Pith. sign in

REVIEW 2 major objections 3 minor 47 references

Diffeomorphic Markov Chain Monte Carlo: fast mixing for heavy-tailed distributions

T0 review · 2 major / 3 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A radial contraction onto a ball turns heavy-tailed MCMC into uniformly ergodic sampling on a bounded domain, with $O(d^3)$ mixing for Student-$t$ targets.

desk verdict Solid uniform-ergodicity theory for heavy tails via a radial pull-back to a ball, but the fast-mixing claims for multimodal targets outrun what the implemented Möbius tuning can actually deliver; still deserves a serious referee. read the letter →

arxiv 2608.04284 v1 pith:BZOWFXCP submitted 2026-08-04 stat.CO math.PR

Miha Brešar , Aleksandar Mijatović This is my paper · ORCID
classification stat.COmath.PR MSC 60J1060J2065C05
keywords heavy-taileddistributionsuniformergodicitynon-asymptoticmixingtimeHit-and-RunBallWalkdiffeomorphiccontractionsamplervariationalinferencelog-concavesampling
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 introduces the Diffeomorphic Contraction Sampler (DCS), which pulls a target density on $\mathbb{R}^d$ back onto a Euclidean ball $B(R)$ through a radial diffeomorphism and then runs a standard sampler, such as Hit-and-Run or Ball Walk, on the bounded ball. Its central claim is that whenever the target has a finite polynomial moment in the sense $\limsup_{|x|\to\infty}\pi(x)|x|^{d+\beta}<\infty$ for some $\beta>0$, DCS is uniformly ergodic on $\mathbb{R}^d$ for several standard ball kernels. For Student-$t$ targets, a stronger set of assumptions on the pull-back density yields a cold-start mixing time of order $d^3$ for the Hit-and-Run variant, in contrast to the exponential-in-$d$ mixing of Gaussian Random Walk Metropolis. The paper also reports numerical evidence that DCS outperforms the No-U-Turn Sampler and spherical projection samplers on heavy-tailed and funnel-geometry benchmarks. The reason this matters is that heavy tails combined with high dimension make standard samplers mix slowly and can break central limit theorems for ergodic averages.

What carries the argument

The load-bearing object is the radial diffeomorphism $RC_\beta(y)=y/(1-|y|^\beta)^{1/\beta}$ from the unit ball to $\mathbb{R}^d$, whose Jacobian has a singularity of order $d+\beta$ at the boundary. Composed with a ball automorphism $T_\omega$ (in examples, a Möbius transformation) and parameters $(\mu,R)$, it defines $RC_{\mathrm{total}}(z)=\mu+R\,RC_\beta(T_\omega(z/R))$, and the pull-back density is $\pi_B(z)=\pi(RC_{\mathrm{total}}(z))\,J_{RC_{\mathrm{total}}}(z)$. The Jacobian singularity cancels polynomial tails, making $\pi_B$ bounded on $B(R)$; the automorphism preconditions skew and multimodal geometry. Total-variation distance is invariant under this diffeomorphism, so mixing on the ball transfers exactly to mixing on $\mathbb{R}^d$.

What would settle it

For a Student-$t$ target in $d$ dimensions with $v\ge2$, run DCS-HnR from a cold start with the paper's recommended $R^2=d+v$ and measure the total-variation mixing time $\tau_X(\varepsilon,\delta_0)$; if $\tau_X$ grows faster than a polynomial of degree $d^3$ in $d$, the central non-asymptotic claim fails. A second check is to evaluate the Hessian $\nabla^2 U_B$ of the pulled-back potential at points approaching $\partial B(R)$: a negative eigenvalue anywhere would violate Assumption (Conv) and place the example outside Theorem 3.6.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a class of samplers whose mixing behaviour is governed by the boundary singularity of a chosen radial contraction rather than by the tail decay of the target. Theorem 3.1 states that if $\pi$ is continuous and satisfies $\limsup_{|x|\to\infty}\pi(x)|x|^{d+\beta}<\infty$ for some $\beta>0$, then Algorithm (DCS) using the DCS-HnR, DCS-BW, DCS-RWM, or covariance-floored DCS-DW kernel is uniformly ergodic on $\mathbb{R}^d$. Under additional assumptions on the pulled-back density $\pi_B$ — convexity of $U_B=-\log\pi_B$, isotropy, and warm or cold start conditions — Theorem 3.6 gives non-asymptotic mixing-time bounds, and Example B.1 verifies these assumptions for Student-$t$ targets with $R^2=d+v$, yielding a mixing time of order $d^3$ for DCS-HnR from a cold start.

