pith. sign in

arxiv: 2604.23229 · v1 · submitted 2026-04-25 · 🧮 math.ST · stat.TH

Solidarity of Spectral Gaps for Component-Wise Markov Chains

Pith reviewed 2026-05-08 07:12 UTC · model grok-4.3

classification 🧮 math.ST stat.TH
keywords spectral gapsMarkov chain Monte CarloGibbs samplerdeterministic scanrandom scancomponent-wise updatessolidarity principleMALA
0
0 comments X

The pith

A block-wise contraction condition ensures that random-scan and deterministic-scan component-wise Markov chains have spectral gaps that are either both positive or both zero, differing by at most polynomial factors in the number of blocks.

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

The paper establishes a solidarity principle for the L2 spectral gaps of component-wise MCMC algorithms including Gibbs samplers. Under a block-wise contraction condition on the updates, the random-scan and deterministic-scan versions either both converge or both fail to converge, with gaps differing by factors polynomial in the number of blocks. This linkage allows transfer of convergence results between scan types. The authors apply it to derive a dimension-polynomial spectral gap bound for a deterministic-scan conditional MALA algorithm targeting multivariate Gaussians, relating the rate to the precision matrix and block step sizes.

Core claim

Under this condition, the spectral gaps of the random-scan and deterministic-scan versions of the Gibbs and component-wise chains are either simultaneously positive or simultaneously zero. Moreover, the spectral gaps differ by at most polynomial factors in the number of blocks. As an application, a deterministic-scan conditional MALA for multivariate Gaussian targets is studied, yielding a spectral gap bound that is polynomial in dimension by combining the condition with known random-scan Gibbs bounds.

What carries the argument

The block-wise contraction condition for the component-wise updates, which links the convergence behavior across different scanning strategies by controlling contraction of the target distribution.

If this is right

  • Positive spectral gap results for random-scan Gibbs samplers immediately imply corresponding results for deterministic-scan versions up to polynomial factors.
  • Convergence analysis of deterministic-scan conditional MALA on Gaussians reduces to properties of the precision matrix and chosen block step sizes.
  • High-dimensional sampling problems become more tractable when bounds can be transferred between scan strategies without separate proofs.
  • The choice between random and deterministic scans need not affect the qualitative convergence behavior of component-wise chains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Deterministic-scan methods, often used in practice for their predictability, can inherit convergence guarantees from their random-scan counterparts once the contraction condition is verified.
  • The polynomial dependence suggests that in very high dimensions the practical difference between scans may remain modest for mixing times.
  • Similar solidarity arguments could be tested for other component-wise Metropolis-Hastings variants beyond Gibbs and MALA.

Load-bearing premise

The component-wise block updates satisfy a contraction condition that controls the L2 distance to the target.

What would settle it

An explicit target distribution and update rule satisfying the block-wise contraction condition for which one scan type has a positive L2 spectral gap while the other has a zero gap, or where the gaps differ by a factor super-polynomial in the number of blocks.

Figures

Figures reproduced from arXiv: 2604.23229 by Galin Jones, Qian Qin, Youngwoo Kwon.

Figure 1
Figure 1. Figure 1: The web of implications for spectral gaps. Dashed arrows indicate known results: view at source ↗
read the original abstract

Deterministic-scan and random-scan component-wise Markov chain Monte Carlo algorithms, such as Gibbs samplers and conditional Metropolis-Hastings, are popular approaches for sampling from multivariate distributions. A long-standing open question is to determine the conditions under which these algorithms have similar convergence rates. A block-wise contraction condition for the component-wise updates is used to establish a solidarity principle for the $L^2$ spectral gaps of the associated Markov chains. Specifically, under this condition, the spectral gaps of the random-scan and deterministic-scan versions of the Gibbs and component-wise chains are either simultaneously positive or simultaneously zero. Moreover, the spectral gaps differ by at most polynomial factors in the number of blocks. As an application of the general results, a deterministic-scan conditional Metropolis-adjusted Langevin algorithm (MALA) for multivariate Gaussian targets is studied. The block-wise contraction condition is combined with known spectral gap bounds for the random-scan Gibbs sampler to obtain a spectral gap bound that is polynomial in dimension. The result is used to clarify how the convergence rate of the conditional MALA depends on the precision matrix of the Gaussian target and the step sizes of the block-wise MALA updates.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proves a solidarity principle for the L² spectral gaps of random-scan and deterministic-scan versions of component-wise Markov chains (Gibbs samplers and conditional Metropolis-Hastings) under an explicitly stated block-wise contraction condition on the component-wise updates. Under this condition the gaps are simultaneously positive or simultaneously zero and differ by at most polynomial factors in the number of blocks. The result is applied to a deterministic-scan conditional MALA targeting multivariate Gaussians, where the contraction condition is combined with known random-scan Gibbs spectral-gap bounds to produce a spectral-gap lower bound that is polynomial in dimension; the dependence of the rate on the precision matrix and on the block-wise step sizes is then characterized.

