PaperlayerView exact revision on arXiv

Source-linked AI summary

Quantitative Target Convergence and Uniform-in-Time Propagation of Chaos for Langevin-Regularized SVGD

Sayan Banerjee, Dohyeon Kim

arXiv:2608.28827v1stat.MLcs.LGmath.PR

TL;DR

The paper addresses convergence and uniform-in-time propagation of chaos for Langevin-regularized SVGD despite non-small interactions and generally noncontractive couplings. It combines entropy identities, synchronous coupling, and moving-product entropy to obtain last-iterate target convergence and polynomial uniform-in-time bounds.

  • Problem

    Langevin-regularized SVGD lacks the small-interaction contractivity used in standard models, while deterministic SVGD has limited last-iterate and uniform-in-time guarantees.

  • Method

    The paper combines mean-field and finite-particle entropy identities with synchronous coupling and moving-product entropy to control convergence and propagation of chaos.

  • Results

    The analysis yields exponential last-iterate convergence to the target and polynomial uniform-in-time propagation-of-chaos rates for empirical KSD, W2^2, fixed-marginal total variation, and W2^2.

  • Takeaways & Limitations

    The bounds apply to last iterates in physical time without requiring the Stein interaction to be small or the particle dynamics to be contractive.

  • Takeaways & Limitations

    The estimates are proved for fixed ε > 0, and their uniform-in-time polynomial exponents degenerate as ε approaches zero.

Abstract

from arXiv · show

We establish quantitative convergence to the target and uniform-in-time propagation of chaos for Langevin-regularized Stein variational gradient descent. The Stein interaction need not be small relative to the confining Langevin drift and does not generally yield a contractive particle coupling. At the mean-field level, the Stein and Langevin components dissipate the same relative entropy in the kernel-induced Stein and $2$-Wasserstein geometries, producing the squared kernel Stein discrepancy and relative Fisher information. Under a log-Sobolev inequality for the target, this yields exponential last-iterate convergence. We also derive a finite-particle entropy identity relative to the product target, giving exponential-in-time convergence of the empirical measure up to polynomial sampling errors. For propagation of chaos, we develop two complementary finite-time approaches. A synchronous coupling, combined with exponential moment estimates for the nonlinear mean-field diffusion, yields explicit single-exponential bounds in Wasserstein distance and kernel Stein discrepancy (KSD). Moving-product entropy gives joint-law relative entropy control relative to the evolving mean-field product law and, through entropy superadditivity and concentration, fixed-marginal relative entropy and total variation bounds and empirical KSD estimates. Under an additional $T_2$ inequality for the initial law, it also yields Wasserstein bounds. Combining these finite-time estimates with target convergence at a logarithmic cutoff time gives polynomial uniform-in-time propagation of chaos rates in expectation for empirical KSD and $W_2^2$, and for fixed-marginal total variation and $W_2^2$. All bounds control the last iterate in physical time. We also compare the two finite-time mechanisms and identify regimes in which each gives the sharper polynomial exponent.

1 Introduction

Langevin-regularized SVGD augments deterministic Stein transport with noise, addressing mode collapse and the lack of general last-iterate, uniform-in-time guarantees. The paper develops entropy-based target convergence and polynomial uniform-in-time propagation-of-chaos results for physical-time last iterates.

  • Background: SVGD transports particles using a kernelized velocity combining attraction toward high target-probability regions with repulsive interaction.Its mean-field evolution is an entropy gradient flow in a kernel-dependent Stein geometry.
  • Background: Deterministic SVGD can suffer mode collapse because kernel repulsion may weaken for high-dimensional or well-separated multimodal targets.Particles may underestimate target variance or concentrate on only some modes.
  • Langevin regularization: Langevin regularization adds diffusion to the interacting dynamics, yielding the continuous-time counterpart of noisy-SVGD recursion.The resulting system is studied alongside its nonlinear McKean–Vlasov equation.
  • Long-time convergence: The Stein and Langevin components dissipate relative entropy through kernel-induced Stein and Wasserstein geometries, with the latter producing Fisher-information decay under an LSI.This supports exponential convergence of the last iterate rather than only time-averaged convergence.
  • Propagation of chaos: Uniform-in-time propagation of chaos is obtained by combining synchronous coupling and moving-product entropy with a logarithmic cutoff argument.The resulting polynomial-in-N rates concern last iterates at every physical time, unlike time-averaged guarantees.
  • Main contributions: The paper derives finite-particle entropy stabilization relative to the product target without requiring the Stein interaction to be small or convex.Tensorized target inequalities yield pointwise target convergence with polynomial particle-number errors.

2. Moving-product entropy, synchronous coupling, and polynomial uniform-in-time

The paper develops synchronous-coupling and moving-product-entropy approaches to finite-time propagation of chaos, then combines them with target convergence at a logarithmic cutoff to obtain polynomial uniform-in-time last-iterate bounds.

