Edge Universality for Inhomogeneous Random Matrices II: Markov Chain Comparison and Critical Statistics

This paper extends edge universality results for inhomogeneous random matrices to subcritical and critical sparsity regimes by establishing new Markov chain comparison conditions that link universal edge statistics to the comparability of underlying variance-profile Markov chains, thereby enabling the analysis of diverse models like random band and Hankel matrices beyond classical random matrix theory.

The Gist

Imagine you are a detective trying to predict the behavior of a massive, chaotic crowd. In the world of Random Matrix Theory, this "crowd" is a giant grid of numbers (a matrix) where the values are chosen somewhat randomly. Mathematicians have long been obsessed with the "edge" of this crowd—the largest numbers in the grid.

For decades, the rule was simple: if the crowd is well-mixed (everyone talks to everyone equally), the largest numbers follow a very specific, famous pattern called the Tracy-Widom distribution. It's like a universal law of physics for these numbers.

But real life isn't perfectly mixed. Sometimes, people only talk to their neighbors (like in a city block), or the rules of interaction change depending on where you are. This is the world of Inhomogeneous Random Matrices. For a long time, we didn't know if the "Universal Law" still applied when the crowd was messy and structured.

This paper, the second in a series, solves a major mystery: What happens when the crowd is not well-mixed?

Here is the breakdown of their discovery, using simple analogies:

1. The Core Idea: The "Chain Reaction" Metaphor

The authors realized that the behavior of these giant number grids isn't about the numbers themselves, but about how they are connected.

Imagine the matrix entries are people in a room.

• The Variance Profile: This is the "social map." It tells you who is likely to talk to whom. Some people are chatty (high variance), others are quiet.
• The Markov Chain: This is the path a rumor takes. If Person A tells a rumor to Person B, who does B tell? Does the rumor spread quickly to the whole room (mixing), or does it get stuck in a small corner?

The paper's big breakthrough is a principle they call "One CLT, One Statistics."

• The Analogy: Think of the "rumor" spreading through the social map as a random walk.
  • If the rumor spreads fast and everywhere (like a viral tweet), the crowd behaves like the classic, well-mixed model. The edge numbers follow the famous Tracy-Widom law.
  • If the rumor spreads slowly or gets stuck in loops (like a whispering game in a long hallway), the crowd behaves differently. The edge numbers follow new, exotic patterns that no one had seen before.

The Takeaway: You don't need to know the exact personality of every person in the room. You just need to know the shape of the social network (the Markov chain). If two different crowds have the same "rumor-spreading speed," their largest numbers will behave exactly the same, even if the people inside are totally different.

2. The Three Regimes: A Traffic Light System

The authors mapped out exactly what happens based on how "connected" the matrix is. They found three distinct phases, like traffic lights:

• 🟢 Green Light (Supercritical / Well-Mixed):

  • The Scene: The bandwidth (how far the "rumor" can jump) is huge. Everyone is connected to everyone.
  • The Result: The classic Tracy-Widom law wins. The system is "universal." It doesn't matter if the matrix is a bit messy; the mixing is so strong that it washes out the details.
  • Analogy: A crowded concert where everyone can hear the band. The music sounds the same regardless of where you stand.

• 🔴 Red Light (Subcritical / Isolated):

  • The Scene: The bandwidth is tiny. People only talk to their immediate neighbors. The "rumor" gets stuck.
  • The Result: The system breaks the universal law. The largest numbers become independent of each other (like a Poisson process). They act like random, isolated events.
  • Analogy: A quiet library where people only whisper to the person sitting next to them. The noise level is just a collection of random, unrelated whispers.

• 🟡 Yellow Light (Critical / The Transition Zone):

  • The Scene: This is the most exciting part. The bandwidth is just right—it's neither fully mixed nor fully isolated. It's the "Goldilocks" zone.
  • The Result: A new, hybrid universe emerges. The statistics here are a unique blend of the Green and Red lights. The authors discovered a whole new family of mathematical distributions that interpolate between the two.
  • Analogy: A busy subway station during rush hour. People are moving, but they are also bumping into each other in specific patterns. The flow isn't chaotic, but it isn't perfectly smooth either. It has its own unique rhythm.