Significance. If the solidarity result holds, it supplies a general, reusable device for transferring spectral-gap information between scan orders in component-wise MCMC without having to derive separate bounds for each ordering. The Gaussian MALA application demonstrates that the device yields concrete, dimension-polynomial guarantees and clarifies the role of the precision matrix and step-size choice. The approach reuses existing random-scan bounds rather than re-deriving them, which is a methodological strength for this class of problems.

major comments (2)
  1. [§3] §3 (solidarity theorem): the precise statement of the block-wise contraction condition (presumably Eq. (3.1) or (3.2)) is load-bearing; it would be useful to see an explicit verification that the condition is satisfied by the conditional MALA updates on the Gaussian target, including the dependence on the chosen step sizes.
  2. [Application section] Application section (conditional MALA): the final spectral-gap bound is stated to be polynomial in dimension, but the degree of the polynomial is not made explicit; because the solidarity result only guarantees “at most polynomial factors,” the overall degree should be tracked through the composition with the known random-scan Gibbs bound.
minor comments (2)
  1. The abstract claims the gaps “differ by at most polynomial factors in the number of blocks”; the main text should state the explicit degree (or at least an upper bound on it) for the reader’s convenience.
  2. Notation for the random-scan versus deterministic-scan operators should be introduced once and used consistently; occasional switches between “RS”/“DS” and full phrases make some passages harder to follow.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading, positive assessment of the significance, and constructive comments. We address the two major comments point by point below and will incorporate revisions to improve clarity.

read point-by-point responses
  1. Referee: [§3] §3 (solidarity theorem): the precise statement of the block-wise contraction condition (presumably Eq. (3.1) or (3.2)) is load-bearing; it would be useful to see an explicit verification that the condition is satisfied by the conditional MALA updates on the Gaussian target, including the dependence on the chosen step sizes.

    Authors: We agree that an explicit verification strengthens the application. In the revised manuscript we will add a dedicated paragraph (or short appendix subsection) that directly verifies the block-wise contraction condition for the conditional MALA updates on the multivariate Gaussian target. The verification will be carried out component-wise, explicitly displaying the dependence on the chosen step sizes and on the eigenvalues of the precision matrix, confirming that the condition holds whenever the step sizes are chosen sufficiently small relative to the local curvature. revision: yes

  2. Referee: [Application section] Application section (conditional MALA): the final spectral-gap bound is stated to be polynomial in dimension, but the degree of the polynomial is not made explicit; because the solidarity result only guarantees “at most polynomial factors,” the overall degree should be tracked through the composition with the known random-scan Gibbs bound.

    Authors: We acknowledge the value of tracking the degree. The solidarity theorem already states that the factors are polynomial in the number of blocks (here equal to dimension d), with the precise exponent depending on the contraction constants appearing in the block-wise condition. Composing with the known random-scan Gibbs bound therefore yields an overall polynomial whose degree is the sum of the solidarity exponent and the degree of the random-scan bound. In the revision we will add an explicit remark after the statement of the final bound that records this additive structure and notes the dependence on the precision-matrix eigenvalues through the random-scan bound; we do not claim a universal numerical degree because it necessarily varies with the target covariance and step-size choice. revision: partial

