REVIEW 3 major objections 6 minor 40 references
From Minimax Optimal Importance Sampling to Uniformly Ergodic Importance-tempered MCMC
T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read The paper proves that the target is the minimax importance-sampling proposal unless an atom exceeds 1/2, and that importance-tempered MCMC is uniformly ergodic for polynomial-tail targets exactly when 1/γ<β<(γ−2)/γ.
desk verdict Two solid contributions, but the 'only if' direction of the main ergodicity theorem rests on a fixable Fatou gap. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
Two objects carry the argument. The first is the worst-case risk functional $R(Q,L^2(\Pi))=\sup_f \Pi(f^2w)$ over functions with $\Pi(f)=0$ and $\Pi(f^2)=1$, where $w=\pi/q$ is the importance weight; on atomless spaces this reduces to the essential supremum of $w$, while on finite spaces it is either the largest weight or the unique root of the equation $\sum_i \pi_i/(w_i-\lambda)=0$, and minimizing it yields the 'cut the heavy atom to $1/2$' rule. The second is the continuous-time chain $(Y_t)$ whose embedded chain is the tempered Metropolis--Hastings chain and whose holding time at $x$ is exponential with mean $w(x)\propto \pi(x)^{1-\beta}$; its generator is $(Ag)(x)=w(x)^{-1}\int[g(y)-g(x)]T(x,dy)$. Uniform ergodicity is proved through drift conditions $(AV)(x)\le -\alpha V(x)$ outside a compact set, with $V$ bounded, nondecreasing, and strictly concave on each tail: Theorem 3 shows the drift rate is controlled by $|V''(x)|/w(x)$, so the importance weight drives the chain out of the tails. The necessity of $\beta<(\gamma-2)/\gamma$ is proved with the unbounded drift $V(x)=\log(1+|x|)$, showing that no bounded drift can work at or beyond the threshold.
What would settle it
Run the continuous-time chain $(Y_t)$ for the target $\pi(x)\propto(1+|x|)^{-5}$ with a random-walk proposal and $\beta=0.55$ versus $\beta=0.65$; the theorem says the first is uniformly ergodic and the second is not, so the total-variation distance from stationarity should decay at a rate independent of $Y_0$ for $\beta=0.55$ but not for $\beta=0.65$. For the independent-sampling half, compute the worst-case asymptotic variance over zero-mean unit-variance $f$ when one atom has mass $0.8$: the predicted minimum, $4(0.8)(0.2)=0.64$, is attained exactly by the trial with atom mass $1/2$.
Extended reading notes
Core claim
The paper establishes two characterizations. First, for independent sampling, the worst-case asymptotic variance of the self-normalized importance sampling estimator over all zero-mean, unit-variance functions is minimized by using the target as the trial distribution if and only if no atom carries probability greater than $1/2$; when an atom has mass $p>1/2$, the minimax trial puts probability $1/2$ on that atom and rescales the target density on the remaining space, lowering the worst-case risk to $4p(1-p)$. An analogous near-optimal construction handles continuous targets concentrated on a small set. Second, for importance-tempered MCMC, the paper shows that the continuous-time chain built from a random-walk Metropolis--Hastings chain with stationary density $\pi(x)^\beta$ is uniformly ergodic for $\pi(x)\propto(1+|x|)^{-\gamma}$ if and only if $1/\gamma<\beta<(\gamma-2)/\gamma$. Such a $\beta$ exists exactly when $\gamma>3$, and uniform ergodicity means the chain forgets its starting point at a rate independent of the start, so time averages can be computed without discarding burn-in.
Load-bearing premise
The load-bearing premise is that uniform ergodicity of the continuous-time chain, whose holding times are random importance weights, reliably describes the discrete self-normalized estimator used in practice; the paper treats the continuous-time convergence rate as a proxy and concedes that with a poor initialization the discrete estimator's variance can still decay slowly in the number of jumps.
Editorial extensions
If this is right
- For any target, a practitioner who cares about worst-case error over all square-integrable functions can safely use the target itself as the trial unless an atom exceeds $1/2$; for a heavy atom of mass $p$, the minimax proposal is explicit: put probability $1/2$ on that atom and renormalize the target density elsewhere.
- For a continuous posterior concentrated in a set $A$ with $\Pi(A)$ close to one, a trial density that puts roughly half its mass on $A$ has worst-case asymptotic variance of order $1-\Pi(A)$ for functions with small oscillation on $A$, compared with unit variance for direct sampling.
