A Diffusion Process Perspective on Posterior Contraction Rates for Parameters

Martin J. Wainwright; Michael I. Jordan; Nhat Ho; Peter Bartlett; Wenlong Mou

REVIEW 3 major objections 3 minor 42 references

A Diffusion Process Perspective on Posterior Contraction Rates for Parameters

T0 review · 3 major / 3 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read This paper argues that posterior contraction rates for Bayesian parameters are governed by the stationary moments of a Langevin diffusion and, in weakly concave models, by a single fixed-point equation linking the population likelihood's…

desk verdict The diffusion framework is a fresh, promising direction, but Theorem 2's proof has a sign error in Lemma 1 and an unproved Lq convergence step; the promised BvM result also isn't there. read the letter →

arxiv 1909.00966 v2 pith:OW3O4T5Q submitted 2019-09-03 math.ST stat.TH

Wenlong Mou , Nhat Ho , Martin J. Wainwright , Peter Bartlett , Michael I. Jordan This is my paper · ORCID

classification math.STstat.TH MSC 62F1562F1260J6062G20

keywords posteriorcontractionrateLangevindiffusionstochasticdifferentialequationweakconcavityBayesianlogisticregressionsingleindexmodelover-specifiedGaussianmixtureBernstein-vonMises

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 a general route to posterior contraction rates for parameters by viewing the posterior as the stationary law of a Langevin stochastic differential equation and bounding moments of that process. In strongly concave population log-likelihoods the method yields the parametric rate $\sqrt{(d+\log(1/\delta)+B)/(n\mu)}$. In weakly concave models the posterior radius is controlled by the unique positive solution of the fixed-point equation $\psi(z)=\varepsilon(n,\delta)\zeta(z)z+(B+d\log(1/\delta))/n$, where $\psi$ measures how strongly the population likelihood pulls toward the true parameter and $\zeta$ controls the gradient fluctuation between sample and population likelihoods. The same calculation produces concrete rates for Bayesian logistic regression, polynomial-link single index models, and over-specified Gaussian mixtures, including a dimension dependence $d^{1/4}$ for the mixture case.

What carries the argument

The central object is the Langevin SDE whose stationary law is the posterior, combined with Itô-calculus moment control for $\mathbb{E}\|\theta_t-\theta^*\|_2^p$. For the weakly concave analysis, the paper defines auxiliary functions $\nu_p(r)=\psi(r^{1/(p-1)})r^{(p-2)/(p-1)}$ and $\tau_p$ by $\tau_p(r^{p-1}\zeta(r))=r^{p-2}\psi(r)$; Assumption W.3 makes these functions convex, so Jensen's inequality turns the moment recursion into the limiting equation (9). The convexity of $r\mapsto\psi(\xi(r))$, where $\xi$ inverts $r\mapsto r\zeta(r)$, is what guarantees uniqueness of the positive solution.

What would settle it

A concrete check would be to exhibit a Bayesian model satisfying Assumptions W.1, W.2, and W.4 whose local curvature and perturbation growth violate W.3—for instance local $\psi(r)=r^2$ with $\zeta(r)=r^2$—and then compute the posterior mass outside the radius predicted by equation (9). A direct numerical or exact calculation for moderate $n$ would show whether inequality (10) still holds; if it fails, the theorem's scope is refuted, and if it holds, W.3 is not necessary.

Watch

Extended reading notes

Core claim

The paper's central claim is that posterior contraction is a diffusion-moment phenomenon: because the posterior density is the stationary distribution of the SDE $d\theta_t=\frac12\nabla F_n(\theta_t)dt+\frac{1}{2n}\nabla\log\pi(\theta_t)dt+\frac{1}{\sqrt{n}}dB_t$, bounding $\mathbb{E}\|\theta_t-\theta^*\|_2^p$ along the path and passing to the limit $t\to\infty$ controls posterior mass near $\theta^*$. Under strong concavity of the population log-likelihood $F$, Theorem 1 gives a contraction radius of order $\sqrt{d/(n\mu)}$ plus a perturbation term $\varepsilon_2(n,\delta)/\mu$. Under weak concavity, Theorem 2 shows that the radius is the unique positive solution of the nonlinear equation $\psi(z)=\varepsilon(n,\delta)\zeta(z)z+(B+d\log(1/\delta))/n$, with $\psi$ encoding the weak-concavity geometry and $\zeta$ the stochastic perturbation growth; the result is applied to logistic regression, single index models with $g(r)=r^p$ and $\theta^*=0$, and over-specified location Gaussian mixtures.

