Title: What Could Push the Edge Far from the Bulk of the Spectrum of Random Matrices?
We survey here the various mechanisms we know that can break the well known fact that for a random matrix the top eigenvalues stick to the bulk. We will survey the influence of heavy tails, of finite rank perturbations and of sparseness of the entries. This will use results from joint works with Antonio Auffinger, Jinho Baik, Alice Guionnet and Sandrine Peche.
Title: Linear Statistics of Eigenvalues and Stein's Method
I will present an abstract central limit theorem with total variation error bounds for linear statistics of eigenvalues of random matrices. The technique, based on Stein's method, gives a unified tool for a large class of matrices. I will show applications to Wigner, Wishart, Double Wishart, and Toeplitz matrices. The CLT for Toeplitz matrices, as well as the error bounds for the other cases, are new results.
Title: The Rank of Random Symmetric Matrice
Let Q(n,p) be a random symmetric matrices whose entries above the main diagonal are independently either 1 (with probability p) or 0 (with probability 1-p). We will examine the following questions (whose answer will depend on p):
(1) What is the singularity probability of Q? (2) If Q is likely to be singular, how can we (almost surely) describe the dependent sets of rows?
Joint with Van Vu (Rutgers) and Terence Tao (UCLA).
Title: Random Matrix Theory: Origins, Theory, Applications and some Open Problems
The speaker will describe how RMT emerged as a tool in theoretical physics and in mathematics. He will also discuss some of the techniques that arise in the analysis of RMT statistics and illustrate the appplication of RMT with some examples. He will also describe a few open problems.
Title: From Random Matrices to Random Analytic Functions
Peres and Virag proved that the zeros of the power series a_0+za_1+z^2a_2+... with i.i.d. standard complex Gaussian coefficients, is a determinantal point process on the unit disk, invariant in distribution under isometries of the hyperbolic plane. Extending this result, I show that the singular points of the matrix-valued power series A_0+zA_1+z^2A_2+..., where A_i are k x k matrices with i.i.d. standard complex Gaussian entries, is also a determinantal process in the disk. This gives a unified framework in which to view the result of Peres and Virag and a well-known theorem of Ginibre on Gaussian random matrices.
Title: Fast Computation of Low Rank Matrix Approximations using Randomized Matrices
We study the problem of computing a matrix of low rank that approximates a given input matrix well. This can typically be dominated by many matrix-vector multiplications, whose cost grows linearly with the number of entries in the source matrix. We show that several randomized matrix simplifications, such as randomized sparsification and quantization, can substantially improve the running time of this computation without greatly impacting the quality of the low rank approximation. We present perturbation analysis bounding the influence that the introduced errors can have, and see empirically that the results are even better than the bounds imply, suggesting further work to more tightly characterize the affect of perturbations on low rank approximations.
Title: Random Matrices with Linear Structure
Traditionally the theory of random matrices has focused mainly on matrices whose entries are all independent random matrices, except possibly for a self-adjointness constraint. In a 1999 survey, Bai suggested studying the asymptotic spectral behavior of random matrices which lie in one of several smaller linear subspaces of the space of matrices, in particular Toeplitz, Hankel and Markov matrices. I will discuss recent results on these and other ensembles of random matrices with linear structure.
Title: Dyck Paths and Random Matrices: the Moment Method
We will speak about the moment approach in random matrix theory, developed by A. Soshnikov for instance to prove universality of the fluctuations of the largest eigenvalues of certain ensembles of random matrices. If Dyck paths are suitable objects to handle traces of Wigner matrices, the suitables tools to investigate traces of sample covariance matrices are "Narayana paths", that is Dyck paths with a certain number of odd up steps.
Title: Norm of the Inverse of a Random Matrix
Let $A$ be an $n \times n$ matrix with independent entries. It is well known that under appropriate moment assumptions, the norm of $A$ is of order $\sqrt{n}$ with high probability. It was long believed that the same holds for $A^{-1}$. This was only known for Gaussian matrices, and was conjectured for Bernoulli matrices (with independent $\pm 1$ entries). We prove this conjecture in full generality, for matrices with general independent entries, and under mild moment assumptions. Moreover, for any $\e > e^{-c n}$ we obtain a bound on the norm of $A^{-1}$, which holds with probability $1-\e$, and show that such bound is optimal.
We will also extend these results to rectangular random matrices. (Joint with R. Vershynin).
Title: Free Probability and Random Matrices
I will give an introduction to free probability theory, with emphasis on its relation to combinatorics (via non-crossing partitions) and random matrices.
Title: Graph Sparsification by Effective Resistances
A sparsifier of a graph is a sparse subgraph whose Laplacian quadratic form approximates the quadratic form of the original. Sparsifiers provide a natural notion of what it means to approximate a graph by a sparse subgraph, and may be used as preconditioners for solving systems of linear equations. We show that one can obtain a high-quality sparsifier of a graph by sampling every edge of the graph with probability proportional to its effective resistance. We then explain how these sparsifiers can be constructed in nearly-linear time.
This is joint work with Nikhil Srivastava.