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper's central result is a solidarity theorem for L2 spectral gaps of random-scan versus deterministic-scan component-wise chains (Gibbs and conditional MH), derived directly from an explicitly stated block-wise contraction condition on the component-wise updates. This condition functions as an assumption, not a fitted or self-derived quantity. The MALA application then combines the solidarity result with external, pre-existing spectral-gap bounds for random-scan Gibbs samplers on Gaussians; those bounds are cited as known and independent of the present work. No equation reduces by construction to its own inputs, no parameter is fitted and then relabeled as a prediction, and no load-bearing step rests on a self-citation chain. The derivation is therefore self-contained against the stated assumption and external literature.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the block-wise contraction condition (a domain assumption) together with standard background facts about Markov operators, L2 spectral gaps, and Gibbs samplers; no free parameters or invented entities are introduced.

axioms (1)
  • domain assumption Block-wise contraction condition on component-wise updates
    This is the key hypothesis used to prove that random-scan and deterministic-scan spectral gaps are simultaneously positive or zero.

pith-pipeline@v0.9.0 · 5499 in / 1277 out tokens · 43440 ms · 2026-05-08T07:12:18.090999+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

57 extracted references · 57 canonical work pages

  1. [1]

    On random-and systematic-scan samplers.Biometrika, 103(3):719–726, 2016

    Christophe Andrieu. On random-and systematic-scan samplers.Biometrika, 103(3):719–726, 2016

  2. [2]

    Uniform ergodicity of the iterated conditional SMC and geometric ergodicity of particle Gibbs samplers.Bernoulli, pages 842–872, 2018

    Christophe Andrieu, Anthony Lee, and Matti Vihola. Uniform ergodicity of the iterated conditional SMC and geometric ergodicity of particle Gibbs samplers.Bernoulli, pages 842–872, 2018

  3. [3]

    Scalability of Metropolis-within- Gibbs schemes for high-dimensional Bayesian models.Journal of the Royal Statistical Society Series B: Statistical Methodology, 2026

    Filippo Ascolani, Gareth O Roberts, and Giacomo Zanella. Scalability of Metropolis-within- Gibbs schemes for high-dimensional Bayesian models.Journal of the Royal Statistical Society Series B: Statistical Methodology, 2026

  4. [4]

    Therateofconvergenceinthemethodof alternatingprojections.St.PetersburgMathematicalJournal,23(3):413–434,2012

    CatalinBadea,SophieGrivaux,andVladimirMüller. Therateofconvergenceinthemethodof alternatingprojections.St.PetersburgMathematicalJournal,23(3):413–434,2012. Translation from Algebra i Analiz 23 (2011), no. 1, 1–30

  5. [5]

    On the computational complexity of MCMC- based estimators in large samples.The Annals of Statistics, pages 2011–2055, 2009

    Alexandre Belloni and Victor Chernozhukov. On the computational complexity of MCMC- based estimators in large samples.The Annals of Statistics, pages 2011–2055, 2009

  6. [6]

    Brooks, Andrew Gelman, Galin L

    Stephen P. Brooks, Andrew Gelman, Galin L. Jones, and Xiao-Li Meng.Handbook of Markov Chain Monte Carlo. Chapman & Hall, London, 2011

  7. [7]

    Austin Brown and Galin L. Jones. Lower bounds on the rate of convergence for accept-reject- based Markov chains in Wasserstein and total variation distances.Bernoulli, 31:1908–1928, 2025

  8. [8]

    Markov chains for exploring posterior distributions

    Kung Sik Chan and Charles J. Geyer. Comment on “Markov chains for exploring posterior distributions”.The Annals of Statistics, 22:1747–1758, 1994

  9. [9]

    Optimal dimension dependence of the Metropolis-Adjusted Langevin Algorithm

    Sinho Chewi, Chen Lu, Kwangjun Ahn, Xiang Cheng, Thibaut Le Gouic, and Philippe Rigollet. Optimal dimension dependence of the Metropolis-Adjusted Langevin Algorithm. In Proceedings of Thirty Fourth Conference on Learning Theory, volume 134 ofProceedings of Machine Learning Research, pages 1260–1300. PMLR, 2021

  10. [10]

    SolidarityofGibbssamplers: The spectral gap.The Annals of Applied Probability, 35(1):142–157, 2025

    IwonaChlebicka,KrzysztofŁatuszyński,andBłażejMiasojedow. SolidarityofGibbssamplers: The spectral gap.The Annals of Applied Probability, 35(1):142–157, 2025

