute $n\times n$ matrix-matrix products. While the cost of an iteration of the AdaRBFGS methods is dominated by the cost of a $n\times n$ matrix by $n\times q$ matrix product. As a result, and because we set $q = \sqrt{n},$ this is difference of $n^3$ to $n^{2+1/2}$ in iteration cost, which clearly shows on the larger dimensional instances. \subsection{Conclusion of numeric experiments} Through our extensive numeric experiments, it is clear that the AdaRBFGS methods are highly efficient at calculating a low precision approximate inverse of a positive definite matrix. Furthermore, in many of these experiments the Newton-Schulz and MR method suffer from an initially slow convergence, particularly so on large-scale problems, see Figures~\ref{fig:pdsynth}c,~\ref{fig:LIBSVM}d,~\ref{fig:UF1}b, ~\ref{fig:UF1}c and~\ref{fig:UF1}d for example. But after a sufficient number of iteration, the asymptotic second order convergence rate of the Newton-Schulz and MR method sets in, see Figures~\ref{fig:pdsynth}a,~\ref{fig:pdsynth}b,~\ref{fig:LIBSVM}a ,~\ref{fig:LIBSVM}b and~\ref{fig:LIBSVM}c for example. These experiments also indicate that the AdaRBFGS method enjoys super linear convergence as can be seen, for example, in Figure~\ref{fig:pdsynth} the residual decreases superlinearly in both time and flops. Yet we have still to prove that the AdaRBFGS methods are even linearly convergent. Thus it is now an open question whether or not the AdaRBFGS convergent linearly or superlinearly. \section{Summary} We developed a family of stochastic methods for iteratively inverting matrices, with a specialized variant for asymmetric, symmetric and positive definite matrices. The methods have two dual viewpoints, a sketch-and-project viewpoint which is an extension of the least-change formulation of the quasi-Newton methods, and a constrain-and-approximate viewpoint which is related to the approximate inverse preconditioning (API) methods. The equivalence between these two viewpoints reveals a new connection between the quasi-Newton and the API methods, which were previously considered to be unrelated. Under mild conditions, we prove convergence rates through two different perspectives, the convergence of the expected norm of the error, and the norm of the expected error. Our convergence theorems are general enough to accommodate discrete samplings and continuous samplings, though we only explore discrete sampling here in more detail. For discrete samplings, we determine a probability distribution for which the convergence rates are equal to a scaled condition number, and thus are easily interpretable. Furthermore, for discrete sampling, we determining a practical optimized sampling distribution, that is obtained by minimizing an upper bound on the convergence rate. We develop new randomized block variants of the quasi-Newton updates, including the BFGS update, complete with convergence rates, and provide new insights into these methods using our dual viewpoint. For positive definite matrices, we develop an Adaptive Randomized BFGS methods (AdaRBFGS), which in large-scale numerical experiments, prove to be orders of magnitude faster (in time and flops) then the self-conditioned minimal residual method and the Newton-Schulz method. In particular, only the AdaRBFGS methods are able to approximately invert the $20,958 \times 20,958$ ridge regression matrix based on the \texttt{real-sim} data set in reasonable time and flops. This work opens up many possible venues for future work, including, developing methods that use continuous random sampling, implementing a limited memory approach akin to the LBFGS~\cite{Nocedal1980} method, which could maintain an operator that serves as an approximation to the inverse. As recently shown in~\cite{Gower2015c}, an analogous method applied to linear systems converges with virtually no assumptions on the system matrix. This can be extended to calculating the pseudoinverse matrix, something we leave for future work. \section{Appendix: Optimizing an Upper Bound on the Convergence Rate} \begin{lemma}\label{lem:fracsum} Let $a_1,\dots,a_r$ be positive real numbers. Then \[\left[ \frac{\sqrt{a_1}}{\sum_{i=1}^r \sqrt{a_i}},\ldots, \frac{\sqrt{a_n}}{\sum_{i=1}^r \sqrt{a_i}} \right] =\arg\min_{p\in \Delta_r} \sum_{i=1}^r \frac{a_i}{p_i}.\] \end{lemma} \begin{proof} Incorporating the constraint $\sum_{i=1}^r p_i =1$ into the Lagrangian we have \[\min_{p\geq 0} \sum_{i=1}^r \frac{a_i}{p_i} + \mu\sum_{i=1}^r (p_i-1),\] where $\mu \in \mathbb{R}.$ Differentiating in $p_i$ and setting to zero, then isolating $p_i$ gives \begin{equation} \label{eq:muaipi} p_i = \sqrt{\frac{a_i}{\mu}}, \quad \mbox{for }i=1,\ldots r.\end{equation} Summing over $i$ gives \[ 1 = \sum_{i=1}^r \sqrt{\frac{a_i}{\mu}} \quad \Rightarrow \quad \mu = \left(\sum_{i=1}^r \sqrt{a_i}\right)^2.\] Inserting this back into~\eqref{eq:muaipi} gives $p_i = \sqrt{a_i}/\sum_{i=1}^r \sqrt{a_i}.$ \end{proof} \section{Appendix: Numerical Experiments with the Same Starting Matrix}\label{app:numerics} We now investigate the empirical convergence of the methods MR, AdaRBFGS\_cols and AdaRGFBS\_gauss when initiated with the same starting matrix $X_0 = I,$ see Figures~\ref{fig:LIBSVM2},~\ref{fig:UFapp1} and~\ref{fig:UFapp2}. We did not include the Newton-Schultz method in these figures because it diverged on all experiments when initiated from $X_0 =I.$ Again we observe that, as the dimension grows, only the two variants of the AdaRBFGS are capable of inverting the matrix to the desired $10^{-2}$ precision in a reasonable amount of time. Furthermore, the AdaRBFGS\_gauss variant had the overall best best performance. \begin{figure} \centering \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim= 40 300 50 300, clip ]{aloi-M0-1} \caption{\texttt{aloi}} \end{subfigure} \\% \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim=40 300 50 300, clip ]{protein-M0-1} \caption{\texttt{protein}} \end{subfigure} \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim=40 300 50 300, clip ]{gisette-scale-M0-1} \caption{\texttt{gisette\_scale}} \end{subfigure} \\% \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim= 40 300 50 300, clip ]{real-sim-M0-1} \caption{\texttt{real\_sim}} \end{subfigure} \caption{The performance of Newton-Schulz, MR, AdaRBFGS\_gauss and AdaRBFGS\_cols methods on the Hessian matrix of four LIBSVM test problems: (a) \texttt{aloi}: $(m;n)=(108,000;128)$ (b) \texttt{protein}: $(m; n)=(17,766; 357)$ (c) \texttt{gisette\_scale}: $(m;n)=(6000; 5000)$ (d) \texttt{real-sim}: $(m;n)=(72,309; 20,958)$. The starting matrix $X_0=I$ was used for all methods.} \label{fig:LIBSVM2} \end{figure} \begin{figure} \centering \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim=40 300 50 300, clip ]{Bates-Chem97ZtZ-M0-1} \caption{\texttt{Bates/Chem97ZtZ}} \end{subfigure} \\ \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim= 40 300 50 300, clip ]{FIDAP-ex9-M0-1} \caption{\texttt{FIDAP/ex9}} \end{subfigure} \\% \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim=40 300 50 300, clip ]{Nasa-nasa4704-M0-1} \caption{\texttt{Nasa/nasa4704}} \end{subfigure} \\%\hspace{0.2\textwidth} \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim=40 300 50 300, clip ]{HB-bcsstk18-M0-1} \caption{\texttt{HB/bcsstk18}} \end{subfigure}% \caption{The performance of Newton-Schulz, MR, AdaRBFGS\_gauss and AdaRBFGS\_cols on (a) \texttt{Bates-Chem97ZtZ}: $n= 2\,541$, (b) \texttt{FIDAP/ex9}: $n = 3,\,363 $, (c) \texttt{Nasa/nasa4704}: $n= 4\,,704$, (d) \texttt{HB/bcsstk18}: $n=11,\,948$. The starting matrix $X_0=I$ was used for all methods. } \label{fig:UFapp1} \end{figure} \begin{figure} \centering \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim=40 300 50 300, clip ]{Pothen-bodyy4-M0-1} \caption{\texttt{Pothen/bodyy4}} \end{subfigure} \\ \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim=40 300 50 300, clip ]{ND-nd6k-M0-1} \caption{\texttt{ND/nd6k}} \end{subfigure} \\% \begin{subfigure}[t]{0.65\textwidth} \centering \includegraphics[width = \textwidth, trim=40 300 50 300, clip ]{GHS-psdef-wathen100-M0-1} \caption{\texttt{GHS\_psdef/wathen100}} \end{subfigure}% \caption{The performance of Newton-Schulz, MR, AdaRBFGS\_gauss and AdaRBFGS\_cols on (a) \texttt{Pothen/bodyy4}: $n =17,\,546$ (b) \texttt{ND/nd6k}: $n=18,\,000$ (c) \texttt{GHS\_psdef/wathen100}: $n = 30, \,401$. The starting matrix $X_0=I$ was used for all methods. } \label{fig:UFapp2} \end{figure} \chapter{Conclusion and Future Work} \chaptermark{Conclusion and Future Work} \label{ch:Conclusion} { \epigraph{\emph{Fuils an bairns soud never see things hauf duin.} \\ It needs powers of perception and mature judgement to visualise the result of an enterprise. }{Scottish proverb.} \let\clearpage\relax } This thesis laid the foundational work for a class of randomized methods for solving linear systems, equipt with convergence analysis, and general enough to accommodate several existing methods (Randomized Kaczmarz and Coordinate descent), but also allows for the design of completely new methods with continuous samplings, optimized samplings and block variants. Thus there is still much to explore in designing new randomized methods for solving linear systems. One particularly promising direction is to use new sophisticated sketching matrices $S$, such as the Walsh-Hadamard matrix~\cite{Lu2013,Pilanci2014}, to design new practical and efficient methods. Using duality theory, we redeveloped the sketch-and-project method through a dual perspective which we refer to as the Stochastic Dual Ascent (SDA) method. This perspective allowed us to extend the application of the sketch-and-project methods to that of finding the projection of a given vector onto the solution space of a linear system. This extensions enables us to solve the distributed consensus problem and it reveals that a standard randomized gossip algorithm is a special case of the sketch-and-project methods. Applying the same sketch and project principle to the inverse equations ($AX=I$ or $XA=I$), we developed new randomized methods for inverting nonsymmetric and symmetric matrices. These randomized methods can be viewed as randomized quasi-Newton updates. Through a dual perspective, we establish a new connection connection between the quasi-Newton methods and the approximate inverse preconditioning methods. Furthermore, we design a highly efficient method, AdaRBFGS, for calculating approximate inverses of positive definite matrices. This opens up many avenues for developing stochastic preconditioning and variable metric methods. Indeed, the AdaRBFGS method has already been used as the basis of a new stochastic variable metric method~\cite{GowerGold2016}. Perhaps the most exciting direction for future work is to extend the sketch-and-project framework to solve linear equations in other settings. For instance, our framework could be extended to solve linear equations in a general Euclidean place. This would open up several new application areas including linear matrix equations such Sylvester equation, Lyapunov equation and more~\cite{Simoncini2014}. Finally, perhaps the sketch-and-project framework can be extended to solving linear equations defined by bounded linear operators between two Hilbert spaces. Such a development would lead to new randomized methods for solving differential and integral equations. \chapter*{Author's Publications \markboth{Author's Publications}{}} \vspace{.3in} \begin{spacing}{1.6} \noindent \small During the course of the PhD program the candidate has co-authored six papers~\cite{Gower2012Sparsity,Gower2013,Gower2014c,Gower2015,Gower2015c,Gower2016}, all of which he was the first author and main contributor. This thesis is based on three of these papers, with a one-to-one correspondence between a chapter and a paper according to the following table. \begin{center} \begin{tabular}{|c||ccc|} \hline Chapter & \ref{ch:linear_systems} & \ref{ch:SDA} & \ref{ch:inverse} \\ \hline Paper & \cite{Gower2015} & \cite{Gower2015c} & \cite{Gower2016} \\ \hline \end{tabular} \end{center} \vspace{1.75in} \noindent Robert Mansel Gower\\ 29th of February 2016 \end{spacing} \clearpage \chapter*{List of Abbreviations \markboth{List of Abbrevations}{}} \small \begin{tabular}{ll} ASAP & As soon as possible \\ RDY & Ready\\ \end{tabular}\\ [6pt] \clearpage \chapter*{Author's Declaration \markboth{Author's Declaration}{}} \vspace{.3in} \begin{spacing}{1.6} \noindent \small I declare that this thesis has been composed solely by myself and that it has not been submitted, either in whole or in part, in any previous application for a degree. Except where otherwise acknowledged, the work presented is entirely my own. During the course of the PhD program I co-authored six papers~\cite{Gower2012Sparsity,Gower2014,Gower2014c,Gower2015,Gower2015c,Gower2016}, all of which I was the first author. This thesis is based on three of these papers, with the table below indicating which chapters are on which papers. I confirm that I contributed to all the results within the papers on which this thesis is based. \begin{center} \begin{tabular}{|c||cccc|} \hline Chapter & \ref{ch:introduction} & \ref{ch:linear_systems} & \ref{ch:SDA} & \ref{ch:inverse} \\ \hline Paper & \cite{Gower2015,Gower2015c,Gower2016} &\cite{Gower2015} & \cite{Gower2015c} & \cite{Gower2016} \\ \hline \end{tabular} \end{center} \vspace{1.75in} \noindent Robert Mansel Gower\\ 29th of February 2016 \end{spacing} \clearpage \chapter*{Lay Summary \markboth{Lay summary}{}} \small This thesis explores the design and analysis of methods (algorithms) for solving two common problems: solving linear systems of equations and inverting matrices. Many engineering and quantitative tasks require the solution of one of these two problems. In particular, the need to solve linear systems of equations is ubiquitous in essentially all quantitative areas of human endeavour, including industry and science. Specifically, linear systems are a central problem in numerical linear algebra, and play an important role in computer science, mathematical computing, optimization, signal processing, engineering, numerical analysis, computer vision, machine learning, and many other fields. This thesis proposes new methods for solving large dimensional linear systems and inverting large matrices that use tools and ideas from probability. The advent of large dimensional linear systems of equations, based on big data sets, poses a challenge. On these large linear systems, the traditional methods for solving linear systems can take an exorbitant amount of time. To address this issue we propose a new class of randomized methods that are capable of quickly obtaining approximate solutions. This thesis lays the foundational work of this new class of randomized methods for solving linear systems and inverting matrices. The main contributions are providing a framework to design and analyze new and existing methods for solving linear systems. In particular, our framework unites many existing methods. For inverting matrices we also provide a framework for designing and analysing methods, but moreover, using this framework we design a highly competitive method for computing an approximate inverse of truly large scale positive definite matrices. Our new method often outperforms previously known methods by several orders of magnitude on large scale matrices. \newpage \chapter*{Abstract \markboth{Abstract}{}} \small Probabilistic ideas and tools have recently begun to permeate into several fields where they had traditionally not played a major role, including fields such as numerical linear algebra and optimization. One of the key ways in which these ideas influence these fields is via the development and analysis of randomized algorithms for solving standard and new problems of these fields. Such methods are typically easier to analyze, and often lead to faster and/or more scalable and versatile methods in practice. This thesis explores the design and analysis of new randomized iterative methods for solving linear systems and inverting matrices. The methods are based on a novel sketch-and-project framework. By sketching we mean, to start with a difficult problem and then randomly generate a simple problem that contains all the solutions of the original problem. After sketching the problem, we calculate the next iterate by projecting our current iterate onto the solution space of the sketched problem. The starting point for this thesis is the development of an archetype randomized method for solving linear systems. Our method has six different but equivalent interpretations: sketch-and-project, constrain-and-approximate, random intersect, random linear solve, random update and random fixed point. By varying its two parameters -- a positive definite matrix (defining geometry), and a random matrix (sampled in an i.i.d. fashion in each iteration) -- we recover a comprehensive array of well known algorithms as special cases, including the randomized Kaczmarz method, randomized Newton method, randomized coordinate descent method and random Gaussian pursuit. We also naturally obtain variants of all these methods using blocks and importance sampling. However, our method allows for a much wider selection of these two parameters, which leads to a number of new specific methods. We prove exponential convergence of the expected norm of the error in a single theorem, from which existing complexity results for known variants can be obtained. However, we also give an exact formula for the evolution of the expected iterates, which allows us to give lower bounds on the convergence rate. We then extend our problem to that of finding the projection of given vector onto the solution space of a linear system. For this we develop a new randomized iterative algorithm: {\em stochastic dual ascent (SDA)}. The method is dual in nature, and iteratively solves the dual of the projection problem. The dual problem is a non-strongly concave quadratic maximization problem without constraints. In each iteration of SDA, a dual variable is updated by a carefully chosen point in a subspace spanned by the columns of a random matrix drawn independently from a fixed distribution. The distribution plays the role of a parameter of the method. Our complexity results hold for a wide family of distributions of random matrices, which opens the possibility to fine-tune the stochasticity of the method to particular applications. We prove that primal iterates associated with the dual process converge to the projection exponentially fast in expectation, and give a formula and an insightful lower bound for the convergence rate. We also prove that the same rate applies to dual function values, primal function values and the duality gap. Unlike traditional iterative methods, SDA converges under virtually no additional assumptions on the system (e.g., rank, diagonal dominance) beyond consistency. In fact, our lower bound improves as the rank of the system matrix drops. By mapping our dual algorithm to a primal process, we uncover that the SDA method is the dual method with respect to the sketch-and-project method from the previous chapter. Thus our new more general convergence results for SDA carry over to the sketch-and-project method and all its specializations (randomized Kaczmarz, randomized coordinate descent...etc). When our method specializes to a known algorithm, we either recover the best known rates, or improve upon them. Finally, we show that the framework can be applied to the distributed average consensus problem to obtain an array of new algorithms. The randomized gossip algorithm arises as a special case. In the final chapter, we extend our method for solving linear system to inverting matrices, and develop a family of methods with specialized variants that maintain symmetry or positive definiteness of the iterates. All the methods in the family converge globally and exponentially, with explicit rates. In special cases, we obtain stochastic block variants of several quasi-Newton updates, including bad Broyden (BB), good Broyden (GB), Powell-symmetric-Broyden (PSB), Davidon-Fletcher-Powell (DFP) and Broyden-Fletcher-Goldfarb-Shanno (BFGS). Ours are the first stochastic versions of these updates shown to converge to an inverse of a fixed matrix. Through a dual viewpoint we uncover a fundamental link between quasi-Newton updates and approximate inverse preconditioning. Further, we develop an adaptive variant of the randomized block BFGS (AdaRBFGS), where we modify the distribution underlying the stochasticity of the method throughout the iterative process to achieve faster convergence. By inverting several matrices from varied applications, we demonstrate that AdaRBFGS is highly competitive when compared to the well established Newton-Schulz and approximate preconditioning methods. In particular, on large-scale problems our method outperforms the standard methods by orders of magnitude. The development of efficient methods for estimating the inverse of very large matrices is a much needed tool for preconditioning and variable metric methods in the big data era. \newpage \chapter*{List of Symbols \markboth{List of Symbols}{}} \vspace{-.1in} \small \noindent \begin{tabular}{ll} $\dotprod{x,y}$ & The standard Euclidean inner product of $x,y \in \mathbb{R}^n$, $\dotprod{x,y} = \sum_{i=1}^n x_i y_i$\\ $\norm{x}_2$ & The standard Euclidean norm of $x \in \mathbb{R}^n,$ $\norm{x}_2 = \sqrt{\dotprod{x,x}}$\\ $\norm{x}_B$ & The norm defined by symmetric positive definite $B \in \mathbb{R}^{n\times n}$, $\norm{x}_B = \sqrt{\dotprod{Bx,x}}$\\ $e^i$ & The $i$th coordinate vector in $\mathbb{R}^m$\\ $f_i$ & The $i$th coordinate vector in $\mathbb{R}^n$\\ $\dim (V)$ & The dimension of $V$ where $V$ is a subspace\\ $\Rank{M}$ & The rank of $M$, where $M$ is a matrix\\ $\Null{M}$ & The nullspace of $M$, $\Null{M}=\{x \, | \, Mx =0 \}$\\ $\myRange{M}$ & The rangespace of $M$, e.g. if $M \in \mathbb{R}^{m\times n}$ then $\myRange{M}=\{Mx \, | \, x \in \mathbb{R}^n\}$\\ $\lambda_{\max}(M)$ & The largest eigenvalue of $M$\\ $\lambda_{\min}(M)$ & The smallest eigenvalue of $M$\\ $\lambda^+_{\min}(M)$ & The smallest nonzero eigenvalue of $M$ (assuming that $M$ is a nonzero matrix) \\ $M^{\dagger}$ & The Moore-Penrose pseudoinverse of $M$.\\ $\Tr{M}$ & The trace of $M$, e.g. if $M \in \mathbb{R}^{n\times n}$ then $\Tr{M} = \sum_{i=1}^n M_{ii}$ \\ $\norm{M}_F$ & The Frobenius norm of $M$, $\norm{M}_F = \sqrt{\Tr{M^\top M}}$ \\ $\norm{M}_2$ & The spectral norm of $M$, $\norm{M}_2 = \max_{\norm{v}_2=1} \norm{Mv}_2 = \sqrt{\lambda_{\max}(M^\top M)}$ \\ $\norm{M}_B$ & $\eqdef \norm{B^{1/2}MB^{-1/2}}_2 = \max_{\norm{v}_B=1} \norm{Mv}_B $ \\ $\norm{M}_B^*$ & $\eqdef \norm{B^{1/2}MB^{1/2}}_2$\\[0.3cm] \end{tabular} \noindent {\bf \large Special Matrices}\\[0.3cm] \begin{tabular}{lcl} $A$ & $\eqdef$ & The $m \times n$ real system matrix (In Chapter~\ref{ch:inverse} we use $m=n$)\\ $S$ & $\eqdef$ & A random $m\times q$ matrix\\ $B$ & $\eqdef$ & A symmetric positive definite $n \times n$ matrix\\ $Z$ & $\eqdef$ & $A^\top S(S^\top A B^{-1}A^\top S)^{\dagger}S^\top A$ \\ $H$ & $\eqdef$ & $\mathbf{E}_{S\sim {\cal D}} \left[ S\left(S^\top AB^{-1}A^\top S\right)^{\dagger}S^\top\right]$ \end{tabular} \clearpage \chapter*{Acknowledgments \markboth{Acknowledgments}{}} \vspace{-.1in} \small \noindent I am immensely grateful to my supervisor Dr. Peter Richt\'arik. Peter has been a mentor to me on all fronts of being a researcher. From the very process of developing novel research directions, to clear and elegant mathematical writing, delivering the best presentations, and the many facets of applying for a job academia. Peter is my role model for being a researcher and supervisor. Furthermore, his attempts at besting me at table tennis were also admirable. I would like to thank Prof. Jacek Gondzio for numerous research discussions, support and collaboration. Many thanks to my second supervisor Dr. Andreas Grothey for accompanying the progress of the PhD through the years and career discussions. I am grateful to my examination committee, Prof. Nicholas J. Higham and Prof. Ben Leimkuhler for their many detailed pointers, suggestions for improvement and for their time and insightful questions. I am indebted to the school of Mathematics of the University of Edinburgh, for not only providing me with a wonderful work environment, funding yearly trips to the Firbush sports center, but, most of all, for directly funding my PhD. I am most grateful to the Dr. Laura Wisewell Travel scholarship for funding my conference travels expenses in 2013 and 2015. The chance to participate in international conferences, with all the experts in my area in one place, was invaluable. One of the many highlights of my PhD years was a chance encounter Prof. Donald Goldfarb at the Optimization and Big Data 2015 Workshop in Edinburgh. There we discovered that we both have impeccable taste in research projects, and had independently arrived at the same extension to Don's hailed BFGS method. Prof. Donald Goldfarb has since been an inspiration to me, a great enthusiastic collaborator, and warm person in general. Thanks also to Don for having me over at Columbia University, I look forward to our continued work together. My brother, Dr. Artur Gower, has been a constant support throughout my PhD studies and throughout my life (he even ventured into the world just before me to check that the coast was clear). Being a year ahead of me on a PhD program in Ireland, Art has given me the heads-up on how to write good science, applying for funding, and finally how to land a job in science. Art went so far as to read draft's of my papers, proposals and co-authored a paper with me. For his ricochet advice, originally directed toward my brother Artur but then finding its way to me, I would like to thank Prof. Michel Destrade. Thanks to my many friends in Edinburgh, without whom life in Edinburgh would have been as grey as the weather and stone that envelopes it. In particular, to my cohort (in order of their appearance) Pablo Gonzalez Brevis, Kimon Fountoulakis, Tim Schultz, Hanyi Chen, Jakub Kone\v{c}n\'y, Dominik Csiba and Nicolas Loizou I extend a heartfelt thanks for many discussions and their friendship. I would specially like to thank my current and past flatmates Tarek Alabbas and Clara Vergez. Thanks Tarek for your friendship through all the years and being my fellow breakfast musketeer. The only reason I made it to work at any reasonable time was down to Clara banging on my bedroom door early in the morning for breakfast. Clara you have been an amazing flatmate, friend, and salsa co-star; thanks bro-ster. Furthermore, I want to thank my mother Elza Maria, who made my existence possible, and for her unconditional and immeasurable love and support. Finally, many thanks to Jess (aka leao fofinha), for her loving support, companionship, being my sous-chef, deputy nutritionist and just being awesome in general. \vfill \begin{figure}[!h] \centering \includegraphics[width=2cm, natwidth=610,natheight=642]{nerv-old-w.jpg} \\ {\small It takes some Nerv to do a PhD} \end{figure} \clearpage \chapter[Appendix 1]{All the material that didn't make it into the thesis} \chaptermark{Appendix A} \label{ch:AppendixA} \textbf{ABSTRACT}\\ \small{Chapter abstract} \vfill \small{\textbf{Author Contributions:} During long nights postgraduate A compiled this appendix} \newpage \section[Stuff related to Chapter 1]{Sadly there is a word limit of 70,000 in the main thesis :-()} \subsection{Stuff} \label{A1:Stuff} \section*{#1}% \markboth{#1}{#1}} \DeclareFieldFormat{sentencecase}{\MakeSentenceCase{#1}} \renewbibmacro*{title}{% \ifthenelse{\iffieldundef{title}\AND\iffieldundef{subtitle}} {} {\ifthenelse{\ifentrytype{article}\OR\ifentrytype{inbook}% \OR\ifentrytype{incollection}\OR\ifentrytype{inproceedings}% \OR\ifentrytype{inreference}} {\printtext[title]{% \printfield[sentencecase]{title}% \setunit{\subtitlepunct}% \printfield[sentencecase]{subtitle}}}% {\printtext[title]{% \printfield[titlecase]{title}% \setunit{\subtitlepunct}% \printfield[titlecase]{subtitle}}}% \newunit}% \printfield{titleaddon}} \renewcommand*{\bibfont}{\small} \renewbibmacro*{volume+number+eid}{% \printfield{volume}% \setunit*{\addnbspace \printfield{number}% \setunit{\addcomma\space}% \printfield{eid}} \DeclareFieldFormat[article]{number}{\mkbibparens{#1}}

https://arxiv.org/abs/1912.05162

Explicit Hilbert's Irreducibility Theorem in Function Fields

We prove a quantitative version of Hilbert's irreducibility theorem for function fields: If $f(T_1,\ldots, T_n,X)$ is an irreducible polynomial over the field of rational functions over a finite field $\mathbb{F}_q$ of characteristic $p$, then the proportion of $n$-tuples $(t_1,\ldots, t_n)$ of monic polynomials of degree $d$ for which $f(t_1,\ldots, t_n,X)$ is reducible out of all $n$-tuples of degree $d$ monic polynomials is $O(dq^{-d/2})$.

\section{Introduction} Hilbert's irreducibility theorem says that if $f(T_1,\ldots, T_n,X)\in \mathbb{Q}[T,X]$ is a multivariate irreducible polynomial of positive degree in $X$, then the set \[ H_{f} = \{ (t_1,\ldots, t_n)\in \mathbb{Q}^n : f(t_1,\ldots, t_n,X) \in \mathbb{Q}[X] \textnormal{ is irreducible}\}\subseteq \mathbb{Q}^n \] is non-empty. We call such a set a (basic) \emph{Hilbert set}. The original motivation of Hilbert was in the inverse Galois problem, see for example \cite{Serre}. Since then the theorem found numerous applications to different problems in algebra and number theory; e.g., for constructing elliptic curves of high rank, see \cite{Serre2}. There are several types of qualitative results that indicate that any Hilbert set $H_f$ is big (see \cite[\S 13]{FJ} for reference): $H_f$ is infinite; $H_f$ is Zariski dense; $H_f$ is $S$-adic dense, there exist many finite sets of primes $S$ for which $H_f$ is $S$-adic open. A more quantitative variant is by counting: for $N\geq 1$, set \[ H_f(N) = \#( H_f \cap \{ (t_1,\ldots, t_n)\in \mathbb{Z}^n : N\leq t_i< 2 N ,\ i=1,\ldots, n\}) \] to be the number integral tuples in $H$ with entries in $[N,2N)$. S.~D.~Cohen \cite{Cohen} proved that \begin{equation}\label{eq:SDCohen} \frac{H_f(N)}{N^n} = 1+O_f\Big(\frac{\log N}{\sqrt{N}}\Big), \qquad N\to \infty. \end{equation} Here the notation $O_f$ means that the implied constant may depend on $f$, but not on $N$. The polynomial $f(T,X)=X^2-(T_1+T_2+\cdots +T_n)$ shows that the error term is optimal (up to the logarithmic factor). The qualitative results mentioned above were proved also in the global function field setting. However to the best of our knowledge, \eqref{eq:SDCohen} has not been proved in function fields. An additional challenge arising in this setting is having to deal with inseparable polynomials, a phenomenon that does not occur in number fields. The goal of this work is to prove a function field analogue. We denote by $\mathbb{F}_q$ the finite field of $q$ elements. We let $\mathbb{F}_q(u)$ be the field of rational functions in $u$ with coefficients in $\mathbb{F}_q$. The set $M_{d}\subseteq \mathbb{F}_q(u)$ of monic polynomials of degree $d$ takes the role of the interval $[N,2N)$ with $q^d=\#M_d$ being the analog of $N = \#[N,2N)\cap \mathbb{Z}$. \begin{theorem}\label{thm:main} Let $K = \mathbb{F}_q(u)$ be of characteristic $p$, let $f(T_1,\ldots, T_n,X) \in K[T_1,\ldots, T_n,X]$ be an irreducible polynomial of positive degree in $X$. Let $H_f(d)$ be the number of tuples $(t_1(u),\ldots, t_n(u))\in M_d^n$ with $f(t_1,\ldots, t_n,X) \in K[X]$ irreducible. Then, \begin{equation}\label{eq:main} \frac{H_f(d)}{q^{dn}} = 1+O_{f,q}(dq^{-d/2}), \end{equation} where the implicit constant may depend on $f,q$. \end{theorem} To compare with \eqref{eq:SDCohen} we note that $q^d$ corresponds to $N$ and $d$ to $\log N$, so the error term in \eqref{eq:main} is the perfect analog of \eqref{eq:SDCohen}. \begin{remark} In fact we will show that if one fixes $\deg_{\mathbf{T},X}f$ and a constant $\lambda$ then a similar bound holds uniformly whenever $\deg_uf\le\lambda d$. See Theorem \ref{thm:nonsep} for the precise statement.\end{remark} When $f$ is separable in the variable $X$, we will prove \eqref{eq:main} using the function-field large sieve inequality due to Hsu \cite{Hsu96}, a proof that goes in the same lines of the proof of \eqref{eq:SDCohen}. When $f$ is inseparable, those ideas seem to fail. Instead we use an approach of Uchida \cite{Uchida}. Theorem~\ref{thm:main} has many consequences. The first is to Galois theory, which states that if $L/K(T_1,\ldots, T_n)$ is a Galois extension, then for most $(t_1,\ldots, t_n)\in M_d^n$, the Galois group remains the same under the specialization $T_i\mapsto t_i$. See Corollary~\ref{cor:sep} for a precise statement. Theorem 1.1 gives a new proof to the following assertions, that also can be derived from \cite[Thm 13.3.5]{FJ}. \begin{corollary} Let $f(T_1,\ldots, T_n,X)\in K[T_1,\ldots, T_n,X]$ be irreducible of positive degree in $X$ and $H_f$ the set of $(t_1,\ldots, t_n) \in \mathbb{F}_q[u]^n$ for which $f(t_1,\ldots,t_n ,X)\in \mathbb{F}_q(u)[X]$ is irreducible. \begin{enumerate} \item $H_f$ contains a Zariski dense set of prime $n$-tuples; i.e., for every nonzero $g(T_1,\ldots, T_n)$, there exists $(t_1,\ldots,t_n) \in H_f$ such that $g(t_1,\ldots,t_n )\neq 0$ and $t_i\in \mathbb{F}_q[u]$ is irreducible for all $i$. \item If $S$ is a finite set of primes of $\mathbb{F}_q(u)$, then $H_f$ is $S$-adically dense. \end{enumerate} \end{corollary} The implication of the corollary from Theorem~\ref{thm:main} is immediate: the primes have density bigger than $q^{-\epsilon}$, fo any $\epsilon$ and any $S$-adic open set has positive density; hence those sets cannot be contained in the complement of $H_f$ by Theorem~\ref{thm:main}. \subsection*{Acknowledgments} The authors thank Dan Haran for some illuminating discussions of $p$-th powers in characteristic $p$ and Arno Fehm for helpful remarks that improved the presentation of the paper. The first named author was partially supported by grant no.\ 702/19 of the Israel Science Foundation and the second by grant no.\ 2507/19 of the Israel Science Foundation. \section{Set-up and preliminaries} Let $\mathbb{F}_q$ be a finite field with $q$ elements and of characteristic $p$, let $K=\mathbb{F}_q(u)$ be the field of rational functions in $u$. We denote by $M_d$ the subset of $K$ of monic polynomials of degree $d$. We denote tuples by bolded letters, e.g.\ we write $\mathbf{t}$ for $(t_1,\ldots, t_n)$ and $\mathbf{T}$ for $(T_1,\ldots,T_n)$. To a polynomial $f\in K[\mathbf{T},X]$ in $n+1$ variables over the field $K$ we attach the following data. Let $\deg_\mathbf{T} f$ (resp. $\deg_{\mathbf{T},X}f$) denote the total degree of $f$ in the variables $T_1,\ldots,T_n$ (resp. $T_1,\ldots,T_n,X$) and denote $\deg_u(f)=\max\left(\deg_u\tilde{f},\deg_ua\right)$, where $f=\tilde{f}/a,\tilde{f}\in\mathbb{F}_q[u,\mathbf{T},X],a\in\mathbb{F}_q[u]$ is a reduced fraction. It is convenient to work with polynomials in $\mathbb{F}_q[u,\mathbf{T},X]$ that are monic in $X$. The following lemmas allow us to reduce to this case. \begin{lemma}\label{lem:monic} Let $f\in K[\mathbf{T},X]$ be a polynomial, let $a(u)\in \mathbb{F}_q[u]$ be the common denominator of the coefficients of $f$, $\tilde{f}=af\in\mathbb{F}_q[u,\mathbf{T},X]$, $b\in \mathbb{F}_q[u,\mathbf{T}]$ the leading coefficient of $\tilde{f}$ (in the variable $X$) and $r=\deg_X f$. Then the polynomial \[ g (\mathbf{T},X) = b^{\deg_Xf-1} \tilde{f}\Big(\mathbf{T}, \frac{X}{b}\Big) \in \mathbb{F}_q[u,\mathbf{T},X] \] satisfies: \begin{enumerate} \item $g$ is monic in $X$. \item $\deg_X g=\deg_X f$, $\deg_\mathbf{T} g\le\deg_\mathbf{T} f\cdot\deg_Xf$ and $\deg_ug\le\deg_u f\cdot\deg_X f$ \item For all $\mathbf{t} \in M_d^n$ with $b(\mathbf{t})\neq 0$ each of the polynomials $g(\mathbf{t},X)$ and $f(\mathbf{t},X)$ is irreducible iff the other is. \end{enumerate} \end{lemma} \begin{proof} Straightforward from the definition of $g$.\end{proof} The following lemma will be useful in several instances. \begin{lemma}\label{lem:bound} Let $h(\mathbf{T}) \in K[T_1,\ldots, T_n]$ be a nonzero polynomial. Then the number of tuples $\mathbf{t}\in M_{d}^n$ such that $h(\mathbf{t})=0$ is at most $\sum_j \deg_{T_j} h \cdot q^{d(n-1)}$. \end{lemma} \begin{proof} We apply induction. If $n=1$, then $h(T)$ has at most $\deg_T h$ solutions in $K$, hence in $M_{d}$. Assume $n\geq 2$ and write $h(\mathbf{T}) = \sum_{i} h_i(T_1,\ldots, T_{n-1})T_n^i$, where not all of the $h_i$-s are zero. Then, by induction, \[ \#\{ \mathbf{t}\in M_{d}^n:h(\mathbf{t})=0\} \leq \sum_{\substack{\mathbf{t}'\in M_{d}^{n-1} \\ \forall i:h_i(\mathbf{t}') =0}} q^d + \sum_{\substack{\mathbf{t}'\in M_{d}^{n-1} \\ \exists i:h_i(\mathbf{t}') \neq 0}} \deg_{T_n} h \le \sum_j\deg_{T_j}h\cdot q^{n-1}, \] as needed. \end{proof} The last two lemmas immediately imply the following result which allows us to reduce to monic polynomials with coefficients in $\mathbb{F}_q[u,\mathbf{T}]$: \begin{corollary}\label{cor:reduction} Let $f\in K[\mathbf{T},X]$ be a polynomial and $g(\mathbf{T},X)\in \mathbb{F}_q[u,\mathbf{T},X]$ as in Lemma~\ref{lem:bound}. Then for all $\mathbf{t}\in M_{d}^n$ but at most $\sum\deg_{T_j}h\cdot q^{d(n-1)}$, each of $f(\mathbf{t},X)$ and $g(\mathbf{t},X)$ is irreducible iff the other is. \end{corollary} \section{Separable Polynomials} In this section we assume that $f\in \mathbb{F}_q[u,\mathbf{T},X]$ is irreducible and that it is separable and monic in $X$. Throughout this and the next section we fix a bound on $\deg_{\mathbf{T},X}f$ and all asymptotic notation will have implied constants that may depend on this bound. We do not fix the degree in $u$ as we are looking for a result which is uniform in $f,q$ as long as $\deg_uf$ grows at most linearly with $d$. Let $L$ be the splitting field of $f$ over $K(\mathbf{T})$ and let $\mathbb{F}_{q^r} = \overline{\mathbb{F}}_q\cap L$ be the algebraic closure of $\mathbb{F}_q$ in $L$. We have the following diagram of fields: \[ \xymatrix{ & L \ar@{-}[r]&L\overline{\mathbb{F}}_q \\ K(\mathbf{T})\ar@{-}[r] & K \mathbb{F}_{q^r}(\mathbf{T})\ar@{-}[r]\ar@{-}[u]& K\overline{\mathbb{F}}_q (\mathbf{T})\ar@{-}[u] \\ \mathbb{F}_q\ar@{-}[r] \ar@{-}[u]& \mathbb{F}_{q^r}\ar@{-}[r]\ar@{-}[u] & \overline{\mathbb{F}}_q\ar@{-}[u] } \] We denote $G={\rm Gal}(L/K(\mathbf{T}))$ and $G_{{\rm geom}}={\rm Gal}(L/K\mathbb{F}_{q^r}(\mathbf{T})) \cong {\rm Gal}(L\overline{\mathbb{F}}_q/K\overline{\mathbb{F}}_q(\mathbf{T}))$. We have the fundamental exact sequence \begin{equation}\label{eq:fund_seq} \xymatrix@1{1\ar[r]&G_{{\rm geom}}\ar[r]&G\ar[r]& \mathbb{Z}/ r\mathbb{Z}\ar[r] &1} \end{equation} where $\mathbb{Z}/r\mathbb{Z} \cong{\rm Gal}(\mathbb{F}_{q^r}/\mathbb{F}_q)$ and under this identification the map $G\to \mathbb{Z}/r\mathbb{Z}$ is the restriction-of-automorphism map. \begin{remark}\label{remark:r}By a result of Guralnick \cite{Guralnick} any cyclic quotient of a transitive permutation group of degree $N$ has order $\le N$ and therefore since $\mathbb{Z}/r\mathbb{Z}$ is a quotient of $G$ which acts transitively on the roots of $f$ we must have $r\le\deg_Xf$.\end{remark} If $\mathbf{t} \in K^n$ is an $n$-tuple for which $f(\mathbf{t},X)$ is separable in $X$, then we denote by $L_{\mathbf{t}}$ the splitting field of $f(\mathbf{t},X)$ over $K$ and we note that $\mathbb{F}_{q^r}\subseteq L_{\mathbf{t}}$. We further denote $G_{\mathbf{t}} = {\rm Gal}(L_{\mathbf{t}}/K)$ and $G_{{\rm geom},\mathbf{t}} = {\rm Gal}(L_{\mathbf{t}}/K\mathbb{F}_{q^r})$. We have that $G_{\mathbf{t}}\leq G$ and $G_{{\rm geom},\mathbf{t}}\leq G_{{\rm geom}}$, in a canonical way up to conjugation, and $G/G_{{\rm geom}} \cong G_{\mathbf{t}}/G_{{\rm geom},\mathbf{t}} \cong \mathbb{Z}/r \mathbb{Z}$. Hence the following observation is immediate: \begin{lemma}\label{lem:GGg} $G_{\mathbf{t}} \cong G$ if and only if $G_{{\rm geom},\mathbf{t}}\cong G_{{\rm geom}}$. \end{lemma} Let $P(u)\in \mathbb{F}_{q}[u]$ be a prime polynomial; that is to say, irreducible and monic. Write $|P|=q^{\deg P}$ for the norm of $P$. We may reduce the coefficients of $f$ modulo $P$, and we denote the resulting polynomial by $f_P\in \mathbb{F}_{|P|}[\mathbf{T},X]$. If in addition $f_P$ is separable in $X$, then we define $L_P$, $G_{P}$ and $G_{{\rm geom},P}$ in a similar fashion to the definition of $L$, $G$ and $G_{{\rm geom}}$ (with $\mathbb{F}_{|P|}$ taking the role of $K$). Since $f$ is monic in $X$, we have that $\deg_X f=\deg_X f_P$, so $G_{P}$ and $G_{{\rm geom}, P}$ embed in $G$ and $G_{{\rm geom}}$, respectively, and this embedding is canonical, up to conjugation in $G$. Consider the set of primes of good reduction in the following sense: \[ \mathcal{P}_f = \{ P\in \mathbb{F}_q[u] : P \textnormal{ prime},\ \ r\mid \deg P,\ f_P \textnormal{ separable in $X$}, \ G_{{\rm geom}} \cong G_{{\rm geom}, P}\}. \] If $P\in \mathcal{P}_f$, then $\mathbb{F}_{q^r}\subseteq \mathbb{F}_{|P|}$, so $L_P$ is regular over $\mathbb{F}_{|P|}$, and we have the diagram of fields: \[ \xymatrix{ L_P \ar@{-}[r]&L_P\overline{\mathbb{F}}_q \\ \mathbb{F}_{|P|}(\mathbf{T})\ar@{-}[r]\ar@{-}[u]^{G_{{\rm geom}}}& \overline{\mathbb{F}}_q(\mathbf{T}) \ar@{-}[u]^{G_{{\rm geom}}} \\ \mathbb{F}_{|P|}\ar@{-}[r] \ar@{-}[u]& \overline{\mathbb{F}}_q\ar@{-}[u] } \] \begin{lemma}\label{lem:primes} Let $\lambda>0$ be a constant. The number of (monic) irreducible polynomials $P\in\mathbb{F}_q[u]$ with $\deg P\ge \frac 1\lambda\deg_uf$, $r|\deg P$ and such that $P\not\in\mathcal{P}_f$ is $O_{\lambda,\deg_{\mathbf{T},X}f}(1)$. \end{lemma} \begin{proof} We fix $\lambda>0$. Since $f$ is separable in $X$, there is a nonozero coefficient $g\in \mathbb{F}_q[u,\mathbf{T}]$ of $X^i$ for some $i\not\equiv 0\mod p$. If $r|\deg P$ and $f_P$ is not separable, then $P$ must divide all the coefficients of $g$. Since we assume $\lambda \deg P\ge \deg_u f \geq \deg_ug$, there are at most $O_\lambda(1)$ such $P$ (since the same coefficient of $g$ is divisible by all of them). It remains to show that if $r|\deg P\ge \deg_uf/\lambda$ and $f_P$ is separable we have $G_{{\rm geom}} \cong G_{{\rm geom}, P}$ with $O_{\lambda,\deg_{X,\mathbf{T}} f}(1)$ many exceptions. For this we choose a monic irreducible polynomial $h\in K\mathbb{F}_{q^r}[\mathbf{T},X]$ whose root generates $L$ ($h$ is a resolvent of $f$). We can choose $h$ with $\deg_u h=O(\deg_uf)$, $\deg_X h = |G_{{\rm geom}}|$ and $h$ is monic in $X$. The resolvent $h$ is geometrically irreducible. If $h_P$ is geometrically irreducible then $G_{{\rm geom}, P } \cong G_{{\rm geom}}$. Let $c_1,\ldots,c_m$ be the coefficients of $h$. By a theorem of Noether (see the beautiful exposition by Geyer \cite[Theorem 5.3.1]{Geyer}) there exist $O_{\deg_{\mathbf{T},X} f}(1)$ polynomials $F_j\in\mathbb{Z}[v_1,\ldots,v_m]$ of degree $O(1)$ such that $F_j(c_1,\ldots,c_m)\neq 0$ and $h_P$ is absolutely irreducible unless $F_j(c_1,\ldots,c_m)\equiv 0\pmod P$ for some $j$. Since $\deg_uF_j(c_1,\ldots,c_m)=O(\deg_uf)$ there are only $O(1)$ such $P$ with $\deg P\ge \deg_uf/\lambda$ if we assume $\deg_uf\le \lambda d$. For any other $P$ we have $G_{{\rm geom}, P } \cong G_{{\rm geom}}$. \end{proof} Take $P\in \mathcal{P}_f$. Every $\mathbf{a} \in \mathbb{F}_{|P|}^n$ with $f_P(\mathbf{a},X)\in \mathbb{F}_{|P|}[X]$ separable gives rise to the Frobenius conjugacy class \[ C_{\mathbf{a}} \subseteq G_{P} = G_{{\rm geom},P} \cong G_{{\rm geom}}. \] Now fix a conjugacy class $C\subseteq G_{{\rm geom}}$ and let \[ \Omega_P =\Omega_P(C) = \{ \mathbf{a} \in \mathbb{F}_{|P|}^n : C_{\mathbf{a}} \neq C\}. \] (The set $\Omega_P$ also contains those $\mathbf{a}$ for which $f_P(\mathbf{a}, X)$ is not separable.) The explicit Chebotarev density theorem for function fields (see Theorem 3 in \cite{Entin}, for a version with the desired uniformity) gives that \[ \frac{\# \Omega_{P}}{|P|^n} = \Big(1- \frac{|C|}{|G_{{\rm geom}}|}\Big) \cdot \big(1+O_{\deg_{\mathbf{T},X} f}\big(|P|^{-1/2}\big)\big). \] In particular, if $G_{{\rm geom}}$ is non-trivial, then $|C|\neq |G_{{\rm geom}}|$, so if we take any $\frac{|C|}{|G|}<c<1$, then for every $P\in \mathcal{P}_f$ of sufficiently large degree, we have \begin{equation}\label{eq:OmegaPbds} \frac{\# \Omega_{P}}{|P|^n}\leq c. \end{equation} \begin{proposition} \label{prop:sep} Fix a constant $\lambda>0$ and assume $\deg_u f\leq \lambda d$. Let $B_C(d)$ be the number of tuples $\mathbf{t} \in M_d^n$ such that $G_{{\rm geom},\mathbf{t}} \cap C\neq \emptyset$. Then we have \[ \frac{B_C(d)}{q^{dn}} = O_{\deg_{\mathbf{T},X} f,\lambda}\left(q^{-d/2+r} d\right). \] with $r\leq \deg_Xf$ as defined above. \end{proposition} \begin{proof} Let $\Omega = \Omega(d,C,f)$ be the set of tuples $\mathbf{t} \in M_d^n$ such that $G_{{\rm geom},\mathbf{t}} \cap C=\emptyset$. Let $\mathbf{t}\in M_{d}^n$ and $P\in \mathcal{P}_f$. Since $r\mid \deg P$, we have that $P$ is totally split in $K\mathbb{F}_{q^r}$. Moreover, $f(\mathbf{t},X)\mod P = f_P(\mathbf{a},X)$, where $\mathbf{a}:= \mathbf{t} \mod P$. If we assume that $f_P(\mathbf{a},X)$ is separable, then we get that the Frobenius class of $P$ in $L_{\mathbf{t}}$ equals the Frobenius class of $\mathbf{a}$ in $L_{P}$, which we earlier denoted by $C_{\mathbf{a}}$. Hence if $\mathbf{t}\in \Omega$, then $C_{\mathbf{a}}\neq C$ and so \[ \mathbf{t}\mod P\in \Omega_P \] for any $P\in \mathcal{P}_f$. We apply the large sieve inequality for function fields \cite[Theorem 3.2]{Hsu96} \footnote{In the notation of \cite[Theorem 3.2]{Hsu96} we take $X=\Omega, N=d-1, f=(t^d,\ldots,t^d),K=d/2$, $\alpha_P=c$ if $P\in\mathcal{P}_f$ and $\alpha_P=1$ otherwise.} to conclude that \[ \# \Omega \leq \frac{q^{dn}}{L}, \] where \[ L \geq 1+\sum_{\substack{P\in \mathcal{P}_f\\ \deg P\leq d/2}} \frac{1-c}{c} \gg \# \{P\in \mathcal{P}_f: \deg P\leq d/2\}. \] Finally, let $s$ to be an integer so that $d/2-r \le rs \leq d/2$ (if it is not positive the assertion of the proposition is trivial). If $d$ is large, by Lemma~\ref{lem:primes}, $\mathcal{P}_f$ contains all but $O(1)$ primes of degree $rs$, so \[ \# \{ P \in \mathcal{P}_f : \deg P\leq d/2\} \geq \pi_{q}(rs)-O(1)\sim \frac{q^{rs}}{rs} \gg\frac{q^{d/2-r}}{d/2}. \] Here $\pi_q(rs)$ denotes the number of primes of degree $rs$ and we use the Prime Polynomial Theorem (see e.g.\ \cite[Theorem~2.2]{Rosen}) which says that $\pi_q(rs) = \frac{q^{rs}}{rs} + O(q^{-rs/2})$. We conclude that \[ \frac{\#\Omega}{q^{dn}} \ll q^{-d/2+r}d. \] Since $B_{C}(d)=\#\Omega$ the proof is done. \end{proof} \begin{corollary}\label{cor:sep} Fix $\lambda>0$ and assume $\deg_uf\le\lambda d$. Let $B(d)$ be the number of $n$-tuples $\mathbf{t} \in M_{d}^n$ such that $G_{\mathbf{t}} \not\cong G$. Then \[ \frac{B(d)}{q^{dn}} = O_{\lambda,\deg_{\mathbf{T},X}f} (q^{-d/2+r}d), \] with $r\leq \deg_X f$ as defined above. \end{corollary} \begin{proof} By Lemma~\ref{lem:GGg}, if $G_{\mathbf{t}}\not \cong G$, then $G_{{\rm geom},\mathbf{t}}\not \cong G_{{\rm geom}}$. If $G_{{\rm geom},\mathbf{t}}$ is isomorphic to a proper subgroup of $G_{{\rm geom}}$, then there exists a conjugacy class $C$ such that $G_{{\rm geom},\mathbf{t}}\cap C \neq \emptyset$ (it is an elementary exercise in group theory to show that every proper subgroup of a finite group is disjoint from some conjugacy class). Applying Proposition~\ref{prop:sep}, the union bound, and the fact that there are $O_{\deg_{X}f}(1)$ conjugacy classes in $G_{{\rm geom}}$, we conclude that \[ B(d) \ll_{\deg_{\mathbf{T},X}f,\lambda} q^{-d/2+r}d, \] as needed. \end{proof} \begin{theorem}\label{thm:uniform} Let $K=\mathbb{F}_{q}(u)$, let $f(T_1,\ldots, T_n,X)\in K[T_1,\ldots,T_n,X]$ be an irreducible polynomial that is separable and of positive degree in $X$. Let $H(d)$ be the number of tuples $\mathbf{t} \in M_{d}^n$ with $f(\mathbf{t},X)$ irreducible, separable, and $\deg_X f(\mathbf{t},X)=\deg_X f(\mathbf{T},X)$. Let $\lambda>0$ be a constant and assume that $\deg_uf\le\lambda d$. Then \[ \frac{H(d)}{q^{dn}} = 1+O_{\deg_{T_1,\ldots,T_n,X}f,\lambda} \left(dq^{-d/2+r}\right), \] with $r\leq \deg_X f$ as defined above. \end{theorem} \begin{proof} We fix $\lambda$ and a bound on $\deg_{\mathbf{T},X}f$. First we observe that by Remark \ref{remark:r} we have $r\le\deg_Xf$. By Corollary~\ref{cor:reduction} we may assume w.l.o.g.\ that $f\in \mathbb{F}_q[u,\mathbf{T},X]$ and that $f$ is monic in $X$. Let $\Sigma$ be the set of tuples $\mathbf{t} \in M_{d}^n$ for which $f(\mathbf{t},X)$ is inseparable; i.e., the zero set of all the coefficients of $X^i$ for $i\not\equiv 0\mod p$. By Lemma~\ref{lem:bound}, $\#\Sigma = O(q^{d(n-1)})$. For $\mathbf{t}\not\in \Sigma$, the action of $G_{\mathbf{t}}$ on the roots of $f(\mathbf{t},X)$ coincides with the action the decomposition group $G_{\mathbf{t}}\cong D_{\mathbf{t}}\leq G$ on the roots of $f(\mathbf{T},X)$, hence if $G_{\mathbf{t}}\cong G$, the action is transitive, and $f(\mathbf{t},X)$ is irreducible. Corollary~\ref{cor:sep} then completes the proof. \end{proof} \section{Inseparable polynomials} In this part, we let $f(\mathbf{T},X)\in K [T_1,\ldots, T_n,X]$ be an irreducible polynomial that is inseparable in $X$. Our treatment is inspired from Uchida's work \cite{Uchida} and its presentation in \cite{FJ}. We will use the following criterion of Uchida for irreducibility of inseparable polynomials, see \cite[Lemma 12.4.1]{FJ}. \begin{lemma}\label{lem:irrpthpower} Let $F$ be a field of positive characteristic $p>0$, let $g\in F[X]$ be an irreducible monic polynomial. Assume that at least one of the coefficients of $g$ is not a $p$-th power in $F$. Then $g(X^{p^\alpha})$ is irreducible for all $\alpha>0$. \end{lemma} We shall need a technical lemma. \begin{lemma}\label{lem:pthpower} Let $h(\mathbf{T}) \in \mathbb{F}_q[u,\mathbf{T}]$ be a polynomial which is not a $p$-th power. Let $A(d)$ be the number of $n$-tuples $\mathbf{t} \in M_d$ such that $h(\mathbf{t})$ is a $p$-th power. Fix a constant $\lambda>0$. Then as long as $\deg_u h\le\lambda d$ we have \[ \frac{A(d)}{q^{dn}} \ll_{\deg_\mathbf{T} h,\lambda} q^{-d/2+1}. \] \end{lemma} \begin{proof} We fix $\lambda>0$ and a bound on $\deg_\mathbf{T} h$. Since every element in $\mathbb{F}_q$ is a $p$-th power and since $h$ is not a $p$-th power, $h$ is not a polynomial in $v^p$ for at least one of the variables $v\in \{u,T_1,\ldots, T_n\}$. Equivalently, $\frac{\partial h}{\partial v}\neq 0$. Similarly, for $\mathbf{t}\in M_{d}^n$, $h(\mathbf{t})$ is a $p$-th power if and only if $d (h(\mathbf{t}))/du=0$. If $\frac{\partial h}{\partial T_i}=0$ for all $i$, hence $\frac{\partial h}{\partial u} \neq 0$, then by the chain rule, for all $\mathbf{t} \in M_{d}^{n}$ we have \[ \frac{d (h(\mathbf{t}))}{d u} = \frac{\partial h}{\partial u} (\mathbf{t}). \] By Lemma~\ref{lem:bound}, \[ \frac{A(d)}{q^{dn}} = \frac{\#\{\mathbf{t}\in M_d^n : \frac{\partial h}{\partial u} (\mathbf{t})=0\} }{q^{dn}}\ll q^{-d}, \] and the proof is done. Next assume that $\frac{\partial h}{\partial T_i}\neq 0$ for some $i$. We can reduce this case to the univariate case (i.e.\ $n=1$): To ease notation assume $i=n$. For $t_1,\ldots, t_{n-1}\in M_d^{n-1}$ consider the polynomial $g(T) = h(t_1,\ldots, t_{n-1},T)\in\mathbb{F}_q[u,T]$. Since \[ \frac{\partial g}{\partial T} = \frac{\partial h}{\partial T_n} (t_1,\ldots, t_{n-1},T) \] we get by Lemma~\ref{lem:bound} that ${\partial g}/{\partial T}\neq 0$ for all $(t_1,\ldots, t_{n-1})\in M_d^{n-1}$ but $O(q^{d(n-2)})$ such $(n-1)$-tuples. So $n$-tuples $\mathbf{t}\in M_{d}^n$ for which ${\partial g}/{\partial T}=0$, contribute at most $O(q^{d(n-1)})$ to the total number of $n$-tuples with $h(\mathbf{t})$ a $p$-th power. Therefore, we may assume that ${\partial g}/{\partial T}\neq 0$ and it suffices to prove that under this assumption \begin{equation}\label{eq:count} \#\{ t\in M_d : g(t) \in K^p\} = O(q^{d/2+1}), \end{equation} in order to conclude the proof. Take $r(u) \in \mathbb{F}_q[u]$ irreducible of degree $\lfloor \frac d2 \rfloor$ such that $g,\partial g/\partial T\not\equiv 0\pmod r$. Such an $r$ exists since by our assumptions $\deg_ug,\deg_u(\partial g/\partial T)=O(\deg_uh)=O(d)$ and so for all but $O(1)$ irreducible $r$ of degree $\lfloor \frac d2 \rfloor$ one of the coefficients of $g$ (and of $\partial g/\partial T$) is not divisible by $r$. Let $\mathcal{A} = \mathbb{F}_q[u]/r^2$ and let $\mathcal{S}=\{c^p : c\in \mathcal A\}$ be the set of $p$-th powers modulo $r^2$. We have $\#\mathcal{S} = O\left(q^{\lfloor d/2\rfloor}\right)$: Indeed, $p$-th powers modulo $r^2$ can come either from the non-invertibles $r\mathcal A$ or from the subgroup $\left(\mathcal A^*\right)^p$ of $\mathcal A^*$ and both sets are of size $O\left(q^{\lfloor d/2\rfloor}\right)$. For $t\in M_d$, if $g(t) \in K^p$, then $g(t) \mod r^2\in \mathcal S$. Since $\partial g/\partial T\not\equiv 0\pmod r$ we have \[ \#\left\{t\in M_d:\frac{\partial g}{\partial T}(t)\equiv 0\pmod r\right\}=O\left(q^{d-\deg r}\right)=O\left( q^{d/2+1}\right). \] Hence it suffices to show that \[ \label{noncritical}\#\left\{t\in M_d:g(t)\bmod r^2\in\mathcal S,\frac{\partial g}{\partial T}(t)\not\equiv 0\pmod r\right\}=O\left( q^{d/2+1}\right). \] Fix $s\in \mathcal S$. Since $g\not\equiv 0\pmod r$ there are $O(1)$ solutions to $g(\tau)\equiv s\pmod r$. If $\tau\in\mathbb{F}_q[u]/r$ is a root of $g\bmod r$ and $\frac{\partial g}{\partial T}(\tau)\not\equiv 0\pmod r$ then by Hensel's lemma, there are $O(1)$ solutions to $g(\tau') \equiv s\pmod {r^2}$ with $\tau'\equiv\tau\pmod r$ and each of them lifts to $q^{d-2\lfloor d/2\rfloor}$ solutions of $g(t)\equiv s\pmod{r^2},t\in M_d$. So the total number of solutions to $g(t)\equiv s\pmod{r^2},\frac{\partial g}{\partial T}(t)\not\equiv 0\pmod{r^2},t\in M_d$ is bounded by $\#\mathcal{S}\cdot q=O(q^{d/2+1})$, as required. \end{proof} Now we may deduce the quantitative Hilbert's irreducibility theorem for inseparable polynomials. \begin{theorem}\label{thm:nonsep} Let $f(\mathbf{T},X)\in K [T_1,\ldots, T_n,X]$ be an irreducible polynomial. Let $H_f(d)$ be the number of tuples $\mathbf{t}\in M_{d}^n$ for which $f(\mathbf{t},X)$ is irreducible. Fix a constant $\lambda>0$ and assume $\deg_u f\le\lambda d$. Then we have \[ \frac{H_f(d)}{q^{dn}} = 1 + O_{\deg_{\mathbf{T},X}f,\lambda}(d q^{-d/2+r}), \] where $r\leq deg_X(f)$ is the degree of the algebraic closure $\mathbb{F}_q^r$ of $\mathbb{F}_q$ in the splitting field of $f$ over $K(\mathbf{T})$. \end{theorem} \begin{proof} If $f$ is separable in $X$ this is Theorem \ref{thm:uniform}. Hence we assume that $f$ is inseparable in $X$. By Corollary~\ref{cor:reduction}, we may assume w.l.o.g.\ that $f\in \mathbb{F}_q[u,\mathbf{T},X]$ and that $f$ is monic in $X$. Since $f$ is inseparable, we may write $f(\mathbf{T},X) = g(\mathbf{T},X^{p^{\alpha}})$ for some $g$ that is irreducible, separable in $X$ and of degree $\deg_X g\leq \deg_Xf$ and for some $\alpha>0$. Let $\mathbb{F}_{q^r}$ be the algebraic closure of $\mathbb{F}_q$ inside a splitting field of $g$ over $K(\mathbf{T})$. Write $g(\mathbf{T},X) = \sum_i g_i(\mathbf{T}) X^i$, with $g_i\in \mathbb{F}_q[u,\mathbf{T}]$. Had all the $g_i$ been $p$-th powers, we would have that $f$ is reducible. Hence there exists $g_i$ which is not a $p$-th power. If $g_i(\mathbf{t})$ is not a $p$-th power and $g(\mathbf{t},X)$ is irreducible, then by Lemma~\ref{lem:irrpthpower}, $f(\mathbf{t},X)=g(\mathbf{t},X^{p^{\alpha}})$ is irreducible. Applying Theorem~\ref{thm:uniform} and Lemma~\ref{lem:pthpower}, gives that this happens for all but $O_{\deg_{\mathbf{T},X}f,\lambda}(dq^{d(n-1/2)+r})$ of the $\mathbf{t}$-s (as long as $\deg_uf\le\lambda d$), hence the proof is complete. \end{proof}

https://arxiv.org/abs/0904.3740

On adding a list of numbers (and other one-dependent determinantal processes)

Adding a column of numbers produces "carries" along the way. We show that random digits produce a pattern of carries with a neat probabilistic description: the carries form a one-dependent determinantal point process. This makes it easy to answer natural questions: How many carries are typical? Where are they located? We show that many further examples, from combinatorics, algebra and group theory, have essentially the same neat formulae, and that any one-dependent point process on the integers is determinantal. The examples give a gentle introduction to the emerging fields of one-dependent and determinantal point processes.

\section{Introduction}\label{sec1} Consider the task of adding a single column of digits: \begin{equation*}\begin{array}{rrr} 7&\cdot&7\\ 9&\cdot&6\\ 4&&0\\ 8&\cdot&8\\ 3&&1\\ 6&&7\\ 1&\cdot&8\\ 6&\cdot&4\\ 8&&2\\ \cline{1-1} 52&& \end{array}\end{equation* We have put a dot to the right of a digit whenever the succeeding addition leads to a carry. The remainder (mod 10) is written to the right of the dot. Thus $7+9=16$ results in a dot by the 7 and a remainder of 6. At the end, the total is found by adding the number of dots (here, five) and appending the final remainder (here, 2) for a total of 52. The dots are a standard bookkeeping device used (for example) in the Trachtenberg system of speed addition \citep{cm}. How many carries are typical and how are they distributed? Common sense suggests that about half the places have carries and that if a carry occurs, it is less likely that the next addition gives a carry. Investigating these questions when digits are chosen uniformly at random leads to interesting mathematics. As additional motivation, look at the remainder column in the example above. Observe that there is a carry on the left (and so a dot) if and only if there is a descent on the right (a sequence $x_1,x_2,\dots$ has a descent at $i$ if $x_i>x_{i+1}$). Further, if the original column of digits is independent and uniformly distributed, so is the remainder column. Thus, the results about carries give the distribution of the descent pattern in a random sequence of digits. \ref{sec2} derives the distribution theory of carries by elementary arguments. There are nice formulae for the chance of any pattern of carries, and the process is one-dependent so a variety of probabilistic limit theorems are available. \ref{sec3} introduces determinantal point processes, shows that the carries process is determinantal, and illustrates how standard tools for determinantal point processes apply to carries. \ref{newsec4} reviews the literature on stationary one-dependent processes, shows that all of these are determinantal and further that essentially all the neat formulae for carries and descents have versions for any stationary one-dependent process. Connections with symmetric function theory are developed. A large class of examples arising from the work of Polishchuk and Positselski \cite{pp} on Koszul algebras is shown to yield many natural examples (including the original carries process). \ref{sect5} gives combinatorial examples: descents in permutations from both uniform and non-uniform distributions, the connectivity set of a permutation, and binomial posets. \ref{sect6} generalizes from adding numbers to multiplying random elements in a finite group. For central extensions (e.g., the quaternions), there is again a carries process that is stationary, one-dependent, and determinantal. Thus explicit formulae and limit theorems are available. \ref{sect7} contains proofs of some of our more technical results. It shows that any one-dependent (possibly non-stationary) point process on the integers is determinantal. It also constructs a family of one-dependent determinantal processes (generalizing many examples in earlier sections), and computes its correlation kernel. \section{Probability theory for carries and descents}\label{sec2} Throughout this section we work base $b$ and so with the alphabet $\mathcal{B}=\{0,1,\dots,b-1\}$. Let $B_1,B_2,\dots,B_n$ be a sequence of randomly chosen elements of $\mathcal{B}$ (independent and identically distributed). There is a \textit{descent at $i$} if $B_i>B_{i+1},\ 1\leq i\leq n-1$. Let $X_i$ be $1$ or $0$ as there is a descent at $i$ and $D=\{i:X_i=1\}$ be the descent set of $B_1,B_2,\dots,B_n$. The following probabilistic facts are elementary. They are stated for carries but verification is easier using descents. \begin{fact} (Single carries) \quad For any $i\in[n-1]$, \begin{equation*} P(X_i=1)=\dfrac12-\dfrac1{2b}=\frac{\binom{b}{2}}{b^2}. \end{equation*} Thus, when $b=10$, the chance of a carry is $.45$. When $b=2$, the chance of a carry is $.25$. For any base, Var$(X_i)=\tfrac14-\tfrac1{4b^2}$. \end{fact} \begin{fact} (Runs of carries) \quad For any $i$ and $j$ with $1\leq i<i+j\leq n$, \begin{equation*} P\left(X_i=X_{i+1}=\dots=X_{i+j-1}=1\right)=\binom{b}{j+1}\bigg/b^{j+1}. \end{equation*} Thus, a run of $b$ or more carries is impossible. Further, \begin{equation*} \text{Cov}\left(X_i,X_{i+1}\right)=E\left(X_iX_{i+1}\right)-E(X_i)E(X_{i+1})=-\dfrac1{12}\left(1-\dfrac1{b^2}\right). \end{equation*} \end{fact} \begin{fact} (Stationary one-dependence) \quad The distribution of $\{X_i\}_{i\in[n-1]}$ is stationary: for $J\subseteq[n-1],\ i\in[n-1]$ with $J+i\subseteq[n-1]$, the distribution of $\{X_j\}_{j\in J}$ is the same as the distribution of $\{X_j\}_{j\in J+i}$. Further, the distribution of $\{X_i\}_{i\in[n-1]}$ is one-dependent: if $J\subseteq[n-1]$ has $j_1,j_2\in J\Rightarrow|j_i-j_2|>1$, then $\{X_j\}_{j\in J}$ are jointly independent binary random variables with $P(X_j=1)=\tfrac12-\tfrac1{2b}$. The literature on $m$-dependent random variables is extensive. In particular, a classical central limit theorem \cite{hr} shows the following: \begin{thm} For $n \geq 2$, the total number of carries $T_{n-1}=X_1+\dots+X_{n-1}$ has mean $(n-1)(\tfrac12-\tfrac1{2b})$, variance $\frac{n+1}{12}(1-\tfrac1{b^2})$ and, normalized by its mean and variance, $T_{n-1}$ has a standard normal limiting distribution for $n$ large. \end{thm} {\it Remarks.} \begin{enumerate} \item An $O(n^{-1/2})$ error bound in this central limit theorem can be proved using the dependency graph approach to normal approximation by Stein's method \cite{cs}, \cite{der}. \item Stationary pairwise independent processes can fail to obey the central limit theorem \cite{brad}, \cite{jan}. \end{enumerate} \end{fact} \begin{fact} ($k$-point correlations) \quad For $A\subseteq[n-1]$, let \[\rho(A)=P\{X_i=1\text{ for }i\in A\}.\] For $|A|=k$, the $\rho(A)$ are called $k$-point correlations for the point process $X_1,\dots,X_{n-1}$. They are basic descriptive units for general point processes \cite{dv}. For the carries and descent process, they are simple to describe. Break $A\subseteq[n-1]$ into disjoint, non-empty blocks of consecutive integers $A=A_1\cup A_2\cup\dots\cup A_k$. Thus $A=\{2,3,5,6,7,11\}=\{2,3\}\cup\{5,6,7\}\cup\{11\}$. From Facts 2 and 3 above, \begin{equation*} \rho(A)=\prod_{i=1}^k \left[ \binom{b}{a_i+1}\bigg/b^{a_i+1} \right] \qquad\text{if }A=\bigcup_{i=1}^kA_i\text{ with }|A_i|=a_i. \end{equation*} \end{fact} \begin{fact} (Determinant formula) \quad Let $\epsilon_1,\epsilon_2,\dots,\epsilon_{n-1}$ be a fixed sequence in $\{0,1\}$. Then \begin{equation*} P\left\{X_1=\epsilon_1,\dots,X_{n-1}=\epsilon_{n-1}\right\}=\dfrac1{b^n} \cdot \text{det}\binom{s_{j+1}-s_i+b-1}{b-1}. \end{equation*} Here, if there are exactly $k$ $1$'s in the $\epsilon$-sequence at positions $s_1<s_2<\dots<s_k$, the determinant is of a $(k+1)\times(k+1)$ matrix with $(i,j)$ entry $\binom{s_{j+1}-s_i+b-1}{b-1}$ for $0\leq i,j\leq k$ with $s_0=0,\ s_{k+1}=n$. \end{fact} \begin{example} If $n=8,\ \epsilon_1=1,\ \epsilon_2=\epsilon_3=\epsilon_4=0,\ \epsilon_5=1,\ \epsilon_6=\epsilon_7=0$, the matrix is \begin{equation*}\begin{pmatrix} \binom{1+b-1}{b-1}&\binom{5+b-1}{b-1}&\binom{8+b-1}{b-1}\\ 1&\binom{4+b-1}{b-1}&\binom{7+b-1}{b-1}\\ 0&1&\binom{3+b-1}{b-1} \end{pmatrix}. \end{equation*} When $b=2$, this is $\left(\begin{smallmatrix}2&6&9\\1&5&8\\0&1&4\end{smallmatrix}\right)$ with determinant $9$. Thus the chance of a carry at exactly positions 1 and 5 when adding eight binary numbers is $9/2^8\doteq.03516$. When $b=10$, this chance becomes $1,042,470/10^8\doteq.0104$. This is not so much smaller than in the binary case. \label{ex21} \end{example} \begin{proof} (Of fact 5) We follow Stanley \cite[p.~69]{sta2}. Let $\alpha_n(S)$ be the number of sequences with descent set contained in $S$ and $\beta_n(S)$ the number of sequences with descent set equal to $S$. Then \begin{equation*} \alpha_n(S)=\sum_{T\subseteq S}\beta_n(T),\qquad \beta_n(S)=\sum_{T\subseteq S}(-1)^{|S-T|}\alpha_n(T). \end{equation*} There is a simple formula for $\alpha_n(S)$: the number of weakly increasing sequences with entries in $\{0,1,\dots,b-1\}$ of length $l$ is $\binom{l+b-1}{b-1}$ using the usual stars and bars argument (place $b-1$ bars into $l+b-1$ places, put symbol $0$ to the left of the first bar, symbol $1$ between the first and second bar, and so on, with symbol $b-1$ to the right of the last bar). Then \begin{equation*} \alpha_n(S)=\binom{s_1+b-1}{b-1}\binom{s_2-s_1+b-1}{b-1}\dots\binom{n-s_k+b-1}{b-1}, \end{equation*} because any such sequence of length $n$ is constructed by concatenating nondecreasing sequences of length $s_1, s_2-s_1,\dots,n-s_k$. Therefore \begin{equation*} \beta_n(S)=\sum_{1 \leq i_1<i_2<\dots<i_j\leq k} (-1)^{k-j} f(0,i_1)f(i_1,i_2)\dots f(i_j,k+1) \end{equation*} with $f(i,j)=\binom{s_j-s_i+b-1}{b-1}$ (with $s_0=0,\ s_{k+1}=n$). From Stanley's discussion, $\beta_n(S)$ is the determinant of the $(k+1)\times(k+1)$ matrix with $(i,j)$ entry $f(i,j+1),\ 0\leq i,j\leq k$. \end{proof} \begin{example} From the above development, the chance that the sum of $n$ base $b$ digits has \textit{no} carries is $\binom{n+b-1}{b-1}/b^n$. When $n=8,\ b=2$, this is $9/2^8\doteq.03516$. When $n=8,\ b=10$, this is $\binom{17}{9}/10^8\doteq.00024$. We feel lucky when adding eight numbers with no carries. These calculations show that we are lucky indeed. \label{ex22} \end{example} \section{Determinantal point processes}\label{sec3} Let $\x$ be a finite set. A point process on $\x$ is a probability measure $P$ on the $2^{|\x|}$ subsets of $\x$. For example, if $\x=\{1,2,\dots,n-1\}$, a point process is specified by recording where the carries occur in adding $n$ base $b$ numbers as in \ref{sec2}. One simple way to specify $P$ is via its so-called correlation functions. For $A\subseteq\x$, let \begin{equation*} \rho(A)=P\{S:S\supseteq A\}. \end{equation*} The collection of numbers $\{\rho(A)\}$ uniquely determine $P$ using inclusion/exclusion. A point process is \textit{determinantal} with kernel $K(x,y)$ if \begin{equation*} \rho(A)=\text{det}\left(K(x,y)\right)_{x,y\in A}. \end{equation*} On the right is the determinant of the $|A|\times|A|$ matrix with $(x,y)$ entry $K(x,y)$ for $x,y\in A$. Determinantal point processes were introduced by Macchi \cite{ma} to model the distribution of fermions. See \cite{dv} for a textbook account, \cite{sos} for a survey, \cite{hkpv} for probabilistic developments and \cite{ly} for many combinatorial examples. We warn the reader that these last references essentially deal with symmetric kernels $K(x,y)$ whereas all of our examples involve non-symmetric kernels. There has been an explosive development of determinantal point processes over the past few years. Some reasons for this include: \begin{enumerate} \item[(a)] A raft of natural processes (including all those in the present paper) turn out to be determinantal. These include: the eigenvalues of various ensembles of random matrices (Dyson \cite{dy1,dy2,dy3}, Mehta \cite{meh}); the presence or absence of edges if a random spanning tree is chosen in a graph (Burton--Pemantle \cite{bp}, Lyons \cite{ly}); random tilings and growth models (Johansson \cite{j2}); the structure of random partitions chosen from a variety of measures (e.g., the Plancherel measure on the symmetric group $S_n$) (Borodin--Okounkov--Olshanski \cite{boo}); random dimers on bipartite planar graphs (Kenyon \cite{ke}); and the zeros of a random analytic function $\sum_{n=0}^\infty a_n z^n$ with i.i.d. complex Gaussian coefficients $a_n$ \cite{pv}. The list goes on extensively. \item[(b)] Specifying a kernel may be much easier than specifying a measure on all $2^{|\x|}$ subsets. Further, the correlation functions allow easy computation of quantities of interest. For example, if $X_x$ is $1$ or $0$ as the random set includes $x$ or not: \begin{align*} E(X_x)&=K(x,x),\\ \text{Cov}(X_x,X_y)&=\text{det}\begin{pmatrix} K(x,x)&K(x,y)\\K(y,x)&K(y,y) \end{pmatrix}-K(x,x)K(y,y)\\ &=-K(x,y)K(y,x). \end{align*} If $K(x,y)=K(y,x)$, then the correlation is $\leq 0$, a distinctive feature of determinantal point processes. \item[(c)] If the matrix $(K(x,y))_{x,y\in\x}$ has all real eigenvalues $\{\lambda_x\}_{x\in\x}$, many theorems become available for $N$, the total number of points in a realization of the process. These include: \begin{itemize} \item $N$ is distributed as a sum $\sum_{x\in\x}Y_x$ with $\{Y_x\}$ independent $0/1$ random variables having $P(Y_x=1)=\lambda_x$. \item Let $\mu$ and $\sigma$ denote the mean and variance of $N$. These are available in terms of $K$ from (b) above. Then the following refined form of the central limit theorem holds: \begin{equation*} \left|P\left\{\frac{N-\mu}{\sigma}\leq x\right\}-\dfrac1{\sqrt{2\pi}}\int_{-\infty}^x e^{-t^2/2}dt\right|\leq \frac{.80}{\sigma} \end{equation*} \item The probability density $P(n)=P(N=n)$ is a Polya-frequency function. In particular the density is unimodal. Further, $|\text{mode}-\text{mean}|\leq1$. \item There is a useful algorithm for simulating from $K$ (here assuming $K(x,y)=K(y,x)$). \end{itemize} Most of these properties follow from the fact that the generating function of $N$, \begin{equation*} E(x^N)=\text{det}\left(I+(x-1) K\right) \end{equation*} has all real zeros. See Pitman \cite{Pi} or Hough, et al. \cite{hkpv} for details. \item[(d)] If $K_n$ is a sequence of such kernels converging to a kernel $K$, then, under mild restrictions, the associated point processes converge. See Soshnikov \cite{sos} or Johansson and Nordenstam \cite{jn} for precise statements and remarkable examples. A simple example is given in \ref{sect5} (Example \ref{exam41}). \end{enumerate} A main result of our paper is that the carries and descent processes, and indeed any one-dependent point processes on the integers, is determinantal. The special case of carries is stated here; a proof of the general case is in \ref{sect7}. \begin{thm} The point process of carries and descents $P_n$ given in \ref{sec2} is determinantal with correlation kernel $K(i,j)=k(j-i)$ where \begin{equation*} \sum_{m\in \mathbb{Z}}k(m)t^m=\dfrac1{1-(1-t)^b}. \end{equation*} \label{thm31} \end{thm} \begin{example} Take $b=2$. Then $\tfrac1{1-(1-t)^2}=\tfrac1{2t(1-t/2)}$. Replacing $t$ by $ct$ changes $k(m)$ to $k(m)c^m$ and $K(i,j)=k(j-i)$ to $c^{j-i}k(j-i)=c^{-i}K(i,j)c^j$. Conjugating $K$ by a diagonal matrix does not change determinants or the correlation functions. Thus setting $\tfrac{t}2=z$, the generating function becomes \begin{equation*} \dfrac14\left(\dfrac1{z}+1+z+z^2+\dots\right). \end{equation*} When e.g., $n-1=5$, the kernel is \begin{equation} \left(\dfrac14\right)\begin{bmatrix} 1&1&1&1&1\\ 1&1&1&1&1\\ 0&1&1&1&1\\ 0&0&1&1&1\\ 0&0&0&1&1\\ \end{bmatrix}. \label{totpos} \end{equation} From this, $\rho(A)={\footnotesize\begin{cases}0&\text{if $A$ has two consecutive entries}\\ \left(\tfrac14\right)^{|A|}&\text{otherwise}\end{cases}}$. This is a manifestation of the one-dependence together with the fact that (for $b=2$), two consecutive carries are impossible. It follows that the \textit{possible} configurations are all binary strings of length $n-1$ with no two consecutive $1$'s. This is a standard coding of the Fibonacci number $F_{n+1}$ (if $1,1,2,3,5$ are the first five Fibonacci numbers). For example, when $n=4$, the possible configurations are $000,001,010,100,101$. The eigenvalues of any finite matrix of the form (\ref{totpos}) are real. This follows from the fact that these matrices are totally positive, see \cite{ed}, and the well known fact that the eigenvalues of a totally positive matrix are real (and nonnegative), see e.g. Corollary 6.6 of \cite{an}. The classification of \cite{ed} also implies that the correlation kernels for $b\ge 3$ are not totally positive (for large enough matrix size), and we do not know if their eigenvalues are real. \label{ex31} \end{example} \begin{example} Take $b=3$. A straightforward expansion leads to \begin{eqnarray*} \sum_{m=-\infty}^\infty k(m)t^m & = & \dfrac1{1-(1-t)^3}= \frac{1}{3t \left( 1-t+ \frac{t^2}{3} \right)} \\ & = & \frac{ \left(1+t+\frac{2t^2}{3}+\frac{t^3}{3}+\frac{t^4}{9} \right)} {3t \left(1+\frac{t^6}{27} \right)}. \label{31} \end{eqnarray*} Thus \begin{itemize} \item $k(n)=0$ for $n<-1$, \item $k(-1)=\frac{1}{3}$, \item $k(n)=0$ for $n=4+6j,\ 0\leq j<\infty$. For other $n\geq2$, \item $k(n)=(-1)^{\lfloor\tfrac{n+1}6\rfloor}\left(\tfrac13\right)^{\lfloor \tfrac{n+3}2 \rfloor} 2^{\delta(n)}$, where $\delta(n)=1$ if $n=1 \mod 6$ and $0$ else. \end{itemize} For example, the first few values are: \begin{equation*}\begin{small}\begin{array}{r|rrrrrrrrrrrrrrrrrr} m&-1&0&1&2&3&4&5&6&7&8&9&10&11&12&13&14&15&16\\\hline k(m)&\tfrac13&\tfrac13&\tfrac29&\tfrac19&\tfrac1{27}&0&-\tfrac1{3^4}&-\tfrac1{3^4} &-\tfrac2{3^5}&-\tfrac1{3^5}&-\tfrac1{3^6}&0& \tfrac1{3^7}&\tfrac1{3^7}&\tfrac2{3^8}&\tfrac1{3^8}&\tfrac1{3^9}&0 \end{array}\end{small} \end{equation*} It is amusing to see that the two-variable kernel $K(x,y)=k(y-x)$ reproduces the correlation functions found for this problem in \ref{sec2} (Fact 4). Thus for $n=7$, \begin{equation*} (K(x,y))_{i,j=1}^6= \begin{bmatrix} 1/3 & 2/9 & 1/9 & 1/27& 0 &-1/81\\ 1/3& 1/3 & 2/9& 1/9 & 1/27& 0\\ 0& 1/3& 1/3 & 2/9 & 1/9& 1/27\\ 0&0& 1/3& 1/3& 2/9& 1/9 \\ 0&0&0&1/3&1/3&2/9\\ 0&0&0&0&1/3&1/3\end{bmatrix} \end{equation*} \label{ex32} \end{example} \section{One-dependent processes}\label{newsec4} Let $\{X_i\}_{i=0}^N$ be a binary stochastic process. Here $0<N\leq\infty$. The process is \textit{one-dependent} if $\{X_j\}_{j=0}^{i-1},\ \{X_j\}_{j=i+1}^N$ are independent for all $i$. A natural example: let $\{Y_i\}_{i=0}^{N+1}$ be independent uniform random variables on $[0,1]$. Let $h:[0,1]^2\to\{0,1\}$ be a measurable function. Then $X_i=h(Y_i,Y_{i+1}),\ 0\leq i\leq N$, is a one-dependent process called a \textit{two-block factor}. Much of the literature on one-dependence is for stationary processes. However, if $Y_i$ are independent but not identically distributed or $h=h_i$ depends on $i$, the associated $X_i$ is still one-dependent. Our main results in \ref{sect7} are for general one-dependent processes. We first describe some results for the stationary case. \subsection*{Stationary one-dependent processes} A book-length review of this subject by de Valk \cite{vdv} collects together many of the results, so we will be brief, focusing on later research. One of the first problems studied is the question of whether all one-dependent processes are two-block factors. Counterexamples were found \cite{vdv} though it appears that most ``natural'' one-dependent processes are two-block factors. For example, let $f\in L^2(\pi)$ be a function on the unit circle. Let $K(i,j)=\hat{f}(i-j),\ i,j\in\mathbb{Z}$, $\hat{f}$ the Fourier transform. Macchi \cite{ma} and Soshnikov \cite{sos} show that $K(i,j)$ is the kernel for a determinantal point process on $\mathbb{Z}$. Lyons and Steif \cite{ls} and Shirai and Takahashi \cite{st1,st2} develop remarkable properties of the associated point process. If $f(\theta)=ae^{-i\theta}+b+ce^{i\theta}$, the point process is one-dependent and \cite{brom} shows it is a two-block factor. See also \cite{cf}. A stationary one-dependent process is determined by the numbers $a_i=P(X_1=X_2=\dots=X_{i-1}=1),\ a_1=1$. Indeed, if $f=f_1,\dots,f_{i-1};e=e_{i+1},\dots,e_{n-1}$ are binary strings, $P(f,0,e)+P(f,1,e)=P(f)P(e)$. From this $P(f,0,e)=P(f)P(e)-P(f,1,e)$. By similar reductions, the probability of any pattern of occurrences can be reduced to a polynomial in the $a_i$. For example, $P(0,0,0)=1-3a_2+a_2^2+2a_3-a_4$. There is a remarkable, simple formula for this polynomial as a minor of a Toeplitz matrix. We learned this from \citep[Chap.~7]{pp}. A generalization to the non-stationary case is in \ref{sect7}. \begin{thm} \label{newthm41} For a stationary one-dependent process, let $a_i=P(X_1=X_2=\dots=X_{i-1}=1),\ a_1=1$. Let $t_1,\dots,t_{n-1}$ be a binary string with $j$ zeros at positions $S = \{s_1<\cdots<s_k\} \subseteq[n-1]$. Then \[ P(t_1,\cdots,t_{n-1})= \det\left(a_{s_{j+1}-s_{i}} \right)_{i,j=0}^{k} .\] Here the determinant is of a $k+1$ by $k+1$ matrix, and one sets $s_0=0,s_{k+1}=n, a_0=1,$ and $a_i=0$ for $i<0$. \end{thm} \begin{proof} Theorem \ref{newthm41} follows from Proposition 2.2 and Theorem 4.2 in \citep[Chap.~7]{pp}, after elementary manipulations. \end{proof} {\it Remark:} It follows from the discussion on \citep[p.~140]{pp} that if all of the determinants (for all $n$ and subsets $S$) in Theorem \ref{newthm41} are non-negative, then a stationary, one-dependent process with these $a_i$ values exists. \begin{example} For any $n$, the sequence with $n-1$ ones has $k=0$ zeros. The relevant matrix is then $1$ by $1$ with entry $a_n$, giving this probability. \label{newex42} \end{example} \begin{example} To compute $P(0,0,0)$ one applies the theorem with $n=4$ and $S=\{1,2,3\}$. It follows that \[ P(0,0,0) = \det \begin{bmatrix} 1&a_2&a_3&a_4\\ 1& 1&a_2&a_3\\ 0&1& 1&a_2\\ 0&0&1&1 \end{bmatrix} = 1-3a_2+a_2^2+2a_3-a_4,\] as above. \label{newex43} \end{example} Our next result expresses the determinant in Theorem \ref{newthm41} as a skew-Schur function of ribbon (also called rim-hook) type. Background on skew-Schur functions can be found in \cite[Sec.~1.5]{mac}. For further references and recent work, one can consult \cite{btw} or \cite{lp}. We will use the elementary symmetric functions $e_r$ in countably many variables defined as $e_0=1$ and \[ e_r = \sum_{1 \leq i_1 < i_2 < \dots < i_r} x_{i_1} x_{i_2} \cdots x_{i_r}.\] From \cite[p.~20]{mac} these are algebraically independent, so there is a homomorphism of the ring of symmetric functions to $\mathbb{R}$ which sends each $e_i$ to $a_i$ from Theorem \ref{newthm41}. \begin{thm} \label{skewschur} With notation as in Theorem \ref{newthm41}, let $\lambda$ and $\mu$ be the partitions defined by \[ \lambda_i=n-s_{i-1}-k+i-1 \ , \ \mu_i=n-s_{i}-k+i-1, \ \ 1 \leq i \leq k+1.\] Let $\lambda',\mu'$ denote the transpose partitions of $\lambda$ and $\mu$ and let $s_{\lambda'/\mu'}$ denote the corresponding skew-Schur function, obtained by specializing the elementary symmetric functions $e_i$ to equal $a_i$. Then \[ P(t_1,\cdots,t_{n-1})= s_{\lambda'/\mu'}.\] \end{thm} \begin{proof} First note that the $\lambda_i,\mu_i$ defined above are all non-negative. From Macdonald \cite[p.~71]{mac}, \begin{equation} \label{jac} s_{\lambda'/\mu'} = \det(e_{\lambda_i-\mu_j-i+j})_{i,j=1}^{k+1},\end{equation} and $e_r$ is the $r$th elementary symmetric function. When each $e_r$ is specialized to equal $a_r$, the quantity $s_{\lambda'/\mu'}$ becomes \[ \det(a_{\lambda_i-\mu_j-i+j})_{i,j=1}^{k+1} = \det(a_{s_j-s_{i-1}})_{i,j=1}^{k+1},\] as desired. \end{proof} One reason why Theorem \ref{skewschur} is interesting is that there are non-obvious equalities between ribbon skew-Schur functions. The paper \cite{btw} characterizes when two ribbon skew-Schur functions are equal; analogous results for more general skew-Schur functions are in \cite{RSW}. In particular, combining the results of \cite{btw} with Theorem \ref{skewschur} one immediately obtains the fact that a stationary one-dependent process is invariant under time reversal, i.e. \[ P(t_1,\cdots,t_n) = P(t_n,\cdots,t_1) \] for all $t_i$. For another proof of invariance under time reversal, see \cite[p.~139]{pp}. The determinantal formulae of Theorems \ref{newthm41} and \ref{skewschur} suggest that a determinantal point process is lurking nearby. This is indeed the case. In \ref{sect7} (see Corollary \ref{cor73}) we prove the following. \begin{cor} A stationary one-dependent process as in Theorem \ref{newthm41} is determinantal with kernel $K(x,y)=k(y-x)$ with \begin{equation*} \sum_{n\in\mathbb{Z}}k(n)z^n=-1/\sum_{j=1}^\infty a_jz^j. \end{equation*} \label{cor2} \end{cor} A remarkable development, connecting stationary one-dependent processes to algebra appears in \citep[Chap.~7]{pp}. They consider a graded algebra $A_0\oplus A_1\oplus A_2+\dots$ with $A_0=k$, a ground field, and $A_iA_j\subseteq A_{i+j}$. For example, the space $k[x]$ of polynomials in one variable has $A_0=k$, and $A_i$ spanned by $x^i$. They assume that each $A_i$ is a finite dimensional vector space of dimension $\dim(A_i)$. The algebra is \textit{quadratic} if $A_0=k$, the algebra is generated by elements of $A_1$, and the relations defining the algebra are in $A_2$. For example, the commutative polynomial ring $k[x_1,\dots,x_n]$ is generated by $x_1,\dots,x_n$ and the quadratic relations $x_ix_j-x_jx_i=0$. Note that algebras need not be commutative. A technical growth condition on the $\dim(A_i)$ which we will not explain here yields the Koszul algebras. These include \textit{many} natural algebras occurring in mathematics (see \cite{fr} for a survey). As a simple example, consider the commutative polynomial ring generated by $x_1,x_2,\dots,x_n$ with the additional relations $x_ix_j=0$ for $(i,j)\in E$, with $E$ the edge set of an undirected graph on $\{1,2,\dots,n\}$ (loops allowed). This is Koszul \citep[Chap.~2, Cor.~4.3]{pp}. Polishchuk and Positselski \citep[Chap.~7, Cor.~4.3]{pp} prove the following remarkable result. \begin{thm} \label{deep} To every Koszul algebra $A$ one can assign a stationary, one-dependent process via \begin{equation*} P(X_1=X_2=\dots=X_{i-1}=1)=\frac{\dim(A_i)}{\dim(A_1)^i} \qquad\text{for }i=1,2,\dots. \end{equation*} \label{newthm42} \end{thm} Part of the reason that Theorem \ref{deep} is substantive is that the quantities $P(X_1=X_2=\dots=X_{i-1}=1)$ can not assume arbitrary values; the determinants in Theorem \ref{newthm41} must be non-negative. \begin{example} Consider the commutative polynomial algebra generated by $x_1,x_2,\dots x_b$ and the additional relations $x_i^2=0,\ 1\leq i\leq b$. The degree $i$ part $A_i$ is spanned by square free monomials and has $\dim(A_i)=\binom{b}{i},\ 0\leq i<\infty$. From Fact 2 of \ref{sec2}, we see that the associated one-dependent process is precisely the carries process of mod $b$ addition from our introduction. Note that since only the $\dim(A_i)$ matter, the algebra in this example can be replaced by the exterior algebra generated by $x_1,x_2,\dots, x_b$ and the relations $x_ix_j=-x_jx_i$ for all $i,j$. \label{newex44} \end{example} \begin{example} Consider the commutative polynomial algebra generated by $x_1,x_2,\dots, x_b$. The degree $i$ part $A_i$ is spanned by monomials of degree $i$ and has $\dim(A_i)=\binom{b+i-1}{i}$. From Example \ref{ex22} of \ref{sec2}, the associated one-dependent process is precisely the complement of the carries process of mod $b$ addition. More generally, in \ref{sect7} we show that the particle-hole involution of a (perhaps non-stationary) one-dependent process with respect to any subset is one-dependent. \label{newex45} \end{example} \begin{example} A PBW algebra is an algebra of the form \[ k[x_1,\cdots,x_n]/\langle x_ix_j: (i,j) \in S \subset [1,n]^2 \rangle,\] where the variables $x_1,\cdots,x_n$ do not commute. Here it is useful to think of $S$ as a directed graph with vertex set $\{1,\cdots,n\}$, loops allowed. From \cite[p.~84]{pp}, PBW algebras are Koszul. As noted in \citep[Chap.~7, Prop.~5.1]{pp}, the one-dependent process associated to a PBW algebra can be described as follows: pick $U_1,U_2,\cdots$ i.i.d. in $\{1,\cdots,n\}$ and let $X_i=h(U_i,U_{i+1})$, where $h(i,j)=0$ if and only if $(i,j) \in S$. Indeed, the dimensions satisfy $\dim(A_0)=1, \dim(A_1)=n$, $\dim(A_i)=$number of paths of length $i-1$ in the complement of $S$, and so the chance of $i-1$ consecutive $1$'s in the point process is equal to $\dim(A_i)/\dim(A_1)^i$. The above argument shows that PBW algebras give rise to two-block-factor processes. In fact they are dense in the set of all two-block factors \cite[Chap.~7, Prop.~5.2]{pp}. For a generalization of these PBW processes, as well as a simple argument that they are determinantal, see Remark 5 after Theorem \ref{thm52}. \end{example} \begin{example} Consider $2n$ points $x_1,x_2,\dots, x_{2n}$ in general position in projective space $\mathbb{P}^n$. The coordinate ring of this projective variety is known to be Koszul \citep[p.~42]{pp} with $\dim(A_0)=1,\dim(A_1)=n+1,\dim(A_i)=2n$ for $i\geq2$. Thus $P(X_1=X_2=\dots=X_{i-1}=1)=2n/(n+1)^i$ defines a one-dependent process. When $n=1$, this is fair coin tossing. When $n=3$, it is the carries process for multiplying quaternions introduced in \ref{sect6}. For general $n$, these one-dependent processes can be described as follows (and illustrates what one might consider to be a ``probabilistic'' description). Let $\{1,2,\cdots,n\}$ be ordered cyclically and let $*$ be another symbol. Choose $U_i$ i.i.d. in $\{*,1,\cdots,n\}$. Let $X_i=h(U_i,U_{i+1})$ where $h(*,x)=0$ for all $x$, $h(x,*)=1=h(i,i+1)$ for $x \neq *$, and $h(i,j)=0$ otherwise. To see that this works, note that $P(X_1=1)=2n/(n+1)^2$, since one has to choose $x$ different from $*$ first and then $*$ or $x+1$ next. Similarly, for the chance of $i-1$ $1$'s in a row, the first choice can be anything but $*$, then the next $i-2$ choices are determined and the last can be one of two. \label{newex46} \end{example} We are certain that natural Koszul algebras will lead to natural point processes. We note further that \citep[Sec.~7.6]{pp} shows how natural operations on algebras preserve the Koszul property. These include the operations of union and complements that we work with in \ref{sect7}. \section{Descents in permutations}\label{sect5} A permutation $\sigma\in S_n$, the symmetric group, has a descent at $i$ if $\sigma(i)>\sigma(i+1)$. The set of such $i$ forms the descent set $D(\sigma)$. If the base $b$ in previous sections is large and $n$ stays fixed, a string of $n$ digits will have no repeated values and the descent theory of \ref{sec2} and \ref{sec3} above becomes descent theory for random permutations. This is a venerable subject. Stanley \cite{sta1,sta2.1} reviews the basics. Some modern highlights are Solomon's descent algebra \cite{sol}, the connections with the free Lie algebra \cite{gars,reu}, Gessel's theory of enumerating permutations by descents and cycle structure \cite{pd77,ges}, quasi-symmetric functions and the theory of riffle shuffling \cite{ful02,sta3}. Any Coxeter group has its own descent theory \cite{sol}. There is some indication that arithmetic carries can be carried over \cite{pd173}. This section introduces three examples where the descent set can be shown to be a one-dependent determinantal point process: uniform choice of permutations, non-uniform choice from the Mallows model, and independent trials. Then it shows that the closely related notion of the connectivity set of a permutation also yields a determinantal process. Finally, connections with binomial posets are mentioned. \begin{example} \label{des1} \textbf{The descent set of a uniformly chosen permutation.}\quad Consider the formula of \ref{sec2} Fact 5 for the chance that a random $b$-ary string of length $n$ has descents exactly at $S\subseteq[n-1]$. Passing to the limit, as $b\nearrow\infty$ using $\binom{a+b-1}{b-1}\sim\frac{b^a}{a!}$ gives a classical formula for the chance that a uniformly chosen $\sigma\in S_n$ has descents exactly at $S$: \begin{equation} P_n\left(D(\sigma)=S\right)=\text{det}\left[1\big/\left(s_{j+1}-s_i\right)!\right]. \label{41} \end{equation} For $S=1\leq s_1<s_2<\dots<s_k\leq n-1$, the determinant is of a $(k+1)\times(k+1)$ matrix with $(i,j)$ entry $1/(s_{j+1}-s_i)!, (i,j)\in[0,k]\times[0,k]$ and $s_0=0, s_{k+1}=n$. This formula is originally due to MacMahon \cite{mcm}. See Stanley \citep[p.~69]{sta2} or Gessel and Viennot \cite{gv} for modern proofs. \label{exam41} \end{example} Equation \eqref{41} allows us to see that the descent set of a uniformly random permutation, treated as a point process, is determinantal. The proof follows from \eqref{41} and Corollary \ref{cor1} in \ref{sect7}. \begin{thm} There exists a stationary point process on $\mathbb{Z}$, call it $P$, such that its restriction to any interval of length $n-1$ coincides with $P_n$ of \eqref{41}. The process $P$ is one-dependent and determinantal. Its correlation kernel $K(x,y)=k(y-x)$ with \begin{equation} \sum_{m\in\mathbb{Z}}k(m)z^m=\dfrac1{1-e^{z}}. \label{42} \end{equation} \label{thm41} \end{thm} \begin{rems}\ \begin{enumerate} \item Theorem \ref{thm41} can also be seen by passing to the limit in Theorem \ref{thm31}. Replacing $t$ in the generating function $1/(1-(1-t)^b)$ by $-z/b$ and letting $b\nearrow\infty$ gives \eqref{42}. \item The Bernoulli numbers $B_n$ are defined by $z/(e^z-1)=\sum_{n=0}^\infty B_nz^n/n!$. These are very well-studied. It is known that $B_{2i+1}=0$ for $i\geq1$, $B_0=1, B_1=-\tfrac12, B_2=\tfrac16, B_4=-\tfrac1{30}, B_6=\tfrac1{42}, \dots$. We see that $k(m)=-B_{m+1}/(m+1)!$. \item From these calculations, the kernel $K$ is \begin{equation*} K=\begin{bmatrix} \tfrac12&-\tfrac1{12}&0 &\tfrac1{720}&\cdots\\ -1 &\tfrac12 &-\tfrac1{12}&0 &\ddots\\ &-1 &\ddots &\ddots &\ddots\\ & &\ddots &\ddots &\ddots\\ & & &\ddots &\ddots \end{bmatrix}\ . \end{equation*} \item The correlation functions of $P$ (and $P_n$) can be computed explicitly by passage to the limit from Fact 5 of \ref{sec2}. Let $A=A_1\cup A_2\cup\dots\cup A_k$ be a decomposition of the finite set $A\subseteq \mathbb{Z}$ into disjoint blocks of adjacent integers. If $|A_i|=a_i$, \begin{equation*} \rho(A)=\prod_{i=1}^k\dfrac1{(a_i+1)!}. \end{equation*} \item For any $n=1,2,\dots$, let $K_n(i,j)=k(j-i)$ be the $n\times n$ top left corner block of $K$. Let $d(\pi)$ denote the number of descents of $\pi$ and let \[ A_{n+1} = \sum_{\pi \in S_{n+1}} x^{d(\pi)} \] be the $(n+1)$st Eulerian polynomial (see e.g., \citep[p.~22]{sta2}). Then item (c) preceding Theorem \ref{thm31} yields that \begin{equation*} A_{n+1}(x)=\text{det}\left(I+(x-1)K_n\right)(n+1)! \end{equation*} It is known \citep[p.~292]{com} that $A_{n+1}(x)$ has all real zeros $\alpha_1,\alpha_2,\dots,\alpha_n$. From the development in \ref{sec3}, $N_{n+1}$, the number of descents in a random permutation from $S_{n+1}$, is the sum of $n$ independent Bernoulli random variables with success rates $(1-\alpha_j)^{-1},\ 1\leq j\leq n$. One easily computes \begin{align*} E(N_{n+1})=\text{tr}(K_n)=\dfrac{n}2,\qquad\text{Var}(N_{n+1})&=\text{tr} \left(K_n-K_n^2\right)\\ &=\dfrac{n}2-\left(\frac{5n}{12}-\dfrac16\right)= \frac{n+2}{12}. \end{align*} It follows by Harper's method (\cite{har},\cite{Pi}) that one has a central limit theorem: \begin{equation*} \frac{N_{n+1}-\tfrac{n}2}{\sqrt{\frac{n+2}{12}}}\Longrightarrow\mathcal{N}(0,1). \end{equation*} \item Chebikin \cite{ch} studied the ``alternating descent set'' $A(\sigma)$ of a random permutation, defined as the set of positions at which the permutation has an {\it alternating descent}, that is an ordinary descent if $i$ is odd, or an ascent if $i$ is even. Combining Lemma 2.3.1 of \cite{ch} with the argument used to prove \eqref{41}, one obtains that for $|S|=k$, \[ P_n(A(\sigma)=S) = \det[E_{s_{j+1}-s_i}/(s_{j+1}-s_i)!]_{i,j=0}^k,\] where $E_n$ is the nth Euler number defined by $\sum_{n \geq 0} E_n z^n/n! = \tan(z) + \sec(z),$ and $s_0=0, s_{k+1}=n$. Applying our Theorem \ref{thm51} in \ref{sect7}, it follows that that $P_n$ is obtained by restricting to any interval of length $n-1$ the stationary, one-dependent, determinantal process with correlation kernel $K(x,y)=k(y-x)$ with \[ \sum_{m \in \mathbb{Z}} k(m) z^m= \frac{1}{1-(\tan(z)+\sec(z))^{-1}}.\] \item Modern combinatorics suggests a host of potential generalizations of Theorem \ref{thm41}. Let $P$ be a partial order on $[n]$. A linear extension of $P$ is a permutation $\sigma\in S_n$ such that if $i$ is less than $j$ according to $P$, then $\sigma(i)<\sigma(j)$. Let $\mathcal{L}(P)$ denote the set of linear extensions of $P$. For each $\sigma\in\mathcal{L}(P)$, let \begin{align*} D(\sigma)&=\{i:\sigma(i)>\sigma(i+1)\},&\text{the descent set of }\sigma,\\ \wedge(\sigma)&=\{i:\sigma(i-1)<\sigma(i)>\sigma(i+1)\},&\text{the peak set of }\sigma. \end{align*} Choosing $\sigma$ uniformly in $\mathcal{L}(P)$ gives two point processes. This section has focused on ordinary descents, that is, $P$ is the trivial poset with no restrictions. There are many indications that ``natural'' posets will give rise to determinantal point processes. For background and first results, see Brenti \cite{bren} or Stembridge \cite{stem}. \item Finally, we mention that it would be interesting to find analogs of Theorem \ref{thm41} for other Coxeter groups. For the hyperoctahedral group $B_n$ consisting of the $2^nn!$ signed permutations, descents can be defined using the linear ordering \[ 1<2<\dots<n<-n<\dots<-2<-1.\] Say that \begin{enumerate} \item $\sigma$ has a descent at position $i$ $(1 \leq i \leq n-1)$ if $\sigma(i) > \sigma(i+1)$. \item $\sigma$ has a descent at position $n$ if $\sigma(n)<0$. \end{enumerate} For example, $-4 \ -1 \ 3 \ 2 \ 5 \ \in B_3$ has descent set $\{2,3,5\}$. Reiner \cite{Re} shows that the chance that a random element of $B_n$ has descent set $\{s_1,\cdots,s_r\} \subseteq \{1,2,\cdots,n\}$ is $\det(a(i,j))_{i,j=0}^r$ where \[ a(i,j) = \left\{ \begin{array}{ll} \frac{1}{(s_{j+1}-s_i)!} & \mbox{if $0 \leq j \leq r-1$}\\ \frac{1}{2^{n-s_i}(n-s_i)!} & \mbox{if $j=r$} \end{array} \right.\] Here $s_0=0$. Theorem \ref{thm51} in \ref{sect7} shows that the resulting processes is determinantal (but not stationary). \end{enumerate} \end{rems} \begin{example} \label{des2} \textbf{Descents from non-uniform distributions on permutations.}\quad In a variety of applications, \textit{non}-uniform distributions are used on permutations. A widely used model is \begin{equation} P_\theta(\sigma)=Z^{-1}\theta^{d(\sigma,\sigma_0)},\qquad\sigma\in S_n. \label{43} \end{equation} Here, $0<\theta\leq1$ is a fixed concentration parameter, $\sigma_0\in S_n$ is a fixed location parameter, $d(\sigma,\sigma_0)$ is a metric on $S_n$ and $Z=\sum_\sigma\theta^{d(\sigma,\sigma_0)}$ is the normalizing constant. These are called ``Mallows models through the metric $d$.'' When $\theta=1$, $P_1$ is the uniform distribution. For $0<\theta<1$, $P_\theta$ is largest at $\sigma_0$ and falls off as $\sigma$ moves away from $\sigma_0$. See \cite{cri}, \cite{dr} or \cite{mar} for background and references. \label{exam42} \end{example} Perhaps the most widely used metric is \begin{equation}\begin{aligned} d(\sigma,\sigma_0)&=\text{minimum number of pairwise adjacent transpositions}\\ &\qquad\text{to bring $\sigma$ to $\sigma_0$},\\ &=I\left(\sigma\sigma_0^{-1}\right),\text{ the number of inversions in }\sigma \sigma_0^{-1}. \end{aligned}\label{44} \end{equation} This is called ``Kendall's tau'' in the statistics literature. References, extensions and properties are in \cite[Sect.~4]{d}, \cite{flig}, \cite{mal}. We mention that the normalizing constant of the Mallows model through Kendall's tau is given by $Z=\prod_{i=1}^n \frac{\theta^i-1}{\theta-1}$ \cite[p.~21]{sta2}. With $\theta,\sigma_0$ and $d$ fixed, one may ask any of the usual questions of applied probability and enumerative combinatorics: ``Picking a permutation randomly from $P_\theta(\cdot)$, what is the distribution of the cycle structure, longest increasing subsequence, \textellipsis?'' For general metrics, these are open research problems. Even the algorithmic task of sampling from $P_\theta$ leads to difficult problems. See \cite{dh} or \cite{d}. For Mallows model $P_\theta$ through Kendall's tau \eqref{44} and $\sigma_0=\text{id}$, we show that the descent set of $\sigma,\ D(\sigma)=\{i:\sigma(i+1)<\sigma(i)\}$ forms a determinantal point process with simple properties. In what follows we use the $q$-analog notation that $i_q=\frac{q^i-1}{q-1}$ and $n!_q=\prod_{i=1}^n i_q$, and take $\theta=q$ as is conventional in combinatorial work in the subject. \begin{prop} Let $P_q$ $(0<q<1)$ be Mallows model through Kendall's tau \eqref{43}, \eqref{44} on $S_n$. \begin{itemize} \item[(a)] The chance that a random permutation chosen from $P_\theta$ has descent set $s_1<s_2<\dots<s_k$ is \begin{equation*} \emph{det}\left[\frac{1}{(s_{j+1}-s_i)!_q}\right]_{i,j=0}^k, \end{equation*} with $s_0=0$ and $s_{k+1}=n$. \item[(b)] The point process associated to $D(\sigma)$ is stationary, one-dependent, and determinantal with kernel $K(x,y)=k(y-x)$, where \begin{equation*} \sum_{m \in \mathbb{Z}} k(m)z^m=\dfrac1{1-\left(\sum_{m=0}^\infty z^m\big/m!_q \right)^{-1}}. \end{equation*} \item[(c)] The chance of finding $k$ descents in a row is $q^{\binom{k+1}{2}}/(k+1)!_q$. In particular, the chance of a descent at any given location is $q/(q+1)$. The number of descents has mean $\mu(n,q)=\frac{(n-1)q}{q+1}$ and variance $q \left[ \frac{(q^2-q+1)n-q^2+3q-1}{(q^2+q+1)(q+1)^2} \right]$. Normalized by its mean and variance, the number of descents has a limiting standard normal distribution. \end{itemize} \label{prop1} \end{prop} \begin{proof} Part (a) of Proposition \ref{prop1} follows from Stanley \cite[Ex.~2.2.5]{sta2}. Part (b) follows from Corollary \ref{cor1} in \ref{sect7} and elementary calculations. To compute the chance of $k$ consecutive descents, note by stationarity that this is the chance of the permutation $\sigma(i)=k+2-i$ in $S_{k+1}$ under the $P_q$ measure on $S_{k+1}$. This probability is $q^{\binom{k+1}{2}}/(k+1)!_q$, as needed. One calculates from part (b) that \[ k(-1)=1, \ k(0)=\frac{q}{q+1}, \ k(1)=\frac{q^2}{(q+1)^2(q^2+q+1)}.\] Recall from the remarks preceding Theorem \ref{thm31} that \[ E(x^N)=\text{det}\left(I+(x-1) K\right),\] where $N$ is the total number of particles. Thus \[ E(|D(\sigma)|) = tr(K_{n-1}) = (n-1)k(0), \] \[ Var(|D(\sigma)|) = tr(K_{n-1}-K_{n-1}^2) = (n-1)k(0) - (n-1)k(0)^2 - 2(n-2)k(1)k(-1),\] and the proof of part (c) is complete. \end{proof} {\it Remarks.} \begin{enumerate} \item From part (c), the correlation functions are given explicitly, just as in remark 4 following Theorem \ref{thm41}. \item It would be interesting to have a direct combinatorial proof of the one-dependence in Proposition \ref{prop1}. \end{enumerate} \begin{example} \textbf{Descents for independent trials}\quad The carries in the carries process have an equivalent description in terms of descents in an independent and uniform sequence. In this example we generalize to descents in a sequence $Y_1,Y_2,\dots,Y_n$ with $Y_i$ independent and identically distributed with $P(Y_1=j)=p_j$. Here $0\leq j\leq b-1$ to match previous sections (and of course $0\leq p_j\leq1, \sum p_j=1$), but in fact $b=\infty$ can be allowed. The descents are related to the order statistics for discrete random variables \cite{abn,eld} but we have not seen this problem previously treated in the probability or statistics literature. Of course, if $Y_i$ are independent and identically distributed from a continuous distribution, the descent theory is the same as the descent theory for a random permutation. \label{exam54} \end{example} It is useful to have the complete homogeneous symmetric functions \begin{equation*} h_k(x_1,x_2,\dots,x_m)=\sum_{1\leq i_1\leq\dots\leq i_k\leq m}x_{i_1}x_{i_2}\dots x_{i_k}. \end{equation*} \begin{thm} If $Y_1,Y_2,\dots,Y_n$ are independent with common probability $P(Y_1=j)=p_j$, $0\leq j\leq b-1$, the descents in $Y_1,Y_2,\dots,Y_n$ form a stationary one-dependent point process on $[n-1]$ with \begin{equation*} P_n(S)=\emph{det}\left(h_{s_{j+1}-s_i}\right)_{0\leq i,j\leq k}\qquad\text{with }s_0=0,s_{k+1}=n, \end{equation*} for $S=\{s_1<s_2<\dots<s_k\}$. The $h_j$ are evaluated at $p_0,\dots,p_{b-1}$. Moreover, this process is determinantal with $K(i,j)=k(j-i)$ and $\sum_{l=-\infty}^\infty k(l)z^l=\tfrac1{1-\prod_{i=0}^{b-1}(1-p_iz)}$. \label{thm52} \end{thm} \begin{proof} Let $\alpha_n(S)=P_n(\text{des}(Y_1,Y_2,\dots,Y_n)\subseteq S)=h_{s_1}h_{s_2-s_1}\dots h_{n-s_k}$. The determinant formula follows from the argument for Fact 5 in \ref{sec2}. Now, Corollary \ref{cor1} implies that $\hat{k}(z)=\tfrac1{1-\hat{e}(z)^{-1}}$, with $\hat{e}(z)=\sum_{l=0}^\infty h_l z^l=\prod_{i=0}^{b-1}(1-p_iz)^{-1}$ from e.g., Macdonald \cite[p.~21]{mac}. \end{proof} \begin{rems}\ \begin{enumerate} \item Setting $p_i=1/b,\ \hat{k}(z)=\tfrac1{1-(1-z/b)^b}$. This is equivalent to our result in Theorem \ref{thm31}, by changing variable $z$ to $z/b\to t$, which doesn't change correlations. \item From the theorem, the associated point process is one-dependent. If $X_i=1$ or $0$ as there is a descent at $i$ or not, then clearly \[ P(X_1=X_2=\dots=X_i=1) = e_{i+1}(p_0,\cdots,p_{b-1}),\] where $e_r(x_1,\cdots,x_m) = \sum_{1 \leq i_1<\cdots<i_r \leq m} x_{i_1} x_{i_2} \cdots x_{i_m}$ is the $r$th elementary symmetric function. For example, \begin{align*} P(X_i=1)&=\dfrac12-\dfrac12\sum_ip_i^2,\\ P(X_i=X_{i+1}=1)&=\dfrac16-\dfrac12\sum_ip_i^2 + \dfrac13\sum_ip_i^3. \end{align*} This allows a straightforward formulation (and proof) of the central limit theorem for the number of descents. As in \ref{sec2}, simple expressions for the correlation functions are available in terms of $P(X_1=X_2=\dots=X_i=1)$. \item As we show in Corollary \ref{skew2}, one can express $P_n(S)$ as a skew-Schur function of ribbon type. For special choices of the $p_i$'s, this symmetric function interpretation may facilitate computation. For example, \cite[Sect.~1.5]{mac} gives $s_{\lambda/\mu}=\sum_{\nu} c_{\mu \nu}^\lambda s_{\nu}$ with $c_{\mu \nu}^\lambda$ the Littlewood--Richardson coefficients and $s_{\nu}$ the Schur function. If $p_i=zq^i$ ($i=0,1,\cdots,b-1$, $z$ a normalizing constant), then $s_{\nu}$ is non-0 only if $\nu$ has length $\leq b$, in which case \begin{equation*} s_{\nu}=z^{|\nu|}q^{n(\nu)}\prod_{x\in \nu}\frac{1-q^{b+c(x)}}{1-q^{h(x)}}; \end{equation*} see Macdonald \cite[p.~44]{mac} for an explanation of notation. \item Coming back to carries, if a column of numbers is chosen independently on $\{0,1,\dots,b-1\}$ from common distribution $\{p_i\}_{i=0}^{b-1}$, the remainder column after addition evolves as a Markov chain with starting distribution $\{p_i\}$ and transition matrix $P(i,j)=p_{j-i}$ (indices mod $b$). This is a circulant with known eigenvalues and eigenvectors. We do not know the distribution of carries. \item There is a sweeping generalization of this example which leads to a large collection of determinantal point processes. Let $R$ be an arbitrary subset of $[N]\times[N]$. Fix a probability distribution $\theta_1,\theta_2,\dots,\theta_N$ on $[N]$. A point process $P_n$ on $[n-1]$ arises as follows: pick $Y_1,Y_2,\dots,Y_n$ independently from $\{\theta_i\}$. Let $X_i={\footnotesize\begin{cases}1&\text{if }(Y_i,Y_{i+1}) \notin R\\0&\text{otherwise}\end{cases}},\ 1\leq i\leq n-1$. Let $S=\{i:X_i=1\}$. Using essentially the arguments above, it is proved in Brenti \cite[Th.~7.32]{bren} that \begin{equation*} P_n(S)=\text{det}\left[h_{s_{j+1}-s_i}^R\right]_{i,j=0}^k\qquad\text{for }S=\{s_1<\dots<s_k\}, s_0=0, s_{k+1}=n, \end{equation*} where $h_j^R=h_j^R(\theta_1,\dots,\theta_N):=P_j(\emptyset)$. Here as usual, $h_0^R=1,h_j^R=0$ for $j<0$. This falls into the domain of Theorem \ref{thm51} and its corollary. It follows that $P_n$ is determinantal with kernel $K(i,j)=k(j-i),\ \sum_{l \in \mathbb{Z}} k(l)z^l=\tfrac1{\left(1-\left(\sum_{j=0}^\infty h_j^Rz^j\right)^{-1}\right)}$. Our examples are the special case $R(i,j)={\footnotesize\begin{cases}1&\text{if }i\leq j\\0&\text{otherwise}\end{cases}}$. We are certain that there are many other interesting cases. \end{enumerate} \end{rems} \begin{example} \label{connect} \textbf{Connectivity set of a permutation.}\quad Stanley \cite{sta4} studied a ``dual'' notion to the descent set of a permutation, which he called the connectivity set. For $\sigma \in S_n$, the connectivity set $C(\sigma)$ is defined as the set of $i$, $1 \leq i \leq n-1$, such that $\sigma \in S_{1,\cdots,i} \times S_{i+1,\cdots,n}$. For example, the permutation $\sigma= 3 \ 1 \ 2 \ 5 \ 4 \ 6$ satisfies $C(\sigma)=\{3,5\}$. The connectivity set also arises in the analysis of quicksort, where it is called the set of splitters of a permutation \cite[Sec.~2.2]{wil}. We will prove that the connectivity set of a random element of $S_n$ is a determinantal point process and determine its correlation kernel. For this the following two facts are helpful: \begin{enumerate} \item A permutation $\sigma \in S_n$ is called {\it indecomposable} or {\it connected} if $C(\sigma)=\emptyset$. Comtet \cite[Exer.~VII.16]{com} shows that the number $f(n)$ of connected permutations in $S_n$ satisfies \begin{equation} \sum_{n \geq 1} f(n)x^n = 1 - \frac{1}{\sum_{n \geq 0} n!x^n}. \label{genf} \end{equation} (An asymptotic expansion of $f(n)$ is given in \cite[Exer.~VII.17]{com}). \item \cite[Prop.~1.1]{sta4} Letting $S=\{s_1<s_2<\cdots<s_k\}$ be a subset of $\{1,2,\cdots,n-1\}$, \begin{equation} \label{cont} \left| \{\sigma \in S_n: S \subseteq C(\sigma) \} \right| = s_1! (s_2-s_1)! \cdots (s_k-s_{k-1})! (n-s_k)!.\end{equation} \end{enumerate} The following is our main result. \begin{thm} Let $C(\sigma)$ denote the connectivity set of a permutation $\sigma$ chosen uniformly at random from $S_n$. The point process corresponding to $C(\sigma) \cup \{0,n\}$ is determinantal, with state space $\{0,\cdots,n\}$ and correlation kernel $K(x,y)$ satisfying \[ \begin{array}{ll} K(0,y)= 1 & \mbox{all $y$}\\ K(n,y)= \delta_{n,y} & \mbox{all $y$} \\ K(x,y)= \frac{1}{{n \choose x}} & \mbox{$0,n \neq x \leq y$}\\ K(x,y)= \frac{1}{{n \choose x}} - \frac{1}{{n-y \choose n-x}} & \mbox{$0,n \neq x>y$} \end{array} \] \end{thm} Note that in the statement of the theorem, $0,n$ are always points of the process. \begin{proof} The first step is to observe that $C(\sigma) \cup \{0,n\}$ can be obtained as a trajectory of a certain Markov chain, started at $0$, with transition probabilities \begin{equation} P(i,j) = \frac{(n-j)!f(j-i)}{(n-i)!}. \label{trans} \end{equation} Here we take $f(0)=0$, so $P(i,i)=0$ for all $i$. To prove \eqref{trans}, note that $C(\sigma)=\{i_1,\cdots,i_k\}$ if and only if the following events $E_1,\cdots,E_{k+1}$ occur: \begin{itemize} \item $E_1$: $\{\pi(1),\cdots,\pi(i_1)\}=\{1,\cdots,i_1\}$ and $\pi$ restricted to $\{1,\cdots,i_1\}$ is indecomposable. \item $E_2$: $\{\pi(i_1+1),\cdots,\pi(i_2)\}=\{i_1+1,\cdots,i_2\}$ and $\pi$ restricted to $\{i_1+1,\cdots,i_2\}$ is indecomposable. \item $\cdots$ \item $E_{k+1}$: $\{\pi(i_k+1),\cdots,\pi(n)\}=\{i_k+1,\cdots,n\}$ and $\pi$ restricted to $\{i_k+1,\cdots,n\}$ is indecomposable. \end{itemize} Letting $f(n)$ denote the number of indecomposable permutations in $S_n$, the probability of $E_1$ is clearly $\frac{f(i_1)(n-i_1)!}{n!}$. The probability of $E_2$ given $E_1$ is $\frac{f(i_2-i_1)(n-i_2)!}{(n-i_1)!}$ and the probability of $E_3$ given $E_1,E_2$ is $\frac{f(i_3-i_2)(n-i_3)!}{(n-i_2)!}$, etc., as claimed. Now \cite[Thm.~1.1]{bo} implies that the point process on $C(\sigma) \cup \{0,n\}$ is determinantal with correlation kernel \begin{equation} \label{ck} K(x,y) = \delta_{0,x} + Q(0,x) - Q(y,x) \end{equation} where \[ Q=P+P^2+P^3+ \cdots.\] To compute $Q$, let $[x^r] g(x)$ denote the coefficient of $x^r$ in a series $g(x)$, and note that \begin{eqnarray*} P^2(i,j) & = & \sum_l P(i,l)P(l,j)\\ & = & \sum_l \frac{(n-l)!f(l-i)}{(n-i)!} \frac{(n-j)!f(j-l)}{(n-l)!}\\ & = & \frac{(n-j)!}{(n-i)!} [x^{j-i}] \left(1 - \frac{1}{\sum_{n \geq 0} n!x^n} \right)^2. \end{eqnarray*} The last equality used \eqref{genf}. A similar computation shows that \[ P^r(i,j) = \frac{(n-j)!}{(n-i)!} [x^{j-i}] \left(1 - \frac{1}{\sum_{n \geq 0} n!x^n} \right)^r,\] and thus \begin{eqnarray*} Q(i,j) & = & \frac{(n-j)!}{(n-i)!} [x^{j-i}] \frac{\left(1 - \frac{1}{\sum_{n \geq 0} n!x^n} \right)}{\left(\sum_{n \geq 0} n!x^n \right)^{-1}}\\ & = & \frac{(n-j)!}{(n-i)!} [x^{j-i}]\left(\sum_{n \geq 1} n!x^n \right). \end{eqnarray*} Thus $Q(i,j)=1/{n-i \choose n-j}$ if $i<j$, and is $0$ otherwise. The theorem follows from this and \eqref{ck}. \end{proof} {\it Remarks.} \begin{enumerate} \item The connectivity set $C(\sigma)$ is a simple example of a determinantal process which is not one-dependent. Indeed, from \eqref{cont}, $P(1 \in C(\sigma))=\frac{1!(n-1)!}{n!}$, $P(3 \in C(\sigma))=\frac{3!(n-3)!}{n!}$, and $P(1,3 \in C(\sigma)) = \frac{1!2!(n-3)!}{n!} \neq P(1 \in C(\sigma)) P(3 \in C(\sigma))$. \item Unlike the other point processes considered in this paper, the expected number of points tends to $0$ as $n \rightarrow \infty$. Indeed, applying \eqref{cont} gives that \[ E|C(\sigma)| = \sum_{i=1}^{n-1} P(i \in C(\sigma)) = \sum_{i=1}^{n-1} \frac{1}{{n \choose i}} \rightarrow 0 \] as $n \rightarrow \infty$. \end{enumerate} \end{example} \subsection*{Binomial posets In Doubilet--Rota--Stanley \cite{drs}, binomial posets were introduced as a unifying mechanism to ``explain'' the many forms of generating functions that appear in enumerative combinatorics. Briefly, $\mathcal{X}$ is a binomial poset if for every interval $[x,y]$, the length of all maximal chains is the same (say $n_{xy}$) and the number of maximal chains in an interval of length $n$ does not depend on the particular interval. This number $B(n)$ is called the factorial function. For the usual Boolean algebra of subsets, $B(n)=n!$. For the subspaces of a vector space over a finite field, $B(n)=n!_q$ ($q$-factorial). There are many further examples. Stanley \cite{sta1} has shown that many enumerative formulae generalize neatly to the setting of binomial posets. Some of these developments give new determinantal point processes. Here is an example. \begin{example} \label{int} (union of descent sets) \begin{thm} Pick $\sigma_1,\sigma_2,\dots,\sigma_n$ independently from Mallows model through Kendall's tau \eqref{43} on the symmetric group $S_n$. Let $S=\{s_1<\cdots<s_k\}$ be the union of their descent sets. Then \begin{equation*} P_n(S)=\emph{det}\left[\dfrac1{\left(\left(s_{j+1}-s_i\right)!_q\right)^r}\right], \end{equation*} where $s_0=0, s_{k+1}=n$. The associated point process is stationary, one-dependent, and determinantal with $K(x,y)=k(y-x)$ for $\hat{k}(z)=\tfrac1{1-1/\hat{e}(z)}$ where $\hat{e}(z)=\sum_{l=0}^\infty z^l/(l!_q)^r$. \label{thm53} \end{thm} \begin{proof} The formula for $P_n(S)$ is Corollary 3.2 of \cite{sta1}, so the theorem follows from one of our main results proved in \ref{sect7} (Corollary \ref{cor1}). \end{proof} When $q=1$ and $r=2,\ \hat{e}(z)=I_0(2 \sqrt{z})$, the classical modified Bessel function. Feller \cite[\S II.7]{fel} develops a host of connections with stochastic processes. \begin {rem} The intersection of two independent point processes with correlation functions $\rho^1(A),\rho^2(A)$ is a point process with correlation function $\rho^1(A)\rho^2(A)$. If both processes are determinantal, this is a product of determinants, but in general the intersection or union of determinantal processes is not determinantal. In Subsection \ref{all}, we show that all one-dependent processes are determinantal. Since the intersection of descent sets of independent permutations is one-dependent, and taking complements preserves one-dependence, this gives another proof that the union process in Theorem \ref{thm53} is determinantal. \end{rem} \end{example} We have not pursued other examples but again believe there is much else to be discovered. \section{More general carries}\label{sect6} Consider the quaternions $Q_8=\{\pm1,\pm i,\pm j,\pm k\}$. The center $Z$ of $Q_8$ is $\{\pm1\}$. Choose coset representatives $X=\{1,i,j,k\}$ for $Z$ in $Q_8$ so any element can be uniquely represented as $g=zx$. The coset representatives are multiplied by the familiar rule $k\curvearrowright i\curvearrowright j \curvearrowright k$ so $ij=k$ and $kj=-i$, etc. If we multiply $g_1g_2\dots g_k=(z_1x_1)(z_2x_2)\dots(z_kx_k)$, the $z_i\in\{\pm1\}$ can all be moved to the left and we must multiply $x_1 x_2\cdot\dots\cdot x_k$, keeping track of the ``carries'', here $\pm1$. Evidently, if $\{g_i\}$ are chosen uniformly at random in $Q_8$, both $\{z_i\}$ and $\{x_i\}$ are independent and uniform in $Z$ and $X$. Thus we have the following problem: choose $X_1,X_2,\dots,X_k$ uniformly in $X$ and multiply as $X_1$, $X_1X_2$, $(X_1X_2)X_3$, \dots, \begin{center} $X_1$\\ $X_2$\\ $X_3$\\ $\vdots$\\ $X_k$ \end{center} This gives a process of remainders and carries as in \ref{sec1}. \begin{example} \begin{equation*}\begin{array}{rrr} k&\cdot&k\\ k&&1\\ i&\cdot&i\\ i&&1\\ k&\cdot&k\\ j&\cdot&i\\ k&\cdot&j\\ i&&k \end{array}\qquad\text{gives }(-1)^5k=-k. \end{equation*} It is almost obvious that the carries form a stationary, one-dependent, two-block process with $P(i-1$ ones in a row) $=\tfrac{6}{4^{i}},\ 2\leq i<\infty$. Further, the remainders in the second column are independent and uniform on $\{1,i,j,k\}$ and there is a simple ``descent'' rule which determines the joint law of the dots (Example \ref{ex72} below). We also mention, by comparison with Example \ref{newex46} in \ref{newsec4}, that the carries process for the quaternions is the same point process that arises from the coordinate ring of 6 generic points in projective space $\mathbb{P}^3$. \label{ex71} \end{example} One natural generalization where all goes through is to consider a finite group $G$ and a normal subgroup $N$ contained in the center of the group. The factor group $F=G/N$ has elements labeled $1,\sigma,\tau,\dots$. We may choose coset representatives $t(1)=1,t(\sigma),t(\tau),\dots$ and any $g=nt(\sigma)$, uniquely. While sometimes $t(\sigma)t(\tau)=t(\sigma \tau)$, in many cases this fails; but we may choose correction factors $f(\sigma,\tau)$ in $N$ (often called a ``factor set'') so that $t(\sigma)t(\tau)=t(\sigma \tau)f(\sigma,\tau)$. Once $t(\sigma)$ are chosen, the $f(\sigma,\tau)$ are forced. \begin{example} If $G=C_{100},\ N=C_{10}$ (thought of as a subgroup $\{0,10,20,\dots,90\}$), the natural choice of coset representatives for $G/N \cong \{0,1,\dots,9\}$ is $t(i)=i\in G$. Of course, \begin{equation*} t(i)+t(j)=t(i+j)+f(i,j),\qquad\text{with }f(i,j)=\begin{cases} 1&\text{if }i+j\geq10\\ 0&\text{if }0\leq i+j<10.\end{cases} \end{equation*} It is natural to ask if the choice of cosets matters. To see that it can, consider $C_{10}$ in $C_{100}$ and choose coset representatives as $0,11,22,43,44,45,46,47,48,49$. The sum of $11$ and $22$ requires a carry of $90$. A lovely exposition of carries as cocycles is in \cite{is}. \end{example} If $g=nt(\sigma)$ is chosen uniformly at random, then $n$ and $t(\sigma)$ are independent and uniform. Multiplying a sequence of $g_i=n_it_i$ can be done by first multiplying the $t_i$, keeping track of the carries, and then multiplying the $n_i$ and carries in any order. Of course, here the carries are in $N$, not necessarily binary. Given $t_1,t_2,\dots,t_k$, we may form a two-column array with $t_i$ in the first column, the successive remainders $r_1,r_2,\dots r_k$ in the second column and ``carries'' $f_1,f_2,\dots f_{k-1}$ (elements in $N$) placed in between. \begin{lem} If coset representatives $t_i,\ 1\leq i\leq k$, are chosen uniformly, then \begin{enumerate} \item the remainder process $r_i,\ 1\leq i\leq k$, is uniform and independent, and \item the carries $f_i,\ 1\leq i\leq k-1$, form a stationary, one-dependent, two-block factor. \end{enumerate} \label{lem71} \end{lem} \begin{proof}\ \begin{enumerate} \item Since $r_1=t_1$, $r_1$ is uniform. Successive $r_i$ are formed by multiplying $r_{i-1}$ by $t_i$. There is a unique choice of $t_i$ giving $r_i$, so $r_i$ is uniform and independent of $r_1,\dots, r_{i-1}$. \item Consider two successive remainders and the unique $t$ giving rise to them: \begin{equation*}\begin{array}{cc} &r_{i-1}\\ t_i&r_i\end{array} \end{equation*} Since $t_i$ is uniquely determined, $r_{i-1}t_i=r_if(r_{i-1},t_i)$ is uniquely determined. It follows that the $f_i$ process is a two-block process: generate $r_1,r_2\dots$ uniformly and independently, set $f_i=h(r_i,r_{i+1})$, with $h(r,r')=f(r,t)$ where $r^{-1}r' \in tN$ determines $t$ uniquely. Because two-block processes are one-dependent, this completes the proof.\qedhere \end{enumerate} \end{proof} \begin{cor} With the notation of the lemma, define a binary process $B_1,\dots,B_{k-1}$ as \begin{equation*} B_i=\begin{cases} 1&\text{if }f_i\neq\emph{id}\\ 0&\text{if }f_i=\emph{id}. \end{cases}\end{equation*} Then $\{B_i\}$ is a stationary, one-dependent, two-block factor. \label{cor71} \end{cor} \begin{example} \label{ex72} As in Example \ref{ex71}, let $G=Q_8,\ N=\{\pm1\}, \ F=G/N\cong C_2\times C_2$. The natural choice of coset representatives $\{1,i,j,k\}$ gives rise to the following two-block representation: choose $U_1,U_2,\dots$ uniformly and independently in $\{1,i,j,k\}$. Let $B_i=h(U_i,U_{i+1})$ with $h(1,x)=1$ (all $x$), $h(x,1)=-1 (x\neq1),\ h(i,j)=h(j,k)=h(k,i)=-1,\ h(i,k)=h(k,j)=h(j,i)=1,\ h(x,x)=1,\ (x\neq1)$. \end{example} \begin{example}(dihedral group) Let $D_8=\langle x,y:x^2=y^2=(xy)^4=1\rangle$. If $z=xy$, this $8$-element group has center $N=\{1,z^2=-1\}$. Choosing coset representatives $1,x,y,z$, the cosets multiply as \begin{equation*}\begin{array}{c|cccc} &1&x&y&z\\\hline 1&1&x&y&z\\ x&x&1&z&y\\ y&y&-z&1&-x\\ z&z&-y&x&-1 \end{array}\end{equation*} From this, elementary manipulations show that $P(i$ ones in a row) $=1/4^i$. Thus, the carries process is independent with $P(B_i=1)=1/4, \ P(B_i=0)=3/4$ for all $i$. We mention that $D_8$ can also be represented as the extension of the normal subgroup $C_4$ by the factor group $C_2$. Here, coset representatives for $C_4$ can be chosen so that there are no carries (i.e. the extension ``splits''). \label{ex73} \end{example} \begin{example}(extensions of $C_2$ by $C_m$) A central extension of $N=C_2$ by $C_m$ is abelian. It follows that when $m$ is odd, $G=C_{2m}$ is the only central extension of $N=C_2$ by $C_m$, and when $m$ is even, there are two central extensions $C_2\times C_m$ and $C_{2m}$. The extension $C_2\times C_m$ splits and choosing $\{(0,i), 0 \leq i \leq m-1\}$ as the coset representatives, there are no carries. For $C_{2m}$, with $C_2\cong\{0,m\}$, choose coset representatives $0,1,2,\dots,m-1$. Thus \begin{equation*} f(i,j)=\begin{cases} 1&\text{if }i+j\geq m\\ 0&\text{if }0\leq i+j<m.\end{cases}\end{equation*} As for usual addition, there is a carry if and only if there is a descent in the remainder column. Thus \begin{equation*} P(B_1=B_2=\dots=B_{i-1}=1)=\frac{\binom{m}{i}}{m^i}. \end{equation*} \label{ex74} \end{example} \begin{example} Let $G=C_2 \times C_2 \times C_2$ and let $N=\{(0,0,0),(1,1,1)\}$. With coset representatives $(0,0,0),(1,0,0),(0,1,0),(1,1,0)$, there are never carries, but with coset representatives $(0,0,0),(1,0,0),(0,1,0),(0,0,1)$, there are carries. This example shows that the one-dependent process $B_1,B_2,\cdots$ depends not only on $G$ and $N$, but also on the choice of coset representatives. \label{ex75} \end{example} We have not embarked on a systematic study of carries for finite groups and believe that there is much more to do. We do note that by a result in the next section (Theorem \ref{rdet}), the above processes, being one-dependent, are determinantal. The basic carries argument works for infinite groups as well. Let $G$ be a locally compact group and $H$ a closed normal subgroup. Suppose that $H$ is in the center of $G$ and that $G/H$ is compact. Choose coset representatives $t(\sigma) \in G$ for $\sigma \in G/H$. As in the finite case, these define factor sets $f(\sigma,\tau)$ by $t(\sigma) t(\tau) = t(\sigma \tau) f(\sigma,\tau)$. Write $g=nt$. Since $G/H$ is compact, it has an invariant probability measure. Choosing $t_1,t_2,\cdots$ independently from this measure and multiplying as above gives remainders and a carries process. Just as above, the remainders are independent and uniformly distributed in $G/H$ and the carries process (with values in $N$) is a one-dependent, two-block factor process. \begin{example} A lovely instance of this set-up explains a classical identity. We begin with the motivation and then translate. Let $\sigma$ be a permutation with number of descents $d(\sigma)$. It is known that, for $\sigma$ chosen from the uniform distribution on $S_n$, \begin{equation} P\left( d(\sigma)=j \right) = P \left( j \leq U_1+ \cdots +U_n <j+1 \right). \label{918} \end{equation} On the right, $U_1,U_2,\cdots,U_n$ are independent uniforms on $[0,1]$. The density and distribution function for $U_1+\cdots+U_n$ was derived by Laplace; see \cite{fel}. Foata \cite{fo} proved \eqref{918} by combining Laplace's calculation with an identity of Worpitzky. Richard Stanley \cite{sta2.5} gives a bijective proof involving an elegant dissection of the $n$-dimensional hypercube. Jim Pitman \cite{Pi} gives the following ``proof from the book'' which is a continuous version of our carries argument from \ref{sec1}: form two columns \begin{equation*}\begin{array}{rr} U_1 & V_1\\ U_2 & V_2 \\ \cdot & \cdot \\ \cdot &\cdot \\ U_n & V_n \end{array}\end{equation*} On the left are independent uniforms on $[0,1]$. On the right are their remainders when added mod 1; so $V_1=U_1$, $V_2=U_1+U_2$ (mod $1$), $\cdots$. The $V_i$ are similarly independent uniforms on $[0,1]$. Place a dot at position $i$ every time the partial sum $U_1+\cdots+U_{i+1}$ crosses an integer. Call these dots carries. As in the discrete case, there is a dot at position $i$ if and only if there is a descent $V_{i+1}<V_i$. The number of dots is the integer part of $U_1+\cdots+U_n$ and also the number of descents, proving \eqref{918}. Of course, the distribution of the descent {\it process} is the same as the carries process. In the language of group theory, let $G=\mathbb{R}, N=\mathbb{Z}$, and $G/N \cong S_1$, the circle group. Choose coset representatives as $[0,1)$ and factor sets in $\{0,1\}$. \end{example} \section{Proofs and generalizations} \label{sect7} This section proves two of our main results for general one-dependent processes (we do not assume stationarity in this section). In Subsection \ref{all}, it is shown that all one-dependent point processes on $\mathbb{Z}$ are determinantal, a result which is new even in the stationary case. Subsection \ref{class} proves that a point process $P$ on a finite set $\x$, with $P$ given as a certain-shaped determinant, is one-dependent and determinantal. This covers quite a few examples from previous sections and is particularly useful in situations (such as Example \ref{des2}) where the one-dependence is not apriori obvious. \subsection{One dependent processes are determinantal} \label{all} For a random point process on a discrete set $\x$, we define the correlation function $\rho$ by \[ \rho(A) = P \{ S:S \supseteq A \}.\] Then one dependence on (a segment of) $\mathbb{Z}$ is equivalent to the condition that $\rho(X \cup Y) = \rho(X) \rho(Y)$ whenever $dist(X,Y) \geq 2$. \begin{thm} \label{rdet} Any one-dependent point process on (a segment of $\mathbb{Z}$) is determinantal. Its correlation kernel can be written in the form $K(x,y)=$ \[ \begin{array}{ll} 0 & \mbox{if $x-y \geq 2$}\\ -1 & \mbox{if $x-y = 1$}\\ \sum_{r=1}^{y-x+1} (-1)^{r-1} \sum_{x=l_0<l_1<\cdots<l_r=y+1} \rho \left([l_0,l_1) \right) \rho \left([l_1,l_2) \right) \cdots \rho \left([l_{r-1},l_r) \right) & \mbox{if $x \leq y$} \end{array} \] Here the notation $[a,b)$ stands for $\{a,a+1,\cdots,b-1\}$. \end{thm} For example, $K(x,x)=\rho(\{x\})$, $K(x,x+1)=\rho(\{x,x+1\})-\rho(\{x\}) \rho(\{x+1\})$, etc. \begin{proof} By one-dependence, it is enough to verify that with $K$ as above, \[ \det[K(x+i,x+j)]_{i,j=0}^{y-x} = \rho \left([x,y+1) \right) \] for any $x \leq y$. We use induction on $y-x$. For $y=x$ the statement is trivial. Otherwise, one has that \[ \det[K(x+i,x+j)]_{0}^{y-x} = \det \begin{bmatrix} K(x,x)& K(x,x+1) & \dots & K(x,y)\\ -1 & K(x+1,x+1) &\dots &\vdots\\ & -1 &\ddots&\vdots\\ & & -1 & K(y,y) \end{bmatrix}. \] When expanding this determinant, various numbers of $-1$'s from the subdiagonal can be used. Observe that if we do not use the $-1$ in position $(i,i-1)$, then we can compute the corresponding contribution, because if we replace that $-1$ by $0$, the determinant splits into the product of two each of which is computable by the induction hypothesis. Similarly, if we insist on not using several $-1$'s, then the contribution is the product of several determinants. Thus we obtain by inclusion-exclusion that \begin{equation} \begin{aligned} \det[K(x+i,x+j)]_{0}^{y-x} & = K(x,y) + \sum_{x<l<y+1} \rho \left([x,l)\right) \rho \left([l,y+1)\right) \\ & - \sum_{x<l_1<l_2<y+1} \rho \left([x,l_1)\right) \rho \left([l_1,l_2)\right) \rho \left([l_2,y+1)\right) + \cdots. \end{aligned} \label{eqv} \end{equation} The $K(x,y)$ term corresponds to using all $-1$'s, the sum over $l$ corresponds to not using a $-1$ at least at location $(l,l-1)$, the sum over $l_1,l_2$ corresponds to not using a $-1$ at least at locations $(l_j,l_j-1), j=1,2$, etc. Our definition of $K(x,y)$ is such that the right hand side of \eqref{eqv} is equal to $\rho \left([x,y+1)\right)$, as desired. \end{proof} To state some corollaries of Theorem \ref{rdet}, we use the concept of particle-hole involution. Essentially, given a subset $\cal{N}$ of $\x$, the involution maps a point configuration $S \subset \x$ to $S \bigtriangleup \cal{N}$ (here $\bigtriangleup$ is the symbol for symmetric difference). This map leaves intact the particles of $S$ outside of $\cal{N}$, and inside $\cal{N}$ it loses the particles and picks up the ``holes'' (points of $\cal{N}$ free of particles). \begin{cor} The class of one-dependent processes is closed under the operations of particle-hole involution on any fixed subset, intersections of independent processes, and unions of independent processes. \end{cor} \begin{proof} For intersections, the claim follows from the definitions. For particle-hole involutions, note that the property of being one-dependent follows from the kernel begin 0 on the second subdiagonal and below. On the other hand, for determinantal point processes the ``complementation principle'' \citep[Sec.~A.3]{boo} says that the particle-hole involution can be implemented by the following change in the kernel: \[ \begin{bmatrix} A& B \\ C& D \end{bmatrix} \rightarrow \begin{bmatrix} A& B \\ -C& I-D \end{bmatrix}, \] where the block structure corresponds to the splitting into the noninverted and inverted parts. Clearly, this keeps the property of having $0$'s on the second subdiagonal and below intact. Finally, unions can be reduced to intersections by the particle-hole involution. \end{proof} {\it Remark}. The class of determinantal point processes is {\it not} closed under intesections/unions. \vspace{.1in} Let us now look at the translation invariant case. Then $\rho \left([x,y) \right) = \rho_{y-x}$ is a function of $y-x$ only, and $\rho_k$ is the chance of $k$ consecutive ones. Set \[ R(z) = 1+z+\sum_{k \geq 1} \rho_k z^{k+1}.\] \begin{cor} \label{cor73} In the translation invariant case, the kernel $K(x,y)=k(y-x)$ is also translation invariant, and \[ \sum_{n \in \mathbb{Z}} k(n)z^n = \frac{1}{1-R(z)}.\] \end{cor} \begin{proof} \begin{eqnarray*} \frac{1}{1-R(z)} & = & - \frac{1}{z} \cdot \frac{1}{1+\rho_1 z+\rho_2 z^2 + \cdots} \\ & = & -\frac{1}{z} \left[1 - \sum_{m \geq 1} \rho_m z^m + \left(\sum_{m \geq 1} \rho_m z_m \right)^2 - \cdots \right] \\ & = & - \frac{1}{z} + \frac{1}{z} \left[ \sum_{m \geq 1} \rho_m z^m - \left(\sum_{m \geq 1} \rho_m z^m \right)^2 + \cdots \right], \end{eqnarray*} which agrees with the formula of Theorem \ref{rdet}. \end{proof} The generating function $R(z)$ also behaves well with respect to the particle-hole involution on $\mathbb{Z}$. \begin{prop} The particle-hole involution on $\mathbb{Z}$ with $\cal{N}=\mathbb{Z}$ replaces $R(z)$ by $1/R(-z)$. \end{prop} \begin{proof} We have \[ \frac{1}{1-1/R(-z)} = \frac{R(-z)}{R(-z)-1} = 1 - \frac{1}{1-R(-z)}.\] Hence by Corollary \ref{cor73}, changing $R(z)$ to $\tilde{R}(z)=1/R(-z)$ leads to the following change in the correlation kernel: \[ \tilde{k}(n)= \delta_{0,n} - (-1)^n k(n).\] This is equivalent to $\tilde{K}(x,y) = \delta_{x,y}- \frac{(-1)^x}{(-1)^y}K(x,y)$. This is the same as $K \rightarrow I-K$, which corresponds to the particle-hole involution \cite[Sec.~A.3]{boo}. \end{proof} \begin{example} The descent process on $\mathbb{Z}$ (Example \ref{des1}) corresponds to $\rho_m=\frac{1}{(m+1)!}$; thus $R(z)=e^z$. The particle-hole involution is given by $\tilde{R}(z)=1/e^{-z}=e^z$, which is the same. \end{example} \begin{example} \label{bor2} The intersection of $r$ independent descent processes on $\mathbb{Z}$ corresponds to $\rho_m= \frac{1}{(m+1)!^r}$, hence \[ R_{\cap}^{(r)}(z) = \sum_{m \geq 0} \frac{z^m}{(m!)^r}.\] The correlation kernel is given by $\sum_{n \in \mathbb{Z}} k(n)z^n = \frac{1}{1-R_{\cap}^{(r)}(z)}$. \end{example} \begin{example} The union of $r$ independent descent processes is the particle-hole involution of Example \ref{bor2}. Thus $R_{\cup}^{(r)}(z)=1/R_{\cap}^{(r)}(-z)$ and \[ \sum_{n \in \mathbb{Z}} k(n) z^n = \frac{1}{1-1/R_{\cap}^{(r)}(-z)}.\] Replacing $z$ by $-z$ doesn't affect correlations, and we recover Example \ref{int}. \end{example} \subsection{A class of determinantal processes} \label{class} For any $n=2,3,\dots$, consider a probability measure $P_n$ on all subsets $S=\{s_1<s_2<\dots<s_k\}\subseteq[n-1]$ given by \begin{equation} P_n(S)=h(n)\,\text{det}\left[e(s_i,s_{j+1})\right]_{i,j=0}^k, \label{51} \end{equation} for some $h:\mathbb{N}\to\mathbb{C}$ and $e:\mathbb{N}\times\mathbb{N}\to\mathbb{C}$ with the notation $s_0=0, s_{k+1}=n$. We assume that $e(i,j)=0$ for $i>j$, and that $e(i,i)=1, e(i,i+1)>0$ for all $i$. \begin{thm} If \eqref{51} holds for some fixed $n$ with $e(i,j)=0$ for $i>j$ and $e(i,i)=1,e(i,i+1)>0$ for all $i$, then $P_n$ is a determinantal, one-dependent process with correlation functions \begin{equation*} \rho(A)=P_n\{S:S\supseteq A\}=\emph{det}\left[K(a_i,a_j)\right]_{i,j=1}^m \qquad\text{for }A=\{a_1,a_2,\dots,a_m\}, \end{equation*} with correlation kernel \begin{equation} \label{kform} K(x,y)=\delta_{x,y}+(E^{-1})_{x,y+1}, \end{equation} where $E$ is the upper triangular matrix $E=[e(i-1,j)]_{i,j=1}^n$, \begin{equation*} E=\begin{bmatrix} e(0,1)&e(0,2)&\dots &e(0,n)\\ &e(1,2)&\dots &e(1,n)\\ & &\ddots&\vdots\\ & & &e(n-1,n) \end{bmatrix}\ . \end{equation*} In addition, $h(n)=(\det\,E)^{-1}=(e(0,1)e(1,2)\dots e(n-1,n))^{-1}$. \label{thm51} \end{thm} \begin{cor} Assume further that $e(i,j)=e(j-i)$. Then the point process is stationary. If $\hat{e}(z)=\sum_{l=0}^\infty e(l)z^l$, then $K(x,y)=k(y-x)$ and \begin{equation*} \hat{k}(z):=\sum_{l=-\infty}^\infty k(l)z^l=\dfrac1{1-1/\hat{e}(z)}. \end{equation*} \label{cor1} \end{cor} \begin{proof} [Proof of Theorem \ref{thm51}] Set $L=[e(i-1,j)+\delta_{i-1,j}]_{i,j=1}^n$. Thus $L$ appears as \begin{equation*} L=\begin{bmatrix} e(0,1)&e(0,2)&\dots &e(0,n)\\ 1 &e(1,2)&\dots &\vdots\\ &1 &\ddots&\vdots\\ & &1 &e(n-1,n) \end{bmatrix}. \end{equation*} For any function $f:\{1,\dots,n-1\}\to\mathbb{C}$, by \eqref{51}, \begin{align*} & E\left(\prod_{s_i \in S} f(s_i)\right)\\ &= \sum_{\substack{0<s_1<\dots<s_k<n\\k=0,\dots,n-1}}P_n\left(\{s_1<s_2<\dots<s_k\} \right)f(s_1)\dots f(s_k)\\ &=h(n)\sum_{\substack{0<s_1<\dots<s_k<n\\k=0,\dots,n-1}}\text{det} \left[L\binom{1,s_1+1,\dots,s_k+1}{s_1,\dots,s_k,n}\right]f(s_1)\dots f(s_k)\\ &=h(n) \cdot \text{det} \begin{bmatrix} f(1)e(0,1)& f(2) e(0,2)& \dots & f(n-1) e(0,n-1) & e(0,n)\\ f(1)-1 & f(2) e(1,2)& \dots & f(n-1) e(1,n-1) & e(1,n)\\ & f(2) - 1 & \ddots & f(n-1) e(2,n-1) & e(2,n)\\ & & \ddots & \dots & \dots \\ & & & f(n-1) -1 & e(n-1,n) \end{bmatrix}\\ &=h(n) \cdot\text{det}\left[\begin{bmatrix} 0&&&\\ -1&\ddots&&\\ &\ddots&\ddots&\\ &&-1&0 \end{bmatrix}+L\begin{bmatrix} f(1)&&&\\ &\ddots&&\\ &&f(n-1)&\\ &&&1 \end{bmatrix}\right]. \end{align*} This is the generating functional of $P_n$. In the second equality, the determinant is of the minor of $L$ with rows $1,s_1+1,\cdots,s_k+1$ and columns $s_1,\cdots,s_k,n$. In the third inequality, the $2^{n-1}$ possible summands correspond to choosing which of the first $n-1$ matrix columns use the $-1$ coming from $f(i)-1$ in the determinant expansion. If $f=1+g$, the generating functional can be expressed in terms of the correlation functions: \begin{eqnarray*} & & 1+\sum_{\substack{s_1<\dots<s_m\\m=1,2,\dots}}\rho_m(s_1,\dots,s_m)g(s_1)\dots g(s_m)\\ & = & E\left[\prod_{s_i \in S} \left(1+g(s_i)\right)\right]\\ &= & h(n)\,\text{det}\left(E+L\begin{bmatrix} g(1)&&&\\ &\ddots&&\\ &&g(n-1)&\\ &&&0 \end{bmatrix}\right)\\ &= & h(n)\,\text{det}(E)\cdot\text{det}\left(I+E^{-1}L\cdot\begin{bmatrix} g(1)&&&\\ &\ddots&&\\ &&g(n-1)&\\ &&&0 \end{bmatrix}\right). \end{eqnarray*} This holds for all $g$. First take $g=0$ to see $h(n)\,\text{det}(E)=1$. Next note the expansion: if $M$ is $n\times n$, then $\text{det}[I+M]=\sum_S \text{det}(M(S))$ with the sum over all $2^n$ subsets of $[n]$, and $M(S)$ the minor with rows and columns in $S$. This gives that $\rho_m(s_1,\dots,s_m)=\text{det}[K(s_i,s_j)]_{i,j=1}^m$, where $K(x,y)=(E^{-1}L)_{xy}$. Then \begin{equation*} E^{-1}L=E^{-1}\left(\begin{bmatrix} 0&&&\\ 1&\ddots&&\\ &\ddots&\ddots&\\ &&1&0 \end{bmatrix}+E\right)=I +E^{-1} \begin{bmatrix} 0&&&\\ 1&\ddots&&\\ &\ddots&\ddots&\\ &&1&0 \end{bmatrix},\qedhere \end{equation*} and the proof of \eqref{kform} is complete. Finally, we note that since $K(x,y)$ vanishes below the first subdiagonal, $P_n$ is one-dependent. \end{proof} \begin{proof}[Proof of Corollary] Here $E$ is a Toeplitz matrix with symbol $(\hat{e}(z)-1)/z$. It is triangular, so $E^{-1}$ is a Toeplitz matrix with symbol $z/(\hat{e}(z)-1)$. Thus $K$ is a Toeplitz matrix with symbol $\hat{k}(z)=1+\tfrac1{\hat{e}(z)-1}=\tfrac1{1-(\hat{e}(z))^{-1}}$. \end{proof} We also note that in the translation invariant case, there is an expression for $P_n(S)$ in terms of skew Schur functions of ribbon type. \begin{cor} \label{skew2} As in Corollary \ref{cor1}, assume further that $e(i,j)=e(j-i)$. Let $\lambda$ and $\mu$ be the partitions defined by \[ \lambda_i=n-s_{i-1}-k+i-1 \ , \ \mu_i=n-s_{i}-k+i-1, \ \ 1 \leq i \leq k+1.\] Let $\lambda',\mu'$ denote the transpose partitions of $\lambda$ and $\mu$ and let $s_{\lambda'/\mu'}$ denote the corresponding skew-Schur function, obtained by specializing the elementary symmetric functions $e_i$ to equal $e(i)$. Then $P_n(S) = \frac{1}{e(1)^n} s_{\lambda'/\tau'}$. \end{cor} \begin{proof} By assumption, formula \eqref{51} becomes \[ P_n(S)=h(n) \cdot \text{det}\left(e_{s_{j+1}-s_i}\right)_{i,j=0}^k.\] Now use the argument of Theorem \ref{skewschur}, together with the identification of the normalizing constant $h(n)=1/e(1)^n$ in Theorem \ref{thm51}. \end{proof} \begin{rems}\ \begin{enumerate} \item Another approach to Theorem \ref{thm51} is via the theory of conditional $L$-ensembles in Borodin--Rains \cite[Prop.~1.2]{br}. \item In the translation invariant case $e(i,j)=e(j-i)$, Stanley \cite[p.~90, Ex.~14]{sta2} shows that \begin{equation*} P_n\left([n-1]\right)=\text{chance that all sites are occupied} \end{equation*} is the coefficient of $z^{n}$ in the power series $h(n)\hat{e}(-z)^{-1}$. To prove this using symmetric function theory, note from Corollary \ref{skew2} that \[ P_n\left([n-1]\right)=h(n) \cdot s_{(n)}.\] Here the Schur function $s_{(n)}=h_n$ (where $h_n$ is the $n$th complete homogeneous symmetric function and $h(n)$ is the normalizing constant in \eqref{51}). If $E(z)=\sum_{r=0}^\infty e_r z^r$ is the generating function for elementary symmetric functions and $H(z)=\sum_{r=0}^\infty h_{r}z^r$, Macdonald \cite[p.~21, (2.6)]{mac} shows that $H(z)E(-z)=1$. This gives Stanley's formula since in Corollary \ref{skew2}, the value of $e_r$ is $e(r)$. \end{enumerate} \end{rems} \begin{example} \textbf{(Descents in a random sequence)}\quad From Fact 5 of \ref{sec2}, we are in the Toeplitz case with $e(j)=\binom{j+b-1}{b-1}, \hat{e}(z)=(1-z)^{-b}$. Applying Corollary \ref{cor1} yields Theorem \ref{thm31}. \label{exam51} \end{example} \begin{example} \textbf{(Descents in a uniform permutation)}\quad From MacMahon's formula \eqref{41}, we are again in the Toeplitz case with $e(j)=1/j!, \hat{e}(z)=e^{z}$. As in Example \ref{ex31}, one can replace $z$ by $-z$ without changing determinants or the correlations functions. This proves Theorem \ref{thm41}. \label{exam52} \end{example} \begin{example} \textbf{(Descents in a non-uniform permutation)}\quad From Stanley's formula (part a of Proposition \ref{prop1}), we are again in the Toeplitz case with $e(j)=1/j!_q$ ($q$-factorial). An identity of Euler allows one to write \[ \hat{e}(z) = \prod_{m \geq 0} \frac{1}{1-z(1-q)q^m} \] when $0<q<1,|z|<1$, but the elementary description of the correlation functions given in Proposition \ref{prop1} seems more useful. \label{exam53} \end{example} \section*{Acknowledgments} Borodin was partially supported by NSF grant DMS 0707163. Diaconis was partially supported by NSF grant DMS 0505673. Fulman was partially supported by NSF grant DMS 0802082 and NSA grant H98230-08-1-0133.

https://arxiv.org/abs/2007.13196

Best low-rank approximations and Kolmogorov n-widths

We relate the problem of best low-rank approximation in the spectral norm for a matrix $A$ to Kolmogorov $n$-widths and corresponding optimal spaces. We characterize all the optimal spaces for the image of the Euclidean unit ball under $A$ and we show that any orthonormal basis in an $n$-dimensional optimal space generates a best rank-$n$ approximation to $A$. We also present a simple and explicit construction to obtain a sequence of optimal $n$-dimensional spaces once an initial optimal space is known. This results in a variety of solutions to the best low-rank approximation problem and provides alternatives to the truncated singular value decomposition. This variety can be exploited to obtain best low-rank approximations with problem-oriented properties.

\section{Introduction}\label{sec:introduction} The problem of approximating a given matrix by another matrix of a lower rank is labeled as the problem of low-rank approximation (of matrices). It aims to obtain a more compact representation of data with limited loss of information. Low-rank approximation of matrices is ubiquitous in applications: discretization of partial differential equations, principal component analysis, image processing, data mining, and machine learning, to name a few; see, e.g., \cite{Kumar:2017} for a survey. In particular, it plays an important role in matrix completion \cite{Cai:2010}, which finds in the so-called \emph{Netflix problem} one of its most well-known applications \cite{Hallinan:2016}. In this paper we consider the classical problem of \emph{best} low-rank approximation of matrices measured in the spectral norm. Let $A$ be an $m\times m$ real matrix of rank $r$, then we seek rank-$n$ matrices $R_n$, $n<r$, such that \begin{align* \|A-R_n\|\leq \|A-B\|, \end{align*} for any $m\times m$ matrix $B$ of rank $n$, and where $\|\cdot\|$ is the operator norm induced by the Euclidean norm, i.e., the spectral norm. The singular value decomposition (SVD) is an essential tool for analyzing and solving the best low-rank approximation problem; see, e.g., \cite[Chapter~3]{Blum:2020}. Let $A=U \Sigma V^T$ be any SVD of $A$, i.e., $\Sigma$ is the diagonal matrix whose diagonal entries, $$ \sigma_1\geq \sigma_2\geq\dots\geq \sigma_r>\sigma_{r+1}=\dots= \sigma_m=0, $$ are the singular values of $A$, and $U$ and $V$ are orthonormal matrices. We further let $\mathbf{u}_j$ and $\mathbf{v}_j$ denote the $j$-th column vector of $U$ and $V$. If $n<r$, then the Eckhart--Young theorem \cite[Theorem~2.4.8]{Golub:2013} states that the rank-$n$ matrix \begin{equation}\label{eq:ba-SVD} R_n=\sum_{i=1}^n\sigma_i\mathbf{u}_i\mathbf{v}_i^T \end{equation} satisfies \begin{equation}\label{eq:rank-n-approx-gen} \|A-R_n\|=\min_{\operatorname{rank}(B)=n} \|A-B\|=\sigma_{n+1}, \end{equation} and is thus a best rank-$n$ approximation to $A$ in the spectral norm. However, in many applications one is interested in finding low-rank approximations that preserve certain \emph{structures} in the original matrix $A$, i.e., structured low-rank approximation \cite{Chu:2003,Higham:1989,Markovsky:2008,Ottaviani:2014, Ishteva:2014,Grussler:2018}. Preserving these structures could exclude the matrix $R_n$ in \cref{eq:ba-SVD} from being a suitable approximation, and in general one looks for \emph{near-best} approximations that preserve these structures. In this paper we provide a classification of other best low-rank approximations to $A$ than $R_n$ in \cref{eq:ba-SVD}. One could then search among these matrices for best low-rank approximations that have the desired structure or other problem-oriented properties. In fact, the special case of best rank-$1$ approximations to Hankel matrices has already been considered in \cite{Antoulas:1997}; see also \cite{Knirsch:preprint} where further results and efficient algorithms for structured best rank-$1$ approximations to Hankel matrices can be found. We also remark that the problem of finding best low-rank approximations in other (entry-wise) matrix norms has been studied in \cite{Pinkus:2012} and \cite{Georgieva:2017}. Observe that the matrix $R_n$ in \cref{eq:ba-SVD} is clearly not unique if $\sigma_n=\sigma_{n+1}>0$ and it is then straightforward to find other best rank-$n$ approximations to $A$. If $\sigma_n>\sigma_{n+1}>0$ it is known that the matrix in \cref{eq:ba-SVD} is the unique best rank-$n$ approximation to $A$ in the Frobenius norm; see, e.g., \cite[Section~7.4.2]{Horn:2013}. However, as argued by Tropp \cite[p.~122]{Tropp:2015}, error bounds in the Frobenius norm are not always useful in cases of practical interest and can even be completely ``vacuous''; see also \cite{Li:2017,Musco:2015} for a similar argument. It is therefore more desirable to look for low-rank approximations in the spectral norm. For this norm the problem has infinitely many solutions whenever $\sigma_{n+1}>0$, because any matrix of the form \begin{equation} \label{eq:no-unique-spectral} \sum_{i=1}^n(\sigma_i+\epsilon_i)\mathbf{u}_i\mathbf{v}_i^T, \quad -\sigma_{n+1}\leq\epsilon_i\leq\sigma_{n+1}, \end{equation} solves \cref{eq:rank-n-approx-gen}. In this paper we look for more general solutions of the form $\sum_{i=1}^n\mathbf{x}_i\mathbf{y}_i^T$ with $\mathbf{x}_i,$ $\mathbf{y}_i\in \mathbb{R}^m$, other than \cref{eq:ba-SVD} and \cref{eq:no-unique-spectral}, to the best low-rank approximation problem in \cref{eq:rank-n-approx-gen}. Our approach to finding other best rank-$n$ approximations to $A$ consists of two steps: first we relate this problem to Kolmogorov $n$-widths \cite{Kolmogorov:36} and then we solve the $n$-width problem. The Kolmogorov $n$-width of a set in a normed linear space is the minimal distance to the given set from all possible $n$-dimensional subspaces. An $n$-dimensional (sub)space is \emph{optimal} when it realizes this minimal distance. We provide a classification of all the optimal $n$-dimensional spaces for the image of the Euclidean unit ball under $A$, which can be recognized as an $r$-dimensional ellipsoid in $\mathbb{R}^m$. It turns out that the corresponding Kolmogorov $n$-width equals $\sigma_{n+1}$ and that any orthonormal basis in such $n$-dimensional optimal space generates a best rank-$n$ approximation to $A$. This results in a large variety of best rank-$n$ approximations beyond the truncated SVD solution in \cref{eq:ba-SVD}, and can be exploited to obtain low-rank approximations with problem-oriented properties. As a byproduct of our results we classify all $n$-dimensional spaces that achieve the minimum in the following min-max formulation for the singular values of~$A$: \begin{equation}\label{eq:min-max} \sigma_{n+1} = \min_{\mathbb{X}_n}\max_{\mathbf{z}\perp \mathbb{X}_n} \sqrt{\frac{\mathbf{z}^TAA^T\mathbf{z}}{\mathbf{z}^T\mathbf{z}}}. \end{equation} This formula is a direct consequence of the Courant--Fischer theorem \cite[Section~7.3]{Horn:2013}. It is easily verified that $\mathbb{X}_n={\rm span}\{\mathbf{u}_1,\ldots,\mathbf{u}_n\}$ achieves the minimum in \cref{eq:min-max}. However, as already pointed out in \cite{Karlovitz:73,Karlovitz:76}, this space is unique only in very special cases. For further relations between the $n$-width and matrix theory we refer the reader to the survey paper \cite{Pinkus:79}, and for further $n$-width results in general to the book \cite{Pinkus:85}. In this paper we restrict our attention to the case where the $(n+1)$-st singular value is non-zero and unique, i.e., \begin{equation} \label{eq:sigmas-unique-weak} \sigma_n>\sigma_{n+1}>\sigma_{n+2}\geq0. \end{equation} Besides the above discussion, this assumption is taken to simplify the exposition since it ensures that the $(n+1)$-st left singular vector of $A$ is unique (up to multiplication by constants). All our findings can be easily extended to rectangular matrices $A$ of rank $r$. The remainder of this paper is organized as follows. \Cref{sec:low-rank} states the definitions of Kolmogorov $n$-widths and optimal spaces for the image of the Euclidean unit ball by $A$ and connects them with best rank-$n$ approximations to $A$. Some known necessary or sufficient conditions for a subspace to be optimal are recalled in \cref{sec:optimal-subspaces}. \Cref{sec:optimality-crit} is the core of the paper and provides characterizations of optimal subspaces by means of some optimality criteria. We discuss them in detail for the important case of best rank-$1$ approximation in \cref{sec:1-width}. Some alternative optimality criteria are collected in \cref{sec:optimality-crit-alt}. \Cref{sec:tp-matrices,sec:seq-subspaces}, inspired by similar results for integral operators in $L^2$, present a simple explicit construction to obtain a sequence of optimal $n$-dimensional subspaces once an initial optimal subspace is given. This construction can be exploited to obtain alternative best rank-$n$ spectral approximations for any matrix $A$. Some concluding remarks are collected in \cref{sec:conclusion}. \section{Kolmogorov $n$-widths and rank-$n$ approximations}\label{sec:low-rank} Let $A$ be an $m \times m$ real matrix of rank $r$, and define the subset of $\mathbb{R}^m$, \begin{align*} {\cal A} &:= \{A\mathbf{x} : \mathbf{x} \in \mathbb{R}^m,\, \|\mathbf{x}\| \le 1\}, \end{align*} where $\| \cdot \|$ is the Euclidean norm in $\mathbb{R}^m$. Note that ${\cal A}$ can be recognized as a (filled) $r$-dimensional ellipsoid in $\mathbb{R}^m$, where the line segments $[-\sigma_i \mathbf{u}_i, \sigma_i \mathbf{u}_i]$, $i=1,\ldots,r$, are its principal axes. The spectral norm of $A$ is the induced operator norm given by \begin{align*} \|A\|:=\max_{\|\mathbf{x}\|\leq1} \|A\mathbf{x}\|, \end{align*} and it can be shown that $\|A\|=\|A^T\|=\sigma_1$. For an $n$-dimensional subspace $\mathbb{X}_n$ of $\mathbb{R}^m$, where $0 \le n \le m$, we define the distance to ${\cal A}$ from $\mathbb{X}_n$ by \begin{equation}\label{eq:E} E({\cal A}, \mathbb{X}_n) :=\max_{\mathbf{a} \in {\cal A}} {\rm dist}(\mathbf{a},\mathbb{X}_n)= \max_{\mathbf{a} \in {\cal A}} \min_{\mathbf{x} \in \mathbb{X}_n} \|\mathbf{a}-\mathbf{x}\|. \end{equation} Then, the Kolmogorov $n$-width of ${\cal A}$, relative to the Euclidean norm in $\mathbb{R}^m$, is defined by $$ d_n({\cal A}) := \min_{\mathbb{X}_n} E({\cal A}, \mathbb{X}_n). $$ A subspace $\mathbb{X}_n$ of $\mathbb{R}^m$ is called an optimal subspace for ${\cal A}$ provided that $$E({\cal A}, \mathbb{X}_n) = d_n({\cal A}). $$ Here the $0$-dimensional subspace $\mathbb{X}_0$ of $\mathbb{R}^m$ is $\{0\}$. We can determine the $n$-width of ${\cal A}$ for any $n=0,\ldots,m$ as follows. Let $P_n$ be the orthogonal projection onto $\mathbb{X}_n$. Then, \begin{equation}\label{eq:max0} \begin{aligned} E({\cal A}, \mathbb{X}_n) &= \max_{\mathbf{a} \in {\cal A}} \|\mathbf{a}-P_n\mathbf{a}\| = \max_{\|\mathbf{x}\| \le 1} \|(I-P_n)A\mathbf{x}\|=\|(I-P_n)A\|\\ &=\|A^T(I-P_n)\| =\max_{\mathbf{x}\neq 0} \frac{\|A^T(I-P_n)\mathbf{x}\|}{\|\mathbf{x}\|}, \end{aligned} \end{equation} where we have used that the spectral norm of a matrix equals the spectral norm of its adjoint. By letting $\mathbf{x}=\mathbf{y}\oplus\mathbf{z}$ for $\mathbf{y}\in\mathbb{X}_n$ and $\mathbf{z}\perp\mathbb{X}_n$ one can check that the last maximum in \eqref{eq:max0} is achieved for $\mathbf{y}=0$. This implies that \begin{equation}\label{eq:max} E({\cal A}, \mathbb{X}_n) = \max_{\mathbf{z}\perp \mathbb{X}_n} \frac{\|A^T\mathbf{z}\|}{\|\mathbf{z}\|} = \max_{\mathbf{z}\perp \mathbb{X}_n} \sqrt{\frac{\mathbf{z}^TAA^T\mathbf{z}}{\mathbf{z}^T\mathbf{z}}}. \end{equation} Now, using the definition of $d_n({\cal A})$, together with \cref{eq:min-max} and \cref{eq:max}, we observe that \begin{equation}\label{eq:width=sv} d_n({\cal A}) = \sigma_{n+1}, \quad n=0,1,\ldots, m-1. \end{equation} We also note that it easily follows from the definition of the $n$-width that $d_m({\cal A}) = 0$, due to the fact that the only choice of a subspace of $\mathbb{R}^m$ of dimension $m$ is $\mathbb{X}_m = \mathbb{R}^m$. Thus, we have $$ (d_0({\cal A}),d_1({\cal A}),\ldots, d_m({\cal A})) = (\sigma_1,\sigma_2,\ldots,\sigma_m,0), $$ and, as mentioned in the introduction, $\mathbb{X}_n={\rm span}\{\mathbf{u}_1,\ldots,\mathbf{u}_n\}$ is an optimal space for ${\cal A}$. The relation between Kolmogorov $n$-widths and rank-$n$ approximations is contained in the next two theorems. \begin{theorem}\label{thm:low-rank-1} Assume that the vectors $\mathbf{x}_i$, $i=1,\ldots,n$, are orthonormal, and define $\mathbf{y}_i:=A^T\mathbf{x}_i$, $i=1,\ldots,n$. If $\mathbb{X}_n:={\rm span}\{\mathbf{x}_1,\ldots,\mathbf{x}_n\}$, then \begin{equation*} \|A-\sum_{i=1}^n\mathbf{x}_i\mathbf{y}_i^T\| = E({\cal A},\mathbb{X}_n), \end{equation*} and, consequently, the matrix $\sum_{i=1}^n\mathbf{x}_i\mathbf{y}_i^T$ is a best rank-$n$ approximation to $A$ if and only if the subspace $\mathbb{X}_n$ is optimal for ${\cal A}$. \end{theorem} \begin{proof} Let $P_n$ be the orthogonal projection onto $\mathbb{X}_n$. It follows from \cref{eq:max0} that \begin{align*} E({\cal A},\mathbb{X}_n) &=\max_{\|\mathbf{x}\|= 1} \|A\mathbf{x}- P_nA\mathbf{x}\| =\max_{\|\mathbf{x}\|= 1} \|A\mathbf{x}- \sum_{i=1}^n(\mathbf{x}_i^TA\mathbf{x})\mathbf{x}_i\| \\ &=\max_{\|\mathbf{x}\|= 1} \|A\mathbf{x}- \sum_{i=1}^n((A^T\mathbf{x}_i)^T\mathbf{x})\mathbf{x}_i\| = \max_{\|\mathbf{x}\|= 1} \|A\mathbf{x}- \sum_{i=1}^n\mathbf{x}_i\mathbf{y}_i^T\mathbf{x}\| \\ &=\|A-\sum_{i=1}^n\mathbf{x}_i\mathbf{y}_i^T\|. \end{align*} Since $d_n({\cal A})=\sigma_{n+1}$, the result follows. \end{proof} We remark that the above theorem can be considered as an extension of an observation in \cite{Pinkus:2012}. Define the subset ${\cal A}_T:= \{A^T\mathbf{x} : \mathbf{x} \in \mathbb{R}^m,\, \|\mathbf{x}\| \le 1\}$ and observe that $d_n({\cal A}_T)=\sigma_{n+1}$ since $A^T$ has the same singular values as $A$. The following result can be obtained by a similar argument as for \cref{thm:low-rank-1}. \begin{theorem}\label{thm:low-rank-2} Assume that the vectors $\mathbf{y}_i$, $i=1,\ldots,n$, are orthonormal, and define $\mathbf{x}_i:=A\mathbf{y}_i$, $i=1,\ldots,n$. If $\mathbb{Y}_n:={\rm span}\{\mathbf{y}_1,\ldots,\mathbf{y}_n\}$, then \begin{equation*} \|A-\sum_{i=1}^n\mathbf{x}_i\mathbf{y}_i^T\| = E({\cal A}_T,\mathbb{Y}_n), \end{equation*} and, consequently, the matrix $\sum_{i=1}^n\mathbf{x}_i\mathbf{y}_i^T$ is a best rank-$n$ approximation to $A$ if and only if the subspace $\mathbb{Y}_n$ is optimal for ${\cal A}_T$. \end{theorem} We remark that if $\mathbb{X}_n$ is an optimal subspace for ${\cal A}$ then it follows from the results of \cref{sec:seq-subspaces} that $A^T(\mathbb{X}_n)={\rm span}\{A^T\mathbf{x}_1,\ldots,A^T\mathbf{x}_n\}$ is an optimal space for ${\cal A}_T$. Thus, the $\mathbf{y}_i$, $i=1,\ldots,n$, in \cref{thm:low-rank-1} span an optimal space for ${\cal A}_T$ whenever $\mathbb{X}_n$ is optimal for ${\cal A}$. A similar observation holds for \cref{thm:low-rank-2} and we refer the reader to \cref{sec:seq-subspaces} for the details. The classical truncated SVD approximation to $A$ can be recovered by taking either $\mathbf{x}_i=\mathbf{u}_i$, $i=1,\ldots,n$, in \cref{thm:low-rank-1} or $\mathbf{y}_i=\mathbf{v}_i$, $i=1,\ldots,n$, in \cref{thm:low-rank-2}. From the above theorems we observe that a classification of all the optimal spaces for ${\cal A}$ and ${\cal A}_T$ leads to a classification of several best low-rank approximations to $A$. Such a classification is the goal of the remainder of this paper. Equivalence between best rank-$n$ approximation and optimality of the corresponding subspaces for the Kolmogorov $n$-width has been shown under the assumptions of either \cref{thm:low-rank-1} or \cref{thm:low-rank-2} (see also \cref{pro:n=1_symmetry}). It is an open question whether this equivalence holds more generally. \section{Optimal subspaces}\label{sec:optimal-subspaces} Let us start searching for optimal subspaces for ${\cal A}$. From now on we assume that the singular values of $A=U \Sigma V^T$ satisfy \cref{eq:sigmas-unique-weak}. Here we recall some optimality conditions from Karlovitz \cite{Karlovitz:76}. The following condition is necessary for the optimality of a subspace; see \cite[Theorem~1]{Karlovitz:76} for a proof. \begin{theorem}\label{thm:Karlovitz-nec} Given $n<r$, if $\mathbb{X}_n$ is an optimal subspace for ${\cal A}$, then $\mathbb{X}_n \perp \mathbf{u}_{n+1}$. \end{theorem} As mentioned in the introduction, under the assumption \cref{eq:sigmas-unique-weak} the left singular vector $\mathbf{u}_{n+1}$ is unique (up to multiplication by constants). In general, if there are multiple equal singular values for $A$, then an optimal subspace $\mathbb{X}_n$ must be orthogonal to a certain subspace spanned by the left singular vectors of $A$; see \cite[Theorem~1]{Karlovitz:76} for the details. Note that in the special case $r=m$ and $n=m-1$, \cref{thm:Karlovitz-nec} implies the uniqueness of the optimality of \begin{equation}\label{eq:optimal:N-1} \mathbb{X}_{m-1}={\rm span}\{\mathbf{u}_1,\ldots,\mathbf{u}_{m-1}\}. \end{equation} In addition to the necessary condition in \cref{thm:Karlovitz-nec}, Karlovitz also proved a sufficient condition for optimality. Roughly speaking, it states that any subspace ``sufficiently close'' to the optimal space ${\rm span}\{\mathbf{u}_1,\ldots\mathbf{u}_n\}$ must be optimal whenever it satisfies the necessary condition of \cref{thm:Karlovitz-nec}. The precise condition is stated in the following theorem; see \cite[Theorem~1]{Karlovitz:76} for a proof. \begin{theorem}\label{thm:Karlovitz-suff} Given $n < \min\{m-1,r\}$, if $\mathbb{X}_n\perp \mathbf{u}_{n+1}$ and \begin{equation}\label{ineq:Karlovitz} \sum_{i=1}^n\|\mathbf{u}_i-P_n\mathbf{u}_i\|^2\sigma^2_i\leq \sigma^2_{n+1}-\sigma^2_{n+2}, \end{equation} where $P_n$ is the orthogonal projection onto $\mathbb{X}_n$, then $\mathbb{X}_n$ is an optimal subspace for ${\cal A}$. \end{theorem} \section{Optimality criteria}\label{sec:optimality-crit} With the aim of deriving novel conditions for optimality of subspaces, we first provide a characterization of the distance $E({\cal A},\mathbb{X}_n)$. \begin{lemma}\label{lem:Eisaneigenvalue} Let $P_n$ be the orthogonal projection onto $\mathbb{X}_n$. The distance $E({\cal A},\mathbb{X}_n)$ is equal to the square root of the largest eigenvalue of \begin{equation}\label{eq:B} \Sigma^2 - \Sigma U^T P_n U\Sigma. \end{equation} \end{lemma} \begin{proof} First note that $P_n = P_n^2 = P_n^T P_n$. Similar to \cite[Theorem~2.3]{Melkman:78}, by using \cref{eq:E} and the definition of ${\cal A}$ we deduce that \begin{align*} E({\cal A},\mathbb{X}_n)^2 &= \max_{\|\mathbf{x}\|\leq 1}\|A\mathbf{x}-P_nA\mathbf{x}\|^2 = \max_{\mathbf{x}\neq 0}\frac{((I-P_n)A\mathbf{x},(I-P_n)A\mathbf{x})}{(\mathbf{x},\mathbf{x})} \\ &=\max_{\mathbf{x}\neq 0}\frac{(A^T(I-P_n)A\mathbf{x},\mathbf{x})}{(\mathbf{x},\mathbf{x})}, \end{align*} and so $E({\cal A},\mathbb{X}_n)$ is the square root of the largest eigenvalue of $M:=A^T(I-P_n)A$. From the SVD of $A$ we see that $M = VBV^T$, where $B:=\Sigma^2 - \Sigma U^T P_n U\Sigma$ is the matrix in \cref{eq:B}. Since $B$ is a similarity transformation of $M$, they share the same eigenvalues. \end{proof} The characterization of $E({\cal A},\mathbb{X}_n)$ in \cref{lem:Eisaneigenvalue} forms the basis for our optimality criteria. Let \begin{equation*} C_{n+1} := \sigma_{n+1}^2 I - \Sigma^2 + \Sigma U^T P_n U \Sigma, \end{equation*} and let $C_{n+1}[i_1,\ldots,i_k]$ denote the $k \times k$ submatrix of $C_{n+1}$ consisting of the rows and columns with indices $i_1,\ldots,i_k$. \begin{lemma}\label{lem:optimality_criterion} The subspace $\mathbb{X}_n$ is optimal for ${\cal A}$ if and only if $\mathbb{X}_n \perp \mathbf{u}_{n+1}$ and $C_{n+1}[1,\ldots,n,n+2,\ldots,m]$ is positive semi-definite. \end{lemma} \begin{proof} Suppose $\mathbb{X}_n$ is optimal for ${\cal A}$. Then, from \cref{eq:width=sv} we deduce that $E({\cal A},\mathbb{X}_n)= \sigma_{n+1}$, and by \cref{lem:Eisaneigenvalue} we have that $C_{n+1}$ is positive semi-definite. Conversely, if $C_{n+1}$ is positive semi-definite, then using again the same lemma we can conclude that $\mathbb{X}_n$ is optimal for ${\cal A}$. Moreover, by \cref{thm:Karlovitz-nec}, $\mathbb{X}_n \perp \mathbf{u}_{n+1}$ and the $(n+1)$-st row and $(n+1)$-st column of $C_{n+1}$ are zero, and so $C_{n+1}$ is positive semi-definite if and only if $C_{n+1}[1,\ldots,n,n+2,\ldots,m]$ is positive semi-definite. \end{proof} \begin{proposition}\label{prop:optimality_criterion2} The subspace $\mathbb{X}_n$ is optimal for ${\cal A}$ if and only if $\mathbb{X}_n \perp \mathbf{u}_{n+1}$ and \begin{equation}\label{eq:mycond} \det(C_{n+1}[{\cal J}]) \ge 0, \end{equation} for any set of indices ${\cal J}\subseteq \{1,\ldots,n,n+2,\ldots,m\}$ such that $\{1,\ldots,n\} \cap {\cal J} \ne \emptyset$. \end{proposition} \begin{proof} By the previous lemma, $\mathbb{X}_n$ is optimal for ${\cal A}$ if and only if $\mathbb{X}_n \perp \mathbf{u}_{n+1}$ and the matrix $C_{n+1}[1,\ldots,n,n+2,\ldots,m]$ is positive semi-definite. The latter is equivalent to the two conditions \begin{equation}\label{eq:condother} \det(C_{n+1}[{\cal J}]) \ge 0, \quad {\cal J}\subseteq \{n+2,\ldots,m\}, \end{equation} and \cref{eq:mycond}. Thus, to complete the proof it is sufficient to show that \cref{eq:condother} holds for all $\mathbb{X}_n$, i.e., that $C_{n+1}[n+2,n+3,\ldots,m]$ is positive semi-definite for any $\mathbb{X}_n$. To see this, let $\mathbf{x} = [0,\ldots,0,x_{n+2},\ldots,x_m]^T \in \mathbb{R}^m$. Then, noting that $P_n = P_n^2 = P_n^T P_n$, \begin{equation}\label{eq:posdef} \mathbf{x}^T C_{n+1} \mathbf{x} = \sum_{i=n+2}^m (\sigma_{n+1}^2 - \sigma_i^2) x_i^2 + \|P_n U \Sigma \mathbf{x} \|^2 \ge 0, \end{equation} and thus $C_{n+1}[n+2,n+3,\ldots,m]$ is indeed positive semi-definite. \end{proof} Alternatively, we can consider a sufficient condition for optimality that involves checking the sign of only $n$ determinants. \begin{corollary}\label{thm:optimality_sufficiency} The subspace $\mathbb{X}_n$ is optimal for ${\cal A}$ if $\mathbb{X}_n \perp \mathbf{u}_{n+1}$ and \begin{equation}\label{eq:mycondstrict} \det(C_{n+1}[k,k+1,\ldots,n,n+2,\ldots,m]) > 0, \quad k=1,2,\ldots,n. \end{equation} \end{corollary} \begin{proof} The subspace $\mathbb{X}_n$ is optimal for ${\cal A}$ if $\mathbb{X}_n \perp \mathbf{u}_{n+1}$ and the matrix \linebreak $C_{n+1}[1,\ldots,n,n+2,\ldots,m]$ is positive definite. The latter is equivalent to the two conditions \begin{equation}\label{eq:condotherstrict} \det(C_{n+1}[k,k+1,\ldots,m]) > 0, \quad k= n+2,\ldots,m, \end{equation} and \cref{eq:mycondstrict}. But \cref{eq:condotherstrict} holds for any $\mathbb{X}_n$ since inequality \cref{eq:posdef} is strict unless $x_{n+2} = x_{n+3} = \cdots = x_m = 0$. \end{proof} Let us now express the subspace $\mathbb{X}_n$ in the form \begin{equation}\label{eq:X_W} \mathbb{X}_n={\rm span}\{\mathbf{x}_1,\ldots,\mathbf{x}_n\}, \end{equation} where $\mathbf{x}_1,\ldots,\mathbf{x}_n$ are orthonormal vectors in $\mathbb{R}^m$. Then, the projection $P_n$ equals $XX^T$ where $X\in\mathbb{R}^{m,n}$ is the matrix whose columns are $\mathbf{x}_1,\ldots,\mathbf{x}_n$. We can express these vectors in the basis $\mathbf{u}_1,\ldots,\mathbf{u}_m$, and write \begin{equation* \mathbf{x}_j = \sum_{i=1}^m w_{ij} \mathbf{u}_i, \quad j=1,\ldots,n, \end{equation*} for coefficients $w_{ij} \in \mathbb{R}$. Letting $W\in\mathbb{R}^{m,n}$ be the matrix $[w_{ij}]_{i=1,\ldots,m,j=1,\ldots,n}$, we find that $$ X = U W, $$ and it follows that $$ P_n = U W W^T U^T, $$ and therefore, that \begin{equation}\label{eq:C_W} C_{n+1} = \sigma_{n+1}^2 I - \Sigma^2 + \Sigma W W^T \Sigma. \end{equation} Note that $W = U^T X$, which implies $$ W^T W = X^T X = I, $$ and so the columns $\mathbf{w}_1,\ldots,\mathbf{w}_n$ of $W$ are orthonormal. We can then further sharpen the condition of \cref{prop:optimality_criterion2} by making use of a matrix determinant identity. \begin{lemma}\label{lem:det-formula} Suppose $\mathbb{X}_n$ is as in \cref{eq:X_W} and $C_{n+1}$ as in \cref{eq:C_W}. If ${\cal J}\subseteq \{1,\ldots,n,n+2,\ldots,m\}$ is any set of indices, then \begin{align*} \det(C_{n+1}[{\cal J}]) = \det(M_{\cal J})\prod_{k\in {\cal J}}(\sigma_{n+1}^2-\sigma_k^2), \end{align*} where $M_{\cal J} = [m_{ij}]_{i,j=1,\ldots,n}$ has the elements \begin{align}\label{eq:M-mat} m_{ij} = \sigma_{n+1}^2\sum_{k\in {\cal J}}\frac{w_{ki} w_{kj}} {\sigma_{n+1}^2-\sigma_k^2}+ \sum_{k\not\in {\cal J}}w_{ki} w_{kj}. \end{align} \end{lemma} \begin{proof} Let ${\cal I}_k:=\{1,\ldots,k\}$, and let $W[{\cal J},{\cal I}_n]$ be the submatrix of $W$ consisting of the rows with indices in ${\cal J}$ and columns with indices in ${\cal I}_n$. We use the fact that for any non-singular matrix $F \in \mathbb{R}^{m,m}$ and any matrix $G \in \mathbb{R}^{m,n}$, it holds that \begin{equation*} \det(F+GG^T)=\det(I+G^T F^{-1} G)\det(F); \end{equation*} see~\cite[Theorem~18.1.1]{Harville:1997}. Applying this identity with \begin{equation*} F:= (\sigma_{n+1}^2 I - \Sigma^2)[{\cal J}],\quad G:= \Sigma[{\cal J}]W[{\cal J},{\cal I}_n], \end{equation*} we find that $$ \det(C_{n+1}[{\cal J}]) =\det(I+ W[{\cal J},{\cal I}_{n}]^TD W[{\cal J},{\cal I}_{n}])\det(F), $$ where $D:=\Sigma[{\cal J}] F^{-1}\Sigma[{\cal J}]$ is the diagonal matrix given by \begin{align*} D_{kk} = \frac{\sigma_k^2}{\sigma_{n+1}^2-\sigma_k^2}, \quad k\in {\cal J}. \end{align*} Moreover, we find that \begin{align*} I &= \left[\sum_{k\in {\cal I}_m}w_{ki} w_{kj}\right]_{i,j=1,\ldots,n}, \\ W[{\cal J},{\cal I}_{n}]^TD W[{\cal J},{\cal I}_{n}] &= \left[ \sum_{k\in {\cal J}}\frac{\sigma_k^2 w_{ki} w_{kj}} {\sigma_{n+1}^2-\sigma_k^2}\right]_{i,j=1,\ldots,n}, \end{align*} and therefore $M_{\cal J}=I+ W[{\cal J},{\cal I}_{n}]^TD W[{\cal J},{\cal I}_{n}]$ since $$ m_{ij} = \sum_{k\in {\cal J}} w_{ki} w_{kj} + \sum_{k\not\in {\cal J}} w_{ki} w_{kj} + \sum_{k\in {\cal J}}\frac{\sigma_k^2 w_{ki} w_{kj}} {\sigma_{n+1}^2-\sigma_k^2}. $$ Finally, since $F$ is diagonal, we have $$\det(F)= \prod_{k\in {\cal J}}(\sigma_{n+1}^2-\sigma_k^2),$$ and the result follows. \end{proof} \begin{theorem}\label{thm:optimality_criterion3} The subspace $\mathbb{X}_n$ is optimal for ${{\cal A}}$ if and only if $\mathbb{X}_n \perp \mathbf{u}_{n+1}$ and for all sets of indices ${\cal J}\subseteq \{1,\ldots,n,n+2,\ldots,m\}$ such that $\{1,\ldots,n\} \cap {\cal J} \ne \emptyset$ we have \begin{equation* (-1)^s \det(M_{\cal J}) \geq 0, \end{equation*} where $s$ is the cardinality of $\{1,\ldots,n\} \cap {\cal J}$ and $M_{\cal J}$ is the matrix given in \cref{eq:M-mat}. \end{theorem} \begin{proof} From \cref{prop:optimality_criterion2} we know that $\mathbb{X}_n$ is optimal for ${{\cal A}}$ if and only if $\mathbb{X}_n \perp \mathbf{u}_{n+1}$ and for all sets of indices ${\cal J}\subseteq \{1,\ldots,n,n+2,\ldots,m\}$ such that $\{1,\ldots,n\} \cap {\cal J} \ne \emptyset$, we have \begin{equation*} \det(C_{n+1}[J]) = \det(M_{\cal J})\prod_{k\in {\cal J}}(\sigma_{n+1}^2-\sigma_k^2) \ge 0. \end{equation*} Now, since the singular values satisfy \cref{eq:sigmas-unique-weak} we find that \begin{align*} (-1)^s \prod_{k\in {\cal J}}(\sigma_{n+1}^2-\sigma_k^2) > 0, \end{align*} which gives the result. \end{proof} There is a freedom in the choice of the basis $\mathbf{x}_1,\ldots,\mathbf{x}_n$ for the space $\mathbb{X}_n$ in \cref{eq:X_W}, and this freedom will affect the matrices in the above optimality criterion. Looking at the sufficient condition in \cref{thm:Karlovitz-suff}, a natural candidate for a basis of $\mathbb{X}_n$ seems to be $P_n\mathbf{u}_1,\ldots,P_n\mathbf{u}_n$, as long as they are linearly independent. If they are, then they can be orthonormalized by a Gram--Schmidt process before being used in \cref{thm:optimality_criterion3}. Let us now prove that $P_n\mathbf{u}_1,\ldots,P_n\mathbf{u}_n$ are in fact linearly independent whenever $\mathbb{X}_n$ is optimal. \begin{proposition} Let $P_n$ be the orthogonal projection onto $\mathbb{X}_n$. If $\mathbb{X}_n$ is optimal for ${\cal A}$, then $P_n \mathbf{u}_1,\ldots,P_n \mathbf{u}_n$ are linearly independent. \end{proposition} \begin{proof} Suppose, on the contrary, that there are coefficients $c_1,\ldots,c_n \in \mathbb{R}$, not all zero, such that $$ \sum_{i=1}^n c_i P_n \mathbf{u}_i = 0. $$ Then, $$ P_n \left(\sum_{i=1}^n c_i \mathbf{u}_i\right) = 0, $$ which we can write as $$ P_n U \mathbf{c} = 0, $$ where $\mathbf{c} = [c_1,\ldots,c_n,0,\ldots,0]^T \in \mathbb{R}^m$. Let $\mathbf{y}\in \mathbb{R}^m$ be such that $ \Sigma\mathbf{y}= \mathbf{c}$. Then, $$ P_n U \Sigma \mathbf{y} = 0, $$ and therefore, $$ \mathbf{y}^T (\Sigma^2 - \Sigma U^T P_n U \Sigma) \mathbf{y} = \mathbf{y}^T \Sigma^2 \mathbf{y}. $$ Since not all the coefficients $c_1,\ldots,c_n$ are zero, not all the coefficients $y_1,\ldots,y_n$ are zero. Therefore, we can form the Rayleigh quotient of $B:=\Sigma^2 - \Sigma U^T P_n U \Sigma$ and $\mathbf{y}$, and we find $$ \frac{\mathbf{y}^T (\Sigma^2 - \Sigma U^T P_n U \Sigma) \mathbf{y}}{\mathbf{y}^T \mathbf{y}} = \frac{\mathbf{y}^T \Sigma^2 \mathbf{y}}{\mathbf{y}^T \mathbf{y}} = \frac{\sum_{i=1}^n y_i^2 \sigma_i^2} {\sum_{i=1}^n y_i^2} \ge \sigma_n^2, $$ and so $E({\cal A},\mathbb{X}_n) \ge \sigma_n$ and $\mathbb{X}_n$ is not optimal for ${{\cal A}}$ (see \cref{lem:Eisaneigenvalue}). \end{proof} \section{Optimality for the $1$-width and best rank-$1$ approximation} \label{sec:1-width} For the $1$-width we can derive an explicit form of the optimality criterion in \cref{thm:optimality_criterion3}. Suppose $\mathbb{X}_1 = {\rm span}\{\mathbf{x}_1\}$ for some $\mathbf{x}_1 = \sum_{i=1}^mw_i\mathbf{u}_i \in \mathbb{R}^m$ with $\|\mathbf{x}_1\| = 1$. \begin{theorem}\label{thm:optimality_criterion_n=1} The subspace $\mathbb{X}_1$ is optimal for ${{\cal A}}$ if and only if $w_2 = 0$ and \begin{equation}\label{ineq:optimality_criterion_n=1} \sum_{i=3}^m \frac{w_i^2}{\sigma_2^2-\sigma_i^2}\leq \frac{w_1^2}{\sigma_1^2-\sigma_2^2}. \end{equation} \end{theorem} \begin{proof} Note that for $n=1$ the matrix $M_{\cal J}$ in \cref{eq:M-mat} is a scalar. Using \cref{thm:optimality_criterion3}, the subspace $\mathbb{X}_1$ is optimal if and only if $w_2 = 0$ and $$M_{\cal J} = \sigma_2^2\sum_{i\in {\cal J}}\frac{w_i^2}{\sigma_2^2-\sigma_i^2} + \sum_{i\not\in {\cal J}}w_i^2\leq 0, $$ for any subset ${\cal J}$ of $\{1,3,\ldots,m\}$ that contains $1$. Since $w_2 = 0$, this is equivalent to \begin{align}\label{ineq:optimality_criterion_J} \sigma_2^2\sum_{i\in {\cal J}}\frac{w_i^2}{\sigma_2^2-\sigma_i^2}+ \sum_{i\in {\cal K}}w_i^2\leq 0, \end{align} where ${\cal K} = \{1,3,\ldots,m\} \setminus J$. Now, if ${\cal J}=\{1,3,\ldots,m\}$, then ${\cal K} = \emptyset$ and \cref{ineq:optimality_criterion_J} is equivalent to \cref{ineq:optimality_criterion_n=1}. If, on the other hand, ${\cal J}$ is a strict subset of $\{1,3,\ldots,m\}$, then \begin{align*} \sigma_2^2\sum_{i\in {\cal J}}\frac{w_i^2}{\sigma_2^2-\sigma_i^2}+ \sum_{i\in {\cal K}}w_i^2\leq \sigma_2^2\sum_{i\in {\cal J}}\frac{w_i^2}{\sigma_2^2-\sigma_i^2}+ \sum_{i\in {\cal K}}\frac{\sigma_2^2}{\sigma_2^2-\sigma_i^2} w_i^2 = \sigma_2^2\sum_{\substack{i=1\\ i\neq 2}}^m\frac{w_i^2}{\sigma_2^2-\sigma_i^2} \leq 0, \end{align*} since $\sigma_2\geq \sigma_2-\sigma_j$ for any $j \in \{3,\ldots,m\}$. This concludes the proof. \end{proof} Observe that by combining the above result with either \cref{thm:low-rank-1} or \cref{thm:low-rank-2} we obtain a characterization of several best rank-$1$ approximations to $A$. We remark that a condition similar to \cref{ineq:optimality_criterion_n=1} was found by Antoulas \cite{Antoulas:1997} in the special case of rank-1 approximation to Hankel matrices. The optimality criterion in \cref{ineq:optimality_criterion_n=1} is trivially satisfied by the classical optimal space ${\rm span}{\{\mathbf{u}_1\}}$ and it provides a characterization of ``how far'' a one-dimensional space can deviate from ${\rm span}{\{\mathbf{u}_1\}}$ and still remain optimal. Specifically, let $\mathbf{x}_1 = \sum_{i=1}^mw_i\mathbf{u}_i \in \mathbb{R}^m$ with $\|\mathbf{x}_1\| = 1$, then \cref{thm:optimality_criterion_n=1} shows that if $\mathbb{X}_1= {\rm span}\{\mathbf{x}_1\}$ is optimal for ${\cal A}$ and $A\neq 0$, then $w_1\neq 0$. Indeed, if $w_1=0$, then from \cref{ineq:optimality_criterion_n=1} we have that $w_i=0$, $i=2,\ldots,m$ and so $\mathbf{x}_1=0$. The space $\mathbb{X}_1=\{0\}$ can only be optimal for the $1$-width of ${\cal A}$ if $\sigma_1=\sigma_2$, which contradicts assumption \cref{eq:sigmas-unique-weak}. Let us now compare the result in \cref{thm:optimality_criterion_n=1} with the sufficient condition of Karlovitz (\cref{thm:Karlovitz-suff}). Note that \cref{ineq:optimality_criterion_n=1} is equivalent to \begin{equation}\label{ineq:optimality_criterion2} \sum_{i=3}^m \frac{\sigma_1^2-\sigma_i^2}{\sigma_2^2-\sigma_i^2}w_i^2\leq 1, \end{equation} by using $w_1^2=1-\sum_{i=3}^mw_i^2$ and $w_2=0$. On the other hand, for $n=1$, the left-hand side of \cref{ineq:Karlovitz} equals \begin{equation*} \|\mathbf{u}_1-(\mathbf{u}_1,\mathbf{x}_1)\mathbf{x}_1\|^2\sigma_1^2 = (\|\mathbf{u}_1\|^2-(\mathbf{u}_1,\mathbf{x}_1)^2)\sigma_1^2 = (1-w_1^2)\sigma_1^2= \sum_{i=3}^mw_i^2\sigma_1^2, \end{equation*} and so, condition \cref{ineq:Karlovitz} is equivalent to \begin{equation}\label{ineq:Karlovitz_n=1} \sum_{i=3}^m\frac{\sigma_1^2}{\sigma_2^2-\sigma_3^2}w_i^2\leq 1. \end{equation} Since the singular values are decreasing, we have \begin{equation}\label{ineq:comparison_Karlovitz} \frac{\sigma_1^2-\sigma_i^2}{\sigma_2^2-\sigma_i^2}\leq \frac{\sigma_1^2}{\sigma_2^2-\sigma_3^2}, \quad i=3,\ldots,m, \end{equation} and condition \cref{ineq:Karlovitz_n=1} implies \cref{ineq:optimality_criterion2}, as expected. However, we note that the case $i=3$ in \cref{ineq:comparison_Karlovitz} is a strict inequality if $\sigma_3> 0$. Thus, for $n=1$, the sufficient condition in \cref{thm:Karlovitz-suff} is stronger than necessary whenever $\sigma_3>0$. \begin{example}\label{ex:n=1_3D} Let $m=3$ and consider the space $\mathbb{X}_1 = {\rm span}\{\mathbf{x}_1\}$ for some $\mathbf{x}_1 = w_1\mathbf{u}_1+w_2\mathbf{u}_2+w_3\mathbf{u}_3$, with $\|\mathbf{x}_1\| = 1$. From \cref{thm:optimality_criterion_n=1} it follows that $\mathbb{X}_1$ is optimal for ${\cal A}$ if and only if $w_2=0$ and \begin{align}\label{ineq:optimality-3D} w_3^2 \le \frac{\sigma_2^2 - \sigma_3^2}{\sigma_1^2 - \sigma_3^2}. \end{align} Now, let $w_1=\cos(\alpha)$, $w_2=0$ and $w_3=\sin(\alpha)$, where $\alpha$ is the angle between $\mathbb{X}_1$ and the classical optimal space ${\rm span}\{\mathbf{u}_1\}$. Condition \cref{ineq:optimality-3D} is then equivalent to \begin{align}\label{ineq:optimality-3D-angle} |\alpha|\leq \extreme{\alpha}:=\arcsin\left(\sqrt{\frac{\sigma_2^2 - \sigma_3^2}{\sigma_1^2 - \sigma_3^2}}\right). \end{align} Thus, $\mathbb{X}_1$ is optimal for ${\cal A}$ if and only if it is rotated in the $(\mathbf{u}_1,\mathbf{u}_3)$-plane with an angle less than or equal to $\extreme{\alpha}$ from the $\mathbf{u}_1$-axis. An illustration of this is given in \cref{fig:optimality-3D} for $\mathbf{u}_i=\mathbf{e}_i$, $i=1,2,3$. \end{example} \begin{figure} \center \begin{tikzpicture}[scale=2] \fill[black!10] (0,0) ellipse (2cm and 1cm); \draw[thick] (0,0) ellipse (2cm and 1cm); \draw[dashed,thick] (0,0)--(2,0) node[midway,above]{$\sigma_1$}; \draw[dashed,thick] (0,0)--(0,1) node[midway,left]{$\sigma_3$}; \draw[thick] (-2,-1)--(2,1) node[right,below]{$\mathbb{X}_1$}; \begin{scope} \path[clip] (0,0)--(2,1)--(2,0); \draw (0,0) circle (0.5cm); \node[above right] at (0.24,-0.02) {$\alpha$}; \end{scope} \end{tikzpicture} \caption{The $(\mathbf{e}_1,\mathbf{e}_3)$ cross-section of ${\cal A}$ in the case $m=3$. The space $\mathbb{X}_1$ is optimal for ${\cal A}$ if and only if $|\alpha|\leq\extreme{\alpha}$ in \cref{ineq:optimality-3D-angle}.}\label{fig:optimality-3D} \end{figure} \begin{example} \label{ex:n=1_Hankel_angle} Similar to an example in \cite{Antoulas:1997} we consider the $3\times3$ matrix \begin{equation*} A = \begin{bmatrix} 1 & 0 & 1/4 \\ 0 & 1/4 & 0 \\ 1/4 & 0 & 1 \end{bmatrix}. \end{equation*} Note that this is a symmetric matrix with Hankel structure. It is easy to verify that $A = U \Sigma U^T$, with \begin{equation*} \sigma_1 = \frac{5}{4},\ \mathbf{u}_1 = \frac{1}{\sqrt{2}}\begin{bmatrix} 1 \\ 0 \\ 1 \end{bmatrix}; \quad \sigma_2 = \frac{3}{4},\ \mathbf{u}_2 = \frac{1}{\sqrt{2}}\begin{bmatrix} 1 \\ 0 \\ -1 \end{bmatrix}; \quad \sigma_3 = \frac{1}{4},\ \mathbf{u}_3 = \begin{bmatrix} 0 \\ 1 \\ 0 \end{bmatrix}. \end{equation*} From \cref{ex:n=1_3D} we deduce that any space $\mathbb{X}_1= {\rm span}\{\mathbf{x}_1\}$ is optimal for ${\cal A}$ if and only if it is rotated in the $(\mathbf{u}_1,\mathbf{u}_3)$-plane with an angle less than or equal to \begin{equation*} \extreme{\alpha} = \arcsin\bigl(1/\sqrt{3}\bigr)\approx 35.26^\circ \end{equation*} from the $\mathbf{u}_1$-axis. The maximum angle $\extreme{\alpha}$ corresponds to the unit vector \begin{equation*} \extreme{\mathbf{x}}_1=\frac{\sqrt{2}\mathbf{u}_1+ \mathbf{u}_3}{\sqrt{3}} =\frac{1}{\sqrt{3}}\begin{bmatrix} 1 \\ 1 \\ 1 \end{bmatrix}, \end{equation*} which will be an interesting choice for structure-preserving approximation (see \cref{ex:n=1_Hankel_range}). \end{example} If $A$ is a symmetric matrix, then the low-rank approximations in \cref{thm:low-rank-1,thm:low-rank-2} do not, in general, result in a symmetric approximation to $A$. As we shall see in the next proposition, if given a proper choice of the scaling factor, then each unit vector satisfying the optimality criterion in \cref{thm:optimality_criterion_n=1} provides a symmetric best rank-$1$ approximation to a symmetric matrix $A$ (at least in the case $m=3$). We remark that the next result is very similar to \cite[Theorem 3.1]{Antoulas:1997}. Specifically, if $A$ is a Hankel matrix, then \cite[Theorem 3.1]{Antoulas:1997} provides a characterization of best rank-$1$ approximations to $A$ that preserve the Hankel structure. This characterization was later generalized to best rank-$1$ Hankel approximations to a symmetric matrix $A$ in \cite[Theorem 4.1]{Knirsch:preprint}. \begin{proposition} \label{pro:n=1_symmetry} Let $n=1$ and $m=3$. Let $A$ be a symmetric matrix and let $\mathbf{x}_1=\sum_{i=1}^3w_i\mathbf{u}_i$ be a unit vector such that $\mathbb{X}_1:={\rm span}\{\mathbf{x}_1\}$ is optimal for ${\cal A}$. Then, for any $\sigma_{\mathbf{x}_1}\in \mathbb{R}$ such that \begin{equation} \label{eq:range} \mathfrak{L}(\mathbf{x}_1):= \frac{(\sigma_1-\sigma_2)(\sigma_2-\sigma_3)}{(\sigma_2-\sigma_3)-(\sigma_1-\sigma_3)w_3^2} \leq \sigma_{\mathbf{x}_1} \leq \frac{(\sigma_1+\sigma_2)(\sigma_2+\sigma_3)}{(\sigma_2+\sigma_3)+(\sigma_1-\sigma_3)w_3^2}=:\mathfrak{U}(\mathbf{x}_1), \end{equation} we have \begin{equation}\label{eq:conjecture} \|A - \sigma_{\mathbf{x}_1} \mathbf{x}_1 \mathbf{x}_1^T\|=\sigma_2. \end{equation} \end{proposition} \begin{proof} Without loss of generality, we can restrict ourselves to the case of $A$ being a diagonal matrix $\Sigma$ and $\mathbf{u}_i=\mathbf{e}_i$, the elements of the canonical basis. Proving equality \cref{eq:conjecture} is equivalent to showing that the maximum modulus of the eigenvalues of the matrix $\Sigma - \sigma_{\mathbf{x}_1} \mathbf{x}_1 \mathbf{x}_1^T$ is equal to $\sigma_2$. Since $\mathbb{X}_1$ is optimal for ${\cal A}$, we know from \cref{thm:optimality_criterion_n=1} that $w_2=0$. Therefore, the eigenvalues of $\Sigma - \sigma_{\mathbf{x}_1} \mathbf{x}_1 \mathbf{x}_1^T$ are given by $\sigma_2$ and by the eigenvalues of the submatrix obtained by removing the second row and the second column, i.e., \begin{equation} \label{eq:Mw} \begin{bmatrix} \sigma_1-\sigma_{\mathbf{x}_1} w_1^2 & -\sigma_{\mathbf{x}_1} w_1 w_3 \\ -\sigma_{\mathbf{x}_1} w_1 w_3 & \sigma_3-\sigma_{\mathbf{x}_1} w_3^2 \end{bmatrix}. \end{equation} Then, proving equality \cref{eq:conjecture} is equivalent to showing that the eigenvalues of the matrix in \cref{eq:Mw} are less than or equal to $\sigma_2$ in modulus. A direct computation shows that its two (real) eigenvalues are given by $$ \lambda_{\pm}=\frac{\sigma_1+\sigma_3-\sigma_{\mathbf{x}_1} \pm\sqrt{(\sigma_1-\sigma_3-\sigma_{\mathbf{x}_1})^2+ 4w_3^2(\sigma_1-\sigma_3)\sigma_{\mathbf{x}_1}}}{2}. $$ Imposing $ -\sigma_2\leq \lambda_{\pm}\leq \sigma_2 $ results in the range \cref{eq:range} for $\sigma_{\mathbf{x}_1}$. \end{proof} Let $m=3$. Recall from \cref{ineq:optimality-3D}--\cref{ineq:optimality-3D-angle} that $\mathbb{X}_1$ is optimal for ${\cal A}$ if and only if $w_2=0$ and $$w_3^2 \le \sin^2(\extreme{\alpha}):=\frac{\sigma_2^2 - \sigma_3^2}{\sigma_1^2 - \sigma_3^2}.$$ Set $\extreme{\mathbf{x}}_1:=\cos(\extreme{\alpha})\mathbf{u}_1+\sin(\extreme{\alpha})\mathbf{u}_3$, then one can check that \begin{equation*} \mathfrak{L}(\mathbf{x}_1)\leq \mathfrak{L}(\extreme{\mathbf{x}}_1)=\sigma_1+\sigma_3= \mathfrak{U}(\extreme{\mathbf{x}}_1)\leq \mathfrak{U}(\mathbf{x}_1), \end{equation*} for any $\mathbf{x}_1$ such that its span is optimal for ${\cal A}$. Therefore, the range of values in \cref{eq:range} for the scaling factor $\sigma_{\mathbf{x}_1}$ is always non-empty. In particular, it always contains the value $\sigma_1+\sigma_3$. This means that there always exists at least one best low-rank approximation in any optimal space for ${\cal A}$ (with $m=3$ and $n=1$). The classical truncated SVD approximation to $A$ corresponds to $\mathbf{x}_1=\mathbf{u}_1$, and in this case we have $\sigma_1-\sigma_2\leq \sigma_{\mathbf{u}_1}\leq \sigma_1+\sigma_2$. This is in agreement with \cref{eq:no-unique-spectral}. \begin{example} \label{ex:n=1_Hankel_range} As a continuation of \cref{ex:n=1_Hankel_angle}, consider again the matrix \begin{equation}\label{eq:A_Hankel_3} A = \begin{bmatrix} 1 & 0 & 1/4 \\ 0 & 1/4 & 0 \\ 1/4 & 0 & 1 \end{bmatrix}. \end{equation} According to \cref{pro:n=1_symmetry}, any choice $\mathbf{x}_1=\cos(\alpha)\mathbf{u}_1+\sin(\alpha)\mathbf{u}_3$, with $|\alpha|\leq\extreme{\alpha}$, leads to a range of best rank-$1$ approximations to $A$ that are symmetric, i.e., \begin{equation*} \sigma_{\mathbf{x}_1} \mathbf{x}_1 \mathbf{x}_1^T = \frac{\sigma_{\mathbf{x}_1}}{2} \begin{bmatrix} \cos^2(\alpha) & \sqrt{2} \cos(\alpha) \sin(\alpha) & \cos^2(\alpha) \\ \sqrt{2} \cos(\alpha) \sin(\alpha) & 2 \sin^2(\alpha) & \sqrt{2} \cos(\alpha) \sin(\alpha) \\ \cos^2(\alpha) & \sqrt{2} \cos(\alpha) \sin(\alpha) & \cos^2(\alpha) \end{bmatrix}, \end{equation*} for any $\sigma_{\mathbf{x}_1}\in \mathbb{R}$ such that \begin{equation*} \frac{1}{2-4\sin^2(\alpha)} \leq \sigma_{\mathbf{x}_1} \leq \frac{2}{1+\sin^2(\alpha)}. \end{equation*} The specific choice $\extreme{\mathbf{x}}_1$, corresponding to the maximum angle $\extreme{\alpha}$, gives a best rank-$1$ approximation that even preserves the Hankel structure of $A$, i.e., \begin{equation*} \sigma_{\extreme{\mathbf{x}}_1} \extreme{\mathbf{x}}_1 \extreme{\mathbf{x}}_1^T = \frac{1}{2} \begin{bmatrix} 1 & 1 & 1 \\ 1 & 1 & 1 \\ 1 & 1 & 1 \end{bmatrix}, \end{equation*} since $\sigma_{\extreme{\mathbf{x}}_1}=3/2$. Similarly, the approximation obtained by taking the angle $-\extreme{\alpha}$ preserves the Hankel structure as well. According to \cite[Theorem~3.1]{Antoulas:1997}, these matrices are the only two Hankel-preserving best rank-$1$ approximations to $A$ in \cref{eq:A_Hankel_3}. \end{example} As shown in \cite{Antoulas:1997}, it is not always possible to find a Hankel-preserving best rank-$1$ approximation to a Hankel matrix $A$. When this is not possible one can ask the question of how well one can approximate $A$ with rank-$1$ Hankel matrices, and this has been studied in \cite{Knirsch:preprint}. \section{Alternative optimality criteria}\label{sec:optimality-crit-alt} In this section we provide some alternative optimality criteria that are useful in the case of large $n$. While this is not relevant for low-rank approximation, these results are still of independent interest for the Kolmogorov $n$-width. To simplify the exposition, we will in this section only consider matrices $A$ that are of full rank, i.e., $r=m$. Recall that a necessary condition for an $n$-dimensional space $\mathbb{X}_n$ to be optimal for the $n$-width is that it is orthogonal to $\mathbf{u}_{n+1}$ (see \cref{thm:Karlovitz-nec}). This implies that the only optimal space for $n=m-1$ is given in \cref{eq:optimal:N-1}. Suppose now that $n \le m-2$ and that $\mathbb{X}_n$ is orthogonal to $\mathbf{u}_{n+1}$. Let us denote the orthogonal complement of $\mathbb{X}_n \oplus \mathbf{u}_{n+1}$ in $\mathbb{R}^m$ by $\mathbb{Y}_{m-n-1}$, and suppose that we can represent it in the form \begin{equation* \mathbb{Y}_{m-n-1}={\rm span}\{\mathbf{y}_1,\ldots,\mathbf{y}_{m-n-1}\}, \end{equation*} where $\mathbf{y}_1,\ldots,\mathbf{y}_{m-n-1}$ are orthonormal vectors in $\mathbb{R}^m$. We can express these vectors as \begin{equation* \mathbf{y}_j = \sum_{i=1}^m q_{ij} \mathbf{u}_i, \quad j=1,\ldots,m-n-1, \end{equation*} for coefficients $q_{ij} \in \mathbb{R}$, $j=1,\ldots,m-n-1$, where now $$ q_{n+1,j} = 0, \quad j=1,\ldots, m-n-1. $$ Denoting by $Q\in\mathbb{R}^{m,m-n-1}$ the matrix \begin{equation} \label{eq:W-alt} [q_{ij}]_{i=1,\ldots,m,j=1,\ldots,m-n-1}, \end{equation} we obtain the following alternative characterization of optimality for $\mathbb{X}_n$. \begin{lemma}\label{lem:Eisaneigenvalue_alt_reduce} Let $Q$ be the matrix in \cref{eq:W-alt}. The subspace $\mathbb{X}_n$ is optimal for ${{\cal A}}$ if and only if $\mathbb{X}_n \perp \mathbf{u}_{n+1}$ and the largest eigenvalue of \begin{equation* Q^T \Sigma^2 Q \end{equation*} is at most $\sigma_{n+1}^2$. \end{lemma} \begin{proof} Recall from \cref{eq:max} that $$ E({\cal A}, \mathbb{X}_n) = \max_{\mathbf{z} \perp \mathbb{X}_n} \sqrt{\frac{\mathbf{z}^TAA^T\mathbf{z}}{\mathbf{z}^T\mathbf{z}}}. $$ Following the argument of Karlovitz in \cite[Theorem~1]{Karlovitz:76}, any $\mathbf{z}$ orthogonal to $\mathbb{X}_n$ can be expressed uniquely as $\mathbf{z} = \mathbf{y} \oplus \mathbf{x}$, where $\mathbf{y} \in \mathbb{Y}_{m-n-1}$ and $\mathbf{x} \in {\rm span}\{\mathbf{u}_{n+1}\}$. Then, \begin{equation*} \frac{\mathbf{z}^T A A^T \mathbf{z}}{\mathbf{z}^T \mathbf{z}} = \frac{\mathbf{x}^T \mathbf{x} \sigma_{n+1}^2 + \mathbf{y}^T A A^T \mathbf{y}} {\mathbf{x}^T \mathbf{x} + \mathbf{y}^T \mathbf{y}} \le \max \left\{ \sigma_{n+1}^2, \frac{\mathbf{y}^T A A^T \mathbf{y}} {\mathbf{y}^T \mathbf{y}} \right\}, \end{equation*} since it is a convex combination of $\sigma_{n+1}^2$ and $ \frac{\mathbf{y}^T A A^T \mathbf{y}} {\mathbf{y}^T \mathbf{y}}$. We conclude that \begin{equation*} E({\cal A}, \mathbb{X}_n) = \max \left\{ \sigma_{n+1}, \max_{\mathbf{y} \in \mathbb{Y}_{m-n-1}} \sqrt{\frac{\mathbf{y}^TAA^T\mathbf{y}}{\mathbf{y}^T\mathbf{y}}} \right\}. \end{equation*} Any $\mathbf{y} \in \mathbb{Y}_{m-n-1}$ can be represented as $$ \mathbf{y} = \sum_{j=1}^{m-n-1} c_j \mathbf{y}_j, $$ for coefficients $c_1,\ldots,c_{m-n-1}$. Setting $\mathbf{c} := [c_1,\ldots,c_{m-n-1}]^T$, we have $$ \mathbf{y}^T \mathbf{y} = \sum_{j=1}^{m-n-1} c_j^2 = \mathbf{c}^T \mathbf{c}. $$ Setting $Y := [\mathbf{y}_1,\ldots,\mathbf{y}_{m-n-1}]$, we also have $$ \mathbf{y} = Y \mathbf{c} = U Q \mathbf{c}, $$ and so \begin{equation*} \mathbf{y}^T A A^T \mathbf{y} = \mathbf{c}^T Q^T U^T A A^T U Q \mathbf{c} = \mathbf{c}^T Q^T \Sigma^2 Q \mathbf{c}. \end{equation*} Therefore, \begin{equation*} \max_{\mathbf{y} \in \mathbb{Y}_{m-n-1}} \frac{\mathbf{y}^T A A^T \mathbf{y}}{\mathbf{y}^T \mathbf{y}} = \max_{\mathbf{c} \in \mathbb{R}^{m-n-1}} \frac{\mathbf{c}^T Q^T \Sigma^2 Q\mathbf{c}}{\mathbf{c}^T\mathbf{c}}, \end{equation*} which is the largest eigenvalue of $Q^T \Sigma^2 Q$. \end{proof} Suppose now that $n = m-2$ and that $\mathbb{X}_{m-2}$ is orthogonal to $\mathbf{u}_{m-1}$. Let $\mathbb{Y}_1$ be the orthogonal complement to $\mathbb{X}_{m-2} \oplus \mathbf{u}_{m-1}$ in $\mathbb{R}^m$. Let $\mathbf{y}_1$ be a unit vector in $\mathbb{Y}_1$ (which is unique up to a change of sign). We can express $\mathbf{y}_1$ in the basis $\mathbf{u}_1,\ldots,\mathbf{u}_m$, and write \begin{equation* \mathbf{y}_1 = \sum_{i=1}^m q_i \mathbf{u}_i, \end{equation*} for coefficients $q_1,\ldots,q_m \in \mathbb{R}$ such that $\sum_{i=1}^m q_i^2 = 1$ and $q_{m-1} = 0$. \begin{theorem}\label{thm:N-2} The subspace $\mathbb{X}_{m-2}$ is optimal if and only if $q_{m-1}=0$ and $$ \sum_{\substack{i=1 \\ i \ne m-1}}^m q_i^2 \sigma_i^2 \le \sigma_{m-1}^2. $$ \end{theorem} \begin{proof} This is just an application of \cref{lem:Eisaneigenvalue_alt_reduce} for $n=m-2$, in which case the matrix $Q^T\Sigma^2 Q$ has the single element \begin{align*} \sum_{\substack{i=1 \\ i \ne m-1}}^m q_i^2 \sigma_i^2. \end{align*} \end{proof} \begin{example}\label{ex:n=1_3D_alt} Let $m=3$ and let $\mathbb{X}_1$ be a $1$-dimensional subspace of $\mathbb{R}^3$ that is orthogonal to $\mathbf{u}_2$, and let $\mathbf{y}_1 = q_1 \mathbf{u}_1 + q_3 \mathbf{u}_3$ be a unit vector orthogonal to $\mathbb{X}_1$. From \cref{thm:N-2} it follows that $\mathbb{X}_1$ is optimal for ${\cal A}$ if and only if $$ q_1^2 \sigma_1^2 + q_3^2 \sigma_3^2 \le \sigma_2^2. $$ If $$ \mathbb{X}_1 = {\rm span}\{\cos(\alpha)\mathbf{u}_1+\sin(\alpha)\mathbf{u}_3\}, $$ then $$ \mathbb{Y}_1 = {\rm span}\{-\sin(\alpha)\mathbf{u}_1+\cos(\alpha)\mathbf{u}_3\}, $$ and $(q_1,q_3) = \pm (-\sin(\alpha),\cos(\alpha))$, thus the optimality condition can be expressed as $$ \sin^2(\alpha) \le \frac{\sigma_2^2 - \sigma_3^2} {\sigma_1^2 - \sigma_3^2}. $$ This agrees with \cref{ex:n=1_3D}. \end{example} \section{Totally positive matrices}\label{sec:tp-matrices} Melkman and Micchelli studied the $n$-width problem for a certain class of matrices, and in this section we compare their results with the optimality criteria in \cref{sec:optimality-crit,sec:1-width}. If $A$ is strictly totally positive, i.e., all its minors are positive, then two optimal spaces for ${\cal A}$ are constructed in \cite[Section~4]{Melkman:78}. These two spaces are in general different from the classical optimal space ${\rm span}\{\mathbf{u}_1,\ldots,\mathbf{u}_n\}$. We will describe the first of these optimal spaces here. The second will be discussed in the next section. When $A$ is strictly totally positive it follows from a theorem of Gantmacher and Krein \cite{Gantmacher:2002} that the singular values are positive and distinct, $$\sigma_1>\sigma_2>\dots>\sigma_m>0,$$ and the right singular vectors of $A$ have the following sign properties, \begin{equation}\label{eq:STP-sign} S^+(\mathbf{v}_{n+1})=S^-(\mathbf{v}_{n+1})=n, \quad n=0,\ldots,m-1. \end{equation} Here $S^-(\mathbf{v})$ denotes the actual sign changes of the vector $\mathbf{v}$, where zero components are discarded and $S^+(\mathbf{v})$ is the maximum number of sign changes obtainable by adding $1$ or $-1$ to the zero components of $\mathbf{v}$. It follows from \cref{eq:STP-sign} that $v_{n+1,1}v_{n+1,m}\neq 0$ and we can assume, without loss of generality, that $v_{n+1,1}>0$. Moreover, using \cref{eq:STP-sign}, there exist indices $0=\ell_0<\ell_1<\dots<\ell_n<\ell_{n+1}=m$, denoting the sign changes in $\mathbf{v}_{n+1}$, i.e, such that $$ v_{n+1,i}(-1)^j\geq 0, \quad \ell_j<i\leq \ell_{j+1},\quad j=0,1,\ldots,n. $$ To simplify the exposition, let us assume that the vector $\mathbf{v}_{n+1}$ has no zero components; see \cite[Section~4]{Melkman:78} for the general case. The index $\ell_j$ is then the index before the sign change, i.e., such that $v_{n+1,\ell_j}v_{n+1,\ell_j+1}<0$. For each $j=1,2,\ldots,n$, define the $m$-dimensional vector $\mathbf{s}_j$ by \begin{equation*} s_{j,k}:=\begin{cases} 1/|v_{n+1,k}|, &k=\ell_j,\ell_{j}+1, \\ 0, &\text{otherwise}. \end{cases} \end{equation*} Then, $\mathbf{s}_j\perp\mathbf{v}_{n+1}$ for each $j=1,\ldots,n$, and Melkman and Micchelli proved the following result \cite[Theorem~3.1]{Melkman:78}. \begin{theorem}\label{thm:Melkman} If $A$ is a strictly totally positive matrix, then \begin{equation}\label{eq:Melkman1} \mathbb{X}_n^1:={\rm span}\{A\mathbf{s}_1,\ldots,A\mathbf{s}_n\} \end{equation} is an optimal subspace for ${\cal A}:=\{A\mathbf{x}: \|\mathbf{x}\|\leq 1\}$. \end{theorem} As a consequence of the above result, if we use a Gram--Schmidt process to find an orthonormal basis for $\mathbb{X}_n^1$, then we immediately obtain a best rank-$n$ approximation to $A$ by applying \cref{thm:low-rank-1}. Note that the space $\mathbb{X}_n^1$ in \cref{eq:Melkman1} satisfies the necessary condition $\mathbb{X}_n^1\perp\mathbf{u}_{n+1}$ (see \cref{thm:Karlovitz-nec}) since $\mathbf{s}_j\perp\mathbf{v}_{n+1}$ for each $j=1,\ldots,n$. \begin{example} Consider the case $n=1$ and $m=3$. In view of \cref{thm:optimality_criterion_n=1} and \cref{ex:n=1_3D} it would be interesting to check how far the optimal subspace in \cref{eq:Melkman1} is from the classical space ${\rm span}\{\mathbf{u}_1\}$ for different choices of $A$. Let us take what is perhaps one of the simplest possible choices of a strictly totally positive matrix, the Vandermonde matrix obtained by interpolating at the points $1,2,3$: \begin{equation* A= \begin{bmatrix} 1 & 1 & 1 \\ 1 & 2 & 4 \\ 1 & 3 & 9 \end{bmatrix}. \end{equation*} In this case, it can be checked that the angle between \cref{eq:Melkman1} and the space spanned by $\mathbf{u}_1$ is less than $0.171^\circ$, while the maximum angle for an optimal space as in \cref{ex:n=1_3D} is greater than $6.695^\circ$. \end{example} \section{Sequence of optimal subspaces}\label{sec:seq-subspaces} In \cref{thm:optimality_criterion3} we obtained an equivalent condition for optimality that allowed us to classify all optimal spaces of dimension $n=1$ for any matrix $A$ in \cref{thm:optimality_criterion_n=1}. However, as $n$ increases it becomes trickier to apply the optimality criterion in \cref{thm:optimality_criterion3} for an arbitrary matrix $A$. On the other hand, as we saw in the last section, there exist matrices where one can obtain an optimal $n$-dimensional space for ${\cal A}$ using specific properties of the matrix $A$. In this section we prove that, given some initial optimal space $\mathbb{X}_n^1$, we can obtain a whole sequence of optimal spaces $\mathbb{X}_n^p$, $p\geq 1$. Moreover, this sequence converges to the classical optimal space as $p\to\infty$. The arguments here hold for any matrix $A$ and are based on those found in \cite{Floater:2017,Floater:2018,Sande:2019} for an integral operator in $L^2$. Let $\mathbb{X}_n^1$ and $\mathbb{Y}_n^1$ be any $n$-dimensional subspaces of $\mathbb{R}^m$, and define the sequence of subspaces $\mathbb{X}_n^p$ and $\mathbb{Y}_n^p$ by \begin{equation}\label{eq:seq} \mathbb{X}_n^{p}:=A(\mathbb{Y}_n^{p-1}),\quad \mathbb{Y}_n^{p}:=A^T(\mathbb{X}_n^{p-1}), \quad p=2,3,\ldots. \end{equation} Then, similar to \cite[Lemma~1]{Floater:2017}, we have the following lemma. \begin{lemma}\label{lem:lifting} For any matrix $A$ and any subspaces $\mathbb{X}_n^1$ and $\mathbb{Y}_n^1$, we have \begin{align*} E({\cal A}, \mathbb{X}_n^p) &\le E({\cal A}_T, \mathbb{Y}_n^{p-1}), \\ E({\cal A}_T, \mathbb{Y}_n^p) &\le E({\cal A}, \mathbb{X}_n^{p-1}), \end{align*} for all $p\geq 2$. \end{lemma} \begin{proof} The two inequalities are analogous and so we only prove the last one. Let $P_n$ be the orthogonal projection onto $\mathbb{X}_n^{p-1}$. Then, the image of $A^TP_n$ is $\mathbb{Y}_n^p=A^T(\mathbb{X}_n^{p-1})$ and so \begin{align*} E({\cal A}_T,\mathbb{Y}_n^p) \leq \max_{\|\mathbf{x}\|\leq1} \|(A^T-A^T P_n)\mathbf{x}\| = \max_{\|\mathbf{x}\|\leq1} \|(A-P_n A)\mathbf{x}\| = E({\cal A},\mathbb{X}_n^{p-1}). \end{align*} \end{proof} Since $d_n({\cal A})=d_n({\cal A}_T)=\sigma_{n+1}$, we can apply \cref{lem:lifting} in an induction argument on $p$ to obtain the following theorem. \begin{theorem}\label{thm:optimal-seq} Suppose the subspace $\mathbb{X}_n^1$ is optimal for ${\cal A}$ and $\mathbb{Y}_n^1$ is optimal for ${\cal A}_T$. Then, \begin{itemize} \item the subspaces $\mathbb{X}_n^p$ in \cref{eq:seq} are optimal for ${\cal A}$, and \item the subspaces $\mathbb{Y}_n^p$ in \cref{eq:seq} are optimal for ${\cal A}_T$, \end{itemize} for all $p\geq 2$. \end{theorem} \begin{proof} Assume $\mathbb{X}_n^{p-1}$ is optimal for ${\cal A}$ and $\mathbb{Y}_n^{p-1}$ is optimal for ${\cal A}_T$. Then, using \cref{lem:lifting}, we have \begin{align*} E({\cal A}, \mathbb{X}_n^p) &\le E({\cal A}_T, \mathbb{Y}_n^{p-1})=d_n({\cal A}_T)=d_n({\cal A}), \\ E({\cal A}_T, \mathbb{Y}_n^p) &\le E({\cal A}, \mathbb{X}_n^{p-1})=d_n({\cal A})=d_n({\cal A}_T), \end{align*} and so $\mathbb{X}_n^{p}$ is optimal for ${\cal A}$ and $\mathbb{Y}_n^{p}$ is optimal for ${\cal A}_T$. The result now follows from induction on $p$. \end{proof} Note that for $p\geq 2$, the spaces $\mathbb{X}_n^p$ and $\mathbb{Y}_n^p$ could in general have dimension less than $n$, but they are still optimal for the $n$-width problem whenever $\mathbb{X}_n^1$ and $\mathbb{Y}_n^1$ are optimal. In fact, if $\mathbb{X}_n^p$ has dimension $k$, $0\leq k<n$, then $d_k({\cal A})$ must equal $d_n({\cal A})$ by definition of the $n$-width. \begin{example} Let $A$ be a strictly totally positive matrix. Then, by definition, $A^T$ is also strictly totally positive, and if we construct the vectors $\mathbf{t}_j$, $j=1,\ldots,n$, in a way analogous to the $\mathbf{s}_j$ in the previous section, it follows from \cref{thm:Melkman} that \begin{equation*} \mathbb{Y}_n^1:={\rm span}\{A^T\mathbf{t}_1,\ldots,A^T\mathbf{t}_n\} \end{equation*} is optimal for ${\cal A}_T$. Using \cref{thm:optimal-seq} we then have that, for $p\geq 1$, the spaces \begin{equation}\label{eq:optimal-seq-TP} \mathbb{X}_n^p=\begin{cases} {\rm span}\{(AA^T)^i A\mathbf{s}_1,\ldots, (AA^T)^i A\mathbf{s}_n\}, & p = 2i+1, \\ {\rm span}\{(AA^T)^{i+1}\mathbf{t}_1,\ldots, (AA^T)^{i+1}\mathbf{t}_n\}, & p=2i+2, \end{cases} \end{equation} are optimal for ${\cal A}$. Moreover, we can apply \cref{thm:low-rank-1} to an orthonormal basis for any of the above subspaces $\mathbb{X}_n^p$, $p\geq 1$, to obtain a best rank-$n$ approximation to $A$. Similarly for $\mathbb{Y}_n^p$ and \cref{thm:low-rank-2}. We remark that the space $\mathbb{X}_n^2$ in \cref{eq:optimal-seq-TP} is the second optimal space found by Melkman and Micchelli. \end{example} \begin{example} Let us compare the result of \cref{thm:optimal-seq} with the optimality criteria in \cref{sec:1-width}. For simplicity we consider the case $n=1$, $m=3$ and $A=\Sigma$. We further assume that the unit vector $\mathbf{x}_1$ is at the boundary of satisfying the optimality criteria in \cref{sec:1-width}. More precisely, we let $\mathbf{x}_1=\sum_{j=1}^3w_j\mathbf{u}_j$, and using \cref{ineq:optimality-3D}, we assume that \begin{align*} w_1^2 = \frac{\sigma_1^2 - \sigma_2^2}{\sigma_1^2 - \sigma_3^2}, \quad w_2=0, \quad w_3^2 = \frac{\sigma_2^2 - \sigma_3^2}{\sigma_1^2 - \sigma_3^2}. \end{align*} It then follows from \cref{thm:optimality_criterion_n=1} that ${\rm span}\{\mathbf{x}_1\}$ is optimal for ${\cal A}$. Now, let $\mathbf{y}_1=A\mathbf{x}_1/\|A\mathbf{x}_1\|$. From \cref{thm:optimal-seq} we know that ${\rm span}\{\mathbf{y}_1\}$ is also optimal for ${\cal A}$. Moreover, if we let $\mathbf{y}_1=\sum_{j=1}^3z_j\mathbf{u}_j$, then $z_2=0$ and \begin{align*} z_3^2 = \frac{\sigma_3^2w_3^2}{\sigma_1^2w_1^2+ \sigma_3^2w_3^2} = \frac{\sigma_2^2 - \sigma_3^2}{\sigma_1^2 - \sigma_3^2 + s}< \frac{\sigma_2^2 - \sigma_3^2}{\sigma_1^2 - \sigma_3^2} = w_3^2, \end{align*} where $s=(\sigma_1^2/\sigma_3^2-1)(\sigma_1^2-\sigma_2^2)> 0$. Thus, $\mathbf{y}_1$ is closer to the first singular vector (or in this case, eigenvector) $\mathbf{u}_1=\mathbf{e}_1$ than $\mathbf{x}_1$. We will look closer at this property in the next theorem. \end{example} Note that the definition of the spaces $\mathbb{X}_n^p$ and $\mathbb{Y}_n^p$ in \cref{eq:seq} is very similar to the (block) power method for eigenvalue approximation. The following result, based on \cite[Theorem~7.1]{Sande:2019}, should therefore not come as a surprise for anyone familiar with this method. \begin{theorem}\label{thm:eig} Suppose $\mathbb{X}_n^1$ is optimal for ${\cal A}$ and $\mathbb{Y}_n^1$ is optimal for ${\cal A}_T$. Let $P_{n,p}$ be the orthogonal projection onto $\mathbb{X}_n^p$ and $\Pi_{n,p}$ be the orthogonal projection onto $\mathbb{Y}_n^p$. Then, \begin{align*} \|(I-P_{n,p})\mathbf{u}_j\|, \|(I-\Pi_{n,p})\mathbf{v}_j\| \leq \left(\frac{\sigma_{n+1}}{\sigma_j}\right)^p, \quad j=1,2,\ldots,n, \end{align*} and consequently, \begin{align*} \mathbb{X}_n^p \xrightarrow[p\to\infty]{}{\rm span}\{\mathbf{u}_1,\ldots\mathbf{u}_n\}, \quad \mathbb{Y}_n^p \xrightarrow[p\to\infty]{}{\rm span}\{\mathbf{v}_1,\ldots\mathbf{v}_n\}. \end{align*} \end{theorem} The above result follows from the next lemma and so we will postpone the proof. To ease notation we define the two function classes ${\cal A}^p$ and ${\cal A}^p_T$, for $p\geq 1$, by ${\cal A}^1:={\cal A}$, ${\cal A}^1_T:={\cal A}_T$ and \begin{equation}\label{eq:Ar} {\cal A}^p:=A({\cal A}^{p-1}_T),\quad {\cal A}^p_T:=A^T({\cal A}^{p-1}), \end{equation} for $p\geq 2 $. Using an argument similar to the proofs of \cite[Lemma~1]{Floater:2018} and \cite[Lemma~2]{Sande:2020} we have the following result. \begin{lemma}\label{lem:seq-sigma} If $\mathbb{X}_n^1$ is optimal for ${\cal A}$ and $\mathbb{Y}_n^1$ is optimal for ${\cal A}_T$, then \begin{align*} E({\cal A}^p, \mathbb{X}_n^p) = E({\cal A}^p_T, \mathbb{Y}_n^p) = (\sigma_{n+1})^p. \end{align*} \end{lemma} \begin{proof} Let $P_{n,p}$ be the orthogonal projection onto $\mathbb{X}_n^p$ and $\Pi_{n,p}$ be the orthogonal projection onto $\mathbb{Y}_n^p$. Then, the matrix \begin{align*} (I-P_{n,p})A\Pi_{n,p-1} = 0, \end{align*} since $A\Pi_{n,p-1}\mathbf{x}\in \mathbb{X}_n^p$ for any vector $\mathbf{x}\in\mathbb{R}^m$. If we now let the matrix $B$ be defined by $B:=A^T(AA^T)^i$ for $p=2i+2$ and $B:=(A^TA)^i$ for $p=2i+1$, then \begin{align*} E({\cal A}^p, \mathbb{X}_n^p) &= \|(I-P_{n,p})AB\| = \|(I-P_{n,p})A(I-\Pi_{n,p-1})B\| \\ &\leq \|(I-P_{n,p})A\|\,\|(I-\Pi_{n,p-1})B\| = \sigma_{n+1}\,E({\cal A}_T^{p-1}, \mathbb{Y}_n^{p-1}), \end{align*} since $\mathbb{X}_n^p$ is optimal for ${\cal A}$ by \cref{thm:optimal-seq}. By a similar argument we have \begin{align*} E({\cal A}_T^p, \mathbb{Y}_n^p) = \sigma_{n+1}\,E({\cal A}^{p-1}, \mathbb{X}_n^{p-1}), \end{align*} and the result follows from induction on $p$. \end{proof} From the definitions of ${\cal A}^p$ and ${\cal A}^p_T$ in \cref{eq:Ar} we deduce that $d_n({\cal A}^p)=d_n({\cal A}^p)=(\sigma_{n+1})^p$. It thus follows from \cref{lem:seq-sigma} that if $\mathbb{X}_n^1$ is optimal for ${\cal A}$ and $\mathbb{Y}_n^1$ is optimal for ${\cal A}_T$ then $\mathbb{X}_n^p$ is optimal for ${\cal A}^p$ and $\mathbb{Y}_n^p$ is optimal for ${\cal A}_T^p$. In fact, using the arguments of \cite[Section~4]{Floater:2018} one can show that if $\mathbb{X}_n^1$ is optimal for ${\cal A}$ and $\mathbb{Y}_n^1$ is optimal for ${\cal A}_T$ then $\mathbb{X}_n^p$ is optimal for ${\cal A}^s$ and $\mathbb{Y}_n^p$ is optimal for ${\cal A}_T^s$ for all $p\geq s\geq 1$. \begin{proof}[Proof of \cref{thm:eig}] The two cases are analogous and so we only consider the case $\|(I-P_{n,p})\mathbf{u}_j\|$. Using the definition of the spectral norm and \cref{lem:seq-sigma} we have \begin{equation}\label{ineq:optimal} \begin{aligned} \|(I-P_{n,p})(AA^T)^i\mathbf{x}\| &\leq \|(I-P_{n,p})(AA^T)^i\|\,\|\mathbf{x}\| \\ &= E({\cal A}^p, \mathbb{X}_n^p) = (\sigma_{n+1})^p, &&\quad p=2i, \\ \|(I-P_{n,p})A(A^TA)^i\mathbf{x}\| &\leq \|(I-P_{n,p})A(A^TA)^i\|\,\|\mathbf{x}\| \\ &= E({\cal A}^p, \mathbb{X}_n^p) = (\sigma_{n+1})^p, &&\quad p=2i+1, \end{aligned} \end{equation} for any unit vector $\mathbf{x}\in \mathbb{R}^m$. We first consider $p=2i$. Then, for any $j=1,\ldots,n$ we have \begin{align*} \|(I-P_{n,p})\mathbf{u}_j\| = \|(I-P_{n,p})\frac{1}{\sigma_j^p}(AA^T)^i\mathbf{u}_j\| = \frac{1}{\sigma_j^p}\|(I-P_{n,p})(AA^T)^i\mathbf{u}_j\|, \end{align*} and by letting $\mathbf{x}=\mathbf{u}_j$ in \cref{ineq:optimal} we obtain \begin{align*} \|(I-P_{n,p})\mathbf{u}_j\| \leq \left(\frac{\sigma_{n+1}}{\sigma_j}\right)^p. \end{align*} A similar argument proves the case $p=2i+1$. \end{proof} \section{Conclusions} \label{sec:conclusion} We have addressed the problem of best rank-$n$ approximations to a given matrix $A$ in the spectral norm, and we have shown that the problem can be related to the concept of Kolmogorov $n$-widths and corresponding optimal spaces. More precisely, any orthonormal basis in an optimal $n$-dimensional space for the image of the Euclidean unit ball under $A$ generates a best rank-$n$ approximation to $A$. This results in a variety of best low-rank approximations that are different from the truncated SVD. In this perspective, we have laid out explicit characterizations of optimal subspaces of any dimension, and presented a complete description of all the optimal one-dimensional subspaces. Furthermore, we have provided a simple construction to obtain a sequence of optimal $n$-dimensional subspaces once an initial optimal subspace is known. The paper features an explicit theoretical contribution. The task to retrieve useful information while maintaining the underlying physical feasibility often necessitates the search for low-rank approximations with/without specific properties/structures of the data matrix \cite{Antoulas:1997,Chu:2003,Higham:1989,Markovsky:2008,Ottaviani:2014}. In this context, the results we have presented may also have a practical impact. However, we have not considered here the problem of finding efficient algorithms to compute our approximations. We note, on the other hand, that in the special case of Hankel matrices such algorithms have been considered in \cite{Knirsch:preprint}. \section*{Acknowledgements} This work was supported by the Beyond Borders Programme of the University of Rome Tor Vergata through the project ASTRID (CUP E84I19002250005) and by the MIUR Excellence Department Project awarded to the Department of Mathematics, University of Rome Tor Vergata (CUP E83C18000100006). C. Manni, E. Sande and H. Speleers are members of Gruppo Nazionale per il Calcolo Scientifico, Istituto Nazionale di Alta Matematica.

https://arxiv.org/abs/1908.03533

Strong external difference families in abelian and non-abelian groups

Strong external difference families (SEDFs) have applications to cryptography and are rich combinatorial structures in their own right; until now, all SEDFs have been in abelian groups. In this paper, we consider SEDFs in both abelian and non-abelian groups. We characterize the order of groups possessing admissible parameters for non-trivial SEDFs, develop non-existence and existence results, several of which extend known results, and present the first family of non-abelian SEDFs. We introduce the concept of equivalence for EDFs and SEDFs, and begin the task of enumerating SEDFs. Complete results are presented for all groups up to order $24$, underpinned by a computational approach.

\section{Introduction} There has been considerable recent interest in strong external difference families (SEDFs), which have applications to cryptography and are rich combinatorial structures in their own right (see \cite{BaJiWeZh}, \cite{HuPa}, \cite{JeLi}, \cite{MaSt},\cite{PaSt}, \cite{WeYaFuFe}). Most of this activity has centred around identifying parameters for which a SEDF exists or does not exist, and obtaining constructions via the application of classic techniques such as cyclotomy. Up till now, all SEDFs have been in abelian groups. We are motivated by the following questions: what is the situation for SEDFs in general finite groups, not simply abelian groups? For given parameters, which groups contain SEDFs with these parameters? How many ``different'' SEDFs exist with the same parameters? In this paper, we introduce the notion of equivalence for SEDFs, and characterise the order of groups with admissible parameters for SEDFs. We present a recursive framework for constructing families of SEDFs with $\lambda=1$ (and related generalized SEDFs) in cyclic groups, which encompasses known results on SEDFs and GSEDFs. We present the first non-abelian SEDFs: a construction for an infinite non-abelian family using dihedral groups. We establish the existence of at least two non-equivalent $(k^2+1, 2, k,1)$-SEDFs for every $k>2$. Finally, we begin the task of enumerating SEDFs, and present complete results for all groups up to order $24$, underpinned by a computational approach using constraint satisfaction programming. \section{Strong External Difference Families} External difference families were first introduced in \cite{OgKuStSi} in relation to AMD codes, while strong EDFs were subsequently introduced in \cite{PaSt} to correspond to strong AMD codes. They were defined in abelian groups. In this paper, we extend the concept of EDF and SEDF in the natural way to any group of order $n$, whether abelian or non-abelian. The definitions precisely correspond to the originals, with the removal of the word ``abelian''. Since the differences are defined in terms of ordered pairs, there is no ambiguity in this definition. For abelian groups, additive notation is generally used; however when we focus on the non-abelian and general cases, we will adopt multiplicative notation. \begin{definition}[External difference family] Let $\G$ be a group of order $n$. An {\bf $(n,m,k,\lambda)$-external difference family} (or {\bf $(n,m,k,\lambda)$-EDF}) is a set of $m \geq 2$ disjoint $k$-subsets of $\G$, say $A_1, \dots , A_m$, such that the multiset \begin{equation*} M=\{ xy^{-1}: x \in A_i, y \in A_j, i \neq j \} \end{equation*} comprises $\lambda$ occurrences of each non-identity element of $\G$. \end{definition} \begin{definition}[Strong external difference family] Let $\G$ be a group of order $n$. An {\bf $(n,m,k,\lambda)$-strong external difference family} (or {\bf $(n,m,k,\lambda)$-SEDF}) is a set of $m \geq 2$ disjoint $k$-subsets of $\G$, say $A_1, \dots , A_m$, such that, for every $i$, $1 \leq i \leq m$, the multiset \begin{equation*} M_i=\{ xy^{-1}: x \in A_i, y \in\cup_{j \neq i} A_j \} \end{equation*} comprises $\lambda$ occurrences of each non-identity element of $\G$. (An $(n,m,k,\lambda)$-SEDF is, by definition, an $(n,m,k,m\lambda)$-EDF.) \end{definition} For every group $\G$, there is at least one SEDF. This has $k=1$ and is often referred to as the \emph{trivial} SEDF. \begin{example}\label{singletons} Let $\G=\{ g_1, \ldots, g_n \}$ be a group of order $n$. Then $A_i=\{g_i\}$ for $1 \leq i \leq n$ is an $(n,n,1,1)$-SEDF over $\G$. \end{example} To our knowledge, the existing literature contains no examples of non-trivial non-abelian SEDFs. The only known explicit construction for non-abelian EDFs is given in \cite{HuPa2}; these EDFs have the \emph{bimodal} property and are not SEDFs. In \cite{Bu1}, new existence results on disjoint difference families (DDFs) in non-abelian groups are established; since $(n,k,k-1)$-DDFs correspond to certain \emph{near-complete} EDFs (see \cite{ChDi} for the proof in the abelian setting), these should guarantee further non-abelian EDFs, but analysis of parameters shows that these will not be strong. An EDF may be viewed as one generalization of a SEDF; a different generalization is a GSEDF, defined in \cite{PaSt} and further explored in \cite{LuNiCa}. We extend this also to the non-abelian setting. \begin{definition}[Generalized strong external difference family] Let $\G$ be a group of order $n$. An {\bf $(n,m;k_1, \ldots, k_m;\lambda_1, \ldots, \lambda_m)$-generalized strong external difference family} (or {\bf $(n,m;k_1, \ldots, k_m;\lambda_1, \ldots, \lambda_m)$-GSEDF}) is a set of $m \geq 2$ disjoint $k_i$-subsets of $\G$, say $A_1, \dots , A_m$, such that, for every $i$, $1 \leq i \leq m$, the multiset \begin{equation*} M_i=\{ xy^{-1}: x \in A_i, y \in\cup_{j \neq i} A_j \} \end{equation*} comprises $\lambda_i$ occurrences of each non-identity element of $\G$. (An $(n,m;k, \ldots, k;\lambda, \ldots, \lambda)$-GSEDF is, by definition, an $(n,m,k,\lambda)$-SEDF.) \end{definition} \section{Equivalence for EDFs and SEDFs} In order to enumerate and compare EDFs and SEDFs, it is essential to have a notion of what it means for two such objects to be equivalent. Equivalence of EDFs has not been previously discussed in the literature; we will adopt a definition consistent with the definition of equivalence for difference families and difference sets (see the Handbook of Combinatorial Designs, Part VI, 16.93, 18.8 \cite{CoDi}). In preparation for this definition, we establish the following: \begin{proposition}\label{translate} Let $\G$ be a group and let $g \in G$. \begin{itemize} \item[(i)] If $\A=\{A_1, \ldots, A_m\}$ is an $(n,m,k,\lambda)$-EDF over $\G$, then the left and right translates $g \A$ and $\A g$ of $\A$ also form $(n,m,k,\lambda)$-EDFs over $\G$. \item[(ii)] If $\A=\{A_1, \ldots, A_m\}$ is an $(n,m,k,\lambda)$-SEDF over $\G$, then the left and right translates $g\A$ and $\A g$ of $\A$ also form $(n,m,k,\lambda)$-SEDFs over $\G$. \end{itemize} \end{proposition} \begin{proof} First consider right translates: by the reversal rule for inverses, $(xg)(yg)^{-1}=(xg)(g^{-1}y^{-1})=x(g g^{-1})y^{-1}=xy^{-1}$ for $x,y \in \G$ . So for any $A_i g (A_j g)^{-1}$ ($i \neq j$), the multiset of differences is precisely the same as the multiset corresponding to $A_i (A_j)^{-1}$. For left translates, $(gx)(gy)^{-1}=gx y^{-1} g^{-1}$, the conjugate of $xy^{-1}$ by $g^{-1}$. So the multiset of differences corresponding to $(g A_i)(g A_j)^{-1}$ is the set of conjugates of the differences $A_i (A_j)^{-1}$ by (fixed) $g^{-1} \in \G$. For an EDF $\A$, the union of multisets of such differences for all $i \neq j$ comprises each non-identity group element $\lambda$ times; since conjugation by $g^{-1}$ is an automorphism of the group $\G$ (and in particular fixes the identity), the corresponding union for $g \A$ also comprises each non-identity group element $\lambda$ times. An analogous argument holds if $\A$ is an SEDF. \end{proof} \begin{proposition}\label{auto} Let $\G$ be a group, let $\A=\{A_1, \ldots, A_m\}$ be an $(n,m,k,\lambda)$-EDF (respectively, SEDF) over $\G$ and let $\alpha$ be an automorphism of $\G$, Then $\alpha(\A)=\{ \alpha(A_1), \ldots, \alpha(A_m)\}$ forms an $(n,m,k,\lambda)$-EDF (respectively, SEDF) over $\G$. \end{proposition} \begin{proof} For $x \in \alpha(A_i)$ and $y \in \alpha(A_j)$, $j \neq i$, consider the difference $xy^{-1}$. Here $x=\alpha(a_i)$ and $y=\alpha(a_j)$ for some $a_i \in A_i, a_j \in A_j$. Then $xy^{-1}=\alpha(a_i) (\alpha(a_j))^{-1}=\alpha(a_i {a_j}^{-1})$ since $\alpha$ is an automorphism. For an EDF $\A$, the union of multisets of external differences $A_i(A_j)^{-1}$ for all $i \neq j$ comprises each non-identity group element $\lambda$ times. Since $\alpha$ is an automorphism of the group $\G$, the union of multisets of external differences $\alpha(A_i) (\alpha(A_j))^{-1}=\alpha(A_i A_j^{-1}) \,(i \neq j)$ also comprises each non-identity group element $\lambda$ times. An analogous argument holds if $\A$ is an SEDF. \end{proof} We now make the following definition: \begin{definition}\label{equivalent} \begin{itemize} \item[(i)] Let $\G$ be a group of order $n$, and let $\A=\{A_1, \ldots, A_m\}$ and $\A^{\prime}=\{A^{\prime}_1, \ldots, A^{\prime}_m\}$ be two $(n,m,k,\lambda)$-EDFs (respectively, SEDFs) over $\G$. We shall say that $\A$ is \emph{equivalent} to $\A^{\prime}$ if there is an automorphism $\alpha$ of $G$ and elements $g,h \in \G$ such that, for all $1 \leq i \leq m$, $A^{\prime}_i=h(\alpha(A_i))g$. \item[(ii)] Two $(n,m,k,\lambda)$-EDFs (respectively, SEDFs) $\A=\{A_1, \ldots, A_m\}$ (in group $\G_1$) and $\A^{\prime}=\{A^{\prime}_1, \ldots, A^{\prime}_m\}$ (in group $\G_2$) are \emph{equivalent} if there is an isomorphism $\alpha$ between $\G_1$ and $\G_2$ and elements $g,h \in \G_2$ such that, for all $1 \leq i \leq m$, $A^{\prime}_i= h(\alpha(A_i))g$. \end{itemize} \end{definition} Propositions \ref{translate} and \ref{auto} guarantee that, if $A$ is an $(n,m,k,\lambda)$-EDF (respectively, SEDF) and $A^{\prime}$ is equivalent to $A$, then $A^{\prime}$ is also an $(n,m,k,\lambda)$-EDF (respectively, SEDF). This definition can be extended to GSEDFs (and other EDF-like structures) in the natural way. \begin{example}\label{equiv:ex} \begin{itemize} \item[(i)] Let $\G=(\mathbb{Z}_{5}, +)$. The $(5,2,2,1)$-SEDF $(\{0,1\},\{2,4\})$ is equivalent to the SEDF $(\{1,4\}, \{2,3\})$ by translating by $2$, and to the SEDF $(\{0,2\}, \{3,4\})$ by applying the automorphism $x \mapsto 2x$. \item[(ii)] Let $\G=(\mathbb{Z}_{17}, +)$. The $(17,2,4,1)$-SEDF $(\{0,1,4,5\},\{6,8,14,16\})$ is equivalent to the SEDF $(\{1,4,13,16\},\{2,8,9,15\})$ by applying the automorphism $x \mapsto 3x$ then translating by $1$. \end{itemize} \end{example} \section{Non-existence and existence in general groups} The following theorem expresses necessary relationships between the parameters of an EDF or SEDF. Although originally established (\cite{OgKuStSi},\cite{PaSt}) in the context of abelian groups, the conditions remain valid in any finite group. \begin{proposition}\label{par_rel} \begin{itemize} \item[(1)] Necessary conditions for the existence of an $(n,m,k,\lambda)$-EDF in a group $\G$ of order $n$ are \begin{itemize} \item[(i)] $m \geq 2$; \item[(ii)] $n \geq mk$; \item[(iii)] $ \lambda(n-1)=k^2 m(m-1)$. \end{itemize} \item[(2)]Necessary conditions for the existence of an $(n,m,k,\lambda)$-SEDF in a group $\G$ of order $n$ are \begin{itemize} \item[(i)] $m \geq 2$; \item[(ii)] $n \geq mk$; \item[(iii)] $ \lambda(n-1)=k^2 (m-1)$. \end{itemize} \end{itemize} Parameter sets that satisfy these conditions will be called \emph{admissible}. \end{proposition} \begin{proof} In each case, (i) is in the definition, while (ii) is due to the fact that the $m$ $k$-sets are disjoint. Part (iii) follows from double-counting the ordered pairs which correspond to the external differences; this does not assume commutativity of the group operation. \end{proof} Analysis of the conditions of Proposition \ref{par_rel} can be used to rule-out certain sets of parameters for SEDFs. Since the analysis is purely number theoretical on the parameter equation, these results apply equally to abelian and non-abelian groups. The following results from the literature hold for any group of order $n$. \begin{lemma}\label{lambdalem} \begin{itemize} \item For an $(n,m,k,\lambda)$-SEDF, $k=1$ and $\lambda=1$; or $k>1$ and $\lambda<k$ (\cite{HuPa}); \item If $\mathrm{gcd}(k,n-1)=1$, then a non-trivial $(n,m,k,\lambda)$-SEDF does not exist (\cite{JeLi}). \end{itemize} \end{lemma} The following new result significantly restricts the orders of SEDF-containing groups. An integer is said to be \emph{square-free} if it is divisible by no square other than $1$. \begin{proposition}\label{square-free} Let $n-1$ be square-free. Then there exists no non-trivial $(n,m,k,\lambda)$-SEDF. \end{proposition} \begin{proof} Let $n-1$ be square-free. Suppose there exists an $(n,m,k,\lambda)$-SEDF. By Proposition \ref{par_rel}(iii), we have $\lambda(n-1)=k^2(m-1)$, which rearranges to $n-1=\frac{ k^2(m-1)}{\lambda}$. Clearly the RHS must be an integer as $n-1$ is, ie $k^2(m-1)$ is divisible by $\lambda$. Moreover, $n-1$ is square-free. If $p^{\alpha}$ is a prime power divisor of $k$, then it must be a divisor of $\lambda$ (otherwise $\frac{k^2(m-1)}{\lambda}$ would be divisible by $p^2$). So $\lambda=ak$ for some positive integer $a$, i.e. $k$ divides $\lambda$. But by Lemma \ref{lambdalem}, for any SEDF we must have either $k=1$ and $\lambda=1$ or $k>1$ and $\lambda<k$. Since $k \mid \lambda$, we have $k \leq \lambda$ and so it is not possible to have $k>1$. So $k=1$ and $\lambda=1$, ie $n=m$, and the SEDF is trivial. \end{proof} We prove a partial converse of Proposition \ref{square-free}. \begin{proposition}\label{squareful} Let $n (>2)$ be such that $n-1$ is not square-free. Then there exists a set of admissible parameters $(n,m,k,\lambda)$ for a non-trivial SEDF. \end{proposition} \begin{proof} Write $n-1=a^2 \cdot b$, where $a^2$ is the largest square dividing $n-1$ and $b \geq 1$. By definition, $a>1$. Take the tuple of parameters $(n,m,k,\lambda)=(n, b+1, a, 1)$: we will show that this satisfies all three conditions of Proposition \ref{par_rel} (2), and the non-trivial condition $k>1$. Taking $\lambda=1$ reduces the parameter equation from Proposition \ref{par_rel} (2)(iii) to $n-1=k^2(m-1)$, ie $a^2b=k^2(m-1)$. We may set $a=k$ and $b=m-1$ to obtain results which satisfy this equation. From above $k=a>1$. Since $b \geq 1$, $m \geq 2$. We must check that $n \geq km$; in fact we prove that $n-1=a^2 b \geq km= a(b+1)$. We have $$ a^2b \geq a(b+1) \Leftrightarrow ab \geq b+1 \Leftrightarrow b(a-1) \geq 1$$ which holds since $a \geq 2$ and $b \geq 1$. \end{proof} For a valid group order $n$, there may of course be other sets of admissible parameters for an $(n,m,k, \lambda)$-SEDF, beyond those in the above proof. Table \ref{table1} contains all admissible SEDF parameter sets for $n$ up to $64$. These were found computationally. \begin{table}[h] \centering \begin{tabular}{cccc|cccc|cccc|cccc|cccc} $n$ & $m$ & $k$ & $\lambda$ & $n$ & $m$ & $k$ & $\lambda$ & $n$ & $m$ & $k$ & $\lambda$ & $n$ & $m$ & $k$ & $\lambda$ & $n$ & $m$ & $k$ & $\lambda$ \\ \hline 5 & 2 & 2 & 1 & 26 & 3 & 5 & 2 & 37 & 5 & 3 & 1 & 49 & 10 & 4 & 3 & 55 & 3 & 9 & 3 \\ \cline{1-4} 9 & 2 & 4 & 2 & 26 & 4 & 5 & 3 & 37 & 5 & 6 & 4 & 49 & 13 & 2 & 1 & 55 & 3 & 18 & 12 \\ \cline{13-16} 9 & 3 & 2 & 1 & 26 & 5 & 5 & 4 & 37 & 6 & 6 & 5 & 50 & 2 & 7 & 1 & 55 & 4 & 6 & 2 \\ \cline{1-8} 10 & 2 & 3 & 1 & 28 & 2 & 9 & 3 & 37 & 9 & 3 & 2 & 50 & 2 & 14 & 4 & 55 & 4 & 12 & 8 \\ 10 & 3 & 3 & 2 & 28 & 3 & 9 & 6 & 37 & 10 & 2 & 1 & 50 & 2 & 21 & 9 & 55 & 5 & 9 & 6 \\ \cline{1-4}\cline{9-12} 13 & 2 & 6 & 3 & 28 & 4 & 3 & 1 & 41 & 2 & 20 & 10 & 50 & 3 & 7 & 2 & 55 & 7 & 3 & 1 \\ 13 & 4 & 2 & 1 & 28 & 4 & 6 & 4 & 41 & 3 & 10 & 5 & 50 & 3 & 14 & 8 & 55 & 7 & 6 & 4 \\ \cline{1-4} 17 & 2 & 4 & 1 & 28 & 7 & 3 & 2 & 41 & 6 & 4 & 2 & 50 & 4 & 7 & 3 & 55 & 13 & 3 & 2 \\ \cline{5-8}\cline{17-20} 17 & 2 & 8 & 4 & 29 & 2 & 14 & 7 & 41 & 11 & 2 & 1 & 50 & 5 & 7 & 4 & 57 & 2 & 28 & 14 \\ \cline{9-12} 17 & 3 & 4 & 2 & 29 & 8 & 2 & 1 & 45 & 2 & 22 & 11 & 50 & 6 & 7 & 5 & 57 & 3 & 14 & 7 \\ \cline{5-8} 17 & 4 & 4 & 3 & 33 & 2 & 8 & 2 & 45 & 12 & 2 & 1 & 50 & 7 & 7 & 6 & 57 & 8 & 4 & 2 \\ \cline{9-12}\cline{13-16} 17 & 5 & 2 & 1 & 33 & 2 & 16 & 8 & 46 & 2 & 15 & 5 & 51 & 2 & 10 & 2 & 57 & 15 & 2 & 1 \\ \cline{1-4}\cline{17-20} 19 & 2 & 6 & 2 & 33 & 3 & 4 & 1 & 46 & 3 & 15 & 10 & 51 & 2 & 20 & 8 & 61 & 2 & 30 & 15 \\ 19 & 3 & 3 & 1 & 33 & 3 & 8 & 4 & 46 & 6 & 3 & 1 & 51 & 3 & 5 & 1 & 61 & 4 & 10 & 5 \\ 19 & 3 & 6 & 4 & 33 & 4 & 8 & 6 & 46 & 6 & 6 & 4 & 51 & 3 & 10 & 4 & 61 & 6 & 6 & 3 \\ 19 & 5 & 3 & 2 & 33 & 5 & 4 & 2 & 46 & 11 & 3 & 2 & 51 & 3 & 15 & 9 & 61 & 16 & 2 & 1 \\ \cline{1-4}\cline{9-12}\cline{17-20} 21 & 2 & 10 & 5 & 33 & 7 & 4 & 3 & 49 & 2 & 12 & 3 & 51 & 4 & 10 & 6 & 64 & 2 & 21 & 7 \\ 21 & 6 & 2 & 1 & 33 & 9 & 2 & 1 & 49 & 2 & 24 & 12 & 51 & 5 & 5 & 2 & 64 & 3 & 21 & 14 \\ \cline{1-8} 25 & 2 & 12 & 6 & 37 & 2 & 6 & 1 & 49 & 3 & 12 & 6 & 51 & 5 & 10 & 8 & 64 & 8 & 3 & 1 \\ 25 & 3 & 6 & 3 & 37 & 2 & 12 & 4 & 49 & 4 & 4 & 1 & 51 & 7 & 5 & 3 & 64 & 8 & 6 & 4 \\ 25 & 4 & 4 & 2 & 37 & 2 & 18 & 9 & 49 & 4 & 8 & 4 & 51 & 9 & 5 & 4 & 64 & 15 & 3 & 2 \\ \cline{13-16}\cline{17-20} 25 & 7 & 2 & 1 & 37 & 3 & 6 & 2 & 49 & 4 & 12 & 9 & 53 & 2 & 26 & 13 & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} \\ \cline{1-4} 26 & 2 & 5 & 1 & 37 & 3 & 12 & 8 & 49 & 5 & 6 & 3 & 53 & 14 & 2 & 1 & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} \\ \cline{13-16} 26 & 2 & 10 & 4 & 37 & 4 & 6 & 3 & 49 & 7 & 4 & 2 & 55 & 2 & 18 & 6 & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} & \multicolumn{1}{l}{} \end{tabular} \caption{All admissible SEDF parameter sets for orders 1 to 64} \label{table1} \end{table} It is not the case that an SEDF will necessarily exist for all admissible parameters. Some parameter sets can be ruled-out in certain classes of group using combinatorial or algebraic arguments. The following result summarises non-existence results for SEDFs in abelian groups: \begin{proposition}\label{nonexistence} A non-trivial abelian $(n,m,k,\lambda)$-SEDF does not exist in each of the following cases: \begin{itemize} \item[(1)] $m \in \{3,4\}$ (\cite{MaSt}); \item[(2)] $m>2$ and $n$ prime (\cite{MaSt}); \item[(3)] $m>2$ and $n=pq$ for distinct primes $p,q$ (\cite{BaJiWeZh}); \item[(4)] $m>2$ and $\lambda=2$ (\cite{HuPa}); \item[(5)] $m>2$ and $\lambda>1$ and $\frac{\lambda(k-1)(m-2)}{(\lambda-1)k(m-1)}>1$ (\cite{HuPa}); \item[(6)] $m>2$ and there is a prime $p$ dividing $n$ such that $\mathrm{gcd}(km,p)=1$ and $m$ is not congruent to $2$ mod $p$ (\cite{BaJiWeZh}); \item [(7)] $m>2$ and $n$ a prime power, when $\G$ is cyclic (\cite{BaJiWeZh} and \cite{LeLiPr}); \item[(8)] $m>2$ and $n$ is the product of at most three (not necessarily distinct) primes, except possibly when $\G=C_p^3$ and $p$ is a prime greater than $3 \times 10^{12}$ (\cite{LeLiPr}). \end{itemize} \end{proposition} Of these, (4) and (5) employ direct combinatorial arguments which rely on the commutativity of the operation, while the others use character theory in abelian groups. It is therefore possible that SEDFs with these forbidden parameters may exist in non-abelian groups. The following is a summary of the main constructive existence results known for abelian groups. All known constructions for families of SEDFs have $m=2$. \begin{proposition} \label{SEDFexistence} An $(n,m,k, \lambda)$-SEDF exists in the group $\G$ in the following cases: \begin{itemize} \item[(1)] $(n,m,k,\lambda)=(k^2+1, 2, k, 1)$ and $\G=\mathbb{Z}_{k^2+1}$ (\cite{PaSt}); sets given by $\{0,1,\ldots,k-1 \}, \{k,2k,\ldots,k^2\}$. \item[(2)] $(n,m,k,\lambda)=(n, 2, \frac{n-1}{2}, \frac{n-1}{4})$ and $n$ is congruent to $1$ mod $4$, provided there exists an $(n,\frac{n-1}{2},\frac{n-5}{4},\frac{n-1}{4})$ partial difference set in $\G$ (\cite{HuPa}). In particular, when $n$ is a prime power, appropriate sets are given by the non-zero squares and non-squares in the multiplicative group of the finite field of order $n$. \item[(3)] $(n,m,k,\lambda)=(q,2,\frac{q-1}{4},\frac{q-1}{16})$, where $q=16 t^2+1$ is a prime power and $t$ an integer (\cite{BaJiWeZh}); the sets are cyclotomic classes in the finite field of order $q$. \item[(4)] $(n,m,k,\lambda)=(p,2,\frac{p-1}{6},\frac{p-1}{36})$, where $p=108 t^2+1$ is a prime and $t$ an integer (\cite{BaJiWeZh}); the sets are cyclotomic classes in the finite field of order $p$. \end{itemize} \end{proposition} Note that different constructions may produce equivalent SEDFs for certain specific parameter sets. For example, when $n=5$, Example \ref{equiv:ex} shows that Proposition \ref{SEDFexistence} (1) and (2) yield equivalent SEDFs, although this is not the case for larger values of $n$. For $\lambda=1$, a parameter characterization for the abelian case is given in \cite{PaSt}: \begin{proposition}\label{PaStlambda=1} There exists an abelian $(n,m,k,1)$-SEDF if and only if $m=2$ and $n=k^2+1$ or $k=1$ and $m=n$. \end{proposition} The proof of the the forward direction relies on the commutativity of the operation, while for the reverse direction, constructions are provided by the trivial SEDF and Proposition \ref{SEDFexistence} (1). Finally, we mention the following characterization for the \emph{near-complete} abelian case (when $n=mk+1$), given in \cite{JeLi}. An alternative construction for a $(243,11,22,20)$-SEDF, using cyclotomic methods, is given in \cite{WeYaFuFe}. \begin{proposition}\label{nearcomplete} Let $\A=\{ A_1, \ldots, A_m \}$ be a collection of $k$-sets ($k>1$) which partition the non-identity elements of an abelian group $\G$. Then $\A$ is a non-trivial $(n,m,k, \lambda)$-SEDF if and only if \begin{itemize} \item[(i)] $(n,m,k,\lambda)=(n,2,\frac{n-1}{2}, \frac{n-1}{4})$, $n$ is congruent to $1$ mod $4$, and $A_1$ is a non-trivial regular $(n,\frac{n-1}{2}, \frac{n-5}{4}, \frac{n-1}{4})$ partial difference set in $\G$; or \item[(ii)] $(n,m,k,\lambda)=(243,11,22,20)$ and each $A_j$ is a non-trivial regular $(243,22,1,2)$ partial difference set in $\G$ for $1 \leq j \leq 11$. \end{itemize} \end{proposition} We note that $(243,11,22,20)$ are the only parameters with $m>2$ for which a SEDF is known to exist. \section{Abelian SEDFs with $\lambda=1$: recursive constructions and equivalence} In this section, we consider the situation when $\lambda=1$. By Proposition \ref{PaStlambda=1}, all non-trivial abelian $(n,m,k,1)$-SEDFs are $(k^2+1,2,k,1)$-SEDFs. We show that the known construction of Proposition \ref{SEDFexistence}(1) is actually one amongst numerous different constructions for SEDFs with $\lambda=1$ in cyclic groups, and we provide recursive techniques to obtain these (which also provide constructions for GSEDFs). Moreover, we demonstrate the existence of at least two non-equivalent $(k^2+1,2, k, 1)$-SEDFs for any composite $k$ (in the subsequent section we will extend this to any $k>2$). Throughout, we will write $\mathbb{Z}_n$ additively; its elements will be $\{0,1,\ldots,n-1\}$ and we fix the natural ordering $0< \cdots < n-1$. We begin by presenting a new explicit construction for a $(k^2+1,2,k,1)$-SEDF in $(\mathbb{Z}_{k^2+1}, +)$ for even $k$, and show it is not equivalent to the known construction for $k>2$. Part of its subtraction table is shown in Table \ref{tableNewZ}; the entry in the row labelled by $x$ and the column labelled by $y$ is $x-y$. \begin{theorem}\label{newZconstruction} Let $\G=(\mathbb{Z}_{k^2+1},+)$, where $k=2a$ for some positive integer $a \geq 1$. \begin{itemize} \item[(i)] Let $B_1=\{0,1, \ldots, a-1\} \cup \{2a, 2a+1, \ldots 3a-1\}$ and let $B_2= \bigcup_{i=1}^a \{ (4i-1)a, 4ia\}$; then $\{B_1, B_2\}$ is a $(k^2+1,2,k,1)$-SEDF in $\G$. \item[(ii)] For $k>2$, $\{ B_1, B_2\}$ is not equivalent to the SEDF in $\mathbb{Z}_{k^2+1}$ obtained from the construction of Proposition \ref{SEDFexistence} (1). \end{itemize} \end{theorem} \begin{proof} (i) The multiset $B_2-B_1$ is the union $\cup_{i=1}^a (D_i \cup E_i)$ where \begin{itemize} \item $D_i=\{ (4i-1)a, 4ia\}-\{2a,2a+1,\ldots,3a-1\}=\{(4i-4)a+1, (4i-4)a+2, \ldots, (4i-2)a\}$ \item $E_i=\{ (4i-1)a, 4ia \} - \{ 0,1, \ldots, a-1 \}=\{ (4i-2)a+1, (4i-2)a+2, \ldots, 4ia \}$. \end{itemize} Clearly this comprises each non-zero element of $\mathbb{Z}_{4a^2+1}$ precisely once. The same is necessarily true for the multiset $B_1-B_2$, which consists of the negatives of these elements.\\ (ii) Let $S_1$ be the $(k^2+1,2,k,1)$-SEDF in $\G$ with sets $A_1=\{0, 1,\ldots, 2a-2, 2a-1\}$ and $A_2=\{2a, 4a, \ldots, (4a-2)a, 4a^2 \}$ from Proposition \ref{SEDFexistence} (1). Denote by $S_2$ the SEDF $\{B_1, B_2 \}$ from part (i). We shall show that, for $k>2$, $S_1$ and $S_2$ are not equivalent. Recall that the automorphisms of $\mathbb{Z}_n$ are of the form $x \mapsto cx$ where $c$ is a unit of $\mathbb{Z}_n$. By Definition \ref{equivalent}, $S_1$ is equivalent to $S_2$ if there exists a function $f(x)=\alpha x + \beta$, with $\alpha \in \mathbb{Z}_{k^2+1}^*$ (the group of units of $\mathbb{Z}_{k^2+1}$) and $\beta \in \mathbb{Z}_{k^2+1}$, such that $f(S_1)=S_2$. We prove that no such $f$ can exist for $a>1$. Observe that for $a>1$, each block of each SEDF contains at least four elements. Aiming for a contradiction, assume that there exists such an $f$. First suppose that $f$ maps $A_1$ onto $B_1$. Fix an ordering of the elements of $A_1$ as they are listed in the theorem statement. We show that the images under $f$ of any sequence $\{x,x+1,x+2\}$ of three consecutive elements of $A_1$ must all lie in $C_1=\{0,1,\ldots, a-1\}$ or all lie in $C_2=\{2a, \ldots, 3a-1\}$. Observe that $f(x)+f(x+2)=(\alpha x + \beta)+(\alpha(x+2)+\beta)=\alpha(2x+2)+ 2 \beta= 2 f(x+1)$. Suppose that one of $f(x)$ and $f(x+2)$ lies in $C_1$ and the other in $C_2$. Then $2a \leq f(x)+f(x+2) \leq 4a-2 (< 4a^2+1)$, ie $a \leq f(x+1) \leq 2a-1$; this is impossible since $B_1=\{0,1, \ldots, 3a-1\} \setminus \{a, \ldots, 2a-1\}$. So both of $x$ and $x+2$ are mapped to the same $C_i$ ($i \in \{1,2\}$). If this is $C_1$, then $1 \leq f(x)+f(x+2) \leq 2a+3 (<4a^2+1)$, i.e. $1 \leq f(x+1) < a-2$, so $f(x+1) \in C_1$. If this is $C_2$, then $4a+1 \leq f(x)+f(x+2) \leq 6a-3 (<4a^2+1)$, i.e. $1 \leq f(x+1) < 3a-2$, so $f(x+1) \in C_2$. Thus all three elements are mapped to the same $C_i$. However, repeated applications of this result to $\{0,1,2\}, \{1,2,3\}, \ldots, \{2a-3, 2a-2, 2a-1\}$ in $A_1$ shows that all $2a$ elements of $A_1$ must be mapped to the $a$ elements of one of the $C_i$, which is impossible. Hence $f$ cannot map $A_1$ onto $B_1$. Now, suppose $f$ maps $A_2$ onto $B_1$. Observe that $A_2=2a(A_1+1)$. Since $f(A_2)=B_1$, we have $B_1=f(A_2)=\alpha A_2 + \beta = 2a \alpha A_1 + (2 a \alpha + \beta)$. Now, $2a$ is a unit of $\G$ (with inverse $-2a)$ and $\alpha$ is a unit by definition, so $2a \alpha$ is a unit. Thus there is a function $g(x)= 2a\alpha x+ (2 a \alpha + \beta)$ such that $2a\alpha$ is a unit, $2a \alpha+\beta$ is a group element and $g(A_1)=B_1$. But we have proved above that this cannot occur; again a contradiction. So no such $f$ exists, and $S_1$ and $S_2$ are non-equivalent. \end{proof} \begin{table}[h] \renewcommand{\arraystretch}{1.2} \centering \setlength\tabcolsep{3pt} \begin{tabular}{|c|cccc|cccc|} \hline & $0$ & $1$ & $\ldots$ & $a-1$ & $2a$ & $2a+1$ & $\ldots$ & $3a-1$ \\ \hline $3a$ & $3a$ & $3a-1$ & $\ldots$ & $2a+1$ & $a$ & $a-1$ & $\ldots$ & $1$ \\ $4a$ & $4a$ & $4a-1$ & $\ldots$ & $3a+1$ & $2a$ & $2a-1$ & $\ldots$ & $a+1$ \\ \hline $7a$ & $7a$ & $7a-1$ & $\ldots$ & $6a+1$ & $5a$ & $5a-1$ & $\ldots$ & $4a+1$ \\ $8a$ & $8a$ & $8a-1$ & $\ldots$ & $7a+1$ & $6a$ & $6a-1$ & $\ldots$ & $5a+1$\\ \hline $\vdots$ & $\vdots$ & $\ldots$ & $\ldots$ & $\vdots$ & $\vdots$ & $\ldots$ & $\ldots$ & $\vdots$\\ \hline $(4i-1)a$ & $(4i-1)a$ & $(4i-1)a-1$ & $\ldots$ & $(4i-2)a+1$ & $(4i-3)a$ & $(4i-3)a-1$ & $\ldots$ & $(4i-4)a+1$ \\ $4ia$ & $4ia$ & $4ia-1$ & $\ldots$ & $(4i-1)a+1$ & $(4i-2)a$ & $(4i-2)a-1$ & $\ldots$ & $(4i-3)a+1$\\ \hline $\vdots$ & $\vdots$ & $\ldots$ & $\ldots$ & $\vdots$ & $\vdots$ & $\ldots$ & $\ldots$ & $\vdots$\\ \hline $(4a-1)a$ & $(4a-1)a$ & $(4a-1)a-1$ & $\ldots$ & $(4a-2)a+1$ & $(4a-3)a$ & $(4a-3)a-1$ & $\ldots$ & $(4a-4)a+1$ \\ $4a^2$ & $4a^2$ & $4a^2-1$ & $\ldots$ & $(4a-1)a+1$ & $(4a-2)a$ & $(4a-2)a-1$ & $\ldots$ & $(4a-3)a+1$\\ \hline \end{tabular} \caption{Table showing $B_2-B_1$ for the construction in Theorem \ref{newZconstruction}} \label{tableNewZ} \end{table} \begin{example}\label{newZex} Using the construction of Theorem \ref{newZconstruction}, \begin{itemize} \item a $(17,2,4,1)$-SEDF in $\mathbb{Z}_{17}$ is given by $B_1=\{0,1,4,5\}$, $B_2=\{6,8,14,16\}$; \item a $(37,2,6,1)$-SEDF in $\mathbb{Z}_{37}$ is given by $B_1=\{0,1,2,6,7,8\}$, $B_2=\{9,12,21,24,33,36\}$. \end{itemize} \end{example} Observe that $k=2$ is excluded from the above theorem, since in this case we obtain a $(5,2,2,1)$-SEDF in $\mathbb{Z}_5$ with $B_1=\{0,2\}$, $B_2=\{3,4\}$. By Example \ref{equiv:ex}, this is equivalent to the $(5,2,2,1)$-SEDFs constructed using Proposition \ref{SEDFexistence} (1) and (2). We have the following recursive result for SEDFs with $\lambda=1$ in cyclic groups, which encompasses both previous constructions and provides a means of generating new examples. \begin{theorem}\label{recursive} Let $S=\{A_1, A_2\}$ be a $(k^2+1,2,k,1)$-SEDF in $\mathbb{Z}_{k^2+1}$ with $A_1=\{x_1, \ldots, x_k\}$ and $A_2=\{y_1, \ldots, y_k\}$, such that $x_i<y_j$ for all $1 \leq i,j \leq k$. Let $a \in \mathbb{N}$. Then we can obtain from $S$ an $((ak)^2+1, 2, ak,1)$-SEDF $S^{\prime}=\{B_1,B_2\}$ in $\mathbb{Z}_{(ak)^2+1}$. The blocks are given by \begin{itemize} \item $B_1= \bigcup_{i=1}^k \{ a x_i + \alpha: 0 \leq \alpha \leq a-1 \}$ \item $B_2=\bigcup_{i=1}^k \{ a(y_i+ k^2 \beta) : 0 \leq \beta \leq a-1 \}$. \end{itemize} (Here $x_i,y_i$ denote the elements of $\mathbb{Z}_{(ak)^2+1}$ with these labels.) \end{theorem} \begin{proof} For a fixed pair of indices $(i,j)$ ($1 \leq i,j \leq k$), consider the multiset of differences $$ D_{ij}=\{ ay_i - (ax_j + \alpha): 0 \leq \alpha \leq a-1 \}$$ i.e. $\{a(y_i-x_j)+ \alpha: 0 \leq \alpha \leq a-1 \} = \{a(y_i-x_j), a(y_i-x_j)-1, \ldots, a(y_i-x_j-1)+1 \}$. As $(i,j)$ runs through all possible pairs, $(y_i-x_j)$ takes each value $1, \ldots, k^2$ precisely once, since $S$ is a $(k^2+1,2,k,1)$-SEDF in $\mathbb{Z}_{k^2+1}$ and all differences $y_i-x_j$ lie in the set $\{1, \ldots, k^2\}$ since $x_i<y_j$ for all $1 \leq i,j \leq k$. Thus $\cup_{i,j} D_{ij}$ contains each element from $1$ to $ak^2$ in $\mathbb{Z}_{(ak)^2+1}$. Now, for given $(i,j)$, the multiset $\{ a(y_i+ k^2 \beta)-(a x_j+\alpha): 0 \leq \alpha \leq a-1\}= ak^2 \beta+D_{ij}$, and so for a fixed $\beta \in \{0, \ldots, a-1\}$, the union $\Delta_{\beta}$ of all these multisets as $(i,j)$ runs through all possible pairs yields all elements from $1+ak^2 \beta$ to $ak^2+ak^2 \beta=ak^2 (\beta+1)$ of $\mathbb{Z}_{(ak)^2+1}$, precisely once each. Finally, observe that the union $\cup _{\beta=0}^{a-1} \Delta_{\beta}$ comprises all elements of $\mathbb{Z}_{(ak)^2+1}$ from $1$ to $(ak)^2$, once each. Since this union equals $B_2-B_1$ (and $B_1-B_2$ consists of the negatives of these elements), we have established that $S^{\prime}$ has the required property. \end{proof} Table \ref{tableRecursive} illustrates part of the subtraction table for Theorem \ref{recursive}. The elements of $B_2$ have been re-ordered to follow the proof approach. \begin{table}[h] \centering \resizebox{\columnwidth}{!}{ \begin{tabular}{|c|cccc|c|cccc|} \hline & $a x_1$ & $a x_1+1$ & $\ldots$ & $ax_1+(a-1)$ & \ldots & $a x_k$ & $a x_k+1$ & $\ldots$ & $ax_k+(a-1)$ \\ \hline $ay_1$ & $a(y_1- x_1)$ & $a (y_1-x_1)-1$ & $\ldots$ & $a(y_1-x_1-1)+1$ & $\ldots$ & $a (y_1-x_k)$ & $a (y_1-x_k)-1$ & $\ldots$ & $a(y_1-x_k-1)+1$ \\ $\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ & $\ldots$ & $\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ \\ $ay_k$ & $a(y_k- x_1)$ & $a (y_k-x_1)-1$ & $\ldots$ & $a(y_k-x_1-1)+1$ & $\ldots$ & $a (y_k- x_k)$ & $a (y_k-x_k)-1$ & $\ldots$ & $a(y_k-x_k-1)+1$ \\ \hline $a(k^2+y_1)$ & $a(k^2+y_1- x_1)$ & $a (k^2+y_1-x_1)-1$ & $\ldots$ & $a(k^2+y_1-x_1-1)+1$ & $\ldots$ & $a (k^2+y_1-x_k)$ & $a (k^2+y_1-x_k)-1$ & $\ldots$ & $a(k^2+y_1-x_k-1)+1$ \\ $\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ & $\ldots$ & $\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ \\ $a(k^2+y_k)$ & $a(k^2+y_k- x_1)$ & $a (k^2+y_k-x_1)-1$ & $\ldots$ & $a(k^2+y_k-x_1-1)+1$ & $\ldots$ & $a (k^2+y_k- x_k)$ & $a (k^2+y_k-x_k)-1$ & $\ldots$ & $a(k^2+y_k-x_k-1)+1$ \\ \hline $\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ & $\ldots$ & $\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ \\ \hline $a((a-1)k^2+y_1)$ & $a((a-1)k^2+y_1- x_1)$ & $\ldots$ & $\ldots$ & $\ldots$ & $\ldots$ & $a ((a-1)k^2+y_1-x_k)$ & $\ldots$ & $\ldots$ & $\ldots$ \\ $\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ & $\ldots$ & $\vdots$ & $\vdots$ & $\vdots$ & $\vdots$ \\ $a((a-1)k^2+y_k)$ & $a((a-1)k^2+y_k- x_1)$ & $\ldots$ & $\ldots$ & $\ldots$ & $\ldots$ & $a ((a-1)k^2+y_k- x_k)$ & $\ldots$ & $\ldots$ & $\ldots$ \\ \hline \end{tabular} } \caption{Table showing $B_2-B_1$ for the construction in Theorem \ref{recursive}} \label{tableRecursive} \end{table} \begin{example} Using the construction of Theorem \ref{recursive}: \begin{itemize} \item Let $S$ be the trivial $(2,2,1,1)$-SEDF $A_1=\{0\}$, $A_2=\{1\}$ in $\mathbb{Z}_2$. Then $S^{\prime}$ is the $(a^2+1,2,a,1)$-SEDF $B_1=\{0,1,\ldots,a-1\}$, $B_2=\{a,2a,\ldots,a^2\}$ in $\mathbb{Z}_{a^2+1}$ of Proposition \ref{SEDFexistence} (1), from the paper \cite{PaSt}. \item If $S$ is the $(5,2,2,1)$-SEDF $\{0,1\}, \{2,4\}$ in $\mathbb{Z}_5$, then $S^{\prime}$ is the $(4a^2+1, 2, 2a, 1)$-SEDF in $\mathbb{Z}_{4a^2+1}$ from Proposition \ref{SEDFexistence} (1), whereas if $S$ is $\{0,2\}, \{3,4\}$ in $\mathbb{Z}_5$ then $S^{\prime}$ is the $(4a^2+1, 2, 2a, 1)$-SEDF in $\mathbb{Z}_{4a^2+1}$ from Theorem \ref{newZconstruction}. This shows that, if the recursion is performed using two equivalent SEDFs as $S$, the resulting $S^{\prime}$'s need not be equivalent. \end{itemize} \end{example} We may now prove the following result, which generalizes Theorem \ref{newZconstruction}. \begin{theorem}\label{kcomposite} For every $\mathbb{Z}_{k^2+1}$ with $k$ composite, there are at least two non-equivalent $(k^2+1, 2, k, 1)$-SEDFs. \end{theorem} \begin{proof} Since $k$ is composite, we can write $k=ra$ for some $a, r \geq 2$. Consider the following two $(r^2+1, 2, r, 1)$-SEDFs in $\mathbb{Z}_{r^2+1}$: \begin{itemize} \item $S_1=\{ \{0,1,\ldots, r-1\}, \{r, 2r, \ldots, r^2\} \}$; \item $S_2=rS_1=\{ \{0,r,\ldots, r^2-r\}, \{r^2-r+1, \ldots, r^2-1, r^2\} \}$. \end{itemize} Each of these satisfies the requirements of Theorem \ref{recursive}, so we may use the recursive construction of Theorem \ref{recursive} with $S_1$ and $S_2$, and $a(\geq 2)$, to obtain two $((ra)^2+1, 2, ra, 1)$-SEDFs (say $T_1$ and $T_2$ respectively). Although $S_1$ and $S_2$ are themselves equivalent, we prove that $T_1$ and $T_2$ are non-equivalent SEDFs in $\mathbb{Z}_{k^2+1}$. This generalizes the proof strategy of Theorem \ref{newZconstruction}. Applying Theorem \ref{recursive} to $S_1$ yields $T_1=\{B_1, B_2\}$, where $B_1=\{0,1,\ldots,ar-1\}$ and $B_2=\{ar, 2ar, \ldots, (ar)^2\}$. (This is the SEDF that would be obtained by setting $k=ar$ in the construction of Proposition \ref{SEDFexistence} (1).) Applying Theorem \ref{recursive} to $S_2$ yields $T_2=\{U_1, U_2\}$, where $U_1=\cup_{i=0}^{r-1} F_i$ and $F_i=\{iar, iar+1, \ldots, iar+(a-1)=a(ir+1)-1 \}$. Recall that the automorphisms of a cyclic group $\mathbb{Z}_n$ are of the form $x \mapsto cx$ where $c$ is a unit of $\mathbb{Z}_n$. By Definition \ref{equivalent}, $T_1$ is equivalent to $T_2$ if there exists a function $f(x)=\alpha x + \beta$, with $\alpha \in \mathbb{Z}_{k^2+1}^*$ (the group of units of $\mathbb{Z}_{k^2+1}$) and $\beta \in \mathbb{Z}_{k^2+1}$, such that $f(T_1)=T_2$. We prove that no such $f$ can exist. Observe that, since $a,r \geq 2$, each block of each SEDF contains at least four elements. Any such $f$ would have to map one of $B_1$ or $B_2$ to $U_1$. We claim that, in fact, it is sufficient to show that there is no $f$ which maps $B_1$ onto $U_1$. For, suppose we have established this claim, and suppose there is an $f$ which maps $B_2$ onto $U_1$. Observe that $B_2=ar(B_1+1)$. Since $f(B_2)=U_1$, we have $U_1=f(B_2)=\alpha B_2 + \beta = (\alpha ar) B_1 + (\alpha ar + \beta)$. Now, $ar$ is a unit of $\G$ (with inverse $-ar)$ and $\alpha$ is a unit by definition, so $\alpha ar$ is a unit. Thus there is a function $g(x)= \alpha ar x+ (\alpha ar+\beta)$ such that $\alpha ar$ is a unit, $\alpha ar+\beta$ is a group element and $g(B_1)=U_1$. But we have proved above that this cannot occur; a contradiction. So we now suppose, aiming for a contradiction, that there is an $f(x)= \alpha x+ \beta$ such that $f(B_1)=U_1=F_0 \cup F_1 \cup \cdots \cup F_{r-1}$. For any set $\{x, x+1, x+2\}$ of elements of $B_1$, their images under $f$ must satisfy $$ f(x+1)+f(x+1)=f(x)+f(x+2).$$ Note that all sums of pairs of elements of $U_1$ are strictly less than (under the specified ordering) $2ar^2-1$, i.e. less than $a^2r^2$ (since $a,r \geq 2$), and so do not reduce modulo $a^2 r^2+1$. Now, let $i,j,k$ be the indices such that $f(x+1) \in F_i$, $f(x) \in F_j$ and $f(x+2) \in F_k$ (here $i,j,k$ are integers with $0 \leq i,j,k, \leq r-1$). Then $$ 2iar \leq f(x+1)+f(x+1) \leq 2iar+2a-2< (2i+1)ar $$ since $ar \geq 2a> 2a-2$. Moreover, $$ (j+k)ar \leq f(x)+f(x+2) \leq (j+k)ar+2a-2< (j+k+1)ar. $$ Now, if $j+k \geq 2i+1$, then $f(x)+f(x+2) \geq (2i+1)ar$ whereas $f(x+1)+f(x+1) < (2i+1)ar$, and similarly if $j+k \leq 2i-1$ then $f(x)+f(x+2) < 2iar$ whereas $f(x+1)+f(x+1) \geq 2iar$. So we must have $j+k=2i$. If any two of $\{i,j,k\}$ are equal, then the third is equal to both. Hence there are two possibilities: $i=j=k$ or all of $\{i,j,k\}$ are distinct and form a three-term arithmetic progression. This latter case is possible only when $r>2$. We now consider the elements $\{0,1,2\} \subset B_1$. Let $f(0) \in F_j$, $f(1) \in F_i$ and $f(2) \in F_k$. If $i=j=k$, then all of $\{0,1,2\}$ are mapped to $F_i$ by $f$. Then $\{1,2,3\} \subset B_1$ has two of its members mapped to $F_i$ and hence also $f(3) \in F_i$. Repeating this process, we see that all $ar$ elements of $B_1$ are mapped to $F_i$, a contradiction since $|F_i|=a$ and $r \geq 2$. If $r=2$, the proof is complete. Otherwise, for $r \geq 3$, it must be the case that all of $\{i,j,k\}$ are distinct. Since $2i=j+k$, one of $\{j,k\}$ must be less than $i$ and the other greater than $i$. Suppose first that $k<i<j$. Let $f(3) \in F_m$ and consider $\{1,2,3\}$; we have $i+m=2k$ and $\{i,k\}$ are distinct. Thus all three of $\{i,k,m\}$ are distinct and $j-i=i-k=k-m$, ie $m<k<i<j$. Repeating for all $0,1,\ldots, ar-1 \in B_1$, we deduce that all $ar$ elements must lie in distinct $F_i$'s. But there are only $r$ possible indices, and $a \geq 2$, so this is impossible. The situation when $j<i<k$ is precisely analogous. Thus all cases lead to a contradiction, and so there does not exist a function $f$ such that $f(B_1)=U_1$. This completes the proof that $T_1$ and $T_2$ are non-equivalent. \end{proof} In fact, the recursive process of Theorem \ref{recursive} can be performed for GSEDFs with $m=2$ and $\lambda_1=\lambda_2=1$, and by appropriate choice of parameters we can build new SEDFs using GSEDFs which are not themselves SEDFs. A recursive construction for GSEDFs was given in \cite{LuNiCa}; our result differs from this in that the value of $\lambda$ remains unchanged, and the new group is a larger cyclic group rather than a cross-product. \begin{theorem}\label{GSEDFrecursive} Let $S=\{A_1, A_2\}$ be a $(st+1,2; s,t; 1,1)$-GSEDF in $\mathbb{Z}_{st+1}$ with $A_1=\{x_1, \ldots, x_s\}$ and $A_2=\{y_1, \ldots, y_t\}$, such that $x_i<y_j$ for all $1 \leq i \leq s$ and $1 \leq j \leq t$. Let $a,b \in \mathbb{N}$. Then we can obtain from $S$ an $(abst+1, 2; as, bt; 1,1)$-GSEDF $S^{\prime}=\{B_1,B_2\}$ in $\mathbb{Z}_{abst+1}$. The blocks are given by \begin{itemize} \item $B_1= \cup_{i=1}^s \{ a x_i + \alpha: 0 \leq \alpha \leq a-1 \}$ \item $B_2=\cup_{i=1}^t \{ a(y_i+ \beta st) : 0 \leq \beta \leq b-1 \}$. \end{itemize} (here $x_i,y_i$ denote the elements of $\mathbb{Z}_{abst+1}$ with these labels). \end{theorem} \begin{proof} The proof is precisely analogous to that of Theorem \ref{recursive}. \end{proof} \begin{corollary} Let $k=as=bt$ for some $a,b,s,t \in \mathbb{N}$, such that there exists an $(st+1,2; s,t; 1,1)$-GSEDF $\{A_1, A_2\}$ in $\mathbb{Z}_{st+1}$ with $x<y$ for all $x \in A_1$ and $y \in A_2$. Then an $(k^2+1, 2; k,1)$-SEDF can be obtained from $S$. \end{corollary} An example of a GSEDF construction which can be used in the recursion is given by $S=(\{0,1\ldots,s-1\}, \{s, 2s, \ldots, ts \})$, presented in Theorem 4.4 of \cite{LuNiCa}. The equivalent GSEDF $T=(\{0,t, \ldots (s-1)t\}, \{(s-1)t+1, (s-1)t+2, \ldots, st\}$ (obtained via the automorphism $x \mapsto tx$) may also be used. There is a very wide range of specific SEDF constructions obtainable from Theorem \ref{GSEDFrecursive}. We illustrate with a specific example. \begin{example} \begin{itemize} \item[(i)] Taking $s=2$ and $t=3$ with $T=(\{0,3\}, \{4,5,6\})$ in $\mathbb{Z}_7$, we obtain the following general construction in $\mathbb{Z}_{k^2+1}$ for any $k$ such that $k=2a=3b$ for some $a,b \in \mathbb{N}$, ie for any $k$ which is a multiple of $6$: \begin{itemize} \item $B_1=\{0,1,\ldots, a-1\} \cup \{3a,3a+1, \ldots, 4a\}$ \item $B_2=\{4a,5a,6a \} \cup \{10a, 11a, 12a\} \cup \cdots \cup \{(6b-2)a, (6b-1)a , 6ab\}$ \end{itemize} \item[(ii)] Taking $k=12$, i.e. $a=6$ and $b=4$, we obtain the following $(145,2,12,1)$-SEDF in $\mathbb{Z}_{145}$: \begin{itemize} \item $B_1=\{0,1,2,3,4,5,18,19,20,21,22,23\}$ \item $B_2=\{24,30,36,60,66,72,96,102,108,132,138,144\}$ \end{itemize} \end{itemize} \end{example} \section{Non-abelian SEDFs with $\lambda=1$} In the non-abelian setting, it is not known whether every $(n,m,k,1)$-SEDF must be a $(k^2+1, 2, k, 1)$-SEDF. The forward direction of Proposition \ref{PaStlambda=1} does not apply here, so it is possible that there may exist non-abelian $(k^2(m-1)+1, m, k, 1)$-SEDFs with $m>2$, although no examples are currently known. However, $(k^2+1, 2, k,1)$ is an admissible parameter set. Unlike in the abelian case, there does not exist a non-abelian group of order $k^2+1$ for every value of $k$ - for example there exists no non-abelian group of order $k^2+1$ for $k \in \{2,4,6,8,10 \}$. In this section, we exhibit an infinite family of non-abelian $(k^2+1, 2, k, 1)$-SEDFs for $k$ odd. We will consider the dihedral group $D_n$ of order $n$ as being generated by the elements $s$ (reflection) and $r$ (rotation by $\frac{4 \pi}{n}$), so that $D_n=\{ s^i r^j: 0 \leq i \leq 1, 0 \leq j < \frac{n}{2} \}$ and the generators satisfy $s^2=1$, $r^{\frac{n}{2}}=1$ and $sr=r^{-1}s$. First, we present two examples, which illustrate the general construction to follow. \begin{example}\label{dihedralex} \begin{itemize} \item[(i)] Let $k=3$ and $\G=D_{10}$; take $A_1=\{e,s,r\}$ and $A_2=\{sr, r^3, sr^4\}$. \item[(ii)] Let $k=5$ and $\G=D_{26}$; take $A_1=\{e,s,r,sr,r^2\}$ and $A_2=\{sr^2, r^5, sr^7,r^{10}, sr^{12}\}$. \end{itemize} \vspace{5mm} Table \ref{table2} demonstrates the multiset of external differences for the $(26, 2, 5, 1)$-SEDF over $D_{26}$; the entry in the row labelled by $x$ and the column labelled by $y$ is $xy^{-1}$. The sets have been ordered to emphasise the structure. \begin{table}[h] \renewcommand{\arraystretch}{1.2} \centering \setlength\tabcolsep{3pt} \begin{tabular}{|c|ccc:cc|cc:ccc|} \hline & $e$ & $r$ & $r^2$ & $s$ & $sr$ & $r^5$ & $r^{10}$ & $sr^2$ & $sr^7$ & $sr^{12}$ \\ \hline $e$ & & & & & & $r^8$ & $r^3$ & $sr^2$ & $sr^7$ & $sr^{12}$ \\ $r$ & & & & & & $r^9$ & $r^4$ & $sr$ & $sr^6$ & $sr^{11}$ \\ $r^2$ & & & & & & $r^{10}$ & $r^5$ & $s$ & $sr^5$ & $sr^4$ \\ \hdashline $s$ & & & & & & $sr^8$ & $sr^3$ & $r^2$ & $r^7$ & $r^{12}$\\ $sr$ & & & & & & $sr^9$ & $sr^4$ & $r$ & $r^6$ & $r^{11}$ \\ \hline $r^5$ & $r^5$ & $r^4$ & $r^3$ & $sr^8$ & $sr^9$ & & & & & \\ $r^{10}$ & $r^{10}$ & $r^9$ & $r^8$ & $sr^3$ & $sr^4$ & & & & & \\ \hdashline $sr^2$& $sr^2$ & $sr$ & $s$ & $r^{11}$ & $r^{12}$ & & & & & \\ $sr^7$ & $sr^7$ & $sr^6$ & $sr^5$ & $r^6$ & $r^7$ & & & & & \\ $sr^{12}$ & $sr^{12}$ & $sr^{11}$ & $sr^{10}$ & $r$ & $r^2$ & & & & & \\ \hline \end{tabular} \caption{Table of $(26, 2, 5, 1)$-SEDF over $D_{26}$} \label{table2} \end{table} \end{example} \begin{theorem}\label{thm:dihedral} Let $k>1$ be odd. Let $\G=D_{k^2+1}$, the dihedral group of order $n=k^2+1$. There exists a $(k^2+1,2,k,1)$-SEDF in $\G$. \\ Specifically, $\A=\{ A_1, A_2 \}$ is a $(k^2+1,2,k,1)$-SEDF where \begin{itemize} \item $A_1=\{e,s, r, sr, r^2, sr^2,\ldots, r^{\frac{k-1}{2}} \}$, \mbox{i.e. } $\{r^i: 0 \leq i \leq \frac{k-1}{2}\} \cup \{sr^j: 0 \leq j \leq \frac{k-3}{2} \}$. \item $A_2=\{ sr^{\frac{k-1}{2}}, r^k, sr^{\frac{k-1}{2}} r^k, r^{2k}, sr^{\frac{k-1}{2}} r^{2k}, \ldots, r^{\frac{k(k-1)}{2}}, sr^{\frac{k-1}{2}}r^{\frac{k(k-1)}{2}} \}$,\\ \mbox{ i.e. } $\{r^{ik}: 1 \leq i \leq \frac{k-1}{2} \} \cup \{ sr^{jk+\frac{k-1}{2}}: 0 \leq j \leq \frac{k-1}{2} \}$. \end{itemize} \end{theorem} \begin{proof} For disjoint subsets $X,Y$ of a group $\G$, let $\Delta(X,Y)=\{xy^{-1}:x \in X, y \in Y\}$. It suffices to show that the multiset $\Delta(A_1,A_2)$ comprises each non-identity group element precisely once, since the multiset $\Delta(A_2,A_1)$ consists of the inverses of this multiset. For $x,y \in \G$, $xy^{-1}$ equals: \begin{itemize} \item $r^{i-j}$ if $x=r^i$ and $y=r^j$; \item $sr^{i-j}$ if $x=sr^i$ and $y=r^j$; \item $r^{j-i}$ if $x=sr^i$ and $y=sr^j$; \item $sr^{j-i}$ if $x=r^i$ and $y=sr^j$. \end{itemize} Note that $r^{(\frac{k^2+1}{2})}=e$. Obtaining the elements $r^i$ ($1 \leq i \leq \frac{k^2-1}{2}$) once each is equivalent to showing that in the additive group $(\mathbb{Z}_{\frac{k^2+1}{2}}, +)$, the multiset of differences between the pair of sets $\{0,1,\ldots,\frac{k-1}{2} \}$ and $\{k, 2k, \ldots, \frac{k(k-1)}{2} \}$ and the multiset of differences between the pair of sets $\{\frac{k-1}{2}, \frac{3k-1}{2} ,\ldots,\frac{k^2-1}{2} \}$ and $\{0, 1, \ldots, \frac{k-3}{2} \}$ together comprise each non-zero element of $\mathbb{Z}_{\frac{k^2+1}{2}}$ once each. It can be verified that this is the case: \begin{itemize} \item $(jk+\frac{k-1}{2})-\{0,1,\ldots,\frac{k-3}{2} \}=\{jk+1, jk+2, \ldots, jk+\frac{k-1}{2}\}$ where $0 \leq j \leq \frac{k-1}{2}$; \item $ \{0, 1, \ldots, \frac{k-1}{2} \} - k(\frac{k-1}{2}-j) = \{jk+\frac{k+1}{2}, jk+\frac{k+3}{2}, \ldots, jk+k \}$ where $0 \leq j \leq \frac{k-3}{2}$. \end{itemize} For the second part, observe that $-k(\frac{k-1}{2}-j)=jk+(\frac{k+1}{2})$. Obtaining the elements $sr^i$ ($0 \leq i \leq \frac{k^2-1}{2}$) once each corresponds to showing that the multiset of differences between the pair of sets $\{0, 1, \ldots, \frac{k-3}{2} \}$ and $\{k, 2k, \ldots, \frac{k(k-1)}{2} \}$ and the multiset of differences between the pair $\{\frac{k-1}{2}, \frac{3k-1}{2} ,\ldots,\frac{k^2-1}{2} \}$ and $\{0,1,\ldots,\frac{k-1}{2} \}$ together comprise all the elements of $\mathbb{Z}_{\frac{k^2+1}{2}}$ once each: \begin{itemize} \item $(jk+\frac{k-1}{2})-\{0,1,\ldots,\frac{k-1}{2} \}=\{jk, jk+1, \ldots, jk+\frac{k-1}{2}\}$ where $0 \leq j \leq \frac{k-1}{2}$; \item $ \{0, 1, \ldots, \frac{k-3}{2} \} - k(\frac{k-1}{2}-j)=\{jk+\frac{k+1}{2}, jk+\frac{k+3}{2}, \ldots, jk+(k-1) \}$ where $0 \leq j \leq \frac{k-3}{2}$. \end{itemize} The proof is demonstrated in Table \ref{table3}, which displays the multiset of external differences $\Delta(A_1,A_2)$. The entry in the row labelled by $x$ and the column labelled by $y$ is $xy^{-1}$. Each non-identity group element occurs precisely once in this table. \begin{table}[h] \renewcommand{\arraystretch}{1.9} \centering \setlength\tabcolsep{3pt} \begin{tabular}{|c|ccccc:ccccc|} \hline & $r^k$ & $r^{2k}$ & \ldots & $r^{k(\frac{k-3}{2})}$ & $r^{k(\frac{k-1}{2})}$ & $sr^{(\frac{k-1}{2})}$ & $sr^{(\frac{3k-1}{2})}$ & \ldots & $sr^{(\frac{(k-2)k-1}{2})}$ & $sr^{(\frac{k^2-1}{2})}$ \\ \hline $r^0=e$ & $r^{(\frac{k^2-2k+1}{2})}$ & $r^{(\frac{k^2-4k+1}{2})}$ & \ldots & $r^{(\frac{3k+1}{2})}$ & $r^{(\frac{k+1}{2})}$ & $sr^{(\frac{k-1}{2})}$ & $sr^{(\frac{3k-1}{2})}$ & \ldots & $sr^{(\frac{(k-2)k-1}{2})}$ & $sr^{(\frac{k^2-1}{2})}$ \\ $r$ & $r^{(\frac{k^2-2k+3}{2})}$ & \ldots & \ldots & \ldots & $r^{(\frac{k+3}{2})}$ & $sr^{(\frac{k-3}{2})}$ & \ldots & \ldots & \ldots & $sr^{(\frac{k^2-3}{2})}$ \\ $\vdots$ & \vdots & \ldots & \ldots & \ldots & \vdots & \vdots & \ldots& \ldots & \ldots & \vdots \\ $r^{(\frac{k-5}{2})}$ & $r^{k(\frac{k-1}{2})-2}$ & \ldots & \ldots & $r^{2k-2}$ & $r^{k-2}$ & \vdots & \ldots & \ldots & \ldots & \vdots \\ $r^{(\frac{k-3}{2})}$ & $r^{k(\frac{k-1}{2})-1}$ & $r^{k(\frac{k-3}{2})-1}$ & \ldots & $r^{2k-1}$ & $r^{k-1}$ & $sr$ & $sr^{k+1}$ & \ldots & \ldots & $sr^{(\frac{k(k-1)}{2}+1)}$\\ $r^{(\frac{k-1}{2})}$ & $r^{k(\frac{k-1}{2})}$ & $r^{k(\frac{k-3}{2})}$ & \ldots & $r^{2k}$& $r^k$ & $s$ & $sr^k$ & \ldots & \ldots & $sr^{(\frac{k(k-1)}{2})}$ \\ \hdashline $sr^0=s$ & $sr^{(\frac{k^2-2k+1}{2})}$ & $sr^{(\frac{k^2-4k+1}{2})}$ & \ldots & $sr^{(\frac{3k+1}{2})}$ & $sr^{(\frac{k+1}{2})}$ & $r^{(\frac{k-1}{2})}$ & $r^{(\frac{3k-1}{2})}$ & \ldots & $r^{(\frac{(k-2)k-1}{2})}$ & $r^{(\frac{k^2-1}{2})}$ \\ $sr$ & $sr^{(\frac{k^2-2k+3}{2})}$ & \ldots & \ldots & \ldots & $sr^{(\frac{k+3}{2})}$ & $r^{(\frac{k-3}{2})}$ & $r^{(\frac{3k-3}{2})}$ & \ldots & \ldots & $ r^{(\frac{k^2-3}{2})}$ \\ $\vdots$ & \vdots & \ldots & \ldots & \ldots & \vdots & \vdots & \ldots & \ldots & \ldots & \vdots\\ $sr^{(\frac{k-5}{2})}$ & $sr^{k(\frac{k-1}{2})-2}$ & \ldots & \ldots & $sr^{2k-2}$ & $sr^{k-2}$ &$ r^2 $ & $r^{k+2}$ & \ldots & \ldots & $r^{(\frac{k(k-1)}{2}+2)}$ \\ $sr^{(\frac{k-3}{2})}$ & $sr^{k(\frac{k-1}{2})-1}$ & $sr^{k(\frac{k-3}{2})-1}$ & \ldots & $sr^{2k-1}$ & $sr^{k-1}$ & $r$ & $r^{k+1}$ & \ldots & $r^{(\frac{k(k-3)}{2}+1)}$ & $r^{(\frac{k(k-1)}{2}+1)}$ \\ \hline \end{tabular} \caption{Table showing $\Delta(A_1,A_2)$ for the $(k^2+1, 2, k, 1)$-SEDF over $D_{k^2+1}$ of Theorem \ref{thm:dihedral}} \label{table3} \end{table} \end{proof} We are now able to prove the following result. \begin{corollary} For every $k>2$, there exist at least two non-equivalent $(k^2+1, 2, k, 1)$-SEDFs. \end{corollary} \begin{proof} If $k$ is odd, we may take the construction of Proposition \ref{SEDFexistence}(1) in $\mathbb{Z}_{k^2+1}$ and the construction of Theorem \ref{thm:dihedral} in $D_{k^2+1}$. These are non-equivalent as the groups are not isomorphic. Otherwise, $k$ is even and $k \geq 4$, i.e. $k=2a$ for some $a \geq 2$, and so two non-equivalent constructions in $\mathbb{Z}_{k^2+1}$ are guaranteed by Theorem \ref{kcomposite}. \end{proof} We end this section with a first step towards addressing the question: are the only non-trivial SEDFs with $\lambda=1$ those with $m=2$? \begin{definition} Let $\G$ be a group of order $n$. \begin{itemize} \item Let $\A$ be a set of $m \geq 2$ disjoint $k$-subsets of $\G$, say $A_1, \dots , A_m$. We say that $\A$ is an $(n,m,k,\lambda)$-coEDF if the multiset \begin{equation*} N=\{ y^{-1}x : x \in A_i, y \in A_j, i \neq j \} \end{equation*} comprises $\lambda$ occurrences of each nonzero element of $\G$. \item Let $\A$ be a set of $m \geq 2$ disjoint $k$-subsets of $\G$, say $A_1, \dots , A_m$. We say that $\A$ is an $(n,m,k,\lambda)$-coSEDF if, for every $i$, $1 \leq i \leq m$, the multiset \begin{equation*} N_i=\{ y^{-1}x: x \in A_i, y \in \cup_{j \neq i} A_j\} \end{equation*} comprises $\lambda$ occurrences of each nonzero element of $\G$. \end{itemize} \end{definition} Clearly if $\G$ is an abelian group, coEDFs and coSEDFs coincide precisely with EDFs and SEDFs. An $(n,m,k,\lambda)$-coSEDF must satisfy precisely the same parameter conditions as an $(n,m,k,\lambda)$-SEDF. For a set $X$ in a group $\G$, we denote $X^{-1}=\{x^{-1}: x \in X \}$. For a collection $\A$ of sets $\{A_1, \ldots, A_m \}$ in a group $\G$, we denote $\A^{-1}=\{ A_1^{-1}, \ldots, A_m^{-1}\}$. The following result is immediate. \begin{lemma} $\A=\{A_1, \ldots, A_m \}$ is an $(n,m,k,\lambda)$-EDF (respectively, SEDF) if and only if $\A^{-1}=\{A_1^{-1}, \ldots, A_m^{-1} \}$ is an $(n,m,k,\lambda)$-coEDF (respectively, coSEDF). \end{lemma} \begin{example}\label{ex:dihedralex} The $(10,2,3,1)$-SEDF $\A$ from Example \ref{dihedralex} in $D_{10}$, with sets $A_1=\{e,s,r\}$ and $A_2=\{sr,r^3,sr^4\}$ is also a $(10,2,3,1)$-coSEDF. Unlike in the abelian case, the subtraction tables corresponding to the SEDF and coSEDF are different. Here, $\A^{-1}$ comprises sets $A_1^{-1}=\{e,s,r^4\}$ and $A_2^{-1}=\{sr, r^2, sr^4\}$. It can be checked directly that this is also a $(10,2,3,1)$-SEDF in $D_{10}$. \end{example} \begin{proposition} Let $\G$ be a group of order $n$. Suppose that $\A=\{A_1, \ldots, A_m \}$ is an $(n,m,k,1)$-SEDF and an $(n,m,k,1)$-coSEDF (equivalently, that $\A$ and $\A^{-1}$ are both $(n,m,k,1)$-SEDFs). Then either $m=2$ or $k=1$. \end{proposition} \begin{proof} Suppose that $m>2$ and $k>1$. Let $N^{\prime}_{ij}=\{ y^{-1}x: x \in A_i, y \in A_j \}$. The multiset union $$\bigcup_{1 \leq i \leq m, \, 1 \leq j \leq n} N^{\prime}_{ij}$$ comprises every non-identity element of $\G$ precisely $m$ times. The multiset union $$\bigcup_{i \neq 1, \, j \neq 1} N^{\prime}_{ij}$$ comprises $m-2$ copies of each non-identity element of $\G$, since $\bigcup_{j \neq 1} N^{\prime}_{1j}(=N_1)$ comprises each non-identity element once, as does $\bigcup_{i \neq 1} N^{\prime}_{i1}$ (the set of inverses of $N_1$). Let $x \neq y \in A_1$ (this is possible since $k>1$). Set $\alpha=y^{-1}x$. Since $\alpha$ is a non-identity element of $\G$ and $\A$ is an $(n,m,k,1)$-coSEDF in $\G$, there exist $u \in A_i$ and $v \in A_j$ for some $i, j \in \{2, \ldots, m\}$ with $i \neq j$ such that $v^{-1}u=\alpha$ (from above, this is possible since $m>2$). Then $v^{-1}u=\alpha=y^{-1}x$, which rearranges to $v^{-1}=y^{-1}x u^{-1}$, i.e. $y v^{-1}=x u^{-1}$. However, $\A$ is an $(n,m,k,1)$-SEDF, and this equality corresponds to two equal distinct external differences arising out of $A_1$ - a contradiction. \end{proof} This result generalizes the forward direction of Proposition \ref{PaStlambda=1}. It indicates that, if a non-trivial non-abelian $(n,m,k,1)$-SEDF exists with $m>2$, it cannot be a $(n,m,k,1)$-coSEDF. \section{Computational approach and results} In this section, we discuss the computational approach which allowed us to find, and classify, all SEDFs in groups up to order 24. \subsection{Algorithm description and analysis} \begin{algorithm} \caption{SEDF Finding} \label{sedf:checkpart}\begin{algorithmic}[1] \Procedure{CheckPartial}{$G,n,m,\lambda, L$} \For{$i1 \in [1..m]$} \State $Count \gets [0,0,\dots,0]$ \Comment{$Count$ is length $n$} \For{$i2 \in [1..i1-1,i1+1..m]$} \For{$j1 \in L[i1]$} \For{$j2 \in L[i2]$} \State$Count[j1*j2^{-1}]\ \mathrel{{+}{=}}\ 1$ \If{$Count[j1*j2^{-1}] > \lambda$} \State\Return \textsc{False} \EndIf \EndFor \EndFor \EndFor \EndFor \State \Return \textsc{True} \EndProcedure \end{algorithmic} \label{sedf:search}\begin{algorithmic}[1] \Procedure{SEDFSearch}{$G,n,m,k,\lambda,L,p$} \If{\textbf{not} \textsc{CheckPartial}$(G,n,m,\lambda,L)$} \State\Return \Comment{Too many occurrences of some difference} \EndIf \If{$m*k-(\sum s \in L. |s|) > n - p$} \State\Return \Comment{Not enough values remain to fill the SEDF} \EndIf \If{$\forall s \in L. |s| = k$} \State\textbf{Output}$(L)$ \Comment{Found a valid, complete SEDF} \State\Return \EndIf \State{\textsc{SEDFSearch}$(G,n,m,k,\lambda,L,p+1)$} \Comment{Try SEDFs without G[p]} \For{$i \in [1..m]$} \If{$|L[i]| = k$} \State\textbf{Continue} \Comment{Continue on to next value for $i$} \EndIf \State{$L^{\prime} := L$}\Comment{Make copy of $L$} \State{Add($L^{\prime}[i], G[p])$} \State{\textsc{SEDFSearch}$(G,n,m,k,\lambda,L',p+1)$} \If{$|L[i]| = 0$} \State\Return \Comment{Only try adding a value to one empty set} \EndIf \EndFor \EndProcedure \end{algorithmic} \label{sedf:search}\begin{algorithmic}[1] \Procedure{SEDFSearchStart}{$G,n,m,k,\lambda$} \Comment{$G$ is interpreted as a list, with identity first} \State{$L \gets [\{G[1]\},\{\},\dots,\{\}]$} \Comment{$L$ has length $m$} \State{\textsc{SEDFSearch}{$(G,n,m,k,\lambda,L,2)$}} \EndProcedure \end{algorithmic} \end{algorithm} Computational search results were obtained via a constraint-style backtrack search. Preprocessing of the group information was performed using GAP (\cite{GAP}), in particular its \textit{SmallGroups} library, while the search for SEDFs was implemented in Java, using a recursive depth-first search algorithm Algorithm \ref{sedf:search} shows how we search for SEDFs. This algorithm will return all \((n,m,k,\lambda)\)-SEDFs in a group $G$. This algorithm assumes that $\lambda(n-1) = k^2(m-1)$ and that the elements of \(G\) are stored as a list, with the identity of \(G\) as the first element. As \(G\) is a list we will use the notation \(G[i]\) to access the \(i^{th}\) member of \(G\). We order the elements of $G$ using their position in this list. We build up the SEDF as a list of sets called \(L\). We build it one value at a time, by inserting elements of \(G\) into the members of \(L\). The algorithm begins with \textsc{SEDFSearchStart}. This function partially breaks the symmetry of left-shifting and right-shifting SEDFs, by requiring that the identity is always in the SEDF. The function \textsc{SEDFSearch} builds the SEDF \(L\) incrementally. This function accepts a partially constructed SEDF \(L\), and a current position \(p\) in the group \(G\). This function will find all SEDFs which can be made by adding values from \(G[p]\) onwards to \(L\). \textsc{SEDFSearch} begins by trying to prove it is impossible to extend \(L\) to a valid SEDF. Firstly Line 2 calls \textsc{CheckPartial}, which calculates the differences between each pair of sets in \(L\) -- if any value occurs more than \(\lambda\) times we can reject \(L\). Line 4 then checks whether there are enough remaining values which could be inserted into \(L\) to complete the SEDF, and rejects if not. In Line 6 we check if each element of \(L\) contains \(k\) values. In this case we have a complete candidate SEDF. As we know that $\lambda(n-1) = k^2(m-1)$ and in \textsc{CheckPartial} we calculated $k^2(m-1)$ differences and counted at most $ \lambda$ occurrences of each of the \(n-1\) non-identity elements of \(G\), then we must have counted exactly \(\lambda\) differences for each non-identity element, and therefore found a SEDF. If we reach Line 8, we know we have a partially built SEDF, $L$. Firstly we try recursively building SEDFs which extend \(L\) and do not contain \(G[p]\). After that, we try building SEDFs where we add \(G[p]\) to each element of \(L\) which does not yet contain \(k\) elements. One concern here is that we do not want to produce the same SEDF multiple times, where the only difference is that the elements of \(L\) are in a different order. Lemma \ref{lem:order} shows that by ensuring \(L\) is always ordered, we avoid producing the same SEDF multiple times. Line 16 gurantees that we only try adding a value to an empty element of \(L\) once per call to \textsc{SEDFSearch}. \begin{lemma}\label{lem:order} During the execution of \textsc{SEDFSearchStart}, the argument \(L\) of \\ \( \textsc{SEDFSearch}(G,n,m,k,\lambda,L,p) \) satisfies the conditions that \begin{itemize} \item there exists some \(i \in \{1, \ldots, m\}\) such that \(\{ j: L[j] \neq \emptyset\} = \{1,\ldots, i\}\) and \(\{ j: L[j]=\emptyset\} = \{i+1,\ldots, m\}\), and \item \(\forall j \in \{1, \ldots, i-1\}. min(L[j]) < min(L[j+1])\) \end{itemize} by ordering the elements of \(G\) (which we consider as a list) under the order \(x < y\) if \(x\) comes before \(y\). \end{lemma} \begin{proof} We will proceed by induction. The first time \textsc{SEDFSearch} is called from \textsc{SEDFSearchStart}, only the first element of \(L\) is not the empty set, so the conditions follow trivially. The first time \textsc{SEDFSearch} recursively calls itself on Line 9 it passes the same value for \(L\). The subsequent recursive calls on Line 15 pass \(L\) with \(G[p]\) added to one member of \(L\). If \(G[p]\) is inserted into a non-empty member of \(L\), then since this value is larger than any value already in \(L\), it will not affect the minimum, and the conditions hold. Otherwise, \(G[p]\) is added to an empty member of \(L\). This will be done for the first empty element of \(L\), where \(G[p]\) will be larger than any other value in any member of \(L\), and so the conditions are again satisfied. After this, Line 16 will cause \textsc{SEDFSearch} to immediately return and not insert \(G[p]\) into any later elements of \(L\). \end{proof} After \textsc{SEDFSearchState} has produced a list of SEDFs, we produce a final list of non-equivalent SEDFs using the Images package \cite{JeJoPfWa} in GAP. \subsection{Summary of results} Table \ref{table4} lists all admissible parameters sets for groups of order at most $24$. \begin{table}[h] \begin{center} \begin{tabular}{c|c|cc|cc|cc|ccccc|cccc|cc} $n$ & 5 & 9 & & 10 & & 13 & & 17 & & & & & 19 & & & & 21 & \\ \hline $m$ & 2 & 2& 3 & 2& 3 & 2& 4& 2& 2& 3& 4& 5& 2& 3& 3& 5& 2& 6 \\ \hline $k$ & 2 & 4& 2 & 3& 3 & 6& 2& 4& 8& 4& 4& 2& 6& 3& 6& 3& 10& 2 \\ \hline $\lambda$ & 1 & 2& 1 & 1& 2 & 3& 1& 1& 4& 2& 3& 1& 2& 1& 4& 2& 5& 1 \\ \end{tabular} \caption{All admissible SEDF parameters for $n \leq 24$} \label{table4} \end{center} \end{table} For abelian groups, the following parameter sets are ruled-out by results from the literature: \begin{itemize} \item Proposition \ref{nonexistence}(1): $(9,3,2,1)$, $(10,3,3,2)$, $(13,4,2,1)$, $(17,3,4,2)$, $(17,4,4,3)$, $(19,3,3,1)$, $(19,3,6,4)$; \item Proposition \ref{nonexistence}(2): $(17,5,2,1)$, $(19,5,3,2)$; \item Proposition \ref{PaStlambda=1}: $(21,6,2,1)$; \item Theorem 4.3 of \cite{JeLi}: $(19,2,6,2)$, $(21,2,10,5)$. \end{itemize} Table \ref{table5} summarizes all non-equivalent abelian SEDFS found by computer search in groups of order up to $24$. All found SEDFs can be classified in terms of constructions in the literature; the cases are as follows: \begin{itemize} \item[(a)] Proposition \ref{SEDFexistence} (1): construction $\{0,1,\ldots,k-1\}$, $\{k,2k, \ldots, k^2 \}$ given in \cite{PaSt}; \item[(b)] Proposition \ref{SEDFexistence} (2): construction corresponding to Paley difference sets, given in \cite{HuPa}: for prime power $n=q$, take the squares/non-squares of the multiplicative group of $\mathbb{F}_q$; \item[(c)] Proposition \ref{SEDFexistence} (3): construction using cyclotomic classes in the multiplicative group of $\mathbb{F}_q$, given in \cite{BaJiWeZh}. \item[(d)] Theorem \ref{newZconstruction}: construction $\{0,1, \ldots, a-1\} \cup \{2a, 2a+1, \ldots 3a-1\}$, $\cup_{i=1}^a \{ (4i-1)a, 4ia\}$. \end{itemize} \begin{table}[ht] \renewcommand{\arraystretch}{1.7} \begin{center} \begin{tabular}{c|c|c|c|c} Parameters &{ Abelian group} & No of non-equiv. SEDFs& Example & Case \\ \hline $(5, 2 ,2 , 1 )$ & $\mathbb{Z}_5$ & 1 & $\{0,1\}, \{2,4\}$ & (a), (b), (d) \\ \hline \multirow{2}{*}{$(9,2,4,2)$}& $\mathbb{Z}_9$ & 0 & - & - \\ & $\mathbb{Z}_ 3 \times \mathbb{Z}_3$ & 1 & \makecell {$\{(1,0),(0,1,)(2,0),(0,2)\},$ \\ $ \{(1,1),(1,2),(2,1),(2,2)\}$} & (b) \\ \hline $(10, 2 ,3 , 1 )$ & $\mathbb{Z}_{10} $ & 1 & $ \{0,1,2\}, \{3,6,9\}$ & (a) \\ \hline $(13, 2 , 6 ,3)$ & $\mathbb{Z}_{13}$ & 1 & \makecell {$\{1,3,4,9,10,12\},$ \\ $ \{ 2,5,6,7,8,11 \}$} & (b) \\ \hline $(17, 2 , 4 ,1)$ & $\mathbb{Z}_{17}$ & 2 & \makecell {$\{0,1,2,3\},\{4,8,12,16\}$ \\ $ \{1,4,13,16\},\{2,8,9,15 \}$} & \makecell {(a) \\ (c), (d)} \\ \hline $(17, 2 , 8 ,4)$ & $\mathbb{Z}_{17}$ & 1 & \makecell {$\{1,2,4,8,9,13,15,16\},$ \\ $ \{ 3,5,6,7,10,11,12,14 \}$} & (b) \\ \end{tabular} \caption{Non-equivalent SEDFs in abelian groups of order up to $24$} \label{table5} \end{center} \end{table} There is no $(9,2,4,2)$-SEDF in the cyclic group of order $9$; this is ruled-out by Theorem 4.2 of \cite{JeLi}. The two non-equivalent $(17,2,4,1)$-SEDFs are those guaranteed by Theorem \ref{kcomposite}; here the SEDF from the cyclotomic construction of Proposition \ref{SEDFexistence} (3) is equivalent to that from Theorem \ref{newZconstruction} (see Example \ref{equiv:ex}). For non-abelian groups of order $n \leq 24$, we cannot rule-out any of the parameter sets from Table \ref{table4} using results from the existing literature. However, there are no non-abelian groups of orders 5, 9, 13, 17, or 19. Therefore non-trivial SEDFs in groups of order up to $24$ are possible only for two orders, $10$ and $21$. In each case there is one non-abelian group: the dihedral group $D_{10}$ of order 10 and the semi-direct product ${\mathbb{Z}_7 \rtimes \mathbb{Z}_3}$ of order 21. The non-trivial parameter sets for these are $(10,2,3,1)$, $(21,2,10,5)$ and $(21,6,2,1)$. The results are shown in Table \ref{table6}. \begin{table}[ht] \renewcommand{\arraystretch}{1.4} \begin{center} \begin{tabular}{c|c|c|c|c} Parameters &{ Non-abelian group} & No of non-equiv. SEDFs& Example & Case \\ \hline $(10, 2 ,3 , 1 )$ & $D_{10}$ & 1 & $\{e,s,r \}, \{sr, r^3, sr^4\}$ & Theorem \ref{thm:dihedral} \\ \hline $(21, 2,10,5 )$ & ${\mathbb{Z}_7 \rtimes \mathbb{Z}_3}$ & 0 & - & - \\ \hline $(21, 6 ,2,1 )$ & ${\mathbb{Z}_7 \rtimes \mathbb{Z}_3}$ & 0 & - & - \\ \end{tabular} \caption{Non-equivalent SEDFs in non-abelian groups of order up to $24$} \label{table6} \end{center} \end{table} \section{Concluding remarks and open problems} In this paper, we have begun to explore the questions: for given parameters, which finite groups (abelian or non-abelian) contain SEDFs with these parameters? What are the possible structures of these SEDFs? How many non-equivalent SEDFs exist with the same parameters? The concept of equivalence also raises new questions about SEDF results in the literature - for example, how many non-equivalent $(243,11,22,20)$-SEDFs exist? The case when $\lambda=1$ possesses a richer structural landscape than previously realised, with new SEDFs found in both abelian and non-abelian groups, and a wide range of abelian constructions obtainable recursively. For no other fixed value of $\lambda$ are explicit families of SEDFs known; all other SEDF constructions have $\lambda$ as a function of $n$. It is an open question whether families can be found for fixed values of $\lambda>1$. Further constructions for non-abelian SEDFs would be of interest. Does there exist a SEDF with parameters which cannot be realised in an abelian group but can be obtained in a non-abelian setting? This is of particular interest as the known range of possible parameters for abelian SEDFs becomes increasingly constrained (eg. in \cite{LeLiPr} it is conjectured that the known $(243,11,22,20)$-SEDF is the only abelian SEDF with $m>2$). Do there exist non-abelian SEDFs with $\lambda=1$ and $m>2$? Finally, we may ask whether certain phenomena observable in the computational results for SEDFs in groups of order up to $24$, are representative of a more general situation. Is it the case that in a group of order $p^2$ for $p$ an odd prime, no SEDF with $m=2$ can exist in the cyclic group of order $p^2$? In \cite{BaJiWeZh}, it is proved that there does not exist a $(p^2,m,k,\lambda)$-SEDF in an abelian group $\G$ with $p$ prime, $k>1$, $m>2$ and $\G$ cyclic; however we know of no such result in the literature for $m=2$. Does the finite field construction using squares and non-squares yield a unique (up to equivalence) $(p^2,2, \frac{p-1}{2}, \frac{p-1}{4})$-SEDF? \subsection*{Acknowledgements} Thanks to Maura Paterson for her helpful comments on the paper, and to Gemma Crowe and Ailsa Robertson for related discussions. The second author is supported by a Royal Society University Research Fellowship.

https://arxiv.org/abs/2002.03448

Kelly Criterion: From a Simple Random Walk to Lévy Processes

The original Kelly criterion provides a strategy to maximize the long-term growth of winnings in a sequence of simple Bernoulli bets with an edge, that is, when the expected return on each bet is positive. The objective of this work is to consider more general models of returns and the continuous time, or high frequency, limits of those models.

