Skip to content
DraconiticResearch
Search Sources API
197,768 studies31 sourcesUpdated 16 min ago
PreprintarXiv licencearXiv

Spectral Gap Inheritance for Metropolis-Hastings with Reversible Proposals

  • Minhyeok Kim
  • Vinayak Rao
Open at sourcePDFDOICite

Abstract

We study Metropolis-Hastings kernels whose proposal kernels are reversible with respect to a reference measure, and establish sufficient conditions for such kernels to inherit a nonzero right spectral gap from their proposal kernel. Our conditions require the importance weight of the target relative to the reference measure to be bounded above, but do not require it to be bounded away from zero. Our approach combines a localized Holley-Stroock argument, which yields a local Poincare inequality, with a Lyapunov drift condition. As our main application, we analyze the preconditioned Crank-Nicolson (pCN) algorithm for sampling from measures on a Hilbert space that are absolutely continuous with respect to a Gaussian reference measure. Under local boundedness and quasiconvexity assumptions on the potential, we establish a dimension-independent L2-spectral gap for this Metropolis-Hastings algorithm, thereby covering cases in which the local Lipschitz condition used by Hairer et al. (2014) is not satisfied.

Related

  1. Convergence of Independence Metropolis-Hastings Algorithms in High Dimensions: Iteration Lengths, Cutoff Profiles, and Diagnostic Unreliability
    Austin Brown
    arXivPreprintCC BY 4.0Full textStatisticsarXiv
  2. Generalizing Metropolis-Hastings for DAG Sampling
    Daniele Nikzad, Mikko Koivisto
    arXivPreprintarXiv licenceFull textStatisticsarXiv
  3. When Metropolis and Hastings Meet Bradley and Terry: Exact MCMC From Preference Voting
    Ariel Smogorghevski, Nir Rosenfeld, Yaniv Romano
    arXivPreprintCC BY-NC-SA 4.0Full textStatisticsarXiv
  4. Fixed-Horizon Error Rates for Array-RQMC Metropolis-Hastings
    Jiarui Du, Josef Dick, Zexin Pan +1
    arXivPreprintCC BY 4.0Full textStatisticsarXiv
  5. A spectral gap for Metropolis-adjusted Langevin algorithm with a uniformly randomized step size
    Qian Qin
    arXivPreprintarXiv licenceFull textProbabilityStatisticsarXiv

Cite

BibTeXRIS
BibTeX
@misc{kim2026spectral,
  title = {{Spectral Gap Inheritance for Metropolis-Hastings with Reversible Proposals}},
  author = {Minhyeok Kim and Vinayak Rao},
  howpublished = {arXiv},
  year = {2026},
  month = nov,
  doi = {10.48550/arxiv.2611.02381},
  eprint = {2611.02381},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.02381},
}
RIS
TY  - UNPB
TI  - Spectral Gap Inheritance for Metropolis-Hastings with Reversible Proposals
AU  - Minhyeok Kim
AU  - Vinayak Rao
PY  - 2026
DA  - 2026/11/02
PB  - arXiv
DO  - 10.48550/arxiv.2611.02381
UR  - https://arxiv.org/abs/2611.02381
AB  - We study Metropolis-Hastings kernels whose proposal kernels are reversible with respect to a reference measure, and establish sufficient conditions for such kernels to inherit a nonzero right spectral gap from their proposal kernel. Our conditions require the importance weight of the target relative to the reference measure to be bounded above, but do not require it to be bounded away from zero. Our approach combines a localized Holley-Stroock argument, which yields a local Poincare inequality, with a Lyapunov drift condition. As our main application, we analyze the preconditioned Crank-Nicolson (pCN) algorithm for sampling from measures on a Hilbert space that are absolutely continuous with respect to a Gaussian reference measure. Under local boundedness and quasiconvexity assumptions on the potential, we establish a dimension-independent L^2-spectral gap for this Metropolis-Hastings algorithm, thereby covering cases in which the local Lipschitz condition used by Hairer et al. (2014) is not satisfied.
ER  -