Localization Lengths of Power-Law Random Band Matrices

Abstract: We study large $N\times N$ power-law random band matrices $H=(H_{ij})$ with centered complex Gaussian entries, where the variances satisfy a power-law decay $\mathbb{E}|H_{ij}|^2\propto (|i-j|/W+1)^{-1-α}$, for some exponent $α>-1$ and bandwidth $W\gg 1$. We establish the following lower bounds, with high probability, on the localization length $\ell$ of bulk eigenvectors in the different regimes of $α$: (1) $\ell=N$ if $-1<α<0$; (2) $\ell \ge W^{C}$ for any large constant $C>0$ if $0 < α<1$; (3) $\ell \ge W^{α/(α-1)}$ if $1 < α<2$; (4) $\ell \ge W^{2}$ if $ α> 2$. These results verify the physical conjecture of arXiv:cond-mat/9604163 on the delocalized side. The main difficulty in the proof lies in handling the interplay between the non-mean-field nature of the model and the slow decay of the variance profile. To address this issue, a key technical ingredient is a new dynamical analysis of $T$-variables formed from pairs of resolvent entries of $H$. In contrast to the fundamental works on regular random band matrices with fast-decaying variances in arXiv:2501.01718 and arXiv:2506.06441, this approach does not rely on higher-order resolvent loops.

• The paper provides nearly sharp lower bounds on localization lengths across various decay regimes for power-law random band matrices.
• It employs novel resolvent flow analysis and loop contraction methods to control slow-decay variances and achieve high-probability bounds.
• The findings validate longstanding conjectures, extending quantum unique ergodicity and universality results to non-exponential decay settings.

Mathematical Analysis of Localization in Power-Law Random Band Matrices

Introduction and Context

The paper "Localization Lengths of Power-Law Random Band Matrices" (2604.12248) presents a comprehensive mathematical study of localization-delocalization properties of eigenstates in power-law random band matrices (PRBMs). These matrices generalize classical random band matrices (RBMs) by assigning the variance of off-diagonal elements according to a slow, power-law decay rather than rapid, compact, or exponential decay. Specifically, in the model considered, the variance structure between entries $H_{ij}$ is given by

$$\mathbb{E}|H_{ij}|^2 \propto \left(\frac{|i-j|}{W}+1 \right)^{-1-\alpha}$$

with exponent $\alpha > -1$ and bandwidth $W \gg 1$. This interpolates between mean-field (Wigner-type) matrices and physically motivated models of long-range "hopping" in disordered quantum systems, connecting the study to quantum chaos, Anderson localization, and related universality phenomena.

PRBMs have received extensive attention in physics, especially due to their conjectured rich phase diagram, multifractality, and superdiffusive/diffusive transport predictions for wave packets, yet a rigorous mathematical account—especially away from rapidly decaying variance profiles—has remained incomplete. This work fills substantial gaps, particularly settling, on the delocalized side, conjectures for the scaling of localization lengths in all power-law decay regimes.

Main Results: Localization Length Scaling

The localization length $\ell$ characterizes the spatial scale of eigenstate delocalization. The central rigorous result is sharp lower bounds for $\ell$ (with high probability) for bulk eigenstates, providing verification of physics conjectures for all $\alpha > -1$. The refined scaling results are:

$\alpha$ Regime	Localization Length $\ell$	Scaling Behavior
$-1 < \alpha < 0$	$$\mathbb{E}|H_{ij}|^2 \propto \left(\frac{|i-j|}{W}+1 \right)^{-1-\alpha}$$0	Fully delocalized
$$\mathbb{E}|H_{ij}|^2 \propto \left(\frac{|i-j|}{W}+1 \right)^{-1-\alpha}$$1	$$\mathbb{E}|H_{ij}|^2 \propto \left(\frac{|i-j|}{W}+1 \right)^{-1-\alpha}$$2 for $$\mathbb{E}|H_{ij}|^2 \propto \left(\frac{|i-j|}{W}+1 \right)^{-1-\alpha}$$3	Stretched delocalization
$$\mathbb{E}|H_{ij}|^2 \propto \left(\frac{|i-j|}{W}+1 \right)^{-1-\alpha}$$4	$$\mathbb{E}|H_{ij}|^2 \propto \left(\frac{|i-j|}{W}+1 \right)^{-1-\alpha}$$5	Interpolating regime
$$\mathbb{E}|H_{ij}|^2 \propto \left(\frac{|i-j|}{W}+1 \right)^{-1-\alpha}$$6	$$\mathbb{E}|H_{ij}|^2 \propto \left(\frac{|i-j|}{W}+1 \right)^{-1-\alpha}$$7	1D RBM scaling

These results match the phase diagram conjectured in [Mirlin, Fyodorov et al., "Power-law random banded matrices"], confirming that for $$\mathbb{E}|H_{ij}|^2 \propto \left(\frac{|i-j|}{W}+1 \right)^{-1-\alpha}$$8 the system is fully delocalized (infinite localization length in the regime $$\mathbb{E}|H_{ij}|^2 \propto \left(\frac{|i-j|}{W}+1 \right)^{-1-\alpha}$$9), while for large $\alpha > -1$0 the scaling matches that of classical 1D RBMs.

