Rank-uniform local law for Wigner matrices

Giorgio Cipolloni, László Erdős, Dominik Schröder

In the world of random matrix theory, Wigner matrices—large matrices with independent, identically distributed entries—are the gold standard for modeling complex quantum systems. A fundamental goal in this field is to understand the "local law," which describes how the matrix's resolvent (a key operator related to its eigenvalues and eigenvectors) behaves at very small scales as the matrix size $N$ grows to infinity.

Historically, researchers have relied on two distinct types of local laws: the averaged law (useful for full-rank observables) and the isotropic law (optimized for rank-one observables). In this paper, Cipolloni, Erdős, and Schröder bridge this gap by proving a new rank-uniform local law that handles observables of any rank optimally, providing a unified framework for understanding eigenvector fluctuations.

TL;DR

The authors prove a new "rank-uniform" local law for Wigner matrices that provides optimal error bounds for test observables of any rank. They use this to establish a Central Limit Theorem for eigenvector overlaps, showing that these overlaps fluctuate according to a Gaussian distribution for a much broader class of matrices than previously known.

What Problem Does This Paper Address?

To study the spectral properties of a Wigner matrix $W$, mathematicians often look at its resolvent $G(z) = (W-z)^{-1}$. A "local law" essentially states that as $N$ becomes large, $G(z)$ starts behaving like a simple scalar matrix $m(z)I$.

However, we usually test this behavior against a deterministic "test matrix" $A$ by looking at the quantity $\langle (G(z) - m(z))A \rangle$.

  • The Gap: If $A$ is the identity matrix (full rank), the "averaged local law" gives a very tight bound. If $A$ is a rank-one projection (like $yx^*$), the "isotropic local law" is used.
  • The Challenge: If $A$ has an intermediate rank (e.g., rank $\sqrt{N}$), neither of the existing laws is optimal. The averaged law overestimates the error, while the isotropic law doesn't capture the tracial cancellations. There was no single, unified estimate that worked across the entire spectrum of possible ranks for $A$.

Methods

The paper introduces several technical innovations to achieve this unification:

  1. Hilbert-Schmidt Scaling: The core insight is that the error in the local law should be measured using the Hilbert-Schmidt norm of the traceless part of the test matrix ($|A^\circ|_2$) rather than just the standard operator norm. This allows the bound to "sense" the effective rank of the observable.
  2. Master Inequalities: Traditional proofs in this field often rely on complex Feynman diagram expansions, which become unmanageable for "multi-resolvent" chains (products of multiple $G$'s and $A$'s). The authors instead develop a system of "master inequalities"—a recursive way to bound higher-order error terms without drawing thousands of diagrams.
  3. Dynamical Approach (DBM): To prove the Central Limit Theorem for eigenvectors, they use Dyson Brownian Motion (DBM). They evolve the matrix through a stochastic process and show that the eigenvector overlaps satisfy a "stochastic eigenstate equation," which naturally leads to Gaussian fluctuations.
  4. Green Function Comparison (GFT): Finally, they use comparison techniques to show that the results derived from the dynamical DBM model hold true for the original, static Wigner matrices.

Datasets

As a theoretical mathematics paper, no empirical datasets were used. The results are proven analytically for the general class of Wigner matrices (real symmetric or complex Hermitian) under standard moment conditions.

Key Insights and Ideas

  • Optimal Interpolation: The new rank-uniform law (Equation 1.4 in the paper) perfectly interpolates between the averaged and isotropic regimes. It shows that these two seemingly different laws are actually two sides of the same coin.
  • Removing "Unnatural" Conditions: Previous proofs of the Central Limit Theorem for eigenvector overlaps required the test matrix $A$ to have either very high rank or very low rank. This paper removes those restrictions, requiring only that the "effective rank" of $A$ is much larger than 1.
  • The Power of Tracelessness: The authors highlight that "traceless" observables ($A$ where the sum of diagonal elements is zero) provide a $\sqrt{\eta}$ improvement in error. By decomposing any matrix into its trace and its traceless part, they can apply the optimal bound to each piece.

Key Results

  • The Rank-Uniform Law: They establish a bound where the error is controlled by: $$\frac{N^\xi}{N\eta} |\langle A \rangle| + \frac{N^\xi \rho^{1/2} \langle |A^\circ|^2 \rangle^{1/2}}{N\sqrt{\eta}}$$ This formula is "rank-uniform" because it automatically adjusts based on whether $A$ is dominated by its trace (high rank) or its traceless fluctuations (low rank).

  • Central Limit Theorem (QUE): They prove that for any bulk eigenvector $u_i$ and a deterministic matrix $A$, the normalized overlap: $$\sqrt{\frac{\beta N}{2 \langle |A^\circ|^2 \rangle}} (\langle u_i, A u_i \rangle - \langle A \rangle) \implies \mathcal{N}(0, 1)$$ This confirms that the "Quantum Unique Ergodicity" (the idea that eigenvectors look like random noise) holds with Gaussian fluctuations for virtually any deterministic observable.

Limitations and Future Work

  • Bulk vs. Edge: The Central Limit Theorem for eigenvectors is proven for the "bulk" of the spectrum (the middle part of the semicircular distribution). The authors note that the behavior at the "edges" (where the eigenvalues taper off) is more complex and is deferred to future research.
  • Symmetry Constraints: While the law is general, the detailed proof of the Central Limit Theorem focuses on the real symmetric case, noting that the complex Hermitian case is analogous but requires slightly different bookkeeping for the non-zero second moments.

Summary updated

Abstract

We prove a general local law for Wigner matrices which optimally handles observables of arbitrary rank and thus it unifies the well-known averaged and isotropic local laws. As an application, we prove that the quadratic forms of a general deterministic matrix $A$ on the bulk eigenvectors of a Wigner matrix has approximately Gaussian fluctuation. For the bulk spectrum, we thus generalize our previous result [arXiv:2103.06730] valid for test matrices $A$ of large rank as well as the result of Benigni and Lopatto [arXiv:2103.12013] valid for specific small rank observables.
BibTeX
@inproceedings{cipolloni2022rank,
  author = {Giorgio Cipolloni and László Erdős and Dominik Schröder},
  title = {{Rank-uniform local law for Wigner matrices}},
  booktitle = {Forum},
  year = {2022},
  doi = {10.1017/fms.2022.86},
}
Figure