  11. [11]

    A cubic algorithm for computing Gaussian volume

    Ben Cousins and Santosh Vempala. A cubic algorithm for computing Gaussian volume. In Proceedings of the twenty-fifth annual ACM-SIAM symposium on discrete algorithms, pages 1215–1228. SIAM, 2014. 20

  12. [12]

    Doss, James M

    Charles R. Doss, James M. Flegal, Galin L. Jones, and Ronald C. Neath. Markov chain Monte Carlo estimation of quantiles.Electronic Journal of Statistics, 8:2448–2478, 2014

  13. [13]

    Springer Series in Operations Research and Financial Engineering

    Randal Douc, Éric Moulines, and Pierre Priouret.Markov Chains. Springer Series in Operations Research and Financial Engineering. Springer, Cham, 2018

  14. [14]

    Wainwright, and Bin Yu

    Raaz Dwivedi, Yuansi Chen, Martin J. Wainwright, and Bin Yu. Log-concave sampling: Metropolis-Hastingsalgorithmsarefast.JournalofMachineLearningResearch,20(183):1–42, 2019

  15. [15]

    Comparisontheoremsforthemixingtimesofsystematic and random scan dynamics

    JasonGaitondeandElchananMossel. Comparisontheoremsforthemixingtimesofsystematic and random scan dynamics. InProceedings of the 2026 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 2575–2587. SIAM, 2026

  16. [16]

    Improved concentration bounds for Gaussian quadratic forms.arXiv preprint arXiv:1911.05720, 2019

    Robert E Gallagher, Louis JM Aslett, David Steinsaltz, and Ryan R Christ. Improved concentration bounds for Gaussian quadratic forms.arXiv preprint arXiv:1911.05720, 2019

  17. [17]

    Equivalences of geometric ergodicity of markov chains.Journal of Theoretical Probability, 37(2):1230–1256, 2024

    Marco A Gallegos-Herrada, David Ledvinka, and Jeffrey S Rosenthal. Equivalences of geometric ergodicity of markov chains.Journal of Theoretical Probability, 37(2):1230–1256, 2024

  18. [18]

    MengxiGao,GarethORoberts,andAndiQWang.WeakPoincaréinequalitiesfordeterministic- scan Metropolis-within-Gibbs samplers.arXiv preprint arXiv:2602.14692, 2026

  19. [19]

    Gelfand and Adrian F

    Alan E. Gelfand and Adrian F. M. Smith. Sampling-based approaches to calculating marginal densities.Journal of the American Statistical Association, 85(410):398–409, 1990

  20. [20]

    Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images.IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-6(6):721–741, 1984

    Stuart Geman and Donald Geman. Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images.IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-6(6):721–741, 1984

  21. [21]

    On the central limit theorem for geometrically ergodic Markov chains

    Olle Häggström. On the central limit theorem for geometrically ergodic Markov chains. Probability Theory and Related Fields, 132:74–82, 2005

  22. [22]

    Monte Carlo sampling methods using Markov chains and their applications

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

  23. [23]

    B. He, C. De Sa, I. Mitliagkas, and C. Ré. Scan order in Gibbs sampling: models in which it matters and bounds on how much.Advances in Neural Information Processing Systems, 29, 2016

  24. [24]

    GilbertHelmberg.IntroductiontoSpectralTheoryinHilbertspace.CourierDoverPublications, 2008

  25. [25]

    Component-wise markov chain monte carlo: Uniform and geometric ergodicity under mixing and composition.Statistical Science, pages 360–375, 2013

    Alicia A Johnson, Galin L Jones, and Ronald C Neath. Component-wise markov chain monte carlo: Uniform and geometric ergodicity under mixing and composition.Statistical Science, pages 360–375, 2013

  26. [26]

    Variable transformation to obtain geometric ergodicity in the random-walk Metropolis algorithm.The Annals of Statistics, pages 3050–3076, 2012

    Leif T Johnson and Charles J Geyer. Variable transformation to obtain geometric ergodicity in the random-walk Metropolis algorithm.The Annals of Statistics, pages 3050–3076, 2012. 21

  27. [27]