\section{Introduction} Consider repeatedly engaging in a game of chance where one side has an edge and seeks to optimize the betting in a way that ensures maximal long-term growth rate of the overall wealth. This problem was posed and analyzed by John Kelly \cite{Kelly56} at Bell Labs in 1956; the solution was implemented and tested in a variety of setting by a successful mathematician, gambler, and hedge fund manager Ed Thorp \cite{Thorp06} over the period from the 60's to the early 00's. As a motivating example, consider betting on a biased coin toss where the return $r$ is a random variable with distribution \begin{equation} \label{return1} \mathbb{P}(r=1) = p,\ \ \ \mathbb{P}(r=-1)=1- p; \end{equation} in what follows, we refer to this as the {\em simple Bernoulli model}. The condition to have an edge in this setting becomes $1/2<p\leq 1$ or, equivalently, \begin{equation} \label{edge1} \mathbb{E}[r] =2p-1>0. \end{equation} We plan on being able to make a large sequence of bets on this biased coin, resulting in an iid sequence of returns $\{r_k\}_{k\geq 1}$ with the same distribution as $r$, and ask how much we should bet so as to maximize long term wealth, given that we are compounding our returns. Assume we are betting with a fixed exposure $f$, that is, each bet involves a fixed fraction $f$ of the overall wealth, and $f \in [0,1]$. Practically, $f \geq 0$ means {\bf no shorting} and $f \leq 1$ means {\bf no leverage}, which we refer to as the {\bf NS-NL} condition. Then, starting with the initial amount $W_0$, the total wealth at time $n=1,2,3,\ldots$ is the following function of $f$: \begin{equation*} W_n^f = W_0 \prod_{k=1}^n\big(1 + f r_k\big). \end{equation*} For the long-term compounder wishing to maximize their long term wealth, a natural and equivalent goal would be to find the strategy $f=f^*$ maximizing the long-term growth rate \begin{equation} \label{rate-dt} g_r(f):= \lim_{n\rightarrow \infty}\frac{1}{n} \ln \frac{W_n^f}{W_0}. \end{equation} By direct computation, $$ g_r(f) = \lim_{n\rightarrow \infty}\frac{1}{n}\sum_{k=1}^n \ln(1+fr_k) = \mathbb{E}\ln(1+fr)=p\ln(1+f) + (1-p)\ln(1-f), $$ where the second equality follows by the law of large numbers, and therefore, after solving $g_r'(f^*)=0$, \begin{equation} \label{optimal1} f^* = 2p-1,\ \ \max_{f \in [0,1]} g_r(f)= g_r(f^*)=p\ln \frac{p}{1-p}+(2-p)\ln(2-2p); \end{equation} note that the edge condition \eqref{edge1} ensures that $f^*$ is an admissible strategy and $g_r(f^*)>0$. For more discussions of this result see \cite{Thorp06}. Our objective in this paper is to derive analogues of \eqref{optimal1} in the following situations: \begin{enumerate} \item the distribution of returns is a more general random variable; \item the compounding is continuous in time; \item the compounding is high frequency, leading to a continuous-time limit. \end{enumerate} In particular, we consider several scenarios when the returns are described by L\'evy processes, which addresses some of Thorp's questions regarding fat-tailed distributions in finance \cite{Thorp08}. In what follows, we write $\xi\overset{d}{=}\eta$ to indicate equality in distribution for two random variables, and $X\overset{{\mathcal{L}}}{=} Y$ to indicate equality in law (as function-valued random elements) for two random processes. For $x>0$, $\lfloor x \rfloor$ denotes the largest integer less than or equal to $x$. To simplify the notations, we always assume that $W_0=1$. \section{Discrete Compounding: General Distribution of Returns} Assume that the returns on each bet are independent random variables $r_k,\ k\geq 1,$ with the same distribution as a given random variable $r$, and let \begin{equation} \label{wealth2} W_n^f = \prod_{k=1}^n\big(1 + f r_k\big) \end{equation} denote the corresponding wealth process. We also keep the NS-NL condition on admissible strategies: $f\in [0,1]$. For the wealth process $W^f$ to be well-defined, we need the random variable $r$ to have the following properties: \begin{align} \label{r1} &\mathbb{P}(r\geq -1)=1;\\ \label{r2} &\mathbb{P}(r>0)>0,\ \mathbb{P}(r<0)>0;\\ \label{r3} &\mathbb{E}|\ln(1+r)|<\infty. \end{align} Condition \eqref{r1} quantifies the idea that a loss in a bet should not be more than $100\%$. Condition \eqref{r2} is basic non-degeneracy: both gains and losses are possible. Condition \eqref{r3} is a minimal requirement to define the long-term growth rate of the wealth process. The key object in this section will be the function \begin{equation} \label{F(f)} g_r(f)=\mathbb{E}\ln(1+fr). \end{equation} In particular, the following result shows that $g_r(f)$ is the long term growth rate of the wealth process $W^f$. \begin{prop} \label{prop:LLN} If \eqref{r1} and \eqref{r3} hold and $g_r(f)\not=0$, then, for every $f\in[0,1]$, the wealth process $W^f$ has an asymptotic representation \begin{equation} \label{eq:LLN-g1} W^f_n=\exp\Big( n g_r(f)\big(1+\varepsilon_n\big)\Big), \end{equation} where \begin{equation} \label{asymp0} \lim_{n\to \infty} \varepsilon_n=0 \end{equation} with probability one. \end{prop} \begin{proof} By \eqref{wealth2}, we have \eqref{eq:LLN-g1} with \begin{equation} \label{err0} \varepsilon_n=\frac{1}{ng_r(f)}\sum_{k=1}^n\bigg( \ln(1+fr_k)-g_r(f)\bigg), \end{equation} and then \eqref{asymp0} follows by \eqref{r3} and the strong law of large numbers. \end{proof} A stronger version of \eqref{r3} leads to a more detailed asymptotic of $W_n^f$. \begin{thm} \label{th:asympt1} Assume that \eqref{r1} holds and \begin{equation} \label{r3-3} \mathbb{E}|\ln(1+r)|^2<\infty. \end{equation} Then then, for every $f\in[0,1]$, the wealth process $W^f$ has an asymptotic representation \begin{equation} \label{eq:LLN-g2} W^f_n=\exp\Big( n g_r(f)+\sqrt{n}\big(\sigma_r(f)\zeta_n+ \epsilon_n\big)\Big), \end{equation} where $\zeta_n,\ n\geq 1, $ are standard Gaussian random variables, \begin{equation*} \sigma_r(f)=\Big(\mathbb{E}\big[\ln^2(1+fr)\big]-g_r^2(f)\Big)^{1/2}, \end{equation*} and \begin{equation*} \lim_{n\to \infty}\epsilon_n=0 \end{equation*} in probability. \end{thm} \begin{proof} With $\varepsilon_n$ from \eqref{err0}, the result follows by the Central Limit Theorem: $$ ng_r(f)\,\varepsilon_n= \sqrt{n}\left(\frac{1}{\sqrt{n}} \sum_{k=1}^n\bigg(\ln(1+fr_k)-g_r(f)\bigg)\right)= \sqrt{n}\big(\sigma_r(f)\zeta_n+ \epsilon_n\big). $$ \end{proof} Because the Central Limit Theorem gives convergence in distribution, the random variables $\zeta_n$ in \eqref{eq:LLN-g2} can indeed depend on $n$. Additional assumptions about the distribution of $r$ \cite[Theorem 1]{Zolotarev-AE} lead to higher-order asymptotic expansions and a possibility to have $\lim_{n\to \infty}\epsilon_n=0$ with probability one. The following properties of the function $g_r$ are immediate consequences of the definition and the assumptions \eqref{r1}--\eqref{r3}: \begin{prop} \label{prop:BasicF} The function $f\mapsto g_r(f)$ is continuous on the closed interval $[0,1]$ and infinitely differentiable in $(0,1)$. In particular, \begin{equation} \label{DF} \frac{dg_r}{df}(f)=\mathbb{E}\left[\frac{r}{1+fr}\right],\ \ \frac{d^2g_r}{df^2}(f)=-\mathbb{E}\left[\frac{r^2}{(1+fr)^2}\right]<0. \end{equation} \end{prop} \begin{cor} \label{cor1} The function $g_r$ achieves its maximal value on $[0,1]$ at a point $f^*\in [0,1]$ and $g_r(f^*)\geq 0$. If $ g_r(f^*)> 0$, then $f^*$ is unique. \end{cor} \begin{proof} Note that $g_r(0)=0$ and, by \eqref{DF}, the function $g_r$ is strictly concave (or convex up) on $[0,1]$. \end{proof} While concavity of $g_r$ implies that $g_r$ achieves a unique global maximal value at a point $f^{**}$, it is possible that the domain of the function $g_r$ is bigger than the interval $[0,1]$ and $f^{**}\notin [0,1]$. A simple way to exclude the possibility $f^{**}<0$ is to consider returns $r$ that are not bounded from above: $\mathbb{P}(r>c)>0$ for all $c>0$: in this case, the function $g_r(f)=\mathbb{E}\ln(1+fr)$ is not defined for $f<0$. Similarly, if $\mathbb{P}(r<-1+\delta)>0$ for all $\delta>0$, then the function $g_r$ is not defined for $f>1$, excluding the possibility $f^{**}>1.$ Below are more general sufficient conditions to ensure that the point $f^*\in [0,1]$ from Corollary \ref{cor1} is the point of global maximum of $g_r$: $f^*=f^{**}$. \begin{prop} \label{prop:glob1} If \begin{align} \label{global1-l} &\lim_{f\to 0+} \mathbb{E}\left[\frac{r}{1+fr}\right]>0 \ \ \ \ \ {\rm and}\\ \label{global1-r} &\lim_{f\to 1-} \mathbb{E}\left[\frac{r}{1+fr}\right]<0, \end{align} then there is a unique $f^*\in (0,1)$ such that $$ g_r(f)< g_r(f^*) $$ for all $f$ in the domain of $g_r$. \end{prop} \begin{proof} Together with the intermediate value theorem, conditions \eqref{global1-l} and \eqref{global1-r} imply that there is a unique $f^*\in (0,1)$ such that $$ \frac{dg_r}{df}(f^*)=0. $$ It remains to use strong concavity of $g_r$. \end{proof} Because $r\geq -1$, the expected value $\mathbb{E}[r]$ is always defined, although $\mathbb{E}[r]=+\infty$ is a possibility. Thus, by \eqref{DF}, condition \eqref{global1-l} is equivalent to the intuitive idea of an edge: $$ \mathbb{E}[r]>0, $$ which, similar to \eqref{edge1}, guarantees that $g_r(f)>0$ for some $f\in (0,1)$. Condition \eqref{global1-r} can be written as $$ \mathbb{E}\left[\frac{r}{1+r}\right]<0, $$ with the convention that the left-hand side can be $-\infty$. This condition does not appear in the simple Bernoulli model, but is necessary in general, to ensure that the edge is not too big and leveraged gambling ($f^*>1$) does not lead to an optimal strategy. As an example, consider the {\em general Bernoulli model} with \begin{equation} \label{GenBern} \mathbb{P}(r=-a)=1-p,\ \ \mathbb{P}(r=b)=p,\ \ 0<a\leq 1,\ b>0,\ 0<p<1. \end{equation} The function $$ g_r(f)=p\ln (1+fb) + (1-p)\ln(1-fa) $$ is defined on $(-1/b, 1/a)$, achieves the global maximum at $$ f^*=\frac{p}{a}-\frac{1-p}{b}, $$ and $$ g_r(f^*)=p\ln p +(1-p)\ln (1-p) +\ln\frac{a+b}{a} - (p-1)\ln \frac{b}{a}; $$ we know that $g_r(f^*)\geq 0$, even though it is not at all obvious from the above expression. The NS-NL condition $f^*\in [0,1]$ becomes $$ \frac{a}{a+b}\leq p \leq \min\left( \frac{ab}{a+b}\left(1+\frac{1}{b}\right), 1\right), $$ and it is now easy to come up with a model in which $f^*>1$: for example, take $$ a=0.1, \ b=0.5,\ p=0.5 $$ so that $f^*=4$. Given that a gain and a loss in each bet are equally likely, but the amount of a gain is five times as much as that of a loss, a large value of $f^*$ is not surprising, although economical and financial implications of this type of leveraged betting are potentially very interesting and should be a subject of a separate investigation. Because of the logarithmic function in the definition of $g_r$, the distribution of $r$ can have a rather heavy right tail and still satisfy \eqref{r3}. For example, consider \begin{equation} \label{Ch0} r=\eta^2-1, \end{equation} where $\eta$ has standard Cauchy distribution with probability density function $$ h_{\eta}(x)=\frac{1}{\pi(1+x^2)},\ \ \ -\infty< x<+\infty. $$ Then $$ g_r(f)=\frac{2}{\pi}\int_0^{+\infty} \frac{\ln\big((1-f)+fx^2\big)}{1+x^2}\, dx= 2\ln\big(\sqrt{f}+\sqrt{1-f}\big), $$ where the second equality follows from \cite[(4.295.7)]{Gradshtein-Ryzhyk}. As a result, we get a closed-form answer $$ f^*=\frac{1}{2},\ g_r(f^*)=\ln 2. $$ A general way to ensure \eqref{r1}--\eqref{r3} is to consider \begin{equation} \label{expo-model} r=e^{\xi}-1 \end{equation} for some random variable $\xi$ such that $\mathbb{P}(\xi>0)>0,\ \mathbb{P}(\xi<0)>0$, and $\mathbb{E}|\xi|<\infty$; note that \eqref{Ch0} is a particular case, with $\xi=\ln\eta^2$. Then \eqref{global1-l} and \eqref{global1-r} become, respectively, \begin{align} \label{global1-l-e} &\mathbb{E}e^{\xi}>1 \ \ \ \ \ {\rm and}\\ \label{global1-r-e} &\mathbb{E}e^{-\xi}>1. \end{align} For example, if $\xi$ is normal with mean $\mu\in \mathbb{R}$ and variance $\sigma^2>0$, then $$ \mathbb{E}e^{\xi}=e^{\mu+(\sigma^2/2)}, \ \ \mathbb{E}e^{-\xi}=e^{-\mu+(\sigma^2/2)}, $$ and \eqref{global1-l-e}, \eqref{global1-r-e} are equivalent to \begin{equation} \label{log-nrmal} -\frac{\sigma^2}{2}<\mu<\frac{\sigma^2}{2}, \end{equation} which, when interpreted in terms of returns, can indeed be considered as a ``reasonable'' edge condition: large values of $|\mu|$ do create a bias in one direction. Note that the corresponding $f^*$ is not available in closed form, but can be evaluated numerically. \section{Continuous Compounding and a Case for L\'{e}vy Processes} \label{sec:CC} Continuous time compounding includes discrete compounding as a particular case and makes it possible to consider more general types of return processes. The objective of this section is to show that continuous time compounding that leads to a non-trivial and non-random long-term growth rate of the resulting wealth process effectively forces the return process to have independent increments. The two main examples of such process are sums of iid random variables from the previous section and the L\'{e}vy processes. Writing \eqref{wealth2} as \begin{equation} \label{wealth1-dt} W_{n+1}^f-W_n^f=\big(fW_n^f\big)\,r_{n+1}, \end{equation} we see that a natural continuous time version of \eqref{wealth1-dt} is \begin{equation} \label{wealth1-ct} dW_{t}^f=fW_t^fdR_{t} \end{equation} for a suitable process $R=R_t,\ t\geq 0$ on a stochastic basis \begin{equation*} \mathbb{F}=\Big(\Omega, \mathcal{F},\ \{\mathcal{F}_t\}_{t\geq 1}, \mathbb{P}\Big) \end{equation*} satisfying the usual conditions \cite[Definition I.1.1]{Protter}. We interpret \eqref{wealth1-ct} as an integral equation \begin{equation} \label{wealth1-ct-i} W_{t}^f=1+f\int_0^tW_s^fdR_{s}; \end{equation} recall that $W_0^f=1$ is the standing assumption. Then the Bichteler-Dellacherie theorem \cite[Theorem III.22]{Protter} implies that the process $R$ must be a semi-martingale (a sum of a martingale and a process of bounded variation) with trajectories that, at every point, are continuous from the right and have limits from the left. Furthermore, if we allow the process $R$ to have discontinuities, then, by \cite[Theorem II.36]{Protter}, we need to modify \eqref{wealth1-ct-i} further: \begin{equation*} W_{t}^f=1+f\int_0^tW_{s-}^fdR_{s}, \end{equation*} where $$ W_{s-}=\lim_{\varepsilon\to 0, \varepsilon>0} W_{s-\varepsilon}, $$ and, assuming $R_0=0$, the process $W^f$ becomes the Dol\'{e}ans-Dade exponential \begin{equation} \label{DDE-1} W^f_t=\exp\left(fR_t-\frac{f^2\langle R^c\rangle_t}{2} \right)\prod_{0<s\leq t} (1+f\triangle R_s)\,e^{-f\triangle R_s}. \end{equation} In \eqref{DDE-1}, $\langle R^c\rangle$ is the quadratic variation process of the continuous martingale component of $R$ and $\triangle R_s=R_s-R_{s-}$. A natural analog of \eqref{r1} is \begin{equation} \label{r1-ct} \triangle R_s\geq -1, \end{equation} and then \eqref{DDE-1} becomes \begin{equation} \label{DDE-2} W^f_t=\exp\left(fR_t-\frac{f^2\langle R^c\rangle_t}{2}+ \sum_{0<s\leq t}\Big( \ln (1+f\triangle R_s) -f\triangle R_s\Big)\right). \end{equation} To proceed, let us assume that the trajectories of $R$ are continuous: $\triangle R_s=0$ for all $s$ so that \begin{equation*} W^f_t=\exp\left(fR_t-f^2\langle R^c\rangle_t\right). \end{equation*} If, similar to \eqref{rate-dt}, we define the long-term growth rate $g_R(f)$ by \begin{equation} \label{gr-ct0} g_R(f)=\lim_{t\to \infty} \frac{\ln W^f_t}{t}, \end{equation} then we need the limits \begin{equation} \label{trip0-cont} \mu:=\lim_{t\to \infty} \frac{R_t}{t},\ \ \ \sigma^2:=\lim_{t\to \infty} \frac{\langle R^c\rangle_t}{t} \end{equation} to exist with probability one and with non-random numbers $\mu,\sigma^2.$ Being a semi-martingale without jumps, the process $R$ has a representation \begin{equation} \label{sm-cont} R_t=A_t+R^c_t, \end{equation} where $A$ is process of bounded variation; cf. \cite[Theorem II.2.34]{LimitTheoremsforStochasticProcesses}. Then \eqref{trip0-cont} imply that, for large $t$, \begin{equation} \label{linear0} A_t\approx \mu t,\ \ \langle R^c\rangle_t\approx \sigma^2 t, \end{equation} that is, a natural way to achieve \eqref{trip0-cont} is to consider the process $R$ of the form \begin{equation*} R_t=\mu t+\sigma\,B_t, \end{equation*} where $\sigma>0$ and $B=B_t$ is a standard Brownian motion. Then \begin{equation} \label{DDE-3-cont} W^f_t=\exp\left(f\mu t+f \sigma\,B_t-\frac{f^2\sigma^2 t}{2}\right), \end{equation} and we come to the following conclusion: {\em continuous time compounding with a continuous return process effectively implies that the wealth process is a geometric Brownian motion}. The long-term growth rate \eqref{gr-ct0} becomes \begin{equation} \label{gr-ctL-cont} g_R(f)=f\mu-\frac{f^2\sigma^2}{2}, \end{equation} so that \begin{equation*} f^*=\frac{\mu}{\sigma^2},\ \ g_R(f^*)=\frac{\mu^2}{2\sigma^2}, \end{equation*} and the NS-NL condition is \begin{equation*} 0<\mu<\sigma^2. \end{equation*} Even though these results are not especially sophisticated, we will see in the next section (Theorem \ref{prop0}) that the process \eqref{DDE-3-cont} naturally appears as the continuous-time, or high frequency, limit of discrete-time compounding for a large class of returns. On the other hand, if we assume that the process $R$ is purely discontinuous, with jumps $\triangle R_{k}=r_k$ at times $s=k\in \{1,2,3,\ldots\}$, then $$ R_t=0,\ t\in (0,1),\ R_t=\sum_{k=1}^{\lfloor t \rfloor} r_k,\ t\geq 1, $$ and \eqref{DDE-1} becomes \eqref{wealth2}. Accordingly, we will now investigate the general case \eqref{DDE-1} when the process $R$ has both a continuous component and jumps. To this end, we use \cite[Proposition II.1.16]{LimitTheoremsforStochasticProcesses} and introduce the jump measure $\mu^R=\mu^R(dx,ds)$ of the process $R$ by putting a point mass at every point in space-time where the process $R$ has a jump: \begin{equation} \label{JumpMeasure} \mu^R(dx,ds) = \sum_{s > 0}\delta_{(\triangle R_s,s)}(dx, ds); \end{equation} note that both the time $s$ and size $\triangle R_s$ of the jump can be random. In particular, with \eqref{r1-ct} in mind, \begin{equation} \label{sum1} \sum_{0<s\leq t} \Big(\ln (1+f\triangle R_s) -f\triangle R_s\Big)= \int_0^t-\!\!\!\!\!\!\int_{-1}^{+\infty} \big(\ln(1+fx)-fx\big)\mu^R(dx,ds); \end{equation} here and below, \begin{equation} \label{zint} -\!\!\!\!\!\!\int_a^b, \ \ \ a<0<b, \end{equation} stands for $$ \int\limits_{(a,0)\bigcup(0,b)}. $$ By \cite[Proposition II.2.9 and Theorem II.2.34]{LimitTheoremsforStochasticProcesses}, and keeping in mind \eqref{r1-ct}, we get the following generalization of \eqref{sm-cont}: \begin{equation} \label{sm-general} \begin{split} R_t&=A_t+R^c_t+ \int_0^t-\!\!\!\!\!\!\int_{1}^{+\infty} x\mu^R(dx,ds)\\ &+ \int_0^t-\!\!\!\!\!\!\int_{-1}^1 x\big(\mu^R(dx,ds)-\nu(dx,s)da_s\big), \end{split} \end{equation} where $a=a_t$ is a predictable non-decreasing process and $\nu=\nu(dx,t)$ is the non-negative random time-dependent measure on $(-1,0)\bigcup(0,+\infty)$ with the property \begin{equation*} -\!\!\!\!\!\!\int_{-1}^{+\infty}\min(1,x^2)\nu(dx,t)\leq 1 \end{equation*} for all $t\geq 0$ and $\omega\in \Omega$. Moreover \begin{align} \label{trip-A} A_t&=\int_0^t \mu_s\, da_s \ {\rm \ for\ some\ predictable \ process} \ \mu=\mu_t,\\ \label{trip-B} \langle R^c\rangle_t&=\int_0^t \sigma^2_s\,da_s\ {\rm \ for\ some\ predictable \ process} \ \sigma=\sigma_t, \end{align} and the process $$ t \mapsto \int_0^t-\!\!\!\!\!\!\int_{-1}^{+\infty} h(x)\big(\mu^R(dx,ds)-\nu(dx,s)da_s\big) $$ is a martingale for every bounded measurable function $h=h(x)$ such that $\limsup_{x\to 0}|h(x)|/|x|<\infty$. To proceed, we assume that $$ \mathbb{E} \int_0^t-\!\!\!\!\!\!\int_{-1}^{+\infty} |\ln(1+x)|\, \nu(dx,s)\,da_s<\infty,\ t>0, $$ which is a generalization of condition \eqref{r3}. Then, by \cite[Theorem II.1.8]{LimitTheoremsforStochasticProcesses}, the process $$ t\mapsto \int_0^t-\!\!\!\!\!\!\int_{-1}^{+\infty} \ln(1+x)\big(\mu^R(dx,ds)-\nu(dx,s)da_s\big) $$ is a martingale. Next, we combine \eqref{DDE-2}, \eqref{sum1}, and \eqref{sm-general}, and re-arrange the terms so that the logarithm of the wealth process becomes \begin{equation} \label{logW-g} \begin{split} \ln W_t^f&=fA_t+fR^c_t-\frac{f^2}{2}\langle R^c_t \rangle -f\int_0^t-\!\!\!\!\!\!\int_{-1}^1x\nu(dx,s)da_s\\ &+ \int_0^t-\!\!\!\!\!\!\int_{-1}^{+\infty} \ln(1+fx)\nu(dx,s)da_s+M^f_t, \end{split} \end{equation} where $$ M^f_t=\int_0^t-\!\!\!\!\!\!\int_{-1}^{+\infty} \ln(1+fx)\big(\mu^R(dx,ds)-\nu(dx,s)da_s\big). $$ In general, for equality \eqref{logW-g} to hold, we need to make an additional assumption \begin{equation} \label{nu-int1} -\!\!\!\!\!\!\int_{-1}^{1}x\nu(dx,t)<\infty \end{equation} for all $t\geq 0$ and $\omega\in \Omega$. In the particular case \eqref{wealth2}, \begin{itemize} \item $a_s=\lfloor s \rfloor$ is the step function, with unit jumps at positive integers, so that $da_s$ is the collection of point masses at positive integers; \item $\nu(dx,s)=F^R(dx),$ where $F^R$ is the cumulative distribution function of the random variable $r$, so that \eqref{nu-int1} holds automatically; \item $\mu_t=g_f(r)+\int_{-1}^1 x\,F^R(dx)$, $R^c_t=0$, $\sigma_t=0$; \item $M^f_t=\sum_{0<k\leq t}\big(\ln(1+fr_k)-g_r(f)\big)$; \item condition \eqref{LogVar-Levy} is \eqref{r3-3}. \end{itemize} A natural way to reconcile \eqref{linear0} with \eqref{trip-A}, \eqref{trip-B} is to take $\mu_t=\mu$, $\sigma_t=\sigma$ for some non-random numbers $\mu\in \mathbb{R}$, $\sigma\geq 0$, and a non-random non-decreasing function $a=a_t$ with the property \begin{equation} \label{limit-a} \lim_{t\to+\infty}\frac{a_t}{t}=1. \end{equation} Then, to have a non-random almost-sure limit $$ \lim_{t\to \infty} \frac{1}{t}\int_0^t\int_{-1}^{+\infty} \varphi(x) \nu(dx,s)da_s $$ for a sufficiently rich class of non-random test functions $\varphi$, we have to assume that there exists a non-random non-negative measure $F^R=F^R(dx)$ on $(-1,0)\bigcup(0,+\infty)$ such that \begin{equation} \label{intF1-1} -\!\!\!\!\!\!\int_{-1}^{+\infty}\min( |x|,1)\, F^R(dx)<\infty \end{equation} and, for large $s$, $$ \nu(dx,s)\approx F^R(dx). $$ As a result, if \begin{equation} \label{LevyMeasure0} \nu(dx,s)= F^R(dx) \end{equation} for all $s$, then \begin{equation} \label{triple-nr1} A_t=\mu a_t,\ \langle R^c\rangle_t=\sigma^2a_t,\ \nu(dx,t)=F^R(dx)a_t \end{equation} are all non-random, and \cite[Theorem II.4.15]{LimitTheoremsforStochasticProcesses} implies that $R$ is a {\em process with independent increments}. Furthermore, \eqref{triple-nr1} and the strong law of large numbers for martingales imply \begin{equation*} \mathbb{P}\left(\lim_{t\to \infty} \frac{R^c_t}{t}=0\right)=1; \end{equation*} cf. \cite[Corollary 1 to Theorem II.6.10]{LSh-M}. Similarly, if \begin{equation} \label{LogVar-Levy} -\!\!\!\!\!\!\int_{-1}^{+\infty} \ln^2(1+x)\, F^R(dx)<\infty, \end{equation} then $M^f$ is a square-integrable martingale and \begin{equation*} \mathbb{P}\left( \lim_{t\to \infty} \frac{M^f_t}{t}=0\right)=1. \end{equation*} Writing $$ \bar{\mu}=\mu-\int_{-1}^{1}x F^R(dx) $$ the long-term growth rate \eqref{gr-ct0} becomes \begin{equation} \label{gr-ctL} g_R(f)=f\bar{\mu}-\frac{f^2\sigma^2}{2}+ -\!\!\!\!\!\!\int_{-1}^{\infty} \ln(1+fx) F^R(dx), \end{equation} which does include both \eqref{F(f)} and \eqref{gr-ctL-cont} as particular cases. By direct computation, the function $f\mapsto g_R(f)$ is concave and the domain of the function contains $[0,1]$. Similar to Proposition \ref{prop:glob1}, we have the following result. \begin{thm} \label{th-LevyRate} Consider continuous-time compounding with return process \begin{equation} \label{Return-Levy0} \begin{split} R_t&=A_t+R_t^c+ \int_0^t-\!\!\!\!\!\!\int_{1}^{+\infty} x\mu^R(dx,ds)\\ &+ \int_0^t-\!\!\!\!\!\!\int_{-1}^1 x\big(\mu^R(dx,ds)-\nu(dx,s)da_s\big), \end{split} \end{equation} where the random measure $\mu^R$ is from \eqref{JumpMeasure}, and assume that equalities \eqref{limit-a} and \eqref{triple-nr1} hold. If $F^R$ satisfies \eqref{intF1-1}, \eqref{LogVar-Levy}, and \begin{align*} &\lim_{f\to 0+} -\!\!\!\!\!\!\int_{-1}^{\infty} \frac{x}{1+fx}\, F^R(dx)>-\bar{\mu},\\ &\lim_{f\to 1-} -\!\!\!\!\!\!\int_{-1}^{\infty} \frac{x}{1+fx}\, F^R(dx)<\sigma^2-\bar{\mu}, \end{align*} then the long-term growth rate is given by \eqref{gr-ctL}, and there exists a unique $f^*\in (0,1)$ such that $$ g_R(f)< g_R(f^*) $$ for all $f$ in the domain of $g_R$. \end{thm} By the Lebesgue decomposition theorem, the measure corresponding to the function $a=a_t$ has a discrete, absolutely continuous, and singular components. With \eqref{limit-a} in mind, a natural choice of the discrete component is $a_t=\lfloor t \rfloor$, which, as we saw, corresponds to discrete compounding discussed in the previous section. A natural choice of the absolutely continuous component is \begin{equation*} a_t=t. \end{equation*} Then \begin{equation*} A_t=\mu t,\ R^c_t=\sigma B_t,\ \nu(dx,t)da_t=F^R(dx)\,dt, \end{equation*} where $B$ is a standard Brownian motion. By \cite[Corollary II.4.19]{LimitTheoremsforStochasticProcesses}, we conclude that the process $R$ has independent and stationary increments, that is, {\em $R$ is a L\'{e}vy process}. In this case, equality \eqref{Return-Levy0} is known as the L\'{e}vy-It\^{o} decomposition of the process $R$; cf. \cite[Theorem 19.2]{Sato}. We do not consider the singular case in this paper and leave it for future investigation. \section{Continuous Limit of Discrete Compounding} \subsection{A (Simple) Random Walk Model} Following the methodology in \cite[Section 7.1]{Thorp06}, we assume compounding a sufficiently large number $n$ of bets in a time period $[0,T]$. The returns $r_{n,1}, r_{n,2},\ldots $ of the bets are \begin{equation} \label{return-nk} r_{n,k} = \frac{\mu}{n} + \frac{\sigma}{\sqrt{n}}\,\xi_{n,k} \end{equation} for some $\mu>0$, $\sigma>0$ and independent identically distributed random variables $\xi_{n,k},\ k=1,2,\ldots,$ with mean $0$ and variance $1$. The classical simple random walk corresponds to $\mathbb{P}(\xi_{n,k}=\pm 1)=1/2$ and can be considered a {\em high frequency} version of \eqref{return1}. Similar to \eqref{r1}, we need $r_{n,k}\geq -1$, which, in general, can only be achieved with {\em uniform boundedness} of $\xi_{n,k}$: \begin{equation} \label{eq:bnd1} |\xi_{n,k}|\leq C_0, \end{equation} and then, with no loss of generality, we assume that $n$ is large enough so that \begin{equation} \label{small-r} |r_{n,k}|\leq \frac{1}{2}. \end{equation} Similar to \eqref{edge1}, a condition to have an edge is $$ \mathbb{E}[r_{n,k}] = \frac{\mu}{n} > 0, $$ and, similar to \eqref{wealth2}, given $n$ bets per unit time period, with exposure $f \in [0,1]$ in each bet, we get the following formula for the total wealth $W_t^{n,f}$ at time $t\in (0,T]$ assuming $W_0=1$: \begin{equation} \label{Wntf} W_t^{n,f} = \prod_{k=1}^{\lfloor nt \rfloor}\big(1+fr_{n,k}\big); \end{equation} $\lfloor nt \rfloor$ denotes the largest integer less than $nt$. Let $B=B_t,\ t\geq 0,$ be a standard Brownian motion on a stochastic basis $(\Omega, \mathcal{F}, \{\mathcal{F}_t\}_{t\geq 0},\mathbb{P})$ satisfying the usual conditions, and define the process \begin{equation} \label{wealth2-cont} W^f_t= \exp\left(\left(f \mu - \frac{f^2\sigma^2}{2}\right)t + f \sigma B_t\right). \end{equation} Note that \eqref{wealth2-cont} is a particular case of \eqref{DDE-3-cont}. \begin{thm} \label{prop0} For every $T>0$ and every $f\in [0,1]$, the sequence of processes $ \big(W_t^{n,f},\ n\geq 1,\ t\in [0,T]\big)$ converges in law to the process $W^f=W^f_t,\ t\in [0,T]$. \end{thm} \begin{proof} Writing $$ Y^{n,f}_t= \ln W_t^{n,f}, $$ the objective is to show weak convergence, as $n\to \infty$, of $Y^{n,f}$ to the process $$ Y^f_t=\left( f \mu - \frac{f^2\sigma^2}{2}\right)t + f \sigma B_t,\ \ t\in [0,T]. $$ The proof relies on the method of predictable characteristics for semimartingales from \cite{LimitTheoremsforStochasticProcesses}. More specifically, we make suitable changes in the proof of Corollary VII.3.11. By \eqref{wealth2-cont} $$ Y^{n,f}_t=\sum_{k=1}^{\lfloor nt \rfloor} \ln(1+fr_{n,k}). $$ Then \eqref{return-nk} and \eqref{eq:bnd1} imply \begin{equation*} \mathbb{E}\bigg( Y^{n,f}_t - \mathbb{E}Y^{n,f}_t\bigg)^4 \leq \frac{C_0^4\sigma^4}{n^2}\big(nT+3nT(nT-1)\big) \leq {3C_0^4\sigma^4T^2}, \end{equation*} from which uniform integrability of the family $\{Y^{n,f}_t,\ n\geq 1, \ t\in [0,T]\}$ follows. Then, by \cite[Theorem VII.3.7]{LimitTheoremsforStochasticProcesses}, it suffices to establish the following: \begin{align} \label{mean0} \lim_{n\to \infty}& \sup_{t \leq T} \left|\lfloor nt \rfloor \mathbb{E}\big[\ln(1+fr_{n,1})\big] - \left(f\mu-\frac{f^2\sigma^2}{2}\right)t \right| =0,\\ \label{var0} \lim_{n\to \infty} & \lfloor nt \rfloor \left(\mathbb{E} \big(\ln(1+fr_{n,1})\big)^2 - \bigg(\mathbb{E}\big[\ln(1+fr_{n,1})\big]\bigg)^2 \right)= f^2 \sigma^2 t,\ t\in [0,T],\\ \label{jump0} \lim_{n\to \infty} &\lfloor nt \rfloor \mathbb{E} \big[\phi\big(\ln(1+fr_{n,1})\big)\big] = 0,\ t\in [0,T]. \end{align} Equality \eqref{jump0} must hold for all functions $\phi=\phi(x), \ x\in \mathbb{R},$ that are continuous and bounded on $\mathbb{R}$ and satisfy $\phi(x)=o(x^2), \ x\to 0$, that is, \begin{equation} \label{phi0} \lim_{x\to 0} \frac{\phi(x)}{x^2}=0. \end{equation} Equalities \eqref{mean0} and \eqref{var0} follow from $$ r_{n,1}^2=\frac{\sigma^2}{n}\, \xi_{n,1}^2 + \frac{2\mu\sigma\xi_{n,1}}{n^{3/2}}+\frac{\mu^2}{n^2}, $$ together with \eqref{small-r} and an elementary inequality $$ \left| \ln(1+x)-x-\frac{x^2}{2} \right| \leq |x|^3,\ |x|\leq \frac{1}{2}. $$ In particular, \begin{equation*} \mathbb{E} \big[\big(\ln(1+fr_{n,1})\big)^2\big] = \frac{f^2\sigma^2}{n}+o(1/n), \ n\to +\infty. \end{equation*} To establish \eqref{jump0}, note that \eqref{phi0} and \eqref{return-nk} imply $$ \phi\big(\ln (1+fr_{n,1})\big) = o(1/n), \ n\to +\infty. $$ \end{proof} Similar to \eqref{rate-dt}, we define the long-term continuous time growth rate \begin{equation*} g(f)=\lim_{t\to \infty} \frac{1}{t}\ln W^f_t. \end{equation*} Then a simple computation show that $$ g(f)=f\mu - \frac{f^2\sigma^2}{2}, $$ and so \begin{equation} \label{optf-rv} f^*=\frac{\mu}{\sigma^2} \end{equation} achieves the maximal long-term continuous time growth rate \begin{equation} \label{optf-rv1} g(f^*)=\frac{\mu^2}{2\sigma^2}. \end{equation} The NS-NL condition $f^*\in [0,1]$ holds if $0\leq \mu\leq \sigma^2$, which, to the order $1/n$, is consistent with \eqref{global1-l} and \eqref{global1-r}, when applied to \eqref{return-nk}: $$ \mathbb{E}[r_{n,k}]=\frac{\mu}{n},\ \ \mathbb{E}\left[\frac{r_{n,k}}{1+r_{n,k}}\right] =\frac{\mu-\sigma^2}{n}+o(n^{-1}). $$ The wealth process \eqref{wealth2-cont} is that of someone who is ``continuously" placing bets, that is, adjusts the positions instantaneously, and, for large $n$, is a good approximation of high frequency betting \eqref{Wntf}. In general, when the returns are given by \eqref{return-nk}, a direct optimization of \eqref{Wntf} with respect to $f$ will not lead to a closed-form expression for the corresponding optimal strategy $f_n^*$ , but Theorem \ref{prop0} implies that, for sufficiently large $n$, \eqref{optf-rv} is an approximation of $f_n^*$ and \eqref{optf-rv1} is an approximation of the corresponding long-term growth rate. As an illustration, consider the high-frequency version of the simple Bernoulli model \eqref{return1}: \begin{equation} \label{return1-hf} \mathbb{P}\left( r_{n,k}=\frac{\mu}{n}\pm \frac{\sigma}{\sqrt{n}}\right)=\frac{1}{2}, \end{equation} which, for fixed $n$, is is a particular case of the general Bernoulli model \eqref{GenBern} with $p=q=1/2$, $$ a=\frac{\sigma}{\sqrt{n}}-\frac{\mu}{n},\ b=\frac{\sigma}{\sqrt{n}}+\frac{\mu}{n}. $$ Then, by direct computation, $$ f_n^*=\frac{\mu}{\sigma^2-(\mu^2/n)}\to \frac{\mu}{\sigma^2},\ n\to \infty, $$ and $$ \lim_{n\to \infty} g_{r_n}(f_n^*)=\frac{\mu^2}{2\sigma^2}. $$ Even though the analysis of the proof of Theorem \ref{prop0} shows that the convergence is uniform in $f$ on compact sub-sets of $(0,1)$, the proof that $\lim_{n\to \infty} f^*_n=f^*$ would require a version of Theorem \ref{prop0} with $T=+\infty$, which, for now, is not available. With natural modifications, Theorem \ref{prop0} extends to the setting \eqref{expo-model}. \begin{thm} \label{prop0-1} Assume that \begin{equation*} r_{n,k}+1=\exp\left(\frac{b}{n}+\frac{\sigma}{\sqrt{n}}\,\xi_{n,k}\right), \end{equation*} where $b \in \mathbb{R}$, $\sigma>0$, and, for each $n\geq 1$, $k\leq n$, the random variables $\xi_{n,k}$ are independent and identically distributed, with zero mean, unit variance, and, for every $a>0$, \begin{equation} \label{moment1} \lim_{n\to\infty}n\, \mathbb{E}\big[|\xi_{n,1}|^2I(|\xi_{n,1}|>a\sqrt{n})\big]=0. \end{equation} Then the conclusion of Theorem \ref{prop0} holds with $$ \mu=b+\frac{\sigma^2}{2}. $$ \end{thm} \begin{proof} Even though a formal Taylor expansion suggests $$ r_{n,k}=\frac{\mu}{n} + \frac{\sigma}{\sqrt{n}}\,\xi_{n,k}+o(1/n) $$ we cannot apply Theorem \ref{prop0} directly because the random variables $\xi_{n,k}$ are not necessarily uniformly bounded. Still, condition \eqref{moment1} makes it possible to verify conditions \eqref{mean0}--\eqref{jump0}. \end{proof} Condition \eqref{moment1} is clearly satisfied when $\xi_{n,k},\ k=1,2,\ldots,$ are iid standard normal, which corresponds to \begin{equation} \label{discr-ret-gauss} r_{n,k}=\frac{P_{k/n} - P_{(k-1)/n}}{P_{(k-1)/n}} \end{equation} and \begin{equation} \label{GBM-returns} P_t=e^{bt+\sigma B_t}. \end{equation} Thus, while the exponential model \eqref{expo-model} with log-normal returns is not solvable in closed form, the high-frequency version leads to the (approximately) optimal strategy \begin{equation} \label{st-line} f^*=\frac{b}{\sigma^2}+\frac{1}{2}, \end{equation} and, under \eqref{log-nrmal}, the NS-NL condition holds: $f^*\in (0,1)$. Numerical experiments with $\sigma=1$ and $n=10$ show that the values of the corresponding optimal $f^*_{10}$ are very close to those given by \eqref{st-line} for all $b\in (-1/2,1/2)$. Informally, both Theorems \ref{prop0} and \ref{prop0-1} can be considered as particular cases of the delta method for the Donsker theorem with drift: if the sequence of processes $$ t\mapsto \sum_{k=1}^{\lfloor nt \rfloor} \xi_{n,k} $$ converges, as $n\to \infty$, to the processes $t\mapsto bt+\sigma B_t$ and $\varphi=\varphi(x)$ is a suitable function with $\varphi(0)=0$, then one would expect the sequence of processes $$ t\mapsto \sum_{k=1}^{\lfloor nt \rfloor} \varphi\big(\xi_{n,k}\big) $$ to converge to the process $$ t\mapsto \left(\varphi'(0)b+\frac{\varphi''(0)\sigma^2}{2}\right)t+|\varphi'(0)|\sigma B_t. $$ \subsection{Beyond the Log-Normal Limit} With the results of Section \ref{sec:CC} in mind, we consider the following generalization of \eqref{discr-ret-gauss}: \eqref{GBM-returns}: $$ r_{n,k} = \frac{P_{k/n} - P_{(k-1)/n}}{P_{(k-1)/n}}, \ k=1,2,\ldots, $$ where the process $P=P_t,\ t\geq 0$, has the form $P_t=e^{R_t}$, and $R=R_t$ is a L\'evy process. In other words, \begin{equation} \label{dctr-gen} r_{n,k}={e}^{R_{k/n}-R_{(k-1)/n}} - 1. \end{equation} As in \eqref{Return-Levy0}, the process $R=R_t$ can be decomposed into a drift, diffusion/small jump, and large jump components according to the L\'{e}vy-It\^{o} decomposition \cite[Theorem 19.2]{Sato}: \begin{equation} \label{Levy-main} R_t={\mu}t+\sigma\,B_t+ \int_0^t-\!\!\!\!\!\!\int_{-1}^1 x\big(\mu^R(dx,ds)-F^R(dx)ds\big)+ \int_0^t\int_{|x|>1} x\mu^R(dx,ds); \end{equation} we continue to use the notation $-\!\!\!\!\!\int$ first introduced in \eqref{zint}. Now that the process $R_t$ is exponentiated, \begin{itemize} \item there is no need to assume that $\triangle R_t\geq -1$ ; \item the analog of \eqref{LogVar-Levy} becomes $\mathbb{E}|R_1|<\infty$. \end{itemize} Equality \eqref{Levy-main} has a natural interpretation in terms of financial risks \cite{Thorp03}: the drift represents the edge (``guaranteed'' return), diffusion and small jumps represent small fluctuations of returns, and the large jump component represents (sudden) large changes in returns. Similar to \eqref{Wntf}, the corresponding wealth process is \begin{equation} \label{wp-100} W_{t}^{n,f} = \prod_{k = 1}^{\lfloor nt \rfloor} \big(1 + f r_{n,k}\big). \end{equation} We have the following generalization of Theorem \ref{prop0-1}. \begin{thm} Consider the family of processes $W^{n,f}=W_t^{n,f}$, $t\in [0,T]$, $n\geq 1$, $f\in [0,1],$ defined by \eqref{wp-100}. If $r_{n,k}$ is given by \eqref{dctr-gen}, with $P_t=e^{R_t}$, and $R=R_t$ is a L\'{e}vy process with representation \eqref{Levy-main} and $\mathbb{E}|R_1|< \infty$, then, for every $f\in [0,1]$ and $T>0$, $$ \lim_{n\to \infty}W^{n,f} \overset{{\mathcal{L}}}{=} W^f $$ in $\mathbb{D}((0,T))$, where \begin{equation} \label{WP-generalLP} \begin{split} W^f_t &=\exp\left(f R_t + \frac{ f(1-f)\sigma^2}{2}\, t\right.\\ &+ \left. \int_0^t -\!\!\!\!\!\!\int_{\mathbb{R}} \Big[\ln\big(1+f(e^x - 1)\big) - f x\Big] \mu^R(dx, ds)\right). \end{split} \end{equation} \end{thm} \begin{proof} By \eqref{dctr-gen} and \eqref{wp-100}, \begin{equation*} \ln W_t^{n,f} = \sum_{k=1}^{\lfloor nt \rfloor}\ln\bigg( 1 + f \big( {e}^{R_{k/n}-R_{(k-1)/n}} - 1\big)\bigg). \end{equation*} \underline{Step 1:} For $s \in \big(\frac{k-1}{n}, \frac{k}{n}\big]$, let \begin{equation} \label{rnks} r_s^{n,k} = {e}^{R_s - R_{(k-1)/n}} - 1, \end{equation} and apply the It\^o's formula \cite[Theorem II.32]{Protter} to the process $$ s\mapsto \ln\big(1+f r_s^{n,k}\big),\ \ s \in \Bigg(\frac{k-1}{n}, \frac{k}{n}\Bigg]. $$ The result is \begin{equation*} \begin{split} \ln\big(1+f r_s^{n,k}\big) &= \int_{\frac{k-1}{n}}^s \frac{f (1 + r_{u-}^{n,k})}{1+f r_{u-}^{n,k}}\, dR_u + \frac{\sigma^2}{2}\int_{\frac{k-1}{n}}^s \frac{f(1-f)(1+r_{u-}^{n,k})}{\big(1+fr_{u-}^{n,k}\big)^2}\, du \\ & + \int_{\frac{k-1}{n}}^s -\!\!\!\!\!\!\int_{\mathbb{R}} \bigg[\ln\big(1-f+f{e}^{x} (r_{u-}^{n,k}+ 1)\big) \\ & - \ln(1+fr_{u-}^{n,k}) - x \frac{f(1+r_{u-}^{n,k})}{1+f r_{u-}^{n,k}}\bigg]\mu^R(dx, du). \end{split} \end{equation*} \underline{Step 2:} Putting $s = \frac{k}{n}$ in the above equality and summing over $k$, we derive the following expression for $ \ln W_t^{n,f}$: \begin{align} \notag \ln W_t^{n,f} &= \sum_{k=1}^{\lfloor nt \rfloor} \bigg(\int_{\frac{k-1}{n}}^{\frac{k}{n}} h^{(1)}_{n,k}(s) \, dR_s + \int_{\frac{k-1}{n}}^{\frac{k}{n}} h^{(2)}_{n,k}(s) \, ds + \int_{\frac{k-1}{n}}^{\frac{k}{n}} -\!\!\!\!\!\!\int_{\mathbb{R}} h^{(3)}_{n,k}(s,x) \mu^R(dx, du)\bigg)\\ \label{main-integrals} &= \int_0^t H^{(1)}_{n,t}(s) \, dR_s + \int_0^t H^{(2)}_{n,t}(s) \, ds + \int_0^t-\!\!\!\!\!\!\int_{\mathbb{R}} H^{(3)}_{n,t}(s,x) \mu^R(dx, ds)\,, \end{align} where \begin{equation*} \begin{split} \displaystyle h^{(1)}_{n,k}(s) & = \frac{f(1+r_{s-}^{n,k})}{1+fr_{s-}^{n,k}}\,,\ \ h^{(2)}_{n,k}(s) = \frac{\sigma f(1-f)}{2}\frac{1+r_{s-}^{n,k}}{(1+fr_{s-}^{n,k})^2}\,, \\ h^{(3)}_{n,k}(s,x)&= \ln\big(1-f+fe^x(r_{s-}^{n,k}+1) \big) - \ln(1+fr_{s-}^{n,k}) - fx\frac{1+r_{s-}^{n,k}}{1+fr_{s-}^{n,k}}\,;\\ H^{(i)}_{n,t}(s) &= \sum_{k=1}^{\lfloor nt \rfloor} h^{(i)}_{n,k}(s) \mathbf{1}_{(\frac{k-1}{n}, \frac{k}{n}]}(s), \ i = 1, 2;\ \ H^{(3)}_{n,t}(s,x) = \sum_{k=1}^{\lfloor nt \rfloor} h^{(3)}_{n,k}(s,x)\mathbf{1}_{(\frac{k-1}{n}, \frac{k}{n}]}(s). \end{split} \end{equation*} \underline{Step 3:} Because $$ \lim_{n\to \infty,\, k/n\to s} R_{(k-1)/n}=R_{s-}, $$ equality \eqref{rnks} implies $$ \lim_{n\to +\infty,\, k/n\to s} r^{n,k}_{s-}= 0 $$ for all $s$. Consequently, we have the following convergence in probability: \begin{align*} \lim_{n\to +\infty} H^{(1)}_{n,t}(s)=f,\ & \lim_{n\to +\infty} H^{(2)}_{n,t}(s)=\frac{\sigma^2 f(1-f)}{2},\\ &\lim_{n\to +\infty} H^{(2)}_{n,t}(s,x)=\ln\big(1+f(e^x-1)\big)-fx. \end{align*} To pass to the corresponding limits in \eqref{main-integrals}, we need suitable bounds on the functions $H^{(i)}$, $i=1,2,3$. Using the inequalities $$ 0<\frac{1+y}{1+ay}\leq \frac{1}{a},\ \ 0<\frac{1+y}{(1+ay)^2} \leq \frac{1}{4a(1-a)}, \ \ \ y>-1,\ a\in (0,1), $$ we conclude that $$ 0<h^{(1)}_{n,k}(s)\leq 1,\ 0<h^{(2)}_{n,k}(s)\leq \sigma^2, $$ and therefore \begin{equation} \label{ubound1-2} 0<H^{(1)}_{n,t}(s)\leq 1,\ 0<H^{(2)}_{n,t}(s)\leq \sigma^2. \end{equation} Similarly, for $f\in (0,1)$ and $y > -1$, \begin{equation} \label{H3bnd00} \left|\ln \frac{1-f+fe^x(y+1)}{1+fy}-fx\frac{1+y}{1+fy}\right|\leq 2\big(|x|\wedge |x|^2\big), \end{equation} so that $$ |h^{(3)}_{n,k}(s,x)|\leq 2\big(|x|\wedge |x|^2\big) $$ and \begin{equation} \label{ubound3} \big|H^{(3)}_{n,t}(s)\big| \leq 2\big(|x|\wedge |x|^2\big). \end{equation} To verify \eqref{H3bnd00}, fix $f\in (0,1)$ and $y>-1$, and define the function $$ z(x) = \ln \frac{1-f+fe^x(y+1)}{1+fy},\ x\in \mathbb{R}. $$ By direct computation, \begin{equation*} \begin{split} z(0) &= 0, \\ z'(x) &= \frac{fe^x(y+1)}{1-f + fe^x(y+1)}=1-\frac{1-f}{1-f + fe^x(y+1)},\\ z'(0) &= \frac{f(y+1)}{1+fy}, \end{split} \end{equation*} so that, using the Taylor formula, \begin{equation} \label{H3bnd} \ln \frac{1-f+fe^x(y+1)}{1+fy}-fx\frac{1+y}{1+fy} =z(x)-z(0)-xz'(0)=\int_0^x(x-u)z''(u)du. \end{equation} It remains to notice that $$ 0\leq z'(x)\leq 1, \ \ 0\leq z''(x)\leq 1, $$ and then \eqref{H3bnd00} follows from \eqref{H3bnd}. With \eqref{ubound1-2} and \eqref{ubound3} in mind, the dominated convergence theorem \cite[Theorem IV.32]{Protter} makes it possible to pass to the limit in probability in \eqref{main-integrals}; the convergence in the space $\mathbb{D}$ then follows from the general results of \cite[Section IX.5.12]{LimitTheoremsforStochasticProcesses}. \end{proof} The following is a representation of the long-term growth rate of the limiting wealth process $W^f$. \begin{thm} \label{th:gr-rate-LP-gen} Let $R=R_t$ be a L\'{e}vy process with representation \eqref{Levy-main}. If $\mathbb{E}|R_1| < \infty$, then the process $W^f=W^f_t$ defined in \eqref{WP-generalLP} satisfies \begin{equation} \label{GR-gen-LP} \begin{split} \lim_{t\to +\infty} \frac{\ln W_t^f}{t}&=f \bigg(\mu + \int_{|x|>1} x F^R(dx)\bigg) + \frac{ f(1-f) \sigma^2}{2}\\ &+ \int_{\mathbb{R}}\big[\ln\big(1 + f ({e}^x - 1) \big) - f x \big] F^R(dx). \end{split} \end{equation} \end{thm} \begin{proof} By \eqref{WP-generalLP}, $$ \frac{\ln W_t^f}{t}= f \frac{R_t}{t} + \frac{ f(1-f)\sigma^2}{2}+ \frac{1}{t}\int_0^t \int_{\mathbb{R}} \Big[\ln\big(1+f(e^x - 1)\big) - f x\Big] \mu^R(dx, ds). $$ It remains to apply the law of large numbers for L\'evy processes \cite[Theorem 36.5]{Sato}. \end{proof} If, in addition, we assume that $$ -\!\!\!\!\!\!\int_{-1}^1 |x|F^R(dx)<\infty, $$ that is, the small-jump component of $R$ has bounded variation, then, after a change of variables and re-arrangement of terms, \eqref{GR-gen-LP} becomes \eqref{gr-ctL}. On the other hand, equality \eqref{gr-ctL} is derived for a wider class of return processes that includes L\'{e}vy processes as a particular case. Similar to Proposition \ref{prop:glob1}, we also have the following result. \begin{thm} \label{th-LevyRate1} In the setting of Theorem \ref{th:gr-rate-LP-gen}, denote the right-hand side of \eqref{GR-gen-LP} by $g_R(f)$ and assume that \begin{align*} &\lim_{f\to 0+} \int_{\mathbb{R}} \left(\frac{e^x - 1}{1+f(e^x - 1)} - x\right)\, F^R(dx)>-\bigg(\mu + \frac{\sigma^2}{2} + \int_{|x|>1} x F^R(dx)\bigg),\\ &\lim_{f\to 1-} \int_{\mathbb{R}} \left(\frac{e^x - 1}{1+f(e^x - 1)} - x\right)\, F^R(dx)< -\bigg(\mu + \int_{|x|>1} x F^R(dx)\bigg), \end{align*} Then there exists a unique $f^*\in (0,1)$ such that $$ g_R(f)< g_R(f^*) $$ for all $f$ in the domain of $g_R$. \end{thm} \section{Continuous Limit of Random Discrete Compounding} The objective of this section is to analyze high frequency limits for betting {\em in business time}. In other words, the number of bets is not known a priori, so that a natural model of the corresponding wealth process is \begin{equation} \label{weath-mgen} W_{t}^{n,f}=\prod_{k=1}^{\lfloor \Lambda_{n,t}\rfloor} (1+fr_{n,k}) \end{equation} where, for each $n$, the process $t\mapsto \Lambda_{n,t}$ is a subordinator, that is, a non-decreasing L\'{e}vy process, independent of all $r_{n,k}$. To study \eqref{weath-mgen}, we will follow the methodology in \cite{KZZ}, where convergence of processes is derived after {\em assuming} a suitable convergence of the random variables. The main result in this connection is as follows. \begin{thm} \label{th:Levy-gen0} Consider the following objects: \begin{itemize} \item random variables $X_{n,k},\ n,k\geq 1$ such that $\{X_{n,k},\ k\geq 1\}$ are iid for each $n$, with mean zero and, for some $\beta\in [0,1]$, $m_n:=\Big(\mathbb{E}|X_{n,1}|^{\beta}\Big)^{1/\beta}<\infty$; \item random processes $\Lambda_n=\Lambda_{n,t}$, $n\geq 1,\ t\geq 0,$ such that, for each $n$, $\Lambda_n$ is a subordinator independent of $\{X_{n,k},\ k\geq 1\}$ with the properties $\Lambda_{n,0}=0$, and for some numbers $0<\delta,\delta_1\leq 1$ and $C_n>0$, $\Big(\mathbb{E}\Lambda_{n,t}^{\delta}\Big)^{1/\delta} \leq C_n t^{\delta_1/\delta}$. \end{itemize} Assume that there exist infinitely divisible random variables $Y$ and $U$ such that $$ \lim_{n\to \infty} \sum_{k=1}^n X_{n,k} \overset{d}{=} \bar{Y},\ \lim_{n\to \infty} \frac{\Lambda_{n,1}}{n} \overset{d}{=} \bar{U}. $$ If \begin{equation} \label{dens} \sup_{n} \Big(C_n m_n^{\beta}\Big)<\infty, \end{equation} then, as $n\to \infty$, the sequence of processes $$ t\mapsto \sum_{k=1}^{\lfloor \Lambda_{n,t}\rfloor} X_{n,k},\ t\in [0,T], $$ converges, in the Skorokhod topology, to the process $Z=Z_t$ such that $Z_t=Y_{U_t}$, where $Y$ and $U$ are independent L\'{e}vy processes satisfying $Y_1\overset{d}{=}\bar{Y}$ and $U_1\overset{d}{=}\bar{U}$. \end{thm} The proof is a word-for-word repetition of the arguments leading to \cite[Theorem 1]{KZZ}: the result of \cite{GnF}, together with the assumptions of the theorem, implies $$ \lim_{n\to \infty} \sum_{k=1}^{\lfloor \Lambda_{n,1}\rfloor} X_{n,k}\overset{d}{=} Z_1, $$ and therefore the convergence of finite-dimensional distributions for the corresponding processes; together with condition \eqref{dens}, this implies the convergence in the Skorokhod space. Because we deal exclusively with L\'{e}vy processes, it is possible to avoid the heavy machinery from \cite{LimitTheoremsforStochasticProcesses}. We now consider the wealth process \eqref{weath-mgen} and apply Theorem \ref{th:Levy-gen0} with $$ X_{n,k}= \ln (1+fr_{n,k})-\mathbb{E}\ln (1+fr_{n,k}). $$ On the one hand, convergence to infinitely divisible distributions other than normal is a very diverse area, with a variety of conditions and conclusions; cf. \cite[Chapter XVII, Section 5]{FellerII} or a summary in \cite[Section 16.2]{Klenke}. On the other hand, optimal strategy \eqref{optf-rv} seems to persist. For example, assume that the returns $r_{n,k}$ are as in \eqref{return-nk}, and let $\Lambda_{n,t}=S_{n^{\alpha}t}$, where $\alpha\in (0,1]$ and $S=S_t$ is the L\'{e}vy process such that $S_1$ has the $\alpha$-stable distribution with both scale and skewness parameters equal to $1$. Recall that an $\alpha$-stable L\'{e}vy process $L^{\alpha}=L^{\alpha}_t$ satisfies the following equality in distribution (as processes): \begin{equation} \label{scale-L} L^{\alpha}_{\gamma t}\overset{{\mathcal{L}}}{=} \gamma^{1/\alpha}L^{\alpha}_t, \ \gamma>0. \end{equation} Then $$ \Lambda_{n,t}\overset{{\mathcal{L}}}{=} nS_{t} $$ and, in the notations of Theorem \ref{th:Levy-gen0}, $\bar{Y}$ is normal with mean zero and variance $\sigma^2$. Keeping in mind that $$ \mathbb{E}\ln (1+fr_{n,k}) = \mathbb{E}\ln (1+fr_{n,1})= \Big(f\mu-\frac{f^2\sigma^2}{2}\Big)n^{-1}+o(n^{-1}), $$ we repeat the arguments from \cite[Example 1]{KZZ} to conclude that $$ \lim_{n\to \infty} \ln W^{n,f}_t \overset{{\mathcal{L}}}{=} \Big(f\mu-\frac{f^2\sigma^2}{2}\Big)S_t + Z_t, $$ where $Z_1$ has symmetric $2\alpha$-stable distribution. By \eqref{scale-L}, $$ S_t\overset{d}{=} t^{1/\alpha}S_1,\ \lim_{t\to +\infty} t^{-1/\alpha}Z_t \overset{d}{=} \lim_{t\to +\infty} t^{-1/(2\alpha)}Z_1 \overset{d}{=} 0, $$ and the ``natural'' long term growth rate becomes $$ \lim_{t\to \infty}t^{-1/\alpha}\Big(\lim_{n\to \infty} \ln W^{n,f}_t\Big) \overset{d}{=} \Big(f\mu-\frac{f^2\sigma^2}{2}\Big)S_1, $$ which is random, but, for each realization of $S$, is still maximized by $f^*$ from \eqref{optf-rv}. Therefore, if the time with which we compound our wealth is random, then the growth rate is also random as we don't know when we will stop compounding, yet it is still maximized by a deterministic fraction. Note that, for the purpose of this computation, the (stochastic) dependence between the processes $S$ and $Z$ is not important. \section{Conclusions And Further Directions} The NS-NL condition $f^*\in [0,1]$ can fail in many situations. Even in the simple Bernoulli model, if $p<1/2$, then the short position $f^*=2p-1$ achieves positive long-time wealth growth: $$ g_r(f^*)=p\ln \frac{p}{1-p}+(2-p)\ln(2-2p)=\ln 2 +p\ln p+(1-p)\ln(1-p)>0. $$ Note that $-p\ln p-(1-p)\ln(1-p)$ is the Shannon entropy of the Bernoulli distribution, and the largest value of the entropy is $\ln 2$, corresponding to $p=1/2$. When the edge is too big (cf. \eqref{GenBern}), then $f^*>1$, that is, leveraged gambling leads to bigger long-time wealth growth than any NS-NL strategy. The economical and financial implications of $f^*\notin [0,1]$ are beyond the scope of our investigation and must be studied in a broader context of risk tolerance: even when $f^*\in (0,1)$, a certain fraction of $f^*$ can be a smarter strategy, cf. \cite[Section 7.3]{Thorp06}. A related observation, to be further studied in the future, is that high-frequency betting can lead to a more aggressive strategy than the ``low frequency'' counterpart. For example, comparing \eqref{return1} and \eqref{return1-hf}, we see that $\mu=2p-1$ and $\sigma^2=4p(1-p) < 1$ when $p\not=1/2$. As a result, by \eqref{optf-rv}, the optimal strategy for \eqref{return1-hf} with large $n$ is $f^*\approx (2p-1)/(4p(1-p))> 2p-1$; recall that $f^*=2p-1$ is the optimal strategy for the simple Bernoulli model \eqref{return1}. On the other hand, numerical simulations suggest that, in the log-normal model \eqref{discr-ret-gauss}, \eqref{GBM-returns}, high-frequency compounding does not always lead to larger $f^*$. Other problems warranting further investigation include \begin{enumerate} \item A dynamic strategy $f=f(t)$ with a predictable process $f$; \item A portfolio of bets, with a vector of strategies $\mathbf{f}=(f_1,\ldots, f_N)$. \end{enumerate} \bibliographystyle{amsplain}

https://arxiv.org/abs/2209.11307

A combinatorial bound on the number of distinct eigenvalues of a graph

The smallest possible number of distinct eigenvalues of a graph $G$, denoted by $q(G)$, has a combinatorial bound in terms of unique shortest paths in the graph. In particular, $q(G)$ is bounded below by $k$, where $k$ is the number of vertices of a unique shortest path joining any pair of vertices in $G$. Thus, if $n$ is the number of vertices of $G$, then $n-q(G)$ is bounded above by the size of the complement (with respect to the vertex set of $G$) of the vertex set of the longest unique shortest path joining any pair of vertices of $G$. The purpose of this paper is to commence the study of the minor-monotone floor of $n-k$, which is the minimum of $n-k$ among all graphs of which $G$ is a minor. Accordingly, we prove some results about this minor-monotone floor.

\section{Motivation and background} The Inverse Eigenvalue Problem for a Graph (IEPG) starts with a pattern of zero and nonzero constraints for a real symmetric matrix, described by a graph $G$ on $n$ vertices, and asks what spectra are possible for the set of $n \times n$ matrices $\mathcal{S}(G)$ that exhibit this pattern. Two narrower questions about the possible spectra of matrices in $\mathcal{S}(G)$ yield important matrix-theoretic graph parameters that motivate what is studied here. Each of these parameters has a natural combinatorial bound. Firstly, the highest possible nullity, denoted by $\mathrm{M}(G)$, is also equal to the maximum multiplicity that can be obtained by any eigenvalue, and this is bounded by a process called zero forcing, giving $\mathrm{M}(G) \le \mathrm{Z}(G)$, where $Z(G)$ is the zero forcing number of $G$. (See, for example, \cite{HLS2022} for the definition of zero forcing number.) Secondly, the lowest possible number of distinct eigenvalues is denoted by $q(G)$, and this is bounded by the number of vertices in a unique shortest path, which we denote by $\usp{G}$, thus giving $q(G) \ge \usp{G}$ \cite[Theorem 3.2]{MinimumDistinct}. (The parameter $q(G)$ has been studied, for example, in \cite{F10,MinimumDistinct,Analysis,Nordhaus}.) As stated, these combinatorial bounds are in opposite directions, but in fact the matrix parameter that will be of interest is $n - q(G)$, which can be thought of as the maximum number of ordered eigenvalue coincidences, and is therefore called the \emph{maximum spectral equality}. Expressed this way, the combinatorial bound is $n - q(G) \le n - \usp{G}$, involving a combinatorial quantity $n - \usp{G}$ that will be called the \emph{spectator number} of $G$ and will be denoted by $\uspc{G}$. The motivation for this project concerns the behavior of the four graph parameters $\mathrm{M}(G)$, $\mathrm{Z}(G)$, $n - q(G)$, and $n - \usp{G}$ with respect to subgraphs, and more generally with respect to graph minors. The set $\mathcal{S}(G)$ is a manifold, an open subset of a vector space, and the process of removing an edge from the graph, forcing a zero entry in the matrix, collapses a coordinate direction of the enveloping vector space, reducing the dimension of $\mathcal{S}(G)$ by one. One might expect, generically, that removing degrees of freedom in the matrix would not allow greater nullity and would also not allow more eigenvalue coincidences. In other words, one might expect both $\mathrm{M}(G)$ and $n - q(G)$ to be weakly decreasing as edges are deleted, and might expect a similar subgraph monotonicity for the related combinatorial bounds $\mathrm{Z}(G)$ and $n - \usp{G}$. In most cases this expected behavior is observed to hold, but there are counterexamples---an extreme example being that the graph with the fewest edges of all on $n$ vertices, the edgeless graph $\overline{K_n} = nK_1$, is the only graph that achieves a value as high as $n$ for either $\mathrm{M}(G)$ or $\mathrm{Z}(G)$ and is the only graph that achieves a value as high as $n - 1$ for either $n - q(G)$ or $n - \usp{G}$. (This extreme example is also maximally disconnected, but there also exist, for any of these four parameters, connected counterexample graphs to which an edge can be added while decreasing the value of the parameter. See, e.g., Example 2.9 in \cite{Parameters} for an example for $\mathrm{M}(G)$ and $\mathrm{Z}(G)$, and see, e.g., Figures 6.1 and 6.2 of \cite{MinimumDistinct} for an example for $n-q(G)$ and $n-\usp{G}$.) For three of these four parameters, namely $\mathrm{M}(G)$, $\mathrm{Z}(G)$, and $n - q(G)$, there is an established variant of the graph parameter that not only exhibits the generically expected behavior of weakly decreasing as edges are deleted, but also is monotone with respect to graph minors more generally. The purpose of this paper is to complete the set of four by introducing, and commencing the study of, a canonically chosen modification of $n - \usp{G}$ that achieves monotonicity with respect to graph minors. Thus, we study the minor-monotone floor of $n - \usp{G}=\uspc{G}$, that is, the minimum of $\uspc{H}$ among all graphs $H$ containing $G$ as a minor. We denote this minor-monotone floor by $\uspcf{G}$ and call it the \emph{spectator floor} of $G$. One of our main results is that, in order to find a graph $G'$ containing $G$ as a minor and such that $\uspcf{G}=\uspc{G'}$, one need only add edges to $G$. In \cref{Minor Operations and the Spectator Floor}, we will prove the following. \begin{theorem} \label{lem:no decontract-intro} For every graph $G$, there is a graph $G'$ with the same number of vertices as $G$ such that $G$ is a subgraph of $G'$ and such that $\uspcf{G}=\uspc{G'}$. \end{theorem} Another main result is that the spectator floor is additive over connected components. We prove the following result in \cref{Disconnected Graphs}. \begin{theorem} \label{prop:components-intro} For any graph $G=G_1\sqcup G_2$ with no edges connecting $G_1$ to $G_2$, \[\uspcf{G}=\uspcf{G_1}+\uspcf{G_2}\] \end{theorem} \cref{trees} gives some results about the spectator floor of trees. In \cref{calculation of spectator floor}, we present an algorithm to calculate the spectator floor of simple graphs. In \cref{sec:minimal}, we study graphs with small spectator floors. In particular, we determine the minor minimal graphs (whether or not parallel edges are allowed) that have spectator floor $k$, when $k=1$ and when $k=2$. In \cref{minor max graphs}, we characterize the minor maximal graphs with a given spectator floor, subject to a restriction on the number of vertices in the graph and the number of parallel edges between any pair of vertices. Finally, \cref{further questions} presents some questions for further research. Before moving on to \cref{Minor Operations and the Spectator Floor}, we continue this introduction with some additional preliminary information. \subsection{Definitions} We allow a graph $G$ to have multiple edges (unless $G$ is explicitly stated to be a simple graph) but not to have loops at its vertices. Moreover, we assume all graphs are nonempty, that is every graph has at least one vertex. The multigraph convention for matrix entries is that each edge of $G$ contributes additively a non-zero amount to the matrix entry, which in particular allows the contributions from multiple edges to cancel. The convention is also that diagonal entries are unconstrained. Concretely, then, given a graph whose vertices are the index set $\{1, \dots, n\}$, the space $\mathcal{S}(G)$ of matrices conforming to the pattern of $G$ is the set of all real symmetric $n \times n$ matrices $A = [a_{ij}]$ such that \begin{itemize} \item if $i \ne j$ and there is no edge in $G$ connecting $i$ to $j$, then $a_{ij} = 0$, \item if $i \ne j$ and there is exactly one edge in $G$ connecting $i$ to $j$, then $a_{ij} \ne 0$, \item if $i \ne j$ and there is more than one edge in $G$ connecting $i$ to $j$, then $a_{ij}$ is unconstrained, and \item diagonal entries $a_{ii}$ are unconstrained. \end{itemize} The following definitions lead up to the promised naming and explanation of $n - \usp{G}$, together with its minor-monotone floor. Let $G$ be a graph on $n$ vertices. We start with some standard definitions. \begin{itemize} \item A \emph{walk} in $G$ from $u$ to $v$ is a sequence of edges $(e_1, \dots, e_k)$ from $G$ and a sequence of vertices $(u=v_1, v_2, \dots, v_{k}, v_{k+1}=v)$ from $G$ such that each edge $e_i$ has endpoints $v_i$ and $v_{i+1}$. The \emph{length} of the walk is the length $k \ge 0$ of the sequence of edges. \item A \emph{path} in $G$ is a walk all of whose vertices are distinct. The length of a path is the number of edges $k$, but the \emph{order} of a path is the number of vertices $k + 1$. \item A \emph{shortest path} in $G$ is a path in $G$ from $u$ to $v$ of order $k + 1$ such that no path in $G$ from $u$ to $v$ has order $k$ or smaller. \item A \emph{unique shortest path} in $G$ is a shortest path $P$ in $G$ from $u$ to $v$ of order $k + 1$ such that every path in $G$ from $u$ to $v$ of order exactly $k + 1$ is identical to $P$. The graph $G$ may be a multigraph, but no edge in a unique shortest path can have other edges parallel to it, because two paths with a different sequence of edges are not considered identical, even if the sequence of vertices is the same. \end{itemize} We now introduce the following definitions. \begin{itemize} \item A \emph{parade} in $G$ is a unique shortest path in $G$ that achieves the largest possible order for a unique shortest path in $G$. \item The \emph{parade number} of a graph $G$, denoted $\usp{G}$, is the number of vertices in some parade in $G$. \item The \emph{spectator number} of a graph $G$, denoted $\uspc{G}$, is the number of vertices outside some parade in $G$, hence $\uspc{G}=n-\usp{G}$. \item The \emph{spectator floor} of a graph $G$, denoted $\uspcf{G}$, is the minor-monotone floor of $\uspc{G}$; in other words, the minimum value of $\uspc{H}$ over the set of all graphs $H$ of which $G$ is a minor. \end{itemize} \subsection{Matrix-theoretic graph parameters related to \texorpdfstring{$q(G)$}{q(G)}} The graph parameters $q_M(G)$ and $q_S(G)$, introduced in \cite{SSP2017}, satisfy $q(G) \le q_M(G) \le q_S(G)$. For these variants of $q(G)$, matrices are restricted to those satisfying the Strong Multiplicity Property (SMP) or the Strong Spectral Property (SSP), respectively. A consequence of \cite{SSP2020} is that the maximum SMP spectral equality, $n - q_M(G)$, and maximum SSP spectral equality, $n - q_S(G)$, are minor-monotone graph parameters. A consequence of \cite{SSP2017} is that $n - q_M(G)$ and $n - q_S(G)$ each take the sum over components of a disconnected graph, in contrast for example to minor-monotone matrix nullity graph parameters that tend to take the maximum over components. Since $q(G) \le q_M(G) \le q_S(G)$, we also have the inequalities \[ n - q_S(G) \le n - q_M(G) \le n - q(G). \] The existence of a minor-monotone lower bound invites inquiry into the minor-monotone floor of the maximum spectral equality, which by general properties of minor-monotone floors and ceilings shares the same lower bound: \[ n - q_M(G) \le \lfloor n - q(G) \rfloor \le n - q(G). \] \medskip \noindent {\bf Combinatorial bounds.} Given a graph $G$ on $n$ vertices, if the vertices $u$ and $v$ are connected by a unique shortest path on $k$ vertices, then it is straightforward to show, for any matrix $A \in \mathcal{S}(G)$, that the powers $I = A^0, A^1, A^2, \dots, A^{k-1}$ form a linearly independent set in the vector space of symmetric $n \times n$ matrices, implying $q(G) > k - 1$ since the linear combination of powers in the minimal polynomial of a matrix is zero. This yields the result that $q(G) \ge k$, where $k$ is the number of vertices in any unique shortest path. This bound suggests the definition of the \emph{spectator number} given above. In particuar, the spectator number is the minimum cardinality of a vertex set that is complementary, relative to the set of all vertices of $G$, to the set of vertices in a unique shortest path in $G$. The above inequality $q(G) \ge k$, satisfied whenever there exist $k$ vertices forming a unique shortest path, guarantees that the spectator number is an upper bound for the maximum spectral equality $n - q(G)$, just as the combinatorial parameter $Z(G)$ is an upper bound for the matrix parameter $M(G)$. Maximum nullity gives a bound on maximum spectral equality, \[ M(G) - 1 \le n - q(G), \] because any one eigenvalue of multiplicity $k$ induces $k - 1$ eigenvalue equalities on its own. In a parallel way, but for entirely different and combinatorial reasons, the quantity $Z(G) - 1$ is a lower bound for the spectator number. Unlike zero forcing, whose computation is in general NP-hard \cite{Hardness}, the minimum spectator number can be computed in polynomial time for any graph (see, e.g., \cref{USP-in-poly-time}). \medskip \noindent {\bf Bounds for minor-monotone floors.} The minor-monotone floor of a graph parameter on a graph $G$ is the minimum of the parameter over the infinite collection of graphs $H$ of which $G$ is a minor. When a parameter $\beta(G)$ is minor-monotone, it is known that for any fixed $k$ the class of graphs with $\beta(G) \ge k$ can be recognized in polynomial time, and in fact can be recognized in linear time in the special case that the complete list of minor-minimal graphs for $\beta(G) \ge k$ is known and all of them happen to be planar graphs \cite{B93}. In \cref{sec:minimal}, we show that the minor-minimal graphs with spectator floor $1$ and $2$ are all planar. The parameter $\uspcf{G}$ extends in a natural way to multigraphs, signed graphs up to negation, and signed multigraphs up to negation, all of which are categories in which minor-minimal sets are guaranteed to be finite. The signed variant is in terms of \emph{monotone shortest paths}; a path of length $k$ from vertex $u$ to vertex $v$ in a signed graph $G$ is a monotone shortest path if there is no path from $u$ to $v$ of length shorter than $k$, and if every other path from $u$ to $v$ of length $k$ has the same product of edge signs. \section{Minor Operations and the Spectator Floor} \label{Minor Operations and the Spectator Floor} A \emph{minor} of a graph $H$ is obtained from $H$ by a sequence of the following operations: \begin{enumerate} \item Deletion of an isolated vertex \item Deletion of an edge, denoted $H\backslash e$ \item Contraction of an edge $e$ with no edges in parallel with it, denoted $H/e$. (If the minor is to be a simple graph, then the edge cannot be in a triangle.) \end{enumerate} A minor obtained by performing exactly one of these operations is called an \emph{elementary minor}. When calculating the spectator floor, we take the minimum of the spectator number over any graph that can be made by performing the three minor operations above in reverse. The following terminology will be useful to describe the operation that reverses contraction. \begin{definition}\label{def:decontraction} If $G$ and $H$ are graphs and $e\in E(H)$ such that $G=H/e$, then we call the operation used to obtain $H$ from $G$ a \emph{decontraction} of a vertex. \end{definition} Note that there may be many ways to decontract a vertex. Therefore, decontraction is not well-defined without additional context. The main result of this section is \cref{lem:no decontract}, which tells us that in order to calculate the spectator floor, it is not necessary to add vertices or to perform decontraction. The only necessary operation is adding edges. The following lemmas, \cref{lem:no-delete-vertex,lem:isolated,lem:no-isolated,lem:still-usp,lem:no-endpoints,lem:e not in pert}, are all in support of the main result. The lemma below tells us that when taking a minor, it is not necessary to perform step 1 (deletion of an isolated vertex), and that we may order the steps so that all instances of step 3 (contraction) appear before all instances of step 2 (edge deletion). \begin{lemma} \label{lem:no-delete-vertex} Let $G$ and $H$ be graphs such that $G$ is a minor of $H$. If $H$ has no isolated vertices, then there are sets of edges $C$ and $D$ such that $G\cong H/C\backslash D$. \end{lemma} \begin{proof} We need to show that it is not necessary to delete any isolated vertices in order to obtain $G$ from $H$. Since $H$ has no such vertices, such a vertex can only appear after all edges incident with that vertex have been deleted. But, in that case, all but one of the edges incident with the vertex can be deleted and the remaining edge contracted instead. (This procedure works since we are assuming $G$ is nonempty.) \end{proof} The next lemma tells us that the presence of isolated vertices does not affect the spectator floor value. \begin{lemma} \label{lem:isolated} Let $G+v$ be a graph with an isolated vertex $v$, and let $G$ be the graph obtained from $G+v$ by deleting $v$. Then $\uspcf{G+v}=\uspcf{G}$. Moreover, if $G$ is a minor of a graph $H$ and $\uspcf{G}=\uspc{H}$, then there is a graph $H^+$, with exactly one more vertex than $H$, such that $\uspcf{G+v}=\uspc{H^+}$ and $G+v$ is a minor of $H^+$. \end{lemma} \begin{proof} Since $G$ is a minor of $G+v$, we have $\uspcf{G}\leq\uspcf{G+v}$. Let $P$ be a parade in $H$. Let $H^+$ be obtained from $H$ by adding a vertex $v$ of degree $1$ adjacent to one of the endpoints of $P$. Let $P^+$ be the path in $H^+$ obtained from $P$ by adding the vertex $v$. Since $P$ is a parade in $H$, it has $|V(H)|-\uspcf{G}$ vertices. Therefore, $P^+$ has $|V(H^+)|-\uspcf{G}$ vertices. Thus, $\uspc{H^+}\leq\uspcf{G}$. But $G$ is a minor of $H^+$. Therefore, the reverse inequality holds, and we have $\uspc{H^+}=\uspcf{G}$. Since $G+v$ is a minor of $H^+$, we have $\uspcf{G+v}\leq\uspc{H^+}=\uspcf{G}\leq\uspcf{G+v}$. Therefore, we have equality, and the result holds. \end{proof} The next lemma tells us that if $H$ is a graph that realizes the spectator floor value of a graph $G$ with no isolated vertices, then $H$ also has no isolated vertices. \begin{lemma} \label{lem:no-isolated} Suppose $G$ is a graph without isolated vetices, and suppose $G$ is a minor of $H$. If $\uspcf{G}=\uspc{H}$, then $H$ has no isolated vertices. \end{lemma} \begin{proof} Suppose otherwise for a contradiction. Let $W$ be the set of isolated vertices of $H$. It is not difficult to see that deletion of such a vertex results in a graph with spectator number reduced by $1$. (This is because $G$ is neither empty nor $K_1$. Therefore, $H\neq K_1$.) Therefore, $\uspc{H-W}=\uspc{H}-|W|$. Since no vertex of $G$ is isolated, $G$ is a minor of $H-W$. This contradicts the assumption that $\uspcf{G}=\uspc{H}$. \end{proof} The next definition gives us a very precise definition of what it means to contract an edge in a path. This will be of much use in further proofs. \begin{definition} \label{def:P/e} Let $P$ be a path and $e$ an edge in a graph $G$. We define $P/e$ to be the subgraph of $G/e$ obtained from $P$ by contracting $e$. More precisely, if $v$ and $w$ are the endpoints of $e$ and $v'$ is the vertex that results from contracting $e$, then: \begin{itemize} \item If $P$ contains $e$, then $P/e$ is the path whose vertex set is $(V(P)-\{v,w\})\cup\{v'\}$ and whose edge set is $E(P)-\{e\}$. \item If $P$ contains exactly one endpoint of $e$ (say $v$), then $P/e$ is the path whose vertex set is $(V(P)-\{v\})\cup\{v'\}$ and whose edge set is $E(P)$. \item If $P$ contains both endpoints of $e$, but not $e$ itself, then $P/e$ is the graph (with exactly one cycle) whose vertex set is $(V(P)-\{v,w\})\cup\{v'\}$ and whose edge set is $E(P)$. \item If $P$ contains neither endpoint of $e$, then $P/e=P$. \end{itemize} \end{definition} The next lemma tells us that the property of being a unique shortest path is preserved after contraction of an edge in the path. \begin{lemma} \label{lem:still-usp} Let $G=H/e$. If $P$ is a unique shortest path in $H$ and $P$ contains $e$, then $P/e$ is a unique shortest path in $G$. \end{lemma} \begin{proof} Let $P$ be a unique shortest path in $H$, and suppose that the endpoints of $P$ are $v,w$ and $length(P)=\ell$. Let $Q_1,Q_2,Q_2,...,Q_N$ be all of the other paths from $v$ to $w$. Since $P$ is a unique shortest path, each $Q_i$ has $length(Q_i)\geq \ell+1$. For a path $Q_i$ that includes both endpoints of $e$ but does not include $e$, $Q_i/e$ is not a path in $G$, but every path that connects $v,w$ in $G=H/e$ is $P/e$ or $Q_i/e$ for some $i$. Clearly, $length(P/e)=\ell-1$. On the other hand, for each $Q_i/e$ that is a path, we have either $length(Q_i/e)=length(Q_i)\geq\ell+1$ or $length(Q_i/e)=length(Q_i)-1\geq \ell$. In either case, $P/e$ is the unique shortest path connecting $v,w$. \end{proof} The next lemma tells us that if contracting an edge in a parade increases the spectator number of the graph, then the resulting contracted parade is no longer a parade after contraction. \begin{lemma} \label{lem:no-endpoints} Let $G=H/e$. If $\uspc{H}<\uspc{G}$ and $P$ is a parade in $H$ and $P$ contains edge $e$, then $P/e$ is not a parade in $G$. \end{lemma} \begin{proof} Suppose not. Suppose $\uspc{H}<\uspc{G}$ and $P$ is a parade in $H$ and $P$ contains edge $e$ and $P/e$ is a parade in $G$. Then \[\usp{G}=|P/e|=|P|-1=\usp{H}-1\] However, since $\uspc{H}<\uspc{G}$, we also have \[|H|-\usp{H}<|G|-\usp{G}\] \[|G|+1-\usp{H}<|G|-\usp{G}\] \[1+\usp{G}<\usp{H}\] \[\usp{G}<\usp{H}-1\] \[\usp{G}\leq \usp{H}-2\] This is a contradiction. \end{proof} The next lemma tells us that contracting an edge contained in a parade cannot increase the spectator number of the graph. \begin{lemma} \label{lem:e not in pert} If $P$ is a parade in $H$ and $P$ contains both of the endpoints of edge $e$, then $\uspc{H}\geq\uspc{H/e}$. \end{lemma} \begin{proof} Since $P$ is a parade in $H$, note that $P$ must contain the edge $e$. Otherwise, $P$ is not a unique shortest path. Suppose for a contradiction that $\uspc{H}<\uspc{H/e}$. Call $H/e$ by the name $G$. We know from \cref{lem:no-endpoints} that $P/e$ is not a parade in $G$. But $P/e$ is still a unique shortest path by \cref{lem:still-usp}, so it must not be a longest unique shortest path anymore. However, there must be some parade in $G$. Let us call that $Q/e$, so that we can call the uncontracted version in $H$ by the name $Q$. Supposing $|P|=\ell$ in $H$, then $|Q|\leq \ell$ because $P$ was a parade in $H$. And since $P/e$ is not long enough to be a parade in $G$, $|Q/e|>\ell-1$. Putting these together, we get \[\ell-1<|Q/e|\leq |Q|\leq \ell.\] This implies that \[|Q/e|= |Q|= \ell.\] (That is, $Q/e=Q$, and $Q$ does not contain $e$.) Hence $\usp{G}=\usp{H}$. However, $\uspc{H}<\uspc{G}$ implies $\usp{G}\leq \usp{H}-2$. This is a contradiction. \end{proof} Finally, we have the main result of this section, which tells us that for any graph $G$, we can always find a supergraph $G'$ with the same number of vertices that realizes the spectator floor value of $G$. This is one of the major results of the paper and supports further results in other sections. We now prove \cref{lem:no decontract-intro} restated below. \begin{theorem} \label{lem:no decontract} For every graph $G$, there is a graph $G'$ with the same number of vertices as $G$ such that $G$ is a subgraph of $G'$ and such that $\uspcf{G}=\uspc{G'}$. \end{theorem} \begin{proof} By \cref{lem:isolated}, it suffices to consider the case where $G$ has no isolated vertices. Let $H$ be a graph such that $G$ is a minor of $H$ and such that $\uspcf{G}=\uspc{H}$. By \cref{lem:no-isolated}, $H$ has no isolated vertices. Therefore, by \cref{lem:no-delete-vertex}, there are sets $C,D$ of edges such that $G\cong H/C\backslash D$, without the necessity to delete vertices. Now, suppose that $H$ is such that $|C|$ is minimal among all such graphs. Suppose for a contradiction that $C\neq\emptyset$. For some edge $e\in C$, consider the graph $H/e$. Let $u$ and $v$ be the endpoints of $e$, and let $v'$ be the vertex that results from contracting $e$. By minimality of $C$, we have $\uspc{H}<\uspc{H/e}$. (Otherwise, the set of edges that need to be contracted from $H/e$ to obtain $G$ is smaller than $C$.) Let $P$ be a parade of $H$. By \cref{lem:e not in pert}, $P$ does not contain both endpoints of $e$. Therefore, $P$ is also a path in $H/e$. However, since $H/e$ has one fewer vertex than $H$, the fact that $\uspc{H}<\uspc{H/e}$ implies that $\usp{H/e}\leq\usp{H}-2$. This implies that $P$ is not a parade in $H/e$. Since $P$ is longer than the parades of $H/e$, there must be at least one path $Q\neq P$ in $H/e$ with the same length and endpoints as $P$. Let $\{P_1,\ldots,P_t\}$ be the set of all such paths $Q$. Since $P$ is a parade in $H$, these paths $P_1,\ldots P_t$ cannot be paths in $H$. Therefore, for each of these paths $P_i$ for $1\leq i\leq t$, there is a path $P_i'$ in $H$, containing $e$, such that $P_i'/e=P_i$. \begin{claim} \label{3vertices} $|V(P)|=\usp{H}\geq3$ \end{claim} \begin{subproof} Contracting an edge cannot result in an empty graph. Therefore, $|V(H/e)|\geq1$, implying that $\usp{H/e}\geq1$. Thus, since $\usp{H}\geq\usp{H/e}+2$, we have $\usp{H}\geq3$. \end{subproof} We will add an edge $f$ to $H/e$ to result in a graph $F$ such that $\uspc{F}\leq\uspc{H}$, contradicting the minimality of $H$. Let $x=v_0,e_1,v_1,\ldots,e_r,v_r=y$ be the succession of vertices and edges of $P$. We claim that each $P_i$ has only one subpath that diverges from $P$. In other words, we claim the following. \begin{claim} \label{1divergence} For each integer $i$ with $1\leq i\leq t$, there are nonnegative integers $m(i)< n(i)\leq r$, such that $P_i$ contains the vertices and edges $x=v_0,e_1,v_1,\ldots,e_{m(i)},v_{m(i)},v_{n(i)},e_{n(i)+1},\ldots,e_r,v_r=y$ but does not contain the vertices and edges $e_{m(i)+1},v_{m(i)+1},\ldots,v_{n(i)-1},e_{n(i)}$. \end{claim} \begin{subproof} Suppose for a contradiction that there are positive integers $m<n\leq p<q$ such that $P_i$ contains the vertices $v_m$, $v_n$, $v_p$, and $v_q$ but such that the subpaths of $P$ and $P_i$ from $v_m$ to $v_n$ are internally disjoint and such that the subpaths of $P$ and $P_i$ from $v_p$ to $v_q$ are internally disjoint. Without loss of generality, let $v'$ (obtained by contracting $e$) be on the subpath of $P_i$ from $v_m$ to $v_n$. (This includes the possibility that $v'=v_m$ or $v'=v_n$, in which case $v'$ is in $P$.) Let $Q$ be the graph consisting of the subpath of $P$ from $v_0$ to $v_p$, the subpath of $P_i$ from $v_p$ to $v_q$, and the subpath of $P$ from $v_q$ to $v_r$. Note that $Q$ contains a path in $H$ from $v_0$ to $v_r$ and therefore must have more vertices than $P$. Thus, the subpath of $P_i$ from $v_p$ to $v_q$ is longer than the subpath of $P$ from $v_p$ to $v_q$. Since $P$ and $P_i$ have the same length, this implies that the subpath of $P_i$ from $v_m$ to $v_n$ is shorter than the subpath of $P$ from $v_m$ to $v_n$. Let $R$ be the graph consisting of the subpath of $P$ from $v_0$ to $v_m$, the subpath of $P_i$ from $v_m$ to $v_n$, and the subpath of $P$ from $v_n$ to $v_r$. Thus $R$ contains a path in $H/e$ from $v_0$ to $v_r$ that is shorter than $P$. This implies the existence of a path in $H$ from $v_0$ to $v_r$ that is no longer than $P$, a contradiction. \end{subproof} Recall that, for each $i$ with $1\leq i\leq t$, the path $P_i$ in $H/e$ is $P_i'/e$, where $P_i'$ is a path in $H$ containing $e$, and recall that $v'$ is the resulting vertex when $e$ is contracted. By \cref{1divergence}, there is exactly one subpath of each $P_i'$ whose intersection with $P$ is the endpoints of the subpath. Let $p(i)$ be the length (number of edges) of the subpath of $P_i$ from $v_{m(i)}$ to $v'$, and let $q(i)$ be the length of the subpath of $P_i$ from $v'$ to $v_{n(i)}$. Let the sequence of vertices of $P_i$ be $x=v_0,v_1,\ldots,v_{m(i)},u_1,\ldots,u_{p(i)-1},u_{p(i)}=v'=w_{q(i)},w_{q(i)-1},\ldots,w_1,v_{n(i)},\ldots,v_r=y$. (See \cref{fig:PandP'}; note that it is possible $v_{m(i)}=v'$ or $v_{n(i)}=v'$, in which case $p(i)=0$ or $q(i)=0$, respectively.) \begin{figure}[!htbp] \[\begin{tikzpicture}[x=1cm, y=1cm] \node[vertex][fill,inner sep=1pt,minimum size=1pt] (v_0) at (-5,3)[label=left:$x= v_0$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (v_1) at (-4,3)[label=above:$v_1$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (vmi) at (-3,3)[label=above:$v_{m(i)}$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (vmi+1) at (-2,3)[label=below:$v_{m(i)+1}$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (vni-1) at (2,3)[label=below:$v_{n(i)-1}$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (vni) at (3,3)[label=above:$v_{n(i)}$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (vr-1) at (4,3)[label=above:$v_{r-1}$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (v_r) at (5,3)[label=right:$v_r= y$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (u_1) at (-2,2)[label=left:$u_1$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (up-1) at (-1,1)[label=left:$u_{p(i)-1}$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (v') at (0,0)[label=below:$u_{p(i)}= v'= w_{q(i)}$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (wq-1) at (1,1)[label=right:$w_{q(i)-1}$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (w_1) at (2,2)[label=right:$w_1$] {}; \draw(v_0.center) to (v_1.center) to (-3.7,3); \draw[dotted,semithick] (-3.6,3) to (-3.4,3); \draw (-3.3,3) to (vmi) to (vmi+1) to (-1,3); \draw[dotted,semithick] (-.25,3) to (.25,3); \draw (1,3) to (vni-1) to (vni) to (3.3,3); \draw[dotted,semithick] (3.6,3) to (3.4,3); \draw (3.7,3) to (vr-1) to (v_r); \draw(vni) to (w_1);\draw (vmi) to (u_1) to (-1.7,1.7); \draw[dotted,semithick] (-1.6,1.6) to (-1.4,1.4); \draw (-1.3,1.3) to (up-1) to (v') to (wq-1) to (1.3,1.3); \draw[dotted,semithick] (1.4,1.4) to (1.6,1.6); \draw (1.7,1.7) to (w_1) to (vni); \end{tikzpicture}\] \caption{The paths $P$ and $P_i$} \label{fig:PandP'} \end{figure} Based on this labeling of the vertices, we see that $|P|=|P_i|=1+m(i)+p(i)-1+q(i)+r-(n(i)-1)=m(i)+p(i)+q(i)+r-n(i)+1$. Recall that $P_i$ and $P_j$ have the same length, for $i,j\in\{1,\ldots,t\}$. Therefore, for all $i,j\in\{1,\ldots,t\}$, we have \[m(i)+p(i)+q(i)-n(i)=m(j)+p(j)+q(j)-n(j).\] \begin{claim} For all $i,j\in\{1,\ldots,t\}$, we have $m(i)+p(i)=m(j)+p(j)$ and $q(i)-n(i)=q(j)-n(j)$. \end{claim} \begin{subproof} Suppose for a contradiction that $m(i)+p(i)<m(j)+p(j)$. Since $m(i)+p(i)+q(i)-n(i)=m(j)+p(j)+q(j)-n(j)$, we have $q(i)-n(i)>q(j)-n(j)$. Consider the subgraph of $H/e$ consisting of the subpath of $P_i$ from $x$ to $v'$ and the subpath of $P_j$ from $v'$ to $y$. This subgraph contains a path from $x$ to $y$ whose number of vertices is at most $m(i)+p(i)+q(j)+r-n(j)+1<m(j)+p(j)+q(j)+r-n(j)+1=|P|$. (This subgraph \emph{is} such a path if the intersection of the paths used to form it consists only of $v'$.) Regardless of whether this path contains $v'$, it implies the existence of a path in $H$ from $x$ to $y$ whose length is no longer than that of $P$, a contradiction. Thus, we have proved the first equality of the claim. The second equality follows from the first equality and the fact that $m(i)+p(i)+q(i)-n(i)=m(j)+p(j)+q(j)-n(j)$. \end{subproof} \begin{claim} \label{mlessn} For all $i,j\in\{1,\ldots,t\}$, we have $m(i)<n(j)$. \end{claim} \begin{subproof} Suppose for a contradiction that $m(i)\geq n(j)$. It follows from \cref{1divergence} that the number of vertices on the subpaths of $P$ and $P_i$ from $v_{m(i)}$ to $v_{n(i)}$ must be equal. Similarly, the number of vertices on the subpaths of $P$ and $P_j$ from $v_{m(j)}$ to $v_{n(j)}$ must be equal. The number of vertices of the subpath of $P$ from $v_{m(i)}$ to $v_{n(i)}$ is $n(i)-m(i)+1$, and the number of vertices of the subpath of $P$ from $v_{m(j)}$ to $v_{n(j)}$ is $n(j)-m(j)+1$. The number of vertices of the subpath of $P_i$ from $v_{m(i)}$ to $v_{n(i)}$ is $p(i)+q(i)+1$, and number of vertices of the subpath of $P_j$ from $v_{m(j)}$ to $v_{n(j)}$ is $p(j)+q(j)+1$. Thus, we have $n(i)-m(i)=p(i)+q(i)$ and $n(j)-m(j)=p(j)+q(j)$. Consider the subgraph of $H/e$ consisting of the subpath of $P_j$ from $x$ to $v'$ and the subpath of $P_i$ from $v'$ to $y$. This subgraph has at most $m(j)+p(j)+q(i)+r-n(i)+1$ vertices. Therefore, it contains a path $\widehat{P}$ from $x$ to $y$ such that $|\widehat{P}|\leq m(j)+p(j)+q(i)+r-n(i)+1$. This path must be at least as long as $P$. Therefore, $r+1\leq m(j)+p(j)+q(i)+r-n(i)+1$, implying that $n(i)-m(j)\leq p(j)+q(i)$. On the other hand, $n(i)-m(j)\geq n(i)-m(i)+n(j)-m(j)=p(i)+q(i)+p(j)+q(j)$. This leads to a contradiction unless $p(i)=q(j)=0$. Thus, $v_{m(i)}=v'=v_{n(j)}$ is a vertex of $P$. This implies $m(i)=n(j)$. Let $u$ be the endpoint of $e$ not on $P$. Consider the subgraph of $H$ consisting of the subpath of $P_j'$ from $x$ to $u$ and the subpath of $P_i$ from $u$ to $y$. This subgraph contains a path from $x$ to $y$ whose number of vertices is at most $m(j)+p(j)+q(i)+r-n(i)+1=m(i)+p(i)+q(j)+r-n(j)+1=r+1=|P|$, a contradiction. \end{subproof} Let $m=\max\{m(i):1\leq i\leq t\}$ and $n=\min\{n(i):1\leq i\leq t\}$. By \cref{mlessn}, we have $m<n$. Recall from \cref{3vertices} that $|P|\geq3$. Therefore, either $n\geq2$ or $m+2\leq r$. (Otherwise, $m<n<2$, which implies $m=0$. If $m=0$ and $m+2>r$, then $r<2$, implying $|P|=r+1<3$.) Without loss of generality, we assume $m+2\leq r$, reversing the order of the vertices and edges of $P$ if necessary. Thus, $v_{m+2}$ is a vertex of $P$. Construct the graph $F$ from $H/e$ by adding the edge $f$ joining $v_m$ and $v_{m+2}$. We claim that the resulting path $P_F$ whose sequence of vertices is $v_0,\ldots,v_m,v_{m+2},\ldots,v_r$ is the unique shortest path between $x$ and $y$ in $F$. Suppose otherwise for a contradiction. Then $F$ contains a path $P_F'\neq P_F$ from $x$ to $y$ with $|P_F'|\leq|P_F|=|P|-1$. Suppose $P_F'$ is a path in $H/e$. Then $H$ contains a path $Q$ from $x$ to $y$ with at most $|P|$ vertices such that either $Q=P_F'$ or $Q=P_F'/e$. Since $P$ is the unique shortest path from $x$ to $y$ in $H$, we muat have $Q=P$. Thus, $P_F'=Q=P$ since $P$ does not contain $e$. This contradicts the assumption that $|P_F'|\leq|P|-1$. Therefore, $P_F'$ is not a path in $H/e$. Thus, $P_F'$ contains the edge $f$. Let $P_F''$ be the path obtained from $P_F'$ by replacing edge $f$ with the subpath of $P$ from $v_m$ to $v_{m+2}$. Then $|P_F''|=|P_F'|+1\leq|P|$. Since $P_F''$ is a path in $H/e$, this implies that $P_F''=P_i$ for some $i\in\{1,\ldots,t\}$. Since $v_m$ is a vertex of $P_i$, we must have $m(i)=m$. Since $v_{m+2}$ is a vertex of $P_i$, we must have $n(i)\in\{m+1,m+2\}$. But then $P_F'=P_F$, a contradiction. Thus, we have shown that $P_F$ is the unique shortest path between $x$ and $y$ in $F$. Therefore, $\usp{F}\geq|V(P)|-1=\usp{H}-1$. Thus, $\uspc{F}=|V(H/e)|-\usp{F}=|V(H)|-1-\usp{F}\leq|V(H)|-1-\usp{H}+1=\uspc{H}$, contradicting the minimality of $H$. \end{proof} \section{Disconnected Graphs} \label{Disconnected Graphs} In this section, we establish that the spectator floor is additive over disconnected graph components. We proved \cref{prop:components} independently of \cref{lem:no decontract} and have decided to include both proofs in their entirety. However, it is interesting to note that either result could be used to prove the other. We will now prove \cref{prop:components-intro} restated below. \begin{theorem} \label{prop:components} For any graph $G=G_1\sqcup G_2$ with no edges connecting $G_1$ to $G_2$, \[\uspcf{G}=\uspcf{G_1}+\uspcf{G_2}\] \end{theorem} \begin{proof} Let $G=G_1\sqcup G_2$, where $G$ contains no edges connecting $G_1$ to $G_2$. The statement is trivially true when either $G_1$ or $G_2$ is empty, so assume that neither $G_1$ nor $G_2$ is empty. For $i=1,2$, let $F_i$ be a graph that realizes $\uspcf{G_i}$; that is, $F_i$ is such that $G_i$ is a minor of $F_i$, and $F_i$ has a unique shortest path $P_i$ with $|F_i|-\uspcf{G_i}$ vertices. Let $F$ be $F_1\sqcup F_2$ with an edge $e$ added to make a path $P$ that connects $P_1$ and $P_2$ end-to-end. Then $G$ is a minor of $F$, and $P$ is a unique shortest path in $F$ with \[|F_1|-\uspcf{G_1}+|F_2|-\uspcf{G_2}=|F|-\uspcf{G_1}-\uspcf{G_2}\] vertices. Hence $\uspc{F}\leq \uspcf{G_1}+\uspcf{G_2}$ and so $\uspcf{G}\leq \uspcf{G_1}+\uspcf{G_2}$. Now, contrary to the statement of the result, suppose that $\uspcf{G}< \uspcf{G_1}+\uspcf{G_2}$. Then there is some graph $H$, of which $G$ is a minor, such that $\uspc{H}=\uspcf{G}<\uspcf{G_1}+\uspcf{G_2}$. Since $G$ is a minor of $H$, we can build $H$ from $G$ by a sequence of decontracting vertices, adding edges, and/or adding isolated vertices. We will now construct a partition of $H$ into three sets: a subgraph $H_1$, a subgraph $H_2$, and a set of ``bridge'' edges $B$. Let us start by defining $H_1$ as $G_1$ and $H_2$ as $G_2$. Next, for each time that a vertex is decontracted while transforming $G$ into $H$, any vertices or edges that are doubled by the decontraction will remain in the same set that they were in before. Whenever a new edge is added, if it connects two vertices in the same $H_i$, that edge will be added to $H_i$. If the new edge connects a vertex from $H_1$ to a vertex from $H_2$, the edge will be added to the bridge edge set $B$. Whenever a new isolated vertex is added, it will be randomly placed into either $H_1$ or $H_2$. At the end of this process, we have that all the vertices of $H$ are in either $H_1$ or $H_2$, all the edges of $H$ are in $H_1$, $H_2$, or $B$, and no vertex or edge is in more than one of these. Furthermore, $G_1$ is a minor of $H_1$ and $G_2$ is a minor of $H_2$. Let $P$ be a path in $H$ that contains $\usp{H}$ vertices. $P$ cannot be entirely within either $H_1$ or $H_2$ for the following reasons. Suppose, without loss of generality, that $P$ is a subgraph of $H_1$. Since $G$ is a minor of $H_1\cup G_2$, that means $\uspcf{G}\leq \uspcf{H_1\cup G_2}$. Since $H_1$ and $G_2$ are disjoint, by the same arguments used for $G,G_1,G_2$ previously, we can conclude that $\uspcf{H_1\cup G_2}\leq \uspcf{H_1}+\uspcf{G_2}$. Hence, $\uspcf{G}\leq \uspcf{H_1}+\uspcf{G_2}$. This implies that \begin{align*} \uspcf{G}\leq \uspc{H_1}+\uspcf{G_2} &=|H_1|-\usp{H_1}+\uspcf{G_2}\\ &=|H_1|-\usp{H}+\uspcf{G_2}\\ &=|H_1|+|H_2|-\usp{H}+\uspcf{G_2}-|H_2|\\ &=|H|-\usp{H}+\uspcf{G_2}-|H_2|\\ &=\uspc{H}+\uspcf{G_2}-|H_2| \end{align*} By definition, $\uspcf{G}=\uspc{H}$, so the above inequality simplifies to $0\leq \uspcf{G_2}-|H_2|$. Furthermore, $\uspcf{G_2}\leq \uspc{G_2}=|G_2|-\usp{G_2}$. Hence, $0\leq |G_2|-\usp{G_2}-|H_2|$. However, since $G_2$ is a minor of $H_2$, the quantity $|G_2|-|H_2|\leq 0$. Hence, we have $0\leq -\usp{G_2}$, which implies that $G_2$ is an empty graph, in contradiction to our assumption that neither $G_1$ nor $G_2$ is empty. Returning to our discussion of the graph $H$, we now know that the path $P$ with $\usp{H}$ vertices must contain vertices from both $H_1$ and $H_2$, and so $P$ must contain at least one edge in $B$. We will now show that this leads to a contradiction as well. Suppose that $P$ contains $m$ edges in $B$, where $m\geq 1$. Assume without loss of generality that $P$ has at least one end in $H_1$. Then we can break $P$ up into disjoint subpaths $Q_1,Q_2,...,Q_j$ in $H_1$ and $R_1,R_2,...,R_k$ in $H_2$, such that $P$ consists of $Q_1$ followed by $R_1$ followed by $Q_2$ followed by $R_2$, etc., with these subpaths linked together by edges in the bridge $B$. (See \cref{disconnected_diagram}.) Note that if $m$ is even, then $j=k+1$ and $m=2k$, whereas if $m$ is odd, then $j=k$ and $m=2k-1$. \begin{figure}[!htbp] \[\begin{tikzpicture}[x=1cm, y=1cm] \fill[gray!20,rounded corners] (-5,2) rectangle (-1,6); \fill[gray!20] (-5,1) rectangle (-1,3); \fill[gray!20,rounded corners] (5,2) rectangle (1,6); \fill[gray!20] (5,1) rectangle (1,3); \shade[ left color = gray!20, right color = white, shading angle = 0 ] (5,0.5) rectangle (1,1); \shade[ left color = gray!20, right color = white, shading angle = 0 ] (-5,0.5) rectangle (-1,1); \draw[rounded corners] (-1,1) -- (-1,6) -- (-5,6) -- (-5,1); \draw[rounded corners] (1,1) -- (1,6) -- (5,6) -- (5,1); \draw[dotted,semithick] (1,1) -- (1,0.5); \draw[dotted,semithick] (5,1) -- (5,0.5); \draw[dotted,semithick] (-1,1) -- (-1,0.5); \draw[dotted,semithick] (-5,1) -- (-5,0.5); \node at (-3.5,6)[label=above:$H_1$] {}; \node at (3.5,6)[label=above:$H_2$] {}; \node at (0,6)[label=above:$B$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a1) at (-3,5.5)[label=below:$a_1$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b1) at (-1.5,5.5)[label=below left:$b_1$]{}; \draw (a1) to (-2.5,5.5); \draw[dotted, semithick] (-2.5,5.5) to (-2,5.5); \draw (-2,5.5) to (b1); \node at (-4,5.5) {$Q_1$}; \node at (-4,4) {$Q_2$}; \node at (-4,2) {$Q_3$}; \node at (4,5) {$R_1$}; \node at (4,3) {$R_2$}; \node at (4,1) {$R_3$}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (c1) at (1.5,5.5)[label=below right:$c_1$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (d1) at (1.5,4.5)[label=below right:$d_1$]{}; \draw (c1) to (2.5,5.5); \draw (2.5,4.5) to (d1); \draw (2.5,5.5) arc (90:45:0.5); \draw (2.5,4.5) arc (-90:-45:0.5); \draw[dotted,semithick] (2.5,5.5) arc (90:-90:0.5); \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a2) at (-1.5,4.5)[label=below left:$a_2$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b2) at (-1.5,3.5)[label=below left:$b_2$]{}; \draw (a2) to (-2.5,4.5); \draw (-2.5,3.5) to (b2); \draw (-2.5,4.5) arc (90:135:0.5); \draw (-2.5,3.5) arc (-90:-135:0.5); \draw[dotted,semithick] (-2.5,4.5) arc (90:270:0.5); \draw (b1) to[in=150,out=30] (c1); \draw (a2) to[in=150,out=30] (d1); \node[vertex][fill,inner sep=1pt,minimum size=1pt] (c2) at (1.5,3.5)[label=below right:$c_2$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (d2) at (1.5,2.5)[label=below right:$d_2$]{}; \draw (c2) to (2.5,3.5); \draw (2.5,2.5) to (d2); \draw (2.5,3.5) arc (90:45:0.5); \draw (2.5,2.5) arc (-90:-45:0.5); \draw[dotted,semithick] (2.5,3.5) arc (90:-90:0.5); \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a3) at (-1.5,2.5)[label=below left:$a_3$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b3) at (-1.5,1.5)[label=below left:$b_3$]{}; \draw (a3) to (-2.5,2.5); \draw (-2.5,1.5) to (b3); \draw (-2.5,2.5) arc (90:135:0.5); \draw (-2.5,1.5) arc (-90:-135:0.5); \draw[dotted,semithick] (-2.5,2.5) arc (90:270:0.5); \draw (b2) to[in=150,out=30] (c2); \draw (a3) to[in=150,out=30] (d2); \node[vertex][fill,inner sep=1pt,minimum size=1pt] (c3) at (1.5,1.5)[label=below right:$c_3$]{}; \draw (b3) to[in=150,out=30] (c3); \draw (c3) to (2.5,1.5); \draw (2.5,1.5) arc (90:45:0.5); \draw[dotted,semithick] (2.5,1.5) arc (90:0:0.5); \draw[dashed] (b1) to[out=-45,in=45] (a2); \draw[dashed] (b2) to[out=-45,in=45] (a3); \draw[dashed] (d1) to[out=-135,in=135] (c2); \draw[dashed] (d2) to[out=-135,in=135] (c3); \end{tikzpicture}\] \caption{A generalized diagram of the subgraph of $H$ induced by the path $P$. Dashed lines indicate edges that will be added to form the graphs $H_1'$, $H'$, $H_2''$, and $H''$.}\label{disconnected_diagram} \end{figure} Furthermore, let us label the first vertex of each $Q_i$ by the name $a_i$, and the last vertex of each $Q_i$ by $b_i$, the first vertex of each $R_i$ by $c_i$, and the last vertex of each $R_i$ by $d_i$. Hence, each vertex $b_i$ is connected to the vertex $c_i$ by an edge in $B$, except possibly for $b_j$, and each vertex $d_i$ is connected to the vertex $a_{i+1}$ by an edge in $B$, except possibly for $d_k$. Let $p$ stand for the number of vertices in $P$, $q$ the total number of vertices in $Q_1,Q_2,...,Q_j$, and $r$ the total number of vertices in $R_1,R_2,...,R_k$. Hence $p=q+r$. By assumption, $\uspc{H}=|H|-p<\uspcf{G_1}+\uspcf{G_2}$. Since $|H|-p=|H_1|-q+|H_2|-r$, it must be the case that either $|H_1|-q<\uspcf{G_1}$ or $|H_2|-r<\uspcf{G_2}$ (since, if both of these were false, it would contradict the original inequality). However, we will now show that neither of these is true. Observe that since $P$ is a unique shortest path in $H$, there cannot be an edge in $H_1$ connecting any $b_i$ to $a_{i+1}$, because that would form a shorter path. Likewise, there cannot be an edge in $H_2$ connecting any $d_i$ to $c_{i+1}$. Let $H_1'$ (respectively, $H'$) be formed by taking $H_1$ (respectively, $H$) and adding an edge from each $b_i$ to $a_{i+1}$, except for $b_j$, the last one. (Note that if $j=1$, then $H_1'=H_1$.) These edges are indicated by dashed lines in \cref{disconnected_diagram}. Since $P$ was a unique shortest path in $H$ and none of these new edges were present previously, we now have a shorter unique path connecting the endpoints of $P$. Let us call the portion of this new unique shortest path that lies in $H_1'$ by the name $P'$. (Unless $j=1$, then let $P'$ be the portion of $P$ in $H_1$.) Since any subpath of a unique shortest path is also a unique shortest path, $P'$ is a unique shortest path with $q$ vertices contained in $H_1'$, hence $\uspc{H_1'}\leq |H_1'|-q=|H_1|-q$. Furthermore, since $G_1$ is a minor of $H_1'$, we have $\uspcf{G_1}\leq \uspc{H_1'}\leq |H_1|-q$. Now, let $H_2''$ (respectively, $H''$) be formed by taking $H_2$ (respectively, $H$) and adding an edge from each $d_i$ to $c_{i+1}$, except for $d_k$, the last one. (Note that if $k=1$, then $H_2''=H_2$.) These edges are indicated by dashed lines in \cref{disconnected_diagram}. Since $P$ was a unique shortest path in $H$ and none of these new edges were present previously, we now have a shorter unique path connecting the endpoints of $P$. Let us call the portion of this new unique shortest path that lies in $H_2''$ by the name $P''$. (Unless $k=1$, then let $P''$ be the portion of $P$ in $H_2$.) Since any subpath of a unique shortest path is also a unique shortest path, $P''$ is a unique shortest path with $r$ vertices contained in $H_2''$, hence $\uspc{H_2''}\leq |H_2''|-r=|H_2|-r$. Furthermore, since $G_2$ is a minor of $H_2''$, we have $\uspcf{G_2}\leq \uspc{H_2''}\leq |H_2|-r$. \end{proof} The following is a direct consequence of the last proposition. \begin{cor} \label{cor:components} For any graph $G$ that consists of connected components $G_1, G_2, ..., G_n$, \[\uspcf{G}=\uspcf{G_1}+\uspcf{G_2}+\cdots+\uspcf{G_n}\] \end{cor} \section{Trees} \label{trees} This section contains results that apply to trees, with the exception of \cref{diam.bound} and \cref{lem:diameter}, which offer bounds that apply to all graphs, and were inspired by the consideration of trees. We begin by noting that since $\usp{G}$ is the number of vertices in the longest unique shortest path and $\diam(G)$ is the length of the longest shortest path in $G$, which may or may not be unique, we have the following: \begin{obs}\label{diam.bound} For any graph $G$, $\usp{G} \leq \diam(G) + 1$ and so $\uspc{G} \geq |G| - \diam(G) - 1$. \end{obs} Next, we parlay \cref{diam.bound} into an exact formula for the spectator floor of a tree. \begin{theorem}\label{tree.uspcf} For a tree $T$, $\uspcf{T}=|T|-\diam(T)-1$. \end{theorem} \begin{proof} Let $T$ be a tree. Since all paths in a tree are unique, $\usp{T}=\diam(T)+1$, and so $\uspc{T}=|T|-\diam(T)-1$. Suppose that $\uspcf{T}<|T|-\diam(T)-1$. Then there exists some graph $G$ which has $T$ as a minor, such that $\uspcf{T}=\uspc{G}<|T|-\diam(T)-1$. The bound $\usp{G} \leq \diam(G)+1$ implies that $|G|-\diam(G)-1\leq \uspc{G}$. Then we have \[|G|-\diam(G)-1<|T|-\diam(T)-1\] and rearranging this, we get \[|G|-|T|<\diam(G)-\diam(T).\] The quantity on the left, $|G|-|T|$, is the number of decontractions performed to transform $T$ into $G$. Note that one decontraction can increase the diameter of a graph by at most 1. Furthermore, adding an edge cannot increase the diameter of a graph. Therefore, the quantity $\diam(G)-\diam(T)$ must be less than or equal to the number of decontractions performed to transform $T$ into $G$. This is a contradiction. \end{proof} The next theorem is important in establishing \cref{cor:diam-paths}, which tells us the conditions under which a tree is minor minimal for a given value of the spectator floor. First we need two definitions. A \emph{diametric path} of a graph is a path whose length is the diameter. The endpoints of such a path are called a \emph{diametric pair} of vertices. \begin{theorem} For any tree $T$, there exists a non-empty elementary minor of $T$ with the same spectator floor as $T$ if and only if there exists an edge $e$ of $T$ such that contracting $e$ reduces the diameter of $T$. \end{theorem} \begin{proof} The reverse implication is the easy direction: Suppose that $e$ is an edge such that the contracted tree $T^\prime = T/e$ has diameter less than the diameter of $T$. The distance between vertices in a tree is given by the unique path that joins them, and so contracting $e$ in $T$ reduces the distance of any pair by exactly $1$ in the case that $e$ belongs to the unique path joining the pair, and leaves the distance unchanged otherwise. In particular, the difference between the diameter of $T^\prime$ and the diameter of $T$ can be at most one, and so $\diam(T^\prime) = \diam(T) - 1$. The number of vertices also differs by exactly $1$, and so the elementary minor $T^\prime$ of $T$ satisfies \[ \uspcf{T^\prime}=|T^\prime|-\diam(T^\prime)-1 = |T|-\diam(T)-1 = \uspcf{T}. \] For the forward implication, there are three sorts of elementary minors. Since $T$ is a connected tree, deletion of an isolated vertex can only happen in the case $T=K_1$, leaving the empty graph as a minor. Contraction of an edge $e$ that leaves the spectator number unchanged must reduce the diameter by $1$ by the same calculation as above. This leaves only the case of edge deletion. Assume by way of contradiction, then, firstly that there does exist a specific edge $e$ whose deletion produces a disjoint union of trees $T_1$ and $T_2$ satisfying \[ \uspcf{T_1 \cup T_2} = \uspcf{T_1} + \uspcf{T_2} = \uspcf{T}, \] and secondly that no edge contraction reduces the diameter of $T$. Using the diameter formula to substitute for $\uspcf{T_1}$, $\uspcf{T_2}$, and $\uspcf{T}$ produces an equation \[ |T_1| - \diam(T_1) - 1 + |T_2| - \diam(T_2) - 1 = |T| - \diam(T) - 1 \] whose simplification \[ \diam(T) = \diam(T_1) + \diam(T_2) + 1 \] implies the strict inequality \[ \diam(T) > \diam(T_1) \] since $\diam(T_2) \ge 0$. If no edge contraction reduces the diameter of $T$, then in particular the contraction $T^\prime = T/e$ has the same diameter as $T$. Let $p^\prime$ and $q^\prime$ be a diametric pair of vertices in $T^\prime$; then $e$ cannot be part of the unique path that joins their preimages $p$ and $q$ in $T$. It follows that $p$ and $q$ are in the same component $T_1$ or $T_2$ of $T\setminus e$. Without loss of generality, both are in $T_1$, and thus that the diameter of $T_1$ is at least the diameter of $T$: \[ \diam(T) \le \diam(T_1). \] This contradicts the previous strict inequality and completes the proof. \end{proof} \begin{cor} \label{cor:diam-paths} A tree $T$ is minor-minimal for the spectator floor if and only if no edge lies in the intersection of all diametric paths in $T$. \end{cor} \begin{proof} The contraction of an edge $e$ reduces the diameter of $T$ if and only $e$ lies in the intersection of all diametric paths in $T$, and $T$ is minor-minimal for the spectator floor if and only if no elementary minor of $T$ has the same spectator floor. \end{proof} As a result of \cref{cor:diam-paths}, we have the following, which tells us that star graphs are minor minimal for a given value of the spectator floor. \begin{cor} \label{cor:K1k+2} For $k\geq1$, the graph $K_{1,k+2}$ is minor-minimal among graphs with spectator floor $k$. \end{cor} \begin{proof} By \cref{tree.uspcf}, $\uspcf{K_{1,k+2}}=k+3-2-1=k$. Since $k+2\geq3$, no edge of $K_{1,k+2}$ lies in all of the diametric paths of $K_{1.k+2}$. Therefore, by \cref{cor:diam-paths}, $K_{1,k+2}$ is minor-minimal among graphs with spectator floor $k$. \end{proof} In the final result of this section, we slightly improve upon the bound of \cref{diam.bound}. \begin{theorem} \label{lem:diameter} For every graph $G$, we have $\uspcf{G}\geq|G|-\diam(G)-1$. Moreover, if $\usp{G}\leq\diam(G)$, then $\uspcf{G}\geq|G|-\diam(G)$. \end{theorem} \begin{proof} By \cref{lem:no decontract}, there is a supergraph $G'$ of $G$ such that $\uspcf{G}=\uspc{G'}$ and $|G|=|G'|$. Since $G'$ is obtained from $G$ by adding edges but no vertices, we have $\diam(G')\leq\diam(G)$. Recalling \cref{diam.bound}, we have $\uspcf{G}=\uspc{G'}\geq|G'|-\diam(G')-1\geq|G|-\diam(G)-1$. Now, suppose $\usp{G}\leq\diam(G)$. Consider a parade in $G'$, and let $u$ and $v$ be its endpoints. We will show that the length of this parade in $G'$ is at most $\diam(G)-1$. If the distance between $u$ and $v$ in $G$ is at most $\diam(G)-1$, then the distance between $u$ and $v$ in $G'$ must be at most $\diam(G)-1$ also. On the other hand, consider the case that the distance in $G$ between $u$ and $v$ is $\diam(G)$. Since $\usp{G}\leq\diam(G)$, every parade in $G$ has at most $\diam(G)$ vertices, which implies that every parade in $G$ has length at most $\diam(G)-1$. Thus, the shortest paths joining $u$ and $v$ in $G$ are not unique. These paths are all still present in $G'$. Thus, since $u$ and $v$ are joined by a unique shortest path in $G'$, this path must have length at most $\diam(G)-1$. Thus, we have $\usp{G'}\leq\diam(G)$, implying that $\uspcf{G}=\uspc{G'}\geq|G'|-\diam(G)=|G|-\diam(G)$. \end{proof} \section{Calculation of the Spectator Floor of Simple Graphs} \label{calculation of spectator floor} If $A$ is the adjacency matrix of a graph $G$ and $k$ is a nonnegative integer, then the $(i, j)$-entry of $A^k$ counts the number of distinct walks of length exactly $k$ from vertex $i$ to vertex $j$ in $G$, including for example any paths of order $k + 1$. This leads to the following efficient way of computing the parade number of a given graph. \begin{obs}\label{USP-in-poly-time} Let $G$ be a connected graph and let $k$ range from $0$ to $n$. Then \[\usp{G} = 1 + \max\{k: (A+2I)^k\text{ has a 1 in some entry}\}.\] \end{obs} We use $(A + 2I)^k$ rather than $A^k$ in order to include a contribution of at least $2A^j$ for all $j < k$, which ensures that an entry equal to $1$ represents not just a unique walk but a unique shortest walk, and therefore a unique shortest path. The calculation terminates once every entry is strictly greater than $1$. Recall that the binomial expansion of $(A+2I)^k$ includes all powers of $A$ from $A^0=I$ through $A^k$ so the $(i,j)$-entry of $(A+2I)^k$ is a weighted accumulation of the number of walks from $i$ to $j$ of length at most $k$. The need of $2I$ rather than $I$ can be seen by considering $K_1$; the parade number of $K_1$ is one but the adjacency matrix is $[0]$ and $([0] + I)^k = [1]$ for all $k$. Note that the choice of the multiplier 2 is arbitrary, any integer greater than one is sufficient. By Theorem \ref{lem:no decontract}, in order to compute the spectator floor of $G$, we need only compare the spectator number of $G$ to the spectator number of each super graph of $G$. Moreover, by Corollary \ref{cor:components}, the spectator floor sums over connected components. This leads to Algorithm \ref{algo:spec-floor}, which calculates the spectator floor of every connected graph on at most $N$ vertices. Once the spectator floor number has been calculated for all connected graphs on at most $N$ vertices, one can than use Algorithm \ref{algo:minor-minimal} to determine which of those graphs are minor-minimal with respect to the spectator floor number. These algorithms were implemented in a SageMath \cite{sagemath} program to calculate the spectator floor of simple graphs and to determine which simple graphs are minor-minimal, with the code available at \cite{spec_floor_github_repo}. Note that while calculating the spectator floor for a fixed graph on a small number of vertices is relatively quick, the program takes over four days to run over all connected simple graphs on at most 10 vertices and would take over 400 days to process the connected simple graphs on 11 vertices since the complex nested nature of the subgraph-supergraph relationship does not allow for a simple parallelization of Algorithm \ref{algo:spec-floor}. Since this program takes so long to run and produces a large data set, that data and functions to access that data are also available in the GitHub repository at \cite{spec_floor_github_repo}. \SetKwComment{Comment}{/* }{ */} \RestyleAlgo{ruled} \begin{algorithm}[!h] \caption{Calculate $\uspcf{G}$ for all connected graphs on at most $N$ vertices}\label{algo:spec-floor} \KwData{$N$, maximum number of vertices} \KwResult{All connected graphs on up to $N$ vertices with their spectator floor number} \emph{$L \gets$ empty list}\; \For{$n\leq N$} { \For{$e = n(n-1)/2, \dots, n-1$} { \ForAll{connected graphs $G$ on $n$ vertices and $e$ edges} { $\usp{G} \gets 1 + \max\{k: (A+2I)^k\text{ has a 1 in some entry}\}$;\ $\uspc{G} \gets n - \usp{G}$; $\uspcf{G} \leftarrow \uspc{G}$\Comment*[r]{Initialize the value of $\uspcf{G}$} \ForAll{the $\widehat{G}$, super graphs of $G$} { \If{$\uspc{\widehat{G}} \leq \uspc{G}$} { $\uspcf{G} \leftarrow \uspc{\widehat{G}}$; } } append $(G, \uspcf{G})$ pair to $L$; } } } \Return{$L$} \end{algorithm} \RestyleAlgo{ruled} \begin{algorithm}[!h] \caption{Identify minor-minimal graphs}\label{algo:minor-minimal} \KwData{The graph-spectator floor pairs resulting from Algorithm \ref{algo:spec-floor}} \KwResult{All minor-minimal (w.r.t $\uspcf{G}$) connected graphs on up to $N$ vertices} \emph{$L \gets$ empty list}\; \For{$n\leq N$} { \For{$e = n(n-1)/2, \dots, n-1$} { \ForAll{connected graphs $G$ on $n$ vertices and $e$ edges} { Determine the minors of $G$; \ForEach{minor $H$ of $G$} {\If{$\uspcf{H} = \uspcf{G}$} { go to top of forall section\Comment*[r]{$G$ is not minor-minimal} } } append $G$ to $L$\Comment*[r]{$G$ is minor-minimal} } } } \Return{$L$} \end{algorithm} \section{Graphs of Spectator Floor 0, 1, and 2} \label{sec:minimal} In this section, we give the complete list of minor-minimal graphs of spectator floor 0, 1, and 2. Along the way, we characterize those graphs with $\uspc{G} > 0$, $\uspc{G} > 1$, and $\uspc{G} > 2$, allowing quick recognition of such graphs. We begin with graphs of spectator floor 0. \begin{prop} \label{lem:path floor} Let $G$ be a graph. Then $\uspcf{G}=0$ if and only if $G$ is a disjoint union of paths. \end{prop} \begin{proof} By definition of $\uspc{G}$ and $\uspcf{G}$, it is clear that a path $P$ has $\uspc{P}=0$ and $\uspcf{P}=0$. First suppose $G$ is the disjoint union of paths $P_1,P_2,\ldots P_t$, and let the endpoints of $P_i$ be $x_i$ and $y_i$. Form a supergraph $G'$ of $G$ by adding edges joining $y_i$ and $x_{i+1}$, for $1\leq i\leq t-1$. Since $G'$ is a path, we have $\uspc{G'}=0$ and therefore $\uspcf{G}=0$. Now, suppose $\uspcf{G}=0$. There must be a graph $G'$ such that $G$ is a minor of $G'$ and such that all vertices of $G'$ are contained in a unique shortest path $P$. Any edge added in parallel to an edge of $P$ causes $P$ to no longer be unique. If any other edge is added, it causes $P$ to not be a shortest path. Therefore, $G'$ must be a path. Since $G$ is a minor of a path, $G$ must be a disjoint union of paths. \end{proof} We now turn our attention to minor-minimal graphs with spectator floor $1$. The first results is a corollary that follows from \cref{lem:path floor}. \begin{cor} \label{cor:min-multi-1} The complete list of minor-minimal graphs with spectator floor $1$ is $C_2$ and $K_{1,3}$. The complete list of minor-minimal simple graphs with spectator floor $1$ is $K_3$ and $K_{1,3}$. \end{cor} \begin{proof} For the first statement, if $G$ does not have spectator floor 0, then by \cref{lem:path floor} $G$ is not a disjoint union of paths, and so $G$ either has a vertex of degree at least $3$, or $G$ contains a cycle. If $G$ has a vertex of degree at least $3$, then it contains $K_{1,3}$ as a subgraph; if $G$ contains a cycle, then $G$ con be contracted to $C_2$ (for multigraphs) or $K_3$ (for simple graphs). \end{proof} Next we begin our investigation of minor-minimal graphs with spectator floor 2. The following lemma will be used later to prove that certain graphs have spectator floor 2. \begin{lemma} \label{lem:figures} All of the graphs in \cref{fig:sharevertex,fig:shareedges,fig:one-cycle,fig:one-cycle2} have spectator floor $1$. \end{lemma} \begin{proof} Consider the graphs in \cref{fig:sharevertex,fig:shareedges,fig:one-cycle,fig:one-cycle2}. (The vertex labels and captions will be used in the proof of \cref{thm:min-multi-2} below.) If we ignore the vertex labels, we note that all of these graphs are subgraphs of the graph in \cref{fig:sharevertex}. By \cref{lem:path floor}, all of these graphs have spectator floor at least $1$. It is clear that the graph in \cref{fig:sharevertex} has spectator number $1$. Therefore, all of the graphs in \cref{fig:sharevertex,fig:shareedges,fig:one-cycle,fig:one-cycle2} have spectator floor at most $1$. \end{proof} \begin{figure}[!htbp] \centering \begin{subfigure}[b]{.47 \textwidth} \centering \begin{tikzpicture}[x=.9cm, y=1cm] \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a) at (0,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b) at (1,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (c) at (2,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (u) at (3,0) [label=below:$u$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (w) at (4,0) [label=below:$w$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (d) at (5,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (e) at (6,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (f) at (7,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (v) at (3.5,1) [label=above:$v$] {}; \path (a) edge (b) (c) edge (u) (u) edge (w) (w) edge (d) (e) edge (f) (u) edge[bend right=20] (v) (u) edge[bend left=20] (v) (w) edge[bend right=20] (v) (w) edge[bend left=20] (v); \draw (b) to (1.3,0); \draw (1.7,0) to (c); \draw [dotted,semithick] (1.4,0) -- (1.65,0); \draw (d) to (5.3,0); \draw (5.7,0) to (e); \draw [dotted,semithick] (5.4,0) -- (5.65,0); \end{tikzpicture} \caption{$G$ if two cycles share exactly one vertex} \label{fig:sharevertex} \end{subfigure} \begin{subfigure}[b]{.47\textwidth} \centering \begin{tikzpicture}[x=.9cm, y=1cm] \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a) at (0,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b) at (1,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (c) at (2,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (u) at (3,0) [label=below:$u$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (v) at (4,0) [label=below:$v$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (d) at (5,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (e) at (6,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (f) at (7,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (w) at (3.5,1) [label=above:$w$] {}; \path (a) edge (b) (c) edge (u) (u) edge (v) (w) edge (v) (v) edge (d) (e) edge (f) (u) edge[bend right=20] (w) (u) edge[bend left=20] (w); \draw (b) to (1.3,0); \draw (1.7,0) to (c); \draw [dotted,semithick] (1.4,0) -- (1.65,0); \draw (d) to (5.3,0); \draw (5.7,0) to (e); \draw [dotted,semithick] (5.4,0) -- (5.65,0); \end{tikzpicture} \caption{$G$ if it has at least two cycles} \label{fig:shareedges} \end{subfigure} \begin{subfigure}[b]{.47\textwidth} \centering \begin{tikzpicture}[x=.9cm, y=1cm] \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a) at (0,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b) at (1,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (c) at (2,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (u) at (3,0) [label=below:$u$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (d) at (4,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (e) at (5,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (f) at (6,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (v) at (3,1) [label=above:$v$] {}; \path (a) edge (b) (c) edge (u) (u) edge (d) (e) edge (f) (u) edge[bend right=20] (v) (u) edge[bend left=20] (v); \draw (b) to (1.3,0); \draw (1.7,0) to (c); \draw [dotted,semithick] (1.4,0) -- (1.65,0); \draw (d) to (4.3,0); \draw (4.7,0) to (e); \draw [dotted,semithick] (4.4,0) -- (4.65,0); \end{tikzpicture} \caption{One possibility if $G$ has exactly one cycle} \label{fig:one-cycle} \end{subfigure} \begin{subfigure}[b]{.47\textwidth} \centering \begin{tikzpicture}[x=.9cm, y=1cm] \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a) at (0,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b) at (1,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (c) at (2,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (u) at (3,0) [label=below:$u$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (v) at (4,0) [label=below:$v$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (d) at (5,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (e) at (6,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (f) at (7,0){}; \path (a) edge (b) (c) edge (u) (v) edge (d) (e) edge (f) (u) edge[bend right=20] (v) (u) edge[bend left=20] (v); \draw (b) to (1.3,0); \draw (1.7,0) to (c); \draw [dotted,semithick] (1.4,0) -- (1.65,0); \draw (d) to (5.3,0); \draw (5.7,0) to (e); \draw [dotted,semithick] (5.4,0) -- (5.65,0); \end{tikzpicture} \caption{Another possibility if $G$ has exactly one cycle} \label{fig:one-cycle2} \end{subfigure} \caption{Some graphs with spectator floor 1} \label{fig:specfloor1} \end{figure} Our first pair of results for spectator floor 2 are concerning simple graphs. The lemma below gives a list of minor-minimal simple graphs with spectator floor 2, and in the following \cref{prop:minimaml-simple} we will also show that this is the complete list of such graphs. \begin{lemma} \label{lem:simple-minimal} The following graphs are minor-minimal among simple graphs with spectator floor $2$: $K_3\sqcup K_3$, $K_3\sqcup K_{1,3}$, $K_{1,3}\sqcup K_{1,3}$, $C_4$, $K_{1,4}$, the $3$-sun (see \cref{fig:3sun}), and the long $Y$, (see \cref{fig:longY}). \end{lemma} \begin{figure}[!htbp] \begin{subfigure}[b]{.47\textwidth} \[\begin{tikzpicture}[x=1cm, y=1cm] \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a) at (0,0) [label=above:$$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b) at (1,0) [label=left:$$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (c) at (2,0) [label=below:$$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (d) at (3,0) [label=below:$$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (e) at (1.5,1) [label=above:$$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (f) at (1.5,2) [label=below:$$] {}; \path (a) edge (b) (b) edge (c) (c) edge (d) (b) edge (e) (c) edge (e) (e) edge (f); \end{tikzpicture}\] \caption{The $3$-sun} \label{fig:3sun} \end{subfigure} \begin{subfigure}[b]{.47\textwidth} \[\begin{tikzpicture}[x=1cm, y=1cm] \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a) at (0,0) [label=above:$$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b) at (0,-1) [label=left:$$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (c) at (0,-2) [label=below:$$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (d) at (-.7,.7) [label=below:$$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (e) at (-1.4,1.4) [label=above:$$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (f) at (.7,.7) [label=below:$$] {}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (g) at (1.4,1.4) [label=above:] {}; \path (a) edge (b) (b) edge (c) (a) edge (d) (d) edge (e) (a) edge (f) (f) edge (g); \end{tikzpicture}\] \caption{The long $Y$} \label{fig:longY} \end{subfigure} \caption{Two graphs from \cref{lem:simple-minimal}} \end{figure} \begin{proof} It follows from \cref{prop:components} that a disconnected graph is minor-minimal among simple graphs with spectator floor $2$ if and only if it has exactly two components, each of which have spectator floor $1$. Therefore, by \cref{cor:min-multi-1}, $K_3\sqcup K_3$, $K_3\sqcup K_{1,3}$, $K_{1,3}\sqcup K_{1,3}$ are all minor-minimal among simple graphs with spectator floor $2$. The fact that $K_{1,4}$ is minor-minimal among simple graphs with spectator floor $2$ follows directly from \cref{cor:K1k+2}. It follows from \cref{tree.uspcf,cor:diam-paths} that the long $Y$ is also minor-minimal among simple graphs with spectator floor $2$. Note that $\usp{C_4}=\diam(C_4)=2$. By the second statement of \cref{lem:diameter}, we have $\uspcf{C_4}\geq4-2=2$. Since $\uspcf{C_4}\geq\usp{C_4}=2$, we have $\uspcf{C_4}=2$. To see that $C_4$ is minor-minimal, note that deletion of any edge, results in a path, which has spectator floor $0$ and that contraction of any edge results in a triangle, which has spectator floor $1$. Let $G$ be the $3$-sun, and note that $|G|=6$ and $\diam(G)=3$. By the first statement in \cref{lem:diameter}, we have $\uspcf{G}\geq6-3-1=2$. Since $\uspcf{G}\leq\uspc{G}=2$, we have $\uspcf{G}=2$. To see that the $3$-sun is minor-minimal, first recall from \cref{lem:isolated} that isolated vertices have no effect on the spectator floor of a graph. If we delete any edge from the $3$-sun and then disregard any isolated vertices that may result, we obtain a subgraph of a graph of the form given in \cref{fig:sharevertex}. Thus, deletion of any edge of the $3$-sun results in a graph with spectator floor $1$. We now consider the effect of contracting an edge from the $3$-sun. Since we want to show that the $3$-sun is minor-minimal among simple graphs, we should immediately simplify once if any parallel edges result from the contraction. One can easily check that, if any edge is contracted and the resulting graph is simplified, then the result is a subgraph of a graph of the form given in \cref{fig:sharevertex}. Thus, the resulting graph has spectator floor $1$. \end{proof} We are now prepared to show that the list of graphs in \cref{lem:simple-minimal} is in fact the complete list of minor-minimal simple graphs with spectator floor 2. \begin{prop} \label{prop:minimaml-simple} The complete list of minor-minimal (simple) graphs with spectator floor $2$ is $K_3\sqcup K_3$, $K_3\sqcup K_{1,3}$, $K_{1,3}\sqcup K_{1,3}$, $C_4$, $K_{1,4}$, the long $Y$, and the $3$-sun. \end{prop} \begin{proof} Suppose for a contradiction that $G$ is a minor-minimal simple graph with spectator floor $2$. If $G$ is not connected, then by \cref{prop:components}, $G$ is the disjoint union of graphs $G_1$ and $G_2$, each with spectator floor $1$. The minor-minimal graphs with spectator floor $1$ are $K_3$ and $K_{1,3}$. Therefore, $G$ is $K_3\sqcup K_3$, $K_3\sqcup K_{1,3}$, or $K_{1,3}\sqcup K_{1,3}$. Thus, we may assume that $G$ is connected. Since $G$ is minor-minimal, it does not contain $C_4$ or $K_{1,4}$ as a minor. Therefore, the maximum degree of $G$ is $3$, and $G$ has no cycle of length greater than $3$. We now show that $G$ has at most one triangle. If $G$ has two disjoint triangles, then $G$ has $K_3\sqcup K_3$ as a minor. If two triangles of $G$ share exactly one vertex, then that vertex has degree $4$. If two triangles share an edge, then $G$ contains $C_4$. Therefore, $G$ has at most one triangle. If $G$ has two vertices of degree $3$ that are not contained in a triangle, then by contracting a path joining the vertices, we obtain $K_{1,4}$ as a minor. Thus, either \begin{itemize} \item[(i)] $G$ has exactly one triangle, and all vertices not in the triangle have degree $1$ or $2$, or \item[(ii)] $G$ is a tree with at most one vertex of degree $3$, and all other vertices of $G$ have degree $1$ or $2$. \end{itemize} We first consider case (i). If every vertex of the triangle has degree $3$, then $G$ has the $3$-sun as a minor. Otherwise, $G$ consists of a path with one additional vertex forming a triangle with two vertices in the path. Then $\uspc{G}=1$, and we have a contradiction. Now we consider case (ii). If $G$ has no vertex of degree $3$, then $G$ is a path, and $\uspc{G}=0$, a contradiction. Thus, we may assume that $G$ has exactly one vertex $v$ of degree $3$. If all three vertices adjacent to $v$ have degree $2$, then $G$ has the long $Y$ as a minor. Otherwise, $G$ consists of a path with one additional vertex adjacent to one vertex on the path. Then $\uspc{G}=1$, and we have a contradiction. \end{proof} Our next pair of results for spectator floor 2 are concerning the more general case of multigraphs. We first need to define the graphs $H_1$ through $H_5$. \begin{definition} \label{def:minor-minimal} Let $H_1$ be obtained from $K_3$ by doubling every edge. Let $H_2$ be obtained from $K_{1,3}$ by doubling two of the edges. Let $H_3$ be the $1$-sum of $K_3$ and $C_2$, and let $H_4$ be obtained by contracting one edge of the triangle in the $3$-sun. Let $H_5$ be the graph shown in \cref{fig:rocket}. \begin{figure}[!htbp] \[\begin{tikzpicture}[x=1cm, y=1cm] \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a) at (0,0.5){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b) at (0,1.5){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (c) at (1,1){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (d) at (2,1){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (e) at (3,1){}; \path (a) edge (c) (b) edge (c) (d) edge (c) (e) edge[bend right=20] (d) (e) edge[bend left=20] (d); \end{tikzpicture}\] \caption{$H_5$} \label{fig:rocket} \end{figure} \end{definition} The lemma below gives a list of minor-minimal (multi)graphs with spectator floor 2, and in the following \cref{thm:min-multi-2} we will also show that this is the complete list of such graphs. \begin{lemma} \label{lem:multi-minimal} The following graphs are all minor-minimal among graphs with spectator floor $2$: $C_2\sqcup C_2$, $C_2\sqcup K_{1,3}$, $K_{1,3}\sqcup K_{1,3}$, $C_4$, $H_1$, $H_2$, $H_3$, $H_4$, $H_5$, $K_{1,4}$, and the long $Y$. \end{lemma} \begin{proof} It follows from \cref{prop:components} that a disconnected graph is minor-minimal among simple graphs with spectator floor $2$ if and only if it has exactly two components, each of which have spectator floor $1$. Therefore, by \cref{cor:min-multi-1}, $C_2\sqcup C_2$, $C_2\sqcup K_{1,3}$, and $K_{1,3}\sqcup K_{1,3}$ are all minor-minimal among graphs with spectator floor $2$. Every single-edge deletion and every single-edge contraction of $C_4$, $K_{1,4}$, and the long $Y$ is a simple graph. Therefore, \cref{lem:simple-minimal} implies that these graphs are minor-minimal among graphs with spectator floor $2$. One can check that $\diam(H_1)=1$, that $\diam(H_2)=\diam(H_3)=2$, and that $\diam(H_4)=\diam(H_5)=3$. One can also check that $\usp{H_1}=1$, that $\usp{H_2}=\usp{H_3}=2$, and that $\usp{H_4}=\usp{H_5}=3$. It follows that, if $G\in\{H_1,H_2,H_3,H_4,H_5\}$, then $\uspc{G}=2$. Thus, $\uspcf{G}\leq2$. Moreover, since $\usp{G}=\diam(G)$, \cref{lem:diameter} implies that $\uspcf{G}\geq|G|-\diam(G)=2$. Therefore, $\uspcf{G}=2$. To show that $H_1$, $H_2$, $H_3$, $H_4$, and $H_5$ are minor-minimal, we must show that every single-edge deletion and every single-edge contraction from each of these graphs is a graph with spectator number less than $2$. Since every edge of $H_1$ is in parallel with another edge, no edge can be contracted. If an edge is deleted from $H_1$, then the resulting graph is a subgraph of a graph of the form given in \cref{fig:sharevertex}, which has spectator number $1$. If $G\in\{H_2,H_3,H_4,H_5\}$, then one can check that every single-edge deletion and every single-edge contraction of $G$ is a subgraph of a graph of the form given in \cref{fig:sharevertex}, which has spectator number $1$. \end{proof} Next, we have a couple of technical lemmas that will support the proof of \cref{thm:min-multi-2}. \begin{lemma} \label{lem:max2edges} Let $G$ be a graph such that two or more edges join vertices $v$ and $w$. If $e$ is one of these edges, then $\usp{G\backslash e}\geq \usp{G}$ and $\uspc{G\backslash e}\leq \uspc{G}$. Moreover, if three or more edges join vertices $v$ and $w$, then $\usp{G\backslash e}=\usp{G}$ and $\uspc{G\backslash e}=\uspc{G}$. \end{lemma} \begin{proof} No unique shortest path in $G$ contains $e$ since there is at least one other edge in parallel with $e$. Therefore, every unique shortest path in $G$ is also a unique shortest path in $G\backslash e$. The only case where a unique shortest path in $G\backslash e$ is not a unique shortest path in $G$ is if $G\backslash e$ has exactly one edge joining $v$ and $w$ and this path contains that edge. Therefore, the set of unique shortest paths of $G$ is a subset of the set of unique shortest paths of $G\backslash e$, implying that $\usp{G\backslash e}\geq \usp{G}$ and $\uspc{G\backslash e}\leq \uspc{G}$. Moreover, if three or more edges join $v$ and $w$ in $G$, then two or more edges join $v$ and $w$ in $G\backslash e$. Therefore, the set of unique shortest paths of $G$ is equal to the set of unique shortest paths of $G\backslash e$, implying that $\usp{G\backslash e}=\usp{G}$ and $\uspc{G\backslash e}=\uspc{G}$. \end{proof} \begin{lemma} \label{lem:max2edgesfloor} Let $G$ be a graph such that three or more edges join vertices $v$ and $w$. If $e$ is one of these edges, then $\uspcf{G\backslash e}=\uspcf{G}$. \end{lemma} \begin{proof} Since $G\backslash e$ is a minor of $G$, we have $\uspcf{G\backslash e}\leq\uspcf{G}$. Now, let $H$ be a graph containing $G\backslash e$ as a minor such that $\uspcf{G\backslash e}=\uspc{H}$. By \cref{lem:no decontract}, we may assume that $G\backslash e$ is obtained from $H$ by deleting edges. Therefore, there are at least two edges joining $v$ and $w$ in $H$. Add an additional edge joining $v$ and $w$ in $H$ to form the graph $H^+$, which contains $G$ as a minor. By \cref{lem:max2edges}, we have $\uspc{H}=\uspc{H^+}$. We then have $\uspcf{G\backslash e}=\uspc{H}=\uspc{H^+}\geq\uspcf{G}$ \end{proof} In our final result for this section, we are able to prove that the list of graphs in \cref{lem:multi-minimal} is in fact the complete list of minor-minimal (multi)graphs of spectator floor 2. \begin{theorem} \label{thm:min-multi-2} The complete list of minor-minimal graphs with spectator floor $2$ is $C_2\sqcup C_2$, $C_2\sqcup K_{1,3}$, $K_{1,3}\sqcup K_{1,3}$, $C_4$, $H_1$, $H_2$, $H_3$, $H_4$, $H_5$, $K_{1,4}$, and the long $Y$. \end{theorem} \begin{proof} Let $G$ be a minor-minimal graph with spectator floor $2$, and suppose for a contradiction that $G$ is not one of the graphs given in the statement of the result. By \cref{lem:max2edgesfloor}, there are at most two edges joining each pair of vertices of $G$. \begin{claim} \label{connected} $G$ is connected. \end{claim} \begin{subproof} If $G$ is not connected, then by \cref{prop:components}, $G$ is the disjoint union of graphs $G_1$ and $G_2$, each with spectator floor $1$. By \cref{cor:min-multi-1}, the minor-minimal graphs with spectator floor $1$ are $C_2$ and $K_{1,3}$. Therefore, $G$ is $C_2\sqcup C_2$, $C_2\sqcup K_{1,3}$, or $K_{1,3}\sqcup K_{1,3}$, each of which are graphs given in the statement of the result. \end{subproof} Since $G$ does not contain $C_4$ as a minor, we have the following. \begin{claim} \label{cycle} $G$ has no cycle of length at least $4$. \end{claim} Since $G$ does not contain $K_{1,4}$ as a minor, we have the following. \begin{claim} \label{neighborhood} $G$ has no vertex with a neighborhood of cardinality at least $4$. \end{claim} \begin{claim} \label{vertex} $G$ has no pair of disjoint cycles and no pair of cycles that share exactly one vertex. \end{claim} \begin{subproof} Since $C_2\sqcup C_2$ is not a minor of $G$, there is no pair of disjoint cycles in $G$. Since $H_3$ is not a minor of $G$, if two cycles share exactly one vertex, both cycles must have length $2$. Now, suppose for a contradiction that $G$ contains two copies of $C_2$ as subgraphs and that these copies share exactly one vertex $v$. Let $u$ and $w$ be the other vertices in these copies of $C_2$. Because $H_2$ is not a minor of $G$, we have $N_G(v)=\{u,w\}$. Because $H_4$ is not a minor of $G$, we have $|N_G(u)-\{v,w\}|\leq1$ and $|N_G(w)-\{u,v\}|\leq1$. Moreover, since $C_4$ and $H_1$ are not minors of $G$, there is no path from $u$ to $w$ in $G-v$ except possibly at most one edge joining $u$ and $w$. Finally, because $H_5$ is not a minor of $G$, no vertex in $V(G)-\{u,v,w\}$ has a neighborhood of cardinality greater than $2$. Therefore, $G$ is a subgraph of the graph in \cref{fig:sharevertex}. This graph has spectator number $1$, Thus $\uspcf{G}=1$, a contradiction. \end{subproof} \begin{claim} \label{triangles} No pair of triangles of $G$ share exactly one edge. \end{claim} \begin{subproof} Otherwise, the union of these triangles contains a cycle of length $4$, violating \cref{cycle}. \end{subproof} \begin{claim} \label{onecycle} $G$ has at most one cycle. \end{claim} \begin{subproof} Suppose for a contradiction that $G$ has more than one cycle. We know that each pair of vertices of $G$ is joined by at most two edges. Therefore, \cref{cycle,vertex,triangles} imply that $G$ contains a triangle with vertices $u$, $v$, and $w$ with a second edge joining $u$ and $w$. By \cref{vertex}, there is no path in $G-w$ from $u$ to $v$ other than the edge $uv$. Similarly, there is no path in $G-u$ from $w$ to $v$ other than the edge $wv$. We can also see that there is no path from $u$ to $w$ in $G-v$ other than the two edges joining $u$ and $w$. This is because $G$ has no cycle of length at least $4$ and at most two edges joining $u$ and $w$. Therefore, there are subgraphs $G_u$, $G_v$, and $G_w$ of $G$ such that, for each vertex $x$ in $G_i$, the path from $x$ to $\{u,v,w\}$ has endpoints $x$ and $i$. (For each $i\in\{u,v,w\}$, we have $i\in V(G_i)$.) This path must be unique; otherwise, \cref{vertex} is violated. Thus, each of $G_u$, $G_v$, and $G_w$ is a tree. Moreover, if we denote by $F$ the set of four edges both of whose endpoints are in $\{u,v,w\}$, then $G$ is the graph whose vertex set is $V(G_u)\sqcup V(G_v)\sqcup V(G_w)$ and whose edge set is $E(G_u)\sqcup E(G_v)\sqcup E(G_w)\sqcup F$. For $i\in\{u,v,w\}$, vertex $i$ has at most one neighbor in $G_i$. Otherwise, $G$ has $K_{1,4}$ as a minor. Similarly, since $G$ does not have $K_{1,4}$ as a minor, no vertex in $V(G_i)-\{i\}$ has more than two neighbors. If $u$ and $w$ have neighbors in $G_u$ and $G_w$, respectively, then $G$ contains $H_4$ as a minor. Therefore, either $N_G(u)=\{v,w\}$ or $N_G(w)=\{u,v\}$. Without loss of generality, let $N_G(w)=\{u,v\}$. Then $G$ is a subgraph of a graph of the form given in \cref{fig:shareedges}. This graph has spectator number $1$, Thus $\uspcf{G}=1$, a contradiction. \end{subproof} Therefore, one of the following holds. \begin{itemize} \item[(i)] $G$ is a tree, \item[(ii)] $G$ has exactly one cycle, which is a triangle, \item[(iii)] $G$ has exactly one cycle, which is a $C_2$, or \end{itemize} In cases (i) and (ii), $G$ is a simple graph. Since $G$ is not any of the graphs listed above, and since $G$ is connected, \cref{prop:minimaml-simple} implies that $G$ is the $3$-sun. However, by contracting an edge of the triangle in the $3$-sun, we obtain $H_4$. Therefore, $G$ has exactly one cycle, which is a $C_2$. Let $u$ and $v$ be the vertices of $C_2$. Because $H_5$ is not a minor of $G$, every vertex in $V(G)-\{u,v\}$ has degree $1$ or $2$. Since $K_{1,4}$ is not a minor of $G$, we have $|N_G(u)-\{v\}|\leq2$ and $|N_G(v)-\{u\}|\leq2$. Moreover, either $|N_G(u)-\{v\}|<2$ or $|N_G(v)-\{u\}|<2$. Without loss of generality, let $|N_G(v)-\{u\}|<2$. If $|N_G(v)-\{u\}|=0$, then $G$ is a subgraph of a graph of the form shown in \cref{fig:one-cycle}. Thus, we have have $\uspc{G}=\uspcf{G}=1$, a contradiction. If $|N_G(v)-\{u\}|=1$, then since $H_4$ is not a subgraph of $G$, we must also have $|N_G(u)-\{v\}|=1$. Therefore, $G$ is a subgraph of a graph of the form given in \cref{fig:one-cycle2}, which is isomorphic to a subgraph of a graph of the form shown in \cref{fig:shareedges}. The graph in \cref{fig:shareedges} has spectator number $1$. Therefore, $\uspcf{G}=1$, a contradiction. Therefore, by contradiction, we conclude that the result holds. \end{proof} \section{Minor Maximal Graphs} \label{minor max graphs} To this point in the paper, we have discussed minor minimal graphs at some length; in this sense, we have only looked downwards. We shall now look upwards and consider the possibilities of minor maximal graphs of a given spectator floor. We begin with the observation that, without additional restrictions, there are no minor maximal graphs with a given spectator floor. This is so because, given any graph $G$, if we add a new isolated vertex to $G$ in order to obtain $G'$, then $G$ is a minor of $G'$, and by \cref{prop:components}, $G'$ has the same spectator floor as $G$. Hence, we can construct an infinite chain of graphs which are above $G$ and have the same spectator floor. Therefore, in order to obtain any meaningful information, we must search for minor maximal graphs amongst subsets of graphs which are restricted in some way. The most natural restriction to make is on the number of vertices and on the number of parallel edges allowed, for which we can completely characterize the minor maximal graphs. In order to present the result, we first must present new definitions. \begin{definition}\label{def:crowded_parade} If $p\geq2$, then a \emph{crowded $p$-parade} is a graph $G$ with a parade of $p$ vertices that has the following properties: \begin{itemize} \item Every vertex outside the parade is connected to exactly two vertices in the parade, and those two parade vertices are adjacent to each other. \item Any two vertices outside the parade that are adjacent to a common parade vertex are also adjacent to each other. \end{itemize} A \emph{crowded $1$-parade} is a graph whose simplification is a complete graph. \end{definition} \begin{definition}\label{def:sat_crowded_parade} If $p\geq2$, an \emph{$m$-saturated crowded $p$-parade} is a crowded $p$-parade where every edge which is not in the parade has $m$ parallel copies, including itself. An \emph{$m$-saturated crowded $1$-parade} is a graph such that every pair of vertices is joined by exactly $m$ edges. \end{definition} \begin{figure}[!htbp] \[\begin{tikzpicture}[x=1cm, y=1cm] \node[vertex][fill] (a) at (0,0){}; \node[vertex][fill] (b) at (2,0){}; \node[vertex][fill] (c) at (4,0){}; \node[vertex][fill] (d) at (6,0){}; \node[vertex][fill] (e) at (8,0){}; \node[vertex][fill] (f) at (10,0){}; \node[vertex][fill] (g) at (12,0){}; \node[vertex][fill] (bc) at (3,1.5){}; \node[vertex][fill] (cd1) at (5,2){}; \node[vertex][fill] (cd2) at (5,1){}; \node[vertex][fill] (de) at (7,1.5){}; \node[vertex][fill] (fg1) at (11,2){}; \node[vertex][fill] (fg2) at (10.4,1.5){}; \node[vertex][fill] (fg3) at (11.6,1.5){}; \draw (a) to (g); \draw[double distance=2] (b) to (bc) to (c); \draw[double distance=2] (cd1) to (c) to (cd2) to (d) to (cd1); \draw[double distance=2] (cd1) to (cd2); \draw[double distance=2] (d) to (de) to (e); \draw[double distance=2] (bc) to (cd1) to (de) to (cd2) to (bc); \draw[double distance=2] (f) to (fg1) to (g); \draw[double distance=2] (f) to (fg2) to (g); \draw[double distance=2] (f) to (fg3) to (g); \draw[double distance=2] (fg1) to (fg2) to (fg3) to (fg1); \end{tikzpicture}\] \caption{An example of a 2-saturated crowded 7-parade.} \label{fig:crowded_parade} \end{figure} With these definitions in hand, we will show that a graph is minor maximal amongst graphs of a given spectator floor value if and only if the graph is a crowded parade. Before presenting this characterization of minor maximal graphs, however, we first have two technical lemmas that we will use repeatedly in the proof of the characterization. In the first lemma, we use the notation $d_S(x,y)$ to refer to the distance between any vertices $x$ and $y$ within any graph $S$. \begin{lemma}\label{lem:maximals} Let $G$ be a graph that is minor maximal amongst graphs with $n$ vertices, at most $m$ parallel edges between any given vertices, and spectator floor $k$. Let $a,b$ be the endpoints of a parade in $G$. Let $H$ be $G$ with one edge added between some vertices $v$ and $w$ where no edge was present in $G$. Then at least one of the following is true: \[d_H(a,b)=d_G(a,v)+1+d_G(w,b)\] \[d_H(a,b)=d_G(a,w)+1+d_G(v,b)\] \[\] \end{lemma} \begin{proof} By \cref{lem:no decontract}, there is a graph $G'$ with the same number of vertices as $G$ such that $G$ is a subgraph of $G'$ and such that $\uspcf{G}=\uspc{G'}$. Further, we may also assume that $G'$ has at most $m$ parallel edges between any given vertices. (It follows from \cref{lem:max2edges} that adding an edge in parallel with an edge in $G$ can only increase the spectator number.) Since we assumed that $G$ is maximal amongst such graphs, we conclude that $G'=G$. Hence $\uspc{G}=k$. Now consider the graph $H$, which is the same as $G$ but with an edge added between $v$ and $w$, which are some vertices of $G$ that are not adjacent in $G$. Since $G$ is a proper subgraph of $H$, and $H$ is a graph with $n$ vertices and at most $m$ parallel edges between any given vertices, and $G$ is maximal amongst graphs with $n$ vertices, at most $m$ parallel edges between any given vertices, and spectator floor $k$, we conclude that $H$ does not have spectator floor $k$. Furthermore, since $G$ is a subgraph of $H$, $\uspcf{H}>\uspcf{G}$. Thus, $\uspc{H}>\uspc{G}$. Since $\uspc{G}=k$, $G$ has a parade $P$ of $n-k$ vertices. Let us call the endpoints of $P$ by the names $a$ and $b$. Since $H$ still contains $P$ but $\uspc{H}\neq k$, we conclude that $P$ is no longer a longest unique shortest path in $H$. There cannot be a longer unique shortest path than $P$ in $H$, however, since that would result in $\uspc{H}<\uspc{G}$, so it must be that $P$ is not a unique shortest path in $H$. There are two possibilities: either $P$ is still a shortest path in $H$ but no longer unique, or $P$ is no longer a shortest path in $H$. Either way, there must be a new shortest path in $H$ between $a$ and $b$ that is of the same or shorter length than $P$; let us call this new path $Q$. $Q$ must contain the edge $\{v,w\}$, since that edge is the only difference between $G$ and $H$. Then, there are two possibilities. If $Q$ connects $a,v,w,b$ in that order, then we have: \[d_H(a,b)=d_G(a,v)+1+d_G(w,b)\] Otherwise, if $Q$ connects $a,w,v,b$ in that order, then we have: \[d_H(a,b)=d_G(a,w)+1+d_G(v,b)\] \end{proof} We now present another technical lemma to be used in proving the characterization of minor maximal graphs. This lemma tells us that minor maximal graphs must be connected. \begin{lemma}\label{lem:maxls_are_connected} Let $G$ be a graph that is minor maximal amongst graphs with $n$ vertices, at most $m$ parallel edges between any given vertices, and spectator floor $k$. Then $G$ is connected. \end{lemma} \begin{proof} Suppose to the contrary that $G$ is not connected, but rather consists of connected components $G_1,G_2,...,G_r$. We know from the proof of \cref{lem:maximals} that $\uspc{G}=\uspcf{G}=k$, and from \cref{prop:components} we know that \[\uspcf{G}=\uspcf{G_1}+\uspcf{G_2}+\cdots+\uspcf{G_r}\] For each $i=1,2,...,r$, let $H_i$ be a supergraph of $G_i$ that realizes $\uspcf{G_i}$ without adding any additional vertices; that is, $\uspc{H_i}=\uspcf{G_i}$. We know such $H_i$ exist from \cref{{lem:no decontract}}. Then let $H$ be a supergraph of $H_1 \cup H_2 \cup \cdots \cup H_r$ in which we add $r-1$ edges to the graph in order to take one parade from each of the $H_i$ and connect them into one long parade. Then we have that $H$ is a supergraph of $G$ and $H$ still has $n$ vertices. Furthermore, due to the way we constructed $H$, we have \begin{align*} \uspc{H} &=\uspc{H_1}+\uspc{H_2}+\cdots+\uspc{H_r} \\ & =\uspcf{G_1}+\uspcf{G_2}+\cdots+\uspcf{G_r} \\ & =\uspcf{G}=\uspc{G} \end{align*} Since $H$ is a supergraph of $G$, it follows that $\uspcf{H}\geq \uspcf{G}=\uspc{G}$. Furthermore, since $\uspc{H}=\uspc{G}$, it follows that $\uspcf{H}\leq \uspc{G}$. Hence $\uspcf{H}=\uspc{G}=k$. Thus, since $H$ has $n$ vertices and spectator floor $k$, and is a supergraph of $G$, and since we assumed $G$ was maximal amongst such graphs, it must be that $G=H$. Since we assumed that $G$ was disconnected and $H$ is connected by construction, this is a contradiction. \end{proof} Before presenting the main results of this section, we prove a lemma that takes care of a special case. \begin{lemma} \label{lem:1-parade} Let $m\geq2$. Then $G$ is minor maximal amongst graphs with $n$ vertices, at most $m$ parallel edges between any given pair of vertices, and spectator floor $n-1$, if and only if $G$ is an $m$-saturated crowded $1$-parade. \end{lemma} \begin{proof} Note that there is exactly one $m$-saturated crowded $1$-parade $H$ with $n$ vertices. Since $m\neq1$, no edge in $H$ is a unique shortest path. This implies that the unique shortest paths of $H$ each contain only one vertex. Therefore, $\uspc{H}=1$, implying that $\uspcf{H}=1$. Every graph containing $H$ as a minor either has more than $n$ vertices or a pair of vertices with more than $m$ edges joining them. Therefore, $H$ is maximal. Conversely, note that every graph $G$ with $n$ vertices and at most $m$ parallel edges between any given pair of vertices is a minor of $H$. Therefore, the only minor maximal graph amongst graphs with $n$ vertices, at most $m$ parallel edges between any given pair of vertices, and spectator floor $n-1$ is $H$. \end{proof} We now present the first of the two main results of this section, which together provide a complete characterization of the minor maximal graphs with a given spectator floor value. Note that complete graphs are both $1$-saturated $1$-parades and $1$-saturated $2$-parades. This gives some intuition for the reason the first sentence of the theorem is needed. \begin{theorem} Let $k\leq n-2$ and $m\geq1$, or let $k=n-1$ and $m\geq2$. If $G$ is minor maximal amongst graphs with $n$ vertices, at most $m$ parallel edges between any given vertices, and spectator floor $k$, then $G$ is an $m$-saturated crowded $(n-k)$-parade. \end{theorem} \begin{proof} By \cref{lem:1-parade}, the result holds when $k=n-1$. Therefore, we may assume that $k\leq n-2$. Suppose that $G$ is minor maximal amongst graphs with $n$ vertices, at most $m$ parallel edges between any given vertices, and spectator floor $k$. We will show that $G$ is an $m$-saturated crowded $(n-k)$-parade. As was shown in the proof of \cref{lem:maximals}, $\uspc{G}=k$ and so $G$ has a parade $P$ of $n-k\geq2$ vertices. We will call the endpoints of $P$ by the names $a$ and $b$. We will now show that $P$ has the properties in the definition of a crowded $(n-k)$-parade, \cref{def:crowded_parade}. Let $v$ be a vertex outside the parade $P$. We will consider cases based on the number of vertices in $P$ that are adjacent to $v$. Suppose $v$ is adjacent to 3 or more vertices in $P$. Then two of the parade vertices that $v$ is adjacent to have a distance of 2 or more within the parade. Hence going through $v$ would provide a path of the same or lesser length between those two vertices as compared to the parade. This contradicts the properties of a parade. So $v$ cannot be adjacent to 3 or more vertices in the parade. Now suppose that $v$ is adjacent to 2 vertices in $P$, and those two vertices have a distance of 2 or more along the parade. The same argument from the last paragraph applies; this is a contradiction. Hence, if $v$ is adjacent to exactly 2 vertices in the parade, then those two vertices must also be adjacent to each other. Next, suppose that $v$ is adjacent to exactly 1 vertex in $P$. There are two cases: either $v$ is adjacent to an endpoint of $P$, or not. Suppose $v$ is adjacent to an endpoint of the parade; without loss of generality, let us assume $v$ is adjacent to $a$. Let $x$ be the vertex of $P$ that is adjacent to $a$, and let $H$ be the graph made from $G$ by adding edge $\{v,x\}$, as shown in \cref{fig:maxl_pf_1}. \begin{figure}[!htbp] \begin{subfigure}[b]{.47\textwidth} \[\begin{tikzpicture}[x=1cm, y=1cm] \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a) at (0,0)[label=below:$a$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (x) at (1,0)[label=below:$x$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (y1) at (2,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (y2) at (3,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b) at (4,0)[label=below:$b$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (v) at (0,1)[label=left:$v$]{}; \draw (a) to (y1); \draw (y1) to (2.3,0); \draw[dotted,semithick] (2.4,0) to (2.6,0); \draw(2.7,0) to (y2); \draw (y2) to (b); \draw (a) to (v); \end{tikzpicture}\] \caption{The subgraph of $G$ induced by $P\cup\{v\}$.} \end{subfigure}\begin{subfigure}[b]{.47\textwidth} \[\begin{tikzpicture}[x=1cm, y=1cm] \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a) at (0,0)[label=below:$a$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (x) at (1,0)[label=below:$x$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (y1) at (2,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (y2) at (3,0){}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b) at (4,0)[label=below:$b$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (v) at (0,1)[label=left:$v$]{}; \draw (a) to (y1); \draw (y1) to (2.3,0); \draw[dotted,semithick] (2.4,0) to (2.6,0); \draw(2.7,0) to (y2); \draw (y2) to (b); \draw (a) to (v); \draw (v) to (x); \end{tikzpicture}\] \caption{The subgraph of $H$ induced by $P\cup\{v\}$.} \end{subfigure} \caption{}\label{fig:maxl_pf_1} \end{figure} Then by \cref{lem:maximals}, one of the following is true: \[d_H(a,b)=d_G(a,v)+1+d_G(x,b)=1+d_G(a,x)+d_G(x,b) \geq 1+d_G(a,b)\] \[d_H(a,b)=d_G(a,x)+1+d_G(v,b)=1+d_G(a,v)+d_G(v,b)\geq 1+d_G(a,b)\] However, since $G$ is a subgraph of $H$, we must have $d_H(a,b)\leq d_G(a,b)$. This is a contradiction. Now we consider the situation where $v$ is adjacent to exactly one vertex of $P$, and that vertex is not an endpoint of $P$. Let $w$ be the vertex of $P$ adjacent to $v$, and let $w$ and $x$ be the vertices of $P$ adjacent to $w$, where $x$ is closer to $a$ than $y$ is. Now consider the graph $X$, which is the same as $G$ but with an edge from $v$ to $x$ added, and the graph $Y$, which is the same as $G$ but with an edge from $v$ to $y$ added. $G$, $X$, and $Y$ are as shown in \cref{fig:maxl_pf_2}. \begin{figure}[!htbp] \centering \begin{subfigure}[b]{.3\textwidth} \[\begin{tikzpicture}[x=0.8cm, y=1cm] \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a) at (0,0)[label=below:$a$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (x) at (1,0)[label=below:$x$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (w) at (2,0)[label=below:$w$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (y) at (3,0)[label=below:$y$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b) at (4,0)[label=below:$b$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (v) at (2,1)[label=left:$v$]{}; \draw (a) to (0.3,0); \draw [dotted,semithick] (0.4,0) to (0.6,0); \draw (0.7,0) to (x); \draw (x) to (y); \draw (y) to (3.3,0); \draw[dotted,semithick] (3.4,0) to (3.6,0); \draw (3.7,0) to (b); \draw (v) to (w); \end{tikzpicture}\] \caption{The subgraph of $G$ induced by $P\cup\{v\}$.} \end{subfigure}\;\;\;\;\begin{subfigure}[b]{.3\textwidth} \[\begin{tikzpicture}[x=0.8cm, y=1cm] \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a) at (0,0)[label=below:$a$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (x) at (1,0)[label=below:$x$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (w) at (2,0)[label=below:$w$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (y) at (3,0)[label=below:$y$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b) at (4,0)[label=below:$b$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (v) at (2,1)[label=left:$v$]{}; \draw (a) to (0.3,0); \draw [dotted,semithick] (0.4,0) to (0.6,0); \draw (0.7,0) to (x); \draw (x) to (y); \draw (y) to (3.3,0); \draw[dotted,semithick] (3.4,0) to (3.6,0); \draw (3.7,0) to (b); \draw (v) to (w); \draw (v) to (x); \end{tikzpicture}\] \caption{The subgraph of $X$ induced by $P\cup\{v\}$.} \end{subfigure}\;\;\;\;\begin{subfigure}[b]{.3\textwidth} \[\begin{tikzpicture}[x=0.8cm, y=1cm] \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a) at (0,0)[label=below:$a$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (x) at (1,0)[label=below:$x$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (w) at (2,0)[label=below:$w$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (y) at (3,0)[label=below:$y$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b) at (4,0)[label=below:$b$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (v) at (2,1)[label=left:$v$]{}; \draw (a) to (0.3,0); \draw [dotted,semithick] (0.4,0) to (0.6,0); \draw (0.7,0) to (x); \draw (x) to (y); \draw (y) to (3.3,0); \draw[dotted,semithick] (3.4,0) to (3.6,0); \draw (3.7,0) to (b); \draw (v) to (w); \draw (v) to (y); \end{tikzpicture}\] \caption{The subgraph of $Y$ induced by $P\cup\{v\}$.} \end{subfigure} \caption{}\label{fig:maxl_pf_2} \end{figure} For $X$, \cref{lem:maximals} tells us that one of \cref{eqn:QX1,eqn:QX2}, below, is true. \begin{align}\label{eqn:QX1} \begin{split} d_X(a,b)&=d_G(a,v)+1+d_G(x,b)\\ &=d_G(a,v)+d_G(v,w)+d_G(x,b)\\ &\geq d_G(a,w)+d_G(x,b)\\ &=d_G(a,b)+1 \end{split} \end{align} Note that the second strict inequality in \cref{eqn:QX2} is because an off-parade path between two parade vertices is longer than the parade route. \begin{align}\label{eqn:QX2} \begin{split} d_X(a,b)&=d_G(a,x)+1+d_G(v,b)\\ &=d_G(a,x)+1+d_G(x,v)-d_G(x,v)+d_G(v,b)\\ &>d_G(a,x)+d_G(x,v)+d_G(v,b)-1\\ &> d_G(a,b)-1 \end{split} \end{align} Since $G$ is a subgraph of $X$, we must have $d_X(a,b)\leq d_G(a,b)$, therefore \cref{eqn:QX1} is a contradiction, and so \cref{eqn:QX2} is true and implies $d_X(a,b)=d_G(a,b)$. For $Y$, \cref{lem:maximals} tells us that one of the following is true: \begin{equation}\label{eqn:QY1} d_Y(a,b)=d_G(a,v)+1+d_G(y,b) \end{equation} \begin{equation}\label{eqn:QY2} d_Y(a,b)=d_G(a,y)+1+d_G(v,b) \end{equation} However, \cref{eqn:QY2} leads to a contradiction by the same logic as for \cref{eqn:QX1}. Hence \cref{eqn:QY1} is true and by the same logic as for \cref{eqn:QX2} it implies $d_Y(a,b)=d_G(a,b)$. When we replace the left-hand sides of \cref{eqn:QX2,eqn:QY1} with $d_G(a,b)$ and add them together, we get the following. Note that use of strict inequality is because an off-parade path between two parade vertices is longer than the parade route. \begin{align}\label{eqn:QXQY} \begin{split} 2d_G(a,b)&=d_G(a,v)+d_G(v,b)+d_G(a,x)+d_G(y,b)+2\\ &>d_G(a,b)+d_G(a,x)+d_G(y,b)+2\implies\\ d_G(a,b)&>d_G(a,x)+d_G(y,b)+2\\ &=d_G(a,b) \end{split} \end{align} which is a contradiction. Thus we conclude that in $G$, a vertex $v$ that is not in $P$ cannot be adjacent to exactly 1 vertex in $P$. For the last case, let us consider vertices of $G$ that are not in $P$ and are not adjacent to any vertices in $P$. We will show these cannot exist. First note that $G$ must be connected because of \cref{lem:maxls_are_connected}. Let $d_G(v,P)$ denote the distance from any vertex $v$ to the path $P$; in other words, \[d_G(v,P):=\min_{p\in P}\{d_G(v,p)\}\] Since $G$ is connected, $d_G(v,P)$ is finite for all $v\in G$. For any vertex $v$ that is not in $P$ and not adjacent to $P$, either $d_G(v,P)=2$ or $d_G(v,P)>2$. If $d_G(v,P)>2$, then there is a path of length $d_G(v,P)$ from $v$ to a vertex $p\in P$, and so there is some vertex $v'$ along that path that has $d_G(v',P)=2$. Hence, if there are any vertices of $G$ that are not in $P$ and are not adjacent to $P$, then there is some vertex $v$ with $d_G(v,P)=2$. We will show this leads to a contradiction. Since $d_G(v,P)=2$, there is some vertex, call it $w$, that is adjacent to $v$ and adjacent to $P$. We have already shown in this proof that if $w$ is adjacent to any vertex of $P$, then it is adjacent to exactly two vertices of $P$, which are also adjacent to each other. Let us call the two vertices of $P$ that $w$ is adjacent to by the names $x$ and $y$, and the endpoints of $P$ by the names $a$ and $b$. As before, we shall assume the $a$ is the endpoint that is closer to $x$. Now consider the graph $X$, which is the same as $G$ but with an edge from $v$ to $x$ added, and the graph $Y$, which is the same as $G$ but with an edge from $v$ to $y$ added. The situation is shown in \cref{fig:maxl_pf_3}. \begin{figure}[!htbp] \centering \begin{subfigure}[b]{.3\textwidth} \[\begin{tikzpicture}[x=1cm, y=1cm] \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a) at (0,0)[label=below:$a$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (x) at (0.85,0)[label=below:$x$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (w) at (1.5,1)[label=right:$w$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (y) at (2.15,0)[label=below:$y$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b) at (3,0)[label=below:$b$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (v) at (1.5,2)[label=right:$v$]{}; \draw (a) to (0.3,0); \draw [dotted,semithick] (.33,0) to (.52,0); \draw (.55,0) to (x); \draw (y) to (2.45,0); \draw[dotted,semithick] (2.48,0) to (2.67,0); \draw (2.7,0) to (b); \draw (v) to (w) to (x) to (y) to (w); \end{tikzpicture}\] \caption{The subgraph of $G$ induced by $P\cup\{v,w\}$.} \end{subfigure}\;\;\;\;\begin{subfigure}[b]{.3\textwidth} \[\begin{tikzpicture}[x=1cm, y=1cm] \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a) at (0,0)[label=below:$a$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (x) at (0.85,0)[label=below:$x$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (w) at (1.5,1)[label=right:$w$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (y) at (2.15,0)[label=below:$y$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b) at (3,0)[label=below:$b$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (v) at (1.5,2)[label=right:$v$]{}; \draw (a) to (0.3,0); \draw [dotted,semithick] (.33,0) to (.52,0); \draw (.55,0) to (x); \draw (y) to (2.45,0); \draw[dotted,semithick] (2.48,0) to (2.67,0); \draw (2.7,0) to (b); \draw (v) to (w) to (x) to (y) to (w); \draw (v) to (x); \end{tikzpicture}\] \caption{The subgraph of $X$ induced by $P\cup\{v,w\}$.} \end{subfigure}\;\;\;\;\begin{subfigure}[b]{.3\textwidth} \[\begin{tikzpicture}[x=1cm, y=1cm] \node[vertex][fill,inner sep=1pt,minimum size=1pt] (a) at (0,0)[label=below:$a$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (x) at (0.85,0)[label=below:$x$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (w) at (1.5,1)[label=left:$w$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (y) at (2.15,0)[label=below:$y$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (b) at (3,0)[label=below:$b$]{}; \node[vertex][fill,inner sep=1pt,minimum size=1pt] (v) at (1.5,2)[label=left:$v$]{}; \draw (a) to (0.3,0); \draw [dotted,semithick] (.33,0) to (.52,0); \draw (.55,0) to (x); \draw (y) to (2.45,0); \draw[dotted,semithick] (2.48,0) to (2.67,0); \draw (2.7,0) to (b); \draw (v) to (w) to (x) to (y) to (w); \draw (v) to (y); \end{tikzpicture}\] \caption{The subgraph of $Y$ induced by $P\cup\{v,w\}$.} \end{subfigure} \caption{}\label{fig:maxl_pf_3} \end{figure} These graphs are not the same as the previous $X$ and $Y$, but all of the arguments from before regarding \cref{eqn:QX1,eqn:QX2,eqn:QY1,eqn:QY2} still apply. Most of \cref{eqn:QXQY} applies as well, with the exception of the last line, where we used the fact that $d_G(a,x)+d_G(y,b)+2=d_G(a,b)$, which is no longer true. Now we have $d_G(a,x)+d_G(y,b)+2=d_G(a,b)+1$, which still leads to a contradiction when applied to the penultimate line of \cref{eqn:QXQY}. Thus we conclude that there are no vertices of $G$ that are not in $P$ and not adjacent to $P$. Every vertex of $G$ which is not in $P$ must be adjacent to two adjacent vertices in $P$. This fulfills the first condition for $G$ to be a crowded $(n-k)$-parade. We will now show that the second condition is also true. Let $v$ and $v'$ be two vertices outside $P$ that are adjacent to a common vertex $w$ in $P$, and suppose $v,v'$ are not adjacent to each other. Let $F$ be the graph made from $G$ by adding an edge between $v$ and $v'$. Now we can assume without loss of generality that the first equation from \cref{lem:maximals} is true (if the second equation is the true one, just swap the names of $a$ and $b$). Then we have the following. Note that to get from line 3 to line 4 of the next equation, we are applying the fact that an off-parade path between two parade vertices is strictly longer than the parade route. \begin{align*}\label{eqn:Fdist} d_F(a,b)&=d_G(a,v)+1+d_G(v',b)\\ &=d_G(a,v)+d_G(v,w)-d_G(v,w)+1-d_G(w,v')+d_G(w,v')+d_G(v',b)\\ &=d_G(a,v)+d_G(v,w)-1+d_G(w,v')+d_G(v',b)\\ &\geq d_G(a,w)+d_G(w,b)+1\\ &=d_G(a,b)+1 \end{align*} However, $G$ is a subgraph of $F$, so we also have $d_F(a,b)\leq d_G(a,b)$. This is a contradiction. Hence, $G$ is a crowded $(n-k)$-parade. We will now argue that $G$ is $m$-saturated. For any edge $e$ of $G$ that is not in $P$, if there are fewer than $m$ parallel copies of $e$, then we can add another parallel copy of $e$ without changing any distances in $G$, and without creating any new paths between the endpoints of $P$ that have the same length as $P$. Hence, adding such an edge does not change the spectator floor value of the graph. Given that $G$ is maximal amongst graphs with $n$ vertices, at most $m$ parallel copies of each edge, and spectator floor $k$, then, $G$ must already contain all such edges. Hence $G$ is an $m$-saturated crowded $(n-k)$-parade.\end{proof} We will now present the second of the two main results in this section, which is the converse of the last theorem, showing that we have a complete characterization of the minor maximal graphs. \begin{theorem}\label{thm:maxl_2} Let $k\leq n-2$ and $m\geq1$, or let $k=n-1$ and $m\geq2$. If $G$ is an $m$-saturated crowded $(n-k)$-parade, then $G$ is minor maximal amongst graphs with $n$ vertices, at most $m$ parallel edges between any given vertices, and spectator floor $k$. \end{theorem} \begin{proof} By \cref{lem:1-parade}, the result holds when $k=n-1$. Therefore, we may assume that $k\leq n-2$. Suppose that $G$ is an $m$-saturated crowded $(n-k)$-parade, but $G$ is not minor maximal amongst graphs with $n$ vertices, at most $m$ parallel edges between any given vertices, and spectator floor $k$. Then there is some other graph, call it $G'$, which is minor maximal on that set and of which $G$ is a proper subgraph. Since $G'$ is minor maximal, it must be an $m$-saturated $(n-k)$-parade as well. Let us call the parade for which $G$ fulfills the definition of a crowded parade, \cref{def:crowded_parade}, by the name $P$. First note that if $m>1$, then there are no unique paths in $G'$ except $P$ and its subpaths, hence for $m>1$, $P$ is still the parade for which $G'$ is a crowded parade. Now suppose $m=1$ (in other words, we are working with simple graphs only) and suppose $P$ is either not a parade in $G'$ or not a parade for which $G'$ is crowded. Let us call the parade for which $G'$ is a crowded parade by the name $P'$, and the endpoints of $P$ by the names $a,b$ and of $P'$ by the names $a',b'$. Since $P\neq P'$, at least one of $a',b'$ is not in $P$. Since $G$ is a subgraph of $G'$, $d_{G'}(v,w)\leq d_G(v,w)$ for all pairs of vertices $v,w$. So $d_G(a',b')\geq d_{G'}(a',b')=d_G(a,b)=n-k-1$. A generalized figure of graph $G$ is shown in \cref{fig:gen_crowded_parade}, where path $P$ consists of the vertices labeled $v_1=a$ to $v_{n-k}=b$. Nodes labeled $K_{m_i}$ represent complete subgraphs on $m_i$ vertices, and bold edges represent a complete set of edges between the connected structures. It should be understood, however, that it is possible for some $m_i$ to be 0, in which case $K_{m_i}$ is an empty graph, and there are no edges connecting $K_{m_{i-1}}$, $K_{m_{i}}$, and $K_{m_{i+1}}$. For example, in the previous \cref{fig:crowded_parade}, we had $\{m_1,m_2,m_3,m_4,m_5,m_6\}=\{0,1,2,1,0,3\}$. \begin{figure}[!htbp] \[\begin{tikzpicture}[x=1cm, y=1cm] \node[vertex][fill] (a) at (0,0)[label=below:${v_1=a}$]{}; \node[vertex][fill] (b) at (2,0)[label=below:$v_2$]{}; \node[vertex][fill] (c) at (4,0)[label=below:$v_3$]{}; \node (d) at (6,0){}; \node[vertex][fill] (e) at (8,0)[label=below:$v_{n-k-2}$]{}; \node[vertex][fill] (f) at (10,0)[label=below:$v_{n-k-1}$]{}; \node[vertex][fill] (g) at (12,0)[label=below:${v_{n-k}=b}$]{}; \node[vertex][] (ab) at (1,1.5){$K_{m_1}$}; \node[vertex][] (bc) at (3,1.5){$K_{m_2}$}; \node[vertex][] (cd) at (5,1.5){$K_{m_3}$}; \node[vertex][] (ef) at (9,1.5){$K_{m_{n-k-2}}$}; \node[vertex][] (fg) at (11,1.5){$K_{m_{n-k-1}}$}; \draw (a) to (c); \draw(c) to (5,0); \draw[dotted,semithick] (5.5,0) to (6.5,0); \draw (7,0) to (e); \draw (e) to (g); \draw[very thick] (ab) to (bc) to (cd); \draw[very thick](cd) to (6,1.5); \draw[very thick, dotted] (6.5,1.5) to (7.5,1.5); \draw[very thick] (7.75,1.5) to (ef); \draw[very thick] (ef) to (fg); \draw[very thick] (a) to (ab) to (b) to (bc) to (c) to (cd); \draw[very thick] (e) to (ef) to (f) to (fg) to (g); \draw[very thick] (cd) to (5.4,.9); \draw[very thick, dotted] (5.5, .75) to (5.65,.525); \draw[very thick] (5.75,.375) to (d); \draw[very thick] (de) to (7.3,1.05); \draw[very thick, dotted] (7.4,.9) to (7.55, .675); \draw[very thick] (7.65,.525) to (e); \end{tikzpicture}\] \caption{A generalized figure of $G$, a simple crowded $(n-k)$-parade.} \label{fig:gen_crowded_parade} \end{figure} For graph $G$, since $d_G(a',b')\geq n-k-1$, it must be the case that either $a'$ or $b'$ is in $K_{m_1}$ or $K_{m_{n-k-1}}$, which are on opposite ends of the graph. Furthermore, in order for the path $P'$ from $a'$ to $b'$ to be a unique shortest path in $G'$, it is necessary that $K_{m'_2}$ and $K_{m'_{n-k-2}}$ be empty, where by $K_{m'_i}$ we mean the subgraph in $G'$ corresponding to $K_{m_i}$ in $G$. If these are not empty, then there would be many alternate routes of equal length connecting $a'$ to $b'$. Since $K_{m'_2}$ and $K_{m'_{n-k-2}}$ are empty, and $G'$ is simple, there are graph symmetries identifying any vertex in $K_{m'_1}$ with any other vertex in $K_{m'_1}$ as well as $v_1$. Likewise, there are graph symmetries identifying any vertex in $K_{m'_{n-k-1}}$ with any other vertex in $K_{m'_{n-k-1}}$ and $v_{n-k}$. Thus, there is a graph symmetry identifying $P$ with $P'$. Therefore, $P$ is still a parade in $G'$. Moreover, we can assume that $P$ is the parade for which $G'$ is a crowded parade. In any case, we now have that both $G$ and $G'$ are crowded $(n-k)$-parades on the same number of vertices and are crowded on the same parade $P$, and $G$ is a proper subgraph of $G'$. Thus $G'$ contains at least one edge $e$ that is not in $G$. There are three cases: either $e$ connects two vertices in $P$, or it connects a vertex in $P$ to a vertex not in $P$, or it connects two vertices not in $P$. If $e$ connects two vertices in $P$, then $P$ would not be a parade of $n-k$ vertices in $G'$, so this is a contradiction. If $e$ connects a vertex in $P$ to a vertex not in $P$, then this provides an alternate path of the same or shorter length between the endpoints of $P$, in contradiction to $P$ being a parade. Finally, if $e$ connects two vertices not in $P$, this also provides an alternate path of the same or shorter length between the endpoints of $P$. In any case, we get a contradiction. The theorem is thus proven. \end{proof} \section{Further Questions} \label{further questions} There are many additional questions that one may consider in this line of research. In this paper, we have characterized the minor-minimal graphs $G$ with $\uspcf{G}=k$ for $k=1$ and $k=2$. For larger, $k$, consider the following. \begin{question} Can we characterize the minor-minimal graphs $G$ with $\uspcf{G}=k$ for $k=3$? $k=4$? Etc. \end{question} Algorithmic questions related to the spectator number and spectator floor of a graph have not been considered in this paper, but we hope to work on some of these questions in the future. \begin{question} Can the minor-monotone floor of the spectator number be computed in polynomial time? \end{question} \begin{question} In the algorithm for calculating the minor-monotone floor of the spectator number, what are the optimal edges to add? \end{question} \begin{question} Can the minor-monotone floor of the spectator number be computed with a ``greedy algorithm"? (That is, can we add a set of edges to a graph $G$ to obtain a supergraph $H$ such that each time we add an edge the spectator number is weakly decreasing? strictly decreasing?) \end{question} A related, but distinct question is the following. \begin{question} If $\uspc{G}=k$ and $\uspcf{G}=m$, can we always find a supergraph $F$ of $G$ that achieves $\uspc{F}=i$ for all $i$ such that $m\leq i \leq k$? \end{question} Finally, one can consider how well this bound relates to our original motivation. \begin{question} This problem originated as a bound on the inverse eigenvalue problem. How good of a bound is this, and when does it fail? Etc. \end{question} \section*{Acknowledgements} This work started at the MRC workshop “Finding Needles in Haystacks: Approaches to Inverse Problems using Combinatorics and Linear Algebra”, which took place in June 2021 with support from the National Science Foundation and the American Mathematical Society. The authors are grateful to the organizers of this meeting. In particular, this material is based upon work supported by the National Science Foundation under Grant Number DMS 1916439.

https://arxiv.org/abs/1209.5360

Balanced Allocations and Double Hashing

Double hashing has recently found more common usage in schemes that use multiple hash functions. In double hashing, for an item $x$, one generates two hash values $f(x)$ and $g(x)$, and then uses combinations $(f(x) +k g(x)) \bmod n$ for $k=0,1,2,...$ to generate multiple hash values from the initial two. We first perform an empirical study showing that, surprisingly, the performance difference between double hashing and fully random hashing appears negligible in the standard balanced allocation paradigm, where each item is placed in the least loaded of $d$ choices, as well as several related variants. We then provide theoretical results that explain the behavior of double hashing in this context.

\section{Introduction} \label{sec:introduction} The standard balanced allocation paradigm works as follows: suppose $n$ balls are sequentially placed into $n$ bins, where each ball is placed in the least loaded of $d$ uniform independent choices of the bins. Then the maximum load (that is, the maximum number of balls in a bin) is $\frac{\log \log n}{\log d} + O(1)$, much lower than the $\frac{\log n}{\log \log n} (1 +o(1))$ obtained where each ball is placed according to a single uniform choice \cite{ABKU}. The assumption that each ball obtains $d$ independent uniform choices is a strong one, and a reasonable question, tackled by several other works, is how much randomness is needed for these types of results (see related work below). Here we consider a novel approach, examining balanced allocations in conjunction with {\em double hashing}. In the well-known technique of standard double hashing for open-addressed hash tables, the $j$th ball obtains two hash values, $f(j)$ and $g(j)$. For a hash table of size $n$, $f(j) \in [0,n-1]$ and $g(j) \in [1,n-1]$. Successive locations $h(j,k) = f(j)+kg(j) \bmod n$, $k=0,1,2,\ldots...$, are tried until an empty slot is found. As discussed later in this introduction, double hashing is extremely conducive to both hardware and software implementations and is used in many deployed systems. In our context, we use the double hashing approach somewhat differently. The $j$th ball again obtains two hash values $f(j)$ and $g(j)$. The $d$ choices for the $j$th ball are then given by $h(j,k) = f(j) + k g(j) \bmod n$, $k=0,1,\ldots,d-1$, and the ball is placed in the least loaded. We generally assume that $f(j)$ is uniform over $[0,n-1]$, $g(j)$ is uniform over all numbers in $[1,n-1]$ relatively prime to $n$, and all hash values are independent. (It is convenient to consider $n$ a prime, or take $n$ to be a power of 2 so that the $g(j)$ are uniformly chosen random odd numbers, to ensure the $h(j,k)$ values are distinct.) It might appear that limiting the space of random choices available to the balls in this way might change the behavior of this random process significantly. We show that this is not the case both in theory and in practice. Specifically, by ``essentially indistinguishable'', we mean that, empirically, for any constant $i$ and sufficiently large $n$ the fraction of bins of load $i$ is well within the difference expected by experimental variance for the two methods. Essentially indistinguishable means that in practice for even reasonable $n$ one cannot readily distinguish the two methods. By ``vanishing'' we mean that, analytically, for any constant $i$ the asymptotic fraction of bins of load $i$ for double hashing differs only by $o(1)$ terms from fully independent choices with high probability. A related key result is that $O(\log \log n)$ bounds on the maximum load hold for double hashing as well. Surprisingly, the difference between $d$ fully independent choices and $d$ choices using double hashing are essentially indistinguishable for sufficiently large $n$ and vanishing asymptotically. \footnote{To be clear, we do not mean that there is {\em no} difference between double hashing and fully random hashing in this setting; there clearly is and we note a simple example further in the paper. As we show, analytically in the limit for large $n$ the difference is vanishing (Theorem~\ref{mainthm} and Corollary~\ref{cormain}), and for finite $n$ the results from our experiments demonstrate the difference is essentially indistinguishable (Section~\ref{sec:sims}).} As an initial example of empirical results, Table~\ref{table_example} below shows the fraction of bins of load $x$ for various $x$ taken over 10000 trials, with $n=2^{14}$ balls thrown into $n$ bins using $d=3$ and $d=4$ choices, using both double hashing and fully random hash values (where for our proxy for ``random'' we utilize the standard approach of simply generating successive random values using the drand48 function in C initially seeded by time). Most values are given to five decimal places. The performance difference is essentially indistinguishable, well within what one would expect simply from variance from the sampling process. \begin{table*}[thp] \begin{subfigure}[h]{0.45\textwidth} \centering \begin{tabular}{|c|c|c|} \hline Load & Fully Random & Double Hashing \\ \hline 0 & 0.17693 & 0.17691 \\ 1 & 0.64664 & 0.64670 \\ 2 & 0.17592 & 0.17589 \\ 3 & 0.00051 & 0.00051 \\ \hline \end{tabular} \caption{3 choices, $n=2^{14}$ balls and bins} \end{subfigure} \qquad \begin{subfigure}[h]{0.45\textwidth} \centering \begin{tabular}{|c|c|c|} \hline Load & Fully Random & Double Hashing \\ \hline 0 & 0.14081 & 0.14081 \\ 1 & 0.71840 & 0.71841 \\ 2 & 0.14077 & 0.14076 \\ 3 & $2.25 \cdot 10^{-5}$ & $2.29\cdot 10^{-5}$ \\ \hline \end{tabular} \caption{4 choices, $n=2^{14}$ balls and bins} \end{subfigure} \caption{An initial example showing the performance of double hashing compared to fully random hashing. In our tables, the row with load $x$ gives the fraction of the bins that have load $x$ over all trials. So over 10000 trials of throwing $n=2^{14}$ balls into $2^{14}$ bins using 3 choices and double hashing, the fraction of bins with load 0 was $0.17691$. \label{table_example}} \end{table*} More extensive empirical results appear in Appendix~\ref{sec:sims}. In particular, we also consider two extensions to the standard paradigm: V\"{o}cking's extension (sometimes called $d$-left hashing), where the $n$ bins are split into $d$ subtables of size $n/d$ laid out left to right, the $d$ choices consist of one uniform independent choice in each subtable, and ties for the least loaded bin are broken to the left \cite{vocking}; and the continuous variation, where the bins represent queues, and the balls represent customers that arrive as a Poisson process and have exponentially distributed service requirements \cite{Mitzenmacher}. We again find empirically that replacing fully random choices with double hashing leads to essentially indistinguishable results in practice.\footnote{We encourage the reader to examine these experimental results. However, because we recognize some readers are as a rule uninterested in experimental results, we have moved them to an appendix.} In this paper, we provide theoretical results explaining why this would be the case. There are multiple methods available that can yield $O(\log \log n)$ bounds on the maximum load when $n$ balls are thrown into $n$ bins in the setting of fully random choices. We therefore first demonstrate how some previously used methods, including the layered induction approach of \cite{ABKU} and the witness tree approach of \cite{vocking}, readily yield $O(\log \log n)$ bounds; this asymptotic behavior is, arguably, unsurprising (at least in hindsight). We then examine the key question of why the difference in empirical results is vanishing, a much stronger requirement. For the case of fully random choices, the asymptotic fraction of bins of each possible load can be determined using fluid limit methods that yield a family of differential equations describing the process behavior \cite{Mitzenmacher}. It is not a priori clear, however, why the method of differential equations should necessarily apply when using double hashing, and the primary result of this paper is to explain why it in fact applies. The argument depends technically on the idea that the ``history'' engendered by double hashing in place of $d$ fully random hash functions has only a vanishing (that is, $o(1)$) effect on the differential equations that correspond to the limiting behavior of the bin loads. We believe this resolution suggests that double hashing will be found to obtain the same results as fully random hashing in other additional hash-based structures, which may be important in practical settings. We argue these results are important for multiple reasons. First, we believe the fact that moving from fully random hashing to double hashing does not change performance for these particular balls and bins problems is interesting in its own right. But it also has practical applications; multiple-choice hashing is used in several hardware systems (such as routers), and double hashing both requires less (pseudo-)randomness and is extremely conducive to implementation in hardware \cite{chunkstash,heileman2005caching}. (As we discuss below, it may also be useful in software systems.) Both the fact that double hashing does not change performance, and the fact that one can very precisely determine the performance of double hashing for load balancing simply using the same fluid limit equations as have been used under the assumption of fully random hashing, are therefore of major importance for designing systems that use multiple-choice methods (and convincing system designers to use them). Finally, as mentioned, these results suggest that using double hashing in place of fully random choices may similarly yield the same performance in other settings that make use of multiple hash functions, such as for cuckoo hashing or in error-correcting codes, offering the same potential benefits for these problems. We have explored this issue further in a subsequent (albeit already published) paper \cite{MT}, where there remain further open questions. In particular, we have not yet found how to use the fluid limit analysis used here for these other problems. Finally, it has been remarked to us that all of our arguments here apply beyond double hashing; any hashing scheme where the $d$ choices for a ball are made so that they are pairwise independent and uniform would yield the same result by the same argument. That is, if for a given ball with $d$ choices $h_1,h_2,\ldots,h_d$, for any distinct bins $b_1$ and $b_2$ we have for all $1\leq i,j \leq d, i \neq j$: $$\Pr(h_i = b_1) = 1/n \mbox{ and }$$ $$\Pr(h_i = b_1 \mbox{ and } h_j = b_2) = \frac{1}{{n \choose 2}},$$ then our results apply. Unfortunately, we do not know of any actual scheme besides double hashing in practical use with these properties; hence we focus on double hashing throughout. \subsection{Related Work} The balanced allocations paradigm, or the power of two choices, has been the subject of a great deal of work, both in the discrete balls and bins setting and in the queueing theoretic setting. See, for example, the survey articles \cite{KSurvey,TwoSurvey} for references and applications. Several recent works have considered hashing variations that utilize less randomness in place of assuming perfectly random hash functions; indeed, there is a long history of work on universal hash functions \cite{CarterWe79}, and more recently min-wise independent hashing \cite{mwi}. Specific recent related works include results on standard one-choice balls and bins problems \cite{SHF}, hashing with linear probing with limited independence \cite{ppr}, and tabulation hashing \cite{pt}; other works involving balls and bins with less randomness include \cite{godfrey,peres}. As another example, Woelfel shows that a variation of V\"{o}cking's results hold using simple hash functions that utilize a collection of $k$-wise independent hash functions for small $k$, and a random vector requiring $o(n)$ space \cite{Woelfel}. Another related work in the balls and bins setting is the paper of Kenthapadi and Panigrahy \cite{kpbalanced}, who consider a setting where balls are not allowed to choose any two bins, but are forced to choose two bins corresponding to an edge on an underlying random graph. In the same paper, they also show that two random choices that yield $d$ bins are sufficient for similar $O(\log \log n)$ bounds on maximum loads that one obtains with $d$ fully random choices, where in their case each random choice gives a contiguous block of $d/2$ bins. Interestingly, the classical question regarding the average length of an unsuccessful search sequence for standard double hashing in an open address hash table when the table load is a constant $\alpha$ has been shown to be, up to lower order terms, $1/(1-\alpha)$, showing that double hashing has essentially the same performance as random probing (where each ball would have its own random permutation of the bins to examine, in order, until finding an empty bin) when using traditional hash tables \cite{bradford2007probabilistic,guibas1978analysis,lueker1993more}. These results appear to have been derived using different techniques than we utilize here; it could be worthwhile to construct a general analysis that applies for both schemes. A few papers have recently suggested using double hashing in schemes where one would use multiple hash functions and shown little or no loss in performance. For Bloom filters, Kirsch and Mitzenmacher \cite{kirsch:bbb}, starting from the empirical analysis by Dillinger and Manolios \cite{dillinger3312bfp}, prove that using double hashing has negligible effects on Bloom filter performance. This result is closest in spirit to our current work; indeed, the type of analysis here can be used to provide an alternative argument for this phenomenon, although the case of Bloom filters is inherently simpler. Several available online implementations of Bloom filters now use this approach, suggesting that the double hashing approach can be significantly beneficial in software as well as hardware implementations.\footnote{See, for example, \url{http://leveldb.googlecode.com/svn/trunk/util/bloom.cc}, \url{https://github.com/armon/bloomd}, and \url{http://hackage.haskell.org/packages/archive/bloomfilter/1.0/doc/html/bloomfilter.txt}.} Bachrach and Porat use double hashing in a variant of min-wise independent sketches \cite{bachrach2010fast}. The reduction in randomness stemming from using double hashing to generate multiple hash values can be useful in other contexts. For example, it is used in \cite{MVad} to improve results where pairwise independent hash functions are sufficient for suitably random data; using double hashing requires fewer hash values to be generated (two in place of a larger number), which means less randomness in the data is required. Finally, in work subsequent to the original draft of this paper \cite{MT}, we have empirically examined double hashing for other algorithms such as cuckoo hashing, and again found essentially no empirical difference between fully random hashing and double hashing in this and other contexts. However, theoretical results for these settings that prove this lack of difference are as of yet very limited. Arguably, the main difference between our work and other related work is that in our setting with double hashing we find the empirical results are essentially indistinguishable in practice, and we focus on examining this phenomenon. \section{Initial Theoretical Results} We now consider formal arguments for the excellent behavior for double hashing. We begin with some simpler but coarser arguments that have been previously used in multiple-choice hashing settings, based on majorization and witness trees. While our witness tree argument dominates our majorization argument, we present both, as they may be useful in considering future variations, and they highlight how these techniques apply in these settings. In the following section, we then consider the fluid limit methodology, which best captures the result we desire here, namely that the load distributions are essentially the same with fully random hashing and double hashing. However, the fluid limit methodology captures results about the fraction of bins with load $i$, for every constant value $i$, and does not readily provide $O(\log \log n)$ bounds (without specialized additional work, which often depend on the techniques used below). The reader conversant with balanced allocation results utilizing majorization and witness trees may choose to skip this section. \subsection{A Majorization Argument} We first note that using double hashing with two choices and using random hashing with two distinct hash values per ball are equivalent. With this we can provide a simple argument, showing the seemingly obvious fact that using double hashing with $d > 2$ choices is at least as good as using 2 random choices. This in turn shows that double hashing maintains $\log \log n +O(1)$ maximum load in the standard balls and bins setting. Our approach uses a standard majorization and coupling argument, where the coupling links the random choices made by the processes when using double hashing and using random hashing while maintaining the fidelity of both individual processes. (See, e.g., \cite{ABKU,Steger}, or \cite{majorization} for more background on majorization.) Let $\vec{x}=(x_1,\ldots,x_n)$ be a vector with elements in non-increasing order, so $x_1 \geq x_2 \ldots \geq x_n$, and similarly for $\vec{y}=(y_1,\ldots,y_n)$. We say that $\vec{x}$ majorizes $\vec{y}$ if $\sum_{i=1}^n x_i = \sum_{i=1}^n y_i$ and, for $j < n$, $\sum_{i=1}^j x_i \geq \sum_{i=1}^j y_i$. For two Markovian processes $X$ and $Y$, we say that $X$ stochastically majorizes $Y$ if there is a coupling of the processes $X$ and $Y$ so that at each step under the coupling the vector representing the state of $X$ majorizes the vector representing the state of $Y$. We note that because we use the loads of the bins as the state, the balls and bins processes we consider are Markovian. We make use of the following simple and standard lemma. (See, for example, \cite[Lemma 3.4]{ABKU}.) \begin{lemma} \label{lem:maj0} If $\vec{x}$ majorizes $\vec{y}$ for vectors $\vec{x}$, $\vec{y}$ of positive integers, and $e_i$ represents a unit vector with a 1 in the $i$th entry and 0 elsewhere, then $\vec{x}+e_i$ majorizes $\vec{y}+e_j$ for $j \geq i$. \end{lemma} \begin{theorem} Let process $X$ be the process where $m$ balls are placed into $n$ bins with two distinct random choices, and $Y$ be the corresponding scheme with $d > 2$ choices using double hashing. Then $X$ stochastically majorizes $Y$. \end{theorem} \begin{proof} At each time step, we let $\vec{x}(t)$ and $\vec{y}(t)$ be the vectors corresponding to the loads sorted in decreasing order. We inductively claim that $\vec{x}(t)$ majorizes $\vec{y}(t)$ at all time steps under the coupling of the processes where if the $a$th and $b$th bins in the sorted order for $X$ are chosen, the $a$th and $b$th bins in the sorted order for $Y$ are chosen as the first two choices, and then the remaining choices are determined by double hashing. That is, the $d$ hash choices are such that the gap between successive choices is $b-a$, so the choices are $a$, $b$, $2b-a$, $3b-2a$, and so on (modulo the size of the table). Clearly $\vec{x}(0)$ majorizes $\vec{y}(0)$ as the vectors are equal. It is simple to check that this process maintains the majorization using Lemma~\ref{lem:maj0}, as the coordinate that increases in $\vec{y}(t)$ at each step is deeper in the sorted order than the coordinate that increases in $\vec{x}(t)$. \end{proof} As two random choices stochastically majorizes $d$ choices from double hashing under this coupling, we see that $$\Pr(x_1 \geq c) \geq \Pr(y_1 \geq c)$$ for any value $c$. Since the seminal result of \cite{ABKU} shows that using two choices gives a maximum load of $\log \log n +O(1)$ with high probability, we therefore have this corollary. \begin{corollary} The maximum load using $d > 2$ choices and double hashing for $n$ balls and $n$ bins is $\log \log n +O(1)$ with high probability. \end{corollary} We note that similarly, when using double hashing, we can show that using $d$ choices stochastically majorizes using $d+1$ choices. \subsection{A Witness Tree Argument} \label{sec:witness} It is well known that $d>2$ choices performs better than 2 choices for multiple-choice hashing; while the maximum load remains $O(\log \log n)$, the constant factor depends on $d$, and can be important in practice. Our simple majorization argument does not provide this type of bound, so to achieve it, we next utilize the witness tree approach, following closely the work of V\"{o}cking \cite{vocking}. (See also \cite{simplified} for related arguments.) While we discuss the case of insertions only, the arguments also apply in settings with deletions as well; see \cite{vocking} for more details. Similarly, here we consider only the standard balls and bins setting of $n$ balls and $n$ bins with $d \geq 3$ being a constant, but similar results for $m = cn$ balls for some constant $c$ can also be derived by simply changing the ``base case'' at the leaves of the witness tree accordingly, and similar results for V\"{o}cking's scheme can be derived by using the ``unbalanced'' witness tree used by V\"{o}cking \cite{vocking} in place of the balanced one. These methods allow us to prove statements of the following form: \begin{theorem} \label{thm:vresult} Suppose $n$ balls are placed into $n$ bins using the balanced allocation scheme with double hashing as described above. Then with $d$ choices the maximum load is $\log \log n / \log d + O(d)$ with high probability. \end{theorem} We note that, while V\"{o}cking obtains a bound of $\log \log n / \log d + O(1)$, we have an $O(d)$ term that appears necessary to handle the leaves in our witness tree. (A similar issue appears to arise in \cite{Woelfel}.) For constant $d$ these are asymptotically the same; however, an $O(1)$ additive term is more pleasing both theoretically and potentially in practice. How we deviate from V\"{o}cking's argument is explained below. \begin{proof} Following \cite{vocking}, we define a witness tree, which is a tree-ordered (multi)set of balls. Each node in the tree represents a ball, inserted at a certain time; the $i$th inserted ball corresponds to time $i$ in the natural way. The ball represented by the root $r$ is placed at time $t$, and a child node must have been inserted at a time previous to its parent. A leaf node in V\"{o}cking's argument is {\em activated} if each of the $d$ locations of the corresponding ball contains at least three balls when it is inserted. An edge $(u,v)$ is activated if when $v$ is the $i$th child of $u$, then the $i$th location of $u$'s ball is the same as one of the locations of $v$'s ball. A witness tree is activated if all of its leaf nodes and edges are activated. Following V\"{o}cking's approach, we first bound the probability that a witness tree is activated for the simpler case where the nodes of the witness trees represent distinct balls. The argument then can be generalized to deal with witness trees where the same ball may appear multiple times. As this follows straightforwardly using the technical approach in \cite{vocking}, we do not provide the full argument here. We now explain where we must deviate from V\"{o}cking's argument. The original argument utilizes the fact that most $n/3$ bins have load at least 3, deterministically. As leaf nodes in V\"{o}cking's argument are required to have all $d$ choices of bins have load at least 3 to be activated, a leaf node corresponding to a ball with $d$ choices of bins is activated with probability at most $3^{-d}$, and a collection of $q$ leaf nodes are all activated with probability $3^{-dq}$. However, this argument will not apply in our case, because the choices of bins are not independent when using double hashing, and depending on which bins are loaded, we can obtain very different results. For example, consider a case where the first $n/3$ bins have load at least 3. The fraction of choices using double hashing where all $d$ bins have load at least 3 is significantly more than $3^{-d}$, which would be the probability if $n/3$ bins with load 3 were randomly distributed. Indeed, for a newly placed ball $j$, if $f(j)$ and $g(j)$ are both less than $n/(3(d+1))$, all $d$ choices will have load at least 3, and this occurs with probability at least $(9(d+1)^2)^{-1}$. While such a configuration is unlikely, the deterministic argument used by V\"{o}cking no longer applies. We modify the argument to deal with this issue. In our double hashing setting, let us call a leaf active if either \vspace{-0.05in} \begin{itemize} \item Some ball in the past has two or more of the bins at this leaf among its $d$ choices. \vspace{-0.1in} \item All the $d$ bins chosen by this ball have previously been chosen by $4d$ previous balls. \end{itemize} \vspace{-0.05in} The probability that any previous ball has hit two or more of the bins at the leaf is $O(d^4n^{-1})$: there are ${d \choose 2}$ pairs of bins from the $d$ choices at the leaf; at most $d(d-1)$ pairs of positions within the $d$ choices where that pair could occur in any previous ball; at most $n$ possible previous balls; and each bad choice that leads that previous ball to have a specific pair of bins in a specific pair of positions occurs with probability $1/(n(n-1))$. Once we exclude this case, we can consider only balls that hit at most one of the $d$ bins associated with the leaf. For any time corresponding to a leaf, we bound the probability that any specific bin has been chosen by $4d$ or more previous balls. We note by symmetry that the probability any specific ball chooses a specific bin is $d/n$. The probability in question is then at most $${n \choose 4d}\left (\frac{d}{n} \right)^{4d} \leq \frac{d^{4d}}{(4d)!} < \left ( \frac{e}{4} \right )^d,$$ which is less than $\frac{1}{3}$ whenever $d \geq 3$. Further, once we consider the case of previous balls that choose two or more bins at this leaf separately, the events that the $d$ bins chosen by this ball have previously been chosen by $4d$ previous balls are negatively correlated. Hence, we find the probability a specific leaf node is activated is less than $3^{-d}$. However, following \cite{vocking}, we need to consider a collection of $q$ leaves and show the probability that they are all active is at most $3^{-dq}$. We will do this below by using Azuma's inequality to show the fraction of choices of hash values from double hashing that lead to an activated ball is less than $3^{-d}$ with high probability. As balls corresponding to leaves independently choose their hash values, this result suffices. Let $S$ be the set of pairs of hash values that generate $d$ values that would activate a leaf at time $n$. We have $\mathbb{E}[|S|] < \left ( \frac{e}{4} \right )^d n(n-1) + cd^4(n-1)$ for some constant $c$, so $\mathbb{E}[|S|] > (3^{-d} - \gamma)n(n-1)$ for some constant $\gamma$ and large enough $n$. Consider the Doob martingale obtained by revealing the bins for the balls one at a time. Each ball can change the final value of $S$ by at most $dn$, since the bin where any ball is placed is involved in less than $dn$ choices of pairs. Azuma's inequality (e.g., \cite[Section 12.5]{MU}) then yields $$\Pr(|S| > 3^{-d}n(n-1)) \leq \mbox{exp}(-\delta n)$$ for a constant $\delta$ that depends on $d$ and $\gamma$. It follows readily that the fraction of pairs of hash values that activate a leaf is at most $3^{-d}$ with very high probability throughout the process; by conditioning on this event, we can continue with V\"{o}cking's argument. (The conditioning only adds an exponentially small additional probability to the probability the maximum load exceeds our bound.) Specifically, we note for there to be a bin of load $L+4d$, there must be an activated witness tree of depth $L$. We can bound the probability that some witness tree (with distinct balls) of depth $L$ is activated. The probability an edge is activated is the probability a ball chooses a specific bin, which as previously noted is $d/n$. As all balls are distinct, the probability that a witness tree of $m$ balls has all edges activated is $(d/n)^{m-1}$, and as we have shown the probability of all leaves being activated is bounded above by $3^{-dq}$ where $q=d^L$ is the number of leaves. Following \cite{vocking}, as there are at most $n^m$ ways of choosing the balls for the witness tree, the probability that there exists an active witness tree is at most \begin{eqnarray*} n^m \left ( \frac{d}{n} \right)^{m-1} 3^{-dq} & \leq & n \cdot d^{2q} \cdot 3^{-dq} \\ & \leq & n \cdot 2^{-q} \\ & = & n \cdot 2^{-d^{L}}. \end{eqnarray*} Hence choosing $L \leq \log_d \log_2 n + \log_d(1+\alpha)$ guarantees a maximum load of $L + 4d$ with probability $O(n^{-\alpha})$. \end{proof} \section{The Fluid Limit Argument} \label{sec:fluid} We now consider the fluid limit approach of \cite{MitzenmacherThesis}. (A useful survey of this approach appears in \cite{Diaz}.) The fluid limit approach gives equations that describe the asymptotic fraction of bins with each possible integer load, and concentration around these values follows from martingale bounds (e.g., \cite{EK,Kurtz,Wormald}). Values can easily be determined numerically, and prove highly accurate even for small numbers of balls and bins. We show that the same equations apply even in the setting of double hashing, giving a theoretical justification for our empirical findings in Appendix~\ref{sec:sims}. This approach can be easily extended to other multiple choice processes (such as V\"{o}cking's scheme and the queuing setting). We emphasize that the fluid limit approach does not, in itself, yield bounds of the type that the maximum load is $O(\log \log n)$ with high probability naturally; rather, it says that for any constant integer $i$, the fraction of bins of load $i$ is concentrated around the value obtained by the fluid limit. One generally has to do additional work -- generally similar in nature to the arguments in the proceeding sections -- to obtain $O(\log \log n)$ bounds. As we already have an $O(\log \log n)$ bound from alternative techniques, here our focus is on showing the fluid limits are the same under double hashing and fully random hashing, which explains our empirical findings. (We show one could achieve an $O(\log \log n)$ bound from the results of this section -- actually bound of $\log_d \log_2 n + O(1)$ -- in Appendix~\ref{sec:followon}.) The standard balls and bins fluid limit argument runs as follows. Let $X_i(t)$ be a random variable denoting the number of bins with load {\em at least} $i$ after $tn$ balls have been thrown; hence $X_0(0) =n$ and $X_i(0) = 0$ for all $i \geq 1$. Let $x_i(t) = X_i(t)/n$. For $X_i$ to increase when a ball is thrown, all of its choices must have load at least $i-1$, but not all of them can have load at least $i$. Hence for $i \geq 1$ $$\mathbb{E}[X_i(t + 1/n) - X_i(t)] = (x_{i-1}(t))^d - (x_{i}(t))^d.$$ Let $\Delta(x_i) = x_i(t + 1/n) - x_i(t)$ and $\Delta(t) = 1/n$. Then the above can be written as: $$\mathbb{E} \left [ \frac{\Delta(x_i)}{\Delta(t)} \right ] = (x_{i-1}(t))^d - (x_{i}(t))^d.$$ In the limit as $n$ grows, we can view the limiting version of the above equation as $$\frac{dx_i}{dt} = x_{i-1}^d - x_{i}^d,$$ where we remove the $t$ on the right hand side as the meaning is clear. Again, previous works \cite{EK,Kurtz,Wormald} justify how the Markovian load balancing process converges to the solution of the differential equations.\footnote{In particular, the technical conditions corresponding to Wormald's result \cite[Theorem 1]{Wormald} hold, and this theorem gives the appropriate convergence; we explain further in our Theorem~\ref{mainthm}.} Specifically, it follows from Wormald's theorem \cite[Theorem 1]{Wormald} that $$X_i(t) = nx_i(t) + o(n)$$ with probability $1-o(1)$, or equivalently that the fraction of balls of load $i$ is within $o(1)$ of the result of the limiting differential equations with probability $1-o(1)$. These equations allow us to compute the limiting fraction of bins of each load numerically, and these results closely match our simulations, as for example shown in Table~\ref{table_6}. \begin{table*}[thp] \centering \begin{tabular}{|c|c|c|c|} \hline Tail load & Fluid Limit & Fully Random & Double Hashing \\ \hline $\geq 1$ & 0.8231 & 0.8231 & 0.8231 \\ $\geq 2$ & 0.1765 & 0.1764 & 0.1764 \\ $\geq 3$ & 0.00051 & 0.00051 & 0.00051 \\ \hline \end{tabular} \caption{3 choices, fluid limit ($n=\infty$) vs. $n=2^{14}$ balls and bins} \label{table_6} \end{table*} Given our empirical results, it is natural to conclude that these differential equations must also necessarily describe the behavior of the process when we use double hashing in place of standard hashing. The question is how can we justify this, as the equations were derived utilizing the independence of choices, which is not the case for double hashing. We now prove that, for constant number of choices $d$, constant load values $i$, and a constant time $T$ (corresponding to $Tn$ total balls), the loads of the bins chosen by double hashing behave essentially the same as though the choices were independent, in that, with high probability over the entire course of the process, $$\mathbb{E}[X_i(t + 1/n) - X_i(t)] = (x_{i-1}(t))^d - (x_{i}(t))^d +o(1);$$ that is, the gap is only in $o(1)$ terms. This suffices for \cite[Theorem 1]{Wormald} (specifically condition (ii) of \cite[Theorem 1]{Wormald} allows such $o(1)$ differences). The result is that double hashing has no effect on the fluid limit analysis. (Again, we emphasize our restriction to constant choices $d$, constant load values $i$, and constant time parameter $T$.) Our approach is inspired by the work of Bramson, Lue, and Prabhakar \cite{Bramson}, who use a similar approach to obtain asymptotic independence results in the queueing setting. However, there the concern was on limiting independence in equilibrium with general service time distributions, and the choices of queues were assumed to be purely random. We show that this methodology can be applied to the double hashing setting. \begin{lemma} \label{lem:mainlemma} When using double hashing, with high probability over the entire course of the process, $$\mathbb{E}[X_i(t + 1/n) - X_i(t)] = (x_{i-1}(t))^d - (x_{i}(t))^d +o(1).$$ \end{lemma} \begin{proof} We refer to the {\em ancestry list} of a bin $b$ at time $t$ as follows. The list begins with the balls $z_1,z_2,\ldots,z_{g(b,t)}$ that have had bin $b$ as one of their choices, where $g(b,t)$ is the number of balls that have chosen bin $b$ up to time $t$. Note that each $z_i$ is associated with a corresponding time $t_i$ and $d-1$ other bin choices. For each $z_i$, we recursively add the list of balls that have chosen each of those $d-1$ bins up to time $t_i$, and so on recursively. We also think of the bins associated with these balls as being part of the ancestry list, where the meaning is clear. It is clear that the ancestry list gives all the necessary information to determine the load of the bin $b$ at time $t$ (assuming the information regarding choices is presented in such a way to include how placement will occur in case of ties; e.g., the bin choices are ordered by priority). We note that the ancestry list holds more information (and more balls and bins) than the witness trees used by V\"{o}cking (and by us in Section~\ref{sec:witness}). In what follows below let us assume $n$ is prime for convenience (we explain the difference if $n$ is not prime in footnotes). We claim that for asymptotic independence of the load among a collection of $d$ bins at a specific time when a new ball is placed, it suffices to show that these ancestry lists are small. Specifically, we start with showing in Lemma~\ref{lem:branching} that all ancestry lists contain only $O(\log n)$ associated bins with high probability. We then show as a consequence in Lemma~\ref{lem:small} that the ancestry lists of the bins associated with a newly placed ball have no bins in common with high probability. This last fact allows us to complete the main lemma, Lemma~\ref{lem:mainlemma}. \begin{lemma} \label{lem:branching} The number of bins in the ancestry list of every bin after the first $Tn$ steps is at most $O(\log n)$ with high probability. \end{lemma} \begin{proof} We view the growth of the ancestry list as a variation of the standard branching process, by going backward in time. Let $B_{0} = 1$ correspond to size of an initial ancestry list of a bin $b$, consisting of the bin itself. If the $(Tn)$th ball thrown has $b$ as one of its $d$ choices, then $d-1$ additional bins are added to the ancestry list, and we then have $B_{1} = d$; otherwise we have no change and $B_{1} = 1$. (Note that when measuring the size of the ancestry list in bins, each bin is counted only once, even if it is associated with multiple balls.) If the $(Tn-1)$st ball thrown has a bin in the ancestry list as one of its $d$ choices, then (at most) $d-1$ bins are added to the ancestry list, and we set $B_2 = B_1 + d-1$; otherwise, we have $B_2 = B_1$. We continue to add to the ancestry list with at each step $B_i = B_{i-1} + d-1$ or $B_i = B_{i-1}$, depending on whether the $(Tn-i+1)$st ball has one of it choices as a bin on the ancestry list, or not. This process is {\em almost} equivalent to a Galton-Watson branching process where in each generation, each existing element produces 1 offspring with probability $1-d/n$ (or equivalently, moves itself into the next generation), or produces $d$ offspring (adding $d-1$ new elements) with probability $d/n$. The one issue is that the production of offspring are not independent events; at most $d-1$ elements are added at each step in the process. (There is also the issue that perhaps fewer than $d-1$ elements are added when elements are added to the ancestry list; for our purposes, it is pessimistic to assume $d-1$ offspring are produced.) Without this dependence concern, standard results on branching process would give that $E[B_{Tn}] = (1+d(d-1)/n)^{Tn} \leq e^{Td(d-1)}$, which is a constant. Further, we could apply (Chernoff-like) tail bounds from Karp and Zhang \cite[Theorem 1]{KZ}, which states the following: for a supercritical finite time branching process $\left \{ Z_n \right \}$ over $n$ time steps starting with $Z_0 =1$, with mean offspring per element $\mathbb{E}[Z_1] = \rho >1$, and with $\mathbb{E}[e^{Z_1}] < \infty$, there exists constants $c_1$ and $c_2$ such that $$\Pr(Z_n > \gamma \rho^n) < c_1 e^{-c_2 \gamma}.$$ In our setting, that would give that there exists constants $c_1$ and $c_2$ such that $$\Pr(B_{Tn} > \gamma (1+d(d-1)/n)^{Tn} ) < c_1 e^{-c_2 \gamma}.$$ This would give our desired $O(\log n)$ high probability bound on the size of the ancestry list. To deal with this small deviation, it suffices to consider a modified Galton-Watson process where each element produces $d$ offspring with probability $d'/n$; we shall see that $d'= d+1$ suffices. Let $B'$ be the resulting size of this Galton Walton process. From the above we have that $B' < c \log n$ with high probability for some suitable constant $c$. Our original desired ancestry list process is dominated by a process where $B_i = \min(B_{i-1} + d-1,n)$ with probability $\min(B_{i-1}d/n,1)$ and $B_i = B_{i-1}$ otherwise, and this process is in turn dominated for values of $B_i$ up to $c \log n$ by a Galton-Waston branching process where the constant $d'$ satisfies $$1-(1-d'/n)^x \geq dx/n$$ for all $1 \leq x \leq c \log n$, so that at every stage the Galton-Watson process is more likely to have at least $d-1$ new offspring (and may have more). We see $d' = d+1$ suffices, as $$1-(1-(d+1)/n)^x = x(d+1)/n - O(dx^2/n^2)$$ which is greater than $dx/n$ for $n$ sufficiently large when $x$ is $O(\log n)$. The straightforward step by step coupling of the processes yields that $$\Pr(B_{Tn} > c \log n) \leq \Pr(B' > c \log n),$$ giving our desired bound. We also suggest a slightly cleaner alternative, which may prove useful for other variations: embed the branching process in a continuous time branching process. We scale time so that balls are thrown as a Poisson process or rate $n$ per unit time over $T$ time units. Each element therefore generates $d-1$ new offspring at time instants that are exponentially distributed with mean $1/d$ (the average time before a ball hits any bin on the ancestry list). Again, assuming $d-1$ new offspring is a pessimistic bound. If we let $C_t$ be the number of elements at time $t$ (starting from 1 element at time 0), it is well known (see, e.g., \cite[p.108 eq. (4)]{AN}, and note that generating $d-1$ new offspring is equivalent to ``dieing'' and generating $d$ offspring) that for such a process, $$\mathbb{E}[C_t] = e^{td(d-1)}.$$ In our case, we run to a fixed time $T$ and $\mathbb{E}[C_T] = e^{Td(d-1)}$, a constant. Indeed, in this specific case, the generating function for the distribution of the number of elements is known (see, e.g., \cite[p.109]{AN}), allowing us to directly apply a Chernoff bound. Specifically, $$\mathbb{E}[s^{C_t}] = se^{-dt}[1-(1-e^{-d(d-1)t})s^{d-1}]^{-1/(d-1)}.$$ Hence we have \begin{eqnarray*} \Pr(C_T > \gamma e^{Td(d-1)}) & = & \Pr(e^{C_T} > e^{\gamma e^{Td(d-1)}}) \\ & \leq & e^{-\gamma e^{Td(d-1)}} \mathbb{E}[e^{C_T}] \\ & \leq & c_3 e^{-c_4 \gamma} \end{eqnarray*} for constants $c_3$ and $c_4$ that depend on $d$ and $T$. Hence, this gives that the size of the ancestry list as viewed from the setting of the continuous branching process is $O(\log n)$ with high probability. The last concern is that running the continuous process for time $Tn$ does not guarantee that $Tn$ balls are thrown; this can be dealt with by thinking of the process running for a slightly longer time $T' > T$. That is, choose $T'= T +\epsilon$ for a small constant $\epsilon$. Standard Chernoff bounds on the Poisson random variables then guarantee that at least $Tn$ balls are then thrown with high probability, and the size of the ancestry lists are stochastically monotonically increasing with the number of balls thrown. Changing to $T'$ time units maintains that each ancestry list is $O(\log n)$ with high probability. Finally, by choosing the constant in the $O(\log n)$ term appropriately, we can achieve a high enough probability to apply a union bound so that this holds for all ancestry lists simultaneously with high probability. \end{proof} We now use Lemma~\ref{lem:branching} to show the following. \begin{lemma} \label{lem:small} The bins in the ancestry lists of the $d$ choices are disjoint with probability $1-\eta$ for $\eta = O(d^2 \log^2 n/n) = o(1)$. \end{lemma} \begin{proof} Let {\cal F} be the probability that the bins are disjoint, and let ${\cal E}$ be the event that no pair of the $d$ choices were previously chosen by the same ball. If ${\cal E}$ occurs, the ancestry lists are clearly not disjoint. Hence we wish to bound $$\Pr(F) \leq \Pr({\cal E}) + \Pr({\cal F} | \neg{\cal E}).$$ Consider any two of the $d$ bins chosen by the ball being placed. Each of the up to $Tn$ previous balls have $O(d^2)$ ways of choosing those two bins as two of their $d$ choices (e.g., picking that bin as the 2nd and 4th choice, for example), and the probability of choosing those two bins for each possible pair of choice positions is $O(1/n^2)$.\footnote{If $n$ is not prime, this probability is $O(1/n\phi(n))$, where $\phi$ is the Euler totient function counting the number of numbers less than $n$ that are relatively prime to $n$. We note $\phi(n)$ is usually $\Omega(n)$ and is always $\Omega(n/\log \log n)$, so this does not affect our argument substantially.} There are ${d \choose 2}$ pairs of balls, so by a union bound $\Pr({\cal E})$ is $O(Td^4/n^2)$. Now suppose that no pair of the $d$ bins were previously chosen by the same ball. Suppose the bins for each of the ancestry lists of the $d$ choices are ordered in some fixed fashion (say according to decreasing ball time, randomly permuted for each ball). We consider the probability that the $i$th bin in the ancestry list of one bin matches the $j$th bin in another. Since the lists do not share any ball in common, the $j$th bin in the second list matches the $i$th bin in the first list with probability only $O(1/n)$, as even conditioned on the value of the $i$th bin on the first list, the $j$th bin on the second list is uniform over $\Omega(n)$ possibilities.\footnote{Again, for $n$ not prime, we may use $\Omega(\phi(n))$ possibilities.} We now condition on all of the $d$ ancestry lists being of size $O(\log n)$; from Lemma~\ref{lem:branching}, this can be made to occur with any inverse polynomial probability by choosing the constant factor in the $O(\log n)$ term, so we assume this bound on ancestry list sizes. In his case, the probability of a match among any of the $d$ bins is only $O(d^2 \log^2 n/n)$ in total, where the $d^2$ factor is from the $d \choose 2$ possible ways of choosing bins, and the $\log^2 n$ term follows the bound on the size ancestry lists. Hence $\Pr({\cal F} | \neg{\cal E})$ is $O(d^2 \log^2 n/n)$, and the total probability that the ancestry lists of the $d$ choices are {\em not} disjoint is $\eta = O(d^2 \log^2 n/n) = o(1)$. \end{proof} We now show that this yields the Lemma~\ref{lem:mainlemma}. To clarify this, consider bins $b_1,b_2,\ldots,b_d$ that were chosen by a ball at some time $t+1/n$. (Recall our scaling of time.) The probability that all $d$ bins have load at least $i$ at that time is equivalent to the probability that each bin $b_j$ has a corresponding ancestry list $A_j$ showing that it has load $i$ at some time $u_j \leq t$. Fix a collection of ancestry lists $A_j$, and let $E_j$ be the event defined by ``bin $b_j$ has ancestry list $A_j$''. If these ancestry lists have disjoint sets of bins, then the corresponding balls in each ancestry list occur at different times and have no intersecting bins, and as such $$ \Pr \left ( \cap_j E_j \right ) = \prod_j \Pr(E_j).$$ For constant $i$, $t$, and $d$, the probability that all $d$ bins have load at least $i$ is constant. Hence, if the probability that the ancestry lists for the $d$ bins intersect at any bin is $\eta = o(1)$, we have asymptotic independence. Specifically, let $\cal X$ be the set of collections of $d$ ancestry lists for balls $b_1,b_2,\ldots,b_d$ that yield that each bin has load at least $i$ at time $t$, let $\cal Y$ be the subset of collections in $\cal X$ where the $d$ ancestry lists have no bins in common, and for a collection $Z$ in $\cal X$ let $E_j(Z)$ be the corresponding event defined by ``bin $b_j$ has ancestry list $A_j$ in collection $Z$''. Then \begin{eqnarray*} \sum_{Z \in {\cal X}} \Pr \left ( \cap_j E_j(Z) \right ) &= & \left [ \sum_{Z \in {\cal Y}} \Pr \left ( \cap_j E_j(Z) \right ) \right ] + o(1) \\ & = & \sum_{Z \in {\cal Y}} \left (\prod_j \Pr E_j(Z) \right ) + o(1) \\ & = & \sum_{Z \in {\cal X}} \left (\prod_j \Pr E_j(Z) \right ) + o(1). \end{eqnarray*} Here the first line uses that the $d$ ancestry lists intersect somewhere with probability $o(1)$; the second lines uses that for ancestry lists in $\cal Y$ we probability of the intersection is the product of the probabilities; and the third line is again because the the collections $Z$ in ${\cal X} - {\cal Y}$ have total probability $o(1)$. Hence up to an $o(1)$ term, the behavior is the same as if the $d$ choices were independent (with respect to all bins having load at least $i$). Thus $$\mathbb{E}[X_i(t + 1/n) - X_i(t)] = (x_{i-1}(t))^d - (x_{i}(t))^d +o(1)$$ as needed. \end{proof} As a result of Lemma~\ref{lem:mainlemma}, we have the following theorem, generalizing the differential equations approach for balanced allocations to the setting of double hashing. \begin{theorem} \label{mainthm} Let $i$, $d$, and $T$ be constants. Suppose $Tn$ balls are sequentially thrown into $n$ bins with each ball having $d$ choices obtained from double hashing and each ball being placed in the least loaded bin (ties broken randomly). Let $X_i(T)$ be the number of bins of load at least $i$ after the balls are thrown. Let $x_i(t)$ be determined by the family of differential equations $$\frac{dx_i}{dt} = x_{i-1}^d - x_{i}^d,$$ where $x_0(t) = 1$ for all time and $x_i(0) = 0$ for $i \geq 1$. Then with probability $1-o(1)$, $$\frac{X_i(T)}{n} = x_i(T) + o(1).$$ \end{theorem} \begin{proof} This follows from the fact that $$\mathbb{E}[X_i(t + 1/n) - X_i(t)] = (x_{i-1}(t))^d - (x_{i}(t))^d +o(1),$$ and applying Wormald's result \cite[Theorem 1]{Wormald}. We remark that Theorem 1 of \cite{Wormald} includes other technical conditions that we briefly consider here. The first condition is that $|X_i(t + 1/n) - X_i(t)|$ is bounded by a constant; all such values here are bounded by 1. The second (and only challenging) condition exactly corresponds to our statement that $\mathbb{E}[X_i(t + 1/n) - X_i(t)] = (x_{i-1}(t))^d - (x_{i}(t))^d +o(1)$ over the course of the process. The third condition is our functions on the right hand side, that is $(x_{i-1}(t))^d - (x_{i}(t))^d$, are continuous and satisfy a Lipschitz condition on an open neighborhood containing the path of the process. These functions are continuous on the domain where all $x_i \in [0,1]$ up to the value $i$ being considered, and they satisfy the Lipschitz condition as \begin{eqnarray*} |(x_{i-1}(t))^d - (x_{i}(t))^d| \! \! & \leq & \! \! |x_{i-1}(t) - x_{i}(t)| \sum_{j=0}^{d-1} (x_{i-1}(t))^{j}(x_i(t))^{d-1-j} \\ \! \! & \leq & \! \! d|(x_{i-1}(t)) - (x_{i}(t))|, \end{eqnarray*} taking note that all $x_i,x_{i-1}$ values are in the interval $[0,1]$. Hence the conditions for Wormald's theorem are met. \end{proof} The following corollary, based on the known fact that the result of Theorem~\ref{mainthm} also holds in the setting of fully random hashing \cite{MitzenmacherThesis}, states that the difference between fully random hashing and double hashing is vanishing. \begin{corollary} \label{cormain} Let $i$, $d$, and $T$ be constants. Consider two processes, where in each $Tn$ balls are sequentially thrown into $n$ bins with each ball having $d$ choices and each ball being placed in the least loaded bin (ties broken randomly), In one process, the $d$ choices are fully random; in the other, the $d$ choices are made by double hashing. Then with probability $1-o(1)$, the fraction of bins with load $i$ differ by an $o(1)$ additive term. \end{corollary} Given the results for the differential equations, it is perhaps unsurprising that one can use these methods to obtain, for example, a maximum load of $\log \log n/ \log d+ O(1)$ maximum load for $n$ balls in $n$ bins, using the related layered induction approach of \cite{ABKU}. While we suggest this is not the main point (given Theorem~\ref{thm:vresult}), we provide further details in Appendix~\ref{sec:followon}. \section{Conclusion} We have first demonstrated empirically that using double hashing with balanced allocation processes (e.g., the power of (more than) two choices), surprisingly, does not noticeably change performance when compared with fully random hashing. We have then shown that previous methods can readily provide $O(\log \log n)$ bounds for this approach. However, explaining why the fraction of bins of load $k$ for each $k$ appears the same requires revisiting the fluid limit model for such processes. We have shown, interestingly, that the same family of differential equations applies for the limiting process. Our argument should extend naturally to other similar processes; for example, the analysis can similarly be made to apply in a straightforward fashion for the differential equations for V\"{o}cking's $d$-left scheme \cite{MV}. This opens the door to the interesting possibility that double hashing can be suitable for other problem or analyses where this type of fluid limit analysis applies, such as low-density parity-check codes \cite{LMSS}. Here, however, the asymptotic independence required was aided by the fact that we were looking at the history of the process, allowing us to tie the ancestry lists to a corresponding branching process. Whether similar asymptotic independence can be derived for other problems remains to be seen. For other problems, such as cuckoo hashing, the fluid limit analysis, while an important step, may not offer a complete analysis. Even for load balancing problems, fluid limits do not straightforwardly apply for the heavily loaded case where the number of balls is superlinear in the number of bins \cite{Steger}, and it is unclear how double hashing performs in that setting. So again, determining more generally where double hashing can be used in place of fully random hashing without significantly changing performance may offer challenging future questions. \section*{Acknowledgments} The author thanks George Varghese for the discussions which led to the formulation of this problem, and thanks Justin Thaler for both helpful conversations and offering several suggestions for improving the presentation of results. \bibliographystyle{plain}

https://arxiv.org/abs/2203.03772

The product structure of squaregraphs

A squaregraph is a plane graph in which each internal face is a $4$-cycle and each internal vertex has degree at least 4. This paper proves that every squaregraph is isomorphic to a subgraph of the semi-strong product of an outerplanar graph and a path. We generalise this result for infinite squaregraphs, and show that this is best possible in the sense that "outerplanar graph" cannot be replaced by "forest".

\section{Introduction} \label{Introduction} \footnotetext[3]{School of Mathematics, Monash University, Melbourne, Australia (\texttt{\{robert.hickingbotham,david.wood\}@monash.edu}). Research of R.H.\ supported by an Australian Government Research Training Program Scholarship. Research of D.W.\ supported by the Australian Research Council.} \footnotetext[4]{Institute of Theoretical Informatics, Karlsruhe Institute of Technology, Germany (\texttt{\{paul.jungeblut,laura.merker2\}@kit.edu}).} \renewcommand{\thefootnote}{\arabic{footnote}} A \defn{squaregraph} is a plane graph\footnote{A \defn{plane graph} is a graph embedded in the plane with no crossings. The word `face' refers to the subgraph on the boundary of the face. A graph is \defn{outerplanar} if it is isomorphic to a plane graph where every vertex is on the outer-face.} in which each internal face is a $4$-cycle and each internal vertex has degree at least $4$. These graphs were introduced in 1973 by \citet{SZP73}. They have many interesting structural and metric properties. For example, \citet{BCE10} showed that squaregraphs are median graphs and are thus partial cubes, and that every squaregraph can be isometrically embedded\footnote{A graph $H$ can be \defn{isometrically embedded} into a graph $G$ if there exists an isomorphism $\phi$ from $V(H)$ to a subgraph of $G$ such that $\dist_H(u,v)=\dist_G(\phi(u),\phi(v))$ for all $u,v\in V(H)$.} into the cartesian product\footnote{The following are the standard graph products. For graphs $ G $ and $ H $, the \defn{cartesian product} $ G \boxempty H $ is the graph with vertex-set $ V(G) \times V(H) $ with an edge between two vertices $ (v,w) $ and $ (v',w') $ if $ v=v' $ and $ ww' \in E(H) $, or $ w=w' $ and $ vv' \in E(G) $. The \defn{direct product} $ G \times H $ is the graph with vertex-set $ V(G) \times V(H) $ with an edge between two vertices $ (v,w) $ and $ (v',w') $ if $ vv' \in E(G) $ and $ ww' \in E(H) $. The \defn{strong product} $ G \boxtimes H := (G\boxempty H)\cup (G\times H)$. } of five trees. See the survey by \citet{BC08} for background on metric graph theory. The primary contribution of this paper is the following product structure theorem for squaregraphs, as illustrated in \cref{fig:squaregraph-product}. For graphs $G$ and $H$, the \defn{semi-strong product} \defn{$ G \Bow H $} is the graph with vertex-set $ V(G) \times V(H) $ with an edge between two vertices $ (v,w) $ and $ (v',w') $ if $ v=v' $ and $ ww' \in E(H) $, or $ vv' \in E(G) $ and $ ww' \in E(H) $; see \citep{GRW76,HLL21} for example. Note that \[G \times H \,\sse\, G \,\Bow\, H \,\sse\, G \boxtimes H.\] We write \defn{$H \subsetsim G$} to mean that $H$ is isomorphic to a subgraph of $G$. \begin{restatable}{thm}{squaregraphs} \label{squaregraphs} For every squaregraph $G$ there is an outerplanar graph $H$ and a path $P$ such that $G\subsetsim H\Bow P$. \end{restatable} Note that since a path is bipartite, $H \Bow P$ is also bipartite. \begin{figure}[h] \centering \includegraphics{product} \caption A squaregraph $ G $ (left) isomorphic to a subgraph of the semi-strong product $ H \Bow P $ of an outerplanar graph $ H $ and a path $P$ (right). } \label{fig:squaregraph-product} \end{figure} We in fact prove a more general sufficient condition for a plane graph to have such a product structure which implies \cref{squaregraphs}; see \cref{srtw2-bfs} in \cref{SectionUB}. The second contribution of this paper is to show that \cref{squaregraphs} is best possible in the sense that ``outerplanar graph'' cannot be replaced by ``forest''. Moreover, this lower bound holds for strong products. In fact, we prove that for every integer $\ell\in\mathbb{N}$ there is a squaregraph $G$ such that for any graph $H$ and path $P$, if $G\subsetsim H\boxtimes P\boxtimes K_\ell$ then $H$ contains a cycle (and is therefore not a forest). This result actually follows from a stronger lower bound for bipartite graphs, which has other interesting consequences; see \cref{BipartiteLower} in \cref{SectionLB}. Also note that \cref{squaregraphs} cannot be strengthened by replacing ``outerplanar graph'' by ``graph with bounded pathwidth''. Indeed, \citet{BDJMW} showed that for every $k \in\mathbb{N}$ there is a tree $T$ (which is a squaregraph) such that for any graph $H$ and path $P$, if $T\subsetsim H \boxtimes P$ then $\pw(H)\geq k$. In \cref{squaregraphs} it is natural to ask whether there is such an outerplanar graph $H$ independent of $G$. This leads to the study of infinite squaregraphs, previously investigated by \citet{BCE10}. Our final contribution is an extension of \cref{squaregraphs} in which we show that every (possibly infinite) squaregraph is isomorphic to a subgraph of $O\Bow \overrightarrow{P}$, where $O$ is the universal outerplanar graph and $\overrightarrow{P}$ is the 1-way infinite path; see \cref{Infinite}. Before proving the above results, we provide further motivation by putting \cref{squaregraphs} in context. The study of the product structure of graph classes emerged with the following seminal result by \citet{DJMMUW20}, now called the \emph{Planar Graph Product Structure Theorem}. This result describes planar graphs in terms of the strong product of graphs with bounded treewidth\footnote{A \defn{tree-decomposition} of a graph $G$ is a collection $(B_x\subseteq V(G):x\in V(T))$ of subsets of $V(G)$ (called \defn{bags}) indexed by the nodes of a tree $T$, such that (a) for every edge $uv\in E(G)$, some bag $B_x$ contains both $u$ and $v$, and (b) for every vertex $v\in V(G)$, the set $\{x\in V(T):v\in B_x\}$ induces a non-empty subtree of $T$. The \defn{width} of a tree-decomposition is the size of the largest bag minus~$1$. The \defn{treewidth} of a graph $G$, denoted by \defn{$\tw(G)$}, is the minimum width of a tree-decomposition of $G$. A \defn{path-decomposition} of a graph $G$ is a tree decomposition $(B_x\subseteq V(G):x\in V(T))$ where $T$ is a path. The \defn{pathwidth} of a graph $G$, denoted by \defn{$\pw(G)$}, is the minimum width of a path-decomposition of $G$.} and a path. A connected graph has treewidth at most 1 if and only if it is a tree. Treewidth measures how similar a graph is to a tree and is an important parameter in algorithmic and structural graph theory; see \cite{HW17,Reed97}. Graphs with bounded treewidth are considered to be a relatively simple class of graphs. \begin{thm}[\cite{UWY,DJMMUW20}]\label{PGPST} For every planar graph $G$ there is a graph $H$ of treewidth at most $6$ and a path $P$ such that $G\subsetsim H \boxtimes P$. \end{thm} The original version of the Planar Graph Product Structure Theorem by \citet{DJMMUW20} had ``treewidth at most $8$'' instead of ``treewidth at most 6''. \citet{UWY} proved \cref{PGPST} with ``treewidth at most $6$''. Since outerplanar graphs have treewidth at most $2$, \Cref{squaregraphs} is stronger than \cref{PGPST} in the case of squaregraphs. \Cref{squaregraphs} is also stronger than \cref{PGPST} in the sense that \Cref{squaregraphs} uses $\Bow$ whereas \cref{PGPST} uses $\boxtimes$. That said, as explained in \cref{Preliminaries}, it is well-known that in the case of bipartite planar graphs $G$, the proof of \cref{PGPST} can be adapted to show that $G\subsetsim H\Bow P$. Product structure theorems are useful since they reduce problems on a complicated class of graphs (such as planar graphs or squaregraphs) to a simpler class of graphs (bounded treewidth graphs such as outerplanar graphs). They have been the key tool to resolve several open problems regarding queue layouts~\citep{DJMMUW20}, nonrepetitive colourings~\citep{DEJWW20}, centered colourings~\citep{DFMS21}, clustered colourings~\citep{DEMWW22}, adjacency labellings~\citep{BGP20,DEJGMM21,EJM}, vertex rankings~\citep{BDJM}, twin-width~\citep{BKW}, odd colourings~\citep{DMO}, and infinite graphs \cite{HMSTW}. Similar product structure theorems are known for other classes including graphs with bounded Euler genus~\cite{DJMMUW20,DHHW}, apex-minor-free graphs~\cite{DJMMUW20}, $(g,d)$-map graphs~\cite{DMW}, $(g,\delta)$-string graphs~\cite{DMW}, $(g,k)$-planar graphs~\cite{DMW}, powers of planar graphs~\cite{DMW,HW21b}, $k$-semi-fan-planar graphs~\cite{HW21b} and $k$-fan-bundle planar graphs~\cite{HW21b}. \subsection{Preliminaries} \label{Preliminaries} We consider undirected simple graphs $G$ with vertex-set $V(G)$ and edge-set $E(G)$. Unless stated otherwise, graphs are finite. Undefined terms and notation can be found in Diestel's textbook~\citep{Diestel5}. For $m,n \in \mathbb{Z}$ with $m \leq n$, let $[m,n]:=\{m,m+1,\dots,n\}$ and $[n]:=[1,n]$. Let $P_n$ denote a path on $n$ vertices. For graphs $G$ and $H$, the \defn{complete join $G+H$} is the graph obtained by the disjoint union of~$G$ and~$H$ by adding all edges between~$G$ and~$H$. For a graph $G$ with $A,B\sse V(G)$, let \defn{$G[A,B]$} be the subgraph of $G$ with $V(G[A,B]):=A \cup B$ and $E(G[A,B]):=\{uv \in E(G):u \in A, v\in B\}$. A \defn{matching} $M$ in a graph $G$ is a set of edges in $G$ such that no two edges in $M$ have a common endvertex. A matching $M$ \defn{saturates} a set $S\sse V(G)$ if every vertex in $S$ is incident to some edge in $M$. A \defn{model} of $H$ in $G$ is a function $\mu$ with domain $V(H)$ such that: $\mu(v)$ is a connected subgraph of $G$; $\mu(v)\cap \mu(w)=\emptyset$ for all distinct $v,w\in V(H)$; and $\mu(v)$ and $\mu(w)$ are adjacent for every edge $vw \in E(H)$. If, for some $s \in \mathbb{N}_0$, there is a model $\mu$ of $H$ in $G$ such that $|V(\mu(v))|\leq s$ for each $v \in V(H)$, then $H$ is an \defn{$s$-small minor} of $G$. In a plane graph $G$, a vertex is \defn{outer} if it is on the outer-face of $G$ and is \defn{inner} otherwise. Let \defn{$I_G$} denote the set of inner vertices in $G$. Let $G$ be a graph. A \defn{partition} of $G$ is a set $\Pcal$ of sets of vertices in $G$ such that each vertex of $G$ is in exactly one element of $\Pcal$. Each element of $\Pcal$ is called a \defn{part}. The \defn{quotient} of $\Pcal$ (with respect to $G$) is the graph, denoted by \defn{$G/\Pcal$}, with vertex set $\Pcal$ where distinct parts $A,B\in \mathcal{P}$ are adjacent in $G/\Pcal$ if and only if some vertex in $A$ is adjacent in $G$ to some vertex in $B$. An \defn{$H$-partition} of $G$ is a partition $\Pcal=(A_x:x \in V(H))$ where $H\cong G/\Pcal$. For an $H$-partition $(A_x:x\in V(H))$ of $G$, for each subgraph $J\sse G$ the quotient $\tilde{H}$ of the partition $(A_x\cap V(J):x\in V(H),A_x\cap V(J)\neq\emptyset)$ is called the \defn{sub-quotient} for $J$. Note that $\tilde{H}$ is a subgraph of $H$. A \defn{layering} of a graph $G$ is an ordered partition $\mathcal{L}:=(L_0,L_1,\dots)$ of $V(G)$ such that for every edge $vw \in E(G)$, if $v \in L_i$ and $w \in L_j$, then $|i-j|\leq 1$. $\mathcal{L}$ is a \defn{\textsc{bfs}-layering} (of $G$) if $L_0 = \{r\}$ for some \defn{root vertex} $r\in V(G)$ and $L_i=\{v\in V(G):\dist_G(v,r)=i\}$ for all $i\geq 1$. A path $P$ is \defn{vertical} (with respect to $\mathcal{L}$) if $|V(P)\cap L_i|\leq 1$ for all $i\geq 0$. A \defn{layered partition} $(\Pcal,\mathcal{L})$ of a graph $G$ consists of a partition $\Pcal$ and a layering~$\mathcal{L}$ of $G$. If $\Pcal$ is an $H$-partition, then $(\Pcal,\mathcal{L})$ is a \defn{layered $H$-partition}. If $\Pcal=(A_x:x\in V(H))$, then the \defn{width} of $(\Pcal,\mathcal{L})$ is $\max\{|A_x\cap L|:x\in V(H), L \in \mathcal{L}\}$. Layered partitions of width at most $1$ are \defn{thin}. Layered partitions were introduced by \citet{DJMMUW20} who observed the following connection to strong products (which follows directly from the definitions). \begin{obs}[\cite{DJMMUW20}]\label{OrthogonalPartitions} For all graphs $G$ and $H$, $G \subsetsim H\boxtimes P\boxtimes K_{\ell}$ for some path $P$ if and only if $G$ has a layered $H$-partition $(\Pcal,\mathcal{L})$ with width at most $\ell$. \end{obs} We have the following analogous observation for $\Bow$ (which also follows directly from the definitions). \begin{obs}\label{BowPartitions} For all graphs $G$ and $H$, $G \subsetsim (H\boxtimes K_\ell) \Bow P$ for some path $P$ if and only if $G$ has a layered $H$-partition $(\Pcal,\mathcal{L})$ with width at most $\ell$, such that each $L \in \mathcal{L}$ is an independent set in $G$. \end{obs} In \cref{BowPartitions} we may use $G \subsetsim (H \boxtimes K_{\ell}) \Bow P $ instead of $G \subsetsim H\boxtimes K_{\ell} \boxtimes P$ when each $L\in\Lcal$ is an independent set, since no edges in $G$ correspond to edges in $H\boxtimes K_{\ell} \boxtimes P$ of the form $ (v,x,w)(v',y,w) $ where $vv'\in E(H)$, $x,y\in V(K_{\ell})$ and $w \in V(P)$. As mentioned in \cref{Introduction}, it is well-known that in the case of bipartite planar graphs $G$, the proof of \cref{PGPST} can be adapted to show that $G\subsetsim H\Bow P$ for some graph $H$ of treewidth at most $6$ and for some path $P$. To see this, we may assume that $G$ is edge-maximal bipartite planar. Thus $G$ is connected, and each face is a 4-cycle. Let $\Lcal=(L_0,L_1,\dots)$ be a \textsc{bfs}-layering of $G$. So each $L_i$ is an independent set. Each face can be written as $(a,b,c,d)$ where $a\in L_i$ and $b,d\in L_{i+1}$ and $c\in L_i \cup L_{i+2}$, for some $i\geq 0$. Let $G'$ be the planar triangulation obtained from $G$ by adding the edge $bc$ across each such face. Thus $(L_0,L_1,\dots)$ is a layering of $G'$. The proof of \cref{PGPST} shows that $G'$ has a partition $\Pcal$ such that $\tw(G/\Pcal)\leq 6$ and $(\Pcal,\Lcal)$ is a thin layered partition. By construction, $(\Pcal,\Lcal)$ is a layered partition of $G$. By \cref{BowPartitions}, $G \subsetsim H \Bow P$. A \defn{red-blue colouring} of a bipartite graph $G$ is a proper vertex $2$-colouring of $G$ with colours `red' and `blue'. \section{Sufficient Conditions} \label{SectionUB} In this section we prove \cref{squaregraphs}. We first prove the following, more general sufficient condition for a plane graph to be isomorphic to a subgraph of the strong or semi-strong product of an outerplanar graph and a path. Afterwards, we show that this more general result implies \cref{squaregraphs}. \begin{thm}\label{srtw2-bfs} Let~$G$ be a plane graph with inner vertices~$I_G$. If $ G $ has a layering $\mathcal{L}= (L_0,L_1,\dots)$ such that $G[L_{i - 1},L_i]$ has a matching saturating $L_{i - 1}\cap I_G$ for each $i \in [n]$, then $G \subsetsim H \boxtimes P$ for some outerplanar graph $H$ and path $P$. Moreover, if $V(L_i)$ is an independent set for all $L_i \in \mathcal{L}$, then $G \subsetsim H \Bow P$. \end{thm} \begin{proof} By \cref{OrthogonalPartitions,BowPartitions}, it suffices to show that $G$ has a thin layered $H$-partition $\Pcal$ (with respect to $\mathcal{L}$) for some outerplanar graph $H$. For each $i \in [n]$, let $E_{i}$ be a matching in $G[L_{i-1},L_i]$ that saturates $ L_{i - 1}\cap I_G$. For vertices $u \in L_{i-1}$ and $v \in L_i$ and an edge $uv\in E_{i}$, we say that $u$ is the \defn{parent} of $v$ and $v$ is the \defn{child} of $u$. Observe that each vertex $u \in L_{i-1}\cap I_G$ has exactly one child and each vertex $v \in L_i$ has at most one parent. Let $J$ be the subgraph of $G$ where $V(J)=V(G)$ and $E(J)=\bigcup_{i \in [n]} E_i$. Let $X$ be a connected component of $J$. Choose the maximum $j \in [0,n]$ such that there exists some vertex $v \in V(X)\cap L_j$. Vertex~$v$ must be outer because each vertex in $L_j \cap I_G$ is adjacent in~$J$ to some vertex in~$L_{j+1}$. As illustrated in \cref{fig:path-contraction}, since each vertex in $X$ has at most one child and at most one parent, $X$ is a vertical path with respect to $\mathcal{L}$. \begin{figure}[!ht] \centerin \includegraphics[page = 3]{path-contraction-leveled} \qquad \includegraphics[page = 6]{path-contraction-leveled} \caption Left: A squaregraph with a \textsc{bfs}-layering and a partition $ \Pcal $ into vertical paths (thick orange). The vertical paths are constructed from matchings between consecutive layers, where the leftmost vertex in $ L_i $ is chosen for each inner vertex in $ L_{i-1} $. Right: The lower endpoint of each path is on the outer-face, so when each path is contracted we obtain an outerplanar graph. } \label{fig:path-contraction} \end{figure} Let $\Pcal$ be the partition of $G$ determined by the connected components of $J$. Let $H=G/\Pcal$ be the quotient of $\Pcal$. Since each part in $\Pcal$ is a vertical path with respect to $\mathcal{L}$, it follows that $(\Pcal,\mathcal{L})$ is a thin layered $H$-partition. It remains to show that $H$ is outerplanar. Since each part in $\Pcal$ is connected, $H$ is a minor of $G$ and is therefore planar. Since each part of $\Pcal$ contains a vertex on the outer-face, contracting each part of $\Pcal$ into a single vertex gives a plane embedding of $H$ with each vertex on the outer-face; see \cref{fig:path-contraction}. Therefore $H$ is outerplanar. \end{proof} We now work towards showing that squaregraphs satisfy the conditions for \cref{srtw2-bfs}. A plane graph $G$ is \defn{leveled} if the edges are straight line-segments and vertices are placed on a sequence of horizontal lines, $(L_0,L_1,\dots)$, called \defn{levels}, such that each edge joins two vertices in consecutive levels. If, in addition, we allow straight-line edges between consecutive vertices on the same level, then $G$ is \defn{weakly leveled}. Observe that the levels in a weakly leveled plane graph $G$ define a layering of $G$. Leveled plane graphs were first introduced by \citet{STT81}, and have since been well studied \cite{BDDEW19}. For a weakly leveled plane graph $G$ with levels $(L_0,L_1,\dots)$ and a vertex $v\in L_i$, the \defn{up-degree} of $v$ is $|N_G(v)\cap L_{i-1}|$ and the \defn{down-degree} of $v$ is $|N_G(v)\cap L_{i+1}|$. We now give a more natural condition that forces our desired matching between two consecutive levels. \begin{lem}\label{WeaklyLeveledUp} Let~$G$ be a weakly leveled plane graph with inner vertices $I_G$. If each vertex in $I_G$ has down-degree at least $2$, then $G \subsetsim H \boxtimes P$ for some outerplanar graph $H$ and path $P$. Moreover, if $G$ is a leveled plane graph, then $G \subsetsim H \Bow P$. \end{lem} \begin{proof} Let $(L_0,L_1,\dots)$ be the levels of $G$. Observe that if $G$ is a leveled plane graph, then $V(L_i)$ is an independent set for all $i\geq 0$. For each $ i \in [n]$, let $E_i$ be the set of edges in $G[L_{i - 1},L_i]$ between each vertex $v \in L_{i-1}\cap I_G$ and its leftmost neighbour in $ L_i $; see \cref{fig:path-contraction}. For the sake of contradiction, suppose there exists a vertex $u\in L_{i - 1}\cup L_i$ that is incident to two edges in $E_i$. By construction, each vertex in $L_{i-1}\cap I_G$ is incident to at most one edge in $E_i$ so $u\in L_i$. Let $x$ and $y$ be the neighbours of $u$ in $L_{i-1}$, where $x$ is to the left of $y$. Since $x$ has down-degree at least $2$, $x$ is adjacent to a vertex $v$ that is to the right of $u$. However, this contradicts $G$ being weakly leveled plane since $uy$ and $vx$ cross; see \cref{fig:LevelCrossing}. Therefore, $E_i$ is a matching that saturates $L_{i-1}\cap I_G$. The claim therefore follows by \cref{srtw2-bfs}. \end{proof} \begin{figure}[h!] \centering \includegraphics[width=0.21\textwidth]{LevelCrossing2} \caption{Contradiction in the proof of \Cref{WeaklyLeveledUp}.} \label{fig:LevelCrossing} \end{figure} We are ready to prove \cref{squaregraphs} which we restate here for convenience. \squaregraphs* \begin{proof} We may assume that $G$ is connected (since if each component of $G$ has the desired product structure, then so does $G$). By taking a \textsc{bfs}-layering of $G$ rooted at any vertex $r$ on the outer-face, \citet{BDDEW19} showed that $G$ is isomorphic to a leveled plane graph. Without loss of generality, assume $G$ is leveled plane with corresponding levels $(L_0,L_1,\dots)$. Below we show that every inner vertex in $G$ has up-degree at most 2. Since each inner vertex has degree at least $4$, each inner vertex has down-degree at least $2$. The result thus follows from \cref{WeaklyLeveledUp}. For the sake of contradiction, suppose there exists an inner vertex with up-degree at least $3$. Let $ i \in [n] $ be minimum such that there is a vertex $ v \in L_i\cap I_G $ with up-degree at least $3$. Let $ u_1, u_2, u_3 $ be neighbours of $v$ in $L_{i-1}$ ordered left to right. Since the levels are defined by a \textsc{bfs}-layering, there is an $(u_1,r)$-path and an $(u_3,r)$-path that does not contain $ u_2 $; see \cref{fig:three-parents}. Hence, $ u_2 $ is an inner vertex of $G$ and thus has degree at least $4$. However, by planarity, $ v $ is the only neighbour of $ u_2 $ in $ L_i $. Since $ u_2 $ has no neighbours in $L_{i-1}$ (as $G$ is leveled plane), $u_2$ has three neighbours in $ L_{i - 2} $, which contradicts the minimality of $ i $, as required. \end{proof} \begin{figure}[!ht] \centering \includegraphics{three-parents} \caption{Vertex $ v \in L_i $ with three neighbours $ u_1, u_2, u_3 $ in the preceding layer $ L_{i - 1} $. Since $ u_2 $ is an inner vertex, it has degree at least 4.} \label{fig:three-parents} \end{figure} We now give an application of \cref{squaregraphs}. A colouring $\phi$ of a graph $G$ is \defn{nonrepetitive} if for every path $v_1,\dots,v_{2h}$ in $G$, there exists $i \in [h]$ such that $\phi(v_i)\neq \phi(v_{i+h})$. The \defn{nonrepetitive chromatic number}, $\pi(G)$, is the minimum number of colours in a nonrepetitive colouring of $G$. Nonrepetitive colourings were introduced by \citet{AGHR02} and have since been widely studied; see the survey \citep{Wood21}. \citet{KP-DM08} showed that $\pi(G)\leq 4^{\tw(G)}$ for every graph $G$. Building upon this result, \citet{DEJWW20} proved the following: \begin{lem}[\citep{DEJWW20}]\label{NonrepProduct} For any graph $H$ and path $P$, if $G \subsetsim H \boxtimes P$ then $\pi(G)\leq 4^{\tw(H)+1}$. \end{lem} Using (a variation of) \cref{PGPST,NonrepProduct}, \citet{DEJWW20} resolved a long-standing conjecture of \citet{AGHR02} by showing that planar graphs $G$ have bounded nonrepetitive chromatic number; in particular, $\pi(G) \leq 768$. When $G$ is a squaregraph, \Cref{squaregraphs,NonrepProduct} imply that $\pi(G)\leq 4^3=64$. \section{Tightness}\label{SectionLB} In this section, we show that \cref{squaregraphs} is tight by proving a lower bound for the product structure of bipartite graphs. The \defn{row treewidth} of a graph $G$ is the minimum integer $k$ such that $G\subsetsim H \boxtimes P$ for some graph $H$ with treewidth $k$ and path $P$ \cite{BDJMW}. \cref{PGPST} says that every planar graph has row treewidth at most $6$. \citet{DJMMUW20} showed that the maximum row treewidth of planar graphs is at least $3$. They in fact proved the following stronger result. \begin{thm}[\cite{DJMMUW20}] \label{rtwLB} For all $k,\ell\in\mathbb{N}$ with $k\geq 2$ there is a graph $G$ with pathwidth $k$ such that for any graph $H$ and path $P$, if $G \subsetsim H \boxtimes P \boxtimes K_{\ell}$ then $K_{k+1}\subsetsim H$ and thus $H$ has treewidth at least $k$. Moreover, if $k=2$ then $G$ is outerplanar, and if $k=3$ then $G$ is planar. \end{thm} \cref{squaregraphs} says that squaregraphs have row treewidth at most 2. We show that this bound is tight by proving \cref{BipartiteLower} which is an analogous result to \cref{rtwLB} for bipartite graphs. As an introduction to the key ideas in the proof of \cref{BipartiteLower}, we first establish \cref{LowerBoundSubgraph} which is a slight generalisation of \cref{rtwLB}. We need the following lemma for finding long paths in quotient graphs. \begin{lem}\label{lem:BaseCaseLB} Let $ G $ be a graph and $(A_x : x\in V(H))$ be an $H$-partition of $G$ such that $|A_x|\leq a$ for all $x \in V(H)$. Then for every $w \in V(H)$ and $ n \in \N $, there is a sufficiently large $ n' \in \N $ such that if $ G $ contains a path on $ n' $ vertices, then $ H-w$ contains a path on $ n $ vertices. \end{lem} \begin{proof} Let $m$ be sufficiently large compared to $n$ and let $n':=(a+1)am+a$. Suppose $G$ has a path on $n'$ vertices. Let $G'=G-A_w$. Since $|V(P)\cap A_w|\leq a$, $P$ is split into at most $a + 1$ disjoint subpaths in $G'$. Thus, there is a path $P_{\max}$ in $G'$ with at least $am$ vertices. Let $\tilde{H}$ be the sub-quotient of $H$ with respect to $P_{\max}$. Observe that $\tilde{H}$ is connected and that $|V(\tilde{H})|\geq am/a = m $. Moreover, $\tilde{H}\sse H-w$ since $A_w\cap V(P_{\max})=\emptyset$. Now $\tilde{H}$ has maximum degree at most $2a$ since every vertex in $P_{\max}$ has degree at most~$2$. Thus, since $m$ is sufficiently large, $\tilde{H}$ contains a path on at least $n$ vertices, as required. \end{proof} The following result generalises \cref{rtwLB} (which is the $n= 2$ case). \begin{prop}\label{LowerBoundSubgraph} For all $k, \ell,n \in \N$ there exists a graph $G$ with pathwidth at most $k+1$ such that for any graph $H$ and path $P$, if $G \subsetsim H \boxtimes P \boxtimes K_{\ell}$ then $P_n+K_k\subsetsim H$. \end{prop} \begin{proof} We proceed by induction on $k\geq 1$. Let $ n' $ be sufficiently large compared to $n$. Let~$G^{(1)}$ be the graph obtained from a path on $n'$ vertices plus a dominant vertex $v$. Observe that $ G^{(1)} $ has radius 1 and pathwidth at most $2$. Suppose $ G^{(1)} \subsetsim H \boxtimes P \boxtimes K_{\ell}$ for some graph~$H$ and path~$P$. By \cref{OrthogonalPartitions}, there is a layered $H$-partition $(A_x : x \in V(H))$ of~$G$ of width~$\ell$. Let $w\in V(H)$ be such that $v \in A_w$. Since $G^{(1)}$ has radius $1$, every layering of $G^{(1)}$ consists of at most three layers so $|A_x|\leq 3\ell$ for all $x \in V(H)$. By \cref{lem:BaseCaseLB} and since $n'$ is sufficiently large, $ H-w $ contains a path on $ n $ vertices. As~$v$ is dominant in $ G^{(1)} $, $w$ is also dominant in $ H $. Thus $P_n+K_1 \subsetsim H$. Now suppose $k > 1$ and let $G^{(k-1)}$ be a graph that satisfies the induction hypothesis for $k-1$. Let~$G^{(k)}$ be obtained by taking $3\ell$ disjoint copies of~$G^{(k-1)}$ plus a dominant vertex~$v$. Then $G^{(k)}$ has pathwidth at most $k+1$. As in the base case, let $(A_x : x \in V(H))$ be a layered~$H$-partition of~$G^{(k)}$ of width~$\ell$. Let $w \in V(H)$ be such that $v \in A_w$. Since $G^{(k)}$ has radius $1$, it follows that $|A_x-\{v\}|\leq 3\ell-1$ for all $x \in V(H)$. Thus, there is a copy of $G^{(k-1)}$ that contains no vertices from~$A_w$. Now consider the sub-quotient $\tilde{H}$ of $H$ with respect to this copy of $G^{(k-1)}$. By induction, $P_n+K_{k-1}\subsetsim \tilde{H}$. Since~$v$ is dominant in $G^{(k)}$, $w$ is dominant in $H$ and thus $P_n+K_k\subsetsim H$, as required. \end{proof} Note that in \cref{LowerBoundSubgraph}, the graph $G^{(1)}$ is outerplanar and the graph $G^{(2)}$ is planar for every $n \in \N$. We now prove our main lower bound which is a bipartite version of \cref{LowerBoundSubgraph}. \begin{thm} \label{BipartiteLower} For all $i,j,k, \ell,n \in \N$ where $i+j=k$, there exists a bipartite graph $G$ with pathwidth at most $k+1$ such that for any graph $H$ and path $P$, if $G \subsetsim H \boxtimes P \boxtimes K_{\ell}$ then $P_n+K_{i,j}$ is a $2$-small minor of $H$. \end{thm} \begin{proof} Let $P_n=(a_1,\dots,a_n)$ be a path on $n$ vertices. Let $B=\{b_1,\dots,b_i\}$ and $C=\{c_1,\dots,c_j\}$ be the bipartition of $V(K_{i,j})$. We proceed by induction on $k$ with the following hypothesis: for every $i,j,k, \ell,n \in \N$ where $i+j=k$, there exists a red-blue coloured connected bipartite graph $G$, such that for any graph $H$, if $(A_x : x \in V(H))$ is a layered $H$-partition of $G$ of width $\ell$, then $H$ contains a model $\mu$ of $P_n+K_{i,j}$ such that for each $u \in V(P_n+K_{i,j})$ we have $|V(\mu(u))|\leq 2$ and $\bigcup_{a \in V(\mu(u))}A_a$ contains: \begin{enumerate} \item a red vertex when $u\in B$; \item a blue vertex when $u \in C$; and \item a red and a blue vertex when $u \in V(P_n)$. \end{enumerate} The claimed theorem follows by \cref{OrthogonalPartitions}. For~$k = 1$ we may assume that $i=1$ and $j=0$. Let $ n' $ be sufficiently large and let~$G^{(1,0)}$ be the bipartite graph obtained from a red-blue coloured path $P_G=(u_1,\dots,u_{n'})$ on $n'$ vertices plus a red vertex $v$ adjacent to all the blue vertices in $V(P_G)$. Observe that $ G^{(1,0)} $ has radius 2 and pathwidth at most $2$. Let $(A_x : x \in V(H))$ be a layered $H$-partition of~$G^{(1,0)}$ of width~$\ell$. Let $w\in V(H)$ be such that $v \in A_w$. Then $A_w$ contains a red vertex. Since $G^{(1,0)}$ has radius $2$, every layering of $G^{(1,0)}$ has at most five layers, so $|A_x|\leq 5\ell$ for all $x \in V(H)$. By \cref{lem:BaseCaseLB} and since $n'$ is sufficiently large, $ H-w $ contains a path $P_H=(a_1',\dots,a_{2n}')$ on $ 2n $ vertices. Now for every edge $a'_i a'_{i+1}\in E(P_H)$, there exists $j \in [n'-1]$ such that $u_j,u_{j+1}\in A_{a_i'}\cup A_{a_{i+1}'}$. As such, $A_{a_i'}\cup A_{a_{i+1}'}$ contains a red and a blue vertex. For all $i \in [n]$, let $\mu(a_i)=H[\{a_{2i-1}',a_{2i}'\}]$ and $\mu(b_1)=\{w\}$. Then $\mu$ is a model of $P_n+K_{1,0}$ in $H$ which satisfies the induction hypothesis. Now suppose $k > 1$ and that there is a red-blue coloured connected bipartite graph $G^{(i-1,j)}$ such that for any graph $H$, if $(A_x : x \in V(H))$ is a layered $H$-partition of $G$ of width $\ell$, then $H$ contains a model $\tilde{\mu}$ of $P_n+K_{i-1,j}$ where $|V(\tilde{\mu}(u))|\leq 2$ for all $u \in V(P_n+K_{i-1,j})$ and $\bigcup_{a \in V(\mu(u))}A_a$ contains a red vertex when $u\in B$; a blue vertex when $u \in C$; and a red and a blue vertex when $u \in V(P_n)$. Let~$G^{(i,j)}$ be obtained by taking $5\ell$ copies of~$G^{(i-1,j)}$ plus a red vertex~$v$ that is adjacent to all the blue vertices. Then $G^{(i,j)}$ has radius $2$ and pathwidth at most $k+1$. As in the base case, let $(A_x : x \in V(H))$ be a layered~$H$-partition of~$G^{(i,j)}$ of width~$\ell$. Let $w \in V(H)$ be such that $v \in A_w$. Then $A_w$ contains a red vertex. Since $G^{(i,j)}$ has radius $2$, $|A_w-\{v\}|\leq 5\ell-1$. Thus, there is a copy of $G^{(i-1,j)}$ that contains no vertices from~$A_w$. Now consider the sub-quotient $\tilde{H}$ of $H$ with respect to this copy of $G^{(i-1,j)}$. By induction, $\tilde{H}$ contains a model $\tilde{\mu}$ which satisfies the induction hypothesis. Let $\mu(b_i)=\{w\}$ and $\mu(v)=\tilde{\mu}(v)$ for all $v \in V(P_n+K_{i-1,j})$. Since~$v$ is adjacent to all the blue vertices in $G$, $w$ is adjacent to a vertex in $\bigcup_{a\in V(\mu(u))}A_a$ whenever $u \in V(P_n)\cup C$. Thus $\mu$ is a model of $P_n+K_{i,j}$ in $H$ which satisfies the induction hypothesis, as required. \end{proof} \begin{figure}[!h] \centering \includegraphics[width=0.85\textwidth]{BipartiteLowerBound} \caption{The graphs $G^{(1,0)}$ and $G^{(1,1)}$ from \cref{BipartiteLower}.} \label{fig:BipartiteLower} \end{figure} We now highlight several consequences of \cref{BipartiteLower}. First, when $i=1$ and $j=0$, the graph $G^{(1,0)}$ is an outerplanar squaregraph as illustrated in \cref{fig:BipartiteLower}. Since $P_2+K_{1,0}$ is a $3$-cycle, we have the following: \begin{cor} For every $\ell \in \N$, there exists a squaregraph $G$ such that for any graph $H$ and path $P$, if $G \subsetsim H \boxtimes P \boxtimes K_{\ell}$ then $H$ contains a cycle of length at most 6. \end{cor} Thus \Cref{squaregraphs} is best possible in the sense that ``outerplanar graph'' cannot be replaced by ``forest''. Second, when $i=j=1$, the graph $G^{(1,1)}$ is a bipartite planar graph, as illustrated in \cref{fig:BipartiteLower}. Since $P_2+K_{1,1}\cong K_4$ which has treewidth $3$, we have the following: \begin{cor} For every $\ell \in \N$, there exists a bipartite planar graph $G$ such that for any graph $H$ and path $P$, if $G \subsetsim H \boxtimes P \boxtimes K_{\ell}$ then $H$ contains a 2-small minor of $K_4$ and thus $\tw(H)\geq 3$. \end{cor} Therefore, the maximum row treewidth of bipartite planar graphs is at least $3$. We conclude this section with the following open problem: what is the maximum row treewidth of bipartite planar graphs? As in the case of (non-bipartite) planar graphs, the answer is in $\{3,4,5,6\}$. \section{Infinite Squaregraphs} \label{Infinite} In this section by `graph' we mean a graph $G$ with $V(G)$ finite or countably infinite. \citet{HMSTW} showed how \cref{PGPST} can be used to construct a graph that contains every planar graph as a subgraph and has several interesting properties. Here we adapt their methods to construct an analogous graph that contains every squaregraph as a subgraph. \citet{BCE10} gave several equivalent definitions of an infinite squaregraph. The following definition suits our purposes. Let $G$ be a locally finite\footnote{A graph $G$ is \defn{locally finite} if every vertex of $G$ has finite degree.} graph. For every vertex $v$ of $G$ and every $r\in\mathbb{N}$ the subgraph $G[ \{ w\in V(G): \dist_G(v,w)\leq r\} ]$ is called a \defn{ball}. Since $G$ is locally finite, every ball is finite. An infinite graph $G$ is a \defn{squaregraph} if it is locally finite and every ball in $G$ is a squaregraph. Let $\overrightarrow{P}$ be the 1-way infinite path, which has vertex-set $\mathbb{N}_0$ and edge-set $\{ \{i,i+1\} : i \in\mathbb{N}_0 \}$. It is well known that there is a \defn{universal} outerplanar graph $O$. This means that $O$ is outerplanar and every outerplanar graph is isomorphic to a subgraph of $O$. See Theorem~4.14 in \citep{HMSTW} for an explicit definition of $O$. \begin{thm} \label{InfiniteSquaregraph} Every squaregraph is isomorphic to a subgraph of $O\Bow \overrightarrow{P}$. \end{thm} \cref{InfiniteSquaregraph} follows from \cref{squaregraphs} and the next lemma, which is an adaptation of Lemma~5.3 in \citep{HMSTW}. \begin{lem} Let $H$ be a graph. Let $G$ be a locally finite graph such that $B\subsetsim H\Bow \overrightarrow{P}$ for every ball $B$ in $G$. Then $G\subsetsim H \Bow \overrightarrow{P}$. \end{lem} \begin{proof}[Proof Sketch] Fix $v\in V(G)$. For $n\in\mathbb{N}_0$, let $V_n :=\{w\in V(G): \dist_G(v,w)=n\}$ and $G_n:=G[ V_0\cup V_1 \cup\dots \cup V_n]$. So $G_n$ is a finite ball in $G$. By assumption, $G_n \subsetsim H \Bow \overrightarrow{P}$. Let $X_n$ be the set of all thin layered $H$-partitions $(\Pcal,\mathcal{L})$ of $G_n$, such that $L$ is an independent set in $G_n$ for each $L \in \mathcal{L}$. By \cref{BowPartitions}, $X_n\neq\emptyset$. Since $G_n$ is finite and connected, $X_n$ is finite. For each $n\in\mathbb{N}$ and for each $(\Pcal,\mathcal{L})\in X_n$, if $\Pcal':= \{ Y \setminus V_n : Y \in \Pcal, Y \setminus V_n\neq\emptyset\}$ and $\Lcal':= \{ L \setminus V_n : L \in \Lcal, Y \setminus V_n\neq\emptyset\}$ then $(\Pcal',\Lcal')\in X_{n-1}$ (since $G_{n-1}$ is connected). By K\H{o}nig's Lemma, there is an infinite sequence $(\Pcal_0,\Lcal_0), (\Pcal_1,\Lcal_1), (\Pcal_2,\Lcal_2), \dots$ where $\Pcal_{n-1}=\Pcal'_{n}$ and $\Lcal_{n-1}=\Lcal'_{n}$ for each $n\in\mathbb{N}$. By construction, $\Pcal_{n-1}$ is a `sub-partition' of $\Pcal_{n}$ and $\Lcal_{n-1}$ is a `sub-partition' of $\Lcal_{n}$. Let $\Pcal:= \bigcup_{n\in\mathbb{N}_0} \Pcal_n$ and $\Lcal:= \bigcup_{n\in\mathbb{N}_0} \Lcal_n$. Then $(\Pcal,\Lcal)$ is a thin layered $H$-partition of $G$; see \citep{HMSTW} for details. By \cref{BowPartitions}, $G\subsetsim H \Bow \overrightarrow{P}$. \end{proof} \subsection*{Acknowledgement} This research was initiated at the workshop, \emph{Geometric Graphs and Hypergraphs}, 30 August -- 3 September 2021, organised by Torsten Ueckerdt and Lena Yuditsky. Thanks to the organisers and other participants for creating a productive environment. \fontsize{11}{12} \selectfont \let\oldthebibliography=\thebibliography \let\endoldthebibliography=\endthebibliography \renewenvironment{thebibliography}[1] \begin{oldthebibliography}{#1 \setlength{\parskip}{0.3ex \setlength{\itemsep}{0.3ex }{\end{oldthebibliography}} \bibliographystyle{DavidNatbibStyle}

https://arxiv.org/abs/1310.4112

Subalgebras of the Fomin-Kirillov algebra

The Fomin-Kirillov algebra $\mathcal E_n$ is a noncommutative quadratic algebra with a generator for every edge of the complete graph on $n$ vertices. For any graph $G$ on $n$ vertices, we define $\mathcal E_G$ to be the subalgebra of $\mathcal E_n$ generated by the edges of $G$. We show that these algebras have many parallels with Coxeter groups and their nil-Coxeter algebras: for instance, $\mathcal E_G$ is a free $\mathcal E_H$-module for any $H\subseteq G$, and if $\mathcal E_G$ is finite-dimensional, then its Hilbert series has symmetric coefficients. We determine explicit monomial bases and Hilbert series for $\mathcal E_G$ when $G$ is a simply-laced finite Dynkin diagram or a cycle, in particular showing that $\mathcal E_G$ is finite-dimensional in these cases. We also present conjectures for the Hilbert series of $\mathcal E_{\tilde{D}_n}$, $\mathcal E_{\tilde{E}_6}$, and $\mathcal E_{\tilde{E}_7}$, as well as for which graphs $G$ on six vertices $\mathcal E_G$ is finite-dimensional.

