arXiv subjects
Publications and source records attributed to Konstantin Tikhomirov.
A well known conjecture of Burr and Erdos asserts that the Ramsey number $r(Q_n)$ of the hypercube $Q_n$ on $2^n$ vertices is of the order $O(2^n)$. In this paper, we show that $r(Q_n)=O(2^{2n-c n})$ for a universal constant $c>0$, improving upon the previous best known bound $r(Q_n)=O(2^{2n})$, due to Conlon, Fox and Sudakov.
Sharp conditions for the presence of spectral outliers are well understood for Wigner random matrices with iid entries. In the setting of inhomogeneous symmetric random matrices (i.e., matrices with a non-trivial variance profile), the corresponding problem has been considered only recently. Of special interest is the setting of sparse inhomogeneous matrices since sparsity is both a key feature and a technical obstacle in various aspects of random matrix theory. For such matrices, the largest of the variances of the entries has been used in the literature as a natural proxy for sparsity. We contribute sharp conditions in terms of this parameter for an inhomogeneous symmetric matrix with sub-Gaussian entries to have outliers. Our result implies a ``structural'' universality principle: the presence of outliers is only determined by the level of sparsity, rather than the detailed structure of the variance profile.
Determining the capacity $α_c$ of the Binary Perceptron is a long-standing problem. Krauth and Mezard (1989) conjectured an explicit value of $α_c$, approximately equal to .833, and a rigorous lower bound matching this prediction was recently established by Ding and Sun (2019). Regarding the upper bound, Kim and Roche (1998) and Talagrand (1999) independently showed that $α_c$ < .996, while Krauth and Mezard outlined an argument which can be used to show that $α_c$ < .847. The purpose of this expository note is to record a complete proof of the bound $α_c$ < .847. The proof is a conditional first moment method combined with known results on the spherical perceptron
Given a positive integer $n$, an unlabeled graph $G$ on $n$ vertices, and a vertex $v$ of $G$, let $N_G(v)$ be the subgraph of $G$ induced by vertices of $G$ of distance at most one from $v$. We show that there are universal constants $C,c>0$ with the following property. Let the sequence $(p_n)_{n=1}^\infty$ satisfy $n^{-1/2}\log^C n\leq p_n\leq c$. For each $n$, let $Γ_n$ be an unlabeled $G(n,p_n)$ Erdös-Rényi graph. Then with probability $1-o_n(1)$, any unlabeled graph $\tilde Γ_n$ on $n$ vertices with $\{N_{\tilde Γ_n}(v)\}_{v}=\{N_{Γ_n}(v)\}_{v}$ must coincide with $Γ_n$. This establishes $\tilde Θ(n^{-1/2})$ as the transition range for the density parameter $p_n$ between reconstructability and non-reconstructability of Erdös-Rényi graphs from their $1$-neighborhoods, and resolves a problem of Gaudio and Mossel.
Let $A$ be an $n\times n$ matrix with mutually independent centered Gaussian entries. Define \begin{align*} σ^*:=\max\limits_{i,j\leq n}\sqrt{{\mathbb E}\,|A_{i,j}|^2}, \quad σ:=\max\bigg(\max\limits_{j\leq n}\sqrt{{\mathbb E}\,\|{\rm col}_j(A)\|_2^2}, \max\limits_{i\leq n}\sqrt{{\mathbb E}\,\|{\rm row}_i(A)\|_2^2}\bigg). \end{align*} Assume that $σ\geq n^\varepsilon\,σ^*$ for a constant $\varepsilon>0$, and that a complex number $z$ satisfies $|z|=Ω(σ)$. We prove that $$ s_{\min}(A-z\,{\rm Id}) \geq |z|\,\exp\bigg(-n^{o(1)}\,\Big(\frac{\sqrt{n}\,σ^*}σ\Big)^2\bigg) $$ with probability $1-o(1)$. Without extra assumptions on $A$, the bound is optimal up to the $n^{o(1)}$ multiple in the power of exponent. We discuss applications of this estimate in context of empirical spectral distributions of inhomogeneous non-Hermitian random matrices.
