Local Random Matrix Theory (LRMT)

• Local Random Matrix Theory (LRMT) is a framework describing fine-scale spectral statistics and operator matrix elements in high-dimensional systems with physical locality and structural constraints.
• It predicts universal behaviors such as sine-kernel and Airy-kernel statistics through precise analysis based on local laws and the matrix Dyson equation.
• LRMT extends its methodology to structured ensembles, dynamical quantum systems, and neural network Hessians, linking spectral universality to phenomena like ETH and learning dynamics.

Local Random Matrix Theory (LRMT) is a conceptual and technical framework developed to describe universal, fine-scale spectral statistics and operator matrix elements in high-dimensional quantum and statistical systems where physical locality, or other structural constraints, impose correlations among matrix elements beyond standard Wigner–Dyson random matrix ensembles. By focusing on matrix statistics at the scale of individual eigenvalue spacings (“local laws”), LRMT provides precise predictions and universality results applicable to quantum many-body systems with local interactions, classical and non-Hermitian random matrices, deformed or correlated ensembles, and out-of-equilibrium settings. It establishes the conditions under which local spectral universality (e.g., sine-kernel or Airy-kernel statistics) and the eigenstate thermalization hypothesis (ETH) arise, and explicates their breakdown when locality, structure, or heavy tails cause ergodicity to fail.

1. Operator Expansion, Behemoths, and Local Law Scaling

A cornerstone of LRMT in many-body quantum systems is the explicit expansion of physical operators in terms of highly nonlocal rank-one “Behemoth” projectors $$\hat\Omega_{\mathbf{n n}'} = |\mathbf{n}\rangle\langle\mathbf{n}'|$$ indexed by basis configurations. In ergodic eigenstate bases, Behemoth matrix elements $X_{AB} = \psi_A^*(\mathbf{n})\psi_B(\mathbf{n}')$ are products of random amplitudes, leading to off-diagonal statistics $$P_{\text{Behemoth}}(x) = \frac{1}{\pi\sigma^2}K_0(|x|/\sigma^2)$$, with $\sigma^2 = 1/\mathcal{D}$ and width scaling $w \sim \mathcal{D}^{-1}$. This “super-ETH” scaling is singular (non-Gaussian), and underpins a theory where arbitrary local or nonlocal operators are viewed as sums of Behemoths:

$\hat O = \sum_{k=1}^M \hat\Omega_k$

By the central limit theorem, if $M = O(\mathcal{D})$, as for true local observables, the resulting matrix elements $Y_{AB}$ are Gaussian with width $w \sim \mathcal{D}^{-1/2}$, reproducing the ETH scaling $w \sim e^{-S/2}$. For increased “density” $X_{AB} = \psi_A^*(\mathbf{n})\psi_B(\mathbf{n}')$0, $X_{AB} = \psi_A^*(\mathbf{n})\psi_B(\mathbf{n}')$1, LRMT predicts sub-ETH scaling $X_{AB} = \psi_A^*(\mathbf{n})\psi_B(\mathbf{n}')$2, $X_{AB} = \psi_A^*(\mathbf{n})\psi_B(\mathbf{n}')$3, interpolating between ETH and more delocalized operator statistics. These predictions are in quantitative agreement with numerical results for non-integrable quantum systems, except in regimes of integrability or many-body localization where the universality breaks down as anticipated (Khaymovich et al., 2018).

2. Local Laws, Bulk and Edge Universality, and the Matrix Dyson Equation

A fundamental achievement of LRMT is the proof of local spectral laws for Hermitian and non-Hermitian random matrices: the empirical spectral densities and local correlation functions, after appropriate rescaling, converge on scales of the mean level spacing to deterministic universal limits (e.g., sine-kernel in the bulk, Airy-kernel at the edge). The analytic backbone is the matrix Dyson equation (MDE):

