Bayesian Regularization · 2010-stat-9

Source-reported assessment from the ICM catalogue snapshot. This is not an independent review by Frontier Science.

The exact problem remains unresolved as of July 31, 2026. The literature proves the displayed upper bound for the hierarchical squared-exponential prior, but I found no citable result proving a matching lower bound, proving that every achievable rate must contain a divergent logarithmic factor, or identifying the optimal logarithmic exponent for density estimation over Hölder log-densities.

• Aad W. van der Vaart and J. Harry van Zanten, “Adaptive Bayesian Estimation Using a Gaussian Random Field with Inverse Gamma Bandwidth,” The Annals of Statistics 37(5B), 2009, 2655–2675. DOI: 10.1214/08-AOS678; arXiv:0908.3556. Theorem 2.1 proves contraction in Hellinger distance at $n^{-\alpha/(2\alpha+d)}(\log n)^{(4\alpha+d)/(4\alpha+2d)}$ for the exact prior and density model in the statement. The discussion immediately after the theorem explicitly conjectures that some logarithmic factor is necessary and says that the displayed exponent may be nonoptimal; it supplies no lower bound.

• Aad W. van der Vaart and J. Harry van Zanten, “Information Rates of Nonparametric Gaussian Process Methods,” Journal of Machine Learning Research 12, 2011, 2095–2119. Direct URL: https://www.jmlr.org/papers/v12/vandervaart11a.html. Theorem 8 gives a genuine slow-rate lower bound for an unrescaled, fixed squared-exponential Gaussian-process prior in Gaussian regression. It does not apply to the non-Gaussian scale mixture obtained by assigning the inverse-Gamma bandwidth prior, and it does not concern logistic Gaussian-process density estimation. Thus it establishes that lower bounds can be sharp in a related oversmoothing problem but does not resolve the stated conjecture.

• Amine Hadji and Botond Szabó, “Can We Trust Bayesian Uncertainty Quantification from Gaussian Process Priors with Squared Exponential Covariance Kernel?”, SIAM/ASA Journal on Uncertainty Quantification 9(1), 2021, 185–230. DOI: 10.1137/19M1253010; arXiv:1904.01383. Appendix Theorem A.5 proves an adaptive contraction upper bound, up to logarithmic factors, for a hierarchical approximately squared-exponential sequence prior in the Gaussian white-noise model. Its principal lower/negative results concern frequentist coverage of credible sets, not the sharp contraction rate of the exact density prior in the statement.

• Xiao Fang and Anindya Bhadra, “Posterior Concentration for Gaussian Process Priors under Rescaled and Hierarchical Matérn and Confluent Hypergeometric Covariance Functions,” Electronic Journal of Statistics 19(2), 2025. DOI: 10.1214/25-EJS2417; arXiv:2312.07502. Theorems 3.10–3.13 obtain exact minimax contraction without a logarithmic loss for hierarchical Matérn and confluent-hypergeometric priors in fixed-design regression. The paper contrasts these finite-smoothness priors with the hierarchical squared-exponential prior, for which it continues to cite only the older “up to a logarithmic factor” rate. Because the covariance family and statistical model differ, these results do not answer the stated density-estimation problem.

Established progress therefore consists of the original near-minimax upper bound and sharp or exact results for neighboring priors and observation models. The missing step is a lower bound for the exact inverse-Gamma-rescaled squared-exponential density prior that either forces a positive logarithmic loss and determines its sharp exponent, or a proof that the logarithm can be removed or improved.