- For one-dimensional polynomial-tail targets with $\gamma>3$, choosing $\beta\in(1/\gamma,(\gamma-2)/\gamma)$ makes the continuous-time importance-tempered chain uniformly ergodic, so its long-run time averages are insensitive to the initial state and burn-in can be discarded.
- Within that window the importance weight $\pi(x)^{1-\beta}$ makes the generator's drift rate large in the tails, so the continuous-time chain spends bounded expected time outside any fixed central interval and the whole trajectory can be used.
- Mixing two self-normalized estimators built from different trial distributions cannot beat the minimax trial's worst-case risk, so the explicit optimal proposal is not improved by averaging trial distributions.
Reading between the lines
- The same minimax logic suggests a practical recipe for Bayesian posteriors: estimate a high-probability credible set $A$ and use a trial density that puts about half its mass on $A$, reweighting the complement; the paper proves near-optimality for functions controlled on $A$, but choosing the set and the exact split is left to the user.
- The sharp window $1/\gamma<\beta<(\gamma-2)/\gamma$ suggests a testable multivariate extension for spherically symmetric heavy-tailed targets, where the upper threshold should shift with dimension; the paper does not treat $d>1$.
- Uniform ergodicity of $Y_t$ does not guarantee fast variance decay for the discrete estimator from a cold start, since the number of jumps before entering the central region can still be unbounded; a safe implementation would monitor accumulated importance weight early in the run.
- For polynomial tails with $\gamma>4$, the value $\beta=1/2$ lies inside the ergodicity window, so the theory points to a simple default temperature in that regime; the paper does not draw this recommendation.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies two related problems. First, for self-normalized importance sampling, it proves a minimax characterization: the target distribution is the minimax optimal trial distribution if and only if no atom carries more than half the total mass; when a large atom exists, the optimal trial downweights that atom to probability 1/2. The paper also gives a continuous-space version for targets concentrated on a small set, and a negative result for multiple importance sampling. Second, it analyzes 'importance-tempered' MCMC, where the chain is run with stationary density proportional to pi(x)^beta and the bias is corrected by importance weights. The continuous-time version of the chain is shown, under a drift condition with bounded concave Lyapunov functions, to be uniformly ergodic for super-exponential targets (any beta in (0,1)) and for polynomial-tailed targets pi(x) proportional to (1+|x|)^{-gamma} when 1/gamma < beta < (gamma-2)/gamma. The main iff result is Theorem 4, which claims necessity of the upper bound beta < (gamma-2)/gamma. Numerical experiments illustrate the absence of burn-in and variance reduction.
Significance. If fully established, the results would be a valuable contribution to both importance sampling theory and MCMC convergence analysis. The minimax characterization in Theorems 1-2 is clean and apparently novel, and the large-atom result gives a concrete, non-obvious prescription for trial distributions. The continuous-time drift analysis in Theorem 3 is a useful technique: using bounded concave Lyapunov functions to exploit the importance-weight factor in the generator is an original idea, and Proposition 4's sufficient condition for uniform ergodicity of heavy-tailed targets is a strong result. The paper is also well structured: the main derivations are from stated assumptions with no fitted constants, and the numerical studies are consistent with the theory. However, the necessity proof of Theorem 4 has a load-bearing gap (an unjustified Fatou/limsup step), and the proof of Theorem 1 has a technical gap for countably infinite atom spaces. Both are plausibly repairable, but the manuscript is not yet fully rigorous as written.
major comments (3)
- [Section 3.3, proof of Theorem 4, after Eq. (29)] The proof uses the step limsup_t E_x[V(X_{t∧τ_D})] ≤ E_x[limsup_t V(X_{t∧τ_D})] = E_x[V(X_{τ_D})], citing Fatou's lemma. For nonnegative random variables, Fatou's lemma gives E[liminf] ≤ liminf E, not the limsup inequality used here. The limsup version requires uniform integrability or a domination argument, and none is supplied. In particular, the regime τ_D>t contributes E_x[V(X_t); τ_D>t], which need not vanish as t→∞ without additional control. As written, the argument only yields a lower bound on liminf E_x[V(X_{t∧τ_D})], which is compatible with the desired contradiction and does not establish log(1+D) ≥ E_x[V(X_{τ_D})]. Since this is the sole support for the 'only if' direction of Theorem 4, it is a load-bearing correctness gap. A possible repair is to use uniform ergodicity to obtain an exponentially decaying bound on P_x(τ_D>t) and the bounded-jump structure to control V(X_t) on the survival event, or to replace the necessity argument with a direct drift-based proof.
