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

1.1 Background: Deterministic and Langevin-regularized SVGD

• Deterministic SVGD can suffer mode collapse, and its general finite-particle guarantees concern time averages rather than last-iterate convergence in physical time.Kernel repulsion may be too weak in high dimensions or for well-separated multimodal targets, while existing general finite-particle bounds decay algebraically with the averaging horizon.Source: §1.1 · pp. 2–3
• Langevin-regularized SVGD dissipates relative entropy through both kernel Stein discrepancy and relative Fisher information; a target log-Sobolev inequality gives exponential convergence.The Stein term acts in the kernel-induced geometry, while the Fisher-information term retains the Wasserstein-2 gradient-flow structure.Source: §1.1 · p. 3

1.3 Main contributions

• Combining finite-time entropy and synchronous-coupling estimates with a cutoff argument yields polynomial-in-N uniform-in-time propagation-of-chaos rates in expectation.The bounds concern the last iterate at every physical time, rather than time-averaged empirical measures.Source: §1.1 · p. 3

1.2 Propagation of chaos and the long-time problem

• The Stein drift need not be small relative to the confining Langevin drift, so the dynamics generally do not contract distances between solutions.The paper nevertheless obtains polynomial uniform-in-time propagation-of-chaos rates in strong metrics through a different form of stability.Source: §1.2 · p. 5

1.3 Main contributions

• The mean-field and particle systems admit exponential-in-time, polynomial-in-N target-convergence bounds under a target log-Sobolev inequality, without requiring a convex Stein interaction.The finite-particle relative-entropy identity is with respect to the product target; a kernel-regime condition bounds its diagonal correction uniformly in N, and the bounds include uniform second-moment control.Source: §1.3 · p. 5

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

• Synchronous coupling pairs particles with independent nonlinear McKean–Vlasov copies driven by the same Brownian motions, reducing averaged squared distance to a Grönwall estimate.Its random coefficient and forcing depend on the independent copies; exponential moment bounds close the estimate in expectation.Source: §2 · pp. 5–6
• Moving-product relative entropy yields a differential inequality in which diffusion-generated relative Fisher information absorbs the interaction error between particle and mean-field drifts.An exponential law-of-large-numbers estimate closes the inequality; entropy superadditivity and Pinsker’s inequality then give fixed-marginal KL and total-variation estimates for 1 ≤ k ≤ N/2.Source: §2 · p. 6
• A logarithmic cutoff combines finite-time PoC and target-convergence estimates to give polynomial uniform-in-time rates in KSD, W2, and total variation for the last iterate at every physical time.The bounds are not time-averaged; the section contrasts them with the time-averaged measure treated in [BBK26].Source: §2 · p. 6
• Under an additional T2 inequality for the initial law, moving-product entropy also gives finite-time and uniform-in-time W2 PoC bounds for empirical distributions and particle marginals.The two methods have qualitatively similar rates, but moving-entropy bounds can be sharper for kernels with large spatial derivatives or strong oscillations, while synchronous coupling can improve rates with favorable pathwise contraction.Source: §2 · pp. 6–7

• Following [BBK26], the setup uses a symmetric positive-definite scalar kernel and its associated vector-valued RKHS to define the Stein velocity.Source: §2 · p. 8
• The Stein witness is defined by a Bochner integral, and under either kernel assumption it is defined for every ρ ∈ P1(Rd), with Sπ(π) = 0.Source: §2 · pp. 8–9
• KSDπ is symmetric, obeys the triangle inequality, and satisfies KSDπ(ρ, π) = KSD(ρ∥π).Source: §2 · p. 9
• The continuous-time interacting diffusion corresponds to the discrete-time noisy-SVGD recursion [PBS24], and its associated flow is a Langevin-regularized RKHS projected gradient flow of relative entropy.Source: §2 · p. 9
• The section defines Kullback–Leibler divergence and relative Fisher information for probability laws, with Fisher information set to +∞ when the required Sobolev condition fails.Source: §2 · pp. 9–10