\section{Introduction} The Fomin-Kirillov algebra $\mathcal E_n$ \cite{FK} is a certain noncommutative algebra with generators $x_{ij}$ for $1 \leq i < j \leq n$ that satisfy a simple set of quadratic relations. \iffalse \begin{definition}The \emph{Fomin-Kirillov algebra} $\mathcal E_n$ is the quadratic algebra (say, over $\mathbb{Q}$) with generators $x_{ij}=-x_{ji}$ for $1 \leq i < j \leq n$ with the following relations: \begin{itemize} \item $x_{ij}^2 = 0$ for distinct $i,j$; \item $x_{ij}x_{kl} = x_{kl}x_{ij}$ for distinct $i,j,k,l$; \item $x_{ij}x_{jk}+x_{jk}x_{ki}+x_{ki}x_{ij}=0$ for distinct $i,j,k$. \end{itemize} \end{definition} \fi While it was originally introduced as a tool to study the structure constants for Schubert polynomials, since then the Fomin-Kirillov algebra and its generalizations have received much attention from the perspectives of both combinatorics and algebra: see, for instance, \cite{Bazlov, FP, Kirillov, KirillovMaeno, Lenart, LenartMaeno, Majid, MPP, MS, P, Vendramin}. But despite its simple presentation, even some basic questions about $\mathcal E_n$ have eluded an answer thus far, such as whether or not it is finite-dimensional for $n \geq 6$. In order to better understand the structure of $\mathcal E_n$, we consider the following subalgebras: \[ \parbox{14cm}{for any graph $G$ on vertices $1, 2, \dots, n$, the \emph{Fomin-Kirillov algebra ${\mathcal E_G}$ of $G$} is the subalgebra of $\mathcal E_n$ generated by $x_{ij}$ for all edges $\overline{ij}$ in $G$.} \] While this definition might seem hopelessly ingenuous, our initial computations using the algebra package \texttt{bergman} \cite{bergman} revealed that, remarkably, whenever $G$ is a graph with at most five vertices, $\mathcal E_G$ has a one-dimensional top degree component, and in fact its Hilbert series has symmetric coefficients. (See the Appendix for these computations.) Our current study is therefore dedicated to an investigation of these beautiful, yet mysterious algebras. \medskip The first half of this paper is devoted to proving structural properties of Fomin-Kirillov algebras. We prove that any finite-dimensional $\mathcal E_G$ has a Hilbert series with symmetric coefficients, as well as that $\mathcal E_G$ is a free $\mathcal E_H$-module whenever $H$ is a subgraph of $G$. We also demonstrate that the Fomin-Kirillov algebras exhibit a striking amount of structure, much of which parallels Coxeter groups and nil-Coxeter algebras: for instance, we describe analogues of minimal coset representatives, descent sets, Bruhat order, and long words. The key tools to proving these facts can be derived from a braided Hopf algebra structure on $\mathcal E_n$ \cite{FP, MS}---in this context, the subalgebras $\mathcal E_G$ are \emph{(left) coideal subalgebras}. Coideal subalgebras are important objects in the study of Hopf algebras, and they sometimes possess freeness and symmetry properties analogous to the ones that we exhibit for $\mathcal E_G$ \cite{Masuoka}. There also exists a certain symmetric bilinear form on $\mathcal E_n$ whose nondegeneracy (known for $n \leq 5$ \cite{Grana}) implies that $\mathcal E_n$ is a special type of braided Hopf algebra called a Nichols algebra \cite{AS, MS}. Though Nichols algebras have been studied since \cite{Nichols}, the many parallels to Coxeter groups demonstrated here have not been observed in the literature on Nichols algebras to our knowledge. In our exposition below, we will not assume familiarity with braided Hopf algebras or Nichols algebras. We refer the reader to \cite{AS} for more details about these objects. \medskip A particularly exciting part of this study is the abundance of graphs $G$ for which $\mathcal E_G$ is finite-dimensional---a much larger class than finite-dimensional Coxeter groups---and the combinatorial mysteries awaiting discovery here. In the second half of this paper, we present results obtained so far about these finite-dimensional algebras. Specifically, we determine explicit monomial bases and Hilbert series for $\mathcal E_G$ when $G$ is a simply-laced Dynkin diagram or a cycle. Here, another surprising connection between the Fomin-Kirillov algebras and Coxeter groups appears: for $G$ a simply-laced Dynkin diagram, the dimension of $\mathcal E_G$ is equal to the dimension of the Weyl group of $G$ divided by the index of connection. In the case when $G$ is a cycle on $n$ vertices, we describe $\mathcal E_G$ explicitly as a quotient of the nil-Coxeter algebra of the affine symmetric group, showing that it is essentially a $q=0$ version of the twisted product of the group algebra of the symmetric group and the ring of coinvariants. We also present intriguing conjectures for the Hilbert series of $\mathcal E_{\tilde{D}_n}$, $\mathcal E_{\tilde{E}_6}$, and $\mathcal E_{\tilde{E}_7}$, as well as for which graphs $G$ on six vertices $\mathcal E_G$ is finite-dimensional. \medskip This paper is organized as follows. In Section 2, we introduce $\mathcal E_n$ and briefly describe some examples of $\mathcal E_G$. In Section 3, we discuss some structural properties of $\mathcal E_n$ and prove that whenever $\mathcal E_G$ is finite-dimensional, its Hilbert series has symmetric coefficients. In Section 4, we prove that when $H$ is a subgraph of $G$, $\mathcal E_G$ is a free $\mathcal E_H$-module. We also show that $\mathcal E_n$ has a tensor product decomposition with factors given by certain complementary $\mathcal E_G$. In Section 5, we discuss Coxeter groups and nil-Coxeter algebras, as well as their relationships and similarities to the subalgebras $\mathcal E_G$. In Section 6, we describe $\mathcal E_G$ for $G$ a simply-laced Dynkin diagram, computing its Hilbert series in each case. In Section 7, we describe $\mathcal E_G$ when $G$ is a cycle (that is, an affine Dynkin diagram of type $\tilde{A}_{n-1}$). Finally, we close in Section 8 with some open questions and conjectures to guide further research. We also include an Appendix containing the Hilbert series of $\mathcal E_G$ for all connected graphs $G$ on at most five vertices. \section{Preliminaries} We begin with the definition of the Fomin-Kirillov algebra \cite{FK}. \begin{defn} The \emph{Fomin-Kirillov algebra} $\mathcal E_n$ is the quadratic algebra (say, over $\mathbb Q$) with generators $x_{ij}=-x_{ji}$ for $1 \leq i < j \leq n$ with the following relations: \begin{itemize} \item $x_{ij}^2 = 0$ for distinct $i,j$; \item $x_{ij}x_{kl} = x_{kl}x_{ij}$ for distinct $i,j,k,l$; \item $x_{ij}x_{jk}+x_{jk}x_{ki}+x_{ki}x_{ij}=0$ for distinct $i,j,k$. \end{itemize} \end{defn} Let $V$ be the vector space spanned by the generators $x_{ij}$. Then $\mathcal E_n$ is a quotient of the tensor algebra $T(V) = \bigoplus_{n \geq 0} V^{\otimes n}$, the free associative algebra on the generators of $\mathcal E_n$. Since the relations are homogeneous, $\mathcal E_n$ is graded with respect to the usual degree. We will denote the degree $d$ part by $\mathcal E_n^d$. Note that $\mathcal E_n$ has another grading with respect to the symmetric group $\S_n$: define the $\S_n$-degree of $x_{ij}$ to be $\sigma_{ij} \in \S_n$, the transposition switching $i$ and $j$, and extend this $\S_n$-degree to all monomials in $T(V)$ by multiplicativity. Since each of the relations in $\mathcal E_n$ is homogeneous with respect to $\S_n$-degree, this gives an $\S_n$-grading on $\mathcal E_n$. We will write $\sigma_P$ for the $\S_n$-degree of a homogeneous element $P \in \mathcal E_n$. (We will always specify when we mean $\S_n$-degree; the use of ``degree'' unqualified will refer to the usual notion of degree.) There is a third grading related to the support of each monomial. For a monomial $m \in \mathcal E_n$, define $\Pi(m)$ to be the coarsest set partition of $[n]$ for which $i$ and $j$ lie in the same part if $x_{ij}$ or $x_{ji}$ appears in $m$. For instance, $\Pi(x_{12}x_{23}x_{45}x_{31}) = 123|45$. Then all of the relations of $\mathcal E_n$ are homogeneous with respect to $\Pi$. Note that $\Pi(m_1m_2)$ is the common coarsening of $\Pi(m_1)$ and $\Pi(m_2)$. The set of relations of $\mathcal E_n$ is also symmetric with respect to the indices. In other words, for all $\sigma \in \S_n$, there exists an automorphism of $\mathcal E_n$ given by $\sigma(x_{ij}) = x_{\sigma(i)\sigma(j)}$. Also note that $\mathcal E_n$ is isomorphic to its opposite algebra: the linear map $\operatorname{rev}\colon T(V) \to T(V)$ sending any monomial to the product of the same generators but in the reverse order preserves the set of relations and thus gives an antiautomorphism of $\mathcal E_n$. \begin{rmk} The natural category for $\mathcal E_n$ is the \emph{Yetter-Drinfeld category} over $\mathbb Q[S_n]$, which is the braided monoidal category consisting of $S_n$-graded $S_n$-modules $M = \bigoplus_{\sigma \in S_n} M_\sigma$ satisfying $\sigma(M_\pi) \subset M_{\sigma\pi\sigma^{-1}}$. \end{rmk} \subsection{Hilbert series} \label{sec-hilbert} Let $\mathcal H_n(t)$ be the Hilbert series of $\mathcal E_n$. Values of $\mathcal H_n(t)$ for small values of $n$ are given as follows. (To simplify expressions, we will often write $[k] = 1+t+t^2+\cdots+t^{k-1}$.) \begin{align*} \mathcal H_1(t) =&\; 1\\ \mathcal H_2(t) =&\; [2]\\ \mathcal H_3(t) =&\; [2]^2[3]\\ \mathcal H_4(t) =&\; [2]^2[3]^2[4]^2\\ \mathcal H_5(t) =&\; [4]^4[5]^2[6]^4\\ \mathcal H_6(t) =&\; 1+15t+125t^2+765t^3+3831t^4+16605t^5+64432t^6\\ &+228855t^7+755777t^8+2347365t^9+6916867t^{10}+\cdots \end{align*} A closed form for $\mathcal H_n$ is not known for $n \geq 6$. It is not even known whether $\mathcal E_n$ is finite-dimensional for $n \geq 6$. \subsection{Subalgebras} In order to better understand $\mathcal E_n$, we will study its subalgebras. Such algebras are mentioned by Kirillov in the introduction of \cite{Kirillov}, and analogues for other root systems are also studied in unpublished work of Bazlov and Kirillov \cite{Kirillov2}. \begin{defn} For any graph $G$ with vertex set $[n]$, the \emph{Fomin-Kirillov algebra} ${\mathcal E_G}$ of $G$ is the subalgebra of $\mathcal E_n$ generated by $x_{ij}$ for all edges $\overline{ij}$ in $G$. \end{defn} We will write $\mathcal E_G^d$ for the degree $d$ part of $\mathcal E_G$ and $\mathcal E_G^+ = \bigoplus_{d \geq 1} \mathcal E_G^d$ for the positive degree part of $\mathcal E_G$. \medskip Note that by this definition, $\mathcal E_n=\mathcal E_{K_n}$. If graphs $G$ and $G'$ are isomorphic, then so are the algebras $\mathcal E_G$ and $\mathcal E_{G'}$ since we can apply the automorphism of $\mathcal E_n$ that permutes the indices appropriately. Moreover, since $\mathcal E_n$ is a subalgebra of $\mathcal E_{n+1}$ (as can be seen from considering $\Pi$-degree), $\mathcal E_G$ does not depend on the choice of $n$---that is, it does not change if we add or remove isolated vertices. Finally, if $G$ and $H$ are graphs on disjoint vertex sets, then all variables in $\mathcal E_G$ commute with all variables in $\mathcal E_H$, so $\mathcal E_{G+H} \cong \mathcal E_G \otimes \mathcal E_H$. \medskip Given a subalgebra $\mathcal E_G$, we will write $\mathcal H_G(t)$ for its Hilbert series. Values of $\mathcal H_G$ for all connected graphs $G$ with at most five vertices are given in the Appendix. Computations were performed using the algebra package \texttt{bergman} \cite{bergman}. Although one might not necessarily expect $\mathcal H_G(t)$ to be particularly nice in general, a quick glance at the Appendix shows that in fact $\mathcal H_G(t)$ is a product of cyclotomic polynomials for all $G$ with at most five vertices. \subsection{Examples} \label{sec-examples} Note that it is not easy to give a presentation of $\mathcal E_G$: while $\mathcal E_n$ is defined by quadratic relations, $\mathcal E_G$ will usually have minimal relations that are not quadratic, possibly of much higher degree. (We may sometimes omit the word ``minimal'' when referring to minimal relations.) To illustrate this, we start with a few examples. \begin{ex}[The Dynkin diagram $A_3$]\label{ex-a3} Let $G=A_3$ be the path with two edges, as shown in Figure~\ref{fig-a3}. The only quadratic relations in $\mathcal E_{A_3}$ are those that come from the definition of $\mathcal E_3$, namely $\mathsf{a}^2=\mathsf{b}^2=0$. However, $\mathcal E_{A_3}$ has other relations: in $\mathcal E_3$, there is a quadratic relation involving the third edge $\mathsf{c}$: \begin{equation} \label{eq1} 0=\mathsf{ab}+\mathsf{bc}+\mathsf{ca} \tag{$*$} \end{equation} Multiplying \eqref{eq1} on the right by $\mathsf{a}$ gives \[0=\mathsf{aba}+\mathsf{bca}+\mathsf{caa}=\mathsf{aba}+\mathsf{bca},\] while multiplying \eqref{eq1} on the left by $b$ gives \[0=\mathsf{bab}+\mathsf{bbc}+\mathsf{bca}=\mathsf{bab}+\mathsf{bca}.\] Equating these, we deduce that $\mathsf{aba}=\mathsf{bab}$ in $\mathcal E_{A_3}$, which we call a \emph{braid relation}. We will see in Theorem~\ref{thm-a} that these generate all relations, so that \[\mathcal E_{A_3} = \langle \mathsf{a},\mathsf{b} \mid \mathsf{a}^2=\mathsf{b}^2=0, \;\; \mathsf{aba}=\mathsf{bab}\rangle.\] Then $\mathcal E_{A_3}$ is the \emph{nil-Coxeter algebra} of type $A_2$. It has basis $\{\mathrm{id}, \mathsf{a}, \mathsf{b}, \mathsf{ab}, \mathsf{ba}, \mathsf{aba}\}$ and Hilbert series \[\mathcal H_{A_3}(t) = 1+2t+2t^2+t^3 = (1+t)(1+t+t^2) = [2][3].\] \end{ex} \begin{figure} \begin{center} \begin{tikzpicture} \node[v] (i) at (0,0) {}; \node[v] (j) at (2,0) {}; \node[v] (k) at (4,0) {}; \draw[thick, ->] (i) -- node[above]{$\mathsf{a}$} (j); \draw[thick, ->] (j) -- node[above]{$\mathsf{b}$} (k); {\draw[dashed, thick, ->] (k.south) to [out = -135, in=-45] node[below]{$\mathsf{c}$} (i.south);} \end{tikzpicture} \end{center} \caption{\label{fig-a3} The Dynkin diagram $A_3$ consists of the two solid edges. The label on an edge directed from vertex $i$ to vertex $j$ represents the generator $x_{ij}$.} \end{figure} \begin{ex}[The star on four vertices] \label{ex-star3} Consider the star graph $K_{1,3}$ on four vertices, as shown on the left of Figure~\ref{fig-star}. A presentation of $\mathcal E_{K_{1,3}}$ is given by three quadratic relations: \[\mathsf{a}^2=\mathsf{b}^2=\mathsf{c}^2=0;\] three braid relations: \[\mathsf{aba}+\mathsf{bab}=\mathsf{bcb}+\mathsf{cbc}=\mathsf{cac}+\mathsf{aca}=0;\] and two \emph{claw relations}: \[\mathsf{abca}+\mathsf{bcab}+\mathsf{cabc}=\mathsf{acba}+\mathsf{bacb}+\mathsf{cbac}=0.\] \end{ex} The braid and claw relations are special cases of the following \emph{cyclic relations}. \begin{lemma} \label{lemma-cyclic}\cite[Lemma 7.2]{FK} For $m=3, \ldots, n,$ and any distinct $a_1, \ldots, a_m \in [n]$ the following relation holds in $\mathcal E_n$: \[\sum_{i=2}^m x_{a_1, a_i}x_{a_1, a_{i+1}}\cdots x_{a_1, a_m}x_{a_1, a_2}x_{a_1,a_3}\cdots x_{a_1, a_i}=0.\] \end{lemma} However, even star graphs have minimal relations that are not of this type. \begin{ex}[The star on five vertices] \label{ex-star4} For the star graph $K_{1,4}$ on five vertices, as shown in the middle of Figure~\ref{fig-star}, $\mathcal E_{K_{1,4}}$ has a presentation consisting of four quadratic relations, six braid relations, eight claw relations, six quartic cyclic relations, and three sextic relations of the following form: \[\mathsf{abacdc}-\mathsf{abcdca}+\mathsf{acdcba}+\mathsf{bacdcb}-\mathsf{bcdcab}-\mathsf{cabadc}+\mathsf{cdabac}-\mathsf{cdcaba}+\mathsf{dabacd}-\mathsf{dcabad}=0.\] \end{ex} \begin{figure} \begin{center} \vc{ \begin{tikzpicture} \node[v] (1) at (0,0){}; \node[v] (2) at (-1,1){}; \node[v] (3) at (1,1){}; \node[v] (5) at (1, -1){}; \draw[thick, ->] (1) -- node[above right]{$\mathsf{a}$} (2); \draw[thick, ->] (1) -- node[below right]{$\mathsf{b}$} (3); \draw[thick, ->] (1) -- node[below left]{$\mathsf{c}$} (5); \end{tikzpicture} } \quad\quad\quad \vc{ \begin{tikzpicture} \node[v] (1) at (0,0){}; \node[v] (2) at (-1,1){}; \node[v] (3) at (1,1){}; \node[v] (4) at (-1,-1){}; \node[v] (5) at (1, -1){}; \draw[thick, ->] (1) -- node[above right]{$\mathsf{a}$} (2); \draw[thick, ->] (1) -- node[below right]{$\mathsf{b}$} (3); \draw[thick, ->] (1) -- node[below left]{$\mathsf{c}$} (5); \draw[thick, ->] (1) --node[above left]{$\mathsf{d}$} (4); \end{tikzpicture} } \quad\quad\quad \vc{ \begin{tikzpicture} \node[v] (1) at (0,0){}; \node[v] (2) at (1.5,0){}; \node[v] (3) at (1.5,-1.5){}; \node[v] (4) at (0,-1.5){}; \draw[thick, ->] (1) -- node[above]{$\mathsf{a}$} (2); \draw[thick, ->] (2) -- node[right]{$\mathsf{b}$} (3); \draw[thick, ->] (3) -- node[below]{$\mathsf{c}$} (4); \draw[thick, ->] (4) --node[left]{$\mathsf{d}$} (1); \end{tikzpicture} } \end{center} \caption{\label{fig-star} On the left, the star $K_{1,3}$ on four vertices. In the middle, the star $K_{1,4}$ on five vertices. On the right, the $4$-cycle $\tilde{A}_3$.} \end{figure} \begin{ex}[The 4-cycle $\tilde{A}_3$] \label{ex-a3tilde} For the 4-cycle $\tilde{A}_3$, as shown on the right of Figure~\ref{fig-star}, $\mathcal E_{\tilde{A}_3}$ has a presentation consisting of four quadratic relations, four braid relations, and the following three relations: \[\mathsf{abc}+\mathsf{bcd}+\mathsf{cda}+\mathsf{dab} = 0,\] \[\mathsf{cba}+\mathsf{dcb}+\mathsf{adc}+\mathsf{bad} = 0,\] \[\mathsf{abda}+\mathsf{bcab}+\mathsf{cdbc}+\mathsf{dacd}+\mathsf{acbd} + \mathsf{bdac} = 0.\] \end{ex} The complexity of the relations in $\mathcal E_G$ increases quickly as the number of edges increases. Despite this, the Fomin-Kirillov algebras often seem to be relatively well-behaved. \section{Structure} In this section, we will describe some structural properties of $\mathcal E_n$ and $\mathcal E_G$. In particular, we will show that if $\mathcal E_G$ is finite-dimensional, then its Hilbert series has symmetric coefficients. \subsection{A bilinear form} The Fomin-Kirillov algebra $\mathcal E_n$ admits an action on itself defined as follows. \begin{prop} \label{prop-delta} \cite{FK} There exists a unique linear map $\Delta_{ab}\colon \mathcal E_n \to \mathcal E_n$ satisfying \[\Delta_{ab}(x_{ij}) = \begin{cases} 1, &\text{if $i=a$, $j=b$;}\\-1, & \text{if $i=b$, $j=a$;}\\0,&\text{otherwise;}\end{cases}\] and $\Delta_{ab}(PQ) = \Delta_{ab}(P)\cdot Q +\sigma_{ab}(P)\cdot \Delta_{ab}(Q)$. The operators $\Delta_{ab}$ satisfy the relations of $\mathcal E_n$, so they describe an action of $\mathcal E_n$ on itself. \end{prop} We will write $\Delta_P$ for the operator corresponding to an element $P \in \mathcal E_n$. In other words, if $P = x_{i_1j_1}x_{i_2j_2}\cdots x_{i_kj_k}$, then we let $\Delta_P = \Delta_{i_1j_1}\Delta_{i_2j_2}\cdots \Delta_{i_kj_k}$, and then we extend to all of $\mathcal E_n$ by linearity. Note that the operators $\Delta_P$ intertwine the automorphisms $\sigma \in \S_n$ in the following way: $\sigma(\Delta_P(Q))=\Delta_{\sigma(P)}(\sigma(Q))$. We can think of $\Delta_{ab}$ as having degree $-1$ and $\S_n$-degree $\sigma_{ab}$: if $P$ is homogeneous with respect to both degree and $\S_n$-degree, then $\Delta_{ab}P$ has degree $\deg P-1$ and $\S_n$-degree $\sigma_{ab}\sigma_P$. Similarly, there exists a dual (right) action of $\mathcal E_n$ on itself, defined as follows. The proof is essentially the same as that of Proposition~\ref{prop-delta} (and can also be deduced from Proposition~\ref{prop-pairing}). \begin{prop} \label{prop-nabla} There exists a unique linear map $\nabla_{ab}\colon \mathcal E_n \to \mathcal E_n$ (acting on the right) satisfying $(x_{ij})\nabla_{ab} =\Delta_{ab}(x_{ij}) $ and $(PQ)\nabla_{ab} = P\cdot (Q)\nabla_{ab} + (P)(\sigma_Q\nabla_{ab})\cdot Q$, where $\sigma_Q\nabla_{ab} = \nabla_{\sigma_Q(a)\sigma_Q(b)}$. The operators $\nabla_{ab}$ satisfy the relations of $\mathcal E_n$, so they describe an action of $\mathcal E_n$ on itself. \end{prop} We will similarly write $\nabla_P$ for the operator corresponding to an element $P \in \mathcal E_n$. The following lemma will be useful for performing calculations involving $\Delta_{ab}$ and $\nabla_{ab}$. It follows easily from repeated use of the Leibniz rules given in Propositions~\ref{prop-delta} and \ref{prop-nabla}. \begin{lemma} \label{lemma-leibniz} For a monomial $P= p_1 \cdots p_d \in \mathcal E_n$, write $P = P_k^L p_k P_k^R$. Then \begin{align*} \Delta_{ab}(P) &= \sum_{k=1}^d \Delta_{ab}(p_k) \cdot \sigma_{ab}(P_k^L) P_k^R, \text{ and}\\ (P)\nabla_{ab} &= \sum_{k=1}^d (p_k)(\sigma_{P_k^R}\nabla_{ab}) \cdot P_k^LP_k^R. \end{align*} \end{lemma} \begin{ex} \label{ex-delta} Here is a brief example of how to apply the $\Delta_{ij}$ and $\nabla_{ij}$ operators. \begin{align*} \Delta_{12}(x_{12}x_{23}x_{31}) &= \Delta_{12}(x_{12})\cdot x_{23}x_{31} + x_{21}\cdot \Delta_{12}(x_{23})\cdot x_{31} + x_{21}x_{13}\cdot \Delta_{12}(x_{31})\\ &= x_{23}x_{31},\\ (x_{12}x_{23}x_{31})\nabla_{12} &= x_{12}x_{23} \cdot (x_{31})\nabla_{12} + x_{12} \cdot (x_{23})\nabla_{32} \cdot x_{31} + (x_{12})\nabla_{23} \cdot x_{23}x_{31}\\ &= -x_{12}x_{31}. \end{align*} \end{ex} These actions are dual in the following sense. \begin{prop} \label{prop-pairing} If $P$ and $Q$ are homogeneous of the same degree, then $\Delta_P(Q) = \Delta_Q(P) = (P)\nabla_Q = (Q)\nabla_P$. This defines a symmetric bilinear form $\langle P, Q \rangle$ on $\mathcal E_n$. With respect to this form, the operators $\Delta_P$ and $\nabla_P$ are adjoint to right and left multiplication by $P$, respectively. \end{prop} \begin{proof} We induct on the degree $d$ of $P$ and $Q$. Choose monomials $P=p_1\cdots p_d$ and $Q=q_1\cdots q_d$, and write $P' = P_d^L= p_1\cdots p_{d-1}$ and $Q=Q_k^Lq_kQ_k^R$ as in Lemma~\ref{lemma-leibniz}. Then \[\Delta_P(Q) = \Delta_{P'} \Delta_{p_d}(Q) = \sum_{k=1}^d \Delta_{P'} (\sigma_{p_d}(Q_k^L)\cdot Q_k^R) \cdot \Delta_{p_d}(q_k).\] By induction, this equals \begin{equation} \label{eq-pairing} \sum_{k=1}^d\Delta_{ \sigma_{p_d}Q_k^L}\Delta_{Q_k^R}(P')\cdot \Delta_{q_k}(p_d) =\sum_{k=1}^d\Delta_{Q_k^L}(\sigma_{p_d}(\Delta_{Q_k^R}(P'))) \cdot \Delta_{q_k}(p_d). \tag{$*$} \end{equation} The $k$th term in the sum is only nonzero if $p_d = \pm q_k$, that is, if $\sigma_{p_d} = \sigma_{q_k}$. We therefore find that $\Delta_P(Q)$ equals \[\sum_{k=1}^d\Delta_{Q_k^L}(\sigma_{q_k}(\Delta_{Q_k^R}(P'))) \cdot \Delta_{q_k}(p_d) = \Delta_Q(P' \cdot p_d) = \Delta_Q(P).\] Also by induction, the left side of \eqref{eq-pairing} equals \[ \sum_{k=1}^d(P')\nabla_{ \sigma_{p_d}Q_k^L}\nabla_{Q_k^R}\cdot (p_d)\nabla_{q_k} = (P' \cdot p_d)\nabla_Q = (P)\nabla_Q. \] All that remains is to show the adjointness properties: \begin{align*} \langle P_1P_2, Q\rangle &= \Delta_{P_1P_2}(Q) = \Delta_{P_1}(\Delta_{P_2}(Q)) = \langle P_1, \Delta_{P_2}(Q)\rangle\\[2mm] \langle P_1P_2, Q\rangle &= (Q)\nabla_{P_1P_2} = ((Q)\nabla_{P_1})\nabla_{P_2} = \langle P_2, (Q)\nabla_{P_1} \rangle. \qedhere \end{align*} \end{proof} \begin{ex} By Example~\ref{ex-delta} and Proposition~\ref{prop-pairing}, \[\langle x_{12}x_{13}x_{12}, x_{12}x_{23}x_{31} \rangle = \langle x_{12}x_{13}, \Delta_{12}(x_{12}x_{23}x_{31}) \rangle = \langle x_{12}x_{13}, x_{23}x_{31} \rangle.\] Similarly, \[\langle x_{12}x_{13}, x_{23}x_{31}\rangle = \langle (x_{12}x_{13})\nabla_{23}, x_{31} \rangle = \langle -x_{13}, x_{31} \rangle = 1.\] \end{ex} Note that if $P$ and $Q$ are both homogeneous with respect to both the usual degree and $\S_n$-degree, then $\langle P, Q \rangle = 0$ unless $P$ and $Q$ have the same degree and $\sigma_P = \sigma_Q^{-1}$. \theoremstyle{plain} \newtheorem*{conj-nichols}{Conjecture \ref{conj-nichols}} \begin{conj} \label{conj-nichols} \cite{MS} The bilinear form $\langle\cdot,\cdot\rangle$ is nondegenerate on $\mathcal E_n$. \end{conj} This conjecture is equivalent to $\mathcal E_n$ being a special type of braided Hopf algebra called a Nichols algebra, which in this case is the quotient of $\mathcal E_n$ by the kernel of the bilinear form (see \cite{AS}). It is known that Conjecture~\ref{conj-nichols} holds for $n \leq 5$ \cite{Grana}. \subsection{Coproduct} The Fomin-Kirillov algebra has the structure of a braided Hopf algebra. This was noted in \cite{MS} and can be derived from the Hopf algebra structure of the twisted version of the algebra described in \cite{FP}. For more information about braided Hopf algebras, see \cite{AS}. We describe only the coproduct here, as we will need it later. The tensor product $\mathcal E_n \otimes \mathcal E_n$ has a braided product structure given by \[(P_1 \otimes Q_1)(P_2 \otimes Q_2) = (P_1 \sigma_{Q_1}(P_2)) \otimes (Q_1Q_2)\] for monomials $P_1$, $P_2$, $Q_1$, and $Q_2$. Then the coproduct $\Delta\colon \mathcal E_n \to \mathcal E_n \otimes \mathcal E_n$ is defined to be the braided homomorphism such that $\Delta(x_{ij}) = x_{ij} \otimes 1 + 1 \otimes x_{ij}$. Let $\mathcal E_n^\vee$ be the graded dual of $\mathcal E_n$, that is, the direct sum of the duals of each graded piece of $\mathcal E_n$. Then $\Delta$ defines an action of $\mathcal E_n^\vee$ on $\mathcal E_n$ as follows: if $p^\vee \in \mathcal E_n^\vee$ and $Q \in \mathcal E_n$, we let $p^\vee * Q = \sum p^\vee(Q_{(1)}^i) \cdot Q_{(2)}^i$, where $\Delta(Q) = \sum Q_{(1)}^i \otimes Q_{(2)}^i$. \begin{ex} We calculate $\Delta(x_{12}x_{23})$ to be \[(x_{12} \otimes 1+1 \otimes x_{12})(x_{23} \otimes 1 + 1 \otimes x_{23}) = x_{12}x_{23} \otimes 1 + x_{12}\otimes x_{23} + x_{13} \otimes x_{12} + 1 \otimes x_{12}x_{23}.\] Note that $(1 \otimes x_{12})(x_{23} \otimes 1) = x_{13} \otimes x_{12}$ due to the braiding. Let $\{x_{ij}^\vee\} \subset \mathcal E_n^\vee$ be the dual basis to $\{x_{ij}\} \subset \mathcal E_n^1$. Then $x_{13}^\vee* x_{12}x_{23}=x_{12}$. \end{ex} \begin{rmk} If $x_{ij}^\vee$ is an element of the dual basis as above, then $x_{ij}^\vee * Q = \operatorname{rev} ((\operatorname{rev} Q)\nabla_{ij}$). \end{rmk} \subsection{Properties of subalgebras} It is important to note how our subalgebras $\mathcal E_G$ behave with respect to the operators and bilinear form described above. The following lemma follows easily from the definitions of these operators. \begin{lemma} \label{lemma-delta} \begin{enumerate}[(a)] \item If $\overline{ij} \not \in G$, then $\Delta_{ij}(\mathcal E_G)=0$. \item For any $\nabla_{ij}$ and any graph $G$, $(\mathcal E_G) \nabla_{ij} \subset \mathcal E_G$. \item The coproduct $\Delta$ sends any element of $\mathcal E_G$ into $\mathcal E_n \otimes \mathcal E_G$. \item The left action of $\mathcal E_n^\vee$ on $\mathcal E_n$ restricts to an action on $\mathcal E_G$. \item If $H$ is a subgraph of $G$, then $\Delta(\mathcal E_H^+\mathcal E_G) \subset \mathcal E_H^+\mathcal E_n \otimes \mathcal E_G + \mathcal E_n \otimes \mathcal E_H^+\mathcal E_G$. \end{enumerate} \end{lemma} In the language of braided Hopf algebras, Lemma~\ref{lemma-delta}(c) says that $\mathcal E_G$ is a \emph{left coideal subalgebra} of $\mathcal E_n$. Likewise, Lemma~\ref{lemma-delta}(e) implies that $\mathcal E_H^+\mathcal E_n$ is a \emph{coideal} of $\mathcal E_n$. One important consequence is the following. \begin{lemma} \label{lemma-orthogonal} Let $G_1$ be a graph on $n$ vertices and $G_2$ its complement. Then the left ideal $\mathcal E_n\mathcal E_{G_1}^+$ is orthogonal to $\mathcal E_{G_2}$ with respect to $\langle \cdot, \cdot \rangle$. \end{lemma} \begin{proof} This follows using the adjointness of right multiplication by $x_{ij}$ and the left action of $\Delta_{ij}$ for $\overline{ij} \in G_1$ together with Lemma~\ref{lemma-delta}(a). \end{proof} \subsection{Finite dimensionality} \label{sec-finitedim} In this section, we will show that if $\mathcal E_G$ is finite-dimensional, then its Hilbert series must have symmetric coefficients. (This was proven for $\mathcal E_n$ in \cite{MS}.) We begin with a definition motivated by the theory of Coxeter groups. \begin{definition} For $w\in \mathcal E_n$, the \emph{right descent set} of $w$, denoted $R(w)$, is the graph containing edge $\overline{ij}$ whenever $wx_{ij} = 0$. Similarly, define the \emph{left descent set} $L(w)$ as the graph containing $\overline{ij}$ whenever $x_{ij}w=0$. \end{definition} We now use these descent sets to prove a key lemma regarding the action of $\mathcal E_n^\vee$ on $\mathcal E_n$. \begin{lemma} \label{lemma-integral} For all $P \in \mathcal E_n$, $\mathcal E_n^\vee * P$ is a left $\mathcal E_{L(P)}$-module. \end{lemma} \begin{proof} Let $Q=q^\vee * P$ be an arbitrary element of $\mathcal E_n^\vee * P$. For $\overline{ij} \in L(P)$, we compute $\sigma_{ij}q^\vee * x_{ij}P$, where $\sigma_{ij}q^\vee \in \mathcal E_n^\vee$ is given by $(\sigma_{ij}q^\vee)(R) = q^\vee(\sigma_{ij}R)$ for all $R \in \mathcal E_n$. If $\Delta(P) = \sum_k P_{(1)}^k \otimes P_{(2)}^k$, then \begin{align*} \Delta(x_{ij}P) &= \Delta(x_{ij})\Delta(P)\\ &= (x_{ij}\otimes 1 + 1 \otimes x_{ij}) \cdot \sum_k P_{(1)}^k \otimes P_{(2)}^k\\ &= \sum_k x_{ij}P_{(1)}^k \otimes P_{(2)}^k + \sum_k \sigma_{ij}P_{(1)}^k \otimes x_{ij}P_{(2)}^k. \end{align*} Thus \begin{align*} \sigma_{ij}q^\vee * x_{ij}P &= \sum_k (\sigma_{ij}q^\vee)(x_{ij}P^k_{(1)})\cdot P_{(2)}^k + \sum_k (\sigma_{ij}q^\vee)(\sigma_{ij}P_{(1)}^k)\cdot x_{ij}P_{(2)}^k\\ &=-\sum_kq^\vee(x_{ij}\sigma_{ij}P^k_{(1)}) \cdot P^k_{(2)} + x_{ij}\cdot \sum_k q^\vee(P_{(1)}^k) \cdot P_{(2)}^k\\ &=-r^\vee * P + x_{ij}Q, \end{align*} where $r^\vee \in \mathcal E_n^\vee$ is given by $r^\vee(R) =q^\vee(x_{ij}\sigma_{ij}R)$. Since $\overline{ij} \in L(P)$, $x_{ij}P=0$, so $x_{ij}Q = r^\vee * P$. \end{proof} One can similarly show that $\mathcal E_n^\vee * P$ is a right $\mathcal E_{R(P)}$-module. (See Proposition~\ref{prop-monicmcr} for a similar calculation.) As a direct consequence of Lemma~\ref{lemma-integral}, we have the following result. \begin{prop} \label{prop-monic} Suppose that $G \subset L(P)$ for some $P \in \mathcal E_n$. Then $\mathcal E_G \subset \mathcal E_n^\vee * P$, and $\mathcal E_G$ is finite-dimensional. If $P \in \mathcal E_G$, then $\mathcal E_n^\vee * P = \mathcal E_G$, and $P$ spans the top degree part of $\mathcal E_G$. \end{prop} \begin{proof} We may assume that $P$ is homogeneous of degree $d$. Then the only component of $\Delta(P)$ that lies in $\mathcal E_n^d \otimes \mathcal E_n^0$ is $P \otimes 1$. Thus $1 \in \mathcal E_n^\vee * P$. By Lemma~\ref{lemma-integral}, we must have that $\mathcal E_G \subset \mathcal E_n^\vee * P$. But clearly every element of $\mathcal E_n^\vee * P$ has degree at most $d$, so $\mathcal E_G$ has bounded degree and is therefore finite-dimensional. If $P \in \mathcal E_G$, then by Lemma~\ref{lemma-delta}, $\mathcal E_n^\vee * P \subset \mathcal E_G$, so we must have $\mathcal E_n^\vee * P = \mathcal E_G$. But the highest degree part of $\mathcal E_n^\vee * P$ has degree $d$, and since the homogeneous part of $\Delta(P)$ in $\mathcal E_n^0 \otimes \mathcal E_n^d$ is $1 \otimes P$, it follows that $P$ spans $\mathcal E_G^d$. \end{proof} We can now prove the main theorem of this section. \begin{thm} \label{thm-finitedim} Let $G$ be a graph such that $\mathcal E_G$ is finite-dimensional. Then \begin{enumerate}[(a)] \item the top degree component of $\mathcal E_G$ is spanned by a single monomial $w_0^G$ of degree $d_0$; \item for any $Q \in \mathcal E_G$, there exists $p^\vee \in \mathcal E_n^\vee$ such that $Q = p^\vee * w_0^G$; \item for any nonzero $Q \in \mathcal E_G$, there exists a monomial $P \in \mathcal E_G$ such that $PQ = w_0^G$; \item the coefficients of the Hilbert series $\mathcal H_G(t)$ are symmetric; \item the subwords of $w_0^G$ span $\mathcal E_G$; and \item $\operatorname{rev}(w_0^G) = \pm w_0^G$. \end{enumerate} \end{thm} \begin{proof} These all follow easily from Proposition~\ref{prop-monic}: For (a), any element $w_0^G$ of top degree $d_0$ satisfies $x_{ij}w_0^G=0$ for all $\overline{ij} \in G$, so $w_0^G$ spans $\mathcal E_G^{d_0}$. For (b), this is equivalent to $\mathcal E_n^\vee * w_0^G = \mathcal E_G$. For (c), let $P \in \mathcal E_G$ be a maximum degree monomial such that $PQ$ is nonzero. Then $x_{ij}PQ=0$ for all $\overline{ij} \in G$, so $PQ$ must be a multiple of $w_0^G$. For (d), the bilinear form $\mathcal E_G^d \otimes \mathcal E_G^{d_0-d} \to \mathbb Q$ that sends $P \otimes Q$ to the coefficient of $w_0^G$ in $PQ$ is nondegenerate by (c), so $\mathcal E_G^d$ and $\mathcal E_G^{d_0-d}$ have the same dimension. For (e), any element of $\mathcal E_n^\vee * w_0^G = \mathcal E_G$ lies in the span of the subwords of $w_0^d$ by the definition of the $*$ action. For (f), $\operatorname{rev}(w_0^G) = cw_0^G$ for some constant $c$, and since $\operatorname{rev}$ is an involution, $c=\pm 1$. \end{proof} \begin{ex} For $G=K_{1,3}$, as in Example~\ref{ex-star3}, the lexicographically minimal choice for $w_0^G$ is $\mathsf{abacabac}$. For $G=K_{1,4}$, as in Example~\ref{ex-star4}, the lexicographically minimal choice is $w_0^G=\mathsf{abacabacdabacabacdabadcabacd}$. \end{ex} \begin{rmk} Theorem~\ref{thm-finitedim} implies that $\mathcal E_G$ is a \emph{Frobenius algebra} when finite-dimensional: its Frobenius form is the bilinear form given in the proof of part (d). \end{rmk} The results in Theorem~\ref{thm-finitedim} are analogues of known results about finite Coxeter groups. See Section~\ref{sec-coxeter} for more discussion of this relationship. \section{Tensor product decomposition} \subsection{Subgraphs} We begin this section by describing the relationship between $\mathcal E_G$ and $\mathcal E_H$ when $H$ is a subgraph of $G$. \begin{thm} \label{thm-subgraph} Let $H$ be a subgraph of $G$. Then $\mathcal E_G$ is a free (left or right) $\mathcal E_H$-module. Specifically, $\mathcal E_G \cong \mathcal E_H \otimes (\mathcal E_G/\mathcal E_H^+\mathcal E_G)$ as left $\mathcal E_H$-modules and $\mathcal E_G \cong (\mathcal E_G/\mathcal E_G\mathcal E_H^+) \otimes \mathcal E_H$ as right $\mathcal E_H$-modules. \end{thm} \begin{proof} We prove just that $\mathcal E_G$ is a free left $\mathcal E_H$-module; the other result follows by passing to the opposite algebra. Let $I=\mathcal E_H^+\mathcal E_G$, and let the projection map be $\pi\colon \mathcal E_G\to \mathcal E_G/I$. We first claim that if $f\colon \mathcal E_G/I \to \mathcal E_G$ is any degree-preserving ($\mathbb Q$-linear) section and $\mu$ is the multiplication map, then $\varphi = \mu \circ (\mathrm{id} \otimes f) \colon \mathcal E_H \otimes (\mathcal E_G/I) \to \mathcal E_G$ is surjective. We will prove by induction that $\mathcal E_G^d$ lies in the image of $\varphi$. Note that the image of $\varphi$ is clearly closed under left multiplication by $\mathcal E_H$. Then if $\mathcal E_G^d$ lies in the image, so does the degree $d+1$ part of $I$. Since any element of $\mathcal E_G^{d+1}$ differs from an element in the image of $f$ by an element of $I$ of degree $d+1$, it follows that $\mathcal E_G^{d+1}$ also lies in the image, completing the induction. Next we show that $\varphi$ is injective. Choose bases $\{h_i\}$ of $\mathcal E_H$ and $\{\bar{g}_j\}$ of $\mathcal E_G/I$, and suppose that $\varphi(\sum_{i,j} c_{ij} h_i \otimes \bar{g}_j)= \sum_{i,j} c_{ij} h_ig_j= 0$ for some constants $c_{ij}$ not all zero, where $g_j = f(\bar{g}_j)$. By restricting to the degree $d$ part, we may assume that $\deg h_i + \deg \bar{g}_j = d$ for all $i$ and $j$. Find $i'$ such that some $c_{i'j}$ is nonzero with $h_{i'}$ of minimum degree $d'$. Let $\{h_{i}^\vee \mid \deg h_i \leq d\} \subset \mathcal E_H^\vee$ be the dual basis to $\{h_i \mid \deg h_i \leq d\} \subset \mathcal E_H$, and extend each $h_i^\vee$ to an element of $\mathcal E_n^\vee$ arbitrarily. We claim that \[0 = \pi(h_{i'}^\vee * \textstyle\sum_{i,j} c_{ij}h_ig_j) = \textstyle\sum_j c_{i'j}\bar{g}_j.\] This will be a contradiction since the $\bar{g}_j$ are linearly independent. To see why the claim is true, consider any term $c_{ij}h_ig_j$ with $c_{ij}$ nonzero. Then $\Delta(h_ig_j) = \Delta(h_i) \cdot \Delta(g_j)$. Since $\mathcal E_n \otimes I$ is a right ideal of $\mathcal E_n \otimes \mathcal E_G$ and $\Delta(h_i)$ is congruent to $h_i \otimes 1$ modulo $\mathcal E_n \otimes I$, we find that \[\pi(h_{i'}^\vee * h_ig_j) = (h_{i'}^\vee \otimes \pi)(\Delta(h_ig_j))= (h_{i'}^\vee \otimes \pi)((h_i \otimes 1) \cdot \Delta(g_j)).\] Since $\deg h_i \geq d'$, $(h_i \otimes 1) \cdot \Delta(g_j)$ can only have a nonzero component in $\mathcal E_n^{d'} \otimes \mathcal E_G^{d-d'}$ when $\deg h_i = d'$, in which case this component is $(h_i \otimes 1)(1 \otimes g_j) = h_i \otimes g_j$. Applying $h_{i'}^\vee \otimes \pi$ then gives 0 unless $i = i'$, in which case it gives $\bar{g}_j$, as desired. This completes the proof. \end{proof} \begin{cor} Let $H$ be a subgraph of $G$. Then $\mathcal H_H(t)$ divides $\mathcal H_G(t)$, and their quotient has positive coefficients. \end{cor} \begin{proof} The quotient is the Hilbert series of $\mathcal E_G/\mathcal E_G\mathcal E_H^+$. \end{proof} \begin{rmk} As noted in Lemma~\ref{lemma-delta}(c), $\mathcal E_G$ is a left coideal subalgebra of the braided Hopf algebra $\mathcal E_n$. Compare Theorem~\ref{thm-subgraph} to the results of \cite{Masuoka}, which gives several conditions that imply that a Hopf algebra is a free module over a (left) coideal subalgebra. \end{rmk} In light of Theorem~\ref{thm-subgraph}, we make the following definition. \begin{defn} \label{defn-mcr} A subset $M \subset \mathcal E_G$ is a set of left (resp. right) \emph{minimal coset representatives} for $\mathcal E_H$ if it is a basis of $\mathcal E_G$ as a right (resp. left) $\mathcal E_H$-module. \end{defn} Equivalently, by Theorem~\ref{thm-subgraph}, the projections of left (resp. right) minimal coset representatives give a basis of $\mathcal E_G/\mathcal E_G\mathcal E_H^+$ (resp. $\mathcal E_G/\mathcal E_H^+\mathcal E_G$). For this reason, we will sometimes abuse terminology and consider left minimal coset representatives to be elements of $\mathcal E_G/\mathcal E_G\mathcal E_H^+$. \begin{ex} Let $G=A_3$ be the path with two edges as in Example~\ref{ex-a3}, and let $H$ be the subgraph containing only edge $\mathsf{a}$. Then since $\mathcal E_H$ has basis $\{\mathrm{id}, \mathsf{a}\}$ and $\mathcal E_G$ has basis $\{\mathrm{id}, \mathsf{a}, \mathsf{b}, \mathsf{ab}, \mathsf{ba}, \mathsf{aba}\}$, a set of left minimal coset representatives is $\{\mathrm{id}, \mathsf{b}, \mathsf{ab}\}$. \end{ex} We use the term ``minimal coset representatives'' by analogy to the case of Coxeter groups and their corresponding nil-Coxeter algebras (see Section~\ref{sec-coxeter} for more details). We will typically take the elements of $M$ to be represented by monomials. Using Theorem~\ref{thm-subgraph}, we can prove the following lemma, which will be useful in the next section. \begin{lemma} \label{lemma-intersect} Let $H$ be a subgraph of $G$. Then $\mathcal E_H^+\mathcal E_n \cap \mathcal E_G = \mathcal E_H^+\mathcal E_G$. \end{lemma} \begin{proof} Let $M$ and $N$ be sets of right minimal coset representatives for $\mathcal E_H$ in $\mathcal E_G$ and for $\mathcal E_G$ in $\mathcal E_n$, respectively. Then any element $x \in \mathcal E_n$ can be written uniquely in the form $x=\sum h_{m,n}mn$, where $h_{m,n} \in \mathcal E_H$, $m \in M$, and $n \in N$. Thus $\{mn \mid m \in M, n \in N\}$ is a set of right minimal coset representatives for $\mathcal E_H$ in $\mathcal E_n$. If $x \in \mathcal E_H^+\mathcal E_n$, then each $h_{m,n}$ has positive degree, while if $x \in \mathcal E_G$, then $h_{m,n}=0$ unless $n$ is a constant. Thus if both hold, then we can write $x=\sum h_{m,n}m$ with $h_{m,n} \in \mathcal E_H^+$, so $x \in \mathcal E_H^+\mathcal E_G$. \end{proof} \subsection{Finite rank} In the event that $\mathcal E_G$ has finite rank as an $\mathcal E_H$-module, we can show that there is essentially a unique minimal coset representative of maximum degree. (If $\mathcal E_G$ itself is finite-dimensional, then this follows from Theorem~\ref{thm-finitedim}.) The general result will follow from the following proposition, akin to Proposition~\ref{prop-monic}. \begin{prop} \label{prop-monicmcr} Let $H$ be a subgraph of $G$. Suppose $P \in \mathcal E_G$ such that $P \not\in \mathcal E_H^+\mathcal E_G$ but $Px_{ij} \in \mathcal E_H^+\mathcal E_G$ for all $x_{ij} \in \mathcal E_G$. Then $P$ spans the top degree part of $\mathcal E_G/\mathcal E_H^+\mathcal E_G$. \end{prop} \begin{proof} Let $(\mathcal E_n^\vee)^H$ be the set of all $q^\vee \in \mathcal E_n^\vee$ such that $q^\vee(\mathcal E_H^+\mathcal E_n) = 0$. By Lemma~\ref{lemma-delta}(e), $q^\vee * \mathcal E_H^+\mathcal E_G \subset \mathcal E_H^+\mathcal E_G$, so $(\mathcal E_n^\vee)^H$ gives a left action $*$ on $\mathcal E_G/\mathcal E_H^+\mathcal E_G$. We may assume that $P$ is homogeneous of degree $d$. We claim that $(\mathcal E_n^\vee)^H * P$ spans $\mathcal E_G/\mathcal E_H^+\mathcal E_G$. First, since $P\not\in \mathcal E_H^+\mathcal E_G$, by Lemma~\ref{lemma-intersect}, $P \not\in \mathcal E_H^+\mathcal E_n$, so there exists a homogeneous element $q^\vee \in (\mathcal E_n^\vee)^H$ such that $q^\vee(P)=1$. Then since the component of $\Delta(P)$ in $\mathcal E_n^d \otimes \mathcal E_G^0$ is $P \otimes 1$, $q^\vee* P=1$, so $1 \in (\mathcal E_n^\vee)^H * P$. Then the claim will follow if we can show that the span of $(\mathcal E_n^\vee)^H * P$ in $\mathcal E_G/\mathcal E_H^+\mathcal E_G$ is a right $\mathcal E_G$-module. Let $Q = q^\vee * P \in (\mathcal E_n^\vee)^H * P$, and let $x_{ij}\in \mathcal E_G$. Write $\Delta(P) = \sum_k P^k_{(1)} \otimes P^k_{(2)}$, so that \[\Delta(Px_{ij}) = \sum_k P^k_{(1)}\sigma_{P^k_{(2)}}(x_{ij}) \otimes P^k_{(2)} + \sum_k P^k_{(1)} \otimes P^k_{(2)}x_{ij}.\] Then \begin{align*} q^\vee * (Px_{ij}) &= \sum_k q^\vee(P^k_{(1)}\sigma_{P^k_{(2)}}(x_{ij})) \cdot P^k_{(2)} + \sum_k q^\vee(P^k_{(1)}) \cdot P^k_{(2)}x_{ij}\\ &= -r^\vee * P + Qx_{ij}, \end{align*} where $r^\vee \in \mathcal E_n^\vee$ is defined (for $\S_n$-homogeneous $R$) by $r^\vee(R) = -q^\vee(R \sigma_R^{-1}\sigma_P(x_{ij}))$. If $R$ lies in $\mathcal E_H^+\mathcal E_n$, then so does $R \sigma_R^{-1}\sigma_P(x_{ij})$. Hence $q^\vee \in (\mathcal E_n^\vee)^H$ implies $r^\vee \in (\mathcal E_n^\vee)^H$. Then since $Px_{ij} \in \mathcal E_H^+\mathcal E_G$, it follows that $r^\vee * P$ and $Qx_{ij}$ are congruent modulo $\mathcal E_H^+\mathcal E_G$. Thus $(\mathcal E_n^\vee)^H * P$ spans $\mathcal E_G/\mathcal E_H^+\mathcal E_G$. Since the component of $\Delta(P)$ in $\mathcal E_n^0 \otimes \mathcal E_G^d$ is $1 \otimes P$, the top degree part of $(\mathcal E_n^\vee)^H * P$ is spanned by $P$, which gives the result. \end{proof} As an easy consequence, we get the following theorem analogous to Theorem~\ref{thm-finitedim}. \begin{thm} Let $H$ be a subgraph of $G$ such that $\mathcal E_G$ has finite rank as an $\mathcal E_H$-module, and let $M$ be a set of right minimal coset representatives. Then \begin{enumerate}[(a)] \item $M$ has a unique element $m_0$ of top degree; \item for any $m \in M$, $m = q^\vee * m_0$ in $\mathcal E_G/\mathcal E_H^+\mathcal E_G$ for some $q^\vee \in \mathcal E_n^\vee$ with $q^\vee(\mathcal E_H^+\mathcal E_n) = 0$; and \item for any $m \in M$, there exists $g \in \mathcal E_G$ such that $m_0 = mg$ in $\mathcal E_G/\mathcal E_H^+\mathcal E_G$. \end{enumerate} \end{thm} \begin{proof} By Proposition~\ref{prop-monicmcr}, the only $m \in M$ such that $mx_{ij} \in \mathcal E_H^+\mathcal E_G$ for all $x_{ij} \in \mathcal E_G$ has maximum degree in $M$ (which exists since $\mathcal E_G$ has finite rank), and this element spans the top degree of $\mathcal E_G/\mathcal E_H^+\mathcal E_G$ so must be unique. Parts (b) and (c) then follow as in the proof of Theorem~\ref{thm-finitedim}(b) and (c). \end{proof} \subsection{Complementary graphs} \label{sec-complementary} In some cases, the tensor product decomposition described in Theorem~\ref{thm-subgraph} is particularly simple. \begin{thm} \label{thm-tensor} Let $G$ be a graph, and let $G_1$ and $G_2$ be complementary subgraphs of $G$ such that any two vertices in the same connected component of $G_2$ have the same neighbors in $G_1$. Then the multiplication map $\mu\colon \mathcal E_{G_1} \otimes \mathcal E_{G_2} \to \mathcal E_G$ is an isomorphism of $\mathcal E_{G_1}$-$\mathcal E_{G_2}$-bimodules. In particular, $\mathcal H_{G_1}(t)\cdot \mathcal H_{G_2}(t) = \mathcal H_G(t)$. \end{thm} As a special case of this theorem, we have the following corollary. \begin{cor} \label{cor-tensor} Let $G_1$ be a complete multipartite graph on $n$ vertices, and let $G_2$ be its complement, a disjoint union of complete graphs. Then $\mathcal E_n \cong \mathcal E_{G_1} \otimes \mathcal E_{G_2}$. \end{cor} The case when $G_1 = K_{1, n-1}$ and $G_2 = K_{n-1}$ was proven in \cite{FP, MS}. \begin{proof}[Proof of Theorem~\ref{thm-tensor}] By our choice of $G_1$ and $G_2$, if $\overline{ij} \in G_1$ and $\overline{jk} \in G_2$, then $\overline{ik} \in G_1$. We first show that $\mu$ is surjective. By Theorem~\ref{thm-subgraph}, it suffices to show that the map $\mathcal E_{G_1} \to \mathcal E_G/\mathcal E_G\mathcal E_{G_2}^+$ is surjective. Choose a monomial $p_1\cdots p_d \in \mathcal E_G^d$. We show by induction on $d$ that it lies in the image of this map. Since the image is closed under left multiplication by $\mathcal E_{G_1}$, we are done if $p_1 \in \mathcal E_{G_1}$. Then suppose $p_1 \in \mathcal E_{G_2}$. By induction, we may assume that $p_2 \in \mathcal E_{G_1}$. If $p_1$ commutes with $p_2$, then $p_1p_2 \cdots p_d = p_2 p_1 \cdots p_d$, and then we are again done because $p_2 \in \mathcal E_{G_1}$. Otherwise, if $p_1 = x_{jk}$ and $p_2=x_{ij}$, then rewrite $x_{jk}x_{ij} = x_{ij}x_{ik} + x_{ik}x_{jk}$. Since $x_{ij}, x_{ik} \in \mathcal E_{G_1}$, we are again done by induction. To show that $\mu$ is injective, by Theorem~\ref{thm-subgraph}, it suffices to show that the map $\mathcal E_{G_2} \to \mathcal E_G/\mathcal E_{G_1}^+\mathcal E_G$ is injective, that is, that $\mathcal E_{G_2}$ intersects $\mathcal E_{G_1}^+\mathcal E_G$ trivially. But this holds because the $\Pi$-degree of any $\Pi$-homogeneous element of $\mathcal E_{G_2}$ has each part contained in a connected component of $G_2$, but this is not the case for any (nonzero) element of $\mathcal E_{G_1}^+\mathcal E_G$. \end{proof} In fact, it appears that the class of complementary graphs $G_1$ and $G_2$ for which $\mathcal E_G \cong \mathcal E_{G_1} \otimes \mathcal E_{G_2}$ is much more general than Corollary~\ref{cor-tensor} implies. Using Theorem~\ref{thm-tensor} and computations for small graphs, we can prove the following partial result on when such a tensor product decomposition holds. \begin{cor} \label{cor-complementary} Let $G_1$ be a graph on $n$ vertices and $G_2$ its complement. The class of graphs $G_1$ for which $\mathcal E_n \cong \mathcal E_{G_1} \otimes \mathcal E_{G_2}$ holds (as in Theorem~\ref{thm-tensor}) contains all graphs with at most five vertices and is closed under disjoint unions and complementation. \end{cor} \begin{proof We checked the claim for graphs with at most five vertices using \texttt{bergman}. Closure under complementation follows since $\mathcal E_n$ and all $\mathcal E_G$ are isomorphic to their opposite algebras. For disjoint unions, suppose that the tensor product decomposition exists for graphs $G_1 \cup G_2 = K_m$ and $H_1 \cup H_2 = K_n$. Then the complement of $G_1+H_1$ in $K_{m+n}$ is $L=(G_2+H_2) \cup K_{m,n}$. By Theorem~\ref{thm-tensor}, \[ \mathcal E_{m+n} \cong \mathcal E_{K_m+K_n} \otimes \mathcal E_{K_{m,n}} \cong \mathcal E_{G_1+H_1} \otimes \mathcal E_{G_2+H_2}\otimes \mathcal E_{K_{m,n}} \cong \mathcal E_{G_1+H_1} \otimes \mathcal E_L.\qedhere\] \end{proof} However, the tensor product decomposition does not hold in general. \begin{figure} \begin{tabular}{ccc} \begin{tikzpicture}[scale=0.7] \node[v] (1) at (-0.5,0){}; \node[v] (2) at (1,0){}; \node[v] (3) at (2,1){}; \node[v] (4) at (2,-1){}; \node[v] (5) at (3, 0){}; \node[v] (6) at (4.5, 0){}; \draw(1) node[above]{1} -- (2)node[above]{2} -- (3)node[above]{3} -- (5)node[above]{5} -- (6)node[above]{6} (2)--(4)node[below]{4}--(5) (2)--(5); \end{tikzpicture} &\quad& \begin{tikzpicture}[scale=0.7] \node[v] (1) at (-0.5,0){}; \node[v] (2) at (1,0){}; \node[v] (3) at (2,1){}; \node[v] (4) at (2,-1){}; \node[v] (5) at (3, 0){}; \node[v] (6) at (4.5, 0){}; \draw(1) node[above]{2} -- (2)node[above]{6} -- (3)node[above]{3} -- (5)node[above]{1} -- (6)node[above]{5} (2)--(4)node[below]{4}--(5) (3)--(4) (2)--(5); \end{tikzpicture}\\ \begin{tikzpicture}[scale=0.7] \node[v] (1) at (0,0){}; \node[v] (2) at (18:1.5){}; \node[v] (3) at (90:1.5){}; \node[v] (4) at (162:1.5){}; \node[v] (5) at (234:1.5){}; \node[v] (6) at (306:1.5){}; \draw (2) node[right]{2}--(3)node[above]{1}--(4)node[left]{5}--(5)node[below]{4}--(6)node[below]{3}--(2)--(1)node[below]{6}--(4); \end{tikzpicture} &\quad& \begin{tikzpicture}[scale=0.7] \node[v] (1) at (0,0){}; \node[v] (2) at (18:1.5){}; \node[v] (3) at (90:1.5){}; \node[v] (4) at (162:1.5){}; \node[v] (5) at (234:1.5){}; \node[v] (6) at (306:1.5){}; \draw (2) node[right]{3}--(3)node[above]{1}--(4)node[left]{4}--(5)node[below]{2}--(6)node[below]{5}--(2)--(1)node[below]{6}--(4) (1)--(3); \end{tikzpicture} \end{tabular} \caption{\label{fig-counter} Either of the graphs on the left can be used for $G_1$ in Proposition~\ref{prop-counter}. On the right are their complements.} \end{figure} \begin{prop} \label{prop-counter} Let $G_1$ be either of the two graphs on the left of Figure~\ref{fig-counter} and $G_2$ its complement (shown on the right). Then $\mathcal E_6 \not \cong \mathcal E_{G_1} \otimes \mathcal E_{G_2}$. \end{prop} \begin{proof} Note that $G_1$ has the property that its complement $G_2$ is isomorphic to $G_1$ but with an extra edge connecting two twin vertices. Hence by Theorem~\ref{thm-tensor}, $\mathcal H_{G_2}(t) = \mathcal H_{G_1}(t) \cdot (1+t)$. Then if $\mathcal E_6 \cong \mathcal E_{G_1} \otimes \mathcal E_{G_2}$, we would be able to solve for $\mathcal E_{G_1}$ using the first few terms of the Hilbert series for $\mathcal E_6$ as given in Section~\ref{sec-hilbert} to find that \begin{align*} \mathcal H_{G_1}(t) &= \left(\frac{\mathcal H_6(t)}{1+t}\right)^{1/2}\\ &= \left(\frac{1 + 15t + 125t^2 + 765t^3 + 3831t^4 + 16605t^5 + 64432t^6 + 228855t^7+\cdots}{1+t}\right)^{1/2}\\ &= \left(1+14t+111t^2+654t^3+3177t^4+13428t^5+51004t^6+177851t^7+\cdots\right)^{1/2}\\ &= 1+7t+31t^2+110t^3+338t^4+938t^5+2408t^6+\tfrac{11623}{2}t^7+\cdots, \end{align*} which is impossible due to the coefficient of $t^7$. \end{proof} It would be interesting to try to classify for which graphs and their complements a tensor product decomposition as in Corollary~\ref{cor-tensor} holds. \subsection{Computation} We include a brief discussion of a method for computing minimal coset representatives. This method of computation was used to achieve some of the results in the next section as well as the conjectures we will present later. Let $G$ be a graph and $e$ an edge not in $G$. Denote $G \cup \{e\}$ by $G'$. Suppose that we wish to compute a set of minimal coset representatives for $\mathcal E_{G}$ in $\mathcal E_{G'}$, that is, a basis for $\mathcal E_{G'}/\mathcal E_{G'}\mathcal E_{G}^+$. Unfortunately, without prior knowledge of the relations of $\mathcal E_{G'}$ (which would naively require an expensive noncommutative Gr\"obner basis calculation for $\mathcal E_n$), one cannot easily determine whether an element of $\mathcal E_{G'}$ lies in $\mathcal E_{G'}\mathcal E_{G}^+$. We can, however, give a simple sufficient condition for an element of $\mathcal E_{G'}$ not to lie in $\mathcal E_{G'}\mathcal E_{G}^+$. Let $H$ be the complement of $G'$ and $H' = H \cup \{e\}$ the complement of $G$. By Lemma~\ref{lemma-orthogonal}, every element of $\mathcal E_{G'}\mathcal E_{G}^+$ is orthogonal to $\mathcal E_{H'}$. Hence, any element of $\mathcal E_{G'}$ that pairs nontrivially with some element of $\mathcal E_{H'}$ does not lie in $\mathcal E_{G'}\mathcal E_{G}^+$. But we need not even pair with all elements of $\mathcal E_{H'}$: any element of $\mathcal E_{H'}\mathcal E_H^+$ is orthogonal to every element of $\mathcal E_{G'}$. Hence we only need pair with elements that are linearly independent in $\mathcal E_{H'}/\mathcal E_{H'}\mathcal E_H^+$. This suggests the following algorithm to calculate linearly independent sets of minimal coset representatives for $\mathcal E_G$ in $\mathcal E_{G'}$ and for $\mathcal E_H$ in $\mathcal E_{H'}$ simultaneously: \begin{alg}\label{alg-mcr} Let $G$ and $H$ be graphs and $e$ an edge such that $G \sqcup H \sqcup \{e\}=K_n$. Let $G' = G \cup \{e\}$ and $H' = H \cup \{e\}$, and set $M^0 = N^0 = \{\mathrm{id}\}$. For $d \geq 0$: \begin{itemize} \item Construct a matrix with rows indexed by $x_{ij}p$ for $x_{ij} \in \mathcal E_{G'}$ and $p \in M^d$, columns indexed by $x_{kl}q$ for $x_{kl} \in \mathcal E_{H'}$ and $q \in N^d$, and entries $\langle x_{ij}p, x_{kl}q \rangle$. \item Set $M^{d+1}$ to be the indices of a maximal set of linearly independent rows and likewise $N^{d+1}$ for columns. \end{itemize} Then $M^d$ and $N^d$ are subsets of degree $d$ minimal coset representatives for $\mathcal E_{G}$ inside $\mathcal E_{G'}$ and $\mathcal E_H$ inside $\mathcal E_{H'}$. (In other words, $M^d$ is linearly independent modulo $\mathcal E_{G'}\mathcal E_G^+$, and likewise $N^d$ modulo $\mathcal E_{H'}\mathcal E_H^+$.) \end{alg} \begin{rmk} Algorithm~\ref{alg-mcr} can be used to give a lower bound on $\mathcal H_{G'}(t)/\mathcal H_G(t)$, but this will not in general be exact. However, it is usually quite accurate for small graphs. For instance, one can show that if Conjecture~\ref{conj-nichols} holds and $\mathcal E_{H'} \otimes \mathcal E_G \cong \mathcal E_n$ (as in Theorem~\ref{thm-tensor}), then Algorithm~\ref{alg-mcr} will give a complete set of minimal coset representatives for $\mathcal E_G$ in $\mathcal E_{G'}$. In particular, the algorithm is exact for all graphs on at most five vertices. \end{rmk} \section{Coxeter groups and nil-Coxeter algebras} \label{sec-coxeter} In this section, we will describe various ways in which the subalgebras $\mathcal E_G$ share similar properties to Coxeter groups, or more specifically, to their nil-Coxeter algebras. \subsection{Definitions} We first recall the definition of (simply-laced) Coxeter groups and nil-Coxeter algebras as well as some of their basic properties. For further details, see \cite{hump} or \cite{BjornerBrenti}. Let $D$ be a graph, which we will refer to in this context as a \emph{simply-laced Dynkin diagram}. \begin{defn} Given a simply-laced Dynkin diagram $D$, the \emph{Coxeter group} $(W, S)$ of $D$ is the group generated by $S$, the set of \emph{simple reflections} $s_i$ for each vertex $i$ of $D$, and relations \begin{itemize} \item $s_i^2 = 1$ for any vertex $i$ of $D$; \item $s_is_j = s_js_i$ for nonadjacent vertices $i,j$ of $D$; \item $s_is_js_i = s_js_is_j$ for adjacent vertices $i,j$ of $D$ (called a \emph{braid relation}). \end{itemize} \end{defn} \begin{defn} Given an element $w \in W$, a \emph{reduced word} (or \emph{reduced expression}, or \emph{reduced decomposition}) $s_{i_1}s_{i_2}\cdots s_{i_\ell}$ for $w$ is a minimum length expression of $w$ as a product of generators $s_i$. The \emph{length} of $w$, denoted $\ell(w)$, is the length of any reduced word for $w$. We say that $w = u \cdot v$ is a \emph{reduced factorization} if $\ell(w) = \ell(u)+\ell(v)$. \end{defn} Given a Coxeter group, one can define the corresponding nil-Coxeter algebra as follows. \begin{definition} Given a simply-laced Dynkin diagram $D$, the \emph{nil-Coxeter algebra $\mathcal N$} is the associative algebra with a generator $t_i$ for each vertex $i$ of $D$ and relations \begin{itemize} \item $t_i^2 = 0$ for any vertex $i$ of $D$; \item $t_{i}t_{j} = t_j t_i$ for nonadjacent vertices $i,j$ of $D$; \item $t_{i}t_{j}t_{i} = t_{j}t_{i}t_{j}$ for adjacent vertices $i,j$ of $D$. \end{itemize} \end{definition} Note the similarity of this definition to that of $\mathcal E_n$. Given an element $w \in W$, we can define an element $t_w \in \mathcal N$ by choosing any reduced word $w = s_{i_1}s_{i_2} \cdots s_{i_\ell}$ and letting $t_w = t_{i_1}t_{i_2} \cdots t_{i_\ell}$. The element $t_w$ does not depend on the choice of reduced word. \begin{prop} The nil-Coxeter algebra $\mathcal N$ has basis $\{t_w \mid w \in W\}$ with multiplication given by \[t_ut_v = \begin{cases} t_{uv},&\text{if $\ell(u)+\ell(v) = \ell(uv)$;}\\ 0,& \text{otherwise.}\end{cases}\] \end{prop} For this reason, we will sometimes abuse notation and use the same letters to refer to both elements of $W$ and elements of $\mathcal N$. \subsection{Line graphs} We briefly describe how to relate the subalgebra $\mathcal E_G$ to a certain (twisted) nil-Coxeter algebra. \begin{definition} Given a graph $G$, let $L(G)$ denote the line graph of $G$, which we will think of as a simply-laced Dynkin diagram. Given a directed graph $G'$, the \emph{twisted nil-Coxeter algebra} of $L(G')$ is the associative algebra with a generator for each edge of $G'$ and relations \begin{itemize} \item $\mathsf{e} = \mathsf{f}$ for edges $\mathsf{e},\mathsf{f}$ of $G'$ with the same ends and direction; \item $\mathsf{e} = -\mathsf{f}$ for any directed 2-cycle $\mathsf{e},\mathsf{f}$ in $G'$; \item $\mathsf{e}^2 = 0$ for any edge $\mathsf{e}$ of $G'$; \item $\mathsf{ef} = \mathsf{fe}$ for edges $\mathsf{e}, \mathsf{f}$ of $G'$ that do not share an end; \item $\mathsf{efe} = \mathsf{fef}$ for edges $\mathsf{e}, \mathsf{f}$ of $G'$ that form a directed path; \item $\mathsf{efe} = -\mathsf{fef}$ for edges $\mathsf{e}, \mathsf{f}$ of $G'$ that share one end but do not form a directed path. \end{itemize} We refer to these last two relations as \emph{positive} and \emph{negative braid relations}, respectively. \end{definition} Note that we have abused notation slightly since the algebra depends on $G'$ and not just the line graph $L(G')$. The nonzero monomials of the twisted nil-Coxeter algebra of $L(G')$ are in bijection with the nonzero monomials of the nil-Coxeter algebra of $L(G)$, where $G$ is the underlying simple undirected graph of $G'$. Since the generators of $\mathcal E_G$ satisfy the same versions of the braid relation satisfied by the generators of the twisted nil-Coxeter algebra, we have the following proposition. \begin{proposition}\label{p Coxeter to FK} For an undirected graph $G$, let $G'$ be any directed graph whose underlying simple undirected graph is $G$. The algebra $\mathcal E_G$ is a quotient of the twisted nil-Coxeter algebra of $L(G')$. Moreover, if the edges of $G$ can be directed so that the indegree and outdegree at each vertex are at most one, then $\mathcal E_G$ is a quotient of the nil-Coxeter algebra of $L(G)$. \end{proposition} It is important to note that the twisted nil-Coxeter algebra of $L(G')$ is almost always infinite-dimensional since, among the Dynkin diagrams of finite type, only those of type $A_n$ are line graphs. Nevertheless, we will use Proposition~\ref{p Coxeter to FK} in our discussion of $\mathcal E_{\tilde{A}_{n-1}}$, utilizing the fact that the line graph of a cycle is again a cycle. \subsection{Analogy to Fomin-Kirillov algebras} Many of the results proved about $\mathcal E_G$ in the previous sections are analogues of the following facts about Coxeter groups and nil-Coxeter algebras. Throughout this section, let $D$ be a simply-laced Dynkin diagram, $(W, S)$ its Coxeter group, and $\mathcal N$ its nil-Coxeter algebra. There is a notion of \emph{Bruhat order} in $W$ defined as follows. Let $T = \{wsw^{-1} \mid w \in W, s \in S\}$ be the set of \emph{reflections} of $W$. The \emph{Bruhat order $<$} of $W$ is the transitive closure of the relations $tw < w$ for $t\in T$ such that $\ell(tw) < \ell(w)$. The \emph{left weak order $<_L$} of $W$ is the transitive closure of the relations $sw <_L w$ for $s\in S$ such that $\ell(sw) < \ell(w)$. An equivalent way of formulating Bruhat order is that $v \leq w$ if some reduced word for $w$ contains a substring that gives a reduced word for $v$; in fact, if this is true for some reduced word for $w$, then it is true for any reduced word for $w$. We can also formulate left weak order similarly: $v \leq_L w$ if there exists a reduced word for $w$ that has a final substring that is a reduced word for $v$ (but this depends on the reduced word for $w$). The analogue of Bruhat order in the Fomin-Kirillov algebra setting is, for $P, Q \in \mathcal E_G$: \[ Q \leq P \text{\quad if } Q \in \mathcal E_n^\vee * P. \] From the definition of the action of $\mathcal E_n^\vee$, it follows that if $Q \leq P$, then $Q$ lies in the span of the subwords of $P$. \begin{rmk} If Conjecture~\ref{conj-nichols} holds, then $Q \leq P$ if and only if $Q=(P)\nabla_x$ for some $x \in \mathcal E_n$. \end{rmk} The analogue of left weak order in the Fomin-Kirillov algebra setting is, for $P, Q \in \mathcal E_G$: \[ Q \leq_L P \text{\quad if there exists } x \in \mathcal E_G \text{ such that } xQ=P. \] Theorem~\ref{thm-finitedim} is then the analogue of the following facts about $W$ and $\mathcal N$: if $W$ is finite, then: \begin{enumerate}[(a)] \item there is a unique element $w_0 \in W$ of maximum length (so the top degree part of $\mathcal N$ is one-dimensional); \item $w\leq w_0$ for all $w\in W$; \item $w \leq_L w_0$ for all $w\in W$; \item the Poincar\'e series of $W$ (and hence the Hilbert series of $\mathcal N$) is symmetric; \item every element of $W$ is a subword of any reduced expression for $w_0$; and \item $w_0 = w_0^{-1}$. \end{enumerate} \begin{rmk} Although $Q \leq P$ is well-defined (independent of the choice of expression for $P$), the span of the subwords of $P$ is not well-defined in general. For instance, in $\mathcal E_3$, $x_{12}x_{23}x_{12} = x_{12}x_{13}x_{23}$, but clearly $x_{13}$ does not lie in the span of the subwords of $x_{12}x_{23}x_{12}$. However, Theorem~\ref{thm-finitedim} shows that this notion is well-defined when $P = w_0^G$. Likewise, in $\mathcal E_G$, $Q \leq_L P$ does not imply $Q \leq P$ in general: for instance, $x_{13}x_{23} \leq_L x_{12}x_{13}x_{23}$, but $x_{13}x_{23} \not\leq x_{12}x_{13}x_{23}$. But again, Theorem~\ref{thm-finitedim} shows that these notions coincide when $P = w_0^G$. \end{rmk} The notion of descent sets introduced in Section~\ref{sec-finitedim} is also the direct analogue of a notion from Coxeter groups: the \emph{left descent set} $L(w)$ of an element $w \in W$ is the set of all $s_i \in S$ such that $\ell(s_iw)<\ell(w)$ (or in $\mathcal N$, $s_iw=0$). (Define $R(w)$ similarly.) Then Proposition~\ref{prop-monic} is the analogue of the following statement about Coxeter groups: \begin{itemize} \item if $W$ has an element $w$ such that $L(w) = S$, then $W$ is finite-dimensional and $w$ is the longest element of $W$. \end{itemize} We can also now prove the following results about descent sets in $\mathcal E_G$, which are again analogues of facts about Coxeter groups. \begin{prop} If $w \in \mathcal E_G$, then $L(w),R(w)\subset G$. \end{prop} \begin{proof} Suppose $e = \overline{ij} \in L(w)$. Then Corollary~\ref{cor-tensor} implies that $\mathcal E_e \otimes \mathcal E_H \cong \mathcal E_n$, where $H=K_n \backslash \{e\}$. Thus left multiplication by $x_{ij}$ is injective on $\mathcal E_H$, so since $x_{ij}w = 0$, we must have that $w \not\in \mathcal E_H$. Thus $G \not\subset H$, so $e \in G$. \end{proof} \begin{proposition}\label{p descent set longword} Let $w \in \mathcal E_G$, and suppose $H \subset L(w)$. Then one can write $w = w_0^H g$ for some $g \in \mathcal E_G$. (Similarly, if $H' \subset R(w)$, then $w = g' w_0^{H'}$ for some $g' \in \mathcal E_G$.) \end{proposition} Recall that by Proposition~\ref{prop-monic}, $\mathcal E_H$ is necessarily finite-dimensional. \begin{proof} By Theorem~\ref{thm-subgraph}, there is a unique expression for $w$ of the form $w= \sum_{m \in M} h_mm$ for some $h_m \in \mathcal E_H$, where $M$ is a set of right minimal coset representatives for $\mathcal E_H$ in $\mathcal E_G$. For any $\overline{ij} \in H$, since $\mathcal E_G$ is a free left $\mathcal E_H$-module, $0=x_{ij}w=\sum_{m \in M} (x_{ij}h_m)m$ implies that $x_{ij} h_m=0$ for all $m\in M$. Hence by Proposition~\ref{prop-monic}, $h_m = c_mw_0^H$ for some constant $c_m$, so $w = w_0^H (\sum_{m \in M} c_mm)$. \end{proof} Along similar lines, we can prove the following result. \begin{prop} Let $w \in \mathcal E_G$, and suppose $w_0^Hw = 0$ for some $H \subset G$ with $\mathcal E_H$ finite-dimensional. Then $w \in \mathcal E_H^+\mathcal E_G$. \end{prop} \begin{proof} As in Proposition~\ref{p descent set longword}, we can write $w= \sum_{m \in M} h_mm$. Multiplying by $w_0^H$ gives $0= \sum_{m \in M} (w_0^H h_m)m$. Since $M$ is a set of minimal coset representatives, $w_0^H h_m=0$ for all $m$. But then each $h_m$ has degree at least 1, so $w \in \mathcal E_H^+\mathcal E_G$. \end{proof} Finally, Coxeter groups have a notion of parabolic subgroups: for $J \subseteq S$, the \emph{parabolic subgroup} $W_J$ is the subgroup of $W$ generated by the set $J$. The pair $(W_J, J)$ is equal to the Coxeter group of the induced subgraph of $D$ with vertex set $J$. Let $\mathcal N_J$ denote the nil-Coxeter algebra of $W_J$. The analogues of parabolic subgroups for $\mathcal E_G$ are the subalgebras $\mathcal E_H$ as $H$ ranges over subgraphs of $G$. However the fact about parabolic subgroups just mentioned does not hold---the minimal relations satisfied by the generators of $\mathcal E_H$ are not easily determined from those of $\mathcal E_G$. Despite this, we still have that Theorem~\ref{thm-subgraph} is the analogue of the following fact about parabolic subgroups: \begin{itemize} \item Each right (resp. left) coset $W_Jw$ (resp. $wW_J$) of $W_J$ contains a unique element of minimal length called a \emph{minimal coset representative}. Let $\leftexp{J}{W}$ (resp. $W^J$) denote the set of all such elements. Then $\mathcal N$ is a free left (resp. right) $\mathcal N_J$-module and as a left (resp. right) $\mathcal N_J$-module has basis $\leftexp{J}{W}$ (resp. $W^J$). \end{itemize} Although there exist many parallels between Coxeter groups and Fomin-Kirillov algebras, we do not yet have a good analogue for many of the geometric notions associated with Coxeter groups, such as reflections or root systems. For instance, for finite Coxeter groups, the length of the long word equals the number of positive roots in the corresponding root system, but we know of no combinatorial object that determines the maximum degree that occurs in $\mathcal E_G$. \section{Dynkin diagrams} In this section, we will describe $\mathcal E_G$ when $G$ is a (simply-laced) Dynkin diagram of finite type. In particular, we will show that any minimal relation of $\mathcal E_G$ in these cases is of one of three types: it is either a \emph{quadratic relation} inherited from $\mathcal E_n$, a \emph{braid relation} of the form $\mathsf{aba}+\mathsf{bab}=0$ between two edges that share a vertex, or a \emph{claw relation} of the form $\mathsf{abca}+\mathsf{bcab}+\mathsf{cabc}=0$ among three edges that share a vertex. (See the examples in Section~\ref{sec-examples} and Lemma~\ref{lemma-cyclic}.) \subsection{The case $\mathcal E_{A_n}$} Define $A_n$ to be the path on $n$ vertices (that is, the Dynkin diagram of type $A_n$). Then, in accordance with Example~\ref{ex-a3}, we have the following theorem. \begin{thm} \label{thm-a} The only minimal relations in $\mathcal E_{A_n}$ are the quadratic and braid relations. Its Hilbert series is \[\mathcal H_{A_n}(t) = [2][3][4] \cdots [n].\] In other words, $\mathcal E_{A_n} \cong \mathcal N_n$, the \textit{nil-Coxeter algebra} of type $A_{n-1}$. \end{thm} This follows easily from the existence of the divided difference representation of $\mathcal E_n$ given in \cite{FK}, but we present a different proof as it will be related to our investigation of $\mathcal E_{\tilde{A}_{n-1}}$ later. \begin{proof} Since $\ensuremath{\mathcal N}_{n}$ is defined using the quadratic and braid relations, $\mathcal E_{A_n}$ is a quotient of $\ensuremath{\mathcal N}_{n}$. We will show that the projection $\Theta\colon \ensuremath{\mathcal N}_{n} \to \mathcal E_{A_n}$ is an isomorphism by showing that the image under $\Theta$ of the basis $\{w\mid w \in \S_n\}$ of $\ensuremath{\mathcal N}_{n}$ is linearly independent. We prove this by the following pairing computation: \begin{equation*}\label{e pairing Sn} \text{ for $w,v\in \S_n$,} \qquad\qquad \langle \Theta(w), \operatorname{rev}(\Theta(v)) \rangle = \begin{cases} 1 & \text{ if } w = v,\\ 0 & \text{ if } w\neq v. \end{cases} \end{equation*} This pairing is zero unless the $\S_n$-degree of $\Theta(w)$ and $\Theta(v)$ are the same, which gives the second case. To prove $\langle \Theta(w), \operatorname{rev}(\Theta(w)) \rangle = 1$, we induct on the length of $w$. Let $P = \Theta(w) = p_1\cdots p_d$ be an expression for $\Theta(w)$ as a product of generators of $\mathcal E_{A_n}$, and write $P = P^L_jp_jP^R_j$. Then by Lemma \ref{lemma-leibniz} and Proposition \ref{prop-pairing}, \[ \langle P,\operatorname{rev} (P)\rangle =\sum_{j=1}^d (p_j) (\sigma_{P^R_j}\nabla_{p_d}) \cdot \langle P^L_j P^R_j, \operatorname{rev}(P^L_d) \rangle.\] By the strong exchange condition for $\S_n$, there is a unique index $j$ for which $P^L_jP^R_j$ has the same $\S_n$-degree as $P^L_d$; clearly this index is $j=d$. Hence only the $j=d$ term in the sum can be nonzero, so we find that $\langle P, \operatorname{rev}(P) \rangle = \langle P^L_d, \operatorname{rev}(P^L_d)\rangle = 1$ by induction. \end{proof} \subsection{The case $\mathcal E_{D_n}$} The next simplest case is that of the Dynkin diagram of type $D_n$ (which we will simply call $D_n$). We denote the edges of $D_n$ by $\mathsf{a}$, $\mathsf{b}$, $\mathsf{1}$, \dots, $\mathsf{n-3}$ as shown in Figure~\ref{fig-d}. (We will use these labels to denote both the edges of the graph and the corresponding variables.) \begin{figure} \begin{center} \begin{tikzpicture}[scale = 1.2] \node[v] (a) at (150:1){}; \node[v] (b) at (-150:1){}; \node[v] (1) at (0,0){}; \node[v] (2) at (1,0){}; \node[v] (3) at (2,0){}; \node[v] (4) at (4,0){}; \node[v] (5) at (5,0){}; \node at (0,-1.6){}; \draw[thick, ->] (5)--node[above]{${\scriptstyle \mathsf{n-3}}$}(4); \draw[thick] (4)--(3.5,0); \draw[thick, ->] (2.5,0)--(3); \draw[thick, ->] (3)--node[above]{${\scriptstyle \mathsf{2}}$}(2); \draw[thick, ->] (2)--node[above]{${\scriptstyle \mathsf{1}}$}(1); \draw[thick, ->] (1)--node[above]{${\scriptstyle \mathsf{a}}$}(a); \draw[thick, ->] (1)--node[below]{${\scriptstyle \mathsf{b}}$}(b); \draw[thick, loosely dotted] (2.7,0)--(3.4,0); \end{tikzpicture} \qquad \begin{tikzpicture}[scale = 1.2] \node[v] (a) at (150:1){}; \node[v] (b) at (-150:1){}; \node[v] (1) at (0,0){}; \node[v] (2) at (1,0){}; \node[v] (3) at (2,0){}; \node[v] (4) at (4,0){}; \node[v] (5) at (5,0){}; \node at (0,-1.6){}; \draw[thick, ->] (5)--node[below, near end]{${\scriptstyle \mathsf{(n-3)'}}$}(4); \draw[thick, loosely dotted] (2.7,0)--(3.4,0); \draw[thick, ->] (5) to [out = 165, in=15, near end] node[above]{${\scriptstyle \mathsf{3'}}$}(3); \draw[thick, ->] (5) to [out =150, in=35, near end] node[above]{${\scriptstyle \mathsf{2'}}$}(2); \draw[thick, ->] (5) to [out =135, in=45, near end] node[above]{${\scriptstyle \mathsf{1'}}$}(1); \draw[thick, ->] (5) to [out =120, in=35, near end] node[above]{${\scriptstyle \mathsf{a'}}$}(a); \draw[thick, ->] (5) to [out =-135, in=-15] node[above]{${\scriptstyle \mathsf{b'}}$}(b); \draw[dashed] (4)--(3.5,0) (2.5,0)--(3)--(2)--(1)--(a) (1)--(b); \end{tikzpicture} \end{center} \caption{\label{fig-d} On the left, the Dynkin diagram $D_n$ with edge labels. On the right, the star $K_{1,n-1}$ with primed edge labels to be used in Lemma~\ref{lemma-d3}.} \end{figure} The main result of this section is the following theorem. \begin{thm} \label{thm-d} The only minimal relations in $\mathcal E_{D_n}$ are the quadratic, braid, and claw relations. Its Hilbert series is \[\mathcal H_{D_n}(t) = [n][n-1] \cdot [4][6][8]\cdots[2n-4].\] \end{thm} In order to prove Theorem~\ref{thm-d}, we will construct a set of minimal coset representatives for $\mathcal E_{D_{n-1}}$ in $\mathcal E_{D_{n}}$. For $n \geq 3$, let \begin{align*} M_n=\{&\mathrm{id}, \quad \mathsf{n-3}, \quad \mathsf{(n-4)(n-3)}, \quad \dots, \quad \mathsf{(n-3)!_{\scriptscriptstyle\swarrow} }, \\ &\mathsf{a(n-3)!_{\scriptscriptstyle\swarrow} }, \quad \mathsf{b(n-3)!_{\scriptscriptstyle\swarrow} },\quad \mathsf{ab(n-3)!_{\scriptscriptstyle\swarrow} }, \quad \mathsf{ba(n-3)!_{\scriptscriptstyle\swarrow} }, \quad \mathsf{aba(n-3)!_{\scriptscriptstyle\swarrow} }, \\ & \mathsf{1aba(n-3)!_{\scriptscriptstyle\swarrow} },\quad \mathsf{2!_{\scriptscriptstyle\searrow} aba(n-3)!_{\scriptscriptstyle\swarrow} },\quad \dots,\quad \mathsf{(n-3)!_{\scriptscriptstyle\searrow} aba(n-3)!_{\scriptscriptstyle\swarrow} } \}, \end{align*} where $\mathsf{i!_{\scriptscriptstyle\swarrow}}=\mathsf{12\cdots(i-1)i}$ and $\mathsf{i!_{\scriptscriptstyle\searrow}}=\mathsf{i(i-1)\cdots 21}$. Note that $M_n$ can be given the structure of a (graded) partially ordered set in which $m'$ covers $m$ if there exists an edge variable $\mathsf{e}$ such that $m' = \mathsf{e}\,m$. Thus we can label the edges in the Hasse diagram of $M_n$ by generators of $\mathcal E_{D_n}$, and each element $m$ of $M_n$ is (up to sign) the right-to-left product of the edge labels along a saturated chain that starts at the minimal element $\mathrm{id}$ and ends at $m$. See Figure~\ref{fig-mcr} for the case $M_5$. \begin{figure} \begin{center} \begin{tikzpicture} \node[v] (1) at (0,0){}; \node[v] (2) at (0,1){}; \node[v] (3) at (0,2){}; \node[v] (4) at (-.75, 2.75){}; \node[v] (5) at (.75, 2.75){}; \node[v] (6) at (-.75, 3.75){}; \node[v] (7) at (.75, 3.75){}; \node[v] (8) at (0, 4.5){}; \node[v] (9) at (0, 5.5){}; \node[v] (10) at (0,6.5){}; \draw(1)node[left]{$\mathrm{id}$}--(2)node[left]{$\mathsf{2}$}--(3)node[left]{$\mathsf{12}$}--(4)node[left]{$\mathsf{a12}$}--(6)node[left]{$\mathsf{ba12}$}--(8)node[left]{$\mathsf{aba12}$}--(9)node[left]{$\mathsf{1aba12}$}--(10)node[left]{$\mathsf{21aba12}$} (3)--(5)node[right]{$\mathsf{b12}$}--(7)node[right]{$\mathsf{ab12}$}--(8); \end{tikzpicture} \end{center} \caption{\label{fig-mcr} Hasse diagram for $M_5$, the (left) minimal coset representatives for $D_4$ inside $D_5$.} \end{figure} When $n=3$, $D_3=A_3$, and $\mathcal E_{D_3}$ has basis $M_3 = \{\mathrm{id}, \mathsf{a}, \mathsf{b}, \mathsf{ab}, \mathsf{ba}, \mathsf{aba}\}$ by Theorem~\ref{thm-a}. For $n \geq 4$, we claim that $M_n$ is a set of (left) minimal coset representatives for $\mathcal E_{D_{n-1}}$ in $\mathcal E_{D_{n}}$. We proceed in two steps. Let $I_n = \mathcal E_{D_n}\mathcal E_{D_{n-1}}^+$. \begin{lemma} \label{lemma-d1} The set $M_n$ spans $\mathcal E_{D_n}/I_n$. \end{lemma} In other words, we can write any element of $\mathcal E_{D_n}$ in a \emph{normal form} as a linear combination of monomials, each of which is a product of a minimal coset representative and a monomial in $\mathcal E_{D_{n-1}}$. The proof of this lemma essentially gives a straightening algorithm for $\mathcal E_{D_n}$. \begin{proof} We induct on $n$. Assume $n \geq 4$. Consider any monomial $m \in \mathcal E_{D_n}$. If $m$ does not contain any instance of the variable $\mathsf{n-3}$, then it either lies in $I_n$ or is the identity element $\mathrm{id} \in M_n$. If $m$ contains exactly one occurrence of $\mathsf{n-3}$, then for it not to lie in $I_n$, it must equal $A\mathsf{(n-3)}$ for some $A \in \mathcal E_{D_{n-1}}$. By induction, $A$ is congruent modulo $I_{n-1}$ to a linear combination of elements in $M_{n-1}$. Since $\mathsf{n-3}$ commutes with $\mathcal E_{D_{n-2}}$, we have $I_{n-1}\cdot \mathsf{(n-3)} \subset I_n$, so $m$ is congruent modulo $I_n$ to a linear combination of elements of $M_{n-1} \cdot \mathsf{(n-3)} \subset M_n$. If $m$ has more than one occurrence of $\mathsf{n-3}$, as above we may assume that it ends with a substring of the form $\mathsf{(n-3)}A\mathsf{(n-3)}$, where $A \in \mathcal E_{D_{n-1}}$. By the previous paragraph, we may assume that $A \in M_{n-1}$. If $A=\mathrm{id}$, this clearly vanishes. Otherwise, either $A=B\mathsf{(n-4)}$ for some $B \in M_{n-2}$ or $A = \mathsf{(n-4)!_{\scriptscriptstyle\searrow} aba(n-4)!_{\scriptscriptstyle\swarrow} }$. In the first case, by the braid relation \[\mathsf{(n-3)}A\mathsf{(n-3)} = B\mathsf{(n-3)(n-4)(n-3)} = B\mathsf{(n-4)(n-3)(n-4)} \in I_n,\] so $m$ will also lie in $I_n$. It remains to consider the case when $m$ ends in $X=\mathsf{(n-3)!_{\scriptscriptstyle\searrow} aba(n-3)!_{\scriptscriptstyle\swarrow} }$. We claim that $\mathsf{a}X=X\mathsf{b}$, $\mathsf{b}X=X\mathsf{a}$, and $\mathsf{i}X = X\mathsf{i}$ for $\mathsf{i} = \mathsf{1}, \mathsf{2}, \dots, \mathsf{n-4}$. This shows that either $m = X \in M_n$ or $m \in I_n$, which will complete the proof. To show $X\mathsf{a}=\mathsf{b}X$, since $\mathsf{a}$ and $\mathsf{b}$ commute with $\mathsf{2}, \mathsf{3}, \dots, \mathsf{n-3}$, it suffices to show that $\mathsf{1aba1a} = \mathsf{b1aba1}$. This follows from the quadratic, braid, and claw relations because \[\mathsf{1aba1a} = \mathsf{1ab1a1} = (\mathsf{ab1a}+\mathsf{b1ab})\mathsf{a1} = \mathsf{ab1aa1}+\mathsf{b1aba1} = \mathsf{b1aba1}.\] Similarly, $X\mathsf{b}=\mathsf{a}X$. Finally, for $\mathsf{i}=\mathsf{1}, \mathsf{2}, \dots, \mathsf{n-4}$, since $\mathsf{i}$ commutes with $\mathsf{i+2}, \dots, \mathsf{n-3}$, it suffices to show that $\mathsf{(i+1)!_{\scriptscriptstyle\searrow} aba(i+1)!_{\scriptscriptstyle\swarrow} i} = \mathsf{i(i+1)!_{\scriptscriptstyle\searrow} aba(i+1)!_{\scriptscriptstyle\swarrow} }$. This holds because \begin{multline*} \mathsf{(i+1)!_{\scriptscriptstyle\searrow} aba(i-1)!_{\scriptscriptstyle\swarrow} i(i+1)i} = \mathsf{(i+1)!_{\scriptscriptstyle\searrow} aba(i-1)!_{\scriptscriptstyle\swarrow} (i+1)i(i+1)} \\ = \mathsf{(i+1)i(i+1)(i-1)!_{\scriptscriptstyle\searrow} aba(i+1)!_{\scriptscriptstyle\swarrow} } = \mathsf{i(i+1)i(i-1)!_{\scriptscriptstyle\searrow} aba(i+1)!_{\scriptscriptstyle\swarrow} }.\qedhere \end{multline*} \end{proof} Note that in order to put an element of $\mathcal E_{D_n}$ into normal form, we used only the quadratic, braid, and claw relations. \medskip To show linear independence, we first show that the highest degree minimal coset representative is nonzero. \begin{lemma}\label{lemma-d3} The highest degree element $X = \mathsf{(n-3)!_{\scriptscriptstyle\searrow} aba(n-3)!_{\scriptscriptstyle\swarrow} }\in M_n$ does not lie in $I_n$. \end{lemma} Let $K_{1, n-1}$ be the star centered at the end vertex in $D_n$, and label its edges using primed symbols as in Figure~\ref{fig-d}. (Note that, for convenience, we let $\mathsf{(n-3)} = \mathsf{(n-3)'}$.) Define $\mathsf{i'!_{\scriptscriptstyle\searrow} }$ and $\mathsf{i'!_{\scriptscriptstyle\swarrow} }$ analogously to the unprimed versions. \begin{proof} By Lemma~\ref{lemma-orthogonal}, $I_n$ is orthogonal to $\mathcal E_{K_{1, n-1}}$. Hence we need only show that $X$ is not orthogonal to some element of $\mathcal E_{K_{1, n-1}}$. Let $X' = \mathsf{(n-3)'!_{\scriptscriptstyle\searrow} a'b'a'(n-3)'!_{\scriptscriptstyle\swarrow} }$. We will show that $\langle X, X' \rangle = (-1)^{n-1}$. For $n=3$, an easy computation shows that $\langle \mathsf{aba}, \mathsf{aba}\rangle = 1$. For $n\geq 4$, let $Y = \mathsf{(n-4)!_{\scriptscriptstyle\searrow} aba(n-4)!_{\scriptscriptstyle\swarrow} }$ and $Y' = \mathsf{(n-4)'!_{\scriptscriptstyle\searrow} a'b'a'(n-4)'!_{\scriptscriptstyle\swarrow} }$, so that $X=\mathsf{(n-3)}Y\mathsf{(n-3)}$ and $X'=\mathsf{(n-3)}Y'\mathsf{(n-3)}$. Since \[ \Delta_{\mathsf{n-3}}(X') = Y'\mathsf{(n-3)}+ \sigma_{\mathsf{n-3}}(\mathsf{(n-3)}Y') = Y'\mathsf{(n-3)} - \mathsf{(n-3)}\sigma_{\mathsf{n-3}}(Y'), \] we have \[ \langle X, X' \rangle = \langle \mathsf{(n-3)}Y, \Delta_{\mathsf{n-3}}(X') \rangle = \langle\mathsf{(n-3)}Y, Y'\mathsf{(n-3)}\rangle - \langle \mathsf{(n-3)}Y, \mathsf{(n-3)}\sigma_{\mathsf{n-3}}(Y') \rangle. \] But $\mathsf{(n-3)}Y$ ends in $\mathsf{n-4}$ (or $\mathsf{a}$ if $n=4$), which does not occur in $Y'\mathsf{(n-3)}$, so the first term vanishes. Thus \[ \langle X, X' \rangle = -\langle \mathsf{(n-3)}Y, \mathsf{(n-3)}\sigma_{\mathsf{n-3}}(Y') \rangle = - \langle (\mathsf{(n-3)}Y)\nabla_{\mathsf{n-3}}, \sigma_{\mathsf{n-3}}(Y')\rangle. \] But in fact $(\mathsf{(n-3)}Y)\nabla_{\mathsf{n-3}} = Y$. To see this, note that none of the edges in $Y$ contain the end vertex in $D_n$, meaning $\nabla_{\mathsf{n-3}}$ will not act on any variable in $Y$, and also that $\sigma_{Y}(\mathsf{n-3}) = \mathsf{n-3}$. It follows that $\langle X, X' \rangle = - \langle Y, \sigma_{\mathsf{n-3}}(Y')\rangle$. But note that the edges appearing in $\sigma_{\mathsf{n-3}}(Y')$ are contained in the star centered at the end vertex of $D_{n-1}$. Thus $\langle Y, \sigma_{\mathsf{n-3}}(Y')\rangle$ is the computation for $n-1$, so the result follows by induction. \end{proof} \begin{lemma}\label{lemma-d2} The set $M_n$ is linearly independent in $\mathcal E_{D_n}/I_n$. \end{lemma} \begin{proof} Note that the elements of $M_n$ all differ in either degree or $\S_n$-degree. Therefore, they are linearly independent unless one of them lies in $I_n$. But each element of $M_n$ is a right factor of the longest element $X$, so since $X$ does not lie in the left ideal $I_n$, none of the elements do. \end{proof} \begin{proof} [Proof of Theorem~\ref{thm-d}] By Lemmas~\ref{lemma-d1} and \ref{lemma-d2}, $M_n$ is a set of minimal coset representatives for $\mathcal E_{D_{n-1}}$ in $\mathcal E_{D_{n}}$. Since the only relations needed in Lemma~\ref{lemma-d1} were the quadratic, braid, and claw relations, Lemma~\ref{lemma-d2} implies that these generate all relations in $\mathcal E_{D_n}$. We can now prove \[\mathcal H_{D_n}(t) = [n][n-1] \cdot [4][6][8]\cdots [2n-4]\] by induction on $n$. The case $n=3$ follows from Theorem~\ref{thm-a}. For $n>3$, since $M_n$ is a set of minimal coset representatives and $\mathcal H_{M_n}(t) = [n](1+t^{n-2}) = \frac{[n][2n-4]}{[n-2]}$, we have \begin{align*} \mathcal H_{D_n}(t) &= \mathcal H_{M_n}(t) \cdot \mathcal H_{D_{n-1}}(t)\\ &= \frac{[n][2n-4]}{[n-2]} \cdot [n-1][n-2] \cdot [4][6] \cdots [2n-6]\\ &= [n][n-1] \cdot [4][6] \cdots [2n-4]. \qedhere \end{align*} \end{proof} \subsection{The cases $\mathcal E_{E_n}$ for $n=6$, $7$, and $8$} In this section, we will describe $\mathcal E_{E_n}$ for $n=6$, $7$, and $8$, where $E_n$ is the Dynkin diagram of type $E_n$ as shown in Figure~\ref{fig-e}. \begin{figure} \begin{center} \dr{(0,0)node[v]{}--(1,0)node[v]{}--(2,0)node[v]{}--(3,0)node[v]{}--(4,0)node[v]{} (2,0)--(2,1)node[v]{}} \quad\quad\quad \dr{(0,0)node[v]{}--(1,0)node[v]{}--(2,0)node[v]{}--(3,0)node[v]{}--(4,0)node[v]{}--(5,0)node[v]{} (2,0)--(2,1)node[v]{}} \\[5mm] \dr{(0,0)node[v]{}--(1,0)node[v]{}--(2,0)node[v]{}--(3,0)node[v]{}--(4,0)node[v]{}--(5,0)node[v]{}--(6,0)node[v]{} (2,0)--(2,1)node[v]{}} \end{center} \caption{\label{fig-e} The Dynkin diagrams $E_6$, $E_7$, and $E_8$.} \end{figure} \begin{thm} The only minimal relations in $\mathcal E_{E_6}$, $\mathcal E_{E_7}$, and $\mathcal E_{E_8}$ are the quadratic, braid, and claw relations. Their Hilbert series are \begin{align*} \mathcal H_{E_6}(t) &= \frac{[4][5][6]^2[8][9]}{[3]},\\ \mathcal H_{E_7}(t) &= \frac{[6]^2[8][9][10][12][14]}{[3]},\\ \mathcal H_{E_8}(t) &= \frac{[6][8][10][12][14][15][18][20][24]}{[3][5]}. \end{align*} \end{thm} Our proof of this theorem is largely computational. For this reason, we only summarize the basic strategy. \begin{proof} For each $n=6, 7, 8$, we first compute using \texttt{bergman} a purported set of minimal coset representatives for $\mathcal E_{E_{n-1}}$ in $\mathcal E_{E_n}$ (note $E_5=D_5$) assuming that the only relations are the quadratic, braid, and claw relations. This gives an upper bound on $\mathcal H_{E_n}(t)/\mathcal H_{E_{n-1}}(t)$. We then use Algorithm~\ref{alg-mcr} to compute a subset of the minimal coset representatives for $\mathcal E_{E_{n-1}}$ in $\mathcal E_{E_n}$. This gives a lower bound on $\mathcal H_{E_n}(t)/\mathcal H_{E_{n-1}}(t)$. In each case, the upper and lower bounds coincide, giving the desired Hilbert series. \end{proof} This method does not work for $E_9$ (also known as the affine Dynkin diagram $\tilde{E}_8$) since $\mathcal E_{E_9}$ appears to require relations other than the quadratic, braid, and claw relations and is, in any case, beyond our current computational abilities. \subsection{Connection to Weyl groups} If we combine the results on $A_n$, $D_n$, and $E_n$ ($n \leq 8$) with known results about Weyl groups, we arrive at the following theorem. \begin{thm}\label{thm-weyl} Let $G$ be the graph of a simply-laced Dynkin diagram. Then the relations in $\mathcal E_G$ are generated by quadratic, braid, and claw relations, and \[\mathcal H_G(t) = \frac{W_G(t)}{C_G(t)},\] where $W_G(t)$ is the Poincar\'e series of the corresponding Weyl group $W$ and $C_G(t)$ is the characteristic polynomial of a Coxeter element in $W$. Moreover, $\dim \mathcal E_G = |W|/f$, where $f=C_G(1)$ is the index of connection (the index of the root lattice in the weight lattice). \end{thm} However, we know of no explanation of this fact that does not require explicitly computing the requisite Hilbert series first. \section{The case ${\affineA{n-1}}$} In this section, we will describe $\mathcal E_{{\affineA{n-1}}}$, where ${\affineA{n-1}}$ is the cycle on $n$ vertices (named after the affine Dynkin diagram). Label the edges $\mathsf{0}, \mathsf{1}, \dots, \mathsf{n-1}$ in order as shown in Figure~\ref{fig-cyc}. For convenience, we will take these labels modulo $n$ so that for any integers $i$ and $j$, the edges $\mathsf{i}$ and $\mathsf{j}$ are equal if and only if $i \equiv j \pmod n$. \begin{figure} \begin{center} \begin{tikzpicture}[scale = 1.5] \node[v] (0) at (90:1){}; \node[v] (1) at (54:1){}; \node[v] (2) at (18:1){}; \node[v] (3) at (-18:1){}; \node[v] (k-1) at (-90:1){}; \node[v] (k) at (-126:1){}; \node[v] (k+1) at (-162:1){}; \node[v] (n-1) at (126:1){}; \draw[thick, ->] (0)--node[above]{${\scriptstyle\mathsf{0}}$}(1); \draw[thick, ->] (1)--node[right]{${\scriptstyle\mathsf{1}}$}(2); \draw[thick, ->] (2)--node[right]{${\scriptstyle\mathsf{2}}$}(3); \draw[thick, ->] (0)--node[below left]{${\scriptstyle\mathsf{a}}$}(2); \draw[thick, ->] (n-1)--node[above, near start]{${\scriptstyle\mathsf{n-1}}$}(0); \draw[thick, ->] (0)--node[left]{${\scriptstyle\mathsf{b}}$}(k); \draw[thick, ->] (k-1)--node[below left, near start]{${\scriptstyle\mathsf{n-k-1}}$}(k); \draw[thick, ->] (k)--node[below left]{${\scriptstyle\mathsf{n-k}}$}(k+1); \draw[thick, loosely dotted] (-36:1) to [bend left = 18] (-72:1) (180:1) to [bend left = 18](144:1); \end{tikzpicture} \end{center} \caption{\label{fig-cyc} The $n$-cycle ${\affineA{n-1}}$ with edges labeled $\mathsf{0}, \dots, \mathsf{n-1}$. The additional edges $\mathsf{a}$ and $\mathsf{b}$ will be used in the proofs of Lemma~\ref{l elem sym vanish} and Proposition~\ref{p elem sym vanish}.} \end{figure} Since the line graph of ${\affineA{n-1}}$ is isomorphic to itself, Proposition~\ref{p Coxeter to FK} shows that $\mathcal E_{\affineA{n-1}}$ is a quotient of the nil-Coxeter algebra $\widetilde\mathcal N_n$ of type ${\affineA{n-1}}$. The main result of this section will be to describe the kernel of the quotient map explicitly as well as a set of minimal coset representatives for $\mathcal E_{A_n}$ in $\mathcal E_{\affineA{n-1}}$. As a corollary, we will obtain the following result. (For a more precise statement, see Theorem~\ref{t cycle}.) \begin{thm} \label{thm-cyc} The algebra $\mathcal E_{\affineA{n-1}}$ has a presentation consisting of quadratic relations, braid relations, and $n-1$ additional relations of degrees $k(n-k)$ for $1 \leq k \leq n-1$. The Hilbert series of $\mathcal E_{\affineA{n-1}}$ is given by \[ \mathcal H_{\affineA{n-1}}(t)=[n]\cdot\prod^{n-1}_{k=1}[k(n-k)]. \] In particular, $\dim \mathcal E_{\affineA{n-1}} = n! \cdot (n-1)!$, and the top degree of $\mathcal E_{\affineA{n-1}}$ is $\binom{n+1}{3}$. \end{thm} It will often be more convenient to work with an extended version of $\mathcal E_{\affineA{n-1}}$ defined as follows: $\widehat{\mathcal E}_{\affineA{n-1}}$ is the twisted algebra $\Pi\cdot\mathcal E_{\affineA{n-1}}$ generated by $\Pi\cong\mathbb Z/ n\mathbb Z$ and $\mathcal E_{\affineA{n-1}}$, where the generator $\pi \in \Pi$ satisfies $\pi \mathsf{i}=\mathsf{(i+1)}\pi$. Enlarging the algebra $\mathcal E_{\affineA{n-1}}$ to $\widehat{\mathcal E}_{\affineA{n-1}}$ is quite harmless---if $B$ is any basis of $\mathcal E_{\affineA{n-1}}$, then $\{\pi^i b\mid 0 \leq i \leq n-1,\; b\in B\}$ is a basis of $\widehat{\mathcal E}_{\affineA{n-1}}$. We may think of $\pi$ as having degree 0 and $\S_n$-degree equal to the automorphism of ${\affineA{n-1}}$ sending each vertex to the next adjacent vertex clockwise. Hence $\sigma_\pi(\mathsf{i}) = \mathsf{i+1} = \pi \mathsf{i} \pi^{-1}$. We begin by presenting some background on the (extended) affine symmetric group. We will then explicitly describe the relations of $\mathcal E_{\affineA{n-1}}$. This will allow us to find minimal coset representatives for $\mathcal E_{A_n} = \mathcal N_n$ in $\mathcal E_{\affineA{n-1}}$ and thereby prove the main result. \subsection{Background} Here we review some background on the geometry of the extended affine symmetric group. What follows is a somewhat cursory treatment; for a more thorough treatment of this material, see \cite{Haiman}. The \emph{affine symmetric group} $(\widetilde\S_n, K)$ is the Coxeter group with simple reflections $K=\{s_0, s_1, \dots, s_{n-1}\}$ whose Dynkin diagram is an $n$-cycle. Here $\S_n = (\S_n, S)$ is the symmetric group generated by $S=\{s_1, \dots, s_{n-1}\}$. The \emph{extended affine symmetric group} $\ensuremath{\widehat \S_n}$ is the semidirect product $\Pi \ltimes \ensuremath{\widetilde{\S}_n}$, where $\Pi$ is the cyclic group of order $n$ generated by $\pi$ satisfying $\pi s_i = s_{i+1}\pi$ (indices taken modulo $n$). The groups $\S_n$ and $\ensuremath{\widetilde{\S}_n}$ can be realized as affine reflection groups as follows. Let $\{\epsilon_1, \dots, \epsilon_n\}$ be the standard basis of $\mathbb R^n$ with the usual inner product $(\cdot,\cdot)$. Let $V \subset \mathbb R^n$ be the subspace spanned by $\alpha_i = \epsilon_i-\epsilon_{i+1}$ for $i = 1, \dots, n-1$. We identify $V$ in the natural way with $\mathbb R^n/\mathbb R\varepsilon$, where $\varepsilon=\epsilon_1+\cdots+\epsilon_n$. The $\alpha_i$ are the simple roots of the root system $\Phi = \Phi^+ \cup \Phi^-$ of type $A_{n-1}$, where $\Phi^+ = \{\epsilon_i-\epsilon_j \mid 1 \leq i < j \leq n\}$. For a root $\alpha \in \Phi$ and $k \in \mathbb Z$, denote by $h_{\alpha- k\delta}$ the (affine) hyperplane $\{x \in V \mid (x, \alpha)=k\}$, and let $s_{\alpha-k\delta}$ be the reflection over $h_{\alpha-k\delta}$. Then the map sending $s_i$ to the reflection $s_{\alpha_i}$ gives a faithful representation of $\S_n$. Denoting the highest root by $\bar\alpha = \epsilon_1-\epsilon_n$ and sending $s_0$ to the reflection $s_{\alpha_0}$ over the hyperplane $h_{\alpha_0} = h_{-\bar\alpha+\delta}=\{x \in V \mid (x, \epsilon_1-\epsilon_n)=1\}$ extends this to a faithful representation of $\widetilde\S_n$. The connected components of $V-\bigcup_{\alpha \in \Phi^+} h_{\alpha}$ are called \emph{chambers}, and the connected components of $V-\bigcup_{\alpha\in\Phi^+, k\in \mathbb Z} h_{\alpha-k\delta}$ are called \emph{alcoves}. The actions of $\S_n$ and $\ensuremath{\widetilde{\S}_n}$ are simply transitive on chambers and alcoves, respectively. We define \begin{align*} \mathbf C_0 &= \{x \in V \mid (\alpha_i, x)>0, \text{ for } i=1, \dots, n-1\}\\ \mathbf A_0 &= \{x \in \mathbf C_0 \mid (\bar\alpha, x)<1\} \end{align*} to be the \emph{fundamental chamber} and \emph{fundamental alcove}. A point in the closure of $\mathbf C_0$ is called \emph{dominant}. The set of all affine transformations that preserve the set of alcoves can be identified with $\ensuremath{\widehat \S_n}$: let $Y$ denote the weight lattice $\mathbb Z^n/\mathbb Z\varepsilon \subset V$. Then $\ensuremath{\widehat \S_n} = Y \rtimes \S_n$, where elements of $Y$ are treated as translations. In this context, $\ensuremath{\widetilde{\S}_n} = \mathbb Z\Phi \rtimes \S_n$, where $\mathbb Z\Phi \subset Y$ is the root lattice. We will denote elements $Y \subset \ensuremath{\widehat \S_n}$ using the multiplicative notation $y^\lambda = y_1^{\lambda_1}\cdots y_n^{\lambda_n}$ for $\lambda = \lambda_1\epsilon_1 + \cdots + \lambda_n\epsilon_n \in Y$. Note that $y_1y_2\cdots y_n = \mathrm{id}$. Here $\Pi$ is the stabilizer of $\mathbf A_0$, and we will take $\pi = y_1s_1s_2\cdots s_{n-1}$ as a generator of $\Pi$. Recall that if $w$ is an element of a Coxeter group, then the \emph{length} $\ell(w)$ of $w$ is the minimal length of an expression for $w$ as a product of simple reflections. For $w \in \ensuremath{\widehat \S_n}$, we set $\ell(w) = \ell(v)$, where $v \in \ensuremath{\widetilde{\S}_n}$ is the unique element such that $w = \pi^k v$ for some $k$. Equivalently, $\ell(w)$ is the number of hyperplanes $h_{\alpha-k\delta}$ separating $\mathbf A_0$ from $w^{-1}(\mathbf A_0)$. We will sometimes abuse notation by identifying $\S_n$, $\ensuremath{\widetilde{\S}_n}$, and $\ensuremath{\widehat \S_n}$ with their nil-Coxeter counterparts $\mathcal N_n$, $\ensuremath{\widetilde{\mathcal N}}_n$, and $\ensuremath{\widehat{\mathcal N}}_n$. Note that a monomial $y^\lambda$, when considered as an element of $\ensuremath{\widehat{\mathcal N}}_n$, is assumed to be a reduced expression (and therefore nonzero). We will write $\widetilde\Theta\colon \ensuremath{\widetilde{\mathcal N}}_n \to \mathcal E_{\affineA{n-1}}$ and $\widehat\Theta\colon \ensuremath{\widehat{\mathcal N}}_n \to \widehat \mathcal E_{\affineA{n-1}}$ for the canonical surjections sending $s_i \mapsto \mathsf{i}$. We will sometimes abuse notation by writing $\mathsf{i}$ for $s_i \in \mathcal N_n$ or omitting $\widetilde\Theta$ or $\widehat\Theta$ when convenient. We will also write $w_0$ for the longest element of $\S_n$. \subsection{Relations} In this section, we will describe the extra relations that occur in $\mathcal E_{\affineA{n-1}}$. Consider the translations $y_1, \dots, y_n \in \widehat\S_n$ as described above. Let us write each translation $y_{i_1}y_{i_2}\cdots y_{i_k} \in \widehat\S_n$ (for $1 \leq i_1<\cdots <i_k\leq n$) as a reduced word---we will see below that this has length $k(n-k)$. Let $e_k(y_1, \dots, y_n)$ be the sum of these words in $\widehat\mathcal N_n$. \begin{proposition}\label{p elem sym vanish} The image of $e_k(y_1, \dots, y_n)$ vanishes in $\widehat{\mathcal E}_{\affineA{n-1}}$ for $k = 1, \dots, n-1$. Therefore, in $\mathcal E_{\affineA{n-1}}$, \[R_k = R_k({\affineA{n-1}}) = \pi^{-k} e_k(y_1, \dots, y_n) = 0.\] \end{proposition} Note that $\operatorname{rev}(R_k) = R_{n-k}$. Before we prove this result, we will give an explicit description of $R_k$ in terms of the generators of $\mathcal E_{{\affineA{n-1}}}$. Consider a \emph{Grassmannian permutation} $w \in \S_n$ with $w_1<w_2<\dots<w_k$ and $w_{k+1}<w_{k+2}<\dots<w_n$, i.e., $w$ has at most one descent appearing at position $k$. Equivalently, $w$ is a minimal coset representative of $\S_k \times \S_{n-k}$ in $\S_n$. We associate to $w$ the partition $\lambda = (w_k-k, \dots, w_1-1)$ and write $w = \gamma(\lambda)$. This gives a bijection between such $w$ and partitions $\lambda$ with at most $k$ parts, each of size at most $n-k$. There is a nice description of the set of reduced words of a Grassmannian permutation called the $\delta$-rule, due to Winkel \cite{Winkel}. For a partition $\lambda$ as above, let $\delta_\lambda$ be the tableau of shape $\lambda$ with entry $k+j-i$ in the box at row $i$, column $j$. The $\delta$-rule says that the reduced words of $\gamma(\lambda)$ are obtained by, starting with the tableau $\delta_\lambda$, successively removing outer corners and recording the entries until all the entries are removed. For example, $s_6s_2s_4s_5s_3s_4$ is the reduced expression for $\gamma(3,2,1,0)$ corresponding to the sequence \[ \begin{array}{cc} {\tiny \delta_{(3,2,1,0)}\;=\;\tableau{4&5&6\\3&4\\2}\quad\to\quad \tableau{4&5\\3&4\\2} \quad\to\quad \tableau{4&5\\3&4} \quad\to\quad \tableau{4&5\\3} \quad\to\quad \tableau{4\\3} \quad\to\quad \tableau{4}}\,. \end{array} \] Given a partition $\mu \subset \lambda$, define $\gamma(\lambda/\mu) = \gamma(\lambda) \gamma(\mu)^{-1}$, and let $\delta_{\lambda/\mu}$ be the skew subtableau of $\delta_\lambda$ of shape $\lambda/\mu$. It follows from the $\delta$-rule that the reduced expressions for $\gamma(\lambda/\mu)$ are obtained by removing entries of $\delta_{\lambda/\mu}$ just like the $\delta$-rule. It is easy to check from the definitions that $y_1 \cdots y_k = \pi^{k}\cdot \gamma(\Omega)$, where $\Omega=(n-k)^k$, the partition with $k$ parts of size $n-k$. Now given any $1 \leq i_1 < i_2 < \dots < i_k \leq n$, let $\lambda = (i_k-k, \dots, i_1-1)$. Then $y_{i_1}\cdots y_{i_k}$ has the reduced factorization \begin{equation} \label{e y reduced} y_{i_1}\cdots y_{i_k} = \gamma(\lambda) y_1 \cdots y_k \gamma(\lambda)^{-1} = \gamma(\lambda) \cdot \pi^k \cdot \gamma(\Omega/\lambda). \tag{$\spadesuit$} \end{equation} Therefore \[ \pi^{-k} y_{i_1}\cdots y_{i_k} = \pi^{-k} \gamma(\lambda) \pi^k \cdot \gamma(\Omega/\lambda) = \sigma_{\pi}^{-k}(\gamma(\lambda)) \cdot \gamma(\Omega/\lambda). \] Summing over all $\lambda \subset \Omega$ gives $R_k$. \begin{ex} If $n=5$, then \begin{align*} R_1 &= \mathsf{4321} + \mathsf{0\cdot 432} + \mathsf{10 \cdot 43} + \mathsf{210 \cdot 4} + \mathsf{3210},\\ R_2 &= \mathsf{342312} + \mathsf{0 \cdot 34231} + \mathsf{10 \cdot 3421} + \mathsf{40 \cdot 3423} + \mathsf{210 \cdot 321}\\ & \qquad + \mathsf{410 \cdot 342} + \mathsf{4210 \cdot 32} + \mathsf{0410 \cdot 34} + \mathsf{04210 \cdot 3} + \mathsf{104210},\\ R_3 &= \mathsf{234123} + \mathsf{0 \cdot 23412} + \mathsf{10 \cdot 2312} + \mathsf{40 \cdot 2341} + \mathsf{410 \cdot 231}\\ & \qquad + \mathsf{340 \cdot 234} + \mathsf{0410 \cdot 21} + \mathsf{3410 \cdot 23} + \mathsf{30410 \cdot 2} + \mathsf{430410},\\ R_4 &= \mathsf{1234}+ \mathsf{0 \cdot 123} + \mathsf{40 \cdot 12} + \mathsf{340 \cdot 1} + \mathsf{2340}. \end{align*} See also Example~\ref{ex-a3tilde} for the case $n=4$. \end{ex} We now prove the following special case of Proposition~\ref{p elem sym vanish}. \begin{lemma} \label{l elem sym vanish} The following relations hold in $\mathcal E_{{\affineA{n-1}}}$: \begin{align*} R_1&= \sum_{i=0}^{n-1} \mathsf{(i-1)(i-2)\cdots 0(n-1)(n-2)\cdots(i+1)}=0,\\ R_{n-1}&= \sum_{i=0}^{n-1} \mathsf{(i+1)(i+2)\cdots(n-1)01\cdots(i-1)}=0. \end{align*} \end{lemma} \begin{proof} We prove the $R_1$ relation, the $R_{n-1}$ relation being similar. We induct on $n$. For the base case $n=3$, $R_1 = \mathsf{21} + \mathsf{02} + \mathsf{10}$ is one of the quadratic relations of $\mathcal E_{\tilde{A}_2}$. For $n>3$, consider the additional edge $\mathsf{a}$ as drawn in Figure~\ref{fig-cyc}, and let $C$ be the cycle $\mathsf{2}, \mathsf{3}, \dots, \mathsf{n-1}, \mathsf{a}$. For $i=2, 3, \dots, n-1$, let \[X_i = \mathsf{(i-1)(i-2)\cdots 2\cdot a\cdot (n-1)(n-2)\cdots (i+1)}.\] Then by induction, $R_1(C) = \mathsf{(n-1)(n-2) \cdots 2} + \sum_{i=2}^{n-1} X_i=0$. Using the commutation relations and the quadratic relation $\mathsf{0a} + \mathsf{a1} = \mathsf{10}$, \[\mathsf{0}\cdot X_i + X_i \cdot \mathsf{1} = \mathsf{(i-1)\cdots 2}\cdot (\mathsf{0a}+\mathsf{a1}) \cdot \mathsf{(n-1)\cdots (i+1)} = \mathsf{(i-1)\cdots 210(n-1) \cdots (i+1)}.\] It now follows easily that $R_1({\affineA{n-1}})=\mathsf{0} \cdot R_1(C) + R_1(C) \cdot \mathsf{1}=0$, as desired. \end{proof} Using Lemma~\ref{l elem sym vanish}, we can complete the proof of Proposition~\ref{p elem sym vanish}. \begin{proof}[Proof of Proposition \ref{p elem sym vanish}] Lemma \ref{l elem sym vanish} establishes the cases $k=1$ and $k= n-1$. Suppose that $n>3$ and $k \in \{2,\dots, n-2\}$. To prove that $e_k(y_1,\dots, y_n)$ is zero in $\widehat{\mathcal E}_{\affineA{n-1}}$, write \[ e_k(y_1,\dots, y_n) = e_k(y_1, \dots, y_{n-1}) + e_{k-1}(y_1, \dots, y_{n-1})y_n \] We will show that \begin{equation}\label{e ek split} - e_k(y_1, \dots, y_{n-1}) = e_{k-1}(y_1, \dots, y_{n-1})y_n. \tag{$*$} \end{equation} in $\widehat{\mathcal E}_{{\affineA{n-1}} \cup \{\mathsf{b}\}}$, where $\mathsf{b}$ is the additional edge shown in Figure~\ref{fig-cyc}. Maintain the notation of \eqref{e y reduced} with $\Omega = (n-k)^k$ and $\Omega^L = (n-k-1)^k$. The left side of \eqref{e ek split} contains terms $y_{i_1}\cdots y_{i_k}$ for which $i_k \neq n$. Then $\lambda_1 = i_k-k < n-k$, so $\gamma(\Omega/\lambda) = \mathsf{(n-k)(n-k+1)\cdots (n-1)} \cdot \gamma(\Omega^L/\lambda)$ is a reduced factorization (corresponding to first removing the rightmost column of $\delta_{\Omega/\lambda}$), which yields \begin{equation} \label{e y reduced ik less n} -y_{i_1} \cdots y_{i_k} = -\gamma(\lambda) \pi^k \mathsf{(n-k)(n-k+1)\cdots (n-1)} \gamma(\Omega^L/\lambda).\tag{$**$} \end{equation} Using the $R_k$ relation of the cycle $\mathsf{n-k}, \dots, \mathsf{n-1}, \mathsf{b}$, which is \[ -\mathsf{(n-k)\cdots (n-1)} = \sum_{i=1}^k \mathsf{(n-i+1)\cdots (n-1) \cdot b \cdot (n-k) \cdots (n-i-1)}, \] we can expand the right side of \eqref{e y reduced ik less n} into $k$ monomials. The left side of \eqref{e ek split} then expands into $k\binom{n-1}{k}$ monomials of the form \begin{multline*} \gamma(\lambda) \pi^k \mathsf{(n-i+1) \cdots (n-1) \cdot b \cdot (n-k) \cdots (n-i-1)} \gamma(\Omega^L/\lambda)\\ =\gamma(\lambda) \mathsf{(k-i+1) \cdots (k-1)}\cdot \pi^k \mathsf{ b \cdot (n-k) \cdots (n-i-1)} \gamma(\Omega^L/\lambda), \end{multline*} where $\lambda \subset \Omega^L$ and $1 \leq i \leq k$. But $\gamma(\lambda)\mathsf{(k-i+1)\cdots (k-1)}$ ranges over all minimal coset representatives $\S_{n-1}^J$ for the parabolic subgroup $\S_{k-1} \times \S_1 \times \S_{n-k-1}$ (generated by $J=\{s_1, \dots, s_{k-2}, s_{k+1}, \dots, s_{n-2}\}$) in $\S_{n-1}$: indeed, $\gamma(\lambda)$ are the minimal coset representatives for $\S_k \times \S_{n-k-1}$ in $\S_{n-1}$, while $\mathsf{(k-i+1)\cdots (k-1)}$ are the ones for $\S_{k-1} \times \S_1$ in $\S_k$. A similar argument shows that each $\mathsf{(n-k)\cdots (n-i-1)}\gamma(\Omega^L/\lambda)$ is a reduced expression. Hence \[-e_k(y_1, \dots, y_{n-1}) = \sum_{w \in \S_{n-1}^J} w\pi^k\mathsf{b}w',\] where $w'$ is the unique element of $\S_n$ such that the $\S_n$-degree of $w\pi^k \mathsf{b}w'$ is the identity. Similarly, the right side of \eqref{e ek split} contains terms $y_{i_1}\cdots y_{i_k}$ for which $i_k = n$ so that $\lambda_1 = n-k$. Then $\lambda^B=(\lambda_2, \dots, \lambda_k) \subset \Omega^B = (n-k)^{k-1}$, so we find (by removing the top row of $\delta_\lambda$ last) that $\gamma(\lambda) = \gamma(\lambda^B) \cdot \mathsf{(n-1)(n-2) \cdots k}$. Thus \begin{align*} y_{i_1}\cdots y_{i_k} &= \gamma(\lambda^B)\mathsf{(n-1)(n-2)\cdots k}\pi^k \gamma(\Omega^B/\lambda^B)\\ &= \gamma(\lambda^B)\pi^k\mathsf{(n-k-1)(n-k-2)\cdots 0} \gamma(\Omega^B/\lambda^B). \end{align*} Then using the $R_1$ relation of the cycle $\mathsf{0}, \mathsf{1}, \dots, \mathsf{n-k-1}, \mathsf{b}$, the right side of \eqref{e ek split} expands into $(n-k)\binom{n-1}{k-1}$ monomials of the form \begin{multline*} \gamma(\lambda^B)\pi^k \mathsf{(i-1) \cdots 0\cdot b \cdot (n-k-1) \cdots (i+1)}\gamma(\Omega^B/\lambda^B) \\ = \gamma(\lambda^B) \mathsf{(k+i-1) \cdots k} \cdot \pi^k \mathsf{b \cdot(n-k-1) \cdots (i+1)}\gamma(\Omega^B/\lambda^B) \end{multline*} for $\lambda^B \subset \Omega^B$ and $0 \leq i \leq n-k-1$. But as above, $\gamma(\lambda^B) \mathsf{(k+i-1) \cdots k}$ ranges over $\S_{n-1}^J$ since $\gamma(\lambda^B)$ are the minimal coset representatives for $\S_{k-1} \times \S_{n-k}$ in $\S_{n-1}$ while $\mathsf{(k+i-1) \cdots k}$ are the ones for $\S_1 \times \S_{n-k-1}$ in $\S_{n-k}$. Thus as above we have \[e_{k-1}(y_1, \dots, y_n)y_n = \sum_{w \in \S_{n-1}^J} w\pi^k\mathsf{b}w' = -e_k(y_1, \dots, y_{n-1}).\qedhere\] \end{proof} \begin{example} We illustrate the proof of Proposition \ref{p elem sym vanish} in the case $n = 5$, $k=2$. There, using the relation $\mathsf{210} = \mathsf{10b} + \mathsf{0b2} + \mathsf{b21}$: \begin{align*} \pi^{-2} e_1(y_1, y_2, y_3, y_4)y_5 \quad = & \phantom{{}+{}} \mathsf{210321} + \mathsf{421032} + \mathsf{042103} + \mathsf{104210}\\[1.5mm] =&\phantom{{}+{}}\mathsf{10b321 + 410b32 + 0410b3 + 10410b}\\ &+\mathsf{0b2321 + 40b232 + 040b23 + 1040b2}\\ &+\mathsf{b21321 + 4b2132 + 04b213 + 104b21}, \end{align*} Similarly, using the relation $\mathsf{-34} = \mathsf{4b} + \mathsf{b3}$: \begin{align*} -\pi^{-2}e_2(y_1, y_2, y_3, y_4) \quad = & - \mathsf{342312 - 034231 - 103421 - 403423 - 410342 - 041034}\\[1.5mm] =&\phantom{{}+{}}\mathsf{4b2312 + 04b231 + 104b21 + 404b23 + 1404b2 + 04104b}\\ &+ \mathsf{b32312 + 0b3231 + 10b321 + 40b323 + 140b32 + 0410b3}\\[1.5mm] =&\phantom{{}+{}}\mathsf{4b2132 + 04b213 + 104b21 + 040b23 + 1040b2 + 10410b}\\ &+ \mathsf{b21321 + 0b2321 + 10b321 + 40b232 + 410b32 + 0410b3}. \end{align*} The last equality uses only braid and commutation relations. We see directly in this case that the twelve terms that appear in each case are the same. \end{example} \subsection{Primitive elements} Let $I_0 \subset \widehat\mathcal N_n$ be the (two-sided) ideal generated by the relations $R_k$, or equivalently by $e_k(y_1, \dots, y_n)$. By Proposition~\ref{p elem sym vanish}, $\widehat\mathcal E_{\affineA{n-1}}$ is a quotient of $\widehat \mathcal N_n/I_0$. To show that they are isomorphic, we first describe a basis of $\widehat \mathcal N_n/I_0$ in terms of a subset of $\widehat\S_n$ which we call \emph{primitive elements}, previously studied in \cite{LJantzen, Xi, SW, BlasiakFactor, BlasiakCyclage}. These primitive elements will turn out to form a set of minimal coset representatives of $ \mathcal N_n = \mathcal E_{A_n}$ in $\widehat\mathcal E_{\affineA{n-1}}$. We will work with a geometric description of primitive elements from \cite{LJantzen}. Let a \emph{box} be a connected component of $H - \bigcup_{i \in [n-1], k \in \mathbb Z} h_{\alpha_i - k\delta}$. We denote by $\mathbf{B}_0$ the box containing $\mathbf{A}_0$, which lies between the hyperplanes $h_{\alpha_i}$ and $h_{\alpha_i - \delta}$ for $i \in [n-1]$. An element $w$ in $\ensuremath{\widehat \S_n}$ is \emph{primitive} if $w^{-1}(\mathbf{A}_0) \subseteq \mathbf{B}_0$. Let $D^S$ denote the set of primitive elements of $\ensuremath{\widehat \S_n}$. Note that since $\mathbf B_0$ is contained in the fundamental chamber $\mathbf C_0$, every primitive element is an element of $\widehat\S_n^S$, that is, a (left) minimal coset representative of $\S_n$ in $\widehat\S_n$. The elements of $D^S$ are in bijection with elements of $\S_n$: for any $w \in \S_n$, there is a unique $y^\lambda$ such that $y^\lambda w \in D^S$. Also, since $\pi$ stabilizes $\mathbf A_0$, $v \in D^S$ if and only if $\pi v \in D^S$. \begin{example}\label{ex SL3} The primitive elements of $\widehat{\S}_3$, expressed as products of simple reflections (top line) and using the fact that $\widehat{S}_3 = Y\rtimes \S_3$ (bottom line), are: \[ { \setlength\arraycolsep{15pt} \begin{array}{cccccc} \mathrm{id} & \pi & \pi^2 & \pi s_0& \pi^2 s_0 & \pi^3 s_0\\ \mathrm{id}&y_1s_1s_2&y_1y_2s_2s_1&y_2s_2&y_1y_3s_1&y_1^2y_2s_1s_2s_1 \end{array} } \] \end{example} \subsection{Spanning} We will now show how to construct a spanning set of $\widehat \mathcal N_n/I_0$ from the primitive elements. We will see in Theorem~\ref{t cycle} that it is actually a basis. We will first need the following lemma about reduced factorizations. \begin{lemma} \label{lemma-redfac} Suppose $\lambda \in Y$ is a dominant weight and $w \in \widehat\S_n^S$. Then $y^\lambda \cdot w^{-1}$ is a reduced factorization. \end{lemma} \begin{proof} It suffices to show that no hyperplane that separates $y^{-\lambda}(\mathbf A_0)$ and $\mathbf A_0$ also separates $\mathbf A_0$ and $w^{-1}(\mathbf A_0)$. Since $(-\lambda, \alpha) \leq 0$ for any $\alpha \in \Phi^+$, any hyperplane separating $y^{-\lambda}(\mathbf A_0)$ from $\mathbf A_0$ has the form $h_{\alpha-k\delta}$, where $\alpha\in \Phi^+$ and $k \leq 0$. Likewise, $w^{-1}(\mathbf A_0)$ lies in $\mathbf C_0$, so any hyperplane separating $\mathbf A_0$ and $w^{-1}(\mathbf A_0)$ has the form $h_{\alpha-k\delta}$ for some $\alpha \in \Phi^+$ and $k>0$. \end{proof} Let $\mathcal N_n^d$ be the degree $d$ part of $\mathcal N_n$, and let $\mathcal N_n^+ = \bigoplus_{d>0} \mathcal N_n^d$. \begin{lemma} \label{lemma-primitive span} If $x \in \widehat\S_n^S$ but $x \not\in D^S$, then $x \in I_0+\widehat\mathcal N_n\mathcal N_n^+$. \end{lemma} \begin{proof} By our choice of $x$, $x^{-1}(\mathbf A_0)$ lies in a box $\mathbf B = y^\lambda(\mathbf B_0) \neq \mathbf B_0$ contained in the fundamental chamber. Hence by Lemma~\ref{lemma-redfac}, $x^{-1}$ has a reduced factorization of the form $y^\lambda\cdot v^{-1}$, where $\lambda$ is a nonzero dominant weight and $v \in D^S$. Thus it suffices to show that the image of $y^{\lambda}$ lies in $I_0 + \mathcal N_n^+\widehat\mathcal N_n$. Choose any $k$ such that $\lambda_k>\lambda_{k+1}$. The only term of $e_k(y_1, \dots, y_n)$ that corresponds to a dominant weight is $y^\mu = y_1\cdots y_k$; hence all other terms lie in $\mathcal N_n^+\widehat\mathcal N_n$ and so $y^\mu \in I_0 + \mathcal N_n^+\widehat\mathcal N_n$. By Lemma~\ref{lemma-redfac}, $y^{\lambda} = y^\mu \cdot y^{\lambda-\mu}$ is a reduced factorization, so the result follows. \end{proof} \begin{prop} \label{prop-primitive span} The image of $\{vw \mid v \in D^S, w \in \mathcal N_n\}$ spans $\widehat\mathcal N_n/I_0$. \end{prop} \begin{proof} Since $\{xw \mid x \in \widehat\S_n^S, w \in \mathcal N_n\}$ is a basis for $\widehat\mathcal N_n$, it spans $\widehat\mathcal N_n/I_0$. Suppose $w \in \mathcal N_n^{d}$. By Lemma~\ref{lemma-primitive span}, if $x \not \in D^S$, then $xw \in I_0 + \widehat\mathcal N_n\mathcal N_n^+w \subset I_0 + \widehat\mathcal N_n\mathcal N_n^{d+1}$. Hence $\widehat\mathcal N_n\mathcal N_n^d / \widehat\mathcal N_n\mathcal N_n^{d+1}$ is spanned by $\{vw \mid v \in D^S, w \in \mathcal N_n^d\}$ and elements of $I_0$. (In particular, $\widehat\mathcal N_n\mathcal N_n^{\ell(w_0)}$ is spanned by $\{vw_0 \mid v \in D^S\}$.) The desired result then follows from a straightforward induction on $\ell(w_0)-d$. \end{proof} \subsection{Linear independence} In this section, we will prove that the images of the primitive elements under $\widehat\Theta$ are linearly independent as (left) minimal coset representatives of $\mathcal E_{A_n} = \mathcal N_n$ in $\widehat \mathcal E_{\affineA{n-1}}$ by computing appropriate evaluations of the bilinear form $\langle \cdot, \cdot \rangle$. First we recall some facts about Bruhat order in $\widetilde\S_n$. A product of simple reflections $s_{i_1}s_{i_2}\cdots s_{i_d}$ can be visualized by the alcove walk (see, e.g., \cite{Ram}) \[ \mathbf D_0 = \mathbf{A}_0,\quad \mathbf D_1 = s_{i_1}(\mathbf{A}_0),\quad \mathbf D_2 = s_{i_1}s_{i_2}(\mathbf{A}_0),\quad\dots,\quad \mathbf D_d = s_{i_1}s_{i_2}\cdots s_{i_d}(\mathbf{A}_0). \] Here $\mathbf D_j$ is the reflection of $\mathbf D_{j-1}$ across one of its facets, namely the one contained in the hyperplane $h_{\beta^j}= s_{i_1}\cdots s_{i_{j-1}}(h_{\alpha_{i_j}})$. Suppose we omit a simple reflection $s_{i_j}$ from the product $s_{i_1}\cdots s_{i_d}$, giving the product $s_{i_1}\cdots \widehat{s_{i_j}}\cdots s_{i_d}$. Then the resulting alcove walk is obtained from that of $s_{i_1}\cdots s_{i_d}$ by reflecting the second part of the walk across $h_{\beta^j}$. More precisely, its alcove walk is \[ \mathbf{D}_0, \dots,\,\mathbf{D}_{j-1},\, s_{\beta^j}(\mathbf{D}_{j+1}),\,\dots,\, s_{\beta^j}(\mathbf{D}_d). \] It is well known that $s_{i_1}\cdots s_{i_d}$ is a reduced expression if and only if its alcove walk never crosses a given hyperplane more than once. We claim that \begin{equation}\label{etext crossed twice} \parbox{14cm}{ If $s_{i_1}\cdots s_{i_d}$ and $s_{i_1}\cdots \widehat{s_{i_j}}\cdots s_{i_d}$ are reduced expressions, then $h_{\beta^j}$ is either the first or last hyperplane parallel to $h_{\beta^j}$ crossed in the alcove walk of $s_{i_1}\cdots s_{i_d}$. } \tag{$\diamondsuit$} \end{equation} If not, then the alcove walk of $s_{i_1}\cdots s_{i_d}$ crosses $h_{\beta^j+\delta}$ and $h_{\beta^j-\delta}$. But then the alcove walk of $s_{i_1}\cdots \widehat{s_{i_j}}\cdots s_{i_d}$ crosses one of them twice (since $s_{\beta^j}(h_{\beta^j+\delta}) = h_{\beta^j-\delta}$). \begin{lemma}\label{l rho pairing} Suppose $w = p_1p_2\cdots p_d \in \widetilde\S_n$ is a reduced expression with $p_1\cdots p_{\ell(w_0)} = w_0$ and $w^{-1}w_0 = p_dp_{d-1} \cdots p_{\ell(w_0)+1}$ primitive. If $j \in [d]$ is an index such that $p_1\cdots \widehat{p_j} \cdots p_d$ is a reduced expression and $p_j = \sigma_{p_{j+1}\cdots p_d}(p_d)$, then $j=d$. \end{lemma} \begin{proof} Since $p_j = \sigma_{p_{j+1}\cdots p_d}(p_d)$, it follows that $p_j \cdots p_{d-1}$ and $p_{j+1} \cdots p_{d}$ have the same $\S_n$-degree. Hence $p_1 \cdots \widehat{p_j} \cdots p_d$ and $p_1 \cdots p_{d-1}$ have the same $\S_n$-degree, so they differ by a translation in $\widetilde \S_n$. Defining $\beta^j$ as in the discussion above, we have that $p_1 \cdots \widehat{p_j} \cdots p_d = s_{\beta^j}w$ and $p_1 \cdots p_{d-1} = s_{\beta^d}w$, so the hyperplanes $h_{\beta^j}$ and $h_{\beta^d}$ are parallel. By \eqref{etext crossed twice} and the assumptions of the lemma, $h_{\beta^j}$ must be the first or last hyperplane parallel to $h_{\beta^j}$ crossed in the alcove walk of $p_1 p_2\cdots p_d$. If it is the last, then $d=j$ and we are done, so assume it is the first. Since $p_1\cdots p_{\ell(w_0)}$ is a reduced expression for $w_0$, the first $\ell(w_0)$ hyperplanes crossed in the alcove walk of $p_1\cdots p_d$ contain all $\ell(w_0)$ different hyperplane directions. Therefore, $j\leq \ell(w_0)$. Set $y=p_1\cdots \widehat{p_j}\cdots p_{\ell(w_0)}$, which is a reduced expression for $y$. Therefore $\ell(y)=\ell(w_0)-1$, so it follows that $y= s_cw_0$ for some simple reflection $s_c \in \S_n$. Hence $s_{\beta^j}w = p_1 \cdots \widehat{p_j} \cdots p_d = s_cw$, so $\beta^j = \alpha_c$. Since $w^{-1}w_0$ is primitive, $w_0w(\mathbf A_0) \subset \mathbf B_0$, or equivalently $w(\mathbf A_0) \subset w_0(\mathbf B_0)$. The region $w_0(\mathbf{B}_0)$ is also a box and lies between $h_{\alpha_i}$ and $h_{\alpha_i+\delta}$ for $i\in [n-1]$. We conclude that the only hyperplane separating $\mathbf{A}_0$ and $w(\mathbf{A}_0)$ and parallel to $h_{\alpha_c}$ is $h_{\alpha_c}$ itself. Hence $j=d$, as desired. \end{proof} We can now prove the following result along the same lines as the proof of Theorem~\ref{thm-a}. For $w, w' \in \widetilde \mathcal N_n$, write $\langle w, w' \rangle = \langle \widetilde\Theta(w), \widetilde\Theta(w') \rangle$. \begin{prop}\label{p primitive pairing} If $v \in \ensuremath{\widetilde{\mathcal N}}_n$ is primitive, then $\langle w_0\operatorname{rev} (v), vw_0\rangle=1$. \end{prop} \begin{proof} First note that if $p_1\cdots p_d$ is not a reduced expression in $\ensuremath{\widetilde{\S}_n}$, then $\langle p_1\cdots p_d, Q\rangle=0$ for all $Q\in \mathcal E_n$. (This is because $p_1\cdots p_d = 0$ in $\widetilde\mathcal N_n$ and hence also in $\mathcal E_{\affineA{n-1}}$.) Let $p_1\cdots p_d=w_0\operatorname{rev}(v)$ as in Lemma \ref{l rho pairing}. If $v = \mathrm{id}$, then $\langle w_0, w_0 \rangle = 1$ by the proof of Theorem~\ref{thm-a}. Otherwise, by Lemma \ref{lemma-leibniz}, Proposition \ref{prop-pairing}, and Lemma \ref{l rho pairing}, \begin{align*} \langle p_1\cdots p_d,\;p_d \cdots p_1\rangle &=\sum_{j=1}^d (p_j) (\sigma_{p_{j+1} \cdots p_d}\nabla_{p_d}) \cdot \langle p_1 \cdots \widehat{p_j} \cdots p_d,\; p_{d-1}p_{d-2}\cdots p_1 \rangle\\ &=\langle p_1\cdots p_{d-1},\; p_{d-1} \cdots p_1\rangle. \end{align*} But $p_1 \cdots p_{d-1} = w_0\operatorname{rev} (v')$ for some primitive element $v'$: since $v$ is primitive, the alcove walk of $\operatorname{rev}(v)$ does not leave $\mathbf B_0$, so neither does that of $\operatorname{rev}(v')$. The result then follows by induction. \end{proof} It is now straightforward to prove that the primitive elements satisfy an appropriate linear independence property. \begin{prop} \label{prop-primitive ind} The images of the primitive elements $D^S \subset \widehat\mathcal N_n$ under $\widehat\Theta$ form a subset of the (left) minimal coset representatives of $\mathcal N_n = \mathcal E_{A_n}$ in $\widehat\mathcal E_{\affineA{n-1}}$, i.e., they are linearly independent modulo $\widehat\mathcal E_{\affineA{n-1}} \mathcal E_{A_n}^+$. \end{prop} \begin{proof} We need to show that if $\sum_{v \in D^S} c_v\widehat\Theta(v) \in \widehat\mathcal E_{\affineA{n-1}} \mathcal E_{A_n}^+$ for some $c_v \in \mathbb Q$, then all the $c_v$ vanish. We may assume that all $v$ lie in $\widetilde\mathcal N_n$. Multiplying the above expression by $w_0$, we have that \begin{equation}\label{e linear ind FK} \sum_{ v\in D^S} c_{v} \widehat{\Theta} (v w_0) =0. \tag{$*$} \end{equation} Since all elements of $D^S$ have different $\S_n$-degree, $\langle w_0 \operatorname{rev}(v'), vw_0 \rangle=0$ unless $v=v'$, in which case it equals 1 by Lemma~\ref{l rho pairing}. Thus pairing \eqref{e linear ind FK} with $\widehat\Theta(w_0 \operatorname{rev}(v))$ gives $c_v=0$. \end{proof} \subsection{Proof of main theorem} Combining the results of the previous sections, we can now prove Theorem~\ref{thm-cyc}. It is an immediate consequence of the following more explicit result. \begin{thm} \label{t cycle} The algebra $\widehat\mathcal E_{\affineA{n-1}}$ is isomorphic to $\widehat\mathcal N_n/I_0$, where $I_0$ is the ideal generated by $e_k(y_1, \dots, y_n)$ for $k=1, \dots, n-1$. Moreover, $D^S$ is a set of (left) minimal coset representatives for $\mathcal N_n = \mathcal E_{A_{n}}$ in $\widehat\mathcal E_{\affineA{n-1}}$. \end{thm} \begin{proof} By Proposition~\ref{p elem sym vanish}, $\widehat\mathcal E_{\affineA{n-1}}$ is a quotient of $\widehat\mathcal N_n/I_0$. By Proposition~\ref{prop-primitive span}, $\{vw \mid v \in D^S, w \in \mathcal N_n\}$ spans $\widehat\mathcal N_n/I_0$, and by Proposition~\ref{prop-primitive ind} and Theorem~\ref{thm-subgraph}, its image is linearly independent in $\widehat\mathcal E_{\affineA{n-1}}$. It follows that it must be a basis of both. \end{proof} \begin{cor} The Hilbert series of $\mathcal E_{\affineA{n-1}}$ is given by \[ \mathcal{H}_{\affineA{n-1}}(t)=[n]\cdot\prod^{n-1}_{i=1}[i(n-i)]. \] \end{cor} \begin{proof} By Remark 1.5 and Lemma 2.2 of \cite{SW}, the length generating function for the subset $D^S\subset \ensuremath{\widehat{\mathcal N}}_n$ is given by \[ \sum_{v\in D^S }t^{\ell(v)}= n\cdot \prod_{i=1}^{n-1}\frac{1-t^{i(n-i)}}{1-t^{i}}. \] Hence by Theorem \ref{t cycle}, the Hilbert series of $\widehat{\mathcal E}_{\affineA{n-1}}$ is given by \[ \sum_{\substack{v \in D^S\\ w\in\ensuremath{\mathcal N}_n}} t^{\ell(v w)} = [n]!\cdot n\cdot \prod_{i=1}^{n-1}\frac{1-t^{i(n-i)}}{1-t^i} =n\cdot\prod_{i=1}^{n}\frac{1-t^i}{1-t}\cdot \prod_{i=1}^{n-1}\frac{1-t^{i(n-i)}}{1-t^i} =n\cdot [n]\cdot\prod_{i=1}^{n-1}[i(n-i)]. \] Since the Hilbert series of $\widehat{\mathcal E}_{\affineA{n-1}}$ is equal to $n$ times the Hilbert series of $\mathcal E_{\affineA{n-1}}$, the desired result follows. \end{proof} \begin{rmk}\label{r coinvariants} Let $I$ be the ideal of $\mathbb Q[y_1, \dots, y_n]$ generated by symmetric polynomials of positive degree. The quotient of $\mathbb Q [y_1, y_2, \dots, y_n]$ by $I$ is called the \emph{ring of coinvariants} and is known to have dimension $n!$. The reference \cite{BlasiakCyclage} identifies a subalgebra $\pH_n$ of the extended affine Hecke algebra $\eH_n$ of type $A$. The subalgebra $\pH_n$ is a $q$-analogue of $\mathbb Q\S_n \ltimes \mathbb Q [y_1,\dots,y_n]$ and inherits a canonical basis from that of $\eH_n$. Let $\mathcal{I}$ denote the two-sided ideal of $\pH_n$ generated by the symmetric polynomials of positive degree in the Bernstein generators. Corollary $6.7$ of \cite{BlasiakCyclage} states that the $\pH_n$-module $\pH_n \ensuremath{C^{\prime}}_{w_0}/ \mathcal{I}\ensuremath{C^{\prime}}_{w_0}$ has canonical basis $\{\ensuremath{C^{\prime}}_{vw_0} \mid v\in D^S\}$, where $\ensuremath{C^{\prime}}_w$ denotes the canonical basis element labeled by $w\in\ensuremath{\widehat \S_n}$. The basis $\{vw_0 \mid v\in D^S\}$ of $\ensuremath{\widehat{\mathcal N}}_n w_0/I_0 w_0$ is essentially the $q=0$ specialization of this canonical basis. (It is not exactly the $q=0$ specialization because the canonical basis of \cite{BlasiakCyclage} is for the $G=GL_n$ extended version of the affine Weyl group, whereas here we have used the $G=SL_n$ version.) \end{rmk} \section{Questions and conjectures} We conclude with some questions and conjectures to guide further research. \begin{question} What is the Hilbert series $\mathcal H_G(t)$ of $\mathcal E_G$? \end{question} Towards this end, we have a conjecture for the Hilbert series of the affine Dynkin diagrams $\tilde{D}_n$ as shown in Figure~\ref{fig-dntilde}. This conjecture was obtained by using Algorithm~\ref{alg-mcr} to compute minimal coset representatives for $\mathcal E_{D_n}$ in $\mathcal E_{\tilde D_n}$. Here, $[n]!! = [n][n-2][n-4]\cdots$. \begin{figure} \begin{center} \begin{tikzpicture} \draw (150:1)node[v]{}--(0,0)node[v]{}--(1,0)node[v]{}--(1.5,0) (-150:1)node[v]{}--(0,0); \draw[thick, loosely dotted] (1.5,0)--(2.5,0); \draw (2.5,0)--(3,0)node[v]{}--(4,0)node[v]{}--+(30:1)node[v]{} (4,0)--+(-30:1)node[v]{}; \end{tikzpicture} \end{center} \caption{\label{fig-dntilde}The affine Dynkin diagram $\tilde D_n$ with $n+1$ vertices.} \end{figure} \begin{conj} \label{conj-dtilde} The Hilbert series for $\mathcal E_{\tilde{D}_n}$ is \[ \frac{[2n-2]!!}{[2n-3]!!} \cdot \frac{[n][n+1]}{[2]^2[n-1][n-2]} \cdot \left[\frac{n^2-n}{2}\right]^2 \cdot \left[\frac{n^2-n-2}{2}\right]^2 \cdot \prod_{i=1}^{n-3} [i(2n-i-1)]. \] In particular, $\dim \mathcal E_{\tilde D_n} = (n+1)^2\cdot 2^{2n-8}$, and the top degree of $\mathcal E_{\tilde D_n}$ is $\frac13n(n-1)(2n-1)$. \end{conj} For small values of $n$ (including $n=3$, for which $\tilde{D}_3=\tilde{A}_3$), this gives the following: \begin{align*} n=3&&&[3]^2[4]^2\\ n=4&&&\frac{[4]^2[5]^2[6]^4}{[2]^2[3]^2}\\ n=5&&&\frac{[6]^2[8]^2[9]^2[10]^2[14]}{[2][3]^2[7]}\\ n=6&&&\frac{[6]^2[8][10]^2[14]^2[15]^2[18][24]}{[2][3][5]^2[9]}\\ n=7&&&\frac{[4][8]^2[10][12]^2[20]^2[21]^2[22][30][36]}{[2][3][5]^2[9][11]} \end{align*} \begin{figure} \begin{center} \dr{(-2,0)node[v]{}--(-1,0)node[v]{}--(0,0)node[v]{}--(1,0)node[v]{}--(2,0)node[v]{} (0,0)--(0,1)node[v]{}--(0,2)node[v]{}} \qquad\quad \dr{(-3,0)node[v]{}--(-2,0)node[v]{}--(-1,0)node[v]{}--(0,0)node[v]{}--(1,0)node[v]{}--(2,0)node[v]{}--(3,0)node[v]{} (0,0)--(0,1)node[v]{} (0,2)node{}}\\[5mm] \dr{(-2,0)node[v]{}--(-1,0)node[v]{}--(0,0)node[v]{}--(1,0)node[v]{}--(2,0)node[v]{}--(3,0)node[v]{} --(4,0)node[v]{}--(5,0)node[v]{} (0,0)--(0,1)node[v]{}} \end{center} \caption{\label{fig-entilde}The affine Dynkin diagrams $\tilde E_6$, $\tilde E_7$, and $\tilde E_8$.} \end{figure} We also have conjectures regarding the affine Dynkin diagrams $\tilde{E}_6$ and $\tilde{E}_7$, as shown in Figure~\ref{fig-entilde}. \begin{conj} \label{conj-etilde} The Hilbert series for $\mathcal E_{\tilde{E}_6}$ and $\mathcal E_{\tilde{E}_7}$ are: \begin{align*} \mathcal H_{\tilde{E}_6}(t) &= \frac{[6][9][12][14]^2[16]^2[21][22][30]^2}{[3]^2[4][7][11]},\\ \mathcal H_{\tilde{E}_7}(t) &= \frac{[6][8][10][12][14][18][24][27][32][34][48][49][52][66][75]}{[3][4][5][7][9][11][13][17]}. \end{align*} \end{conj} As for the remaining affine Dynkin diagram, $\mathcal E_{\tilde{E}_8}$ appears finite-dimensional but too large for us to confidently infer a possible Hilbert series. Even more basic is the following question. \begin{question} For which graphs $G$ is $\mathcal E_G$ finite-dimensional? \end{question} While $\mathcal E_G$ is finite-dimensional for all graphs $G$ on at most five vertices, it appears that some graphs on six vertices may have $\mathcal E_G$ infinite-dimensional. While we are not yet able to prove that any $\mathcal E_G$ is infinite-dimensional, our computations do suggest the following conjecture. \begin{conj} \label{conj-infinite} Let $G$ be a graph on six vertices. Then $\mathcal E_G$ is infinite-dimensional if and only if it contains a subgraph isomorphic to one of the graphs shown in Figure~\ref{fig-infinite}. \end{conj} \begin{figure} \begin{center} \dr{(0,0)node[v]{}--(90:1)node[v]{} (0,0)--(18:1)node[v]{} (0,0)--(162:1)node[v]{} (0,0)--(234:1)node[v]{} (0,0)--(306:1)node[v]{}} \quad\quad\quad \dr{(1,0)node[v]{}--(0,0)node[v]{}--(-1,0)node[v]{}--(-1,-1)node[v]{}--(0,-1)node[v]{}--(0,0)--(0,1)node[v]{}} \quad\quad\quad \dr{(1,0)node[v]{}--(60:1)node[v]{}--(120:1)node[v]{}--(180:1)node[v]{}--(240:1)node[v]{}--(300:1)node[v]{}--(1,0)--(-1,0)} \quad\quad\quad \dr{(1,0)node[v]{}--(60:1)node[v]{}--(120:1)node[v]{}--(180:1)node[v]{}--(240:1)node[v]{}--(300:1)node[v]{}--(1,0) (60:1)--(-60:1) (120:1)--(-120:1)}\\[.5cm] \dr{(18:1)node[v]{}--(90:1)node[v]{}--(162:1)node[v]{}--(234:1)node[v]{}--(306:1)node[v]{}--(18:1)--(0,0)node[v]{}--(162:1)}\quad\quad\quad \dr{(0,0)node[v]{}--(1,0)node[v]{}--(2.5,0)node[v]{}--(1.75,.8)node[v]{}--(1,0)--(1.75,-.8)node[v]{}--(2.5,0)--(3.5,0)node[v]{}}\quad\quad\quad \dr{(-1,0)node[v]{}--(0,0)node[v]{}--(1,1)node[v]{}--(2,0)node[v]{}--(1,-1)node[v]{}--(0,0) (1,1)--(1,0)node[v]{}--(1,-1)} \end{center} \caption{\label{fig-infinite}Minimal graphs $G$ on six vertices for which $\mathcal E_G$ appears to be infinite-dimensional. See Conjecture~\ref{conj-infinite}.} \end{figure} In particular, this would imply that $\mathcal E_6$ is infinite-dimensional. When $\mathcal E_G$ is finite-dimensional, it appears that its Hilbert series is especially nice. \begin{conj} If $\mathcal E_G$ is finite-dimensional, then $\mathcal H_G(t)$ is a product of cyclotomic polynomials. \end{conj} This holds for all Hilbert series we have been able to compute, and it also appears plausible for those Hilbert series that we cannot compute in full. Note that the corresponding statement is also true for all finite Coxeter groups. In Section~\ref{sec-coxeter}, we discussed many similarities between Coxeter groups and Fomin-Kirillov algebras. \begin{question} What other results about Coxeter groups have analogues for Fomin-Kirillov algebras? Are there corresponding notions to reflections/root systems/etc.? \end{question} We have studied the cases when $G$ is a Dynkin diagram of finite type because $\mathcal E_G$ appears to be relatively simple in these cases. However, despite results such as Theorem~\ref{thm-weyl}, we know of no direct explanation for the apparent connection between $\mathcal E_G$ and the corresponding Weyl group in these cases. \begin{question} Is there a uniform explanation for the structure of $\mathcal E_G$ when $G$ is a (simply-laced) Dynkin diagram of finite type? What if $G$ is an affine Dynkin diagram? \end{question} Our next question is related to the discussion in Section~\ref{sec-complementary}. By Corollary~\ref{cor-complementary}, there is a large class of graphs for which $\mathcal E_{G_1} \otimes \mathcal E_{G_2} \cong \mathcal E_n$, but by Proposition~\ref{prop-counter}, this does not hold for all pairs of complementary graphs. \begin{question} For which complementary graphs $G_1$ and $G_2$ is it true that $\mathcal E_{G_1} \otimes \mathcal E_{G_2} \cong \mathcal E_n$? In general, is the multiplication map $\mathcal E_{G_1} \otimes \mathcal E_{G_2} \to \mathcal E_n$ always injective? What is the structure of $\mathcal E_n$ as an $\mathcal E_{G_1}$-$\mathcal E_{G_2}$-bimodule? \end{question} Our next set of questions concerns relations in $\mathcal E_G$. \begin{question} Is there a straightforward description of the relations of $\mathcal E_G$? \end{question} In all of the examples we have discussed above, the minimal relations have a very special form: their $\S_n$-degrees are always automorphisms of the support of the relation (the graph containing the edges that appear in the relation). \begin{question} Must the $\S_n$-degree of a minimal relation be an automorphism of the support of the relation? \end{question} Also for many of the examples considered above, the minimal coset representatives of $\mathcal E_H$ inside $\mathcal E_G$ are relatively simple: the choice of minimal coset representatives is unique up to scalar factors (if we stipulate that they be representable by monomials). This allows us to describe a poset structure on the minimal coset representatives using the analogue of weak order. Then Theorem~\ref{thm-subgraph} suggests the following question. \begin{question} For which graphs $H \subset G$ does $\mathcal E_G/\mathcal E_H^+\mathcal E_G$ have a unique monomial basis (up to scalar factors)? \end{question} Finally, recall Conjecture~\ref{conj-nichols}. \begin{conj-nichols} \cite{MS} The bilinear form $\langle\cdot,\cdot\rangle$ is nondegenerate on $\mathcal E_n$. \end{conj-nichols} The quotient of $\mathcal E_n$ by the kernel of this bilinear form is a certain type of braided Hopf algebra called a \emph{Nichols algebra} (see \cite{AS}). Nichols algebras exist in much greater generality than the examples just mentioned. For instance, the analogues of Fomin-Kirillov algebras for types other than $A$ as defined in \cite{KirillovMaeno} can be described in terms of Nichols algebras \cite{Bazlov}. As such, it would be interesting to investigate the following question. \begin{question} To what extent can one extend the results of this paper to Fomin-Kirillov algebras of other types or more general Nichols algebras? \end{question} \section{Acknowledgments} The authors would like to thank John Stembridge for his valuable input throughout the duration of this research, as well as Thomas Lam and Sergey Fomin for interesting discussions. \section{Appendix} Here we give the Hilbert series $\mathcal H_G(t)$ for any connected graph on at most five vertices along with its top degree and total dimension. Note that $[k] = 1+t+t^2+\dots+t^{k-1}$. \bigskip \tikzset{every picture/.append style={scale=0.6}} \begin{longtable}{c>{\qquad$}c<{$\qquad}cc} $G$&\mbox{Hilb}&deg&dim\\ \hline\hline \dr{(0,0) node[v]{} -- (1,0)node[v]{}} &[2]&1&2\\ \hline \dr{(0,0) node[v]{} -- (1,0)node[v]{} -- (2,0)node[v]{}} &[2][3]&3&6\\ \dr{(0,0) node[v]{} -- (1,0)node[v]{} -- +(120:1)node[v]{} -- cycle} &[2]^2[3]&4&12\\ \hline \dr{(0,0) node[v]{} -- (1,0) node[v]{} -- (2,0) node[v]{} -- (3,0) node[v]{}} &[2][3][4]&6&24\\ \dr{(0,0) node[v]{} -- +(150:1) node[v]{} +(-150:1) node[v]{}--(0,0)--(1,0) node[v]{}} &[3][4]^2&8&48\\ \dr{(0,0) node[v]{} -- +(150:1) node[v]{} -- +(-150:1) node[v]{}--(0,0)--(1,0) node[v]{}} &[2][3][4]^2&9&96\\ \dr{(0,0)node[v]{}--(1,0)node[v]{}--(1,1)node[v]{}--(0,1)node[v]{}--cycle} &[3]^2[4]^2&10&144\\ \dr{(0,0)node[v]{}--(1,0)node[v]{}--(1,1)node[v]{}--(0,1)node[v]{}--cycle--(1,1)} &[2][3]^2[4]^2&11&288\\ \dr{(0,0)node[v]{}--(1,0)node[v]{}--(1,1)node[v]{}--(0,1)node[v]{}--cycle--(1,1) (1,0)--(0,1)} &[2]^2[3]^2[4]^2&12&576\\ \hline \dr{(0,0)node[v]{}--(1,0)node[v]{}--(2,0)node[v]{}--(3,0)node[v]{}--(4,0)node[v]{}} &[2][3][4][5]&10&120\\ \dr{(0,0) node[v]{} -- +(150:1) node[v]{} +(-150:1) node[v]{}--(0,0)--(1,0) node[v]{}--(2,0)node[v]{}} &[4]^2[5][6]&15&480\\ \dr{(0,0)node[v]{} -- ++(45:1)node[v]{} -- +(45:1)node[v]{} +(135:1)node[v]{} -- +(0,0) -- +(-45:1)node[v]{}} &[2]^{-2}[3]^{-2}[4]^2[5]^2[6]^4&28&14400\\ \dr{(0,0)node[v]{} -- ++(150:1)node[v]{} -- ++(-90:1)node[v]{} -- (0,0) -- (1,0)node[v]{} -- (2,0)node[v]{}} &[2][4]^2[5][6]&16&960\\ \dr{(0,0)node[v]{}--(-1,0)node[v]{}--+(-150:1)node[v]{}--(-1,-1)node[v]{}--(-1,0)(-1,-1)--(0,-1)node[v]{}}&[4]^2[5][6]^2&20&2880\\ \dr{(0,0)node[v]{}--++(36:1)node[v]{}--++(-36:1)node[v]{}--++(-108:1)node[v]{}--++(-1,0)node[v]{}--cycle}&[4]^2[5][6]^2&20&2880\\ \dr{(0,0)node[v]{}--(0,1)node[v]{}--(-1,1)node[v]{}--(-1,0)node[v]{}--(0,0)--(1,0)node[v]{}}&[2]^{-1}[4]^2[5][6]^3&24&8640\\ \dr{(0,0)node[v]{} -- ++(-150:1)node[v]{} -- ++(-150:1)node[v]{} -- ++(90:1)node[v]{} -- ++(-30:1) -- ++(-30:1)node[v]{}} &[2]^{-1}[3]^{-2}[4]^2[5]^2[6]^4&29&28800\\ \dr{(2,0)node[v]{}--(1,0)node[v]{}--(0,0)node[v]{}--(0,1)node[v]{}--(1,1)node[v]{}--(1,0)node[v]{} (0,0)--(1,1)}&[4]^2[5][6]^3&25&17280\\ \dr{(0,0)node[v]{}--++(150:1)node[v]{}--++(0,-1)node[v]{}--++(30:1)--++(30:1)node[v]{}--++(0,-1)node[v]{}--cycle}&[3]^{-2}[4]^2[5]^2[6]^4&30&57600\\ \dr{(0,0)node[v]{}--(-1,0)node[v]{}--(-1,1)node[v]{}--(0,1)node[v]{}--(30:1)node[v]{}--(0,0)--(0,1)node[v]{}} &[2]^{-1}[3]^{-1}[4]^3[5][6]^4&30&69120\\ \dr{(2,0)node[v]{}--(1,0)node[v]{}--(0,0)node[v]{}--(0,1)node[v]{}--(1,1)node[v]{}--(1,0)node[v]{} (1,0)--(0,1)}&[2]^{-1}[3]^{-1}[4]^2[5]^2[6]^4&31&86400\\ \dr{(0,0)node[v]{}--(1,0)node[v]{}--(1,1)node[v]{}--(0,1)node[v]{}--cycle--(1,1) (.5,.5)node[v]{}} &[2]^{-3}[3]^{-1}[4]^4[5]^2[6]^4&35&345600\\ \dr{(1,0)node[v]{}--(0,0)node[v]{}--(-1,0)node[v]{}--(-1,1)node[v]{}--(0,1)node[v]{}--(0,0)--(-1,1) (-1,0)--(0,1)} &[3]^{-1}[4]^2[5]^2[6]^4&32&172800\\ \dr{(120:1)node[v]{}--(-1,0)node[v]{}--(0,0)node[v]{}--(120:1)--(60:1)node[v]{}--(1,0)node[v]{}--(0,0)--(60:1)} &[2]^{-1}[3]^{-1}[4]^3[5]^2[6]^4&34&345600\\ \dr{(0,0)node[v]{}--(1,0)node[v]{}--(1,1)node[v]{}--(0,1)node[v]{}--cycle--(1,1) (.5,.5)node[v]{}--(1,0)} &[2]^{-2}[3]^{-1}[4]^4[5]^2[6]^4&36&691200\\ \dr{(0,0)node[v]{}--(-1,0)node[v]{}--(0,1)node[v]{}--(0,0)--(-1,1)node[v]{}--(0,1)--(30:1)node[v]{}--(0,0)} &[2]^{-2}[3]^{-1}[4]^4[5]^2[6]^4&36&691200\\ \dr{(-1,0)node[v]{}--(-1,1)node[v]{}--(0,1)node[v]{}--(30:1)node[v]{}--(0,0)node[v]{}--(-1,0)--(0,1)--(0,0)--(-1,1)} &[2]^{-1}[3]^{-1}[4]^4[5]^2[6]^4&37&1382400\\ \dr{(0,0)node[v]{}--(1,0)node[v]{}--(1,1)node[v]{}--(0,1)node[v]{}--(0,0)--(1,1) (1,0)--(0,1) (.5,.5)node[v]{}} &[2]^{-2}[4]^4[5]^2[6]^4&38&2073600\\ \dr{(306:1)node[v]{}--(18:1)node[v]{}--(90:1)node[v]{}--(162:1)node[v]{}--(234:1)node[v]{} --(18:1)--(162:1)--(306:1)--(90:1)--(234:1)} &[2]^{-1}[4]^4[5]^2[6]^4&39&4147200\\ \dr{(306:1)node[v]{}--(18:1)node[v]{}--(90:1)node[v]{}--(162:1)node[v]{}--(234:1)node[v]{} --(18:1)--(162:1)--(306:1)--(90:1)--(234:1)--(306:1)} &[4]^4[5]^2[6]^4&40&8294400 \end{longtable}

https://arxiv.org/abs/1204.5704

A variant of Touchard's Catalan number identity

It is well known that the Catalan number C_n counts dissections of a regular (n+2)-gon into triangles. Here we count such dissections by number of triangles that contain two sides of the polygon among their three edges, leading to a combinatorial interpretation of the identity C_n =sum_{1<=k<=n/2} 2^{n-2k} n-choose-2k C_k (k(n+2))/(n(n-1)), and illustrating its connection with Touchard's identity.

\section{Introduction} \vspace*{-5mm} Consider a regular polygon of $n+2$ sides with one side designated the base. It is a classic result that there are the Catalan number $C_{n}$ ways to insert noncrossing diagonals connecting nonadjacent vertices of the polygon so as to dissect it into triangles (see illustration in Figure \ref{fig}). Each such dissection contains $n-1$ diagonals and $n$ triangles. When $n\ge 2$, each triangle may have 0, 1, or 2 sides in common with the polygon. Let $u_{n,k}$ denote the number of dissections in which precisely $k$ triangles contain 2 sides of the polygon. In any dissection, the number of such 2-polygon-side triangles ranges from a minimum of 2 (provided $n\ge 2$) to a maximum of $\lfloor(n+2)/2\rfloor$. Our main result is that $u_{n,k+1}=2^{n-2k} {n \choose 2k} C_{k}\,\frac{k(n+2)}{n(n-1)}$, yielding the apparently new identity \begin{equation}\label{main} \hspace*{30mm} C_{n} =\sum_{1 \,\le\,k \,\le\, n/2} 2^{n-2k} {n \choose 2k} C_{k}\,\frac{k(n+2)}{n(n-1)},\hspace*{20mm} n\ge 2. \end{equation} This identity is reminiscent of Touchard's identity \cite{A091894}, \begin{equation}\label{touchard} \hspace*{38mm} C_{n+1} =\sum_{0 \,\le\,k \,\le\, n/2} 2^{n-2k} {n \choose 2k} C_{k},\hspace*{27mm} n\ge 0, \end{equation} and indeed we will see a connection between them. To obtain an expression for $u_{n,k}$, it is convenient to color the base of the polygon blue and the remaining edges black, and let $v_{n,k}$ denote the number of dissections in which $k$ triangles contain two \emph{black} edges of the polygon. In Section 2, we express the \{$u_{n,k}$\} directly in terms of the \{$v_{n,k}$\}. In Section 3, we show bijectively that $v_{n+1,k+1}$ is actually the summand $2^{n-2k} {n \choose 2k} C_{k}$ in (\ref{touchard}), incidentally giving another combinatorial interpretation of Touchard's identity. Section 4 then establishes the main result. \section[A relation between u(n,k) and v(n,k)]{\protect{A relation between $\mbf{u_{n,k}}$ and $\mbf{v_{n,k}}$}} \vspace*{-5mm} Clearly, $u_{1,1}=1,\ u_{2,2}=2,\ u_{3,2}=5$. For $n\ge 4$, let us count the contribution to $u_{n,k}$ according to the positive vertex $r$ of the triangle that contains the base, after labelling the vertices of the polygon $r=-1,0,1,...,n$ counterclockwise from the left endpoint of the base as illustrated in Figure \ref{fig}. We find that the contribution to $u_{n,k}$ for both $r=1$ and $r=n$ is $v_{n-1,k-1}$, and for $2\le r \le n-1$, the contribution is $ \sum_{k-\frac{n-r+1}{2} \le\, j\, \le \frac{r}{2}}v_{r-1,j}\,v_{n-r,k-j}. $ Hence, \[ u_{n,k} = 2 v_{n-1,k-1} + \sum_{r=2}^{n-1}\:\sum_{k-\frac{n-r+1}{2} \le\, j\, \le \frac{r}{2}}v_{r-1,j}\,v_{n-r,k-j}, \] valid for $n \ge 4 ,\ 2 \le k \le \frac{n + 2}{2}$. Similarly, we find a recurrence for $v_{n,k}$, \[ v_{n,k} = 2 v_{n-1,k} + \sum_{r=2}^{n-1}\:\sum_{k-\frac{n-r+1}{2} \le\, j\, \le \frac{r}{2}}v_{r-1,j}\,v_{n-r,k-j}, \] that involves the same double sum. Eliminating the double sum in the two equations leads to the relation \begin{equation}\label{eq2} u_{n,k} = v_{n,k} + 2v_{n-1,k-1}-2v_{n-1,k}, \end{equation} which in fact holds for all $n,k$. \section{A bijection} \label{bij} \vspace*{-5mm} It is well known that $2^{n-1-2k} {n-1 \choose 2k} C_{k}$ is the number of Dyck paths that contain $k\ DDU$'s, where $U$ denotes an upstep and $D$ a downstep. (See \cite{dyck2004} for a bijective proof.) We now present a bijection from polygon dissections to Dyck paths which makes it visually obvious that the triangles containing two black sides, taken in clockwise order from the base, except that the last one is ignored, correspond to the $DDU$'s, taken left to right, in the Dyck path. This bijection is simply the composition of the following 3 well known bijections, (1) the Erdelyi-Etherington bijection from triangle-dissections of a polygon to binary trees \cite[p.\,171]{ec2}, (2) the standard bijection from binary trees to ordered trees (Knuth's ``natural'' correspondence \cite[Section 2.3.2]{acp1}), and (3) the (trivial) ``glove'' bijection from ordered trees to Dyck paths. Here is an illustration. \setcounter{equation}{0} \vspace*{-1mm} \begin{equation}\label{fig}\notag \textrm{Figure 1} \end{equation} \begin{center} \begin{pspicture}(-1,-4)(10,2.5 \psset{unit=1.4cm} \psdots(-2,0)(-1.62,1.18)(-.62,1.9)(.62,1.9)(1.62,1.18)(2,0)(1.62,-1.18)(.62,-1.9)(-.62,-1.9)(-1.62,-1.18) \pspolygon(-2,0)(-1.62,1.18)(-.62,1.9)(.62,1.9)(1.62,1.18)(2,0)(1.62,-1.18)(.62,-1.9)(-.62,-1.9)(-1.62,-1.18)(-2,0) \psline[linecolor=blue,linewidth=1pt](-.62,-1.9)(.62,-1.9) \psline(-.62,-1.9)(-.62,1.9) \psline(.62,-1.9)(.62,1.9) \psline(-.62,1.9)(-2,0)(-.62,-1.9) \psline(.62,-1.9)(1.62,1.18)(1.62,-1.18) \psline(-.62,-1.9)(.62,1.9) \rput(-.8,-2.1){\textrm{{\footnotesize $-1$}}} \rput(.75,-2.1){\textrm{{\footnotesize $0$}}} \rput(1.8,-1.18){\textrm{{\footnotesize $1$}}} \rput(2.2,0){\textrm{{\footnotesize $2$}}} \rput(-1.8,-1.18){\textrm{{\normalsize $n$}}} \rput(1.75,1.5){$\ddots$} \rput(-2.2,0.1){\vdots} \rput(2.9,-.5){$\longrightarrow$} \psline[linecolor=red,linewidth=1pt](0,-2.4)(0,-1.4)(-.3,0.5)(-1,0)(-1.5,1) \psline[linecolor=red,linewidth=1pt](-1,0)(-1.5,-1) \psline[linecolor=red,linewidth=1pt](0,-1.4)(1,0)(1.3,-.5)(1.8,0) \rput(0,-2.8){\textrm{{\footnotesize a dissection into $n=8$ triangles}}} \rput(6,-2.8){\textrm{{\footnotesize left-planted binary tree}}} \psset{dotscale=1.5} \psdots[linecolor=red](0,-2.4)(0,-1.4)(-.3,0.5)(-1,0)(-1.5,1)(-1.5,-1)(1,0)(1.3,-.5)(1.8,0) \psset{unit=1.0cm} \psdots[linecolor=red](5,1)(6,0)(7,1)(7,-1)(8,-2)(9,-3)(9,-1)(10,0)(9,1) \psline[linecolor=red](5,1)(6,0)(7,1) \psline[linecolor=red](6,0)(7,-1)(8,-2)(9,-3) \psline[linecolor=red](8,-2)(9,-1)(10,0)(9,1) \psset{unit=1.4cm} \psset{dotscale=1.6} \psdots[linecolor=red](-1.5,1)(-1.5,-1) \psset{unit=1.0cm} \psset{dotscale=1.2} \psdots[linecolor=red](5,1)(7,1) \rput(11.7,-.5){$\xrightarrow[45^{\circ}]{\mathrm{rotate}}$} \end{pspicture} \end{center} \begin{center} \begin{pspicture}(-1,-1)(14,5 \psset{unit=1.0cm} \psdots(-1,0)(-1,1)(-1,2)(-1,3)(-1,4)(0,3)(0,1)(1,1)(1,2) \psline(-1,0)(-1,1)(-1,2)(-1,3)(-1,4) \psline(-1,3)(0,3) \psline(-1,1)(0,1)(1,1)(1,2) \rput(3,2){$\xrightarrow[\textrm{east$\:\rightarrow\:$west edges}]{\textrm{swing down}}$} \psdots(5,0)(5,1)(5,2)(5,3)(5,4)(6,3)(6,1)(7,1)(7,2) \psline(6,3)(5,2)(5,3)(5,4) \psline(6,1)(5,0)(5,1)(5,2) \psline(5,0)(7,1)(7,2) \rput(8.2,2){$\xrightarrow{\textrm{``prettify''}}$} \psdots(9.5,4)(9.5,3)(10,2)(10,1)(10.5,3)(11,0)(11,1)(12,1)(12,2) \psline(9.5,4)(9.5,3)(10,2)(10.5,3) \psline(10,2)(10,1)(11,0)(11,1) \psline(11,0)(12,1)(12,2) \rput(11,-0.5){\textrm{{\footnotesize ordered tree}}} \rput(13,2){$\longrightarrow$} \psset{dotscale=1.8} \psdots(-1,4)(0,3)(5,4)(6,3)(9.5,4)(10.5,3) \end{pspicture} \end{center} \vspace*{-3mm} \Einheit=0.6cm \[ \Pfad(-8,0),3333443444343344\endPfad \SPfad(-8,0),1111111111111111\endSPfad \DuennPunkt(-8,0) \DuennPunkt(-7,1) \DuennPunkt(-6,2) \DuennPunkt(-5,3) \NormalPunkt(-4,4) \DuennPunkt(-3,3) \DuennPunkt(-2,2) \NormalPunkt(-1,3) \DuennPunkt(0,2) \DuennPunkt(1,1) \DuennPunkt(2,0) \DuennPunkt(3,1) \DuennPunkt(4,0) \DuennPunkt(5,1) \DuennPunkt(6,2) \DuennPunkt(7,1) \DuennPunkt(8,0) \] \begin{center} \begin{pspicture}(-1,-1)(0,0 \psset{unit=1.0cm} \rput(-.5,0.3){\textrm{{\footnotesize Dyck path}}} \rput(0,-.5){\textrm{ {\normalsize Bijection from triangle dissections to Dyck paths}}} \end{pspicture} \end{center} \vspace*{-5mm} \noindent The last step is the glove bijection: walk clockwise around the tree starting from the root and record an upstep (resp. downstep) each time an edge is traversed upward (resp. downward). Or, more picturesquely, burrow up the edges from the root to form a multi-fingered glove and fan out the fingers. Thus each edge in the tree corresponds to a matching upstep and downstep in the path. The illustrated dissection has 3 triangles containing two black sides of the polygon; all but the last are highlighted using enlarged dots, and they show up in the Dyck path as vertices initiating a descent of 2 or more downsteps followed by an upstep, that is, they correspond to $DDU$s in the Dyck path, as claimed. \vspace*{-2mm} \section{Conclusion} \vspace*{-5mm} The preceding section shows that $v_{n,k} = 2^{n+1-2k}\binom{n-1}{2k-2}C_{k-1}$. Substituting into (\ref{eq2}), we find \[ u_{n,k}=2^{n+1-2k}\left(\binom{n-2}{2k-3}C_{k-1}+\binom{n-2}{2k-4}4C_{k-2}\right), \] which simplifies to \[ u_{n,k+1}=2^{n-2k} \binom{n}{2k}C_{k} \frac{(n+2)k}{n(n-1)}. \] Now sum over $k$ to obtain (\ref{main}). The first few values of $u_{n,k}$ are given in the following table. \[ \begin{array}{c|ccccc} n^{\textstyle{\,\backslash \,k}} & 2 & 3 & 4 & 5 \\ \hline 2& 2 & & & \\ 3& 5 & & & \\ 4& 12 & 2 & & \\ 5& 28 & 14 & & \\ 6& 64 & 64 & 4 & \\ 7& 144 & 240 & 45 & \\ 8& 320 & 800 & 300 & 10 \\ 9& 704 & 2464 & 1540 & 154 \\ \end{array} \] \vspace*{2mm} \centerline{{\normalsize Table of values of $u_{n,k}$ }} \vspace*{4mm} \textbf{Added in proof} \ Tewodros Amdeberhan informs me that he has recently discovered an identity equivalent to (\ref{main}), namely \[ \frac{2n}{n+3} C_{n+1} =\sum_{0 \,\le\,k \,\le\, (n-1)/2} 2^{n-2k} {n \choose 2k+1} C_{k}\,\frac{2k+1}{k+2}, \] and has observed that subtracting the latter from Touchard's identity (\ref{touchard})(multiplied by 2) gives an alternating sum expression for the super ballot number $6/(n+3)C_{n+1}$, sequence \htmladdnormallink{A007054}{http://oeis.org/A007054} in OEIS. \textbf{Acknowledgement of priority}\ \ Alon Regev pointed out to me that the main results of this paper have previously been obtained by Hurtado and Noy \cite{ears}.

https://arxiv.org/abs/1710.00442

Virtual Element Methods on Meshes with Small Edges or Faces

We consider a model Poisson problem in $\R^d$ ($d=2,3$) and establish error estimates for virtual element methods on polygonal or polyhedral meshes that can contain small edges ($d=2$) or small faces ($d=3$).

\section{Introduction}\label{sec:Introduction} Let ${\Omega}\subset \mathbb{R}^d$ ($d=2,3$) be a bounded polygonal/polyhedral domain and $f\in L_2({\Omega})$. The Poisson problem with the homogeneous Dirichlet boundary condition is to find $u\in H^1_0({\Omega})$ such that \begin{equation}\label{eq:Poisson} a(u,v)=(f,v) \qquad\forall\,v\in H^1_0({\Omega}), \end{equation} where \begin{equation}\label{eq:aDef} a(u,v)=\int_{\Omega} \nabla u\cdot\nabla v\,dx \end{equation} and $(\cdot,\cdot)$ is the inner product for $L_2({\Omega})$. Here and throughout the paper we follow standard notation for differential operators, function spaces and norms that can be found for example in \cite{Ciarlet:1978:FEM,ADAMS:2003:Sobolev,BScott:2008:FEM}. \par Problem \eqref{eq:Poisson} can be solved by virtual element methods \cite{BBCMMR:2013:VEM,AABMR:2013:Projector} on polygonal or polyhedral meshes. It has been observed in numerical experiments that the convergence rates for the virtual element methods do not deteriorate noticeably even in the presence of small edges or faces (cf. \cite{AABMR:2013:Projector,BLR:2016:VEM,BDR:2017:VEM}). Our goal is to establish error estimates that justify these numerical results for the virtual element methods introduced in \cite{AABMR:2013:Projector}. \par We will develop error estimates that are based on general shape regularity assumptions on the subdomains in the polygonal or polyhedral meshes. For the two dimensional problem, we assume that (i) each polygonal subdomain is star-shaped with respect to a disc whose diameter is comparable to the diameter of the subdomain and (ii) the number of edges of the subdomains is uniformly bounded. For the three dimensional problem, we assume that (i) each polyhedral subdomain is star-shaped with respect to a ball whose diameter is comparable to the diameter of the subdomain; (ii) the number of faces of the subdomains is uniformly bounded; and (iii) the faces of the subdomains satisfy the two dimensional shape regularity assumptions. Our error estimates are optimal up to at most a logarithmic factor that involves the ratio of the lengths of the longest edge and the shortest edge of each subdomain in two dimensions and a similar ratio over the edges of the faces on the subdomains in three dimensions. \par The rest of the paper is organized as follows. We begin with a star-shaped condition in Section~\ref{sec:SS}. Then we treat the two dimensional case in Section~\ref{sec:LocalVEM2D} and Section~\ref{sec:Poisson2D}, where the analysis benefits from the techniques developed in \cite{BLR:2016:VEM} and \cite{BGS:2017:VEM2}. The extension to the three dimensional Poisson problem is presented in Section~\ref{sec:Poisson3D}. We end with some concluding remarks in Section~\ref{sec:Conclusions}. \par In order to avoid the proliferation of constants, we will often use the notation $A\lesssim B$ to represent the statement that $A\leq (\text{constant}) B$, where the positive constant is independent of mesh sizes. The notation $A\approx B$ is equivalent to $A\lesssim B$ and $B\lesssim A$. The precise dependence of the hidden constants will be declared in the text. \par To minimize the technicalities, we also assume that ${\Omega}$ is convex so that the solution of \eqref{eq:Poisson} belongs to $H^2({\Omega})$ by elliptic regularity \cite{Grisvard:1985:EPN,Dauge:1988:EBV}. \section{A Star-Shaped Condition}\label{sec:SS} Let $D$ be a bounded open polygon ($d=2$) or a bounded open polyhedron $(d=3)$, and $\hspace{1pt}h_{{\scriptscriptstyle D}}$ be the diameter of $D$. \par We assume that \begin{equation}\label{eq:SA} \text{$D$ is star-shaped with respect to a disc/ball $\fB_{\!\ssD}\subset D$ with radius $\rho_{\!{_\ssD}}\hspace{1pt}h_{{\scriptscriptstyle D}}$}. \end{equation} \par We will denote by $\tilde\fB_{\!\ssD}$ the disc/ball concentric with $\fB_{\!\ssD}$ whose radius is $\hspace{1pt}h_{{\scriptscriptstyle D}}$. It is clear that \begin{equation}\label{eq:discs} \fB_{\!\ssD}\subset D\subset \tilde\fB_{\!\ssD}. \end{equation} \par Below are some consequences of the star-shaped condition \eqref{eq:SA}. The hidden constants in Section~\ref{subsec:Sobolev}--Section~\ref{subsec:PF} only depend on $\rho_{\!{_\ssD}}$, while those in Section~\ref{subsec:DE} also depend on $k$. \subsection{Sobolev Inequalities}\label{subsec:Sobolev} It follows from \eqref{eq:SA} that \begin{alignat}{3} \|\zeta\|_{{L_\infty(D)}}&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-(d/2)}\|\zeta\|_{L_2(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}^{1-(d/2)}|\zeta|_{{H^1(D)}} +\hspace{1pt}h_{{\scriptscriptstyle D}}^{2-(d/2)}|\zeta|_{H^2(D)}&\qquad&\forall\, \zeta\in H^2(D),\label{eq:Sobolev}\\ \intertext{and in the case where $d=2$,} \|\zeta\|_{{L_\infty(D)}}&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|\zeta\|_{L_2(D)}+|\zeta|_{{H^1(D)}} +\hspace{1pt}h_{{\scriptscriptstyle D}}^{1/2}|\zeta|_{H^{3/2}(D)}&\qquad&\forall\, \zeta\in H^{3/2}(D).\label{eq:Sobolev2} \end{alignat} Details can be found in \cite[Lemma~4.3.4]{BScott:2008:FEM} and \cite[Section~6]{DS:1980:BH}. \subsection{Bramble-Hilbert Estimates}\label{subsec:BH} Condition \eqref{eq:SA} also implies the following Bramble-Hilbert estimates \cite{BH:1970:Lemma}: \begin{alignat}{3} \inf_{q\in\mathbb{P}_\ell}|\zeta-q|_{H^m(D)}&\lesssim h_D^{\ell+1-m}|\zeta|_{H^{\ell+1}(D)} &\qquad&\forall\,\zeta\in H^{\ell+1}(D), \,\ell=0,\ldots,k,\;\text{and}\; m\leq \ell, \label{eq:BHEstimates}\\ \inf_{q\in\mathbb{P}_\ell}|\zeta-q|_{H^m(D)}& \lesssim h_D^{\ell+\frac12-m}|\zeta|_{H^{\ell+\frac12}(D)} &\qquad&\forall\,\zeta\in H^{\ell+\frac12}(D), \,\ell=0,\ldots,k,\;\text{and}\; m\leq \ell. \label{eq:BHEstimates2} \end{alignat} Details can be found in \cite[Lemma~4.3.8]{BScott:2008:FEM} and \cite[Section~6]{DS:1980:BH}. \subsection{A Lipschitz Isomorphism between $D$ and $\fB_{\!\ssD}$}\label{subsec:Lipschitz} \par In view of the star-shaped condition \eqref{eq:SA}, there exists a Lipschitz isomorphism $\Phi:\fB_{\!\ssD}\longrightarrow D$ such that both $|\Phi|_{W^{1,\infty}(\fB_{\!\ssD})}$ and $|\Phi^{-1}|_{W^{1,\infty}(D)}$ are bounded by a constant that only depends on $\rho_{\!{_\ssD}}$ (cf. \cite[Section~1.1.8]{Mazya:2011:Sobolev}). \par It follows that % \begin{align} |D|&\approx \hspace{1pt}h_{{\scriptscriptstyle D}}^d,\label{eq:G1}\\ |\partial D|&\approx \hspace{1pt}h_{{\scriptscriptstyle D}}^{d-1},\label{eq:G2} \end{align} where $|D|$ is the area of $D$ ($d=2$) or the volume of $D$ ($d=3$), and $|\partial D|$ is the arclength of $\partial D$ ($d=2$) or the surface area of $D$ ($d=3$). \par Moreover we have (cf. \cite[Theorem~4.1]{Wloka:1987:PDE}) \begin{alignat}{3} \|\zeta\|_{L_2(\partial D)}&\approx \|\zeta\circ\Phi\|_{L_2(\partial\fB_{\!\ssD})} &\qquad&\forall\,\zeta\in L_2(\partial D),\label{eq:L2Iso1}\\ \|\zeta\|_{L_2(D)}&\approx \|\zeta\circ\Phi\|_{L_2(\fB_{\!\ssD})} &\qquad&\forall\,\zeta\in L_2(D),\label{eq:L2Iso2}\\ |\zeta|_{H^1(\partial D)}&\approx|\zeta\circ\Phi|_{H^1(\partial\fB_{\!\ssD})} &\qquad&\forall\,\zeta\in H^1(\partial D),\label{eq:H1Iso1}\\ |\zeta|_{{H^1(D)}}&\approx|\zeta\circ\Phi|_{H^1(\fB_{\!\ssD})} &\qquad&\forall\,\zeta\in {H^1(D)},\label{eq:H1Iso2}\\ |\zeta|_{H^{1/2}(\partial D)}&\approx |\zeta\circ\Phi|_{H^{1/2}(\partial\fB_{\!\ssD})} &\qquad&\forall\,\zeta\in H^{1/2}(\partial D). \label{eq:HalfIso} \end{alignat} \subsection{Poincar\'e-Friedrichs Inequalities}\label{subsec:PF} The Bramble-Hilbert estimate \eqref{eq:BHEstimates} and the geometric estimates \eqref{eq:G1}--\eqref{eq:G2} imply the following Poincar\'e-Friedrichs inequalities: \begin{alignat}{3} h_D^{-(d/2)} \|\zeta\|_{L_2(D)}&\lesssim h_D^{-d}\Big|\int_{D} \zeta\,dx\Big| +\hspace{1pt}h_{{\scriptscriptstyle D}}^{1-(d/2)}|\zeta|_{{H^1(D)}} &\qquad&\forall\,\zeta\in {H^1(D)}, \label{eq:PF1}\\ h_D^{-(d/2)} \|\zeta\|_{L_2(D)}&\lesssim h_D^{-(d-1)}\Big|\int_{\partial D} \zeta\,ds\Big| +\hspace{1pt}h_{{\scriptscriptstyle D}}^{1-(d/2)}|\zeta|_{{H^1(D)}} &\qquad&\forall\,\zeta\in {H^1(D)}.\label{eq:PF2} \end{alignat} \subsection{Estimates for $|\cdot|_{H^{1/2}(\partial D)}$}\label{subsec:Half} On the circle $\partial \fB_{\!\ssD}$, we have a standard estimate \begin{equation*} |\zeta|_{H^{1/2}(\partial\fB_{\!\ssD})}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1/2}\|\zeta\|_{L_2(\partial\fB_{\!\ssD})} +\hspace{1pt}h_{{\scriptscriptstyle D}}^{1/2}|\zeta|_{H^1(\partial \fB_{\!\ssD})} \qquad\forall\,\zeta\in H^1(\partial \fB_{\!\ssD}). \end{equation*} It then follows from a Poincar\'e-Friedrichs inequality for a circle that \begin{equation*} |\zeta|_{H^{1/2}(\partial\fB_{\!\ssD})}=|\zeta-\bar\zeta|_{H^{1/2}(\partial \fB_{\!\ssD})} \lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1/2}\|\zeta-\bar\zeta\|_{L_2(\partial\fB_{\!\ssD})} +\hspace{1pt}h_{{\scriptscriptstyle D}}^{1/2}|\zeta|_{H^1(\partial \fB_{\!\ssD})} \lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{1/2}|\zeta|_{H^1(\partial\fB_{\!\ssD})}, \end{equation*} where $\bar\zeta$ is the mean of $\zeta$ over $\partial \fB_{\!\ssD}$. Therefore, in view of \eqref{eq:H1Iso1} and \eqref{eq:HalfIso}, we have % \begin{equation}\label{eq:HalfAndOne} |\zeta|_{H^{1/2}(\partial D)}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{1/2}|\zeta|_{H^1(\partial D)} \qquad\forall\,\zeta\in H^1(\partial D). \end{equation} \par Similarly, it follows from \eqref{eq:H1Iso2}, \eqref{eq:HalfIso} and the trace theorem for $H^1(\fB_{\!\ssD})$ that \begin{alignat}{3} |\zeta|_{H^{1/2}(\partial D)}&\lesssim |\zeta|_{{H^1(D)}} &\qquad&\forall\,\zeta\in {H^1(D)}.\label{eq:HalfTrace} \end{alignat} \subsection{Trace Inequalities}\label{subsec:Trace} It follows from \eqref{eq:L2Iso1}, \eqref{eq:L2Iso2}, \eqref{eq:H1Iso2} and standard (scaled) trace inequalities for $H^1(\fB_{\!\ssD})$ that \begin{alignat}{3} \|\zeta\|_{L_2(\partial D)}^2&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|\zeta\|_{L_2(D)}^2+\hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{{H^1(D)}}^2 &\qquad&\forall\,\zeta\in {H^1(D)}.\label{eq:Trace} \end{alignat} \par We also have trace inequalities for the $H^1$ norm on $\partial D$ that require a different derivation. \begin{lemma}\label{lem:CZ1} Let $e$ be an edge of $D\subset \mathbb{R}^2$. We have \begin{equation* \hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{H^1(e)}^2\lesssim |\zeta|_{H^1(D)}^2+\hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{H^{3/2}(D)}^2 \qquad \forall\,\zeta\in H^{3/2}(D). \end{equation*} \end{lemma} \begin{proof} By scaling we can assume $\hspace{1pt}h_{{\scriptscriptstyle D}}=1$. Without loss of generality we may also assume that \begin{equation}\label{eq:CZ1Est0} \int_D \zeta\,dx=0. \end{equation} \par The existence of the Lipschitz isomorphism $\Phi:D\longrightarrow \fB_{\!\ssD}$ implies that the domain $D$ satisfies a uniform cone condition \cite[Section~4.8]{ADAMS:2003:Sobolev}, with one reference cone and a finite cover of $\bar D$ that contains a fixed number of congruent open discs. Furthermore, the angle and the height of the reference cone and the radius of the open discs only depend on $\rho_{\!{_\ssD}}$. It follows that there exists a Calderon-Zygmund extension operator $E:H^1(D)\longrightarrow H^1(\mathbb{R}^2)$ (cf. \cite[Theorem~5.28]{ADAMS:2003:Sobolev}) such that $E$ maps $H^2(D)$ into $H^2(\mathbb{R}^2)$ and the restriction of $E\zeta$ to $D$ equals $\zeta$. Moreover we have \begin{equation*} \|E\zeta\|_{H^1(\mathbb{R}^2)}\lesssim \|\zeta\|_{H^1(D)} \quad\forall\,\zeta\in{H^1(D)} \quad\text{and}\quad \|E\zeta\|_{H^2(\mathbb{R}^2)}\lesssim \|\zeta\|_{H^2(D)} \quad\forall\,\zeta\in{H^2(D)}. \end{equation*} It follows from the interpolation of Sobolev spaces \cite[Chapter~7]{ADAMS:2003:Sobolev} that \begin{align}\label{eq:CZ1Est1} \|E\zeta\|_{H^{3/2}(\mathbb{R}^2)}\lesssim \|\zeta\|_{H^{3/2}(D)} \lesssim |\zeta|_{H^1(D)} +|\zeta|_{H^{3/2}(D)} \qquad\forall\,\zeta\in H^{3/2}(D), \end{align} where we have used \eqref{eq:PF2} and \eqref{eq:CZ1Est0}. \par Let $e$ be an edge of $D$, $\tilde e$ be the infinite line that contains $e$ and $G$ be a half-plane that borders $\tilde e$. Then we have, by \eqref{eq:CZ1Est1} and the trace theorem \cite[Theorem~8.1]{Wloka:1987:PDE}, \begin{align* |\zeta|_{H^1(e)}&=|E\zeta|_{H^1(e)} \leq |E\zeta|_{H^1(\tilde e)} \lesssim \|E\zeta\|_{H^{3/2}(G)} \lesssim |\zeta|_{H^1(D)}+|\zeta|_{H^{3/2}(D)}. \end{align*} \end{proof} \par The proof of the following result is similar. \begin{lemma}\label{lem:CZ2} Let $F$ be a face of $D\subset \mathbb{R}^3$. We have \begin{equation* \hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{H^1(F)}^2\lesssim |\zeta|_{H^1(D)}^2+\hspace{1pt}h_{{\scriptscriptstyle D}}^{2}|\zeta|_{H^2(D)}^2 \qquad \forall\,\zeta\in H^{2}(D). \end{equation*} \end{lemma} \subsection{A Lifting Operator}\label{subsec:Lifting} It follows from \eqref{eq:L2Iso2}, \eqref{eq:H1Iso2}, \eqref{eq:HalfIso} and the inverse trace theorem for $H^1(\fB_{\!\ssD})$ (cf. \cite[Theorem~8.8]{Wloka:1987:PDE}) that there exists a linear operator $\mathrm {Tr}^\dag:H^{1/2}(\partial D)\longrightarrow {H^1(D)}$ such that \begin{equation* \mathrm {Tr}^\dag \zeta=\zeta\;\text{on}\;\partial D \quad\text{and}\quad \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|\mathrm {Tr}^\dag \zeta\|_{L_2(D)}+|\mathrm {Tr}^\dag \zeta|_{{H^1(D)}}\leq C|\zeta|_{H^{1/2}(\partial D)}, \end{equation*} where the constant $C$ depends only on $\rho_{\!{_\ssD}}$. % \subsection{Some Estimates for Polynomials}\label{subsec:DE} Let $\mathbb{P}_k$ be the space of polynomials of total degree $\leq k$ in $d$ variables. We obtain the following estimates by using the equivalence of norms on finite dimensional vector spaces and scaling. \begin{lemma}\label{lem:DiscreteEstimates} We have \begin{alignat}{3} \|p\|_{L_2(\partial D)}^2&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|p\|_{L_2(D)}^2&\qquad&\forall\,p\in\mathbb{P}_k, \label{eq:DiscreteEstimate0}\\ |p|_{{H^1(D)}}&\lesssim h_D^{-1}\|p\|_{L_2(D)} &\qquad&\forall\,p\in \mathbb{P}_k, \label{eq:DiscreteEstimate1}\\ \|p\|_{L_\infty(D)}&\lesssim |\bar{p}_{_{\partial D}}|+\hspace{1pt}h_{{\scriptscriptstyle D}}^{1-(d/2)} |p|_{{H^1(D)}}&\qquad&\forall\,p\in\mathbb{P}_k, \label{eq:DiscreteEstimate2}\\ \|p\|_{L_\infty(D)}&\lesssim |\bar{p}_{_{D}}|+\hspace{1pt}h_{{\scriptscriptstyle D}}^{1-(d/2)} |p|_{{H^1(D)}}&\qquad&\forall\,p\in\mathbb{P}_k, \label{eq:DiscreteEstimate3} \end{alignat} where $\bar p_{_{\partial D}}$ is the mean of $p$ over $\partial D$ and $\bar p_{_D}$ is the mean of $p$ over $D$. \end{lemma} \begin{proof} In view of \eqref{eq:discs}, we have \begin{align}\label{eq:DS2} \|p\|_{L_\infty(D)}\leq \|p\|_{L_\infty(\tilde\mathfrak{B}_{{\scriptscriptstyle D}})} &\lesssim \|p\|_{L_\infty(\mathfrak{B}_{{\scriptscriptstyle D}})}\\ &\lesssim (\text{diam}\,\mathfrak{B}_{{\scriptscriptstyle D}})^{-d/2}\|p\|_{L_2(\mathfrak{B}_{{\scriptscriptstyle D}})} \leq (\text{diam}\,\mathfrak{B}_{{\scriptscriptstyle D}})^{-d/2}\|p\|_{L_2(D)}.\notag \end{align} The estimate \eqref{eq:DiscreteEstimate0} then follows from \eqref{eq:G2} and \eqref{eq:DS2}: \begin{align*} \|p\|_{L_2(\partial D)}^2 \lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{d-1}\|p\|_{L_\infty(D)}^2 \lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|p\|_{L_2(D)}^2. \end{align*} \par Similarly, we have \begin{equation*} |p|_{{H^1(D)}}\leq |p|_{H^1(\tilde\mathfrak{B}_{{\scriptscriptstyle D}})}\lesssim |p|_{H^1(\mathfrak{B}_{{\scriptscriptstyle D}})} \lesssim (\text{diam}\,\mathfrak{B}_{{\scriptscriptstyle D}})^{-1}\|p\|_{L_2(\mathfrak{B}_{{\scriptscriptstyle D}})}\leq (\text{diam}\,\mathfrak{B}_{{\scriptscriptstyle D}})^{-1}\|p\|_{L_2(D)}, \end{equation*} which together with \eqref{eq:SA} implies \eqref{eq:DiscreteEstimate1}. \par The estimates \eqref{eq:DiscreteEstimate2} and \eqref{eq:DiscreteEstimate3} follow immediately from \eqref{eq:G1}--\eqref{eq:G2}, \eqref{eq:PF1}--\eqref{eq:PF2} and \eqref{eq:DS2}. \end{proof} % \begin{lemma}\label{lem:Polynomial} Given any $p\in \mathbb{P}_{k-2}$ $(k\geq2)$, there exists $q\in\mathbb{P}_k$ such that $\Delta q=p$ and \begin{equation*} \|\nabla q\|_{L_2(B)}\leq C (\mathrm{diam}\, B)\|p\|_{L_2(B)} \end{equation*} where $B\subset \mathbb{R}^d$ is any ball and the positive constant $C$ depends only on $k$. \end{lemma} \begin{proof} By scaling it suffices to treat the case where $B$ is a unit ball. Since $\Delta$ maps $\mathbb{P}_k$ onto $\mathbb{P}_{k-2}$, there exists an operator $\Delta^\dag:\mathbb{P}_{k-2}\longrightarrow \mathbb{P}_k$ such that $\Delta \Delta^\dag$ is the identity operator on $\mathbb{P}_{k-2}$, and we can take $q=\Delta^\dag p$. The lemma follows from the observation that both $\|p\|_{L_2(B)}$ and $\|\Delta^\dag p\|_{L_2(B)}$ are norms on $\mathbb{P}_{k-2}$ together with a standard inverse estimate \cite{Ciarlet:1978:FEM,BScott:2008:FEM}. \end{proof} \par \section{Local Virtual Element Spaces in Two Dimensions}\label{sec:LocalVEM2D} In this section we obtain properties of the local virtual element spaces that will be used in the stability and error analyses in Section~\ref{sec:Poisson2D}. \par Let the space $\mathbb{P}_k(D)$ be the restriction of $\mathbb{P}_k$ to $D$ and the space of $\mathbb{P}_k(\p D)$ be defined by \begin{equation*} \mathbb{P}_k(\p D)=\{v\in C(\partial D):\,v\big|_e\in \mathbb{P}_k(e) \;\text{for all}\;e\in\cE_{\ssD}\}, \end{equation*} where $C(\partial D)$ is the space of continuous functions on $\partial D$, $\mathbb{P}_k(e)$ is the restriction of $\mathbb{P}_k$ to the edges $e$, and $\cE_{\ssD}$ is the set of the edges of $D$. The length of an edge $e$ is denoted by $h_e$. \subsection{The Projection $\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}$ and the Space $\cQ^k(D)$}\label{subsec:PODCQD} The Sobolev space ${H^1(D)}$ is a Hilbert space under the inner product \begin{equation}\label{eq:InnerProduct} (\!(\zeta,\eta)\!)=(\nabla \zeta,\nabla \eta) +\Big(\int_{\partial D}\zeta\,ds\Big)\Big(\int_{\partial D} \eta\,ds\Big). \end{equation} The projection operator from ${H^1(D)}$ onto $\mathbb{P}_k(D)$ with respect to $(\!(\cdot,\cdot)\!)$ is denoted by $\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}$, i.e., $\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} \zeta\in \mathbb{P}_k(D)$ satisfies \begin{equation}\label{eq:PODDef} (\!(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta,q)\!)=(\!(\zeta,q)\!) \qquad\forall\,q\in\mathbb{P}_k(D). \end{equation} In particular we have \begin{equation}\label{eq:Projection} \Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} p=p \qquad\forall\,p\in\mathbb{P}_k(D). \end{equation} \par It is straight-forward to check that \eqref{eq:PODDef} is equivalent to \begin{alignat}{3} \int_D \nabla (\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta)\cdot\nabla q\,dx&=\int_D \nabla \zeta\cdot \nabla q\,dx\label{eq:POD1}\\ &=\int_{\partial D} \zeta(n\cdot\nabla q)\,ds-\int_D \zeta(\Delta q)\,dx &\qquad&\forall\,q\in \mathbb{P}_k(D),\notag \end{alignat} together with \begin{alignat}{3} \int_{\partial D} \Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta\,ds&=\int_{\partial D}\zeta\,ds. \label{eq:POD2} \end{alignat} \par For $k\geq 1$, the virtual element space $\cQ^k(D)\subset {H^1(D)}$ is defined by the following conditions: $v\in {H^1(D)}$ belongs to $\cQ^k(D)$ if and only if (i) the trace of $v$ on $\partial D$ belongs to $\mathbb{P}_k(\p D)$, (ii) the distribution $-\Delta v$ belongs to $\mathbb{P}_k(D)$, and (iii) we have \begin{equation}\label{eq:Condition3} \Pi_{k,D}^0\hspace{1pt} v-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v\in \mathbb{P}_{k-2}(D), \end{equation} where $\Pi_{k,D}^0\hspace{1pt}$ is the projection from $L_2(D)$ onto $\mathbb{P}_k(D)$ and $\mathbb{P}_{-1}(D)=\{0\}$. \begin{remark}\label{rem:Continuity}{\rm It follows from elliptic regularity for bounded Lipschitz domains \cite[Section~1.2]{Kenig:1994:CBMS} and conditions (i) and (ii) in the definition of $\cQ^k(D)$ that $\cQ^k(D)\subset C(\bar D)$.} \end{remark} \begin{remark}\label{rem:dof}{\rm The dimension of $\cQ^k(D)$ is the sum of the dimension of $\mathbb{P}_k(\p D)$ and the dimension of $\mathbb{P}_{k-2}(D)$ (cf. \cite{AABMR:2013:Projector}). The degrees of freedom consist of (i) the values of $v$ at the vertices of $D$ and at the points in the interior of the edges that together determine $\mathbb{P}_k(\p D)$, and (ii) the moments of $\Pi_{k-2,D}^0 v$. The set of the nodes in (i) will be denoted by $\mathcal{N}_{\p D}$.} \end{remark} \begin{remark}\label{rem:Computable} {\rm It follows from \eqref{eq:POD1} and \eqref{eq:POD2} that the polynomial $\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v$ can be computed in terms of the degrees of freedom of $v\in\cQ^k(D)$. Moreover, the polynomial $\Pi_{k,D}^0\hspace{1pt} v$ can also be computed through \eqref{eq:Condition3}.} \end{remark} \subsection{A Minimum Energy Principle}\label{subsec:MEP} The following minimum energy principle is useful for bounding the $H^1$ norm of a virtual element function. \begin{lemma}\label{lem:MEP} The inequality \begin{equation* |v|_{{H^1(D)}}\leq |\zeta|_{{H^1(D)}} \end{equation*} holds for any $v\in\cQ^k(D)$ and $\zeta\in {H^1(D)}$ such that $\zeta-v=0$ on $\partial D$ and $\Pi_{k,D}^0\hspace{1pt}(\zeta-v)=0$. \end{lemma} \begin{proof} It follows from condition (ii) in the definition of $\cQ^k(D)$ that \begin{equation*} \int_D \nabla v\cdot\nabla (\zeta-v)\,dx=\int_D (-\Delta v) (\zeta-v)dx=0 \end{equation*} and hence $ |\zeta|_{H^1(D)}^2=|\zeta-v|_{H^1(D)}^2+|v|_{H^1(D)}^2. $ \end{proof} \subsection{A Maximum Principle}\label{subsec:MP} We begin with a result from \cite[Lemma~3.3]{BLR:2016:VEM}. \begin{lemma}\label{lem:BLR} There exists a positive constant $C$, depending only on $\rho_{\!{_\ssD}}$ and $k$, such that \begin{equation}\label{eq:BLR} \|\Delta v\|_{L_2(D)}\leq C\hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|\nabla v\|_{L_2(D)} \qquad\forall\,v\in\cQ^k(D). \end{equation} \end{lemma} \begin{proof} By scaling we may assume $\hspace{1pt}h_{{\scriptscriptstyle D}}=1$. \par Let $\phi\geq0$ be a smooth (bump) function supported on the disc $\fB_{\!\ssD}$ with radius $\rho_{\!{_\ssD}}$ such that \begin{equation}\label{eq:BumpFunction} \int_D \phi\,dx=1, \quad |\phi|\lesssim 1 \quad \text{and} \quad |\nabla\phi(x)|\lesssim 1. \end{equation} We have, by the equivalence of norms on finite dimensional vector spaces, scaling and \eqref{eq:discs}, \begin{equation}\label{eq:Fundamental1} \|p\|_{L_2(D)}^2\leq \|p\|_{L_2(\tilde\fB_{\!\ssD})}^2\lesssim \|p\|_{L_2(\fB_{\!\ssD})}^2 \lesssim \int_{\fB_{\!\ssD}} p^2\phi\,dx \qquad\forall\,p\in\mathbb{P}_k. \end{equation} \par Since $\Delta v\in\mathbb{P}_k$, it follows from \eqref{eq:DiscreteEstimate1} (with $\hspace{1pt}h_{{\scriptscriptstyle D}}=1$), \eqref{eq:Fundamental1} and integration by parts that \begin{align*} \|\Delta v\|_{L_2(D)}^2&\lesssim \int_{\fB_{\!\ssD}} (\Delta v)^2\phi\,dx\\ &= -\int_{\fB_{\!\ssD}} \nabla v\cdot(\phi\nabla(\Delta v)+(\Delta v)\nabla\phi\big)dx\\ &\lesssim \|\nabla v\|_{L_2(D)} \big(\|\nabla(\Delta v)\|_{L_2(D)}+\|\Delta v\|_{L_2(D)}\big) \lesssim \|\nabla v\|_{L_2(D)}\big(\|\Delta v\|_{L_2(D)}\big), \end{align*} which implies \eqref{eq:BLR} (with $\hspace{1pt}h_{{\scriptscriptstyle D}}=1$). \end{proof} The following maximum principle will be used in the analysis of the interpolation operator in Section~\ref{subsec:Interpolation} (cf. Lemma~\ref{lem:1DInterpolationError}), and in the stability and error analyses for virtual element methods in three dimensions (cf. \eqref{eq:3DIDtbar} and Lemma~\ref{lem:3DSDBdd}). \begin{lemma}\label{lem:MaximumPrinciple} There exists a positive constant $C$, depending only on $\rho_{\!{_\ssD}}$ and $k$, such that \begin{equation*} \|v\|_{{L_\infty(D)}}\leq C\big[ \|v\|_{L_\infty(\partial D)}+|v|_{H^1(D)}\big] \qquad\forall\,v\in \cQ^k(D). \end{equation*} \end{lemma} \begin{proof} There exists $q\in \mathbb{P}_{k+2}$ such that \begin{equation}\label{eq:MP2} \Delta q=\Delta v \quad\text{and}\quad \|\nabla q\|_{L_2(\fB_{\!\ssD})} \lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\|\Delta v\|_{L_2(\fB_{\!\ssD})} \end{equation} by Lemma~\ref{lem:Polynomial} (with $p=\Delta v\in\mathbb{P}_k$). \par Without loss of generality we may assume that the mean of $q$ over $D$ is zero. Therefore we have \begin{equation* \|q\|_{{L_\infty(D)}}\lesssim |q|_{{H^1(D)}} \lesssim |q|_{H^1(\tilde\fB_{\!\ssD})} \lesssim |q|_{H^1(\fB_{\!\ssD})} \lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\|\Delta v\|_{L_2(\fB_{\!\ssD})} \lesssim |v|_{{H^1(D)}}\notag \end{equation*} by \eqref{eq:discs}, \eqref{eq:DiscreteEstimate3}, Lemma~\ref{lem:BLR}, \eqref{eq:MP2} and scaling. \par It then follows from the maximum principle for the harmonic function $v-q$ (cf. \cite{Evans:2010:PDE}) that \begin{align*} \|v\|_{{L_\infty(D)}}&\leq \|v-q\|_{{L_\infty(D)}}+\|q\|_{{L_\infty(D)}}\\ &\leq \|v-q\|_{L_\infty(\partial D)}+\|q\|_{{L_\infty(D)}} \lesssim \|v\|_{L_\infty(\partial D)}+|v|_{H^1(D)}. \end{align*} \end{proof} \subsection{The Semi-norm ${|\!|\!|\cdot|\!|\!|_{k,D}}$}\label{subsec:Norm} The semi-norm $|\!|\!|\cdot|\!|\!|_{k,D}$ on ${H^1(D)}$ is defined by \begin{equation}\label{eq:tbarNormDef} |\!|\!| \zeta|\!|\!|_{k,D}^2=\|\Pi_{k-2,D}^0\zeta\|_{L_2(D)}^2+ \hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{e\in\cE_{\ssD}}\|\Pi_{k-1,e}^0 \zeta\|_{L_2(e)}^2, \end{equation} where $\Pi_{k-1,e}^0$ is the orthogonal projection from $L_2(e)$ onto $\mathbb{P}_{k-1}(e)$. \par It follows from \eqref{eq:Trace} and \eqref{eq:tbarNormDef} that \begin{equation}\label{eq:tbarBdd} |\!|\!|\zeta|\!|\!|_{k,D}\leq C\big(\|\zeta\|_{L_2(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{{H^1(D)}}\big) \qquad\forall\, \zeta\in {H^1(D)}, \end{equation} where the positive constant $C$ depends only on $\rho_{\!{_\ssD}}$ and $k$. \par We also have, by \eqref{eq:G2} and a standard estimate for polynomials in one variable, \begin{align}\label{eq:tbarBdd2} |\!|\!| v|\!|\!|_{k,D}&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\|v\|_{L_\infty(\partial D)} +\|\Pi_{k-2,D}^0 v\|_{L_2(D)}\\ &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\Big(\sum_{p\in\mathcal{N}_{\p D}} v^2(p)\Big)^\frac12 +\|\Pi_{k-2,D}^0 v\|_{L_2(D)} \qquad \forall\,v\in\cQ^k(D),\notag \end{align} where the hidden constant depends only on $\rho_{\!{_\ssD}}$ and $k$. \subsection{Estimates for $\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}$}\label{subsec:PODEstimates} \par All the hidden constants in this subsection depend only on $\rho_{\!{_\ssD}}$ and $k$. Besides the obvious stability estimate \begin{equation}\label{eq:PODStability1} |\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta|_{{H^1(D)}}\leq |\zeta|_{{H^1(D)}}\qquad\forall\,\zeta\in {H^1(D)} \end{equation} that follows from \eqref{eq:POD1}, we also have a stability estimate for $\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}$ in terms of $\|\cdot\|_{L_2(D)}$ and the semi-norm $|\!|\!|\cdot|\!|\!|_{k,D}$. \begin{lemma}\label{lem:PODStability2} We have \begin{equation* \|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta\|_{L_2(D)}\lesssim |\!|\!| \zeta|\!|\!|_{k,D} \qquad\forall\,\zeta\in {H^1(D)}. \end{equation*} \end{lemma} \begin{proof} It follows from \eqref{eq:POD1} that \begin{align*} &\int_D \nabla(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} \zeta)\cdot\nabla(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta)\,dx =\int_{\partial D}\zeta n\cdot\nabla(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta)ds -\int_D \zeta \Delta (\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta)dx\\ &\hspace{40pt}\leq \Big(\sum_{e\in\cE_{\ssD}}\|\Pi_{k-1,e}^0\zeta\|_{L_2(e)}^2\Big)^\frac12 \Big(\sum_{e\in\cE_{\ssD}}\|\nabla\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta\|_{L_2(e)}^2\Big)^\frac12\\ &\hspace{80pt}+\|\Pi_{k-2,D}^0\zeta\|_{L_2(D)}\|\Delta(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta)\|_{L_2(D)}, \end{align*} and we have, by \eqref{eq:DiscreteEstimate0} and \eqref{eq:DiscreteEstimate1}, \begin{align*} \sum_{e\in\cE_{\ssD}}\|\nabla\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta\|_{L_2(e)}^2\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|\nabla \Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta\|_{L_2(D)}^2 \quad\text{and}\quad \|\Delta(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta)\|_{L_2(D)}&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|\nabla \Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta\|_{L_2(D)}. \end{align*} It follows that \begin{equation*} \|\nabla\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta\|_{L_2(D)} \lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\Big(\hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{e\in\cE_{\ssD}}\|\Pi_{k-1,e}^0\zeta\|_{L_2(e)}^2 +\|\Pi_{k-2,D}^0\zeta\|_{L_2(D)}^2\Big)^\frac12 =\hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}|\!|\!|\zeta|\!|\!|_{k,D}. \end{equation*} \par Moreover \eqref{eq:G2}, \eqref{eq:POD2} and \eqref{eq:tbarNormDef}, imply \begin{align*} \Big|\int_{\partial D} \Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} \zeta\,ds\Big| =\Big|\sum_{e\in\cE_{\ssD}}\int_e \Pi_{0,e}^0\zeta\,ds\Big| &\leq \sum_{e\in\cE_{\ssD}}\sqrt{h_e}\|\Pi_{k-1,e}^0\zeta\|_{L_2(e)}\\ &\lesssim \sqrt{\hspace{1pt}h_{{\scriptscriptstyle D}}} \Big(\sum_{e\in\cE_{\ssD}}\|\Pi_{k-1,e}^0\zeta\|_{L_2(e)}^2\Big)^\frac12\leq |\!|\!|\zeta|\!|\!|_{k,D}. \end{align*} \par Finally we have, by \eqref{eq:PF2}, \begin{align*} \|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta\|_{L_2(D)}&\leq \Big|\int_{\partial D} \Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} \zeta\,ds\Big| + \hspace{1pt}h_{{\scriptscriptstyle D}} \|\nabla\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta\|_{L_2(D)} \lesssim |\!|\!| \zeta|\!|\!|_{k,D}. \end{align*} \end{proof} \par We can now establish error estimates for $\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}$. \par \begin{lemma}\label{lem:PODErrors} We have \begin{alignat}{3} \|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} \zeta\|_{L_2(D)}&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{\ell+1}|\zeta|_{H^{\ell+1}(D)} &\qquad&\forall\,\zeta\in H^{\ell+1}(D),\,0\leq\ell\leq k, \label{eq:LTwoPOD}\\ |\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta|_{{H^1(D)}}&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{\ell}|\zeta|_{H^{\ell+1}(D)} &\qquad&\forall\,\zeta\in H^{\ell+1}(D),\,1\leq\ell\leq k, \label{eq:HOnePOD}\\ |\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} \zeta|_{H^2(D)}&\lesssim h_D^{\ell-1}|\zeta|_{H^{\ell+1}(D)}&\qquad& \forall\,\zeta\in H^{\ell+1}(D),\,1\leq\ell\leq k.\label{eq:HTwoPOD} \end{alignat} \end{lemma} \begin{proof} The estimate \eqref{eq:HOnePOD} follows immediately from \eqref{eq:BHEstimates}, \eqref{eq:Projection} and \eqref{eq:PODStability1}. \par In view of \eqref{eq:DiscreteEstimate1} and \eqref{eq:PODStability1}, we have \begin{equation*} |\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta|_{H^2(D)}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta|_{{H^1(D)}}\leq\hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}|\zeta|_{{H^1(D)}}, \end{equation*} which together with \eqref{eq:BHEstimates} and \eqref{eq:Projection} implies \eqref{eq:HTwoPOD}. \par Similarly we have, by \eqref{eq:tbarBdd} and Lemma~\ref{lem:PODStability2} \begin{equation*} \|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta\|_{L_2(D)}\lesssim |\!|\!| \zeta|\!|\!|_{k,D} \lesssim \|\zeta\|_{L_2(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{{H^1(D)}}, \end{equation*} which together with \eqref{eq:BHEstimates} and \eqref{eq:Projection} implies \eqref{eq:LTwoPOD}. \end{proof} \subsection{Estimates for $\Pi_{k,D}^0\hspace{1pt}$}\label{subsec:PDZEstimates} All the hidden constants in this subsection only depend on $\rho_{\!{_\ssD}}$ and $k$. We have an obvious stability estimate \begin{equation}\label{eq:PDZLTwo} \|\Pi_{k,D}^0\hspace{1pt} \zeta\|_{L_2(D)}\leq \|\zeta\|_{L_2(D)} \qquad\forall\,\zeta\in L_2(D) \end{equation} % and an obvious relation \begin{equation}\label{eq:PDZInvariance} \Pi_{k,D}^0\hspace{1pt} q=q\qquad\forall\,q\in \mathbb{P}_k(D). \end{equation} % \par It follows from \eqref{eq:BHEstimates}, \eqref{eq:PDZLTwo} and \eqref{eq:PDZInvariance} that \begin{equation} \label{eq:PDZLTwoError} \|\zeta-\Pi_{k,D}^0\hspace{1pt}\zeta\|_{L_2(D)}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{\ell+1}|\zeta|_{H^{\ell+1}(D)} \qquad\forall\,\zeta\in H^{\ell+1}(D),\, 0\leq\ell\leq k. \end{equation} \par We also have a stability estimate for $\Pi_{k,D}^0\hspace{1pt}$ in $|\cdot|_{{H^1(D)}}$. \begin{lemma}\label{lem:PDZHOne} We have % \begin{equation}\label{eq:PDZHOne} |\Pi_{k,D}^0\hspace{1pt} \zeta|_{{H^1(D)}}\lesssim |\zeta|_{{H^1(D)}} \qquad\forall\,\zeta\in {H^1(D)}. \end{equation} \end{lemma} \begin{proof} This is a consequence of \eqref{eq:DiscreteEstimate1}, \eqref{eq:PODStability1}, \eqref{eq:LTwoPOD} and \eqref{eq:PDZLTwoError}: \begin{align*} |\Pi_{k,D}^0\hspace{1pt} \zeta|_{{H^1(D)}}&\leq |\Pi_{k,D}^0\hspace{1pt} \zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta|_{{H^1(D)}}+|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta|_{{H^1(D)}}\\ &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|\Pi_{k,D}^0\hspace{1pt} \zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta\|_{L_2(D)}+|\zeta|_{{H^1(D)}}\\ &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\big(\|\Pi_{k,D}^0\hspace{1pt} \zeta-\zeta\|_{L_2(D)}+\|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta\|_{L_2(D)}\big) +|\zeta|_{{H^1(D)}} \lesssim |\zeta|_{{H^1(D)}}. \end{align*} \end{proof} \par We can then derive error estimates for $\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}$ by combining the Bramble-Hilbert estimates and Lemma~\ref{lem:PDZHOne}. \begin{lemma}\label{lem:PDZErrors} We have \begin{alignat}{3} |\zeta-\Pi_{k,D}^0\hspace{1pt}\zeta|_{{H^1(D)}}&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^\ell|\zeta|_{H^{\ell+1}(D)} &\qquad&\forall\,\zeta\in H^{\ell+1}(D), \, 1\leq\ell\leq k, \label{eq:PDZHOneError}\\ |\zeta-\Pi_{k,D}^0\hspace{1pt}\zeta|_{H^2(D)}&\lesssim h^{\ell-1}|\zeta|_{H^{\ell+1}(D)} &\qquad&\forall\,\zeta\in H^{\ell+1}(D), \, 1\leq\ell\leq k.\label{eq:PDZHTwoError} \end{alignat} \end{lemma} \begin{proof} In view of \eqref{eq:PDZInvariance}, the estimate \eqref{eq:PDZHOneError} follows from \eqref{eq:BHEstimates} and \eqref{eq:PDZHOne}. \par Similarly the estimate \eqref{eq:PDZHTwoError} follows from \eqref{eq:BHEstimates}, \eqref{eq:PDZInvariance} and the inequality \begin{equation*} |\Pi_{k,D}^0\hspace{1pt}\zeta|_{H^2(D)}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}|\Pi_{k,D}^0\hspace{1pt}\zeta|_{H^1(D)} \lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}|\zeta|_{H^1(D)} \end{equation*} obtained from \eqref{eq:DiscreteEstimate1} and \eqref{eq:PDZHOne}. \end{proof} The following is another useful estimate. \begin{lemma}\label{lem:PDZcQD} We have \begin{equation*} \|\Pi_{k,D}^0\hspace{1pt} v\|_{L_2(D)}\lesssim C |\!|\!| v|\!|\!|_{k,D} \qquad\forall\,v\in \cQ^k(D). \end{equation*} \end{lemma} \begin{proof} Let $v\in\cQ^k(D)$ be arbitrary. It follows from \eqref{eq:Condition3} that \begin{align*} \|\Pi_{k,D}^0\hspace{1pt} v\|_{L_2(D)}^2&=\|\Pi_{k-2,D}^0 v\|_{L_2(D)}^2+\|(\Pi_{k,D}^0\hspace{1pt}-\Pi_{k-2,D}^0)v\|_{L_2(D)}^2\\ &=\|\Pi_{k-2,D}^0 v\|_{L_2(D)}^2+\|(\Pi_{k,D}^0\hspace{1pt}-\Pi_{k-2,D}^0)\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v\|_{L_2(D)}^2\\ &\leq \|\Pi_{k-2,D}^0 v\|_{L_2(D)}^2+\|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v\|_{L_2(D)}^2, \end{align*} which together with \eqref{eq:tbarNormDef} and Lemma~\ref{lem:PODStability2} completes the proof. \end{proof} \subsection{Inverse Estimates}\label{subsec:InverseEstimate} These are estimates that bound the norm $|v|_{H^1(D)}$ of a virtual element function $v\in\cQ^k(D)$ in terms of $\|\Pi_{k-2,D}^0 v\|_{L_2(D)}$ and norms that only involve the boundary data of $v$. They are crucial for the stability analysis of virtual element methods in Section~\ref{subsec:DiscreteProblem}. \par We begin with a key lemma. \begin{lemma}\label{lem:Fundamental} There exists a positive constant $C$ depending only on $\rho_{\!{_\ssD}}$ and $k$, such that \begin{equation}\label{eq:Fundamental} |v|_{{H^1(D)}}\leq C\big[ \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}|\!|\!| v|\!|\!|_{k,D}+|v|_{H^{1/2}(\partial D)}\big]. \end{equation} \end{lemma} \begin{proof} By scaling we may assume $\hspace{1pt}h_{{\scriptscriptstyle D}}=1$. \par Let $\mathrm {Tr}^\dag$ be the lifting operator from Section~\ref{subsec:Lifting}. The function $w=\mathrm {Tr}^\dag v\in H^1(D)$ satisfies $w=v$ on $\partial D$ and \begin{equation}\label{eq:w} \|w\|_{H^1(D)}\lesssim |v|_{H^{1/2}(D)}. \end{equation} \par Let $\phi$ be the same (bump) function in the proof of Lemma~\ref{lem:MaximumPrinciple} and $\zeta=w+p\phi$, where the polynomial $p\in \mathbb{P}_k(D)$ is determined by \begin{equation*} \int_D (\zeta-v)q\,dx=0 \qquad\forall\,q\in \mathbb{P}_k(D), \end{equation*} or equivalently \begin{equation}\label{eq:Fundamental2} \int_D pq\phi\,dx=\int_D (v-w)q\,dx=\int_D (\Pi_{k,D}^0\hspace{1pt} v-w)q\,dx \qquad\forall\,q\in \mathbb{P}_k(D). \end{equation} Then we have \begin{equation}\label{eq:Fundamental3} |v|_{{H^1(D)}}\leq |\zeta|_{{H^1(D)}} \end{equation} by Lemma~\ref{lem:MEP} and, in view of \eqref{eq:DiscreteEstimate1}, \eqref{eq:BumpFunction} and \eqref{eq:w}, \begin{equation}\label{eq:Fundamental4} |\zeta|_{{H^1(D)}}\leq |w|_{{H^1(D)}}+|p\phi|_{{H^1(D)}} \lesssim |w|_{{H^1(D)}}+\|p\|_{L_2(D)}\lesssim |v|_{H^{1/2}(\partial D)}+\|p\|_{L_2(D)}. \end{equation} \par Note that \eqref{eq:Fundamental1} and \eqref{eq:Fundamental2} imply \begin{equation*} \|p\|_{L_2(D)}\lesssim \|\Pi_{k,D}^0\hspace{1pt} v-w\|_{L_2(D)} \end{equation*} and hence \begin{equation}\label{eq:Fundamental5} \|p\|_{L_2(D)} \lesssim \|\Pi_{k,D}^0\hspace{1pt} v\|_{L_2(D)}+\|w\|_{L_2(D)}\lesssim |\!|\!| v|\!|\!|_{k,D}+|v|_{H^{1/2}(\partial D)} \end{equation} by Lemma~\ref{lem:PDZcQD} and \eqref{eq:w}. \par The estimate \eqref{eq:Fundamental} (with $\hspace{1pt}h_{{\scriptscriptstyle D}}=1$) follows from \eqref{eq:Fundamental3}--\eqref{eq:Fundamental5}. \end{proof} \begin{lemma}\label{lem:InverseEstimate1} There exists a positive constant $C$, depending only $\rho_{\!{_\ssD}}$ and $k$, such that \begin{equation}\label{eq:InverseEstimate1} |v|_{{H^1(D)}}\leq C \big[\hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}|\!|\!| v|\!|\!|_{k,D} +\hspace{1pt}h_{{\scriptscriptstyle D}}^{1/2}\|\partial v/\partial s\|_{L_2(\partial D)}\big] \qquad\forall\,v\in\cQ^k(D), \end{equation} where $\partial v/\partial s$ is a tangential derivative of $v$. \end{lemma} \begin{proof} The estimate \eqref{eq:InverseEstimate1} follows immediately from \eqref{eq:HalfAndOne} and Lemma~\ref{lem:Fundamental} . \end{proof} \begin{lemma}\label{lem:InverseEstimate2} There exists a positive constant $C$, depending only on $\rho_{\!{_\ssD}}$, $|\cE_{\ssD}|$ and $k$, such that \begin{equation}\label{eq:InverseEstimate2} |v|_{{H^1(D)}}\leq C \big[ \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}|\!|\!| v|\!|\!|_{k,D} +\sqrt{\ln(1+\ \tau_{\!\ssD})}\|v\|_{L_\infty(\partial D)} \big]\qquad\forall\,v\in \cQ^k(D), \end{equation} where \begin{equation}\label{eq:TauD} \tau_{\!\ssD}= \frac{\max_{e\in\cE_{\ssD}}h_e}{\min_{e\in\cE_{\ssD}}h_e}. \end{equation} \end{lemma} \begin{proof} According to \cite[Lemma~5.1]{BLR:2016:VEM}, we have \begin{equation}\label{eq:HalfAndInfty} |v|_{H^{1/2}(\partial D)}\leq C \sqrt{\ln(1+\tau_{\!\ssD})}\|v\|_{L_\infty(\partial D)} \qquad\forall\,v\in\cQ^k(D), \end{equation} where the positive constant $C$ only depends on $\rho_{\!{_\ssD}}$, $|\cE_{\ssD}|$ and $k$. \par The estimate \eqref{eq:InverseEstimate2} follows from Lemma~\ref{lem:Fundamental} and \eqref{eq:HalfAndInfty}. \end{proof} \par Combining \eqref{eq:tbarBdd2} and \eqref{eq:InverseEstimate2}, we have the following corollary. \begin{corollary}\label{cor:InverseEstimate3} There exists a positive constant $C$, depending only on $\rho_{\!{_\ssD}}$, $|\cE_{\ssD}|$ and $k$, such that \begin{equation*} |v|_{H^1(D)}\leq C\big[ \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|\Pi_{k-2,D}^0 v\|_{L_2(D)}+\sqrt{\ln(1+\tD)} \|v\|_{L_\infty(\partial D)}\big] \qquad\forall\,v\in\cQ^k(D). \end{equation*} \end{corollary} \subsection{The Interpolation Operator}\label{subsec:Interpolation} For $s>1$ the interpolation operator $I_{k,D}:H^s(D)\longrightarrow\cQ^k(D)$ is defined by the condition that $\zeta$ and $I_{k,D}\zeta$ share the same degrees of freedom, i.e., $I_{k,D}\zeta$ agrees with $\zeta$ at the nodes in $\mathcal{N}_{\p D}$ and \begin{equation}\label{eq:IDDef} \Pi_{k-2,D}^0I_{k,D}\zeta=\Pi_{k-2,D}^0\zeta. \end{equation} \par It is clear that \begin{equation}\label{eq:IDInvariance} I_{k,D} q=q\qquad\forall\,q\in\mathbb{P}_k(D), \end{equation} and by a standard estimate for polynomials in one variable, \begin{equation}\label{eq:TrivialBdd} \|I_{k,D} \zeta\|_{L_\infty(\partial D)}\leq C \max_{p\in\mathcal{N}_{\p D}}|\zeta(p)| \leq\|\zeta\|_{L_\infty(\partial D)}\qquad \forall\,\zeta\in H^s(D) \;\text{and}\; s>1, \end{equation} where the positive constant $C$ only depends on $k$. \par For the three dimensional Poisson problem, if the solution belongs to $H^{\ell+1}({\Omega})$, then its restriction to a face $F$ of a polyhedral subdomain belongs to $H^{\ell+\frac12}(F)$. Therefore below we also consider the interpolants of functions in $H^{\ell+\frac12}(D)$. \par We begin with several stability estimates for the interpolation operator. \begin{lemma}\label{lem:IDtbar} There exists a positive constant $C$, depending only on $\rho_{\!{_\ssD}}$ and $k$, such that \begin{alignat}{3} \tbarI_{k,D}\zeta|\!|\!|_{k,D}&\leq C\big[\|\zeta\|_{L_2(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{H^1(D)} +\hspace{1pt}h_{{\scriptscriptstyle D}}^2|\zeta|_{H^2(D)}\big] &\qquad&\forall\,\zeta\in H^2(D),\label{eq:IDtbar1}\\ \tbarI_{k,D}\zeta|\!|\!|_{k,D}&\leq C\big[\|\zeta\|_{L_2(D)} +\hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{H^1(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}^{3/2}|\zeta|_{H^{3/2}(D)}\big] &\qquad&\forall\,\zeta\in H^{3/2}(D).\label{eq:IDtbar2} \end{alignat} \end{lemma} \begin{proof} Let $\zeta\in {H^2(D)}$ (resp., $H^{3/2}(D)$) be arbitrary. From \eqref{eq:G2}, \eqref{eq:tbarBdd2}, \eqref{eq:IDDef} and \eqref{eq:TrivialBdd}, we have \begin{align* \tbarI_{k,D}\zeta|\!|\!|_{k,D}&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\|I_{k,D}\zeta\|_{L_\infty(\partial D)} +\|\Pi_{k-2,D}^0I_{k,D}\zeta\|_{L_2(D)}\\ &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\|\zeta\|_{L_\infty(\partial D)}+\|\Pi_{k-2,D}^0 \zeta\|_{L_2(D)}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\|\zeta\|_{L_\infty(D)},\notag \end{align*} which together with \eqref{eq:Sobolev} (resp., \eqref{eq:Sobolev2}) implies \eqref{eq:IDtbar1} (resp., \eqref{eq:IDtbar2}). \end{proof} \begin{lemma}\label{lem:IDFundamental1} We have \begin{align} |I_{k,D}\zeta|_{{H^1(D)}}&\lesssim |\zeta|_{{H^1(D)}}+\hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{H^2(D)} \label{eq:IDHOne}\\ \intertext{for all $\zeta\in{H^2(D)}$, and} |I_{k,D}\zeta|_{{H^1(D)}}&\lesssim |\zeta|_{{H^1(D)}}+\hspace{1pt}h_{{\scriptscriptstyle D}}^{1/2}|\zeta|_{H^{3/2}(D)}\label{eq:IDHOneHalf} \end{align} for all $\zeta\in H^{3/2}(D)$, where the hidden constants only depend on $\rho_{\!{_\ssD}}$, $|\cE_{\ssD}|$ and $k$. \end{lemma} \begin{proof} Let $\zeta\in H^2(D)$ be arbitrary and $\bar\zeta_D$ be the mean of $\zeta$ over $D$. Since $I_{k,D}\bar\zeta_D=\bar\zeta_D$, it follows from \eqref{eq:InverseEstimate1} that \begin{align*} |I_{k,D}\zeta|_{{H^1(D)}}^2&=|I_{k,D}(\zeta-\bar\zeta_D)|_{H^1(D)}^2\\ &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-2}\tbarI_{k,D}(\zeta-\bar\zeta_D)|\!|\!|_{k,D}^2 +\hspace{1pt}h_{{\scriptscriptstyle D}}\|\partial (I_{k,D}\zeta)/\partial s\|_{L_2(\partial D)}^2. \end{align*} We have, by a standard interpolation estimate in one dimension and \eqref{eq:Trace} (applied to the first order derivatives of $\zeta$), \begin{align*} \hspace{1pt}h_{{\scriptscriptstyle D}}\|\partial(I_{k,D} \zeta)/\partial s\|_{L_2(\partial D)}^2&\lesssim\sum_{e\in\cE_{\ssD}} \hspace{1pt}h_{{\scriptscriptstyle D}}\|\partial\zeta/\partial s\|_{L_2(e)}^2\\ &\lesssim \sum_{e\in\cE_{\ssD}}\big[|\zeta|_{H^1(D)}^2+ \hspace{1pt}h_{{\scriptscriptstyle D}}^2|\zeta|_{H^2(D)}^2\big]\lesssim |\zeta|_{H^1(D)}^2+\hspace{1pt}h_{{\scriptscriptstyle D}}^2|\zeta|_{H^2(D)}^2. \end{align*} \par These two estimates together with \eqref{eq:PF2} and \eqref{eq:IDtbar1} imply \eqref{eq:IDHOne}. \par Similarly we obtain \eqref{eq:IDHOneHalf} by replacing \eqref{eq:Trace} with the estimate in Lemma~\ref{lem:CZ1}. \end{proof} \begin{lemma}\label{lem:IDFundamental2} We have \begin{align} \|I_{k,D}\zeta\|_{L_2(D)}&\lesssim \|\zeta\|_{L_2(D)} +\hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{{H^1(D)}}+\hspace{1pt}h_{{\scriptscriptstyle D}}^2|\zeta|_{H^2(D)} \label{eq:IDLTwo}\\ \intertext{for all $\zeta\in {H^2(D)}$, and} \|I_{k,D}\zeta\|_{L_2(D)}&\lesssim \|\zeta\|_{L_2(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{{H^1(D)}}+\hspace{1pt}h_{{\scriptscriptstyle D}}^{3/2}|\zeta|_{H^{3/2}(D)} \label{eq:IDLTwoHalf} \end{align} for all $\zeta\in H^{3/2}(D)$, where the hidden constants only depend on $\rho_{\!{_\ssD}}$, $|\cE_{\ssD}|$ and $k$. \end{lemma} \begin{proof} From Lemma~\ref{lem:PODStability2} and \eqref{eq:LTwoPOD} we have \begin{align}\label{eq:LTwoID} \|I_{k,D}\zeta\|_{L_2(D)}&\lesssim \|I_{k,D}\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta\|_{L_2(D)} +\|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta\|_{L_2(D)}\\ &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}|I_{k,D}\zeta|_{H^1(D)}+\tbarI_{k,D}\zeta|\!|\!|_{k,D}.\notag \end{align} The estimates \eqref{eq:IDLTwo} and \eqref{eq:IDLTwoHalf} follow from \eqref{eq:LTwoID}, Lemma~\ref{lem:IDtbar} and Lemma~\ref{lem:IDFundamental1}. \end{proof} \par We can now derive error estimates for the interpolation operator. \begin{lemma}\label{lem:InterpolationError} We have, for $1\leq \ell\leq k$, \begin{alignat}{3} \|\zeta-I_{k,D}\zeta\|_{L_2(D)}+\|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta\|_{L_2(D)} &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{\ell+1}|\zeta|_{H^{\ell+1}(D)}&\qquad& \forall\,\zeta\in H^{\ell+1}(D), \label{eq:LTwoPODID}\\ |\zeta-I_{k,D}\zeta|_{{H^1(D)}}+|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta|_{{H^1(D)}} &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{\ell}|\zeta|_{H^{\ell+1}(D)}&\qquad& \forall\,\zeta\in H^{\ell+1}(D), \label{eq:HOnePODID}\\ |\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta|_{H^2(D)}&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{\ell-1}|\zeta|_{H^{\ell+1}(D)} &\qquad&\forall\,\zeta\in H^{\ell+1}(D), \label{eq:HTwoPODID} \end{alignat} and \begin{alignat}{3} \|\zeta-I_{k,D} \zeta\|_{L_2(D)}&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{\ell+\frac12}|\zeta|_{H^{\ell+\frac12}(D)} &\qquad&\forall\,\zeta\in H^{\ell+\frac12}(D), \label{eq:HalfIDLTwoError}\\ |\zeta-I_{k,D} \zeta|_{H^1(D)}&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{\ell-\frac12}|\zeta|_{H^{\ell+\frac12}(D)} &\qquad&\forall\,\zeta\in H^{\ell+\frac12}(D), \label{eq:HalfIDHOneError} \end{alignat} where the hidden constants only depend on $\rho_{\!{_\ssD}}$, $|\cE_{\ssD}|$ and $k$. \end{lemma} \begin{proof} In view of \eqref{eq:tbarBdd2}, Lemma~\ref{lem:PODStability2}, \eqref{eq:IDtbar1} and \eqref{eq:IDLTwo}, we have, for any $\zeta\in H^2(D)$, \begin{align*} \|I_{k,D}\zeta\|_{L_2(D)}+\|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} I_{k,D}\zeta\|_{L_2(D)} &\lesssim \|I_{k,D}\zeta\|_{L_2(D)}+\tbarI_{k,D}\zeta|\!|\!|_{k,D}\\ &\lesssim \|\zeta\|_{L_2(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{{H^1(D)}}+\hspace{1pt}h_{{\scriptscriptstyle D}}^2|\zeta|_{H^2(D)}, \end{align*} which together with \eqref{eq:BHEstimates}, \eqref{eq:Projection} and \eqref{eq:IDInvariance} implies \eqref{eq:LTwoPODID}. \par From \eqref{eq:PODStability1} and \eqref{eq:IDHOne} we also have, for any $\zeta\in H^2(D)$, \begin{align*} & |I_{k,D}\zeta|_{H^1({\Omega})}+|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta|_{{H^1(D)}}\leq 2|I_{k,D}\zeta|_{{H^1(D)}} \lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|\zeta\|_{L_2(D)}+|\zeta|_{{H^1(D)}}+\hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{H^2(D)}, \end{align*} which together with \eqref{eq:BHEstimates}, \eqref{eq:Projection} and \eqref{eq:IDInvariance} implies \eqref{eq:HOnePODID}. \par Similarly, by using the relation \begin{align*} |\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta|_{H^2(D)}&=|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta-\Pi_{1,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta|_{H^2(D)}\\ &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta-\Pi_{1,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta|_{H^1(D)}\\ &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}|I_{k,D}\zeta|_{H^1(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}|\zeta|_{H^1(D)} \lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-2}\|\zeta\|_{L_2(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}|\zeta|_{H^1(D)}+|\zeta|_{H^2(D)} \end{align*} that follows from \eqref{eq:DiscreteEstimate1}, \eqref{eq:PODStability1} and \eqref{eq:IDHOne}, we can establish \eqref{eq:HTwoPODID} through \eqref{eq:BHEstimates}, \eqref{eq:Projection} and \eqref{eq:IDInvariance}. \par Finally we obtain \eqref{eq:HalfIDLTwoError} and \eqref{eq:HalfIDHOneError} by replacing \eqref{eq:BHEstimates} (resp., \eqref{eq:IDtbar1}, \eqref{eq:IDHOne} and \eqref{eq:IDLTwo}) with \eqref{eq:BHEstimates2} (resp., \eqref{eq:IDtbar2}, \eqref{eq:IDHOneHalf} and \eqref{eq:IDLTwoHalf}) in the arguments for \eqref{eq:LTwoPODID} and \eqref{eq:HOnePODID}. \end{proof} \par The proof for the following result is similar. \begin{lemma}\label{lem:LTwoIDSpecial} There exists a positive constant $C$, depending only on $\rho_{\!{_\ssD}}$, $|\cE_{\ssD}|$ and $k$, such that % \begin{equation* \|I_{k,D} \zeta-\Pi_{1,D}^0I_{k,D} \zeta\|_{L_2(D)}\leq C \hspace{1pt}h_{{\scriptscriptstyle D}}^2|\zeta|_{H^2(D)} \qquad\forall\,\zeta\in H^2(D). \end{equation*} \end{lemma} \par We also have interpolation error estimates in the $L_\infty$ norm. \begin{lemma}\label{lem:1DInterpolationError} There exists a positive constant $C$, depending only on $\rho_{\!{_\ssD}}$, $N$ and $k$, such that \begin{alignat}{3} \|\zeta-I_{k,D}\zeta\|_{L_\infty(D)}&\leq C \hspace{1pt}h_{{\scriptscriptstyle D}}^\ell|\zeta|_{H^{\ell+1}(D)}&\qquad& \forall\,\zeta\in H^{\ell+1}(D) \;\text{and}\; 1\leq\ell\leq k,\label{eq:1DInterpolationError}\\ \|\zeta-I_{k,D}\zeta\|_{L_\infty(D)}&\leq C\hspace{1pt}h_{{\scriptscriptstyle D}}^{\ell-\frac12}|\zeta|_{H^{\ell+\frac12}(D)} &\qquad& \forall\,\zeta\in H^{\ell+\frac12}(D) \;\text{and}\;1\leq\ell\leq k. \label{eq:1DInterpolationErrorHalf} \end{alignat} \end{lemma} \begin{proof} It follows from \eqref{eq:Sobolev}, Lemma~\ref{lem:MaximumPrinciple}, \eqref{eq:TrivialBdd} and \eqref{eq:IDHOne} that, for any $\zeta\in {H^2(D)}$, \begin{equation*} \|I_{k,D}\zeta\|_{L_\infty(D)}\lesssim \|I_{k,D}\zeta\|_{L_\infty(\partial D)}+|I_{k,D}\zeta|_{H^1(D)} \lesssim\hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|\zeta\|_{L_2(D)}+|\zeta|_{H^1(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{H^2(D)}, \end{equation*} which together with \eqref{eq:BHEstimates} and \eqref{eq:IDInvariance} imply \eqref{eq:1DInterpolationError}. \par The proof for \eqref{eq:1DInterpolationErrorHalf} is similar, but with \eqref{eq:Sobolev} (resp., \eqref{eq:BHEstimates} and \eqref{eq:IDHOne}) replaced by \eqref{eq:Sobolev2} (resp., \eqref{eq:BHEstimates2} and \eqref{eq:IDHOneHalf}). \end{proof} \subsection{The Null Space of $\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}$}\label{subsec:Kernel} Let $\mathcal{N}(\POD)=\{v\in\cQ^k(D):\,\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v=0\}$ be the null space of the projection $\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}$. The inverse estimates in Section~\ref{subsec:InverseEstimate} can be simplified for functions in $\mathcal{N}(\POD)$. \begin{lemma}\label{lem:KerPOD} There exists a positive constant $C$, depending only on $\rho_{\!{_\ssD}}$ and $k$, such that \begin{equation}\label{eq:KerPOD} |\!|\!| v|\!|\!|_{k,D}^2\leq C \hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{e\in\cE_{\ssD}}\|\Pi_{k-1,e}^0 v\|_{L_2(e)}^2 \qquad\forall\,v\in \mathcal{N}(\POD). \end{equation} \end{lemma} \begin{proof} We can assume $k\geq2$ since $\Pi_{k-2,D}\hspace{1pt}v=0$ for $k=1$. \par Let $v\in\mathcal{N}(\POD)$ be arbitrary. It follows from \eqref{eq:POD1} that \begin{equation}\label{eq:KerPOD1} \int_D v(\Delta q)\,dx=\int_{\partial D}v(n\cdot\nabla q)\,ds \qquad\forall \,q\in \mathbb{P}_k(D). \end{equation} \par According to \eqref{eq:discs} and Lemma~\ref{lem:Polynomial}, given any $p\in\mathbb{P}_{k-2}$, there exists $q\in\mathbb{P}_k$ such that $\Delta q=p$ and \begin{equation}\label{eq:KerPOD12} \|\nabla q\|_{L_2(D)}\leq\|\nabla q\|_{L_2(\tilde\fB_{\!\ssD})}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\|p\|_{L_2(\tilde\fB_{\!\ssD})} \lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\|p\|_{L_2(\fB_{\!\ssD})}\leq \hspace{1pt}h_{{\scriptscriptstyle D}}\|p\|_{L_2(D)}. \end{equation} \par It follows from \eqref{eq:DiscreteEstimate0}, \eqref{eq:KerPOD1} and \eqref{eq:KerPOD12} that \begin{align*} \int_D vp\,dx &\leq \Big(\sum_{e\in\cE_{\ssD}}\|\Pi_{k-1,e}^0 v\|_{L_2(e)}^2\Big)^\frac12 \Big(\sum_{e\in\cE_{\ssD}}\|\nabla q\|_{L_2(e)}^2\Big)^\frac12\\ &\lesssim \Big(\sum_{e\in\cE_{\ssD}}\|\Pi_{k-1,e}^0 v\|_{L_2(e)}^2\Big)^\frac12 \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1/2} \|\nabla q\|_{L_2(D)}\\ &\lesssim \Big(\sum_{e\in\cE_{\ssD}}\|\Pi_{k-1,e}^0 v\|_{L_2(e)}^2\Big)^\frac12 \hspace{1pt}h_{{\scriptscriptstyle D}}^{1/2} \|p\|_{L_2(D)} \end{align*} and hence \begin{equation}\label{eq:KerPod13} \|\Pi_{k-2,D}^0v\|_{L_2(D)}=\max_{p\in\mathbb{P}_{k-2}}\int_D vp\,dx /\|p\|_{L_2(D)} \lesssim \Big(\hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{e\in\cE_{\ssD}}\|\Pi_{k-1,e}^0 v\|_{L_2(e)}^2\Big)^\frac12. \end{equation} \par The estimate \eqref{eq:KerPOD} follows from \eqref{eq:tbarNormDef} and \eqref{eq:KerPod13}. \end{proof} \begin{lemma}\label{lem:BdryEst} For any $v\in\mathbb{P}_k(\p D)$ that vanishes at some point on $\partial D$, we have \begin{equation*} \|v\|_{L_2(\partial D)} \leq C \hspace{1pt}h_{{\scriptscriptstyle D}} \|\partial v/\partial s\|_{L_2(\partial D)}, \end{equation*} where $\partial v /\partial s$ denotes a tangential derivative of $v$ along $\partial D$ and the positive constant $C$ only depends on $k$. \end{lemma} \begin{proof} It follows from a Poincar\'e-Friedrichs inequality on the circle $\partial\fB_{\!\ssD}$ that \begin{equation*} \|\zeta\|_{L_2(\partial\fB_{\!\ssD})}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{H^1(\partial\fB_{\!\ssD})} \end{equation*} for any $\zeta\in H^1(\partial\fB_{\!\ssD})$ that vanishes at some point on $\partial\fB_{\!\ssD}$. The lemma then follows from \eqref{eq:L2Iso1} and \eqref{eq:H1Iso1}. \end{proof} \par Note that every $v\in \mathcal{N}(\POD)$ must vanish at some point on $\partial{\Omega}$ because $\int_{\partial D}v\,ds=0$ by \eqref{eq:POD2}. Therefore, in view of Lemma~\ref{lem:KerPOD} and Lemma~\ref{lem:BdryEst}, the inverse estimates \eqref{eq:InverseEstimate1} and \eqref{eq:InverseEstimate2} can be simplified to \begin{alignat*}{3} |v|_{{H^1(D)}}&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{1/2}\|\partial v/\partial s\|_{L_2(\partial D)} &\qquad&\forall\,v\in \mathcal{N}(\POD)\\ \intertext{with a hidden constant depending only on $\rho_{\!{_\ssD}}$ and $k$, and} |v|_{{H^1(D)}}&\lesssim \sqrt{\ln(1+\tau_{\!\ssD})}\|v\|_{L_\infty(\partial D)} &\qquad&\forall\,v\in \mathcal{N}(\POD) \end{alignat*} with a hidden constant that also depends on $|\cE_{\ssD}|$. \par Hence we have \begin{alignat}{3} |v|_{H^1(D)}^2&=|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v|_{H^1(D)}^2+|v-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v|_{H^1(D)}^2\label{eq:EnergyBdd1}\\ &\lesssim |\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v|_{H^1(D)}^2+\hspace{1pt}h_{{\scriptscriptstyle D}}\|\partial (v-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v)/\partial s\|_{L_2(\partial D)}^2 &\qquad&\forall\,v\in\cQ^k(D),\notag\\ \intertext{with a hidden constant depending only on $\rho_{\!{_\ssD}}$ and $k$, and also} |v|_{H^1(D)}^2&\lesssim |\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v|_{H^1(D)}^2+\ln(1+\tau_{\!\ssD})\|v-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v\|_{L_\infty(\partial D)}^2 &\qquad&\forall\,v\in\cQ^k(D),\label{eq:EnergyBdd2} \end{alignat} with a hidden constant that also depends on $|\cE_{\ssD}|$. \section{The Poisson Problem in Two Dimensions}\label{sec:Poisson2D} In this section we consider virtual element methods for the Poisson problem \eqref{eq:Poisson} in two dimensions and establish error estimates under two global shape regularity assumptions. \par Let $\mathcal{T}_h$ be a triangulation of the convex polygon ${\Omega}\subset\mathbb{R}^2$ by polygonal subdomains, where $h=\displaystyle\max_{D\in\mathcal{T}_h}\hspace{1pt}h_{{\scriptscriptstyle D}}$ is the mesh parameter. The global virtual finite element space $\cQ^k_h$ is given by $$\cQ^k_h=\{v\in H^1_0({\Omega}):\, v\big|_D\in\cQ^k(D) \quad \forall\,D\in\mathcal{T}_h\}.$$ The space of (discontinuous) piecewise polynomials of degree $\leq k$ with respect to $\mathcal{T}_h$ is denoted by $\cP^k_h$. \par The operators $\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} :H^1({\Omega})\longrightarrow \cP^k_h$, $\Pi_{k,h}^0:L_2({\Omega})\longrightarrow \cP^k_h$ and $I_{k,h}:H^2({\Omega})\cap H^1_0({\Omega})\longrightarrow \cQ^k_h$ are defined in terms of their local counterparts, i.e., $$(\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\zeta)\big|_D=\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(\zeta\big|_D), \quad (\Pi_{k,h}^0\zeta)\big|_D=\Pi_{k,D}^0\hspace{1pt}(\zeta\big|_D) \quad \text{and} \quad (I_{k,h}\zeta)\big|_D=I_{k,D}(\zeta\big|_D). $$ \par The piecewise $H^1$ norm with respect to $\mathcal{T}_h$ is given by \begin{equation}\label{eq:PiecewiseHOneNOrm} |v|_{h,1}=\Big(\sum_{D\in\cT_h} |v|_{{H^1(D)}}^2\Big)^\frac12. \end{equation} \subsection{Global Shape Regularity Assumptions} \label{subsec:GlobalShapeRegularity} We assume that the local shape regularity assumption \eqref{eq:SA} is satisfied by all $D\in\mathcal{T}_h$ and impose the following global regularity assumptions. \par\noindent {\em Assumption 1}\quad There exists a positive number $\rho\in (0,1)$, independent of $h$, such that \begin{equation}\label{eq:rho} \rho_{\!{_\ssD}}\geq\rho \qquad\forall\,D\in\mathcal{T}_h. \end{equation} \par\noindent {\em Assumption 2}\quad There exists a positive integer $N$, independent of $h$, such that \begin{equation}\label{eq:N} |\cE_{\ssD}|\leq N\qquad\forall\,D\in\mathcal{T}_h. \end{equation} \par The hidden constants in the rest of Section~\ref{sec:Poisson2D} will only depend on $\rho$, $N$ and $k$. \subsection{The Discrete Problem}\label{subsec:DiscreteProblem} Let the local stabilizing bilinear form $S^D_1(\cdot,\cdot)$ and $S^D_2(\cdot,\cdot)$ be defined by \begin{align} S^D_1(w,v)&=\sum_{\mathcal{N}_{\p D}} w(p)v(p),\label{eq:SD1}\\ S^D_2(w,v)&=\hspace{1pt}h_{{\scriptscriptstyle D}}(\partial w/\partial s,\partial v/\partial s)_{L_2(\partial D)},\label{eq:SD2} \end{align} where $\partial v/\partial s$ denotes a tangential derivative of $v$ along $\partial D$. \begin{remark}\label{rem:SD}{\rm The local stability bilinear form $S^D_1(\cdot,\cdot)$ is the boundary part of the local stability bilinear form in \cite{BBCMMR:2013:VEM}. The bilinear form $S^D_2(\cdot,\cdot)$ was introduced in \cite{WRR:2016:VEM}.} \end{remark} \begin{remark}\label{rem:AlternativeSD2}{\rm We can also use the bilinear form $\tilde S^D_2(\cdot,\cdot)$ defined by \begin{equation*} \tilde S^D_2(w,v)=\sum_{e\in\cE_{\ssD}}h_e(\partial w/\partial s,\partial v/\partial s)_{L_2(e)} \end{equation*} } \end{remark} \begin{remark}\label{rem:NormEquivalence} {\rm By the equivalence of norms on finite dimensional vector spaces, we have $$\sum_{p\in\mathcal{N}_e}v^2(p)\approx \sum_{e\in\cE_{\ssD}}\|v\|_{L_\infty(e)}^2 \approx \|v\|_{L_\infty(\partial D)}^2 \qquad \forall\, v\in \mathbb{P}_k(\p D).$$ } \end{remark} \par The discrete problem for \eqref{eq:Poisson} is to find $u_h\in\cQ^k_h$ such that \begin{equation}\label{eq:DiscreteProblem} a_h(u_h,v)=(f,\Xi_h v) \qquad \forall\,v\in\cQ^k_h, \end{equation} where \begin{align} a_h(w,v)&=\sum_{D\in\cT_h} \big[a^D(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} w,\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v) +S^D(w-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} w,v-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v)\big],\label{eq:ah}\\ a^D(w,v)&=\int_D \nabla w\cdot\nabla v\,dx,\label{eq:aD} \end{align} $S^D(\cdot,\cdot)$ is either $S^D_1(\cdot,\cdot)$ or $S^D_2(\cdot,\cdot)$, and $\Xi_h:\cQ^k_h\longrightarrow\cP^k_h$ is given by \begin{equation}\label{eq:Xih} \Xi_h=\begin{cases} \Pi_{1,h}^0&\qquad\text{if $k=1,2,$}\\[4pt] \Pi_{k-2,h}^0 &\qquad\text{if $k\geq 3$}.\\ \end{cases} \end{equation} \par It follows from \eqref{eq:EnergyBdd1} and \eqref{eq:EnergyBdd2} that \begin{equation}\label{eq:Stability} |v|_{H^1({\Omega})}^2 \lesssim \alpha_h a_h(v,v) \qquad\forall\,v\in \cQ^k_h, \end{equation} where \begin{equation}\label{eq:kappah} \alpha_h=\begin{cases}\displaystyle \ln(1+\max_{D\in\mathcal{T}_h}\tau_{\!\ssD})&\qquad \text{if $S^D(\cdot,\cdot)=S^D_1(\cdot,\cdot)$}\\ 1&\qquad\text{if $S^D(\cdot,\cdot)=S^D_2(\cdot,\cdot)$} \end{cases}\;. \end{equation} The well-posedness of the discrete problem follows from the stability estimate \eqref{eq:Stability}. \par We will use the following properties of $\Xi_h$ in the error analysis. \begin{lemma}\label{lem:Xih} We have, for $1\leq\ell\leq k$, \begin{alignat}{3} (f,w-\Xi_h w)&\lesssim h^\ell|f|_{H^{\ell-1}({\Omega})} |w|_{H^1({\Omega})} &\qquad&\forall\,f\in H^{\ell-1}({\Omega}),\;w\in\cQ^k_h, \label{eq:Xih1}\\ (f,I_{k,h}\zeta-\Xi_hI_{k,h}\zeta)&\lesssim h^{\ell+1}|f|_{H^{\ell-1}({\Omega})}|\zeta|_{H^2({\Omega})}&\qquad& \forall f\in H^{\ell-1}({\Omega}),\;\zeta\in H^2({\Omega}).\label{eq:Xih2} \end{alignat} \end{lemma} \begin{proof} In view of the relation \begin{align*} (f,w-\Xi_h w)&=(f-\Pi_{k-2,h}^0 f,w-\Xi_h w)\\ &\leq \|f-\Pi_{k-2,h}^0 f\|_{L_2({\Omega})}\| w-\Xi_h w\|_{L_2({\Omega})} \leq \|f-\Pi_{\ell-2,h}^0 f\|_{L_2({\Omega})} \|w-\Pi_{0,h}^0 w\|_{L_2({\Omega})}, \end{align*} the estimate \eqref{eq:Xih1} follows from \eqref{eq:PDZLTwoError}. \par Similarly we have \begin{align*} (f,I_{k,h}\zeta-\Xi_hI_{k,h}\zeta)&=(f-\Pi_{k-2,h}^0 f,I_{k,h}\zeta-\Xi_hI_{k,h}\zeta)\\ &\leq \|f-\Pi_{k-2,h}^0 f\|_{L_2({\Omega})}\|I_{k,h}\zeta-\Xi_hI_{k,h}\zeta\|_{L_2({\Omega})}\\ &\leq \|f-\Pi_{\ell-2,h}^0f\|_{L_2({\Omega})} \|I_{k,h}\zeta-\Pi_{1,h}^0I_{k,h}\zeta\|_{L_2({\Omega})}, \end{align*} and the estimate \eqref{eq:Xih2} follows from \eqref{eq:PDZLTwoError} and Lemma~\ref{lem:LTwoIDSpecial}. Note that this is the reason why $\Xi_h$ is chosen to be $\Pi_{1,h}^0$ for $k=2$ instead of $\Pi_{k-2,h}^0=\Pi_{0,h}^0$. \end{proof} \subsection{An Abstract Error Estimate in the Energy Norm} \label{subsec:AbstractEnergyError} Let $\|\cdot\|_h=\sqrt{a_h(\cdot,\cdot)}$ be the mesh-dependent energy norm. Note that \eqref{eq:Stability} implies \begin{equation}\label{eq:HOneVEM} |v|_{H^1({\Omega})}\lesssim \sqrt{\alpha_h}\|v\|_h \qquad\forall\,v\in\cQ^k_h. \end{equation} \par The discrete problem \eqref{eq:DiscreteProblem} is defined in terms of a non-inherited symmetric positive definite bilinear form. We have a standard error estimate (cf. \cite[Lemma~10.1.7]{BScott:2008:FEM} and \cite{BSS:1972:NC}) \begin{equation}\label{eq:Standard} \|u-u_h\|_h\leq \inf_{v\in\cQ^k_h}\|u-v\|_h+\sup_{v\in \cQ^k_h} \frac{a_h(u,v)-(f,\Xi_h v)}{\|v\|_h}. \end{equation} The key is to control the numerator on the right-hand side of \eqref{eq:Standard}. \par In view of \eqref{eq:Poisson}, \eqref{eq:aDef} and \eqref{eq:POD1} we can write, for any $v\in\cQ^k_h$, \begin{align*} a_h(u,v)&=\sum_{D\in\cT_h} \big[a^D(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u,\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v)+S^D(u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u,v-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v)\big]\\ &=\sum_{D\in\cT_h} \big[a^D(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u,v)+S^D(u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u,v-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v)\big]\\ &=\sum_{D\in\cT_h} \big[a^D(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u-u,v)+S^D(u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u,v-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v)\big]+(f,v)\\ &=\sum_{D\in\cT_h} \big[a^D(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u-u,v-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v)+S^D(u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u,v-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v)\big]+(f,v), \end{align*} and hence, by \eqref{eq:PODStability1}, \eqref{eq:ah} and \eqref{eq:HOneVEM}, \begin{align}\label{eq:Numerator} a_h(u,v)-(f,\Xi_h v)&=\sum_{D\in\cT_h} \big[a^D(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u-u,v-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v) +S^D(u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u,v-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v)\big]\notag\\ &\hspace{30pt}+(f,v-\Xi_h v),\notag\\ &\lesssim \Big(\sum_{D\in\cT_h}|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u-u|_{{H^1(D)}}|v-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v|_{{H^1(D)}}\Big) +\|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u\|_h\|v\|_h\\ &\hspace{30pt}+\Big(\sup_{w\in\cQ^k_h}\frac{(f,w-\Xi_h w)}{|w|_{H^1({\Omega})}}\Big) \sqrt{\alpha_h}\|v\|_h\notag\\ &\lesssim \|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u\|_h\|v\|_h+ \Big(|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u|_{h,1}+\sup_{w\in\cQ^k_h} \frac{(f,w-\Xi_h w)}{|w|_{H^1({\Omega})}}\Big)\sqrt{\alpha_h}\|v\|_h.\notag \end{align} \par Putting \eqref{eq:Standard} and \eqref{eq:Numerator} together we arrive at the estimate \begin{equation}\label{eq:EnergyError} \|u-u_h\|_h\lesssim \|u-I_{k,h} u\|_h+\|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u\|_h +\sqrt{\alpha_h}\Big(|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u|_{h,1}+\sup_{w\in\cQ^k_h} \frac{(f,w-\Xi_h w)}{|w|_{H^1({\Omega})}}\Big). \end{equation} \par Below we will derive concrete error estimates under the assumption that the solution $u$ of \eqref{eq:Poisson} belongs to $H^{\ell+1}({\Omega})$ for $1\leq\ell\leq k$. \subsection{Concrete Error Estimates in the Energy Norm} \label{subsec:ConcreteEnergyErrors} The terms on the right-hand side of \eqref{eq:EnergyError} are estimated as follows. \par\medskip\noindent{\em Estimate for the Term Involving $f$} \par\smallskip Since $u\in H^{\ell+1}({\Omega})$ and $f=-\Delta u$, we have, by \eqref{eq:Xih1}, \begin{equation}\label{eq:RHSError} \sup_{w\in\cQ^k_h}\frac{(f,w-\Xi_h w)}{|w|_{H^1({\Omega})}} \lesssim h^{\ell}|u|_{H^{\ell+1}({\Omega})}. \end{equation} \par\medskip\noindent{\em Estimate for $|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u|_{h,1}$} \par\smallskip It follows directly from \eqref{eq:HOnePOD} that \begin{equation}\label{eq:PODError} |u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u|_{h,1}=\Big(\sum_{D\in\cT_h} |u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u|_{H^1(D)}^2\Big)^\frac12 \lesssim h^\ell|u|_{H^{\ell+1}({\Omega})}. \end{equation} \par\medskip\noindent{\em Estimate for $\|u-I_{k,h} u\|_h$} \par\smallskip We will establish the estimate \begin{equation}\label{eq:InterpolationError} \|u-I_{k,h} u\|_h\lesssim h^{\ell}|u|_{H^{\ell+1}(\O)} \end{equation} for both choices of $S^D(\cdot,\cdot)$. \par In the case where $S^D(\cdot,\cdot)=S^D_1(\cdot,\cdot)$, it follows from \eqref{eq:PODStability1}, \eqref{eq:SD1} and \eqref{eq:ah} that \begin{align}\label{eq:InterError1} \|u-I_{k,h} u\|_h^2&\lesssim \sum_{D\in\cT_h} |\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-I_{k,D} u)|_{H^1(D)}^2+\sum_{D\in\cT_h} \|(u-I_{k,D} u)-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-I_{k,D} u)\|_{L_\infty(\partial D)}^2\notag\\ &\lesssim \sum_{D\in\cT_h} \big(|u-I_{k,D} u|_{H^1(D)}^2+\|u-I_{k,D} u\|_{L_\infty(\partial D)}^2\big)\\ &\hspace{40pt}+\sum_{D\in\cT_h} |\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-I_{k,D} u)|_{L_\infty(D)}^2,\notag \end{align} and we have \begin{align}\label{eq:InterError2} \sum_{D\in\cT_h} |\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-I_{k,D} u)|_{L_\infty(D)}^2&\lesssim \sum_{D\in\cT_h}\big(\,(\,\overline{\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-I_{k,D} u)}\,)_{\partial D}^2 +|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-I_{k,D} u)|_{H^1(D)}^2\big)\notag\\ &\leq \sum_{D\in\cT_h}\big(\,(\,\overline{u-I_{k,D} u}\,)_{\partial D}^2+|u-I_{k,D} u|_{H^1(D)}^2\big)\\ &\lesssim \sum_{D\in\cT_h} \big(\|u-I_{k,D} u\|_{L_\infty(\partial D)}^2 +|u-I_{k,D} u|_{H^1(D)}^2\big),\notag \end{align} by \eqref{eq:DiscreteEstimate2}, \eqref{eq:Condition3} and \eqref{eq:PODStability1}. \par The estimate \eqref{eq:InterpolationError} now follows from \eqref{eq:HOnePODID}, \eqref{eq:1DInterpolationError}, \eqref{eq:InterError1} and \eqref{eq:InterError2}. \par In the case where $S^D(\cdot,\cdot)=S^D_2(\cdot,\cdot)$, it follows from \eqref{eq:SD2} and \eqref{eq:ah} that \begin{align*} \|u-I_{k,h} u\|_h^2&\lesssim \sum_{D\in\cT_h}|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-I_{k,D} u)|_{H^1(D)}^2 + \sum_{D\in\cT_h}\hspace{1pt}h_{{\scriptscriptstyle D}} \|\partial [\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-I_{k,D} u)]/\partial s\|_{L_2(\partial D)}^2\notag\\ &\hspace{30pt}+{\sum_{D\in\cT_h}\hspace{1pt}h_{{\scriptscriptstyle D}} \|\partial(u-I_{k,D} u)/\partial s\|_{L_2(\partial D)}^2}. \end{align*} We have \begin{equation*} \sum_{D\in\cT_h}|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-I_{k,D} u)|_{H^1(D)}^2\leq \sum_{D\in\cT_h}|u-I_{k,D} u|_{H^1(D)}^2\lesssim h^{2\ell}|u|_{H^{\ell+1}(\O)}^2 \end{equation*} by \eqref{eq:PODStability1} and \eqref{eq:HOnePODID}, and hence, in view of \eqref{eq:DiscreteEstimate0} (applied to the first order derivatives of the polynomial $\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-I_{k,D} u)$), \begin{align*} \sum_{D\in\cT_h}\hspace{1pt}h_{{\scriptscriptstyle D}} \|\partial [\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-I_{k,D} u)]/\partial s\|_{L_2(\partial D)}^2&\lesssim \sum_{D\in\cT_h} \|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-I_{k,D} u)|_{H^1(D)}^2\lesssim h^{2\ell}|u|_{H^{\ell+1}(\O)}^2. \end{align*} Finally it follows from a standard interpolation error estimate in one variable and \eqref{eq:HalfTrace} (applied to the $\ell$-th order derivatives of $u$) that \begin{align*} \sum_{D\in\cT_h}\hspace{1pt}h_{{\scriptscriptstyle D}} \|\partial(u-I_{k,D} u)/\partial s\|_{L_2(\partial D)}^2&\lesssim \sum_{D\in\cT_h}\hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{e\in\cE_{\ssD}}h_e^{2\ell-1} \,|\partial^\ell u/\partial s^\ell|_{H^{1/2}(e)}^2 \lesssim h^{2\ell}|u|_{H^{\ell+1}(\O)}^2. \end{align*} \par Together these estimates imply \eqref{eq:InterpolationError}. \goodbreak \par\medskip\noindent{\em Estimate for $\|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u\|_h$} \par\smallskip We will show that the estimate \begin{equation}\label{eq:ProjectionError} \|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u\|_h\lesssim h^\ell|u|_{H^{\ell+1}(\O)} \end{equation} holds for both choices of $S^D(\cdot,\cdot)$. \par In the case where $S^D(\cdot,\cdot)=S^D_1(\cdot,\cdot)$, it follows from \eqref{eq:Sobolev}, Lemma~\ref{lem:PODErrors}, \eqref{eq:SD1} and \eqref{eq:ah} that \begin{align*}\label{eq:ProjectionEst2} \|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u\|_h^2 &\lesssim \sum_{D\in\cT_h} \|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u\|_{L_\infty(\partial D)}^2\\ &\lesssim \sum_{D\in\cT_h}\big[\hspace{1pt}h_{{\scriptscriptstyle D}}^{-2}\|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u\|_{L_2(D)}^2+|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u|_{{H^1(D)}}^2 +\hspace{1pt}h_{{\scriptscriptstyle D}}^2|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u|_{H^2(D)}^2\big]\notag\\ &\lesssim h^{2\ell}|u|_{H^{\ell+1}({\Omega})}^2.\notag \end{align*} \par In the case where $S^D(\cdot,\cdot)=S^D_2(\cdot,\cdot)$, we have \begin{align*} \|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u\|_h^2&=\sum_{D\in\cT_h}\hspace{1pt}h_{{\scriptscriptstyle D}} \|\partial(u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u)/\partial s\|_{L_2(D)}^2\\ &\lesssim \sum_{D\in\cT_h} \big[|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u|_{{H^1(D)}}^2+\hspace{1pt}h_{{\scriptscriptstyle D}}^2|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u|_{H^2(D)}^2\big] \lesssim h^{2\ell}|u|_{H^{\ell+1}({\Omega})}^2\notag \end{align*} by \eqref{eq:Trace} (applied to the first order derivatives of $u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u$) and Lemma~\ref{lem:PODErrors}. \par Putting \eqref{eq:EnergyError}--\eqref{eq:InterpolationError} and \eqref{eq:ProjectionError} together, we arrive at the following result. \begin{theorem}\label{thm:ConcreteEnergyError} Assuming the solution $u$ of \eqref{eq:Poisson} belongs to $H^{\ell+1}({\Omega})$ for $\ell$ between $1$ and $k$, we have \begin{equation*} \|u-u_h\|_h\leq C\sqrt{\alpha_h} h^\ell|u|_{H^{\ell+1}({\Omega})}, \end{equation*} where $\alpha_h$ is defined in \eqref{eq:kappah} and the positive constant $C$ only depends on $\rho$, $k$ and $N$. \end{theorem} \par We have similar estimates for the computable approximate solutions $\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h$ and $\Pi_{k,h}^0 u_h$. \begin{theorem}\label{thm:ComputableEnergyError} Assuming the solution $u$ of \eqref{eq:Poisson} belongs to $H^{\ell+1}({\Omega})$ for some $\ell$ between $1$ and $k$, we have \begin{equation}\label{eq:ComputableEnergyError} |u-u_h|_{H^1({\Omega})}+\sqrt{\alpha_h}\big[|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h|_{h,1} +|u-\Pi_{k,h}^0 u|_{h,1}\big] \leq C \alpha_h h^\ell |u|_{H^{\ell+1}({\Omega})}, \end{equation} where $\alpha_h$ is defined in \eqref{eq:kappah} and the positive constant $C$ only depends on $\rho$, $N$ and $k$. \end{theorem} \begin{proof} In view of \eqref{eq:ah} and Theorem~\ref{thm:ConcreteEnergyError}, we have \begin{equation*} |\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} (u-u_h)|_{h,1}\leq \|u-u_h\|_h\lesssim \sqrt{\alpha_h}h^{\ell}|u|_{H^{\ell+1}({\Omega})}, \end{equation*} which together with \eqref{eq:PODError} implies \begin{equation*} |u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h|_{h,1}\leq |u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u|_{h,1}+|\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}(u-u_h)|_{h,1} \lesssim \sqrt{\alpha_h}h^{\ell}|u|_{H^{\ell+1}({\Omega})}. \end{equation*} It follows from this estimate and \eqref{eq:PDZHOne} that \begin{equation*} |u-\Pi_{k,h}^0 u|_{h,1}\leq |u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u|_{h,1}+|\Pi_{k,h}^0(\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u-u)|_{h,1} \lesssim |u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u|_{h,1}\lesssim \sqrt{\alpha_h}h^{\ell}|u|_{H^{\ell+1}(\O)}. \end{equation*} \par Finally we have, by \eqref{eq:HOnePODID}, \eqref{eq:HOneVEM}, \eqref{eq:InterpolationError} and Theorem~\ref{thm:ConcreteEnergyError}, \begin{align*} |u-u_h|_{H^1({\Omega})}&\leq |u-I_{k,h} u|_{H^1({\Omega})}+|I_{k,h} u-u_h|_{H^1({\Omega})}\\ &\lesssim |u-I_{k,h} u|_{H^1({\Omega})}+\sqrt{\alpha_h}\|I_{k,h} u-u_h\|_h\\ &\leq |u-I_{k,h} u|_{H^1({\Omega})}+\sqrt{\alpha_h} \big(\|u-I_{k,h} u\|_h+\|u-u_h\|_h\big) \lesssim \alpha_h h^{\ell}|u|_{H^{\ell+1}(\O)}. \end{align*} \end{proof} \goodbreak \begin{remark}\label{rem:HOneErrors}{\rm In the case where $S^D(\cdot,\cdot)=S^D_1(\cdot,\cdot)$, the estimates for $|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h|_{h,1}$ and $|u-\Pi_{k,h}^0 u_h|_{h,1}$ are better than the estimate for $|u-u_h|_{H^1({\Omega})}$. } \end{remark} \subsection{Error Estimates in the $L_2$ Norm}\label{subsec:L2Error} We begin with two lemmas involving $S^D(\cdot,\cdot)$. \begin{lemma}\label{lem:SDEstimate} We have \begin{equation}\label{eq:SDEstimate} \sum_{D\in\cT_h} S^D(\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D} \zeta,\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D} \zeta)\lesssim h^2|\zeta|_{H^2({\Omega})}^2\qquad\forall\, \zeta\in H^2({\Omega})\cap H^1_0({\Omega}). \end{equation} \end{lemma} \begin{proof} It follows from \eqref{eq:Trace} (applied to the first order partial derivatives of $\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta$), Lemma~\ref{lem:InterpolationError} and \eqref{eq:SD2} that \begin{align*} &\sum_{D\in\cT_h} S^D_2(\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D} \zeta,\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D} \zeta) =\sum_{D\in\cT_h}\hspace{1pt}h_{{\scriptscriptstyle D}}\|\partial(\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D} \zeta)/\partial s\|_{L_2(\partial D)}^2\\ &\hspace{60pt}\lesssim \sum_{D\in\cT_h} \big[|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta|_{{H^1(D)}}^2+ \hspace{1pt}h_{{\scriptscriptstyle D}}^2|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta|_{H^2(D)}^2\big] \lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^2|\zeta|_{H^2({\Omega})}^2, \end{align*} and we have \begin{align*} &\sum_{D\in\cT_h} S^D_1(\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta,\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D} \zeta)\lesssim \sum_{D\in\cT_h}\|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D} \zeta\|_{L_\infty(\partial D)}^2\\ &\hspace{30pt}\lesssim \sum_{D\in\cT_h}\big[\hspace{1pt}h_{{\scriptscriptstyle D}}^{-2}\|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D} \zeta\|_{L_2(D)}^2 +|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D} \zeta|_{{H^1(D)}}^2+\hspace{1pt}h_{{\scriptscriptstyle D}}^2|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D} \zeta|_{H^2(D)}^2\big]\\ &\hspace{30pt}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^2|\zeta|_{H^2({\Omega})}^2 \end{align*} by \eqref{eq:Sobolev}, Lemma~\ref{lem:InterpolationError} and \eqref{eq:SD1}. \end{proof} \begin{lemma}\label{lem:POSD} Assuming that $u\in H^{\ell+1}({\Omega})$ for some $\ell$ between $1$ and $k$, we have \begin{equation}\label{eq:POSD} \sum_{D\in\cT_h} S^D(u_h-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h,u_h-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h)\lesssim \alpha_h^2 h^{2\ell}|u|_{H^{\ell+1}(\O)}^2. \end{equation} \end{lemma} \begin{proof} This is a consequence of \eqref{eq:ah}, \eqref{eq:ProjectionError}, Theorem~\ref{thm:ConcreteEnergyError} and Theorem~\ref{thm:ComputableEnergyError}: \begin{align*} &\sum_{D\in\cT_h} S^D(u_h-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h,u_h-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h)= \|u_h-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h\|_h^2\\ &\hspace{100pt}\lesssim \|u_h-u\|_h^2+\|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u\|_h^2 +\|\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}(u-u_h)\|_h^2\\ &\hspace{100pt}=\|u_h-u\|_h^2+\|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u\|_h^2+|u-u_h|_{H^1({\Omega})}^2 \lesssim \alpha_h^2 h^{2\ell}|u|_{H^{\ell+1}(\O)}^2. \end{align*} \end{proof} \par We can now prove a consistency estimate. \begin{lemma}\label{lem:ConsistencyError} Assuming that $u\in H^{\ell+1}({\Omega})$ for some $\ell$ between $1$ and $k$, we have \begin{equation*} a(u-u_h,I_{k,h}\zeta)\leq C{\alpha_h} h^{\ell+1}|u|_{H^{\ell+1}({\Omega})} |\zeta|_{H^2({\Omega})} \qquad \forall\, \zeta\in H^2({\Omega})\cap H^1_0({\Omega}), \end{equation*} where $\alpha_h$ is defined in \eqref{eq:kappah} and the positive constant $C$ only depends on $\rho$, $k$ and $N$. \end{lemma} \begin{proof} We have, by \eqref{eq:Poisson}, \eqref{eq:aDef}, \eqref{eq:POD1} and \eqref{eq:DiscreteProblem}--\eqref{eq:aD}, \begin{align* &a(u-u_h,I_{k,h}\zeta)=a(u,I_{k,h}\zeta)-\sum_{D\in\cT_h} a^D(u_h,I_{k,D}\zeta)\notag\\ &\hspace{40pt}=(f,I_{k,h}\zeta)-\sum_{D\in\cT_h} a^D(u_h,\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta) -\sum_{D\in\cT_h} a^D(u_h,I_{k,D}\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta)\\ &\hspace{40pt}=(f,I_{k,h}\zeta)-a_h(u_h,I_{k,D}\zeta) +\sum_{D\in\cT_h} S^D(u_h-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h,I_{k,D}\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta)\\ &\hspace{80pt}+\sum_{D\in\cT_h} a^D(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h-u_h,I_{k,D}\zeta -\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta)\\ &\hspace{40pt}=(f,I_{k,h}\zeta-\Xi_hI_{k,h}\zeta) +\sum_{D\in\cT_h} S^D(u_h-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h,I_{k,D}\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta)\\ &\hspace{80pt}+\sum_{D\in\cT_h} a^D(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h-u_h,I_{k,D}\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta), \end{align*} and the three terms on the right-hand side can be estimated as follows. \par We have \begin{equation*} |(f,I_{k,h}\zeta-\Xi_hI_{k,h}\zeta)|\lesssim h^{\ell+1}|u|_{H^{\ell+1}(\O)}|\zeta|_{H^2({\Omega})} \end{equation*} by \eqref{eq:Xih2}, \begin{equation*} \sum_{D\in\mathcal{T}_h} {S^D(u_h-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h,I_{k,D}\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta)} \lesssim {\alpha_h}h^{\ell+1}|u|_{H^{\ell+1}(\O)}|\zeta|_{H^2({\Omega})} \end{equation*} by Lemma~\ref{lem:SDEstimate} and Lemma~\ref{lem:POSD}, and \begin{align*} & \sum_{D\in\mathcal{T}_h} a^D(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u-u_h,I_{k,D}\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} I_{k,D}\zeta)\\ &\hspace{40pt}\leq \big[|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u-u|_{h,1}+|u-u_h|_{H^1({\Omega})}\big] \big[|I_{k,D}\zeta-\zeta|_{H^1({\Omega})}+|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta|_{h,1}\big]\\ &\hspace{40pt}\lesssim {\alpha_h} h^{\ell+1}|u|_{H^{\ell+1}(\O)}|\zeta|_{H^2({\Omega})} \end{align*} by Lemma~\ref{lem:InterpolationError}, \eqref{eq:PODError} and Theorem~\ref{thm:ComputableEnergyError}. \end{proof} \begin{theorem}\label{thm:uhLTwoError} Assuming $u\in H^{\ell+1}({\Omega})$ for some $\ell$ between $1$ and $k$, there exists a positive constant $C$, depending only on $\rho$, $N$ and $k$, such that \begin{equation}\label{eq:uhLTwoError} \|u-u_h\|_{L_2({\Omega})}\leq C{\alpha_h}h^{\ell+1}|u|_{H^{\ell+1}({\Omega})}, \end{equation} where $\alpha_h$ is defined in \eqref{eq:kappah}. \end{theorem} \begin{proof} Let $\zeta\in H^1_0({\Omega})$ be defined by \begin{equation*} a(v,\zeta)=(v,u-u_h) \qquad\forall\,v\in H^1_0({\Omega}). \end{equation*} We have % \begin{equation}\label{eq:Duality1} \|u-u_h\|_{L_2({\Omega})}^2=a(u-u_h,\zeta) = a(u-u_h,\zeta-I_{k,h}\zeta)+a(u-u_h,I_{k,h}\zeta), \end{equation} and since ${\Omega}$ is convex, \begin{equation}\label{eq:Duality2} \|\zeta\|_{H^2({\Omega})}\leq C_{\Omega} \|u-u_h\|_{L_2({\Omega})} \end{equation} by elliptic regularity \cite{Grisvard:1985:EPN,Dauge:1988:EBV}. \par The first term on the right-hand side of \eqref{eq:Duality1} satisfies \begin{equation}\label{eq:Duality3} a(u-u_h,\zeta-I_{k,h}\zeta)\leq |u-u_h|_{H^1({\Omega})} |\zeta-I_{k,h}\zeta|_{H^1({\Omega})}\\ \lesssim h|u-u_h|_{H^1({\Omega})}|\zeta|_{H^2({\Omega})} \end{equation} by \eqref{eq:HOnePODID}, and then \eqref{eq:uhLTwoError} follows from Theorem~\ref{thm:ComputableEnergyError}, Lemma~\ref{lem:ConsistencyError} and \eqref{eq:Duality1}--\eqref{eq:Duality3}. \end{proof} \par We have similar $L_2$ error estimates for the computable approximations $\Pi_{k,h}^0 u_h$ and $\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h$. \begin{theorem}\label{thm:ComputableLTwoErrors} Assuming $u\in H^{\ell+1}({\Omega})$ for some $\ell$ between $1$ and $k$, there exists a positive constant $C$, depending only on $\rho$, $N$ and $k$, such that \begin{equation*} \|u-\Pi_{k,h}^0 u_h\|_{L_2({\Omega})} +\|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h\|_{L_2({\Omega})}\leq C{\alpha_h}h^{\ell+1}|u|_{H^{\ell+1}({\Omega})}, \end{equation*} where $\alpha_h$ is defined in \eqref{eq:kappah}. \end{theorem} \begin{proof} The estimate for $\Pi_{k,h}^0 u_h$ follows from \eqref{eq:PDZLTwoError}, Theorem~\ref{thm:uhLTwoError} and the relation \begin{align*} \|u-\Pi_{k,h}^0 u_h\|_{L_2({\Omega})}&\leq \|u-\Pi_{k,h}^0 u\|_{L_2({\Omega})} +\|\Pi_{k,h}^0(u-u_h)\|_{L_2({\Omega})}\\ &\leq \|u-\Pi_{k,h}^0 u\|_{L_2({\Omega})}+\|u-u_h\|_{L_2({\Omega})}. \end{align*} \par For the approximation $\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h$, we have \begin{equation}\label{eq:LTwoPODError1} \|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h\|_{L_2({\Omega})}\leq\|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}I_{k,h} u\|_{L_2({\Omega})} +\|\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}(I_{k,h} u-u_h)\|_{L_2({\Omega})} \end{equation} and, in view of \eqref{eq:Trace}, \eqref{eq:tbarNormDef}, and Lemma~\ref{lem:PODStability2}, \begin{align}\label{eq:LTwoPODError2} \|\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}(I_{k,h} u-u_h)\|_{L_2({\Omega})}^2&\lesssim \sum_{D\in\mathcal{T}_h}|\!|\!| I_{k,D} u-u_h|\!|\!|_{k,D}^2\notag\\ &\lesssim \sum_{D\in\mathcal{T}_h}\big(\hspace{1pt}h_{{\scriptscriptstyle D}}\|I_{k,D} u-u_h\|_{L_2(\partial D)}^2 +\|\Pi_{k-2,D}^0 (I_{k,D} u-u_h)\|_{L_2(D)}^2\big)\\ &\lesssim\sum_{D\in\mathcal{T}_h} \big(\|I_{k,D} u-u_h\|_{L_2(D)}^2 +\hspace{1pt}h_{{\scriptscriptstyle D}}^2|I_{k,D} u-u_h|_{{H^1(D)}}^2\big)\notag\\ &\lesssim \|u-u_h\|_{L_2({\Omega})}^2+ \|u-I_{k,h} u\|_{L_2({\Omega})}^2 +h^2|u-I_{k,h} u|_{H^1({\Omega})}^2\notag\\ &\hspace{50pt}+h^2|u-u_h|_{H^1({\Omega})}^2.\notag \end{align} \par The estimate for $\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h$ follows from Lemma~\ref{lem:InterpolationError}, Theorem~\ref{thm:ComputableEnergyError}, Theorem~\ref{thm:uhLTwoError} and \eqref{eq:LTwoPODError1}--\eqref{eq:LTwoPODError2}. \end{proof} \subsection{Error Estimates for $u_h$ in the $L_\infty$ Norm}\label{subsec:uhLInftyError} Here we consider a $L_\infty$ error estimate for $u_h$ over the edges of $\mathcal{T}_h$, where $u_h$ is computable. We will treat the two choices of $S^D(\cdot,\cdot)$ separately. The set of all the edges in $\mathcal{T}_h$ will be denoted by $\mathcal{E}_h$. \subsubsection{The Case where $S^D(\cdot,\cdot)=S^D_2(\cdot,\cdot)$} \label{eq:subsubsec:uhLInftySD2} We have the following result for this choice of $S^D(\cdot,\cdot)$. \begin{theorem}\label{thm:MaxNormBdryEst2} Assuming that the solution $u$ of \eqref{eq:Poisson} belongs to $\in H^{\ell+1}({\Omega})$ for some $\ell$ between $1$ and $k$, we have \begin{equation*} \max_{e\in\mathcal{E}_h}\|u-u_h\|_{L_\infty(e)}\leq C h^{\ell}|u|_{H^{\ell+1}(\O)}, \end{equation*} where the positive constant $C$ only depends on $\rho$, $N$ and $k$. \end{theorem} \begin{proof} First we observe that, by \eqref{eq:SD2}, \eqref{eq:ah} and Theorem~\ref{thm:ConcreteEnergyError}, \begin{equation}\label{eq:EdgeEstimate} \sum_{D\in\cT_h}\sum_{e\in\cE_{\ssD}}\hspace{1pt}h_{{\scriptscriptstyle D}}\|\partial[(u-u_h)-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} (u-u_h)]/\partial s\|_{L_2(e)}^2 \lesssim \|u-u_h\|_h^2\lesssim h^{2\ell}|u|_{H^{\ell+1}(\O)}^2. \end{equation} \par We can connect any point in $e\in\mathcal{E}_h$ to $\partial{\Omega}$, where $u-u_h=0$, by a path along the edges in $\mathcal{E}_h$. Therefore it follows from a direct calculation (or a Sobolev inequality in one variable) that \begin{align}\label{eq:PathEstimate} \|u-u_h\|_{L_\infty(e)}^2&\lesssim \sum_{D\in\cT_h}\sum_{e\in\cE_{\ssD}} h_e\|\partial(u-u_h)/\partial s\|_{L_2(e)}^2\notag\\ &\lesssim \sum_{D\in\cT_h}\sum_{e\in\cE_{\ssD}}\hspace{1pt}h_{{\scriptscriptstyle D}}\|\partial[(u-u_h)-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} (u-u_h)]/\partial s\|_{L_2(e)}^2\\ &\hspace{40pt}+\sum_{D\in\cT_h}\sum_{e\in\cE_{\ssD}} \hspace{1pt}h_{{\scriptscriptstyle D}}\|\partial[\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-u_h)]/\partial s\|_{L_2(e)}^2\qquad\qquad\forall\,e\in\mathcal{E}_h,\notag \end{align} and we have, by \eqref{eq:Trace}, \eqref{eq:DiscreteEstimate1}, \eqref{eq:PODStability1} and Theorem~\ref{thm:ComputableEnergyError}, \begin{align}\label{eq:TrickyEstimate} &\sum_{D\in\cT_h}\sum_{e\in\cE_{\ssD}} \hspace{1pt}h_{{\scriptscriptstyle D}}\|\partial[\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-u_h)]/\partial s\|_{L_2(e)}^2\notag\\ &\hspace{80pt}\lesssim \sum_{D\in\cT_h} \big(|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-u_h)|_{H^1(D)}^2+\hspace{1pt}h_{{\scriptscriptstyle D}}^2|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-u_h)|_{H^2(D)}^2\big)\\ &\hspace{80pt}\lesssim\sum_{D\in\cT_h} |\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-u_h)|_{H^1(D)}^2\lesssim h^{2\ell}|u|_{H^{\ell+1}(\O)}^2. \notag \end{align} \par The estimates \eqref{eq:EdgeEstimate}--\eqref{eq:TrickyEstimate} together imply \begin{equation*} \|u-u_h\|_{L_\infty(e)}\leq h^\ell|u|_{H^{\ell+1}(\O)} \qquad\forall\,e\in\mathcal{E}_h. \end{equation*} \end{proof} \subsubsection{The Case where $S^D(\cdot,\cdot)=S^D_1(\cdot,\cdot)$} \label{eq:subsubsec:uhLInftySD1} We will establish an analog of Theorem~\ref{thm:MaxNormBdryEst2} under the additional assumption that $\mathcal{T}_h$ is quasi-uniform, i.e., there exists a positive constant $\gamma$ independent of $h$ such that \begin{equation}\label{eq:gamma} \hspace{1pt}h_{{\scriptscriptstyle D}}\geq \gamma h \qquad\forall\,\gamma\in\mathcal{T}_h. \end{equation} \begin{theorem}\label{thm:MaxNormBdryEst1} Assuming $\mathcal{T}_h$ is quasi-uniform and the solution $u$ of \eqref{eq:Poisson} belongs to $H^{\ell+1}({\Omega})$ for some $\ell$ between $1$ and $k$, we have \begin{equation*} \max_{e\in\mathcal{E}_h}\|u-u_h\|_{L_\infty(e)}\leq C\ln(1+\max_{D\in\mathcal{T}_h}\tau_{\!\ssD}) h^{\ell}|u|_{H^{\ell+1}(\O)}, \end{equation*} where the positive constant $C$ only depends on $\rho$, $N$, $\gamma$ and $k$. \end{theorem} \begin{proof} Let $D\in\mathcal{T}_h$ be arbitrary. First we observe that, by \eqref{eq:SD1}, Remark~\ref{rem:NormEquivalence}, \eqref{eq:ah}, \eqref{eq:InterpolationError}, Theorem~\ref{thm:ConcreteEnergyError} and Theorem~\ref{thm:ComputableEnergyError}, \begin{align}\label{eq:SD1Edge1} &\|(I_{k,D} u-u_h)-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(I_{k,D} u-u_h)\|_{L_\infty(\partial D)}\notag\\ &\hspace{80pt}\lesssim \|I_{k,h} u-u_h\|_h\\ &\hspace{80pt} \lesssim \|I_{k,h} u-u\|_h+\|u-u_h\|_h \lesssim [\ln(1+\max_{D\in\mathcal{T}_h}\tau_{\!\ssD})]^\frac12h^\ell|u|_{H^{\ell+1}(\O)}.\notag \end{align} \par Furthermore, it follows from \eqref{eq:G2}, \eqref{eq:DiscreteEstimate3}, \eqref{eq:PODStability1}, \eqref{eq:LTwoPODID}, Theorem~\ref{thm:ConcreteEnergyError}, Theorem~\ref{thm:ComputableLTwoErrors} and \eqref{eq:gamma} that \begin{align}\label{eq:SD1Edge2} &\|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(I_{k,D} u-u_h)\|_{L_\infty(\partial D)}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(I_{k,D} u-u_h)\|_{L_2(D)}+ |\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(I_{k,D} u-u_h)|_{H^1(D)}\notag\\ &\hspace{60pt}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\big(\|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D} u-u\|_{L_2(D)} +\|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h\|_{L_2(D)}\big)+|I_{k,D} u-u_h|_{H^1(D)}\\ &\hspace{60pt}\lesssim \ln(1+\max_{D\in\mathcal{T}_h}\tau_{\!\ssD}) h^\ell|u|_{H^{\ell+1}(\O)}.\notag \end{align} \par The theorem follows from \eqref{eq:1DInterpolationError}, \eqref{eq:SD1Edge1}, \eqref{eq:SD1Edge2} and the triangle inequality. \end{proof} \subsection{Error Estimates for $\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h$ and $\Pi_{k,h}^0 u_h$ in the $L_\infty$ Norm} \label{subsec:ComputableLInftyError} \par Again we treat the two choices of $S^D(\cdot,\cdot)$ separately. \subsubsection{The Case where $S^D(\cdot,\cdot)=S^D_2(\cdot,\cdot)$} \label{eq:subsubsec:ComputableLInftySD2} For this choice of $S^D(\cdot,\cdot)$, we can establish the following result without assuming that $\mathcal{T}_h$ is quasi-uniform. \begin{theorem}\label{thm:LInftyErrorsSD2} Assuming the solution $u$ of \eqref{eq:Poisson} belongs to $H^{\ell+1}({\Omega})$ for some $\ell$ between $1$ and $k$, there exists a positive constant $C$, depending only on $\rho$, $N$ and $k$, such that \begin{equation*} \|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h\|_{L_\infty({\Omega})}+\|u-\Pi_{k,h}^0 u_h\|_{L_\infty({\Omega})} \leq C h^{\ell}|u|_{H^{\ell+1}(\O)}. \end{equation*} \end{theorem} \begin{proof} For any $D\in\mathcal{T}_h$, we have, by \eqref{eq:Sobolev}, \eqref{eq:DiscreteEstimate1} and \eqref{eq:PODStability1}, \begin{align*} \|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h\|_{{L_\infty(D)}} &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h\|_{L_2(D)}+|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h|_{H^1(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h|_{H^2(D)}\\ &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h\|_{L_2(D)}+|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h|_{H^1(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u|_{H^2(D)}\\ &\hspace{50pt}+ \hspace{1pt}h_{{\scriptscriptstyle D}}|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} (u-u_h)|_{H^2(D)}\\ &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h\|_{L_2(D)}+|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h|_{H^1(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u|_{H^2(D)}\\ &\hspace{40pt}+|u-u_h|_{H^1(D)}, \end{align*} and \begin{align*} \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h\|_{L_2(D)}&\leq \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|u-u_h\|_{L_2(D)}+ \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|u_h-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h\|_{L_2(D)}\\ &\lesssim \big(\|u-u_h\|_{L_\infty(\partial D)}+|u-u_h|_{H^1(D)}\big) +|u_h-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h|_{H^1(D)} \end{align*} by \eqref{eq:G2}, \eqref{eq:PF2} and \eqref{eq:POD2}. These two estimates together with Lemma~\ref{lem:PODErrors}, Theorem~\ref{thm:ComputableEnergyError} and Theorem~\ref{thm:MaxNormBdryEst2} imply the estimate for $u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h$. \par The estimate for $u-\Pi_{k,h}^0 u_h$ can be derived similarly. We have, by \eqref{eq:Sobolev}, \eqref{eq:DiscreteEstimate1} and \eqref{eq:PDZHOne}, \begin{align*} \|u-\Pi_{k,D}^0\hspace{1pt} u_h\|_{{L_\infty(D)}} &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|u-\Pi_{k,D}^0\hspace{1pt} u_h\|_{L_2(D)}+|u-\Pi_{k,D}^0\hspace{1pt} u_h|_{H^1(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}|u-\Pi_{k,D}^0\hspace{1pt} u|_{H^2(D)}\\ &\hspace{40pt}+|u-u_h|_{H^1(D)}, \end{align*} and \begin{align*} \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|u-\Pi_{k,D}^0\hspace{1pt} u_h\|_{L_2(D)}&\leq \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|u-u_h\|_{L_2(D)}+ \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|u_h-\Pi_{k,D}^0\hspace{1pt} u_h\|_{L_2(D)}\\ &\lesssim \big(\|u-u_h\|_{L_\infty(\partial D)}+|u-u_h|_{H^1(D)}\big)+ |u_h-\Pi_{k,D}^0\hspace{1pt} u_h|_{H^1(D)} \end{align*} by \eqref{eq:G2}, \eqref{eq:PF1} and \eqref{eq:PF2}. The estimate for $u-\Pi_{k,h}^0 u_h$ now follows from Lemma~\ref{lem:PDZErrors}, Theorem~\ref{thm:ComputableEnergyError} and Theorem~\ref{thm:MaxNormBdryEst2}. \end{proof} \subsubsection{The Case where $S^D(\cdot,\cdot)=S^D_1(\cdot,\cdot)$} \label{eq:subsubsec:ComputableLInftySD1} The following analog of Theorem~\ref{thm:LInftyErrorsSD2} is proved by the same arguments but with Theorem~\ref{thm:MaxNormBdryEst2} replaced by Theorem~\ref{thm:MaxNormBdryEst1}. \begin{theorem}\label{thm:LInftyErrorsSD1} Assuming $\mathcal{T}_h$ is quasi-uniform and the solution $u$ of \eqref{eq:Poisson} belongs to $H^{\ell+1}({\Omega})$ for some $\ell$ between $1$ and $k$, there exists a positive constant $C$, depending only on $\rho$, $N$, $\gamma$ and $k$, such that \begin{equation*} \|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h\|_{L_\infty({\Omega})}+\|u-\Pi_{k,h}^0 u_h\|_{L_\infty({\Omega})} \leq C \ln(1+\max_{D\in\mathcal{T}_h}\tau_{\!\ssD}) h^{\ell}|u|_{H^{\ell+1}(\O)}. \end{equation*} \end{theorem} \section{Virtual Element Methods for the Poisson Problem in Three Dimensions}\label{sec:Poisson3D} The analysis of virtual element methods in three dimensions follows the same strategy as in two dimensions and many of the results in Section~\ref{sec:LocalVEM2D} and Section~\ref{sec:Poisson2D} carry over by identical arguments. We will only provide details for estimates that require different derivations. \par Let $\mathcal{T}_h$ be a polyhedral mesh on ${\Omega}$. The set of the faces of a subdomain $D\in\mathcal{T}_h$ is denoted by $\mathcal{F}_\ssD$ and the set of the edges of $F$ is denoted by $\cE_{\ssF}$. The set of all the faces of $\mathcal{T}_h$ is denoted by $\mathcal{F}_h$ and the set of all the edges of $\mathcal{T}_h$ is denoted by $\mathcal{E}_h$. \subsection{Shape Regularity Assumptions in Three Dimensions} \label{subsec:Shape3D} We impose the following shape regularity assumptions on $\mathcal{T}_h$, where $\hspace{1pt}h_{{\scriptscriptstyle D}}$ is the diameter of $D$. \par\smallskip\noindent {\em Assumption 1} \quad There exists $\rho\in (0,1)$, independent of $h$, such that every polyhedron $D\in\mathcal{T}_h$ is star-shaped with respect to a ball $\fB_{\!\ssD}$ with radius $\geq \rho \hspace{1pt}h_{{\scriptscriptstyle D}}$. \par\smallskip\noindent {\em Assumption 2} \quad There exists a positive integer $N$, independent of $h$, such that $|\mathcal{F}_\ssD|\leq N$ for all $D\in\mathcal{T}_h$. \par\smallskip\noindent {\em Assumption 3}\quad The shape regularity assumptions in Section~\ref{subsec:GlobalShapeRegularity} are satisfied by all the faces in $\mathcal{F}_h$, with the same $\rho$ from Assumption~1 and the same $N$ from Assumption~2. \par\smallskip All the hidden constants below will only depend on $\rho$, $N$ and $k$. \par Let $D$ be a polyhedron in $\mathcal{T}_h$. We can define the inner product $(\!(\cdot,\cdot)\!)$ by \eqref{eq:InnerProduct} where the infinitesimal arc-length $ds$ is replaced by the infinitesimal surface area $dS$. Then the projection operator $\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}:H^1(D)\longrightarrow\mathbb{P}_k(D)$ is defined by \eqref{eq:PODDef} and \begin{equation}\label{eq:3DPODStability1} |\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta|_{H^1(D)}\leq |\zeta|_{H^1(D)}\qquad\forall\,\zeta\in H^1(D). \end{equation} The projection from ${L_2(D)}$ to $\mathbb{P}_k(D)$ is again denoted by $\Pi_{k,D}^0\hspace{1pt}$. \par The results in Section~\ref{sec:SS} are valid for $D\in\mathcal{T}_h$ under Assumption 1. Consequently the results in Section~\ref{subsec:PODEstimates} and Section~\ref{subsec:PDZEstimates} are also valid provided the semi-norm $|\!|\!|\cdot|\!|\!|_{k,D}$ is defined by the following analog of \eqref{eq:tbarNormDef}: \begin{equation}\label{eq:tbarNorm3D} |\!|\!| \zeta|\!|\!|_{k,D}^2=\|\Pi_{k-2,D}^0 \zeta\|_{L_2(D)}^2 +\hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\mathcal{F}_\ssD}\|\Pi_{k-1,F}^0 \zeta\|_{L_2(F)}^2, \end{equation} where $\Pi_{k-1,F}^0$ is the projection from $L_2(F)$ onto $\mathbb{P}_{k-1}(F)$. \par We have the following estimates for $\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}$ and $\Pi_{k,D}^0\hspace{1pt}$: \begin{equation}\label{eq:3DLTwoPODPDZ} \|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} \zeta\|_{L_2(D)}+ \|\zeta-\Pi_{k,D}^0\hspace{1pt}\zeta\|_{L_2(D)} \lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{\ell+1}|\zeta|_{H^{\ell+1}(D)} \quad\forall\,\zeta\in H^{\ell+1}(D),\,0\leq\ell\leq k, \end{equation} and for $1\leq\ell \leq k$, \begin{alignat}{3} |\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}\zeta|_{{H^1(D)}}+ |\zeta-\Pi_{k,D}^0\hspace{1pt}\zeta|_{{H^1(D)}}&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{\ell}|\zeta|_{H^{\ell+1}(D)} &\quad&\forall\,\zeta\in H^{\ell+1}(D), \label{eq:3DHOnePODPDZ}\\ |\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} \zeta|_{H^2(D)}+|\zeta-\Pi_{k,D}^0\hspace{1pt}\zeta|_{H^2(D)}& \lesssim h_D^{\ell-1}|\zeta|_{H^{\ell+1}(D)} &\quad& \forall\,\zeta\in H^{\ell+1}(D).\label{eq:3DHTwoPODPDZ} \end{alignat} \par The analogs of Lemma~\ref{lem:PODStability2} and \eqref{eq:tbarNorm3D} lead to the estimate \begin{equation}\label{eq:3DPODStability2} \|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} \zeta\|_{L_2(D)}^2\lesssim |\!|\!|\zeta|\!|\!|_{k,D}^2 \lesssim \|\zeta\|_{L_2(D)}^2+\hspace{1pt}h_{{\scriptscriptstyle D}}\|\zeta\|_{L_2(\partial D)}^2\qquad \forall\,\zeta\in H^1(D), \end{equation} and we also have the following analog of \eqref{eq:PDZHOne}: \begin{equation}\label{eq:3DPDZHOne} |\Pi_{k,D}^0\hspace{1pt} \zeta|_{H^1(D)}\lesssim |\zeta|_{H^1(D)} \qquad\forall\,\zeta\in{H^1(D)}. \end{equation} \subsection{The Local Virtual Element Space $\bm\cQ^k(D)$} \label{subsec:Local3D} The space $\cQ^k(\p D)$ of continuous piecewise (two dimensional) virtual element functions of order $\leq k$ on $\partial D$ is defined by \begin{equation}\label{eq:cQbDDef} \cQ^k(\p D)=\{v\in C(\partial D):\,v\big|_F\in \mathcal{Q}^k(F) \quad\forall\,F\in \mathcal{F}_\ssD\}. \end{equation} \par For $k\geq 1$, the virtual element space $\cQ^k(D)\subset {H^1(D)}$ is defined by the following conditions: $v\in {H^1(D)}$ belongs to $\cQ^k(D)$ if and only if (i) the trace of $v$ on $\partial D$ belongs to $\cQ^k(\p D)$, (ii) the distribution $-\Delta v$ belongs to $\mathbb{P}_k(D)$, and (iii) condition \eqref{eq:Condition3} is satisfied. \begin{remark}\label{rem:3DContinuity}{\rm Since the restriction of $v\in\cQ^k(D)$ to $\partial D$ belongs to $C(\partial D)$ and $-\Delta v\in \mathbb{P}_k(D)$, the virtual element function $v$ is also continuous on $\bar D$ (cf. \cite[Section~1.2]{Kenig:1994:CBMS}).} \end{remark} \begin{remark}\label{rem:3Ddofs}{\rm The degrees of freedom of $\cQ^k(D)$ (cf. \cite{AABMR:2013:Projector}) consist of (i) the values of $v$ at the vertices of $D$ and nodes on the interior of each edge of $D$ that determine a polynomial of degree $k$ on each edge of $D$, (ii) the moments of $\Pi_{k-2,F}^0 v$ on each face $F$ of $D$, and (iii) the moments of $\Pi_{k-2,D}^0 v$ on $D$.} \end{remark} \begin{remark}\label{rem:3DComputable}{\rm For $v\in\cQ^k(D)$ and $F\in\mathcal{F}_\ssD$, the polynomial $\Pi_{k,F}^0 v$ can be computed in terms of the degrees of freedom of $v\big|_F$ (cf. Remark~\ref{rem:Computable}). Therefore the polynomial $\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v$ can be computed in terms of the degrees of freedom of $v\in\cQ^k(D)$ through \eqref{eq:POD1} and \eqref{eq:POD2}. The polynomial $\Pi_{k,D}^0\hspace{1pt} v$ can then be computed through \eqref{eq:Condition3}. } \end{remark} \begin{remark}\label{rem:DAndF}{\rm Under Assumption~3 in Section~\ref{subsec:Shape3D}, the results in Section~\ref{sec:LocalVEM2D} (with $D$ replaced by $F$) are valid for the restriction of $v\in \cQ^k(D)$ to any face $F$ of $D$. } \end{remark} \par The three dimensional analogs of Lemma~\ref{lem:MEP}, Lemma~\ref{lem:Fundamental} and Lemma~\ref{lem:InverseEstimate1} lead to the estimate \begin{equation}\label{eq:3DInverse} |v|_{H^1(D)}^2\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-2}|\!|\!|\zeta|\!|\!|_{k,D}^2 +\hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\mathcal{F}_\ssD}\|\nabla_{\!\!F} v\|_{L_2(F)}^2 \qquad\forall\,v\in\cQ^k(D), \end{equation} where $\nabla_{\!\!F}$ is the two dimensional gradient operator on the face $F$, and we also have an analog of \eqref{eq:KerPod13}: \begin{equation}\label{eq:3DKerPOD} \|\Pi_{k-2,D}^0 v\|_{L_2(D)}^2\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD} \|\Pi_{k-1,F}^0 v\|_{L_2(F)}^2 \qquad\forall\,v\in \mathcal{N}(\POD). \end{equation} Hence we have, by \eqref{eq:tbarNorm3D}, \eqref{eq:3DInverse} and \eqref{eq:3DKerPOD}, \begin{equation}\label{eq:3DInverse2} |v|_{H^1(D)}^2\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\sum_{F\in\mathcal{F}_\ssD}\|\Pi_{k-1,F}^0 v\|_{L_2(F)}^2\\ +\hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\mathcal{F}_\ssD}\|\nabla_{\!\!F} v\|_{L_2(F)}^2\qquad\forall\,v\in\mathcal{N}(\POD). \end{equation} \par The interpolation operator $I_{k,D}:H^2(D)\longrightarrow \cQ^k(D)$ is defined by the condition that $I_{k,D}\zeta$ and $\zeta$ share the same degrees of freedom. In particular we have \begin{equation}\label{eq:3DIDInvariance} I_{k,D} q=q\qquad\forall\,q\in \mathbb{P}_k(D). \end{equation} \par Note that \begin{equation}\label{eq:ConsistentID} I_{k,F} (\zeta\big|_F)=(I_{k,D}\zeta)\big|_F \qquad\forall\,F\in\mathcal{F}_\ssD, \end{equation} and hence, in view of \eqref{eq:G2}, Lemma~\ref{lem:MaximumPrinciple} and \eqref{eq:tbarNorm3D}, \begin{align}\label{eq:3DIDtbar} |\!|\!| I_{k,D}\zeta|\!|\!|_{k,D}^2&=\|\Pi_{k-2,D}^0 (I_{k,D}\zeta)\|_{L_2(D)}^2 +\hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD}\|\Pi_{k-1,F}^0 (I_{k,D}\zeta)\|_{L_2(F)}^2\notag\\ &=\|\Pi_{k-2,D}^0\zeta\|_{L_2(D)}^2+\hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD} \|\Pi_{k-1,F}^0(I_{k,F}\zeta)\|_{L_2(F)}^2\\ &\lesssim \|\zeta\|_{L_2(D)}^2+\hspace{1pt}h_{{\scriptscriptstyle D}}\SumFh_F^2\|I_{k,F}\zeta\|_{L_\infty(F)}^2\notag\\ &\lesssim \|\zeta\|_{L_2(D)}^2+\hspace{1pt}h_{{\scriptscriptstyle D}}\SumFh_F^2\big(\|I_{k,F}\zeta\|_{L_\infty(\partial F)}^2 +\|\nabla_{\!\!F}(I_{k,F}\zeta)\|_{L_2(F)}^2\big).\notag \end{align} \par The error estimates for $I_{k,D}$ rely on the following analog of \eqref{eq:IDHOne}, where $\tau_{\!\ssF}$ is defined by replacing $D$ by $F$ in \eqref{eq:TauD}. \begin{lemma}\label{lem:3DIDHOne} We have \begin{equation}\label{eq:3DIDHOne} |I_{k,D}\zeta|_{H^1(D)}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}\|\zeta\|_{L_2(D)} +|\zeta|_{H^1(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{H^2(D)} \end{equation} for all $\zeta\in{H^2(D)}$. \end{lemma} \begin{proof} Let $\zeta \in H^2(D)$ be arbitrary. It follows from \eqref{eq:3DInverse} and \eqref{eq:3DIDtbar} that \begin{align*} |I_{k,D}\zeta|_{H^1(D)}^2 &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-2}\|\zeta\|_{L_2(D)}^2+\hspace{1pt}h_{{\scriptscriptstyle D}} \sum_{F\in\FD} \|I_{k,F} \zeta\|_{L_\infty(\partial F)}^2 +\hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD} \|\nabla_{\!\!F}(I_{k,F}\zeta)\|_{L_2(F)}^2. \end{align*} We have, by \eqref{eq:Sobolev} and \eqref{eq:TrivialBdd}, \begin{align}\label{eq:IFLinfty} \hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD}\|I_{k,F}\zeta\|_{L_\infty(\partial F)}^2&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD} \|\zeta\|_{L_\infty(\partial F)}^2\\ &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\|\zeta\|_{L_\infty(D)}^2\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-2}\|\zeta\|_{L_2(D)}^2 +|\zeta|_{H^1(D)}^2+\hspace{1pt}h_{{\scriptscriptstyle D}}^2|\zeta|_{H^2(D)}^2,\notag \end{align} and by \eqref{eq:HalfTrace}, Lemma~\ref{lem:CZ2} and \eqref{eq:IDHOneHalf}, \begin{align}\label{eq:IFHone} \hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD} \|\nabla_{\!\!F}(I_{k,F}\zeta)\|_{L_2(F)}^2&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD}\big(|\zeta|_{H^1(F)}^2+h_F |\zeta|_{H^{3/2}(F)}^2\big)\\ &\lesssim \sum_{D\in\cT_h}\big( |\zeta|_{H^1(D)}^2+\hspace{1pt}h_{{\scriptscriptstyle D}}^{2}|\zeta|_{H^2(D)}^2\big). \notag \end{align} \end{proof} \par Note that \eqref{eq:3DIDtbar}, \eqref{eq:IFLinfty} and \eqref{eq:IFHone} imply \begin{equation}\label{eq:3DIDtbar2} \tbarI_{k,D}\zeta|\!|\!|_{k,D}\lesssim \|\zeta\|_{L_2(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{H^1(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}^2|\zeta|_{H^2(D)} \quad\forall\,\zeta\in{H^2(D)}, \end{equation} and hence we have, in view of \eqref{eq:3DLTwoPODPDZ}, \eqref{eq:3DPODStability2}, \eqref{eq:3DIDtbar} and \eqref{eq:3DIDHOne}, \begin{align}\label{eq:3DIDLTwo} \|I_{k,D}\zeta\|_{L_2(D)}&\leq \|I_{k,D}\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta\|_{L_2(D)}+\|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta\|_{L_2(D)}\notag\\ &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}|I_{k,D}\zeta|_{H^1(D)}+\tbarI_{k,D}\zeta|\!|\!|_{k,D}\\ &\lesssim \big(\|\zeta\|_{L_2(D)} +\hspace{1pt}h_{{\scriptscriptstyle D}}|\zeta|_{H^1(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}^2|\zeta|_{H^2(D)}\big)\notag \end{align} for all $\zeta\in H^2(D)$. \par In view of \eqref{eq:3DIDInvariance}, the following analogs of \eqref{eq:LTwoPODID}--\eqref{eq:HTwoPODID}, where $\zeta\in H^{\ell+1}(D)$ and $1\leq\ell\leq k$, can be obtained by combining the Bramble-Hilbert estimates \eqref{eq:BHEstimates} with the stability estimates \eqref{eq:3DPODStability1}, \eqref{eq:3DPODStability2}, \eqref{eq:3DIDHOne}, \eqref{eq:3DIDtbar2} and \eqref{eq:3DIDLTwo}. \begin{alignat}{3} \|\zeta-I_{k,D}\zeta\|_{L_2(D)}+\|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta\|_{L_2(D)} &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{\ell+1}|\zeta|_{H^{\ell+1}(D)} \label{eq:3DLTwoPODID}\\ |\zeta-I_{k,D}\zeta|_{{H^1(D)}}+|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta|_{{H^1(D)}} &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{\ell}|\zeta|_{H^{\ell+1}(D)} \label{eq:3DHOnePODID}\\ |\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta|_{H^2(D)}&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{\ell-1}|\zeta|_{H^{\ell+1}(D)} \label{eq:3DHTwoPODID} \end{alignat} \begin{remark}\label{rem:3DLInftyInterpolation}{\rm We also have the following analog of \eqref{eq:1DInterpolationError}: \begin{equation*} \|\zeta-I_{k,D}\zeta\|_{L_\infty(D)}\lesssim h^{\ell-\frac12}|u|_{H^{\ell+1}(D)} \end{equation*} for all $\zeta\in H^{\ell+1}(D)$ and $1\leq \ell\leq k$. The proof uses Lemma~\ref{lem:MaximumPrinciple} (which is valid in three dimensions) and the arguments for \eqref{eq:1DInterpolationError}. But we do not need this estimate in the error analysis. } \end{remark} \subsection{The Discrete Problem}\label{subsec:DP3D} Let the global virtual element space $\cQ^k_h$ be defined by $$\cQ^k_h=\{v\in H^1_0({\Omega}):\,v\big|_D\in\cQ^k(D)\quad\forall\,D\in\mathcal{T}_h\}.$$ The discrete problem for \eqref{eq:Poisson} is to find $u_h\in\cQ^k_h$ such that \begin{equation*} a_h(u_h,v)=(f,\Xi_h v) \qquad\forall\,v\in\cQ^k_h, \end{equation*} where $\Xi_h$ is defined as in \eqref{eq:Xih}, \begin{equation*} a_h(w,v)=\sum_{D\in\cT_h}\big[\big(\nabla(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} w),\nabla(\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v)\big)_{L_2(D)}+S^D(w-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} w,v-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v)\big], \end{equation*} and the local stabilizing bilinear form $S^D(\cdot,\cdot)$ is given by \begin{align} S^D(w,v)&=\hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD}\Big(h_F^{-2}(\Pi_{k-2,F}^0 w,\Pi_{k-2,F}^0 v)_{L_2(F)}+ \sum_{p\in\mathcal{N}_{\p F}}w(p)v(p)\Big).\label{eq:3DSD} \end{align} Here $\mathcal{N}_{\p F}$ is the set of the nodes along $\partial F$ associated with the degrees of freedom of a virtual element function. \begin{lemma}\label{lem:3DSDBdd} There exists a positive constant $C$, depending only on $\rho$, $N$ and $k$, such that \begin{alignat}{3} |v|_{H^1(D)}^2&\leq C \big [\ln(1+\max_{F\in\mathcal{F}_\ssD}\tau_{\!\ssF})\big] S^D(v,v) &\qquad&\forall\,v\in\mathcal{N}(\POD). \label{eq:3DSDBdd} \end{alignat} \end{lemma} \begin{proof} Let $v\in\mathcal{N}(\POD)$ be arbitrary. We have, by \eqref{eq:G2}, \eqref{eq:PF2}, Lemma~\ref{lem:MaximumPrinciple}, Corollary~\ref{cor:InverseEstimate3} and \eqref{eq:3DInverse2}, \begin{align*} |v|_{H^1(D)}^2&\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD} \big(\hspace{1pt}h_{{\scriptscriptstyle D}}^{-2}\|v\|_{L_2(F)}^2 +\|\nabla_{\!\!F} v\|_{L_2(F)}^2\big)\notag\\ &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD}\Big(\hspace{1pt}h_{{\scriptscriptstyle D}}^{-2}h_F^2\big(\|v\|_{L_\infty(\partial F)}^2 +\|\nabla_{\!\!F} v\|_{L_2(F)}^2\big)+\|\nabla_{\!\!F} v\|_{L_2(F)}^2\Big)\\ &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD}\big(\|v\|_{L_\infty(\partial F)}^2+\|\nabla_{\!\!F} v\|_{L_2(F)}^2\big)\\ &\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD}\big(h_F^{-2}\|\Pi_{k-2,F}^0 v\|_{L_2(F)}^2 +\ln(1+\tau_{\!\ssF})\|v\|_{L_\infty(\partial F)}^2\big), \end{align*} which together with Remark~\ref{rem:NormEquivalence} and \eqref{eq:3DSD} implies \eqref{eq:3DSDBdd}. \end{proof} \par It follows from \eqref{eq:3DSDBdd} that we have an analog of \eqref{eq:Stability}: \begin{equation}\label{eq:3DStability} |v|_{H^1({\Omega})}^2\leq 2\sum_{D\in\cT_h}\big[|\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} v|_{H^1({\Omega})}^2 +|v-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v|_{{H^1(D)}}^2\big]\lesssim \beta_h a_h(v,v) \qquad\forall\,v\in \cQ^k_h, \end{equation} where \begin{equation}\label{eq:lambdah} \beta_h= \ln(1+\max_{F\in\mathcal{F}_h}\tau_{\!\ssF}). \end{equation} Hence the discrete problem is well-posed. \begin{remark}\label{rem:3DInfo}{\rm The constants in the error estimates for the virtual element methods will only depend on $\rho$, $N$, $k$ and $\beta_h$. Therefore the existence of small faces in $\mathcal{T}_h$ does not affect the performance of the method. It is only the relative sizes of the edges on each face that matter. } \end{remark} \par Note that the estimates in Lemma~\ref{lem:Xih} are also valid for ${\Omega}\subset \mathbb{R}^3$. \subsection{Error Estimates in the Energy Norm}\label{subsec:3DEnergyError} The abstract error estimate \begin{equation}\label{eq:3DAbstractEnergyError} \|u-u_h\|_h\lesssim \|u-I_{k,h} u\|_h+\|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u\|_h +\sqrt{\beta_h}\Big(|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u|_{h,1}+ \sup_{w\in\cQ^k_h}\frac{(f,w-\Xi_hw)}{|w|_{H^1({\Omega})}}\Big) \end{equation} is obtained by the same arguments as in Section~\ref{subsec:AbstractEnergyError}, where $|\cdot|_{h,1}$ is defined in \eqref{eq:PiecewiseHOneNOrm}. \par We will derive concrete error estimates under the assumption that $u$ belongs to $H^{\ell+1}({\Omega})$ for $1\leq\ell\leq k$. Since the estimate \begin{equation}\label{eq:3DEasy} |u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u|_{h,1}+ \sup_{w\in\cQ^k_h}\frac{(f,w-\Xi_hw)}{|w|_{H^1({\Omega})}} \lesssim h^\ell|u|_{H^{\ell+1}(\O)} \end{equation} remains the same, we only need to estimate $\|u-I_{k,h} u\|_h$ and $\|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u\|_h$. \par It follows from \eqref{eq:G1}, \eqref{eq:3DPODStability1}, \eqref{eq:ConsistentID} and \eqref{eq:3DSD} that \begin{align}\label{eq:3DInterpolationError1} \|u-I_{k,h} u\|_h^2&\lesssim \sum_{D\in\cT_h}|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-I_{k,D} u)|_{H^1(D)}^2 +\sum_{D\in\cT_h}\hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD} h_F^{-2}\|u-I_{k,D} u\|_{L_2(F)}^2\notag\\ &\hspace{60pt}+\sum_{D\in\cT_h}\hspace{1pt}h_{{\scriptscriptstyle D}}\SumFh_F^{-2}\|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-I_{k,D} u)\|_{L_2(F)}^2\notag\\ &\hspace{90pt}+\sum_{D\in\cT_h} \hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD} \|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} (u-I_{k,D} u)\|_{L_\infty(\partial F)}^2\\ &\lesssim \sum_{D\in\cT_h} |u-I_{k,D} u|_{H^1(D)}^2 +\sum_{D\in\cT_h}\hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD} h_F^{-2}\|u-I_{k,F} u\|_{L_2(F)}^2\notag\\ &\hspace{60pt}+\sum_{D\in\cT_h} \hspace{1pt}h_{{\scriptscriptstyle D}}\|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} (u-I_{k,D} u)\|_{L_\infty(D)}^2,\notag \end{align} and we have, by \eqref{eq:HalfTrace} and \eqref{eq:HalfIDLTwoError}, \begin{equation}\label{eq:3DInterpolationError2} \sum_{D\in\cT_h}\hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD} h_F^{-2}\|u-I_{k,F} u\|_{L_2(F)}^2 \lesssim \sum_{D\in\cT_h} \hspace{1pt}h_{{\scriptscriptstyle D}} \sum_{F\in\FD} h_F^{2\ell-1}|u|_{H^{\ell+(1/2)}(F)}^2 \lesssim h^{2\ell}|u|_{H^{\ell+1}(\O)}^2. \end{equation} Moreover the estimates \eqref{eq:G1}, \eqref{eq:DiscreteEstimate3} and \eqref{eq:3DPODStability1} imply \begin{align}\label{eq:3DInterpolationError3} &\sum_{D\in\cT_h} \hspace{1pt}h_{{\scriptscriptstyle D}}\|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} (u-I_{k,D} u)\|_{L_\infty(D)}^2 \notag\\ &\hspace{60pt} \lesssim \sum_{D\in\cT_h} \hspace{1pt}h_{{\scriptscriptstyle D}}^{-2}\|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} (u-I_{k,D} u)\|_{L_2(D)}^2 +\sum_{D\in\cT_h}|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} (u-I_{k,D} u)|_{H^1(D)}^2\\ &\hspace{60pt}\lesssim \sum_{D\in\cT_h} \hspace{1pt}h_{{\scriptscriptstyle D}}^{-2}\big(\|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u-u)\|_{L_2(D)}^2 +\|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D} u)\|_{L_2(D)}^2\big)\notag\\ &\hspace{100pt}\notag+\sum_{D\in\cT_h}|u-I_{k,D} u|_{H^1(D)}^2.\notag \end{align} \par Combining \eqref{eq:3DLTwoPODPDZ}, \eqref{eq:3DLTwoPODID}, \eqref{eq:3DHOnePODID} and \eqref{eq:3DInterpolationError1}--\eqref{eq:3DInterpolationError3}, we obtain \begin{equation}\label{eq:3DInterpolationError} \|u-I_{k,h} u\|_h^2\lesssim h^{2\ell}|u|_{H^{\ell+1}(\O)}^2, \end{equation} which is the analog of \eqref{eq:InterpolationError}. \par From \eqref{eq:Sobolev}, \eqref{eq:G1}, \eqref{eq:3DLTwoPODPDZ}--\eqref{eq:3DHTwoPODPDZ} and \eqref{eq:3DSD}, we find \begin{align}\label{eq:3DProjectionError} \|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u\|_h^2&\lesssim \sum_{D\in\cT_h}\hspace{1pt}h_{{\scriptscriptstyle D}}\SumFh_F^{-2}\|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u\|_{L_2(F)}^2 +\sum_{D\in\cT_h} \hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD} \|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u\|_{L_\infty(\partial F)}^2\notag\\ &\lesssim \sum_{D\in\cT_h} \hspace{1pt}h_{{\scriptscriptstyle D}}\|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u\|_{L_\infty(D)}^2\\ &\lesssim \sum_{D\in\cT_h}\big( \hspace{1pt}h_{{\scriptscriptstyle D}}^{-2}\|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u\|_{L_2(D)}^2 +|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u|_{H^1(D)}^2+\hspace{1pt}h_{{\scriptscriptstyle D}}^2|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u|_{H^2(D)}^2\big)\notag\\ &\lesssim h^{2\ell}|u|_{H^{\ell+1}(\O)}^2,\notag \end{align} which is the analog of \eqref{eq:ProjectionError}. \par The estimates \eqref{eq:3DAbstractEnergyError}, \eqref{eq:3DEasy}, \eqref{eq:3DInterpolationError} and \eqref{eq:3DProjectionError} lead to the following analog of Theorem~\ref{thm:ConcreteEnergyError}. \begin{theorem}\label{thm:3DConcreteEnergyError} Assuming the solution $u$ of \eqref{eq:Poisson} belongs to $H^{\ell+1}({\Omega})$ for $\ell$ between $1$ and $k$, we have \begin{equation}\label{eq:3DConcreteEnergyError} \|u-u_h\|_h\leq C \sqrt{\beta_h}h^\ell|u|_{H^{\ell+1}(\O)}. \end{equation} where $\beta_h$ is defined in \eqref{eq:lambdah} and the positive constant $C$ depends only on $\rho$, $N$ and $k$. \end{theorem} \par The following analog of Theorem~\ref{thm:ComputableEnergyError} on the computable approximate solutions $\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h$ and $\Pi_{k,h}^0 u_h$ is obtained by the same arguments. \begin{theorem}\label{thm:3DComputableEnergyError} Assuming the solution $u$ of \eqref{eq:Poisson} belongs to $H^{\ell+1}({\Omega})$ for $\ell$ between $1$ and $k$, there exists a positive constant $C$, depending only on $\rho$, $N$ and $k$, such that \begin{equation}\label{eq:3DComputableEnergyError} |u-u_h|_{H^1({\Omega})}+\sqrt{\beta_h}\big[|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h|_{h,1} +|u-\Pi_{k,h}^0 u|_{h,1}\big]\leq C \beta_h h^\ell |u|_{H^{\ell+1}({\Omega})}, \end{equation} where $\beta_h$ is defined in \eqref{eq:lambdah}. \end{theorem} \subsection{Error Estimates in the $L_2$ Norm}\label{subsec:3DL2Error} We begin with an analog of Lemma~\ref{lem:SDEstimate}. \begin{lemma}\label{lem:3DSDEstimate} We have \begin{equation* \sum_{D\in\cT_h} S^D(\zeta-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}I_{k,h} \zeta,\zeta-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}I_{k,h} \zeta)\lesssim h^2|\zeta|_{H^2({\Omega})}^2\qquad\forall \zeta\in H^2({\Omega})\cap H^1_0({\Omega}). \end{equation*} \end{lemma} \begin{proof} It follows from \eqref{eq:Sobolev}, \eqref{eq:3DLTwoPODID}--\eqref{eq:3DHTwoPODID} and \eqref{eq:3DSD} that \begin{align*} &\sum_{D\in\cT_h} S^D(\zeta-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} I_{k,h}\zeta,\zeta-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} I_{k,h}\zeta)\\ &\hspace{30pt}\lesssim \sum_{D\in\cT_h}\hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD}\big(h_F^{-2}\|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta\|_{L_2(F)}^2 +\|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta\|_{L_\infty(\partial F)}^2\big)\\ &\hspace{30pt}\lesssim \sum_{D\in\cT_h}\hspace{1pt}h_{{\scriptscriptstyle D}}\|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta\|_{L_\infty(\partial D)}^2\\ &\hspace{30pt}\lesssim \sum_{D\in\cT_h} \big(\hspace{1pt}h_{{\scriptscriptstyle D}}^{-2}\|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta\|_{L_2(D)}^2+ |\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta|_{H^1(D)}^2+\hspace{1pt}h_{{\scriptscriptstyle D}}^2|\zeta-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D}\zeta|_{H^2(D)}^2\big)\\ &\hspace{30pt}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^2|\zeta|_{H^2({\Omega})}^2. \end{align*} \end{proof} \goodbreak \par The same arguments as in the proof of Lemma~\ref{lem:POSD} lead to the following result. \begin{lemma}\label{lem:3DPOSD} We have \begin{equation*} \sum_{D\in\cT_h} S^D(u_h-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h,u_h-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h)\lesssim \beta_h^2 h^{2\ell}|u|_{H^{\ell+1}(\O)}^2. \end{equation*} \end{lemma} \par With Lemma~\ref{lem:3DSDEstimate} and Lemma~\ref{lem:3DPOSD} in hand, we obtain the following analog of Theorem~\ref{thm:uhLTwoError} and Theorem~\ref{thm:ComputableLTwoErrors} by identical arguments. \begin{theorem}\label{thm:3DLTwoErrors} Assuming $u\in H^{\ell+1}({\Omega})$ for some $\ell$ between $1$ and $k$, there exists a positive constant $C$, depending only on $N$, $k$ and $\rho$, such that \begin{align*} \|u-u_h\|_{L_2({\Omega})}+\|u-\Pi_{k,h}^0 u_h\|_{L_2({\Omega})}+ \|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h\|_{L_2({\Omega})}\leq C\beta_h h^{\ell+1}|u|_{H^{\ell+1}({\Omega})}, \end{align*} where $\beta_h$ is defined in \eqref{eq:lambdah}. \end{theorem} \subsection{Error Estimate in the $L_\infty$ Norm}\label{subsec:3DInftyError} We will derive $L_\infty$ error estimates under the additional assumption that $\mathcal{T}_h$ is quasi-uniform (cf. \eqref{eq:gamma}). We begin with an analog of Theorem~\ref{thm:MaxNormBdryEst1}. \begin{theorem}\label{thm:uhInftyBdryEst} Assuming $\mathcal{T}_h$ is quasi-uniform and the solution $u$ of \eqref{eq:Poisson} belongs to $H^{\ell+1}({\Omega})$ for $\ell$ between $1$ and $k$, we have \begin{equation}\label{eq:InftyBdryEst} \max_{e\in\mathcal{E}_h}\|u-u_h\|_{L_\infty(e)}\leq C\beta_h h^{\ell}|u|_{H^{\ell+1}(\O)}, \end{equation} where $\beta_h$ is defined in \eqref{eq:lambdah} and the positive constant $C$ only depends on $\rho$, $N$, $\gamma$ and $k$. \end{theorem} \begin{proof} It follows from Remark~\ref{rem:NormEquivalence} that \begin{equation*} \sum_{D\in\cT_h}\hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD}\|(I_{k,D} u-u_h)-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} (I_{k,D} u-u_h)\|_{L_\infty(\partial F)}^2 \lesssim \|I_{k,h} u-u_h\|_h^2 \end{equation*} and hence, for any $D\in\mathcal{T}_h$ and $F\in\mathcal{F}_\ssD$, \begin{equation}\label{eq:3DBdryEst1} \|(I_{k,D} u-u_h)-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} (I_{k,D} u-u_h)\|_{L_\infty(\partial F)}^2\lesssim \beta_h\, h^{2\ell-1}|u|_{H^{\ell+1}(\O)}^2 \end{equation} by \eqref{eq:3DInterpolationError}, Theorem~\ref{thm:3DConcreteEnergyError} and the quasi-uniformity of $\mathcal{T}_h$, \par For any $D\in\mathcal{T}_h$ and $F\in\mathcal{F}_\ssD$, we have, by \eqref{eq:G1}, \eqref{eq:DiscreteEstimate3}, \eqref{eq:3DPODStability1}, \eqref{eq:3DLTwoPODID}, \eqref{eq:3DHOnePODID}, Theorem~\ref{thm:3DLTwoErrors} and the quasi-uniformity of $\mathcal{T}_h$, \begin{align}\label{eq:3DBdryEst2} &\|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(I_{k,D} u-u_h)\|_{L_\infty(F)}^2\notag\\ &\hspace{40pt}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-3}\|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(I_{k,D} u-u_h)\|_{L_2(D)}^2+ \hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(I_{k,D} u-u_h)|_{H^1(D)}^2\notag\\ &\hspace{40pt}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-3}\big(\|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}I_{k,D} u-u\|_{L_2(D)}^2 +\|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h\|_{L_2(D)}^2\big) +\hspace{1pt}h_{{\scriptscriptstyle D}}^{-1}|I_{k,D} u-u_h|_{H^1(D)}^2\\ &\hspace{40pt}\lesssim \beta_h^2 h^{2\ell-1}|u|_{H^{\ell+1}(\O)}^2.\notag \end{align} \par Finally we have, for any $D\in\mathcal{T}_h$ and $F\in\mathcal{F}_\ssD$, \begin{equation}\label{eq:3DBdryEst3} \|u-I_{k,D} u\|_{L_\infty(F)}^2=\|u-I_{k,F} u\|_{L_\infty(F)}^2\lesssim h_F^{2\ell-1}|u|_{H^{\ell+\frac12}(F)}^2 \lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{2\ell-1}|u|_{H^{\ell+1}(D)}^2 \end{equation} by \eqref{eq:HalfTrace}, \eqref{eq:1DInterpolationErrorHalf} and \eqref{eq:ConsistentID}. \par % The estimate \eqref{eq:InftyBdryEst} then follows from \eqref{eq:3DBdryEst1}--\eqref{eq:3DBdryEst3} and the triangle inequality. \end{proof} \par We also have estimates for the computable approximate solutions $\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h$ and $\Pi_{k,h}^0 u_h$. \begin{theorem}\label{thm:3DMaxNormBdryEst} Assuming $\mathcal{T}_h$ is quasi-uniform and the solution $u$ of \eqref{eq:Poisson} belongs to $H^{\ell+1}({\Omega})$ for $\ell$ between $1$ and $k$, there exists a positive constant $C$, depending only on $\rho$, $N$, $\gamma$ and $k$, such that \begin{equation}\label{eq:3DLInftyError} \|u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h\|_{L_\infty({\Omega})}+\|u-\Pi_{k,h}^0 u_h\|_{L_\infty({\Omega})}\leq C \beta_h h^{\ell-(1/2)}|u|_{H^{\ell+1}(\O)}, \end{equation} where $\beta_h$ is defined in \eqref{eq:lambdah}. \end{theorem} \begin{proof} For any $D\in\mathcal{T}_h$, we have, by \eqref{eq:Sobolev}, \eqref{eq:DiscreteEstimate1} and \eqref{eq:3DPODStability1}, \begin{align*} &\|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h\|_{{L_\infty(D)}}\\ &\hspace{40pt}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-\frac32}\|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h\|_{L_2(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}^{-\frac12}|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h|_{H^1(D)} +\hspace{1pt}h_{{\scriptscriptstyle D}}^\frac12|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h|_{H^2(D)}\\ &\hspace{40pt}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-\frac32}\|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h\|_{L_2(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}^{-\frac12}|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h|_{H^1(D)} +\hspace{1pt}h_{{\scriptscriptstyle D}}^\frac12|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u|_{H^2(D)}\\ &\hspace{70pt}+\hspace{1pt}h_{{\scriptscriptstyle D}}^\frac12|\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt}(u-u_h)|_{H^2(D)}\\ &\hspace{40pt}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-\frac32}\|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h\|_{L_2(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}^{-\frac12}|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u_h|_{H^1(D)} +\hspace{1pt}h_{{\scriptscriptstyle D}}^\frac12|u-\Pi_{k,D}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} u|_{H^2(D)}\\ &\hspace{60pt}+\hspace{1pt}h_{{\scriptscriptstyle D}}^{-\frac12}|u-u_h|_{H^1(D)}, \end{align*} which together with \eqref{eq:3DHTwoPODPDZ}, Theorem~\ref{thm:3DComputableEnergyError}, Theorem~\ref{thm:3DLTwoErrors} and the quasi-uniformity of $\mathcal{T}_h$ implies the estimate for $u-\Pi_{k,h}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}} u_h$. \par Similarly we have, by \eqref{eq:Sobolev}, \eqref{eq:DiscreteEstimate1} and \eqref{eq:3DPDZHOne}, \begin{align*} &\|u-\Pi_{k,D}^0\hspace{1pt} u_h\|_{{L_\infty(D)}}\\ &\hspace{40pt}\lesssim \hspace{1pt}h_{{\scriptscriptstyle D}}^{-\frac32}\|u-\Pi_{k,D}^0\hspace{1pt} u_h\|_{L_2(D)}+\hspace{1pt}h_{{\scriptscriptstyle D}}^{-\frac12}|u-\Pi_{k,D}^0\hspace{1pt} u_h|_{H^1(D)} +\hspace{1pt}h_{{\scriptscriptstyle D}}^\frac12|u-\Pi_{k,D}^0\hspace{1pt} u|_{H^2(D)}\\ &\hspace{70pt}+\hspace{1pt}h_{{\scriptscriptstyle D}}^{-\frac12}|u-u_h|_{H^1(D)}, \end{align*} which together with \eqref{eq:3DHTwoPODPDZ}, Theorem~\ref{thm:3DComputableEnergyError}, Theorem~\ref{thm:3DLTwoErrors} and the quasi-uniformity of $\mathcal{T}_h$ implies the estimate for $u-\Pi_{k,h}^0 u_h$. \end{proof} \section{Concluding Remarks}\label{sec:Conclusions} We have developed error estimates for virtual element methods for the model Poisson problem in two and three dimensions that provide justifications for existing numerical results for polygonal (or polyhedral) meshes with small edges (or faces). \par For the two dimensional problem, the convergence of the virtual element method based on the stabilizing bilinear form $S^D_2(\cdot,\cdot)$ is optimal under the shape regularity assumptions in Section~\ref{subsec:GlobalShapeRegularity}. Under the additional assumption that the edges of any subdomain in a polygonal mesh are comparable to one another, convergence of the virtual element method based on the stabilizing bilinear form $S^D_1(\cdot,\cdot)$ is also optimal. \par For the three dimensional problem, the convergence of the virtual element method is optimal if, in addition to the assumptions in Section~\ref{subsec:Shape3D}, we also assume that the edges of any face in the polyhedral mesh are comparable to one another. \par The results in this paper can be extended to virtual element methods with the stabilizing bilinear form \begin{equation*} S^D(w,v)=(\Pi_{k-2,D}^0 w,\Pi_{k-2,D}^0 v)+\sum_{p\in\mathcal{N}_{\p D}}w(p)v(p) \end{equation*} in two dimensions, and the stabilizing bilinear form \begin{align*} S^D(w,v)&=(\Pi_{k-2,D}^0 w,\Pi_{k-2,D}^0 v)+\hspace{1pt}h_{{\scriptscriptstyle D}}\sum_{F\in\FD}\Big(h_F^{-2}(\Pi_{k-2,F}^0 w,\Pi_{k-2,F}^0 v)_{L_2(F)}+ \sum_{p\in\mathcal{N}_{\p F}}w(p)v(p)\Big).\label{eq:3DSD} \end{align*} in three dimensions. The stability for these virtual element methods is automatic and the error analysis also does not pose any new difficulties. \par The results in this paper can also be extended to virtual element methods ($k\geq2$) where the inner product \eqref{eq:InnerProduct} is replaced by the inner product \begin{equation*} (\!(\zeta,\eta)\!)=(\nabla \zeta,\nabla \eta) +\Big(\int_{D}\zeta\,dx\Big)\Big(\int_{D} \eta\,dx\Big). \end{equation*} \par We note that error estimates for the Poisson problem on general polygonal or polyhedral domains can also be obtained by the techniques developed in this paper. \par Finally it would be interesting to construct a three dimensional analog of the stabilizing bilinear form $S^D_2(\cdot,\cdot)$ defined in Section~\ref{subsec:DiscreteProblem} so that the convergence of the virtual element methods is optimal for polyhedral meshes with arbitrarily small faces and edges, and $L_\infty$ error estimates can be established without assuming the meshes are quasi-uniform. We conjecture that such a bilinear form can be defined by \begin{align*} S^D(v,w)&=\SumFh_F(\nabla_{\!\!F}\Pi_{k,F}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v,\nabla_{\!\!F}\Pi_{k,F}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} w)_{L_2(F)}\\ &\hspace{40pt}+\SumFh_F\sum_{e\in\cE_{\ssF}} h_e\big(\partial (v-\Pi_{k,F}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} v)/\partial s,\partial (w-\Pi_{k,F}^{\raise 1pt\hbox{$\scriptscriptstyle\nabla$}}\hspace{1pt} w)/\partial s\big)_{L_2(e)}. \end{align*}

https://arxiv.org/abs/2112.10171

Newtonian mechanics in a Riemannian manifold

The work done by Isaac Newton more than three hundred years ago, continues being a path to increase our knowledge of Nature. To better understand all the ideas behind it, one of the finest ways is to generalize them to wider situations. In this report we make a review of one of these enlargements, the one that bears the mechanical systems from the elementary homogeneous three dimensional Euclidean space to the more abstract geometry of a Riemannian manifold.

\section{Introduction} Mechanics is an ancient wisdom of humankind. It is not possible to build the Egyptian Pyramids, for example, without some more or less organised intuitive ideas of mechanics. And they were made near five thousand years ago. Archimedes, 287--212 BC, was one of those that best exploited mechanical ideas as an interesting instrument not only to conceive several devices but to proof geometric results. As a science, mechanics began to emerge in modern times, in XVI and XVII centuries, when Galileo, Kepler and Newton tried to explain the motion of the planets they knew in the solar system using the old intuitive mechanical and geometric ideas and the carefully collected observations of the positions of the planets annotated along the time. Isaac Newton, 1643-1727, was the real founder of what we know as classical mechanics or analytical mechanics, as Lagrange, 1736-1813, called it. The Newton's {\sl Principia}, \cite{Newton}, have been developed along more than three hundred years and we continue increasing their understanding and applications. Near every scientific generation has a novel approach to Newtonian mechanics and has increased the range of applications. In fact, analytical mechanics has arrived to be a basic tool to found the description of our Universe in a systematic form and it can be said that all our modern knowledge of Nature has been developed following ideas coming at the end from Newtonian mechanics. Geometry has always been related to mechanics. They have influenced each other in a very enhancing way along the time, using both disciplines as a source of problems and methods to state and solve questions in the other. And more and more, new geometric techniques have been introduced in mechanics to obtain new results, or for better understanding older ones, and applications. In particular, differential geometry is a powerful tool to clarify the deep dependencies between different expressions of the same problem, or solution, and the relations among magnitudes, due to the intrinsic formulation of the theory, that is the independence of the used coordinate system to express the equation under consideration. Indeed, it has proven to be a very adequate way to express the mechanical concepts in a short and more comprehensible form. As far as we know, the oldest reference to the name ``Mechanics in a Riemannian manifold" is used by R. Hermann in \cite{HE1968}, where one chapter has precisely this title. But clearly there are previous authors using Riemannian techniques to tackle mechanical problems. We necessarily need to cite Eisenhart, \cite{EISEN-1928} and Synge, \cite{SYN-1926, SYN-1928}, most of these works inspired by Levi--Civita, the real pionneer. Other modern references are \cite{AM-78, Ar-89,BLOCH2015, BULE2005,CC-2005,Ol-02} where this name appears together with geometric mechanics. We will use all of them without specific citation. It is interesting to note the special effort made by some authors, specially A.D. Lewis in \cite{LEWIS, Lewis2018}, in order to justify the way going from classical to geometric mechanics and the usefulness of this last approach. At this point, it is important to say that this survey is by no means an historical review of classical or geometric mechanics. A very nice survey on historical aspects of geometrical mechanics with an extensive bibliography is \cite{Leon-2017}. Some other historical remarks are contained in \cite{Ar-89} in the introduction of the different chapters. The aim of this review is to develop some aspects of Newtonian mechanics in a Riemannian setting, hence without reference to the usual vector, or affine, space structure for the configuration manifold. With respect to present developments and applications we will give only some of them with names of people involved and general references in the corresponding sections, references which contain a large amount of more specific references. With this ideas in mind, the organization of the paper is as follows: \begin{description} \item[Section 2]: Notations, definitions and dynamical equation of the trajectories including the Lagrangian formulation for conservative and more general systems. \item[Sections 3 and 4]: We study constrained systems. First with holonomic constraints and secondly with nonholonomic ones. The corresponding d'Alembert principles and the subsequent dynamical equations are stated, both in the Riemannian form and in the Euler-Lagrange setting. \item[Section 5]: Is a short section for non autonomous systems. \item[Section 6]: Some classical subjects in Newtonian mechanics are included: Hamilton-Jacobi vector fields and geodesic fields and Hamilton-Jacobi equation in a Lagrangian setting. We finish the section with an approximation to stationary Euler equation for fluids as a Hamilton--Jacobi equation for a Newtonian system and comments and references on the relation between solutions to the Hamilton--Jacobi equation and to the associated Schr\"odinger equation for this kind of systems. \item[Section 7]: Comments and references on other topics in this approach and applications: symmetries and Noether's theorem and control of mechanical systems. \item[Section 8]: Conclusions. \end{description} Observe that Sections 2, 3 and 4 are the general theory while sections 6 and 7 are devoted to some specific developments and applications where the Riemannian approach is significantly enhancing. As general references on differential and Riemannian geometry we recommend \cite{Con2001,Lee2013,Lee2018}. With respect to analytical mechanics and geometric mechanics, see \cite{AM-78, Ar-89,Arovas2014, Ga-70,GPS-01,JS-98,LL-76, LM-sgam,MR-99,Ol-02, SC-71, Sch-2005, So-ssd}. A recent reference on the deep relations between geometry and physics is \cite{CIMM-2015}, not only with classical mechanics but including quantum physics. As is usual in this approach we consider that our manifolds and mappings are of $\mathcal{C}^\infty$--class. Einstein index summation convention is also assumed. \section{Newtonian dynamical systems} \subsection{Definitions and dynamical equation} From a mathematical point of view a \textbf{Newtonian mechanical system} is a triple $(Q,{\bf g},\omega)$, where \begin{enumerate} \item $Q$ is a differentiable manifold ($\dim\, Q=n$). \item ${\bf g}$ is a {\sl Riemannian metric} in $Q$. Then $(Q,{\bf g})$ is a Riemannian manifold. \item $\omega$ is a differential 1-form in $Q$, called the \textbf{work form}. \end{enumerate} Being ${\bf g}$ a Riemannian metric, the work form $\omega\in{\mit\Omega}^1(Q)$ is associated to a unique vector field ${\rm F}\in\mathfrak{X} (Q)$ such that $\mathop{i}\nolimits({\rm F}) {\bf g}=\omega$. We call ${\rm F}$ the \textbf{the field of forces} of the system. Clearly we can determine the system with the form of work or with its force field. In this case we denote the system as $(Q,{\bf g},{\rm F})$. With this elements we have a differential equation: we look for curves, $\gamma\colon[a,b]\subset\mathbb{R}\to Q$, solutions to the equation \begin{equation}\label{Neweq} \nabla_{\dot{\gamma}}\dot\gamma ={\rm F}\circ\gamma \end{equation} where $\nabla$ is the {\sl Levi-Civita connection} associated to the Riemannian metric ${\bf g}$. Recall that, given a Riemannian metric ${\bf g}$ on the manifold $Q$, the Levi-Civita connection $\nabla$ is the unique linear connection which is symmetrical, that is with null torsion tensor field, and Riemannian, that is $\nabla_Z({\bf g}(X,Y))= {\bf g}(\nabla_ZX,Y) + {\bf g}(X,\nabla_ZY)$, for every $X,Y,Z\in\mathfrak{X}(Q)$, or, what is the same $\nabla_Z{\bf g}$ for every $Z\in\mathfrak{X}(Q)$. Equation (\ref{Neweq}) is called the \textbf{Newton equation}, or dynamical equation, of the system. A {\bf trajectory} of the system is a curve solution to the Newton equation. The manifold $Q$ is the \textbf{configuration space} of the system. If $\dim Q=n$ we say that the system has $n$ \textbf{degrees of freedom}. Its tangent bundle ${\rm T} Q$ is the \textbf{phase space of coordinates--velocities} and the \textbf{phase space of coordinates--momenta} is the cotangent bundle ${\rm T}^*Q$. Both phase spaces are also called {\bf state spaces}. As we are describing physical systems, it must exist {\bf observables} in order to make measures and obtain results from the state of the system. The observables of the system with configuration space $(Q,{\bf g})$ are the real algebra ${\rm C}^\infty ({\rm T} Q)$ of smooth functions defined on the phase space. Equivalently the algebra ${\rm C}^\infty ({\rm T}^*Q)$. The value of an observable $f\in{\rm C}^\infty ({\rm T} Q)$ on a state $(q,v)\in{\rm T} Q$ is the real number $f(q,v)$. \bigskip \noindent{\bf Comment}: If the above definitions and equation are the Riemannian image of the Newtonian mechanics, they must contain in some sense the three Newton laws contained in his ``{\sl Principia Mathematica}", see \cite{Newton}, and it is so: \begin{enumerate} \item Equation (\ref{Neweq}) is no more than the classical $ma=F$, that is ``mass by acceleration is equal to force" written as $$ m\frac{{\rm d}}{{\rm d} t}v=F\, , $$ where the constant $m$ is contained in the covariant derivative, that is in the metric, which is usually called, as we will see in the sequel, the {\bf kinetic energy metric}. \item Note that if ${\rm F}=0$, then the dynamical trajectories are the geodesic curves of the metric ${\bf g}$. This correspond to the first Newton Law ({\sl Inertia Law}\/). This is what, at present, is stated as the existence of \textbf{inertial systems} of coordinates: those coordinate systems where the Newton laws are fulfilled. In our differential geometry formulation not preferred coordinates are used and Inertia Law is a consequence of the dynamical equation. \item With respect to the \textsl{third law}, or action and reaction law, there are long and complicated discussions about its rank of application. For detailed comments see \cite{SPIVAK-2004,Spivak2010} and references therein. It seems that its appropriate domain is in the search of models for different types of forces between bodies in contact or to state the relation of one system and the universe surrounding it. As our systems are studied as isolated ones, this law has no significant meaning in this geometric approach where every system is described and studied by its own structure given by the three defining elements $(Q,{\bf g},\omega)$ without relation to other external elements. \end{enumerate} \bigskip If $(U,\varphi=(x^i))$ is a local chart in $Q$, and $\{\Gamma^k_{ij}\}$ are the corresponding Christoffel symbols of the Levi--Civita connection $\nabla$, then the dynamical equation is locally given by $$ \ddot\gamma^k+\Gamma_{ij}^k\dot\gamma^i\dot\gamma^j={\rm F}^k\circ\gamma \ . $$ The relation in coordinates between $\omega$ and $F$ is as follows: if, $\omega=\omega_i{\rm d} x^i$ and \(\displaystyle {\rm F}={\rm F}^i\derpar{}{x^i}\), then we have that $$ \omega_i=g_{ij}{\rm F}^j \quad , \quad {\rm F}^i=g^{ij}\omega_j \ , $$ where $g^{ij}$ are the components of the inverse matrix of ${\bf g}$ in this local chart (the ``inverse metric''). It is well known that this relation between $\omega$ and $F$ is a particular case of the so called musical diffeomorphisms associated to the Riemannian metric ${\bf g}$, defined by: \begin{eqnarray*} \flat:{\rm T} Q\to{\rm T}^*Q & &\qquad (q,v)\mapsto (q,\mathop{i}\nolimits(v){\bf g}) \\ \vspace{2mm}\sharp:{\rm T}^* Q\to{\rm T} Q& &\qquad (q,\omega)\mapsto (q,\mathop{i}\nolimits(\omega){\bf g}^{-1})\, , \end{eqnarray*} and using these mappings we can write the dynamical equation in dual form: If $\gamma$ is the trajectory of the system, then $\mathop{i}\nolimits(\dot\gamma){\bf g}$ is called the \textbf{linear momentum} and, as $\nabla$ is the Levi--Civita connection, we have that $$ \nabla_{\dot{\gamma}}(\mathop{i}\nolimits(\dot\gamma){\bf g}) =\mathop{i}\nolimits(\nabla_{\dot{\gamma}}\dot\gamma){\bf g}=\mathop{i}\nolimits({\rm F}\circ\gamma){\bf g}=\mathop{i}\nolimits({\rm F}){\bf g} \circ\gamma=\omega\circ\gamma\, , $$ that is: the dynamical equation (\ref{Neweq}) is equivalent to the dual form: \begin{equation}\label{newtondual} \nabla_{\dot{\gamma}}(\mathop{i}\nolimits(\dot\gamma){\bf g}) =\omega\circ\gamma\,. \end{equation} Then, as if ${\rm F}=0$ then $\omega=0$, we have that the linear momentum $\mathop{i}\nolimits(\dot\gamma){\bf g}$ is conserved along the motion which is the dual statement of the inertial law. \subsection{Euler-Lagrange equations} Given a Riemannian manifold $(Q,{\bf g})$, we can associate a natural function defined in the tangent bundle, the so called \textbf{kinetic energy}, defined by: $$ \begin{array}{ccccc} K & \colon & {\rm T} Q & \longrightarrow & \mathbb{R} \\ & & (q,v) &\mapsto & \frac{1}{2}{\bf g}(v,v)\, , \end{array} $$ whose local expression is $$ K(q^i,v^j)=\frac{1}{2}g_{ij}(q)v^iv^j\, . $$ For a Newtonian system $(Q,{\bf g}, {\rm F})$, with $\omega=\mathop{i}\nolimits({\rm F})g$, and using the kinetic energy $K$, we can transform the Newton equation into a new form which is easier to state when we know the elements defining the system. \begin{teor}: Let $(Q,{\bf g},\omega)$ a Newtonian mechanical system, and $\gamma\colon [a,b]\subset\mathbb{R}\to Q$ a smooth curve contained in the domain $U\subset Q$ of a chart $(U,\varphi=(q^i))$ of $Q$. Then, $\gamma$ is a solution to the dynamical equation (\ref{Neweq}) if and only if, it satisfies the equations \begin{equation} \frac{d}{d t}\left(\derpar{K}{v^j}\circ\dot\gamma\right)- \derpar{K}{q^j}\circ\dot\gamma= (\omega\circ\gamma)\left(\derpar{}{q^j}\right)=\omega_j \circ\gamma =g_{ij}F^i \circ\gamma\ , \label{eel} \end{equation} for every $j$, which are called \textbf{Euler-Lagrange equations of the second kind} of the system. \end{teor} These equations are usually written as: $$ \frac{d}{d t}\left(\derpar{K}{v^j}\right)- \derpar{K}{q^j}=\omega_j =g_{ij}F^i \, . $$ The proof is a direct calculus on the function $K$ in local coordinates and using the local expression of the Christoffel symbols of the Levi--Civita connection: $$ [kl,j]=g_{ij}\Gamma^i_{kl}= \frac{1}{2}\left(\derpar{g_{jk}}{q^l}+\derpar{g_{jl}}{q^k}-\derpar{g_{lk}}{q^j}\right) \ . $$ Why have we changed the intrinsic dynamical equation (\ref{Neweq}) into this expression in local coordinates? These equations were obtained by Lagrange in 1788 and published in \cite{Lagrange} and are related with variational calculus. They are easier to calculate for a particular system than the dynamical equation because you don't need to know the Christoffel symbols of the connection. In fact they have the same ``formal" theoretical expression in every coordinate system, that is, you need to apply the same rule to obtain them independently of the used coordinates, called ``generalised coordinates" by Lagrange. \bigskip To finish this section a comment on other kinds of forces. In this initial section we have preferred to keep us in the case of simple forces, depending only on the position coordinates, but it is usual that the mechanical forces, or the work forms, depend not only on the position coordinates but also on the time and the velocities. For the case of time depending forces, see Section 5 where a short introduction with references is given. The dependency on the velocities is the case of dissipative systems or electromagnetic, Lorentz, forces. Geometrically this means that $\omega\in{\mit\Omega}^1(Q,\tau_Q)$ and ${\rm F}\in\mbox{\fr X} (Q,\tau_Q)$, that is they are forms, or vector fields, along the natural projection of the corresponding phase space on the configuration manifold. In this case the only change we need to do in the dynamical equations is to write $\omega\circ\dot{\gamma}$, or ${\rm F}\circ\dot{\gamma}$, instead of $\omega\circ\gamma$ or ${\rm F}\circ\gamma$, respectively. For a detailed study of general forces in mechanics in a geometric way you can see \cite{God-69, Leon-2021}. \subsection{Conservative systems} There is a special kind of Newtonian systems, those which are of conservative, or mechanical Lagrangian, type. A Newtonian system $(Q,{\bf g},\omega)$ is \textbf{conservative} if the work form is exact, that is, there exists $V\in{\rm C}^\infty (Q)$ such that $\omega =-{\rm d} V$. The negative sign is a customary tradition in Physics. In this case, the function $V$ is called the \textbf{potential energy} of the system. Observe that in this case the force vector field is ${\rm F}=-{\rm grad}\ V$. For these systems $(Q,{\bf g},\omega=-{\rm d} V)$, the \textbf{total energy} or \textbf{mechanical energy} of the system is the function $E\in{\rm C}^\infty({\rm T} Q$ defined as $$ \begin{array}{ccccc} E & \colon & {\rm T} Q & \longrightarrow & \mathbb{R} \\ & & (q,v) &\mapsto & K(q,v)+(\tau^*_QV)(q,v)\,. \end{array} $$ We usually write $E=K+V$. As a direct consequence of the definition we have: \begin{teor} \textbf{(Mechanical energy conservation)}: Let $(Q,{\bf g},\omega=-{\rm d} V)$ be a conservative Newtonian mechanical system, then the mechanical energy $E$ is invariant, that is constant along the trajectories of the system. \end{teor} ({\sl Proof\/})\quad If $\gamma\colon [a,b]\subset\mathbb{R}\to Q$ is a solution to the Newton equation, then \begin{equation} \nabla_{\dot\gamma}\dot\gamma={\rm F}\circ\gamma \quad , \quad \mathop{i}\nolimits({\rm F}){\bf g}= \omega=-{\rm d} V \ , \label{eqdinsis} \end{equation} then we have that \begin{eqnarray*} \frac{d (E\circ\dot\gamma)}{d t} &=& \nabla_{\dot\gamma}(E\circ\dot\gamma)= \nabla_{\dot\gamma}\left(\frac{1}{2}{\bf g}(\dot\gamma,\dot\gamma)+V\circ\gamma\right) \\ \vspace{2mm}&=&{\bf g}(\nabla_{\dot\gamma}\dot\gamma,\dot\gamma)+{\rm d} V(\dot\gamma)=0 \ . \end{eqnarray*} \qed In this case, the Lagrange equations have a simpler expression. Consider the function ${\cal L}=K-\tau_Q^*V\in\mathcal{C}^\infty({\rm T} Q)$, called {\bf Lagrangian} of the system. The Euler--Lagrange equations for $(Q,{\bf g},\omega=-{\rm d} V)$ are $$ \frac{d}{d t}\left(\derpar{K}{v^j}\circ\dot\gamma\right)- \derpar{K}{q^j}\circ\dot\gamma=\omega_j\circ\gamma= (-{\rm d} V\circ\gamma)\left(\derpar{}{q^j}\right)= -\derpar{(\tau_Q^*V)}{q^j}\circ\gamma \ , $$ which we can write as \begin{equation} \frac{d}{d t}\left(\derpar{{\cal L}}{v^j}\circ\dot\gamma\right)- \derpar{{\cal L}}{q^j}\circ\dot\gamma=0 \ , \label{eel1e} \end{equation} recalling that \(\displaystyle\derpar{(\tau_Q^*V)}{v^j}=0\). From now on we write ${\cal L}=K-V$ for simplicity. The local expression of ${\cal L}$ is $$ {\cal L} (q,v)=(K-V)(q,v)= \frac{1}{2}g_{ij}(q)v^iv^j-V(q)\, . $$ These conservative systems $(Q,{\bf g},\omega=-{\rm d} V)$ are called {\bf simple mechanical systems} and the associated Lagrangians {\bf natural lagrangians}. \section{Holonomic constrained systems. D'Alembert principle} \protect\label{sdnlh} A relevant topic in classical mechanics is the study of systems with holonomic or nonholonomic constraints. A {\bf constraint} is a restriction in the motion of the system. This restriction may be in the configuration space: the system is obliged to move in a particular subset of this manifold. Or it may restrict the possible velocities to be reached, to be used to move. From our geometric approach, the first ones, called {\bf holonomic constraints}, are defined by a submanifold of the configuration space where the system must remain. The other, or {\bf nonholonomic constraints}, by a submanifold of the phase space of coordinate--velocities allowing to reach all the positions in the configuration manifold. To write the equations of motion we need to add to our postulates some new idea related to the submanifold of constraints. There are several approaches to tackle this problem. The older and best stablished is {\bf d'Alembert principle}, which in geometric terms is specially clarifying on the motion of these kinds of systems. For an historical approach to constrained systems, in particular nonholonomic ones, see \cite{Leon-2012} and the extended bibliography contained there. As d'Alembert principle is not the only one used to obtain the equations of motion, see the previous reference for other known principles and relations among them. For a nice discussion on d'Alembert principle in different situations with a geometric viewpoint, see \cite{MARLE-98}. \subsection{Holonomic constraints. Holonomic d'Alembert principle} Let $(Q,{\bf g},\omega)$ be a Newtonian mechanical system and ${\rm F}\in\mbox{\fr X} (Q)$ its force field. Let ${\bf S}$ be a submanifold of $Q$, usually called {\bf submanifold of holonomic constraints}, and $j\colon ${\bf S}$\hookrightarrow Q$ the natural embedding. We intend to describe the dynamics of the given system when it is obliged to evolve in the submanifold ${\bf S}$ of the configuration space, hence with some restriction in the coordinates the system can reach. To force this behaviour, it is compulsory to apply a new force field ${\rm R}$, called {\bf constraint force}, which obliges the system to remain in ${\bf S}$. In general, such force depends, not only on the position, but also on the velocity; then ${\rm R}\in\mbox{\fr X} (Q,\tau_Q)$ and, moreover, we don't know it; in fact it is a new unknown to find. Then we have a new dynamical equation for curves $\gamma\colon [a,b]\subset\mathbb{R}\to {\bf S}$, which is \begin{equation} \nabla_{\dot\gamma}\dot\gamma ={\rm F}\circ\gamma+{\rm R}\circ\dot\gamma \ . \label{eqdinS} \end{equation} To solve this problem, we introduce the so called \textbf{d'Alembert principle}: The constraint force ${\rm R}$ is orthogonal to the submanifold ${\bf S}$; that is, for every $q\in {\bf S}$ and for every $ u,v\in{\rm T}_q{\bf S}$, we have ${\bf g}(u,{\rm R}(q,v))=0$, supposing that ${\rm R}$ depends on the velocities. We impose that the constraint force is orthogonal to the constraint submanifold. How to obtain the equation of motion and the expression of the constraint force? If $g_{\bf S}=j^*g$, let $\nabla^{\bf S}$ be the Levi-Civita connection in the Riemannian manifoldt $({\bf S},g_{\bf S})$. Then, we have the following natural geometric elements and consequences: \begin{description} \item[a)] For every $q\in {\bf S}$ the orthogonal decomposition ${\rm T}_qQ={\rm T}_q{\bf S}\oplus ({\rm T}_q{\bf S})^\perp $ and the orthogonal projections, \vspace{-2mm} $$ \pi_{\bf S}(q)=\colon {\rm T}_qQ\rightarrow{\rm T}_q{\bf S} \quad , \quad \pi^\perp_{\bf S}(q)=\colon {\rm T}_qQ\rightarrow ({\rm T}_q{\bf S})^\perp \ , $$ \item[b)] The global orthogonal projections \vspace{-2mm} $$ \pi_{\bf S}=\colon {\rm T} Q\vert_ {\bf S}\rightarrow{\rm T} {\bf S} \quad , \quad \pi^\perp_{\bf S}=\colon {\rm T} Q\vert_ {\bf S}\rightarrow {\rm T} {\bf S}^\perp \ . $$ Thus d'Alembert principle is reduced to $\pi_{\bf S}\circ{\rm R}=0$. \item[c)] The Levi--civita connection on the submanifold ${\bf S}$ satisfies $\nabla^ {\bf S}=\pi_ {\bf S}\circ\nabla$. This can be directly proved because $\pi_ {\bf S}\circ\nabla$ is a $g_{\bf S}$--Riemannian symmetrical connection. \end{description} Now, taking the dynamical equation (\ref{eqdinS}) and splitting up it into the tangent and orthogonal components with respect to ${\bf S}$, we obtain respectively \begin{eqnarray} \pi_ {\bf S}(\nabla_{\dot\gamma}\dot\gamma) &=& \pi_ {\bf S}\circ{\rm F}\circ\gamma+\pi_ {\bf S}\circ{\rm R}\circ\dot\gamma = \pi_ {\bf S}\circ{\rm F}\circ\gamma \ , \label{eqdinsplit1} \\ \pi_ {\bf S}^\perp(\nabla_{\dot\gamma}\dot\gamma) &=& \pi_ {\bf S}^\perp\circ{\rm F}\circ\gamma+\pi_ {\bf S}^\perp\circ{\rm R}\circ\dot\gamma = \pi_ {\bf S}^\perp\circ{\rm F}\circ\gamma +{\rm R}\circ\dot\gamma \ . \label{eqdinsplit2} \end{eqnarray} and, denoting by ${\rm F}^{\bf S}=\pi_ {\bf S}\circ{\rm F}\in\mbox{\fr X} ({\bf S})$ the projection of ${\rm F}$ on ${\bf S}$, equation \eqref{eqdinsplit1} is simply \begin{equation} \nabla^{\bf S}_{\dot\gamma}\dot\gamma = {\rm F}^{\bf S}\circ\gamma \ ; \label{eqdinresS} \end{equation} that is: the dynamical equation of the Newtonian mechanical system $({\bf S},g_{\bf S},\omega_{\bf S})$, where $\omega_{\bf S}=\mathop{i}\nolimits({\rm F}^{\bf S})g_{\bf {\bf S}}$. Observe that solutions to equation (\ref{eqdinresS}) are curves $\gamma\colon [a,b]\subset\mathbb{R}\to {\bf S}$ such that, introducing each of them into equation (\ref{eqdinsplit2}), allows us to calculate the constraint force ${\rm R}$ for that trajectory $\gamma$, obtaining \begin{equation}\label{constraintR} \nabla_{\dot\gamma}\dot\gamma -\nabla^{\bf S}_{\dot\gamma}\dot\gamma= {\rm F}\circ\gamma-{\rm F}^{\bf S}\circ\gamma +{\rm R}\circ\dot\gamma \ . \end{equation} Then we know ${\rm R}\circ\dot\gamma\in\mbox{\fr X} (Q,\dot\gamma)$. Notice that we can calculate the constraint force only on every trajectory of the system but {\bf not} as a vector field depending on the velocities: we need to calculate one trajectory using equation (\ref{eqdinresS}) and then we can obtain the constraint force on this trajectory by means of equation (\ref{constraintR}) where the only unknown is ${\rm R}\circ\dot\gamma$. \bigskip As in the case of equation (\ref{newtondual}) we can dualise equation (\ref{eqdinsplit1}) and we obtain $$ \nabla^{\bf S}_{\dot\gamma}(\mathop{i}\nolimits(\dot\gamma){\bf g}_{\bf S})=\omega_{\bf S}\circ\gamma $$ being $\omega_{\bf S}=j^*\omega$. In fact we can state the {\bf dual d'Alembert principle} as follows: If $\rho=\mathop{i}\nolimits({\rm R})g$, then $j^*\rho =0$. And by means of the dual of the above orthogonal projections in the cotangent bundle we can obtain the value of $\rho$ on every trajectory of the system and hence of {\rm R}. This dual form of d'Alembert principle is also called ``Principle of virtual work": the integral of the work form $\rho$ along any piece of a possible trajectory $\gamma$ of the system is zero. \bigskip {\bf Examples:} \begin{enumerate} \item {\bf Systems with one constraint}: Let ${\bf S}=\{ q\in Q \ ;\ \varphi (q)=0\}$, with $\varphi\in{\rm C}^\infty (Q)$ and suppose that ${\rm d}\varphi(q)\neq 0$, for every $q\in {\bf S}$, hence ${\bf S}$ is a hypersurface of $Q$. Let $X\in\mbox{\fr X} (Q)$ such that $\mathop{i}\nolimits (X){\bf g}={\rm d}\varphi$, then $X$ is orthogonal to $ {\bf S}$. In this case we have $$ \pi^\perp_ {\bf S}({\rm F})=\frac{{\bf g}({\rm F},X)}{{\bf g}(X,X)}X= \frac{{\rm d}\varphi ({\rm F})}{\| {\rm d}\varphi\|^2}X \ ,\qquad {\rm F}^ {\bf S}=\pi_ {\bf S}({\rm F})={\rm F}-\frac{{\rm d}\varphi ({\rm F})}{\| {\rm d}\varphi\|^2}X \ . $$ This allows us to find the trajectories of the system as solutions to the differential equation $$ \nabla^ {\bf S}_{\dot\gamma}\dot\gamma= {\rm F}\circ\gamma-\frac{{\rm d}\varphi ({\rm F})}{\| {\rm d}\varphi\|^2}X \ . $$ Once we have a trajectory solution $\gamma$, the constraint force is given by the equation $$ \nabla_{\dot\gamma}\dot\gamma-\nabla^ {\bf S}_{\dot\gamma}\dot\gamma= \frac{{\rm d}\varphi ({\rm F})}{\| {\rm d}\varphi\|^2}\circ\gamma-{\rm R}\circ\dot\gamma \ , $$ where the only unknown is ${\rm R}\circ\dot\gamma$. These expressions are related to the second fundamental form of the hypersurface ${\bf S}$. \item {\bf Systems with several constraints}: Consider now $ {\bf S}=\{ q\in Q \ ;\ \varphi_1(q)=0,\ldots ,\varphi_h(q)=0\}$, with $\varphi_1\ldots ,\varphi_h\in{\rm C}^\infty (Q)$, such that ${\rm d}\varphi_1(q),\ldots{\rm d}\varphi_h(q)$ are linearly independent at every point $q\in {\bf S}$ (we assume that $ {\bf S}$ is not empty). Let $\moment{Z}{1}{n-h}\in\mbox{\fr X} (Q)$ such that: \begin{description} \item[i)] $i(Z_a)d\varphi_{\beta}=0$, for $1\leq \beta\leq h,\,\,1\leq a\leq n-h$, \item[ii)]$ {\bf g}(Z_a, Z_b)=0, a\not= b$, $1\leq a,b\leq n-h$. \end{description} To obtain these vector fields $Z_a$, it is enough to take vector fields $\moment{X}{1}{n-h}\in\mbox{\fr X} (Q)$ satisfying the first condition (a linear equation) and apply the well known {\sl Gramm-Schmidt method}. In this situation, we have that $$ \pi_ {\bf S}({\rm F})=\sum_{a=1}^{n-h}\frac{{\bf g}({\rm F},Z_a)}{{\bf g}(Z_a,Z_a)}Z_a $$ hence, as in the previous case, we obtain the dynamical equation and the expression of the constraint force along every trajectory solution. \end{enumerate} The above examples can be taken as local coordinate expression for a general submanifold of the configuration manifold. \subsection{Euler-Lagrange equations for holonomic constraints} We have shown that for a Newtonian mechanical system $(Q,{\bf g},\omega)$ constrained to move on the submanifold $j\colon {\bf S}\hookrightarrow Q$, the dynamics is given by the Newtonian mechanical system $( {\bf S},g_ {\bf S},\omega_ {\bf S})$. To write the corresponding Euler-Lagrange equation of this last system, take a local chart $(U,q^i)$ in $ {\bf S}$ and the corresponding natural lifting $(\tau_Q^{-1}(U),q^i,v^i)$ to ${\rm T} {\bf S}$. Then we have \begin{equation} \frac{d}{d t}\left(\derpar{K_ {\bf S}}{v^k}\circ\dot\gamma\right)- \derpar{K_ {\bf S}}{q^k}\circ\dot\gamma= (g_ {\bf S})_{ik}{(\rm F^ {\bf S})}^i \ , \label{equno} \end{equation} where $K_ {\bf S}\in{\rm C}^\infty({\rm T} {\bf S})$ is the {\sl kinetic energy} of the system $( {\bf S},g_ {\bf S},\omega_ {\bf S})$. It i {\bf S} easy to show that $K_ {\bf S}=({\rm T} j)^*K$. If the dynamical system is conservative, that is $\omega=-{\rm d} V$, then $$ \omega_ {\bf S}=j^*\omega=-j^*{\rm d} V=-{\rm d} j^*V \ , $$ and the above equation takes the expression $$ \frac{d}{d t}\left(\derpar{{\cal L}_ {\bf S}}{v^j}\circ\dot\gamma\right)- \derpar{{\cal L}_ {\bf S}}{q^j}\circ\dot\gamma=0 \ , $$ where ${\cal L}_ {\bf S}:=({\rm T} j)^*{\cal L}=({\rm T} j)^*(K-V)$. Notice that the constraint force is {\bf not} in these equations. This was one of the innovations developed by Lagrange, to obtain the dynamical equations without mention to the constraint force. In fact, if $(W,x^i)$ is a local chart in $Q$, and $(U,q^j)$ is another in ${\bf S}$, both adapted to the inclusion map $j\colon U\hookrightarrow W$, hence $U={\bf S}\cap W$, given by the local expression $x^i=f^i(q)$, and hence \(\displaystyle \dot x^i=\derpar{f^i}{q^j}\dot q^j\). This shows that it is enough to know the Lagrangian function ${\cal L}$ of the unconstrained system, to introduce these last expressions of $x^i,\dot x^i$ in the Euler-Lagrange equations of the unconstrained system and, by direct derivation, to obtain the Euler-Lagrange equations of this constrained system using the local coordinates $(q^i,\dot{q}^i)$ of ${\rm T}{\bf S}$, the real phase space of the system. See \cite{Ar-89} for interesting comments on this topic. \subsection{Product systems} Suppose we have a family of Newtonian systems, $(Q_\mu,{\bf g}_\mu, {\rm F}_\mu)$, $\mu=1,\ldots, N$. If the force fields ${\rm F}_\mu$ depend only on the corresponding configuration manifold, that is ${\rm F}_\mu\in \mbox{\fr X}(Q_\mu)$, then we have a family of systems of differential equations for a curve $\gamma=(\gamma_1,\ldots,\gamma_N)$, decomposed into $N$ non--coupled equations, one for every $\gamma_\mu$. But if the force fields depend on the manifold product, ${\rm F}_\mu(q_1,\ldots,q_N)$ instead of ${\rm F}_\mu(q_\mu)$, then we have a coupled family of differential equations. In some places these systems are called in interaction. We can represent the situation as another Newtonian system $(Q,{\bf g},\omega)$ with the following elements as configuration manifold, Riemannian metric, work form and force field respectively: $$ Q=\prod_{\mu=1}^NQ_\mu , \qquad {\bf g}=\oplus_{\mu=1}^N{\bf g}_\mu\, ,\qquad\omega=(\omega_1,\ldots ,\omega_N)\,,\qquad {\rm F}=({\rm F}_1,\ldots ,{\rm F}_N) , $$ where in fact $\omega_\mu\in\Omega^1(Q_\mu, \pi_\mu)$, ${\rm F}_\mu\in\mbox{\fr X} (Q_\mu,\pi_\mu)$, being $\pi_\mu\colon \prod_{\nu=1}^N Q_\nu\to Q_\mu$ the natural projections. As we have a new Newtonian system, we can consider the case of constrained one, that is a holonomic constrained system: the system is obliged to move in a submanifold ${\bf S}\subset Q$. The constraint force is also decomposed as ${\rm R}=({\rm R}_1,\ldots,{\rm R}_N)$, where ${\rm R}_\mu\in \mbox{\fr X} (Q_\mu,\pi_\mu)$, $\mu=1,\ldots,N$. The dynamical equation has the same form: $$ \nabla_{\dot\gamma}\dot\gamma={\rm F}\circ\gamma +{\rm R}\circ\dot\gamma \ ; $$ for curves $\gamma\colon [a,b]\subset\mathbb{R}\to S$, $\gamma=(\gamma_1,\ldots,\gamma_N)$, $\gamma_\mu\colon [a,b]\subset\mathbb{R}\to Q_\mu$, or the corresponding constraint submanifold in the case of holonomic constraints. \section{Nonholonomic constrains. Nonholonomic d'Alembert principle} \protect\label{slnh} As far as we know, the description of this kind of systems is not contained in the literature with this Riemannian approach, hence we develop them in detail. For a classical approach see for example \cite{GPS-01,Som-1952}. Other geometric approaches can be seen in \cite{CLMM-2002, GMM-2003,LEWIS1998,Lewis2020}. For an extended bibliography on this topic, for both classical and geometric approaches, see \cite{Leon-2017}. In \cite{Koiller-2019}, there is a dual standpoint using Cartan equivalence that can be directly written in our Riemannian approach. \subsection{Nonholonomic constrained systems} Let $(Q,{\bf g},\omega)$ be a Newtonian mechanical system, ${\rm F}\in\mbox{\fr X} (Q)$ the force field. Let $C$ be a submanifold of ${\rm T} Q$, $j_C\colon C\hookrightarrow {\rm T} Q$ the natural embedding, and suppose that $\tau_Q(C)=Q$. In this situation $C$ is called a {\bf submanifold of nonholonomic constraints}. We want to describe the dynamics of the system when it is constrained to evolve in the submanifold $C$ of the phase space. The constrained system is given by $(Q,{\bf g},F,C)$. Notice that the system is not restricted in the configuration manifold, that is in the positions, but in the possible velocities to move with. As in the holonomic case, to solve this problem we suppose that there exists a {\bf constraint force} ${\rm R}$, usually depending on the velocities, that is ${\rm R}\in\mbox{\fr X} (Q,\tau_Q)$, which forces the system to move in $C$. This constraint force is unknown. Then the Newton dynamical equation is given for curves $\gamma\colon [a,b]\subset\mathbb{R}\to Q$ such that satisfy \begin{enumerate} \item $\dot\gamma (t)\in C$, $ t\in[a,b]$. \item $\nabla_{\dot\gamma}\dot\gamma ={\rm F}\circ\gamma+{\rm R}\circ\dot\gamma$. \end{enumerate} And we need to state conditions allowing us to find the trajectories of the system and calculate ${\rm R}$\footnote{Arnold Sommerfeld, says that this force $\mathrm{R}$ is a ``geometric force'' versus ${\rm F}$ which is an ``applied force''. See \cite{Som-1952}.}. In order to state the nonholonomic d'Alembert principle, we need some geometric preliminaries. Let $(q,v)\in C$. The condition assumed on $C$, $\tau_Q (C)=Q$, tells us that the dimension of the subspace of ${\rm V}_{(q,v)}({\rm T} Q)$ which is tangent to $C$, does not depend on the point $(q,v)$. Let $$ {\rm T}_{(q,v)}^VC={\rm V}_{(q,v)}({\rm T} Q)\cap{\rm T}_{(q,v)}C=\{ w\in{\rm V}_{(q,v)}({\rm T} Q)\ ;\ w\in{\rm T}_{(q,v)}C\} $$ be the vertical subspace tangent to $C$. This is a vector subbundle of ${\rm T}({\rm T} Q)$ and we can write ${\rm T}^VC={\rm V}({\rm T} Q)|_{C}\cap{\rm T} C$ as vector bundles over the manifold $C$. For $(q,v)\in {\rm T} Q$, consider the vertical lifting from the point $q\in Q$ to $(q,v)$ given, as usual, by \begin{align*} \lambda_q^{(q,v)}\colon&{\rm T}_qQ\to{\rm V}_{(q,v)}({\rm T} Q)\\ &\quad u_q\,\mapsto\quad\lambda_q^{(q,v)}(u_{q}):\phi\mapsto\lim_{t\rightarrow 0}\frac{\phi(q,v+tu)-\phi(q,v)}{t}, \end{align*} that is, the directional derivative of $\phi\in{\rm C}^\infty ({\rm T} Q)$ along $u_{q}$ at the point $(q,v)\in{\rm T} Q$. As $\lambda_q^{(q,v)}$ is an isomorphism from ${\rm T}_{q}Q$ to $V_{(q,v)}({\rm T} Q)$, let $({\rm T}_{(q,v)}^VC)_q$ the inverse image of ${\rm T}_{(q,v)}^VC\subset V_{(q,v)}({\rm T} Q)$ by $\lambda_q^{(q,v)}$. Then ${\rm T}_{q}Q=({\rm T}_{(q,v)}^VC)_q\oplus({\rm T}_{(q,v)}^VC)^\perp_q$, being this one an orthogonal decomposition with respect to ${\bf g}$. Now we introduce the \textbf{Nonholonomic d'Alembert principle}: The constraint force ${\rm R}\in\mbox{\fr X} (Q,\tau_Q)$ satisfies that $$ {\rm R}(q,v)\in({\rm T}^V_{(q,v)}C)^\perp_q \ , $$ that is, ${\bf g}({\rm R}(q,v),w)=0$, for every $w\in ({\rm T}^V_{(q,v)}C)_q$. The constraint force in $(q,v)\in C$ is orthogonal to the subspace $({\rm T}^V_{(q,v)}C)_q\subset{\rm T}_q Q$ of tangent vectors at $q$ whose vertical lifting is tangent to $C$. In the classical physics literature, the elements in $({\rm T}^V_{(q,v)}C)_q$ are called {\bf virtual velocities}. \bigskip \noindent{\bf Comment}: If there are no constraints, that is $C={\rm T} Q$, then for every $(q,v)\in C$ we have that ${\rm T}_{(q,v)}^VC=V_{(q,v)}({\rm T} Q)$, hence $({\rm T}^V_{(q,v)}C)_q={\rm T}_{q}Q$ and $({\rm T}^V_{(q,v)}C)^\perp_q=\{0\}$, that is ${\rm R}(q,v)=0$ and there is no constraint force. \bigskip This principle allows to obtain the expression of the constraint force and the dynamical equations of the trajectories of the system as we will see in the next paragraphs. \subsection{Relation between the constraint force and the constraints} In order to do this, we need to characterize the subspace $({\rm T}^V_{(q,v)}C)_q$ in relation with the \textbf{constraints}, defined as the functions vanishing on the submanifold $C$, that is functions $\phi:{\rm T} Q\to\mathbb{R}$ such that $j_C^*\phi =0$. \medskip First, let $\phi\in{\rm C}^\infty ({\rm T} Q)$, and consider the 1-form ${\rm d}^V\phi\in{\mit\Omega}^1(Q,\tau_Q)$ defined by $$ ({\rm d}^V\phi(q,v))(u)= ({\rm d}\phi\circ\lambda_q^{(q,v)})(u)= {\rm d}\phi (\lambda_q^{(q,v)}(u)) , \quad (q,v)\in{\rm T} Q, \quad u\in{\rm T}_qQ \ ; $$ whose expression in a local natural chart $(q^i,v^i)$ of ${\rm T} Q$ is \(\displaystyle{\rm d}^V\phi=\derpar{\phi}{v^i}{\rm d} q^i\). We have the following result: \begin{prop}\label{verticalvectors} Let $(q,v)\in C$. \begin{enumerate} \item If $w\in{\rm T}_qQ$, then \(\displaystyle w\in({\rm T}_{(q,v)}^VC)_q\) if, and only if, $({\rm d}^V\phi(q,v))(w)=0$, for every $\phi\in{\rm C}^\infty ({\rm T} Q)$ such that $j_C^*\phi =0$, that is for every constraint. \item Let $(({\rm T}_{(q,v)}^VC)_q)^o=\{\alpha\in{\rm T}^{*}_{q}Q; \alpha(w)=0, \forall w\in({\rm T}_{(q,v)}^VC)_q\}\subset{\rm T}^{*}_{q}Q$ be the annihilator of $({\rm T}_{(q,v)}^VC)_q$, then $ (({\rm T}_{(q,v)}^VC)_q)^o=\{{\rm d}^V\phi(q,v);\forall \phi\in{\rm C}^\infty({\rm T} Q), j_C^*\phi=0\}. $ \item If $w\in{\rm T}_qQ$, then \(\displaystyle w\in({\rm T}_{(q,v)}^VC)_q\) if, and only if, $\mathop{i}\nolimits(w){\bf g} \in(({\rm T}_{(q,v)}^VC)_q)^o$. \end{enumerate} \end{prop} ({\sl Proof\/})\quad To prove the first item let $w\in{\rm T}_qQ$, then we have \begin{eqnarray*} &w\in({\rm T}_{(q,v)}^VC)_q& \Leftrightarrow \\ &\lambda_q^{(q,v)}(w)\in{\rm T}_{(q,v)}^VC &\Leftrightarrow\\ \lambda_q^{(q,v)}(w)(\phi)=0, &\forall \phi\in{\rm C}^\infty({\rm T} Q), \mathrm{with} \,\, j_C^*\phi=0 &\Leftrightarrow\\ {\rm d}\phi(\lambda_q^{(q,v)}(w))=0, &\forall \phi\in{\rm C}^\infty({\rm T} Q), \mathrm{with} \,\, j_C^*\phi=0 &\Leftrightarrow \\ ({\rm d}^V\phi(q,v))(w)=0,&\forall \phi\in{\rm C}^\infty({\rm T} Q), \mathrm{with} \,\, j_C^*\phi=0.& \end{eqnarray*} The second item is a direct consequence of the first and the third can be obtained from the definitions. \qed \begin{corol} Let ${\rm R}\in\mbox{\fr X} (Q,\tau_Q)$ and $(q,v)\in C$; then ${\rm R}(q,v)\in({\rm T}^V_{(q,v)}C)^\perp_p$ if, and only if, $\mathop{i}\nolimits({\rm R}(q,v)){\bf g}\in(({\rm T}_{(q,v)}^VC)_q)^o$. \end{corol} Usually the submanifold $C$ is given by the annihilation of a finite family of constraints, functions defined in ${\rm T} Q$. We are going to characterize $(({\rm T}_{(q,v)}^VC)_q)^o$ using these constraints. \bigskip From here to the end of this section, we suppose that the submanifold $C$ is defined by the vanishing of $r$ functions $\{\phi^i\}$, with $r<n=\dim\, Q$, satisfying the condition $$ \mathrm{rank}\,\left(\derpar{\coor{\phi}{1}{r}}{\coor{v}{1}{n}}\right)=r\, . $$ Then $\dim\, C=2n-r$ and we have: \begin{prop} Let $(q,v)\in C$, then \begin{enumerate} \item $\dim {\rm T}^V_{(q,v)}C=n-r$. \item $({\rm T}^V_{(q,v)}C)_{q}=\{w\in {\rm T}_{q}Q; ({\rm d}^V\phi^i(q,v))(w)=0, i=1,\ldots,r \}$. \item The subspace $(({\rm T}_{(q,v)}^VC)_q)^o$ is generated by $\{{\rm d}^V\phi^1(q,v),\ldots,{\rm d}^V\phi^r(q,v)\}$. Or what is the same, if $\alpha\in{\rm T}_{q}^{*}Q$ satisfies $\alpha |_{({\rm T}^V_{(q,v)}C)_{q}}=0$, then $\alpha$ is a linear combination of ${\rm d}^V\phi^1(q,v),\ldots,{\rm d}^V\phi^r(q,v)$. \end{enumerate} \end{prop} ({\sl Proof\/})\quad Let $(q^{i}, v^{i})$ a natural coordinate system on $TQ$. \begin{enumerate} \item The assumed condition \(\displaystyle {\rm rank}\,\left(\derpar{\coor{\phi}{1}{r}}{\coor{v}{1}{n}}\right)=r\) implies that, up to a change of order in the coordinates $\coor{q}{1}{n}$, we can suppose that $$ \det\, \left(\derpar{\coor{\phi}{1}{r}}{\coor{v}{1}{r}}\right)\not= 0. $$ Then $(q^{1},\ldots,q^{n},\phi^1,\ldots,\phi^r,v^{r+1},\ldots,v^{n})$ is a local coordinate system of ${\rm T} Q$ by the Inverse Function Theorem. The vector space $V_{q,v}({\rm T} Q)$ is generated by $$ \left\{\derpar{}{\phi^1},\ldots,\derpar{}{\phi^r},\derpar{}{v^{r+1}},\ldots,\derpar{}{v^n}\right\}_{(q,v)}\, , $$ and the subspace ${\rm T}^V_{(q,v)}C\subset V_{q,v}({\rm T} Q)$ is generated by $$ \left\{\derpar{}{v^{r+1}},\ldots,\derpar{}{v^n}\right\}_{(q,v)}\, . $$ Hence the first item is proved. \item The inclusion part is proved in the first item of Proposition \ref{verticalvectors} and the equality is a matter of dimensions. \item The previous items imply that $\{{\rm d}^V\phi^1(q,v),\ldots,{\rm d}^V\phi^r(q,v)\}$ is a basis of $ (({\rm T}_{(q,v)}^VC)_q)^o$. \end{enumerate} \qed Then, as a corollary we obtain: \begin{prop} For $(q,v)\in C$, the form $\eta\in{\mit\Omega}^1(Q,\tau_Q)$ satisfies \(\displaystyle j^*_C\eta\vert_{({\rm T}^V_{(q,v)}C)_q}=0\) if and only if there exist $\moment{\lambda}{1}{r}\in{\rm C}^\infty ({\rm T} Q)$ such that $\eta=\lambda_\alpha{\rm d}^V\phi^\alpha=\lambda_1{\rm d}^V\phi^1+\ldots+\lambda_r{\rm d}^V\phi^r$ \end{prop} ({\sl Proof\/})\quad Because for every $(q,v)\in C$, we have that $\eta(q,v)\in(({\rm T}_{(q,v)}^VC)_q)^o$. \qed \bigskip In the case that the constraints define only locally the submanifold $C$, then the above results are valid only in the corresponding open set. The last Proposition allows us to state the so called \rm{Dual d'Alembert nonholonomic principle}: {\it The work form $\mathop{i}\nolimits({\rm R}){\bf g}$ corresponding to the constraint force ${\rm R}$ annihilates the virtual velocities of the system}. And, as an immediate result, we have: \begin{corol} If ${\rm R}$ is the nonholonomic constraint force, then there exist $\lambda_1,\ldots,\lambda_r\in{\rm C}^\infty ({\rm T} Q)$ such that $$ \mathop{i}\nolimits({\rm R}){\bf g}=\lambda_\alpha{\rm d}^V\phi^\alpha=\lambda_\alpha\derpar{\phi^\alpha}{v^j}{\rm d} q^j \ , $$ and, as a consequence, $$ {\rm R}=\lambda_\alpha\derpar{\phi^\alpha}{v^j}g^{jk}\derpar{}{q^k} \ . $$ \end{corol} \begin{definition} The functions $\lambda_1,\ldots,\lambda_r$ are called \textbf{Lagrange multipliers} of the nonholonomic system. \end{definition} \noindent{\bf Comment}: Observe that in the case of holonomic constraints we could not obtain the global expression of the constraint force, we obtain the constraint form along any particular trajectory of the system, but in the present situation we can obtain one expression for ${\rm R}$ as a vector field depending on the velocities and on the Lagrange multipliers. This is because holonomic constraints are not a particular case of the nonholonomic ones. See also the comment following equation (\ref{constraintR}). \bigskip \noindent{\bf Important particular case}: The submanifold $C\subset{\rm T} Q$ is a linear subbundle of ${\rm T} Q$. \begin{enumerate} \item In this case $C$ is defined by the annihilation of a family of differential forms, that is, we have $\omega^{\alpha}\in\Omega^{1}(Q)$, $\alpha=1,\ldots,r$, linearly independent at every point of $Q$, and $$ C=\{(q,v)\in{\rm T} Q;\, \omega^{\alpha}_{q}(v)=0,\alpha=1,\ldots,r\}\, . $$ In local coordinates, if $\omega^{\alpha}=a^{\alpha}_{j}(q){\rm d} q^{j}$, then $\phi^\alpha=a^{\alpha}_{j}(q)v^{j}$, that is the constraints are linear in the velocities, and the expression of the constraint force is $$ {\rm R}=\lambda_\alpha a^\alpha_{j}g^{jk}\derpar{}{q^k} \ . $$ \item Alternatively we can suppose that the subbundle $C$ is given as a regular distribution $\mathcal{D}$, the distribution annihilated by $\{\omega^{\alpha}, \alpha=1,\ldots,r\}$. If $(q,v)\in\mathcal{D}$, by linearity we have that ${\rm T}_{(q,v)}^VC=\lambda_{q}^{(q,v)} (\mathcal{D}_{q})$, hence $({\rm T}_{(q,v)}^VC)_{q}=\mathcal{D}_{q}$ and $({\rm T}_{(q,v)}^VC)^\perp_q=\mathcal{D}^\perp_q$, then the constraint force ${\rm R}$ is orthogonal to $\mathcal{D}$. This is the situation usually considered in the classical books on mechanics. \item If the distribution $\mathcal{D}$ is integrable and $(q,v)\in\mathcal{D}$ is the initial condition of the dynamical equation for the solution $\gamma$, then the image of $\gamma$ is contained in the integral submanifold of $\mathcal{D}$ passing throught the point $q\in Q$, because $\dot\gamma(t)\in\mathcal{D}_{\gamma(t)}$ for every $t$. The constraint force ${\rm R}$, orthogonal to $\mathcal{D}$, obliges the system to move on the integral submanifolds of the constraint distribution $\mathcal{D}$. \end{enumerate} \bigskip \noindent{\bf Comments}: \begin{enumerate} \item We can understand the solution as follows: if we have $\phi:{\rm T} Q\to\mathbb{R}$, an only constraint, there is an associated 1-form, ${\rm d}^V\phi$, which gives a ``constraint force'' $R^{\phi}$ such that $\mathop{i}\nolimits(R^{\phi}){\bf g}$ is proportional to ${\rm d}^V\phi$. If we have $r$ independent constraints $\{\phi^\alpha\}$, then we have the corresponding constraint forces, $R^{\phi^\alpha}$, and the subbundle generated by them, $\{R^{\phi^\alpha}\}$, and the resultant constraint force ${\rm R}$ is contained in this subbundle. \item For these systems, d'Alembert principle says that if the system moves ``along the vertical fibres'' of C, the work realised by the constraint force, the integral along the trajectory, is null. This is called the {\bf virtual work principle} as alternative to d'Alembert principle. \end{enumerate} \subsection{Dynamical equations for nonholonomic systems} We finish this section giving the expressions of the dynamical equations of nonholonomic constrained systems $(Q,{\bf g},\omega, C)$ in different significant cases of $C\subset {\rm T} Q$, the submanifold of nonholonomic constraints. \begin{enumerate} \item If the submanifold of constraints $C$ is locally defined by the annihilation of $r$ functions $\{\phi^\alpha\}$, and they are independent constraints, then the dynamical equation is $$ \nabla_{\dot\gamma}\dot\gamma ={\rm F}\circ\gamma+\lambda_\alpha\mathop{i}\nolimits({\rm d}^V\phi^\alpha)g^{-1}\circ\dot\gamma \ , $$ or, in dual form, $$ \nabla_{\dot\gamma}(\mathop{i}\nolimits(\dot\gamma){\bf g})= \omega\circ\gamma+\lambda_\alpha{\rm d}^V\phi^\alpha\circ\dot\gamma \ . $$ These equations together with the constraints defining $C$, $\phi^1=0,\ldots,\phi^r=0$, are a system of $n+r$ equations with $n+r$ unknowns: the components of the trajectory $\gamma$ and the Lagrange multipliers $\lambda_\alpha$. Observe that some of them are the dynamical equations and the remaining ones are the constraint functions. The corresponding Euler-Lagrange equations are $$ \frac{d}{d t}\left(\derpar{K}{v^j}\circ\dot\gamma\right)- \derpar{K}{q^j}\circ\dot\gamma= \omega_j\circ\gamma +\lambda_\alpha\derpar{\phi^\alpha}{v^j}\circ\dot\gamma \quad , \quad (j=1,\ldots ,n) $$ because $\omega=\omega_k{\rm d} q^k$, with $\omega_k=g_{kj}{\rm F}^j$. \item If the system is conservative, then $\omega=-{\rm d} V$, where $V\in{\rm C}^\infty (Q)$ is the potential function. In this case we can introduce the Lagrangian function ${\cal L}=K-\tau_Q^*V$ and we have $$ \frac{d}{d t}\left(\derpar{{\cal L} }{v^j}\circ\dot\gamma\right)- \derpar{{\cal L}}{q^j}\circ\dot\gamma= \lambda_\alpha\derpar{\phi^\alpha}{v^j}\circ\dot\gamma \ . $$ As above, these equations, together with the constraints defining $C$, are also a system of $n+r$ equations with $n+r$ unknowns. \item If $C$ is a vector subbundle, then $\phi^\alpha=a^{\alpha}_{j}(q)v^{j}$ and the Euler-Lagrange equations are $$ \frac{d}{d t}\left(\derpar{K}{v^j}\circ\dot\gamma\right)- \derpar{K}{q^j}\circ\dot\gamma= \omega_j\circ\gamma +f_ka^k_j\circ\dot\gamma \quad , \quad (j=1,\ldots ,n)\, . $$ Or in the case of conservative systems $$ \frac{d}{d t}\left(\derpar{{\cal L} }{v^j}\circ\dot\gamma\right)- \derpar{{\cal L}}{q^j}\circ\dot\gamma= \lambda_\alpha a^\alpha_j\circ\dot\gamma \ . $$ \end{enumerate} If $C$ is an affine subbundle, then $\phi^\alpha=a^{\alpha}_{j}(q)v^{j}+b^\alpha(q)$ and the expression of the Euler-Lagrange dynamical equations is the same as above. \section{Non-autonomous Newtonian systems} In some interesting cases the force field acting on a Newtonian system depends not only on the positions and the velocities but also on time. They are called {\bf non-autonomous} or {\bf time-depending} systems. In the following paragraphs we will try to extend the above geometric formulation to this situation. The geometric model appropriate to this case is the following: A \textbf{non-autonomous Newtonian mechanical system} is a triple $(\mathbb{R}\times Q,{\bf g},{\rm F})$, where $(Q,{\bf g})$ is a Riemannian manifold and the force field is ${\rm F}\in\mbox{\fr X} (Q,\pi_2)$, with $\pi_2\colon\mathbb{R}\times Q\to Q$; that is, $$ \xymatrix{&{\rm T} Q\ar[d]^{\tau_Q}\\{\mathbb{R}\times Q}\ar[ur]^{{\rm F}}\ar[r]^{\quad\pi_2}&\, Q\,.} $$ Moreover, if the force field depends on the velocities, then ${\rm F}\in\mbox{\fr X} (Q,\tau_Q\circ\rho_2)$, with $\rho_2\colon\mathbb{R}\times{\rm T} Q\to {\rm T} Q$; that is, $$ \xymatrix{& &{\rm T} Q\ar[d]^{\tau_Q}\\{\mathbb{R}\times {\rm T} Q}\ar[urr]^{{\rm F}}\ar[r]^{\quad\rho_2}&{\rm T} Q\ar[r]^{\tau_Q}&\, Q\,.} $$ The Newton equations are written in the usual way: \begin{itemize} \item In the case that the force does not depend on the velocities $$ \nabla_{\dot\gamma}\dot\gamma ={\rm F}\circ\bar\gamma $$ where $\bar\gamma=(t,\gamma)\colon I\subset\mathbb{R}\to I\times Q$. We can also use the dual form by means of the corresponding work form $\omega\in{\mit\Omega}^1(Q,\pi_2)$. \item If the force field depends on the velocities $$ \nabla_{\dot\gamma}\dot\gamma ={\rm F}\circ\bar{\dot\gamma} \ , $$ where $\bar{\dot\gamma}=(t,\dot\gamma)\colon I\subset\mathbb{R}\to I\times{\rm T} Q$. As above we can use the corresponding work form $\omega\in{\mit\Omega}^1(Q,\tau_Q\circ\rho_2)$ and obtain the equations in the dual form. \end{itemize} The Euler-Lagrange equations are the same as usual but the second term depends on time $t\in\mathbb{R}$. In particular, if the work form depends on time, $\omega\in{\mit\Omega}^1(Q,\pi_2)$, we say that the system is \textsl{conservative} if there exists $V\colon\mathbb{R}\times Q\to \mathbb{R}$, such that $\omega=-{\rm d} V_t$, where $V_t\colon Q\to Q$ is defined by $V_t(p):=V(t,p)$, for every $p\in Q$ and $t\in\mathbb{R}$. In this situation we can define the Lagrangian function ${\cal L}=K-V$, depending on time, and the Euler-Lagrange equation are as usual $$ \frac{d}{d t}\left(\derpar{{\cal L}}{v^i}\circ\bar{\dot\gamma}\right)- \derpar{{\cal L}}{q^i}\circ\bar{\dot\gamma}= 0 \ . $$ The case of time depending constrained systems, both holonomic and nonholonomic, can be directly formulated with the adequate changes. The constraint force also depends on time. It is interesting to note that if the Lagrangian function is time-depending, then the system is not conservative, that is, the energy function is not conserved along the trajectories of the system. In fact it is easy to show that $$ \frac{dE\circ \dot\gamma}{dt}=-\frac{\partial L}{\partial t}\circ \dot\gamma\, . $$ using the definition of the total energy from the Lagrangian, $E=K+V$. Observe that, in the above comments we have supposed that the dependency on the time is in the forces, but there are systems where it is in the kinetic energy, that is in the Riemannian metric. This is the case of variable mass systems, like a rocket whose mass changes while the combustion goes; see \cite{Fox-67} and \cite{Cve-16} for other interesting examples. There are also constrained systems whose constraints depend on time, see for example \cite{Ar-89}. Other different applications in mechanics of time depending Riemanian metrics can be seen in \cite{SarPrin-2010,MesSarCram-2011,SarPrinMes-kup-2012} and references therein. Some problems in geometry, mechanics and relativity consider the case of Riemannian metrics depending on parameters, the time for example when we have an evolution problem, but they are out of the scope of this survey. \section{Geodesic fields, Hamilton-Jacobi equation and applications} Hamilton--Jacobi equation is a fundamental tool to obtain significant results in the study of Lagrangian and Hamiltonian systems, but for a Newtonian system we have not, in general, this tool. Following ideas contained in \cite{CGMMR-06}, we develop in this section the associated notion, which we call Hamilton--Jacobi vector fields. They can be associated with non conservative forces, that is when we haven't a Lagrangian or Hamiltonian function. We compare both situation, the classical and the Newtonian, and give some particular applications in different fields. \subsection{Hamilton--Jacobi vector fields} Consider a Newtonian system $(Q,{\bf g},{\rm F})$ and the associated dynamical equation for curves $\gamma:I\subseteq \mathbf{R}\longrightarrow Q$ \begin{equation}\label{DIN} \nabla_{\dot{\gamma}(t)}\dot{\gamma}(t)={\rm F}(\gamma(t))\,. \end{equation} This is a second order differential equation for the curves $\gamma$, hence it is associated to a second order vector field in the phase space ${\rm T} Q$. Following ideas of Jacobi, we try to reduce the order of this differential equation and we will obtain the Hamilton--Jacobi equation for Newtonian systems in this Riemannian approach. For a similar approach in the Lagrangian setting, see \cite{CGMMR-06}. In \cite{BLMDMM-2012} you can see a more general approach in the atmosphere of skew symmetric algebroids. First we begin with a geometric result on vector fields on the configuration space $Q$. \begin{teor} The vector field $X\in\mathfrak{X}(Q)$ satisfies the equation \begin{equation}\label{HJL} \nabla_X X={\rm F} \end{equation} if and only if its integral curves $\gamma:I\rightarrow Q$ are trajectories of the above Newtonian system, that is they satisfy equation (\ref{DIN}). We say that $X$ is a \textbf{Hamilton-Jacobi vector field} associated to the Newtonian system $(Q,{\bf g},{\rm F})$. \end{teor} \begin{proof} For any $p\in Q$, let $\gamma$ be the integral curve of $X$, $\dot{\gamma}=X\circ\gamma$, with initial condition $\gamma(0)=p$. Suppose that $\gamma$ satisfies (\ref{DIN}). Then at $t=0$ we have: \begin{equation} \nabla_{\dot{\gamma}(0)}\dot{\gamma}(t)={\rm F}(\gamma(0)) \end{equation} that is \begin{equation} \nabla_{X(p)}X={\rm F}(p)\,. \end{equation} But the point $p\in Q$ is arbitrary, hence $\nabla_{X}X={\rm F}$ as we wanted. On the other side, if $X$ satisfies $\nabla_{X}X={\rm F}$ and $\gamma$ is an integral curve of $X$, then $$ \nabla_{\dot{\gamma}(t)}\dot{\gamma}(t)= (\nabla_{X}X)(\gamma(t))={\rm F}(\gamma(t))\,, $$ and the curve $\gamma$ satisfies equation (\ref{DIN}). \end{proof}\qed \bigskip \noindent{\bf Comments}: \begin{enumerate} \item Notice that by this result, we obtain solutions of a second order differential equation, that is integral curves of a vector field in ${\rm T} Q$, as integral curves of vector fields in the manifold $Q$, any vector field solution to the equation $\nabla_X X={\rm F}$, that is first order differential equations. But, given a solution $X\in\mathfrak{X}(Q)$, we do not obtain from $X$ all the integral curves of the second order equation, we only obtain those with initial conditions $(q,u)\in{\rm T} Q$ of the form $(q,X(q))$. In fact, for every solution $X$, we obtain a family of curves solution to the dynamical equation (\ref{DIN}): they are those solution curves contained in the submanifold of ${\rm T} Q$ defined by the graph of $X$ as a section of the natural projection $\tau_Q:{\rm T} Q\rightarrow Q$. It can be proved that this submanifold is invariant by the second order differential system associated to the dynamical equation. See \cite{CGMMR-06} for more details. \item If ${\rm F}=0$, we have the family of vector fields satisfying the equation $\nabla_X X=0$, the so called {\bf geodesic vector fields}. Their integral curves are geodesic curves for the Levi--Civita connection $\nabla$, they are solutions to the equation $\nabla_{\dot{\gamma}(t)}\dot{\gamma}(t)=0$. Some interesting properties of these vector field are in \cite{CMM--2021} and references therein. \end{enumerate} \subsection{Classical Hamilton-Jacobi equation in Lagrangian form} Let $(Q,{\bf g},\omega=-{\rm d} V)$ be a conservative Newtonian system with force field ${\rm F}=-\mathrm{grad}\,V$. The Lagrangian function is $L=K-V$ and the total energy function is $E=K+V$. Suppose the vector field $X$ is a Hamilton--Jacobi vector field of the system, $\nabla_X X={\rm F}$, then $$ \mathcal{L}_X (E\circ X)=\mathcal{L}_X\left(\frac{1}{2}\textbf{g}(X,X)+V\right)= $$ $$ =\textbf{g}(\nabla_X X,X)+\mathrm{d}V(X)= \textbf{g}(F,X)+\mathrm{d}V(X)=0 $$ which is a conservation of energy theorem for the Hamilton--Jacobi vector fields. And following this last result we have: \bigskip \begin{teor} (Hamilton--Jacobi equation) Suppose that the vector field $X\in\mathfrak{X}(Q)$ satisfyes the condition $\mathrm{d}(\mathop{i}\nolimits(X) \textbf{g})=0$, for example if $X$ has a potential function (see the comments below). Then the following conditions for $X$ are equivalent: \begin{enumerate} \item $\nabla_X X=F=-\mathrm{grad}\,V$. \item $\mathrm{d}(E\circ X)=0$. \end{enumerate} \end{teor} \begin{proof} Let $Y$ be a vector field on $Q$, then: \vspace{-2mm} \begin{eqnarray*} \mathrm{d}(E\circ X)(Y)&=&\mathcal{L}_Y(E\circ X)=\mathcal{L}_Y\left(\frac{1}{2}(\textbf{g}(X,X)+V\right)=\textbf{g}(\nabla_Y X,X)+\mathrm{d}V(Y)\\ &=&\textbf{g}(\nabla_X X,Y)+\mathrm{d}V(Y)=\textbf{g}(\nabla_X X-F,Y)\,, \end{eqnarray*} where the fourth identity is consequence of $$ \mathrm{d}(i_X \textbf{g})(Y,Z)=\textbf{g}(\nabla_Y X,Z)-\textbf{g}(\nabla_Z X,Y) $$ directly obtained from the definition of the exterior differential and the symmetry of the connection $\nabla$. But $Y$ is arbitrary, hence we have the equivalence. \end{proof} \qed \bigskip \noindent{\bf Comments}: \begin{enumerate} \item Notice that the condition $\mathrm{d}(E\circ X)=0$ is equivalent to say that $E\circ X$ is a constant in every connected component of $Q$, that is $$E(q^1,\ldots,q^n, X^1,\ldots,X^n)=\mathrm{constant}\,,$$ which is no more that the Hamilton--Jacobi equation but in Lagrangian form. Its solutions are the vector fields $X\in\mathfrak{X}(Q)$ satisfying $\nabla_X X={\rm F}$ and $\mathrm{d}(\mathop{i}\nolimits(X) \textbf{g})=0$. \item If we want to obtain the classical form of the Hamilton--Jacobi equation, that is the Hamiltonian form, we need to construct the Hamiltonian function, $H:{\rm T}^*Q\to\mathbb{R}$, and change the vector field $X$ by the closed differential form $\alpha=\mathop{i}\nolimits(X) \textbf{g}$. Then, locally, $\alpha={\rm d} S$ and we obtain the classical equation. The closedness of the form is related to the special condition we have imposed to the vector field $X$ and it is equivalent to say that the image of the form $\alpha$ in ${\rm T}^*Q$ is a Lagrangian submanifold of ${\rm T}^*Q$ with its natural symplectic structure (see \cite{CGMMR-06}). \item Condition $\mathrm{d}(i_X \textbf{g})=0$ for the vector field $X$ is equivalent to say that the image of the map $X:Q\to\mathrm{T}Q$ is a Lagrangian submanifold of the symplectic manifold $\mathrm{T}Q$ with the 2-Lagrangian form associated to $L=K-V$, which is a regular Lagrangian. It is related with the previous item, as $\alpha=i_X \textbf{g}$, and the Legendre transformation associated to the kinetic energy. Once again, see \cite{CGMMR-06} for more details. \end{enumerate} \subsection{Jacobi metric} Consider a conservative mechanical system defined on the Riemannian manifold $(Q,{\bf g})$ with potential energy function $V \colon Q \to \mathbb{R}$. Recall that the energy is given by $E(v_q) = \frac12 {\bf g}(v_q,v_q) + V(q)$ and it is a constant of the motion. Suppose $E_0 > V(q)$ for all $q \in Q$. We define the {\bf Jacobi metric} as $$ {\bf g}_0 = (E_0-V) g . $$ It is well known that the solutions $\gamma$ of the Newton equation $\nabla_{\dot\gamma} \dot\gamma = -\mathrm{grad}\, V \circ \gamma$ with fixed energy $E_0$ are, under convenient reparametrization, the geodesic lines of~${\bf g}_0$. In \cite{God-69} there are interesting comments on this topic, and you can see a nice new approach of the same question in \cite{CMM--2021}. \subsection{Applications} \subsubsection{Euler equation for fluids as a Hamilton--Jacobi equation of a Newtonian system} In this paragraph we interpret the Euler equation for stationary fluids as a Hamilton--Jacobi equation for a Newtonian system. Let $U\subset\mathbf{R}^3$ an open set. The classical Euler equation for a time--depending vector field $X\in\mathfrak{X}(U)$, in Cartesian coordinates is given as $$ \frac{\partial X}{\partial t}+<X,\nabla>X=-\nabla p $$ where $p:U\times\mathbf{R}\to\mathbf{R}$ is a function, the pressure on the fluid. If we consider the case where the fluid is incompressible, then we include the condition $\mathrm{div}\,X=0$, that is the vector field is conservative. For a point $(x,t)$, the tangent vector $X(x,t)\in{\rm T}_x U$ is the velocity of the fluid particle in the point $x$ at the moment $t$. In a Riemannian manifold $(M,\textbf{g})$, the corresponding Euler equation for a time--depending vector field $X\in\mathfrak{X}(M)$ is written as: $$ \frac{\partial X}{\partial t}+\nabla_X X=-\mathrm{grad}\, p\, . $$ The condition of being conservative for $X$ is $L_X \mathrm{d}V=0$, where $\mathrm{d}V$ is the volume element associated to $\textbf{g}$. The equivalence of both equations in direct in local coordinates. For a stationary motion, that is constant in time, the last equation is: $$ \nabla_X X=-\mathrm{grad}\, p $$ that is, the Lagrangian Hamilton--Jacobi equation (\ref{HJL}) with ${\rm F}=-\mathrm{grad}\, p$, corresponding to the conservative Newtonian system $(M,{\bf g},\omega=-{\rm d} p)$. \subsubsection{Hamilton--Jacobi equation and Schr\"odinger equation. } Given a Newtonian conservative system $(M,{\bf g},\omega=-{\rm d} V)$ and a vector field $X\in\mathfrak{X}(M)$ which is a gradient, $X=-\mathrm{grad}\,S$, consider the following three conditions: \begin{enumerate} \item $X$ is a Hamilton--Jacobi vector field for the given system, that is $\nabla_X X=-\mathrm{grad}\, V$ or equivalently $E\circ \mathrm{grad}\,S=E_o=\mathrm{constant}$. \item $\mathrm{div} X=0$, hence $\mathrm{div}\, \mathrm{grad}\,S=\Delta S=0$, where $\Delta$ is the Laplacian operator for the Riemannian metric ${\bf g}$. \item The function $\Psi= \exp(iS)$ satisfies the Schr\"odinger equation $$ \left(-\frac{1}{2}\Delta+V\right)\Psi=E_o\Psi\, . $$ \end{enumerate} Then two of them imply the other one. This is stated in \cite{ALOMU-2014} and is a nice relation between classical and quantum worlds, specially highlighted with this Riemannian approach to mechanics because all the elements included in the statement have a clear geometric meaning. For more explanations of these relations and interesting ideas see \cite{CMMGM-2016,MARMO-2009}. See also \cite{CHERO2001} for a Riemannian approach to quantum mechanics. We only include the first paper of a series of four. \section{Other interesting topics} In this section we give some references about topics where this Riemannian approach has given a modern revival of results and applications. We limite ourselves to give some information on some significant authors and recommend some of their publications and web pages as a source of information and update on these topics in the Riemannian approach. In any case, it is necessary to say that this is not a full list of groups or researchers which have contributed to the study of these problems with this approach. \subsection{Symmetry and conserved quantities} Symmetries of the system, associated conserved quantities and reduction by the action of symmetry groups are deep ideas in the study of mechanical systems from the very beginning. In our approach, the Riemannian structure gives a natural set of symmetries, the isometries of the metric and, infinitesimally, the Killing vector fields. For example, in the case that the configuration manifold is $\mathbb{R}^3$, invariance by translations and rotations give rise to linear and angular momenta in the classic terminology. We do not enter in historical references on a so extended subject, which from a geometric viewpoint goes back to S. Lie in nineteenth century, with well recognized groups working on different topics and with a large amount of significant deep results. As an example we cite the work of G. Marmo and collaborators making geometry and symmetry fundamental tools to understand physical problems. See \cite{CIMM-2015} and the references therein as a book collecting their work along the years. Other well known books on the topic have been written by J. Marsden and coauthors or A. Bloch and others authors. See, e.g. \cite{BLOCH2015} and the personal web pages of the authors, the web page of J. Marsden, specially maintained since he passed away in 2010, where most of his books and research are open accessible and are a source of information and results on this topic. These given references contain extended lists of names and topics related to the use of symmetry to understand problems in mechanics and other problems. Hence we only give an example of a general result, a kind of {\bf Noether theorem} for the situation under study, that is mechanics with the Riemannian approach. \medskip Suppose we have a Newtonian system $(M,{\bf g},{\rm F})$. For a vector field $X\in\mbox{\fr X}(M)$, consider the function ${\bf I}_X=\widehat{\mathop{i}\nolimits(X){\bf g}}:{\rm T} M\to\mathbb{R}$ defined by ${\bf I}_X(v_q)={\bf g}(X_q,v_q)$. Then we have \begin{enumerate} \item If $X$ is a Killing vector field, that is $\mathcal{L}_X{\bf g}=0$, and $\sigma:I\subset\mathbb{R}\to M$ is a geodesic line, that is $\nabla_{\dot\sigma}{\dot\sigma}=0$, then ${\bf I}_X$ is constant along $\dot\sigma$, that is \begin{equation} \frac{{\rm d}}{{\rm d} t}({\bf I}_X\circ\dot\sigma)(t)=0\,. \end{equation} In other words, if $X$ is a Killing vector field, then ${\bf I}_X$ is a conserved quantity along the geodesic lines. \item If $X$ is a Killing vector field satisfying ${\bf g}(X,{\rm F})=0$, and $\gamma:I\subset\mathbb{R}\to M$ is a trajectory of the system $(M,{\bf g},{\rm F})$, that is $\nabla_{\dot\gamma}{\dot\gamma}={\rm F}\circ\gamma$, then ${\bf I}_X$ is constant along $\dot\gamma$, that is \begin{equation} \frac{{\rm d}}{{\rm d} t}({\bf I}_X\circ\dot\gamma)(t)=0\,. \end{equation} In other words, if $X$ is a Killing vector field orthogonal to the force field ${\rm F}$, then ${\bf I}_X$ is a conserved quantity along the geodesic lines. \end{enumerate} Both results are consequence of the following equation \begin{equation} \frac{{\rm d}}{{\rm d} t}\left({\bf g}(X(\gamma(t),\dot\gamma(t))\right)=\nabla_{\dot\gamma}({\bf g}(X,\dot\gamma))={\bf g}(\nabla_{\dot\gamma}X,\dot\gamma)(t)+{\bf g}(X,\nabla_{\dot\gamma}\dot\gamma)(t)\, , \end{equation} for every curve $\gamma$ and the well known property that $X$ is a Killing vector field if and only if ${\bf g}(\nabla_YX,Y)=0$ for every $Y\in\mbox{\fr X}(M)$ as can be seen from the following chain of identities: \begin{eqnarray*} {\bf g}(\nabla_YX,Y)&=&{\bf g}(\nabla_XY+\mathcal{L}_YX,Y)={\bf g}(\nabla_XY,Y) -{\bf g}(\mathcal{L}_XY,Y)\\ &=&\frac{1}{2}\mathcal{L}_X({\bf g}(Y,Y))-\left(\frac{1}{2}\mathcal{L}_X({\bf g}(Y,Y))-\frac{1}{2}(\mathcal{L}_X{\bf g})(Y,Y)\right)\\ &=&\frac{1}{2}(\mathcal{L}_X{\bf g})(Y,Y)\, , \end{eqnarray*} obtained using the properties of the Levi--Civita connection. \medskip \noindent{\bf Comment}: Which is the origin of the condition ${\bf g}(X,{\rm F})=0$ in item 2? Consider the case where the Newtonian system is conservative, then ${\rm F}=-{\rm grad}\, V$, hence the above orthogonality condition is equivalent to say that $\mathcal{L}_X V=0$, and this last condition, together with $X$ being Killing, implies that the Lagrangian $L=T-V$ is invariant by $X$, then the above result is a kind of generalization of the classical Noether theorem for Lagrangian systems. \subsection{Ideas on control of mechanical systems} Control of mechanical systems and robotics is a broad field of study both from the theoretical and the applied viewpoint. The geometric approach has a wide development at least from the eighties in the last century. Lagrangian, Hamiltonian and Riemannian approaches are used depending on the taste of the different authors and the closeness to the applications. In particular, Riemannian treatment is specially used in robotics. The control systems are modelled as a Newtonian system $(M,{\bf g},{\rm F})$ together with a set of control vector fields, or input forces, $F^1,\ldots,F^k$, we can modulate with some coefficients. The control equation is given as: $$ \nabla_{\dot\gamma}{\dot\gamma}={\rm F}\circ\gamma+\sum_{i=1}^k u_i F^i\,. $$ The coefficients $u_i$ are the input controls and, usually, are functions of time belonging to a specific set of functions and taking values in a specific domain $(u_1,\ldots,u_k)\in U\subset\mathbb{R}^k$. The full development of this ideas applied to control of mechanical systems can be seen in \cite{BULE2005}. It contains not only the state of the art and the work developed by the authors but a detailed bibliography of other authors work. Applications in robotics and other fields, and recent developments, can be seen in the included references and in the web pages of the authors. In particular the importance of the so called {\bf symmetric product}, associated to the Levi--Civita connection, in the study of controllability of Newtonian control systems. The controllability of the above control mechanical system depends on properties of the symmetric algebra generated by the products $\ll F^i,F^j\gg=\nabla_{F^i}F^j+\nabla_{F^j}F^i$ of the control vector fields. Furthermore, it contains a complete description of the motion of a rigid body in this Riemannian approach. In reference \cite{CORTES2002}, Ph. D. dissertation of the author, and in the subsequent research, there are several studies and applications in different specific fields of control of mechanical systems, in particular from the Riemannian point of view. In his personal web page there is a complete account of his developments and collaborators. Optimal control of Newtonian systems is another topic where this Riemannian approach gives specific insight. Once again, the web pages of the above authors, in particular A. D. Lewis, are a good source of information on this topic. See also \cite{mbl-mcml-2009, CMZ-10}, and references therein, for an study of problems of optimal control in mechanical systems and the existence of special kind of solutions. Another significant author working on this topic is Arjan van der Schaft whose web page gives account of its developments with collaborators in the subject of control of mechanical and electromechanical systems. The use of different geometrical tools is permanent including the Riemannian approach. \subsection{Other developments} There are other interesting topics not included in this manuscript, for example forces depending on higher order derivatives, as elastic forces, used in mechanical models of continuous media. Sometimes these models include constraints depending also on second or third order derivatives. In \cite{Neimark-Fufaev}, a classical reference in nonholonomic systems, there are several examples of this type of systems. The approach in this book is classical and there will be interesting to develop some of the chapters in a more geometric terms, in particular in a Riemannian setting. See \cite{CILM-2004} for an interesting an pioonnering geometrical approach. Moreover, not any reference is made in this manuscript to numerical methods of integration of dynamical equations, including those called ``geometric integrators" whose origin is in some geometric formulations of classical mechanics. \section{Conclusions} We have developed the analytical mechanics from a Riemannian geometry perspective including systems with constraints both holonomic and nonholonomic and non--autonomous systems. As specific topics with clear insight with our viewpoint, apart from the general theory, we include Hamilton--Jacobi vector fields and equation, the Jacobi metric and a specific Noether theorem for infinitesimal symmetries of the system. As applications we give the Euler equation for fluids as a Newtonian conservative system and one relation between specific solutions to the Hamilton--Jacobi equation an Schr\"odinger equation. To finish we include some comments on control and optimal control of mechanical systems from a Riemannian perspective. There are more topics, and authors working on them, where the Riemannian approach gives a special insight and sure they will continue giving new results an applications in different fields. \section*{Acknowledgements} \hspace{6 mm}Some ideas of this work come from a course given along the years in our Faculty of Mathematics and Statistics at the UPC for the students of the Mathematics degree. Both the course and this document would not have been possible without the permanent collaboration and friendship of my colleagues Narciso Rom\'an--Roy and Xavier Gr\`acia along the years. My deep thanks to them. My former Ph.\,D. students Javier Yaniz--Fern\'andez and Mar\'ia Barbero--Li\~n\'an worked on control and optimal control of mechanical systems with this Riemannian approach. To both my reconnaissance for their dedication. The author also acknowledges the financial support from the Spanish Ministerio de Ciencia, Innovaci\'on y Universidades project PGC2018-098265-B-C33 and to the Secretary of University and Research of the Ministry of Business and Knowledge of the Catalan Government project 2017--SGR--932. I also thank the referee for his careful reading of the manuscript and his comments that have allowed the author to improve some parts of the work.

https://arxiv.org/abs/1203.0690

Asymptotic normality of integer compositions inside a rectangle

Among all restricted integer compositions with at most $m$ parts, each of which has size at most $l$, choose one uniformly at random. Which integer does this composition represent? In the current note, we show that underlying distribution is, for large $m$ and $l$, approximately normal with mean value $\frac{ml}{2}$.