  • Finite-time mechanisms: Two complementary approaches quantify finite-time propagation of chaos: synchronous coupling and moving-product entropy.The coupling uses shared Brownian motions and exponential moment estimates, while the entropy method controls the evolving joint law relative to the mean-field product law.
  • Finite-time mechanisms: Entropy superadditivity and Pinsker’s inequality convert moving-product entropy control into fixed-marginal relative entropy and total variation estimates.The argument uses an exponential law-of-large-numbers estimate to close the entropy differential inequality.
  • Uniform-in-time results: A logarithmic cutoff combines finite-time propagation-of-chaos estimates with target convergence to yield polynomial uniform-in-time rates in KSD, W2, and total variation.These estimates concern the last iterate at every physical time rather than a time-averaged empirical measure.
  • Uniform-in-time results: Under an additional initial-law T2 inequality, moving-product entropy also yields empirical and fixed-marginal W2 bounds.The two methods have qualitatively similar rates, but their polynomial exponents depend on kernel regularity and dynamics.
  • Method comparison: Moving-entropy bounds can be sharper for kernels with large spatial derivatives or strong oscillations, whereas synchronous coupling can be better under favorable pathwise contractivity.The paper does not assume a corresponding uniform-in-particle log-Sobolev inequality for the noisy-SVGD invariant law.
  • Assumptions and scope: The results require core target and initial-law conditions together with dissipativity and sub-Gaussian initial-tail assumptions, with additional kernel conditions varying by bound.The stronger tail information is exchanged for polynomial uniform-in-time estimates in expectation.

2 Setup and notation

The paper formulates Langevin-regularized SVGD through Stein operators, RKHS witnesses, entropy quantities, and interacting diffusions, linking the particle system to a nonlinear mean-field process.

  • Stein geometry: The Stein velocity is defined through a Bochner-integrated Stein witness in a vector-valued RKHS generated by the kernel.The associated KSD compares probability laws through the Stein witness norm.
  • Stein geometry: The Langevin–Stein operator and KSD connect the particle approximation to the mean-field projected gradient flow of relative entropy.The KSD satisfies symmetry and a triangle inequality in the stated construction.
  • Particle and mean-field dynamics: The noisy-SVGD particle diffusion is the continuous-time interacting dynamics associated with a discrete-time noisy-SVGD recursion.The corresponding nonlinear process supplies the mean-field comparison system.
  • Information quantities: Relative entropy and relative Fisher information are defined for probability laws, with Pinsker’s inequality providing a conversion from entropy to total variation.The paper also fixes conventions for constants, empirical measures, and marginal particle laws.

3 Assumptions and well-posedness

Under target, kernel, regularity, and concentration assumptions, the particle and nonlinear systems are globally well posed with finite finite-time second moments, while stronger tails support the paper’s uniform-in-time expectation bounds.

  • Core assumptions: A target log-Sobolev inequality implies π ∈ T2(1/α), while the additional condition µ0 ∈ T2(C0) is used only for the moving-entropy W2 results.These are distinct target and initial-law transportation conditions.
  • Core assumptions: The particle kernel assumptions require symmetric positive definiteness with specified C4 or C3 regularity, bounded derivatives, and a diagonal correction.The two assumption families support KSD- and Wasserstein-oriented estimates, respectively.
  • Concentration regime: Strong concentration assumes dissipativity together with sub-Gaussian initial tails and supplies the concentration estimates used in both propagation-of-chaos approaches.It is not needed for target convergence, full-law target entropy estimates, or uniform second-moment bounds.
  • Scope and limitations: The results trade stronger tail information for stronger polynomial uniform-in-time estimates in expectation, with localization under weaker moment assumptions identified as a possible extension.The assumptions are comparable to, but not inclusive of, those used for deterministic SVGD results.
  • Well-posedness: Under the core assumptions and either kernel condition, the particle system and nonlinear SDE admit unique global strong solutions with finite second moments on every finite time interval.The proof uses local Lipschitz and growth bounds, truncation, moment estimates, and Grönwall arguments.

4 Quantitative convergence to the target

The mean-field and finite-particle dynamics converge exponentially toward the target under a target log-Sobolev inequality, with entropy dissipation controlling KSD, Fisher information, and Wasserstein error. The finite-particle results include diagonal-correction conditions and sampling-error floors, while all stated bounds are pointwise in physical time.