Load-bearing premise

For the advertised $O(d^3)$ cold-start guarantee, the pull-back density $\pi_B$ must be log-concave, near-isotropic, and satisfy cold-start level-set conditions; these are verified for Student-$t$ targets in the paper, but for general or multimodal targets a suitable preconditioning automorphism is only shown to exist, not constructed.

Editorial extensions

If this is right

  • Uniform ergodicity holds for every target satisfying (A-$\beta$), regardless of how slow the polynomial tail is, so the usual moment restrictions of stereographic samplers are removed.
  • For Student-$t$ targets, DCS-HnR mixes in $O(d^3)$ steps from a cold start, whereas Gaussian Random Walk Metropolis has an exponential-in-$d$ bound; this is the first polynomial cold-start bound of this type.
  • Any improvement in sampling algorithms on the Euclidean ball immediately becomes an improvement for DCS, because the pull-back reduces the problem to sampling on a ball.
  • The VI-tuned Möbius automorphism is meant to make the pull-back density near-uniform; when it succeeds, the non-asymptotic bounds of Theorem 3.6 apply even for skewed or multimodal targets.
  • Numerical experiments indicate the method reaches comparable accuracy to tuned NUTS or spherical samplers in a fraction of the wall-clock time on heavy-tailed posteriors and constrained GARCH-type boundaries.

Reading between the lines

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

  • Because the TV distance is unchanged by the diffeomorphism, any future sampler with tighter bounds on the ball would automatically transfer those bounds to the full heavy-tailed problem; the paper's use of existing ball samplers is not the only possible choice.
  • The cold-start guarantee depends on the automorphism $T_\omega$ being found; for multimodal targets the paper guarantees existence via triangular transport but not a constructive recipe, so an empirical check of whether VI finds such a map on a given target is a direct test of the $O(d^3)$ claim.
  • The tail-decay trichotomy in Remark 3.11 suggests an adaptive rule: estimate the tail index of $\pi$ and choose $\beta$ slightly smaller, so that (A-$\beta$) and asymptotic convexity hold; the paper notes this preference but does not propose an estimator.
  • One could extend DCS to gradient-based samplers by handling the boundary singularity of $\nabla U_B$; the paper lists this as an open problem, but its resolution would likely improve practical performance in regimes where random-walk proposals are diffusive.
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

2 major / 3 minor

Summary. The paper proposes the Diffeomorphic Contraction Sampler (DCS), which maps a target on R^d to a density on a Euclidean ball via a radially contracting diffeomorphism composed with an automorphism of the ball, runs a convex-body sampler (Ball Walk, Hit-and-Run, RWM, Dikin walk, or projected Langevin) on the pull-back density, and pushes the samples forward. Theorem 3.1 establishes uniform ergodicity for every continuous target satisfying limsup_{|x|→∞} π(x)|x|^{d+β} < ∞, for the HnR, BW, RWM, and covariance-floored Dikin kernels. Theorem 3.6 imports non-asymptotic mixing bounds under convexity/log-concavity, isotropy, and warm/cold start conditions on the pull-back density. Example B.1 verifies these conditions for Student-t targets and yields an O(d^3) cold-start mixing bound for DCS-HnR. Numerical experiments compare DCS favourably with NUTS and with stereographic samplers on skewed Student-t, PosteriorDB funnel-type, and infinite-moment targets.

