Numerical ranges of non-normal random matrices: elliptic Ginibre and non-Hermitian Wishart ensembles

This paper characterizes the limiting geometry of numerical ranges for fundamental non-Hermitian random matrix ensembles, demonstrating that elliptic Ginibre and chiral variants converge to ellipses while non-Hermitian Wishart matrices and products of multiple elliptic Ginibre matrices form non-elliptic envelopes.

The Gist

Imagine you are looking at a complex machine, like a giant, chaotic orchestra of numbers. In the world of mathematics, these machines are called matrices.

For a long time, mathematicians mostly studied "well-behaved" machines (called normal matrices). These are like a perfectly tuned piano: if you know the notes (the eigenvalues or spectrum), you know everything about the sound. The notes tell you the whole story.

But in the real world, things are often messy and "non-normal." These are like a jazz band where the instruments are slightly out of sync, or a weather system where a tiny change in wind speed creates a hurricane. In these cases, knowing just the "notes" (the eigenvalues) isn't enough. You need to understand the Numerical Range.

What is the "Numerical Range"?

Think of the Numerical Range as the "shadow" or the "safety zone" of the machine.

• If you shine a light on a spinning, wobbling top, the shadow it casts on the wall is its numerical range.
• For a perfect, stable top (a normal matrix), the shadow is just a dot (the eigenvalue).
• For a wobbly, chaotic top (a non-normal matrix), the shadow is a big, fuzzy shape. This shape tells you how much the machine might wobble, how sensitive it is to a tiny push, and where it might go if things go wrong.

This paper is about mapping the shapes of these "shadows" for three specific types of chaotic, random machines.

---

The Three Machines Studied

The authors looked at three different ways to build these random machines, each controlled by a "dial" called $\tau$ (tau). This dial turns the machine from "perfectly stable" (Hermitian) to "completely chaotic" (non-Hermitian).

1. The Elliptic Ginibre Ensemble (The Stretchy Balloon)

• The Setup: Imagine taking a perfect circle of numbers and stretching it.
• The Result: As the authors proved, no matter how much you stretch or twist this machine (adjusting the dial $\tau$), the shadow it casts is always a perfect Ellipse (an oval).
• The Analogy: It's like a balloon being squeezed. It changes shape from a circle to a long oval, but it always stays a smooth, simple oval. The authors found the exact formula for how long and wide this oval gets.

2. The Chiral Elliptic Ginibre Ensemble (The Twisted Ribbon)

• The Setup: This is a more complex version, like a ribbon that has been twisted and folded. It's used in physics to describe things like quarks (tiny particles).
• The Result: Even with this extra complexity, the shadow is still an ellipse!
• The Twist: However, if you twist the ribbon too much (change the parameters), the "cloud" of numbers inside can split into two separate islands. But the outer boundary (the numerical range) that contains them all remains a single, smooth oval. It's like a figure-eight cloud where the outer edge is still just one big oval.

3. The Non-Hermitian Wishart Ensemble (The Weird Blob)

• The Setup: This machine is built by multiplying two different random machines together. It's like mixing two different soups.
• The Result: This is the surprise! The shadow is NOT an ellipse.
• The Analogy: Imagine you expect a smooth oval, but instead, you get a shape that looks like a slightly squashed, weirdly curved blob. It has "flat" spots and sharp curves that an ellipse doesn't have. The authors had to invent a new, complicated mathematical recipe (involving quartic polynomials) to describe this weird shape. It's the "odd one out" that refuses to be a simple oval.

---

The "Multiplication" Surprise

The paper also looked at what happens if you multiply these machines together (e.g., Machine A $\times$ Machine B $\times$ Machine C).

• The Intuition: You might think multiplying more chaotic machines would make the shadow grow wildly and unpredictably.
• The Reality: Surprisingly, when you multiply enough of these random machines together (specifically, 2 or more), the shape of the shadow becomes a perfect Circle again!
• The Analogy: It's like spinning a chaotic top. If you spin it once, it wobbles in a weird oval. But if you spin it twice or three times in a row, the chaos averages out, and the wobble settles into a perfect, spinning circle. The authors calculated exactly how big this circle gets as you add more machines to the chain.