- [Section 2.2, proof of Theorem 1] The proof asserts that, when the essential supremum of w on the non-atomic part is below 1, there exists an atom x* with w(x*) = ess sup_{x∈X} w(x) ≥ 1. For a countably infinite set of atoms this is not guaranteed: the weights on atoms can approach a supremum without attaining it (e.g., w(a_n) = 2 − 1/n). The subsequent argument, including the inequality Π|_{E^c}(w) ≤ w(x*) and the decomposition before Eq. (7), requires that w(x*) be the global maximum of w. The gap is fixable by taking a sequence of atoms with weights approaching the essential supremum and passing to the limit, but as written the proof of Theorem 1 is incomplete. Since Theorem 1 is a central result of Section 2, this requires repair; note also that the proof of Proposition 1 in Section 2.3 repeats the same attainment argument and inherits the issue.
- [Section 3.3 and Remark 3] The ergodicity theorems (Theorem 4 and Propositions 3-4) are proved only for proposal densities with compact support, specifically the truncated normal in Eq. (25) and Condition (ii) of Theorem 3. The introduction and abstract, however, present the result for 'the Metropolis--Hastings algorithm' generally, and the numerical experiments in Section 4 use an untruncated N(0, 3^2) proposal. Remark 3 states that the truncation is a technical convenience and that the untruncated case can be handled with additional constraints on D, but no proof or precise statement is provided. This is a scope gap between the theoretical claims and the motivating application. I ask the authors to either supply the extension to unbounded symmetric proposals or explicitly state that the theorem is for bounded-support proposals and adjust the abstract and claims accordingly.
minor comments (6)
- [Section 2.3, paragraph before Eq. (11)] The definitions 'wX = π/qY and wY = π/qY' should be 'w_X = π/q_X' and 'w_Y = π/q_Y'.
- [Section 3.3, paragraph before Proposition 3] 'In Proposition 2, we allow polynomially decaying tails' should refer to Proposition 4, not Proposition 2.
- [Lemma 4 statement] The drift condition is stated for '∀ x∈(−∞,D]∪[D,∞)', which is all of R; the intended statement is clearly '∀ x∈(−∞,−D]∪[D,∞)'.
- [Remark 5] The citation for the stereographic projection sampler appears to be incorrect: reference [38] is Yang, Wainwright and Jordan (2016), while the stereographic sampler is discussed in [37] (Yang, Latuszyński and Roberts, 2024). Please clarify the citation.
- [Proof of Theorem 4, application of Meyn--Tweedie Theorem 16.2.2] The argument that uniform ergodicity of the T-skeleton chain implies sup_x E_x[τ_D] < ∞ for the continuous-time chain is plausible but needs a brief justification: the skeleton chain may jump over the compact set [-D,D] between skeleton times, so one should consider hitting a slightly enlarged set such as [-D-ξ, D+ξ].
- [Abstract and Remark 5] The abstract's claim that importance tempering can 'essentially eliminate the need for burn-in' should be read in light of Remark 5, which correctly notes that uniform ergodicity of the continuous-time chain Y_t does not directly imply a uniform finite-sample bound for the discrete estimator eΠ_{β,n}(f) when initialization is poor. Consider adding a qualifier in the abstract to avoid overstating the practical conclusion.
Circularity Check
No significant circularity: the minimax and ergodicity theorems are proven from stated conditions; the only self-citation is a peripheral variance-comparison used for motivation.
full rationale
The paper's central claims are self-contained. The minimax results (Theorems 1 and 2, Proposition 2) are derived by computing the worst-case asymptotic variance R(Q,L2(Pi)) directly from definitions, by exhibiting a universal lower bound via Lemma 3, and by verifying that the constructed trial distribution attains that bound. No fitted constant, data-dependent tuning, or assumed conclusion enters the derivation. The ergodicity results (Theorem 3, Propositions 3 and 4, Theorem 4) are established from explicit drift conditions on the generator of the continuous-time chain, with the drift function V chosen from the stated tail assumptions; the sufficiency and necessity directions are argued internally rather than imported from the author's own prior work. The only self-citation is [40, Lemma 2], used in Section 3.1 to assert that the asymptotic variance of the continuous-time estimator is at least that of the discrete self-normalized estimator and hence that convergence of Y_t can serve as a proxy for the efficiency of ePi_{beta,n}. That comparison is peripheral to the proofs of the main theorems and is explicitly hedged in Remark 5, which concedes that with poor initialization the variance of ePi_{beta,n} may still decay slowly in n. Thus no derivation step reduces by construction: no quantity is defined in terms of the quantity it is used to predict, no fitted parameter is relabeled as a prediction, and no load-bearing claim relies solely on the author's own prior result. A possible technical gap in the Fatou step of Theorem 4 would be a correctness concern, not a circularity concern.
