PaperlayerView exact revision on arXiv

Source-linked AI summary

Anchored Stein Variational Gradient Descent Algorithm for Non-smooth Sampling

Thanh Dang, Matthew X. Burns, Jiaming Liang

arXiv:2610.36018v1stat.MLcs.LG

TL;DR

For sampling when the target potential may be nonsmooth, the paper introduces anchored SVGD using a smoother reference potential and anchor weight. It establishes the KL descent characterization under low regularity and proves a nonasymptotic convergence bound in anchored KSD.

  • Problem

    Standard SVGD updates depend on the target score, while this paper addresses sampling targets whose potential U may be nonsmooth.

    Source: §1 · p. 2

  • Method

    Anchored SVGD uses a continuously differentiable reference potential U0 and weight a = e^(U-U0) to define deterministic particle evolution through anchored KL steepest descent.

    Source: §1 · pp. 2–3

  • Results

    The paper establishes the KL first variation and steepest descent direction under locally Lipschitz U, and proves a nonasymptotic bound on time-averaged anchored KSD.

    Source: §1 · pp. 3–5

  • Takeaways & Limitations

    The anchored KSD includes standard KSD when U = U0, and the framework permits analysis with a locally Lipschitz target potential and continuously differentiable reference potential.

    Source: §1 · p. 3

    The convergence theorem requires stated assumptions including a bound on |∆U0(x)| by C∆d and conditions on the initial particle configuration and density.

    Source: §1 · p. 4

Abstract

from arXiv · show

We introduce a Stein variational framework for sampling from a target density $p\propto e^{-U}$ when the target potential $U$ is possibly non-smooth. The key idea is to introduce a smoother reference potential $U_0$ and the anchor weight $a=e^{U-U_0}$. The resulting anchored Stein variational gradient descent (SVGD) algorithm is able to target $p\propto e^{-U}$ without computing the gradient of $U$ and include standard SVGD as a special case. In particular, we allow the target potential $U$ to be locally Lipschitz while the reference potential $U_0$ is continuously differentiable, along with other suitable assumptions. Under this low regularity regime, we establish a Kullback-Leibler first variation formula for anchored perturbations, identify the corresponding steepest descent direction in the reproducing kernel Hilbert space and prove a nonasymptotic convergence bound for the anchored SVGD algorithm in terms of an anchored kernelized Stein discrepancy. Finally, we also provide guidance on constructing a suitable reference potential $U_0$ based on $U$.

1 Introduction and main result

Anchored SVGD uses a smoother reference potential to sample targets whose potentials may be nonsmooth, and the paper establishes its KL descent characterization and a nonasymptotic convergence bound. The analysis applies when U is locally Lipschitz and U0 is continuously differentiable.

1 Introduction and main result

  • Anchored SVGD targets p ∝ e^-U using reference potential U0 and weight a = e^(U-U0), replacing stochastic diffusion with deterministic particle evolution.The algorithm is motivated by anchored Langevin dynamics and selects particle evolution through the anchored KL steepest-descent principle.Source: §1 · p. 2
  • The target potential U may be locally Lipschitz while the reference potential U0 is continuously differentiable, and the paper establishes a KL first variation formula under these assumptions.This formula identifies the steepest descent direction in the anchored RKHS class.Source: §1 · p. 3
  • Theorem 1.1 establishes that anchored perturbations yield a well-defined flow and a KL functional that is finite and absolutely continuous over every bounded time interval.The flow’s pushforward density satisfies a continuity equation.Source: §1 · p. 3
  • Theorem 1.2 identifies the steepest descent direction for KL divergence among anchored velocity fields ψ = aϕ with ϕ in the product RKHS.The paper also states a De Bruijn-type identity for this setting.Source: §1 · p. 3
  • Theorem 1.4 gives a nonasymptotic bound on the time-averaged anchored KSD of empirical measures along the discrete anchored SVGD trajectory.The anchored KSD contains standard KSD as the special case U = U0.Source: §1 · pp. 4–5
  • The paper describes convolution smoothing of U as a way to construct reference potential U0 and develops flow and particle-update bounds to support its convergence analysis.The convergence proof uses a discrete descent estimate for KL divergence relative to p0 ∝ e^-U0.Source: §1 · pp. 4–5

2 Preliminaries

The preliminaries define the function spaces and flows used in the analysis, then introduce an anchored kernelized Stein discrepancy and describe how to construct a smoother reference potential. They also state conditions under which the discrepancy detects the target and the smoothing construction meets required assumptions.

