Source-linked AI summary
Measuring Sample Quality with Diffusions
Jackson Gorham, Andrew B. Duncan, Sebastian J. Vollmer, Lester Mackey
TL;DR
Multivariate Stein factors and characterizing operators have been analyzed for relatively few continuous targets, limiting the available scope of Stein-based sample-quality measures. This paper develops diffusion-based operators and explicit bounds, then establishes convergence-determining discrepancies with a near-linear Wasserstein relationship that improves on prior analyses even for strongly log-concave targets.
Problem
Few multivariate continuous targets have established characterizing operators and Stein factor bounds, although these are central to Stein-based convergence measurement.
Source: §1 · p. 3Method
The paper builds characterizing Stein operators from Itô diffusions and derives explicit multivariate Stein factor bounds for fast-coupling diffusions.
Source: §1 · p. 3Results
The Langevin Stein discrepancy determines convergence for smooth, distantly dissipative targets via Wasserstein bounds, and the paper reports improvement over prior analyses even for strongly log-concave targets.
Source: §1 · pp. 2–3Takeaways & Limitations
For biased MCMC hyperparameter selection, the Stein discrepancy selects an intermediate step size that matches the surrogate Wasserstein minimum and yields samples closest to the reference sample.
Source: §5.1 · p. 18, §5.2 · p. 20The paper restricts its attention to Stein operators arising from diffusion generators, although it notes that more general operators need not correspond to diffusion processes.
Source: §6 · p. 23
Abstract
from arXiv · showhide
Stein's method for measuring convergence to a continuous target distribution relies on an operator characterizing the target and Stein factor bounds on the solutions of an associated differential equation. While such operators and bounds are readily available for a diversity of univariate targets, few multivariate targets have been analyzed. We introduce a new class of characterizing operators based on Ito diffusions and develop explicit multivariate Stein factor bounds for any target with a fast-coupling Ito diffusion. As example applications, we develop computable and convergence-determining diffusion Stein discrepancies for log-concave, heavy-tailed, and multimodal targets and use these quality measures to select the hyperparameters of biased Markov chain Monte Carlo (MCMC) samplers, compare random and deterministic quadrature rules, and quantify bias-variance tradeoffs in approximate MCMC. Our results establish a near-linear relationship between diffusion Stein discrepancies and Wasserstein distances, improving upon past work even for strongly log-concave targets. The exposed relationship between Stein factors and Markov process coupling may be of independent interest.
1. Introduction.
The paper seeks a tractable measure of how weighted samples approximate expectations under a continuous target when exact expectations are unavailable. It extends multivariate Stein analysis with diffusion-based operators and Stein factor bounds, then develops convergence-determining discrepancies for broader target classes.
1. Introduction.
- The goal is a computationally tractable quality measure that can detect convergence and non-convergence of weighted samples approximating an intractable target.The sample points may come from any random or deterministic mechanism.Source: §1 · pp. 1–2
- The Langevin Stein discrepancy determines convergence for smooth, distantly dissipative targets through explicit upper and lower bounds by Wasserstein distances.Distant dissipativity covers Gaussian mixtures and robust Student’s t regression posteriors.Source: §1 · p. 2
- For heavy-tailed targets, the paper introduces multivariate Stein operators based on general Itô diffusions as an alternative to the Langevin operator.The authors note that heavy-tailed distributions are never distantly dissipative.Source: §1 · p. 2
- The paper establishes a near-linear relationship between diffusion Stein discrepancies and standard Wasserstein distances, improving on prior analyses even for strongly log-concave targets.The introduction states that constants in the bounds are determined by Theorem 7 and Proposition 8.Source: §1 · pp. 2–3
- Explicit Stein factor bounds from fast-coupling Itô diffusions extend multivariate Stein analysis beyond its concentration on Gaussian, Dirichlet, and strongly log-concave targets.The paper connects Stein factors to Markov process coupling times.Source: §1 · p. 3
- Applications use the discrepancies to select biased-sampler hyperparameters, compare quadrature rules, and quantify bias-variance tradeoffs in approximate MCMC.The paper also verifies sufficient conditions for fast coupling for log-concave, heavy-tailed, and multimodal targets.Source: §1 · p. 3
2. Stein’s method.
Section 2 constructs diffusion-based Stein discrepancies and connects their convergence to Wasserstein convergence. It derives multivariate Stein factor bounds from diffusion coupling and gives conditions for bounding discrepancies in both directions.
