Matrix Ornstein–Uhlenbeck Diffusion

• Matrix Ornstein–Uhlenbeck diffusion is a class of matrix-valued stochastic processes defined by linear drift and additive noise, extending the classic scalar OU process.
• It underpins studies in random matrix theory, eigenvalue dynamics, and convergence to equilibrium through functional inequalities and spectral gap analysis.
• Applications include Bayesian parameter inference, convergence rate analysis, and modeling of spectral evolution in high-dimensional statistics.

A matrix Ornstein–Uhlenbeck diffusion refers to a class of stochastic processes where the state variable is matrix-valued and evolves under both linear drift and noise, generalizing the scalar Ornstein–Uhlenbeck process to multivariate, operator-valued, or random matrix contexts. These processes arise in random matrix theory, multivariate probability, high-dimensional statistics, and the study of spectrum evolution in dynamical ensembles.

1. Hermitian Matrix Ornstein–Uhlenbeck Dynamics

Let $M_t$ denote an $n\times n$ Hermitian matrix-valued process governed by the stochastic differential equation (SDE)

$$\mathrm{d}M_t = -{\tfrac12} M_t\,\mathrm{d}t + \mathrm{d}B_t, \qquad M_0 = m_0,$$

where $B_t$ is Hermitian matrix Brownian motion (entries have independent real and imaginary Brownian motions). In "mean-field" random matrix scaling,

$$\mathrm dM_t = \sqrt{\tfrac{2}{n}}\,\mathrm dB_t - M_t\,\mathrm dt, \quad M_0 = m_0.$$

The process is reversible with respect to the Gaussian Unitary Ensemble (GUE)

$${\rm GUE}_n(\tfrac1n) = \frac1{Z_n} e^{-\tfrac n2\,\mathrm{Tr}(M^2)}\,\mathrm dM,$$

with a spectral gap of $1$ under normalization (Boursier et al., 2021).

2. Induced Eigenvalue (Dyson) Diffusion and the β-Hermite Ensemble

Diagonalizing $M_t = U_t\Lambda_t U_t^*$, the eigenvalues $$\Lambda_t = \mathrm{diag}(\lambda_1(t),\ldots,\lambda_n(t))$$ evolve according to the Dyson–Ornstein–Uhlenbeck process (DOU), a system of interacting diffusions: $$\mathrm d\lambda_i = \mathrm dW_i + \sum_{j\ne i} \frac{\beta/2}{\lambda_i-\lambda_j}\,\mathrm dt - \tfrac12 \lambda_i\,\mathrm dt,$$ where $W_i$ are independent scalar Brownian motions. This describes Brownian motion on a symmetric chamber with global repulsion (β-Coulomb interaction) and linear drift. The equilibrium law is the $\beta$-Hermite (Coulomb-gas) ensemble

$$p(\lambda) \propto \exp\left(-\tfrac{\beta}{4}\sum_i \lambda_i^2 + \beta \!\!\sum_{i<j} \ln|\lambda_i - \lambda_j|\right).$$

For $\beta=2$ this is the GUE eigenvalue law (Boursier et al., 2021). The process generalizes to non-Hermitian and other symmetry classes (Blaizot et al., 2015).

3. Functional Inequalities and Quantifying Convergence to Equilibrium

Convergence to equilibrium is analyzed via multiple distances:

• Total-variation (TV) distance $\|\mu-\nu\|_{\mathrm{TV}}$
• Relative entropy (Kullback–Leibler divergence) $D_{\mathrm{KL}}(\nu\mid\mu)$
• $\chi^2$-divergence
• Fisher information $I(\nu\mid\mu)$
• Wasserstein-2 distance $W_2(\mu,\nu)$

Convexity of the energy functional ensures log-concavity of the invariant law, yielding:

• Poincaré inequality (spectral gap $1$)
• Log-Sobolev inequality: $$D_{\mathrm{KL}}(\nu\mid P) \le \frac{1}{2n} I(\nu\mid P)$$
• Talagrand's $T_2$-inequality: $W_2^2(\nu, P)\le\frac1{n} D_{\mathrm{KL}}(\nu|P)$

These inequalities provide exponential decay in all the considered distances, which is crucial for establishing rapid and robust mixing to equilibrium in high dimension (Boursier et al., 2021).

