Back to search
2208.04506

Second Order Ensemble Langevin Method for Sampling and Inverse Problems

Ziming Liu, Andrew Stuart, Yixuan Wang

correctmedium confidence
Category
math.DS
Journal tier
Strong Field
Processed
Sep 28, 2025, 12:56 AM

Audit review

The paper’s Proposition 5.1 states and proves the closed ODEs for the mean and block covariances, classifies all steady states (unique positive-definite equilibrium and a family of degenerate equilibria), and establishes local exponential stability in B-whitened variables with a rate depending only on γ, independent of B and c. The candidate solution reproduces the same moment equations, obtains the same steady-state structure via the projection X characterization, performs the same whitening, and derives the same cubic characteristic polynomial λ^3 + 3γ λ^2 + (2γ^2 + 6) λ + 4γ for the 3×3 covariance block. It adds a Routh–Hurwitz check and a standard small-data bootstrap to articulate the open basin and uniform-in-(B,c) rate. All core steps align with the paper’s derivations and claims, including the block structure of the Jacobian and the independence of B and c after whitening. Hence both are correct and essentially the same proof, with the model supplying slightly more explicit stability folklore details. See the statement of (5.2) and Proposition 5.1 and its proof in Section 8.4, including the whitened system (8.1) and the cubic eigenvalue relation, as well as the independence from B and c and hyperbolicity claims in the paper.

Referee report (LaTeX)

\textbf{Recommendation:} minor revisions

\textbf{Journal Tier:} strong field

\textbf{Justification:}

The paper gives a correct and clear analysis of the linear case of covariance-preconditioned underdamped mean-field Langevin dynamics. It derives closed moment equations, classifies equilibria, and establishes local exponential stability with rates independent of the problem after whitening. Minor additions could improve readability and rigor in the nonlinear step, but the core results are sound and valuable for sampling in inverse problems.