$X_{AB} = \psi_A^*(\mathbf{n})\psi_B(\mathbf{n}')$4

where $X_{AB} = \psi_A^*(\mathbf{n})\psi_B(\mathbf{n}')$5 is the variance (or self-energy) operator, $X_{AB} = \psi_A^*(\mathbf{n})\psi_B(\mathbf{n}')$6 encodes deterministic structure, and $X_{AB} = \psi_A^*(\mathbf{n})\psi_B(\mathbf{n}')$7 approximates the resolvent $X_{AB} = \psi_A^*(\mathbf{n})\psi_B(\mathbf{n}')$8. Under mean-field scaling and mild moment assumptions, local laws are established down to scales $X_{AB} = \psi_A^*(\mathbf{n})\psi_B(\mathbf{n}')$9:

• Entrywise and isotropic estimates:

$$P_{\text{Behemoth}}(x) = \frac{1}{\pi\sigma^2}K_0(|x|/\sigma^2)$$0

$$P_{\text{Behemoth}}(x) = \frac{1}{\pi\sigma^2}K_0(|x|/\sigma^2)$$1

• Averaged control:

$$P_{\text{Behemoth}}(x) = \frac{1}{\pi\sigma^2}K_0(|x|/\sigma^2)$$2

The stability properties of the MDE govern the robustness of these estimates (Erdos, 2019). When the MDE's solution is regular (i.e., the associated operator is invertible with controlled norm), universality in local correlation functions and eigenvalue gap statistics (bulk sine-kernel, edge Airy kernel / Tracy–Widom laws) follows via Green-function comparison and Dyson Brownian motion techniques (Sodin, 2014, Kriecherbauer et al., 2015).

3. Extensions to Structured and Localized Ensembles

LRMT encompasses a broad class of structured models, including:

• Heavy-tailed ensembles: Lévy matrices with tail index $$P_{\text{Behemoth}}(x) = \frac{1}{\pi\sigma^2}K_0(|x|/\sigma^2)$$3 exhibit a sharp GOE $$P_{\text{Behemoth}}(x) = \frac{1}{\pi\sigma^2}K_0(|x|/\sigma^2)$$4 Poisson transition as a function of $$P_{\text{Behemoth}}(x) = \frac{1}{\pi\sigma^2}K_0(|x|/\sigma^2)$$5 and energy. For $$P_{\text{Behemoth}}(x) = \frac{1}{\pi\sigma^2}K_0(|x|/\sigma^2)$$6, the bulk spectrum is delocalized with GOE local statistics; for $$P_{\text{Behemoth}}(x) = \frac{1}{\pi\sigma^2}K_0(|x|/\sigma^2)$$7, a mobility edge emerges, separating Gaussian and Poisson phases. Essential singularities in finite-size crossovers are a generic LRMT phenomenon for ensembles with incipient localization (Tarquini et al., 2015).
• Band and sparse matrices: LRMT predicts nonuniversal crossover regimes and slow convergence to universal statistics as bandwidth shrinks or sparsity increases, due to the emergence of “Goldstone modes” and rare-region effects (Knowles et al., 2014).
• Classical non-Hermitian products: The local circular law for products $$P_{\text{Behemoth}}(x) = \frac{1}{\pi\sigma^2}K_0(|x|/\sigma^2)$$8 of deterministic and random matrices holds up to optimal scales for test points outside the spectral unit circle, with explicit error bounds (Xi et al., 2016).
• p-adic random matrices: LRMT has been extended to $$P_{\text{Behemoth}}(x) = \frac{1}{\pi\sigma^2}K_0(|x|/\sigma^2)$$9-adic contexts, where the limiting local point processes for singular numbers of random matrix products are described via explicit contour integrals and $\sigma^2 = 1/\mathcal{D}$0-series, interpolating between classical kernel regimes (Peski, 2023).

4. Dynamical and Many-Body Applications: Energy Transport and ETH

Recent developments connect LRMT to dynamical phenomena in quantum systems. In random quantum chains with strictly local Hilbert space terms, the dynamics of conserved densities (e.g., energy) and out-of-time-ordered correlators (OTOCs) can be mapped, in the large local dimension limit, to single-particle random walks on appropriate graphs. This results in explicit formulas:

$\sigma^2 = 1/\mathcal{D}$1

Spectral densities interpolate between arcsine (free convolution) and Gaussian limits, with Bethe lattice analogs in intermediate regimes. Exact diagonalization for $\sigma^2 = 1/\mathcal{D}$2 confirms emergence of diffusive scaling ($\sigma^2 = 1/\mathcal{D}$3) for energy autocorrelators. Such LRMT frameworks provide minimal, analytically soluble settings for universal hydrodynamics of chaos and thermalization (Pollock et al., 7 Feb 2025).