\section{Introduction} An integer composition of a nonnegative integer $n$ is, informally, a way of writing $n$ as a sum of nonnegative integers $\pi_1,\dotsc,\pi_k$, for some $k\ge 0$. Let $h_{l,m}(n)$ denote the number of integer compositions of the nonnegative integer $n$ with at most $m$ parts, each of which has size at most $l$ (`compositions inside a rectangle'). Recently, Sagan (2009) \cite{sagan} has shown that the sequence \begin{align*} h_{l,m}:= \bigl(\,h_{l,m}(0),\dotsc,h_{l,m}(lm)\,\bigr), \end{align*} is unimodal. In Figure \ref{fig:hlm}, we plot this sequence for $l=2$, $m=5$; $l=6$, $m=5$; and $l=6$, $m=20$. \begin{figure*}[!ht] \centering \includegraphics[scale=0.12]{hlm2_5.jpg} \includegraphics[scale=0.12]{hlm6_5.jpg} \includegraphics[scale=0.12]{hlm6_20.jpg} \caption { The sequences $h_{l,m}(0),\dotsc,h_{l,m}(lm)$ for $l=2$, $m=5$ (left), $l=6$, $m=5$ (middle) and $l=6$, $m=20$ (right). } \label{fig:hlm} \end{figure*} Apparently, as $l$ and $m$ increase, $h_{l,m}$ looks more and more `Gaussian'. This suggests a probabilistic interpretation of $h_{l,m}(n)$, according to which the normalized values $\frac{h_{l,m}(n)}{\sum_{i=0}^{lm}h_{l,m}(i)}$, $n=0,\dotsc,lm$, denote the probabilities that a uniform randomly chosen integer composition with at most $m$ parts, each of which has size at most $l$, represents the integer $n$. In the current note, we show that these probabilities follow, for large $l$ and $m$, approximately a normal distribution with mean value $\frac{lm}{2}$ and variance $m\frac{(l+1)^2-1}{12}$. Thereby, we first define \emph{multinomial triangles} as a generalization of Pascal's triangle and characterize their entries, \emph{polynomial coefficients}, as generalizations of the well-studied binomial coefficients (Section \ref{sec:triangles}), whereupon we outline a recently found relationship between polynomial coefficients and specificially restricted integer compositions (Section \ref{sec:integer}). The latter, with various types of restrictions, have attracted much attention in recent years (cf. \cite{chinn}, \cite{eger}, \cite{heubach}, \cite{hitczenko}, \cite{malandro}, \cite{schmutz}, \cite{shapcott}). For example, Malandro \cite{malandro} determines asymptotic formulas for $L$-restricted integer compositions --- $L$ being an arbitrary finite set --- and Shapcott \cite{shapcott} and Schmutz and Shapcott \cite{schmutz} find a lognormal distribution for part products of restricted integer compositions. Hitczenko and Stengle \cite{hitcz2} derive the expected number of distinct part sizes of unrestricted random compositions. Restricted and unrestricted integer compositions have a variety of applications, ranging from the theory of patterns \cite{heubach2} to monotone paths in two-dimensional lattices (\cite{kimberling}), alignments between strings (\cite{eger4}), and the distribution of the sum of discrete integer-valued random variables (\cite{eger2}). Then, in Section \ref{sec:main}, we state our main theorem, asymptotic normality of compositions inside a rectangle, which we prove in Section \ref{sec:proof}. In the conclusion, we discuss generalizations of the analyzed setting where part sizes are restricted to lie within arbitrary finite sets. While our main result, perceived rightly, might be considered not very surprising, the steps that lead to it (Lemmas \ref{lemma:exact} to \ref{lemma:approx}) may be judged interesting on their own (and are certainly novel) because they specify the exact distribution of the random variable $X_{l,m}$ that sums the parts of a randomly chosen integer composition from a rectangle of size $l\times m$, and give an elegant characterization of it in terms of the distribution of the sum of independent uniform random variables and an ``error term'' that quadratically tends toward zero. \section{Multinomial triangles and polynomial coefficients}\label{sec:triangles} In generalization to binomial triangles, $(l+1)$-nomial triangles, $l\ge 0$, are defined in the following way. Starting with a $1$ in row zero, construct an entry in row $k$, $k\ge 1$, by adding the overlying $(l+1)$ entries in row $(k-1)$ (some of these entries are taken as zero if not defined); thereby, row $k$ has $(kl+1)$ entries. For example, the monomial ($l=0$), binomial ($l=1$), trinomial ($l=2$) and quadrinomial triangles ($l=3$) start as follows, \begin{table}[!h] \begin{tabular}{r} 1\\ 1\\ 1\\ 1 \end{tabular}\hspace{0.5cm} \begin{tabular}{rrrr} 1\\ 1 & 1\\ 1 & 2 & 1\\ 1 & 3 & 3 & 1\\ \end{tabular}\hspace{0.5cm} \begin{tabular}{rrrrrrr} 1\\ 1 & 1 & 1\\ 1 & 2 & 3 & 2 & 1\\ 1 & 3 & 6 & 7 & 6 & 3 & 1\\ \end{tabular}\hspace{0.5cm} \begin{tabular}{rrrrrrrrrr} 1\\ 1 & 1 & 1 & 1\\ 1 & 2 & 3 & 4 & 3 & 2 & 1\\ 1 & 3 & 6 & 10 & 12 & 12 & 10 & 6 & 3 & 1\\ \end{tabular} \end{table} In the $(l+1)$-nomial triangle, entry $n$, $0\le n\le kl$, in row $k$, which we denote by $\binom{k}{n}_{l+1}$ and refer to as \emph{polynomial coefficient} (cf. Caiado (2007) \cite{caiado}, Comtet (1974) \cite{comtet}), has the following interpretation. It is the coefficient of $x^n$ in the expansion of \begin{align}\label{eq:multinomial_coeff1} (1+x+x^2+\dotsc+x^l)^k = \sum_{n=0}^{kl} \binom{k}{n}_{l+1}x^n. \end{align} Also note that, by its definition, $\binom{k}{n}_{l+1}$ satisfies the following recursion \begin{align} \binom{k}{n}_{l+1} = \sum_{j=0}^l \binom{k-1}{n-j}_{l+1}. \end{align} \section{Integer compositions and polynomial coefficients}\label{sec:integer} An {\bf integer composition} of a nonnegative integer $n$ is a tuple $\pi=(\pi_1,\dotsc,\pi_k)$, $k\ge 0$, of nonnegative integers such that \begin{align*} n = \pi_1+\dotsc+\pi_k \end{align*} where the $\pi_i$'s are called \emph{parts}, and $k$ is the \emph{number of parts}.\footnote{Compositions where some parts are allowed to be zero are sometimes called \emph{weak compositions}.} Let $\struct{C}(n,k,a,b)$ denote the set of restricted compositions of $n$ into $k$ parts $\pi_i$ with $a\le \pi_i\le b$, where $a,b\in\mathbb{N}\cup\set{\infty}$ such that $0\le a\le b$, and let $c(n,k,a,b)$ denote its size, $c(n,k,a,b)=\length{\struct{C}(n,k,a,b)}$. For example, for $n=5$, $k=2$, $a=0$, $b=\infty$, we have \begin{align*} 5 = 5+0=0+5=4+1=1+4=3+2=2+3, \end{align*} and thus $c(5,2,0,\infty)=6$. The following results are well-known. \begin{align} c(n,k,0,\infty) &= \binom{n+k-1}{k-1}\label{eq:1}\\ c(n,k,1,\infty) &= \binom{n-1}{k-1}\label{eq:2}\\ c(n,k,a,\infty) &= c(n-ka,k,0,\infty) = \binom{n-ka+k-1}{k-1}\label{eq:3}. \end{align} Moreover, in recent work, Eger (2012) \cite{eger} has shown, more generally, a simple relationship between the number of restricted integer compositions and polynomial coefficients, namely, \begin{align}\label{eq:eger} c(n,k,a,b) &= \binom{k}{n-ka}_{b-a+1}. \end{align} \section{Main theorem}\label{sec:main} Let $m$ be a positive integer and let $l$ be a nonnegative integer. Denote by $h_{l,m}(n)$ the number of integer compositions of the integer $n$ with at most $m$ parts $p$, each of which has size at most $l$, i.e. $0\le p\le l$. Let $X_{l,m}$ be the random variable that takes on the integer $n$, for $0\le n\le lm$, with probability \begin{align*} \frac{h_{l,m}(n)}{\sum_{i=0}^{lm} h_{l,m}(i)}. \end{align*} \begin{theorem}\label{theorem:main} Let $\mu_{l,m}=\frac{ml}{2}$ and let $\sigma_{l,m}^2 = \frac{(l+1)^2-1}{12}$. Then \begin{align*} \frac{X_{l,m}-m\mu_{l,m}}{\sigma_{l,m} \sqrt{m}}\goesto \struct{N}(0,1) \quad\text{as } l,m\goesto\infty. \end{align*} \end{theorem} Our strategy for proving Theorem \ref{theorem:main} is as follows. First, we determine the exact distribution of $X_{l,m}$ in Lemma \ref{lemma:exact}. Then we derive the exact distribution of the sum of $m$ independently and uniformly distributed random variables in Lemma \ref{lemma:iid}, which is, by the Central Limit Theorem, asymptotically a normal distribution. Next, Lemmas \ref{lemma:central} and \ref{lemma:asymptotic} provide inequalities and upper bounds that we require in Lemma \ref{lemma:approx}, where we show that the distribution of $X_{l,m}$ can be represented, roughly, as the sum of two parts: the distribution of the sum $S_1+\dotsc+S_m$ of $m$ independently distributed uniform random variables (derived in Lemma \ref{lemma:iid}) and an ``error term'' that converges quadratically toward zero in $l$. \section{Proof of the main theorem}\label{sec:proof} \begin{lemma}\label{lemma:exact} Let $i$, $1\le i\le m$, be the smallest index such that $n\le il$. Then, \begin{align*} P[X_{l,m}=n] = \frac{1}{(l+1)^m-1}\frac{l}{l+1}\sum_{j=i}^m\binom{j}{n}_{l+1}. \end{align*} \end{lemma} \begin{proof} By definition, $h_{l,m}(n)=\sum_{j=1}^m c(n,j,0,l)=\sum_{j=1}^m\binom{j}{n}_{l+1}$, where the last equality follows from \eqref{eq:eger}. Moreover, $c(n,j,0,l)$ is obviously zero when $j<i$ since $n>(i-1)l$. Finally, the number of integers representable by $j$ parts, each between $0$ and $l$, is obviously $(l+1)^j$. Therefore, \begin{align*} \sum_{i=0}^{lm}h_{l,m}(i) =\sum_{i=0}^{lm} \sum_{j=1}^m c(i,j,0,l) = \sum_{j=1}^m\sum_{i=0}^{lm}c(i,j,0,l) = \sum_{j=1}^m(l+1)^j = \frac{l+1}{l}{\bigl((l+1)^m-1\bigr)}. \end{align*} Hence, \begin{align*} P[X_{l,m}=n]=\frac{h_{l,m}(n)}{\sum_{i=0}^{lm} h_{l,m}(i)}= \frac{1}{(l+1)^m-1}\frac{l}{l+1}\sum_{j=i}^m\binom{j}{n}_{l+1}. \end{align*} \end{proof} \begin{lemma}\label{lemma:iid} Denote by $S^{(m)}_l$ the sum $S_1+\dotsc+S_m$ of independent uniform random variables $S_j$, $j=1,\dotsc,m$, each taking values from the set $\set{0,\dotsc,l}$. The distribution of $S^{(m)}_l$ is given by \begin{align*} P[S^{(m)}_l=n] = \Bigl(\frac{1}{l+1}\Bigr)^m\binom{m}{n}_{l+1}. \end{align*} \end{lemma} \begin{proof} See Caiado \cite{caiado}, Eger \cite{eger2}. \end{proof} \begin{remark} Note that the expected value and the variance of $S_l^{(m)}$ in Lemma \ref{lemma:iid} are given by \begin{align*} \Exp[S^{(m)}_l] = m\Exp[S_j] = \frac{ml}{2},\quad\quad\Var[S^{(m)}_l] = m\Var[S_j] = m\frac{(l+1)^2-1}{12}. \end{align*} Also note that, by the Central Limit Theorem, the distribution of $S_l^{(m)}$ is asymptotically normal. \end{remark} Now, we prove a fact well-known for binomial coefficients, namely, that the `central' coefficient majorizes the remaining coefficients in a given row in the (multinomial) triangle. \begin{lemma}\label{lemma:central} Let $k\ge 0$ and $l\ge 0$ be integers. For all integers $n$ such that $0\le n\le kl$, \begin{align*} \binom{k}{n}_{l+1} \le \binom{k}{\lfloor\frac{kl}{2}\rfloor}_{l+1}. \end{align*} \end{lemma} \begin{proof} By the representation of $\binom{k}{n}_{l+1}$ as $\binom{k}{n}_{l+1}=\sum_{j=0}^{l}\binom{k-1}{n-j}_{l+1}$ we find for $n\ge 1$ \begin{align}\label{eq:repr} \binom{k}{n}_{l+1} = \binom{k}{n-1}_{l+1}+\Bigl[\binom{k-1}{n}_{l+1}-\binom{k-1}{n-l-1}_{l+1}\Bigr]. \end{align} Moreover, it is easy to show that polynomial coefficients are symmetric in the following sense, \begin{align*} \binom{k}{n}_{l+1} = \binom{k}{kl-n}_{l+1}. \end{align*} Therefore it suffices to show that the sequence $\binom{k}{0}_{l+1},\binom{k}{1}_{l+1},\dotsc,\binom{k}{\lfloor\frac{kl}{2}\rfloor}_{l+1}$ is non-decreasing. But by \eqref{eq:repr} this easily follows inductively, using the row number $k$ as induction variable. Importantly, note that, in \eqref{eq:repr}, if $n\le \lfloor\frac{kl}{2}\rfloor$, then $\binom{k-1}{n}_{l+1}$ is defined and greater than zero for all $k\ge 2$ since then $n\le \lfloor\frac{kl}{2}\rfloor\le (k-1)l$. \end{proof} In the following lemma, we write $a_k\sim b_k$ as a short-hand for $\lim_{k\goesto\infty} \frac{a_k}{b_k}=1$. Also note that the following lemma is a generalization of Stirling's approximation to the central binomial coefficient. \begin{lemma}\label{lemma:asymptotic} For all fixed $l$, \begin{align*} \binom{k}{\lfloor\frac{kl}{2}\rfloor}_{l+1}\sim \frac{(l+1)^{k}}{\sqrt{2\pi k\frac{(l+1)^2-1}{12}}}. \end{align*} \end{lemma} \begin{proof} See Eger \cite{eger3}. \end{proof} \begin{lemma}\label{lemma:approx} For all $l$ and $m$ and for all $n$ such that $0\le n\le ml$, \begin{align*} P[X_{l,m}=n] = \gamma_{l,m}P[S^{(m)}_l=n]+e_{l,m}, \end{align*} where $e_{l,m}$ is an ``error term'' that satisfies \begin{align*} 0\le e_{l,m} \le O(l^{-2}) \end{align*} and $\gamma_{l,m}$ satisfies \begin{align*} \gamma_{l,m}=\bigl(1+O(\inv{l})\bigr)^{-1}. \end{align*} \end{lemma} \begin{proof} Let $i$, $1\le i\le m$, be the smallest index such that $n\le il$. Moreover, define $\alpha_{l,m}$ as $\alpha_{l,m}=\frac{1}{(l+1)^m-1}\frac{l}{l+1}$ and note that $\alpha_{l,m}=\gamma_{l,m}\frac{1}{(l+1)^m}$, where $\gamma_{l,m}=\inv{(1+1/l)}$ (ignoring the $(-1)$ in the denominator of $\alpha_{l,m}$). Then \begin{align*} P[X_{l,m}=n] &= \alpha_{l,m}\sum_{j=i}^m \binom{j}{n}_{l+1} = \alpha_{l,m}\binom{m}{n}_{l+1}+\alpha_{l,m}\sum_{j=i}^{m-1}\binom{j}{n}_{l+1} = \gamma_{l,m}P[S^{(m)}_l=n]+e_{l,m}, \end{align*} where we define $e_{l,m}=\alpha_{l,m}\sum_{j=i}^{m-1}\binom{j}{n}_{l+1}$. Obviously, $e_{l,m}\ge 0$. Moreover, by Lemmas \ref{lemma:central} and \ref{lemma:asymptotic} \begin{align}\label{eq:start} e_{l,m}&\le \alpha_{l,m}\sum_{j=i}^{m-1}\binom{j}{\lfloor\frac{jl}{2}\rfloor}_{l+1} \le \alpha_{l,m}O(1)\sum_{j=i}^{m-1}\frac{(l+1)^j}{\sqrt{2\pi j\frac{(l+1)^2-1}{12}}}. \end{align} Now, \begin{align*} \frac{(l+1)^j}{\sqrt{2\pi j\frac{(l+1)^2-1}{12}}} = O(1)\cdot\frac{(l+1)^{j}}{\sqrt{j\bigl((l+1)^2-1\bigr)}}, \end{align*} so that \begin{align*} \sum_{j=i}^{m-1}\frac{(l+1)^j}{\sqrt{2\pi j\frac{(l+1)^2-1}{12}}} = \sum_{j=i}^{m-1}O(1)\frac{(l+1)^{j}}{\sqrt{j}\sqrt{(l+1)^2-1}} \le O(1)\sum_{j=i}^{m-1}(l+1)^{j-1} = O(1)\frac{(l+1)^{i-1}\bigl[(l+1)^{m-i}-1\bigr]}{l}, \end{align*} whence, continuing from \eqref{eq:start}, \begin{equation}\label{eq:shape} \begin{split} e_{l,m}&\le \alpha_{l,m}O(1)\sum_{j=i}^{m-1}\frac{(l+1)^j}{\sqrt{2\pi j\frac{(l+1)^2-1}{12}}} \le O(1)\frac{(l+1)^{i-2}}{(l+1)^m-1}\bigl[(l+1)^{m-i}-1\bigr] \\&\le O(1)\Bigl((l+1)^{-2}-(l+1)^{i-m-2}\Bigr) \le O(1)(l+1)^{-2}. \end{split} \end{equation} \end{proof} In Table \ref{table:approx}, we show the decrease of $e_{l,m}$ in Lemma \ref{lemma:approx} as $l$ increases. Obviously, our bound is apparently quite well, as in fact $e_{l,m}$ seems to approximately quadratically decay in $l$. In Figure \ref{fig:xlm_slm}, the distributions of $X_{l,m}$ and $S_{l}^{(m)}$ for different values of $l$ and $m$ are plotted. The variable $X_{l,m}$ has a particular distributional shape that can be inferred from the proof of Lemma \ref{lemma:approx}. For small values $n$ the distribution of $X_{l,m}$ tends to be larger than that of $S_l^{(m)}$ --- $e_{l,m}$ is relatively larger as can be seen from Equation \eqref{eq:shape} --- while this relation is reversed for large $n$. \begin{table}[!h] \centering \begin{tabular}{l||rr|rr} & \multicolumn{2}{|c|}{$m=10$} & \multicolumn{2}{|c}{$m=20$} \\\hline\hline $l=1$ & $0.0471$ & & $0.0240$ &\\ $l=2$ & $0.0191$ & $2.46$ & $0.0093$ & $2.57$\\ $l=4$ & $0.0064$ & $2.94$ & $0.0031$ & $2.96$\\ $l=8$ & $0.0019$ & $3.25$ & $9.5016\times 10^{-4}$& $3.30$\\ $l=16$ & $5.5909\times 10^{-4}$ & $3.56$ & $2.6494\times 10^{-4}$& $3.58$\\ $l=32$ & $1.4871\times 10^{-4}$ & $3.75$ & $7.0291\times 10^{-5}$& $3.76$\\ $l=64$ & $3.8399\times 10^{-5}$ & $3.82$ & $1.8126\times 10^{-5}$& $3.87$\\ \end{tabular}\hspace{0.5cm} \caption{ Maximum over absolute differences $\abs{P[X_{l,m}=n]-P[S^{(m)}_l=n]}$, $n=0,\dotsc,lm$, for $m=10$ and $m=20$ and varying $l$. We also specify the factor of decrease in these differences between successive $l$ values. } \label{table:approx} \end{table} \begin{figure*}[!ht] \centering \includegraphics[scale=0.12]{distM10L2.jpg} \includegraphics[scale=0.12]{distM10L4.jpg} \includegraphics[scale=0.12]{distM10L8.jpg} \caption { The distributions of $X_{l,m}$ and $S^{(m)}_l$ for $m=10$ and $l=2$ (left), $l=4$ (middle), and $l=8$ (right). } \label{fig:xlm_slm} \end{figure*} \section{Conclusion} The choice of the restrictions $0\le p\le l$ for parts $p$ of integer compositions has, although illustrating a model case, largely been arbitrary. In fact, similar results as Theorem \ref{theorem:main} would hold for any finite set $L=\set{a_1,\dotsc,a_k}$ as range for part sizes. For $L=\set{a,a+1,\dotsc,b}$, $0\le a\le b$, we find simple closed form solutions of the asymptotic distribution of $X_{L,m}$, where we define $X_{L,m}$ (and other variables such as $S_{L}^{(m)}$) as a generalization of $X_{l,m}$ above with $X_{l,m}=X_{\set{0,\dotsc,l},m}$. For example, in this case, $S^{(m)}_L$ has exact distribution \begin{align*} \Bigl(\frac{1}{b-a+1}\Bigr)^m\binom{m}{n-ma}_{b-a+1}, \end{align*} (cf. Eger (2012) \cite{eger2}) with expected value $\frac{m(a+b)}{2}$ and is, by the Central Limit Theorem, asymptotically normally distributed. Conversely, the distribution of $X_{L,m}$ allows a similar representation as in Lemma \ref{lemma:exact}, as a sum of quantities $\binom{j}{n-ja}_{b-a+1}$ and a normalizing term, from which we can straightforwardly derive a decomposition of $X_{L,m}$ as in Lemma \ref{lemma:approx}, with bounds obtained from Lemmas \ref{lemma:central} and \ref{lemma:asymptotic}. As a final remark, note that our results entail a `Stirling' like formula for $h_{l,m}(n)$. By definition $P[X_{l,m}=n]=\frac{h_{l,m}(n)}{\sum_{i=0}^{lm} h_{l,m}(i)}$, and equating this quantity at its asymptotic mean value $\frac{ml}{2}$ with the corresponding normal density leads to \begin{align*} h_{l,m}(\frac{ml}{2}) \sim \frac{\bigl((l+1)^m-1\bigr)\frac{l+1}{l}}{\sqrt{2\pi m \frac{(l+1)^2-1}{12}}}. \end{align*}

https://arxiv.org/abs/1909.03192

Analytic Solution of the Time-Optimal Control of a Double Integrator from an Arbitrary State to the State-space Origin

This brief note presents known results about the minimum-time control of a double integrator system from an arbitrary initial state to the state-space origin (minimum-time regulation problem, or special problem). The main purpose of this note is didactical. Results are presented in all details and following a step by step procedure.

\section{Introduction} The theory of minimum-time control of linear systems is for the most part maturely established~\cite[115--187]{Pontryagin:1962aa}~\cite[127--158]{Lee1967}\cite[395--426]{athans1966}\cite[83--173]{Bolt1971}\cite[248--249]{Kirk1998}~\cite[127--158]{Lee1967}. Among linear systems, the double integrator is widely studied as it constitutes useful model for many dynamic phenomena encountered in engineering and science~\cite{Rao2001}. This note collects and presents in all details results that are originally found in many textbooks, including for instance~\cite[115--187]{Pontryagin:1962aa},~\cite{athans1966}. Results are here presented and demonstrated in a step by step fashion, that is deemed particularly useful for students. \section{System Dynamics} A one-d.o.f. \emph{double-integrator} system is governed by the following equation \begin{equation} \label{doubleintegrator} \left\lbrace \begin{array}{l} I \ydd \!\left( t \right)= C\!\left( t \right)\\ |C\!\left( t \right)| \le C_{\text{max}} \end{array} \right. , \end{equation} where $I\in \mathcal{R}$ is the inertia parameter, $y\!\left( t \right)\in\mathcal{R}$ is the displacement, $C\!\left( t \right)\in\mathcal{R}$ is the control and $C_{\text{max}} $ is the maximum magnitude of the control. Without loosing generality, the equations above can be rewritten in a convenient scaled form. In particular, by transforming the displacement variable to the new variable $x\!\left( t \right)\in\mathcal{R}$ and the control variable to the new variable $u \!\left( t \right)\in\mathcal{R}$ as follows \begin{equation} x\!\left( t \right) = \displaystyle\frac{I}{C_{\text{max}}} \, y \!\left( t \right), \qquad u\!\left( t \right) =\displaystyle \frac{C\!\left( t \right)}{C_{\text{max}}}, \end{equation} the system of Eq.~\ref{doubleintegrator} can be equivalently written as \begin{equation} \label{doubleintegrator2} \left\lbrace \begin{array}{l} \xdd \!\left( t \right)= u\!\left( t \right)\\ |u\!\left( t \right)| \le 1 \end{array} \right. , \end{equation} or, in state-space form, as \begin{equation} \label{DI_statespace} \left\lbrace \begin{array}{l} \xsd= {\bf A} \xs+{\bf B} u(t)\\ |u\!\left( t \right)| \le 1 \end{array} \right. , \end{equation} where \begin{equation} \label{DI_AB_matrix} {{\bf x} \!\left( t \right)}=\left\lbrack \begin{array}{c} \x \\ \xd \end{array} \right\rbrack; \hskip.1cm {\bf A} = \left\lbrack \begin{array}{cc} 0&1 \\ 0&0 \end{array} \right\rbrack; \hskip.1cm {\bf B} = \left\lbrack \begin{array}{c} 0 \\ 1 \end{array} \right\rbrack. \end{equation} \section{Solution of the Special Problem} This section reports known results regarding the time-optimal control of the double integrator between an arbitrary initial state and the state-space origin. This problem can be referred as \emph{No-rest to rest control to the origin} or as \emph{Special Problem}. \begin{thm}\label{teoremaDINR2R} {\textit {\textbf {(Double Integrator: Solution of the Minimum-time optimal control problem from an arbitrary state to the origin)}}}\\ Assume boundary states \begin{equation} \label{DI_boundaryconditionNR2R} {{\bf x} (0)}=\boldsymbol{\xi}\triangleq\left\lbrack \begin{array}{c}x_0 \\ \dot x_0 \end{array} \right\rbrack; \hskip.5cm {\bf{x}}(T^*)= \bf{0}=\left\lbrack \begin{array}{c}0 \\ 0 \end{array} \right\rbrack. \end{equation} Define \begin{equation} \label{definitionofF} F({\bf x} )=F(\x,\xd)\triangleq \x+ {\rm sgn} \!\left( \xd \right)\frac{ \xd^2}{2}. \end{equation} In particular, it is named \emph{switching curve} the following curve composed of two arcs of semi-parabolas joined at the origin \begin{equation} \label{switchcurve_DI} F({\bf x} )=0. \end{equation} Define, furthermore, \begin{align} \label{DINR2RF0} &F_0\!\!\! &&\triangleq F(\boldsymbol{\xi}) = F(\x_0,\xd_0) =x_0+ {\rm sgn} \!\left( \xd_0 \right) \frac{ \xd_0^2}{2},\\ &\Sigma_0\!\!\! &&\triangleq {\rm sgn} (F_0). \label{DINR2RSigma0def} \end{align} The optimal solution of the problem is as follows (a) if $F_0=0\leftrightarrow\Sigma_0=0$, the optimal control history is \begin{equation} \label{controlonswitchcurve} u^* \!\left( t \right) = u^* = -{\rm sgn} (\xd_0 ), \hskip.4cm \, t \in [0, T^*], \end{equation} with \begin{equation}\label{mintimeDINR2R1} T^*=-u^*\,\xd_0, \end{equation} or equivalently \begin{equation}\label{DINR2RdiseqTstar0} T^*=|\xd_0|. \end{equation} Furthermore, the optimal trajectory is \begin{equation} \label{stateequationevo1} {{\bf x}^* \!\left( t \right)}=\left\lbrack \begin{array}{c} \x^*(t) \\ \dot{x}^*(t) \end{array} \right\rbrack= \left\lbrack \begin{array}{l} \displaystyle u^*\frac{ t^2}{2} + \xd_0 t +x_0\\u^* t + \xd_0 \end{array} \right\rbrack, \quad t\in[0,T^*], \end{equation} which corresponds in the phase-plane to the parabola \begin{equation} \label{parabola1} \left ( \x-{1 \over 2} u^* \xd^2 \right ) =\left ( \x_0-{1 \over 2} u^* \xd^2_0 \right ). \end{equation} \vskip.05cm (b) if $F_0\neq0\leftrightarrow \Sigma_0=\pm1$, the optimal control history is \begin{equation} \label{controloutofswitchcurve} u^* \!\left( t \right) = \left\{ \begin{alignedat}{2} u^*_1 &= -\Sigma_0, \quad &t &\in [0, \Delta_1)\\ u^*_2 &= -u^*_1 = \Sigma_0,\quad &t &\in [\Delta_1,T^*] \end{alignedat} \right., \end{equation} where the duration of each ``bang'' is~\footnote{In the particular case when $(x_0\neq 0,\,\xd_0=0)$ it yields $\Delta_1=\Delta_2=\Lambda_0=\sqrt{|x_0|}$.} \begin{equation} \label{mintimeDINR2R2 \left\lbrace \begin{array}{l} \displaystyle \D_1= \Lambda_0+\Sigma_0\, \xd_0 \\ \displaystyle \D_2=\Lambda_0 \end{array} \right., \end{equation} with \begin{equation} \label{DINR2Rlambdef} \Lambda_0\triangleq\sqrt{\Sigma_0 \,\x_0+\frac{\xd^2_0}{2} }, \end{equation} where \begin{equation}\label{DINR2Rlambineq} \Sigma_0 \,\x_0+{1 \over 2} \xd^2_0 > 0, \quad \forall \,{\bf{x}}(0)\neq{\bf{0}}, \end{equation} and therefore \begin{equation} \label{DINR2Rlambdeflt0} \Lambda_0 \in \mathcal{R}^+, \quad \forall \,{\bf{x}}(0)\neq{\bf{0}}, \end{equation} Moreover, the following inequality yields (which is more restrictive than the inequality implied in Eq.~\ref{DINR2Rlambdeflt0}) \begin{equation}\label{DINR2Rlambineq33} \Lambda_0> |\xd_0|, \quad \quad \forall \,\left(\Sigma_0,\xd_0 \right) :\!\!| \,\, \Sigma_0\, \xd_0 = -|\xd_0| <0, \end{equation} which immediately yields \begin{equation} \D_1>0. \end{equation} It yields, finally, \begin{equation} \label{DINR2RTstarexplicit} T^*= \D_1+ \D_2= 2\Lambda_0+\Sigma_0\, \xd_0, \end{equation} which enjoys the property \begin{equation} \label{DINR2RdiseqTstar} T^*> |\xd_0|. \end{equation} Furthermore, the optimal trajectory is \begin{equation} \label{stateequationevo2} {{\bf x}^* \!\left( t \right)}=\left\lbrack \begin{array}{c} \x^*(t) \\ \dot{x}^*(t) \end{array} \right\rbrack= \left\lbrack \begin{array}{l} \displaystyle u^*\frac{ t^2}{2} + \xd_0 t +x_0\\u^* t + \xd_0 \end{array} \right\rbrack, t\in[0,\Delta_1), \end{equation} \begin{equation} \label{stateequationevo22} {{\bf x}^* \!\left( t \right)}=\left\lbrack \begin{array}{c} \x^*(t) \\ \dot{x}^*(t) \end{array} \right\rbrack= \left\lbrack \begin{array}{l} \displaystyle u^*\frac{ t^2}{2} + \xd_s t +x_s\\u^* t + \xd_s \end{array} \right\rbrack, t\in(\Delta_1,\Delta_2], \end{equation} where \begin{equation} \label{intersectionpoint_DI1} \left\lbrace \begin{array}{l} \displaystyle \x_s={1 \over 2} \left( \x_0 +{1 \over 2} \Sigma_0 \,\xd^2_0 \right) \\ \displaystyle \xd_s= -\Sigma_0 \Lambda_0 \end{array} \right., \end{equation} are the coordinate of the switch point $S$ (see also Fig.~\ref{DI_traj_nr2r}). The optimal trajectory corresponds in the phase-plane to the union of the following two arcs of parabolas, connected at the switch point $S$. The first arc of parabola $p_1$ runs between the initial state and the switch point, and the parabola has equation \begin{equation} \label{parabola1} p_1: \quad \left ( \x-{1 \over 2} u^* \xd^2 \right ) =\left ( \x_0-{1 \over 2} u^* \xd^2_0 \right ). \end{equation} The second arc of parabola $p_2$ runs between the switch point and the final state (origin of the phase-plane) and the parabola has equation \begin{equation} \label{parabola12} p_2: \quad \left ( \x-{1 \over 2} u^* \xd^2 \right ) =\left ( \x_s-{1 \over 2} u^* \xd^2_s \right ). \end{equation} Finally, notably, the time elapsed between any two successive points $A$ and $B$ both on $p_1$ equates the difference in ordinate ($\dot{x}$) between the two points, i.e. (See also Fig.1) \begin{equation} \label{eq30} \Delta t_{BA}|_{p_1}= \frac{\dot{x}_B-\dot{x}_A}{-\Sigma_0}, \end{equation} and analogously for any two successive points $C$ and $D$ both on $p_2$, \begin{equation} \Delta t_{DC}|_{p_2}= \frac{\dot{x}_D-\dot{x}_C}{\Sigma_0}. \end{equation} Finally the time elapsed between any two successive points $B$ on $p_1$ and $C$ on $p_2$ is the sum of the time elapsed between $B$ and the switch point $S$ along $p_1$ and the time elapsed between $S$ and $C$ along $p_2$, i.e. \begin{equation}\label{eq32} \Delta t_{CB}= \Delta t_{SB}|_{p_1}+\Delta t_{CS}|_{p_2}. \end{equation} \end{thm} \begin{figure}[h!] \includegraphics[width=\linewidth]{Figure1.pdf} \caption{Double integrator: sample no-rest to rest (i.e. special problem) optimal state trajectory for an initial state such that $F_0>0$ (initial state: $\boldsymbol \xi=[2,1]'$) \label{DI_traj_nr2r} \end{figure} Notably, since Theorem 1 is valid for an arbitrary initial condition, it gives the feedback optimal control synthesis for regulating to zero a double-integrator system. {\bf Proof of Theorem \ref{teoremaDINR2R}}:\\ The system of Eq. \ref{DI_statespace} is linear time invariant, normal and has real eigenvalues. General theorems valid for this class of systems, guarantee that the time optimal control sequence exists, is unique, and has at most one control switch~\cite[th.6-5, p.399-420]{athans1966}. The detailed demonstration of Theorem 1 is found below by exploiting Pontryagin's principle together with geometric analysis. For a control history and controlled trajectory to be optimal, Pontryagin's principle requires that exists a constant multiplier $\rho^*\ge0$ and a costate $\po \in \mathcal{R}^n, t\in[0,T^*]$ such that the following conditions are safisfied~\cite[p.94 and p.108]{Heinz1}\cite[th.6-4, p.396]{athans1966}\cite[p.18]{Pontryagin:1962aa}: \begin{enumerate} \item{\emph{Non-triviality of the multipliers: } \begin{equation} \left(\rho^*,\po \right)\neq \left(0,{\bf{0}}\right). \end{equation}} \item{\emph{Canonical Equations: }$\xso$ and $\po$ satisfy \begin{equation} \begin{array}{l}\label{problem1canonicalequations} \xsdo= \displaystyle \frac{\partial H}{\partial {\bf{p}}} ={\bf{A}} \xso +{\bf{B}} \u*\\\\ \pdo=\displaystyle- \frac{\partial H}{\partial {\bf{x}}}=- {\bf{A}}' \po, \end{array} \end{equation} with boundary-state conditions as in Eq.~\ref{DI_boundaryconditionNR2R}, and Hamiltonian function given by \begin{equation} \label{eqregulatorhamiltonian} H(\rho,{\bf{p}},{\bf{x}}, {\bf{u}}) =\rho+\p' ({\bf{A}} \xs+{\bf{B}} {\bf{u}}(t)). \end{equation} The second of Eqs.~\ref{problem1canonicalequations} yields \begin{equation} \label{costateevo} \po=\e^{-{\bf{A}}' t}{\bf\,{p^*}}(0). \end{equation} } \item{\emph{Minimum Condition: }The Hamiltonian has an absolute minimum at ${\bf{u}}(t)=\u*$ \begin{equation} \label{eq8} H(\rho^*,{\bf{p}}^*,{\bf{x}}^*, {\bf{u}}^*) \le H(\rho^*,{\bf{p}}^*,{\bf{x}}^*, {\bf{u}}) \quad \forall \, t \in [0, T^*], \end{equation} which requires \begin{equation} \label{regulatoroptimalcontrol} \u*=-{\rm sgn}\!\left ( {\bf{B}}' {\bf{p}^*}(t) \right ), \quad \forall t\in[0,T^*]. \end{equation}} \item{\emph{Transversality Condition: }In case of minimum-time control and fixed boundary states the Hamiltonian is zero at the end-poin~\cite[187]{Kirk1998}. Furthermore, the value of the costate at the initial and final time is free.} \item{\emph{Stationarity of the Hamiltonian: } If conditions 2 and 3 above are satisfied~\cite[p.36]{Geering2007}, by taking into account Eq.\ref{eqregulatorhamiltonian}, it yields \begin{equation} \frac{d H}{dt}=\frac{\partial H}{\partial t}=0 \rightarrow H=constant. \end{equation} Therefore, by considering condition 4, it yields \begin{equation} \label{zerohamiltonian} H(\rho^*,\po,\xso, \u*) = 0, \,\,\, \forall \, t \in [0, T^*]. \end{equation} } \end{enumerate} By taking into account Eqs.~\ref{DI_statespace} and~\ref{DI_AB_matrix}, and Eqs.~\ref{eqregulatorhamiltonian},~\ref{costateevo} and~\ref{regulatoroptimalcontrol}, it yields \begin{equation} \po=\left\lbrack \begin{array}{c} p_x^*(t) \\ p^*_{\dot{x}}(t) \end{array} \right\rbrack=\left\lbrack \begin{array}{c} p^*_{x0} \\ p^*_{{\dot{x}}0} -p^*_{x0}\,t\end{array} \right\rbrack, \end{equation} \begin{equation} \label{hamiltonianforallt} H=\rho^*+p^*_{x0} \dot{x} + \left(p^*_{{\dot{x}}0} -p^*_{x0} t \right) u^*(t), \end{equation} and \begin{equation} \label{regulatoroptimalcontrolexpl} u^*(t) =-{\rm sgn}\!\left ( p^*_{{\dot{x}}0} -p^*_{x0}\,t \right ). \end{equation} Therefore the optimal control is either $+1$ or $-1$ with at most one switch because the argument of the sign function in Eq.~\ref{regulatoroptimalcontrolexpl} is linear in the time variable, and therefore crosses zero at most once. Moreover, the condition of optimality in Eq.~\ref{zerohamiltonian} becomes \begin{equation} \label{zerohamiltonianspecific} H=0,\,\quad t \in [0, T^*]. \end{equation} By considering an interval of time $t\in[0,t]$ during which $u^*$ is constant, by integrating in time Eq.~\ref{DI_statespace}, it yields \begin{equation} \label{stateequationevo} {{\bf x}^* \!\left( t \right)}=\left\lbrack \begin{array}{c} \x(t) \\ \xd(t) \end{array} \right\rbrack= \left\lbrack \begin{array}{l} \displaystyle u^*\frac{ t^2}{2} + \xd_0 t +x_0\\u^* t + \xd_0 \end{array} \right\rbrack, \end{equation} which, by solving the first equation for $t$ and substituting the result into the second equation and by taking into account that $u^*=1/u^*=\pm 1$, yields the following optimal control path on the phase-plane \begin{equation} \label{parabolas} \left ( \x-{1 \over 2} u^* \xd^2 \right ) =\left ( \x_0-{1 \over 2} u^* \xd^2_0 \right ). \end{equation} Eq.~\ref{parabolas} represents a parabola on the phase-plane having $x$ as abscissa-axis coordinate and $\dot{x}$ as ordinate-axis coordinate. The parabola has vertex of coordinates $(\x_0-{1 \over 2} u^* \xd^2_0, 0)$ and axis of symmetry coincident with the abscissa axis $(\xd=0)$. When $u^*=+1$, the parabola has positive concavity toward the positive end of the abscissa. When $u^*=-1$, the parabola has negative concavity toward the positive end of the abscissa. In both cases the parabola is run in a clock-wise fashion by the representative point of the system state on the phase plane. Furthermore, a curve named {\it switching curve} on the phase-plane can be defined with Eq.~\ref{switchcurve_DI}. The switching curve is the union of the two semi-parabolas passing through the phase-plane origin and occupying second and fourth quadrant (see Fig.~\ref{DI_traj_nr2r}). Each semi-parabola is an element of one of the two families of parabolas in Eq.~\ref{parabolas}. Assume first that $F_0=0$, where $F_0$ is defined in Eq.~\ref{DINR2RF0}, i.e. assume that the initial state belongs to the switching curve. In this case, the following controlled state path equation on the phase-plane \begin{equation} \label{stateevonoswitch} \x-u \frac{ \xd^2}{2}=0, \end{equation} with the constant control history $u=-{\rm sgn} \!\left( \xd_0 \right)$, i.e., in other words, the path equation $F({\bf x} )=0$, satisfies the state equation. In fact, it satisfies Eqs.~\ref{stateequationevo} or the equivalent Eq.~\ref{parabolas}. Furthermore, Eq.~\ref{stateevonoswitch} satisfies the boundary conditions in Eq.~\ref{DI_boundaryconditionNR2R}. It satisfies the initial condition because, by hypothesis, $F_0=0$, and the final condition since Eq.~\ref{stateevonoswitch} becomes an identity when ${{\bf x}}={\bf{0}}$. Finally, by considering the following initial values for the costate \begin{equation}~\label{initialcostatecasef00} p^*_{{\dot{x}}0} ={\rm sgn} \!\left( \xd_0 \right)\,\rho^*, \qquad p^*_{x0}=0, \qquad \forall \rho^*, \end{equation} the zero-value condition for the Hamiltonian in Eq.~\ref{zerohamiltonianspecific} holds true, as it can be immediately verified by considering the two possible cases of ${\rm sgn} \!\left( \xd_0 \right)=\pm1$. By substituting the values of initial costate variables in Eq.~\ref{initialcostatecasef00} into the argument of the sign in Eq.~\ref{regulatoroptimalcontrolexpl}, it remains confirmed that there is no switch of the control during this optimal maneuver. \newline By observing the phase-plane geometry, it is immediate to discern that the optimal state trajectory is an arc of one of the two parabolas belonging to the switching curve. In particular it is the arc between the point $(\x_0,\xd_0)$ and the origin. Therefore, the optimal control history is as reported in Eq.~\ref{controlonswitchcurve}, with $T^*$ given by Eqs.~\ref{mintimeDINR2R1}, which is obtained by substituting the value of $\dot x(T^*)$ from Eq.~\ref{DI_boundaryconditionNR2R} into the second of Eqs.~\ref{stateequationevo}. Assume now that $F_0\neq 0$. By observing the phase-plane geometry and considering that the control can only be ``bang-bang'' and there can only be one switch, it is immediate to discern that the optimal state trajectory is the union of two arcs of parabolas. The first arc of parabola goes from the point $(\x_0,\xd_0)$ to the point $S$ where the switching curve is intersected, the second one, which belongs to the switching curve, goes from the point $S$ to the origin (see Fig. \ref{DI_traj_nr2r} ). If $F>0$, i.e. if the initial state is above the switching curve, the first control ``bang'' is $-1$, the second one is $+1$ and the switch point is on the arc of the switching curve in the fourth quadrant of the phase-plane. Viceversa, if $F>0$, i.e. if the initial state is below the switching curve, the first control ``bang'' is $+1$, the second one is $-1$ and the switch point is on the arc of the switching curve in the second quadrant of the phase-plane.Therefore, the optimal control history is as reported in Eq.~\ref{controloutofswitchcurve}. The switch point $S$ has coordinates \begin{equation} \label{intersectionpoint_DI} \left\lbrace \begin{array}{l} \displaystyle \x_s={1 \over 2} \left( \x_0 +{1 \over 2} \Sigma_0 \,\xd^2_0 \right) \\ \displaystyle \xd_s= -\Sigma_0 \Lambda_0 \end{array} \right., \end{equation} as obtained by solving the algebraic system of Eq.~\ref{parabolas} and Eq.~\ref{switchcurve_DI}, after substituting $x$ with $x_s$, $\dot{x}$ with $\dot{x}_s$, $u^*$ with $u^*_1$ from Eq.~\ref{controloutofswitchcurve}, and ${\rm sgn} \!\left( \dot{x}_s \right)$ with $-\Sigma_0$. This latter expression is also used to make explicit the discrimination between $+1$ and $-1$ in the equation $\xd_s=\pm \Lambda_0$, resulting from solving the algebraic system. The duration of the control ``bangs'' is as in Eq.~\ref{mintimeDINR2R2}. In particular, the expression of $\D_1$ is obtained from the second of Eqs.~\ref{stateequationevo} by substituting $t$ with the symbol $\D_1$, $\xd \!\left( t \right)$ with the value of $\xd_s$ from Eq.~\ref{intersectionpoint_DI} and $u^*$ with the value of $u^*_1$ from Eq.~\ref{controloutofswitchcurve}. Finally, the expression of $\D_2$ is obtained from the same equation by substituting $t$ with the symbol $\D_2$, $\xd \!\left( t \right)$ with zero, $u^*$ with the value of $u^*_2$ from Eq.~\ref{controloutofswitchcurve} and $\xd_0$ with the value of $\xd_s$ from Eq.~\ref{intersectionpoint_DI}. Furthermore, the properties in Eqs.~\ref{DINR2Rlambineq},~\ref{DINR2Rlambineq33}, and \ref{DINR2RdiseqTstar} are proven by exhaustive analysis of validity for all possible sign combinations. See Tables~\ref{tableanalysisDINR2R1},~\ref{tableanalysislambdaeq2} and~\ref{tableanalysisDINR2R2}, respectively. Finally, by considering the following initial values for the costate \begin{equation}~\label{initialcostatecasef00} \displaystyle p^*_{{\dot{x}}0} = p^*_{x0}\,\Delta_1, \qquad p^*_{x0}=\frac{\rho^*}{\Lambda_0\Sigma_0}\, \qquad \forall \rho^*, \end{equation} the zero-value condition for the Hamiltonian in Eq.~\ref{zerohamiltonianspecific} holds true, as it can be immediately verified. The values in Eq.~\ref{initialcostatecasef00} are obtained by considering the algebraic system formed by the equation obtained by imposing to zero the argument of the sign function in Eq.~\ref{regulatoroptimalcontrolexpl} with $t=t_s=\Delta_1$, and by the equation obtained by imposing to zero the Hamiltonian at the initial time (Eq.~\ref{hamiltonianforallt} with $t=0$). Finally, Eqs~\ref{eq30} to ~\ref{eq32} are immediately demonstrated by integrating in time the first of Eq.~\ref{doubleintegrator2} between the points of interest and solving for the time. \begin{flushleft} \hfill \bf Q.e.d. \end{flushleft} \renewcommand{\arraystretch}{1.4} \begin{table}[t!] \setlength{\tabcolsep}{9pt} \begin{tabular}{@{}lllc@{}} \toprule $\boldsymbol{\Sigma_0}$ & $\boldsymbol{\xd_0}$ & \thead{\bf{Eq.~\ref{DINR2Rlambineq} }\\\bf{yields}}& \thead{\bf{Eq.~\ref{DINR2Rlambineq} }\\\bf{is}}\\ \cmidrule(rl){3-3} \midrule \addlinespace 1 & $>0$ &$x_0>-\displaystyle \frac{\xd_0}{2} $ & True*\\ \addlinespace 1 & $<0$ &$x_0>-\displaystyle \frac{\xd_0}{2} $ & True*\\ \addlinespace -1 & $>0$ &$x_0<\displaystyle \frac{\xd_0}{2} $ & True*\\ \addlinespace -1 & $<0$ &$x_0<\displaystyle \frac{\xd_0}{2} $ & True*\\ \bottomrule \end{tabular} \vspace{0.05in} \caption{\emph{Analysis of validity of Eq.~\ref{DINR2Rlambineq} }} \label{tableanalysisDINR2R1} \small{*[Because it agrees with Eq.~\ref{DINR2RF0}]} \end{table} \begin{table}[t!] \setlength{\tabcolsep}{6pt} \begin{tabular}{@{}llllc@{}} \toprule $\boldsymbol{\Sigma_0}$ & $\boldsymbol{\xd_0}$ & \multicolumn{2}{c}{\bf{Eq.~\ref{definitionofF} yields}} & \thead{\bf{Eq.~\ref{DINR2Rlambineq33}}\\\bf{is}}\\ \cmidrule(rl){3-4} \midrule \addlinespace 1 & $<0$ & $\x_0 > \displaystyle \frac{\xd_0}{2}\rightarrow$ &$ \Sigma_0 x_0 >\displaystyle \frac{\xd_0}{2}$ &True* \\ \addlinespace -1 & $>0$ & $\x_0 <- \displaystyle \frac{\xd_0}{2}\rightarrow$ &$ \Sigma_0 x_0 > \displaystyle \frac{\xd_0}{2}$ &True* \\ \bottomrule \end{tabular} \vspace{0.05in} \caption{\emph{Analysis of validity of Eq.~\ref{DINR2Rlambineq33} }} \label{tableanalysislambdaeq2} \small{*[As obtained by considering together Eqs.~\ref{definitionofF} and~\ref{DINR2Rlambdef}]} \end{table} \begin{table}[t!] \begin{tabular}{@{} l l l l l @{}} \toprule $\boldsymbol{\Sigma_0}$ & $\boldsymbol{\xd_0}$ & \multicolumn{2}{c}{\bf{Eqs.~\ref{DINR2RTstarexplicit} and~\ref{DINR2RdiseqTstar} yields}}& \thead{\bf{Eq.~\ref{DINR2RdiseqTstar}}\\\bf{is}}\\ \cmidrule(rl){3-4} \midrule 1 & $>0$ &$2\Lambda_0+\xd_0>\xd_0 \rightarrow$&$\Lambda_0 >0$ & True*\\ \addlinespace 1 & $<0$ &$2\Lambda_0-|\xd_0|>|\xd_0| \rightarrow$&$\Lambda_0 >|\xd_0|$ & True**\\ \addlinespace -1 & $>0$ &$2\Lambda_0-|\xd_0|>|\xd_0| \rightarrow$&$\Lambda_0 >|\xd_0|$ & True**\\ \addlinespace -1 & $<0$ &$2\Lambda_0-\xd_0>-\xd_0\rightarrow$&$ \Lambda_0 >0$ & True*\\ \bottomrule \end{tabular} \vspace{0.05in} \caption{\emph{Analysis of validity of Eq.~\ref{DINR2RdiseqTstar}}} \label{tableanalysisDINR2R2} \small{*[Because it agrees with Eq.~\ref{DINR2Rlambdeflt0}]\\ {**}[Because it agrees with Eq.~\ref{DINR2Rlambineq33}]} \end{table} \bibliographystyle{plain}

https://arxiv.org/abs/1203.4200

Residues and Telescopers for Rational Functions

We give necessary and sufficient conditions for the existence of telescopers for rational functions of two variables in the continuous, discrete and q-discrete settings and characterize which operators can occur as telescopers. Using this latter characterization, we reprove results of Furstenberg and Zeilberger concerning diagonals of power series representing rational functions. The key concept behind these considerations is a generalization of the notion of residue in the continuous case to an analogous concept in the discrete and q-discrete cases.

\section{Introduction}\label{SECT:intro} Residues have played a ubiquitous and important role in mathematics and their use in combinatorics has had a lasting impact~(e.g., \cite{Flajolet_Sedgewick}). In this paper we will show how the notion of residue and its generalizations lead to new results and a recasting of known results concerning telescopers in the continuous, discrete and~$q$-discrete cases. As an introduction to our point of view and our results, let us consider the problem of finding a differential telescoper for a rational function of two variables. Let~$k$ be a field of characteristic zero,~$k(t,x)$ the field of rational functions of two variables and~$D_t = \partial/\partial_t$ and~$D_x=\partial/\partial_x$ the usual derivations with respect to~$t$ and~$x$, respectively. Given~$f \in k(t,x)$, we wish to find a nonzero operator $L \in k(t)\langle D_t\rangle$, the ring of linear differential operators in~$D_t$ with coefficients in $k(t)$, and an element~$g \in k(t,x)$ such that~$L(f) = D_x(g)$. We may consider~$f$ as an element of~$\overline{K}(x)$ where~$\overline{K}$ is the algebraic closure of~$K = k(t)$. As such, we may write \begin{equation} \label{eqn0} f = p + \sum_{i=1}^m \sum_{j=1}^{n_i}\frac{\alpha_{i,j}}{(x-\beta_i)^j}, \end{equation} where $p\in {K}[x]$, the $\beta_i$ are the roots in $\overline{K}$ of the denominator of $f$ and the $\alpha_{i,j}$ are in~$\overline{K}$. Note that the element $\alpha_{i,1}$ is the usual \emph{residue} of~$f$ at~$\beta_i$. Using Hermite reduction (\cite[p.~39]{BronsteinBook} {or Section~\ref{SUBSECT:cres} below}), one sees that a rational function $h \in K(x)$ is of the form $h = D_x(g)$ for some $g \in K(x)$ if and only if all residues of $h$ are zero. Therefore to find a telescoper for $f$ it is enough to find a nonzero operator $L \in K\langle D_t \rangle$ such that $L(f)$ has only zero residues. For example assume that $f$ has only simple poles, i.e., $f = \frac{a}{b}, a,b \in K[x]$, $\deg_xa < \deg_x b$ and $b$ squarefree. We then know that the Rothstein-Trager resultant \cite{Trager1976, Rothstein1977} \[ R := {\rm resultant}_x(a-zD_x(b), b)\in K[z] \] is a polynomial whose roots are the residues at the poles of~$f$. Given a squarefree polynomial in $K[z]=k(t)[z]$, differentiation with respect to $t$ and elimination allow one to construct a nonzero linear differential operator $L \in k(t)\langle D_t \rangle$ such that $L$ annihilates the roots of this polynomial. Applying $L$ to each term of \eqref{eqn0} one sees that $L(f)$ has zero residues at each of its poles. Applying Hermite reduction to $L(f)$ allows us to find a $g$ such that $L(f) = D_x(g)$. The main idea in the method described above is that nonzero residues are the obstruction to being the derivative of a rational function and one constructs a linear operator to remove this obstruction. This idea is the basis of results in \cite{CKS2012} where it is shown that the problem of finding differential telescopers for rational functions in~$m$ variables is equivalent to the problem of finding telescopers for algebraic functions in~$m-1$ variables and where a new algorithm for finding telescopers for algebraic functions in two variables is given. For a precise problem description, let~$k(t,x)$ be as above and~$D_t$ and~$D_x$ be the derivations defined above. We define shift operators~$S_t$ and~$S_x$ as \[S_t(f(t,x)) = f(t+1,x) \quad \text{and} \quad S_x(f(t,x)) = f(t, x+1)\] and~$q$-shift operators (for~$q\in k$ not a root of unity)~$Q_t$ and~$Q_x$ as \[Q_t(f(t,x) = f(qt,x)\quad \text{and} \quad Q_x(f(t,x)) = f(t,qx)).\] Let~$\Delta_x$ and~$\Delta_{q, x}$ denote the difference and~$q$-difference operators~$S_x-1$ and~$Q_x-1$, respectively. In this paper, we give a solution to the following problem \begin{center} \begin{minipage}[t]{10cm} \underline{Existence Problem for Telescopers.}\,\, {\it For any~$\partial_t\in \{D_t, S_t, Q_t\}$ and~$\partial_x\in \{D_x, \Delta_x, \Delta_{q,x}\}$ find necessary and sufficient conditions on elements~$f \in k(t,x)$ that guarantee the existence of a nonzero linear operator~$L(t,\partial_t)$ in~$\partial_t$ with coefficients in~$k(t)$ (a {\emph{telescoper}}) and an element~$g \in k(t,x)$ (a {\emph{certificate}}) such that \[L(t,\partial_t)(f) = \partial_x(g).\]} \end{minipage} \end{center} \vspace{-0.3cm} As we have shown above, when~$\partial_t = D_t$ and~$\partial_x = D_x$, a telescoper and certificate exist for any~$f\in k(t,x)$. This is not necessarily true in the other cases. In the case when~$\partial_t = S_t$ and~$\partial_x = \Delta_x$, Abramov and Le~\cite{AbramovLe2002} showed that there is no telescoper for the rational function~$1/(t^2+x^2)$ and presented a necessary and sufficient condition for the existence of telescopers. Later, Abramov gave a general criterion for the existence of telescopers for hypergeometric terms~\cite{Abramov2003}. The~$q$-analogs were achieved in the works by Le~\cite{Le2001} and by Chen et al. \cite{Chen2005}. Our approach in this paper represents a unified way of solving the Existence Problem for Telescopers (for rational functions) in these and the remaining cases. In particular, we will first identify in each case the appropriate notion of ``residues'' which will be elements of $\overline{k(t)}$, the algebraic closure of~$k(t)$. We will show that for any~$f \in k(t,x)$ and~$\partial_x \in \{D_x, \Delta_x, \Delta_{q,x}\}$, there exists a~$g \in k(t,x)$ such that~$f = \partial_x(g)$ if and only if all the ``residues'' vanish. We will then show that to find a telescoper, it is necessary and sufficient to find an operator~$L(t,\partial_t)$ that annihilates all of the residues. This necessary and sufficient condition has several applications. For example, our results reduce the Existence Problem for Telescopers to the problem of finding necessary and sufficient conditions that guarantee the existence of operators that annihilate algebraic functions and we present a solution to this latter problem. Our approach also gives termination criteria for the Zeilberger method~\cite{Almkvist1990, Zeilberger1990, Zeilberger1991} and also a strategy for finding telescopers and certificates, which has been successfully used in the continuous case in~\cite{CKS2012}. In addition, these criteria together with the results in~\cite{Hardouin2008, Schneider2010} can be used to determine if indefinite sums and integrals satisfy (possibly nonlinear) differential equations (see Example~\ref{EX:transcendental}). The rest of the paper is organized as follows. In Section~\ref{SECT:residues} we define the notions of residues relevant to the discrete and~$q$-discrete cases and show that for any~$f \in k(t,x)$ and~$\partial_x \in \{D_x, \Delta_x, \Delta_{q,x}\}$, there exists a~$g \in k(t,x)$ such that~$f = \partial_x g$ if and only if all the residues vanish. In Section~\ref{SECT:algfuns} we characterize those algebraic functions in~$\overline{k(t)}$ for which there exist annihilating linear operators~$L(t,S_t)$ or $L(t,Q_t)$ as well as prove some ancillary results useful in succeeding sections. In Section~\ref{SECT:telescoper}, we solve the Existence Problem for Telescopers as well as characterize when a linear operator is a telescoper. Using this latter characterization, we can give a proof, using our approach, of the theorem of Furstenberg~\cite{Furstenberg1967} stating that the diagonal of a rational power series in two variables is an algebraic function. We also discuss a recent example of Ekhad and Zeilberger~\cite{EZ2011} in the context of the results of this paper. The final Appendix contains proofs of the characterizations stated in Section~\ref{SECT:algfuns}. \section{Residues}\label{SECT:residues} Let~$K$ be a field of characteristic zero and~$K(x)$ be the field of rational functions in~$x$ over~$K$. Let~$\overline{K}$ denote the algebraic closure of~$K$. Let~$q\in K$ be such that~$q^i\neq 1$ for any nonzero~$i\in \bZ$, i.e., $q$ is not a root of unity. As in the Introduction, we define the derivation~$D_x$, shift operator~$S_x$, and~$q$-shift operator~$Q_x$ on~$K(x)$, respectively, as \[D_x(f(x))=\frac{d(f(x))}{dx}, \quad S_x(f(x))= f(x+1), \quad \text{and}\quad Q_x(f(x))=f(qx) \] for all~$f\in K(x)$. Let~$\Delta_x$ and~$\Delta_{q, x}$ denote the difference and~$q$-difference operators~$S_x-1$ and~$Q_x-1$, respectively. A rational function~$f\in K(x)$ is said to be \emph{rational integrable} (resp.\ \emph{summable, $q$-summable}) in~$K(x)$ if there exists~$g\in K(x)$ such that~$f=D_x(g)$ (resp.\ $f=\Delta_x(g)$, $f = \Delta_{q, x}(g)$). This section is motivated by the well known result (Proposition~\ref{PROP:ratint} below) that characterizes rational integrability in terms of vanishing residues. In the remainder of this section we describe other types of ``residues'' and how they can be used to give necessary and sufficient conditions for summability and~$q$-summability. \subsection{Continuous residues}\label{SUBSECT:cres} Let~$f=a/b\in K(x)$ with~$a, b\in K[x]$ and~$\gcd(a, b)=1$. Then~$f$ can be uniquely written in its partial fraction decomposition \begin{equation}\label{EQ:cparfrac} f = p + \sum_{i=1}^m \sum_{j=1}^{n_i} \frac{\alpha_{i,j}}{(x-\beta_i)^j}, \end{equation} where~$p\in K[x]$, { $m, n_i\in \bN$, $\alpha_{i,j}, \beta_i\in \overline{K}$, and~$\beta_j$'s are roots of~$b$. From any of the usual proofs of partial fraction decompositions, one sees that all the~$\alpha_{i, j}$'s are in~$K(\beta_1, \ldots, \beta_m)$. \begin{define}[Continuous residue]\label{DEF:cres} Let~$f\in K(x)$ be of the form~\eqref{EQ:cparfrac}. The value~$\alpha_{i,1}\in \ok$ is called the \emph{continuous residue} of~$f$ at~$\beta_i$ (with respect to~$x$), denoted by~$\operatorname{cres}_x(f, \beta_i)$. \end{define} \noindent Note that the continuous residue is just the usual residue in complex analysis. We will define other kinds of residues below but when we refer to a residue without further modification, we shall mean the continuous residue. Although the following is well known (see~\cite[Proposition 2.1]{VanDerPut2001}) we include it since this result is the motivation and model for the considerations that follow. \begin{prop}\label{PROP:ratint} Let~$f=a/b\in K(x)$ be such that~$a, b\in K[x]$ and~$\gcd(a, b)=1$. Then~$f$ is rational integrable in~$K(x)$ if and only if the residue~$\operatorname{cres}_x(f, \beta)$ is zero for any root~$\beta\in \ok$ of~$b$. \end{prop} \begin{proof} Suppose that~$f$ is rational integrable in~$K(x)$, i.e.,\ $f= D_x(g)$ for some~$g$ in~$K(x)$. Writing~$g$ in its partial fraction decomposition and differentiating each term, one sees that all the residues of~$D_x(g)$ are~$0$. Conversely, if all residues of~$f$ at its poles are zero, then~$f$ can be written as \[f = p + \sum_{i=1}^m \sum_{j=2}^{n_i} \frac{\alpha_{i,j}}{(x-\beta_i)^j},\] where~$p\in K[x]$, $\alpha_{i,j}, \beta_i\in \overline{K}$, and~$\beta_j$'s are roots of~$b$. Note that any polynomial is rational integrable in~$K(x)$, and for all~$i, j$ with~$1\leq i\leq m$ and~$2\leq j\leq n_i$, \[\frac{\alpha_{i,j}}{(x-\beta_i)^j} = D_x\left(\frac{(1-j)^{-1}\alpha_{i, j}}{(x-\beta_i)^{j-1}}\right).\] Then~$f=D_x(g)$, where~$g$ is of the form \[g = \tilde{p} + \sum_{i=1}^m \sum_{j=2}^{n_i} \frac{(1-j)^{-1}\alpha_{i, j}}{(x-\beta_i)^{j-1}} \quad \text{for some~$\tilde{p}\in K[x]$.}\] For each irreducible factor $p$ of~$b$, the sum in~$g$ is a symmetric function of those~$\beta_i$'s that are roots of~$p$. From this one concludes that $g$ lies in~$K(x)$. Thus,~$f$ is rational integrable in~$K(x)$. \end{proof} \subsection{Discrete residues}\label{SUBSECT:dres} Given a rational function, Matusevich~\cite{Matusevich2000} found a necessary and sufficient condition for its rational summability. Moreover, one can algorithmically decide whether a rational function is rational summable or not using methods in~\cite{Abramov1975, Abramov1989, AbramovYu1991, Abramov1995, Abramov1995b, Paule1995b, Pirastu1995a, Pirastu1995b}. Here, we present a rational summability criterion via a discrete analogue of residues. To this end, we first recall some terminology from~\cite{Abramov1975, Paule1995b} and~\cite[Chapter 2]{vdPutSinger1997}. For an element~$\alpha \in \ok$, we call the subset~$\alpha + \bZ$ the~\emph{$\bZ$-orbit} of~$\alpha$ in~$\ok$, denoted by~$[\alpha]$. For a polynomial~$b\in K[x]\setminus K$, the value \[\max\{i\in \bZ \mid \text{$\exists \, \alpha, \beta \in \ok$ such that~$i=\alpha-\beta$ and~$b(\alpha)=b(\beta)=0$}\}\] is called the~\emph{dispersion} of~$b$ with respect to~$x$, denoted by~$\operatorname{disp}_x(b)$. The polynomial~$b$ is said to be~\emph{shift-free} with respect to~$x$ if~$\operatorname{disp}_x(b)=0$. Let~$f=a/b\in K(x)$ be such that~$a, b\in K[x]$ and~$\gcd(a, b)=1$. Over the field~$\ok$, $f$ can be decomposed into the form \begin{equation}\label{EQ:dparfrac} f = p + \sum_{i=1}^m \sum_{j=1}^{n_i} \sum_{\ell=0}^{d_{i,j}} \frac{\alpha_{i,j,\ell}}{(x-(\beta_i+\ell))^{j}}, \end{equation} where~$p\in K[x]$, $m, n_i, d_{i,j}\in \bN$, $\alpha_{i, j, \ell}, \beta_i\in \ok$, and~$\beta_i$'s are in distinct $\bZ$-orbits. \begin{define}[Discrete residue]\label{DEF:dres} Let~$f\in K(x)$ be of the form~\eqref{EQ:dparfrac}. The sum~$\sum_{\ell=0}^{d_{i,j}}\alpha_{i,j, \ell}$ is called the \emph{discrete residue} of~$f$ at the $\bZ$-orbit $[\beta_i]$ of multiplicity~$j$ (with respect to~$x$), denoted by~$\operatorname{dres}_x(f, [\beta_i], j)$. \end{define} \begin{lemma}\label{LM:dres} Let~$f=\sum_{\ell=0}^d \alpha_{\ell}/(x-(\beta+\ell))^s$ be such that~$d, s\in \bN$ and~$\alpha_{\ell}, \beta\in \ok$. Then~$f$ is rational summable in~$\ok(x)$ if and only if the sum~$\sum_{\ell=0}^d \alpha_{\ell}$ is zero that is, if and only if~$\operatorname{dres}_x(f, [\beta], s)=0$. \end{lemma} \begin{proof} Suppose that the sum~$\sum_{\ell=0}^d \alpha_{\ell}$ is zero. We show that~$f$ is rational summable in~$\ok(x)$. To this end, we proceed by induction on~$d$. In the base case when~$d=0$, $f$ is clearly rational summable in~$\ok(x)$ since~$f=0$. Suppose that the assertion holds for~$d=m$ with~$m\geq 0$. Note that \begin{align*} \frac{\alpha_{m+1}}{(x-(\beta+m+1))^s} = \Delta_x \left(-\frac{\alpha_{m+1}}{(x-(\beta+m+1))^s}\right) + \frac{\alpha_{m+1}}{(x-(\beta+m))^s}. \end{align*} This implies that \[\sum_{\ell=0}^{m+1} \frac{\alpha_{\ell}}{(x-(\beta+\ell))^s} = \Delta_x \left(-\frac{\alpha_{m+1}}{x-(\beta+m+1)^s}\right) + \sum_{\ell=0}^{m} \frac{\tilde{\alpha}_{\ell}}{(x-(\beta+\ell))^s},\] where~$\tilde{\alpha}_{\ell} = \alpha_{\ell}$ if~$0\leq \ell \leq m-1$ and~$\tilde{\alpha}_{m} = \alpha_{m+1} + \alpha_{m}$. By definition, the sum~$\sum_{\ell=0}^m \tilde{\alpha}_{\ell}$ is still zero. The induction hypothesis then implies that there exists~$\tilde{g}\in \ok(x)$ such that \[\sum_{\ell=0}^{m} \frac{\tilde{\alpha}_{\ell}}{(x-(\beta+\ell))^s} = \Delta_x(\tilde{g}).\] So~$f=\Delta_x(g)$ with~$g=\tilde{g} - \alpha_{m+1}/(x-(\beta+m+1))^s\in \ok(x)$. For the opposite implication, we assume to the contrary that the sum~$\sum_{\ell=0}^d \alpha_\ell$ is nonzero. Without loss of generality, we can assume that~$\alpha_0 \neq 0$. Write~$\alpha_0 = \bar{\alpha}_0 + \tilde{\alpha}_0$ such that~$\tilde{\alpha}_0 + \sum_{\ell=1}^{d} \alpha_\ell =0$. Since~$\sum_{\ell=0}^d \alpha_\ell\neq 0$, $\bar{\alpha}_0\neq 0$. By the assertion shown above, there exists~$\tilde{g}\in \ok(x)$ such that \[f = \frac{\bar{\alpha}_0}{(x-\beta)^s} + \Delta_x(\tilde{g}).\] Since~$\operatorname{disp}_x((x-\beta)^s)=0$ and~$\bar{\alpha}_0\neq 0$, ${\bar{\alpha}_0}/{(x-\beta)^s}$ is not rational summable by~\cite[Lemma 3]{Matusevich2000} or~\cite[Lemma 6.3]{Hardouin2008}. Then~$f$ is not rational summable in~$\ok(x)$. This completes the proof. \end{proof} \begin{prop}\label{PROP:ratsum} Let~$f=a/b\in K(x)$ be such that~$a, b\in K[x]$ and~$\gcd(a, b)=1$. Then~$f$ is rational summable in~$K(x)$ if and only if the discrete residue $\operatorname{dres}_x(f, [\beta], j)$ is zero for any $\bZ$-orbit~$[\beta]$ with~$b(\beta)=0$ of any multiplicity~$j\in \bN$. \end{prop} \begin{proof} Let~$f\in K(x)$ be decomposed into the form~\eqref{EQ:dparfrac}. If the discrete residue of~$f$ at any~$\bZ$-orbit of any multiplicity is zero, then Lemma~\ref{LM:dres} implies that for all~$i, j$ with~$1\leq i \leq m$ and~$1\leq j \leq n_i$, the sum \[\sum_{\ell=0}^{d_{i,j}} \frac{\alpha_{i,j,\ell}}{(x-(\beta_i+\ell))^{j}} = \Delta_x(g_{i,j})\quad \text{for some~$g_{i, j}\in \ok(x)$.}\] Since any polynomial is rational summable, there exists~$\tilde{p}\in K[x]$ such that~$p = \Delta_x(\tilde{p})$. So~$f = \Delta_x(\tilde{p} + g)$, where~$g=\sum_{i=1}^m\sum_{j=1}^{n_i} g_{i,j}$. Arguing as in Proposition~\ref{PROP:ratint}, one sees that for each irreducible factor $p$ of~$b$, the sum in~$f$ is a symmetric function of those~$\beta_i$'s that are roots of~$p$. From this one concludes that the sum is in~$K(x)$ and that~$f$ is rational summable in~$K(x)$. Suppose that~$f$ is rational summable in~$K(x)$, i.e., $f=\Delta_x(g)$ for some~$g\in K(x)$. Over the field~$\ok$, we decompose~$g$ into the form~\eqref{EQ:dparfrac}. For all~$i, j$ with~$1\leq i \leq m$ and~$1\leq j \leq n_i$, the linearity of~$\Delta_x$ implies that \[\Delta_x\left(\sum_{\ell=0}^{d_{i,j}} \frac{\alpha_{i,j,\ell}}{(x-(\beta_i+\ell))^{j}} \right) = \sum_{\ell=0}^{d_{i,j}+1} \frac{\tilde{\alpha}_{i,j,\ell}}{(x-(\tilde{\beta}_i +\ell))^{j}}, \] where~$\tilde{\beta}_i = \beta_i -1$, $\tilde{\alpha}_{i,j, 0}={\alpha}_{i, j, 0}$, $\tilde{\alpha}_{i, j, d_{i,j}+1} = -\alpha_{i, j, d_{i, j}}$, and~$\tilde{\alpha}_{i, j, \ell} = \alpha_{i,j, \ell}-\alpha_{i, j, \ell-1}$ for~$1\leq \ell \leq d_{i, j}$. Then the residue~$\operatorname{dres}_x(f, [\tilde{\beta}_i], j)=\sum_{\ell=0}^{d_{i, j}}\tilde{\alpha}_{i,j,\ell}=0$ for all~$i, j$. This completes the proof. \end{proof} \begin{remark} Proposition~\ref{PROP:ratsum} is also known in literature (see~\cite[Theorem 10]{Matusevich2000} or~\cite[Corollary 1]{Marshall2005}). We have recast the known proofs in our terms to show the relevance of discrete residues. \end{remark} \subsection{$q$-discrete residues}\label{SUBSECT:qres} Given a rational function, the $q$-analogue of Abramov's algorithm in~\cite{Abramov1995b} can decide whether it is rational $q$-summable or not. Here, we present a $q$-analogue of Proposition~\ref{PROP:ratsum} in terms of a $q$-discrete analogue of residues. To this end, we first recall some terminology from~\cite{Abramov1975, Abramov1989, Abramov1995b}. For an element~$\alpha \in \ok$, we call the subset~$\{\alpha \cdot q^i \mid i \in \bZ\}$ of~$\ok$ the~\emph{$q^\bZ$-orbit} of~$\alpha$ in~$\ok$, denoted by~$[\alpha]_q$. For a polynomial~$b\in K[x]\setminus K$, the value \[\max\{i\in \bZ \mid \text{$\exists$ nonzero~$\alpha, \beta \in \ok$ such that~$\alpha=q^i\cdot \beta$ and~$b(\alpha)=b(\beta)=0$}\}\] is called the~\emph{$q$-dispersion} of~$b$ with respect to~$x$, denoted by~$\operatorname{qdisp}_x(b)$. For~$b=\lambda x^n$ with~$\lambda \in K$ and~$n\in \bN\setminus\{0\}$, we define~$\operatorname{qdisp}_x(b)=+\infty$. The polynomial~$b$ is said to be~\emph{$q$-shift-free} with respect to~$x$ if~$\operatorname{qdisp}_x(b)=0$. Let~$f=a/b\in K(x)$ be such that~$a, b\in K[x]$ and~$\gcd(a, b)=1$. Over the field~$\ok$, $f$ can be uniquely decomposed into the form \begin{equation}\label{EQ:qparfrac} f = c + xp_1 + \frac{p_2}{x^s} + \sum_{i=1}^m \sum_{j=1}^{n_i} \sum_{\ell=0}^{d_{i,j}} \frac{\alpha_{i,j,\ell}}{(x-q^\ell \cdot \beta_i)^{j}}, \end{equation} where~$c\in K$, $p_1, p_2 \in K[x]$, $m, n_i\in \bN$ are nonzero, $s, d_{i,j}\in \bN$, $\alpha_{i, j, \ell}, \beta_i\in \ok$, and~$\beta_i$'s are nonzero and in distinct $q^\bZ$-orbits. \begin{define}[$q$-discrete residue]\label{DEF:qres} Let~$f\in K(x)$ be of the form~\eqref{EQ:qparfrac}. The sum $\sum_{\ell=0}^{d_{i,j}}q^{-\ell \cdot j} \alpha_{i,j, \ell}$ is called the \emph{$q$-discrete residue} of~$f$ at the~$q^\bZ$-orbit $[\beta_i]_q$ of multiplicity~$j$ (with respect to~$x$), denoted by~$\operatorname{qres}_x(f, [\beta_i]_q, j)$. In addition, we call the constant~$c$ the \emph{$q$-discrete residue} of~$f$ at infinity, denoted by~~$\operatorname{qres}_x(f, \infty)$. \end{define} We summarize some basic facts concerning rational $q$-summability in the next lemma. For a detailed proof, one can see~\cite[\S 3]{Abramov1995b}. \begin{lemma}\label{LM:ratqsum} Let~$p, p_1, p_2\in K[x]$, $c\in K$, and~$s\in \bN\setminus \{0\}$ be as in \eqref{EQ:qparfrac}. Then \begin{enumerate} \item $\deg_x(\Delta_{q, x}(p)) = \deg_x(p)$. \item If~$c$ is nonzero, then~$c$ is not rational $q$-summable in~$K(x)$. \item $f = xp_1 + p_2/x^s$ is rational $q$-summable in~$K(x)$. \end{enumerate} \end{lemma} The following lemma is a $q$-analogue of Lemma~\ref{LM:dres} and its proof proceeds in a similar way. \begin{lemma}\label{LM:qres} Let~$f=\sum_{\ell=0}^d \alpha_{\ell}/(x-q^\ell\cdot \beta )^s$ be such that~$d, s\in \bN$,~$\alpha_{\ell}, \beta\in \ok$, and~$\beta$ is nonzero. Then~$f$ is rational $q$-summable in~$\ok(x)$ if and only if the sum~$\sum_{\ell=0}^d q^{-\ell\cdot s}\alpha_{\ell}$ is zero, that is, if and only if~$\operatorname{qres}_x(f, [\beta]_q, s)=0$. \end{lemma} \begin{proof} Suppose that the sum~$\sum_{\ell=0}^d q^{-\ell\cdot s}\alpha_{\ell}$ is zero. We show that~$f$ is rational $q$-summable in~$\ok(x)$. To this end, we proceed by induction on~$d$. In the base case when~$d=0$, $f$ is clearly rational $q$-summable since~$f=0$. Suppose that the assertion holds for~$d=m$ with~$m\geq 0$. Note that \begin{align*} \frac{\alpha_{m+1}}{(x-q^{m+1}\beta)^s} = \Delta_{q, x} \left(-\frac{\alpha_{m+1}}{(x-q^{m+1}\beta)^s}\right) + \frac{q^{-s}\alpha_{m+1}}{(x-q^m\beta)^s}. \end{align*} This implies that \[\sum_{\ell=0}^{m+1} \frac{\alpha_{\ell}}{(x-q^{\ell}\beta)^s} = \Delta_{q, x} \left(-\frac{\alpha_{m+1}}{(x-q^{m+1}\beta)^s}\right) + \sum_{\ell=0}^{m} \frac{\tilde{\alpha}_{\ell}}{(x-q^{\ell}\beta)^s},\] where~$\tilde{\alpha}_{\ell} = \alpha_{\ell}$ if~$0\leq \ell \leq m-1$ and~$\tilde{\alpha}_{m} = q^{-s}\alpha_{m+1} + \alpha_{m}$. From the definition and assumption on the~$\alpha_\ell$'s, the sum~$\sum_{\ell=0}^m q^{-\ell\cdot s}\tilde{\alpha}_{\ell}$ is zero. The induction hypothesis then implies that there exists~$\tilde{g}\in \ok(x)$ such that \[\sum_{\ell=0}^{m} \frac{\tilde{\alpha}_{\ell}}{(x-q^{\ell}\beta)^s} = \Delta_{q, x}(\tilde{g}).\] So~$f=\Delta_{q, x}(g)$ with~$g=\tilde{g} - \alpha_{m+1}/(x-q^{m+1}\beta)^s\in \ok(x)$. For the opposite implication, we assume to the contrary that the sum~$\sum_{\ell=0}^d q^{-\ell \cdot s}\alpha_\ell$ is nonzero. Without loss of generality, we can assume that~$\alpha_0 \neq 0$. Write~$\alpha_0 = \bar{\alpha}_0 + \tilde{\alpha}_0$ such that~$\tilde{\alpha}_0 + \sum_{\ell=1}^{d} q^{-\ell\cdot s}\alpha_\ell =0$. Since~$\sum_{\ell=0}^d q^{-\ell\cdot s}\alpha_\ell\neq 0$, $\bar{\alpha}_0 \neq 0$. By the assertion shown above, there exists~$\tilde{g}\in \ok(x)$ such that \[f = \frac{\bar{\alpha}_0}{(x-\beta)^s} + \Delta_{q, x}(\tilde{g}).\] Since~$\operatorname{qdisp}_x((x-\beta)^s)=0$ and~$\bar{\alpha}_0\neq 0$, ${\bar{\alpha}_0}/{(x-\beta)^s}$ is not rational summable by~\cite[Lemma 6.3]{Hardouin2008}. Then~$f$ is not rational $q$-summable in~$\ok(x)$. This completes the proof. \end{proof} \begin{prop}\label{PROP:ratqsum} Let~$f=a/b\in K(x)$ be such that~$a, b\in K[x]$ and~$\gcd(a, b)=1$. Then~$f$ is rational $q$-summable in~$K(x)$ if and only if the~$q$-discrete residues $\operatorname{qres}_x(f, \infty)$ and~$\operatorname{qres}_x(f, [\beta]_q, j)$ are all zero for any $q^\bZ$-orbit~$[\beta]_q$ with~$\beta \neq 0$ and~$b(\beta)=0$ of any multiplicity~$j\in \bN$. \end{prop} \begin{proof} Let~$f\in K(x)$ be decomposed into the form~\eqref{EQ:qparfrac}. If the residue of~$f$ at any $q^\bZ$-orbit~$[\beta]_q, \beta \neq 0$, of any multiplicity is zero, then Lemma~\ref{LM:qres} implies that for all~$i, j$ with~$1\leq i \leq m$ and~$1\leq j \leq n_i$, the sum \[\sum_{\ell=0}^{d_{i,j}} \frac{\alpha_{i,j,\ell}}{(x-q^{\ell}\beta_i)^{j}} = \Delta_{q, x}(g_{i,j})\quad \text{for some~$g_{i, j}\in \ok(x)$.}\] Since the rational function~$xp_1 + \frac{p_2}{x^s}$ in~\eqref{EQ:qparfrac} is rational $q$-summable by Lemma~\ref{LM:ratqsum}, there exists~$u\in K(x)$ such that~$xp_1 + p_2/x^s = \Delta_{q, x}(u)$. So~$f = \Delta_{q, x}(u + g)$, where~$g=\sum_{i=1}^m\sum_{j=1}^{n_i} g_{i,j}$. As in Proposition~\ref{PROP:ratsum}, we see that~$g \in K(x)$ and therefore that $f$ is rational~$q$-summable in~$K(x)$. Suppose that~$f$ is rational $q$-summable in~$K(x)$, i.e., $f=\Delta_{q, x}(g)$ for some~$g\in K(x)$. Over the field~$\ok$, we decompose~$g$ into the form~\eqref{EQ:qparfrac}. For all~$i, j$ with~$1\leq i \leq m$ and~$1\leq j \leq n_i$, the linearity of~$\Delta_{q, x}$ implies that \[\Delta_{q, x}\left(\sum_{\ell=0}^{d_{i,j}} \frac{\alpha_{i,j,\ell}}{(x-q^{\ell}\beta_i)^{j}} \right) = \sum_{\ell=0}^{d_{i,j}+1} \frac{\tilde{\alpha}_{i,j,\ell}}{(x-q^{\ell}\tilde{\beta}_i)^{j}}, \] where~$\tilde{\beta}_i = q^{-1}\beta_i$, $\tilde{\alpha}_{i,j, 0}=q^{-j}{\alpha}_{i, j, 0}$, $\tilde{\alpha}_{i, j, d_{i,j}+1} = -\alpha_{i, j, d_{i, j}}$, and~$\tilde{\alpha}_{i, j, \ell} = q^{-j}\alpha_{i,j, \ell}-\alpha_{i, j, \ell-1}$ for~$1\leq \ell \leq d_{i, j}$. Then the residue~$\operatorname{qres}_x(f, [\tilde{\beta}_i]_q, j)=\sum_{\ell=0}^{d_{i, j}}q^{-\ell\cdot j} \tilde{\alpha}_{i,j,\ell}=0$ for all~$i, j$. Since~$\Delta_{q, x}(c)=0$ for any constant~$c\in k$, the residue of~$f$ at infinity is zero. This completes the proof. \end{proof} \subsection{Residual forms}\label{SUBSECT:resform} In terms of residues, we will present a normal form of a rational function in the quotient space~$K(x)/\partial_x (K(x))$ with~$\partial_x\in \{D_x, \Delta_x, \Delta_{q, x}\}$. Let~$f\in K(x)$. If~$f$ is of the form~\eqref{EQ:cparfrac}, then we can reduce it to \[f = D_x(g) + r, \quad \text{where~$r=\sum_{i=1}^m \frac{\operatorname{cres}_x(f, \beta_i)}{x-\beta_i}$}.\] Note that~$r$ actually lies in~$K(x)$. We call such an~$r$ the \emph{residual form} of~$f$ with respect to~$D_x$. Similarly, residual forms with respect to~$\Delta_x$ and~$\Delta_{q, x}$ are respectively \[r=\sum_{i=1}^m\sum_{j=1}^{n_i} \frac{\operatorname{dres}_x(f, [\beta_i], j)}{(x-\beta_i)^j}, \quad \text{where ~$\beta_i$'s in distinct $\bZ$-orbits}.\] and \[r=c + \sum_{i=1}^m\sum_{j=1}^{n_i} \frac{\operatorname{qres}_x(f, [\beta_i]_q, j)}{(x-\beta_i)^j}, \quad \text{where~$c\in K$ and~$\beta_i$'s in distinct $q^\bZ$-orbits}.\] Such a residual form for a rational function is unique up to taking a different representative from orbits. One can compute residual forms without introducing algebraic extensions of~$K$ by algorithms in~\cite{Hermite1872, Ostrogradsky1845, Horowitz1971, Paule1995b, Pirastu1995a, Pirastu1995b, Abramov1995b}. \section{Algebraic functions}\label{SECT:algfuns} As early as 1827, Abel already observed that an algebraic function satisfies a linear differential equation with polynomial coefficients~\cite[p.\ 287]{Abel1881}. The annihilating differential equations are important in the study of algebraic functions and their series expansions~\cite{Comtet1964, Chudnovsky1986, Chudnovsky1987}. Algorithms for constructing differential annihilators for algebraic functions have been developed in~\cite{Cockle1861, Harley1862, CormierSingerTragerUlmer2002, Nahay2003, BCLSS2007}. It is not true that any algebraic function satisfies a linear or a~$q$-linear recurrence. In this section we characterize those algebraic functions that satisfy such equations and prove a few lemmas concerning algebraic solutions of first order linear and~$q$-linear recurrences. In the next section, we will see how this restriction on algebraic solutions of such recurrences is responsible for the essential difference between the continuous problems and the~($q$-)discrete ones. Let~$k$ be an algebraically closed field of characteristic zero. Let~$q\in k$ be such that~$q^i\neq 1$ for any~$i\in \bZ\setminus\{0\}$. Let~$k(t)$ be the field of all rational functions in~$t$ over~$k$. On the field~$k(t)$, we let~$D_t$, $S_t$, and~$Q_t$ denote the derivation, shift operator, and $q$-shift operator with respect to~$t$, respectively. Let~$k(t)\langle D_t \rangle$ (resp.\ $k(t)\langle S_t \rangle$, $k(t)\langle Q_t \rangle$) denote the ring of linear differential (resp.\ recurrence, $q$-recurrence) operators over~$k(t)$. We recall the following fact for reference later. One can find its proof in~\cite[p.\ 339]{Harley1862} or~\cite[p.\ 267]{Comtet1964}. \begin{prop}\label{PROP:aflde} Let~$\alpha(t)$ be an element of the algebraic closure of~$k(t)$. Then there exists a nonzero operator~$L(t, D_t)\in k(t)\langle D_t\rangle$ such that~$L(\alpha)=0$. \end{prop} {As mentioned above, the situation is different if we consider the linear \linebreak ($q$-)recurrence equations for algebraic functions and the following results show that requiring an algebraic function~$f$ to satisfy such a recurrence equation severely restricts~$f$. \begin{prop}\label{PROP:afrde} Let~$\alpha(t)$ be an element in the algebraic closure of~$k(t)$. If there exists a nonzero operator~$L(t, S_t)\in k(t)\langle S_t \rangle$ such that~$L(\alpha)=0$, then~$\alpha\in k(t)$. \end{prop} \begin{prop}\label{PROP:aflqe} Let~$\alpha(t)$ be an element in the algebraic closure of~$k(t)$. If there exists a nonzero operator~$L(t, Q_t)\in k(t)\langle Q_t \rangle$ such that~$L(\alpha)=0$, then~$\alpha\in k(t^{1/n})$ for some positive integer~$n$. \end{prop} We have included complete proofs (and references to other proofs) of these results in the Appendix. In the next section, algebraic functions will appear as residues of bivariate rational functions and these functions will satisfy certain first order linear ($q$-)recurrence relations. The following lemmas characterize the form of these functions. Although these characterizations can be derived from Propositions~\ref{PROP:afrde} and~\ref{PROP:aflqe}, we will give more elementary proofs. Abusing notation, we let~$S_t$ and~$Q_t$ denote arbitrary extensions of~$S_t$ and~$Q_t$ to automorphisms of~$\overline{k(t)}$, the algebraic closure of~$k(t)$. \begin{lemma}\label{LM:const} Let~$n$ be a positive integer. \begin{enumerate} \item[(i)] If~$f\in \overline{k(t)}$ and~$S_t^n(f) = f$, then~$f \in k$. \item[(ii)] If~$f\in \overline{k(t)}$ and~$Q_t^n(f) = f$, then~$f \in k$. \item[(iii)] If~$f\in \overline{k(t)}$ and~$D_t(f)=0$, then~$f \in k$. \end{enumerate} \end{lemma} \begin{proof}~$(i)$. We begin by showing that if~$f\in {k(t)}$ and~$S_t^n(f) = f$ then~$f \in k$. If~$f\notin k$, then there exists an element~$a\in k$ such that~$a$ is a pole or zero of~$f$. In this case the infinite set~$\{a + in \ | \ i \in \bZ \}$ will also consist of poles or zeroes, an impossibility since~$f$ is a rational function. Now assume that~$f\in \overline{k(t)}$ and~$S_t^n(f) = f$. Let~$Y^\lambda + a_{\lambda-1} Y^{\lambda-1} + \ldots + a_0$ be the minimal polynomial of~$f$ over~$k(t)$. We then have that~$Y^\lambda + S_t^n(a_{\lambda-1}) Y^{\lambda-1} + \ldots + S_t^n(a_0)$ is also the minimal polynomial of~$f(t) = S_t^n(f(t))$. Therefore~$S_t^n(a_i) = a_i$ for all~$i = \lambda -1, \ldots , 0$. This implies that the~$a_i \in k$. Since $k$ is algebraically closed,~$f \in k$. $(ii)$. We again begin by showing that if~$f\in {k(t)}$ and~$Q_t^n(f) = f$ then~$f \in k$. Assume~$f \notin k$ and let~$a \in k$ be a nonzero pole or zero of~$f$. We then have that the set~$\{aq^{in} \ | \ i \in \bZ \}$ consists of poles or zeroes. Since~$q$ is not a root of unit, this set is infinite and we get a contradiction as before. Therefore,~$f = ct^m$ for some~$m \in \bZ$. Since~$f(q^nt) = f(t)$, we have~$q^{nm} = 1$, a contradiction. Therefore~$f\in k$. Now assume that~$f\in \overline{k(t)}$ and~$Q_t^n(f) = f$. An argument similar to that in 1.~shows that 2.~holds. $(iii)$. This assertion follows from Lemma~3.3.2~(i) of~\cite[Chapter 3]{BronsteinBook} and the assumption that~$k$ is algebraically closed. \end{proof} \begin{lemma}\label{LM:commuting} Let~$E \subset F$ be fields of characteristic zero with~$F$ algebraic over~$E$. Let~$\sigma$ be an automorphism of~$F$ such that~$\sigma(E) \subset E$ and let~$\delta$ be a derivation of~$F$ such that~$\delta(E) \subset E$. If~$\delta\sigma(f) = \sigma\delta(f)$ for all~$f \in E$, then~$\delta\sigma(f) = \sigma\delta(f)$ for all~$f \in F$.\end{lemma} \begin{proof} One can verify that~$\sigma^{-1}\delta\sigma$ is a derivation on~$F$ such that~$\sigma^{-1}\delta\sigma(E) \subset E$. Therefore~$\sigma^{-1}\delta\sigma - \delta$ is a derivation on~$f$ that is zero on~$E$. From the uniqueness of extensions of derivations to algebraic extensions, we have that~$\sigma^{-1}\delta\sigma - \delta$ is zero on~$F$, which yields the result.\end{proof} } \begin{lemma}\label{LM:cstdq} Let~$\alpha(t)$ be an element in the algebraic closure of~$k(t)$. If there exists a nonzero~$n\in \bN$ such that~$S_t^n(\alpha)= q^m\alpha$ for some~$m\in \bZ$, then~$m=0$ and~$\alpha(t)\in k$. \end{lemma} \begin{proof} Let~$\delta = D_t$. Lemma~\ref{LM:commuting} implies that~$S_t^n\delta = \delta S_t^n$ on~$\overline{k(t)}$. Therefore,~$S_t^n(\delta \alpha) = q^m \delta\alpha$. One see that this implies that~$S_t^n(\delta\alpha/\alpha) = \delta\alpha/\alpha$, so by Lemma~\ref{LM:const}~$\delta\alpha = c \alpha$ for some~$c \in k$. Assume that~$\alpha \notin k$ and therefore that~$\delta\alpha \neq 0$ and~$c \neq 0$. Let~$P(Y) = Y^\lambda + a_{\lambda-1} Y^{\lambda-1} + \ldots + a_0$ be the minimal polynomial of~$\alpha$ over~$k(t)$. Applying~$\delta$ to~$P(\alpha)$, one sees that \[Y^\lambda + \frac{\delta a_{\lambda-1} + (\lambda-1)c }{\lambda c} Y^{\lambda -1} + \ldots + \frac{\delta a_0}{\lambda c}\] is also the minimal polynomial of~$\alpha$ over~$k(t)$. Therefore \[ \frac{\delta a_0}{a_0} = \lambda c .\] Since~$a_0 \in k(t)$, we may write~$a_0 = d\prod(t-e_i)^{\mu_i}$, where~$d, e_i \in k, \mu_i \in \bZ$. Therefore \[\sum \frac{\mu_i}{t-e_i} = \lambda c\] contradicting the uniqueness of partial fraction decomposition. This contradiction implies that $\alpha \in k$. From the equation~$S_t^n(\alpha)=q^m \alpha$ we get~$q^m=1$. Therefore~$m=0$ since~$q$ is not root of unity. \end{proof} \begin{lemma}\label{LM:intlin} Let~$\alpha(t)$ be an element in the algebraic closure of~$k(t)$. If there exists a nonzero~$n\in \bZ$ such that~$S_t^n(\alpha)-\alpha = m$ for some~$m\in \bZ$, then~$\alpha(t)=\frac{m}{n} t + c$ for some~$c\in k$. \end{lemma} \begin{proof} Let~$\beta(t) = \frac{m}{n} t$. Since~$S_t^n(\beta) - \beta = m$, we have that $S_t^n(\alpha - \beta) - (\alpha - \beta) = 0$. Therefore Lemma~\ref{LM:const} implies that~$\alpha = \beta + c = \frac{m}{n}t + c$ for some~$c \in k$. \end{proof} \begin{lemma}\label{LM:cstqd} Let~$\alpha(t)$ be an element in the algebraic closure of~$k(t)$. If there exists a nonzero~$n\in \bZ$ such that~$Q_t^n(\alpha)-\alpha=m$ for some~$m\in \bZ$, then~$m=0$ and~$\alpha(t)\in k$. \end{lemma} \begin{proof}Let$\delta = tD_t$. One has that~$\delta Q_t = Q_t \delta$ on~$k(t)$ so Lemma~\ref{LM:commuting} implies that~$\delta Q_t = Q_t \delta$ on~$\overline{k(t)}$. We then also have~$\delta Q_t^n = Q_t^n \delta$ on~$\overline{k(t)}$ so~$Q_t^n(\delta \alpha) - \delta \alpha = 0$. Lemma~\ref{LM:const} implies~$\delta\alpha \in k$. Suppose that~$\delta \alpha = c$ for~$c\in k$. Then~$D_t(\alpha) = c/t$. If~$\text{Tr}: k(t)(\alpha) \rightarrow k(t)$ is the trace mapping, then ~$D_t(\text{Tr}(\alpha)) = \lambda c/t$ for some nonzero~$\lambda \in \bN$. By Proposition~\ref{PROP:ratint}, we have~$\lambda c =0$ and then~$c=0$. Now~$\alpha\in k$ follows from the third assertion of Lemma~\ref{LM:const}. \end{proof} \begin{lemma}\label{LM:qintlin} Let~$\alpha(t)$ be an element in the algebraic closure of~$k(t)$. If there exists a nonzero~$n\in \bZ$ such that~$Q_t^n(\alpha)= q^m\alpha$ for some~$m\in \bZ$, then~$\alpha(t)=c t^{\frac{m}{n}}$ for some~$c\in k$. \end{lemma} \begin{proof} Let~$\beta(t) = t^\frac{m}{n}$. We then have that \[Q_t^n(\frac{\alpha}{\beta}) = \frac{\alpha}{\beta}\] so $\alpha/\beta = c \in k$ by Lemma~\ref{LM:const}, that is,~$\alpha = c t^\frac{m}{n}$.\end{proof} \section{Telescopers}\label{SECT:telescoper} In Section~\ref{SECT:residues}, we see that nonzero residues are the obstruction for a rational function to being rational integrable (resp.\ summable, $q$-summable). In this section, we consider whether we can use a linear operator, a so-called~\emph{telescoper}, to remove this obstruction if an extra parameter is available. The importance of telescopers in the study of special functions and combinatorial identities have been shown in the work by Zeilberger and his collaborators~\cite{Zeilberger1990, Almkvist1990, WilfZeilberger1990a, WilfZeilberger1990b, Wilf1992}. Let~$k(t, x)$ be the field of rational functions in~$t$ and~$x$ over~$k$. On the field~$k(t, x)$, we have derivations~$D_t, D_x$, shift operators~$S_t, S_x$, and $q$-shift operators~$Q_t, Q_x$. The linear operators used below will be in the ring~$k(t)\langle D_t \rangle$, $k(t)\langle S_t \rangle$, or~$k(t)\langle Q_t \rangle$. For a rational function~$f\in k(t, x)$, we wish to solve the Existence Problem for Telescopers stated in the Introduction, that is, we want to decide the existence of linear operators~$L(t, \partial_t)$ with~$\partial_t \in \{D_t, S_t, Q_t\}$ such that \begin{equation}\label{EQ:tele} L(t, \partial_t)(f)=\partial_x(g) \end{equation} for some~$g\in k(t, x)$ and~$\partial_x \in \{D_x, \Delta_x, \Delta_{q, x}\}$. According to the different choices of~$L$ and~$\partial_x$, we have nine types of telescopers in general, see Table~\ref{tab:ninetelepb}. \begin{center} \renewcommand{\arraystretch}{1.2} \tabcolsep4.5pt \begin{table}[ht] \begin{tabular}{|c|c|c|c|} \hline $(L, \partial_x) $ & $D_x$ & $\Delta_x$ & $\Delta_{q, x}$ \\ \hline $k(t)\langle D_t \rangle $ & $L(t, D_t)(f) = D_x(g)$ & \underline{$L(t, D_t)(f) = \Delta_x(g)$} & \underline{$L(t, D_t)(f) = \Delta_{q, x}(g)$} \\ $k(t)\langle S_t \rangle$ & \underline{$L(t, S_t)(f) = D_x(g)$} & $L(t, S_t)(f) = \Delta_x(g)$ & \underline{$L(t, S_t)(f) = \Delta_{q, x}(g)$}\\ $k(t)\langle Q_t \rangle$ & \underline{$L(t, Q_t)(f) = D_x(g)$} & \underline{$L(t, Q_t)(f) = \Delta_x(g)$} & $L(t, Q_t)(f) = \Delta_{q, x}(g)$\\[0.05in] \hline \end{tabular} \caption{Nine different types of telescoping equations}\label{tab:ninetelepb} \end{table} \end{center} \vspace{-0.5cm} The existence problem of telescopers is related to the termination of Zeilberger-style algorithms and has been studied in~\cite{AbramovLe2002, Abramov2003, Chen2005, CCFL2010} but, to our knowledge, our results concerning telescopers of the six types underlined in the above table are new. In this section, we will present a unified way to solve this problem for rational functions by using the knowledge in the previous sections. Before the investigation of the existence of telescopers, we first present some preparatory lemmas for later use. \begin{define} Let~$\sim$ be an equivalence relation on a set~$R$ and~$\sigma: R\rightarrow R$ be a bijection. The relation~$\sim$ is said to be~\emph{$\sigma$-compatible} if \[\sigma(r_1) \sim \sigma(r_2) \, \, \Leftrightarrow \, \, r_1 \sim r_2 \quad \text{for all~$r_1, r_2\in R$.}\] \end{define} If the equivalence relation~$\sim$ is compatible with a bijection~$\sigma$ on~$R$, then a bijection on the quotient set~$R/\sim$ can be naturally induced by~$\sigma$, for which we still use the name~$\sigma$. We denote by~$[t]$ the equivalence class of~$t$ in~$R/\sim$. \begin{prop}\label{PROP:periodic} Let~$\sigma: R\rightarrow R$ be a bijection and~$\sim$ be a~$\sigma$-compatible equivalence relation on the set~$R$. Let~$T=\{[t_1], \ldots, [t_n]\}\subset R/\sim$. If for any~$i\in \{1, \ldots, n\}$, there exists nonzero~$m_i\in \bN$ such that~$\sigma^{m_i}([t_i])\in T$, then there exists nonzero~$m\in \bN$ such that~$\sigma^m([t_i])=[t_i]$ for all~$i\in \{1, \ldots, n\}$. \end{prop} \begin{proof} Let~$\tilde{m}$ be the least common multiple of~$m_i$'s. Then~$\sigma^{\tilde{m}}$ is a permutation on the finite set~$T$. Since any permutation on a finite set is idempotent, there exists an~$s\in \bN$ such that~$\sigma^{\tilde{m}s}$ is an identity on~$T$. Taking~$m=\tilde{m}s$ completes the proof. \end{proof} We will specialize Proposition~\ref{PROP:periodic} to different bijections and equivalence relations. The following examples show how to perform specializations. \begin{example}\label{EX:intlin} Let~$R$ be the algebraic closure of~$k(t)$. The equivalence relation~$\sim$ on~$R$ is defined by $\alpha_1 \sim \alpha_2$ if and only if~$\alpha_1-\alpha_2\in \bZ$. We take the shift mapping~$\sigma(\alpha(t))=\alpha(t+1)$ as the bijection. Let~$T=\{[\alpha_1], \ldots, [\alpha_n]\}$ be such that for any~$i\in \{1, \ldots, n\}$, $\sigma^{m_i}([\alpha_i])\in T$ for some nonzero~$m_i\in \bN$. By Proposition~\ref{PROP:periodic}, there exists nonzero~$m\in \bN$ such that~$\sigma^m(\alpha_i)-\alpha_i \in \bZ$ for all~$i\in \{1, \ldots, n\}$. Applying Lemma~\ref{LM:intlin} to~$\alpha_i$ yields~$\alpha_i = \frac{n_i}{m} t +c_i$ for some~$n_i\in \bZ$ and~$c_i\in k$. \end{example} \begin{example}\label{EX:intqlin} Let~$R$ be the algebraic closure of~$k(t)$. The equivalence relation~$\sim$ on~$R$ is defined by $\alpha_1 \sim \alpha_2$ if and only if~$\alpha_1/\alpha_2\in q^{\bZ}$. We take the $q$-shift mapping~$\sigma(\alpha(t))=\alpha(qt)$ as the bijection. Let~$T=\{[\alpha_1]_q, \ldots, [\alpha_n]_q\}$ be such that for any~$i\in \{1, \ldots, n\}$, $\sigma^{m_i}([\alpha_i])\in T$ for some nonzero~$m_i\in \bN$. By Proposition~\ref{PROP:periodic}, there exists nonzero~$m\in \bN$ such that~$\sigma^m(\alpha_i)/\alpha_i \in q^{\bZ}$ for all~$i\in \{1, \ldots, n\}$. Applying Lemma~\ref{LM:qintlin} to~$\alpha_i$ yields~$\alpha_i = c_i t^{{n_i}/{m}}$ for some~$n_i\in \bZ$ and~$c_i\in k$. \end{example} \subsection{Existence of telescopers}\label{SUBSECT:existence} The first result about the existence of telescopers was shown by Zeilberger in~\cite{Zeilberger1990} based on the theory of holonomic D-modules. In the following, we will study the existence problems from the residual point of view. For rational functions, the existence of telescopers is related to the properties of residues and the commutativity between the residue mappings and linear operators. Starting from the simplest, we consider the telescoping relation~$L(t, D_t)(f)=D_x(g)$ for a given rational function~$f\in k(t, x)$. Given~$\beta\in \overline{k(t)}$, view the residue mapping~$\operatorname{cres}_x(\underline{ \ \ }, \beta)$ as a $\overline{k(t)}$-linear transformation from~$\overline{k(t)}(x)$ to~$\overline{k(t)}$. For any~$\alpha, \beta\in \overline{k(t)}$, we have \[D_t\left(\frac{\alpha}{x-\beta}\right) = \frac{D_t(\alpha)}{x-\beta} + \frac{\alpha D_t(\beta)}{(x-\beta)^2}.\] Then~$\operatorname{cres}_x(D_t(f), \beta) = D_t(\operatorname{cres}_x(f, \beta))$ for any~$f\in \overline{k(t)}(x)$ and~$\beta\in \overline{k(t)}$. Assume that~$f=a/b$ with~$a, b \in k[t, x]$ and~$\gcd(a, b)=1$. Let~$\beta_1, \ldots, \beta_m$ be the roots of~$b$ in~$\overline{k(t)}$. For each root~$\beta_i$, the continuous residue $\operatorname{cres}_x(f, \beta_i) \in \overline{k(t)}$ is annihilated by a linear differential operator~$L_{i}\in k(t)\langle D_t \rangle$ by Proposition~\ref{PROP:aflde}. Let~$L(t, D_t)$ be the least common left multiple (LCLM) of the~$L_i$'s. Then we have~$L(\operatorname{cres}_x(f, \beta_i))= \operatorname{cres}_x(L(f), \beta_i)=0$ for all~$i$ with~$1\leq i \leq m$. So~$L(f)$ is rational integrable with respect to~$x$ by Proposition~\ref{PROP:ratint}. In summary, we have the following theorem. \begin{theorem}\label{THM:cc} For any~$f\in k(t, x)$, there exists a nonzero operator~$L\in k(t)\langle D_t \rangle$ such that~$L(f)= D_x(g)$ for some~$g\in k(t, x)$. \end{theorem} However, the situation in other cases turns out to be more involved. For the rational function~$f=1/(t^2+x^2)$, Abramov and Le~\cite{Le2001, AbramovLe2002} showed that there is no telescoper in~$k(t)\langle S_t \rangle$ such that~$L(f)=\Delta_x(g)$ for any~$g\in k(t, x)$. In other cases, there are two main reasons for non-existence: one is the non-commutativity between linear operators~$\partial_t\in \{D_t, S_t, Q_t\}$ and residue mappings, the other is that not all algebraic functions would satisfy linear ($q$)-recurrence relations. So it is natural that rational functions are of special forms if telescopers exist. Let~$f\in k(t, x)$ and~$\partial_x\in \{D_x, \Delta_x, \Delta_{q, x}\}$. Then~$f = \partial_x(g) + r$ with~$g, r\in k(t, x)$ and~$r$ being the residual form of~$f$ with respect to~$\partial_x$ {(see Section~\ref{SUBSECT:resform})}. Since linear operators~$L(t, \partial_t)$ with~$\partial_t\in \{D_t, S_t, Q_t\}$ commute with the linear operator~$\partial_x\in \{D_x, \Delta_x, \Delta_{q, x}\}$, a rational function has a telescoper if and only if its residual form does. From now on, we always assume that the given rational function is in its residual form. We will also use the fact~\cite[Lemma 1]{AbramovLe2002} that the sum~$f_1+f_2$ has a telescoper if both~$f_1$ and~$f_2$ do. To be more precise, if~$L_1, L_2$ are telescopers for~$f_1, f_2$, respectively, then the LCLM of~$L_1, L_2$ is a telescoper for~$f_1+f_2$. \subsubsection{Telescopers with respect to~$D_x$}\label{SUBSECT:existencediff} Let~$f\in k(t, x)$ be a residual form, that is, \begin{equation}\label{EQ:cresform} f = \sum_{i=1}^m \frac{\alpha_i}{x-\beta_i}, \quad \text{where~$\alpha_i, \beta_i\in \overline{k(t)}$ and the~$\beta_i$ are pairwise distinct.} \end{equation} \begin{theorem}\label{THM:dc} Let~$f\in k(t, x)$ be as in~\eqref{EQ:cresform}. Then~$f$ has a telescoper~$L$ in~$k(t)\langle S_t\rangle$ such that~$L(t, S_t)(f)= D_x(g)$ for some~$g\in k(t, x)$ if and only if all the~$\beta_i$ are in~$k$. \end{theorem} \begin{proof} Suppose that there exists a nonzero~$L\in k(t)\langle S_t\rangle$ such that~$L(t, S_t)(f)=D_x(g)$ for some~$g\in k(t, x)$. Write~$L=\sum_{\ell=0}^{\rho} e_{\ell}S_t^{\ell}$ with~$e_{\ell}\in k(t)$ and~{$e_\rho = 1$}. Then \[L(f) = \sum_{\ell=0}^{\rho} \sum_{i=1}^m \frac{e_{\ell}S_t^{\ell}(\alpha_i)}{x-S_t^{\ell}(\beta_i)}.\] Assume that~$\ell_0$ is the first index in~$\{0, 1, \ldots, \rho\}$ such that~$e_{\ell_0}\neq 0$. Since~$L(f)$ is rational integrable in~$k(t, x)$ with respect to~$D_x$, all residues of~$L(f)$ are zero by Proposition~\ref{PROP:ratint}. In particular, the set~$T=\{S_t^{\ell_0}(\beta_1), \ldots, S_t^{\ell_0}(\beta_m)\}$ satisfies the property that for any~$i\in \{1, \ldots, m\}$, there exists nonzero~$m_i\in \bN$ such that~$S_t^{\ell_0+m_i}(\beta_i)\in T$. By taking equality as the equivalence relation and the shift mapping as the bijection in Proposition~\ref{PROP:periodic}, there exists nonzero~$m\in \bN$ such that~$S_t^{\ell_0+m}(\beta_i)= \beta_i$ for all~$i\in \{1, \ldots, m\}$. By Lemma~\ref{LM:const}~(i) and the assumption that~$k$ is algebraically closed, all the~$\beta_i$ are in~$k$. For the opposite implication, it suffices to show that each fraction~$\alpha_i/(x-\beta_i)$ with~$\beta_i\in k$ has a telescoper in~$k(t)\langle S_t\rangle$. According to the process of partial fraction decomposition, ~$\alpha_i\in k(t)(\beta_i)$ for any~$i$ with~$1\leq i\leq m$. Then~$\alpha_i\in k(t)$, which is annihilated by the operator~$L_i = S_t - {\alpha_i(t+1)}/{\alpha_i(t)}$. Moreover, $L_i(\alpha_i/(x-\beta_i))=L_i(\alpha_i)/(x-\beta_i)=0$. So the LCLM of the~$L_i$'s is a telescoper for~$f$. This completes the proof. \end{proof} \begin{theorem}\label{THM:qc} Let~$f\in k(t, x)$ be as in~\eqref{EQ:cresform}. Then~$f$ has a telescoper~$L$ in~$k(t)\langle Q_t\rangle$ such that~$L(t, Q_t)(f)= D_x(g)$ for some~$g\in k(t, x)$ if and only if all the~$\beta_i$ are in~$k$. \end{theorem} \begin{proof} The proof proceeds in a similar way as above replacing~$S_t$ by~$Q_t$ and Lemma~\ref{LM:const}~(i) by~Lemma~\ref{LM:const}~(ii). \end{proof} \begin{example} Let~$f=1/(x+t)$. Since the root of~$x+t$ in~$\overline{k(t)}$ is~$t$, which is not in~$k$, $f$ has no telescoper in either~$k(t)\langle S_t\rangle$ or~$k(t)\langle Q_t\rangle$ with respect to~$D_x$ by Theorems~\ref{THM:dc} and~\ref{THM:qc}. \end{example} \subsubsection{Telescopers with respect to~$\Delta_x$}\label{SUBSECT:existenceshift} Let~$f\in k(t, x)$ be of the form \begin{equation}\label{EQ:dresform} f = \sum_{i=1}^m\sum_{j=1}^{n_i} \frac{\alpha_{i, j}}{(x-\beta_i)^j}, \end{equation} where~$\alpha_{i,j}, \beta_i\in \overline{k(t)}$, $\alpha_{i, n_i}\neq 0$, and the~$\beta_i$ are in distinct~$\bZ$-orbits. \begin{theorem}\label{THM:cd} Let~$f\in k(t, x)$ be as in~\eqref{EQ:dresform}. Then~$f$ has a telescoper~$L$ in~$k(t)\langle D_t\rangle$ such that~$L(t, D_t)(f)= \Delta_x(g)$ for some~$g\in k(t, x)$ if and only if all the~$\beta_i$ are in~$k$. \end{theorem} \begin{proof} Suppose that there exists a nonzero~$L\in k(t)\langle D_t\rangle$ such that~$L(t, D_t)(f)=\Delta_x(g)$ for some~$g\in k(t, x)$. Write~$L=\sum_{\ell=0}^{\rho} e_{\ell}D_t^{\ell}$ with~$e_{\ell}\in k(t)$. By induction on~$\ell$, we get \[D_t^{\ell}\left(\frac{\alpha_{i, n_i}}{(x-\beta_i)^{n_i}}\right) = \frac{(n_i)_{\ell}\alpha_{i, n_i}(D_t(\beta_i))^{\ell}}{(x-\beta_i)^{n_i+\ell}} + \, \, \text{lower terms,}\] where~$(n_i)_{\ell} = n_i (n_i+1) \cdots (n_i+\ell-1)$. Then we have \[L(f) = \sum_{i=1}^m \frac{(n_i)_{\rho}\alpha_{i, n_i}(D_t(\beta_i))^{\rho}}{(x-\beta_i)^{n_i+\rho}} +\,\, \text{lower terms.}\] Since~$L(f)$ is rational summable with respect to~$\Delta_x$ and the~$\beta_i$ are in distinct~$\bZ$-orbits, we get~$(n_i)_{\rho}\alpha_{i, n_i}(D_t(\beta_i))^{\rho}=0$ for all~$i\in \{1, \ldots, m\}$ by Proposition~\ref{PROP:ratsum}. Since~$\alpha_{i, n_i}\neq 0$ and~$(n_i)_{\rho}>0$, $D_t(\beta_i)=0$, which implies that~$\beta_i\in k$ {by Lemma~\ref{LM:const}~(iii)}. For the opposite implication, the proof is similar to that of Theorem~\ref{THM:dc}. Let~$L_{i,j}$ be the operator~$D_t - D_t(\alpha_{i, j})/\alpha_{i,j}\in k(t)\langle D_t \rangle$. Then the LCLM of the~$L_{i, j}$ is a telescoper for~$f$ with respect to~$\Delta_x$. \end{proof} \begin{example}\label{EX:transcendental} Let \[f = \frac{1}{x^2 - t} =\frac{1}{2\sqrt{t}}\left(\frac{1}{x - \sqrt{t}} - \frac{1}{x +\sqrt{t}}\right).\] Note that $f$ is already in residual form with respect to~$\Delta_x$. By Theorem~\ref{THM:cd}, there is no linear differential operator~$L(t,D_t) \in k(t)\langle D_t\rangle$ and $g \in k(t,x)$ such that $L(t,D_t)f = \Delta_x(g)$. Furthermore, Proposition~3.1 in~\cite{Hardouin2008} and the descent argument similar to that given in the proof of Corollary~3.2 of~\cite{Hardouin2008} (or Section 1.2.1 of~\cite{DH2011}) implies that the sum \[F(t,x) = \sum_{i=1}^{x -1}\frac{1}{{ i}^2-t} \ \ \mbox { (satisfying~$S_x(F) - F = f$) }\] satisfies no polynomial differential equation $P(t,x,F,D_tF,D_t^2F, \ldots ) = 0$. \end{example} The following theorem is the same as~\cite[Theorem 1]{AbramovLe2002}. We give an alternative proof using the knowledge developed in previous sections. \begin{theorem}\label{THM:dd} Let~$f\in k(t, x)$ be as in~\eqref{EQ:dresform}. Then~$f$ has a telescoper~$L$ in~$k(t)\langle S_t\rangle$ such that~$L(t, S_t)(f)= \Delta_x(g)$ for some~$g\in k(t, x)$ if and only if all the~$\beta_i=r_i t +c_i$ with~$r_i\in \bQ$ and~$c_i\in k$. \end{theorem} \begin{proof} Suppose that there exists a nonzero~$L\in k(t)\langle S_t\rangle$ such that~$L(t, S_t)(f)=\Delta_x(g)$ for some~$g\in k(t, x)$. Write~$L=\sum_{\ell=0}^{\rho} e_{\ell}S_t^{\ell}$ with~$e_{\ell}\in k(t)$ and~$e_0\neq 0$. For any~$\lambda \in \{1, \ldots, m\}$, we consider the rational function \[f_{\lambda} = \sum_{i=1}^m \frac{\alpha_{i, n_{\lambda}}}{(x-\beta_i)^{n_{\lambda}}},\quad \text{where~$\alpha_{\lambda, n_{\lambda}}\neq 0$ by assumption} .\] Without loss of generality, we may assume that the other~$\alpha_{i, n_{\lambda}}$ with~$i\neq \lambda$ are also nonzero. Since the shift operators~$S_t, S_x$ preserve the multiplicity, we have~$L(f_{\lambda})=\Delta_x(g_{\lambda})$ for some~$g_{\lambda}\in k(t, x)$. By Proposition~\ref{PROP:ratsum}, all the residues of~$L(f_{\lambda})$ are zero. We now use the notation and analysis of Example~\ref{EX:intlin}. We see that the set~$T=\{[\beta_1], \ldots, [\beta_m]\}$ satisfies the property that for any~$i \in \{1, \ldots, m\}$, there exists a nonzero~$m_i$ such that~$S_t^{m_i}([\beta_i])\in T$. As in Example~\ref{EX:intlin}, we conclude that~$\beta_i=\frac{p_i}{m} t +c_i$ with~$p_i, m\in \bZ$ and~$c_i\in k$. The opposite implication follows from the fact that the linear operator \[L_{i, j}=\alpha_{i, j}(t)S_t^{m}-\alpha_{i,j}(t+m)\] is a telescoper for the fraction~$f_{i, j} = \alpha_{i, j}/(x-(\frac{p_i}{m} t +c_i))^j$ with respect to~$\Delta_x$ since~$\operatorname{dres}(L_{i, j}(f_{i, j}), [\frac{p_i}{m} t +c_i], j)=0$. Then the LCLM of the~$L_{i, j}$ is a telescoper for~$f$ with respect to~$\Delta_x$. \end{proof} \begin{theorem}\label{THM:qd} Let~$f\in k(t, x)$ be as in~\eqref{EQ:dresform}. Then~$f$ has a telescoper~$L$ in~$k(t)\langle Q_t\rangle$ such that~$L(t, Q_t)(f)= \Delta_x(g)$ for some~$g\in k(t, x)$ if and only if all the~$\beta_i$ are in~$k$. \end{theorem} \begin{proof} Suppose that there exists a nonzero~$L\in k(t)\langle Q_t\rangle$ such that~$L(t, Q_t)(f)=\Delta_x(g)$ for some~$g\in k(t, x)$. Write~$L=\sum_{\ell=0}^{\rho} e_{\ell}Q_t^{\ell}$ with~$e_{\ell}\in k(t)$ and~$e_0\neq 0$. For any~$\lambda \in \{1, \ldots, m\}$, we consider the rational function \[f_{\lambda} = \sum_{i=1}^m \frac{\alpha_{i, n_{\lambda}}}{(x-\beta_i)^{n_{\lambda}}},\quad \text{where~$\alpha_{\lambda, n_{\lambda}}\neq 0$ by assumption} .\] Without loss of generality, we may assume that the other~$\alpha_{i, n_{\lambda}}$ with~$i\neq \lambda$ are also nonzero. Since the operators~$Q_t, S_x$ preserve the multiplicity, we have~$L(f_{\lambda})=\Delta_x(g_{\lambda})$ for some~$g_{\lambda}\in k(t, x)$. By Proposition~\ref{PROP:ratsum}, all the residues of~$L(f_{\lambda})$ are zero. We shall again use the reasoning and notation in Example~\ref{EX:intlin} where~$[ \ \ ]$ is an equivalence class of the equivalence relation that~$\alpha_1\sim \alpha_2$ in~$\overline{k(t)}$ if~$\alpha_1-\alpha_2\in \bZ$. In particular, the set~$T=\{[\beta_1], \ldots, [\beta_m]\}$ satisfies the property that for any~$i \in \{1, \ldots, m\}$, there exists a nonzero~$m_i$ such that~$Q_t^{m_i}([\beta_i])\in T$. Taking the shift mapping~$Q_t$ as the bijection, Proposition~\ref{PROP:periodic} and Lemma~\ref{LM:cstqd} imply that~$\beta_i\in k$ for all~$i$ with~$1\leq i \leq m$. The opposite implication follows from the fact that the linear operator \[L_{i, j}=\alpha_{i, j}(t)Q_t-\alpha_{i,j}(qt)\] is a telescoper for the fraction~$f_{i, j} = \alpha_{i, j}/(x-\beta_i)^j$ with respect to~$\Delta_x$ since $\operatorname{dres}(L_{i, j}(f_{i, j}), [\beta_i], j)=0$. Then the LCLM of the~$L_{i, j}$ is a telescoper for~$f$ with respect to~$\Delta_x$. \end{proof} \subsubsection{Telescopers with respect to~$\Delta_{q, x}$}\label{SUBSECT:existenceqshift} Let~$f\in k(t, x)$ be of the form \begin{equation}\label{EQ:qresform} f = c + \sum_{i=1}^m\sum_{j=1}^{n_i} \frac{\alpha_{i, j}}{(x-\beta_i)^j}, \end{equation} where~$c\in k(t)$, $\alpha_{i,j}, \beta_i\in \overline{k(t)}$, $\alpha_{i, n_i}\neq 0$, and the~$\beta_i$ are in distinct~$q^\bZ$-orbits. \begin{theorem}\label{THM:cq} Let~$f\in k(t, x)$ be as in~\eqref{EQ:qresform}. Then~$f$ has a telescoper~$L$ in~$k(t)\langle D_t\rangle$ such that~$L(t, D_t)(f)= \Delta_{q, x}(g)$ for some~$g\in k(t, x)$ if and only if all the~$\beta_i$ are in~$k$. \end{theorem} \begin{proof} The proof proceeds in the same way as that in Theorem~\ref{THM:cd}. \end{proof} \begin{theorem}\label{THM:dq} Let~$f\in k(t, x)$ be as in~\eqref{EQ:qresform}. Then~$f$ has a telescoper~$L$ in~$k(t)\langle S_t\rangle$ such that~$L(t, S_t)(f)= \Delta_{q, x}(g)$ for some~$g\in k(t, x)$ if and only if all the~$\beta_i$ are in~$k$. \end{theorem} \begin{proof} Suppose that there exists a nonzero~$L\in k(t)\langle S_t\rangle$ such that~$L(t, S_t)(f)=\Delta_{q, x}(g)$ for some~$g\in k(t, x)$. Write~$L=\sum_{\ell=0}^{\rho} e_{\ell}S_t^{\ell}$ with~$e_{\ell}\in k(t)$ and~$e_0\neq 0$. For any~$\lambda \in \{1, \ldots, m\}$, we consider the rational function \[f_{\lambda} = \sum_{i=1}^m \frac{\alpha_{i, n_{\lambda}}}{(x-\beta_i)^{n_{\lambda}}},\quad \text{where~$\alpha_{\lambda, n_{\lambda}}\neq 0$ by assumption} .\] Without loss of generality, we may assume that the other~$\alpha_{i, n_{\lambda}}$ with~$i\neq \lambda$ are also nonzero. Since the operators~$S_t, Q_x$ preserve the multiplicity, we have~$L(f_{\lambda})=\Delta_{q, x}(g_{\lambda})$ for some~$g_{\lambda}\in k(t, x)$. By Proposition~\ref{PROP:ratqsum}, all the residues of~$L(f_{\lambda})$ are zero. We now use the reasoning and notation in Example~\ref{EX:intqlin}. In particular, the set~$T=\{[\beta_1]_q, \ldots, [\beta_m]_q\}$ satisfies that for any~$i \in \{1, \ldots, m\}$, there exists a nonzero~$m_i$ such that~$S_t^{m_i}([\beta_i]_q)\in T$. Taking the shift mapping~$S_t$ as bijection, Proposition~\ref{PROP:periodic} and Lemma~\ref{LM:cstdq} imply that~$\beta_i\in k$ for all~$i$ with~$1\leq i \leq m$. The opposite implication follows from the fact that~$c(t)$ is annihilated by the operator~$L_0 =c(t)S_t - c(t+1)$ and the linear operator \[L_{i, j}=\alpha_{i, j}(t)S_t-\alpha_{i,j}(t+1)\] is a telescoper for the fraction~$f_{i, j} = \alpha_{i, j}/(x-\beta_i)^j$ with respect to~$\Delta_{q, x}$ since $\operatorname{dres}(L_{i, j}(f_{i, j}), [\beta_i]_q, j)=0$. Then the LCLM of the~$L_0$ and~$L_{i, j}$ is a telescoper for~$f$ with respect to~$\Delta_{q, x}$. \end{proof} The following theorem is a $q$-analogue of Theorem~\ref{THM:dd}, which has also been shown in~\cite[Theorem 1]{Le2001}. \begin{theorem}\label{THM:qq} Let~$f\in k(t, x)$ be as in~\eqref{EQ:qresform}. Then~$f$ has a telescoper~$L$ in~$k(t)\langle Q_t\rangle$ such that~$L(t, Q_t)(f)= \Delta_{q, x}(g)$ for some~$g\in k(t, x)$ if and only if all the~$\beta_i=c_i t^{r_i}$ with~$r_i\in \bQ$ and~$c_i\in k$. \end{theorem} \begin{proof} Suppose that there exists a nonzero~$L\in k(t)\langle Q_t\rangle$ such that~$L(t, Q_t)(f)=\Delta_{q, x}(g)$ for some~$g\in k(t, x)$. Write~$L=\sum_{\ell=0}^{\rho} e_{\ell}Q_t^{\ell}$ with~$e_{\ell}\in k(t)$ and~$e_0\neq 0$. For any~$\lambda \in \{1, \ldots, m\}$, we consider the rational function \[f_{\lambda} = \sum_{i=1}^m \frac{\alpha_{i, n_{\lambda}}}{(x-\beta_i)^{n_{\lambda}}},\quad \text{where~$\alpha_{\lambda, n_{\lambda}}\neq 0$ by assumption} .\] Without loss of generality, we may assume that the other~$\alpha_{i, n_{\lambda}}$ with~$i\neq \lambda$ are also nonzero. Since the $q$-shift operators~$Q_t, Q_x$ preserve the multiplicity, we have~$L(f_{\lambda})=\Delta_{q, x}(g_{\lambda})$ for some~$g_{\lambda}\in k(t, x)$. By Proposition~\ref{PROP:ratqsum}, all the residues of~$L(f_{\lambda})$ are zero. In particular, the set~$T=\{[\beta_1]_q, \ldots, [\beta_m]_q\}$ satisfies that for any~$i \in \{1, \ldots, m\}$, there exists a nonzero~$m_i$ such that~$Q_t^{m_i}([\beta_i]_q)\in T$. By the analysis in Example~\ref{EX:intqlin}, we conclude that~$\beta_i=c_it^{{p_i}/{m}}$ with~$p_i, m\in \bZ$ and~$c_i\in k$. The opposite implication follows from the fact that~$c(t)$ is annihilated by the operator~$L_0 =cS_t - c(t+1)$ and the linear operator \[L_{i, j}=\alpha_{i, j}(t)Q_t^{m}-q^{-jp_i}\alpha_{i,j}(q^mt)\] is a telescoper for the fraction~$f_{i, j} = \alpha_{i, j}/(x-(c_it^{{p_i}/{m}}))^j$ with respect to~$\Delta_{q, x}$ since~$\operatorname{qres}(L_{i, j}(f_{i, j}), [c_it^{{p_i}/{m}}]_q, j)=0$. Then the LCLM of the~$L_0$ and~$L_{i, j}$ is a telescoper for~$f$ with respect to~$\Delta_{q, x}$. \end{proof} The necessary and sufficient conditions for the existence of telescopers enable us to decide the termination of the Zeilberger algorithm for rational-function inputs. After reducing the given rational function into a residual form, one can detect the existence by investigating the denominator. For instance, we could check whether the denominator factors into two univariate polynomials respectively in~$t$ and~$x$ in the case when~$\partial_t=D_t$ and~$\partial_x=\Delta_x$. Combining the existence criteria with the Zeilberger algorithm yields a complete algorithm for creative telescoping with rational-function inputs. \subsection{Characterization of telescopers}\label{SUBSECT:chartele} We have shown that telescopers exist for a special class of rational functions. Now, we will characterize the linear differential and ($q$-)recurrence operators that could be telescopers for rational functions. Using such a characterization, we will give a direct algebraic proof of a theorem of Furstenberg stating that the diagonal of a rational power series in two variables is algebraic~\cite{Furstenberg1967}. In all of these considerations, residues are still the key. For a rational function~$f\in k(t, x)$, all of the telescopers for~$f$ in~$k(t)\langle D_t\rangle$ form a left ideal in~$k(t)\langle D_t\rangle$, denoted by~$\mathcal{T}_f$. Since the ring~$k(t)\langle D_t\rangle$ is a left Euclidean domain, the monic telescoper of minimal order generates the left ideal~$\mathcal{T}_f$, and we call this generator~\emph{the minimal telescoper} for~$f$. \begin{theorem}\label{THM:telecc} Let~$L(t, D_t)$ be a linear differential operator in~$k(t)\langle D_t\rangle$. Then~$L$ is a telescoper for some~$f\in k(t, x)\setminus D_x(k(t, x))$ such that~$L(f)= D_x(g)$ with~$g\in k(t, x)$ if and only if $L(y(t))=0$ has a nonzero solution algebraic over~$k(t)$. Moreover, if~$L$ is the minimal telescoper for~$f$, then all solutions of~$L(y(t))=0$ are algebraic over~$k(t)$. \end{theorem} \begin{proof} Suppose that there exists~$f\in k(t, x)\setminus D_x(k(t, x))$ such that~$L(f)= D_x(g)$ for some~$g\in k(t, x)$. Since~$f$ is not rational integrable with respect to~$x$, $f$ has a nonzero residue by Proposition~\ref{PROP:ratint}. Since~$L$ is a telescoper for~$f$ with respect to~$D_x$, $L$ vanishes at all residues of~$f$. So~$L(y(t))=0$ has a nonzero algebraic solution in~$\overline{k(t)}$ because any residue of a rational function in~$k(t, x)$ is algebraic over~$k(t)$. Conversely, if~$\alpha\in \overline{k(t)}$ is a nonzero algebraic solution of~$L(y(t))=0$ with minimal polynomial~$P\in k[t, x]$, then~$L$ is a telescoper for the rational function~$f=xD_x(P)/P$ with respect to~$D_x$. Let~$a/b\in k(t, x)$ be the residual form of~$f$ with respect to~$D_x$. All of the residues of~$a/b$ are roots of the polynomial~$R(t, z) = \mbox{resultant}_x(b, a-zD_x(b))\in k(t)[z]$. By the method in~\cite[\S 2]{CormierSingerTragerUlmer2002}, one can construct the minimal operator~$L_R$ in~$k(t)\langle D_t\rangle$ such that~$L_R(\alpha(t))=0$ for all roots of~$R$ in~$\overline{k(t)}$. Moreover, the solutions space of~$L_R$ is spanned by the roots of~$R$. Since~$L_R$ vanishes at all residues of~$f$, $L_R$ is a telescoper for~$f$. If~$L$ is the minimal telescoper for~$f$, then~$L$ divides~$L_R$ on the right. Thus, all solutions of~$L(y(t))=0$ are solutions of~$L_R(y(t))=0$, and therefore algebraic over~$k(t)$. \end{proof} The diagonal~$\operatorname{diag}(f)$ of a formal power series~$f=\sum_{i, j\geq 0} f_{i, j}t^ix^j\in k[[t, x]]$ is defined by \[\operatorname{diag}(f) = \sum_{i\geq 0} f_{i, i} t^i\in k[[t]].\] Using the characterization of telescopers in Theorem~\ref{THM:telecc}, we now give a proof of a theorem of Furstenberg that the diagonal of a rational power series in two variables is algebraic~\cite{Furstenberg1967}. For other proofs, see the papers~\cite{Fliess1974, Gessel1980, Haiman1993} and Stanley's book~\cite[Theorem 6.3.3]{Stanley1999}. Let~$\mathcal {F}=k((x))$ be the quotient field of~$k[[x]]$ and~$\mathcal {F}[[t]]$ be the formal power series over~$\mathcal {F}$. We use the notation~$[x^{-1}](a)$ to denote the coefficient of~$x^{-1}$ in~$a\in \mathcal {F}$. For a formal power series~$g=\sum_{i\geq 0} a_i(x)t^i \in \mathcal {F}[[t]]$, we define \[[x^{-1}](g)=\sum_{i\geq 0} ([x^{-1}](a_i))t^i\in k[[t]],\] and two derivations \[D_t(g) = \sum_{i\geq 0} i a_i(x) t^{i-1}, \quad D_x(g) = \sum_{i\geq 0} D_x(a_i)t^{i}.\] The ring~$\mathcal {F}[[t]]$ then becomes a $k[t, x]\langle D_t, D_x\rangle$-module. By definition, we have \[[x^{-1}](D_t(g)) = D_t([x^{-1}](g)) \quad \text{and} \quad [x^{-1}](t^i(g)) = t^i([x^{-1}](g)) \] for all~$i\in \bN$. By induction, we have~$L([x^{-1}](g))=[x^{-1}](L(g))$ for all~$L\in k[t]\langle D_t\rangle$. Since~$[x^{-1}](D_x(a))=0$ for any~$a\in \mathcal {F}$, we get~$[x^{-1}](D_x(g))=0$ for any~$g\in \mathcal {F}[[t]]$. Let~$f=\sum_{i, j\geq 0} f_{i, j}t^ix^j$ be a formal power series in~$k[[t, x]]$. Then~$F=f(x, t/x)/x$ is in~$\mathcal {F}[[t]]$. Applying~$[x^{-1}]$ to~$F$ yields \[[x^{-1}](F) = [x^{-1}](\sum_{i, j\geq 0} f_{i, j} x^{i-j-1}t^j) = \sum_{j\geq 0} f_{j, j} t^j = \operatorname{diag}(f).\] If~$L\in k[t]\langle D_t\rangle$ be such that~$L(F)=D_x(G)$ for some~$G\in \mathcal {F}[[t]]$, then applying~$[x^{-1}]$ to both sides of~$L(F)=D_x(G)$ yields~$L(\operatorname{diag}(f))=0$. In summary, we have the following lemma. \begin{lemma}\label{LM:diag} Let~$f\in k[[t, x]]$ and~$F=f(x, t/x)/x\in \mathcal {F}[[t]]$. If~$L\in k[t]\langle D_t\rangle$ is a telescoper for~$F$ such that~$L(F)=D_x(G)$ with~$G\in \mathcal {F}[[t]]$, then~$L(\operatorname{diag}(f))=0$. \end{lemma} In the following, we prove Furstenberg's diagonal theorem. \begin{theorem}[Furstenberg, 1967]\label{THM:diag} Let~$f\in k[[t, x]]\cap k(t, x)$. Then the diagonal of~$f$ is a power series algebraic over~$k(t)$. \end{theorem} \begin{proof} Let~$F=f(x, t/x)/x$. Since~$f$ is a rational function in~$k(t, x)$, so is~$F$. Let~$L\in k(t)\langle D_t \rangle$ be the minimal telescoper for~$F$. Since multiplying by an element of~$k[t]$ commutes with the derivation~$D_x$, we can always assume that the coefficients of~$L$ are polynomials in~$k[t]$. By Theorem~\ref{THM:telecc}, all of the solutions of~$L(y(t))=0$ are algebraic over~$k(t)$. So the diagonal of~$f$ is algebraic over~$k(t)$ since~$L(\operatorname{diag}(f))=0$ by Lemma~\ref{LM:diag}. \end{proof} The following example is borrowed from the recent paper by Ekhad and Zeilberger~\cite{EZ2011}, from which one can see how Zeilberger's method of creative telescoping plays a role in solving concrete problems in combinatorics. \begin{example} Let~$s(n)$ be the number of binary words of length~$n$ for which the number of occurrences of~$00$ is the same as that of~$01$ as subwords. Stanley~\cite{Stanley2011} asked for a proof of the following formula \begin{equation}\label{EQ:sf} S(t) \triangleq \sum_{n=0}^{\infty} s(n) t^n = \frac{1}{2}\left(\frac{1}{1-t} + \frac{1+2t}{\sqrt{(1-t)(1-2t)(1+t+2t^2)}} \right). \end{equation} We first show that the generating function~$S(t)$ is an algebraic function over~$k(t)$. The key ingredient is the Goulden-Jackson cluster method~\cite{GJ1979}. Noonan and Zeilberger~\cite{NoonanZeilberger1999} gave an elegant survey of this method together with an efficient implementation. Let~$\mathcal {W}$ be the set of all binary words and let~$\tau_{00}(w), \tau_{01}(w)$ be the numbers of occurrences of~$00$ and~$01$ in~$w\in \mathcal {W}$, respectively. Ekhad and Zeilberger~\cite{EZ2011} define the generating function \[f(t, y, z) = \sum_{w\in \mathcal {W}} t^{\text{length(w)}} y^{\tau_{00}(w)}z^{\tau_{01}(w)}.\] Loading the package~{\sf DAVID{\_}IAN} created by Noonan and Zeilberger to {\sf Maple}, typing {\sf GJstDetail([0, 1], \{[0, 0], [0, 1]\}, t, s)}, and replacing~$s[0, 0], s[0, 1]$ by~$y, z$, respectively, we get an explicit form of~$f(t, y, z)$, \[f(t, y, z) = \frac{(1-y)t +1}{(y-z)t^2-(1+y)t+1},\] which is a rational function of three variables. By definition, the desired generating function~$S(t)$ is the coefficient of~$x^{-1}$ in~$F(t, x) := x^{-1}f(t, x, x^{-1})$. Since~$\tau_{00}(w)$ and~$\tau_{01}(w)$ are bounded by~${\text{length(w)}}$, the function~$F(t, x)$ is an element in the ring~$k((x))[[t]]$. Therefore, the coefficient~$[x^{-1}](F)$ is annihilated by any telescoper for~$F$ in~$k[t]\langle D_t\rangle$. By Theorem~\ref{THM:telecc}, the function~$S(t)$ must be an algebraic function over~$k(t)$. By typing~{DETools[Zeilberger](F, t, x, Dt)} in {\sf Maple}, we get the minimal telescoper~$L$ for~$F$, which is \begin{align*} L =& \left( -1+5\,t-13\,{t}^{2}-30\,{t}^{4}+23\,{t}^{3}+40\,{t}^{5}-40\,{t}^{6}+16\,{t}^{7} \right) {{\it Dt}}^{2} \\ &\, \, + \left( 80\,{t}^{6}-168\,{t}^ {5}+152\,{t}^{4}-88\,{t}^{3}+24\,{t}^{2}-2\,t+2 \right) {\it Dt}\\ &\, \, +48\,{ t}^{5}-72\,{t}^{4}+48\,{t}^{3}-12\,{t}^{2}-6\,t. \end{align*} To show Stanley's formula~\eqref{EQ:sf}, it suffices to verify that~$S(t)$ satisfies the equation~$L(y(t)) = 0$, and check the two initial condition:~$y(0) =1$ and~$D_t(y)(0) = 2$. Moreover, we could also rediscover Stanley's formula by solving the differential equation. Thanks to Zeilberger's method, many classical combinatorial identities now can be proved and rediscovered automatically all by computer. \end{example} Except the case when~$\partial_t=D_t$ and~$\partial_x=D_x$ as above, we will show that telescopers for non-integrable or non-summable rational functions in~$k(t, x)$ have at least one nonzero rational solution in~$k(t)$. Of these 8 cases, 6 follow easily from an examination of some of the proofs above. These cases are considered in Theorem~\ref{THM:telemixed}. The remaining two cases require a slightly more detailed proof and are considered in Theorem~\ref{THM:teleddqq}. \begin{theorem}\label{THM:telemixed} Let~$L\in k(t)\langle \partial_t\rangle$ and~$f\in k(t, x)$ satisfy one of the following conditions: \begin{enumerate} \item $\partial_t = D_t$ and~$f\notin \Delta_x(k(t, x))$; \item $\partial_t = D_t$ and~$f\notin \Delta_{q, x}(k(t, x))$; \item $\partial_t = S_t$ and~$f\notin D_x(k(t, x))$; \item $\partial_t = S_t$ and~$f\notin \Delta_{q, x}(k(t, x))$; \item $\partial_t = Q_t$ and~$f\notin D_x(k(t, x))$; \item $\partial_t = Q_t$ and~$f\notin \Delta_x(k(t, x))$. \end{enumerate} Then~$L(t, \partial_t)$ is a telescoper for some~$f\in k(t, x)$ if and only if~$L(y(t))=0$ has a nonzero rational solution in~$k(t)$. \end{theorem} \begin{proof} Suppose that~$L(y(t))=0$ has a nonzero rational solution~$r(t)$ in~$k(t)$. Then~$L$ is a telescoper for~$f=r(t)/x$ and~$f$ satisfies the assumption above. For the opposite implication, Theorems~\ref{THM:cd}, \ref{THM:cq}, \ref{THM:dc}, \ref{THM:dq}, \ref{THM:qc} and~\ref{THM:qd} imply that the residual form of~$f$ is of the form~$a/b$ such that~$b=b_1(t)b_2(x)$ with~$b_1\in k[t]$ and~$b_2\in k[x]$. Then \[\frac{a}{b} = \sum_{i=1}^m \sum_{j=1}^{n_i} \frac{\alpha_{i, j}}{(x-\beta_i)^j}, \] where~$\alpha_{i,j}\in k(t)$ and~$\beta_i\in k$ are in distinct ($q$-)orbits. If~$L$ is a telescoper for~$f$, then~$L$ is also a telescoper for~$a/b$. Since all the~$\beta_i$ are free of~$t$, we have \[L(a/b) = \sum_{i=1}^m \sum_{j=1}^{n_i} \frac{L(\alpha_{i, j})}{(x-\beta_i)^j}=\partial_x(g), \quad \text{where~$\partial_x\in \{D_x, \Delta_x, \Delta_{q, x}\}$}.\] By Propositions~\ref{PROP:ratint}, \ref{PROP:ratsum}, and~\ref{PROP:ratqsum}, we have~$L(\alpha_{i,j})=0$. Since~$a/b$ is not zero, at least one of the~$\alpha_{i, j}$ is nonzero. Thus~$L(y(t))=0$ has at least one nonzero rational solution in~$k(t)$. \end{proof} \begin{theorem}\label{THM:teleddqq} Let~$L\in k(t)\langle \partial_t\rangle$ and~$f\in k(t, x)$ satisfy one of the following conditions: $(1)$~$\partial_t = S_t$ and~$f\notin \Delta_{x}(k(t, x))$; $(2)$~$\partial_t = Q_t$ and~$f\notin \Delta_{q, x}(k(t, x))$. Then~$L(t, \partial_t)$ is a telescoper for some~$f\in k(t, x)$ if and only if~$L(y(t))=0$ has a nonzero rational solution in~$k(t)$. \end{theorem} \begin{proof} Suppose that~$L(y(t))=0$ has a nonzero rational solution~$r(t)$ in~$k(t)$. Then~$L$ is a telescoper for~$f=r(t)/x$ and~$f$ satisfies the assumption above. For the opposite implication, we only prove the assertion for the first case, that is, when~$L$ and~$f$ satisfies the condition~$(1)$. The remaining assertion follows in a similar manner. Theorem~\ref{THM:dd} implies that the residual form~$a/b$ of~$f$ can be decomposed into \[\frac{a}{b} = \sum_{i=1}^m \sum_{j=1}^{n_i} \frac{\alpha_{i, j}}{(x-\beta_i)^j},\] where~$\alpha_{i, j}\in k(t)$ and~$\beta_i =\frac{\lambda_i}{\mu_i}t+c_i$ with~$c_i\in k$, $\lambda_i\in \bZ$ and~$\mu_i\in \bN$ such that~$\gcd(\lambda_i, \mu_i)=1$ and the~$\beta_i$ are in distinct~$\bZ$-orbits. If~$L\in k(t)\langle D_t\rangle$ is a telescoper for~$f$, then~$L$ is a telescoper for~$a/b$. Moreover, $L$ is a telescoper for each fraction~$f_{i, j} = {\alpha_{i, j}}/{(x-\beta_i)^j}$. We claim that the operator~$L_{i, j} := \alpha_{i,j}(t)S_t^{\mu_i} - \alpha_{i, j}(t+\mu_i)\in k(t)\langle D_t \rangle$ is the minimal telescoper for~$f_{i, j}$ with respect to~$\Delta_x$. In fact, $L_{i, j}$ is a telescoper for~$f_{i,j}$ as shown in the proof of~Theorem~\ref{THM:dd}. It remains to show the minimality. Assume that there exists a telescoper~$\tilde{L}_{i, j}$ of order less than~$\mu_i$ for~$f_{i, j}$. Write~$\tilde{L}_{i, j}=\sum_{\ell=0}^{\mu_i-1} e_{\ell}S_t^{\ell}$. Then \[\tilde{L}_{i, j}(f_{i, j}) = \sum_{\ell=0}^{\mu_i-1} \frac{e_{\ell} \alpha_{i, j}(t+\ell)}{(x-(\frac{\lambda_i}{\mu_i}t+\frac{\lambda_i}{\mu_i}\ell + c_i))^j}.\] Since~$\gcd(\lambda_i, \mu_i)=1$ and~$\ell\in \{0, \ldots, \mu_i-1\}$, the values~$\frac{\lambda_i}{\mu_i}t+\frac{\lambda_i}{\mu_i}\ell + c_i$ are in distinct~$\bZ$-orbits. If~$\tilde{L}_{i, j}(f_{i, j})$ is rational summable, then all the residues~$e_{\ell} \alpha_{i, j}(t+\ell)$ are zero by Proposition~\ref{PROP:ratsum}. Since~$\alpha_{i,j} \neq 0$, we have~$\tilde{L}_{i,j}$ is a zero operator. The claim holds. Since~$L$ is a telescoper for~$f_{i, j}$, $L_{i, j}$ divides~$L$ on the right. Note that the rational function~$\alpha_{i, j}\in k(t)$ is a nonzero solution of~$L_{i, j}(y(t))=0$. Thus, $L$ has at least one nonzero rational solution in~$k(t)$. \end{proof} {\section*{Appendix}\label{SECT:appendix} In this appendix, we present proofs of Propositions~\ref{PROP:afrde} and~\ref{PROP:aflqe}. Let~$K \subset E$ be difference fields of characteristic zero with automorphism~$\sigma$ and assume that the constants $E^\sigma$ of~$E$ are in~$K$. Furthermore assume that~$E$ is algebraically closed. \begin{lemma}\label{LM:shiftclose} Let~$u\in E$ be algebraic over~$K$ and assume that~$u$ satisfies a homogeneous linear difference equation over~$K$. Then there exists a field~$F \subset E$ with~$\sigma(F) = F$, $K \subset F$, $[F : K]< \infty$, and~$u \in F$. \end{lemma} \begin{proof} Let $u$ satisfy \begin{equation}\label{eqn1} \sigma^n(u) + b_{n-1} \sigma^{n-1}(u) + \cdots + b_0u = 0 \end{equation} with~$b_i \in K, b_0 \neq 0$ and let~$F = K(u, \sigma(u), \ldots , \sigma^{n-1}(u))$. We have that~$[F:K] < \infty$ since for any~$i$, $\sigma^i(u)$ is algebraic over~$K$. To see that~$\sigma(F) \subset F$ it is enough to show that~$\sigma^i(u) \in F$ for all~$i$. This is certainly true for~$i = 0, \ldots n$. If~$i > n$, apply~$\sigma^{i-n}$ to equation~\eqref{eqn1} and proceed by induction to conclude~$\sigma^{i}(u) \in F$. If~$i <0$ apply~$\sigma^{i}$ and proceed by induction to conclude~$\sigma^i(u) \in F$. \end{proof} \begin{lemma}\label{LM:aflre1} Let~$K = k(t)$, where~$k$ is algebraically closed. Let~$(E, \sigma)$ be a difference field such that~$K \subset E$, $\sigma(t) = t+1$ and~$[E : K] < \infty$. The~$E = K$. \end{lemma} \begin{proof} Let~$n = [E:K]$ and~$g$ be the genus of~$E$. The Riemann-Hurwitz formula (see~\cite[p.\ 106]{Chevalley1951} or~\cite[p.\ 125]{Fulton1989}) yields \begin{equation}\label{RHfmla} 2g-2 = -2n + \sum_P(e(P) - 1), \end{equation} where the sum is over all places~$P$ of~$E$ and~$e(P)$ is the ramification index of~$P$ with respect to~$K$. There are only a finite number of places~$Q$ of~$K$ over which places of~$E$ ramify and the automorphism~$\sigma$ leaves the set of such places invariant. On the other hand, the only finite set of places of~$K$ that is left invariant by~$\sigma$ is the place at infinity. Therefore, if~$P$ is a place of~$E$ with~$e(P) > 1$, then~$P$ lies above the place at infinity. Note that for any place~$Q$ of~$K$, Theorem~1 of~\cite[p.\ 52]{Chevalley1951} implies (under our assumptions) that \begin{equation}\label{ramsum} \sum_{\mbox{$P$ lies above~$Q$}} e(P) = n. \end{equation} Therefore we have \begin{eqnarray*} 2g-2 & = & -2n + \sum_{\mbox{$P$ lies above~$\infty$}}(e(P) - 1)\\& =& -2n +n-t \\ & =& -n-t, \end{eqnarray*} where~$t$ is the number of places above infinity. Since~$n$ and~$t$ are both positive integers and~$g$ is nonnegative, we must have~$g=0$ and~$n = t = 1$. In particular, since~$n = 1$, we have~$E= K$. \end{proof} \noindent{\underline{\emph{Proof of Proposition~\ref{PROP:afrde}.}} Suppose that~$\alpha(t)$ satisfies the linear recurrence relation \[S_t^n(\alpha) + a_{n-1} S_t^{n-1} (\alpha) + \cdots + a_0 \alpha = 0,\] where~$a_i\in k(t)$. By Lemma~\ref{LM:shiftclose}, the field~$E = k(t)(\alpha, S_t(\alpha), \ldots, S_t^{n-1}(\alpha)) \subset \overline{k(t)}$ is a difference field extension of~$k(t)$. Since~$[E:k(t)]<\infty$, $E=k(t)$ by Proposition~\ref{LM:aflre1}. Thus~$\alpha\in k(t)$. \hfill $\Box$ \begin{remark} Proposition~\ref{PROP:afrde} has been shown in~\cite[Theorem 1]{Benzaghou1992},~\cite[Prop.\ 4.4]{vdPutSinger1997} and~\cite[Theorem 5.2]{Bell2008}. The proof in~\cite[Theorem 5.2]{Bell2008} is based on analytic properties of algebraic functions.\\[0.1in] In this proposition, we assume that~$\alpha(t)$ satisfies a polynomial equation over~$k(t)$ and \underline{lies in a field}. This latter condition cannot be weakened without weakening the conclusion. For example, the sequence~$y=(-1)^n$ satisfies $y^2-1 = 0$ but~$k(t)[y]$ is a ring with zero divisors. The above references give a complete characterization of sequences satisfying both linear recurrences and polynomial equations. \end{remark} The following result is a~$q$-analogue of Lemma~\ref{LM:aflre1}. \begin{lemma}\label{LM:aflqre1} Let~$K = k(t)$, where~$k$ is algebraically closed. Let~$(E, \sigma)$ be a difference field such that~$K \subset E$, $\sigma(t) = qt$ with~$q \in k\setminus\{0\}$ and not a root of unity, and~$[E:K] < \infty$. Then~$E = k(t^{{1}/{n}})$ for some positive integer~$n$. \end{lemma} \begin{proof} Let~$[E:K] = n$ and~$g$ be the genus of~$E$. We again consider the set of places of~$K$ over which places of~$E$ ramify. This set is left invariant by~$\sigma$ and so must be a subset of the set containing the place at~$0$ and the place at~$\infty$. Therefore, ramification can occur only at~$0$ and~$\infty$. Equations~\eqref{RHfmla} and~\eqref{ramsum} imply \begin{eqnarray*} 2g-2 & = & -2n + \sum_{\mbox{~$P$ lies above~$0$}}(e(P) - 1) + \sum_{\mbox{$P$ lies above~$\infty$}}(e(P) - 1)\\ &=&-2n +2n-t_0-t_\infty \\ &=& -t_0-t_\infty \end{eqnarray*} where~$t_0, t_\infty$ are the number of places above~$0$ and~$\infty$. Since~$t_0$ and~$t_\infty$ are positive and~$g$ is nonnegative, we must have that~$g=0$ and~$t_0 = t_\infty = 1$. Therefore, $E$ has one place~$P_0$ over~$0$ with~$e_{P_0} = n$ and one place~$P_\infty$ over~$\infty$ with~$e_{P_\infty} = n$. Writing divisors multiplicatively, Riemann's Theorem (\cite[p.\ 22]{Chevalley1951}) implies that \begin{equation*} l(P_0P^{-1}_\infty) \geq d(P_0P^{-1}_\infty) -g+1 = 0-0+1 = 1 \end{equation*} where~$l(P_0P^{-1}_\infty)$ is the dimension of the space of elements of~$E$ which are~$\equiv 0 \mod {P_0P^{-1}_\infty}$. Note that since the degree~$P_0P^{-1}_\infty$ is~$0$, this latter condition implies that any such element has~$P_0P^{-1}_\infty$ as its divisor. Therefore, there exists an element~$y \in E$ whose divisor is~$P_0P^{-1}_\infty$. Note that the element~$t$ has divisor~$P_0^nP^{-n}_\infty$ and therefore~$y^nt^{-1}$ must be in~$k$. Therefore~$y = c\, t^{{1}/{n}}$ for some~$c \in k$. Finally, Theorem 4 of~\cite[p.\ 18]{Chevalley1951} states that~$[E:k(y)]$ equals the degree of the divisor of zeros of~$y$, that is, $[E:k(y)] = 1$. Therefore~$ E = k(y) = k(t^{{1}/{n}})$. \end{proof} \noindent{\underline{\emph{Proof of Proposition~\ref{PROP:aflqe}.}} Suppose that~$\alpha(t)$ satisfies the linear $q$-recurrence relation \[Q_t^n(\alpha) + a_{n-1} Q_t^{n-1} (\alpha) + \cdots + a_0 \alpha = 0,\] where~$a_i\in k(t)$. By Lemma~\ref{LM:shiftclose}, the field~$E = k(t)(\alpha, Q_t(\alpha), \ldots, Q_t^{n-1}(\alpha)) \subset \overline{k(t)}$ is a difference field extension of~$k(t)$. Since~$[E:k(t)]<\infty$, $E=k(t^{1/n})$ by Lemma~\ref{LM:aflqre1}. Thus~$\alpha\in k(t^{1/n})$.} \hfill $\Box$ \bibliographystyle{plain}

https://arxiv.org/abs/1406.1984

Discrete Hardy-type Inequalities

This paper studies the Hardy-type inequalities on the discrete intervals. The first result is the variational formulas of the optimal constants. Using these formulas, one may obtain an approximating procedure and the known basic estimates of the optimal constants. The second result, which is the main innovation of this paper, is about the factor of basic upper estimates. An improved factor is presented, which is smaller than the known one and is best possible. Some comparison results are included for comparing the optimal constants on different intervals.

\section{Introduction} For given two constants $p$ and $q$ with $1< p\leqslant q< \infty$, two positive sequences $\mathbf{u}$ and $\mathbf{v}$ on a discrete interval $[1, N]:= \{1, 2, \dots, N \}$ with $N\leqslant +\infty$, here is the discrete Hardy-type inequality: \begin{equation}\label{Hardy} \left[\sum_{n= 1}^N u_n \left(\sum_{i=1}^n x_i \right)^q \right]^{1/q} \leqslant A \left(\sum_{n=1}^N v_n x_n^p \right)^{1/p}, \end{equation} where $\mathbf{x}$ is an arbitrary non-negative sequence on $[1, N]$. For saving notations, the constant $A$ is assumed to be optimal. The purpose of this paper is two-fold. First, we give some variational formulas of the optimal constants. The primary applications of the variational formulas are the approximating procedure and the basic estimates. It is necessary to review the advance of the basic estimates in recent research, cf. \cite{Bliss, Chen4, Chen3, Maz'ya, Opic}. In continuous case, the following result, due to B. Opic \rf{Opic}{Theorem 1.14} and V. G. Maz'ya \rf{Maz'ya}{Theorem 1, pp. 42-43}, is well known \bg{equation}\label{basic} B \leqslant A \leqslant \tilde{k}_{q, p} B, \end{equation} where $B$ is a quantity described by $N$, $p$, $q$, $\mathbf{u}$ and $\mathbf{v}$, the factor $\tilde{k}_{q, p}$ is a constant defined by $p$ and $q$: \bg{equation} \label{tilde k_qp} \tilde{k}_{q,p} = \left(1+ \frac{q}{p^*} \right)^{1/q} \left(1+ \frac{p^*}{q} \right)^{1/p^*}, \end{equation} here $p^*$ is the conjugate number of $p$, i.e. $1/p + 1/{p^*}= 1$. In particular, $\tilde{k}_{p,p}= p^{1/p}(p^*)^{1/{p^*}}$. Afterwards, Chen \rf{Chen3}{Theorem 2.1} gets the same conclusions through the variational formulas of the optimal constants. Furthermore, there is an approximating procedure \rf{Chen3}{Theorem 2.2} based on the variational formulas, which can improve the estimates of the optimal constants step by step. In discrete context, when $p=q$, Chen, Wang and Zhang \cite{Chen2} arrive the corresponding variational formulas and basic estimates which are similar to (\ref{basic}), of course, $B$ must be adjusted appropriately in discrete case. When $p\ne q$, Mao \rf{Mao}{Proposition A.1} gets the similar result, but the factor of basic upper estimates is $p^{1/q} (p^*)^{1/p^*}$, which is a little coarser than $\tilde{k}_{q,p}$. Our first destination is to show the corresponding variational formulas in discrete context with the condition of $p\neq q$. Later on, as applications of these formulas, we obtain the basic estimates and the approximating procedure. Overall, these results can be regarded as an extension of the studies in continuous context \cite{Chen3}. Second, we study the upper bounds of the basic estimates of the optimal constants in discrete case. Our result is the factor $\tilde{k}_{q, p}$ in (\ref{basic}) can be improved to $k_{q, p}$: \bg{equation} \label{k_qp} k_{q,p} = \left(\frac{r}{B(\frac{1}{r}, \frac{q- 1}{r})} \right)^{1/p- 1/q}, \end{equation} where $B(a, b)= \int_0^1 x^{a-1} (1- x)^{b- 1} \text{\rm d} x$ is the Beta function and $r= q/p -1$. Moreover, our result shows that the factor is best possible and is consistent with the result of continuous case. In continuous case, the improvement has been worked out, cf. \rf{Bennett3}{Theorem 8}, \rf{Manakov}{Theorem 2}, and \rf{Kufner}{pp. 45-47}. The key is the result of Bliss \cite{Bliss}, which gives an integral inequality that the optimal constants can be attained. However, the analogue of the conclusion in the discrete context is nontrivial, as mentioned in \rf{Bennett3}{page 170, two lines above (61)}, \textquotedblleft I have been unable to prove the discrete analogue of Theorem 8\textquotedblright (here the last result is the continuous case). We are lucky to be able to prove this conclusion which constitutes the second part of this paper. When $p=q$, it is well known that the factor $\tilde{k}_{q,p}$ is sharp (see for instance \rf{Hardy_book}{Theorem 326 and 327}). Note that if we allow $q\rightarrow p$, by the identity $$ \lim_{r\rightarrow 0^{+}} B^r \left(\frac{a}{r}, \frac{b}{r} \right) = \frac{a^a b^b}{(a+ b)^{a+ b}}, \qquad (a,b >0), $$ we have \bg{equation}\label{k_pp} k_{p,p}= \tilde{k}_{p,p}= p^{1/p}(p^*)^{1/{p^*}}. \end{equation} It means our improved factor is consistent with the original one when $p= q$. Thus, our main results are devoted to the case of $p< q$. It is a long time for the research about Hardy-type inequalities, which are one of the major themes of harmonic analysis and represent useful tools e.g. in the theory and practice of differential equations, in the theory of approximation etc. From probabilistic consideration, these inequalities are important tools to study the convergence rate of the corresponding processes. These are the origin and motive of this study. This paper is organized as follows. The rest of this section, we give some notations and definitions, then illustrate the main results. In Section 2 and Section 3, we prove the conclusions on discrete half line (i.e. $N= \infty$). The case of finite interval (i.e. $N< \infty$) will be handled unitedly in the final section, which gives some comparison results for the optimal constants and their basic estimates on different intervals. For the simplicity of illustration, we need some notations. Let $\hat{v}_i = v_i^{1- p^*}, 1\leqslant i\leqslant N$. For any sequence $\mathbf{x}$ on $[1, N]$, define an operator $H$: \bg{equation}\label{H} H\mathbf{x}(n)= \begin{cases} 0, & n= 0, \\ \sum_{i=1}^n x_i, & n=1, 2, \cdots, N, \end{cases} \end{equation} which means the partial summation of $\mathbf{x}$. The following notations are used frequently: $$ \alpha} \def\bz{\beta \wedge \bz = \min \{ \alpha} \def\bz{\beta, \bz \}, \qquad \alpha} \def\bz{\beta \vee \bz = \max \{ \alpha} \def\bz{\beta, \bz \}. $$ Set $$ {\scr A}} \def\bq{{\scr B}[1, N]= \{\mathbf{x}: \text{$x_1>0$, and $x_i\geqslant 0$ for $i=2, \dots, N$} \}. $$ In this article, we use the convention $1/0 = \infty$. Then the optimal constant $A$ can be denoted by the following variational formula: \begin{equation}\label{A} A = \sup_{\mathbf{x} \in {\scr A}} \def\bq{{\scr B}[1, N]} \frac{\left[\sum_{n= 1}^N u_n \left(\sum_{i=1}^n x_i \right)^q \right]^{1/q}}{\left(\sum_{n= 1}^N v_n x_n^p \right)^{1/p}} = \sup_{\mathbf{x} \in {\scr A}} \def\bq{{\scr B}[1, N]} \frac{\| H\mathbf{x} \|_{l^q(u)}}{\| \mathbf{x} \|_{l^p(v)}}, \end{equation} where $\| \mathbf{x} \|_{l^q(u)}= \left[ \sum_{n=1}^{N} u_n x_n^q \right]^{1/q}$ and similarly to $\| \mathbf{x} \|_{l^p(v)}$. For upper estimates, define the single summation operator $I^*$ and the double summation operator $I\!I^*$ as: \begin{align} I_n^*(\mathbf{x}) &= \frac{\hat{v}_n}{x_n} \left(\sum_{i=n}^N u_i (H\mathbf{x}(i))^{q/p^*} \right)^{p^*/q}, \label{I*} \\ I\!I_n^*(\mathbf{x}) &= \frac{1}{H\mathbf{x}(n)} \sum_{i=1}^n \hat{v}_i \left(\sum_{j=i}^N u_j (H\mathbf{x}(j))^{q/p^*} \right)^{p^*/q}, \label{II*} \end{align} with domain ${\scr A}} \def\bq{{\scr B}[1, N]$. For lower estimates, there are some differences: \begin{align} I_n (\mathbf{x})&= \frac{\hat{v}_n}{x_n} \left(\sum_{i=n}^N u_i (H\mathbf{x}(i))^{q- 1} \right)^{p^*- 1}, \label{I} \\ I\!I_n (\mathbf{x})&= \frac{1}{H \mathbf{x} (n)} \sum_{i=1}^n \hat{v}_i \left(\sum_{j=i}^N u_j (H\mathbf{x}(j))^{q- 1} \right)^{p^*- 1}. \label{II} \\ \end{align} It is easy to see that $I\!I^*= I\!I$ and $I^*= I$ when $p= q$. To avoid the non-summability problem, the domain of $I$ and $I\!I$ have to be modified to: $$ {\scr A}} \def\bq{{\scr B}_0 [1, N]= \left\{\mathbf{x} \in {\scr A}} \def\bq{{\scr B}[1, N]: \sum_{i=1}^N v_i x_i^p < \infty \right\}. $$ With these notations, we can give the main conclusions of the variational formulas of the optimal constants. \nnd\begin{thm1}\label{Var} {\cms The optimal constant $A$ in the Hardy-type inequality {\rm (\ref{Hardy})} satisfies (i) upper estimates: \begin{equation} A\leqslant \inf_{\mathbf{x}\in {\scr A}} \def\bq{{\scr B}[1, N]} \left(\sup_{n\in [1, N]} I\!I_n^*(\mathbf{x}) \right)^{1/p^*} = \inf_{\mathbf{x}\in {\scr A}} \def\bq{{\scr B}[1, N]} \left(\sup_{n\in [1, N]} I_n^*(\mathbf{x}) \right) ^{1/p^*}. \end{equation} (ii) lower estimates: \begin{equation} \aligned A&\geqslant \sup_{\mathbf{x}\in {\scr A}} \def\bq{{\scr B}_0 [1, N]} \|\mathbf{x}\|_{l^p(v)}^{p/q -1} \left(\inf_{n\in [1,N]} I\!I_n (\mathbf{x})\right)^{{(p-1)}/q} \\ &\geqslant \sup_{\mathbf{x}\in {\scr A}} \def\bq{{\scr B}_0 [1, N]} \|\mathbf{x}\|_{l^p(v)}^{p/q -1} \left(\inf_{n\in [1,N]} I_n (\mathbf{x})\right)^{{(p-1)}/q}. \\ \endaligned \end{equation} } \end{thm1} Using Theorem \ref{Var} on a appropriate test function, we can obtain the basic estimates (\ref{basic}). To be specific, that is: \nnd\begin{crl1}\label{basic estimates} {\cms The inequality {\rm (\ref{Hardy})} holds for every $\mathbf{x}\in {\scr A}} \def\bq{{\scr B}[1, N]$ if and only if $B< \infty$, where \bg{equation}\label{B} B= \sup_{n\in [1, N]} \left( \sum_{i=1}^n \hat{v}_i \right)^{1/p^*} \left(\sum_{j=n}^N u_j \right)^{1/q}. \end{equation} Moreover, we have $$ B\leqslant A\leqslant \tilde{k}_{q,p}B, $$ where $\tilde{k}_{q,p}$ is defined as (\ref{tilde k_qp}), which is independent of $\mathbf{u}$, $\mathbf{v}$ and $N$. } \end{crl1} Roughly speaking, the conclusion of Corollary \ref{basic estimates} is from the first iteration through an appropriate test function. Moreover, we can improve the estimates step by step through multiple iterations on this test function. The following corollary is based on this idea, which is of great significance to numerical computation. \nnd\begin{crl1}\label{approximating} {\cms (i) Define $$ \aligned x^{(1)}_n &= (H \mathbf{\hat{v}} (n))^\alpha} \def\bz{\beta - (H \mathbf{\hat{v}} (n- 1))^\alpha} \def\bz{\beta, \\ x^{(m+ 1)}_n &= \hat{v}_n \left(\sum_{i= n}^N u_i (H \mathbf{x}^{(m)} (i))^{q/{p^*}} \right)^{{p^*}/q}, \endaligned $$ where $\alpha} \def\bz{\beta= q/(p^*+ q)$. If $\dz_1:= \sup_{n\in [1, N]} \left( I\!I_n (\mathbf{x}^{(1)}) \right)^{1/p^*}< \infty$, define a sequence as: \bg{equation} \dz_m= \sup_{n\in [1, N]} \left( I\!I^*_n(\mathbf{x}^{(m)})\right)^{1/p^*}, \quad m=2, 3, \dots. \end{equation} Otherwise, define $\dz_m \equiv \infty$. Then $\dz_m$ is a non-increasing sequence (denote $\dz_\infty$ be the limit of $\dz_m$) and we have $$ A\leqslant \dz_\infty \leqslant \cdots \leqslant \dz_1 \leqslant \tilde{k}_{q,p} B. $$ (ii) Fix $k\in [1, N]$, define $$ \aligned y^{(k, 1)}_n &= \begin{cases} \hat{v}_n, & 1\leqslant n\leqslant k, \\ 0, & n> k, \end{cases} \\ y^{(k, m+1)}_n &= \hat{v}_n \left(\sum_{i=n}^N \left(H \mathbf{y}^{(k, m)} (i) \right)^{q- 1} \right)^{p^*- 1}, \endaligned $$ and define a sequence as $$ \aligned \widetilde{\dz}_m&= \sup_{k\in [1, N]} \big\| \mathbf{y}^{(k, m)} \big\|_{l^p (v)}^{p/q -1} \left(\inf_{n\in [1, N]} I\!I_n \left( \mathbf{y}^{(k, m)} \right) \right)^{{(p- 1)}/q}, \\ \overline{\dz}_m&= \sup_{k\in [1, N]} \frac{\left[\sum_{n= 1}^N u_n \left( H \mathbf{y}^{(k, m)} (n) \right)^q \right]^{1/q}}{\left[\sum_{n=1}^N v_n \left(y_n^{(k, m)} \right)^p \right]^{1/p}}. \\ \endaligned $$ Then we have $A\geqslant \widetilde{\dz}_m \vee \overline{\dz}_m$ for all $m\geqslant 1$. } \end{crl1} Another main result of this paper is about the factor in (\ref{basic}). Just like the continuous case, the factor of the basic upper estimates can be improved. Furthermore, we can prove the improved factor is best possible. This result is described in detail below. \nnd\begin{thm1}\label{kqpB} {\cms The basic upper estimates can be improved to: \begin{equation} A\leqslant k_{q, p}B, \end{equation} where $k_{q,p}$ is defined as (\ref{k_qp}). In particular, when $N= \infty$ and $\sum_{i=1}^\infty \hat{v}_i = \infty$, the factor $k_{q,p}$ is sharp. } \end{thm1} \section{Proof of Theorem \ref{Var} } In this section, we always assume $N= \infty$, and the case of the finite interval will be discussed in Section \ref{Interval}. The first proposition is about the property of the sequences which reach the equality case of (\ref{Hardy}). This result is useless in the proof of Theorem \ref{Var}, but is necessary to study the property of the optimal constants. \nnd\begin{prp1}\label{decreasing} {\cms If the optimal constant $A$, appearing in (\ref{Hardy}), is attained by some non-negative sequence $\mathbf{x}$. Define $w_n = \hat{v}_n^{-1} x_n$, then the sequence $\mathbf{w}$ is decreasing. } \end{prp1} \medskip \noindent {\bf Proof}. With the definition of $\mathbf{w}$, we can rewrite the Hardy- type inequalities (\ref{Hardy}) as: \bg{equation}\label{Hardy*} \left[ \sum_{n=1}^\infty u_n \left(\sum_{i=1}^n \hat{v}_i w_i \right)^q \right]^{1/q}\leqslant A \left(\sum_{n=1}^\infty \hat{v}_n w_n^p \right)^{1/p}, \end{equation} and $A$ is attained at $\mathbf{w}$. The idea in the remainder of this proof is from Bennett \rf{Bennett1}{Section 3}. Using reduction to absurdity, assume there exist integers $i$ and $j$ with $1\leqslant i< j< \infty$, but $w_i< w_j$. We can construct a new sequence $\mathbf{w'}$ from $\mathbf{w}$ as \begin{equation}\label{decreasing1} w'_i= w'_j= w_0, \end{equation} where $w_0$ satisfies \begin{equation}\label{decreasing2} (\hat{v}_i+ \hat{v}_j) w_0^p= \hat{v}_i w_i^p+ \hat{v}_j w_j^p. \end{equation} On the one hand, from (\ref{decreasing2}), the right side of (\ref{Hardy*}) is unchanged when $\mathbf{w}$ is replaced by $\mathbf{w'}$. On the other hand, since $p> 1$ and (\ref{decreasing2}), we have \bg{equation}\label{decreasing3} w_i< w_0< w_j. \end{equation} \bg{equation}\label{decreasing4} \hat{v}_i w_i + \hat{v}_j w_j < (\hat{v}_i + \hat{v}_j) w_0. \end{equation} Combine (\ref{decreasing3}) and (\ref{decreasing4}), we obtain $$ \sum_{k=1}^n \hat{v}_k w_k < \sum_{k=1}^n \hat{v}_k w'_k, \qquad \forall n \geqslant 1. $$ It means the left side of (\ref{Hardy*}), increases strictly when $\mathbf{w}$ is replaced by $\mathbf{w'}$. Hence $A$ is not attained at $\mathbf{w}$, which is a contradiction. \quad $\square$ \medskip \noindent {\bf Proof of Theorem \ref{Var}.} The briefing of the proof of Theorem \ref{Var} is given as follows. (a) First, we need to verify the relation of the single summation operator $I^*$ and the double summation operator $I\!I^*$: $$ \inf_{\mathbf{x} \in {\scr A}} \def\bq{{\scr B}[1, \infty)} \left[\sup_{n\in [1, \infty)} I\!I_n^*(\mathbf{x}) \right]^{1/p^*} = \inf_{\mathbf{x} \in {\scr A}} \def\bq{{\scr B}[1, \infty)} \left[\sup_{n\in [1, \infty)} I_n^*(\mathbf{x}) \right]^{1/p^*}. $$ For any $\mathbf{x} \in {\scr A}} \def\bq{{\scr B}[1, \infty)$, as an application of the proportional property, we get $$ \aligned \sup_{n\in [1, \infty)} I\!I_n^*(\mathbf{x}) &= \sup_{n\in [1, \infty)} \frac{1}{H \mathbf{x} (n)} \left[ \sum_{i= 1}^{n} \hat{v}_i \left(\sum_{j=i}^{\infty} u_j (H \mathbf{x} (j))^{q/{p^*}} \right)^{{p^*}/q}\right] \\ &\leqslant \sup_{n\in [1, \infty)} \frac{1}{x_n} \left[ \hat{v}_n \left(\sum_{i=n}^{\infty} u_i (H \mathbf{x} (i))^{q/{p^*}} \right)^{{p^*}/q}\right] \\ &= \sup_{n\in [1, \infty)} I_n^*(\mathbf{x}). \endaligned $$ Hence, we have $$ \inf_{\mathbf{x} \in {\scr A}} \def\bq{{\scr B}[1, \infty)} \left[\sup_{n\in [1, \infty)} I\!I_n^*(\mathbf{x}) \right]^{1/p^*} \leqslant\inf_{\mathbf{x} \in {\scr A}} \def\bq{{\scr B}[1, \infty)} \left[\sup_{n\in [1, \infty)} I_n^*(\mathbf{x}) \right]^{1/p^*}. $$ On the other hand, for any $\mathbf{x}\in {\scr A}} \def\bq{{\scr B}[1, \infty)$, define $$ y_n= \hat{v}_n \left(\sum_{i= n}^{\infty} u_i (H \mathbf{x} (i))^{q/{p^*}} \right)^{{p^*}/q}. $$ Obviously, $y_n> 0$ on $[1, \infty)$, then $\mathbf{y} \in {\scr A}} \def\bq{{\scr B}[1, \infty)$. Again, using the proportional property, we have $$ \aligned \sup_{n\in [1, \infty)} I_n^*(\mathbf{y}) &= \sup_{n\in [1, \infty)} \left[ \frac{\sum_{i=n}^\infty u_i (H \mathbf{y} (i))^{q/{p^*}}}{\sum_{i=n}^\infty u_i (H \mathbf{x} (i))^{q/{p^*}}} \right]^{{p^*}/q} \\ &\leqslant \sup_{n\in [1, \infty)} \frac{1}{H \mathbf{x} (n)} \sum_{i= 1}^n \hat{v}_i \left(\sum_{j= i}^\infty u_j (H \mathbf{x} (j))^{q/{p^*}} \right)^{{p^*}/q} \\ &= \sup_{n\in [1, \infty)} I\!I_n^*(\mathbf{x}). \endaligned $$ Since $\mathbf{x}$ is arbitrary, we obtain the conclusion we need. (b) The next step is to show the upper estimates of the optimal constants. Assume $A$ is attained at a non-negative sequence $\mathbf{a}$. For each positive sequence $\mathbf{h}$, as an application of the H\"older inequality and the H\"older-Minkowski inequality, we have $$ \aligned \sum_{n= 1}^{\infty} u_n (H \mathbf{a} (n))^q &= \sum_{n= 1}^{\infty} u_n \left(\sum_{i= 1}^n a_i v_i^{1/p} h_i^{-1} v_i^{-1/p} h_i \right)^q \\ &\leqslant \sum_{n= 1}^{\infty} u_n \left(\sum_{i= 1}^n a_i^p v_i h_i^{-p} \right)^{q/p} \left(\sum_{k= 1}^n v_k^{-{p^*}/{p}} h_k^{p^*} \right)^{q/{p^*}} \\ &\leqslant \left\{ \sum_{n= 1}^{\infty} a_n^p v_n h_n^{-p} \left[\sum_{i= n}^{\infty} u_i \left(\sum_{k= 1}^i \hat{v}_k h_k^{p^*} \right)^{q/{p^*}} \right]^{p/q} \right\}^{q/p}. \endaligned $$ At the last step, we use the H\"older-Minkowski inequality, which needs the condition $p< q$. In particular, when $p= q$, it is Fubini theorem. Now, making a power $1/q$, we get \begin{align}\label{==1} \left[\sum_{n= 1}^{\infty} u_n (H \mathbf{a} (n))^q \right]^{1/q} &\leqslant \left\{ \sum_{n= 1}^{\infty} a_n^p v_n h_n^{-p} \left[\sum_{i= n}^{\infty} u_i \left(\sum_{k= 1}^i \hat{v}_k h_k^{p^*} \right)^{q/{p^*}} \right]^{p/q} \right\}^{1/p} \notag \\ &\leqslant \sup_{n\in [1, \infty)} \left[ \frac{1}{h_n^q} \sum_{i=n}^{\infty} u_i \left(\sum_{k= 1}^i \hat{v}_k h_k^{p^*} \right)^{q/{p^*}} \right]^{1/q} \left(\sum_{j= 1}^{\infty} v_j a_j^p \right)^{1/p}. \end{align} For any $\mathbf{x} \in {\scr A}} \def\bq{{\scr B}[1, \infty)$, let $$ h_n= \left( \sum_{i=n}^{\infty} u_i (H \mathbf{x} (i))^{q/{p^*}} \right)^{1/q}, $$ by the proportional property, we have $$ \aligned \sup_{n\in [1, \infty)} & \left[\frac{1}{h_n^q} \sum_{i=n}^{\infty} u_i \left(\sum_{j= 1}^i \hat{v}_j h_j^{p^*} \right)^{q/{p^*}} \right]^{1/q} \\ &\leqslant \sup_{n\in [1, \infty)} \left[\frac{1}{H \mathbf{x} (n)} \cdot \sum_{i= 1}^n \hat{v}_i \left(\sum_{j= i}^{\infty} u_j (H \mathbf{x} (j))^{q/p^*} \right)^{p^*/q} \right]^{1/p^*} \\ &= \sup_{n\in [1, \infty)} {I\!I_n^*(\mathbf{x})}^{1/p^*}. \endaligned $$ Inserting this formula into (\ref{==1}), we obtain $$ A = \frac{\left( \sum_{n= 1}^\infty u_n (H \mathbf{a} (n))^q \right)^{1/q}}{\left( \sum_{n= 1}^\infty v_n a_n^p \right)^{1/p}}\leqslant \sup_{n\in [1, \infty)} {I\!I_n^*(\mathbf{x})}^{1/p^*}. $$ Since $\mathbf{x}$ is arbitrary, it follows that $$ A \leqslant \inf_{\mathbf{x} \in {\scr A}} \def\bq{{\scr B}[1, \infty)} \sup_{n \in [1, \infty)} {I\!I_n^*(\mathbf{x})}^{1/p^*}. $$ (c) For the lower estimates, again, we consider the relation of $I$ and $I\!I$ first. For any $\mathbf{x} \in {\scr A}} \def\bq{{\scr B}_0 [1, \infty)$, we need to show: $$ \inf_{n\in [1, \infty)} I_n (\mathbf{x}) \leqslant \inf_{n\in [1, \infty)} I\!I_n (\mathbf{x}). $$ In fact, with the help of proportional property, an argument similar to the one used in (a) can show this result. Next, we should show the variational formulas of $A$. Since the summability of sequence $\mathbf{x}$, without loss of generality, we may assume $\sum_{i= 1}^\infty v_i x_i^p =1$. Hence our next step is to proof $$ \sup_{\mathbf{x} \in \tilde{{\scr A}} \def\bq{{\scr B}}[1, \infty) } \inf_{n\in [1, \infty)} I\!I_n(\mathbf{x})^{(p- 1)/q} \leqslant A, $$ where $\tilde{{\scr A}} \def\bq{{\scr B}}[1, \infty)= \{ \mathbf{x} \in {\scr A}} \def\bq{{\scr B}_0 [1, \infty): \sum_{i=1}^\infty v_i x_i^p = 1 \}$. We would begin with the classical variational formulas (\ref{A}) of the optimal constants. For any $\mathbf{x} \in \tilde{{\scr A}} \def\bq{{\scr B}}[1, \infty)$, define $$ y_n= \hat{v}_n \left(\sum_{i= n}^\infty u_i (H \mathbf{x} (i))^{q-1} \right)^{p^*- 1}, $$ then we have \bg{equation}\label{==2} A \geqslant \frac{\| H \mathbf{y} \|_{l^q (u)}}{\| \mathbf{y} \|_{l^p(v)}} = \frac{\left[\sum_{n= 1}^\infty u_n (H \mathbf{y} (n))^{q} \right]^{1/q}}{\left[\sum_{n= 1}^\infty \hat{v}_n \left(\sum_{i=n}^\infty u_i (H \mathbf{x} (i))^{q-1} \right)^{p^*} \right]^{1/p}}. \end{equation} Consider the denominator of (\ref{==2}), according to Fubini theorem and the definition of $\mathbf{y}$, we obtain \begin{align}\label{==3} \sum_{i=1}^\infty & y_i \left(\sum_{j= i}^\infty u_j (H \mathbf{x} (j))^{q-1} \right) = \sum_{j=1}^\infty u_j (H \mathbf{y} (j)) (H \mathbf{x} (j))^{q-1}\notag \\ &\leqslant \left[\sum_{j= 1}^\infty u_j (H \mathbf{y} (j))^{q/p} (H \mathbf{x} (j))^{q/{p^*}} \right]^{p/q} \left[\sum_{j= 1}^\infty u_j (H \mathbf{x} (j))^q \right]^{(q-p)/q}. \end{align} The last step is based on the H\"older inequality, which needs the condition $p< q$. Moreover, since $\sum_{i= 1}^\infty v_i x_i^p = 1$ we have \bg{equation}\label{==4} \left[\sum_{j= 1}^\infty u_j (H \mathbf{x} (j))^q \right]^{(q-p)/q} \leqslant A^{q- p}. \end{equation} Combining (\ref{==2}), (\ref{==3}), (\ref{==4}) and using the proportional property, we obtain $$ A \geqslant \left[\frac{\sum_{i=1}^\infty u_i (H \mathbf{y} (i))^q}{\sum_{i=1}^\infty u_i (H \mathbf{y} (i))^{q/p} (H \mathbf{x} (i))^{q/{p^*}}} \right]^{p/{q^2}} \geqslant \inf_{n \in [1, \infty)} \left[ \frac{H \mathbf{y} (n)}{H \mathbf{x} (n)} \right]^{(p-1)/q}. $$ By the definition of $\mathbf{y}$, we get $$ \inf_{n\in [1, \infty)} I\!I_n(\mathbf{x})^{(p-1)/q} \leqslant A. $$ Since $\mathbf{x}$ is arbitrary, we obtain the variational formulas of $A$. The proof is completed in the case $N= \infty$. \quad $\square$ \medskip \noindent {\bf Proof of Corollary \ref{basic estimates}.} With the help of Theorem \ref{Var}, we can obtain the basic estimates if we choose an appropriate test function. We consider the upper estimates first. Before the proof, we need some preparations. Given an increasing positive sequence $\mathbf{\Phi} \def\qqz{\Psi}$ on $[1, N]$, for any $n\in [1, N- 1]$ and $0< \alpha} \def\bz{\beta< 1$, we assert \begin{align}\label{==5} \sum_{i=n+ 1}^{N} \left[\left(\frac{\Phi} \def\qqz{\Psi_i}{\Phi} \def\qqz{\Psi_n}\right)^\alpha} \def\bz{\beta- \left(\frac{\Phi} \def\qqz{\Psi_{i- 1}}{\Phi} \def\qqz{\Psi_n}\right)^\alpha} \def\bz{\beta \right] \left(\frac{\Phi} \def\qqz{\Psi_i}{\Phi} \def\qqz{\Psi_n}\right)^{-1} \leqslant \frac{\alpha} \def\bz{\beta}{1- \alpha} \def\bz{\beta}. \end{align} In fact, it can be proved by induction. Assume $n=N- 1$. Let $y= \Phi} \def\qqz{\Psi_N/ \Phi} \def\qqz{\Psi_{N- 1}$, then $y\geqslant 1$ (since $\mathbf{\Phi} \def\qqz{\Psi}$ is increasing). Through simple calculations, we know the function $$ f(x)= (x^\alpha} \def\bz{\beta- 1) x^{-1}, \qquad x\geqslant 1, $$ reaches the maximum when $x= \left(\frac{1}{1- \alpha} \def\bz{\beta}\right)^{1/\alpha} \def\bz{\beta}$. Hence $$ \aligned \left[\left(\frac{\Phi} \def\qqz{\Psi_N}{\Phi} \def\qqz{\Psi_{N- 1}}\right)^\alpha} \def\bz{\beta- 1 \right] \left(\frac{\Phi} \def\qqz{\Psi_N}{\Phi} \def\qqz{\Psi_{N- 1}}\right)^{-1} &= (y^\alpha} \def\bz{\beta- 1) y^{-1} \\ &\leqslant (1- \alpha} \def\bz{\beta)^{1/ \alpha} \def\bz{\beta}\left(\frac{\alpha} \def\bz{\beta}{1- \alpha} \def\bz{\beta}\right) \\ &\leqslant \frac{\alpha} \def\bz{\beta}{1- \alpha} \def\bz{\beta}. \endaligned $$ For any $1\leqslant m\leqslant N- 1$, assume the inequality (\ref{==5}) is true when $n= m$. Now consider $n=m- 1$. Let $y= \Phi} \def\qqz{\Psi_{m}/ \Phi} \def\qqz{\Psi_{m- 1}$, then $y\geqslant 1$. By the assumption, we have $$ \aligned \sum_{i= m}^N & \left[\left(\frac{\Phi} \def\qqz{\Psi_i}{\Phi} \def\qqz{\Psi_{m- 1}}\right)^\alpha} \def\bz{\beta - \left(\frac{\Phi} \def\qqz{\Psi_{i- 1}}{\Phi} \def\qqz{\Psi_{m- 1}}\right)^\alpha} \def\bz{\beta \right] \left(\frac{\Phi} \def\qqz{\Psi_i}{\Phi} \def\qqz{\Psi_{m- 1}}\right)^{-1} \\ &= \left(\frac{\Phi} \def\qqz{\Psi_{m}}{\Phi} \def\qqz{\Psi_{m- 1}}\right)^{\alpha} \def\bz{\beta- 1} \sum_{i=m}^{N} \left[\left(\frac{\Phi} \def\qqz{\Psi_i}{\Phi} \def\qqz{\Psi_{m}}\right)^\alpha} \def\bz{\beta - \left(\frac{\Phi} \def\qqz{\Psi_{i- 1}}{\Phi} \def\qqz{\Psi_{m}}\right)^\alpha} \def\bz{\beta \right] \left(\frac{\Phi} \def\qqz{\Psi_i}{\Phi} \def\qqz{\Psi_{m}}\right)^{-1} \\ &\leqslant \left(\frac{\Phi} \def\qqz{\Psi_{m}}{\Phi} \def\qqz{\Psi_{m- 1}}\right)^{\alpha} \def\bz{\beta- 1} \left[\frac{\alpha} \def\bz{\beta}{1- \alpha} \def\bz{\beta}+ 1- \left(\frac{\Phi} \def\qqz{\Psi_{m- 1}}{\Phi} \def\qqz{\Psi_{m}}\right)^\alpha} \def\bz{\beta \right] \\ &= \frac{1}{1- \alpha} \def\bz{\beta} y^{\alpha} \def\bz{\beta- 1}- y^{-1}. \endaligned $$ Again, by simple calculations, we know the function $$ f(x)= \frac{1}{1- \alpha} \def\bz{\beta} x^{\alpha} \def\bz{\beta- 1}- x^{-1}, \qquad x\geqslant 1, $$ reachs the maximum $\dfrac{\alpha} \def\bz{\beta}{1- \alpha} \def\bz{\beta}$ when $x= 1$. Hence $$ \sum_{i= m}^N \left[\left(\frac{\Phi} \def\qqz{\Psi_i}{\Phi} \def\qqz{\Psi_{m- 1}}\right)^\alpha} \def\bz{\beta - \left(\frac{\Phi} \def\qqz{\Psi_{i- 1}}{\Phi} \def\qqz{\Psi_{m- 1}}\right)^\alpha} \def\bz{\beta \right] \left(\frac{\Phi} \def\qqz{\Psi_i}{\Phi} \def\qqz{\Psi_{m- 1}}\right)^{-1} \leqslant \frac{\alpha} \def\bz{\beta}{1- \alpha} \def\bz{\beta}. $$ By induction, we prove this assertion. We should notice that the right side of (\ref{==5}) is independent of $N$, then (\ref{==5}) also be true when $N \rightarrow \infty$. Let $\alpha} \def\bz{\beta\in (0, 1)$ be an undetermined parameter. Having the inequality (\ref{==5}) in hand, we can prove: \bg{equation}\label{==6} \left(\sum_{i= n}^\infty u_i (H \mathbf{\hat{v}} (i))^{\alpha} \def\bz{\beta q/p^*} \right)^{1/q}\leqslant B (H \mathbf{\hat{v}} (n))^{(\alpha} \def\bz{\beta- 1)/p^*} \left(\frac{1}{1- \alpha} \def\bz{\beta}\right)^{1/q}, \quad 1\leqslant n< \infty. \end{equation} With the definition of $B$, for any $n$, we have $(\sum_{i=n}^\infty u_i)^{1/q}\leqslant B (H \mathbf{\hat{v}} (n))^{-1/{p^*}}$. Let $\Phi} \def\qqz{\Psi_n= (H \mathbf{\hat{v}} (n))^{q/{p^*}}$, summation by parts, we have $$\aligned \sum_{i= n}^\infty u_i \Phi} \def\qqz{\Psi_i^\alpha} \def\bz{\beta &= \Phi} \def\qqz{\Psi_n^\alpha} \def\bz{\beta \left(\sum_{i=n}^\infty u_i \right) + \sum_{i=n+ 1}^\infty \left(\Phi} \def\qqz{\Psi_i^\alpha} \def\bz{\beta- \Phi} \def\qqz{\Psi_{i- 1}^\alpha} \def\bz{\beta \right) \left(\sum_{j=i}^\infty u_j \right) \\ &\leqslant B^q \Phi} \def\qqz{\Psi_n^{\alpha} \def\bz{\beta- 1} + B^q \sum_{i=n+ 1}^\infty \left(\Phi} \def\qqz{\Psi_i^\alpha} \def\bz{\beta- \Phi} \def\qqz{\Psi_{i- 1}^\alpha} \def\bz{\beta \right) \Phi} \def\qqz{\Psi_i^{-1} \\ &= B^q \Phi} \def\qqz{\Psi_n^{\alpha} \def\bz{\beta- 1} \left\{ 1+ \sum_{i=n+ 1}^\infty \left[\left(\frac{\Phi} \def\qqz{\Psi_i}{\Phi} \def\qqz{\Psi_n}\right)^\alpha} \def\bz{\beta- \left(\frac{\Phi} \def\qqz{\Psi_{i- 1}}{\Phi} \def\qqz{\Psi_n}\right)^\alpha} \def\bz{\beta \right] \left(\frac{\Phi} \def\qqz{\Psi_i}{\Phi} \def\qqz{\Psi_n}\right)^{-1} \right\} \\ &\leqslant B^q \Phi} \def\qqz{\Psi_n^{\alpha} \def\bz{\beta- 1} \left(\frac{1}{1- \alpha} \def\bz{\beta} \right). \endaligned$$ Making a power $1/q$, we obtain the inequality (\ref{==6}). Now, we are ready to prove the upper bounds of the basic estimates. For any $n\in[1, \infty)$, Let $x_n = (H \mathbf{\hat{v}} (n))^{\alpha} \def\bz{\beta} - (H \mathbf{\hat{v}} (n- 1))^{\alpha} \def\bz{\beta}$, we have $$ \aligned {I_n^*(\mathbf{x})}^{1/p^*} &= \left[\frac{\hat{v}_n}{(H \mathbf{\hat{v}} (n))^{\alpha} \def\bz{\beta} - (H \mathbf{\hat{v}} (n- 1))^{\alpha} \def\bz{\beta}} \left(\sum_{i= n}^\infty u_i (H \mathbf{\hat{v}} (i))^{\alpha} \def\bz{\beta q /p^*} \right)^{{p^*}/q} \right]^{1/{p^*}} \\ &\leqslant \left[\frac{1}{\alpha} \def\bz{\beta (H \mathbf{\hat{v}} (n))^{\alpha} \def\bz{\beta- 1} } \left(\sum_{i= n}^\infty u_i (H \mathbf{\hat{v}} (i))^{\alpha} \def\bz{\beta q/ p^*} \right)^{{p^*}/q} \right]^{1/{p^*}} \\ &= \alpha} \def\bz{\beta^{-1/{p^*}} \left(H \mathbf{\hat{v}} (n) \right)^{(1- \alpha} \def\bz{\beta)/{p^*}} \left(\sum_{i= n}^\infty u_i (H \mathbf{\hat{v}} (i))^{\alpha} \def\bz{\beta q/{p^*}} \right)^{1/q} \\ &\leqslant B \alpha} \def\bz{\beta^{-1/{p^*}} (1- \alpha} \def\bz{\beta)^{-1/q}. \endaligned $$ An easy calculation shows that the function $$ f(x)= x^{-1/{p^*}} (1- x)^{-1/q}, \qquad 0< x<1, $$ reaches the maximum $$ \tilde{k}_{q, p}= \left(1+ \frac{q}{p^*}\right)^{1/q} \left(1+ \frac{p^*}{q}\right)^{1/p^*} $$ when $x= \dfrac{q}{p^*+ q}$. Hence we take $\alpha} \def\bz{\beta= \dfrac{q}{p^*+ q}$, then we have $$ \sup_{n\in [1, \infty)} \left( I_n^* (\mathbf{x}) \right)^{1/p^*}\leqslant \tilde{k}_{q, p} B. $$ By Theorem \ref{Var}, we get the basic upper estimates. The basic lower estimates are more straightforward. For any $n\in [1, \infty)$, we can choose a test sequence as $$ x_i^{(n)}= \begin{cases} \hat{v}_i, & 1\leqslant i\leqslant n, \\ 0, & n< i< \infty. \end{cases} $$ It is obvious that $\mathbf{x}^{(n)} \in {\scr A}} \def\bq{{\scr B}_0 [1, \infty)$, then by (\ref{A}) we have $$ \aligned A&\geqslant \sup_{n\in [1, \infty)} \frac{\left[\sum_{i= 1}^\infty u_i (H \mathbf{x}^{(n)} (i))^q \right]^{1/q}}{\left[ \sum_{i= 1}^\infty v_i \left(x_i^{(n)} \right)^p \right]^{1/p}} \\ &= \sup_{n\in [1, \infty)} \left(\sum_{i= 1}^n \hat{v}_i \right)^{1/p^*} \left[ \left(\sum_{i= 1}^n \hat{v}_i \right)^{-q} \left( \sum_{i= 1}^{n- 1} u_i \left(H \mathbf{\hat{v}} (i) \right)^q \right) + \sum_{i= n}^\infty u_i \right]^{1/q} \\ &\geqslant B. \endaligned $$ This completes the proof of Corollary \ref{basic estimates}. \quad $\square$ \medskip \medskip \noindent {\bf Proof of Corollary \ref{approximating}.} By the proportional property, we can obtain the monotonicity of \{$\dz_n$\}. The approximating sequence $\{ \dz_n \}$ comes from the upper estimates of the variational formula, and $\{\widetilde{\dz}_n\}$ comes from the lower one. These results are the simple applications of Theorem \ref{Var}. The sequence $\{\overline{\dz}_n\}$ is the straightforward application of the classical variational formula (\ref{A}). \quad $\square$ \medskip \section{Proof of Theorem \ref{kqpB}} Again, we assume $N= \infty$, and the case of finite interval will be discussed in Section \ref{Interval}. We begin with the following well-known lemma. \nnd\begin{lmm1}\label{compare} {\cms Let $\mathbf{a}$, $\mathbf{b}$ be sequences with non-negative entries. If $$ \sum_{k=i}^\infty a_k \leqslant \sum_{k=i}^\infty b_k, \qquad (\forall ~ i=1,2, \cdots) $$ then for any increasing non-negative sequence $\mathbf{c}$, we have $$ \sum_{k=1}^\infty a_k c_k \leqslant \sum_{k=1}^\infty b_k c_k. $$ } \end{lmm1} \medskip \noindent {\bf Proof}. Set $c_0= 0$. Summation by parts, we have $$ \aligned \sum_{k=1}^\infty a_k c_k &= \sum_{k=1}^\infty \left(\sum_{i=1}^k (c_i- c_{i-1}) \right) a_k = \sum_{i=1}^\infty \left(\sum_{k=i}^\infty a_k \right) (c_i- c_{i-1}) \\ &\leqslant \sum_{i=1}^\infty \left(\sum_{k=i}^\infty b_k \right) (c_i- c_{i-1})= \sum_{k=1}^\infty b_k c_k. \endaligned $$ This completes the proof of Lemma \ref{compare}. \quad $\square$ \medskip The conclusion of Lemma \ref{compare} is about increasing sequences, analogously, there is a conclusion corresponding to the decreasing sequences, cf. \rf{Bennett2}{Lemma 1}. The following lemma is due to Bliss \cite{Bliss}, which gives a special Hardy-type inequality in the continuous case. \nnd\begin{lmm1}\label{Bliss_lemma} {\cms For any non-negative real function $f(x)$, we have \bg{equation} \left(\int_0^\infty \frac{1}{x^{q- r}} \left(\int_0^x f(t) \text{\rm d} t \right)^q \text{\rm d} x \right)^{1/q}\leqslant k_{q,p} \left(\frac{p^*}{q} \right)^{1/q} \left(\int_0^\infty f^p(x) \text{\rm d} x \right)^{1/p}, \end{equation} where $r= q/p -1$ and $k_{q,p}$ is the optimal constant, which is defined as (\ref{k_qp}). Moreover, the optimal constant is attained when $$ f(x)= \frac{c}{(d \cdot x^r + 1)^{(r+ 1)/r}}, $$ where $c$ and $d$ are non-negative constants. } \end{lmm1} \medskip \noindent { {\bf Proof of Theorem \ref{kqpB}}. By Corollary \ref{basic estimates}, it is obvious that $A= \infty$ if $B= \infty$. To avoid this trivial case, we assume $B< \infty$. (a) First we consider the case that $H \mathbf{\hat{v}} (\infty)= \lim_{n \rightarrow \infty} H \mathbf{\hat{v}} (n) =\infty$. Similar to Proposition \ref{decreasing}, we can rewrite the Hardy-type inequalities (\ref{Hardy}) as: $$ \sum_{n=1}^\infty u_n \left(\sum_{i=1}^n \hat{v}_i x_i \right)^q \leqslant A^q \left(\sum_{n=1}^\infty \hat{v}_n x_n^p \right)^{q/p}. $$ Define sequence $\mathbf{\tilde{u}}$ \bg{equation}\label{tilde_u} \tilde{u}_n = B^q \left( (H \mathbf{\hat{v}} (n))^{-q/p^*}- (H \mathbf{\hat{v}} (n+1))^{-q/p^*} \right), \qquad n\geqslant 1. \end{equation} By direct summation and $H \mathbf{\hat{v}} (\infty)= \infty$, we have $$ \sum_{i= n}^\infty \tilde{u}_i = B^q \left(H \mathbf{\hat{v}} (n)^{-q/p^*} - H \mathbf{\hat{v}} (\infty)^{-q/p^*} \right)= B^q H \mathbf{\hat{v}} (n)^{-q/p^*} \geqslant \sum_{i= n}^\infty u_i. $$ Applying Lemma \ref{compare}, for any non-negative sequence $\mathbf{x}$, we obtain \bg{equation}\label{kqp1} \sum_{n= 1}^\infty u_n \left(\sum_{i=1}^n \hat{v}_i x_i \right)^q \leqslant \sum_{n= 1}^\infty \tilde{u}_n \left(\sum_{i=1}^n \hat{v}_i x_i \right)^q. \end{equation} The next is to show $$ \sum_{n= 1}^\infty \tilde{u}_n \left(\sum_{i=1}^n \hat{v}_i x_i \right)^q \leqslant A^q \left(\sum_{n=1}^\infty \hat{v}_n x_n^p \right)^{q/p}. $$ In order to use Lemma \ref{Bliss_lemma}, we should construct a function which connects summation with integration. Defined function $f: [0, \infty) \rightarrow [0, \infty)$ \bg{equation}\label{prp1} f(x)= \begin{cases} x_n, & H \mathbf{\hat{v}} (n-1) \leqslant x< H \mathbf{\hat{v}} (n), \\ 0, & x\geqslant \sup_n H \mathbf{\hat{v}} (n). \end{cases} \end{equation} It is clear that \bg{equation}\label{prp2} \sum_{i=1}^\infty \hat{v}_i x_i^p = \int_{0}^\infty f^p(x) \text{\rm d} x, \end{equation} and \bg{equation}\label{prp3} \sum_{i=1}^n \hat{v}_i x_i\leqslant \int_0^\alpha} \def\bz{\beta f(t) \text{\rm d} t, \end{equation} where $H \mathbf{\hat{v}} (n) \leqslant \alpha} \def\bz{\beta< H \mathbf{\hat{v}} (n+ 1)$. For convenience, write $\tilde{u}_0= 0, \hat{v}_0= 0$. Applying (\ref{prp3}), Lemma \ref{Bliss_lemma} and (\ref{prp2}), we see that \begin{align} \allowdisplaybreaks \sum_{n=0}^\infty \tilde{u}_n \left(\sum_{k=0}^n \hat{v}_k x_k \right)^q &= \sum_{n=0}^\infty B^q \left( (H \mathbf{\hat{v}} (n))^{-q/p^*}- (H \mathbf{\hat{v}} (n+ 1))^{-q/p^*} \right) \left(\sum_{k= 0}^n \hat{v}_k x_k \right)^q \notag \\ &= \sum_{n=0}^\infty B^q \left(\frac{q}{p^*} \int_{H \mathbf{\hat{v}} (n)}^{H \mathbf{\hat{v}} (n+ 1)} x^{-q/p^* - 1} \text{\rm d} x\right) \left(\sum_{k= 0}^n \hat{v}_k x_k \right)^q \notag \\ &\leqslant \frac{q}{p^*} B^q \left(\sum_{n=0}^\infty \int_{H \mathbf{\hat{v}} (n)}^{H \mathbf{\hat{v}} (n+ 1)} x^{-q/p^* - 1} \left(\int_0^x f(t) \text{\rm d} t\right)^q \text{\rm d} x \right) \notag \\ &= \frac{q}{p^*} B^q \int_0^\infty x^{r- q} \left(\int_0^x f(t) \text{\rm d} t \right)^q \text{\rm d} x \quad (r=q/p - 1)\notag \\ &\leqslant B^q k_{q, p}^q \left(\int_0^\infty f^p(x) \text{\rm d} x \right)^{q/p} \notag \\ &= B^q k_{q, p}^q \left(\sum_{i=0}^\infty \hat{v}_i x_i^p \right)^{q/p}. \notag \end{align} By the definition of the optimal constant, we have $A\leqslant k_{q, p} B$. (b) To show the factor of the basic upper estimate is best possible, we attempt to mimic the extremal function in Lemma \ref{Bliss_lemma}: $$ f(x)= \frac{c}{\left(dx^r+ 1\right)^{(r+1)/r}} \quad \text{and} \quad \int_0^x f(t) \text{\rm d} t = \frac{cx}{\left(dx^r +1 \right)^{1/r}}, $$ where $c$ and $d$ are arbitrary positive constants. We start with this form, set $$ u_n= n^{-q/p^*}- (n+1)^{-q/p^*}, \qquad v_n \equiv 1, $$ and $$ x_n = \frac{c n}{(n^r+ d)^{1/r}}- \frac{c (n-1)}{((n-1)^r+ d)^{1/r}}. $$ Obviously, the form of $\mathbf{x}$ comes from the difference of $\int_0^x f(t) \text{\rm d} t$. In this case, we have $B= 1$. Here we are free to choose $c$ and $d$, however, no matter what choices, there is some loss of precision between integrals and series. But by direct calculation, we find that this loss becomes negligible when $c/d \rightarrow 0$. Without loss of generality, we choose $c= 1$ and let $d$ be a positive and large enough real number. Next, we calculate the left and the right side of (\ref{Hardy}). The calculation of the right side of (\ref{Hardy}) is direct. By the definition of $\mathbf{x}$, we have \begin{align}\label{right} \sum_{n= 1}^\infty x_n^p &= \sum_{n=1}^\infty \left[\frac{n}{(n^r + d)^{1/r}} -\frac{n- 1}{((n- 1)^r + d)^{1/r}} \right]^p \notag \\ &= \sum_{n=1}^\infty \left[\int_{n- 1}^n \frac{d}{(x^r+ d)^{1/r +1}} \text{\rm d} x \right]^p \notag \\ &\leqslant \sum_{n= 1}^\infty \int_{n- 1}^n \left(\frac{d}{(x^r+ d)^{1/r +1}} \right)^p \text{\rm d} x \notag \\ &= r^{-1} d^{(1- p)/r} B\left(\frac{1}{r}, \frac{q- 1}{r}\right). \end{align} The left side of (\ref{Hardy}) is difficult. First, we assert there is a large enough integer $N$ such that \bg{equation}\label{left1} \int_N^{\infty} x^{-q/{p^*}- 1} \left[ \frac{x^q}{(x^r+ d)^{q/r}} \right] \text{\rm d} x \leqslant \int_1^{\infty} (x+ 1)^{-q/{p^*}- 1} \left[ \frac{x^q}{(x^r+ d)^{q/r}} \right] \text{\rm d} x. \end{equation} In fact, we have $$ \int_N^{\infty} x^{-q/{p^*}- 1} \left[ \frac{x^q}{(x^r+ d)^{q/r}} \right] \text{\rm d} x \leqslant \int_N^{\infty} x^{-q/{p^*}- 1} \text{\rm d} x = \frac{p^*}{q} N^{-q/{p^*}}. $$ The existence of $N$ is obvious since the left side of (\ref{left1}) decreases to $0$ as $N \uparrow \infty$. Fix this sufficiently large integer $N$, then the left side of (\ref{left1}) is calculable. Using the integral transform $s^{-1}= d^{-1} x^r +1$, we have \bg{equation}\label{left2} \int_N^{\infty} x^{-q/{p^*}- 1} \left[ \frac{x^q}{(x^r+ d)^{q/r}} \right] \text{\rm d} x = r^{-1} d^{- \frac{q}{rp^*}} B\left(\frac{1+r}{r}, \frac{q- r- 1}{r}, \frac{d}{N^r+ d}\right), \end{equation} where $B(a, b, x)$ is the incomplete Beta function: $$ B(a, b, x)= \int_{0}^{x} s^{a- 1} (1- s)^{b- 1} \text{\rm d} s. $$ Applying the mean value theorem, (\ref{left1}) and (\ref{left2}), we have \begin{align}\label{left3} \int_1^\infty & \left[x^{-q/{p^*}}- (x+ 1)^{-q/{p^*}} \right] \frac{x^q}{(x^r+ d)^{q/r}} \text{\rm d} x \notag \\ & \geqslant \int_1^\infty \frac{q}{p^*} (x+ 1)^{-q/{p^*}- 1} \left[\frac{x^q}{(x^r+ d)^{q/r}} \right] \text{\rm d} x \notag \\ & \geqslant d^{- \frac{q}{rp^*}} \left(\frac{q}{p^*} \right) r^{-1} B\left(\frac{1+r}{r}, \frac{q- r- 1}{r}, \frac{d}{N^r+ d}\right). \end{align} Now, it's very easy to calculate the optimal constants. Using the relation $$ B(a+ 1, b- 1)= \frac{a}{b- 1} B(a, b), $$ it follows from (\ref{A}), (\ref{right}) and (\ref{left3}) that $$ \aligned A^q &\geqslant \left[\sum_{n=1}^{\infty} \left(n^{-q/p^*}- (n+ 1)^{-q/p^*} \right) (H \mathbf{x} (n))^q \right] \left( \sum_{n=1}^\infty x_n^p \right)^{- q/p} \\ &\geqslant \left[\int_{1}^{\infty} \left[x^{-q/p^*} - (x+ 1)^{-q/p^*} \right] \frac{x^q}{(x^r+ d)^{\frac{q}{r}}} \text{\rm d} x \right] \left( \sum_{n=1}^\infty x_n^p \right)^{- q/p} \\ &\geqslant \left(\frac{q}{p^*}\right) r^{q/p - 1} \cdot B\left(\frac{1+ r}{r}, \frac{q- 1- r}{r}, \frac{d}{N^r+ d} \right) \cdot B\left(\frac{1}{r}, \frac{q- 1}{r}\right)^{- q/p} \\ &\rightarrow k_{q,p}^q \qquad \text{(as $d \rightarrow \infty$)}. \endaligned $$ Hence, the factor of basic upper estimate is best possible. (c) The final step is to remove condition $H \mathbf{\hat{v}} (\infty)= \infty$. We use Proposition \ref{C. estimates} of section \ref{Interval}. Fix $N_0 < \infty$. For given $\mathbf{u}$ and $\mathbf{v}$ on $[1, \infty)$, we define $\mathbf{u}^{N_0}$ and $\mathbf{v}^{N_0}$ to be the restriction of $\mathbf{u}$ and $\mathbf{v}$ on $[1, N_0]$. Then define $$ \overline{u}_n = \begin{cases} u_n^{N_0}, & 1\leqslant n \leqslant N_0, \\ 0, & n> N_0, \end{cases} $$ and $$ \overline{v}_n = \begin{cases} v_n^{N_0}, & 1\leqslant n\leqslant N_0, \\ 1, & n> N_0. \end{cases} $$ Obviously, we have $H \mathbf{\overline{v}} (\infty) = \infty$. Applying the result of (a), we have $$ A(\mathbf{\overline{u}}, \mathbf{\overline{v}}) \leqslant k_{q, p} B(\mathbf{\overline{u}}, \mathbf{\overline{v}}). $$ By Proposition \ref{C. estimates}, we get $$ A(\mathbf{u}^{N_0}, \mathbf{v}^{N_0}) \leqslant k_{q, p} B(\mathbf{u}^{N_0}, \mathbf{v}^{N_0}). $$ The assertion follows by letting $N_0\rightarrow \infty$. This completes the proof of Theorem \ref{kqpB} in the case $N= \infty$. \quad $\square$ \medskip Review part (a) of the proof of Theorem (\ref{kqpB}), when \bg{equation}\label{v} N= \infty, \qquad H \mathbf{\hat{v}} (\infty) =\infty, \end{equation} we give the method to construct $\mathbf{u}$ from $\mathbf{v}$ such that the Hardy-type inequalities (\ref{Hardy}) hold with these $\mathbf{u}$ and $\mathbf{v}$. The part (b) show the optimal constant reaches the upper bound of the basic estimate. It means that the basic upper estimate with the improved factor $k_{q,p}$ holds for a large class of $(\mathbf{u}, \mathbf{v})$. The original idea of this construction is from Chen \rf{Chen3}{Proposition 4.5}. To distinguish it from Theorem \ref{kqpB}, we give the following proposition. \nnd\begin{prp1}\label{Bliss_prp} {\cms For any positive sequences $\mathbf{v}$ and constant $0< C< \infty$, the discrete Hardy-type inequalities (\ref{Hardy}) hold on $[1, \infty)$ with \bg{equation} \tilde{u}_n = C^q \left( (H \mathbf{\hat{v}} (n))^{-q/p^*}- (H \mathbf{\hat{v}} (n+1))^{-q/p^*} \right), \qquad n\geqslant 1, \end{equation} and its optimal constant $A$ satisfies \bg{equation} A \leqslant k_{q, p} C, \end{equation} where $k_{q,p}$ is defined as (\ref{k_qp}). Moreover, when $N$ and $\mathbf{\hat{v}}$ satisfy (\ref{v}), the upper bound is sharp with $C= B$. } \end{prp1} \section{Hardy-type Inequalities on Interval}\label{Interval} In this section, we study the comparison results of the optimal constants and the basic estimates on different intervals. In continuous case, the corresponding comparison results have been done by Chen \rf{Chen3}{Appendix}. With these results, we can get the complete proofs of Theorem \ref{Var} and Theorem \ref{kqpB}. Before specific discussion, we need some notations. Fix two natural numbers $N$ and $N'$ with $N< N'$. Given two positive sequences $\mathbf{u}$ and $\mathbf{v}$ on $[1, N]$, we can extend them to $[1, N']$ as follows: \bg{equation}\label{u'} u'_i = \begin{cases} u_i, & 1\leqslant i\leqslant N, \\ 0, & N< i\leqslant N'; \end{cases} \end{equation} \bg{equation}\label{v'} v'_i = \begin{cases} v_i, & 1\leqslant i\leqslant N, \\ \#, & N< i\leqslant N', \end{cases} \end{equation} where $\#$ means arbitrary positive numbers. Denote $A_N(\mathbf{u}, \mathbf{v})$ be the optimal constant of the Hardy-type inequalities (\ref{Hardy}) in the interval $[1, N]$ with sequences $\mathbf{u}$ and $\mathbf{v}$, and similar for $B_N(\mathbf{u}, \mathbf{v})$. The first result is a comparison for the optimal constants on different intervals. \nnd\begin{prp1}\label{C. constant} {\cms Given two positive sequences $\mathbf{u}'$ and $\mathbf{v}'$ on $[1, N']$. Use $\mathbf{u}$ and $\mathbf{v}$ to denote their restrictions to $[1, N]$. Then we have $A_N(\mathbf{u}, \mathbf{v}) \uparrow A_{N'}(\mathbf{u}', \mathbf{v}')$ as $N \uparrow {N'} \leqslant \infty$. In particular, if the inequality (\ref{Hardy}) holds on $[1, N']$, then it also holds with the same constant $A_{N'}(\mathbf{u}', \mathbf{v}')$ on $[1, N]$. } \end{prp1} \medskip \noindent {\bf Proof}. (a) Given a non-negative sequence $\mathbf{x}$ on $[1, N]$, we can extend to $[1, N']$ by setting \bg{equation}\label{extension of x} x'_i = \begin{cases} x_i & 1\leqslant i\leqslant N, \\ 0 & N< i\leqslant N'. \end{cases} \end{equation} Then we have $$ \aligned \left[ \sum_{n=1}^N u_n \left(H \mathbf{x} (n) \right)^q \right]^{1/q} &= \left[ \sum_{n=1}^{N'} u'_n \left(H \mathbf{x'} (n) \right)^q \right]^{1/q} \\ &\leqslant A_{N'}(\mathbf{u}', \mathbf{v}') \left[ \sum_{n=1}^{N'} v'_n {x'}_n^p \right]^{1/p} \\ &= A_{N'}(\mathbf{u}', \mathbf{v}') \left[ \sum_{n=1}^{N} v_n {x}_n^p \right]^{1/p}. \endaligned $$ It means that $A_N(\mathbf{u}, \mathbf{v}) \leqslant A_{N'}(\mathbf{u}', \mathbf{v}')$. (b) Our next goal is to show the convergence. First we consider the case that $\sum_{n=1}^{N'} u'_n = \infty$. Clearly, in this case we have $N'= \infty$ and $A_{N'}(\mathbf{u}', \mathbf{v}')= \infty$. Besides, restricting to $[1, n]$ and choosing $\mathbf{x}= (1, 0, \dots, 0)$, we obtain $$ A_n(\mathbf{u}, \mathbf{v}) \geqslant \left(\sum_{i=1}^n u_i \right)^{1/q} v_1^{- 1/p} \rightarrow \infty, \qquad \text{as $n\rightarrow \infty$}. $$ Hence the convergence holds in this case. (c) Let $\sum_{n=1}^{N'} u'_n < \infty$. For every non-negative sequence $\mathbf{x}$ on $[1, N']$ with $\sum_{n = 1}^{N'} v'_n x_n^p < \infty$, we get $$ \frac{\left[\sum_{n= 1}^N u_n (H \mathbf{x} (n))^q \right]^{1/q}}{\left(\sum_{n= 1}^N v_n x_n^p \right)^{1/p}} \rightarrow \frac{\left[\sum_{n= 1}^{N'} u'_n (H \mathbf{x} (n))^q \right]^{1/q}}{ \left( \sum_{n= 1}^{N'} v'_n x_n^p \right)^{1/p}} \leqslant A_{N'}(\mathbf{u}', \mathbf{v}'), $$ as $N \uparrow N'$. With (\ref{A}), for every $\varepsilon} \def\xz{\xi> 0$, we can choose a sequence $\mathbf{x}$ such that $$ A_{N'}(\mathbf{u}', \mathbf{v}') \leqslant \frac{\left[\sum_{n= 1}^{N'} u'_n (H \mathbf{x} (n))^q \right]^{1/q}}{\left(\sum_{n= 1}^{N'} v'_n x_n^p \right)^{1/p}} + \varepsilon} \def\xz{\xi. $$ Then we can choose $N$ closed to $N'$ such that $$ \frac{\left[\sum_{n= 1}^{N'} u'_n (H \mathbf{x} (n))^q \right]^{1/q}}{\left(\sum_{n= 1}^{N'} v'_n x_n^p \right)^{1/p}} \leqslant \frac{\left[\sum_{n= 1}^N u_n (H \mathbf{x} (n))^q \right]^{1/q}}{\left(\sum_{n= 1}^N v_n x_n^p \right)^{1/p}} + \varepsilon} \def\xz{\xi. $$ Hence, we get $$ A_N(\mathbf{u}, \mathbf{v}) \leqslant A_{N'}(\mathbf{u}', \mathbf{v}') \leqslant \frac{\left[\sum_{n= 1}^N u_n (H \mathbf{x} (n))^q \right]^{1/q}}{\left(\sum_{n= 1}^N v_n x_n^p \right)^{1/p}} + 2\varepsilon} \def\xz{\xi \leqslant A_N(\mathbf{u}, \mathbf{v}) + 2\varepsilon} \def\xz{\xi. $$ It means that the convergence holds. \quad $\square$ \medskip The following result is about the factor in the basic estimates. \nnd\begin{prp1}\label{C. estimates} {\cms Given two positive sequences $\mathbf{u}$ and $\mathbf{v}$ on $[1, N]$, $\mathbf{u}'$ and $\mathbf{v}'$, defined by (\ref{u'}) and (\ref{v'}), are the extensions on $[1, N']$. Suppose that $A_{N'}(\mathbf{u}', \mathbf{v}')\leqslant k B_{N'}(\mathbf{u}', \mathbf{v}')$ for a universal constant $k$, then we have $A_N(\mathbf{u}, \mathbf{v})\leqslant k B_N(\mathbf{u}, \mathbf{v})$. } \end{prp1} \medskip \noindent {\bf Proof}. Given a sequence $\mathbf{x}$ on $[1, N]$, we can extend it from $[1, N]$ to $[1, N']$ by (\ref{extension of x}). Then we have $$ \aligned \left( \sum_{n=1}^N u_n (H \mathbf{x} (n))^q \right)^{1/q} &= \left( \sum_{n=1}^{N'} u'_n (H \mathbf{x}' (n))^q \right)^{1/q} \\ &\leqslant A_{N'}(\mathbf{u}', \mathbf{v}') \left( \sum_{n= 1}^{N'} v'_n {x'}_n^p \right)^{1/p} \\ &\leqslant k B_{N'}(\mathbf{u}', \mathbf{v}') \left( \sum_{n= 1}^{N'} v'_n {x'}_n^p \right)^{1/p} \\ &= k B_{N'}(\mathbf{u}', \mathbf{v}') \left( \sum_{n= 1}^{N} v_n x_n^p \right)^{1/p}. \endaligned $$ With the definition of the extensions (\ref{u'}) and (\ref{v'}), we can easily check that $$ B_{N'}(\mathbf{u}', \mathbf{v}')= B_{N}(\mathbf{u}, \mathbf{v}). $$ It follows that $$ \left( \sum_{n=1}^N u_n (H \mathbf{x} (n))^q \right)^{1/q} \leqslant k B_{N}(\mathbf{u}, \mathbf{v}) \left( \sum_{n= 1}^{N} v_n x_n^p \right)^{1/p}. $$ Hence $A_N(\mathbf{u}, \mathbf{v})\leqslant k B_N(\mathbf{u}, \mathbf{v})$ as required. \quad $\square$ \medskip With the help of Proposition \ref{C. constant} and Proposition \ref{C. estimates}, we know the variational formulas of the optimal constants, the basic estimates and the improved factor of the basic upper estimates are true when $N< \infty$. So far, we complete the proofs of our main results. The following result gives an opposite view of Proposition \ref{C. constant}: from some local sub-intervals to the whole interval. It gives us an approximating procedure for the unbounded interval. \nnd\begin{prp1}\label{C. interval} {\cms Given two positive sequences $\mathbf{u}$ and $\mathbf{v}$ on $[1, N]$, extend them to $[1, N']$ by (\ref{u'}) and (\ref{v'}). Then we have $A_N (\mathbf{u}, \mathbf{v}) = A_{N'} (\mathbf{u}', \mathbf{v}')$. } \end{prp1} \medskip \noindent {\bf Proof}. For any sequence $\mathbf{x} \in {\scr A}} \def\bq{{\scr B}[1, N]$, let $\mathbf{x}'$ be the extension of $\mathbf{x}$ from $[1, N]$ to $[1, N']$ by (\ref{extension of x}). Obviously, we have $\mathbf{x}' \in {\scr A}} \def\bq{{\scr B}[1, N']$. The inequalities in $[1, N']$ are $$ \| H \mathbf{x}' \|_{l^q(u')} \leqslant A_{N'} (\mathbf{u}', \mathbf{v}') \| \mathbf{x}' \|_{l^p(v')}. $$ With (\ref{u'}), (\ref{v'}) and (\ref{extension of x}), it follows that $$ \| H \mathbf{x} \|_{l^q(u)} \leqslant A_{N'} (\mathbf{u}', \mathbf{v}') \| \mathbf{x} \|_{l^p(v)}. $$ Because $\mathbf{x}$ is arbitrary, it implies that $A_N (\mathbf{u}, \mathbf{v}) \leqslant A_{N'} (\mathbf{u}', \mathbf{v}')$. Conversely, for any $\mathbf{x} \in {\scr A}} \def\bq{{\scr B}[1, N']$, we have $$ \aligned \left(\sum_{n=1}^{N'} {u'}_n (H \mathbf{x} (n))^q \right)^{1/q} &= \left(\sum_{n=1}^{N} u_n (H \mathbf{x} (n))^q \right)^{1/q} \\ &\leqslant A_N (\mathbf{u}, \mathbf{v}) \left( \sum_{n=1}^N v_n \mathbf{x}_n^p \right)^{1/p} \\ &\leqslant A_N (\mathbf{u}, \mathbf{v}) \left( \sum_{n=1}^{N'} v'_n \mathbf{x}_n^p \right)^{1/p}. \endaligned $$ This implies that $A_{N'} (\mathbf{u}', \mathbf{v}') \leqslant A_N (\mathbf{u}, \mathbf{v})$ and then the equality holds. \quad $\square$ \medskip \section{Examples}\label{examples} As mentioned in introduction, Hardy-type inequalities play important role in probability theory. The first example is from birth-death processes which is standard having constant birth and death rates, cf. \rf{Chen5}{Example 5.3}. We present this example to illustrate the power of out results. \nnd\begin{xmp1}\label{example1} {\cms Let $p=q=2$ and $N= \infty$. For $n \geqslant 1$, let $u_n = \gamma} \def\kz{\kappa^n$, $v_n= b \gamma} \def\kz{\kappa^n$, where $\gamma} \def\kz{\kappa$ and $b$ are constants with $\gamma} \def\kz{\kappa< 1$ and $b> 0$. Then $$ B < \widetilde{\dz}_1= \overline{\dz}_1 < A = \dz_1 < 2B, $$ where $B= \dfrac{1}{\sqrt{b} (1- \gamma} \def\kz{\kappa)}$, $\widetilde{\dz}_1= \overline{\dz}_1= \dfrac{\sqrt{1+ \gamma} \def\kz{\kappa}}{\sqrt{b} (1- \gamma} \def\kz{\kappa)}$, $A = \dz_1= \dfrac{1}{\sqrt{b}(1- \sqrt{\gamma} \def\kz{\kappa})}$. Moreover, the optimal constant is attained at sequence $$ a_n= \gamma} \def\kz{\kappa^{(-n +1)/2} \left[n - (n- 1) \gamma} \def\kz{\kappa^{1/2} \right], \qquad n\geqslant 1. $$ } \end{xmp1} \medskip \noindent {\bf Proof}. (a) First, $B$ is easy to calculate. By the definition, we have $$ B= \sup_{n\in [1, \infty)} \left(\sum_{i=1}^n b^{-1} \gamma} \def\kz{\kappa^{- i} \right)^{1/2} \left(\sum_{j= n}^{\infty} \gamma} \def\kz{\kappa^{j}\right)^{1/2} = \frac{1}{\sqrt{b} (1- \gamma} \def\kz{\kappa)}. $$ Next, by (\ref{k_pp}), we have $k_{2, 2}= 2$. By Corollary \ref{basic estimates}, we obtain the basic estimates of the optimal constants: \bg{equation}\label{EX13} \frac{1}{\sqrt{b} (1- \gamma} \def\kz{\kappa)} \leqslant A \leqslant \frac{2}{\sqrt{b} (1- \gamma} \def\kz{\kappa)}. \end{equation} (b) To compute $\dz_1$, we use Corollary \ref{approximating}. Let $$ x_n^{(1)}= (H \mathbf{\hat{v}} (n))^{1/2}- (H \mathbf{\hat{v}} (n- 1))^{1/2}, \qquad n\geqslant 1, $$ and then $$ H \mathbf{x}^{(1)} (n) = (H \mathbf{\hat{v}} (n))^{1/2} = \left[\frac{\gamma} \def\kz{\kappa^{-n}- 1}{b (1- \gamma} \def\kz{\kappa)}\right]^{1/2}. $$ For convenience, we use $\fz_n= \gamma} \def\kz{\kappa^{-n} -1$ in the following. By direct computations, we have \begin{align}\label{EX11} I\!I_n^* \left(\mathbf{x}^{(1)}\right) &= \frac{1}{H \mathbf{x}^{(1)} (n)} \sum_{i=1}^n \hat{v}_i \left(\sum_{j=i}^\infty u_j \left(H \mathbf{x}^{(1)} (j) \right) \right) \notag \\ &= \frac{b^{-3/2}}{(1- \gamma} \def\kz{\kappa)^{1/2}} \frac{1}{H \mathbf{x}^{(1)} (n)} \sum_{i=1}^n \gamma} \def\kz{\kappa^{-i} \left(\sum_{j=i}^\infty \gamma} \def\kz{\kappa^j \fz_j^{1/2} \right) \notag \\ &= \frac{\fz_n^{-1/2}}{b (1- \gamma} \def\kz{\kappa)} \left[\sum_{j=1}^n \gamma} \def\kz{\kappa^{j} \fz_j^{3/2} + \fz_n \sum_{j=n+ 1}^\infty \gamma} \def\kz{\kappa^{j} \fz_j^{1/2}\right]. \end{align} At the last step, we exchange the order of summation. From (\ref{EX11}), it is easy to check that $I\!I_n^* \left(\mathbf{x}^{(1)}\right)$ reaches the maximum when $n \rightarrow \infty$. Hence, by L'Hospital's rule, we obtain $$ \aligned \dz_{1}^2 &= \sup_{n\in [1, \infty)} I\!I_n^* \left(\mathbf{x}^{(1)} \right) \\ &= \frac{1}{b (1- \gamma} \def\kz{\kappa)} \left[ \lim_{n\rightarrow \infty} \fz_n^{-1/2} \sum_{j=1}^n \gamma} \def\kz{\kappa^{j} \fz_j^{3/2} + \lim_{n\rightarrow \infty} \fz_n^{1/2} \sum_{j=n+ 1}^\infty \gamma} \def\kz{\kappa^{j} \fz_j^{1/2} \right] \\ &= \frac{1}{b (1- \gamma} \def\kz{\kappa)} \left[ \frac{1}{1- \sqrt{\gamma} \def\kz{\kappa} } + \frac{\sqrt{\gamma} \def\kz{\kappa}}{1- \sqrt{\gamma} \def\kz{\kappa}} \right] \\ &= \frac{1}{b (1- \sqrt{\gamma} \def\kz{\kappa})^2}. \endaligned $$ (c) Similarly, we use Corollary \ref{approximating} to compute $\overline{\dz}_1$ and $\widetilde{\dz}_1$. Fix $k> 0$, let $$ y_n^{(k, 1)}= \begin{cases} b^{-1} \gamma} \def\kz{\kappa^{-n}, & n\leqslant k, \\ 0, & n>k, \end{cases} $$ and then $$ H \mathbf{y}^{(k, 1)} (n) = \frac{\gamma} \def\kz{\kappa^{-(n \wedge k)} - 1}{b (1- \gamma} \def\kz{\kappa)}= \frac{\fz_{n \wedge k}}{b (1- \gamma} \def\kz{\kappa)}. $$ By lots of tedious calculations, we have $$ \aligned I\!I_n \left( \mathbf{y}^{(k, 1)} \right) &= \frac{1}{H \mathbf{y}^{(k, 1)} (n)} \sum_{i=1}^n \hat{v}_i \left(\sum_{j= i}^\infty u_j \left(H \mathbf{y}^{(k, 1)} \right) \right) \\ &= \frac{1}{b \fz_{n \wedge k}} \sum_{i=1}^n \gamma} \def\kz{\kappa^{-i} \left(\sum_{j= i}^\infty \gamma} \def\kz{\kappa^{j} \fz_{j \wedge k} \right) \\ &= \frac{1}{b (1- \gamma} \def\kz{\kappa)} \left[\frac{1+ \gamma} \def\kz{\kappa}{1- \gamma} \def\kz{\kappa} - \frac{2 (n \wedge k)}{\fz_{n \wedge k}} + (k- n)\vee 0 - \frac{\gamma} \def\kz{\kappa^{k+ 1} \fz_{(n- k) \vee 0}}{1- \gamma} \def\kz{\kappa} \right]. \endaligned $$ Next, note that $I\!I_n \left( \mathbf{y}^{(k, 1)} \right)$ reaches the minimum when $n= k$, and then $$ \aligned \widetilde{\dz}_1^2 &= \sup_{k\in [1, \infty)} \inf_{n\in [1, \infty)} I\!I_n \left(\mathbf{y}^{(k, 1)} \right) \\ &= \sup_{k\in [1, \infty)} \frac{1}{b (1- \gamma} \def\kz{\kappa)} \left(\frac{1+ \gamma} \def\kz{\kappa}{1- \gamma} \def\kz{\kappa} - \frac{2 k}{\fz_{k}} \right) \\ &= \frac{1}{b (1- \gamma} \def\kz{\kappa)} \lim_{k\rightarrow \infty} \left(\frac{1+ \gamma} \def\kz{\kappa}{1- \gamma} \def\kz{\kappa} - \frac{2 k}{\fz_{k}} \right) \\ &= \frac{1+ \gamma} \def\kz{\kappa}{b (1- \gamma} \def\kz{\kappa)^2}. \endaligned $$ Now, we consider $\overline{\dz}_1$. Since $$ \sum_{n=1}^\infty v_n \left(y_n^{(k, 1)} \right)^2 = \frac{\fz_k}{b(1- \gamma} \def\kz{\kappa)}, $$ and $$ \sum_{n=1}^\infty u_n \left( H \mathbf{y}^{(k, 1)} (n) \right)^2 = \frac{1}{b^2 (1- \gamma} \def\kz{\kappa)^2} \left[\sum_{n= 1}^k \gamma} \def\kz{\kappa^n \fz_n^2 + \frac{\gamma} \def\kz{\kappa^{k+ 1} \fz_k^2}{1- \gamma} \def\kz{\kappa} \right], $$ we have $$ \aligned \overline{\dz}_1^2&= \sup_{k\in [1, \infty)} \frac{1}{b(1- \gamma} \def\kz{\kappa)} \left[\fz_k^{-1} \sum_{n=1}^k \gamma} \def\kz{\kappa^n \fz_n^2 + \frac{\gamma} \def\kz{\kappa- \gamma} \def\kz{\kappa^{k+ 1}}{1- \gamma} \def\kz{\kappa} \right] \\ &= \frac{1}{b(1- \gamma} \def\kz{\kappa)} \left[ \lim_{k\rightarrow \infty} \fz_k^{-1} \sum_{n=1}^k \gamma} \def\kz{\kappa^n \fz_n^2 + \frac{\gamma} \def\kz{\kappa }{1- \gamma} \def\kz{\kappa} \right] = \frac{1+ \gamma} \def\kz{\kappa}{b (1- \gamma} \def\kz{\kappa)^2}. \endaligned $$ In the last step, the L'Hospital's rule is used to calculate the limitation of $k$. (d) So far, by Corollary \ref{approximating}, we obtain the estimates of the optimal constants, which is more precise than the basic estimates (\ref{EX13}) \bg{equation}\label{EX12} \frac{\sqrt{1+ \gamma} \def\kz{\kappa}}{\sqrt{b} (1- \gamma} \def\kz{\kappa)} \leqslant A \leqslant \frac{1}{\sqrt{b} (1- \sqrt{\gamma} \def\kz{\kappa})}. \end{equation} In fact, the optimal constant can be accurately calculated. Let $a_n= \gamma} \def\kz{\kappa^{(-n +1)/2} \left[n - (n- 1) \gamma} \def\kz{\kappa^{1/2} \right] (n\geqslant 1)$, then $$ H \mathbf{a} (n)= n \gamma} \def\kz{\kappa^{(-n+ 1)/2}. $$ Here we want to use $\mathbf{a}$ instead of $\mathbf{y}^{(k, 1)}$ to get the lower estimates. However, it is easy to check that $\mathbf{a}$ is non-summability. It means that Theorem \ref{Var} is invalid. By the classical variational formula (\ref{A}) and the L'Hospital's rule, we have $$ \aligned A^2 &\geqslant \frac{\sum_{n= 1}^\infty \gamma} \def\kz{\kappa^n a_n^2 }{\sum_{n=1}^\infty b \gamma} \def\kz{\kappa^n \left(a_n - a_{n-1} \right)^2} \\ &= b^{-1} \lim_{n\rightarrow \infty} \frac{n^2}{\left[n- (n- 1) \gamma} \def\kz{\kappa^{1/2} \right]^2} \\ &= \frac{1}{b (1- \sqrt{\gamma} \def\kz{\kappa})^2}. \endaligned $$ As a consequence, we obtain $A = \dfrac{1}{\sqrt{b} (1- \sqrt{\gamma} \def\kz{\kappa})}$. \quad $\square$ \medskip To distinguish the first example, the second one is about the nonlinear situation $p \neq q$, which is from proof of Theorem \ref{kqpB}. The optimal constant is clear in this example. \nnd\begin{xmp1}\label{example2} {\cms Let $p \neq q$ and $N= \infty$. For $n\geqslant 1$, let $u_n= n^{-q/{p^*}}- (n+ 1)^{-q/{p^*}}$, $v_n \equiv 1$. Then (1) The optimal constant is $A= k_{q, p}$, which is attained at sequence $\mathbf{x}$: $$ x_n= \frac{cn}{(n^r+ d)^{1/r}} - \frac{c(n- 1)}{((n-1)^r+ d)^{1/r}}, \qquad n \geqslant 1, $$ where $r=q/p - 1$, $k_{q, p}$ is defined as (\ref{k_qp}), $c$ and $d$ are arbitrary positive constants. (2) The basic estimates and the approximating procedure are $$ B \leqslant \overline{\dz}_1 \vee \widetilde{\dz}_1 \leqslant A = k_{q, p} B \leqslant \dz_1, $$ where $B=1$, $\overline{\dz}_1 \geqslant 1$, $\widetilde{\dz}_1 \geqslant 1$ and $\dz_1 \leqslant \left(1+ \dfrac{q}{p^*} \right)^{1/q+ 1/p^*}$. } \end{xmp1} \medskip \noindent {\bf Proof}. The first part has been done in Theorem \ref{kqpB}. The remainder of this proof is to compute $\dz_1$, $\overline{\dz}_1$ and $\widetilde{\dz}_1$. To compute $\dz_1$, let $$ x_n^{(1)}= n^{q/(p^*+ q)} - (n- 1)^{q/(p^*+ q)}, n\geqslant 1, $$ then we have $$ \aligned I\!I_n^* \left(\mathbf{x}^{(1)}\right) &= n^{- \frac{q}{p^*+q}} \sum_{i= 1}^n \left\{\sum_{j=i}^{\infty} \left[j^{-\frac{q}{p^*}}- (j+ 1)^{- \frac{q}{p^*}} \right] j^{\frac{q^2}{p^*(p^*+ q)}} \right\}^{p^*/q} \\ &\leqslant n^{- \frac{q}{p^*+q}} \sum_{i= 1}^n \left\{ \left(\frac{q}{p^*} \right) \sum_{j=i}^{\infty} \int_j^{j+1} x^{\frac{q^2}{p^*(p^*+ q)}- \frac{q}{p^*}- 1} \text{\rm d} x \right\}^{p^*/q} \\ &= n^{- \frac{q}{p^*+q}} \sum_{i= 1}^n \left[ \left(\frac{p^*+ q}{p^*} \right) i^{- \frac{q}{p^*+ q}} \right]^{p^*/q} \\ &\leqslant n^{- \frac{q}{p^*+q}} \left(\frac{p^*+ q}{p^*} \right)^{p^*/q} \left(1+ \int_1^n x^{- \frac{p^*}{p^*+ q}} \text{\rm d} x\right) \\ &= \left(1+ \frac{q}{p^*} \right)^{p^*/q + 1}. \\ \endaligned $$ Therefore, we obtain $$ \dz_1 = \sup_{n\in [1, \infty)} \left[ I\!I_n^* \left(\mathbf{x}^{(1)} \right)\right]^{1/p^*} \leqslant \left(1+ \frac{q}{p^*} \right)^{1/q + 1/p^*}. $$ To compute $\overline{\dz}_1$ and $\widetilde{\dz}_1$, let $$ y_n^{(k, 1)} = \begin{cases} 1, & n\leqslant k, \\ 0, & n> k. \end{cases} $$ Obviously, we have $H \mathbf{y}^{(k, 1)} (n) = n \wedge k$, $\|\mathbf{y}^{(k, 1)} \|_{l^p(v)} = k^{1/p}$ and $$ \aligned \|H \mathbf{y}^{(k, 1)} \|_{l^q(u)} &= \left[ \sum_{n=1}^\infty u_n (n \wedge k)^q \right]^{1/q} \\ &= \left[ \sum_{n=1}^{k- 1} \left(n^{-\frac{q}{p^*}}- (n+1)^{-\frac{q}{p^*}} \right) n^q + k^{q/p} \right]^{1/q}. \endaligned $$ Hence, we obtain $$ \aligned \overline{\dz}_1 &= \sup_{k\in [1, \infty)} \frac{\|H \mathbf{y}^{(k, 1)} \|_{l^q(u)}}{\|\mathbf{y}^{(k, 1)} \|_{l^p(v)}} \\ &= \sup_{k\in [1, \infty)} \left[k^{-q/p} \sum_{n=1}^{k- 1} \left(n^{-q/p^*}- (n+1)^{-q/p^*} \right) n^q + 1 \right]^{1/q} \\ &\geqslant 1. \endaligned $$ Now, we consider $\widetilde{\dz}_1$. With directly calculating, we have $$ \aligned I\!I_n & \left(\mathbf{y}^{(k, 1)}\right) = \inf_{n\in [1, \infty)} \frac{1}{n\wedge k} \sum_{i=1}^n \left[\sum_{j=i}^{\infty} u_j \left(j \wedge k\right)^{q- 1} \right]^{p^*- 1}\\ &= \frac{1}{n \wedge k} \sum_{i=1}^{k \wedge n} \left[k^{q/p - 1} + \sum_{j=i}^{k-1} u_j j^{q-1} \right]^{p^*- 1} + \mathbbm{1}_{\{n> k\}} k^{\frac{q- p}{p- 1}} \sum_{i=k +1}^n i^{-q/p}. \endaligned $$ Obviously, $I\!I_n \left(\mathbf{y}^{(k, 1)}\right)$ is increasing when $n\geqslant k$. It means that $I\!I_n \left(\mathbf{y}^{(k, 1)}\right)$ reaches its minimum at $n\in [1, k]$. Thus, we obtain $$ \aligned \inf_{n\in [1, \infty)} I\!I_n \left(\mathbf{y}^{(k, 1)}\right) &= \inf_{n\leqslant k} \frac{1}{n} \sum_{i=1}^{n} \left[k^{q/p - 1} + \sum_{j=i}^{k-1} u_j j^{q-1} \right]^{p^*- 1} \\ &\geqslant \inf_{n\leqslant k} \left[k^{q/p - 1} + \sum_{j=n}^{k-1} u_j j^{q-1} \right]^{p^*- 1} \\ &= k^{(q/p - 1)(p^*- 1)}. \endaligned $$ Therefore, we obtain $$ \aligned \widetilde{\dz}_1 &= \sup_{k\in [1, \infty)} k^{1/q- 1/p} \left( \inf_{n\in [1, \infty)} I\!I_n \left(\mathbf{y}^{(k, 1)}\right) \right)^{(p- 1)/q} \\ &\geqslant \sup_{k\in [1, \infty)} k^{1/q- 1/p} \left[k^{(q/p - 1)(p^*- 1)} \right]^{(p- 1)/q}= 1. \qquad \square \endaligned $$ \noindent {\bf Acknowledgements} $\quad$ This paper is based on the series of studies of my supervisor Prof. M. F. Chen. Heartfelt thanks are given to my supervisor for his careful guidance and helpful suggestions. Thanks are also given to Prof. Y. H. Mao, Prof. F. Y. Wang and Prof. Y. H. Zhang for their comments and suggestions, which lead to lots of improvements of this paper. The research is supported by NSFC (Grant No. 11131003) and by the ``985'' project from the Ministry of Education in China.

https://arxiv.org/abs/math/0701940

Monochromatic triangles in two-colored plane

We prove that for any partition of the plane into a closed set $C$ and an open set $O$ and for any configuration $T$ of three points, there is a translated and rotated copy of $T$ contained in $C$ or in $O$. Apart from that, we consider partitions of the plane into two sets whose common boundary is a union of piecewise linear curves. We show that for any such partition and any configuration $T$ which is a vertex set of a non-equilateral triangle there is a copy of $T$ contained in the interior of one of the two partition classes. Furthermore, we give the characterization of these "polygonal" partitions that avoid copies of a given equilateral triple. These results support a conjecture of Erdos, Graham, Montgomery, Rothschild, Spencer and Straus, which states that every two-coloring of the plane contains a monochromatic copy of any nonequilateral triple of points; on the other hand, we disprove a stronger conjecture by the same authors, by providing non-trivial examples of two-colorings that avoid a given equilateral triple.

\section{Introduction} Euclidean Ramsey theory addresses the problems of the following kind: assume that a finite configuration $X$ of points is given; for what values of $c$ and $d$ is it true that every coloring of the $d$-dimensional Euclidean space by $c$ colors contains a monochromatic congruent copy of $X$? The first systematic treatise on this theory appears in 1973 in a series of papers \cite{erdi,erdii,erdiii} by Erd\H os, Graham, Montgomery, Rothschild, Spencer and Straus. Since that time, many strong results have been obtained in this field, often related to high-dimensional configurations (see, e.g., \cite{fra,kri1,kri2,mat} or the survey \cite{gra}); however, there are basic `low-dimensional' problems that remain open. In this paper, we consider the special case when $d=2$, $c=2$ and $|X|=3$; in other words, we study the configurations of three points in the Euclidean plane colored by two colors. We use the term \emph{triangle} to refer to any set of three points, including collinear triples of points, which we call \emph{degenerate} triangles. An $(a,b,c)$-triangle is a triangle whose edges, in anti-clockwise order, have respective lengths $a$, $b$ and~$c$. A $(1,1,1)$-triangle is also called a \emph{unit triangle}. We say that a set of points $X\subseteq \mathbb{R}^2$ is a \emph{copy} of a set of points $Y\subseteq\mathbb{R}^2$, if $X$ can be obtained from $Y$ by translations and rotations in the plane. A \emph{coloring} is a partition of $\mathbb{R}^2$ into two sets $\B$ and $\W$. The elements of $\B$ and $\W$ are called \emph{black points} and \emph{white points}, respectively. We use the term \emph{boundary of $\chi$} to refer to the common boundary of the sets $\B$ and $\W$. Given a coloring $\chi=(\B,\W)$, we say that a set of points $X$ is \emph{monochromatic}, if $X\subseteq\B$ or $X\subseteq\W$. We say that a coloring $\chi$ \emph{contains} a triangle $T$, if there exists a monochromatic set $T'$ which is a copy of $T$; otherwise, we say that $\chi$ \emph{avoids} $T$. A coloring that avoids the unit triangle is easy to obtain: consider a coloring $\chi^*$ that partitions the plane into alternating half-open strips of width $\frac{\sqrt{3}}{2}$; formally, a point $(x,y)$ is black if an only if $n\sqrt{3}<y\le\left(n+\frac{1}{2}\right)\sqrt{3}$ for some integer $n$. It can be easily checked that $\chi^*$ avoids the unit triangle. We can even change the color of some of the points on the boundaries of the strips without creating any monochromatic unit triangle. Erd\H os et al.~\cite[Conjecture~1]{erdiii} have conjectured that this is essentially the only example of colorings avoiding a given triangle: \begin{con}[Erd\H os et al.\ \cite{erdiii}]\label{con-silna} For every triangle $T$ and every coloring $\chi$, if $\chi$ avoids $T$, then $T$ is an equilateral $(l,l,l)$-triangle and $\chi$ is equal to an $l$-times scaled copy of the coloring $\chi^*$ defined above, up to possible modifications of the colors of the points on the boundary of the strips. \end{con} In Section~\ref{sec-poly} of this paper, we present a counterexample to this conjecture, and define a general class of colorings (which includes $\chi^*$ as a special case) that avoid the unit triangle. On the other hand, the following conjecture by Erd\H os et al. \cite[Conjecture~3]{erdiii} remains open: \begin{con} [Erd\H os et al.\ \cite{erdiii}]\label{con-slaba} Every coloring $\chi$ contains every nonequilateral triangle $T$. \end{con} In the past, it has been shown that Conjecture~\ref{con-slaba} holds for special types of triangles $T$ (see, e.g., \cite{erdiii,gra,sha}). Our approach is different: we prove that the conjecture is valid for a restricted class of colorings $\chi$ and arbitrary $T$. In Section~\ref{sec-uzav}, we show that every coloring that partitions $\mathbb{R}^2$ into a closed set and an open set contains every triangle $T$. Then, in Section~\ref{sec-poly}, we consider \emph{polygonal} colorings, whose boundary is a union of piecewise linear curves (see page~\pageref{def-poly} for the precise definition). We show that Conjecture~\ref{con-slaba} holds for the polygonal colorings, but there are polygonal counterexamples to the stronger Conjecture~\ref{con-silna}. In fact, we are able to characterize all these polygonal counterexamples. The following lemma from \cite{erdiii} offers a useful insight into the topic of monochromatic triangles in two-colored plane: \begin{lem} \label{lem-osm} Let $\chi$ be a coloring of the plane. The following holds: \begin{enumerate}[(i)] \item If $\chi$ contains an $(a,a,a)$-triangle for some $a>0$, then $\chi$ contains an $(a,b,c)$-triangle, for every $b,c>0$ such that $a,b,c$ satisfy the (possibly degenerate) triangle inequality. \item If $\chi$ contains an $(a,b,c)$-triangle, then $\chi$ contains an $(x,x,x)$-triangle for some $x\in\{a,b,c\}$. \end{enumerate} \end{lem} \begin{figure} \begin{center} \includegraphics[scale=0.9]{lem1_obr.eps} \end{center} \caption[Proof of Lemma~\ref{lem-osm}]{The illustration of the proof of Lemma~\ref{lem-osm}} \end{figure}\label{fig-osm} \begin{proof} The essence of the proof is the configuration in Figure~\ref{fig-osm}. The configuration consists of two $(a,a,a)$-triangles $ABC$ and $A'B'C'$, two $(b,b,b)$-triangles $ADB'$ and $A'D'B$ and two $(c,c,c)$-triangles $BDC'$ and $B'D'C$. To prove the first part of the lemma, assume, for a given $\chi$, that there is a monochromatic $(a,a,a)$-triangle $ABC$, and choose arbitrary $b$ and $c$ satisfying triangle inequality with $a$. Assume that $A$, $B$ and $C$ are all black. Furthermore, assume for contradiction that no $(a,b,c)$-triangle is monochromatic. Considering the configuration in Fig.~\ref{fig-osm}, we deduce that the points $B'$, $D$ and $D'$ are all white, otherwise one of the $(a,b,c)$-triangles $BAD$, $CAB'$ and $CBD'$ would be monochromatic. Then, $A'$ is black, due to $B'A'D'$, and $C'$ is white, due to $C'A'B$. It follows that $C'B'D$ is monochromatic, a contradiction. The second part is proved by an analogous argument: assume that $BAD$ is an all-white monochromatic triangle and that the statement does not hold. Then $B'$, $C$ and $C'$ are all black, due to $ADB'$, $ABC$ and $BDC'$. $A'$ is white, due to $A'B'C'$; $D'$ is black, due to $A'D'B$, and $B'D'C$ is monochromatic. This concludes the proof. \end{proof} From Lemma~\ref{lem-osm}, we obtain directly the following facts: \begin{cor}\label{cor-osm} For every coloring $\chi$ the following holds: \begin{enumerate}[(i)] \item $\chi$ contains every triangle if and only if $\chi$ contains every equilateral triangle. \item $\chi$ contains every non-equilateral triangle if and only if there is an $a_0>0$ such that $\chi$ contains the equilateral $(a,a,a)$-triangle for all values of $a>0$ different from $a_0$. \item $\chi$ contains an $(a,b,c)$-triangle if and only if $\chi$ contains a $(b,a,c)$-triangle. \end{enumerate} \end{cor} \section{Coloring by closed and open sets}\label{sec-uzav} The aim of this section is to prove the following result: \begin{thm}\label{thm-uzav} Let $\chi=(\B,\W)$ be a coloring such that $\B$ is closed and $\W$ is open. Then $\chi$ contains every triangle $T$. \end{thm} By Corollary~\ref{cor-osm}, it suffices to prove Theorem~\ref{thm-uzav} for the case when $T$ is an arbitrary equilateral triangle. Moreover, since scaling does not affect the topological properties of $\B$ and $\W$, we only need to consider the case when $T$ is the unit triangle. Before stating the proof, we introduce a definition and prove an auxiliary result. \begin{defi} Let $\varepsilon>0$. An $(a,b,c)$-triangle whose edge-lengths satisfy $1-\varepsilon\le a,b,c \le 1+\varepsilon$ is called \emph{an $\varepsilon$-almost unit triangle}. \end{defi} Suppose that an orthogonal coordinate system is given in the plane. For $a>0$, let $Q(a)$ be the closed square with vertices ${(a,a),(-a,a),(-a,-a),(a,-a)}$. \begin{prop}\label{prop-almost} Let $Q(3)=\B \cup \W$ be a decomposition of the square $Q(3)$ into two disjoint sets such that there is no monochromatic unit triangle in $Q(3)$. Then for every $\varepsilon > 0$ both $\B$ and $\W$ contain an $\varepsilon$-almost unit triangle. \end{prop} \begin{proof} Let $\varepsilon$ be a given positive number. Assume that we are given a partition $\B\cup\W=Q(3)$ such that $Q(3)$ does not contain any monochromatic unit triangle. For contradiction, assume that one of the classes, wlog the class $\B$, does not contain any $\varepsilon$-almost unit triangle. There is a white point $S$ and a black point $R$ in $Q(1)$ such that $|R-S| < \varepsilon$ (otherwise the whole square $Q(1)$ would be monochromatic). Let $\cC$ be the unit circle centered at $S$. For every $\alpha \in \mathbb{R}$, let $K(\alpha)$ denote the point of $\cC$ with coordinates $(x_S+\cos(\alpha), y_S+\sin(\alpha))$, where $(x_S,y_S)$ are the coordinates of $S$. Note that the distance between $R$ and any point on $\cC$ is always in the interval $(1-\varepsilon, 1+\varepsilon)$; thus, for every $\alpha$, the points $K(\alpha)$ and $K(\alpha + \frac{\pi}{3})$ must have different colors, otherwise they would form a monochromatic white unit triangle with $S$ or a monochromatic black $\varepsilon$-almost unit triangle with $R$. Let $K(\alpha_0)$ be a white point, then $K(\alpha_0 + \frac{\pi}{3})$ is black (see Fig.~\ref{fig-uzav}). Note that for every $\alpha \in (\alpha_0-\varepsilon, \alpha_0+\varepsilon)$ the distance between $K(\alpha)$ and $K(\alpha_0 + \frac{\pi}{3})$ is in the interval $(1-\varepsilon,1+\varepsilon)$, so the whole arc $\{K(\alpha); \alpha \in (\alpha_0-\varepsilon, \alpha_0+\varepsilon)\}$ is white. Let $A=\{K(\alpha); \alpha \in (\beta_1, \beta_2)\}$ be the maximal open white arc of $\cC$ containing the point $K(\alpha_0)$. Then the whole arc $\{K(\alpha); \alpha \in (\beta_1+\frac{\pi}{3},\beta_2+\frac{\pi}{3})\}$ is black. By definition of $A$, there exists $\beta \in (\beta_2, \beta_2 + \frac{\varepsilon}{2})$ such that $K(\beta)$ is black. There also exists $\gamma \in (\beta_2+\frac{\pi}{3} - \frac{\varepsilon}{2}, \beta_2+\frac{\pi}{3})$ such that $K(\gamma)$ is black. But then $(\gamma-\beta) \in (\frac{\pi}{3}-\varepsilon,\frac{\pi}{3})$, so the distance between the black points $K(\beta)$ and $K(\gamma)$ is in the interval $(1-\varepsilon, 1)$, hence the three points $R, K(\beta),K(\gamma)$ form a black $\varepsilon$-almost unit triangle\thinspace ---\thinspace a contradiction. \end{proof} \begin{figure} \begin{center}\includegraphics{uzav_obr}\end{center} \caption{Illustration of the proof of Proposition~\ref{prop-almost}}\label{fig-uzav} \end{figure} We are now ready to prove the main result of this section. \begin{proof}[Proof of Theorem~\ref{thm-uzav}] Let $\chi=(\B,\W)$ be a coloring, with $\B$ closed. By Corollary~\ref{cor-osm}, it is sufficient to show that $\chi$ contains the unit triangle. Assume, for contradiction, that this is not the case. Let $\B_0 = Q(3) \cap \B$ and let $\W_0=Q(3)\cap\W$. Clearly, neither $\B_0$ nor $\W_0$ contain the unit triangle, so by Proposition~\ref{prop-almost}, both these sets contain $\varepsilon$-almost unit triangles for every $\varepsilon>0$. In particular, the set $\B_0$ contains, for every $n\in\mathbb{N}$, a $\frac{1}{n}$-almost unit triangle $X_nY_nZ_n$. Since $\B_0$ is a compact set, the set $\B_0^3 = \B_0\times\B_0\times\B_0$ is compact as well. The sequence $\{(X_n,Y_n,Z_n); n \in \mathbb{N}\}$ is an infinite sequence of points in $\B_0^3$, so there exists a convergent subsequence $\{(X_{n_k},Y_{n_k},Z_{n_k}); k \in \mathbb{N}\}$. Let $(X,Y,Z) \in \B_0^3$ be its limit. Then $X,Y,Z\in \B$ are limits of the sequences $\{X_{n_k};k \in \mathbb{N}\}$, $\{Y_{n_k};k \in \mathbb{N}\}$, and $\{Z_{n_k};k \in \mathbb{N}\}$, respectively. The Euclidean distance is a continuous function of two variables, so $|X-Y| = \lim_{k \to \infty} |X_{n_k}-Y_{n_k}|=1$, similarly $|Y-Z|=|Z-X|=1$. Thus, $\{X,Y,Z\}$ is a black unit triangle in $Q(3)$, which is a contradiction. \end{proof} \section{Polygonal colorings}\label{sec-poly} Throughout this section, $\cC(A)$ denotes the unit circle with center $A$, and $\cD(A)$ denotes the closed unit disc with center $A$. In this section, we consider \emph{polygonal} colorings of the plane, defined as follows: \begin{defi}\label{def-poly} A coloring $\chi=(\B,\W)$ is said to be \emph{polygonal}, if it satisfies the following conditions (see an example in Fig.~\ref{fig-poly}): \begin{figure}[ht] \begin{center}\includegraphics{ob1_pol1}\end{center} \caption{Example of a polygonal coloring}\label{fig-poly} \end{figure} \begin{itemize} \item Each of the two sets $\B$ and $\W$ is contained in the closure of its interior. \item The boundary of $\chi$ (denoted by $\cB$) is a union of straight line segments (called \emph{boundary segments}). Two boundary segments may only intersect at their endpoints. We allow these segments to be unbounded, i.e., a boundary segment may in fact be a half-line or a line. An endpoint of a boundary segment is called a \emph{boundary vertex}. We may assume that if exactly two boundary segments meet at a boundary vertex, then the two segments do not form a straight angle, because otherwise they could be replaced with a single boundary segment. Note that with this condition, the boundary segments and boundary vertices of $\chi$ are determined uniquely. \item Every bounded region of the plane is intersected by only finitely many boundary segments (which implies that every bounded region contains only finitely many boundary vertices). \end{itemize} \end{defi} Note that these conditions imply that a sufficiently small disc around an interior point of a boundary segment is separated by the boundary segment into two halves, one of which is colored black and the other white. Note also that we make no assumptions about the colors of the points on the boundary~$\cB$. We say that a coloring $\chi'$ is a \emph{twin} of a coloring $\chi$ if the two colorings have the same boundary and they assign the same colors to the points outside this boundary. The main aim of this section is to prove that every polygonal coloring contains every nonequilateral triangle, and to characterize the polygonal colorings that avoid an equilateral triangle. To achieve this, we need the following definition: \begin{defi}\label{def-strip} A coloring $\chi=(\B,\W)$ is called \emph{zebra-like} if it has the following form: the boundary of $\chi$ is a disjoint union of infinitely many continuous curves $\cL_i; i\in\mathbb{Z}$ with the following properties (see Fig.~\ref{fig:except}): \begin{enumerate}[(a)] \item There is a unit vector $\vec x$ such that for every $i\in\mathbb{Z}$, $\cL_i+\vec x=\cL_i$. In other words, the $\cL_i$ are invariant upon a translation of length 1. \item For every $i\in\mathbb{Z}$, the curve $\cL_{i+1}$ is a translated copy of $\cL_i$. Moreover, there is a unit vector $\vec y$ orthogonal to $\vec x$, so that \[ \cL_{i+1}=\cL_i+\frac{1}{2}\vec x+\frac{\sqrt{3}}{2}\vec y. \] In other words, for an arbitrary boundary point $X\in\cL_i$, the points $Y=X+\vec x$ and $Z=X+\frac{1}{2}\vec x+\frac{\sqrt{3}}{2}\vec y$ belong to the boundary as well. Note that $XYZ$ is a unit triangle, and that $Y\in\cL_i$ and $Z\in\cL_{i+1}$. \item For every $i\in\mathbb{Z}$, the interior of the region delimited by $\cL_i\cup\cL_{i+1}$ is colored with a different color than the interior of the region delimited by $\cL_{i-1}\cup\cL_{i}$. \item For two points $A$ and $B$, let $\theta_{AB}$ denote the size of the acute angle formed by the segment $AB$ and the vector $\vec x$. For every $i\in\mathbb{Z}$ and every two points $A\in\cL_i$ and $B\in\cL_{i+1}$, the following holds: $\|AB\|>1$ if and only if $\theta_{AB}<\frac{\pi}{3}$. This last condition can also be stated in the following equivalent form: Let $A\in\cL_i$ be an arbitrary point on the boundary. Let $B_1=A-\frac{1}{2}\vec x+\frac{\sqrt{3}}{2}\vec y$ and $B_2=A+\frac{1}{2}\vec x+\frac{\sqrt{3}}{2}\vec y$ (the two points $B_1, B_2$ belong to $\cL_{i+1}$ by the previous conditions), and let $A'=A+\sqrt{3}\vec y$ (so that $A'\in\cL_{i+2}$). Under these assumptions, the portion of $\cL_{i+1}$ between $B_1$ and $B_2$ is contained inside of the closed lens-shaped region $\cD(A)\cap\cD(A')$ and no other point of $\cL_{i+1}$ is inside this region. \end{enumerate} \end{defi} We stress that a zebra-like coloring is not necessarily polygonal. \begin{figure}[htb] \begin{center} \includegraphics[scale=0.7]{except_coloring} \end{center} \caption{The boundary of a zebra-like coloring} \label{fig:except} \end{figure} \subsection{The result} The following theorem is the main result of this section: \begin{thm}\label{thm-poly} For a polygonal coloring $\chi$, the following conditions are equivalent: \begin{enumerate} \item[(C1)] The coloring $\chi$ is a zebra-like polygonal coloring. \item[(C2)] The coloring $\chi$ has a twin $\chi'$ which avoids the unit triangle. \item[(C3)] For every monochromatic unit triangle $ABC$, at least one of the three points $A,B$ and $C$ belongs to the boundary of $\chi$. \end{enumerate} \end{thm} Clearly, the condition (C2) of Theorem~\ref{thm-poly} implies the condition (C3), so we only need to prove that (C1) implies (C2) and that (C3) implies (C1). The proof is organized as follows: we first prove that (C3)$\Rightarrow$(C1). This part of the proof proceeds in several steps: first of all, we use the condition (C3) to describe the set $\cB(\chi)\cap\cC(A)$, where $A$ is a boundary point. Then we apply a continuity argument to extend this information into a global description of~$\chi$. Next, in Theorem~\ref{thm-obarv}, we prove that every (not necessarily polygonal) zebra-like coloring has a twin that avoids the unit triangle, which shows that (C1)$\Rightarrow$(C2), completing the proof of Theorem~\ref{thm-poly}. In the last part of this section, we show that Theorem~\ref{thm-poly} implies that every polygonal coloring contains a monochromatic copy $T$ of a given non-equilateral triangle, with the vertices of $T$ avoiding the boundary. \subsection{The proof} We begin with an auxiliary lemma: \begin{lem}\label{lem-triprimky} Let $q_1, q_2, q_3$ be (not necessarily distinct) lines in the plane, not all three parallel. Then exactly one of the following possibilities holds: \begin{enumerate} \item The lines $q_1, q_2, q_3$ intersect at a common point and every two of them form an angle $\frac{\pi}{3}$. \item There exist only finitely many unit triangles $ABC$ such that $A \in q_1$, $B \in q_2$ and $C \in q_3$. \end{enumerate} \end{lem} \begin{proof} It can be easily checked that the two conditions cannot hold simultaneously: in fact, if the three lines satisfy the first condition, then for every point $A\in q_1$ whose distance from the other two lines is at most 1 there are points $B\in q_2$ and $C\in q_3$ such that $ABC$ is a unit triangle. We now show that at least one of the two conditions holds. Since the three lines are not all parallel, we may assume that neither $q_1$ nor $q_2$ is parallel to $q_3$. Consider a Cartesian coordinate system whose $y$\nobreakdash-axis is $q_3$. There exist real numbers $a_1, a_2, b_1, b_2$ such that for $i \in \{1,2\}$ we have $q_i=\{(x,y)\in \mathbb{R}^2; y=a_ix+b_i\}$. Let $ABC$ be a unit triangle with $A=(x_1,y_1) \in q_1$, $B=(x_2,y_2) \in q_2$ and $C\in q_3$, and assume that $A,B,C$ are in the counter-clockwise order (the other case is symmetric). Then $C=(\frac{x_1+x_2}{2},\frac{y_1+y_2} {2})+\frac{\sqrt{3}}{2}(y_1-y_2, x_2-x_1)$. The point $C$ lies on $q_3$, which implies the following equality: \begin{equation} \frac{x_1+x_2}{2}+\frac{\sqrt{3}(y_1-y_2)}{2}=0 \label{equa1} \end{equation} Points $A$ and $B$ are at the distance $1$, from which we get \begin{equation} (x_1-x_2)^2+(y_1-y_2)^2=1 \label{equa2} \end{equation} By combining \eqref{equa1} and \eqref{equa2} and eliminating $y_1, y_2$ we get \[ \left(\frac{x_1+x_2}{2}\right)^2={\frac{3}{4}}\left(1-(x_1-x_2)^2\right), \] which yields \begin{equation} x_1^2+x_2^2-x_1x_2={\frac{3}{4}}. \label{equa3} \end{equation} Substituting $y_1=a_1x_1+b_1$ and $y_2=a_2x_2+b_2$ into \eqref{equa1} gives \begin{equation} {\frac{1+\sqrt{3}a_1}{2}}x_1 + {\frac{1-\sqrt{3}a_2}{2}}x_2 + {\frac{\sqrt{3}}{2}}(b_1-b_2)=0 \label{equa4} \end{equation} If both $\frac{1+\sqrt{3}a_1}{2}$ and $\frac{1-\sqrt{3}a_2}{2}$ are equal to zero, then the equality \eqref{equa4} degenerates and we get that $a_1=-\frac{1}{\sqrt{3}}$, $a_2=\frac{1}{\sqrt{3}}$ and $b_1=b_2$, so the first case of the statement holds. In the other case, suppose (wlog) that $\frac{1+\sqrt{3}a_1}{2} \neq 0$. From \eqref{equa4} we can obtain that $x_1=cx_2+d$ for some reals $c,d$. By substituting it into \eqref{equa3} we get a quadratic equation for the variable $x_2$, where the leading coefficient is equal to $c^2-c+1=(c-\frac{1}{2})^2+\frac{3}{4} > 0$, so there exist at most two possible values for $x_2$, thus at most two possible locations of $B$ and at most four possible unit triangles $ABC$. \end{proof} Throughout the rest of this section, we assume that $\chi$ is a fixed polygonal coloring satisfying the condition (C3) of Theorem~\ref{thm-poly}. Every boundary segment can be regarded as a common edge of two (possibly unbounded) polygonal regions, one of which is white and the other black. We choose an orientation of the boundary segments in the following way: a boundary segment with endpoints $A$ and $B$ is directed from $A$ to $B$ if the white region adjacent to this segment is on the left hand side from the point of view of an observer walking from $A$ to~$B$. \begin{defi}\label{def-vhodny} A boundary point $A\in\cB$ is called \emph{feasible}, if $A$ is not a boundary vertex, and the unit circle $\cC(A)$ does not contain any boundary vertex. An \emph{infeasible} point is a point on the boundary that is not feasible. \end{defi} We may easily see that every bounded subset of the plane contains only finitely many infeasible points. The first step in the proof of the main result is the description of the set of all the boundary points at the unit distance from a given feasible point $A$. Let $A$ be a fixed feasible point, let $s$ be the boundary segment containing $A$. The set $\cB\cap\cC(A)$ is finite, by the definition of polygonal coloring; on the other hand, this set is nonempty, otherwise we could find two points $B,C$ of $\cC(A)$ such that $ABC$ is a unit triangle, with $B$ and $C$ in the interior of the same color class. By shifting the triangle $ABC$ slightly in a suitable direction, we would obtain a monochromatic unit triangle avoiding the boundary, which is forbidden by the condition (C3). In the following arguments, we will use a Cartesian coordinate system whose origin is the point $A$, and whose $x$-axis is parallel to $s$ and has the same orientation. We shall assume that the $x$-axis and the segment $s$ is directed left-to-right and the $y$-axis is directed bottom-to-top. Assuming this coordinate system, we let $P(\alpha,A)$ denote the point of $\cC(A)$ with coordinates $\left(\cos(\alpha),\sin(\alpha)\right)$. If no ambiguity arises, we write $P(\alpha)$ instead of $P(\alpha, A)$. \begin{lem} Let $B=P(\alpha)$ be an arbitrary element of $\cB\cap\cC(A)$, let $t$ be the boundary segment containing $B$ (the segment $t$ is determined uniquely, because $A$ is a feasible point). Then the segments $s$ and $t$ are parallel. \end{lem} \begin{proof} For contradiction, assume that $s$ and $t$ are not parallel, let $\sigma\in(0,\pi)$ be the angular slope of $t$ with respect to the coordinate system established above, i.e., $\sigma$ is the angle formed by the lines containing $s$ and $t$. First of all, note that the point $C=P(\alpha+\frac{\pi}{3})$ lies on the boundary $\cB$; otherwise, a sufficiently small translation of the unit triangle $ABC$ in a suitable direction would yield a counterexample to condition (C3) (here we use the assumption that $s$ and $t$ are not parallel). Let $u$ be the boundary segment containing $C$, and let $\tau$ be the angular slope of $u$. Secondly, we may deduce that $\{\sigma,\tau\}=\{\frac{\pi}{3},\frac{2\pi}{3}\}$, and the three lines containing $s$, $t$ and $u$ all meet at one point. If this were not the case, then by Lemma~\ref{lem-triprimky} there would be only finitely many unit triangles with vertices belonging to the three segments $s$, $t$ and $u$. Thus, we could find a unit triangle $A'B'C'$ with $A'\in s$, $B'\in t$ and $C'\not\in\cB$, which is impossible, by the argument presented in the previous paragraph. By repeating this argument with $\{\alpha+\frac{i\pi}{3};\ i=1,\dots,5\}$ in place of $\alpha$, we obtain the following conclusions: \begin{itemize} \item The six points $\{P(\alpha+\frac{i\pi}{3});\ i=1,\dotsc,6\}$ all belong to the boundary $\cB$. \item The lines passing through the boundary segments containing these six points all meet at one point. \item The boundary segments containing $P(\alpha)$, $P(\alpha+\frac{2\pi}{3})$ and $P(\alpha+\frac{4\pi}{3})$ all have the same slope. \end{itemize} This is a contradiction, because three parallel segments intersecting a circle in three distinct points cannot belong to a single line, and two parallel lines do not intersect. \end{proof} \begin{lem}\label{lem-pipul} $P(\frac{\pi}{2})\not\in\cB$, $P(-\frac{\pi}{2})\not\in\cB$. \end{lem} \begin{proof} For contradiction, assume that $B=P(\frac{\pi}{2})\in\cB$ (the case of $P(-\frac{\pi}{2})$ is symmetric), let $t$ denote the boundary segment containing $B$. Let $C=P(\frac{\pi}{6})$. We distinguish the following cases: \begin{itemize} \item The segment $t$ has the same orientation as the segment $s$. In this case, by applying a rotation around the center $C$ and then, if $C\in\cB$, a suitable translation, we may transform the triple $ABC$ into a monochromatic triple with vertices avoiding the boundary, contradicting (C3). \item The segments $s$ and $t$ have opposite orientations (i.e., $t$ is oriented right-to-left, which means that there is a white region touching $t$ from below); furthermore, either $C\in\cB$ or $C$ is in the interior of the white color. In such case, we may rotate the configuration $ABC$ around the center of the segment $AB$ to obtain a unit triangle in the interior of the white color. \item The segments $s$ and $t$ have opposite orientations and the point $C$ is in the interior of the black color. Let $\theta$ be the maximal angle with the properties that for every $\alpha\in (\frac{\pi}{2},\frac{\pi}{2}+\theta)$ the point $P(\alpha)$ lies in the interior of the white color and for every $\alpha\in(\frac{\pi}{6},\frac{\pi}{6}+\theta)$ the point $P(\alpha)$ lies in the interior of the black color. The value of $\theta$ is well defined, and by the previous assumptions, $0<\theta<\frac{\pi}{3}$. Let $B'=P(\frac{\pi}{2}+\theta)$ and $C'=P(\frac{\pi}{6}+\theta)$. By the maximality of $\theta$, at least one of the two points lies on the boundary, and the boundary segment passing through this point is directed left-to-right (see Fig.~\ref{fig-pipul}). As in the first case of this proof, we may rotate and translate the configuration $AB'C'$ to obtain a monochromatic unit triangle. \end{itemize} In all cases we get a contradiction. \end{proof} \begin{figure}[ht!] \begin{center} \includegraphics{lemma3} \end{center} \caption{Illustration of the proof of Lemma~\ref{lem-pipul}} \label{fig-pipul} \end{figure} The previous two lemmas imply that if $A$ is a feasible point, then no boundary segment is tangent to $\cC(A)$. \begin{lem}\label{lem-bilapod} Let $B=P(\alpha)\in\cB$ be a point on the boundary, let $t$ be the boundary segment containing this point. If $\alpha\in (\frac{\pi}{6}, \frac{5\pi}{6})$ or $\alpha\in(\frac{7\pi}{6},\frac{11\pi}{6})$, then $s$ and $t$ have opposite orientation. If $|\alpha|<\frac{\pi}{6}$ or $|\alpha-\pi|<\frac{\pi}{6}$, then $s$ and $t$ have the same orientation. \end{lem} \begin{proof} We first consider the case $\alpha\in(\frac{\pi}{6}, \frac{5\pi}{6})$ or $\alpha\in(\frac{7\pi}{6},\frac{11\pi}{6})$. The proof is analogous to the proof of the first part of Lemma~\ref{lem-pipul}: if $t$ had the same orientation as $s$, we could take $C=P(\frac{\pi}{3}+\alpha)$ and then by rotating and translating the unit triangle $ABC$ we would get a contradiction. Note that the condition $\alpha\in(\frac{\pi}{6}, \frac{5\pi}{6})\cup(\frac{7\pi}{6},\frac{11\pi}{6})$ guarantees that $C$ is either the leftmost or the rightmost point of the triangle $ABC$, so whenever we start rotating the triangle $ABC$ around $C$, the two points $A,B$ move into the interior of the same color. The case $|\alpha|<\frac{\pi}{6}$ or $|\alpha-\pi|<\frac{\pi}{6}$ can be proven analogously. \end{proof} \begin{lem}\label{lem-pitre} $P(\alpha)\in\cB$ if and only if $P(\alpha+\frac{\pi}{3})\in\cB$. \end{lem} \begin{proof} It suffices to prove one implication, the other case is symmetric. Assume that for some $\alpha$ we have $P(\alpha)\in\cB$ and $P(\frac{\pi}{3}+\alpha)\not\in\cB$. Let $B=P(\alpha)$, $C=P(\frac{\pi}{3}+\alpha)$, and let $t$ be the boundary segment containing $B$. We consider the following cases: \begin{itemize} \item If $s$ and $t$ have opposite orientation, we may rotate $ABC$ around the center of $AB$ to obtain a monochromatic unit triangle in the interior of one color (see Fig.~\ref{fig-pitre}). Here we use the fact that $\alpha\neq\frac{\pi}{2}$, which follows from Lemma~\ref{lem-pipul}. \item If $s$ and $t$ have the same orientation, a small translation in a suitable direction transforms $ABC$ into a monochromatic unit triangle. \end{itemize} In both cases we get a contradiction. \end{proof} \begin{figure}[Ht!] \begin{center} \includegraphics{lemma5.eps} \end{center} \caption{Illustration of the proof of Lemma~\ref{lem-pitre}} \label{fig-pitre} \end{figure} \begin{lem} For every $\theta$ there is exactly one value of $\alpha\in[\theta,\theta+\frac{\pi}{3})$ such that $P(\alpha)\in\cB$. \end{lem} \begin{proof} By Lemma~\ref{lem-pitre}, if the statement holds for some value of $\theta$, it holds for all other values of $\theta$ as well. Thus, it is enough to prove the lemma for $\theta=\frac{\pi}{2}$. Clearly, there is at least one $\alpha\in[\frac{\pi}{2},\frac{5\pi}{6})$ such that $P(\alpha)\in\cB$; otherwise, the set $\cC(A)\cap\cB$ would be empty, which is impossible. Assume that there are $\alpha$ and $\alpha'$ such that $\frac{\pi}{2}\le\alpha<\alpha'<\frac{5\pi}{6}$ with $P(\alpha)\in\cB$ and $P(\alpha')\in\cB$. Let us fix $\alpha$ and $\alpha'$ as small as possible. Let $t$ and $t'$ be the boundary segments containing $P(\alpha)$ and $P(\alpha')$. The circle $\cC(A)$ consists of alternating black and white arcs and one of these arcs has $P(\alpha)$ and $P(\alpha')$ for endpoints. It follows that one of the segments $t$, $t'$ has the same orientation as the segment~$s$, contradicting Lemma~\ref{lem-bilapod}. \end{proof} Before we proceed with the proof of the main result, we summarize the lemmas proved so far (and introduce some related notation) in the following claim (see Fig.~\ref{fig-sum}): \begin{cla}\label{tvrz-sum} Let $A\in\cB$ be an arbitrary feasible point. The circle $\cC(A)$ intersects the boundary $\cB$ at exactly six points, which form the vertex set of a regular hexagon. These six points will be denoted by $P_0(A),\dotsc,P_5(A)$, where $P_i(A)= P(\alpha+\frac{i\pi}{3},A)$ with $\alpha\in\left( -\frac{\pi}{6},\frac{\pi}{6} \right)$ (this determines $P_i(A)$ uniquely). The boundary segments containing the six points $P_i(A)$ are all parallel to the boundary segment $s$ containing the point $A$. The boundary segments containing the points $P_0(A)$ and $P_3(A)$ have the same orientation as $s$, whereas the boundary segments containing $P_1(A)$, $P_2(A)$, $P_4(A)$ and $P_5(A)$ have opposite orientation. \end{cla} \begin{figure}[Ht!] \begin{center} \includegraphics[scale=.9]{tvrz7.eps} \end{center} \caption{Illustration of Claim~\ref{tvrz-sum}} \label{fig-sum} \end{figure} Now we use Claim~\ref{tvrz-sum} to get more global information about the boundary. \begin{lem} \label{lem-uhel} Let $u_1$ and $u_2$ be two boundary segments that share a common endpoint $X$. The size of the convex angle formed by these two segments is greater than $\frac{2\pi}{3}$. \end{lem} \begin{figure}[ht] \begin{center}\includegraphics{ob2_le11_mod}\end{center} \caption{Illustration of the proof of Lemma~\ref{lem-uhel}}\label{fig-uhel} \end{figure} \begin{proof} For contradiction, assume that for some $u_1$, $u_2$ and $X$, the statement of the lemma does not hold (see Fig.~\ref{fig-uhel}). We may assume that the convex angle determined by $u_1$ and $u_2$ does not contain any other boundary segment with endpoint $X$. Furthermore, we may assume that the segment $u_1$ is directed from $X$ to the other endpoint. For $0<t<|u_1|$, let $A(t) \in u_1$ denote the point with $|A(t)-X|=t$ and let $A'(t)=P_4(A(t))$. There exists $\varepsilon >0$ such that for all $0<t<\varepsilon$ the points $A(t)$ are feasible, the points $A'(t)$ are feasible as well and lie on a common boundary segment. By our assumption, the convex angles between the ray $A(t)A'(t)$ and the segments $u_1, u_2$ directed from $X$ are both greater than $\frac{\pi}{2}$. It follows that if $t$ is sufficiently small, the tangent to the circle $\cC(A(t))$ at $A(t)$ intersects both segments $u_1, u_2$ and so does the circle $\cC(A(t))$, contradicting Claim~\ref{tvrz-sum}. \end{proof} An important consequence of Lemma~\ref{lem-uhel} is that no three boundary segments share a common endpoint. Hence, every connected component of the boundary is either an infinite piecewise linear curve, or a simple closed piecewise linear curve (i.e. the boundary of a simple polygon). We will call these cuves \emph{boundary components} or simply \emph{components}. \begin{defi} Let $A$ be a point on the boundary. For $t\in\mathbb{R}$, let $A(t)$ denote the point of the same boundary component as $A$, such that the directed length of the part of the boundary starting at $A$ and ending at $A(t)$ is equal to~$t$. $A(t)$ is clearly a continuous function of $t$. If $A(t)$ is a feasible point, we let $p_i(t)=P_i(A(t))$, for $i=0,\dotsc,5$. \end{defi} It is easy to see that the functions $p_i$ are continuous on a sufficiently small neighborhood of every value of $t$ for which $A(t)$ is a feasible point. Our next aim is to show that these functions can be extended into continuous functions by suitably defining the values of $p_i(t)$ when $A(t)$ is not feasible. It is not obvious that the functions $p_i$ can be extended in this way: the definition of $P_i(A(t))$ uses the Cartesian system whose $x$-axis is parallel with the boundary segment containing $A(t)$. Hence, if $A_1$ and $A_2$ are two feasible points belonging to two distinct boundary segments of the same boundary component, it might not be immediately clear that $P_i(A_1)$ belongs to the same boundary component as $P_i(A_2)$. The next lemma shows that these technical difficulties can be overcome. \begin{lem}\label{lem-limita} Let $A(t_0)$ be an infeasible point. For every $i=0,\dotsc,5$, there is a point $P_i\in\cB$ such that \[ \lim_{t\to t_0-}p_i(t)=P_i=\lim_{t\to t_0+} p_i(t) \] This means that if we define $p_i(t_0)=P_i$, then $p_i$ is continuous at $t_0$. \end{lem} \begin{proof} It is sufficient to prove the lemma for $i=0$, because $p_i(t)$ is clearly a continuous function of $A(t)$ and $p_0(t)$. Since every boundary segment contains only finitely many infeasible points, we may choose a sufficiently small $\varepsilon>0$, such that for every $t$ from the open interval $(t_0-\varepsilon,t_0)$ the points $A(t)$ are feasible and they all belong to a single boundary segment $u_1$, and similarly, for every $t'\in (t_0,t_0+\varepsilon)$ the points $A(t')$ are feasible, and they belong to a single boundary segment $u_2$. If the segments $u_1$ and $u_2$ are distinct, then $A(t_0)$ is their common endpoint. Note that for $t \in(t_0-\varepsilon,t_0)$, the points $p_0(t)$ all belong to a single boundary segment $v_1$, otherwise some of the $A(t)$ would not be feasible. By Claim~\ref{tvrz-sum}, the segment $v_1$ is parallel and consistently oriented with $u_1$. Similarly, for $t'\in(t_0,t_0+\varepsilon)$ the points $p_0(t')$ belong to a single boundary segment $v_2$, parallel and consistently oriented with $u_2$. We do not know yet whether $v_1$ and $v_2$ appear consecutively on the same component of the boundary. Let $B=\lim_{t\to t_0-} p_0(t)$ (clearly, the limit exists, because the points $\{p_0(t);\, t\!\in\!(t_0-\varepsilon,t_0)\}$ form an open segment whose endpoint is $B$). See Fig.~\ref{fig-limita}. \begin{figure}[ht] \begin{center}\includegraphics{ob3_le12_mod}\end{center} \caption{Illustration of the proof of Lemma~\ref{lem-limita}}\label{fig-limita} \end{figure} For $t\in (t_0-\varepsilon,t_0)$, let us fix $\alpha\in(-\frac{\pi}{6},\frac{\pi}{6})$ such that $p_0(t)=P(\alpha,A(t))$, i.e., $\alpha$ is the (signed) measure of the angle between the segment $u_1$ and the segment $A(t)p_0(t)$. Note that $\alpha$ does not depend on the choice of $t$. The circle $\cC(A(t_0))$ intersects the boundary at $B$. Let $w$ be the boundary segment starting at $B$ and directed away from $B$. By Lemma~\ref{lem-uhel}, the convex angles determined by $v_1$ and $w$ and by $u_1$ and $u_2$ have size at least $\frac{2\pi}{3}$, which implies that the convex angle $\alpha'$ between $u_2$ and $BA(t_0)$ is acute and the convex angle between $w$ and $BA(t_0)$ is obtuse. Thus, for $t'\in(t_0,t_0+\varepsilon)$ the circle $\cC(A')$ (where $A'=A(t')$) intersects the segment $w$ at a point $B'=p_i(t')$. From Claim~\ref{tvrz-sum} it follows that $w$ is parallel to $u_2$. Also, the segment $A'B'$ is parallel to the segment $A(t_0)B$, which is in turn parallel to any of the segments $A(t)p_0(t)$, for $t\in(t_0-\varepsilon,t_0)$. To finish the proof of this lemma, we need to show that $B'=p_0(t')$ (as opposed to $B'=p_i(t')$ for some $i\neq 0$), i.e., we need to prove that the angle $\alpha'$ determined by the segment $u_2$ and the segment $A'B'$ falls into the range $(-\frac{\pi}{6},\frac{\pi}{6})$. We have observed that $\alpha'\in(-\frac{\pi}{2},\frac{\pi}{2})$. This leaves us with the following three possibilities: either $B'=p_5(t')$, or $B'=p_1(t')$, or $B'=p_0(t')$. However, the former two possibilities are ruled out by the fact that the segment $w$ is oriented consistently with the segment $u_2$. This concludes the proof. \end{proof} \begin{lem}\label{lem-posun} Let $i\in\{0,\dotsc,5\}$, let $A\in\cB$ be an arbitrary boundary point. All the unit segments of the form $A(t)p_i(t)$ have the same slope, independently of the choice of $t$. \end{lem} \begin{proof} The slope of $A(t)p_i(t)$ (as a function of $t$) is constant in a neighborhood of every $t$ for which $A(t)$ is feasible. Moreover, this slope is a continuous function of $t$, which follows from Lemma~\ref{lem-limita}. Hence the function is constant on the whole range. \end{proof} Lemma~\ref{lem-posun} shows that every translation that maps a feasible point $A$ to the point $P_i(A)$ also maps the boundary component containing $A$ onto the boundary component containing $P_i(A)$ (which may be the same component). Composing such translations (or their inverses) we conclude that the translations that send $P_i(A)$ to $P_j(A)$ have the same component-preserving property. For the proof of Lemma~\ref{lem-usek}, we will need a slight extension of Claim~\ref{tvrz-sum} to infeasible points: \begin{cla}\label{cla-infeasible_body} Let $A\in\cB$ be an arbitrary infeasible point. \begin{enumerate}[(i)] \item At each of the six points $P_0(A), P_1(A), \dots, P_5(A)$ the circle $\cC(A)$ properly crosses the corresponding boundary component, i.e., in a sufficiently small neighborhood of such point, the circle $\cC(A)$ separates the boundary component into two portions, one lying inside $\cC(A)$ and the other one lying outside $\cC(A)$. \item There are no more proper crossings of $\cC(A)$ with boundary components. (However, $\cC(A)$ may touch the boundary at some other points.) \item The boundary components containing the points $P_0(A)$ and $P_3(A)$ have the same orientation as the component containing $A$, whereas the boundary components containing $P_1(A)$, $P_2(A)$, $P_4(A)$ and $P_5(A)$ have opposite orientation. \end{enumerate} \end{cla} \begin{proof} The first two statements follow from the fact that $\cC(A)$ has the same number of proper crossings with the boundary as the circle $\cC(A(t))$, where $A(t)$ is a feasible point sufficiently close to $A$. The third statement follows from Claim~\ref{tvrz-sum} applied to the point $A(t)$. \end{proof} \begin{lem}\label{lem-usek} Let $A\in\cB$ be an arbitrary boundary point. For the sake of brevity, let us write $P_i$ instead of $P_i(A)$, $\cC$ instead of $\cC(A)$ and $\cD$ instead of $\cD(A)$ in the statement and proof of this lemma. The point $P_1$ belongs to the same boundary component as $P_2$, the point $P_0$ belongs to the same boundary component as $A$ and $P_3$, and the point $P_4$ belongs to the same boundary component as $P_5$. The four portions of the boundary that connect $P_1$ with $P_2$, $P_0$ with $A$, $A$ with $P_3$, and $P_4$ with $P_5$ are all translated copies of a single piecewise linear curve. These four portions of the boundary are all contained in the closed unit disc with center~$A$. \end{lem} \begin{proof} It suffices to show that the boundary component that enters inside $\cD$ at $P_1$ leaves $\cD$ at $P_2$. The rest of the statement follows from Lemma~\ref{lem-posun}. Let $\cL$ be the boundary component that contains $P_1$. Let us follow $\cL$ from $P_1$ in the direction of its orientation, i.e., into the interior of the unit disc $\cD$, and let $X$ be the first point where $\cL$ leaves $\cC$. We observe the following: \begin{itemize} \item $X$ is neither $P_3$ nor $P_5$, because in these points, the boundary is oriented into the interior of the disc $\cD$. \item $X$ is not the point $P_0$: if $X=P_0$, then the translation $P_0\mapsto A$ would map the fragment of the boundary between $P_1$ and $P_0$ onto a fragment directed from $P_2$ to $A$. Similarly, the translation $P_1\mapsto A$ would map the fragment $P_1P_0$ onto a fragment directed from $P_5$ to $A$. This is impossible, because two different boundary fragments of equal length cannot both end at $A$. \item $X$ is not $P_4$: if $X$ were equal to $P_4$, we would consider the boundary component that enters into the interior of $\cC$ at the point $P_3$. Since this boundary component cannot intersect the boundary fragment between $P_1$ and $P_4$, it must leave the interior of $\cC$ at the point $P_2$. However, this is symmetric to the previous case and leads to contradiction in the same way. \item Having excluded all other possibilities, we know that $X=P_2$. \end{itemize} Let $U$ denote the fragment of $\cL$ between $P_1$ and $P_2$. By definition, this fragment properly crosses $\cC$ only at its endpoints. Applying a symmetric argument, we find that the boundary fragment from $P_5$ to $P_4$ (which is a translated copy of $U$) properly crosses $\cC$ only in its endpoints. Translating $U$ appropriately, we obtain the boundary fragments connecting $P_3$ with $A$ and $A$ with $P_0$. This concludes the proof. \end{proof} From the previous lemmas, we readily obtain the following claim. \begin{cla}\label{cla-nutne} The condition (C3) of Theorem~\ref{thm-poly} implies the condition (C1). \end{cla} \begin{proof} We check that the coloring $\chi$ satisfies the conditions of Definition~\ref{def-strip}. Let $\vec x$ denote the unit vector $\overrightarrow{AP_0}$ and let $\vec y$ be a unit vector orthogonal to $\vec x$. By Lemma~\ref{lem-usek}, every component of the boundary is a piecewise linear $\vec x$-periodic curve and if $\cL$ is a boundary component, then any other component is a translate of $\cL$ by an integral multiple of the vector $\overrightarrow{AP_1}=\frac{1}{2}\vec x+\frac{\sqrt{3}}{2}\vec y$. Let $\vec z$ denote this last vector and let $\cL_i=\cL_0+i\vec z$, $i\in\mathbb{Z}$, where $\cL_0$ is a boundary component chosen arbitrarily. We have $\cB=\bigcup_{i\in\mathbb{Z}}\cL_i$. The condition $(d)$ of Definition~\ref{def-strip} follows from Lemma~\ref{lem-usek}. \end{proof} It remains to show that the condition (C1) implies (C2). This is the easier part of the proof. In fact, we prove a more general claim: \begin{thm}\label{thm-obarv} Every zebra-like coloring has a twin that avoids the unit triangle. \end{thm} \begin{proof} Let $\chi$ be a zebra-like coloring, let $\cL_i$, $\vec x$ and $\vec y$ be as in Definition~\ref{def-strip}. Let $\vec z=\frac{1}{2}\vec x+\frac{\sqrt{3}}{2}\vec y$. Let $\chi'$ be the twin coloring of $\chi$ such that the points of $\cL_i$ are black in $\chi'$ if $i$ is even and white if $i$ is odd. Observe that by the definition of the coloring, the color of a point $P$ is equal to the color of $P+\vec x$ and different from the color of $P+\vec z$. Now assume that $ABC$ is a monochromatic unit triangle, wlog the three points are black. By the previous observation, no edge of the triangle forms an angle of size $\frac{\pi}{3}$ (or $\frac{2\pi}{3}$) with the vector $\vec x$. It follows that exactly one of the three edges (wlog the edge $AB$) forms with $\vec x$ an angle whose size falls into the range $(\frac{\pi}{3},\frac{2\pi}{3})$. We claim that the three points $A,B,C$ all belong to a single connected component of the black color: otherwise one of the two edges $AC$ and $BC$ would have to intersect (at least) two curves $\cL_i$ and $\cL_{i+1}$. By the definition of the coloring, the distance between the two points of intersection is greater than~1, contradicting the fact that $ABC$ is a unit triangle. We now deduce that $\|AB\|<1$: let $\ell$ be the line containing the segment $AB$. Note that the line $\ell$, as well as any other line not parallel with $\vec x$, must intersect all the curves $\cL_i$. Let $A'B'$ be the segment obtained as the convex hull of the intersection of $\ell$ with the closure of the black component containing $A$ and $B$. By the definition of the coloring, $\|A'B'\|\le 1$. Moreover, since the two points $A'$ and $B'$ belong to two adjacent boundary curves $\cL_i$ and $\cL_{i+1}$, they have different colors. Hence, the segment $AB$ is a proper subset of the segment $A'B'$, and $\|AB\|<1$. This shows that $ABC$ is not a unit triangle\thinspace ---\thinspace a~contradiction. \end{proof} This concludes the proof of Theorem~\ref{thm-poly}. Next, we present a simple corollary, which shows that every polygonal coloring of the plane contains any nonequilateral triangle. \subsection{Nonequilateral triangles} The following result is a direct consequence of Theorem~\ref{thm-poly}, by an easy modification of the proof of Lemma~\ref{lem-osm}. \begin{thm}\label{thm-osm} Let $XYZ$ be a nonequilateral triangle, let $\chi$ be a polygonal coloring. There is a monochromatic copy $X'Y'Z'$ of the configuration $XYZ$, such that none of the three points $X',Y'$ and $Z'$ belongs to the boundary of $\chi$. \end{thm} \begin{proof} Let $a, b$ and $c$ be the lengths of the three edges of $XYZ$. Wlog, assume that $a\neq b$. From Theorem~\ref{thm-poly} it follows that no polygonal coloring can simultaneously avoid copies of equilateral triangles of two different sizes. Hence, we may assume that $\chi$ contains a monochromatic equilateral triangle $ABC$ with edges of length $a$ whose vertices avoid the boundary of $\chi$. Assume that the three points $A$, $B$ and $C$ are all black. Consider the configuration of eight points on Fig.~\ref{fig-osm}. As discussed in the proof of the first part of Lemma~\ref{lem-osm}, every coloring of the five points $D$, $A'$, $B'$, $C'$ and $D'$ yields a monochromatic $(a,b,c)$-triangle. Furthermore, we may assume that the eight points all avoid the boundary of $\chi$, otherwise we might shift the configuration slightly to move the points away from the boundary, without changing the color of $ABC$ (recall that $A, B$ and $C$ already belong to the interior of the black color). This concludes the proof. \end{proof} \section{Concluding remarks} The Conjecture~\ref{con-slaba} remains wide open, despite the indirect support from the results of this paper, as well as from earlier research. It might happen that the validity of this conjecture would depend on the particular choice of set-theoretical axioms. Such issues do not arise in this paper, since our proof techniques are very elementary. Unfortunately, these elementary techniques do not offer much hope for broad generalizations. It might nevertheless be possible to extend our results about polygonal colorings to some broader class of colorings, e.g., the colorings by monochromatic regions bounded by continuous curves. Colorings of this kind have already been studied in the context of the related problem of the chromatic number of the plane (see \cite{woo}). The zebra-like colorings provide a hitherto unknown example of colorings that avoid an equilateral triangle. We are not aware of any other examples of colorings avoiding a given triangle, but we do not dare to make any conjectures about the uniqueness of our construction, because our understanding of non-polygonal colorings is rather limited. \section*{Acknowledgments} We appreciate the useful discussions with Zden\v ek Dvo\v r\'ak, Jan Kratochv\'\i l, Martin Tancer, Pavel Valtr and Tom\'a\v s Vysko\v cil.

https://arxiv.org/abs/1810.02671

On the number of vertices of projective polytopes

Let $X$ be a configuration of $n$ points in $\mathbb{R}^d$.What is the maximum number of vertices that $conv(T(X))$ can have among all the possible permissible projective transformations $T$?In this paper, we investigate this and connected questions. After presenting several upper bounds, we study a closely related problem (via Gale transforms) concerning the number of minimal Radon partitions of a set of points. We then present some bounds for this number that enable us to partially answer a question due to Pach and Szegedy. We also discuss another related problem concerning the size of topes in arrangements of hyperplanes.

