On the spectral radius of the ratio of Girko matrices

This paper proves that the spectral radius of the ratio of two independent Girko matrices, when scaled by the square root of the dimension, converges to a universal heavy-tailed distribution as the dimension tends to infinity, a result established through Girko Hermitization and local law estimates that reveals the model's remarkable mathematical accessibility compared to single Girko matrices.

The Gist

Imagine you are standing in a vast, crowded room filled with people. In the world of mathematics, this room is a "random matrix," and the people are numbers (specifically, complex numbers).

Usually, mathematicians study what happens when you look at just one of these rooms. They ask: "If I pick a random person, how far are they from the center of the room?" This is called the spectral radius. It's a bit like asking, "Who is the loudest person in the room?" or "Who is standing the furthest away from the door?"

For a long time, figuring out the behavior of the "loudest person" in these random rooms was incredibly difficult, especially when the people weren't behaving nicely (when the numbers had "heavy tails," meaning a few wild outliers could show up).

The Big Idea of This Paper

The authors of this paper decided to look at a different, slightly stranger scenario. Instead of looking at one room, they looked at the ratio of two rooms.

Imagine you have two identical, chaotic crowds (let's call them Crowd A and Crowd B). You take every person in Crowd A and divide them by the person standing in the exact same spot in Crowd B.

• The Result: You get a new, third crowd.
• The Problem: This new crowd is messy. The people are no longer independent; they are tangled up with each other. It's like a dance where everyone is holding hands with someone from the other group.

The authors wanted to know: "As these crowds get infinitely huge, how far does the furthest person in this new, mixed-up crowd stand from the center?"

The Surprising Discovery: The "Spherical" Secret

Here is the magic trick the authors found:

1. The Inversion Trick:
   Imagine the room is actually a giant sphere (like a beach ball). The center of the room is the "South Pole," and the furthest edge is the "North Pole."
   The authors realized that if you flip the beach ball inside out (an inversion), the person standing at the North Pole (the furthest away) swaps places with the person at the South Pole (the closest to the center).

   • Why this matters: It is much easier to study the person standing right next to the door (the South Pole) than the person way out in the middle of the ocean (the North Pole). By flipping the problem, they turned a "hard" question about the edge into an "easy" question about the center.

2. The Universal Pattern:
   They discovered that no matter how the original crowds behaved (as long as they weren't too crazy), the final mixed crowd always settled into the same pattern.

   • The Analogy: Imagine pouring sand from two different types of buckets (one with fine sand, one with coarse sand) into a funnel. No matter what the sand looked like at the start, once it passes through the funnel and settles, it always forms the exact same shape.
   • In math terms, this shape is called the Spherical Ensemble. It's a "heavy-tailed" distribution, meaning there's a higher chance of finding a "wild" outlier far away than in normal distributions.

3. The "Gaussian" Shortcut:
   To prove this, they used a clever mathematical shortcut. They said, "Let's pretend the original crowds were made of perfect, bell-curve Gaussian numbers (the most 'normal' kind of randomness)."

   • They proved that if the result works for the perfect Gaussian case, it works for almost any other case, provided the "moments" (the basic statistical averages) match up to the fourth degree.
   • It's like saying: "If I can predict the weather using a perfect, theoretical model, and the real world is only slightly different, my prediction is still accurate."

The "Heavy Tail" Surprise

The most exciting part of their finding is about the "furthest person" (the spectral radius).

• Normal Expectation: In most random systems, the furthest person stays within a predictable distance.
• The Reality Here: Because this is a ratio of two random matrices, the "furthest person" can be extremely far away. The paper shows that the probability of finding someone very far away follows a "heavy tail."
  • Metaphor: In a normal crowd, if you walk 100 steps, you might find someone. If you walk 1,000 steps, it's very unlikely. But in this specific "ratio crowd," if you walk 1,000 steps, you might still find someone, and the chance of finding them drops off very slowly. It's like a party where a few people are always wandering off to the edge of the universe.

Why Should You Care?

