Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Stability of Mean-Field Variational Inference

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The mean-field variational approximation moves Lipschitz-continuously with the target distribution.

desk verdict New dimension-free Lipschitz stability for MFVI with a clean transport-map proof, but Theorem 2.3's reward bound rests on an identity that only holds for the target measure, so one main result is unproven as written. read the letter →

arxiv 2506.07856 v1 pith:TJUIXFIQ submitted 2025-06-09 math.PR math.FAmath.STstat.MLstat.TH

Shunan Sheng , Bohan Wu , Alberto González-Sanz , Marcel Nutz This is my paper
classification math.PRmath.FAmath.STstat.MLstat.TH MSC 90C2549Q2262F1549N80
keywords mean-fieldvariationalinferencestronglog-concavityWassersteindistanceoptimaltransportLipschitzstabilityBernstein–vonMisesempiricalBayesdistributedstochasticcontrol
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper establishes that the mean-field variational inference (MFVI) optimizer, the product measure closest to a target distribution in KL divergence, is stable under perturbations of the target when the target is strongly log-concave. The main result, Theorem 2.1, bounds the 2-Wasserstein distance between the optimizers by the inverse log-concavity constant times the $L^2$ difference of the target gradient potentials, with no dependence on dimension. Under stronger regularity, the optimizer is also differentiable in the target potential, with the derivative characterized by a weak PDE. These results matter because they turn heuristic robustness of variational Bayes into quantitative guarantees: small changes in prior or likelihood, or errors in estimated parameters, propagate proportionally into the approximate posterior, and the bounds remain useful in high dimension.

What carries the argument

The central object is the lifted MFVI functional $F_V(T) = -\sum_{i=1}^d \int \log T_i'(u_i)\,\rho_1(du_i) + \int V(T(u))\,\rho(du)$, defined on the convex cone $H_+$ of maps in the Gaussian Sobolev space $H(\rho)$ with $T_i' > 0$ a.e. Convexity of this functional turns the non-convex KL minimization over product measures into a calculus-of-variations problem, so the stability bound follows from comparing first variations at two optimizers, and the differentiability theorem follows from coercivity of the second-variation bilinear form $B_{\theta_0}$. Caffarelli's contraction theorem supplies the a priori bounds $1/\sqrt{\beta} \le (T^\theta_i)' \le 1/\sqrt{\alpha}$ that make the integrals finite.

What would settle it

For two Gaussians $N(0, \sigma^2 I_d)$ and $N(0, \tilde\sigma^2 I_d)$ the MFVI optimizers are explicit Gaussians with independent coordinates, making every term in Theorem 2.1 computable in closed form; checking the inequality $W_2 \le \alpha^{-1}\|\nabla V - \nabla\tilde V\|_{L^2(\nu^*)}$ over a grid of $\sigma$, $\tilde\sigma$, and $\alpha$ would settle the bound in the simplest nontrivial case, and any violation with $\alpha > 0$ would refute the theorem. A second check: construct a sequence of strongly log-concave targets with $\alpha \to 0$ and watch whether the optimizer jumps discontinuously under fixed-size potential perturbations, which would identify the curvature assumption as the essential one.

Watch

Extended reading notes

Core claim

On its own terms, the paper proves a dimension-free Lipschitz stability theorem: for two $C^2$ strongly log-concave targets $\pi, \tilde\pi$ with potentials $V, \tilde V$ and a common modulus $\alpha$, the MFVI optimizers satisfy $W_2(\tilde\nu^*, \nu^*) \le \alpha^{-1}\|\nabla\tilde V - \nabla V\|_{L^2(\tilde\nu^*)}$. This is achieved by lifting the non-convex MFVI problem to a convex optimization over optimal transport maps from a fixed Gaussian reference, where the optimizer is the componentwise increasing map $T = (T_1, \dots, T_d)$. The proof uses the first-order optimality condition at both optimizers and Caffarelli's contraction theorem to control the map differences. The second main theorem shows that $\theta \mapsto T^\theta$ is Fr\'echet differentiable in a Gaussian Sobolev space and characterizes the derivative as the unique weak solution of a PDE with bilinear form $B_{\theta_0}$; the Lax--Milgram theorem supplies existence and uniqueness via the $\alpha$-coercivity of that form.