3 Assumptions and well-posedness

• The target LSI implies π ∈ T2(1/α), whereas the additional initial-law condition µ0 ∈ T2(C0) is Assumption 7.8 and is used only in Section 7.3.The initial law is also required to have a smooth positive density under the core assumptions.Source: §3 · p. 10
• Assumptions 3.2 and 3.3 provide alternative kernel regularity conditions for the results, while Assumption 3.4 defines a strong-concentration regime.The concentration assumption is not needed for target convergence, the full-law target entropy estimate, or the uniform second-moment bound; it supports concentration and exponential-moment estimates in the finite-time arguments.Source: §3 · p. 11
• The assumptions are not inclusive of those used for deterministic W1 and W2 results; the paper uses a target LSI and sub-Gaussian initial-law information for stronger polynomial uniform-in-time estimates in expectation.The cited comparison notes that deterministic results instead use specialized Langevin-semigroup stability assumptions and normalized or unnormalized bilinear–Matérn kernels [BBK26].Source: §3 · p. 11
• Under Assumption 3.1 and either Assumption 3.2 or 3.3, the particle system and nonlinear SDE have unique global strong solutions with finite second moments on every finite time interval.The proof uses local Lipschitz and growth bounds, moment estimates to rule out explosion, and a synchronous-coupling argument for pathwise uniqueness of the nonlinear SDE.Source: §3 · pp. 11–13

Mean-field target entropy

• Under the target log-Sobolev inequality, mean-field relative entropy decays exponentially, yielding pointwise convergence to the target in W2 and KSD.The mean-field entropy dissipation combines the Stein and Langevin components; the Wasserstein and KSD convergence results are pointwise in time.Source: Mean-field target entropy · pp. 13–14

Finite-particle target entropy

• The finite-particle law has a target-entropy dissipation identity that, with target log-Sobolev control, gives exponential entropy decay from product initialization.The theorem assumes Assumption 3.1 and either Assumption 3.2 or 3.3; the diagonal-correction condition in Assumption 3.2 is one route to the result.Source: Finite-particle target entropy · pp. 15–16

4.2 Results in Langevin KSD

• Finite-particle empirical KSD converges pointwise to the target up to an N^-1/2 sampling floor.The pointwise estimate follows from full-law target-entropy control and tensorized T2; unlike deterministic SVGD, no time averaging is needed.Source: §4.2 · pp. 16–17

4.3 Results in W2 distance

• The empirical measure converges pointwise to the target in W2 under Assumption 3.1 and either kernel assumption in Assumption 3.2 or 3.3.The argument uses full-law target entropy and the target T2 inequality implied by the log-Sobolev inequality, without a KSD-to-W2 comparison.Source: §4.3 · pp. 17, 19
• Under Proposition 4.10’s assumptions, the mean-field and particle second moments satisfy a uniform-in-time bound.Proposition 4.11 provides pointwise second-moment control, alongside an integrated bound over each finite time horizon.Source: §4.3 · pp. 18–19

5.1 Exponential moment estimates under target and mean-field laws

• The target has moments of every polynomial order, while the nonlinear law has uniform sub-Gaussian moments and yields uniform-in-time empirical W2 floors.The nonlinear-law moment bound assumes 3.1 and 3.4 together with either kernel assumption; the empirical W2 result applies to the target and uniformly over the nonlinear laws.Source: §5.1 · pp. 20–21

5.2 Stein-velocity fluctuations

• The Stein-velocity defect satisfies an exponential law of large numbers with constants independent of particle number and time.The result holds under Assumptions 3.1 and 3.4, together with either kernel assumption.Source: §5.2 · p. 22

5.3 Synchronous coupling and metric estimates

• Synchronous coupling, using the same Brownian motions, yields a finite-time particle-to-mean-field discrepancy bound with constants independent of particle number and time.The argument uses an exponential moment bound under the independent nonlinear product path law.Source: §5.3 · pp. 23–24

5.4 Finite-time propagation of chaos in W2

• Finite-time empirical and fixed-marginal W2 propagation-of-chaos estimates follow under Assumptions 3.1, 3.3, and 3.4.The empirical estimate compares the particle empirical measure with the nonlinear-law empirical measure; the fixed-marginal estimate applies to every number of coordinates.Source: §5.4 · p. 26

5.5 Finite-time propagation of chaos in KSD

• Under Assumptions 3.1, 3.2, and 3.4, the section establishes finite-time KSD propagation-of-chaos estimates.The estimate is obtained by combining the synchronous-coupling discrepancy with a centered Hilbert-space empirical-average term.Source: §5.5 · p. 26