    On the Markov chain central limit theorem.Probability Surveys, 1:299–320, 2004

    Galin L Jones. On the Markov chain central limit theorem.Probability Surveys, 1:299–320, 2004

  28. [28]

    Fixed-widthoutputanalysisfor Markov chain Monte Carlo.Journal of the American Statistical Association, 101:1537–1547, 2006

    GalinL.Jones,MuraliHaran,BrianS.Caffo,andRonaldNeath. Fixed-widthoutputanalysisfor Markov chain Monte Carlo.Journal of the American Statistical Association, 101:1537–1547, 2006

  29. [29]

    Jones and James P

    Galin L. Jones and James P. Hobert. Honest exploration of intractable probability distributions via Markov chain Monte Carlo.Statistical Science, 16:312–334, 2001

  30. [30]

    Jones, Gareth O

    Galin L. Jones, Gareth O. Roberts, and Jeffrey S. Rosenthal. Convergence of conditional Metropolis-Hastings samplers.Advances in Applied Probability, 46(2):422–445, 2014

  31. [31]

    Sampling according to the multivariate normal density

    Ravi Kannan and Guangxing Li. Sampling according to the multivariate normal density. In Proceedings of 37th Conference on Foundations of Computer Science, pages 204–212. IEEE, 1996

  32. [32]

    Aphasetransitioninsampling from restricted Boltzmann machines.The Annals of Applied Probability, 36(2):1615–1652, 2026

    YoungwooKwon,QianQin,GuanyangWang,andYuchenWei. Aphasetransitioninsampling from restricted Boltzmann machines.The Annals of Applied Probability, 36(2):1615–1652, 2026

  33. [33]

    Bounds on the L2 spectrum for Markov chains and Markov processes: a generalization of Cheeger’s inequality.Transactions of the American mathematical society, 309(2):557–580, 1988

    Gregory F Lawler and Alan D Sokal. Bounds on the L2 spectrum for Markov chains and Markov processes: a generalization of Cheeger’s inequality.Transactions of the American mathematical society, 309(2):557–580, 1988

  34. [34]

    Isoperimetry and gaussian analysis

    Michel Ledoux. Isoperimetry and gaussian analysis. InLectures on Probability Theory and Statistics: Ecole d’Eté de Probabilités de Saint-Flour XXIV—1994, pages 165–294. Springer, 2006

  35. [35]

    American Mathematical Soc., 2017

    David A Levin and Yuval Peres.Markov Chains and Mixing times, volume 107. American Mathematical Soc., 2017

  36. [36]

    Hit-and-run mixes fast.Mathematical Programming, 86(3):443–461, 1999

    László Lovász. Hit-and-run mixes fast.Mathematical Programming, 86(3):443–461, 1999

  37. [37]

    Random walks in a convex body and an improved volume algorithm.Random Structures & Algorithms, 4(4):359–412, 1993

    László Lovász and Miklós Simonovits. Random walks in a convex body and an improved volume algorithm.Random Structures & Algorithms, 4(4):359–412, 1993

  38. [38]

    Xavier Mak and James P. Hobert. Extensions of the solidarity principle of the spectral gap for Gibbs samplers to their blocked and collapsed variants. 2026

  39. [39]

    Equation of state calculations by fast computing machines.The Journal of Chemical Physics, 21(6):1087–1092, 1953

    Nicholas Metropolis, Arianna W Rosenbluth, Marshall N Rosenbluth, Augusta H Teller, and Edward Teller. Equation of state calculations by fast computing machines.The Journal of Chemical Physics, 21(6):1087–1092, 1953

  40. [40]

    Prentice-HallEnglewood Cliffs, NJ, 1977

    BenNoble,JamesWDaniel,etal.AppliedLinearAlgebra,volume3. Prentice-HallEnglewood Cliffs, NJ, 1977

  41. [41]

    Convergence bounds for Monte Carlo Markov chains

    Qian Qin. Convergence bounds for Monte Carlo Markov chains. InHandbook of Markov Chain Monte Carlo, pages 492–526. Chapman and Hall/CRC, 2024. 22

  42. [42]

    On spectral gap decomposition for Markov chains.arXiv:2504.01247, 2025

    Qian Qin. On spectral gap decomposition for Markov chains.arXiv:2504.01247, 2025

  43. [43]

