Skip to main content

Lund University Publications

LUND UNIVERSITY LIBRARIES

Variational shrinkage and excess risk in sparse normal means

Javed, Farrukh LU (2026) In Statistics and Probability Letters 238.
Abstract

We study variational shrinkage in the sparse normal means model and compare it with the corresponding exact Bayes shrinkage rule. The focus is on the excess quadratic risk of the variational posterior mean. Using a squared-loss decomposition, we show that this excess risk is bounded by the discrepancy between the variational and exact Bayes estimators together with the benchmark Bayes risk. This yields a simple criterion for asymptotic equivalence, since the two estimators have asymptotically the same quadratic risk whenever the variational shrinkage rule is sufficiently close to the exact Bayes rule. Simulations and a semi-synthetic Golub leukemia example illustrate how the risk gap behaves under exact sparsity, approximate sparsity,... (More)

We study variational shrinkage in the sparse normal means model and compare it with the corresponding exact Bayes shrinkage rule. The focus is on the excess quadratic risk of the variational posterior mean. Using a squared-loss decomposition, we show that this excess risk is bounded by the discrepancy between the variational and exact Bayes estimators together with the benchmark Bayes risk. This yields a simple criterion for asymptotic equivalence, since the two estimators have asymptotically the same quadratic risk whenever the variational shrinkage rule is sufficiently close to the exact Bayes rule. Simulations and a semi-synthetic Golub leukemia example illustrate how the risk gap behaves under exact sparsity, approximate sparsity, and real-data-derived sparse signals.

(Less)
Please use this url to cite or link to this publication:
author
organization
publishing date
type
Contribution to journal
publication status
published
subject
keywords
Bayes shrinkage, Excess risk, High-dimensional inference, Sparse normal means, Variational inference
in
Statistics and Probability Letters
volume
238
article number
110846
publisher
Elsevier
external identifiers
  • scopus:105040120813
ISSN
0167-7152
DOI
10.1016/j.spl.2026.110846
language
English
LU publication?
yes
id
b955b484-65fd-422c-8e32-b31510d657df
date added to LUP
2026-08-13 11:15:26
date last changed
2026-08-13 11:15:35
@article{b955b484-65fd-422c-8e32-b31510d657df,
  abstract     = {{<p>We study variational shrinkage in the sparse normal means model and compare it with the corresponding exact Bayes shrinkage rule. The focus is on the excess quadratic risk of the variational posterior mean. Using a squared-loss decomposition, we show that this excess risk is bounded by the discrepancy between the variational and exact Bayes estimators together with the benchmark Bayes risk. This yields a simple criterion for asymptotic equivalence, since the two estimators have asymptotically the same quadratic risk whenever the variational shrinkage rule is sufficiently close to the exact Bayes rule. Simulations and a semi-synthetic Golub leukemia example illustrate how the risk gap behaves under exact sparsity, approximate sparsity, and real-data-derived sparse signals.</p>}},
  author       = {{Javed, Farrukh}},
  issn         = {{0167-7152}},
  keywords     = {{Bayes shrinkage; Excess risk; High-dimensional inference; Sparse normal means; Variational inference}},
  language     = {{eng}},
  publisher    = {{Elsevier}},
  series       = {{Statistics and Probability Letters}},
  title        = {{Variational shrinkage and excess risk in sparse normal means}},
  url          = {{http://dx.doi.org/10.1016/j.spl.2026.110846}},
  doi          = {{10.1016/j.spl.2026.110846}},
  volume       = {{238}},
  year         = {{2026}},
}