2.1 Notations and function spaces

  • The notation covers Lipschitz and Sobolev spaces, RKHSs, and Lipschitz flows whose pushforwards define transported densities.Source: §2.1 · pp. 6–7

2.2 Anchored kernelized Stein discrepancy

  • The anchored KSD is zero only when q equals p almost everywhere, and it reduces to the standard KSD when U equals U0.Source: §2.2 · pp. 7–8

2.3 Constructing reference potential U0

  • A locally Lipschitz target potential can be smoothed using general densities, including Gaussian smoothing, under conditions that relax those imposed in earlier Gaussian-convolution guidance.Source: §2.3 · p. 9

Define the smoothing

  • A locally Lipschitz U with the stated gradient-difference and growth conditions yields a smoothed U0 satisfying the cited assumptions.The growth condition requires inf_x U(x) > 0 and |∇U(x)| ≤ A0 U(x)^α0, with α0 in [0, 1/2).Source: Define the smoothing · p. 9, §2.3 · p. 9

3 KL first variation formula and optimal direction field

This section establishes the KL first variation for anchored perturbations under locally Lipschitz target potentials, with regularity and integrability conditions justifying the analysis. It develops supporting transport and RKHS results, identifies the steepest KL-decrease direction in the RKHS, and establishes a De Bruijn-type identity.

3.1 Proof overview for the KL first variation formula

  • The paper rigorously derives the KL first variation under locally Lipschitz U, where the continuity equation is only distributionally defined and the formal substitution is unjustified.The smoothness-based derivation is explicitly described as formal; the rigorous argument addresses this low-regularity gap.Source: §3.1 · p. 10

3.2 General transport results

  • For bounded uniformly Lipschitz transport fields, a unique flow exists with controlled displacement and bi-Lipschitz bounds, and its Jacobian is positive with an almost-everywhere time evolution.The flow also supports change-of-variables and pushforward density formulas.Source: §3.2 · pp. 10–11
  • The transport change-of-variables formula defines pushforward densities that preserve probability mass.The lemma gives an almost-everywhere density formula and states that a probability density remains a probability density under the flow.Source: §3.2 · p. 11
  • For q ∈ L1(Rd), the passage states that t ↦ q_t is continuous from I, but the cited text ends before specifying the codomain or the topology of continuity.Source: §3.2 · p. 11

3.3.1 Proof of Theorem 1.1

  • The anchored field aϕ is bounded and globally Lipschitz, yielding a well-posed ODE and an associated global flow.This follows from the stated assumptions on the test field and anchor.Source: §3.3.1 · p. 14

3.3 KL first variation formula

  • Under the stated assumptions, the anchored perturbation’s KL functional is C1 and its derivative at zero is the first variation in Theorem 1.1.The assumptions include locally Lipschitz U, continuously differentiable U0, a bounded Lipschitz anchor a, and moment and integrability conditions.Source: §3.1 · p. 10, §3.3 · p. 14, §3.3.1 · p. 16

3.4 Optimal direction in the RKHS

  • Under the kernel assumptions, RKHS functions have coordinate derivatives, and the vector-valued RKHS embeds continuously into bounded Lipschitz fields.Source: §3.4 · p. 17
  • The RKHS vector fields embed continuously into the bounded-Lipschitz space, so the perturbation flow is well defined under the cited existence-and-uniqueness result.Source: §3.4 · p. 17, §3.4.1 · p. 18

3.4.1 Proof of Theorem 1.2

  • The KL map along the perturbation is differentiable at zero, as follows from the bounded-Lipschitz embedding and the first variation theorem.Source: §3.4.1 · p. 18
  • The proof represents field values and coordinate derivatives through RKHS inner products with the kernel and its derivative.Source: §3.4.1 · p. 18
  • The H-valued maps used in the variation are Bochner integrable because their expected norms are finite under the kernel and moment assumptions.Source: §3.4.1 · p. 18
  • Bounded linear functionals commute with the Bochner integrals in the proof, allowing the variation to be expressed for each fixed RKHS coordinate.Source: §3.4.1 · p. 18
  • The supremum of the KL variation over the RKHS unit ball equals the norm of h, attained by the stated optimal direction, yielding steepest KL decrease and a De Bruijn-type identity.Source: §3.4.1 · p. 18

4 Anchored Stein variational gradient descent algorithm

Section 4 defines anchored SVGD updates and develops a descent estimate using the reference density as a Lyapunov choice. Its proof of Theorem 1.4 applies that estimate under the stated assumptions to obtain the theorem’s bound.

4 Anchored Stein variational gradient descent algorithm

  • Algorithm 1 updates particle i by adding its step size εℓ times F_i(x^ℓ), evaluated at the full configuration, for i = 1, . . . , N and ℓ = 0, 1, . . . , L − 1.Source: §4 · p. 19
  • The descent estimate is stated for a general positive reference density, then specialized to p0 ∝ e^−U0 for anchored SVGD.The section identifies p0 as the natural Lyapunov reference in the anchored setting.Source: §4 · p. 19
  • The descent estimate requires the listed regularity assumptions and a positive reference density, with the proof specializing to p0.Proposition 4.3 assumes Assumptions 3.3(A2), 3.4(K2), and 4.1(S1), along with a constant step size η.Source: §4 · p. 19, §4.1 · p. 25

4.1 Proof of Theorem 1.4

  • The convergence proof bounds a KL term by (2γd + 2/e + log 2)/(Lη) before deriving the bound in Theorem 1.4.The proof uses the choices of p0 and η from the theorem’s setup.Source: §4.1 · pp. 25, 27

A Sampling targets with non-smooth potentials

The section surveys several approaches to sampling from distributions with non-smooth potentials. These include derivative-free random walks, modified Langevin methods, auxiliary-variable samplers, and interacting-particle methods.