2. Stein’s method.
- Stein’s method builds a computable discrepancy by choosing a target-mean-zero operator, solving its Stein equation, then bounding that discrepancy.The operator and its domain define the discrepancy, avoiding explicit integration under the target.Source: §2 · p. 4
2.1. Identifying a Stein operator.
- An invariant-measure Itô diffusion yields a mean-zero Stein operator that depends on the target density only through its log-gradient.The classical Stein set supplies a domain of functions with bounded, Lipschitz derivatives.Source: §2.1 · pp. 5–7
2.2. Lower bounding the diffusion Stein discrepancy.
- Fast-coupling Itô diffusions yield explicit uniform multivariate Stein factor bounds, while Wasserstein decay gives lower bounds relating discrepancy to Wasserstein distance.The explicit Wasserstein bound from discrepancy is established for constant covariance and stream matrices; additional smoothness assumptions extend the relation to nonconstant diffusion coefficients. For d ≥ 2, even a Gaussian target can have Lipschitz test functions whose Stein solutions have infinite M2.Source: §2.2 · pp. 7–10
2.3. Upper bounding the diffusion Stein discrepancy.
- When the drift and combined diffusion and stream coefficients are Lipschitz and sufficiently integrable, Wasserstein convergence of Qn to P implies discrepancy convergence to zero.The upper bound compares coefficient values at a draw from Qn and a draw from P.Source: §2.3 · p. 10
2.4. Extension to non-uniform Stein sets.
- Non-uniform Stein sets allow bounds tailored to known Stein factors, while the classical discrepancy is strongly equivalent to any such non-uniform discrepancy.When Stein factors are unknown or difficult to compute, the paper recommends the parameter-free classical or graph Stein set.Source: §2.3 · p. 10, §2.4 · p. 11
3. Sufficient conditions for Wasserstein decay.
The section gives sufficient conditions for exponential Wasserstein decay of Itō diffusions, then applies them to uniform and distant dissipativity. Examples show these conditions cover log-concave, multimodal, and heavy-tailed targets, including a non-constant diffusion coefficient.
3.1. Uniform dissipativity.
- For an Itô diffusion whose drift is −k/2-one-sided Lipschitz in G, whenever k′ < k the diffusion exhibits exponential Wasserstein decay at rate e−(k−k′)t/2 √(λmax(G)/λmin(G)).Source: §3.1 · pp. 11–12
3.2. Distant dissipativity, constant σ.
- For a Lipschitz drift, dissipativity at large distances—κ(r) ≥ k for some k > 0 for all sufficiently large r—ensures exponential Wasserstein decay under the stated conditions.When κ(r) ≥ 1/2 for r > 2Δ, the diffusion enjoys exponential Wasserstein decay [22, Lem. 1] and a Stein-discrepancy upper bound on Wasserstein distance.Source: §3.2 · p. 13
3.3. Distant dissipativity, non-constant σ.
- With non-constant diffusion coefficients, the Riemannian Langevin diffusion satisfies the Wasserstein-decay preconditions when VᵀV ≻ 0.The argument obtains a lower bound on κ(r) beyond a threshold, leading to exponential Wasserstein decay and a Wasserstein lower bound on the Stein discrepancy.Source: §3.3 · p. 15
4. Computing Stein discrepancies.
Spanner Stein discrepancies replace an infinite-dimensional optimization with constraints on sample points and selected pairs. Their equivalence to classical discrepancies preserves convergence-detection properties, while spanners and decoupled linear programs make computation tractable.