\section{Introduction} We consider the following question \begin{quote} Given a set of $n$ points in general position $X \subset \mathbb R^d$, what is the maximum number of $k$-faces that $conv(T(X))$ can have among all the possible {\em permissible projective transformations} $T$? \end{quote} More precisely, recall that a \emph{projective transformation} $T:\mathbb{R}^{d} \rightarrow \mathbb{R}^{d}$ is a function such that $T(x)=\frac{Ax+b}{\langle c,x \rangle + \delta}$, where $A$ is a linear transformation of $\mathbb{R}^{d}$, $b, c \in \mathbb{R}^{d}$ and $\delta \in \mathbb{R},$ is such that at least one of $c\neq 0$ or $\delta \neq 0.$ $T$ is said to be \emph{permissible} for a set $X\subset \mathbb{R}^{d}$ if and only if $\langle c,x \rangle + \delta \neq 0$ for all $x\in X$. Let $d\ge k\ge 0$ be integers and let $X$ be a set of points in $\mathbb R^d$, we define \begin{equation}\label{eq:def-h} h_X(d,k)= \max\limits_T \left\{f_k(conv(T(X)))\right\} \end{equation} maximum taken over all possible permissible projective transformations $T$ of $X$ and where $f_k(P)$ denotes the number of $k$-faces in the polytope $P$. We consider the function $H_k(n,d)$ defined as, $$H_k(n,d)=\min\limits_{X \subset \mathbb R^d, |X|=n}\left\{h_X(d,k)\right\}.$$ In this paper, we focus our attention to the values of $H_{0}(n,d),$ and $H_{d-1}(n,d)$. These are closely related and they are natural generalizations of the well-known McMullen's problem \cite{LARMAN72}: \emph{What is the largest integer $\nu(d)$ such that any set of $\nu(d)$ points in general position, $X \subset \mathbb R^{d},$ can de mapped by a permissible projective transformation onto the vertices of a convex polytope?} The best known bounds are \begin{equation}\label{eq:boundsmcmullen} 2d+1 \leq \nu(d) < 2d + \left\lceil \frac{d+1}{2} \right\rceil. \end{equation} The lower bound was given by Larman \cite{LARMAN72} while the upper bound was provided by Ramirez Alfonsin \cite{RAMIREZALFONSIN2001}. In the same spirit, the following function has also been investigated: \begin{quote} $\nu(d,k):=$ the largest integer such that any set of $\nu(d,k)$ points in general position in $\mathbb R^d$ can be mapped, by a permissible projective transformation onto the vertices of a $k$--neighbourly polytope. \end{quote} Garc\'{i}a-Col\'{i}n \cite{NGC2015} proved that, for each $2 \leq k \leq \left\lfloor \frac{d}{2}\right\rfloor$ \begin{equation}\label{eq:boundsmcmullengen} d + \left\lceil \frac{d}{k} \right\rceil +1 \leq \nu(d, k) < 2d-k+1. \end{equation} We now define the following closely related function, which will be useful to study $H_{0}(n,d).$ Let $t\ge 0$ be an integer, we define \begin{quote} $n(d,t):=$ the largest integer such that any set of $n(d,t)$ points in general position in $\mathbb R^d$ can be mapped, by a permissible projective transformation onto the vertices of a convex polytope with at most $t$ points in its interior. \end{quote} \begin{remark}\label{relationf_0andn(d,t)} The functions $\nu(d),$ $\nu(d,k)$ and $n(d,t)$ are closely related to $H_0(n(d,t),d),$ in particular: \begin{enumerate}[(a)] \item $n(d,0)=\nu(d)$ where $\nu(d)$ is the desired maximum value in McMullen's problem \item $H_0(n(d,t),d)=n(d,t)-t$. \end{enumerate} \end{remark} The function $n(d,t)$ will allows us to study $H_0(n,d)$ in a more general setting in terms of oriented matroids. Our main contribution is the following results: \begin{theorem}\label{boundsforf_o} Let $d,l$ and $n$ be integers. Then, {\scriptsize $$H_0(n,d)\left\{\begin{array}{ll} =2 & \text{ if } d=1,\ n\ge 2,\\ =5 & \text{ if } d=2, \ n\ge 5, \\ \le 7 & \text{ if } d=3, \ n\ge 7,\\ =n & \text{ if } d\ge 2, \ n\le 2d+1,\\ \le n-1& \text{ if } d\ge 4, \ n\ge 2d+\lceil\frac{d+1}{2}\rceil,\\ \le n-2 & \text{ if } d\ge 4, \ n\ge 2d+\lceil\frac{d+1}{2}\rceil+1,\\ \le n-(l+2) & \text{ if } d\ge 4, \ {2d+3+l(d-2) \leq n < 2d+3+(l+1)(d-2)}, \ l\ge 1.\\ \end{array}\right.$$} \end{theorem} \begin{theorem}\label{boundsforf_d-1} Let $d,l$ and $n$ be integers. Then, {\scriptsize $$H_{d-1}(n,d)\left\{\begin{array}{ll} =2 & \text{ if } d=1, \ n\ge 2,\\ =5 & \text{ if } d=2, \ n\ge 5, \\ \le 10 & \text{ if } d=3, \ n\ge 7,\\ \le f_{d-1}(C_d(n)) & \text{ if } d\ge 2, \ n\le 2d+1,\\ \le f_{d-1}(C_d(n-1))& \text{ if } d\ge 4, \ n\ge 2d+\lceil\frac{d+1}{2}\rceil,\\ \le f_{d-1}(C_d(n-2)) & \text{ if } d\ge 4, \ n\ge 2d+\lceil\frac{d+1}{2}\rceil+1,\\ \le f_{d-1}(C_d(n-l-2)) & \text{ if } d\ge 4, \ {2d+3+l(d-2) \leq n < 2d+3+(l+1)(d-2)}, \ l\ge 1.\\ \end{array}\right.$$} where $C_d(n)$ denotes the $d$--dimensional {\em cyclic polytope} with $n$ vertices. Moreover, $H_{d-1}(n,d)\ge n(d-1)-(d+1)(d-2)$ when $d\ge 2, n\le 2d+1$. \end{theorem} In the next section, we give some straightforward values and bounds. In Section \ref{sec:lawrence}, we discuss the aforementioned oriented matroid setting and recall some notions and results on the special class of {\em Lawrence oriented matroids} needed for the rest of the paper. In Section \ref{cotassupmatroides}, we prove our main results. In Section \ref{sec:arrangements}, we discuss the following interpretation of $H_{0}(n,d)$ in terms of arrangements of (pseudo)hyperplanes: \begin{quote} What is the maximum size that a {\em tope} in \underline{any} given simple arrangement of $n$ (pseudo) hyperplanes in $\mathbb{P}^d$ can have? \end{quote} The latter is closely related to the so-called Las Vergnas' conjecture and yields to further bounds for $H_0(n,d)$ (Proposition \ref{lowerboundvd=2}). In Section \ref{sec:anapplication}, we shall provide an answer to the following question due to Pach and Szegedy \cite{PACH2003}: \begin{question}\label{question:ps} Given $n$ points in general position in the plane, colored red and blue, maximize the number of multicolored 4-tuples with the property that the convex hull of its red elements and the convex hull of its blue elements have at least one point in common. In particular, show that when the maximum is attained, the number of red and blue elements are roughly the same. \end{question} We will first present a connection between $H_{d-1}(n,d)$ and the notion of {\em minimal Radon partitions} by using the relationship between {\em Gale transforms} and projective transformations (Theorem \ref{lem:radon}). The latter induces an upper bound for the number of minimal Radon partitions (i.e $(d+2)$--tuples) induced by a red/blue partition of a set of $n$ points in general position in $\mathbb{R}^d$. We then show that attaining the maximum does not depend on the number of red/blue points, but on the projective class of the configuration (Theorem \ref{thm:unbalanced}). We end by discussing some questions and open problems. \section{Some basic results}\label{sec:basic} The well-known Upper Bound Theorem \cite{MCMULLEN1971187} states that $f_{k-1}(P) \leq f_{k-1}(C_d(n))$ for all $1\leq k \leq d$ among all simplicial (convex) polytopes $P \subset \mathbb R^d$ with $n$ vertices where $C_d(n)$ is the $d$--dimensional {\em cyclic polytope} with $n$ vertices, that is, a polytope formed as a convex hull of $n$ distinct points in the moment curve $x(t):=(t,t^2,\dots ,t^d)$. A natural upper bound for $H_{d-1}(n,d)$ is thus given by \begin{equation}\label{eq:upperbound} H_{d-1}(n,d) \leq f_{d-1}(C_{d}(n)) \text{ for all } n\ge 1. \end{equation} Analogously, the Lower Bound Theorem \cite{BARNETTE1973, BILLERA1981} states that $f_k(P)\ge f_k(P_d(n))$ for all $1\leq k \le d-1$ among all simplicial (convex) polytopes $P \subset \mathbb R^d$ with $n$ vertices, where $P_d(n)$ is a $d$-dimensional {\em stacked polytope} with $n$ vertices, that is, a polytope formed from a simplex by repeatedly gluing another simplex onto one of its facets. We thus have \begin{equation}\label{eqn:lowerbound2} f_{d-1}(P_d(n))\le H_{d-1}(n,d) \end{equation} Moreover, we may deduce that \begin{equation}\label{upperboundcyclicp} f_{d-1}(P_d(H_0(n,d)))\leq H_{d-1}(n,d)\leq f_{d-1}(C_d(H_0(n,d))). \end{equation} \begin{proposition}\label{easybounds} Let $d,n\ge 1$ be integers. Then, $$H_0(n,d)\left\{\begin{array}{ll} =n & \text{ if } n\le 2d+1,\\ <n & \text{ if } n \ge 2d+\lceil\frac{d+1}{2}\rceil,\\ =n & \text{ if } d \text{ is even and } n\le d+3,\\ =n & \text{ if } d \text{ is odd and } n\le d+4.\\ \end{array}\right.$$ \end{proposition} \begin{proof} Let $n \leq 2d+1$. By the lower bound of $\nu(d)$ given in \eqref{eq:boundsmcmullen}, it follows that any set of points of cardinality $n$ can be mapped to the vertices of a convex polytope by a permissible projective transformations, therefore $H_{0}(n,d)=n$. If $n \ge 2d + \lceil \frac{d+1}{2} \rceil$ then by the upper bound of $\nu(d)$ given in \eqref{eq:boundsmcmullen}, there exist a set of $n$ points such that cannot be mapped to the vertices of a convex polytope by any permissible projective transformation, therefore $H_0(n,d)\le n-1$. If $d\ge 2$ is even (resp. is odd) and $n \leq d+3$ (resp. $n\leq d+4$) then, by taking $k=\lfloor \frac{d}{2}\rfloor$, the lower bound of \eqref{eq:boundsmcmullengen}, all polytopes with less than $d+3$ (resp. $d+4$) vertices are in the projective class of a neighbourly polytope (i.e. $k$--neighbourly for $k=\lfloor \frac{d}{2}\rfloor$). The last two equalities follow by using the same argument as above. \end{proof} The following result is an easy consequence of Proposition \ref{easybounds} and Equation \eqref{upperboundcyclicp}: \begin{proposition}\label{easybounds1} Let $d,n\ge 1$ be integers. Then, $$H_{d-1}(n,d)\left\{\begin{array}{ll} \le f_{d-1}(C_{d}(n)) & \text{ if } n\le 2d+1,\\ \le f_{d-1}(C_{d}(n-1)) & \text{ if } n \ge 2d+\lceil\frac{d+1}{2}\rceil,\\ =f_{d-1}(C_d(n)) & \text{ if } d \text{ is even and } n\le d+3,\\ =f_{d-1}(C_d(n)).& \text{ if } d \text{ is odd and } n\le d+4.\\ \end{array}\right.$$ Moreover, $f_{d-1}(P_d(n))\le H_{d-1}(n,d)$ when $n\le 2d+1$.\\ \end{proposition} \section{Oriented matroid setting}\label{sec:lawrence} Recall that an oriented matroid $M$ is \emph{acyclic} if it does not contain positive circuits (otherwise, $M$ is called \emph{cyclic}). We say that an element $e$ of an oriented acyclic matroid is \emph{interior} if there exists a signed circuit $C=(C^+,C^-)$ with $C^-=\{e\}$. Cordovil and Da Silva \cite{CORDOVIL1985157} proved that a permissible projective transformation on $n$ points in $\mathbb R^d$ corresponds to an acyclic reorientation of its corresponding oriented matroid $M$ of rank $r=d+1$ and conversely. A natural generalization of $n(d,t)$ in terms of oriented matroids is given by the following function. \begin{quote} $\bar n(d,t):=$ the largest integer such that for any uniform matroid $M$ of rank $d+1$ with $\bar n(d,t)$ elements there is an acyclic reorientation of $M$ with at most $t$ interior elements. \end{quote} We notice that $n(d,t)=\bar n(d,t)$ in the case when $M$ is {\em realizable}. In this section we shall provide examples of uniform oriented matroids such that for any acyclic reorientation there are at least $t+1$ interior elements, thus providing upper bounds on $\bar{n}(d,t)$. To this end, we briefly outline some facts of the special class, {\em Lawrence oriented matroids,} (see \cite{OM1999,RAMIREZALFONSIN2001} for further details and proofs). A {\em Lawrence oriented matroid} (LOM) ${M}$ of rank $r$ on the totally ordered set $E=\{1,\ldots,n\}$, $r\le n$, is a uniform oriented matroid obtained as the union of $r$ uniform oriented matroids ${M}_1,\ldots,{M}_r$ of rank $1$ on $(E,<)$. LOM can also be defined via the signature of their bases, that is, via their chirotope $\chi$. Indeed, the chirotope $\chi$ corresponds to some LOM, ${M}_A,$ if and only if there exists a matrix $A=(a_{i,j})$, $1\le i\le r$, $1\le j\le n$ with entries from $\{+1,-1\}$ (where the $i$th row corresponds to the chirotope of the oriented matroid $\mathcal{M}_i$) such that \begin{equation}\label{eq3} \chi(B)=\prod_{i=1}^{r} a_{i,j_i} \end{equation} where $B$ is an ordered $r$-tuple, $j_1\le\ldots\le j_r,$ of elements of $E$. \begin{remark} We highlight some useful Properties of LOM. \begin{enumerate}[(a)] \item Acyclic LOM are realizable as configurations of points (since they are unions of realizable oriented matroids). \item LOM are closed under minors and duality. \item The LOM corresponding to the reorientation of an element $c \in E,$ $_{\bar c}M_A$ is obtained by reversing the sign of all the coefficients of a column $c$ in $A$. \end{enumerate} \end{remark} From now, we will denote $A=A_{r,n}$ as a matrix with entries $a_{i,j}\in \{+1,-1\}$, $1\le i\le r$, $1\le j\le n$. Some of the following definitions and lemmas, which highlight the properties of $A$, and facilitate the study of these type of matroids, where introduced and proved in \cite{RAMIREZALFONSIN2001}: A \emph{Top Travel}, denoted as $TT,$ in $A$ is a subset of the entries of \(A$, \linebreak $\{ [a_{1, 1}, a_{1, 2},\dots ,a_{1, j_{1}}] , [a_{2, j_{1}}, a_{2, j_{1}+1},\dots , a_{2, j_{2}}] , \dots , [a_{s, j_{s-1}}, a_{s, j_{s-1}+1},\dots , a_{s, j_{s}}] \}$, with the following characteristics: \begin{enumerate} \item$ a_{i, j_{i-1}} \times a_{i, j}= 1, \quad \forall \quad\ j_{i-1} \leq j < j_{i};$ \item$ a_{i, j_{i-1}} \times a_{i, j_{i}}= -1;$ and \item either \begin{enumerate} \item $1\leq s < r$; then $j_{s} = n$ or \item $s=r$ and $ j_{s} \leq n.$ \end{enumerate} \end{enumerate} A \emph{Bottom Travel}, denoted as $BT$, in $A$ is a subset of the entries of \(A$, \linebreak $\{ [a_{r, n}, a_{r, n-1},\dots ,a_{r, j_{r}}] , [a_{r-1, j_{r}}, a_{r-1, j_{r}-1},\dots , a_{r-1, j_{r-1}}] , \dots , [a_{s, j_{s-1}}, a_{s, j_{s-1}+1},\dots , a_{s, j_{s}}] \}$, with the following characteristics: \begin{enumerate} \item $ a_{i, j_{i+1}} \times a_{i, j}= 1, \quad \forall \quad\ j_{i} < j \leq j_{i+1};$ \item $ a_{i, j_{i+1}} \times a_{i, j_{i}}= -1;$ and \item either \begin{enumerate} \item $1 < s \leq r$; then $j_{s} = 1$ or \item $s=1$ and $ 1 \leq j_{s}.$ \end{enumerate} \end{enumerate} Note that every matrix $A$ has exclusively one $TT$ and one $BT$ and they carry surprising information about $M_{A}$. In \cite{RAMIREZALFONSIN2001} was proved that if $A = A_{r,n}$ then the following conditions are equivalent: \begin{enumerate} \item$M_{A}$ is \emph{cyclic}; \item $TT$ ends at $ a_{r,s}$ for some $ 1 \leq s < n$; and \item $BT$ ends at $ a_{1, s'}$ for some $ 1< s' \leq n$. \end{enumerate} We say that $TT$ and $BT$ are \emph{parallel} at column $k$ in $A$ if $a_{i,k-1},a_{i,k},a_{i,k+1}\in TT$ and either $a_{i,k-1},a_{i,k},a_{i,k+1}\in BT$ or $a_{i+1,k-1},a_{i+1,k},a_{i+1,k+1}\in BT$, with $2\le k\le n-1$, $1\le i\le r$. We have \cite{RAMIREZALFONSIN2001} that for $A = A_{r,n}$, then $k$ is an interior element of $M_A$ if and only if \begin{itemize} \item[(a)] $BT=(a_{r,n},\ldots,a_{1,2},a_{1,1})$ for $k=1$, \item[(b)] $TT=(a_{1,1},\ldots,a_{r,n-1},a_{r,n})$ for $k=n$, \item[(c)] $TT$ and $BT$ are parallel at $k$ for $2\le k\le n-1$. \end{itemize} The above imply that we can identify acyclic reorientations and interior elements of $M_{A}$ by studying the behaviour of the $TT$ and $BT$ in the re-orientations of $A$. \begin{example} Let $M_A$ be the LOM associated to the matrix $A$ given in Figure \ref{ejemplo}. We notice that $M_A$ is acyclic and that $4$, $5$ and $6$ are interior elements. \begin{figure}[htb] \begin{center} \includegraphics[width=.4\textwidth]{ejemplo.pdf} \caption{Top and Bottom travels in matrix $A$.} \label{ejemplo} \end{center} \end{figure} \end{example} Interestingly, all possible re-orientations of the matroid can be identified with yet another simple object; A \emph{Plain Travel} in $A$, denoted as PT, is a subset of the entries of \(A$ which satisfies: \[PT=\{ [a_{1, 1}, a_{1, 2},\dots ,a_{1, j_{1}}] , [a_{2, j_{1}}, a_{2, j_{1}+1},\dots , a_{2, j_{2}}] , \dots , [a_{s, j_{s-1}}, a_{s, j_{s-1}+1},\dots , a_{s, j_{s}}] \} \] with \( 2\leq j_{i-1} \leq j_{i} \leq n \quad \forall \quad 1 \leq i \leq r ,\;\; 1< s \leq r$ and $j_{s} = n$. We have {\cite{RAMIREZALFONSIN2001}} that there is a bijection between the set of all plain travels of $A$ and the set of all acyclic reorientations of $M_{A}$. The correspondence between an acyclic reorientation of $M_{A}$ and a $PT$ of $A$ is given by the subset of columns of $A$ (i.e. a subset of elements of $E$) that have to be reoriented in order to transform the top travel of $A,$ $TT,$ into $PT$. Finally, we recall another useful object that can be constructed from the matrix $A,$ the \emph{chessboard} $B[A]$ which is a black and white board of size $(r-1)*(n-1)$, such that the square $s(i,j)$ has its upper left hand corner at the intersection of row $i$ and column $j$. A square $s(i,j)$, with $1 \leq i \leq r-1$ and $ 1 \leq j \leq n-1$, will be said to be {\em black} if the product of the entries $a_{i,j}, a_{i,j+1}, a_{i+1,j}, a_{i+1,j+1}$ is $-1$, and {\em white} otherwise. \begin{proposition} \cite{RAMIREZALFONSIN2001} \label{prop:chessboard} The following properties establish a link between chessboards, reorientations of a LOM, and travels: \begin{enumerate}[(a)] \item $B[A]$ is invariant under reorientations of $M_A$. \item If in a pair of consecutive columns there is one black square between the top travel $TT$ and the bottom travel $BT$, they follow symmetrically opposite paths through the entries of the matrix; in other words, if $TT$ makes a single horizontal movement from $a_{i,j}$ to $a_{i,j+1}$ and continues its movement forward in the same row (i.e $a_{i,j} = a_{i,j+1}$) , then $BT$ goes from $a_{i+h,j+1}$ to $a_{i+h,j}$ and moves vertically to $a_{i+h-1, j}$ (i.e. $a_{i+h,j+1}\neq a_{i+h,j}$), with $h\geq{1}$, and vice versa. \end{enumerate} \end{proposition} \section{Upper bounds for $\bar{n}(d,t)$}\label{cotassupmatroides} \subsection{Small dimensions} We first show that $\bar{n}(d,t)\le 2d+1+t$ for $d=2,3$ and every $t\ge0$. The following remark will be very useful throughout this section. \begin{remark} \label{dim1} Any acyclic reorientation of a rank $2$ oriented matroid on $n$ elements has $n-2$ interior elements. \end{remark} Given a matrix $A=A_{n,r}$, let $A^+_{i,j}$ be the submatrix of $A$ that results after removing rows $i+1,\ldots, r$ and columns $j+1,\ldots, n$. Similarly, let $A^-_{i,j}$ be the submatrix of $A$ resulting after the removal of rows $1,\ldots, i-1$ and columns $1,\ldots,j-1,$. \begin{theorem}\label{dimension2} $\bar{n}(2,t)<t+6$ for every $t\ge 0$. \end{theorem}\begin{proof} Let $A=A_{3,t+6}$ be such that the corresponding chessboard $B[A]$ has exactly one black square for each column and let $PT$ be any plane travel in $A$. We shall prove that the corresponding $\mathcal{A}$ in which $PT$ is the Top Travel has at least $t+1$ interior elements. Let $j$ be the smallest number such that column $j$ is not an interior element in $\mathcal{A}$ and there are not vertical movements in column $j$ of $PT$ neither of $BT$. If $j$ does not exists, as $PT$ and $BT$ can make at most $2$ vertical movements each, then $\mathcal{A}$ would have at least $t+2$ interior elements. Hence, we may suppose that $j$ exists. By the rules of Property \ref{prop:chessboard}, $PT$ and $BT$ make a vertical movement in column $j+1$ and $j-1$, respectively. Notice by the definition of $j$ that $PT$ and $BT$ arrives in column $j$ at row $1$ and $3$, respectively, otherwise $j$ would be an interior element. Then, each interior element of $A^+_{2,j-1}$ and $A^-_{2,t+6-j}$ is an interior element of $\mathcal{A}$. Therefore, $A^+_{2,j-1}$ has $j-3$ interior elements and $A^-_{2,t+6-j}$ has $t+6-j-2$ interior elements by Remark \ref{dim1}, concluding the proof. \end{proof} Given a matrix $A=A_{r,n}$, we say that a chess board $B[A]$ has the \emph{sequence} $(x_1,x_2,\ldots,x_{r-1})$ if the square $s(i,j)$ is black if and only if $\sum\limits_{k=0}^{i-1}x_{k}+1\le j\le \sum\limits_{k=0}^{i}x_{k}$ with $1 \leq i \leq r-1$ and $1 \leq j \leq n-1$, where we define $x_0=0$. \begin{example} Figure \ref{figurehypothesis} illustrates a chessboard with sequence $(2,3,2,3)$. \end{example} \begin{theorem}\label{dimension3} $\bar{n}(3,t)<t+8$ for $t\ge 0$. \end{theorem}\begin{proof} Let $A=A_{4,t+8}$ be such that the corresponding chessboard $B[A]$ has a sequence $(2,t+3,2)$ and let $PT$ be any plane travel in $A$. We prove that the corresponding $\mathcal{A}$ in which $PT$ is the Top Travel has at least $t+1$ interior elements. Let $j$ be the smallest number such that column $j$ is not an interior element in $\mathcal{A}$ and there are not vertical movements in column $j$ of $PT$ neither of $BT$. If $j$ does not exists, as $PT$ and $BT$ can make at most $3$ vertical movements each, then $\mathcal{A}$ would have at least $t+2$ interior elements. Hence, we may suppose that $j$ exists. By the rules of Property \ref{prop:chessboard}, $PT$ and $BT$ make a vertical movement in column $j+1$ and $j-1$, respectively. By the definition of $j$, $TT$ arrives in column $j$ at row $1$ or $BT$ arrives in column $j$ at row $4$, otherwise $j$ would be an interior element. Suppose without loss of generality that $TT$ arrives in column $j$ at row $1$. Then, $BT$ arrives in column $j$ at row $2$ or $3$ (see Figure \ref{dim3}). \begin{figure}[htb] \begin{center} \includegraphics[width=.2\textwidth]{fig_dim3caso1.pdf} \caption{$BT$ in blue and $TT$ in red.} \label{dim3}\end{center} \end{figure} Notice that each interior element of $A^+_{3,j-1}$ and $A^-_{2,t+8-j}$ is an interior element of $\mathcal{A}$. If $BT$ arrives in column $j$ at row $3$, by the rules of Property \ref{prop:chessboard}, $A^+_{3,j-1}$ has $j-4$ interior elements and $A^-_{2,t+8-j}$ has $t+5-j$ interior elements, concluding the proof in this case. If $BT$ arrives in column $j$ at row $2$, by the rules of Property \ref{prop:chessboard} and by Remark \ref{dim1}, one can check that $A^+_{3,j-1}$ has $j-3$ interior elements and $A^-_{2,t+8-j}$ has at least $t+4-j$ interior elements, concluding the proof. \end{proof} \subsection{High dimensions $d$} In what follows, we will consider different matrices $A=A_{r,h(r)}$ where $h(r)$ is a strictly increasing function. If $B[A]$ has the sequence $(x_1,x_2,\ldots,x_{r-1})$, we will consider functions, $h(r),$ where $h(1)=1$, $\sum\limits_{k=1}^{m-1}x_{k}+1\le h(m)\le \sum\limits_{k=1}^{m}x_{k}+1$ if $2\le m\le r-1$ and $\sum\limits_{k=1}^{r-1}x_{k}+1\le h(r)$. For every $1\le m\le r-1$, we will say that the element $a_{m,h(m)}$ is the \emph{$m$--th corner} of $A$. \begin{example} Figure \ref{figurehypothesis} illustrates a chessboard with $h(r)=2(r-1)+\lceil\frac{r}{2}\rceil$. \end{example} The following lemmas will be very useful. \begin{lemma}\label{lemmageneral1} Let $A=A_{r,h(r)}$ be a matrix with $r\ge 3$ such that $B[A]$ has the sequence $(x_1,x_2,\ldots,x_{r-1})$, with $x_i\ge1$, $1\le i\le r-1$. Suppose that $TT$ always passes strictly above all corners after the $1$--st corner, then \begin{itemize} \item [(i)] $a_{i,h(2)}\in BT$ for some $i\le \max\{2,2r-h(r-1)+h(2)-3\}$, \item [(ii)] $a_{i,h(2)}\in BT$ for some $i\le \max\{2,2r-h(r)+h(2)-1\}$ if $TT$ and $BT$ do not share steps from columns $h(2)$ to $h(r)$. \end{itemize} \end{lemma} \begin{proof} We will proceed by induction on $r\ge3$. Using the rules of Property \ref{prop:chessboard}, one can check that the lemma holds for $r=3$. Suppose that the result holds for $r-1$ and we show it for $r\ge4$. If $BT$ arrives at the $m$--th corner for $3\le m\le r-1$, the result follows by the induction hypothesis. Moreover, if $BT$ arrives at the $2$--th corner, the result follows. Similarly, the result follows if $BT$ arrives at $a_{i,h(m)}$ for $2\le m\le r-1$ and $i\le m$. Then, we may suppose that $BT$ always passes strictly below the $m$--th corner for every $2\le m\le r-1$. Notice that $TT$ makes exactly $h(r)-h(2)$ horizontal movements and at most $r-1$ vertical movements, from right to left, to arrive at $a_{1,h(2)}$. We know, by the rules of construction of $TT$ that for each vertical movement we must also count one horizontal movement. So, $TT$ makes at least $h(r)-h(2)-(r-1)$ single horizontal movements, from right to left, until $a_{1,h(2)}$ is attained. On the other hand, as $TT$ always passes strictly above all corners after the $1$--st corner and $BT$ always passes strictly below the $m$--th corner for every $2\le m\le r-1$, then $TT$ and $BT$ do not share steps from columns $h(2)$ to $h(r-1)$. However, $TT$ and $BT$ could share at most $h(r)-h(r-1)-2$ steps from columns $h(r-1)+1$ to $h(r)$ (see Figure \ref{figurehypothesis}). \begin{figure}[htb] \begin{center} \includegraphics[width=.35\textwidth]{fig_hypothesislemma2.pdf} \caption{The chessboard $B[A]$ with sequence $(2,3,2,3)$ and $h(r)=2(r-1)+\lceil\frac{r}{2}\rceil$. The points represent the corners of $A=A_{5,h(5)}$. We observe that $TT$ and $BT$ share the final step.} \label{figurehypothesis} \end{center} \end{figure} Therefore, by the rules of Property \ref{prop:chessboard}, for each single horizontal movement that $TT$ does not share with $BT$, $BT$ makes a vertical movement. So, $BT$ makes at least $h(r)-h(2)-(r-1)-(h(r)-h(r-1)-2)=h(r-1)-h(2)-r+3$ vertical movements, from right to left, until column $h(2)$ is attained. Hence, $BT$ arrives at $a_{i,h(2)}$ for some $i\le \max\{2,r-(h(r-1)-h(2)-r+3)\}$ concluding the first part of the proof. If $TT$ and $BT$ do not share steps from columns $h(2)$ to $h(r)$, then $BT$ makes at least $h(r)-h(2)-(r-1)$ vertical movements, from right to left, until column $h(2)$ is attained, concluding that $BT$ arrives at $a_{i,h(2)}$ for some $i\le \max\{2,r-(h(r)-h(2)-(r-1))\}$. \end{proof} \begin{lemma}\label{lemmageneral2}Let $A=A_{r,h(r)}$ be a matrix and suppose that $TT$ always passes strictly above all corners after the $1$--st corner. Then the following holds: \begin{itemize} \item [(i)] If $B[A]$ has a sequence $(x_1,x_2,\ldots,x_{r-1})$, $x_i\ge2$ for odd $i$, $x_j\ge3$ for even $j$ and $h(m)=\sum\limits_{k=0}^{m-1}x_{k}+1$ for every $1\le m\le r$, then $a_{1,1},a_{1,2}\in BT$ when $r\ge4$. When $r=3$, $a_{1,1}, a_{1,2}\in BT$, or column $h(r)$ is an interior element and $a_{2,1},a_{1,1}\in BT$. \item [(ii)] If $B[A]$ has a sequence $(x_1,x_2,\ldots,x_{r-1})$, $x_i\ge3$ for odd $i$, $x_j\ge2$ for even $j$ and $h(m)=\sum\limits_{k=0}^{m-1}x_{k}+1$ for every $1\le m\le r$, then $a_{1,1},a_{1,2}\in BT$ when $r\ge3$. \item [(iii)] If $B[A]$ has a sequence $(2,t+3,2,t+1,t+1,\ldots,t+1)$ for some $t\ge2$, $h(2)=t+3$, $h(3)=t+6$ and $h(m)=(t+1)(m-3)+7$ for $4\le m\le r$ for every $4\le m\le r$, then $a_{i,t+3}\in BT$ for some $i\le 2$ when $r\ge4$ (see Figure \ref{cotageneral2}). \item [(iv)] If $B[A]$ has a sequence $(t+1,\ldots,t+1,2,t+3,2)$ for some $t\ge2$, $h(2)=t+3$, $h(r)=(t+1)(r-3)+7$ and $TT$ and $BT$ do not share steps from columns $h(2)$ to $h(r)$, then $a_{i,t+3}\in BT$ for some $i\le 2$ when $r\ge5$. \item [(v)] If $B[A]$ has a sequence $(2,4,2,3,2,3,\ldots)$ and $h(m)=2(m-1)+\lceil\frac{m}{2}\rceil+1$ for every $2\le m\le r$, then $a_{i,4}\in BT$ for some $i\le 2$ when $r\ge4$. \item [(vi)] If $r\ge6$ is even, $B[A]$ has a sequence $(2,3,2,3,\ldots,2,4,2)$, $h(2)=4$, $h(r)=2(r-1)+\lceil\frac{r}{2}\rceil+1$ and $TT$ and $BT$ do not share steps from columns $h(2)$ to $h(r)$, then $a_{i,4}\in BT$ for some $i\le 2$. \end{itemize} \end{lemma} \begin{proof} We prove (i) and (ii) for $B[A]$ with sequences $(2,3,2,3\ldots)$ and $(3,2,3,2\ldots)$, respectively, since the general case holds as a consequence. (i) By the sequence of $B[A]$, we observe that $h(m)=2(m-1)+\lceil\frac{m}{2}\rceil$ for $1\le m\le r$. Then, as $2r-h(r-1)+h(2)-3=4-\lceil\frac{r-1}{2}\rceil\le 2$ when $r\ge4$, we obtain by Lemma \ref{lemmageneral1} (i) that $a_{i,h(2)}\in BT$ for some $i\le2$. Since $a_{1,i}\in TT$ for $i\le4$ and $h(2)=3$, we conclude by the rules of Property \ref{prop:chessboard} that $a_{1,1},a_{1,2}\in BT$. If $r=3$, one can check that $a_{1,1}, a_{1,2}\in BT$, or column $6$ is an interior element and $a_{1,1}, a_{2,1}\in BT$ (see Figure \ref{r3}). \begin{figure}[htb] \begin{center} \includegraphics[width=.2\textwidth]{fig_remarkr3.pdf} \caption{Case $r=3$ when column $6$ is an interior element and $a_{1,1}, a_{2,1}\in BT$. The points represent the corners.} \label{r3} \end{center} \end{figure} (ii) By the sequence of $B[A]$, we observe that $h(m)=2(m-1)+\lceil\frac{m+1}{2}\rceil$ for $1\le m\le r$. Then, as $2r-h(r-1)+h(2)-3=5-\lceil\frac{r}{2}\rceil\le 3$, we obtain by Lemma \ref{lemmageneral1} (i) that $a_{i,h(2)}\in BT$ for some $i\le3$. Since $a_{1,i}\in TT$ for $i\le5$ and $h(2)=4$, we conclude by the rules of Property \ref{prop:chessboard} that $a_{1,1},a_{1,2}\in BT$. (iii) As $2r-h(r-1)+h(2)-3=2$ when $r=4$ and $2r-h(r-1)+h(2)-3=2r-(t+1)(r-4)-7+t\le 7-r\le2$ when $r\ge5$ and $t\ge2$, we obtain by Lemma \ref{lemmageneral1} (i) that $a_{i,h(2)}\in BT$ for some $i\le 2$. (iv) As $2r-h(r)+h(2)-1=2r-(t+1)(r-3)+t-5\le 6-r\le1$ when $r\ge5$ and $t\ge2$, we obtain by Lemma \ref{lemmageneral1} (ii) that $a_{i,h(2)}\in BT$ for some $i\le 2$. (v) As $2r-h(r-1)+h(2)-3=4-\lceil\frac{r-1}{2}\rceil\le2$ when $r\ge4$, we obtain by Lemma \ref{lemmageneral1} (i) that $a_{i,h(2)}\in BT$ for some $i\le 2$. (vi) As $2r-h(r)+h(2)-1=2r-2(r-1)-\frac{r}{2}+2=4-\frac{r}{2}\le1$ when $r\ge6$ is even, we obtain by Lemma \ref{lemmageneral1} (ii) that $a_{i,h(2)}\in BT$ for some $i\le 2$. \end{proof} From now on, denote as $A^+_m$ and $A^-_m$ the matrices $A^+_{m,h(m)}$ and $A^-_{m,h(m)}$, respectively (see Figure \ref{cotageneral2}). We are now ready to tackle the problem for any $d\ge4$ and $t\ge2$. \begin{theorem}\label{cotageneral} $\bar{n}(d,t)<2d+(t-1)(d-2)+3$ for $d\ge4$ and $t\ge 2$. \end{theorem} \begin{proof} Let $A=A_{r,h(r)}$ be a matrix where $h(r)$ is defined as $h(2)=t+3$, $h(3)=t+6$, $h(m)=(t+1)(m-3)+7$ for $4\le m\le r$ and $B[A]$ with sequence $(2,t+3,2,t+1,t+1,\ldots,t+1)$ for $t\ge2$ (see Figure \ref{cotageneral2}). \begin{figure}[htb] \begin{center} \includegraphics[width=.6\textwidth]{fig_cotageneral2.pdf} \caption{A matrix $A=A_{6,h(6)}=A_{6,19}$ and the submatrices $A^+_4$ and $A^-_4$. The chessboard $B[A]$ has sequence $(2,t+3,2,t+1,t+1)$ for $t=3$. The points represent the corners of $A$ associated to the function $h(r)$ of Theorem \ref{cotageneral}.} \label{cotageneral2}\end{center} \end{figure} We shall show by induction on $r$ that for every $r\ge2$ and for any plain travel $PT$ in $A$, the corresponding $\mathcal{A}$ in which $PT$ is the Top Travel has at least $t+1$ interior elements. In particular, as $h(r)=(t+1)(d-2)+7=2d+(t-1)(d-2)+3$ for $r\ge5$, we will prove the theorem for $d\ge4$ and $t\ge2$. We observe that $\mathcal{A}$ has $t+1$ interior elements when $r=2$ since $\mathcal{A}$ is a $2\times (t+3)$ matrix (Remark \ref{dim1}). For $r=3$ and $4$, the result follows by Theorems \ref{dimension2} and \ref{dimension3}, respectively, since the chessboards considered in these theorems coincide with $B[A]$. Thus, assume that the theorem holds for $r-1$ and we show it for $r\ge 5$. Suppose that the $m$--st corner is the last corner that $PT$ meets in $A$, for some $1\le m\le r-1$. If $PT$ always passes strictly below the $i$--st corner for $i>m$, then there would be at least $t+2$ interior elements in $\mathcal{A}$ (from columns $h(r-1)$ to $h(r)$). Hence, we may suppose that $PT$ always passes strictly above the $i$--st corner for $i>m$. We have the following cases. \emph{Case $m\le r-3.$} First suppose that $m=1$. Then $a_{i,t+3}\in BT$ for some $i\le 2$ by Lemma \ref{lemmageneral2} (iii), concluding by the rules of Property \ref{prop:chessboard} that there are at least $t+1$ interior elements in $\mathcal{A}$. Now suppose that $2\le m\le r-3.$ As $\mathcal{A}^-_m$ has at least $4$ rows, applying Lemma \ref{lemmageneral2} (i) (when $m=3$) and Lemma \ref{lemmageneral2} (ii) (when $m\neq 3$) on submatrix $\mathcal{A}^-_m$, we obtain that $a_{m,h(m)},a_{m,h(m)+1}\in BT$. Thus, the theorem holds by induction hypothesis on $\mathcal{A}^+_m$ since $BT$ restricted in $\mathcal{A}^+_m$ is also the Bottom travel of $\mathcal{A}^+_m$ and each interior element of $\mathcal{A}^+_m$ is an interior element of $\mathcal{A}$. \emph{Case $m=r-2.$} As $\mathcal{A}^-_m$ has $3$ rows, $a_{m,h(m)},a_{m,h(m)+1}\in BT$, or column $h(r)$ is an interior element and $a_{m+1,h(m)}, a_{m,h(m)}\in BT$ by Lemma \ref{lemmageneral2} (i). If $a_{m,h(m)},a_{m,h(m)+1}\in BT$, the theorem holds by induction hypothesis on $\mathcal{A}^+_m$. If column $h(r)$ is an interior element and $a_{m+1,h(m)}, a_{m,h(m)}\in BT$, notice that each interior element of $\mathcal{A}^+_m$ is an interior element of $\mathcal{A}$, except for column $h(m)$. Thus, $\mathcal{A}$ has at least $t$ interior elements from columns $1$ to $h(m)$ by induction hypothesis on $\mathcal{A}^+_m$ and one interior element in column $h(r)$. \emph{Case $m=r-1.$} First suppose that $BT$ arrives at the $k$-th corner for some $2\le k\le r-1$. As each interior element of $\mathcal{A}^+_k$ and $\mathcal{A}^-_m$ is an interior element of $\mathcal{A}$, except for (maybe) columns $h(k)$ and $h(m)$, $\mathcal{A}$ has at least $t$ interior elements from columns $1$ to $h(k)$ by induction hypothesis on $\mathcal{A}^+_k$ and at least $t-1\ge 1$ interior elements from columns $h(m)$ to $h(r)$ by Remark \ref{dim1} (since $\mathcal{A}^-_m$ is a $2\times (t+2)$ matrix), concluding the proof in this case. Similarly, the proof holds if $BT$ arrives at $a_{i,h(k)}$ for $2\le k\le r-2$ and $i\le k$. Now suppose that $BT$ passes always below the $i$--st corner, for every $i\ge2$. In particular, as $BT$ does not arrives at the $m$-th corner, every interior element of $\mathcal{A}^-_m$ is an interior element of $\mathcal{A}$, concluding by Remark \ref{dim1} that $\mathcal{A}$ has $t$ interior elements from columns $h(m)+1$ to $h(r)$. So, we may suppose that $TT$ and $BT$ do not share steps from columns $1$ to $h(2)$, otherwise the theorem holds. Also, if $a_{m,h(m)-1}\in PT$, then column $h(m)$ is an interior element and the theorem holds. So, we may suppose that $a_{m,h(m)-1}\not\in PT$. Hence, $a_{r,h(m)-2}\in BT$ by the rules of Property \ref{prop:chessboard}. Let $\mathcal{A}^{'}$ be the matrix obtained by turning the matrix $\mathcal{A}$ upside down. We observe that $BT$ and $PT$ are the Top and Bottom Travels of $\mathcal{A}^{'}$, respectively. Let define the $i$--st corners of $\mathcal{A}^{'}$ as $a_{r-i+1,h(r-i+1)}$ for $i\neq 2$ and define the $2$--st corner of $\mathcal{A}^{'}$ as $a_{m,h(m)-1}$. Notice that $BT$ always passes strictly above all corners of $\mathcal{A}^{'}$ after the $1$--st corner of $\mathcal{A}^{'}$. Moreover, $B[A']$ has the same sequence and the same $2$--st corner as that considered in Lemma \ref{lemmageneral2} (iv). Hence, as $TT$ and $BT$ do not share steps from columns $h(1)$ to $h(m)-1$, we know by Lemma \ref{lemmageneral2} (iv) that $a_{i,h(m)-1}\in PT$ for some $i\ge r-1$, but this is a contradiction since we had assumed that $a_{m,h(m)-1}\not\in PT$ and clearly $a_{r,h(m)-1}\not\in PT$, concluding the proof. \end{proof} Figure \ref{contraexample1} (a) shows that the chessboard considered in Theorem \ref{cotageneral} can not be used to prove $\bar{n}(d,1)<2d+(t-1)(d-2)+3$ for $d=4$ and $t=1$. In fact, this example can be generalized in order to show that this chessboard can not be used to prove $\bar{n}(d,t)<2d+3$ for $d\ge4$. \begin{theorem}\label{tpequeno} $\bar{n}(d,1)<2d+\lceil\frac{d+1}{2}\rceil+1$ for $d\ge4$. \end{theorem} \begin{proof} Let $A=A_{r,h(r)}$ be a matrix where $h(r)$ is defined as $h(m)=2(m-1)+\lceil\frac{m}{2}\rceil+1$ for every $2\le m\le r$ and $B[A]$ has sequence $(2,4,2,3,2,3,\ldots)$. We shall show by induction on $r$ that for every $r\ge2$ and for any plain travel $PT$ in $A$, the corresponding $\mathcal{A}$ in which $PT$ is the Top Travel has at least $2$ interior elements. In particular, as $h(r)=2d+\lceil\frac{d+1}{2}\rceil+1$ for $r\ge5$, we will prove the theorem for $d\ge4$. We observe that $\mathcal{A}$ has at least $2$ interior elements when $r=2$ since $\mathcal{A}$ is a $2\times4$ matrix (Remark \ref{dim1}). For $r=3$ and $4$, the result follows by Theorems \ref{dimension2} and \ref{dimension3}, respectively (applying them for $t=1$), since the chessboards considered in these theorems coincide with $B[A]$. Thus, assume the theorem holds for $r-1$ and we show it for $r\ge 5$. Suppose that the $m$--st corner is the last corner that $PT$ meets in $A$, for some $1\le m\le r-1$. If $PT$ always passes strictly below the $i$--st corner for $i>m$, then there would be at least $3$ interior elements in $\mathcal{A}$ (from columns $h(r-1)$ to $h(r)$). Hence, we may suppose that $PT$ always passes strictly above the $i$--st corner for $i>m$. If $m=1$, then $a_{i,4}\in BT$ for some $i\le 2$ by Lemma \ref{lemmageneral2} (v), concluding by the rules of Property \ref{prop:chessboard} that there are at least $2$ interior elements in $\mathcal{A}$. We omit the proof of the cases $2\le m\le r-2$, since they are analogous to those in the proof of Theorem \ref{cotageneral}. Now, consider the case $m=r-1$. If $r$ is odd, $B[A]$ has a sequence $(2,4,2,3,2,3,\ldots,2,3)$ and one can verify that there are at least $2$ interior elements in $\mathcal{A}$ or $a_{m,h(m)},a_{m,h(m)+1} \in BT$. In both cases the theorem holds. Now suppose that $r$ is even and then $B[A]$ has a sequence $(2,4,2,3,2,3,\ldots,3,2)$. If $a_{m,h(m)},a_{m,h(m)+1} \in BT$, the result follows by induction hypothesis on $\mathcal{A}^+_m$. Then, suppose from now that the above does not hold. Hence, each interior element of $\mathcal{A}^-_m$ is an interior element of $\mathcal{A}$ concluding by Remark \ref{dim1} that $\mathcal{A}$ has one interior element from columns $h(m)+1$ to $h(r)$ (since $\mathcal{A}^-_m$ is a $2\times 3$ matrix). First suppose that $BT$ arrives at the $k$-th corner for some $2\le k\le r-1$. As each interior element of $\mathcal{A}^+_k$ is an interior element of $\mathcal{A}$, except for (maybe) column $h(k)$, $\mathcal{A}$ has at least one interior element from columns $1$ to $h(k)$ by induction hypothesis on $\mathcal{A}^+_k$ and since $\mathcal{A}$ has one interior element from columns $h(m)+1$ to $h(r)$, the theorem holds in this case. Similarly, the proof holds if $BT$ arrives at $a_{i,h(k)}$ for $2\le k\le r-2$ and $i\le k$. Now suppose that $BT$ passes always below the $i$--st corner, for every $i\ge2$. If $TT$ and $BT$ share steps from columns $1$ to $h(2)$, then $\mathcal{A}$ has at least one interior element from columns $1$ to $h(2)$ and since $\mathcal{A}$ has one interior element from columns $h(m)+1$ to $h(r)$, the theorem holds. Then, we may suppose that $TT$ and $BT$ does not share steps from columns $1$ to $h(2)$. Also, if $a_{m,h(m)-1}\in PT$, then column $h(m)$ is an interior element and the theorem holds. So, we may suppose that $a_{m,h(m)-1}\not\in PT$. Hence, $a_{r,h(m)-2}\in BT$ by the rules of Property \ref{prop:chessboard}. Let $\mathcal{A}^{'}$ be the matrix obtained by turning the matrix $\mathcal{A}$ upside down. We observe that $BT$ and $PT$ are the Top and Bottom Travels of $\mathcal{A}^{'}$, respectively. Let define the $i$--st corners of $\mathcal{A}^{'}$ as $a_{r-i+1,h(r-i+1)}$ for $i\neq 2$ and define the $2$--st corner of $\mathcal{A}^{'}$ as $a_{m,h(m)-1}$. Notice that $BT$ always passes strictly above all corners of $\mathcal{A}^{'}$ after the $1$--st corner of $\mathcal{A}^{'}$. Moreover, $B[A']$ has the same sequence and the same $2$--st corner as that considered in Lemma \ref{lemmageneral2} (vi). Hence, as $TT$ and $BT$ do not share steps from columns $h(1)$ to $h(m)-1$, we know by Lemma \ref{lemmageneral2} (vi) that $a_{i,h(m)-1}\in PT$ for some $i\ge r-1$, but this is a contradiction since we had assumed that $a_{m,h(m)-1}\not\in PT$ and clearly $a_{r,h(m)-1}\not\in PT$, concluding the proof. \end{proof} The chessboard considered in Theorem \ref{cotageneral} can not be extended to a chessboard with sequence $(2,t+3,2,3,2,3,\ldots)$ in order to prove $\bar{n}(d,t)<2d+\lceil\frac{d+1}{2}\rceil+t$. Figures \ref{contraexample1} (b) and (c) provide examples of this phenomena for $d=5, t=2$, and for $d=4, t=3$, respectively. In fact, these examples can be generalized in order to show that this chessboard can not be used to prove $\bar{n}(d,t)<2d+\lceil\frac{d+1}{2}\rceil+t$ for odd $d\ge5, t\ge2$ and for even $d\ge 4, t\ge3$. The upper bound given in Theorem \ref{cotageneral} can be improved when $d$ is even. \begin{theorem}\label{dpar} $\bar{n}(d,t)<2d+(t-1)\frac{d}{2}+3$ for even $d\ge4$ and $t\ge 2$. \end{theorem} This theorem can be proved by making one final tweak to the chessboard defined previously. The latter is a bit technical and requires some extra work in the same flavour as above. This will be done in the Annex. \subsection{Proofs of the main results.} \begin{proof}[Proof of Theorem \ref{boundsforf_o}.] Recall that $n(d,t)= \bar{n}(d,t)$ and that $H_0(n(d,t),d)=n(d,t)-t$. \begin{itemize} \item $d=1, \ n\ge 2$. We clearly have that $H_0(n,1)=2$ since every convex set in dimension $1$ has as support only two vertices. \item $d=2, \ n\ge 5$. By Corollary \ref{igualdadv}, we have $n(2,t)=5+t$ for any integer $t\ge 0$. So, by Remark \ref{relationf_0andn(d,t)}, we have $H_0(t+5,2)=5$ for any integer $t\ge 0$ or, equivalently, $H_0(n,2)=5$ for any integer $n\ge 5$. \item $d=3, \ n\ge 7$. By Theorem \ref{dimension3}, we have $n(3,t)\le 7+t$ for any integer $t\ge 0$. So, by Remark \ref{relationf_0andn(d,t)}, we have $H_0(t+7,3)\le 7$ for any integer $t\ge 0$ or, equivalently, $H_0(n,3)\le 7$ for any integer $n\ge 7$. \smallskip \item $d\ge 2, \ n\le 2d+1$. By Proposition \ref{easybounds}, $H_0(n,d)=n.$ \item $d\ge 4, \ n \ge 2d+\lceil\frac{d+1}{2}\rceil.$ By Proposition \ref{easybounds}, $H_0(n,d)< n.$ \item $d\ge 4, \ n\ge 2d+\lceil\frac{d+1}{2}\rceil+1.$ By Theorem \ref{tpequeno} $\bar{n}(d,1)< 2d+\lceil\frac{d+1}{2}\rceil+1$. Therefore, $H_0(n,d) < n-1.$ \item $d\ge 4, \ 2d+3+l(d-2) \leq n < 2d+3+(l+1)(d-2),$ $l\geq 1.$ By Theorem \ref{cotageneral}, $\bar{n}(d,t)<2d+(t-1)(d-2)+3$ for every $t\ge 2.$ Since $H_0(n(d,t),d)=\bar{n}(d,t)-t$ then, for a given $n,$ it would be enough to work out $t$ such that $2d+(t-2)(d-2)+3 \leq {n}<2d+(t-1)(d-2)+3,$ in order to conclude $H_0(n,d) < n-t.$ It is not hard to see that we may take $t=\lfloor\frac{n-2d-3}{d-2}\rfloor+1\ge2$ and it follows that $H_0(n,d)\le n-(\lfloor\frac{n-2d-3}{d-2}\rfloor+2)$ for every $n\ge3d+1$. Finally, this can be expressed as; if $2d+3+l(d-2) \leq n < 2d+3+(l+1)(d-2)$ for some $l\ge1$, then $H_0(n,d) \le n-(l+2)$. \end{itemize} \end{proof} \begin{proof}[Proof of Theorem \ref{boundsforf_d-1}] The bounds are obtained by combining Proposition \ref{easybounds}, Theorem \ref{boundsforf_o} and inequality \eqref{upperboundcyclicp}. \end{proof} \section{Arrangement of pseudo-hyperplanes}\label{sec:arrangements} Recall that a {\em projective} $d$-arrangement of $n$ (pseudo) hyperplanes $\mathcal{H}(d,n)$ is a finite collection of pseudo-hyperplanes in the projective space $\mathbb{P}^d$ such that no point belongs to every hyperplane of $\mathcal{H}(d,n)$. Any such arrangement, $\mathcal{H}$ decomposes $\mathbb{P}^d$ into a $d$--dimensional cell complex. A cell of dimension $d$ is usually called a \emph{tope} of the arrangement $\mathcal{H}$. The well-known Topological Representation Theorem due to Folkman and Lawrence \cite{FOLKMAN1978199} states that loop-free oriented matroids of rank $d+1$ on $n$ elements (up to isomorphism) are in one-to-one correspondence with arrangements of pseudo-hyperplanes in the projective space $\mathbb{P}^{r-1}$ (up to topological equivalence). A $d$-arrangement is said to be \emph{simple} if no $d+1$ hyperplanes have a common point, i.e., the hyperplanes are in general position. It is known that simple $d$-arrangements of $n$ (pseudo) hyperplanes correspond to \emph{uniform} oriented matroids of rank $d+1$ on $n$ elements. If we denote as $\mathcal{H}=\{h_i\}_{1\le i\le n}$ an arrangement of (pseudo)hyperplanes and $M_{\mathcal H}$ as its corresponding oriented matroid, where $e_i$ is the element of $M_{\mathcal H}$ corresponding to (pseudo)hyperplane $h_i$, then, it is known that an acyclic reorientation of $M_{\mathcal A}$ having $\{e_{i_1},e_{i_2},\ldots,e_{i_l}\}$, with $l\le n$, as interior elements corresponds to a tope in ${\mathcal H}$ which is bordered precisely by the (pseudo)hyperplanes $h_j\not\in\{h_{i_1},h_{i_2},\ldots,h_{i_l}\}$. The \emph{size} of a tope is the number of (pseudo)hyperplanes bordering it. The function $\bar n(d,t)$ can thus be defined in terms of hyperplane arrangements as follows \begin{quote} $\bar n(d,t):=$ the largest integer $m$ such that any simple arrangement of $m$ (pseudo) hyperplanes in $\mathbb{P}^d$ contains a tope of size at least $m-t$. \end{quote} The latter is related to the following conjecture by Las Vergnas: \begin{quote}\emph{Every arrangement of (pseudo) hyperplanes in $\mathbb{P}^d$ admits a tope of size $d+1$ (a simplex).} \end{quote} This conjecture is known to be true when the arrangements are hyperplanes i.e., for realizable oriented matroids \cite{Roudneff1988, Shannon1979}. We are able to give a lower bound for $\bar{n}(d,t)$ when $d=2$. \begin{proposition}\label{lowerboundvd=2} Every simple arrangement of at least $5$ (pseudo) lines in $\mathbb{P}^2$ has a tope of size at least $5$. \end{proposition} \begin{proof} The proof is by induction on the set of $n$ (pseudo) lines. By Equation \ref{eq:boundsmcmullengen}, any arrangement of $5$ (pseudo) lines in $\mathbb{P}^2$ has a tope of size $5$ and thus the proposition holds for $n=5$. We suppose the result true for $n'<n$ and will prove that any arrangement $H$ of $n\ge6$ (pseudo) lines in $\mathbb{P}^2$ has a tope of size at least $5$. Let $l\in H$, then by induction $H-{l}$ has a tope $T$ of size at least $5$ in $\mathbb{P}^2$. If $l$ does not intersect $T$, then $T$ is a tope of $H$ of size at least $5$ in $\mathbb{P}^2$. Otherwise, $l$ divides $C$ into two topes, and since $H$ is simple then one of these two topes is of size at least $5$. \end{proof} \begin{corollary}\label{corolowerboundvd=2} $5+t\le \bar{n}(2,t)$ for every $t\ge 0$. \end{corollary} By combining this result with Theorem \ref{dimension2}, we obtain the following. \begin{corollary}\label{igualdadv} $\bar{n}(2,t)=5+t$ for any $t\ge0$. \end{corollary} \section{Pach and Szegedy's question}\label{sec:anapplication} In order to investigate Question \ref{question:ps}, we may consider a more general problem. Let $X$ be a finite set of points in $\mathbb R^d$. We say that $A, B$ is a {\em Radon partition} of $X$ if $X=A\cup B$, $A\cap B=\emptyset$ and $conv(A) \cap conv(B)\neq \emptyset$. Let $X=\{x_1,\dots,x_n\}\subset\mathbb R^d,$ $n\geq d+2$ be a set of points in general position. For a partition $X=A\cup B$ we define $r_{X}(A,B)$ as the number of $(d+2)$--size subsets $S \subset X$ such that $conv(A\cap S) \cap conv(B \cap S) \neq \emptyset$, that is, the partition $A, B$ induces a Radon partition of $S$. Let $$r(X)= \max\limits_{ \{ (A, B) | A\cup B = X \}} r_{X}(A,B).$$ We define $$r(d,n)=\min_{ X \subset \mathbb{R}^d, |X|=n} r(X).$$ \begin{theorem}\label{lem:radon} Let $d,n\ge 1$ be integers. Then, $$r(d,n)=H_{d'-1}(n,d') \text{ where } d=d'-n+2.$$ \end{theorem} In order to prove this theorem we need to take a geometric detour on the relationship between faces of convex polytopes, simplices embracing the origin and Radon partitions. There is an old tradition of using Gale transforms to study facets of convex polytopes \cite{GRUN2003}, by studying simplices embracing the origin (this equivalence was further extended by Larman \cite{LARMAN72} to studying Radon partitions of points in space). Given a finite set of points, $X=\{x_{1}, \ldots,x_{n}\}$ such that the dimension of their affine span is $r$, the set $D(X)=\{ \alpha=(\alpha_{1},\ldots,\alpha_{n}) | \; \sum_{i=1}^{n}\alpha_{i} x_{i}=0,\; \sum_{i=1}^{n}\alpha_{i}=0 \}\subset \mathbb{R}^{n}$ of its \emph{affine dependences} is of dimension $n-r-1.$ Let $\{a_{i}= (\alpha_{1, i},\ldots,\alpha_{n, i})\}_{i=1}^{n-r-1}$ be a basis of $D(X),$ then the set $\bar X =\{ \bar x_{j}= (\alpha_{j, 1}, \ldots \alpha_{j, n-r-1})\}_{j=1}^{n}$ is a \emph{Gale transform} of $X$. It is emphasized that $\bar{X}$ is \emph{a} Gale transform of $X,$ rather than \emph{the} Gale transform of $X,$ because the resulting points depend on the specific choice of basis for $D(X).$ Still, different Gale transforms of the same set of points are linearly equivalent. The \emph{Gale diagram} of $X$ is its normalized Gale transform, that is: the set of points $\hat{X}=\{\hat{x_i}= \frac{\bar{x}_i}{\lVert \bar{x}_i \rVert} \text{ if } \bar{x}_i\neq 0, \hat{x_i}=0 \text{ otherwise} \; | \; \bar{x}_i \in \bar{X}\}\in \mathbb{S}^{n-r-2}.$ We will state some of the properties of Gale transforms and diagrams (we refer the reader to \cite{GRUN2003} for proofs and a full introduction to Gale transforms) Let ${X}=\{{x_{1}}, \ldots, {x_{n}}\}$ be a set of $n$ points in $\mathbb{R}^{d}$ and let $\hat{X}$ ($\bar X$) be its Gale diagram (transform), then the following statements hold: (a) The $n$ points of $X$ are in general position in $\mathbb{R}^{d}$ if and only if the n-tuple $\hat{X}$ ($\bar X$) consists of $n$ points in linearly general position in $\mathbb{R}^{n-d-1}.$ (b) Faces of $conv(X)$ are in one to one correspondence with simplices of $\hat{X}$ ($\bar {X}$) that contain $0$ in their convex hull. More precisely, $Y \subset X$ is a face of $conv(X)$ iff $0 \in relint \: conv(\hat{X}\setminus \hat{Y})$ ($0 \in relint \: conv(\bar{X}\setminus \bar{Y})$). (c) $X$ is projectively equivalent to a set of points $Y$ (by a permissible projective transformation) if and only if there is a non zero vector $\epsilon=(\epsilon_{1},\ldots, \epsilon_{n}) \in \{1,-1\}^{n}$ ($\lambda=(\lambda_{1},\ldots, \lambda_{n}) \in \mathbb{R}^{n}$) such that $\hat{y}_i=\epsilon_i \hat{x}_i$ ($\bar{y}_i=\lambda_{i}\bar{x}_i$). Given a set of $n\geq d+2$ points in general position $X=\{x_1,\dots,x_n\}\subset\mathbb R^d$ and a partition $A \cup B =X,$ where $A \cap B= \emptyset$ we will construct its \emph{affine projection} into the unit sphere as follows; $\tilde{X}= \{\tilde{x}_i= \mathbf{I}(x_i) \frac{(x_i; 1)}{\lVert (x_i; 1) \rVert} | x_i \in X \} \subset \mathbb{S}^{d},$ where $\mathbf{I}(x_i)=1$ when $x_i \in A$ and $\mathbf{I}(x_i)=-1$ when $x_i \in B.$ By using basic linear algebra it can be easily be proved that (d) $A, B$ is a Radon partition of $X$ if and only if $0 \in conv(\tilde{A}\cup \tilde{B}).$ We have gathered now all the tools necessary to prove Theorem \ref{lem:radon}. \begin{proof}[Proof of Theorem \ref{lem:radon}.] By (d) , we know that if we consider the affine projection of $X$ into $\mathbb{S}^{d},$ $\tilde{X},$ we have that $conv(A\cap S) \cap conv(B \cap S) \neq \emptyset$ if and only if $0 \in conv( \tilde{S}).$ So, if we let let $\tilde{X}_\epsilon=\{\epsilon_1 \tilde{x}_1, \ldots, \epsilon_n \tilde{x}_n\}$ for $\epsilon=(\epsilon_1, \ldots, \epsilon_n) \in \{1, -1\}^n$ and define $\rho(\tilde{X}_\epsilon)$ as the number of subsets $\tilde{S}\subset \tilde{X}_\epsilon$ such that $0 \in conv(\tilde{S})$ and $\rho(\tilde{X})= \max_{\epsilon \in \{1, -1\}^n} {\rho(\tilde{X}_\epsilon)},$ then we can find the value of $\rho(d,n)=\min_{ \{\tilde{X} \subset \mathbb{S}^d, |\tilde{X}|=n\}}\rho(\tilde{X})$. We thus have that $r(d,n)=\rho(d,n)$, we observe that $\tilde{X} \subset \mathbb{S}^{d} \subset \mathbb{R}^{d+1}$ while $X\subset\mathbb R^d.$ Now, recall that the set $\tilde{X} \subset \mathbb{S}^{d}$ can be considered to be the Gale diagram of a set of points in $X' \subset \mathbb R^{n-d-2}$ where each $\tilde{X}_\epsilon=\{\epsilon_1 \tilde{x}_1, \ldots, \epsilon_n \tilde{x}_n\}$ corresponds to a permissible projective transformation of $X'.$ Therefore, each $(d+2)$--set $ \tilde{S} \subset \tilde{X}$ such that $0 \in conv( \tilde{S})$ is in one to one correspondence with a cofacet of $X'$ (i.e. the corresponding set $X'\setminus S'$ is a facet of $X').$ \end{proof} As a consequence of Lemma \ref{lem:radon}, together with Corollary \ref{boundsforf_d-1}, we obtain. {\scriptsize $$r(d,n)\left\{\begin{array}{ll} =1 & \text{ if } n= d+2,\\ =2 & \text{ if } n= d+3,\\ =5 & \text{ if } n= d+4,\\ \le 10 & \text{ if } n= d+5,\\ \le f_{n-d-3}(C_{n-d-2}(n) & \text{ if } n\ge 2d+3,\\ \le f_{n-d-3}(C_{n-d-2}(n-1)) & \text{ if } n \leq \frac{5d+8}{3},\\ \le f_{n-d-3}(C_{n-d-2}(n-2)) & \text{ if } n \leq \frac{5d+6}{3},\\ \le f_{n-d-3}(C_{n-d-2}(n-(l+2))),\ \ 1 \leq l & \text{ if } d+4+ \frac{d-3}{l+2} < n \leq d+4 +\frac{d-1}{l+1}.\\ \end{array}\right.$$} Moreover, $r(d,n)\ge 8$ when $n=d+5$. The thorough reader may have noticed that we have not addressed yet the second part of Question \ref{question:ps}, that is, \begin{quote} is the partition of $X$ balanced when the maximum is attained ? \end{quote} The following result answers this question negatively. \begin{theorem} \label{thm:unbalanced} Let $X\subset \mathbb{R}^d,$ be a set of points in general position, if $A, B$ is a partition such that its number of induced minimal Radon partitions is maximal, then the size of $A$ and $B$ can be as different as possible. \end{theorem} \begin{proof} Let $A,B$ be a partition that attains the maximum number of induced minimal Radon partitions for a set $X,$ $r(X),$ and let $\tilde{X}_\epsilon$ be it's corresponding affine projection, and $X'$ be the point configuration whose Gale diagram is $\tilde{X}_\epsilon$, as defined before. Note that $X$ will have the maximum number of induced partitions for a set of its size if and only if $X'$ is in the projective class of a neighbourly polytope. This will not be the case in general. Thus we argue that we can have some set, $X,$ that will be in the projective class of a set of points, $X',$ that cannot be made convex by any projective transformation. In particular, if $X'$ cannot be made convex, there is a point $x'$ in the interior of $conv(X').$ By (b), the set $\tilde{X}_\epsilon \setminus \tilde{x'}$ is such that $0 \not \in conv(\tilde{X}_\epsilon \setminus \tilde{x'}),$ where $\tilde{x'}$ is the point of $\tilde{X}_\epsilon$ corresponding to $x'.$ Hence, there is a hyperplane $H$ through the origin such that $H^+ \cap \tilde{X}_\epsilon=\{ \tilde{x'}\}$ and $H^- \cap \tilde{X}_\epsilon=\tilde{X}_\epsilon \setminus \tilde{x'}.$ Now, observe that the affine projection of $X$ was arbitrarily constructed making the hyperplane $aff(X) \subset \mathbb{R}^{d+1}$ be parallel to the hyperplane $x_{d+1}=0.$ However, we can "lift" the configuration $\tilde{X}_\epsilon \subset \mathbb{R}^{d+1}$ by choosing a hyperplane $H'$ which is parallel to $H$ by translating it in the direction of $H$'s unit normal vector to its "north pole" and then defining $Y=\{y=lin(\tilde{x}) \cap H' | \tilde{x} \in \tilde{X}_\epsilon\},$ where $lin(\tilde{x})$ is the line through the origin and $\tilde{x}.$ Furthermore, we can define a partition $A_Y, B_Y$ of the set $Y$ as follows $y=lin(\tilde{x}) \cap H' \in A_Y$ when $\tilde{x} \in H^+,$ $y \in B_Y$ otherwise. Here $|A_Y|=1$ and $|B_Y|=n-1.$ We claim that the number of induced minimal Radon partitions has to be the same as for $X, A$ and $B,$ i.e. $r(X)=r(Y).$ Recall that a subset $T$ of $Y$ is an induced Radon partition if and only if its corresponding set in $\tilde{X}_\epsilon$ contains the origin. So, no partition of $Y$ can lead to a selection $\epsilon' \in \{1, -1\}^n$ such that the number of simplices embracing the origin for $\tilde{X}_\epsilon'$ will be greater than that of $\tilde{X}_\epsilon,$ because it would contradict the maximality of the selection of $\epsilon.$ So here we have it. We've built a set $Y$ in $\mathbb{R}^d$ and a partition $A_Y, B_Y$ such that its number of induced minimal Radon partitions is maximal but $|A_Y|$ and $|B_Y|$ are as different as they can be. \end{proof} Let us highlight the following two facts. \begin{fact} The maximum number of minimal Radon partitions induced by a partition into two sets, $A, B$ of a set of points in general position $X$ depends on the projective class of the set and it is independent of the balance of the cardinality of $A$ and $B.$ \end{fact} The above follows as the set $X$ and $Y$ in the proof of Theorem \ref{thm:unbalanced} are projectively equivalent. \begin{fact} If a set of $n$ points in general position $X \subset \mathbb{R}^d$ is such that its affine projection is in the projective class of a neighbourly polytope then, there is a \emph{balanced} partition $A,B$ such that the number of induced Radon partitions is the maximum possible for a configuration of $n$ points. \end{fact} This can be easily argued as follows: Let $\tilde{X},$ be the affine projection of $X$ and let $\epsilon \in \{1, -1\}^n$ such that the set $\tilde{X}_\epsilon$ corresponds to the Gale diagram of a neighbourly polytope. Recall that a set of vertices is the face of a polytope if and only if the vertices corresponding to its complement in the Gale diagram contain the origin in its convex hull and that, by definition, every set of $\lfloor \frac{d'}{2}\rfloor$ points is a face of a $d'$--dimensional neighbourly polytope; hence, every plane through the origin $H$ is such that $|H^+ \cap \tilde{X}_\epsilon|>\lfloor \frac{d'}{2}\rfloor$ and $|H^- \cap \tilde{X}_\epsilon|>\lfloor \frac{d'}{2}\rfloor.$ In this case, the dimension of the neighbourly polytope associated to $\tilde{X}_\epsilon$ is $d'=n-d-2.$ If we now construct $A$ and $B$ as usual ($x \in A$ iff $\tilde{x}_\epsilon$ is in the north hemisphere, $x \in B$ otherwise), the cardinality of both $A$ and $B$ are strictly greater than $\lfloor \frac{n-d-2}{2}\rfloor,$ thus the partition is somewhat balanced. \section{Concluding remarks}\label{sec:concluding-rem} Corollary \ref{igualdadv} states that $\bar{n}(2,t)=5+t$ for any $t\ge0$, in other words, any simple arrangement of $n\ge5$ (pseudo) lines in $\mathbb{P}^2$ contains a tope of size at least $5$. Moreover, for each $n\ge5$ there exists an example of a simple arrangement of $n$ hyperplanes with all its topes having size at most $5$. The latter suggests the following \begin{question}\label{conjeturav(d,t)=2d+t} Let $d\ge2$ and $t\ge 0$ be integers. Is it true that $\bar{n}(d,t)=2d+1+t?$ In other words, is it true that any simple arrangement of $n\ge 2d+1$ (pseudo) hyperplanes in $\mathbb{P}^d$ contains a tope of size at least $2d+1$ and conversely, for any $n\ge 2d+1$ there exists a simple arrangement of $n$ (pseudo) hyperplanes in $\mathbb{P}^d$ with every tope of size at most $2d+1$? \end{question} Or, alternatively, \begin{question}\label{conjeturac(d)} Let $d\ge2$ and $t\ge 0$ be integers. Is there a constant $c(d)\ge 1$ such that $\bar{n}(d,t)=2d+1+c(d)t ?$ \end{question} In the case $d=3$, we know (by Theorem \ref{dimension3}) that $\bar{n}(3,t)\le t+7$ for any integer $t\ge 1$, i.e., for any $n\ge 7$ there exists a simple arrangement of $n$ (pseudo) hyperplanes in $\mathbb{P}^3$ with every tope of size at most $7$. \begin{conjecture}\label{questionv(3,t)=t+7} $\bar{n}(3,t)= t+7$ for any integer $t\ge 1$. In other words, any simple arrangement of at least $7$ (pseudo) hyperplanes in $\mathbb{P}^3$ always contains a tope of size at least $7$. \end{conjecture} \bibliographystyle{amsplain}

https://arxiv.org/abs/1301.5020

Generalized cover ideals and the persistence property

Let $I$ be a square-free monomial ideal in $R = k[x_1,\ldots,x_n]$, and consider the sets of associated primes ${\rm Ass}(I^s)$ for all integers $s \geq 1$. Although it is known that the sets of associated primes of powers of $I$ eventually stabilize, there are few results about the power at which this stabilization occurs (known as the index of stability). We introduce a family of square-free monomial ideals that can be associated to a finite simple graph $G$ that generalizes the cover ideal construction. When $G$ is a tree, we explicitly determine ${\rm Ass}(I^s)$ for all $s \geq 1$. As consequences, not only can we compute the index of stability, we can also show that this family of ideals has the persistence property.

\section{Introduction} Let $I$ be an ideal of the polynomial ring $R = k[x_1,\ldots,x_n]$ with $k$ a field. A prime ideal $P \subseteq R$ is an {\it associated prime} of $I$ if there exists an element $T \in R$ such that $I:\langle T \rangle = P$. The {\it set of associated primes} of $I$, denoted ${\rm Ass}(I)$, is the set of all prime ideals associated to $I$. We shall be interested in the sets ${\rm Ass}(I^s)$ as $s$ varies. Brodmann \cite{B} proved that there exists an integer $s_0$ such that ${\rm Ass}(I^s) = {\rm Ass}(I^{s_0})$ for all integers $s \geq s_0$. The least such integer $s_0$ is called the {\it index of stability}, and following \cite{HQ}, we denote it by {\rm astab}$(I)$. We are interested in the following problem which arises from Brodmann's result: determine {\rm astab}$(I)$ in terms of the invariants of $R$ and $I$. Little is known about this problem, and in particular, there are few results providing exact calculations of {\rm astab}$(I)$. An upper bound on {\rm astab}$(I)$ for any monomial ideal $I$ was given by Hoa \cite{H}. This bound is quite large and is in terms of the number of variables in the ring, the number of minimal generators of the ideal and the maximal degree of a minimal generator. Even when $I$ is a square-free monomial ideal, determining ${\rm astab}{(I)}$ remains a challenging problem. A lower bound for ${\rm astab}{(I)}$ was given in \cite{FHVT} in terms of the chromatic number of a hypergraph constructed from the primary decomposition of $I$. When $I$ is the edge ideal of a graph (a quadratic square-free monomial ideal), Chen, Morey and Sung \cite{CMS} provide an upper bound on {\rm astab}$(I)$. However, the recent work of \cite{BHR,FHVT,FHVTconj,HQ,HRV,MMV} has suggested a possible answer. In particular, Herzog and Qureshi \cite{HQ} posit that the bound ${\rm astab}(I) \leq \dim R-1 = n-1$ should hold for square-free monomial ideals $I$ (this bound is significantly smaller than that given in \cite{H}). Brodmann's results also suggest the following secondary question: which ideals satisfy the {\it persistence property}, that is, for which ideals does the containment ${\rm Ass}(I^s)\subseteq {\rm Ass}(I^{s+1})$ hold for all $s \geq 1$? Recently, Kaiser, Stehl\'ik, and \u Skrekovski \cite{KSS} have shown that not all square-free monomial ideals have this property. In light of this result, it is an interesting question to determine which square-free monomial ideals have the persistence property. Results in this direction have shown that the persistence property holds for many classes of square-free monomial ideals, including square-free principal Borel ideals \cite{A}, edge ideals \cite{MMV}, the cover ideals of perfect graphs \cite{FHVT}, and polymatroidal ideals \cite{HRV}. In this paper, we introduce a family of square-free monomial ideals (generalizing the notion of a cover ideal) that can be associated to a finite simple graph $G$, and study the associated primes of their powers. More formally, suppose that $G$ is a finite simple graph on the vertex set $V_G = \{x_1,x_2,\ldots,x_n\}$ with edge set $E_G$. For any $x \in V_G$, we let $N(x) = \{y ~|~ \{x,y\} \in E_G\}$ denote the set of {\it neighbours of $x$}. By identifying the vertex $x_i$ with the variable $x_i$ in $R$, we define the following ideals. \begin{definition}\label{partial} Fix an integer $t \geq 1$. The {\it partial $t$-cover ideal} of $G$ is the monomial ideal \[J_t(G) = \bigcap_{x \in V_G} \left(\bigcap_{\{x_{i_1},\ldots,x_{i_t}\} \subseteq N(x)} \langle x,x_{i_1},\ldots,x_{i_t} \rangle \right).\] \end{definition} \noindent When $t=1$, our construction is simply the cover ideal of a finite simple graph $G$ (see Section 2 for more details). Recall that a graph is a {\it tree} if it has no induced cycles. Our main result is to show that for when $G$ is a tree, we can compute the index of stability of $J_t(G)$, and show that this family has the persistence property. \begin{theorem}\label{maintheorem} Let $G = (V_G,E_G)$ be a tree on $n$ vertices and fix any integer $t \geq 1$. Then the partial $t$-cover ideal $J_t(G)$ satisfies the persistence property. Furthermore \[{\rm astab}(J_t(G)) = \left\{ \begin{array}{ll} 1 & \mbox{if $t=1$} \\ \min\{s ~|~ s(t-1) \geq \Delta(G) -1 \} & \mbox{if $t > 1$} \end{array} \right. \] where $\Delta(G)$ is the maximal degree of $G$, i.e., the largest degree of a vertex of $G$. \end{theorem} \noindent In fact, we prove a stronger result (Theorem \ref{maintheoremtrees}) by determining the elements of ${\rm Ass}(J_t(G)^s)$ for all $s \geq 1$. Note that $\Delta(G) \leq n-1$, so the upper bound suggested by Herzog and Qureshi also holds for this family. Our paper is structured as follows. In Section 2, we review the required ingredients of associated primes and describe some of the properties of $J_t(G)$. In Section 3, we specialize to the case that $G = K_{1,n}$ is the star graph. These graphs will play an important role in our proof of Theorem \ref{maintheorem}; we also use these graphs to answer a question raised in \cite{FHVT}. Section 4 is devoted to the proof of our main result. \noindent {\bf Acknowledgements.} Some of the results of Section 3 first appeared in \cite{Bhat}. {\em Macaulay2} \cite{Mt} was used for computer experiments. We thank T. H\`a and C. Francisco for their feedback. The third author acknowledges the support of an NSERC Discovery Grant. \section{Preliminaries} We continue to use the terminology and definitions introduced in the previous section. Throughout this paper, $\mathcal{G}(I)$ denotes the unique set of minimal generators of a monomial ideal $I$. For any $W = \{x_{i_1},\ldots,x_{i_s}\} \subseteq V_G$, we let $x_W = x_{i_1}\cdots x_{i_s} \in R$. We first explain the significance of the name partial $t$-cover ideal in Definition \ref{partial}. A {\it vertex cover} of a graph $G$ is a subset $W \subseteq V_G$ which satisfies the following property: for any $x \in V_G$, either $x \in W$ or $N(x) \subseteq W$. In other words, all the edges containing $x$ are covered. We generalize this definition: a {\it partial $t$-cover} is a subset $W \subseteq V_G$ which satisfies the following property: for any $x \in V_G$, either $x \in W$ or there exists some subset $S \subseteq N(x)$ with $|S| = |N(x)|-t+1$ and $S \subseteq W$. That is, for each $x \in V_G$, all, but perhaps $t-1$ of the edges containing $x$, are covered by $W$. When $t =1$, this is simply the definition of a vertex cover. The following lemma justifies our choice of name for $J_t(G)$. \begin{lemma} \label{genspartial} Let $G = (V_G,E_G)$ be a finite simple graph and $t \geq 1$ an integer. Then \[J_t(G) = \langle x_W ~|~ \mbox{$W \subseteq V_G$ is a partial $t$-cover} \rangle.\] \end{lemma} \begin{proof} Let $m \in \mathcal{G}(J_t(G))$, and so $m = x_W$ for some $W \subseteq V_G$. Suppose $W$ is not a partial $t$-cover. Then there exists a vertex $x$ such that $x \not\in W$, and for all $S \subseteq N(x)$ with $|S| = |N(x)|-t+1$, there is some $x_j \in S \setminus W$. We claim that there are $t$ neighbours of $x$ not in $W$. Let $S_1 = \{x_1,\ldots,x_{|N(x)|-t+1}\}$. Because $W$ is not a partial $t$-cover, let $x_{i_1} \in S_1 \setminus W$. Set $S_2 = (S_1 \setminus \{x_{i_1}\}) \cup \{x_{|N(x)|-t+2}\}$. Again, $W$ is not a partial $t$-cover, so there exists $x_{i_2} \in S_2 \setminus W$. We repeat $t$ times and find $t$ neighbours of $x$, say $\{x_{i_1},\ldots,x_{i_t}\}$, that do not appear in $W$. It then follows that $m = x_W \not\in \langle x,x_{i_1},\ldots,x_{i_t} \rangle$ since none of these variables appear in $x_W$. But this contradicts the fact that $m \in J_t(G) \subseteq \langle x,x_{i_1},\ldots,x_{i_t} \rangle$. Therefore $W$ is a partial $t$-cover. For the converse, let $x_W$ be any square-free monomial which corresponds to a partial $t$-cover. Rewrite $J_t(G)$ as \footnotesize \[J_t(G) = \left(\bigcap_{x \in W} \left(\bigcap_{\{x_{i_1},\ldots,x_{i_t}\} \subseteq N(x)} \langle x,x_{i_1},\ldots,x_{i_t} \rangle \right)\right) \cap \left( \bigcap_{x \in V_G \setminus W} \left(\bigcap_{\{x_{i_1},\ldots,x_{i_t}\} \subseteq N(x)} \langle x,x_{i_1},\ldots,x_{i_t} \rangle \right)\right) .\] \normalsize If $x \in W$, then $x_W \in \langle x,x_{i_1},\ldots,x_{i_t}\rangle$, so $x_W$ is in the first intersection. If $x \not\in W$, then there exists a subset $S \subseteq N(x)$ with $|N(x)|-t+1$ elements such that $S \subseteq W$. But then for any subset $T \subseteq N(x)$ with $|T| = t$, $S \cap T \neq \emptyset$. This implies that $x_W \in \langle x,x_{i_1},\ldots,x_{i_t} \rangle$ for each subset $\{x_{i_1},\ldots,x_{i_t}\}$ of $N(x)$ of size $t$. So $x_W$ is in the second intersection, thus completing the proof. \end{proof} \begin{remark} The Alexander dual (see \cite{MS} for the definition) of $J_t(G)$ is also of interest: \[ I_t(G):= J_t(G)^\vee = \sum_{x \in V_G} \langle xx_{i_1}\cdots x_{i_t} ~|~ \{x_{i_1},\dots, x_{i_t}\}\subseteq N(x)\rangle. \] If $t = 1$, then $I_1(G)$ is the edge ideal of $G$, and if $t=2$, then $I_2(G)$ is the 2-path ideal of $G$ (see \cite{CD} for the definition). The ideals $I_t(G)$ can be viewed as generalized edge ideals. In a future paper, we will investigate some of the properties of $I_t(G)$. \end{remark} We turn to the relevant results on associated primes of square-free monomial ideals. Via the technique of localization, and using the fact that localization and taking powers commute, we simply need to determine when the maximal ideal is an associated prime of a monomial ideal. The following lemma justifies this reduction. The proof is similar to the proof of \cite[Lemma 2.11]{FHVT}, so is omitted. Given a graph $G = (V_G,E_G)$ and subset $P \subseteq V_G$, we write $G_P$ for the {\it induced graph} on $P$, i.e., the graph with vertex set $P$, and edge set $E_{G_{P}} = \{e \in E_G ~|~ e \subseteq P\}$. \begin{lemma}\label{localization} Let $G$ be a graph on the vertex set $\{x_1, \dots, x_n\}$, and let $J_t(G)$ be the partial $t$-cover ideal of $G$. The following are equivalent: \begin{enumerate} \item[$(i)$] $P= \langle x_{i_1}, \dots, x_{i_r}\rangle \in {\rm Ass}(J_t(G)^s)$ in $R = k[x_1,\ldots,x_n]$ \item[$(ii)$] $P=\langle x_{i_1}, \dots, x_{i_r}\rangle \in {\rm Ass} (J_t(G_P)^s)$ in $R_P = k[x_{i_1},\ldots,x_{i_r}]$. \end{enumerate} \end{lemma} The next lemma shows $P \in {\rm Ass}(J_t(G)^s)$ gives a necessary condition on the graph $G_P$. \begin{lemma}\label{connected} Let $G$ be a graph on the vertex set $\{x_1, \dots, x_n\}$, and let $J_t(G)$ be the partial $t$-cover ideal of $G$. If $P =\langle x_{i_1}, \dots, x_{i_r}\rangle \in {\rm Ass}(J_t(G)^s)$, then $G_P$ is connected. \end{lemma} \begin{proof} By Lemma \ref{localization}, it is enough to show that if $\langle x_1,\ldots,x_n \rangle \in {\rm Ass}(J_t(G)^s)$ for some $s$, then $G$ is connected. Suppose $G$ is not connected, i.e., $G = G_1 \cup G_2$ with $G_1 \cap G_2 = \emptyset$. After relabeling the vertices, we can assume the vertices of $G_1$ are $\{y_1,\ldots,y_a\}$ and the vertices of $G_2$ are $\{z_1,\ldots,z_b\}$. If $m \in \mathcal{G}(J_t(G))$, then $m = m_ym_z$ where $m_y$ is a square-free monomial in the $y$ variables, and $m_z$ is a square-free monomial in the $z$ variables, and furthermore, we must have $m_y \in \mathcal{G}(J_t(G_1))$, and $m_z \in \mathcal{G}(J_t(G_2))$. Because $\langle x_1,\ldots,x_n \rangle = \langle y_1,\ldots,y_a,z_1, \ldots,z_b \rangle$, and $\langle x_1,\ldots,x_n \rangle \in {\rm Ass}(J_t(G)^s)$, there exists a monomial $T \not\in J_t(G)^s$ such that \begin{eqnarray*} Ty_1 &= &m_1\cdots m_sM ~~\mbox{with $m_i \in \mathcal{G}(J_t(G))$} \\ & = & m_{y,1}m_{z,1} \cdots m_{y,s}m_{z,s}M_yM_z ~~ \mbox{with $m_i = m_{y,i}m_{z,i}$} \end{eqnarray*} where $m_{y,i} \in \mathcal{G}(J_t(G_1))$ and $m_{z,i} \in \mathcal{G}(J_t(G_2))$, and $M_y$ (respectively $M_z$) is a monomial in the $y$ variables (respectively the $z$ variables). So, $T = (m_{z,1}\cdots m_{z,s}M_z)T'$ where $T'$ is a monomial in the $y$ variables. But we also know that $Tz_1 \in J_t(G)^s$, so a similar argument allows us to write $T = (u_{y,1}\cdots u_{y,s}U_y)T''$ where $T''$ is a monomial in the $z$ variables, $U_y$ is a monomial in the $y$ variables, and each $u_{y,j} \in \mathcal{G}(J_t(G_1))$. But this means \begin{eqnarray*} T &= &(m_{z,1}\cdots m_{z,s}M_z)(u_{y,1}\cdots u_{y,s}U_y) = (u_{y,1}m_{z,1})\cdots (u_{y,s}m_{z,s})U_yM_z. \end{eqnarray*} Now each $u_{y,i}m_{z,i} \in \mathcal{G}(J_t(G))$, so $T \in J_t(G)^s$, a contradiction. Thus $G$ is connected. \end{proof} Section 3 focuses on {\it star graphs} $G = K_{1,n}$. These are the graphs with vertex set $V_G = \{z,x_1,\ldots,x_n\}$ and edge set $E_G = \{\{z,x_i\} ~|~ 1 \leq i \leq n\}$. The generators of $J_t(K_{1,n})$, as described by the next lemma, follow directly from the definitions: \begin{lemma} \label{generators} Let $G = K_{1,n}$ with $V = \{z,x_1,\ldots,x_n\}$, and let $n \geq t \geq 1$. Then \[J_t(G) = \langle z \rangle + \langle x_{j_1}\cdots x_{j_{n-t+1}} ~|~ \{j_1,\ldots,j_{n-t+1}\} \subseteq \{1,\ldots,n\} \rangle.\] \end{lemma} The next example explains what we know about ${\rm Ass}(J_t(K_{1,n})^s)$ when $t=1$; the situation for $t \geq 2$ is explored in the next section. \begin{example}\label{caset=1} Let $G = K_{1,n}$ and $t=1$. By Lemma \ref{generators}, $J_1(G) = \langle z,x_1x_2\cdots x_n \rangle$. But this is a complete intersection, so for all $s \geq 1$, \[{\rm Ass}(J_1(G)^s) = {\rm Ass}(J_1(G)) = \{ \langle z,x_i \rangle ~|~ 1 \leq i \leq n\}. \] There are at least two ways to prove this result. For any complete intersection $J$, $J^s = J^{(s)}$, the $s$-th symbolic power of $J$ (see \cite{ZS}) and thus ${\rm Ass}(J^s) = {\rm Ass}(J)$ for all $s \geq 1$. Alternatively, Gitler, Reyes, and Villarreal have shown \cite[Corollary 2.6]{GRV} that $J_1(G)$ is normal, i.e., $J_1(G)^s = \overline{J_1(G)^s}$, whenever $G$ is a bipartite graph, whence the conclusion again follows. Because ${\rm astab}(J_1(G)) = 1$, $J_1(G)$ has the persistence property. \end{example} \section{Star graphs} Fix integers $n \geq t \geq 1$. In this section we will completely describe the sets ${\rm Ass}(J_t(G)^s)$ when $G = K_{1,n}$. We use our results to give a new answer to a question raised by Francisco, H\`a, and the third author in \cite{FHVT}. Our main result is a corollary of the following theorem: \begin{theorem}\label{maintheoremstar} Fix integers $n \geq t \geq 1$ and let $G = K_{1,n}$ be the star graph on $V_G = \{z,x_1,\ldots,x_n\}$. Set $J_t = J_t(G)$. The following are equivalent: \begin{enumerate} \item[$(i)$] $\langle z,x_1,\ldots,x_n \rangle \in{\rm Ass}(J_t^s)$ \item[$(ii)$] $s(t-1) \geq n-1$. \end{enumerate} \end{theorem} We postpone the proof, but record its consequences: \begin{corollary}\label{corstar} Fix integers $n \geq t \geq 1$ and let $G = K_{1,n}$ be the star graph on $V_G = \{z,x_1,\ldots,x_n\}$. For any $s \geq 1$, \[{\rm Ass}(J_t(G)^s) = \left. \left\{\langle z,x_{i_1},\ldots,x_{i_r} \rangle ~\right|~ t \leq r \leq \min\{n,s(t-1)+1\}\right \}.\] Moreover, \[{\rm astab}(J_t(G)) = \left\{ \begin{array}{ll} 1 & \mbox{if $t=1$} \\ \min\{s ~|~ s(t-1) \geq n -1 \} & \mbox{if $t > 1$.} \end{array} \right.\] \end{corollary} \begin{proof} The result on ${\rm astab}(J_t(G))$ follows from the first statement. Let $\mathcal{P}$ denote the set on the right hand side of the first statement. Let $P \in {\rm Ass}(J_t(G)^s)$. Because $G_P$ is connected by Lemma \ref{connected}, $P = \langle z,x_{i_1},\ldots,x_{i_r} \rangle$, i.e., $P$ cannot be generated by a subset of $x$ variables. Note that this means that $G_P = K_{1,r}$ for some $r$. Either $P$ is a minimal prime of $J_t(G)$, or contains a minimal prime of $J_t(G)$, thus showing showing that $t \leq r$. By Lemma \ref{localization}, $\langle z,x_{i_1},\ldots,x_{i_r} \rangle \in {\rm Ass}(J_t(G_P)^s)$, and so by Theorem \ref{maintheoremstar}, $s(t-1) \geq r-1$, i.e., $r \leq s(t-1) +1$. Also, it is clear that $r \leq n$, so $P \in \mathcal{P}$. Conversely, suppose that $P = \langle z,x_{i_1},\ldots,x_{i_r} \rangle \in \mathcal{P}$. Abusing notation, let $P \subseteq V_G$ denote the corresponding vertices. After localizing at $P$, $P \in {\rm Ass}(J_t(G_P)^s)$ by Theorem \ref{maintheoremstar} since $s(t-1) \geq r-1$. Lemma \ref{localization} then gives $P \in {\rm Ass}(J_t(G)^s)$. \end{proof} To prove Theorem \ref{maintheoremstar} we require some information about our annihilator. \begin{lemma}\label{annlemma} Fix integers $n \geq t \geq 1$ and let $G = K_{1,n}$ be the star graph on $V_G = \{z,x_1,\ldots,x_n\}$. Set $J_t = J_t(G)$. Suppose that there exists a monomial $T \in k[z,x_1,\ldots,x_n]$, $T \notin J_t^s$, such that $J_t^s:\langle T \rangle = \langle z,x_1,\ldots,x_n \rangle$. If $T = z^eT'$ where $z \nmid T'$, then $T \mid z^e(x_1 \cdots x_n)^{s-e-1}.$ \end{lemma} \begin{proof} It suffices to prove that $T' | (x_1\cdots x_n)^{s-e-1}$. Suppose that there exists some $x_i$ such that $x_i^{s-e} | T'$. Now $x_iT = z^ex_iT' \in J_t^s$, so \[z^ex_iT' = m_1m_2 \cdots m_sM ~~\mbox{with $M \in k[z,x_1,\ldots,x_n]$ and $m_i \in \mathcal{G}(J_t)$}.\] We cannot have $x_i|M$. If it did, then we could cancel $x_i$ from both sides and have $T =z^eT' = m_1\cdots m_s(M/x_i) \in J_t^s$, which contradicts the fact that $T \not\in J_t^s$. So, the variable $x_i$ appears at least $s-e+1$ times in $z^ex_iT'$, and thus, must appear in at least $s-e+1$ of $m_1,\ldots,m_s$, because each $m_j$ is square-free. In particular, we can assume $m_1 = x_im'_1$. This means at most $e-1$ of $m_1,\ldots,m_s$ can be equal to $z$ (no minimal generator of $J_t$ is divisible by both $z$ and $x_i$ by Lemma \ref{generators}). So, $z$ must divide $M$, i.e., $M = zM'$. So, to summarize, \[z^ex_iT' = m_1m_2\cdots m_sM = (x_im'_1)m_2\cdots m_s(zM').\] If we cancel $x_i$ from both sides, we get \[T = z^eT' = (m'_1)m_2\cdots m_s(zM').\] But $m_2,\ldots,m_s,z \in \mathcal{G}(J_t)$, which means $T \in J_t^s$. This is our desired contradiction. \end{proof} We are now ready to prove Theorem \ref{maintheoremstar}. \begin{proof} (of Theorem \ref{maintheoremstar}) Note that if $t=1$, then Example \ref{caset=1} implies $\langle z, x_1,\ldots,x_n \rangle \in {\rm Ass}(J_1(G)^s)$ if and only if $n=1$ if and only if $0 = s(t-1) \geq n-1$. So, we assume $t > 1$. $(i) \Rightarrow (ii)$. If $\langle z,x_1,\ldots,x_n \rangle \in {\rm Ass}(J_t^s)$, then there exists a monomial $T \notin J_t^s$ such that $J_t^s: \langle T \rangle = \langle z,x_1,\ldots,x_n\rangle$. Rewrite $T$ as $T = z^eT'$ where $z \nmid T'$. We now claim that \begin{equation}\label{specialmonomial} z^e(x_1\cdots x_{n-t+2})^{s-e}(x_{n-t+3}\cdots x_n)^{s-e-1} \in J_t^{s+1}. \end{equation} Indeed, by Lemma \ref{annlemma}, $z^eT' | z^e(x_1\cdots x_n)^{s-e-1}$. Now $x_1z^eT' \in J_t^s$, which means $$z^ex_1^{s-e}(x_2\cdots x_n)^{s-e-1} \in J_t^s.$$ But $x_2\cdots x_{n-t+2} \in J_t$, so multiplying these two elements together gives us the desired element in $J_t^{s+1}$. We proceed by a degree argument. By \eqref{specialmonomial} there exist generators $m_1, \ldots, m_{s+1}$ of $J_t$ such that \[z^e(x_1\cdots x_{n-t+2})^{s-e}(x_{n-t+3}\cdots x_n)^{s-e-1} =m_1 \cdots m_{s+1}M.\] By Lemma \ref{generators}, $f$ of these generators are of the form $z$, and the remaining $s+1-f$ generators are of degree $n-t+1$ and have the form $x_{j_1}\cdots x_{j_{n-t+1}}$ for some $\{j_1,\ldots,j_{n-t+1}\} \subseteq \{1,\ldots,n\}$. Note that we must have $f \leq e$, and thus, looking at the degree of the generators in the $x$ variables, we must have \[ (s+1-f)(n-t+1) \leq (n-t+2)(s-e) + (t-2)(s-e-1) = (s-e)n - (t-2).\] Expanding out the left hand side gives \[sn-st+s+n-t+1-fn+ft-f \leq sn -en -t + 2.\] Removing $sn$ and $-t$ from both sides and using the fact that $-en \leq -fn$ and $0 \leq f(t-1)$ gives $-st+s+n \leq 1$, which implies $s(t-1) \geq n-1$, as desired. $(ii) \Rightarrow (i)$ Let $s_0 = \min\{s ~|~ s(t-1) \geq n-1 \}$. We first show that $\langle z,x_1,\ldots,x_n \rangle \in {\rm Ass}(J_t^{s_0})$. We construct our annihilator as follows. Write out the variables $x_1,\ldots,x_n$ as a repeating sequence, i.e., \begin{equation}\label{word} x_1,x_2,\ldots,x_n,x_1,x_2,\ldots,x_n,x_1,x_2,\ldots,x_n,x_1,\ldots. \end{equation} Let $T$ be the product of the first $s_0(n-t+1)-1$ variables in this sequence, that is, \[T = \underbrace{x_1x_2\cdots x_nx_1x_2 \cdots x_nx_1 \cdots x_j}_{s_0(n-t+1)-1}.\] The monomial $T \not\in J_t^{s_0}$. We can see this by a degree argument because $J_t$ is generated by monomials in the $x$ variables of degree $n-t+1$. We make the crucial observation that the index $j$ of the last variable in $T$ has the property that $n-t+1 \leq j \leq n$. To see this, note that after $n-t+1$ steps in the sequence \eqref{word} we are at vertex $x_{n-t+1}$, after $2(n-t+1)$ steps in the sequence \eqref{word}, we are at the vertex $x_{n-2(t-1)} = x_{n-2t+2}$, after $3(n-t+1)$ steps, we are at $x_{n-3t+3}$, ..., and finally, after $(s_0-1)(n-t+1)$ steps, we are at vertex $x_{n-(s_0-1)(t-1)} = x_{n-s_0t+s_0+t-1}$. By our choice of $s_0$, $-s_0t+s_0 \leq -n+1$, so $n-s_0t+s_0+t-1 \leq t$. In fact, after $(s_0-1)$ steps of size $(n-t+1)$ in our sequence \eqref{word}, this is the first time we arrive at an index $\leq t$. At the same time, by our choice of $s_0$, we have $(s_0-1)(t-1) < n-1$, so we are at an index $\geq 1$. When constructing $T$, we go an additional $n-t$ steps in the sequence. This means that we arrive at an index between $n-t+1$ and $n$. We next show $J_t^{s_0}:\langle T \rangle = \langle z,x_1,\ldots,x_n \rangle$. Now $zT \in J_t^{s_0}$. To see this, note that $z$ is a minimal generator of $J_t$, and every $n-t+1$ consecutive variables in \eqref{word} is also a generator of $J_t$. Thus, the product of the first $(s_0-1)(n-t+1)$ elements of \eqref{word} is in $J_t^{s_0-1}$, and so $z \in J_t^{s_0}:\langle T \rangle $. Now take $x_i$ with $i \in \{1,\ldots, n\}$. To show $x_iT \in J_t^{s_0}$, take the first $s_0(n-t+1)-1$ variables in \eqref{word}, and insert $x_i$ after its first appearance, i.e., \[ x_1,x_2,\ldots,x_i,x_i,x_{i+1},\ldots,x_n,x_1,x_2,\ldots,x_n,x_1,x_2,\ldots,x_n,x_1,\ldots,x_j. \] Think of these variables as being placed around a circle. Starting at the second $x_i$, move around the circle, grouping $n-t+1$ variables together. Because we have $s_0(n-t+1)$ variables, we end up with $s_0$ groups. Because the index of $j$ is between $n-t+1$ and $n$, each group will consist of $n-t+1$ distinct variables, and thus, by Lemma \ref{generators}, when we multiply each group of $n-t+1$ distinct variables together, we have a generator of $J_t$. But this means that $x_iT \in J_t^{s_0}$ since $x_iT$ is expressed as a product of $s_0$ generators. Thus, $\langle z,x_0,\ldots,x_n \rangle \subseteq J_t^{s_0}:\langle T \rangle \subsetneq \langle 1 \rangle$, which completes the proof for the case $s_0$. Now suppose that $s > s_0$. Let $e = s-s_0$ and let $T$ be as above. We will show that $J_t^{s}:\langle z^eT \rangle = \langle z,x_1,\ldots,x_n \rangle$. By a degree argument $z^eT \not\in J_t^s$, but $z(z^eT) \in J_t^s$ because, as noted above, $T \in J_t^{s_0-1}$ and $z^{e+1} \in J_t^{e+1}$. Similarly, $x_iz^eT \in J_t^{s}$ because $z^e \in J_t^e$, and as above, $x_iT \in J_t^{s_0}$. Hence $J_t^{s}:\langle z^eT \rangle = \langle z,x_1,\ldots,x_n \rangle$. \end{proof} \subsection{An application} Corollary \ref{corstar} allows us to answer a question raised by Francisco, H\`a, and the third author \cite{FHVT}. We first recall some terminology. A {\it hypergraph} $\mathcal{H}$ is a pair of sets $\mathcal{H} = (\mathcal{X},\mathcal{E})$ where $\mathcal{X} = \{x_1,\ldots,x_n\}$ and $\mathcal{E}$ is a collection of subsets $\{E_1,\ldots,E_t\}$ with each $E_i \subseteq \mathcal{X}$. We call $\mathcal{H}$ a {\it simple} hypergraph if $|E_i| \geq 2$ for all $i$, and if $E_i \subseteq E_j$, then $i=j$. (When each $|E_i| = 2$, then $\mathcal{H}$ is a finite simple graph.) As in the case of graphs, we say a subset $W \subseteq \mathcal{X}$ is a {\it vertex cover} if $W \cap E \neq \emptyset$ for all $E \in \mathcal{E}$. In a manner analogous to the cover ideal, we can define the cover ideal of $\mathcal{H}$: \[J(\mathcal{H}) = \langle x_W ~|~ W = \{x_{i_1},\ldots,x_{i_t}\} \subseteq \mathcal{X} ~~\mbox{is a vertex cover} \rangle.\] A {\it colouring} of $\mathcal{H}$ is an assignment of a colour to each vertex of $\mathcal{X}$ so that no edge $E$ is mono-coloured, i.e., each edge must contain at least two vertices of different colours. The {\it chromatic number} of $\mathcal{H}$, denoted $\chi(\mathcal{H})$, is the least number of colours required to colour $\mathcal{H}$. The chromatic number provides a lower bound on the index of stability of $J(\mathcal{H})$. \begin{theorem}[{\cite[Corollary 4.9]{FHVT}}] For any finite simple hypergraph $\mathcal{H}$, \[\chi(\mathcal{H}) -1 \leq {\rm astab}(J(\mathcal{H})).\] \end{theorem} It was asked in \cite[Question 4.10]{FHVT} if for each $m \geq 1$, there exists a hypergraph $\mathcal{H}_m$ with $\chi(\mathcal{H}_m)-1+m \leq {\rm astab}(J(\mathcal{H}_m))$, that is, could the index of stability be arbitrarily larger than the chromatic number. Wolff \cite{W} showed that this is the case, even if $\mathcal{H}$ is a finite simple graph. Wolff's family of graphs requires $5m-1$ vertices. We can use Corollary \ref{corstar} to give another answer to this question which only requires $m+3$ vertices. \begin{theorem} Fix an $m \geq 1$, and let $\mathcal{H}_m = (\mathcal{X}_{m},\mathcal{E}_{m})$ where $\mathcal{X}_m = \{z,x_1,\ldots,x_{m+2}\}$ and $\mathcal{E}_m = \{\{z,x_i,x_j\} ~|~ 1 \leq i < j \leq {m+2}\}$. Then \[\chi(\mathcal{H}_m)-1+m \leq {\rm astab}(J(\mathcal{H}_m)).\] \end{theorem} \begin{proof} First, $\chi(\mathcal{H}_m) = 2$ because each $x_i$ can be assigned the same colour, and $z$ can be given a different colour. Note that $J(\mathcal{H}_m) = J_2(K_{1,m+2})$. By Corollary \ref{corstar}, ${\rm astab}(J(\mathcal{H}_m)) = {\rm astab}(J_2(K_{1,m+2})) \geq m+1 = (2-1)+m = \chi(\mathcal{H}_m)-1+m$. \end{proof} \section{Associated primes of Generalized cover ideals of trees} In this section we completely determine the associated primes of the ideals $J_t(\Gamma)^s$ when $\Gamma$ is a {\it tree}, that is, a graph with no induced cycles. Theorem \ref{maintheorem} will follow directly from this result. We begin by stating the main theorem of this section: \begin{theorem} \label{maintheoremtrees} Fix an integer $t \geq 1$ and let $\Gamma$ be a tree on $n$ vertices. Then for all $s \geq 1$, \[{\rm Ass}(J_t(\Gamma)^s) = \left. \left\{ P = \langle x_{i_0},x_{i_1},\ldots,x_{i_r} \rangle ~\right|~ \Gamma_P = K_{1,r} ~\mbox{with $ t \leq r \leq \min\{n,s(t-1)+1\}$}\right\}.\] \end{theorem} \noindent In other words, a prime is associated to $J_t(\Gamma)^s$ if and only if the corresponding induced subgraph in $\Gamma$ is a star of a particular size. We require the following lemma which can be found in \cite[Proposition 4.1]{JK}). This lemma will gives us some insight into the generators of $J_t(\Gamma)$. \begin{lemma}\label{specialvertex} For any tree $\Gamma$, there exists a vertex $x$ such that all, but possibly one, of its neighbours have degree $1$. \end{lemma} We fix some notation to be used throughout the remainder of this paper. Let $\Gamma$ be a tree, and let $x$ be the vertex of Lemma \ref{specialvertex} with neighbours $y_1,\ldots,y_d$. We can assume that $\deg y_1 = \cdots = \deg y_{d-1} = 1$ and $\deg y_d \geq 1$. Using this notation, we have: \begin{lemma} \label{structure} Let $\Gamma$ be a tree with partial $t$-cover ideal $J_t(\Gamma)$. If $m \in \mathcal{G}(J_t(\Gamma))$, then $m$ has one of the following forms: \begin{enumerate} \item[$(i)$] $m = y_{i_1}\cdots y_{i_{d-t+1}}m'$ \item[$(ii)$] $m = xm'$ \item[$(iii)$] $m = xy_dm'$ \end{enumerate} where in each case, $m'$ is not divisible by any of the variables $y_1, \dots, y_d$, $x$. \end{lemma} \begin{proof} By Lemma \ref{genspartial}, the minimal generators of $J_t(\Gamma)$ correspond to the minimal partial $t$-covers of $\Gamma$. The result will follow if we look at the corresponding statement for minimal partial $t$-covers of $\Gamma$. Let $W$ be a minimal partial $t$-cover of $\Gamma$. First, suppose that $x \not\in W$. By definition, $W$ must contain a subset $S \subseteq N(x)$ of size $|N(x)|-t+1 = d-t+1$. Because $N(x) = \{y_1,\ldots,y_d\}$, let us say that $S = \{y_{i_1},\ldots,y_{i_{d-t+1}}\}$. It now suffices to show that $W \setminus S$ does not contain any other neighbours of $x$. If $t=1$, then $S = N(x)$, so this is clear. So, suppose that $t \geq 2$, and suppose that there is some $y_j \in N(x)\cap (W \setminus S)$. There are two cases to consider: $j \neq d$ and $j = d$. If $j \neq d$, then Lemma \ref{specialvertex} gives $\deg y_j = 1$. We claim that $(W \setminus \{y_j\})$ is also a partial $t$-cover of $\Gamma$, thus contradicting the minimality of $W$. Indeed, take any vertex $z$ of $\Gamma$. Because $y_j$ is only adjacent to $x$, for any vertex $z \not\in \{y_j,x\}$, either $z$ is in $(W \setminus \{y_j\}) \subseteq W$ or all but perhaps $t-1$ of the neighbours of $z$ are in $(W \setminus \{y_j\}) \subseteq W$. We know that $x \not\in W$, but because $S \subseteq (W\setminus \{y_j\}) \subseteq W$, we know that all but perhaps $t-1$ of the neighbours of $x$ are in $(W \setminus \{y_j\})$. Finally, although $y_j \not\in W$, all but perhaps $t-1 \geq 2-1 =1$ of its neighbours belong to $W$. But since $y_j$ only has the neighbour $x$, $(W\setminus \{y_j\})$ is also a partial $t$-cover. If $j = d$, then we can simply repeat the above argument to show that $(W \setminus \{y_{i_1}\})$ (remove one the vertices of $S$, but keep $y_d$) creates a smaller partial $t$-cover. Now consider the case that $x \in W$. It suffices to show that $\{y_1,\ldots,y_{d-1}\} \cap W = \emptyset$. Then we will have the form $(ii)$ if $y_d \not\in W$, and the form $(iii)$ if $y_d \in W$. Suppose that $y_j \in \{y_1,\ldots,y_{d-1}\} \cap W$. We claim that $(W \setminus \{y_j\})$ would also be a partial $t$-cover. By Lemma \ref{specialvertex}, $\deg y_j = 1$, and $y_j$ is only adjacent to $x$. As argued above, for any vertex $z \not\in \{y_j,x\}$, either $z$ or all but perhaps $t-1$ of its neighbours will belong to $(W \setminus \{y_j\})$. The vertex $x$ is in $(W \setminus \{y_j\})$, and as for $y_j$, although $y_j \not\in (W \setminus \{y_j\})$, the unique edge containing $y_j$ is covered by $x$. So $(W \setminus \{y_j\})$ is a partial $t$-cover, contradicting the minimality of $W$. \end{proof} \begin{proof} (of Theorem \ref{maintheoremtrees}) Let $\mathcal{P}$ denote the set on the right. Lemma \ref{localization} and Corollary \ref{corstar} imply that every induced star graph of $\Gamma$ of the appropriate size will contribute an associated prime; more precisely, we already have $\mathcal{P} \subseteq {\rm Ass}(J_t(\Gamma)^s)$. It therefore suffices to show that if $P \in {\rm Ass}(J_t(\Gamma)^s)$, then $\Gamma_P$ is a star graph. Corollary \ref{corstar} and Lemma \ref{localization} then imply the condition on the size of the star graph, thus showing $P \in \mathcal{P}$. We let $J = J_t(\Gamma)$. If $P \in \operatorname{Ass}(J^s)$, by Lemma \ref{localization} we can assume that $\Gamma_P = \Gamma$ and by Lemma \ref{connected}, we can assume that $\Gamma$ is connected. Because $\Gamma$ is a tree, so is $\Gamma_P$. So, we can apply Lemma \ref{specialvertex}. That is, we can assume that there is a vertex $x$ with neighbours $y_1,\ldots, y_d$ such that $\deg y_1 = \cdots =\deg y_{d-1} = 1$, and $\deg y_d \geq 1$ in $\Gamma_P$. It suffices to show that $\deg y_d= 1$. Since $\Gamma_P$ is connected, this would mean $\Gamma_P = K_{1,d}$. So, suppose $y_d$ has a neighbour, say $w \neq x$. We thus have $P = \langle y_1,\ldots,y_d,x,w,\ldots \rangle$. We now want to build a contradiction from this information. Since $P \in \operatorname{Ass}(J^s)$, there exists a monomial $T\notin J^s$ such that $J^s:\langle T\rangle = P$. Because $w \in P$, \[Tw = m_1\cdots m_sM ~~\mbox{with $m_i \in \mathcal{G}(J)$}.\] By Lemma \ref{structure}, a generator of $J$ has one of three forms. Let's say that $a$ of $m_1,\ldots,m_s$ are of type $(i)$, $b$ of $m_1\ldots,m_s$ are of type $(ii)$, and $c$ are of type $(iii)$. We then have \[T = T'y_1^{e_1}y_2^{e_2}\cdots y_{d-1}^{e_{d-1}}y_d^{e_d+c}x^{b+c}\] where $e_1+\dots+e_d = (d-t+1)a$ and $a+b+c=s$. Without loss of generality we may assume that $e_1= \max\{e_1, \dots, e_{d-1}\}$. We now consider $Ty_1$. Since $y_1 \in P$, $Ty_1 \in J^s$, that is, \[Ty_1 = u_1\cdots u_sU ~~\mbox{with $u_j \in \mathcal{G}(J)$}.\] First, note that $y_1$ does not divide $U$, since if it did we would then have $T = u_1\cdots u_s(U/y_1) \in J^s$, a contradiction. Since $Ty_1 = T'y_1^{e_1+1}y_2^{e_2}\cdots y_{d-1}^{e_{d-1}}y_d^{e_d+c}x^{b+c}$, this means that (at least) $e_1+1$ of the generators $u_1,\ldots,u_s$ are divisible by $y_1$. We may assume that after reordering that these generators are $u_1,\ldots,u_{e_1+1}$. We next observe that $x$ also does not divide $U$. To see why, suppose that $U = xU'$. As noted above, $u_1 = y_1y_{i_2} \cdots y_{i_{d-t+1}}m$ for some monomial $m$ not divisible by $x$. Note that $(u_1x)/y_1 = xy_{i_2} \cdots y_{i_{d-t+1}}m$ will also be a non-minimal generator of $J$. This means that \begin{eqnarray*} Ty_1 &= &u_1\cdots u_sU = (y_1y_{i_2} \cdots y_{i_{d-t+1}}m)u_2\cdots u_s(xU') \\ & = & (xy_{i_2} \cdots y_{i_{d-t+1}}m)u_2\cdots u_s(y_1U'). \end{eqnarray*} If we now cancel $y_1$ from both sides, this implies that $T \in J^s$, a contradiction. So $x$ cannot divide $U$, and thus at least $b+c$ of $u_1,\ldots,u_s$ are divisible by $x$. By Lemma \ref{structure}, they cannot be among $u_1,\ldots,u_{e_1+1}$ since these are all divisible by $y_1$. Let us say that they are $u_{e_1+2},\ldots,u_{e_1+b+c+1}$. To summarize, we now have \begin{eqnarray*} Ty_1 &=& \underbrace{u_1\cdots u_{e_1+1}}_{\mbox{all divisible by $y_1$}}\cdot \underbrace{u_{e_1+2}\cdots u_{e_1+1+b+c}}_{\mbox{all divisible by $x$}}\cdots u_s U. \end{eqnarray*} We finish the proof by counting the degrees of the variables $y_2, \ldots, y_d$ in $Ty_1$. There are two cases to consider: ({\it Case 1}) there is a generator among $u_1, \ldots, u_s$ of type $(iii)$; and ({\it Case 2}) there is no generator among $u_1, \ldots, u_s$ of type $(iii)$. {\it Case 1:} Suppose there is some $u_j = xy_dm$. Then $y_d$ must divide every generator among $u_1, \dots, u_s$ of type $(i)$. To see why, suppose that there is some generator $u_r = y_1y_{i_2}\cdots y_{i_{d-t+1}}m'$ with $y_{i_\ell} \neq y_d$ for all $2\leq \ell \leq d-t+1$. \begin{eqnarray*} Ty_1 = u_1\cdots u_r \cdots u_j \cdots u_sU &=& u_1 \cdots (y_1y_{i_2}\cdots y_{i_{d-t+1}}m')\cdots (xy_dm) \cdots u_sU\\ &=& u_1 \cdots (xm') \cdots (y_{i_2}\cdots y_{i_{d-t+1}}y_dm)\cdots u_s (y_1U). \end{eqnarray*} Note that $xm', y_{i_2}\cdots y_{i_{d-t+1}}y_dm \in\mathcal{G}(J)$. If we cancel $y_1$ from both sides, we get $T \in J^s$, which is a contradiction. Similarly, suppose that there is some generator $u_r = y_{i_1}\dots y_{i_{d-t+1}}m'$ with $y_{i_\ell} \neq y_1, y_d$ for all $1\leq \ell \leq d$, and let $u_1 = y_1y_{k_2}\cdots y_{k_{d-t}}y_dm''$ (since $u_1$ is divisible by $y_1$, it must also be divisible by $y_d$ by above). Since $y_{i_\ell} \neq y_1$ for all $1\leq \ell \leq d-t+1$, there is some variable among $y_{i_1}, \ldots, y_{i_{d-t+1}}$ which does not divide $u_1$. Without loss of generality, assume that $y_{i_1}$ does not divide $u_1$. Then \begin{eqnarray*} Ty_1 &=& u_1\cdots u_r \cdots u_j \cdots u_sU\\ &=& (y_1y_{k_2}\cdots y_{k_{d-t}}y_dm'') \cdots (xy_dm) \cdots (y_{i_1}y_{i_2}\cdots y_{i_{d-t+1}}m')\cdots u_sU\\ &=& (y_{i_1}y_{k_2}\cdots y_{k_{d-t}}y_dm'') \cdots (y_{i_2}\cdots y_{i_{d-t+1}}y_dm) \cdots (xm')\cdots u_s (y_1U). \end{eqnarray*} The monomials $xm'$, $y_{i_2}\cdots y_{i_{d-t+1}}y_dm$, and $y_{i_1}y_{k_2}\cdots y_{k_{d-t}}y_dm''$ are generators of $J$, so if we cancel $y_1$ from both sides this leads to the contradiction $T \in J^s$. So if there is some $u_j = xy_dm$, then every generator of type $(i)$ among $u_1, \dots, u_s$ is divisible by $y_d$. Now consider the monomials $u_1, \ldots, u_{e_1+1}$. After relabeling, we may assume that $y_1, \ldots, y_j$ and $y_d$ divide all of $u_1, \ldots, u_{e_1+1}$, and that each of the remaining variables $y_{j+1}, \ldots, y_{d-1}$ do not divide at least one of the generators $u_1, \dots, u_{e_1+1}$. We now count the number of times that the variables $y_{j+1}, \ldots, y_{d-1}$ occur in the generators $u_1, \ldots, u_s$. Each of $u_1, \ldots u_{e_1+1}$ are divisible by exactly $d-t+1$ of the variables $y_1, \ldots, y_d$ including $y_1, \ldots, y_j$ and $y_d$. Therefore exactly $d-t+1-(j+1) = d-t-j$ of the variables $y_{j+1}, \ldots, y_{d-1}$ divide each of $u_1, \ldots, u_{e_1+1}$. In addition, the variables $y_{j+1}, \ldots, y_{d-1}$ may divide each of the monomials $u_{e_1+b+c+2}, \ldots, u_s$ (there are $s-(e_1+b+c+1) = a-e_1-1$ such monomials). Since $y_d$ divides every generator of type $(i)$ in the list $u_1, \ldots, u_s$, at most $d-t$ of the variables $y_{j+1}, \ldots, y_{d-1}$ divide each of the generators $u_{e_1+b+c+2}, \ldots, u_s$. In total, the number of times that the variables $y_{j+1}, \ldots, y_{d-1}$ divide the monomials $u_1, \ldots, u_s$ is at most \[ (d-t-j)(e_1+1) + (d-t)(a-e_1-1) = (d-t)a - j e_1 -j. \] On the other hand, since $T = T'y_1^{e_1}\cdots y_{d-1}^{e_{d-1}}y_d^{e_d+c}x^{b+c}$, the number of times that the variables $y_{j+1}, \ldots, y_{d-1}$ divide $T$ is at least \begin{eqnarray*} e_{j+1}+ \cdots + e_{d-1} &=& e_1+\cdots+e_d - (e_1+\cdots +e_j+e_d)\\ &=& (d-t+1)a - (e_1+\cdots + e_j + e_d). \end{eqnarray*} Since $e_1 = \max\{e_1, e_2, \ldots, e_{d-1}\}$ we have $e_1+e_2+\cdots+e_j\leq je_1$. So \begin{eqnarray*} (d-t+1)a - (e_1+\cdots + e_j + e_d) &\geq& (d-t+1)a - (je_1 + e_d)\\ &=& (d-t+1)a -je_1-e_d. \end{eqnarray*} And since $e_d$ is the number of times that the variable $y_d$ appears among the (square-free) monomials $m_1, \dots, m_a$ we have $a\geq e_d$. So \begin{eqnarray*} (d-t+1)a -je_1-e_d &\geq&(d-t+1)a - je_1 -a = (d-t)a -je_1. \end{eqnarray*} Since $j\geq 1$, this number is larger than the number of times that the variables $y_{j+1}, \ldots, y_{d-1}$ divide $u_1, \ldots , u_s$. Therefore, there must be some $y_k$ with $j+1\leq k\leq d-1$ which divides $U$. Let $U = y_kU'$. By assumption, there is some monomial among $u_1, \ldots, u_{e_1+1}$ which is not divisible by $y_k$. Without loss of generality, say $y_k \nmid u_1$. Then $u_1 = y_1y_{i_2}\cdots y_{i_{d-t+1}}m'$ for some monomial $m'$ with $y_{i_\ell}\neq y_k$ for all $2\leq \ell \leq d-t+1$. Then \begin{eqnarray*} Ty_1 = u_1\dots u_s U &=& (y_1y_{i_2}\cdots y_{i_{d-t+1}}m') u_2\cdots u_s(y_kU')\\ &=& (y_ky_{i_2}\cdots y_{i_{d-t+1}}m') u_2 \cdots u_s(y_1U'). \end{eqnarray*} Since $y_ky_{i_2}\cdots y_{i_{d-t+1}}m' \in \mathcal{G}(J)$ this implies that $T \in J^s$, which is a contradiction. So $w \not\in P$, and thus $\Gamma_P = K_{1,d}$, as desired. {\it Case 2:} Suppose that no generator among $u_1, \dots, u_s$ is of the form $xy_dm'$ (which implies $c=0$). Assume again that each of the variables $y_1, \dots, y_j$ with $1 \leq j < d$ divides each of the monomials $u_1, \dots, u_{e_1+1}$ and that the variables $y_{j+1}, \ldots, y_{d-1}$ do not. Note that $y_d$ may or may not divide every monomial in $u_1,\ldots,u_{e+1}$. We will count the variables $y_{j+1}, \dots, y_d$. We saw in the previous case that we arrive at a contradiction if we assume that the variable $y_d$ divides every minimal generator of type $(i)$ in the list $u_1, \dots, u_s$. Therefore we may assume that there is some monomial of type $(i)$ among $u_1, \ldots , u_{e_1+1}, u_{e_1+b+2}, \dots, u_s$ which is of type $(i)$ and which is not divisible by $y_d$. Now $(d-t+1-j)$ of the variables $y_{j+1}, \dots, y_d$ divide each of the monomials $u_1, \dots, u_{e_1+1}$. In addition, at most $(d-t+1)$ of the variables $y_{j+1}, \dots, y_d$ divide each of the monomials $u_{e_1+b+2}, \dots, u_s$. In total the number of times that the variables $y_{j+1}, \dots, y_d$ divide the monomials $u_1, \dots, u_s$ is at most \[ (d-t+1-j)(e_1+1)+(d-t+1)(s-(b+e_1+1)) = (d-t+1)a -je_1 -j.\\ \] On the other hand, since $T = T'y_1^{e_1}\cdots y_{d-1}^{e_{d-1}}y_d^{e_d}x^{b}$ (because $c=0$ in this case), the number of times the variables $y_{j+1}, \dots, y_d$ divide $Ty_1$ is at least \begin{eqnarray*} e_{j+1}+\cdots+ e_d &=& e_1+\cdots+e_d - (e_1+\cdots +e_j)\\ &=& (d-t+1)a -(e_1+\cdots+e_j)\\ &\geq& (d-t+1)a -je_1 \end{eqnarray*} because $e_1 = \max\{e_1, \dots, e_{d-1}\}$. Since $j\geq1$, this number is strictly greater than the number of times that $y_{j+1}, \ldots, y_d$ divide the monomials $u_1, \ldots, u_s$. Therefore there is some $y_k$ with $j+1\leq k\leq d$ which divides $U$. If $k \neq d$, then we know that there is some monomial among $u_1, \ldots, u_{e_1+1}$ which is not divisible by $y_k$. Without loss of generality we may assume that $y_k$ does not divide $u_1 = y_1y_{i_2}\cdots y_{i_{d-t+1}}m'$. Then \begin{eqnarray*} Ty_1 = u_1\cdots u_s U &=& (y_1y_{i_2}\cdots y_{i_{d-t+1}}m') u_2\cdots u_s(y_kU')\\ &=& (y_ky_{i_2}\cdots y_{i_{d-t+1}}m') u_2 \cdots u_s(y_1U'). \end{eqnarray*} Since $y_ky_{i_2}\cdots y_{i_{d-t+1}}m' \in \mathcal{G}(J)$, this implies that $T \in J^s$ which is a contradiction. Finally, assume that none of $y_{j+1}, \dots, y_{d-1}$ divide $U$. Then $y_d$ must divide $U$. Let $U = y_dU'$. If there is some monomial among $u_1, \dots, u_{e_1+1}$ which is not divisible by $y_d$ then we arrive at a contradiction as above. If $y_d$ divides each of $u_1, \dots, u_{e_1+1}$, then there is some monomial in the list $u_{e_1+b+2}, \dots, u_s$ which is not divisible by $y_d$. Without loss of generality, assume $u_s$ is not divisible by $y_d$. So $u_s = y_{k_1}\cdots y_{k_{d-t+1}}m$, where $y_{k_\ell} \neq y_d$ for all $1 \leq \ell \leq d-t+1$ and $u_1 = y_1 y_{i_2}\cdots y_{i_{d-t}}y_dm'$. Since $y_d$ divides $u_1$ and does not divide $u_s$ there is at least one of the variables $y_{k_1}, \ldots, y_{k_{d-t+1}}$ which does not divide $u_1$. Assume that $y_{k_1}$ does not divide $u_1$. Then \begin{eqnarray*} Ty_1 &=& u_1\cdots u_sU\\ &=& (y_1 y_{i_2}\cdots y_{i_{d-t}}y_dm')u_2\cdots u_{s-1}(y_{k_1}y_{k_2}\cdots y_{k_{d-t+1}}m)(y_dU')\\ &=&(y_{k_1} y_{i_2}\cdots y_{i_{d-t}}y_dm')u_2\cdots u_{s-1}(y_{k_2}\cdots y_{k_{d-t+1}}y_dm)(y_1U'). \end{eqnarray*} Since $y_{k_1} y_{i_2}\cdots y_{i_{d-t}}y_dm'$ and $y_{k_2}\cdots y_{k_{d-t+1}}y_dm$ are also minimal generators of $J$, this implies that $T$ is an element of $J^s$ which is a contradiction. Therefore the associated prime $P$ cannot be of the form $P = \langle y_1, \dots, y_d, x, w, \ldots \rangle$. In other words, $\deg (y_d) = 1$, so $\Gamma_P = K_{1, d}$ is a star graph as desired. \end{proof} We can now prove Theorem \ref{maintheorem}. \begin{proof}(of Theorem \ref{maintheorem}) The persistence property is immediate from our description of the sets ${\rm Ass}(J_t(\Gamma)^s)$ in Theorem \ref{maintheoremtrees}. When $t=1$, ${\rm astab}(J_1(\Gamma))=1$ since $\Gamma$ is bipartite. So the result follows from \cite{GRV}. When $t \geq 2$, let $x$ be a vertex with $\deg x = \Delta(\Gamma)$, i.e., a vertex of maximal degree. Let $P = \{x\} \cup N(x)$. Then $\Gamma_P = K_{1,\Delta(\Gamma)}$. If we abuse notation, and let $P$ also denote the ideal generated by the variables corresponding to the vertices in $P$, then $P \in {\rm Ass}(J_t(\Gamma)^s)$ if and only if $s(t-1) \geq \Delta(\Gamma)-1$. So ${\rm astab}(J_t(\Gamma)) \geq \min\{s ~|~ s(t-1) \geq \Delta(\Gamma)-1\}$. Let $s_0 = \min\{s ~|~ s(t-1) \geq \Delta(\Gamma)-1\}$ and suppose that ${\rm astab}(J_t(\Gamma)) > s_0$. Because $J_t(\Gamma)$ has the persistence property, that means that there is a $P \in {\rm Ass}(J_t(\Gamma)^s) \setminus {\rm Ass}(J_t(\Gamma)^{s_0})$ with $s > s_0$. We can assume $s$ is the smallest such integer with this property. By Theorem \ref{maintheoremtrees}, $\Gamma_P = K_{1,r}$, and by Theorem \ref{maintheoremstar}, we must have $s(t-1) \geq r-1$. Since $P \not\in {\rm Ass}(J_t(\Gamma)^{s_0})$, we must have $s_0(t-1) \not\geq r-1$. But this means that $r > \Delta(\Gamma)$, which implies that $\Gamma$ has a vertex of degree greater than $\Delta(\Gamma)$, a contradiction. \end{proof}

https://arxiv.org/abs/2207.04701

Spectral radius and edge-disjoint spanning trees

The spanning tree packing number of a graph $G$, denoted by $\tau(G)$, is the maximum number of edge-disjoint spanning trees contained in $G$. The study of $\tau(G)$ is one of the classic problems in graph theory. Cioabă and Wong initiated to investigate $\tau(G)$ from spectral perspectives in 2012 and since then, $\tau(G)$ has been well studied using the second largest eigenvalue of the adjacency matrix in the past decade. In this paper, we further extend the results in terms of the number of edges and the spectral radius, respectively; and prove tight sufficient conditions to guarantee $\tau(G)\geq k$ with extremal graphs characterized. Moreover, we confirm a conjecture of Ning, Lu and Wang on characterizing graphs with the maximum spectral radius among all graphs with a given order as well as fixed minimum degree and fixed edge connectivity. Our results have important applications in rigidity and nowhere-zero flows. We conclude with some open problems in the end.

\section{Introduction} In this paper, we only consider simply graphs unless otherwise stated and $k$ always denotes a positive integer. The study of edge-disjoint spanning trees has been shown to be very important to graphs and has many applications in fault-tolerance networks as well as network reliability~\cite{Cunningham,Hobbs91}. Thus it is quite interesting to explore how many edge-disjoint spanning trees in a given graph. The \textit{spanning tree packing number} (or simply \textit{STP number}) of a graph $G$, denoted by $\tau(G)$, is the maximum number of edge-disjoint spanning trees contained in $G$. Nash-williams~\cite{Nash-Williams} and Tutte~\cite{Tutte} independently discovered a fundamental theorem that characterizes graphs $G$ with $\tau(G)\ge k$ (See Theorem~\ref{lem::2.9} in the next section). The \textit{edge connectivity} of $G$, denote by $\kappa'(G)$, is the minimum cardinality of an edge cut of $G$. It is known that $\kappa'(G)$ and $\tau(G)$ are closely related. In fact, the fundamental theorem of Nash-williams~\cite{Nash-Williams} and Tutte~\cite{Tutte} implies that if $\kappa'(G)\ge 2k$, then $\tau(G)\ge k$. For more about the spanning tree packing number, we refer readers to the survey~\cite{Palmer01} by Palmer. \medskip The number of spanning trees has also been well studied from spectral perspectives. For a graph $G$, let $A(G)$ denote the adjacency matrix of $G$ and let $\lambda_i(G)$ denote the $i$th largest eigenvalue of $A(G)$. In particular, the largest eigenvalue of $A(G)$ is called the \textit{spectral radius} of $G$ and is denoted by $\rho(G)$. Denote by $D(G)$ the diagonal matrix of vertex degrees of $G$. The \textit{Laplacian matrix} of $G$ is defined as $L(G)=D(G)-A(G)$. The Laplacian matrix is positive semidefinite and we order its eigenvalues as $\mu_{1}\geq\mu_{2}\geq\ldots\geq\mu_{n-1}\geq\mu_{n}=0$. The well-known Matrix-Tree Theorem of Kirchhoff \cite{Kirchhoff} indicates that the number of spanning trees (not necessarily edge-disjoint) of a graph $G$ with $n$ labelled vertices is $\frac{\prod_{i=1}^{n-1}\mu_{i}}{n}$. For edge-disjoint spanning trees, Seymour proposed the following problem in private communication to Cioab\u{a} as mentioned in~\cite{Wong}. \begin{prob}\label{prob1} Let $G$ be a connected graph. Determine the relationship between $\tau(G)$ and the eigenvalues of $G$. \end{prob} Inspired by the Kirchhoff's Matrix-Tree Theorem and Problem~\ref{prob1}, Cioab\u{a} and Wong \cite{Wong} started to study the spanning tree packing number via the second largest eigenvalue of the adjacency matrix. They proved that for a $d$-regular connected graph $G$, $\tau(G)\geq k$ if $\lambda_{2}(G)<d-\frac{2(2k-1)}{d+1}$ for $d\geq 2k\geq 4$, and further conjectured that the sufficient condition can be improved to $\lambda_{2}(G)<d-\frac{2k-1}{d+1}$. In the same paper, they did the preliminary work of this conjecture for $k=2,3$ and gave examples to show the bound is best possible. Later, Gu et al.~\cite{Gu-Lai} extended the conjecture to graphs that are not necessarily regular, that is, $\tau(G)\geq k$ if $\lambda_2(G)< \delta-\frac{2k-1}{\delta+1}$ for $\delta\geq 2k\geq 4$. They confirmed the conjecture for $k=2,3$ and obtained a partial result that $\lambda_2(G)< \delta-\frac{3k-1}{\delta+1}$ suffices. This conjecture was completely settled in 2014 by Liu et al.~\cite{Liu-Hong} who proved a stronger result, which also implied the truth of the conjecture of Cioab\u{a} and Wong \cite{Wong}. Most recently, the result in~\cite{Liu-Hong} has been shown to be essentially best possible in \cite{coppww22} by constructing extremal graphs. On the other hand, the result was extended to a fractional version by Hong et al.~\cite{HGLL16}, and improved by Liu, Lai and Tian \cite{Liu-Lai} with a Moore function. Motivated by Problem~\ref{prob1} and the above results, we study the spanning tree packing number by means of the spectral radius of graphs. We first investigate an extremal result for $\tau(G)\ge k$. Let $e(G)$ denote the number of edges in $G$. \begin{thm}\label{thm::edgenumber} Let $G$ be a connected graph with minimum degree $\delta\geq 2k$ and order $n\geq 2\delta+2$. If $e(G)\geq {\delta+1\choose2}+{n-\delta-1\choose 2} +k$, then $\tau(G)\geq k$. \end{thm} The condition in Theorem~\ref{thm::edgenumber} is tight. Denote by $K_n$ the complete graph on $n$ vertices, and $\mathcal{G}_{n,n_1}^{i}$ the set of graphs obtained from $K_{n_1}\cup K_{n-n_{1}}$ by adding $i$ edges between $K_{n_1}$ and $K_{n-n_{1}}$. Notice that any graph $G$ in $\mathcal{G}_{n,\delta+1}^{k-1}$ has exactly ${\delta+1\choose2} + {n-\delta-1\choose 2} +k-1$ edges but $\tau(G)<k$. \medskip We then focus on a spectral analogue. The corresponding spectral problem is much harder. Let $B_{n,\delta+1}^{i}$ be a graph obtained from $K_{\delta+1}\cup K_{n-\delta-1}$ by adding $i$ edges joining a vertex in $K_{\delta+1}$ and $i$ vertices in $K_{n-\delta-1}$. We succeed in discovering a sufficient condition for $\tau(G)\geq k$ via the spectral radius, and characterize the unique spectral extremal graph $B_{n,\delta+1}^{k-1}$ among the structural extremal graph family $\mathcal{G}_{n,\delta+1}^{k-1}$. \begin{thm}\label{thm::1.2} Let $k\geq 2$, and let $G$ be a connected graph with minimum degree $\delta\geq 2k$ and order $n\geq 2\delta+3$. If $\rho(G)\geq \rho(B_{n,\delta+1}^{k-1})$, then $\tau(G)\geq k$ unless $G\cong B_{n,\delta+1}^{k-1}$. \end{thm} Our proofs are novel and as an application, we will use similar proof techniques to settle a conjecture of Ning, Lu and Wang \cite{Ning}. Let $\mathcal{A}_{n}^{\kappa',\delta}$ be a set of graphs of order $n$ with minimum degree $\delta$ and edge connectivity $\kappa'$. For $0\leq \kappa'\leq 3$, Ning, Lu and Wang \cite{Ning} determined that the unique extremal graph with the maximum spectral radius is $B_{n,\delta+1}^{\kappa'}$, and they proposed the following conjecture. \begin{conj}[Ning, Lu and Wang~\cite{Ning}]\label{conj1} For $4\leq \kappa'<\delta$, $B_{n,\delta+1}^{\kappa'}$ is the graph with the maximum spectral radius in $\mathcal{A}_{n}^{\kappa',\delta}$. \end{conj} We confirm this conjecture for $n\ge 2\delta +4$ as described below. \begin{thm}\label{thm::1.3} Let $G\in \mathcal{A}_{n}^{\kappa',\delta}$ where $4\leq\kappa'<\delta$ and $n\geq 2\delta+4$. Then $\rho(G)\leq \rho(B_{n,\delta+1}^{\kappa'})$, with equality if and only if $G\cong B_{n,\delta+1}^{\kappa'}$. \end{thm} The proofs of Theorems~\ref{thm::edgenumber} and \ref{thm::1.2} as well as the one of Theorem~\ref{thm::1.3} will be presented in the next two sections, respectively. Since edge-disjoint spanning trees have many applications, we will list some of them in Section~\ref{sec::app}, including rigidity and nowhere-zero flows. Some concluding remarks will be made in Section~\ref{sec::remarks}. \section{Proofs of Theorems~\ref{thm::edgenumber} and \ref{thm::1.2}} In this section, we present the proofs of Theorems~\ref{thm::edgenumber} and \ref{thm::1.2}. We first list several lemmas that will be used in the sequel. The following sharp upper bound on the spectral radius was obtained by Hong, Shu and Fang ~\cite{HSF} and Nikiforov~\cite{V.N}, independently. \begin{lem}[\cite{HSF,V.N}]\label{lem::2.1} Let $G$ be a graph on $n$ vertices and $m$ edges with minimum degree $\delta\geq 1$. Then $$\rho(G) \leq \frac{\delta-1}{2}+\sqrt{2 m-n \delta+\frac{(\delta+1)^{2}}{4}},$$ with equality if and only if $G$ is either a $\delta$-regular graph or a bidegreed graph in which each vertex is of degree either $\delta$ or $n-1$. \end{lem} \begin{lem}[\cite{HSF,V.N}]\label{lem::2.2} For nonnegative integers $p$ and $q$ with $2q \leq p(p-1)$ and $0 \leq x \leq p-1$, the function $f(x)=(x-1) / 2+\sqrt{2 q-p x+(1+x)^{2} / 4}$ is decreasing with respect to $x$. \end{lem} Recall that $\mathcal{G}_{n,n_1}^{i}$ is the set of graphs obtained from $K_{n_1}\cup K_{n-n_{1}}$ by adding $i$ edges between $K_{n_1}$ and $K_{n-n_{1}}$. \begin{lem}\label{lem::2.3} Let $k\geq 2$ and let $G\in \mathcal{G}_{n,\delta+1}^{k-1}$ where $n\geq 2\delta+3$ and $\delta\geq 2k$. Then $$n-\delta-2< \rho(G)< n-\delta-1.$$ \end{lem} \begin{proof} Note that $G$ contains $K_{\delta+1}\cup K_{n-\delta-1}$ as a proper spanning subgraph and $n\geq 2\delta+3$. Then $\rho(G)>\rho(K_{\delta+1}\cup K_{n-\delta-1})= n-\delta-2$. Since $\delta\geq 2k\geq 4$, it follows that \begin{equation*} \begin{aligned} e(G)&={\delta+1\choose 2}+{n-\delta-1\choose 2}+k-1\\ &\leq {\delta+1\choose 2}+{n-\delta-1\choose 2}+\frac{\delta}{2}-1\\ &=\frac{n^2}{2}-\frac{(2\delta+3)n}{2}+\delta^2+\frac{5\delta}{2}. \end{aligned} \end{equation*} Combining this with Lemmas \ref{lem::2.1} and \ref{lem::2.2}, we have \begin{equation*} \begin{aligned} \rho(G)&\leq \frac{\delta-1}{2}+\sqrt{n^2 + (-3\delta - 3)n + \frac{9\delta^2}{4}+ \frac{11\delta}{2}+\frac{1}{4}}\\ &= \frac{\delta-1}{2}+\sqrt{\left(n-\frac{3\delta}{2}-\frac{1}{2}\right)^2-2(n-2\delta)}\\ &< \frac{\delta-1}{2}+ \left(n-\frac{3\delta}{2}-\frac{1}{2}\right)~~(\mbox{since $n\geq 2\delta\!+\!3$})\\ &=n-\delta-1. \end{aligned} \end{equation*} This completes the proof. \end{proof} It is known that if $G$ is connected, then $A(G)$ is irreducible. By the Perron-Frobenius theorem(cf. \cite[Section 8.8]{C.G-2}), the Perron vector $x$ is a positive eigenvector of $A(G)$ with respect to $\rho(G)$. For any $v\in V(G)$, let $N_{G}(v)$ and $d_G(v)$ be the neighborhood and degree of $v$ in $G$, respectively. Let $N_{G}[v]=N_{G}(v)\cup\{v\}$. \begin{lem}[\cite{H.L-1}]\label{lem::2.4} Let $G$ be a connected graph, and let $u,v$ be two vertices of $G$. Suppose that $v_{1},v_{2},\ldots,v_{s}\in N_{G}(v)\backslash N_{G}(u)$ with $s\geq 1$, and $G^*$ is the graph obtained from $G$ by deleting the edges $vv_{i}$ and adding the edges $uv_{i}$ for $1\leq i\leq s$. Let $x$ be the Perron vector of $A(G)$. If $x_{u}\geq x_{v}$, then $\rho(G)<\rho(G^*)$. \end{lem} For any vertex $v\in V(G)$ and any subset $S\subseteq V(G)$, let $N_{S}(v)=N_{G}(v)\cap S$ and $d_{S}(v)=|N_{S}(v)|$. \begin{lem}\label{lem::2.5} Let $k\geq 2$ and let $G\in \mathcal{G}_{n,\delta+1}^{k-1}$ where $n\geq 2\delta+3$ and $\delta\geq 2k$. Then $\rho(G)\leq \rho(B_{n,\delta+1}^{k-1})$, with equality if and only if $G\cong B_{n,\delta+1}^{k-1}$. \end{lem} \begin{proof} Suppose that $G'$ is the graph that attains the maximum spectral radius in $\mathcal{G}_{n,\delta+1}^{k-1}$. For every $G\in \mathcal{G}_{n,\delta+1}^{k-1}$, we have \begin{equation}\label{equ::1} \begin{aligned} \rho(G)\leq \rho(G'). \end{aligned} \end{equation} We partition $V(G')$ into $V_1\cup V_2$ with $V_1=V(K_{\delta+1})=\{u_1,u_2,\ldots,u_{\delta+1}\}$ and $V_2=V(K_{n-\delta-1})=\{v_{1},v_2,\ldots,v_{n-\delta-1}\}$. Let $x$ be the Perron vector of $A(G')$, and let $\rho'=\rho(G')$. Without loss of generality, we assume that $x_{u_{i+1}}\leq x_{u_{i}}$ and $x_{v_{j+1}}\leq x_{v_{j}}$ for $1\leq i\leq \delta$ and $1\leq j\leq n-\delta-2$. We assert that $N_{G'}(u_{i+1})\subseteq N_{G'}[u_i]$ and $N_{G'}(v_{j+1})\subseteq N_{G'}[v_j]$ for $1\leq i\leq \delta$ and $1\leq j\leq n-\delta-2$. If not, suppose that there exist $i,j$ with $i<j$ such that $N_{G'}(u_j)\nsubseteq N_{G'}[u_i]$. Let $w\in N_{G'}(u_j)\backslash N_{G'}[u_i]$ and $G^{*}=G'-wu_{j}+wu_{i}$. Note that $x_{u_{j}}\leq x_{u_{i}}$. Then $\rho(G^{*})>\rho(G')$ by Lemma \ref{lem::2.4}, which contradicts the maximality of $\rho'$. This implies that $N_{G'}(u_{i+1})\subseteq N_{G'}[u_i]$ for $1\leq i\leq \delta$. Similarly, we can deduce that $N_{G'}(v_{j+1})\subseteq N_{G'}[v_j]$ for $1\leq j\leq n-\delta-2$. Furthermore, we have $N_{V_2}(u_{i+1})\subseteq N_{V_2}(u_i)$ and $N_{V_1}(v_{j+1})\subseteq N_{V_1}(v_j)$ for $1\leq i\leq \delta$ and $1\leq j\leq n-\delta-2$. Let $d_{V_2}(u_1) =r$, $d_{V_2}(u_2)=t$, and $d_{V_1}(v_1)=s$. Again by the maximality of $\rho'$ and Lemma \ref{lem::2.4}, we have $N_{V_2}(u_1)=\{v_1,v_2,\dots,v_r\}$, $N_{V_2}(u_2)=\{v_1,v_2,\dots,v_t\}$ and $N_{V_1}(v_1)=\{u_1,u_2,\dots,u_s\}$. If $r=k-1$ or $s=1$, then $G'\cong B_{n,\delta+1}^{k-1}$. Combining this with (\ref{equ::1}), we can deduce that $\rho(G)\leq\rho(B_{n,\delta+1}^{k-1})$, with equality if and only if $G\cong B_{n,\delta+1}^{k-1}$, as required. Next we consider $r\leq k-2$ and $s\geq 2$ in the following. Note that $x_{v_i}=x_{v_{r+1}}$ for $r+1\leq i\leq n-\delta-1$, $x_{v_j}\geq x_{v_{r+1}}$ for $2\leq j\leq r$. Then, by $A(G')x=\rho' x$, we have \begin{eqnarray*} \rho' x_{v_{r+1}} =x_{v_1}+\sum_{2\leq i\leq r}x_{v_i}+(n-\delta-r-2)x_{v_{r+1}}\geq x_{v_1}+(n-\delta-3)x_{v_{r+1}}. \end{eqnarray*} Note that $G'\in\mathcal{G}_{n,\delta+1}^{k-1}$. Then $\rho'> n-\delta-2$ by Lemma \ref{lem::2.3}, and hence \begin{eqnarray} x_{v_{r+1}}\geq\frac{x_{v_1}}{\rho'-(n-\delta-3)}. \label{equ::2} \end{eqnarray} Assume that $E_1=\{u_1v_{i}|~r+1\leq i\leq k-1\}$ and $E_2=\{u_{i}v_j\in G' |~ 2\leq i\leq s, 1\leq j\leq t\}$. Let $G''=G'-E_2+E_1$. Clearly, $G''\cong B_{n,\delta+1}^{k-1}$. Let $y$ be the Perron vector of $A(G'')$, and let $\rho''=\rho(G'')$. By $A(G'')y=\rho'' y$, we have \begin{eqnarray*} &\rho''y_{u_2}=y_{u_1}+(\delta-1)y_{u_2}. \end{eqnarray*} Note that $G''\in\mathcal{G}_{n,\delta+1}^{k-1}$ and $n\geq2\delta+3$. Again by Lemma \ref{lem::2.3}, $\rho''> n-\delta-2> \delta-1$, and hence \begin{eqnarray} y_{u_2}=\frac{y_{u_1}}{\rho''-\delta+1}. \label{equ::3} \end{eqnarray} Since $G',G''\in \mathcal{G}_{n,\delta+1}^{k-1}$, by Lemma \ref{lem::2.3}, $\rho''\!-\!\rho'>-1$. Note that $x_{u_1}\geq x_{u_i}$ for $2\leq i\leq s$, $x_{v_1}\geq x_{v_j}$ for $2\leq j\leq t$. Combining this with (\ref{equ::2}), (\ref{equ::3}) and $\rho''\!-\!\rho'>-1$, we have \begin{equation*} \begin{aligned} y^{T}(\rho''-\rho')x &=y^{T}(A(G'')-A(G'))x\\ &=\sum_{u_1v_i\in E_1}(x_{u_{1}}y_{v_{i}}\!+\!x_{v_{i}}y_{u_{1}})\! -\sum_{u_iv_j\in E_2}\!(x_{u_{i}}y_{v_{j}}\!+\!x_{v_{j}}y_{u_{i}})\\ &\geq(k\!-\!1-r)(x_{u_1}y_{v_1}+x_{v_{r+1}}y_{u_1}-x_{u_1}y_{v_1}-x_{v_1}y_{u_2}) ~~(\mbox{since $r\leq k-2$})\\ &=(k\!-\!1-r)(x_{v_{r+1}}y_{u_1}-x_{v_1}y_{u_2})\\ &\geq(k\!-\!1-r)x_{v_{1}}y_{u_1}\left(\frac{1}{\rho'-(n-\delta-3)}-\frac{1}{\rho''-\delta+1}\right) ~~(\mbox{by (\ref{equ::2}) and (\ref{equ::3})})\\ &=\frac{(k\!-\!1-r)x_{v_{1}}y_{u_1}}{(\rho'-(n-\delta-3))(\rho''-\delta+1)}(\rho''-\rho'+n-2\delta-2)\\ &>0~~(\mbox{since $k\geq r\!+\!2$, $\rho'>n\!-\!\delta\!-\!2$, $\rho''>\delta\!-\!1$ and $n\geq 2\delta\!+\!3$}), \end{aligned} \end{equation*} and hence $\rho''>\rho'$, which contradicts the maximality of $\rho'$. This completes the proof. \end{proof} For $X,Y\subseteq V(G)$, we denote by $E_{G}(X,Y)$ the set of edges with one endpoint in $X$ and one endpoint in $Y$, and $e_{G}(X,Y)=|E_{G}(X,Y)|$. \begin{lem}\label{lem::2.6} Let $k\geq 2$, and let $G\in \mathcal{G}_{n,a}^{k-1}$ where $n\geq 2a$, $a\geq \delta+2$ and $\delta\geq 2k$. Then $\rho(G)<\rho(B_{n,\delta+1}^{k-1}).$ \end{lem} \begin{proof} Let $G\in \mathcal{G}_{n,a}^{k-1}$. Then $V(G)=V(K_a)\cup V(K_{n-a})$ and $e_{G}(V(K_a),V(K_{n-a}))=k-1$. We partition $V(K_a)$ into $A_1\cup A_{2}$ and $V(K_{n-a})$ into $B_1\cup B_{2}$ such that $|A_1|=|B_1|=\delta+1$, $e_{G}(A_2, V(K_{n-a}))=0$ and $e_{G}(B_2, V(K_{a}))=0$. Since $n\geq 2a$, $a\geq \delta+2$ and $\delta\geq 2k$, it follows that $|A_2|=|V(K_a)|-|A_1|\geq 1$ and $|B_2|=|V(K_{n-a})|-|B_1|\geq 1$. Let $x$ be the Perron vector of $A(G)$. If $\sum_{v\in A_1}x_{v}<\sum_{v\in V(K_{n-a})}x_{v}$, then let $G'$ be a graph obtained from $G$ by deleting all edges between $A_1$ and $A_2$, and adding all possible edges between $A_2$ and $V(K_{n-a})$. Clearly, $G'\in \mathcal{G}_{n,\delta+1}^{k-1}$ and \begin{equation*} \begin{aligned} \rho(G')-\rho(G)&\geq x^{T}(A(G')-A(G))x\\ &=2\sum_{v\in A_2}x_{v}\left(\sum_{v\in V(K_{n-a})}x_{v}-\sum_{v\in A_1}x_{v}\right)\\ &>0, \end{aligned} \end{equation*} and hence $\rho(G')>\rho(G)$. Combining this with Lemma \ref{lem::2.5}, we have $\rho(B_{n,\delta+1}^{k-1})\geq\rho(G')>\rho(G)$, as required. If $\sum_{v\in A_1}x_{v}\geq\sum_{v\in V(K_{n-a})}x_{v}$, since $\sum_{v\in A_2}x_{v}>0$ and $\sum_{v\in B_2}x_{v}>0$, we have $$\sum_{v\in V(K_{a})}x_{v}>\sum_{v\in A_1}x_{v}\geq\sum_{v\in V(K_{n-a})}x_{v}>\sum_{v\in B_1}x_{v}.$$ Let $G''$ be a graph obtained from $G$ by deleting all edges between $B_1$ and $B_2$, and adding all possible edges between $V(K_{a})$ and $B_2$. One can verify that $G''\in \mathcal{G}_{n,\delta+1}^{k-1}$ and \begin{equation*} \begin{aligned} \rho(G'')-\rho(G)&\geq x^{T}(A(G'')-A(G))x\\ &=2\sum_{v\in B_2}x_{v}\left(\sum_{v\in V(K_a)}x_{v}-\sum_{v\in B_1}x_{v}\right)\\ &>0, \end{aligned} \end{equation*} and hence $\rho(G'')>\rho(G)$. Similarly, $\rho(B_{n,\delta+1}^{k-1}) \geq\rho(G'') >\rho(G)$. This completes the proof. \end{proof} \begin{lem}\label{lem::2.7} Let $a$ and $b$ be two positive integers. If $a\geq b$, then $${a\choose 2}+{b\choose 2}< {a+1\choose 2}+{b-1\choose 2}.$$ \end{lem} \begin{proof} Note that $a\geq b$. Then \begin{equation*} \begin{aligned} &{a+1\choose 2}+{b-1\choose 2}\!-\!{a\choose 2}\!-\!{b\choose 2}=a-b+1>0. \end{aligned} \end{equation*} Thus the result follows.\end{proof} Denote by $\partial_{G}(X)=E_{G}(X,V(G)-X)$. \begin{lem}[Lemma~2.8 of \cite{Gu-Lai}]\label{lem::2.8} Let $G$ be a graph with minimum degree $\delta$ and $U$ be a non-empty proper subset of $V(G)$. If $|\partial_{G}(U)|\leq \delta-1$, then $|U|\geq \delta+1$. \end{lem} For any partition $\pi$ of $V(G)$, let $E_{G}(\pi)$ denote the set of edges in $G$ whose endpoints lie in different parts of $\pi$, and let $e_{G}(\pi)=|E_{G}(\pi)|$. A part is \textit{trivial} if it contains a single vertex. The following fundamental theorem on spanning tree packing number of a graph was established by Nash-Williams \cite{Nash-Williams} and Tutte \cite{Tutte}, independently. \begin{thm}[Nash-williams~\cite{Nash-Williams} and Tutte~\cite{Tutte}] \label{lem::2.9} Let $G$ be a connected graph. Then $\tau(G)\geq k$ if and only if for any partition $\pi$ of $V(G)$, $$e_{G}(\pi)\geq k(t-1),$$ where $t$ is the number of parts in the partition $\pi$. \end{thm} For $X\subseteq V(G)$, let $G[X]$ be the subgraph of $G$ induced by $X$, and let $e_{G}(X)$ be the number of edges in $G[X]$. Now, we shall give the proofs of Theorems \ref{thm::edgenumber} and \ref{thm::1.2}. \begin{proof}[\bf Proof of Theorem \ref{thm::edgenumber}] Assume to the contrary that $\tau(G)\leq k-1$. By Theorem \ref{lem::2.9}, there exists a partition $\pi$ of $V(G)$ with $t_1$ trivial parts $v_1,v_2,\ldots, v_{t_1}$ and $t_2$ nontrivial parts $V_1,V_2,\ldots,V_{t_2}$ such that \begin{equation}\label{equ::4} \begin{aligned} e_{G}(\pi)\leq k(t-1)-1, \end{aligned} \end{equation} where $t=t_1+t_2$. If the partition $\pi$ contains only one part, then $e_{G}(\pi)\leq -1$ by (\ref{equ::4}), which is impossible because $e_{G}(\pi)=0$. This implies that $t\geq 2$. We assert that $t_2\geq 2$. If not, suppose that $t_2\leq 1$. Then $t_1\geq t-1$. Note that $d_{G}(v_i)\geq \delta$ for $1\leq i\leq t_1$. Combining this with $\delta\geq 2k$, we have \begin{equation*} \begin{aligned} e_{G}(\pi)\geq \frac{1}{2}\sum_{1\leq i\leq t_1}d_{G}(v_i)\geq \frac{1}{2}\delta t_1\geq k(t-1), \end{aligned} \end{equation*} which contradicts (\ref{equ::4}). This implies that $t_2\geq 2$. If the partition $\pi$ contains at most one nontrivial part, say $V_j$ ($1\leq j\leq t_2$), such that $|\partial_{G}(V_j)|\leq \delta-1$. Then $|\partial_{G}(V_i)|\geq \delta$ for all $i\in\{1,\ldots,t_2\}\backslash\{j\}$. Since $G$ is connected, it follows that $|\partial_{G}(V_j)|\geq 1$, and hence \begin{equation*} \begin{aligned} 2e_{G}(\pi)&= \sum_{1\leq i\leq t_2}|\partial_{G}(V_i)|+\sum_{1\leq j\leq t_1}d_{G}(v_j)\\ &\geq (t_2-1)\delta+1+\delta t_1\\ &= \delta(t-1)+1~~(\mbox{since $t=t_1+t_2$})\\ &\geq 2k(t-1)+1 ~~(\mbox{since $\delta\geq 2k$}), \end{aligned} \end{equation*} which also contradicts (\ref{equ::4}). Therefore, the partition $\pi$ contains at least two nontrivial parts, say $V_1,V_2$, such that $|\partial_{G}(V_i)|\leq \delta-1$ for $i=1,2$. Furthermore, by Lemma~\ref{lem::2.8}, we obtain $|V_i|\geq \delta+1$ for $i=1,2$. If $|V_1|=\max\{|V_1|,|V_2|,\ldots,|V_{t_2}|\}$ or $|V_2|=\max\{|V_1|,|V_{2}|,\ldots,|V_{t_2}|\}$, since $|V_{i}|\geq\delta+1$ and $|V_{j}|\geq2$ for $i=1,2$ and $3\leq j\leq t_2$, by Lemma \ref{lem::2.7}, \begin{equation*} \begin{aligned} \sum_{1\leq i\leq t_2}e_{G}(V_i)&\leq {\delta+1\choose 2}+(t_2-2){2\choose 2}+{n-(t+\delta+t_2-3)\choose 2}\\ &\leq{\delta+1\choose 2} +{n-(t+\delta-1)\choose 2} ~~(\mbox{since $t_2\geq 2$}). \end{aligned} \end{equation*} If there exists a nontrivial part, say $V_j$, such that $|V_j|=\max\{|V_1|,|V_2|,\ldots,|V_{t_2}|\}$ for some $3\leq j\leq t_2$. Similarly, $$\sum_{1\leq i\leq t_2}e_{G}(V_i)\leq 2{\delta+1\choose 2} +{n-(2\delta+t-1)\choose 2}.$$ Note that $|V_1|\geq \delta+1$ and $|V_2|\geq \delta+1$. Then $2\leq t\leq n-2\delta$. Combining this with (\ref{equ::4}) and $\sum_{1\leq i\leq t_1}e_{G}(v_i)=0$, we have \begin{equation*} \begin{aligned} e(G)&=\sum_{1\leq i\leq t_2}e_{G}(V_i)+\sum_{1\leq i\leq t_1}e_{G}(v_i)+e_{G}(\pi)\\ &\leq \max\left\{{\delta\!+\!1\choose 2}\!+\!{n\!-\!(t\!+\!\delta\!-\!1)\choose 2},2{\delta\!+\!1\choose 2}\!+\!{n\!-\!(2\delta\!+\!t\!-\!1)\choose 2}\right\}\!+\!k(t\!-\!1)\!-\!1\\ &\leq{\delta\!+\!1\choose 2}\!+\!{n\!-\!(t\!+\!\delta\!-\!1)\choose 2}+k(t\!-\!1)\!-\!1~~(\mbox{since $t\leq n-2\delta$})\\ &=\frac{t^2}{2}+(-n+\delta-\frac{1}{2}+k) t+\delta^2-n\delta+ \frac{n^2}{2}+\frac{n}{2}-k-1. \end{aligned} \end{equation*} Let $g(t)=\frac{t^2}{2}+(-n+\delta-\frac{1}{2}+k) t+\delta^2-n\delta+ \frac{n^2}{2}+\frac{n}{2}-k-1$. We take the derivative of $g(t)$. Thus $$g'(t)=t+\delta+k-n-\frac{1}{2}\leq k-\delta-\frac{1}{2}<0$$ by the facts that $\delta\geq 2k$, $k\geq 1$ and $t\leq n-2\delta$. This implies that $g(t)$ is decreasing with respect to $2\leq t\leq n-2\delta$, and hence $$e(G)\leq {\delta\!+\!1\choose2}\!+\!{n\!-\!\delta\!-\!1\choose 2}+k-1,$$ a contradiction. This completes the proof.\end{proof} \begin{proof}[\bf Proof of Theorem \ref{thm::1.2}] Assume to the contrary that $\tau(G)\leq k-1$. By Theorem~\ref{lem::2.9}, there exists a partition $\pi$ of $V(G)$ with $t_1$ trivial parts $v_1,v_2,\ldots, v_{t_1}$ and $t_2$ nontrivial parts $V_1,V_2,\ldots,V_{t_2}$ such that \begin{equation}\label{equ::5} \begin{aligned} e_{G}(\pi)\leq k(t-1)-1, \end{aligned} \end{equation} where $t=t_1+t_2$. By using the same analysis as Theorem \ref{thm::edgenumber}, we can deduce that the partition $\pi$ contains at least two nontrivial parts, say $V_1,V_2$, such that $|V_i|\geq\delta+1$ for $i=1,2$. First suppose that $t=2$. This implies that the partition $\pi$ consists of two nontrivial parts $V_1,V_2$. Then $e_{G}(\pi)=e_{G}(V_1,V_2)\leq k-1$ by (\ref{equ::5}). Clearly, $G$ is a spanning subgraph of some graph $H$ in $\mathcal{G}_{n,|V_1|}^{k-1}$. Then \begin{equation}\label{equ::6} \rho(G)\leq \rho(H), \end{equation} with equality if and only if $G\cong H$. Note that $\min\{|V_1|,|V_2|\}\geq\delta+1$. Combining this with Lemmas~\ref{lem::2.5} and \ref{lem::2.6} as well as (\ref{equ::6}), we conclude that $$\rho(G)\leq\rho(B_{n,\delta+1}^{k-1}),$$ with equality if and only if $G\cong B_{n,\delta+1}^{k-1}$. However, this is impossible because $\rho(G)\geq\rho(B_{n,\delta+1}^{k-1})$ and $G\ncong B_{n,\delta+1}^{k-1}$. Now suppose that $t\geq 3$. Note that $\rho(G)\geq\rho(B_{n,\delta+1}^{k-1})>\rho(K_{n-\delta-1})=n-\delta-2$. Then by Lemmas \ref{lem::2.1} and \ref{lem::2.2}, \begin{equation}\label{equ::7} e(G)> \frac{n^2}{2}-\frac{(2\delta+3)n}{2}+(\delta+1)^{2}. \end{equation} Since $\min\{|V_1|,|V_2|\}\geq\delta+1$, by using a similar analysis as Theorem \ref{thm::edgenumber}, it follows that \begin{equation*} \begin{aligned} e(G)&=\sum_{1\leq i\leq t_2}e_{G}(V_i)+\sum_{1\leq i\leq t_1}e_{G}(v_i)+e_{G}(\pi)\\ &\leq \max\left\{{\delta\!+\!1\choose 2}\!+\!{n\!-\!(t\!+\!\delta\!-\!1)\choose 2},2{\delta\!+\!1\choose 2}\!+\!{n\!-\!(2\delta\!+\!t\!-\!1)\choose 2}\right\}\!+\!k(t\!-\!1)\!-\!1\\ &\leq{\delta\!+\!1\choose 2}\!+\!{n\!-\!(t\!+\!\delta\!-\!1)\choose 2}+k(t\!-\!1)\!-\!1~~(\mbox{since $t\leq n-2\delta$})\\ &= \frac{n^2}{2}-\frac{(2t+2\delta-1)n}{2}+\frac{(t+2\delta+2k-1)t}{2}+\delta^2-k-1. \end{aligned} \end{equation*} Combining this with (\ref{equ::7}) and $t\geq 3$, we have \begin{equation}\label{equ::8} n< \delta+k+\frac{t+1}{2}+\frac{k-1}{t-2}. \end{equation} Suppose that $f(t)=\delta+k+\frac{t+1}{2}+\frac{k-1}{t-2}$. One can verify that $f(t)$ is convex for $t>0$ and its maximum in any closed interval is attained at one of the ends of this interval. Note that $3\leq t\leq n-2\delta$. Then \begin{equation*} \begin{aligned} f(t)&\leq \max\left\{\delta+2k+1, \frac{n+1}{2}+k+\frac{k-1}{n-2\delta-2}\right\} \leq \frac{n-1}{2}+2k~(\mbox{since $n\geq 2\delta +3$}). \end{aligned} \end{equation*} Combining this with (\ref{equ::8}) and $\delta\geq 2k$, we can deduce that $n<4k-1\leq 2\delta -1$, a contradiction. This completes the proof. \end{proof} \section{Proof of Theorem \ref{thm::1.3}}\label{sec::econ} The proof idea of Theorem~\ref{thm::1.3} is quite similar to that of Theorem~\ref{thm::1.2}. To present the proof, we need several lemmas below. \begin{lem}[Theorem~3.1 of \cite{Ning}]\label{lem::3.1} If $G_0$ is a graph with the maximum spactral radius in $\mathcal{A}_{n}^{\kappa',\delta}$, where $1\leq\kappa'<\delta$, then $G_0\in \mathcal{G}_{n,\delta+1}^{\kappa'}$. \end{lem} Recall that $\mathcal{G}_{n,n_1}^{i}$ is the set of graphs obtained from $K_{n_1}\cup K_{n-n_{1}}$ by adding $i$ edges between $K_{n_1}$ and $K_{n-n_{1}}$, and $B_{n,\delta+1}^{i}$ is the graph obtained from $K_{\delta+1}\cup K_{n-\delta-1}$ by adding $i$ edges joining a vertex in $K_{\delta+1}$ and $i$ vertices in $K_{n-\delta-1}$. We can extend the results of Lemmas \ref{lem::2.3} and \ref{lem::2.5}. \begin{lem}\label{lem::3.2} Let $G\in\mathcal{G}_{n,\delta+1}^{\kappa'}$, where $n\geq 2\delta+4$ and $4\leq \kappa'< \delta$. Then $$n-\delta-2< \rho(G)< n-\delta.$$ \end{lem} \begin{proof} Note that $G$ contains $K_{\delta+1}\cup K_{n-\delta-1}$ as a proper spanning subgraph and $n\geq 2\delta+4$. Then $\rho(G)>\rho(K_{\delta+1}\cup K_{n-\delta-1})= n-\delta-2$. Since $G\in\mathcal{G}_{n,\delta+1}^{\kappa'}$ and $\delta> \kappa'$, we have \begin{equation*} \begin{aligned} e(G)&={\delta+1\choose 2}+{n-\delta-1\choose 2}+\kappa'\\ &\leq {\delta+1\choose 2}+{n-\delta-1\choose 2}+\delta-1\\ &=\frac{n^2}{2}-\frac{(2\delta+3)n}{2}+\delta^2+3\delta. \end{aligned} \end{equation*} Combining this with Lemmas \ref{lem::2.1} and \ref{lem::2.2}, we have \begin{equation*} \begin{aligned} \rho(G)&\leq \frac{\delta-1}{2}+\sqrt{n^2 + (-3\delta - 3)n + \frac{9\delta^2}{4}+ \frac{13\delta}{2}+\frac{1}{4}}\\ &= \frac{\delta-1}{2}+\sqrt{\left(n-\frac{3\delta}{2}+\frac{1}{2}\right)^2-4(n-2\delta )}\\ &< \frac{\delta-1}{2}+ \left(n-\frac{3\delta}{2}+\frac{1}{2}\right)~~(\mbox{since $n\geq 2\delta\!+\!4$})\\ &=n-\delta. \end{aligned} \end{equation*} Thus $n-\delta-2< \rho(G)< n-\delta$, as required.\end{proof} By Lemma \ref{lem::3.2} and using the same analysis as the proof of Lemma \ref{lem::2.5}, we easily obtain the following result. \begin{lem}\label{lem::3.3} Let $G\in \mathcal{G}_{n,\delta+1}^{\kappa'}$ where $4\leq \kappa'< \delta$ and $n\geq 2\delta+4$. Then $\rho(G)\leq \rho(B_{n,\delta+1}^{\kappa'})$, with equality if and only if $G\cong B_{n,\delta+1}^{\kappa'}$. \end{lem} \begin{proof}[\bf Proof of Theorem \ref{thm::1.3}] The result follows from Lemmas~\ref{lem::3.1} and \ref{lem::3.3}. \end{proof} \section{Some applications}\label{sec::app} Edge-disjoint spanning trees are closely related to many graph properties, such as collapsibility, supereulerianity, spanning connectivity, nowhere-zero flows, group connectivity, rigidity, and others \cite{CDGG22,LL19,Palmer01}. We introduce several applications of our main result in spectral graph theory. Spectral conditions of classic rigidity have been studied in \cite{CDGG22,CDG21,FHL22}. We present spectral conditions for two other variations of rigidity in the next two subsections. To conclude, we also present applications in nowhere-zero flows at the end of this section. \subsection{Body-and-bar rigidity} A \textit{body-and-bar frameworks} in $\mathbb{R}^d$ is a framework of $d$-dimension rigid bodies that are connected by fixed-length bars attached at points of their surfaces (see \cite{Tay84} for more details). Informally, we say a graph $G$ is \textit{body-bar rigid in $\mathbb{R}^d$} if there exists a generic rigid body-bar framework in $\mathbb{R}^d$. Instead of a formal definition, we present the following characterization. \begin{thm}[Tay~\cite{Tay84}] A graph $G$ is body-bar rigid in $\mathbb{R}^d$ if and only if it contains $\frac{d(d+1)}{2}$ edge-disjoint spanning trees. \end{thm} By the above theorem and Theorem~\ref{thm::1.2}, we have the following spectral condition for body-bar rigidity. \begin{pro} Let $k=\frac{d(d+1)}{2}$, and let $G$ be a connected graph with minimum degree $\delta\geq 2k$ and order $n\geq 2\delta+3$. If $\rho(G)\geq \rho(B_{n,\delta+1}^{k-1})$, then $G$ is body-bar rigid in $\mathbb{R}^d$ unless $G\cong B_{n,\delta+1}^{k-1}$. \end{pro} \subsection{Rigidity on surfaces of revolution} Here we assume that the joints of our framework are restricted to lie on a smooth surface $\mathcal{M} \subset \mathbb{R}^3$, and $\mathcal{M}$ is an \textit{irreducible surface}; i.e.,~$\mathcal{M}$ is the zero set of an irreducible rational polynomial $h(x,y,z) \in \mathbb{Q}[X,Y,Z]$. The framework $(G,p)$ with $p(v) \in \mathcal{M}$ for every $v \in V(G)$ is \textit{rigid on $\mathcal{M}$} if there exists $\varepsilon >0$ such that if $(G,p)$ is equivalent to $(G,q)$ and $\|p(v)-q(v)\|<\epsilon$ and $q(v) \in \mathcal{M}$ for every $v \in V(G)$, then $(G,p)$ is congruent to $(G,q)$. An irreducible surface is called an \textit{irreducible surface of revolution} if it can be generated by rotating a continuous curve about a fixed axis. See \cite{NixonOwenPower12} for more information. \begin{thm}[Nixon, Owen and Power \cite{NixonOwenPower12,NixonOwenPower14}] \label{thm::nop} Let $\mathcal{M}$ be an irreducible surface of revolution. Then a graph $G$ is rigid on $\mathcal{M}$ if and only if one of the following: \\$(i)$ $G$ is a complete graph, \\$(ii)$ $\mathcal{M}$ is a sphere and $G$ contains a spanning Laman graph, \\$(iii)$ $\mathcal{M}$ is a cylinder and $G$ contains two edge-disjoint spanning trees, or \\$(iv)$ $\mathcal{M}$ is not a cylinder or a sphere and $G$ contains two edge-disjoint spanning subgraphs $G_1,G_2$, where $G_1$ is a tree and every connected component of $G_2$ contains exactly one cycle. \end{thm} By the above theorem, the results in \cite{FHL22} actually imply spectral conditions for the rigidity on sphere. By Theorem~\ref{thm::1.2}, we have the following spectral condition for rigidity on irreducible surfaces of revolution that is not a sphere. \begin{pro} Let $G$ be a connected graph with minimum degree $\delta\geq 4$ and order $n\geq 2\delta+3$. If $\rho(G)\geq \rho(B_{n,\delta+1}^{1})$, then $G$ is rigid on any irreducible surface of revolution that is not a sphere unless $G\cong B_{n,\delta+1}^{1}$. \end{pro} \begin{proof} By Theorem~\ref{thm::1.2}, $G$ contains two edge-disjoint spanning trees. Since $\delta\ge 4$, we have $|E(G)|\ge n\delta/2\ge 2n$. Thus $G$ has extra edges that are not in the two edge-disjoint spanning trees. It follows that $G$ contains a spanning tree and a spanning subgraph with exactly one cycle that are edge-disjoint. By Theorem~\ref{thm::nop}, the result follows. \end{proof} \subsection{Nowhere-zero flows} The theory of integer flows was initiated by Tutte. For an orientation $D$ of a graph $G$ and for each vertex $v$, let $E_D^+(v)$ and $E_D^-(v)$ be the set of edges oriented away from $v$ and the set of edges oriented into $v$, respectively. A \textit{nowhere-zero $k$-flow} of a graph $G$ is an orientation $D$ together with a function $f: E(G)\rightarrow \{\pm 1, \pm 2,\ldots, \pm(k-1)\}$ such that $\sum_{e\in E_D^+(v)} f(e) = \sum_{e\in E_D^-(v)} f(e)$ for each vertex $v$. As a generalization of nowhere-zero $k$-flow, a \textit{nowhere-zero circular $k/d$-flow}, introduced by Goddyn, Tarsi and Zhang~\cite{GTZ98}, is a nowhere-zero $k$-flow such that the range of $f$ is contained in $\{\pm d, \pm (d+1),\ldots, \pm(k-d)\}$. The \textit{flow index $\phi(G)$} of a graph $G$ is the least rational number $r$ such that $G$ admits a nowhere-zero circular $r$-flow. The central problems in this research area are the three well-known flow conjectures proposed by Tutte. For the application of spanning tree packing number in nowhere-zero flow theory, we refer readers to~\cite{LL19}. In particular, it is proved in~\cite{LXZ07} that if $\tau(G)\ge 3$, then $\phi(G)<4$; and in~\cite{HLL18} that if $\tau(G)\ge 4$, then $\phi(G)\le 3$. Thus, by Theorem~\ref{thm::1.2}, we can easily obtain sufficient conditions of flow index via spectral radius, however, these spectral conditions might not be best possible. It is worth studying flow index directly from spectral perspectives in the future. \section{Concluding remarks}\label{sec::remarks} Theorem~\ref{thm::1.2} actually implies that $B_{n,\delta+1}^{\tau}$ is the unique graph that has the maximum spectral radius among all graphs of fixed order $n$ with minimum degree $\delta$ and spanning tree packing number $\tau$. Let $G$ be a minimum graph with $\tau(G)\ge k$ and of order $n$, that is, $G$ consists of exactly $k$ edge-disjoint spanning trees with no extra edges. This implies that $\tau(G)=k$ and $e(G)=k(n-1)$. We are interested in the maximum possible spectral radius of $G$. \begin{prob}\label{prob::minkst} Let $G$ be a minimum graph with $\tau(G)\ge k$ and of order $n$. For each $n\ge 4$ and each $k\ge 2$, determine the maximum possible spectral radius of $G$ and characterize extremal graphs. \end{prob} To attack Problem~\ref{prob::minkst}, notice that $\delta(G)\ge k$ and $e(G)=k(n-1)$, we can easily obtain an upper bound on $\rho(G)$ by Lemma~\ref{lem::2.1}. However, this upper bound may not be tight, since for most values of $n$, $G$ cannot be $k$-regular or a bidegreed graph in which each vertex is of degree either $k$ or $n-1$. To attain the maximum spectral radius, it seems that $G$ is obtained from a $K_{2k}$ by continuously adding a vertex and $k$ incident edges step by step until $G$ has $n$ vertices, and in particular, we guess the graph is $K_{k}\nabla (K_{k}\cup (n-2k)K_1)$, where $\nabla$ and $\cup$ denote the join and the union of two graphs, respectively. We leave this as an open question. Nash-Williams~\cite{Nash64} ever studied the forest covering problem, seeking the minimum number of forests that cover the entire graph. This is like a dual problem of spanning tree packing. The {\it arboricity} $a(G)$ is the minimum number of edge-disjoint forests whose union equals $E(G)$. \begin{thm}[Nash-Williams~\cite{Nash64}]\label{thm::5.1} Let $G$ be a connected graph. Then $a(G)\le k$ if and only if for any subgraph $H$ of $G$, $|E(H)|\le k(|V(H)|-1)$. \end{thm} Naturally, we have the following problem. \begin{prob}\label{prob::arbo} Find a tight spectral radius condition for a graph $G$ of order $n$ with $a(G)\le k$ and characterize extremal graphs. \end{prob} When $n\leq 2k$, Problem~\ref{prob::arbo} is trivial. Any graph $G$ of order $n\leq 2k$ has the property $a(G)\le k$ and so the extremal graph is $K_n$. To see this, for any subgraph $H$ of $G$, we can deduce that $|E(H)|\leq {|V(H)|\choose 2} \leq k(|V(H)|-1)$ when $n\leq 2k$, and the conclusion follows from Theorem~\ref{thm::5.1}. The situation becomes more involved for $n\geq 2k+1$. In fact, this case is even stronger than Problem~\ref{prob::minkst}, since if $G$ consists of exactly $k$ edge-disjoint spanning trees with no extra edges, we have $a(G) =\tau(G)=k$. Notice that $a(K_{k}\nabla (K_{k}\cup (n-2k)K_1))=k$ and $e(K_{k}\nabla (K_{k}\cup (n-2k)K_1))=k(n-1)$, thus we guess that the extremal graph w.r.t. the spectral radius is also $K_{k}\nabla (K_{k}\cup (n-2k)K_1)$. This is left for possible future work. \iffals \section{Some possible problems} \red{This is a temporary section for discussing and listing related problems.} \begin{prob} If a simple graph $G$ consists of exactly $k$ edge-disjoint spanning trees with maximum possible spectral radius, what is the structure of $G$? \end{prob} \red{For example, if $k=2$, it seems $G$ can be obtained from a $K_4$ by continuously adding a vertex and two edges in each step? In general, $G$ can be obtained from a $K_{2k}$ by continuously adding a vertex and $k$ edges step by step?} \medskip A {\bf $c$-forest} is a forest with exactly $c$ components. \begin{prob}[Cioab\u{a} and Wong~\cite{Wong}] Find a (tight) spectral condition for a graph $G$ that contains $k$ edge-disjoint spanning $c$-forests. \end{prob} \begin{prob} Find a (tight) spectral condition for a graph $G$ that contains two edge-disjoint spanning subgraphs $G_1,G_2$, where $G_1$ is a tree and every connected component of $G_2$ contains exactly one cycle. \red{It has an application in rigidity theory.} \end{prob} \begin{prob} Find a (tight) spectral condition for a graph $G$ so that $\tau(G-e)\ge k$ for every edge $e$ of $G$. (\red{Is this feasible?}) \end{prob} \begin{prob} Denoted by $tG$ the multigraph obtained from $G$ by replacing every edge with $t$ parallel edges. Find a (tight) spectral condition for a graph $G$ so that $\tau(tG)\ge k$ (or $\tau(tG-e)\ge k$). \red{It has an application in rigidity theory.} \end{prob} \begin{prob} Any generalized edge connectivity version of Theorem~\ref{thm::3.1}? \end{prob} Motivated by Conjecture~\ref{conj1}, we have the following problem. To attack this problem, we need a result similar to Lemma~\ref{lem::3.1}. \begin{prob} Let $\mathcal{H}_{n}^{\tau,\delta}$ be a set of graphs of order $n$ with minimum degree $\delta$ and STP number $\tau$. What is the graph with maximum spectral radius in $\mathcal{H}_{n}^{\tau,\delta}$? \red{This can be solved by Theorem~\ref{thm::1.2}.} \end{prob} \f \section*{Acknowledgements} Xiaofeng Gu was supported by a grant from the Simons Foundation (522728), and Huiqiu Lin was supported by the National Natural Science Foundation of China (No. 12011530064) and Natural Science Foundation of Shanghai (No. 22ZR1416300).

https://arxiv.org/abs/1302.2076

On measures of symmetry and floating bodies

We consider the following measure of symmetry of a convex n-dimensional body K: $\rho(K)$ is the smallest constant for which there is a point x in K such that for partitions of K by an n-1-dimensional hyperplane passing through x the ratio of the volumes of the two parts is at most $\rho(K)$. It is well known that $\rho(K)=1$ iff K is symmetric. We establish a precise upper bound on $\rho(K)$; this recovers a 1960 result of Grunbaum. We also provide a characterization of equality cases (relevant to recent results of Nill and Paffenholz about toric varieties) and relate these questions to the concept of convex floating bodies.

\section{\@startsection {section}{1}{\z@}{-3.5ex plus-1ex minus -.2ex}{2.3ex plus.2ex}{\reset@font\large\bf}} \def\subsection{\@startsection{subsection}{2}{\z@}{-3.25ex plus-1ex minus-.2ex}{-.1em}{\reset@font\large\bf}} \def\subsubsection{\@startsection{subsubsection}{3}{\z@}{-3.25ex plus -1ex minus-.2ex}{-.1em}{\reset@font\normalsize\bf}} \makeatother \date{} \title{ On measures of symmetry and floating bodies} \author{Stanislaw J. Szarek (Cleveland \& Paris) } \defS^{n-1}{S^{n-1}} \def\rightarrow{\rightarrow} \newcommand{{\mathbb{R}}}{{\mathbb{R}}} \newcommand{\R ^{+}}{{\mathbb{R}} ^{+}} \newcommand{\R ^n}{{\mathbb{R}} ^n} \newcommand{{\mathbb{N}}}{{\mathbb{N}}} \newtheorem{fact}{Fact \newtheorem{thm}[fact]{Theorem} \newtheorem{prop}[fact]{Proposition} \newtheorem{lemma}[fact]{Lemma} \newtheorem{cor}[fact]{Corollary} \newtheorem{ntn}[fact]{Notation} \newtheorem{dfn}[fact]{Definition} \newtheorem{example}[fact]{Example} \newtheorem{claim}[fact]{Claim} \begin{document} \maketitle \begin{center} {\small{\sl Dedicated to Olek Pe{\l}czy{\'n}ski, a teacher and a friend}} \end{center} \begin{abstract} We consider the following measure of symmetry of a convex $n$-dimensional body $K$: $\rho(K)$ is the smallest constant for which there is a point $x \in K$ such that for partitions of $K$ by an $n-1$-dimensional hyperplane passing through $x$ the ratio of the volumes of the two parts is $\le \rho(K)$. It is well known that $\rho(K)=1$ iff $K$ is symmetric. We establish a precise upper bound on $\rho(K)$; this recovers a 1960 result of Gr\"unbaum. We also provide a characterization of equality cases (relevant to recent results of Nill and Paffenholz about toric varieties) and relate these questions to the concept of convex floating bodies. \end{abstract} \noindent {\bf 1. Introduction.} The structure of general $n$-dimensional (bounded) convex bodies is understood much less than that of the symmetric ones. For example, if we endow each of these classes with the appropriate version of the Banach-Mazur distance, then the asymptotic order (as $n \rightarrow \infty$) of the diameter of the resulting compactum has been known for the symmetric case since the 1981 seminal Gluskin's paper \cite{Gl}, while the corresponding problem in the general case is wide open, with the first non-trivial results having been obtained only in the last several years \cite{BLPS, R}. Various invariants have been proposed to explain the difference between these two classes; see, e.g., \cite{Gr1} for an early survey of related work. One such measure of symmetry (or rather of asymmetry) was mentioned to the author by Olek Pe{\l}czynski around the turn of the millennium. Given convex body $K \subset \R ^n$ and $x \in K$, consider all partitions of $K$ into two parts by an $n-1$-dimensional hyperplane $H$ passing through $x$, and let $\rho(K,x)$ be the largest ratio of volumes of the two parts. Next, let $\rho(K) :=\min_{x\in K}{\rho(K,x)}$. Clearly $\rho(K) = 1$ if $K$ is centrally symmetric. (The reverse implication is also true but nontrivial, even for the larger class of star-shaped bodies; see \cite{S1} for details and \cite{Gro} for additional references.) Olek's question was to establish a dimension-free upper bound on $\rho(K)$. I came up with an argument (which also established a precise upper bound on $\rho(K)$ for $K \subset \R ^n$ and characterized bodies, for which that upper bound -- call it $\rho_n$ -- is attained) and wrote it up some time afterwards, but then realized that the question had been considered and solved by Gr\"unbaum in 1960 \cite{Gr0} and so I abandoned the project. However, it transpired very recently \cite{BB, NP} that Gr\"unbaum's result and a characterization of $K \subset \R ^n$ for which $\rho(K)=\rho_n$ were relevant to problems in toric geometry. Moreover, since the question in \cite{Gr0} was stated slightly differently than above, it led to an apparently non-equivalent analysis of equality cases (see \cite{Gr0}, Remark 4(i), p. 1260), less suitable -- at least without additional work -- for the applications considered in \cite{NP}. Accordingly, I am posting the manuscript (with added references and minor editorial changes). \medskip \noindent {\bf 2. More background and the results.} The parameter $\rho( K )$ is related to another geometrical concept, the convex floating bodies of $K$. [To give meaning to the formulae and to avoid artificial anomalies, it should be understood that $K$ -- and all bodies above and in what follows -- is convex, compact and has a nonempty interior.] Let $\delta \in (0, 1/2]$; slightly modifying the original definition from \cite{SW1}, let us denote by $K^\delta$ the intersection of all half-spaces whose complements contain at most the proportion $\delta$ of the volume of $K$. As is easy to see, if we call $\phi(K)$ the largest number $\delta$ for which the convex floating body $K^\delta$ is nonempty, then $\phi(K)=(\rho(K)+1)^{-1}$. The existence of a universal (i.e., independent of $n$ and $K$) strictly positive lower bound for $\phi(K)$ (and, analogously, universal upper bound for $\rho( K )$) has been a part of the folklore for some time \cite{Gi, SW2}. Here we prove the following ``isometric" result. \begin{thm} Let $n \in {\mathbb{N}} $ and let $K \subset \R ^n$ be a convex, compact body with nonempty interior. Let $c$ be the centroid of $K$ and let $H$ be an $n-1$-dimensional hyperplane which passes through $c$, and thus divides $K$ into two parts. Then the ratio of volumes of the two parts is $ \le (1+1/n)^n-1=: \rho_n$ (which is $< e -1 < 1.7183$). Moreover, we have an equality iff $K$ is a ``pyramid," i.e., $K = {\rm conv}(\{v\} \cup B)$, where $B$ is an $n-1$-dimensional convex body (the ``base") and $v$ the vertex, and $H$ is parallel to $B$. \end{thm} Similar statements about $\rho(K)$ and $\phi(K)$ are then simple consequences. \begin{cor} In the notation and under the hypotheses of Theorem 1 we have \smallskip \noindent {\rm (i) } $\rho(K) \le \rho_n$; moreover, $\rho(K,c) \le \rho_n$, with equality iff $K$ is a pyramid and the maximizing hyperplane $H$ is parallel to a base of the pyramid. \smallskip \noindent {\rm (ii) } $\phi(K) \ge (1+1/n)^{-n} =:\delta_n$; moreover, $c \in K^{\delta_n}$ with $c \in \partial K^{\delta_n}$ iff $K$ is a pyramid. \end{cor} The estimates $\rho(K) \le \rho_n$ and $\phi(K) \ge \delta_n$ are, in general, best possible as seen from the example of a simplex. \medskip \noindent {\bf 3. The proofs.} {\em Proof of Theorem 1 } Without loss of generality we may assume that the centroid $c$ of $K$ is at the origin. Let $\theta \in S^{n-1}$, $H = \theta^\perp$ and consider the function $f = f_\theta : {\mathbb{R}} \rightarrow \R ^{+}$ defined by \begin{equation} \label{sections} f(t) := {\rm vol_{n-1}} (K \cap (H+t\theta). \end{equation} It then follows that $f$ is upper semi-continuous (since $K$ is closed) and supported on some bounded interval $[-a, b]$ with $a, b>0$. (In fact it will follow from our arguments -- and is likely well known -- that the ratio $|a|/|b| \in [1/n,n]$.) Clearly, ${\rm vol_{n}} (K) = \int_{-a}^b {f(t) \, dt }$ and \begin{equation} \label{rhoeq} \rho(K,c) = \max_{\theta \in S^{n-1}} {\frac{\int_0^b {f(t) \, dt}}{\int_{-a}^0 {f(t) \, dt}}}. \end{equation} Additionally, the centroid being at the origin is equivalent to \begin{equation} \label{centroid} \int_{-a}^b {tf(t) \, dt } = 0 . \end{equation} The only other property of $f$ we shall need is that $h := f^{1/{(n-1)}}$ is concave on $[-a,b]$, which is a consequence of the Brunn-Minkowski inequality. We note that the concavity implies continuity on $(-a,b)$ and lower semi-continuity, hence continuity on $[-a,b]$. The Theorem will follow easily from the following two claims. \begin{claim} Let $n \in {\mathbb{N}}$ and $a,b > 0$. For any $\theta \in S^{n-1}$ and for any continuous function $f : [-a,b] \rightarrow \R ^{+}$ such that $h :=f^{1/{(n-1)}}$ is concave, there exists a closed convex body $K \subset \R ^n$ such that $f$ is obtained from $K, \theta$ via {\rm (\ref{sections})}. If, additionally, {\rm (\ref{centroid})} holds, then $K$ may be chosen so that its centroid is at the origin. Finally, $f$ is affine with $f(-a)=0$ (or $f(b)=0$) iff $K$ is a pyramid and $\theta $ is perpendicular to a base $B$; in that case, if {\rm (\ref{centroid})} also holds, the ratio from {\rm (\ref{rhoeq})} equals $(1+1/n)^n-1$. \end{claim} \begin{claim} Let $a, b >0$. Among continuous functions on $[-a,b]$, strictly positive on $(-a,b)$, verifying {\rm (\ref{centroid})} and such that $h :=f^{1/{(n-1)}}$ is concave on $[-a,b]$, the largest value of the ratio appearing in {\rm (\ref{rhoeq})} is achieved iff $h$ is affine on $[-a,b]$ with $h(-a) = 0$. \end{claim} Claim 3 shows that investigating $\rho(K)$ and the ratio from (\ref{rhoeq}) for functions verifying our assumptions are fully equivalent. Its proof is based on elementary geometric considerations. To construct $K$ starting from $f$, we choose any $n-1$-dimensional convex body $B_0$ in $H = \theta^\perp$ with ${\rm vol_{n-1}} (B_0)=1$ and $0 \in B_0$, and set $K:= \bigcup_{t \in [-a,b]} t\theta + h(t)B_0$. The ``only if" part in the last assertion follows from the analysis of equality cases in the Brunn-Minkowski inequality (it occurs ``essentially iff" the two sets are homothetic, see \cite{S1}, Theorem 6.1.1). The details are left to the reader. \medskip The proof of Claim 4 is also elementary, but less obvious. We will use the following lemma, variants of which exist in the literature. Similar arguments were employed (independently of this note) in \cite{F}, see also \cite{FG} for a more conceptualized application of closely related phenomena. \begin{lemma} Let $M >0$, $m \in {\mathbb{R}}$ and $n \in {\mathbb{N}}$. We consider the set $\mathcal{H}$ of functions $h : {\mathbb{R}} \rightarrow \R ^{+}$ which verify \newline {\rm (i) } the support of $h$ is the interval $[0,b]$ (for some $b>0$) \newline {\rm (ii) } $h$ is continuous and concave on $[0,b]$ \newline {\rm (iii) } $h(0)=1$ and the right derivative of $h$ at $0$ is $\leq m$ \newline {\rm (iv) } if $f := h^{n-1}$, then $\int_0^b {tf(t) \, dt } = M$. \newline The set $\mathcal{H}$ is nonempty iff $m \ge -1/\sqrt{Mn(n+1)}$. In that case, set $\mu(h):= \int_0^b {f(t) \, dt}$ for $h \in \mathcal{H}$. The minimal value of $\mu(h)$ is attained iff $h$ is affine on its support $[0,b]$ with $f(b)=0$. The maximal value of $\mu(h)$ is attained iff $h(t) = 1+mt $ on the support of $h$. \end{lemma} \noindent {\em Proof of Claim 4.} Since the ratio in (\ref{rhoeq}) doesn't change if $f(\cdot)$ is replaced by $\alpha f(\beta\,\cdot)$, it is enough to consider $f$'s whose support verifies $[-1/n,1/n] \subset [-a,b] \subset [-1,1]$ and such that $f(1)=1$ (for the first inclusion, see the comments in the paragraph following (\ref{sections}) and at the very end of this note). Concavity of $f^{1/(n-1)}$ gives then a lower bound on $\int_{-a}^0 {f(t) \, dt}$ and an upper bound on $\|f\|_{\infty}$ (both dependent on $n$). It follows that the set of the functions $f$ in question is compact (say, in the $L_1$ metric, which is relevant here) and hence that the supremum of the ratio in (\ref{rhoeq}) is attained. (This can also be proved in a variety of ways, including from the John's theorem.) Let $f_0$ be such an extremal function (with support $[-a_0,b_0]$), we shall show that it must be of the form indicated in Claim 4. Indeed, if $f_0$ was not affine on $[-a_0,0]$ with $f(-a_0)=0$, we could apply the Lemma with $h(\cdot) = f_0(-\,\cdot)^{1/{(n-1)}}: [0,a_0] \rightarrow \R ^{+}$, $M=-\int_{-a_0}^0 {tf_0(t) \, dt}$ and $m$ equal to the right derivative of $h$ at $0$ to obtain an extremal $h_1$ (supported on a possibly different interval $[0,a_1]$) for which $\mu(h_1) < \mu(h_0)$. Defining $f_1$ to coincide with $f_0$ on $[0,b_0]$ and with $h_1(-\, \cdot)^{n-1}$ on $[-a_1,0]$ we would get a function for which the ratio from (\ref{rhoeq}) was strictly larger than for $f_0$. At the same time, the conditions (iii) and (iv) from the Lemma (together with our choices of $m, M$ would assure that, respectively, $f_1^{1/{(n-1)}}$ was concave on $[-a_1,b_0]$ and that (\ref{centroid}) was satisfied. This shows that $f_0(t) = (1+t/a_0)^{n-1}$ for $t \in [-a_0,0]$. A similar argument applied to $h= f_0^{1/{(n-1)}}: [0,b_0] \rightarrow \R ^{+}$, $m=1/a_0$ and the same $M$ shows that $f_0$ is affine on the entire interval $[-a_0,b_0]$, which concludes the proof of the Claim. (In fact we showed directly that affine $h$'s give the extremal value of the ratio from (\ref{rhoeq}), so we didn't really need to know that the supremum was attained.) \hfill $\Box$ \bigskip \noindent {\em Sketch of the proof of Lemma 5.} First, let us point out that the condition $m \geq -1/\sqrt{Mn(n+1)}$ is equivalent to $\mathcal{H} \neq \emptyset$. Indeed, this is easily deduced from the observation that if $h$ is supported on $[0,b]$ and defined there by $h(t) = 1-t/b$, then we have the relationship $b=\sqrt{Mn(n+1)}$. The assertion of the Lemma is then ``essentially obvious from physical considerations.'' To obtain maximal ``mass" $\mu(h)$ for a given ``moment" $M$ (the constraint (iv)), we need to place the mass as closely to the axis $t=0$ as allowed by the concavity condition (ii) and by (iii). To minimize the mass for a fixed moment, we need to place the mass as far from $t=0$ as possible subject to (ii) and (iii). This is easily formalized. A very similar argument allows also to determine the largest possible value of $b$ for a given $M$, and the smallest possible value for $b$ (if at all possible) given $m$ and $M$, and to subsequently deduce that the ratio of the two is at most $n$, thus implying the bounds on the ratio $|a|/|b|$ stated in the paragraph following (\ref{sections}) and mildly used in the proof of Claim 4. \hfill $\Box$ \medskip \noindent{\small {\em Acknowledgement} \ Supported in part by grants from NSF~(U.S.A.) and by an International Cooperation Grant from Min. des Aff. Etrang. (France) \& KBN (Poland).} \small

https://arxiv.org/abs/0912.5000

Classification of Q-trivial Bott manifolds

A Bott manifold is a closed smooth manifold obtained as the total space of an iterated $\C P^1$-bundle starting with a point, where each $\C P^1$-bundle is the projectivization of a Whitney sum of two complex line bundles. A \emph{$\Q$-trivial Bott manifold} of dimension $2n$ is a Bott manifold whose cohomology ring is isomorphic to that of $(\CP^1)^n$ with $\Q$-coefficients. We find all diffeomorphism types of $\Q$-trivial Bott manifolds and show that they are distinguished by their cohomology rings with $\Z$-coefficients. As a consequence, we see that the number of diffeomorphism classes in $\Q$-trivial Bott manifolds of dimension $2n$ is equal to the number of partitions of $n$. We even show that any cohomology ring isomorphism between two $\Q$-trivial Bott manifolds is induced by a diffeomorphism.

\section{Introduction} A Bott tower of height $n$ is a sequence of $\mathbb{C}P^1$-bundles \begin{equation} \label{eqn:Bott tower} B_n\stackrel{\pi_n}\longrightarrow B_{n-1} \stackrel{\pi_{n-1}}\longrightarrow \dots \stackrel{\pi_2}\longrightarrow B_1 \stackrel{\pi_1}\longrightarrow B_0=\{\text{a point}\}, \end{equation} where each $\pi_i\colon B_i\to B_{i-1}$ for $i=1,\dots,n$ is the projectivization of a Whitney sum of two complex line bundles over $B_{i-1}$. We call $B_i$ an \emph{$i$-stage Bott manifold} and are concerned with the diffeomorphism type of the $n$-stage Bott manifold $B_n$. Note that even if two Bott towers of height $n$ are different, their $n$-stage Bott manifolds can be diffeomorphic. If the fiber bundles in \eqref{eqn:Bott tower} are all trivial, then $B_n$ is diffeomorphic to $(\mathbb{C}P^1)^n$. It is shown in \cite{Ma-Pa-2008} that if the cohomology ring of $B_n$ is isomorphic to that of $(\mathbb{C}P^1)^n$ with $\mathbb{Z}$-coefficients as graded rings, then $B_n$ is diffeomorphic to $(\mathbb{C} P^1)^n$ and moreover the fiber bundles in \eqref{eqn:Bott tower} are all trivial. We say that $B_n$ is \emph{$\mathbb{Q}$-trivial} if its cohomology ring is isomorphic to that of $(\mathbb{C}P^1)^n$ with $\mathbb{Q}$-coefficients as graded rings. In this paper, we shall find all diffeomorphism types of $\mathbb{Q}$-trivial Bott manifolds and show that they are diffeomorphic if and only if their cohomology rings with $\mathbb{Z}$-coefficients are isomorphic as graded rings (Theorem~\ref{theorem:Q-trivial Bott manifold}). As a consequence, we see that the number of diffeomorphism classes in $\mathbb{Q}$-trivial Bott manifolds of dimension $2n$ is equal to the number of partitions of $n$. We also prove that any automorphism of the cohomology ring of a $\mathbb{Q}$-trivial Bott manifold is induced by a diffeomorphism. This implies that any cohomology ring isomorphism between two $\mathbb{Q}$-trivial Bott manifolds is induced by a diffeomorphism since we already establish that the diffeomorphism types of $\mathbb{Q}$-trivial Bott manifolds are distinguished by their cohomology rings. Our study is motivated by the so-called \emph{cohomological rigidity problem} for toric manifolds. A toric manifold is a non-singular compact complex algebraic variety with an algebraic torus action having a dense orbit. The cohomological rigidity problem for toric manifolds asks whether the topological types of toric manifolds are distinguished by their cohomology rings or not (see \cite{Ma-Su-2008}). This problem is open, but we have some affirmative partial solutions to the problem for (generalized) Bott manifolds in \cite{Ma-Pa-2008}, \cite{Ch-Ma-Su-2010}, \cite{ch-su-pre} and \cite{Ch-Par-Su-pre}. The result of this paper provides another affirmative evidence to the problem for Bott manifolds. One can consider the real analogue of Bott towers and Bott manifolds, but the cohomological rigidity for \emph{real} Bott manifolds is established with $\mathbb{Z}/2$-coefficients, see \cite{Ka-Ma-2009} and \cite{Masuda-2008}. This paper is organized as follows. In Section \ref{section:Bott manifolds}, we review Bott manifolds and prepare several lemmas to prove our main theorems. We find all diffeomorphism types of $\mathbb{Q}$-trivial Bott manifolds in Section \ref{section : Q-trivial Bott manifold} and prove the cohomological rigidity for $\mathbb{Q}$-trivial Bott manifolds in Section \ref{section: cohomological rigidity of Q-trivial}. Section~\ref{sect:auto} is devoted to proving that any automorphism of the cohomology ring of a $\mathbb{Q}$-trivial Bott manifold is induced by a diffeomorphism. Throughout this paper, cohomology is taken with $\mathbb{Z}$-coefficient unless otherwise stated. \section{Cohomology of Bott manifolds} \label{section:Bott manifolds} We begin with recalling some general facts on projective bundles. Let $\pi\colon E\to B$ be a complex vector bundle over a smooth manifold $B$ and let $P(E)$ be the projectivization of $E$. \begin{lemma}\cite[Lemma 2.1]{Ch-Ma-Su-2010} \label{lemm:line} Let $B$ and $E$ be as above and let $L$ be a complex line bundle over $B$. We denote by $E^*$ the complex vector bundle dual to $E$. Then both $P(E^*)$ and $P(E\otimes L)$ are isomorphic to $P(E)$ as fiber bundles over $B$, in particular, they are diffeomorphic. \end{lemma} \begin{proof} We shall reproduce the proof given in \cite{Ch-Ma-Su-2010} for the reader's convenience sake. Choose a Hermitian metric $\langle\ ,\ \rangle$ on $E$, which is anti-$\mathbb{C}$-linear on the first entry and $\mathbb{C}$-linear on the second entry, and define a map $\tilde b\colon E\to E^*$ by $\tilde b(u):=\langle u,\ \rangle$. This map is not $\mathbb{C}$-linear but anti-$\mathbb{C}$-linear, so it induces a map $b\colon P(E)\to P(E^*)$, which gives an isomorphism as fiber bundles. For each $x\in B$, we choose a non-zero vector $v_x$ from the fiber of $L$ over $x$ and define a map $\tilde c\colon E\to E\otimes L$ by $\tilde c(u_x):=u_x\otimes v_x$ where $u_x$ is an element of the fiber of $E$ over $x$. The map $\tilde c$ depends on the choice of $v_x$'s but the induced map $c\colon P(E)\to P(E\otimes L)$ does not because $L$ is a line bundle. It is easy to check that $c$ gives an isomorphism of $P(E)$ and $P(E\otimes L)$ as fiber bundles over $B$. \end{proof} \begin{remark} \label{rema:line} The bundle map $b\colon P(E)\to P(E^*)$ does not preserve the canonical complex structures on the fibers and the pullback of the tautological line bundle over $P(E^*)$ by $b$ is complex conjugate to the tautological line bundle over $P(E)$ since $\tilde b$ is anti-$\mathbb{C}$-linear. On the other hand, the bundle map $c\colon P(E)\to P(E\otimes L)$ above preserves the canonical complex structure on the fibers and pulls back the tautological line bundle over $P(E\otimes L)$ to that over $P(E)$. \end{remark} If $H^{odd}(B)=0$ (and this is the case for Bott manifolds), then $H^*(P(E))$ is a free module over $H^*(B)$ via $\pi^*\colon H^*(B)\to H^*(P(E))$ and the Borel-Hirzebruch formula \cite[(2) on p.515]{bo-hi58} tells us that \begin{equation} \label{eqn:BH} H^\ast(P(E)) = H^\ast(B)[x]/\big(\sum_{i=0}^m (-1)^ic_{i}(E)x^{m-i}\big), \end{equation} where $m$ is the fiber dimension of $E$, $c_i(E)$ denotes the $i$-th Chern class of $E$, and $x$ denotes the first Chern class of the tautological line bundle over $P(E)$. Moreover, the tangent bundle $T_fP(E)$ along the fibers of $P(E)\to B$ admits a canonical complex structure since each fiber is a complex projective space, and with this complex structure its total Chern class is given by \begin{equation} \label{eqn:tf} c(T_fP(E))=\sum_{i=0}^m(1-x)^{m-i}c_i(E). \end{equation} Now we consider the Bott tower \eqref{eqn:Bott tower}. Each fiber bundle $\pi_j\colon B_j\to B_{j-1}$ for $j=1,\dots,n$ is the projectivization of a Whitney sum of two complex line bundles by definition and we may assume that one of the two line bundles is trivial by Lemma~\ref{lemm:line}. Therefore, one can express \[ \text{$B_j=P(\underline{\mathbb{C}}\oplus\gamma^{\alpha_j})$ with $\alpha_j\in H^2(B_{j-1})$,} \] where $\underline{\mathbb{C}}$ denotes the trivial complex line bundle and $\gamma^{\alpha_j}$ denotes the complex line bundle over $B_{j-1}$ with $\alpha_j$ as the first Chern class. Note that $\alpha_1=0$ since $B_0$ is a point. Let $x_j$ be the first Chern class of the tautological line bundle over $B_j$. Then it follows from \eqref{eqn:BH} that $$ H^\ast(B_j) = H^\ast(B_{j-1})[x_j]/ \big(x_j^2 = \alpha_j x_j\big). $$ Using this formula inductively on $j$ and regarding $H^*(B_j)$ as a graded subring of $H^*(B_n)$ through the projections in \eqref{eqn:Bott tower}, we see that \begin{equation} \label{eqn:HBn} H^*(B_n)=\mathbb{Z}[x_1, \ldots, x_n]/\big(x_j^2 = \alpha_jx_j\mid j=1,\dots,n\big). \end{equation} Sometimes it is convenient and helpful to express \[ \alpha_j=\sum_{i=1}^{j-1} A^i_j x_i \quad\text{with $A^i_j\in \mathbb{Z}$} \] and form an upper triangular matrix of size $n$ with zero diagonals: $$ A=\left( \begin{array}{ccccc} 0 & A^1_2 & A^1_3 &\cdots & A^1_n \\ & 0 & A^2_3 & \cdots & A^2_n \\ & & \ddots & \ddots& \vdots \\ & & &0 &A^{n-1}_n\\ & & & &0 \end{array} \right). $$ Let $S^1$ and $S^3$ denote the unit sphere of $\mathbb{C}$ and $\mathbb{C}^2$ respectively. Using the matrix $A$, one can describe $B_n$ as the quotient of $(S^3)^n$ by a free action of $(S^1)^n$ defined by \begin{equation} \label{eqn:quotient} \begin{split} &(t_1,\dots,t_n)\cdot \big((z_1,w_1),\dots,(z_j,w_j),\dots,(z_n,w_n)\big)\\ =&\big((t_1z_1,t_1w_1),\dots,(t_jz_j,(\prod_{i=1}^{j-1}t_i^{-A^i_j})t_jw_j),\dots,(t_nz_n,(\prod_{i=1}^{n-1}t_i^{-A^i_n})t_nw_n)\big) \end{split} \end{equation} where $(t_1,\dots,t_n)\in (S^1)^n$ and $(z_j,w_j)$ denotes the coordinate of the $j$th component of $(S^3)^n$. In fact, the projections $$(S^3)^n\to (S^3)^{n-1}\to\dots \to S^3\to \text{\{a point\}}$$ defined by dropping the last factor at each stage induces the Bott tower \eqref{eqn:Bott tower}. The next lemma and corollary are tricks to simplify algebraic computations. An ordered pair $(z, \bar{z})$ of elements in $H^2(B_n)$ is said to be \emph{vanishing} if $z\bar{z} = 0$ and \emph{primitive} if both $z$ and $\bar{z}$ are primitive. Note that $(x_j,x_j-\alpha_j)$ is a primitive vanishing pair for each $j$ since $x_j^2=\alpha_jx_j$. \begin{lemma}\label{lemma:masuda's tricky} A primitive vanishing pair $(z, \bar{z})$ is of the form $$ (ax_j+ u, \pm (a(x_j - \alpha_j) - u)) $$ for some $j$, where $a$ is a non-zero integer, $u$ is a linear combination of $x_i$'s with $i<j$, and $u(u+a \alpha_j) = 0$. \end{lemma} \begin{proof} Set $z= ax_j + u$ (resp. $\bar{z} = bx_k + v$), where $a$ (resp. $b$) is a non-zero integer and $u$ (resp. $v$) is a linear combination of $x_i$'s with $i<j$ (resp. $i<k$). If $k \neq j$, then $ab x_j x_k$ term in $z\bar{z}$ survives in $H^\ast(B_n)$ because of \eqref{eqn:HBn}, hence $k=j$. Therefore, \begin{equation} \label{eqn:zz} 0=z\bar{z} = abx_j^2 + (av + bu)x_j + uv = (ab \alpha_j + av + bu)x_j + uv. \end{equation} Since $u$ and $v$ are linear combinations of $x_i$'s with $i<j$, the identity \eqref{eqn:zz} implies that \begin{equation} \label{eqn:vanish} \text{$ab \alpha_j + av +bu =0$\quad and\quad $uv = 0$.} \end{equation} The former identity in \eqref{eqn:vanish} shows that $bu$ is divisible by $a$. However $u$ is not divisible by any nontrivial factor of $a$ since $z=ax_j+u$ is primitive. Hence $a | b$. Similarly, $av$ is divisible by $b$ and hence $b | a$. Therefore, $b = \pm a$ and hence $v = \mp (u +a\alpha_j)$ by the former identity of \eqref{eqn:vanish}. This proves the first statement in the lemma because $\bar z=bz_j+v$. The last identity in the lemma follows from the latter identity of \eqref{eqn:vanish} since $v=u+a\alpha_j$ up to sign. \end{proof} \begin{corollary}\label{corollary: squared zero elements} A square zero primitive element in $H^2(B_n)$ is either $x_j - \frac{1}{2}\alpha_j$ or $2x_j - \alpha_j$ up to sign for some $j$, where $\alpha_j^2 = 0$ in both cases. In particular, the number of square zero primitive elements in $H^2(B_n)$ up to sign is equal to the number of $\alpha_j$'s with $\alpha_j^2 = 0$. \end{corollary} \begin{proof} Since $z = \bar{z}$ in the proof of Lemma~\ref{lemma:masuda's tricky}, either $2u = -a \alpha_j$ or $2x_j = \alpha_j$. But the latter case does not occur since $\alpha_j$ is a linear combination of $x_i$'s with $i<j$. Hence, $2u = -a\alpha_j$. Thus, it follows from the primitiveness of $z$ that $z$ must be either $x_j-\frac{1}{2}\alpha_j$ or $2x_j - \alpha_j$ up to sign. Since $u(u+a\alpha_j)=0$ and $2u=-a\alpha_j$, we have $\alpha_j^2=0$, proving the corollary. \end{proof} \section{$\mathbb{Q}$-trivial Bott manifolds} \label{section : Q-trivial Bott manifold} The purpose of this section is to classify $\mathbb{Q}$-trivial Bott manifolds. We freely use the notation in Section~\ref{section:Bott manifolds}. \begin{proposition}\label{proposition:Q-trivial Bott manifold} $B_n$ is $\mathbb{Q}$-trivial if and only if $\alpha_j^2 = 0$ in $H^*(B_n)$ for all $j=1, \ldots, n$. In particular, if $B_n$ is $\mathbb{Q}$-trivial, then every Bott manifold $B_j$ in the tower \eqref{eqn:Bott tower} is $\mathbb{Q}$-trivial. \end{proposition} \begin{proof} If $\alpha_j^2 =0$, then $( x_j - \frac{\alpha_j}{2})^2=0$ in $H^\ast(B_n; \mathbb{Q})$ because $x_j^2=\alpha_j x_j$. Since $ x_j - \frac{\alpha_j}{2}$ for $j=1,\dots,n$ generate $H^*(B_n;\mathbb{Q})$ as a graded ring, this shows that $B_n$ is $\mathbb{Q}$-trivial. Conversely, if $B_n$ is $\mathbb{Q}$-trivial, there are $n$ primitive elements in $H^2(B_n)$ up to sign whose square vanish. By Corollary \ref{corollary: squared zero elements}, the number of $\alpha_j$'s whose square vanish is also $n$, which implies the converse. \end{proof} \begin{example} \label{example:Hirzebruch surface} For $a\in \mathbb{Z}$, let $\Sigma_a = P(\underline{\mathbb{C}} \oplus \gamma^{ax_1})$, where $\gamma^{ax_1}$ is the complex line bundle over $\mathbb{C}P^1=B_1$ whose first Chern class is $ax_1\in H^2(\mathbb{C} P^1)$. $\Sigma_a$ is called a \emph{Hirzebruch Surface}, which was first studied by Hirzebruch in \cite{Hirzebruch-1951}. Note that $$ H^\ast(\Sigma_a;\mathbb{Z}) = \mathbb{Z}[x_1, x_2]/ (x_1^2=0,\ x_2^2 =ax_1), $$ so that $\alpha_1 = 0$ and $\alpha_2 = ax_1$ in this case. Since the squares of $\alpha_1$ and $\alpha_2$ are both $0$, $\Sigma_a$ is $\mathbb{Q}$-trivial. As is well-known, $\Sigma_a$ is diffeomorphic to $\mathbb{C}P^1 \times \mathbb{C}P^1$ if $a$ is even and to $\mathbb{C}P^2 \sharp \overline{\mathbb{C}P^2}$ if $a$ is odd. \end{example} Denote $\mathcal{H}_1 = \mathbb{C}P^1, \mathcal{H}_2 = \Sigma_1$ and let $\pi_2 : \mathcal{H}_2 \to \mathcal{H}_1$ be the canonical projection. We consider the pullback bundle $\pi_3 : \mathcal{H}_3 \to \mathcal{H}_2$ of $\pi_2:\mathcal{H}_2 \to \mathcal{H}_1$ via $\pi_2$; \begin{equation} \label{eqn:H32} \begin{CD} \mathcal{H}_3 @>\rho_3>> \mathcal{H}_2 = P(\underline{\mathbb{C}} \oplus \gamma^{x_1}) \\ @VV{\pi_3}V @VV{\pi_2}V\\ \mathcal{H}_2 = P(\underline{\mathbb{C}} \oplus \gamma^{x_1}) @>{\pi_2}>> \mathcal{H}_1 = \mathbb{C}P^1 \end{CD} \end{equation} where $\rho_3$ denotes the induced bundle map. Then $\mathcal{H}_3$ is a $3$-stage Bott manifold, in fact, $\mathcal{H}_3 = P(\underline{\mathbb{C}} \oplus \gamma^{x_1})$ where $\underline{\mathbb{C}}$ and $\gamma^{x_1}$ are both regarded as complex line bundles over $\mathcal{H}_2$. Therefore, the matrix corresponding to the Bott tower $$\mathcal{H}_3\stackrel{\pi_3}\longrightarrow \mathcal{H}_2\stackrel{\pi_2}\longrightarrow \mathcal{H}_1\stackrel{\pi_1}\longrightarrow \{\text{a point}\}$$ is given by $$ \left( \begin{array}{ccc} 0 & 1 & 1 \\ & 0 & 0 \\ & & 0 \\ \end{array} \right). $$ Since the pullback of the tautological line bundle over $\mathcal{H}_2$ by $\rho_3$ in \eqref{eqn:H32} is the tautological line bundle over $\mathcal{H}_3$, we have $\rho_3^*(x_2)=x_3$, while $\rho_3^*(x_1)=x_1$ which follows from the commutativity of the diagram \eqref{eqn:H32}. Inductively, we shall define $\mathcal{H}_n$ as follows: \begin{equation} \label{eqn:suyoung_tower} \begin{CD} \mathcal{H}_n @>\rho_n>> \mathcal{H}_{n-1} @>\rho_{n-1}>> \dots @>\rho_4>> \mathcal{H}_3 @>\rho_3>> \mathcal{H}_2\\ @VV{\pi_n}V @VV{\pi_{n-1}}V @. @VV{\pi_3}V @VV{\pi_2}V\\ \mathcal{H}_{n-1} @>{\pi_{n-1}}>> \mathcal{H}_{n-2} @>{\pi_{n-2}}>> \dots @>\pi_3>> \mathcal{H}_2 @>{\pi_2}>> \mathcal{H}_1. \end{CD} \end{equation} Note that \begin{equation} \label{eqn:rhon} \mathcal{H}_n\stackrel{\pi_n}\longrightarrow \mathcal{H}_{n-1}\stackrel{\pi_{n-1}}\longrightarrow \dots \stackrel{\pi_2}\longrightarrow \mathcal{H}_1 \stackrel{\pi_1}\longrightarrow \{\text{a point}\} \end{equation} is a Bott tower of height $n$ corresponding to the $n\times n$-matrix \begin{equation} \label{eqn:matrixHn} \left( \begin{array}{ccccc} 0 & 1 & 1 &\cdots & 1 \\ & 0 & 0 & \cdots & 0 \\ & & 0 & \cdots & 0 \\ & & & \ddots & \vdots \\ & & & & 0 \\ \end{array} \right) \end{equation} and \begin{equation} \label{eqn:Hn} H^\ast(\mathcal{H}_n) = \mathbb{Z}[x_1, \ldots, x_n]/ (x_1^2 =0,\ x_j^2 =x_1x_j \text{ for }j=2, \ldots, n), \end{equation} so that $\alpha_1=0$ and $\alpha_j = x_1$ for all $j=2, \ldots, n$. Since $\alpha_j^2 = 0$ for any $j$, $\mathcal{H}_n$ is a $\mathbb{Q}$-trivial Bott manifold by Proposition \ref{proposition:Q-trivial Bott manifold}. We also note that $\rho_j\colon \mathcal{H}_j\to \mathcal{H}_{j-1}$ $(j>2)$ is a bundle map and pulls back the tautological line bundle over $\mathcal{H}_{j-1}$ to that of $\mathcal{H}_j$, so that \begin{equation} \label{eqn:rhoj} \begin{split} &\rho_j^*(x_{j-1})=x_j \quad\text{for $j>2$, while}\\ &\rho_j^*(x_1)=x_1\quad \text{by the commutativity of \eqref{eqn:suyoung_tower}.} \end{split} \end{equation} \begin{lemma} \label{lemm:mod2} Square zero primitive elements in $H^2(\mathcal{H}_n)$ are \[ \text{$\pm x_1$ and $\pm(2x_j-x_1)$ for $j>1$.} \] In particular, their mod 2 reductions are equal to the mod 2 reduction of $x_1$. \end{lemma} \begin{proof} Since $\alpha_1=0$ and $\alpha_j=x_1$ for $j>1$ in \eqref{eqn:Hn}, the lemma is an immediate consequence of Corollary~\ref{corollary: squared zero elements}. \end{proof} Note that the mod 2 reduction of a square zero element of $H^2(\mathcal{H}_n)$ is either zero or equal to the mod 2 reduction of $x_1$ by Lemma~\ref{lemm:mod2}. \begin{lemma} \label{lemma:bundle over H_n} If $\alpha$ is a square zero element in $H^2(\mathcal{H}_n)$, then \[ P(\underline{\mathbb{C}}\oplus\gamma^\alpha)\cong \begin{cases} P(\underline{\mathbb{C}}\oplus\underline{\mathbb{C}})=\mathcal{H}_n\times \mathcal{H}_1 \quad &\text{if $\alpha=0$ in $H^2(\mathcal{H}_n)\otimes\mathbb{Z}/2$,}\\ P(\underline{\mathbb{C}}\oplus\gamma^{x_1})=\mathcal{H}_{n+1} \quad&\text{if $\alpha=x_1$ in $H^2(\mathcal{H}_n)\otimes\mathbb{Z}/2$,} \end{cases} \] as bundles over $\mathcal{H}_n$. \end{lemma} \begin{proof} By Lemma~\ref{lemm:mod2}, $\alpha$ is either $ax_1$ or $a(2x_j-x_1)$ for $j>1$, where $a$ is an integer. Thus it suffices to prove \begin{enumerate} \item \label{item:1} $P(\gamma^{ax_1}\oplus \underline{\mathbb{C}})\cong P(\gamma^{(a+2b)x_1} \oplus \underline{\mathbb{C}})$ as bundles for any $b \in \mathbb{Z}$, \item \label{item:2} $P(\gamma^{a( 2x_j-x_1)} \oplus \underline{\mathbb{C}})\cong P(\gamma^{-ax_1} \oplus \underline{\mathbb{C}})$ as bundles for any $j>1$. \end{enumerate} We first prove (1). By Lemma~\ref{lemm:line} we have \begin{equation*} P(\gamma^{ax_1}\oplus \underline{\mathbb{C}})\cong P((\gamma^{ax_1}\oplus \underline{\mathbb{C}})\otimes \gamma^{bx_1})=P(\gamma^{(a+b)x_1}\oplus\gamma^{bx_1})\quad\text{as bundles}. \end{equation*} Therefore it suffices to prove \begin{equation} \label{eqn:(1)} P(\gamma^{(a+b)x_1}\oplus\gamma^{bx_1})\cong P(\gamma^{(a+2b)x_1}\oplus\underline{\mathbb{C}})\quad\text{as bundles}. \end{equation} All line bundles involved in \eqref{eqn:(1)} are the pullback of line bundles over $\mathcal{H}_1$ by a composition of the projections $\pi_i$'s in the tower \eqref{eqn:rhon}. Therefore it suffices to prove \eqref{eqn:(1)} when the base space is $\mathcal{H}_1$. But then the two vector bundles $\gamma^{(a+b)x_1}\oplus\gamma^{bx_1}$ and $\gamma^{(a+2b)x_1}\oplus\underline{\mathbb{C}}$ in \eqref{eqn:(1)} are isomorphic because their total Chern classes are same and complex vector bundles over $\mathcal{H}_1=\mathbb{C}P^1$ are classified by their total Chern classes as is well-known. The proof of (2) is similar to that of (1). By Lemma~\ref{lemm:line} we have \begin{equation*} P(\gamma^{a(2x_j-x_1)}\oplus \underline{\mathbb{C}})\cong P((\gamma^{a(2x_j-x_1)}\oplus \underline{\mathbb{C}})\otimes \gamma^{-ax_j})=P(\gamma^{a(x_j-x_1)}\oplus\gamma^{-ax_j}). \end{equation*} Therefore it suffices to prove \begin{equation} \label{eqn:(2)} P(\gamma^{a(x_j-x_1)}\oplus\gamma^{-ax_j})\cong P(\gamma^{-ax_1}\oplus\underline{\mathbb{C}})\quad\text{as bundles}. \end{equation} As remarked at \eqref{eqn:rhoj}, $\rho_i\colon \mathcal{H}_i\to \mathcal{H}_{i-1}$ for $i>2$ is a bundle map and pulls back the tautological line bundle over $\mathcal{H}_{i-1}$ to that over $\mathcal{H}_i$ so that $\rho_i^*(x_{i-1})=x_i$. Therefore $\gamma^{x_j}$ is the pullback of $\gamma^{x_2}$ over $\mathcal{H}_2$ by a composition of the bundle maps $\rho_i$'s. Moreover $\rho_i^*(x_1)=x_1$ as noted before. Therefore it suffices to prove \eqref{eqn:(2)} when $j=2$ and the base space is $\mathcal{H}_2$. But then the two vector bundles $\gamma^{a(x_j-x_1)}\oplus\gamma^{-ax_j}$ and $\gamma^{-ax_1}\oplus\underline{\mathbb{C}}$ in \eqref{eqn:(2)} are isomorphic because their total Chern classes are same and complex vector bundles of complex dimension two over $\mathcal{H}_2$ are classified by their total Chern classes. In fact the last assertion follows from an exact sequence $$ [\mathcal{H}_2,U/U(2)] \to [\mathcal{H}_2,BU(2)] \to [\mathcal{H}_2,BU]=K(\mathcal{H}_2) $$ induced from a fibration $U/U(2)\to BU(2)\to BU$. Here $[\mathcal{H}_2,U/U(2)]=0$ because $\mathcal{H}_2$ is of real dimension 4 and $U/U(2)$ is 4-connected and $K(\mathcal{H}_2)$ is torsion free since $H^{odd}(\mathcal{H}_2)=0$, so that elements in $[\mathcal{H}_2,BU(2)]$ can be distinguished by their Chern classes. \end{proof} \section{Cohomological rigidity of $\mathbb{Q}$-trivial Bott manifolds} \label{section: cohomological rigidity of Q-trivial} For $n \in \mathbb{N}$, a finite sequence $\lambda = (\lambda_1, \ldots, \lambda_m )$ of positive integers is called a \emph{partition} of $n$ if $\sum_{1 \leq i \leq m} \lambda_i = n$ and $\lambda_1 \geq \cdots \geq \lambda_m \geq 1$. We define $\mathcal{H}_\lambda$ by $$ \mathcal{H}_\lambda := \mathcal{H}_{\lambda_1} \times \cdots \times \mathcal{H}_{\lambda_m}. $$ For instance, $(\mathbb{C}P^1)^n$ is $\mathcal{H}_{(1, \ldots, 1)}$ and $\mathcal{H}_n$ is $\mathcal{H}_{(n)}$. Note that \begin{equation} \label{eqn:tensor} \text{$H^\ast(\mathcal{H}_{\lambda})=H^\ast(\mathcal{H}_{\lambda_1}) \otimes \cdots \otimes H^\ast(\mathcal{H}_{\lambda_m})$.} \end{equation} \begin{theorem} \label{theorem:Q-trivial Bott manifold} \begin{enumerate} \item An $n$-stage $\mathbb{Q}$-trivial Bott manifold is diffeomorphic to $\mathcal{H}_{\lambda}$ for some partition $\lambda$ of $n$. \item Let $\lambda$ and $\lambda'$ be two partitions of $n$. If $H^\ast(\mathcal{H}_\lambda)$ is isomorphic to $H^\ast(\mathcal{H}_{\lambda'})$ as graded rings, then $\lambda = \lambda'$. \end{enumerate} \noindent Therefore, $\mathbb{Q}$-trivial Bott manifolds are distinguished by their cohomology rings with $\mathbb{Z}$-coefficients and the number of diffeomorphism classes in $n$-stage Bott manifolds is equal to the number of partitions of $n$. \end{theorem} \begin{proof} (1) We prove the statement (1) by induction on $n$. Let $B_n$ be an $n$-stage Bott manifold in the tower \eqref{eqn:Bott tower} and suppose that $B_n$ is $\mathbb{Q}$-trivial. When $n=1$, the statement is trivial since $B_1=\mathbb{C} P^1=\mathcal{H}_1$. Assume the statement (1) holds for $(n-1)$-stage $\mathbb{Q}$-trivial Bott manifolds. Then, since $B_{n-1}$ is also $\mathbb{Q}$-trivial by Proposition~\ref{proposition:Q-trivial Bott manifold}, we may assume that $B_{n-1}=\mathcal{H}_\mu$ for some partition $\mu$ of $n-1$ by the induction assumption and $B_n = P(\gamma^{\alpha_n} \oplus \underline{\mathbb{C}})$ with $\alpha_n \in H^2(\mathcal{H}_{\mu})$. We note that $\alpha_n^2 = 0$ by Proposition~\ref{proposition:Q-trivial Bott manifold} because $B_n$ is $\mathbb{Q}$-trivial. If $\alpha_n=0$, then $B_n=\mathcal{H}_\mu\times \mathcal{H}_1$ and the theorem holds in this case. Suppose $\alpha_n\not=0$. Then $\alpha_n$ must sit in $H^2(\mathcal{H}_{\mu_j})$ for some component $\mu_j$ of the partition $\mu$ in \eqref{eqn:tensor} with $\lambda$ replaced by $\mu$ because otherwise $\alpha_n^2$ cannot vanish. Therefore the line bundle $\gamma^{\alpha_n}$ over $\mathcal{H}_\mu$ can be obtained by pulling back a line bundle over $\mathcal{H}_{\mu_j}$. It follows that $B_n$ is diffeomorphic to \[ P(\gamma^{\alpha_n}\oplus\underline{\mathbb{C}})\times \prod_{i\not=j}\mathcal{H}_{\mu_i} \] where $\gamma^{\alpha_n}$ is regarded as a line bundle over $\mathcal{H}_{\mu_j}$, $\mu_i$ runs over all components of $\mu$ different from $\mu_j$. Then the statement (1) follows from Lemma~\ref{lemma:bundle over H_n}. (2) Any (non-zero) square zero element in $H^2(\mathcal{H}_\lambda)$ sits in $H^2(\mathcal{H}_{\lambda_i})$ for some component $\lambda_i$ of $\lambda$ as noted above and it follows from Lemma~\ref{lemm:mod2} that the mod 2 reductions of a square zero primitive element in $H^2(\mathcal{H}_{\lambda_i})$ and that in $H^2(\mathcal{H}_{\lambda_j})$ are same if and only if $i=j$. Therefore, if $\varphi : H^\ast(\mathcal{H}_{\lambda}) \to H^\ast(\mathcal{H}_{\lambda'})$ is a graded ring homomorphism, then all square zero primitive elements in $H^2(\mathcal{H}_{\lambda_i})$ map into $H^2(\mathcal{H}_{\lambda'_j})$ by $\varphi$ for some component $\lambda'_j$ of $\lambda'$. Since the square zero primitive elements in $H^2(\mathcal{H}_{\lambda_i})$ generate $H^*(\mathcal{H}_{\lambda_i})$ over $\mathbb{Q}$, this implies that $\varphi(H^*(\mathcal{H}_{\lambda_i}))$ is contained in $H^*(\mathcal{H}_{\lambda'_j})$. If $\varphi$ is in particular an isomorphism, then this together with \eqref{eqn:tensor} implies the statement (2). \end{proof} \begin{remark} One can show that $\mathcal{H}_\lambda$'s, in other words $\mathbb{Q}$-trivial Bott manifolds, can be distinguished by their cohomology rings even with $\mathbb{Z}/2$- or $\mathbb{Z}_{(2)}$-coefficients. It is not true that all Bott manifolds can be distinguished by their cohomology rings with $\mathbb{Z}/2$-coefficients (e.g. 3-stage Bott manifolds are such examples, see \cite{Ch-Ma-Su-2010}), but it might be true with $\mathbb{Z}_{(2)}$-coefficients, see \cite{ch-su-pre}. \end{remark} \section{Automorphisms of $\mathbb{Q}$-trivial Bott manifolds} \label{sect:auto} By Theorem~\ref{theorem:Q-trivial Bott manifold} we may assume that an $n$-stage Bott manifold is $\mathcal{H}_\lambda$ where $\lambda$ is a partition of $n$. In this section we shall study the group $\Aut(H^*(\mathcal{H}_\lambda))$ of graded ring automorphisms of $H^*(\mathcal{H}_\lambda)$ and prove the following. \begin{theorem} \label{theo:diffeo} Any element of $\Aut(H^*(\mathcal{H}_\lambda))$ is induced from a diffeomorphism of $\mathcal{H}_\lambda$. \end{theorem} Since $\mathbb{Q}$-trivial Bott manifolds are distinguished by their cohomology rings by Theorem~\ref{theorem:Q-trivial Bott manifold}, the theorem above implies the following. \begin{corollary} \label{coro:diffeo} Any cohomology ring isomorphism between two $\mathbb{Q}$-trivial Bott manifolds is induced from a diffeomorphism. \end{corollary} The rest of this section is devoted to the proof of Theorem~\ref{theo:diffeo}. Remember that the square zero primitive elements in $H^2(\mathcal{H}_n)$ are $\pm x_1$ and $\pm(2x_j-x_1)$ for $j>1$ by Lemma~\ref{lemm:mod2}. \begin{lemma} An automorphism of $H^*(\mathcal{H}_n)$ permutes $\pm x_1$ and $\pm(2x_j-x_1)$ for $j>1$ up to sign. On the other hand, any permutation of $\pm x_1$ and $\pm(2x_j-x_1)$ for $j>1$ up to sign induces an automorphism of $H^*(\mathcal{H}_n)$. Therefore, $\Aut(H^*(\mathcal{H}_n))$ is isomorphic to a semi-direct product $(\mathbb{Z}/2)^n\rtimes \frak S_n$ where $\frak S_n$ denotes the symmetric group on $n$ letters and the action of $\frak S_n$ on $(\mathbb{Z}/2)^n$ is the natural permutation of factors of $(\mathbb{Z}/2)^n$. \end{lemma} \begin{proof} The first statement is obvious. Suppose that $\varphi$ is a permutation of $\pm x_1$ and $\pm(2x_j-x_1)$ for $j>1$ up to sign. Then $\varphi(x_1)=\pm x_1$ or $\pm(2x_k-x_1)$ for some $k>1$. In any case one can easily check that if we extend $\varphi$ linearly, then $\varphi(x_i)$ is integral (i.e., a linear combination of $x_\ell$'s over $\mathbb{Z}$) for any $i$. For instance, if \[ \varphi(x_1)=2x_k-x_1,\ \varphi(2x_i-x_1)=x_1,\ \varphi(2x_j-x_1)=-(2x_\ell-x_1) \text{ for $j\not=i$}, \] then a simple computation shows that \[ \varphi(x_i)=x_k\text{ and } \varphi(x_j)=x_k-x_\ell. \] Thus the linear extension of $\varphi$ defines an endomorphism of $H^2(\mathcal{H}_n)$. Moreover, one can also check that $\varphi(x_1)^2=0$ and $\varphi(x_j)^2=\varphi(x_1)\varphi(x_j)$ for $j>1$. This ensures that $\varphi$ extends to a graded ring endmorphism $\overline{\varphi}$ of $H^*(\mathcal{H}_n)$ since the ideal in \eqref{eqn:Hn} is generated by $x_1^2$ and $x_j^2-x_1x_j$ for $j>1$. Similarly, $\varphi^{-1}$ induces a graded ring endomorphism $\overline{\varphi^{-1}}$ of $H^*(\mathcal{H}_n)$ and clealry $\overline{\varphi^{-1}}$ gives the inverse of $\overline{\varphi}$, so $\overline{\varphi}$ is an automorphism of $H^*(\mathcal{H}_n)$. This proves the lemma. \end{proof} We write $\lambda=(d_1^{a_1},\dots,d_k^{a_k})$ where $d_1>\dots>d_k$ and $d_i^{a_i}$ denotes $a_i$ copies of $d_i$ for $i=1,\dots,k$. Then \[ H^*(\mathcal{H}_\lambda)=\bigotimes_{i=1}^k H^*(\mathcal{H}_{d_i})^{\otimes a_i}. \] The proof of (2) in Theorem~\ref{theorem:Q-trivial Bott manifold} shows that an automorphism of $H^*(\mathcal{H}_\lambda)$ maps factors of $H^*(\mathcal{H}_{d_i})^{\otimes a_i}$ to themselves for each $i$, so that \begin{equation} \label{eqn:AutH} \Aut(H^*(\mathcal{H}_\lambda))=\prod_{i=1}^k\Aut(H^*(\mathcal{H}_{d_i})^{\otimes a_i})=\prod_{i=1}^k\Aut(H^*(\mathcal{H}_{d_i}))^{a_i}\rtimes \frak S_{a_i} \end{equation} where the action of $\frak S_{a_i}$ on $\Aut(H^*(\mathcal{H}_{d_i}))^{a_i}$ is the natural permutation of factors of $\Aut(H^*(\mathcal{H}_{d_i}))^{a_i}$. A permutation of factors of $\Aut(H^*(\mathcal{H}_{d_i}))^{a_i}$ is induced from a permutation of factors of $\mathcal{H}_{d_i}^{a_i}$, which is a diffeomorphism, so it suffices to prove Theorem~\ref{theo:diffeo} when $\lambda=(n)$ by \eqref{eqn:AutH}. We first prove it when $n=2$. \begin{lemma} \label{lemm:H2} Any element of $\Aut(H^*(\mathcal{H}_2))$, which permutes $\pm x_1$ and $\pm(2x_2-x_1)$ up to sign, is induced from a diffeomorphism of $\mathcal{H}_2$. \end{lemma} \begin{proof} As remarked in Example~\ref{example:Hirzebruch surface}, $\mathcal{H}_2=\Sigma_1$ is diffeomorphic to $\mathbb{C}P^2\#\overline{\mathbb{C}P^2}$. Let $u$ and $v$ be elements of $H_2( \mathbb{C}P^2\#\overline{\mathbb{C}P^2})$ represented by a canonical submanifold $\mathbb{C}P^1$ in $\mathbb{C}P^2$ and $\overline{\mathbb{C}P^2}$ respectively. They are a basis of $H_2( \mathbb{C}P^2\#\overline{\mathbb{C}P^2})$. (Through the Poincar\'e duality, $u$ and $v$ correspond to $x_2$ and $x_1-x_2$ up to sign since the self-intersection numbers of $u$ and $v$ are $\pm 1$ while squares of $x_2$ and $x_2-x_1$ are a cofundamental class $x_1x_2$ up to sign.) It suffices to show that any permutation of $\pm u$ and $\pm v$ up to sign can be represented by a diffeomorphism of $\mathbb{C}P^2\#\overline{\mathbb{C}P^2}=\mathcal{H}_2$ since the number of those permutations is 8 which agrees with the number of elements in $\Aut(H^*(\mathcal{H}_2))\cong (\mathbb{Z}/2)^2\rtimes \frak S_2$. We consider two involutions $s$ and $t$ on $\mathbb{C}P^2$ defined by \[ s\colon [z_1,z_2,z_3]\to [\bar{z}_1,\bar{z}_2,\bar{z}_3],\qquad t\colon [z_1,z_2,z_3]\to [z_1,z_2,-z_3] \] where $[z_1,z_2,z_3]$ denotes the homogenous coordinate of $\mathbb{C}P^2$ and $\bar{z}$ denotes the complex conjugate of a complex number $z$. Observe that \begin{enumerate} \item $s$ leaves the submanifold $\mathbb{C}P^1=\{z_3=0\}$ of $\mathbb{C}P^2$ invariant, reverses an orientation on the $\mathbb{C}P^1$ and the fixed point set of $s$ is $\mathbb{R} P^2$, \item the induced action of $t$ on $H_\ast(\mathbb{C}P^2)$ is trivial and the fixed point set of $t$ is the disjoint union of $\mathbb{C}P^1=\{z_3=0\}$ and a point $[0,0,1]$. \end{enumerate} \noindent {\bf Type 1.} We consider the involution $s$ on both $\mathbb{C}P^2$ and $\overline{\mathbb{C}P^2}$. Choose a point from the fixed set $\mathbb{R} P^2$ in $\mathbb{C}P^2$ and $\overline{\mathbb{C}P^2}$ respectively and take equivariant connected sum of $\mathbb{C}P^2$ and $\overline{\mathbb{C}P^2}$ around the chosen points. Then the resulting involution on $\mathbb{C}P^2\#\overline{\mathbb{C}P^2}$ sends $(u,v)$ to $(-u,-v)$. \noindent {\bf Type 2.} We consider the involution $s$ on $\mathbb{C}P^2$ and $t$ on $\overline{\mathbb{C}P^2}$. Choose a point from the fixed set $\mathbb{R} P^2$ in $\mathbb{C}P^2$ and a point from the fixed set $\mathbb{C}P^1$ in $\overline{\mathbb{C}P^2}$ and take equivariant connected sum of $\mathbb{C}P^2$ and $\overline{\mathbb{C}P^2}$ around the chosen points. Then the resulting involution on $\mathbb{C}P^2\#\overline{\mathbb{C}P^2}$ sends $(u,v)$ to $(-u,v)$. \noindent {\bf Type 3.} $\mathbb{C}P^2\#\overline{\mathbb{C}P^2}$ is obtained by removing an open disk $D$ from $\mathbb{C}P^2$ and $\overline{\mathbb{C}P^2}$ respectively and gluing together along the boundary $S^3$ via the identity map, so that it admits a reflection with respect to the $S^3$, which maps $\mathbb{C}P^2\backslash D$ to $\overline{\mathbb{C}P^2}\backslash D$. This reflection sends $(u,v)$ to $(v,u)$. Combining the diffeomorphisms of the three types above, one can realize any element of $\Aut(H^*(\mathcal{H}_2))$ by a diffeomorphism of $\mathcal{H}_2$. \end{proof} We shall prove that any element of $\Aut(H^*(\mathcal{H}_n))$ is induced from a diffeomorphism of $\mathcal{H}_n$ for any $n$ by induction on $n$, so that the proof of Theorem~\ref{theo:diffeo} will be completed. For that we prepare three lemmas. We regard $H^*(\mathcal{H}_{j})$ for $j<n$ as a subring of $H^*(\mathcal{H}_n)$ as usual and remember that $\pm x_1$ and $\pm(2x_j-2x_1)$ for $j>1$ are all the square zero primitive elements in $H^2(\mathcal{H}_n)$. \begin{lemma} \label{lemm:ext} Let $\psi$ be an element of $\Aut(H^*(\mathcal{H}_{j}))$ for $j<n$. If $\psi$ is induced from a diffeomorphism of $\mathcal{H}_{j}$, then there is a diffeomorphism of $\mathcal{H}_n$ whose induced automorphism of $H^*(\mathcal{H}_n)$ preserves the subring $H^*(\mathcal{H}_{j})$ and agrees with the given $\psi$ on $H^*(\mathcal{H}_j)$. \end{lemma} \begin{proof} Let $f_j$ be a diffeomorphism of $\mathcal{H}_j$ whose induced automorphism of $H^*(\mathcal{H}_j)$ is $\psi$. The pullback of the bundle \begin{equation} \label{eqn:bundle} \text{$\mathcal{H}_{j+1}=P(\underline{\mathbb{C}}\oplus \gamma^{\alpha_{j+1}})\stackrel{\pi_{j+1}}\longrightarrow \mathcal{H}_j$} \end{equation} by $f_j$ is of the form $P(\underline{\mathbb{C}}\oplus\gamma^{f_j^*(\alpha_{j+1})})\to \mathcal{H}_j$ but this is isomorphic to \eqref{eqn:bundle} by Lemma~\ref{lemma:bundle over H_n} since $\alpha_{j+1}^2=0=f_j^*(\alpha_{j+1})^2$ and the mod 2 reductions of $\alpha_{j+1}$ and $f_j^*(\alpha_{j+1})$ are same. It follows that there is a bundle automorphism $f_{j+1}$ of \eqref{eqn:bundle} which covers $f_j$. Since $f_{j+1}$ covers $f_j$, the automorphism $f_{j+1}^*$ of $H^*(\mathcal{H}_{j+1})$ induced by $f_{j+1}$ preserves the subring $H^*(\mathcal{H}_j)$ and agrees with $f_j^*$ on it. Repeating this argument for $f_{j+1}$ in place of $f_j$, we get a diffeomorphism $f_{j+2}$ of $\mathcal{H}_{j+2}$ which covers $f_{j+1}$ and so on. Then the last diffeomorphism $f_n$ of $\mathcal{H}_n$ is the desired one. \end{proof} \begin{lemma} \label{lemm:xn} There is a diffeomorphism of $\mathcal{H}_n$ whose induced automorphism of $H^*(\mathcal{H}_n)$ is the identity on the subring $H^*(\mathcal{H}_{n-1})$ and maps $x_n$ to $-x_n+x_1$ (equivalently maps $2x_n-x_1$ to $-(2x_n-x_1)$). \end{lemma} \begin{proof} Since the dual bundle of $\underline{\mathbb{C}}\oplus\gamma^{x_1}$ is isomorphic to $\underline{\mathbb{C}}\oplus\gamma^{-x_1}$, the proof of Lemma~\ref{lemm:line} shows that we have a bundle map \[ \text{$b\colon \mathcal{H}_n=P(\underline{\mathbb{C}}\oplus \gamma^{x_1})\to P(\underline{\mathbb{C}}\oplus\gamma^{-x_1})$} \] which covers the identity map on $\mathcal{H}_{n-1}$. The pullback of the tautological line bundle $\eta_-$ over $P(\underline{\mathbb{C}}\oplus\gamma^{-x_1})$ by $b$ is complex conjugate to the tautological line bundle $\eta_+$ over $P(\underline{\mathbb{C}}\oplus\gamma^{x_1})$ (see Remark~\ref{rema:line}); so we obtain \begin{equation} \label{eqn:c1} b^*(x)=-x_n \end{equation} where $x=c_1(\eta_-)$ and $x_n=c_1(\eta_+)$ by the definition of $x_n$. On the other hand, the proof of Lemma~\ref{lemm:line} shows that we have a bundle isomorphism \[ c\colon P(\underline{\mathbb{C}}\oplus\gamma^{-x_1})\to P((\underline{\mathbb{C}}\oplus\gamma^{-x_1})\otimes\gamma^{x_1})=P(\gamma^{x_1}\oplus\underline{\mathbb{C}})=\mathcal{H}_n \] which preserves the complex structures on each fiber. Therefore it induces a \emph{complex} vector bundle isomorphism $T_fP(\underline{\mathbb{C}}\oplus\gamma^{-x_1})\to T_fP(\gamma^{x_1}\oplus\underline{\mathbb{C}})$ between their tangent bundles along the fibers. According to the Borel-Hirzebruch formula \eqref{eqn:tf}, their first Chern classes are respectively $-2x-x_1$ and $-2x_n+x_1$, so \begin{equation} \label{eqn:c2} c^*(-2x_n+x_1)=-2x-x_1. \end{equation} Since the map $c$ covers the identity map on $\mathcal{H}_{n-1}$, $c^*(x_1)=x_1$. It follows from \eqref{eqn:c2} that $c^*(x_n)=x+x_1$. This together with \eqref{eqn:c1} shows that \begin{equation} \label{eqn:bc} b^*(c^*(x_n))=-x_n+x_1 \end{equation} because $b^*(x_1)=x_1$ which follows from the fact that $b$ covers the identity map on $\mathcal{H}_{n-1}$. The identity \eqref{eqn:bc} shows that the composition $c\circ b$ is the desired diffeomorphism. \end{proof} \begin{lemma} \label{lemm:xnxj} There is a diffeomorphism of $\mathcal{H}_n$ whose induced automorphism of $H^*(\mathcal{H}_n)$ interchanges $x_i$ and $x_j$ for $i,j>1$ and fixes $x_k$ for $k\not=i,j$. \end{lemma} \begin{proof} It suffices to show that there is a diffeomorphism $g_i$ of $\mathcal{H}_n$ for each $i>1$ whose induced automorphism of $H^*(\mathcal{H}_n)$ interchanges $x_i$ and $x_{i+1}$ and fixes $x_k$ for $k\not=i,i+1$, because the desired diffeomorphism can be obtained by composing those diffeomorphisms. Remember that $\mathcal{H}_{i+1}$ is obtained as the fiber product \[ \begin{CD} \mathcal{H}_{i+1} @>\rho_{i+1}>> \mathcal{H}_i\\ @VV\pi_{i+1}V @VV\pi_iV\\ \mathcal{H}_i @>\pi_i>> \mathcal{H}_{i-1}. \end{CD} \] Permuting the coordinates of $\mathcal{H}_i\times \mathcal{H}_i$ preserves the subset $\mathcal{H}_{i+1}$ and defines a diffeomorphism $\tau_{i+1}$ of $\mathcal{H}_{i+1}$. One notes that $\tau_{i+1}^*(x_i)=\rho_{i+1}^*(x_i)=x_{i+1}$ and $\tau_{i+1}^*(x_k)=x_k$ for $k<i$. Since $\pi_{i+1}\circ \tau_{i+1}=\pi_{i+1}$, the diffeomorphism $\tau_{i+1}$ naturally extends to a diffeomorphism $\tau_{i+2}$ of $\mathcal{H}_{i+2}$ and finally extends to a diffeomorphism $g_i$ of $\mathcal{H}_n$ because of \eqref{eqn:suyoung_tower}. Since $\tau_{i+1}^*(x_1)=x_1$, the pullback of the line bundle $\gamma^{x_1}$ over $\mathcal{H}_{i+1}$ is isomorphic to $\gamma^{x_1}$ itself. This implies that $\tau_{i+2}^*(x_{i+2})=x_{i+2}$ because $x_{i+2}$ is the first Chern class of the tautological line bundle over $P(\underline{\mathbb{C}}\oplus \gamma^{x_1})$. Therefore $g_i^*$ fixes $x_{i+2}$ since $g_i$ is an extension of $\tau_{i+2}$. Similarly, $g_i^*$ fixes $x_k$ for $k>i+1$. Thus $g_i$ is the desired diffeomorphism. \end{proof} \begin{remark} As remarked at \eqref{eqn:quotient}, one can regard $\mathcal{H}_n$ as the quotient of $(S^3)^n$ by a free action of $(S^1)^n$ associated with the matrix \eqref{eqn:matrixHn}. Then interchanging the $i$-th factor and the $j$-th factor of $(S^3)^n$ produces a desired diffeomorphism in Lemma~\ref{lemm:xnxj}. \end{remark} Now we shall prove that any element of $\Aut(H^*(\mathcal{H}_n))$ is induced from a diffeomorphism of $\mathcal{H}_n$ for any $n$ by induction on $n$. This claim is established for $n=2$ by Lemma~\ref{lemm:H2}. Suppose the claim holds for $n-1$. Let $\varphi$ be an element of $\Aut(H^*(\mathcal{H}_n))$. Then $\varphi$ permutes square zero primitive elements $\pm x_1, \pm(2x_j-x_1)$ $(j>1)$ up to sign. We distinguish three cases. {\bf Case 1.} The case where $\varphi(2x_n-x_1)=\pm(2x_n-x_1)$. In this case $\varphi$ preserves the subring $H^*(\mathcal{H}_{n-1})$ and let $\psi$ be the restriction of $\varphi$ to $H^*(\mathcal{H}_{n-1})$. By Lemma~\ref{lemm:ext} there is a diffeomorphism $f$ of $\mathcal{H}_n$ whose induced automorphism $f^*$ of $H^*(\mathcal{H}_n)$ agrees with $\psi$ on $H^*(\mathcal{H}_{n-1})$. Then the composition $(f^{-1})^*\circ \varphi$ is the identity on $H^*(\mathcal{H}_{n-1})$, so we may assume that $\varphi$ is the identity on $H^*(\mathcal{H}_{n-1})$. If $\varphi(2x_n-x_1)=2x_n-x_1$, then $\varphi$ is the identity so that it is induced from the identity diffeomorphism of $\mathcal{H}_n$. If $\varphi(2x_n-x_1)=-(2x_n-x_1)$, then $\varphi$ is induced from a diffeomorphism of $\mathcal{H}_n$ by Lemma~\ref{lemm:xn}. {\bf Case 2.} The case where $\varphi(2x_n-x_1)=\pm(2x_j-x_1)$ for some $1<j<n$. By Lemma~\ref{lemm:xnxj} there is a diffeomorphism $g$ of $\mathcal{H}_n$ whose induced automorphism $g^*$ of $H^*(\mathcal{H}_n)$ interchanges $x_j$ and $x_n$ and fixes $x_k$ for $k\not=j,n$. Therefore the composition $g^*\circ \varphi$ is an automorphism treated in Case 1, so that $g^*\circ \varphi$ is induced from a diffeomorphism of $\mathcal{H}_n$ by Case 1 and hence so is $\varphi$. {\bf Case 3.} The case where $\varphi(2x_n-x_1)=\pm x_1$. By Lemma~\ref{lemm:H2} and Lemma~\ref{lemm:ext}, there is a diffeomorphism $h$ of $\mathcal{H}_n$ whose induced automorphism $h^*$ of $H^*(\mathcal{H}_n)$ maps $x_1$ to $2x_2-x_1$. Therefore the composition $h^*\circ \varphi$ is an automorphism treated in Case 2, so that it is induced from a diffeomorphism of $\mathcal{H}_n$ and hence so is $\varphi$. This completes the proof of the desired claim and hence Theorem~\ref{theo:diffeo}. \medskip {\bf Concluding remark.} The cohomological rigidity problem asks whether two toric manifolds are diffeomorphic (or homeomorphic) if their cohomology rings are isomorphic. More strongly, it is asked in \cite{Ma-Su-2008} whether any cohomology ring isomorphism between two toric manifolds is induced from a diffeomorphism. We may call this problem the \emph{strong cohomological rigidity problem} for toric manifolds. Corollary~\ref{coro:diffeo} gives a supporting evidence to the problem and the authors do not know any counterexample to the problem. \bigskip \bibliographystyle{amsplain}

https://arxiv.org/abs/1911.01151

Successive shortest paths in complete graphs with random edge weights

Consider a complete graph $K_n$ with edge weights drawn independently from a uniform distribution $U(0,1)$. The weight of the shortest (minimum-weight) path $P_1$ between two given vertices is known to be $\ln n / n$, asymptotically. Define a second-shortest path $P_2$ to be the shortest path edge-disjoint from $P_1$, and consider more generally the shortest path $P_k$ edge-disjoint from all earlier paths. We show that the cost $X_k$ of $P_k$ converges in probability to $2k/n+\ln n/n$ uniformly for all $k \leq n-1$. We show analogous results when the edge weights are drawn from an exponential distribution. The same results characterise the collectively cheapest $k$ edge-disjoint paths, i.e., a minimum-cost $k$-flow. We also obtain the expectation of $X_k$ conditioned on the existence of $P_k$.

\section{Introduction} It is a standard problem to find the shortest $s$--$t$\xspace path in a graph, i.e., the cheapest path $P_1$ between specified vertices $s$ and $t$, and its cost $X_1$, where the cost of a path is the sum of the costs of its edges. We will use the terms ``cost'' and ``weight'' interchangeably, and reserve ``length'' for the number of edges in a path. Consider the complete graph $G=K_n$ with each edge $\set{u,v}$ having weight $w(u,v)$, where the $w(u,v)$ are i.i.d.\ random variables with exponential distribution $\Exp(1)$ or uniform distribution $U(0,1)$ (we consider both versions). In this random setting, a well-known result of Janson~\cite{Janson123} is that as $n \to \infty$, \begin{equation}\label{eq:svante-X1} \frac{X_1}{\ln n / n } \pto 1 . \end{equation} We define the second cheapest path $P_2$, with cost $X_2$, to be the cheapest $s$--$t$\xspace path edge-disjoint from $P_1$, and in general define $P_k$, with cost $X_k$, to be the cheapest $s$--$t$\xspace path edge-disjoint from $P_1 \cup \cdots \cup P_{k-1}$, provided such a path exists. We also think of this as finding path $P_k$ after the preceding paths' edges have been removed. Our question is how the costs $X_k$ behave in the limit as $n \to \infty$ (this limit is implicit throughout). Our main result is the following. \begin{theorem}\label{Tmain} In the complete graph $K_n$ with i.i.d\xperiod uniform $U(0,1)$ edge weights, with $X_k$ the cost of the $k$th cheapest path, \begin{align} \frac{X_k}{2k/n + \ln n / n} & \pto 1 \label{eq:UnifMain} \end{align} uniformly for all $k \leq n-1$. That is, for any $\varepsilon>0$, asymptotically almost surely, for every $k =1,\ldots,n-1$, \begin{align}\label{Xkbounds} 1-\varepsilon &\leq \frac{X_k}{2k/n + \ln n / n} \leq 1+\varepsilon . \end{align} \end{theorem} Naturally, with $k=1$, \cref{eq:UnifMain} recovers Janson's result \cref{eq:svante-X1}, since $2/n=o(\ln n/n)$. As discussed shortly, in contrast to many cases, the result for the uniform distribution does not extend immediately to all distributions with positive density at 0. However, we have a corresponding result for exponentially distributed edge weights. Given an edge-weight distribution, let $W\os k$ be the (random) weight of the $k$th cheapest edge out of a vertex (the $k$th order statistic of $n-1$ edge weights). \begin{theorem}\label{Texp} In the complete graph $K_n$ with i.i.d\xperiod exponential edge weights with mean 1, \begin{align} \frac{X_k}{2 \E W\os k + \ln n / n} & \pto 1 \label{eq:ExpMain} \end{align} uniformly for all $k \leq n-1$. \end{theorem} \noindent We give the guiding intuition behind the formula \cref{eq:ExpMain} in \cref{sec:paths-intution}. Note that $\E W\os k =\sum_{i=1}^k \tfrac{1}{n-i}$ in the exponential case (see e.g. \cref{lemma:edge-orderstat}). In the uniform case, $\E W\os k = k/n$, so \cref{eq:UnifMain} in \cref{Tmain} can also be written as \cref{eq:ExpMain}. Rather than finding the $k$ successive cheapest paths, we may alternatively wish to find the $k$ edge-disjoint paths of \emph{collective} minimum cost. Equivalently, where every edge of $G$ has capacity~$1$, we may be interested in the minimum-cost $k$-flow from $s$ to $t$ in $G$. The following remark shows that this problem leads to essentially the same costs. (The analogous ``collective'' problem for minimum spanning trees is solved in \cite{FrJo}, and \cite{JaSoMST} shows that for MSTs, the ``successive'' version leads to strictly larger costs.) \begin{remark}\label{rmk:pathsFk} In the complete graph $K_n$ with i.i.d.\ edge weights with distribution $U(0,1)$ or exponential with mean 1, the minimum-cost $k$-flow has cost $F_k$ satisfying \begin{align} \label{kflow} \frac{F_k}{ \sum_{i=1}^k (2 \E W_{(i)}+ \ln n/n) } \pto 1 \end{align} uniformly for all $k \leq n-1$. \end{remark} As in \cref{Xkbounds}, the statement consists of high-probability upper and lower bounds. The upper bounds here, for the two models, follow immediately from the upper bounds of \cref{eq:UnifMain} and \cref{eq:ExpMain}. The lower bounds follow from the lower bound on $S_k \coloneqq \sum_{i=1}^k X_k$ (see \cref{Skdef}) in \cref{Sklower} and its analogue for the exponential case, as those bounds hold for any set of $k$ edge-disjoint paths. (The main work in \cref{lowerbound}, not needed here, is to extract lower bounds on $X_k$ from the lower bounds on $S_k$.) \begin{remark}\label{rmk:existence} $P_k$ is always defined for all $k \leq n/2$, but, at least for $n$ even, may be undefined for all $k>n/2$. \end{remark} \begin{proof} There are $n-2$ length-2 $s$--$t$\xspace paths. Any path $P_k$ can destroy (share an edge with) at most two such paths (since $P_k$ uses just one edge incident to each of $s$ and $t$). Also, the single-edge path \ensuremath{\set{s,t}}\xspace is destroyed only by the path $P_k$ consisting of just this edge. So, for $P_1,\ldots,P_{k}$ to destroy all length-1 and length-2 paths requires $k \geq (n-2)/2+1=n/2$, so for $k \leq n/2$, certainly path $P_k$ exists. Conversely, a construction described in 1892 by Lucas \cite[pp.~162--164]{Lucas}, which he attributes to Walecki, shows that a complete graph $K_{2r}$ can be decomposed into $r$ edge-disjoint Hamilton paths whose $2r$ terminals are all distinct. For $n$ even, decompose $G=K_n \setminus \set{s,t}$ in this way, then link $s$ to one ``start'' terminal of each such path and $t$ to the other ``end'' terminal, giving $(n-2)/2$ edge-disjoint $s$--$t$\xspace paths. The edge \ensuremath{\set{s,t}}\xspace gives another path, for $n/2$ paths in all. The only edges not used by these paths are a star from $s$ to the Hamilton paths' end terminals, and another star from $t$ to their start terminals, and as there are no other unused edges to connect these two stars, there is no further $s$--$t$\xspace path. With nonzero probability, the edge weights are such that $P_1,\ldots,P_{n/2}$ are these $n/2$ paths, so that $P_{n/2+1}$ does not exist. \end{proof} \cref{rmk:existence} implies that, at least for $n$ even, $\E[X_k]$ is undefined for $k > n/2$. The following theorem establishes $\E X_k$ for $k \leq n/2$, and for all $k \leq n-1$, gives the expectation conditioned on the (high-probability) event that $P_k$ exists. \begin{theorem}\label{thm:expectation} In both the uniform and exponential models, for $k \leq n-1$, a.a.s.\ $P_k$ exists, and \begin{align} \E[X_k \mid P_k \textnormal{ exists}] &=(1+o(1))(2\E W\os k + \ln n / n), \label{EX} \end{align} uniformly in $k$. \end{theorem} For $k \leq n/2$, by \cref{rmk:existence} the conditioning is null, so it is immediate from \cref{thm:expectation} that $E[X_k]=(1+o(1))(2\E W\os k + \ln n / n)$. \subsection{Intuition}\label{sec:paths-intution} The intuitive picture is that path $P_k$ should use the $k$th cheapest edges out of $s$ and $t$, whose costs are denoted $W\os k^s$ and $W\os k^t$ respectively. Then, if we ignore previous paths' use of other edges in $G\setminus \set{s,t}$, by \cref{eq:svante-X1} the opposite endpoints of these two edges should be connected by a path of cost about $\ln n/n$. This suggests that $X_k \leq W\os k^s+W\os k^t + \ln n/n$, and this is our guiding intuition. Obviously, the path $P_k$ does not have to use the $k$th cheapest edge, its middle section may cost more or less than $\ln n / n$, and as earlier paths use up edges, the costs of these middle sections may rise. It is true, though, that $\sum_{i=1}^k X_i \geq \sum_{i=1}^{k-1} \parens{W\os k^s + W\os k^t}$ (summing only to ${k-1}$ on the right-hand side to avoid doubly counting edge \ensuremath{\set{s,t}}\xspace), and we use this in proving the lower bounds on $X_k$ (in \cref{lowerbound} for uniform and \cref{ExpLB} for exponential) and, more surprisingly, in proving the upper bounds on $X_k$ for large $k$ (in \cref{largekUB} generically, the details treated in \cref{unifUB,ExpBounds}). Our upper bounds are obtained by reasoning as follows. Janson~\cite{Janson123} analyses the shortest $s$--$t$\xspace path, and shortest-path tree (SP tree or SPT) on $s$, in the randomly edge-weighted graph $G=K_n$, showing that the cost of $P_1$ is asymptotically almost surely, almost exactly $\ln n/n$. When the path $P_1$ is deleted, this prunes away a root-level branch of the SP tree. The SP tree is a uniform random tree, and using known properties of such trees (see for example~\cite{Su}) it is not hard to show that what remains of the SP tree is likely to be large; capitalising on this we can find an almost equally cheap path $P'_2$. This line of argument also shows that there remains a cheap path after deleting $P'_2$, but we need to know what happens when we delete the true second-shortest path $P_2$, and at this point the argument fails because it gives no characterisation of $P_2$, only of $P_2'$. We do know, however, that $P_2$ is cheap (no more expensive than $P'_2$), and of course uses just one edge incident to each of $s$ and $t$, and we will show that deleting \emph{any} {edge set} with these properties (including $P_2$ as a possibility) must still leave a cheap path $P'_3$, and so forth. This ``adversarial'' deletion argument is developed in \cref{adversary} to prove \cref{Tmain}. \subsection{Context} The question fits with a broad research theme on optimisation (and satisfiability) problems on random structures. The novel element here is the ``robustness'' aspect of finding cheap structures even after the cheapest has been removed, and in this we were motivated by a recent study by Janson and Sorkin~\cite{JaSoMST} of the same question for successive minimum spanning trees (MSTs), again for $K_n$ with uniform or exponential random edge weights. The results for shortest paths and MSTs are dramatically different. For MSTs, it is a celebrated result of Frieze~\cite{FriezeMST} that as $n \to \infty$ the cost of the MST $T_1$ satisfies $w(T_1) \pto \zeta(3) \eqdef \sum_{k=1}^{\infty} 1/k^3$, and~\cite{JaSoMST} shows that each subsequent tree's cost has $w(T_k) \pto \gamma_k$ with the $\gamma_k$ strictly increasing (and $2k-2\sqrt k <\gamma_k<2k+2\sqrt k$). That is very different from the case here, for paths, where for $k=o(\ln n)$ we have $X_k$ asymptotically equal to $X_1$. Further context is given in the discussion of open problems in \cref{otherModels}. \subsection{Edge weight distributions} As remarked earlier, in many contexts (including for the length $X_1$ of a shortest path) the result for any distribution with positive density at 0 follows immediately from that for the uniform distribution $U(0,1)$, but that is not the case for the successive paths considered here. \begin{remark}\label{remark:blackbox} Janson proves the $X_1$ case in the exponential model but provides standard ``black-box'' reasoning that it holds also for the uniform distribution, for any distribution with density 1 at 0 (i.e., with cumulative distribution function (CDF) $\Pr(X \leq x) = x+o(x)$ for $x \downto 0$), and, after simple rescaling, for any distribution with positive density at 0. Simply, if there is a path of cost $o(1)$ in some such model, each edge $w$ must also cost $o(1)$, and, coupling with the uniform distribution by replacing $w$ with $w'=F(w)$, with $F$ the CDF, $w' \leq (1+o(1)) w$, and thus the same path is similarly cheap in the uniform model. By the same token, if a path is cheap in any model, the same path has asymptotically the same cost in any other model, and thus the cheapest paths have asymptotically the same cost. \end{remark} \begin{remark}\label{remark:openbox} In our setting this argument does not apply: to find path $P_k$ we must know the nature of the $k-1$ previous paths; their costs are not enough. For $k=o(n)$, however, the standard argument applies within our proofs, since the proofs rely only on edges of cost $o(1)$. However, for larger $k$ there are genuine difficulties. Our argument for the exponential case, in \cref{ExpBounds}, largely parallels that for uniform but requires new calculations for the upper bound, and one new idea for the lower bound (in \cref{ExpLB}). It is not clear for what other edge-weight distributions (even those with density 1 at 0) \cref{eq:ExpMain} will hold. \end{remark} \section{Open problems} \subsection{Poisson multigraph model} The issue of possible non-existence of paths $P_k$ for $k>n/2$ (see \cref{rmk:existence}) is obviated if, as in \cite{JaSoMST}, we work in a Poisson multigraph model. Here, each pair of vertices $\set{u,v}$ of $K_n$ is joined by infinitely many edges, whose weights are drawn from a Poisson process of rate 1 (so that the cheapest $\set{u,v}$ edge has exponentially distributed cost of mean 1). By construction, in this model every $s$--$t$\xspace path is always available (possibly at a higher cost). \begin{conjecture} In the Poisson multigraph model, $ \frac{X_k}{2k/n + \ln n / n} \pto 1 $ uniformly for all $k \leq n-1$, and $ \frac{\E X_k}{2k/n + \ln n / n} \to 1 $ for all $k \leq n-1$. \end{conjecture} \noindent Actually, in this model there is no need to stop at $k=n-1$, but it is not clear how far out we can go (especially preserving uniform convergence). \subsection{Other models} \label{otherModels} Most narrowly, it would be interesting to characterise successive shortest paths that are vertex-disjoint rather than edge-disjoint, and (in the style of \Cref{rmk:pathsFk} for edge-disjoint paths) the $k$ vertex-disjoint paths of collective minimum cost. In this model, guessing that path lengths stay around $\log n$, we would expect $P_k$ to be defined up to $k$ about $n / \log n$. More broadly, it would be interesting to explore different edge-weight distributions, different structures, and different graphs. As noted earlier, we have results for uniformly and exponentially distributed edge weights, but not for arbitrary distributions. As mentioned, results for the single shortest path follow by standard arguments for any distribution with positive density near~0. For a distribution with density tending to 0 or $\infty$ at 0, shortest paths were studied in \cite{BH2012}. In particular, they consider the case when edge weights are i.i.d\xperiod and have the same distribution as $Z^{p}$, where $Z \sim \Exp(1)$ and $p>0$ is a fixed parameter; in this setting, the shortest path has length $p \ln n$ and its cost is $\ln n/n^p$ times a $p$-dependent constant. A variant where the edge-weight distribution may depend on $n$ is studied in \cite{Eckhoff}. To what distributions does \cref{Texp} extend? Restricting to distributions with positive density near~0, the arguments in \cref{ExpBounds} should immediately extend for all $k=o(n)$. For larger $k$, the ``middle'' of each path should remain short, so the issue is the edges incident on $s$ and $t$ in $P_k$. Certainly \cref{eq:ExpMain} will fail if the order statistics of edges incident to $s$ are not concentrated, for example if the edge distribution is a mixture of $U(0,1)$ and an atom at 2 or (for a continuous example) a mixture of $U(0,1)$ and the Pareto distribution with CDF $1-1/x$ for $x\geq 1$. It might be true that \cref{eq:ExpMain} holds more generally if the expectation $2\E W\os k$ is replaced by $W\os k^s+W\os k^t$. However, to obtain the needed lower bound for the exponential model (see \cref{ExpBounds}), we had to address the fact that the $k$th path does not necessarily use the edges of cost $W\os k^s$ and $W\os k^t$; we also needed exponential-specific calculations for the upper bound. One could explore other structural models. Minimum spanning trees (MSTs) have already been explored in \cite{JaSoMST} for the successive version and in \cite{FrJo} for the collective version. But for many other models the single cheapest structure is well studied but the successive and collective extensions have not been explored: this includes perfect matchings in complete bipartite graphs $K_{n,n}$ \cite{AldousAP,WastlundAP}, perfect matchings in complete graphs $K_n$ \cite{WastlundPM}, and Hamilton cycles (i.e., the Travelling Salesman Problem) in $K_n$ \cite{Was2010}. One could also consider graphs other than complete graphs, in the style of studies of the MST in a random regular graph \cite{BFM1998}, and of first-passage percolation in \ER random graphs \cite{Bhamidi} and hypercubes \cite{Martinsson}. \section{Upper bound for small \tp{$k$}{k}}\label{sec:k-small} In this section we prove the upper bound of \cref{Tmain} for all $k=o(\sqrt n)$; larger values are treated in the next section. As discussed in the introduction, we can characterise the cheapest path $P_1$ and subsequent paths that are \emph{cheap} but not necessarily \emph{cheapest}, putting us at a loss to characterise what remains on deletion of a subsequent \emph{cheapest} path. We address this in this section. Given $k$, we show a construction of a subgraph $\ensuremath{R}\xspace=R^{(k)}$ of $G$ designed so that, as we will show in turn, its $s$--$t$\xspace paths are all cheap, and no deletion of edges from $\ensuremath{R}\xspace$ subject to certain constraints can destroy all these paths. We show that the union of the $k$ shortest paths satisfies these constraints, so that there remains a cheap $s$--$t$\xspace path in $\ensuremath{R}\xspace$ and thus in $G$, and use this to prove \cref{Tmain}. \begin{figure}[h] \centering \includegraphics[width=0.7\linewidth]{Rsmall} \caption{Cartoon of a robust subgraph $\ensuremath{R}\xspace$ of $G$, showing the vertices $s$ and $t$, their respective structures $R_s$ and $R_t$ including shortest-path trees represented by triangles (some ``failed'' and thus not shown), and the cheap edges connecting triangles in $R_s$ and $R_t$. Vertices $s$ and $t$ have down-degree (number of children) $r_0$, and vertices at levels 1 and 2 (in $R_s$ and $R_t$) have down-degrees $r_1$ and $r_2$ respectively. }\label{Frobust} \end{figure} Specifically, we will define a structure $\ensuremath{R}\xspace$, sketched in \cref{Frobust}, that has many cheap and spread-out paths between $s$ and $t$, within which we will always find a cheap path. A crucial point is that each step of the construction occurs in a complete induced subgraph of $G$ of size $n-o(n)$ with all edges unconditioned. We will show, assuming that \begin{align}\label{indHyp} X_i \leq (1+\eps) \parens { \frac{2i}n+\frac{\ln n}n } \end{align} for all $i\leq k$, that the same holds for $i={k+1}$. We will do so by showing that after deleting $k$ paths, each of cost $\leq (1+\eps) (2k/n+\ln n/n)$ from $G$, some or all of whose edges may lie in $R$, there remains a path in $R$ satisfying the same cost bound, and so this must also be true of $P_{k+1}$. \medskip Consistent with this approach, and because to prove convergence in probability it suffices to consider an arbitrarily small, fixed $\varepsilon$ (see around \cref{Xkbounds}), throughout this section we assume that $\varepsilon>0$ is fixed. Thus, in the $n \to \infty$ limit implicit throughout, \begin{align} \varepsilon=\Theta(1) , \label{epsconst} \end{align} and $\varepsilon$ (and functions of $\varepsilon$) may be absorbed into the constants implicit in any Landau-notation expression. \begin{remark} \label{warning} Most of the calculations below hold for any $\varepsilon>0$, but a few (\cref{Bheavy} and \cref{pathlength} for example) hold only for $\varepsilon$ sufficiently small. This is not restrictive here, in proving convergence in probability, but to characterise expectation, \cref{expSmallk} requires $\varepsilon$ to be a large constant (to assure sufficiently small failure probabilities). The proof of \cref{lem:large-eps} addresses the changes needed. \end{remark} Before going into detail let us sketch the construction of $R$. We first build up a tree $R_s$ on $s$, starting from $s$ at level 0, the opposite endpoints of edges out of level $i$ forming level $i+1$. We will always choose ``cheap'' edges, but not always the cheapest ones, as explained later. From $s$ we will choose $k+\rr0$ cheap edges; from each of these $k+\rr0$ level-1 vertices we choose $\rr1$ cheap edges; from each of the $(k+\rr0)\rr1$ level-2 vertices we choose $\rr2$ cheap edges; and on each of the $(k+\rr0)\rr1 \rr2$ level-3 vertices we construct a shortest-path tree comprising $\dd$ vertices. We do a similar construction on $t$ to form $R_t$. Finally, we link $R_s$ and $R_t$ using cheap edges between their shortest-path trees. The values of the parameters $\rr0$, $\rr1$, $\rr2$ and $d$ are given in ~\cref{k0},~\cref{k1},~\cref{k2} and~\cref{ddef}, and it is confirmed in \cref{sec:Rsize} that the construction uses only a small fraction of $G$'s vertices, \begin{align}\label{n'} \card{V(R)} &= O((k+\rr0) \rr1 \rr2 d) = o(n) , \end{align} a fact we rely on in the construction. We will repeatedly use the following Chernoff bound, which in fact holds under more general conditions; see for example~\cite[Theorem 1, eq.~(4)]{Janson2001}. \begin{lemma}\label{lemma:BinDev} Let $X \sim \Bi(n,p)$ be a binomial random variable with mean $\lambda = np$. Then for any $\varepsilon>0$, $\Pr(X< (1-\varepsilon)\lambda) \leq \exp(-\varepsilon^2 \lambda/2)$. \end{lemma} \subsection{Cheap paths are short} We show that, w.h.p\xperiod} % {with high probability, every cheap path in $G$ is also short. The following lemma asserts the contrapositive. The result is used in~\cref{Bany} to restrict the number of edges the adversary can delete. \begin{lemma}\label{LLenBd} In both the uniform and exponential models, with probability $1-\OO{n^{-1.9}}$, simultaneously for all $l$ with $\ln n \leq l < n$, every $s$--$t$\xspace path of length $l$ has cost $\geq l/(19 n)$. \end{lemma} \begin{proof} We start with the uniform distribution. Here, with $X=\sum_{i=1}^{l} X_i$, $X_i \sim U(0,1)$ i.i.d\xperiod, $X$ has the Irwin-Hall distribution and it is a standard result that $\Pr(X \leq a) \leq a^l/l!$ (see for example \cite[eq.\ 8]{SorkinClique}). Recall that Stirling's approximation is also a lower bound. Thus, \begin{align*} \Prr{X \leq \frac{l}{19n}} &\leq \frac{(l/19n)^l}{l!} \leq \frac{(l/19n)^l}{\sqrt{2\pi l} \, (l/e)^l} < \parens{\frac{e}{19n}}^l. \end{align*} The cost of a fixed path of length $l$ has the same law as $X$. Over the $\leq n^l$ choices for such a path, the number $M_l$ of ``cheap paths'' (of cost $<l/(19n)$) satisfies (by Markov's inequality) \[ \Pr(M_l>0) \leq \E M_l \leq n^l \Prr{X \leq \frac{l}{19n}} \leq n^l \parens{\frac{e}{19n}}^l = \parens{\frac e{19}}^l . \] Summing over $l \geq \ln n$, the probability that there is a cheap path of any such length is $\OO{{(e/19)}^{\ln n}} = \OO{n^{-1.9}}$. Since an $\Exp(1)$ random weight $X'$ can be obtained from a $U(0,1)$ r.v.\ $X$ by setting $X' = -\ln(1-X)>X$, the exponential weight stochastically dominates the uniform, so the result for uniform immediately implies that for exponential. \end{proof} \subsection{Adversarial edge deletions}\label{adversary} As noted in the introduction, we introduce an edge-deleting adversary whose powers allow it to delete the paths $P_1,\ldots,P_k$, but which is more easily characterised than those paths are. We now specify what the adversary is permitted to do. Let \begin{align}\label{sdef} s = s(k) & \coloneqq 2k+\ln n . \end{align} (From context it should be easy to distinguish this use of $s$ from that as the source of an $s$--$t$\xspace path.) Let $w_0$ be the ``target cost'' of a path, namely \begin{align}\label{w0def} w_0 = w_0(k) & \coloneqq \frac s n = \frac{2k}n+\frac{\ln n}n . \end{align} Define a ``heavy'' edge to be one of cost \begin{align}\label{heavydef} & \geq \tfrac1{11} \varepsilon w_0. \end{align} Assuming that each of $P_1,\ldots,P_k$ has weight $\leq (1+\eps) w_0$, the \emph{number of heavy edges} in $P_1 \cup \cdots \cup P_k$ is at most \begin{align} \frac{k (1+\eps) w_0}{\tfrac1{11} \varepsilon w_0} < \frac{12k}{\varepsilon} < \frac{12 s}{\varepsilon} . \label{Bheavy} \end{align} Also, modulo the one-time failure probability $\OO{n^{-1.9}}$ from \cref{LLenBd}, by that lemma each path has length at most \begin{align} (1+\eps) w_0 \cdot 19n < 20 s . \label{pathlength} \end{align} Thus, the length of all $k$ paths taken together (i.e., the number of edges in $P_1 \cup \dots \cup P_k$) is at most \begin{align} 20 k s < 10 s^2 . \label{Bany} \end{align} And of course the $k$ paths include \begin{align} \label{Bincident} &\text{exactly $k$ edges incident on each of $s$ and $t$.} \end{align} Subject to these assumptions --- that each of $P_1,\ldots,P_k$ has weight $\leq (1+\eps) w_0$ and that the high-probability conclusion of \cref{LLenBd} holds --- $P_1 \cup \cdots \cup P_k$ satisfies all three of the constraints \cref{Bheavy}, \cref{Bany}, and \cref{Bincident} on heavy edges, all edges, and ``incident'' edges. An adversary who can delete any edge set subject to these constraints is able to delete $P_1 \cup \cdots \cup P_k$, which is all we require. However, to simplify analysis we will give the adversary even more power. At the root of $R$ we will allow the adversary to delete edges subject only to \cref{Bincident}; at level 1, additional edges subject only to the ``heavy-edge budget'' \cref{Bheavy}; and at levels 2 and 3 and for ``middle'' edges, additional edges subject only to the ``edge-count budget'' \cref{Bany}. \medskip We will show how to choose the parameters of $R$ so that every $s$--$t$\xspace path in $R$ has cost $\leq (1+\eps) w_0$, and so that $R$ is ``robust'': after the adversarial deletions, at least one path remains. Specifically, we will arrange that there remains a path in which the ``root'' edge incident to $s$ costs $\leq \tfrac k n + \frac19 \varepsilon w_0$, the edge out of level 1 is heavy but has cost $\leq \frac19 \varepsilon w_0$, the edge out of level 2 may be light or heavy and also has cost $\leq \frac19 \varepsilon w_0$, the path through the SP tree has total cost $\leq \frac12 \tfrac {\ln n} n + \frac19 \varepsilon w_0$, the central edge joining this to the opposite SP tree adds cost $\leq \frac19 \varepsilon w_0$, and the continuation of this path to $t$ has the symmetrical properties. It is immediate that such a path has total cost $\leq (2k+\ln n)/n + 9 \cdot \tfrac19 \varepsilon w_0 = (1+\eps) w_0$. (But see \cref{Rpathcost} for confirmation, after the construction is detailed.) \medskip \subsection{Level 0, cheapest edges}\label{level0} On $s$, add to $R$ the $k+\rr0$ edges of lowest cost, excluding $\set{s,t}$ from consideration, with \begin{align}\label{k0} \rr0 = \ceil{\tfrac1{10} \varepsilon s} = \Theta(s) . \end{align} Consider this step a \emph{failure} if any selected edge has cost greater than $\tfrac k n+\frac19 \varepsilon w_0$. There are $n'=n-2=(1-o(1))n$ edges under consideration, with weights i.i.d\xperiod $U(0,1)$, and failure occurs iff the number $X$ of edges with weights in the interval $[0, \tfrac k n+\frac19 \varepsilon w_0]$ is smaller than $k+\rr0$. Note that $X \sim \Bi(n', \tfrac k n+\frac19 \varepsilon w_0)$, thus $\E X = (1-o(1)) \, (k+\frac19 \varepsilon s)$, and failure means that $X <k+\rr0$, i.e., that \begin{align*} \frac{X}{\E X} &< (1+o(1)) \, \frac{k+\rr0}{k+\frac19 \varepsilon s} = (1+o(1)) \, \frac{k+\frac1{10} \varepsilon s}{k+\frac19 \varepsilon s} , \intertext{which by $s>2k$ is} &< (1+o(1)) \, \frac{k+\frac1{10} \varepsilon \cdot 2k}{k+\frac19 \varepsilon \cdot 2k} = (1+o(1)) \, \frac{(1+\frac2{10} \varepsilon)k}{(1+\frac29 \varepsilon)k} < 1-\tfrac1{50}\varepsilon = 1-\Omega(\varepsilon) . \end{align*} By \cref{lemma:BinDev}, then, the probability of failure is \begin{align} \Pr(X < (1-\Omega(\varepsilon)) \E X) &\leq \exp(-\Omega(\varepsilon^2) \E X/2) \leq \exp(-\Omega(\varepsilon^2 \cdot \varepsilon s)) \leq \exp(-\Theta(s)) , \label{level0fail} \end{align} the final expression using that $\varepsilon$ is constant (see \cref{epsconst}). So, modulo the given failure probability, every selected edge incident on $s$ has cost $\leq \tfrac k n+\frac19 \varepsilon w_0$, and after the adversarial deletion of $k$ of these edges, $\rr0$ remain. The selection of edges conditions the costs of the other edges incident on $s$, but none will play any role in the analysis. The purpose of the next two levels is to expand the number of edges to the point where the adversary cannot delete all of them, because of the heavy-edge budget \cref{Bheavy} for edges out of level 1, and the edge-count budget \cref{Bany} for edges out of level 2 and beyond. At the same time, we try to minimise the number of vertices introduced into the construction so that it will remain $o(n)$ for as large a $k$ as possible. \subsection{Level 1, cheapest heavy edges}\label{level1} From each neighbour $v$ of $s$ along the edges just added, add to $R$ the \begin{align}\label{k1} \rr1 \coloneqq \ceil{10,000/\varepsilon^2}= \Theta(1) \end{align} cheapest \emph{heavy} edges from $v$ to any of the $n'=n(1-o(1))$ vertices not yet added (see \cref{n'}), as before also excluding vertex $t$. Consider this step a \emph{failure} if any added edge has cost greater than $\frac19 \varepsilon w_0$. For each neighbour $v$ there are $n'$ edges under consideration, with weights i.i.d\xperiod $U(0,1)$, and failure occurs iff the number $X$ of edges with weights in the interval $[\frac1{11} \varepsilon w_0, \frac19 \varepsilon w_0]$ is smaller than $\rr1$. Note that $X \sim \Bi(n', (\frac19-\frac1{11}) \varepsilon w_0)$, thus $\E X = (1-o(1)) \, (\frac19-\frac1{11}) \varepsilon s = \Theta(\varepsilon s)$. Failure means that $X < \rr1 < \E X/2$, so by \cref{lemma:BinDev} the probability of failure for a given $v$ is $\leq \exp(-\Theta(\varepsilon s))$. The number of level-1 vertices $v$ is $k+\rr0 = O(s)$, so by the union bound the probability of any failure is \begin{align}\label{level1fail} \leq O(s) \exp(-\Theta(\varepsilon s)) & \leq \exp(-\Theta(s)), \end{align} by suitable adjustment of the constants implicit in $\Theta$. This edge selection conditions the costs of the other edges incident on each $v$, but none will play any role in the analysis. The adversary must leave $\rr0$ edges out of the root, expanding to \[ \rr0 \rr1 \geq \tfrac1{10} \varepsilon s \cdot 10,000/\varepsilon^2 = 1,000 s/\varepsilon \] (heavy) edges out of level 1, of which (by~\cref{Bheavy}) he can delete at most $12 s/\varepsilon$, leaving (very generously calculated) at least \begin{align}\label{K1} \Rsub1 & \coloneqq 120 s/\varepsilon = \Theta(\rr0 \rr1) \end{align} edges out of level 1. The vertices at the opposite endpoints of these edges constitute level 2. \subsection{Level 2, cheapest edges}\label{level2} From each level 2 vertex $v$ in turn, add to $R$ the cheapest $\rr2$ edges to any of the $n'=n(1-o(1))$ vertices not yet added, again also excluding vertex $t$ from consideration. Here choose $\rr2$ so as to make \begin{align}\label{K2} \Rsub2 \coloneqq \Rsub1 \rr2 = 12 s^2 , \end{align} namely taking \begin{align} \rr2 = \frac{12 s^2}{\Rsub1} = \frac{12 s^2}{120 s/\varepsilon} = \tfrac1{10} \varepsilon s = \Theta(\varepsilon s) . \label{k2} \end{align} Consider this step a \emph{failure} if any added edge has cost greater than $\frac19 \varepsilon w_0$. For each neighbour $v$ there are $n'=(1-o(1))n$ edges under consideration, with weights i.i.d\xperiod $U(0,1)$, and failure occurs iff the number $X$ of edges with weights in the interval $[0, \frac19 \varepsilon w_0]$ is smaller than $\rr2$. Note that $X \sim \Bi(n', \frac19 \varepsilon w_0)$, thus $\E X = (1-o(1)) \, \frac19 \varepsilon s = \Theta(\varepsilon s)$. Failure means that $X < \rr2 < 0.99 \E X$, so by \cref{lemma:BinDev} the probability of failure for a given $v$ is $\leq \exp(-\Theta(\varepsilon s))$. The number of level-2 vertices $v$ is $(k+\rr0) \rr1 = \OO{s}$ so by the union bound the probability of any failure is \begin{align}\label{level2fail} & \leq \exp(-\Theta(\varepsilon s)) \end{align} This edge selection conditions the costs of the other edges incident on each $v$, but none will play any role in the analysis. The adversary had to leave at least $\Rsub1$ edges out of level 1, expanding to $\Rsub1 \rr2 = \Rsub2 = 12 s^2$ edges out of level 2, of which by~\cref{Bany} he can delete at most $10 s^2$, leaving at least $2 s^2$ edges out of level 2. The vertices at the opposite endpoints of these edges constitute level 3. \subsection{Level 3, shortest-path trees}\label{level3} We now grow each level-3 vertex $v$ to a tree $T_v$ with $d$ vertices, including $v$, choosing \begin{align} d \coloneqq \ceil{ \sqrt{\frac{n \ln n}{2 s^3}} \; } < \sqrt n. \label{ddef} \end{align} We grow these trees one after another, always working within the $n'=n(1-o(1))$ vertices not yet added, and again always excluding vertex t from consideration. Controlling the lengths of the paths in $T_v$ would allow a choice of $d$ as large as $\sqrt n$, but we make it smaller to keep the number of vertices in $R$ as small as possible (and thus keep it to $o(n)$ for $s$ as large as possible). Here it will be convenient to work with exponentially rather than uniformly distributed edge weights. There are various easy ways to arrange this. We do so by temporarily replacing each uniform weight $w$ with a weight $w' = -\ln (1-w)$; it is standard that these transformed weights are exponentially distributed, and that $w' \geq w$. We construct a shortest-path tree (SPT) of order $d$ using the transformed weights; it will not be an SPT for the original weights, but its paths will be short under the original weights, which is all that we care about. Define the distance $\operatorname{dist}(u,v)$ between two vertices to be the cost of a minimum-weight path between them, and define the radius $\operatorname{rad}(T_v)$ of an SPT $T_v$ to be the maximum distance from $v$ to any vertex in $T_v$. The radius is described by the following claim, which we phrase in a generic setting with $n$ vertices and a root vertex $s$. \begin{claim} \label{treeradius} In a complete graph $K_n$ with i.i.d\xperiod exponential edge weights with mean 1, the radius $X=\operatorname{rad}(T_s)$ of a shortest-path tree $T_s$ of order $d$ is \begin{align}\label{treeX} X = \sum_{i=1}^{d-1} X_i , \end{align} where the $X_i$ are independent random variables with $X_i \sim \Exp(i(n-i))$. \end{claim} \begin{proof} Following~\cite{Janson123}, think of the process of finding shortest paths from $s$ to other vertices as first-passage percolation or ``infection spreading'' starting from $s$. Let $L\coloneqq L(r)$ be the set of vertices within radius (distance) $r$ of $s$; we think of gradually increasing $r$, starting with $r=0$ where $L=\set s$. It is well known that each edge $(v,u) \in L(r) \times (V\setminus L(r))$ has exponentially distributed weight $W$ conditioned by $W+\operatorname{dist}(s,v) \geq r$, and that these random weights are independent. This can be seen by imagining that infection has spread to radius $r$ from $s$, including to the vertex $v$ and additionally a length $r-\operatorname{dist}(s,v)$ further along the edge $(v,u)$, and appealing to the memoryless property of the exponential distribution; it can also be verified by analysing Dijkstra's algorithm in this randomised setting. It follows that the distance $X_1$ to the vertex nearest $s$ is distributed as $X_1 \sim \Exp(n-1)$; the additional distance to the next vertex is $X_2$ with $X_2 \sim \Exp(2(n-2))$ and independent of $X_1$ (for total distance $X_1+X_2$); and when there are $i$ vertices in the tree, the additional distance to the next is $X_i \sim \Exp(i(n-i))$, with all the $X_i$ independent, for total distance as claimed. \end{proof} \emph{We will only use trees} $T_v$ whose radius is $X \leq (1+\tfrac29 \varepsilon) \, \tfrac12 \ln n/n$ $< \tfrac12 \ln n/n + \tfrac19 \varepsilon w_0$. Call a tree a failure (and do not include it in the structure $R$) if $X > (1+\tfrac29 \varepsilon) \tfrac12 \ln n/n$. Declare the construction of level 3 a \emph{failure} if more than $0.01 s^2$ trees fail. Since~\cref{treeX} is monotone increasing in $d$, the larger the $d$, the greater the probability of failure, so in the next paragraphs we will pessimistically take $d$ to be $\sqrt n$ (ignoring integrality since $\sqrt n$ is large). In this case, applying \cref{treeradius} to $T_v$, constructed in a complete graph of order $n'=n(1-o(1))$, the expectation of $X$ is \begin{align} \mu \coloneqq \E X &= \sum_{i=1}^{d-1} \frac1{i(n'-i)} = \frac{1+o(1)}{n} \sum_{i=1}^{d-1} \frac1{i} = (1+o(1)) \frac{\ln d}n = (1+o(1)) \frac12 \frac{\ln n}n \label{treeDia} . \end{align} Thus, failure of $T_v$ implies that \begin{align}\label{treefail1} \frac X \mu &> 1+\frac15 \varepsilon . \end{align} To bound the probability of this event we require one more lemma (also used later in proving \cref{lemma:edge-orderstat}). \begin{lemma}[{\cite[Theorem 5.1]{JansonExpTail}}]\label{exptail} Let $X=\sum_{i=1}^{n} X_i$ with $X_i \sim \Exp(a_i)$ independent rate-$a_i$ random variables, where $a_i \geq 0$. Write $a_{*} \coloneqq \min_{i} a_i$ and $\mu \coloneqq \E X = \sum_{i=1}^n \frac 1{a_i}$. Then: \newline for any $\lambda = 1+\varepsilon > 1$, \begin{align}\label{exp.1} \Prob(X \geq \lambda \mu) & \leq \lambda^{-1} e^{-a_{*} \mu (\lambda - 1 - \ln \lambda)} \leq \exp(-\Omega(\a_* \mu)) \end{align} for any $\lambda = 1- \varepsilon < 1$, \begin{align}\label{exp.2} \Prob(X \leq \lambda \mu) & \leq e^{-a_{*} \mu (\lambda - 1 - \ln \lambda)} \leq \exp(-\Omega(\a_* \mu)), \end{align} and for any $\varepsilon>0$, \begin{align}\label{exp.3} \Prob( \abs{X-\mu} \geq \varepsilon \mu) & \leq 2 \exp(-\Omega(\a_* \mu )). \end{align} The constants implicit in the $\Omega(\cdot)$ expressions are positive and only depend on $\varepsilon$. \end{lemma} \begin{proof} The inequalities in \cref{exp.1} and \cref{exp.2} in terms of $\lambda$ are directly from~\cite[Theorem 5.1]{JansonExpTail}. The remaining expressions, including \cref{exp.3}, follow immediately. \end{proof} From \cref{exp.1} of \cref{exptail}, the probability of the event in \cref{treefail1} (and thus that of $T_v$ failing) is at most \begin{align} \Pr\big( {X-\mu} > \frac15 \varepsilon \mu \big) & \leq \exp(-\Omega(\a_* \mu) ) = \exp( -\Omega(\ln n)), \label{treeFail} \end{align} using that $a_* = n' = (1-o(1))n$, $\mu$ is given by \cref{treeDia}, and $\varepsilon=\Theta(1)$. The total number of trees built is $N=(k+\rr0) \rr1 \rr2$, which, with reference to \cref{sdef}, \cref{k0}, \cref{k1}, and \cref{k2}, is $\Theta(s^2)$. By \cref{treeFail}, each tree independently fails with at most some probability $p=o(1)$. Thus, the number of trees surviving dominates $\Bi(N,1-p)$, with expectation $\lambda=N(1-p)=N(1-o(1))$. Failure at level 3 means that at least $0.01 s^2=\Theta(N)$ trees fail, equivalently the number surviving is at most some $\lambda(1-\Theta(1))$, which by \cref{lemma:BinDev} has probability \begin{align}\label{level3fail} & \exp(-\Omega(s^2)) . \end{align} \begin{remark}\label{rem1} When construction of a tree $T_v$ rooted at a level-3 vertex $v$ is finished, the edge between any vertex $a$ of $T_v$ and any vertex $b$ in $V' \setminus V(T_v)$ has weight $w(a,b)$ that --- still in the uniform model with edge weights temporarily transformed to be exponentially distributed --- is exponentially distributed conditional upon being $\geq \operatorname{rad}(T_v)-d(v,a)$. Equivalently, the edge $\set{a,b}$ gives a $v$-to-$b$ path (through $a$) with cost $\operatorname{rad}(T_v)+X_{a,b}$, where the ``excess'' $X_{a,b}$ has simple exponential distribution $X_{a,b} \sim \Exp(1)$ (with no conditioning). Furthermore, the $X_{a,b}$ are independent, over all choices of $a$ and $b$. \end{remark} Call $R_s$ the now-complete construction on $s$. Note that there is no conditioning on edges between the remaining vertices; in particular, the SPT infection process (or equivalently Dijkstra's algorithm) as described in \cref{treeradius} never looked at edges between uninfected vertices. \subsection{Symmetric construction on vertex \tp{$t$}{t}} Just as we have constructed $R_s$, we now make a similar construction $R_t$ for vertex $t$, with the same branching factors out of levels 0, 1, and 2 and similar SPTs on level-3 vertices. Since the number $n'$ of vertices available after constructing $R_s$ still satisfies $n' = (1-o(1))n$, and because the construction on $s$ did not look at nor condition any edge between these vertices, the construction on $t$ enjoys the same properties as that on $s$. \subsection{Edges between the trees on \tp{$s$}{s} and \tp{$t$}{t}} It remains only to complete paths between $s$ and $t$, which we do by adding cheap edges (where present) between the SPTs in $R_s$ and those in $R_t$. Let $T_u$ be an SPT rooted at a level-3 vertex $u$ of $R_s$, and $T_v$ one rooted at a level-3 vertex $v$ of $R_t$. Let $a$ and $b$ be any vertices in $T_u$ and $T_v$ respectively. By \cref{rem1}, edge $\set{a,b}$ gives a $u$-to-$b$ path with cost $\operatorname{rad}(T_u)+X_{a,b}$, the collection of all the excesses $X_{a,b}$ being i.i.d\xperiod each with distribution $X_{a,b} \sim \Exp(1)$. Thus, $\set{a,b}$ gives a $u$-to-$v$ path with cost $\leq \operatorname{rad}(T_u) + X_{a,b} + \operatorname{rad}(T_v)$. Select, and add to the full construction $R$, any such ``middle edge'' $\set{a,b}$ having $X_{a,b} \leq \frac19 \varepsilon w_0$. This completes the construction of $R$. \subsection{Order of \tp{$R$}{R}, failure probability, and path costs}\label{sec:Rsize} It is worth first confirming that the construction uses, as claimed, $o(n)$ vertices. The number of vertices used is of order $(k+\rr0)\rr1\rr2 d$, which by \cref{k0}, \cref{k1}, \cref{k2}, and \cref{ddef} is $O(s^2 d)$. Recalling from~\cref{ddef} that $d = \ceil{ \sqrt{\lfrac{n \ln n}{2 s^3}} \;}$, as long as the ceiling function does not affect the order of $d$, the total number of vertices is $O(s^2 d) = O(\sqrt{n s \ln n})$, which is $o(n)$ for $s=o(n/\ln n)$. However, the ceiling function does affect the order of $d$ when $n \ln n/2s^3 < 1$, i.e., when $s> {(\frac12 n \ln n)}^{1/3}$; in this case, $d=1$, the total number of vertices used is $O(s^2)$, and this is still $o(n)$ if $s=o(\sqrt n)$. Taking the two cases together, the construction is valid up to any $s=o(\sqrt n)$, or equivalently for any $k=o(\sqrt n)$. Failures at levels 0, 1, and 2 each occur w.p.\ $\leq \exp(-\Theta(s))$ (by \cref{level0fail}, \cref{level1fail}, and \cref{level2fail}), and at level 3 w.p.\ $\leq \exp(-\Omega(s^2))$ (by \cref{level3fail}), so by the union bound the probability of any failure is $\leq \exp(-\Theta(s))$. We now confirm that, assuming that the construction was successful, any $s$--$t$\xspace path in $R$ \emph{through successful SPTs} has cost $\leq (1+\varepsilon) w_0$. (Remember that there may be some unsuccessful SPTs.) By assumption of success, any level-0 edge on $s$ or $t$ has cost $\leq \tfrac k n+\frac19 \varepsilon w_0$, any level-1 edge has cost $\leq \frac19 \varepsilon w_0$, and any level-2 edge also has cost $\leq \frac19 \varepsilon w_0$. Each successful level-3 tree $T$ in $R_s$ or $R_t$ has radius $\operatorname{rad}(T) \leq (1+\frac29 \varepsilon) \frac12 \ln n/n \leq \frac12 \ln n/n + \frac19 \varepsilon w_0$, and each selected middle edge $\set{a,b}$ connects the roots of two trees at an excess cost (above the sum of the two radii) of $X_{a,b} \leq \frac19 \varepsilon w_0$. The total of the 9 upper bounds in question is \begin{align}\label{Rpathcost} 2 \cdot \frac k n + 2 \cdot \frac12 \ln n/n + 9 \cdot \frac19 \varepsilon w_0 &= (1+\varepsilon) w_0 . \end{align} \subsection{Robustness of \tp{$R$}{R}} We now show that, after the deletion of the $k$ cheapest paths in $G$, there remains at least one path in $R$ (that uses successful SPTs). Recall from \cref{adversary} that deletion of the $k$ cheapest paths in $G$ is conservatively modeled as an adversarial deletion subject to: \cref{Bincident}, the deletion of exactly $k$ edges incident on each of $s$ and $t$; \cref{Bheavy}, the number of heavy edges deleted at level 1; and \cref{Bany}, the total number of edges deleted elsewhere in $R$ (at levels 2 and 3, and joining $R_s$ and $R_t$). Without loss of generality we may assume that the adversary does not delete an edge within an SPT $T$, nor a middle edge from such a tree to a facing one, since deleting the level-2 edge into the level-3 root of $T$ destroys more paths in $R$ at the same budgetary cost. By the assumption of success, there are at most $0.01 s^2$ failed SPTs on each of $s$ and $t$, and for simplicity we will deal with them by imagining all trees to be successful but allowing the adversary his choice of this many SPTs to delete; by the argument above we can model this as deletion of edges into the roots of these trees, and simply add $0.02 s^2$ to this budget. Let us now allow the adversary to delete $k$ edges from each of $s$ and $t$, $12s/\varepsilon$ edges out of level 1 for each (double-counting the heavy-edge budget), and $10.02 s^2$ edges out of level 2 for each (again double-counting). Can he destroy all $s$--$t$\xspace paths? We have not yet made any high-probability structural assertion about the middle edges, so this is a probabilistic question: what is the probability, over the randomness still present in the middle edges, that there is an adversarial deletion destroying all paths? Of the $k+\rr0=O(s)$ edges on $s$, the adversary chooses $k$ to delete; there are at most $2^{O(s)}$ ways to do so. Any choice leaves $\Theta(\rr0 \rr1) = \Theta(s)$ edges out of level 1, of which the adversary is able to delete a positive fraction, again in at most $2^{O(s)}$ ways. Any choice leaves $\Theta(s^2)$ edges out of level 2, of which the adversary is able to delete a positive fraction, in at most $2^{O(s^2)}$ ways. The adversary makes a similar set of choices on $t$, but still this comes to just $2^{O(s^2)}$ possible outcomes in all. A given deletion choice destroys all paths precisely if it leaves no middle edge of excess $\leq \frac19 \varepsilon w_0$. (Remember that, w.l.o.g., we have excluded deletions in and between the SPTs at level~3.) By construction, any deletion choice leaves $\Theta(s^2)$ edges out of level 2 and thus, by \cref{ddef}, $\Theta(s^2 d) = \Omega(\sqrt{n s \ln n})$ vertices in SPTs in each of $R_s$ and $R_t$, for $\Omega(n s \ln n)$ potential middle edges. A middle edge is \emph{selected} if its excess cost (in the exponential model) is $w'=-\ln(1-w) \leq \frac19 \varepsilon w_0$, i.e., if $1-w \geq \exp(-\frac19 \varepsilon w_0)$, thus is \emph{rejected} with probability $\exp(-\frac19 \varepsilon w_0)$. There is no path only if every potential edge is rejected, which happens w.p.\ $\leq \exp(-\frac19 \varepsilon w_0 \cdot n s \ln n) = \exp(-\Omega(s^2 \ln n))$. Taking the union bound over all adversarial choices, the probability than any choice leaves no paths is \begin{align} \label{smallkAdversaryFailure} 2^{O(s^2)} \exp(-\Omega(s^2 \ln n)) &= \exp(-\Omega(s^2 \ln n)) . \end{align} This is dominated by the failure probabilities $\exp(-\Theta(s))$ for other steps. \subsection{Success for each \tp{$k$}{k}, and for all \tp{$k$}{k}}\label{sec:small-success} We have shown that, for any $k=o(\sqrt n)$, subject to an absence of failures, we can generate a robust structure $R^{(k)}$ in which, after adversarial deletions, there remains an $s$--$t$\xspace path of cost $\leq (1+\varepsilon) w_0(k)$. (Remember that $w_0$ and $s$ are simple functions of $k$, per~\cref{sdef} and~\cref{w0def}. Here we retain the argument $k$ we usually suppress.) There are two types of failures possible. The first is that the graph fails \cref{LLenBd}'s conclusion that ``cheap paths are short''; this occurs w.p.\ $\OO{n^{-1.9}}$. The second is that $R^{(k)}$ is not robust; this occurs w.p.\ $\OO{\exp(-\Omega(s(k)))}$. Assume success in generating $R^{(k)}$. We claim that $P_1,\ldots,P_{k+1}$ all have cost $\leq (1+\varepsilon) w_0(k)$ (call this ``cheap''). Suppose not. Then there is some $i\leq k$ for which $P_1,\ldots,P_i$ are cheap but $P_{i+1}$ is not. Our adversary's budget allows it to delete $P_1,\ldots,P_i$, and by assumption of success this leaves a cheap path $P$ in $R^{(k)}$. Thus there is a cheap $i+1$st path in $G$, a contradiction. It follows that \emph{for each} $k$, $X_{k+1} \leq (1+\varepsilon) w_0(k)$ with probability \begin{align} 1-O(n^{-1.9}) -O(\exp(-\Omega(s(k)))) . \label{ksucceeds} \end{align} \medskip A simple calculation shows that w.h.p\xperiod} % {with high probability $X_{k+1} \leq (1+\varepsilon) w_0(k)$ \emph{simultaneously for all $k$} in this range, proving the upper bound in \cref{Xkbounds}. By the union bound, the probability of failure to build a robust structure for \emph{any} $k$ is at most \begin{align} \sum_{k=0}^{\infty} \exp(-\Omega(s(k))) &\leq \ln n \exp(-\Omega(\ln n)) + \sum_{k=\ln n}^{\infty} \exp(-\Omega(k)) \notag \\ &= \exp(-\Omega(\ln n)) = n^{-\Omega(1)}. \label{eq:all-k-small} \end{align} Including the probability of failure in applying \cref{LLenBd}, the total failure probability is $O(n^{-1.9} + n^{-\Omega(1)}) = o(1)$. \subsection{Limitation to small \tp{$k$}{k}} \label{small-k-limitation} We have established \cref{Tmain} up to any $k=o(\sqrt n)$, and the construction of $R^{(k)}$ was tailored to such values. For levels using heavy edges, fanout is limited to $O(s)$. On the other hand, the meet-in-the-middle argument requires that each side grow large, to $\Omega(\sqrt{n/s})$. Thus, for small $k$, a more-than-constant number of levels is needed. Summing heavy edges over this many levels would exceed the target weight $(1+\varepsilon) w_0$, so light edges are needed. The adversary may delete $\Theta(s^2)$ light edges, so the construction must contain at least this many. The construction explicitly required each light edge to lead to a new vertex, and we do not readily see how to do otherwise as long as we are using shortest-path trees, thus intrinsically limiting $s$ (thus $k$) to $O(\sqrt n)$. For larger $k$, however, we can obtain sufficient heavy-edge fanout in constant depth, permitting a simpler construction described in \cref{largekUB}. \section{Edge order statistics}\label{sec:order-stat} In this section we establish results on order statistics needed in later sections. Let $\set{W\os k}_{k=1}^{n-1}$ be the order statistics of $n-1$ i.i.d\xperiod random variables, variously uniform $U(0,1)$ or exponential $\Exp(1)$. We choose $n-1$ rather than $n$ as the parameter both because many expressions are more natural in this parametrisation, and because this way $W\os k$ is the cost of the $k$th cheapest edge incident to a fixed vertex $v \in K_n$. The following lemma is used in \cref{pathweightsdetail}. \begin{lemma}\label{intervals} Let $l=n^{-0.99}$. Consider the unit interval $[0,1]$ with $n$ points placed uniformly and independently at random. Then w.h.p\xperiod} % {with high probability every interval of length at least $l'\geq l$ contains at least $0.99 l' n$ points. \end{lemma} \begin{proof} Partition the unit interval into contiguous intervals $I_i$ each of length $L\coloneqq l/1000$, using $\floor{1/L}$ such intervals (possibly leaving a small interval near 1 not covered). Any interval $I$ of length $l' \geq l$ has at least a $998/1000$ fraction of its length covered by intervals $I_i \subset I$, and we will show that w.h.p\xperiod} % {with high probability every interval $I_i$ contains at least $0.999 L n$ points (that is, at least a $0.999$ fraction of the expectation). If so, it follows that $I$ has at least $0.999 \cdot 0.998 l' n \geq 0.99 l' n$ points. The distribution of the number of points in each interval $I_i$ of length $L$ follows the binomial distribution $\Bi(n, L)$. By \cref{lemma:BinDev}, \begin{align*} \Prob \parens{\Bi(n, L) \leq 0.999 L n} \leq \exp(-\Omega(L n)) , \end{align*} where the sign in the $\Omega$ is taken as positive. The probability that any interval $I_i$ contains less than $0.999$ points is, by the union bound, at most, \begin{align}\label{eq:intervals} \floor{1/L} \cdot \exp(-{\Omega(L n)}) = \exp(-\Omega(n^{0.01})) = o(1) \end{align} as desired. \end{proof} The following lemma is used in \cref{eq:implies-main} and \cref{eq:edge-orderstat}. \begin{lemma}\label{lemma:edge-orderstat} Let $\set{W\os k}_{k=1}^{n-1}$ be the order statistics of $n-1$ i.i.d\xperiod random variables, either all uniform $U(0,1)$ or all exponential $\Exp(1)$. For any $\varepsilon > 0$ and $a= a(n) = \omega(1)$, w.h.p\xperiod} % {with high probability \[ 1-\varepsilon \leq \frac{W\os k}{\E W\os k} \leq 1+\varepsilon \] simultaneously for all $k$ in the range $a \leq k \leq n-1$. \end{lemma} \begin{proof} Without loss of generality, we may assume that $a \leq n/10$. \textbf{Exponential case}. It is standard that, where $Z_i \sim \Exp(i)$ are independent exponential r.v.s, we may generate the $W\os k$ as \begin{align} W\os k &= \sum_{i=1}^{k} Z_{n-i} . \label{ordersumexp} \end{align} Using a superscripted $E$ to highlight the exponential model, $W\os k$ has mean \begin{align} \label{muX} \mu_k = \mu_k^{(E)} \coloneqq \E W\os k &= \sum_{i=1}^k \frac{1}{n-i} = H(n-1)-H(n-k-1) \asymp \ln(n)-\ln(n-k) ; \end{align} the change by 1 in the logarithms' arguments avoids $\ln(0)$ when $k=n-1$ and remains asymptotically correct. By~\cref{exp.3}, \begin{align} \Prob( \abs{ W\os k - \mu_k} \geq \varepsilon \mu) & \leq 2 \exp \parens{-\Omega((n-k) \mu_k)} .\label{eq:ExpSum} \end{align} By the union bound, it suffices to show that the sum over $k$ from $a$ to $n-1$ of the RHS of~\cref{eq:ExpSum} is $o(1)$. We treat the sum in two ranges. For $k \leq {\frac n2}$, $(n-k)\mu_k \geq \frac n 2 \cdot \frac k n = \frac k 2$. Thus, \begin{align} \sum_{k=a}^{{\lfrac n2}} \exp \parens{-\Omega((n-k) \mu_k)} \; \leq \; \sum_{k=a}^{{\lfrac n2}} \exp \parens{- \Omega(k)} \; \leq \; O(a e^{-\Omega(a)}) \to 0, \end{align} since $a = \omega(1)$. For $k > \frac n 2$, for brevity let $\bar{k}=n-k$. Then $\mu_k \asymp \ln n-\ln(\bar{k})$ by~\cref{muX} and \begin{align} \notag \sum_{\kbar=1}^{n/2} \exp\parens{ -\Omega((n-k) \mu_k) } &= \sum_{\kbar=1}^{n/2} \exp \parens{ - \bar{k} \, \Omega(\ln n-\ln(\bar{k}))} = \sum_{\kbar=1}^{n/2} \parens{\frac{\bar{k}}{n}}^{\Omega(\bar{k})} \\ & \leq \parens{\frac1n}^{\Omega(1)} \sum_{\kbar=1}^{n/2} \bar{k}^{\Omega(1)} \parens{\frac \bar{k} n}^{\Omega(\bar{k}-1)} = n^{-\Omega(1)} = o(1) , \end{align} where the explicit inequality factors out the $\bar{k}=1$ term, from which, since $\bar{k}/n \leq 1/2$, the later terms decrease geometrically. This concludes the exponential case. \medskip \textbf{Uniform case}: Let $U_i \sim U(0,1)$ be i.i.d\xperiod uniform random variables and $W_i \sim \Exp(1)$ i.i.d\xperiod exponential random variables. Because the exponential distribution has CDF $F(x) = 1-\exp(-x)$, we may couple the two sets of variables as $U_i = F(W_i)$ or equivalently $W_i = f(U_i)$ with $f(x) = F^{-1}(x) = -\ln(1-x)$. Because $f$ is increasing, $W\os k = f(U\os k)$. Now using superscript $U$ to distinguish the uniform model, the mean is well known to be \begin{align}\label{muU} \mu_k = \mu_k^{(U)} & \coloneqq \E U_{(k)} = \frac kn \end{align} We want to show that with high probability, for all $k$ in the range $a \leq k \leq n-1$, \[ (1-\varepsilon) \mu_k^{(U)} \leq U_{(k)} \leq (1+\varepsilon) \mu_k^{(U)} \] or equivalently, \[ f\parens{ (1-\varepsilon) \mu_k^{(U)}} \leq W\os k \leq f\parens{ (1+\varepsilon) \mu_k^{(U)}} . \] From the exponential case already proved, taking error bound $\varepsilon/2$, we know that w.h.p\xperiod} % {with high probability, for all $k$, \[ (1-\varepsilon/2) \mu_k^{(E)} \leq W\os k \leq (1+\varepsilon/2) \mu_k^{(E)} , \] so it suffices to show that, for all $k$ (deterministically), \[ f\parens{ (1-\varepsilon) \mu_k^{(U)}} \leq (1-\varepsilon/2) \mu_k^{(E)} \quad \text{ and } \quad f\parens{ (1+\varepsilon) \mu_k^{(U)}} \geq (1+\varepsilon/2) \mu_k^{(E)} . \] This is so. Using~\cref{muU},~\cref{muX}, and convexity of $f$, \[ f\left((1-\varepsilon) \mu_k^{(U)} \right) = f((1-\varepsilon)k/n) \leq (1-\varepsilon) f(k/n) = (1-\varepsilon) \ln \left( \frac{n}{n-k} \right) \leq (1-\varepsilon/2) \mu_k^{(E)} ; \] \[ f\left((1+\varepsilon) \mu_k^{(U)} \right) = f((1+\varepsilon)k/n) \geq (1+\varepsilon) f(k/n) = (1+\varepsilon) \ln \left( \frac{n}{n-k} \right) \geq (1+\varepsilon/2) \mu_k^{(E)} . \] \end{proof} \section{Upper bound for large \tp{$k$}{k}, sketch}\label{largekUB} \subsection{Introduction} \label{largekIntro} To address larger values of $k$ we use a different construction, generating $s$--$t$\xspace paths of length 4. A straightforward extension of the previous argument to this construction would let us get up to $k=n-f(n)$ for an arbitrarily slowly growing function $f$, but not to $k=n-1$ because it requires ${k+1}/\varepsilon^2$ edges incident on each of $s$ and $t$ (thus requires that ${k+1}/\varepsilon^2 \leq n-1$). Getting all the way to $k=n-1$ requires a couple of additional ideas. Again, we will introduce an adversary with a cost budget that with high probability exceeds the cost of the first $k$ cheapest paths. First, we observe that much of the adversary's cost budget must be spent on edges incident to $s$ and $t$, leaving less to delete other edges, thus allowing a smaller structure $R$ to be sufficiently robust. In particular, the $k$ cheapest paths from $s$ to $t$ must use edges incident on $s$ of total weight at least $\sum_{i=1}^k W\os i^s$ where \begin{align} \label{Wkv} W\os i^v \end{align} is the cost of the $i$th cheapest edge incident to $v$. (We may omit the superscript when it is either generic or clear from context.) One technical detail is that, where $R$ includes the $k+\rr0$ cheapest edges incident to $s$, we will control $W\os{\kk}-W\os k$ directly, using results on order statistics from \cref{sec:order-stat}, rather than through a high-probability upper bound on $W\os{\kk}$ and a high-probability lower bound on $W\os k$. Finally, it is no longer adequate to allow path costs to exceed their nominal values by an $\varepsilon=\Theta(1)$ factor, as such large excesses would swell the adversary's budget too quickly, so we more tightly control the excess cost of each path $P_k$ as a function of $k$ (and $n$, implicitly). The details later will be clearer if we sketch the argument now, with most details but without the calculations. We will argue for $k$ from $n^{4/10}$ to $n-1$. (We must start with some $k=o(n^{1/2})$ since that is as far as the ``small $k$'' argument extended, and we need $k=\omega(n^{1/3})$ since below this the new construction's path costs would exceed the $2k/n$ target.) \subsection{Structure \tp{$R$}{R}} \label{structR} \cref{figR2} illustrates the robust structure $R=R^{(k)}$ after adversarial deletion of root edges, as discussed in \cref{robustness} below. The construction is based on parameters $\rr0 = \rr0(k)$ and $\varepsilon_k$ to be defined later. Start with $R$ consisting of just the vertices $s$ and $t$. Add to $R$ the $k+\rr0$ edges incident on $s$ of lowest cost, and let $V_s$ be the set of opposite endpoints of these edges. Do the same for $t$, generating vertex set $V_t$. Take $M \coloneqq V(G)\setminus \set{s,t}$ as a collection of ``middle vertices''. Note that $V_s$, $V_t$, and $M$ may well have vertices in common, but our analysis will use a subgraph of $R$ where the relevant subsets of these three sets are disjoint, and it is easier to understand the construction imagining them to be disjoint. Add to $R$ each edge $e$ in $M \times V_s$ and $M \times V_t$ that is ``heavy but not too heavy'', with cost $W(e) \in (\eps_k, 2\eps_k)$. This concludes the construction of the structure $R$. \begin{figure} \centering \vspace*{2cm} \includegraphics[width=0.6\textwidth]{Rlarge} \caption{The robust structure $R=R^{(k)}$ after adversarial deletion of $k$ edges on $s$, leaving $\rr0$ edges to some vertices $V'_s \subseteq V_s$, and likewise for $t$ and $V'_t$. The middle vertices are pruned to $M'=M\setminus (V_s'\cup V_t')$, and edges from $M'$ to $V_s'$ and $V_t'$ are in $R$ if they have weight between $\varepsilon_k$ and $2\varepsilon_k$. Here, edges from just one representative vertex $v \in M'$ are illustrated. }\label{figR2} \end{figure} \subsection{Path weights} \label{pathweights} It is immediate that every $s$--$t$\xspace path in $R$ has cost at most \begin{align} W^s_{(k+\rr0)} +2\eps_k+2\eps_k+W^t_{(k+\rr0)} . \label{path1} \end{align} We will show (in \cref{eq:Wkk0} for uniform and \cref{eq:expWkk0} for exponential) that, subject to the non-occurrence of certain unlikely failure events, \cref{path1} is at most \begin{align} W\os{\kp}^s+W\os{\kp}^t+7\eps_k . \label{path+} \end{align} We will show in \cref{robustness} that, after deletion of the first $k$ paths, there remains an $s$--$t$\xspace path in $R$ (again subject to non-occurrence of unlikely failure events), whereupon it follows that \begin{align} X_{k+1} & \leq W\os{\kp}^s+W\os{\kp}^t+7\eps_k . \label{XkWok} \end{align} \subsection{Adversary} \label{subadv} We define an adversary who is ``sufficiently strong'' to delete the first $k$ paths. For $k \leq n^{4/10}$, taking $\varepsilon=0.1$, \cref{ksucceeds} implies that w.p.\ \begin{align} 1-O(n^{-1.9}) -O(\exp(-\Omega(n^{4/10}))) &= 1-O(n^{-1.9}) = 1-o(1), \label{Pn0.4} \end{align} we have that \begin{align} X_k &\leq X_{n^{4/10}} \leq 3 n^{4/10} / n . \label{Xn0.4} \end{align} For $k>n^{4/10}$, further assume the absence of the failure events alluded to just above, so that \cref{XkWok} holds. Then, hiding a sum of the $\ln n/n$ terms of~\cref{Xkbounds} in the $o(\,)$ term below, \begin{align} \sum_{i=1}^{k} X_i &= \sum_{i=1}^{n^{4/10}} X_i + \sum_{i=n^{4/10}+1}^k X_i \notag \\ &\leq \frac{3{(n^{4/10})}^2}{n} + \sum_{i=n^{4/10}+1}^{k} X_i \notag \\ &\leq 3n^{-2/10} + \sum_{i=n^{4/10}+1}^{k} \parens{W\os i^s+W\os i^t+7\varepsilon_{i-1}} \notag \\ & =: U_k . \label{Ukdef} \end{align} Thus, the first $k$ paths' edges have total weight at most $U_k$. Furthermore, the first $k$ paths' edges incident on $s$ and $t$ are all distinct except, possibly, for the edge \ensuremath{\set{s,t}}\xspace. Therefore, not counting edge $s$--$t$\xspace at all, the cost of these ``incident'' edges is at least \begin{align} \label{Ik} I_k & \coloneqq \sum_{i=1}^{k-1} \parens{ W\os i^s+W\os i^t } . \end{align} (In proving \cref{expkbig} we will use a slightly different lower bound $I_k$ on the weight of the incident edges.) It follows that the first $k$ paths' ``middle edges'' (edges other than the incident edges) cost at most $U_k-I_k$. We will explicitly define a budget $B_k$ satisfying \begin{align}\label{budget} B_k & \geq U_k-I_k . \end{align} We will allow the adversary to delete any $k$ edges in $G$ incident on each of $s$ and $t$, possibly including the edge $s$--$t$\xspace (enough to let it delete the incident edges of the first $k$ paths), and to delete any other edges in $G$ of total cost at most $B_k$ (enough to let it delete the middle edges of the first $k$ paths). Thus, the adversary is sufficiently strong to delete the first $k$ paths. The adversary's allowable deletions in $G$ mean that also in $R$ it deletes at most $k$ edges incident on each of $s$ and $t$, and middle edges of total cost at most $B_k$. \subsection{Budgets \tp{$B_k$}{B\_k}} \label{budgets} The budgets $B_k$ will be defined explicitly in the details. For the model with uniformly distributed edge weights we will do so in two ranges of $k$, corresponding to \cref{kmedium,kbig}, and likewise in the model with exponentially distributed edge weights, corresponding to \cref{expkmedium,expkbig}. For \cref{kbig,expkbig} we will establish \cref{budget} directly. For \cref{kmedium,expkmedium} we will establish \cref{budget} by the following reasoning; we will only need to check \cref{Bksuff}, \cref{Bkbase}, and \cref{epsfit} below. We will show that the budgets satisfy \begin{align}\label{Bksuff} B_{k+1} &\geq B_k + 8\eps_k . \end{align} (Roughly speaking, given $B_k$ we will set $\varepsilon_k$ as small as possible while keeping $R^{(k)}$ robust to the adversary with budget $B_k$. Then, we will set $B_{k+1}$ as small as possible, namely by taking equality in \cref{Bksuff}. Behind the scenes, we derive $B_k$ by solving the differential-equation equivalent of \cref{Bksuff} satisfied with equality.) We will show that \cref{budget} is satisfied in the base case, by showing that \begin{align} \label{Bkbase} B_k &\geq U_k \quad \text{for $k=n^{4/10}$} . \end{align} Then, \cref{budget} is established for all $k$ by induction on $k$: \begin{align} U_{k+1}-I_{k+1} &= (U_{k+1}-U_k)-(I_{k+1}-I_k)+ (U_k-I_k) \notag \intertext{which by \cref{Ukdef}, \cref{Ik}, and the inductive hypothesis \cref{budget} is} & \leq (\Wo {k+1}^s+\Wo {k+1}^t+7\eps_k)-(W\os k^s+W\os k^t) + B_k \notag \\ & \leq B_k + 8\eps_k \eqnote{see below} \label{epsfit} \\ & \leq B_{k+1} \eqnote{by \cref{Bksuff}} \label{Bkp} . \end{align} To justify \cref{epsfit} it suffices to show that $W\os{\kp}-W\os k$ is at most $0.1 \varepsilon_k$, and we do so in \cref{epsfitpf} for the uniform case and in \cref{expepsfitpf} for the exponential case. In both cases, $\rr0=\omega(1)$, and $W\os k+\rr0-W\os{\kp}=O(\varepsilon_k)$ (as used in going from \cref{path1} to \cref{path+}), making this conclusion unsurprising.% \footnote{In proving \cref{kbig,expkbig} we will set $\rr0=1$, so this reasoning does not apply. Indeed, in \cref{expkbig} (the large-$k$ exponential case) \cref{epsfit} would be false --- $W\os{\kp}-W\os k$ can be much larger than $\varepsilon_k$ --- but (to reiterate) it is not needed there, as we establish \cref{budget} directly. } \subsection{Robustness of \tp{$R$}{R}} \label{robustness} We wish to make $R$ robust against the adversary, so that after the deletions just described, $R$ should retain an $s$--$t$\xspace path w.h.p\xperiod} % {with high probability, so that \cref{XkWok} holds and $X_{k+1}$ is small. It will suffice to show that, to delete all $s$--$t$\xspace paths in $R$, \begin{align}\label{robustblurb} \parbox{0.8 \textwidth}{\emph{after deletion of $k$ edges incident on each of $s$ and $t$, an adversary would still have to delete middle edges of total cost more than $B_k$,}} \end{align} and thus it is powerless to do so. Obtaining this robustness requires choosing $\varepsilon_k$ sufficiently large in the construction. With reference to \cref{figR2}, on deletion of any $k$ edges on each of $s$ and $t$, the level-1 sets are in effect pruned to $V'_s$ and $V'_t$, each of cardinality $\rr0$. Should $V'_s$ and $V'_t$ have vertices in common, or if $t \in V'_s$ or $s \in V'_t$, then there is an $s$--$t$\xspace path. So, assume that $V'_s$ and $V'_t$ are disjoint and do not contain $s$ nor $t$. Consider only middle vertices $M' \subseteq M$ not appearing in $V'_s$ nor $V'_t$, i.e., $M'=M \setminus \set{V'_s \cup V'_t}$. We will have $\rr0=o(n)$, so $\card{M'} = n-2-2\rr0 > 0.99n$. Note that edges in $M' \times V'_s$, $M' \times V'_t$, $\set s \times V'_s$, and $\set t \times V'_t$ are all distinct. Consider a choice of the $k$ deletions on $s$ and $t$ to be fixed in advance. (We will eventually take a union bound over all such choices.) The weights of edges in $M' \times V'_s$ and $M' \times V'_t$ have not even been observed yet, so each has (unconditioned) $U(0,1)$ distribution, all are independent (by distinctness of the edges), and thus each such edge is included in $R$ with probability $\eps_k$, independently. A vertex $v \in M'$ is connected to $V'_s$ by \begin{align} \label{Zvs} Z_v^s & \sim \Bi(\rr0,\eps_k) \end{align} edges, with mean \begin{align}\label{kmedlambda} \lambda \coloneqq \E Z_v^s = \rr0 \eps_k . \end{align} Define $Z_v^t$ symmetrically, and note that $Z_v^s$ and $Z_v^t$ are i.i.d\xperiod. Intuitively, if $\lambda$ is small, $Z_v^s$ is usually 0, is 1 with probability about $\lambda$, and rarely any larger value. So, the probability that $v$ is connected to both $V'_s$ and $V'_t$ is about $\lambda^2$, in which case to destroy $s$--$t$\xspace paths through $v$ the adversary must delete an edge of cost at least $\varepsilon_k$. So, to delete all $s$--$t$\xspace paths, over the nearly $n$ vertices in $M'$ the adversary would have to delete edges of expected total weight at least \begin{align} \varepsilon_k \, n \, \lambda^2 \label{kmedtotalweight} . \end{align} We will choose $\varepsilon_k$ so that \begin{align} \label{epsk} \varepsilon_k \, n \, \lambda^2 > B_k , \end{align} which hopefully will ensure (see \cref{failprob}) that a path must remain (i.e., that $R$ is robust). Let us give a back-of-the-envelope calculation. In the uniform case we expect $W\os k$ to be about $k/n$, so letting $\rr0= \varepsilon_k n$ means that $W\os{\kk}-W\os k$ will be about $\varepsilon_k$, justifying \cref{path+}. Then \cref{kmedlambda} gives $\lambda=\eps_k^2 n$, so \cref{epsk} indicates that we need to take $\eps_k^5 n^3 > B_k$. As noted after \cref{Bksuff}, roughly speaking, we obtain $B_k$ and $\eps_k$ by solving this and \cref{Bksuff} with equality as a system of differential equations. \begin{remark} \label{failprob} This intuitive argument proves to be essentially sound, but to make it rigorous will take some work. Chiefly, $\Pr(Z_v^s > 0)$ is of course not exactly $\E Z_v^s = \lambda$ even when $\lambda$ is small, and we will also have to consider the case when $\lambda$ is large. Also, where the intuition is based on expectations, we must calculate the probability of the ``failure'' event that all paths can be deleted at a cost less than $B_k$. Finally, we must take the union bound of this failure event over all choices of root edges at $s$ and $t$ (but, as in the small-$k$ case, this turns out to change nothing). \end{remark} \section{Upper bound for large \tp{$k$}{k}, uniform model} \label{unifUB} In this section we fill in the details of the steps from \cref{largekUB} and show that they conclude the proof of the upper bound in \cref{Tmain}. Specifically, to control the \emph{path weights} (these emphasised keywords match section titles) we must show that \cref{path1} is at most \cref{path+}. For the \emph{adversary} we need only show \cref{budget}; as noted earlier, for large $k$ (\cref{kbig}) we will do this directly, while for medium $k$ (\cref{kmedium}) we will argue that the \emph{budgets} $B_k$ satisfy \cref{Bkbase} and \cref{epsfit}. And for \emph{robustness} we will prove that the probability of failure is small (i.e., it is unlikely that the adversary can destroy all $s$--$t$\xspace paths in $R^{(k)}$). \subsection{Claims, and implications for \cref{Tmain}}\label{uniclaims} We first state the two precise claims we make for large $k$, in two ranges. We use symbolic constants $C_B$, $C_\eps$, $C'_B$, and $C'_\varepsilon$ in the claims and the proofs, as it makes the calculations clearer. Whenever we encounter an inequality that the constants must satisfy, we will highlight with a parenthetical ``check'' that they do so. \begin{claim}\label{kmedium} For $k \in [n^{4/10}, n-14\sqrt n \,]$, let ${B_k} = \parens{C_B k n^{-3/5}+{(2n^{-1/5})}^{4/5}}^{5/4}$ and $\eps_k =C_\eps n^{-3/5} {{B_k}}^{1/5}$, with $C_B=32$ and $C_\eps=5$. Then, asymptotically almost surely, simultaneously for all $k$ in this range, \begin{equation}\label{eq:XkWk} X_{k+1} \leq W\os{\kp}^s + W\os{\kp}^t + 8 \eps_k. \end{equation} \end{claim} \noindent\textbf{Remark:} In proving \cref{kmedium} we will set \begin{align}\label{k0med} \rr0 & \coloneqq \varepsilon_k n . \end{align} From the definitions of $B_k$ and $\eps_k$ in \cref{kmedium}, both are increasing in $k$, and we will make frequent use of the following inequalities. For $n$ sufficiently large, \newline \noindent \begin{minipage}[t]{.5\textwidth} \begin{align} {B_k} &= \Theta \parens{ k^{5/4} n^{-3/4} + n^{-{1/5}}} \label{eq:Bk} \\ B_k &\leq B_n \leq 1.01 C_B^{5/4} n^{1/2} \label{eq:Bkupper}\\ B_k &\geq B_{n^{4/10}} \geq 2n^{-{1/5}} \label{eq:Bklower} \end{align} \end{minipage}% \begin{minipage}[t]{.5\textwidth} \begin{align} \eps_k &= \Theta \parens{ k^{1/4} n^{-3/4}+ n^{-{16/25}} } \label{eq:ek} \\ \eps_k &\leq \varepsilon_n \leq 1.01 C_\eps C_B^{1/4} n^{-1/2} \label{eq:ekupper} \\ \eps_k &\geq \varepsilon_{n^{4/10}} \geq 1.14 C_\eps n^{-{16/25}} \label{eq:eklower} . \end{align} \end{minipage}% \begin{claim}\label{kbig} For $k \in ( n-14\sqrt n, n-2]$, let \begin{align} B_k&=C_B' \sqrt n \quad \text{and} \quad \eps_k = C'_\varepsilon n^{-1/6} \label{kbigparams} \end{align} with $C'_B=78$ and $C'_\varepsilon=5$. Then, asymptotically almost surely, simultaneously for all $k$ in this range, \[ X_{k+1} \leq W\os{\kp}^s + W\os{\kp}^t + 8 \eps_k. \] \end{claim} \noindent\textbf{Remark:} In proving \cref{kbig} we will set $\rr0 \coloneqq 1$. Note that here $B_k$ and $\varepsilon_k$ are constants independent of $k$, but we retain the subscript for consistency with the notation of \cref{largekIntro}. We will prove the two claims shortly. \medskip \begin{proof}[Proof of the upper bound of \cref{Xkbounds} in \cref{Tmain}] Given $\varepsilon > 0$ from \cref{Tmain}, apply \cref{lemma:edge-orderstat} to the order statistics $W\os k^s$ and $W\os k^t$ with $\varepsilon$ in the lemma as our $\varepsilon/2$ and $a=n^{4/10}$. Then by \cref{kmedium} w.h.p\xperiod} % {with high probability, simultaneously for all $k \in [n^{4/10}+1, n-14\sqrt n]$, \begin{equation}\label{eq:implies-main} X_\k \leq W\os k^s + W\os k^t + 8\varepsilon_{k-1} \leq (1+\varepsilon/2) 2k/n + 8\varepsilon_{k-1} \leq (1+\varepsilon) (2k/n + \ln n/n); \end{equation} the key point is that $\varepsilon_{{k-1}} \leq \eps_k = o(k/n)$, which follows from~$\cref{eq:ek}$. Specifically, by \cref{eq:ek}, $\eps_k/(k/n) = O(k^{-3/4} n^{1/4}+k^{-1}n^{9/25})$, which by $k \geq n^{4/10}$ is $O(n^{-0.3}n^{0.25}+n^{-4/10}n^{0.36})=o(1)$. Likewise, by \cref{kbig}, inequality~\cref{eq:implies-main} holds w.h.p\xperiod} % {with high probability simultaneously for all $k \in [n-14\sqrt n, n-2]$. Again, we need only show that $\eps_k=o(k/n)$, which holds because here $k/n=\Theta(1)$ while by definition $\eps_k = o(1)$. \end{proof} We now prove the two claims, by filling in the details for \cref{structR,robustness}. \subsection{Structure \tp{$R$}{R}} With reference to \cref{structR}, all that we need to confirm is that $k+\rr0 \leq n-1$. For \cref{kmedium}, by hypothesis $k \leq n-14 \sqrt n$, and provided that $1.01 C_\eps C_B^{1/4} \leq 13$ (check), by~\cref{eq:ekupper} $\eps_k \leq 13 n^{-1/2}$, whereupon $\rr0 = \eps_k n \leq 13 \sqrt n$. For \cref{kbig}, with $\rr0=1$, $k+\rr0 \leq n-1$ is immediate. \subsection{Path weights} \label{pathweightsdetail} With reference to \cref{pathweights}, we establish that the bound \cref{path+} holds w.h.p\xperiod} % {with high probability simultaneously for all $k \geq n^{4/10}$. With $W\os k$ representing the cost of the $k$th cheapest edge incident on some fixed vertex (which we will take to be $s$ and then $t$ in turn), it suffices to show that \begin{align}\label{eq:Wkk0} W\os{\kk} &\leq W\os{\kp} + 1.1\eps_k \end{align} holds with high probability for all $k \geq n^{4/10}$. For \cref{kbig}, with $\rr0=1$, \cref{eq:Wkk0} is immediate. For \cref{kmedium}, with $\rr0=\varepsilon_k n$, generate the variables $W\os k$ by placing $n-1$ points uniformly at random on the unit interval $I$, associating $W\os k$ with the $k$th smallest point. It suffices to show that, w.h.p\xperiod} % {with high probability, each interval $(W\os{\kp}, W\os{\kp}+1.1\eps_k)$ contains at least $\rr0$ points. For all $k \in [n^{4/10}, n-14\sqrt n-1\,]$, \cref{kmedium} has $\eps_k \geq n^{-0.99}$ by~\cref{eq:eklower}, so \cref{intervals} shows that w.p.\ $1-\exp(-\Omega(n^{0.01}))$, every interval of length $\geq 1.1\eps_k$ in $[0,1]$ contains at least $\rr0 \eqdef \eps_k n$ points, and in particular this holds for all the intervals $(W\os{\kp}, W\os{\kp}+1.1\eps_k)$. We assume henceforth that the graph $G$ is ``good'' in the sense that \cref{eq:Wkk0} holds for all $k \geq n^{4/10}$ for vertices $s$ and $t$, and that for all $k \leq n^{4/10}$ we have the upper bounds on $X_k$ from~\cref{Xkbounds}, as proved to hold w.h.p.\ in \cref{sec:k-small}. \subsection{Adversary} With reference to \cref{subadv}, we need only verify \cref{budget}, and this will be done in the next subsection. \subsection{Budgets \tp{$B_k$}{B\_k}} With reference to \cref{budgets}, we first establish \cref{epsfit}. This follows from \begin{align}\label{epsfitpf} \Wo {k+1}^s-W\os k^s & \leq 0.1 \eps_k . \end{align} The reasoning for this is the same as for \cref{eq:Wkk0}: each interval of length $0.1 \eps_k$ contains at least one point. The parameters are trivial to check. Next, we show that the parameters of \cref{kmedium} satisfy \cref{budget}, for which as argued in \cref{budgets} it suffices to show that they satisfy \cref{Bksuff} and \cref{Bkbase}. We start with \cref{Bkbase}, the base case. Here $k=n^{4/10}$, $B_k \geq 3n^{-2/10}$ from~\cref{eq:Bklower}, and $U_k = 3n^{-2/10}$ from \cref{Ukdef}, establishing \cref{Bkbase}. To establish \cref{Bksuff}, first note that $\frac{\partial}{\partial k} {B_k} = \tfrac 54 C_B {{B_k}}^{1/5}n^{-3/5}$ is an increasing function. Then, \[ B_{k+1}-{B_k} \geq \frac{\partial}{\partial k} {B_k} = \frac 54 C_B {{B_k}}^{1/5}n^{-3/5} = \frac 54 \frac {C_B}{C_\eps} \eps_k \geq 8\eps_k, \] since $C_B \geq \tfrac{8 \cdot 4}{5} C_\eps$ (check). We now establish \cref{budget} for the parameters of \cref{kbig}. With ${k^\star}= \floor{n-14\sqrt n}$, the point where \cref{kmedium} ends and just before \cref{kbig} begins, the previous case showed that $B_{k^\star} \geq U_{k^\star} - I_{k^\star}$, and by~\cref{eq:Bkupper} $B_{k^\star} \leq 77 \sqrt n$. Then, for $k$ from ${k^\star}+1$ to $n-2$, \begin{align} U_k - I_k &= (U_{k^\star} - I_{k^\star}) + [(U_k-U_{k^\star}) - (I_k - I_{k^\star})] \notag \\&\leq B_{k^\star} + \left[ \sum_{i={k^\star}+1}^k (W\os i^s + W\os i^t + 7 \varepsilon_{i-1}) - \sum_{i={k^\star}}^{k-1} (W\os i^s + W\os i^t)\right] \eqnote{see \cref{Ukdef} and \cref{Ik}} \notag \\&\leq B_{k^\star} + \sum_{i={k^\star}+1}^{n-2} 7 \varepsilon_{i-1} + (W\os k^s + W\os k^t - \Wo{k^\star}^s - \Wo{k^\star}^t) \notag \\&\leq 77 \sqrt n + (14\,\sqrt n) \cdot 7 C'_\varepsilon n^{-1/6} + 2 \eqnote{see \cref{kbigparams}} \notag \\& \leq 78 \sqrt n \notag \\& \leq B_k \eqnote{see \cref{kbigparams}}, \label{budgetsClaim2} \end{align} using that $C_B' \geq 78$ (check). \subsection{Minimum of two binomial variables} Before addressing robustness of the structure $R$, we require a lemma (\cref{lemma:BinMin}) on the minimum $Z$ of two i.i.d\xperiod binomial $\Bi(n,p)$ random variables. There is a genuine difference in the cases when the common mean $\lambda=np$ is large or small: if $\lambda$ is large then $Z$ is likely to be close to $\lambda$, making $\E Z = \Theta(\lambda)$; if $\lambda$ is small then $Z$ will most often be 0, occasionally 1 (with probability about $\lambda^2$), and rarely anything larger, making $\E Z = \Theta(\lambda^2)$. The lemma relies on the following property of the median of a binomial random variable. (A weaker form of~\cref{Lbigla} and thus of \cref{lemma:BinMin} can be obtained from \cref{lemma:BinDev} in lieu of using the median.) \begin{theorem}[Hamza~{\cite[Theorem 2]{Hamza1999}}] A binomial random variable $X$ has median satisfying $|\operatorname{Med}(X)-\E X| \leq \ln 2$. \end{theorem} \noindent In this discrete setting $\operatorname{Med}(X)$ is not unique: it can be any value $m$ for which $\Pr(X \leq m) \geq 1/2$ and $\Pr(X \geq m) \geq 1/2$. \cite{Hamza1999} defines it uniquely as the smallest integer $m$ such that $\Pr(X \leq m) >1/2$; as desired, this gives $\Pr(X \geq \operatorname{Med}(X)) = 1-\Pr(X \leq \operatorname{Med}(X)-1) \geq 1-1/2=1/2$. (For other results on the binomial median see Kaas and Buhrman~\cite{Kaas1980}, in particular, Corollary~1. Stronger results for the Poisson distribution are given by Choi~\cite{Choi1994}, proving a conjecture of Chen and Rubin, and by Adell and Jodr\'a~\cite{Adell2005}.) \begin{lemma}\label{lemma:BinMin} Let $Z_1, Z_2$ be i.i.d\xperiod $\Bi(n, p)$ random variables, $Z \coloneqq \min(Z_1, Z_2)$ and $\lambda \coloneqq \E Z_1 = np$. \begin{enumerate}[(1)] \item If $\lambda \geq 2$, then \begin{align} \Prob(Z \geq 0.65 \lambda) &> 1/4. \label{Lbigla} \end{align} \item If $\lambda \leq 2$, then \begin{align} \Prob(Z \geq 1) &> 0.18 \lambda^2 . \label{Lsmallla} \end{align} \end{enumerate} \end{lemma} \begin{proof} In the first case, \[ \operatorname{Med}(Z_1) \geq \lambda-\ln 2 = \frac{\lambda-\ln 2}\lambda \lambda \geq \frac{2-\ln 2}2 \lambda \geq 0.65 \lambda , \] so $\Prob \parens{Z_1 \geq 0.65 \lambda} \geq \Prob \parens{Z_1 \geq \operatorname{Med}(Z_1)} \geq 1/2$. The same holds of course for $Z_2$, and the result follows by independence. In the second case we again use independence, and here \begin{equation*} \Prob \parens{Z_1 \geq 1} = 1-{(1-p)}^n \geq 1-\exp(-\lambda) = \frac{1-\exp(-\lambda)}{\lambda} \cdot \lambda \geq 0.43\lambda . \end{equation*} The last inequality comes from minimising $\frac{1-\exp(-x)}{x}$ over $0 \leq x \leq 2$; the function is decreasing so the minimum is at $x=2$. \end{proof} \subsection{Robustness in \cref{kmedium}} \label{kmedrobust} With reference to \cref{robustness}, let us complete the robustness argument for \cref{kmedium}, showing that \cref{robustblurb} holds with high probability. Here we have taken $\rr0 = \varepsilon_k n$, so that the number of edges from a middle vertex to $V'_S$ (see \cref{Zvs}) is $Z^s_v \sim \Bi(\varepsilon_k n, \varepsilon_k)$, with mean $\lambda=\rr0 \varepsilon_k = \varepsilon_k^2 n$ (see \cref{kmedlambda}). Recall that if $\lambda$ is small we expect (see \cref{kmedtotalweight}) that to destroy all paths the adversary will have to delete edges of total weight at least $\varepsilon_k \, n \, \lambda^2 = \eps_k^5 n^3$, which will exceed $B_k$. And, if $\lambda$ is large, then each $Z_v$ will have expectation close to $\lambda=\eps_k^2 n$, for a total cost $\eps_k n$ times larger, namely $\eps_k^3 n^2$, and again this exceeds $B_k$. We now replace these rough calculations with detailed probabilistic ones, applying \cref{lemma:BinMin} to $Z_v$ in the two cases of $\lambda$ small and large. For the adversary to delete all $s$--$t$\xspace paths via $v$, he must delete at least \[ Z_v\coloneqq \min(Z_v^s, Z_v^t) \] edges, and to destroy all paths he must delete at least \[ N\coloneqq \sum_{v \in M'} Z_v \] edges. As described in \cref{robustness}, we imagine a fixed deletion of $k$ edges on each of $s$ and $t$ giving neighbour sets $V'_s$ and $V'_t$ and a set $M'$ of middle vertices; we will eventually take a union bound over all such choices. \medskip \noindent \tmbf{If $\lambda \geq 2$}, then by \cref{lemma:BinMin}, for each $v \in M'$, $\Pr(Z^s_v \geq 0.65 \lambda) \geq 1/4$. Thus, $N$ stochastically dominates $0.65\lambda \cdot \Bi(0.99n, 1/4)$, with expectation $> 0.1608 \lambda n$. We shall consider it a \emph{failure} if $N \leq 0.16 \lambda n$. Assuming success, since each edge costs at least $\eps_k$ to delete, it costs at least $0.16 \eps_k \lambda n = 0.16 \eps_k^3 n^2$ to delete them all. This exceeds $B_k$: \begin{align*} \frac{0.16 \cdot \eps_k^3 n^2}{B_k} &= 0.16 \cdot C_\eps^3 n^{-9/5} B_k^{-2/5} n^2 \eqnote{by definition of $\varepsilon_k$} \\& \geq 0.15 \cdot C_\eps^3 C_B^{-1/2} n^{1/5} n^{-1/5} \eqnote{by~\cref{eq:Bkupper}} \\ &> 1, \end{align*} using that $0.15 \cdot C_\eps^3 C_B^{-1/2}>1$ (check). Failure means that $N/(0.65 \lambda) \sim \Bi(0.99n, 1/4) \leq (0.16/0.65)n$. Noting that $0.99 \cdot 1/4 > 0.16/0.65$, by \cref{lemma:BinDev}, the probability of failure is $\exp(-\Omega(n))$. By the union bound, the total of the failure probabilities, over all rounds (values of $k$) and all adversary choices of the $k$ root edges at $s$ and $t$, is small: \begin{align} \label{case1failure} \sum_k {{\binom{k+\rr0}{\rr0}}^2} & \cdot \exp(-\Omega(n)) \\& \leq \sum_k \, {(n^{\rr0})}^2 \exp(-\Omega(n)) \notag \\ &= \sum_k \exp\parens{2\eps_k n \ln n-\Omega(n)} \quad\text{(by $\rr0=\eps_k n$)} \notag \\&\leq n \exp(-\Omega(n)) \quad\text{(using $\eps_k n = O(n^{1/2})$ from~\cref{eq:ekupper})} \notag \\& = o(1) . \notag \end{align} \medskip \noindent \tmbf{If $\lambda < 2$}, then by \cref{lemma:BinMin} $N$ stochastically dominates $\Bi(0.99n, 0.18\lambda^2)$, with expectation $> 0.175 \lambda^2 n$. We shall consider it a \emph{failure} if $N \leq 0.17 \lambda^2 n = 0.17 \eps_k^4 n^3$. Each edge costs at least $\eps_k$ to delete. Assuming success, it thus costs at least $0.17 \eps_k^5 n^3$ to delete them all, which exceeds $B_k$: \begin{align*} \frac{0.17 \eps_k^5 n^3}{B_k} &= 0.17 C_\eps^5 \eqnote{by definition of $\varepsilon_k$} \\ &> 1, \end{align*} using that $0.17 C_\eps^5>1$ (check). By \cref{lemma:BinDev}, the probability of {failure} is \begin{equation}\label{eq:N2} \Prob \parens{ N \leq 0.17 \eps_k^4 n^3} = \exp(-\Omega(\eps_k^4 n^3)). \end{equation} Over all rounds (values of $k$) and adversary choices of edges incident to $s$ and $t$, the total failure probability is at most \begin{align} \sum_k {\binom{k+\rr0}{\rr0}}^2 & \cdot \Prob \parens{N < 0.17 \eps_k^4 n^2} \notag \\ &\leq \sum_k \exp\parens{2\eps_k n \ln n- \exp(-\Omega(\eps_k^4 n^3))} \notag \\&\leq n \exp(-\Omega(\eps_k^4 n^3)) , \notag \intertext{because $\eps_k n \ln n$ is dominated by $\eps_k^4 n^3$: the latter is larger by a factor $\eps_k^3 n^2/\ln n$, which by~\cref{eq:eklower} is $\Omega(n^{-48/25}n^2/\ln n)=\Omega(n^{2/25}/\ln n)=\omega(1)$. Continuing, this is} &\leq n \exp(-\Omega(n^{11/25})) \eqnote{invoking~\cref{eq:eklower} again} \notag \\& = o(1) . \label{case2failure} \end{align} \bigskip \subsection{Robustness in \cref{kbig}} \label{kbigrobust} Again, our aim is to establish robustness of $R$ by showing that \cref{robustblurb} holds with high probability, and the argument is similar to but simpler than that of \cref{kmedrobust}. Since $\rr0=1$, both $V'_s$ and $V'_t$ have size 1. For a vertex $v \in M'$, let $Z_v$ be the number of paths from $V'_s$ to $V'_t$ via $v$. There is only one such possible path, hence \[ Z_v \sim \Bern \parens{\eps_k^2}. \] To destroy all $s$--$t$\xspace paths the adversary must delete at least \[ N\coloneqq \sum_{v \in M'} Z_v \] edges. $N$ stochastically dominates $\Bi(0.99n, \eps_k^2)$, which has expectation $0.99 \eps_k^2 n$. We declare the event $N \leq 0.98 \eps_k^2 n$ a \emph{failure}. Assuming success, destroying all $s$--$t$\xspace paths would cost at least $\eps_k N \geq 0.98 \eps_k^3 n$. This exceeds $B_k$, since \begin{equation*} \frac{0.98 \eps_k^3 n}{B_k} = \frac{0.98 {C'_\varepsilon}^3}{C'_B}, \end{equation*} and $0.98 {C'_\varepsilon}^3 > C'_B$ (check). The probability of failure is \begin{equation}\label{eq:N3} \Prob \parens{ N \leq 0.98 \eps_k^2 n} = \exp(-\Omega(\eps_k^2 n)) = \exp(-\Omega(n^{2/3})). \end{equation} Over all rounds and adversary choices, using that $\binom{k+\rr0}{\rr0} = \binom{k+1}1 \leq n$, the total failure probability is at most \begin{align} \sum_k {\binom{k+\rr0}{\rr0}}^2 & \cdot \Prob(N \leq 0.98 \varepsilon^2 n) \notag \\ &\leq (14\sqrt n) \, n^2 \, \exp (-\Omega(n^{2/3})) \eqnote{by~\cref{eq:N3}} \label{eq:unifailure3} \\& = o(1) . \notag \end{align} \section{Lower bound}\label{lowerbound} In this section, we establish the lower bound in \cref{Xkbounds} of \cref{Tmain}. \cref{LBsmallk} establishes the lower bound on $X_k$ directly for $k \leq \sqrt{\ln n}$. Values $k \geq \sqrt{\ln n}$ are treated in the subsequent parts. In \cref{LBrunning}, \cref{lemma:S_k-lower} establishes a lower bound on the running totals $S_k$, \begin{align}\label{Skdef} S_k \coloneqq \sum_{i=1}^{k} X_i . \end{align} In \cref{LBlargek}, \cref{lemma:X_k-lower} obtains a lower bound on $X_k$ using \cref{lemma:S_k-lower}'s lower bound on $S_k$, the previously established upper bound on $X_k$ from \cref{Tmain}, and the monotonicity of $X_k$. \subsection{Lower bound for small \tp{$k$}{k}} \label{LBsmallk} We begin with $k \leq \sqrt{\ln n}$. For any fixed $\varepsilon>0$, we know from~\cite{Janson123} that w.h.p\xperiod} % {with high probability \begin{align}\label{X1lower} X_1 &> (1-\varepsilon/2)\frac{\ln n}{n} . \end{align} Assuming that~\cref{X1lower} holds, it follows immediately, and deterministically, that for all $k \leq \sqrt{\ln n}$, \begin{align}\label{Xklower1} X_k &\geq X_1 \geq (1-\varepsilon/2)\frac{\ln n}{n} \geq (1-\varepsilon)\frac{2k+\ln n}{n} . \end{align} The first inequality holds because the sequence $X_k$ is monotone increasing, the next by assumption on $X_1$, the next by $k = o(\ln n)$. \subsection{Lower bound on the running totals} \label{LBrunning} \begin{lemma}\label{lemma:S_k-lower} For any $\varepsilon>0$, w.h.p\xperiod} % {with high probability, simultaneously for every $k \leq n-1$, \begin{align}\label{Sklower} S_k &\geq (1-\eps)\sum_{i=1}^{k}\left(\frac{2i+\ln n}{n}\right). \end{align} \end{lemma} \begin{proof} Write $W\os i^s$ and $W\os i^t$ for the order statistics of edge weights out of $s$ and $t$, respectively. By \cref{lemma:edge-orderstat}, w.h.p\xperiod} % {with high probability, \begin{equation}\label{eq:edge-orderstat} W\os k^s, W\os k^t \in \left[\left(1-\varepsilon/2 \right) \frac k n, \left(1+\varepsilon/2 \right) \frac k n \right] \quad \text{ for all } k \geq \sqrt{\ln n} , \end{equation} and we will assume throughout the proof that~\cref{eq:edge-orderstat} holds. We prove the assertion in two ranges of $k$. \medskip\noindent\tmbf{For $\ln^{11/10} n \leq k \leq n-1$,} the $k$ paths must use at least $k-1$ edges on each of $s$ and $t$, all distinct ($k$ edges each, ignoring the edge $\set{s,t}$ if it is used). Then, using~\cref{eq:edge-orderstat}, we get that w.h.p\xperiod} % {with high probability, for all $k$ in the range, \begin{align} S_k &\geq \sum_{i=1}^{k-1} \left( W\os i^s + W\os i^t \right) \geq \sum_{i=\sqrt{\ln n}}^{k-1} (1-\varepsilon /2) \frac{2 i}n \notag \\& = (1-\varepsilon/2) \parens{ \sum_{i=1}^{k} {\frac{2i+\ln n}n} - \sum_{i=1}^{k} \frac{\ln n}n - \sum_{i=1}^{\sqrt{\ln n}-1} {\frac{2i}n} - {\frac{2k}n} } \label{SbigK1} \\& \geq (1-o(1)) (1-\varepsilon/2) \sum_{i=1}^{k} {\frac{2i+\ln n}n} \label{SbigK2} \eqnote{see below} \\&\geq (1-\eps)\sum_{i=1}^{k}\left(\frac{2i+\ln n}{n}\right) \label{SbigK} . \end{align} To justify \cref{SbigK2} it suffices to show that the first sum in \cref{SbigK1} is of strictly larger order than the other terms. The first sum is at least $\sum_{i=k/2}^{k} 2i/n = \Omega(k^2/n)$, which since $k \geq \ln^{11/10}n$ is also $\Omega(k \ln^{11/10}n/n)$ and $\Omega(\ln^{22/10}n/n)$; we will use all three formulations. The second term is of order $O(k \ln n/n)$, negligible compared with the middle formulation. The third term is $O(\ln^{2/3}n/n)$, negligible compared with the last formulation. And the fourth term, of order $O(k/n)$, is negligible compared with the first formulation. \medskip\noindent\tmbf{For $1 \leq k \leq \ln^{11/10} n$,} let $\delta=\varepsilon/3$ and let $G'=G-s-t$. Let $N_s$ and $N_t$ be the endpoints of the cheapest $\ln^3 n$ edges out of $s$ and $t$ respectively. Note that these sets are independent of the edge weights of $G'$. If any path $P_i$, $i \leq k$, uses a root edge (edge incident on $s$ or $t$) \emph{not} among the $\ln^3 n$ cheapest edges of $s$ or $t$, then by~\cref{eq:edge-orderstat} this edge costs at least $(1-\varepsilon) \ln^3 n / n$, thus $S_k \geq (1-\varepsilon) \ln^3 n / n$. Then \cref{Sklower} follows because this is larger than the RHS of~\cref{Sklower}, namely $\Theta((k^2+k\ln n)/n) = O(\ln^{11/5} n/n)$ for this range of $k$. Thus we may assume that for all $i \leq k$, each path $P_i$ goes via some $s' \in N_s, \; t' \in N_t$. For $s' \in N_s, \; t' \in N_t$, define $A(s',t')$ to be the event that $t'$ is one of the ${(n-2)}^{1-\delta}$ nearest vertices of $s'$, by cost, in $G'$. Clearly, for each pair $s', t'$, $\Pr(A(s',t')) = {(n-2)}^{-\delta}$. Let $A$ be the union of these events, i.e., the event that any such pair has this property. By the union bound, \[ \Prob(A) \leq {\left(\ln^3 n \right)}^2 {(n-2)}^{-\delta} = o(1) . \] We assume henceforth that $A$ does not hold: the $\ln^3n$ cheapest root edges at $s$ and $t$ do not happen to sample any ``nearest'' pairs in $G'$. By assumption that $A$ does not hold, in the \emph{exponential} model (where each edge is i.i.d.\ $\Exp(1)$) for $G'$, for each $s' \in N_s, \; t' \in N_t$, the distance $d(s',t')$ stochastically dominates $Y \sim \sum_{i=1}^{n^{1-\delta}} \Exp(i(n-2-i))$ by \cref{treeX}. We have $\E Y = (1+o(1))(1-\delta) \ln n / n$ by ~\cref{treeDia} (just adjusting its last equation where the value of $d$ is substituted in). Applying \cref{exptail}'s \cref{exp.3} with $\mu=\E Y$ as above, $a^\star=n-3$, and $\lambda=1-\delta$, that in the exponential model $G'$, \begin{align*} \Prob & \left(d_{G'}(s',t') \leq (1-\delta) \frac{(1+o(1))(1-\delta) \ln n}{n} \right) \; \leq \; \exp\parens{ -\Theta(n \cdot \ln n/n \cdot \delta^2) } \; = \; n^{-\Theta(1)}. \end{align*} Since ${(1+o(1))(1-\delta)}^2 \geq (1-\tfrac34 \varepsilon)$, by the union bound this implies, still in the exponential model $G'$, \begin{equation}\label{eq:not-too-close} \Prob\Big( \exists s' \in N_s, t' \in N_t \colon d_{G'}(s',t') \leq (1-\tfrac34 \varepsilon) \ln n / n \Big) \; \leq \; {\left(\ln^3 n \right)}^2 n^{-\Theta(1)} \; = \; o(1) . \end{equation} By standard coupling arguments (see \cref{remark:blackbox}), this also implies that \cref{eq:not-too-close} holds in the \emph{uniform} model $G$ in which we are working. Thus w.h.p\xperiod} % {with high probability, for all $s' \in N_s, t' \in N_t$, we have $d_{G'}(s',t') \geq (1-\tfrac34 \varepsilon) \ln n$; assume this holds. We already assumed that each path $P_i$, $i\leq k$, goes via some $s' \in N_s, t' \in N_t$, so its non-root edges contribute at least $d_{G'}(s',t') \geq (1-\tfrac34 \varepsilon) \ln n/n$ to $S_k$. Then, for all $k$ in this range, \begin{align} S_k &\geq \sum_{i=1}^{k} (1-\tfrac34 \varepsilon) \frac{\ln n}n + \sum_{i=1}^{k-1} \left( W\os i^s + W\os i^t \right) \notag \\& \geq \sum_{i=1}^{k} (1-\tfrac34 \varepsilon) \frac{\ln n}n + (1-\tfrac12 \varepsilon) \sum_{i=\sqrt{\ln n}}^{k-1} \frac{2i}n \eqnote{by \cref{eq:edge-orderstat}} \label{lbx1} \\& \geq (1-\eps)\sum_{i=1}^{k}\left(\frac{2i+\ln n}{n}\right). \notag \end{align} To justify the final inequality, rewrite the second sum in \cref{lbx1} as $\sum_{i=1}^{k} \frac{2i}n - \frac{2k}n - \sum_{i=1}^{\sqrt{\ln n}-1} \frac{2i}n$ and observe that both its second term, $2k/n$, and its final term, which is of order $O(\sqrt{\ln n}^2/n)$, are negligible compared with the first sum in \cref{lbx1}, which is of order $\Omega(k \ln n/n)$. \end{proof} \subsection{Lower bound for large \tp{$k$}{k}} \label{LBlargek} \begin{lemma}\label{lemma:X_k-lower} For any $\varepsilon>0$, w.h.p\xperiod} % {with high probability, simultaneously for every $k \in [\sqrt{\ln n}, n-1]$, \[ X_k \geq (1-\varepsilon)\parens{\frac{2k + \ln n}{n} }. \] \end{lemma} \begin{proof} Let $\delta=\varepsilon^2/9$ and define \begin{equation} c_k = \frac{2k+\ln n}{n}, \quad L_k = (1-\delta)\sum_{i=1}^k c_i, \quad U_k = (1+\delta) \sum_{i=1}^k c_i. \end{equation} W.h.p\xperiod} % With high probability, simultaneously for all $k$, $S_k \geq L_k$ (by \cref{lemma:S_k-lower}) and $S_k \leq U_k$ (by the upper bound of~\cref{Tmain}, already proved). Henceforth, assume that both hold, so $L_k \leq S_k \leq U_k$. The rest of the argument is deterministic. For any positive integer $t<k$, using that $X_k$ is monotone increasing, we have \begin{align} t X_k & \geq X_k +\cdots+ X_{k-t+1} \notag \\& = S_k - S_{k-t} \notag \\& \geq L_k - U_{k-t} \label{LBform} . \end{align} Thus \begin{align*} X_k &\geq \frac1t \parens{L_k - U_{k-t}} = \frac1t \parens{ (1-\delta)\sum_{i=1}^k c_i - (1+\delta)\sum_{i=1}^{k-t} c_i } \\&\geq \frac1t \parens{ \sum_{i=k-t+1}^k c_i -2\delta \sum_{i=1}^{k} c_i } \geq \frac1t \parens{ t c_{k-t} - 2\delta k c_k } = c_{k-t} - \frac{2 \delta k c_k}{t} \\&= c_k- \frac{2t}{n} - \frac{2 \delta k c_k}{t} \\& \geq c_k- \frac{t c_k}{k} - \frac{2\delta k}{t} c_k \quad\text{(using that $c_k/k>2/n$)} \\&= c_k \parens{ 1- \frac{t}{k} - \frac{2\delta k}{t} } . \end{align*} Ignoring integrality for a moment, setting $t=k\sqrt{2\delta}$ would make the last expression $c_k (1- 2\sqrt{2\delta})$. Since this $t=\Theta(k)=\omega(1)$, rounding it can be seen to change the expression by a factor $1+o(1)$, so we may safely write \begin{align*} X_k &\geq c_k (1-3\sqrt{\delta}) = (1-\varepsilon) \frac{2\k + \ln n}{n}. \end{align*} \end{proof} \section{Exponential model} \label{ExpBounds} In this section we prove \cref{Texp}, the analogue of \cref{Tmain} for exponentially distributed edge weights. For small $k$, results for the exponential case follow from those for the uniform. We first argue that the upper bound of \cref{Tmain} also holds in the exponential case for any $k=o(n)$. Couple the two models, so that any edge of weight $w=o(1)$ in one model has cost $w'=w(1+o(1))$ in the other. The uniform-model upper-bound constructions in \cref{sec:k-small} (for $k=o(n^{1/2})$) and \cref{largekUB,unifUB} (for larger $k$) only use edges of weight $o(1)$ (when $k=o(n)$), and therefore the same upper bounds hold for the exponential model; the multiplicative difference of $1+o(1)$ can be subsumed into the factor $1+\varepsilon$ already present. (In the construction of \cref{largekUB,unifUB}, the ``middle edges'' are of cost $o(1)$ for \emph{all} $k$, but the ``incident edges'' have larger cost for $k$ large. In particular, for large $k$, \cref{eq:Wkk0} will no longer hold in the exponential case until we adjust $\rr0$ and $\varepsilon_k$ appropriately.) For the lower bound too, the argument in \cref{lowerbound} carries over for all $k=o(n)$. The lower bounds $L_k$ on the prefix sums $S_k$ derived in \cref{LBsmallk,LBrunning} carry over to the exponential case because the edge costs are equal to within $1+o(1)$ factors in the two models. The upper bounds $U_k$ on the prefix sums are simply the sums of the individual upper bounds on $X_k$, and we have just argued that these change only by a $1+o(1)$ factor. \cref{LBlargek} only uses $L_k$ and $U_k$ to derive lower bounds on $X_k$, so with these both changed only by $1+o(1)$ factors, its results carry over verbatim. Our task, then, is to prove the upper and lower bounds in \cref{Texp} for larger $k$. For the upper bound, arguing for $k> n^{0.4}$ (there is no advantage to a larger starting value), we use the same approach as for the uniform model in \cref{largekUB}. For the lower bound, we argue for $k \geq n^{9/10}$. Unfortunately, the method used in \cref{lowerbound} for the uniform distribution does not extend; let us explain why. The lower bound there came from \cref{LBform}, $t X_k \geq L_k - U_{k-t}$, valid for any functions $L$ and $U$ with $L_k \leq S_k \leq U_k$. Here, we would take $L_k$ as the sum $I_k$ of incident edges as in \cref{Ik} and $U_k$ as the sum of the $X_k$ upper bounds as in \cref{XkWok}. Recall that we defined $B_k$ so that $B_k \geq U_k-L_k$, as in \cref{budget}. Then we can rewrite the previous lower bound approach as $X_k \geq \frac1t (L_k-U_{k-t}) \geq \frac1t (L_k-L_{k-t}) + \frac1t (L_{k-t}-U_{k-t}) \geq \frac1t \sum_{i=k-t}^{k-1} W\os i - \frac1t B_{k-t}$. For large $k$, $W\os k$ and therefore $X_k$ are $\Theta(\ln n)$. Since the $B_k$ grow to size $\Theta(n^{1/2})$ (in the exponential case as well as the uniform case), we are thus limited by the second term to $t=\Omega(n^{1/2+o(1)})$. However, from \cref{muX}, such a large value of $t$ would mean that the average given by the first term is significantly different from $W\os k$. The desired lower bound would be immediate if we could claim that $P_k$ necessarily used the $k$th cheapest edge on $s$ (of cost $W\os k^s$) or a later one, and likewise for $t$. We will prove something close to this. We argue in \cref{ExpLB} that every pair of vertices (excluding both $s$ and $t$) is joined by a path of cost at most $\delta$ (for some small $\delta$ to be specified) that is edge-disjoint from \emph{all} $P_i$, $i=1,\ldots,n-1$. We will show that this implies that path $P_k$ uses an edge on $s$ that is at most $\delta$ cheaper than $W\os k^s$, and likewise for $t$, yielding a sufficient lower bound. \subsection{Claims, and implications for \cref{Texp}} \label{ExpUpper} In order to establish upper bounds on $X_k$ in the exponential model, we use the same structure $R^{(k)}$ as described in \cref{structR}. Then \cref{XkWok} follows as before, and we can continue to define $U_k$ as in \cref{Ukdef}. For convenience define \begin{align}\label{kbardef} \bar{k} &= n-k . \end{align} As before we will treat $k$ in two ranges, and we start now with the smaller range. \begin{claim}\label{expkmedium} For $k \in [n^{4/10}, n-\sqrt{n} \,]$, let \begin{align}\label{ExpBk} B_k \coloneqq \parens{ \frac{2 n^{1/25}+C_B(n^{3/5}- \bar{k}^{3/5})}{n^{1/5}} }^{5/4} \quad \text{and} \quad \varepsilon_k \coloneqq C_\eps B_k^{1/5} n^{-1/5} \bar{k}^{-2/5} , \end{align} with $C_B=44$ and $C_\eps=4$. Then, asymptotically almost surely, \begin{equation}\label{ExpXkWk} X_{k+1} \leq W\os{\kp}^s + W\os{\kp}^t + 8 \eps_k. \end{equation} \end{claim} \noindent\textbf{Remark:} In proving \cref{expkmedium} we will set \begin{align}\label{expk0med} \rr0 & \coloneqq \eps_k \bar{k} . \end{align} because it roughly equates $\Wo{k+\rr0}-W\os k$ and $\eps_k$; see \cref{muX}. In this regime integrality is not an issue: $\rr0$ is large, per \cref{eq:Expr0}. It is clear that both $B_k$ and $\varepsilon_k$ in \cref{ExpBk} are increasing in $k$, even over the larger range $k \in [0,n]$. We will make use of the following bounds, holding for $n$ sufficiently large. Here, \cref{eq:ExpBkUpper} uses that at $k=n-\Theta(\sqrt{n})$, $\bar{k}^{3/5}$ dominates $2n^{1/25}$, while \cref{eq:ExpBkLower} takes $k=0$. \begin{align} B_k &\leq B_{n-\sqrt{n}} \leq C_B^{5/4} n^{1/2} \label{eq:ExpBkUpper}\\ B_k &\geq B_{n^{4/10}} \geq 2n^{-{1/5}} \label{eq:ExpBkLower} \\ \eps_k & \leq C_\eps B_k^{1/5} n^{-1/5} n^{-1/2 \cdot 2/5} \leq C_\eps C_B^{1/4} n^{-3/10} \label{eq:ExpEkUpper} \\ \eps_k & \geq C_\eps {(B_{n^{4/10}})}^{1/5} \, n^{-1/5} \, \bar{k}^{-2/5} \geq C_\eps n^{-{6/25}} \bar{k}^{-2/5} \label{eq:ExpEkLower} \\ \rr0 &= \bar{k} \eps_k \stackrel{\cref{eq:ExpEkLower}}{\geq} C_\eps n^{-6/25} \bar{k}^{3/5} \geq C_\eps n^{3/50} . \label{eq:Expr0} \end{align} \begin{claim}\label{expkbig} For $k \in (n-\sqrt{n}, n-2 \,] $, let \begin{align}\label{ExpBk2} B_k \coloneqq C_B' \sqrt n \quad \text{and} \quad \varepsilon_k \coloneqq C'_\varepsilon n^{-1/6} , \end{align} with $C_B=115$ and $C_\eps=5$. Then, asymptotically almost surely, simultaneously for all $k$ in this range, \begin{align} X_{k+1} &\leq W\os{\kp}^s + W\os{\kp}^t + 8 \eps_k. \label{ExpXkUB} \end{align} \end{claim} \noindent\textbf{Remark:} In proving \cref{expkbig} we will set \begin{align}\label{expk0big} \rr0 & \coloneqq 1 . \end{align} As in \cref{kbig}, $B_k$ and $\varepsilon_k$ are constants independent of $k$, but we retain the subscript for consistency with the notation of \cref{largekIntro}. \begin{proof}[Proof of the upper bounds in \cref{Texp}] Analogous to the argument in \cref{uniclaims}, it is sufficient to check that $\eps_k = o(\E W\os k)$. Since $\E W\os k \sim \ln \parens{\frac{n}{n-k}} \geq \frac kn$ (see \cref{muX}), it is enough to show that $\varepsilon_k = o(k/n).$ For $k \leq n^{0.99} = o(n)$, by first-order approximation, \begin{align}\label{n35} n^{3/5}-\bar{k}^{3/5} &\eqdef n^{3/5}-(n-k)^{3/5} \sim \tfrac 35 n^{-2/5} k , \end{align} so $ B_k = \Theta \parens{n^{-1/5} + n^{-3/4} k^{5/4}} $. Hence, from \cref{expkmedium}, specifically \cref{ExpBk}, \begin{align} \varepsilon_k = \Theta( (n^{-1/25} + n^{-3/20}k^{1/4}) n^{-1/5} n^{-2/5} ) = \Theta( n^{-16/25} + n^{-3/4}k^{1/4}) = o(k/n) \end{align} as $k \geq n^{4/10}$. For $k > n^{0.99}$, we have in \cref{expkmedium} that $\varepsilon_k = O(n^{-3/10})$ by \cref{eq:ExpBkUpper}, and so $\varepsilon_k = o(k/n)$, while in \cref{expkbig}, $\varepsilon_k = \Theta(n^{-1/6}) =o(k/n)$. \end{proof} \subsection{Path weights} To show inequality \cref{path+} it suffices to show that \begin{align}\label{eq:expWkk0} W\os{\kk} - W\os{\kp} \leq 1.1 \eps_k. \end{align} In \cref{expkbig}, we have defined $\rr0 \coloneqq 1$, so \cref{eq:expWkk0} is trivial. For \cref{expkmedium}, $\Delta \coloneqq W\os{\kk}-W\os{\kp}$ has the same distribution as $\sum_{i=k+2}^{k+\rr0} X(n-i)$, where $X(a) \sim \Exp(a)$ and these variables are all independent. Thus $\Delta$ is stochastically dominated by the sum of $\rr0-1$ independent random variables $X(\bar{k}-\rr0)$. Since $\rr0 = \bar{k} \varepsilon_k$, we have that $\E \Delta \leq \rr0/(\bar{k}-\rr0) = \eps_k/(1-\eps_k)$, and from \cref{exptail} it follows that $\Pr(\Delta > 1.1 \eps_k) = O(\exp(-\Theta(\rr0)))$. From \cref{eq:Expr0}, by the union bound, there is a negligible chance that \cref{path+} fails in any round. \subsection{Budgets in \cref{expkmedium}} \label{ExpC1budgets} As before, we need to define a $B_k$ satisfying \cref{budget} and, as before, $\eps_k$ can be guessed from \cref{epsk}, then checked to satisfy yield robustness as in \cref{kmedrobust,kbigrobust}. The base case, confirming \cref{Bkbase}, is given by $k=n^{4/10}$, where by \cref{eq:ExpBkLower} \begin{align} B_k \geq 2 n^{-1/5} \eqdef U_k . \end{align} To verify \cref{Bksuff}, it is straightforward to check that $\frac{\partial^2}{\partial k^2} {B_k}$ is positive, so $\frac{\partial}{\partial k} {B_k}$ is increasing, and \[ B_{k+1}-{B_k} \geq \frac{\partial}{\partial k} {B_k} = \frac 54 \frac 35 \frac {C_B}{C_\eps} = \frac 34 \frac {C_B}{C_\eps} \geq 8\eps_k , \] since $\frac 34 \frac {C_B}{C_\eps} \geq 8$ (check). Finally, we establish \cref{epsfit}. We show that w.h.p\xperiod} % {with high probability for all $k$ in the range, \begin{align}\label{expepsfitpf} \Delta &\coloneqq \Wo {k+1}^s-W\os k^s \leq 0.1 \eps_k . \end{align} Note that $\Delta \sim \Exp(\bar{k}-1)$, so \begin{align*} \Pr \parens{\Delta > 0.1\varepsilon_k} = \exp \parens{-0.1 \eps_k \cdot (\bar{k}-1)} = \exp(-\Omega(\rr0)) = \exp(-n^{\Omega(1)}) \end{align*} by \cref{eq:Expr0}. Then, by the union bound there is a negligible chance that \cref{expepsfitpf} fails for any $k$. \subsection{Robustness in \cref{expkmedium}} With reference to \cref{robustness}, we complete the robustness argument for \cref{expkmedium}, showing that \cref{robustblurb} holds with high probability. Here we have taken $\rr0 = \eps_k \bar{k}$, so the number of edges from a middle vertex to $V'_S$ (see \cref{Zvs}) is $Z^s_v \sim \Bi(\eps_k \bar{k}, \eps_k)$, with mean \begin{align}\label{Expla} \lambda &= \rr0 \eps_k = \eps_k^2 \bar{k} \end{align} (see \cref{kmedlambda}). Recall that if $\lambda$ is small we expect (see \cref{kmedtotalweight}) that to destroy all paths the adversary will have to delete edges of total weight at least $\eps_k \, n \, \lambda^2 = \eps_k^5 n \bar{k}^2$, which will exceed $B_k$. And, if $\lambda$ is large, then each $Z_v$ will have expectation close to $\lambda=\eps_k^2 \bar{k}$, for a total cost $\eps_k n$ times larger, namely $\eps_k^3 n \bar{k}$, and again this exceeds $B_k$. We now show the details of these rough calculations, including the probabilistic details, applying \cref{lemma:BinMin} to $Z_v$ in the two cases of $\lambda$ small and large. For the adversary to delete all $s$--$t$\xspace paths via $v$, he must delete at least \[ Z_v\coloneqq \min(Z_v^s, Z_v^t) \] edges, and to destroy all paths he must delete at least \[ N\coloneqq \sum_{v \in M'} Z_v \] edges. As described in \cref{robustness}, we imagine a fixed deletion of $k$ edges on each of $s$ and $t$, giving neighbour sets $V'_s$ and $V'_t$ and a set $M'$ of middle vertices, eventually taking a union bound over all such choices. \medskip \noindent \tmbf{If $\lambda \geq 2$}, then by \cref{lemma:BinMin}, for each $v \in M'$, $\Pr(Z^s_v \geq 0.65 \lambda) \geq 1/4$. Thus, $N$ stochastically dominates $0.65\lambda \cdot \Bi(0.99n, 1/4)$, with expectation $> 0.1608 \lambda n$. We shall consider it a \emph{failure} if $N \leq 0.16 \lambda n$. Assuming success, since each edge costs at least $\eps_k$ to delete, it costs at least $0.16 \eps_k \lambda n = 0.16 \eps_k^3 n \bar{k}$ to delete them all. This exceeds $B_k$: \begin{align*} \frac{0.16 \, \eps_k^3 n \bar{k}}{B_k} &= 0.16 \, C_\eps^3 n^{-3/5} \bar{k}^{-6/5} B_k^{-2/5} n \bar{k} \eqnote{by definition of $\varepsilon_k$} \\&= 0.16 \, C_\eps^3 B_k^{-2/5} n^{2/5} \bar{k}^{-1/5} \\& \geq 0.15 \, C_\eps^3 C_B^{-1/2} n^{1/5} n^{-1/5} \eqnote{by~\cref{eq:ExpBkUpper}} \\ &> 1, \end{align*} using that $0.15 \cdot C_\eps^3 C_B^{-1/2}>1$ (check). Failure means that $N/(0.65 \lambda) \sim \Bi(0.99n, 1/4) \leq (0.16\lambda n)/(0.65 n)=(0.16/0.65)n$. Noting that $0.99 \cdot 1/4 > 0.16/0.65$, by \cref{lemma:BinDev}, the probability of failure is $\exp(-\Omega(n))$. By the union bound, the total of the failure probabilities, over all rounds and all adversary choices of the $k$ root edges at $s$ and $t$, is small: \begin{align} \label{expcase1failure} \sum_k {{\binom{k+\rr0}{\rr0}}^2} & \cdot \exp(-\Omega(n)) \\& \leq \sum_k \, {(n^{\rr0})}^2 \exp(-\Omega(n)) \notag \\ &= \sum_k \exp\parens{2\eps_k \bar{k} \ln n-\Omega(n)} \quad\text{(by $\rr0=\eps_k \bar{k}$)} \notag \\&\leq n \exp(-\Omega(n)) = o(1), \notag \end{align} the penultimate inequality using $\eps_k \bar{k} = O(n^{7/10})$ by~\cref{eq:ExpEkUpper}. \medskip \noindent \tmbf{If $\lambda < 2$}, then by \cref{lemma:BinMin} $N$ stochastically dominates $\Bi(0.99n, 0.18\lambda^2)$, with expectation $> 0.175 \lambda^2 n$. We shall consider it a \emph{failure} if $N \leq 0.17 \lambda^2 n = 0.17 \eps_k^4 n \bar{k}^2$. Each edge costs at least $\eps_k$ to delete. Assuming success, it thus costs at least $0.17 \eps_k^5 n \bar{k}^2$ to delete them all, which exceeds $B_k$: \begin{align*} \frac{0.17 \eps_k^5 n \bar{k}^2}{B_k} &= 0.17 C_\eps^5 \eqnote{by definition of $\varepsilon_k$} \\ &> 1, \end{align*} using that $0.17 C_\eps^5>1$ (check). By \cref{lemma:BinDev}, the probability of {failure} is \begin{equation}\label{eq:expN2} \Prob \parens{ N \leq 0.17 \eps_k^4 n \bar{k}^2} = \exp(-\Omega(\eps_k^4 n \bar{k}^2)). \end{equation} Over all rounds and adversary choices of edges incident to $s$ and $t$, the total failure probability is at most \begin{align} \sum_k {\binom{k+\rr0}{\rr0}}^2 & \cdot \Prob \parens{N < 0.17 \eps_k^4 n \bar{k}^2} \notag \\ &\leq \sum_k \exp\parens{2\eps_k \bar{k} \ln n- \exp(-\Omega(\eps_k^4 n \bar{k}^2))} \notag \\&\leq n \exp(-\Omega(\eps_k^4 n \bar{k}^2)) , \notag \intertext{% because $\eps_k^4 n \bar{k}^2$ is larger than $\eps_k \bar{k}$ by a factor $\eps_k^3 n \bar{k}$, which by~\cref{eq:ExpEkLower} is $\Omega(n^{-18/25} \bar{k}^{-6/5} n \bar{k})=\Omega(n^{7/25} \bar{k}^{-1/5}) = \Omega(n^{2/25})$. Continuing, this is} \notag \\ &\leq n \exp(-\Omega(n^{1/25} \, \bar{k}^{2/5})) \eqnote{invoking~\cref{eq:ExpEkLower} again} \label{expcase2failure} \\& = o(1) . \notag \end{align} \subsection{Budgets in \cref{expkbig}} \label{expbudgetsbig} We now establish \cref{budget} for the parameters of \cref{expkbig}. \cref{ExpC1budgets} showed that \cref{budget} holds for $k$ up to ${k^\star} \coloneqq \floor{n-\sqrt{n}}$, the point where \cref{expkmedium} ends and just before \cref{expkbig} begins, so in particular $B_{k^\star} \geq U_{k^\star} - I_{k^\star}$. For the regime of \cref{expkbig}, we redefine $I_k$ from \cref{Ik}. Recall that $I_k$ is a lower bound on the edges incident to $s$ and $t$ used by the first $k$ paths. Previously, the sum defining $I_k$ in \cref{Ik} went to $k-1$ to avoid double counting the \ensuremath{\set{s,t}}\xspace edge. In this regime, however, we need the sum to go $k$, as the $W\os i$ increase rapidly. The weight of the \ensuremath{\set{s,t}}\xspace edge is distributed as $\Exp(1)$, thus w.h.p\xperiod} % {with high probability it costs at most $n^{0.01}$. For $k>{k^\star}$, define \begin{align} \label{Ikdefnew} I_k & \coloneqq \sum_{i=1}^k \parens{W\os k^s + W\os k^t} - n^{0.01}, \end{align} so that w.h.p\xperiod} % {with high probability $I_k$ is a lower bound on the incident edges: the $n^{0.01}$ term resolves the potential double-counting of \ensuremath{\set{s,t}}\xspace. We are now ready to check that \cref{budget} holds. Following the derivation of \cref{budgetsClaim2}, for $k$ from ${k^\star}+1$ to $n-2$, \begin{align} U_k - I_k &= (U_{k^\star} - I_{k^\star}) + [(U_k-U_{k^\star}) - (I_k - I_{k^\star})] \notag \\ &\leq B_{k^\star} + \sum_{k={k^\star}+1}^{n-2} {7 \eps_k} - (\Wo{{k^\star}}^s+\Wo{{k^\star}}^t-n^{0.01}) \eqnote{see \cref{Ukdef}, \cref{Ik}, and \cref{Ikdefnew}} \notag \\&\leq 114 \sqrt n + \sqrt n \cdot 7 C'_\varepsilon n^{-1/6} + n^{0.01} \eqnote{see \cref{eq:ExpBkUpper} and \cref{ExpBk2}} \notag \\& \leq 115 \sqrt n \notag \\& \leq B_k \eqnote{see \cref{ExpBk2}}, \label{ExpBudget} \end{align} using that $C_B' \geq 115$ (check). \subsection{Robustness in \cref{expkbig}} \label{exprobustbig} Again, our aim is to establish robustness of $R$ by showing that \cref{robustblurb} holds with high probability, and the argument is similar to but simpler than that for robustness in \cref{expkmedium}. Since $\rr0=1$, both $V'_s$ and $V'_t$ have size 1. For a vertex $v \in M'$, let $Z_v$ be the number of paths from $V'_s$ to $V'_t$ via $v$. There is only one such possible path, hence \[ Z_v \sim \Bern \parens{\eps_k^2}. \] To destroy all $s$--$t$\xspace paths the adversary must delete at least \[ N\coloneqq \sum_{v \in M'} Z_v \] edges. $N$ stochastically dominates $\Bi(0.99n, \eps_k^2)$, with expectation at least $0.99 \eps_k^2$. We declare the event $N \leq 0.98 \eps_k^2 n$ a \emph{failure}. Assuming success, destroying all $s$--$t$\xspace paths would cost at least $\eps_k N \geq 0.98 \eps_k^3 n$. This exceeds $B_k$, since by \cref{ExpBk2} and $0.98 {C'_\varepsilon}^3 > C'_B$ (check), \begin{equation*} \frac{0.98 \eps_k^3 n}{B_k} = \frac{0.98 {C'_\varepsilon}^3}{C'_B} >1 . \end{equation*} The probability of failure is \begin{equation}\label{eq:expN3} \Prob \parens{ N \leq 0.98 \eps_k^2 n} = \exp(-\Omega(\eps_k^2 n)) = \exp(-\Omega(n^{2/3})). \end{equation} Over all rounds and adversary choices, using that $\binom{k+\rr0}{\rr0} = \binom{k+1}1 \leq n$, the total failure probability is at most \begin{align} \sum_k {\binom{k+\rr0}{\rr0}}^2 & \cdot \Prob(N \leq 0.98 \varepsilon^2 n) \notag \\ &\leq \sqrt n \, n^2 \, \exp (-\Omega(n^{2/3})) \eqnote{by~\cref{eq:expN3}} \label{eq:expfailure3} \\& = o(1) . \notag \end{align} \subsection{Lower bound} \label{ExpLB} As argued in the introduction of this section, for any $k=o(n)$, the lower bound follows from the uniform case. Thus it is sufficient if we show the lower bound for $k \geq n^{9/10}$, which we do now. \begin{remark}\label{remdelta} With high probability, for every pair of vertices $u$ and $v$ in $G' = G-s-t$, there is a $u$--$v$ path in $G'$ of cost at most $\delta = 20n^{-1/6}$ that is edge-disjoint from $P_1, \ldots, P_{n-1}$. \end{remark} \begin{proof} The proof of \cref{expkbig} showed that w.h.p\xperiod} % {with high probability, for all $k$ in the claim's range (up to $k=n-2$), there is a cheap $s$--$t$\xspace path (of cost given by \cref{ExpXkUB}) disjoint from $P_1,\ldots,P_k$, \emph{because} for a given pair of neighbours $u,v$ of $s$ and $t$, there is a $u$--$v$ path in $G'$ that is edge-disjoint from these $k$ paths and has cost at most $4 \eps_k = 20n^{-1/6} \eqdef \delta$ (see \cref{ExpBk2}). The existence of a $k+1$st $s$--$t$\xspace path limits $k$ to $n-2$ since after that there are no new neighbours $u$ and $v$ of $s$ and $t$, but the rest of the argument extends to $k={n-1}$. In particular, extending the definition \cref{ExpBk2} of $B_k$ and $\eps_k$ to $k={n-1}$, the derivation of \cref{ExpBudget} extends without change and shows that the budget $B_{n-1}$ covers the middle edges of all paths $P_1,\ldots,P_{n-1}$, and the robustness argument also extends and shows \cref{eq:expN3} to hold for $k={n-1}$. Since the failure probability in \cref{eq:expN3} is exponentially small, and there are fewer than $n^2$ pairs $\set{u,v}$ in $G'$, w.h.p\xperiod} % {with high probability there is a cheap path (of cost $\leq \delta$) for every pair. \end{proof} For the remainder of this section we assume that the high-probability conclusion of \cref{remdelta} holds. Let $H_k^s$ be the weight of the heaviest edge incident to $s$ used by the first $k$ paths, and let $L_k^s$ be the weight of the lightest edge incident to $s$ \emph{not} used by the first $k$ paths. Define $H_k^t$ and $L_k^t$ likewise. We claim that for all $k$ from 1 to ${n-1}$, with $\delta=20n^{-1/6}$ as in \cref{remdelta}, \begin{align}\label{HLdelta} H_k^s - L_k^s \leq \delta . \end{align} We argue by contradiction. Given $k$, let $P_i$, $i\leq k$, be the path using the edge of weight $H^s_k$. By \cref{remdelta}, we can construct an $s$--$t$\xspace path $Q$ whose $s$-incident edge is the one of weight $L^s_k$, whose $t$-incident edge is the same as that of $P_i$, and whose middle edges cost at most $\delta$ and are not used in $P_1,\ldots,P_{n-1}$. This path $Q$ is cheaper than $P_i$: its $s$-incident edge is cheaper by $H^s_k-L^s_k > \delta$, its $t$-incident edge has the same cost, and its middle edges (costing at most $\delta$) cost at most $\delta$ more than those of $P_i$. Also, $Q$ is edge-disjoint from the first $i-1$ paths: its $s$-incident edge $L^s_k$ is not used even by the first $k$ paths, the middle edges are disjoint from those of all $n-1$ paths, and its $t$-incident edge is that used by $P_i$ (so not used by a previous path). Thus, $Q$ should have been chosen in preference to $P_i$, a contradiction, establishing \cref{HLdelta}. Trivially, $H_k^s \geq W\os k^s$. Thus, from \cref{HLdelta}, \begin{align}\label{LW} L^s_k \geq H_k^s - \delta \geq W\os k^s - \delta . \end{align} For $k \leq n-2$, the edge of $P_{k+1}$ incident to $s$ costs at least $L_k^s$ and the edge incident to $t$ at least $L_k^t$. If $P_{k+1}$ is not the single-edge path $\set{s,t}$ these two edges are distinct, so that $X_{k+1} \geq L^s_k + L^t_k$. If $P_{k+1}$ is the single-edge path $\set{s,t}$ then $P_k$ is not, and $X_{k+1} \geq X_k \geq L^s_{k-1} + L^t_{k-1}$. Either way, by \cref{LW}, \begin{align} X_{k+1} &\geq L^s_{k-1} + L^t_{k-1} \notag \\ &\geq \Wo{k-1}^s + \Wo{k-1}^t - 2\delta. \label{bigkLB1} \end{align} Recall that we are concerned here with $k \geq n^{9/10}$. By \cref{lemma:edge-orderstat}, for all such $k$, and for any $\gamma>0$, w.h.p\xperiod} % {with high probability $W\os k \geq (1-\gamma) \EWW k$. Since the exponential random variable is stochastically greater than the uniform, $\EWW k > k/n = \Omega(n^{-1/10})$, while $\delta = 20n^{-1/6} = o(\EWW k)$. From \cref{muX} it is clear that $\EWW {k-1} \asymp \EWW {k+1}$ (for any $k=\omega(1)$), and we subsume the asymptotic error into the constant $\gamma$. Thus, from \cref{bigkLB1}, for any $\gamma>0$, w.h.p\xperiod} % {with high probability, for all $k \geq n^{9/10}$, \begin{align*} X_k \geq (1-\gamma) 2 \EWW k , \end{align*} completing the proof of the lower bound in \cref{Texp}. \section{Expectation}\label{sec:expectation} In this section we prove \cref{thm:expectation}. We treat the uniform and exponential models at the same time. Let $\evp k$ be the event that $P_k$ exists. Clearly $\Pr(\evp k) \geq \Pr(\evp {n-1})$. By \cref{Tmain} (for the uniformly random model) and \cref{Texp} (for the exponential model), $\Pr(\evp {n-1}) = 1-o(1)$. This establishes the first part of the theorem. Then, let $\mu_k = 2\E W\os k + \ln n /n$ (so for the uniform model, $\mu_k=w_0(k)$). It suffices to show that \begin{align} \E[X_k \mid \evp k] &= (1+o(1)) \mu_k \label{Emu} \end{align} uniformly in $k$. First, we show the lower bound implicit in \cref{Emu}. Fix $\varepsilon > 0$. Let $\evl k$ be the event that (jointly) $P_k$ exists and $X_k \geq (1-\varepsilon) \mu_k$. By \cref{Tmain} (for the uniform model) and \cref{Texp} (for the exponential model), $\evl k$ holds with probability $1-o(1)$ uniformly in $k$. Thus, \begin{align*} E[X_k \mid \evp k] &\geq \Pr(\evl k) \E[X_k \mid \evp k \wedge \evl k ] \geq (1-o(1)) \, (1-\varepsilon) \mu_k. \end{align*} Since this holds for any $\varepsilon$, we have that \[ \E[X_k \mid \evp k] \geq (1-o(1)) \mu_k. \] We now establish the corresponding upper bound. \subsection{Small \tp{$k$}{k}} \label{expSmallk} First, we consider the range $k \leq n^{4/10}$. We will need the following lemma in \cref{EXkU}. \begin{lemma}\label{lem:large-eps} There exists an absolute constant $C>0$ such that, for all $\varepsilon>C$, in both the exponential and uniform models, for all $k=o(\sqrt n)$ the probability of the event \begin{align}\label{eq:indC} X_k > (1+\varepsilon) \mu_k \end{align} is $\OO{n^{-1.9}}$. \end{lemma} \begin{proof} By the reasoning given in the introduction of \cref{ExpBounds}, it is sufficient to show the result in the uniform case, where $\mu_k=\frac{2\k + \ln n}{n}$. We use the same argument as developed in \cref{sec:k-small}, where we prove \cref{Tmain} up to $k=o(\sqrt n)$. Our argument in \cref{sec:k-small} (see \cref{indHyp}) was that for any sufficiently small $\varepsilon>0$, \begin{align}\label{eq:ind-large} \text{if } X_i \leq (1+\eps) \parens { \frac{2i}n+\frac{\ln n}n } \text{for all $i \leq k$, then w.h.p\xperiod} % {with high probability the same holds for $i=k+1$.} \end{align} We proved this by constructing a structure $R=R^{(k)}$ in $G$, in which after deleting $k$ paths, each of cost $\leq (1+\eps) (2k/n+\ln n/n)$ from $G$, w.h.p\xperiod} % {with high probability there remains a path in $R$ satisfying the same cost bound. By \cref{ksucceeds}, the probability of failure was $\OO{n^{-1.9}}+\exp(-\Theta(s(k)))$. This does not suffice since for $k$ small the second term may exceed $\OO{n^{-1.9}}$ (recall $s = 2k + \ln n$). To prove the lemma, we will show that for some sufficiently \emph{large} constant $\varepsilon$, the failure probability in \cref{eq:ind-large} is $\OO{n^{-1.9}}$. As noted in \cref{warning}, a few parts of the argument developed in \cref{sec:k-small} rely on $\varepsilon$ being sufficiently small, and here we will detail the changes needed. Principally, we will make one modification (a simplification) to \cref{sec:k-small}'s construction of $R$. We will also track the dependence of key Landau-notation expressions on $\varepsilon$. \medskip Recall from \cref{sdef,w0def} that $s=2k + \ln n$ and $w_0 = s/n$. Parallelling the structure of \cref{sec:k-small}, we start by reviewing the adversary's edge-count budget. This was given by \cref{Bany} which, through its dependence on \cref{pathlength}, held only for sufficiently {small} $\varepsilon$. For sufficiently large $\varepsilon$, modulo the one-time failure probability $\OO{n^{-1.9}}$ from \cref{LLenBd}, each of the first $k$ paths has length $\leq (1+\varepsilon) w_0 \cdot 19 n < 20 s \varepsilon$, and the total length of the first $k$ paths is at most \begin{align}\label{eq:budget-large} 20 k s \varepsilon < 10 s^2 \varepsilon , \end{align} so we now take this to be the adversary's budget. We build level-0 edges of $R$ exactly as in \cref{level0}, and using the same parameter $\rr0$. That is, we add the cheapest $k+\rr0$ edges incident on $s$, with $\rr0= \ceil{\tfrac1{10} \varepsilon s}$ as in \cref{k0}; the opposite endpoints of these edges are the level-1 vertices. Recall that we declared this step a failure if the number $X$ of edges with weights in the interval $[0, \tfrac k n+\frac19 \varepsilon w_0]$ is smaller than $k+\rr0$. Note that $X \sim \Bi(n', \tfrac k n+\frac19 \varepsilon w_0)$, thus $\E X = (1-o(1)) \, (k+\frac19 \varepsilon s)$, and failure means that $X <k+\rr0$, i.e., that \begin{align*} \frac{X}{\E X} = (1+o(1)) \, \frac{k+\frac1{10} \varepsilon s}{k+\frac19 \varepsilon s} \leq \frac{10}{11} \end{align*} for $\varepsilon$ sufficiently large. Then, analogously to \cref{level0fail}, the failure probability by \cref{lemma:BinDev} is at most \begin{align} \Pr(X < \tfrac{10}{11} \E X) &\leq \exp(-\Omega(\E X)) \leq \exp(-\Omega(\varepsilon s)). \label{eq:large-level0fail} \end{align} We skip constructing level-1 edges as in \cref{level1}, instead setting the level-2 vertices identical to level-1 vertices. (There are no edges between these levels; we have ``level 2'' only to keep the level numbering the same as before.) We build level-2 edges exactly as before, with the same parameter $\rr2$, linking to each level-2 vertex its cheapest $\rr2= \frac{1}{10} \varepsilon s$ neighbours (which become the level-3 vertices). The calculations in \cref{level2} hold for any $\varepsilon>0$, and from \cref{level2fail} the probability of any failure on this level is \begin{align}\label{eq:large-level1fail} \leq \exp{-\Theta(\varepsilon s)}. \end{align} The adversary's deletions of edges incident on $s$ must leave $r_0$ vertices at level 1 (a.k.a.\ level 2), thus $\rr0 \rr2 = \varepsilon^2 s^2/100$ edges leading to level~3. By \cref{eq:budget-large} the adversary is allowed to delete at most $10 s^2 \varepsilon$ edges, so for $\varepsilon$ sufficiently large, at least $2 s^2$ level-3 vertices remain; this is the same as before, and will continue to suffice. From level~3 we construct shortest-path trees just as in \cref{level3}, whose calculations hold for any $\varepsilon > 0$. To recapitulate, these trees are built to a size \cref{ddef} independent of $\varepsilon$, the calculations made are valid for all $\varepsilon$, and the result (here as in \cref{sec:k-small}) is that each tree fails with some probability $o(1)$, but the level as a whole fails only if at least $0.01 s^2$ trees fail, which occurs with probability only $\exp(-\Omega(s^2))$ (see \cref{level3fail}). This concludes the modified construction of $R$. The remainder of the argument is unchanged from \cref{sec:k-small}. In the absence of failures, the maximum weight of any $s$--$t$\xspace path in $R$ remains at most $(1+\varepsilon) w_0$ per \cref{Rpathcost} (indeed, a little less as we've skipped the level-1 edges). The number of successful level-3 trees is $\Omega(s^2)$ as before, and the calculations leading to the probability that an adversary can destroy all cheap paths in $R$ are unaffected: this probability remains $\exp(-\Omega(s^2 \ln n))$ as in \cref{smallkAdversaryFailure}, which is dominated by other failure probabilities. Tallying up, as in \cref{sec:small-success}, we have a one-time failure probability of $\OO{n^{-1.9}}$ from \cref{LLenBd}. Out of levels 0, 2 and 3 we have failure probabilities given respectively by \cref{eq:large-level0fail}, \cref{eq:large-level1fail} and \cref{level3fail}, namely $\exp(-\Omega( \varepsilon s))$, $\exp(-\Omega( \varepsilon s))$ and $\exp(-\Omega( s^2))$. Since $s > \ln n$, for some $\varepsilon$ sufficiently large, the net failure probability is $\OO{n^{-1.9}}$, as claimed. \end{proof} Let $C$ be the constant in \cref{lem:large-eps}. Separately, fix any sufficiently small $\varepsilon>0$. Let \begin{align*} U_1 &= [0, (1+\varepsilon){\mu_k}), \\ U_2 &= [(1+\varepsilon){\mu_k}, C{\mu_k}), \\ U_3 &= [C{\mu_k}, \infty). \end{align*} Let $\mathcal{A}_i$ be the event that $X_k \in U_i$. By \cref{Tmain}, $\Pr(\mathcal{A}_1) = 1-o(1)$ and $\Pr(\mathcal{A}_2)=o(1)$, and by \cref{lem:large-eps}, $\Pr(\mathcal{A}_3)=O(n^{-1.9})$. Since here we are considering $k \leq n^{4/10} \leq n/2$, with reference to the proof of \cref{rmk:existence}, one possible choice for $P_k$ is some path of length 2 (there must remain at least one such), and thus, deterministically, \begin{align}\label{eq:length2} X_k \leq W_s + W_t , \end{align} where $W_v$ denotes most expensive edge out of $v$ ($W_v = W^v_{\os{n-1}}$ in the notation of \cref{Wkv}). In the uniform model, \cref{eq:length2} means that, deterministically, $X_k \leq 2$. Then, \begin{align} \E[X_k] &=\Pr(\mathcal{A}_1) \E[X_k \mid \mathcal{A}_1] + \Pr(\mathcal{A}_2) \E[X_k \mid \mathcal{A}_2] + \Pr(\mathcal{A}_3) \E[X_k \mid \mathcal{A}_3] \notag \\ &\leq (1-o(1)) \cdot (1+\varepsilon) {\mu_k} + o(1) \cdot (1+C){\mu_k} + O(n^{-1.9}) \cdot 2 \notag \\ &\leq (1+\varepsilon+o(1)){\mu_k} , \label{EXkU} \end{align} since $\mu_k > \ln n/n$. As this holds for arbitrarily small $\varepsilon > 0$, \begin{align} \E[X_k] \leq (1+o(1)) {\mu_k} . \label{eq:expectLB} \end{align} For the exponential model the same argument applies, once we control $\E[X_k \mid \mathcal{A}_3]$. We make use of the following inequality. Let $Z$ be a random variable with CDF $F$, and $\mathcal{A}$ be an event with $\Pr(\mathcal{A}) = \alpha$. Then, \begin{align}\label{eq:F} \E[Z \mid \mathcal{A}] \leq \E[Z \mid Z > F^{-1} (1-\alpha) ]. \end{align} In the case that $Z$ is an exponential random variable with rate $\lambda$, $F(z)=1-\exp(-\lambda z)$, so $F^{-1} (1-\alpha) = -\ln(\alpha)/\lambda$. By the memoryless property of the exponential, the RHS of \cref{eq:F} is $\E[Z]+F^{-1} (1-\alpha)$, giving \begin{align} \E[Z \mid \mathcal{A}] & \leq \frac{1-\ln(\alpha)}{\lambda} . \label{ExpCondExp} \end{align} Recall from \cref{ordersumexp} that $W_v = \sum_{i=1}^{n-1} Z_i$ where $Z_i \sim \Exp(i)$. Condition on the event $\mathcal{A}_3$, taking $\alpha \coloneqq \Pr(\mathcal{A}_3) = O(n^{-1.9})$. By \cref{ExpCondExp}, \begin{align} \E[W_k \mid \mathcal{A}_3] =\sum_{i=1}^{n-1} \E [Z_i \mid \mathcal{A}_3] \leq \sum_{i=1}^{n-1} \frac{1-\ln(\alpha)}{i} \sim (1-\ln(\alpha)) \ln n = O(\ln^2 n). \label{eq:Walpha} \end{align} By \cref{eq:length2}, \cref{eq:Walpha} and linearity of expectation, \begin{align} \Pr(\mathcal{A}_3) \E[X_k \mid \mathcal{A}_3] \leq \alpha \E[W_s + W_t \mid \mathcal{A}_3] = 2 \alpha O(\ln^2 n) = O(n^{-1.9} \, \ln^2 n) , \label{Exp2Path} \end{align} which is $o(\mu_k)$ since $\mu_k > \ln n/n$. Thus \cref{EXkU} holds also for the exponential model (the change to the middle line of the calculation affects nothing), whereupon so does \cref{eq:expectLB}. \subsection{Large \tp{$k$}{k}} For $k \geq n^{4/10}$, we gather the failure events in \cref{largekUB}. First, we have $X_{n^{4/10}} \leq 3 n^{4/10} / n$ with failure probability $O(n^{-1.9})$, from \cref{Xn0.4} and \cref{Pn0.4}. Then, we have to check two types of failures: failure of \cref{path+} to be an upper bound on \cref{path1} (because the edge order statistics are not as expected), and violation of \cref{XkWok} (because $R$ fails to be robust against the adversary). Failure of \cref{path+} as an upper bound is, in the uniform model, checked through violation of \cref{eq:Wkk0}, the paragraph after \cref{eq:Wkk0} showing failure to occur w.p.\ at most $\exp(-\Omega(n^{0.01}))$. Likewise, in the exponential model it is checked in and following \cref{eq:expWkk0}, with a failure probability of $O(\exp(-\Omega(n^{3/50})))$. The failure probability of \cref{XkWok} in the uniform model is calculated for three cases: near \cref{case1failure} as $n \exp(-\Omega(n))$, near \cref{case2failure} as $n \exp(-\Omega(n^{11/25}))$, and near \cref{eq:unifailure3} as $14 n^{5/2} \exp(-\Omega(n^{2/3}))$. The failure probability in the exponential model is also calculated for three cases: near \cref{expcase1failure} as $n \exp(-\Omega(n))$, near \cref{expcase2failure} as $n \exp(-\Omega(n^{1/25}))$, and near \cref{eq:expfailure3} $n^{5/2} \exp(-\Omega(n^{2/3}))$. Thus, the failure probabilities for \cref{path+} and \cref{XkWok} are all $O(\exp(-n^{0.01}))$, so the probability of any failure affecting any $k>n^{4/10}$ is $O(n^{-1.9})$. Let \begin{align*} U_1 &= [0, (1+\varepsilon){\mu_k}) \\ U_2 &= [(1+\varepsilon){\mu_k}, \infty), \end{align*} and let $\mathcal{A}_i$ be the event that $P_k$ exists and $X_k \in U_i$. Thus $\Pr(\mathcal{A}_1)=1-o(1)$ and $\Pr(\mathcal{A}_2) = O(n^{-1.9})$. Conditioning on the event $\evp k$ that $P_k$ exists, this path clearly has cost \[ X_k \leq Z \coloneqq \sum_{v \in V(G)} W_v \] (analogous to \cref{eq:length2}). In the uniform model, deterministically, $Z \leq n$. In the exponential model, the event $\mathcal{A}_2$ here has the same probability as event $\mathcal{A}_3$ in \cref{expSmallk}, so we may reuse \cref{eq:Walpha}, obtaining \begin{align*} \E[Z \mid \mathcal{A}_2] &= \sum_{v \in V(G)} \E[W_v \mid \mathcal{A}_2] = n \, O(\ln^2 n) = o(n^{1.1}) . \end{align*} Thus, in both the uniform and exponential cases, \begin{align} \E[X_k \mid \evp k] &=\Pr(\mathcal{A}_1) \E[X_k \mid \mathcal{A}_1] + \Pr(\mathcal{A}_2) \E[X_k \mid \mathcal{A}_2] \notag \\ &\leq (1-o(1)) \cdot (1+\varepsilon) {\mu_k} + O(n^{-1.9}) \cdot o(n^{1.1}) \notag \\ &=(1-o(1))(1+\varepsilon){\mu_k} , \label{EXkUlarge} \end{align} since $\mu_k > 2k/n > n^{-6/10} = \omega(n^{-0.8})$. As this holds for arbitrarily small $\varepsilon > 0$, for all $k \geq n^{4/10}$, \begin{align} \E[X_k \mid \evp k] \leq (1+o(1)) \mu_k , \label{eq:expectUB} \end{align} completing the proof. \section*{Acknowledgements} We thank Alan Frieze and Wes Pegden for an initial discussion of the second-shortest path, and Alan for noticing that minimum-cost $k$-flow (\cref{rmk:pathsFk}) was not an open problem but immediately implied by our other results. We also thank two anonymous referees for helpful suggestions. \printbibliography \end{document}