Assumptions & free parameters
assumptions (6)
- standard math Atomless probability spaces admit measurable subsets of any prescribed probability (Sierpinski's theorem).
- standard math Drift-and-minorization criteria for continuous-time Markov chains (Down, Meyn, and Tweedie): a drift condition (AV) <= -alpha V with bounded V outside a petite compact set implies uniform ergodicity and exponential hitting-time moments.
- standard math Uniform ergodicity of the T-skeleton chain implies bounded expected hitting times of compact sets (Meyn and Tweedie Theorem 16.2.2).
- domain assumption The continuous-time chain Y_t with exponential holding times of mean w(x) is a faithful proxy for the discrete self-normalized estimator ePi_{beta,n}.
- domain assumption Proposal density kappa is symmetric, non-increasing on [0,infinity), and supported on [-xi,xi] (truncated normal in Propositions 3 and 4 and Theorem 4).
- domain assumption Target pi is one-dimensional, strictly positive, with non-increasing tails.
Cite this review
Pith. "Pith review of From Minimax Optimal Importance Sampling to Uniformly Ergodic Importance-tempered MCMC." pith.science (2026). https://pith.science/paper/RDAEXWTM
@misc{pith2026250619186,
author = {Pith},
title = {Pith review of: From Minimax Optimal Importance Sampling to Uniformly Ergodic Importance-tempered MCMC},
year = {2026},
howpublished = {\url{https://pith.science/paper/RDAEXWTM}},
note = {Machine review of arXiv:2506.19186}
}
abstract
We make two closely related theoretical contributions to the use of importance sampling schemes. First, for independent sampling, we prove that the minimax optimal trial distribution coincides with the target if and only if the target distribution has no atom with probability greater than $1/2$, where "minimax" means that the worst-case asymptotic variance of the self-normalized importance sampling estimator is minimized. When a large atom exists, it should be downweighted by the trial distribution. A similar phenomenon holds for a continuous target distribution concentrated on a small set. Second, we argue that it is often advantageous to run the Metropolis--Hastings algorithm with a tempered stationary distribution, $\pi(x)^\beta$, and correct for the bias by importance weighting. The dynamics of this "importance-tempered" sampling scheme can be described by a continuous-time Markov chain. We prove that for one-dimensional targets with polynomial tails, $\pi(x) \propto (1 + |x|)^{-\gamma}$, this chain is uniformly ergodic if and only if $1/\gamma < \beta < (\gamma - 2)/\gamma$. These results suggest that for target distributions with light or polynomial tails of order $\gamma > 3$, importance tempering can improve the precision of time-average estimators and essentially eliminate the need for burn-in.