  • Mean-field convergence: Mean-field relative entropy dissipates through the Stein and Langevin components, yielding exponential convergence under the target log-Sobolev inequality.The resulting controls involve squared KSD and relative Fisher information.
  • Mean-field convergence: Pointwise mean-field convergence holds in both W2 and KSD under the stated regularity assumptions.The W2 result follows from the target LSI and does not require a KSD-to-W2 comparison.
  • Finite-particle convergence: The full particle-law entropy relative to π⊗N dissipates exponentially, providing the basis for finite-particle target convergence.Product initialization gives HN(0) = N KL(µ0∥π), while tensorization transfers the target LSI to the product law.
  • Assumptions and scope: The particle entropy argument requires a diagonal-correction condition, although a broader entropy-variational replacement enlarges only the target-entropy part.The replacement does not remove the uniform growth condition needed for finite-time propagation of chaos.
  • Finite-particle convergence: Finite-particle convergence to the target in empirical KSD and W2 is pointwise in physical time and includes polynomial sampling errors.The empirical KSD estimate uses tensorized T2, while the W2 estimate combines full-law entropy with empirical sampling error.
  • Assumptions and scope: Unlike deterministic SVGD’s time-averaged control, Langevin regularization yields exponential last-iterate decay and transfers it to empirical KSD up to an N^-1/2 sampling floor.The same pointwise physical-time perspective applies to the finite-particle target estimates.

5 Finite-time propagation of chaos

Finite-time propagation of chaos is obtained through exponential moment and concentration estimates combined with two complementary mechanisms: synchronous coupling and metric-specific empirical comparisons. These yield quantitative W2 and KSD bounds for empirical measures and fixed marginals under the stated assumptions.

  • Moment and concentration estimates: Uniform sub-Gaussian moments of the mean-field law support time-uniform concentration estimates for empirical sampling and Stein-velocity fluctuations.The target has moments of every polynomial order under the strong-concentration regime.
  • Moment and concentration estimates: An exponential law of large numbers controls the Stein-velocity defect uniformly in particle number and physical time.The estimate relies on centered independent fluctuations, bounded-kernel assumptions, and uniform exponential-square moments.
  • Synchronous coupling: Synchronous coupling uses identical Brownian motions for interacting particles and independent mean-field copies, so the Brownian terms cancel in the discrepancy dynamics.Exponential moment estimates for the mean-field paths close the resulting finite-time bound.
  • Synchronous coupling: The synchronous-coupling estimates produce finite-time empirical W2 propagation-of-chaos bounds and fixed-marginal W2 bounds.The empirical estimate combines coupling discrepancy with the i.i.d. empirical W2 floor, while fixed-marginal bounds use labelled coordinate couplings.
  • Metric estimates: The same coupling framework yields finite-time KSD propagation-of-chaos estimates under the bounded-kernel assumptions.The proof combines the RKHS triangle inequality, coupling control, and concentration of independent centered Hilbert-valued fluctuations.

6 Uniform-in-Time Propagation of Chaos using Cutoff

The cutoff method combines finite-time propagation-of-chaos estimates with long-time target convergence to obtain polynomial uniform-in-time bounds. Single-exponential finite-time growth is essential for logarithmic cutoffs, and the resulting rates cover KSD, empirical W2, and fixed-marginal W2.

  • Cutoff framework: The cutoff argument combines finite-time propagation-of-chaos estimates with long-time target estimates to obtain polynomial uniform-in-time rates in N.The comparison is performed through the fixed target π using an abstract discrepancy framework.
  • Cutoff framework: Single-exponential finite-time growth yields polynomial rates after the logarithmic cutoff, whereas double-exponential growth would yield only a negative power of log N.This distinguishes the present coupling estimate from the deterministic-SVGD regime discussed in the comparison remark.
  • KSD and Wasserstein results: Theorem 6.3 establishes uniform-in-time propagation of chaos in Langevin KSD under the stated assumptions.
  • KSD and Wasserstein results: Theorem 6.4 establishes uniform-in-time empirical W2 propagation of chaos under the corresponding assumptions.
  • Comparison: The empirical-W2 route can lose a square root and inherit the curse of dimensionality through the empirical Wasserstein rate rN,d.
  • Fixed-marginal results: Theorem 6.9 establishes uniform-in-time fixed-marginal W2 chaos for fixed k and N ≥ 2k.

7 PoC using the moving-product entropy method

The moving-product entropy method controls the joint particle law relative to the evolving mean-field product law and converts this control into marginal KL, total variation, KSD, and conditional W2 estimates. Its finite-time growth can improve uniform-in-time exponents when kernel regularity makes synchronous coupling less favorable.