https://arxiv.org/abs/2204.11084

Decompositions of functions defined on finite sets in $\mathbb{R}^d$

A finite subset $M \subset \mathbb{R}^d$ is basic, if for any function $f \colon M \to \mathbb{R}$ there exists a collection of functions $f_1, \ldots, f_d \colon \mathbb{R} \to \mathbb{R}$ such that for each element $(x_1, \ldots, x_d)\in M$ we have $f(x_1, \ldots, x_d) = f_1(x_1) + \ldots + f_d(x_d)$. For certain finite sets, we prove a criterion for a set to be basic, and we show that it cannot be extended to the general case. In addition, we interpret the above criterion in terms of doubly-weighted graphs and give an estimation for the number of elements in certain basic and non-basic subsets.

\section{Introduction} The concept of \textit{basic subsets} arises in connection with Hilbert's thirteenth problem on the superposition of continuous functions. The first time it was introduced in an explicit form by Sternfeld in 1989 \cite{Sternfeld1989}. He called a~subset $M \subset \mathbb{R}^d$ \emph{(continuously) basic}, if for any continuous function $f \colon M \to \mathbb{R}$ there exists a collection of continuous functions $f_1, \ldots, f_d \colon \mathbb{R} \to \mathbb{R}$ such that \begin{equation}\label{eq: basic set condition} f(x_1, \ldots, x_d) = f_1(x_1) + \ldots + f_d(x_d) \end{equation} for each element $(x_1, \ldots, x_d)\in M$. The origin of this concept goes back to 1958~\cite{Arnold1958} when Arnold raised the question equivalent to the following: what are basic subsets in the case $d=2$? Sternfeld showed that a closed bounded subset $M\subset\mathbb{R}^2$ is basic if and only if $M$ does not contain arbitrary long arrays, where an \emph{array} is a (finite or infinite) sequence of points $(x_i,y_i)$ on the plane such that $x_i=x_{i+1}$, $y_i\ne y_{i+1}$ for odd $i$ and $y_i=y_{i+1}$, $x_i\ne x_{i+1}$ for even $i$. In this paper, we focus on finite subsets of $\mathbb{R}^d$, where $d\geqslant2$. In this case, the condition of continuity can be omitted. Without loss of generality, we identify finite subsets of~$\mathbb{R}^d$ with integer points inside $d$-dimensional cube~$[n]^d$, where $[n] = \{1,\ldots,n\}$. We establish the following estimation for the number of elements in basic subsets. \begin{theorem}\label{th: M is basic => |M| < dn - (d-2)} If $M \subset [n]^d$ is a basic subset, then $$ |M| \leqslant dn - (d-1). $$ This boundary cannot be improved. \end{theorem} On the other hand, we estimate the number of elements in non-basic subsets that are minimal in the sense of inclusion (to be precise, we call a~non-basic subset $M\subset[n]^d$ \emph{minimal}, if any proper subset $K\subset M$ is basic). In order to obtain non-trivial results, we suppose that each layer of $[n]^d$ contains an element of a~subset, where by \emph{layers} we mean hyperplanes orthogonal to the coordinate axes, so that each layer can be determined by the equation $x_i=j$, where $i\in[d]$ and $j\in[n]$. \begin{theorem}\label{th: M is non-basic => 2n-1 < |M| < dn - (d-3)} If $M \subset [n]^d$ is a minimal non-basic subset such that every layer of~$[n]^d$ has a non-empty intersection with $M$, then $$ 2n \leqslant |M| \leqslant dn - (d-2). $$ \end{theorem} The condition for a subset $M \subset [n]^d$ to be basic is equivalent to the consistency of the corresponding system of linear equations with integer coefficients. In particular, there exists an algorithm, polynomial in $|M|$, for determining whether the subset $M$ is basic. For $d=2$, however, one can indicate a~simpler algorithm based on the following criterion: a finite subset $M \subset [n]^2$ is basic if and only if $M$ does not contain a \emph{closed array}, that is, a~non-trivial finite array with the coinciding first and last points \cite{Skopenkov2010}. The reader can think about this criterion in the following way. Let us color the points of a closed array in two colors, say, color odd points in red and even points in blue respectively. Then every layer contains the same number of red and blue points. Extending the idea of coloring onto $\mathbb{R}^d$, we get the following theorem. \begin{theorem} \label{th_set_to_graph} Let $M\subset[n]^d$ be a subset containing two or zero elements in every layer. Then $M$ is non-basic if and only if there exists a non-empty subset $K\subset M$ and a coloring of $K$ in two colors such that every layer contains the same number of elements of each color. \end{theorem} It turns out that the same ideas can be used for studying doubly-weighted graphs as well. In particular, we establish the following result which is of independent interest. \begin{theorem} \label{th_basis_graph} A graph $G = (V, E)$ does not contain a bipartite connected component if and only if for any vertex weight function $w_V \colon V \to \mathbb{R}$ there exists an edge weight function $w_E \colon E \to \mathbb{R}$ such that the weight of any vertex is equal to the sum of weights of the edges incident to this vertex: \[ w_V(v) = \sum\limits_{e\colon v\in e} w_E(e). \] \end{theorem} Assigning colors to elements of a finite subset $M\subset[n]^d$ is a particular case of \emph{weight functions} $M\to\mathbb{Z}$. Identifying red and blue colors with values of the set $\{-1,1\}$, one can notice that if a non-basic subset $M$ satisfies Theorem~\ref{th_set_to_graph}, then, for a subset $K\subset M$, the sum of the weight function values taken over a fixed layer is zero. This observation leads us to the following definition. We will call $f\colon M\to\mathbb{Z}$ \emph{annihilation function of~$M$}, if for each layer $L$: \[ \sum\limits_{\EuScript X\in L\cap M}f(\EuScript X) = 0. \] It turns out that basic and minimal non-basic subsets admit the following descriptions in terms of their annihilation functions. \begin{lemma}\label{lemma: M is non-basic <=> annihilation weight function} A subset $M \subset [n]^d$ is non-basic if and only if there exists a~non-trivial annihilation function of $M$. \end{lemma} \begin{lemma}\label{lemma: M is min. non-basic <=> annihilation weight function is unique} If a non-basic subset $M \subset [n]^d$ is minimal, then the annihilation function is unique up to multiplying by a constant. \end{lemma} As we have just seen, when a non-basic subset $M$ satisfies conditions of Theorem~\ref{th_set_to_graph}, one can choose an annihilation function of $M$ whose values are in $\{-1,0,1\}$. In general, however, the values are not bounded. More precisely, consider the annihilation function of a minimal non-basic subset~$M$ whose values are setwise coprime integers (we will call such annihilation functions \emph{irreducible}). Then the following result takes place. \begin{theorem}\label{th_irreducible_annihilation_function_is_unbounded} For any positive integer $m$ there exist a positive integer $n$ and a minimal non-basic subset $M\subset[n]^3$ whose irreducible annihilation function~$f$ has the value $f(\EuScript X)=m$ for some $\EuScript X \in M$. \end{theorem} Since the coefficients of irreducible annihilation functions are unbounded, one cannot expect to have a criterion similar to Theorem~\ref{th_set_to_graph} in the general case. Nevertheless, one can obtain a simplification by considering an annihilation function $f$ of a subset $M\subset[n]^d$ as a function $f\colon [n]^d \to \mathbb{Z}$ such that $f(\EuScript X)=0$ as long as $\EuScript X\notin M$. The simplest non-trivial annihilation functions of $[n]^d$ correspond to the simplest closed arrays, i.e. rectangles (we call such annihilation functions \emph{simple}, see Definition~\ref{def: simple annihilation function}). It turns out that any annihilation function can be generated as a sum of simple annihilation functions. This observation gives us a rather simple way to construct non-trivial non-basic subsets of $[n]^d$ as domains of annihilation functions of $[n]^d$. \begin{theorem} \label{thBoyarov} Every annihilation function of $[n]^d$ can be decomposed into a~finite sum of simple annihilation functions. \end{theorem} The structure of the paper is the following. In Section~\ref{Section:algebra_app}, we translate the concept of finite basic subsets into the language of systems of linear equations and study their properties. Section~\ref{Section:estimations}, Section~\ref{Section:graphs_app} and Section~\ref{Section:structure_of_non-basic_sets} are devoted to the proofs of Theorems~\ref{th: M is basic => |M| < dn - (d-2)} and~\ref{th: M is non-basic => 2n-1 < |M| < dn - (d-3)}, Theorems~\ref{th_set_to_graph} and~\ref{th_basis_graph}, Theorems~\ref{th_irreducible_annihilation_function_is_unbounded} and~\ref{thBoyarov} respectively. We end the paper by Section~\ref{Section:conclusion} discussing possible directions for further research. \section{Finite basic subsets and systems of linear equations} \label{Section:algebra_app} Let $M$ be a subset of $[n]^d$. Then, in algebraic terms, condition~\eqref{eq: basic set condition} for~$M$ to be basic corresponds to the consistency of the system of $|M|$~linear equations in $dn$~variables. Indeed, for every $i \in [d]$, the variable $x_i$ generates $n$~different layers $x_i = 1$, $\ldots$, $x_i = n$. For every $j \in [n]$, j-th layer corresponds to the unique value~$f_i(j)$ that can be interpreted as a~variable~$X_{ij}$. Thus, \eqref{eq: basic set condition} is equivalent to the system of linear equations of the form \begin{equation}\label{eq: basic set condition, algebraic form} f(\EuScript X) = f(x_1,\ldots,x_d) = \sum\limits_{i=1}^d X_{ix_i}, \end{equation} where $\EuScript X = (x_1,\ldots,x_d) \in M$. In particular, if this system is consistent for any function $f\colon M\to\mathbb{R}$, then $M$ is basic, and vice versa. \begin{notation}\label{notation: A_M} We denote $A_M$ the matrix of system~\eqref{eq: basic set condition, algebraic form}. \end{notation} \begin{remark} The set of all functions $f\colon M \to \mathbb{R}$ form a vector space of dimension~$|M|$ with the basis of \emph{indicator functions} $\textbf{1}_{\EuScript X}$, \[ \textbf{1}_{\EuScript X}(\EuScript Y) = \left\{\begin{array}{cl} 1, & \EuScript Y = \EuScript X \\ 0, & \EuScript Y \ne \EuScript X, \end{array}\right. \] where $\EuScript X,\EuScript Y \in M$. Indeed, every function $f$ can be represented as a linear combination of the indicator functions: \[ f = \sum\limits_{\EuScript X \in M}f(\EuScript X)\textbf{1}_{\EuScript X}. \] \end{remark} \begin{proof}[Proof of Lemma~\ref{lemma: M is non-basic <=> annihilation weight function}] A subset $M \subset [n]^d$ is non-basic if and only if the rows of the matrix~$A_M$ of system~\eqref{eq: basic set condition, algebraic form} are linearly dependent. In other words, there exists a non-trivial linear combination of rows which equals zero. Since the rows and columns of~$A_M$ correspond to the elements of $M$ and the layers respectively, coefficients of this linear combination are the values of annihilation function. Note, that the entries of~$A_M$ are zeroes and ones, therefore, it is possible to choose a~linear combination with integer coefficients. \end{proof} \begin{proof}[Proof of Lemma~\ref{lemma: M is min. non-basic <=> annihilation weight function is unique}] The existence of two different annihilation functions corresponds to the existence of two different linear combinations of rows of~$A_M$ that equal zero. The latter implies that we can construct a non-trivial linear combination of rows such that at least one of its coefficients is zero. Hence, there is a proper subset $K\subset M$ which is non-basic. \end{proof} \begin{remark} The converse of Lemma~\ref{lemma: M is min. non-basic <=> annihilation weight function is unique} does not hold. For example, let $$ M=\{(1,1,1),(2,1,1),(1,2,1),(1,1,2),(2,2,1)\}. $$ Then $M$ is not minimal since we can eliminate $(1,1,2)$ and obtain a~closed plane array (see Fig.~\ref{figure: non-minimal non-basic set}). Nevertheless, any annihilation function of $M$ is a~product of $$ f = \textbf{1}_{(1,1,1)} - \textbf{1}_{(2,1,1)} + \textbf{1}_{(2,2,1)} - \textbf{1}_{(1,2,1)} $$ and some constant. We can see that in this example $f(1,1,2) = 0$. In fact, the uniqueness of the annihilation function implies minimality of a non-basic subset $M$, if we additionally require that $f(\EuScript X)\ne0$ for any $\EuScript X\in M$. \end{remark} \begin{center} \begin{tikzpicture}[scale=1.2,x={(1cm,0cm)},y={(0.54cm,0.36cm)},z={(0cm,1cm)},line width=.5pt] \begin{scope} \coordinate (a1) at (0,0,0); \coordinate (a2) at (1,0,0); \coordinate (a3) at (1,1,0); \coordinate (a4) at (0,1,0); \coordinate (b1) at (0,0,1); \coordinate (b2) at (1,0,1); \coordinate (b3) at (1,1,1); \coordinate (b4) at (0,1,1); \draw [line width=.8pt] (b1)--(b4)--(b3)--(b2)--(b1)--(a1)--(a2)--(b2); \draw [line width=.8pt] (a2)--(a3)--(b3); \draw [dashed] (a1)--(a4); \draw [dashed] (b4)--(a4)--(a3); \foreach \EuScript P in {b2,b3,b4}{ \filldraw[white] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \foreach \EuScript P in {a1,a2,a3,a4,b1} \filldraw (\EuScript P) circle (1.7pt); \refstepcounter{ris} \draw (0.5,0.5,-0.6) node {Figure \arabic{ris}.\label{figure: non-minimal non-basic set}}; \end{scope} \end{tikzpicture} \end{center} \begin{lemma}\label{lemma: one point in layer => basic} A subset $M \subset [n]^d$ is basic, if for any non-empty subset $K \subset M$ there is a layer containing only one element from $K$. \end{lemma} \begin{proof} In terms of matrix $A_M$, the conditions of Lemma~\ref{lemma: one point in layer => basic} mean that any collection of rows of $A_M$ generates a submatrix having a column with the unique non-zero entry. Hence, there is no non-trivial linear combination of rows of $A_M$ which equals zero. This implies that $M$ is basic. \end{proof} \section{Estimations for basic and non-basic subsets} \label{Section:estimations} The main goal of this section is to prove Theorems~\ref{th: M is basic => |M| < dn - (d-2)} and~\ref{th: M is non-basic => 2n-1 < |M| < dn - (d-3)}. \begin{lemma}\label{lemma: estimations} Let $n\in\mathbb{N}$ and $Y_{ij}$ be coordinates in $\mathbb{R}^{dn}$, where $i\in[d]$ and $j\in[n]$. If the entries of each of $m$ vectors from $\mathbb{R}^{dn}$ satisfy $(d-1)$ equations \begin{equation}\label{eq: subspace condition} Y_{11} + \ldots + Y_{1n} = \ldots = Y_{d1} + \ldots + Y_{dn} \end{equation} and $m > dn-(d-1)$, then these $m$ vectors are linearly dependent. \end{lemma} \begin{proof} Each vector is contained in the subspace $V \subset\mathbb{R}^{dn}$ determined by condition~\eqref{eq: subspace condition}. Hence, $\dim V = dn - (d-1) < m$, which implies that any $m$~vectors in $V$ are linearly dependent. \end{proof} \begin{proof}[Proof of Theorem~\ref{th: M is basic => |M| < dn - (d-2)}] Let us suppose that $|M| > dn - (d-1)$. Then the rows of the matrix~$A_M$ satisfy conditions of Lemma~\ref{lemma: estimations} for $m=|M|$. Hence, they are linearly dependent, meaning, that $M$ is not basic. To show that it is impossible to improve the inequality, it is sufficient to present an appropriate example. For this purpose, set $M$ to be as follows: $$ M = \{(k,1,\ldots,1), (1,k,\ldots,1), \ldots, (1,1,\ldots,k) \mid k\in[n]\} $$ (Figure~\ref{figure: maximal basic set in 4x4x4} illustrates the set $M$ in the case $d=3$, $n=4$). Then $M$ is basic by Lemma~\ref{lemma: one point in layer => basic} and $|M| = dn - (d-1)$. \end{proof} \begin{center} \begin{tikzpicture}[scale=1.2,x={(1cm,0cm)},y={(0.3cm,0.2cm)},z={(0cm,1cm)},line width=.5pt] \begin{scope} \coordinate (a) at (0,0,0); \coordinate (b) at (1,0,0); \coordinate (c) at (2,0,0); \coordinate (d) at (3,0,0); \coordinate (e) at (0,1,0); \coordinate (f) at (0,2,0); \coordinate (g) at (0,3,0); \coordinate (h) at (0,0,1); \coordinate (i) at (0,0,2); \coordinate (j) at (0,0,3); \foreach \EuScript P in {0,1,2,3}{ \draw [line width=.8pt] (0,\EuScript P,3)--++(3,0,0)--++(0,0,-3); \draw [line width=.8pt] (0,0,\EuScript P)--++(3,0,0)--++(0,3,0); } \foreach \EuScript P in {0,1,2}{ \draw [line width=.8pt] (\EuScript P,0,0)--++(0,0,3)--++(0,3,0); \draw [dashed] (0,\EuScript P+1,3)--++(0,0,-3)--++(3,0,0); \draw [dashed] (\EuScript P,0,0)--++(0,3,0)--++(0,0,3); } \foreach \EuScript P in {1,2}{ \draw [dashed] (0,0,\EuScript P)--++(0,3,0)--++(3,0,0); \foreach \EuScript Q in {1,2}{ \draw [dashed] (0,\EuScript P,\EuScript Q)--++(3,0,0); \draw [dashed] (\EuScript P,0,\EuScript Q)--++(0,3,0); \draw [dashed] (\EuScript P,\EuScript Q,0)--++(0,0,3); } } \foreach \EuScript X in {0,1,2,3}{ \foreach \EuScript Y in {0,1,2,3}{ \foreach \z in {0,1,2,3}{ \filldraw[white] (\EuScript X,\EuScript Y,\z) circle (1.7pt); \draw (\EuScript X,\EuScript Y,\z) circle (1.7pt); } } } \foreach \EuScript P in {a,b,c,d,e,f,g,h,i,j} \filldraw (\EuScript P) circle (1.7pt); \refstepcounter{ris} \draw (1.5,1.5,-0.7) node {Figure \arabic{ris}.\label{figure: maximal basic set in 4x4x4}}; \end{scope} \end{tikzpicture} \end{center} \begin{remark} Slightly modifying Lemma~\ref{lemma: estimations}, we can prove that if a basic subset~$M$ belongs to a~parallelepiped of size $n_1 \times \ldots \times n_d$, then $$ |M| \leqslant (n_1 + \ldots + n_d) - (d-1). $$ \end{remark} \begin{proof}[Proof of Theorem~\ref{th: M is non-basic => 2n-1 < |M| < dn - (d-3)}] To obtain the upper bound, it is sufficient to use Theorem~\ref{th: M is basic => |M| < dn - (d-2)}. To obtain the lower bound, suppose that $|M|<2n$. This supposition immediately implies that there exists a layer $L$ such that $|L\cap M|<2$. Due to the conditions of the statement, $L\cap M\ne\emptyset$, hence, there is a single element $\EuScript X\in L\cap M$. As a consequence, if $M$ is non-basic, so is $M\setminus\{\EuScript X\}$. Therefore, $M$ is not minimal, which contradicts our supposition. \end{proof} \begin{remark} \label{Remark: lower bound is reachable} The lower bound of Theorem~\ref{th: M is non-basic => 2n-1 < |M| < dn - (d-3)} cannot be improved. To ensure, consider $$ M = \{(k,k,\ldots,k), (k+1,k,\ldots,k) \mid k\in[n-1]\} \cup \{n,n,\ldots,n\} \cup \{1,n,\ldots,n\}. $$ (see Fig.~\ref{figure: minimal non-basic set in 4x4x4}). We can see, that $|M|=2n$ and $M$ is non-basic, since it is annihilated by $$ f = \sum\limits_{k=1}^{n-1} \Big(\textbf{1}_{(k,k,\ldots,k)} - \textbf{1}_{(k+1,k,\ldots,k)}\Big) + \Big(\textbf{1}_{(n,n,\ldots,n)} - \textbf{1}_{(1,n,\ldots,n)}\Big). $$ \begin{center} \begin{tikzpicture}[scale=1.2,x={(1cm,0cm)},y={(0.3cm,0.2cm)},z={(0cm,1cm)},line width=.5pt] \begin{scope} \coordinate (a) at (0,0,0); \coordinate (b) at (1,0,0); \coordinate (c) at (1,1,1); \coordinate (d) at (2,1,1); \coordinate (e) at (2,2,2); \coordinate (f) at (3,2,2); \coordinate (g) at (3,3,3); \coordinate (h) at (0,3,3); \foreach \EuScript P in {0,1,2,3}{ \draw [line width=.8pt] (0,\EuScript P,3)--++(3,0,0)--++(0,0,-3); \draw [line width=.8pt] (0,0,\EuScript P)--++(3,0,0)--++(0,3,0); } \foreach \EuScript P in {0,1,2}{ \draw [line width=.8pt] (\EuScript P,0,0)--++(0,0,3)--++(0,3,0); \draw [dashed] (0,\EuScript P+1,3)--++(0,0,-3)--++(3,0,0); \draw [dashed] (\EuScript P,0,0)--++(0,3,0)--++(0,0,3); } \foreach \EuScript P in {1,2}{ \draw [dashed] (0,0,\EuScript P)--++(0,3,0)--++(3,0,0); \foreach \EuScript Q in {1,2}{ \draw [dashed] (0,\EuScript P,\EuScript Q)--++(3,0,0); \draw [dashed] (\EuScript P,0,\EuScript Q)--++(0,3,0); \draw [dashed] (\EuScript P,\EuScript Q,0)--++(0,0,3); } } \foreach \EuScript X in {0,1,2,3}{ \foreach \EuScript Y in {0,1,2,3}{ \foreach \z in {0,1,2,3}{ \filldraw[white] (\EuScript X,\EuScript Y,\z) circle (1.7pt); \draw (\EuScript X,\EuScript Y,\z) circle (1.7pt); } } } \foreach \EuScript P in {a,b,c,d,e,f,g,h} \filldraw (\EuScript P) circle (1.7pt); \refstepcounter{ris} \draw (1.5,1.5,-0.7) node {Figure \arabic{ris}.\label{figure: minimal non-basic set in 4x4x4}}; \end{scope} \end{tikzpicture} \end{center} On the other hand, it is not clear whether it is possible to improve the upper bound or not. For instance, if we add an element to the basic subset $$ M = \{(k,1,\ldots,1), (1,k,\ldots,1), \ldots, (1,1,\ldots,k) \mid k\in[n]\} $$ discussed in the proof of Theorem~\ref{th: M is basic => |M| < dn - (d-2)}, then we get a non-basic set which is not minimal. Say, if we add $\EuScript X=(x_1,\ldots,x_d)$, where $x_k>1$ for all $k\in[d]$, then the set $M\cup\{\EuScript X\}$ is annihilated by the function $$ (d-1)\mathbf{1}_{(1,1,\ldots,1)} + \mathbf{1}_{\EuScript X} - \big(\mathbf{1}_{(x_1,1,\ldots,1)} + \mathbf{1}_{(1,x_2,\ldots,1)} + \ldots + \mathbf{1}_{(1,1,\ldots,x_d)}\big), $$ and hence, $M\cup\{\EuScript X\}$ has a non-basic subset of size $d+2$. \end{remark} \section{Criterion for certain subsets to be basic} \label{Section:graphs_app} The main goal of this section is to prove Theorem~\ref{th_set_to_graph} and Theorem~\ref{th_basis_graph}. The second part of the section assumes that the reader is familiar with the basics of hypergraph theory. For an extensive account of this topic, we refer, for example, to~\cite{Bretto2013}. \begin{proof}[Proof of Theorem~\ref{th_set_to_graph}] Let us assume that there is an appropriate coloring of a~subset $K\subset M$. Then this coloring corresponds to the annihilation function $f\colon K\to\{\pm 1\}$. Extending $f$ to the set $M$ by defining $f(\EuScript X)=0$ for $\EuScript X\in M\setminus K$, we obtain an annihilation function of $M$. Hence, by Lemma~\ref{lemma: M is non-basic <=> annihilation weight function}, $M$~is non-basic. Conversely, let $M$ be non-basic. By Lemma~\ref{lemma: M is non-basic <=> annihilation weight function} this implies that there exists a~non-trivial annihilation function $f$ of $M$. Denote $K\subset M$ to be the domain of $f$ and define a function $g\colon M\to\{-1,0,1\}$ by \[ g(\EuScript X) = \left\{ \begin{array}{rl} f(\EuScript X)/|f(\EuScript X)|, & \mbox{if } \EuScript X\in K\\ 0, & \mbox{if } \EuScript X\in M\setminus K. \end{array} \right. \] Since every layer contains two or zero elements of $M$, the function $g$ is annihilation. Hence, it provides us a desired coloring. \end{proof} A subset $M\subset[n]^d$ can be naturally considered as a~hypergraph whose vertices are elements of~$M$ and hyperedges are subsets of elements of $M$ that lie in the same layer. We denote this hypergraph~$G(M)$. If every layer contains two or zero elements of $M$, then the hypergraph becomes a~graph, possibly with multiple edges. The notion of finite basic subsets can be naturally extended to hypergraphs as follows. \begin{definition} \label{basis_graph} We call a hypergraph $G = (V, E)$ \emph{basic}, if for any vertex weight function $w_V \colon V \to \mathbb{R}$ there exists an edge weight function $w_E \colon E \to \mathbb{R}$ such that the weight of any vertex is equal to the sum of the weights of the edges incident to this vertex. In other words, for any $v \in V$: \begin{equation}\label{eq: basic graph condition} w_V(v) = \sum\limits_{e\colon v \in e} w_E(e). \end{equation} \end{definition} \begin{lemma} \label{lemma: set_to_hypergraph} A subset $M\subset[n]^d$ is basic if and only if the corresponding hypergraph~$G(M)$ is basic. \end{lemma} \begin{proof} It follows directly from the observation that there is a natural bijection between the function $f$ in Eq.~\eqref{eq: basic set condition} and the vertex weight function $w_V$ in Eq.~\eqref{eq: basic graph condition}. Thus, a collection of functions $f_1, \ldots, f_d$ determines the edge weight function $w_E$ and vice versa. \end{proof} \begin{definition} \label{co-boundary} Let $G = (V, E)$ be a hypergraph. The \emph{co-boundary} of a~vertex $v \in V$ is an edge weight function $\delta_v$ defined be the formula $$ \delta_v(e) = \left\{\begin{array}{cc} 1, & v \in e \\ 0, & v \notin e. \end{array}\right. $$ In other words, the co-boundary $\delta_v$ is the indicator function of the subset of all edges incident to $v$. Also, the reader can interpret it as a row of the incidence matrix of $G$ corresponding to the vertex $v$. \end{definition} \begin{remark} The notion of co-boundary comes from \cite{Prasolov2006}, although there it has a~slightly different form. Note, that by the co-boundary of a vertex of a~hypergraph we also mean a particular case of the co-boundary of a vertex of a graph. \end{remark} \begin{lemma} \label{claim1} A hypergraph $G = (V, E)$ is basic if and only if the co-boundaries of its vertices are linearly independent in $\mathbb{R}^{|E|}$. \end{lemma} \begin{proof Let $V = \{v_1,\ldots,v_n\}$, $E = \{e_1,\ldots,e_m\}$ and $A$ be the incidence matrix of the hypergraph $G$. Then $G$ is basic if and only if for any vertex weight function $w_V$ there exists an edge weight function $w_E$ such that $$ A\begin{pmatrix} w_E(e_1) \\ \vdots \\ w_E(e_m) \\ \end{pmatrix} = \begin{pmatrix} w_V(v_1) \\ \vdots \\ w_V(v_n) \\ \end{pmatrix}, $$ meaning, that rows of $A$ (i.e. co-boundaries) are linearly independent. \end{proof} \begin{lemma} \label{claim2} The co-boundaries of vertices of a graph $G=(V,E)$ are linearly independent in $\mathbb{R}^{|E|}$ if and only if $G$ does not contain a bipartite connected component. \end{lemma} \begin{proof Assume that there is a linear dependence \begin{equation}\label{eq: dependence} \sum \limits_{i = 1}^{n} \lambda_i \delta_{v_i} = 0. \end{equation} If there is an edge between vertices $v_i$ and $v_j$, then $\lambda_i = -\lambda_j$. Thus, vertices in every connected component are divided into two equivalence classes with coefficients $\lambda_i$ and $-\lambda_i$ respectively. Vertices from each of the classes are adjacent only to vertices from the other class. Hence, this component of the graph is bipartite. Conversely, if $G$ contains a bipartite component $C$ with parts $A$ and $B$, then there exists a linear dependence of form~\eqref{eq: dependence} with $$ \lambda_i = \left\{\begin{array}{rl} 1, & v_i \in A \\ -1, & v_i \in B \\ 0, & v_i \notin C. \end{array}\right. $$ \end{proof} \begin{proof}[Proof of Theorem~\ref{th_basis_graph}] It is sufficient to apply Lemma~\ref{claim1} and Lemma~\ref{claim2}. \end{proof} \begin{corollary}\label{cor:graph_to_set} Let the intersection of a subset $M\subset[n]^d$ with each layer consist of two or zero points. Then $M$ is basic if and only if the corresponding graph $G(M)$ does not contain a bipartite connected component. \end{corollary} \begin{example} \label{example1} Let $M = \{(2,1,1), (1,2,1), (1,1,2), (2,2,2)\}$ (Fig.~\ref{figure: basic set, 4 points}). Then the corresponding graph $G(M)$ is the complete graph $K_4$ with four vertices. Since $K_4$ is connected and not bipartite, in accordance with Corollary~\ref{cor:graph_to_set} the set $M$ is basic. \end{example} \begin{center} \begin{tikzpicture}[scale=1.2,x={(1cm,0cm)},y={(0.54cm,0.36cm)},z={(0cm,1cm)},line width=.5pt] \begin{scope} \coordinate (a1) at (0,0,0); \coordinate (a2) at (1,0,0); \coordinate (a3) at (1,1,0); \coordinate (a4) at (0,1,0); \coordinate (b1) at (0,0,1); \coordinate (b2) at (1,0,1); \coordinate (b3) at (1,1,1); \coordinate (b4) at (0,1,1); \draw [line width=.8pt] (b1)--(b4)--(b3)--(b2)--(b1)--(a1)--(a2)--(b2); \draw [line width=.8pt] (a2)--(a3)--(b3); \draw [dashed] (a1)--(a4); \draw [dashed] (b4)--(a4)--(a3); \foreach \EuScript P in {a1,a3,b2,b4}{ \filldraw[white] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \foreach \EuScript P in {a2,a4,b1,b3} \filldraw (\EuScript P) circle (1.7pt); \refstepcounter{ris} \draw (0.5,0.5,-0.6) node {Figure \arabic{ris}.\label{figure: basic set, 4 points}}; \end{scope} \end{tikzpicture} \end{center} \section{Structure of non-basic subsets}\label{Section:structure_of_non-basic_sets} By Lemma~\ref{lemma: M is min. non-basic <=> annihilation weight function is unique}, if a non-basic subset $M \subset [n]^d$ is minimal, then there exists the unique annihilation function up to multiplying by a constant. Since the matrix~$A_M$ has integer entries, it is possible to choose the annihilation function $f$ with integer values that are setwise coprime integers, so that the notion of irreducible annihilation function is well-defined. The name \emph{irreducible} is chosen by analogy with integer fractions. \begin{example} \label{example: non-basic 5 points in 2x2x2} If, as it is shown in Fig.~\ref{figure: non-basic set, 5 points}, $$ M = \{(1,1,1), (2,1,1), (1,2,1), (1,1,2), (2,2,2)\}, $$ then the irreducible annihilation function of $M$ has the following form: $$ f = 2\cdot\textbf{1}_{(1,1,1)} - \textbf{1}_{(2,1,1)} - \textbf{1}_{(1,2,1)} - \textbf{1}_{(1,1,2)} + \textbf{1}_{(2,2,2)} $$ \end{example} \begin{center} \begin{tikzpicture}[scale=1.2,x={(1cm,0cm)},y={(0.54cm,0.36cm)},z={(0cm,1cm)},line width=.5pt] \begin{scope} \coordinate (a1) at (0,0,0); \coordinate (a2) at (1,0,0); \coordinate (a3) at (1,1,0); \coordinate (a4) at (0,1,0); \coordinate (b1) at (0,0,1); \coordinate (b2) at (1,0,1); \coordinate (b3) at (1,1,1); \coordinate (b4) at (0,1,1); \draw [line width=.8pt] (b1)--(b4)--(b3)--(b2)--(b1)--(a1)--(a2)--(b2); \draw [line width=.8pt] (a2)--(a3)--(b3); \draw [dashed] (a1)--(a4); \draw [dashed] (b4)--(a4)--(a3); \foreach \EuScript P in {a3,b2,b4}{ \filldraw[white] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \foreach \EuScript P in {a1,a2,a4,b1,b3} \filldraw (\EuScript P) circle (1.7pt); \refstepcounter{ris} \draw (0.5,0.5,-0.6) node {Figure \arabic{ris}.\label{figure: non-basic set, 5 points}}; \end{scope} \end{tikzpicture} \end{center} \begin{proof}[Proof of Theorem~\ref{th_irreducible_annihilation_function_is_unbounded}] Let $a_1,\ldots,a_m,b_1,\ldots,b_m,c_1,\ldots,c_m,d,e$ be different integers. Define the set $M$ to consist of the following $6m+2$ elements (see Fig.~\ref{figure: non-basic set, 6k+2 points}): \begin{itemize} \item $3m$ points with coordinates $(d,a_k,a_k)$, $(b_k,d,b_k)$, $(c_k,c_k,d)$, $k \in [m]$; \item $3m$ points with coordinates $(e,a_k,a_{k+1})$, $(b_k,e,b_{k+1})$, $(c_k,c_{k+1},e)$,\\ $k \in [m]$ (here, we assume that $a_{m+1}=a_1$, $b_{m+1}=b_1$ and $c_{m+1}=c_1$); \item $1$ point with coordinates $(d,d,d)$; \item $1$ point with coordinates $(e,e,e)$. \end{itemize} \begin{center} \begin{tikzpicture}[line width=.8pt] \begin{scope} \filldraw[rounded corners, green, opacity = 0.2] (6.0,2.3) -- (-0.3,1.2) .. controls (-0.5,0.5) and (3.5,0.5) .. (3.3,1.2)--(6.0,2)--cycle; \filldraw[rounded corners, yellow, opacity = 0.4] (5.5,2.3) -- (3.7,1.2) .. controls (3.5,0.5) and (7.5,0.5) .. (7.3,1.2)--cycle; \filldraw[rounded corners, red, opacity = 0.2] (5.0,2.3) -- (11.3,1.2) .. controls (11.5,0.5) and (7.5,0.5) .. (7.7,1.2)--(5.0,2)--cycle; \draw[rounded corners] (6.0,2.3) -- (-0.3,1.2) .. controls (-0.5,0.5) and (3.5,0.5) .. (3.3,1.2)--(6.0,2)--cycle; \draw[rounded corners] (5.5,2.3) -- (3.7,1.2) .. controls (3.5,0.5) and (7.5,0.5) .. (7.3,1.2)--cycle; \draw[rounded corners] (5.0,2.3) -- (11.3,1.2) .. controls (11.5,0.5) and (7.5,0.5) .. (7.7,1.2)--(5.0,2)--cycle; \filldraw[rounded corners, red, opacity = 0.2] (6.0,-1.3) -- (-0.3,-0.2) .. controls (-0.5,0.5) and (3.5,0.5) .. (3.3,-0.2)--(6.0,-1)--cycle; \filldraw[rounded corners, yellow, opacity = 0.4] (5.5,-1.3) -- (3.7,-0.2) .. controls (3.5,0.5) and (7.5,0.5) .. (7.3,-0.2)--cycle; \filldraw[rounded corners, green, opacity = 0.2] (5.0,-1.3) -- (11.3,-0.2) .. controls (11.5,0.5) and (7.5,0.5) .. (7.7,-0.2)--(5.0,-1)--cycle; \draw[rounded corners] (6.0,-1.3) -- (-0.3,-0.2) .. controls (-0.5,0.5) and (3.5,0.5) .. (3.3,-0.2)--(6.0,-1)--cycle; \draw[rounded corners] (5.5,-1.3) -- (3.7,-0.2) .. controls (3.5,0.5) and (7.5,0.5) .. (7.3,-0.2)--cycle; \draw[rounded corners] (5.0,-1.3) -- (11.3,-0.2) .. controls (11.5,0.5) and (7.5,0.5) .. (7.7,-0.2)--(5.0,-1)--cycle; \foreach \EuScript X in {0,...,2}{ \foreach \EuScript Y in {0,...,2}{ \draw (4*\EuScript X+\EuScript Y,0)--++(1,1); } \draw (4*\EuScript X,1)--++(3,-1); } \foreach \EuScript X in {0,...,11}{ \draw (\EuScript X,0)--++(0,1); \filldraw (\EuScript X,0) circle (1.7pt); \filldraw (\EuScript X,1) circle (1.7pt); } \filldraw (5.5,2) circle (1.7pt); \filldraw (5.5,-1) circle (1.7pt); \refstepcounter{ris} \draw (5.5,-1.8) node {Figure \arabic{ris}.\label{figure: non-basic set, 6k+2 points}}; \end{scope} \end{tikzpicture} \end{center} Then the function $f$ taking the values $1$, $-1$, $-m$ and $m$ at the points of first, second, third and fourth groups respectively is irreducible annihilation. Now we show that the set $M$ is minimal. Indeed, $2m$ points with coordinates $(d,a_k,a_k)$ and $(e,a_k,a_{k+1})$ constitute a cycle of order $2m$, whose elements belong (or do not belong) to a minimal subset $N$ of $M$ simultaneously. Two other cycles behave the same way. To finish the proof, we need to mention that points $(d,d,d)$ and $(e,e,e)$ belong to $N$ for sure, and if $(e,e,e)$ belongs to $N$, then each of three cycles does as well. Hence, all $6m+2$ points are in~$N$ and $N=M$. \end{proof} \begin{definition}\label{def: simple annihilation function} Let $\EuScript P,\EuScript Q,\EuScript R,\EuScript S$ be the consecutive vertices of a rectangle whose sides are parallel to the coordinate axes. A \emph{simple annihilation function} is $$ f_{\EuScript P\EuScript Q\EuScript R\EuScript S} = \textbf{1}_{\EuScript P} - \textbf{1}_{\EuScript Q} + \textbf{1}_{\EuScript R} - \textbf{1}_{\EuScript S}. $$ \end{definition} By double induction on $d$ and $n$, one can verify that every annihilation function of $[n]^d$ can be decomposed into a~finite sum of simple annihilation functions (Theorem~\ref{thBoyarov}). The proof is routine and will be omitted. \begin{example} \label{example: non-basic 5 points in 2x2x2 revisited} For the set $M$ from Example~\ref{example: non-basic 5 points in 2x2x2}, its annihilation function $$ f = 2\cdot\textbf{1}_{(1,1,1)} - \textbf{1}_{(2,1,1)} - \textbf{1}_{(1,2,1)} - \textbf{1}_{(1,1,2)} + \textbf{1}_{(2,2,2)} $$ is decomposed as the sum of simple annihilation functions $$ \textbf{1}_{(1,1,1)} - \textbf{1}_{(1,2,1)} + \textbf{1}_{(2,2,1)} - \textbf{1}_{(2,1,1)}, $$ $$ \textbf{1}_{(1,1,1)} - \textbf{1}_{(1,1,2)} + \textbf{1}_{(2,1,2)} - \textbf{1}_{(2,1,1)} $$ and $$ \textbf{1}_{(2,1,1)} - \textbf{1}_{(2,2,1)} + \textbf{1}_{(2,2,2)} - \textbf{1}_{(2,1,2)}. $$ \end{example} \section{Conclusion} \label{Section:conclusion} As we have seen in Section~\ref{Section:estimations}, Theorem~\ref{th: M is non-basic => 2n-1 < |M| < dn - (d-3)} provides some bounds for the number of elements in certain non-basic subsets of $[n]^d$. While the lower bound is reachable (see Remark~\ref{Remark: lower bound is reachable}), it is not clear, whether the upper bound and the intermediate values are reachable as well. This observation leads us to the following question. \begin{question}\label{question: reachability} Are the intermediate values in Theorem~\ref{th: M is non-basic => 2n-1 < |M| < dn - (d-3)} reachable? In other words, is it true that for any integer $k$, $2n < k \leqslant dn - (d-2)$, there exists a~minimal non-basic subset $M\subset[n]^d$ of size $k$ such that every layer of~$[n]^d$ has a non-empty intersection with $M$? \end{question} Figures~\ref{figure: non-basic sets in 3x3x3},~\ref{figure: non-basic sets, n=4, |M|=8,9} and~\ref{figure: non-basic sets, n=4, |M|=10,11} shows that for $d=3$, $n\in\{3,4\}$ the answer to Question~\ref{question: reachability} is positive (here, a point color indicates the value of the irreducible annihilation function: red, blue, green and black are reserved for $-1$, $1$, $-2$ and $2$ respectively). \begin{center} \begin{tikzpicture}[scale=1.2,x={(1cm,0cm)},y={(0.4cm,0.3cm)},z={(0cm,1cm)},line width=.5pt] \begin{scope} \coordinate (a) at (2,0,0); \coordinate (b) at (2,1,0); \coordinate (c) at (1,0,1); \coordinate (d) at (1,2,1); \coordinate (e) at (0,1,2); \coordinate (f) at (0,2,2); \foreach \EuScript P in {0,1,2}{ \draw [line width=.8pt] (0,\EuScript P,2)--++(2,0,0)--++(0,0,-2); \draw [line width=.8pt] (0,0,\EuScript P)--++(2,0,0)--++(0,2,0); } \foreach \EuScript P in {0,1}{ \draw [line width=.8pt] (\EuScript P,0,0)--++(0,0,2)--++(0,2,0); \draw [dashed] (0,\EuScript P+1,2)--++(0,0,-2)--++(2,0,0); \draw [dashed] (\EuScript P,0,0)--++(0,2,0)--++(0,0,2); } \draw [dashed] (0,0,1)--++(0,2,0)--++(2,0,0); \draw [dashed] (0,1,1)--++(2,0,0); \draw [dashed] (1,0,1)--++(0,2,0); \draw [dashed] (1,1,0)--++(0,0,2); \foreach \EuScript X in {0,1,2}{ \foreach \EuScript Y in {0,1,2}{ \foreach \z in {0,1,2}{ \filldraw[white] (\EuScript X,\EuScript Y,\z) circle (1.7pt); \draw (\EuScript X,\EuScript Y,\z) circle (1.7pt); } } } \foreach \EuScript P in {a,d,e}{ \filldraw[red] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \foreach \EuScript P in {b,c,f}{ \filldraw[blue] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \end{scope} \begin{scope}[xshift=4cm] \coordinate (a) at (2,0,0); \coordinate (b) at (2,2,1); \coordinate (c) at (1,1,1); \coordinate (d) at (1,2,2); \coordinate (e) at (0,0,0); \coordinate (f) at (0,1,2); \coordinate (g) at (0,2,2); \foreach \EuScript P in {0,1,2}{ \draw [line width=.8pt] (0,\EuScript P,2)--++(2,0,0)--++(0,0,-2); \draw [line width=.8pt] (0,0,\EuScript P)--++(2,0,0)--++(0,2,0); } \foreach \EuScript P in {0,1}{ \draw [line width=.8pt] (\EuScript P,0,0)--++(0,0,2)--++(0,2,0); \draw [dashed] (0,\EuScript P+1,2)--++(0,0,-2)--++(2,0,0); \draw [dashed] (\EuScript P,0,0)--++(0,2,0)--++(0,0,2); } \draw [dashed] (0,0,1)--++(0,2,0)--++(2,0,0); \draw [dashed] (0,1,1)--++(2,0,0); \draw [dashed] (1,0,1)--++(0,2,0); \draw [dashed] (1,1,0)--++(0,0,2); \foreach \EuScript X in {0,1,2}{ \foreach \EuScript Y in {0,1,2}{ \foreach \z in {0,1,2}{ \filldraw[white] (\EuScript X,\EuScript Y,\z) circle (1.7pt); \draw (\EuScript X,\EuScript Y,\z) circle (1.7pt); } } } \foreach \EuScript P in {g} \filldraw (\EuScript P) circle (1.7pt); \foreach \EuScript P in {b,d,e,f}{ \filldraw[red] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \foreach \EuScript P in {a,c}{ \filldraw[blue] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \refstepcounter{ris} \draw (1,1,-0.7) node {Figure \arabic{ris}.\label{figure: non-basic sets in 3x3x3}}; \end{scope} \begin{scope}[xshift=8cm] \coordinate (a) at (2,0,0); \coordinate (b) at (2,0,1); \coordinate (c) at (2,1,0); \coordinate (d) at (1,0,0); \coordinate (e) at (1,2,2); \coordinate (f) at (0,1,2); \coordinate (g) at (0,2,1); \coordinate (h) at (0,2,2); \foreach \EuScript P in {0,1,2}{ \draw [line width=.8pt] (0,\EuScript P,2)--++(2,0,0)--++(0,0,-2); \draw [line width=.8pt] (0,0,\EuScript P)--++(2,0,0)--++(0,2,0); } \foreach \EuScript P in {0,1}{ \draw [line width=.8pt] (\EuScript P,0,0)--++(0,0,2)--++(0,2,0); \draw [dashed] (0,\EuScript P+1,2)--++(0,0,-2)--++(2,0,0); \draw [dashed] (\EuScript P,0,0)--++(0,2,0)--++(0,0,2); } \draw [dashed] (0,0,1)--++(0,2,0)--++(2,0,0); \draw [dashed] (0,1,1)--++(2,0,0); \draw [dashed] (1,0,1)--++(0,2,0); \draw [dashed] (1,1,0)--++(0,0,2); \foreach \EuScript X in {0,1,2}{ \foreach \EuScript Y in {0,1,2}{ \foreach \z in {0,1,2}{ \filldraw[white] (\EuScript X,\EuScript Y,\z) circle (1.7pt); \draw (\EuScript X,\EuScript Y,\z) circle (1.7pt); } } } \foreach \EuScript P in {h} \filldraw (\EuScript P) circle (1.7pt); \foreach \EuScript P in {a}{ \filldraw[green] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \foreach \EuScript P in {e,f,g}{ \filldraw[red] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \foreach \EuScript P in {b,c,d}{ \filldraw[blue] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \end{scope} \end{tikzpicture} \end{center} \begin{center} \begin{tikzpicture}[scale=1.2,x={(1cm,0cm)},y={(0.3cm,0.2cm)},z={(0cm,1cm)},line width=.5pt] \begin{scope} \coordinate (a) at (0,0,0); \coordinate (b) at (1,0,0); \coordinate (c) at (1,1,1); \coordinate (d) at (2,1,1); \coordinate (e) at (2,2,2); \coordinate (f) at (3,2,2); \coordinate (g) at (3,3,3); \coordinate (h) at (0,3,3); \foreach \EuScript P in {0,1,2,3}{ \draw [line width=.8pt] (0,\EuScript P,3)--++(3,0,0)--++(0,0,-3); \draw [line width=.8pt] (0,0,\EuScript P)--++(3,0,0)--++(0,3,0); } \foreach \EuScript P in {0,1,2}{ \draw [line width=.8pt] (\EuScript P,0,0)--++(0,0,3)--++(0,3,0); \draw [dashed] (0,\EuScript P+1,3)--++(0,0,-3)--++(3,0,0); \draw [dashed] (\EuScript P,0,0)--++(0,3,0)--++(0,0,3); } \foreach \EuScript P in {1,2}{ \draw [dashed] (0,0,\EuScript P)--++(0,3,0)--++(3,0,0); \foreach \EuScript Q in {1,2}{ \draw [dashed] (0,\EuScript P,\EuScript Q)--++(3,0,0); \draw [dashed] (\EuScript P,0,\EuScript Q)--++(0,3,0); \draw [dashed] (\EuScript P,\EuScript Q,0)--++(0,0,3); } } \foreach \EuScript X in {0,1,2,3}{ \foreach \EuScript Y in {0,1,2,3}{ \foreach \z in {0,1,2,3}{ \filldraw[white] (\EuScript X,\EuScript Y,\z) circle (1.7pt); \draw (\EuScript X,\EuScript Y,\z) circle (1.7pt); } } } \foreach \EuScript P in {a,c,e,g}{ \filldraw[red] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \foreach \EuScript P in {b,d,f,h}{ \filldraw[blue] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \end{scope} \begin{scope}[xshift=5cm] \coordinate (a) at (0,3,3); \coordinate (b) at (0,3,2); \coordinate (c) at (0,2,3); \coordinate (d) at (1,3,3); \coordinate (e) at (1,0,1); \coordinate (f) at (2,2,0); \coordinate (g) at (3,1,2); \coordinate (h) at (2,1,1); \coordinate (i) at (3,0,0); \foreach \EuScript P in {0,1,2,3}{ \draw [line width=.8pt] (0,\EuScript P,3)--++(3,0,0)--++(0,0,-3); \draw [line width=.8pt] (0,0,\EuScript P)--++(3,0,0)--++(0,3,0); } \foreach \EuScript P in {0,1,2}{ \draw [line width=.8pt] (\EuScript P,0,0)--++(0,0,3)--++(0,3,0); \draw [dashed] (0,\EuScript P+1,3)--++(0,0,-3)--++(3,0,0); \draw [dashed] (\EuScript P,0,0)--++(0,3,0)--++(0,0,3); } \foreach \EuScript P in {1,2}{ \draw [dashed] (0,0,\EuScript P)--++(0,3,0)--++(3,0,0); \foreach \EuScript Q in {1,2}{ \draw [dashed] (0,\EuScript P,\EuScript Q)--++(3,0,0); \draw [dashed] (\EuScript P,0,\EuScript Q)--++(0,3,0); \draw [dashed] (\EuScript P,\EuScript Q,0)--++(0,0,3); } } \foreach \EuScript X in {0,1,2,3}{ \foreach \EuScript Y in {0,1,2,3}{ \foreach \z in {0,1,2,3}{ \filldraw[white] (\EuScript X,\EuScript Y,\z) circle (1.7pt); \draw (\EuScript X,\EuScript Y,\z) circle (1.7pt); } } } \foreach \EuScript P in {a} \filldraw (\EuScript P) circle (1.7pt); \foreach \EuScript P in {b,c,d,h,i}{ \filldraw[red] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \foreach \EuScript P in {e,f,g}{ \filldraw[blue] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \end{scope} \refstepcounter{ris}\label{figure: non-basic sets, n=4, |M|=8,9} \draw (5.0,-1.8) node {Figure \arabic{ris}.}; \end{tikzpicture} \end{center} \begin{center} \begin{tikzpicture}[scale=1.2,x={(1cm,0cm)},y={(0.3cm,0.2cm)},z={(0cm,1cm)},line width=.5pt] \begin{scope} \coordinate (a) at (0,3,3); \coordinate (b) at (3,0,0); \coordinate (c) at (3,3,2); \coordinate (d) at (2,3,0); \coordinate (e) at (3,1,3); \coordinate (f) at (1,0,3); \coordinate (g) at (0,2,0); \coordinate (h) at (0,0,1); \coordinate (i) at (2,1,1); \coordinate (j) at (1,2,2); \foreach \EuScript P in {0,1,2,3}{ \draw [line width=.8pt] (0,\EuScript P,3)--++(3,0,0)--++(0,0,-3); \draw [line width=.8pt] (0,0,\EuScript P)--++(3,0,0)--++(0,3,0); } \foreach \EuScript P in {0,1,2}{ \draw [line width=.8pt] (\EuScript P,0,0)--++(0,0,3)--++(0,3,0); \draw [dashed] (0,\EuScript P+1,3)--++(0,0,-3)--++(3,0,0); \draw [dashed] (\EuScript P,0,0)--++(0,3,0)--++(0,0,3); } \foreach \EuScript P in {1,2}{ \draw [dashed] (0,0,\EuScript P)--++(0,3,0)--++(3,0,0); \foreach \EuScript Q in {1,2}{ \draw [dashed] (0,\EuScript P,\EuScript Q)--++(3,0,0); \draw [dashed] (\EuScript P,0,\EuScript Q)--++(0,3,0); \draw [dashed] (\EuScript P,\EuScript Q,0)--++(0,0,3); } } \foreach \EuScript X in {0,1,2,3}{ \foreach \EuScript Y in {0,1,2,3}{ \foreach \z in {0,1,2,3}{ \filldraw[white] (\EuScript X,\EuScript Y,\z) circle (1.7pt); \draw (\EuScript X,\EuScript Y,\z) circle (1.7pt); } } } \foreach \EuScript P in {a,b} \filldraw (\EuScript P) circle (1.7pt); \foreach \EuScript P in {c,d,e,f,g,h}{ \filldraw[red] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \foreach \EuScript P in {i,j}{ \filldraw[blue] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \end{scope} \begin{scope}[xshift=5cm] \coordinate (a) at (0,3,3); \coordinate (b) at (3,0,0); \coordinate (c) at (1,2,3); \coordinate (d) at (1,2,2); \coordinate (e) at (1,1,1); \coordinate (f) at (2,2,1); \coordinate (g) at (2,3,0); \coordinate (h) at (3,0,2); \coordinate (i) at (0,0,1); \coordinate (j) at (3,3,1); \coordinate (k) at (0,1,0); \coordinate (l) at (2,3,0); \foreach \EuScript P in {0,1,2,3}{ \draw [line width=.8pt] (0,\EuScript P,3)--++(3,0,0)--++(0,0,-3); \draw [line width=.8pt] (0,0,\EuScript P)--++(3,0,0)--++(0,3,0); } \foreach \EuScript P in {0,1,2}{ \draw [line width=.8pt] (\EuScript P,0,0)--++(0,0,3)--++(0,3,0); \draw [dashed] (0,\EuScript P+1,3)--++(0,0,-3)--++(3,0,0); \draw [dashed] (\EuScript P,0,0)--++(0,3,0)--++(0,0,3); } \foreach \EuScript P in {1,2}{ \draw [dashed] (0,0,\EuScript P)--++(0,3,0)--++(3,0,0); \foreach \EuScript Q in {1,2}{ \draw [dashed] (0,\EuScript P,\EuScript Q)--++(3,0,0); \draw [dashed] (\EuScript P,0,\EuScript Q)--++(0,3,0); \draw [dashed] (\EuScript P,\EuScript Q,0)--++(0,0,3); } } \foreach \EuScript X in {0,1,2,3}{ \foreach \EuScript Y in {0,1,2,3}{ \foreach \z in {0,1,2,3}{ \filldraw[white] (\EuScript X,\EuScript Y,\z) circle (1.7pt); \draw (\EuScript X,\EuScript Y,\z) circle (1.7pt); } } } \foreach \EuScript P in {a,b}{ \filldraw[black] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \foreach \EuScript P in {c}{ \filldraw[green] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \foreach \EuScript P in {g,h,i,j,k}{ \filldraw[red] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \foreach \EuScript P in {d,e,f}{ \filldraw[blue] (\EuScript P) circle (1.7pt); \draw (\EuScript P) circle (1.7pt); } \end{scope} \refstepcounter{ris} \draw (5.0,-1.8) node {Figure \arabic{ris}.\label{figure: non-basic sets, n=4, |M|=10,11}}; \end{tikzpicture} \end{center} Surprisingly, in all known examples, including the sets shown in Figs.~\ref{figure: minimal non-basic set in 4x4x4}--\ref{figure: non-basic sets, n=4, |M|=10,11}, the values of irreducible annihilation functions present a specific behavior. This allows us to state the following conjecture. \begin{conjecture} If $M \subset [n]^3$ is a minimal non-basic subset such that every layer of $[n]^3$ has a non-empty intersection with $M$, then its irreducible annihilation function $f$ satisfies $$ \sum\limits_{\EuScript X\in M} |f(\EuScript X)| = 2\big(|M| - n\big). $$ \end{conjecture} As we mentioned before, due to Theorem~\ref{th_irreducible_annihilation_function_is_unbounded}, there is no reason to expect the existence of simple criterion (similar to Theorem~\ref{th_set_to_graph}) for a subset of $[n]^d$ to be basic in the general case $d\geqslant3$. Still, it does not mean that there is no simplification at all, and it would be interesting to find one, at least for $d=3$ or for the case of small values of~$n$. Another possible direction for research would come from the generalization of the initial problem to hypergraphs. As we have seen in Section~\ref{Section:graphs_app}, the concept of basic hypergraphs admits almost the same interpretation in algebraic terms as the one of basic subsets. We can make this similarity even deeper as follows. Let $G = (V, E)$ be a~hypergraph. Define a linear map $\Psi \colon \mathbb{R}^{|V|} \to \mathbb{R}^{|E|}$ on indicators, $$ \Psi(\textbf{1}_{v}) = \sum\limits_{e\colon v \in e} \textbf{1}_e = \delta_v, $$ and extend it on $\mathbb{R}^{|V|}$ by linearity. In other words, for any $v \in V$, we set the image of the indicator function~$\textbf{1}_{v}$ to be the co-boundary $\delta_v$. Then the analogue of Lemma~\ref{lemma: M is non-basic <=> annihilation weight function} is that the hypergraph is basic if and only if $\ker\Psi$ is trivial, while the analogue of Lemma~\ref{lemma: M is min. non-basic <=> annihilation weight function is unique} is that for minimal non-basic hypergraphs we have $\dim\ker\Psi=1$. At the same time, the analogue of Theorem~\ref{th: M is basic => |M| < dn - (d-2)} is that for a basic hypergraph $G = (V, E)$, on has $|V| \leqslant |E|$ (which is trivial). It is natural to state the following general question. \begin{question}\label{question: basic hypergraph} What are the conditions for a hypergraph to be basic? \end{question} For now, this question in its generality is open. \section{Acknowledgements} \label{Section:acknowledgements} We thank A.B.~Skopenkov for useful discussions and criticism, N.~Volkov for searching examples of non-basic subsets and I.~Boyarov for proving the weaker version of Theorem~\ref{thBoyarov}. This work was partly supported by Russian Science Foundation Grant N~22-11-00177.

https://arxiv.org/abs/1201.6035

How Accurate is inv(A)*b?

Several widely-used textbooks lead the reader to believe that solving a linear system of equations Ax = b by multiplying the vector b by a computed inverse inv(A) is inaccurate. Virtually all other textbooks on numerical analysis and numerical linear algebra advise against using computed inverses without stating whether this is accurate or not. In fact, under reasonable assumptions on how the inverse is computed, x = inv(A)*b is as accurate as the solution computed by the best backward-stable solvers. This fact is not new, but obviously obscure. We review the literature on the accuracy of this computation and present a self-contained numerical analysis of it.

\section{Introduction} Can you accurately compute the solution to a linear equation $Ax=b$ by first computing an approximation $V$ to $A^{-1}$ and then multiplying $b$ by $V$ (\texttt{x=inv(A){*}b} in Matlab)? Unfortunately, most of the literature provides a misleading answer to this question. Many textbooks, including recent and widely-used ones, mislead the reader to think that \texttt{x=inv(A){*}b} is less accurate than \texttt{x=A\textbackslash{}b}, which computes the $LU$ factorization of $A$ with partial pivoting and then solves for $x$ using the factors~\emph{\cite[p. 31]{ForsytheMalcolmMoler}}, \emph{\cite[p. 53]{Moler}}, \emph{\cite[p. 50]{Heath}}, \emph{\cite[p. 77]{OLeary}}, \emph{\cite[p. 166]{ConteDeBoor}}, \emph{\cite[pp. 184, 235, and 246]{Stewart}}. Other textbooks warn against using a computed inverse for performance reasons without saying anything about accuracy. If you still dare use \texttt{x=inv(A){*}b} in Matlab code, Matlab's analyzer issues a wrong and misleading warning~\cite{MLint-R2010b}. As far as we can tell, only two sources in the literature present a correct analysis of this question. One is almost 50 years old~\cite[pp. 128--129]{Wilkinson}, and is therefore hard to obtain and somewhat hard to read. The other is recent, but relegates this analysis to the solution of an exercise, rather than including it in the 27-page chapter on the matrix inverse\emph{~\cite[p. 559; see also p. 260]{Higham}}; even though the analysis there shows that \texttt{x=inv(A){*}b} is as accurate as \texttt{x=A\textbackslash{}b}, the text ends by stating that {}``multiplying by an explicit inverse is simply not a good way to solve a linear system''. The reader must pay careful attention to the analysis if he or she is to answer our question correctly. Our aim in this article is to clarify to researchers (and perhaps also to educators and students) the numerical properties of a solution to $Ax=b$ that is obtained by multiplying by a computed inverse. We do not present new results; we present results that are known, but not as much as they should be. Computing the inverse requires more arithmetic operations than computing an $LU$ factorization. We do not address the question of computational efficiency, but we do note that there is evidence that using the inverse is sometimes preferable from the performance perspective~\cite{DitkowskiFibichGavish}. It also appears that explicit inverses are sometimes used when the inverse must be applied in hardware, as in some MIMO radios~\cite{eberli08,StuderEtAl2011}. The numerical analysis in the literature and in this paper does not apply as-is to these computations, because hardware implementations typically use fixed-point arithmetic rather than floating point. Still, the analysis that we present here provides guiding principles to all implementations (e.g., to solve for the rows of the inverse using a backward-stable solver), and it may also provide a template for an analysis of fixed-point implementations or alternative inversion algorithms. The rest of this paper is organized as follows. Section~\ref{sec:A-Loose-Bound} presents the naive numerical analysis that probably led many authors to claim that \texttt{x=inv(A){*}b} is inaccurate; the analysis is correct, but the error bound that it yields is too loose. Section~\ref{sec:Tightening-the-Bound} presents a much tighter analysis, due to Wilkinson; Higham later showed that this bound holds even in the componentwise sense. Section~\ref{sec:Left-and-Right} explains another aspect of computed inverses that is not widely appreciated: that they are typically good for applying either from the left or from the right, but not both. Even when \texttt{x=inv(A){*}b} is accurate, \texttt{x} is usually not backward stable; Section~\ref{sec:backward-stability-of-xv} discusses conditions under which \texttt{x} is also backward stable. To help the reader fully understand all of these results, Section~\ref{sec:Numerical-Examples} demonstrates them using simple numerical experiments. We present concluding remarks in Section~\ref{sec:Closing-Remarks}. \section{\label{sec:A-Loose-Bound}A Loose Bound} Why did the inverse acquire its bad reputation? Good inversion methods produce a computed inverse $V$ that is, at best, \emph{conditionally} accurate, \begin{equation} \frac{\left\Vert V-A^{-1}\right\Vert }{\left\Vert A^{-1}\right\Vert }=O(\kappa(A)\epsilon_{\text{machine}})\;.\label{eq:conditionally-accurate-V} \end{equation} We cannot hope for an unconditional bound of $O(\epsilon_{\text{machine}})$ on the relative forward error. Some inversion methods guarantee conditional accuracy (for example, computing the inverse column by column using a backward stable linear solver). In particular, \noun{lapack}'s \texttt{xGETRI} satisfies (\ref{eq:conditionally-accurate-V}), and also a componentwise conditional bound~\cite[p. 268]{Higham}. That is, each entry in the computed inverse that \texttt{xGETRI} produces is conditionally accurate. It appears that Matlab's \texttt{inv} function also satisfies (\ref{eq:conditionally-accurate-V}). Let's try to use (\ref{eq:conditionally-accurate-V}) to obtain a bound on the forward error $\|x_{V}-x\|$. Multiplying $b$ by $V$ in floating point produces $x_{V}$ that satisfies $x_{V}=\left(V+\Delta\right)b$ for some $\Delta$ with $\|\Delta\|/\|V\|=O(\epsilon_{\text{machine}})$. Denoting $\Gamma=V-A^{-1}$, we have \begin{eqnarray*} x_{V} & = & \left(V+\Delta\right)b\\ & = & \left(A^{-1}+\Gamma+\Delta\right)b\\ & = & \left(A^{-1}+\Gamma+\Delta\right)Ax\\ & = & x+\Gamma Ax+\Delta Ax\;, \end{eqnarray*} so \begin{eqnarray} \left\Vert x_{V}-x\right\Vert & \leq & \left\Vert \Gamma\right\Vert \left\Vert A\right\Vert \left\Vert x\right\Vert +\left\Vert \Delta\right\Vert \left\Vert A\right\Vert \left\Vert x\right\Vert \nonumber \\ & \leq & O(\kappa(A)\epsilon_{\text{machine}})\left\Vert A^{-1}\right\Vert \left\Vert A\right\Vert \left\Vert x\right\Vert +O(\epsilon_{\text{machine}})\left\Vert V\right\Vert \left\Vert A\right\Vert \left\Vert x\right\Vert \nonumber \\ & = & O(\kappa^{2}(A)\epsilon_{\text{machine}})\left\Vert x\right\Vert +O(\epsilon_{\text{machine}})\left\Vert V\right\Vert \left\Vert A\right\Vert \left\Vert x\right\Vert \;.\label{eq:loose-accuracy-bound-pre} \end{eqnarray} Unless $A$ is so ill conditioned that the left-hand side of (\ref{eq:conditionally-accurate-V}) is larger than $1$ (any constant would do), $\|V\|=\Theta(\|A^{-1}\|)$. Therefore, \begin{equation} \left\Vert x_{V}-x\right\Vert \leq O(\kappa^{2}(A)\epsilon_{\text{machine}})\left\Vert x\right\Vert \;.\label{eq:loose-bound} \end{equation} In contrast, solving $Ax=b$ using a backward stable solver such as one based on the $QR$ factorization (or on an $LU$ factorization with partial pivoting provided there is no growth) yields $x_{\text{backward-stable}}$ for which \begin{equation} \left\Vert x_{\text{backward-stable}}-x\right\Vert \leq O(\kappa(A)\epsilon_{\text{machine}})\left\Vert x\right\Vert \;.\label{eq:backward-stable-accuracy} \end{equation} The bound (\ref{eq:loose-bound}) is correct, but it is just an upper bound on the error, and it turns out that it is loose by a factor of $\kappa(A)$. It appears that this easy-to-derive but loose bound gave the matrix inverse its bad reputation. In fact, $x_{V}$ satisfies an accuracy bound just like (\ref{eq:backward-stable-accuracy}). \section{\label{sec:Tightening-the-Bound}Tightening the Bound} The bound (\ref{eq:loose-bound}) is loose because of a single term, $\left\Vert \Gamma\right\Vert \left\Vert A\right\Vert $, which we used to bound the norm of $\Gamma A$. The other term in the bound, $O(\kappa(A)\epsilon_{\text{machine}})\left\Vert x\right\Vert $, is tight. The key insight is that rows of $\Gamma=V-A^{-1}$ tend to lie mostly in the directions of left singular vectors of $A$ that are associated with small singular values. The smaller the singular value of $A$, the stronger the influence of the corresponding singular vector (or singular subspace) on the rows of $\Gamma$. Therefore, the norm of the product of $\Gamma$ and $A$ is much smaller than the product of the norms; $A$ shrinks the strong directions of the error matrix $\Gamma$. This explains why the norm of $\Gamma A$ is small. This relationship between the singular vectors of $A$ and $\Gamma$ depends on a backward stability criterion on $V$, which we define and analyze below. Suppose that we use a backward stable solver to compute the rows of $V$ one by one by solving $v_{i}A=e_{i}$ where $e_{i}$ is row $i$ of $I$. Each computed row satisfies \[ v_{i}\left(A+\Xi_{i}\right)=e_{i} \] with $\|\Xi_{i}\|/\|A\|=O(\epsilon_{\text{machine}})$. Rearranging the equation, we obtain \[ VA-I=\left[\begin{array}{c} -v_{1}\Xi_{1}\\ \vdots\\ -v_{n}\Xi_{n} \end{array}\right]\;, \] so $\|VA-I\|=O(\|V\|\|A\|\epsilon_{\text{machine}})=O(\kappa(A)\epsilon_{\text{machine}})$. For a componentwise version of this bound and related bounds for other methods of computing $V$, see~\cite[section 14.3]{Higham}. This is the key to the conditional accuracy of $x_{V}$. Since $\Gamma A=(V-A^{-1})A=VA-I$, the norm of $\Gamma A$ is $O(\kappa(A)\epsilon_{\text{machine}})$. We therefore have the following theorem. \begin{thm} \label{thm:inv-accurate}Let $Ax=b$ be a linear system with a coefficient matrix that satisfies $\kappa(A)\epsilon_{\text{machine}}=O(1)$. Assume that $V$ is an approximate inverse of $A$ that satisfies $\|VA-I\|=O(\kappa(A)\epsilon_{\text{machine}})$. Then the floating-point product $x_{V}$ of $V$ and $b$ satisfies \[ \frac{\left\Vert x_{V}-x\right\Vert }{\left\Vert x\right\Vert }=O\left(\kappa(A)\epsilon_{\text{machine}}\right)\;. \] \end{thm} The essence of this analysis appears in Wilkinson's 1963 monograph~\cite[pp. 128--129]{Wilkinson}. \begin{comment} The reference to Wilkinson has already been given above. \end{comment} Wilkinson did not account for the rounding errors in the multiplication $Vb$, which are not asymptotically significant, but otherwise his analysis is complete and correct. \section{\label{sec:Left-and-Right}Left and Right Inverses} In this article, we multiply $b$ by the inverse from the left to solve $Ax=b$. This implies that the approximate inverse $V$ should be a good left inverse. Indeed, we have seen that a $V$ with a small left residual $VA-I$ guarantees a conditionally accurate solution $x_{V}$. Whether $V$ is also a good right inverse, in the sense that $AV-I$ is small, is irrelevant for solving $Ax=b$. If we were trying to solve $x^{T}A=b^{T}$, we would need a good right inverse. Wilkinson noted that if rows of $V$ are computed using $LU$ with partial pivoting, then $V$ is usually both a good left inverse and a good right inverse, but not always~\cite[page~113]{Wilkinson}. Du~Croz and Higham show matrices for which this is not the case, but they also note that such matrices are the exception rather than the rule~\cite{DuCrozHigham}. Other inversion methods tend to produce a matrix that is either a left inverse or a right inverse but not both. A good example is Newton's method. If one iterates with $V^{(t)}=(2I-V^{(t-1)}A)V^{(t-1)}$ then $V^{(t)}$ converges to a left inverse. If one iterates with $V^{(t)}=V^{(t-1)}(2I-AV^{(t-1)})$ then $V^{(t)}$ converges to a right inverse. Strassen's inversion formula~\cite{BaileyFergusonStrassenInv,Strassen69} sometimes produces an inverse that is neither a good left inverse nor a good right inverse~\cite[Section~26.3.2]{Higham}. \section{\label{sec:backward-stability-of-xv}Multiplication by the Inverse is (Sometimes) Backward Stable} The next section presents a simple example in which the computed solution $x_{V}$ is conditionally accurate but not backward stable. In this section we show that under certain conditions, the solution is also backward stable. The analysis also clarifies in what ways backward stability can be lost. Suppose that we use a $V$ that is a good right inverse, $\|AV-I\|=O(\kappa(A)\epsilon_{\text{machine}})$. We can produce such a $V$ by solving for its columns using a backward-stable solver. We have \begin{eqnarray*} Ax_{V}-b & = & A(V+\Delta)b-b\\ & = & \left(AV-I\right)b+A\Delta b \end{eqnarray*} for some $\Delta$ with $\|\Delta\|/\|V\|=O(\epsilon_{\text{machine}})$. Here too, the $\Delta$ term does not influence the asymptotic upper bound. The assumption that $V$ is a good right inverse bounds the other term, \begin{eqnarray} \left\Vert Ax_{V}-b\right\Vert & \leq & \left\Vert AV-I\right\Vert \left\Vert b\right\Vert +\left\Vert A\right\Vert \left\Vert \Delta\right\Vert \left\Vert b\right\Vert \nonumber \\ & \leq & O(\kappa(A)\epsilon_{\text{machine}})\left\Vert b\right\Vert +O(\epsilon_{\text{machine}})\left\Vert A\right\Vert \left\Vert V\right\Vert \left\Vert b\right\Vert \nonumber \\ & = & O(\kappa(A)\epsilon_{\text{machine}})\left\Vert b\right\Vert \;.\label{eq:loose-accuracy-bound-pre-1} \end{eqnarray} The relative backward error is given by the expression $\|Ax_{V}-b\|/(\|A\|\|x_{V}\|+\|b\|)$~\cite{RigalGaches}. Filling in the bound on the norm of the residual, we obtain \begin{eqnarray*} \frac{\left\Vert Ax_{V}-b\right\Vert }{\|A\|\|x_{V}\|+\|b\|} & \leq & \frac{\left\Vert Ax_{V}-b\right\Vert }{\|A\|\|x_{V}\|}\\ & = & O\left(\frac{\|A\|\|A^{-1}\|\epsilon_{\text{machine}}\|b\|}{\|A\|\|x_{V}\|}\right)\\ & = & O\left(\frac{\|A^{-1}\|\|b\|}{\|x_{V}\|}\epsilon_{\text{machine}}\right)\;. \end{eqnarray*} If we assume that $V$ is a reasonable enough left inverse so that at least $\|x_{V}\|$ is close to $\|x\|$ (that is, if the forward error is $O(1)$), then a solution $x$ that has norm close to $\|A^{-1}\|\|b\|$ guarantees backward stability to within $O(\epsilon_{\text{machine}})$. Let $A=L\Sigma R^{*}$ be the SVD of $A$, so \begin{eqnarray*} x & = & A^{-1}b\\ & = & R\Sigma^{-1}L^{*}b\\ & = & \sum_{i}\frac{L_{i}^{*}b}{\sigma_{i}}R_{i}\;, \end{eqnarray*} where $L_{i}$ and $R_{i}$ are the left and right singular vectors of $A$. If $L_{n}^{*}b=O(\|b\|)$, then $\|x\|\geq\left|L_{n}^{*}b\right|/\sigma_{n}=O(\|A^{-1}\|\|b\|)$ and $x_{V}$ is backward stable. If the projection of $b$ on $L_{n}$ is not large but the projection on, say, $L_{n-1}$ is large and $\sigma_{n-1}$ is close to $\sigma_{n}$, the solution is still backward stable, and so on. Perhaps more importantly, we have now identified the ways in which $x_{V}$ can fail to be backward stable: \begin{enumerate} \item $V$ is not a good right inverse, or \item $V$ is such a poor left inverse that $\|x_{V}\|$ is much smaller than $\|x\|$, or \item the projection of $b$ on the left singular vectors of $A$ associated with small singular values is small. \end{enumerate} The next section shows an example that satisfies the last condition. \section{\label{sec:Numerical-Examples}Numerical Examples} Let us demonstrate the theory with a small numerical example. We set $n=256$, $\sigma_{1}=10^{4}$ and $\sigma_{n}=10^{-4}$, generate a random matrix $A$ with $\kappa(A)=10^{8}$, and generate its inverse. The matrix and the inverse are produced by matrix multiplications, and each multiplication has at least one unitary factor, so both are accurate to within a relative error of about $\epsilon_{\text{machine}}$. We also compute an approximate inverse $V$ using \noun{Matlab}'s \texttt{inv} function. \begin{lyxcode} {[}L,dummy,R{]}~=~svd(randn(n));~ svalues~=~logspace(log10(sigma\_1),~log10(sigma\_n),~n); S~=~diag(svalues); invS~=~diag(svalues.\textasciicircum{}-1); A~=~L~{*}~S~{*}~R'; AccurateInv~=~R~{*}~invS~{*}~L'; V~=~inv(A); \end{lyxcode} The approximate inverse $V$ is only conditionally accurate, as predicted by (\ref{eq:conditionally-accurate-V}), but its use as a left inverse leads to a conditionally small residual. \begin{lyxcode} Gamma~=~V~-~AccurateInv; norm(Gamma)~/~norm(AccurateInv) ~~ans~=~3.4891e-09 norm(V~{*}~A~-~eye(n)) ~~ans~=~1.6976e-08 \end{lyxcode} We now generate a random right hand-side $b$ and use the inverse to solve $Ax=b$. The result is backward stable to within a relative error of $\epsilon_{\text{machine}}$. \begin{lyxcode} b~=~randn(n,~1); x~=~R~{*}~(invS~{*}~(L'~{*}~b)); xv~=~V~{*}~b; norm(A~{*}~xv~-~b)~/~(norm(A)~{*}~norm(xv)~+~norm(b)) ~~ans~=~8.8078e-16~ \end{lyxcode} Obviously, the solution should be conditionally accurate, and it is. \begin{lyxcode} norm(xv~-~x)~/~norm(x) ~~ans~=~3.102e-09~ \end{lyxcode} We now perform a similar experiment, but with a random $x$, which leads to a right-hand side $b$ which is nearly orthogonal to the left singular vectors of $A$ that correspond to small singular values; now the solution is only conditionally backward stable. \begin{lyxcode} x~=~randn(n,~1); b~=~L~{*}~(S~{*}~(R'~{*}~x)); xv~=~V~{*}~b; norm(A~{*}~xv~-~b)~/~(norm(A)~{*}~norm(xv)~+~norm(b)) ~~ans~=~2.1352e-10 \end{lyxcode} Theorem~\ref{thm:inv-accurate} predicts that the solution should still be conditionally accurate. It is. \noun{Matlab}'s backslash operator, which is a linear solver based on Gaussian elimination with partial pivoting, produces a solution with a similar accuracy. \begin{lyxcode} norm((A\textbackslash{}b)~-~x)~/~norm(x) ~~ans~=~4.0801e-09 norm(xv~-~x)~/~norm(x) ~~ans~=~4.5699e-09 \end{lyxcode} The magic is in the special structure of rows of $\Gamma$. Figure~\ref{fig:proj-Gamma-on-sign-vectors} displays this structure graphically. We can see that a row of $\Gamma$ is almost orthogonal to the left singular vectors of $A$ associated with large singular values, and that the magnitude of the projections increases with decreasing singular values. If we produce an approximate inverse with the same magnitude of error as in \texttt{inv(A)} but with a random error matrix, it will not solve $Ax=b$ conditionally accurately. \begin{lyxcode} BadInv~=~AccurateInv~+~norm(Gamma)~{*}~randn(n); xv~=~BadInv~{*}~b; norm(A~{*}~xv~-~b)~/~(norm(A)~{*}~norm(xv)~+~norm(b))~\\ ~~ans~=~0.075727 norm(xv~-~x)~/~norm(x) ~~ans~=~0.83552 \end{lyxcode} \begin{figure} \begin{centering} \includegraphics[width=0.75\textwidth]{proj_Gamma_on_singvects} \par\end{centering} \caption{\label{fig:proj-Gamma-on-sign-vectors}The magnitude of the projections of three rows of $\Gamma$ (the first, last, and middle, but all rows produce similar plots) on the left singular vectors of $A$, as a function of the corresponding singular values of $A$.} \end{figure} \section{\label{sec:Closing-Remarks}Closing Remarks} Solving a linear system of equations $Ax=b$ using a computed inverse $V$ produces a conditionally accurate solution, subject to an easy to satisfy condition on the computation of $V$. Using Gaussian elimination with partial pivoting or a $QR$ factorization produces a solution with errors that have the same order of magnitude as those produced by $V$. If the right-hand side $b$ does not have any special relationship to the left singular subspaces of $A$, then the solution produced by $V$ is also backward stable (under a slightly different technical condition on $V$), and hence as good as a solution produced by GEPP or $QR$. As far as we know, this result is new. If $b$ is close to orthogonal to the left singular subspaces of $A$ corresponding to small singular values, then the solution produced by $V$ is conditionally accurate, but usually not backward stable. Whether this is a significant defect or not depends on the application. In most applications, it is not a serious problem. One difficulty with a conditionally-accurate solution that is not backward stable is that it does not come with a certificate of conditional accuracy. We normally take a small backward error to be such a certificate. There might be applications that require a backward stable solution rather than an accurate one. Strangely, this is exactly the case with the computation of $V$ itself; the analysis in this paper relies on rows being computed in a backward-stable way, not on their forward accuracy. We are not aware of other cases where this is important. \bibliographystyle{plain}

https://arxiv.org/abs/1809.10263

Counting Shellings of Complete Bipartite Graphs and Trees

A shelling of a graph, viewed as an abstract simplicial complex that is pure of dimension 1, is an ordering of its edges such that every edge is adjacent to some other edges appeared previously. In this paper, we focus on complete bipartite graphs and trees. For complete bipartite graphs, we obtain an exact formula for their shelling numbers. And for trees, we relate their shelling numbers to linear extensions of tree posets and bound shelling numbers using vertex degrees and diameter.

\section{Introduction} In combinatorial topology, shelling of a simplicial complex is a very useful and important notion that has been well-studied. \begin{definition}\label{def:shellingcomplex} An (abstract) simplicial complex $\Delta$ is called \textit{pure} if all of its maximal simplicies have the same dimension. Given a finite (or countably infinite) simplicial complex $\Delta$ that is pure of dimension $d$, a \textit{shelling} is a total ordering of its maximal simplicies $C_1,C_2,\ldots$ such that for every $k>1$, $C_k\cap\left(\bigcup_{i=1}^{k-1}C_i\right)$ is pure of dimension $d-1$. A simplicial complex that admits a shelling is called \textit{shellable}. \end{definition} Shellable complexes enjoy many strong algebraic and topological properties. For example, a shellable complex is homotopy equivalent to a wedge sum of spheres, thus not allowing torsion in its homology. The study of shellability in its combinatorial aspects has turned out to be very fruitful as well. The arguably earliest notable result that polytopes are shellable is due to Brugesser and Mani (Section 8 of \cite{ziegler2012lectures}). Later on, Bjorner and Wachs developed theories on lexicographic shellability (Section 12 of \cite{kozlov2007combinatorial}). In particular, shellable posets, which are posets whose order complexes are shellable, are studied and powerful notions such as $EL$-shellability and $CL$-shellability are invented. In a recent work, testing shellability is proved to be NP-complete \cite{goaoc2017shellability}. As there is rich literature on shellability, little work has been done on counting the number of shellings for a specific simplicial complex. It is generally believed that if a simplicial complex is shellable, then it usually admits a lot of shellings, but no precise arguments are given. In this paper, we investigate the problem of counting shellings, aiming to start a new line of research. We restrict our attentions to finite simplicial complexes that are pure of dimension 1, namely, undirected graphs, where interesting combinatorial arguments are already taking place. Let's first reformulate Definition~\ref{def:shellingcomplex} in the language of graph theory. \begin{definition}[Graph Shelling]\label{def:shellinggraph} Given an undirected graph $G=(V,E)$, where $V$ is the vertex set of $G$ and $E$ is the edge set of $G$, a \textit{shelling} of $G$ is a total ordering of the edge set $\sigma\in\S_E$, where $\S$ stands for symmetric group, such that $\sigma(1),\ldots,\sigma(k)$ form a connected subgraph of $G$ for all $k=1,\ldots,|E|$. \end{definition} We will adopt the following notation throughout the paper. \begin{definition} For a graph $G$, let $F(G)$ denote the number of shellings of $G$. \end{definition} Clearly, a graph admits a shelling if and only if it is connected, which is equivalent to $F(G)>0$. A few results are already known. \begin{theorem}[\cite{MO297411}] Let $K_n$ be the complete graph on $n$ vertices. Then $$F(K_n)=\frac{2^{n-2}}{C_{n-1}}\binom{n}{2}!$$ where $C_{n-1}=\binom{2n-2}{n-1}/n$ is the $(n-1)^{th}$ Catalan number. \end{theorem} As an overview for the paper, in Section~\ref{sec:cbg}, we will give an explicit formula for the number of shellings of complete bipartite graphs, resolving a MathOverflow question \cite{MO297385}; in Section~\ref{sec:trees}, we will provide methods to compute the number of shellings of trees and obtain some upper and lower bounds for them. \section{Complete Bipartite Graphs}\label{sec:cbg} Denote $K_{m,n}$ as the complete bipartite graph with part sizes $m$ and $n$. The following is our main theorem. \begin{theorem}\label{thm:bipartite} $$F(K_{m,n})=\frac{m!n!(mn)!}{(m+n-1)!}.$$ \end{theorem} The formula in Theorem~\ref{thm:bipartite} is conjectured in the MathOverflow post \cite{MO297385}. Partial progress has been made. In particular, Lemma~\ref{lem:bipartiteStanley}, given by Richard Stanley, serves as an important tool for our computation. \begin{lemma}\label{lem:bipartiteStanley} $F(K_{m,n})$ is equal to the following expression: $$m!n!(mn-1)!\sum_{\alpha}\frac{b_{1}b_{2}\cdots b_{m+n-2}}{b_{m+n-2}(b_{m+n-2}+b_{m+n-3})\cdots (b_{m+n-2}+b_{m+n-3}+\ldots+b_1)},$$ where the sum is over all sequences $\alpha = (a_1,a_2,\ldots,a_{m+n-2})$ of $(m-1)$ 0's and $(n-1)$ 1's, and $$b_{i} = 1 + |\{1\leq j\leq i: a_j \neq a_i\}|.$$ \end{lemma} \begin{proof} Let $\sigma$ be a shelling of $K_{m,n}$. In each part of $K_{m,n}$, consider the order of the appearance of the vertices. Here, we say that vertex $v$ appears in $\sigma$ at time $t$ if $t$ is the first index such that $v\in \sigma(t)$. There are $m!$ ways to choose such order in the part of size $m$ and $n!$ ways in the part of size $n$. Fix the order of vertex appearance in each part to be $(u_0,u_1,\ldots,u_{m-1}), (v_0,v_1,\ldots,v_{n-1})$, respectively. Consider a fixed order of appearance of all $(m+n)$ vertices $w = w_{-1}w_0\ldots w_{m+n-2}.$ Note that $\sigma(1)$ must be the edge $e_0=(u_0,v_0)$, so $\{w_{-1},w_0\} = \{u_0,v_0\}.$ For $1\leq i\leq m+n-2$, define $$a_{i} = \begin{cases} 0, & \text{ if } w_{i} = u_j \text{ for some }j, \\ 1, & \text{ if } w_{i} = v_k \text{ for some }k, \end{cases}$$ and $$b_i = 1 + |\{1\leq j\leq i: a_j \neq a_i\}|.$$ Now, for each $w_i$ ($i\geq 1$), consider the first edge $e_i$ incident to $w_i$ in $\sigma$. This edge must be of the form $(w_i,w_j)$ where $j<i$ and $w_i, w_j$ are in different parts of $K_{m,n}$. There are $b_{i}$ choices for this edge. Thus, there are $b_1b_2\cdots b_{m+n-2}$ ways to choose $e_1,e_2,\ldots,e_{m+n-2}.$ We further fix the edges $e_0,e_1,\ldots,e_{m+n-2}$. Note that the rest of the $b_{m+n-2}$ edges incident to $w_{m+n-2}$ must appear after $e_{m+n-2}$ in $\sigma$, so there are $(b_{m+n-2}-1)!$ ways to arrange these edges. After making this arrangement, the edges which are incident to $w_{m+n-3}$ and not yet arranged must appear after $e_{m+n-3}$, so there are $$(b_{m+n-2}+1)(b_{m+n-2}+2)\cdots (b_{m+n-2}+b_{m+n-3}+1) = \frac{(b_{m+n-2}+b_{m+n-3}+1)!}{b_{m+n-2}!}$$ ways to arrange them (since there are already $b_{m+n-2}$ edges arranged after $e_{m+n-3}$). Similarly, for each $i$, after making the arrangement of all edges incident to vertices appearing after $w_i$, there are $$\frac{(b_{m+n-2}+b_{m+n-3}+\ldots + b_i + 1)!}{(b_{m+n-2}+b_{m+n-3}+\ldots + b_{i+1})!}$$ ways to arrange all the edges which are incident to $w_i$ and not yet arranged. Therefore, after fixing $e_0,e_1,\ldots,e_{m+n-2}$, the number of shellings is $$\prod_{i=1}^{m+n-2}\frac{(b_{m+n-2}+\ldots + b_i + 1)!}{(b_{m+n-2}+\ldots + b_{i+1})!} = \frac{(mn-1)!}{b_{m+n-2}(b_{m+n-2}+b_{m+n-3})\cdots (b_{m+n-2}+\ldots+b_1)}.$$ Combining all discussions above, we obtain Lemma~\ref{lem:bipartiteStanley}. \end{proof} We first prove a few lemmas which are essential to Theorem~\ref{thm:bipartite}. These lemmas involve binomial coefficients whose entries are not necessarily integers. For this reason, in the rest of this section, we will use the generalized binomial coefficient $$\binom{x}{y} = \frac{\Gamma(x+1)}{\Gamma(y+1)\Gamma(x-y+1)},$$ where $\Gamma$ is the Gamma function that extends the factorial function. In particular, if $y$ is a positive integer, $$\binom{x}{y} = \frac{x(x-1)\cdots (x-y+1)}{y!}.$$ \begin{lemma}\label{lem:story} For positive integers $x,y$ and positive real numbers $z,w$ such that $w-z\geq x$ is a postive integer, $$\sum_{j=x}^{w-z}\binom{j}{y}\binom{w-j}{z}=\sum_{i=\max\{0,x+y+z-w\}}^y \binom{x}{i}\binom{w-x+1}{z+y-i+1}.$$ \end{lemma} \begin{proof} We first prove this lemma assuming that $z,w$ are both integers. Consider the following problem: we want to arrange $(y+z+1)$ letter A's in $(w+1)$ positions, such that each position has at most one A and there are at most $y$ A's in the first $x$ positions. The number of such arrangements is $$\sum_{i=0}^y \binom{x}{i}\binom{w-x+1}{z+y-i+1} = \sum_{i=\max\{0,x+y+z-w\}}^y \binom{x}{i}\binom{w-x+1}{z+y-i+1}$$ by considering the number of A's in the first $x$ positions. On the other hand, consider the position of the $(y+1)^{th}$ A. It must be at some position $p>x$. For a fixed $p$, there are $\binom{p-1}{y}$ ways to arrange the first $y$ A's and $\binom{w-p+1}{z}$ ways to arrange the last $z$ A's, so the total number of such arrangements is $$\sum_{p=x+1}^{w-z+1}\binom{p-1}{y}\binom{w-p+1}{z} = \sum_{j=x}^{w-z}\binom{j}{y}\binom{w-j}{z}.$$ Thus, Lemma~\ref{lem:story} follows under additional assumption. For the general case, we fix $z' = w-z \in \mathbb{N}$. Lemma~\ref{lem:story} is equivalent to \begin{equation}\label{eq:story} \sum_{j=x}^{z'}\binom{j}{y}\binom{w-j}{z'-j}=\sum_{i=\max\{0,x+y-z'\}}^y \binom{x}{i}\binom{w-x+1}{z'-x-y+i}. \end{equation} Both sides of Equation~\eqref{eq:story} are polynomials in $w$ of degree at most $z'$. From our previous discussion, every positive integer greater than $z'$ is a root of ~\eqref{eq:story}. Thus, the two sides of ~\eqref{eq:story} agree as polynomials in $w$ and the proof is complete. \end{proof} Lemma~\ref{lem:story} serves to prove the following lemma, which will be crucial in calculating the sum in Lemma~\ref{lem:bipartiteStanley}. \begin{lemma}\label{lem:binomialComputation} For positive integers $k<n$ and $s<m+n-k-1$, \begin{equation}\label{eq:lemma} \begin{split} &\sum_{t=s+1}^{m+n-k-1}(t-n+k+1)(t+2)(t+3)\cdots(t+k)\binom{\frac{mn}{n-k}+n-k-t-2}{\frac{mk}{n-k}-1} \\ =&\frac{m}{m+n-k}(s+2)(s+3)\cdots(s+k+1)\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}}, \end{split} \end{equation} where general binomial coefficients are used. \end{lemma} \begin{proof} First, note that $$(t-n+k+1)(t+2)(t+3)\cdots (t+k) = k!\bigg[\binom{t+k}{k} +\frac{k-n}{k}\binom{t+k}{k-1}\bigg].$$ We shall split the sum in the left hand side of $(\ref{eq:lemma})$ based on the equation above. Applying Lemma~\ref{lem:story} with replacements $x=s+k+1$, $y=k$, $z=\frac{mk}{n-k}-1$, $w=\frac{mn}{n-k}+n-2$ (notice that $w-z = m+n-1$ is a positive integer), we obtain $$\sum_{j = s+k+1}^{m+n-1}\binom{j}{k}\binom{\frac{mn}{n-k}+n-2-j}{\frac{mk}{n-k}-1} = \sum_{i=i_0}^{k} \binom{s+k+1}{i}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-i},$$ where $i_0 = \max\{0, s+2k+2-m-n\}.$ Writing $t=j-k$, we have $$\sum_{t = s+1}^{m+n-k-1}\binom{t+k}{k}\binom{\frac{mn}{n-k}+n-k-t-2}{\frac{mk}{n-k}-1} = \sum_{i=i_0}^{k} \binom{s+k+1}{i}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-i}.$$ Similarly, replacing $x=s+k+1$, $y=k-1$, $z=\frac{mk}{n-k}-1$, $w=\frac{mn}{n-k}+n-2$ in Lemma~\ref{lem:story}, \begin{align*} \sum_{t=s+1}^{m+n-k-1}\binom{t+k}{k-1}\binom{\frac{mn}{n-k}+n-k-t-2}{\frac{mk}{n-k}-1} &= \sum_{i=i_1}^{k-1} \binom{s+k+1}{i}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-i-1}\\ &=\sum_{i=i_1+1}^{k} \binom{s+k+1}{i-1}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-i}, \end{align*} where $i_1 = \max\{0, s+2k+1-m-n\}.$ Therefore, the left hand side of (\ref{eq:lemma}) \begin{equation*} \begin{split} &\frac{1}{k!}\sum_{t=s+1}^{m+n-k-1}(t-n+k+1)(t+2)(t+3)\cdots(t+k)\binom{\frac{mn}{n-k}+n-k-t-2}{\frac{mk}{n-k}-1} \\ =& \sum_{i=i_0}^{k} \binom{s+k+1}{i}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-i}+\frac{k-n}{k}\sum_{i=i_1+1}^{k} \binom{s+k+1}{i-1}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-i}. \end{split} \end{equation*} We claim that the following identity (\ref{eq:inductionLemma}) holds for all $i_0\leq \ell\leq k$. \begin{equation}\label{eq:inductionLemma} \begin{split} & \sum_{i=i_0}^{\ell} \binom{s+k+1}{i}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-i}+\frac{k-n}{k}\sum_{i=i_1+1}^{\ell} \binom{s+k+1}{i-1}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-i} \\ =&\frac{\frac{mk}{n-k}+k-\ell}{\frac{mk}{n-k}+k}\binom{s+k+1}{\ell}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-\ell}. \end{split} \end{equation} There are two cases: $i_0=0$ and $i_0>0$. \noindent\textbf{Case 1.} $i_0 = i_1 = 0.$ In this case, the left hand side of (\ref{eq:inductionLemma}) is \begin{equation*} \binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k}+\sum_{i=1}^\ell \bigg[\binom{s+k+1}{i}+\frac{k-n}{k}\binom{s+k+1}{i-1}\bigg]\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-i}. \end{equation*} Induct on $\ell$. When $\ell=0$, both sides of (\ref{eq:inductionLemma}) are equal to $\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k}.$ Assume that (\ref{eq:inductionLemma}) holds for $\ell-1$ and consider $\ell$ case. Then, the formula above becomes \begin{align*} & \bigg[\binom{s+k+1}{\ell}+\frac{k-n}{k}\binom{s+k+1}{\ell-1}\bigg]\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-\ell} + \\ &\qquad\frac{\frac{mk}{n-k}+k-\ell+1}{\frac{mk}{n-k}+k}\binom{s+k+1}{\ell-1}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-\ell+1} \\ =&\bigg[\binom{s+k+1}{\ell}+\frac{k-n}{k}\cdot\frac{\ell}{s+k+2-\ell}\binom{s+k+1}{\ell}\bigg]\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-\ell} + \\ &\ \frac{\frac{mk}{n-k}+k-\ell+1}{\frac{mk}{n-k}+k}\frac{\ell}{s+k+2-\ell}\binom{s+k+1}{\ell}\frac{m+n+\ell-2k-s-2}{\frac{mk}{n-k}+k-\ell+1}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-\ell} \\ =& \left(1+\frac{(k-n)\ell}{k(s+k+2-\ell)}+ \frac{(m+n+\ell-2k-s-2)\ell}{(\frac{mk}{n-k}+k)(s+k+2-\ell)}\right)\binom{s+k+1}{\ell}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-\ell} \\ =& \frac{\frac{mk}{n-k}+k-\ell}{\frac{mk}{n-k}+k}\binom{s+k+1}{\ell}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-\ell}. \end{align*} Thus, (\ref{eq:inductionLemma}) follows by induction. \noindent\textbf{Case 2.} $i_0 = s+2k+2-m-n > 0$ and $i_1 = i_0 - 1.$ We can simplify the left hand side of (\ref{eq:inductionLemma}) as \begin{equation*} \sum_{i=i_0}^{\ell} \bigg[\binom{s+k+1}{i}+\frac{k-n}{k}\binom{s+k+1}{i-1}\bigg]\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-i}. \end{equation*} Induct on $\ell$. When $\ell = i_0$, \begin{align*} &\bigg[\binom{s+k+1}{i_0}+\frac{k-n}{k}\cdot \frac{i_0}{s+k+2-i_0}\binom{s+k+1}{i_0}\bigg]\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-i_0} \\ =& \frac{\frac{mk}{n-k}+k-i_0}{\frac{mk}{n-k}+k}\binom{s+k+1}{i_0}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-i_0}, \end{align*} as desired. The inductive step $(\ell-1) \Rightarrow \ell$ holds by the same calculation as the previous case $i_0=0$. \begin{comment} \begin{align*} (\ref{eq:simplification2}) =& \bigg[\binom{s+k+1}{l}+\frac{k-n}{k}\binom{s+k+1}{l-1}\bigg]\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-l} + \\ &\qquad\frac{\frac{mk}{n-k}+k-l+1}{\frac{mk}{n-k}+k}\binom{s+k+1}{l-1}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-l+1} \\ =& \frac{\frac{mk}{n-k}+k-l}{\frac{mk}{n-k}+k}\binom{s+k+1}{l}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-l}. \end{align*} \end{comment} Thus, the claim follows by induction. In particular, when $\ell=k$, (\ref{eq:inductionLemma}) becomes \begin{equation*} \begin{split} & \sum_{i=i_0}^{k} \binom{s+k+1}{i}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-i}+\frac{k-n}{k}\sum_{i=i_1+1}^{k} \binom{s+k+1}{i-1}\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}+k-i} \\ =&\frac{1}{k!}\cdot\frac{m}{m+n-k}(s+2)(s+3)\cdots(s+k+1)\binom{\frac{mn}{n-k}+n-k-s-2}{\frac{mk}{n-k}}. \end{split} \end{equation*} Therefore, the proof of this lemma is complete. \end{proof} Now we are ready to prove Theorem~\ref{thm:bipartite}. \begin{proof}[Proof of Theorem~\ref{thm:bipartite}] According to Lemma~\ref{lem:bipartiteStanley}, it suffices to show that \begin{equation*} (mn-1)!\sum_{\alpha}\frac{b_{1}b_{2}\cdots b_{m+n-2}}{b_{m+n-2}(b_{m+n-2}+b_{m+n-3})\cdots (b_{m+n-2}+b_{m+n-3}+\ldots+b_1)} = \frac{(mn)!}{(m+n-1)!}. \end{equation*} where the sum is over all sequences $\alpha=(a_1,a_2,\ldots,a_{m+n-2})$ consisting of $(m-1)$ 0's and $(n-1)$ 1's. Suppose $a_{r_1}=a_{r_2}=\ldots=a_{r_{n-1}} = 1$ where $1\leq r_1<r_2<\ldots<r_{n-1}\leq m+n-2$. Denote $r_0 = 0$. Then for $k=1,2,\ldots,n-1$, \begin{comment} \begin{align*} b_1 = b_2 = \ldots = b_{r_1-1} = 1,&\ b_{r_1} = r_1,\\ \vdots \qquad & \\ b_{r_{k-1}+1} = \ldots = b_{r_k-1} = k,&\ b_{r_k} = r_{k}-k+1,\\ \vdots \qquad & \\ b_{r_{n-1}+1} = \ldots = b_{m+n-2} = n. \end{align*} \end{comment} $$b_{r_{k-1}+1} = b_{r_{k-1}+2} = \cdots = b_{r_k-1} = k, b_{r_k} = r_{k}-k+1$$ and $$b_{r_{n-1}+1} = \cdots = b_{m+n-2} = n.$$ Therefore, \begin{equation*} \prod_{i=1}^{m+n-2}b_i = n^{m+n-2-r_{n-1}}\prod_{j=1}^{n-1} (r_j -j+1) j^{r_{j}-r_{j-1}-1}. \end{equation*} For $1\leq i\leq m+n-2$, write $c_i = b_{m+n-2}+\ldots + b_{i}$, then $$c_{m+n-2} = n, c_{m+n-3} = 2n, \ldots, c_{r_{n-1}+1} = mn+n(n-2-r_{n-1})$$ $$\implies c_{m+n-2}c_{m+n-3}\cdots c_{r_{n-1}+1} = n^{m+n-2-r_{n-1}}\Gamma(m+n-1-r_{n-1})$$ \begin{comment} $$c_{r_{n-1}} = mn +(n-1)(n-2-r_{n-1}),\ldots, c_{r_{n-2}+1} = mn+(n-1)(n-3-r_{n-2}),$$ \begin{align*} \implies c_{r_{n-1}}\cdots c_{r_{n-2}+1} &= (n-1)^{r_{n-1}-r_{n-2}}\prod_{i=r_{n-2}+1}^{r_{n-1}} (\frac{mn}{n-1}+n-2-i) \\ &= (n-1)^{r_{n-1}-r_{n-2}}\frac{\Gamma(\frac{mn}{n-1}+n-2-r_{n-2})}{\Gamma(\frac{mn}{n-1}+n-2-r_{n-1})}. \end{align*} \end{comment} For $k=1,2,\ldots,n-1,$ we have $$c_{r_{k}} = mn+k(k-1-r_k),\ldots, c_{r_{k-1}+1} = mn+k(k-2-r_{k-1})$$ \begin{align*} \implies c_{r_{k}}\cdots c_{r_{k-1}+1} &= k^{r_{k}-r_{k-1}}\prod_{i=r_{k-1}+1}^{r_{k}} (\frac{mn}{k}+k-1-i) \\ &= k^{r_{k}-r_{k-1}}\frac{\Gamma(\frac{mn}{k}+k-1-r_{k-1})}{\Gamma(\frac{mn}{k}+k-1-r_{k})}. \end{align*} \begin{comment} $$c_{r_1} = mn-r_1,\ldots, c_1 = mn-1.$$ $$\implies c_{r_1}\cdots c_1 = \frac{\Gamma(mn)}{\Gamma(mn-r_1)}.$$ \end{comment} Denote $r_n = m+n-2$, we have \begin{equation*} \begin{split} \prod_{i=1}^{m+n-2}c_i &= \prod_{j=1}^{n} j^{r_{j}-r_{j-1}}\frac{\Gamma(\frac{mn}{j}+j-1-r_{j-1})}{\Gamma(\frac{mn}{j}+j-1-r_{j})} \\ &=(mn-1)!\bigg(\prod_{j=1}^{n} j^{r_{j}-r_{j-1}}\bigg)\bigg(\prod_{k=1}^{n-1}\frac{\Gamma(\frac{mn}{k+1}+k-r_{k})}{\Gamma(\frac{mn}{k}+k-1-r_{k})}\bigg) \end{split} \end{equation*} Comparing the product of $b_i$'s and $c_i$'s, we obtain $$(mn-1)!\prod_{i=1}^{m+n-2}\frac{b_i}{c_i} = \frac{1}{(n-1)!}\prod_{k=1}^{n-1}(r_k-k+1)\frac{\Gamma(\frac{mn}{k}+k-1-r_{k})}{\Gamma(\frac{mn}{k+1}+k-r_{k})}.$$ \begin{align*} \implies (mn-1)!\sum_{\alpha}\prod_{i=1}^{m+n-2}\frac{b_i}{c_i} =& \frac{1}{(n-1)!}\sum_{1\leq r_1<\ldots<r_{n-1}\leq m+n-2}\prod_{j=1}^{n-1}(r_j-j+1)\frac{\Gamma(\frac{mn}{j}+j-1-r_{j})}{\Gamma(\frac{mn}{j+1}+j-r_{j})} \\ =& \frac{1}{(n-1)!}\sum_{1\leq r_1<\ldots<r_{n-1}\leq m+n-2}\prod_{j=1}^{n-1}R_j, \end{align*} where $R_j = (r_j-j+1)\frac{\Gamma(\frac{mn}{j}+j-1-r_{j})}{\Gamma(\frac{mn}{j+1}+j-r_{j})}.$ We claim that the sum \begin{equation}\label{eq:inductionTheorem} \begin{split} &\frac{1}{(n-1)!}\sum_{1\leq r_1<\ldots<r_{n-1}\leq m+n-2}\prod_{j=1}^{n-1}R_j \\ =&\frac{(m+k)!\Gamma(\frac{m(n-k)}{k})}{(m+n-1)!k!(n-k-1)!}\sum_{1\leq r_1<\ldots<r_{k}\leq m+k-1}(r_k-k+1)\frac{(r_k+n-k)!}{(r_k+1)!}\binom{\frac{mn}{k}+k-2-r_k}{\frac{m(n-k)}{k}-1}\prod_{j=1}^{k-1}R_j \end{split} \end{equation} for all $1\leq k\leq n-1$. To prove this claim, we reversely induct on $k$. When $k=n-1$, the right hand side of (\ref{eq:inductionTheorem}) \begin{align*} &\frac{\Gamma(\frac{m}{n-1})}{(n-1)!}\sum_{1\leq r_1<\ldots<r_{n-1}\leq m+n-2}(r_{n-1}-n+2)\frac{\Gamma(\frac{mn}{n-1}+n-2-r_{n-1})}{\Gamma(\frac{m}{n-1})\Gamma(m+n-1-r_{n-1})}\prod_{j=1}^{n-2}R_j \\ =& \frac{1}{(n-1)!}\sum_{1\leq r_1<\ldots<r_{n-1}\leq m+n-2}\prod_{j=1}^{n-1}R_j, \end{align*} as desired. Assume that the claim holds for $k+1$, then the left hand side of (\ref{eq:inductionTheorem}) becomes \begin{equation}\label{eq:inductionStep1} \frac{(m+k+1)!\Gamma(\frac{m(n-k-1)}{k+1})}{(m+n-1)!(k+1)!(n-k-2)!}\sum_{1\leq r_1<\ldots<r_{k+1}\leq m+k}(r_{k+1}-k)\frac{(r_{k+1}+n-k-1)!}{(r_{k+1}+1)!}\binom{\frac{mn}{k+1}+k-1-r_{k+1}}{\frac{m(n-k-1)}{k+1}-1}\prod_{j=1}^{k}R_j \end{equation} Setting $t=r_{k+1},s= r_{k}, k\rightarrow n-k-1$ in Lemma~\ref{lem:binomialComputation}, we have \begin{align*} &\sum_{1\leq r_1<\ldots<r_{k+1}\leq m+k}(r_{k+1}-k)\frac{(r_{k+1}+n-k-1)!}{(r_{k+1}+1)!}\binom{\frac{mn}{k+1}+k-1-r_{k+1}}{\frac{m(n-k-1)}{k+1}-1}\prod_{j=1}^{k}R_j \\ =& \sum_{1\leq r_1<\ldots<r_k\leq m+k-1}\bigg[\bigg(\prod_{j=1}^{k}R_j\bigg)\sum_{r_{k+1} = r_k+1}^{m+k}(r_{k+1}-k)\frac{(r_{k+1}+n-k-1)!}{(r_{k+1}+1)!}\binom{\frac{mn}{k+1}+k-1-r_{k+1}}{\frac{m(n-k-1)}{k+1}-1}\bigg] \\ =& \sum_{1\leq r_1<\ldots<r_k\leq m+k-1}\bigg[\bigg(\prod_{j=1}^{k}R_j\bigg)\frac{m}{m+k+1}(r_k+2)(r_k+3)\cdots(r_k+n-k)\binom{\frac{mn}{k+1}+k-r_k-1}{\frac{m(n-k-1)}{k+1}}\bigg] \\ =& \sum_{1\leq r_1<\ldots<r_k\leq m+k-1}\frac{m}{m+k+1}\cdot\frac{(r_k+n-k)!}{(r_k+1)!}\binom{\frac{mn}{k+1}+k-r_k-1}{\frac{m(n-k-1)}{k+1}}\prod_{j=1}^{k}R_j. \end{align*} Thus, \begin{align*} (\ref{eq:inductionStep1}) &= \frac{(m+k+1)!\Gamma(\frac{m(n-k-1)}{k+1})}{(m+n-1)!(k+1)!(n-k-2)!}\sum_{1\leq r_1<\ldots<r_k\leq m+k-1}\frac{m(r_k+n-k)!}{(m+k+1)(r_k+1)!}\binom{\frac{mn}{k+1}+k-r_k-1}{\frac{m(n-k-1)}{k+1}}\prod_{j=1}^{k}R_j \\ &=\frac{(m+k)!\Gamma(\frac{m(n-k-1)}{k+1})}{(m+n-1)!(k+1)!(n-k-2)!} \sum_{1\leq r_1<\ldots<r_k\leq m+k-1}\frac{m(r_k+n-k)!}{(r_k+1)!}\frac{\Gamma(\frac{mn}{k+1}+k-r_k)}{\Gamma(\frac{m(n-k-1)}{k+1}+1)\Gamma(m+k-r_k)}\prod_{j=1}^{k}R_j \\ &=\frac{(m+k)!}{(m+n-1)!k!(n-k-1)!}\cdot \\ &\qquad \sum_{1\leq r_1<\ldots<r_k\leq m+k-1}\frac{(r_k+n-k)!}{(r_k+1)!}\frac{\Gamma(\frac{mn}{k+1}+k-r_k)}{\Gamma(m+k-r_k)} (r_k-k+1)\frac{\Gamma(\frac{mn}{k}+k-1-r_{k})}{\Gamma(\frac{mn}{k+1}+k-r_{k})}\prod_{j=1}^{k-1}R_j\\ &=\frac{(m+k)!\Gamma(\frac{m(n-k)}{k})}{(m+n-1)!k!(n-k-1)!}\sum_{1\leq r_1<\ldots<r_{k}\leq m+k-1}(r_k-k+1)\frac{(r_k+n-k)!}{(r_k+1)!}\binom{\frac{mn}{k}+k-2-r_k}{\frac{m(n-k)}{k}-1}\prod_{j=1}^{k-1}R_j. \end{align*} and the claim follows by (reverse) induction. In particular, when $k=1$, $(\ref{eq:inductionTheorem})$ becomes \begin{equation*} \frac{(m+1)!(mn-m-1)!}{(m+n-1)!(n-2)!}\sum_{r_1=1}^m r_1\frac{(r_1+n-1)!}{(r_1+1)!}\binom{mn-r_1-1}{mn-m-1}. \end{equation*} Again, setting $t=r_1, s=0, k=n-1$ in Lemma~\ref{lem:binomialComputation}, $$\sum_{r_1=1}^m r_1\frac{(r_1+n-1)!}{(r_1+1)!}\binom{mn-r_1-1}{mn-m-1} = \frac{m}{m+1}\cdot n!\binom{mn-1}{mn-m}.$$ Therefore, \begin{align*} (\ref{eq:inductionTheorem}) &= \frac{(m+1)!(mn-m-1)!}{(m+n-1)!(n-2)!}\cdot\frac{m}{m+1}\cdot n!\binom{mn-1}{mn-m} \\ &= \frac{(mn)!}{(m+n-1)!}. \end{align*} Therefore, the proof of Theorem~\ref{thm:bipartite} is complete. \end{proof} \section{Trees}\label{sec:trees} \subsection{Tree Shelling Number Computation} \ Trees are one of the most fundamental type of graphs. However, unlike the complete bipartite graph case, there is no simple formula for tree shelling numbers. The goal of this section is to give a relatively easy method to compute the number of shellings of a tree. Throughout this section, let $T$ be a tree with $n$ vertices and $n-1$ edges. We first focus on computing the number of shellings of rooted trees, whose definition is given below. \begin{definition}\label{def:shellingOfRootedTree} Let $v$ be a vertex of $T$. The rooted tree induced by $T$ and rooted at $v$ is denoted as $T_v$. A \textit{shelling of rooted tree} $T_v$ is a shelling $\sigma$ of $T$ such that $\sigma(1)$ is an edge incident to $v$. \end{definition} The following definitions are used to efficiently describe structures in a (rooted) tree. \begin{definition}\label{def:parent} Let $T_v$ be a tree rooted at vertex $v$. We say a vertex $u$ is a \textit{parent} of vertex $w$ (and $w$ is a \textit{child} of $u$) if $(w,u)$ is an edge and $u$ lies closer to the root than $w$. A \textit{descending path} from $u$ to $w$ in the rooted tree $T_v$ is a structure $$u-v_1-v_2-\cdots -v_r-w$$ where each vertex is a parent of the subsequent vertex. We say $u$ is an \textit{ancestor} of $w$ (and $w$ is a \textit{descendant} of $u$) if there exists a descending path from $u$ to $w$. \end{definition} \begin{definition}\label{def:rootedSubtree} Let $u,v\in T.$ The (rooted)\textit{ subtree of $T_v$ rooted at $u$}, denoted as $T_v(u)$, is a subgraph of $T$ rooted at $u$ and induced by the set of vertices $$\{w\in T: w \text{ is a descendant of } u \text{ in }T_v\}.$$ See Figure~\ref{fig:definition} for an example. \end{definition} \begin{figure}[h] \begin{tikzpicture}[scale=1] \draw (2,-1)--(0,0)--(-2,-1); \draw (0,0)--(0.5,-1); \draw (-1.5,-2)--(-2,-1)--(-2.5,-2); \draw (0.5,-2)--(0.5,-1)--(1.5,-2); \draw (-0.5,-2)--(0.5,-1); \draw (-3,-3)--(-2.5,-2)--(-2,-3); \draw (-1.5,-3)--(-1.5,-2); \draw (0,-3)--(0.5,-2)--(1,-3); \draw (1.75,-3)--(1.5,-2); \node at (0,0){$\bullet$}; \node at (-2,-1){$\bullet$}; \node at (2,-1){$\bullet$}; \node at (0.5,-1){$\bullet$}; \node at (-2.5,-2){$\bullet$}; \node at (-1.5,-2){$\bullet$}; \node at (0.5,-2){$\bullet$}; \node at (-0.5,-2){$\bullet$}; \node at (1.5,-2){$\bullet$}; \node at (-3,-3){$\bullet$}; \node at (-2,-3){$\bullet$}; \node at (-1.5,-3){$\bullet$}; \node at (1,-3){$\bullet$}; \node at (0,-3){$\bullet$}; \node at (1.75,-3){$\bullet$}; \draw (0,0) node[above]{$v$}; \draw (-2,-1) node[above]{$u$}; \draw (-2.2,-2.2) ellipse (1.2 and 1.7); \draw (-3.1,-1) node[left]{$T_v(u)$}; \end{tikzpicture} \caption{Definition of $T_v(u)$.} \label{fig:definition} \end{figure} For a tree $T$, the edge set of $T$ is denoted as $E(T)$. The vertex set of $T$ is denoted as $V(T)$, or $T$ for simplicity. Accordingly, $|T|$ is the number of vertices in $T$. The same notations are used for rooted trees. The following proposition provides a way to calculate the number of shellings of a rooted tree $T_v$ based on the size of its rooted subtrees. \begin{proposition}\label{prop:rootedTreeCounting} $$F(T_v) = \frac{n!}{\prod_{u\in T} |T_v(u)|}.$$ \end{proposition} \begin{proof} The proposition holds for $n=2$ by regular check. Assume that it holds for $n-1$ and consider a tree $T$ with $n$ vertices. Suppose the neighbors of $v$ are $u_1, u_2,\ldots,u_r.$ For $1\leq i\leq r,$ define $T^{(i)}$ to be the tree $T_{v}(u_i)$ with an additional edge $(u_i,v)$. Given fixed shellings $\sigma_i$ of $T_v^{(i)}$ for all $1\leq i\leq r$, we can construct shellings of $T_v$ by merging $\sigma_i$'s together while preserving the order of each $\sigma_i$. Every shelling of $T_v$ can be uniquely constructed in this way. Therefore, $$F(T_v) = \binom{|E(T)|}{|E(T^{(1)})|,|E(T^{(2)})|,\ldots,|E(T^{(r)})|}\prod_{i=1}^r F(T_v^{(i)})$$ By induction hypothesis, \begin{align*} F(T_v^{(i)}) &= \frac{|T^{(i)}|!}{\prod_{w\in T^{(i)}}|T_v^{(i)}(w)|} = \frac{|T^{(i)}|!}{|T^{(i)}_v(v)|\prod_{w\neq v, w\in T^{(i)}}|T_v(w)|} \\ &=\frac{|E(T^{(i)})|!}{\prod_{w\neq v, w\in T^{(i)}}|T_v(w)|}. \end{align*} Therefore, \begin{align*} F(T_v) &= \frac{|E(T)|!}{|E(T^{(1)})|!\cdots |E(T^{(r)})|!}\prod_{i=1}^r\frac{|E(T^{(i)})|!}{\prod_{w\neq v, w\in T^{(i)}}|T_v(w)|} \\ &= \frac{|E(T)|!}{\prod_{w\neq v, w\in T} |T_v(w)|} \\ &= \frac{n!}{\prod_{w\in T} |T_v(w)|}. \end{align*} and the induction is complete. \end{proof} \begin{corollary}\label{cor:rootedTreeRatio} Suppose that $(u,v)$ is an edge of $T$, then $$\frac{F(T_v)}{F(T_u)} = \frac{|T_u(v)|}{|T_v(u)|} = \frac{|T_u(v)|}{n-|T_u(v)|}.$$ \end{corollary} \begin{proof} For any vertex $w\neq u,v$, $T_u(w)$ and $T_v(w)$ are the same subtree of $T_v$. Therefore, by Proposition~\ref{prop:rootedTreeCounting}, \begin{align*} \frac{F(T_v)}{F(T_u)} &= \frac{\prod_{w\in T} |T_u(w)|}{\prod_{w\in T} |T_v(w)|} = \frac{|T_u(v)|\cdot |T_u(u)|}{|T_v(u)|\cdot |T_v(v)|} \\ &= \frac{|T_u(v)|}{|T_v(u)|} = \frac{|T_u(v)|}{n-|T_u(v)|}. \end{align*} \end{proof} Corollary ~\ref{cor:rootedTreeRatio} establishes a simple relationship between the number of shellings of $T$ rooted at adjacent edges. In this way, by only calculating $F(T_v)$ for a single vertex $v$, one can quickly derive $F(T_u)$ for all $u\in T.$ For example, suppose $T$ is a path of length $n-1$, as shown in figure~\ref{fig:path}. Then $F(T_{v_1}) = 1$, and $$F(T_{v_{i+1}}) = \frac{|T_{v_i}(v_{i+1})|}{n-|T_{v_i}(v_{i+1})|}F(T_{v_i}) = \frac{n-i}{i}F(T_{v_i})$$ by corollary~\ref{cor:rootedTreeRatio}. This gives $F(T_{v_i}) = \binom{n-1}{i-1}$ for all $i=1,2,\ldots,n.$ \begin{figure}[h] \begin{tikzpicture}[scale=1] \draw(0,0)--(4,0); \draw(5,0)--(7,0); \node at (0,0){$\bullet$}; \node at (1,0){$\bullet$}; \node at (2,0){$\bullet$}; \node at (3,0){$\bullet$}; \node at (7,0){$\bullet$}; \node at (6,0){$\bullet$}; \node[draw=none] (ellipsis2) at (4.5,0) {$\cdots$}; \draw (0,0) node[below]{$v_1$}; \draw (1,0) node[below]{$v_2$}; \draw (2,0) node[below]{$v_3$}; \draw (3,0) node[below]{$v_4$}; \draw (6,0) node[below]{$v_{n-1}$}; \draw (7,0) node[below]{$v_n$}; \end{tikzpicture} \caption{A path of length $n-1$. The shelling number is $2^{n-2}$.} \label{fig:path} \end{figure} Finally, the following proposition relates the number of shellings of $T$ with that of its rooted trees. \begin{proposition}\label{prop:treeCounting} $$F(T) = \frac{1}{2}\sum_{v\in T} F(T_v).$$ \end{proposition} \begin{proof} Note that any shelling of $T$ beginning with edge $(u,v)$ is counted as a shelling of both $T_u$ and $T_v$. Thus, Proposition~\ref{prop:treeCounting} follows. \end{proof} \begin{example}\label{ex:path} By Proposition~\ref{prop:treeCounting} and the discussion under Corollary~\ref{cor:rootedTreeRatio}, the number of shellings of a path of length $n-1$ is $$\frac{1}{2}\sum_{i=1}^n \binom{n-1}{i-1} = 2^{n-2}.$$ \end{example} \subsection{Bounds on Tree Shelling Number} \ The goal of this section is to give several bounds of tree shelling numbers based on various parameters of a graph, such as vertex degree and diameter. A trivial upper bound is $(n-1)!$, since every shelling is also a permutation of edges. The upper bound is achieved when $T$ is a star, in which every two edges are adjacent to each other. Here are the main theorems of the section. \begin{theorem}\label{thm:lowerBoundDegree} \begin{equation*} F(T) \geq \prod_{v\in T} d(v)!, \end{equation*} where $d(v)$ is the degree of a vertex $v$ in $T$. The equality holds if and only if $T$ is a path of length $n-1$ or a star. \end{theorem} \begin{remark} A weaker lower bound $F(T)\geq\prod_{v\in T}\big(d(v)-1\big)!$ can be shown easily by observation. However, an extra factor of $\prod_{v\in T}d(v)$ in Theorem~\ref{thm:lowerBoundDegree} requires much more efforts. \end{remark} \begin{theorem}\label{thm:upperBoundDiameter} Suppose the diameter of $T$ is $\ell$. When $\ell$ is even, $$F(T)\leq \frac{2(n-1-\frac{\ell}{2})!}{(\frac{\ell}{2})!}\bigg[\binom{n-2}{\frac{\ell}{2}}+\sum_{i=0}^{\frac{\ell}{2}-1}\binom{n-1}{i}\bigg].$$ When $\ell$ is odd, $$F(T) \leq \frac{(n-\frac{\ell+3}{2})!}{(\frac{\ell+1}{2})!}\bigg[(n-1-\ell)\binom{n-2}{\frac{\ell-1}{2}}+n\sum_{i=0}^{\frac{\ell-1}{2}}\binom{n-1}{i}\bigg].$$ The equality holds if and only if $T$ has the following form: there exists a path $$v_0 - v_1 - \cdots -v_\ell$$ such that every edge not in this path is adjacent to $v_{\lfloor \frac{\ell}{2}\rfloor}.$ \end{theorem} Before proving Theorem~\ref{thm:lowerBoundDegree}, it is worth noticing the following inequality, which relates the number of shellings of $T$ and $T_v$. \begin{lemma}\label{lem:weightInequality} Let $v$ be a vertex in $T$ and $\ell$ be the length of the longest descending path in $T_v$. Then $$F(T)\leq \bigg[\sum_{k=0}^{\ell-1}\binom{n-2}{k}\bigg] F(T_v).$$ In particular, $F(T)\leq 2^{n-2}F(T_v).$ \end{lemma} \begin{proof} Let $L = v - v_1 - v_2 -\cdots - v_\ell$ be the longest descending path in $T_v$. Consider the following operations on $T$: \begin{enumerate} \item[1.] Suppose $i\leq \ell-2$ is the first index such that $v_i$ has a children $v'\neq v_{i+1}$ in $T_v$. Remove $T_v(v')$ and attach it on $v_{i+1}$ (i.e., children of $v'$ become children of $v_{i+1}$). Furthermore, remove edge $(v',v_i)$ and add a new edge $(v',v_{i+1}).$ This operation is illustrated in Figure~\ref{fig:pathAdjustmentForWeight}. \item[2.] Repeat step 1 until no further operations can be performed. \end{enumerate} \begin{figure}[h] \begin{tikzpicture}[scale=1] \draw(0,0)--(2,0); \draw(3,0)--(6,0); \draw(7,0)--(8,0); \draw(4,0)--(4.5,0.866); \draw(4,0)--(4.707,-0.707); \draw(4,0)--(4.259,-0.966); \node at (0,0){$\bullet$}; \node at (1,0){$\bullet$}; \node at (4,0){$\bullet$}; \node at (5,0){$\bullet$}; \node at (4.5,0.866){$\bullet$}; \node at (4.707,-0.707){$\bullet$}; \node at (4.259,-0.966){$\bullet$}; \node at (8,0){$\bullet$}; \draw[rotate around={-15:(4.5,-0.866)}] (5.5,-0.866) ellipse (1.5 and 0.5); \draw[rotate around={15:(4.5,0.866)}] (5.5,0.866) ellipse (1.5 and 0.5); \node[draw=none] (ellipsis2) at (2.5,0) {$\cdots$}; \node[draw=none] (ellipsis2) at (6.5,0) {$\cdots$}; \draw (0,0) node[below]{$v$}; \draw (1,0) node[below]{$v_1$}; \draw (4,0) node[below left]{$v_i$}; \draw (5,0) node[below]{$v_{i+1}$}; \draw (4.5,0.866) node[above]{$v'$}; \draw (5.5,1.1) node[]{$T_{v}(v')$}; \draw (8,0) node[below]{$v_\ell$}; \end{tikzpicture} $$\Big\Downarrow$$ \begin{tikzpicture}[scale=1] \draw(0,0)--(2,0); \draw(3,0)--(6,0); \draw(7,0)--(8,0); \draw(5,0)--(4.5,0.866); \draw(4,0)--(4.707,-0.707); \draw(4,0)--(4.259,-0.966); \node at (0,0){$\bullet$}; \node at (1,0){$\bullet$}; \node at (4,0){$\bullet$}; \node at (5,0){$\bullet$}; \node at (4.5,0.866){$\bullet$}; \node at (4.707,-0.707){$\bullet$}; \node at (4.259,-0.966){$\bullet$}; \node at (8,0){$\bullet$}; \draw[rotate around={-15:(4.5,-0.866)}] (5.5,-0.866) ellipse (1.5 and 0.5); \draw[rotate around={36:(5,0)}] (6,0) ellipse (1.5 and 0.5); \node[draw=none] (ellipsis2) at (2.5,0) {$\cdots$}; \node[draw=none] (ellipsis2) at (6.5,0) {$\cdots$}; \draw (0,0) node[below]{$v$}; \draw (1,0) node[below]{$v_1$}; \draw (4,0) node[below left]{$v_i$}; \draw (5,0) node[below]{$v_{i+1}$}; \draw (4.5,0.866) node[above]{$v'$}; \draw (5.8,0.6) node[]{$T_{v}(v')$}; \draw (8,0) node[below]{$v_\ell$}; \end{tikzpicture} \caption{Operation on $T$: moving edges away from root.} \label{fig:pathAdjustmentForWeight} \end{figure} Such operations preserve the length of the longest descending path in $T_v$ and would eventually stop within finite steps. Let $T^{(k)}$ be the tree after $k$th operation. For $u\in T$, define the weight of $u$ in $T^{(k)}$ $$W_k(u) = \frac{F(T^{(k)}_u)}{F(T^{(k)}_v)}.$$ We claim that the sum of weights of all vertices is non-decreasing after each operation, i.e. \begin{equation}\label{eq:weightInequality} \sum_{u\in T}W_k(u) \leq \sum_{u\in T}W_{k+1}(u). \end{equation} It suffices to prove the claim for $k=0.$ By Corollary~\ref{cor:rootedTreeRatio}, suppose $(u,w)$ is an edge in $T^{(k)}$, then \begin{equation}\label{eq:weightRatio} \frac{W_k(u)}{W_k(w)} = \frac{|T^{(k)}_w(u)|}{|T^{(k)}_u(w)|} = \frac{|T^{(k)}_w(u)|}{n-|T^{(k)}_w(u)|}. \end{equation} Therefore, suppose $v-u_1-u_2-\cdots - u_r = u$ is a path in $T^{(k)}_v$, then \begin{equation*} W_k(u) = \prod_{j=1}^r \frac{|T^{(k)}_v(u_j)|}{n-|T^{(k)}_v(u_j)|}. \end{equation*} Note that for all $u\neq v', v_{i+1}$, $|T^{(1)}_v(u)| = |T_v(u)|$. For $w\not\in T_v(v')\cup T_v(v_{i+1})$, $v',v_{i+1}$ are not on the path from $v$ to $w$, so \begin{equation*} W_0(w) = W_1(w). \end{equation*} Write $|T_v(v'))|=a, |T_v(v_{i+1})| = b$, then $|T^{(1)}_v(v')| = 1$, $|T^{(1)}_v(v_{i+1})| = |T_v(v')|+|T_v(v_{i+1})| = a+b.$ By (\ref{eq:weightRatio}), $$W_0(v') = \frac{a}{n-a}W_0(v_i).$$ $$W_0(v_{i+1}) = \frac{b}{n-b}W_0(v_i).$$ $$W_1(v_{i+1}) = \frac{a+b}{n-a-b}W_1(v_i).$$ Since $W_0(v_i) = W_1(v_i)$, \begin{equation*} W_0(v')+W_0(v_{i+1})\leq W_1(v_{i+1}). \end{equation*} For $w\in T_v(v')\setminus\{v'\}$, by (\ref{eq:weightRatio}), \begin{equation*} \frac{W_1(w)}{W_1(v_{i+1})} = \frac{W_0(w)}{W_0(v')} \implies W_0(w)\leq W_1(w). \end{equation*} Similarly, for $w\in T_v(v_{i+1})\setminus\{v_{i+1}\}$, \begin{equation*} \frac{W_1(w)}{W_1(v_{i+1})} = \frac{W_0(w)}{W_0(v_{i+1})}\implies W_0(w)\leq W_1(w). \end{equation*} Therefore, we conclude that $$\sum_{w\in T} W_0 (w)\leq \sum_{w\in T} W_1(w),$$ and (\ref{eq:weightInequality}) is proved. Finally, suppose the operation stops after step $M$, then $T^{(M)}$ is the tree where all vertices not in $L$ are incident to $v_{\ell-1}$. Thus, by~\ref{eq:weightRatio}, \begin{align*} \sum_{u\in T} W_{M}(u) &= W_M(v)+W_M(v_1)+\cdots+W_M(v_{\ell-1}) + (n-\ell)W_M(v_\ell) \\ &=\sum_{i=0}^{\ell-1}\binom{n-1}{i} + (n-l)\frac{\binom{n-1}{\ell-1}}{n-1}\\ &=2\sum_{i=0}^{\ell-1}\binom{n-2}{i}. \end{align*} According to equation (\ref{eq:weightInequality}), Proposition~\ref{prop:treeCounting}, $$\frac{F(T)}{F(T_v)} = \frac{1}{2}\sum_{u\in T}W_0(u) \leq \frac{1}{2}\sum_{u\in T}W_M(u) = \sum_{i=0}^{\ell-1}\binom{n-2}{i},$$ so the proof is complete. \end{proof} Now we are ready to prove Theorem~\ref{thm:lowerBoundDegree} and ~\ref{thm:upperBoundDiameter}. \begin{proof}[Proof of Theorem~\ref{thm:lowerBoundDegree}] Induct on $|T|$. When $|T| = 2$, $F(T) = 2 = \prod_{v\in T} d(v)!$. The equality holds if and only if $T$ is a path (in this case $T$ is also a star). Assume that statement holds for all $|T|<n$, consider the case where $|T| = n$. If $|T|$ is a path of length $n-1$, then by Example~\ref{ex:path}, $$F(T) = 2^{n-2} = \prod_{v\in T} d(v)!,$$ as desired. Suppose that $T$ is not a single path, then there exists a vertex $v$ of degree $d\geq 3.$ Let $u_1,u_2,\ldots,u_d$ be vertices adjacent to $v$ and write $|T_v(u_i)| = s_i$ for $i=1,2,\ldots,d$. Assume $s_1\leq s_2\leq \ldots \leq s_d.$ Let $T'$ be the subtree of $T$ obtained by removing all vertices in $T_v(u_1)$ and all edges incident to those vertices. Let $T''$ be the subtree of $T$ induced by edges in $E(T)\setminus E(T').$ See Figure~\ref{fig:theoremLowerBound} for illustration. \begin{figure}[h] \begin{tikzpicture}[scale=1] \draw(0,0)--(1,0); \draw(0,0)--(0,1); \draw(0,0)--(0,-1); \draw(0,0)--(-0.866,-0.5); \draw(0,0)--(-0.866,0.5); \node at (0,0){$\bullet$}; \node at (1,0){$\bullet$}; \node at (0,1){$\bullet$}; \node at (0,-1){$\bullet$}; \node at (-0.866,-0.5){$\bullet$}; \node at (-0.866,0.5){$\bullet$}; \draw (1.5,0) ellipse (2 and 0.5); \draw (-0.7,0) ellipse (1.3 and 2); \draw (0,0) node[below right]{$v$}; \draw (1,0) node[below]{$u_1$}; \draw (0,1) node[above]{$u_2$}; \draw (-0.866,0.5) node[above]{$u_3$}; \draw (0,-1) node[below]{$u_d$}; \draw (2,0) node[]{$T''$}; \draw (-1.2,0) node[]{$T'$}; \end{tikzpicture} \caption{Merging a shelling of $T'$ and $T''_v$ to a shelling of $T$.} \label{fig:theoremLowerBound} \end{figure} Suppose $\sigma'$ is a shelling of $T'$ and $\sigma''$ a shelling of $T''_v$. Consider the following method to merge $\sigma'$ and $\sigma''$ into $\sigma$, a permutation of $E(T)$, such that (i) the order of edges in $\sigma'$ and in $\sigma''$ are preserved; (ii) $\sigma''(1) = (v,u_1)$ is not one of the first $s_d$ edges after merge. Note that $\sigma$ must be a shelling of $T$, since at least one of $\{\sigma'(k): 1\leq k\leq s_d\}$ is incident to $v$ and $(v,u_1)$ is adjacent to some previous edges in $\sigma$. For fixed $\sigma'$ and $\sigma''$, the number of $\sigma$ constructed by the above merging method is $$\binom{n-1-s_d}{|E(T'')|} = \binom{s_1+s_2+\cdots s_{d-1}}{s_1}.$$ Therefore, \begin{equation}\label{eq:shellingExtensionNumber} F(T) \geq F(T')F(T''_v)\binom{s_1+s_2+\cdots+ s_{d-1}}{s_1}. \end{equation} Note that the shellings of $T$ constructed above do not include those whose first edge is $(v,u_1)$, so we can replace ``$\geq$" with ``$>$" in (\ref{eq:shellingExtensionNumber}). Furthermore, by Lemma~\ref{lem:weightInequality}, $$F(T''_v) \geq \frac{F(T'')}{2^{s_1-1}}.$$ By induction hypothesis, $$F(T')F(T'') \geq \frac{1}{d}\prod_{u\in T} d(u)!.$$ Thus, (\ref{eq:shellingExtensionNumber}) implies $$F(T) > \frac{1}{2^{s_1-1}d}\binom{s_1+s_2+\cdots+s_{d-1}}{s_1}\prod_{u\in T}d(u)!.$$ If for some choices of $v$ with degree $d\geq 3$, $\binom{s_1+s_2+\cdots+s_{d-1}}{s_1} \geq 2^{s_1-1}d$, then $F(T) > \prod_{u\in T} d(u)!$ and equality never holds. If not, for all choices of $v$, $\binom{s_1+s_2+\cdots+s_{d-1}}{s_1} < 2^{s_1-1}d.$ By Lemma~\ref{app:ineq1}, $s_1 = s_2 = \cdots = s_{d-1} = 1.$ Therefore, $T$ must be the following type of trees: for every vertex $v$ of degree $d(v) \geq 3$, it connects at least $d(v) - 1$ leaves. If $T$ is a star, then $F(T) = (n-1)!$ is an equality case. If not, $T$ has the form shown in Figure~\ref{fig:Stars}, where $v_0$ and $v_m$ are the only two possible vertices with degree at least 3. \begin{figure}[h] \begin{tikzpicture}[scale=1] \draw(0,0)--(2,0); \draw(3,0)--(4,0); \draw(0,0)--(0,1); \draw(0,0)--(0,-1); \draw(0,0)--(-0.707,0.707); \draw(0,0)--(-1,0); \draw(4,0)--(4,-1); \draw(4,0)--(4,1); \draw(4,0)--(5,0); \draw(4,0)--(4.707,0.707); \node[draw=none] (ellipsis2) at (2.5,0) {$\cdots$}; \node at (0,0){$\bullet$}; \node at (1,0){$\bullet$}; \node at (0,1){$\bullet$}; \node at (0,-1){$\bullet$}; \node at (-0.707,0.707){$\bullet$}; \node at (-1,0){$\bullet$}; \node at (4,0){$\bullet$}; \node at (5,0){$\bullet$}; \node at (4,1){$\bullet$}; \node at (4,-1){$\bullet$}; \node at (4.707,0.707){$\bullet$}; \node[rotate around={45:(0,0)}] (ellipsis1) at (-0.55,-0.45) {$\vdots$}; \node[rotate around={45:(0,0)}] (ellipsis1) at (4.5,-0.5) {$\cdots$}; \draw (0,0) node[below right]{$v_0$}; \draw (1,0) node[below]{$v_1$}; \draw (4,0) node[below left]{$v_m$}; \end{tikzpicture} \caption{The only type of trees that satisfy case 2 condition.} \label{fig:Stars} \end{figure} Suppose $d(v_0) = d_1$, $d(v_m) = d_2$ with $2\leq d_1\leq d_2$. If $m=1$, by Proposition~\ref{prop:rootedTreeCounting}, ~\ref{prop:treeCounting}, and Lemma~\ref{app:ineq2}, \begin{align*} F(T) &= \frac{d_1^2+d_2^2+d_1d_2-d_1-d_2}{d_1d_2}(d_1+d_2-2)! \\ &\geq 2\cdot(d_1+d_2-2)!\\ &\geq d_1!d_2! = \prod_{u\in T} d(u)!. \end{align*} The equality holds only if $d_1 = d_2 = 2$ and $T$ is a single path. Now suppose $m\geq 2.$ Consider the following type of shelling of $T$: The first $m-1$ edges of $\sigma$ consist of $\{(v_i,v_{i+1}): 0\leq i\leq m-2\}$. The number of shellings of such type is $$2^{m-2}(d_1-1)!(d_2-1)!\binom{d_1+d_2-1}{d_2}.$$ Similarly, the number of shellings whose first $m-1$ edges consist of $\{(v_i,v_{i+1}): 1\leq i\leq m-1\}$ is $$2^{m-2}(d_1-1)!(d_2-1)!\binom{d_1+d_2-1}{d_1}.$$ Thus by Lemma~\ref{app:ineq3}, \begin{align*} F(T) &\geq 2^{m-2}(d_1-1)!(d_2-1)!\bigg[\binom{d_1+d_2-1}{d_2}+\binom{d_1+d_2-1}{d_1}\bigg] \\ &= 2^{m-2}(d_1-1)!(d_2-1)!\binom{d_1+d_2}{d_1}\\ &> 2^{m-1}d_1!d_2! = \prod_{u\in T}d(u)! \end{align*} unless $d_1=2, d_2\leq 4$, in which cases we have: \begin{itemize} \item $(d_1,d_2) = (2,2).$ $F(T) = 2^{n-2} = \prod_{u\in T}d(u)!.$ In this equality case, $T$ is a single path. \item $(d_1,d_2) = (2,3).$ $F(T) = 2^{n-1}-2 > 3!\cdot 2^{n-4} = \prod_{u\in T}d(u)!.$ \item $(d_1,d_2) = (2,4).$ $F(T) = 6(2^{n-2}-n+1) > 4!\cdot 2^{n-5} = \prod_{u\in T}d(u)!.$ \end{itemize} By induction, the proof of Theorem~\ref{thm:lowerBoundDegree} is complete. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:upperBoundDiameter}] Let $v_0 - v_1 -\cdots - v_\ell$ be a longest path in $T$. Firstly, we reduce the problem to the case where all edges in $T$ are incident to $\{v_1,v_2,\ldots,v_{\ell-1}\}$. If not, construct a new tree $T'$ by removing every edge $e$ not incident to $\{v_i:1\leq i\leq \ell-1\}$ and adding a corresponding edge incident to $v_j$, where $v_j$ is the closest vertex from $e$ among $L$. Every shelling of $T$ is still a shelling of $T'$ by considering the corresponding edges. Thus, $F(T)\leq F(T')$ while the longest path remains the same. Under this assumption, denote $V' = T\setminus \{v_0,v_1,\ldots,v_\ell\}.$ Consider the following operations: \begin{enumerate} \item[1.] Let $i$ be the smallest index such that $v_i$ has degree $\geq 3.$ If $i<\frac{\ell}{2}$, we remove all edges of the form $(v_i,u)$ for $u\in V'$ and add edges $(u,v_{i+1}).$ \item[2.] Repeat step 1 until no further operations can be performed. \item[3.] Let $j$ be the largest index such that $v_j$ has degree $\geq 3.$ If $j>\frac{\ell}{2}$, we remove all edges of the form $(v_j,u)$ for $u\in V'$ and add edges $(u,v_{j-1}).$ \item[4.] Repeat step 3 until no further operations can be performed. \end{enumerate} \begin{figure}[h] \begin{tikzpicture}[scale=1.3] \draw(1,0)--(7,0); \draw(2.293,0.707)--(3,0)--(3.707,0.707); \draw(3,0)--(3,1); \draw(4,-1)--(4,0); \draw(4.5,-0.866)--(5,0)--(5.5,-0.866); \draw(6,0)--(6,1); \node at (1,0){$\bullet$}; \node at (2,0){$\bullet$}; \node at (3,0){$\bullet$}; \node at (4,0){$\bullet$}; \node at (5,0){$\bullet$}; \node at (6,0){$\bullet$}; \node at (7,0){$\bullet$}; \node at (2.293,0.707){$\bullet$}; \node at (3,1){$\bullet$}; \node at (3.707,0.707){$\bullet$}; \node at (4,-1){$\bullet$}; \node at (4.5,-0.866){$\bullet$}; \node at (5.5,-0.866){$\bullet$}; \node at (6,1){$\bullet$}; \draw (1.5,0) node[below]{5}; \draw (2.5,0) node[below]{2}; \draw (3.5,0) node[below]{4}; \draw (4.5,0) node[below]{6}; \draw (5.5,0) node[below]{10}; \draw (6.5,0) node[below]{11}; \draw (2.5,0.5) node[below]{1}; \draw (2.95,0.55) node[right]{8}; \draw (3.5,0.5) node[below]{3}; \draw (4,-0.5) node[left]{9}; \draw (4.7,-0.5) node[left]{7}; \draw (5.3,-0.5) node[right]{13}; \draw (6,0.5) node[right]{12}; \end{tikzpicture} $$\Big\Downarrow$$ \begin{tikzpicture}[scale=1.3] \draw(1,0)--(7,0); \draw(3.293,0.707)--(4,0)--(4.707,0.707); \draw(4,0)--(4,1); \draw(4,-1)--(4,0); \draw(4.5,-0.866)--(5,0)--(5.5,-0.866); \draw(6,0)--(6,1); \node at (1,0){$\bullet$}; \node at (2,0){$\bullet$}; \node at (3,0){$\bullet$}; \node at (4,0){$\bullet$}; \node at (5,0){$\bullet$}; \node at (6,0){$\bullet$}; \node at (7,0){$\bullet$}; \node at (3.293,0.707){$\bullet$}; \node at (4,1){$\bullet$}; \node at (4.707,0.707){$\bullet$}; \node at (4,-1){$\bullet$}; \node at (4.5,-0.866){$\bullet$}; \node at (5.5,-0.866){$\bullet$}; \node at (6,1){$\bullet$}; \draw (1.5,0) node[below]{10}; \draw (2.5,0) node[below]{6}; \draw (3.5,0) node[below]{4}; \draw (4.5,0) node[below]{2}; \draw (5.5,0) node[below]{5}; \draw (6.5,0) node[below]{11}; \draw (3.5,0.5) node[below]{1}; \draw (3.95,0.55) node[right]{8}; \draw (4.5,0.5) node[below]{3}; \draw (4,-0.5) node[left]{9}; \draw (4.7,-0.5) node[left]{7}; \draw (5.3,-0.5) node[right]{13}; \draw (6,0.5) node[right]{12}; \end{tikzpicture} \caption{An example of operation on $T$: moving edges towards middle. The shellings are indicated by the number on the edges. $g$ maps a shelling of the first tree to a shelling of the second tree.} \label{fig:theoremUpperBound} \end{figure} Suppose the above operations end in step $M$. Let $T^{(t)}$ be the tree after $t^{\text{th}}$ operation. We claim that for all $t<M$, \begin{equation*} F(T^{(t+1)}) \geq F(T^{(t)}). \end{equation*} It suffices to prove the case when $t = 0.$ By symmetry, we can assume $i<\frac{\ell}{2}$. Let $V_i$ be the set of vertices adjacent to $v_i$ in $T$ except $v_{i-1}, v_{i+1}$. Define \begin{align*} S_{T\cap T^{(1)}} &:= \{\sigma\text{ is a shelling of }T: \exists u\in V_i, (v_i,u) \text{ appears after }(v_i,v_{i+1}) \text{ in } \sigma\}, \\ S_{T^{(1)}\cap T} &:= \{\tau\text{ is a shelling of }T^{(1)}: \exists u\in V_i, (v_{i+1},u) \text{ appears after }(v_i,v_{i+1}) \text{ in } \tau\}, \\ S_{T\setminus T^{(1)}} &:= \{\sigma\text{ is a shelling of }T: \exists u\in V_i, (v_i,u) \text{ appears before }(v_i,v_{i+1}) \text{ in } \sigma\},\\ S_{T^{(1)}\setminus T} &:= \{\tau\text{ is a shelling of }T^{(1)}: \exists u\in V_i, (v_{i+1},u) \text{ appears before }(v_i,v_{i+1}) \text{ in } \tau\}. \end{align*} Note that there is a bijection between $S_{T\cap T^{(1)}}$ and $S_{T^{(1)}\cap T}$ by replacing edges of the form $(v_i,u)$ in every $\sigma\in S_{T\cap T^{(1)}}$ with $(v_{i+1},u)$, for all $u\in V_i$. Thus, $|S_{T\cap T^{(1)}}|=|S_{T^{(1)}\cap T}|$ and $$F(T^{(1)}) - F(T) = |S_{T^{(1)}\setminus T}| - |S_{T\setminus T^{(1)}}|.$$ Consider a function $g: S_{T\setminus T^{(1)}} \rightarrow S_{T^{(1)}\setminus T}$. For $\sigma \in S_{T\setminus T^{(1)}}$, define $\tau = g(\sigma)$ as follows. If $\sigma(k) = (v_i, u)$ for some $u\in V_i$, $\tau(k) = (v_{i+1}, u)$; if $\sigma(k) = (v_j,u)$ for some $j\neq i$ and $u\in V'$, $\tau(k) = (v_j,u)$. It remains to define $\tau(k)$'s where $\sigma(k)$ is an edge of $L$. Write $e(j) = (v_{j},v_{j+1})$ for $j=0,1,\ldots,\ell-1.$ For $1\leq r\leq \ell$, suppose $\sigma(k_r) = e(j_r)$ where $k_1<k_2<\cdots < k_\ell.$ Define $\tau(k_r)$ inductively: when $r=1$, $\tau(k_1) = e(2i-j_1).$ When $r\geq 2$, $$\tau(k_{r}) = \begin{cases} e(j_{r}), & \text{ if } \{\tau(k_1), \tau(k_2),\ldots, \tau(k_{r-1})\} =\{e(j_1), e(j_2),\ldots, e(j_{r-1})\} \\ e(2i-j_{r}), & \text{ otherwise.} \end{cases}$$ The idea is that $g$ maps edges not in $L$ to the corresponding edges. For edges in $L$, $g$ acts as a reflection with respect to $e(i)$, until the reflection image matches with the preimage. An example of $g$ is in Figure~\ref{fig:theoremUpperBound}. We check the following properties of $g$: \begin{itemize} \item $g$ is well-defined. We first note that for any $r\leq \ell$, both $\{\sigma(k_1),\sigma(k_2),\ldots,\sigma(k_r)\}$ and $\{\tau(k_1),\tau(k_2),\ldots,\tau(k_r)\}$ form a path $P_r$ and $P^{(1)}_r$ in $L$, respectively. Since $j_1 \leq i$, the right endpoint of $P^{(1)}_r$ is never on the left side of the right endpoint of $P_r$ (assuming that $L$ is a horizontal path with left endpoint $v_0$ and right endpoint $v_\ell$, as illustrated in Figure~\ref{fig:theoremUpperBound}). Furthermore, since the ``branching edges" of $T$ (edges in $E(T)\setminus E(L)$) are not on the left side of $v_i$, every branching edge adjacent to $P_r$ must be adjacent to $P^{(1)}_r$. Thus, $\tau$ is a shelling of $T^{(1)}$. Moreover, $\tau\in S_{T^{(1)}\setminus T}$ by the correspondence between $(v_i,u)\in \sigma$ and $(v_{i+1},u)\in \tau$ for all $u\in V_i$. Therefore, $g$ is well-defined. \item $g$ is injective. Suppose $g(\sigma) = \tau.$ By the definition of $g$, $\sigma(k)$ is uniquely determined whenever $\tau(k) \not\in L.$ Suppose $\tau(k_r) = e(i_r)$ for $1\leq r\leq \ell$. we can recover $\sigma(k_r)$ from $\tau$: $\sigma(k_1) = e(2i-i_1)$. When $r\geq 2$, $$\sigma(k_{r}) = \begin{cases} e(i_{r}), & \text{ if } \{\sigma(k_1), \sigma(k_2),\ldots, \sigma(k_{r-1})\} =\{e(i_1), e(i_2),\ldots, e(i_{r-1})\} \\ e(2i-i_{r}), & \text{ otherwise.} \end{cases}$$ Therefore, $\sigma$ is uniquely determined by $\tau$ and $g$ is injective. \end{itemize} Since $g$ is injective, $|S_{T^{(1)}\setminus T}| \geq |S_{T\setminus T^{(1)}}|$ and thus $F(T^{(1)}) \geq F(T).$ Finally, note that $T^{(M)}$ is the tree where all edges not in $L$ are incident to $v_{\lfloor\frac{\ell}{2}\rfloor}$. By Proposition~\ref{prop:rootedTreeCounting} and ~\ref{prop:treeCounting}, $$F(T^{(M)})= \begin{cases} \frac{2(n-1-\frac{\ell}{2})!}{(\frac{\ell}{2})!}\bigg[\binom{n-2}{\frac{\ell}{2}}+\sum_{i=0}^{\frac{\ell}{2}-1}\binom{n-1}{i}\bigg], &\text{ if }\ell \text{ is even,} \\ \frac{(n-\frac{\ell+3}{2})!}{(\frac{\ell+1}{2})!}\bigg[(n-1-\ell)\binom{n-2}{\frac{\ell-1}{2}}+n\sum_{i=0}^{\frac{\ell-1}{2}}\binom{n-1}{i}\bigg], &\text{ if } \ell \text{ is odd.} \end{cases} $$ Thus, the proof of inequality is complete. Futhermore, we shall prove that $g$ is surjective only if $T$ is isomorphic to $T^{(M)}.$ If not, then there are two cases: \noindent \textbf{Case 1.} $i< \frac{\ell-1}{2}$. In this case, $2i<\ell-1$. Thus, for every $\sigma \in S_{T\setminus T^{(1)}}$, $g(\sigma)(1)\neq e(\ell-1) = (v_{\ell-1},v_\ell)$. However, there exists $\tau\in S_{T^{(1)}\setminus T}$ whose first edge is $(v_{\ell-1},v_{\ell})$, contradiction! \noindent \textbf{Case 2.} $i= \frac{\ell-1}{2}$ and there exists another vertex $v_j$ of degree at least 3. Suppose $(v_j, u)$ is an edge not in $L$, then for every $\sigma \in S_{T\setminus T^{(1)}}$, $g(\sigma)(1)$ cannot be this edge. However, there exists $\tau\in S_{T^{(1)}\setminus T}$ whose first edge is $(v_j,u)$, contradiction! Therefore, $g$ is surjective only if $T$ is isomorphic to $T^{(M)}$, so $$F(T) = F(T^{(M)})$$ if and only if $T$ is isormorphic to $T^{(M)}$. This completes the proof of Theorem~\ref{thm:upperBoundDiameter}. \end{proof} \section*{Acknowledgements} This research was carried out as part of the 2018 Summer Program in Undergraduate Research (SPUR) of the MIT Mathematics Department. The authors would like to thank Prof. Richard Stanley for suggesting the project and Prof. Ankur Moitra and Prof. Davesh Maulik for helpful suggestions. \bibliographystyle{plain}

https://arxiv.org/abs/1611.01153

On Perfectness of Intersection Graph of Ideals of $\mathbb{Z}_n$

In this paper, we characterize the positive integers $n$ for which intersection graph of ideals of $\mathbb{Z}_n$ is perfect.

\section{Introduction} The idea of associating graphs to algebraic structures for characterizing the algebraic structures with graphs and vice versa dates back to Bosak \cite{bosak}. Till then, a lot of research, e.g., \cite{graph-ideal,anderson-livingston,badawi,power2,mks-ideal,angsu-comm-alg-1,angsu-lin-mult-alg,angsu-comm-alg-2,angsu-jaa,survey2} has been done in connecting graph structures to various algebraic objects like groups, rings, vector spaces etc. However, the most prominent among them are the zero-divisor graphs \cite{anderson-livingston} and intersection graph of ideals of rings \cite{mks-ideal}. Recently, authors in \cite{weakly-perfect} proved that intersection graph of ideals of $\mathbb{Z}_n$ is weakly perfect for all $n>0$. In this paper, we characterize the values of $n$ for which the intersection graph of ideals of $\mathbb{Z}_n$ is perfect. In particular, we prove the following theorem. \begin{theorem*} The intersection graph of ideals of $\mathbb{Z}_n$ is perfect if and only if $n={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$ where $p_i$'s are distinct primes and $\alpha_i \in \mathbb{N}\cup\{0\}$, i.e., the number of distinct prime factors of $n$ is less than or equal to $4$. \end{theorem*} \section{Definition, Preliminaries and Known Results} In this section, for convenience of the reader and also for later use, we recall some definitions, notations and results concerning elementary graph theory and intersection graph of ideals of a ring. For undefined terms and concepts the reader is referred to \cite{west-graph-book}. By a graph $G=(V,E)$, we mean a non-empty set $V$ and a symmetric binary relation (possibly empty) $E$ on $V$. The set $V$ is called the set of vertices and $E$ is called the set of edges of $G$. Two element $u$ and $v$ in $V$ are said to be adjacent if $(u,v) \in E$. $H=(W,F)$ is called an {\it induced subgraph} of $G$ if $\phi \neq W \subseteq V$ and $F$ consists of all the edges between the vertices in $W$ in $G$. A complete subgraph of a graph $G$ is called a {\it clique}. A {\it maximal clique} is a clique which is maximal with respect to inclusion. The {\it clique number} of $G$, written as $\omega(G)$, is the maximum size of a clique in $G$. The {\it chromatic number} of $G$, denoted as $\chi(G)$, is the minimum number of colours needed to label the vertices so that the adjacent vertices receive different colours. It is easy to observe that $\omega(G)\leq \chi(G)$. A graph $G$ is said to be {\it weakly perfect} if $\omega(G)= \chi(G)$ and it is said to be {\it perfect} if $\omega(H)= \chi(H)$ for all induced subgraphs $H$ of $G$. Chudnovsky {\it et.al.} \cite{spgt} in 2004 settled a long standing conjecture regarding perfect graphs and provided a characterization of perfect graphs. {\theorem[Strong Perfect Graph Theorem] \label{spgt} \cite{spgt} A graph $G$ is perfect if and only if neither $G$ nor its complement contains an odd cycle of length at least $5$ as an induced subgraph.} Let $R$ be a ring. The intersection graph of ideals of $R$ (introduced in \cite{mks-ideal}), denoted by $G(R)$, consists of all non-trivial ideals as vertices and two ideals $I$ and $J$ are adjacent if and only if $I \cap J\neq \{0\}$. Throughout this paper, we take the ring $R$ to be $\mathbb{Z}_n$, the ring of integers modulo $n$. We know that $\mathbb{Z}_n$ is a principal ideal ring and each of its ideals is generated by $\overline{m}\in \mathbb{Z}_n$ where $m$ is a factor of $n$. For convenience, we denote this ideal by $(m)$. Also without loss of generality, whenever we take an ideal $(m)$ of $\mathbb{Z}_n$, we assume that $m$ is a factor of $n$. It was proved in \cite{weakly-perfect} proved that intersection graph of ideals of $\mathbb{Z}_n$ is weakly perfect, i.e., $\omega(G( \mathbb{Z}_n))=\chi(G( \mathbb{Z}_n))$ for all $n>0$. \section{Perfectness of Intersection Graph of Ideals of $\mathbb{Z}_n$} In this section, we prove some preparatory results and subsequently use them to prove the main theorem of the paper. {\proposition \label{lcm-lemma} Let $G(\mathbb{Z}_n)$ be the intersection graph of ideals of $\mathbb{Z}_n$ and $(a)$ and $(b)$ be two ideals in $\mathbb{Z}_n$ such that $a \mid n$ and $b\mid n$. Then $(a)$ and $(b)$ are adjacent in $G(\mathbb{Z}_n)$ if and only if $lcm(a,b)$ is a factor of $n$ and $1<lcm(a,b)<n$.}\\ \\ \noindent {\bf Proof: } Since $\mathbb{Z}_n$ is isomorphic to $\mathbb{Z}/n\mathbb{Z}$ as ring via the correspondence $\overline{a}\leftrightarrow a+n\mathbb{Z}$, the ideal $(a)$ in $\mathbb{Z}_n$ corresponds to the ideal $\langle a \rangle + n\mathbb{Z}$ in $\mathbb{Z}/n\mathbb{Z}$ where $\langle a \rangle$ denote the set of integer multiples of $a$. Now, let $(a)\sim (b)$ in $G(\mathbb{Z}_n)$, i.e., $(a)\cap(b)\neq \{\overline{0}\}$. Since $a \mid n$ and $b\mid n$, we have $lcm(a,b)\mid n$. On the other hand, using the correspondence described above, we have $\langle a \rangle + n\mathbb{Z} \cap \langle b \rangle + n\mathbb{Z}\neq \{n\mathbb{Z}\}$. But, we know that $\langle a \rangle + n\mathbb{Z} \cap \langle b \rangle + n\mathbb{Z}=\langle lcm(a,b) \rangle + n\mathbb{Z}$. Hence, we have $\langle lcm(a,b) \rangle + n\mathbb{Z} \neq \{n\mathbb{Z}\}$. This, together with the fact that $lcm(a,b)\mid n$, implies that $1<lcm(a,b)<n$. Conversely, let $lcm(a,b)$ is a factor of $n$ and $1<lcm(a,b)<n$. Clearly, $\overline{0} \neq \overline{lcm(a,b)} \in (a) \cap (b)$ in $\mathbb{Z}_n$ and hence $(a)\sim (b)$ in $G(\mathbb{Z}_n)$. \hfill \rule{2mm}{2mm} {\theorem \label{not-perfect} Let $n={p_1}^{\alpha_1}{p_2}^{\alpha_2}\cdots {p_k}^{\alpha_k}$. If $k\geq 5$, then $G(\mathbb{Z}_n)$ is not perfect.}\\ \\ \noindent {\bf Proof: } Let $n={p_1}^{\alpha_1}{p_2}^{\alpha_2}\cdots {p_5}^{\alpha_5}.s$ where $s=1$ if $k=5$ and $s={p_6}^{\alpha_6}\cdots {p_k}^{\alpha_k}$ if $k>5$. Consider the cycle $C$ given by $({p_1}^{\alpha_1}{p_2}^{\alpha_2} {p_3}^{\alpha_3}s)\sim({p_2}^{\alpha_2}{p_3}^{\alpha_3} {p_4}^{\alpha_4}s)\sim({p_3}^{\alpha_3}{p_4}^{\alpha_4} {p_5}^{\alpha_5}s)\sim({p_4}^{\alpha_4}{p_5}^{\alpha_5} {p_1}^{\alpha_1}s)\sim({p_5}^{\alpha_5}{p_1}^{\alpha_1}{p_2}^{\alpha_2} s)\sim ({p_1}^{\alpha_1}{p_2}^{\alpha_2} {p_3}^{\alpha_3}s)$. Simple calculation using Proposition \ref{lcm-lemma} shows that $C$ is an induced $5$-cycle in $G(\mathbb{Z}_n)$ and hence by Theorem \ref{spgt}, $G(\mathbb{Z}_n)$ is not perfect.\hfill \rule{2mm}{2mm} {\theorem \label{no-odd-hole} Let $n={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$. Then $G(\mathbb{Z}_n)$ does not contain any induced cycle of length greater than $4$.}\\ \\ \noindent {\bf Proof: } Let, if possible $G(\mathbb{Z}_n)$ contains an induced cycle $C$ of length greater than $4$, say $(a_1)\sim(a_2)\sim(a_3)\sim(a_4)\sim(a_5)\sim \cdots \sim (a_1)$. By Proposition \ref{lcm-lemma}, we have $$lcm(a_1,a_3)=lcm(a_1,a_4)=lcm(a_2,a_4)=lcm(a_2,a_5)=lcm(a_3,a_5)=n.$$ {\it Claim: $gcd(a_1,a_3)>1$.} If possible, let $gcd(a_1,a_3)=1$. Since $gcd(a_1,a_3)\cdot lcm(a_1,a_3)=a_1a_3$, we have $lcm(a_1,a_3)=a_1a_3=n$. Note that as $lcm(a_3,a_5)=n$, we have $n=\frac{a_1a_3}{gcd(a_1,a_3)}=\frac{a_3a_5}{gcd(a_3,a_5)}$, i.e., $a_1\cdot gcd(a_3,a_5)=a_5\cdot gcd(a_1,a_3)$, i.e., $a_5=a_1\cdot gcd(a_3,a_5)$, i.e., $a_5$ is a multiple of $a_1$. Now as $a_1$ and $a_3$ are coprime and their lcm is $n$, without loss of generality, two cases may arise: either $a_1={p_1}^{\alpha_1}{p_2}^{\alpha_2}; a_3={p_3}^{\alpha_3}{p_4}^{\alpha_4}$ or $a_1={p_1}^{\alpha_1};a_3={p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$. If $a_1={p_1}^{\alpha_1}{p_2}^{\alpha_2}; a_3={p_3}^{\alpha_3}{p_4}^{\alpha_4}$, we have $a_5={p_1}^{\alpha_1}{p_2}^{\alpha_2}\cdot s$ for some natural number $s$ such that $a_5 \mid n$. Also as $lcm(a_1,a_4)=n$, we have $a_4={p_3}^{\alpha_3}{p_4}^{\alpha_4}\cdot t$, for some natural number $t$ such that $a_4 \mid n$. Thus $lcm(a_4,a_5)=n$ contradicting Proposition \ref{lcm-lemma} and the fact that $a_4\sim a_5$ in $C$. If $a_1={p_1}^{\alpha_1};a_3={p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$, similarly we have $a_5={p_1}^{\alpha_1}\cdot s$ and $a_4={p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4} \cdot t$ and hence $lcm(a_4,a_5)=n$ thereby leading to a contradiction. Thus by combining above two cases, we have $gcd(a_1,a_3)>1$. Thus we have $lcm(a_1,a_3)=n$ and $gcd(a_1,a_3)>1$ with $a_1 \mid n$ and $a_3\mid n$. Without loss of generality, let $p_1$ be a common factor of $a_1$ and $a_3$ and let $a_1={p_1}^x\cdot s$ and $a_3={p_1}^y \cdot t$ where $p_1$ is coprime with $s$ and $t$. Now, if $max\{x,y\}<\alpha_1$, then $lcm(a_1,a_3)<n$, a contradiction. Thus either $x=\alpha_1$ or $y=\alpha_1$, i.e., for any common prime divisor $p_i$ of $a_1$ and $a_3$, either ${p_i}^{\alpha_i} \mid a_1$ or ${p_i}^{\alpha_i} \mid a_3$ or both. Also as $lcm(a_1,a_3)=n$, all the ${p_i}^{\alpha_i}$ are factors of either $a_1$ or $a_3$ or both. Thus, without loss of generality, the forms of $a_1$ and $a_3$ are as follows: either $$\mathsf{Case ~ 1:}~ a_1={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\beta_3}{p_4}^{\beta_4};a_3={p_1}^{\beta_1}{p_2}^{\beta_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$$ or $$\mathsf{Case ~ 2:}~a_1={p_1}^{\alpha_1}{p_2}^{\beta_2}{p_3}^{\beta_3}{p_4}^{\beta_4};a_3={p_1}^{\beta_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$$ or $$\mathsf{Case ~ 3:}~a_1={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\beta_3}{p_4}^{\beta_4};a_3={p_1}^{\beta_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$$ or $$\mathsf{Case ~ 4:}~a_1={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\beta_4};a_3={p_1}^{\beta_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$$ where $\beta_i < \alpha_i$. Note that in first two cases, $a_1$ and $a_3$ do not share any ${p_i}^{\alpha_i}$ as common factor. In the third case, they share only one ${p_i}^{\alpha_i}$ as common factor and in the fourth case, they share two ${p_i}^{\alpha_i}$'s as common factor. {\it Case 1: ($a_1={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\beta_3}{p_4}^{\beta_4};a_3={p_1}^{\beta_1}{p_2}^{\beta_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$)} Since $lcm(a_1,a_4)=n$, we have $a_4={p_1}^{\gamma_1}{p_2}^{\gamma_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$ where $\gamma_1\leq \alpha_1,\gamma_2 \leq \alpha_2$ and $(\gamma_1,\gamma_2)\neq (\alpha_1,\alpha_2)$. Again, since $lcm(a_3,a_5)=n$, we have $a_5={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\delta_3}{p_4}^{\delta_4}$ where $\delta_3\leq \alpha_3,\delta_4 \leq \alpha_4$ and $(\delta_3,\delta_4)\neq (\alpha_3,\alpha_4)$. Hence, we have $lcm(a_4,a_5)=n$, a contradiction to the fact that $a_4 \sim a_5$. Thus Case 1 is an impossibility. {\it Case 2: ($a_1={p_1}^{\alpha_1}{p_2}^{\beta_2}{p_3}^{\beta_3}{p_4}^{\beta_4};a_3={p_1}^{\beta_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$)} Since $lcm(a_1,a_4)=n$, we have $a_4={p_1}^{\gamma_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$ where $\gamma_1< \alpha_1$. Again, since $lcm(a_3,a_5)=n$, we have $a_5={p_1}^{\alpha_1}{p_2}^{\delta_2}{p_3}^{\delta_3}{p_4}^{\delta_4}$ where $\delta_i\leq \alpha_i$ and $(\delta_2,\delta_3,\delta_4)\neq (\alpha_2,\alpha_3,\alpha_4)$. Hence, we have $lcm(a_4,a_5)=n$, a contradiction to the fact that $a_4 \sim a_5$. Thus Case 2 is an impossibility. {\it Case 3: ($a_1={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\beta_3}{p_4}^{\beta_4};a_3={p_1}^{\beta_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$)} Since $lcm(a_1,a_4)=n$, we have ${p_3}^{\alpha_3}{p_4}^{\alpha_4}\mid a_4$. Again, since $lcm(a_3,a_5)=n$, we have ${p_1}^{\alpha_1}\mid a_5$. Now, as $lcm(a_2,a_5)=n$, we have either ${p_2}^{\alpha_2}\mid a_2$ or ${p_2}^{\alpha_2}\mid a_5$. But if ${p_2}^{\alpha_2}\mid a_5$, then we have $lcm(a_4,a_5)=n$, a contradiction. Thus, we have ${p_2}^{\alpha_2}\mid a_2$. Again, as $lcm(a_2,a_4)=n$, we have either ${p_1}^{\alpha_1}\mid a_2$ or ${p_1}^{\alpha_1}\mid a_4$. If ${p_1}^{\alpha_1}\mid a_2$, then $lcm(a_2,a_3)=n$, a contradiction. On the other hand, if ${p_1}^{\alpha_1}\mid a_4$, then $lcm(a_3,a_4)=n$, a contradiction. Thus Case 3 is an impossibility. {\it Case 4: ($a_1={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\beta_4};a_3={p_1}^{\beta_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$)} Since $lcm(a_1,a_4)=n$, we have ${p_4}^{\alpha_4}\mid a_4$. Now, as $lcm(a_2,a_4)=n$, we have either ${p_1}^{\alpha_1}\mid a_2$ or ${p_1}^{\alpha_1}\mid a_4$. If ${p_1}^{\alpha_1}\mid a_2$, then $lcm(a_2,a_3)=n$, a contradiction. On the other hand, if ${p_1}^{\alpha_1}\mid a_4$, then $lcm(a_3,a_4)=n$, a contradiction. Thus Case 4 is an impossibility. Thus, combining all the cases we conclude that $G(\mathbb{Z}_n)$ does not contain any induced cycle of length greater than $4$.\hfill \rule{2mm}{2mm} {\theorem \label{no-odd-antihole} Let $n={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$. Then $\overline{G(\mathbb{Z}_n)}$, the complement of $G(\mathbb{Z}_n)$, does not contain any induced cycle of length greater than $4$.}\\ \\ \noindent {\bf Proof: } Let, if possible $\overline{G(\mathbb{Z}_n)}$ contains an induced cycle $C$ of length greater than $4$, say $(a_1)\sim(a_2)\sim(a_3)\sim(a_4)\sim \cdots \sim(a_t)\sim (a_1)$ with $t\geq 5$. Then, by Proposition \ref{lcm-lemma}, $lcm(a_1,a_2)=lcm(a_2,a_3)=lcm(a_3,a_4)=\cdots =lcm(a_t,a_1)=n$. [{\it Claim: $gcd(a_2,a_3)>1$}] If possible, let $gcd(a_2,a_3)=1$. Since $lcm(a_2,a_3)=n$, we have $n=a_2a_3$. Thus without loss of generality, either $$a_2={p_1}^{\alpha_1}{p_2}^{\alpha_2};a_3={p_3}^{\alpha_3}{p_4}^{\alpha_4} \mbox{ or } a_2={p_1}^{\alpha_1};a_3={p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$$ If $a_2={p_1}^{\alpha_1}{p_2}^{\alpha_2};a_3={p_3}^{\alpha_3}{p_4}^{\alpha_4}$, as $lcm(a_3,a_4)=lcm(a_1,a_2)=n$, we have $a_1={p_3}^{\alpha_3}{p_4}^{\alpha_4}\cdot s$ and $a_4={p_1}^{\alpha_1}{p_2}^{\alpha_2}\cdot t$ for some positive integer $s,t$. But this implies that $lcm(a_1,a_4)=n$, i.e., $a_1\sim a_4$ in $\overline{G(\mathbb{Z}_n)}$, a contradiction. On the other hand, if $a_2={p_1}^{\alpha_1};a_3={p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$, as $lcm(a_3,a_4)=lcm(a_1,a_2)=n$, we have $a_1={p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}\cdot s$ and $a_4={p_1}^{\alpha_1}\cdot t$ for some positive integer $s,t$. But this implies that $lcm(a_1,a_4)=n$, i.e., $a_1\sim a_4$ in $\overline{G(\mathbb{Z}_n)}$, a contradiction. Hence the claim is true. Now, we have $lcm(a_2,a_3)=n$ and $gcd(a_2,a_3)>1$ with $a_2 \mid n$ and $a_3\mid n$. Without loss of generality, let $p_1$ be a common factor of $a_2$ and $a_3$ and let $a_2={p_1}^x\cdot s$ and $a_3={p_1}^y \cdot t$ where $p_1$ is coprime with $s$ and $t$. Now, if $max\{x,y\}<\alpha_1$, then $lcm(a_2,a_3)<n$, a contradiction. Thus either $x=\alpha_1$ or $y=\alpha_1$, i.e., for any common prime divisor $p_i$ of $a_2$ or $a_3$, either ${p_i}^{\alpha_i} \mid a_2$ or ${p_i}^{\alpha_i} \mid a_3$ or both. Also as $lcm(a_2,a_3)=n$, all the ${p_i}^{\alpha_i}$ are factors of either $a_2$ or $a_3$. Thus, without loss of generality, the forms of $a_2$ and $a_3$ are as follows: either $$\mathsf{Case ~ 1:}~ a_2={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\beta_3}{p_4}^{\beta_4};a_3={p_1}^{\beta_1}{p_2}^{\beta_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$$ or $$\mathsf{Case ~ 2:}~a_2={p_1}^{\alpha_1}{p_2}^{\beta_2}{p_3}^{\beta_3}{p_4}^{\beta_4};a_3={p_1}^{\beta_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$$ or $$\mathsf{Case ~ 3:}~a_2={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\beta_3}{p_4}^{\beta_4};a_3={p_1}^{\beta_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$$ or $$\mathsf{Case ~ 4:}~a_2={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\beta_4};a_3={p_1}^{\beta_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$$ where $\beta_i < \alpha_i$. Note that in first two cases, $a_2$ and $a_3$ do not share any ${p_i}^{\alpha_i}$ as common factor. In the third case, they share only one ${p_i}^{\alpha_i}$ as common factor and in the fourth case, they share two ${p_i}^{\alpha_i}$'s as common factor. {\it Case 1: ($a_2={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\beta_3}{p_4}^{\beta_4};a_3={p_1}^{\beta_1}{p_2}^{\beta_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$)} Since $lcm(a_1,a_2)=lcm(a_3,a_4)=n$, we have ${p_3}^{\alpha_3}{p_4}^{\alpha_4} \mid a_1$ and ${p_1}^{\alpha_1}{p_2}^{\alpha_2}\mid a_4$. But this implies $lcm(a_1,a_4)=n$, i.e., $a_1\sim a_4$ in $\overline{G(\mathbb{Z}_n)}$, a contradiction and hence Case 1 is an impossibility. {\it Case 2: ($a_2={p_1}^{\alpha_1}{p_2}^{\beta_2}{p_3}^{\beta_3}{p_4}^{\beta_4};a_3={p_1}^{\beta_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$)} Since $lcm(a_1,a_2)=lcm(a_3,a_4)=n$, we have ${p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}\mid a_1$ and ${p_1}^{\alpha_1}\mid a_4$. But this implies $lcm(a_1,a_4)=n$, a contradiction and hence Case 2 is an impossibility. {\it Case 3: ($a_2={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\beta_3}{p_4}^{\beta_4};a_3={p_1}^{\beta_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$)} Since $lcm(a_1,a_2)=n$, we have ${p_3}^{\alpha_3}{p_4}^{\alpha_4} \mid a_1$. Also, since $lcm(a_t,a_1)=n$, either ${p_1}^{\alpha_1}\mid a_1$ or ${p_1}^{\alpha_1}\mid a_t$. If ${p_1}^{\alpha_1}\mid a_1$, then we have ${p_1}^{\alpha_1}{p_3}^{\alpha_3}{p_4}^{\alpha_4} \mid a_1$ which implies $lcm(a_1,a_3)=n$, i.e., $a_1\sim a_3$ in $\overline{G(\mathbb{Z}_n)}$, a contradiction. On the other hand, if ${p_1}^{\alpha_1}\mid a_t$, we have $lcm(a_t,a_3)=n$, i.e., $a_t \sim a_3$ in $\overline{G(\mathbb{Z}_n)}$, a contradiction. Thus combining both the possibilities, Case 3 is an impossibility. {\it Case 4: ($a_2={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\beta_4};a_3={p_1}^{\beta_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$)} Since $lcm(a_1,a_2)=n$, we have ${p_4}^{\alpha_4} \mid a_1$. Also, since $lcm(a_t,a_1)=n$, either ${p_1}^{\alpha_1}\mid a_1$ or ${p_1}^{\alpha_1}\mid a_t$. If ${p_1}^{\alpha_1}\mid a_1$, then we have ${p_1}^{\alpha_1}{p_4}^{\alpha_4} \mid a_1$ which implies $lcm(a_1,a_3)=n$, i.e., $a_1\sim a_3$ in $\overline{G(\mathbb{Z}_n)}$, a contradiction. On the other hand, if ${p_1}^{\alpha_1}\mid a_t$, we have $lcm(a_t,a_3)=n$, i.e., $a_t \sim a_3$ in $\overline{G(\mathbb{Z}_n)}$, a contradiction. Thus combining both the possibilities, Case 4 is an impossibility. Thus, combining all the cases we conclude that $\overline{G(\mathbb{Z}_n)}$ does not contain any induced cycle of length greater than $4$.\hfill \rule{2mm}{2mm} Finally, with Theorems \ref{spgt}, \ref{not-perfect}, \ref{no-odd-hole} and \ref{no-odd-antihole} in hand, we are now in a position to prove the main result of this paper. \begin{theorem*} The intersection graph of ideals of $\mathbb{Z}_n$ is perfect if and only if $n={p_1}^{\alpha_1}{p_2}^{\alpha_2}{p_3}^{\alpha_3}{p_4}^{\alpha_4}$ where $p_i$'s are distinct primes and $\alpha_i \in \mathbb{N}\cup\{0\}$, i.e., the number of distinct prime factors of $n$ is less than or equal to $4$. \end{theorem*} \noindent {\bf Proof: } Clearly, Theorem \ref{not-perfect} shows that the condition is necessary. For the sufficiency part, first with the help of Theorems \ref{no-odd-hole} and \ref{no-odd-antihole}, along with Theorem \ref{spgt}, we conclude that the intersection graph of ideals of $\mathbb{Z}_n$ is perfect if $n$ has exactly four distinct prime factors. The proofs for the cases when $n$ has exactly three, two or one distinct prime factors follows similarly by suitably taking some of the $\alpha_i$'s to be zero. \hfill \rule{2mm}{2mm} \section*{Acknowledgement} The author is thankful to Sabyasachi Dutta and Jyotirmoy Pramanik for some fruitful discussions on the paper. The research is partially funded by NBHM Research Project Grant, (Sanction No. 2/48(10)/2013/ NBHM(R.P.)/R\&D II/695), Govt. of India.

https://arxiv.org/abs/1508.02851

On interval edge-colorings of bipartite graphs of small order

An edge-coloring of a graph $G$ with colors $1,\ldots,t$ is an interval $t$-coloring if all colors are used, and the colors of edges incident to each vertex of $G$ are distinct and form an interval of integers. A graph $G$ is interval colorable if it has an interval $t$-coloring for some positive integer $t$. The problem of deciding whether a bipartite graph is interval colorable is NP-complete. The smallest known examples of interval non-colorable bipartite graphs have $19$ vertices. On the other hand it is known that the bipartite graphs on at most $14$ vertices are interval colorable. In this work we observe that several classes of bipartite graphs of small order have an interval coloring. In particular, we show that all bipartite graphs on $15$ vertices are interval colorable.

\section{Introduction} In this paper we consider only finite, undirected graphs, without loops and multiple edges. $V(G)$ and $E(G)$ denote the sets of vertices and edges, respectively. The degree of the vertex $v \in V(G)$ is denoted by $d_G(v)$. The concepts and notations not defined here can be found in \cite{West}. A proper edge-coloring of a graph $G$ is a coloring of the edges of $G$ such that no two adjacent edges receive the same color. If $\alpha$ is a proper edge-coloring of $G$ and $v \in V(G)$, then by $S(v,\alpha)$ we denote the set of colors of the edges incident to $v$. The largest color of $S(v,\alpha)$ is denoted by $\overline{S}(v,\alpha)$. A proper edge-coloring of a graph $G$ with colors $1,...,t$ is called an interval $t$-coloring if all colors are used, and for any vertex $v \in V(G)$, the set $S(v,\alpha)$ is an interval of integers. A graph $G$ is interval colorable if it has an interval $t$-coloring for some $t \in \mathbb{N}$. The set of all interval colorable graphs is denoted by $\mathfrak{N}$. The concept of interval edge-coloring of graphs was introduced by Asratian and Kamalian \cite{AK1987,AK1994}. In \cite{K1989}, Kamalian proved that all complete bipartite graphs and trees are interval colorable. In \cite{P2010,PKT2013}, it was shown that $n$-dimensional cubes have interval $t$-coloring if and only if $ n \leq t \leq \frac{n(n+1)}{2}$. In \cite{Sevastjanov1990}, Sevastjanov proved that it is an NP-complete problem to decide whether a bipartite graph has an interval coloring or not. \begin{figure}[t] \label{fig19} \begin{center} \includegraphics[width=8cm]{fig3.jpg} \caption{Interval non-colorable bipartite graph $F$ on $19$ vertices} \vspace{1.7cm} \end{center} \end{figure} In \cite{Hansen1992} Hansen proved that every bipartite graph $G$ with $\Delta(G) \leq 3$ is interval colorable. In \cite{JensenToft} Jensen and Toft formulated the following question: is there an interval non-colorable bipartite graph $G$ with $4 \leq \Delta(G) \leq 12$? A partial answer to this question was given by Petrosyan and the first author in \cite{PK2014}. In particular they constructed two interval non-colorable bipartite graphs, $G$ and $H$, where $|V(G)|=20$, $\Delta(G)=11$, and $|V(H)|=19$, $\Delta(H)=12$. Another example of interval non-colorable bipartite graph $F$ on $19$ vertices (Fig. \ref{fig19}) was discovered by Mirumyan in 1989, but was not published, and was independently found by Giaro, Kubale and Malafiejski in \cite{GKM1999}. On the other hand, in \cite{G1999} it was shown that all bipartite graphs on at most $14$ vertices are interval colorable. Based on these results the following problem was posed in \cite{PK2014}: \begin{problem} Is there a bipartite graph $G$ with $15 \leq |V(G)| \leq 18$ and $G \notin \mathfrak{N}$? \end{problem} In this work we partially solve this problem by showing that all bipartite graphs on at most $15$ vertices are interval colorable. We also show that some other classes of bipartite graphs of small order are interval colorable. \section{Related work} In 1999, Giaro \cite{G1999} used computer search to show that the following result holds. \begin{theorem} \label{Giaro14} All bipartite graphs of order at most $14$ are interval colorable. \end{theorem} In \cite{GKM1999} it was observed that the both parts of an interval non-colorable bipartite graph should be relatively large. \begin{theorem} \label{GiaroBipartition} If $G$ is a bipartite graph with bipartition $(X,Y)$ and $\min\{|X|,|Y|\} \leq 3$, then $G \in \mathfrak{N}$. \end{theorem} Several algorithms for finding a generalized version of interval edge-colorings of graphs are presented and compared in \cite{BHD2009}. \section {Implementation details} We use a computer search to look for interval non-colorable graphs of small order. First we generate a set of candidate graphs, then we try to color them using distributed computing system, and finally we double check the obtained colorings for possible errors. \subsection{Graph generation} We use the \textbf{nauty} package \cite{McKay} to generate the graphs. In particular, \textbf{genbg} program from the package generates all bipartite graphs according to the given bipartition. Moreover, it allows to specify the minimum and maximum degrees of the graphs, whether the generated graphs should be connected and several other options. \subsection{Distributed computing} In order to color the generated graphs we use CrowdProcess, a web-based distributed computing system \cite{CrowdProcess}. CrowdProcess provides an easy to use interface and REST API to submit a program written in JavaScript language together with a list of tasks as a JSON file. It distributes the program and the tasks to multiple computers (between 1000 and 5000 computers during our experiments), gathers the results and passes back through a web interface or an API. Most of the graphs are colored in less than 1 millisecond, so we send up to 200 graphs for each computer. \subsection{Coloring algorithm} We represent the bipartite graph $G$ with bipartition $(X,Y)$ and its coloring as a biadjacency matrix $B(G) = (b_{ij})_{n \times m}$, where $|X|=n$ and $|Y|=m$. Here $b_{ij}$ is the color of the edge joining the $i$-th vertex of $X$ and the $j$-th vertex of $Y$, if the edge exists, and is set to $0$ otherwise. We use backtracking to fill in the matrix with colors. At each step we calculate the set of possible colors for the current matrix cell (by taking into account already colored edges). We color the edge by a randomly selected color from the set of possible colors and move to the next cell. If for some edge the set of possible colors is empty, we return to the previous edge and change the color (if there still exist a color in the set of possible colors). The algorithm stops when all the edges are colored, or when all the possibilities are tested and the graph has no interval coloring. In practice the latter never happens because CrowdProcess puts a 5 minute time limit on the computation per computer. \subsection{Verification} After downloading the colorings from CrowdProcess we use a C++ program to verify the colorings. We discovered one case when the coloring returned from the CrowdProcess contained an error. We do not know how such an error could occur. We use one more C++ program to gather the graphs which were not colored (due to an error in the coloring or timeout) and send them again. We repeat this process until all the graphs are colored. \section{Results} Let $\mathfrak{F}$ be some set of bipartite graphs. Denote by $C(\mathfrak{F})$ the set of all connected bipartite graphs from $\mathfrak{F}$ having minimum degree at least $2$ and having at least $4$ vertices on each of its parts. Also let $M(\mathfrak{F})$ be the set of all graphs obtained by taking any graph $G \in \mathfrak{F}$ and removing any vertex from it. The following lemma holds. \begin{figure}[t] \label{fig19-2} \begin{center} \includegraphics[width=8cm]{fig6.eps} \caption{Interval non-colorable bipartite graph $H$ on $19$ vertices} \vspace{1.7cm} \end{center} \end{figure} \begin{lemma} \label{mainLemma} If for any set of bipartite graphs $\mathfrak{F}$, all graphs from the sets $M(\mathfrak{F})$ and $C(\mathfrak{F})$ are interval colorable, then all graphs from $\mathfrak{F}$ are also interval colorable. \end{lemma} Let $G$ be a bipartite graph from the set $\mathfrak{F}$. If $G \in C(\mathfrak{F})$, then it is interval colorable. Otherwise it is disconnected, has a minimum degree $1$, or has less than $4$ vertices on at least one of its parts. If $G$ is disconnected, then each of its connected components belongs to $M(\mathfrak{F})$, so the union of the colorings of the connected components will be a coloring of $G$. If there exists some pendant edge $uv \in E(G)$, where $d_G(v)=1$, then we take the coloring $\alpha$ of the graph $G-v$ (which belongs to $M(\mathfrak{F})$) and color the edge $uv$ by the color $\overline{S}(u,\alpha) + 1$. Finally, if one of the parts of $G$ has less than $4$ vertices, then $G$ is interval colorable by the Theorem \ref{GiaroBipartition}. \endproof \begin{theorem} \label{th15} All bipartite graphs on $15$ vertices are interval colorable. \end{theorem} Let $\mathfrak{F}$ be the set of all bipartite graphs on $15$ vertices. All graphs from the set $M(\mathfrak{F})$ are interval colorable due to the Theorem \ref{Giaro14}. According to the Lemma \ref{mainLemma} it is sufficient to show that all graphs from the set $C(\mathfrak{F})$ are interval colorable. The number of graphs in the set $C(\mathfrak{F})$ is 288 643 868. We color them all by using a computer algorithm described in the previous section. Some details of the performed computation are presented in Table \ref{table15}. \endproof \begin{theorem} \label{th4x} All bipartite graphs having $4$ vertices on one part and up to $15$ vertices on the other part are interval colorable except for the graph $G_1$ in Fig. \ref{fig19}. \end{theorem} Let $\mathfrak{F}_{i,j}$ be the set of all bipartite graphs with bipartition $(X,Y)$ where $|X|=i$ and $|Y| = j$, $i, j \in \mathbb{N}$. Note that $M(\mathfrak{F}_{i,j})=\mathfrak{F}_{i-1,j} \cup \mathfrak{F}_{i,j-1}$, for any $i,j>1$. We need to prove that all graphs from the sets $\mathfrak{F}_{4,j}$, $12 \leq j \leq 15$, are interval colorable except for the graph $F$ from the Fig. \ref{fig19}. Note that all the graphs from the sets $\mathfrak{F}_{3,j}$ are interval colorable due to the Theorem \ref{GiaroBipartition}. All graphs from the set $\mathfrak{F}_{4,11}$ are also interval colorable due to the Theorem \ref{th15}. We use the computer algorithm described in the previous section to color all graphs from the sets $C(\mathfrak{F}_{4,j})$, $12 \leq j \leq 15$, except for the graph $F$. The details of the computation are presented in Table \ref{table4x}. To complete the proof, we iteratively apply the Lemma \ref{mainLemma} to the sets $\mathfrak{F}_{4,j}$, for $j=12,13,14,15$. \endproof \begin{table}[t] \renewcommand{\arraystretch}{1.2} \label{table15} \begin{center} \begin{tabular}{|c|c|c|} \hline No. of vertices & No. of graphs & CPU hours\\ \hline 4 / 11 & 16308 & 3.04 \\ \hline 5 / 10 & 1583646 & 146.35 \\ \hline 6 / 9 & 43739172 & 340.51\\ \hline 7 / 8 & 243304742 & 15537.42\\ \hline \end{tabular} \caption{Details of the computation for coloring bipartite graphs of order $15$. CPU hours are reported by CrowdProcess} \vspace{0.3cm} \end{center} \end{table} \begin{table}[t] \renewcommand{\arraystretch}{1.2} \label{table4x} \begin{center} \begin{tabular}{|c|c|c|} \hline No. of vertices & No. of graphs & CPU hours\\ \hline 4 / 12 & 29515 & 4.96 \\ \hline 4 / 13 & 51616 & 19.19 \\ \hline 4 / 14 & 87609 & 96.95\\ \hline 4 / 15 & 144766 & N/A \\ \hline \end{tabular} \caption{Details of the computation for coloring bipartite graphs having $4$ vertices on one part and $j$ vertices on the other part, $12 \leq j \leq 15$. CPU hours are reported by CrowdProcess} \vspace{1.7cm} \end{center} \end{table} \section{Future work} Different algorithms can be tried to find interval colorings of the graphs even faster. In fact, the experiments show that most of the graphs have many different interval colorings and possibly some approximation algorithms will be sufficient to find colorings the easily. So far we have tried algorithms based on simulated annealing with no luck. Currently we are working on coloring the bipartite graphs on $16$ vertices. The number of such graphs after filtering out easily colorable ones (similar to the case of $15$-vertex bipartite graphs) is 12 322 367 816. We have colored more than 98\% of these graphs, but still it cannot be excluded that there exist interval non-colorable graphs on $16$ vertices. \section{Acknowledgements} We would like to thank the CrowdProcess team for donating practically unlimited computational time to us and for the excellent technical support. Also we would like to thank UtopianLab coworking space for providing broadband internet access. Finally, we would like to thank P. A. Petrosyan for many valuable comments and ideas. This work was made possible by a research grant from the Armenian National Science and Education Fund (ANSEF) based in New York, USA.

https://arxiv.org/abs/1004.4953

The Number of Eigenvalues of a Tensor

Eigenvectors of tensors, as studied recently in numerical multilinear algebra, correspond to fixed points of self-maps of a projective space. We determine the number of eigenvectors and eigenvalues of a generic tensor, and we show that the number of normalized eigenvalues of a symmetric tensor is always finite. We also examine the characteristic polynomial and how its coefficients are related to discriminants and resultants.

\section{Introduction} In the current numerical analysis literature, considerable interest has arisen in extending concepts that are familiar from linear algebra to the setting of multilinear algebra. One such familiar concept is that of an eigenvalue of a square matrix. Several authors have explored definitions of eigenvalues and eigenvectors for higher-dimensional tensors, and they have argued that these notions are useful for wide range of applications \cite{Lim}. Our approach in this paper is based on the definition of {\em E-eigenvalues} of tensors introduced by Liqun Qi in \cite{NQWW,Qi}. Throughout this paper, eigenvalue will mean E-eigenvalue, as defined in Definition~\ref{def:eig}. We fix two positive integers $m$ and $n$, and we consider order-$m$ tensors $A = (a_{i_1 i_2 \cdots i_m})$ of format $\,n \times n \times \cdots \times n\,$ with entries in the field of complex numbers $\mathbb C$. \begin{defn} \label{def:eig} \rm Let $x$ be in $\mathbb C^n$ and $A$ a tensor as above. Adopting the notation introduced in \cite{NQWW, Qi}, we define $Ax^{m-1}$ to be the vector in $\mathbb C^n$ whose $j$-th coordinate is the scalar \begin{equation} \label{eq:Am} (Ax^{m-1})_{j} \quad = \quad \sum_{i_2 = 1}^n \cdots \sum_{i_m=1}^n a_{j i_2 \cdots i_m} x_{i_2}\cdots x_{i_m}. \end{equation} If $\lambda$ is a complex number and $x \in \mathbb C^n$ a non-zero vector such that $A x^{m-1} = \lambda x$, then $\lambda$ is an \emph{eigenvalue} of~$A$ and $x$ is an \emph{eigenvector} of $A$. We will refer to the pair of $\lambda$ and~$x$ as an \emph{eigenpair}. Two eigenpairs $(\lambda,x)$ and $(\lambda', x')$ of the same tensor $A$ are considered to be {\em equivalent} if there exists a complex number $t \neq 0$ such that $t^{m-2}\lambda = \lambda'$ and $t x = x'$. \end{defn} If $m=2$ then $Ax^1$ is the ordinary matrix-vector product, and Definition \ref{def:eig} recovers the familiar eigenvalues and eigenvectors of a square matrix $A$. In that case, our notion of equivalence amounts to rescaling the eigenvector, but the eigenvalue is uniquely determined. For $m \geq 3$, Qi \cite{Qi} normalizes the eigenvectors $x$ of $A$ by additionally requiring $x \cdot x = 1$. When $x \cdot x$ is non-zero, this has the effect of choosing two distinguished representatives, related by $\lambda' = (-1)^m \lambda$ and $x' = -x$, from each equivalence class. In particular, when $m$ is even, the eigenvalue is uniquely determined, and when $m$ is odd, it is determined up to sign. However, since equivalence classes with $x \cdot x = 0$ are not allowed by Qi's normalization, his definition does not strictly generalize the classical eigenvalues of a matrix. We will call eigenvalues $\lambda$ with an eigenvector satisfying $x \cdot x = 1$ \emph{normalized eigenvalues} of the tensor~$A$. In Section~6 of~\cite{Qi}, Qi considers an alternative normalization by requiring $x \cdot \overline x = 1$, where $\overline x$ is the complex conjugate. This reduces the equivalence classes from two real dimensions to one real dimension. One can still get an equivalent eigenpair $(t^{m-2}\lambda, t x)$ for any complex number $t$ with unit modulus. Yet another normalization, based on the $p$-norm over the real numbers $\mathbb R$, was introduced by Lek-Heng Lim in his variational approach \cite{Lim}. For most of this paper, we prefer not to choose any normalization whatsoever. Instead, we depend on the notion of equivalence in Definition~\ref{def:eig} in order to have a finite number of equivalence classes of eigenpairs in the generic case. This equivalence is a generalization of the usual ambiguity of eigenvectors of an $n {\times} n$-matrix $A$, which at best, are only unique up to scaling. The following theorem generalizes, from $m=2$ to $m \geq 3$, the familiar linear algebra fact that an $n {\times} n$-matrix $A$ has precisely $n$ eigenvalues over the complex numbers. \begin{theorem}\label{thm:count} If a tensor $A$ has finitely many equivalence classes of eigenpairs over~$\mathbb C$ then their number, counted with multiplicity, is equal to $\,((m-1)^n - 1)/(m-2)$. If the entries of $A$ are sufficiently generic, then all multiplicities are equal to $1$, so there are exactly $\,((m-1)^n - 1)/(m-2)\,$ equivalence classes of eigenpairs. \end{theorem} For the case when the tensor order $m$ is even, the above formula was derived in \cite[Theorem~3.4]{NQWW} by means of a detailed analysis of the Macaulay matrix for the multivariate resultant. For arbitrary $m$, it was stated as a conjecture in line 2 on page 1228 of \cite{NQWW}. We here present a short proof of this conjecture that is based on techniques from toric geometry \cite{CS,fulton}. The rest of this paper is organized as follows. In Section~\ref{sec:count}, we prove Theorem~\ref{thm:count}. In Section~\ref{sec:characteristic}, we investigate the characteristic polynomial. Tensors with vanishing characteristic polynomial are interpreted as singular tensors. In Section~\ref{sec:dynamics}, we relate eigenvalues of tensors to dynamics on projective space. Finally, in Section~\ref{sec:symmetric}, we specialize to the case of symmetric tensors. We show that, in that case, the set of normalized eigenvalues is always finite. \section{Intersections in a Weighted Projective Space}\label{sec:count} We shall formulate the problem of computing the eigenvalues and eigenvectors of the tensor~$A$ as an intersection problem in the $n$-dimensional weighted projective space \begin{equation*} X \,\,\, = \,\,\, \mathbb P(1,1,\ldots, 1, m-2). \end{equation*} The textbook definition of~$X$ can be found, for example, in \cite[\S 2.0]{CS} and \cite[page 35]{fulton}. Points in~$X$ are represented by vectors of complex numbers $(u_1: \cdots: u_n: \lambda)$, not all zero, modulo the rescaling $(t u_1:\cdots: t u_n: t^{m-2} \lambda)$ for any non-zero complex number~$t$. The corresponding algebraic representation of our weighted projective space is $\, X = \operatorname{Proj}(R)$, where $R = \mathbb C[x_1, \ldots, x_n, \lambda]$ is the polynomial ring with $x_1,\ldots,x_n$ having degree~$1$ and $\lambda$ having degree~$m-2$. The following proof uses basic toric intersection theory as in \cite[Ch.~5]{fulton}. \begin{proof}[Proof of Theorem \ref{thm:count}] For $m = 2$, the expression $((m-1)^n-1)/(m-2)$ simplifies to $ n$, which is the number of eigenvalues of an ordinary $n {\times} n$-matrix. Hence we shall now assume that $\,m \geq 3$. For a fixed tensor~$A$, the $n$ equations determined by $\,Ax^{m-1} = \lambda x\,$ correspond to $n$ homogeneous polynomials of degree $m-1$ in our graded polynomial ring $R$. Since $R$ is generated in degree~$m-2$, the line bundle $\mathcal O_X(m-2)$ is very ample. The corresponding lattice polytope~$\Delta$ is an $n$-dimensional simplex with vertices at $(m-2)e_i$ for $1 \leq i \leq n$, and $e_{n+1}$, where the $e_i$ are the basis vectors in $\mathbb R^{n+1}$. The affine hull of~$\Delta$ is the hyperplane $\,x_1+\cdots + x_n + (m{-}2)\lambda = m{-} 2$. The normalized volume of this simplex equals \begin{equation} \label{eq:simplexvolume} {\rm Vol}(\Delta) \,\, = \,\, (m-2)^{n-1} . \end{equation} The lattice polytope $\Delta$ is smooth, except at the vertex~$e_{n+1}$, where it is simplicial with index $m-2$. Therefore, the projective toric variety $X$ is simplicial, with precisely one isolated singular point corresponding to the vertex~$e_{n+1}$. By \cite[p.~100]{fulton}, the variety $X$ has a rational Chow ring $A^*(X)_{\mathbb Q}$, which we can use to compute intersection numbers of divisors on~$X$. Our system of equations $Ax^{m-1} = \lambda x$ consists of $n$~polynomials of degree $m-1$ in~$R$. Let $D$ be the divisor class corresponding to $\mathcal O_X(m-1)$, and let $H$ be the very ample divisor class corresponding to $\mathcal O_X(m-2)$. The volume formula (\ref{eq:simplexvolume}) is equivalent to $\,H^n = (m-2)^{n-1}\,$ in $A^*(X)_{\mathbb Q}$, and we compute the self-intersection number of $D$ as the following rational number: \begin{equation*} D^n \quad = \quad \left(\frac{m-1}{m-2} \cdot H\right)^{\!\! n} \quad = \quad \left(\frac{m-1}{m-2}\right)^{\!\! n} \! \cdot (m-2)^{n-1} \quad = \quad \frac{(m-1)^n}{m-2}. \end{equation*} From this count we must remove the trivial solution $\{x=0\}$ of $\,Ax^{m-1} = \lambda x$. That solution corresponds to the singular point $e_{n+1}$ on~$X$. Since that point has index $m-2$, the trivial solution counts for $1/(m-2)$ in the intersection computation, as shown in \cite[p.~100]{fulton}. Therefore the number of non-trivial solutions in $X$ is equal to \begin{equation} \label{nicenumber} D^n \, - \, \frac{1}{m-2} \quad \, = \, \quad \frac{(m-1)^n-1}{m-2}, \end{equation} Therefore, when the tensor $A$ admits only finitely many equivalence classes of eigenpairs, then their number, counted with multiplicities, coincides with the positive integer in (\ref{nicenumber}). In Example~\ref{ex:diagonal} below we exhibit a tensor $A$ which attains the upper bound (\ref{nicenumber}). For that $A$, each solution $(x,\lambda)$ has multiplicity $1$. It follows that the same holds for generic $A$. \end{proof} \begin{remark} An alternative presentation of our proof is to perform the substitution $\lambda = \tilde\lambda^{m-2}$ in the equations $Ax^{m-1} = \lambda x$. This makes the system of equations homogeneous of degree~$m-1$. B\'ezout's theorem says that there are generically $(m-1)^n$ solutions in the projective space $\mathbb P^n$. If we remove the trivial solution, this leaves $(m-1)^n - 1$ solutions. Assuming that none of these have $\tilde\lambda = 0$, the orbits formed by multiplying $\tilde \lambda$ by $e^{2 \pi i/(m-2)}$ each yield the same value of $\lambda = \tilde \lambda^{m-2}$ and $x$. Thus, there are $((m - 1)^n - 1)/(m - 2)$ classes. The delicate point in such a proof would be to argue that the solution to $A x^{m-1} = \tilde \lambda^{m-2} x$ has multiplicity $m-2$ even when $\tilde \lambda = 0$. In effect, toric geometry conveniently does the bookkeeping in the correspondence between solutions to $A x^{m-1} = \lambda x$ and solutions to $A x^{m-1} = \tilde \lambda^{m-2} x$. \end{remark} \begin{ex}\label{ex:diagonal} Let $A$ be the diagonal tensor of order $m$ and size $n$ defined by setting $A_{i i \ldots i} = 1$ and all other entries zero. An eigenpair $(\lambda,x)$ is a solution to the equations \begin{equation} \label{eq:binomials} x_i^{m-1} \,\,=\,\, \lambda x_i \qquad \hbox{for $\,1 \leq i \leq n$.} \end{equation} All non-trivial solutions in $X = \mathbb P(1,\ldots,1,m-2)$ satisfy $\lambda \not= 0 $. By rescaling, we can assume that $\lambda = 1$. Fix the root of unity $\zeta = e^{2\pi i/(m-2)}$, and let $S = \{0, \ldots, m-3, *\}$. For any string $\sigma$ in $S^n$ other than the all $*$s string, we define $x_i = \zeta^{\sigma_i}$ if $\sigma_i$ is an integer and $x_i = 0$ if $\sigma_i = *$. This defines $(m-1)^n - 1$ eigenpairs. However, some of these are equivalent. Incrementing each integer in our string by $1$ modulo $m-2$ corresponds to multiplying our eigenvector by~$\zeta$. Thus, we have defined $((m-1)^n - 1)/(m-2)$ equivalence classes of eigenvalues and eigenvectors. These are all equivalence classes of solutions to (\ref{eq:binomials}). More generally, suppose that $A$ is a diagonal tensor with $A_{ii\ldots i}$ equal to some non-zero complex number~$a_i$. Then the eigenpairs are similarly given by $\lambda = 1$ and $x_i = a_i^{1/{m-2}}\zeta^{\sigma_i}$ or $x_i = 0$, as above, where $a_i^{1/{m-2}}$ is a fixed root of~$a_i$. In particular, for generic $a_i$ all $((m-1)^n - 1)/(m-2)$ eigenpairs will have distinct normalized eigenvalues. \qed \end{ex} From Theorem~\ref{thm:count}, we get the following result guaranteeing the existence of real eigenpairs. \begin{corollary}\label{cor:real-eig} If $A$ has real entries and either $m$ or $n$ is odd, then $A$ has a real eigenpair. \end{corollary} \begin{proof} When either $m$ or $n$ is odd, then one can check that the integer $\,((m-1)^n - 1)/(m-2)\,$ in Theorem~\ref{thm:count} is odd. This implies that $A$ has a real eigenpair by \cite[Corollary 13.2]{fulton-intersection}. \end{proof} Corollary~\ref{cor:real-eig} is sharp, in the sense that there exist real tensors with no real eigenpairs whenever both $m$ and~$n$ are even. We illustrate this in the following example. \begin{ex} Let $m$ be even, $n=2$, and $A$ the $2 {\times} \cdots {\times} 2$ tensor which is zero except for the entries $a_{12\cdots2} = 1$ and $a_{21\cdots1} = -1$. The eigenpairs of $A$ are the solutions to the equations: \begin{align*} x_2^{m-1} &\,=\, \lambda x_1 \\ -x_1^{m-1} &\,=\, \lambda x_2. \end{align*} Eliminating $\lambda$, we obtain $x_1^m + x_2^m = 0$, which has no non-zero real solutions for even~$m$. For $n$ any even integer, let $B$ be the tensor whose $n/2$ diagonal $2\times \cdots \times 2$ blocks are the tensor~$A$ above, and which is zero elsewhere. A non-trivial eigenpair must be an eigenpair for at least one of the blocks, and therefore cannot be real. Thus, $B$ has no real eigenpairs. \qed \end{ex} \section{Characteristic Polynomial and Singular Tensors}\label{sec:characteristic} The {\em characteristic polynomial} $\phi_A(\lambda)$ of a generic tensor $A$ was defined as follows in \cite{NQWW, Qi}. Consider the univariate polynomial in $\lambda $ that arises by eliminating the unknowns $x_1,\ldots,x_n$ from the system of equations $A x^{m-1} = \lambda x$ and $x \cdot x = 1$. If $m$ is even then this polynomial equals $\phi_A(\lambda)$. If $m$ is odd then this polynomial has the form $\phi_A(\lambda^2)$, i.e.\ the characteristic polynomial evaluated at $\lambda^2$. With these definitions, Theorem \ref{thm:count} implies the following: \begin{corollary} \label{cor:charpol} For a generic tensor $A$, the characteristic polynomial $\phi_A(\lambda)$ is irreducible and has degree $((m-1)^n-1)/(m-2)$. Hence this is the number of normalized eigenvalues. \end{corollary} For any particular tensor $A$, the {\em characteristic polynomial} $\phi_A(\lambda)$ is obtained by specializing the entries in the coefficients of the generic characteristic polynomial. Ni {\it et al.} \cite{NQWW} expressed $\phi_A(\lambda)$ as a Macaulay resultant, which implies a formula as a ratio of determinants. For the present work, we used Gr\"obner-based software to compute the characteristic polynomials of various tensors. It is tempting to surmise that all zeros of the characteristic polynomial $\phi_A(\lambda)$ are normalized eigenvalues of the tensor $A$. This statement is almost true, but not quite. There is some subtle fine print, to be illustrated by Example \ref{ex:fineprint} below. Qi \cite[Question 1]{Qi} asked whether the set of normalized eigenvalues of a tensor is either finite or all of~$\mathbb C$. We answer this question by showing a tensor where neither of these alternatives holds: \begin{ex} \label{ex:fineprint} Consider the complex $2 \times 2 \times 2$ tensor $A$ whose nonzero entries are \begin{equation*} a_{111} = a_{221} \, =\, 1 \quad \hbox{and} \quad a_{112} = a_{222} \,=\, i = \sqrt{-1} . \end{equation*} We claim that any complex number other than $0$ is a normalized eigenvalue of $A$, but $0$ is not a normalized eigenvalue. The equations for an eigenvalue and eigenvector of~$A$ are $$ x_1^2 + ix_1 x_2 \,= \, \lambda x_1 \quad \hbox{and} \quad x_1 x_2 + ix_2^2 \,= \, \lambda x_2. $$ For any $\lambda \neq 0$ we obtain a matching eigenvector that also satisfies $\,x \cdot x = 1\,$ by taking \begin{equation*} x = \left( \frac{\lambda^2 + 1}{2\lambda}, \,\frac{\lambda^2 - 1}{2 i \lambda} \right). \end{equation*} Hence $\lambda$ is a normalized eigenvalue. However, if $\lambda = 0$, then an eigenvector must satisfy $$ x_1^2 + i x_1 x_2 \, =\, 0 \quad \hbox{and} \quad x_1 x_2 + i x_2^2 \, = \, 0. $$ These imply that $\,x \cdot x = x_1^2 + x_2^2 \,$ is zero, so $\lambda = 0$ cannot be a normalized eigenvalue. \qed \end{ex} However, we have the following weaker statement: \begin{proposition}\label{prop:normalized-eigen} The set of normalized eigenvalues of a tensor is either finite or it consists of all complex numbers in the complement of a finite set. \end{proposition} \begin{proof} The set $\mathcal{E}(A)$ of normalized eigenvalues $\lambda$ of the tensor $A$ is defined by the condition \begin{equation*} \exists \,x \in \mathbb C^n \,\,: \,\,A x^{m-1} = \lambda x \,\,\, \hbox{and} \,\,\, x \cdot x = 1. \end{equation*} Hence $\mathcal{E}(A)$ is the image of an algebraic variety in $\mathbb C^{n+1}$ under the projection $(x,\lambda) \mapsto \lambda$. Chevalley's Theorem states that the image of an algebraic variety under a polynomial map is constructible, that is, defined by a Boolean combination of polynomial equations and inequations. We conclude that the set $\mathcal{E}(A)$ of normalized eigenvalues is a constructible subset of $\mathbb C$. This means that $\mathcal{E}(A)$ is either a finite set or the complement of a finite set. \end{proof} The relationship between the normalized eigenvalues and the characteristic polynomial is summarized in the following proposition. \begin{proposition} \label{prop:sing} For a tensor~$A$, each of the following conditions implies the next: \begin{enumerate} \item The set $\,\mathcal E(A)$ of all normalized eigenvalues consists of all complex numbers \item The set $\,\mathcal{E}(A)$ is infinite. \item The characteristic polynomial $\phi_A(\lambda)$ vanishes identically. \end{enumerate} \end{proposition} \begin{proof} Clearly, (1) implies (2). By the projection argument in the proof above, the zero set in $\mathbb C$ of the characteristic polynomial $\phi_A(\lambda)$ contains the set $\mathcal{E}(A)$. Hence (2) implies (3). \end{proof} From Example~\ref{ex:fineprint}, we see that (2) does not necessarily imply (1), and in Example~\ref{ex:dreizwei}, we shall see that (3) does not imply (2). Qi defines a singular tensor to be one for which the first statement of Proposition~\ref{prop:sing} holds. However, we suggest that the last condition is a better definition: a \emph{singular tensor} is a tensor such that $\phi_A(\lambda)$ vanishes identically. This definition has the advantage that the limit of singular tensors is again singular. In particular, the set of all singular tensors is a closed subvariety in the $n^m$-dimensional tensor space $\mathbb C^{n \times \cdots \,\times n}$. Its defining polynomial equations are the coefficients of the characteristic polynomial $\phi_A(\lambda)$ where $A$ is the tensor whose entries $a_{i_1 \cdots i_n}$ are indeterminates. \begin{example} \label{ex:dreizwei} Let $m = 3$ and $n=2$. Here $A = (a_{ijk})$ is a general tensor of format $2 \times 2 \times 2$. The characteristic polynomial $\phi_A$ is obtained by eliminating $x_1$ and $x_2$ from the ideal $$ \langle \, a_{111} x_1^2 + (a_{112} + a_{121}) x_1 x_2 + a_{122} x_2^2 - \lambda x_1\, , \, a_{211} x_1^2 + (a_{212} + a_{121}) x_1 x_2 + a_{222} x_2^2 - \lambda x_2 \,, \, x_1^2 + x_2^2 - 1 \, \rangle . $$ We find that $\phi_A$ has degree $3$, as predicted by Theorem \ref{thm:count}. Namely, the elimination yields $$ \phi_A(\lambda^2) \quad = \quad C_2 \lambda^6 \, + \,C_4 \lambda^4 \,+\, C_6 \lambda^2 + C_8, $$ where $C_i$ is a certain homogeneous polynomial of degree $i$ in the eight unknowns $a_{ijk}$. The set of singular $2 {\times} 2 {\times} 2$-tensors is the variety in the projective space $\mathbb P^7 = \mathbb P(\mathbb C^{2 \times 2 \times 2})$ given by \begin{equation} \label{eq:ideal} \langle C_2, C_4, C_6, C_8 \rangle \,\,\subset \,\, \mathbb C[a_{111},a_{112}, \ldots,a_{222}] \end{equation} This is an irreducible variety of codimension~$2$ and degree~$4$, but the ideal (\ref{eq:ideal}) is not prime. The constant coefficient $C_8$ is the square of a quartic. That quartic is the {\em Sylvester resultant} $$ {\rm Res}_x (A x^2) \quad = \quad {\rm det} \begin{pmatrix} \, a_{111} & a_{112} + a_{121} & a_{122} & 0 \, \\ \, 0 & a_{111} & a_{112} + a_{121} & a_{122} \, \\ \, a_{211} & a_{212} + a_{221} & a_{222} & 0 \, \\ \, 0 & a_{211} & a_{212} + a_{221} & a_{222} \, \end{pmatrix}. \qquad $$ The leading coefficient of the characteristic polynomial is a sum of squares $$ C_2 \, \,\,= \, \,\, (-a_{111} + a_{122} + a_{212} + a_{221} )^2 \,+\, ( a_{112} + a_{121} + a_{211} - a_{222})^2 $$ This indicates that among singular $2 {\times} 2 {\times} 2$-tensors those with real entries are scarce. Indeed, the real variety of (\ref{eq:ideal}) is the union of two linear spaces of codimension~$4$, with defining ideal \begin{equation*} \langle a_{122}, a_{211}, a_{112}+a_{121}-a_{222}, a_{212}+a_{221}-a_{111} \rangle \,\cap \, \langle a_{111} - a_{122}, a_{211} - a_{222}, a_{112} + a_{121}, a_{212} + a_{221} \rangle. \end{equation*} This explains why the singular tensor in Example \ref{ex:fineprint} had to have a non-real coordinate. We now look more closely at the real singular tensors defined by the second ideal in this intersection. These are tensors $A$ for which $\,Ax^2 = \bigl( a_{111}(x^2 + y^2) \,,\, a_{211}(x^2 +y^2) \bigr)$. It is easy to see that, so long as $a_{111}$ and~$a_{211}$ are not both zero, the only normalized eigenvector is \begin{equation*} \left( \frac{a_{111}}{\sqrt{a_{111}^2 + a_{211}^2}}, \frac{a_{211}}{\sqrt{a_{111}^2 + a_{211}^2}} \right), \mbox{ which has eigenvalue } \lambda = \sqrt{a_{111}^2 + a_{211}^2}. \end{equation*} In particular, the number of eigenvalues of such a tensor must be finite. This example shows that (3) does not imply (2) in Proposition~\ref{prop:sing}. \qed \end{example} The reader will not have failed to notice that the notion of ``singular'' used here (and in \cite{Qi}) is more restrictive than the one familiar from the classical case $m=2$. Indeed, a matrix is singular if it has $\lambda = 0$ as an eigenvalue, or, equivalently, if the constant term of the characteristic polynomial vanishes. That constant term is a power of the resultant $\,{\rm Res}_x(A x^{m-1})$, and its vanishing means that the homogeneous equations $A x^{m-1} = 0$ have a non-trivial solution $x \in \mathbb P^{n-1}$. This holds when the tensor $A$ is singular but not conversely. Yet another possible notion of singularity for a tensor $A$ arises from its {\em hyperdeterminant} ${\rm Det}(A)$, as defined in \cite{gkz}. For example, the hyperdeterminant of a $2 {\times} 2 {\times} 2 $-tensor equals $$ \begin{matrix} {\rm Det}(A) &=& a_{122}^2 a_{211}^2 +a_{121}^2 a_{212}^2 +a_{112}^2 a_{221}^2 +a_{111}^2 a_{222}^2 \qquad \qquad \qquad \qquad \qquad \qquad \qquad \\ & & - 2 a_{121} a_{122} a_{211} a_{212} -2 a_{112} a_{122} a_{211} a_{221} -2 a_{112} a_{121} a_{212} a_{221} -2 a_{111} a_{122} a_{211} a_{222} \\ & & - 2 a_{111} a_{121} a_{212} a_{222} - 2 a_{111} a_{112} a_{221} a_{222} +4 a_{111} a_{122} a_{212} a_{221} + 4 a_{112} a_{121} a_{211} a_{222} . \end{matrix} $$ The hyperdeterminant vanishes if the hypersurface defined by the multilinear form associated with $A$ has a singular point in $(\mathbb P^{n-1})^m$. This property is unrelated to the coefficients of the characteristic polynomial $\phi_A(\lambda)$. In particular, $\,{\rm Det}(A) \not= {\rm Res}_x(A x^{m-1}) $. For an example take the $2 {\times} 2 {\times} 2$-tensor $A$ all of whose entries are $a_{ijk} = 1$ is not singular but ${\rm Det}(A) = 0$. On the other hand, the following tensor $A$ is singular but has ${\rm Det}(A) = -1$: \begin{equation*} a_{111} = -1, \, a_{112} = 0, \, a_{121} = 0, \, a_{122} = -1, \,\, a_{211} = 1,\, a_{212} = -1, \, a_{221} = 0, \, a_{222} = 1-i. \end{equation*} This highlights the distinction between our setting here and that in \cite[Proposition 2]{Lim}. \section{Dynamics on Projective Space}\label{sec:dynamics} The purpose of this short section is to point out a connection to dynamical systems. Dynamics on projective space is a well-established field of mathematics \cite{BJ, ivashkovich}. We believe that the interpretation of eigenpairs of tensors in terms of fixed points on $\mathbb P^{n-1}$ could be of interest to applied mathematicians as a new tool for modeling and numerical computations. We consider the map $\psi_A$ defined by the formula $\psi_A(x) = Ax^{m-1}$. This is a rational map from complex projective space $\mathbb P^{n-1}$ to itself. The fixed points of the map $\psi_A \colon \mathbb P^{n-1} \dashrightarrow \mathbb P^{n-1}$ are exactly the eigenvectors of the tensor $A$ with non-zero eigenvalue, and the base locus of $\psi_A$ is the set of eigenvectors with eigenvalue zero. In particular, the map $\psi_A$ is defined everywhere if and only if $0$ is not an eigenvalue of~$A$. Note that every such rational map arises from some tensor~$A$, but the tensor is not unique. Indeed, $A$ has $n^m$ entries while the map is determined by $n$ polynomials, which have only $n \binom{n +m -2}{ m-1}$ distinct coefficients. For instance, in Example~\ref{ex:dreizwei}, with $m=3,n=2$, the eight entries of the tensor translate into six distinct coefficients of the two binary quadrics that specify the self-map of the projective line $\,\psi_A\colon \mathbb P^1 \dashrightarrow \mathbb P^1$. It is instructive to revisit the classical case $m=2$, where $\psi_A\colon \mathbb P^{n-1} \dashrightarrow \mathbb P^{n-1}$ is a linear map. The condition that every eigenvalue of the matrix~$A$ is zero is equivalent to saying that $A$ is {\em nilpotent}, that is, some matrix power of~$A$ is zero. Geometrically, this means that some iterate of the rational map $\psi_A$ is defined nowhere in projective space $\mathbb P^{n-1}$. We use the same definition for tensors: A is {\em nilpotent} if some iterate of $\psi_A$ is nowhere defined. \begin{proposition} If the tensor $A$ is nilpotent then $0$ is the only eigenvalue of $A$. The converse is not true: there exist tensors with only eigenvalue $0$ that are not nilpotent. \end{proposition} \begin{proof} Suppose $\lambda \not= 0 $ is an eigenvalue and $x \in \mathbb C^n\backslash \{0\}$ a corresponding eigenvector. Then $x$ represents a point in $\mathbb P^{n-1}$ that is fixed by $\psi_A$. Hence it is fixed by every iterate $\psi_A^{(r)}$ of $\psi_A$. In particular, $\psi_A^{(r)}$ is defined at (an open neighborhood) of $x \in\mathbb P^{n-1}$, and $A$ is not nilpotent. Let $A$ be the $2 {\times} 2 {\times} 2$-tensor with $a_{111} = a_{211} = a_{212} = 1$ and the other five entries zero. The eigenpairs of $A$ are the solutions to $\, x_1^2 \, = \, \lambda x_1 \,$ and $\, x_1^2 + x_1x_2 \, = \, \lambda x_2 $. Up to equivalence, the only eigenpair is $x = (0, 1)$ and $\lambda = 0$. However, the self-map $\psi_A$ on $\mathbb P^1$ is dominant. To see this, note that $\psi_A$ acts by translation on the affine line $\,\mathbb A^1 = \{ x_1 \neq 0\}$ because $(x_1^2: x_1^2 + x_1x_2) = (x_1: x_1+x_2)$. All iterates of $\psi_A$ are defined on $\mathbb A^1$, i.e.~there are no base points with $x_1 \not= 0$, and hence $A$ is not nilpotent. \end{proof} The example in the previous proof works because the two binary quadrics in $Ax^{m-1}$ have $x_1$ as a common factor. Indeed, whenever $n=2$, an eigenvector~$x$ has eigenvalue zero if and only if $x$ is a solution to a common linear factor of the two binary forms of $Ax^{m-1}$. However, for $n \geq 3$, this is no longer true. Work in dynamics by Ivashkovic \cite[Theorem~1]{ivashkovich} implies that, for $n=3$, one can construct tensors~$A$ such that zero is the only eigenvalue, the polynomials in $Ax^{m-1}$ have no common factors, but $A$ is not nilpotent. \begin{example} This example is taken from \cite[Example~4.1]{ivashkovich}. Let $m=n=3$ and take $A$ to be any tensor whose corresponding map is the Cremona transformation $$ \psi_A \colon \mathbb P^2 \dashrightarrow \mathbb P^2 \,:\, (x_1,x_2,x_3) \mapsto (x_1 x_2, x_1 x_3, 2 x_2 x_3).$$ This map has no fixed points, but it is not nilpotent. The base locus of $\psi_A$ consists of the three points $(1,0,0)$, $(0,1,0)$, and~$(0,0,1)$, and, up to scaling, these are the only eigenvectors of~$A$, all with eigenvalue~$0$. \qed \end{example} \section{Symmetric Tensors}\label{sec:symmetric} Of particular interest in numerical multilinear algebra is the situation when the tensor $A$ is symmetric and has real entries. Here $A$ being {\em symmetric} means that the entries $a_{i_1 i_2 \cdots i_n}$ are invariant under permuting the $n$ indices $i_1, i_2,\ldots,i_n$. Each symmetric $n {\times} \cdots {\times} n$ tensor $A$ of order~$m$ corresponds to a unique homogeneous polynomial~$f(x)$ of degree~$m$ in~$n$ unknowns. The symmetric case is of interest because a real polynomial~$f(x)$ of even degree~$m$ is positive semidefinite if and only if every real eigenpair of the corresponding symmetric tensor~$A$ has non-negative eigenvalue~\cite[Theorem~5(a)]{Qi-sym}. This is illustrated in Example~\ref{ex:motzkin}. In the notation of \cite{Qi-sym}, the relation between the tensor and the polynomial is written as \begin{equation} \label{eqn:iswrittenas} Ax^m \,\,=\,\, mf(x) \quad \hbox{and} \quad A x^{m-1} \,\, =\,\, \nabla f(x) , \end{equation} where $Ax^m$ is defined to be \begin{equation*} \sum_{i_1 = 1}^n \sum_{i_2 = 1}^n \cdots \sum_{i_m = 1}^n a_{i_1 \ldots i_m} x_{i_1} x_{i_2} \cdots x_{i_m} = x \cdot Ax^{m-1}. \end{equation*} The first equation in~(\ref{eqn:iswrittenas}) follows from the second because $x \cdot \nabla f(x) = mf(x)$. Note that the second equation in~(\ref{eqn:iswrittenas}) says that the coordinates of the gradient of $f(x)$ are precisely the entries of $A x^{m-1}$. The gradient $\nabla f(x)$ vanishes at a point $x$ in $\mathbb P^{n-1}$ if and only if $x$ is a singular point of the hypersurface in $\mathbb P^{n-1}$ defined by the polynomial $f(x)$. This implies: \begin{corollary} \label{cor:sing2} The singular points of the projective hypersurface $\{x \in \mathbb P^{n-1}:f(x) = 0\}$ are precisely the eigenvectors of the corresponding symmetric tensor $A$ which have eigenvalue~$0$. \end{corollary} The other eigenvectors of $A$ can also be characterized in terms of the polynomial $f(x)$. \begin{proposition} Fix a non-zero $\lambda$ and suppose $m \geq 3$. Then $x \in \mathbb C^n$ is a normalized eigenvector with eigenvalue $\lambda$ if and only if $x$ is a singular point of the affine hypersurface defined by the polynomial \begin{equation}\label{eqn:eig-char} f(x) - \frac{\lambda}{2} x \cdot x - \left(\frac{1}{m} - \frac{1}{2}\right) \lambda. \end{equation} \end{proposition} \begin{proof} The gradient of~(\ref{eqn:eig-char}) is $\nabla f - \lambda x = Ax^{m-1} - \lambda x$, so every singular point $x$ is an eigenvector with eigenvalue~$\lambda$. Furthermore, if we substitute $f(x) = \frac{1}{m} x \cdot \nabla f = \frac{\lambda}{m} x\cdot x\,$ into (\ref{eqn:eig-char}), then we obtain $x\cdot x = 1$. This argument is reversible: if $x$ is a normalized eigenvector of $A$ then $x \cdot x = 1$ and $\nabla f(x) = \lambda x$, and this implies that (\ref{eqn:eig-char}) and its derivatives vanish. \end{proof} \begin{corollary} The characteristic polynomial $\phi_A(\lambda)$ is a factor of the discriminant of (\ref{eqn:eig-char}). \end{corollary} Here we mean the classical multivariate discriminant \cite{gkz} of an inhomogeneous polynomial of degree $m$ in $n$ variables $x$ evaluated at (\ref{eqn:eig-char}), where $\lambda$ is regarded as a parameter. Besides the characteristic polynomial $\phi_A(\lambda)$, this discriminant may contain other irreducible factors. \begin{example}[\em Discriminantal representation of the characteristic polynomial of a symmetric tensor] If $n=2$ and $m=3$ then the discriminant at bivariate cubic in (\ref{eqn:eig-char}) equals $\,\lambda^4 \cdot \phi_A(\lambda)$. If $n=2$ and $m=4$ then we evaluate the discriminant of the ternary quartic using Sylvester's formula \cite[\S 3.4.D]{gkz}. The output has the discriminant of binary quartic as an extraneous factor: $$ \hbox{\rm Discriminant of (\ref{eqn:eig-char})} \quad = \quad \bigl(\phi_A(\lambda) \bigr)^2 \cdot \lambda^9 \cdot \hbox{\rm Discriminant of $f(x)$} $$ It would be interesting to determine the analogous factorization for arbitrary $m$ and $n$. \qed \end{example} \smallskip The subject of this paper is the number of normalized eigenvalues of a tensor. In Section~2 we gave an upper bound for that number under the hypothesis that the number is finite. Remarkably, this hypothesis is not needed if we restrict our attention to symmetric tensors. \begin{theorem}\label{thm:finite-norm-eig} Every symmetric tensor $A$ has at most $((m-1)^n-1)/(m-2)$ distinct normalized eigenvalues. This bound is attained for generic symmetric tensors $A$. \end{theorem} \begin{proof} It suffices to show that the number of normalized eigenvalues of \underbar{every} symmetric tensor $A$ is finite. Recall from the proof of Theorem~\ref{thm:count} that the set of eigenpairs is the intersection of $n$ linearly equivalent divisors on a weighted projective space. Since these divisors are ample, each connected component of the set of eigenpairs contributes at least one to the intersection number. Therefore, the number of connected components of eigenpairs can be no more than $((m-1)^n - 1)/(m-2)$. We conclude that the number of normalized eigenvalues of $A$, if finite, must be bounded above by that quantity as well. Finally, Example~\ref{ex:diagonal} shows that the bound is tight. We now prove that the number of normalized eigenvalues of a symmetric tensor $A$ is finite. Let $S$ be the affine hypersurface in $\mathbb C^n$ defined by the equation $x_1^2 + \cdots + x_n^2 = 1$. We claim that a point $x \in S$ is an eigenvector of~$A$ if and only if $x$ is a critical point of $f$ restricted to~$S$, in which case, the corresponding eigenvalue $\lambda$ equals $\frac{1}{m}f(x)$. By definition, a point $x \in S$ is a critical point of $f \vert_S$ if and only if the gradient $\nabla (f \vert_S)$ is zero at~$x$. The latter condition is equivalent to the gradient $\nabla f$ being a multiple of $\nabla(x_1^2 + \cdots + x_n^2 - 1) = 2x$. This is exactly the definition of an eigenvector. Finally, if $x\in S$ is a critical point of $f \vert_S$, then $mf(x) = x \cdot \nabla f(x) = \lambda x \cdot x = \lambda$, and hence $\,\lambda = \frac{1}{m}f(x)$. Finally, to prove Theorem~\ref{thm:finite-norm-eig}, we note that, by generic smoothness~\cite[Cor. III.10.7]{Hartshorne}, a polynomial function on a smooth variety has only finitely many critical values. Equivalently, Sard's theorem in differential geometry says that the set of critical values of a differentiable function has measure zero, so, by Proposition~\ref{prop:normalized-eigen}, that set must be finite. \end{proof} We note two subtleties about Theorem~\ref{thm:finite-norm-eig}. First, it does not imply that the characteristic polynomial of every symmetric tensor is non-trivial. Second, the result is intrinsically tied to the normalization $x \cdot x = 1$. We begin with an example of the first. \begin{example} Let $A$ be the symmetric $2 \times 2 \times 2$ tensor with \begin{equation*} a_{111} = -2i \,,\quad a_{112} = a_{121} = a_{211} = 1 \,,\quad a_{122} = a_{212} = a_{221} = 0 \,, \quad a_{222} = 1. \end{equation*} Then, up to equivalence, the only eigenvectors are $(0, 1)$ with eigenvalue~$1$ and $(1, i)$ with eigenvalue~$0$. Note that the second cannot be rescaled to be a normalized eigenvector, so the only normalized eigenvalue is~$1$. However, the characteristic polynomial of~$A$ is identically zero. The reason is that, for a small perturbation of $A$, the perturbation of the eigenvector~$(1,i)$ can take on any given normalized eigenvalue. \qed \end{example} No analogue of Theorem~\ref{thm:finite-norm-eig} is possible with the alternative normalization of requiring $x \cdot \overline x = 1$. In this case, each equivalence class yields infinitely many eigenvalues, which nonetheless have the same magnitude. However, the following example shows that the magnitudes of the eigenvalues with $x \cdot \overline x = 1$ may still be an infinite set. \begin{example} Let $A$ be the symmetric $3 {\times} 3 {\times} 3$ tensor whose non-zero entries are \begin{equation*} a_{111} = 2 \quad \hbox{and} \quad a_{122} = a_{212} = a_{221} \,=\, a_{133} = a_{313} = a_{331} \,=\, 1. \end{equation*} The eigenpairs of $A$ are the solutions to the equations \begin{align*} 2 x_1^2 + x_2^2 + x_3^2 &\,=\, \lambda x_1, \\ 2 x_1 x_2 &\,=\, \lambda x_2, \\ 2 x_1 x_3 &\,=\, \lambda x_3. \end{align*} For any $\alpha \in \mathbb C$, the vector $\,x = (1, i \alpha, \alpha)$ is an eigenvector with eigenvalue $\lambda = 2$. Rescaling, $x/\sqrt{x \cdot \overline x}$ is an eigenvector with unit length and eigenvalue \begin{equation*} \frac{2}{\sqrt{1 + 2 \lvert \alpha \rvert}}. \end{equation*} The magnitude of this eigenvalue can be any real number in the interval $(0,2\,]$. Note that the family of eigenvectors above all have $x \cdot x = 1$, so $\lambda = 2$ is the only normalized eigenvalue. \qed \end{example} One application of eigenvalues of symmetric tensors is that these can be used to decide whether a polynomial $f$ is {\em positive semidefinite}, i.e., whether $f(x) \geq 0$ for all $x \in \mathbb R^n$. \begin{example}\label{ex:motzkin} The {\em Motzkin polynomial} $f(x,y,z) = z^6 + x^4 y^2 + x^2 y^4 - 3 x^2 y^2 z^2$ is a well-known example of a positive semidefinite polynomial which cannot be written as a sum of squares. Let $A$ be the corresponding $3 {\times} 3 {\times} 3 {\times} 3 {\times} 3 {\times} 3 $-tensor. This tensor has $25$ eigenvalues, counting multiplicities, six less than our upper bound of $31$. Disregarding multiplicities, there are only four distinct eigenvalues. All four are real and they are equal to: $0$ (with multiplicity~$14$), $3/32$ (with multiplicity~$8$), $3/2$ (with multiplicity~$2$), and $6$ (with multiplicity~$1$). By~\cite[Theorem 5(a)]{Qi-sym}, this confirms the fact that the Motzkin polynomial $f$ is positive semidefinite. \qed \end{example} \bigskip {\bf Acknowledgments.} We thank Tamara Kolda for inspiring this project, with a question she asked us at the {\em Berkeley Optimization Day} on March 6, 2010. Both authors were supported in part by the National Science Foundation (DMS-0456960 and DMS-0757207). \bigskip

https://arxiv.org/abs/1509.07908

Helly-type theorems for the diameter

We study versions of Helly's theorem that guarantee that the intersection of a family of convex sets in $R^d$ has a large diameter. This includes colourful, fractional and $(p,q)$ versions of Helly's theorem. In particular, the fractional and $(p,q)$ versions work with conditions where the corresponding Helly theorem does not. We also include variants of Tverberg's theorem, Bárány's point selection theorem and the existence of weak epsilon-nets for convex sets with diameter estimates.

\section{Introduction} Quantitative results in combinatorial geometry have recently caught new interest. Those surrounding Helly's theorem have as aim to show that \textit{given a finite family of convex set in $\mathds{R}^d$, if the intersection of every small subfamily is large, then the intersection of the whole family is also large} \cite{Amenta:2015tp}. When the size of a convex set is measured according to a function that varies discretely, such as the number of points in a lattice, there are very sharp results (e.g. \cite{Aliev:2014va}). However, when the size of a convex set is measured according to a function that varies continuously, such as the volume or diameter, the behaviour changes considerably. The first results of this kind were presented by B\'ar\'any, Katchalski and Pach \cite{Barany:1982ga, Barany:1984ed}, where Helly-type theorems were made regarding the volume and diameter of the intersection of families of convex sets. They showed that, given a finite family of convex sets in $\mathds{R}^d$, if the intersection of every $2d$ of them has volume at least one, one can obtain lower bounds on the volume of the intersection, and the same holds for the diameter. The constant $2d$ is optimal, but the downside is that the guarantee of the diameter or volume of the intersection decreases quickly with the dimension. Helly's theorem has an impressive number of variations and generalisations (see, for instance, the surveys \cite{Danzer:1963ug,Eckhoff:1993uy, Matousek:2002td, Wenger:2004uf, Amenta:2015tp}). Thus, it is natural to determine which results can be extended in this quantitative framework. For the volume, several advances have been made in this direction \cite{Naszodi:2015vi, DeLoera:2015wp, Soberon:2015tsa}. This includes optimising the original result by B\'ar\'any, Katchalski and Pach, and finding colourful versions, fractional versions and $(p,q)$ type theorems. These are classical variations of Helly's theorem found in \cite{Barany:1982va}, \cite{Katchalski:1979bq} and \cite{Alon:1992gb}, respectively. The aim of this paper is to present analogues to these results for the diameter. For example, the guarantee of the size of the intersection can be improved if we are willing to check larger families. Regarding the diameter, the following result makes this clear. \begin{theoremp}[Helly's theorem for diameter, De Loera et al. {\cite[Thm 1.5]{DeLoera:2015wp}}] Let $d$ be a positive integer and $1>\delta>0$, then, there is an integer $n=n(\diam, d, \delta)$ such that for any finite family $\mathcal{F}$ of convex sets in $\mathds{R}^d$, if the intersection of every subfamily of size $n$ has diameter greater than or equal to one, then $\diam (\cap \mathcal{F}) \ge 1- \delta$. Moreover, $n(\diam, d, \delta) = \Omega_d (\delta^{-(d-1)/2})$. \end{theoremp} Given two functions $g(d,\delta)$ and $f(d,\delta)$, we say $g(d,\delta)=\Omega_d (f(\delta))$ if, for any fixed $d$, $g(d,\delta) = \Omega (f(\delta))$, and similarly with other notation for asymptotic bounds. For an upper bound to the result above, one can apply the main theorem of \cite{Langberg:2009go} to get $n(\diam, d, \delta) = O(\delta^{-d/2})$. The equivalent result for volume {\cite[Thm 1.4]{DeLoera:2015wp}} has similar upper and lower bounds in terms of $\delta$, giving $n(\vol, d, \delta)= \Theta_d (\delta^{-(d-1)/2})$. In the same spirit as Lov\'asz's generalisation of Helly's theorem \cite{Barany:1982va}, we show a ``colourful version'' of De Loera et al.'s diameter Helly in Section \ref{section-colourful}. Moreover, Theorem \ref{theorem-colourful-diameter} implies an asymptotically optimal bound for the diameter as well: $n(\diam, d, \delta) = \Theta_d (\delta^{-(d-1)/2})$. Asymptotic result as above hold for a very general family of functions, and are closely related to the approximability of convex sets by polyhedra \cite{Amenta:2015tp}. If we allow for a loss of diameter $\delta$, other versions of Helly's theorem can be recreated. In particular, we show a version of Alon and Kleitman's $(p,q)$ theorem \cite{Alon:1992gb} for the diameter. The $(p,q)$ theorem was conjectured originally by Hadwiger and Gr\"unbaum \cite{Hadwiger:1957we}, and asks if a slight weakening of Helly's condition is still enough information to bound the number of points needed to intersect all members of a finite family of convex sets in $\mathds{R}^d$. A lucid description of the theorem and its variations is contained in a survey by Eckhoff \cite{Eckhoff:2003ed}, and more recent results are summarised in \cite{Amenta:2015tp}. \begin{theorem}[$(p,q)$ theorem for diameter]\label{theorem-p,q-diameter} Let $p \ge q \ge 2d$ be positive integers and $1 > \delta > 0$. Then, there is a $c = c(p,q,d,\delta)$ such that for any finite family $\mathcal{F}$ of at least $p$ convex sets in $\mathds{R}^d$ of diameter at least one each, if out of every $p$ sets in $\mathcal{F}$, there are $q$ of them whose intersection has diameter at least one, then we can find $c$ convex sets $K_1, K_2, \ldots, K_c$ of diameter at least $1-\delta$ such that every set in $\mathcal{F}$ contains at least one $K_i$. \end{theorem} In the volumetric version \cite[Thm. 1.2]{Soberon:2015tsa}, the lower bound on $q$ depends on heavily on $\delta$. The proof of the original $(p,q)$ theorem is a tour de force of combinatorial geometry, and requires many classic results. In order to prove Theorem \ref{theorem-p,q-diameter}, we give diameter version of these as well. Notice that if one stubbornly refuses to check large families as Helly's theorem for the diameter requires, the result above gives non-trivial consequences with the same condition as B\'ar\'any, Katchalski and Pach used. \begin{corollary}\label{corollary-chido} Let $d$ be a positive integer and $1 > \delta > 0$. Let $\mathcal{F}$ be a finite family of convex sets in $\mathds{R}^d$ such that the intersection of every $2d$ of them has diameter greater than or equal to one. Then, $\mathcal{F}$ may be split into $c(2d,2d,d,\delta)$ parts such that the diameter of the intersection of each of them is at least $1-\delta$. \end{corollary} The loss of diameter $\delta$ is necessary in Theorem \ref{theorem-p,q-diameter} and Corollary \ref{corollary-chido}. This is shown in section \ref{section-remarks} with a construction. The rest of the paper is organised as follows. In section \ref{section-width} we show Helly-type results for the property \textit{having $v$-width at least one}, for some fixed direction $v$. In section \ref{section-colourful} we show a colourful version of Helly's theorem for the diameter. In section \ref{section-fractional} we prove diameter versions of the fractional Helly theorem \cite{Katchalski:1979bq}, Tverberg's theorem \cite{Tverberg:1966tb}, B\'ar\'any's selection theorem (sometimes called the ``first selection lemma'') \cite{Barany:1982va} and the existence of weak $\varepsilon$-nets for convex sets \cite{Alon:2008ek} in order to prove Theorem \ref{theorem-p,q-diameter}. Finally, in section \ref{section-remarks} we include some remarks and open problems. \section{Results for fixed direction width}\label{section-width} If instead of looking at the diameter, one is interested in the width in a fixed direction $v$, we can obtain similar Helly-type to the ones mentioned in the introduction. The original proofs for Helly-type theorems can be translated to this setting with minimal effort. However, since some of them are useful for the results regarding the diameter, we present them completely here. Let $v$ be a unit vector in $\mathds{R}^d$. Given a compact convex set $K \subset \mathds{R}^d$, we say that $p \in K$ is a $v$-directional minimum if $\langle v, p \rangle \le \langle v, x\rangle$ for all $x \in K$, where $\langle \cdot, \cdot \rangle$ denotes the usual dot product. We define a $v$-directional maximum similarly. Throughout the rest of the paper we will assume that all convex sets we work with are compact and their boundary contains no segments. This guarantees that the $v$-directional minimums for the sets and their non-empty finite intersections exist and are unique. Standard approximation techniques show that there is no loss of generality. Given a compact convex set $K$, we define its $v$-width as $\langle q, v\rangle - \langle p, v\rangle$ where $q, p$ are its $v$-directional maximum and $v$-direction minimum, respectively. \begin{theorem}[Helly for $v$-width]\label{theorem-v-width-basic-helly} Let $v$ be a unit vector in $\mathds{R}^d$ and $\mathcal{F}$ be a finite family of convex sets in $\mathds{R}^d$ such that the intersection of every $2d$ sets of $\mathcal{F}$ has a $v$-width greater than or equal to one. Then, the $v$-width of $\cap \mathcal{F}$ is greater than or equal to one. \end{theorem} \begin{proof} Let $A$ be a subfamily of size $d$ whose $v$-directional minimum $p$ maximizes $\langle p, v\rangle$. Given any other set $K _0 \in \mathcal{F}$, let us show $p \in K_0$. We know that $A \cup \{K_0\}$ must be intersecting, so let $u$ be a point of the intersection. The minimality of $p$ implies $\langle u, v\rangle \ge \langle p, v\rangle$. If we denote by $K_1, K_2, \ldots, K_d$ the sets in $A$, every $d$-tuple of $A \cup \{K_0\}$ must be intersecting. We call $p_i$ the $v$-directional minimum of $(A \cup \{K_0\})\setminus \{K_i\}$. We know that $\langle p_i, v\rangle \le \langle p, v\rangle$ for each $i$ (also notice $p_0 = p$). By convexity, there is a point $u_i \in (A \cup \{K_0\})\setminus \{K_i\}$ such that $\langle u_i, v \rangle = \langle p, v \rangle$ for each $i$. This gives us $d+1$ points in the hyperplane $\{y : \langle y , v\rangle = \langle p, v \rangle \}$, of dimension $d-1$. By Radon's lemma \cite{Radon:1921vh}, these points can be partitioned into two sets $B, C$ such that $\conv (B) \cap \conv (C) \neq \emptyset$. Let $p'$ be a point in $\conv (B) \cap \conv (C) \neq \emptyset$. It is immediate that $p' \in K_i$ for all $0 \le i \le d$. Thus, $p = p'$ and we have $p \in K_0$, as desired. Let $B$ be a subfamily of size $d$ with minimal $v$-directional maximum $q$. Again, every set in the family contains $q$. Since the $v$-width of $A \cup B$ is at least one, and this is realised by the segment $[p,q]$, the $v$-width of $\cap \mathcal{F}$ is at least one. \end{proof} \begin{theoremp}[Colourful Carath\'eodory for two points]\label{theorem-colourful-caratheodory} Given $2d$ sets of points $S_1, S_2, \ldots, S_{2d}$, and a set $S$ of two points $x,y$ such that $\{x,y\} \subset \conv (S_i)$ for each $1 \le i \le 2d$, there is a choice of points $s_1 \in S_1, \ldots, s_{2d} \in S_{2d}$ such that $\{x,y\} \in \conv \{s_1, \ldots, s_{2d}\}$ \end{theoremp} This is a particular case of Theorem 1.3 in \cite{DeLoera:2015wp}. \begin{theorem}[Colourful Helly for $v$-width] Let $\mathcal{F}_1, \mathcal{F}_2, \ldots, \mathcal{F}_{2d}$ be finite families of convex sets in $\mathds{R}^d$, considered as colour classes. If the $v$-width of the intersection of every rainbow choice $F_1 \in \mathcal{F}_1, \ldots, F_{2d} \in \mathcal{F}_{2d}$ is at least one, then there is a colour class $\mathcal{F}_i$ such that the $v$-width of $\cap \mathcal{F}_i$ is at least one. \end{theorem} \begin{proof} A systematic way to obtain a colourful Helly theorem from a ``monochromatic'' version was presented in \cite[Thm. 5.3]{DeLoera:2015tc}. Thus, it is sufficient to check the conditions of that result. Let $\mathcal{P}(K)$ stand for \textit{``$K$ has $v$-width at least one''}. Then, the following properties are satisfied. \begin{itemize} \item $\mathcal{P}$ is a Helly property (Theorem \ref{theorem-v-width-basic-helly}), with Helly number $2d$. \item $\mathcal{P}$ is a monotone property. i.e. if $K \subset K'$ then $\mathcal{P}(K)$ implies $\mathcal{P}(K')$. \item Let $v'$ be a unit vector in $\mathds{R}^d$ sufficiently close to $v$, but different. We consider a $v'$-semispace a set of the form $\{x \in \mathds{R}^d : \langle x, v' \rangle \le \alpha \}$ for some $\alpha$. Then, for every compact convex $K$ without segments in its boundary such that $\mathcal{P}(K)$ holds, there is a containment minimal $v'$-semispace $H$ such that $\mathcal{P}(K \cap H)$ holds. Moreover, if we denote by $p, q$ the $v$-directional minimum and $v$-directional maximum of $K \cap H$ (which exist and are unique since $v' \neq v$), then every closed convex subset $K' \subset K \cap H$ with $\mathcal{P'}$ satisfies that $\conv\{p,q\} \subset K'$. \end{itemize} Having these properties, \cite[Thm. 5.3]{DeLoera:2015tc} implies the result we were seeking. \end{proof} The same idea leads to a fractional version. \begin{theorem}\label{theorem-helly-v-width} Let $\alpha >0 $, $d$ a positive integer and $v$ a unit vector in $\mathds{R}^d$. Then, there is a positive constant $\beta$ depending only on $\alpha, d$ such that for any family $\mathcal{F}$ of $n$ convex sets in $\mathds{R}^d$ such that the intersection of at least $\alpha {{n}\choose{2d}}$ of the $2d$-tuples has $v$-width greater than or equal to one, there is a subfamily $\mathcal{F}'$ of $\mathcal{F}$ of cardinality at least $\beta n$ such that its intersection has $v$-width at least one. \end{theorem} \begin{proof} Let $v' \neq v$ be a unit vector in $\mathds{R}^d$ sufficiently close to $v$. For each subfamily $A$ of $\mathcal{F}$ of cardinality $2d-1$ whose intersection has $v$-width at least one, let $H_A$ be the containment-minimal $v'$-semispace such that $\cap (A \cup \{H_A\})$ has $v$-width at least one. Notice that if we consider $p_A, q_A$ the $v$-directional maximum and minimum of $\cap (A \cup \{H_A\})$ respectively (which exist and are unique since $v' \neq v$), then every convex set $C$ of $v$-width at least one such that $C \subset \cap (A \cup \{H_A\})$ satisfies $\{p_A, q_A\} \subset C$. For each subfamily $B$ of $2d$ sets of $\mathcal{F}$ whose intersection has $v$-width greater than or equal to one, let $A$ be its subfamily of size $2d-1$ with containment-maximal $H_A$. Then, Theorem \ref{theorem-v-width-basic-helly} implies that the intersection of $B \cup \{H_A\}$ has $v$-width at least one, so all the sets in $B$ contain $\{p_A, q_A\}$ Consider the function that assigns to each $2d$-tuple $B$ with $v$-width at least one a $(2d-1)$-tuple $A$ as above. Since a positive fraction of the $2d$-tuples satisfy this property, a direct double counting argument shows that there is a $(2d-1)$-tuple $A_0$ which was assigned at least $\beta n$ times for some positive $\beta$ depending only $d$ and $\alpha$. Thus, at least $\beta n$ sets contain $\{p_{A_0}, q_{A_0}\}$, as desired. \end{proof} With the results above, the methods in Section \ref{section-fractional} can be carried out verbatim to the $v$-width case to prove the following result. \begin{theorem}[$(p,q)$ theorem for $v$-width] Let $p \ge q \ge 2d$ be positive integers and $v$ a unit vector in $\mathds{R}^d$. Then, there is a $c' = c'(p,q,d)$ such that for any finite family $\mathcal{F}$ of at least $p$ convex sets of $v$-width at least one each, if out of every $p$ sets in $\mathcal{F}$, there are $q$ of them whose intersection has $v$-width at least one, then we can find $c'$ convex sets $K_1, K_2, \ldots, K_{c'}$ of $v$-width at least one such that every set in $\mathcal{F}$ contains at least one $K_i$.\end{theorem} \section{Colourful Helly for the diameter}\label{section-colourful} \begin{theorem}\label{theorem-colourful-diameter} There is an $n'=n'\left(\diam, d, \delta\right)$ such that for any $n'$ finite families $\mathcal{F}_1, \mathcal{F}_2, \ldots, \mathcal{F}_{n'}$ of convex sets in $\mathds{R}^d$, considered as colour classes, if the intersection of every colourful choice $F_1 \in \mathcal{F}_1, \ldots, F_{n'} \in \mathcal{F}_{n'}$ has diameter at least one, then there is a colour class $\mathcal{F}_i$ with $\diam (\cap \mathcal{F}_i) \ge 1- \delta$. Moreover, $n'(\diam , d, \delta ) = \Theta_d \left(\delta^{-(d-1)/2}\right)$. \end{theorem} If $\mathcal{F}_1 = \ldots = \mathcal{F}_{n'}$, we obtain the monochromatic result, and the upper bound matches the one mentioned in the introduction. The equivalent result for the volume \cite[Thm. 1.5]{Soberon:2015tsa} has a worse bound $n'(\vol, d, \delta) = O_d(\delta^{-(d^2-1)/4})$. \begin{proof} Given a point $x \in S^{d-1}$, we denote by $C_{\delta}(x)$ the cap \[ C_{\delta}(x):= \left\{y \in S^{d-1}: \langle x, y \rangle \ge 1-{\delta} \right\}. \] We denote by $c_{\delta}$ its measure under the usual probability Haar measure of $S^{d-1}$. It is known that $c_{\delta} = \Omega (\delta^{(d-1)/2})$ {\cite[Lemma 2.3]{ball1997elementary}}. Let $m= \left\lfloor\frac{1}{c_{\delta / 4}}\right\rfloor$ and consider $n' = 2dm$. Assume that $\diam \mathcal{F}_i < 1-\delta$ for all $i$. We look for a contradiction. We can find $v_1, v_2, \ldots, v_{m}$ directions in $S^{d-1}$ such that for any $v \in S^{d-1}$, there is a $v_j$ with $\langle v, v_j \rangle \ge 1-\delta$. In order to see this, take a set of points in $S$ of maximal cardinality such the caps $C_{\delta /4} (x)$ for $x \in S$ are have pairwise disjoint interiors. By counting surface area one gets $|S| \le \lfloor{(c_{\delta / 4})^{-1}}\rfloor$. However, if there was a direction $v$ not satisfying the conditions, an elementary geometric argument shows that we would be able to include $v$ in $S$, contradicting its maximality. For each $v_j$, consider $2d$ colour classes associated to it. Since $\diam (\mathcal{F}_i) < 1- \delta$ for all $i$, then their $v_j$-widths are also smaller than $1-\delta$. Thus, by Theorem \ref{theorem-helly-v-width}, there must be a rainbow choice of these $2d$ colours such that the $v_j$-width of its intersection is strictly smaller than $1-\delta$. Take $X$ to be the union of all these $2d$-tuples. Notice that the $v_j$-width of $\cap X$ is strictly smaller than $1-\delta$ for all $v_j$. Let $\lambda = \diam (\cap X) \ge 1$, and $v$ a direction realising it. Thus, $X$ contains a segment parallel to $v$ of length $\lambda$. Let $v_j$ be such that $\langle v_j, v \rangle \ge 1- \delta$. This implies that the $v_j$-width of $X$ is at least $1-\delta$, a contradiction. \end{proof} \section{Fractional and $(p,q)$ results for the diameter}\label{section-fractional} In order to prove Theorem \ref{theorem-p,q-diameter}, we need to recreate the results needed for the proof of the original $(p,q)$ theorem for the diameter. Simplified versions of Alon and Kleitman's method can be found in \cite{Alon:1996uf, Matousek:2002td}. There are two main ingredients needed. One is a fractional Helly theorem and the second is the existence for weak $\epsilon$-nets for convex sets of small size. Their equivalents are Theorem \ref{theorem-fractional} and \ref{theorem-weak-nets} described below. The structure of the proof we present here is the same. However, some definitions, such as the one for weak $\varepsilon$-net, must be adapted. In the case of volume, it is possible to recreate these results using properties of floating bodies \cite{Soberon:2015tsa}. Namely, given a convex set $K$ of volume one, and $\varepsilon >0$, we define its floating body $K_{\varepsilon}$ as \[ K_{\varepsilon} = K \setminus \cup \{H : H \ \mbox{is a halfsapce}, \vol (H \cap K) \le \varepsilon \}. \] There estimates on $\vol (K_{\varepsilon})$ allow for the proofs to work \cite{Barany:2010cy}. For the diameter, there is no similar ``floating body''. However, pigeonhole arguments on the directions realising the diameter, as in section \ref{section-colourful}, are sufficient. Let us begin with a fractional Helly for diameter in the same spirit as \cite{Katchalski:1979bq}. \begin{theorem}[Fractional Helly for the diameter]\label{theorem-fractional} Let $\alpha >0$, $1> \delta >0$ and $d$ a positive integer. Then, there is a positive constant $\beta$ depending only on $\alpha, d, \delta$ such that for any finite family $\mathcal{F}$ of $n$ convex sets in $\mathds{R}^d$ such that the intersection of at least $\alpha {{n}\choose{2d}}$ of the $2d$-tuples has diameter greater than or equal to one, there is a subfamily $\mathcal{F}'$ of $\mathcal{F}$ of cardinality at least $\beta n'$ such that its intersection has diameter at least $1-\delta$. \end{theorem} The equivalent result for volume \cite[Thm. 1.4]{Soberon:2015tsa} has the disadvantage that the size of the subfamilies needed to check grows as $\delta$ decreases. Namely, it needs to check families of size $O(\delta^{-(d^2-1)/4})$, which is much worse than the requirements of the volumetric Helly theorem. Theorem \ref{theorem-fractional} is an example of a fractional Helly-type theorem which goes beyond its corresponding Helly theorem. Such examples have appeared previously for set systems of bounded VC-dimension \cite{Matousek:2004cs}, for convexity on the integer lattice \cite{Anonymous:PHt9HPGF} or for checking the existence of hyperplane transversals in $\mathds{R}^d$ \cite{Alon:1995fs}. \begin{proof}[Proof of Theorem \ref{theorem-fractional}] Consider the usual probability Haar measures on $S^{d-1}$. For $y \in S^{d-1}$, let $C_\delta (y)$ be the set of points $x \in S^{d-1}$ such that $\langle x, y \rangle \ge 1-\delta$. Let $c_{\delta}$ be the measure of $C_{\delta} (y)$. A double counting argument shows that for any set $D$ of points in $S^{d-1}$, there must be a subset $D'$ of cardinality at least $c_{\delta} |D|$ and a point $v$ in $S^{d-1}$ such that $D' \subset C_{\delta} (v)$. For each such $2d$-tuple $B \subset \mathcal{F}$, consider a direction $v_B$ realising its diameter. Each of these directions can be represented by an antipodal pair on $S^{d-1}$. Using the observation above, there must be a direction $v$ and set of at least $2 \alpha c_{\delta}{{n}\choose{2d}}$ of the $2d$-tuples of $\mathcal{F}$ whose intersections have $v$-width greater than or equal to $1-\delta$. Applying Theorem \ref{theorem-helly-v-width}, we are done. \end{proof} In order to get to the existence of weak $\varepsilon$-nets, we start by getting results showing that given a set of objects in $\mathds{R}^d$, there is a point $p$ that is ``sufficiently well surrounded'' by them. The first result of this type is Tverberg's theorem. Tverberg's theorem \cite{Tverberg:1966tb} says that given enough points in $\mathds{R}^d$, they can be split into $m$ parts such that the convex hulls of the parts intersect. A point in this intersection is in some sense ``very deep'' within the original set of points. In order to recreate this for the diameter, we need to work with sets with large diameter instead of points, giving the following statement. \begin{theorem}[Tverberg's theorem for diameter]\label{theorem-tverberg-diameter} Let $d, m$ be positive integer, $1>\delta > 0$ and $n = \left\lfloor 4d^2(m-1)c_{\delta}^{-1}\right\rfloor+1$. Given a family $\mathcal{T} = \{T_1, T_2, \ldots, T_{n}\}$ of sets in $\mathds{R}^d$ of diameter greater than or equal to one each, there is a partition of them into $m$ parts $A_1, A_2, \ldots, A_m$ so that \[ \diam \left(\bigcap_{i=1}^m \conv(\cup A_i)\right) \ge 1- \delta. \] \end{theorem} \begin{proof} By a double counting as before, there is a subfamily $\mathcal{T}' \subset \mathcal{T}$ of cardinality greater than or equal to $4d^2(m-1)+1$ and a direction $v$ such that the $v$-width of every member of $\mathcal{T}'$ is at least $1-\delta$. Now consider \[ \mathcal{F} = \{\conv (\cup \mathcal{G}): \mathcal{G} \subset \mathcal{T}', |\mathcal{G}| = (m-1)(2d-1)2d+1 \}. \] Notice that every family forming an element of $\mathcal{F}$ is missing at most $2d(m-1)$ members of $\mathcal{T}'$. Thus, the intersection of every $2d$ of them contains some $T_i \in \mathcal{T'}$ which in turn implies that the $v$-width of their intersection is at least $1-\delta$. Thus, by Theorem \ref{theorem-v-width-basic-helly} the $v$-width of $\cap \mathcal{F}$ is at least $1-\delta$. Take two points $x, y \in \cap \mathcal{F}$ that realise its $v$-width. Every half-space containing either of them has non-empty intersection with at least $2d(m-1)+1$ sets of $\mathcal{T}'$. Otherwise, it would contradict the fact that they are contained in $\cap \mathcal{F}$. Thus, $x,y$ are contained in the convex hull of $\cup \mathcal{T}'$. By the colourful Carath\'eodory theorem for two points (see Section \ref{section-width}) with $\mathcal{F}_i = \cup {\mathcal T}'$ for $1 \le i \le 2d$, the set $\{x,y\}$ is contained in the convex hull of $2d$ points of $\cup \mathcal{T}'$. If we remove the sets $T_i$ that generated these points and set them aside in a set called $A_1$, we have that every half-space containing either of $x,y$ has non-empty intersection with at least $2d(m-2)+1$ sets of what is left of $\mathcal{T}'$. Thus, we can continue this process and generate the desired sets $A_1, A_2, \ldots, A_m$ inductively. \end{proof} Using Tverberg's theorem and colourful Carath\'edory, one can prove B\'ar\'any's selection theorem \cite{Barany:1982va}, also called the ``first selection lemma'' in \cite{Matousek:2002td}. It says that, given a finite set $S$ of points in $\mathds{R}^d$, there is a point $p$ in a positive fraction of the simplices spanned by $S$. This result holds in much more general settings, by either replacing the word ``simplex'' by the image of a different operator or requiring additional properties on the simplices containing the point \cite{Gromov:2010eb, Karasev:2012bj, Pach:1998vx, Magazinov:2015td}. For our purposes we only need a diameter version of the original result by B\'ar\'any. \begin{theorem}[Selection lemma for diameter]\label{theorem-selection-diameter} Let $d$ be a positive integer and $1 > \delta > 0$. There is a constant $\lambda = \lambda (\delta, d)$ such that for any finite family $\mathcal{F}$ of convex sets in $\mathds{R}^d$ of diameter one each, there is a set $K$ of diameter at least $1-\delta$ such that $K \subset \conv (\cup A)$ for at least $\lambda {{|\mathcal{F}|}\choose{2d}}$ subsets $A \subset \mathcal{F}$ of cardinality $2d$. Moreover, $\lambda = \Omega_d( \delta ^{d(d-1)})$. \end{theorem} \begin{proof} First, there is a subfamily $\mathcal{F}' \subset \mathcal{F}$ of cardinality at least $c_{\delta} |\mathcal{F}|$ and a direction $v$ such that the $v$-width of every member of $\mathcal{F}'$ is at least $1-\delta$. By the second part of the proof of Theorem \ref{theorem-tverberg-diameter}, there is a partition of $\mathcal{F}'$ into at least $\frac{|\mathcal{F}'|-1}{4d^2}+1$ parts such that the intersection of the convex hull of the union of the parts contains a set $K$ of $v$-width at least $1-\delta$. Moreover, we may assume that $K$ is the convex hull of two points. Colour each part by a different colour. By the colourful Carath\'eodory theorem for two points (see section \ref{section-width}), for each $2d$-tuple of colours, there is an heterochromatic set whose convex hull contains $K$. Thus, up to constants in the dimension there are at least \[ {{\frac{|\mathcal{F}'|-1}{4d^2}+1}\choose{2d}} \sim (4d)^{-2d}{{c_{\delta}|\mathcal{F}|}\choose{2d}} \sim c_{\delta}^{2d}(4d)^{-2d}{{|\mathcal{F}|}\choose{2d}} \] subsets $A \subset \mathcal{F}$ of cardinality $2d$ such that $K \subset \conv(\cup A)$. \end{proof} The final ingredient needed is the existence of weak $\varepsilon$-nets for diameters. The original results aims to find, for a given $S \subset \mathds{R}^d$, a set $T$ whose cardinality depends only on $\varepsilon$ and $d$ that intersects the convex hull of every subset of $S$ of cardinality at least $\varepsilon|S|$ \cite{Alon:2008ek}. For our purposes we need both $S$ and $T$ to be families of sets with large diameter. \begin{theorem}[Weak $\varepsilon$-nets for diameter]\label{theorem-weak-nets} Let $d$ be a positive integer, $1 > \delta > 0$, $1 > \varepsilon > 0$. Then, there is a constant $m = m(d, \delta, \varepsilon)$ such that for any finite family $\mathcal{F}$ of sets of diameter one each in $\mathds{R}^d$, there are $m$ sets $K_1, K_2, \ldots, K_m$ of diameter at least $1-\delta$ each such that for any subfamily $\mathcal{F}' \subset \mathcal{F}$ with $|\mathcal{F}'| \ge \varepsilon|\mathcal{F}|$, there is an $i$ satisfying $K_i \subset \conv (\cup \mathcal{F}')$. Moreover, $m(d, \delta, \varepsilon) = O_d (\varepsilon^{-2d}\cdot\delta^{-d(d-1)})$. \end{theorem} \begin{proof} We construct the set $\mathcal{K}=\{K_1, \ldots, K_m\}$ inductively, starting with an empty set. Let $T$ be the number of $2d$-tuples $A \subset \mathcal{F}$ such that $\conv (\cup A)$ contains no set in $\mathcal{K}$. If there is a subfamily $\mathcal{F}'$ with $|\mathcal{F}'| \ge \varepsilon|\mathcal{F}|$ such that $\conv (\cup \mathcal{F}')$ does not contain any set of $\mathcal{K}$, we can apply Theorem \ref{theorem-selection-diameter} to $\mathcal{F}'$. Thus, we can find a set $K$ contained in the convex hull of the union of at least $\lambda {{|\mathcal{F}'|}\choose{2d}} \sim \lambda \varepsilon^{2d} {{|\mathcal{F}|}\choose{2d}}$ different subsets $A \subset \mathcal{F}$ of cardinality $2d$, effectively reducing $T$ by that number if we add $K$ to $\mathcal{K}$. The process cannot be repeated more than $O_d\left((\lambda \varepsilon^{2d})^{-1}\right)$ times, as desired. \end{proof} We call $\mathcal{K}$ a diameter weak $\varepsilon$-net for the pair $(\mathcal{F} , \delta)$. At this point we have all the ingredients needed to prove Theorem \ref{theorem-p,q-diameter}. The proof of this theorem relies on the linear programming technique by Alon and Kleitman. For this, we need the following definitions. We consider $\mathcal{C}_{d, \delta} = \{F \subset \mathds{R}^d: \diam (F) \ge 1- \delta, \ F \ \mbox{is convex}\}$. Then, given a finite family of convex sets $\mathcal{F}$ in $\mathds{R}^d$, we define \begin{itemize} \item the diameter transversal number $\tau_{\delta} (\mathcal{F})$ as the minimum $\sum_{C \in \mathcal{C}_{d,\delta}} w(C)$ over all functions $w:\mathcal{C}_{d,\delta} \to \{0,1\}$ such that \[ \sum_{{C:C \subset F, \ C \in \mathcal{C}_{d,\delta}}} w(C) \ge 1 \] for all $F \in \mathcal{F}$, \item the fractional diameter transversal number $\tau^*_{\delta} (\mathcal{F})$ as the minimum $\sum_{C \in \mathcal{C}_{d,\delta}} w(C)$ over all functions $w:\mathcal{C}_{d,\delta} \to [0,1]$ such that \[ \sum_{C: C \subset F, \ C \in \mathcal{C}_{d,\delta}} w(C) \ge 1 \] for all $F \in \mathcal{F}$, and \item the fractional diameter packing number $\nu^*_k (\mathcal{F})$ as the maximum $\sum_{F \in \mathcal{F}} w(F)$ for $w: \mathcal{F} \to [0,1]$ such that \[ \sum_{F: C \subset F, \ F \in \mathcal{F}} w(F) \le 1 \] for all $C \in \mathcal{C}_{d,\delta}$. \item Given two finite families $\mathcal{F}$, ${\mathcal T}$ of convex sets in $\mathds{R}^d$, we say ${\mathcal T}$ is a $(1-\delta)$-diameter transversal for $\mathcal{F}$ if every set in ${\mathcal T}$ has diameter at least $1- \delta $ and every set in $\mathcal{F}$ contains at least one set in ${\mathcal T}$. Note that if $w:\mathcal{C}_{d,\delta} \to \{0,1\}$ is a function satisfying the condition of $\tau_{\delta}(\mathcal{F})$, then the family $\{K\in \mathcal{C}_{d, \delta}: w(K)=1\}$ is a $(1-\delta)$-diameter transversal of $\mathcal{F}$. \item We refer to the conditions of Theorem \ref{theorem-p,q-diameter} as the diameter $(p,q)$ condition. \end{itemize} \begin{lemma}\label{lemma-one} Let $\mathcal{F}$ be a finite family of convex sets in $\mathds{R}^d$, all with diameter at least one. Then, $\tau_{\delta} (\mathcal{F})$ is bounded by a function depending only $\tau^*_{\delta/2}(\mathcal{F})$, $d$ and $\delta$. \end{lemma} \begin{proof} Consider a function $w:\mathcal{C}_{d, \delta/2} \to [0,1]$ which realises $\tau^{*}_{\delta/2}$. Without loss of generality, we may assume that $w$ has finite support and only has rational values. Let $M$ be the common denominator of $w(K)$ for all $K \in \mathcal{C}_{d, \delta/2}$. Let ${\mathcal T}$ be the family that formed by the disjoint union of $M\cdot w(K)$ copies of $K$, for each $K \in \mathcal{C}_{d, \delta/2}$. Now consider $\mathcal{K}$ a diameter weak $\left(\frac{1}{\tau^*_{\delta/2}(\mathcal{F})}\right)$-net of $({\mathcal T}, \delta/2)$, as in Theorem \ref{theorem-weak-nets}. Notice that the diameter of every member of $\mathcal{K}$ is at least $(1-\frac{\delta}{2})^2 \ge 1-\delta$. By the definition of $\tau^*_{\delta/2}$, for $F \in \mathcal{F}$, the number of copies of sets in ${\mathcal T}$ which are contained in $F$ is at least $(\tau^*_{\delta/2} (\mathcal{F}))^{-1} |{\mathcal T}|$. Thus, there is an element of $\mathcal{K}$ contained in $F$. In other words, $\mathcal{K}$ is a $(1-\delta)$-diameter transversal to to $\mathcal{F}$, so $\tau_{\delta} (\mathcal{F}) \le |\mathcal{K}|$, which in turn is only bounded by a function of $\tau^*_{\delta/2} (\mathcal{F})$, $d$ and $\delta$. \end{proof} \begin{lemma}\label{lemma-two} If $p \ge q \ge 2d$ and $\mathcal{F}$ is a finite family of convex sets with the diameter $(p,q)$ condition, then $\nu^*_{\delta/2} (\mathcal{F})$ is bounded by a function that depends only on $p,q,d,\delta$. \end{lemma} \begin{proof} Let $w: \mathcal{F} \to [0,1]$ be a function that realises $\nu^*_{\delta/2} (\mathcal{F})$. We may assume without loss of generality that $w(C)$ is rational for all $C \in \mathcal{F}$. Let $w(C) = \frac{n_C}{m}$ where $m$ is the common denominator for all $w(C)$ for $C \in \mathcal{F}$. Let $\mathcal{F}'$ be the family consisting of $n_C$ copies of $C$ for each $C \in \mathcal{F}$ and let $N= |\mathcal{F}'|$. Note that $\frac{N}{m}= \sum_{C \in \mathcal{F}} \frac{n_C}{m} = \nu^*_{\delta/2}(\mathcal{F})$. The family $\mathcal{F}'$ satisfies the diameter $((q-1)(p-1)+1,q)$ property. This comes immediately from the fact that every $[(q-1)(p-1)+1]$-tuple from $\mathcal{F}'$ contains either $q$ copies of the same set or $p$ different sets of $\mathcal{F}$. In either case we have a $q$-tuple whose intersection has diameter at least one. However, since $q\ge 2d$ this implies that there is a positive fraction of the $2d$-tuples of $\mathcal{F}'$ whose intersection has diameter at least one. Theorem \ref{theorem-fractional} implies then that there is a positive fraction $\beta$ depending only on $p,q,\delta$ such that there is a set $K_0$ of diameter at least $1-\frac{\delta}{2}$ contained in the intersection of at least $\beta N$ sets of $\mathcal{F}'$. Thus \[ 1 \ge \sum_{C \in \mathcal{F}: K_0 \subset C}w(C) = \sum_{C \in \mathcal{F}: K_0 \subset C} \frac{n_C}{m} \ge \frac{1}m \cdot \beta N = \beta \nu^*_{\delta/2}(\mathcal{F}). \] This implies $\nu^*_{\delta/2} (\mathcal{F}) \le \frac{1}{\beta}$, as desired. \end{proof} \begin{proof}[Proof of Theorem \ref{theorem-p,q-diameter}] As in the Alon-Kleitman proof of the $(p,q)$ theorem, linear programming duality implies that $\nu_{\delta/2}^*(\mathcal{F}) = \tau^*_{\delta/2}(\mathcal{F})$. Thus, the lemmata \ref{lemma-one} and \ref{lemma-two} finish the proof. \end{proof} \section{Remarks and open problems}\label{section-remarks} The Helly-type results we obtain for the diameter here improve upon their volumetric equivalent. However, we see no reason they should not hold for the same strength in that setting. Corollary \ref{corollary-chido} seems like a good result to test for that purpose. \begin{question} Is there a constant $r(d, \delta)$ such that any finite family $\mathcal{F}$ of convex sets in $\mathds{R}^d$ such that the intersection of every $2d$ of them has volume at least one can be partitioned into $r(d, \delta)$ parts so that the volume of the intersection of each part is at least $1-\delta$? \end{question} Let us construct an example to show that the diameter loss $\delta$ is needed in Corollary \ref{corollary-chido}, Theorem \ref{theorem-p,q-diameter} and Theorem \ref{theorem-fractional}. \begin{claim} For any $k$, there is a family $\mathcal{F}$ of $2dk+1$ convex sets in $\mathds{R}^d$ such that the intersection of any $2d$ of them has diameter at least one and for any partition of $\mathcal{F}$ into $k$ parts, there is one whose intersection has diameter strictly smaller than one. \end{claim} \begin{proof} Let $n={{2k+1}\choose{2d}}$, and for each $2d$-tuple $A \subset \{1,2,\ldots, 2kd+1\}$, let $v_A$ be a pair of antipodal points in $\frac12S^{d-1}$. For each $i \in \{1,2,\ldots, 2kd+1\}$, let \[ K_i = \conv \left\{v_A: A \in {{[2kd+1]}\choose{2d}}, \ i \in A \right\}. \] For any $2d$-tuple of sets, by construction their intersection contains some $v_A$, so the diameter is at least one. Given a partition of $\{K_i : i\in [2kd+1]\}$ into $k$ parts, there must be one, say $P$, of cardinality at least $2d+1$. For any $A \in {{[2kd+1]}\choose{2d}}$, there is an $K_i \in P$ for which $i \not\in A$, so $v_A \not\in \conv (\cap P)$. Thus, $\cap P$ is contained in the interior of $\conv (\frac12S^{d-1})$ and is closed, so its diameter is strictly less than one. If one wants shaper estimates, we can choose the antipodal points $v_A$ so that the circular caps $C_{\delta} (v_A)$ are pairwise disjoint for some $\delta$, similarly to the argument of Theorem \ref{theorem-colourful-diameter}. \end{proof} The original conjecture by B\'ar\'any, Katchalski and Pach is still open, so we state it again to bring more attention to it. \begin{conjecture}[B\'ar\'any, Katchalski, Pach \cite{Barany:1982ga}] Let $\mathcal{F}$ be a finite family of convex sets such that the intersection of every $2d$ of them has diameter at least one. Then, $\diam (\cap \mathcal{F}) \ge c d^{-1/2}$ for some absolute constant $c$. \end{conjecture} For the conjecture above, the best lower bound on $\diam (\cap \mathcal{F})$ is $O(d^{-2d})$ \cite{Barany:1982ga}. \bibliographystyle{amsalpha}

https://arxiv.org/abs/2209.10045

New Lower Bounds for Cap Sets

A cap set is a subset of $\mathbb{F}_3^n$ with no solutions to $x+y+z=0$ other than when $x=y=z$. In this paper, we provide a new lower bound on the size of a maximal cap set. Building on a construction of Edel, we use improved computational methods and new theoretical ideas to show that, for large enough $n$, there is always a cap set in $\mathbb{F}_3^n$ of size at least $2.218^n$.