4.1. Spanner Stein discrepancies.
- A t-spanner graph yields a discrepancy between the classical value and κ_d t² times that value, preserving equivalence up to a dimension- and norm-dependent factor.Graph Stein sets constrain function values at sample points and smoothness on graph edges, rather than throughout the domain.Source: §4.1 · p. 16
4.2. Decoupled linear programs.
- For the ℓ1 norm, the graph optimization decouples into d parallel linear programs; the recommended measure uses a 2-spanner diffusion Stein discrepancy.Spanner construction leads to a finite-dimensional convex problem with O(n) variables and O(κ′_d n) constraints.Source: §4.2 · p. 17, §4.1 · p. 17
5. Numerical illustrations.
The numerical examples use diffusion Stein discrepancies to assess sample quality across biased samplers and quadrature rules. They show how the measure distinguishes missed modes, selects a useful sampler step size, tracks a bias-variance trade-off, and compares deterministic with random quadrature.
5. Numerical illustrations.
- The numerical illustrations deploy the discrepancies to select hyperparameters for biased samplers, quantify approximate-MCMC bias-variance trade-offs, and compare deterministic and random quadrature rules.The experiments use settings with a notion of surrogate ground truth for external validation.Source: §5 · pp. 17–18
5.1. A simple example.
- For Gaussian-mixture samples, the discrepancy falls at an n^-1/2 rate, while samples from one component remain bounded away from zero.Larger mode separation requires larger samples to distinguish the mixture from the single-mode sequence; the associated Wasserstein decay rate and Stein-to-Wasserstein bound also degrade.Source: §5.1 · p. 18, Figure 1 · p. 18
5.2. Selecting sampler hyperparameters.
- The Stein discrepancy selects step size ϵ = 10^-4, which also minimizes the surrogate Wasserstein measure and yields samples closest to the MARLA reference.ESS instead selects ϵ = 10^-2, producing overdispersed samples; ϵ = 10^-7 produces underdispersed samples due to slow mixing.Source: Figure 2 · p. 19, §5.2 · p. 20
5.3. Quantifying a bias-variance trade-off.
- At a small computational budget, the ARWMH sample with ϵ = 0.2 has higher Stein quality, but it is eventually overtaken by ϵ = 0.1, which has smaller asymptotic bias.The discrepancy results are comparable to expectation-error measures based on a surrogate chain, without requiring that additional chain.Source: §5.3 · pp. 20–21, Figure 3 · p. 21
5.4. Comparing quadrature rules.
- For larger sample sizes, FCFW underperforms FW and QMC in the diagnostics, while i.i.d. samples typically have lower median quality than deterministic counterparts.For the 1-Lipschitz test function h1, QMC gives a better quadrature estimate than FCFW.Source: §5.4 · p. 21, Figure 4 · p. 22
6. Connections and conclusions.
The discussion situates diffusion Stein discrepancies among alternative sample-quality measures and outlines extensions and scope boundaries. It also notes that the proof of Proposition 3 uses integrability and divergence-theorem arguments.
6. Connections and conclusions.
- The discussion contrasts finite-known-expectation diagnostics with MMDs, which detect arbitrary distributional biases but are computable only when their kernels can be integrated under the target.In practice, MMD is often approximated using a separate trustworthy sample from the target.Source: §6 · pp. 22–23
- Diffusion Stein operators are presented for pairing with RKHS unit balls or random-feature functions, beyond the graph and classical Stein sets considered here.The paper restricts its own focus to operators arising from diffusion generators.Source: §6 · p. 23
- The discussion is restricted to the Wasserstein distance, which requires finite means; a parallel Dudley-metric development could address targets with undefined means.The authors also note that the Lipschitz condition on drift and diffusion coefficients may be relaxable.Source: §6 · p. 24
APPENDIX A: PROOF OF PROPOSITION 3
- The proof of Proposition 3 applies dominated convergence and the divergence theorem on expanding balls, then uses the coarea formula and integrability of a and c.These are proof steps for establishing the proposition, rather than a numerical procedure.Source: Appendix A · p. 24
APPENDIX B: PROOF OF THEOREM 5
The proof establishes the Poisson-equation solution through strong continuity of the diffusion semigroup on a weighted function space. It then uses semigroup derivative bounds to establish regularity and Stein-factor bounds under additional smoothness conditions.