6 Uniform-in-Time Propagation of Chaos using Cutoff

• The cutoff requires only a comparison through the fixed target, and the resulting uniform-in-time PoC rates decay polynomially in N.The argument adapts the cutoff technique of [BBK26] and uses single-exponential finite-time growth; double-exponential growth would produce only a negative power of log N [BBK26].Source: §6 · p. 27

6.1 Uniform-in-time PoC in KSD

• Theorem 6.3 establishes uniform-in-time propagation of chaos in KSD under Assumptions 3.1, 3.2, and 3.4.The cutoff framework gives polynomial decay in N when the finite-time estimate grows at most single-exponentially; a double-exponential estimate would yield only a negative power of log N [BBK26].Source: §6.1 · p. 27, §6 · p. 27

6.2 Uniform-in-time PoC in W2

• Theorem 6.4 establishes uniform-in-time propagation of chaos for the empirical measure in W2 under Assumptions 3.1, 3.3, and 3.4.When both kernel assumptions hold, the induced KSD estimate loses a square root and has worse rates than Theorem 6.3 because of the empirical Wasserstein rate’s dimensional dependence.Source: §6.2 · p. 28

6.3 Uniform-in-time PoC for fixed-marginal laws

• Theorem 6.9 establishes uniform-in-time W2 propagation of chaos for fixed-k marginals when N ≥ 2k under Assumptions 3.1, 3.3, and 3.4.The section obtains fixed-marginal bounds from full-law entropy control using entropy superadditivity for exchangeable laws, and compares short-time and long-time estimates.Source: §6.3 · pp. 28, 30

7 PoC using the moving-product entropy method

• Following, the method controls joint-law entropy against an evolving product law using an exponential law of large numbers, with potential advantages over synchronous coupling for less regular kernels.The moving-product entropy bound is the basis for the estimates developed in the subsections that follow.Source: §7 · p. 30

7.1 Fixed-particle marginal estimates

• Entropy superadditivity and Pinsker yield finite-time fixed-marginal KL and total variation control, plus uniform-in-time total variation for fixed k, but not uniform-in-time KL.The cutoff argument cannot transfer the finite-time KL estimate to uniform time because relative entropy lacks a triangle inequality.Source: §7 · p. 30, §7.1 · pp. 32–33

7.2 Moving-product entropy-based PoC in Langevin KSD

• Uniform-in-N,t exponential-moment concentration of the Stein witness, combined with the moving-entropy bound, yields quantitative finite-time KSD propagation-of-chaos estimates.The argument combines the concentration result with the entropy variational inequality.Source: §7.2 · pp. 33, 35

7.3 Alternative entropy-to-Wasserstein estimates

• Under an initial T2 condition, the entropy route gives finite-time empirical and fixed-marginal W2 estimates and alternative uniform-in-time W2 bounds.It first propagates a time-dependent T2 inequality to the nonlinear law.Source: §7.3 · pp. 36–38

7.4 Comparison of the coupling and entropy routes

• Entropy gives the better cutoff rate when its finite-time growth exponent is smaller; coupling is better when L_cpl < L_ent, though neither estimate need be sharp.Coupling does not require a transportation inequality for the initial law; entropy controls joint-law relative entropy and yields marginal relative entropy and total variation, but its conversion to Wasserstein requires µ0 ∈ T2(C0). The empirical W2^2 rate may still be determined by the sampling term rN,d.Source: §7.4 · p. 39

7.4.1 A narrow-kernel regime favoring entropy

• For narrow kernels, the entropy cutoff exponent remains h-independent, whereas coupling yields an N^-cε,α,a,dh^2 uniform-in-time rate for fixed ε and h.This compares powers of N for each fixed h, not an h-uniform multiplicative constant; finite-particle target-estimate constants may diverge as h ↓ 0.Source: §7.4.1 · p. 41

7.4.2 A contractive regime favoring coupling

• For a constant kernel, direct synchronous coupling gives the optimal N^-1 error, while the entropy cutoff yields a fixed-marginal rate N^-ε^2/(β^2+ε^2).For any fixed admissible λ, the entropy cutoff exponent is strictly below 1, so the direct coupling bound is sharper in this example.Source: §7.4.2 · p. 42
• For kernel families with controlled interaction amplitude but Klip(h) ≍ h^-p, entropy has uniform-in-h cutoff growth while the coupling bound contains ε^-1h^-2p growth.A model-specific coupling estimate can also give a uniform fixed-marginal N^-1 bound, while the general entropy cutoff exponent is below 1 when L_ent > 0.Source: §7.4.2 · p. 42

