Banded non-Hermitian random matrices, neural networks, and eigenvalue degeneracies

This paper investigates two-banded, non-Hermitian random matrices inspired by sparse neural networks, revealing how the competition between random sign disorder and directional bias drives distinct delocalization transitions and creates complex spectral structures, including loops of extended states and specific eigenvalue degeneracies, in both SSH chain and ladder models.

The Gist

Imagine a giant, circular train track made of two parallel rails. This track represents a simplified model of a neural network (a brain-like system) where the "stations" are neurons. In this specific model, the connections between these stations have two special rules:

1. The "Dale's Law" Rule: Every station is either purely an "exciter" (pushing the train forward) or an "inhibitor" (braking the train). They are assigned randomly, like flipping a coin for every station. This creates a chaotic, disordered environment.
2. The "Directional Bias" Rule: There is a wind blowing along the track. This wind makes it easier to move in one direction and harder in the other.

The scientists in this paper are studying what happens to the "energy" (or activity) of this train system when you mix the chaos of the random stations with the wind of the directional bias. They looked at two different track layouts: a SSH Chain (a zig-zagging double track) and a Ladder (two straight parallel tracks connected by rungs).

Here is what they found, explained through simple analogies:

1. The Battle Between Chaos and Wind

Think of the random stations as potholes that trap the train. If there is no wind, the train gets stuck in these potholes everywhere. The energy is "localized," meaning it stays stuck in small, specific spots and doesn't travel.

However, when you turn up the wind (the directional bias), it starts to push the train out of the potholes.

• The Result: The energy begins to "delocalize," meaning it spreads out and travels freely along the track.
• The Shape: As the wind gets stronger, the places where the energy can travel freely form loops (like rings) in a complex mathematical map. The places where the energy is still stuck (localized) sit either inside or outside these rings.

2. The Two Different Tracks Behave Differently

Even though both tracks have the same rules (random potholes + wind), they react to the wind in very different ways.

The SSH Chain (The Zig-Zag Track):

• The "Magic Moment": As you increase the wind, the four separate rings of traveling energy slowly expand. At a very specific wind speed, all four rings crash into each other at the center and merge into one big ring.
• The "Exceptional Point": The paper calls this crash an Exceptional Point. Imagine a magic trick where two different things (like a red ball and a blue ball) suddenly become the exact same object and lose their individual identities. At this specific wind speed, the system's behavior changes drastically, and the "holes" in the middle of the rings disappear.

The Ladder Model (The Parallel Tracks):

• The "Two-Stage" Reaction: This track is more stubborn. As you increase the wind, the energy starts to spread, but it doesn't merge everything at once.
  • Stage 1: First, the outer rings of energy expand, but they leave a core of "stuck" energy in the middle. The rings grow, but they don't swallow the center yet.
  • Stage 2: Only when the wind gets very strong (past a specific "Diabolic Point") does a second ring of traveling energy appear from the center, pushing the stuck energy out.
• The "Diabolic Point": The paper calls the moment where the two rings meet a Diabolic Point. Unlike the SSH chain, the two things merging here remain distinct (like two separate balls touching but not becoming one). It's a "point of degeneracy" where the energy levels match, but the underlying structure stays separate.

3. Predicting the Path

The scientists didn't just watch the trains; they built a mathematical "speedometer" called the Lyapunov exponent.

• Think of this as a map showing how fast the wind can push a train out of a pothole.
• They found that the rings of traveling energy always form exactly where the "wind speed" matches the "pothole strength." If you know the math of the potholes, you can predict exactly where the traveling rings will appear, and their computer simulations proved this was 100% accurate.

4. What Happens at the Edges? (Open Boundaries)

So far, we assumed the track was a perfect circle (no start or end). But what if the track has a start and an end?

• The Skin Effect: In these non-Hermitian systems, the wind doesn't just push the train; it pushes all the trains to pile up at one end of the track (the "skin").
• If the wind blows to the right, all the trains pile up on the right wall. If it blows left, they pile up on the left. This happens even if the track is full of random potholes.
• The SSH Surprise: For the zig-zag SSH track, if the wind is weak and the potholes are arranged a certain way, the trains don't just pile up; they get stuck in "edge modes" right at the very ends of the track, similar to how a special type of rope knot holds tight only at the ends.

Summary

The paper explores how chaos (random connections) and direction (biased flow) fight each other in a neural network model.

• Chaos tries to trap energy in small spots.
• Direction tries to free the energy and make it flow.
• The SSH Chain and the Ladder are two different ways these forces interact. The Chain merges its flow patterns all at once (an "Exceptional Point"), while the Ladder does it in two distinct steps (a "Diabolic Point").
• The scientists proved that they can predict exactly where the energy will flow using a mathematical "wind speed" calculation, and they showed that if the track has ends, the energy will inevitably pile up at the edges.