Load-bearing premise

The proofs assume the target log-density is strongly convex with a known positive modulus $\alpha$; if $\alpha$ is zero, unknown, or the target is not log-concave, the Lipschitz constant blows up and the uniqueness and transport-map representation used throughout are no longer guaranteed.

Editorial extensions

If this is right

  • If a Bayesian posterior satisfies strong log-concavity, replacing the prior by a surrogate prior moves the variational posterior by at most the $L^2$ gradient difference of the log-priors divided by the posterior curvature, with no factor of dimension; this gives a computable uncertainty interval for any Lipschitz statistic without sampling the true posterior.
  • In empirical Bayes, an estimator $\hat\theta$ with error $|\hat\theta - \theta|$ propagates into the MFVI posterior with an explicit Lipschitz constant, so the approximation error of $\nu^*_{\hat\theta}$ relative to the true posterior $\pi_\theta$ is controlled by $|\hat\theta - \theta|$ plus the known MFVI approximation gap of the true target.
  • The quantitative Bernstein--von Mises bound $W_2^2(\nu^*_n, \gamma^*_n) \le 4b_n^2 d/(\alpha_n^3 n)$ implies that, when $d/n \to 0$ and curvature constants stay bounded, the mean-field posterior is close to a Gaussian at the mode and the variational mean attains the same $O(\sqrt{d/n})$ error as the mode.
  • Since the ELBO functional is also Lipschitz, the optimal value of distributed stochastic control problems is stable under changes in the utility function, with an explicit bound in terms of gradient differences of the utilities.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the lifting idea transfers to non-product variational families, such as structured covariance approximations, analogous stability and differentiability results should hold for those families, with the tangent space of the family replacing the product-map cone; the paper hints at this by calling the MFVI result a stepping stone for structured variational inference.
  • The $\alpha^{-1}$ constant suggests a sharp transition in robustness as log-concavity weakens: for nearly flat or multimodal targets, the mean-field optimizer may fail to be unique and small perturbations could produce large jumps, so practitioners should treat variational-Bayes robustness certificates only in the strongly log-concave regime.
  • A testable practical extension would use the derivative PDE to build a first-order correction to the MFVI optimizer under parameter changes, avoiding a full refit; the paper's $o(|\theta - \theta_0|)$ guarantee makes this correction certified in Wasserstein distance, and its accuracy could be benchmarked on Bayesian linear regression with varying precision parameter.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper develops a linearized optimal transport lifting of the mean-field variational inference (MFVI) problem for strongly log-concave targets, establishing an equivalence with a convex optimization problem over Gaussian Sobolev transport maps. On this basis it proves three main results: Theorem 2.1, a dimension-free Lipschitz stability bound for MFVI optimizers under perturbations of the target potential; Theorem 2.3, a stability bound for the optimal ELBO/reward; and Theorem 2.5, Frechet differentiability of the optimizer with respect to the target potential together with a PDE characterization of the derivative. The paper also derives explicit stability bounds, several Bayesian robustness applications, a quantitative Bernstein-von Mises theorem for MFVI, and an application to distributed stochastic control.

Significance. If the results hold, this is a substantial contribution: the lifting approach is elegant and promising, the dimension-free bound in Theorem 2.1 is a genuine improvement over dimension-dependent alternatives, and the quantitative Bernstein-von Mises theorem for MFVI is new. The proof of Theorem 2.1 is clean and self-contained, and the density bound in Lemma 6.1 together with the Lp estimates in Proposition 5.2 are useful technical tools in their own right. However, the proof of Theorem 2.3 rests on a false Stein-type identity, and the proof of Theorem 2.5 inherits a gap from that identity as well as an admissibility problem for its test functions. These are load-bearing issues for two of the three advertised main results, so the paper cannot be accepted in its current form.

