The anisotropic local law for sample covariance matrices under quadratic-form concentration

Technical Summary: The Anisotropic Local Law for Sample Covariance Matrices Under Quadratic-Form Concentration

1. Problem Statement

The paper addresses the spectral analysis of sample covariance matrices $K = N^{-1}XX^*$ in the proportional regime where the dimension $n$ and sample size $N$ are comparable ($n \asymp N$). The columns $x_1, \dots, x_N \in \mathbb{R}^n$ are independent, centered random vectors with a common covariance matrix $\Sigma = \mathbb{E}[x_i x_i^*]$. Crucially, the authors do not assume the columns have independent coordinates (the "separable" case); instead, the coordinates may exhibit strong, nonlinear dependencies.

The primary goal is to establish the anisotropic local law, which provides a precise approximation of the resolvent $R(z) = (K - z)^{-1}$ for complex spectral parameters $z = E + i\eta$ down to the optimal scale $\eta \ge N^{-1+\tau}$. Specifically, the paper seeks to prove that for deterministic unit vectors $u, v$:
$$|u^* R(z) v - u^* \Pi(z) v| \prec \Psi(z)$$
where $\Pi(z)$ is a deterministic equivalent derived from the deformed Marchenko–Pastur law, and $\Psi(z)$ is the optimal error rate.

Prior work by Fan, Ma, Paquette, and Wang (2026) [FMPW26] established this law under the assumption of quadratic-form concentration (Assumption 2) but required an additional, restrictive hypothesis on the decay of higher-order cumulant tensors (Assumption 3). The central question motivating this work is whether the quadratic-form concentration assumption alone is sufficient to imply the anisotropic local law, thereby removing the need for structural assumptions on higher-order cumulants.

2. Methodology

The proof adapts the zig-zag strategy, a dynamical method introduced by Cipolloni, Erdős, and Schröder [CES24], to the setting of sample covariance matrices with dependent coordinates. The strategy involves evolving the random matrix along a stochastic flow while simultaneously moving the spectral parameter along a characteristic of the associated Dyson equation.

Key Technical Components:

• Block Linearization: Instead of working directly with $K$, the authors linearize the problem using a Hermitian block matrix $L(X; Z, \beta)$ involving the data matrix $X$. This allows the use of the Matrix Dyson Equation (MDE) framework. The deterministic equivalent is constructed as a block-diagonal solution $M$ to the MDE.
• Covariance-Preserving Ornstein–Uhlenbeck (OU) Flow: The columns of the matrix evolve according to an OU flow $dX_t = -\frac{1}{2}X_t dt + \Sigma^{1/2} dB_t$. This flow preserves the covariance structure $\Sigma$ while introducing Gaussian smoothing.
• The "Zig" Step (Forward Comparison): The resolvent evolves along the OU flow while the source parameters $(Z, \beta)$ move along a specific characteristic curve derived from the MDE. The motion of the source is chosen to cancel the leading-order Itô correction. The error is controlled using Ward identities and a static fluctuation-averaging argument adapted from [FMPW26].
• The "Zag" Step (Reverse Comparison): This is the novel contribution of the paper. The source is held fixed, and the flow is reversed from the smoothed state back to the original ensemble. Unlike previous proofs that rely on cumulant expansions of individual matrix entries (which fail for dependent coordinates), this step treats each column as a whole.
  • The authors differentiate the fluctuation moments $\mathbb{E}|u^*(R - M)u|^{2L}$ along the OU flow.
  • Using the Schur complement formula, the dependence of the resolvent on a specific column $x_i$ (conditioned on the others) is expressed as a rational function of a linear form $x_i^* b_i$ and a quadratic form $x_i^* A_i x_i$.
  • The generator of the OU flow decomposes into a sum of one-column generators $L_i$. The authors show that $L_i$ preserves the class of observables formed by these linear and quadratic forms.
  • Crucially, the control of these forms relies only on the quadratic-form concentration assumption (Assumption 2). No expansion into coordinates or assumptions on higher-order cumulants are required. The leading cancellations arise from the centering of these forms ($\mathbb{E}[x_i^* b_i] = 0$ and $\mathbb{E}[x_i^* A_i x_i - \text{Tr}(\Sigma A_i)] = 0$).

3. Key Contributions and Results

Main Theorem (Theorem 2.4 / Theorem 1.1):
Under Assumption 1 (regularity of $\Sigma$ and the proportional regime) and Assumption 2 (quadratic-form concentration and polynomial norm moments), the optimal anisotropic local law holds:
$$|u^* (R(z) - \Pi(z)) v| \prec \sqrt{\frac{\text{Im } m_0(z)}{N\eta}} + \frac{1}{N\eta}$$
uniformly for $z$ in regular spectral domains bounded away from the origin. The result applies to the full linearized resolvent and implies the same bound for the standard sample covariance resolvent $R(z)$ and the Gram resolvent $\tilde{R}(z)$.

Removal of Cumulant Assumptions:
The paper explicitly removes the higher-order cumulant tensor assumption (Assumption 3) required in [FMPW26]. This resolves the open question raised in that work regarding whether quadratic-form concentration is sufficient for the anisotropic law.

Applications and Examples:
The authors verify that their assumptions hold for a broad class of non-separable models where coordinates are nonlinearly dependent, including:

• Convex Concentration: Vectors satisfying a dimension-free logarithmic Sobolev inequality or convex concentration with subpolynomial constants.
• Log-Concave Columns: Independent samples from centered log-concave distributions (e.g., uniform on convex bodies).
• Nonlinear Tilts of Gaussian Measures: Distributions of the form $e^{-\frac{1}{2}x^*\Lambda x - W(x)}$ where $W$ has a bounded Hessian, allowing for dense coupling of coordinates without structural sparsity.
• Deep Random Features: Models of the form $x = \sigma_L(T_L \dots \sigma_1(T_1 y))$ with Lipschitz activations and fixed depth.
• Spherical $p$-Spin Models: High-temperature Gibbs measures on the sphere. The authors note that for the 4-spin model, the fourth cumulant tensor is disordered and violates the [FMPW26] cumulant condition, yet the anisotropic local law holds under their new framework.

Corollaries:
As a direct consequence of the anisotropic law, the paper establishes the delocalization of eigenvectors in every deterministic direction for eigenvalues in the regular part of the spectrum.

4. Significance and Claims

The paper claims to provide a new, dynamical proof of the anisotropic local law for real columns that does not rely on the restrictive separability assumption or higher-order cumulant structures.

• Generality: By relying solely on quadratic-form concentration, the result applies to "disordered-cumulant" vectors where previous methods failed. This significantly broadens the scope of random matrix theory to include complex, nonlinearly dependent high-dimensional distributions found in modern statistics and machine learning (e.g., deep neural networks and random feature models).
• Methodological Innovation: The "whole-column" comparison step in the zag estimate represents a significant technical departure from entrywise cumulant expansions. It demonstrates that the Schur complement structure of the resolvent, combined with OU flow differentiation, is sufficient to control fluctuations without expanding the random vectors into their coordinates.
• Optimality: The error rate $\Psi(z)$ and the spectral scale $\eta \ge N^{-1+\tau}$ are optimal, matching the best-known results for the separable case (Knowles and Yin [KY17]).

The authors remain modest regarding open problems, noting that extending the result to domains approaching the origin (the hard edge) and proving edge universality for general dependent columns remain open challenges requiring further analysis.