← All problems

Non-asymptotic convergence problem for stereographic MCMC

For stereographic projection MCMC, and likewise for the sub-Cauchy projection sampler, determine explicit finite-time bounds on tmix(ε;μ,P):=inf⁡{t≥0:∥μPt−π∥TV≤ε},t_{\mathrm{mix}}(\varepsilon;\mu,P):=\inf\{t\ge 0:\lVert\mu P^t-\pi\rVert_{\mathrm{TV}}\le\varepsilon\}, where π\pi is a high-dimensional target distribution and μ\mu is an initial distribution. In particular, establish conditions on π\pi and μ\mu under which tmix(ε;μ,P)t_{\mathrm{mix}}(\varepsilon;\mu,P) is bounded polynomially in the dimension for a broad class of light- and heavy-tailed targets.

References

Source: arXiv

Progress summary

Refreshed
Open · solution reported

A new paper claims to settle the question by proving finite-time guarantees for stereographic samplers, but the proof has not been independently checked.

The problem asks for non-asymptotic convergence guarantees for stereographic MCMC applied to high-dimensional, potentially heavy-tailed targets. An earlier paper established asymptotic and geometric-ergodicity results, but left the relevant finite-time behavior unresolved.

Known results

  • The earlier stereographic-MCMC work proves uniform ergodicity under tail conditions, including multivariate Student targets with degrees of freedom ν≥d\nu \ge d.
  • It analyzes certain high-dimensional product and isotropic targets, while identifying limitations for heavier-tailed cases such as ν<d\nu<d.

October 2026 claimed resolution

An arXiv preprint by Federica Milinanni, Tyler Farghly, and Jun Yang claims polynomial non-asymptotic mixing bounds for stereographic and sub-Cauchy projection samplers, together with a weighted isoperimetric inequality with optimal dimension dependence. The claim applies only under stated conditions on the target and initial distributions, and no independent assessment, correction, withdrawal, or retraction was found.

Current status (as of October 2026): A preprint claims to resolve the non-asymptotic convergence question, but its proof is unverified; the result should therefore be treated as claimed rather than settled.

Sources

Solutions 0

No solutions have been posted yet.