[Paper Review] Distribution functions for largest eigenvalues and their applications

Craig A. Tracy, Harold Widom|ArXiv.org|Oct 17, 2002

Random Matrices and ApplicationsMathematics48 references108 citations

TL;DR

This paper establishes that the limiting distribution of the largest eigenvalue in Gaussian random matrix ensembles (GOE, GUE, GSE) converges to universal laws described by a Painlevé II transcendental function. These distributions, expressed via Fredholm determinants of the Airy kernel and related to integrable systems, emerge universally across diverse systems in statistical physics, growth processes, random tilings, and statistics, with explicit formulas for Fβ(s) and their moments provided.

ABSTRACT

It is now believed that the limiting distribution function of the largest eigenvalue in the three classic random matrix models GOE, GUE and GSE describe new universal limit laws for a wide variety of processes arising in mathematical physics and interacting particle systems. These distribution functions, expressed in terms of a certain Painlevé II function, are described and their occurences surveyed.

Motivation & Objective

• To establish the universal limiting distribution of the largest eigenvalue in the three classic Gaussian random matrix ensembles (GOE, GUE, GSE).
• To demonstrate that this limiting distribution is described by a Painlevé II function and expressed via Fredholm determinants of the Airy kernel.
• To survey the emergence of these universal laws in diverse physical and stochastic systems beyond random matrix theory.
• To provide explicit formulas and statistical properties (mean, variance, skewness, kurtosis) for the limiting distribution functions Fβ(s).
• To establish universality beyond Gaussian ensembles, showing that the same limiting laws arise in Wigner matrices, growth processes, random tilings, and queuing systems.

Proposed method

• Derives the limiting distribution Fβ(s) as the N→∞ limit of the cumulative distribution F_{N,β}(t) for the largest eigenvalue in N×N random matrices.
• Uses the Airy kernel K_Airy(x,y) = [Ai(x)Ai'(y) - Ai'(x)Ai(y)] / (x-y) on L^2(s,∞) to express the GUE case as F_2(s) = det(I - K_Airy) = exp(-∫_s^∞ (x-s)q^2(x)dx).
• Relies on the solution q(s) of the Painlevé II equation q'' = s q + 2 q^3 with asymptotic condition q(s) ~ Ai(s) as s→∞.
• Extends results to GOE and GSE via relations F_1(s) = exp(-1/2 ∫_s^∞ q(x)dx) × F_2(s)^{1/2} and F_4(s/√2) = cosh(1/2 ∫_s^∞ q(x)dx) × F_2(s)^{1/2}.
• Applies Riemann-Hilbert methods and orthogonal polynomial techniques to prove universality for general potentials V(A) in the β=2 case.
• Demonstrates that the same limiting distribution F_2(s) appears in non-Gaussian systems such as Wigner matrices, growth models, random tilings (Aztec diamond), and queuing networks.

Experimental results

Research questions

• RQ1 What is the limiting distribution of the largest eigenvalue in the Gaussian Unitary Ensemble (GUE) as N→∞?
• RQ2 How do the limiting laws for the largest eigenvalue in GOE and GSE relate to the GUE case and to Painlevé II functions?
• RQ3 To what extent are these limiting distributions universal across different random matrix ensembles and non-Gaussian models?
• RQ4 In what physical and stochastic systems does the F_2(s) distribution emerge as a universal limit?
• RQ5 Can the Fredholm determinant representation of F_2(s) be extended to other systems like growth processes or random tilings?

Key findings

• The limiting distribution F_2(s) for the GUE ensemble is given by F_2(s) = det(I - K_Airy) = exp(-∫_s^∞ (x-s)q^2(x)dx), where q is the solution to the Painlevé II equation q'' = s q + 2 q^3 with q(s) ~ Ai(s) as s→∞.
• For the GOE (β=1), the limiting distribution is F_1(s) = exp(-1/2 ∫_s^∞ q(x)dx) × F_2(s)^{1/2}, and for the GSE (β=4), F_4(s/√2) = cosh(1/2 ∫_s^∞ q(x)dx) × F_2(s)^{1/2}.
• The mean, standard deviation, skewness, and kurtosis of F_2(s) are μ_2 = -1.77109, σ_2 = 0.9018, S_2 = 0.224, and K_2 = 0.093.
• Universality holds for general unitarily invariant potentials V(A) in the β=2 case, with the GUE limit F_2(s) emerging generically, though fine-tuning can produce new universality classes.
• The F_2(s) distribution appears universally in non-random-matrix systems: in the edge of the spectrum of Wigner matrices, in the Airy process, in random tilings of the Aztec diamond, and in the queuing model of D(k,n).
• In the queuing model with Poisson service times, the normalized departure time D(⌊ xn⌋,n) converges in distribution to F_2(s) as n→∞, with explicit constants c_1 and c_2 depending on x.