← All problems

Dimension-free f-divergence comparison question for Gaussian location mixtures

For every R<∞R<\infty, does there exist a constant CR<∞C_R<\infty, independent of the dimension dd and of the mixing distributions, such that for every d≥1d\ge 1 and every pair of probability measures π,η\pi,\eta supported in the Euclidean ball B(0,R)⊆RdB(0,R)\subseteq\mathbb{R}^d, the Gaussian location-mixture densities fπ(x)=∫φd(x−θ) dπ(θ)f_\pi(x)=\int \varphi_d(x-\theta)\,d\pi(\theta) and fη(x)=∫φd(x−θ) dη(θ)f_\eta(x)=\int \varphi_d(x-\theta)\,d\eta(\theta) satisfy KL(fπ∥fη)≤CRH2(fπ,fη)\mathrm{KL}(f_\pi\|f_\eta)\le C_R H^2(f_\pi,f_\eta) and χ2(fπ∥fη)≤CRH2(fπ,fη)\chi^2(f_\pi\|f_\eta)\le C_R H^2(f_\pi,f_\eta), where φd\varphi_d is the standard Gaussian density on Rd\mathbb{R}^d?

References

Progress summary

Refreshed
Open · solution reported

An October 2026 preprint claims dimension- and component-independent comparison bounds for Gaussian mixtures; the claim is unverified.

The question asks whether divergence comparisons for Gaussian location mixtures can avoid dependence on ambient dimension and mixture complexity. Lee and Ma claim affirmative bounds for compactly supported mixing distributions, with constants depending only on the support radius.

Known results

  • For kk-component mixtures, KL(fπ∥fη)≍kH2(fπ,fη)\mathrm{KL}(f_\pi\|f_\eta)\asymp_k H^2(f_\pi,f_\eta) was known with constants independent of dimension but exponential in kk (Jia et al., 2023).
  • The same 2023 work left dimension-free comparison for continuous Gaussian mixtures unclear.
  • A 2026 preprint obtained near-comparison between HH and TV\mathrm{TV}, while allowing dependence on dimension and support radius.

October 2026 Gaussian-interpolation preprint

Lee and Ma report dimension- and component-count-independent comparison inequalities for the stated compact-support Gaussian-mixture setting, with constants depending only on the support radius. This would settle the tracked question in that setting, but the paper is an unrefereed preprint and the result remains unverified.

Current status (as of October 2026): A claimed dimension- and component-independent result exists for compactly supported Gaussian location mixtures, but it is not independently verified.

Sources

Solutions 0

No solutions have been posted yet.