The size-Ramsey number $\hat r(G')$ of a graph $G'$ is defined as the smallest integer $m$ so that there exists a graph $G$ with $m$ edges such that every $2$-coloring of the edges of $G$ contains a monochromatic copy of $G'$. Answering a question of Beck, Rodl and Szemeredi showed that for every $n\geq 1$ there exists a graph $G'$ on $n$ vertices each of degree at most three, with the size-Ramsey number at least $cn\log^{\frac{1}{60}}n$ for a universal constant $c>0$. In this note we show that a modification of Rodl and Szemeredi's construction leads to a bound $\hat r(G')\geq cn\,\exp(c\sqrt{\log n})$.
Let $n\geq 1$, $K>0$, and let $X=(X_1,X_2,\dots,X_n)$ be a random vector in $\mathbb{R}^n$ with independent $K$--subgaussian components. We show that for every $1$--Lipschitz convex function $f$ in $\mathbb{R}^n$ (the Lipschitzness with respect to the Euclidean metric), $$ \max\big(\mathbb{P}\big\{f(X)-{\rm Med}\,f(X)\geq t\big\},\mathbb{P}\big\{f(X)-{\rm Med}\,f(X)\leq -t\big\}\big)\leq \exp\bigg( -\frac{c\,t^2}{K^2\log\big(2+\frac{ n}{t^2/K^2}\big)}\bigg),\quad t>0, $$ where $c>0$ is a universal constant. The estimates are optimal in the sense that for every $n\geq \tilde C$ and $t>0$ there exist a product probability distribution $X$ in $\mathbb{R}^n$ with $K$--subgaussian components, and a $1$--Lipschitz convex function $f$, with $$ \mathbb{P}\big\{\big|f(X)-{\rm Med}\,f(X)\big|\geq t\big\}\geq \tilde c\,\exp\bigg( -\frac{\tilde C\,t^2}{K^2\log\big(2+\frac{n}{t^2/K^2}\big)}\bigg). $$ The obtained deviation estimates for subgaussian variables are in sharp contrast with the case of variables with bounded $\|X_i\|_{ψ_p}$--norms for $p\in[1,2)$.
For reversible Markov chains on finite state spaces, we show that the modified log-Sobolev inequality (MLSI) can be upgraded to a log-Sobolev inequality (LSI) at the surprisingly low cost of degrading the associated constant by $\log (1/p)$, where $p$ is the minimum non-zero transition probability. We illustrate this by providing the first log-Sobolev estimate for Zero-Range processes on arbitrary graphs. As another application, we determine the modified log-Sobolev constant of the Lamplighter chain on all bounded-degree graphs, and use it to provide negative answers to two open questions by Montenegro and Tetali (2006) and Hermon and Peres (2018). Our proof builds upon the `regularization trick' recently introduced by the last two authors.
This paper deals with the problem of graph matching or network alignment for Erdős--Rényi graphs, which can be viewed as a noisy average-case version of the graph isomorphism problem. Let $G$ and $G'$ be $G(n, p)$ Erdős--Rényi graphs marginally, identified with their adjacency matrices. Assume that $G$ and $G'$ are correlated such that $\mathbb{E}[G_{ij} G'_{ij}] = p(1-α)$. For a permutation $π$ representing a latent matching between the vertices of $G$ and $G'$, denote by $G^π$ the graph obtained from permuting the vertices of $G$ by $π$. Observing $G^π$ and $G'$, we aim to recover the matching $π$. In this work, we show that for every $\varepsilon \in (0,1]$, there is $n_0>0$ depending on $\varepsilon$ and absolute constants $α_0, R > 0$ with the following property. Let $n \ge n_0$, $(1+\varepsilon) \log n \le np \le n^{\frac{1}{R \log \log n}}$, and $0 < α< \min(α_0,\varepsilon/4)$. There is a polynomial-time algorithm $F$ such that $\mathbb{P}\{F(G^π,G')=π\}=1-o(1)$. This is the first polynomial-time algorithm that recovers the exact matching between vertices of correlated Erdős--Rényi graphs with constant correlation with high probability. The algorithm is based on comparison of partition trees associated with the graph vertices.