major comments (3)
  1. [Section 4.3, Eq. (47)] The proof of Theorem 2.3 uses the identity E_{nu*}[Delta V(X) - <grad V(X), grad V(X)>] = 0, citing [25, Lemma 4.0.1]. That is Stein's identity for the target Gibbs measure pi proportional to e^{-V}, not for the MFVI optimizer nu*, which is a product measure and coincides with pi only in dimension one. For a product nu*, integration by parts gives E_{nu*}[Delta V] = -E_{nu*}[<grad V, grad log nu*>], which is not equal to E_{nu*}||grad V||^2 in general. This identity is used to assert that the integral of ||grad V(T^V)||^2 with respect to rho equals the integral of Delta V with respect to nu*, and is bounded by beta d. This controls the gradient term in the bound on B and is essential for the final inequality (49) and (8). Since the asserted estimate is not justified, Theorem 2.3 is not established as written.
  2. [Section 5, Lemma 5.1] The proof of Lemma 5.1 invokes the same identity (47) to show that the linear functional L_{T^theta,V_theta} is bounded on H(rho), and then uses this to extend the first-order optimality condition from smooth compactly supported directions to all R in H(rho). Because (47) is invalid for non-product MFVI optimizers, the boundedness step is unsupported and the differentiability proof of Theorem 2.5 inherits the gap. A repair may be possible, for instance via the density bound in Lemma 6.1 together with the Poincare inequality for strongly log-concave measures giving E_{nu*}||grad V_theta||^2 <= d beta^2 / alpha, but such an argument is not present and would change the constants in the proof.
  3. [Section 5, Proposition 5.2, Step 2] The test functions R_i defined in Step 2 of the proof of Proposition 5.2 have weak derivative |S_i'|^{p-2} S_i'. For R to belong to H(rho), this derivative must lie in L^2(rho_1), i.e. one needs S_i' in L^{2p-2}(rho_1). At that point the proof only has S in L^p from Step 1 and S' in L^2 from Remark 4.3, so the claimed admissibility R in H(rho) is not established. Consequently the bound (62) on the L^p norm of (T^theta)' - (T^theta0)' is not proved as written. Since Theorem 2.5 uses Proposition 5.2 in Step 2, this is another load-bearing gap in the differentiability argument; an L^{2p-2} a priori estimate or an extension of the bilinear relation (BR) to the relevant larger space is needed, and the manuscript supplies neither.
minor comments (4)
  1. [Example 3.5] The displayed formula for the epsilon-contamination bound contains a malformed expression with an unresolved radical and misplaced fraction; it should be rewritten as ||(q/p_epsilon)(grad log q - grad log p)||_{L^2(nu*)} with clear notation for the density ratio.
  2. [Corollary 2.4(i)] The exponent in the displayed bound, 2^{2p+d-4}/2, is ambiguous; using explicit parentheses such as 2^{(2p+d-4)/2} would improve readability.
  3. [Section 6, Lemma 6.1] In the proof of Lemma 6.1, the sentence 'As nu in P(R)^{otimes d}' should read 'As nu* in P(R)^{otimes d}' so that the second-moment estimate is applied to the optimizer.
  4. [Section 2.1.2] The normalization condition on V and tilde V in Theorem 2.3 is stated without comment; it would help to note explicitly that this ensures the mean of tilde V - V under tilde nu* is zero so that the Poincare inequality applies to the last term of (49).

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the MFVI stability bounds follow from a proved equivalence and first-order optimality, not from fitting or self-citation.

full rationale

Theorem 2.1 is derived from Lemma 4.1 (equivalence of MFVI and L-MFVI) and Lemma 4.2 (first-order condition), both proved in the paper. The bound W2(ν̃*,ν*) ≤ α^{-1}∥∇Ṽ-∇V∥_{L2(ν̃*)} follows from strong monotonicity of ∇V and the Cauchy-Schwarz inequality; the right-hand side is not an input but a computed norm at the optimizer. There are no fitted parameters; Corollary 2.4 only makes the unobservable norm explicit via Lemma 6.1. Theorem 2.5 uses Lax-Milgram on a coercive bilinear form; the derivative PDE is a characterization, not assumed. The only author self-citation is [41] (González-Sanz and Sheng) in Remark 2.8, where it is described as related to the weak formulation and not used in any proof. The uniqueness of the MFVI optimizer is imported from [55] (Lacker-Mukherjee-Yeung), whose authors do not overlap with the present paper, and is used as an external theorem. One flagged issue is a correctness gap, not circularity: in the proof of Theorem 2.3, the identity E_{ν*}[ΔV - ⟨∇V,∇V⟩] = 0 (Eq. 47, citing [25]) is Stein's identity for the target Gibbs measure and fails for the mean-field optimizer ν* unless ν*=π; this affects the proof of Theorem 2.3 and the boundedness argument in Lemma 5.1, but it is a mathematical error in an external lemma application, not a reduction of the result to its own assumptions. The central claim (Theorem 2.1) does not rely on this identity.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper's results rely on standard optimal transport theory (Brenier's theorem, Caffarelli's contraction theorem), the known existence/uniqueness of MFVI optimizers for log-concave targets ([55]), and Gaussian Sobolev space tools. No fitted parameters or invented entities are introduced. The strong log-concavity and smoothness of the target potentials are domain assumptions inherited from the problem setup.

