Renormalization group for spectral collapse in random matrices with power-law variance profiles

This paper proposes a renormalization group framework that utilizes a size-dependent normalization to collapse eigenvalue densities of random matrix ensembles with power-law variance profiles, deriving fixed-point equations and Beta functions to demonstrate spectral collapse across different system sizes.

The Gist

The Big Problem: Comparing Apples to Oranges

Imagine you are studying a complex system, like a city's traffic network, a brain's neural connections, or a stock market. You collect data and turn it into a giant grid of numbers (a matrix) to see how the different parts interact.

The problem is that these systems come in different sizes. One study might look at 100 neurons, while another looks at 10,000. When you look at the "spectrum" (a map of the system's stability and behavior) of the small system and the big system, they look completely different. The big one is huge and spread out; the small one is tiny and cramped.

It's like trying to compare a photo of a single ant to a photo of an entire anthill. If you just look at the raw pictures, you can't tell if the ants are behaving differently or if the difference is just because one picture is zoomed in and the other is zoomed out.

The Solution: A "Renormalization Group" (RG) Recipe

The authors propose a new way to compare these systems, borrowing a tool from physics called the Renormalization Group (RG).

Think of the RG approach as a universal zoom lens.

1. The Goal: We want to see the "shape" of the system's behavior, regardless of how many parts (N) the system has.
2. The Trick: Instead of keeping the picture size fixed, we adjust the "zoom" (a normalization factor) as the system gets bigger. We force the "average energy" or "bandwidth" of the system to stay the same size, no matter how many ants or neurons we add.
3. The Result: When you apply this zoom, the messy, different-sized spectra "collapse" onto a single, smooth curve. Suddenly, the 100-neuron system and the 10,000-neuron system look like they are following the exact same rule.

The Two Experiments: Wigner and Wishart

To test this recipe, the authors used two classic mathematical models that act like "test tubes" for complex systems:

• The Wigner Ensemble: Think of this as a web where every node is connected to every other node with a certain strength.
• The Wishart Ensemble: Think of this as a dataset where you have rows of observations (like daily stock prices) and columns of variables.

In both cases, they introduced a twist: Power-Law Variance.
Imagine the connections in the web aren't all equal strength. Instead, the connections near the "start" of the list are very strong, and they get weaker and weaker as you go down the list, following a specific mathematical rule (a power law). This mimics real life, where a few "super-connections" often dominate a system (like a few famous genes or a few highly connected people in a social network).

The "Beta Function": The Flow of the Zoom

The authors didn't just find a zoom lens; they figured out exactly how the zoom needs to change as the system grows. They call this the Beta function.

Imagine you are walking down a hill (the RG flow):

• Steep Hill (Relevant): If the power-law exponent is low, the "zoom" changes rapidly as you add more data. The system is very sensitive to its size.
• Flat Hill (Marginal): At a specific "sweet spot" (exponent = 0.5), the zoom barely changes. The system is in a delicate balance.
• Dead Flat (Irrelevant): If the exponent is high, the zoom stops changing almost entirely. The system becomes so dominated by the few strong connections at the top that adding more weak connections at the bottom doesn't change the overall picture.

What They Found

1. The Collapse Works: When they applied their "running zoom" to computer simulations, the jagged, different-sized spectra lined up perfectly into a single, smooth curve.
2. It's Robust: It didn't matter if the numbers in the matrix were generated by a bell curve (Gaussian), a coin flip (Rademacher), or other distributions. As long as the "power-law" structure was there, the collapse happened.
3. The Math Checks Out: They derived complex equations (fixed-point equations) to predict what the curve should look like. Their computer simulations matched these predictions almost perfectly.

Why This Matters (According to the Paper)

The paper argues that this method gives us a way to compare complex systems of different sizes on an "equal footing."

• Stability: If you know the "collapsed" shape of a system, you can predict when it will become unstable (like a bridge collapsing or a neural network going haywire) without needing to know the exact size of the system.
• Universal Rules: It suggests that despite the chaos of complex systems, there are universal rules governing how they behave, provided you look at them through the right "RG lens."

In short: The paper provides a mathematical "universal translator" that lets us compare small and large complex systems by adjusting the scale, revealing that underneath the size differences, they often follow the same fundamental patterns.

Technical Summary

1. Problem Statement

In the analysis of complex systems (e.g., neural networks, gene correlations, material couplings), data is often represented by large matrices of size $N \times N$. A key challenge arises when comparing spectral distributions (eigenvalue densities) across different system sizes $N$. Standard Random Matrix Theory (RMT) provides asymptotic normalizations (like Wigner's variance scaling or Marčenko–Pastur aspect ratios) that work for specific regimes (bulk or microscopic edges) but fail to provide a general rule for collapsing entire spectral curves across varying $N$.

Specifically, the paper addresses ensembles where the matrix entries have power-law variance profiles (heterogeneous variances decaying as $j^{-\alpha}$). In such systems, the spectral support and density shape change drastically with $N$ under standard normalization, making direct comparison impossible. The goal is to develop a method to extract size-independent system properties by finding a "natural spectral scale" that allows spectra from different $N$ to collapse onto a single universal curve.