Load-bearing premise

The weakly concave result depends on two technical inequalities relating the likelihood's curvature curve $\psi$ and the error-growth curve $\zeta$; they are verified for the paper's three examples but not derived from more basic assumptions. If a model does not satisfy them, the fixed-point radius is not guaranteed.

Editorial extensions

If this is right

For strongly concave population log-likelihoods, the posterior contracts at the parametric $\sqrt{d/n}$ rate with explicit dependence on the prior concentration constant $B$ and the strong-concavity constant $\mu$, without requiring a bounded parameter space.
For weakly concave models, the posterior radius is the unique solution of $\psi(z)=\varepsilon(n,\delta)\zeta(z)z+(B+d\log(1/\delta))/n$; for local power forms $\psi(r)=r^\alpha$ and $\zeta(r)=r^\beta$ this yields rates of order $(d/n)^{1/(2(\alpha-\beta-1))}$ or $(d/n)^{1/\alpha}$, whichever is larger.
Bayesian logistic regression has posterior contraction rate $(d/n)^{1/2}$.
Bayesian single index models with polynomial link $g(r)=r^p$, $p\ge2$, and true parameter $\theta^*=0$ have posterior contraction rate $(d/n)^{1/(2p)}$.
Over-specified location Gaussian mixtures have posterior contraction rate of order $(d/n)^{1/4}$, including a novel $d^{1/4}$ dimension dependence and no boundedness assumption on the parameter space.
Because the SDE moment bounds are non-asymptotic in time, the same technique also addresses approximate posterior distributions produced by Langevin sampling procedures.

Reading between the lines

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

Beyond the paper: the same moment-control scheme should give finite-time contraction bounds for Langevin Monte Carlo samplers, since inequality (33) is non-asymptotic in $t$; the paper notes the link to approximate posteriors but does not develop sampler-specific mixing rates.
Beyond the paper: equation (9) reads as an oracle-style trade-off between statistical bias, $\psi^{-1}((B+d\log(1/\delta))/n)$, and stochastic fluctuation, $\varepsilon\zeta(z)z$; one could try to match $\psi$ and $\zeta$ to a model's Fisher-information degeneracy to conjecture minimax lower bounds of the same shape.
Beyond the paper: the differential inequalities in Assumption W.3 look like a shape condition on the pair $(\psi,\zeta)$: $\psi$ must be sufficiently more convex than $\zeta$ at every scale. Models with locally flat likelihoods paired with heavy-tailed gradient noise fall outside the theorem, even if the fixed-point equation still has a solution.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The diffusion-process framework here is genuinely new and worth engaging with, but the main weak-concavity theorem is not proven as written. The stress-test note is correct: Lemma 1 in Appendix B.1 is internally inconsistent as stated—non-increasing with φ(c)=0 and φ(t)≥0 for t>c forces φ≡0—and the function φ in Section 5.2 has the opposite sign. The proof then asserts that the process (θ_t) converges in Lq for arbitrarily large q, which Proposition 1 does not supply; it only gives L2 convergence of densities. Those two gaps are load-bearing, since they produce the limit that drives the contraction bound.

What is genuinely good: the central idea of bounding posterior moments through the Langevin SDE is a real departure from sieve-based arguments. Theorem 1 in the strongly concave case is clean and delivers the expected (d/n)^{1/2} rate with explicit dimension and prior dependence. The d^{1/4} scaling for over-specified mixture locations appears new, and the examples are worked out in substantial detail—the perturbation bounds in the appendices are real work. The assumptions W.1–W.4 are stated clearly and the examples do verify them.

