Abstract
This research introduces a novel framework for describing the evolution of the spectrum in massive random matrices. By shifting the focus from spectral measures to their spatial primitives—cumulative distribution functions—the authors utilize viscosity solution theory to provide elegant proofs for existence, uniqueness, and convergence. The central innovation is the concept of "spectral dominance," which yields a powerful comparison principle for non-local transport equations.
1. Introduction: The Challenge of Spectral Evolution
In the study of large random matrices, a central objective is to understand how the spectrum—the set of eigenvalues—evolves as matrix coefficients undergo stochastic processes, such as Brownian motion. As the matrix size $N$ tends toward infinity, the individual, fluctuating eigenvalues smooth out into a continuous, deterministic density known as the limit spectral measure.
Mathematically, this density evolution is governed by non-local, singular partial differential equations (PDEs). Historically, proving the uniqueness of solutions to these PDEs has been a significant hurdle, often requiring arduous, case-by-case analysis tailored to specific models like Dyson Brownian motion or Wishart processes. This paper addresses this gap by providing a "master key": a generalized framework that ensures uniqueness across a broad class of models.
2. A Shift in Perspective: From Measures to Primitives
The core innovation of this work lies in a clever change of variables. Instead of analyzing the spectral measure $\mu$ directly, the authors focus on its spatial primitive, the cumulative distribution function (CDF) $F(t, x)$:
$$F(t, x) = \int_{-\infty}^x d\mu_t(y)$$
This function counts the proportion of eigenvalues below a threshold $x$ at time $t$. While the original equation for the measure $\mu$ is highly singular and mathematically recalcitrant, the derived transport equation for $F$ is significantly better behaved. Crucially, because $F$ is inherently non-decreasing, the non-local operator in the resulting PDE satisfies a "maximum principle"—a fundamental requirement for the application of viscosity solution theory.
3. The Comparison Principle and Viscosity Solutions
To handle the potential lack of smoothness in solutions, the authors employ the theory of viscosity solutions. This powerful framework allows for the definition of "correct" solutions even when traditional derivatives do not exist. The cornerstone of this approach is the "comparison principle," derived from the concept of spectral dominance.
In the discrete setting, a matrix $A$ is spectrally dominated by matrix $B$ if every ordered eigenvalue of $A$ is less than or equal to the corresponding eigenvalue of $B$. The paper demonstrates that this property translates into a comparison principle for the limit PDE: if one initial distribution is dominated by another, their respective solutions maintain this ordering over time. This principle is the "secret sauce" that guarantees the uniqueness of the viscosity solution.
4. Key Results and Unification
The spectral dominance framework yields several significant results:
- Existence and Uniqueness: The paper proves that for a general class of random matrix models, there exists a unique viscosity solution to the limit equation.
- Rigorous Convergence: The authors provide a streamlined proof showing that as $N \to \infty$, the discrete $N$-particle system rigorously converges to the unique continuous viscosity solution.
- Geometric Interpretation: The "Dyson flow" is shown to be a contraction in Wasserstein metric spaces, offering a geometric perspective on how these spectral distributions converge.
- Handling Singularities: The framework naturally identifies "physical" solutions, such as the Wigner semi-circle law, even when starting from singular initial data like a Dirac mass.
5. Conclusion and Future Directions
By unifying the treatment of various models under the banner of spectral dominance, this work simplifies the landscape of random matrix theory. While the current focus is on one-dimensional real spectra, the authors suggest that future research will aim to extend these results to complex spectra in two dimensions and explore systems involving time-dependent boundary conditions or external control mechanisms.
Summary updated