assumptions (5)
  • standard math Brenier's theorem: optimal transport maps exist and are gradients of convex functions
    Used in Lemma 4.1 to parametrize measures by transport maps.
  • standard math Caffarelli's contraction theorem: the optimal map from standard Gaussian to an α-log-concave target is 1/√α-Lipschitz
    Used to show the transport maps lie in H(ρ) and to bound derivatives.
  • standard math Existence and uniqueness of MFVI optimizer for strongly log-concave targets ([55, Theorem 1.1])
    Basis for defining the optimizer; also gives α-log-concavity of ν*.
  • domain assumption Target measures are C2, α-log-concave and β-log-smooth with growth condition (5)
    Defines the admissible class of potentials; used throughout.
  • standard math Lax-Milgram theorem for the bilinear form B_θ0
    Used in Theorem 2.5 to define the derivative as solution of a PDE.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Stability of Mean-Field Variational Inference." pith.science (2026). https://pith.science/paper/TJUIXFIQ

@misc{pith2026250607856,
  author       = {Pith},
  title        = {Pith review of: Stability of Mean-Field Variational Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TJUIXFIQ}},
  note         = {Machine review of arXiv:2506.07856}
}
read the original abstract

Mean-field variational inference (MFVI) is a widely used method for approximating high-dimensional probability distributions by product measures. This paper studies the stability properties of the mean-field approximation when the target distribution varies within the class of strongly log-concave measures. We establish dimension-free Lipschitz continuity of the MFVI optimizer with respect to the target distribution, measured in the 2-Wasserstein distance, with Lipschitz constant inversely proportional to the log-concavity parameter. Under additional regularity conditions, we further show that the MFVI optimizer depends differentiably on the target potential and characterize the derivative by a partial differential equation. Methodologically, we follow a novel approach to MFVI via linearized optimal transport: the non-convex MFVI problem is lifted to a convex optimization over transport maps with a fixed base measure, enabling the use of calculus of variations and functional analysis. We discuss several applications of our results to robust Bayesian inference and empirical Bayes, including a quantitative Bernstein--von Mises theorem for MFVI, as well as to distributed stochastic control.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. The Influence Function of Transport-based Quantiles

    math.ST 2026-07 conditional novelty 8.0 of 10

    The influence function of multivariate transport quantiles has a pole-type singularity in dimension ≥2, so contamination near a quantile level yields unbounded first-order sensitivity.

Reference graph

Works this paper leans on

82 extracted references · 70 canonical work pages · cited by 1 Pith paper

  1. [1]

    C. D. Aliprantis and K. C. Border. Infinite dimensional analysis . Springer, Berlin, third edition, 2006

    2006

  2. [2]

    Arnese and D

    M. Arnese and D. Lacker. Convergence of coordinate ascent variational inference for log-concave measures via optimal transport. arXiv preprint arXiv:2404.08792 , 2024

    arXiv 2024

  3. [3]

    F. Augeri. Nonlinear large deviation bounds with applications to Wigner matrices and sparse Erd¨ os-R´ enyi graphs.Ann. Probab., 48(5):2404–2448, 2020

    2020

  4. [4]

    F. Augeri. A transportation approach to the mean-field approximation. Probab. Theory Related Fields, 180(1-2):1–32, 2021

    work page 2021

  5. [5]