Where it is soft: the abstract promises a non-asymptotic Bernstein-von-Mises guarantee that never appears in the body, and Section 6 explicitly says the weak-concavity setting cannot have a Gaussian approximation. That sentence should be fixed or deleted. There are minor typos as well—the closed-form expression for F_I in Section 4.2 has a stray '2', for instance. The sign error in Lemma 1 looks repairable: the intended lemma should have φ(t)≤0 to the right of c, and the Lq convergence step can plausibly be replaced by a limsup argument. But 'likely repairable' is not the same as proven.

Who is this for: people working on Bayesian asymptotics for weakly identified parameters. The framework is promising and the examples are informative, but as it stands I would not rely on Theorem 2. I would still send it to a serious referee—the method deserves scrutiny and the gaps are plausibly fixable. My recommendation: engage with it, ask for a revision that fixes Lemma 1, supplies a proof of the moment convergence or replaces the limit with limsup, and either delivers the promised BvM result or removes it from the abstract.

Referee Report

3 major / 3 minor

Summary. The paper develops a diffusion-process (Langevin SDE) framework for posterior contraction rates of parameters. The posterior is represented as the stationary distribution of the SDE (5), and posterior concentration is bounded through moment control of the SDE combined with Fatou's lemma. Under strong concavity of the population log-likelihood (Assumptions A, B, S.1, S.2), Theorem 1 gives a contraction radius of order sqrt((d + log(1/delta) + B)/(n mu)) + eps_2/mu. Under weak concavity (Assumptions W.1-W.4), Theorem 2 characterizes the radius as the unique positive solution z* of the nonlinear equation psi(z) = eps(n,delta) zeta(z) z + (B + d log(1/delta))/n. Corollaries derive rates (d/n)^{1/2} for Bayesian logistic regression, (d/n)^{1/(2p)} for polynomial single-index models, and (d/n)^{1/4} for over-specified location Gaussian mixtures. Proofs use Ito calculus, Burkholder-Davis-Gundy inequalities, empirical-process concentration, and several auxiliary lemmas deferred to the appendices.

Significance. If the weak-concavity result can be repaired, the framework is a useful complement to density-based posterior contraction techniques: it avoids bounded parameter spaces and gives explicit dimension dependence. The unified fixed-point equation (9) and the new d^{1/4} scaling for over-specified Gaussian mixtures are valuable contributions, and the perturbation bounds in Appendix A are substantial. The paper also provides detailed proofs for the examples. However, because the proof of Theorem 2 currently contains a contradictory lemma and an unproved convergence step, the main weak-concavity result is not established as written.

major comments (3)

[5.2 / Appendix B.1] Lemma 1 is internally inconsistent as stated. It assumes that phi is non-increasing with phi(c)=0 and phi(t) >= 0 for all t > c; non-increasing together with phi(c)=0 forces phi(t) <= 0 for t > c, so the only function satisfying all three conditions is phi identically zero on [c, infinity). The proof then sets delta = -sup_{s >= c+epsilon} phi(s) < 0, which requires phi(s) < 0 for s > c+epsilon, contradicting the stated sign condition. Moreover, the application in Theorem 2 defines phi(r) = -r + eps tau^{-1}(r) + ((B+(p-1)d)/n) nu^{-1}(r)^{(p-2)/(p-1)}, which is concave with phi(r) > 0 for r < r* and phi(r) < 0 for r > r*; this phi is neither non-increasing nor non-decreasing globally. The intended comparison argument is plausible, but the lemma and its application must be repaired before Theorem 2 is established.
[5.2] The proof asserts that 'the process (theta_t : t >= 0) converges in Lq norm for arbitrarily large q' and uses this to conclude that lim_{t->infinity} R_p(t) exists. This assertion is unsupported. Proposition 1 guarantees only L2 convergence of the densities of theta_t to the posterior density; it does not imply convergence of the process in Lq norm, nor does it imply convergence of polynomial moments without additional uniform integrability arguments. The step lim_{t->infinity} R_p(t) exists and is bounded by r* is therefore not justified as written. This is load-bearing for Theorem 2.
[Abstract / Section 6] The abstract promises 'non-asymptotic versions of a Bernstein-von-Mises guarantee for the posterior', but no such theorem appears anywhere in the body. The only related discussion is in Section 6, which states that under weak concavity the posterior cannot be approximated by a Gaussian. Either a non-asymptotic Bernstein-von-Mises result should be stated and proved (presumably for the locally strongly concave case), or the abstract's claim should be removed.