Why Does This Matter?

You might ask, "Who cares about the shape of a shadow of random numbers?"

1. Stability: In engineering and physics, these shapes tell us if a system will stay stable or crash. If the shadow is too big or weird, a tiny error could cause a disaster.
2. Speed: When computers solve massive equations (like predicting the weather), they use iterative steps. The shape of this numerical range tells us how fast the computer will finish the job.
3. Universality: The paper shows that even though these machines are random, they follow strict, beautiful geometric rules. Nature seems to prefer ovals and circles, even in chaos.

Summary

• Normal Matrices: Shadow = A single dot (The notes).
• Elliptic/Chiral Matrices: Shadow = A perfect Oval (The stretchy balloon).
• Wishart Matrices: Shadow = A weird, non-oval Blob (The squashed soup).
• Products of Matrices: Shadow = A perfect Circle (The spinning top).

The authors successfully mapped these shadows, proving that even in the messy world of random, non-normal matrices, there is a hidden, elegant geometry waiting to be discovered.

Technical Summary

1. Problem Statement and Motivation

In random matrix theory, the behavior of normal matrices is largely governed by their spectral statistics (eigenvalue distributions). However, non-normal matrices exhibit phenomena that spectral data alone cannot capture, such as eigenvector overlaps, sensitivity to perturbations, and the emergence of pseudospectra.

The numerical range (or field of values), defined for a matrix $A \in \mathbb{C}^{N \times N}$ as $W(A) = \{ \langle Ay, y \rangle : \|y\|_2 = 1 \}$, serves as a robust descriptor of these non-normal effects. While the numerical range coincides with the spectrum for normal matrices, it forms a strictly larger convex set for non-normal matrices.

The Core Problem: While the limiting eigenvalue distributions (spectral droplets) of various non-Hermitian ensembles (like the Elliptic Ginibre and Non-Hermitian Wishart) are well-understood, their corresponding limiting numerical ranges have not been systematically characterized, particularly for ensembles interpolating between Hermitian and non-Hermitian regimes.

2. Methodology

The authors employ a combination of random matrix theory, free probability, and geometric analysis to derive the limiting numerical ranges.

• Supporting Half-Plane Characterization: The central tool is the characterization of the numerical range as the intersection of half-planes determined by the largest eigenvalue of the rotated Hermitian part of the matrix. Specifically:
  $$W(A) = \bigcap_{\theta \in [0, 2\pi]} \{ z \in \mathbb{C} : \text{Re}(e^{i\theta} z) \leq \lambda_{\max}(\text{Re}(e^{i\theta} A)) \}$$
  The authors establish that if the spectral norm $\|\text{Re}(e^{i\theta} X_N)\|$ converges almost surely to a deterministic function $\lambda(\theta)$, then the numerical range converges to the intersection of the corresponding half-planes.

• Free Probability and R-Transforms: For the Non-Hermitian Wishart ensemble, the Hermitian part involves sums of independent Wishart matrices. The authors utilize the R-transform from free probability theory to linearize the free additive convolution of the limiting spectral distributions. This allows them to derive the cubic equation governing the Cauchy transform of the sum, from which the support (and thus the numerical range boundary) is extracted via the discriminant.

• Combinatorics of Non-Crossing Partitions: For the product of elliptic Ginibre matrices, the authors use the moment-cumulant formula of free probability. They analyze the moments of the real part of the product using non-crossing pairings (Catalan-like structures) to show that the limiting norm is independent of the non-Hermiticity parameter $\tau$.

• Algebraic Geometry: To prove that the numerical range of the Wishart ensemble is not an ellipse, the authors perform an algebraic analysis of the defining quartic polynomial, showing that an elliptical ansatz leads to a contradiction for general parameters.

3. Key Contributions and Results

The paper analyzes three specific ensembles, characterized by a non-Hermiticity parameter $\tau \in [0, 1]$ and a chiral parameter $\alpha$ (limit of $\nu/N$).