    Convergenceratesoftwo-componentMCMCsamplers.Bernoulli, 29(1):338–363, 2023

    QianQinandGalinL.Jones. Convergenceratesoftwo-componentMCMCsamplers.Bernoulli, 29(1):338–363, 2023

  44. [44]

    Spectral gap bounds for reversible hybrid gibbs chains.The Annals of Statistics, 53(4):1613–1638, 2025

    Qian Qin, Nianqiao Ju, and Guanyang Wang. Spectral gap bounds for reversible hybrid gibbs chains.The Annals of Statistics, 53(4):1613–1638, 2025

  45. [45]

    Roberts and Jeffrey S

    Gareth O. Roberts and Jeffrey S. Rosenthal. Geometric ergodicity and hybrid Markov chains. Electronic Communications in Probability, 2:13–25, 1997

  46. [46]

    Optimal scaling of discrete approximations to Langevindiffusions.JournaloftheRoyalStatisticalSociety: SeriesB(StatisticalMethodology), 60(1):255–268, 1998

    Gareth O Roberts and Jeffrey S Rosenthal. Optimal scaling of discrete approximations to Langevindiffusions.JournaloftheRoyalStatisticalSociety: SeriesB(StatisticalMethodology), 60(1):255–268, 1998

  47. [47]

    Surprising convergence properties of some simple Gibbs samplers under various scans.International Journal of Statistics and Probability, 5(1):51–60, 2015

    Gareth O Roberts and Jeffrey S Rosenthal. Surprising convergence properties of some simple Gibbs samplers under various scans.International Journal of Statistics and Probability, 5(1):51–60, 2015

  48. [48]

    Roberts and Sujit K

    Gareth O. Roberts and Sujit K. Sahu. Updating schemes, correlation structure, blocking and parameterization for the Gibbs sampler.Journal of the Royal Statistical Society: Series B (Statistical Methodology), 59(2):291–317, 1997

  49. [49]

    Flegal, Dootika Vats, and Galin L

    Nathan Robertson, James M. Flegal, Dootika Vats, and Galin L. Jones. Assessing and visualizing simultaneous simulation error.Journal of Computational and Graphical Statistics, 30(2):324–334, 2021

  50. [50]

    International Series in Pure and Applied Mathematics

    Walter Rudin.Functional Analysis. International Series in Pure and Applied Mathematics. McGraw-Hill, New York, 2 edition, 1991

  51. [51]

    Weightedshape-constrainedestimationfortheautocovariance sequence from a reversible markov chain.arXiv:2408.03024, 2024

    HyebinSongandStephenBerg. Weightedshape-constrainedestimationfortheautocovariance sequence from a reversible markov chain.arXiv:2408.03024, 2024

  52. [52]

    On the Lyapunov Foster criterion and Poincaré inequalityforreversibleMarkovchains.IEEETransactionsonAutomaticControl,67(5):2605– 2609, 2021

    Amirhossein Taghvaei and Prashant G Mehta. On the Lyapunov Foster criterion and Poincaré inequalityforreversibleMarkovchains.IEEETransactionsonAutomaticControl,67(5):2605– 2609, 2021

  53. [53]

    Flegal, and Galin L

    Dootika Vats, James M. Flegal, and Galin L. Jones. Multivariate output analysis for Markov chain Monte Carlo.Biometrika, 106:321–337, 2019. 23 A Proofs in Section 2 Proof of Lemma 2.1.Let𝑋=(𝑥 𝑗 , 𝑥−𝑗 ) ∼Π. For any𝑓∈𝐿 2(Π), we have (𝑃 𝑗 𝑓) (𝑥 𝑗 , 𝑥−𝑗 )= ∫ X𝑗 𝑓(𝑦 𝑗 , 𝑥−𝑗 )Π 𝑗 (d𝑦 𝑗 |𝑥 −𝑗 ), where Π 𝑗 (𝑑𝑦 𝑗 |𝑥 −𝑗 ) isaconditionaldistributionof 𝑦 𝑗 given 𝑥−𝑗....