minor comments (3)

[4.2] The displayed closed-form expression for the single-index population log-likelihood F_I(theta) has the wrong sign and an incorrect constant: since Y = epsilon with theta* = 0, one expects -1/2 - ((2p-1)!!/2) ||theta||^{2p} plus a constant, not '1 + ... / 2'. Only the gradient bound (24a) is used later, so this is a typo, but it should be corrected.
[5.1] In the proof of Theorem 1 the definition 'alpha = 1/(2 mu) - eps_1(n,delta) > mu/6' is dimensionally and algebraically suspicious; the later combination of J1 and J5 requires alpha <= 2 mu/3 - eps_1(n,delta) rather than the displayed expression. Please check and correct the displayed definition of alpha.
[4.1] In display (17b) the supremum is written over theta in R^d, but the surrounding text says 'for any r > 0'; the statement should clarify that the bound is uniform over balls B(theta*, r) or over all of R^d, since the two formulations are different.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the contraction rates are proved from stated structural assumptions via diffusion moment bounds, and the example rates are derived, not fitted.

full rationale

The paper's central derivation is self-contained in the relevant sense. Theorem 2 characterizes the contraction radius z*(n,δ) as the unique positive solution of equation (9), but that equation is not assumed as an ansatz: it is obtained from the moment bound (33), which is proved from Itô's formula together with Assumptions W.1–W.3 (weak concavity, perturbation control, and convexity of the auxiliary functions). The rates quoted in Corollaries 2–4 are then computed by verifying W.1–W.4 for each model from explicit population log-likelihood inequalities such as (17a), (24a), (29a) and from uniform gradient perturbation bounds such as (17b), (24b), (29b); none of these bounds is tuned to match the final posterior rate. The only citation with overlapping authorship is [9] (Dwivedi et al.), used in Appendix A.3 to obtain the model-specific bounds (29a) and (29b) for the Gaussian mixture example. That cited work is a separate study of EM convergence rates; its bounds are population/empirical-process facts that do not presuppose the posterior contraction rate proved here, so the citation is real evidence rather than a load-bearing circular premise. The manuscript does contain a separate rigor issue in the proof of Theorem 2 — the unproved assertion that (θ_t) converges in Lq for arbitrarily large q and an apparent sign inconsistency in Lemma 1 — but those are validity/correctness concerns, not circular reductions of the result to its own inputs. Accordingly, no circular step is identified and the circularity score is 0.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new physical or probabilistic entities beyond the diffusion process, which is a standard mathematical tool. The central claim rests on the technical assumptions W.1-W.4 and A-B, which are domain assumptions verified only for the examples. No parameters are fitted to data; the constants in the rates are universal or problem-specific but not calibrated.

assumptions (6)

standard math The Langevin SDE defined in (5) has the posterior distribution as its unique stationary distribution and converges in L2.
This is Proposition 1, quoted from known results on Langevin diffusions [20,26]; it is the bridge between posterior and diffusion used throughout.
standard math Fatou's lemma inequality (6) allows posterior moment bounds to be obtained from limiting diffusion moments.
Invoked in Section 2.2 to justify that bounding E||θ_t||^p bounds posterior moments.
domain assumption The empirical gradient perturbation bounds W.2 and S.2 hold uniformly on balls around θ*.
Stated in Section 3; verified for the examples in Section 4 and appendices, but not proven generally.
domain assumption Assumptions A and B on the prior hold (Lipschitz log-density gradient and a finite B).
Section 2.3; used to control the prior drift term in the moment inequalities.
domain assumption W.3 differential inequalities hold for the model's ψ and ζ.
Section 3.2; needed for convexity of auxiliary functions in Theorem 2's proof.
standard math The fixed-point equation (9) has a unique positive solution under W.4.
Proven in Section 5.2 using convexity of r→ψ(ξ(r)).