A. Elliptic Ginibre Ensemble ($X_e$)

• Model: Interpolates between the complex Ginibre matrix ($\tau=0$) and the Gaussian Unitary Ensemble (GUE, $\tau=1$).
• Result: The limiting numerical range is an ellipse $E_{a,b}$ with semi-axes:
  $$a = \sqrt{2(1+\tau)}, \quad b = \sqrt{2(1-\tau)}$$
• Significance: This confirms that the geometry of the numerical range mirrors the elliptical shape of the eigenvalue droplet but with different scaling factors.

B. Chiral Elliptic Ginibre Ensemble ($X_{ce}$)

• Model: A chiral extension involving rectangular matrices $P$ and $Q$, relevant to QCD with chemical potential.
• Result: Despite the eigenvalue droplet undergoing a topological phase transition (from a connected region to two disjoint components) depending on $\tau$ and $\alpha$, the limiting numerical range remains a single ellipse $E_{a,b}$ with axes:
  $$a = \frac{\sqrt{1+\tau}(\sqrt{1+\alpha}+1)}{\sqrt{2}}, \quad b = \frac{\sqrt{1-\tau}(\sqrt{1+\alpha}+1)}{\sqrt{2}}$$
• Significance: The numerical range "smooths out" the topological complexity of the spectrum, remaining a convex ellipse even when the spectrum splits.

C. Non-Hermitian Wishart Ensemble ($X_w$)

• Model: Defined as $X_w = X_1 X_2^*$, related to sample cross-covariance matrices.
• Result: The limiting numerical range is not an ellipse. It is a convex set defined as the intersection of half-planes $H_\theta$, where the boundary is determined by the largest real root $\lambda(\theta)$ of a specific quartic polynomial $D_\theta(x) = 0$.
  • The polynomial coefficients depend on $\tau$, $\alpha$, and the angle $\theta$.
  • In the maximally non-Hermitian case ($\tau=0$), the range becomes a disk with radius $R \approx 1.665$ (for $\alpha=0$).
• Significance: This is a major finding: unlike the spectral limit (which is a shifted ellipse) and the Ginibre cases, the numerical range here possesses a complex, non-elliptic geometry. The authors prove analytically that it cannot be approximated by an ellipse.

D. Products of Elliptic Ginibre Matrices

• Model: Products of $n$ independent elliptic Ginibre matrices ($X_e^n = X_e^1 \cdots X_e^n$).
• Result: For $n \geq 2$, the limiting numerical range is a disk $D(R_n)$ with radius $R_n$ that is independent of $\tau$.
  $$R_n = \frac{\sqrt{n}}{2} \left( 1 + \sqrt{1 + \frac{8}{n}} \right)^{\frac{3}{2}} \left( 3 + \sqrt{1 + \frac{8}{n}} \right)^{\frac{n-1}{2}}$$
• Significance: This extends the "universality" of the product model's spectrum to its numerical range. It also implies that the numerical range of a product of independent Ginibre matrices is asymptotically identical to that of a power of a single Ginibre matrix, provided the total number of factors is the same.

4. Significance and Implications

1. Beyond Spectral Descriptions: The work demonstrates that while eigenvalue distributions provide a "droplet" shape, the numerical range provides a distinct, often larger, geometric descriptor that captures the non-normal stability and convergence properties of these systems.
2. Geometric Diversity: The paper highlights that non-Hermitian random matrices do not universally yield elliptical numerical ranges. The Non-Hermitian Wishart ensemble serves as a counter-example where the geometry is governed by higher-order algebraic curves (quartic envelopes).
3. Universality in Products: The discovery that the numerical range of products of elliptic Ginibre matrices becomes independent of the non-Hermiticity parameter $\tau$ (for $n \geq 2$) suggests a strong universality class for products of random matrices, extending previous spectral results to the numerical range.
4. Methodological Advancement: The rigorous derivation using the intersection of half-planes and free probability techniques provides a template for analyzing the numerical ranges of other complex non-Hermitian ensembles.

In summary, Byun and Park provide the first explicit characterization of the limiting numerical ranges for fundamental non-Hermitian ensembles, revealing a rich geometric landscape that ranges from simple ellipses to complex non-elliptic envelopes, thereby deepening the understanding of non-normal effects in random matrix theory.