I. Singularity and Determinant of Random Matrices II. Least Singular Value of Random Matrices III. Eigenvalue Distributions of Random Matrices
Title: I. Singularity and Determinant of Random Matrices
I. Consider a large random matrix M in which all entries are
identically and independently drawn from a discrete distribution (e.g.
uniformly chosen at random from {-1,+1}). How likely is M to be
singular (non-invertible)? As a related question, how is the
determinant det(M) of M distributed? We discuss some results of
Kahn-Komlos-Szemeredi, Tao-Vu, Costello-Tao-Vu, and others towards
understanding these questions. The techniques used are a combination
of Euclidean geometry, linear algebra, Fourier analysis, geometry of
numbers, and combinatorics. In particular, the _Littlewood-Offord
problem_ (how often does a discrete random walk return to the
origin?), or more precisely the _inverse Littlewood-Offord problem_
(if a random walk returns often to the origin, what does this tell us
about the walk?) plays a pivotal role.
Title: II. Least Singular Value of Random Matrices
II. In many applications, knowing that a random matrix M is
invertible is not enough; one also needs to understand how large the
inverse is (or to control closely related quantities, such as the
_condition number_ or the _least singular value_). For this more
quantitative task, one needs to use "robust" versions of the
technologies discussed in the first lecture. In particular, the
inverse problem for the _small ball probability_ (how often does a
random walk return to a small ball?) becomes central. We will discuss
some recent work of Rudelson-Vershynin and Tao-Vu in understanding
these issues.
Title: III. Eigenvalue Distributions of Random Matrices
III. In the third lecture we consider the full spectral distribution
of the (complex) eigenvalues of a (non-selfadjoint) random matrix M,
and their limiting value as the size of the matrix goes to infinity.
Due to the lack of self-adjointness, many standard techniques, such as
the moment method, do not work well. Instead one needs to study other
transforms of this distribution, such as the quantity $\log|\det
(M-zI)|$ for various complex numbers z, as pioneered by Girko and then
formalised by Bai, Girko, and others. In order to execute this
strategy for discrete matrices, the results from the previous two
lectures are needed. We survey some recent developments by
Gotze-Tikhomorov, Pan-Zhou, Tao-Vu and others on this topic.
Title: Logconcave Random Graphs
We propose the following model of a random graph on $n$ vertices.
Let F be a distribution in R_+^{n(n-1)/2} with a coordinate for every
pair ij with 1 \le i,j \le n. Then G_{F,p} is the distribution on graphs
with n vertices obtained by picking a random point X from F and
defining a graph on n vertices whose edges are pairs ij for which
X_{ij} \le p. The standard Erd\H{o}s-R\'{e}nyi model is the special
case when F is uniform on the 0-1 unit cube. We determine basic
properties such as the connectivity threshold for quite general distributions.
We also consider cases where the X_{ij} are the edge weights in some
random instance of a combinatorial optimization problem. By choosing
suitable distributions, we can capture random graphs with interesting
properties such as triangle-free random graphs and weighted random
graphs with bounded total weight.
Title: Concentration and the Littlewood-Offord Problem
The classical concentration inequalities of probability theory demonstrate
that nice random variables are typically concentrated about their means.
Less studied but often needed is the reverse phenomenon -- that the
concentration can not be too tight.
The Littlewood-Offord Problem asks: how much can a sum of independent
random variables concentrate? I will discuss a sharp result joint with
Mark Rudelson which clasiffies the obstacles in the Littlewod-Offord
Problems via additive combinatorics. This leads to a proof of two basic
conjectures on invertibility of random matrices, which will be the content
of the talk by Mark Rudelson.
Title: On the Singular Probability of Random Discrete Matrices
Let $n$ be a large integer and $M_n$ be an $n$ by $n$ random matrix
whose entries are independent (but not necessarily identically distributed)
random variables taking at most $n^{o(n)}$ values in the complex numbers.
The main goal of this paper is to prove a general upper bound for the
probability that $M_{n}$ is singular.
For a constant $0< p< 1$ and a constant positive integer $r$, we will
define aproperty called $p$-bounded of exponent $r$.
Our main result shows that if
the entries of $M_n$ satisfy this property, then the probability that $M_n$ is singular is at most
$(p^{1/r} + o(1))^n$. Here are a few sample corollaries of this theorem:
The proof refines the approach from Kahn-Komlos-Szemeredi and Tao-Vu and makes a critical use of Tao-Vu inverse theorem.
(Joint work with J. Bourgain and V. Vu).
Title: Local Dependencies in Random Matrices, the LLN, and Algebraicity
We consider random matrices whose entries are locally dependent.
We prove the convergence of the spectral measure, and characterize the
limit by means of a system of (algebraic) equations. Study of these
equations yields some analytic information on the limit density.
(joint work with G. W. Anderson).
Santosh Vempala, Georgia Tech
Roman Vershynin, UC. Davis
Phillip Matchett Wood
(1) If the distribution of each entry has mass at most p on a single point, then the singular probability is at most
$p+o(1))^{n/2}. in the special case of random Bernoulli matrices, this improves the previous bound $(3/4+o(1))^n$ of Tao and Vu.
(2) If the entries are iid with distribution which puts mass 1/2 on 0 and 1/4 on -1 and 1, then the singular probability is at most
$(1/2+o(1))^n$. This bounds is sharp, as the probability of having a zero row is (1/2+o(1))^n.
Ofer Zeitouni, Weizman Institute
Previous: Program
Workshop Index