In this work, we develop a comparison procedure for the Modified log-Sobolev Inequality (MLSI) constants of two reversible Markov chains on a finite state space. Efficient comparison of the MLSI Dirichlet forms is a well known obstacle in the theory of Markov chains. We approach this problem by introducing a {\it regularized} MLSI constant which, under some assumptions, has the same order of magnitude as the usual MLSI constant yet is amenable for comparison and thus considerably simpler to estimate in certain cases. As an application of this general comparison procedure, we provide a sharp estimate of the MLSI constant of the switch chain on the the set of simple bipartite regular graphs of size $n$ with a fixed degree $d$. Our estimate implies that the total variation mixing time of the switch chain is of order $O_d(n\log n)$. The result is optimal up to a multiple depending on $d$ and resolves a long-standing open problem. We expect that the MLSI comparison technique implemented in this paper will find further applications.
The problem of estimating the smallest singular value of random square matrices is important in connection with matrix computations and analysis of the spectral distribution. In this survey, we consider recent developments in the study of quantitative invertibility in the non-Hermitian setting, and review some applications of this line of research.
Consider the switch chain on the set of $d$-regular bipartite graphs on $n$ vertices with $3\leq d\leq n^{c}$, for a small universal constant $c>0$. We prove that the chain satisfies a Poincaré inequality with a constant of order $O(nd)$; moreover, when $d$ is fixed, we establish a log-Sobolev inequality for the chain with a constant of order $O_d(n\log n)$. We show that both results are optimal. The Poincaré inequality implies that in the regime $3\leq d\leq n^c$ the mixing time of the switch chain is at most $O\big((nd)^2 \log(nd)\big)$, improving on the previously known bound $O\big((nd)^{13} \log(nd)\big)$ due to Kannan, Tetali and Vempala and $O\big(n^7d^{18} \log(nd)\big)$ obtained by Dyer et al. The log-Sobolev inequality that we establish for constant $d$ implies a bound $O(n\log^2 n)$ on the mixing time of the chain which, up to the $\log n$ factor, captures a conjectured optimal bound. Our proof strategy relies on building, for any fixed function on the set of $d$-regular bipartite simple graphs, an appropriate extension to a function on the set of multigraphs given by the configuration model. We then establish a comparison procedure with the well studied random transposition model in order to obtain the corresponding functional inequalities. While our method falls into a rich class of comparison techniques for Markov chains on different state spaces, the crucial feature of the method - dealing with chains with a large distortion between their stationary measures - is a novel addition to the theory.
Graph matching, also known as network alignment, refers to finding a bijection between the vertex sets of two given graphs so as to maximally align their edges. This fundamental computational problem arises frequently in multiple fields such as computer vision and biology. Recently, there has been a plethora of work studying efficient algorithms for graph matching under probabilistic models. In this work, we propose a new algorithm for graph matching: Our algorithm associates each vertex with a signature vector using a multistage procedure and then matches a pair of vertices from the two graphs if their signature vectors are close to each other. We show that, for two Erdős--Rényi graphs with edge correlation $1-α$, our algorithm recovers the underlying matching exactly with high probability when $α\le 1 / (\log \log n)^C$, where $n$ is the number of vertices in each graph and $C$ denotes a positive universal constant. This improves the condition $α\le 1 / (\log n)^C$ achieved in previous work.
We consider the problem of estimating small ball probabilities $\mathbb P\{f(G) \leqslant δ\mathbb Ef(G)\}$ for sub-additive,positively homogeneous functions $f$ with respect to the Gaussian measure. We establish estimates that depend on global parameters of the underlying function which take into account analytic and statistical measures, such as the variance and the $L^1$-norms of its partial derivatives. This leads to dimension-dependent bounds for small ball and lower small deviation estimates for seminorms when the linear structure is appropriately chosen to optimize the aforementioned parameters. Our bounds are best possible up to numerical constants. In all regimes, $\|G\|_\infty = \max_{ i \leqslant n}|g_i|$ arises as an extremal case in this study. The proofs exploit the convexity and hypercontractivity properties of the Gaussian measure.