  • Core method: Moving-product relative entropy compares the joint particle law with the evolving mean-field product law and closes through an exponential law of large numbers.
  • Marginal consequences: Entropy superadditivity and Pinsker’s inequality yield fixed-marginal KL and total variation propagation-of-chaos estimates.
  • KSD estimates: The method also provides an alternate route to empirical KSD propagation of chaos through concentration of the Stein witness map.
  • Wasserstein estimates: Under an additional T2 inequality for the initial law, moving entropy yields finite-time and uniform-in-time W2 bounds for empirical distributions and fixed particle marginals.
  • Limitation: Relative entropy lacks a triangle inequality, so the cutoff argument alone cannot establish uniform-in-time KL propagation of chaos.
  • Method comparison: The entropy route can outperform synchronous coupling when interaction amplitudes remain controlled while higher spatial derivatives become large.

8 Concrete applications: smooth stationary kernels with convex and nonconvex targets

The paper verifies its assumptions for broad smooth stationary-kernel classes, including Gaussian RBF kernels, convex targets, and a nonconvex bounded perturbation of a Gaussian target. These examples inherit the paper’s finite-time and uniform-in-time propagation-of-chaos conclusions.

  • Kernel classes: The covered stationary-kernel class includes the Gaussian RBF kernel.
  • Convex targets: For potentials with suitable lower and upper Hessian bounds, the target satisfies a log-Sobolev inequality and the kernel conditions hold.
  • Consequences: The resulting assumptions support synchronous coupling, moving-product entropy, and propagation-of-chaos results for empirical KSD, W2, and finite marginals.
  • Nonconvex target: A nonconvex target with Aω^2 > m is covered, and all main propagation-of-chaos and uniform-in-time conclusions apply.
  • Nonconvex target: For the nonconvex example, bounded perturbation arguments give a target log-Sobolev constant α = me^-2A.

9 Conclusion and discussion

The discussion identifies extensions and limitations of the theory, especially the need for localization without exponential initial moments and the degeneration of cutoff exponents as ε decreases. A Stein log-Sobolev inequality may remove the mean-field rate degeneration, but additional finite-particle and finite-time control remains necessary.

  • Extensions: Without finite initial exponential moments, localization may yield uniform-in-time propagation of chaos in probability, but this direction is not pursued.
  • Small-noise limitation: The proved estimates are not uniform as ε ↓ 0, and the polynomial uniform-in-time exponents degenerate in the small-noise limit.
  • Possible remedy: A Stein log-Sobolev inequality could keep mean-field target-convergence rates nondegenerate as ε ↓ 0 when its constant is ε-independent.
  • Open requirements: An ε-uniform cutoff theorem additionally requires long-time finite-particle target control and ε-uniform finite-time propagation-of-chaos estimates.
  • Kernel scope: High-frequency coercive kernels may be singular on the diagonal and require new control of self-interactions and empirical fluctuations.

A Entropy evolution equation

The appendix establishes the finite-dimensional relative-entropy evolution identity by localizing the calculation, controlling cutoff remainders with moment and Girsanov estimates, and then removing the cutoff and taking the initial-time limit.

  • Localization: The entropy calculation is localized with a smooth cutoff before separately evaluating drift, diffusion, and remaining contributions.The cutoff equals one on a radius-R ball, vanishes outside radius 2R, and has gradient bounded by C/R.
  • Localization: Linear-growth assumptions and finite-time second-moment estimates control the cutoff remainder, while Girsanov’s theorem supplies the needed integrability.Dominated convergence is then applied as the cutoff radius tends to infinity.
  • Removing the cutoff: Completing the square and applying dominated or monotone convergence passes the localized identity to the limit as R →∞.The nonnegative square term is handled by monotone convergence.
  • Initial-time limit: The resulting entropy identity holds for 0 < s ≤ t and is extended to s = 0 by weak convergence and lower semicontinuity of relative entropy.The proof concludes after letting s decrease to zero.

B Regularity results

The appendix verifies regularity and growth properties for the particle and mean-field drifts and establishes well-definedness of the Stein witness, including the identity Sπ(π) = 0.

  • Drift regularity: The particle and mean-field drifts have the required local spatial regularity, Lipschitz bounds, and at most linear growth.These properties follow from regularity of V, bounded kernel derivatives, and finite-time moment estimates.
  • Drift regularity: The nonlinear process provides a coupling of (µt, µs), yielding 1/2-Hölder continuity in time for the mean-field drift and its spatial derivative.The bound is sup_x {∥vµt(x) − vµs(x)∥ + ∥∇xvµt(x) − ∇xvµs(x)∥} ≤ CT|t − s|1/2.
  • Stein witness: Sπ(π) = 0 under the stated assumptions, established using the reproducing property, cutoff approximation, and dominated convergence.The argument tests the Stein witness against every ϕ ∈ Hd.
  • Stein witness: The Stein witness is strongly measurable and Bochner integrable in the separable RKHS, so Sπ(ρ) is well defined for every ρ ∈ P1(Rd).Continuity of the RKHS-valued kernel maps and boundedness of the relevant derivatives provide the required measurability and integrability.
Loading 2608.28827v1…