    Bakry, I

    D. Bakry, I. Gentil, and M. Ledoux. Analysis and geometry of Markov diffusion operators , volume 348 of Grundlehren der mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer, Cham, 2014

    work page 2014

  6. [6]

    Basak and S

    A. Basak and S. Mukherjee. Universality of the mean-field for the Potts model. Probab. Theory Related Fields, 168(3-4):557–600, 2017

    work page 2017

  7. [7]

    A. Beck. First-order methods in optimization , volume 25 of MOS-SIAM Series on Optimization . Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA; Mathematical Optimization Society, Philadelphia, PA, 2017

    work page 2017

  8. [8]

    Berger and L

    J. Berger and L. M. Berliner. Robust Bayes and empirical Bayes analysis with ϵ-contaminated priors. Ann. Statist., 14(2):461–486, 1986

    work page 1986

Show all 82 references
  1. [9]

    J. O. Berger. Statistical decision theory and Bayesian analysis . Springer Series in Statistics. Springer-Verlag, New York, second edition, 1985

    1985

  2. [10]

    J. O. Berger. An overview of robust Bayesian analysis. Test, 3(1):5–124, 1994. With comments and a rejoinder by the author

    1994

  3. [11]

    P. J. Bickel and B. J. K. Kleijn. The semiparametric Bernstein-von Mises theorem. Ann. Statist., 40(1):206–237, 2012. 40

    2012

  4. [12]

    C. M. Bishop. Pattern recognition and machine learning . Information Science and Statistics. Springer, New York, 2006

    2006

  5. [13]

    D. M. Blei, A. Kucukelbir, and J. D. McAuliffe. Variational inference: a review for statisticians. J. Amer. Statist. Assoc. , 112(518):859–877, 2017

    2017

  6. [14]

    D. M. Blei, A. Y. Ng, and M. I. Jordan. Latent Dirichlet Allocation. J. Mach. Learn. Res. , 3(Jan):993–1022, 2003

    2003

  7. [15]

    V. I. Bogachev. Gaussian measures, volume 62 of Mathematical Surveys and Monographs . American Mathematical Society, Providence, RI, 1998

    1998

  8. [16]

    H. Brezis. Functional analysis, Sobolev spaces and partial differential equations . Universitext. Springer, New York, 2011

    2011

  9. [17]

    L. A. Caffarelli. Monotonicity properties of optimal transportation and the FKG and related inequalities. Comm. Math. Phys. , 214(3):547–563, 2000

    2000

  10. [18]

    Carbonetto and M

    P. Carbonetto and M. Stephens. Scalable variational inference for Bayesian variable selection in regression, and its accuracy in genetic association studies. Bayesian Anal. , 7(1):73–107, 2012

    2012

  11. [19]

    Carmona and F

    R. Carmona and F. Delarue. Probabilistic theory of mean field games with applications. I , volume 83 of Probability Theory and Stochastic Modelling . Springer, Cham, 2018. Mean field FBSDEs, control, and games

    2018

  12. [20]

    Carmona and F

    R. Carmona and F. c. Delarue. Forward-backward stochastic differential equations and controlled McKean-Vlasov dynamics. Ann. Probab., 43(5):2647–2700, 2015

    2015

  13. [21]

    Carmona, M

    R. Carmona, M. Lauri` ere, and Z. Tan. Model-free mean-field reinforcement learning: mean-field MDP and mean-field Q-learning. Ann. Appl. Probab., 33(6B):5334–5381, 2023

    2023

  14. [22]

    Celentano, Z

    M. Celentano, Z. Fan, L. Lin, and S. Mei. Mean-field variational inference with the TAP free energy: Geometric and statistical properties in linear models. arXiv preprint arXiv:2311.08442 , 2023

    2023 arXiv

  15. [23]

    Chatterjee and A

    S. Chatterjee and A. Dembo. Nonlinear large deviations. Adv. Math., 299:396–450, 2016

    2016

  16. [24]

    Chernozhukov, A

    V. Chernozhukov, A. Galichon, M. Hallin, and M. Henry. Monge-Kantorovich depth, quantiles, ranks and signs. Ann. Statist., 45(1):223–256, 2017