Corollary 8.1

• The covered class of standard stationary kernels includes the Gaussian RBF kernel.For these kernels, the corollary verifies the kernel regularity and diagonal-correction conditions used by the theory.Source: §8 · pp. 42–43
• With the stated sub-Gaussian initial-law condition, synchronous coupling and moving-product entropy yield finite-time and uniform-in-time empirical-measure PoC in KSD and W2, as well as finite-marginal PoC.If the initial law also satisfies T2(C0), the alternative entropy-derived Wasserstein bounds apply.Source: §8 · p. 43
• For convex potentials with Hessian bounded below by m and above by L, the target has log-Sobolev constant α = m and the target-convergence and uniform-second-moment results apply.The lower and upper Hessian bounds also ensure a finite-second-moment target and the required regularity conditions.Source: §8 · p. 43

Corollary 8.2 (A nonconvex target covered by the theory)

• The nonconvex example satisfies the bounded-Hessian and kernel assumptions, so the main propagation-of-chaos and uniform-in-time conclusions apply.It is nonconvex when Aω^2 > m; its target has log-Sobolev constant α = me^-2A, and the dissipativity constant is m/2.Source: §8 · pp. 43–44

9.1 Possible finite-moment localization

• Localization may yield uniform-in-time propagation of chaos in probability under finite-moment assumptions, but requires a maximal or stopped fluctuation estimate and does not recover moving-product entropy’s fixed-marginal entropy and total-variation bounds.The proposed approach is in the spirit of [BBK26]; the needed fluctuation estimate is not pursued here.Source: §9.1 · pp. 44–45

9.2 Dependence on the regularization strength

• The estimates are not uniform as ε decreases: entropy decays at rate 2εα, W2 and KSD at rate εα, and finite-time coupling and entropy exponents are of order ε^-1.Concentration constants may also depend on ε, so the cutoff argument’s polynomial uniform-in-time propagation-of-chaos exponents degenerate in the small-noise limit.Source: §9.2 · p. 45
• A Stein log-Sobolev inequality could keep the target-convergence rate nondegenerate as ε decreases, if its constant αS > 0 is independent of ε.The discussion notes recent coercivity results for kernels with sufficiently strong high-frequency behavior [CSW24] [Chi+26].Source: §9.2 · p. 45
• An ε-uniform cutoff theorem would still require finite-particle long-time target control and ε-uniform finite-time propagation-of-chaos estimates; mean-field Stein log-Sobolev alone supplies neither.High-frequency-coercive Riesz, Bessel, or Coulomb kernels may be singular on the diagonal, requiring new self-interaction and fluctuation control, potentially by removing or renormalizing diagonal terms [TH26].Source: §9.2 · p. 45

A Entropy evolution equation

• The appendix establishes a relative-entropy chain rule for two finite-dimensional diffusion laws with a common diffusion coefficient.The result is obtained first for positive times and then extended to time zero.Source: A Entropy evolution equation · pp. 46, 49
• A smooth spatial cutoff localizes the entropy calculation, separating drift, diffusion, and remaining contributions before the cutoff is removed.Linear-growth and finite-time moment bounds, together with integrability and convergence arguments, justify removal of the cutoff.Source: A Entropy evolution equation · pp. 47–49

B Regularity results

• The mean-field drift and its spatial gradient are 1/2-Hölder continuous in time on every finite interval.Their uniform-in-space differences between times t and s are bounded by C_T |t-s|^1/2.Source: B Regularity results · p. 49
• The particle and mean-field configuration-space drifts have local spatial regularity and at most linear growth on finite time intervals.These properties hold under Assumption 3.1 and either kernel assumption.Source: B Regularity results · pp. 49–50
• The RKHS H^d is separable, the Stein witness is strongly measurable, and S_π(ρ) is well defined for every ρ in P1(R^d).The lemma also establishes S_π(π) = 0 and, under Assumption 3.2, additional RKHS regularity for the witness.Source: B Regularity results · pp. 50–51