In the context of ETH, LRMT rigorously establishes that random ensembles of strictly local Hamiltonians and local observables yield exponentially small, Gumbel-distributed (extremal) fluctuations in diagonal matrix elements, in line with the strong ETH. However, ensemble ergodicity is broken due to locality, and sample-to-sample fluctuations in smooth functions persist in the thermodynamic limit (Sugimoto et al., 2020).

5. Anisotropic and Correlated Ensembles

LRMT has been advanced to include anisotropic settings, such as deformed Wigner and sample covariance matrices:

$\sigma^2 = 1/\mathcal{D}$4

for deterministic $\sigma^2 = 1/\mathcal{D}$5 and random $\sigma^2 = 1/\mathcal{D}$6. The anisotropic local law asserts that for any deterministic vectors $\sigma^2 = 1/\mathcal{D}$7, $\sigma^2 = 1/\mathcal{D}$8,

$\sigma^2 = 1/\mathcal{D}$9

where the deterministic approximation $w \sim \mathcal{D}^{-1}$0 is determined by a matrix Dyson or Pastur equation, and the error $w \sim \mathcal{D}^{-1}$1 is sharp. These techniques are also critical for proving edge universality (Tracy–Widom fluctuations), delocalization of eigenvectors, and the BBP transition in signal-plus-noise models with arbitrary deterministic structure (Knowles et al., 2014, Ding et al., 2020).

6. LRMT in Neural Networks and Modern Learning Theory

LRMT is now integral in analyzing the loss surface geometry of deep neural networks. The Hessian at critical points can be modeled as a spiked random-matrix plus “noise” (sum of layerwise random covariances), and LRMT predicts that its empirical spectrum shows a bulk matching the Marchenko–Pastur law, outlier eigenvalues determined by signal strength, and local eigenvalue statistics governed by sine- and Airy-kernels. These results enable adaptive learning-rate schedules (estimating spectrum edges in time), principled preconditioning (using $w \sim \mathcal{D}^{-1}$2 as a Hessian inverse proxy), and identify the universal structure of rare, information-carrying directions in the parameter space (Baskerville et al., 2022).

The physical implication is a direct, quantitative bridge from random-matrix spectral laws—local, edge, and outlier phenomena—to optimally-tuned dynamics and generalization in high-dimensional learning systems.

---

Summary Table: Universal Local Laws in LRMT Contexts

System / Ensemble	Local Law Type	Kernel / Limit
Wigner, Invariant Hermitian	Sine / Airy (bulk/edge)	$w \sim \mathcal{D}^{-1}$3, $w \sim \mathcal{D}^{-1}$4
Lévy matrices, $w \sim \mathcal{D}^{-1}$5	GOE $w \sim \mathcal{D}^{-1}$6 Poisson	Mobility edge, essential singularity
Sample covariance, deformed Wishart	Marchenko–Pastur	MP law, edge: Tracy–Widom
Band/Sparse random (with Goldstone)	Slower universality	Crossover, delayed kernel convergence
Non-Hermitian (circular law, products)	Resolved local law outside circle	Ginibre, rotation-invariant
Strictly local Hamiltonians (spin/boson chains)	Behemoth–sum, CLT	Gaussian (ETH), Bessel functions
Deep NN Hessians (spiked + noise)	Spiked deformed covariance	Outlier + MP bulk + TW edge

---

Through combinatorial analysis, self-consistent equations, advanced probabilistic and field-theoretic arguments, LRMT achieves a quantitative taxonomy of local operator and spectral statistics in both physical and mathematical ensembles, clarifying the conditions under which the celebrated universality of RMT continues to hold, and identifying new fixed points in the space of random matrices that encode locality, sparsity, and structured randomness.