    2017

  17. [25]

    S. Chewi. Log-concave sampling (unfinished draft). 2023

    2023

  18. [26]

    Chewi and A.-A

    S. Chewi and A.-A. Pooladian. An entropic generalization of Caffarelli’s contraction theorem via covariance inequalities. C. R. Math. Acad. Sci. Paris , 361:1471–1482, 2023

    2023

  19. [27]

    J. A. Cuesta-Albertos and C. Matr´ an. Notes on the Wasserstein metric in Hilbert spaces. Ann. Probab., 17:1264–1276, 1989

    1989

  20. [28]

    Dal Maso

    G. Dal Maso. An introduction to Γ-convergence, volume 8 of Progress in Nonlinear Differential Equations and their Applications . Birkh¨ auser Boston, Inc., Boston, MA, 1993

    1993

  21. [29]

    Deb and B

    N. Deb and B. Sen. Multivariate rank-based distribution-free nonparametric testing using measure transportation. J. Amer. Statist. Assoc. , 118(541):192–207, 2023

    2023

  22. [30]

    Diaconis and D

    P. Diaconis and D. Ylvisaker. Conjugate priors for exponential families. Ann. Statist., 7(2):269– 281, 1979

    1979

  23. [31]

    Djellout, A

    H. Djellout, A. Guillin, and L. Wu. Transportation cost-information inequalities and applications to random dynamical systems and diffusions. Ann. Probab., 32(3B):2702–2732, 2004

    2004

  24. [32]

    Durmus and E

    A. Durmus and E. Moulines. High-dimensional Bayesian inference via the unadjusted Langevin algorithm. Bernoulli, 25(4A):2854–2882, 2019

    2019

  25. [33]

    R. Eldan. Taming correlations through entropy-efficient measure decompositions with ap- plications to mean-field approximation. Probab. Theory Related Fields , 176(3-4):737–755, 2020

    2020

  26. [34]

    L. C. Evans. Partial differential equations , volume 19 of Graduate Studies in Mathematics . American Mathematical Society, Providence, RI, second edition, 2010

    2010

  27. [35]

    W. Gabcke. Neue Herleitung und explizite Restabsch¨ atzung der Riemann-Siegel-Formel, 1979. PhD Thesis, Georg-August-Universit¨ at zu G¨ ottingen

    1979

  28. [36]

    Gangbo and R

    W. Gangbo and R. J. McCann. The geometry of optimal transportation. Acta Math. , 41 177(2):113–161, 1996

    1996

  29. [37]

    A. E. Gelfand and A. F. M. Smith. Sampling-based approaches to calculating marginal densities. J. Amer. Statist. Assoc. , 85(410):398–409, 1990

    1990

  30. [38]

    Ghosal and B

    P. Ghosal and B. Sen. Multivariate ranks and quantiles using optimal transport: consistency, rates and nonparametric testing. Ann. Statist., 50(2):1012–1037, 2022

    2022

  31. [39]

    Giacomini, T

    R. Giacomini, T. Kitagawa, and M. Read. Robust Bayesian analysis for econometrics. CEPR Discussion Paper No. DP16488 , 2021

    2021

  32. [40]

    Giordano, T

    R. Giordano, T. Broderick, and M. I. Jordan. Covariances, robustness, and variational Bayes. J. Mach. Learn. Res. , 19:Paper No. 51, 49, 2018

    2018

  33. [41]

    Gonz´ alez-Sanz and S

    A. Gonz´ alez-Sanz and S. Sheng. Linearization of Monge-Amp` ere equations and data science applications. arXiv preprint arXiv:2408.06534 , 2024

    2024

  34. [42]

    Guillin, C

    A. Guillin, C. L´ eonard, L. Wu, and N. Yao. Transportation-information inequalities for Markov processes. Probab. Theory Related Fields, 144(3-4):669–695, 2009

    2009

  35. [43]

    Gustafson

    P. Gustafson. Local sensitivity of posterior expectations. Ann. Statist., 24(1):174–195, 1996

    1996

  36. [44]