2. Methodology

The author proposes a Renormalization Group (RG) approach adapted for Random Matrix Theory. The core methodology involves:

• Running Normalization: Instead of fixing the normalization factor $\gamma$ (which scales the matrix entries), $\gamma$ is allowed to "run" with the system size $N$. The RG condition is defined by fixing a specific spectral moment (e.g., the second moment $m_2$ for Wigner ensembles or the first moment $m_1$ for Wishart ensembles) to a constant value $m^*$.
• Self-Consistent Equations: The author derives deterministic fixed-point equations for the resolvent (Green's function) of the finite-$N$ ensembles using the Matrix Dyson Equation (MDE) and cavity methods. These equations account for the power-law variance weights $w_j = j^{-\alpha}$.
• Decimation Scheme: A coarse-graining procedure is defined via matrix decimation. This involves removing a fraction of matrix indices corresponding to the smallest variance weights (largest indices) to simulate a reduction in system size. The change in the coupling constant under this decimation defines the RG flow.
• Beta Function Analysis: The flow of the coupling constant is characterized by a Beta function, $\beta_{RG}$, which determines whether the variance heterogeneity is relevant, marginal, or irrelevant under the RG flow.
• Callan–Symanzik Equation: In the large-$N$ limit, the invariance of the spectral observables under scale changes is formalized via a Callan–Symanzik equation, linking the change in scale to the change in the running coupling.

3. Key Contributions

A. Theoretical Framework

• RG Formulation for RMT: The paper establishes a rigorous operational framework to apply RG concepts to random matrices, specifically for collapsing spectral densities across sizes.
• Fixed-Point Derivations: It derives self-consistent equations for the resolvent of two generalized ensembles:
  1. Wigner-type: Symmetric matrices with power-law variance profiles.
  2. Wishart-type: Sample covariance matrices with power-law heterogeneity in rows.
• Continuum Limits: The paper derives integral equations for the large-$N$ continuum resolvent, showing how the discrete fixed-point equations converge to rank-one self-energy kernels.

B. Identification of RG Regimes

The analysis reveals three distinct regimes based on the power-law exponent $\alpha$:

1. Relevant ($\alpha < 1/2$): The coupling grows under decimation (or shrinks as $N$ increases). The normalization $\gamma(N)$ scales as a power law $N^{1-2\alpha}$. The spectral density retains a non-trivial bulk structure.
2. Marginal ($\alpha = 1/2$): The system sits at a non-trivial RG fixed point. The normalization scales logarithmically (or remains constant depending on the ensemble), and the flow exhibits logarithmic corrections. The coupling is scale-stationary but non-zero.
3. Irrelevant ($\alpha > 1/2$): The coupling weakens under decimation. The normalization saturates to a constant. The bulk eigenvalue density collapses to a delta mass at zero, leaving only a finite number of non-self-averaging outliers associated with small indices.

C. Spectral Collapse and Universality

• Spectral Collapse: By applying the derived running normalization $\gamma(N)$, the paper demonstrates that eigenvalue densities from simulations of different sizes $N$ collapse perfectly onto a single curve.
• Universality: The paper provides numerical evidence that this collapse is robust across different entry distributions (Gaussian, Rademacher, Student-t), provided the variance is finite. This suggests the spectral density is insensitive to microscopic details once the spectral scale is fixed.

4. Key Results

• Analytical Solutions: The paper provides exact solutions for the running normalization $\gamma(N)$:
  • For Wigner: $\gamma(N) \propto N^{1-2\alpha}$ (for $\alpha < 1$).
  • For Wishart: $\gamma^2(T) \propto T^{1-2\alpha}$ (for $\alpha < 1/2$).
• Beta Functions:
  • Wigner: $\beta_{RG} = 1 - 2\alpha$.
  • Wishart: $\beta_{RG} = -(1 - 2\alpha)$ (sign difference due to definition of scale).
• Simulation Validation: Extensive numerical simulations confirm that:
  • The fixed-point equations accurately predict eigenvalue densities for finite $N$.
  • Raw spectra (with $\gamma=1$) diverge significantly across sizes.
  • Collapsed spectra (with running $\gamma(N)$) overlap perfectly, validating the RG hypothesis.
• Edge Behavior: The paper notes that while the bulk collapses, edge behaviors (like Tracy-Widom fluctuations) may show non-uniform convergence, and outliers in the $\alpha > 1/2$ regime do not self-average.

5. Significance and Implications

• Data Analysis Tool: This approach offers a practical method for analyzing complex systems where data is collected at varying scales (e.g., different numbers of neurons in a recording, different numbers of genes in a dataset). It allows researchers to compare stability, variability, and characteristic scales directly without being misled by system size artifacts.
• Stability and Dynamics: The collapsed spectral distribution allows for the definition of size-independent dynamical thresholds. For instance, stability boundaries (determined by the spectral edge) and relaxation times (determined by the spectral gap) can be set at a reference size and applied universally to systems of any size.
• Theoretical Bridge: It bridges the gap between statistical mechanics (RG flow) and high-dimensional statistics (RMT), providing a new perspective on how heterogeneity in complex networks affects macroscopic spectral properties.
• Future Directions: The paper suggests that this framework can be extended to other ensembles and structural perturbations, potentially defining the "domain of attraction" for collapsed spectral distributions in broader classes of complex systems.

In summary, Fleig's work provides a powerful, analytically tractable, and numerically verified method to normalize and compare random matrix spectra in heterogeneous systems, transforming size-dependent data into size-invariant physical insights.