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Diffusion Process Perspective on Posterior Contraction Rates for Parameters." pith.science (2026). https://pith.science/paper/OW3O4T5Q

@misc{pith2026190900966,
  author       = {Pith},
  title        = {Pith review of: A Diffusion Process Perspective on Posterior Contraction Rates for Parameters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OW3O4T5Q}},
  note         = {Machine review of arXiv:1909.00966}
}

read the original abstract

We analyze the posterior contraction rates of parameters in Bayesian models via the Langevin diffusion process, in particular by controlling moments of the stochastic process and taking limits. Analogous to the non-asymptotic analysis of statistical M-estimators and stochastic optimization algorithms, our contraction rates depend on the structure of the population log-likelihood function, and stochastic perturbation bounds between the population and sample log-likelihood functions. Convergence rates are determined by a non-linear equation that relates the population-level structure to stochastic perturbation terms, along with a term characterizing the diffusive behavior. Based on this technique, we also prove non-asymptotic versions of a Bernstein-von-Mises guarantee for the posterior. We illustrate this general theory by deriving posterior convergence rates for various concrete examples, as well as approximate posterior distributions computed using Langevin sampling procedures.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 41 canonical work pages

[1]

Barron, M

A. Barron, M. Schervish, and L. Wasserman. The consisten cy of posterior distributions in nonparametric problems. Ann. Statist , 27:536–561, 1999. (Cited on page 1.)

work page 1999
[2]

Bhattacharya, D

A. Bhattacharya, D. Pati, , and D. B. Dunson. Anisotropic function estimation using multi-bandwidth Gaussian processes. Annals of Statistics , 42:352–381, 2014. (Cited on page 2.)

work page 2014
[3]

D. Blei, A. Ng, and M. Jordan. Latent Dirichlet allocatio n. J. Mach. Learn. Res , 3:993– 1022, 2003. (Cited on pages 2 and 20.)

work page 2003
[4]

Boucheron, G

S. Boucheron, G. Lugosi, and P. Massart. Concentration Inequalities: A Nonasymptotic Theory of Independence . Oxford University Press, 2016. (Cited on pages 27 and 34.)

work page 2016
[5]

R. J. Carroll and P. Hall. Optimal rates of convergence fo r deconvolving a density. Journal of American Statistical Association , 83:1184–1186, 1988. (Cited on page 10.)

work page 1988
[6]

J. Chen. Optimal rate of convergence for ﬁnite mixture mo dels. Annals of Statistics , 23(1):221–233, 1995. (Cited on pages 3 and 14.)

work page 1995
[7]

de Jonge and J

R. de Jonge and J. H. van Zanten. Adaptive nonparametric B ayesian inference using location-scale mixture priors. Annals of Statistics , 38:3300–3320, 2010. (Cited on page 2.)

work page 2010
[8]

J. L. Doob. Application of the theory of martingales. Actes du ColloqueInternational Le Calcul des Probabilit´ es et ses applications (Lyon, 28 Juin 3 J uillet, 1948) , pages 23–27,

work page 1948

Show all 42 references

[9]

Dwivedi, N

R. Dwivedi, N. Ho, K. Khamaru, M. J. Wainwright, M. I. Jord an, and B. Yu. Singularity, misspeciﬁcation, and the convergence rate of EM. arXiv preprint arXiv:1810.00828, 2018. (Cited on page 31.)

2018 arXiv
[10]

D. A. Freedman. On the asymptotic behavior of Bayes esti mates in the discrete case. Annals of Statistics , 34:13861403, 1963. (Cited on page 1.)