    Gustafson

    P. Gustafson. Local robustness in Bayesian analysis. In Robust Bayesian analysis , volume 152 of Lect. Notes Stat. , pages 71–88. Springer, New York, 2000

    2000

  37. [45]

    Gustafson and L

    P. Gustafson and L. Wasserman. Local sensitivity diagnostics for Bayesian inference. Ann. Statist., 23(6):2153–2167, 1995

    1995

  38. [46]

    Hallin, E

    M. Hallin, E. del Barrio, J. Cuesta-Albertos, and C. Matr´ an. Distribution and quantile functions, ranks and signs in dimension d: a measure transportation approach. Ann. Statist., 49(2):1139–1165, 2021

    2021

  39. [47]

    W. K. Hastings. Monte Carlo sampling methods using Markov chains and their applications. Biometrika, 57(1):97–109, 1970

    1970

  40. [48]

    He and Q.-M

    X. He and Q.-M. Shao. On parameters of increasing dimensions. J. Multivariate Anal. , 73(1):120–135, 2000

    2000

  41. [49]

    Jiang, S

    Y. Jiang, S. Chewi, and A.-A. Pooladian. Algorithms for mean-field variational inference via polyhedral optimization in the Wasserstein space. In S. Agrawal and A. Roth, editors, Proceedings of Thirty Seventh Conference on Learning Theory , volume 247 of Proceedings of Machine...

    2024

  42. [50]

    Jiang, S

    Y. Jiang, S. Chewi, and A.-A. Pooladian. Algorithms for mean-field variational inference via polyhedral optimization in the Wasserstein space. In The Thirty Seventh Annual Conference on Learning Theory, pages 2720–2721. PMLR, 2024

    2024

  43. [51]

    M. I. Jordan, Z. Ghahramani, T. S. Jaakkola, and L. K. Saul. An introduction to variational methods for graphical models. Machine learning, 37:183–233, 1999

    1999

  44. [52]

    Katsevich

    A. Katsevich. Improved scaling with dimension in the Bernstein-von Mises theorem for two statistical models. arXiv:2308.06899, 2023

    2023 arXiv

  45. [53]

    D. P. Kingma and M. Welling. An introduction to variational autoencoders. Found. Trends Mach. Learn., 12(xx):1–18, 2019

    2019

  46. [54]

    D. Lacker. Independent projections of diffusions: Gradient flows for variational inference and optimal mean field approximations. arXiv preprint arXiv:2309.13332 , 2023

    2023 arXiv

  47. [55]

    Lacker, S

    D. Lacker, S. Mukherjee, and L. C. Yeung. Mean field approximations via log-concavity. Int. Math. Res. Not. IMRN , (7):6008–6042, 2024

    2024

  48. [56]

    Lacker and L

    D. Lacker and L. C. Yeung. A characterization of transportation-information inequalities for Markov processes in terms of dimension-free concentration. Ann. Inst. Henri Poincar´ e Probab. Stat., 59(1):364–377, 2023

    2023

  49. [57]

    LeCun, S

    Y. LeCun, S. Chopra, R. Hadsell, M. Ranzato, and F. J. Huang. A tutorial on energy-based learning. Predicting Structured Data, 2006

    2006

  50. [58]

    Lott and C

    J. Lott and C. Villani. Ricci curvature for metric-measure spaces via optimal transport. Ann. of Math. (2) , 169(3):903–991, 2009

    2009

  51. [59]

    Mahajan, N

    A. Mahajan, N. C. Martins, M. C. Rotkowitz, and S. Y¨ uksel. Information structures in optimal decentralized control. In 51st IEEE Conference on Decision and Control (CDC) , pages 42 1291–1306, 2012

    2012

  52. [60]

    J. W. Miller. Asymptotic normality, concentration, and coverage of generalized posteriors. J. Mach. Learn. Res., 22:Paper No. 168, 53, 2021

    2021

  53. [61]

    Mukherjee, B

    S. Mukherjee, B. Sen, and S. Sen. A mean field approach to empirical Bayes estimation in high-dimensional linear regression. arXiv preprint arXiv:2309.16843 , 2023

    2023 arXiv

  54. [62]

    Mukherjee and S

    S. Mukherjee and S. Sen. Variational inference in high-dimensional linear regression. J. Mach. Learn. Res., 23:Paper No. [304], 56, 2022