\section{Introduction} \begin{definition}\label{def:cap} A \emph{cap set} is a set $A \subseteq \mathbb{F}_3^n$ with no solutions to $x+y+z=0$ other than when $x=y=z$, or equivalently $A$ has no 3 distinct elements in arithmetic progression. \end{definition} In this paper, we prove the following result. \begin{theorem} \label{main} There is a cap set in $\mathbb{F}_3^{56232}$ of size \[\binom{11}{7}^{141} \cdot 6^{572} \cdot 12^{572} \cdot 112^{8800} \cdot 37 \cdot 142\] and hence, for large $n$, there is a cap set $A \subseteq \mathbb{F}_3^n$ of size $(2.218021\ldots)^n$. \end{theorem} Since \cite{Pellegrino} in 1970, there have been 2 improvements to the lower bound on the size of a maximal cap set. A lower bound of $(2.210147\ldots)^n$ was given in \cite{CalderbankFishburn}, and then Edel gave an improved lower bound of $(2.217389\ldots)^n$ in \cite{Edel}. In this paper, we obtain the first new lower bound for nearly 2 decades, and prove that a maximal cap set has size at least $(2.218021\ldots)^n$. Our method is based on that of Edel, which in turn is based on the work of Calderbank and Fishburn, and involves taking cap sets which are known to be maximal in a relatively low dimension and then combining them carefully to produce large cap sets in higher dimensions. \bigbreak The upper bound, by contrast, has received significant attention. In \cite{Meshulam}, an upper bound of $\frac{3^n}{n}$ was shown, which was then improved to $\frac{3^n}{n^{1+\varepsilon}}$ for some $\varepsilon > 0$ in \cite{BK}. It was an open problem for over 20 years as to whether a cap set in $\mathbb{F}_3^n$ has size at most $c^n$, for some $c<3$. This was finally solved by Ellenberg and Gijswijt in \cite{EllenbergGijswijt}, who used the polynomial method developed by Croot, Lev and Pach in \cite{CLP} to show that a cap set in $\mathbb{F}_3^n$ has size at most $(2.7552\ldots)^n$. A symmetric version of the polynomial method proof for the upper bound has since been formulated by Tao in \cite{tao_2016}, and Ellenberg and Gijswijt's result was formalised in the Lean theorem prover in \cite{dahmen2019formalizing}. The upper bound has since been improved by an extra factor of $\sqrt{n}$ in \cite{jiang2021improved}. \bigbreak One reason for the interest in cap sets is that they can provide useful insights into analogous problems in more complicated sets. For example, the current best upper bound for Roth's theorem due to Bloom and Sisask in \cite{bloom2020}, which asks how large a subset of integers can be without containing a 3 term arithmetic progression, employs several of the methods which were used in the upper bound for the cap set problem given by Bateman and Katz in \cite{BK}. \smallbreak Cap sets are also of interest in finite geometry, design of experiments and various other problems in combinatorics and number theory. For an excellent survey on the motivation and background of the problem, and its application to several other interesting questions, we recommend the article \cite{grochow2019new}. \subsection*{Structure of the paper} In section 2, we present the extended product construction of Edel, and combine it with improved computational techniques to obtain some new lower bounds. We then introduce a new construction in section 3, which takes the extended product from section 2 up a level, and use this to achieve the best lower bound in this paper. In section 4, we discuss possible ideas for future work, as well as the limitations of our approach. Finally, we describe our computational methods, which make use of a SAT solver, in section 5. \section{The Extended Product Construction} In this section, we describe a construction due to Edel in \cite{Edel}, which extends the ideas from \cite{CalderbankFishburn} to produce larger cap sets. We use Edel's method to construct a cap set in 396 dimensions which gives a lower bound of $(2.217981\ldots)^n$, already an improvement on Edel's lower bound of $(2.217389\ldots)^n$. We first describe a relatively simple construction. \begin{proposition}[Product Caps] \label{def:ProdCap} Let $A \subseteq \mathbb{F}_3^n$, $B \subseteq \mathbb{F}_3^m$ be cap sets. Then, by taking a direct product of $A$ and $B$, there is a cap set of size $|A||B|$ in $\mathbb{F}_3^{n+m}$. \end{proposition} \begin{proof} Define $A \times B = \{(a,b) : a \in A, b \in B\}$. Clearly $|A \times B| = |A||B|$, and we show that $A \times B$ is a cap set. \smallbreak Assume we have a solution to $x+y+z = 0$ in $A \times B$. So $x = (x_a, x_b)$, $y = (y_a, y_b)$ and $z = (z_a, z_b)$ where $x_a, y_a, z_a \in A$ and $x_b, y_b, z_b \in B$. Therefore, $x_a + y_a + z_a = 0 = x_b + y_b + z_b$. Since $A,B$ are both cap sets, we must have $x_a = y_a = z_a$ and $x_b = y_b = z_b$, so $x=y=z$. Hence $A \times B$ has no non trivial solutions to $x+y+z = 0$, and is therefore a cap set. \end{proof} \smallbreak The above proposition shows that there is always a cap set in $\mathbb{F}_3^n$ of size $2^n$, by taking direct products of the set $\{0,1\} \subseteq \mathbb{F}_3$, to form the cap set $\{0,1\}^n \subseteq \mathbb{F}_3^n$. We now use the direct product construction to show how we can derive an asymptotic lower bound for the cap set problem. \begin{proposition} \label{prop:lower bound asymptotic} Let $A \subseteq \mathbb{F}_3^n$ be a cap set of size $c^n$. Then for any $\varepsilon > 0$, there is an $M$ such that for all $m \geq M$, there is a cap set of size greater than $\left(c - \varepsilon\right)^{m}$ in $\mathbb{F}_3^m$. \end{proposition} \begin{proof} Let $A \subseteq \mathbb{F}_3^n$ be a cap set of size $c^n$. For $m > n$, there is $k$ such that $m = nk + r$, where $0 \leq r<n$. Applying the product construction $k$ times to $A$, we have a cap set in $\mathbb{F}_3^m$ of size $c^{nk} = \left(c^{1-r/m}\right)^m$. \smallbreak Given $\varepsilon>0$, we can always choose a large enough $M$ such that $c^{1-n/M} > c - \varepsilon$. Since $r<n$, for all $m \geq M$ we have $\frac{r}{m}< \frac{n}{M}$, and hence $c^{1-r/m} > c - \varepsilon$. Therefore, we have constructed a cap set in $\mathbb{F}_3^m$ of size greater than $(c - \varepsilon)^m$. \end{proof} \begin{remark} This proposition demonstrates that finding an asymptotic lower bound for the cap set problem amounts to finding a cap set $A \subseteq \mathbb{F}_3^n$ where $|A|^{1/n}$ is as large as possible. If $A \subseteq \mathbb{F}_3^n$ and $B \subseteq \mathbb{F}_3^m$ are cap sets, we can use the previous proposition to achieve a lower bound of $|A|^{1/n}$ and $|B|^{1/m}$ respectively. We know we can form the direct product $A \times B$, but since $|A \times B|^{\frac{1}{n+m}} \leq \text{max}\left(|A|^{\frac{1}{n}},|B|^{\frac{1}{m}}\right)$, the bound from $A \times B$ will never beat the better of the bounds from $A$ and $B$. \end{remark} In \cite{Edel}, Edel introduced a new construction based on the work in \cite{CalderbankFishburn}, which one can think of as a sort of twisted product. The idea is simple - if we start with a collection of cap sets, and construct several different direct products, can we take the union of the direct products and still be a cap set? The answer is yes, under certain conditions on the cap sets and the way we combine them. \begin{definition}[Extendable collection] \label{def:extendable} Let $A_0, A_1, A_2 \subseteq \mathbb{F}_{3}^{n}$ be cap sets. We say that this collection of cap sets is \emph{extendable} if the following 2 conditions hold: \begin{enumerate} \item If $x,y\in A_0$ and $z\in A_1\cup A_2$ then $x+y+z \neq 0$. \item If $x\in A_0$, $y\in A_1$ and $z\in A_2$ then $x+y+z\neq 0$. \end{enumerate} Note that taking $x=y$ in condition (1) shows that $A_0$ is disjoint from $A_1$ and $A_2$. \end{definition} \smallbreak \begin{definition}[Admissible set\footnote{To avoid confusion, we note that our terminology and definitions are not the same as those in \cite{Edel}. In particular, the object which Edel calls a `cap' we call a `cap set', and Edel's `capset' is our `admissible set'.}] \label{def:admissible} Let $S \subseteq \{ 0,1,2\}^m$. $S$ is \emph{admissible} if: \begin{enumerate} \item For all distinct $s, s' \in S$, there are coordinates $i$ and $j$ such that $s_i = 0 \neq s_i'$ and $s_j \neq 0 = s_j'$. \item For all distinct $s, s', s'' \in S$, there is a coordinate $k$ such that $\{s_k, s_k', s_k''\} = \{0, 1, 2\}$, $\{0, 0, 1\}$ or $\{0, 0, 2\}$. \end{enumerate} \end{definition} \begin{definition}[Extended product construction] \label{def:extended product} As the name suggests, we can extend an extendable collection of cap sets by an admissible set. The construction is as follows: for $s = (s_1, \ldots, s_m) \in \{0,1,2\}^m$ and $A_0, A_1, A_2 \subseteq \mathbb{F}_{3}^{n}$, we define \[s(A_0, A_1, A_2) = A_{s_1} \times \cdots \times A_{s_m} \subseteq \mathbb{F}_{3}^{nm}.\] If $S \subseteq \{0,1,2\}^{m}$ is an admissible set, we define \[S(A_0, A_1, A_2) = \bigcup_{s \in S} \ s\left(A_0, A_1, A_2\right) \subseteq \mathbb{F}_{3}^{nm}.\] \end{definition} The following lemma, which although rather different in presentation is essentially Lemma 10 of \cite{Edel}, demonstrates the usefulness of these definitions. \begin{lemma} \label{lemma:extended product cap} If $(A_0, A_1, A_2)$ is an extendable collection of cap sets in $\mathbb{F}_{3}^{n}$, and $S \subseteq \{ 0,1,2\}^m$ is an admissible set, then $S(A_0, A_1, A_2)$ is a cap set in $\mathbb{F}_{3}^{nm}$. \end{lemma} \begin{proof} We want to prove that \[\bigcup_{s \in S} s(A_0, A_1, A_2) \] is a cap set, where $s(A_0,A_1, A_2)= A_{s_1} \times \cdots \times A_{s_{m}}$. \smallbreak Suppose we have distinct $x,y,z \in S(A_0, A_1, A_2)$ such that $x+y+z=0$. We have 3 cases, depending on where $x,y,z$ come from. \smallbreak \textbf{Case 1:} By the direct product construction, we know that each $s(A_0,A_1, A_2)$ is a cap set. So, there are no distinct $x,y,z \in s(A_0,A_1, A_2)$ such that $x+y+z=0$. \smallbreak \textbf{Case 2:} Suppose we have $x,y\in s(A_0,A_1, A_2)$ and $z \in s'(A_0,A_1, A_2)$, where $s \neq s'$. So $x=(x_{s_1},\ldots,x_{s_m})$, $y=(y_{s_1},\ldots,y_{s_m})$ and $z=(z_{s_1'},\ldots,z_{s_m'})$. By property (1) of being admissible, there is some coordinate $j$ in which $s_j=0$ and $s_j'\neq 0$. So $x_{s_j}+y_{s_j}+z_{s'_j}=0$ where $x_{s_j},y_{s_j}\in A_0$ and $z_{s'_j} \in A_1 \cup A_2$, contradicting property (1) of extendable. \smallbreak \textbf{Case 3:} Suppose $x,y,z$ come from the distinct vectors $s,s',s''$. By condition (2) of admissible, there is a coordinate $k$ such that $\{s_k, s_k', s_k''\}$ is $\{0, 1, 2\}$, $\{0, 0, 1\}$ or $\{0, 0, 2\}$. If $\{s_k,s_k',s_k''\}=\{0,0,1\}$ or $\{0,0,2\}$, then we have a contradiction of property (1) of extendable, as above. Otherwise, if $\{s_k,s_k',s_k''\}=\{0,1,2\}$, we have a contradiction of property (2) of extendable. \end{proof} We will now discuss some important types of admissible sets, which will be used in our later constructions. \begin{definition}[Recursively admissible set\footnote{Again, note that our terminology is different to Edel's. In \cite{Edel}, these objects are simply called `admissible sets'.}] \label{def:recursive} $S$ is a \emph{recursively admissible} set if $S$ is an admissible set, $|S| \geq 2$ and for all distinct pairs $s, s' \in S$ at least one of the following holds: \begin{enumerate}[label=(\roman*)] \item There are coordinates $i, j$ such that $\{s_i, s_i'\} = \{0, 1\}$ and $\{s_j, s_j'\} = \{0, 2\}$. \item There is a coordinate $k$ such that $s_k = s_k' = 0$. \end{enumerate} \end{definition} Given the name `recursively admissible', the reader may not be too surprised by the flavour of the following lemma. \begin{lemma} \label{lemma:recursive} If $(A_0, A_1, A_2)$ is an extendable collection of cap sets, and $S \subseteq \{0,1,2\}^m$ is a recursively admissible set, then $\left(S\left(A_0, A_1, A_2\right), A_1^m, A_2^m\right)$ is an extendable collection of cap sets. \end{lemma} \begin{proof} $S(A_0, A_1, A_2)$ is a cap set by \eqref{lemma:extended product cap}, and takes the place of $A_0$ from the definition of extendable. First of all, we show that there are no $x,y\in S(A_0, A_1, A_2)$ and $z\in A_1^m\cup A_2^m$ such that $x+y+z=0$. \bigbreak Assume $x,y \in S(A_0, A_1, A_2)$ and $z \in A_1^m \cup A_2^m$. If $x,y \in s(A_0, A_1, A_2)$, then since $|S|\geq 2$, it must be that $s$ has at least one coordinate 0. Let $s_k= 0$, and let the corresponding blocks of $x,y,z$ be $x_k, y_k, z_k$ respectively. Then $x_k, y_k \in A_0$, $z_k \in A_1 \cup A_2$, so by property (1) of $(A_0, A_1, A_2)$ being extendable, $x_k + y_k + z_k \neq 0$. \smallbreak Now assume $x \in s(A_0, A_1, A_2)$ and $y \in s'(A_0, A_1, A_2)$. Since $S$ is recursively admissible, either there is a coordinate $k$ such that $s_k = s'_k = 0$ or there are coordinates $j,k$ such that $\{s_j, s'_j\} = \{0,1\}$ and $\{s_k, s'_k\} = \{0,2\}$. \smallbreak In the first case where $s_k = s'_k = 0$, we have $x_k, y_k \in A_0$ and $z_k \in A_1 \cup A_2$, so by the same reasoning as above $x_k + y_k + z_k \neq 0$ by property (1) for extendable. \smallbreak In the case where $\{s_j, s'_j\} = \{0,1\}$ and $\{s_k, s'_k\} = \{0,2\}$, either $z_j = z_k = 1$ or $z_j = z_k = 2$. Assume that $z_j = z_k = 2$ and $s_j =1$. Then $x_j \in A_1$, $y_j \in A_0$ and $z_j \in A_2$. By property (2) of extendable, we deduce that $x_j + y_j + z_j \neq 0$. \smallbreak Similarly, if $z_j = z_k = 2$ and $s'_j = 1$, then $x \in A_0$, $y \in A_1$ and $z \in A_2$, so $x_j+y_j+z_j \neq 0$ by property (2) of extendable again. Finally, if $z_j = z_k = 1$, then either $s_k = 0$ and $s'_k = 2$, or $s_k = 2$ and $s'_k = 0$, and once again we use property (2) of extendable to show that $x_k + y_k + z_k \neq 0$. \smallbreak Hence, we can never have $x,y\in S(A_0, A_1, A_2)$ and $z\in A_1^m\cup A_2^m$ such that $x+y+z=0$, so condition (1) of extendable holds. \bigbreak Now we want to show that if $x\in S(A_0, A_1, A_2)$, $y\in A_1^m$ and $z\in A_2^m$ then $x+y+z\neq 0$. This is relatively straightforward - by the same reasoning as above, $s$ has a coordinate $k$ where $s_k = 0$. So, $x_k \in A_0$, $y_k \in A_1$ and $z_k \in A_2$. Then, by property (2) of extendable, $x_k + y_k + z_k \neq 0$, so $x + y + z \neq 0$. \smallbreak So we have condition (2) for extendable, and hence $\left(S\left(A_0, A_1, A_2\right), A_1^m, A_2^m\right)$ is an extendable collection of cap sets, as required. \end{proof} In addition to recursively admissible sets, we will also be making use of admissible sets where every element has the same weight. By `weight', we mean the number of non zero entries in a vector. We will also talk about the `support' of a vector, meaning the set of non zero coordinates. \begin{definition}[Constant weight admissible sets] Write $S = I(m,w)$ if $S \subseteq \{0,1,2\}^{m}$ is an admissible set consisting of $\binom{m}{w}$ vectors, each of weight $w$. If in addition $S$ is recursively admissible, write $S = \tilde{I}(m,w)$.\footnote{Once again, we mention the difference between our notation and that of Edel. In \cite{Edel}, the second parameter is not the weight $w$ of each vector, but the number of zeroes $t$. Note that $t = m -w$, and so $\binom{m}{w}$ = $\binom{m}{t}$. It is also worth pointing out that the meanings of $I$ and $\tilde{I}$ are swapped in \cite{Edel} - whereas we use $\tilde{I}$ to denote a stronger property than $I$, Edel uses $\tilde{I}$ to denote the weaker property of being an admissible set (which is referred to there as a capset), and uses $I$ for a recursively admissible set (which Edel simply calls an admissible set).} \end{definition} \smallbreak One good thing about this class of admissible sets is that they satisfy the pairwise condition for admissible automatically. There are $\binom{m}{w}$ different ways to choose $w$ of the $m$ coordinates to be non zero, and any 2 distinct vectors $x,y$ then necessarily have coordinates $i,j$ such that $x_i = 0 \neq y_i$ and $x_j \neq 0 = y_j$. This will turn out to be helpful when we need to find admissible sets later, as we can start with the $\binom{m}{w}$ different support sets, which immediately gives us the first condition for admissible, and then we can view the second condition as a 2-colouring problem on the non zero entries of each vector. \smallbreak Another reason this type of admissible set is useful to work with is that it makes calculating the size of the extended product cap sets relatively simple. \begin{lemma} \label{lemma:size} If we extend a collection of cap sets $(A_0, A_1, A_2)$ by $S = I(m,w) \subseteq \{ 0,1,2\}^m$, where $|A_1| = |A_2|$, then \[|S(A_0, A_1, A_2)| = \binom{m}{w}|A_0|^{m-w} |A_1|^{w} .\] \end{lemma} \begin{proof} Let $s \in S$. If $s_i = 1$ or $s_i = 2$, then $|A_{s_i}| = |A_1|$, and if $s_i = 0$ then $|A_{s_i}| = |A_0|$. Recall that we defined $s(A_0, A_1, A_2) = A_{s_1} \times \cdots \times A_{s_m}$. Since there are $m-w$ zero coordinates and $w$ non zero coordinates in $s$, a simple counting argument gives $|s(A_0, A_1, A_2)| = |A_0|^{m-w} |A_1|^{w}$. Then, as $S(A_0, A_1, A_2) = \bigcup\limits_{s \in S} s(A_0, A_1, A_2)$, the $s(A_0, A_1, A_2)$ are all disjoint and $|S| = \binom{m}{w}$, it follows that $|S(A_0, A_1, A_2)| = \binom{m}{w}|A_0|^{m-w} |A_1|^{w}$. \end{proof} It is finally time for some examples of admissible sets. For our first example, we prove the existence of an important family of recursively admissible sets, by a relatively simple construction due to Edel. We then use a computer search to produce particular examples of admissible sets. \begin{lemma} \label{lemma:m-1} For any $m \geq 2$, there exists a recursively admissible set $\tilde{I}(m,m-1)$. \end{lemma} This is a special case of Lemma 13 in \cite{Edel}, when $c=2$. \begin{proof} Construct a set $S$ as follows: consider the $m$ vectors who have exactly one coordinate $0$, and all others non zero. For each vector, let all entries before the $0$ be $1$, and all entries after the $0$ be $2$. We show S is a recursively admissible set. \smallbreak Let $x,y$ be distinct elements of $S$. Then they must have zeroes in different coordinates, so the pairwise condition for admissible holds: there are $i,j$ such that $x_i = 0 \neq y_i$ and $x_j \neq 0 = y_j$. \smallbreak Without loss of generality, let $i<j$. Then $y_i = 1$ and $x_j = 2$ by our construction, so we also have the condition for recursively admissible. That is, there exist $i,j$ such that $\{x_i, y_i\} = \{0,1\}$ and $\{x_j, y_j\} = \{0,2\}$. \smallbreak Let $x,y,z$ be distinct elements of $S$, with zero in coordinate $i,j,k$ respectively. Without loss of generality, let $i<k<j$. Then $x_k = 2$, $y_k = 1$, $z_k = 0$, so we have a coordinate $k$ such that $\{x_k, y_k, z_k\} = \{0,1,2\}$, which gives the triples condition for admissible. \end{proof} \begin{lemma} \label{computer caps} There exist admissible sets $I(11,7)$, $I(11,6)$ and $I(10,6)$. \end{lemma} The admissible sets can be found on the author's webpage at \url{fredtyrrell.com/cap-sets}. The computational methods we employed, including the use of a SAT solver, are described in section 5. Several other admissible sets were given by Edel, including the $I(10,5)$ used to find the previous lower bound in \cite{Edel}. The admissible sets mentioned in \cite{Edel} can be found on Edel's webpage \cite{EdelSets}. \bigbreak We now present an example of an extendable collection of cap sets in $\mathbb{F}_3^6$. This is exactly the collection used in \cite{Edel} and \cite{CalderbankFishburn}, presented slightly differently. We will summarise the construction - for more details, see Section 3 of \cite{Edel} or Section 2, Figure 3 of \cite{CalderbankFishburn}. \begin{lemma} \label{extendable caps} There is an extendable collection $(A_0, A_1, A_2)$ of cap sets in $\mathbb{F}_3^6$, where $|A_0| = 12$, $|A_1| = |A_2| = 112$. \end{lemma} \begin{proof} \textbf{Constructing the collection:} \smallbreak Consider the following 6 x 10 matrix: \begin{center} $\begin{pmatrix} 1 & 1 & 1 & 1 & 1 & 0 & 0 & 0 & 0 & 0\\ 1 & 1 & 0 & 0 & 0 & 1 & 1 & 1 & 0 & 0\\ 1 & 0 & 1 & 0 & 1 & 0 & 0 & 0 & 1 & 1\\ 0 & 1 & 0 & 1 & 0 & 0 & 1 & 0 & 1 & 1\\ 0 & 0 & 1 & 0 & 1 & 0 & 1 & 1 & 1 & 0\\ 0 & 0 & 0 & 1 & 1 & 1 & 0 & 1 & 0 & 1\\ \end{pmatrix}$ \end{center} This is the incidence matrix of a $(6,3,2)$-design, an example of a balanced incomplete block design, meaning any pair of rows are both 1 in exactly 2 coordinates. Let $D \subseteq \mathbb{F}_{3}^{6}$ be the vectors with non zero entries given in the coordinates corresponding to the 1s in the matrix. By this, we mean $D$ is the set of vectors whose non-zero coordinates are 123, 124, 135, 146, 156, 236, 245, 256, 345 or 346. \smallbreak Let $D'$ be the remaining vectors of $\mathbb{F}_{3}^{6}$ with 3 non zero entries. There are $\binom{6}{3} \times \ 2^3 = 160$ vectors of weight 3, $|D|$ = $2^3 \times 10 = 80$, so $|D| = |D'| = 80$. \smallbreak Let $R$ be the vectors with no zeros, and an even number of 1s, let $R'$ be the other weight 6 vectors, with an odd number of 1s. $|R| = |R'| = 32$. \smallbreak Define $A_1 = D \cup R$, $A_2 = D' \cup R$, and let $A_0$ be the vectors of weight 1. Then $A_0$ is a cap set of size 12 and $A_1, A_2$ are cap sets of size 112 in $\mathbb{F}_{3}^{6}$. Furthermore, $|A_1 \cap A_2|$ = 32 and $A_1 + A_2 = \mathbb{F}_{3}^{6} \setminus A_0$. This all follows from the properties of a block design, and by checking that $D + D'$, $R+R$, $D+R$ and $D' + R$ don't contain any weight 1 vectors. \bigbreak \textbf{This collection is extendable:} \smallbreak Let $x,y \in A_0$. Since all elements of $A_0$ have weight 1, $x+y$ must have weight $0, 1$ or $2$. If $z \in A_1 \cup A_2$ is such that $x + y + z = 0$, then $z$ needs to have the same weight as $x+y$. Since $A_1 \cup A_2$ consists only of vectors of weights 3 or 6, we cannot have such a $z$. So, there are no solutions to $x+y+z=0$ where $x,y \in A_0$ and $z \in A_1 \cup A_2$, which is condition (1) for extendable. \smallbreak Let $x \in A_1$ and $y \in A_2$. Since $A_1 + A_2 = \mathbb{F}_{3}^{6} \setminus A_0$, and $z \in A_0 \iff 2z \in A_0$, there is no $z \in A_0$ such that $x + y + z = 0$. This is condition (2) for extendable, and so we have an extendable collection of cap sets. \end{proof} We are now ready to prove the main result of this section, and obtain our first new lower bound for maximal cap sets. We will use several results and examples from this section, to show the following. \begin{theorem} \label{thm:396} There exists a cap set $A \subseteq \mathbb{F}_3^{396}$ of size \[\binom{11}{7} \cdot 6^4 \cdot 12^4 \cdot 112^{62}.\] \end{theorem} \begin{proof} First, we take the extendable collection $(A_0, A_1, A_2)$ from \eqref{extendable caps}, and extend it by the recursively admissible set $S = \tilde{I}(6,5)$, which exists by \eqref{lemma:m-1}. This gives a cap set $B \subseteq \mathbb{F}_{3}^{36}$ by \eqref{lemma:extended product cap}, where $|B| = 6 \times 112^5 \times 12$ by \eqref{lemma:size}, and an extendable collection $(B, A_{1}^6, A_{2}^6)$ by \eqref{lemma:recursive}. \smallbreak Then we extend our new collection $(B, A_{1}^6, A_{2}^6)$ by $T = I(11,7)$ from \eqref{computer caps}, to produce a cap set in $\mathbb{F}_3^{396}$, which by \eqref{lemma:size} has size \[\binom{11}{7} \cdot (6 \times 112^5 \times 12)^4 \cdot (112^6)^7 .\] \end{proof} \begin{remark} \label{otherbounds} Since $|A|^{1/396} \approx 2.217981$, we see that our cap in $\mathbb{F}_3^{396}$ gives an exponential improvement on the asymptotic lower bound on the size of a cap set. We note that Edel's lower bound uses essentially the same method, but with the admissible sets $S=\tilde{I}(8,7)$ and $T=I(10,5)$. Using either $I(10,6)$ or $I(11,6)$ from \eqref{computer caps} produces a better lower bound than Edel, but neither is better than our new bound of $\approx 2.217981^n$. \end{remark} \section{Extending The Admissible Set Construction} In this section, we will extend Edel's methods, by mimicking the extended product for cap sets to find large admissible sets. This will allow us to construct large admissible sets, much larger than is computationally feasible, which will be used to obtain a lower bound of $2.218^n$. \smallbreak The following proposition is similar to \eqref{def:ProdCap}, where we showed that the direct product of cap sets is a cap set. \begin{proposition} \label{prop: ad prod} If $S, T$ are admissible sets, then so is their direct product $S \times T$. \end{proposition} \begin{proof} \textbf{Pairwise condition:} \\Let $a,b \in S \times T$ be distinct. Then $a,b$ are of the form $(s,t)$, $(s',t')$ for some $s,s' \in S$ and $t,t' \in T$. If $s,s' \in S$ are distinct, there are coordinates $i,j$ such that $s_i = 0 \neq s_i'$ and $s_j \neq 0 = s_j'$ by condition (1) of admissibility for S. The same is true for $t,t' \in T$, and since $(s,t) \neq (s',t')$, we must have $s \neq s'$ or $t \neq t'$. It follows that for all $a \neq b \in S \times T$, we have coordinates $i,j$ such that $a_i = 0 \neq b_i$ and $a_j \neq 0 = b_j$. So condition (1) for admissibility is satisfied. \bigbreak \textbf{Triples condition:} \\Let $a,b,c \in S \times T$ be distinct. As before, we must have $a = (s,t)$, $b = (s',t')$ and $c = (s'',t'')$ for some (not necessarily distinct) $s,s',s'' \in S$ and $t,t',t'' \in T$. \\ \textbf{Case 1:} If $s,s',s''$ or $t,t',t''$ are all distinct, then condition (2) of admissibility follows immediately by the admissibility of $S$ or $T$. \\ \textbf{Case 2:} Assume that neither $s,s',s''$ nor $t,t',t''$ are all distinct. Without loss of generality, we may assume $s = s'$. Since $(s,t) \neq (s',t')$, we cannot also have $t=t'$, but since $t,t',t''$ are not all distinct, without loss of generality $t=t''$. So, we must have: $a = (s,t)$, $b = (s,t')$ and $c = (s'',t)$. Since $a,b,c$ are distinct, we must have $s \neq s''$. By condition (1) of admissibility of S, there is a coordinate $k$ such that $s_k = 0 \neq s''_k$. So, $\{a_k, b_k, c_k\} = \{0,0,1\}$ or $\{0,0,2\}$, hence condition (2) for admissibility also holds for $S \times T$. \end{proof} \begin{remark} Now we have a direct product construction for admissible sets, there are at least 2 natural questions: \begin{enumerate} \item Does the direct product construction allow us to produce better admissible sets, by combining known admissible sets? \item Can we generalise the direct product construction for admissible sets, in an analogous way to the extension construction \eqref{def:extended product} for cap sets? \end{enumerate} \end{remark} The answer to the first question is no: much like the situation for direct products of cap sets, taking a direct product of admissible sets only ever does as well as the best of the individual admissible sets. \smallbreak We focus on the 2nd question - can we improve the product construction for admissible sets, in a similar way to the extended product construction for cap sets? We will answer this question in the affirmative. In particular, in this section we will construct an admissible set in $\{0,1,2\}^{1562}$, which we use to prove \eqref{main} and obtain the lower bound of $2.218^n$. \begin{definition}[Meta-admissible] \label{def:meta ad} We say a set $T \subseteq \{0, 1,2\}^r$ is \emph{meta-admissible} if it is admissible, as in \eqref{def:admissible}. Recall that admissible means: \begin{enumerate} \item For all distinct $t, t' \in T$, there are coordinates $i$ and $j$ such that $t_i = 0 \neq t_i'$ and $t_j \neq 0 = t_j'$. \item For all $t, t', t'' \in T$ distinct, there is a coordinate $k$ such that $\{t_k, t_k', t_k''\} = \{0, 1, 2\}$, $\{0,0,1\}$ or $\{0, 0, 2\}$. \end{enumerate} \end{definition} \begin{definition}[Meta-extendable] \label{def:meta ext} A collection $S_0, S_1, S_2 \subseteq \{0,1,2\}^m$ of admissible sets is \emph{meta-extendable} if: \begin{enumerate} \item For any $s \in S_0$ and $s' \in S_1 \cup S_2$, the weight of $s$ is less than the weight of $s'$, so all the vectors in $S_0$ have more zeroes than any vector in $S_1$ or $S_2$. \item If $x,y \in S_0$ and $z \in S_1 \cup S_2$ then there is a coordinate $k$ such that $\{x_k, y_k, z_k\} = \{0, 1, 2\}$, $\{0,0,1\}$ or $\{0, 0, 2\}$. \item If $x \in S_0$, $y \in S_1$ and $z \in S_2$, then there is a coordinate $k$ such that $\{x_k, y_k, z_k\} = \{0, 1, 2\}$, $\{0,0,1\}$ or $\{0, 0, 2\}$. \end{enumerate} \end{definition} We have a similar construction to \eqref{def:extended product}, but this time we extend a collection of admissible sets rather than cap sets. \begin{definition} \label{def:meta con} If $S_0, S_1, S_2$ are admissible sets, and $T \subseteq \{0, 1,2\}^r$, for each $t = (t_1, \ldots, t_r) \in T$ we define \[t(S_0, S_1, S_2) = S_{t_1} \times \cdots \times S_{t_r}.\] Predictably, we then define \[T(S_0, S_1, S_2) = \bigcup_{t \in T} t\bra{S_0, S_1, S_2} = \bigcup_{t \in T} \bra{S_{t_1} \times \cdots \times S_{t_r}}.\] \end{definition} \bigbreak It will not come as a shock that our new definitions of meta-admissible \eqref{def:meta ad} and meta-extendable \eqref{def:meta ext} allow us to use the extended product construction on admissible sets \eqref{def:meta con} to produce new admissible sets. \begin{lemma} If $S_0, S_1, S_2 \subseteq \{0,1,2\}^m$ is a meta-extendable collection of admissible sets, and $T \subseteq \{0, 1,2\}^r$ is meta-admissible, then $T(S_0, S_1, S_2)$ is an admissible set. \end{lemma} \begin{proof} \textbf{First condition for admissible (pairs):} Let $x,y \in T(S_0, S_1, S_2)$ be distinct. So we know $x \in S_{t_1} \times \cdots \times S_{t_r}$ and $y \in S_{t'_1} \times \cdots \times S_{t'_r}$ for some $t,t' \in T$. There are 2 cases: $t=t'$ or $t \neq t'$. \smallbreak If $t=t'$, we know $x,y \in S_{t_1} \times \cdots \times S_{t_r}$, which is admissible by \eqref{prop: ad prod}. \smallbreak Assume $t \neq t'$. By condition (1) of meta-admissible, there is a coordinate $i$ such that $t_i = 0 \neq t_i'$, and $j$ such that $t_j \neq 0 = t'_j$. Without loss of generality, assume $t'_i < t_j$. So \[x \in S_{t_1} \times \cdots \times S_{0} \times \cdots \times S_{t_j} \times \cdots \times S_{t_r}\] and \[y \in S_{t'_1} \times \cdots \times S_{t'_i} \times \cdots \times S_0 \times \cdots \times S_{t'_r}.\] Let $s \in S_0$, $s' \in S_{t'_i}$. Since $t'_i \neq 0$, by property (1) of meta-extendable $s$ has lower weight than $s'$. By the pigeonhole principle there must be a coordinate $k$ such that $s_{k} = 0 \neq s'_k$, as $s$ has more zero entries than $s'$, so $s$ must be zero somewhere $s'$ is not. Similarly, for $s'' \in S_0$, $s''' \in S_{t_j}$, since $t_j \neq 0$ there is a coordinate $\ell$ such that $s''_{\ell} = 0 \neq s'''_{\ell}$. So, it follows that there are coordinates $i', j'$ such that $x_{i'} = 0 \neq y_{i'}$ and $x_{j'} \neq 0 = y_{j'}$, so we have the first condition for admissibility in this case too. \bigbreak \textbf{Second condition of admissibility (triples):} Let $x,y,z \in T(S_0, S_1, S_2)$ be distinct. If $x,y,z$ all come from the same $t \in T$ then we are done, by the direct product construction. So, we need to consider the other cases: $x,y$ come from $t$ and $z$ comes from $t'$ or $x,y,z$ come from distinct $t, t', t''$ \smallbreak If $x,y$ are from $t$ and $z$ is from $t'$ where $t \neq t'$, then by definition of $T$ being meta-admissible, there is a coordinate $i$ such that $t_i = 0 \neq t'_i$. So, the $i$-th blocks $x_i, y_i$ of $x,y$ are from $S_0$, and the $i$-th block $z_i$ of $z$ is in $S_1 \cup S_2$. Then by condition (2) of meta-extendable applied to $x_i, y_i, z_i$, there is a coordinate $k$ where $\{x_k, y_k, z_k\} = \{0, 1, 2\}$, $\{0,0,1\}$ or $\{0, 0, 2\}$. \smallbreak Finally, if $x,y,z$ from distinct $t, t', t''$ then by condition (2) of meta-admissible, we have a coordinate $k$ where either $\{t_k, t'_k, t''_k\} = \{0, 0, 1\}$, $\{0, 0, 2\}$ or $\{0, 1, 2\}$. \smallbreak In the first case, two of $x_k$, $y_k$, $z_k$ are in $S_0$ and the other one is from $S_1 \cup S_2$, and we are done as above, with condition (2) of meta-extendable. \smallbreak If $\{t_k, t'_k, t''_k\} = \{0, 1, 2\}$, then exactly one of $x_k$, $y_k$, $z_k$ is in each of $S_0$, $S_1$, $S_2$. Then by condition 3 of meta-extendable, we are done. \end{proof} \smallbreak \begin{lemma} \label{metacoll} There is a meta-extendable collection of admissible sets $S_0, S_1, S_2$, where $S_1$ and $S_2$ are both $I(11,7)$ admissible sets, and $S_0 \subseteq I(11,3)$ has size 37. \end{lemma} \begin{proof} We use $S_1 = I(11,7)$ from \eqref{computer caps}, and then we take $S_2$ as the set formed by swapping all 1s and 2s in $S_1$. Note that this is still an admissible set, since the doubles condition is unaffected by swapping nonzero coordinates, and if there was a coordinate $k$ where $\{x_k, y_k, z_k\} = \{0,0,1\}$, $\{0,0,2\}$ or $\{0,1,2\}$, then after swapping 1s and 2s we will have $\{x_k, y_k, z_k\} = \{0,0,1\}$, $\{0,0,2\}$ or $\{0,1,2\}$. \smallbreak Using a computer search, we can find $S_0 \subseteq I(11,3)$ such that $|S_0| = 37$ and $S_0, S_1, S_2$ satisfy the conditions for meta-extendable given in \eqref{def:meta ext}. The admissible set $S_0 \subseteq I(11,3)$ can be found on the author's webpage at \url{fredtyrrell.com/cap-sets}. \end{proof} \begin{lemma} There is an admissible set $T' \subseteq \{0, 1,2\}^{1562}$ such that $|T'| = 142 \cdot 37 \cdot \binom{11}{7}^{141}$ and all $t \in T'$ have weight 990. \end{lemma} \begin{proof} Let $T = I(142,141)$, which exists by \eqref{lemma:recursive}. Using the meta-extendable collection $(S_0, S_1, S_2)$ from \eqref{metacoll} above, we can then produce a new admissible set $T'=T(S_0, S_1, S_2)$. Since $S_0, S_1, S_2 \subseteq \{0,1,2\}^{11}$, and $11 \times 142 = 1562$, we see that $T(S_0, S_1, S_2) \subseteq \{0,1,2\}^{1562}$. \smallbreak Each element $T(S_0, S_1, S_2)$ contains 141 $S_1$ or $S_2$ blocks, and one $S_0$, so has weight $141 \times 7 + 3 = 990$. For each $t \in T$, the set $t(S_0, S_1, S_2)$ has $\binom{11}{7}^{141} \cdot 37$ different vectors, so $|T'| = 142 \cdot 37 \cdot \binom{11}{7}^{141}$. \end{proof} Using this new admissible set, we can finally prove the main result of this paper. \begin{proof}[Proof of theorem \eqref{main}] We use the admissible sets $S=\tilde{I}(6,5)$, which exists by \eqref{lemma:recursive}, and $T' \subseteq \{0, 1,2\}^{1562}$ from the previous lemma. First, we apply the recursively admissible set $S$ to $(A_0, A_1, A_2)$, the 6 dimensional extendable collection of cap sets from \eqref{extendable caps}, to produce the extendable collection $(B, A_1^6, A_2^6)$, where $B$ has size $6 \cdot 112^5 \cdot 12$. Then we apply $T'$, and our final cap set has size \[|T'| \cdot |B|^{1562-990} \cdot (112^6)^{990} = \binom{11}{7}^{141} \cdot 6^{572} \cdot 12^{572} \cdot 112^{8800} \cdot 37 \cdot 142.\] \end{proof} \begin{remark} This cap set in $\mathbb{F}_3^{56232}$ has size $|A| \approx 2.21 \times 10^{19455} \approx 10^{10^{4.3}}$, and since $|A|^{1/56232} \approx 2.218021$, this example gives a slight improvement to the lower bound in \eqref{thm:396}. Similar to the remark \eqref{otherbounds} at the end of section 2, it is possible to find other meta-extendable collections. For example, there is a collection $S_1 = I(11,6)$, $S_2$ is flipped $S_1$ and $S_0 \subseteq I(11,2)$ of size 20. However, none of these give a better bound than the above. \end{remark} \subsection{Summary} Now we have proved all of the lower bounds in this paper, it seems a good idea to present the results we have achieved, alongside the previous lower bounds. Note that the bounds of Edel in \cite{Edel} and Calderbank and Fishburn in \cite{CalderbankFishburn} both come from what we are now calling the extended product construction. \smallbreak \begin{center} \renewcommand{\arraystretch}{1.5} \begin{tabular}{ |c|c|c|c| } \hline \textbf{Bound} & \textbf{Construction} & \textbf{Dimension} & \textbf{Appears in} \\ [1ex] \hline\hline 2 & $\{0,1\}^n$ & All & Trivial \\[1ex] \hline 2.114742\ldots & Maximal cap of size 20 in $\mathbb{F}_3^4$ & 4 & \cite{Pellegrino} \\ [1ex] \hline 2.141127\ldots & Maximal cap of size 45 in $\mathbb{F}_3^5$ & 5 & \cite{10.1006/jcta.2002.3261} \\ [1ex] \hline 2.195514\ldots & Maximal cap of size 112 in $\mathbb{F}_3^6$ & 6 & \cite{dim6} \\ [1ex] \hline 2.210147\ldots & $\tilde{I}(25,24)$ and $I(90,89)$ & 13500 & \cite{CalderbankFishburn} \\ [1ex] \hline 2.217389\ldots & $\tilde{I}(8,7)$ and ${I}(10,5)$ & 480 & \cite{Edel} \\ [1ex] \hline \hline 2.2175608\ldots & $\tilde{I}(7,6)$ and ${I}(10,6)$ & 420 & \eqref{otherbounds} \\[1ex] \hline 2.217950\ldots & $\tilde{I}(7,6)$ and ${I}(11,6)$ & 462 & \eqref{otherbounds} \\[1ex] \hline 2.217981\ldots & $\tilde{I}(6,5)$ and ${I}(11,7)$ & 396 & \eqref{thm:396} \\[1ex] \hline 2.218021\ldots & Meta-extendable collection & 56232 & \eqref{main} \\[1ex] \hline \end{tabular} \end{center} \section{Limits to the admissible set construction} Given that our new lower bounds have come from finding new admissible sets, it is natural to ask what happens if we continue to find more admissible sets. In particular, we would like to know how much we could improve the lower bound by, if all admissible sets were to exist. Edel gave an answer to this question in the very final section of \cite{Edel}, which we present as the following proposition. \begin{proposition} \label{prop:asymptotic limit} For a collection $(A_0, A_1, A_2)$ of extendable cap sets in $\mathbb{F}_3^{n}$, where $|A_1| = |A_2|$, the best admissible sets are those of the form $I(m,m \alpha)$ where $\alpha = \frac{|A_1|}{|A_0|+|A_1|}$ and $m$ is large. Using the extended product construction, the best constant we could achieve in our asymptotic lower bound is $c = \left(|A_0| + |A_1|\right)^{1/{n}}$. \end{proposition} \begin{proof} Let $\alpha \in (0,1)$. If we apply the admissible set $I(m,m \alpha)$ to $(A_0, A_1, A_2)$, we have a cap set in $\mathbb{F}_3^{nm}$ of size \[\binom{m}{m \alpha} \cdot |A_1|^{m \alpha} \cdot |A_0|^{m(1-\alpha)}.\] For large $m$, we can use the well known asymptotic estimate $\binom{m}{m\alpha} \sim 2^{m h(\alpha)}$, where $h(x) = -x \log_2(x) - (1-x)\log_2(1-x)$ is the binary entropy function. Taking logs, applying a change of base and removing constants, we see that we want to maximise the function \[f(x) = x \log\bra{\frac{|A_1|}{|A_0|}} - x\log(x) - (1-x)\log(1-x).\] The derivative is given by $f'(x) = \log\bra{\frac{|A_1|}{|A_0|}} + \log(1-x) - \log(x)$, which has its root at $x = \frac{|A_1|}{|A_0|+|A_1|}$. We can then substitute $w=m \alpha = m\frac{|A_1|}{|A_0|+|A_1|}$ back into our formula from \eqref{lemma:size} for the size of the cap set, giving a cap set in $\mathbb{F}_3^{nm}$ of size $\left(|A_0|+|A_1|\right)^m$. \end{proof} \begin{remark} For the collection of cap sets in $\mathbb{F}_3^6$ from \eqref{extendable caps}, the best admissible sets are those of the form $I\left(m, \frac{28m}{31}\right)$ for large $m$, and the best asymptotic lower bound we could get using these 6 dimensional cap sets is $\left(124^{1/6}\right)^n = (2.233\ldots)^n$. \end{remark} Our experimental evidence, combined with some heuristic arguments (and perhaps a little wishful thinking), lead us to explicitly state the following conjecture, which was implied at the end of \cite{Edel}. \begin{conjecture} \label{conj1} For any $m>w>0$, there always exists an $I(m,w)$ admissible set. Therefore, the maximum size of a cap set in $\mathbb{F}_3^n$ is at least $124^{n/6} \approx 2.233^n$. \end{conjecture} \begin{remark} In addition to all admissible sets in dimensions up to 11, we can prove the existence of admissible sets of weight 2 and 3 for all $m$, via a similar construction to \eqref{lemma:recursive}. These, combined with the admissible sets $I(m,0)$, $I(m,1)$, $I(m,m-1)$ and $I(m,m)$, are all of the admissible sets currently known to exist. \end{remark} The following table shows the lowest dimension admissible set which would be needed to achieve new bounds, using the extended product construction in \eqref{def:extended product} with the cap sets in $\mathbb{F}_3^6$ from \eqref{extendable caps}. \smallbreak \begin{center} \renewcommand{\arraystretch}{1.4} \begin{tabular}{ |c|c|c|c| } \hline \textbf{Bound} & \textbf{Admissible Sets} & \textbf{Dimension} \\[1ex] \hline\hline 2.220\ldots & $\tilde{I}( 5 , 4 )$ and $I( 17 , 11 )$ & 510 \\[1ex] \hline 2.225\ldots & $\tilde{I}( 3 , 2 )$ and $I( 54 , 41 )$ & 972 \\[1ex] \hline 2.230\ldots & $I( 311 , 281 )$ & 1866 \\[1ex] \hline 2.233\ldots & $I( 22948 , 20727 )$ & 137688 \\[1ex] \hline 2.233076\ldots & $I\left(m, \frac{28m}{31}\right)$ for large $m$ & 6$m$ \\[1ex] \hline \end{tabular} \end{center} \smallbreak While small improvements to our bound are possible by constructing better admissible sets, we do not expect that a significant increase in the lower bound is possible by simply finding more admissible sets through a computer search. It is perhaps not a huge surprise then that although Edel was able to find an $I(10,5)$ about 20 years ago, the best we have been able to do is an $I(11,7)$. \section{The SAT Solver} \begin{definition}[Boolean satisfiability] The Boolean satisfiability problems asks whether, given a propositional formula, there exists an assignment of true or false to each variable in the formula such that the formula is true. If this is the case, we say that the formula is \emph{satisfiable}. \end{definition} Significant research has gone into finding efficient algorithms for the Boolean satisfiability problem, known as SAT solvers. We used the kissat SAT solver, which is described in \cite{SAT}. Most SAT solvers, including kissat, take inputs in a format called conjuctive normal form (CNF). \begin{definition}[Conjunctive normal form] A propositional formula is in \emph{conjuctive normal form} if the formula consists of a conjunction of clauses, where each clause is a disjunction of propositional variables or their negations. \end{definition} Our use of the SAT solver requires three steps. First, we convert the problem of a particular admissible set existing into a statement in CNF. Once we have defined the problem in CNF, we can use a SAT solver program to check whether this formula is satisfiable. If our formula is satisfiable, the SAT solver returns the assignment of the variables which satisfy the formula, which we convert back into a set of vectors, producing our admissible set. The first step is the interesting one, which we will discuss in this section. \subsection{CNF algorithm} \subsubsection{Variables} We begin by generating the $\binom{m}{w}$ different support sets. In other words, for each $1 \leq i \leq \binom{m}{w}$ we have a different $S_i \subseteq \{1, \ldots, m\}$, where $|S_i| = w$. We then define a variable for each non zero coordinate in each support set - let $S_i^k$ correspond to the element $k$ in $S_i$, meaning the $k$ coordinate of the $i$th support set. There are $w \cdot \binom{m}{w}$ such variables $S_i^k$. \smallbreak For any pair of support sets, we want to record when a given coordinate is different. So, for each coordinate in the first support set, we define a variable which is true if this coordinate is the same in the second support set as well, and false if it is not. We can represent this variable as $S_{i,j}^k$, corresponding to the pair $S_i, S_j$ of support sets, and the coordinate $k$. This variable $S_{i,j}^k$ is true if coordinate $k$ is in both support sets and $S_i^k = S_j^k$, and is false if coordinate $k$ is only in $S_i$ or $S_i^k \neq S_j^k$. \subsubsection{Reconciling the pair and triple variables} We now add constraints, starting with constraints which combine the single variables and the pairwise variables. This is basically just a common sense constraint, to make sure the variables looking at individual coordinates and the variables looking at pairs of coordinates are compatible. \smallbreak For any pair $S_i$ and $S_j$ of different support sets, for every coordinate $k$ in both $S_i$ and $S_j$ we ask that either $S_{i,j}^k$ is true or $\{S_i^k, S_j^k\} = \{1,2\}$. An equivalent way to phrase this is as follows: take any pair $x,y$ with different support sets. For each coordinate $k$ where both $x_k$ and $y_k$ are non zero we add the constraints $(x_k=y_k) \lor (x_k=1) \lor (y_k=1)$ and $(x_k=y_k) \lor (x_k=2) \lor (y_k=2)$. \smallbreak In other words, if the variable which records when $x_k \neq y_k$ and $x_k$, $y_k >0$ is true, then exactly one of the 2 variables which record whether $x_k$ or $y_k$ are 1 is true. This is essentially just saying that $x_k \neq 0 \neq y_k$ and $x_k \neq y_k$ implies $\{x_k, y_k\} = \{1,2\}$. \subsubsection{Checking the condition on triples} We now ensure every triple $x,y,z$ in our set has a coordinate $k$ such that $\{x_k,y_k,z_k\} = \{0,0,1\}$, $\{0,0,2\}$ or $\{0,1,2\}$. This is the triples condition for admissible sets from \eqref{def:admissible}. \smallbreak Take any 3 distinct support sets, and check all of the coordinates from $1$ to $m$. If a coordinate is in exactly one of the 3 support sets, we are done and don't need to worry about this triple of support sets - we will be automatically guaranteed a coordinate where our triple is $\{0,0,1\}$ or $\{0,0,2\}$. \bigbreak The problem we need to consider is when there is no coordinate in exactly one of the 3 supports. That is, every coordinate in one of the three supports is in at least one of the others. In this case, we need to look at the coordinates in exactly 2 of the supports, and force one of these coordinates to give us $\{0,1,2\}$. If $k$ is in the support of $x,y$, but not $z$, we can add the condition $x_k \neq y_k$, which would mean $\{x_k, y_k, z_k\} = \{0,1,2\}$. \smallbreak We do this for every coordinate in exactly 2 of the support sets of $x,y,z$, and we then take the disjunction of these conditions, meaning we require this to be true for only one coordinate. This produces a constraint which asks for $\{x_k,y_k,z_k\} = \{0,1,2\}$ in at least one coordinate $k$. So, if this is satisfied for all triples without a coordinate in exactly one of the three supports, we are done, and have an admissible set. \subsection{Improving the SAT solver} In order for the SAT solver to return an output in a reasonable time, we add more constraints to the problem, reducing the search space of all potential assignments and hence hopefully allowing the SAT solver to work faster. We need to be clever with our choice of extra conditions, to make algorithm more efficient without turning the problem into one which is impossible. By studying smaller examples of admissible sets, heuristic arguments and a healthy dose of educated guesswork, we tried various combinations of further constraints on our problem. Some made the problem unsatsfiable, and some still did not allow the SAT solver to finish within a reasonable time. However, we used this information to refine our method, and eventually we were successful in producing 3 new admissible sets. The following additional constraints allowed us to find the admissible sets $I(11,7)$, $I(11,6)$ and $I(10,6)$. \subsubsection{Extra constraints for $I(11,7)$} \begin{enumerate}[label=(\roman*)] \item Every vector must have the first 2 non zero entries different - that is, the first 2 non zero coordinates are always $(1,2)$ or $(2,1)$. \item Every vector must have at least one of each digit 1 and 2 in their 4th, 5th or 6th coordinates - they cannot all be 1 or all be 2. \item If the third non zero entry is in coordinates 1 to 7, it is always a 1. \item If the fourth non zero entry is in coordinates 1 to 7, it is always a 2. \end{enumerate} \subsubsection{Extra constraints for $I(11,6)$} \begin{enumerate}[label=(\roman*)] \item Every vector must have the first 2 non zero entries different - that is, the first 2 non zero coordinates are always $(1,2)$ or $(2,1)$. \item Every vector must have at least one of each digit 1 and 2 in the final 3 non zero coordinates - they cannot all be 1 or 2. \item If the third non zero entry is in coordinates 1 to 7, it is always a 1. \end{enumerate} \subsubsection{Extra constraints for $I(10,6)$} \begin{enumerate}[label=(\roman*)] \item If the second non zero entry is in the first 6 coordinates, make it a 1. \item If the third non zero entry is in the first 6 coordinates, make it a 2. \item No vector ends in $(2,2)$. \end{enumerate} \begin{remark} The CNF generation code and the code to transform the SAT output back to vectors can be found on the author's website, at \url{fredtyrrell.com/cap-sets}. We would like to make it clear that our use of the SAT solver is unlikely to be the most efficient way to produce admissible sets, and we would be very interested in suggestions to improve this aspect of our construction, from those with more experience and knowledge of SAT solvers and other computational methods. \end{remark} \section*{Acknowledgements} I am extremely grateful to Thomas Bloom for suggesting the problem and supervising my research. I would like to thank Thomas for his support, encouragement and advice throughout my project. In addition, I thank him for many helpful discussions and useful suggestions in preparing this paper. \smallbreak Thanks also go to Akshat, Albert, Ittihad, Maria, and Yifan for being an excellent audience for my presentations in the Maths Institute, where they provided me with a valuable opportunity to share and discuss my research. \smallbreak Finally, I would like to thank Anubhab, Chris, Claire, Flora, James and Jess for allowing me to try and explain my work to them, with varying levels of success. \smallbreak The work in this paper was completed while the author was employed as a Summer Project Intern in the Mathematical Institute, University of Oxford, under the supervision of Dr Thomas Bloom, and was supported by the supervisor's Royal Society University Research Fellowship. \printbibliography \end{document}

https://arxiv.org/abs/2004.06097

Saturation problems in the Ramsey theory of graphs, posets and point sets

In 1964, Erdős, Hajnal and Moon introduced a saturation version of Turán's classical theorem in extremal graph theory. In particular, they determined the minimum number of edges in a $K_r$-free, $n$-vertex graph with the property that the addition of any further edge yields a copy of $K_r$. We consider analogues of this problem in other settings. We prove a saturation version of the Erdős-Szekeres theorem about monotone subsequences and saturation versions of some Ramsey-type theorems on graphs and Dilworth-type theorems on posets.We also consider semisaturation problems, wherein we allow the family to have the forbidden configuration, but insist that any addition to the family yields a new copy of the forbidden configuration. In this setting, we prove a semisaturation version of the Erdős-Szekeres theorem on convex $k$-gons, as well as multiple semisaturation theorems for sequences and posets.

\section{Introduction} Extremal problems have a long history in combinatorics originating with the results of Mantel~\cite{Mantel} in 1907 and Turán~\cite{Turan} in 1947 determining the maximum number of edges in a triangle- and $K_r$-free, $n$-vertex graph, respectively. Erdős, Hajnal and Moon~\cite{ehm} investigated the dual problem, called the saturation problem, wherein one aims to minimize the number of edges in a $K_r$-free, $n$-vertex graph, such that the addition of any edge yields a copy of $K_r$. Since their initial result, the saturation problem has been considered for a variety of graphs. Of particular note is a theorem of Kászonyi and Tuza~\cite{Tuza}, which showed that for any graph $H$, the minimum number of edges in an $H$-saturated, $n$-vertex graph is at most linear in $n$. Going beyond graphs, saturation problems have been considered in several other settings. A structure which is maximal with respect to some property is said to \emph{saturate} that property. A maximum size saturating structure is called an \emph{extremal structure}, while a minimum size saturating structure is called a \emph{minimal saturating structure}. For intersecting hypergraphs, a saturation version of the Erdős-Ko-Rado theorem~\cite{ekr} (uniform setting) was proven by Füredi~\cite{furedi}. In particular, he showed that for a given uniformity $r$, there exists a family of approximately $3r^2/4$ sets of size $r$ with the property that adding any further set yields a pair of disjoint sets, disproving a conjecture of Meyer~\cite{meyer}. In the nonuniform setting, it is well known that all maximal intersecting families of subsets of an $n$ element set have the same size, namely $2^{n-1}$. This result was extended to the case of families without $k$-matchings by Buci\'c~\emph{et al.}~\cite{bucic}. In the setting of forbidden (induced or non-induced) posets in the Boolean lattice the saturation problem has been investigated by Ferrara~\emph{et al.}~\cite{Ferrara}, and further results in this direction were obtained in~\cite{mar} and~\cite{satsolved}. Parallel with the development of extremal combinatorics, Ramsey theory has been investigated extensively. This topic begins with the seminal result of Ramsey~\cite{ramsey}, which states that for any integers $c,r,k$ there is an integer $N$ such that any $c$-coloring of the edges of an $r$-uniform hypergraph on $N$ vertices contains a monochromatic complete graph of size $k$ in some color. This initial result gave rise to a variety of problems where in place of a complete hypergraph one is given hypergraphs $F_1,F_2,\dots,F_c$, and seeks to minimize the value of $N$ which yields, for all $c$-colorings of the complete $r$-uniform, $N$-vertex hypergraph, a copy of $F_i$ in color $i$ for some $i=1,2,\dots,c$. Ramsey-type problems may be interpreted as extremal problems in the following way. One wishes to maximize the number of vertices $n$ in such a way that there exists a coloring, such that for all $i$, we find no copy of $F_i$ in color $i$ (so $n=N-1$, where $N$ is defined as above). With this interpretation, it becomes natural to ask the corresponding minimal saturation problem, wherein we seek to minimize $n$ such that the hypergraph can be $c$-colored without a monochromatic copy of $F_i$ in color $i$, but if we extend this $c$-colored hypergraph to a $c$-colored hypergraph on $n+1$ vertices, then we have for some $i$, a monochromatic copy of some $F_i$ in color $i$. Finally, we mention that many classical results, such as Dilworth's theorem on posets and the Erd\H{o}s-Szekeres theorem for sequences or cups and caps, can be interpreted as Ramsey-type problems where the allowed colorings of the hypergraph are restricted in some way. As such, we may again consider the corresponding saturation versions of these results. In this paper, we initiate such a study of Ramsey-type saturation problems. We concentrate on well-known settings (graphs, posets, monotone and convex subsets of point sets), and in several cases we manage to prove tight bounds. In addition, we also consider the corresponding semisaturation problems, a notion introduced by F\"{u}redi and Kim~\cite{semi} (also sometimes called oversaturation or strong saturation). In the graph setting the semisaturation problem is the following: Given a graph $F$, what is the minimum number of edges in an $n$-vertex graph $G$ with the property that adding any edge to $G$ yields a copy of $F$ containing that edge. Note that now we allow the graph $G$ to contain $F$ as a subgraph. We will consider semisaturation problems for sequences, cups and caps, posets and point sets as well as for the Ramsey problem on graphs. Note that by definition, the semisaturation number is always at most the saturation number which is in turn at most the extremal (Ramsey) number. In the rest of this section we provide a precise formulation of each of the saturation and semisaturation problems that are considered in the paper and our results for each case. We also briefly summarize the known results about the corresponding extremal problem in order to contrast them with our minimal saturation results. Sections~\ref{section:graph}--\ref{section:conv} contain the proofs of these results. Finally, in Section~\ref{general} we rigorously define the general framework that was hinted at above and illustrate how these problems fit into it. \subsection*{Graphs} Let $\mbox{\ensuremath{\mathcal G}}\xspace$ be the family of (labeled) complete graphs whose edges are colored with $c$ colors (numbered by $1,2,\dots, c$). Given $G, G'\in \mbox{\ensuremath{\mathcal G}}\xspace$, we say $G'$ \emph{extends} $G$ if $G$ is a proper subgraph of $G'$, i.e., $G'$ can be obtained from $G$ by iteratively adding a new vertex and colored edges connecting the new vertex with each of the existing vertices. A member $G$ of $\mbox{\ensuremath{\mathcal G}}\xspace$ is called \emph{$(k_1, k_2, \ldots, k_c)$-saturated} if for every $i\in [c]$, the graph $G$ does not contain a monochromatic $K_{k_i}$ of color $i$, but every $G' \in \mbox{\ensuremath{\mathcal G}}\xspace$ that extends $G$ contains a monochromatic $K_{k_i}$ of color $i$ for some $i$. A graph $G \in \mbox{\ensuremath{\mathcal G}}\xspace$ is called \emph{$(k_1, k_2, \ldots, k_c)$-semisaturated} if for every $G'\in \mbox{\ensuremath{\mathcal G}}\xspace$ that extends $G$, there exists some $i\in [c]$ such that $G'$ contains a copy of a monochromatic $K_{k_i}$ of color $i$ which is not in $G$. Clearly the size of the largest $(k_1, k_2, \ldots, k_c)$-saturated graph in $\mbox{\ensuremath{\mathcal G}}\xspace$, which we denote by $\ram_{\mbox{\ensuremath{\mathcal G}}\xspace}(k_1,\dots, k_c)$, is equal to the usual Ramsey number minus one. Let $\sat_{\mbox{\ensuremath{\mathcal G}}\xspace}(k_1,\dots k_c)$ denote the size of the smallest saturated $G\in \mbox{\ensuremath{\mathcal G}}\xspace$, and finally let $\osat_{\mbox{\ensuremath{\mathcal G}}\xspace}(k_1,\dots k_c)$ denote the size of the smallest $(k_1, \dots, k_c)$-semisaturated $G\in \mbox{\ensuremath{\mathcal G}}\xspace$. From the definition it is clear that \[\osat_{\mbox{\ensuremath{\mathcal G}}\xspace}(k_1,\dots, k_c)\le \sat_{\mbox{\ensuremath{\mathcal G}}\xspace}(k_1,\dots, k_c) \le \ram_{\mbox{\ensuremath{\mathcal G}}\xspace}(k_1,\dots, k_c).\] For convenience, we also use $\sat_{\mbox{\ensuremath{\mathcal G}}\xspace}(k;c)$ to denote $\sat_{\mbox{\ensuremath{\mathcal G}}\xspace}(k_1, k_2, \dots, k_c)$ and $\osat_{\mbox{\ensuremath{\mathcal G}}\xspace}(k;c)$ to denote $\osat_{\mbox{\ensuremath{\mathcal G}}\xspace}(k_1, k_2, \dots, k_c)$ when $k_1 = k_2 = \cdots = k_c = k$. For a fixed $k$ and growing $l$, the following results about Ramsey numbers are known (due to Bohman and Keevash~\cite{BK} and Ajtai, Komlós and Szemerédi~\cite{AKS}, respectively): \[c'_k\frac{l^{\frac{k+1}{2}}}{(\log l)^{\frac{k+1}{2}-\frac{1}{k-2}}}\le \ram_{\mbox{\ensuremath{\mathcal G}}\xspace}(k,l)\le c_k\frac{l^{k-1}}{(\log l)^{k-2}}.\] For the case $l=3$, it is known that \[\ram_{\mbox{\ensuremath{\mathcal G}}\xspace}(k,3)=\Theta\left(\frac{k^2}{\log k}\right).\] The upper bound was proven by Ajtai, Komlós, and Szemerédi~\cite{AKS}; the lower bound was obtained originally by Kim~\cite{Kim}, and subsequently improved by Fiz Pontiveros, Griffiths and Morris~\cite{FGM} and Bohman and Keevash~\cite{BK}. We prove the following results: \begin{theorem}\label{theorem:satgraphs} For two colors, \[\osat_{\mbox{\ensuremath{\mathcal G}}\xspace}(k,l)= \sat_{\mbox{\ensuremath{\mathcal G}}\xspace}(k,l)= (k-1)(l-1),\] and for more than two colors, \[(k_1-1)(k_2+\dots+k_c-2c+3)\le \osat_{\mbox{\ensuremath{\mathcal G}}\xspace}(k_1, \ldots, k_c)\le \sat_{\mbox{\ensuremath{\mathcal G}}\xspace}(k_1, \ldots, k_c)\le (k_1-1)\cdots (k_c-1).\] In the latter lower bound we can exchange $k_1$ with any other $k_i$. \end{theorem} In the case when $k_1=k_2=\dots =k_c=k$, Theorem~\ref{theorem:satgraphs} implies that $\sat_{\mbox{\ensuremath{\mathcal G}}\xspace}(k;c)\le (k-1)^c$. Using an idea of Pálvölgyi~\cite{pdperscomm}, with probabilistic methods we improve this bound in the case when $c$ is large compared to $k$. \begin{theorem}\label{thm:random} $\osat_{\mbox{\ensuremath{\mathcal G}}\xspace}(k;c)\le 48k^2c^{k^2}$. \end{theorem} \subsection*{Posets} In this paper we are also interested in saturation problems on partially ordered sets (posets). Dilworth's theorem~\cite{Dilworth} answers a Ramsey-type problem about posets, since it implies that a poset of size $(k-1)(l-1)+1$ contains either a chain of length $k$ or an antichain of length~$l$. A natural saturation and semisaturation version of this problem can be posed. Let $\mbox{\ensuremath{\mathcal P}}\xspace$ denote the set of all finite posets. Given $P = (S, \leq_P)$, $P'=(S',\leq_{P'})\in \mbox{\ensuremath{\mathcal P}}\xspace$, we say $P'$ \emph{extends} $P$ if $S\subsetneq S'$ and for all $x,y\in S$, $x\leq_{P} y$ if and only if $x\leq_{P'} y$. A poset $P\in \mbox{\ensuremath{\mathcal P}}\xspace$ is $(k,l)$-semisaturated if every poset $P'\in \mbox{\ensuremath{\mathcal P}}\xspace$ extending $P$ contains a $k$-chain or an $l$-antichain which is not completely contained in $P$. Similarly as before, $\osat_{\mbox{\ensuremath{\mathcal P}}\xspace}(k,l)$ denotes the minimum size of such a semisaturated poset. If $P$ is additionally $k$-chain and $l$-antichain free, then we say that $P$ is $(k,l)$-saturated. We use $\sat_{\mbox{\ensuremath{\mathcal P}}\xspace}(k,l)$ to denote the minimum size of such a saturated poset. We also define $\ram_{\mbox{\ensuremath{\mathcal P}}\xspace}(k,l)$ as the maximum size of a $(k,l)$-saturated poset. Using this notation, Dilworth's theorem implies that \[\ram_{\mbox{\ensuremath{\mathcal P}}\xspace}(k,l)=(k-1)(l-1).\] For the semisaturation number of posets we show the following. \begin{theorem}\label{thm:weakGeneralPoset} \[\osat_{\mbox{\ensuremath{\mathcal P}}\xspace}(k,1)=0,\; \osat_{\mbox{\ensuremath{\mathcal P}}\xspace}(k,2)=k-1.\] For $l\ge 3$, we have \[\osat_{\mbox{\ensuremath{\mathcal P}}\xspace}(k,l)= \min(2k+l-5,k+3l-7).\] \end{theorem} For the saturation numbers of general posets, we prove the following theorem. \begin{theorem}\label{thm:satGeneralPoset} \[\sat_{\mbox{\ensuremath{\mathcal P}}\xspace}(k,l)=(k-1)(l-1)=\ram_{\mbox{\ensuremath{\mathcal P}}\xspace}(k,l).\] \end{theorem} When the Ramsey and the saturation numbers are the same we gain further insight into the structure of the saturated objects. For example every saturated object has the same size. Other examples of this kind include the intersecting families of subsets of an $n$ element set mentioned in the introduction and, as we will see, sequences without increasing and decreasing subsequences of given lengths. \subsection*{Monotone point sets and sequences} Another well-known Ramsey-type result is the Erdős-Szekeres theorem on monotone point sets. A point set in general position is said to be monotone increasing (resp. decreasing) if when ordered according to the $x$-coordinates, the $y$-coordinates of the points are monotone increasing (resp. decreasing). The theorem of Erdős and Szekeres~\cite{ES} states that a set of $(k-1)(l-1)+1$ points in general position contains either an increasing subsequence of length $k$ or a decreasing subsequence of length $l$. There is an equivalent formulation of this result in terms of sequences. It states that a sequence of $(k-1)(l-1)+1$ numbers must contain either an increasing subsequence of length $k$ or a decreasing sequence of length $l$. We will work with both of these formulations. We can convert this problem into a saturation problem in the usual way. A sequence $S$ of distinct numbers (resp. point set with distinct $x$- and $y$-coordinates) is called $(k,l)$-saturated if it does not contain an increasing subsequence (resp. subset) of length $k$ or a decreasing subsequence (resp. subset) of length $l$ but any sequence (resp. point set with distinct $x$- and $y$-coordinates) $S'$ that contains $S$ as a subsequence (resp. subset) has either an increasing subsequence (resp. subset) of length $k$ or a decreasing subsequence (resp. subset) of length $l$. The functions $\sat_{\mbox{\ensuremath{\mathcal S}}\xspace}(k,l)$ and $\osat_{\mbox{\ensuremath{\mathcal S}}\xspace}(k,l)$ are defined analogously to as before. Moreover, we define $\ram_{\mbox{\ensuremath{\mathcal S}}\xspace}(k,l)$ to be the maximum size of a $(k,l)$-saturated sequence. With this notation the Erdős-Szekeres theorem states that \[\ram_{\mbox{\ensuremath{\mathcal S}}\xspace}(k,l)=(k-1)(l-1).\] For saturation numbers, we prove the following theorem. \begin{theorem}\label{thm:seqSat} \[\sat_{\mbox{\ensuremath{\mathcal S}}\xspace}(k,l) = (k-1)(l-1)=\ram_{\mbox{\ensuremath{\mathcal S}}\xspace}(k,l).\] \end{theorem} In other words Theorem~\ref{thm:seqSat} says that if a sequence of distinct numbers does not contain an increasing subsequence of length $k$ or a decreasing subsequence of length $l$, then either we can extend the sequence without creating such a subsequence or the length of the sequence is $(k-1)(l-1)$. For semisaturation numbers we have the following theorem. \begin{theorem}\label{thm:weakSatSeq} \[\osat_{\mbox{\ensuremath{\mathcal S}}\xspace}(1,l) = \osat_{\mbox{\ensuremath{\mathcal S}}\xspace}(k,1) = 0.\] For $k, l\geq 2$, we have \[\osat_{\mbox{\ensuremath{\mathcal S}}\xspace}(k,l) = \min(2k+l-5, 2l+k-5).\] \end{theorem} \subsection*{Convex point sets} Finally, we investigate the saturation problem for convex point sets in the plane. If a set of $n$ points is in convex position, then we say that the points form a \emph{convex $n$-set}. A set of $k$ points in convex position is called a $k$-cup (resp. $k$-cap) if the points lie on the graph of a convex (resp. concave) function (possibly multivalued). Let $\ram_{\mbox{\ensuremath{\mathcal C}}\xspace\C}(k,l)$ denote the size of the largest set of points in general position which contains neither a subset forming a $k$-cup nor a subset forming an $l$-cap. Similarly, let $\sat_{\mbox{\ensuremath{\mathcal C}}\xspace\C}(k,l)$ denote the size of the smallest point set which contains neither a subset forming a $k$-cup nor a subset forming an $l$-cap, such that adding any new point yields a $k$-cup or $l$-cap. Finally, let $\osat_{\mbox{\ensuremath{\mathcal C}}\xspace\C}(k,l)$ denote the size of the smallest point set such that adding any new point introduces a new $k$-cup or a new $l$-cap. In 1935, Erdős and Szekeres~\cite{ES} showed that \[\ram_{\mbox{\ensuremath{\mathcal C}}\xspace\C}(k,l)={\binom{k+l-4} {k-2}}.\] While we were not able to obtain non-trivial bounds for the saturation problem, for the semisaturation problem we have the following result. \begin{theorem}\label{thm:cupcapSemiSat} We have \begin{align*} \osat_{\mbox{\ensuremath{\mathcal C}}\xspace\C}(k,3)=\osat_{\mbox{\ensuremath{\mathcal C}}\xspace\C}(3,k)&=k-1.\\ \end{align*} For $k\geq4$, we have \begin{align*} \osat_{\mbox{\ensuremath{\mathcal C}}\xspace\C}(k,4)=\osat_{\mbox{\ensuremath{\mathcal C}}\xspace\C}(4,k)&=2k-2, \end{align*} and for $k\geq 5$ and $l \ge 5$, \[2k+2l-12 \le \osat_{\mbox{\ensuremath{\mathcal C}}\xspace\C}(k,l)\le 2k+2l-10.\] \end{theorem} The original motivation for investigating point sets free of cups and caps was to give an upper bound, namely ${ \binom{2n-4} {n-2}}$, on the maximum number of points in the plane avoiding a convex $n$-gon. Erd\H{o}s and Szekeres provided a lower bound of size $2^{n-1}$, and after a number of subsequent improvements a nearly optimal upper bound of size $2^{n+o(n)}$ was provided by Suk~\cite{suk}. An intriguing problem is to obtain the analogous saturation result. \begin{problem}\label{convex_sat} What is the minimum possible size of a point set in the plane in general position which does not contain a convex $n$-set, but adding any extra point (in general position) creates one? \end{problem} We could not even determine if the answer is polynomial in $n$ or not. Note that if we drop the general position assumption the problem becomes trivial, one can simply take $n-1$ points on a line. For the respective semisaturation problem, let $\osat_{\mbox{\ensuremath{\mathcal C}}\xspace}(n)$ denote the minimum possible size of a point set in the plane general position, such that adding any extra point to it (in general position) creates a new convex $n$-set. With this notation we prove the following theorem. \begin{theorem}\label{thm:convex} \[\osat_{\mbox{\ensuremath{\mathcal C}}\xspace}(n) = 2n-4.\] \end{theorem} Unlike the problem about cups and caps, this problem generalizes easily to higher dimensions. Let $\osat_{\mbox{\ensuremath{\mathcal C}}\xspace,d}(n)$ denote the minimum possible size of a point set in $\mathbb{R}^d$, such that it is in general position and adding one extra point to it (in general position) creates a new convex $n$-set. We obtain the following result. \begin{theorem}\label{thm:osat00} \[\osat_{\mbox{\ensuremath{\mathcal C}}\xspace,d}(n) \ge n-1 + \floor{\frac{n-2}{d}}.\] \end{theorem} \section{Graphs}\label{section:graph} \begin{proof}[Proof of Theorem \ref{theorem:satgraphs}] First we prove the upper bounds. For $c=2$, one sharp construction is when the blue (the first color) edges form the graph consisting of $l-1$ disjoint copies of $K_{k-1}$. Another construction is when the red (the second color) edges form the graph consisting of $k-1$ disjoint copies of $K_{l-1}$. It is easy to see that these two-edge-colored graphs are saturated. It is also easy to generalize these constructions to $c>2$ colors: Start with a graph $G_0$ with a single vertex. Now one by one for each color $i$ with $1\leq i\leq c$, construct a colored graph $G_i$ by replacing each vertex $v_j$ of $G_{i-1}$ with a clique $S_j$ of size $k_i-1$ such that all edges in the clique are in color $i$. The edges between every pair of such cliques $S_{j_1}$ and $S_{j_2}$ are in the same color as the edge $v_{j_1}v_{j_2}$ in $G_{i-1}$. It is not hard to see that the graph $G_c$ obtained by this construction is saturated. Now we prove the lower bound for $c=2$. Take a minimal two-edge-colored semisaturated graph (with respect to a blue $k$-clique and red $l$-clique). Extract maximal complete blue subgraphs (cliques) greedily one by one until we partition all the vertices into cliques (for a more advanced treatment of such greedy partitions see~\cite{gyorikeszegh}). Assume that the first $i$ blue cliques have size at least $k-1$ and the rest have size at most $k-2$. If $i\ge l-1$, then $G$ has at least $(k-1)(l-1)$ vertices and we are done. Otherwise when $i<l-1$, let $p$ be a new vertex, which we add to $G$ and connect with red edges to the first $i$ cliques and blue edges to the rest of the vertices. It is easy to see that in the resulting graph there is neither a blue clique of size $k$ containing $p$ nor a red clique of size $l$ containing $p$. Hence $G$ is not semisaturated, which contradicts our assumption. Finally, we prove the lower bound for $c>2$. Again, take a minimal $c$-edge-colored semisaturated graph $G$. If there is no clique of size $k_1-1$ of the first color in $G$, then we can connect a new vertex $p$ to every vertex with the first color. It follows that $G$ is not semisaturated, giving us a contradiction. Otherwise there exists a clique of size $k_1-1$ with color $1$ in $G$. Take such a clique $S$ and connect $p$ to every vertex in $S$ with the second color. Now $G-S$ must again contain a clique of size $k_1 -1$ with color $1$ otherwise we can connect $p$ to the rest of the vertices in $G$ with color one. We can connect $p$ to the vertices of this clique as well with the second color if $k_2>2$. Repeating this argument, we keep finding additional disjoint cliques of size $k_1-1$ of the first color. When we have $k_2-2$ such cliques, we continue to pull out cliques of size $k_1-1$ of the first color and connect them to $p$ with color $3$. Continuing in this way we find altogether $(k_2-2)+(k_3-2)+\dots+(k_c-2)+1$ cliques of size $k_1-1$ of color $1$, showing that indeed the number of vertices is at least $(k_1-1)(k_2+k_3+\dots+k_c-2c+3)$. As the role of the first color was not used, the same way we can find enough cliques of color $i$. \end{proof} In the case of two colors we have determined the exact bound for $\osat_{\mbox{\ensuremath{\mathcal G}}\xspace}$. Using the following trivial observation we see that the next open case is when $c=3$ and $k_1=k$, $k_2=k_3=3$, for which Theorem~\ref{theorem:satgraphs} gives the lower bound $3(k-1)$ and upper bound $4(k-1)$ for both the saturation and semisaturation problems. \begin{obs}\label{obs:kis2} \[\sat_{\mbox{\ensuremath{\mathcal G}}\xspace}(2,k_1,\dots, k_c)=\sat_{\mbox{\ensuremath{\mathcal G}}\xspace}(k_1,\dots, k_c),\] \[\osat_{\mbox{\ensuremath{\mathcal G}}\xspace}(2,k_1,\dots, k_c)=\osat_{\mbox{\ensuremath{\mathcal G}}\xspace}(k_1,\dots, k_c).\] \end{obs} We state an equivalent formulation of the $c=3$, $k_1=k$, $k_2=k_3=3$ case as an problem. \begin{problem} Is it true that the vertices of every $3$-edge colored complete graph $G$ with $n=4(k-1)$ vertices can be partitioned into three parts, the first part avoiding a $K_{k-1}$ of the first color, the second part avoiding edges of the second color and the third part avoiding edges of the third color? This is equivalent to the $\osat$ problem, in the $\sat$ variant we further assume that $G$ itself avoids $K_{k}$ of the first color and triangles of the second and third colors. \end{problem} We now give a probabilistic argument improving the upper bound in Theorem~\ref{theorem:satgraphs} in some cases. \begin{proof}[Proof of Theorem \ref{thm:random}] Consider a uniform random coloring of the edges of the complete graph $K_{nc}$ with $c$ colors, that is, each edge is assigned one of the $c$ colors uniformly randomly and independently. We claim that the resulting edge-colored graph $G$ is semisaturated with positive probability if $n$ is large enough. To show that $G$ is semisaturated, it suffices to show that the subgraph of $G$ induced by any set of $n$ vertices contains a monochromatic $K_{k-1}$ of each color. Indeed, if we add an extra vertex $q$ to $G$ and color the edges incident to $q$ in any way to get the graph $G'$, then by pigeonhole principle there will be $n$ edges incident to $q$ having the same color $d$. Since the endpoints of those $n$ edges (other than $q$) induce a subgraph in $G$ that contains a monochromatic $K_{k-1}$ of each color, we then can find a new $K_k$ in color $d$ in $G'$. Note that since we pick the color of each edge uniformly randomly and independently, each color class can be considered as an instance of the Erd\H{o}s-R\'enyi random graph $G(nc,\frac{1}{c})$. So we need to bound the probability of $G(nc,\frac{1}{c})$ having $n$ vertices whose induced subgraph does not contain a copy of $K_{k-1}$. We first need many pairwise edge-disjoint copies of $K_k$ in $K_n$. For our purposes the following simple lemma is enough: \begin{lemma}\label{decomp} We can find $\frac{1}{16k^2}n^2$ pairwise edge-disjoint copies of $K_k$ in $K_n$ for any $n\ge 4k^2$. \end{lemma} \begin{proof} Lemma~\ref{decomp} can be proved in many ways. For example it easily follows from the following construction. Let $\{(i,j)|i\in [k], j\in[r]\}$ be the first $kr$ vertices of the $K_n$ where $r$ is a the largest prime such that $kr\le n$. From Bertrand's postulate we know that $r$ is at least $\floor {\frac{n}{2k}}$. Since $n>4k$ we have $\floor{\frac{n}{2k}}\ge \frac{n}{2k}-1\ge \frac{n}{4k}$. Consider the cliques whose vertex set has the following form: $\{(1,a+b),(2,a+2b),\dots, (k,a+bk) \}$ where $a,b\in [r]$ and the second coordinates are understood modulo $r$. It is easy to see that this gives us $\frac{1}{16k^2}n^2$ disjoint $K_k$'s. Indeed, suppose otherwise that two copies share an edge, then for some values $(a,b)\ne(a',b')$, $i\ne j$ we would have $a+ib=a'+ib'$ and $a+jb=a'+jb'$ modulo $r$, implying $(i-j)(b-b')=0$ modulo $r$. As $1\le i,j$ and $i,j\le k\le r$ ($n\ge 4k^2$ implies that $k\le r$), this gives $b=b'$, which in turn implies $a=a'$, a contradiction. \end{proof} Now we can calculate a bound on the probability of $G(n,p)$ containing a $K_{k-1}$ (where $p=1/c$). Let $n\ge 4k^2$ to be chosen later. First we fix $\frac{1}{16k^2}n^2$ disjoint copies of $K_{k-1}$ using Lemma \ref{decomp}. For each copy the probability that it is not in $G(n,p)$ is $1-p^{\binom{k-1}{2}}$. Since the cliques are edge-disjoint, the probability that no $K_{k-1}$ appears at all is at most $\left ( 1-p^{\binom{k-1}{2}}\right )^{\frac{n^2}{16k^2}}\le e^{-p^{\binom{k-1}{2}}\frac{n^2}{16k^2}}$. Returning to the original problem we see that there are $\binom{cn}{n}\le (ec)^n$ ways to choose $n$ vertices out of $nc $. The colors have a symmetric role in the problem. Hence by the union bound, the probability that we can find $n$ vertices and a color such that there is no $K_{k-1}$ of that color among those $n$ vertices is at most \[c(ec)^n e^{-p^{\binom{k-1}{2}}\frac{n^2}{16k^2}}.\] Picking $n=3\log(c)16k^2c^{\binom{k-1}{2}}$ we get \[c(ec)^n e^{-p^{\binom{k-1}{2}}\frac{n^2}{16k^2}}\le e^nc^{n+1}e^{-3\log(c)n}< 1.\] Therefore the probability of the bad cases is less than $1$. So we can find a semisaturated graph on $cn=c\cdot 3\log(c)16k^2c^{\binom{k-1}{2}}\le 48k^2 c^{k^2}$ vertices. \end{proof} \section{Posets}\label{section:poset \begin{proof}[Proof of Theorem \ref{thm:weakGeneralPoset}] If $l=1$, then obviously adding an element to the empty poset will introduce an antichain of size $1$. Thus $\osat_{\mbox{\ensuremath{\mathcal P}}\xspace}(k,1)=0$. Consider the case when $l=2$. If a newly added element is incomparable with any of the elements of the poset, then we find a new antichain of size two. So if $P$ is a poset which is not semisaturated for $l=2$, then we must be able to add an element comparable to all elements of $P$ without introducing a new chain of size $k$. This is possible if and only if $P$ does not contain a chain of size $k-1$. On the other hand, the smallest poset containing a chain of size $k-1$ is the poset containing only this chain on $k-1$ elements, and this poset is clearly semisaturated for $2$-antichains and $k$-chains. Thus $\osat_{\mbox{\ensuremath{\mathcal P}}\xspace}(k,2)=k-1$ (and this is the only semisaturating poset of this size). Now we may assume that $l\ge 3$. We show two semisaturated posets, one with $2k+l-5$ elements and one with $k+3l-7$ elements. For the first construction consider an antichain $A$ of size $l-1$ and then add two chains, $C_1$ and $C_2$, of length $k-2$ to the poset such that the $C_1$ lies below the elements of $A$ and $C_2$ lies above them (see Figure~\ref{fig:theo5}). The resulting poset is semisaturated. Indeed, if we add $p$ such that it is not comparable to any element of $A$, then $A\cup\{p\}$ is an antichain of length $l$. If $p$ lies below an element $a$ in $A$, then $C_2\cup\{a,p\}$ is a chain of length $k$. Similarly if $p$ lies above an element $a$ in $A$ then $C_1\cup\{a,p\}$ is a chain of length $k$. Therefore we cannot add $p$ without creating a chain of size $k$ or an antichain of size $l$. \begin{figure}[!ht] \centering \begin{minipage}{0.4\textwidth} \include{posetconst4} \end{minipage} \begin{minipage}{0.4\textwidth} \include{posetconst3} \end{minipage} \caption{The two constructions for $k=6$, $l=5$.} \label{fig:theo5} \end{figure} The second construction starts with two disjoint antichains $A_1$ and $A_2$ of length $l-1$. Then we add a chain $C$ of length $k-3$ between them, that is, every element of $C$ is above every element of $A_1$ and below every element of $A_2$. Finally we add an antichain $B$ of $l-2$ elements such that they are incomparable to everything. To see that this construction is semisaturated suppose that we can add an element $p$ without creating a chain of size $k$. Then $p$ cannot be above any element of $A_2$ nor below any element of $A_1$. On the other hand $p$ must be comparable to some elements $a_1\in A_1$ and $a_2\in A_2$ otherwise we would get an antichain of length $l$. Consequently, $p$ is above $a_1$ and below $a_2$. We know that $C\cup \{a_1,a_2\}$ is a chain of length $k-1$ so there must be an element $c\in C$ such that $p$ and $c$ are incomparable otherwise $C\cup\{a_1,a_2,p\}$ would be a chain of size $k$. Since $B\cup \{c\}$ is an antichain and $B \cup \{p,c\}$ cannot be an antichain, $p$ must be comparable to some element $q\in B$. If $p$ is above $q$ then $a_2$ is comparable to $q$ through $p$. If $p$ is below $q$ then $a_1$ is comparable to $q$ through $p$. But neither case is possible since $q$ is incomparable to both $a_1$ and $a_2$ in the original poset. To show that we need at least $\min(2k+l-5,k+3l-7)$ elements for semisaturation we start with the following observation. Let $P$ be a semisaturated poset, and let $L$ denote those elements that are the minimal elements of a chain of length $k-1$ in $P$. We claim that $|L|\ge l-1$. To see this, observe that if we add an element $p$ below every element of $P\setminus L$ and incomparable to every element of $L$ (this is possible since $L$ is clearly a downset in the poset), then no chain of length $k$ is created. Since $P$ is semisaturated, it follows that $p$ must be in an antichain of length $l$ which lies in $L\cup \{p\}$. Thus, $|L|\ge l-1$ and $L$ contains an antichain $L'$ of size $l-1$. Similarly we define $U$ to be the set of elements that are maximal elements of a chain of length $k-1$. In the same way we can see that $|U|\ge l-1$, and $U$ contains an antichain $U'$ of size $l-1$. If $U\cap L\ne \emptyset$, then there is a chain of length $2k-3$. Since this chain intersects $L'$ in at most one element, we have at least $2k-3+l-1-1=2k+l-5$ elements. If $U\cap L= \emptyset$ and $1\le k\le 3$, then the number of elements is at least $|U|+|L|\ge 2l-2\ge 2k+l-5$, as required (using that $l\ge 3$). From now on we assume $k\ge 3$ and consider two cases. First suppose that every element of $U$ is comparable to every element of $L$. Then we can add $p$ to the poset such that it lies below every element of $U$, above every element of $L$ and incomparable to every other element. Suppose $p$ creates a new chain $C$ of length $k$. Clearly $C\subset U\cup L\cup \{p\}$. By symmetry we may assume that $|C\cap U|\ge \ceil{\frac{k-1}{2}}$. Let $u$ be the first element in $C$ above $p$. Since $u\in U$ there is a chain $C_2$ of size $k-1$ ending in $u$. Consequently $(C\cap U) \cup C_2$ is a chain of length at least $k-1+\ceil{\frac{k-1}{2}}-1$. In total we have found $|U'\cup L'\cup ((C\cap U) \cup C_2)| $ elements. Any chain intersects any antichain in at most one element, therefore \[|U'\cup L'\cup ((C\cap U) \cup C_2)|\ge l-1+l-1+k-1+\ceil{\frac{k-1}{2}}-1-2=2l+k+\ceil{\frac{k-1}{2}}-6.\] As $\min(k+3l-7,2k+l-5)$ is at most \[\floor{\frac{(2k+l-5)+(k+3l-7)}{2}}=\floor{2l+\frac{3}{2}k-6}=2l+k+\floor{\frac{k}{2}}-6=2l+k+\ceil{\frac{k-1}{2}}-6,\] we have at least $\min(k+3l-7,2k+l-5)$ elements. Suppose now that adding $p$ does not create a new chain of length $k$. Then since $P$ is $(k,l)$-semisaturated, it must happen that adding $p$ creates a new antichain of length $l$. Since $p$ is comparable to the elements of $U$ and $L$ we have found $l-1$ new elements that are not in $L\cup U$. Therefore we have three disjoint antichains of length $l-1$ ($L'$, $U'$ and this antichain containing $p$). We must also have a chain of length $k-1$ and this chain intersects each of the three antichains in at most one element. Therefore we have at least $3(l-1)+k-1-3=k+3l-7$ elements. The only remaining case is when $U\cap L= \emptyset$ and we can find $u\in U$ and $q\in L$ such that $u$ and $q$ are incomparable. This implies that the chains going up from $q$ and going down from $u$ are disjoint. So we have two disjoint chains of length $k-1$, giving us at least $2(k-1)+2(l-1)-4=2k+2l-8\ge 2k+l-5$ elements. \end{proof} \begin{proof}[Proof of Theorem \ref{thm:satGeneralPoset}] Given a poset $P$ with fewer than $(k-1)(l-1)$ elements which contains no $k$-chain and no $l$-antichain, we need to show that we can always add an element $p$ to $P$ in such a way that the resulting poset still avoids $k$-chains and $l$-antichains. If the maximum length of an antichain in $P$ is at most $l-2$, then we can easily add a new element to $P$ incomparable with all elements of $P$, and thus still avoid $k$-chains and $l$-antichains. Suppose now that the size of the maximal antichain is $l-1$. By Dilworth's theorem we can decompose $P$ into $l-1$ chains. By our assumption that we are $k$-chain free, all of them have size at most $k-1$ and at least one of them has size strictly less than $k-1$. Denote one such chain by $C$. For an element $c\in C$ denote by $D_c$ the subchain of $C$ consisting of $c$ and the elements that are below $c$ in $C$. Similarly, $U_c$ is the subchain of $C$ containing $c$ and the elements above $c$ in $C$. First suppose that there is no chain of size $k-1$ in $P$ whose bottom element is the bottom element $q'$ of $C$. Then we add a new element $p$ directly under $q'$ and incomparable with all the elements that are not above $q'$ (Figure~\ref{fig:theo6a}, left side). We claim that $P\cup \{p\}$ still avoids $k$-chains and $l$-antichains and so $P$ was not saturated, a contradiction. Indeed, as the poset can still be partitioned into $l-1$ chains (the former partition of $P$ with the difference that $C$ is extended with $p$ as a new bottom element under $q'$), it follows that there is no antichain of length $l$. Also a chain of length $k$ must have $p$ as its bottom element, and then $q'$ as its element directly above $p$, but then this chain minus $p$ would be a chain of length $k-1$ in $P$ with bottom element $q'$, contradicting our assumption. The case when there is no chain of size $k-1$ in $P$ whose top element is the top element of $C$ is handled similarly. \begin{figure}[!ht] \centering\begin{minipage}{0.4\textwidth} \includegraphics{theorem6fig.pdf} \end{minipage} \begin{minipage}{0.4\textwidth} \includegraphics{theorem6fig3.pdf} \end{minipage} \caption{Adding $p$ to the bottom of $C$ and finding a long chain if $q$ is above $r$.} \label{fig:theo6a} \end{figure} Thus, we may assume that there is a largest element $q$ of $C$ for which there is a chain $C_q$ of size $k-1$ containing $q$ whose part below $q$ (including $q$) coincides with $D_q$. Similarly, there is a smallest element $r$ of $C$ for which there is a chain $C_r$ of size $k-1$ containing $r$ whose part above $r$ (including $r$) coincides with $U_r$. We claim that $r$ is above $q$. Suppose on the contrary that $q$ is above $r$ or that they coincide. Then taking the part of $C_q$ above $q$, the part of $C$ between $q$ and $r$ and the part of $C_r$ below $r$ we get a chain $C'$ whose length is at least $k$, a contradiction. Indeed, the sum of $\abs{C}$ and $\abs{C'}$ is the same as the sum of $\abs{C_q}$ and $\abs{C_r}$, and so as $C_q$ and $C_r$ have $k-1$ elements and $C$ has at most $k-2$ elements, it follows that $C'$ must have at least $k$ elements, a contradiction (see Figure~\ref{fig:theo6a}). Thus $q$ is not the top element of $C$ and there is an element $s$ of $C$ directly above it. Now add a new element $p$ directly above $q$ and below $s$ such that $p$ is incomparable to all elements that are not below $q$ or above $s$ (Figure~\ref{fig:theo6b}). We claim that $P\cup \{p\}$ still avoids $k$-chains and $l$-antichains and thus $P$ was not saturated, a contradiction. The poset $P\cup \{p\}$ can still be partitioned into $l-1$ chains by taking the previous partition except that $C$ is extended with $p$ put between $q$ and $s$. Thus, the resulting poset still avoids $l$-antichains. Suppose now that it contains a $k$-chain, it necessarily contains $p$, and then $q$ is directly below $p$ in the chain and $s$ is directly above $p$ in the chain. Deleting $p$ from this $k$-chain we get a $(k-1)$-chain $C'$ in $P$ which has $q$ and $s$ directly above each other. Let $D'_q$ be the part of $C'$ below $q$ (including $q$) and $U'_{s}$ be the rest of $C'$, that is the part of $C'$ above $s$ (including $s$). By the definition of $q$ there is a chain of size $k-1$ whose bottom part is $D_q$, thus $D'_q$ can have size at most as much as $D_q$ otherwise the top part of this chain extended with $D'_q$ would be a chain of size at least $k$. Again by the maximality of $q$, the chain formed by $D_q$ and $U'_{s}$ can have size at most $k-2$, thus the chain formed by $D'_q$ and $U'_{s}$ can also have size at most $k-2$, but this is exactly $C'$ which was of size $k-1$, a contradiction. \begin{figure}[!ht] \centering \centering\begin{minipage}{0.4\textwidth} \includegraphics[width=5cm]{theorem6fig2.pdf} \end{minipage} \begin{minipage}{0.4\textwidth} \includegraphics[width=5cm]{theorem6fig2b.pdf} \end{minipage} \caption{Inserting $p$ into the poset.} \label{fig:theo6b} \end{figure} \end{proof} \section{Saturation of monotone point sets and sequences}\label{section:strongseq Throughout this section, the term sequence will always refer to a sequence of distinct real numbers. We may, without loss of generality, assume that the numbers in the sequence are positive. \begin{definition} A sequence $S$ is $(k,l)$-saturated if $S$ contains no increasing subsequence of length $k$ nor decreasing subsequence of length $l$, but any sequence $S'$ containing $S$ as a proper subsequence contains either an increasing subsequence of length $k$ or a decreasing subsequence of length $l$. Let $\sat_{\mbox{\ensuremath{\mathcal S}}\xspace}(k,l)$ denote the minimum possible length of a $(k,l)$-saturated sequence. \end{definition} Let $\vec{a} = [a_1, a_2,\dots, a_m]$ be a $(k+1,l+1)$-saturated sequence. We define the function ${\gamma_{\vec{a}}:[m] \to [l] \times [k]}$ by $\gamma_{\vec{a}}(t)=(i,j)$, where $i$ is the length of the longest decreasing subsequence of $\vec{a}$ ending at $a_t$ and $j$ is the length of the longest increasing subsequence of $\vec{a}$ ending at $a_t$. Since for every $n' < n\le m$ we can extend either the longest increasing subsequence or the longest decreasing subsequence ending at $a_{n'}$ by appending $a_n$ to the end, we have the following observation. \begin{obs}\label{obs:inc} If $n'<n$, then at least one coordinate of $\gamma_{\vec{a}}(n)$ is strictly larger than the corresponding coordinate of $\gamma_{\vec{a}}(n')$. \end{obs} Define an $l \times k$ matrix $R_{\vec{a}} = (r_{ij})$ corresponding to the sequence $\vec{a}$ by setting \[r_{ij} = \begin{cases} \gamma_{\vec{a}}^{-1}(i,j) & \textrm{if $(i,j) \in \image(\gamma_{\vec{a}})$,} \\ 0 & \textrm{otherwise.} \end{cases}\] Define an $l \times k$ matrix $V_{\vec{a}} = (v_{ij})$ corresponding to the sequence $\vec{a}$ by setting \[v_{ij}= \begin{cases} a_{\gamma_{\vec{a}}^{-1}(i,j)} & \textrm{if $(i,j) \in \image(\gamma_{\vec{a}})$,} \\ 0 & \textrm{otherwise.} \end{cases} \] Finally, let $W_{\vec{a}}=(w_{ij})$ be the $l \times k$ matrix such that the $i$-th row of $W_{\vec{a}}$ is the $(l+1 -i)$-th row of $V_{\vec{a}}$ for every $i \in [l]$. For example, consider the sequence $\vec{a}=[33,11,22,55,44]$ with $k=3$, $l=2$. In this case $R_{\vec{a}}=\begin{pmatrix} 1 & 0 & 4\\ 2 & 3 & 5 \end{pmatrix}$ and $V_{\vec{a}}=\begin{pmatrix} 33 & 0 & 55\\ 11 & 22 & 44 \end{pmatrix}$ and thus $W_{\vec{a}}=\begin{pmatrix}11 & 22 & 44\\ 33 & 0 & 55 \end{pmatrix}$. \begin{definition} We say that a matrix $(m_{ij})$ is \emph{partially increasing} if all the elements are nonnegative, the positive values are distinct and $i_1\le i_2$, $j_1\le j_2$ implies $m_{i_1j_1}\le m_{i_2j_2}$ for all nonzero elements of the matrix. \end{definition} \begin{lemma}\label{lem:young_tableau} $R_{\vec{a}}$ and $W_{\vec{a}}$ are both partially increasing. \end{lemma} \begin{proof} Observation~\ref{obs:inc} implies that $R_{\vec{a}}$ is partially increasing. To show that $W_{\vec{a}}$ is partially increasing we need to prove that if we take $i_1\ge i_2$ and $j_1\le j_2$, then $v_{i_1,j_1}\le v_{i_2,j_2}$ whenever $v_{i_1,j_1}$ and $v_{i_2,j_2}$ are positive numbers. Assume on the contrary that $v_{i_1,j_1}> v_{i_2,j_2}>0$. Let $n, n'$ be the indices such that $v_{i_1,j_1}=a_n>a_{n'}=v_{i_2,j_2}$. First, if $n<n'$, then a decreasing subsequence of length $i_1$ ending in $a_{n}$ can be extended with $a_{n'}$ and thus the longest increasing subsequence ending in $a_{n'}$ has length at least $i_1+1$, which contradicts that $i_2 \leq i_1$. Second, if $n>n'$, then an increasing subsequence of length $j_2$ ending in $a_{n'}$ can be extended with $a_{n}$ and thus the longest increasing subsequence ending in $a_{n}$ has length at least $j_2+1$, which contradicts that $j_1 \leq j_2$. \end{proof} Let $\mathbb{S}_{k,l}$ be the set of extremal $(k+1, l+1)$-saturated sequences of length $kl$ whose entries are distinct integers in $[kl]$. Observe that when $\vec{a} \in \mathbb{S}_{k,l}$, $R_{\vec{a}}$ and $W_{\vec{a}}$ have all positive entries, and are increasing in both rows and columns by Lemma~\ref{lem:young_tableau}. Before moving on to the proof of Theorem~\ref{thm:seqSat}, we briefly discuss how our results relate to the classification of extremal sequences for the Erd\H{o}s-Szekeres theorem in terms of Young tableaus. As $R_{\vec{a}}$ and $W_{\vec{a}}$ have values from $[kl]$, they correspond to a pair of standard rectangular Young tableaus $\mathbb{Y}_{l,k} \times \mathbb{Y}_{l,k}$ (with entries in $[kl]$). It was observed earlier by Knuth [\cite{Knuth}, Exercise 5.1.4.9] (see also [\cite{Stanley}, Example 7.23.19(b)]) that the set $\mathbb{S}_{k,l}$ is in bijection with the set of pairs of standard Young tableaus $\mathbb{Y}_{l,k} \times \mathbb{Y}_{l,k}$ (with entries in $[kl]$) via the Robinson-Schensted correspondence. Romik~\cite{romik} (see also~\cite{Czabarka-W}) gave an explicit bijection via the function $\phi(\vec{a}) = (R_{\vec{a}},W_{\vec{a}})$. Theorem~\ref{thm:seqSat} shows that all $(k+1,l+1)$-saturated sequences are in fact extremal, i.e., have length $kl$. Hence there is also a bijection between the set of all $(k+1,l+1)$-saturated sequences and the set of pairs of standard rectangular Young tableaus (with entries in $[kl]$). \begin{proof}[Proof of Theorem~\ref{thm:seqSat}] Let $\vec{a} = [a_1, a_2,\dots, a_m]$ be a $(k+1,l+1)$-saturated sequence. By definition $\sat_{\mbox{\ensuremath{\mathcal S}}\xspace}(k+1,l+1) \leq m \leq \ram_{\mbox{\ensuremath{\mathcal S}}\xspace}(k+1,l+1)=kl$. To simplify our notation let $\gamma=\gamma_{\vec{a}}$, $R=R_{\vec{a}}$, $V=V_{\vec{a}}$ and $W=W_{\vec{a}}$. By the Erdős-Szekeres theorem for sequences, $\sat_{\mbox{\ensuremath{\mathcal S}}\xspace}(k+1,l+1) \leq k l$. Hence it suffices to show that $\sat_{\mbox{\ensuremath{\mathcal S}}\xspace}(k+1,l+1) \geq kl$. Let $\vec{a} = [a_1, a_2,\dots, a_m]$ be an arbitrary sequence of length $m < kl$ containing no increasing subsequence of length $k+1$ and no decreasing subsequence of length $l+1$. We need to show that $\vec{a}$ is not saturated. \begin{claim}\label{cl:completion} If a partially increasing matrix $M$ contains a $0$, then that $0$ can be replaced by a positive number so that the resulting matrix is still partially increasing. \end{claim} \begin{proof}[Proof of Claim~\ref{cl:completion}] Let $(i_0,j_0)$ be the position of a $0$ in $M$. If there is no nonzero entry in any position $(i,j)$ with $i_0 \le i$, $j_0 \le j$, then replace 0 with any number larger than all entries of the matrix. Otherwise let $t=\min\limits_{i\ge i_0,j\ge j_0}(M)_{ij}$. Change the value of $M$ at $(i_0,j_0)$ to be $t-\epsilon$ where $\epsilon$ is smaller than the difference of any two nonzero values of $M$ and it is also smaller than $t$. It is easy to see that since $M$ is partially increasing, the new matrix obtained is also partially increasing. \end{proof} By Claim~\ref{cl:completion}, if $r_{i,j}=0$ (and thus $v_{i,j}=w_{l+1-i,j}=0$), we can replace these $0$'s with some positive number (not necessarily integers) such that $R$ and $W$ are still partially increasing. We then relabel the elements of $R$ (if necessary) with an initial sequence of integers in $[k l]$ while respecting their order. Call the resulting matrices $R'$ and $W'$ respectively, we get $V'$ from $W'$ by reversing the order of its rows, that is the $i$-th row of $V'$ is the $(l+1-i)$-th row of $W'$. Continuing the example above we obtain that ${R'=\begin{pmatrix} 1 & 3 & 5\\ 2 & 4 & 6 \end{pmatrix}}$, $V'=\begin{pmatrix} 33 & 50 & 55\\ 11 & 22 & 44 \end{pmatrix}$ and $W'=\begin{pmatrix} 11 & 22 & 44\\ 33 & 50 & 55 \end{pmatrix}$. Now we can extend $\vec{a}$ as follows: Insert the number $(V')_{i,j}$ immediately after the {$((R')_{i,j}-1)${-th}} position of $\vec{a}$. Call the resulting sequence $\vec{b}$ (in our example $\vec{b}=[33,11,50,22,55,44]$). We see that $V'$ records the values in $\vec{b}$ and $R'$ records the order of these values. We want to show that $\vec{b}$ contains no increasing subsequence of length $k+1$ nor a decreasing subsequence of length $l+1$. Suppose $\vec{b}$ contains an increasing subsequence of length $k+1$. Then at least two elements of this subsequence must be in the same column of $V'$ by the pigeonhole principle. Since $V'$ is just $W'$ reversed, it is decreasing in the columns. Hence, the lower one of these two elements is smaller, and since they are in an increasing subsequence it must appear earlier in $\vec{b}$. On the other hand $R'$ is increasing in the columns so the lower must come later, a contradiction. Similarly, $\vec{b}$ does not have a decreasing subsequence of length $l+1$. Hence $\vec{a}$ is not saturated and the proof is complete. \end{proof} \section{Semisaturation of monotone point sets and sequences}\label{section:seq} In this section we use the point set formulation of the problem, the term point set always refers to a set of points in general position (i.e., no two points share a common $x$ or $y$ coordinate). We start with the following trivial lemma. \begin{lemma} \label{lem:intersect} If $I$ is an increasing subset and $D$ is a decreasing subset of the point set $P$, then they intersect in at most one element. \end{lemma} \begin{proof}[Proof of Theorem \ref{thm:weakSatSeq}] Fix some $n\in\mathbb{Z}^{+}$, and assume that $P$ is semisaturated with respect to monotone $n$-sequences. Then we know that any point not already contained in $P$ must combine with $n-1$ points from $P$ to form a monotone $n$-sequence. Note that any subsequence of a monotone sequence must itself be a monotone sequence; as such, our analysis will focus on monotone $(n-1)$-sequences in $P$, and consider which points in the plane can be added to a given $(n-1)$-sequence to produce a monotone $n$-sequence. We say an $(n-1)$-sequence {\em blocks} such points; thus, we can say that $P$ is saturated with respect to monotone $n$-sequences if and only if every point in the plane is either contained in $P$, or blocked by some monotone $(n-1)$-sequence from $P$. Consider some increasing $(k-1)$-sequence $\point[1],\dots,\point[k-1]$. The set of points blocked by the sequence is precisely the union $\cup_{i=1}^{k}\region[i]$ of regions given by \begin{minipage}{0.55\textwidth} \begin{alignat*}{5} &\region[1] &&=& (-\infty,x_1]&\times(-\infty,y_1];\\ &\region[i+1] &&=& [x_i,x_{i+1}]&\times[y_i,y_{i+1}],\text{ for $i=1,\dots,k-2$;}\\ &\region[k] &&=& [x_{k-1},\infty)&\times[y_{k-1},\infty). \end{alignat*} \end{minipage} \begin{minipage}{0.4\textwidth} \includegraphics{squenceproog.pdf} \end{minipage} \smallskip Decreasing sequences behave similarly. For our proof, we focus on points that are outside some fixed axis-parallel rectangle that contains all the points in $P$, and so are only interested in regions $\region[1]$ and $\region[k]$ above. More precisely, we pick bounding values $\lo{x}$, $\hi{x}$, $\lo{y}$, and $\hi{y}$ such that for each point $\coords\inP$ we have that $\lo{x}<x<\hi{x}$ and $\lo{y}<y<\hi{y}$. We will focus on how points lying on the lines forming the boundary of this region are blocked; note that these inequalities guarantee that such points can {\em only} be blocked by being at one end or the other of an increasing $(k-1)$-sequence or decreasing $(l-1)$-sequence. To begin, consider points along the line $y=\lo{y}$; fix any such point $(x,\lo{y})$. Now, since $\lo{y}$ is strictly less than the $y$-coordinate of any point in $P$, we may conclude that if any $(n-1)$-sequence of points $\point[1],\dots,\point[n-1]$ from $P$ block $(x,\lo{y})$, it must be the case either that the sequence is decreasing (and $n=l$) and $x_{n-1}\le x$ or that the sequence is increasing (and $n=k$) and $x\le x_{1}$. Viewed from the opposite perspective, we can see that any decreasing $(l-1)$-sequence $\point[1],\dots,\point[l-1]$ in $P$ blocks the left-bounded interval $[x_{l-1},\infty)$ on the line $y=\lo{y}$. Symmetrically, any increasing $(k-1)$-sequence blocks the right-bounded interval $(-\infty,x_1]$. For the entire line $y=\lo{y}$ to be blocked, then, we need a left-bounded interval and a right-bounded interval that intersect each other; this equates to a decreasing $(l-1)$-sequence and an increasing $(k-1)$-sequence such that the former lies entirely to the left of the latter. Let $\lo{D}$ and $\lo{I}$ be a decreasing $(l-1)$-sequence and increasing $(k-1)$-sequence from $P$, respectively, such that for all $\coords\in\lo{D}$ and all $\coordsP\in\lo{I}$, we have that $x \le x'$. The preceding argument guarantees the existence of such, and furthermore a symmetric argument with respect to the line $y=\hi{y}$ gives us that we can find an increasing $(k-1)$-sequence and a decreasing $(l-1)$-sequence $\hi{I}$ and $\hi{D}$, respectively, in $P$, such that for all $\coords\in\hi{I}$ and all $\coordsP\in\hi{D}$ we have that $x \le x'$. Our claim is that $\abs{\lo{D}\cup\lo{I}\cup\hi{D}\cup\hi{I}}\ge \min(2k+l-5,2l+k-5)$. We break our analysis into three cases. \begin{itemize} \item[Case:] $\lo{D}\cap\hi{D}=\emptyset$. Recall that we have assumed that no two points in $P$ share either a common $x$-value or a common $y$-value; thus, Lemma~\ref{lem:intersect} tells us that $\abs{\lo{I}\cap\lo{D}}\le1$ and $\abs{\lo{I}\cap\hi{D}}\le1$. Thus, we get that \begin{align*} \abs{\lo{D}\cup\lo{I}\cup\hi{D}\cup\hi{I}} &\ge \abs{\lo{I}\cup\lo{D}\cup\hi{D}}\\ &\ge \abs{\lo{I}}+\abs{\lo{D}}+\abs{\hi{D}}-\abs{\lo{I}\cap\lo{D}}-\abs{\lo{I}\cap\hi{D}}-\abs{\lo{D}\cap\hi{D}}\\ &\ge k-1 + 2(l-1) - 2\\ &= 2l+k -5, \end{align*} exactly as desired. \item[Case:] $\lo{I}\cap\hi{I}=\emptyset$. This case is symmetric to the preceding one. Applying Lemma~\ref{lem:intersect} appropriately gives us that \begin{equation*} \abs{\lo{D}\cup\lo{I}\cup\hi{D}\cup\hi{I}} \ge \abs{\lo{D}\cup\lo{I}\cup\hi{I}} \ge 2k + l -5, \end{equation*} once again. \item[Case:] $\abs{\lo{D}\cap\hi{D}},\abs{\lo{I}\cap\hi{I}}>0$. Let $\coords\in\lo{D}\cap\hi{D}$. Now, by our definitions of $\hi{I}$ and $\hi{D}$, we must have that for all $\coordsP\in\hi{I}$, $x' \le x$ holds. Similarly, our definitions of $\lo{I}$ and $\lo{D}$ ensure that for all $\coordsP\in\lo{I}$, we have $x \le x'$. Consider combining this with our assumption that no two points in $P$ share either a common $x$-value or a common $y$-value. Say $\hi{I}$ consists of the point sequence $\point[1],\dots,\point[k-1]$, and $\lo{I}$ consists of $\point[1]',\dots,\point[k-1]'$. Then we must have that \begin{equation*} x_1 < x_2 < \dots < x_{k-1} \le x \le x_1' < x_2' <\dots < x_{k-1}'. \end{equation*} By assumption, however, we have that $\lo{I}\cap\hi{I}\neq\emptyset$; this can only hold, then, if $p_{k-1}=p_{1}'$. Thus, we have that $\lo{I}\cup\hi{I}$ is, in fact, an increasing $(2k-3)$-sequence. A symmetric argument implies that, similarly, $\lo{D}\cup\hi{D}$ is a decreasing $(2l-3)$-sequence. So applying Lemma~\ref{lem:intersect} gives us that \begin{align*} \abs{\lo{D}\cup\lo{I}\cup\hi{D}\cup\hi{I}} & \ge \abs{\lo{D}\cup\hi{D}} + \abs{\lo{I}\cup\hi{I}} -\abs{(\lo{D}\cup\hi{D})\cap(\lo{I}\cup\hi{I})} \\ & \ge (2l-3)+(2k-3) - 1 \ge \min(2k+l-5,2l+k-5), \end{align*} \end{itemize} since $k$ and $l$ are at least $2$. In every case, the above gives us the desired lower bound, namely that $\abs{P}\ge \min(2k+l-5,2l+k-5)$. The upper bound follows from the following simple construction (see Figure~\ref{fig:seq}). \begin{figure}[!ht] \centering \includegraphics{seqconst.pdf} \caption{Construction for semisaturated sequences ($k=6,l=4$).} \label{fig:seq} \end{figure} \begin{construction} We present a construction of size $2k+l-5$, a construction of size $2l+k-5$ is attained by taking this construction for $k'=l$ and $l'=k$ of size $2k'+l'-5=k+2l-5$ and then reversing the order of the $x$-coordinates. Take an increasing $(k-2)$-sequence $p_1,p_2,\dots,p_{k-2}$ and let $p_{k-2}=(a,b)$. Take another increasing $(k-2)$-sequence $p_1',p_2',\dots,p_{k-2}'$, let $p'_1 = (a',b')$ and assume $a<a'$ and $b<b'$. Consider the rectangle defined by the vertices $(a,b),(a,b'),(a',b),(a',b')$. Let $0<\epsilon< \min\left((a'-a)/4,(b'-b)/4\right)$ and consider the rectangle with corners $(a+\epsilon,b+\epsilon),(a+\epsilon,b'-\epsilon),(a'-\epsilon,b+\epsilon),(a'-\epsilon,b'-\epsilon)$. Take a decreasing $(l-1)$-sequence $q_1,q_2,\dots,q_{l-1}$ with $q_1 = (a+\epsilon,b'-\epsilon)$ and $q_{l-1} = (a'-\epsilon,b+\epsilon)$. \qedhere \end{construction} \end{proof} \section{Convex point sets}\label{section:conv Given a point set $S \subseteq \mathbb{R}^d$, we use $\conv(S)$ to denote the {\em convex hull} of $S$, which is the smallest convex set in $\mathbb{R}^d$ that contains $S$, and we use $\interior(S)$ to denote the interior of $S$. First we prove the semisaturation result about cups and caps. \begin{proof}[Proof of Theorem \ref{thm:cupcapSemiSat}] For $k\le 2$ and $l\le 2$ the problem is trivial. Any two points form a 2-cup and also a 2-cap. Hence $\osat_{\mbox{\ensuremath{\mathcal C}}\xspace\C}(2,l)=\osat_{\mbox{\ensuremath{\mathcal C}}\xspace\C}(k,2)=1$. From now on let us assume that $k\ge 3$ and $l\ge 3$. Let $P$ be a point set that is semisaturated for $3$-cups and $l$-caps. Let $L$ be the set of lines determined by the points of $P$. There is an unbounded region $R$ in the plane bounded by parts of the lines that lies below every line of $L$. If we add a point $p$ in $R$ it must create a $3$-cup or an $l$-cap. We can choose $p$ inside $R$ to have smaller $x$ coordinate than any element of $P$, hence we can ensure that $p$ is not part of any $3$-cup. Therefore $p$ must be in a $l$-cap and $P$ must have at least $l-1$ elements. On the other hand a point set forming an $(l-1)$-cap is semisaturated for $k=3$. Indeed if we add a point, then it either creates a 3-cup or any three points form a 3-cap, which means that the whole set is a cap. The $l=3$ case can be handled similarly. Now we will assume that $k\ge 5$ and $l\ge 5$; the case when $k=4$ or $l=4$ will be settled later. We can define $L$ and $R$ as above. If we add $p$ anywhere in $R$ it must create a $k$-cup or an $l$-cap. Since $k\ge 4$ and $p$ lies below the lines of $L$, $p$ can only create an $l$-cap and furthermore $p$ is either the first element of this cap or the last one. We can choose $p$ inside $R$ to have a smaller $x$ coordinate than any element of $P$ (see Figure \ref{fig:cupcap}). This ensures that $p$ is the first element in the $l$-cap it has created. Now we move $p$ continuously inside $R$, all the while increasing its $x$-coordinate, until it has bigger $x$-coordinate than any element of $P$. At this point $p$ cannot be the first element of the $l$-cap it creates, thus during this movement there must be a last moment where $p$ is the first element of some $l$-cap containing it. Clearly the change happens as $p$ passes below an element $p_{below}$ of $P$. Let $x_{below}$ denote the $x$-coordinate of this point. \begin{figure}[!ht] \centering \includegraphics[width=10cm]{cupcap1.pdf} \caption{If $p$ is to the right of $x_{below}$, then it is not the first element of any $4$-cap.} \label{fig:cupcap} \end{figure} If we put $p$ in $R$ slightly after $x_{below}$, then it must extend some $(l-1)$-cap $A_1$ whose points lie to the left of $x_{below}$, except maybe for $p_{below}$. Similarly if we put it slightly before $x_{below}$, then it must extend some $(l-1)$-cap $A_2$ that lies to the right of $x_{below}$. In the same way we can define $p_{above}$, $x_{above}$ and two $(k-1)$-cups $U_1$ and $U_2$ such that they lie on the left and right side of $x_{above}$. In total we have found $|A_1\cup A_2 \cup U_1 \cup U_2|$ elements. Clearly $|A_1\cap A_2|\le 1$ and $|U_1\cap U_2|\le 1$. Since any cup intersects any cap in at most two points we have $|A_i\cap U_j|\le 2$. Also either $x_{below}< x_{above}$ or $x_{below}> x_{above}$ or $x_{below}= x_{above}$. In the first case $A_1\cap U_2=\emptyset$ and in the second case $A_2\cap U_1=\emptyset$, giving us \[|A_1\cup A_2 \cup U_1 \cup U_2|\ge 2k-2+2l-2-|A_1\cap A_2|-|U_1\cap U_2|-3\cdot 2\ge 2k+2l-12.\] If $x_{below}= x_{above}$ we have $|A_1\cap U_2|\le 1$ and $|A_2\cap U_1|\le 1$ so we have \[|A_1\cup A_2 \cup U_1 \cup U_2|\ge 2k-2+2l-2-|A_1\cap A_2|-|U_1\cap U_2|-2\cdot 2-1-1\ge 2k+2l-12.\] For the upper bound we give a construction. First consider the point set shown in Figure~\ref{fig:cupcapconst}. \begin{figure}[!ht] \centering \includegraphics[width=7cm]{cupcap4.pdf} \includegraphics[width=7cm]{cupcap5.pdf} \caption{ Cup-cap semisaturation for $k=l=5$ and for $k=8,l=7$} \label{fig:cupcapconst} \end{figure} This point set consists of 10 points, and it is saturated for 5-cups and 5-caps. We show this by dividing the plane into regions and for each region showing four points of the point set that form either a 5-cup or a 5-cap with any point of the region. Consider the regions in Figure~\ref{fig:cupcapregions}. \begin{figure}[!ht] \centering \includegraphics[width=0.3\textwidth]{cupcap41.pdf} \includegraphics[width=0.3\textwidth]{cupcap42.pdf} \includegraphics[width=0.3\textwidth]{cupcap43.pdf} \includegraphics[width=0.3\textwidth]{cupcap44.pdf} \includegraphics[width=0.3\textwidth]{cupcap45.pdf} \includegraphics[width=0.3\textwidth]{cupcap46.pdf} \includegraphics[width=0.3\textwidth]{cupcap47.pdf} \includegraphics[width=0.3\textwidth]{cupcap48.pdf} \caption{ The regions blocked by the 4-cups and 4-caps.} \label{fig:cupcapregions} \end{figure} In the first seven subfigures we have drawn a region and indicated which four points of the point set blocks that region. In the eight subfigure we have drawn all the regions. As we can see these regions cover half of the plane. Since the point set is centrally symmetric, this is enough. Hence for $k=l=5$ we have semisaturation with $2k+2l-10$ points. Now we will show that this construction can be extended for $k,l\ge 5$. In Figure \ref{fig:cupcapconst} we can see $A_1$, $A_2$, $U_1$ and $U_2$. To get a construction for $k,l$ we just extend $A_1$ to the left and $A_2$ to the right with $l-5$ elements and $U_1$ to the left and $U_2$ to the right with $k-5$ elements. See Figure \ref{fig:cupcapconst} for an example. The resulting configuration will be semisaturated. Considering the same regions as in Figure \ref{fig:cupcapregions} will work. If a region was blocked by a $4$-cup, that $4$-cup is now extended by $k-5$ elements, so we have a blocking $(k-1)$-cup. Similarly if a region was blocked by a $4$-cap, that $4$-cap is now extended by $k-5$ elements, so we have a blocking $(k-1)$-cap. Therefore we have found a semisaturated set with $10+2(k-5)+2(l-5)=2k+2l-10$ points. In the case of $l=4$ and $k\ge 5$ the construction is quite similar. A possible configuration for the $k=5$ case is given in Figure \ref{fig:4case} and the blocked regions are given in Figure \ref{fig:4caseall}. For $k>5$ we can extend $U_1$ and $U_2$ just as we did in Figure \ref{fig:cupcapconst}. We leave the details for the interested readers. \begin{figure}[!ht] \centering \includegraphics[width=0.3\textwidth]{4cupalap.pdf} \includegraphics[width=0.39\textwidth]{4cupalap2.pdf} \caption{Cup-cap semisaturation for $l=4, k = 5$ and for $l=4, k=7$.} \label{fig:4case} \end{figure} \begin{figure}[!ht] \centering \includegraphics[width=0.15\textwidth]{4cuplist1.pdf} \includegraphics[width=0.15\textwidth]{4cuplist2.pdf} \includegraphics[width=0.15\textwidth]{4cuplist3.pdf} \includegraphics[width=0.15\textwidth]{4cuplist4.pdf} \includegraphics[width=0.15\textwidth]{4cuplist5.pdf} \includegraphics[width=0.15\textwidth]{4cuplist6.pdf} \includegraphics[width=0.15\textwidth]{4cupall.pdf} \caption{The regions blocked by the 4-cups and 3-caps.} \label{fig:4caseall} \end{figure} Next we show that at least $2k-2$ points are required to obtain a saturated construction. Indeed, by the same reasoning as in the $l,k \ge 5$ case we can find two cups $U_1,U_2$ of size $k-1$ intersecting in at most one point and two caps $A_1,A_2$ of size $3$ intersecting in at most one point. If $U_1$ and $U_2$ are disjoint, then we have already found $2k-2$ many points, so suppose they intersect in one point $q$. We know that either $A_1$ lies to the left of $q$ or $A_2$ lies to the right of $q$ (it is possible that $A_1$ or $A_2$ contains $q$). In either case we must have one more point since no three points of $U_1$ nor of $U_2$ form a cap. \end{proof} Now we continue with the semisaturation of convex point sets. \begin{proof}[Proof of Theorem~\ref{thm:osat00}] Suppose that to the contrary there is a semisaturated set of points $S$ with $n-1 + \floor{\frac{n-2}{d}}-s$ points in $\mbox{\ensuremath{\mathbb R}}\xspace^d$, for some $s\ge 1$. Denote by $S_1,S_2,\dots,S_m$ the subsets of $S$ that are convex $(n-1)$-sets. If $\cap_{i=1}^m \interior(\conv(S_i)) \neq \varnothing$, then we may add any point in the intersection without yielding $n$ points in convex position. We can also add this point such that the resulting point set is in general position. The interior of a convex set is convex, hence by Helly's theorem it is sufficient to show that the intersection of any $(d+1)$ of the sets from $\{\interior(\conv(S_i))\}_{i\in [m]}$ is nonempty. Consider $d+1$ sets in $\{S_i\}_{i\in [m]}$: $S_{i_1},S_{i_2},\dots,S_{i_{d+1}}$. They each have size $n-1$ and are contained in the point set $S$ of size $n-1 + \floor{\frac{n-2}{d}}-s$, thus we have that \begin{align*} \abs{(S_{i_1}\cap S_{i_2}\cap \dots \cap S_{i_{d+1}})^c} &= \abs{S_{i_1}^c\cup S_{i_2}^c\cup \dots \cup S_{i_{d+1}}^c} \\& \leq \abs{S_{i_1}^c}+ \abs{S_{i_2}^c} + \dots + \abs{S_{i_{d+1}}^c} =(d+1)\left(\floor{\frac{n-2}{d}}-s\right). \end{align*} Therefore \begin{align*} \abs{S_{i_1}\cap S_{i_2}\cap \dots \cap S_{i_{d+1}}}& \ge \left(n-1+\floor{\frac{n-2}{d}}-s\right)-(d+1)\left(\floor{\frac{n-2}{d}}-s\right) \\&= n-1-d\left(\floor{\frac{n-2}{d}}-s\right)\ge1+ds\ge d+1. \end{align*} Since the original point set is in general position, these $d+1$ points in the intersection span a non-degenerate simplex. It follows that the interiors of $\conv(S_{i_1}),\dots, \conv(S_{i_{d+1}})$ intersect, as required. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:convex}] We will construct a set of points $S$ such that $S$ is semisaturated in the plane $\mathbb{R}^2$. Consider a convex polygon $Q =\conv(v_0,v_1,\ldots, v_{2n-3})$ with $2n-4$ sides such that the side $v_i v_{i+1}$ is parallel to the side $v_{n+i-2} v_{n+i-1}$ (where the indices are modulo $2n-4$). For ease of reference, we call the pair of sides $v_i v_{i+1}$, $v_{n+i-2} v_{n+i-1}$ \textit{opposite sides} of $Q$. Let $S=\{v_0,v_1,\dots,v_{2n-3}\}$ be the vertex set of $Q$. We claim that $S$ is semisaturated in $\mathbb{R}^2$. \begin{claim}\label{cl:inside} Let $P$ be any point contained in $Q$ such that $Q\cup\{P\}$ is in general position. Then $S \cup \{P\}$ has a convex $n$-gon with $P$ as one of its vertices. \begin{proof} Let $P$ be an arbitrary point in the interior of $Q$. Consider the chord $v_0v_{n-2}$ (see Figure~\ref{fig:convexfig}). This divides $Q$ into two $(n-1)$-gons $\{ v_{0}, v_{1},\ldots, v_{n-2}, \}$ and $\{ v_{n-2}, v_{n-1},\ldots, v_{0}, \}$. Since the points are in general position, $P$ must lie in the interior of one of these $(n-1)$-gons. Then $P$ and the points of the other $(n-1)$-gon form a convex $n$-set. \end{proof} \end{claim} \begin{claim}\label{cl:outside} Let $P$ be any point not contained in $Q$ such that $Q\cup\{P\}$ is in general position. Then $S \cup \{P\}$ has a convex $n$-gon with $P$ as one of its vertices. \end{claim} We say that a point $A\in Q$ can be \emph{seen} from $P$ if $\overline{PA}\cap Q={A}$. A side $v_i,v_{i+1}$ can be seen from $P$ if all of its points can be seen from $P$. Clearly $P$ can see at most one from a pair of opposite sides. So there are at least $n-2$ sides that cannot be seen from $P$. Each of these sides will be a side of the convex hull $\conv(Q\cup\{P\})$. There will be two additional sides of this convex~hull incident to $P$, so $\conv(Q\cup\{P\})$ has at least $n$ sides. Therefore we can choose $n-1$ points from $S$ such that they form a convex $n$-gon with $P$. \begin{figure}[!ht] \centering \include{convex} \caption{Finding a convex $n$-gon in the extended point set.} \label{fig:convexfig} \end{figure} Claim~\ref{cl:inside} and Claim~\ref{cl:outside} imply that $S$ is semisaturated, hence $\osat_{\mbox{\ensuremath{\mathcal C}}\xspace}(n) \leq 2n-4$. Now we prove the lower bound. Suppose $S$ is a semisaturated set of points in the plane. The set $S$ determines $\binom{|S|}{2}$ lines, which partition the plane into regions. Let $P_1$ be a point in one of the infinite regions (and not between any pair of parallel lines), and let $P_2$ be another point in the opposite infinite region. That is $P_1$ and $P_2$ lie on different sides for each of the $\binom{|S|}{2}$ lines. Since $S$ is semisaturated, there are two sets $S_1,S_2\subset S$ such that $S_1\cup \{P_1\}$ and $S_2\cup \{P_2\}$ are convex $n$-gons. We claim that $|S_1\cap S_2|\le 2$. Suppose $v_1,v_2,v_3 \in S_1\cap S_2$, and assume that $v_1,v_2,v_3,P_1$ is a convex quadrilateral (in that order). Observe that $P_1$ is in the same side as $v_3$ with respect to the line $v_1v_2$ and on the same side as $v_1$ with respect the line $v_2v_3$. On the other hand, $P_2$ is in the opposite side of $v_3$ with respect the line $v_1v_2$ and on the opposite of $v_1$ with respect the line $v_2v_3$. The points $v_1,v_2,v_3,P_2$ cannot form a convex quadrilateral (in any order). Indeed, one of the sides of this quadrilateral must be either $v_1v_2$ or $v_2v_3$, but the line defined by $v_1v_2$ would separate $P_2$ from $v_3$, and the line defined by $v_2v_3$ would separate $P_2$ from $v_1$. This cannot happen since $S_2 \cup \{P_2\}$ is in convex position, a contradiction. \begin{figure}[!ht] \centering \includegraphics{convim.pdf} \caption{Convex $n$-gons containing $P_1$ and $P_2$ intersect in at most two points.} \label{fig:conv2} \end{figure} Hence $|S_1\cup S_2|=|S_1|+|S_2|-|S_1\cap S_2|\ge n-1+n-1-2=2n-4.$ \end{proof} \section{General treatment of saturation questions}\label{general} In this section we provide a general formulation for many of the the problems we have considered in this paper. Given a $c$-edge-colored complete $s$-uniform hypergraph $H=(V,E)$, we say that a subset of vertices $S \subseteq V$ forms a monochromatic complete subhypergraph of $H$ in color $i$ if the ($s$-uniform) subhypergraph induced by $S$ has only hyperedges in color $i$. Many of the problems we considered have the following form. \begin{defi}\label{gengen} Given constants $c$ and $s$, let $\mbox{\ensuremath{\mathcal F}}\xspace_0$ be the family of complete $s$-uniform hypergraphs whose edges are colored with $c$ colors (numbered by $1,2,\dots, c$). For a subfamily $\mbox{\ensuremath{\mathcal F}}\xspace$ of $\mbox{\ensuremath{\mathcal F}}\xspace_0$, a member $F$ of $\mbox{\ensuremath{\mathcal F}}\xspace$ is \emph{saturated} if for every $i$, $F$ does not contain a monochromatic complete subhypergraph of size $k_i$ and color $i$, but every $F' \in \mbox{\ensuremath{\mathcal F}}\xspace$ that extends $F$ contains a monochromatic complete subhypergraph of size $k_i$ of color $i$ for some $i$. $F$ is \emph{semisaturated} if we omit the first condition, that is, if $F \in\mbox{\ensuremath{\mathcal F}}\xspace$ and every $F'\in \mbox{\ensuremath{\mathcal F}}\xspace$ that extends $F$ contains a monochromatic complete subhypergraph of size $k_i$ of color $i$ for some $i$ which is not in $F$. Let $\ram_{\mbox{\ensuremath{\mathcal F}}\xspace}(k_1,\dots k_c)$ denote the size (number of vertices) of the largest saturated $F\in \mbox{\ensuremath{\mathcal F}}\xspace$, and let $\sat_{\mbox{\ensuremath{\mathcal F}}\xspace}(k_1,\dots k_c)$ denote the size of the smallest saturated $F\in \mbox{\ensuremath{\mathcal F}}\xspace$. Finally, let $\osat_{\mbox{\ensuremath{\mathcal F}}\xspace}(k_1,\dots k_c)$ denote the size of the smallest semisaturated $F\in \mbox{\ensuremath{\mathcal F}}\xspace$. \end{defi} \begin{obs} For any $\mbox{\ensuremath{\mathcal F}}\xspace$ and positive integers $k_1, \ldots, k_c$, \[\osat_{\mbox{\ensuremath{\mathcal F}}\xspace}(k_1, \ldots, k_c)\le \sat_{\mbox{\ensuremath{\mathcal F}}\xspace}(k_1, \ldots, k_c) \le \ram_{\mbox{\ensuremath{\mathcal F}}\xspace}(,k_1, \ldots, k_c).\] \end{obs} Note that whenever $\sat=\ram$ holds, all saturated members of $\mbox{\ensuremath{\mathcal F}}\xspace$ have the same size. Thus we gain further insight into the respective Ramsey-type problem as well. Moreover, when $c=2$, one can regard the problem as the first color class forming an (uncolored) hypergraph $H$. Then it follows that a complete subhypergraph in the first color is a complete subhypergraph in $H$ while a complete subhypergraph in the second color is an independent set in $H$. Definition~\ref{gengen} is quite general. In this paper we have introduced saturation problems for graphs, posets, monotone point sets, and cups and caps. All of these fit into this formulation. First, the graph case we get by setting $s=2$ and $\mbox{\ensuremath{\mathcal F}}\xspace=\mbox{\ensuremath{\mathcal F}}\xspace_0$. We get the poset case by setting $c=s=2$ and letting $\mbox{\ensuremath{\mathcal F}}\xspace$ be the family of those $2$-edge-colored graphs that we can obtain as the comparability graph of a poset. We obtain the monotone point set case by setting $c=s=2$ and letting $\mbox{\ensuremath{\mathcal F}}\xspace$ be the family of those $2$-edge-colored graphs that we can obtain from the pairs of elements in a sequence by coloring the increasing pairs red and the decreasing pairs blue. Finally, the cup and cap case we get by setting $c=2$, $s=3$ and letting $\mbox{\ensuremath{\mathcal F}}\xspace$ be the family of those $2$-colored complete $3$-uniform hypergraphs that we can obtain by taking a point set in general position and coloring a triple red if it forms a cup and blue if it forms a cap (note that every triple forms a cup or a cap). The only problem we considered that does not fit into this formulation is the case of convex subsets of points. It is interesting that for both the $2$-colored graph case and the poset case we have $\sat(k,l)=(k-1)(l-1)$, yet we could not find any general reasoning that handles both of these cases at once. It is interesting that the relative behavior of $\sat$, $\osat$ and $\ram$ can vary substantially depending on the setting. Indeed, for graphs $\osat=\sat$ yet $\ram$ is exponential, while for posets and monotone point sets, $\sat$ equals $\ram$ yet $\osat$ is smaller ($\osat$ behaves differently in the latter two cases). \section{Acknowledgements} We thank Tuan Tran for pointing out an inaccuracy in Lemma~\ref{decomp} in an earlier version of this manuscript.

https://arxiv.org/abs/2010.14043

The Teaching Dimension of Kernel Perceptron

Algorithmic machine teaching has been studied under the linear setting where exact teaching is possible. However, little is known for teaching nonlinear learners. Here, we establish the sample complexity of teaching, aka teaching dimension, for kernelized perceptrons for different families of feature maps. As a warm-up, we show that the teaching complexity is $\Theta(d)$ for the exact teaching of linear perceptrons in $\mathbb{R}^d$, and $\Theta(d^k)$ for kernel perceptron with a polynomial kernel of order $k$. Furthermore, under certain smooth assumptions on the data distribution, we establish a rigorous bound on the complexity for approximately teaching a Gaussian kernel perceptron. We provide numerical examples of the optimal (approximate) teaching set under several canonical settings for linear, polynomial and Gaussian kernel perceptrons.

\section{Acknowledgements} Yuxin Chen is supported by NSF 2040989 and a C3.ai DTI Research Award 049755. \section{Experimental Evaluation}\label{appendix: experimentals} In this section, we provide an algorithmic procedure for constructing the $\epsilon$-approximate teaching set, and quantitatively evaluate our theoretical results as presented in \thmref{thm: boundedclassifier} and \thmref{thm: gaussian_main_thm} (cf. \secref{subsec: bounded_error}). Results in this section are supplementary to \figref{fig:example-RBF}. For a qualitative evaluation of the \emph{$\epsilon$-approximated teaching set}, please refer to \figref{fig:Learned_RBF_45c}, which illustrates the learner's Gaussian kernel perceptron learned from the {$\epsilon$-approximated teaching sets} on different classification tasks. \subsection{Experimental Setup} Our experiments are carried out on 4 different datasets: the two-moon dataset (2 interleaving half-circles with noise), the two-circles dataset (a large circle containing a small circle, with noise) from sklearn\footnote{https://scikit-learn.org/stable/modules/classes.html\#module-sklearn.datasets}, the Banana dataset\footnote{https://www.scilab.org/tutorials/machine-learning-–-classification-svm} where the two classes are not perfectly separable, and the Iris dataset\footnote{https://archive.ics.uci.edu/ml/datasets/iris} with one of the three classes removed. For each dataset, the following steps are performed: \begin{enumerate} \item\label{enum: e1} For a given set of data, we, assuming the role of the teacher, find the optimal Gaussian (with $\sigma = 0.9$) separator $\boldsymbol{\theta}^*$ and plot the corresponding boundaries. We estimate the perceptron loss $\textbf{err}(f^*)$ for this separator by summing up the total perceptron loss on the dataset and averaging over the size of the dataset. \item\label{enum: e2} For some $s$, we use the degree $s$ polynomial approximation of the Gaussian separator to determine the approximate polynomial boundaries and select $r = \binom{2+s}{s} - 1$ points on the boundaries such that their images in the polynomial feature space are linearly independent. We make a copy of these points and assign positive labels to one copy and negative labels to the other. In addition, we pick 2 more points arbitrarily, one on each side of the boundaries (i.e. with opposite labels). Thus $\mathcal{TS_{\theta^*}}$ of size $2r$ is constructed. \item\label{enum: e3} Following \assref{assumption: orthogonal}-\ref{assumption: bounded cone}, the Gaussian kernel perceptron learner (with the same $\sigma$ parameter value 0.9) uses only $\mathcal{TS_{\theta^*}}$ to learn a separator $\hat{\boldsymbol{\theta}}$. The perceptron loss $\textbf{err}(\hat{f})$ w.r.t. the original dataset is calculated by averaging the total perceptron loss over the number of points in the dataset. \item\label{enum: e4} Repeat Step 2 and Step 3 for $s$ = $2, 3,\cdots, 12$ and record the perceptron loss (i.e the $\max$ error function as shown in \secref{sec.statement}) for the corresponding teaching set sizes $2r$ (where $r$ = $\binom{2+s} {s} - 1$). Then we plot the error $\left|\textbf{err}(f^*) - \textbf{err}(\hat{f})\right|$ as a function of the teaching set size. \end{enumerate} The corresponding plots for Steps 1-4 are shown in columns (a)-(d) of \figref{fig:panel-RBF}, where for Step 2 (column (b)), the plots all correspond to when $s=5$. \subsection{Implementation Details} In this subsection we provide more algorithmic and numeric details about the implementation of the experiments. First we describe how the first $r-1$ points are generated in \stepref{enum: e2}. Given that the approximate polynomial separator has been found using the kernel and feature map approximation described in \eqnref{eqn: eqn17} and \eqnref{eqn:eqn18}, we are able to plot the corresponding boundaries, and by the same reasoning as in the case of teaching set generation for the polynomial learner, we need to locate points on the boundaries such that their images in the r-dimensional feature space are linearly independent. We achieve this by sampling points on the zero-contour line and row-reducing the matrix formed by the image of all such points. This way, $r-1$ qualified points can be efficiently located. In addition, as discussed in \secref{subsec: bounded_error}, the teaching points are selected within the radius of some small constant multiple of $\sqrt{R}$ consistently across the experiments. In this case, we have arbitrarily picked the constant to be 4. In \stepref{enum: e3}, when the learner learns the separator, we need to ensure \assref{assumption: orthogonal}-\ref{assumption: bounded cone} are satisfied. This is made possible by adding the corresponding constraints to the learner's optimization procedure. Specifically, we need to enforce that 1) the norm of $\hat{\boldsymbol{\theta}}$ is not far from 1, and 2) $\beta$\footnote{We pick two points outside the orthogonal complement of $\mathbb{P}\boldsymbol{\theta}^*$, one with positive label and another with negative label. Thus, in place of $\beta_0$ (as used in \secref{subsec.gaussiankernel}) we use $\beta \in \ensuremath{\mathbb{R}}^2$ here.} and $\gamma$ are bounded absolutely as mentioned in \assref{assumption: bounded cone}. This is achieved by adjusting the specified bound higher or lower as the current-iteration $\hat{\boldsymbol{\theta}}$ norm varies during the optimization procedure. Eventually, we normalize $\hat{\boldsymbol{\theta}}$ and check that the final $\beta$ and $\gamma$ are indeed bounded (i.e. \assref{assumption: bounded cone} is satisfied). Finally, the perceptron loss calculated for each value of $s$ is based on 5 separate runs of \stepref{enum: e2}, while for each run, the learner's kernelized Gaussian perceptron learning algorithm is repeated 5 times. The learner's perceptron loss is then averaged over the 25 epochs of the algorithm to prevent numerical inaccuracies that may arise during the learner's constrained optimization process and possibly the teaching set generation process. \subsection{Results} We present the experiment results in \figref{fig:panel-RBF}. In the right-most plot of each row, the estimates of $\left|\textbf{err}(f^*) - \textbf{err}(\hat{f})\right|$ are plotted against the teaching set sizes $2r$ corresponding to $s=2,\cdots,12$ (as discussed in \secref{subsec: gaussian_kernel_approx}). As can be observed from the shape of the curves in plots of column (d), indeed, our experimental results confirm that the number of teaching examples needed for $\epsilon$-approximate teaching is upper-bounded by $d^{\bigO{\log^2 \frac{1}{\epsilon}}}$ for Gaussian kernel perceptrons. \begin{figure*}[t] \centering \begin{subfigure}[b]{0.24\textwidth} \centering \includegraphics[width=\linewidth]{fig/supp_fig1-1b.png} \label{fig:Teacher_RBf_41a} \end{subfigure} \begin{subfigure}[b]{0.24\textwidth} \centering \includegraphics[width=\linewidth]{fig/supp_fig1-2b.png} \label{fig:Polynomial_approx_41b} \end{subfigure} \begin{subfigure}[b]{0.24\textwidth} \centering \includegraphics[width=\linewidth]{fig/supp_fig1-3b.png} \label{fig:Learned_RBF_41c} \end{subfigure} \begin{subfigure}[b]{0.24\textwidth} \centering \includegraphics[width=\linewidth]{fig/plot1d.png} \label{fig:Epsilon_size_41d} \end{subfigure} \\ \begin{subfigure}[b]{0.24\textwidth} \centering \includegraphics[width=\linewidth]{fig/supp_fig2-1b.png} \label{fig:Teacher_RBf_42a} \end{subfigure} \begin{subfigure}[b]{0.24\textwidth} \centering \includegraphics[width=\linewidth]{fig/supp_fig2-2b.png} \label{fig:Polynomial_approx_42b} \end{subfigure} \begin{subfigure}[b]{0.24\textwidth} \centering \includegraphics[width=\linewidth]{fig/supp_fig2-3b.png} \label{fig:Learned_RBF_42c} \end{subfigure} \begin{subfigure}[b]{0.24\textwidth} \centering \includegraphics[width=\linewidth]{fig/plot2d.png} \label{fig:Epsilon_size_42d} \end{subfigure} \\ \begin{subfigure}[b]{0.24\textwidth} \centering \includegraphics[width=\linewidth]{fig/supp_fig3-1.png} \label{fig:Teacher_RBf_43a} \end{subfigure} \begin{subfigure}[b]{0.24\textwidth} \centering \includegraphics[width=\linewidth]{fig/supp_fig3-2.png} \label{fig:Polynomial_approx_43b} \end{subfigure} \begin{subfigure}[b]{0.24\textwidth} \centering \includegraphics[width=\linewidth]{fig/supp_fig3-3.png} \label{fig:Learned_RBF_43c} \end{subfigure} \begin{subfigure}[b]{0.24\textwidth} \centering \includegraphics[width=\linewidth]{fig/plot3d.png} \label{fig:Epsilon_size_43d} \end{subfigure} \\ \begin{subfigure}[b]{0.24\textwidth} \centering \includegraphics[width=\linewidth]{fig/supp_fig5-1.png} \caption{} \label{fig:Teacher_RBf_45a} \end{subfigure} \begin{subfigure}[b]{0.24\textwidth} \centering \includegraphics[width=\linewidth]{fig/supp_fig5-2.png} \caption{} \label{fig:Polynomial_approx_45b} \end{subfigure} \begin{subfigure}[b]{0.24\textwidth} \centering \includegraphics[width=\linewidth]{fig/supp_fig5-3.png} \caption{} \label{fig:Learned_RBF_45c} \end{subfigure} \begin{subfigure}[b]{0.24\textwidth} \centering \includegraphics[width=\linewidth]{fig/plot4d.png} \caption{} \label{fig:Epsilon_size_45d} \end{subfigure} \caption{Constructing the $\epsilon$-approximate teaching set $\mathcal{TS}_{\boldsymbol{\theta}^*}$ for a Gaussian kernel perceptron learner. Each row corresponds to the numerical results on a different dataset as described in the beginning of \appref{appendix: experimentals}. For each row from left to right: (a) optimal Gaussian boundary and the data set; (b) Teacher identifies a (degree-5) polynomial approximation of the Gaussian decision boundary and finds the $\epsilon$-approximate teaching set $\mathcal{TS}_{\boldsymbol{\theta}^*}$ (marked by cyan plus markers and red dots); (c) Learner learns a Gaussian kernel perceptron from the optimal teaching set in the previous plot; (d) $\abs{\mathcal{TS}}\text{-}\epsilon$ plot for degree-2 to degree-12 polynomial approximation teaching results. The blue curve corresponds to $d^{\bigO{\log^2 \frac{1}{\epsilon}}}$ = $2^{\bigO{\log^2 \frac{1}{\epsilon}}}$ where $d=2$.} \label{fig:panel-RBF} \end{figure*} \section{Gaussian Kernel Perceptron}\label{appendix: gaussian perceptron} In this appendix, we would provide the proofs to the key results: \thmref{thm: boundedclassifier} and \thmref{thm: gaussian_main_thm}, as shown in \secref{subsec: bounded_error}. The key to establishing the results is to provide a constructive procedure for an approximate teaching set. Under the \assref{assumption: orthogonal} and \assref{assumption: bounded cone}, when the Gaussian learner optimizes \eqnref{eqn: bounded} w.r.t the teaching set, any solution $\hat{\boldsymbol{\theta}} \in {\boldsymbol{\mathcal{A}}}_{opt}$ would be $\epsilon$-close to the optimal classifier point-wise, thereby bounding the error on the data distribution ${\mathcal{P}}$ on the input space ${\mathcal{X}}$. We organize this appendix as follows: in \appref{appendixsub: solutionexists} we show that there exists a solution to \eqnref{eqn: bounded}; in \appref{appendixsub: proofofmainthm} we provide the proofs for our key results \thmref{thm: boundedclassifier} and \thmref{thm: gaussian_main_thm}. \paragraph{Truncating the Taylor features of Gaussian kernel.} In \secref{subsec: gaussian_kernel_approx}, we showed the construction of the projection $\mathbb{P}$ such that $\mathbb{P}\Phi$ forms a feature map for the kernel $\Tilde{{\mathcal{K}}}$. We denote the orthogonal projection to $\mathbb{P}$ by $\mathbb{P}^{\bot}$. Thus, we can write $\Phi({\mathbf{x}}) = \mathbb{P}\Phi({\mathbf{x}}) + \mathbb{P}^{\bot}\Phi({\mathbf{x}})$ for any ${\mathbf{x}} \in \mathbb{R}^d$. We discussed the choice of $R$ and $s$. The primary motivation to pick them in the certain way is to retain maximum information in the first $\binom{d+s}{s}$ coordinates of $\Phi(\cdot)$. This is in line with the observation that the eigenvalues of the canonical orthogonal basis~\cite{article} (also eigenvectors) for the Gaussian reproducing kernel Hilbert space ${\mathcal{H}}_{{\mathcal{K}}}$ decays with higher-indexed coordinates, thus the more sensitive eigenvalues are in the first $\binom{d+s}{s}$ coordinates. Thus, if we could show that $\mathbb{P}\hat{\boldsymbol{\theta}}$ is $\epsilon$-approximately close to $\mathbb{P}\boldsymbol{\theta}^*$ where $\hat{\boldsymbol{\theta}} \in {\boldsymbol{\mathcal{A}}}_{opt}$ is a solution to \eqnref{eqn: bounded}, then $\hat{\boldsymbol{\theta}}$ also would be $\epsilon$-approximately close to $\boldsymbol{\theta}^*$. \paragraph{What should be an optimal $R$ vs. choice of the index $s$?} In \secref{subsec: gaussian_kernel_approx}, we solved for $s$ such that $$\frac{1}{(s+1)!}\cdot \paren{R}^{s+1} \le \epsilon$$ If ${\mathbf{x}} \in \mathcal{B}(\sqrt{2\sqrt{R}\sigma^2}, 0)$ then using \lemref{lemma: approxbound} we have \begin{equation} \inmod{\mathbb{P}^{\perp}\Phi({\mathbf{x}})}^2 \le \frac{1}{(s+1)!}\cdot \paren{\sqrt{R}}^{s+1} \le \frac{\epsilon}{\paren{\sqrt{R}}^{s+1}} \le \frac{\epsilon}{\paren{\sqrt{d}}^{s}} \label{eqn: largeeps} \end{equation} where the last inequality follows as $R := \max\left\{\frac{\log^2 \frac{1}{\epsilon}}{e^2},d\right\}$. We define $\epsilon_s := \frac{\epsilon}{\paren{\sqrt{d}}^{s}}$. Note that \eqnref{eqn: largeeps} holds for all ${\mathbf{x}} \in {\mathcal{X}}$\, since $\frac{\norm{{\mathbf{x}}}{{\mathbf{x}}}}{\sigma^2} \le 2\sqrt{R}$. This factor $\paren{\sqrt{d}}^{s}$ in the denominator of $\epsilon_s$ would be useful in nullifying any $\sqrt{r}$ term since $r = \binom{d+s}{s} = \bigO{d^{s}}$. \subsection{Construction of a Solution to \eqnref{eqn: bounded}}\label{appendixsub: solutionexists} In this subsection, we would show that $\eqnref{eqn: bounded}$ has a minimizer $\hat{\boldsymbol{\theta}} \in {\boldsymbol{\mathcal{A}}}_{opt}$ such that $\mathbf{p}(\hat{\boldsymbol{\theta}}) = 0$ where $\mathbf{p}(\cdot)$ is the objective value. Notice that for any $i$ the teaching points $\curlybracket{({\mathbf{z}}_i, 1), ({\mathbf{z}}_i, -1)}$ are correctly classified only if $\hat{\boldsymbol{\theta}}\cdot\Phi({\mathbf{z}}_i) = 0$ and $\hat{\boldsymbol{\theta}}\cdot\Phi({\mathbf{a}}) > 0$. We define the set $\boldsymbol{\mathrm{B}} = \curlybracket{{\mathbf{b}}_1,{\mathbf{b}}_2,\cdots,{\mathbf{b}}_{r}}$ to represent $\{{\mathbf{z}}_i\}_{i=1}^{r-1}\cup \{{\mathbf{a}}\}$ in that order. We define the Gaussian kernel Gram matrix $\textLambda$ corresponding to $\boldsymbol{\mathrm{B}}$ as follows: \begin{equation} \textLambda[i,j] = {\mathcal{K}}({\mathbf{b}}_i,{\mathbf{b}}_j)\quad \forall i,j \in \bracket{r} \end{equation} Since $\{{\mathbf{z}}_i\}_{i=1}^{r-1}$ and ${\mathbf{a}}$ could be chosen from $\mathcal{B}(\sqrt{2\sqrt{R}\sigma^2},0)$ as for any two points ${\mathbf{x}}, {\mathbf{x}}'$ in the teaching set $\frac{\inmod{{\mathbf{x}} - {\mathbf{x}}'}^2}{2\sigma^2} = \bigTheta{\log \frac{1}{\epsilon}}$ thus all the non-diagonal entries of $\textLambda$ could be bounded as $\bigTheta{\epsilon}$. Thus, the non-diagonal entries of $\textLambda$ are upper bounded w.r.t to the choice of $\epsilon$. We denote the concatenated vector of $\gamma$ and $\beta_0$ by $\ensuremath{\boldsymbol{\eta}}$ as $\ensuremath{\boldsymbol{\eta}} := (\gamma^\top,\beta_0)^\top$. Consider the following matrix equation: \begin{equation} \textLambda\cdot \ensuremath{\boldsymbol{\eta}} = \parenb{\underbrace{0,\cdots,0}_{\textnormal{For}\; {\mathbf{z}}_i's},\underbrace{\vphantom{0,\cdots,0}1}_{\textnormal{For}\; {\mathbf{a}}}}^{\top}\label{eqn: satisfyobjective} \end{equation} Notice that any solution $\ensuremath{\boldsymbol{\eta}}$ to \eqnref{eqn: satisfyobjective} has zero objective value for \eqnref{eqn: bounded}. Since $\sum_{i=1}^r \paren{\textLambda[i,r]\cdot\ensuremath{\boldsymbol{\eta}}_i} = \hat{\boldsymbol{\theta}}\cdot\Phi({\mathbf{a}}) > 0$ thus we scale the last component of \eqnref{eqn: satisfyobjective} to 1. First, we observe that \eqnref{eqn: satisfyobjective} has a solution because Gaussian kernel Gram matrix $\textLambda$ to the finite set of points is \tt{strictly positive definite} implying $\textLambda$ is invertible. Thus, there is a unique solution $\ensuremath{\boldsymbol{\eta}}_0 \in \ensuremath{\mathbb{R}}^{r}$ such that: \begin{equation} \textLambda\cdot \ensuremath{\boldsymbol{\eta}}_0 = \nu^{\top}\nonumber \end{equation} where $\nu := (0,0,\cdots,1)$ as shown in \eqnref{eqn: satisfyobjective}. Also, $\ensuremath{\boldsymbol{\eta}}_0^{\top}\cdot\textLambda\cdot\ensuremath{\boldsymbol{\eta}}_0 = \beta_0 > 0$. Now, we need to ensure that $\ensuremath{\boldsymbol{\eta}}_0$ satisfies \assref{assumption: bounded cone}. To analyse the boundedness, we rewrite the above equation as: \begin{equation} \ensuremath{\boldsymbol{\eta}}_0 = \textLambda^{-1}\cdot \nu^{\top}\nonumber \end{equation} To evaluate the entries of $\ensuremath{\boldsymbol{\eta}}_0$, we only need to understand the last column of $\textLambda^{-1}$ (since $\textLambda^{-1}\cdot \nu^{\top}$ contains entries from the last column of $\textLambda^{-1}$). Using the construction of the inverse using the minors of $\textLambda$, we note that $\textLambda^{-1}[i,r] = \frac{1}{\det(\textLambda)}\cdot M_{(i,r)}$, where $M_{(i,r)}$ is the minor of $\textLambda$ corresponding to entry indexed as $(i, r)$ (determinant of the submatrix of $\textLambda$ formed by removing the \tt{i}th row and \tt{r}th column). Note, determinant is an alternating form, which for a square matrix $T$ of dimension $n$ has the explicit sum $\sum_{\sigma \in S^n}sign(\sigma)\cdot\prod_{i}^n u_{i,\sigma(i)}$, where $T[i,j] = u_{i,j}$. Since the non-diagonal entries of $\textLambda$ are bounded by $\epsilon$, thus we can bound the minors. Note, $|M_{r,r}| \ge 1 - \bigO{\epsilon^2} + \cdots + (-1)^{r-1} \bigO{\epsilon^{r-1}}$ and for $i \neq r$ $|M_{i,r}| \le \bigO{\epsilon} + \bigO{\epsilon^2} + \cdots + \bigO{\epsilon^{r-1}}$. Since, $\epsilon$ is sufficiently small, thus $|M_{r,r}|$ majorizes over $|M_{i,r}|$ for $i \neq r$. But then $\ensuremath{\boldsymbol{\eta}}_0 = \frac{1}{\det(\textLambda)}\paren{(-1)^{1+r}M_{1, r}, (-1)^{2+r}M_{2, r}, \cdots, (-1)^{r+r}M_{r, r}}^{\top}$. When we normalize $\boldsymbol{\theta}$, we get $\bar{\ensuremath{\boldsymbol{\eta}}}_0 = \ensuremath{\boldsymbol{\eta}}_0/\inmod{\boldsymbol{\theta}}$. We note that, $\inmod{\boldsymbol{\theta}}^2 = \ensuremath{\boldsymbol{\eta}}_0^{\top}\cdot\textLambda\cdot\ensuremath{\boldsymbol{\eta}}_0 = \beta_0$, implying $\bar{\ensuremath{\boldsymbol{\eta}}}_0 = \ensuremath{\boldsymbol{\eta}}_0/\sqrt{\beta_0}$. Since, $\beta_0 = \frac{1}{\det(\textLambda)}\cdot M_{r,r}$, thus entries of $\bar{\ensuremath{\boldsymbol{\eta}}}_0$ satisfy \assref{assumption: bounded cone}. Thus, we have a solution to \eqnref{eqn: bounded} which satisfies \assref{assumption: bounded cone}. \subsection{Proof of \thmref{thm: boundedclassifier} and \thmref{thm: gaussian_main_thm}}\label{appendixsub: proofofmainthm} In this section, we would establish our key results for the approximate teaching of a Gaussian kernel perceptron. Under the \assref{assumption: orthogonal} and \assref{assumption: bounded cone}, we would show to teach a target model $\boldsymbol{\theta}^*$ $\epsilon$-approximately we only require at most $d^{\bigO{\log^2\frac{1}{\epsilon}}}$ labelled teaching points from ${\mathcal{X}}$. In order to achieve the $\epsilon$-approximate teaching set, we would show that the teaching set $\mathcal{TS}_{\boldsymbol{\theta}^*}$ as constructed in \eqnref{eqn: teaching set} achieves an $\epsilon$-closeness between $f^* = \boldsymbol{\theta}^*\cdot\Phi(\cdot)$ and $\hat{f} = \hat{\boldsymbol{\theta}}\cdot\Phi(\cdot)$ i.e $ \left|f^*({\mathbf{x}}) - \hat{f}({\mathbf{x}})\right| \le \epsilon$ point-wise. Before we move to the proofs of the key results, we state the following relevant lemma which bounds the length (norm) of a vector spanned by a basis with the smoothness condition on the basis as mentioned in \secref{subsec: bounded_error}. \begin{lemma}\label{lemma: maximum norm} Consider the Euclidian space $\ensuremath{\mathbb{R}}^n$. Assume $\curlybracket{{\mathbf{v}}_i}_{i=1}^n$ forms a basis with unit norms. Additionally, for any $i,j$ $\left|{\mathbf{v}}_i\cdot {\mathbf{v}}_j\right| \le \cos{\theta_0}$ where $ \cos{\theta_0} \le \frac{1}{2n}$. Fix a small real scalar $\epsilon > 0$. Now, consider any random vector ${\mathbf{p}} \in \ensuremath{\mathbb{R}}^n$ such that $\forall i \in \bracket{n}$ $|{\mathbf{p}} \cdot {\mathbf{v}}_i| \le \epsilon$. Then the following bound on ${\mathbf{p}}$ holds: $$||{\mathbf{p}}||_2 \le \sqrt{2n}\cdot \epsilon.$$ \end{lemma} \begin{proof} We define $M = \ensuremath{\mathbb{R}}^n$ as the space in which ${\mathbf{p}}$ and ${\mathbf{v}}_i$'s are embedded. Consider another copy of the space $N = \ensuremath{\mathbb{R}}^n$ with standard orthogonal basis $\curlybracket{e_1,\cdots,e_n}$. We define the map $\boldsymbol{\mathrm{W}}: M \longrightarrow N$ as follows: \begin{align*} \boldsymbol{\mathrm{W}} &: M \longrightarrow N\\ q &\mapsto ({\mathbf{v}}_1\cdot q,{\mathbf{v}}_2\cdot q,\cdots,{\mathbf{v}}_n\cdot q) \end{align*} Since $\curlybracket{{\mathbf{v}}_i}_{i=1}^n$ forms a basis, thus $\boldsymbol{\mathrm{W}}$ is invertible. To ease the analysis, we could assume $\epsilon = 1$ (follows by scaling symmetry). Thus, it is clear that $w := \boldsymbol{\mathrm{W}}{\mathbf{p}}$ has all its entries bounded in absolute value by 1. We could write ${\mathbf{p}} = \boldsymbol{\mathrm{W}}^{-1}w $, thus $\inmod{{\mathbf{p}}}^2 = \paren{\boldsymbol{\mathrm{W}}^{-1}w}^{\top}\paren{\boldsymbol{\mathrm{W}}^{-1}w} = w^{\top}\paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}^{-1}w$. Thus, showing the bound for $w^{\top}\paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}^{-1}w$ where $w \in \bracket{-1,1}^n$ suffices. We note that $\paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}$ is a symmetric $(n\times n)$ matrix with diagonal entries 1 and non-diagonal entries bounded in absolute value by $\cos{\theta_0}$. Using convergence of the Neumann series $\sum_{k=0}^{\infty} \paren{\mathbb{Id}_{\paren{n\times n}} - \paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}}^k$ as $\paren{\mathbb{Id}_{\paren{n\times n}} - \paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}}$ is a bounded operator, we have: \begin{align} \left|\paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}^{-1} - \mathbb{Id}_{\paren{n\times n}}\right|_{\ell_{\infty}} &\le \left|\sum_{k=1}^{\infty} \paren{\mathbb{Id}_{\paren{n\times n}} - \paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}}^k\right|_{\ell_{\infty}} \label{eqn: neu1}\\ &\le \sum_{k=1}^{\infty}\left| \paren{\mathbb{Id}_{\paren{n\times n}} - \paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}}^k\right|_{\ell_{\infty}}\label{eqn: neu2}\\ &\le \sum_{k=1}^{\infty} n^{k-1}\cos^{k}{\theta_0}\label{eqn: neu3}\\ &= \frac{\cos{\theta_0}}{1-n\cos{\theta_0}}\label{eqn: neu4} \end{align} where $|B|_{\ell_{\infty}}$ refers to the maximum absolute value of any entry of $B$. \eqnref{eqn: neu1} follows using the Neumann series $\paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}^{-1}$ = $\sum_{k=0}^{\infty} \paren{\mathbb{Id}_{\paren{n\times n}} - \paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}}^k$. \eqnref{eqn: neu2} is a direct consequence of triangle inequality. Since entries of $\paren{\mathbb{Id}_{\paren{n\times n}} - \paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}}^k$ are bounded in absolute value by $\cos{\theta_0}$ thus \eqnref{eqn: neu3} follows. Using a straight-forward geometric sum we get an upper bound on the maximum absolute value of any entry in $\paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}^{-1} -\, \mathbb{Id}_{\paren{n\times n}}$ in \eqnref{eqn: neu4}. Now, we note that \begin{align} w^{\top}\paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}^{-1}w &= w^{\top}\paren{\paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}^{-1} -\, \mathbb{Id}_{\paren{n\times n}}}w + w^{\top}\paren{\mathbb{Id}_{\paren{n\times n}}}w\nonumber\\ &\le \sum_{i,j}w_{ij}\paren{\paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}^{-1} -\, \mathbb{Id}_{\paren{n\times n}}}_{ij}w_{ij} + \inmod{w}^2\nonumber\\ &\le \sum_{i,j}\left|w_{ij}\paren{\paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}^{-1} -\, \mathbb{Id}_{\paren{n\times n}}}_{ij}w_{ij}\right| + \inmod{w}^2\nonumber\\ &\le \sum_{i,j} \left|\paren{\paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}^{-1} -\, \mathbb{Id}_{\paren{n\times n}}}_{ij}\right| + n\label{eqn: b1}\\ &\le \frac{n^2\cos{\theta_0}}{1-n\cos{\theta_0}} + n\label{eqn: b2}\\ & = \frac{n}{1-n\cos{\theta_0}}\label{eqn: b3} \end{align} In \eqnref{eqn: b1} we use $w \in \bracket{-1,1}^n$. \eqnref{eqn: b2} follows using \eqnref{eqn: neu4}. Since we have $\inmod{{\mathbf{p}}}_2^2 = w^{\top}\paren{\boldsymbol{\mathrm{W}}^{\top}\boldsymbol{\mathrm{W}}}^{-1}w$ and $\cos{\theta_0} \le \frac{1}{2n}$, thus using \eqnref{eqn: b3} $$\inmod{{\mathbf{p}}}_2 \le \sqrt{\frac{n}{1-n\cos{\theta_0}}} \le \sqrt{2n}.$$ Scaling the map $\boldsymbol{\mathrm{W}}$ to $\epsilon$ yields the stated claim. \end{proof} Under the \assref{assumption: orthogonal} and \assref{assumption: bounded cone}, and bounded norm of $\boldsymbol{\theta}^*$ and $\hat{\boldsymbol{\theta}}$, we would establish that $\mathcal{TS}_{\boldsymbol{\theta}^*}$ is a $d^{\bigO{\log^2 \frac{1}{\epsilon}}}$ size $\epsilon$-approximate teaching set for $\boldsymbol{\theta}^*$. Before establishing the main result, we show the proof of \thmref{thm: boundedclassifier} below. Using \eqnref{eqn: largeeps}, we note that: \begin{align*} \forall\,{\mathbf{x}} \in {\mathcal{X}}\quad &\inmod{\mathbb{P}^{\perp}\Phi({\mathbf{x}})} \le \sqrt{\epsilon_s} \implies \inmod{\mathbb{P}\Phi({\mathbf{x}})} \ge \sqrt{1-\epsilon_s}\\ \forall\,({\mathbf{z}},y) \in \mathcal{TS}_{\boldsymbol{\theta}^*}\quad &\inmod{\mathbb{P}^{\perp}\Phi({\mathbf{z}})} \le \sqrt{\epsilon_s} \implies \inmod{\mathbb{P}\Phi({\mathbf{z}})} \ge \sqrt{1-\epsilon_s} \end{align*} Now, we could further bound the norms of $\mathbb{P}^{\perp}\hat{\boldsymbol{\theta}}$ and $\mathbb{P}^{\perp}\boldsymbol{\theta}^*$ using triangle inequality and boundedness of $\curlybracket{\alpha_i}_{i=1}^l$ and $\bracket{\beta_0,\gamma}$ (as shown in \assref{assumption: bounded cone}): \begin{align*} \inmod{\mathbb{P}^{\perp}\boldsymbol{\theta}^*} = \inmod{\sum_{i=1}^l \alpha_i\cdot \mathbb{P}^{\perp}\Phi({\mathbf{a}}_i)} &\le \sum_{i=1}^l \inmod{\alpha_i\cdot \mathbb{P}^{\perp}\Phi({\mathbf{a}}_i)} \le \paren{\sum_{i=1}^l \left| \alpha_i\right|}\cdot \sqrt{\epsilon_s} = \boldsymbol{C}_{\epsilon}\cdot \sqrt{\epsilon_s}\\ \inmod{\mathbb{P}^{\perp}\hat{\boldsymbol{\theta}}} = \inmod{\beta_0\cdot \mathbb{P}^{\perp}\Phi({\mathbf{a}}) + \sum_{j=1}^{r-1}\gamma_j\cdot \mathbb{P}^{\perp}\Phi({\mathbf{z}}_j)} &\le \inmod{\beta_0\cdot \mathbb{P}^{\perp}\Phi({\mathbf{a}})} + \sum_{i=1}^{r-1} \inmod{\gamma_i\cdot \mathbb{P}^{\perp}\Phi({\mathbf{z}}_i)} \le \paren{\left|\beta_0\right| + \sum_{i=1}^{r-1} \left| \gamma_i\right|}\cdot \sqrt{\epsilon_s} = \boldsymbol{D}_{\epsilon}\cdot \sqrt{\epsilon_s} \end{align*} \begin{proof}[Proof of \thmref{thm: boundedclassifier}] In the following, we would bound $\left|f^*({\mathbf{x}}) - \hat{f}({\mathbf{x}})\right|$ by $ \sqrt{\epsilon}$. In order to bound the modulus, we would split the difference using $\mathbb{P}$ and $\mathbb{P}^{\bot}$ and then analyze the terms correspondingly. We can write any classifier $f$ as $f({\mathbf{x}}) = \boldsymbol{\theta}\cdot\Phi({\mathbf{x}}) = \mathbb{P}\boldsymbol{\theta}\cdot\mathbb{P}\Phi({\mathbf{x}}) + \mathbb{P}^{\bot}\boldsymbol{\theta}\cdot\mathbb{P}^{\bot}\Phi({\mathbf{x}})$. Thus, we have: \begin{align} \left|f^*({\mathbf{x}}) - \hat{f}({\mathbf{x}})\right| &= \left|\mathbb{P}\boldsymbol{\theta}^*\cdot\mathbb{P}\Phi({\mathbf{x}}) + \mathbb{P}^{\bot}\boldsymbol{\theta}^*\cdot\mathbb{P}^{\bot}\Phi({\mathbf{x}}) - \mathbb{P}\hat{\boldsymbol{\theta}}\cdot\mathbb{P}\Phi({\mathbf{x}}) - \mathbb{P}^{\bot}\hat{\boldsymbol{\theta}}\cdot\mathbb{P}^{\bot}\Phi({\mathbf{x}})\right|\nonumber\\ &\le \left|\mathbb{P}\boldsymbol{\theta}^*\cdot\mathbb{P}\Phi({\mathbf{x}})- \mathbb{P}\hat{\boldsymbol{\theta}}\cdot\mathbb{P}\Phi({\mathbf{x}})\right| + \left|\mathbb{P}^{\bot}\boldsymbol{\theta}^*\cdot\mathbb{P}^{\bot}\Phi({\mathbf{x}})-\mathbb{P}^{\bot}\hat{\boldsymbol{\theta}}\cdot\mathbb{P}^{\bot}\Phi({\mathbf{x}})\right|\label{eqn: split1}\\ &\le \left|\mathbb{P}\boldsymbol{\theta}^*\cdot\mathbb{P}\Phi({\mathbf{x}})-\mathbb{P}\hat{\boldsymbol{\theta}}\cdot\mathbb{P}\Phi({\mathbf{x}})\right| + \inmod{\mathbb{P}^{\bot}\boldsymbol{\theta}^*-\mathbb{P}^{\bot}\hat{\boldsymbol{\theta}}}\cdot\inmod{\mathbb{P}^{\bot}\Phi({\mathbf{x}})}\label{eqn: split2}\\ &\le \underbrace{\left|\mathbb{P}\boldsymbol{\theta}^*\cdot\mathbb{P}\Phi({\mathbf{x}})- \mathbb{P}\hat{\boldsymbol{\theta}}\cdot\mathbb{P}\Phi({\mathbf{x}})\right|}_{\bigstar} +\, \paren{\boldsymbol{C}_{\epsilon}+\boldsymbol{D}_{\epsilon}}\cdot \epsilon_s \label{eqn: split3} \end{align} \eqnref{eqn: split1} follows using triangle inequality. We can further bound $\left|\mathbb{P}^{\bot}\boldsymbol{\theta}^*\cdot\mathbb{P}^{\bot}\Phi({\mathbf{x}})-\mathbb{P}^{\bot}\hat{\boldsymbol{\theta}}\cdot\mathbb{P}^{\bot}\Phi({\mathbf{x}})\right|$ using Cauchy-Schwarz inequality and thus \eqnref{eqn: split2} follows. Using the observations: $||\mathbb{P}^{\bot}\hat{\boldsymbol{\theta}}|| \le \boldsymbol{D}_{\epsilon}\cdot\sqrt{\epsilon_s}$ and $\inmod{\mathbb{P}^{\bot}\boldsymbol{\theta}^*} \le \boldsymbol{C}_{\epsilon}\cdot\sqrt{\epsilon_s}$, we could upper bound $\inmod{\mathbb{P}^{\bot}\boldsymbol{\theta}^*-\mathbb{P}^{\bot}\hat{\boldsymbol{\theta}}}$ by $\paren{\boldsymbol{C}_{\epsilon}+\boldsymbol{D}_{\epsilon}}\cdot \sqrt{\epsilon_s}$. Since ${\mathbf{x}} \in {\mathcal{X}}$ thus $\inmod{\mathbb{P}^{\bot}\Phi({\mathbf{x}})} \le \sqrt{\epsilon_s}$ (as shown in \eqnref{eqn: largeeps}), which gives \eqnref{eqn: split3}. Now, the key is to bound the $(\bigstar)$ appropriately and then the result would be proven. We would rewrite $\mathbb{P}\hat{\boldsymbol{\theta}}\cdot\mathbb{P}\Phi({\mathbf{x}})$ in terms of the basis formed by $\{\mathbb{P}\Phi({\mathbf{z}}_i)\}_{i=1}^{r-1} \cup \{\mathbb{P}\theta^*\}$ (by \assref{assumption: orthogonal} $\{\mathbb{P}\Phi({\mathbf{z}}_i)\}_{i=1}^l$ are linearly independent and orthogonal to $\mathbb{P}\theta^*$). Using the basis, we can write $\mathbb{P}\hat{\boldsymbol{\theta}} = \sum_{i=1}^{r-1}c_i\cdot \mathbb{P}\Phi({\mathbf{z}}_i) + \lambda_r\cdot\mathbb{P}\boldsymbol{\theta}^*$ for some scalars $c_1,c_2,\cdots,\lambda_r$. Alternatively, we could rewrite $\mathbb{P}\hat{\boldsymbol{\theta}} = \beta_0\cdot \mathbb{P}\Phi({\mathbf{a}}) + \sum_{j=1}^{r-1}\gamma_j\cdot \mathbb{P}\Phi({\mathbf{z}}_j)$ where $\beta_0 > 0$ (as shown in \appref{appendixsub: solutionexists}). This could be used to note that $\lambda_r > 0$ because $\mathbb{P}\boldsymbol{\theta}^*\cdot\mathbb{P}\Phi({\mathbf{a}}) > 0$ (cf \secref{subsec.gaussiankernel}). We study the decomposition of $\mathbb{P}\hat{\boldsymbol{\theta}}$ in terms of the basis in order to understand the component of $\mathbb{P}\hat{\boldsymbol{\theta}}$ along $\mathbb{P}\boldsymbol{\theta}^*$. We observe that: \begin{equation} \inmod{\mathbb{P}\hat{\boldsymbol{\theta}}}^2 = \inmod{\sum_{i=1}^{r-1}c_i\cdot \mathbb{P}\Phi({\mathbf{z}}_i)}^2 + \inmod{\lambda_r\cdot\mathbb{P}\boldsymbol{\theta}^*}^2 \label{eqn:perpnorm} \end{equation} Since $\hat{\boldsymbol{\theta}}$ is a solution to \eqnref{eqn: bounded}, $\hat{\boldsymbol{\theta}}\cdot\Phi({\mathbf{z}}_i) = 0$ for any $i \in \bracket{r-1}$. Now, we can write the equation in terms of projections as: \begin{equation} \forall i\quad \mathbb{P}\hat{\boldsymbol{\theta}}\cdot \mathbb{P}\Phi({\mathbf{z}}_i) + \mathbb{P}^{\bot}\hat{\boldsymbol{\theta}}\cdot \mathbb{P}^{\bot}\Phi({\mathbf{z}}_i) = 0\label{eqn: expandeqn} \end{equation} Using Cauchy-Schwarz inequality on the product $|\mathbb{P}^{\bot}\hat{\boldsymbol{\theta}}\cdot \mathbb{P}^{\bot}\Phi({\mathbf{z}}_i)|$ we obtain: \begin{equation} |\mathbb{P}^{\bot}\hat{\boldsymbol{\theta}}\cdot \mathbb{P}^{\bot}\Phi({\mathbf{z}}_i)| \le ||\mathbb{P}^{\bot}\hat{\boldsymbol{\theta}}||\cdot ||\mathbb{P}^{\bot}\Phi({\mathbf{z}}_i)|| \le \boldsymbol{D}_{\epsilon}\cdot\epsilon_s \nonumber \end{equation} Plugging this into \eqnref{eqn: expandeqn}, we get the following bound on $|\mathbb{P}\hat{\boldsymbol{\theta}}\cdot \mathbb{P}\Phi({\mathbf{z}}_i)|$: \begin{equation} |\mathbb{P}\hat{\boldsymbol{\theta}}\cdot \mathbb{P}\Phi({\mathbf{z}}_i)| \le \boldsymbol{D}_{\epsilon}\cdot\epsilon_s \label{eqn: boundproj} \end{equation} We denote $V_{O} := \sum_{i=1}^{r-1}c_i\cdot \mathbb{P}\Phi({\mathbf{z}}_i)$. Notice that $V_{O}$ is the orthogonal projection of $\mathbb{P}\hat{\boldsymbol{\theta}}$ along the subspace $\textbf{span}\langle\mathbb{P}\Phi({\mathbf{z}}_1),\cdots,\mathbb{P}\Phi({\mathbf{z}}_{r-1})\rangle$. Thus, we could rewrite \eqnref{eqn: boundproj} further as: \begin{equation} |\mathbb{P}\hat{\boldsymbol{\theta}}\cdot \mathbb{P}\Phi({\mathbf{z}}_i)| = |\paren{V_{O} + \lambda_r\cdot\mathbb{P}\boldsymbol{\theta}^*}\cdot \mathbb{P}\Phi({\mathbf{z}}_i)| = |V_{O}\cdot \mathbb{P}\Phi({\mathbf{z}}_i)| \le \boldsymbol{D}_{\epsilon}\cdot\epsilon_s \nonumber \end{equation} Notice that $||\mathbb{P}\Phi({\mathbf{z}}_i)|| \ge \sqrt{1-\epsilon_s}$. Hence, component of $V_{O}$ along $\mathbb{P}\Phi({\mathbf{z}}_i)$ is upper bounded by $\frac{\boldsymbol{D}_{\epsilon}\cdot\epsilon_s}{\sqrt{1-\epsilon_s}}$. Since $\curlybracket{\mathbb{P}\Phi({\mathbf{z}}_1),\cdots,\mathbb{P}\Phi({\mathbf{z}}_{r-1})}$ satisfy the conditions of \lemref{lemma: maximum norm} (the smoothness condition mentioned in \secref{subsec.gaussiankernel}) thus we could bound the norm of $V_{O}$ as follows: \begin{equation} ||V_{O}|| \le \sqrt{2(r-1)}\cdot \frac{\boldsymbol{D}_{\epsilon}\cdot\epsilon_s}{\sqrt{1-\epsilon_s}} \label{eqn: finalbound} \end{equation} Using \eqnref{eqn:perpnorm} and \eqnref{eqn: finalbound} we can lower bound the norm of $\lambda_r\cdot\mathbb{P}\boldsymbol{\theta}^*$ as follows: \begin{align} \inmod{\mathbb{P}\hat{\boldsymbol{\theta}}}^2 =& \inmod{\sum_{i=1}^{r-1}c_i\cdot \mathbb{P}\Phi({\mathbf{z}}_i)}^2 +\inmod{\lambda_r\cdot\mathbb{P}\boldsymbol{\theta}^*}^2 = \inmod{V_{O}}^2 +\inmod{\lambda_r\cdot\mathbb{P}\boldsymbol{\theta}^*}^2\nonumber\\ \implies& \inmod{\lambda_r\cdot\mathbb{P}\boldsymbol{\theta}^*}^2 \ge \paren{1-\boldsymbol{D}_{\epsilon}^2\cdot\epsilon_s} - 2\paren{r-1}\cdot \frac{\boldsymbol{D}_{\epsilon}^2\cdot\epsilon_s^2}{\paren{1-\epsilon_s}} \ge 1-2\boldsymbol{D}_{\epsilon}^2\cdot\epsilon_s \label{eqn: actualnorm} \end{align} This follows because $\inmod{\mathbb{P}\hat{\boldsymbol{\theta}}}^2 \ge \paren{1-\boldsymbol{D}_{\epsilon}^2\cdot\epsilon_s} $ as $\inmod{\hat{\boldsymbol{\theta}}} = \bigO{1}$ and $\sqrt{2(r-1)}\cdot \epsilon_s \le \sqrt{2(r-1)}\cdot \frac{\epsilon}{(\sqrt{d})^s} \le \epsilon$. With these observations we can rewrite $\paren{\bigstar}$ as follows: \begin{align} \centering \left|\mathbb{P}\boldsymbol{\theta}^*\cdot\mathbb{P}\Phi({\mathbf{x}})- \mathbb{P}\hat{\boldsymbol{\theta}}\cdot\mathbb{P}\Phi({\mathbf{x}})\right| &= \left|\mathbb{P}\boldsymbol{\theta}^*\cdot\mathbb{P}\Phi({\mathbf{x}})- \sum_{i=1}^{r-1}c_i\cdot \mathbb{P}\Phi({\mathbf{z}}_i)\cdot\mathbb{P}\Phi({\mathbf{x}}) - \lambda_r\cdot\mathbb{P}\boldsymbol{\theta}^*\cdot\mathbb{P}\Phi({\mathbf{x}})\right|\nonumber\\ &\le \left|\mathbb{P}\boldsymbol{\theta}^*\cdot\mathbb{P}\Phi({\mathbf{x}})- \lambda_r\cdot\mathbb{P}\boldsymbol{\theta}^*\cdot\mathbb{P}\Phi({\mathbf{x}})\right| + \left| \sum_{i=1}^{r-1}c_i\cdot \mathbb{P}\Phi({\mathbf{z}}_i)\cdot\mathbb{P}\Phi({\mathbf{x}})\right|\label{eqn: triangle}\\ &\le \sqrt{|\boldsymbol{C}_{\epsilon}^2 - 2\boldsymbol{D}_{\epsilon}^2|}\cdot\sqrt{\epsilon_s} + \inmod{\sum_{i=1}^{r-1}c_i\cdot \mathbb{P}\Phi({\mathbf{z}}_i)}\cdot \inmod{\mathbb{P}\Phi({\mathbf{x}})}\label{eqn: boundfinal1}\\ &\le \sqrt{|\boldsymbol{C}_{\epsilon}^2 - 2\boldsymbol{D}_{\epsilon}^2|}\cdot\sqrt{\epsilon_s} +\sqrt{2(r-1)}\cdot \frac{\boldsymbol{D}_{\epsilon}\cdot\epsilon_s}{\sqrt{1-\epsilon_s}}\label{eqn: boundfinal2}\\ & \le \frac{\sqrt{|\boldsymbol{C}_{\epsilon}^2 - 2\boldsymbol{D}_{\epsilon}^2|}\cdot\sqrt{\epsilon}}{(\sqrt{d})^{s/2}} + 2\boldsymbol{D}_{\epsilon}\cdot\epsilon\label{eqn: boundfinal3}\\ & \le 2\max\left\{\frac{\sqrt{|\boldsymbol{C}_{\epsilon}^2 - 2\boldsymbol{D}_{\epsilon}^2|}}{(\sqrt{d})^{s/2}},\, 2\boldsymbol{D}_{\epsilon}\cdot\sqrt{\epsilon}\right\}\cdot \sqrt{\epsilon}\label{eqn: boundfinal4} \end{align} \eqnref{eqn: triangle} is a direct implication of triangle inequality. In \eqnref{eqn: boundfinal1}, in the first term we note that $\lambda_r > 0$ () $\inmod{\mathbb{P}\boldsymbol{\theta}^*} \ge \sqrt{1-\boldsymbol{C}_{\epsilon}^2\cdot\epsilon_s}$ and use \eqnref{eqn: actualnorm}, and in the second use Cauchy-Schwarz inequality. \eqnref{eqn: boundfinal2} follows using \eqnref{eqn: finalbound} and that $||\mathbb{P}\Phi({\mathbf{x}})||$ is bounded by 1. We could unfold the value of $\epsilon_s \le \frac{\epsilon}{(\sqrt{d})^s}$. This gives us \eqnref{eqn: boundfinal3}. We could rewrite \eqnref{eqn: boundfinal3} to get a bound in terms of $\sqrt{\epsilon}$ to obtain \eqnref{eqn: boundfinal4}. Now, using \eqnref{eqn: split3} and \eqnref{eqn: boundfinal4}, can bound $\left|f^*({\mathbf{x}}) - \hat{f}({\mathbf{x}})\right|$ as follows: \begin{align*} \left|f^*({\mathbf{x}}) - \hat{f}({\mathbf{x}})\right| &\le 2\max\left\{\frac{\sqrt{|\boldsymbol{C}_{\epsilon}^2 - 2\boldsymbol{D}_{\epsilon}^2|}}{(\sqrt{d})^{s/2}},\, 2\boldsymbol{D}_{\epsilon}\cdot\sqrt{\epsilon}\right\}\cdot \sqrt{\epsilon} + \paren{\boldsymbol{C}_{\epsilon}+\boldsymbol{D}_{\epsilon}}\cdot \frac{\epsilon}{(\sqrt{d})^s}\\ &\le 3\max\left\{\frac{\sqrt{|\boldsymbol{C}_{\epsilon}^2 - 2\boldsymbol{D}_{\epsilon}^2|}}{(\sqrt{d})^{s/2}},\, 2\boldsymbol{D}_{\epsilon}\cdot\sqrt{\epsilon},\, \paren{\boldsymbol{C}_{\epsilon}+\boldsymbol{D}_{\epsilon}}\cdot \frac{\sqrt{\epsilon}}{(\sqrt{d})^s} \right \}\cdot \sqrt{\epsilon}\\ &\le 3\boldsymbol{C}'\cdot \sqrt{\epsilon} \end{align*} where $\boldsymbol{C}' := \max\left\{\frac{\sqrt{|\boldsymbol{C}_{\epsilon}^2 - 2\boldsymbol{D}_{\epsilon}^2|}}{(\sqrt{d})^{s/2}},\, 2\boldsymbol{D}_{\epsilon}\cdot\sqrt{\epsilon},\, \paren{\boldsymbol{C}_{\epsilon}+\boldsymbol{D}_{\epsilon}}\cdot \frac{\sqrt{\epsilon}}{(\sqrt{d})^s} \right \}$. \\ Notice that all the terms in $\max\left\{\frac{\sqrt{|\boldsymbol{C}_{\epsilon}^2 - 2\boldsymbol{D}_{\epsilon}^2|}}{(\sqrt{d})^{s/2}},\, 2\boldsymbol{D}_{\epsilon}\cdot\sqrt{\epsilon},\, \paren{\boldsymbol{C}_{\epsilon}+\boldsymbol{D}_{\epsilon}}\cdot \frac{\sqrt{\epsilon}}{(\sqrt{d})^s} \right \}$ are smaller than 1 because of boundedness of $\boldsymbol{C}_{\epsilon}$ and $\boldsymbol{D}_{\epsilon}$. Thus, we have shown a $3C'\cdot\sqrt{\epsilon}$ (where $C'$ is a constant smaller than 1) bound on the point-wise difference of $\hat{f}$ and $f^*$. Now, if we scale the $\epsilon$ and solve for $\epsilon^2/3$, we get the desired bound. Hence, the main claim of \thmref{thm: boundedclassifier} is proven i.e. $ \left|f^*({\mathbf{x}}) - \hat{f}({\mathbf{x}})\right| \le \epsilon$. \end{proof} Now, we would complete the proof of the main result of \secref{subsec.gaussiankernel} which bounds the error incurred by the solution $\hat{\boldsymbol{\theta}} \in {\boldsymbol{\mathcal{A}}}_{opt}(\mathcal{TS}_{\boldsymbol{\theta}^*})$ i.e. \thmref{thm: gaussian_main_thm}. The point-wise closeness of $f^*$ and $\hat{f}$ established in \thmref{thm: boundedclassifier} would be key in bounding the error. We complete the proof as follows: \begin{proof}[Proof of \thmref{thm: gaussian_main_thm}] We show the error analysis when data-points are sampled from the data distribution ${\mathcal{P}}$. \begin{align} \left|\textbf{err}(f^*) - \textbf{err}(\hat{f})\right| &= \left|\expover{({\mathbf{x}},y) \sim {\mathcal{P}}}{\max(-y\cdot f^*({\mathbf{x}}), 0)} - \expover{({\mathbf{x}},y) \sim {\mathcal{P}}}{\max(-y\cdot \hat{f}({\mathbf{x}}), 0)}\right|\label{eqn:final1}\\ &= \left|\expover{({\mathbf{x}},y) \sim {\mathcal{P}}}{\max(-y\cdot f^*({\mathbf{x}}), 0) - \max(-y\cdot \hat{f}({\mathbf{x}}), 0)}\right|\label{eqn:final2}\\ &\le \expover{({\mathbf{x}},y) \sim {\mathcal{P}}}{\left|f^*({\mathbf{x}}) - \hat{f}({\mathbf{x}})\right|}\label{eqn:final3}\\ &\le \epsilon \label{eqn:final4} \end{align} \eqnref{eqn:final1} follows using the definition of $\textbf{err}(\cdot)$ function. Because of linearity of expectation, we get \eqnref{eqn:final2}. In \eqnref{eqn:final3}, we use the observation that modulus of an expectation is bounded by the expectation of the modulus of the random variable $f^*({\mathbf{x}}) - \hat{f}({\mathbf{x}})$. In \thmref{thm: boundedclassifier}, we showed that for any ${\mathbf{x}} \in {\mathcal{X}}$, $\left|f^*({\mathbf{x}}) - \hat{f}({\mathbf{x}})\right| \le \epsilon$. Thus, the main claim follows. \end{proof} \newpage \section{Linear Perceptrons}\label{appendix: linear perceptron} \section{Polynomial Kernel Perceptron}\label{appendix: polynomial perceptron} In this appendix, we would provide the proof for the main result of \secref{subsec.poly} i.e. \thmref{thm: poly_main_theorem}. We would complete the proof by constructing a teaching set for exact teaching. Similar to the proof of \thmref{thm: linear perceptron main result}, the key idea is to find linear independent polynomials on the orthogonal subspace defined by $\boldsymbol{\theta}^* \in {\mathcal{H}}_{{\mathcal{K}}}$. Our \assref{assumption: polyorthogonal} would ensure that there are such $r-1$ linear independent polynomials. Rest of the work follows steps inspired as seen in the proof of \thmref{thm: linear perceptron main result}. We assume that $\boldsymbol{\theta}^*$ is non-degenerate and has at least one point in $\mathbb{R}^d$ classified \tt{strictly} positive, and provide the poof below. \begin{proof}[Proof of \thmref{thm: poly_main_theorem}] First, we would show the construction of a teaching set for a target model $\boldsymbol{\theta}^* \in {\mathcal{H}}_{{\mathcal{K}}}$. Denote by $\mathcal{V}^{\bot}_{\boldsymbol{\theta}^*} \subset {\mathcal{H}}_k$ ($\cong$ ${\mathcal{H}}_{{\mathcal{K}}}$ using \propref{prop: polynomial space} i.e isomorphic as vector spaces) the orthogonal subspace of $\boldsymbol{\theta}^*$. Since $\boldsymbol{\theta}^*$ satisfies \assref{assumption: polyorthogonal}, thus there exists a set of $r-1$ linearly independent vectors (polynomials because of \propref{prop: polynomial space}) of the form $\{\Phi({\mathbf{z}}_i)\}_{i=1}^{r-1}$ in $\mathcal{V}^{\bot}_{\boldsymbol{\theta}^*}$ where $\curlybracket{{\mathbf{z}}_i}_{i=1}^{r-1} \in \mathbb{R}^d$. Note that $\boldsymbol{\theta}^*\cdot\Phi({\mathbf{z}}_i) = 0$. Now, pick ${\mathbf{a}} \in \mathbb{R}^d$ such that $\boldsymbol{\theta}^*\cdot\Phi({\mathbf{a}}) >0$ (assuming non-degeneracy). We note that $\curlybracket{({\mathbf{z}}_i,1)}_{i=1}^{r-1}\cup \curlybracket{({\mathbf{z}}_i,-1)}_{i=1}^{r-1}\cup \curlybracket{({\mathbf{a}},1)}$ forms a teaching set for the decision boundary corresponding to $\boldsymbol{\theta}^*$. Using similar ideas from the proof of \thmref{thm: linear perceptron main result}, we notice that any solution $\hat{\boldsymbol{\theta}}$ to \eqnref{eqn: objectkernel} satisfies $\boldsymbol{\theta}^*\cdot\Phi({\mathbf{z}}_i) = 0$ for the labelled datapoints corresponding to $\Phi({\mathbf{z}}_i)$. Thus, $\hat{\boldsymbol{\theta}}$ doesn't have any component along $\Phi({\mathbf{z}}_i)$. \eqnref{eqn: objectkernel} is minimized if $\hat{\boldsymbol{\theta}}\cdot\Phi({\mathbf{a}}) \ge 0$ implying $\hat{\boldsymbol{\theta}} = t\boldsymbol{\theta}^*$. Thus, under \assref{assumption: polyorthogonal}, we show an upper bound $\bigO{\binom{d+k-1}{k}}$ on the size of a teaching set for $\boldsymbol{\theta}^*$. \end{proof} \newpage \section{Conclusion} We have studied and extended the notion of teaching dimension for optimization-based perceptron learner. We also studied a more general notion of approximate teaching which encompasses the notion of exact teaching. To the best of our knowledge, our exact teaching dimension for linear and polynomial perceptron learner is new; so is the upper bound on the approximate teaching dimension of Gaussian perceptron learner and our analysis technique in general. There are many possible extensions to the present work. For example, one may extend our analysis to relaxing the assumptions imposed on the data distribution for polynomial and Gaussian perceptrons. This can potentially be achieved by analysing the linear perceptron and finding ways to nullify subspaces other than orthogonal vectors. This could enhance the results for both the exact teaching of polynomial perceptron learner to more general case and a tighter bound on the approximate teaching dimension of Gaussian perceptron learner. On the other hand, a natural extension of our work is to understand the approximate teaching complexity for other types of ERM learners, e.g. kernel SVM, kernel ridge, and kernel logistic regression. We believe the current work and its extensions would enrich our understanding of optimal and approximate teaching and enable novel applications. \subsubsection*{\bibname}} \begin{document} \twocolumn[ \aistatstitle{Instructions for Paper Submissions to AISTATS 2021} \aistatsauthor{ Author 1 \And Author 2 \And Author 3 } \aistatsaddress{ Institution 1 \And Institution 2 \And Institution 3 } ] \begin{abstract} The Abstract paragraph should be indented 0.25 inch (1.5 picas) on both left and right-hand margins. Use 10~point type, with a vertical spacing of 11~points. The \textbf{Abstract} heading must be centered, bold, and in point size 12. Two line spaces precede the Abstract. The Abstract must be limited to one paragraph. \end{abstract} \section{GENERAL FORMATTING INSTRUCTIONS} The camera-ready versions of the accepted papers are 8 pages, plus any additional pages needed for references. Papers are in 2 columns with the overall line width of 6.75~inches (41~picas). Each column is 3.25~inches wide (19.5~picas). The space between the columns is .25~inches wide (1.5~picas). The left margin is 0.88~inches (5.28~picas). Use 10~point type with a vertical spacing of 11~points. Please use US Letter size paper instead of A4. Paper title is 16~point, caps/lc, bold, centered between 2~horizontal rules. Top rule is 4~points thick and bottom rule is 1~point thick. Allow 1/4~inch space above and below title to rules. Author descriptions are center-justified, initial caps. The lead author is to be listed first (left-most), and the Co-authors are set to follow. If up to three authors, use a single row of author descriptions, each one center-justified, and all set side by side; with more authors or unusually long names or institutions, use more rows. Use one-half line space between paragraphs, with no indent. \section{FIRST LEVEL HEADINGS} First level headings are all caps, flush left, bold, and in point size 12. Use one line space before the first level heading and one-half line space after the first level heading. \subsection{Second Level Heading} Second level headings are initial caps, flush left, bold, and in point size 10. Use one line space before the second level heading and one-half line space after the second level heading. \subsubsection{Third Level Heading} Third level headings are flush left, initial caps, bold, and in point size 10. Use one line space before the third level heading and one-half line space after the third level heading. \paragraph{Fourth Level Heading} Fourth level headings must be flush left, initial caps, bold, and Roman type. Use one line space before the fourth level heading, and place the section text immediately after the heading with no line break, but an 11 point horizontal space. \subsection{Citations, Figure, References} \subsubsection{Citations in Text} Citations within the text should include the author's last name and year, e.g., (Cheesman, 1985). Be sure that the sentence reads correctly if the citation is deleted: e.g., instead of ``As described by (Cheesman, 1985), we first frobulate the widgets,'' write ``As described by Cheesman (1985), we first frobulate the widgets.'' The references listed at the end of the paper can follow any style as long as it is used consistently. \subsubsection{Footnotes} Indicate footnotes with a number\footnote{Sample of the first footnote.} in the text. Use 8 point type for footnotes. Place the footnotes at the bottom of the column in which their markers appear, continuing to the next column if required. Precede the footnote section of a column with a 0.5 point horizontal rule 1~inch (6~picas) long.\footnote{Sample of the second footnote.} \subsubsection{Figures} All artwork must be centered, neat, clean, and legible. All lines should be very dark for purposes of reproduction, and art work should not be hand-drawn. Figures may appear at the top of a column, at the top of a page spanning multiple columns, inline within a column, or with text wrapped around them, but the figure number and caption always appear immediately below the figure. Leave 2 line spaces between the figure and the caption. The figure caption is initial caps and each figure should be numbered consecutively. Make sure that the figure caption does not get separated from the figure. Leave extra white space at the bottom of the page rather than splitting the figure and figure caption. \begin{figure}[h] \vspace{.3in} \centerline{\fbox{This figure intentionally left non-blank}} \vspace{.3in} \caption{Sample Figure Caption} \end{figure} \subsubsection{Tables} All tables must be centered, neat, clean, and legible. Do not use hand-drawn tables. Table number and title always appear above the table. See Table~\ref{sample-table}. Use one line space before the table title, one line space after the table title, and one line space after the table. The table title must be initial caps and each table numbered consecutively. \begin{table}[h] \caption{Sample Table Title} \label{sample-table} \begin{center} \begin{tabular}{ll} \textbf{PART} &\textbf{DESCRIPTION} \\ \hline \\ Dendrite &Input terminal \\ Axon &Output terminal \\ Soma &Cell body (contains cell nucleus) \\ \end{tabular} \end{center} \end{table} \section{SUPPLEMENTARY MATERIAL} If you need to include additional appendices during submission, you can include them in the supplementary material file. You can submit a single file of additional supplementary material which may be either a pdf file (such as proof details) or a zip file for other formats/more files (such as code or videos). Note that reviewers are under no obligation to examine your supplementary material. If you have only one supplementary pdf file, please upload it as is; otherwise gather everything to the single zip file. You must use \texttt{aistats2021.sty} as a style file for your supplementary pdf file and follow the same formatting instructions as in the main paper. The only difference is that it must be in a \emph{single-column} format. You can use \texttt{supplement.tex} in our starter pack as a starting point. \section{SUBMISSION INSTRUCTIONS} To submit your paper to AISTATS 2021, please follow these instructions. \begin{enumerate} \item Download \texttt{aistats2021.sty}, \texttt{fancyhdr.sty}, and \texttt{sample\_paper.tex} provided in our starter pack. Please, do not modify the style files as this might result in a formatting violation. \item Use \texttt{sample\_paper.tex} as a starting point. \item Begin your document with \begin{flushleft} \texttt{\textbackslash documentclass[twoside]\{article\}}\\ \texttt{\textbackslash usepackage\{aistats2021\}} \end{flushleft} The \texttt{twoside} option for the class article allows the package \texttt{fancyhdr.sty} to include headings for even and odd numbered pages. \item When you are ready to submit the manuscript, compile the latex file to obtain the pdf file. \item Check that the content of your submission, \emph{excluding} references, is limited to \textbf{8 pages}. The number of pages containing references alone is not limited. \item Upload the PDF file along with other supplementary material files to the CMT website. \end{enumerate} \subsection{Camera-ready Papers} If your papers are accepted, you will need to submit the camera-ready version. Please make sure that you follow these instructions: \begin{enumerate} \item Change the beginning of your document to \begin{flushleft} \texttt{\textbackslash documentclass[twoside]\{article\}}\\ \texttt{\textbackslash usepackage[accepted]\{aistats2021\}} \end{flushleft} The option \texttt{accepted} for the package \texttt{aistats2021.sty} will write a copyright notice at the end of the first column of the first page. This option will also print headings for the paper. For the \emph{even} pages, the title of the paper will be used as heading and for \emph{odd} pages the author names will be used as heading. If the title of the paper is too long or the number of authors is too large, the style will print a warning message as heading. If this happens additional commands can be used to place as headings shorter versions of the title and the author names. This is explained in the next point. \item If you get warning messages as described above, then immediately after $\texttt{\textbackslash begin\{document\}}$, write \begin{flushleft} \texttt{\textbackslash runningtitle\{Provide here an alternative shorter version of the title of your paper\}}\\ \texttt{\textbackslash runningauthor\{Provide here the surnames of the authors of your paper, all separated by commas\}} \end{flushleft} Note that the text that appears as argument in \texttt{\textbackslash runningtitle} will be printed as a heading in the \emph{even} pages. The text that appears as argument in \texttt{\textbackslash runningauthor} will be printed as a heading in the \emph{odd} pages. If even the author surnames do not fit, it is acceptable to give a subset of author names followed by ``et al.'' \item The camera-ready versions of the accepted papers are 8 pages, plus any additional pages needed for references. \item If you need to include additional appendices, you can include them in the supplementary material file. \item Please, do not change the layout given by the above instructions and by the style file. \end{enumerate} \subsubsection*{Acknowledgements} All acknowledgments go at the end of the paper, including thanks to reviewers who gave useful comments, to colleagues who contributed to the ideas, and to funding agencies and corporate sponsors that provided financial support. To preserve the anonymity, please include acknowledgments \emph{only} in the camera-ready papers. \subsubsection*{References} References follow the acknowledgements. Use an unnumbered third level heading for the references section. Please use the same font size for references as for the body of the paper---remember that references do not count against your page length total. \subsubsection*{\bibname}} \begin{document} \onecolumn \aistatstitle{Instructions for Paper Submissions to AISTATS 2021: \\ Supplementary Materials} \section{FORMATTING INSTRUCTIONS} To prepare a supplementary pdf file, we ask the authors to use \texttt{aistats2021.sty} as a style file and to follow the same formatting instructions as in the main paper. The only difference is that the supplementary material must be in a \emph{single-column} format. You can use \texttt{supplement.tex} in our starter pack as a starting point, or append the supplementary content to the main paper and split the final PDF into two separate files. Note that reviewers are under no obligation to examine your supplementary material. \section{MISSING PROOFS} The supplementary materials may contain detailed proofs of the results that are missing in the main paper. \subsection{Proof of Lemma 3} \textit{In this section, we present the detailed proof of Lemma 3 and then [ ... ]} \section{ADDITIONAL EXPERIMENTS} If you have additional experimental results, you may include them in the supplementary materials. \subsection{The Effect of Regularization Parameter} \textit{Our algorithm depends on the regularization parameter $\lambda$. Figure 1 below illustrates the effect of this parameter on the performance of our algorithm. As we can see, [ ... ]} \vfill \end{document} \section{Introduction} Machine teaching studies the problem of finding an optimal training sequence to steer a learner towards a target concept \cite{DBLP:journals/corr/ZhuSingla18}. An important learning-theoretic complexity measure of machine teaching is the \emph{teaching dimension} \cite{goldman1995complexity}, which specifies the minimal number of training examples required in the worst case to teach a target concept. Over the past few decades, the notion of teaching dimension has been investigated under a variety of learner's models and teaching protocols (e.g,. \citet{cakmak2012algorithmic,singla2013actively,singla2014near,liu2017iterative,haug2018teaching,tschiatschek2019learner,DBLP:conf/icml/LiuDLLRS18,DBLP:conf/ijcai/KamalarubanDCS19,DBLP:conf/nips/Hunziker0AR0PYS19,DBLP:conf/ijcai/DevidzeMH0S20,DBLP:conf/icml/RakhshaRD0S20}). One of the most studied scenarios is the case of teaching a version-space learner \cite{goldman1995complexity,article:anthony95,zilles2008teaching,doliwa2014recursive,chen2018understanding,mansouri2019preference,pmlr-v98-kirkpatrick19a}. Upon receiving a sequence of training examples from the teacher, a version-space learner maintains a set of hypotheses that are consistent with the training examples, and outputs a \emph{random} hypothesis from this set. As a canonical example, consider teaching a 1-dimensional binary threshold function $f_{\theta^*}(x) = \id{x -\theta^*}$ for $x\in [0,1]$. For a learner with a finite (or countable infinite) version space, e.g., $\theta \in \{\frac{i}{n}\}_{i=0,\dots,n}$ where $n\in \mathbb{Z}^+$ (see \figref{fig:example.1d-vs}), a smallest training set is $\{\left(\frac{i}{n},0\right), \left(\frac{i+1}{n},1\right)\}$ where $\frac{i}{n} \leq \theta^* < \frac{i+1}{n}$; thus the teaching dimension is $2$. However, when the version space is continuous, the teaching dimension becomes $\infty$, because it is no longer possible for the learner to pick out a unique threshold $\theta^*$ with a finite training set. This is due to two key (limiting) modeling assumptions of the version-space learner: (1) all (consistent) hypotheses in the version space are treated equally, and (2) there exists a hypothesis in the version space that is consistent with all training examples. As one can see, these assumptions fail to capture the behavior of many modern learning algorithms, where the best hypotheses are often selected via \emph{optimizing} certain loss functions, and the data is not perfectly separable (i.e. not realizable w.r.t. the hypothesis/model class). To lift these modeling assumptions, a more realistic teaching scenario is to consider the learner as an \emph{empirical risk minimizer} (ERM). In fact, under the realizable setting, the version-space learner could be viewed as an ERM that optimizes the 0-1 loss---one that finds all hypotheses with zero training error. Recently, \citet{JMLR:v17:15-630} studied the teaching dimension of linear ERM, and established values of teaching dimension for several classes of linear (regularized) ERM learners, including support vector machine (SVM), logistic regression and ridge regression. As illustrated in \figref{fig:example.1d-hinge}, for the previous example it suffices to use $\{\left(\theta^*-\epsilon,0\right), \left(\theta^*+\epsilon,1\right)\}$ with any $\epsilon \leq \min(1-\theta^*, \theta^*)$ as training set to teach $\theta^*$ as an optimizer of the SVM objective (i.e., $l$2 regularized hinge loss); hence the teaching dimension is 2. In \figref{fig:example.1d-perceptron}, we consider teaching an ERM learner with perceptron loss, i.e., $\ell(f_\theta(x), y) = \max\left( -y\cdot (x-\theta), 0\right)$ (where $y \in \curlybracket{-1,1}$). If the teacher is allowed to construct \emph{any} training example with \emph{any} labeling\footnote{If the teacher is restricted to only provide consistent labels (i.e., the realizable setting), then the ERM with perceptron loss reduces to the version space learner, where the teaching dimension is $\infty$.} , then it is easy to verify that the minimal training set is $\{(\theta^*, -1), (\theta^*,1)\}$. \begin{figure}[t] \centering \begin{subfigure}[b]{.2\textwidth} \centering \includegraphics[trim={0, 0, 0, 10mm}, width=\linewidth]{fig/illu-threshold-finite.pdf} \vspace{-5mm} \caption{$\textsc{0/1}$ loss} \label{fig:example.1d-vs} \end{subfigure}\qquad \begin{subfigure}[b]{.2\textwidth} \centering \includegraphics[width=\linewidth]{fig/illu-threshold-svm.pdf} \vspace{-5mm} \caption{SVM (hinge loss)} \label{fig:example.1d-hinge} \end{subfigure}\\%\qquad \vspace{2mm} \begin{subfigure}[b]{.2\textwidth} \centering \includegraphics[width=\linewidth]{fig/illu-threshold-perceptron.pdf} \vspace{-5mm} \caption{Perceptron} \label{fig:example.1d-perceptron} \end{subfigure} \caption{Teaching a 1D threshold function to an ERM learner. Training instances are marked in grey. (a) Version-space learner with a finite hypothesis set. (b) SVM and training set $\{\left(\theta^*-\epsilon,0\right), \left(\theta^*+\epsilon,1\right)\}$. (c) ERM learner with (perceptron) loss and training set $\{(\theta^*, 0), (\theta^*,1)\}$. }\label{fig:illu-relaxed-general} \vspace{-3mm} \end{figure} While these results show promise at understanding optimal teaching for ERM learners, existing work \cite{JMLR:v17:15-630} has focused exclusively on the linear setting with the goal to teach the exact hypothesis (e.g., teaching the exact model parameters or the exact decision boundary for classification tasks). Aligned with these results, we establish an upper bound as shown in \secref{subsec.linear}. It remains a fundamental challenge to rigorously characterize the teaching complexity for nonlinear learners. Furthermore, in the cases where exact teaching is not possible with a finite training set, the classical teaching dimension no longer captures the fine-grained complexity of the teaching tasks, and hence one needs to relax the teaching goals and investigate new notions of teaching complexity. In this paper, we aim to address the above challenges. We focus on kernel perceptron, a specific type of ERM learner that is less understood even under the linear setting. Following the convention in teaching ERM learners, we consider the \emph{constructive} setting, where the teacher can construct arbitrary teaching examples in the support of the data distribution. Our contributions are highlighted below, with main theoretical results summarized in Table~\ref{tab:results-overview}. \begin{itemize} \item We formally define approximate teaching of kernel perceptron, and propose a novel measure of teaching complexity, namely the \emph{$\epsilon$-approximate teaching dimension} ($\epsilon$-TD), which captures the complexity of teaching a ``relaxed'' target that is close to the target hypothesis in terms of the expected risk. Our relaxed notion of teaching dimension strictly generalizes the teaching dimension of \citet{JMLR:v17:15-630}, where it trades off the teaching complexity against the risk of the taught hypothesis, and hence is more practical in characterizing the complexity of a teaching task (\secref{sec.statement}).\newline \item We show that exact teaching is feasible for kernel perceptrons with finite dimensional feature maps, such as linear kernel and polynomial kernel. Specifically, for data points in $\mathbb{R}^d$, we establish a $\bigTheta{d}$ bound on the teaching dimension of linear perceptron. Under a mild condition on data distribution, we provide a tight bound of $\bigTheta{\binom{d+k-1}{k}}$ for polynomial perceptron of order $k$. We also exhibit optimal training sets that match these teaching dimensions (\secref{subsec.linear} and \secref{subsec.poly}).\newline \item We further show that for Gaussian kernelized perceptron, exact teaching is not possible with a finite set of hypotheses, and then establish a $d^{\bigO{\log^2 \frac{1}{\epsilon}}}$ bound on the $\epsilon$-approximate teaching dimension (\secref{subsec.gaussiankernel}). To the best of our knowledge, these results constitute the first known bounds on (approximately) teaching a non-linear ERM learner (\secref{sec.theoreticalresults}). \end{itemize} \begin{table}[t!] \centering \scalebox{.88}{ \begin{tabular}{cccc} \toprule & \textbf{linear} & \textbf{polynomial} & \textbf{Gaussian} \\ \midrule TD (exact) & $\bigTheta{d}$ & $\bigTheta{\binom{d+k-1}{k}}$ & $\infty$ \\ $\epsilon$-approximate TD & - & - & $d^{\bigO{\log^2 \frac{1}{\epsilon}}}$ \\ \textbf{Assumption } & - & \ref{assumption: polyorthogonal} & \ref{assumption: orthogonal}, \ref{assumption: bounded cone}\\ \bottomrule \end{tabular} } \caption{Teaching dimension for kernel perceptron}\label{tab:results-overview} \vspace{-3mm} \end{table} \section{Motivation for Assumptions}\label{appendix: motivation} In this appendix, we discuss the motivations and insights for the key \assumref{assumption: polyorthogonal}{assumption: orthogonal}{assumption: bounded cone} made in \secref{subsec.poly} and \secref{subsec.gaussiankernel}. This appendix is organized in the following way: \appref{appsubsec.limitation} discusses \assref{assumption: polyorthogonal} and provides the proofs of \lemref{lemma: exact teaching} and \lemref{lemma: approximate teaching } in the context of polynomial kernel (see \secref{subsec.poly}); \appref{appsubsec.approxassumtions} discusses the \assref{assumption: orthogonal} and \assref{assumption: bounded cone} in the context of Gaussian kernel perceptron (see \secref{subsec.gaussiankernel}). \paragraph{Reformulation of a model $\boldsymbol{\theta}$ as a polynomial form} As noted in \secref{sec.statement}, we consider the reproducing kernel Hilbert space~\cite{learnkernel} ${\mathcal{H}}_{{\mathcal{K}}}$ which could be spanned by the linear combinations of kernel functions of the form ${\mathcal{K}}({\mathbf{x}}, \cdot)$. More concretely, \begin{equation} {\mathcal{H}}_{{\mathcal{K}}} = \condcurlybracket{\sum_{i=1}^m \alpha_i\cdot{\mathcal{K}}({\mathbf{x}}_i,\cdot)}{ m \in \mathbb{N},\, {\mathbf{x}}_i \in {\mathcal{X}},\, \alpha_i \in \ensuremath{\mathbb{R}}, i = 1,\cdots,m}\nonumber \end{equation} Thus, we could write any model $f_{\boldsymbol{\theta}} \in {\mathcal{H}}_{{\mathcal{K}}}$ (parametrized by $\boldsymbol{\theta}$) as $\sum_{i=1}^n \alpha_i\cdot{\mathcal{K}}({\mathbf{x}}_i,\cdot)$ for some $n \in \mathbb{N}$, $ {\mathbf{x}}_i \in {\mathcal{X}}$ for $i \in \bracket{n}$. This interesting because if ${\mathcal{K}}(\cdot,\cdot)$ is a polynomial kernel of degree $k$, then \begin{equation} f_{\boldsymbol{\theta}}({\mathbf{x}}) = \sum_{i=1}^n \alpha_i\cdot{\mathcal{K}}({\mathbf{x}}_i,{\mathbf{x}}) = \sum_{i=1}^n \alpha_i\cdot \normg{{\mathbf{x}}_i}{{\mathbf{x}}}^k = \sum_{i=1}^n \alpha_i\cdot \paren{{\mathbf{x}}_{i1}{\mathbf{x}}_1+\cdots+{\mathbf{x}}_{id}{\mathbf{x}}_d}^k \label{eqn: poly form} \end{equation} where ${\mathbf{x}}_i = \paren{{\mathbf{x}}_{i1},\cdots,{\mathbf{x}}_{id}}$. Thus, $f_{\boldsymbol{\theta}}(\cdot)$ could be reformulated as a homogeneous polynomial of degree $k$ in $d$ variables. Notice that for polynomial kernel in \secref{subsec.poly}, for a target model $\boldsymbol{\theta}^*$ we study the orthogonal projections of the form $\Phi({\mathbf{x}})$ for ${\mathbf{x}} \in {\mathcal{X}}$ such that $\boldsymbol{\theta}^*\cdot\Phi({\mathbf{x}}) = 0$. Alternatively, using \eqnref{eqn: poly form} we wish to solve the polynomial equation: \begin{equation} f_{\boldsymbol{\theta}^*}({\mathbf{x}}) = 0 \implies \sum_{i=1}^n \alpha_i\cdot \paren{{\mathbf{x}}_{i1}{\mathbf{x}}_1+\cdots+{\mathbf{x}}_{id}{\mathbf{x}}_d}^k = 0 \nonumber \end{equation} where we denote $\boldsymbol{\theta}^* := \sum_{i=1}^n \alpha_i\cdot\Phi({\mathbf{x}}_i)$. For \assref{assumption: polyorthogonal} we wish to find $\binom{d+k-1}{k}$ real solutions of the form ${\mathbf{x}}' \in \ensuremath{\mathbb{R}}^d$, of this equation which are linearly independent. It is well-studied in the literature of polynomial algebra that this equation might not satisfy the required assumption. We construct one such model for the proof of \lemref{lemma: approximate teaching }. This reformulation can be extended for sum of polynomial kernels of the form $\sum_{j=1}^s c_j\cdot{\normg{{\mathbf{x}}}{{\mathbf{x}}'}}^j$ where $c_j \ge 0$. In \assref{assumption: orthogonal} the reformulation reduces to a variant of the above polynomial equation i.e. \begin{equation*} \sum_{i=1}^n\alpha_i\paren{\sum_{j=1}^s c_i\cdot \paren{{\mathbf{x}}_{i1}{\mathbf{x}}_1+\cdots+{\mathbf{x}}_{id}{\mathbf{x}}_d}^j} = 0 \end{equation*} So far, we discussed a characterization of the notion of orthogonality for a target model $\boldsymbol{\theta}^*$ in the form of a polynomial equation. This characterization would help us understand \assref{assumption: polyorthogonal} and \assref{assumption: orthogonal}. In \secref{appsubsec.limitation}, we discuss that \assref{assumption: polyorthogonal} is the most natural step to make for exact teaching of a target model. \subsection{ Limitation of Exact Teaching: Polynomial Kernel Perceptron}\label{appsubsec.limitation} In this subsection, we provide the proofs of \lemref{lemma: exact teaching} and \lemref{lemma: approximate teaching } as stated in \secref{subsec.motivation}. These results establish that in the realizable setting, \assref{assumption: polyorthogonal} is required for exact teaching: \lemref{lemma: exact teaching}. Furthermore, there are pathological cases where violation of the assumption leads to models which couldn't be approximately taught: \lemref{lemma: approximate teaching }. \begin{proof}[Proof of \lemref{lemma: exact teaching}] We would prove the result by contradiction. Assume that $\mathcal{TS}_{\boldsymbol{\theta}^*}$ be a teaching set which \tt{exactly} teaches $\boldsymbol{\theta}^*$. $\mathrm{WLOG}$ we enumerate the teaching set as $\mathcal{TS}_{\boldsymbol{\theta}^*} = \curlybracket{\paren{{\mathbf{x}}_1,y_1},\cdots,\paren{{\mathbf{x}}_n,y_n}}$. For the sake of clarity, we would rewrite \eqref{eqn: objectkernel} again \begin{equation} {\boldsymbol{\mathcal{A}}}_{opt}(\mathcal{TS}_{\boldsymbol{\theta}^*}):= \mathop{\rm arg\,min}_{\boldsymbol{\theta} \in {\mathcal{H}}_{{\mathcal{K}}}}\sum_{i=1}^n \max(-y_i\cdot \boldsymbol{\theta}^*\cdot\Phi({\mathbf{x}}_i), 0)\label{eq: polyobject} \end{equation} Denote by $\mathcal{V}^{\bot}_{\boldsymbol{\theta}^*} \subset {\mathcal{H}}_k$ the orthogonal subspace of $\boldsymbol{\theta}^*$. We denote the objective value of \eqnref{eq: polyobject} by ${\mathbf{p}}(\boldsymbol{\theta}) := \sum_{i = 1}^{n} \max(-y_i\cdot\boldsymbol{\theta}\cdot\Phi({\mathbf{x}}_i),\: 0)$. We further define \tt{effective direction of a teaching point} $\paren{{\mathbf{x}}_i,y_i} \in \mathcal{TS}_{\boldsymbol{\theta}^*}$ in the RKHS ${\mathcal{H}}_{{\mathcal{K}}}$ as $\boldsymbol{d}_i := -y_i\cdot\Phi({\mathbf{x}}_i)$. Because of the realizable setting i.e. all teaching points are correctly classified, it is clear that $$-y_i\cdot \boldsymbol{\theta}^*\cdot\Phi({\mathbf{x}}_i) \le 0 \implies \boldsymbol{\theta}^*\cdot \boldsymbol{d}_i \le 0.$$ Since $\boldsymbol{\theta}^*$ violates \assref{assumption: polyorthogonal}, thus $\exists$ a unit normalized direction $\hat{\boldsymbol{d}} \in \mathcal{V}^{\bot}_{\boldsymbol{\theta}^*}$ which can't be spanned by ${\mathcal{S}}_{0} \triangleq \condcurlybracket{\Phi({\mathbf{x}})}{ \Phi({\mathbf{x}}) \in \mathcal{V}^{\bot}_{\boldsymbol{\theta}^*}\,\, \textnormal{for some}\,\, {\mathbf{x}} \in {\mathcal{X}}}$ such that $\hat{\boldsymbol{d}} \perp \mathbf{span}\left\langle{\mathcal{S}}_{0}\right\rangle$. Now, we would show that $\exists \lambda > 0$ (real) such that \begin{equation} \paren{\boldsymbol{\theta}^* + \lambda\hat{\boldsymbol{d}}} \in {\boldsymbol{\mathcal{A}}}_{opt}(\mathcal{TS}_{\boldsymbol{\theta}^*}) \label{eqn: new theta} \end{equation} Notice that for some $i$ if $\boldsymbol{d}_i \in \mathcal{V}^{\bot}_{\boldsymbol{\theta}^*}$ then $(\boldsymbol{\theta}^* + \lambda\hat{\boldsymbol{d}})\cdot \boldsymbol{d}_i = \boldsymbol{\theta}^*\cdot \boldsymbol{d}_i \le 0$. Now, we consider the case when $\boldsymbol{d}_i \notin \mathcal{V}^{\bot}_{\boldsymbol{\theta}^*}$. We could expand $\boldsymbol{d}_i$ as follows: \begin{equation} \boldsymbol{d}_i = a_i\hat{\boldsymbol{d}}^{\perp} + b_i\hat{\boldsymbol{d}} \label{eqn: expand} \end{equation} where $a_i$ and $b_i$ are real scalars and $\hat{\boldsymbol{d}}^{\perp}$ is normalized orthogonal projection of $\boldsymbol{d}_i$ to orthogonal complement (orthogonal subspace) of $\hat{\boldsymbol{d}}$. These constructions are illustrated in \figref{fig:lemma1}. Now, we would compute the following dot product: \begin{align} &(\boldsymbol{\theta}^* + \lambda\hat{\boldsymbol{d}})\cdot \boldsymbol{d}_i \nonumber\\ \implies& \boldsymbol{\theta}^*\cdot \boldsymbol{d}_i + \lambda\hat{\boldsymbol{d}}\cdot (a_i\hat{\boldsymbol{d}}^{\perp} + b_i\hat{\boldsymbol{d}}) \label{eqn: part1}\\ \implies& \boldsymbol{\theta}^*\cdot\boldsymbol{d}_i + \lambda\hat{\boldsymbol{d}}\cdot b_i\hat{\boldsymbol{d}} \label{eqn: part2} \end{align} \eqnref{eqn: part1} follows using \eqnref{eqn: expand}. In \eqnref{eqn: part2} we note that $\hat{\boldsymbol{d}} \perp \hat{\boldsymbol{d}}^{\perp}$. If $b_i \le 0$ then $(\boldsymbol{\theta}^* + \lambda\hat{\boldsymbol{d}})\cdot \boldsymbol{d}_i \le 0$ as $\boldsymbol{\theta}^*\cdot \boldsymbol{d}_i < 0$. If $b_i > 0$, then to ensure $(\boldsymbol{\theta}^* + \lambda\hat{\boldsymbol{d}})\cdot \boldsymbol{d}_i \le 0$, using \eqnref{eqn: part2} we need $$\lambda \le \frac{-\boldsymbol{\theta}^*\cdot \boldsymbol{d}_i}{b_i}$$ Since, $i$ is chosen arbitrarily thus for all the effective directions $\boldsymbol{d}_i \notin \mathcal{V}^{\bot}_{\boldsymbol{\theta}^*}$ where $b_i > 0$, we pick postive scalar $\lambda$ such that: $$\lambda := \min_{i:\, b_i > 0}\frac{-\boldsymbol{\theta}^*\cdot \boldsymbol{d}_i}{b_i}$$ For this choice of $\lambda$ we show that $\boldsymbol{\theta}^* + \lambda\hat{\boldsymbol{d}} \in {\boldsymbol{\mathcal{A}}}_{opt}(\mathcal{TS}_{\boldsymbol{\theta}^*})$. Thus, by definition, $\mathcal{TS}_{\boldsymbol{\theta}^*}$ as stated above can't teach $\boldsymbol{\theta}^*$ exactly. Hence, if $\boldsymbol{\theta}^*$ violates \assref{assumption: polyorthogonal} then we can't teach it exactly in the realizable setting. \end{proof} \begin{figure*}[t] \centering \begin{subfigure}[b]{0.30\textwidth} \centering \includegraphics[width=\linewidth]{fig/lemma1.png} \caption{} \label{fig:lemma1} \end{subfigure} \qquad \begin{subfigure}[b]{0.40\textwidth} \centering \includegraphics[width=\linewidth]{fig/lemma2_plot1.png} \caption{} \label{fig:lemma2} \end{subfigure} \caption{Illustrations for Proofs of \lemref{lemma: exact teaching} and \lemref{lemma: approximate teaching }. (a) For \lemref{lemma: exact teaching}, consider $\boldsymbol{\theta}^*$ as shown. $\mathbf{span}\left\langle{\mathcal{S}}_{0}\right\rangle$ only covers the direction indicated by the dashed blue arrow. $\Tilde{\boldsymbol{\theta}}$ correctly labels not only teaching examples on $\mathcal{V}^{\bot}_{\boldsymbol{\theta}^*}$ and within $\mathbf{span}\left\langle{\mathcal{S}}_{0}\right\rangle$, but also those not on $\mathcal{V}^{\bot}_{\boldsymbol{\theta}^*}$, e.g. along effective directions $d_1$, $d_2$, $d_3$; (b) To visualize the proof idea of \lemref{lemma: approximate teaching }, we demonstrate an example in $\ensuremath{\mathbb{R}}^2$ with feature space of dimension 3 (where $k = 2$). We consider a model $\boldsymbol{\theta}^* = \frac{1}{\sqrt{2}}\cdot\Phi((1,0))+\frac{1}{\sqrt{2}}\cdot\Phi((0,1))$. Since $k$ is even, each point $\textbf{x}$ in $\mathbb{R}^2$ corresponds to a non-negative value of $f_{\boldsymbol{\theta}^*}(\textbf{x})$. This function is plotted as the blue surface in (b) along $z$-axis. Yellow surface represents a threshold of $\epsilon$ along $z$-axis; thus any point above it has value more than $\epsilon$. A red $\delta$-norm ring on the $\ensuremath{\mathbb{R}}^2$-plane denotes the constraint on the norm of a teaching point ($||\Phi({\mathbf{x}})|| = \delta \implies ||{\mathbf{x}}|| = \delta^{1/4}$). If we are constrained to select only points outside of the red $\delta$-norm ring, then the plot illustrates a situation where no points outside the ring satisfies $f_{\boldsymbol{\theta}^*}(\textbf{x}) = \boldsymbol{\theta}^*\cdot\Phi({\mathbf{x}}) < \epsilon$.} \label{fig:lemma1-2} \end{figure*} Now, we provide the proof of \lemref{lemma: approximate teaching } for which we give a construction of a model $\boldsymbol{\theta}^*$ which violates \assref{assumption: polyorthogonal} and also show that it can't be taught arbitrarily $\epsilon$-close approximately. The proof is also illustrated in \figref{fig:lemma2}. \begin{proof}[Proof of \lemref{lemma: approximate teaching }] Assume $\boldsymbol{\theta}^*$ is a target model which violates \assref{assumption: polyorthogonal}. If $\boldsymbol{\theta}^*$ can be taught \tt{approximately} for arbitrarily small $\epsilon > 0$ then $\exists$ $\Tilde{\boldsymbol{\theta}}^*$ which can be taught \tt{exactly} (i.e. satisfies \assref{assumption: polyorthogonal}) such that $$\boldsymbol{\theta}^*\cdot \Tilde{\boldsymbol{\theta}}^* \ge 1-\cos{a_{\epsilon}},\,\, \textnormal{where}\,\, \cos{a_{\epsilon}} = \epsilon$$ if $\boldsymbol{\theta}^*$ and $\Tilde{\boldsymbol{\theta}}^*$ are unit normalized. This implies that if $\Phi({\mathbf{x}}) \in \mathcal{V}_{\Tilde{\boldsymbol{\theta}}^*}^{\perp} \subset {\mathcal{H}}_{{\mathcal{K}}}$ (orthogonal complement of $\Tilde{\boldsymbol{\theta}}^*$) such that $||\Phi({\mathbf{x}})|| \le 1$ then the following holds: \begin{equation} |\boldsymbol{\theta}^*\cdot \Phi({\mathbf{x}})| \le \epsilon \label{eqn: almost orth} \end{equation} Alternatively, we could think of $\Phi({\mathbf{x}})$ as being almost orthogonal to $\boldsymbol{\theta}^*$. Now, we would show a construction of a target model when $k$ has parity even, which not only violates \assref{assumption: polyorthogonal} but can't be taught \tt{approximately} such that \eqnref{eqn: almost orth} holds. The idea is to find $\boldsymbol{\theta}^*$ which doesn't have almost orthogonal projections in ${\mathcal{H}}_{{\mathcal{K}}}$ with norm lower-bounded by $\delta$. Consider the following construction for a target model $\boldsymbol{\theta}^* \in {\mathcal{H}}_{{\mathcal{K}}}$: \begin{equation} \boldsymbol{\theta}^* = \sum_{i=1}^d \frac{1}{\sqrt{d}}\cdot \Phi({\mathbf{e}}_i),\quad ||\boldsymbol{\theta}^*|| = 1 \label{eqn: thetaconstruction1} \end{equation} where $\{{\mathbf{e}}_i\}$'s form the standard basis in $\ensuremath{\mathbb{R}}^d$. Notice that for any ${\mathbf{x}} \in {\mathcal{X}}$, \begin{equation} \boldsymbol{\theta}^*\cdot \Phi({\mathbf{x}}) = \sum_{i=1}^d \frac{1}{\sqrt{d}}\cdot \Phi({\mathbf{e}}_i)\cdot\Phi({\mathbf{x}}) = \sum_{i=1}^d \frac{1}{\sqrt{d}}\cdot {\mathbf{x}}_i^k. \label{eqn: thetaconstruction2} \end{equation} RHS of the above equation is zero only when all the ${\mathbf{x}}_i^k = 0$ since $k$ is even. Thus, the only projection orthogonal to $\boldsymbol{\theta}^*$ is the zero projection in ${\mathcal{H}}_{{\mathcal{K}}}$, thus violates \assref{assumption: polyorthogonal}. Now, we show that $\boldsymbol{\theta}^*$ as constructed in \eqnref{eqn: thetaconstruction1} can't be taught approximately for arbitrarily small $\epsilon > 0$. If ${\mathbf{x}} \in {\mathcal{X}}$ is such that $||\Phi({\mathbf{x}})|| \ge \delta$, then using H\"{o}lder's inequality: \begin{align} \sum_{i=1}^d {\mathbf{x}}_i^2 \le \sum_{i=1}^d 1\cdot {\mathbf{x}}_i^2 \le d^{\frac{k-2}{k}}\paren{\sum_{i=1}^d \paren{{\mathbf{x}}_i^2}^{\frac{k}{2}}}^{\frac{2}{k}} \nonumber \end{align} But we have $\sum_{i=1}^d {\mathbf{x}}_i^2 \ge \delta^{\frac{2}{k}}$. Thus, using \eqnref{eqn: thetaconstruction1}-\eqnref{eqn: thetaconstruction2} \begin{equation} \paren{\sum_{i=1}^d {\mathbf{x}}_i^k} \ge \frac{\delta}{d^{2(k-2)}} \implies \boldsymbol{\theta}^*\cdot\Phi({\mathbf{x}}) \ge \frac{\delta}{d^{2k-\frac{7}{2}}} \nonumber \end{equation} Thus, if $\epsilon < \frac{\delta}{d^{2k-\frac{7}{2}}}$ then $\boldsymbol{\theta}^*\cdot\Phi({\mathbf{x}}) > \epsilon$. This implies that $\Phi({\mathbf{x}})$ can't be chosen almost orthogonal to $\boldsymbol{\theta}^*$ violating \eqnref{eqn: almost orth}. Hence, $\nexists$ $\Tilde{\boldsymbol{\theta}}^*$ arbitrarily close to $\boldsymbol{\theta}^*$ which can be taught exactly. Thus, the construction of $\boldsymbol{\theta}^*$ in \eqnref{eqn: thetaconstruction1} violates \assref{assumption: polyorthogonal} and can't be taught approximately for arbitrarily small $\epsilon > 0$. \end{proof} \paragraph{Is the assumption of lower bound $\delta$ restrictive?} Now, we would argue that the assumption of a lower bound on the norm of the teaching point for \lemref{lemma: approximate teaching } is only for analysis of the proof presented above. Consider the target model $\boldsymbol{\theta}^*$ constructed in \eqnref{eqn: thetaconstruction1}. Consider that $\exists$ $\Tilde{\boldsymbol{\theta}}^*$ which can be taught exactly using arbitrarily small normed teaching points (i.e. lower bound of $\delta$ is violated) such that $$\boldsymbol{\theta}^*\cdot \Tilde{\boldsymbol{\theta}}^* \ge 1-\cos{a_{\epsilon}},\,\, \textnormal{where}\,\, \cos{a_{\epsilon}} = \epsilon$$ for arbitrarily small $\epsilon > 0$. Define the teaching set as $\mathcal{TS}_{\Tilde{\boldsymbol{\theta}}^*}$. But, even if we unit-normalize all the teaching points, call the normalized set $\mathcal{TS}^{unit}_{\Tilde{\boldsymbol{\theta}}^*}$, \eqnref{eqn: objectkernel} is still satisfied. Since in that case for any $({\mathbf{x}}_i,y_i) \in \mathcal{TS}^{unit}_{\Tilde{\boldsymbol{\theta}}^*}$, \eqnref{eqn: almost orth} is violated. Hence, violating the assumption of lower bound on the norm of the teaching points doesn't invalidate the claim of \lemref{lemma: approximate teaching }. \subsection{Approximate Teaching: \assref{assumption: orthogonal} and \assref{assumption: bounded cone}}\label{appsubsec.approxassumtions} As noted in \secref{subsec.gaussiankernel}, the teaching dimension of a Gaussian kernel perceptron learner is $\infty$. This calls for studying these non-linear kernel in the setting of approximate teaching. Inspired by our discussion in the previous subsection, we argue that the underlying assumptions: \assref{assumption: orthogonal} and \assref{assumption: bounded cone} are fairly mild in order to establish strong results stated in \thmref{thm: boundedclassifier} and \thmref{thm: gaussian_main_thm} (cf. \secref{subsec.gaussiankernel}). This appendix subsection is divided into two paragraphs corresponding to the assumptions as follows: \paragraph{Existence of orthogonal linear independent projections: \assref{assumption: orthogonal}.} Notice that the projected polynomial space or the approximated kernel $\Tilde{{\mathcal{K}}}$ is a sum of polynomial kernels. We rewrite \eqnref{eqn: eqn17} for ease of clarity: \begin{equation*} \Tilde{{\mathcal{K}}}({\mathbf{x}}, {\mathbf{x}}') = \mathbi{e}^{-\frac{||{\mathbf{x}}||^2}{2\sigma^2}}\mathbi{e}^{-\frac{||{\mathbf{x}}'||^2}{2\sigma^2}}\sum_{k=0}^{s}\frac{1}{k!}\paren{\frac{\normg{{\mathbf{x}}}{{\mathbf{x}}'}}{\sigma^2}}^k \end{equation*} If we replace $z = \frac{\normg{{\mathbf{x}}}{{\mathbf{x}}'}}{\sigma^2}$, we could write \begin{equation*} \Tilde{{\mathcal{K}}}({\mathbf{x}}, {\mathbf{x}}') = \mathbi{e}^{-\frac{||{\mathbf{x}}||^2}{2\sigma^2}}\mathbi{e}^{-\frac{||{\mathbf{x}}'||^2}{2\sigma^2}}\sum_{k=0}^{s}\frac{1}{k!}\cdot z^k \end{equation*} Since all the coefficients of the polynomial $\sum_{k=0}^{s}\frac{1}{k!}\cdot z^k$ are positive thus if $s$ is even then $\Tilde{{\mathcal{K}}}({\mathbf{x}}, {\mathbf{x}}') > 0$. Thus, if $\boldsymbol{\theta}^* = \sum_{i=1}^l \alpha_i\cdot{\mathcal{K}}({\mathbf{a}}_i, \cdot)$ for some $\{{\mathbf{a}}_i\}_{i=1}^l \subset {\mathcal{X}}$ such that $\alpha_i$'s are positive then \assref{assumption: orthogonal} would be violated. Hence, there is a class of $r$ values for which the assumption would be violated. It is straight-forward to note that \lemref{lemma: exact teaching} could be extended to sum of polynomial kernels. Similar extension for \lemref{lemma: approximate teaching } when the highest degree is of parity even could be established. These results follow by noting the polynomial space of homogeneous polynomials of degree $k$ in $d$ variables is isomorphic \cite{article} to the polynomial space of degree $k$ in $(d-1)$ variables. Since Hilbert space $\Tilde{{\mathcal{K}}}$ is a sum of polynomial kernels thus the extended results hold. This implies that there could be pathological cases where $\mathbb{P}\boldsymbol{\theta}^*$ could not be learnt approximately in $\Tilde{{\mathcal{K}}}$. But this poses a problem because most of the information of a model in terms of the eigenvalues of the orthogonal basis of the Gaussian kernel is contained in the starting indices i.e. $\forall k \le s,\quad \Phi_{k,\boldsymbol{\lambda}}({\mathbf{x}}) = \mathbi{e}^{-\frac{||{\mathbf{x}}||^2}{2\sigma^2}}\cdot \frac{\sqrt{{\mathcal{C}}^k_{\boldsymbol{\lambda}}}}{\sqrt{k!}\sigma^k}\cdot{\mathbf{x}}^{\boldsymbol{\lambda}} $ where $\sum_{i=1}^d\boldsymbol{\lambda}_i = k$. It has been discussed in \appref{appendix: gaussian perceptron}. Since the fixed Hilbert space induced by $\Tilde{{\mathcal{K}}}$ is spanned by these truncated projections, thus \assref{assumption: orthogonal} gives a characterization for approximately teachable models. It is left to be understood if there is a more unified characterization which could incorporate approximately teachable models beyond \assref{assumption: orthogonal}. \paragraph{Boundedness of weights: \assref{assumption: bounded cone}.} It is fairly natural in the sense that in \thmref{thm: boundedclassifier} we are bounding (approximating) the error values of the function point-wise i.e. $f^*$ (using $\hat{f}$) for a fixed $\epsilon$. If for some $\hat{\boldsymbol{\theta}} \in \mathcal{A}_{opt}$, $\hat{f}$ ( $= \hat{\boldsymbol{\theta}}\cdot\Phi(\cdot)$) is unboundedly sensitive to some teaching (training) point, then bounding error becomes stringent. Further, we show that there exists a unique solution up to a positive constant scaling to \eqnref{eqn: bounded} which satisfy the assumption in \appref{appendixsub: solutionexists}. The weights $\{\alpha_i\}_{i=1}^l$ and $[\beta_0,\gamma]$ could be thought of as Lagrange multipliers for Gaussian kernel perceptron. Boundedness of the multipliers is a well-studied problem in the mathematical programming and optimization literature \cite{gauvin,nooshin,dutta,locallipschitz}. Interestingly, \citet{luksan} demonstrated the importance of the boundedness of the Lagrange multipliers for the study of interior point methods for non-linear programming. On the other hand, \assref{assumption: bounded cone} as a regularity condition provides new insights into solving problems where task is to universally approximate the underlying functions as discussed in the proof of \thmref{thm: boundedclassifier} in \appref{appendixsub: proofofmainthm}. \newpage \section{Teaching Dimension for Kernel Perceptron}\label{sec.theoreticalresults} \vspace{-1mm} In this section, we study the generic problem of teaching kernel perceptrons in three different settings:\,1) linear (in \secref{subsec.linear}); 2)\, polynomial (in \secref{subsec.poly}); and Gaussian (in \secref{subsec.gaussiankernel}). Before establishing our main result for Gaussian kernelized perceptrons, we first introduce two important results for linear and polynomial perceptrons inherently connected to the Gaussian perceptron. Our proofs are inspired by ideas from linear algebra and projective geometry as detailed \iftoggle{longversion}{in \appref{appendix:table-of-contents}}{in the supplemental materials}. \vspace{-1mm} \subsection{Homogeneous Linear Perceptron}\label{subsec.linear} In this subsection, we study the problem of teaching a linear perceptron. First, we consider an optimization problem similar to \eqnref{eqn: objectmain} as shown in \citet{JMLR:v17:15-630}:\vspace{-1mm} \begin{equation} {\boldsymbol{\mathcal{A}}}_{opt} := \mathop{\rm arg\,min}_{\boldsymbol{\theta} \in \mathbb{R}^d} \sum_{i = 1}^n \ell(\boldsymbol{\theta}\cdot{{\mathbf{x}}}_i, y_i) + \frac{\lambda}{2}||\boldsymbol{\theta}||^2_{A} \label{eqn: eqn1} \looseness -3 \end{equation} where $\ell(\cdot,\cdot)$ is a convex loss function, $A$ is a positive semi-definite matrix, $||\boldsymbol{\theta}||_{A}$ is defined as $\sqrt{\boldsymbol{\theta}^\top A \boldsymbol{\theta}}$, and $\lambda > 0$. For convex loss function $\ell(\cdot,\cdot)$, Theorem 1~\cite{JMLR:v17:15-630} established a degree-of-freedom lower bound on the number of training items to obtain a unique solution $\boldsymbol{\theta}^*$. Since, the loss function for linear perceptron is convex thus we immediately obtain a lower bound on the teaching dimension as follows: \begin{corollary}\label{cor: linear lower bound} If $A = 0$ and $\lambda = 1$, then \eqnref{eqn: objectmain} can be solved as \eqnref{eqn: eqn1}. Moreover, teaching dimension for decision boundary corresponding to a target model $\boldsymbol{\theta}^*$ is lower-bounded by $\bigOmega{d}$. \end{corollary} Now, we would establish an upper bound on $TD({\boldsymbol{\mathcal{A}}}_{opt},\boldsymbol{\theta}^*)$ for exact teaching of the decision boundary of a target model $\boldsymbol{\theta}^*$. The key idea is to find a set of points which span the orthogonal subspace of $\boldsymbol{\theta}^*$, which we use to force a solution $\hat{\boldsymbol{\theta}} \in {\boldsymbol{\mathcal{A}}}_{opt}$ such that it has a component only along $\boldsymbol{\theta}^*$. Formally, we state the claim of the result with proof as follows: \begin{theorem}\label{thm: linear perceptron main result} Given any target model $\boldsymbol{\theta}^*$, for solving \eqnref{eqn: objectmain} the teaching dimension for the decision boundary corresponding to $\boldsymbol{\theta}^*$ is $\bigTheta{d}$. The following is a teaching set: \begin{align*} {{\mathbf{x}}}_i = {\mathbf{v}}_i,\quad y_i = 1\quad \forall\; i\; \in\; [d-1];\qquad\qquad\qquad\ \ \,\\ \quad {{\mathbf{x}}}_d = -\sum_{i=1}^{d-1} {\mathbf{v}}_i,\quad y_d = 1;\quad {{\mathbf{x}}}_{d+1} = \boldsymbol{\theta}^*,\quad y_{d+1} = 1 \end{align*} where $\{{\mathbf{v}}_i\}_{i=1}^{d}$ is an orthogonal basis for $\mathbb{R}^d$ which extends with ${\mathbf{v}}_d = \boldsymbol{\theta}^*$. \vspace{-1mm} \end{theorem} \begin{proof} Using \corref{cor: linear lower bound}, the lower bound for solving \eqnref{eqn: objectmain} is immediate. Thus, if we show that the mentioned labeled set of training points form a teaching set, then we can show an upper bound which would imply a tight bound of $\bigTheta{d}$ on the teaching dimension for finding the decision boundary. Denote the set of labeled data points as ${\mathcal{D}}$. Denote by ${\mathbf{p}}(\boldsymbol{\theta}) := \sum_{i = 1}^{d+1} \max(-y_i\cdot\boldsymbol{\theta}\cdot{{\mathbf{x}}}_i,\: 0)$. Since $\{{\mathbf{v}}_i\}_{i=1}^{d}$ is an orthogonal basis, thus $\forall \, i \in [d-1]\quad {\mathbf{v}}_i\cdot \boldsymbol{\theta}^* = 0$, thus it is not very difficult to show that ${\mathbf{p}}(t\boldsymbol{\theta}^*) = 0$ for some positive scalar $t$. Note, if $\hat{\boldsymbol{\theta}}$ is a solution to \eqnref{eqn: objectmain} then: \vspace{-1mm} \begin{equation*} \hat{\boldsymbol{\theta}} \in \mathop{\rm arg\,min}_{\boldsymbol{\theta} \in \mathbb{R}^d} \sum_{i = 1}^{d+1} \max(-y_i\cdot\boldsymbol{\theta}\cdot{{\mathbf{x}}}_i,\: 0) \end{equation*} Also, ${\mathbf{p}}(\hat{\boldsymbol{\theta}}) = 0 \implies {{\mathbf{x}}}_i\cdot \hat{\boldsymbol{\theta}} \ge 0\; \forall \, i \in [d]$ but then ${\mathbf{x}}_{d} = - \sum_{i=1}^{d-1} {{\mathbf{x}}}_i$ $\implies \forall \, i \in [d]\quad {{\mathbf{x}}}_i\cdot \hat{\boldsymbol{\theta}} = 0$. Note that, $\hat{\boldsymbol{\theta}}\cdot \boldsymbol{\theta}^* \ge 0$ forces $\hat{\boldsymbol{\theta}} = t\boldsymbol{\theta}^*$ for some positive constant $t$. Thus, ${\mathcal{D}}$ is a teaching set for the decision boundary of $\boldsymbol{\theta}^*$. This establishes the upper bound, and hence the theorem follows. \end{proof} \vspace{-3mm} \paragraph{Numerical example} To illustrate \thmref{thm: linear perceptron main result}, we provide a numerical example for teaching a linear perceptron in $\mathbb{R}^3$, with $\boldsymbol{\theta}^* = (-3,3,5)^\top$ (illustrated in \figref{fig:exp:exact-teaching:linear}). To construct the teaching set, we first obtain an orthogonal basis $\{(0.46, 0.86, -0.24)^\top, (0.76, -0.24, 0.6)^\top\}$ for the subspace orthogonal to $\boldsymbol{\theta}^*$, and add a vector $(-1.22, -0.62, -0.36)^\top$ which is in the exact opposite direction of the first two combined. Finally we add to $\mathcal{TS}$ an arbitrary vector which has a positive dot product with the normal vector, e.g. $(-0.46, 0.46, 0.76)^\top$. Labeling all examples positive, we obtain $\mathcal{TS}$ of size $4$. \subsection{Homogeneous Polynomial Kernelized Perceptron}\label{subsec.poly} In this subsection, we study the problem of teaching a polynomial kernelized perceptron in realizable setting. Similar to \secref{subsec.linear}, we establish an exact teaching bound on the teaching dimension under a mild condition on the data distribution. We consider homogeneous polynomial kernel ${\mathcal{K}}$ of degree $k$ in which for any ${\mathbf{x}}, {\mathbf{x}}' \in \mathbb{R}^d$ \[{\mathcal{K}}({\mathbf{x}}, {\mathbf{x}}') = \paren{\langle{\mathbf{x}}, {\mathbf{x}}'\rangle}^k\] If $\Phi(\cdot)$ denotes the \textit{feature map} for the corresponding RKHS, then we know that the dimension of the map is $\binom{d+k-1}{k}$ where each component of the map can be represented by $\Phi_{\boldsymbol{\lambda}}({\mathbf{x}}) = \sqrt{ \frac{k!}{\prod_{i=1}^d\boldsymbol{\lambda}_i!}}{\mathbf{x}}^{\boldsymbol{\lambda}}$ where $\boldsymbol{\lambda} \in \paren{\ensuremath{\mathbb{N}}\cup \curlybracket{0}}^{d}$ and $\sum_{i} \boldsymbol{\lambda}_i = k$. Denote by ${\mathcal{H}}_{{\mathcal{K}}}$ the RKHS corresponding to the polynomial kernel ${\mathcal{K}}$. We use ${\mathcal{H}}_k := {\mathcal{H}}_k(\mathbb{R}^d)$ to represent the linear space of homogeneous polynomials of degree $k$ over $\mathbb{R}^d$. We mention an important result which shows the RKHS for polynomial kernels is isomorphic to the space of homogeneous polynomials of degree $k$ in $d$ variables. \begin{proposition}[Chapter III.2, Proposition 6 \cite{article}]\label{prop: polynomial space} ${\mathcal{H}}_k = {\mathcal{H}}_{{\mathcal{K}}}$ as function spaces and inner product spaces. \end{proposition} The dimension $\dim \paren{{\mathcal{H}}_k(\mathbb{R}^d)}$ of the linear space of homogeneous polynomials of degree $k$ over $\mathbb{R}^d$ is $\binom{d+k-1}{k}$. Denote by $r := \binom{d+k-1}{k}$. Since ${\mathcal{H}}_{{\mathcal{K}}}$ is a vector space for polynomial kernel ${\mathcal{K}}$, thus for exact teaching there is an obvious lower bound of $\bigOmega{\binom{d+k-1}{k}}$ on the teaching dimension. Before we establish the main result of this subsection we state a mild assumption on the target model we consider for exact teaching which is as follows: \begin{assumption}[Existence of orthogonal polynomials]\label{assumption: polyorthogonal} For the target model $\boldsymbol{\theta}^* \in {\mathcal{H}}_{{\mathcal{K}}}$, we assume that there exist $(r-1)$ linearly independent polynomials on the orthogonal subspace of $\boldsymbol{\theta}^*$ in ${\mathcal{H}}_{{\mathcal{K}}}$ of the form $\left\{\Phi({\mathbf{z}}_i)\right\}_{i=1}^{r-1}$ where $\forall i\; {\mathbf{z}}_i \in {\mathcal{X}} $. \end{assumption} Similar to \thmref{thm: linear perceptron main result}, the key insight in having \assref{assumption: polyorthogonal} is to find independent polynomial on the orthogonal subspace defined by $\boldsymbol{\theta}^*$. We state the claim here with proof established \iftoggle{longversion}{in \appref{appendix: polynomial perceptron}}{in the supplemental materials}. \begin{theorem}\label{thm: poly_main_theorem} For all target models $\boldsymbol{\theta}^* \in {\mathcal{H}}_{{\mathcal{K}}}$ for which the \assref{assumption: polyorthogonal} holds, for solving \eqnref{eqn: objectkernel}, the exact teaching dimension for the decision boundary corresponding to $\boldsymbol{\theta}^*$ is $\bigO{\binom{d+k-1}{k}}$. \end{theorem} \begin{figure*}[t] \centering \begin{subfigure}[b]{0.3\textwidth} \centerin \includegraphics[width=\linewidth]{fig/TS_lin3_2.png} \caption{Linear ($\mathcal{TS}$)}\label{fig:exp:exact-teaching:linear} \label{fig:example.polytope} \end{subfigure} \begin{subfigure}[b]{0.3\textwidth} \centerin \includegraphics[width=\linewidth]{fig/TS_poly2_4.png} \caption{Polynomial ($\mathcal{TS}$)}\label{fig:exp:exact-teaching:polynomial} \end{subfigure} \begin{subfigure}[b]{0.3\textwidth} \centerin \includegraphics[width=\linewidth]{fig/ts_poly2_3d_4.png} \caption{Polynomial (feature space)} \label{fig:exp:feature-space:polynomial} \end{subfigure} \caption{Numerical examples of exact teaching for linear and polynomial perceptrons. Cyan plus marks and red dots correspond to positive and negative teaching examples respectively } \label{fig:exp:exact-teahcing} \end{figure*} \vspace{-2mm} \paragraph{Numerical example} For constructing $\mathcal{TS}$ in the polynomial case, we follow a similar strategy in the higher dimensional space that the original data is projected into. The only difference is that we need to ensure the teaching examples have pre-images in the original space. For that, we adopt a randomized algorithm that solves for $r-1$ boundary points in the original space (i.e. solve for $\theta^*\cdot \Phi(\mathbf{x}) = 0$) , while checking the images of these points are linearly independent. Also, instead of adding a vector in the opposite direction of these points combined, we simply repeat the $r-1$ points in the teaching set, while assigning one copy of them positive labels and the other copy negative labels. Finally, we need one last vector (label it positive) whose image has a positive component in $\theta^*$, and we obtain $\mathcal{TS}$ of size $2r-1$. \figref{fig:exp:exact-teaching:polynomial} and \figref{fig:exp:feature-space:polynomial} demonstrate the above constructive procedure on a numerical example with $d=2$, homogeneous polynomial kernel of degree 2, and $\boldsymbol{\theta}^* = {(1, 4, 4)}^\top$. In \figref{fig:exp:exact-teaching:polynomial} we show the decision boundary (red lines) and the level sets (polynomial contours) of this quadratic perceptron, as well as the teaching set identified via the above algorithmic procedure. In \figref{fig:exp:exact-teaching:polynomial}, we visualize the decision boundary (grey plane) in the feature space (after applying the feature map). The blue surface corresponds to all the data points that have pre-images in the original space $\mathbb{R}^2$. \subsection{Limitations in Exact Teaching of Polynomial Kernel Perceptron}\label{subsec.motivation} In the previous section \secref{subsec.poly}, we imposed the \assref{assumption: polyorthogonal} on the target models $\boldsymbol{\theta}^*$. It turns out that we couldn't do better than this. More concretely, we need to impose this assumption for exact teaching of polynomial kernel perceptron learner. Further, there are pathological cases where violation of the assumption leads to models which couldn't be approximately taught. Intuitively, solving \eqnref{eqn: objectkernel} in the paradigm of exact teaching reduces to nullifying the orthogonal subspace of $\boldsymbol{\theta}^*$ i.e. any component of $\boldsymbol{\theta}^*$ along the subspace is nullified. Since the information of the span of the subspace has to be encoded into the datapoints chosen for teaching, \assref{assumption: polyorthogonal} is a natural step to make. Interestingly, we show that the step is not so stringent. In the realizable setting in which all the teaching points are correctly classified, if we lift the assumption then exact teaching is not possible We state the claim in the following lemma: \begin{lemma}\label{lemma: exact teaching} Consider a target model $\boldsymbol{\theta}^*$ that doesn't satisfy \assref{assumption: polyorthogonal}. Then, there doesn't exist a teaching set $\mathcal{TS}_{\boldsymbol{\theta}^*}$ which exactly teaches $\boldsymbol{\theta}^*$ i.e. for any $\mathcal{TS}_{\boldsymbol{\theta}^*}$ and any real $t > 0$ $${\boldsymbol{\mathcal{A}}}_{opt}\paren{\mathcal{TS}_{\boldsymbol{\theta}^*}} \neq \{t\boldsymbol{\theta}^*\}.$$ \end{lemma} \lemref{lemma: exact teaching} shows that for \tt{exact} teaching $\boldsymbol{\theta}^*$ should satisfy \assref{assumption: polyorthogonal}. Then, the natural question that arises is whether we can achieve arbitrarily $\epsilon$-close \tt{approximate} teaching for $\boldsymbol{\theta}^*$. In other words, we would like to find $\Tilde{\boldsymbol{\theta}}^*$ that satisfies \assref{assumption: polyorthogonal} and is in $\epsilon$-neighbourhood of $\boldsymbol{\theta}^*$. We show a negative result for this when $k$ is even. For this we assume that, the datapoints in the teaching set $\mathcal{TS}_{\Tilde{\boldsymbol{\theta}}^*}$ have lower-bounded norm, call it, $\delta > 0$ i.e. if $({{\mathbf{x}}}_i,y_i) \in \mathcal{TS}_{\Tilde{\boldsymbol{\theta}}^*}$ then $||\Phi({{\mathbf{x}}}_i)|| \ge \delta$. We require this additional assumption only for the purpose of analysis. We would show that it wouldn't lead to any pathological cases where the constructed target model $\boldsymbol{\theta}^*$ incorporates approximate teaching. \begin{lemma}\label{lemma: approximate teaching } Let ${\mathcal{X}} \subseteq \mathbb{R}^d$ and ${\mathcal{H}}_{{\mathcal{K}}}$ be the reproducing kernel Hilbert space such that kernel function ${\mathcal{K}}$ is of degree $k$. If $k$ has parity even then there exists a target model $\boldsymbol{\theta}^*$ which violates \assref{assumption: polyorthogonal} and can't be taught approximately. \end{lemma} The results are discussed in details with proofs \iftoggle{longversion}{in \appref{appendix: motivation}}{in the supplemental materials}. \assref{assumption: polyorthogonal} and the stated lemmas provide insights into understanding the problem of teaching for non-linear perceptron kernels. In the next section, we study Gaussian kernel and the ideas generated here would be useful in devising a teaching set in the paradigm of approximate teaching. \subsection{Gaussian Kernelized Perceptron}\label{subsec.gaussiankernel} In this subsection, we consider the Gaussian kernel. Under mild assumptions inspired by the analysis of teaching dimension for exact teaching of linear and polynomial kernel perceptrons, we would establish as our main result an upper bound on the $\epsilon$-approximate teaching dimension of Gaussian kernel perceptrons using a construction of an $\epsilon$-approximate teaching set. \paragraph{Preliminaries of Gaussian kernel} A Gaussian kernel ${\mathcal{K}}$ is a function of the form \begin{equation} {\mathcal{K}}({\mathbf{x}}, {\mathbf{x}}') = \mathbi{e}^{-\frac{||{\mathbf{x}}-{\mathbf{x}}'||^2}{2\sigma^2}} \label{eqn:eqn11} \end{equation} for any ${\mathbf{x}}, {\mathbf{x}}' \in \mathbb{R}^d$ and parameter $\sigma$. First, we would try to understand the feature map before we find an approximation to it. Notice: \[\mathbi{e}^{-\frac{||{\mathbf{x}}-{\mathbf{x}}'||^2}{2\sigma^2}} = \mathbi{e}^{-\frac{||{\mathbf{x}}||^2}{2\sigma^2}}\mathbi{e}^{-\frac{||{\mathbf{x}}'||^2}{2\sigma^2}}\mathbi{e}^{\frac{\normg{{\mathbf{x}}}{{\mathbf{x}}'}}{\sigma^2}} \] Consider the scalar term $z = \normg{{\mathbf{x}}}{{\mathbf{x}}'}/\sigma^2$. We can expand the term of the product using the Taylor expansion of $\mathbi{e}^z$ near $z = 0$ as shown in \citet{Cotter2011ExplicitAO}, which amounts to $\mathbi{e}^{\frac{\normg{{\mathbf{x}}}{{\mathbf{x}}'}}{\sigma^2}} = \sum_{k=0}^{\infty}\frac{1}{k!}\paren{\frac{\normg{{\mathbf{x}}}{{\mathbf{x}}'}}{\sigma^2}}^k$. We can further expand the previous sum as \begin{align} \mathbi{e}^{\frac{\normg{{\mathbf{x}}}{{\mathbf{x}}'}}{\sigma^2}} &= \sum_{k=0}^{\infty}\frac{1}{k!}\paren{\frac{\normg{{\mathbf{x}}}{{\mathbf{x}}'}}{\sigma^2}}^k \nonumber\\ &= \sum_{k=0}^{\infty}\frac{1}{k!\sigma^{2k}}\bigparen{\sum_{l=1}^d {\mathbf{x}}_l\cdot{\mathbf{x}}'_l}^k \nonumber\\ &= \sum_{k=0}^{\infty}\frac{1}{k!\sigma^{2k}}\sum_{|\boldsymbol{\lambda}|=k} {\mathcal{C}}^k_{\boldsymbol{\lambda}}\cdot{\mathbf{x}}^{\boldsymbol{\lambda}}\cdot({\mathbf{x}}')^{\boldsymbol{\lambda}} \label{eqn:eqn12} \end{align} where ${\mathcal{C}}^k_{\boldsymbol{\lambda}} = \frac{k!}{\prod_{i=1}^d\boldsymbol{\lambda}_i!}$. Thus, we use \eqnref{eqn:eqn12} to obtain explicit feature representation to the Gaussian kernel in \eqnref{eqn:eqn11} as $\Phi_{k,\boldsymbol{\lambda}}({\mathbf{x}}) = \mathbi{e}^{-\frac{||{\mathbf{x}}||^2}{2\sigma^2}}\cdot \frac{\sqrt{{\mathcal{C}}^k_{\boldsymbol{\lambda}}}}{\sqrt{k!}\sigma^k}\cdot{\mathbf{x}}^{\boldsymbol{\lambda}}$. We get the explicit feature map $\Phi(\cdot)$ for the Gaussian kernel with coordinates as specified. Theorem 1 of \citet{haquangminh} characterizes the RKHS of Gaussian kernel. It establishes that $\dim({\mathcal{H}}_{{\mathcal{K}}}) = \infty$. Thus, we note that the exact teaching for an arbitrary target classifier $f^*$ in this setting has an infinite lower bound. This calls for analysing the teaching problem of a Gaussian kernel in the \tt{approximate} teaching setting \paragraph{Definitions and notations for approximate teaching} For any classifier $f \in {\mathcal{H}}_{{\mathcal{K}}}$, we define $\textbf{err}$($f$) = $\expctover{({\mathbf{x}},y) \sim {\mathcal{P}}({\mathbf{x}},y)}{\max(-y\cdot f({\mathbf{x}}),0)}$. Our goal is to find a classifier $f$ with the property that its expected true loss $\textbf{err}$($f$) is as small as possible. In the realizable setting, we assume that there exists an optimal separator $f^*$ such that for any data instances sampled from the data distribution the labels are consistent i.e. ${\mathcal{P}}(y\cdot f^*({\mathbf{x}}) \le 0) = 0$. In addition, we also experiment for the non-realizable setting. In the rest of the subsection, we would study the relationship between the teaching complexity for an optimal Gaussian kernel perceptron for \eqnref{eqn: objectkernel} and $|\textbf{err}(f^*) - \textbf{err}(\hat{f})|$ where $f^*$ is the optimal separator and $\hat{f}$ is the solution to ${\boldsymbol{\mathcal{A}}}_{opt}(\mathcal{TS}_{\boldsymbol{\theta}^*})$ for the constructed teaching set $\mathcal{TS}_{\boldsymbol{\theta}^*}$. \subsubsection{Gaussian Kernel Approximation}\label{subsec: gaussian_kernel_approx} Now, we would talk about finite-dimensional polynomial approximation $\Tilde{\Phi}$ to the Gaussian feature map $\Phi$ via projection as shown in \citet{Cotter2011ExplicitAO}. Consider \begin{align*} \Tilde{\Phi}&: \mathbb{R}^d \longrightarrow \mathbb{R}^q\\ \Tilde{{\mathcal{K}}}({\mathbf{x}}, {\mathbf{x}}') &= \Tilde{\Phi}({\mathbf{x}})\cdot\Tilde{\Phi}({\mathbf{x}}') \end{align*} With these approximations, we consider classifiers of the form $\Tilde{f}({\mathbf{x}}) = \Tilde{\boldsymbol{\theta}}\cdot \Tilde{\Phi}({\mathbf{x}})$ such that $\Tilde{\boldsymbol{\theta}} \in \mathbb{R}^q$. Now, assume that there is a projection map $\mathbb{P}$ such that $\Tilde{\Phi} = \mathbb{P}\Phi$. In \citet{Cotter2011ExplicitAO}, authors used the following approximation to the Gaussian kernel: \begin{equation} \Tilde{{\mathcal{K}}}({\mathbf{x}}, {\mathbf{x}}') = \mathbi{e}^{-\frac{||{\mathbf{x}}||^2}{2\sigma^2}}\mathbi{e}^{-\frac{||{\mathbf{x}}'||^2}{2\sigma^2}}\sum_{k=0}^{s}\frac{1}{k!}\paren{\frac{\normg{{\mathbf{x}}}{{\mathbf{x}}'}}{\sigma^2}}^k \label{eqn: eqn17} \end{equation} This gives the following explicit feature representation for the approximated kernel: \begin{equation} \forall k \le s,\quad \Tilde{\Phi}_{k,\boldsymbol{\lambda}}({\mathbf{x}}) = \Phi_{k,\boldsymbol{\lambda}}({\mathbf{x}}) = \mathbi{e}^{-\frac{||{\mathbf{x}}||^2}{2\sigma^2}}\cdot \frac{\sqrt{{\mathcal{C}}^k_{\boldsymbol{\lambda}}}}{\sqrt{k!}\sigma^k}\cdot{\mathbf{x}}^{\boldsymbol{\lambda}} \label{eqn:eqn18} \end{equation} where $\Phi_{k,\boldsymbol{\lambda}}({\mathbf{x}})$ is the coordinate for Gaussian feature map. Note that the feature map $\Tilde{\Phi}$ defined by the explicit features in \eqnref{eqn:eqn18} has dimension $\binom{d+s}{d}$. Thus, $\mathbb{P}\Phi = \Tilde{\Phi}$ where the first $\binom{d+s}{d}$ coordinates are retained. We denote the RKHS corresponding to $\Tilde{{\mathcal{K}}}$ as ${\mathcal{H}}_{\Tilde{{\mathcal{K}}}}$. A simple property of the approximated kernel map is stated in the following lemma which was proven in \citet{Cotter2011ExplicitAO}. \begin{lemma}[\citet{Cotter2011ExplicitAO}]\label{lemma: approxbound} For the approximated map $\Tilde{{\mathcal{K}}}$, we obtain the following upper bound: \begin{equation} \left|{\mathcal{K}}({\mathbf{x}},{\mathbf{x}}) - \Tilde{{\mathcal{K}}}({\mathbf{x}},{\mathbf{x}})\right| \le \frac{1}{(s+1)!}\paren{\frac{||{\mathbf{x}}||\cdot ||{\mathbf{x}}'||}{\sigma^2}}^{s+1} \label{eqn: approximatekernel} \end{equation} \end{lemma} Note that if $s$ is chosen large enough and the points ${\mathbf{x}}, {\mathbf{x}}'$ are bounded wrt $\sigma^2$, then RHS of \eqnref{eqn: approximatekernel} can be bounded by any $\epsilon > 0$. Since $\left|{\mathcal{K}}({\mathbf{x}},{\mathbf{x}}) - \Tilde{{\mathcal{K}}}({\mathbf{x}},{\mathbf{x}})\right| = \inmod{\mathbb{P}^{\bot}\Phi({\mathbf{x}})}^2$, thus for a Gaussian kernel, information theoretically, the first $\binom{d+s}{s}$ coordinates are highly sensitive. We would try to analyze this observation under some mild assumptions on the data distribution to construct an $\epsilon$-approximate teaching set. As discussed in \iftoggle{longversion}{\appref{appendix: gaussian perceptron}}{ the supplemental materials}, we would find the value of $s$ as if the datapoints are coming from a ball of radius $R := \max\left\{\frac{\log^2 \frac{1}{\epsilon}}{e^2},d\right\}$ in $\mathbb{R}^d$ i.e. $\frac{\inmod{{\mathbf{x}}}^2}{\sigma^2} \le R$. Thus, we wish to solve for the value of $s$ such that $\frac{1}{(s+1)!}\cdot \paren{R}^{s+1} \le \epsilon$. To approximate $s$ we use Sterling's approximation, which states that for all positive integers $n$, we have $$\sqrt{2\pi}n^{n+1/2}e^{-n} \le n! \le en^{n+1/2}e^{-n}.$$ Using the bound stated in \lemref{lemma: approxbound}, we fix the value for $s$ as $e^2\cdot R$. We would assume that $R = \frac{\log^2 \frac{1}{\epsilon}}{e^2}$ since we wish to achieve arbitrarily small $\epsilon$-approximate\footnote{When $R = d$ all the key results follow the same analysis.} teaching set. We define $r:= r(\boldsymbol{\theta}^*, \epsilon) = \binom{d+s}{s}$. \vspace{-1mm} \subsubsection{Bounding the Error}\label{subsec: bounded_error} \vspace{-1mm} In this subsection, we discuss our key results on approximate teaching of a Gaussian kernel perceptron learner under some mild assumptions on the target model $\boldsymbol{\theta}^*$. In order to show $\left|\textbf{err}(f^*) - \textbf{err}(\hat{f})\right| \le \epsilon$ via optimizing to a solution $\hat{\boldsymbol{\theta}}$ for \eqnref{eqn: objectkernel}, we would achieve a point-wise $\epsilon$-closeness between $f^*$ and $\hat{f}$. Specifically, we show that $\left|f^*({\mathbf{x}})-\hat{f}({\mathbf{x}})\right| \le \epsilon$ universally which is similar in spirit to universal approximation theorems~\cite{Liang2017WhyDN, lu2020universal, Yarotsky2017ErrorBF} for neural networks. We prove that this universal approximation could be achieved with $d^{\bigO{\log^2 \frac{1}{\epsilon}}}$ size teaching set. We assume that the input space ${\mathcal{X}}$ is bounded such that $\forall {\mathbf{x}} \in {\mathcal{X}}$\,\, $\frac{\norm{{\mathbf{x}}}{{\mathbf{x}}}}{\sigma^2} \le 2\sqrt{R}$. Since the motivation is to find classifiers which are close to the optimal one point-wise, thus we assume that target model $\boldsymbol{\theta}^*$ has unit norm. As mentioned in \eqnref{eqn: kernelfunction}, we can write the target model $\boldsymbol{\theta}^* \in {\mathcal{H}}_{{\mathcal{K}}}$ as $\boldsymbol{\theta}^* = \sum_{i=1}^l \alpha_i\cdot{\mathcal{K}}({\mathbf{a}}_i, \cdot)$ for some $\{{\mathbf{a}}_i\}_{i=1}^l \subset {\mathcal{X}}$ and $\alpha_i \in \ensuremath{\mathbb{R}}$. The classifier corresponding to $\boldsymbol{\theta}^*$ is represented by $f^*$. \eqnref{eqn: objectkernel} can be rewritten corresponding to a teaching set ${\mathcal{D}} := \curlybracket{\paren{{\mathbf{x}}_i,\; y_i}}_{i=1}^n$ as: \begin{equation} {\boldsymbol{\mathcal{A}}}_{opt} := \mathop{\rm arg\,min}_{\beta \in \ensuremath{\mathbb{R}}^l} \sum_{i = 1}^n \max\bigparen{-y_i\cdot \sum_{j=1}^l \beta_j\cdot{\mathcal{K}}({\mathbf{a}}_j,{\mathbf{x}}_i),\; 0}\label{eqn: bounded1} \end{equation} Similar to \assref{assumption: polyorthogonal} (cf \secref{subsec.poly}), to construct an approximate teaching set we assume a target model $\boldsymbol{\theta}^*$ has the property that for some truncated polynomial space ${\mathcal{H}}_{\Tilde{{\mathcal{K}}}}$ defined by feature map $\Tilde{\Phi}$ there are linearly independent projections in the orthogonal complement of $\mathbb{P}\boldsymbol{\theta}^*$ in ${\mathcal{H}}_{\Tilde{{\mathcal{K}}}}$. More formally, we state the property as an assumption which is discussed in details \iftoggle{longversion}{in \appref{appendix: motivation}}{in the supplemental materials}. \begin{assumption}[Existence of orthogonal classifiers]\label{assumption: orthogonal} For the target model $\boldsymbol{\theta}^*$ and some $\epsilon > 0$, we assume that there exists $r= r(\boldsymbol{\theta}^*, \epsilon)$ such that $\mathbb{P}\boldsymbol{\theta}^*$ has $r-1$ linear independent projections on the orthogonal subspace of $\mathbb{P}\boldsymbol{\theta}^*$ in ${\mathcal{H}}_{\Tilde{{\mathcal{K}}}}$ of the form $\{\Tilde{\Phi}({\mathbf{z}}_i)\}_{i=1}^{r-1}$ such that $\forall i\,\, {\mathbf{z}}_i \in {\mathcal{X}} $. \end{assumption} For the analysis of the key results, we impose a smoothness condition on the linear independent projections $\{\Tilde{\Phi}({\mathbf{z}}_i)\}_{i=1}^{r-1}$ that they are oriented away by a factor of $\frac{1}{r-1}$. Concretely, for any $i,j$ $\left|\Tilde{\Phi}({\mathbf{z}}_i)\cdot\Tilde{\Phi}({\mathbf{z}}_j)\right| \le \frac{1}{2(r-1)}$. This smoothness condition is discussed in the supplemental. Now, we consider the following reformulation of the optimization problem in \eqnref{eqn: bounded1} as follows: \begin{align} \hspace*{-3mm}{\boldsymbol{\mathcal{A}}}_{opt} := \mathop{\rm arg\,min}_{\beta_0 \in \ensuremath{\mathbb{R}},\, \gamma \in \ensuremath{\mathbb{R}}^{r-1}} \sum_{i = 1}^{2r-1} \max\paren{\ell(\beta_0, \gamma, {\mathbf{x}}_i, y_i),\; 0}\label{eqn: bounded} \end{align} where for any $i \in \bracket{2r-1}$ \vspace{-3mm}$$ \ell(\beta_0, \gamma, {\mathbf{x}}_i, y_i) = -y_i\cdot\bigparen{ \beta_0\cdot {\mathcal{K}}({\mathbf{a}},{\mathbf{x}}_i) + \sum_{j=1}^{r-1}\gamma_j\cdot {\mathcal{K}}({\mathbf{z}}_j, {\mathbf{x}}_i)} \nonumber$$ and with respect to the teaching set \begin{align} \mathcal{TS}_{\boldsymbol{\theta}^*} := \curlybracket{\parenb{{\mathbf{z}}_i, 1},\, \parenb{{\mathbf{z}}_i, -1}}_{i = 1}^{r-1} \cup \curlybracket{\parenb{{\mathbf{a}}, 1}}\label{eqn: teaching set} \end{align} where ${\mathbf{a}}$ is chosen such that $\mathbb{P}\boldsymbol{\theta}^*\cdot\mathbb{P}\Phi({\mathbf{a}}) > 0$\footnote{We assume $\boldsymbol{\theta}^*$ is non-degenerate in $\Tilde{{\mathcal{K}}}$ (as for polynomial kernels in \secref{subsec.poly}) i.e. has points ${\mathbf{a}} \in {\mathcal{X}}$ such that $\mathbb{P}\boldsymbol{\theta}^*\cdot\mathbb{P}\Phi({\mathbf{a}}) > 0$ (classified with label 1).} and $\Phi({\mathbf{a}})\cdot\Phi({\mathbf{z}}_i) \le Q\cdot \epsilon$ (where $Q$ is a constant). ${\mathbf{a}}$ could be chosen from a $\mathcal{B}(\sqrt{2\sqrt{R}\sigma^2},0)$ spherical ball in $\ensuremath{\mathbb{R}}^d$. We index the set $\mathcal{TS}_{\boldsymbol{\theta}^*}$ as $\curlybracket{\paren{{\mathbf{x}}_i, y_i}}_{i=1}^{2r-1}$. \eqnref{eqn: bounded} is optimized over $\hat{\boldsymbol{\theta}} = \beta_0\cdot {\mathcal{K}}({\mathbf{a}},\cdot) + \sum_{j=1}^{r-1}\gamma_j\cdot {\mathcal{K}}({\mathbf{z}}_j, \cdot)$ such that $\hat{\boldsymbol{\theta}}\cdot\Phi({\mathbf{a}}) > 0$ and $\{\Phi({\mathbf{z}}_i)\}_{i=1}^{r-1}$ satisfy \assref{assumption: orthogonal} where $\inmod{\hat{\boldsymbol{\theta}}} = \bigO{1}$. \begin{figure*}[t!] \centering \begin{subfigure}[b]{0.28\textwidth} \centering \includegraphics[width=\linewidth]{fig/supp_fig1-1b.png} \caption{Optimal Gaussian boundary} \label{fig:Teacher_RBf} \end{subfigure} \qquad \begin{subfigure}[b]{0.28\textwidth} \centering \includegraphics[width=\linewidth]{fig/supp_fig1-2b.png} \caption{Polynomial approximation} \label{fig:Polynomial_approx} \end{subfigure} \qquad \begin{subfigure}[b]{0.28\textwidth} \centering \includegraphics[width=\linewidth]{fig/supp_fig1-3b.png} \caption{Taught Gaussian boundary} \label{fig:Learned_RBF} \end{subfigure} \caption{Approximate teaching for Gaussian kernel perceptron. (a) Teacher ``receives'' $\boldsymbol{\theta}^*$ by training from the complete data set; (b) Teacher identifies a polynomial approximation of the Gaussian decision boundary and generates the teaching set $\mathcal{TS}_{\boldsymbol{\theta}^*}$ (marked by red dots and cyan crosses); (c) Learner learns a Gaussian kernel perceptron from $\mathcal{TS}_{\boldsymbol{\theta}^*}$.} \label{fig:example-RBF} \end{figure*} Note that any solution to $\eqnref{eqn: bounded}$ can have unbounded norm and can extend in arbitrary directions, thus we make an assumption on the learner which would be essential to bound the error of optimal separator of \eqnref{eqn: bounded}. \begin{assumption}[Bounded Cone]\label{assumption: bounded cone} For the target model $\boldsymbol{\theta}^* = \sum_{i = 1}^l \alpha_i\cdot{\mathcal{K}}({\mathbf{a}}_i,\cdot)$, the learner optimizes to a solution $\hat{\boldsymbol{\theta}}$ for \eqnref{eqn: bounded} with bounded coefficients. Alternatively, the sums $\sum_{i=1}^l \left|\alpha_i\right|$ and $\left|\beta_0\right| + \sum_{j=1}^{r-1}\left|\gamma_j\right|$ are bounded where $\hat{\boldsymbol{\theta}} \in {\mathcal{H}}_{{\mathcal{K}}}$ has the form $\hat{\boldsymbol{\theta}} = \beta_0\cdot {\mathcal{K}}({\mathbf{a}}_j,\cdot) + \sum_{j=1}^{r-1}\gamma_j\cdot {\mathcal{K}}({\mathbf{z}}_j, \cdot)$. \end{assumption} This assumption is fairly mild or natural in the sense that for $\hat{\boldsymbol{\theta}} \in {\boldsymbol{\mathcal{A}}}_{opt}$ as a classifier approximates $\boldsymbol{\theta}^*$ point-wise then they shouldn't be highly (or unboundedly) sensitive to datapoints involved in the classifiers. It is discussed in greater details \iftoggle{longversion}{in \appref{appendix: motivation}}{in the supplemental materials}. We denote by $\boldsymbol{C}_{\epsilon} := \sum_{i=1}^l |\alpha_i|$ and $\boldsymbol{D}_{\epsilon}:= |\beta_0| + \sum_{j=1}^{r-1}|\gamma_j|$. In \iftoggle{longversion}{ \appref{appendixsub: solutionexists}}{ the supplemental materials}, we show that there exists a unique solution (upto a positive scaling) to \eqnref{eqn: bounded} which satisfies \assref{assumption: bounded cone}. We would show that $\mathcal{TS}_{\boldsymbol{\theta}^*}$ is an $\epsilon$-approximate teaching set with $r = d^{\bigO{\log^2 \frac{1}{\epsilon}}}$ on the $\epsilon$-approximate teaching dimension. To achieve this, we first establish the $\epsilon$-\tt{closeness} of $\hat{f}$ (classifier $\hat{f}({\mathbf{x}}) := \hat{\boldsymbol{\theta}}\cdot\Phi({\mathbf{x}})$ where $\hat{\boldsymbol{\theta}} \in {\boldsymbol{\mathcal{A}}}_{opt}$) to $f^*$. Formally, we state the result as follows: \begin{theorem}\label{thm: boundedclassifier} For any target $\boldsymbol{\theta}^* \in {\mathcal{H}}_{{\mathcal{K}}}$ that satisfies \assref{assumption: orthogonal}-\ref{assumption: bounded cone} and $\epsilon > 0$, the teaching set $\mathcal{TS}_{\boldsymbol{\theta}^*}$ constructed for \eqnref{eqn: bounded} satisfies $\left|f^*({\mathbf{x}}) - \hat{f}({\mathbf{x}})\right| \le \epsilon$ for any ${\mathbf{x}} \in {\mathcal{X}}$ and any $\hat{f} \in {\boldsymbol{\mathcal{A}}}_{opt}(\mathcal{TS}_{\boldsymbol{\theta}^*})$. \end{theorem} Using \thmref{thm: boundedclassifier}, we can obtain the main result of the subsection which gives an $d^{\bigO{\log^2 \frac{1}{\epsilon}}}$ bound on $\epsilon$-approximate teaching dimension. We detail the proofs \iftoggle{longversion}{in \appref{appendix: gaussian perceptron}}{in the supplemental materials}: \begin{theorem}\label{thm: gaussian_main_thm} For any target $\boldsymbol{\theta}^* \in {\mathcal{H}}_{{\mathcal{K}}}$ that satisfies \assref{assumption: orthogonal}-\ref{assumption: bounded cone} and $\epsilon > 0$, the teaching set $\mathcal{TS}_{\boldsymbol{\theta}^*}$ constructed for \eqnref{eqn: bounded} is an $\epsilon$-approximate teaching set with $\epsilon$-$TD(\boldsymbol{\theta}^*,{\boldsymbol{\mathcal{A}}}_{opt}) = d^{\bigO{\log^2 \frac{1}{\epsilon}}}$ i.e. for any $\hat{f} \in {\boldsymbol{\mathcal{A}}}_{opt}(\mathcal{TS}_{\boldsymbol{\theta}^*})$, $$\left|\textbf{err}(f^*) - \textbf{err}(\hat{f})\right| \le \epsilon.$$ \end{theorem} \paragraph{Numerical example} \figref{fig:example-RBF} demonstrates the approximate teaching process for a Gaussian learner. We aim to teach the optimal model $\boldsymbol{\theta}^*$ (infinite-dimensional) trained on a pre-collected dataset with Gaussian parameter $\sigma = 0.9$, whose corresponding boundary is shown in \figref{fig:Teacher_RBf}. Now, for approximate teaching, the teacher calculates $\Tilde{\boldsymbol{\theta}}$ using the polynomial approximated kernel (i.e. $\Tilde{{\mathcal{K}}}$, and in this case, k=5) in \eqnref{eqn: eqn17} and the corresponding feature map in \eqnref{eqn:eqn18}. To ensure \assref{assumption: orthogonal} is met while generating teaching examples for $\Tilde{\boldsymbol{\theta}}$, we employ the randomized algorithm (as was used in \secref{subsec.poly}) with the key idea of ensuring that the teaching examples on the boundary are linearly independent in the approximated polynomial feature space, i.e. $\Tilde{{\mathcal{K}}}({\mathbf{z}}_i, {\mathbf{z}}_j) = 0$. Finally, the Gaussian learner receives $\mathcal{TS}_{\boldsymbol{\theta}^*}$ and learns the boundary shown in \figref{fig:Learned_RBF}. Note the slight difference between the boundaries in \figref{fig:Polynomial_approx} and in \figref{fig:Learned_RBF} as the learner learns with a Gaussian kernel. \section{Problem Statement}\label{sec.statement} \paragraph{Basic definitions} We denote by ${\mathcal{X}}$ the input space and ${\mathcal{Y}}:= \{-1,1\}$ the output space. A hypothesis is a function $h: {\mathcal{X}} \to {\mathcal{Y}}$. In this paper, we identify a hypothesis $h_{\boldsymbol{\theta}}$ with its model parameter $\boldsymbol{\theta}$. The hypothesis space ${\mathcal{H}}$ is a set of hypotheses. By training point we mean a pair $\paren{{\mathbf{x}},y} \in {\mathcal{X}} \times {\mathcal{Y}}$. We assume that the training points are drawn from an unknown distribution ${\mathcal{P}}$ over ${\mathcal{X}} \times {\mathcal{Y}}$. A training set is a multiset ${\mathcal{D}}$ = $\curlybracket{\paren{{\mathbf{x}}_1,y_1},\cdots,\paren{{\mathbf{x}}_n,y_n}}$ where repeated pairs are allowed. Let $\mathbb{D}$ denote the set of all training sets of all sizes. A learning algorithm ${\boldsymbol{\mathcal{A}}}: \mathbb{D} \to 2^{{\mathcal{H}}} $ takes in a training set $D \in \mathbb{D}$ and outputs a subset of the hypothesis space ${\mathcal{H}}$. That is, ${\boldsymbol{\mathcal{A}}}$ doesn't necessarily return a unique hypothesis. \vspace{-1mm} \paragraph{Kernel perceptron} Consider a set of training points ${\mathcal{D}} := \curlybracket{\paren{{\mathbf{x}}_i, y_i}}_{i=1}^n$ where ${\mathbf{x}}_i \in \ensuremath{\mathbb{R}}^d$ and hypothesis $\boldsymbol{\theta} \in \mathbb{R}^d$. A linear perceptron is defined as $f_{\boldsymbol{\theta}}({\mathbf{x}}):= \sign(\boldsymbol{\theta}\cdot {\mathbf{x}})$ in homogeneous setting. We consider the algorithm ${\boldsymbol{\mathcal{A}}}_{opt}$ to learn an optimal perceptron to classify ${\mathcal{D}}$ as defined below: \begin{equation} {\boldsymbol{\mathcal{A}}}_{opt}\paren{{\mathcal{D}}} := \mathop{\rm arg\,min}_{\boldsymbol{\theta} \in \ensuremath{\mathbb{R}}^d} \sum_{i = 1}^n \ell(f_{\boldsymbol{\theta}}({\mathbf{x}}_i),y_i). \label{eqn: objectmain} \end{equation} where the loss function $\ell(f_{\boldsymbol{\theta}}({\mathbf{x}}),y) := \max(-y\cdot f_{\boldsymbol{\theta}}({\mathbf{x}}), 0)$. Similarly, we consider the non-linear setting via kernel-based hypotheses for perceptrons that are defined with respect to a kernel operator ${\mathcal{K}}: {\mathcal{X}} \times {\mathcal{X}} \to \ensuremath{\mathbb{R}}$ which adheres to Mercer’s positive definite conditions \cite{vapnik1998statistical}. A kernel-based hypothesis has the form, \begin{equation} f({\mathbf{x}}) = \sum_{i=1}^k{\alpha_i}\cdot{\mathcal{K}}({\mathbf{x}}_i,{\mathbf{x}}) \label{eqn: kernelfunction} \end{equation} where $\forall i\,\, {\mathbf{x}}_i \in {\mathcal{X}}$ and $\alpha_i$ are reals. In order to simplify the derivation of the algorithms and their analysis, we associate a \tt{reproducing kernel Hilbert space} (RKHS) with ${\mathcal{K}}$ in the standard way common to all kernel methods. Formally, let ${\mathcal{H}}_{{\mathcal{K}}}$ be the closure of the set of all hypotheses of the form given in \eqnref{eqn: kernelfunction}. A non-linear kernel perceptron corresponding to ${\mathcal{K}}$ optimizes \eqnref{eqn: objectmain} as follows: \begin{equation} {\boldsymbol{\mathcal{A}}}_{opt}({\mathcal{D}}):= \mathop{\rm arg\,min}_{\boldsymbol{\theta} \in {\mathcal{H}}_{{\mathcal{K}}}}\sum_{i=1}^n \ell(f_{\boldsymbol{\theta}}({\mathbf{x}}_i),y_i)\label{eqn: objectkernel} \end{equation} where $f_{\boldsymbol{\theta}}(\cdot) = \sum_{i=1}^l \alpha_i\cdot {\mathcal{K}}({\mathbf{a}}_i,\cdot)$ for some $\{{\mathbf{a}}_i\}_{i=1}^l \subset {\mathcal{X}}$ and $\alpha_i$ real. Alternatively, we also write $f_{\boldsymbol{\theta}}(\cdot) = \boldsymbol{\theta}\cdot\Phi(\cdot)$ where $\Phi : {\mathcal{X}} \rightarrow {\mathcal{H}}_{{\mathcal{K}}}$ is defined as feature map to the kernel function ${\mathcal{K}}$. A reproducing kernel Hilbert space with ${\mathcal{K}}$ could be decomposed as ${\mathcal{K}}({\mathbf{x}},{\mathbf{x}}') = \normg{\Phi({\mathbf{x}})}{\Phi({\mathbf{x}}')}$ \cite{learnkernel} for any ${\mathbf{x}}, {\mathbf{x}}' \in {\mathcal{X}}$. Thus, we also identify $f_{\boldsymbol{\theta}}$ as $\sum_{i=1}^l \alpha_i\cdot \Phi({\mathbf{a}}_i)$. \paragraph{The teaching problem We are interested in the problem of teaching a target hypothesis $\boldsymbol{\theta}^*$ where a helpful \tt{teacher} provides labelled data points $\mathcal{TS} \subseteq {\mathcal{X}}\times {\mathcal{Y}}$, also defined as a \tt{teaching set}. Assuming the constructive setting \cite{JMLR:v17:15-630}, to teach a kernel perceptron learner the teacher can construct a training set with any items in $\mathbb{R}^d$ i.e. for any $({\mathbf{x}}', y') \in \mathcal{TS}$ we have ${\mathbf{x}}' \in \mathbb{R}^d$ and $y' \in \curlybracket{-1,1}$. Importantly, for the purpose of teaching we do not \tt{assume} that $\mathcal{TS}$ are drawn \tt{i.i.d} from a distribution. We define the teaching dimension for \tt{exact} parameter of $\boldsymbol{\theta}^*$ corresponding to a kernel perceptron as $TD(\boldsymbol{\theta}^*, {\boldsymbol{\mathcal{A}}}_{opt})$, which is the size of the smallest teaching set $\mathcal{TS}$ such that ${\boldsymbol{\mathcal{A}}}_{opt}\paren{\mathcal{TS}} = \{\boldsymbol{\theta}^*\}$. We define teaching of exact parameters of a target hypothesis $\boldsymbol{\theta}^*$ as \tt{exact teaching}. Since, a perceptron is agnostic to norms, we study the problem of teaching a target classifier \tt{decision boundary} where ${\boldsymbol{\mathcal{A}}}_{opt}\paren{\mathcal{TS}} = \{t\boldsymbol{\theta}^*\}$ for some real $t > 0$. Thus, $$TD(\{t\boldsymbol{\theta}^*\}, {\boldsymbol{\mathcal{A}}}_{opt}) = \min_{\text{real}\,p > 0} TD(p\boldsymbol{\theta}^*, {\boldsymbol{\mathcal{A}}}_{opt}).$$ Since it can be stringent to construct a teaching set for decision boundary (see \secref{subsec.gaussiankernel}), exact teaching is not always feasible. We introduce and study \tt{approximate teaching} which is formally defined as: \begin{definition}[$\epsilon$-approximate teaching set] Consider a kernel perceptron learner, with a kernel ${\mathcal{K}}: {\mathcal{X}} \times {\mathcal{X}} \to \ensuremath{\mathbb{R}}$ and the corresponding RKHS feature map $\Phi(\cdot)$. For a target model $\boldsymbol{\theta}^* \in {\mathcal{H}}_{{\mathcal{K}}}$ and $\epsilon > 0$, we say $\mathcal{TS} \subseteq {\mathcal{X}}\times {\mathcal{Y}}$ is an $\epsilon$-approximate teaching set wrt to ${\mathcal{P}}$ if the kernel perceptron $\hat{\boldsymbol{\theta}} \in {\boldsymbol{\mathcal{A}}}_{opt}(\mathcal{TS})$ satisfies \looseness -1 \begin{equation} \left|\expct{\max(-y\cdot f^*({\mathbf{x}}), 0)} - \expct{\max(-y\cdot \hat{f}({\mathbf{x}}), 0)}\right| \le \epsilon \end{equation} where the expectations are over $({\mathbf{x}},y)\sim {\mathcal{P}}$ and $f^*({\mathbf{x}}) = \boldsymbol{\theta}^*\cdot \Phi({\mathbf{x}})$ and $\hat{f}({\mathbf{x}}) = \hat{\boldsymbol{\theta}}\cdot \Phi({\mathbf{x}})$. \end{definition} Naturally, we define approximate teaching dimension as: \begin{definition}[$\epsilon$-approximate teaching dimension]\label{def:approxteaching} Consider a kernel perceptron learner, with a kernel ${\mathcal{K}}: {\mathcal{X}} \times {\mathcal{X}} \to \ensuremath{\mathbb{R}}$ and the corresponding RKHS feature map $\Phi(\cdot)$. For a target model $\boldsymbol{\theta}^* \in {\mathcal{H}}_{{\mathcal{K}}}$ and $\epsilon > 0$, we define $\epsilon$-$TD(\boldsymbol{\theta}^*,{\boldsymbol{\mathcal{A}}}_{opt})$ as the teaching dimension which is the size of the smallest teaching set for $\epsilon$-approximate teaching of $\boldsymbol{\theta}^*$ wrt ${\mathcal{P}}$. \end{definition} According to \defref{def:approxteaching}, exact teaching corresponds to constructing a $0$-approximate teaching set for a target classifier (e.g., the decision boundary of a kernel perceptron). We study linear and polynomial kernelized perceptrons in the exact teaching setting. Under some mild assumptions on the smoothness of the data distribution, we establish approximate teaching bound on approximate teaching dimension for Gaussian kernelized perceptron \section{List of Appendices}\label{appendix:table-of-contents} First, we provide the proofs of our theoretical results in full detail in the subsequent sections. We follow these by the experimental evaluations section. The appendices are summarized as follows: \begin{itemize} \item \appref{appendix: polynomial perceptron} provides the proof of \thmref{thm: poly_main_theorem} \item \appref{appendix: motivation} provides the motivations and key insights into \assumref{assumption: polyorthogonal}{assumption: orthogonal}{assumption: bounded cone} \item \appref{appendix: gaussian perceptron} provides relevant results and proofs of \thmref{thm: boundedclassifier} and \thmref{thm: gaussian_main_thm} (Approximate Teaching Set for Gaussian Learner) \item \appref{appendix: experimentals} provides the experimental evaluations for the theoretical results on various datasets \end{itemize}

https://arxiv.org/abs/1810.11439

Hardy-Littlewood-Sobolev and Stein-Weiss inequalities on homogeneous Lie groups

In this note we prove the Stein-Weiss inequality on general homogeneous Lie groups. The obtained results extend previously known inequalities. Special properties of homogeneous norms play a key role in our proofs. Also, we give a simple proof of the Hardy-Littlewood-Sobolev inequality on general homogeneous Lie groups.

\section{Introduction} Historically, in \cite{HL28}, Hardy and Littlewood considered the one dimensional fractional integral operator on $(0,\infty)$ given by \begin{equation}\label{1Doper} T_{\lambda}u(x)=\int_{0}^{\infty}\frac{u(y)}{|x-y|^{\lambda}}dy,\,\,\,\,0<\lambda<1, \end{equation} and proved the following theorem: \begin{thm}\label{1DHLS28} Let $1<p<q<\infty$ and $u\in L^{p}(0,\infty)$ with $\frac{1}{q}=\frac{1}{p}+\lambda-1$, then \begin{equation} \|T_{\lambda}u\|_{L^{q}(0,\infty)}\leq C \|u\|_{L^{p}(0,\infty)}, \end{equation} where $C$ is a positive constant independent of $u$. \end{thm} The $N$-dimensional analogue of \eqref{1Doper} can be written by the formula: \begin{equation}\label{NDoper} I_{\lambda}u(x)=\int_{\mathbb{R}^{N}}\frac{u(y)}{|x-y|^{\lambda}}dy,\,\,\,\,0<\lambda<N. \end{equation} The $N$-dimensional case of Theorem \ref{1DHLS28} was extended by Sobolev in \cite{Sob38}: \begin{thm}\label{THM:HLS} Let $1<p<q<\infty$, $u\in L^{p}(\mathbb{R}^{N})$ with $\frac{1}{q}=\frac{1}{p}+\frac{\lambda}{N}-1$, then \begin{equation} \|I_{\lambda}u\|_{L^{q}(\mathbb{R}^{N})}\leq C \|u\|_{L^{p}(\mathbb{R}^{N})}, \end{equation} where $C$ is a positive constant independent of $u$. \end{thm} Later, in \cite{StWe58} Stein and Weiss obtained the following two-weight extention of the Hardy-Littlewood-Sobolev inequality, which is known as the Stein-Weiss inequality. \begin{thm}\label{Classiacal_Stein-Weiss_inequality} Let $0<\lambda<N$, $1<p<\infty$, $\alpha<\frac{N(p-1)}{p}$, $\beta<\frac{N}{q}$, $\alpha+\beta\geq0$ and $\frac{1}{q}=\frac{1}{p}+\frac{\lambda+\alpha+\beta}{N}-1$. If $1<p\leq q<\infty$, then \begin{equation} \||x|^{-\beta}I_{\lambda}u\|_{L^{q}(\mathbb{R}^{N})}\leq C \||x|^{\alpha}u\|_{L^{p}(\mathbb{R}^{N})}, \end{equation} where $C$ is a positive constant independent of $u$. \end{thm} The Hardy-Littlewood-Sobolev inequality on the Heisenberg group was obtained by Folland and Stein in \cite{FS74}. In \cite{GMS} the authors studied the Stein-Weiss inequality on the Carnot groups. Note that in \cite{HLZ} the authors also proved an analogue of the Stein-Weiss inequality on the Heisenberg groups. In \cite{ZW} author proved Stein-Weiss inequality on product spaces. In \cite{JD} author proved the Stein-Weiss inequality on the Euclidean half-space. In the works \cite{CF}, \cite{FM}, \cite{MW} and \cite{Per} authors were studied the regularity of fractional integrals on Euclidean spaces. In this note we first give a simple proof for the Hardy-Littlewood-Sobolev inequality on general homogeneous groups, recapturing the result of \cite[Theorem 4.1]{RY} where a much heavier machinery was used. In the proof we follow the method of Stein and Weiss, however, special properties of homogeneous norms of the homogeneous Lie groups play a key role in our calculations. Furthermore, in Theorem \ref{stein-weiss3} we establish the Stein-Weiss on general homogeneous groups based on the integral Hardy inequalities established in \cite{RY}. Of course, the obtained result recovers the previously known results of Abelian (Euclidean), Heisenberg, Carnot groups since the class of the homogeneous Lie groups contains those and since we can work with an arbitrary homogeneous quasi-norm. Note that in this direction systematic studies of different functional inequalities on general homogeneous (Lie) groups were initiated by the paper \cite{RSAM}. We refer to this and other papers by the authors (e.g. \cite{RSY1}) for further discussions. We also note that the best constant in the Hardy-Littlewood-Sobolev inequality on the Heisenberg group is now known, see Frank and Lieb \cite{FL12} (in the Euclidean case this was done earlier by Lieb in \cite{Lie83}). The expression for the best constant depends on the particular quasi-norm used and may change for a different choice of the quasi-norm. The main results of this paper are as follows: \begin{itemize} \item {\bf Hardy-Littlewood-Sobolev inequality}: Let $\mathbb{G}$ be a homogeneous group of homogeneous dimension $Q$ and let $|\cdot|$ be an arbitrary homogeneous quasi-norm on $\mathbb{G}$. Let $1<p<q<\infty,\,\,0<\lambda<Q$, $\frac{1}{q}=\frac{1}{p}+\frac{\lambda}{Q}-1$. Then for all $u\in L^{p}(\mathbb{G})$ and $h\in L^{q'}(\mathbb{G})$ we have \begin{equation}\label{EQ:HLSi1} \left|\int_{\mathbb{G}}\int_{\mathbb{G}}\frac{u(y)h(x)}{|y^{-1} x|^{\lambda}}dxdy\right|\leq C\|u\|_{L^{p}(\mathbb{G})}\|h\|_{L^{q'}(\mathbb{G})}, \end{equation} where $C$ is a positive constant independent of $u$ and $h$. For the formulation similar to that of Theorem \ref{THM:HLS} see Theorem \ref{Trsob}. \item {\bf Stein-Weiss inequality}: Let $\mathbb{G}$ be a homogeneous group of homogeneous dimension $Q$ and let $|\cdot|$ be an arbitrary homogeneous quasi-norm on $\mathbb{G}$. Let $0<\lambda<Q$, $1<p\leq q<\infty$, $\alpha<\frac{Q}{p'}$, $\beta<\frac{Q}{q}$, $\alpha+\beta\geq0$, $\frac{1}{q}=\frac{1}{p}+\frac{\alpha+\beta+\lambda}{Q}-1$, where $\frac{1}{p}+\frac{1}{p'}=1$ and $\frac{1}{q}+\frac{1}{q'}=1$. Then we have \begin{equation}\label{EQ:HLSi2} \left|\int_{\mathbb{G}}\frac{u(y)h(x)}{|x|^{\beta}|y^{-1}x|^{\lambda}|y|^{\alpha}}dxdy\right|\leq C\|u\|_{L^{p}(\mathbb{G})}\|h\|_{L^{q'}(\mathbb{G})}, \end{equation} where $C$ is a positive constant independent of $u$ and $h$. For the formulation similar to that of Theorem \ref{Classiacal_Stein-Weiss_inequality} see Theorem \ref{stein-weiss3}. Although \eqref{EQ:HLSi1} is clearly contained in \eqref{EQ:HLSi2}, we still keep them as separate statements since the Hardy-Littlewood-Sobolev inequality \eqref{EQ:HLSi1} allows for a simple proof which is much more transparent than that of the Stein-Weiss inequality \eqref{EQ:HLSi2}. The present proof is also much simpler than the original proof of \eqref{EQ:HLSi1} in \cite{RY}. \end{itemize} Finally, let us note that the heavier machinery developed in \cite{RY} also yielded a differential version of the Stein-Weiss inequality (which may be also called {Stein-Weiss-Sobolev inequality}), however, in a more special case of graded groups as follows (see \cite[Theorem 5.12]{RY} for details and the proof): \begin{itemize} \item {\bf Differential Stein-Weiss (or Stein-Weiss-Sobolev) inequality}: Let $\mathbb{G}$ be a graded Lie group of homogeneous dimension $Q$ and let $|\cdot|$ be an arbitrary homogeneous quasi-norm. Let $1<p,q<\infty$, $0\leq a<Q/p$ and $0\leq b<Q/q$. Let $0<\lambda<Q$, $0\leq \alpha <a+Q/p'$ and $0\leq \beta\leq b$ be such that $(Q-ap)/(pQ)+(Q-q(b-\beta))/(qQ)+(\alpha+\lambda)/Q=2$ and $\alpha+\lambda\leq Q$, where $1/p+1/p'=1$. Then there exists a positive constant $C=C(Q,\lambda, p, \alpha, \beta, a, b)$ such that \begin{equation}\label{HLS_ineq1_grad} \left|\int_{\mathbb{G}}\int_{\mathbb{G}}\frac{\overline{f(x)}g(y)}{|x|^{\alpha}|y^{-1}x|^{\lambda}|y|^{\beta}}dxdy\right|\leq C\|f\|_{\dot{L}^{p}_{a}(\mathbb{G})}\|g\|_{\dot{L}^{q}_{b}(\mathbb{G})} \end{equation} holds for all $f\in \dot{L}^{p}_{a}(\mathbb{G})$ and $g\in \dot{L}^{q}_{b}(\mathbb{G})$, where $\dot{L}^{p}_{a}(\mathbb{G})$ stands for a homogeneous Sobolev space of order $a$ over $L^p$ on the graded Lie group $\mathbb{G}.$ \end{itemize} \section{Stein-Weiss inequality on homogeneous group} \label{SEC:2} Let us recall that a Lie group (on $\mathbb{R}^{N}$) $\mathbb{G}$ with the dilation $$D_{\lambda}(x):=(\lambda^{\nu_{1}}x_{1},\ldots,\lambda^{\nu_{N}}x_{N}),\; \nu_{1},\ldots, \nu_{n}>0,\; D_{\lambda}:\mathbb{R}^{N}\rightarrow\mathbb{R}^{N},$$ which is an automorphism of the group $\mathbb{G}$ for each $\lambda>0,$ is called a {\em homogeneous (Lie) group}. For simplicity, throughout this paper we use the notation $\lambda x$ for the dilation $D_{\lambda}.$ The homogeneous dimension of the homogeneous group $\mathbb{G}$ is denoted by $Q:=\nu_{1}+\ldots+\nu_{N}.$ Also, in this note we denote a homogeneous quasi-norm on $\mathbb{G}$ by $|x|$, which is a continuous non-negative function \begin{equation} \mathbb{G}\ni x\mapsto |x|\in[0,\infty), \end{equation} with the properties \begin{itemize} \item[i)] $|x|=|x^{-1}|$ for all $x\in\mathbb{G}$, \item[ii)] $|\lambda x|=\lambda |x|$ for all $x\in \mathbb{G}$ and $\lambda>0$, \item[iii)] $|x|=0$ iff $x=0$. \end{itemize} Moreover, the following polarisation formula on homogeneous Lie groups will be used in our proofs: there is a (unique) positive Borel measure $\sigma$ on the unit quasi-sphere $ \mathfrak{S}:=\{x\in \mathbb{G}:\,|x|=1\}, $ so that for every $f\in L^{1}(\mathbb{G})$ we have \begin{equation}\label{EQ:polar} \int_{\mathbb{G}}f(x)dx=\int_{0}^{\infty} \int_{\mathfrak{S}}f(ry)r^{Q-1}d\sigma(y)dr. \end{equation} The quasi-ball centred at $x \in \mathbb{G}$ with radius $R > 0$ can be defined by \begin{equation} B(x,R) := \{y \in\mathbb {G} : |x^{-1} y|< R\}. \end{equation} We refer to \cite{FS1} for the original appearance of such groups, and to \cite{FR} for a recent comprehensive treatment. Let us consider the integral operator \begin{equation} I_{\lambda,|\cdot|}u(x)=\int_{\mathbb{G}}\frac{u(y)}{|y^{-1} x|^{\lambda}}dy,\,\,\,0<\lambda<Q. \end{equation} Note that when $Q>\alpha>0$ and $\lambda=Q-\alpha$ we get the Riesz potential $I_{\lambda,|\cdot|}=I_{Q-\alpha,|\cdot|}$. First we give a short proof of a version of the Hardy-Littlewood-Sobolev inequality on $\mathbb{G}$. \begin{thm}\label{Trsob} Let $\mathbb{G}$ be a homogeneous group of homogeneous dimension $Q$ and let $|\cdot|$ be an arbitrary homogeneous quasi-norm on $\mathbb{G}$. Let $1<p<q<\infty,\,\,0<\lambda<Q$, $\frac{1}{q}=\frac{1}{p}+\frac{\lambda}{Q}-1$, and $u\in L^{p}(\mathbb{G})$. Then we have \begin{equation}\label{rieszsobolev} \|I_{\lambda,|\cdot|}u\|_{L^{q}(\mathbb{G})}\leq C \|u\|_{L^{p}(\mathbb{G})}, \end{equation} where $C$ is a positive constant independent of $u.$ \end{thm} \begin{rem} With the assumptions of Theorem \ref{Trsob} and $h\in L^{q'}(\mathbb{G}),$ we have the following Hardy-Littlewood-Sobolev inequality \begin{equation} \left|\int_{\mathbb{G}}\int_{\mathbb{G}}\frac{u(y)h(x)}{|y^{-1} x|^{\lambda}}dxdy\right|\leq C\|u\|_{L^{p}(\mathbb{G})}\|h\|_{L^{q'}(\mathbb{G})}, \end{equation} where $C$ is a positive constant independent of $u$ and $h$. This gives \eqref{EQ:HLSi1}. \end{rem} \begin{proof}[Proof of Theorem \ref{Trsob}] As in the Euclidean case we will show that there is a constant $C>0$, such that \begin{equation}\label{needtoprove} m\{x:|K*u(x)|>\zeta\}\leq C\frac{\|u\|_{L^{p}(\mathbb{G})}^{q}}{\zeta^{q}}, \end{equation} where $m$ is the Haar measure on $\mathbb{G}$, $K(x)=|x|^{-\lambda}$ and $I_{\lambda,|\cdot|}u(x)=K*u(x)$, where $*$ is convolution. This implies inequality \eqref{rieszsobolev} via the Marcinkiewicz interpolation theorem. Let $K(x)=K_{1}(x)+K_{2}(x)$, where \begin{equation}\label{2.7} K_{1}(x):= \begin{cases} K(x),\,\,\,\text{if}\,\,\,|x|\leq\mu, \\ 0,\,\,\,\text{if}\,\,\,|x|>\mu, \end{cases} \text{and}\,\,\,\, K_{2}(x):= \begin{cases} K(x),\,\,\,\text{if}\,\,\,|x|>\mu, \\ 0,\,\,\,\text{if}\,\,\,|x|\leq\mu. \end{cases} \end{equation} Here $\mu$ is a positive constant. We have $I_{\lambda,|\cdot|}u(x)=K*u(x)=K_{1}*u(x)+K_{2}*u(x)$, so \begin{equation}\label{ocenkamery} m\{x:|K*u(x)|>2\zeta\}\leq m\{x:|K_{1}*u(x)|>\zeta\}+m\{x:|K_{2}*u(x)|>\zeta\}. \end{equation} It is enough to prove inequality \eqref{needtoprove} with $2\zeta$ instead of $\zeta$ in the left-hand side of the inequality. Without loss of generality we can assume $\|u\|_{L^{p}(\mathbb{G})}=1$ and by using Chebychev's and Minkowski's inequalities, we get \begin{multline}\label{2.9} m\{x:|K_{1}*u(x)|>\zeta\}\leq\frac{\int_{|K_{1}*u|>\zeta}|K_{1}*u|^{p}dx}{\zeta^{p}}\\ \leq\frac{\|K_{1}*u\|^{p}_{L^{p}(\mathbb{G})}}{\zeta^{p}}\leq\frac{\|K_{1}\|^{p}_{L^{1}(\mathbb{G})}\|u\|^{p}_{L^{p}(\mathbb{G})}}{\zeta^{p}}=\frac{\|K_{1}\|^{p}_{L^{1}(\mathbb{G})}}{\zeta^{p}}. \end{multline} By using \eqref{EQ:polar} and \eqref{2.7}, we compute \begin{multline} \|K_{1}\|_{L^{1}(\mathbb{G})}=\int_{0<|x|\leq\mu}|x|^{-\lambda}dx=\int_{0}^{\mu}r^{Q-1}r^{-\lambda}dR\int_{\mathfrak{S}}|y|^{-\lambda}d\sigma(y)\\ =|\mathfrak{S}| \int_{0}^{\mu}r^{Q-\lambda-1}dr= \frac{|\mathfrak {S}|}{Q-\lambda}(r^{Q-\lambda}|^{\mu}_{0})=\frac{|\mathfrak {S}|}{Q-\lambda} \mu^{Q-\lambda}, \end{multline} where $|\mathfrak{S}|$ is the $Q - 1$ dimensional surface measure of the unit quasi-sphere $\mathfrak{S}$. By using this in \eqref{2.9}, we obtain \begin{equation} m\{x:|K_{1}*u(x)|>\zeta\}\leq \left(\frac{|\mathfrak {S}|}{Q-\lambda}\right)^{p} \frac{\mu^{(Q-\lambda)p}}{\zeta^{p}}. \end{equation} Similarly by using Young's inequality, \eqref{EQ:polar} and the assumptions, we get \begin{multline} \|K_{2}*u\|_{L^{\infty}(\mathbb{G})}\leq\|K_{2}\|_{L^{p'}(\mathbb{G})}\|u\|_{L^{p}(\mathbb{G})}=\left(\int_{\mu}^{\infty}r^{-\lambda p'}r^{Q-1}dr\int_{\mathfrak{S}}|y|^{-\lambda p'}d\sigma(y)\right)^{\frac{1}{p'}}\\ =\left(\frac{|\mathfrak{S}|}{Q-\lambda p'}\right)^{\frac{1}{p'}}\left(\int_{\mu}^{\infty}r^{Q-\lambda p'-1}dr\right)^{\frac{1}{p'}}=\left(\frac{|\mathfrak{S}|}{Q-\lambda p'}\right)^{\frac{1}{p'}} (r^{Q-\lambda p'}|^{\infty}_{\mu})^{\frac{1}{p'}}\\ =\left(\frac{|\mathfrak{S}|}{\lambda p'-Q}\right)^{\frac{1}{p'}}\mu^{-\frac{Q}{q}}, \end{multline} since from the assumptions, we get $\frac{Q-\lambda p'}{p'}=\frac{Q}{p'}-\lambda=Q(1-\frac{1}{p}-\frac{\lambda}{Q})=-\frac{Q}{q}$. Moreover, if $\left(\frac{|\mathfrak{S}|}{\lambda p'-Q}\right)^{\frac{1}{p'}}\mu^{-\frac{Q}{q}}=\zeta$, then $\mu=\left(\frac{|\mathfrak{S}|}{\lambda p'-Q}\right)^{-\frac{q}{Qp'}} \zeta^{-\frac{\theta}{Q}}$, so we have $\|K_{2}*u\|_{L^{\infty}(\mathbb{G})}\leq\zeta$. Thus, we have $m\{x:|K_{2}*u|>\zeta\}=0.$ Combining these facts with \eqref{ocenkamery}, $\|u\|_{L^{p}(\mathbb{G})}=1$ and the assumptions we establish \begin{multline} m\{x:|K*u|>2\zeta\} \leq \left(\frac{|\mathfrak {S}|}{Q-\lambda}\right)^{p}\frac{\mu^{(Q-\lambda)p}}{\zeta^{p}}\\ =\left(\frac{|\mathfrak {S}|}{Q-\lambda}\right)^{p}\left(\frac{|\mathfrak{S}|}{\lambda p'-Q}\right)^{-\frac{q(Q-\lambda)p}{Qp'}}\zeta^{\frac{-(Q-\lambda)pq}{Q}-p} \leq C\zeta^{\frac{-(Q-\lambda)pq}{Q}-p}=C\zeta^{(\frac{\lambda}{Q}-1)pq-p}\\ =C\zeta^{(\frac{1}{q}-\frac{1}{p})pq-p}=C\zeta^{p-q-p}=C\frac{\|u\|^{q}_{L^{p}(\mathbb{G})}}{\zeta^{q}}. \end{multline} For completeness, let us recall two well-known ingredients. \begin{defn}[\cite{steinbook}]\label{defin} Let $ 1\leq p\leq\infty$, $1\leq q<\infty$ and $V:L^{p}(\mathbb{G})\rightarrow L^{q}(\mathbb{G})$ be a operator, then $V$ is called an operator of $\textit{weak type}$ $(p,q)$ if \begin{equation} m\{x:|Vu|>\zeta\}\leq C\left(\frac{\|u\|_{L^{p}(\mathbb{G})}}{\zeta}\right)^{q},\,\,\,\,\zeta>0, \end{equation} where $C$ is a positive constant and independent by $u$. \end{defn} Let us also recall the classical Marcinkiewicz interpolation theorem: \begin{thm} Let $V$ be sublinear operator of weak type $(p_{k}, q_{k})$ with $1 \leq p_{k} \leq q_k< \infty$, $k = 0, 1$ and $q_0 < q_1$. Then $V$ is bounded from $L^{p}(\mathbb{G})$ to $L^{q}(\mathbb{G})$ with \begin{equation} \frac{1}{p}=\frac{1-\gamma}{p_{0}}+\frac{\gamma}{p_{1}},\,\,\,\frac{1}{q}=\frac{1-\gamma}{q_{0}}+\frac{\gamma}{q_{1}}, \end{equation} for any $0<\gamma<1$, namely, \begin{equation} \|Vu\|_{L^{q}(\mathbb{G})}\leq C \|u\|_{L^{p}(\mathbb{G})}, \end{equation} for any $u\in L^{p}(\mathbb{G})$ and $C$ is a positive constant. \end{thm} From assumptions $\frac{1}{q}=\frac{1}{p}+\frac{\lambda}{Q}-1<\frac{1}{p}$, then $q>p$. According to Definition \ref{defin}, $I_{\lambda,|\cdot|}u$ is of weak type $(p,q)$, so by using the Marcinkiewicz interpolation theorem, we prove \eqref{rieszsobolev}. The proof of Theorem \ref{Trsob} is complete. \end{proof} The following statements will be useful to prove the homogeneous group version of the Stein-Weiss inequality (\cite[Theorem B*]{StWe58}). The next proposition is well-known, see e.g. {\cite[Theorem 3.1.39 and Proposition 3.1.35]{FR}} and historical references therein. \begin{prop}\label{prop_quasi_norm} Let $\mathbb{G}$ be a homogeneous Lie group. Then there exists a homogeneous quasi-norm on $\mathbb{G}$ which is a norm, that is, a homogeneous quasi-norm $|\cdot|$ which satisfies the triangle inequality \begin{equation} |x y|\leq |x| + |y|, \,\,\,\forall x, y \in \mathbb{G}. \end{equation} Furthermore, all homogeneous quasi-norms on $\mathbb{G}$ are equivalent. \end{prop} The next theorem is the integral version of Hardy inequalities on general homogeneous groups that will be instrumental in our proof. \begin{thm}[\cite{RY}]\label{integral_hardy} Let $\mathbb{G}$ be a homogeneous group of homogeneous dimension $Q$ and let $1 < p \leq q < \infty$. Let $W(x)$ and $U(x)$, be positive functions on $\mathbb{G}$. Then we have the following properties: (1) The inequality \begin{equation}\label{5.2} \left(\int_{\mathbb{G}}\left(\int_{B(0,|x|)}f(z)dz\right)^{q}W(x)dx\right)^{\frac{1}{q}}\leq C_{1} \left(\int_{\mathbb{G}}f^{p}(x)U(x)dx\right)^{\frac{1}{p}} \end{equation} holds for all $f\geq0$ a.e. on $\mathbb{G}$ if only if \begin{equation}\label{5.2.1} A_{1}:=\sup_{R>0}\left(\int_{\mathbb{G}\setminus B(0,|x|)}W(x)dx\right)^{\frac{1}{q}}\left(\int_{B(0,|x|)}U^{1-p'}(x)dx\right)^{\frac{1}{p'}}<\infty. \end{equation} (2) The inequality \begin{equation}\label{5.4} \left(\int_{\mathbb{G}}\left(\int_{\mathbb{G}\setminus B(0,|x|)}f(z)dz\right)^{q}W(x)dx\right)^{\frac{1}{q}}\leq C_{2} \left(\int_{\mathbb{G}}f^{p}(x)U(x)dx\right)^{\frac{1}{p}}, \end{equation} holds for all $f \geq 0$ if and only if \begin{equation}\label{5.4.1} A_{2}:=\sup_{R>0}\left(\int_{B(0,|x|)}W(x)dx\right)^{\frac{1}{q}}\left(\int_{\mathbb{G}\setminus B(0,|x|)}U^{1-p'}(x)dx\right)^{\frac{1}{p'}}<\infty. \end{equation} (3) If $\{C_i\}^{2}_{i=1}$ are the smallest constants for which \eqref{5.2} and \eqref{5.4} hold, then \begin{equation} A_{i} \leq C_{i} \leq (p')^{\frac{1}{p'}}p^{\frac{1}{q}} A_{i}, \,\,\,i = 1, 2. \end{equation} \end{thm} Now we formulate the Stein-Weiss inequality on $\mathbb{G}$. \begin{thm}\label{stein-weiss3} Let $\mathbb{G}$ be a homogeneous group of homogeneous dimension $Q$ and let $|\cdot|$ be an arbitrary homogeneous quasi-norm on $\mathbb{G}$. Let $0<\lambda<Q$, $1<p<\infty$, $\alpha<\frac{Q}{p'}$, $\beta<\frac{Q}{q}$, $\alpha+\beta\geq0$, $\frac{1}{q}=\frac{1}{p}+\frac{\alpha+\beta+\lambda}{Q}-1$, where $\frac{1}{p}+\frac{1}{p'}=1$ and $\frac{1}{q}+\frac{1}{q'}=1$. Then for $1<p\leq q<\infty$, we have \begin{equation}\label{stein-weiss} \||x|^{-\beta}I_{\lambda,|\cdot|}u\|_{L^{q}(\mathbb{G})}\leq C \||x|^{\alpha}u\|_{L^{p}(\mathbb{G})}. \end{equation} where $C$ is positive constant and independent by $u$. \end{thm} In inequality \eqref{stein-weiss} with $\alpha=0$ we get the weighted Hardy-Littlewood-Sobolev inequality established in \cite[Theorem 4.1]{RY}. Thus, by setting $\alpha=\beta=0$ we get Hardy-Littlewood-Sobolev inequality on the homogeneous Lie groups. In the Abelian (Euclidean) case ${\mathbb G}=(\mathbb R^{N},+)$, we have $Q=N$ and $|\cdot|$ can be any homogeneous quasi-norm on $\mathbb R^{N}$, so with the usual Euclidean distance, i.e. $|\cdot|=\|\cdot\|_{E}$, Theorem \ref{stein-weiss3} gives the classical result of Stein and Weiss (Theorem \ref{Classiacal_Stein-Weiss_inequality}). \begin{proof}[Proof of Theorem \ref{stein-weiss3}] Define \begin{equation} \||x|^{-\beta}I_{\lambda,|\cdot|}u\|^{q}_{L^{q}(\mathbb{G})}=\int_{\mathbb{G}}\left(\int_{\mathbb{G}}\frac{u(y)}{|x|^{\beta}|y^{-1} x|^{\lambda}}dy\right)^{q}dx=I_{1}+I_{2}+I_{3}, \end{equation} where \begin{equation} I_{1}=\int_{\mathbb{G}}\left(\int_{B\left(0,\frac{|x|}{2}\right)}\frac{u(y)}{|x|^{\beta}|y^{-1} x|^{\lambda}}dy\right)^{q}dx, \end{equation} \begin{equation} I_{2}=\int_{\mathbb{G}}\left(\int_{B(0,2|x|)\setminus B\left(0,\frac{|x|}{2}\right)}\frac{u(y)}{|x|^{\beta}|y^{-1} x|^{\lambda}}dy\right)^{q}dx, \end{equation} and \begin{equation} I_{3}=\int_{\mathbb{G}}\left(\int_{\mathbb{G}\setminus B(0,2|x|)}\frac{u(y)}{|x|^{\beta}|y^{-1} x|^{\lambda}dy}\right)^{q}dx. \end{equation} From now on, in view of Proposition \ref{prop_quasi_norm} we can assume that our quasi-norm is actually a norm. \textbf{Step 1.} Let us consider $I_{1}$. By using Proposition \ref{prop_quasi_norm} and the properties of the quasi-norm with $|y|\leq\frac{|x|}{2}$, we get $$|x|=|x^{-1}|=|x^{-1}y y^{-1}|$$ $$\leq |x^{-1} y|+|y^{-1}|=|y^{-1} x|+|y|$$ $$\leq |y^{-1} x|+\frac{|x|}{2}.$$ Then for any $\lambda>0$, we have $$2^{\lambda}|x|^{-\lambda}\geq |y^{-1} x|^{-\lambda}.$$ Therefore, we get \begin{multline} I_{1}=\int_{\mathbb{G}}\left(\int_{B\left(0,\frac{|x|}{2}\right)}\frac{u(y)}{|x|^{\beta}|y^{-1} x|^{\lambda}}dy\right)^{q}dx\leq 2^{\lambda}\int_{\mathbb{G}}\left(\int_{B\left(0,\frac{|x|}{2}\right)}\frac{u(y)}{|x|^{\beta+\lambda}}dy\right)^{q}dx\\ =2^{\lambda}\int_{\mathbb{G}}\left(\int_{B\left(0,\frac{|x|}{2}\right)}u(y)dy\right)^{q}|x|^{-(\beta+\lambda)q}dx. \end{multline} If condition \eqref{5.2.1} in Theorem \ref{integral_hardy} with $W(x)=|x|^{-(\beta+\lambda)q}$ and $U(y)=|y|^{\alpha p}$ in \eqref{5.2} is satisfied, then we have \begin{equation} I_{1}\leq2^{\lambda}\int_{\mathbb{G}}\left(\int_{B(0,\frac{|x|}{2})}u(y)dy\right)^{q}|x|^{-(\beta+\lambda)q}dx \leq C_{1}\||x|^{\alpha}u\|^{q}_{L^{p}(\mathbb{G})}. \end{equation} Let us verify condition \eqref{5.2.1}. So from the assumption we have $\alpha<\frac{Q}{p'}$, then $$\frac{1}{q}=\frac{1}{p}+\frac{\alpha+\beta+\lambda}{Q}-1<\frac{1}{p}+\frac{\frac{Q}{p'}+\beta+\lambda}{Q}-1=\frac{1}{p}+\frac{1}{p'}+\frac{\beta+\lambda}{Q}-1=\frac{\beta+\lambda}{Q},$$ that is, $Q-(\beta+\lambda)q<0$ and by the using polar decomposition \eqref{EQ:polar}: \begin{multline} \left(\int_{\mathbb{G}\setminus B(0,|x|)}W(x)dx\right)^{\frac{1}{q}}=\left(\int_{\mathbb{G}\setminus B(0,|x|)}|x|^{-(\beta+\lambda)q}dx\right)^{\frac{1}{q}}\\ =\left(\int_{R}^{\infty}\int_{\mathfrak{S}}r^{Q-1}r^{-(\beta+\lambda)q}drd\sigma(y)\right)^{\frac{1}{q}} =\left(|\mathfrak{S}|\int_{R}^{\infty}r^{Q-1-(\beta+\lambda)q}dr\right)^{\frac{1}{q}}\leq C R^{\frac{Q-(\beta+\lambda)q}{q}}. \end{multline} Since $\alpha<\frac{Q}{p'}$, we have $$\alpha p(1-p')+Q>\alpha p(1-p')+\alpha p'=\alpha p+\alpha p'(1-p)=\alpha p -\alpha p=0 .$$ So, $\alpha p(1-p')+Q>0$. Then, let us consider \begin{multline} \left(\int_{ B(0,|x|)}U^{1-p'}(x)dx\right)^{\frac{1}{p'}}=\left(\int_{ B(0,|x|)}|x|^{(1-p')\alpha p}dx\right)^{\frac{1}{p'}}\\=\left(\int^{ R}_{0}\int_{\mathfrak{S}}r^{(1-p')\alpha p}r^{Q-1}drd\sigma(y)\right)^{\frac{1}{p'}} \leq C\left(|\mathfrak{S}|\int^{ R}_{0}r^{(1-p')\alpha p+Q-1}dr\right)^{\frac{1}{p'}}\\ \leq C R^{\frac{(1-p')\alpha p+Q}{p'}}=CR^{\frac{Q-\alpha p'}{p'}}. \end{multline} Moreover, the assumptions imply $$A_{1}=\sup_{R>0}\left(\int_{\mathbb{G}\setminus B(0,|x|)}W(x)dx\right)^{\frac{1}{q}}\left(\int_{ B(0,|x|)}U^{1-p'}(x)dx\right)^{\frac{1}{p'}}\leq C R^{\frac{Q-(\beta+\lambda)q}{q}+\frac{Q-\alpha p'}{p'}}$$ $$=C R^{Q(\frac{1}{q}-\frac{1}{p}-\frac{\alpha+\beta+\lambda}{Q}+1)}=C<\infty,$$ where $C=C(\alpha,\beta,p,\lambda)$ is a positive constant. Then by using \eqref{5.2}, we obtain \begin{equation} I_{1}\leq C\int_{\mathbb{G}}\left(\int_{B\left(0,\frac{|x|}{2}\right)}u(y)dy\right)^{q}|x|^{-(\beta+\lambda)q}dx \leq C_{1}\||x|^{\alpha}u\|^{q}_{L^{p}(\mathbb{G})}. \end{equation} \textbf{Step 2.} As in the previous case $I_{1}$, now we consider $I_{3}$. From $2|x|\leq |y|$, we calculate $$|y|=|y^{-1}|=|y^{-1} x x^{-1}|\leq |y^{-1} x|+|x|$$ $$\leq |y^{-1} x|+\frac{|y|}{2},$$ that is, $$\frac{|y|}{2}\leq |y^{-1} x|.$$ Then, if condition \eqref{5.4.1} with $W(x)=|x|^{-\beta q}$ and $U(y)=|y|^{(\alpha+\lambda)p}$ is satisfied, then we have \begin{multline} I_{3}=\int_{\mathbb{G}}\left(\int_{\mathbb{G}\setminus B(0,2|x|)}\frac{u(y)}{|x|^{\beta}|y^{-1}x|^{\lambda}}dy\right)^{q}dx\leq C\int_{\mathbb{G}}\left(\int_{\mathbb{G}\setminus B(0,2|x|)}\frac{u(y)}{|x|^{\beta}|y|^{\lambda}}dy\right)^{q}dx\\ =C\int_{\mathbb{G}}\left(\int_{\mathbb{G}\setminus B(0,2|x|)}u(y)|y|^{-\lambda}dy\right)^{q}|x|^{-\beta q}dx\leq C\||x|^{\alpha}u\|^{q}_{L^{p}(\mathbb{G})}. \end{multline} Now let us check condition \eqref{5.4.1}. We have \begin{multline} \left(\int_{ B(0,|x|)}W(x)dx\right)^{\frac{1}{q}}=\left(\int_{ B(0,|x|)}|x|^{-\beta q}dx\right)^{\frac{1}{q}}\\=\left(\int_{0}^{R}\int_{\mathfrak{S}}r^{-\beta q}r^{Q-1}drd\sigma(y)\right)^{\frac{1}{q}} \leq C R^{\frac{Q-\beta q}{q}}, \end{multline} where $Q-\beta q>0$, and \begin{multline} \left(\int_{\mathbb{G}\setminus B(0,|x|)}U^{1-p'}(x)dx\right)^{\frac{1}{p'}}=\left(\int_{\mathbb{G}\setminus B(0,|x|)}|x|^{(\alpha+\lambda)(1-p')p}dx\right)^{\frac{1}{p'}}\\=\left(\int_{R}^{\infty}\int_{\mathfrak{S}}r^{Q-1}r^{(\alpha+\lambda)(1-p')p}drd\sigma(y)\right)^{\frac{1}{p'}} \leq C R^{\frac{Q-p'(\alpha+\lambda)}{p'}}, \end{multline} where from $\beta<\frac{Q}{q}$, we obtain $Q-p'(\alpha+\lambda)<0$. Combining these facts we have \begin{multline} A_{2}:=\sup_{R>0}\left(\int_{ B(0,|x|)}W(x)dx\right)^{\frac{1}{\theta}}\left(\int_{\mathbb{G}\setminus B(0,|x|)}U^{1-p'}(x)dx\right)^{\frac{1}{p'}}\leq C R^{\frac{Q-p'(\alpha+\lambda)}{p'}+\frac{Q-\beta q}{q}}\\ =C R^{\frac{Q}{p'}-(\alpha+\beta+\lambda)+\frac{Q}{q}}=C R^{Q(\frac{1}{p'}-\frac{\alpha+\beta+\lambda}{Q}+\frac{1}{q})}=C<\infty, \end{multline} where $C=C(\alpha,\beta,p,\lambda)$ is a positive constant. Then we establish \begin{equation} I_{3}=\int_{\mathbb{G}}\left(\int_{\mathbb{G}\setminus B(0,2|x|)}\frac{u(y)}{|x|^{\beta}|y^{-1} x|^{\lambda}}dy\right)^{q}dx\leq C\||x|^{\alpha}u\|^{q}_{L^{p}(\mathbb{G})}. \end{equation} \textbf{Step 3.} Let us estimate $I_{2}$ now. \textbf{Case 1:} $p<q$. From $\frac{|x|}{2}<|y|<2|x|$, we obtain $$\frac{|y^{-1} x|}{2}\leq \frac{|x|+|y|}{2}= \frac{|x|}{2}+\frac{|y|}{2}<\frac{3}{2}|y|,$$ that is, $$|y^{-1} x|<3|y|.$$ For all $\alpha+\beta\geq0$, we have $$|y^{-1} x|^{\alpha+\beta}< 3^{\alpha+\beta}|y|^{\alpha+\beta}= 3^{\alpha+\beta}|y|^{\alpha}|y|^{\beta}\leq 3^{\alpha+\beta}2^{|\beta|}|x|^{\beta}|y|^{\alpha}.$$ Therefore, \begin{multline*} I_{2}=\int_{\mathbb{G}}\left(\int_{B(0,2|x|)\setminus B\left(0,\frac{|x|}{2}\right)}\frac{u(y)}{|x|^{\beta}|y^{-1} x|^{\lambda}}dy\right)^{q}dx\\ \leq C\int_{\mathbb{G}}\left(\int_{B(0,2|x|)\setminus B\left(0,\frac{|x|}{2}\right)}\frac{|y|^{\alpha}u(y)}{|y^{-1} x|^{\alpha+\beta+\lambda}}dy\right)^{q}dx \\ \leq C \int_{\mathbb{G}}\left(\int_{\mathbb{G}}\frac{|y|^{\alpha}u(y)}{|y^{-1}x|^{\alpha+\beta+\lambda}}dy\right)^{q}dx =C\|I_{\lambda+\alpha+\beta,|\cdot|}\tilde{u}\|^{q}_{L^{q}(\mathbb{G})}, \end{multline*} where $\tilde{u}(x)=|x|^{\alpha}u(x)$. By assumption $\frac{1}{q}-\frac{1}{p}=\frac{\lambda+\alpha+\beta}{Q}-1<0$, then $Q>\lambda+\alpha+\beta$ and by using Theorem \ref{Trsob} with $p<q$, we establish \begin{equation} I_{2}\leq C\|I_{\lambda+\alpha+\beta,|\cdot|}\tilde{u}\|^{q}_{L^{q}(\mathbb{G})}\leq C\|\tilde{u}\|_{L^{p}(\mathbb{G})}^{q}=C\||x|^{\alpha}u\|_{L^{p}(\mathbb{G})}^{q}. \end{equation} \textbf{Case 2:} $p=q$. We decompose $I_{2}$ as \begin{equation} I_{2}=\sum_{k\in \mathbb{Z}}\int_{2^{k}\leq |x| \leq 2^{k+1}}\left(\int_{B(0,2|x|)\setminus B\left(0,\frac{|x|}{2}\right)}\frac{u(y)}{|x|^{\beta}|y^{-1} x|^{\lambda}}dy\right)^{p}dx. \end{equation} From $|x| \leq 2|y| \leq 4 |x|$ and $2^{k} \leq |x| \leq 2^{k+1}$, we have $2^{k-1} \leq |y| \leq 2^{k+2}$ and $0 \leq |y^{-1} x| \leq 3|x| \leq 3 \cdot 2^{k+1}$. By using Young's inequality with $\frac{1}{p}+\frac{1}{r}=1+\frac{1}{q}$ (our case $p=q$, hence $r=1$), we calculate \begin{align*} I_{2}&=\sum_{k\in \mathbb{Z}}\int_{2^{k}\leq |x| \leq 2^{k+1}}\left(\int_{B(0,2|x|)\setminus B\left(0,\frac{|x|}{2}\right)}\frac{u(y)}{|x|^{\beta}|y^{-1} x|^{\lambda}}dy\right)^{p}dx\\& =\sum_{k\in \mathbb{Z}}\int_{2^{k}\leq |x| \leq 2^{k+1}}\left(\int_{B(0,2|x|)\setminus B\left(0,\frac{|x|}{2}\right)}\frac{u(y)}{|y^{-1} x|^{\lambda}}dy\right)^{p}\frac{dx}{|x|^{\beta p}}\\& \leq \sum_{k\in \mathbb{Z}} 2^{-\beta p k}\|u\cdot\chi_{\{2^{k-1}\leq |y|\leq 2^{k+2}\}}*|x|^{-\lambda}\|^{p}_{L^{p}(\mathbb{G})}\\& \leq \sum_{k\in \mathbb{Z}} 2^{-\beta p k} \||x|^{-\lambda}\cdot\chi_{\{0\leq |y|\leq 3\cdot2^{k+1}\}}\|^{p}_{L^{1}(\mathbb{G})}\|u\cdot\chi_{\{2^{k-1}\leq |y|\leq 2^{k+2}\}}\|^{p}_{L^{p}(\mathbb{G})}\\& \leq C \sum_{k\in \mathbb{Z}} 2^{(Q-\lambda-\beta)kp}\|u\cdot\chi_{\{2^{k-1}\leq |y|\leq 2^{k+2}\}}\|^{p}_{L^{p}(\mathbb{G})} =C \sum_{k\in \mathbb{Z}} 2^{\alpha kp}\|u\cdot\chi_{\{2^{k-1}\leq |y|\leq 2^{k+2}\}}\|^{p}_{L^{p}(\mathbb{G})}\\& =C \sum_{k\in \mathbb{Z}} \|2^{\alpha (k-1)}u\cdot\chi_{\{2^{k-1} \leq |y|\leq 2^{k+2}\}}\|^{p}_{L^{p}(\mathbb{G})} \leq C\sum_{k\in \mathbb{Z}}\||y|^{\alpha}u\cdot\chi_{\{2^{k-1}\leq |y|\leq 2^{k+2}\}}\|^{p}_{L^{p}(\mathbb{G})}\\& = C\||x|^{\alpha}u\|^{p}_{L^{p}(\mathbb{G})}. \end{align*} Theorem \ref{stein-weiss3} is proved. \end{proof} \begin{rem} With assumptions Theorem \ref{stein-weiss3} and $h\in L^{q'}(\mathbb{G})$, we have the following Stein-Weiss inequality \begin{equation} \left|\int_{\mathbb{G}}\frac{u(y)h(x)}{|x|^{\beta}|y^{-1}x|^{\lambda}|y|^{\alpha}}dxdy\right|\leq C\|u\|_{L^{p}(\mathbb{G})}\|h\|_{L^{q'}(\mathbb{G})}, \end{equation} where $C$ is a positive constant independent of $u$ and $h$. This gives \eqref{EQ:HLSi2}. \end{rem}

https://arxiv.org/abs/math/9805076

An Introduction to Total Least Squares

The method of ``Total Least Squares'' is proposed as a more natural way (than ordinary least squares) to approximate the data if both the matrix and and the right-hand side are contaminated by ``errors''. In this tutorial note, we give a elementary unified view of ordinary and total least squares problems and their solution. As the geometry underlying the problem setting greatly contributes to the understanding of the solution, we introduce least squares problems and their generalization via interpretations in both column space and (the dual) row space and we shall use both approaches to clarify the solution. After a study of the least squares approximation for simple regression we introduce the notion of approximation in the sense of ``Total Least Squares (TLS)'' for this problem and deduce its solution in a natural way. Next we consider ordinary and total least squares approximations for multiple regression problems and we study the solution of a general overdetermined system of equations in TLS-sense. In a final section we consider generalizations with multiple right-hand sides and with ``frozen'' columns. We remark that a TLS-approximation needs not exist in general; however, the line (or hyperplane) of best approximation in TLS-sense for a regression problem does exist always.

