pith. sign in

arxiv: 2505.02621 · v2 · pith:WTBEHZA4new · submitted 2025-05-05 · 💻 cs.LG · math.OC· stat.ML

Mirror Mean-Field Langevin Dynamics

Anming Gu , Juno Kim This is my paper

Pith reviewed 2026-05-22 16:09 UTC · model grok-4.3

classification 💻 cs.LG math.OCstat.ML
keywords mirror mean-field Langevin dynamicsconstrained optimizationlog-Sobolev inequalitypropagation of chaosmean-field limitsWasserstein spacemirror maps
0
0 comments X

The pith

Mirror mean-field Langevin dynamics optimizes probability measures on constrained convex domains with linear convergence.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces mirror mean-field Langevin dynamics to optimize entropy-regularized functionals over probability measures restricted to a convex subset of Euclidean space. Standard mean-field Langevin dynamics cannot enforce such constraints because its diffusion term acts globally. The new method incorporates a mirror map to adapt the dynamics while keeping the mean-field interaction structure intact. Linear convergence of the continuous process follows from a uniform log-Sobolev inequality that yields a domain-independent contraction rate. The time- and particle-discretized versions satisfy uniform-in-time propagation of chaos, so finite-particle approximations track the mean-field limit over long horizons.

Core claim

The authors propose the mirror mean-field Langevin dynamics (MMFLD) as an extension of mean-field Langevin dynamics to the mirror Langevin framework. This allows optimization of probability measures constrained to a convex subset of R^d. They obtain linear convergence guarantees for the continuous MMFLD via a uniform log-Sobolev inequality, and uniform-in-time propagation of chaos results for its time- and particle-discretized counterpart.

What carries the argument

The mirror mean-field Langevin dynamics, which uses a mirror map to transform the constrained problem into an equivalent unconstrained dynamics in a different geometry while preserving mean-field interactions.

If this is right

  • The continuous MMFLD converges linearly to the minimizer of the entropy-regularized functional under the uniform log-Sobolev condition.
  • Finite-particle discretizations remain close to the mean-field limit uniformly in time.
  • The method applies to constrained mean-field models such as infinite-width neural networks with domain restrictions.
  • Both time discretization and particle discretization preserve the convergence and approximation properties without extra assumptions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the uniform log-Sobolev inequality can be checked for common constraints like the probability simplex, the method becomes immediately usable for many practical problems.
  • The mirror construction may combine with other acceleration techniques already known for mirror descent in finite dimensions.
  • Similar extensions could apply to non-entropy regularizers or to dynamics with additional interaction terms.
  • Numerical tests on low-dimensional constrained measures would clarify whether the theoretical rates appear in practice.

Load-bearing premise

A uniform log-Sobolev inequality holds for the mirror mean-field dynamics on the constrained convex domain.

What would settle it

A specific convex constraint set and functional for which the continuous dynamics exhibits only sublinear convergence or the log-Sobolev constant diverges with the constraint.

read the original abstract

The mean-field Langevin dynamics (MFLD) minimizes an entropy-regularized nonlinear convex functional on the Wasserstein space over $\mathbb{R}^d$, and has gained attention recently as a model for the gradient descent dynamics of interacting particle systems such as infinite-width two-layer neural networks. However, many problems of interest have constrained domains, which are not solved by existing mean-field algorithms due to the global diffusion term. We study the optimization of probability measures constrained to a convex subset of $\mathbb{R}^d$ by proposing the \emph{mirror mean-field Langevin dynamics} (MMFLD), an extension of MFLD to the mirror Langevin framework. We obtain linear convergence guarantees for the continuous MMFLD via a uniform log-Sobolev inequality, and uniform-in-time propagation of chaos results for its time- and particle-discretized counterpart.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes mirror mean-field Langevin dynamics (MMFLD) to optimize entropy-regularized functionals over probability measures supported on a convex constrained domain in R^d. It claims linear convergence of the continuous-time process by invoking a uniform log-Sobolev inequality, together with uniform-in-time propagation of chaos for its time-discretized and particle-discretized versions.