3. Real-World Examples They Tested

To prove their theory, they applied it to three specific "toy models" (simplified versions of real-world problems):

1. Random Band Matrices (The City Block): Imagine a matrix where numbers only interact with neighbors within a certain distance (like a city grid). They showed that as you widen the "band" (allow people to talk to further neighbors), the system smoothly transitions from the "Red Light" (isolated) to the "Green Light" (universal), passing through the "Yellow Light" (critical) phase.
2. Wegner Orbital Model (The Block Party): Imagine a matrix made of blocks of numbers. Some blocks talk to each other, others don't. Depending on how strongly they are coupled, the system switches between behaving like a single giant block or a collection of independent mini-blocks.
3. Hankel-Profile Matrices (The Mirror): Imagine a matrix where the rules depend on the sum of the coordinates (like a mirror reflection). This creates a "bouncing" random walk. They found this creates a completely different type of critical statistics, determined by the geometry of the "mirror."

4. Why This Matters

Before this paper, mathematicians mostly studied "perfect" systems. If a system wasn't perfect, they often gave up or assumed it was too messy to understand.

This paper says: "It's not too messy; it's just a different kind of order."

They provided a universal toolkit. If you have a complex, messy system (like a neural network, a financial market, or a quantum material), you don't need to simulate every single number. You just need to analyze the "social map" (the variance profile).

• If the map mixes fast? Expect the standard Tracy-Widom law.
• If the map is slow or structured? You might be sitting on a brand-new, undiscovered statistical law.

Summary in One Sentence

This paper proves that the behavior of the largest numbers in a messy, structured grid is entirely dictated by how fast a "rumor" can travel through the grid's connections, revealing a rich landscape of new mathematical laws that exist between total chaos and perfect order.

Technical Summary

1. Problem Statement and Context

This paper addresses the spectral edge statistics of inhomogeneous random matrices (IRMs), specifically focusing on regimes where classical universality (the Tracy–Widom law) breaks down.

• Background: Classical Random Matrix Theory (RMT) establishes that for homogeneous ensembles (e.g., Wigner matrices), local eigenvalue statistics at the spectral edge follow the Tracy–Widom distribution, regardless of entry details.
• The Gap: For inhomogeneous matrices (where entry variances vary according to a profile), the behavior is less understood. While Part I of this series [LZ25] characterized the supercritical regime (where mixing is strong enough to restore Tracy–Widom universality), this paper tackles the subcritical and critical regimes.
• Core Question: When the variance profile induces strong locality or slow mixing, what governs the edge statistics? Do new, non-mean-field universality classes emerge?

2. Methodology

The authors develop a rigorous framework based on Markov chain comparison and diagrammatic expansions.

A. Markov Chain Comparison Framework

The central innovation is the "Short-to-Long Comparison" principle. The authors associate the variance profile $\Sigma$ of a random matrix with a Markov transition matrix $P = (\sigma_{ij}^2)$.

• Definition: Two IRMs are comparable if their underlying Markov chains satisfy specific Short-to-Long conditions:
  1. Average Upper Bound: The cumulative transition probabilities are bounded by a sequence $b_n$.
  2. Distance Control: The $\ell_1$ and $\ell_\infty$ distances between the transition probabilities of two chains decay sufficiently fast ($\epsilon_n, \delta_n$).
• Theorem 1.3 (Markov Chain Comparison Theorem): If two IRMs have comparable variance-profile Markov chains, their mixed moments of Chebyshev polynomials (which encode spectral statistics) are asymptotically equivalent. This reduces the problem of proving universality for complex matrices to comparing their underlying Markov chains to a canonical model.

B. Diagrammatic Analysis and Chebyshev Moments

• Chebyshev Expansion: The authors utilize the expansion of traces of powers of the matrix into Chebyshev polynomials of the second kind ($U_n$).
• Ribbon Graphs: Mixed moments are evaluated using Wick's formula, represented as sums over ribbon graphs (diagrams).
• Asymptotic Equivalence: The paper establishes that for diagrams in the subcritical/critical regimes, the diagram functions converge to limits determined by the Local Central Limit Theorem (LCLT) of the underlying Markov chain.
• Key Insight ("One CLT, One Statistics"): The paper posits that every specific LCLT for the variance-profile Markov chain dictates a corresponding universal pattern of edge statistics.