Significance. The uniform ergodicity theorem is a clean and useful result: Lemma A.4 bounds the pull-back density under the moment condition, Lemma A.5 gives m-step minorization from an arbitrary starting point, and Theorem 3.1 follows by a standard minorization argument. The total-variation invariance observation (Proposition A.1, Corollary A.3) correctly reduces the analysis of the pushed-forward chain to sampling on the ball and is a strong structural simplification. The explicit Student-t verification of Assumptions (Conv), (Iso), and (Cold) in Example B.1, with the resulting O(d^3) cold-start Hit-and-Run bound, is a concrete and useful advance over the exponential-in-d RWM bound. The main weakness is that the fast-mixing theory for multimodal targets is existential rather than constructive, and the practical Möbius-tuned algorithm is not covered by Theorem 3.6 for multimodal or funnel targets; the abstract and contribution list should be reworded so that this scope limitation is explicit.

major comments (2)
  1. [§3.2, Example B.2, Remark B.2, §5(IIb)] The fast-mixing claim for multimodal targets is not supported by a constructive transport. Example B.2 invokes a Knothe–Rosenblatt diffeomorphism whose existence is guaranteed by [6] but for which neither an explicit form nor closed-ball boundary regularity is supplied; Remark B.2 and Open Problem (IIb) concede that only existence is known. The practical tuning procedure in Algorithm (VI) optimizes only (μ, R, δ) within the Möbius family (8), and a Möbius automorphism is a homeomorphism of B(1) that cannot make a genuinely multimodal pull-back density log-concave, because Assumption (Conv) forces log-concavity and hence unimodality. Therefore Theorem 3.6 does not cover the implemented VI-tuned algorithm on multimodal or funnel targets, and the numerical results in Section 4.2 are, as the text admits, outside the full-space theory. The abstract and contribution (II) should attribute the fast non-asymptotic guarantee specifically to the Student-t construction in Example B.1 and, conditionally, to target classes for which a suitable automorphism is actually available, rather than to general heavy-tailed or multimodal targets.
  2. [§3.2, Theorem 3.6(d)] The statement of the DCS-LMC bound is ill-typed: the chain X_N lives on R^d and targets ν, but the displayed bound compares L_x(X_N) with ν_B, which is a measure on B(R). The proof also refers to a chain 'started at 0', while the statement takes X_0 = RC_total(0). This should be corrected by writing either the bound for the transformed chain on B(R) or the push-forward bound against ν. The error is local and easy to fix, but it currently makes the formal statement of part (d) unreadable.
minor comments (3)
  1. [§3.2.1, Remark 3.11] Remark 3.11 states that choosing β too large results in 'failure of uniform ergodicity'; Theorem 3.1 gives only a sufficient condition for uniform ergodicity, so the remark should say that condition (A-β) fails rather than asserting non-uniform ergodicity.
  2. [Appendix C.2 and §4.2] The numerical experiments in Section 4.2 implement Hit-and-Run using a univariate slice-sampling transition, while Lemma A.5 and Theorem 3.1 concern the exact-conditional HnR kernel. Appendix C notes this, but the main text should explicitly state that the implemented slice-based kernel is not covered by the uniform-ergodicity or mixing theorems.
  3. [§2.3.1, Algorithm (VI)] The variational objective minimizes reverse KL with respect to the uniform measure on B(1); it does not directly promote Assumptions (Conv), (Iso), or (Cold). A sentence stating that the tuned Möbius family can satisfy these assumptions for unimodal approximately isotropic targets but not in general for multimodal targets would help align the practical tuning step with the theoretical assumptions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DCS convergence theorems are conditional on explicit assumptions and external convex-body sampling bounds; self-citations are not load-bearing.