4. Universal Cutoff Phenomenon in High Dimensions

A sharp cutoff phenomenon occurs as the system size $n\to\infty$: the approach to equilibrium (in TV, KL, Hellinger, Wasserstein metrics) transitions abruptly at a critical time,

$$T_n^{(\mathrm{TV})} = \log(n a_n), \qquad T_n^{(W_2)} = \log(\sqrt{n}a_n),$$

for initial conditions $x_0^n\in[-a_n,a_n]^n$ with $a_n\to0$ slowly. The cutoff profile is explicit in the noninteracting (β=0) case via Mehler formulas. Strikingly, the cutoff time and mixing rates do not depend on β (interaction strength): the DOU process and independent OUs exhibit identical mixing-time scales in high dimension. This β-independence is established through matching upper and lower bounds derived from contraction, coupling, and projection techniques (Boursier et al., 2021).

5. Semigroup Spectrum and Multivariate Ornstein–Uhlenbeck Processes

For a matrix-valued (multivariate) OU process defined by

$$\mathrm{d}X_t = B X_t\,\mathrm{d}t + \Sigma\,\mathrm{d}W_t,\quad X_t\in\mathbb{R}^d,$$

with $B$ diagonalizable (real spectrum $-\beta_i < 0$) and $\Sigma$ of full rank, the generator is

$$L f(x) = \langle B x, \nabla f(x) \rangle + \frac12 \mathrm{Tr}(Q \nabla^2 f(x)), \quad Q = \Sigma\Sigma^T,$$

with unique stationary Gaussian measure $\mu$ and covariance solving the Lyapunov equation. The spectrum is

$$\sigma_p(L) = \left\{ -\langle n, \beta \rangle = -\sum_{i=1}^d n_i\beta_i: n\in\mathbb{N}^d\right\},$$

with corresponding multivariate Hermite eigenfunctions and explicit co-eigenfunctions. The spectrum and eigenfunction multiplicities are independent of the noise covariance $\Sigma$ ("isospectrality") (Sarkar, 21 Feb 2025). In the presence of non-normal drift, the full Jordan decomposition and algebraic/geometric multiplicities can be derived explicitly (Chen et al., 2012).

6. Riemannian and Covariance-Matrix-Valued Generalizations

For the cone of $n\times n$ positive-definite matrices $\mathcal{S}_+(n)$, Riemannian matrix OU diffusions are constructed on the manifold with either Log-Euclidean or Affine-Invariant metric:

• Under the Log-Euclidean metric, a diffeomorphic reduction to linear multivariate OU yields explicit solutions and invariant "log-Gaussian" laws.
• Under the Affine-Invariant metric, the invariant law is a Riemannian Gaussian and solutions require nontrivial geometric constructions (Bui et al., 2021).

The infinitesimal generator is

$$\mathcal{A}f = \left\langle -\frac\theta2 \nabla d^2(\cdot, M), \nabla f\right\rangle + \frac{\sigma^2}{2}\Delta f,$$

and for the Log-Euclidean case, the spectrum matches that of the classic OU: Hermite-polynomial eigenfunctions and explicit spectral gap (Bui et al., 2021).

7. Statistical Inference and Bayesian Estimation

For time series data from multivariate or matrix OU processes, efficient Bayesian inference of drift and diffusion parameters is achieved via sufficient statistics: $$\hat{\Phi} = S_{xy}\,S_{xx}^{-1}, \quad \hat{\Sigma} = \frac{1}{N}[S_{yy} - \hat{\Phi} S_{xy}^T ],$$ enabling closed-form maximum a posteriori (MAP) or maximum likelihood (MLE) estimators, with analytic expressions for uncertainty and model comparison via Laplace approximation (Singh et al., 2017). This framework applies to inertial Brownian oscillators, matrix-valued volatility models, and a wide range of multivariate time series.

---

References:

• Universal cutoff for Dyson Ornstein Uhlenbeck process (Boursier et al., 2021)
• Spectrum of Lévy-Ornstein-Uhlenbeck semigroups on $\mathbb{R}^d$ (Sarkar, 21 Feb 2025)
• Ornstein-Uhlenbeck diffusion of hermitian and non-hermitian matrices—unexpected links (Blaizot et al., 2015)
• Fast Bayesian inference of the multivariate Ornstein-Uhlenbeck process (Singh et al., 2017)
• Inference for partially observed Riemannian Ornstein-Uhlenbeck diffusions of covariance matrices (Bui et al., 2021)
• On the Jordan decomposition for a class of non-symmetric Ornstein-Uhlenbeck operators (Chen et al., 2012)