3. Key Contributions and Results

A. Phase Transitions in Random Band Matrices

The authors apply their framework to Random Band Matrices (RBMs) with power-law variance profiles (related to $\alpha$-stable distributions). They identify three distinct regimes based on the bandwidth $W$ and matrix size $N$:

1. Supercritical Regime ($W \gg N^{1 - 1/(3\alpha)}$): The system mixes rapidly. The edge statistics converge to the classical GOE/GUE Tracy–Widom (Airy) point process.
2. Subcritical Regime ($W \ll N^{1 - 1/(3\alpha)}$): Mixing is weak. The local eigenvalue statistics become asymptotically independent (Poisson-like). The $k$-point correlation measures factorize into a product of 1-point measures.
3. Critical Regime ($W \sim N^{1 - 1/(3\alpha)}$): A novel tricritical point process emerges. This process interpolates between the Airy and Poisson limits, governed by the $\alpha$-stable theta function.

B. Tricritical Analysis with Spikes

The paper extends the analysis to deformed RBMs with finite-rank perturbations (spikes).

• Tricritical Point Process: When the bandwidth, perturbation strength, and spike location are all critically scaled, a new universal point process arises.
• Interpolation: This process unifies the BBP transition (spike emergence in mean-field) with the variance-profile transition (localization/delocalization). As parameters vary, the statistics continuously morph from Airy to Poisson to spike-dominated behaviors.

C. Applications to Prototypical Models

The Markov chain comparison theorem is applied to derive edge statistics for:

1. Wegner Orbital Model: A block-structured model interpolating between localized blocks and delocalized band matrices. The authors identify a sequence of regimes: Frozen (independent blocks) $\to$ Skellam regime (discrete Poisson difference) $\to$ Diffusive (Gaussian) $\to$ Mixing. The Skellam point process is identified as a new universal class.
2. Hankel-Profile Matrices: Matrices where variance depends on $x+y$ (reflection symmetry). The underlying Markov chain is an "alternating walk." This leads to a distinct class of critical edge statistics determined by global geometry and boundary symmetries, differing from the translation-invariant Toeplitz case.

D. Deviation Inequalities

The paper derives non-asymptotic deviation inequalities for the spectral norm $\|X\|$.

• It shows that for inhomogeneous matrices, the fluctuation scale at the edge is not always $N^{-2/3}$ (Tracy–Widom) but depends on the geometry of the variance profile.
• For power-law band matrices, the fluctuation scale is determined by the power-law index $\alpha$, specifically scaling as $W^{-2\alpha/(3\alpha-1)}$.

4. Significance and Implications

• Beyond Mean-Field Universality: The paper fundamentally challenges the notion that Tracy–Widom is the only edge universality class. It demonstrates that inhomogeneity is not just a perturbation but can generate entirely new universality classes (e.g., Skellam, critical theta-function processes) when the underlying Markov chain mixes slowly.
• Universal Reduction Principle: The "One CLT, One Statistics" principle provides a powerful, systematic tool. Instead of analyzing the complex matrix entries directly, one only needs to analyze the Local Central Limit Theorem of the associated Markov chain to predict the spectral edge behavior.
• Phase Diagram Completion: The work completes the phase diagram for edge statistics in inhomogeneous matrices, bridging the gap between the localized (Poisson) and delocalized (Airy) regimes and characterizing the critical transition points.
• Physical Relevance: These results are crucial for understanding disordered systems (Anderson localization), quantum chaos, and high-dimensional data where correlations are structured and non-uniform.

5. Conclusion

Liu and Zou establish a comprehensive theory for the edge statistics of inhomogeneous random matrices. By linking spectral properties to the mixing rates of Markov chains, they prove that the "zoo" of edge statistics is far richer than previously thought, containing critical and non-universal phenomena that are robust against microscopic entry details but sensitive to the macroscopic geometric structure of the variance profile.