full rationale

I walked the derivation chain and found no circular step. Theorem 3.1 is a conditional statement: under the explicit tail condition (A-beta) on the original target pi, and with the pull-back density pi_B bounded via Lemma A.4, the paper proves minorization for each listed ball kernel using only the kernel definitions and standard inequalities. The constants beta, R, mu, and omega are user-chosen or VI-tuned parameters; they are not fitted to the quantities that the theorems predict. Theorem 3.6 is likewise conditional on assumptions (Conv), (Iso), (Warm), (Cold), etc. on pi_B, and its bounds are imported from external log-concave/convex-body sampling results ([11], [28], [32], [33], [34]), not from the paper's own numerical fits. Example B.1 is a constructive verification for Student-t targets: it chooses R^2=d+v and beta=2 and then proves convexity, isotropy, and the cold-start conditions directly; this is a worked example, not a fitted-input-called-prediction. Example B.2 invokes the external Knothe-Rosenblatt transport theorem [6] to establish existence of a diffeomorphism making the pull-back log-concave, and Remark B.2 explicitly flags boundary regularity as a separate condition. The Möbius family in (8) is presented as a practical, tunable ansatz and is not claimed by the theory to make multimodal pull-backs log-concave; the gap between the existence-based theory and the VI-tuned numerical implementation is a scope/correctness concern, not circularity. The only same-author citations ([8], [9]) are used for motivational context and background on heavy-tailed MCMC, and neither carries the weight of the paper's central uniform-ergodicity or mixing-time claims. No equation in the paper reduces a predicted output to a fitted parameter by construction, so the appropriate finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central uniform-ergodicity result rests only on the tail condition and mild smoothness; the fast non-asymptotic results rest on a stack of geometric assumptions on the pull-back density and imported convex-body theorems. The paper is transparent about these assumptions, but the practical VI-tuning phase introduces parameters β, R, μ, and δ whose optimal values are not theoretically guaranteed.

free parameters (4)
  • β = β=2 in Student-t examples, β=0.2 in super-Cauchy example
    Radial contraction exponent chosen by the user; must be smaller than the tail decay rate to preserve uniform ergodicity. Not fitted to data but chosen by hand.
  • R (ball radius) = R^2=d+v in Example B.1; otherwise tuned via VI
    Radius of the sampling ball. In the theory it is arbitrary; in Example B.1 it is chosen to make the transformed density isotropic, and in numerics it is optimized via variational inference.
  • μ (location parameter) = Tuned via VI in Section 4; μ=0 in symmetric examples
    Re-centering parameter in (RC_total). Chosen by hand or via VI, not derived from the theory.
  • δ (Möbius automorphism parameter) = Tuned via VI with Adam in Section 4.3.1; δ=0 (identity) in symmetric examples
    Controls the automorphism of the ball. Its choice affects whether the pull-back density is log-concave and isotropic, but no constructive guarantee is given that VI finds a suitable δ.
assumptions (6)
  • domain assumption Tail condition (A-β): limsup_{|x|→∞} π(x)|x|^{d+β} < ∞
    Used in Theorem 3.1 and Lemma A.4. It is the main condition on the target distribution for uniform ergodicity.
  • domain assumption T_ω is a C^1 diffeomorphism extending to the closed ball B(1)
    Invoked in Section 2.2 and Lemma A.4 to ensure the Jacobian J_{T_ω} is bounded away from zero and infinity at the boundary, preserving the tail behaviour.
  • domain assumption π is continuous and strictly positive on R^d
    Used in Lemma A.4(b) to obtain lower bounds on π_B over compact subsets of B(R).
  • standard math Existing non-asymptotic mixing theorems for log-concave densities on convex bodies
    Theorem 3.6 imports bounds from [11], [24], [28], [32], and [33]. The paper assumes these external results are correct.
  • standard math Knothe-Rosenblatt triangular transport theorem and Dacorogna-Moser prescribed Jacobian theorem
    Example B.2 and Remark B.2 use these theorems to assert existence of an automorphism T_ω that makes a multimodal Student-t mixture transform to a log-concave density on the ball. The map is non-constructive.
  • domain assumption Geometric assumptions (Conv), (Smooth), (Warm), (Iso), (Lip), (Cold) on the pull-back density π_B
    These are the standing assumptions for the non-asymptotic bounds in Theorem 3.6. The paper verifies them for Student-t targets in Example B.1, but they do not hold automatically for arbitrary heavy-tailed targets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Diffeomorphic Markov Chain Monte Carlo: fast mixing for heavy-tailed distributions." pith.science (2026). https://pith.science/paper/BZOWFXCP