This paper is a breakthrough because:

1. It's Universal: It shows that this wild, heavy-tailed behavior isn't a fluke of a specific math problem; it's a fundamental law of nature for this type of ratio.
2. It's Easier Than You Think: The authors found that studying the ratio of two random matrices is actually mathematically easier than studying just one! By flipping the problem (inversion), they turned a nightmare into a solvable puzzle.
3. Real World Applications: While this sounds abstract, "ratios of random matrices" show up in physics (quantum mechanics), signal processing (filtering noise), and even in understanding how complex systems like the internet or financial markets might behave when two chaotic forces interact.

In a Nutshell:
The authors took two chaotic, random crowds, mixed them together, and found that the resulting chaos follows a beautiful, predictable, and "heavy-tailed" pattern. They proved this by flipping the problem upside down (literally, on a sphere) and realizing that the edge of the system behaves just like the center. It's a story about finding order in chaos by looking at it from a different angle.

Technical Summary

1. Problem Statement and Context

The paper investigates the asymptotic behavior of the spectral radius (the maximum modulus of eigenvalues) of a specific class of non-normal random matrices: the ratio of two independent Girko matrices.

• Model: Let $A$ and $B$ be independent $n \times n$ complex random matrices with independent and identically distributed (i.i.d.) entries having mean zero and unit variance. The object of study is the random matrix $M = AB^{-1}$.
• Challenges:
  • Unlike single Girko matrices (where the spectrum converges to the uniform distribution on the unit disk), the ratio $AB^{-1}$ results in a heavy-tailed distribution.
  • The entries of $M$ are dependent, making standard random matrix theory techniques difficult to apply directly.
  • The spectral radius of such heavy-tailed matrices typically diverges or behaves differently than in the "light-tailed" (Gaussian) case.
• Goal: To establish the universality of the high-dimensional fluctuations of the spectral radius of $M$ as $n \to \infty$, specifically showing that the behavior depends only on the first few moments of the entry distribution, not the specific distribution itself.

2. Methodology

The authors employ a sophisticated combination of probabilistic and analytic techniques:

A. Inversion Symmetry and Geometric Interpretation

• Invariance: A crucial observation is that the model $M = AB^{-1}$ is invariant in law under inversion ($M^{-1} \stackrel{d}{=} M$ if $A, B$ are identically distributed, or with swapped laws otherwise).
• Spherical Ensemble: In the Gaussian case (where $A, B$ are complex Ginibre matrices), $M$ is known as the Spherical Ensemble. Its spectrum corresponds to a planar Coulomb gas with a specific potential.
• Stereographic Projection: The authors utilize the connection between the planar spectrum and the two-sphere ($S^2$) via stereographic projection. The inversion symmetry on the plane corresponds to a rotation on the sphere that swaps the origin ($0$) and infinity ($\infty$). This allows the study of the spectral radius (behavior near $\infty$) to be transformed into the study of the inner spectral radius (behavior near $0$).

B. Girko Hermitization

To handle the non-Hermitian nature of $M$, the authors use Girko's Hermitization trick:

• The eigenvalues of $M$ are related to the singular values of the matrix $A - zB$.
• Specifically, $\log|\det(M - z)|$ is expressed via the logarithmic potential of the empirical spectral distribution, which is converted into an integral involving the resolvent (Green's function) of the Hermitized matrix $H_z = \begin{pmatrix} 0 & A-zB \\ (A-zB)^* & 0 \end{pmatrix}$.

C. Local Law and Fourth Moment Matching

• Local Law: The proof relies on precise local law estimates for the Green's function of Wigner-type matrices (specifically $A-zB$) to control the density of states near the origin.
• Fourth Moment Theorem: The authors establish a replacement principle (Theorem 1.6). They prove that the local eigenvalue statistics of a general Girko ratio match those of the Gaussian (Ginibre) ratio, provided the entry distributions satisfy a fourth moment matching condition (Condition C2).
• Ornstein-Uhlenbeck Flow: To prove universality, they employ a continuous interpolation (Ornstein-Uhlenbeck process) between the general Girko matrices and the Gaussian Ginibre matrices. Using Itô's formula and cumulant expansions, they show that the difference in expectations of test functions vanishes as the flow evolves, relying on the cancellation of terms up to the fourth order.