Crucially, these bounds are matched to the conjectured dynamical scaling length of quantum return probabilities, and thus, are sharp up to negligible corrections across regimes.

Methodology and Technical Innovations

Resolvent Flow and Quantum Diffusion

The primary approach is via a dynamical analysis of matrix Brownian motion, interpolating from the zero matrix to the random matrix of interest, and tracking the resolvent (Green's function) flow. This permits control of local statistics down to the optimal scale $\alpha > -1$1 in the complex plane. At each time, the resolvent $\alpha > -1$2 is compared to the deterministic scalar limit from the semicircle law.

A key innovation is the use of so-called $\alpha > -1$3-variables—weighted sums of squared resolvent entries—which dominate the off-diagonal Green function decay and encode quantum diffusion. The control of these $\alpha > -1$4-variables, as well as higher-order matrix loops, is performed without relying on the fast decay properties exploited in earlier regular RBM works. Instead, detailed Fourier analysis of the associated variance matrix and the establishment of new dynamical self-consistent bounds are developed, especially crucial in the slow-decay ($\alpha > -1$5) high-dimensional analog regime.

Loop Contraction and Absence of Mean-Field Hierarchies

For $\alpha > -1$6, the loop hierarchy analysis (used in [Erdős et al.]) requires careful control of polynomial decay in variance matrices, which unlike the exponential case does not permit simple $\alpha > -1$7 norm bootstrap at each step. The authors introduce loop contraction inequalities reminiscent of combinatorial "tree" decay ideas used in high-dimensional Anderson models, enabling sharp control on all scales and ensuring the closure of the contraction scheme even without exponential cutoff.

In $\alpha > -1$8, the failure of a convolution inequality that allows higher-dimension analogies for regular RBMs forces the authors to develop a new truncation mechanism involving only loops and chains of order up to two, and to replace higher-order loops with refined dynamical estimates on $\alpha > -1$9-variables.

Quantum Unique Ergodicity and Universality

Beyond localization length, the analysis establishes quantum unique ergodicity (QUE) of bulk eigenstates whenever complete delocalization occurs ($W \gg 1$0), and proves bulk universality for local eigenvalue statistics. This is shown by tight control on various local Green function averages and exploiting comparison principles to relate PRBM fluctuations to those of GUE or GOE ensembles. The proofs build on recent advances in QUE and bulk universality for short-range RBM with variable profiles, but extend to the much more subtle non-exponential decay setting.

Numerical and Probabilistic Statements

The results are expressed as high-probability bounds (probability at least $W \gg 1$1 for arbitrary $W \gg 1$2 and large $W \gg 1$3) on the maximum $W \gg 1$4 norm of bulk eigenvectors, entrywise Green function deviations, $W \gg 1$5-variable behavior, and loop expectations. The theorems specify explicit polynomial or exponential scaling in all relevant parameters. The methods are robust, allowing precise quantitative QUE statements on appropriate spatial averaging scales for all $W \gg 1$6.

Implications and Extensions

These results rigorously validate the physical heuristics that PRBMs interpolate between mean-field and low-dimensional random operators (e.g., random Schrödinger with long-range hopping), quantify superdiffusive and multifractal transitions, and provide sharp thresholds for the Anderson transition as a function of $W \gg 1$7 and $W \gg 1$8.

Practically, the techniques open the door for:

• Analysis of higher-dimensional power-law RBM, where similar phase diagrams are expected but with different critical exponents.
• Treatment of non-Gaussian or non-Hermitian PRBM models (the method extends in part).
• Finer control of edge eigenvalue statistics and localization for spectral edge scaling.

The developed combinatorial and probabilistic machinery—especially the new dynamical truncation at order two and loop contraction bounds—are likely to be critical in future work on non-mean-field, slowly decaying variance random operators, including for models lacking translation invariance.

Conclusion

By introducing a rich set of new technical and probabilistic tools, this work rigorously derives nearly sharp lower bounds for localization lengths, proves QUE, and bulk universality in power-law RBMs for all $W \gg 1$9 across all asymptotic regimes. The results confirm longstanding physics conjectures on their phase diagram, demonstrate the power of dynamical resolvent analysis beyond the mean-field regime, and set the stage for further progress in long-range random matrix and operator theory (2604.12248).