@misc{pith2026260804284,
  author       = {Pith},
  title        = {Pith review of: Diffeomorphic Markov Chain Monte Carlo: fast mixing for heavy-tailed distributions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BZOWFXCP}},
  note         = {Machine review of arXiv:2608.04284}
}
abstract

We introduce a new class of uniformly ergodic MCMC algorithms, termed Diffeomorphic Contraction Sampler (DCS), and provide fast non-asymptotic mixing guarantees for DCS targeting distributions on $\R^d$ with arbitrarily heavy polynomial tails. DCS provides a solution to a well-known problem for MCMC samplers, which typically struggle with the combination of unbounded high-dimensional state space and vanishing gradients. The DCS pulls back a target on $\R^d$ onto a Euclidean ball $B(R)\subset\R^d$ and then samples from the transformed density on the convex set $B(R)$ via algorithms such as the Ball Walk, Hit-and-Run and others. A radial diffeomorphic contraction is chosen so that the pull-back density on $B(R)$ is bounded, implying uniform ergodicity for \textit{all} targets with a finite polynomial moment. Non-asymptotic bounds for DCS require stronger assumptions such as log-concavity of the pull-back density. In practice, this is achieved approximately by a preconditioned automorphism of the ball $B(R)$, tuned via Variational Inference. Numerical simulation tests demonstrate that the DCS outperforms significantly the No-U-Turns sampler on multi-dimensional heavy-tailed targets arising as real-world posteriors in PosteriorDB benchmark. DCS also numerically outperforms in high-dimensional examples recently developed spherical projection samplers for heavy-tailed target distributions.

Figures

Figures reproduced from arXiv: 2608.04284 by the authors.

Figure 1
Figure 1. Boxplots of ergodic averages across 50 independent chains for a 200- dimensional skewed Student-t target with v = 3. The left (resp. right) panel shows the estimated tail probability E[I{|X| ≥ 50.0}] (resp. expected norm E[|X|]). Red dashed lines indicate the exact analytical ground-truth values of 0.0293 and 19.5234, respectively. The results compare the optimised Algorithm (DCS-BW) with β = 2 (in (RCtotal) and (6)… view at source ↗
Figure 2
Figure 2. Comparison of performance and mean squared error across 50 independ￾ent chains (10 seconds of CPU time each) for the sampling from the heavy-tailed target (14) (a 200-dimensional skewed Student-t target with v = 3). NUTS clearly outperforms both RWM algorithms suggesting that, in spite of its dependence on the gradient (which in this case vanishes in the tails), it remains the state-of-the￾art among base-line algori… view at source ↗
Figure 3
Figure 3. Comparison of average runtimes and mean squared errors across 20 independent chains (500,000 iterations each) for the sampling from the heavy￾tailed skewed Student-t target (14) (100-dimensional skewed Student-t target with v = 3 and skewness parameter α ∈ R d with the first component equal to 100, the second one to −100 and the rest equal to zero). Samplers (DCS-RWM) and (DCS￾BW) achieve lower mean squared errors f… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Average number of generated iterations and Mean Squared Error com￾parison across 50 independent chains (10.0 seconds time budget per chain). The properly tuned projected samplers (β = 0.2) drastically outperform vanilla SCS in both estimation accuracy and computational…
Figure 5
Figure 5. Figure 5: Q-Q plot of β1 in robust logistic regression obtained from samples of algorithms (DCS-BW) (left column) and SCS [19] (right column). The prior is the Student-t distribution with degrees of freedom equal to 4 (top), and 6 (bottom). Shaded regions show variation across 2…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 44 canonical work pages

  1. [6]

    3, 309–335

    Vladimir Igorevich Bogachev, Aleksandr Viktorovich Kolesnikov, and Kirill Vladimirovich Medvedev,Triangular transformations of measures, Sbornik: Mathematics196(2005), no. 3, 309–335

    work page 2005

  2. [1]

    Wang,Weak Poincar´ e inequalities for Markov chains: theory and applications, Ann

