[Paper Review] An Extension of Level-spacing Universality

E. Brézin, S. Hikami|arXiv (Cornell University)|Feb 25, 1997

Random Matrices and Applications73 citations

TL;DR

This paper extends the universality of level-spacing statistics in random matrix theory beyond the Gaussian unitary ensemble (GUE) to Hamiltonians with a deterministic background H₀. Using a novel integral representation based on the Itzykson-Zuber formula, the authors derive exact finite-N expressions for n-point correlation functions as determinants of a kernel K_N(λ,μ), proving that Dyson's short-distance universality and the GUE-level spacing distribution P(s) remain valid regardless of H₀. The key result is the robustness of P(s) under arbitrary non-random perturbations.

ABSTRACT

Dyson's short-distance universality of the correlation functions implies the universality of P(s), the level-spacing distribution. We first briefly review how this property is understood for unitary invariant ensembles and consider next a Hamiltonian H = H_0+ V , in which H_0 is a given, non-random, N by N matrix, and V is an Hermitian random matrix with a Gaussian probability distribution. n-point correlation function may still be expressed as a determinant of an n by n matrix, whose elements are given by a kernel $K(\lambda,\mu)$ as in the H_0=0 case. From this representation we can show that Dyson's short-distance universality still holds. We then conclude that P(s) is independent of H_0.

Motivation & Objective

• To establish the universality of the level-spacing distribution P(s) for random Hamiltonians H = H₀ + V, where H₀ is a non-random matrix and V is a Gaussian random matrix.
• To overcome the breakdown of orthogonal polynomial methods in the presence of a non-zero H₀, which breaks unitary invariance.
• To derive exact, finite-N expressions for n-point correlation functions in the presence of an external source H₀.
• To demonstrate that the short-distance scaling limit of the correlation functions remains universal, leading to a universal P(s) independent of H₀.

Proposed method

• Uses the Itzykson-Zuber integral formula to express the partition function and correlation functions in terms of integrals over 2n variables.
• Applies contour integration techniques to transform the n-point correlation function into a form amenable to determinant representation.
• Identifies a Cauchy determinant structure in the transformed integral, enabling the expression of the n-point function as a determinant of an n×n matrix.
• Derives an explicit integral representation for the kernel K_N(λ, μ) that generalizes the sine kernel in the short-distance scaling limit.
• Verifies consistency conditions, such as the kernel's trace identity and eigenfunction properties, using contour integral representations.
• Demonstrates that the kernel's eigenvalues are Hermite polynomials for n < N, confirming the structure's validity.

Experimental results

Research questions

• RQ1 Does the level-spacing distribution P(s) remain universal when a non-random matrix H₀ is added to a Gaussian random matrix V?
• RQ2 Can the n-point correlation function for H = H₀ + V be expressed as a determinant of a kernel, despite the loss of unitary invariance?
• RQ3 Is the short-distance scaling limit of the correlation function universal, i.e., does it approach the sine kernel independent of H₀?
• RQ4 What is the structure of the kernel K_N(λ, μ) in the presence of an external source H₀, and does it satisfy necessary consistency conditions?
• RQ5 How does the universality of P(s) extend to energy scales away from zero, particularly when ρ(E₀) is finite and non-zero?

Key findings

• The n-point correlation function R_n(λ₁,…,λ_n) is exactly expressible as a determinant of an n×n matrix with elements given by a kernel K_N(λ_i, λ_j), even for finite N and non-zero H₀.
• The kernel K_N(λ, μ) is derived as an explicit integral over 2n variables, generalizing the standard sine kernel in the short-distance scaling limit.
• In the short-distance scaling limit (N→∞, N|λ_i - λ_j| fixed), the kernel approaches the universal sine kernel ˜K(y₁,y₂) = sin[π(y₁−y₂)] / [π(y₁−y₂)], independent of H₀.
• As a direct consequence, the level-spacing distribution P(s) is universal and identical to the GUE result, regardless of the deterministic part H₀.
• The kernel satisfies the consistency condition ∫ KN(λ,μ)KN(μ,ν) dμ = KN(λ,ν), confirming its role in a consistent correlation hierarchy.
• The kernel has N eigenvalues equal to one with Hermite polynomials as eigenfunctions for n < N, and the structure breaks down for n ≥ N, consistent with finite-N effects.