    2022

  55. [63]

    Nayyar, A

    A. Nayyar, A. Mahajan, and D. Teneketzis. Decentralized stochastic control with partial history sharing: A common information approach. IEEE Transactions on Automatic Control , 58(7):1644–1658, 2013

    2013

  56. [64]

    Neiswanger and A

    W. Neiswanger and A. Ramdas. Uncertainty quantification using martingales for misspecified Gaussian processes. In Algorithmic learning theory, volume 132 of Proc. Mach. Learn. Res. (PMLR), page 22. 2021

    2021

  57. [65]

    Opper and D

    M. Opper and D. Saad. Advanced Mean Field Methods: Theory and Practice . MIT press, 2001

    2001

  58. [66]

    F. Otto. The geometry of dissipative evolution equations: the porous medium equation. Comm. Partial Differential Equations , 26(1-2):101–174, 2001

    2001

  59. [67]

    G. Parisi. Mean field theory for spin glasses. Physics Reports, 67:25–28, 1980

    1980

  60. [68]

    C. P. Robert and G. Casella. Monte Carlo statistical methods . Springer Texts in Statistics. Springer-Verlag, New York, second edition, 2004

    2004

  61. [69]

    Santambrogio

    F. Santambrogio. Optimal Transport for Applied Mathematicians , volume 87 of Progress in Nonlinear Differential Equations and their Applications . Birkh¨ auser/Springer, Cham, 2015

    2015

  62. [70]

    Sivaganesan and J

    S. Sivaganesan and J. O. Berger. Ranges of posterior measures for priors with unimodal contaminations. Ann. Statist., 17(2):868–889, 1989

    1989

  63. [71]

    Tierney and J

    L. Tierney and J. B. Kadane. Accurate approximations for posterior moments and marginal densities. J. Amer. Statist. Assoc. , 81(393):82–86, 1986

    1986

  64. [72]

    A. W. van der Vaart. Asymptotic statistics , volume 3 of Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press, Cambridge, 1998

    1998

  65. [73]

    C. Villani. Topics in Optimal Transportation, volume 58 of Graduate Studies in Mathematics . American Mathematical Society, Providence, RI, 2003

    2003

  66. [74]

    Villani et al

    C. Villani et al. Optimal transport: old and new , volume 338. Springer, 2009

    2009

  67. [75]

    M. J. Wainwright and M. I. Jordan. Graphical models, exponential families, and variational inference. Found. Trends Mach. Learn., 1:1–305, 2008

    2008

  68. [76]

    Wang and D

    C. Wang and D. M. Blei. Variational inference in nonconjugate models. J. Mach. Learn. Res. , 14:1005–1031, 2013

    2013

  69. [77]

    G. Wang, A. Sarkar, P. Carbonetto, and M. Stephens. A simple new approach to variable selection in regression, with application to genetic fine mapping. J. R. Stat. Soc. Ser. B. Stat. Methodol., 82(5):1273–1300, 2020

    2020

  70. [78]

    W. Wang, D. Slepˇ cev, S. Basu, J. A. Ozolek, and G. K. Rohde. A linear optimal transportation framework for quantifying and visualizing variations in sets of images. Int. J. Comput. Vis. , 101(2):254–269, 2013

    2013

  71. [79]

    Wang and D

    Y. Wang and D. Blei. Frequentist consistency of variational Bayes. J. Amer. Statist. Assoc. , 114:1147–1161, 2019

    2019

  72. [80]

    J. Yan. Nonlinear large deviations: beyond the hypercube. Ann. Appl. Probab., 30(2):812–846, 2020

    2020

  73. [81]

    Y. Yao, A. Vehtari, D. Simpson, and A. Gelman. Using stacking to average Bayesian predictive distributions (with discussion). Bayesian Anal., 13(3):917–1003, 2018. Including a rejoinder by the authors

    2018

  74. [82]

    Zhang and Y

    Y. Zhang and Y. Yang. Bayesian model selection via mean-field variational approximation. J. R. Stat. Soc. Ser. B. Stat. Methodol. , 86(3):742–770, 2024. 43

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.