    Christophe Andrieu, Anthony Lee, Sam Power, and Andi Q. Wang,Weak Poincar´ e inequalities for Markov chains: theory and applications, Ann. Appl. Probab.36(2026), no. 1, 46–107. MR 5033571

    work page 2026

  3. [2]

    2, 367–389

    Adelchi Azzalini and Antonella Capitanio,Distributions generated by perturbation of symmetry with emphasis on a multivariate skew t-distribution, Journal of the Royal Statistical Society Series B: Statistical Methodology65 (2003), no. 2, 367–389

    work page 2003

  4. [3]

    Necdet Batır,Bounds for the gamma function., Results in Mathematics/Resultate der Mathematik72(2017)

    work page 2017

  5. [4]

    Cameron Bell, Krzystof Latuszy´ nski, and Gareth O Roberts,Adaptive stereographic mcmc, arXiv preprint arXiv:2408.11780 (2024)

    arXiv 2024

  6. [5]

    Blei, Alp Kucukelbir, and Jon D

    David M. Blei, Alp Kucukelbir, and Jon D. McAuliffe,Variational inference: a review for statisticians, J. Amer. Statist. Assoc.112(2017), no. 518, 859–877. MR 3671776

    work page 2017

  7. [7]

    James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Yash Katariya, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang,JAX: composable transformations of Python+NumPy programs, 2018

    work page 2018

  8. [8]

    Miha Breˇ sar, Aleksandar Mijatovi´ c, and Gareth Roberts,Central limit theorem for ergodic averages of Markov chains & the comparison of sampling algorithms for heavy-tailed distributions, arXiv preprint arXiv:2512.18255 (2025)

    work page arXiv 2025

Show all 47 references
  1. [9]

    Diffeomorphic Contraction Sampler: fast mixing for heavy-tailed distributions

    Miha Breˇ sar and Aleksandar Mijatovi´ c,YouTube video “Diffeomorphic Contraction Sampler: fast mixing for heavy-tailed distributions”, DCS Algorithm (theorical convergence guarantees and numerical examples) on the YouTube channel Prob-AM, 2026

    2026

  2. [10]

    Nicolas Brosse, Alain Durmus, ´Eric Moulines, and Marcelo Pereyra,Sampling from a log-concave distribution with compact support with proximal Langevin Monte Carlo, Conference on learning theory, PMLR, 2017, pp. 319–342

    2017

  3. [11]

    Geom.59(2018), no

    S´ ebastien Bubeck, Ronen Eldan, and Joseph Lehec,Sampling from a log-concave distribution with projected Langevin Monte Carlo, Discrete Comput. Geom.59(2018), no. 4, 757–783. MR 3802303 22 MIHA BRE ˇSAR AND ALEKSANDAR MIJATOVI ´C

    2018

  4. [12]

    Geom.75(2026), no

    Yuansi Chen and Ronen Eldan,Hit-and-run mixing via localization schemes, Discrete Comput. Geom.75(2026), no. 3, 747–794. MR 5009943

    2026

  5. [13]

    Sinho Chewi,Log-concave sampling, Book draft available at https://chewisinho. github. io9(2025), 17–18

    2025

  6. [14]

    Bernard Dacorogna and J¨ urgen Moser,On a partial differential equation involving the jacobian determinant, Annales de l’Institut Henri Poincar´ e C, Analyse non lin´ eaire7(1990), no. 1, 1–26

    1990

  7. [15]

    Statist.47(2019), no

    George Deligiannidis, Alexandre Bouchard-Cˆ ot´ e, and Arnaud Doucet,Exponential ergodicity of the bouncy particle sampler, Ann. Statist.47(2019), no. 3, 1268–1287. MR 3911112

    2019

  8. [16]

    George Deligiannidis, Daniel Paulin, Alexandre Bouchard-Cˆ ot´ e, and Arnaud Doucet,Randomized Hamiltonian Monte Carlo as scaling limit of the bouncy particle sampler and dimension-free convergence rates, Ann. Appl. Probab.31(2021), no. 6, 2612–2662. MR 4350970

    2021

  9. [17]