1963
[11]

D. A. Freedman. On the asymptotic behavior of Bayes esti mates in the discrete case.II. Annals of Statistics , 36:454456, 1965. (Cited on page 1.) 34

1965
[12]

Gao and H

C. Gao and H. H. Zhou. Rate exact Bayesian adaptation wit h modiﬁed block priors. Annals of Statistics , 44:318–345, 2016. (Cited on page 2.)

2016
[13]

Ghosal, J

S. Ghosal, J. K. Ghosh, and A. van der Vaart. Convergence rates of posterior distributions. Annals of Statistics , 28:500–531, 2000. (Cited on page 2.)

2000
[14]

Ghosal and A

S. Ghosal and A. van der Vaart. Entropies and rates of con vergence for maximum likelihood and bayes estimation for mixtures of normal dens ities. Annals of Statistics , 29:1233–1263, 2001. (Cited on pages 2 and 12.)

2001
[15]

Ghosal and A

S. Ghosal and A. van der Vaart. Posterior convergence ra tes of Dirichlet mixtures at smooth densities. Annals of Statistics , 35:697–723, 2007. (Cited on page 2.)

2007
[16]

Ho and X

N. Ho and X. Nguyen. Convergence rates of parameter esti mation for some weakly identiﬁable ﬁnite mixtures. Annals of Statistics , 44:2726–2755, 2016. (Cited on page 2.)

2016
[17]

Ishwaran, L

H. Ishwaran, L. F. James, and J. Sun. Bayesian model sele ction in ﬁnite mixtures by marginal density decompositions. Journal of the American Statistical Association , 96:1316–1332, 2001. (Cited on page 14.)

2001
[18]

B. J. K. Kleijn and A. W. van der Vaart. Misspeciﬁcation i n inﬁnite-dimensional Bayesian statistics. Annals of Statistics , 34:837–877, 2006. (Cited on page 2.)

2006
[19]

B. Lindsay. Mixture Models: Theory, Geometry and Applications . In NSF-CBMS Re- gional Conference Series in Probability and Statistics. IM S, Hayward, CA., 1995. (Cited on page 12.)

1995
[20]

X. Mao. Stochastic Diﬀerential Equations and Applications . Elsevier, 2007. (Cited on page 5.)

2007
[21]

McCullagh and J

P. McCullagh and J. A. Nelder. Generalized Linear Models . Chapman and Hall/CRC,
[22]

X. Nguyen. Convergence of latent mixing measures in ﬁni te and inﬁnite mixture models. Annals of Statistics , 4(1):370–400, 2013. (Cited on pages 2, 3, 13, and 14.)

2013
[23]

X. Nguyen. Borrowing strength in hierarchical Bayes: c onvergence of the Dirichlet base measure. Bernoulli, 22:1535–1571, 2016. (Cited on page 2.)

2016
[24]

Pritchard, M

J. Pritchard, M. Stephens, and P. Donnelly. Inference o f population structure using multilocus genotype data. Genetics, 155:945959, 2000. (Cited on page 2.)

2000
[25]

Revuz and M

D. Revuz and M. Yor. Continuous Martingales and Brownian Motion , volume 293. Springer-Verlag, third edition, 1999. (Cited on pages 5 and 14.)

1999
[26]

H. Risken. The Fokker-Planck Equation . Springer, 1996. (Cited on page 5.)

1996
[27]

Rousseau

J. Rousseau. Rates of convergence for the posterior dis tributions of mixtures of Betas and adaptive nonparametric estimation of the density. Annals of Statistics , 38:146–180,
[28]

Rousseau and K

J. Rousseau and K. Mengersen. Asymptotic behaviour of t he posterior distribution in overﬁtted mixture models. Journal of the Royal Statistical Society: Series B (Statist ical Methodology), 73:689–710, 2011. (Cited on pages 7 and 12.) 35

2011
[29]

Schwartz

L. Schwartz. On Bayes procedures. Zeitschrift f˝ ur Wahrscheinlichkeitstheorie und Ver- wandte Gebiete , 4:10–26, 1965. (Cited on page 1.)

1965
[30]

W. Shen, S. R. Tokdar, , and S. Ghosal. Adaptive Bayesian multivariate density estima- tion with Dirichlet mixtures. Biometrika, 100:623640, 2013. (Cited on page 2.)