  54. [54]

    42 Then, for every𝑎∈ (0,1), Φ(𝑀 2 Δ 𝑗 ) ≥𝜖min 1−𝑎 2 , 𝑎2𝜅𝑡 4

    Three-set isoperimetric inequality:For any measurable partition{𝑆1, 𝑆2, 𝑆3} of R𝑁 𝑗 such thatd(𝑆 1, 𝑆2) ≥𝑡, 𝛾𝑁 𝑗 (𝑆 3) ≥𝜅𝑡𝛾 𝑁 𝑗 (𝑆 1)𝛾 𝑁 𝑗 (𝑆 2). 42 Then, for every𝑎∈ (0,1), Φ(𝑀 2 Δ 𝑗 ) ≥𝜖min 1−𝑎 2 , 𝑎2𝜅𝑡 4 . Since MALA consists of a Gaussian proposal followed by a Metropolis step, we first bound the total variation distance between the proposal distribut...

  55. [55]

    For any𝑢, 𝑣∈R 𝑁 𝑗, supposedΔ 𝑗 (𝑢, 𝑣) ≤𝑡=𝜂 √

    Set 𝜖=𝜀(𝜂,Δ 𝑗 )= 2 1+𝛿 max exp ( − tr(Δ 2 𝑗 ) 2(1−𝛿 max) ) −2Φ 𝑁 ∥𝐼−Δ 𝑗 ∥op 2 𝜂 . For any𝑢, 𝑣∈R 𝑁 𝑗, supposedΔ 𝑗 (𝑢, 𝑣) ≤𝑡=𝜂 √

  56. [56]

    By assumption, tr(Δ 2 𝑗 )=ℎ 2 𝑗 tr(𝑄 2 𝑗, 𝑗 ) ≤ 1 10 ≤log(10/9), and since𝛿max ≤1/2, this implies tr(Δ 2 𝑗 ) ≤2(1−𝛿 max)log 5/3 1+𝛿 max

    Since 𝛿max =𝜓 max(Δ 𝑗 ) ≤1/2<1 , Lemma D.7 yields ∥𝑀 2 Δ 𝑗 (𝑢,·) −𝑀 2 Δ 𝑗 (𝑣,·) ∥ TV ≤ ∥𝑀 Δ 𝑗 (𝑢,·) −𝑀 Δ 𝑗 (𝑣,·) ∥ TV ≤1−𝜖 . By assumption, tr(Δ 2 𝑗 )=ℎ 2 𝑗 tr(𝑄 2 𝑗, 𝑗 ) ≤ 1 10 ≤log(10/9), and since𝛿max ≤1/2, this implies tr(Δ 2 𝑗 ) ≤2(1−𝛿 max)log 5/3 1+𝛿 max . Hence the first term in the definition of𝜖 is at least6/5. Also, since∥𝐼−Δ 𝑗 ∥op =1−𝜓 min(Δ 𝑗 ...

  57. [57]

    Hence Φ(𝑀 2 Δ 𝑗 ) ≥𝜖 𝜅𝑡 8 ≥ √︁ 𝜓min(Δ 𝑗 ) 80√𝜋

    Since 𝜓min(Δ 𝑗 ) ≤𝜓 max(Δ 𝑗 )=ℎ 𝑗 𝜓max(𝑄 𝑗, 𝑗 ) ≤ 1 2, we have𝜅𝑡= √︁ 𝜓min(Δ 𝑗 )/𝜋 <1 , which implies(1−𝑎 ★)/2≥1/8 and 𝑎2 ★𝜅𝑡/4=𝜅𝑡/8≤1/8 . Hence Φ(𝑀 2 Δ 𝑗 ) ≥𝜖 𝜅𝑡 8 ≥ √︁ 𝜓min(Δ 𝑗 ) 80√𝜋 . By Lemma D.3, 1− ∥𝑀 2 Δ 𝑗 −Π 𝛾 𝑗 ∥ ≥ Φ(𝑀 2 Δ 𝑗 )2 8 ≥ 𝜓min(Δ 𝑗 ) 51200𝜋 . Finally, by (18), 1− ∥𝑀 Δ 𝑗 −Π 𝛾 𝑗 ∥ ≥ 1− ∥𝑀 2 Δ 𝑗 −Π 𝛾 𝑗 ∥ 2 ≥ 𝜓min(Δ 𝑗 ) 102400𝜋 , 49 which yi...