[Paper Review] Dyson's Brownian-motion model for random matrix theory - revisited. With an Appendix by Don Zagier

Christopher Joyner, Uzy Smilansky|arXiv (Cornell University)|Mar 22, 2015

Random Matrices and Applications11 references3 citations

TL;DR

This paper revisits Dyson's Brownian motion model in random matrix theory by shifting focus from eigenvalues to matrix traces $ t_n = \sum \lambda_\nu^n $, deriving a Fokker-Planck equation for the joint probability density in trace space. It establishes a stationary solution linked to eigenvalue statistics and proves two novel combinatorial identities via residue calculus, offering a new algebraic and dynamical perspective on the Vandermonde determinant's emergence.

ABSTRACT

We offer an alternative viewpoint on Dyson's original paper regarding the application of Brownian motion to random matrix theory (RMT). In particular we show how one may use the same approach in order to study the stochastic motion in the space of matrix traces $t_n = \sum_{ν=1}^{N} λ_ν^n$, rather than the eigenvalues $λ_ν$. In complete analogy with Dyson we obtain a Fokker-Planck equation that exhibits a stationary solution corresponding to the joint probability density function in the space $t = (t_1,\ldots,t_n)$, which can in turn be related to the eigenvalues $λ= (λ_1,\ldots,λ_N)$. As a consequence two interesting combinatorial identities emerge, which are proved algebraically in the appendix. We also offer a number of comments on this version of Dyson's theory and discuss its potential advantages.

Motivation & Objective

• To reframe Dyson's original Brownian motion approach in random matrix theory by focusing on traces $ t_n = \sum \lambda_\nu^n $ instead of eigenvalues $ \lambda_\nu $.
• To derive a Fokker-Planck equation governing the stochastic evolution of the joint probability density in trace space $ \mathbf{t} = (t_1, \dots, t_n) $.
• To establish a stationary solution for this Fokker-Planck equation that corresponds to the known eigenvalue joint probability density in Gaussian ensembles.
• To derive and prove two new combinatorial identities involving traces and eigenvalues using residue calculus, as detailed in the appendix by Don Zagier.

Proposed method

• Formulate the stochastic dynamics of matrix entries as Ornstein-Uhlenbeck processes with specified drift and diffusion coefficients, leading to a Brownian motion in matrix space.
• Transform the eigenvalue-level dynamics (1.4) into trace-level dynamics by expressing $ \delta \lambda_\mu $ in terms of matrix element increments and summing over $ \mu $.
• Derive the first and second moments of the trace dynamics $ \mathbf{E}(\delta t_m) $ and $ \mathbf{E}(\delta t_m^2) $, which define the drift and diffusion in the Fokker-Planck equation.
• Construct the Fokker-Planck equation for the joint probability density $ P(\mathbf{t}; s) $, with a stationary solution corresponding to the Gaussian orthogonal/unitary/symplectic ensemble eigenvalue distribution.
• Use generating functions and residue theory to analyze the trace dynamics, expressing $ \partial T(X)/\partial t_m $ as rational functions and summing over $ m $ to derive identities.
• Prove two key identities: $ \sum_m m \partial T/\partial t_m X^{m-1} = -T'(X)/2 + T(X)^2/2 $ and $ \sum_m m \partial \Delta/\partial t_m X^{m-1} = T(X) + T'(X)/T(X) $, using contour integration and residue calculus.

Experimental results

Research questions

• RQ1 How can Dyson's Brownian motion model be reformulated in terms of matrix traces $ t_n = \sum \lambda_\nu^n $ rather than eigenvalues?
• RQ2 What is the form of the Fokker-Planck equation governing the stochastic evolution of the joint probability density in trace space?
• RQ3 Can the stationary solution of the trace-level Fokker-Planck equation be related to the known eigenvalue joint probability density in Gaussian ensembles?
• RQ4 What new combinatorial identities emerge from the trace dynamics, and how can they be rigorously proven?
• RQ5 How does the use of generating functions and residue calculus enable the derivation of identities involving traces and eigenvalues?

Key findings

• The Fokker-Planck equation for the trace dynamics is derived with drift $ F_m(\mathbf{t}) $ and diffusion $ \beta^{-1} $, leading to a stationary solution that corresponds to the joint eigenvalue distribution of Gaussian ensembles.
• The stationary solution in trace space is shown to be equivalent to the known eigenvalue joint probability density $ P(\boldsymbol{\lambda}) \propto \prod_{\mu<\nu} |\lambda_\mu - \lambda_\nu|^\beta \exp(-\beta/2 \sum \lambda_\mu^2) $, confirming consistency.
• Two new combinatorial identities are proven: $ \sum_m m \partial T/\partial t_m X^{m-1} = -T'(X)/2 + T(X)^2/2 $ and $ \sum_m m \partial \Delta/\partial t_m X^{m-1} = T(X) + T'(X)/T(X) $, where $ T(X) = \sum_m t_m X^{m-1} $.
• The identities are derived using residue calculus and contour integration, demonstrating that the trace dynamics encode deep algebraic structures related to the Vandermonde determinant.
• The method reveals that the eigenvalue repulsion factor $ \prod_{\mu<\nu} |\lambda_\mu - \lambda_\nu|^\beta $ arises naturally from the second-order perturbation term in the trace dynamics, mirroring Dyson's original insight but in trace space.
• The approach provides a new algebraic framework for studying random matrix ensembles, with potential advantages in analyzing higher-order spectral statistics and non-Gaussian ensembles.