Figures
Reference graph
Works this paper leans on
-
[1]
The pseudo-marginal approach for efficient Monte Carlo computations
Christophe Andrieu and Gareth O Roberts. The pseudo-marginal approach for efficient Monte Carlo computations. The Annals of Statistics , 37(2):697–725, 2009
work page 2009
-
[2]
Explicit constraints on the geometric rate of convergence of random walk Metropolis-Hastings
Riddhiman Bhattacharya and Galin L Jones. Explicit constraints on the geometric rate of convergence of random walk Metropolis-Hastings. Bernoulli, 31(3):2042–2076, 2025. 28
work page 2025
-
[3]
Adaptive importance sampling: The past, the present, and the future
Monica F Bugallo, Victor Elvira, Luca Martino, David Luengo, Joaquin Miguez, and Petar M Djuric. Adaptive importance sampling: The past, the present, and the future. IEEE Signal Processing Magazine , 34(4):60–79, 2017
work page 2017
-
[4]
Computational approaches for empirical Bayes methods and Bayesian sensitivity analysis
Eugenia Buta and Hani Doss. Computational approaches for empirical Bayes methods and Bayesian sensitivity analysis. The Annals of Statistics, 39(5):2658–2685, 2011. doi: 10.1214/11-AOS913
-
[5]
Markov Chains
Randal Douc, Eric Moulines, Pierre Priouret, and Philippe Soulier. Markov Chains . Springer, 2018
2018
-
[6]
Exponential and uniform ergodicity of Markov processes
Douglas Down, Sean P Meyn, and Richard L Tweedie. Exponential and uniform ergodicity of Markov processes. The Annals of Probability , 23(4):1671–1691, 1995
work page 1995
-
[7]
Stewart N. Ethier and Thomas G. Kurtz. Markov processes: characterization and convergence. Wiley Series in Probability and Mathematical Statistics. John Wiley & Sons, Inc., New York, 1986. doi: 10.1002/9780470316658
-
[8]
Robert Gramacy, Richard Samworth, and Ruth King. Importance tempering. Statistics and Computing, 20:1–7, 2010
work page 2010
Show all 40 references
-
[9]
Optimal mixture weights in multiple importance sampling
Hera Y He and Art B Owen. Optimal mixture weights in multiple importance sampling. arXiv preprint arXiv:1411.3954 , 2014
2014 arXiv
-
[10]
Convergence of heavy-tailed Monte carlo Markov chain algorithms
Søren F Jarner and Gareth O Roberts. Convergence of heavy-tailed Monte carlo Markov chain algorithms. Scandinavian Journal of Statistics , 34(4):781–815, 2007
2007
-
[11]
Geometric ergodicity of Metropolis algorithms
Søren Fiig Jarner and Ernst Hansen. Geometric ergodicity of Metropolis algorithms. Stochastic processes and their applications , 85(2):341–361, 2000
2000
-
[12]
Jennison
C. Jennison. Discussion on the meeting on the Gibbs sampler and other Markov chain Monte Carlo methods. Journal of the Royal Statistical Society: Series B (Methodolog- ical), 55(1):53–102, 1993
1993
-
[13]
A course in functional analysis and measure theory
Vladimir Kadets. A course in functional analysis and measure theory . Springer, 2018
2018
-
[14]
Methods of reducing sample size in Monte Carlo computations
Herman Kahn and Andy W Marshall. Methods of reducing sample size in Monte Carlo computations. Journal of the Operations Research Society of America , 1(5):263–278, 1953
1953
-
[15]
Importance is important: Generalized Markov chain importance sampling methods
Guanxun Li, Aaron Smith, and Quan Zhou. Importance is important: Generalized Markov chain importance sampling methods. arXiv preprint arXiv:2304.06251 , 2023
2023 arXiv
-
[16]
Dynamically weighted importance sampling in Monte Carlo computa- tion
Faming Liang. Dynamically weighted importance sampling in Monte Carlo computa- tion. Journal of the American Statistical Association , 97(459):807–821, 2002. 29
2002
-
[17]
Monte Carlo strategies in scientific computing , volume 10
Jun S Liu. Monte Carlo strategies in scientific computing , volume 10. Springer, 2001
2001
-
[18]
A theory for dynamic weighting in Monte Carlo computation
Jun S Liu, Faming Liang, and Wing Hung Wong. A theory for dynamic weighting in Monte Carlo computation. Journal of the American Statistical Association , 96(454): 561–573, 2001
2001
-
[19]
Foundations of locally- balanced Markov processes
Samuel Livingstone, Giorgos Vasdekis, and Giacomo Zanella. Foundations of locally- balanced Markov processes. arXiv preprint arXiv:2504.13322 , 2025
2025 arXiv
-
[20]
Optimality in importance sampling: a gentle survey
Fernando Llorente and Luca Martino. Optimality in importance sampling: a gentle survey. arXiv preprint arXiv:2502.07396 , 2025
2025
-
[21]
MCMC-driven importance samplers