APPENDIX B: PROOF OF THEOREM 5
- The transition semigroup is shown to be strongly continuous on L, the weighted space (1 + ∥x∥2^2)C0(Rd), supporting the Poisson-equation argument.The proof uses path continuity for pointwise convergence and weak continuity to obtain strong continuity.Source: Appendix B · p. 25
- The proof constructs the solution uh as an integral of the semigroup applied to a centered test function, then establishes that it solves the Poisson equation and is Lipschitz.It also notes that the chosen Banach space matters: the semigroup fails to be strongly continuous on the weighted Cb space for an Ornstein-Uhlenbeck process.Source: Appendix B · p. 26
- Under additional coefficient and inverse-diffusion assumptions, the proof obtains a bounded continuous Hessian for the semigroup and integrates its bounds to control the solution's Hessian.Further smoothness assumptions yield a Hölder-continuous Hessian bound.Source: Appendix B · pp. 27–28
APPENDIX C: PROOF OF LEMMA 15
The appendix proves semigroup derivative estimates by differentiating the diffusion flow and controlling its first and second variations. These estimates provide the gradient and Hessian regularity and bounds used in the main Stein-factor argument.
APPENDIX C: PROOF OF LEMMA 15
- The proof derives the semigroup gradient bound from the Wasserstein decay rate and differentiates the diffusion flow to express first and second derivatives of Ptf.The first variation solves a linear stochastic differential equation, and the second variation is obtained by differentiating that equation.Source: Appendix C · p. 28
- Derivative-flow moment bounds justify exchanging derivatives and expectations, yielding a continuous Hessian for Ptf under the stated smoothness assumptions.The proof then bounds the Hessian using semigroup estimates and moment controls for the derivative flows.Source: Appendix C · pp. 29–30
C.1. Proof of Lemma 16: Derivative flow bounds.
- The second derivative-flow bound follows from Dynkin's formula, Cauchy-Schwarz and Young's inequalities, and Grönwall's inequality.The resulting bound is used in the semigroup Hessian estimate.Source: Appendix C · p. 31, §C.1 · pp. 32–33
APPENDIX D: PROOF OF THEOREM 6
The proof relates Wasserstein distance to diffusion Stein discrepancy through Stein-equation solutions, then smooths those solutions to obtain the regularity needed for Stein-set bounds. Gaussian convolution preserves the gradient representation, while the Theorem 7 argument uses smoothed derivative bounds.
APPENDIX D: PROOF OF THEOREM 6
- Gaussian convolution produces a smoothed Stein solution with a bounded Lipschitz gradient, while its operator value closely approximates the test function when the smoothing level is small.The approximation uses the Lipschitz continuity of the test function.Source: Appendix D · p. 33
D.1. Proof of Lemma 17: Smoothing by Gaussian convolution.
- For a continuously differentiable function, the gradient of its Gaussian convolution equals the convolution of its gradient.The proof also uses that the convolution does not increase the function's supremum bound.Source: §D.1 · p. 34
APPENDIX E: PROOF OF THEOREM 7
- Theorem 7 starts from a Lipschitz Stein-equation solution and smooths it because a Lipschitz gradient is needed for membership in a scaled classical Stein set.The argument assumes the listed drift and diffusion coefficient bounds are finite.Source: Appendix E · p. 35
E.1. Proof of the first inequality.
- Lemma 17 shows gh,s admits a Lipschitz gradient, ∇gh,s(x)=E[∇gh(x+sG)]; when ∇3b and ∇3σ are bounded and locally Lipschitz, smoothing gives M2(gh,s)=M1(∇gh,s).The argument fixes any ι∈(0,1).Source: Appendix E · p. 35, §E.1 · p. 35
E.2. Proof of the second inequality.
The proof bounds Wasserstein distance by optimizing over the smoothing scale, balancing the Stein discrepancy against smoothing-dependent terms. It treats small and larger discrepancies separately to establish the stated result.
E.2. Proof of the second inequality.