    Gersende Fort and Eric Moulines,V-subgeometric ergodicity for a Hastings-Metropolis algorithm, Statist. Probab. Lett.49(2000), no. 4, 401–410. MR 1796485

    2000

  10. [18]

    6, 721–741

    Stuart Geman and Donald Geman,Stochastic relaxation, gibbs distributions, and the bayesian restoration of images, IEEE Transactions on pattern analysis and machine intelligence (1984), no. 6, 721–741

    1984

  11. [19]

    Sebastiano Grazzi, Sifan Liu, Gareth O Roberts, and Jun Yang,Sub-cauchy sampling: Escaping the dark side of the moon, arXiv preprint arXiv:2601.11066 (2026)

    2026

  12. [20]

    2, 223–242

    Heikki Haario, Eero Saksman, and Johanna Tamminen,An adaptive metropolis algorithm, Bernoulli7(2001), no. 2, 223–242. MR 1828504

    2001

  13. [21]

    Matthew Hoffman, Pavel Sountsov, Joshua V Dillon, Ian Langmore, Dustin Tran, and Srinivas Vasudevan, Neutra-lizing bad geometry in hamiltonian monte carlo using neural transport, arXiv preprint arXiv:1903.03704 (2019)

    2019 arXiv

  14. [22]

    Hoffman and Andrew Gelman,The no-U-turn sampler: adaptively setting path lengths in Hamiltonian Monte Carlo, J

    Matthew D. Hoffman and Andrew Gelman,The no-U-turn sampler: adaptively setting path lengths in Hamiltonian Monte Carlo, J. Mach. Learn. Res.15(2014), 1593–1623. MR 3214779

    2014

  15. [23]

    Jarner and Gareth O

    Søren F. Jarner and Gareth O. Roberts,Convergence of heavy-tailed Monte Carlo Markov chain algorithms, Scand. J. Statist.34(2007), no. 4, 781–815. MR 2396939

    2007

  16. [24]

    3017–3078

    Minhui Jiang and Yuansi Chen,Regularized dikin walks for sampling truncated logconcave measures, mixed iso- perimetry and beyond worst-case analysis, The Thirty Eighth Annual Conference on Learning Theory, PMLR, 2025, pp. 3017–3078

    2025

  17. [25]

    Johnson and Charles J

    Leif T. Johnson and Charles J. Geyer,Variable transformation to obtain geometric ergodicity in the random-walk Metropolis algorithm, Ann. Statist.40(2012), no. 6, 3050–3076. MR 3097969

    2012

  18. [26]

    Ravi Kannan, L´ aszl´ o Lov´ asz, and Mikl´ os Simonovits,Random walks and anO∗(n5)volume algorithm for convex bodies, Random Structures Algorithms11(1997), no. 1, 1–50. MR 1608200

    1997

  19. [27]

    Kingma and Jimmy Ba,Adam: A method for stochastic optimization, International Conference on Learning Representations (ICLR), 2015

    Diederik P. Kingma and Jimmy Ba,Adam: A method for stochastic optimization, International Conference on Learning Representations (ICLR), 2015

    2015

  20. [28]

    3137–3240

    Yunbum Kook and Santosh S Vempala,Gaussian cooling and dikin walks: The interior-point method for logcon- cave sampling, The Thirty Seventh Annual Conference on Learning Theory, PMLR, 2024, pp. 3137–3240

    2024

  21. [29]

    Yunbum Kook, Santosh S Vempala, and Matthew S Zhang,In-and-out: Algorithmic diffusion for sampling convex bodies, Advances in Neural Information Processing Systems37(2024), 108354–108388

    2024

  22. [30]

    3, 1043–1092

    Yin Tat Lee and Santosh S Vempala,Eldan ’s stochastic localization and the kls conjecture: Isoperimetry, concen- tration and mixing, Annals of Mathematics199(2024), no. 3, 1043–1092

    2024

  23. [31]

    4A, 3109–3138

    Samuel Livingstone, Michael Betancourt, Simon Byrne, and Mark Girolami,On the geometric ergodicity of Hamiltonian Monte Carlo, Bernoulli25(2019), no. 4A, 3109–3138