We show that for an $n\times n$ random matrix $A$ with independent uniformly anti-concentrated entries, such that $\mathbb{E} ||A||^2_{HS}\leq K n^2$, the smallest singular value $σ_n(A)$ of $A$ satisfies $$ P\left( σ_n(A)\leq \frac{\varepsilon}{\sqrt{n}} \right) \leq C\varepsilon+2e^{-cn},\quad \varepsilon \ge 0. $$ This extends earlier results of Rudelson and Vershynin, and Rebrova and Tikhomirov by removing the assumption of mean zero and identical distribution of the entries across the matrix, as well as the recent result of Livshyts, where the matrix was required to have i.i.d. rows. Our model covers "inhomogeneus" matrices allowing different variances of the entries, as long as the sum of the second moments is of order $O(n^2)$. In the past advances, the assumption of i.i.d. rows was required due to lack of Littlewood--Offord--type inequalities for weighted sums of non-i.i.d. random variables. Here, we overcome this problem by introducing the Randomized Least Common Denominator (RLCD) which allows to study anti-concentration properties of weighted sums of independent but not identically distributed variables. We construct efficient nets on the sphere with lattice structure, and show that the lattice points typically have large RLCD. This allows us to derive strong anti-concentration properties for the distance between a fixed column of $A$ and the linear span of the remaining columns, and prove the main result.
In this paper, we study the distribution of the minimal distance (in the Hamming metric) of a random linear code of dimension $k$ in $\mathbb{F}_q^n$. We provide quantitative estimates showing that the distribution function of the minimal distance is close ({\it{}superpolynomially} in $n$)to the cumulative distribution function of the minimum of $(q^k-1)/(q-1)$ independent binomial random variables with parameters $\frac{1}{q}$ and $n$. The latter, in turn, converges to a Gumbel distribution at integer points when $\frac{k}{n}$ converges to a fixed number in $(0,1)$. Our result confirms in a strong sense that apart from identification of the weights of proportional codewords, the probabilistic dependencies introduced by the linear structure of the random code, produce a negligible effect on the minimal code weight. As a corollary of the main result, we obtain an improvement of the Gilbert--Varshamov bound for $2<q<49$.
Let $A$ be an $n\times n$ random matrix with independent rows $R_1(A),\dots,R_n(A)$, and assume that for any $i\leq n$ and any three-dimensional linear subspace $F\subset {\mathbb R}^n$ the orthogonal projection of $R_i(A)$ onto $F$ has distribution density $ρ(x):F\to{\mathbb R}_+$ satisfying $ρ(x)\leq C_1/\max(1,\|x\|_2^{2000})$ ($x\in F$) for some constant $C_1>0$. We show that for any fixed $n\times n$ real matrix $M$ we have $${\mathbb P}\{s_{\min}(A+M)\leq t n^{-1/2}\}\leq C'\, t,\quad\quad t>0,$$ where $C'>0$ is a universal constant. In particular, the above result holds if the rows of $A$ are independent centered log-concave random vectors with identity covariance matrices. Our method is free from any use of covering arguments, and is principally different from a standard approach involving a decomposition of the unit sphere and coverings, as well as an approach of Sankar-Spielman-Teng for non-centered Gaussian matrices.
Let $n,k\geq 1$ and let $G$ be the $n\times n$ random matrix with i.i.d. standard real Gaussian entries. We show that there are constants $c_k,C_k>0$ depending only on $k$ such that the smallest singular value of $G^k$ satisfies $$ c_k\,t\leq {\mathbb P}\big\{s_{\min}(G^k)\leq t^k\,n^{-1/2}\big\}\leq C_k\,t,\quad t\in(0,1], $$ and, furthermore, $$ c_k/t\leq {\mathbb P}\big\{\|G^{-k}\|_{HS}\geq t^k\,n^{1/2}\big\}\leq C_k/t,\quad t\in[1,\infty), $$ where $\|\cdot\|_{HS}$ denotes the Hilbert-Schmidt norm.