Fernando Llorente, Ernesto Curbelo, Luca Martino, Victor Elvira, and David Delgado. MCMC-driven importance samplers. Applied Mathematical Modelling , 111:310–331, 2022
2022
-
[22]
Simulating ratios of normalizing constants via a simple identity: a theoretical exploration
Xiao-Li Meng and Wing Hung Wong. Simulating ratios of normalizing constants via a simple identity: a theoretical exploration. Statistica Sinica, pages 831–860, 1996
1996
-
[23]
Rates of convergence of the Hastings and Metropolis algorithms
Kerrie L Mengersen and Richard L Tweedie. Rates of convergence of the Hastings and Metropolis algorithms. The annals of Statistics , 24(1):101–121, 1996
1996
-
[24]
Markov chains and stochastic stability
Sean P Meyn and Richard L Tweedie. Markov chains and stochastic stability . Springer Science & Business Media, 2012
2012
-
[25]
Annealed importance sampling
Radford M Neal. Annealed importance sampling. Statistics and computing , 11(2): 125–139, 2001
2001
-
[26]
Art B. Owen. Monte Carlo theory, methods and examples . https://artowen.su. domains/mc/, 2013
2013
-
[27]
Jump Markov chains and rejection-free Metropolis algorithms
Jeffrey S Rosenthal, Aki Dote, Keivan Dabiri, Hirotaka Tamura, Sigeng Chen, and Ali Sheikholeslami. Jump Markov chains and rejection-free Metropolis algorithms. Computational Statistics, pages 1–23, 2021
2021
-
[28]
Selection of proposal distributions for multiple importance sampling
Vivekananda Roy and Evangelos Evangelou. Selection of proposal distributions for multiple importance sampling. Statistica Sinica, 34:27–46, 2024
2024
-
[29]
Rubinstein
Reuven Y. Rubinstein. Simulation and the Monte Carlo Method . John Wiley & Sons, Inc., New York, NY, USA, 1981. ISBN 9780471089179. doi: 10.1002/9780470316511
1981 doi
-
[30]
On a Metropolis–Hastings importance sampling estimator
Daniel Rudolf and Bj¨ orn Sprungk. On a Metropolis–Hastings importance sampling estimator. Electronic Journal of Statistics , 14(1):857–889, 2020. 30
2020
-
[31]
Markov chain importance sampling — a highly efficient estimator for MCMC
Ingmar Schuster and Ilja Klebanov. Markov chain importance sampling — a highly efficient estimator for MCMC. Journal of Computational and Graphical Statistics , pages 1–9, 2020
2020
-
[32]
Sur les fonctions d’ensemble additives et continues
Wac law Sierpi´ nski. Sur les fonctions d’ensemble additives et continues. Fundamenta Mathematicae, 3(1):240–246, 1922
1922
-
[33]
Honest importance sampling with multiple Markov chains
Aixin Tan, Hani Doss, and James P Hobert. Honest importance sampling with multiple Markov chains. Journal of Computational and Graphical Statistics , 24(3):792–826, 2015
2015
-
[34]
Optimally combining sampling techniques for Monte Carlo rendering
Eric Veach and Leonidas J Guibas. Optimally combining sampling techniques for Monte Carlo rendering. In Proceedings of the 22nd Annual Conference on Computer Graphics and Interactive Techniques, pages 419–428, 1995
1995
-
[35]
Importance sampling type estimators based on approximate marginal Markov chain Monte Carlo
Matti Vihola, Jouni Helske, and Jordan Franks. Importance sampling type estimators based on approximate marginal Markov chain Monte Carlo. Scandinavian Journal of Statistics, 47(4):1339–1376, 2020
2020
-
[36]
Complexity results for MCMC derived from quan- titative bounds
Jun Yang and Jeffrey S Rosenthal. Complexity results for MCMC derived from quan- titative bounds. The Annals of Applied Probability , 33(2):1459–1500, 2023
2023
-
[37]
Stereographic Markov chain Monte Carlo
Jun Yang, Krzysztof Latuszy´ nski, and Gareth O Roberts. Stereographic Markov chain Monte Carlo. The Annals of Statistics , 52(6):2692–2713, 2024
2024
-
[38]
On the computational com- plexity of high-dimensional Bayesian variable selection
Yun Yang, Martin J Wainwright, and Michael I Jordan. On the computational com- plexity of high-dimensional Bayesian variable selection. The Annals of Statistics , 44 (6):2497–2532, 2016
2016
-
[39]
Scalable importance tempering and Bayesian variable selection
Giacomo Zanella and Gareth Roberts. Scalable importance tempering and Bayesian variable selection. Journal of the Royal Statistical Society Series B: Statistical Method- ology, 81(3):489–517, 2019
2019
-
[40]
Rapid convergence of informed importance tempering
Quan Zhou and Aaron Smith. Rapid convergence of informed importance tempering. In International Conference on Artificial Intelligence and Statistics , pages 10939–10965. PMLR, 2022. 31
2022
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.