    2019

  24. [32]

    Comput.35(2006), no

    L´ aszl´ o Lov´ asz and Santosh Vempala,Hit-and-run from a corner, SIAM J. Comput.35(2006), no. 4, 985–1005. MR 2203735

    2006

  25. [33]

    3, 307–358

    ,The geometry of logconcave functions and sampling algorithms, Random Structures Algorithms30(2007), no. 3, 307–358. MR 2309621

    2007

  26. [34]

    L´ aszl´ o Lov´ asz and Santosh S. Vempala,Fast algorithms for logconcave functions: Sampling, rounding, integration and optimization, Proceedings of the 47th Annual IEEE Symposium on Foundations of Computer Science, 2006, pp. 57–68. DIFFEOMORPHIC MONTE CARLO 23

    2006

  27. [35]

    2, 132–158

    Qiu-Ming Luo and Feng Qi,Bounds for the ratio of two gamma functions—from wendel’s and related inequalities to logarithmically completely monotonic functions, Banach Journal of Mathematical Analysis6(2012), no. 2, 132–158

    2012

  28. [36]

    M ˚ ans Magnusson, Paul-Christian B¨ urkner, and Aki Vehtari,PosteriorDB: a database of bayesian posterior in- ference, 2021

    2021

  29. [37]

    M ˚ ans Magnusson, Jakob Torgander, Paul-Christian B¨ urkner, Lu Zhang, Bob Carpenter, and Aki Vehtari,Pos- teriorDB: testing, benchmarking and developing Bayesian inference algorithms, arXiv preprint arXiv:2407.04967 (2024)

    2024 arXiv

  30. [38]

    K. L. Mengersen and R. L. Tweedie,Rates of convergence of the Hastings and Metropolis algorithms, Ann. Statist. 24(1996), no. 1, 101–121. MR 1389882

    1996

  31. [39]

    3, 705–767

    Radford M Neal,Slice sampling, The Annals of Statistics31(2003), no. 3, 705–767

    2003

  32. [40]

    2, 645–682

    Matthew D Parno and Youssef M Marzouk,Transport map accelerated Markov chain monte carlo, SIAM/ASA Journal on Uncertainty Quantification6(2018), no. 2, 645–682

    2018

  33. [41]

    Du Phan, Neeraj Pradhan, and Martin Jankowiak,Composable effects for flexible and accelerated probabilistic programming in numpyro, Program Transformations for ML Workshop at NeurIPS 2019, 2019

    2019

  34. [42]

    8083–8092

    Danilo Jimenez Rezende, George Papamakarios, S´ ebastien Racaniere, Michael Albergo, Gurtej Kanwar, Phiala Shanahan, and Kyle Cranmer,Normalizing flows on tori and spheres, International Conference on Machine Learning, PMLR, 2020, pp. 8083–8092

    2020

  35. [43]

    1, 95–110

    Gareth O Roberts and Richard L Tweedie,Geometric convergence and central limit theorems for multidimensional hastings and metropolis algorithms, Biometrika83(1996), no. 1, 95–110

    1996

  36. [44]

    GO Roberts, A Gelman, and WR Gilks,Weak convergence and optimal scaling of random walk metropolis al- gorithms, The Annals of Applied Probability (1997), 110–120

    1997

  37. [45]

    H. H. Rosenbrock,An automatic method for finding the greatest or least value of a function, Comput. J.3 (1960/61), 175–184. MR 136042

    1960

  38. [46]

    Pedro Teixeira,Dacorogna–moser theorem on the jacobian determinant equation with control of support, Discrete and Continuous Dynamical Systems37(2017), 4071–4089

    2017

  39. [47]

    Roberts,Stereographic Markov chain Monte Carlo, Ann

    Jun Yang, Krzysztof Latuszy´ nski, and Gareth O. Roberts,Stereographic Markov chain Monte Carlo, Ann. Statist. 52(2024), no. 6, 2692–2713. MR 4842823 AppendixA.Proofs: convergence theory for Algorithm (DCS) Recall the definition of the total variation distance of two probabili...

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.