D. Determinantal Point Processes

• In the Gaussian case, the spectrum of the Spherical Ensemble forms a determinantal point process.
• The authors use the convergence of the kernel of this process to the infinite Ginibre ensemble kernel to derive the limiting distribution of the spectral radius.

3. Key Contributions and Results

Theorem 1.2: Local Universality (From Ratio to Infinite Ginibre)

The paper proves that the local eigenvalue statistics of the ratio $M = AB^{-1}$, when zoomed in near any point $\lambda_0 \in \mathbb{C}$ with a scaling factor of $\sqrt{n}$, converge in distribution to the infinite Ginibre determinantal point process ($Gin_\infty$).

• This holds for general Girko matrices satisfying bounded density, finite moments, and fourth-moment matching.
• The limit kernel is $K(z, w) = \sqrt{\gamma(z)\gamma(w)} e^{zw}$, where $\gamma(z) = \frac{1}{\pi}e^{-|z|^2}$.

Theorem 1.3: Spectral Radius Convergence

This is the main result regarding the spectral radius $\rho_{\max}(M)$.

• Scaling: The spectral radius scales as $\sqrt{n}$.
• Limiting Distribution: The normalized spectral radius $\frac{\rho_{\max}(M)}{\sqrt{n}}$ converges in distribution to a random variable $R_\infty = \frac{1}{\sqrt{\min_{k \ge 1} \gamma_k}}$, where $\gamma_k \sim \text{Gamma}(k, 1)$ are independent Gamma variables.
• Heavy Tails: The limiting distribution is heavy-tailed, with a tail decay of $P(R_\infty > x) \sim x^{-2}$. This contrasts with the Gumbel fluctuations seen in single Ginibre matrices.
• Inner Radius: Similarly, the inner spectral radius $\rho_{\min}(M)$ converges to $\sqrt{\min_{k \ge 1} \gamma_k}$.

Theorem 1.6: Replacement Principle

A technical cornerstone of the paper, this theorem states that for any smooth test function $f$ and any point $z_0$, the expectation of functionals of the local spectral measure of a general Girko ratio $M$ is asymptotically identical to that of the Gaussian Spherical Ensemble $M_{Gin}$. This establishes the universality of the high-dimensional fluctuations.

4. Significance and Implications

1. Universality in Heavy-Tailed Models: The paper demonstrates that despite the heavy-tailed nature of the ratio $AB^{-1}$ (which creates dependent entries), the fluctuations of the spectral radius are universal and governed by the same laws as the integrable Gaussian case.
2. Accessibility of the Ratio Model: The authors note a counter-intuitive finding: proving universality for the ratio of matrices is mathematically more accessible than for a single Girko matrix. This is largely due to the inversion symmetry of the ratio model, which equates the "edge" (spectral radius) with the "bulk" (local behavior near zero), allowing the use of powerful local law techniques.
3. Connection to Free Probability: The results bridge the gap between random matrix theory and Free Probability Theory, interpreting the limit via Brown measures and the spherical ensemble as a "matrix Cauchy law."
4. Methodological Advancement: The work extends the "fourth moment theorem" and "Ornstein-Uhlenbeck flow" techniques to the ratio of non-Hermitian matrices, a setting where standard tools often fail due to the lack of independence in the entries of the ratio.

5. Conclusion

The paper provides a rigorous proof that the spectral radius of the ratio of two independent Girko matrices, when scaled by $\sqrt{n}$, converges to a universal heavy-tailed distribution. By leveraging the geometric symmetry of the spherical ensemble and advanced local law techniques, the authors show that the complex fluctuations of this non-normal, heavy-tailed system are entirely determined by the first four moments of the entry distribution, mirroring the behavior of the Gaussian Spherical Ensemble.