2013
[31]

Shen and L

X. Shen and L. Wasserman. Rates of convergence of poster ior distributions. Annals of Statistics, 29:687–714, 2001. (Cited on page 2.)

2001
[32]

A. W. van der Vaart. Asymptotic Statistics . Cambridge University Press, 1998. (Cited on page 20.)

1998
[33]

A. W. van der Vaart and J. Wellner. Weak Convergence and Empirical Processes . Springer-Verlag, New York, NY, 1996. (Cited on page 23.)

1996
[34]

A. W. van der Vaart and J. A. Wellner. Weak Convergence and Empirical Processes: With Applications to Statistics . Springer-Verlag, New York, NY, 2000. (Cited on page 25.)

2000
[35]

M. J. Wainwright. High-Dimensional Statistics: A Non-Asymptotic Viewpoint . Cam- bridge University Press, 2019. (Cited on pages 23, 24, and 27.)

2019
[36]

S. Walker. On suﬃcient conditions for Bayesian consist ency. Annals of Statistics , 90:482– 488, 2003. (Cited on page 1.)

2003
[37]

S. Walker. New approaches to Bayesian consistency. Annals of Statistics , 32:2028–2043,
[38]

S. G. Walker, A. Lijoi, and I. Prunster. On rates of conve rgence for posterior distributions in inﬁnite-dimensional models. Annals of Statistics , 35:738–746, 2007. (Cited on page 2.)

2007
[39]

Y. Wang. Convergence rates of latent topic models under relaxed identiﬁability conditions. Electronic Journal of Statistics , 13:37–66, 2019. (Cited on page 2.)

2019
[40]

Yang and D

Y. Yang and D. B. Dunson. Bayesian manifold regression. Annals of Statistics , 44:876– 905, 2016. (Cited on page 2.)

2016
[41]

Yang and S

Y. Yang and S. T. Tokdar. Minimax-optimal nonparametri c regression in high dimensions. Annals of Statistics , 43:652–674, 2015. (Cited on page 2.) 36

2015
[1989]

(Cited on pages 7 and 9.)

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Barron, M

A. Barron, M. Schervish, and L. Wasserman. The consisten cy of posterior distributions in nonparametric problems. Ann. Statist , 27:536–561, 1999. (Cited on page 1.)

work page 1999

[2] [2]

Bhattacharya, D

A. Bhattacharya, D. Pati, , and D. B. Dunson. Anisotropic function estimation using multi-bandwidth Gaussian processes. Annals of Statistics , 42:352–381, 2014. (Cited on page 2.)

work page 2014

[3] [3]

D. Blei, A. Ng, and M. Jordan. Latent Dirichlet allocatio n. J. Mach. Learn. Res , 3:993– 1022, 2003. (Cited on pages 2 and 20.)

work page 2003

[4] [4]

Boucheron, G

S. Boucheron, G. Lugosi, and P. Massart. Concentration Inequalities: A Nonasymptotic Theory of Independence . Oxford University Press, 2016. (Cited on pages 27 and 34.)

work page 2016

[5] [5]

R. J. Carroll and P. Hall. Optimal rates of convergence fo r deconvolving a density. Journal of American Statistical Association , 83:1184–1186, 1988. (Cited on page 10.)

work page 1988

[6] [6]

J. Chen. Optimal rate of convergence for ﬁnite mixture mo dels. Annals of Statistics , 23(1):221–233, 1995. (Cited on pages 3 and 14.)

work page 1995

[7] [7]

de Jonge and J

R. de Jonge and J. H. van Zanten. Adaptive nonparametric B ayesian inference using location-scale mixture priors. Annals of Statistics , 38:3300–3320, 2010. (Cited on page 2.)

work page 2010

[8] [8]

J. L. Doob. Application of the theory of martingales. Actes du ColloqueInternational Le Calcul des Probabilit´ es et ses applications (Lyon, 28 Juin 3 J uillet, 1948) , pages 23–27,

work page 1948