A Sampling targets with non-smooth potentials

  • Geometric random walks such as the ball walk and hit-and-run require no derivatives of the potential.The passage presents them as approaches to non-smooth sampling.Source: A Sampling targets with non-smooth… · p. 33
  • Non-smooth Langevin methods use proximal updates, potential splitting, smoothing, subgradient updates, or anchoring.The passage also lists auxiliary-variable approaches that alternate conditional sampling steps.Source: A Sampling targets with non-smooth… · p. 33
  • Other approaches connect sampling to stochastic localization and denoising diffusion or use Wasserstein proximal splitting and proximal particle Langevin algorithms.The passage describes the former as using increasingly informative observations and the latter as interacting-particle methods.Source: A Sampling targets with non-smooth… · p. 33

B Proof and supporting lemmas for Section 2.2

The appendix supports Section 2.2 with identities and lemmas for anchored KSD. It relates the discrepancy to an RKHS norm and gives a condition under which a function orthogonal to all RKHS test functions must vanish.

B Proof and supporting lemmas for Section 2.2

  • Anchored KSD is expressed as the RKHS norm of h, with h specified in equation (4).The passage uses this identity in the proof of Theorem 2.2, Part i.Source: B Proof and supporting lemmas for… · p. 33
  • The proof of Theorem 2.2 uses the zero-discrepancy condition together with these lemmas to conclude p = q almost everywhere.The supplied proof passage states this conclusion after identifying the relevant function as zero.Source: B Proof and supporting lemmas for… · pp. 33–34
  • Under the stated universality assumption, orthogonality of h to every RKHS test function implies that h vanishes almost everywhere with respect to ν.The proof uses density of H in C0(Rd) and applies the conclusion componentwise.Source: B Proof and supporting lemmas for… · p. 36

C Proof for Section 2.3

The proof verifies the required regularity assumptions for the smoothed reference potential and anchor weight when the target potential is locally Lipschitz. It establishes bounded derivatives for the reference potential and bounded weak derivatives for the anchor.

Verifying Assumption 3.3(A1)

  • The locally Lipschitz target potential has an almost-everywhere gradient bound, and its convolution with the smoothing kernel yields a continuously differentiable reference potential satisfying the stated gradient-growth condition.The proof uses kernel integrability, quadratic growth of U, and integration by parts to establish the reference-potential bound.Source: C Proof for Section 2.3 · pp. 36–37

Verifying Assumption 3.3(A2)

  • The anchor weight a belongs to W^1,∞(R^d), with its weak gradient bounded using the target-gradient and reference-gradient bounds.The proof combines the bound on the difference between target gradients at nearby points with ∇a = a∇(U − U0) almost everywhere.Source: C Proof for Section 2.3 · p. 37

Verifying Assumption 4.1(S1)

  • The reference potential has a uniformly bounded Hessian, so the stated properties of Assumption 4.1(S1) hold.The proof first establishes the bound almost everywhere and extends it to every x by continuity.Source: C Proof for Section 2.3 · pp. 36, 38

D Supporting lemmas for Section 4

Lemma D.1 develops estimates for the particle drift and an auxiliary function under assumptions on the target, kernel, and reference potential. Combining these estimates supports finite-iteration bounds, including a bound on the quantity E′′ over the specified step intervals.

Part i

  • Under Assumptions 3.3(A2), 3.4(K2), and 4.1(S1), the lemma bounds the particle drift term for every particle configuration.The proof uses kernel-derivative bounds and the reference-potential growth condition.Source: D Supporting lemmas for Section 4 · pp. 38–39

Part ii

  • The auxiliary function h(ℓ) admits a bound with positive constants M and D depending only on the listed model and kernel parameters.The proof combines bounds on the anchor, kernel, and reference potential.Source: D Supporting lemmas for Section 4 · pp. 38, 42

Parts iii–iv

  • Under Part ii’s hypotheses, Part iii combines the estimates to bound the relevant quantities, while Part iv bounds E′′ for every iteration through L and t in [0, η].Part iv assumes x0 ∈ GK and specifies a step-size choice; the displayed bound’s expression is not included in these passages.Source: D Supporting lemmas for Section 4 · pp. 38–39
Loading 2610.36018v1…