Significance. If the uniformity of the LSI constant (independent of the evolving mean-field measure) can be rigorously established under the stated assumptions on the mirror map and interaction kernel, the work would supply useful theoretical guarantees for mean-field optimization on constrained domains, extending existing MFLD results to settings relevant for constrained neural network training and related particle systems.

major comments (2)
  1. [Abstract and convergence analysis] Abstract and convergence analysis: the linear convergence claim for continuous MMFLD is obtained via a uniform log-Sobolev inequality whose constant must remain independent of the mean-field measure μ. The effective potential is the original objective plus the interaction ∫W(x,y)dμ(y); the Bakry–Émery or Holley–Stroock curvature condition then depends on the Hessian of this term, which varies with μ. The manuscript must supply an explicit uniform lower bound on the curvature (or an explicit LSI constant that does not deteriorate with μ) to convert the invocation into a proof of a μ-independent linear rate.
  2. [Section on mirror map and domain constraints] Section on mirror map and domain constraints: it is unclear whether the mirror map properties alone guarantee that the LSI constant remains uniform when the support is restricted to the convex subset; additional regularity assumptions on the interaction W (e.g., uniform strong convexity or bounded Hessian norms) appear necessary but are not stated explicitly as sufficient conditions.
minor comments (1)
  1. [Notation and assumptions] Clarify the precise statement of the uniform LSI (including the dependence on the mirror map) and state all assumptions on W in a single theorem or proposition.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive feedback on the uniformity of the log-Sobolev inequality and the role of assumptions on the mirror map and interaction kernel. We address the major comments point by point below.

read point-by-point responses

  1. Referee: [Abstract and convergence analysis] Abstract and convergence analysis: the linear convergence claim for continuous MMFLD is obtained via a uniform log-Sobolev inequality whose constant must remain independent of the mean-field measure μ. The effective potential is the original objective plus the interaction ∫W(x,y)dμ(y); the Bakry–Émery or Holley–Stroock curvature condition then depends on the Hessian of this term, which varies with μ. The manuscript must supply an explicit uniform lower bound on the curvature (or an explicit LSI constant that does not deteriorate with μ) to convert the invocation into a proof of a μ-independent linear rate.

    Authors: We agree that an explicit uniform lower bound on the curvature is required to obtain a μ-independent linear rate. Under the paper's standing assumptions that the mirror map is α-strongly convex and β-smooth and that the interaction kernel W has Hessian norm bounded by L (independent of μ), the Bakry–Émery curvature of the effective potential is bounded below by α − L. We will add a short lemma in the revised manuscript that states this bound explicitly and derives the corresponding uniform LSI constant, thereby completing the linear-convergence argument. revision: yes

  2. Referee: [Section on mirror map and domain constraints] Section on mirror map and domain constraints: it is unclear whether the mirror map properties alone guarantee that the LSI constant remains uniform when the support is restricted to the convex subset; additional regularity assumptions on the interaction W (e.g., uniform strong convexity or bounded Hessian norms) appear necessary but are not stated explicitly as sufficient conditions.

    Authors: The referee correctly notes that mirror-map properties alone are insufficient. The manuscript implicitly relies on a uniform bound on the Hessian of W to control the perturbation of the curvature on the constrained domain. We will revise the relevant section to list the explicit sufficient conditions on W (bounded Hessian norm and, optionally, uniform strong convexity) and state that these conditions, together with the mirror-map assumptions, guarantee uniformity of the LSI constant on the convex subset. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on standard functional inequalities

full rationale

The paper derives linear convergence for continuous MMFLD from a uniform log-Sobolev inequality and propagation of chaos for discretizations from standard mirror map and mean-field analysis. No step reduces a claimed prediction or result to a fitted parameter, self-definition, or self-citation chain by construction. The uniform LSI is invoked as an assumption on the constrained domain rather than derived from the target convergence rate itself, and the abstract and described results remain self-contained against external benchmarks such as Bakry-Émery criteria and existing MFLD theory. No load-bearing self-citation or ansatz smuggling is exhibited in the provided derivation outline.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claims rest on the convexity of the domain and the existence of a uniform log-Sobolev inequality for the mirror dynamics; these are domain assumptions rather than derived quantities.

axioms (1)
  • domain assumption Uniform log-Sobolev inequality holds for the MMFLD on the constrained convex set
    Directly invoked in the abstract to obtain linear convergence for the continuous dynamics.

pith-pipeline@v0.9.0 · 5666 in / 1045 out tokens · 55315 ms · 2026-05-22T16:09:56.391304+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.