\section{Introduction\label{par1}} \setcounter{equation}{0} This (tutorial) paper grew out of the need to motivate the usual formulation of a ``Total Least Squares problem'' and to explain the way it is solved using the ``Singular Value Decomposition''. Although it is an important generalization of (ordinary) least squares and not more difficult to understand, it is hardly treated in numerical textbooks up to now. In the well-known book of Golub \& Van Loan \cite{gvl} and in \cite{vanhuffel}, the problem is formulated as follows: \par\noindent \beq{problem} \matrix{ \mbox{\sl Given a matrix $A\inI\!\!R^{m\times n} $ with $m>n$ and a vector $\vek b\inI\!\!R^m$,}\cr \mbox{\sl find residuals $E\inI\!\!R^{m\times n} $ and $\vek r\inI\!\!R^m$ that minimize }\cr \mbox{\sl the Frobenius norm $\|(\,E\,|\,\vek r\,)\|_F$ subject to the condition $\vek b+\vek r\in Im(A+E)$. }} \end{equation} \par\noindent It is proposed as a more natural way to approximate the data if both $A$ and $b$ are contaminated by ``errors''. In our opinion, it is not made clear sufficiently well, why this indeed is a natural generalization of the standard least squares problem and why it makes sense to study it. On the other hand, the classroom note of Y. Nievergelt \cite{niever} gives a very nice introduction, but it tells only half of the story in that it considers (multiple) regression only. \par In this note, we shall give a unified view of ordinary and total least squares problems and their solution. As the geometry underlying the problem setting greatly contributes to the understanding of the solution, we shall introduce least squares problems and their generalization via interpretations in both column space and (the dual) row space and we shall use both approaches to clarify the solution. After a study of the least squares approximation for simple regression in section \ref{par2}, we introduce the notion of approximation in the sense of ``Total Least Squares (TLS)'' for this problem in section \ref{par3}. In the next section we consider ordinary and total least squares approximations for multiple regression problems and in section \ref{par5} we study the solution of a general overdetermined system of equations in TLS-sense. In a final section we consider generalizations with multiple right-hand sides and with ``frozen'' columns. We remark that a TLS-approximation needs not exist in general; however, the line (or hyperplane) of best approximation in TLS-sense for a regression problem does exist always. \par As numerical algorithms such as the QR-factorization and the Singular Value Decomposition (SVD) are relatively well-known and nicely implemented in a package like MATLAB, we shall not consider numerical algorithms to compute the solutions effectively. \section{Primal vs.~dual approach} \setcounter{equation}{0} To make clear how both column- and row-space arguments can be used to derive the solution of a least squares problem, we consider least squares in one dimension: \par\noindent \centerline{\sl Given $m$ points $\{x_i~|~ i=1,\,\cdots\,,\,m\}$, find $z\inI\!\!R$ that minimizes the quadratic functional} \beq{intro1} f(z):=\sum_{i=1}^m\,(x_i-z)^2 \,. \end{equation} The function $z\mapsto f(z)$ is a parabola. When we shift its center to the average $\overline x:={1\over m}\sum_{i=1}^m\,x_i$\,, \beq{intro1a} f(z)=\sum_{i=1}^m\,(x_i-z)^2 =\sum_{i=1}^m\,\{\,(x_i-\overline x)^2 +2(x_i-\overline x)(\overline x - z)+(\overline x-z)^2\,\}\,, \end{equation} we see that the sum of double products vanishes. Hence, the average $\overline x$ is the unique minimizer. \par\noindent In the dual approach we consider the data as one point in $\vek x\inI\!\!R^m$. The functional $f(z)$ then measures the square of the Euclidean distance to the point $z\vek e$, \beq{intro2} f(z)=\|\,\vek x - z\vek e\,\|_2^2\,,~~~\mbox{where}~~~ \vek x :=\left(\matrix{~x_1~\cr x_2 \cr \vdots \cr x_m}\right) ~~\mbox{and}~~ \vek e :=\left(\matrix{~1~\cr 1 \cr \vdots \cr 1}\right)\,. \end{equation} \hrule \begin{figure}[htb] \begin{center} \begin{picture}(280,115) \setlength{\unitlength}{.1mm} \put(0,0){\line(2,1){800}} \thicklines \put(200,100){\vector(1,0){500}} \put(202,101){\vector(2,1){400}} \put(600,300){\vector(1,-2){100}} \put(200,60){$\scriptstyle O$} \put(700,60){$\scriptstyle \vek x$} \put(810,400){$\scriptstyle span\{\vek e\}$} \put(560,320){$\scriptstyle z\,\vek e$} \put(650,200){$\scriptstyle \vek x -z\,\vek e$} \end{picture} \end{center} \caption{Vector $\vek x$, its orthogonal projection on $span\{\vek e\}$ and the residual vector $\vek x -z\,\vek e$ in the dual approach.\label{fig1a}} \end{figure} \hrule \par\noindent From fig. \ref{fig1a}, which shows the plane in $I\!\!R^m$ spanned by $\vek x$ and $\vek e$, we find the orthogonal projection of $\vek x$ on $span\{\vek e\}$ as minimizer, \beq{intro3} \overline x={\vek x^T\,\vek e \over \vek e^T\,\vek e }= {1\over m}\sum_{i=1}^m\,x_i\,. \end{equation} We see that both the primal and the dual approach provide the solution in different ways. In the primal approach we use the fact that linear terms vanish by a shift towards the average. In the dual approach we use an orthogonality argument. \section{Simple regression\label{par2}} \setcounter{equation}{0} In the plane $I\!\!R^2$ we are given $m$ data points (abscissae and ordinates) \beq{sregres0} \{(x_i\,,\,y_i)\inI\!\!R^2~|~ i=1,\,\cdots\,,\,m\} \end{equation} that should satisfy the linear (affine) relation $y(x)=a+bx$; find the parameters $a$ and $b$ that provide a ``best fit'', minimizing the sum of squares of the residuals \beq{sregres1} f(a\,,\,b):=\sum_{i=1}^m\,(y_i-a-b\,x_i)^2 \,. \end{equation} We can interpret this as searching the line $\ell:=\{(x,y)\inI\!\!R^2~|~ y=a +b\,x\}$ ``nearest'' to the datapoints, minimizing vertical distances and making the tacit assumption that {\sl model errors in the data-model $y=a+bx$ are confined to the observed $y$-coordinates}, as depicted in fig. \ref{fig1}. \ifnum\themachine=0 \begin{figure}[htb] \hbox{\hskip 37mm \vbox{\vskip 165pt{\special{illustration tlsfig1a.eps scaled 500}}\vskip -25pt}} \caption{\label{fig1} Simple linear regression; {\rm distances are measured along the $y$-axis.}} \end{figure} \else \begin{figure}[htb] \vskip -20pt \hskip43mm\psfig{figure=tlsfig1a.eps,width=210pt,height=160pt} \vskip-15pt \caption{\label{fig1} Simple linear regression; {\rm distances are measured along the $y$-axis.}} \end{figure} \fi \par\noindent Analogously to (\ref{intro1a}) using the centroid $\overline{ \vek z}:=(\overline x\,,\,\overline y)^T=( \,{1\over m}\sum_{i=1}^m\,x_i \,,\,{1\over m}\sum_{i=1}^m\,y_i\,)^T $ we rewrite $f$ and find as before, that the double products vanish, \beq{sregres2} \begin{array}{r c l} \displaystyle f(a\,,\,b):=\sum_{i=1}^m\,(y_i-a-b\,x_i)^2 &=&\displaystyle \sum_{i=1}^m\,\Big(y_i-\overline y +b\,(x_i-\overline x){\Big )}^2 +m\,(\overline y - a-b\,\overline x)^2 \cr ~&\ge&\displaystyle \sum_{i=1}^m\,\Big(y_i-\overline y + b\,(x_i-\overline x)\Big)^2\,,~~~~\forall ~a,\,b\,, \cr \end{array} \end{equation} with equality if $\overline y = a+b\,\overline x$. This implies that the centroid is located on the line: $\overline{\vek z}\in\ell$. Eliminating $a$ it remains to minimize a function of $b$ alone, which is a parabola. Hence the minimizer of (\ref{sregres1}) is \beq{sregres3} b={\sum_{i=1}^m\,(\overline x-x_i)(\overline y-y_i)\over \sum_{i=1}^m\,(\overline x-x_i)^2} ~~~~\mbox{and}~~~~a=\overline y -b\,\overline x\,. \end{equation} \par\noindent {\bf In the dual} approach in $I\!\!R^m$ we interpret $x_i$ and $y_i$ as components of vectors $\vek x$ and $\vek y\in I\!\!R^m$\,, \beq{sregres3a} \vek x :=\left(\matrix{~x_1~\cr x_2 \cr \vdots \cr x_m}\right) ~~~~\vek y :=\left(\matrix{~y_1~\cr y_2 \cr \vdots \cr y_m}\right) ~~~~ \vek e :=\left(\matrix{~1~\cr 1 \cr \vdots \cr 1}\right)~~~~\mbox{and}~~~~ A:=\left(~\vek e~|~\vek x\,\right)\inI\!\!R^{m\times 2}\,. \end{equation} In this setting the functional $f$ measures the square of the distance from $\vek y$ to a linear combination of $\vek e$ and $\vek x$, \beq{sregres4} f(a,\,b)=\|\,\vek y - a\,\vek e-b\,\vek x\,\|_2^2= \|\,\vek y - A\, {a\choose b} \, \|_2^2\,. \end{equation} As in (\ref{intro3}) it is minimized by the orthogonal projection of $\vek y$ on the span of $\vek x$ and $\vek e$ \beq{sregres5} f~\mbox{minimal}~~~~\iff ~~~~\vek y - A\, {a\choose b} ~\perp~ \mbox{Im}(A)\,. \end{equation} If the rank of $A$ is maximal, the solution can be computed, see \cite{gvl}, from the {\sl Normal Equations} or better by an {\sl Orthogonal Factorization} \beq{sregres6} A^TA {a\choose b}=A^T \vek y~~~~{\rm or~better}~~~~ A=QR~~~ {\rm and} ~~~ R {a\choose b}=Q^T\vek y\,. \end{equation} Otherwise we can use the {\sl Singular Value Decomposition} \beq{sregres8} A=U\,\Sigma\,V^T~~~~ {\rm and} ~~~~ {a\choose b}=V\,\Sigma^\dagger\,U^T\vek y\,. \end{equation} \section{Total Least Squares for simple regression\label{par3}} \setcounter{equation}{0} In (\ref{sregres1}) and fig. \ref{fig1} we considered the problem of locating a line nearest to a collection of points, where the distance is measured along the $y$-axis. It looks ``more natural'' to use the (shorter) true Euclidean distance instead, as drawn in fig. \ref{fig2}, which yields the line of {\it Total Least Squares}. \ifnum\themachine=0 \begin{figure}[htb] \hbox{\hskip 37mm \vbox{\vskip 160pt{\special{illustration tlsfig1c.eps scaled 500}}\vskip -25pt}} \caption{\label{fig2}Line of {\it Total Least Squares}: {\rm Model errors are distributed over the $x$- and $y$-coordinates.}} \end{figure} \else \begin{figure}[htb] \vskip -20pt \hskip43mm\psfig{figure=tlsfig1c.eps,width=210pt,height=160pt} \vskip-15pt \caption{\label{fig2}Line of {\it Total Least Squares}: {\rm Model errors are distributed over the $x$- and $y$-coordinates.}} \end{figure} \fi \par\noindent So we consider the {\sl Total Least Squares} problem of finding the line $\ell$ that minimizes the sum of squares of {\bf true} distances: \beq{tlstwo1} f(\ell)~:=~\sum_{i=1}^m\,dist(\,(x_i\,,\,y_i)\,,\, \ell\,)^2 \end{equation} Instead of asking for a line $y=ax+b$, we use the more symmetric form \beq{tlstwo1a} \ell=\{(x,y)\inI\!\!R^2~|~ a+r_1x+r_2y=0\}=\vek w+\vek r^\perp,~~~{\rm with}~~~ \|\vek r\|^2=r_1^2+r_2^2=1, \end{equation} where $\vek w$ is an arbitrary point on the line $\ell$, i.e. $a+r_1w_1+r_2w_2=0$. With this parametrization of $\ell$ we accept the possibility, that $r_2$ may become zero, and hence, that the line cannot be recast in the form $y=\alpha +\beta x$. In the description $\ell=\vek w+\vek r^\perp$, where $\vek r$ is of unit length, the distance from a point $\vek z$ to $\ell$ is given by, see fig. \ref{fig3}, \beq{tlstwo2} ~~~~~dist(\vek z,\ell)={|\vek r^T(\vek z -\vek w)|} ~~~{\rm where}~~~ \ell=\vek w +\vek r^\perp=\{\vek z \in I\!\!R^2 ~|~ \vek r^T(\vek z - \vek w)=0\,\} ~~ {\rm and} ~~\|\vek r\|=1\,. \end{equation} \begin{figure}[htb] \hrule \vspace*{15pt} \begin{center}\setlength{\unitlength}{1pt} \begin{picture}(280,140)(-130,0) \setlength{\unitlength}{.1mm} \put(500,0){\line(-4,3){650}} \thicklines \put(400,75){\vector(-3,4){313}} \put(0,0){\vector(3,4){150}} \put(0,375){\vector(3,4){87}} \put(400,75){\vector(-4,3){400}} \put(0,-35){$\scriptstyle\bf O$} \put(150,170){$\scriptstyle\vek r$} \put(410,75){$\scriptstyle\vek w$} \put(-420,420){$\scriptstyle\ell=\{\vek z \in I\!\!R^2~\vert~(\vek z - \vek w,\vek r)=0\}$} \put(87,500){$\scriptstyle\vek x$} \put(250,290){$\scriptstyle\vek x-\vek w$} \put(40,400){$\scriptstyle dist(\vek x,\ell)= {|\vek r^T(\vek x -\vek w)|}$} \end{picture} \end{center} \caption{\label{fig3}The line $\ell$ in the plane is given as the line through the vector $\vek w$ orthogonal to the vector $\vek r$ of unit length. For a given vector $\vek x$ the difference vector $\vek x -\vek w$ is drawn together with its projection along the line $\ell$ and its orthogonal complement.} \vspace*{10pt} \hrule \end{figure} \par\noindent Hence the TLS problem is to find $\vek r$ and $\vek w$ that minimize the functional \beq{tlstwo3} I(\vek r,\vek w):=\sum_{i=1}^m~ \left(\vek r^T(\vek z_i -\vek w)\right)^2= \sum_{i=1}^m ~\left(r_1\,(x_i-w_1)+r_2\,(y_i-w_2)\right)^2 \end{equation} where $$ \vek z_i={x_i\choose y_i}~~~~\mbox{and}~~~~ \vek r={r_1\choose r_2}\,,~~~~\|\vek r\|^2=r_1^2+r_2^2=1\,. $$ Making the shift to the centroid, as in (\ref{sregres2}) and (\ref{intro1a}), we find again, that the sum of double products vanishes, \beq{tlstwo4} \begin{array}{l c l} I(\vek r,\vek w)&=&\displaystyle \sum_{i=1}^m ~\left(\vek r^T (\vek z_i-\vek w)\vruimte{1.1em}{.6em}\right)^2 \vruimte{0pt}{12pt} \\ &=&\displaystyle \sum_{i=1}^m ~\left(\vek r^T (\vek z_i-\overline{\vek z})\right)^2\, +\sum_{i=1}^m ~2\,\vek r^T (\vek z_i-\overline{\vek z})\, \vek r^T (\overline{\vek z}-\vek w) +\,m(\vek r^T (\overline{\vek z}-\vek w))^2\vruimte{1.9em}{1.6em} \\ &=&\displaystyle I(\vek r,\overline{\vek z}) +\,m(\vek r^T (\overline{\vek z}-\vek w))^2~\ge~ I(\vek r,\overline{\vek z}) \,.\\ \end{array} \end{equation} Clearly, the centroid $\vek{\overline z}:=(\overline x,\overline y)^T$ minimizes the functional $\vek w\mapsto I(\vek r,\,\vek w\,)$ for every $\vek r\inI\!\!R^2$. This implies, that the minimizing line $\ell=\vek{\overline z}+\vek r^\perp$ passes through the centroid (as did the line of simple regression) and that we are left with the reduced minimization problem: \\{\sl Find the vector $\vek r$ with $\|\vek r\|_2=1$ minimizing} \beq{tlstwo5} I(\vek r,\,\vek{\overline z})= \sum_{i=1}^m ~\left(r_1\,(x_i-\overline x)+r_2\,(y_i-\overline y) \vruimte{1.1em}{.2em}\right)^2= \|B\vek r\|_2^2=\vek r^T\,B^T\,B\,\vek r\,, \end{equation} where $B\in I\!\!R^{m\times 2}$ is the matrix \beq{tlstwo5a} B:=\left(\,\vek x-\overline x\,\vek e~|~\vek y-\overline y\,\vek e\,\right)= \left(\matrix{x_1-\overline x & y_{1}-\overline y\cr x_2-\overline x & y_{2}-\overline y\cr \vdots & \vdots\cr x_m-\overline x & y_m-\overline y}\right) \,. \end{equation} The problem of minimizing $\|\,B\,\vek r~\|_2^2$ subject to $\|\,\vek r\,\|_2=1$ is solved by the Singular Value Decomposition of $B$, $$ B=U\,\Sigma\,V^T~~~~\mbox{with}~~~~\Sigma= \left(\matrix{~\sigma_1~&~0~\cr~0~&~\sigma_2~}\right)~~~ \mbox{and}~~~\sigma_1\ge\sigma_2\,. $$ The solution vector $\vek r$ of (\ref{tlstwo5}) is the right singular vector of $B$ corresponding to the smaller singular value of $B$\,. So we conclude: \begin{list}{\alph{enumi}.}{\leftmargin15pt \usecounter{enumi}} \item The solution always exists and is given by the line through the centroid orthogonal to the subdominant singular vector of $B$. \item As $r_2$ can be zero, the solution needs not be expressible in the form $y=\alpha +\beta x$. \item The solution is unique iff $\sigma_1\ne\sigma_2\,.$ \item The shift (\ref{tlstwo4}) to the centroid $\vek z\in\ell$ is the key in finding the solution, as shown in \cite{niever}. \end{list} \ifnum\themachine=0 \begin{figure}[htb] \hbox{\hskip 37mm \vbox{\vskip 160pt{\special{illustration tlsfig1c.eps scaled 500}}\vskip -15pt}} \caption{\label{fig2a}Components $(f_i,g_i)$ are the best approximations of $(x_i,y_i)$ on the line $a+r_1x+r_2y=0$\,.} \vskip-10pt \end{figure} \else \begin{figure}[htb] \vskip -20pt \hskip43mm\psfig{figure=tlsfig1c.eps,width=210pt,height=160pt} \vskip-15pt \caption{\label{fig2a}Components $(f_i,g_i)$ are the best approximations of $(x_i,y_i)$ on the line $a+r_1x+r_2y=0$\,.} \end{figure} \fi \par\noindent {\bf In the dual formulation} we consider the vectors $\vek x$, $\vek y$ and $\vek e$ as in (\ref{sregres3a}) and we describe the line $\ell$ as in (\ref{tlstwo1a}) by $\ell:=\{(\xi,\eta)~|~ a +r_1 \xi+r_2 \eta=0\}$. For $i=1\,\cdots\,m$ we denote by $(f_i,g_i)$ the point on $\ell$ nearest to $(x_i,y_i)$, see fig.~\ref{fig2a}, and by $(\overline f,\overline g):={1\over m}\sum_{i=1}^m (x_i,y_i)$ we denote their average. We define the vectors of first and second components $\vek f$, $\vek g\in I\!\!R^m$, $$ \vek f:=(f_1\,,\,f_2\,,\,\cdots\,,\,f_m)^T~~~ {\rm and} ~~~ \vek g:=(g_1\,,\,g_2\,,\,\cdots\,,\,g_m)^T. $$ These vectors clearly satisfy the relation $a\,\vek e+r_1\,\vek f+r_2\, \vek g=0$. So we can rephrase the minimization problem (\ref{tlstwo1}) as the quest for vectors $\vek f$ and $\vek g$ that minimize the sum of squares of distances \beq{tlstwo6} \begin{array}{r c l} I(a,\vek r):=\sum_{i=1}^m (x_i-f_i)^2 &+&\sum_{i=1}^m (y_i-g_i)^2~=~ \|\,\vek x-\vek f\,\|^2_2+\|\,\vek y-\vek g\,\|^2_2\cr & &\mbox{subject to}~~~~ a\,\vek e+r_1\,\vek f+r_2 \,\vek g=0\,,~~~r_1^2+r_2^2=1.\vruimte{1.9em}{0em} \end{array} \end{equation} Decomposing the vectors in their components in $span\{\vek e\}$ and in the orthogonal complement $\vek e^\perp$ we obtain \beq{tlstwo6a} I(a,\vek r)=\|\,\vek x-\vek f-(\overline x-\overline f)\vek e\,\|^2_2+ \|\,\vek y-\vek g-(\overline y-\overline g)\vek e\,\|^2_2+ m(\overline x-\overline f)^2+m(\overline y-\overline g)^2\,. \end{equation} The contributions from the parts in $span\{\vek e\}$ are minimized by the choice $\overline f=\overline x$ and $\overline g=\overline y$ and the subsidiary condition implies $a+r_1\overline x +r_2\overline y=0$ for that choice. Choosing $\widetilde \vek f:=\vek f-\overline x\,\vek e$ and $\widetilde\vek g :=\vek g-\overline y\,\vek e$ we are left with the problem to minimize in $\vek e^\perp$ the functional: \beq{tlstwo6b} \|\,\vek x-\overline x\,\vek e - \widetilde\vek f\,\|^2_2+ \|\,\vek y-\overline y\,\vek e -\widetilde\vek g\,\|^2_2~~~~\mbox{subject to}~~~~ r_1\,\widetilde\vek f+r_2\,\widetilde\vek g=\vek 0\,. \end{equation} It is not necessary to impose the condition $\widetilde \vek f\,,\,\widetilde \vek g\in\vek e^\perp$, since it is automatically satisfied by the minimizer, because $\vek x-\overline x\,\vek e$ and $\vek x-\overline x\,\vek e$ satisfy this condition. In matrix notation with $B:=\left(\,\vek x -\overline x\,\vek e~|~\vek y - \overline y\,\vek e\,\right)$ and $E:=\left(\,\widetilde \vek f~|~\widetilde \vek g\,\right)$ this minimization problem takes the form \beq{tlstwo7} \mbox{minimize}~~~~~\|\,B-E\,\|^2_F~~~~~~\mbox{subject to}~~~~rank(E)=1\,. \end{equation} From the Singular Value Decomposition of $B$, $$B=\sigma_1\,\vek u_1\,\vek v_1^T+\sigma_2\,\vek u_2\,\vek v_2^T ~~~~\mbox{we find} ~~~~E=\sigma_1\,\vek u_1\, \vek v_1^T\,,~~~~\mbox{provided}~~~~\sigma_1>\sigma_2\,. $$ Hence the total least squares solution is (as before) given by, $$ E\,\vek v_2=\vek 0~~~~\mbox{implying}~~~~\vek r =\vek v_2\,. $$ There is a difference in flavour between both approaches. Whereas the primal formulation (\ref{tlstwo5}) directly produces the minimizing vector, the dual approach (\ref{tlstwo7}) takes a roundabout. The latter provides a minimizing {\sl matrix} $E$; the parameters of the line are found only afterwards as the coefficients in the linear combination of the columns of $E$ that equals zero. \section{Multiple regression\label{par4}} \setcounter{equation}{0} The extension of ordinary and total least squares to multiple regression is almost straightforward. As most ideas in 2D-regression easily carry over, we can be brief about it. We are given the cloud of $m$ datapoints in $I\!\!R^n$ (each point consisting of an ``abscissa'' in $I\!\!R^{n-1}$ and an ordinate in $I\!\!R$), \beq{mregres0a} \{\vek z_i:=(x_1^{(i)},\cdots,x_{n-1}^{(i)},y_i)^T\,\in\,I\!\!R^n ~|~ i=1,\,\cdots\,,\,m\}\,, \end{equation} that should satisfy the linear (affine) model $y(x_1\,\cdots\,x_{n-1})= c_0 + c_1 x_1 + c_2 x_2+\cdots+c_{n-1} x_{n-1}\,.$ In ordinary least squares the parameters are determined by minimizing the functional $J$, \beq{mregres1} J(\vek c):=\sum_{i=1}^m~(y_i-c_0-c_1x_1^{(i)}- \cdots-c_{n-1}x_{n-1}^{(i)})^2\,, ~~~\vek c:=(\,c_{0}\,,\,\cdots\,,\, c_{n-1}\,)^T\,. \end{equation} and we can interpret this as the search for the best fitting hyperplane in $I\!\!R^n$\,, \beq{mregres0} \{ (x_1\,,\,\cdots\,,\,x_{n-1}\,,\,y)^T\inI\!\!R^n\,| \,y=c_0 + c_1 x_1 + c_2 x_2+\cdots+c_{n-1} x_{n-1}\,\}. \end{equation} As in (\ref{sregres2}), the double products vanish by a shift of the center to the centroid, implying $$ J(\vek c)\ge \sum_{i=1}^m~\,\left(y_i-\overline y- c_1(x_1^{(i)}-\overline x_{1})- \cdots-c_{n-1}(x_{n-1}^{(i)}-\overline x_{n-1})\,\right)^2 $$ with equality if $\overline y=c_0+c_1\,\overline x_1+\cdots+c_{n-1}\,\overline x_{n-1}$. Hence, the centroid is in the hyperplane. However, more than one unknown parameter is left and the easy argument of (\ref{sregres3}) cannot be applied directly. On the other hand, the dual approach (in ``column space'') (\ref{sregres3a}-\ref{sregres5}) is straightforward and provides the solution easily. Defining vectors $\vek x_k$ and $\vek y\inI\!\!R^m$ and the matrix $A\inI\!\!R^{m\times n}$, $$ \vek x_k :=\left(\matrix{~x_k^{(1)}~\cr x_k^{(2)} \cr \vdots \cr x_k^{(m)}}\right),~~~ {\vek y}:=\left(\matrix{y_{1}\cr y_{2}\cr\vdots\cr y_m\cr}\right),~~~ {\rm and} ~~~ A:=\left(\vek e\,|\,\vek x_1\,|\,\cdots\,|\,\vek x_{n-1}\right) =\left(\matrix{1 & x_1^{(1)}&\cdots&x_{n-1}^{(1)}\cr 1 & x_1^{(2)}&\cdots&x_{n-1}^{(2)}\cr \vdots & \vdots&~&\vdots\cr 1 & x_1^{(m)}&\cdots&x_{n-1}^{(m)}}\right) $$ the functional (\ref{mregres1}) takes the form: \beq{mregres2} J(\vek c)=\|\,\vek y-c_0\vek e-\cdots-c_{n-1}\vek x_{n-1}\,\|^2= \|\vek y-A\vek c\|_2^2\,. \end{equation} As in (\ref{intro3}) and (\ref{sregres5}) it is minimized by the orthogonal projection of $\vek y$ on the span of $\vek x_1\,\cdots\,\vek x_{n-1}$ and $\vek e$, i.e. on $Im(A)$, \beq{mregres3} f~\mbox{minimal}~~~~\iff ~~~~\vek y - A\, \vek c ~\perp~ \mbox{Im}(A)\,. \end{equation} As before, if the rank of $A$ is maximal, the solution can be computed from the {\sl Normal Equations} or better by an {\sl Orthogonal Factorization}, see \cite{gvl}, \beq{mregres4} A^TA \vek c =A^T \vek y~~~~{\rm or~better}~~~~ A=QR~~~ {\rm and} ~~~ R \vek c=Q^T\vek y\,. \end{equation} Otherwise we can use the {\sl Singular Value Decomposition} \beq{mregres5} A=U\,\Sigma\,V^T~~~~ {\rm and} ~~~~ \vek c=V\,\Sigma^\dagger\,U^T\vek y\,. \end{equation} \par\noindent {\bf The total least squares approximation} minimizes the sum of squares of {\it true} distances. We do not attribute a special position to the $y$-coordinate and describe the hyperplane in $I\!\!R^n$, as in (\ref{tlstwo1a}), by $\vek w + \vek r^\perp$. The functional to minimize is: \beq{mregres6} I(\vek r,\vek w):=\sum_{i=1}^m~ \left(\vek r^T(\vek z_i -\vek w)\right)^2= \sum_{i=1}^m \left(\vek r^T (\vek z_i -\overline{\vek z})\right)^2 +m(\vek r^T (\overline{\vek z}-\vek w))^2 \end{equation} subject to $\|\vek r\|=1\,.$ Since the double products in the second right-hand side cancel, the centroid (again) is in the hyperplane and it minimizes (\ref{mregres6}) for all $\vek r$. We are left with the reduced minimization problem, to find $\vek r$ with $\|\vek r\|_2=1$ minimizing \beq{mregres7} I(\vek r,\,\vek{\overline z})= \|B\vek r\|_2^2 \,,~~~~\mbox{with}~~~~ B:=\left(\matrix{x_1^{(1)}-\overline x_1 &\cdots& x_{n-1}^{(1)}-\overline x_{n-1} & y_{1}-\overline y\cr x_1^{(2)}-\overline x_1 &\cdots& x_{n-1}^{(2)}-\overline x_{n-1} & y_{2}-\overline y\cr \vdots &\ &\vdots& \vdots\cr x_1^{(m)}-\overline x_1 &\cdots& x_{n-1}^{(m)}-\overline x_{n-1} & y_m-\overline y}\right) \,. \end{equation} The solution vector $\vek r$ is the right singular vector of $B$ corresponding to the smallest singular value of $B$. We conclude: \begin{list}{\alph{enumi}.}{\leftmargin15pt \usecounter{enumi}} \item A solution always exists; it is given by the hyperplane through the centroid and orthogonal to the right singular vector belonging to the smallest singular value of matrix $B$. It is not expressible in the form (\ref{mregres0}) if $r_n=0$. \item The solution is unique, iff $\sigma_{n-1}>\sigma_n\,.$ \item The shift of (\ref{mregres6}) to the centroid $\vek z\in\ell$ is the key in finding the solution. \end{list} \par\noindent {\bf In the dual approach} we again consider the hyperplane (\ref{mregres0}), but now the $y$-coordinate has no special position in the defining equation, \beq{mregres8a} \{ (x_1\,,\,\cdots\,,\,x_{n-1}\,,\,y)^T\inI\!\!R^n\,|\, c_0 + c_1 x_1 + c_2 x_2+\cdots+c_{n-1} x_{n-1}+c_n y=0\,\}\,; \end{equation} instead of $c_n = -1$ we require $\sum_{i=1}^n c_i^2=1$. We choose (for each $i$) the point $(f_1^{(i)}\,,\,\cdots\,,\,f_{n-1}^{(i)}\,,\,g_i)^T$ on this hyperplane nearest to the datapoint $\vek z_i$, $(i=1\cdots m)$. The first, second, etc.\ coordinates of these points form in $I\!\!R^m$ the vectors $\vek f_k$ ($k=1\,\cdots\, n-1$) and $\vek g$, $$ \vek f_k=(f_k^{(1)}\,,\,f_k^{(2)}\,,\,\cdots\,,\,f_k^{(m)})^T~~~ {\rm and} ~~~ \vek g=(g_1\,,\,g_2\,,\,\cdots\,,\,g_m)^T\,, $$ which clearly satisfy the relation $c_0\vek e+c_1\vek f_1+\cdots+c_{n-1}\vek f_{n-1}+c_n\vek g=0$\,. The minimization of the sum of squares of distances from the datapoints $\vek z_i$ to the hyperplane can now be reformulated as the problem of finding vectors $\vek f_k$ ($k=1\,\cdots\, n-1$) and $\vek g$ in $I\!\!R^m$ that minimize the functional \beq{mregres8} \|\,\vek y-\vek g\,\|^2_2\,+\,\sum_{k=1}^{n-1}\, \|\,\vek x_k-\vek f_k\,\|^2_2 ~~~~\mbox{ subject to}~~ c_0\,\vek e+c_1\,\vek f_1+\cdots+c_{n-1}\vek f_{n-1}+c_n\,\vek g=0\,, \end{equation} where $\sum_{k=1}^{n}\,c_k^2=1\,.$ As in (\ref{tlstwo6a} -- \ref{tlstwo6b}) we may restrict this minimization problem to $\vek e^\perp$ and eliminate the unknown $c_0=-c_n\overline y_n-\sum_{k=1}^{n-1}\,c_k\overline x_k$ by orthogonalization w.r.t. $\vek e$; essentially this amounts to the same as the shift to the centroid in the primal approach in $I\!\!R^n$. So we find the restricted problem of finding vectors $\vek f_k$ ($k=1\,\cdots\, n-1$) and $\vek g$ that minimize $$ \|\,\vek y-\overline y\,\vek e -\vek g\,\|^2_2\,+ \,\sum_{k=1}^{n-1}\,\|\,\vek x_k-\overline x_k\,\vek e -\vek f_k\,\|^2_2 ~~~~\mbox{subject to}~~~~ c_1\,\vek f_1+\cdots+c_{n-1}\vek f_{n-1}+c_n\,\vek g=0\,. $$ Without imposing it, the minimizing vectors are orthogonal to $\vek e$ automatically, as in (\ref{tlstwo6b}). Defining the matrices $B$ and $E$, $$ B:=\left(\,\vek x_1 -\overline x_1\,\vek e~|~\cdots~|~ \vek x_{n-1} -\overline x_{n-1}\,\vek e~|~ \vek y -\overline y\,\vek e\,\right)~~~ {\rm and} ~~~E:=\left(\,\vek f_1~|~\cdots~|~\vek f_{n-1}~|~\vek g\,\right) $$ we can reformulate the problem as: \beq{mregres9} \mbox{minimize}~~~~~\|\,B-E\,\|^2_F~~~~~~\mbox{subject to}~~~~rank(E)=n-1\,. \end{equation} In this form it is easily solved by the SVD. If $B=\sum_{i=1}^n\,\sigma_i\,\vek u_i\,\vek v_i^T$, then $E=\sum_{i=1}^{n-1}\,\sigma_i\,\vek u_i\,\vek v_i^T$ is a minimizer of (\ref{mregres9}), which is unique, if $\sigma_{n-1}>\sigma_n\,.$ The coefficients $c_1\,,\,\cdots\,,\,c_n$ determining the hyperplane are the coordinates of the right singular vector $\vek v_n$ as before: $$ E\,\vek v_n=\vek 0\,,~~~~\Longrightarrow~~~~ \left(\matrix{c_1\cr\vdots\cr c_n}\right)=\vek v_n\,. $$ \section{General Least Squares\label{par5}} \setcounter{equation}{0} For a given matrix $A\inI\!\!R^{m\times n}$ with $m>n$ and right-hand side $\vek b\inI\!\!R^m$ we consider the problem to find the minimizer $\vek c\inI\!\!R^n$ of the functional \beq{gls1} J(\vek c)~:=~\|\,A\,\vek c\,-\,\vek b\,\|_2^2~~~~~ \mbox{with}~~~\vek c:=\left(\matrix{~c_1~\cr\vdots\cr c_n}\right)\,. \end{equation} where $$A:=\left(\matrix{~a_{1,1}~&~\cdots~&~a_{1,n}~\cr \vdots& &\vdots\cr a_{m,1}~&~\cdots~&~a_{m,n}}\right)\,\inI\!\!R^{m\times n} ~~~~~ {\rm and} ~~~~~ \vek b:=\left(\matrix{~b_1~\cr\vdots\cr b_m}\right)\,\inI\!\!R^m~~~(m\ge n)\,, $$ The difference with (\ref{mregres2}) is, that $A$ needs not contain a column consisting of all ones. The solution is obtained by a column space argument as in (\ref{mregres3}), namely that $J(\vek x)$ is minimal iff $\vek b - A\, \vek x$ is orthogonal to $\mbox{Im}(A)$ and it may be computed by normal equations, QR-factorization or SVD. \par What is interesting for the TLS generalization is the interpretation of (\ref{gls1}) in row space. We have introduced the TLS approximation in the sections \ref{par3} and \ref{par4} as the one that minimizes the sum of squares of the {\em true} distances of $m$ points to a hyperplane, whereas ordinary least squares measures the distances along the $y$-axis. We can interpret (\ref{gls1}) in this sense. The rows of the extended matrix $(A\,|\,-\vek b)$ define a cloud of $m$ points in $I\!\!R^{n+1}\,,$ \beq{gls2} ~~~~~\vek z_k :=(\,a_{k,1}\,,\,\cdots\,,\, a_{k,n}\,,\,-b_k\,)^T\inI\!\!R^{n+1}\,, ~~~\mbox{such that}~~~ \left(\,\vek z_1~|~\cdots~|~\vek z_m\,\right)= \left(\,A~|~ -\vek b\,\right)^T, \end{equation} to which we try to fit a linear function $b(x_1\cdots x_n)=c_1x_1+\cdots+c_nx_n$. In other words, we look for an $n$-dimensional {\em subspace} $\vek{\widehat c}^\perp$ in $I\!\!R^{n+1}$ (and not a hyperplane in $I\!\!R^n$ as in the regression problem), that is nearest to the datapoints (\ref{gls2}), minimizing \beq{gls3} ~~~J(\vek c)=\|\,\left(\,A~|~ -\vek b\,\right)\, \left(\matrix{~\vek c~\cr 1}\right)\,\|^2_2~=~ \sum_{k=1}^m~(\,\vek z_k^T\, \vek{\widehat c}\,)^2~~~{\rm where}~~~ \vek{\widehat c}:=\left(\matrix{\vek c\cr 1}\right)= \left(\matrix{c_1\cr \vdots \cr c_n\cr 1}\right)\inI\!\!R^{n+1}\,. \end{equation} In this sum of squares the quantity $\vek z_k^T\, \vek{\widehat c}$ measures the distance from $\vek z_k$ to $\vek{\widehat c}^\perp$ along the $n\!+\!1$-st coordinate axis. \par\noindent {\bf The Total Least Squares} approximation for the cloud of points (\ref{gls2}) minimizes the sum of squares of {\bf true} distances to the subspace $\vek{\widehat c}^\perp$. As the true distance from $\vek z_k$ to the subspace is given by $\vek z_k^T \vek c /\vek c^T\vek c$\,, see (\ref{tlstwo2}), the TLS-approximation minimizes the functional: \beq{gls4} I(\vek c):= \sum_{k=1}^m~{\left(\,\vek z_k^T\, \vek{\widehat c}\,\right)^2\over \vek{\widehat c}^T\, \vek{\widehat c}}~=~ {\|\left(\,A\,| -\vek b\,\right)\, \vek{\widehat c}\,\|^2 \over \vek{\widehat c}^T\, \vek{\widehat c}} ~~~~{\rm where}~~~~ \vek{\widehat c}:=\left(\matrix{\vek c\cr 1}\right) \end{equation} The fuctional $\vek r \mapsto \|(\,A\,| -\vek b\,)\,\vek r\|^2$ subject to $\|\vek r\|=1$ is minimal, if $\vek r$ is the right singular vector corresponding to the smallest singular value of the matrix $(\,A\,| -\vek b\,)$. Renormalizing the last component to $-1$, {\em if possible}, provides the solution to the TLS problem for the overdetermined system of equations $A\vek x=\vek b$. If the $n\!+\!1$-st component of this right singular vector is zero, no solution exists to the TLS-problem. The solution is unique if $\sigma_n>\sigma_{n+1}$. \par\noindent {\bf Interpretation of TLS in Column Space:} To each point $\vek z_k$ ($k=1\cdots m$) in the cloud (\ref{gls2}) \beq{gls5} \vek z_k=\left(\matrix{~a_{k,1}~\cr\vdots~\cr a_{k,n}\cr -b_k}~\right) ~~~\mbox{corresponds its best approximation} ~~~\vek w_k:= \left(\matrix{~f_{k,1}~\cr\vdots\cr f_{k,n}\cr -g_k}~\right)~ \in~\vek{\widehat c}^\perp\,. \end{equation} The TLS-approximation minimizes the sum of squares of the distances between the (given) points $\vek z_k$ and the points $\vek w_k$ in the subspace $\vek{\widehat c}^\perp$. We can write this sum of squares as the Frobenius norm of a matrix, if we consider the components $f_{k,j}$ as the elements of a matrix $F\inI\!\!R^{m\times n}$, and the components $g_k$ as the components of a vector $\vek g\inI\!\!R^m$. Hence, TLS minimizes \beq{gls6} \sum_{k=1}^m\,\| \vek z_k - \vek w_k\|^2=\|A-F\|^2_F+\|\vek b-\vek g\|^2= \|(A\,|-b)-(F\,| -g)\|^2_F \end{equation} Since the rows of the matrix $E:=(F\,| -g)\inI\!\!R^{m\times (n+1)}$ are orthogonal to $\vek{\widehat c}$, the rank of $E$ is $n$ at most. In other words, TLS minimizes \beq{gls7} \|\,(\,A\,|-\vek b\,) - E\,\|_F^2 ~~~~~\mbox{subject to}~~~~~ E \in I\!\!R^{m\times (n+1)}~~~ {\rm and} ~~~rank(E)\le n\,. \end{equation} We may interpret this as the quest for the solution of the solvable linear system $F\vek c=\vek g$ ``nearest'' to the (unsolvable) system $A\vek x=\vek b$, where ``solvable'' means: $\vek g \in {\rm Im}(F)\,$. \par The minimization problem (\ref{gls7}) is solved by the SVD. If $(\,A\,| -\vek b\,)=\sum_{i=1}^{n+1}\,\sigma_i\vek u_i\vek v_i^T\,,$ then $E=\sum_{i=1}^{n}\,\sigma_i\vek u_i\vek v_i^T$ and the required solution of the TLS-problem is the null-vector $\vek v_{n+1}$ of $E$, i.e.\ the right singular vector $\vek v_{n+1}$ of $(\,A\,|-\vek b\,)$ corresponding to the smallest singular value $\sigma_{n+1}$\,, provided the $n\!+\!1$-st component is non-zero. As stated at the end of section \ref{par3}, the formulation (\ref{gls7}) takes a roundabout in comparison to the equivalent formulation (\ref{gls4}) in that it asks for a minimizing system of equations, instead of the solution $\vek{\widehat c}$ itself. We conclude, that in general a best approximation of the overdetermined system $A\vek x=\vek b$ in TLS-sense may not exist, because we are not satisfied with the subspace as in a problem of regression; we want the equation for the subspace $b=c_1x_1+\cdots+c_nx_n$ to be explicit w.r.t.\ $b$. Furthermore, the solution is not necessarily unique. We shall illustrate this by two examples. \par\noindent {\bf Example 1:} Consider the cloud of 4 points in $I\!\!R^2$: $$ (1,1)\,,~~(-1,1)\,,~~(1,-1)\,,~~ {\rm and} ~~(-1,-1) $$ The LS-approximation is the horizontal line $\{(x,y)~|~ y=0\}$. The TLS-approximation makes the SVD of the matrix $B$, $$ B:=\left(\matrix{~1~&~1~\cr~1&-1~\cr-1~&~1\cr-1~&-1~}\right)= \left(\matrix{~ {1 \over 2} ~&~ {1 \over 2} ~& \sqrt{ {1 \over 2} }&0\cr ~ {1 \over 2} &- {1 \over 2} ~ & 0& \sqrt{ {1 \over 2} }\cr - {1 \over 2} ~ &~ {1 \over 2} & 0& \sqrt{ {1 \over 2} } \cr - {1 \over 2} ~&- {1 \over 2} &\sqrt{ {1 \over 2} }&0 ~ }\right) ~\left(\matrix{~2~&~0~\cr0&2\cr0&0\cr0&0}\right)~ \left(\matrix{~1~&~0~\cr0&1}\right)\,. $$ As both singular values are equal, there is no unicity; every line through the origin provides a solution, as shown in fig.\ \ref{fig4}. The sum of squares of distances from the points to a line with slope $\tan\phi$ is independent of the slope. \medskip \hrule \begin{figure}[htb] \begin{center} \begin{picture}(300,100)(-150,-55) \setlength{\unitlength}{.32mm} \put(-100,-50){\line(2,1){200}} \put(-100,0){\line(1,0){200}} \put(0,0){\circle*{3}} \put(50,50){\circle*{3}}\put(27,49){$\scriptscriptstyle (1,1)$} \put(50,-50){\circle*{3}}\put(22,-53){$\scriptscriptstyle (1,-1)$} \put(-50,50){\circle*{3}}\put(-78,49){$\scriptscriptstyle (-1,1)$} \put(-50,-50){\circle*{3}}\put(-83,-53){$\scriptscriptstyle (-1,-1)$} \put(50,50){\line(1,-2){10}} \put(50,-50){\line(-1,2){30}} \put(50,-50){\line(0,1){100}} \put(-25,-7){$\scriptstyle\varphi$} \put(53,10){$\scriptstyle\tan\,\varphi$} \put(57,40){$\scriptstyle\xi$} \put(27,-25){$\scriptstyle\eta$} \end{picture} \end{center}\vskip-10pt \caption{\label{fig4} Example 1: $\xi^2+\eta^2=(1+\tan\varphi)^2\cos^2 \varphi+(1-\tan\varphi)^2\cos^2\varphi=2$ independent on $\varphi$\,.} \end{figure} \hrule \par\noindent {\bf Example 2:} Solve the following problem in LS-sense and TLS-sense: $$ \left(\matrix{~1~&~0~\cr 0&0\cr0&0}\right)~{x \choose y}= \left(\matrix{~1~\cr1\cr 1}\right) $$ The normal equations for the LS-approximation are: $$ \left(\matrix{~1~&~0~\cr 0&0}\right)~{x \choose y}= \left(\matrix{~1~\cr 0}\right)~~~~~\Longrightarrow ~~~~x=1~~~ {\rm and} ~~~y~\mbox{undetermined}\,. $$ The SVD for TLS-problem is: $$ \def\scriptscriptstyle{1\over 2}{\scriptscriptstyle{1\over 2}} \def\scriptscriptstyle{-{1\over 2}}{\scriptscriptstyle{-{1\over 2}}} \def\scriptscriptstyle{1\over\sqrt 2}{\scriptscriptstyle{1\over\sqrt 2}} \def\scriptscriptstyle{-{1\over\sqrt 2}}{\scriptscriptstyle{-{1\over\sqrt 2}}} \def\scriptscriptstyle{\sqrt{2+\sqrt 2}}{\scriptscriptstyle{\sqrt{2+\sqrt 2}}} \def\scriptscriptstyle{\sqrt{2-\sqrt 2}}{\scriptscriptstyle{\sqrt{2-\sqrt 2}}} B~=~\left(\matrix{~1~&0&~1~\cr0&0&1\cr0&0&1}\right)~=~ \left(\matrix{ \scriptscriptstyle{1\over\sqrt 2} & \scriptscriptstyle{1\over\sqrt 2} & 0 \cr \scriptscriptstyle{1\over 2} & \scriptscriptstyle{-{1\over 2}} & \scriptscriptstyle{-{1\over\sqrt 2}} \cr \scriptscriptstyle{1\over 2} & \scriptscriptstyle{-{1\over 2}} & \scriptscriptstyle{1\over\sqrt 2} }\right) \left(\matrix{ \scriptscriptstyle{\sqrt{2+\sqrt 2}} & 0 & 0 \cr0&\scriptscriptstyle{\sqrt{2-\sqrt 2}}&0\cr0&0&0}\right) \left(\matrix{ \scriptscriptstyle{1\over 2}\scriptscriptstyle{\sqrt{2-\sqrt 2}} & 0 & \scriptscriptstyle{1\over 2}\scriptscriptstyle{\sqrt{2+\sqrt 2}} \cr \scriptscriptstyle{1\over 2}\scriptscriptstyle{\sqrt{2+\sqrt 2}} & 0 & \scriptscriptstyle{-{1\over 2}}\scriptscriptstyle{\sqrt{2-\sqrt 2}} \cr0&1&0}\right)\,. $$ The smallest singular value is 0\,. However, the $3^{\rm rd}$ component of the corresponding right singular vector $(0\,,\,1\,,\,0)^T$ is 0 as well, such that no TLS-solution exists! \section{Generalizations: (a) Multiple RHS\label{par7}} \setcounter{equation}{0} In ordinary least squares there is no difference between the treatment of one and multiple right-hand sides (RHS). In Total Least Squares the column space of the matrix is bent towards the RHS. If there are given several RHS's, we can treat each of them separately and compute the SVD of an extended matrix for each RHS. In a different approach we can try to bend the matrix to all RHS's collectively. So we consider the problem: given $A\inI\!\!R^{m\times n}$ ($m\ge n+p$) and $B\inI\!\!R^{m\times p}$ find $X\inI\!\!R^{n\times p}$ that solves the overdetermined system of equations $A\,X=B$ in TLS-sense. By analogy to (\ref{gls6}) we have to find the solution $X$ of a solvable matrix equation $F\,X=G$ (i.e. ${\rm Im}(G)\subset{\rm Im}(F)$\,) nearest to $A\,X=B$; we have to minimize \beq{gentls1} ~~~\|\,A-F\,\|_F^2~+~\|\,B-G\,\|_F^2 ~~~\mbox{subject to}~~~F \in I\!\!R^{m\times n}\,,~G \in I\!\!R^{m\times p}~ {\rm and} ~F\,X=G\,. \end{equation} Otherwise stated, find an approximation $E=(\,F~|~ G\,) \in I\!\!R^{m\times(n+p)}$ to $(\, A ~|~ B \,)$, such that \beq{gentls2} \|\,(\,A~|~ B\,)-E\,\|_F^2~~~~\mbox{is minimal subject to }~~~~rank(E)=n\,. \end{equation} The solution of (\ref{gentls2}) is constructed by making the SVD of $(\,A~|~ B\,)$: \beq{gentls3} \def\stapelup#1#2{\matrix{{\scriptscriptstyle #2}\cr #1}} \def\stapel#1#2{\matrix{#1\cr{\scriptscriptstyle #2}}} (\, A ~|~ B \,)=U\,\Sigma\,V^T= \left(\, \stapelup{U_1\vruimte{1em}{1em} }{(m\times n)} ~\vrule~\stapelup{ U_2\vruimte{1em}{1em} }{(m\times p)}\,\right)~ \left(\matrix{~\stapelup{\Sigma_{1}}{ (n\times n)}&~0~\cr \vruimte{1em}{0em}~&~\cr 0&\stapel{\Sigma_{2}}{ (p\times p)}}\right)~ \left(\matrix{~\stapelup{V_{1,1}}{(n\times n)}&~\stapelup{V_{1,2}} { (n\times p)}~\cr \vruimte{1em}{0em}~&~\cr \stapel{V_{2,1}}{ (p\times n)}&\stapel{V_{2,2}} { (p\times p)}}\right)^T\,. \end{equation} \par\noindent {\bf Theorem.} If we assume: \begin{list}{\alph{enumi}.}{\leftmargin15pt \usecounter{enumi}} \item $rank(V_{2,2})=p$\,, \item $\Sigma=diag(\sigma_1\,,\,\cdots\,,\,\sigma_n\,,\,\sigma_{n+1} \,,\,\cdots\,,\,\sigma_{n+p}\,)$ with $\sigma_j\ge\sigma_{j+1}$ and $\sigma_n\ne\sigma_{n+1}$\,, \end{list} then the TLS problem (\ref{gentls2}) has the unique solution $X=-V_{1,2}\,V_{2,2}^{-1}\,.$ \medskip\par\noindent {\bf Proof}: From (\ref{gentls3}) and the assumption $\sigma_n>\sigma_{n+1}$ it follows, that the best $rank ~n$ approximation\footnote{see \cite{gvl} theorem 2.5.2} of $(A~|~ B)$ in the Frobenius norm is given by $E$, \beq{gentls4} E:=\left(\, U_1 ~|~ U_2\,\right)~ \left(\matrix{~\Sigma_{1}~&~0~\cr 0&0}\right)~ \left(\matrix{~V_{1,1}~&~V_{1,2}~\cr V_{2,1}&V_{2,2}}\right)^T~ =U_1\,\Sigma_1\,\left(\, V_{1,1}^T ~|~ V_{1,2}^T\,\right)=(F~|~ G)\,, \end{equation} where $F:=U_1\,\Sigma_1\, V_{1,1}^T$ and $G:=U_1\,\Sigma_1\,V_{1,2}^T\,$. The orthogonality of the columns of $V$ implies $$ {V_{1,1}\choose V_{2,1}}^T\,{V_{1,2}\choose V_{2,2}}=(0) ~~~\mbox{and hence}~~~E\,\displaystyle{V_{1,2}\choose V_{2,2}}= F\,V_{1,2}+G\, V_{2,2}=(0)\,. $$ Under the assumption $rank(V_{2,2})=p$ we may conclude, that $X:=-V_{1,2}\,V_{2,2}^{-1}$ solves the approximate equation $FX=G$\,. \hfill\square{6pt}{} \par\noindent {\bf (b) Fixed columns:} In section \ref{par2} we have introduced the simple (bivariate) regression problem and we have shown that it is solved in LS-sense by the LS-solution of the overdetermined system of equations $A{a\choose b}=(\vek e ~|~\vek x){a\choose b}=\vek y$ (cf. \ref{sregres4}). However, as explained in section \ref{par5}, the TLS-solution of this overdetermined system of equations is derived from the SVD of the matrix $(\vek e~|~\vek x~|~\vek y)\inI\!\!R^{m\times 3}$. This differs from the TLS-solution of the regression problem, which is derived from the SVD of $B:=(\vek x -\overline x\vek e~|~\vek y-\overline y\vek e)\inI\!\!R^{m\times 2}$, cf. eq.~(\ref{tlstwo5a}). The reason for this difference is, that the formulation of the regression problem as an overdetermined set of equations $A{a\choose b}=\vek y$ hast lost its geometric interpretation as a line $y=a+bx$ in the $(x,y)$-plane. In the LS-solution this makes no difference since all uncertainty is put in the $\vek y$-column. However, TLS for $A{a\choose b}=\vek y$ puts uncertainty in all three columns $\vek e$, $\vek x$ and $\vek y$, although in the regression problem there is no reason to postulate uncertainty in the ``constant term''. The TLS-solution of the regression problem can be regained from $A{a\choose b}=\vek y$ if we ``freeze'' the first column of $A$ and put uncertainty in the columns $\vek x$ and $\vek y$ only as in eq.~(\ref{tlstwo6}). The solution is obtained by orthogonalization w.r.t. the frozen column $\vek e$. \par This motivates the study of TLS-problem for $A\,X=B$ with frozen columns, see \cite{ghs}, where uncertainty is postulated in a part of the columns of $A$ (LS is a special case, all columns of the matrix being frozen!). So we assume that the matrix $A$ is partitioned in a frozen part $A_1\inI\!\!R^{m\times j}$ and a part $A_2\inI\!\!R^{m\times k}$ containing some uncertainty with $j+k=n$. Given a right-hand side $B\inI\!\!R^{m\times p}$ with $m\ge j+k+p$\,, we seek matrices $X_1\inI\!\!R^{j\times p}$ and $X_2\inI\!\!R^{k\times p}$, such that \beq{gentls5} A_1\,X_1+A_2\,X_2~=~B~~~~~~\mbox{in TLS-sense w.r.t.}~~ A_2~~ {\rm and} ~~B~~~\mbox{keeping $A_1$ fixed.} \end{equation} More precise, minimize among all $C\inI\!\!R^{m\times k}$ and $D\inI\!\!R^{m\times p}$ \beq{gentls6} \|\,A_2-C\,\|_F^2~+~\|\,B-D\,\|_F^2 ~~~~~~\mbox{subject to}~~~~~A_1\,X_1+C\,X_2 =D\,. \end{equation} or otherwise said, subject to the condition $rank(A_1~|~ C~|~ D)=j+k=n$. \par\noindent Guided by the idea of (\ref{tlstwo6}), where we orthogonalized w.r.t.~the frozen column, we find the \\{\bf solution}: \beq{gentls7} \begin{array}{l} \mbox{a. Orthogonalize columns of $A_2$ and $B$ w.r.t. columns of $A_1$}\hspace{5em}\cr \mbox{b. Solve TLS-problem in the orthogonal complement ${\rm Im}(A_1)^\perp$\,.} \end{array} \end{equation} \par\noindent {\bf Proof:} If $A_1$ is of full column rank ($rank(A_1)=j$), we make the QR-factorization $$A_1=U\,{R_1\choose 0}~~~~ \mbox{ with}~~~~ U\inI\!\!R^{m\times m}\mbox{ orthogonal}~~~ {\rm and} ~~~R_1\inI\!\!R^{j\times j}\,. $$ Because the Frobenius norm is orthogonally invariant, the functional (\ref{gentls6}) is equal to \beq{gentls8a} \|\,U^T A_2-U^T C\,\|_F^2~+~\|\,U^T B-U^T D\,\|_F^2\,. \end{equation} Partitioning the matrices in parts consisting of the topmost $j$ rows and the remaining $m-j$ rows respectively, \beq{gentls8b} {A_{12}\choose A_{22}}:=U^T\,A_2,~~~ {B_1\choose B_2}:=U^T\,B ,~~~{C_1\choose C_2}:=U^T\,C, ~~~{D_1\choose D_2}:=U^T\,D \,, \end{equation} we can rewrite the functional as \beq{gentls8} \|\,A_{12}-C_1\,\|_F^2~+~\|\,B_1-D_1\,\|_F^2~+~\|\,A_{22}-C_2\,\|_F^2 ~+~\|\,B_2-D_2\,\|_F^2\,. \end{equation} It has to be minimized subject to the equations $R_1\,X_1+C_1\,X_2=D_1$ and $C_2\,X_2=D_2\,$. If $X_2$ is known, and if we choose $A_{12}=C_1$ and $B_1=D_1$, the first two terms in (\ref{gentls8}) vanish and $X_1$ can be solved from the equation $R_1\,X_1+C_1\,X_2=D_1$. Hence it suffices to minimize \beq{gentls9} \|\,A_{22}-C_2\,\|_F^2~+~\|\,B_2-D_2\,\|_F^2~~~~ \mbox{subject to}~~~~C_2\,X_2=D_2\,. \end{equation} This is solved as eq.~(\ref{gentls2}) by the SVD of $(C_2~|~ D_2)$. \par If $A$ is not of full column rank ($rank(A_1)=r<j$), we use the SVD of $A_1$: $$A_1=U\,\left(\matrix{\Sigma_1 &0\cr0&0}\right)\, {V^T_1\choose V^T_2}~~~~ \mbox{ with}~~~~ U\inI\!\!R^{m\times m},~~~ \Sigma_1\inI\!\!R^{r\times r},~~~V_1\inI\!\!R^{j\times r}, ~~~V_2\inI\!\!R^{j\times(j- r)}\,. $$ With the same partitioning as in (\ref{gentls8b}), but now with the $r$ topmost rows in the upper parts and the remaining $m-r$ rows in the lower parts, we arrive at the minimization of (\ref{gentls8}) subject to the conditions \beq{gentls10} \Sigma_1\,V_1^T\,X_1+C_1\,X_2=D_1~~~~ {\rm and} ~~~~ C_2\,X_2=D_2\,. \end{equation} Choosing $A_{12}=C_1$ and $B_1=D_1$ and solving $X_2$ from (\ref{gentls9}) we can solve $V_1^T\,X_1$ from (\ref{gentls10}). This makes the first two terms in (\ref{gentls8}) zero, such that the problem again is reduced to the form (\ref{gentls2}). As in standard LS-problems in which the matrix is not of full column rank, the part $X_1$ is not uniquely defined; we may add to it any linear combination of the columns of $V_2$\,.\hfill\square{6pt}{}

https://arxiv.org/abs/2103.00310

Alternative proof of upper bound of spanning trees in a graph

We give a proof for sharp estimate for the number of spanning trees using linear algebra and generalize this bound to multigraphs. In addition, we show that this bound is tight for complete graphs. In addition, we give estimates for number of spanning trees in specific type of graphs such as product of graphs or Cartesian product of two graphs.

\section{Introduction} Let $ G = (V,E) $ denote the undirected connected (multi)graph without loops and $ \tau(G) $ denote number of spanning trees in $ G $. In this paper we want to find upper bound of $ \tau(G) $. A common approach for counting spanning trees is \textit{Kirchhoff's Matrix Tree Theorem} and its corollary that $ \tau(G) $ can be expressed in terms of Laplacian eigenvalues. There are many various upper bounds for $\tau(G)$ in terms of number of vertices, number of edges and degrees of vertices. In \cite{narayanan} the following upper bound for number of spanning trees in a graph was proved by induction using stronger result about multigraphs: \begin{thm} Let $ G $ be a simple graph, $ d_1 \leqslant \dots \leqslant d_n $ --- degrees of its vertices. Then $$ \tau(G) \leqslant \frac{(1+d_1)\dots(1+d_n)}{n^2}. $$ \end{thm} This paper is organised as follows. We give an alternative proof of this inequality using linear algebraic techniques. After that we formulate the analogous result for multigraphs and prove its sharpness for complete multigraphs. \section{Preliminary lemmas} \begin{defn} Let $ x = (x_1 \geqslant \dots \geqslant x_n) $ and $ y = (y_1 \geqslant \dots \geqslant y_n) $ be two finite seqences of real numbers. We say that $ x $ \textbf{majorizes} $ y $ ($x \succ y$) iff for every $ k \in [1..n] $ $$ x_1 + \dots + x_k \geqslant y_1 + \dots + y_k $$ and $ x_1 + \dots + x_n = y_1 + \dots + y_n $. \end{defn} \begin{lm}[Karamata's inequality, \cite{kar}] Let $ f \colon I \to \mathbb{R} $ be a real-valued convex function, $ I $ is an interval on the real line, $ x = (x_1 \geqslant \dots \geqslant x_n) $ and $ y = (y_1 \geqslant \dots \geqslant y_n) $ are two finite sequences of numbers in I that $ x \succ y $. Then the following inequality is true: $$ f(x_1) + \dots + f(x_n) \geqslant f(y_1) + \dots + f(y_n). $$ If $ f $ is strictly convex, then the equality holds iff the sequences coincide, i.e. $ x_i = y_i $ for every $ i \in [1..n] $. \end{lm} \begin{cor} If $ x = (x_1, \dots, x_n) $, $ y = (y_1, \dots, y_n) $ are two sequences of positive numbers and $ x \succ y $, then $ x_1 \cdot \ldots \cdot x_n \leqslant y_1 \cdot \ldots \cdot y_n $. \end{cor} \begin{proof} Function $ f(x) = -\log x $ is defined on $ \mathbb{R}_+ $ and it is convex. Applying Karamata's inequality to this function and given sequences we obtain $$ -\log x_1 - \dots -\log x_n \geqslant -\log y_1 -\dots -\log y_n. $$ Reversing the sign and exponentiating with base $ e $, we obtain the desired inequality. \end{proof} \begin{lm}[Schur's inequality\cite{brou}] Let $ A $ be the symmetric real-valued $ n \times n $ matrix with diagonal elements $ d_1 \geqslant \dots \geqslant d_n $ and eigenvalues $ \lambda_1 \geqslant \dots \geqslant \lambda_n $. Then $ d_1 + \dots + d_k \leqslant \lambda_1 + \dots + \lambda_k $ for every $ k \in [1..n] $. In other words, $ \lambda \succ d $. \end{lm} \section{Basic properties about Laplacian eigenvalues} \begin{defn} Let $ G $ be a simple graph (without multiple edges and loops). \textbf{Laplacian} $ L(G) $ of this graph is the following matrix: \begin{equation*} L_{i,j} = \begin{cases} deg(v_i) &\text{if $ i = j $}\\ -1 &\text{if $ i \ne j $ and $ v_iv_j \in E(G)$}\\ 0 &\text{if $ i \ne j $ and $ v_iv_j \notin E(G)$} \end{cases} \end{equation*} \end{defn} \begin{prop} Let $ 0 = \mu_1 \leqslant \mu_2 \leqslant \ldots \leqslant \mu_n $ be the spectrum of $ L(G) $. Then for every $ k\in [1..n] $ $ \mu_k \in [0,n] $. \end{prop} \begin{prop} Let $ 0 = \mu_1 \leqslant \mu_2 \leqslant \ldots \leqslant \mu_n $ be the spectrum of $ L(G) $. Define \textbf{the complement} of G: for every $ v_i, v_j \in V(G) $: \begin{align*} v_iv_j \in E(G) \iff v_iv_j \notin E(\overline{G}); \\ v_iv_j \notin E(G) \iff v_iv_j \in E(\overline{G}). \end{align*} Then the spectrum of $ L(\overline{G}) $ is $ 0 \leqslant n-\mu_n \leqslant \dots \leqslant n-\mu_2 $. \end{prop} \begin{thm}[Kirchhoff's theorem, \cite{kir}] Number of spanning trees in graph $ G $ (denote by $ \tau(G) $) is equal to every cofactor of $ L(G) $. \end{thm} \begin{cor}[\cite{cvet}] If $ 0 = \mu_1 \leqslant \mu_2 \leqslant \dots \leqslant \mu_n $ is the spectrum of $ L(G) $, then $ \tau(G) = \frac{\mu_2\cdot\ldots\cdot\mu_n}{n} $. \end{cor} Now proceed to give the proof of main theorem. \section{Proof of the main theorem} Let us remind the main theorem. \begin{thm} Let $ G $ be a simple graph, $ v(G) = n $, $ d_1 \leqslant \dots \leqslant d_n $ be its degree sequence. Then the following equality holds: $$ \tau(G) \leqslant \frac{(1+d_1)\cdot\ldots\cdot(1+d_n)}{n^2}. $$ \end{thm} \begin{proof} From corollary of Kirchhoff's theorem we know that if $ 0 =\mu_1 \leqslant \mu_2 \leqslant \dots \leqslant \mu_n $ are eigenvalues of $ L(G) $, then $ \tau(G) = \frac{\mu_2 \cdot \ldots \cdot \mu_n}{n} $. Using this equality, we obtain $$ \frac{\mu_2 \cdot \ldots \cdot \mu_n}{n} \leqslant \frac{(1+d_1)\dots(1+d_n)}{n^2}. $$ Hence, we should prove the following inequality: $$ n\cdot\mu_n \cdot \ldots \cdot \mu_2 \leqslant (1+d_n)\cdot \ldots \cdot(1+d_1). $$ Notice that sums of multipliers in both parts are equal. Indeed, $$ n + \mu_n + \dots + \mu_2 = n + \Tr(L(G)) = n + d_1 + \dots + d_n = (d_n+1) + \dots + (d_1+1). $$ Now consider graph $ \overline{G} $ --- the complement of $ G $. Its degree sequence is $ n-1-d_n \leqslant \dots \leqslant n-1-d_1 $ and its spectrum is $ 0 \leqslant n-\mu_n \leqslant \dots \leqslant n-\mu_2 $. Applying Schur's inequality we obtain $$ (n-1-d_1) + \dots + (n-1-d_{k-1}) \leqslant (n-\mu_2) + \dots + (n-\mu_k), $$ which is equivalent to $$ \mu_2 + \dots + \mu_k \leqslant d_1 + \dots + d_{k-1} + (k-1) = (1 + d_1) + \dots + (1 + d_{k-1}). $$ So, $$ n + \mu_n + \dots + \mu_{k+1} \geqslant (1+d_n) + \dots + (1+d_k) $$ Thus, we get that sequence $ (n, \mu_n, \dots, \mu_2) $ majorizes $ (1+d_n, \dots, 1+d_1) $. Now apply the corollary of Karamata's inequality for products and get the desired inequality. \end{proof} \subsection{Generalization for multigraphs} As a corollary, we give the variant of the inequality of the main theorem in case of multigraphs. \begin{defn} Define the \textbf{complement} of multigraph $ G $ with maximal edge multiplicity $ \Delta $ ($ \Delta = \max_{i\ne j} |L_{i,j}(G)| $): for every edge $ e \in E(G) $ $$ \mu_G(e) = k \iff \mu_{\overline{G}}(e) = \Delta - k, $$ where $ \mu_G(e) $ is multiplicity of $ e $ in $ G $. \end{defn} \begin{prop} Let $ G $ be a multigraph with maximal edge multiplicity $ \Delta $, $ 0 = \mu_1 \leqslant \mu_2 \leqslant \dots \leqslant \mu_n $ --- its spectrum. Then for every $ k\in [1..n] $ $ \mu_n \in [0,n\Delta] $. \end{prop} \begin{prop} Let $ G $ be a multigraph with maximal edge multiplicity $ \Delta $, $ 0 = \mu_1 \leqslant \mu_2 \leqslant \dots \leqslant \mu_n $ --- its spectrum. Then the spectrum of $ L(\overline{G}) $ is $ 0 \leqslant n\Delta - \mu_n \leqslant \dots \leqslant n\Delta - \mu_2 $. \end{prop} \begin{thm} Let $ G $ be a multigraph, $ v(G) = n $, $ \Delta $ --- its maximal edge multiplicity and $ d_1 \leqslant \dots \leqslant d_n $ --- its degree sequence. Then the following inequality is true: $$ \tau(G) \leqslant \frac{(\Delta+d_1) \dots (\Delta+d_n)}{\Delta n^2} $$ \end{thm} \begin{proof} We will apply corollary of Karamata's inequality for sequences $ (n\Delta, \mu_n, \dots, \mu_2) $ and $ (\Delta+d_n, \dots, \Delta+d_1) $ in the similar fashion as in proof of the main theorem. Considering the multigraph $ \overline{G} $ and using proposition about its eigenvalues, we obtain the following inequality for every $ k\in[1..n]$: $$ ((n-1)\Delta-d_1) + \dots + ((n-1)\Delta-d_{k-1}) \leqslant (n\Delta - \mu_2) + \dots + (n\Delta - \mu_k), $$ which is equivalent to $$ \mu_2 + \dots + \mu_k \leqslant (d_1 + \Delta) + \dots (d_{k-1} + \Delta). $$ So, $$ n\Delta + \mu_n + \dots + \mu_{k+1} \geqslant (d_n + \Delta) + \dots + (d_k + \Delta). $$ We get that $ (n\Delta, \mu_n, \dots, \mu_2) \succ (d_n+\Delta, \dots, d_1+\Delta) $, as we needed. Application of Karamata's inequality for products completes the proof. \end{proof} Now consider in what graphs the inequality turns into equality. \begin{st} The inequality turns into equality iff $ G $ is a complete multigraph with all edge multiplicities equal to $ \Delta $. \end{st} \begin{proof} We have that $ n\Delta \cdot \mu_n \cdot \ldots \cdot \mu_2 = (d_n+\Delta)\dots(d_1+\Delta) $. It is true iff the sequences coincide, i.e. $ n\Delta = d_n + \Delta $, $ \mu_n = d_{n-1} + \Delta $, $ \dots $, $ \mu_2 = d_1 + \Delta $. From the first equality we obtain that $ d_n = (n-1)\Delta $, and since every edge of $ G $ has multiplicity no greater than $ \Delta $, we get that every edge incident to the vertex with degree $ d_n $ has multiplicity $ \Delta $. Notice that $ \overline{G} $ has at least two connectivity components, thus its second eigenvalue is zero, $ \mu_n = n\Delta $ and $ d_{n-1} = (n-1)\Delta $. So, every edge incident to the vertex with degree $ d_{n-1} $ has multiplicity $ \Delta $. Iterating this process, we obtain that all non-zero Laplacian eigenvalues of $ G $ are equal to $ n\Delta $ and every vertex has degree $ (n-1)\Delta $. Thus, every edge has multiplicity $ \Delta $. \end{proof} \newpage \section{Applications} \subsection{Product of graphs} \begin{defn} Let $ G_1, \ldots, G_n $ be $ n $ graphs. \textbf{The product of $G_i$'s}, which is denoted by $ G_1\nabla G_2 \nabla \ldots \nabla G_n $, is defined in the following way: \begin{align*} G = G_1\nabla G_2 \nabla \ldots \nabla G_n, V(G) = \bigsqcup_{i\in[1..n]}V(G_i), \\ E(G) = \bigcup_{i\in[1..n]}E(G_i) \cup \{v_iv_j\}_{1\leqslant i < j \leqslant n, v_i \in V(G_i), v_j \in V(G_j)} \end{align*} \end{defn} Suppose that $ v(G_i) = v_i $ for every $ i\in [1..n] $, $ V(G) = v $ and $ 0 = \mu^i_1 \leqslant \ldots \leqslant \mu^i_{v_i} $ is spectrum of $ L(G_i) $. Then it is known that $$ \tau(G) = v^{n-2}\prod\limits_{i\in[1..n]}\prod\limits_{j\in[2..v_i]}(v - v_i + \mu^i_j) $$ And since for every $ i\in [1..n] $ every product can be estimated by $ \prod\limits_{j\in[2..v_i]}(v - v_i + \mu^i_j) \leqslant \frac{\prod\limits_{j\in[1..v_i]}(v - v_i + d^i_j)}{(v-v_i)}$. Thus, we have \begin{thm} Let $ G_1 $, $ \ldots $, $ G_n $ be $ n $ graphs, $ v(G_i) = v_i $; $ \sum\limits_{i=1}^{n} v_i = v $, $ G = G_1\nabla\ldots\nabla G_n $. Then the following inequality holds: $$ \tau(G) \leqslant v^{n-2}\frac{\prod\limits_{u\in V(G)}d_G(u)}{\prod\limits_{i\in[1..n]}(v-v_i)} $$ \end{thm} \begin{rem} The equality holds iff $ G_i = \overline{K_{v_i}} $, and hence $ G = K_{v_1,\ldots,v_n} $. \end{rem} \subsection{Cartesian product of graphs} \begin{defn} Let $ G $ and $ H $ be two graphs. \textbf{The Cartesian product of $ G $ and $ H $}, which is denoted by $ G \times H $, is defined in the following way: if $ V(G) = \{u_1, \ldots, u_m\} $ and $ V(H) = \{v_1, \ldots, v_n\} $. Then $ V(G\otimes H) = \{w_{ij}\}_{i\in[1..m], j\in[1..n]} $, $ E(G\otimes H) = \{w_{ij}w_{ik}\}_{i\in[1..m],1\leqslant j<k \leqslant n, v_jv_k\in E(H)} \cup \{w_{ik}w_{jk}\}_{k\in[1..n], 1\leqslant i<j \leqslant m, u_iu_j \in E(G)} $. \end{defn} \begin{lm} Let $ G $ and $ H $ be two graphs, $ 0 = \lambda_1 \leqslant \ldots \leqslant \lambda_m $ is spectrum of $ L(G) $ and $ 0 = \mu_1 \leqslant \ldots \leqslant \mu_n $ is spectrum of $ L(H) $. Then spectrum of $ L(G\times H) $ is $ \{\lambda_i + \mu_j\}_{i\in[1..m], j\in[1..n]} $. \end{lm} Using previous notions, we can obtain the following formula for number of spanning trees in $ G\times H $: $$ \tau(G\times H) = \frac{\lambda_2\cdot\ldots\cdot \lambda_m}{m}\cdot\frac{\mu_2\cdot\ldots\cdot \mu_n}{n}\cdot\prod\limits_{i\in[2..m], j\in[2..n]}(\lambda_i + \mu_j) $$ Considering that first and second multipliers in the RHS are equal to $ \tau(G) $ and $ \tau(H) $, we get $$ \tau(G\times H) = \tau(G)\cdot\tau(H)\cdot\prod\limits_{i\in[2..m], j\in[2..n]}(\lambda_i + \mu_j) $$ Fix $ i\in [2..m] $ Then, if $ \lambda_i > 0 $, we can use Karamata nad Schur's inequalities and obtain $$ (\lambda_i + \mu_2)\cdot\ldots\cdot(\lambda_i + \mu_n) \leqslant \frac{\prod\limits_{j\in[1..n]}(\lambda_i + d_H(v_j))}{\lambda_i} $$ If all such $ \lambda_i $ are strictly positive, then $ G $ is connected and, multiplying these inequalities by all $ i $'s in this interval, we obtain the following: $$ \tau(G\times H) = \tau(G)\cdot\tau(H)\cdot \frac{\prod\limits_{i\in[2..m]}\prod\limits_{j\in[1..n]}(\lambda_i + d_H(v_j))}{\lambda_2\cdot\ldots\cdot\lambda_m} = \frac{\tau(H)}{m}\prod\limits_{j\in[1..n]}\prod\limits_{i\in[2..m]}(\lambda_i + d_H(v_j)) $$ Now fix $ j\in [1..n] $. Then, if $ d_H(V_j) \ne 0 $ $$ \prod\limits_{i\in[2..m]}(\lambda_i + d_H(v_j)) \leqslant \frac{\prod\limits_{k\in[1..m]}(d_G(u_k) + d_H(v_j))}{d_H(v_j)} $$ If there is no isolated vertices in $ H $, then, multiplying these inequalities by all $ j $'s in this interval, we obtain the following: $$ \prod\limits_{j\in[1..n]}\prod\limits_{i\in[2..m]}(\lambda_i + d_H(v_j)) \leqslant \frac{\prod\limits_{i\in[1..m], j\in[1..n]}(d_G(u_i) + d_H(v_j))}{\prod\limits_{j\in[1..n]}d_H(v_j)} $$ hence, $$ \tau(G\times H) \leqslant \tau(H)\cdot\frac{\prod\limits_{i\in[1..m], j\in[1..n]}(d_G(u_i) + d_H(v_j))}{m\cdot\prod\limits_{j\in[1..n]}d_H(v_j)} $$ Now formulate this result as a theorem. \begin{thm} Let $ G $ and $ H $ be two graphs, $ v(G) = m $, $ v(H) = n $, $ G $ is connected and $ H $ has no isolated vertices. Then the following inequality holds: $$ \tau(G\times H) \leqslant \tau(H)\cdot\frac{\prod\limits_{u\in V(G), v\in V(H)}(d_G(u) + d_H(v))}{m\cdot\prod\limits_{v\in V(H)}d_H(v)} $$ \end{thm} \begin{center} \begin{footnotesize} ACKNOWLEDGEMENTS \end{footnotesize} \end{center} The author is grateful to Fedor Petrov and Alexander Polyanski for suggestions and discussions concerned with this paper and their support. The main results are formulated in author's bachelor thesis. \newpage

https://arxiv.org/abs/0809.4621

Explicit constructions of infinite families of MSTD sets

We explicitly construct infinite families of MSTD (more sums than differences) sets. There are enough of these sets to prove that there exists a constant C such that at least C / r^4 of the 2^r subsets of {1,...,r} are MSTD sets; thus our family is significantly denser than previous constructions (whose densities are at most f(r)/2^{r/2} for some polynomial f(r)). We conclude by generalizing our method to compare linear forms epsilon_1 A + ... + epsilon_n A with epsilon_i in {-1,1}.

\section{Introduction} Given a finite set of integers $A$, we define its sumset $A+A$ and difference set $A-A$ by \bea A + A & \ = \ & \{a_i + a_j: a_i, a_j \in A\} \nonumber\\ A - A & = & \{a_i - a_j: a_i, a_j \in A\}, \eea and let $|X|$ denote the cardinality of $X$. If $|A+A| > |A-A|$, then, following Nathanson, we call $A$ an MSTD (more sums than differences) set. As addition is commutative while subtraction is not, we expect that for a `generic' set $A$ we have $|A-A| > |A+A|$, as a typical pair $(x,y)$ contributes one sum and two differences; thus we expect MSTD sets to be rare. Martin and O'Bryant \cite{MO} proved that, in some sense, this intuition is wrong. They considered the uniform model\footnote{This means each of the $2^n$ subsets of $\{1,\dots,n\}$ are equally likely to be chosen, or, equivalently, that the probability any $k \in \{1,\dots,n\}$ is in $A$ is just $1/2$.\label{footnote:unifmodel}} for choosing a subset $A$ of $\{1,\dots,n\}$, and showed that there is a positive probability that a random subset $A$ is an MSTD set (though, not surprisingly, the probability is quite small). However, the answer is very different for other ways of choosing subsets randomly, and if we decrease slightly the probability an element is chosen then our intuition is correct. Specifically, consider the binomial model with parameter $p(n)$, with $\lim_{n\to\infty} p(n) = 0$ and $n^{-1} = o(p(n))$ (so $p(n)$ doesn't tend to zero so rapidly that the sets are too sparse).\footnote{This model means that the probability $k \in \{1,\dots,n\}$ is in $A$ is $p(n)$.} Hegarty and Miller \cite{HM} recently proved that, in the limit as $n \to 0$, the percentage of subsets of $\{1,\dots,n\}$ that are MSTD sets tends to zero in this model. Though MSTD sets are rare, they do exist (and, in the uniform model, are somewhat abundant by the work of Martin and O'Bryant). Examples go back to the 1960s. Conway is said to have discovered $\{0, 2, 3, 4, 7, 11, 12, 14\}$, while Marica \cite{Ma} gave $\{0, 1, 2, 4, 7, 8, 12, 14, 15\}$ in 1969 and Freiman and Pigarev \cite{FP} found $\{0, 1, 2, 4, 5$, $9, 12, 13$, $14, 16, 17$, $21, 24, 25, 26, 28, 29\}$ in 1973. Recent work includes infinite families constructed by Hegarty \cite{He} and Nathanson \cite{Na2}, as well as existence proofs by Ruzsa \cite{Ru1, Ru2, Ru3}. Most of the previous constructions\footnote{An alternate method constructs an infinite family from a given MSTD set $A$ by considering $A_t = \{\sum_{i=1}^t a_i m^{i-1}: a_i \in A\}$. For $m$ sufficiently large, these will be MSTD sets; this is called the base expansion method. Note, however, that these will be very sparse. See \cite{He} for more details.} of infinite families of MSTD sets start with a symmetric set which is then `perturbed' slightly through the careful addition of a few elements that increase the number of sums more than the number of differences; see \cite{He,Na2} for a description of some previous constructions and methods. In many cases, these symmetric sets are arithmetic progressions; such sets are natural starting points because if $A$ is an arithmetic progression, then $|A+A| = |A-A|$.\footnote{As $|A+A|$ and $|A-A|$ are not changed by mapping each $x \in A$ to $\alpha x + \beta$ for any fixed $\alpha$ and $\beta$, we may assume our arithmetic progression is just $\{0,\dots,n\}$, and thus the cardinality of each set is $2n+1$.} In this work we present a new method which takes an MSTD set satisfying certain conditions and constructs an infinite family of MSTD sets. While these families are not dense enough to prove a positive percentage of subsets of $\{1, \dots, r\}$ are MSTD sets, we are able to elementarily show that the percentage is at least $C / r^4$ for some constant $C$. Thus our families are far denser than those in \cite{He,Na2}; trivial counting\footnote{For example, consider the following construction of MSTD sets from \cite{Na2}: let $m, d, k \in \mathbb{N}$ with $m \ge 4$, $1 \le d \le m-1$, $d \neq m/2$, $k \ge 3$ if $d < m/2$ else $k \ge 4$. Set $B = [0,m-1] \backslash \{d\}$, $L = \{m-d, 2m-d, \dots , km-d\}$, $a^\ast = (k + 1)m-2d$ and $A = B \cup L \cup (a^\ast - B) \cup \{m\}$. Then $A$ is an MSTD set. The width of such a set is of the order $km$. Thus, if we look at all triples $(m,d,k)$ with $km \le r$ satisfying the above conditions, these generate on the order of at most $\sum_{k \le r} \sum_{m \le r/k} \sum_{d \le m} 1 \ll r^2$, and there are of the order $2^r$ possible subsets of $\{0,\dots,r\}$; thus this construction generates a negligible number of MSTD sets. Though we write $f(r)/2^{r/2}$ to bound the percentage from other methods, a more careful analysis shows it is significantly less; we prefer this easier bound as it is already significantly less than our method. See for example Theorem 2 of \cite{He} for a denser example.} shows all of their infinite families give at most $f(r)2^{r/2}$ of the subsets of $\{1,\dots, r\}$ (for some polynomial $f(r)$) are MSTD sets, implying a percentage of at most $f(r) / 2^{r/2}$.\\ We first introduce some notation. The first is a common convention, while the second codifies a property which we've found facilitates the construction of MSTD sets. \\ \bi \item We let $[a,b]$ denote all integers from $a$ to $b$; thus $[a,b] = \{n \in \mathbb{Z}: a \le n \le b\}$.\\ \item We say a set of integers $A$ has the property $P_n$ (or is a $P_n$-set) if both its sumset and its difference set contain all but the first and last $n$ possible elements (and of course it may or may not contain some of these fringe elements).\footnote{It is not hard to show that for fixed $0<\alpha\le1$ a random set drawn from $[1,n]$ in the uniform model is a $P_{\lfloor \alpha n\rfloor}$-set with probability approaching $1$ as $n\to\infty$.\label{footnote:beingpn}} Explicitly, let $a=\min{A}$ and $b=\max{A}$. Then $A$ is a $P_n$-set if \bea\label{eq:beingPnsetsum} [2a+n,\ 2b-n] \ \subset\ A+A \eea and \bea\label{eq:beingPnsetdiff} [-(b-a)+n,\ (b-a)-n]\ \subset\ A-A.\eea \ \\ \ \ei We can now state our construction and main result. \begin{thm}\label{thm:mainconstruction} Let $A=L\cup R$ be a $P_n$, MSTD set where $L\subset[1,n]$, $R\subset[n+1,2n]$, and $1,2n\in A$;\footnote{Requiring $1, 2n \in A$ is quite mild; we do this so that we know the first and last elements of $A$.} see Remark \ref{rek:thmisnontrivial} for an example of such an $A$. Fix a $k \ge n$ and let $m$ be arbitrary. Let $M$ be any subset of $[n+k+1, n+k+m]$ with the property that it does not have a run of more than $k$ missing elements (i.e., for all $\ell \in [n+k+1,n+m+1]$ there is a $j \in [\ell, \ell+k-1]$ such that $j\in M$). Assume further that $n+k+1 \not\in M$ and set $A(M;k)=L \cup O_1 \cup M \cup O_2 \cup R'$, where $O_1=[n+1,n+k]$, $O_2=[n+k+m+1,n+2k+m]$ (thus the $O_i$'s are just sets of $k$ consecutive integers), and $R'=R+2k+m$. Then \ben \item $A(M;k)$ is an MSTD set, and thus we obtain an infinite family of distinct MSTD sets as $M$ varies; \item there is a constant $C > 0$ such that as $r\to\infty$ the percentage of subsets of $\{1,\dots,r\}$ that are in this family (and thus are MSTD sets) is at least $C / r^4$. \een \end{thm} \begin{rek}\label{rek:thmisnontrivial} In order to show that our theorem is not trivial, we must of course exhibit at least one $P_n$, MSTD set $A$ satisfying all our requirements (else our family is empty!). We may take the set\footnote{This $A$ is trivially modified from \cite{Ma} by adding 1 to each element, as we start our sets with 1 while other authors start with 0. We chose this set as our example as it has several additional nice properties that were needed in earlier versions of our construction which required us to assume slightly more about $A$.} $A = \{1, 2, 3, 5, 8, 9, 13, 15, 16\}$; it is an MSTD set as \bea A+A & \ = \ & \{2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,\nonumber\\ & & \ \ \ \ \ 22,23,24,25,26,28,29,30,31,32\} \nonumber\\ A-A & \ = \ & \{-15,-14,-13,-12,-11,-10,-8,-7,-6,-5,-4,-3,-2,-1,\nonumber\\ & & \ \ \ \ \ 0,1,2,3,4,5,6,7,8,10,11,12,13,14,15\} \eea (so $|A+A| = 30 >29 = |A-A|$). $A$ is also a $P_n$-set, as \eqref{eq:beingPnsetsum} is satisfied since $[10,24] \subset A+A$ and \eqref{eq:beingPnsetdiff} is satisfied since $[-7,7] \subset A-A$. For the uniform model, a subset of $[1,2n]$ is a $P_n$-set with high probability as $n\to\infty$, and thus examples of this nature are plentiful. For example, of the $1748$ MSTD sets with minimum $1$ and maximum $24$, $1008$ are $P_n$-sets. \end{rek} Unlike other estimates on the percentage of MSTD sets, our arguments are not probabilistic, and rely on explicitly constructing large families of MSTD sets. Our arguments share some similarities with the methods in \cite{He} (see for example Case I of Theorem 8) and \cite{MO}. There the fringe elements of the set were also chosen first. A random set was then added in the middle, and the authors argued that with high probability the resulting set is an MSTD set. We can almost add a random set in the middle; the reason we do not obtain a positive percentage is that we have the restriction that there can be no consecutive block of size $k$ of numbers in the middle that are not chosen to be in $A(M;k)$. This is easily satisfied by requiring us to choose at least one number in consecutive blocks of size $k/2$, and this is what leads to the loss of a positive percentage\footnote{Without this requirement, we could take any $M$ and thus would have a positive percentage work, specifically at least $2^{-(2k+2n)}$.} (though we do obtain sets that are known to be MSTD sets, and not just highly likely to be MSTD sets). The paper is organized as follows. We describe our construction in \S\ref{sec:infinitefamilies}, and prove our claimed lower bounds for the percentage of sets that are MSTD sets in \S\ref{sec:lowerboundspercentage}. We then generalize our construction in \S\ref{sec:generalizingconstr} and explore when there are infinite families of sets satisfying \be \left|\gep_1 A + \cdots + \gep_n A\right| \ > \ \left|\widetilde{\gep}_1 A + \cdots + \widetilde{\gep}_n A\right|, \ \ \ \gep_i, \widetilde{\gep}_i \in \{-1,1\}. \ee We end with some concluding remarks and suggestions for future research in \S\ref{sec:concremfutureresearch}. \section{Construction of infinite families of MSTD sets}\label{sec:infinitefamilies} Let $A\subset [1,2n]$. We can write this set as $A=L\cup R$ where $L\subset[1,n]$ and $R\subset[n+1,2n]$. We have \begin{equation} A+A \ = \ [L+L] \cup [L+R] \cup [R+R] \end{equation} where $L+L\subset[2,2n]$, $L+R \subset[n+2,3n]$ and $R+R\subset [2n+2,4n]$, and \begin{equation} A-A \ = \ [L-R]\cup [L-L] \cup [R-R] \cup [R-L] \end{equation} where $L-R\subset[-1,-2n+1]$, $L-L \subset[-(n-1),n-1]$, $R-R\subset [-(n-1),n-1]$ and $R-L\subset [1,2n-1]$. A typical subset $A$ of $\{1,\dots,2n\}$ (chosen from the uniform model, see Footnote \ref{footnote:unifmodel}) will be a $P_n$-set (see Footnote \ref{footnote:beingpn}). It is thus the interaction of the ``fringe'' elements that largely determines whether a given set is an MSTD set. Our construction begins with a set $A$ that is both an MSTD set and a $P_n$-set. We construct a family of $P_n$, MSTD sets by inserting elements into the middle in such a way that the new set is a $P_n$-set, and the number of added sums is equal to the number of added differences. Thus the new set is also an MSTD set. In creating MSTD sets, it is very useful to know that we have a $P_n$-set. The reason is that we have all but the ``fringe'' possible sums and differences, and are thus reduced to studying the extreme sums and differences. The following lemma shows that if $A$ is a $P_n$, MSTD set and a certain extension of $A$ is a $P_n$-set, then this extension is also an MSTD set. The difficult step in our construction is determining a large class of extensions which lead to $P_n$-sets; we will do this in Lemma \ref{lem:mainconstr}. \begin{lem}\label{lem:stability} Let $A=L\cup R$ be a $P_n$-set where $L\subset[1,n]$ and $R\subset[n+1,2n]$. Form $A'=L \cup M \cup R'$ where $M\subset [n+1,n+m]$ and $R'=R+m$. If $A'$ is a $P_n$-set then $|A'+A'|-|A+A|=|A'-A'|-|A-A|=2m$ (i.e., the number of added sums is equal to the number of added differences). In particular, if $A$ is an MSTD set then so is $A'$. \end{lem} \begin{proof} We first count the number of added sums. In the interval $[2,n+1]$ both $A+A$ and $A'+A'$ are identical, as any sum can come only from terms in $L+L$. Similarly, we can pair the sums of $A+A$ in the region $[3n+1,4n]$ with the sums of $A'+A'$ in the region $[3n+2m+1,4n+2m]$, as these can come only from $R+R$ and $(R+m)+(R+m)$ respectively. Since we have accounted for the $n$ smallest and largest terms in both $A+A$ and $A'+A'$, and as both are $P_n$-sets, the number of added sums is just $(3n+2m+1)-(3n+1)=2m$. Similarly, differences in the interval $[1-2n, -n]$ that come from $L-R$ can be paired with the corresponding terms from $L-(R+m)$, and differences in the interval $[n, 2n-1]$ from $R-L$ can be paired with differences coming from $(R+m)-L$. Thus the size of the middle grows from the interval $[-n+1,n-1]$ to the interval $[-n-m+1,n+m-1]$. Thus we have added $(2n+2m+3)-(2n+3)=2m$ differences. Thus $|A'+A'|-|A+A|=|A'-A'|-|A-A|=2m$ as desired. \end{proof} The above lemma is not surprising, as in it we assume $A'$ is a $P_n$-set; the difficulty in our construction is showing that our new set $A(M;k)$ is also a $P_n$-set for suitably chosen $M$. This requirement forces us to introduce the sets $O_i$ (which are blocks of $k$ consecutive integers), as well as requiring $M$ to have at least one of every $k$ consecutive integers. We are now ready to prove the first part of Theorem \ref{thm:mainconstruction} by constructing an infinite family of distinct $P_n$, MSTD sets. We take a $P_n$, MSTD set and insert a set in such a way that it remains a $P_n$-set; thus by Lemma \ref{lem:stability} we see that this new set is an MSTD set. \begin{lem}\label{lem:mainconstr} Let $A=L\cup R$ be a $P_n$-set where $L\subset[1,n]$, $R\subset[n+1,2n]$, and $1,2n\in A$. Fix a $k \ge n$ and let $m$ be arbitrary. Choose any $M\subset[n+k+1,n+k+m]$ with the property that $M$ does not have a run of more than $k$ missing elements, and form $A(M;k)=L \cup O_1 \cup M \cup O_2 \cup R'$ where $O_1=[n+1,n+k]$, $O_2=[n+k+m+1,n+2k+m]$, and $R'=R+2k+m$. Then $A(M;k)$ is a $P_n$-set. \end{lem} \begin{proof} For notational convenience, denote $A(M;k)$ by $A'$. Note $A'+A' \subset [2, 4n+4k+2m]$. We begin by showing that there are no missing sums from $n+2$ to $3n+4k+2m$; proving an analogous statement for $A'-A'$ shows $A'$ is a $P_n$-set. By symmetry\footnote{Apply the arguments below to the set $2n+2k+m-A'$, noting that $1, 2n+2k+m \in A'$.} we only have to show that there are no missing sums in $[n+2, 2n+2k+m]$. We consider various ranges in turn. We observe that $[n+2, n+k+1]\subset A'+A'$ because we have $1\in L$ and these sums result from $1+O_1$. Additionally, $O_1 + O_1=[2n+2, 2n+2k]\subset A'+A'$. Since $n\le k$ we have $n+k+1\ge 2n+1$, these two regions are contiguous and thus $[n+2, 2n+2k]\subset A'+A'$. Now consider $O_1 + M$. Since $M$ does not have a run of more than $k$ missing elements, the worst case scenario (in terms of getting the required sums) is that the smallest element of $M$ is $n+2k$ and that the largest element is $n+m+1$ (and, of course, we still have at least one out of every $k$ consecutive integers is in $M$). If this is the case then we still have $O_1+M \supset [(n+1)+(n+2k), (n+k) + (n+m+1)]=[2n+2k+1, 2n+k+m+1]$. We had already shown that $A'+A'$ has all sums up to $2n+2k$; this extends the sumset to all sums up to $2n+k+m+1$. All that remains is to show we have all sums in $[2n+k+m+2, 2n+2k+m]$. This follows immediately from $O_1+O_2=[2n+k+m+2,2n+3k+m]\subset A'+A'$. This extends our sumset to include all sums up to $2n+3k+m$, which is well past our halfway mark of $2n+2k+m$. Thus we have shown that $A'+A' \supset [n+2, 3n+4k+2m+1]$.\\ We now do a similar calculation for the difference set, which is contained in $[-(2n+2k+m)+1, (2n+2k+m)-1]$. As we have already analyzed the sumset, all that remains to prove $A$ is a $P_n$-set is to show that $A'-A' \supset [-n-2k-m+1,n+2k+m-1]$. As all difference sets\footnote{Unless, of course, $A$ is the empty set!} are symmetric about and contain $0$, it suffices to show the positive elements are present, i.e., that $A'-A' \supset [1,n+2k+m-1]$. We easily see $[1,k-1] \subset A'-A'$ as $[0,k-1] \subset O_1 - O_1$. Now consider $M - O_1$. Again the worst case scenario (for getting the required differences) is that the least element of $M$ is $n+2k$ and the greatest is $n+m+1$. With this in mind we see that $M - O_1 \supset [(n+2k)-(n+k) , (n+m+1)-(n+1)]=[k,m]$. Now $O_2 - O_1 \supset [(n+k+m+1)-(n+k), (n+2k+m)-(n+1)]=[m+1,2k+m-1]$, and we therefore have all differences up to $2k+m-1$. Since $2n \in A$ we have $2n+2k+m \in A'$. Consider $(2n+2k+m) - O_1 = [n+k+m,n+2k+m-1]$. Since $k\ge n$ we see that $n+k+m \le 2k+m$; this implies that we have all differences up to $n+2k+m-1$ (this is because we already have all differences up to $2k+m-1$, and $n+k+m$ is either less than $2k+m-1$, or at most one larger). \end{proof} \begin{proof}{Proof of Theorem \ref{thm:mainconstruction}(1).} The proof of the first part of Theorem \ref{thm:mainconstruction} follows immediately. By Lemma \ref{lem:mainconstr} our new sets $A(M;k)$ are $P_n$-sets, and by Lemma \ref{lem:stability} they are also MSTD. All that remains is to show that the sets are distinct; this is done by requiring $n+k+1$ is not in our set (for a fixed $k$, these sets have elements $n+1, \dots, n+k$ but not $n+k+1$; thus different $k$ yield distinct sets). \end{proof} \section{Lower bounds for the percentage of MSTDs}\label{sec:lowerboundspercentage} To finish the proof of Theorem \ref{thm:mainconstruction}, for a fixed $n$ we need to count how many sets $\widetilde{M}$ of the form $O_1 \cup M \cup O_2$ (see Theorem \ref{thm:mainconstruction} for a description of these sets) of width $r = 2k+m$ can be inserted into a $P_n$, MSTD set $A$ of width $2n$. As $O_1$ and $O_2$ are just intervals of $k$ consecutive ones, the flexibility in choosing them comes solely from the freedom to choose their length $k$ (so long as $k \ge n$). There is far more freedom to choose $M$. There are two issues we must address. First, we must determine how many ways there are there to fill the elements of $M$ such that there are no runs of $k$ missing elements. Second, we must show that the sets generated by this method are distinct. We saw in the proof of Theorem \ref{thm:mainconstruction}(1) that the latter is easily handled by giving $A(M;k)$ (through our choice of $M$) slightly more structure. Assume that the element $n+k+1$ is \emph{not} in $M$ (and thus not in $A$). Then for a fixed width $r=2k+m$ each value of $k$ gives rise to necessarily distinct sets, since the set contains $[n+1, n+k]$ but not $n+k+1$. In our arguments below, we assume our initial $P_n$, MSTD set $A$ is fixed; we could easily increase the number of generated MSTD sets by varying $A$ over certain MSTD sets of size $2n$. We choose not to do this as $n$ is fixed, and thus varying over such $A$ will only change the percentages by a constant independent of $k$ and $m$. Fix $n$ and let $r$ tend to infinity. We count how many $\widetilde{M}$'s there are of width $r$ such that in $M$ there is at least one element chosen in any consecutive block of $k$ integers. One way to ensure this is to divide $M$ into consecutive, non-overlapping blocks of size $k/2$, and choose at least one element in each block. There are $2^{k/2}$ subsets of a block of size $k/2$, and all but one have at least one element. Thus there are $2^{k/2} - 1 = 2^{k/2} (1 - 2^{-k/2})$ valid choices for each block of size $k/2$. As the width of $M$ is $r-2k$, there are $\lceil \frac{r-2k}{k/2}\rceil \le \frac{r}{k/2}-3$ blocks (the last block may have length less than $k/2$, in which case any configuration will suffice to ensure there is not a consecutive string of $k$ omitted elements in $M$ because there will be at least one element chosen in the previous block). We see that the number of valid $M$'s of width $r-2k$ is at least $2^{r-2k} \left(1 - 2^{-k/2}\right)^{\frac{r}{k/2}-3}$. As $O_1$ and $O_2$ are two sets of $k$ consecutive $1$'s, there is only one way to choose either. We therefore see that, for a fixed $k$, of the $2^r = 2^{m+2k}$ possible subsets of $r$ consecutive integers, we have at least $2^{r-2k} \left(1 - 2^{-k/2}\right)^{\frac{r}{k/2}-3}$ are permissible to insert into $A$. To ensure that all of the sets are distinct, we require $n+k+1 \not\in M$; the effect of this is to eliminate one degree of freedom in choosing an element in the first block of $M$, and this will only change the proportionality constants in the percentage calculation (and \emph{not} the $r$ or $k$ dependencies). Thus if we vary $k$ from $n$ to $r/4$ (we could go a little higher, but once $k$ is as large as a constant times $r$ the number of generated sets of width $r$ is negligible) we have at least some fixed constant times $2^r \sum_{k=n}^{r/4} \frac{1}{2^{2k}} \left(1 - 2^{-k/2}\right)^{\frac{r}{k/2}-3}$ MSTD sets; equivalently, the percentage of sets $O_1 \cup M \cup O_2$ with $O_i$ of width $k \in \{n,\dots, r/4\}$ and $M$ of width $r-2k$ that we may add is at least this divided by $2^r$, or some universal constant times \be\label{eq:keycardsum} \sum_{k=n}^{r/4} \frac1{2^{2k}} \left(1 - \frac1{2^{k/2}}\right)^{\frac{r}{k/2}} \ee (as $k \ge n$ and $n$ is fixed, we may remove the $-3$ in the exponent by changing the universal constant). We now determine the asymptotic behavior of this sum. More generally, we can consider sums of the form \be S(a,b,c;r) \ = \ \sum_{k=n}^{r/4} \frac1{2^{ak}} \left(1 - \frac1{2^{bk}}\right)^{r/ck}. \ee For our purposes we take $a=2$ and $b=c=1/2$; we consider this more general sum so that any improvements in our method can readily be translated into improvements in counting MSTD sets. While we know (from the work of Martin and O'Bryant \cite{MO}) that a positive percentage of such subsets are MSTD sets, our analysis of this sum yields slightly weaker results. The approach in \cite{MO} is probabilistic, obtained by fixing the fringes of our subsets to ensure certain sums and differences are in (or not in) the sum- and difference sets. While our approach also fixes the fringes, we have far more possible fringe choices than in \cite{MO} (though we do not exploit this). While we cannot prove a positive percentage of subsets are MSTD sets, our arguments are far more elementary. The proof of Theorem \ref{thm:mainconstruction}(2) is clearly reduced to proving the following lemma, and then setting $a = 2$ and $b=c=1/2$. \begin{lem}\label{lem:lowerupperbounds} Let \be S(a,b,c;r) \ = \ \sum_{k=n}^{r/4} \frac1{2^{ak}} \left(1 - \frac1{2^{bk}}\right)^{r/ck}. \ee Then for any $\epsilon > 0$ we have \be \frac{1}{r^{a/b}} \ \ll \ S(a,b,c;r) \ \ll \ \frac{(\log r)^{2a+\epsilon}}{r^{a/b}}. \ee \end{lem} \begin{proof} We constantly use $(1 - 1/x)^x$ is an increasing function in $x$. We first prove the lower bound. For $k \ge (\log_2 r)/b$ and $r$ large, we have \be \left(1 - \frac1{2^{bk}}\right)^{r/ck} \ = \ \left(1 - \frac1{2^{bk}}\right)^{2^{bk} \frac{r}{ck 2^{bk}}} \ \ge \ \left(1 - \frac{1}{r}\right)^{r \cdot \frac{b}{c \log_2 r}} \ \ge \ \foh \ee (in fact, for $r$ large the last bound is almost exactly 1). Thus we trivially have \bea S(a,b,c;r) & \ \ge \ & \sum_{k = (\log_2 r)/b}^{r/4} \frac1{2^{ak}} \cdot \foh \ \gg \ \frac1{r^{a/b}}. \eea For the upper bound, we divide the $k$-sum into two ranges: (1) $bn \le bk \le \log_2 r - \log_2 (\log r)^\delta$; (2) $\log_2 r - \log_2(\log r)^\delta \le bk \le br/4$. In the first range, we have \begin{eqnarray} \left(1 - \frac1{2^{bk}}\right)^{r/ck} & \ \le \ & \left(1 - \frac{(\log r)^\delta}{r}\right)^{r/ck}\nonumber\\ & \ \ll \ & \exp\left(-\frac{b(\log r)^\delta}{c \log_2 r}\right) \nonumber\\ & \ \le \ & \exp\left(-\frac{b\log 2}{c} \cdot (\log r)^{\delta -1}\right). \end{eqnarray} If $\delta > 2$ then this factor is dominated by $r^{-\frac{b\log 2}{c} \cdot (\log r)^{\delta - 2}} \ll r^{-A}$ for any $A$ for $r$ sufficiently large. Thus there is negligible contribution from $k$ in range (1) if we take $\delta = 2 + \epsilon/a$ for any $\epsilon > 0$. For $k$ in the second range, we trivially bound the factors $\left(1 - 1/2^{bk}\right)^{r/ck}$ by 1. We are left with \bea \sum_{k \ge \frac{\log_2 r}b - \frac{\log_2(\log r)^\delta}{b}} \ \frac1{2^{ak}} \cdot 1 \ \le \ \frac{(\log r)^{a \delta}}{r^{a/b}} \sum_{\ell = 0}^\infty \frac1{2^{a\ell}} \ \ll \ \frac{(\log r)^{a \delta}}{r^{a/b}}. \eea Combining the bounds for the two ranges with $\delta = 2 + \epsilon/a$ completes the proof. \end{proof} \begin{rek} The upper and lower bounds in Lemma \ref{lem:lowerupperbounds} are quite close, differing by a few powers of $\log r$. The true value will be at least $\left(\frac{\log r}{r}\right)^{a/b}$; we sketch the proof in Appendix \ref{sec:sizeSabcm}. \end{rek} \begin{rek} We could attempt to increase our lower bound for the percentage of subsets that are MSTD sets by summing $r$ from $R_0$ to $R$ (as we have fixed $r$ above, we are only counting MSTD sets of width $2n+r$ where $1$ and $2n+r$ are in the set. Unfortunately, at best we can change the universal constant; our bound will still be of the order $1/R^4$. To see this, note the number of such MSTD sets is at least a constant times $\sum_{r=R_0}^R 2^r / r^4$ (to get the percentage, we divide this by $2^R$). If $r \le R/2$ then there are exponentially few sets. If $r \ge R/2$ then $r^{-4} \in [1/R^4, 16/R^4]$. Thus the percentage of such subsets is still only at least of order $1/R^4$. \end{rek} \section{Generalizing our construction}\label{sec:generalizingconstr} Instead of searching for $A$ such that $|A+A| > |A-A|$, we now consider the more general problem\footnote{We do not consider the most general problem of comparing arbitrary combinations of $A$, contenting ourselves to this special case; see \cite{HM} for some thoughts about such generalizations.} of when \be \left|\gep_1 A + \cdots + \gep_n A\right| \ > \ \left|\widetilde{\gep}_1 A + \cdots + \widetilde{\gep}_n A\right|, \ \ \ \gep_i, \widetilde{\gep}_i \in \{-1,1\}. \ee Consider the generalized sumset \be f_{j_1,\ j_2}(A)\ =\ A+A+\cdots+A-A-A-\cdots-A,\ee where there are $j_1$ pluses\footnote{By a slight abuse of notation, we say there are two sums in $A+A-A$, as is clear when we write it as $\epsilon_1 A + \epsilon_2 A + \epsilon_3 A$.} and $j_2$ minuses, and set $j=j_1 + j_2$. Our notion of a $P_n$-set generalizes, and we find that if there exists one set $A$ with $|f_{j_1,\ j_2}(A)| > |f_{j_1',\ j_2'}(A)|$, then we can construct infinitely many such $A$. Note without loss of generality that we may assume $j_1 \ge j_2$.\footnote{This follows as we are only interested in $|f_{j_1,\ j_2}(A)|$, which equals $|f_{j_2,\ j_1}(A)|$. This is because $B$ and $-B$ have the same cardinality, and thus (for example) we see $A+A-A$ and $-(A-A-A)$ have the same cardinality.} \begin{defi}[$P_n^j$-set.] Let $A \subset [1, k]$ with $1, k, \in A$. We say $A$ is a $P^j_n$-set if any $f_{j_1,\ j_2}(A)$ contains all but the first $n$ and last $n$ possible elements. \end{defi} \begin{rek} Note that a $P_n^2$-set is the same as what we called a $P_n$-set earlier. \end{rek} We expect the following generalization of Theorem \ref{thm:mainconstruction} to hold. \begin{conj}\label{conj:genconstr} For any $f_{j_1,\ j_2}$ and $f_{j_1',\ j_2'}$, if there exists a finite set of integers $A$ which is (1) a $P^j_n$-set; (2) $A \subset [1, 2n]$ and $1, 2n \in A$; and (3) $|f_{j_1,\ j_2}(A)|>|f_{j_1',\ j_2'}(A)|$, then there exists an infinite family of such sets. \end{conj} The difficulty in proving the above conjecture is that we need to find a set $A$ satisfying $|f_{j_1,\ j_2}(A)|>|f_{j_1',\ j_2'}(A)|$; once we find such a set, we can mirror the construction from Theorem \ref{thm:mainconstruction}. Currently we can only find such $A$ for $j \in \{2,3\}$: \begin{thm}\label{thm:genconstr} Conjecture \ref{conj:genconstr} is true for $j \in \{2,3\}$. \end{thm} As the proof is similar to that of Theorem \ref{thm:mainconstruction}, we just highlight the changes. We prove the lemmas below in greater generality than we need for our theorem as this generality is needed to attack Conjecture \ref{conj:genconstr}. The first step is an analogue of Lemma \ref{lem:stability}, the second is proving that a $P_n^2$-set is also a $P_n^j$-set, and the third is constructing sets $A$ (when $j = 3$) to start the construction. \begin{lem}\label{lem:gen1} Let $A=L \cup R$ be a $P_n^j$-set, where $L \subset [1, n], R \subset [n+1, 2n]$. Form $A' =L\cup M\cup R',$ where $M \subset [n+1, n+m]$ and $R' = R+m$. If $A'$ is a $P_n^j$-set, then $|f_{j_1,\ j_2}(A')| -|f_{j_1,\ j_2}(A)| = |f_{j_1',\ j_2'}(A')| -|f_{j_1',\ j_2'}(A)|.$ Thus if $|f_{j_1,\ j_2}(A)|>|f_{j_1',\ j_2'}(A)|$, the same is true for $A'$. \end{lem} \begin{proof} Since $A\subset [1, 2n]$ and is a $P^j_n$-set, we know $f(A) \subset [j_1-2nj_2, 2nj_1-j_2]$ and $[j_1 -2nj_2+n, 2nj_1-j_2 -n] \subset f(A)$. Note any elements in $f(A) \cap [j_1 -2nj_2, j_1-2nj_2+n-1]$ can only come from $L+L+L+ \cdots+L-R-R-R-\cdots-R$. As $A' \subset [1, 2n+m]$, $f(A') \subset [j_1-(2n+m)j_2 , (2n+m)j_1-j_2]$ and $[j_1 -(2n+m)j_2+n, (2n+m)j_1-j_2 -n] \subset f(A)]$. Any elements in $f(A) \cap [j_1 -(2n+m)j_2, j_1-(2n+m)j_2+n-1]$ can only come only from $L+L+L+ \cdots+L-R'-R'-R'-\cdots-R'$, which is simply a translation of $L+L+L+ \cdots+L-R-R-R-\cdots-R$. A similar argument works for the right fringe of $f_{j_1,\ j_2}(A')$. Thus $|f(A')| = |f(A)| +jm$ (this is because the potential width of $f_{j_1,\ j_2}(A')$ is $jm$ more than that of $f_{j_1,\ j_2}(A)$, and the two fringes of these sets are in a 1-1 correspondence). Since $|f_{j_1,\ j_2}(A')| - |f_{j_1,\ j_2}(A)|$ depends only on $j=j_1+j_2$, it holds for any pair of forms with $j$ coefficients, and the lemma is proven. \end{proof} \begin{lem}\label{lem:gen2} For $j \ge 3$, any $P_n^2$-set is also a $P^j_n$-set. \end{lem} \begin{proof} Let A be a $P_n^2$-set, where $A \subset [1, k]$ and $1, k \in A$. Assume $k \geq 2n$. Then $A+A \cap [n+2, 2k-n] = [n+2, 2k-n]$ (as $A$ is a $P_n^2$-set). Let $f_{j_1,\ j_2}$ be a form with $j \ge 3$, and thus either $j_1$ or $j_2$ is at least 2; without loss of generality we assume $j_1 \ge 2$. There is a form $f_{j_1-2,\ j_2}$ such that $f_{j_1-2,\ j_2}(A) +A+A = f_{j_1,\ j_2}(A)$. The proof follows by showing $f_{j_1-2,\ j_2}(\{1, k\})+A+A$ contains all necessary elements, namely $[j_1-kj_2 +n, j_1k - j_2 -n]$. (By $f_{j_1-2,\ j_2}(\{1, k\})$ we mean all numbers of the form $\gep_1 a_1 + \cdots + \gep_{j-2} a_{j-2}$, with the $\gep_i$ the coefficients of the form $f_{j_1-2,j_2}$ and $a_i \in \{1,k\}$.) We have \begin{equation} f_{j_1-2,\ j_2}(\{1, k\})\ \supset \ \{j_1 - 2 -i +k(i-j_2) \;|\; 0\leq i \leq j -2\}. \end{equation} To see this, we first consider $i \le j_1-2$. For such $i$, for the positive summands choose $1$ a total of $j_1-2-i$ times and $k$ a total of $i$ times, while for the negative summands we choose $k$ each of the $j_2$ times. If now $j_1 - 2 < i \le j-2$, for the positive summands we choose $k$ a total of $i-j_2$ times (which is permissible as this is at most $j_1-2$) and we choose 1 the remaining $j_1-2-(i-j_2)$ times, while for the negative summands we choose $1$ all $j_2$ times. This leads to a sum of $k\cdot (i-j_2) + 1\cdot(j_1-2+j_2-i)-1\cdot j_2$, which equals $j_1-2-i+ k(i-j_2)$ as claimed. Unfortunately, this argument fails if $i=j_1-1$ and $j_1=j_2$, as we would then be choosing $k$ from the positive summands negative one times.\footnote{This is the only bad case we need consider, as we know $j_1 \ge j_2$, and the only problem arises when $i-j_2 < 0$.} We are thus left with showing that we may obtain the sum $-1-k$ in this special case. As $j_1=j_2$, we just choose $1$ for the $j_1-2$ positive summands and $-1$ for all but one of the $j_2$ negative summands (where we choose one to be $k$). As $A$ is a $P_n^2$-set, $A+A \supset [n+2, 2k-n]$. Thus \begin{eqnarray} \bigcup_{i=0}^{j-2} [L_i, U_i] & \ \subset \ & f_{j_1-2,\ j_2}(\{1, k\})+A+A, \end{eqnarray} where \begin{eqnarray} L_i & \ = \ & j_1 - 2 -i +k(i-j_2)+ n+2 \nonumber\\ U_i &=& j_1 - 2 -i +k(i-j_2) +2k-n. \end{eqnarray} We see that $L_0 = j_1-kj_2 +n$ and $U_{j-2} = j_1k - j_2 -n$, our two desired endpoints. The proof is completed by showing the intervals $[L_i,U_i]$ cover the desired interval and has no gap with its neighbors. Since $2n \leq k$, we have: \begin{eqnarray} L_i -1 & \ = \ & j_1 -i +k(i-j_2)+ n -1\nonumber\\ &=& (j_1-i+ki-j_2k -1) +n\nonumber\\ & \le & (j_1-i+ki-j_2k -1) +k- n \nonumber\\ &=& j_1 - 2 -(i-1) +k((i-1)-j_2) +2k-n \nonumber\\ &\le & U_{i-1}. \end{eqnarray} Thus there are no gaps between the intervals $[L_{i-1}, U_{i-1}], \; [L_i, U_i]$ and they therefore cover the necessary range. \end{proof} \begin{rek} Note that the above lemma is false if the size of $n$ is unrestricted. To take an extreme example, let $A = \{1, 10\}$ and $n= 9$. Then $A$ is a $P_n^2$-set ($11 \in A+A, \; 0 \in A-A$) but $A$ is not a $P^3_n$-set. \end{rek} \begin{proof}[Proof of Theorem \ref{thm:genconstr}] Lemmas \ref{lem:gen1} and \ref{lem:gen2} imply that the sets described in Lemma \ref{lem:mainconstr} also work in our generalized case. The counting argument of \S\ref{sec:lowerboundspercentage} requires no modification. Thus the theorem is proved \emph{provided} we can find an $A$ to start the process. The following set was obtained by taking elements in $\{2,\dots,49\}$ to be in $A$ with probability\footnote{Note the probability is 1/3 and not 1/2.} $1/3$ (and, of course, requiring $1, 50 \in A$); it took about 300000 sets to find the first one satisfying our conditions: \be A \ = \ \{1, 2, 5, 6, 16, 19, 22, 26, 32, 34, 35, 39, 43, 48, 49, 50\}. \ee To be a $P_{25}^3$-set we need to have $A+A+A \supset [n + 3, 6 n - n] = [28, 125]$ and $A+A-A \supset [-n + 2, 3 n - 1] = [-23, 74]$. A simple calculation shows $A+A+A = [3,150]$, all possible elements, while $A+A-A = [-48, 99] \backslash \{-34\}$ (i.e., every possible element but -34). Thus $A$ is a $P_{25}^3$-set satisfying $|A+A+A| > |A+A-A|$, and thus we have the example we need to prove Theorem \ref{thm:genconstr}. \end{proof} \begin{rek} We could also have taken \be A \ = \ \{1, 2, 3, 4, 8, 12, 18, 22, 23, 25, 26, 29, 30, 31, 32, 34, 45, 46, 49, 50\}, \ee which has the same $A+A+A$ and $A+A-A$. \end{rek} \section{Concluding remarks and future research}\label{sec:concremfutureresearch} One avenue of future research is to complete the proof of Conjecture \ref{conj:genconstr} and give an elementary example of an infinite family of sets satisfying $|f_{j_1,\ j_2}(A)| > |f_{j_1',\ j_2'}(A)|$. We have reason to believe the correct model is to look for $P_n^j$-sets by choosing the numbers $\{2,\dots,2n-1\}$ to be in $A$ with probability $1/j$ (and, of course, requiring $1, 2n \in A$). Unfortunately the density of such sets appears to decrease rapidly with $n$, and to date straightforward computer searches have been unsuccessful when $j=4$. As we shall see below, perhaps a better algorithm would incorporate choosing elements near the fringes (i.e., near $1$ and $2n$) with a different probability than $1/j$. \\ We also observed earlier (Footnote \ref{footnote:beingpn}) that for a constant $0<\alpha \le 1$, a set randomly chosen from $[1,2n]$ is a $P_{\lfloor \alpha n \rfloor}$-set with probability approaching $1$ as $n\to\infty$. MSTD sets are of course not random, but it seems logical to suppose that this pattern continues. \begin{conj}\label{conj:MSTDsP_n} Fix a constant $0<\alpha\le 1/2$. Then as $n\to\infty$ the probability that a randomly chosen MSTD set in $[1,2n]$ containing $1$ and $2n$ is a $P_{\lfloor \alpha n \rfloor}$-set goes to $1$. \end{conj} In our construction and that of \cite{MO}, a collection of MSTD sets is formed by fixing the fringe elements and letting the middle vary. The intuition behind both is that the fringe elements matter most and the middle elements least. Motivated by this it is interesting to look at all MSTD sets in $[1,n]$ and ask with what frequency a given element is in these sets. That is, what is \be \gamma(k;n) \ = \ \frac{\#\{A: k \in A\ {\rm and}\ A\ {\rm is\ an\ MSTD\ set}\}}{\#\{A: \ A\ {\rm is\ an\ MSTD\ set}\}} \ee as $n\to\infty$? We can get a sense of what these probabilities might be from Figure \ref{fig:MSTDfreq}. \begin{figure} \begin{center} \scalebox{1.75}{\includegraphics{freqn100kvaries.eps}} \caption{\label{fig:MSTDfreq}\textbf{Estimation of $\gamma(k,100)$ as $k$ varies from $1$ to $100$ from a random sample of 4458 MSTD sets.}} \end{center}\end{figure} Note that, as the graph suggests, $\gamma$ is symmetric about $\frac{n+1}2$, i.e. $\gamma(k,n)=\gamma(n+1-k,n)$. This follows from the fact that the cardinalities of the sumset and difference set are unaffected by sending $x \to \alpha x + \beta$ for any $\alpha, \beta$. Thus for each MSTD set $A$ we get a distinct MSTD set $n+1-A$ showing that our function $\gamma$ is symmetric. These sets are distinct since if $A=n+1-A$ then $A$ is sum-difference balanced.\footnote{The following proof is standard (see, for instance, \cite{Na2}). If $A = n+1-A$ then \be |A+A| \ = \ |A + (n+1-A)| \ = \ |n+1+(A-A)| \ = \ |A-A|.\ee} From \cite{MO} we know that a positive percentage of sets are MSTD sets. By the central limit theorem we then get that the average size of an MSTD set chosen from $[1,n]$ is about $n/2$. This tells us that on average $\gamma(k,n)$ is about $1/2$. The graph above suggests that the frequency goes to $1/2$ in the center. This leads us to the following conjecture: \begin{conj}\label{conj:50conjecture} Fix a constant $0<\alpha<1/2$. Then $\lim_{n\rightarrow\infty}{\gamma(k,n)}=1/2$ for $\lfloor \alpha n \rfloor \le k \le n - \lfloor \alpha n \rfloor$. \end{conj} \begin{rek} More generally, we could ask which non-decreasing functions $f(n)$ have $f(n)\rightarrow \infty$, $n-f(n)\rightarrow \infty$ and $\lim_{n\rightarrow\infty}{\gamma(k,n)}=1/2$ for all $k$ such that $\lfloor f(n) \rfloor \le k \le n - \lfloor f(n) \rfloor$. \end{rek}

https://arxiv.org/abs/1402.5826

Depth and Stanley Depth of the Canonical Form of a factor of monomial ideals

In this paper we show that the depth and the Stanley depth of the factor of two monomial ideals is invariant under taking a so called canonical form. It follows easily that the Stanley Conjecture holds for the factor if and only if it holds for its canonical form. In particular, we construct an algorithm which simplifies the depth computation and using the canonical form we massively reduce the run time for the sdepth computation.

\section{Introduction} \vskip 1cm Let $K$ be a field and $S=K[x_1,\ldots,x_n]$ be the polynomial ring over $K$ in $n$ variables. A Stanley decomposition of a graded $S-$module $M$ is a finite family $$\mathcal{D} = (S_i, u_i)_{i \in I}$$ in which $u_i$ are homogeneous elements of $M$ and $S_i$ are graded $K-$algebra retract if $S$ for all $i \in I$ such that $S_i \cap \Ann (u_i) = 0$ and $$M = \DSum{i \in I}{} S_iu_i$$ as a graded $K-$vector space. The Stanley depth of $\mathcal{D}$, denoted by $\sdepth (\mathcal{D})$, is the depth of the $S-$module $\DSum{i \in I}{} S_iu_i$. The Stanley depth of $M$ is defined as $$ \sdepth\ (M) :=\max\{\sdepth \ ({\mathcal D})\ |\ {\mathcal D}\; \text{is a Stanley decomposition of}\; I \}.$$ Another definition of sdepth using partitions is given in \cite{HVZ}. Stanley's Conjecture \cite{S} states that the Stanley depth \\ $\sdepth (M)$ is $\geq \depth\ (M)$. Let $J \subsetneq I \subset S$ be two monomial ideals in $S$. In \cite{IKF}, Ichim et. al. studied the sdepth and depth of the factor $\nicefrac{I}{J}$ under polarization and reduced the Stanley's Conjecture to the case when the ideals are monomial squarefree. This is possible the best result from the last years concerning Stanley's depth. It is worth to mention that this result is not very useful for computing sdepth since it introduces a lot of new variables. In the squarefree case there are not many known results about the Stanley conjecture (see for example \cite{PP}). Another result of \cite{IKF} which helps in the sdepth computation is the following proposition, which extends \cite[Lemma 1.1]{Ci}, \cite[Lemma 2.1]{IQ}. \begin{proposition}\label{main} \cite[Proposition 5.1]{IKF} Let $k\in {\mathbb N}$ and $I''$, $J''$ be the monomial ideals obtained from $I$, $J$ in the following way: Each generator whose degree in $x_n$ is at least $k$ is multiplied by $x_n$ and all other generators are taken unchanged. Then $\sdepth_S\nicefrac{I}{J} = \sdepth_S\nicefrac{I''}{J''}$. \end{proposition} Inspired by this proposition we introduce a canonical form of a factor $\nicefrac{I}{J}$ of monomial ideals (see Definition \ref{def: canonical quot}) and we prove easily that sdepth is invariant under taking the canonical form (see Theorem \ref{prop:sdepth}). This leads us to the idea to study also the depth case (see Theorem \ref{prop:depth}). Theorem \ref{m} says that Stanley's Conjecture holds for a factor of monomial ideals if and only if it holds for its canonical form. As a side result, in the depth (respectively sdepth) computation algorithm for $\nicefrac{I}{J}$, one can first compute the canonical form and use the algorithm on this new much more simpler module (see the Appendix). In Example \ref{ex: timings} we conclude that the $\depth$ and $\sdepth$ algorithms are faster when considering the canonical form: using $\textsc{CoCoA}$\cite{Co}, $\textsc{Singular}$\cite{Sing} and Rinaldo's $\sdepth$ computation algorithm \cite{Rina} we see a small decrease in the $\depth$ case timing, but in the $\sdepth$ case the run time is massively reduced. We hope that our algorithm together with the one from \cite{AP} will be used very often in problems concerning monomial ideals. We owe thanks to Y.-H. Shen who noticed our results in a previous arXiv version and showed us the papers of Okazaki and Yanagawa \cite{OY} and \cite{Y}, because they are strongly connected with our topic. Indeed Proposition \ref{main} and Corollary \ref{prop:5.1 for depth } follow from \cite[Theorem 5.2]{OY} (see also \cite[Section 2,3]{OY}). However, our proofs of Lemma \ref{lemma:depth is constant} and Corollary \ref{prop:5.1 for depth } are completely different from those appeared in the quoted papers and we keep them for the sake of our completeness. \vskip 1cm \section{The canonical form of a factor of monomial ideals} \vskip 1cm Let $R=K[x_1,\ldots,x_{n-1}]$ be the polynomial $K$-algebra over a field $K$ and $S := R[x_n]$. Consider $J \subsetneq I\subset R$ two monomial ideals and denote by $G(I)$, respectively $G(J)$, the minimal (monomial) system of generators of $I$, respectively $J$. \begin{definition} \label{def: canonical} {\em The power $x_n^r$ {\em enters in a monomial} $u$ if $x_n^r|u$ but $x_n^{r+1}\nmid u$. We say that $I$ is {\em of type} $(k_1,\ldots,k_s)$ {\em with respect to} $x_n$ if $x_n^{k_i}$ are all the powers of $x_n$ which enter in a monomial of $G(I)$ for $i\in [s]$ and $1\leq k_1<\ldots<k_s$. $I$ is {\em in the canonical form with respect to} $x_n$ if $I$ is of type $(1,\ldots,s)$ for some $s\in {\mathbb N}$. We simply say that $I$ is {\em the canonical form} if it is in the canonical form with respect to all variables $x_1, \ldots, x_n$. } \end{definition} \begin{remark} \label{rem:canonical form}{\em Suppose that $I$ is of type $(k_1,\ldots,k_s)$ with respect to $x_n$. It is easy to get the {\em canonical form} $I'$ of $I$ {\em with respect to} $x_n$: replace $x_n^{k_i}$ by $x_n^i$ whenever $x_n^{k_i}$ enters in a generators of $G(I)$. Applying by recurrence this procedure for other variables we get the {\em canonical form} of $I$, that is with respect to all variables. Note that a squarefree monomial ideal is of type $(1)$ with respect to each $x_i$ and it is in the canonical form with respect to $x_i$, so in this case $I'=I$. } \end{remark} \begin{definition} \label{def: canonical quot} {\em Let $J \subsetneq I \subset S$ two monomial ideals. We say that $\nicefrac{I}{J}$ is {\em of type} $(k_1, \ldots, k_s)$ {\em with respect to } $x_n$ if $x_n^{k_i}$ are all the powers of $x_n$ which enter in a monomial of $G(I) \cup G(J)$ for $i \in [s]$ and $1 \leq k_1 < \ldots < k_s $. All the terminology presented in Definition \ref{def: canonical} will extend automatically to the factor case. Thus we may speak about the {\em canonical form} $\overline{\nicefrac{I}{J}}$ of $\nicefrac{I}{J}$. } \end{definition} \begin{remark} \label{rem:canonical quot} {\em In order to compute the canonical form with respect to $x_n$ of the $(k_1, \ldots, k_s)-$type factor $\nicefrac{I}{J}$, one will replace $x_n^{k_i}$ by $x_n^i$ whenever $x_n^{k_i}$ enters a generator of $G(I) \cup G(J)$. } \end{remark} \begin{example}\label{ex:canonical} {\em We present some examples where we compute the canonical form of a monomial ideal, respectively a factor of two monomial ideals. \\ \begin{enumerate} \item Consider $S = \mathbb Q [x,y]$ and the monomial ideal $I = (x^4,x^3y^7)$. Then the canonical form of $I$ is $I' = (x^2, xy)$. \item Consider $S = \mathbb Q [x,y,z]$, $I = (x^{10}y^5, x^4yz^7,z^7y^3)$ and \\ $J = (x^{10}y^{20}z^2, x^3y^4z^{13}, x^9y^2z^7)$. The canonical form of $\nicefrac{I}{J}$ is $\overline{\nicefrac{I}{J}} = \fracs{(x^4y^5, x^2yz^2, y^3z^2)}{(x^4y^6z, xy^4z^3, x^3y^2z^2)}$. \end{enumerate}} \end{example} The canonical form of a factor of monomial ideals $\nicefrac{I}{J}$ is not usually the factor of the canonical forms of $I$ and $J$ as shows the following example. \begin{example}{\em Let $S = \mathbb Q [x, y]$, $I = (x^4, y^{10}, x^2y^7)$ be and $J = (x^{20}, y^{30})$. The canonical form of $I$ is $I' = (x^2, y^2, xy)$ and the canonical form of $J$ is $J' = (x,y)$. Then $J'\not\subset I'$. But the canonical form of the factor $\nicefrac{I}{J}$ is $\overline{\nicefrac{I}{J}} = \fracs{(x^2,y^2,xy)}{(x^3, y^3)}$. } \end{example} Using Proposition \ref{main}, we see that the Stanley depth of a monomial ideal does not change when considering its canonical form. \begin{theorem}\label{prop:sdepth} Let $I$, $J$ be monomial ideals in $S$ and $\overline{\nicefrac{I}{J}}$ the canonical form of $\nicefrac{I}{J}$. Then $$\sdepth_S \nicefrac{I}{J}=\sdepth_S \overline{\nicefrac{I}{J}}.$$ \end{theorem} The proof goes applying inductively the following lemma. \begin{lemma}\label{lemma: sdepth} Suppose that $\nicefrac{I}{J}$ is of type $(k_1,\ldots,k_s)$ with respect to $x_n$ and $k_j+1<k_{j+1}$ for some $0\leq j<s$ (we set $k_0=0$). Let $G(I')$ (resp. $G(J')$) be the set of monomials obtained from $G(I)$ (resp. $G(J)$) by substituting $x_n^{k_i} $ by $x_n^{k_i-1}$ for $i>j$ whenever $x_n^{k_i}$ enters in a monomial of $G(I)$ (resp. $G(J)$). Let $I'$ and $J'$ be the ideals generated by $G(I')$ and $G(J')$. Then $$\sdepth_S \nicefrac{I}{J}=\sdepth_S \nicefrac{I'}{J'}.$$ \end{lemma} The proof of Lemma \ref{lemma: sdepth} follows from the proof of \cite[Proposition 5.1]{IKF} (see here Proposition \ref{main}). Next we focus on the $\depth \nicefrac{I}{J}$ and $\depth \overline{\nicefrac{I}{J}}$. The idea of the proof of the following lemma is taken from \cite[Section 2]{P}. \begin{lemma}\label{lemma:depth is constant} Let $I_0 \subset I_1 \subset \ldots \subset I_e \subset R$, $J\subset S$, $U_0 \subset U_1 \subset \ldots \subset U_e \subset R$, $V \subset S$ be some graded ideals of $S$, respectively $R$, such that $U_i \subset I_i$ for $0 \leq i \leq e$, $I_e \subset J$, $V \subset J$ and $U_e \subset V$. Consider $T_k = \Sum{i=0}{e}x_n^i I_i S + x_n^k J$ and $W_k = \Sum{i=0}{e}x_n^iU_iS + x_n^kV$ for $k > e$. Then $\depth_S \fracs{T_k}{W_k}$ is constant for all $k>e$. \end{lemma} \begin{proof} Consider the following linear subspaces of $S$: $I := \Sum{i=0}{e}x_n^i I_i$ and $U := \Sum{i=0}{e}x_n^i U_i$. Note that $I$ and $U$ are not ideals in $S$. If $I = U$, then the claim follows easily from the next chain of isomorphisms $\fracs{T_k}{W_k} \iso \fracs{x_n^kJ}{x_n^kJ \cap (I+x_n^kV)S} \iso \fracs{x_n^kJ}{x_n^k(I+V)S} \iso \fracs{J}{(I+V)S}$ for all $k>e$, and hence $\depth_S \fracs{T_k}{W_k}$ is constant for all $k>e$. Assume now that $I \neq U$ and consider the following exact sequence $$0 \to \fracs{J}{V} \xto{\cdot x_n^k} \fracs{T_k}{W_k} \to \fracs{T_k}{W_k + x_n^kJ} \to 0,$$ where the last term we denote by $H_k$. Note that $H_k \iso \fracs{IS}{IS \cap (U+x_n^kJ)S}$ and $IS \cap (U + x_n^kJ)S = US + x_n^kIS$. Since $x_n^kH_k=0$, $H_k$ is a $\nicefrac{S}{(x_n^k)}-$module. Then $\depth_S H_k = \depth_{\nicefrac{S}{(x_n^k)}}H_k=\depth_R H_k$ because the graded maximal ideal $m$ of $R$ generates a zero dimensional ideal in $\nicefrac{S}{(x_n^k)}$. But $H_k$ over $R$ is isomorphic with $\fracs{\oplus_{i=0}^{k-1} I_i}{\oplus_{i=0}^{k-1} U_i} \iso \bigoplus_{i=0}^{k-1} \fracs{I_i}{U_i}$, where $I_i=I_e$ and $U_i=U_e$ for $e<i<k$. It follows that $t:=\depth_S H_k = \min_i\left\{\depth_{R} \fracs{I_i}{U_i}\right\}$. If $\depth_S \fracs{J}{V} = 0$, then the Depth Lemma gives us $\depth_S \fracs{T_k}{W_k} = t = 0$ for all $k>e$ and hence we are done. Therefore we may suppose that $\depth_S \fracs{J}{V} > 0$. Note that $t>0$ implies $\depth_S \fracs{T_k}{W_k} > 0$ by the Depth Lemma since otherwise $\depth_S \fracs{T_k}{W_k} = \depth_{S} \fracs{J}{V} = 0$, which is false. Next we will split the proof in two cases. $\circ$ Case $t = 0$. Let ${\mathcal F}=\big\{i\in \{0,\ldots,e\}\ \big|\ \depth_R\nicefrac{I_i}{U_i}=0\big\}$ and $L_i\subset I_i$ be the graded ideal containing $U_i$ such that $\nicefrac{L_i}{U_i} \iso H_m^0(\nicefrac{I_{i}}{U_{i}})$. If $i\in {\mathcal F}$ and there exists $u\in ( L\cap V)\setminus U_i$ then $(m^s,x_n^k)x_n^i u\subset W_k$ for some $s\in {\mathbb N}$, that is $\depth_S \fracs{T_k}{W_k} = 0$ for all $k > e$. Now consider the case when $L_i\cap V=U_i$ for all $i\in {\mathcal F}$. If $i\in {\mathcal F}$ then note that $L_i\subset L_j$ for $i< j\leq e$. Set $V'=V+L_eS$, $U'=U+ \Sum{i\in {\mathcal F}}{}x_n^i L_i$ and $W'_k := U'S+x_n^kV'=U'S+x_n^kV$ because $x_n^kL_eS\subset U'S$. Consider the following exact sequence $$0 \to \fracs{W'_k}{W_k} \to \fracs{T_k}{W_k} \to \fracs{T_k}{W_k'} \to 0.$$ For the last term we have $H_m^0(\nicefrac{I_j}{U'_j})=0$, $0\leq j\leq e$ and so the new $t>0$, which is our next case. Thus we get $\depth_S \fracs{T_k}{W'_k}>0$ is constant for $k>e$. The first term is isomorphic to $\fracs{U'S}{U'S\cap W_k}$. But $U'S\cap W_k=US+(U'S\cap x_n^kV)$ since $US\subset U'S$. Since $U'S\cap (x_n^kS)=x_n^k (U_e+L_e)S$ and $U_e\subset V$ it follows that $U'S\cap x_n^kV=x_n^kUS+(x_n^kL_eS\cap x_n^kVS)=x_n^kUS$. Consequently, the first term from the above exact sequence is isomorphic with $\fracs{U'S}{US}$. Note that the annihilator of the element induced by some $u\in L_e\setminus V$ in $\nicefrac{U'S}{US}$ contains a power of $m$ and so $\depth_S \fracs{U'S}{US}\leq 1$. The inequality is equality since $x_n$ is regular on $\nicefrac{U'S}{US}$. By the Depth Lemma we get $\depth_S \fracs{T_k}{W_k}=1$ for all $k>e$. $\circ$ Case $t>0$. If $\depth_{R} \fracs{J}{V}\leq t= \depth_S H_k$ then the Depth Lemma gives us again the claim, i.e. $\depth_S \fracs{T_k}{W_k} = \depth_S \fracs{J}{V}$ for all $k>e$. Assume that $\depth_{S} \fracs{J}{V} >t$. Apply induction on $t$, the initial step $t = 0$ being done in the first case. Suppose that $t>0$. Then $\depth_S \fracs{J}{V}>t>0$ implies that $\depth_S \fracs{J}{V}\geq 2$ and so we may find a homogeneous polynomial $f \in m$ that is regular on $\fracs{J}{V}$. Moreover we may find $f$ to be regular also on all $\fracs{I_i}{U_i}$, $i \leq e$. Then $f$ is regular on $\fracs{T_k}{W_k}$. Set $V'' := V + f J$ and $U''_i := U_i + f I_i$ for all $i \leq e$ and set $W''_k := \Sum{i=0}{e}x_n^iU''_iS + x_n^kV''$. By Nakayama's Lemma we get $U'' \neq U$, and therefore $\depth_{R} \fracs{I}{U''} = t-1$ and by induction hypothesis it results that $\depth_S \fracs{T_k}{W_k} = 1 + \depth_S \fracs{T_k}{W''_k} = $ constant for all $k > e$. Finally, note that we may pass from the first case to the second one and conversely. In this way $U$ increases at each step. By Noetherianity at last we may arrive in finite steps to the case $I=U$, which was solved at the beginning. \hfill\ \end{proof} The next corollary is in fact \cite[Proposition 5.1]{IKF} (see Proposition \ref{main}) for depth. It follows easily from Lemma \ref{lemma:depth is constant} but also from \cite[Proposition 5.2]{OY} (see also \cite[Sections 2, 3]{Y}. \begin{corollary}\label{prop:5.1 for depth } Let $e \in \mathbb N$, $I$ and $J$ monomial ideals in $S := K[x_1, \ldots, x_n]$. Consider $I'$ and $J'$ be the monomial ideals obtained from $I$ and $J$ in the following way: each generator whose degree in $x_n$ $\geq e$ is multiplied by $x_n$ and all the other generators are left unchanged. Then $$\depth_S \nicefrac{I}{J} = \depth_S \nicefrac{I'}{J'}.$$ \end{corollary} This leads us to the equivalent result of Theorem \ref{prop:sdepth} for depth. \begin{theorem} \label{prop:depth} Let $I$ and $J$ be two monomial ideals in $S$ and $\overline{\nicefrac{I}{J}}$ the canonical form of $\nicefrac{I}{J}$. Then $$\depth_S \nicefrac{I}{J} = \depth_S \overline{\nicefrac{I}{J}}.$$ \end{theorem} \begin{proof} Assume that $\nicefrac{I}{J}$ is of type $(k_1,\ldots,k_s)$ with respect to $x_n$ and obviously $\overline{\nicefrac{I}{J}}$ is of type $(1,2,\ldots,s)$ with respect to $x_n$. Starting with $\overline{\nicefrac{I}{J}}$, we apply Corollary \ref{prop:5.1 for depth } till we obtain an $\nicefrac{I'_1}{J'_1}$ of type $(k_1,k_1+1,\ldots,k_1+s-1)$ having the same depth as $\overline{\nicefrac{I}{J}}$. We repeat the process until we get $\nicefrac{I'_s}{J'_s}$ of type $(k_1, k_2, \ldots, k_s)$ with respect to $x_n$ with the unchanged depth. Now we iterate and take the next variable. At the very end the claim will follow.\hfill \ \end{proof} Theorem \ref{prop:sdepth} and Theorem \ref{prop:depth} give us the following theorem \begin{theorem}\label{m} The Stanley conjecture holds for a factor of monomial ideals $\nicefrac{I}{J}$ if and only if it holds for its canonical form $\overline{\nicefrac{I}{J}}$. \end{theorem} Using Theorem \ref{prop:depth}, instead of computing the $\depth$ or the $\sdepth$ of $\nicefrac{I}{J}$, $J \subsetneq I \subset S$, we can compute it for the simpler module $\overline{\nicefrac{I}{J}}$. \begin{example} \label{ex: timings} {\em We present the different timings for the depth and sdepth computation algorithms with and without extracting the canonical form. $\textsc{Singular}$\cite{Sing} was used in the depth computations while $\textsc{CoCoA}$ \cite{Co} and Rinaldo's paper\cite{Rina} were used for the Stanley depth computation. \begin{enumerate} \item Consider the ideals from Example \ref{ex:canonical}(2). Timing for $\sdepth \nicefrac{I}{J}$ computation: 22s. Timing for $\sdepth \overline{\nicefrac{I}{J}}$ computation: 74 ms. \item Consider $R = \mathbb Q[x,y,z]$ and $I = (x^{100}yz,x^{50}yz^{50},x^{50}y^{50}z)$. Then the canonical form is $I' = (x^2yz,xyz^2,xy^2z)$. Timing for $\sdepth I$ computation: 13m 3s. Timing for $\sdepth I'$ computation: 21 ms. Notice that the difference in timings is very large. Therefore using the canonical form in the $\sdepth$ computation is a very important optimization step. On the other side, the $\depth$ computation is immediate in both cases. In the last example, the timing difference can be seen. \item Consider $R = \mathbb{Q}[x,y,z,t,v,a_1,\ldots,a_5]$,\\ $I = (v^4x^{12}z^{73},v^{87}t^{21}y^{13},x^{43}y^{18}z^{72}t^{28},vxy,vyz,vzt,vtx,a_1^{7000}, a_2^{413};)$, \\$J = (v^5x^{13}z^{74},v^{88}t^{22}y^{14},x^{44}y^{19}z^{73}t^{29},v^2x^2y^2,v^2y^2z^2,v^2z^2t^2,v^2t^2x^2)$. Timing for $\depth \nicefrac{I}{J}$ computation: 16m 11s. Timing for $\depth \overline{\nicefrac{I}{J}}$ computation: 11m. \end{enumerate} } \end{example} \vskip 1cm \section{Appendix} \vskip 1cm We sketch the simple idea of the algorithm which computes the canonical form of a monomial ideal $I$. This can easily be extended to compute the canonical form of $\nicefrac{I}{J}$ by simple applying it for $G(I) \cup G(J)$ and afterwards extracting the generators corresponding to $I$ and $J$. This was used in Example \ref{ex: timings}. The algorithm is based on Remark \ref{rem:canonical quot}: for each variable $x_i$ we build the list \verb"gp" in which we save the pair $(g,p)$, were $p$ is chosen such that $x_i^p$ enters the $g-$generator of the monomial ideal $I$. This list will be sorted by the powers $p$ as in the following example \begin{example} {\em Consider the ideal $I := (x^{13}, x^{4}y^{7}, y^7z^{10}) \subset \mathbb{Q} [x, y, z]$. Then for each variable we will obtain a different \verb"gp" as shown below: \begin{itemize} \item[$\circ$] For the first variable $x$, \verb"gp" is equal to \begin{tabular}{ |c | c | c |c| } \hline 2 & 4 & 1 & 13 \\ \hline \end{tabular}. Therefore $I$ is of type $(4,13)$ with respect to $x$. Hence, in order to obtain the canonical form with respect to $x$, one has to divide the second generator by $x^{4-1} = x^3$ and the first generator by $x^{13-2} = x^{11}$. After these computation we will get $I_1 = (x^2, xy^7, y^7z^{10})$. Note that $I_1$ is in the canonical form w.r.t. $x$. \item[$\circ$] For the second variable $y$, \verb"gp" is equal to \begin{tabular}{ |c | c | c |c| } \hline 3 & 7 & 2 & 7 \\ \hline \end{tabular}. Similar as above, one has to divide the second and the third generator by $y^6$, and hence it results $I_2 = (x^2, xy, yz^{10})$. Again, $I_2$ is in the canonical form w.r.t. $y$ and $x$. \item[$\circ$] For the last variable $z$, \verb"gp" is equal to \begin{tabular}{ |c | c | } \hline 3 & 10 \\ \hline \end{tabular}. We divide the third generator of $I_2$ by $z^9$ and we get our final result $I' = (x^2, xy, yz)$., which is in the canonical form with respect to all variables. \end{itemize} } \end{example} Based on the above idea, we construct two procedures: \verb"putIn" and \verb"canonical" $-$ the first one constructing the list \verb"gp", and the second one computing the canonical form of a monomial ideal. The proof of correctness and termination is trivial. The procedures were written in the $\textsc{Singular}$ language. \begin{verbatim} proc putIn(intvec v, int power, int nrgen) { if(size(v) == 1) { v[1] = nrgen; v[2] = power; return(v); } int i,j; if(power <= v[2]) { for(j = size(v)+2; j >=3; j--) { v[j] = v[j-2]; } v[1] = nrgen; v[2] = power; return(v); } if(power >= v[size(v)]) { v[size(v)+1] = nrgen; v[size(v)+1] = power; return(v); } for(j = size(v) + 2; (j>=4) && (power < v[j-2]); j = j-2) { v[j] = v[j-2]; v[j-1] = v[j-3]; } v[j] = power; v[j-1] = nrgen; return(v); } proc canonical(ideal I) { int i,j,k; intvec gp; ideal m; intvec v; v = 0:nvars(basering); for(i = 1; i<=nvars(basering); i++) { gp = 0; v[i] = 1; for(j = 1; j<=size(I); j++) { if(deg(I[j],v) >= 1) { gp = putIn(gp,deg(I[j],v),j); } } k = 0; if(size(gp) == 2) { I[gp[1]] = I[gp[1]]/(var(i)^(gp[2]-1)); } else { for(j = 1; j<=size(gp)-2;) { k++; I[gp[j]] = I[gp[j]]/(var(i)^(gp[j+1]-k)); j = j+2; while((j<=size(gp)-2) && (gp[j-1] == gp[j+1]) ) { I[gp[j]] = I[gp[j]]/(var(i)^(gp[j+1]-k)); j = j + 2; } } if(j == size(gp)-1) { if(gp[j-1] == gp[j+1]) { I[gp[j]] = I[gp[j]]/(var(i)^(gp[j+1]-k)); } else { k++; I[gp[j]] = I[gp[j]]/(var(i)^(gp[j+1]-k)); } } } v[i] = 0; } return(I); } \end{verbatim}

https://arxiv.org/abs/1407.5210

Faster convergence rates of relaxed Peaceman-Rachford and ADMM under regularity assumptions

Splitting schemes are a class of powerful algorithms that solve complicated monotone inclusion and convex optimization problems that are built from many simpler pieces. They give rise to algorithms in which the simple pieces of the decomposition are processed individually. This leads to easily implementable and highly parallelizable algorithms, which often obtain nearly state-of-the-art performance.In this paper, we provide a comprehensive convergence rate analysis of the Douglas-Rachford splitting (DRS), Peaceman-Rachford splitting (PRS), and alternating direction method of multipliers (ADMM) algorithms under various regularity assumptions including strong convexity, Lipschitz differentiability, and bounded linear regularity. The main consequence of this work is that relaxed PRS and ADMM automatically adapt to the regularity of the problem and achieve convergence rates that improve upon the (tight) worst-case rates that hold in the absence of such regularity. All of the results are obtained using simple techniques.

\section{Conclusion} In this paper, we provided a comprehensive convergence rate analysis of relaxed PRS and ADMM under various regularity assumptions. By appealing to the examples developed in \cite{davis2014convergence}, we showed that several of the convergence rates cannot be improved. All results follow from some combination of a lemma that deduces convergence rates of summable monotonic sequences (Lemma~\ref{lem:sumsequence}), a simple diagram (Figure~\ref{fig:DRSTR}), and fundamental inequalities (Propositions~\ref{prop:DRSupper},~\ref{prop:DRSlower},~\ref{prop:DRSupperfglipschitz}, and~\ref{prop:fundamentaldiff}) that relate the FPR to the objective error of the relaxed PRS algorithm. Thus, together with~\cite{davis2014convergence}, we have developed a comprehensive convergence rate of the relaxed PRS and ADMM algorithms under the standard regularity assumptions in convex optimization. \begin{acknowledgements}D. Davis' work is partially supported by NSF GRFP grant DGE-0707424. W. Yin's work is partially supported by NSF grants DMS-0748839 and DMS-1317602. \end{acknowledgements} \bibliographystyle{spmpsci} \section{From relaxed PRS to ADMM}\label{sec:DRSADMM} The relaxed PRS algorithm can be applied to problem \eqref{eq:simplelinearconstrained}. To this end we define the Lagrangian: $${\mathcal{L}}_\gamma(x,y;w):=f(x)+g(y)-\dotp{w,Ax+By-b} + \frac{\gamma}{2}\|Ax + By - b\|^2.$$ Section~\ref{sec:DRSADMM} presents Algorithm \ref{alg:DRS} applied to the Lagrange dual of ~\eqref{eq:simplelinearconstrained}, which reduces to the following algorithm: \begin{algorithm}[H] \begin{algorithmic} \Require $w^{-1} \in {\mathcal{H}}, x^{-1} = 0, y^{-1} = 0, \lambda_{-1} = \frac{1}{2}$, $\gamma > 0, (\lambda_j)_{j \geq 0} \subseteq (0, 1]$ \For{$k=-1,~0,\ldots$} \State $y^{k+1} = \argmin_{y} {\mathcal{L}}_\gamma(x^k,y;w^k) + \gamma(2\lambda_k - 1) \dotp{By, (Ax^{k} + By^{k} -b)}$\; \State $w^{k+1} = w^{k} - \gamma (Ax^{k} + By^{k+1} - b) - \gamma(2\lambda_k - 1)(Ax^{k} + By^{k} - b)$\; \State $x^{k+1} = \argmin_{x} {\mathcal{L}}_\gamma(x,y^{k+1};w^{k+1})$\; \EndFor \end{algorithmic} \caption{{Relaxed alternating direction method of multipliers (relaxed ADMM)}} \label{alg:ADMM} \end{algorithm} If $\lambda_k\equiv 1/2$, Algorithm \ref{alg:ADMM} recovers the standard ADMM. It is well known that ADMM is equivalent to DRS applied to the Lagrange dual of Problem~\eqref{eq:simplelinearconstrained} \cite{gabay1983chapter}. Thus, if we let \begin{align*} d_f(w) := f^\ast(A^\ast w) && \mathrm{and} && d_g(w) := g^\ast(B^\ast w) - \dotp{w, b}, \end{align*} then relaxed ADMM is equivalent to relaxed PRS applied to the following problem: \begin{align}\label{eq:dualproblem2} \Min_{w \in {\mathcal{G}}} & \; d_f(w) + d_g(w). \end{align} {We make two assumptions regarding $d_f$ and $d_g$. \begin{assump}[Solution existence]\label{assump:additivesubdual} Functions $f, g : {\mathcal{H}} \rightarrow (-\infty, \infty]$ satisfy \begin{align} \zer(\partial d_f + \partial d_g) \neq \emptyset. \end{align} \end{assump} This is a restatement of Assumption~\ref{assump:additivesub}, which we have used in our analysis of the primal case. \begin{assump}\label{assump:precomposgradient} The following differentiation rule holds: \begin{align*} \partial d_f(x) = A^\ast \circ (\partial f^\ast) \circ A && \mathrm{and} && \partial d_g(x) = B^\ast \circ (\partial g^\ast) \circ B. \end{align*} \end{assump} See \cite[Theorem 16.37]{bauschke2011convex} for conditions that imply this identity, of which the weakest are $0 \in \mathrm{sri}(\range(A^\ast) - \dom(f^\ast))$ and $0 \in \mathrm{sri}(\range(B^\ast) - \dom(g^\ast))$, where $\mathrm{sri}$ is the strong relative interior of a convex set.} This assumption may seem strong, but it is standard in the analysis of ADMM because it implies the form in Proposition~\ref{prop:relaxedADMM}. The next proposition shows how the strong convexity and the differentiability of a closed, proper, and convex function transfer to the dual function. \begin{proposition}[Strong convexity and differentiability of the conjugate]\label{prop:strongconvextodual} Suppose that $f : {\mathcal{H}} \rightarrow (-\infty, \infty]$ is closed, proper, and convex. Then the following implications hold: \begin{enumerate} \item If $f$ is $\mu_f$-strongly convex, then $f^\ast$ is differentiable and $\nabla f$ is $({1}/{\mu_f})$-Lipschitz. \item If $f$ is differentiable and $\nabla f$ is $({1}/{\beta})$-Lipschitz, then $f^\ast$ is $\beta$-strongly convex. \end{enumerate} \end{proposition} \begin{proof} See~\cite[Theorem 18.15]{bauschke2011convex}. \qed\end{proof} With Proposition~\ref{prop:strongconvextodual}, we can characterize the strong convexity and differentiability of the dual functions in terms of $A, B$ and $f$ and $g$. We first recall that a linear map $L : {\mathcal{G}} \rightarrow {\mathcal{G}}$ is \emph{$\alpha$-strongly monotone} if for all $x \in {\mathcal{G}}$, the bound $\dotp{Lx, x}_{\mathcal{G}} \geq \alpha \|x\|_{\mathcal{G}}^2$ holds. \begin{proposition}[Strong convexity and differentiability of the dual]\label{prop:stronglymonotone} The following implications hold: \begin{enumerate} \item If $\nabla f$, (respectively $\nabla g$), is $({1}/{\beta})$-Lipschitz and $AA^\ast$ (respectively $BB^\ast$) is $\alpha$-strongly monotone, then $d_f$ (respectively $d_g$) is $\alpha\beta$-strongly convex. \item If $f$, (respectively $g$) is $\mu$-strongly convex, then $d_f$ (respectively $d_g$) is differentiable and $\nabla d_f$ (respectively $\nabla d_g$) is $({\|A\|^2}/{\mu})$ (respectively $({\|B\|^2}/{\mu})$)-Lipschitz. \end{enumerate} \end{proposition} The proof of Proposition~\ref{prop:stronglymonotone} is straightforward, so we omit it. We note that $AA^\ast$ and $BB^\ast$ are always $0$-strongly monotone. Thus, we assume that $AA^\ast$ and $BB^\ast$ are $\alpha_A$ and $\alpha_B$-strongly monotone, respectively, while allowing the cases $\alpha_A = 0$ and $\alpha_B= 0$. In addition, we use the convention that $\widetilde{\nabla} f$ and $\widetilde{\nabla} g$ are always $({1}/{\beta_f})$, and $({1}/{\beta_g})$-Lipschitz, respectively, by allowing the cases $\beta_f = 0$ and $\beta_g = 0$. We carry the following notation throughout the rest of Section~\ref{sec:DRSADMM}: \begin{align} \mu_{d_f} = \beta_f \alpha_A \geq 0 && \mathrm{and} && \mu_{d_g} = \beta_g \alpha_B \geq 0. \end{align} Thus, $d_f$ and $d_g$ are $\mu_{d_f}$ and $\mu_{d_g}$-strongly convex, respectively. Finally, we always assume that $f$ and $g$ are $\mu_f$ and $\mu_g$-strongly convex, respectively, by allowing $\mu_f=0$ and $\mu_g = 0$. We assume that $\|A\|\|B\| \neq 0$, and denote \begin{align} \beta_{d_f} = \frac{\mu_f}{\|A\|^2} \geq 0 && \mathrm{and} && \beta_{d_g} = \frac{\mu_g}{\|B\|^2} \geq 0. \end{align} If $\beta_{d_f}$ is strictly positive, then $d_f$ is differentiable and $\nabla d_f$ is $(1/\beta_f)$-Lipschitz. A similar result holds for $d_g$. Now we apply Algorithm~\ref{alg:DRS} to the dual problem in Equation~\eqref{eq:dualproblem2}. Given $z^0 \in {\mathcal{H}}$, Lemma~\ref{prop:DRSmainidentity} shows that we need to compute the following vectors for all $k\geq 0$: \begin{align}\label{eq:DRSADMM} \begin{cases} w_{d_g}^k &= \mathbf{prox}_{\gamma d_g}(z^k); \\ w_{d_f}^k &= \mathbf{prox}_{\gamma d_f}(2w_{d_g}^k - z^k); \\ z^{k+1} &= z^k + 2\lambda_k(w_{d_f}^k - w_{d_g}^k). \end{cases} \end{align} A detailed proof of Proposition~\ref{prop:relaxedADMM} recently appeared in \cite[Proposition 11]{davis2014convergence}. \begin{proposition}[Relaxed ADMM]\label{prop:relaxedADMM} Let $z^0 \in {\mathcal{G}}$, and let $(z^j)_{j \geq 0}$ be generated by the relaxed PRS algorithm applied to the dual formulation in Equation~(\ref{eq:dualproblem2}). Choose $w_{d_g}^{-1} = z^0, x^{-1} = 0$ and $y^{-1} = 0$ and $\lambda_{-1} = {1}/{2}$. Then we have the following identities starting from $k = -1$: \begin{align*} y^{k+1} &= \argmin_{y \in {\mathcal{H}}_2} g(y) - \dotp{w_{d_g}^{k},Ax^{k} + By - b} + \frac{\gamma}{2} \|Ax^k + By - b + (2\lambda_k - 1)(Ax^{k} + By^{k} -b) \|^2; \\ w_{d_g}^{k+1} &= w_{d_g}^{k} - \gamma (Ax^{k} + By^{k+1} - b) - \gamma(2\lambda_k - 1)(Ax^{k} + By^{k} - b); \\ x^{k+1} &= \argmin_{x \in {\mathcal{H}}_1} f(x) - \dotp{w_{d_g}^{k+1}, Ax + By^{k+1} - b} + \frac{\gamma}{2} \|Ax + By^{k+1} - b\|^2; \\ w_{d_f}^{k+1} &= w_{d_g}^{k+1} - \gamma (Ax^{k+1} + By^{k+1} - b). \end{align*} \end{proposition} \begin{remark} Proposition~\ref{prop:relaxedADMM} proves that $w_{d_f}^{k+1} = w_{d_g}^{k+1} - \gamma (Ax^{k+1} + By^{k+1} - b)$. Recall that by Equation~(\ref{eq:DRSADMM}), $z^{k+1} - z^k = 2\lambda_k(w_{d_f}^{k} - w_{d_g}^{k})$. Therefore, it follows that \begin{align}\label{eq:ADMMfeasibilityFPR} z^{k+1} - z^k &= -2\gamma \lambda_k( Ax^{k} + By^k - b). \end{align} \end{remark} \begin{center} \begin{table} \centering \begin{tabular}{lll} \hline Function & Primal subgradient & Dual subgradient \\ \hline\hline $g$ & $\widetilde{\nabla} g(y^s) = B^\ast w_{d_g}^s$ & $\widetilde{\nabla} d_g(w_{d_g}^s) = By^s - b$ \\\hline $f$ & $\widetilde{\nabla} f(x^s) = A^\ast w_{d_f}^s$ & $\widetilde{\nabla} d_f(w_{d_f}^s) = Ax^s$\\ \hline \end{tabular} \caption{The main subgradient identities used throughout Section~\ref{sec:DRSADMM}. The letter $s$ denotes a superscript (e.g. $s = k$ or $s = \ast$). See \cite{davis2014convergence} for a proof.}\label{table:ADMMsubgradients} \end{table} \end{center} \subsection{Converting dual inequalities to primal inequalities}\label{sec:convertinequalities} The ADMM algorithm generates $5$ sequences of iterates: \begin{align*} (z^j)_{j \geq 0}, (w_{d_f}^j)_{j \geq 0}, \mbox{ and } (w_{d_g}^j)_{j \geq 0} \subseteq {\mathcal{G}} && \mbox{and} && (x^j)_{j \geq 0} \subseteq {\mathcal{H}}_1, (y^j)_{j \geq 0} \subseteq {\mathcal{H}}_2. \end{align*} In this section we recall some inequalities, which were derived in \cite[Section 8.2]{davis2014convergence}, that relate these sequences to each other through the primal and dual objective functions. In the following propositions, $z^\ast$ will denote a fixed point of $T_{\mathrm{PRS}}$. The point $w^\ast := \mathbf{prox}_{\gamma d_g}(z^\ast)$ is a minimizer of the dual problem in Equation~\eqref{eq:dualproblem2}. Finally, we let $x^\ast$ and $y^\ast$ be defined as in Table~\ref{table:ADMMsubgradients}. \begin{proposition}[ADMM primal upper fundamental inequality]\label{prop:ADMMupper} For all $k \geq 0$, we have the bound \begin{align*} 4\gamma \lambda_k (f(x^k) + g(y^k) &- f(x^\ast) - g(y^\ast))\\ &\leq \|z^k - (z^\ast - w^\ast)\|^2 - \|z^{k+1} - (z^\ast - w^\ast)\|^2 + \left(1- \frac{1}{\lambda_k} \right) \|z^{k} - z^{k+1}\|^2. \numberthis \label{prop:ADMMupper:eq:main} \end{align*} \end{proposition} \begin{proposition}[ADMM primal lower fundamental inequality]\label{prop:ADMMlower} For all $x \in \dom(f)$ and $y \in \dom(g)$, we have the bound: \begin{align}\label{prop:ADMMlower:eq:main} f(x) + g(y) - f(x^\ast) - g(y^\ast) &\geq \dotp{Ax + By - b, w^\ast}. \end{align} \end{proposition} \subsection{Converting dual convergence rates to primal convergence rates}\label{sec:dualtoprimalrates} In this section, we use the inequalities deduced in Section~\ref{sec:convertinequalities} and the convergence rates proved in previous sections to derive convergence rates for the primal objective error and strong convergence of various quantities that appear in ADMM. In addition, we translate the results of the previous sections and use Proposition~\ref{prop:stronglymonotone} to state all theorems in terms of purely primal quantities. We recall the definition of the two auxiliary terms (Equation~\eqref{eq:snotation}): \begin{align}\label{eq:snotationdual} S_{d_f}(w_{d_f}^k, w^\ast) &= \max\left\{\frac{\beta_f\alpha_A}{2}\|w_{d_f}^k - w^\ast\|_{}^2, \frac{\mu_f}{2\|A\|^2}\|Ax^k - Ax^\ast\|_{}^2\right\}, \\ S_{d_g}(w_{d_g}^k, w^\ast) &= \max\left\{\frac{\beta_g\alpha_B}{2}\|w_{d_g}^k - w^\ast\|_{}^2, \frac{\mu_g}{2\|B\|^2}\|By^k - By^\ast\|_{}^2\right\}. \end{align} This form readily follows from Table~\ref{table:ADMMsubgradients}. The following is a direct translation of Theorem~\ref{prop:sumauxilliaryterms} to the current setting. Note that any of the Lipschitz, strong convexity, and strong monotonicity constants may be zero. \begin{theorem}[Primal differentiability and strong convexity]\label{prop:primaldifferentiability} Suppose that $(z^j)_{j \geq 0}$ is generated by Algorithm~\ref{alg:ADMM}. Then \begin{enumerate} \item \label{prop:sumauxilliaryterms:part:1} \textbf{Best iterate convergence:} If $(\lambda_j)_{j \geq 0}$ is bounded away from zero, then $\min_{i=0, \cdots, k}\left\{S_{d_f}(w_{d_f}^{i}, w^\ast)\right\} = o\left(1/(k+1)\right)$ and $\min_{i=0, \cdots, k}\{S_{d_g}(w_{d_g}^{i}, w^\ast)\}= o\left(1/(k+1)\right).$ \item \label{prop:sumauxilliaryterms:part:2} \textbf{Ergodic convergence:} Let $\overline{w}_{d_f}^k = (1/\Lambda_k) \sum_{i=0}^k w_{d_f}^i$, let $\overline{w}_{d_g}^k = (1/\Lambda_k) \sum_{i=0}^k\lambda_i w_{d_g}^i$, let $\overline{x}^k = (1/\Lambda_k) \sum_{i=0}^k x^i$, and let $\overline{y}^k = (1/\Lambda_k) \sum_{i=0}^k\lambda_i y^i$. Then \begin{align*} \max\left\{ \beta_f\alpha_A\left\|\overline{w}_{d_f}^k - w^\ast\right\|^2, \frac{\mu_f}{\|A\|^2}\left\|A\overline{x}^k - Ax^\ast\right\|^2\right\} &+ \max\left\{\beta_g\alpha_B\left\|\overline{w}_{d_g}^k - w^\ast\right\|^2 + \frac{\mu_g}{\|B\|^2}\left\|B\overline{y}^k- By^\ast\right\|^2\right\} \\ &\leq \frac{\|z^0 - z^\ast\|^2}{4 \gamma \Lambda_k}. \end{align*} \item \label{prop:sumauxilliaryterms:part:3} \textbf{General convergence:} If $\underline{\tau} = \inf_{j \geq 0} \lambda_j(1-\lambda_j) > 0$, then $S_{f}(w_{d_f}^{k}, w^\ast) + S_g(w_{d_g}^{k}, w^\ast) = o({1}/{\sqrt{k+1}})$. \end{enumerate} \end{theorem} The following proposition deduces $o(1/(k+1))$ objective error convergence of standard ADMM whenever $g$ is strongly convex, and $\gamma$ is small enough. \begin{theorem}[Strong convexity of $g$]\label{thm:strongconvexadmm} Suppose that $g$ is $\mu_g$-strongly convex. Let $\lambda_k \equiv 1/2$, and let $\gamma < \kappa \beta = \kappa \mu_g/\|B\|^2$ (see Theorem~\ref{thm:differentiableobjective}). Then for all $k \geq 1$, we have the constraint violations convergence rate: \begin{align*} \|Ax^k + By^k - b\|^2 \leq \frac{\beta^2\|w_{d_g}^0 - w^\ast\|^2}{\gamma^2k^2\left(1+\gamma/\beta\right)^2\left(\beta^2 -\gamma^2/\kappa^2\right)} && \mathrm{and} && \|Ax^k + By^k - b\|^2 = o\left(\frac{1}{k^2}\right). \end{align*} Moreover, the primal objective errors satisfy \begin{align*} \frac{-\beta\|w^\ast\|\|w_{d_g}^0 - w^\ast\|}{\gamma k\left(1+{\gamma}/{\beta}\right)\sqrt{\left(\beta^2 -{\gamma^2}/{\kappa^2}\right)}} \leq f(x^k) + g(y^k) - f(x^\ast) - g(y^\ast) &\leq\frac{\beta\|w_{d_g}^0 - w^\ast\| \left(\|z^0 - z^\ast\| + \|w^\ast\|\right)}{\gamma k \left(1+{\gamma}/{\beta}\right)\sqrt{\left(\beta^2 -{\gamma^2}/{\kappa^2}\right)}}, \end{align*} and $|f(x^k) + g(y^k) - f(x^\ast) - g(y^\ast)| = o\left(1/k\right).$ \end{theorem} \begin{proof} The constraint violations rate follows from the identity $z^{k+1} - z^k = -\gamma ( Ax^{k} + By^k - b)$ (Equation~\eqref{eq:ADMMfeasibilityFPR}) and the FPR convergence rate in Theorem~\ref{thm:differentiableFPR}. The lower bound follows from the lower fundamental inequality in Proposition~\ref{prop:ADMMlower} and the FPR convergence rate in Theorem~\ref{thm:differentiableFPR}: \begin{align*} f(x^k) + g(y^k) - f(x^\ast) - g(y^\ast) \stackrel{\eqref{prop:ADMMlower:eq:main}}{\geq} \dotp{Ax^k + By^k - b, w^\ast} &\stackrel{\eqref{eq:ADMMfeasibilityFPR}}{\geq} -\frac{1}{\gamma}\|z^k - z^{k+1}\|\| w^\ast\| \\ &\stackrel{\eqref{eq:differentiableFPR}}{\geq} - \frac{-\beta\|w^\ast\|\|w_{d_g}^0 - w^\ast\|}{\gamma k\left(1+{\gamma}/{\beta}\right)\sqrt{\left(\beta^2 -{\gamma^2}/{\kappa^2}\right)}}. \end{align*} Part~\ref{fact:averagedconvergence:eq:mono} of Fact~\ref{fact:averagedconvergence} bounds the norm: $\|z^{k+1} - (z^\ast - w^\ast)\| \leq \|z^{k+1} - z^\ast\| + \|w^\ast\| \leq \|z^0 - z^\ast\| + \|w^\ast\|$. Therefore, the upper bound follows from the upper fundamental inequality in Proposition~\ref{prop:ADMMupper} and the FPR convergence rate in Theorem~\ref{thm:differentiableFPR}: \begin{align*} f(x^k) + g(y^k) - f(x^\ast) - g(y^\ast) &\stackrel{\eqref{prop:ADMMupper:eq:main}}{\leq} \frac{1}{2\gamma}\left(\|z^k - (z^\ast - w^\ast)\|^2 - \|z^{k+1} - (z^\ast - w^\ast)\|^2 - \|z^{k} - z^{k+1}\|^2\right) \\\ &\stackrel{\eqref{eq:cosinerule}}{\leq} \frac{1}{\gamma}\dotp{ z^{k+1} - (z^\ast - w^\ast), z^k - z^{k+1}} \stackrel{\eqref{eq:differentiableFPR}}{\leq} \frac{\beta\|w_{d_g}^0 - w^\ast\| \left(\|z^0 - z^\ast\| + \|w^\ast\|\right)}{\gamma k \left(1+{\gamma}/{\beta}\right)\sqrt{\left(\beta^2 -{\gamma^2}/{\kappa^2}\right)}}. \end{align*} The little $o$-rate follows because, as the above equations have shown, the objective error is upper and lower bounded by a multiple of the square root of the FPR, which has convergence rate $o(1/k)$ by Theorem~\ref{thm:differentiableFPR}. \qed\end{proof} It would be nice to prove a convergence rate for the ``best iterate" of the sequence of primal objective errors in the style of Theorem~\ref{thm:lipschitzbest}. Unfortunately the fundamental inequalities we developed in Section~\ref{sec:convertinequalities} do not immediately imply such a rate. Now we shift our focus to linear convergence. The following proposition is a direct translation of the main results of Section~\ref{sec:linearconvergence} to the current setting. The interested reader is encouraged to read Appendix~\ref{app:linearconvergenceimplies} to see how the following rates imply convergence rates for the primal and dual objective, and feasibility errors. \begin{theorem}[Linear convergence of Relaxed ADMM]\label{thm:admmlinearconvergence} The following are true: \begin{enumerate} \item \label{thm:admmlinearconvergence:part:gregular} If $\mu_g \beta_g \alpha_B > 0$, then $(z^j)_{j \geq 0}$ converges linearly and \begin{align*} \|z^{k+1} - z^\ast\|^2 \leq \left(1 -\frac {4\gamma\lambda_k\beta_g\alpha_B}{(1 + {\gamma\|B\|^2}/{\mu_g})^2}\right)^{{1}/{2}}\|z^k - z^\ast\|^2. \end{align*} \item \label{thm:admmlinearconvergence:part:fregular} If $\mu_f\beta_f \alpha_A > 0$, then $(z^j)_{j \geq 0}$ converges linearly and \begin{align*} \|z^{k+1} - z^\ast\|^2 \leq \left(1 - \frac{\lambda_k\min\left\{ {4\gamma \beta_f \alpha_A}/{\left(1+{\|A\|^2}/{\mu_f}\right)^2}, (1 - \lambda_k)\right\}}{2}\right)^{{1}/{2}}\|z^k - z^\ast\|^2. \end{align*} \item \label{thm:admmlinearconvergence:part:fgregular} If $\mu_f \beta_g \alpha_B > 0$, then $(z^j)_{j \geq 0}$ converges linearly and \begin{align*} \|z^{k+1} - z^\ast\|^2 \leq \left(1 - \frac{4\lambda_k\min\{ \gamma \beta_g\alpha_B, { \mu_f}/({\|A\|^2\gamma}), (1 - \lambda_k)\}}{3}\right)^{{1}/{2}}\|z^k - z^\ast\|^2. \end{align*} \item\label{thm:admmlinearconvergence:part:gfregular} If $\mu_g \beta_f \alpha_A > 0$, then $(z^j)_{j \geq 0}$ converges linearly and \begin{align*} \|z^{k+1} - z^\ast\|^2 \leq \left(1 - \frac{4\lambda_k\min\{ \gamma \beta_f\alpha_A, { \mu_g}/({\|B\|^2\gamma}), (1 - \lambda_k)\}}{3}\right)^{{1}/{2}}\|z^k - z^\ast\|^2. \end{align*} \end{enumerate} \end{theorem} We can apply Proposition~\ref{prop:linearconvergenceimpliesADMM} to any of the scenarios that appear in Theorem~\ref{thm:admmlinearconvergence} and deduce the rate of linear convergence of the objective error and constraint violations. We leave this application to the reader. Linear convergence of ADMM has been deduced in a variety of scenarios. In \cite{denglinear2012}, the authors prove the linear convergence (in finite dimensions) of a generalized form of ADMM, which allows the possibility of adding proximal terms to the alternating minimization steps that appear in Algorithm~\ref{alg:ADMM}. The four scenarios that appear in \cite[Table 1.1]{denglinear2012}) have some overlap with our results. In the standard version of ADMM, (with no relaxation or extra proximal terms), scenarios 1 and 2 in~\cite[Table 1.1]{denglinear2012} are the finite-dimensional analogues of Part~\ref{thm:admmlinearconvergence:part:gregular} of Theorem~\ref{sec:linearconvergence}. Scenarios 3 and 4 in~\cite[Table 1.1]{denglinear2012} are not covered by our analysis because they require that we treat the structure of $A$ and $B$ more carefully than we have in this section. In addition, Parts~\ref{thm:admmlinearconvergence:part:fregular},~\ref{thm:admmlinearconvergence:part:fgregular}, and~\ref{thm:admmlinearconvergence:part:gfregular} of Theorem~\ref{thm:admmlinearconvergence} are not discussed in \cite{denglinear2012}. Finally, we note that this paper and \cite{denglinear2012} use the opposite update orders in ADMM. They generally lead to different sequences except when at least one of $f$ and $g$ is quadratic \cite{yanyin2014}. Therefore, when comparing the results between the two papers, one must switch $f$ and $g$, as well as $A$ and $B$. \subsection{Convergence rates of summable sequences} The following facts will be key to deducing Convergence rates in Sections~\ref{sec:generalstrongconvexity} and~\ref{sec:lipschitzderivatives}. It originally appeared in \cite[Lemma 3]{davis2014convergence}. \begin{fact}[Summable sequence convergence rates]\label{lem:sumsequence} Suppose that the nonnegative scalar sequences $(\lambda_j)_{j\geq0}$ and $(a_{j})_{j\geq 0}$ satisfy $\sum_{i=0}^\infty \lambda_ia_i < \infty$, and define $\Lambda_k$ as in Equation~\eqref{def:Lambda}. \begin{enumerate} \item \label{lem:sumsequence:part:main} \textbf{Monotonicity:} If $(a_j)_{j \geq0}$ is \emph{monotonically nonincreasing}, then \begin{align}\label{eq:sumsequence-bigo} a_{k} \leq \frac{1}{\Lambda_k}\left(\sum_{i=0}^\infty\lambda_i a_i \right) && and && a_{k} = o\left(\frac{1}{\Lambda_{k} - \Lambda_{\ceil{{k}/{2}}}}\right). \end{align} \item \label{lem:sumsequence:part:b} \textbf{Faster rates:} Suppose $(b_j)_{j\geq0}$ is a nonnegative scalar sequence, that $\sum_{i=0}^\infty b_j < \infty$, and that $\lambda_k a_k \leq b_k - b_{k+1}$ for all $k\geq 0$. Then the following sum is finite: \begin{align} \sum_{i=0}^\infty (i+1)\lambda_ia_i \leq \sum_{i=0}^\infty b_i \end{align} \item \label{lem:sumsequence:part:nonmono} \textbf{No monotonicity:} For all $k \geq 0$, define the sequence of ``best indices'' with respect to $(a_j)_{j\ge 0}$ as \begin{align*} k_{\mathrm{best}} &:= \argmin_i\{a_i | i = 0, \cdots, k\}. \end{align*} Then $(a_{j_{\mathrm{best}}})_{j \geq 0}$ is nonincreasing, and the above bounds continue to hold when $a_k$ is replaced with $a_{k_{\best}}$. \end{enumerate} \end{fact} \subsection{Convergence of the fixed-point residual (FPR)}\label{sec:FPR} We will need to following facts in our analysis below: \begin{fact}[Convergence rates of FPR]\label{fact:averagedconvergence} Let $z^\ast \in {\mathcal{H}}$ be a fixed point of $T_{\mathrm{PRS}}$, and let $(z^j)_{j \geq 0}$ be generated by the relaxed PRS algorithm: $z^{k+1} = (T_{\mathrm{PRS}})_{\lambda_k}z^k. $ Then the following are true (\cite[Theorem 1]{davis2014convergence}): \begin{enumerate} \item $(\|z^j - z^\ast\|)_{j \geq 0}$ is monotonically nonincreasing; \label{fact:averagedconvergence:eq:mono} \item $(\|T_{\mathrm{PRS}} z^j - z^j\|)_{j \geq 0}$ is monotonically nonincreasing, and thus so is $((1/\lambda_j)\|z^{j+1}-z^j\|)_{j \geq 0}$; \label{fact:averagedconvergence:eq:FPRmono} \item If $\lambda_k \equiv \lambda$, then $(\|(T_{\mathrm{PRS}})_\lambda z^j - z^\ast\|)_{j\geq0}$ is monotonically nonincreasing; \item The Fej\'er-type inequality holds: for all $\lambda \in (0, 1]$ \begin{align*} \|(T_{\mathrm{PRS}})_{\lambda}z^k - z^\ast\|^2 &\leq \|z^k - z^\ast\|^2 - \frac{1- \lambda}{\lambda}\|(T_{\mathrm{PRS}})_{\lambda} z^k - z^k\|^2.\numberthis\label{eq:fejer} \end{align*} \item For all $k \geq 0$, let $\tau_k = \lambda_k(1-\lambda_k)$. Then $\sum_{i=0}^\infty \tau_i\|T_{\mathrm{PRS}} z^i - z^i\|^2 \leq \|z^0 - z^\ast\|^2.$ \item If $\underline{\tau} := \inf_{j \geq 0} \lambda_k(1-\lambda_k) > 0$, then the following convergence rates hold: \begin{align}\label{cor:DRSaveragedconvergence:eq:main} \|T_{\mathrm{PRS}} z^{k} - z^k\|^2 \leq \frac{\|z^{0} - z^\ast\|^2}{\underline{\tau}(k+1)} && and && \|T_{\mathrm{PRS}} z^{k} - z^k\|^2 = o\left(\frac{1}{\underline{\tau}(k+1)}\right). \end{align} \end{enumerate} \end{fact} \begin{remark} We call the quantity $\|T_{\mathrm{PRS}} z^{k} - z^k\|^2$ the fixed-point residual (FPR) of the relaxed PRS algorithm. Throughout this paper, we slightly abuse terminology and call the successive iterate difference $\|z^{k+1} - z^k\|^2 = \lambda_k^2\|T_{\mathrm{PRS}} z^k - z^k\|^2$ FPR as well. \end{remark} \subsection{Subgradients} Lemma~\ref{prop:DRSmainidentity} is key to deducing all of the algebraic relations necessary for relating the objective error to the FPR of the relaxed PRS iteration \begin{lemma}\label{prop:DRSmainidentity} Let $z\in {\mathcal{H}}$. Define auxiliary points $x_g := \mathbf{prox}_{\gamma g}(z)$ and $x_f := \mathbf{prox}_{\gamma f}(\mathbf{refl}_{\gamma g}(z))$. Then the identities hold: \begin{align} x_g = z - \gamma \widetilde{\nabla} g(x_g) && \mathrm{and} && x_f &= x_g - \gamma \widetilde{\nabla} g(x_g) - \gamma \widetilde{\nabla} f(x_f). \label{prop:DRSmainidentity:f} \end{align} In addition, each relaxed PRS step $z^+=(T_{\mathrm{PRS}} )_{\lambda}(z)$ has the following representation: \begin{align}\label{eq:DRSmainidentity2} z^+ - z = 2\lambda(x_f- x_g) = -2 \lambda\gamma(\widetilde{\nabla} g(x_g)+\widetilde{\nabla} f(x_f)). \end{align} \end{lemma} \subsection{Fundamental inequalities} Throughout the rest of the paper we will use the following notation: Every function $f$ is $\mu_f$-strongly convex and $\widetilde{\nabla} f$ is $(1/\beta_f)$-Lipschitz. Note that if $\beta_f > 0$, then $f$ is differentiable and $\widetilde{\nabla} f = \nabla f$. However, we also allow the strong convexity or Lipschitz differentiability constants to vanish, in which case $\mu_f = 0$ or $\beta_f = 0$ and $f$ may fail to posses either regularity property. Thus, we always have the inequality \cite[Theorem 18.15]{bauschke2011convex}: \begin{align}\label{eq:strongconvexandlipschitzlowerbound} f(x) &\geq f(y) + \dotp{x-y, \widetilde{\nabla} f(y)} + S_f(x, y), \end{align} where \begin{align}\label{eq:snotation} S_f(x, y) &:= \max\left\{\frac{\mu_f}{2}\|x - y\|^2, \frac{\beta_f}{2}\|\widetilde{\nabla} f(x) - \widetilde{\nabla} f(y)\|^2\right\}. \end{align} Note that there is a slight technicality in that $S_f(x, y)$ is only defined where $\partial f(x) \neq \emptyset$. In particular, we only derive bounds on $S_f(x, y)$ where this is satisfied. The following two fundamental inequalities are straightforward modifications of the fundamental inequalities that appeared in \cite[Propositions 4 and 5]{davis2014convergence}. When these bounds are iteratively applied, they bound the objective error by the sum of a telescoping sequence and a multiple of the FPR. \begin{proposition}[Upper fundamental inequality]\label{prop:DRSupper} Let $z \in {\mathcal{H}}$, let $z^+ = (T_{\mathrm{PRS}})_{\lambda}(z)$, and let $x_f$ and $x_g$ be defined as in Lemma~\ref{prop:DRSmainidentity}. Then for all $x\in \dom(f) \cap \dom(g)$ where $\partial f(x) \neq \emptyset$ and $\partial g(x) \neq \emptyset$, we have \begin{align*} 4\gamma\lambda\big(f(x_f) + g(x_g) &- f(x) - g(x) + S_f(x_f, x) + S_g(x_g, x)\big) \\ &\leq \|z - x\|^2 - \|z^+ - x\|^2 + \left(1 - \frac{1}{\lambda}\right)\|z^{+} - z\|^2. \numberthis \label{prop:DRSupper:eq:main} \end{align*} \end{proposition} In our analysis below, we will use the upper inequality \begin{align*} 4\gamma\lambda\big(f(x_f) + g(x_g) &- f(x^*) - g(x^*) + S_f(x_f, x^*) + S_g(x_g, x^*)\big) \\ &\leq \|z - z^*\|^2 - \|z^+ - z^*\|^2 +2\dotp{z-z^+,z^*-x^*}+ \left(1 - \frac{1}{\lambda}\right)\|z^{+} - z\|^2, \numberthis \label{prop:DRSupper:eq:aux}\end{align*} which is obtained from \eqref{prop:DRSupper:eq:main} by letting $x=x^*$ and applying $\|z - x^*\|^2 - \|z^+ - x^*\|^2=\|z-z^*\|^2-\|z^+-z^*\|^2+2\dotp{z-z^+,z^*-x^*}. $ \begin{proposition}[Lower fundamental inequality]\label{prop:DRSlower} Let $z^\ast$ be a fixed point of $T_{\mathrm{PRS}} $, and let $x^\ast = \mathbf{prox}_{\gamma g}(z^\ast)$. Then for all $x_f \in \dom(f)$ and $x_g\in \dom(g)$, the lower bound holds: \begin{align*} f(x_f) + g(x_g) - f(x^\ast) - g(x^\ast) &\geq \frac{1}{\gamma }\dotp{x_g - x_f, z^\ast - x^\ast} + S_f(x_f, x^\ast) + S_g(x_f, x^\ast). \numberthis \label{prop:DRSlower:eq:main} \end{align*} \end{proposition} \section{Examples} In this section, we apply DRS and ADMM to concrete problems and explicitly bound the associated objective errors and FPR with the convergence rates that we derived in the previous sections. \subsection{Feasibility problems}\label{sec:feasibility} Suppose that $C_f$ and $C_g$ are closed convex subsets of ${\mathcal{H}}$ with nonempty intersection. The goal of the feasibility problem is the find a point in the intersection of $C_f$ and $C_g$. In this section, we present a comparison between MAP and the relaxed PRS algorithm. \subsubsection{Linear convergence} Section~\ref{sec:feasibilityniceintersection} shows that relaxed PRS applied to $f = d_{C_f}^2$ and $g = d_{C_g}^2$ converges linearly whenever $C_f$ and $C_g$ have a sufficiently nice intersection. In addition, \cite{bauschke2014linear} and \cite{phan2014linear} have recently shown that one can achieve linear convergence under the same regularity assumptions on $C_f \cap C_g$ when $f= \iota_{C_f}$ and $g = \iota_{C_f}$. We refer to \cite[Fact 5.8]{bauschke2014linear} for an extensive list of conditions that guarantee (bounded) linear regularity of $\{C_1, C_2\}$. For the readers convenience, we list a few important examples: \begin{enumerate} \item {\bf Subspaces:} If $C_f^\perp + C_g^\perp$ is closed, then $\{C_f, C_g\}$ is linearly regular. \item {\bf Polyhedron:} If $C_f \cap C_g \neq \emptyset$, then $\{C_f, C_g\}$ is linearly regular. \item {\bf Standard constraint qualification:} If the relative interiors of $C_f$ and $C_g$ intersect, then $\{C_f,C_g\}$ is boundedly linearly regular. \end{enumerate} \subsubsection{General convergence}\label{sec:feasibilitygeneralconvergence} In general, we cannot expect linear convergence of relaxed PRS algorithm for the feasibility problem. Indeed, \cite[Theorem 9]{davis2014convergence} constructs a DRS iteration that converges in norm but does so \emph{arbitrarily slowly}. A similar result holds for MAP~\cite{bauschke2009characterizing}. Thus, in \cite{davis2014convergence} the authors focused on other measures of convergence, namely \emph{FPR} and \emph{objective error} rate. The following discussion will utilize the results of \cite{davis2014convergence} to compare the relaxed PRS and MAP algorithms in the absence of regularity. Let $\iota_{C_f}$ and $\iota_{C_g}$ be the indicator functions of $C_f$ and $C_g$. Then $x \in C_f \cap C_g$, if, and only if, $\iota_{C_f}(x) + \iota_{C_g}(x) = 0$, and the sum is infinite otherwise. Thus, a point is in the intersection of $C_f$ and $C_g$ if, and only if, it is the minimizer of the following problem: \begin{align}\label{sec:feasibility:eq:chiminimize} \Min_{x \in {\mathcal{H}}} \iota_{C_f}(x) + \iota_{C_g}(x). \end{align} The relaxed PRS algorithm applied to $f = \iota_{C_f}$ and $g = \iota_{C_g}$ has the following form: given an initial point $z^0 \in {\mathcal{H}}$, for all $k \geq 0$, define \begin{align}\label{sec:feasibility:eq:DRSchi} \begin{cases} x_g^k &= P_{C_g}(z^k); \\ x_f^k &= P_{C_f}(2x_g^k - z^k); \\ z^{k+1} &= z^k + 2\lambda_k(x_f^k - x_g^k). \end{cases} \end{align} In general, the functions $f$ and $g$ are neither differentiable nor strongly convex. Furthermore, they only take on the values $0$ and $\infty$. Thus, we will only discuss FPR convergence rates of relaxed PRS. The FPR identity $x_{f}^k - x_g^k = \frac{1}{2\lambda_k}(z^{k+1} - z^k)$ shows that after $k$ iterations \begin{align}\label{eq:feasibilitybounddistancenonergodic} \max\{ d^2_{C_g}(x_f^k), d^2_{C_f}(x_g^k)\} \leq \|x_f^k - x_g^k\|^2 &\stackrel{\eqref{cor:DRSaveragedconvergence:eq:main}}{=} o\left(\frac{1}{k+1}\right). \end{align} By the convexity of $C_f$ and $C_g$, the ergodic iterates of relaxed PRS satisfy $\overline{x}_f^k = ({1}/{\Lambda_k})\sum_{i=0}^k \lambda_i x_f^i \in C_f$ and $\overline{x}_g^k = ({1}/{\Lambda_k})\sum_{i=0}^k \lambda_i x_g^i \in C_g$. Thus, \cite[Theorem 6]{davis2014convergence} implies the improved bound \begin{align}\label{eq:drsergodicdistancebound} \max\{ d^2_{C_g}(\overline{x}_f^k), d^2_{C_f}(\overline{x}_g^k)\} \leq \|\overline{x}_f^k - \overline{x}_g^k\|^2 &= O\left(\frac{1}{\Lambda_k^2}\right), \end{align} which is optimal by \cite[Proposition 7]{davis2014convergence}. Therefore, after $k$ iterations the relaxed PRS algorithm produces a point in each set with distance of order at most $O({1}/{\Lambda_k})$ from each other. We now shift our focus to the MAP algorithm. First we replace both of the indicator functions with the squared distance functions: $f= \min_{y \in C_f} \|x - y\|^2$ and $g(x) = \min_{y \in C_g} \|x-y\|^2.$ Now recall that $f$ and $g$ are differentiable, the gradient $\nabla g$ is $2$-Lipschitz continuous \cite[Corollary 12.30]{bauschke2011convex}, and relaxed PRS takes the form in Equation~\eqref{eq:DRSfeasibilitygamma}. Specializing to $\gamma = {1}/{2}$ and $\lambda_k \equiv 1$ yields the MAP algorithm (Corollary~\ref{cor:AP}). In this algorithm, the main MAP sequence satisfies $(z^j)_{j \geq 1} \subseteq C_f$, while the auxiliary sequences $(x_f^j)_{j \geq 0}$ and $(x_g^j)_{j \geq 0}$ are not necessarily elements $C_f$ or $C_g$. Therefore, the MAP FPR rate is less useful for estimating distances of the current iterates to $C_f$ and $C_g$ than it is in the relaxed PRS algorithm (See Equation~\eqref{eq:feasibilitybounddistancenonergodic}). Although $\lambda_k \equiv 1$, the map $P_{C_f} P_{C_g}$ is $\alpha$-averaged for some $\alpha < 1$, and, hence, we can still estimate $\|z^{k+1} - z^k\|^2 = o({1}/{(k+1)})$ (Corollary~\ref{cor:AP2}). The ergodic convergence rate in \cite[Theorem 6]{davis2014convergence} (where we use the identity $d_{C_g}(x_g^k) = ({1}/{2})d_{C_g}(z^k)$ and Jensen's inequality) shows that \begin{align} d^2_{C_g}\left(\frac{1}{k+1}\sum_{i=0}^k z^i\right) \leq \frac{2}{k+1}\sum_{i=0}^k d_{C_g}^2(x_g^k) = O\left(\frac{1}{k+1}\right). \end{align} Thus, if we choose $z^0 \in C_f$, the ergodic iterate $({1}/{(k+1)})\sum_{i=0}^k z^i $ is an element of $C_f$ and we can bound its distance from $C_g$. Note that this rate is strictly slower than the rate in Equation~\eqref{eq:drsergodicdistancebound}. Although $d_{C_f}^2$ and $d_{C_g}^2$ are differentiable (Proposition~\ref{prop:factsaboutdistancesquared}), we cannot apply the results of Section~\ref{sec:lipschitzderivatives} to MAP because they require that $(\lambda_j)_{j \geq 0} \subseteq (0, 1)$. Therefore, we cannot use the regularity of $d_{C_f}^2$ and $d_{C_g}^2$ to deduce faster convergence of the AP algorithm. This discussion shows that the convergence rates predicted in \cite{davis2014convergence} for relaxed PRS, which are known to be optimal, are faster than those predicted for MAP. When $C_f$ and $C_g$ intersect nicely (Section~\ref{sec:feasibilityniceintersection}), the rate predicted for MAP is faster (See Corollary~\ref{cor:AP}). In \cite[Section 8]{bauschke2013rate} a similar phenomenon is observed for the case of intersecting subspaces: DRS is faster than MAP for problems with nonregular intersection. It would be highly satisfying to characterize this phenomenon in general. \subsection{Parallelized model fitting and classification}\label{sec:modelfiting} The following scenario appears in \cite[Chapter 8]{boyd2011distributed}. Consider the model fitting problem: Let $M : {\mathbf{R}}^n \rightarrow {\mathbf{R}}^m$ be a \emph{feature matrix}, let $b \in {\mathbf{R}}^m$, be the \emph{output} vector, let $l$ be a \emph{loss function} and let $r$ be a \emph{regularization function}. The goal of the \emph{model fitting problem} is to \begin{align}\label{sec:modelfiting:eq:problem} \Min_{x\in {\mathbf{R}}^n } \; l(Mx -b) + r(x). \end{align} The function $l$ is used to enforce the constraint $Mx = b + \nu$ up to some noise $\nu$ in the measurement, while $r$ enforces the \emph{regularity} of $x$ by incorporating \emph{prior knowledge} of the form of the solution. In this section, we present one way to split Equation~\eqref{sec:modelfiting:eq:problem}. Our discussion extends the one given in~\cite[Section 9.2]{davis2014convergence}, where only convexity of $l$ and $r$ is assumed. \subsubsection{Auxiliary variable} We can split Equation~\eqref{sec:modelfiting:eq:problem} by defining an auxiliary variable for $Mx - b$: \begin{align*} \Min_{x \in {\mathbf{R}}^m, y \in{\mathbf{R}}^n} & \; l\left(y \right) + r(x) \\ \mbox{subject to } & \; Mx - y = b.\numberthis \label{sec:modelfiting:eq:pullAout} \end{align*} We will now analyze the convergence rates predicted in Section~\ref{sec:dualtoprimalrates} for ADMM applied to Problem~\eqref{sec:modelfiting:eq:pullAout}. Our most general convergence result applies to the auxiliary terms: \begin{align*} S_{d_r}(w_{d_r}^k, w^\ast) &= \max\left\{\frac{\beta_r\alpha_M}{2}\|w_{d_r}^k - w^\ast\|_{{\mathbf{R}}^m}^2, \frac{\mu_r}{2\|M\|^2}\|Mx^k - Mx^\ast\|_{{\mathbf{R}}^m}^2\right\}, \\ S_{d_l}(w_{d_l}^k, w^\ast) &= \max\left\{\frac{\beta_l}{2}\|w_{d_l}^k - w^\ast\|_{{\mathbf{R}}^m}^2, \frac{\mu_l}{2}\|y^k - y^\ast\|_{{\mathbf{R}}^n}^2\right\}. \end{align*} Theorem~\ref{prop:primaldifferentiability} shows that the best auxiliary term converges with rate $o(1/(k+1))$, the ergodic auxiliary term converges with rate $O(1/\Lambda_k)$, and the entire sequence of auxiliary terms converges with rate $o(1/\sqrt{k+1})$. Now suppose that $\mu_l > 0$. Then we can bound the distance of $y^k$ to the optimal point $y^\ast: = Mx^\ast - b$: $$\|\overline{y}^k - y^\ast\|^2 = O\left(\frac{1}{\Lambda_k^2}\right).$$ Now let $f = r$, let $g = l$, let $A = M$, and let $B = -I_{{\mathcal{R}}^m}$. If $\gamma <\kappa \mu_l$, then Theorem~\ref{thm:strongconvexadmm} bounds the primal objective error and the FPR: \begin{align*} |l(y^k) + r(x^k) - l(Mx^\ast - b) - r(x^\ast)| = o\left(\frac{1}{{k+1}}\right) && \mathrm{and} && \|Mx^k - b- y^k\|^2 = o\left(\frac{1}{({k+1})^2}\right). \end{align*} In particular, if $l$ is Lipschitz, then $|l(y^k) - l(Mx^k - b)| = o\left(1/({k+1})\right)$. Thus, we have \begin{align*} 0 \leq l(Mx^k - b) + r(x^k) - l(Mx^\ast - b) - r(x^\ast) = o\left(\frac{1}{{k+1}}\right). \end{align*} A similar result holds if $r$ is strongly convex and we assign $g= r$ and $f = l$, etc. We can improve the above sublinear rate to a linear rate in any of the following cases (Theorem~\ref{thm:admmlinearconvergence}): \begin{itemize} \item $r$ is differentiable and strongly convex and $MM^\ast$ is strongly monotone; \item $l$ is differentiable and strongly convex; \item $r$ is differentiable, $MM^\ast$ is strongly monotone, and $l$ is strongly convex; \item $r$ is strongly convex and $l$ is differentiable. \end{itemize} \section{Feasibility Problems with regularity}\label{sec:feasibilityniceintersection} In this section we consider the feasibility problem: $$\mbox{Given two closed convex subsets $C_f$ and $C_g$ of ${\mathcal{H}}$ such that $C_f \cap C_g \neq \emptyset$, find a point $x \in C_f \cap C_g$.}$$ Throughout this section we assume that $\{C_f, C_g\}$ is \emph{boundedly linearly regular}: \begin{definition}[Bounded linear regularity]\label{defi:linearregularity} Suppose that $C_1, \cdots, C_m$ are closed convex subsets of ${\mathcal{H}}$ with nonempty intersection. We say that $\{C_1, \cdots, C_m\}$ is boundedly linearly regular if the following holds: for all $\rho > 0$, there exists $\mu_\rho > 0$ such that for all $x \in B(0, \rho)$, (the open ball centered at the origin with radius $\rho$), we have \begin{align*} d_{C_1\cap \cdots \cap C_m}(x) &\leq \mu_\rho\max\{d_{C_1}(x), \cdots, d_{C_m}(x)\} \end{align*} where for any subset $C \subseteq {\mathcal{H}}$, the distance function $d_C(x):= \inf_{y \in C} \|x -y\|.$ Evidently, if $B(0, \rho) \backslash (C_1 \cap \cdots \cap C_m) \neq \emptyset$, then $\mu_\rho \geq 1$. We say that $\{C_1,\cdots, C_m\}$ is linearly regular if it is boundedly linearly regular and $\mu_\rho$ does not depend on $\rho$, i.e. $\mu_\rho = \mu_\infty < \infty$. \qed\end{definition} Intuitively, (bounded) linear regularity is the following implication: $$\mbox{(close to all of the sets) $\implies$ (close to the intersection).}$$ This property will be key to deducing linear convergence of an application of the relaxed PRS algorithm. See \cite{davis2014convergence} for the feasibility problem when no regularity is assumed. There are several ways to model the feasibility problem, e.g. with $f$ and $g$ given by indicator functions, distance functions, or squared distance functions. In this section, we will model the feasibility problem using squared distance functions: \begin{align*} f(x) := d_{C_f}^2(x) && \mathrm{and} && g(x) := d_{C_g}^2(x). \end{align*} We briefly summarize some properties of squared distance functions. \begin{proposition}[Properties of distance functions]\label{prop:factsaboutdistancesquared} Let $C$ be a nonempty closed convex subset of ${\mathcal{H}}$. Then the following properties hold: \begin{enumerate} \item The function $d_{C}$ is $1$-Lipschitz. \item The function $d_{C}^2$ is differentiable, and $\nabla d_C^2 = 2(I_{{\mathcal{H}}} - P_{C})$. In addition, $\nabla d_C^2$ is $2$-Lipschitz. \item The proximal identity holds: for all $\gamma > 0$, \begin{align*} \mathbf{prox}_{ \gamma d^2_C} &= \frac{1}{2\gamma + 1}I_{{\mathcal{H}}} + \frac{2\gamma}{2\gamma + 1} P_{C}. \end{align*} \end{enumerate} \end{proposition} \begin{proof} For a proof see \cite[Corollary 12.30]{bauschke2011convex}. \qed\end{proof} Given $z^0 \in {\mathcal{H}}$, sequences of implicit stepsize parameters, $(\gamma_{f, j})_{j \geq 0}$, $(\gamma_{g, j})_{j \geq 0}$, and relaxation parameters, $(\lambda_j)_{j \geq 0}$, we consider the iteration: for all $k \geq 0$, let \begin{align}\label{eq:DRSfeasibilitygamma} \begin{cases} x_g^k &= \mathbf{prox}_{\gamma_{g, k} d_{C_g}^2} (z^k); \\ x_f^k &= \mathbf{prox}_{\gamma_{f,k} d_{C_f}^2}(2x_g^k - z^k); \\ z^{k+1} &= z^k + 2\lambda_k(x_f^k - x_g^k). \end{cases} \end{align} If $(\gamma_{f,j})_{j \geq 0}, (\gamma_{g, j})_{j \geq 0} \subseteq (0, {1}/{2}]$ and $\lambda_k \equiv 1$, then the iteration in Equation~\eqref{eq:DRSfeasibilitygamma} is the \emph{underrelaxed MAP} (see \cite{bauschke1996projection} for the parallel product space version and see \cite{bauschke2013method} for the nonconvex case). In particular, Corollary~\ref{cor:AP} (below) shows that when all implicit stepsize parameters are equal to ${1}/{2}$ and all relaxation parameters are $1$, Equation~\eqref{eq:DRSfeasibilitygamma} reduces to the MAP algorithm, where $ P_{C_g}z^k = 2x_g^k - z^k$, and $z^{k+1} = P_{C_f}P_{C_g} z^k$. This was already noticed in \cite[Proposition 2.5]{luke2008finding} for the fixed $\gamma$ case. We now specialize the fundamental inequality in Proposition~\ref{prop:DRSupper} to the feasibility problem. See Appendix~\ref{app:feasibilityniceintersection} for a proof. \begin{proposition}[Upper fundamental inequality for feasibility problem]\label{eq:DRSupperfeas} Suppose that $z\in {\mathcal{H}}$ and $z^+ = (T_{\mathrm{PRS}}^{\gamma_f, \gamma_g})_{\lambda}(z)$. Then for all $x^\ast\in C_f\cap C_g$, \begin{align}\label{eq:DRSupperfeas:eq:main} 8\lambda(\gamma_{f} d^2_{C_f}(x_f) + \gamma_{g}d^2_{C_g}(x_g)) &\leq \|z - x^\ast\|^2 - \|z^{+} - x^\ast\|^2 + \left(1 - \frac{1}{\lambda} \right)\|z^{+} - z\|^2. \end{align} \end{proposition} We are now ready to prove the linear convergence of Algorithm~\eqref{eq:DRSfeasibilitygamma} whenever $\{C_f, C_g\}$ is (boundedly) linearly regular. The proof is a consequence of the upper inequality in Proposition~\ref{eq:DRSupperfeas}. \begin{theorem}[Linear convergence: Feasibility for two sets]\label{thm:linearfeasibility} Suppose that $(z^j)_{j \geq 0}$ is generated by the iteration in Equation~\eqref{eq:DRSfeasibilitygamma}, and that $C_f$ and $C_g$ are (boundedly) linearly regular. Let $\rho > 0$ and $\mu_\rho > 0$ be such that $(z^j)_{j \geq 0} \subseteq B(0, \rho)$ and the inequality \begin{align*} d_{C_f \cap C_g}(x) &\leq \mu_\rho\max\{ d_{C_f}(x), d_{C_g}(x)\} \end{align*} holds for all $x \in B(0, \rho)$. Then $(z^j)_{j \geq 0}$ satisfies the following relation: for all $k \geq 0$, \begin{align}\label{thm:linearfeasibility:eq:decrease} d_{C_f \cap C_g}(z^{k+1}) &\leq C(\gamma_{f, k}, \gamma_{g, k}, \lambda_k, \mu_\rho) \times d_{C_f \cap C_g}(z^k) \end{align} where \begin{align*} C(\gamma_{f, k}, \gamma_{g, k}, \lambda_k, \mu_\rho) &:= \left(1-\frac{4\lambda_k\min\{{\gamma_{g, k}}/{(2\gamma_{g, k} + 1)^2}, {\gamma_{f, k}}/{(2\gamma_{f, k} + 1)^2}\}}{\mu_\rho^2\max\{{16\gamma_{g, k}^2}/{(2\gamma_{g, k}+1)^2}, 1\}} \right)^{{1}/{2}}. \end{align*} In particular, if $\overline{C} = \sup_{j \geq 0}C(\gamma_{f, j}, \gamma_{g, j}, \lambda_j, \mu) < 1$, then $(z^j)_{j \geq 0}$ converges linearly to a point in $x \in C_f \cap C_g$ with rate $\overline{C}$, and \begin{align} \|z^k - x\| &\leq 2d_{C_f\cap C_g}(z^0) \prod_{i=0}^k C(\gamma_{f, i}, \gamma_{g, i}, \lambda_i, \mu). \end{align} \end{theorem} \begin{proof} For simplicity, throughout the proof we will drop the iteration index $k$ and denote $z^+ := z^{k+1}$ and $z := z^k$, etc. Now recall the identities: \begin{align} x_g = \frac{1}{2\gamma_{g} + 1} z + \frac{2\gamma_{g}}{2\gamma_{g}+1}P_{C_g}(z) &&\mathrm{and} && x_f = \frac{1}{2\gamma_{f} + 1} \mathbf{refl}_{\gamma_{g} g}(z) + \frac{2\gamma_{f}}{2\gamma_{f} + 1} P_{C_f} (\mathbf{refl}_{\gamma_{g} g}(z)). \end{align} Thus, $x_g$ is a point on the line segment connecting $P_{C_g}(z)$ and $z$, and $x_f$ is a point on the line segment connecting $\mathbf{refl}_{\gamma_g g}(z)$ and $P_{C_f}(\mathbf{refl}_{\gamma_g g}(z))$. Hence, we have the projection identities: $P_{C_g}z = P_{C_g}x_g$ and $P_{C_f}(\mathbf{refl}_{\gamma_{g} g}(z)) = P_{C_f} x_f$. We can also compute the distances to $C_f$ and $C_g$: \begin{align}\label{thm:linearfeasibility:eq:switchtoz} d_{C_g}^2(x_g) = \frac{1}{(2\gamma_{g}+1)^2}d_{C_g}^2(z) && \mathrm{and} && d_{C_f}^2(x_f) = \frac{1}{(2\gamma_{f} + 1)^2}d^2_{C_f}(\mathbf{refl}_{\gamma_{g} g}(z)). \end{align} We will now bound $d^2_{C_f}(z)$. Because $x_g$ is a point on the line segment connecting $z$ and $P_{C_g}(z)$, Equation~\eqref{thm:linearfeasibility:eq:switchtoz} shows that that $\|z - x_g\| = (2\gamma_g/(2\gamma_g + 1) ) d_{C_g}(z)$. Thus, if $c_1 := c_1(\gamma_{g}) = {4\gamma_{g}}/({2\gamma_{g}+1})$, we have \begin{align}\label{thm:linearfeasibility:eq:reflbound} \|z - \mathbf{refl}_{\gamma_{g} g}(z)\| = 2\|z - x_g\| = c_1d_{C_g}(z). \end{align} Therefore, because $d_{C_f}$ is $1$-Lipschitz and by the convexity of $(\cdot)^2$, \begin{align*} d^2_{C_f}(z) &\leq (\|z - \mathbf{refl}_{\gamma_{g} g}(z)\| + d_{C_f}(\mathbf{refl}_{\gamma_{g} g}(z)))^2 \\ &= (c_1d_{C_g}(z) + d_{C_f}(\mathbf{refl}_{\gamma_{g} g}(z)))^2 \\ &\leq 2\max\{c_1^2, 1\}(d^2_{C_g}(z) + d^2_{C_f}(\mathbf{refl}_{\gamma_{g} g}(z))). \numberthis \label{thm:linearfeasibility:eq:upperboundztoc1} \end{align*} Now we will simplify the upper bound in Equation~\eqref{eq:DRSupperfeas:eq:main} by using Equation~\eqref{thm:linearfeasibility:eq:switchtoz} \begin{align*} 8\lambda\left(\frac{\gamma_{g}}{(2\gamma_{g} + 1)^2}d^2_{C_f}(\mathbf{refl}_{\gamma_{g} g}(z)) + \frac{\gamma_{f}}{(2\gamma_{f} + 1)^2}d^2_{C_g}(z)\right) &+ \|z^{+} - x\|^2 + \left(\frac{1}{\lambda} - 1\right)\|z^{+} - z\|^2 \leq \|z - x\|^2. \numberthis \label{thm:linearfeasibility:eq:upper2} \end{align*} Because ${1}/({2\max\{c_1^2, 1\}}) < 1$, we have \begin{align*} &8\lambda\left(\frac{\gamma_{g}}{(2\gamma_{g} + 1)^2}d^2_{C_f}(\mathbf{refl}_{\gamma_{g} g}(z)) + \frac{\gamma_{f}}{(2\gamma_{f} + 1)^2}d^2_{C_g}(z)\right)\\ &\geq 8\lambda\min\left\{\frac{\gamma_{g}}{(2\gamma_{g} + 1)^2}, \frac{\gamma_{f}}{(2\gamma_{f} + 1)^2}\right\}\left(d^2_{C_f}(\mathbf{refl}_{\gamma_{g} g}(z)) + d^2_{C_g}(z)\right) \\ &\stackrel{\eqref{thm:linearfeasibility:eq:upperboundztoc1}}{\geq}\frac{8\lambda\min\{{\gamma_{g}}/{(2\gamma_{g} + 1)^2}, {\gamma_{f}}/{(2\gamma_{f} + 1)^2}\}}{2\max\{c_1^2, 1\}} \max\{d_{C_f}^2(z), d^2_{C_g}(z)\}. \numberthis \label{thm:linearfeasibility:eq:lower} \end{align*} Now, recall the bounded linear regularity property: for all $x \in B(0, \rho)$, \begin{align*} d_{C_f \cap C_g}(x) &\leq \mu_\rho\max\{ d_{C_f}(x), d_{C_g}(x)\}. \end{align*} Thus, for all $x \in C_f \cap C_g$, the lower bound in Equation~\eqref{thm:linearfeasibility:eq:lower} shows that (where we use $(1/\lambda - 1) \geq 0$ in Equation~\eqref{thm:linearfeasibility:eq:upper2}) \begin{align*} \frac{4\lambda\min\{{\gamma_{g}}/{(2\gamma_{g} + 1)^2}, {\gamma_{f}}/{(2\gamma_{f} + 1)^2}\}}{\mu_\rho^2\max\{c_1^2, 1\}} d^2_{C_f\cap C_g}(z) + \|z^{+} - x\|^2&\stackrel{\eqref{thm:linearfeasibility:eq:upper2}}{\leq} \|z - x\|^2. \end{align*} Hence, if we define $$C(\gamma_{f}, \gamma_{g}, \lambda, \mu_\rho) = \left(1-\frac{4\lambda\min\{{\gamma_{g}}/{(2\gamma_{g} + 1)^2}, {\gamma_{f}}/{(2\gamma_{f} + 1)^2}\}}{\mu_\rho^2\max\{c_1^2, 1\}} \right)^{{1}/{2}}$$ and $ x= P_{C_f \cap C_g}(z)$, then $d_{C_f \cap C_g}(z) = \|z - x\|$ and $d_{C_f \cap C_g}(z^{+}) \leq \|z^{+} - x\|$. Therefore, \begin{align} d_{C_f\cap C_g}(z^{+}) &\leq C(\gamma_{f}, \gamma_{g}, \lambda, \mu_\rho)d_{C_f\cap C_g}(z). \end{align} Linear convergence of $(z^j)_{j \geq 0}$ to a point in $C_f \cap C_g$ follows from \cite[Theorem 5.12]{bauschke2011convex}. The rate follows from Equation~\eqref{thm:linearfeasibility:eq:decrease}. \qed\end{proof} \begin{remark} The recent papers \cite{bauschke2014linear,phan2014linear} have proved linear convergence of DRS applied to $f = \iota_{C_f}$ and $g = \iota_{C_g}$ under the same bounded linear regularity assumption on the pair $\{C_f, C_g\}$. In \cite{bauschke2014linear}, the proof uses the FPR to bound the distance of $z^k$ to the fixed point set of $T_{\mathrm{PRS}}$. Note that for any closed convex set $C$, we have the limit: $\mathbf{prox}_{\gamma d_{C}^2}(x) \rightarrow P_{C}(x)$ as $\gamma \rightarrow \infty$. Thus, the results of \cite{bauschke2014linear} and \cite{phan2014linear} can be seen as the limiting case of our results, but cannot be recovered from Theorem~\ref{thm:linearfeasibility}. Indeed, for any positive $\lambda$ and $\mu$, we have the limit: $C(\gamma', \gamma, \lambda, \mu) {\rightarrow} 1$, as $\gamma, \gamma' \rightarrow \infty.$ \end{remark} \begin{remark} The constant $C(\gamma, \gamma', \lambda, \mu)$ has the following form: \begin{align*} C(\gamma', \gamma, \lambda, \mu) &= \begin{cases} \left(1 - \frac{\lambda(2\gamma + 1)^2\min\left\{{\gamma}/{(2\gamma + 1)^2}, {\gamma'}/{(2\gamma' +1 )^2}\right\}}{4\gamma^2\mu^2}\right)^{{1}/{2}}, & \mbox{if } \gamma \geq \frac{1}{2}; \\ \left(1 - \frac{4 \lambda\min\left\{{\gamma}/{(2\gamma + 1)^2}, {\gamma'}/{(2\gamma' +1 )^2}\right\}}{\mu^2}\right)^{{1}/{2}}, & \mbox{otherwise.} \end{cases} \end{align*} For fixed positive $\gamma, \lambda$ and $\mu$, the function $C(\gamma', \gamma, \lambda, \mu)$ is minimized when $\gamma' = {1}/{2}$. Furthermore, it follows that that $C({1}/{2}, \gamma, \lambda, \mu)$ is minimized over $\gamma$, at $\gamma = {1}/{2}$. Finally, note that $C(\gamma', \gamma, \lambda, \mu)$ is monotonically decreasing in $\lambda$ and monotonically increasing in $\mu$. Thus, in view of Corollary~\ref{cor:AP}, we achieve the minimal constant for MAP: $C({1}/{2}, {1}/{2}, 1, \mu) = \left(1 - {1}/({2\mu^2)}\right)^{{1}/{2}}$. \end{remark} We can use Theorem~\ref{thm:linearfeasibility} to deduce the linear convergence of MAP and give an explicit rate. In \cite[Theorem 3.15]{deutsch2008rate}, the authors show that $\mu$-linear regularity of a finite collection of sets is equivalent to the linear convergence of the method of cyclic projections applied to these sets and, they derive the rate $\left(1 - {1}/({8\mu^2})\right)^{{1}/{2}}$. Corollary~\ref{cor:AP} is a special case of one direction of this result but with a better rate. It is not clear if the rate in \cite[Theorem 3.15]{deutsch2008rate} can be improved for the general cyclic projections algorithm. The rate we show in Corollary~\ref{cor:AP} appears in \cite[Corollary 3.14]{bauschke1993convergence} under the same assumptions. \begin{corollary}[Convergence of MAP] \label{cor:AP} Let $(z^j)_{j \geq 0}$ be generated by the iteration in Equation~\eqref{eq:DRSfeasibilitygamma} with $\gamma_{f, k} \equiv \gamma_{g, k} \equiv {1}/{2}$ and $\lambda_k \equiv1$. Then for all $k \geq 0$, $z^{k+1} = P_{C_f}P_{C_g}z^k$. Thus, MAP is a special case of PRS. Consequently, under the assumptions of Theorem~\ref{thm:linearfeasibility}, the iterates of MAP converge linearly to a point in the intersection of $C_f \cap C_g$ with rate $\left(1 - {1}/{\mu_\rho^2}\right)^{{1}/{2}}$. \end{corollary} \begin{proof} Notice that $x_g^k = ({1}/{2})z^k + ({1}/{2})P_{C_g}z^k$ and $\mathbf{refl}_{({1}/{2})g}(z^k) = P_{C_g}z^k$. Similarly, $x_f^k = ({1}/{2})P_{C_g}(z^k) + ({1}/{2})P_{C_f}P_{C_g}z^k$ and $z^{k+1} = \mathbf{refl}_{({1}/{2}) f}(P_{C_g}z^k) = P_{C_f}P_{C_g}z^k$. We see that $C({1}/{2}, {1}/{2}, 1, \mu) = \left(1 - {1}/({2\mu_\rho^2})\right)^{{1}/{2}}$. We can strengthen this rate to $\left(1 - {1}/{\mu_\rho^2}\right)^{{1}/{2}}$ by observing that in Equation~\eqref{thm:linearfeasibility:eq:upperboundztoc1} we have $d_{C_f}(z^k) = 0$, and so we can set $c_1 = 0$. The proof then follows the same argument. \qed\end{proof} \begin{remark} If $C_f$ and $C_g$ are closed subspaces with Friedrichs angle $\cos^{-1}(c_F)$, \cite[Corollary 11]{bauschke1999strong} shows that $\mu \leq {2}/{\sqrt{1 - c_F}}$. Therefore, Corollary~\ref{cor:AP} predicts that iterates of MAP converges with rate no less than $(({3 + c_F})/{4})^{1/2}$. The actual rate for this problem is $c_F^2$ \cite{aronszajn1950theory,kayalar1988error}. See \cite[Section 7]{bauschke2013rate} for a comparison between DRS and MAP for two subspaces. \end{remark} With this interpretation of MAP we can examine the inconsistent case, $C_f \cap C_g = \emptyset$, from a different perspective than the current literature. A part of the following result appeared in \cite[Theorem 4.8]{bauschke1994dykstra}. In particular, if $x$ satisfies Equation~\eqref{cor:AP2:eq:gap}, then $P_{C_f}x - P_{C_g}x$ is the gap vector of \cite[Theorem 4.8]{bauschke1994dykstra}. \begin{corollary}[Convergence of MAP: infeasible case]\label{cor:AP2} Let $(z^j)_{j \geq 0}$ be generated by MAP, and suppose that $C_f \cap C_g = \emptyset$. If there exists $x \in {\mathcal{H}}$ such that \begin{align}\label{cor:AP2:eq:gap} x - P_{C_f}x = P_{C_g}x - x, \end{align} then $(z^j)_{j \geq 0}$ converges weakly to a point in the following set: \begin{align}\label{eq:APfixedpoints} \{ P_{C_f}x \mid x\in {\mathcal{H}}, x - P_{C_f}x = P_{C_g}x - x\} \subseteq C_f, \end{align} with FPR rate $\|z^{k+1} - z^k\|^2 = o\left({1}/({k+1})\right)$. Furthermore, if $x$ satisfies Equation~\eqref{cor:AP2:eq:gap}, then \begin{align}\label{cor:AP2:eq:sum} \sum_{i=0}^\infty \left(\left\|\frac{1}{2}(z^i - P_{C_g}z^i) - (x - P_{C_g}x)\right\|^2 + \left\|\frac{1}{2}(P_{C_g}z^i - P_{C_f}P_{C_g}z^i) - (x - P_{C_f}x)\right\|^2 \right) < \infty. \end{align} In particular, the vector $P_{C_g}z^k - P_{C_f}P_{C_g}z^k$ strongly converges to the gap vector $P_{C_g} x - P_{C_f} x$, and $$\min_{i = 0, \cdots, k}\left\{\|(P_{C_g}z^{i} - P_{C_f}z^{i}) - (P_{C_g} x - P_{C_f} x)\|^2\right\} = o(1/(k+1)).$$ \end{corollary} \begin{proof} In view of Proposition~\ref{prop:factsaboutdistancesquared}, the condition $x - P_{C_f}x = P_{C_g}x -x$ is equivalent to $ x\in \zer(\nabla d_{C_f}^2 + \nabla d_{C_g}^2)$. The mapping $T_{\mathrm{PRS}}^{{1}/{2}, {1}/{2}} = P_{C_f}P_{C_g}$ is the composition of $({1}/{2})$-averaged maps, and so it is $\alpha$-averaged for some $\alpha < 1$ \cite[Proposition 4.32]{bauschke2011convex}. In addition, \cite[Theorem 1]{davis2014convergence} shows that the FPR satisfies $\|z^{k+1} - z^k\|^2 = o\left({1}/({k+1})\right)$. The set in Equation~\eqref{eq:APfixedpoints} is precisely the set of fixed points of $T_{\mathrm{PRS}}$. Therefore, weak convergence follows from \cite[Proposition 5.15]{bauschke2011convex}. The sum in Equation~\eqref{cor:AP2:eq:sum} is exactly the sum of derivatives $\|\nabla d_{C_g}^2(x_g^k) - \nabla d_{C_g}^2(x)\|^2 + \|\nabla d_{C_f}^2(x_f^k) - \nabla d_{C_f}^2(x)\|^2$, and so it is finite by Proposition~\ref{prop:sumauxilliaryterms}. Finally, strong convergence of $P_{C_g}z^k - P_{C_f}P_{C_g}z^k$ to the gap vector follows from the identity $x - P_{C_f}x = (1/2)(P_{C_g} x - P_{C_f} x)$. The rate is a consequence of~Fact~\ref{lem:sumsequence} and Equation~\eqref{cor:AP2:eq:sum}. \qed\end{proof} \begin{remark}\label{rem:gapvector} Note that that the condition $x - P_{C_f}x = x - P_{C_g}x$ is equivalent to $\|P_{C_g}x - P_{C_f}x\|^2 = \min_{y \in {\mathcal{H}}} (d_{C_f}^2(y) + d_{C_g}^2(y))= \min_{x_f \in C_f, x_g \in C_g} \|x_g - x_f\|^2$. See \cite[Fact 5.1]{bauschke1994dykstra} for conditions that guarantee the infimum is attained in Corollary~\ref{cor:AP2}. \end{remark} See Appendix~\ref{app:feasibilitymultiplesets} for the extension of the results of this section to finite collections of sets. \section{Introduction} The Douglas-Rachford splitting (DRS), Peaceman-Rachford splitting (PRS), and alternating direction method of multipliers (ADMM) algorithms are abstract splitting schemes that solve monotone inclusion and convex optimization problems \cite{lions1979splitting,GlowinskiADMM,gabay1976dual}. The DRS and PRS algorithms solve monotone inclusion problems in which the operator is the sum of two (possibly) simpler operators by accessing each operator individually through its resolvent. The ADMM algorithm solves convex optimization problems in which the objective is the sum of two (possibly) simpler functions with variables linked through a linear constraint via an alternating minimization strategy. The variable splitting that occurs in each of these algorithms can give rise to parallel and even distributed implementations of minimization algorithms \cite{boyd2011distributed,shi2013linear,wei2012distributed}, which are particularly suitable for large-scale applications. Since the 1950s, these methods were largely applied to solving partial differential equations (PDEs) and feasibility problems, and only recently has their power been utilized in (PDE and non-PDE related) image processing, statistical and machine learning, compressive sensing, matrix completion, finance, and control \cite{goldstein2009split,boyd2011distributed}. In this paper, we consider two prototype optimization problems: the unconstrained problem \begin{align}\label{eq:simplesplit} \Min_{x\in {\mathcal{H}}}~ f(x) + g(x) \end{align} where ${\mathcal{H}}$ is a Hilbert space, and the linearly constrained variant \begin{align*} \Min_{x \in {\mathcal{H}}_1,~ y \in {\mathcal{H}}_2} & \; f(x) + g(y) \\ \St~ & \; Ax + By = b \numberthis \label{eq:simplelinearconstrained} \end{align*} where ${\mathcal{H}}_1, {\mathcal{H}}_2$, and ${\mathcal{G}}$ are Hilbert spaces, the vector $b$ is an element of ${\mathcal{G}}$, and $A : {\mathcal{H}}_1 \rightarrow {\mathcal{G}}$ and $B : {\mathcal{H}}_2 \rightarrow {\mathcal{G}}$ are linear operators. Problem~\eqref{eq:simplesplit} models a variety of tasks in signal recovery where one function corresponds to a \emph{data fitting term} and the other enforces \emph{prior knowledge}, such as sparsity, low rank, or smoothness \cite{combettes2011proximal}. In this paper, we apply relaxed PRS (Algorithm~\ref{alg:DRS}) to solve Problem~\eqref{eq:simplesplit}. On the other hand, Problem~\eqref{eq:simplelinearconstrained} models tasks in machine learning, image processing and distributed optimization. The linear constraint can be used to enforce data fitting, but it can also be used to split variables in a way that gives rise to parallel or distributed optimization algorithms \cite{bertsekas1989parallel,boyd2011distributed}. We will apply relaxed ADMM (Algorithm~\ref{alg:ADMM}) to Problem~\eqref{eq:simplelinearconstrained}. \subsection{Goals, challenges, and approaches} This work improves the theoretical understanding of DRS, PRS, and ADMM, as well as their averaged versions. When applied to convex optimization problems, they are known to converge under rather general conditions \cite[Corollary 27.4]{bauschke2011convex}. This work seeks to complement the results of \cite{davis2014convergence}, which are developed under general convexity assumptions, by deriving stronger rates under correspondingly stronger conditions on Problems~\ref{eq:simplesplit} and~\ref{eq:simplelinearconstrained}. One of the main consequences of this work is that the relaxed PRS and ADMM algorithms automatically adapt to the regularity of the problem at hand and achieve convergence rates that improve upon the worst-case rates shown in \cite{davis2014convergence} for the nonsmooth case. Thus, our results offer an explanation of the great performance of relaxed PRS and ADMM observed in practice, and together with \cite{davis2014convergence} we now have a comprehensive convergence rate analysis of the relaxed PRS and ADMM algorithms. \begin{table} \renewcommand{\arraystretch}{1.3} \begin{center} \begin{tabular}{c|c|c|c}\hline Regularity assumption & \multicolumn{2}{|c|}{Objective error} & \multirow{2}{*}{FPR} \\\cline{2-3} beyond convexity & Rate & Type &\\\hline\hline None & \multicolumn{2}{|c|}{not available} & \multirow{5}{*}{$o(1/k)$} \\\cline{1-3} \multirow{2}{*}{Lipschitz $f$ or $g$~\cite{davis2014convergence}} & $o(1/\sqrt{k})$ & nonergodic & \\\cdashline{2-3} & $O(1/k)$ & ergodic$^\dag$ & \\\cline{1-1}\cline{2-3} \multirow{2}{*}{Strongly convex $f$ or $g$} & $o(1/k)$ & best itr. & \\\cdashline{2-3} & $O(1/k)$ & ergodic$^\dag$ & \\\cline{1-3} \multirow{2}{*}{Lipschitz $\nabla g$} & $o(1/k)$ & best itr. & \\\cline{2-4} & $o(1/k)$$^\ddag$ & nonergodic$^\ddag$ & $o(1/k^2)$$^\ddag$ \\\hline Lipschitz $\nabla f$ or $\nabla g$, & \multirow{2}{*}{$O(e^{-k})$} & \multirow{2}{*}{R-linear} & \multirow{2}{*}{$O(e^{-k})$}\\ strongly convex $f$ or $g$ & & & \\\hline $f = d_{C_1}^2$ and $g = d_{C_2}^2$, & \multirow{2}{*}{$O(e^{-k})$} & \multirow{2}{*}{R-linear} & \multirow{2}{*}{$O(e^{-k})$}\\ $\{C_1, C_2\}$ linearly regular & & & \\\hline \end{tabular}\\[5pt] \caption{Summary of convergence rates for \emph{relaxed PRS} with relaxation parameters $\lambda_k\in(\epsilon,1-\epsilon)$, for any $\epsilon>0$. FPR stands for the fixed-point residual $\|T_{\mathrm{PRS}} z^{k} - z^k\|^2$. $^\dag$These two ergodic rates hold for $\lambda_k\in(\epsilon,1]$. $^\ddag$These rates hold for DRS ($\lambda_k=1/2$) and properly bounded step size $\gamma$.\label{tb:rPRS}} \end{center} \end{table} \begin{table} \begin{center} \begin{tabular}{l|c|c|c|c|c}\hline &\multicolumn{3}{c|}{Regularity assumption beyond convexity} & Convergence & \multirow{2}{*}{Type} \\\cline{2-4} & Strongly convex & Lipschitz & Full rank & rate & \\\hline\hline \multirow{4}{*}{1} & \multirow{4}{*}{-} & \multirow{4}{*}{-} & \multirow{4}{*}{-} & $o(1/k)$ & nonergodic feas. \\\cline{5-6} & & & & $O(1/k^2)$ & ergodic feas. \\\cline{5-6} & & & & $o(1/\sqrt{k})$ & nonergodic obj. error \\\cline{5-6} & & & & $O(1/k)$ & ergodic obj. error\\\hline \multirow{2}{*}{2} & \multirow{2}{*}{$g$} & \multirow{2}{*}{-} & \multirow{2}{*}{-} & $o(1/k^2)$ & feasibility \\\cline{5-6} & & & & $o(1/k)$ & objective \\\hline 3 & $g$ & $\nabla g$ & $B$ (row rank) & \multirow{4}{*}{$O(e^{-k})$} & R-linear \\\cline{1-4} 4 & $f$ & $\nabla f$ & $A$ (row rank) & & feasibility, \\\cline{1-4} 5 & $f$ & $\nabla g$ & $B$ (row rank) & & objective error, \\\cline{1-4} 6 & $g$ & $\nabla f$ & $A$ (row rank) & & solution error \\ \hline \end{tabular}\\[5pt] \end{center} \caption{Summary of convergence rates for relaxed ADMM. Feasibility is $\|Ax^k+By^k-b\|^2$, objective error is $(f(x^k)+g(y^k))-(f(x^*)+g(y^*))$, and solution error includes $\|w^k-w^*\|^2$, $\|Ax^k-Ax^*\|^2$, and $\|By^k-By^*\|^2$. Case 1 is from~\cite{davis2014convergence}, where the nonergodic rates hold for relaxation parameters $\lambda_k\in(\epsilon,1-\epsilon)$, for any $\epsilon>0$, and the ergodic rates hold for $\lambda_k\in(\epsilon,1]$. Case 2 also requires a bounded step size. Each of cases 3--6 ensures R-linear convergence. \label{tb:ADMM} } \end{table} In this paper, we derive the convergence rates of the objective error and fixed-point residual (FPR) of relaxed PRS applied to Problem~\eqref{eq:simplesplit}; see Table \ref{tb:rPRS}. In addition, we derive the convergence rates of the constraint violations and objective errors for relaxed ADMM applied to Problem~\eqref{eq:simplelinearconstrained}; see Table \ref{tb:ADMM}. {By appealing to counterexamples in~\cite{davis2014convergence}, several of the rates in Table~\ref{tb:rPRS} can be shown to be tight up to constant factors.} The derived rates are useful for determining how many iterations of the relaxed PRS and ADMM algorithms are needed in order to reach a certain accuracy, to decide when to stop an algorithm, and to compare relaxed PRS and ADMM to other algorithms in terms of their worst-case complexities. \subsection{Notation} In what follows, ${\mathcal{H}}, {\mathcal{H}}_1, {\mathcal{H}}_2, {\mathcal{G}}$ denote (possibly infinite dimensional) Hilbert spaces. In fixed-point iterations, $(\lambda_j)_{j \geq 0} \subset {\mathbf{R}}_+$ will denote a sequence of relaxation parameters, and \begin{equation}\label{def:Lambda}\Lambda_k := \sum_{i=0}^k \lambda_i \end{equation} is its $k$th partial sum. To ease notational memory, the reader may assume that $\lambda_k \equiv (1/2)$ and $\Lambda_k =(k+1)/2$ in the DRS algorithm, or that $\lambda_k \equiv 1$ and $\Lambda_k = (k+1)$ in the PRS algorithm. Given the sequence $(x^j)_{j \geq 0}\subset {\mathcal{H}}$, we let $\overline{x}^k = ({1}/{\Lambda_k})\sum_{i=0}^k \lambda_i x^i$ denote its $k$th average with respect to the sequence $(\lambda_j)_{j \geq 0}$. A convergence result is \emph{ergodic} if it applies to the sequence $(\overline{x}^j)_{j \geq 0}$, and \emph{nonergodic} if it applies to the sequence $(x^j)_{j \geq 0}$. Given a closed, proper, and convex function $f : {\mathcal{H}} \rightarrow (-\infty, \infty]$, the set $\partial f(x)$ denotes its subdifferential at $x$ and $ \widetilde{\nabla} f(x) \in \partial f(x) $ denotes a subgradient. (This notation was used in \cite[Eq. (1.10)]{bertsekas2011incremental}.) The convex conjugate of a closed, proper, and convex function $f$ is $ f^\ast(y) := \sup_{x \in {\mathcal{H}}} \dotp{y, x} - f(x). $ Let $I_{{\mathcal{H}}}: {\mathcal{H}} \rightarrow {\mathcal{H}}$ denote the identity map. For any point $x \in {\mathcal{H}}$ and $\gamma \in {\mathbf{R}}_{++}$, we let $\mathbf{prox}_{\gamma f}(x) := \argmin_{y \in {\mathcal{H}}} f(y) + \frac{1}{2\gamma} \|y - x\|^2$ and $\mathbf{refl}_{\gamma f} := 2\mathbf{prox}_{\gamma f} - I_{{\mathcal{H}}},$ which are known as the \emph{proximal} and \emph{reflection} operators. In addition, we define the PRS operator: \begin{align*} T_{\mathrm{PRS}} &:= \mathbf{refl}_{\gamma f} \circ \mathbf{refl}_{\gamma g}. \end{align*} Let $\lambda > 0$. For every nonexpansive map $T : {\mathcal{H}} \rightarrow {\mathcal{H}}$ we define the averaged map: \begin{align*} T_{\lambda} := (1-\lambda)I_{{\mathcal{H}}} + \lambda T. \end{align*} We call the following identity the \emph{cosine rule}: \begin{align*} \|y-z\|^2+2\dotp{y-x,z-x}=\|y-x\|^2+\|z-x\|^2,\quad\forall x,y,z\in{\mathcal{H}} \numberthis\label{eq:cosinerule}. \end{align*} \subsection{Assumptions} We list the the assumptions used throughout this papers as follows. \begin{assump}[Problem assumptions] Every function we consider is closed, proper, and convex. \end{assump} Unless otherwise stated, a function is not necessarily differentiable. \begin{assump}[Solution existence]\label{assump:additivesub} Functions $f, g : {\mathcal{H}} \rightarrow (-\infty, \infty]$ satisfy \begin{align} \zer(\partial f + \partial g) \neq \emptyset. \end{align} \end{assump} Note that this assumption is slightly stronger than the existence of a minimizer because $\zer(\partial f + \partial g) \neq \zer(\partial (f + g))$, in general \cite[Remark 16.7]{bauschke2011convex}. Nevertheless, this assumption is standard. \begin{assump}[Differentiability] Every differentiable function is Fr{\'e}chet differentiable \cite[Def. 2.45]{bauschke2011convex}. \end{assump} \subsection{The Douglas-Rachford and relaxed Peaceman-Rachford Splitting Algorithms} The results of this paper apply to several operator-splitting algorithms that are all based on the atomic evaluation of the \emph{proximal operator}. By default, all algorithms start from an arbitrary $z^0 \in {\mathcal{H}}$. The Douglas-Rachford splitting (DRS) algorithm applied to minimizing $f+g$ is as follows: \begin{align*} \begin{cases} x_g^k = \mathbf{prox}_{\gamma g}(z^k);\\ x_f^k = \mathbf{prox}_{\gamma f}( 2 x_g^k - z^k);\\ z^{k+1} = z^k + (x_f^k - x_g^k); \end{cases} \quad k = 0, 1, \ldots, \end{align*} which has the equivalent operator-theoretic and subgradient form (Lemma~\ref{prop:DRSmainidentity}): \begin{align*} z^{k+1} &= \frac{1}{2}(I_{{\mathcal{H}}} + T_{\mathrm{PRS}})(z^k) = z^k - \gamma(\widetilde{\nabla} f(x_f^k) + \widetilde{\nabla} g(x_g^k)), \quad k = 0, 1, \ldots, \end{align*} where $\widetilde{\nabla} f(x_f^k)\in\partial f(x_f^k)$ and $\widetilde{\nabla} g(x_g^k)\in \partial g(x_g^k)$. (See Part~\ref{prop:basicprox:part:optprox} of Proposition~\ref{prop:basicprox} for how the notation $\widetilde{\nabla}$ relates to $\mathbf{prox}$.) In the above algorithm, we can replace the $(1/2)$-average of $I_{{\mathcal{H}}}$ and $T_{\mathrm{PRS}}$ with any other weight; this results the \emph{relaxed PRS} algorithm: \begin{algorithm}[H] \begin{algorithmic} \Require $z^0 \in {\mathcal{H}}, ~\gamma > 0, ~(\lambda_j)_{j \geq 0}\subset (0, 1]$ \For{$k=0,~1,\ldots$} \State $z^{k+1} = (1-\lambda_k)z^k + \lambda_k\mathbf{refl}_{\gamma f} \circ \mathbf{refl}_{\gamma g}(z^k) $\; \EndFor \end{algorithmic} \caption{{Relaxed Peaceman-Rachford Splitting (relaxed PRS)}} \label{alg:DRS} \end{algorithm} The special cases $\lambda_k \equiv 1/2$ and $\lambda_k \equiv 1$ are called the DRS and PRS algorithms, respectively. \subsection{Practical implications: a comparison with forward-backward splitting}\label{section:FBSpractical} Suppose that the function $g$ in Problem~\ref{eq:simplesplit} is differentiable and $\nabla g$ is $(1/\beta)$-Lipschitz. Under this smoothness assumption, we can apply FBS algorithm to Problem~\ref{eq:simplesplit}: given $z^0 \in {\mathcal{H}}$, for all $k \geq 0$, define \begin{align*} z^{k+1} = \mathbf{prox}_{\gamma f} (z^k - \gamma \nabla g(z^k)). \end{align*} To ensure convergence, the stepsize parameter $\gamma$ must be strictly less than $2\beta$. Now because the gradient operator is often simpler to evaluate than the proximal operator, it may be preferable to use FBS instead of relaxed PRS whenever one of the objectives is differentiable. From our results, we can give two reasons why it may be preferable to use relaxed PRS over FBS: \begin{enumerate} \item If the Lipschitz constant of the gradient is known, our analysis indicates how to properly choose stepsizes of relaxed PRS so that both algorithms converge with the same rate (Theorem~\ref{thm:differentiableobjective}). In practice, relaxed PRS is often observed to converge faster than FBS, so our results at least indicate that we can do no worse by using relaxed PRS. \item If the Lipschitz constant of the gradient is not known, a line search procedure can be used to guarantee convergence of FBS. If this procedure is more expensive than evaluating the proximal operator, then relaxed PRS should be used. Indeed, Theorem~\ref{thm:lipschitzbest} shows that the ``best iterate" of relaxed PRS will converge with rate $o(1/(k+1))$ regardless of the chosen stepsize, whereas FBS may fail to converge. \end{enumerate} Thus, one of our main contributions is the ``demystification" of parameter choices, and a partial explanation of the perceived practical advantage of relaxed PRS over FBS. \subsection{Basic properties of proximal operators}\label{sec:nonexpansive} The following properties are included in textbooks such as \cite{bauschke2011convex}. \begin{proposition}\label{prop:basicprox} Let $f, g : {\mathcal{H}} \rightarrow (-\infty, \infty)$ be closed, proper, and convex functions, and let $T : {\mathcal{H}} \rightarrow {\mathcal{H}}$ be nonexpansive. The the following are true: \begin{enumerate} \item\label{prop:basicprox:part:optprox} {\bf Optimality conditions of $\mathbf{prox}$:} Let $x \in {\mathcal{H}}$. Then $x^+ = \mathbf{prox}_{\gamma f} (x)$ if, and only if, $$\widetilde{\nabla} f(x^+) :=\frac{1}{\gamma}(x-x^+) \in \partial f(x^+).$$ \item \label{cor:proxcontraction} {\bf The proximal operator $\mathbf{prox}_{\gamma f} : {\mathcal{H}} \rightarrow {\mathcal{H}}$ is ${1}/{2}$-averaged:} \begin{align}\label{cor:proxcontraction:eq:main} \|\mathbf{prox}_{\gamma f}(x) - \mathbf{prox}_{\gamma f}(y) \|^2 &\leq \|x - y\|^2 - \|(x - \mathbf{prox}_{\gamma f}(x)) - (y - \mathbf{prox}_{\gamma f}(y))\|^2. \end{align} \item \label{prop:basicprox:part:nonexpansive} {\bf Nonexpansiveness of the PRS operator:} The operator $\mathbf{refl}_{\gamma f} : {\mathcal{H}} \rightarrow {\mathcal{H}}$ is nonexpansive. Therefore, the composition $T_{\mathrm{PRS}} = \mathbf{refl}_{\gamma f} \circ \mathbf{refl}_{\gamma g}.$ is nonexpansive. \end{enumerate} \end{proposition} \section{Linear convergence}\label{sec:linearconvergence} In this section, we study the convergence rate of relaxed PRS under the assumption \begin{assump}\label{assump:mixed} The gradient of at least one of the functions $f$ and $g$ is Lipschitz, and at least one of the functions $f$ and $g$ is strongly convex. In symbols: $(\mu_f + \mu_g)(\beta_f + \beta_g) > 0$. \end{assump} Linear convergence of relaxed PRS is expected whenever Assumption~\ref{assump:mixed} is true. In addition, by the strong convexity of $f + g$, the minimizer of Problem~\eqref{eq:simplesplit} is unique. The following proposition lists some consequences of linear convergence of the relaxed PRS sequence $(z^j)_{j \geq 0}$. \begin{proposition}[Consequences of linear convergence]\label{prop:linearconvergenceimplies} Let $(C_j)_{j \geq0} \subseteq [0, 1]$ be a positive scalar sequence, and suppose that for all $k \geq0$, \begin{align}\label{prop:linearconvergenceimplies:eq:main} \|z^{k+1} - z^\ast\| \leq C_k\|z^k - z^\ast\|. \end{align} Fix $k \geq 1$. Then \begin{align*} \|x_g^k - x^\ast\|^2 + \gamma^2\|\widetilde{\nabla} g(x_g^k) - \widetilde{\nabla} g(x^\ast)\|^2 \leq \|z^0 - z^\ast\|^2\prod_{i = 0}^{k-1} C_i^2; \\ \|x_f^k - x^\ast\|^2 + \gamma^2\|\widetilde{\nabla} f(x_f^k) - \widetilde{\nabla} f(x^\ast)\|^2 \leq \|z^0 - z^\ast\|^2\prod_{i = 0}^{k-1} C_i^2. \end{align*} If $\lambda <1$, then the FPR rate holds: $\|(T_{\mathrm{PRS}})_{\lambda}z^k- z^k\| \leq \sqrt{\lambda/(1-\lambda)}\|z^0 - z^\ast\|\prod_{i = 0}^{k-1} C_i.$ Consequently, if the gradient $\nabla f$ (respectively $\nabla g$), is $({1}/{\beta})$-Lipschitz and $x^k = x_g^k$ (respectively $x^k = x_f^k$), then \begin{align*} &f(x^k) + g(x^k) - f(x^\ast) - g(x^\ast) \leq \frac{\|z^0 - z^\ast\|^2}{\gamma} \prod_{i=0}^{k-1}C_i^2 \times\begin{cases} 1 , & \mbox{if } \gamma \leq \beta; \\ 1 + \frac{(\gamma - \beta)}{2\beta}, & \mbox{otherwise.}\\ \end{cases} \end{align*} \end{proposition} \begin{proof} The bounds for $x_g^k$ and $x_f^k$ follow because $\|x_g^k - x^\ast\|^2 + \gamma^2\|\nabla g(x_g^k) -\nabla g(x^\ast)\|^2 \leq \|z^k - z^\ast\|^2,$ and $\|x_f^k - x^\ast\|^2 + \gamma^2\|\widetilde{\nabla} f(x_f^k) - \widetilde{\nabla} f(x^\ast)\|^2 \leq \|\mathbf{refl}_{\gamma g}(z^k) - \mathbf{refl}_{\gamma g}(z^\ast)\|^2 \leq \|z^k - z^\ast\|^2$ by Part~\ref{cor:proxcontraction} of Proposition~\ref{prop:basicprox}, the nonexpansiveness of $\mathbf{refl}_{\gamma f}$, and Equation~\eqref{prop:linearconvergenceimplies:eq:main}. The FPR convergence rate follows from the Fej\'er-type inequality in Equation~\eqref{eq:fejer}. Now fix $k \geq 1$, and let $z_\lambda = (T_{\mathrm{PRS}})_\lambda z^k$ for all $\lambda \in [0, 1]$. Then Proposition~\ref{prop:DRSupperfglipschitz} shows that: \begin{align*} &f(x^k) + g(x^k) - f(x^\ast) - g(x^\ast) \\ &\leq \inf_{\lambda \in [0, 1]}\frac{1}{4\gamma \lambda}\begin{cases} \|z^k - z^\ast\|^2 - \|z_\lambda - z^\ast\|^2 +\left(1 + \frac{1}{2\lambda}\left( \frac{\gamma}{\beta} - 1\right)\right)\|z^k - z_\lambda\|^2, & \mbox{if } \gamma \leq \beta; \\ \left(1 + \frac{(\gamma - \beta)}{2\beta}\right)\left(\|z^k - z^\ast\|^2 - \|z_\lambda - z^\ast\|^2 + \|z^k - z_\lambda\|^2\right), & \mbox{otherwise.} \end{cases} \\ &\leq \frac{1}{2\gamma}\begin{cases} \|z^k - z^\ast\|^2+ \|z^k - z_{1/2}\|^2 , & \mbox{if } \gamma \leq \beta; \\ \left(1 + \frac{(\gamma - \beta)}{2\beta}\right)\left(\|z^k - z^\ast\|^2 + \|z^k - z_{1/2}\|^2\right), & \mbox{otherwise.}\\ \end{cases} \numberthis\label{eq:preboundlinear} \end{align*} The objective error rate now follows from Equation~\eqref{eq:preboundlinear} and the FPR convergence rate. \qed\end{proof} Whenever $\sup_{j \geq 0} C_j < 1$, Proposition~\ref{prop:linearconvergenceimplies} gives the linear convergence rates of the sequences $(z^j)_{j \geq 0}$, $(x_g^j)_{j \geq 0}$ and $(x_f^j)_{j \geq 0}$, the subgradient error, the FPR, and the objective error. In the following sections, we will prove Inequality~\eqref{prop:linearconvergenceimplies:eq:main} holds under several different regularity assumptions on $f$ and $g$. In each case we leave it to the reader to apply Proposition~\ref{prop:linearconvergenceimplies}. \subsection{Solely regular $f$ or $g$}\label{sec:strongconvexity} Throughout this subsection, at least one of the functions $f$ and $g$ will carry both regularity properties. In symbols: $\mu_f\beta_f + \mu_g\beta_g > 0$. The following theorem recovers \cite[Proposition 4]{lions1979splitting} as a special case ($\lambda_k \equiv 1/2$). \begin{theorem}[Linear convergence with regularity of $g$]\label{thm:nonergodicstronglipschitzderivative} Let $z^\ast$ be a fixed point of $T_{\mathrm{PRS}}$, let $x^\ast = \mathbf{prox}_{\gamma g}(z^\ast)$, and suppose that $\mu_g\beta_g > 0$. For all $\lambda \in [0, 1]$, let $C(\lambda) := \left(1 - {4\gamma\lambda\mu_g}/{(1 + {\gamma}/{\beta_g})^2}\right)^{{1}/{2}}$. Then for all $k \geq 0$, $\|z^{k+1} - z^\ast\| \leq C(\lambda_k) \|z^k - z^\ast\|.$ \end{theorem} \begin{proof} Theorem~\ref{prop:sumauxilliaryterms} bounds the distance of $x_g^k$ to the minimizer \begin{align*} \frac{8\gamma\lambda_k\mu_g}{2}\|x_g^k - x^\ast\|^2 &\stackrel{\eqref{prop:sumauxilliaryterms:eq:main}}{\leq} \|z^{k} - z^\ast\|^2 - \|z^{k+1} - z^\ast\|^2. \numberthis \label{thm:nonergodicstronglipschitzderivative:eq:2} \end{align*} Now we use the identity $z^{k} = x_g^k + \gamma \nabla g(x_g^k)$ and the Lipschitz continuity of $\nabla g$ to upper bound $\|z^k - z^\ast\|^2$ by a multiple of $\|x_g^k - x^\ast\|^2$: $\|z^k - z^\ast\|^2 \leq \left(1+\gamma/\beta_g\right)^2\|x_g^k - x^\ast\|^2.$ Rearrange Equation~\eqref{thm:nonergodicstronglipschitzderivative:eq:2} with this bound to complete the proof. \qed\end{proof} \begin{remark} For all $\lambda \in [0, 1]$, the constant $C(\lambda)$ is minimal when $\gamma = \beta_g$, i.e. $C(\lambda) = \left( 1- \lambda_k\mu_g\beta_g\right)^{{1}/{2}}$. Furthermore, for any choice of $\gamma$, we have the bound $C(1) \leq C(\lambda)$. In particular, for $g = ({1}/{2})\|\cdot \|^2$, the PRS algorithm converges in one step ($C(1) = 0$). Thus, this rate is tight. \end{remark} The following theorem deduces linear convergence of relaxed PRS whenever $f$ carries both regularity properties. Note that linear convergence of the PRS algorithm ($\lambda_k \equiv 1$) does not follow. \begin{theorem}[Linear convergence with regularity of $f$]\label{prop:regf} Let $z^\ast$ be a fixed point of $T_{\mathrm{PRS}}$, let $x^\ast = \mathbf{prox}_{\gamma g}(z^\ast)$, and suppose that $\mu_f\beta_f > 0$. For all $\lambda \in [0, 1]$, let $$C(\lambda) := \left(1 - (\lambda/2)\min\left\{ {4\gamma \mu_f}/{\left(1+{\gamma}/{\beta_f}\right)^2}, (1 - \lambda)\right\}\right)^{{1}/{2}}.$$ Then for all $k \geq 0$, $\|z^{k+1} - z^\ast\| \leq C(\lambda_k) \|z^k - z^\ast\|.$ \end{theorem} \begin{proof} Theorem~\ref{prop:sumauxilliaryterms} bounds the distance of $x_f^k$ to the minimizer (where we substitute $z^{k+1} - z^k = 2\lambda_k(x_f^k - x_g^k)$) \begin{align*} 4\gamma\lambda_k\mu_f\|x_f^k - x^\ast\|^2 + 4\lambda_k\left(1-\lambda_k \right) \|x_f^k - x_g^k\|^2&\stackrel{\eqref{prop:sumauxilliaryterms:eq:main}}{\leq} \|z^k - z^\ast\|^2 - \|z^{k+1} - z^\ast\|^2. \numberthis \label{prop:regf:eq:first} \end{align*} Recall the identities: \begin{align*} z^k = x_g^k + \gamma \nabla g(x_g^k) = x_f^k - \gamma \nabla f(x_f^k) + 2(x_g^k - x_f^k) && \mathrm {and} && z^\ast = x^\ast - \gamma \nabla f(x^\ast). \end{align*} Therefore, by the convexity of $\|\cdot \|^2$, we can bound the distance of $z^k$ to the fixed point $z^\ast$ \begin{align*} \|z^k - z^\ast\|^2 \leq 2\left(\left(1+\frac{\gamma}{\beta_f}\right)^2\|x_f^k - x^\ast\|^2 + 4\|x_g^k - x_f^k\|^2\right). \numberthis\label{prop:regf:eq:second} \end{align*} Equations~\eqref{prop:regf:eq:first} and~\eqref{prop:regf:eq:second} produce the contraction: \begin{align*} C'\|z^k - z^\ast \|^2 + \|z^{k+1} - z^\ast\|^2 \leq 4\gamma\lambda_k\mu_f\|x_f^k - x^\ast\|^2 + 4\lambda_k(1 - \lambda_k)\|x_f^k - x_g^k\|^2 + \|z^{k+1} - z^\ast\|^2 \leq \|z^k - z^\ast\|^2 \end{align*} where $C' = {(\lambda_k/2)\min\left\{ {4\gamma \mu_f}/{\left(1+{\gamma}/{\beta_f}\right)^2}, (1 - \lambda_k)\right\}}.$ \qed\end{proof} \subsection{Complementary regularity of $f$ and $g$} In this subsection, we assume that $f$ and $g$ share the regularity. In symbols: $\mu_f\beta_g + \mu_g\beta_f > 0$. In this case, linear convergence is expected. To the best of our knowledge, the next result is new. \begin{theorem}[Linear convergence: mixed case]\label{thm:sharedreg} Let $z^\ast$ be a fixed point of $T_{\mathrm{PRS}}$, let $x^\ast = \mathbf{prox}_{\gamma g}(z^\ast)$, and suppose that $\nabla g$, (respectively $\nabla f$), is $({1}/{\beta})$-Lipschitz and $f$, (respectively $g$), is $\mu$-strongly convex. For all $\lambda \in [0, 1]$, let $C(\lambda) := \left(1 - {(4\lambda/3)\min\{ \gamma \mu, { \beta}/{\gamma}, (1 - \lambda)\}}\right)^{{1}/{2}}$. Then for all $k \geq 0$, $\|z^{k+1} - z^\ast\| \leq C(\lambda_k)\|z^k - z^\ast\|.$ \end{theorem} \begin{proof} First assume that $\mu_f\beta_g > 0$. Theorem~\ref{prop:sumauxilliaryterms} bounds the distance of $x_f^k$ to the minimizer and the distance of $\nabla g(x_g^k)$ to the optimal gradient (where we substitute $z^{k+1} - z^k = 2\lambda_k(x_f^k - x_g^k)$): \begin{align*} 4\gamma\lambda_k\mu\|x_f^k - x^\ast\|^2 &+ 4\gamma\lambda_k \beta\|\nabla g(x_g^k) - \nabla g(x^\ast)\|^2 + 4\lambda_k\left(1- \lambda_k\right) \|x_f^k - x_g^k\|^2\\ &\stackrel{\eqref{prop:sumauxilliaryterms:eq:main}}{\leq} \|z^k - z^\ast\|^2 - \|z^{k+1} - z^\ast\|^2. \numberthis\label{thm:sharedreg:eq:first} \end{align*} Recall the identities: \begin{align*} z^k = x_g^k + \gamma \nabla g(x_g^k) = x_f^k + \gamma \nabla g(x_g^k) + (x_g^k - x_f^k) && \mathrm{and} && z^\ast = x^\ast + \gamma \nabla g(x^\ast). \end{align*} Thus, from the convexity of $\|\cdot \|^2$, \begin{align*} \|z^k - z^\ast\|^2 \leq 3\left(\|x_f^k - x^\ast\|^2 + \|\gamma \nabla g(x_g^k) - \gamma \nabla g(x^\ast)\|^2 + \|x_g^k - x_f^k\|^2\right) \numberthis \label{thm:sharedreg:eq:second} . \end{align*} We use Equation~\eqref{thm:sharedreg:eq:second} to bound the distance of $z^k$ to the fixed point $z^\ast$ by the left hand side of Equation~\eqref{thm:sharedreg:eq:first}: \begin{align*} C'\|z^k - z^\ast \|^2 &\leq 4\gamma\lambda_k\mu\|x_f^k - x^\ast\|^2 + 4\lambda_k ({ \beta}/{\gamma})\|\gamma\nabla g(x_g^k) - \gamma\nabla g(x^\ast)\|^2 + 4\lambda_k(1 - \lambda_k)\|x_f^k - x_g^k\|^2 \end{align*} where $C' = {(4\lambda_k/3)\min\{ \gamma \mu, { \beta}/{\gamma}, (1 - \lambda_k)\}}.$ Therefore, we reach the contraction: \begin{align*} \|z^{k+1} - z^\ast\| \leq \left(1 - {(4\lambda_k/3)\min\{ \gamma \mu, { \beta}/{\gamma}, (1 - \lambda_k)\}}\right)^{{1}/{2}}\|z^k - z^\ast\|^2. \end{align*} If $\mu_g\beta_f> 0$, then the proof is nearly identical, but relies on the identity: \begin{align*} z^k = x_g^k + \gamma \widetilde{\nabla} g(x_g^k) = x_g^k - \gamma \nabla f(x_f^k) + (x_g^k - x_f^k). \end{align*} \qed\end{proof} \section{Lipschitz derivatives}\label{sec:lipschitzderivatives} In this section, we study the convergence rate of relaxed PRS under the following assumption. \begin{assump} The gradient of at least one of the functions $f$ and $g$ is Lipschitz. \end{assump} Throughout this section, Fact~\ref{lem:sumsequence} will be used repeatedly to deduce the convergence rates of summable sequences. In general, because we can only deduce the summability and not the monotonicity of the objective errors in Problem~\ref{eq:simplesplit}, we can only show that the smallest objective error after $k$ iterations is of order $o(1/(k+1))$. If $\lambda_k \equiv 1/2$, the implicit stepsize parameter $\gamma$ is small enough, and the gradient of $g$ is $(1/\beta)$-Lipschitz, we show that a sequence that dominates the objective error is monotonic and summable, and deduce a convergence rate for the entire sequence. \subsection{The general case: best iterate convergence rate}\label{sec:lipschitzderivatives:part:general} The next proposition bounds the objective error by a summable sequence. See Appendix~\ref{app:lipschitzderivatives:part:general} for a proof. \begin{proposition}[Fundamental inequality under Lipschitz assumptions]\label{prop:DRSupperfglipschitz} Let $z \in {\mathcal{H}}$, let $z^+ = (T_{\mathrm{PRS}})_{\lambda}z$, let $z^\ast$ be a fixed point of $T_{\mathrm{PRS}}$, and let $x^\ast = \mathbf{prox}_{\gamma g}(z^\ast)$. If $\nabla f$ (respectively $\nabla g$) is $({1}/{\beta})$-Lipschitz, then for $x = x_g$ (respectively $x = x_f$), \begin{align*} &4\gamma\lambda\big(f(x) + g(x) - f(x^\ast) - g(x^\ast)\big) \\ &\leq\begin{cases} \|z - z^\ast\|^2 - \|z^+ - z^\ast\|^2 +\left(1 + \frac{1}{2\lambda}\left( \frac{\gamma}{\beta} - 1\right)\right)\|z - z^+\|^2, & \mbox{if } \gamma \leq \beta; \\ \left(1 + \frac{(\gamma - \beta)}{2\beta}\right)\left(\|z - z^\ast\|^2 - \|z^+ - z^\ast\|^2 + \|z - z^+\|^2\right), & \mbox{otherwise.} \end{cases} \end{align*} \end{proposition} Proposition~\ref{prop:DRSupperfglipschitz} shows that the the objective error is summable whenever $f$ or $g$ is Lipschitz and $(\lambda_j)_{j \geq 0}$ is chosen properly. A direct application of Fact~\ref{lem:sumsequence} yields a convergence rate for the objective error. Depending on the choice of $\gamma$ and $(\lambda_j)_{j \geq 0}$, we can achieve several different rates. In the following Theorem we only analyze a few such choices. \begin{theorem}[Best iterate convergence under Lipschitz assumptions]\label{thm:lipschitzbest} Let $z \in {\mathcal{H}}$, let $z^\ast$ be a fixed point of $T_{\mathrm{PRS}}$, and let $x^\ast = \mathbf{prox}_{\gamma g}(z^\ast)$. Suppose that $\underline{\tau} = \inf_{j \geq 0} \lambda_j(1-\lambda_j) > 0$, and let $\underline{\lambda} = \inf_{j \geq 0}\lambda_j$. If $\nabla f$ (respectively $\nabla g$) is $({1}/{\beta})$-Lipschitz, and $x^k = x_g^k$ (respectively $x^k = x_f^k$), then \begin{align*} \min_{i = 0, \cdots, k}\left\{f(x^{i}) + g(x^{i}) - f(x^\ast) - g(x^\ast)\right\} = o\left(\frac{1}{k+1}\right). \end{align*} \end{theorem} \begin{proof} Fact~\ref{fact:averagedconvergence} proves the following bound: \begin{align*} \inf_{j \geq 0} \frac{1-\lambda_j}{\lambda_j}\sum_{i=0}^\infty \|z^k - z^{k+1}\|^2 \leq \sum_{i=0}^\infty \tau_i\|T_{\mathrm{PRS}} z^i - z^i\|^2 \leq \|z^0 - z^\ast\|^2. \end{align*} Therefore, the proof follows from Part~\ref{lem:sumsequence:part:nonmono} of Lemma~\ref{lem:sumsequence} applied to the summable upper bound in Proposition~\ref{prop:DRSupperfglipschitz}, which bounds the objective error. Note that under different choices of $(\lambda_j)_{j \geq 0}$ and $\gamma$, we get the bounds: \begin{align*} &f(x^{k_{\best}}) + g(x^{k_{\best}}) - f(x^\ast) - g(x^\ast) \\ &\leq \frac{\|z^0 - z^\ast\|^2}{4\gamma\underline{\lambda}(k+1)} \times \begin{cases} 1, & \mbox{if } \gamma \leq \beta \mbox{ and } (\lambda_j)_{j \geq 0} \subseteq \left[ \underline{\lambda}, \frac{1}{2}\left(1- \frac{\gamma}{\beta}\right)\right];\\ 1 + 1/\left(\inf_{j \geq 0} \frac{1-\lambda_j}{\lambda_j}\right), & \mbox{if } \gamma \leq \beta; \\ \left(1 + \frac{(\gamma - \beta)}{2\beta}\right)\left(1 + 1/\left(\inf_{j \geq 0} \frac{1-\lambda_j}{\lambda_j}\right)\right), & \mbox{otherwise.} \end{cases} \end{align*} \qed\end{proof} This result should be compared with the known convergence properties of the FBS algorithm, which has order $o(1/(k+1))$ for a bounded $\gamma$, but may even fail to converge if $\gamma$ is too large. See Section~\ref{section:FBSpractical} for more on the distinction between FBS and relaxed PRS. \subsection{Constant relaxation and better rates}\label{sec:constantrelaxationandbetterates} In this section, we study the convergence rate of DRS under the assumption \begin{assump} The function $g$ is differentiable on ${\mathcal{H}}$, the gradient $\nabla g$ is $({1}/{\beta})$-Lipschitz, and the sequence of relaxation parameters $(\lambda_j)_{j \geq 0}$ is constant and equal to ${1}/{2}$. \end{assump} With these assumptions, we will show that for a special choice of $\theta^\ast$ (Lemma~\ref{lem:howtochoosea}) and for $\gamma$ small enough, the following sequence is monotonic and summable (Propositions~\ref{prop:lipschitzmono} and~\ref{prop:lipschitzsum}): \begin{align}\label{eq:monoandsum} \left(2\gamma\left(f(x_f^j) + g(x_f^j) - f(x) - g(x)\right) +\theta^\ast\gamma^2 \| \nabla g(x_g^{j+1}) - \nabla g(x_g^j)\|^2+ \frac{(1-\theta^\ast)\gamma^2}{\beta^2}\|x_g^{j+1} - x_g^{j}\|^2\right)_{j \geq 0}. \end{align} We then use Fact~\ref{lem:sumsequence} to deduce $f(x_f^j) + g(x_f^j) - f(x) - g(x) = o(1/(k+1))$. There are several other simpler monotonic and summable sequences that dominate the objective error. For example, if we choose $\theta^\ast = 1$, we can drop the last term in Equation~\eqref{eq:monoandsum}, but we can no longer use this sequence to help deduce the convergence rate of the FPR in Theorem~\ref{thm:differentiableFPR}. Thus, we choose to analyze the slightly complicated sequence in Equation~\eqref{eq:monoandsum} in order to provide a unified analysis for all results in this section. We are now ready to deduce the objective error convergence rate for the DRS algorithm when $\nabla g$ is Lipschitz. Our bounds show that $$\mbox{DRS is at least as fast as FBS whenever $\gamma$ is small enough.}$$ Additionally, we show that the convergence rate of the best iterate has essentially the same constant for a large range of $\gamma$. When $\gamma$ is large, the best iterate still enjoys the convergence rate $o(1/(k+1))$, albeit with a larger constant (Theorem~\ref{thm:lipschitzbest}). The rates we derive are the best possible for this algorithm, as shown by \cite[Theorem 12]{davis2014convergence}. Because each step of the relaxed PRS algorithm is generated by a proximal operator, it may seem strange that the choice of stepsize $\gamma$ affects the convergence rate of relaxed PRS. This is certainly not the case for the proximal point algorithm, which achieves an $o(1/(k+1))$ convergence rate by Fact~\ref{lem:sumsequence}. A possible explanation is that the reflection operator of a differentiable function is the composition of averaged operators \begin{align*} \mathbf{refl}_{\gamma g} = (I - \gamma \nabla g) \circ \mathbf{prox}_{\gamma g} \end{align*} whenever $\gamma < 2\beta$, and, therefore, it is averaged \cite[Propositions 4.32 and~4.33]{bauschke2011convex}. Thus, although $T_{\mathrm{PRS}}$ is not necessarily averaged when $f$ or $g$ is differentiable, the individual reflection operators enjoy a stronger contraction property \cite[Proposition 4.25]{bauschke2011convex} as long as $\gamma$ is small enough. As soon as $\gamma$ is too large, we seem to lose monotonicity of various sequences that arise in our analysis. \begin{theorem}[Differentiable function convergence rate]\label{thm:differentiableobjective} Let $\rho \approx 2.2056$ be the positive real root of $x^3 - 2x^2 - 1$. Then \begin{align*} \min_{i=0, \cdots, k} \{f(x_f^{i}) &+ g(x_f^{i}) - f(x^\ast) - g(x^\ast)\} \\ &\leq \frac{1}{2\gamma(k+1)} \begin{cases} \|x_g^0 - x^\ast\|^2, & \mbox{if $\gamma < \rho\beta$};\\ \|x_g^0 - x^\ast\|^2 + \frac{1}{\beta^2+ \gamma^2}\left( \frac{\gamma^3}{\beta} - 2\gamma \beta - \beta^2\right) \|z^0 - z^\ast\|^2, & \mbox{otherwise}; \end{cases} \end{align*} and $\min_{i=0, \cdots, k} \{f(x_f^{i}) + g(x_f^{i}) - f(x^\ast) - g(x^\ast)\} = o\left(1/(k+1)\right).$ Furthermore, if $\kappa$ ($\approx 1.24698$) is the positive root of $x^3 + x^2 -2x - 1$, and $\gamma < \kappa \beta$, then \begin{align*} f(x_f^k) + g(x_f^k) - f(x^\ast) - g(x^\ast) \leq \frac{\|x_g^0 - x^\ast\|^2}{2\gamma(k+1)} && \mathrm{and} && f(x_f^k) + g(x_f^k) - f(x^\ast) - g(x^\ast) = o\left(\frac{1}{k+1}\right). \end{align*} \end{theorem} \begin{proof} To prove the ``$k_{\best}$" bounds, rearrange the upper inequality in Equation~\eqref{prop:fundamentaldiff:eq:main} to \begin{align*} 2\gamma&(f(x_f^k) + g(x_f^k) - f(x^\ast) - g(x^\ast)) \\ &\leq \|x_g^k - x^\ast\| - \|x_g^{k+1} - x^\ast\|^2 + \left(\frac{\gamma^3}{\beta} - 2\gamma \beta \right)\|\nabla g(x_g^{k+1}) - \nabla g(x_g^k)\|^2 - \|x_g^k - x_{g}^{k+1}\|^2 \\ &\leq \|x_g^k - x^\ast\| - \|x_g^{k+1} - x^\ast\|^2 + \left(\frac{\gamma^3}{\beta} - 2\gamma \beta - \beta^2\right)\|\nabla g(x_g^{k+1}) - \nabla g(x_g^k)\|^2, \numberthis \label{thm:differentiableobjective:bound1} \end{align*} where the last line follows from the bound $-\|x_g^k - x_{g}^{k+1}\|^2 \leq -\beta^2 \|\nabla g(x_g^{k}) - \nabla g(x_g^{k+1})\|^2.$ Note that ${\gamma^3}/{\beta} - 2\gamma \beta - \beta^2 \leq 0$ if, and only if, $\gamma \leq \rho\beta$ where $\rho$ is the positive root of $x^3 - 2x^2 - 1$. Therefore, the result follows by summing Equation~\eqref{thm:differentiableobjective:bound1} and applying~Fact~\ref{lem:sumsequence}. If $\gamma \leq \kappa\beta$, then $( {\gamma^3}/{\beta} - 2\gamma \beta + \theta^\ast\gamma^2) \leq 0$ and $(1-\theta^\ast)\gamma^2/\beta^2 \leq 1$. Therefore, Equation~\eqref{prop:lipschitzmono:eq:main} shows that the sequence \begin{align*} \left(2\gamma(f(x_f^j) + g(x_f^j) - f(x) - g(x)) +\theta^\ast\gamma^2 \| \nabla g(x_g^{j+1}) - \nabla g(x_g^j)\|^2+ \frac{(1-\theta^\ast)\gamma^2}{\beta^2}\|x_g^{j+1} - x_g^{j}\|^2\right)_{j \geq 0} \end{align*} is monotonic. In addition, Equation~\eqref{prop:lipschitzsum:eq:main} shows the sum of this sequence is bounded by $\|x_g^0 - x^\ast\|^2$. Therefore, the result follows by~Fact~\ref{lem:sumsequence}. \qed\end{proof} It was recently shown that the FPR convergence rate for the FBS algorithm is $o(1/(k+1)^2))$ \cite[Theorem 3]{davis2014convergence}. We complement this result by showing the same is true for DRS whenever $\gamma$ is small enough. This rate is optimal by \cite[Theorem 12]{davis2014convergence}. \begin{theorem}[Differentiable function FPR rate]\label{thm:differentiableFPR} Suppose that $\gamma < \kappa \beta$ where $\kappa$ ($\approx 1.24698$) is the positive root of $x^3 + x^2 -2x - 1$. Then for all $k \geq 1$, we have \begin{align}\label{eq:differentiableFPR} \|z^k - z^{k+1}\|^2 \leq \frac{\beta^2\|x_g^0 - x^\ast\|^2}{k^2\left(1+{\gamma}/{\beta}\right)^2\left(\beta^2 -{\gamma^2}/{\kappa^2}\right)} && \mathrm{and} && \|z^k - z^{k+1}\|^2 &= o\left(\frac{1}{k^2}\right). \end{align} \end{theorem} \begin{proof} For all $k \geq 1$, let $$\eta = 1- \frac{(1-\theta^\ast)\gamma^2}{\beta^2} = \frac{\beta^2 +(\theta^\ast-1) \gamma^2}{\beta^2}= \frac{\beta^2 - \gamma^2/\kappa^2}{\beta^2}, $$ let $a_{k-1} = (\eta/(1+\gamma/\beta)^2)\|z^{k+1} - z^k\|^2$, and let \begin{align*} b_{k-1} = 2\gamma(f(x_f^k) + g(x_f^k) - f(x^\ast) - g(x^\ast)) + \theta^\ast\gamma^2 \| \nabla g(x_g^{k+1}) - \nabla g(x_g^k)\|^2 +\frac{(1-\theta^\ast)\gamma^2}{\beta^2}\|x_g^{k+1} - x_g^{k} \|^2. \end{align*} Because $z^{k} = x_g^k + \gamma \nabla g(x_g^k)$ and and $\nabla g$ is $(1/\beta)$-Lipschitz, we get \begin{align*} \eta\|z^k - z^{k+1}\|^2 &\leq \eta\left(1+\frac{\gamma}{\beta}\right)^2\|x_g^k - x_g^{k+1}\|^2. \end{align*} Therefore, Equation~\eqref{prop:lipschitzmono:eq:main} shows that for all $k \geq 1$, \begin{align*} a_{k-1} \leq \eta \|x_g^k - x_g^{k+1}\|^2 \leq b_{k-1} - b_{k}. \end{align*} Fact~\ref{lem:sumsequence} applied to the sequences $(a_j)_{j \geq 0}$ and $(b_j)_{j \geq 0}$ with weighting parameters $\lambda_k \equiv 1$, (not to be confused with the constant relaxation parameter of the relaxed PRS algorithm), yields \begin{align*} \sum_{i=0}^\infty (i+1)a_i \leq \sum_{i=0}^\infty b_{i} \stackrel{\eqref{prop:lipschitzsum:eq:main}}{\leq} \|x_g^0 - x^\ast\|^2. \end{align*} \cite[Part 2 of Theorem 1]{davis2014convergence} shows that $(a_j)_{j \geq 0}$ is monotonic. Therefore, the result follows from Fact~\ref{lem:sumsequence}. \qed\end{proof} \begin{remark} Note that the FBS algorithm achieves $o(1/(k+1))$ objective error rate and $o(1/(k+1)^2)$ FPR rate as long as $\gamma < 2\beta$ \cite[Theorem 3]{davis2014convergence}. For the DRS algorithm, our analysis only covers the smaller range $\gamma \leq \kappa \beta$. It is an open question whether $\kappa$ can be improved for the DRS algorithm. \end{remark} \section{Strong convexity}\label{sec:generalstrongconvexity} The following theorem will deduce the convergence of $S_f(x_f^k, x^\ast)$ and $S(x_g^k, x^\ast)$ (see Equation~\eqref{eq:snotation}). In particular, if either $f$ or $g$ is strongly convex and the sequence $(\lambda_j)_{j \geq 0}\subseteq (0, 1]$ is bounded away from zero, then $x_f^k$ and $x_g^k$ converge strongly to a minimizer of $f + g$. Equation~\eqref{prop:sumauxilliaryterms:eq:main} is the main inequality needed to deduce linear convergence of the relaxed PRS algorithm (Section~\ref{sec:linearconvergence}), and it will reappear several times. \begin{theorem}[Auxiliary term bound]\label{prop:sumauxilliaryterms} Suppose that $(z^j)_{j \geq 0}$ is generated by Algorithm~\ref{alg:DRS}. Then for all $k \geq 0$, \begin{align}\label{prop:sumauxilliaryterms:eq:main} 8\gamma\lambda_k(S_{f}(x_f^k, x^\ast) + S_g(x_g^k, x^\ast)) &\leq \|z^{k} - z^\ast\|^2 - \|z^{k+1} - z^\ast\|^2 + \left(1 - \frac{1}{\lambda_k}\right)\|z^{k+1} - z^k\|^2. \end{align} Therefore, $8\gamma\sum_{i=0}^\infty \lambda_k(S_{f}(x_f^i, x^\ast) + S_g(x_g^i, x^\ast)) \leq \|z^0 - z^\ast\|^2$, and \begin{enumerate} \item \label{prop:sumauxilliaryterms:part:1} \textbf{Best iterate convergence:} If $\underline{\lambda} := \inf_{j \geq 0} \lambda_j > 0$, then $\min_{i=0, \cdots, k}\left\{S_{f}(x_f^{i}, x^\ast)\right\} = o\left(1/(k+1)\right)$ and $\min_{i=0, \cdots, k}\left\{S_{g}(x_g^{i}, x^\ast)\right\} = o\left(1/(k+1)\right).$ \item \label{prop:sumauxilliaryterms:part:2} \textbf{Ergodic convergence:} Let $\overline{x}_f^k = (1/\Lambda_k) \sum_{i=0}^k \lambda_ix_f^i$ and $\overline{x}_g^k = (1/\Lambda_k) \sum_{i=0}^k\lambda_i x_g^i$. Then \begin{align*} \overline{S}_{f}(x_f^k, x^\ast)+\overline{S}_{g}(x_g^k, x^\ast)\leq \frac{\|z^0 - z^\ast\|^2}{8 \gamma \Lambda_k} \end{align*} where $ \overline{S}_{f}(x_f^k, x^\ast):=\max\bigg\{\frac{\mu_f}{2}\left\|\overline{x}_f^k - x^\ast\right\|^2,~ \frac{\beta_f}{2}\bigg\|\frac{1}{\Lambda_k}\sum_{i=0}^k \widetilde{\nabla} f(x_f^k) - \widetilde{\nabla} f(x^\ast)\bigg\|^2\bigg\} $ and $\overline{S}_{g}(x_g^k, x^\ast)$ is similarly defined. \item \label{prop:sumauxilliaryterms:part:3} \textbf{Nonergodic convergence:} If $\underline{\tau} = \inf_{j \geq 0} \lambda_j(1-\lambda_j)> 0$, then $S_{f}(x_f^{k}, x^\ast) + S_g(x_g^{k}, x^\ast) = o\left(1/\sqrt{k+1}\right). $ \end{enumerate} \end{theorem} \begin{proof} By assumption, the relaxation parameters satisfy $\lambda_k \leq 1$. Therefore, Equation~\eqref{prop:sumauxilliaryterms:eq:main} is a consequence of the following inequalities: \begin{align*} 8\gamma\lambda_k(S_{f}(x_f^k, x^\ast) + S_g(x_g^k, x^\ast)) &\stackrel{\eqref{prop:DRSlower:eq:main},\eqref{eq:DRSmainidentity2}}{\leq} 4\gamma\lambda_k(f(x_f^k) + g(x_g^k) - f(x^\ast) - g(x^\ast) + S_{f}(x_f^k, x^\ast) + S_g(x_g^k, x^\ast)) \\ &\hspace{30pt}- 2\dotp{z^{k} - z^{k+1}, z^\ast - x^\ast} \\ &\stackrel{\eqref{prop:DRSupper:eq:aux}}{\leq} \|z^{k} - z^\ast\|^2 - \|z^{k+1} - z^\ast\|^2 + \left(1 - \frac{1}{\lambda_k}\right)\|z^{k+1} - z^k\|^2 \\ &\leq \|z^{k} - z^\ast\|^2 - \|z^{k+1} - z^\ast\|^2. \numberthis \label{eq:upperboundonstrongterms} \end{align*} Note that the sum of Equation~\eqref{eq:upperboundonstrongterms} over all $k$ is indeed bounded by $\|z^0 - z^\ast\|^2$. Thus, Part~\ref{prop:sumauxilliaryterms:part:1} follows from Fact~\ref{lem:sumsequence}, and Part~\ref{prop:sumauxilliaryterms:part:2} follows from Jensen's inequality applied to $\|\cdot \|^2.$ Fix $k \geq 0$, let $z_\lambda = (T_{\mathrm{PRS}})_{\lambda}z^k$ for $\lambda \in [0, 1]$, and note that $z_{\lambda} - z^k = \lambda( T_{\mathrm{PRS}} z^k - z^k)$. Fact~\ref{fact:averagedconvergence} shows that $\|z_{\lambda} - z^\ast\| \leq \|z^k - z^\ast\|$ (Equation~\eqref{eq:fejer}) and that the sequence $(\|z^j - z^\ast\|)_{j \geq 0}$ is nonincreasing. Therefore, Part~\ref{prop:sumauxilliaryterms:part:3} is a consequence of the cosine rule, Fact~\ref{fact:averagedconvergence}, Equation~\eqref{prop:sumauxilliaryterms:eq:main}, and the following inequalities: \begin{align*} S_{f}(x_f^{k}, x^\ast) + S_g(x_g^{k}, x^\ast) &\stackrel{\eqref{prop:sumauxilliaryterms:eq:main}}{\leq} \inf_{\lambda \in [0, 1]} \frac{1}{8\gamma \lambda}\left( \|z^{k} - z^\ast\|^2 - \|z_\lambda - z^\ast\|^2 + \left(1 - \frac{1}{\lambda}\right)\|z^{k} - z_\lambda\|^2\right)\\ &\stackrel{\eqref{eq:cosinerule}}{=} \inf_{\lambda \in [0,1]}\frac{1}{8\gamma \lambda}\left(2\dotp{z_\lambda - z^\ast, z^k - z_\lambda} +2\left(1 - \frac{1}{2\lambda}\right) \|z_\lambda - z^k\|^2\right) \\ &\leq \frac{\|z_{1/2} - z^\ast\|\|z^k - z_{1/2}\|}{2\gamma} \leq \frac{\|z^0 - z^\ast\|\|z^k - z_{1/2}\|}{2\gamma} \stackrel{\eqref{cor:DRSaveragedconvergence:eq:main}}{\leq} \frac{\|z^0 - z^\ast\|^2}{4\gamma \sqrt{\underline{\tau}(k+1)}}. \end{align*} The little-$o$ convergence rate follows because $S_{f}(x_f^{k}, x^\ast) + S_g(x_g^{k}, x^\ast)$ is bounded by a multiple of the square root of the FPR. \qed\end{proof} It is not clear whether the ``best iterate" convergence results of Theorem~\ref{prop:sumauxilliaryterms} can be improved to a convergence rate for the entire sequence because the values $S_f(x_f^k, x)$ and $S_g(x_g^k, x)$ are not necessarily monotonic. \section{Technical results from Section~\ref{sec:lipschitzderivatives:part:general}}\label{app:lipschitzderivatives:part:general} The following Theorem will be used several times throughout our analysis. \begin{theorem}[Descent theorem/Baillon-Haddad]\label{thm:descent} Suppose that $g : {\mathcal{H}} \rightarrow (-\infty, \infty]$ is closed, proper, convex, and differentiable. If $\nabla g$ is $({1}/{\beta})$-Lipschitz, then for all $x, y \in {\mathcal{H}}$, we have the upper bound \begin{align}\label{eq:lipschitzderivative} g(x) &\leq g(y) + \dotp{ x- y, \nabla g(y)} + \frac{1}{2\beta} \|x - y\|^2, \end{align} and the cocoercive inequality \begin{align}\label{eq:baillon} \beta\|\nabla g(x) - \nabla g(y)\|^2 &\leq \dotp{x - y, \nabla g(x) - \nabla g(y)}. \end{align} \end{theorem} \begin{proof} See \cite[Theorem 18.15(iii)]{bauschke2011convex} for Equation~\eqref{eq:lipschitzderivative}, and~\cite{baillon1977quelques} for Equation~\eqref{eq:baillon}. \qed\end{proof} \begin{proof}[Proof of Proposition~\ref{prop:DRSupperfglipschitz}] Because $\nabla f$ is $(1/\beta)$-Lipschitz, we have \begin{align*} f(x_g) &\stackrel{\eqref{eq:lipschitzderivative}}{\leq} f(x_f) + \dotp{x_g - x_f, \nabla f(x_f)} + \frac{1}{2\beta}\|x_g - x_f\|^2,\numberthis \label{eq:fxgupperbound}\\ S_f(x_f,x^*)&\stackrel{\eqref{eq:strongconvexandlipschitzlowerbound}}{\ge} \frac{\beta}{2}\|\nabla f(x_f) - \nabla f(x^\ast)\|^2\numberthis \label{eq:Sflowerbound}. \end{align*} We now derive some identities that will be used below to bound $f(x_g) + g(x_g) - f(x^*) - g(x^*)$. By applying the identity $z^\ast - x^\ast = \gamma\widetilde{\nabla} g(x^\ast) = - \gamma\nabla f(x^\ast)$ (Equation~\eqref{eq:gradoptimality}), the cosine rule \eqref{eq:cosinerule}, and Equation \eqref{eq:DRSmainidentity2} multiple times, we have \begin{align*} 2\dotp{z - z^{+}, z^\ast - x^\ast} &+4\gamma \lambda \dotp{x_g - x_f, \nabla f(x_f)} = 4\gamma \lambda \dotp{x_g - x_f, \nabla f(x_f)-\nabla f(x^\ast)} \\ &= 4\lambda \dotp{ \gamma \widetilde{\nabla} g(x_g) +\gamma \nabla f(x_f), \gamma \nabla f(x_f)-\gamma \nabla f(x^\ast)} \\ &= 2\lambda \left( \|x_f - x_g\|^2 + \|\gamma\nabla f(x_f) - \gamma\nabla f(x^\ast)\|^2-\|\gamma \widetilde{\nabla} g(x_g) - \gamma \widetilde{\nabla} g(x^\ast) \|^2\right). \end{align*} By Equation \eqref{eq:DRSmainidentity2} $ \left(1-\frac{1}{\lambda}\right) \|z - z^+\|^2 + 2\lambda\left(\frac{\gamma}{\beta} + 1\right) \|x_g - x_f\|^2 = \left(1 + \frac{ \left( \gamma - \beta\right)}{2\beta\lambda}\right)\|z - z^+\|^2. $ Using the above two identities, we have \begin{align*} &4\gamma\lambda\big(f(x_g) + g(x_g) - f(x^\ast) - g(x^\ast)\big)\\ &\stackrel{\eqref{eq:fxgupperbound}}{\le}4\gamma\lambda\big(f(x_f) + g(x_g) - f(x^\ast) - g(x^\ast)\big)+ 4\gamma\lambda\dotp{x_g - x_f, \nabla f(x_f)} + \frac{2\gamma\lambda}{\beta}\|x_g - x_f\|^2\\ &\stackrel{\eqref{prop:DRSupper:eq:aux}}{\le} \|z - z^*\|^2 - \|z^+ - z^*\|^2 +\big(2\dotp{z-z^+,z^*-x^*}+ 4\gamma\lambda\dotp{x_g - x_f, \nabla f(x_f)}\big) + \left(1 - \frac{1}{\lambda}\right)\|z^{+} - z\|^2\\ & \hspace{20pt}+\frac{2\gamma\lambda}{\beta}\|x_g - x_f\|^2-4\gamma\lambda S_f(x_f,x^*) \\ &= \|z - z^*\|^2 - \|z^+ - z^*\|^2 + \left(\left(1-\frac{1}{\lambda}\right) \|z - z^+\|^2 + 2\lambda\left(\frac{\gamma}{\beta} + 1\right) \|x_g - x_f\|^2 \right) \\ &\hspace{20pt}+ 2\lambda\|\gamma\nabla f(x_f) - \gamma\nabla f(x^\ast)\|^2-4\gamma\lambda S_f(x_f,x^*) -2\lambda \|\gamma \widetilde{\nabla} g(x_g) - \gamma \widetilde{\nabla} g(x^\ast)\|^2\\ &\stackrel{\eqref{eq:Sflowerbound}}{\le}\|z - z^*\|^2 - \|z^+ - z^*\|^2 +\left(1 + \frac{ \left( \gamma - \beta\right)}{2\beta\lambda}\right)\|z - z^+\|^2+2\gamma\lambda(\gamma-\beta)\|\nabla f(x_f) - \nabla f(x^\ast)\|^2. \end{align*} If $\gamma \leq \beta$, we can drop the last term. If $\gamma > \beta$, we apply the upper bound on $S_f(x_f, x)$ in \eqref{prop:sumauxilliaryterms:eq:main} to get \begin{align*} 2\gamma\lambda\left( \gamma - \beta\right)\|\nabla f(x_f) - \nabla f(x^\ast)\|^2 &\leq \frac{ \left( \gamma - \beta\right)}{2\beta}\left(\|z - z^\ast\|^2 - \|z^+ -z^\ast\|^2 + \left(1 - \frac{1}{\lambda}\right) \|z - z^+\|^2\right), \end{align*} and the result follows. If $\nabla g$ is $({1}/{\beta})$-Lipschitz, the argument is symmetric, so we omit the proof. \qed\end{proof} \section{Proofs from Section~\ref{sec:constantrelaxationandbetterates}}\label{app:constantrelaxationandbetterates} The following two results are well known, but we include some of the proofs for completeness. They will help us tighten the bounds that we develop below. \begin{lemma}[Extra contraction of derivative operator]\label{lem:extracontraction} Suppose that $\nabla g$ is $({1}/{\beta})$-Lipschitz, and let $x, y \in {\mathcal{H}}$. If $x^+ = \mathbf{prox}_{\gamma g}(x)$ and $y^+ = \mathbf{prox}_{\gamma f}(y)$, then \begin{align}\label{eq:lipschitzcontraction} \|\nabla g(x^+) - \nabla g(y^+)\|^2 \leq \frac{1}{\gamma^2 + \beta^2} \|x - y\|^2. \end{align} \end{lemma} \begin{proof} From the identity $\gamma \nabla g(x^+) = x - x^+$, the contraction property in Proposition~\ref{prop:basicprox}, and the Lipschitz continuity of $\nabla g$ we have \begin{align*} \beta^2\|\nabla g(x^+) - \nabla g(y^+)\|^2 \leq \|x^+ - y^+\|^2 && \mathrm{and} && \gamma^2 \|\nabla g(x^+) - \nabla g(y^+)\|^2 \leq \|x - y\|^2 - \|x^+ - y^+\|^2 \end{align*} Adding both equations and rearranging proves the result. \qed\end{proof} The following is a direct corollary of the descent theorem (Theorem~\ref{thm:descent}). \begin{corollary}[Joint descent theorem]\label{cor:jointdescent} If $g$ is differentiable and $\nabla g$ is $({1}/{\beta})$-Lipschitz, then for all pairs $x, y \in \dom(f)$, points $z {\mathcal{H}}$, and subgradients $\widetilde{\nabla} f(x) \in \partial f(x)$, we have \begin{align}\label{cor:jointdescent:eq:main} f(x) + g(x) &\leq f(y) + g(y) + \dotp{ x- y, \nabla g(z) + \widetilde{\nabla} f(x)} + \frac{1}{2\beta} \|z - x\|^2. \end{align} \end{corollary} \begin{proof} Inequality \eqref{cor:jointdescent:eq:main} follows from adding the upper bound \begin{align*} g(x) - g(y) &\stackrel{\eqref{eq:lipschitzderivative}}{\leq} g(z) - g(y) + \dotp{ x- z, \nabla g(z)} + \frac{1}{2\beta}\|z - x\|^2 \leq \dotp{x - y, \nabla g(z)} + \frac{1}{2\beta}\|z - x\|^2, \end{align*} with the subgradient inequality: \begin{equation} f(x) \le f(y) + \dotp{x-y,\widetilde{\nabla} f(x)}.\label{eq:subineq} \end{equation} \qed \end{proof} The following theorem develops an alternative fundamental inequality to the one in Proposition~\ref{prop:DRSupperfglipschitz}. \begin{proposition}[Fundamental inequality for differentiable functions]\label{prop:fundamentaldiff} For all $x \in \dom(f)$, \begin{align*} 2\gamma&(f(x_f^k) + g(x_f^k) - f(x) - g(x)) + \left(2\gamma \beta - \frac{\gamma^3}{\beta}\right)\| \nabla g(x_g^{k+1}) - \nabla g(x_g^k)\|^2 + \|x_g^{k+1} - x\|^2 + \|x_g^{k+1} - x_g^{k} \|^2 \\ &\leq \|x_g^k - x\|^2. \numberthis \label{prop:fundamentaldiff:eq:main} \end{align*} \end{proposition} \begin{proof} The following identities are straightforward from Lemma~\ref{prop:DRSmainidentity}: \begin{align*} x_{g}^{k} - x_{g}^{k+1} = \gamma(\nabla g(x_g^{k+1}) + \widetilde{\nabla} f(x_f^k) ) && \mathrm{and} && x_{f}^k - x_{g}^{k+1} = \gamma(\nabla g(x_g^{k+1}) - \nabla g(x_g^{k})). \numberthis \label{prop:fundamentaldiff:mainidentities} \end{align*} Therefore, \begin{align*} 2\gamma&(f(x_f^k) + g(x_f^k) - f(x) - g(x)) + \left(2\gamma \beta - \frac{\gamma^3}{\beta}\right)\| \nabla g(x_g^{k+1}) - \nabla g(x_g^k)\|^2 \\ &\stackrel{\eqref{cor:jointdescent:eq:main}}{\leq} 2\gamma\dotp{x_f^k - x, \widetilde{\nabla} f(x_f^k) + \nabla g(x_g^{k+1})} + \frac{\gamma}{\beta}\|x_f^{k} - x_g^{k+1}\|^2 + \left(2\gamma \beta - \frac{\gamma^3}{\beta}\right)\| \nabla g(x_g^{k+1}) - \nabla g(x_g^k)\|^2 \\ &\stackrel{\eqref{prop:fundamentaldiff:mainidentities}}{=}2\dotp{x_f^k - x, x_g^k - x_g^{k+1}} + 2\gamma \beta\| \nabla g(x_g^{k+1}) - \nabla g(x_g^k)\|^2 \\ &= 2\dotp{x_g^{k+1} - x, x_g^k - x_g^{k+1}} + 2\dotp{x_f^k - x_{g}^{k+1}, x_g^k - x_g^{k+1}} + 2\gamma \beta\| \nabla g(x_g^{k+1}) - \nabla g(x_g^k)\|^2 \\ &\stackrel{\eqref{eq:cosinerule}}{\leq} \|x_g^{k} - x\|^2 - \|x_g^{k+1} - x\|^2 - \|x_g^{k} - x_g^{k+1}\|^2 \\ &\stackrel{\eqref{prop:fundamentaldiff:mainidentities}}{+} 2\gamma\dotp{\nabla g(x_g^{k+1}) - \nabla g(x_g^{k}), x_g^k - x_g^{k+1}} + 2\gamma \beta\| \nabla g(x_g^{k+1}) - \nabla g(x_g^k)\|^2\\ &\stackrel{\eqref{eq:baillon}}{\leq} \|x_g^{k} - x\|^2 - \|x_g^{k+1} - x\|^2 - \|x_g^{k} - x_g^{k+1}\|^2. \numberthis \label{eq:before32} \end{align*} Equation~\eqref{prop:fundamentaldiff:eq:main} now follows by rearranging Equation~\eqref{eq:before32}. \qed\end{proof} The following proposition uses the fundamental inequality in Proposition~\ref{prop:fundamentaldiff} evaluated at the point $x = x_f^{k-1}$ to construct a monotonic sequence that dominates the objective error. We introduce a factor $\theta \in [0, 1]$ that we will optimize in Lemma~\ref{lem:howtochoosea} in order to maximize the range of $\gamma$ for which the sequence remains monotonic. \begin{proposition}[Monotonicity]\label{prop:lipschitzmono} For scalars $\theta \in [0, 1]$ and integers $k \geq 1$, the following bound holds: \begin{align*} 2\gamma&(f(x_f^k) + g(x_f^k) - f(x^\ast) - g(x^\ast)) + \left(2\gamma \beta - \frac{\gamma^3}{\beta}\right) \| \nabla g(x_g^{k+1}) - \nabla g(x_g^k)\|^2 + \|x_g^{k+1} - x_g^{k} \|^2 \\ &\leq 2\gamma(f(x_f^{k-1}) + g(x_f^{k-1}) - f(x^\ast) - g(x^\ast)) + \theta\gamma^2\|\nabla g(x_g^k) - \nabla g(x_g^{k-1})\|^2 + \frac{(1-\theta)\gamma^2}{\beta^2}\|x_g^k - x_g^{k-1}\|^2 \numberthis \label{prop:lipschitzmono:eq:main}. \end{align*} \end{proposition} \begin{proof} Plug $x = x_{f}^{k-1}$ into Equation~\eqref{prop:fundamentaldiff:eq:main} and subtract $f(x^\ast) + g(x^\ast)$ from both sides. Equation~\eqref{prop:lipschitzmono:eq:main} follows from the identity \begin{align*} x_g^k - x_{f}^{k-1} = \gamma (\nabla g(x_g^{k-1}) - \nabla g(x_g^{k})), \end{align*} the bound $\|\nabla g(x_g^k) - \nabla g(x_g^{k-1})\|^2 \leq ({1}/{\beta^2} )\|x_g^k - x_g^{k-1}\|^2$, rearranging, and dropping the positive term $\|x_g^{k+1} - x_f^{k-1}\|^2$. \qed\end{proof} We now choose the factor $\theta$ in order to maximize the range of implicit stepsize parameters $\gamma$ for which the sequence constructed in Proposition~\ref{prop:lipschitzmono} remains monotonic. \begin{lemma}[Maximizing $\gamma$ range]\label{lem:howtochoosea} Let $\beta > 0$, and let \begin{align}\label{eq:kappa} \kappa := \sup\left\{ \frac{\gamma}{\beta} \mid \gamma > 0, \theta \in [0, 1], \theta\gamma^2\leq \left(2\gamma \beta - \frac{\gamma^3}{\beta}\right) , \; \frac{(1-\theta)\gamma^2}{\beta^2} \leq 1 \right\}. \end{align} Then $\kappa$ is the positive root of $x^3 +x^2 - 2x- 1$. Therefore, $(\gamma^\ast, \theta^\ast) = (\kappa\beta, 1 - 1/\kappa^2)$. \end{lemma} \begin{proof} Observe that the constraints on $\theta$ and $\gamma$ are equivalent to following inequalities: \begin{align*} 1+ \frac{2\gamma}{\beta} - \frac{\gamma^2}{\beta^2} - \frac{\gamma^3}{\beta^3} \geq (\theta - 1)\frac{\gamma^2}{\beta^2} +1 \geq 0. \numberthis\label{eq:thetarearrange} \end{align*} The left hand side of Equation~\eqref{eq:thetarearrange} is monotonically decreasing in $\gamma$ for all $\gamma \geq \beta$. Furthermore, if $\gamma = \kappa \beta$, then the left hand side is $0$. Thus, $\gamma^\ast \leq \kappa \beta$. Finally, for every $\gamma \in [0, \kappa \beta]$, the scalar $\theta_\gamma = 1 - \beta^2/\gamma^2$ satisfies $(\theta - 1)(\gamma^2/\beta^2) +1 \geq 0$. Therefore, $(\gamma^\ast, \theta^\ast) = (\kappa\beta, 1 - 1/\kappa^2)$. \qed\end{proof} \begin{remark}\label{rem:kappa} Throughout the rest of the paper, we will let $\kappa = 1/\sqrt{1-\theta^\ast} \approx 1.24698$ where $\theta^\ast$ is defined in Lemma~\ref{lem:howtochoosea}. Note that the inequality constraints in Equation~\eqref{eq:kappa} become equalities for the pair $(\gamma^\ast, \theta^\ast)$. \end{remark} We will need the following bound in several of the proofs below. \begin{prop}[Gradient sum bound] For all $\gamma > 0$ \begin{align}\label{eq:Lipschitzgradientsum} \sum_{i=0}^\infty \|\nabla g(x_g^i) - \nabla g(x_g^{i+1})\|^2 & \leq \frac{1}{\gamma^2 + \beta^2}\|z^0 - z^\ast\|^2. \end{align} \end{prop} \begin{proof} From Lemma~\ref{lem:extracontraction} and the Fej\'er type in equality in Equation~\eqref{eq:fejer}: \begin{align*} \|\nabla g(x_g^k) - \nabla g(x_g^{k+1})\|^2 &\leq ({1}/({\gamma^2 + \beta^2}))\|z^k - z^{k+1}\|^2 \\ &\leq ({1}/({\gamma^2 + \beta^2}))\left(\|z^k - z^\ast\|^2 - \|z^{k+1} - z^\ast\|^2\right). \numberthis\label{prop:lipschitzsum:eq:finalalt} \end{align*} Therefore, the result follows by summing~\eqref{prop:lipschitzsum:eq:finalalt}. \qed\end{proof} The following proposition computes an upper bound of the sum of the sequence in Equation~\eqref{prop:lipschitzmono:eq:main}. \begin{proposition}[Summability]\label{prop:lipschitzsum} If $\gamma < \kappa \beta$, choose $\theta = \theta^\ast$ as in Lemma~\ref{lem:howtochoosea}; otherwise, set $\theta = 1$. Then \begin{align*} \sum_{i=0}^\infty&\left( 2\gamma(f(x_f^k) + g(x_f^k) - f(x^\ast) - g(x^\ast)) + \theta\gamma^2 \| \nabla g(x_g^{k+1}) - \nabla g(x_g^k)\|^2 +\frac{(1-\theta)\gamma^2}{\beta^2}\|x_g^{k+1} - x_g^{k} \|^2\right) \\ &\leq \begin{cases} \|x_g^0 - x^\ast\|^2, & \mbox{if $\gamma < \kappa \beta$};\\ \|x_g^0 - x^\ast\|^2 + \frac{1}{\beta^2+ \gamma^2}\left( \frac{\gamma^3}{\beta} - 2\gamma \beta + \gamma^2 - \beta^2\right) \|z^0 - z^\ast\|^2, & \mbox{otherwise.} \end{cases} \numberthis \label{prop:lipschitzsum:eq:main} \end{align*} \end{proposition} \begin{proof} First note that: $$-\|x_g^k - x_{g}^{k+1}\|^2 \leq -\beta^2 \|\nabla g(x_g^{k}) - \nabla g(x_g^{k+1})\|^2.$$ In addition, for either choice of $\theta$ we have $ (1-\theta)\gamma^2/\beta^2- 1 \leq 0$. Thus, from Equation~\eqref{prop:fundamentaldiff:eq:main} \begin{align*} 2\gamma&(f(x_f^k) + g(x_f^k) - f(x^\ast) - g(x^\ast)) + \theta\gamma^2 \| \nabla g(x_g^{k+1}) - \nabla g(x_g^k)\|^2 +\frac{(1-\theta)\gamma^2}{\beta^2}\|x_g^{k+1} - x_g^{k} \|^2 \\ &\leq \|x_g^k - x^\ast\| - \|x_g^{k+1} - x^\ast\|^2 \\ &+ \left(\frac{\gamma^3}{\beta} - 2\gamma \beta + \theta\gamma^2\right)\|\nabla g(x_g^{k+1}) - \nabla g(x_g^k)\|^2 + \left(\frac{(1-\theta)\gamma^2}{\beta^2}- 1\right)\|x_g^k - x_{g}^{k+1}\|^2 \numberthis \label{prop:lipschitzsum:eq:alternating}\\ &\leq \|x_g^k - x^\ast\| - \|x_g^{k+1} - x^\ast\|^2 + \left(\frac{\gamma^3}{\beta} - 2\gamma \beta + \gamma^2 - \beta^2\right)\|\nabla g(x_g^{k+1}) - \nabla g(x_g^k)\|^2. \end{align*} The last line of Equation~\eqref{prop:lipschitzsum:eq:alternating} is negative if, and only if, $\gamma \leq \kappa \beta$. This proves the first bound in Equation~\eqref{prop:lipschitzsum:eq:main}. The second bound follows from the sum bound in Equation~\eqref{eq:Lipschitzgradientsum}. \qed\end{proof} \section{Proofs from Section~\ref{sec:feasibilityniceintersection}}\label{app:feasibilityniceintersection} In this section, we will vary the implicit stepsize parameter in every iteration. In addition $f$ and $g$ will have separate implicit stepsize parameters. Thus, we augment the $T_{\mathrm{PRS}}$ notation as follows: for all $\gamma_f, \gamma_g > 0$, \begin{align*} T_{\mathrm{PRS}}^{\gamma_f, \gamma_g} &:= \mathbf{refl}_{\gamma_f f} \circ \mathbf{refl}_{\gamma_g g}. \end{align*} The following optimality conditions are well known. They will be needed in Section~\ref{sec:feasibilityniceintersection} because we vary the implicit stepsize parameter $\gamma$. See \cite{davis2014convergence,bauschke2011convex} for a proof. \begin{lemma}[Optimality conditions of $T_{\mathrm{PRS}} $]\label{lem:PRSoptimality} The set of zeros of $\partial f+ \partial g$ is precisely \begin{align} \zer(\partial f + \partial g) &= \{ \mathbf{prox}_{\gamma g}(z) \mid z \in {\mathcal{H}}, T_{\mathrm{PRS}} z = z\}.\numberthis \label{eq:setoptimalitydrs} \end{align} That is, if $z^\ast$ is a fixed point of $T_{\mathrm{PRS}}$, then $x^\ast = x_g^\ast = x_f^\ast$ is a solution to Problem~\ref{eq:simplesplit}, and \begin{align}\label{eq:gradoptimality} z^\ast - x^\ast = \gamma \widetilde{\nabla} g(x^\ast) \in \gamma \partial g(x^\ast). \end{align} Therefore, the set of fixed points of $T_{\mathrm{PRS}}$ is exactly \begin{align*} \left\{ x + \gamma w\mid x \in \zer \left( \partial f + \partial g\right), w\in (-\partial f(x)) \cap \partial g(x)\right\}. \end{align*} \end{lemma} The following propositions study the behavior of $T_{\mathrm{PRS}}^{\gamma_f, \gamma_g}$ as the positive implicit stepsize parameters $\gamma_f $ and $\gamma_g$ vary. \begin{lemma}[Non expansiveness of PRS operator] The operator $T_{\mathrm{PRS}}^{\gamma_f, \gamma_g}$ is nonexpansive. \end{lemma} \begin{proof} This is an immediate consequence of the nonexpansiveness of the reflection mapping (See Part~\ref{prop:basicprox:part:nonexpansive} of Proposition~\ref{prop:basicprox}). \qed\end{proof} The following lemma will be useful for determining the fixed point set of $T_{\mathrm{PRS}}^{\gamma_f, \gamma_g}$. \begin{lemma}[Minimizers of weighted squared distance]\label{lem:differentgammadist} Let $\rho_1, \rho_2 > 0$, and suppose that $C_f \cap C_g \neq \emptyset$. Then the set of minimizers of $\rho_1 d^2_{C_f} + \rho_2 d^2_{C_g}$ is $C_f \cap C_g$. \end{lemma} \begin{proof} The minimal value is attained whenever $x \in C_f \cap C_g$; otherwise, the sum is nonzero. \qed\end{proof} We will now compute the fixed points of $T_{\mathrm{PRS}}^{\gamma_f, \gamma_g}$. \begin{proposition}[Fixed points of PRS operator] The set of fixed points of $T_{\mathrm{PRS}}^{\gamma_f, \gamma_g}$ is $C_f \cap C_g$. \end{proposition} \begin{proof} Let $f' = \gamma_f f$ and let $g' = \gamma_g g$. Then Lemma~\ref{lem:PRSoptimality} combined with Lemma~\ref{lem:differentgammadist} show that the set of fixed points of $T_{\mathrm{PRS}}^{\gamma_f, \gamma_g} = \mathbf{refl}_{f'} \circ \mathbf{refl}_{g'}$ is \begin{align*} \{x + \gamma \nabla g'(x) \mid x \in C_f \cap C_g, \nabla g'(x) = -\nabla f'(x)\}. \end{align*} However, $\nabla g'(x) = 2\gamma_g(x - P_{C_g}(x))= 0$ for all $x \in C_f \cap C_g$, and so the identity holds. \qed\end{proof} Now, we will show that the sequence generated by Equation~\eqref{eq:DRSfeasibilitygamma} is bounded. \begin{proposition}[Boundedness] Suppose that $(z^k)$ is generated by the iteration in Equation~\eqref{eq:DRSfeasibilitygamma}. If $(\lambda_k)_{k \geq 0} \subseteq (0, 1]$, then $(\|z^j - x\|^2)_{j \geq 0}$ is monotonically nonincreasing for any $x \in C_f \cap C_g$. \end{proposition} \begin{proof} Because the set of fixed points of $T_{\mathrm{PRS}}^{\gamma_f, \gamma_g}$ does not depend on $\gamma_f$ and $\gamma_g$, the claim follows directly from the Fej\'er-type inequality in Equation~\eqref{eq:fejer}. \qed\end{proof} We restate the fundamental inequality here for the readers convenience. \begin{proposition}[Upper fundamental inequality for feasibility problem] Suppose that $z\in {\mathcal{H}}$ and $z^+ = (T_{\mathrm{PRS}}^{\gamma_f, \gamma_g})_{\lambda}(z)$. Then for all $x^\ast\in C_f\cap C_g$, \begin{align} 8\lambda(\gamma_{f} d^2_{C_f}(x_f) + \gamma_{g}d^2_{C_g}(x_g)) &\leq \|z - x^\ast\|^2 - \|z^{+} - x^\ast\|^2 + \left(1 - \frac{1}{\lambda} \right)\|z^{+} - z\|^2. \end{align} \end{proposition} \begin{proof}[Proof of Proposition~\ref{prop:factsaboutdistancesquared}] This follows directly from the upper fundamental inequality in Proposition~\ref{prop:DRSupper} (with $\mu_f = \mu_g = 0$, and $\gamma = 1$), applied to the functions $f' = \gamma_f f$ and $g' = \gamma_g g$. Indeed, the gradients $\gamma_{f}\nabla d_{C_f}^2$ and $\gamma_{g}\nabla d_{C_g}^2$ are $2\gamma_{f}$ and $2\gamma_{g}$-Lipschitz ($\beta_{f'} = {1}/({2\gamma_{f}})$ and $\beta_{g'} = {1}/({2\gamma_{g}})$). Furthermore, if $S_{g'}$ and $S_{f'}$ are defined as in Equation~\eqref{eq:snotation}, then \begin{align} S_{g'}(x_{g'}, x^\ast) = \frac{1}{4\gamma_g} \|\gamma_{g}\nabla d_{C_g}^2(x_g) - \gamma_{g}\nabla d_{C_g}^2(x^\ast)\|^2 =\frac{\gamma_{g}}{4}\|2(x_g - P_{C_g}(x_g))\|^2 = \gamma_{g}d_{C_g}^2(x_g), \end{align} and by the same argument, $S_{f'}(x_{f'}, x^\ast) = \gamma_{f} d_{C_f}^2(x_f)$. To summarize, we have \begin{align} \gamma_{f} d^2_{C_f}(x_f) + \gamma_{g}d^2_{C_g}(x_g) + S_{f'}(x_{f'}, x^\ast) + S_{g'}(x_{g'}, x^\ast) &= 2\gamma_{f}d^2_{C_f}(x_f) + 2\gamma_{g}d^2_{C_g}(x_g). \end{align} Therefore, the inequality follows because $d_{C_g}^2(x^\ast) = d_{C_f}^2(x^\ast) = 0$. \qed\end{proof} \section{Extension of results of Section~\ref{sec:feasibilityniceintersection} to multiple sets}~\label{app:feasibilitymultiplesets} The concept of (bounded) linear regularity is defined for any finite number of sets. The following theorem shows that (bounded) linear regularity of a collection of sets is equivalent to the (bounded) linear regularity of a certain pair of sets in a product space. For convenience we set \begin{align} D &:= \{(x, \cdots, x) \mid x \in {\mathcal{H}}\} \subseteq {\mathcal{H}}^m, \end{align} and endow ${\mathcal{H}}^m$ with the canonical norm: $\|(x_1, \cdots, x_m)\|^2 = ({1}/{m})\sum_{i=1}^m \|x_i\|^2$. We will use the boldface notation ${\mathbf{x}} \in {\mathcal{H}}^m$ for an arbitrary vector in ${\mathcal{H}}^m$. Finally, for any ${\mathbf{x}} \in {\mathcal{H}}^m$, we will write ${\mathbf{x}}_{j}$ for the $j$th component of ${\mathbf{x}}$, which is an element of ${\mathcal{H}}$. \begin{theorem}[(Bounded) linear regularity in product spaces] \label{thm:boundedlinearlyregularproduct} Suppose that $C_1, \cdots, C_m$ are closed convex subsets of ${\mathcal{H}}$ with nonempty intersection. Then $\{C_1, \cdots, C_m\}$ is boundedly linearly regular or linearly regular, if, and only if, $\{C_1 \times \cdots \times C_m, D\}$ has the same property in ${\mathcal{H}}^m$ with the canonical norm. In particular, if $\{C_1, \cdots, C_m\}$ is $\mu_\rho$-(boundedly) linearly regular on the ball $B(0, \rho)$, then $\{C_1 \times \cdots \times C_n, D\}$ is $\sqrt{(1+4m\mu_\rho^2)}$-(boundedly) on the ball $B(\mathbf{0}, \rho)$, and \begin{align} d_{(C_1\times \cdots \times C_m) \cap D}({\mathbf{x}}) &\leq \sqrt{(1+ 4m\mu_\rho^2)}\max\{d_{C_1\times \cdots \times C_m}({\mathbf{x}}), d_{D}({\mathbf{x}})\}. \end{align} \end{theorem} \begin{proof} See \cite[Theorem 3.12]{deutsch2008rate}. \qed\end{proof} In this section we model the feasibility problem of the $m$ sets $\{C_{1}, \cdots, C_m\}$ using the following two objective functions on the product space ${\mathcal{H}}^m$: \begin{align*} f(x_1, \cdots, x_m) = \sum_{i=1}^m d_{C_i}^2(x_i) && \mathrm{and} && g(x_1, \cdots, x_n) = d_{D}^2(x_1, \cdots, x_n). \end{align*} In the space ${\mathcal{H}}^m$, the proximal operators of $f$ and $g$ have the following form: \begin{align*} \mathbf{prox}_{\gamma f} ({\mathbf{x}}) = \left(\frac{1}{2\gamma + 1} x_i + \frac{2\gamma}{2\gamma + 1} P_{C_j}x_j\right)_{j = 1}^m && \mathrm{and} && \mathbf{prox}_{\gamma g}({\mathbf{x}}) = \left(\frac{1}{2\gamma + 1}x_j + \frac{2\gamma }{(2\gamma + 1)m}\sum_{i=1}^m x_i\right)_{j=1}^m. \end{align*} We apply the iteration in Equation~\eqref{eq:DRSfeasibilitygamma} with these identities to get the following parallel algorithm: given implicit stepsize parameters $(\gamma_{f, j})_{ j \geq 0}$ and $(\gamma_{g, j})_{j \geq 0}$, relaxation parameters $(\lambda_j)_{j \geq 0} \subseteq (0, 1]$, and an initial point ${\mathbf{z}}^0 \in {\mathcal{H}}^m$, for all $k \geq 0$, define \begin{align*} &\overline{z}^k = \frac{1}{m}\sum_{i=1}^m z_i^k;\\ &\begin{cases} x_{g,i}^k = ({1}/({2\gamma_{g, k} + 1}))z^k_i+ ({2\gamma_{g, k}}/({2\gamma_{g, k} + 1}))\overline{z}^k; & \mbox{For} \\ x_{f,i}^k = ({1}/({2\gamma_{f, k}+1}))(2x_{g, i}^k - z_i^k) + ({2\gamma_{f, k}}/({2\gamma_{f, k} + 1})) P_{C_i}(2x_{g, i}^k - z_i^k); & \boxed{i = 1, \cdots, m} \\ z_{i}^{k+1} = z_i^k + 2\lambda_k (x_{f, i}^k - x_{g, i}^k); & \mbox{in parallel.} \end{cases} \numberthis \label{eq:PPSdistance} \end{align*} Note that the algorithm in Equation~\eqref{eq:PPSdistance} is related to the general algorithm in \cite[Section 8.3]{bauschke1996projectionthesis}. One of the main differences between these two algorithms is that the projection operators are not necessarily evaluated at same point in each iteration ($(2{\mathbf{x}}_g^k - {\mathbf{z}}^k) \notin D$). By changing the metric of the underlying space, e.g. to $\|(x_1, \cdots, x_m)\|^2 = \sum_{i=1}^m w_i \|x_i\|^2$ where $w_i > 0$ are arbitrary weights, we can perform a weighted average of all the projections. In addition, we can assign each set $C_i$ a different implicit stepsize parameter at each iteration. For simplicity we do not pursue these extensions here. The following theorem deduces the linear convergence of the iteration in Equation~\eqref{eq:PPSdistance}. \begin{theorem}[Linear convergence: Feasibility for multiple sets]\label{thm:PPSfeasibility} Suppose that $({\mathbf{z}}^j)_{j \geq 0}$ is generated by the iteration in Equation~\eqref{eq:PPSdistance}, and suppose that $\{C_1, \cdots, C_m\}$ is (boundedly) linearly regular. Let $\rho > 0$ and $\mu_\rho > 0$ be such that $({\mathbf{z}}^j)_{ j \geq 0} \subseteq B(\mathbf{0}, \rho)$ and the inequality \begin{align} d_{C_1 \cap \cdots \cap C_m}(x) &\leq \mu_\rho \max\{d_{C_1}(x),\cdots, d_{C_m}(x)\} \end{align} holds for all $x \in B(0, \rho)$. Then $({\mathbf{z}}^j)_{j \geq 0}$ satisfies the following relation: for all $k \geq 0$, \begin{align} d_{(C_1\times \cdots \times C_m) \cap D}({\mathbf{z}}^{k+1}) &\leq C(\gamma_{f, k}, \gamma_{g, k}, \lambda_k, \mu_\rho) \times d_{(C_1\times \cdots \times C_m) \cap D}({\mathbf{z}}^k) \end{align} where \begin{align*} C(\gamma_{f, k}, \gamma_{g, k}, \lambda_k, \mu_\rho) &:= \left(1-\frac{4\lambda_k\min\{{\gamma_{g, k}}/{(2\gamma_{g, k} + 1)^2}, {\gamma_{f, k}}/{(2\gamma_{f, k} + 1)^2}\}}{(1+4m\mu_\rho^2)\max\{{16\gamma_{g, k}^2}/{(2\gamma_{g, k}+1)^2}, 1\}} \right)^{{1}/{2}}. \end{align*} In particular, if $\overline{C} = \sup_{j \geq 0}C(\gamma_{f, k}, \gamma_{g, k}, \lambda_k, \mu_\rho) < 1$, then $({\mathbf{z}}^j)_{j \geq 0}$ converges linearly to a point in $(C_1\times \cdots \times C_m) \cap D$ with rate $\overline{C}$, and \begin{align} \|{\mathbf{z}}^k - {\mathbf{z}}^\ast\| &\leq 2d_{(C_1\times \cdots \times C_m) \cap D}({\mathbf{z}}^0) \prod_{i=0}^k C(\gamma_{f, i}, \gamma_{g, i}, \lambda_i, \mu_\rho). \end{align} \end{theorem} \begin{proof} This theorem is a direct corollary of Theorem~\ref{thm:linearfeasibility} except that Theorem~\ref{thm:boundedlinearlyregularproduct} is used to calculate the (bounded) linear regularity constant. \qed\end{proof} Finally we derive the following analogue of Corollary~\ref{cor:AP}. \begin{corollary}[Convergence of MAP: Multiple sets]\label{cor:APm} Let $({\mathbf{z}}^j)_{j \geq 0}$ be generated by the iteration in Equation~\eqref{eq:PPSdistance} with $\gamma_{f, k} \equiv \gamma_{g, k} \equiv {1}/{2}$ and $\lambda_k \equiv1$. Define $x^k :=(P_D{\mathbf{z}}^k)_1$. Then for all $k \geq 0$, \begin{align}\label{eq:AAP} x^{k+1} = \frac{1}{m}\sum_{i=1}^m P_{C_i}(x^k). \end{align} Thus, Averaged MAP is a special case of PRS. Consequently, under the assumptions of Theorem~\ref{thm:PPSfeasibility}, $x^k$ converges linearly to a point in the intersection $C_{1} \cap \cdots \cap C_m$ with rate $\left(1 - {1}/({1+4m\mu^2})\right)^{{1}/{2}}$. \end{corollary} \begin{proof} Equation~\eqref{eq:AAP} follows because $\mathbf{refl}_{\gamma g} = P_D$ and $\mathbf{refl}_{\gamma f} = P_{C_1\times \cdots \times C_m}$. In addition, by the nonexpansiveness of $P_{D}$ we have \begin{align*} \|x^k - z^\ast\|^2 = \|(P_D{\mathbf{z}}^k)_{1} - {\mathbf{z}}_1^\ast\|^2 = \frac{1}{m}\sum_{i=1}^k \|(P_D{\mathbf{z}}^k)_i - {\mathbf{z}}_i^\ast\|^2 \leq \|{\mathbf{z}}^k - {\mathbf{z}}^\ast\|^2. \end{align*} By Corollary~\ref{cor:AP} and Theorem~\ref{thm:boundedlinearlyregularproduct}, the sequence $({\mathbf{z}}^j)_{j \geq 0}$ converges linearly with rate $\left(1 - {1}/({1+4m\mu^2})\right)^{{1}/{2}}$. Thus, the rate for $(x^k)_{k \geq 0}$ follows from the rate for $({\mathbf{z}}^j)_{j \geq 0}$. \qed\end{proof} \section{Consequences of linear convergence of ADMM}\label{app:linearconvergenceimplies} The following proposition is a translation of Proposition~\ref{prop:linearconvergenceimplies} to the ADMM setting. \begin{proposition}[Consequences of linear convergence of ADMM]\label{prop:linearconvergenceimpliesADMM} Let $(C_j)_{j \geq0} \subseteq [0, 1]$ be a positive scalar sequence, and suppose that for all $k \geq0$, \begin{align}\label{ADMM:prop:linearconvergenceimplies:eq:main} \|z^{k+1} - z^\ast\| \leq C_k\|z^k - z^\ast\|. \end{align} Fix $k \geq 1$. Then \begin{align*} \|w_{d_g}^k - w^\ast\|^2 + \gamma^2\| By^k - By^\ast\|^2 \leq \|z^0 - z^\ast\|^2\prod_{i = 0}^{k-1} C_i^2; \\ \|w_{d_f}^k - w^\ast\|^2 + \gamma^2\|Ax^k - Ax^\ast\|^2 \leq \|z^0 - z^\ast\|^2\prod_{i = 0}^{k-1} C_i^2. \end{align*} If $\lambda <1$, then the FPR rate holds: $$\|(T_{\mathrm{PRS}})_{\lambda}z^k- z^k\| \leq \sqrt{\frac{\lambda}{1-\lambda}}\|z^0 - z^\ast\|\prod_{i = 0}^{k-1} C_i.$$ Consequently, the following convergence rates for constraint violations and objective errors hold:$$\|Ax^k + By^k - b\|^2 \leq \frac{\|z^0 - z^\ast\|^2}{\gamma^2}\prod_{i = 0}^{k-1} C_i^2,$$ and \begin{align*} \frac{-\|z^0 - z^\ast\|\|w^\ast\|}{\gamma}\prod_{i = 0}^{k-1} C_i \leq f(x^k) + g(y^k) - f(x^\ast) - g(y^\ast) \leq \frac{\left(\|z^0 - z^\ast\| + \|w^\ast\|\right)\|z^0 - z^\ast\|}{\gamma} \prod_{i = 0}^{k-1} C_i. \end{align*} \end{proposition} \begin{proof} The convergence rates for the dual variables, primal variables, and FPR follow from Proposition~\ref{prop:linearconvergenceimplies} and the identities in Table~\ref{table:ADMMsubgradients}. Now fix $k \geq 1$, and let $z_\lambda = (T_{\mathrm{PRS}})_{\lambda}z^k$ for all $\lambda \in [0,1]$. The convergence rate for the constraint violation follows from the identity $z_\lambda - z^k = -2\gamma \lambda( Ax^{k} + By^k - b)$ (Equation~\eqref{eq:ADMMfeasibilityFPR}) and the FPR convergence rate: \begin{align*} \|Ax^k + By^k - b\|^2 = \inf_{\lambda \in [0, 1]} \frac{\|z_\lambda - z^k\|^2}{4\gamma^2\lambda^2}= \inf_{\lambda \in [0, 1]}\frac{\|z^0 - z^\ast\|^2}{4\gamma^2\lambda(1-\lambda)}\prod_{i = 0}^{k-1} C_i^2 = \frac{\|z^0 - z^\ast\|^2}{\gamma^2}\prod_{i = 0}^{k-1} C_i^2 \end{align*} The lower bound on the objective error follows from the fundamental lower inequality in Proposition~\ref{prop:ADMMlower} and the constraint violations rate: \begin{align*} f(x^k) + g(y^k) - f(x^\ast) - g(y^\ast)\stackrel{\eqref{prop:ADMMlower:eq:main}}{\geq} \dotp{Ax^k + By^k - b, w^\ast} \geq -\|Ax^k + By^k - b\|\| w^\ast\| \geq \frac{-\|z^0 - z^\ast\|\|w^\ast\|}{\gamma}\prod_{i = 0}^{k-1} C_i \end{align*} The upper bound on the objective error follows from Proposition~\ref{prop:ADMMupper}, the FPR rate, the bound $\|z_{\lambda} - z^\ast\|^2 \leq \|z^k - z^\ast\|$ (Equation~\eqref{eq:fejer}), the monotonicity of the sequence $(\|z^{j} - z^\ast\|)_{j \geq 0}$ (Part~\ref{fact:averagedconvergence:eq:mono} of Fact~\ref{fact:averagedconvergence}), and the following inequalities: \begin{align*} f(x^k) + g(y^k) &- f(x^\ast) - g(y^\ast)\\ &\stackrel{\eqref{prop:ADMMupper:eq:main}}{\leq} \inf_{\lambda \in [0, 1]} \frac{1}{4\gamma\lambda}\left(\|z^k - (z^\ast - w^\ast)\|^2 - \|z_{\lambda} - (z^\ast - w^\ast)\|^2 + \left(1- \frac{1}{\lambda} \right) \|z^{k} - z_\lambda\|^2\right) \\\ &\stackrel{\eqref{eq:cosinerule}}{\leq}\inf_{\lambda \in [0, 1]} \frac{1}{4\gamma\lambda}\left(2\dotp{ z_\lambda - (z^\ast - w^\ast), z^k - z_\lambda} + 2\left(1- \frac{1}{2\lambda} \right) \|z^{k} - z_\lambda \|^2 \right) \\ &\leq \frac{\left(\|z_{1/2} - z^\ast\| + \|w^\ast\|\right)\|z_{1/2} - z^k\|}{\gamma} \\ &\leq \frac{\left(\|z^0 - z^\ast\| + \|w^\ast\|\right)\|z^0 - z^\ast\|}{\gamma } \prod_{i = 0}^{k-1} C_i. \end{align*} \qed\end{proof} \section{Applications to conic programming}\label{app:conicprogramming} In this section we borrow the setting of \cite{o2013operator}. The goal of linear (LP) and semidefinite (SDP) programming is to minimize a linear function subject to linear and matrix semidefinite constraints, respectively. Thus, in this section we study the following generic primal-dual pair problem \begin{align*} \Min_{x \in {\mathbf{R}}^n} &\; c^T x && & \Max_{y \in {\mathbf{R}}^m} &\; -b^Ty \\ \mbox{subject to} &\; Ax + s = b && &\mbox{subject to} &\; - A^T y + r = c \\ &\; (x, s) \in {\mathbf{R}}^n \times {\mathcal{K}} && && (r, y) \in \{0\}^n \times {\mathcal{K}}^\ast \numberthis\label{eq:conicprogramming} \end{align*} where $c \in {\mathbf{R}}^n$, $b, s \in {\mathbf{R}}^m$, $A : {\mathbf{R}}^n \rightarrow {\mathbf{R}}^m$ is a linear map, ${\mathcal{K}} \subseteq {\mathbf{R}}^m$ is a closed convex cone, and ${\mathcal{K}}^\ast \subseteq {\mathbf{R}}^m$ is the dual cone to ${\mathcal{K}}$. In linear programming ${\mathcal{K}}$, is the positive orthant ${\mathcal{K}} = {\mathbf{R}}^n_+$, and for semidefinite programming, ${\mathcal{K}}$ is the cone of symmetric, positive semidefinite matrices. In \cite{o2013operator}, both optimization problems in Equation~\eqref{eq:conicprogramming} are combined into a single feasibility problem. To this end we introduce slack variables $\tau, \kappa \in {\mathbf{R}}_+$, and the vectors and matrix \begin{align*} u =\begin{bmatrix} x \\ y \\ \tau \end{bmatrix} \in {\mathbf{R}}^{n + m + 1}, \quad v = \begin{bmatrix} r \\ s \\ \kappa\end{bmatrix} \in {\mathbf{R}}^{n + m + 1}, \quad Q = \begin{bmatrix} 0 & A^T & c \\ - A & 0 & b \\ -c^T &- b^T &0 \end{bmatrix} \in {\mathbf{R}}^{(n + m + 1)\times (n+ m + 1)}. \end{align*} In addition, we let ${\mathcal{C}} = {\mathbf{R}}^n \times {\mathcal{K}}^\ast \times {\mathbf{R}}_+$ and ${\mathcal{C}}^\ast = \{0\} \times {\mathcal{K}} \times {\mathbf{R}}_+$. With this notation the goal of the \emph{homogeneous self dual embedding} problem is to find $(u, v) \in {\mathbf{R}}^{n+m + 1}$ such that $Qu = v$ and $(u, v) \in {\mathcal{C}} \times {\mathcal{C}}^\ast$. Throughout this section we denote \begin{align} C_f = {\mathcal{C}} \times {\mathcal{C}}^\ast && \mathrm{and} && C_g = \{(u, v) \in {\mathbf{R}}^{n + m + 1} \times {\mathbf{R}}^{n+m+1} \mid Qu = v\}. \end{align} Our goal is to find a point in the intersection $C_f \cap C_g$. A remarkable trichotomy was derived in \cite{ye1994nl}: Suppose $(u, v) \in C_f \cap C_g$, then \begin{enumerate} \item If $\tau > 0$ and $\kappa = 0$, then $(x/\tau, y/\tau, s/ \tau)$ is a primal dual solution of \label{eq:conicprogramming}. \item If $\tau = 0$ and $\kappa > 0$, then $c^T x + b^Ty < 0$. The case $b^T y < 0$ is a certificate of primal infeasibility, and the case $c^Tx < 0$ is a certificate of dual in feasibility. \item If $\tau = \kappa = 0$, then nothing can be concluded about Equation~\eqref{eq:conicprogramming}. However, if there exists a point $(u', v') \in C_f \cap C_g$ for which $\tau' + \kappa' \neq 0$, then we can choose an initial point $z^0 \in {\mathbf{R}}^{n + m + 1}$ such that DRS applied with $f= \iota_{C_f}$ and $g = \iota_{C_g}$ converges to a point $(u', v') \in C_f\cap C_g$ with $\kappa' + \tau' \neq 0$~\cite{o2013operator}. \end{enumerate} \subsection{Linear programming}\label{sec:linearprogramming} Let us now examine the structure of the sets $C_f$ and $C_g$. For linear programming problems, $C_f = {\mathbf{R}}^n \times {\mathbf{R}}_+^m \times {\mathbf{R}}_+ \times \{0\} \times {\mathbf{R}}_+^m \times {\mathbf{R}}_+$ is a polyhedron, i.e. the intersection of finitely many half planes, and $C_g$ is a linear subspace. In finite dimensional spaces the pair $\{C_f, C_g\}$ is linearly regular in the sense of Definition~\ref{defi:linearregularity} \cite[Remark 5.7.3]{bauschke1996projectionthesis}. We have four different algorithms that we can apply to find a point in $C_f \cap C_g$. The first two are the non parallelized versions of DRS which correspond to function pairs \begin{align}\label{eq:linearprogpairnoparallel} (f = \iota_{C_f}, g = \iota_{C_g}) && \mathrm{and} && (f = d_{C_f}^2, g = d_{C_g}^2). \end{align} Theorem~\ref{thm:linearfeasibility} shows that relaxed PRS applied to the second pair (Equation~\eqref{eq:DRSfeasibilitygamma}) linearly convergence to a point in the intersection $C_f \cap C_g$. Linear convergence of DRS applied to the first pair was shown in \cite{bauschke2014linear}. The projection onto $C_f$ is simple, and so the main computational bottleneck of the algorithm is to project onto $C_g$. There are various tricks that can be employed to speed this step up \cite{o2013operator}, but in some cases it is desirable to break up the linear equations into several sets $C_g = C_{g_1} \cap \cdots \cap C_{g_r}$ where $C_{g_i} \subseteq {\mathbf{R}}^{n + m + 1}$ each encode a small number of linear constraints. The collection $\{C_f, C_{g_1}, \cdots, C_{g_r}\}$ is linearly regular by \cite[Remark 5.7.3]{bauschke1996projectionthesis}, so we can apply Theorem~\ref{thm:boundedlinearlyregularproduct} to show that $\{C_f \times C_{g_1} \times \cdots \times C_{g_r}, D\}$ is linearly regular where $D \subseteq {\mathbf{R}}^{(r+1)(n + m + 1)}$ is the ``diagonal set" of Appendix~\ref{app:feasibilitymultiplesets}. Thus, we can apply DRS or relaxed PRS to either of the following pairs: \begin{align}\label{eq:linearprogpairparallel} (f = \iota_{C_f\times C_{g_1} \times \cdots \times C_{g_r}}, g = \iota_{D}) && \mathrm{and} && (f = d_{C_f\times C_{g_1} \times \cdots \times C_{g_r}}^2, g = d_{D}^2). \end{align} We can deduce linear convergence of the first pair using \cite{bauschke2014linear} and of the second by Theorem~\ref{thm:PPSfeasibility}. In general, the pairs in Equation~\eqref{eq:linearprogpairnoparallel} and~\eqref{eq:linearprogpairparallel} may not perform the same in practice. Thus, we cannot make any prediction about the practical performances of the methods. We can only point to our arguments in Section~\ref{sec:feasibilitygeneralconvergence} that seem to indicate a better performance of the indicator function pair in problems that are badly conditioned. \subsection{Semidefinite programming} For semidefinite programming, ${\mathcal{K}}$ is the cone of positive semidefinite matrices. Note that ${\mathcal{K}}^\ast = {\mathcal{K}}$, i.e. ${\mathcal{K}}$ is self dual \cite[Example 6.25]{bauschke2011convex}. In general, the pair $\{C_f, C_g\}$ is not necessarily (boundedly) linearly regular. The main condition to check is whether the relative interior of $C_f$ intersects the subspace $C_g$ \cite[Theorem 5.6.2]{bauschke1996projectionthesis}. In fact, the relative interior of ${\mathcal{K}}$ in ${\mathbf{R}}^{m}$ is the set of all strictly positive definite matrices, i.e. the set of full rank positive definite matrices. Many problems of interest in semidefinite programming arise from the \emph{lifting} of a non convex problem and desire \emph{low rank} solutions of the associated SDP \cite{goemans1995improved}. Thus, we do not expect the relative interior of $C_f$ to intersect $C_g$ for every SDP. In terms of algorithm choice, we have at least four options to model the feasibility problem (See Equations~\eqref{eq:linearprogpairnoparallel} and~\eqref{eq:linearprogpairparallel}). In particular, when the linear constraints are difficult to solve in unison, we can break them into smaller pieces and solve them exactly. However, the main computational bottleneck of semidefinite programming is the projection onto the semidefinite cone. Unfortunately, there seems to be no way to lighten the cost of this projection. \cut{The convergence rates for relaxed PRS applied to the feasibility problem are linear whenever the relative interior of $C_f$ intersects $C_g$. In terms of ${\mathcal{K}}$ this condition requires that there is a full rank strictly positive definite primal dual pair $(x, y) \in {\mathcal{K}} \times {\mathcal{K}}$. Finally, because we usually do not expect full rank solutions to SDPs,}We refer the reader to Section~\ref{sec:feasibilitygeneralconvergence} and Equations~\eqref{eq:feasibilitybounddistancenonergodic} and~\eqref{eq:drsergodicdistancebound} which show the worst case feasibility convergence rates.