Technical Summary

Technical Summary: Banded Non-Hermitian Random Matrices, Neural Networks, and Eigenvalue Degeneracies

Problem Statement
This work investigates the spectral and localization properties of two-banded, non-Hermitian random matrices inspired by sparse neural networks with circular, one-dimensional topology. The study focuses on the interplay between two specific features inherent to biological neural networks: Dale's Law (constraining neurons to be purely excitatory or inhibitory, leading to random-sign disorder in hopping terms) and sparse connectivity (modeled here as a lattice structure). The authors aim to understand how these constraints, combined with a non-Hermitian directional bias (asymmetric hopping), affect eigenstate localization and eigenvalue degeneracies in multi-banded systems. The study contrasts two paradigmatic models: the non-Hermitian Su-Schrieffer-Heeger (SSH) chain and a non-Hermitian ladder model.

Methodology
The authors employ a combination of analytical techniques and numerical simulations:

1. Model Construction: Two models are defined with periodic boundary conditions (and later open boundaries). Both incorporate random-sign disorder ($\sigma \in \{+1, -1\}$) on the hopping amplitudes to satisfy Dale's Law and a directional bias parameter $g$ that breaks Hermiticity.
   • SSH Chain: Features alternating intracell ($t_+$) and intercell ($t_-$) hoppings with bias $g$.
   • Ladder Model: Consists of two coupled Hatano-Nelson chains with symmetric rung hopping ($u$) and biased intra-chain hopping ($t$).
2. Spectral Analysis: The authors analyze the eigenspectra in the complex plane, focusing on the distribution of eigenvalues and the inverse participation ratio (IPR) to distinguish between localized and extended states. For non-Hermitian systems, a gauge-invariant IPR involving both left and right eigenvectors is utilized.
3. Transfer Matrix Formalism: To predict the contours of delocalization, the eigenvalue problem is recast into recursion relations. The Lyapunov exponents ($\gamma$) associated with products of random transfer matrices are calculated. The delocalization criterion is established as $g = \gamma = \xi^{-1}$, where $\xi$ is the localization length.
4. Singular Value Analysis: The authors utilize the relationship $M = H \cdot \Sigma$ (where $H$ is the pure non-Hermitian model and $\Sigma$ is a diagonal matrix of random signs) to show that the singular values of the disordered matrix are identical to those of the pure model. This provides rigorous bounds on the eigenvalue moduli in the complex plane.

Key Contributions and Results

• Persistence of Degeneracies under Disorder:

  • SSH Chain: In the absence of disorder, tuning the bias $g$ leads to an exceptional point (where eigenvalues and eigenvectors coalesce) at a critical value $g_c$. The authors demonstrate that this exceptional point persists even in the presence of random-sign disorder, manifesting as a two-fold degenerate zero eigenvalue with a single eigenvector.
  • Ladder Model: In contrast, the pure ladder model exhibits a diabolic point (degenerate eigenvalues with distinct eigenvectors) at $g_c$. This diabolic point also survives the introduction of disorder, maintaining a two-fold degenerate zero eigenvalue with a complete set of eigenvectors.

• Distinct Delocalization Phenomena:

  • SSH Chain: As the directional bias $g$ increases, states delocalize starting from within the bands, forming four loops of extended states in the complex plane. These loops expand and eventually merge at the origin when $g = g_c$, creating a hole in the spectrum. For $g > g_c$, all states eventually delocalize.
  • Ladder Model: The delocalization process occurs in two distinct stages. Initially, for $g < g_c$, extended states form four loops that do not open a hole in the spectrum (localized states remain inside). As $g$ exceeds $g_c$, a second set of extended states emerges from the origin, resulting in two concentric loops of extended states separated by a region of localized states. Only for sufficiently large $g$ do all states delocalize.

• Prediction of Extended State Contours:
  The precise contours on which extended states reside are accurately predicted by the Lyapunov exponents derived from the transfer matrix formalism. Numerical diagonalization confirms that extended states lie on contours where the Lyapunov exponent $\gamma(\lambda)$ equals the bias parameter $g$.

• Open Boundary Conditions and Skin Effect:
  Under open boundary conditions, both models exhibit the non-Hermitian skin effect, where eigenstates accumulate at the boundaries.

  • For the ladder model, the eigenvalues remain independent of $g$ (identical to the $g=0$ case), but eigenstates localize at the left (right) boundary for $g > 0$ ($g < 0$).
  • For the SSH chain, eigenvalues are also independent of $g$, but the system can support topological edge states with near-zero eigenvalues if the hopping parameters satisfy $t_+ < t_-$, a feature reminiscent of the clean SSH chain.

Significance
The paper establishes that the interplay between random-sign disorder (Dale's Law) and directional bias in multi-banded non-Hermitian systems leads to rich spectral structures that differ qualitatively from single-band models. The work highlights that the nature of eigenvalue degeneracies (exceptional vs. diabolic points) in the clean limit dictates the topology of the delocalization transition in the disordered system. Specifically, the ladder model's two-stage delocalization and the formation of nested loops of extended states represent a new phenomenon not observed in previously studied single-band non-Hermitian random matrices. The ability to predict these complex spectral contours using Lyapunov exponents provides a robust theoretical framework for understanding localization in sparse, non-Hermitian biological networks. The results suggest that the spatial substructure of neural clusters (modeled here as unit cells) plays a critical role in determining the stability and delocalization properties of network dynamics.