- dW∥·∥2(P, Qn) ≤ inf_s S(Qn, T, G∥·∥2) max(sr, β, Cιs^-ι) + sζ.The bound is stated as equation (34), with the smoothing scale optimized over s > 0.Source: §E.2 · p. 36
- For discrepancies below e^-1, the proof chooses ι = 1/log(1/S(Qn,T,G∥·∥)) to handle the small-discrepancy case.The stated argument uses x^(1/log(x^-1)) ≤ e for x in (0, e^-1).Source: §E.2 · p. 36
- For discrepancies at least e^-1, the proof instead fixes any ι in (0, 1), and combines both cases with equation (34) to obtain the result.Source: §E.2 · p. 36
APPENDIX F: PROOF OF PROPOSITION 8
The proof uses the zero expectation of the Stein operator under the target and bounds discrepancies through couplings. A second bound applies a concave truncation and Fenchel–Young inequality.
APPENDIX F: PROOF OF PROPOSITION 8
- The first advertised inequality follows by applying Fenchel–Young to a coupling and using the boundedness and Lipschitz constraints on the Stein function and its gradient.The proof takes a supremum over functions in the Stein set.Source: Appendix F · p. 37
- The second advertised inequality follows from the first using Jensen’s inequality, Hölder’s inequality, and the definition of Ws,∥·∥.Source: Appendix F · p. 37
- The final claim is bounded by c log(eµ0/c), after defining c as the truncated Wasserstein distance and applying Fenchel–Young, concavity, and Jensen’s inequality.Source: Appendix F · p. 37
APPENDIX G: PROOF OF THEOREM 10
The proof couples two Itô diffusions with a shared Wiener process and uses uniform dissipativity to bound their separation. It then applies the resulting estimate to the transition semigroup.
APPENDIX G: PROOF OF THEOREM 10
- The proof constructs diffusions from x and y using the same Wiener process, then applies Dynkin’s formula to their difference process.Source: Appendix G · p. 37
- Uniform dissipativity bounds the squared separation by the initial squared distance in the G norm.The passage identifies the latter with the squared G-Wasserstein distance between point masses at x and y.Source: Appendix G · p. 37
- The proof proceeds to the transition semigroup and invokes a lower bound involving the minimum eigenvalue of G1.Source: Appendix G · p. 37
APPENDIX H: PROOF OF THEOREM 11
The appendices develop estimates for Theorem 11 and Theorem 7 through coupling, smoothing and interpolation, and derivative-flow bounds. They use these tools to control semigroup derivatives.
APPENDIX H: PROOF OF THEOREM 11
- Theorem 11’s proof couples diffusions from arbitrary starting points and studies their difference process to obtain bounds for the chosen drift parameters.For a concave increasing function f with f(0)=0 and f′(0)=1, the drift term is bounded above by β_t; the argument then invokes results of [22].Source: Appendix H · pp. 37–38
SUPPLEMENTARY APPENDIX I: SMOOTHING AND INTERPOLATION
- Appendix I presents convolution-smoothing and seminorm-interpolation results, including a Lipschitz bound for smoothed functions in terms of the original function’s Hölder continuity.The results support the proof of Theorem 7.Source: SUPPLEMENTARY APPENDIX I: SMOOTHING… · p. 44
J.1. The J1,·,x terms.
- The J.1 bounds use Cauchy–Schwarz, derivative-flow bounds, and the Itô isometry to control terms in the third semigroup derivative estimate.The passage identifies the derivative-flow bound used for the estimate.Source: §J.1 · pp. 46–47
J.2. The J2,·,x terms.
- The J.2 bounds rewrite the unbounded f(Z_t,x) term using equation (29) and the chain rule, with resulting terms bounded in terms of M1(f).One displayed bound is proportional to M1(f).Source: §J.2 · pp. 47–48
J.3. The J3,· terms.
- The final J.3 expression is bounded in terms of M1(f) once the expected squared norm of the third directional derivative flow is bounded.The appendix establishes that flow bound in Section J.5.Source: §J.3 · p. 48
J.4. Semigroup third derivative bound.
- The semigroup third-derivative bound combines the termwise estimates with semigroup gradient and Hessian bounds.The proof adapts [8, Prop. 1.5.1].Source: §J.4 · p. 48
J.5. Third derivative flow bound.
- The third derivative-flow bound is developed using Dynkin’s formula and Grönwall’s inequality.The proof also applies Itô’s formula to the second derivative flow.Source: §J.5 · pp. 49–50