Pré-Publication, Document De Travail Année : 2026

Bernoulli and Gauss Take a Look at the MAP

Résumé

This article discusses two Maximum a Posteriori (MAP) interpretations for state-of-the-art methods used in sparse inverse problems: the joint-MAP and the Marginal-MAP. Canonically rooted in a Bayesian framework, sparsity is modeled by a general spike and slab distribution. The focus is on the recovery of the solution support rather than on signal amplitudes. We study the prominent Bernoulli-Gaussian model leading to NP-hard optimization problems. We show that a judicious re-parametrization of the joint-MAP may indeed be a nice surrogate of the marginal-MAP. Additionally, we explore common continuous relaxations of the support and encompass them under the scope of a parametrized distribution. Upon describing the behavior of a few relaxations, strong links are established between the Bernoulli-Gaussian joint-MAP, marginal-MAP, and well-studied methods such as the Lasso and Sparse Bayesian Learning. Finally, the utilization of randomized rounding for both joint-MAP and marginal-MAP problems yields valuable insights into obtaining sparse solutions with an emphasis on support recovery.

Fichier principal
Vignette du fichier
MMAPvsJMAP.pdf (308.58 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-04782425 , version 1 (14-11-2024)
hal-04782425 , version 2 (23-02-2026)

Licence

Identifiants

  • HAL Id : hal-04782425 , version 2

Citer

Pierre Barbault, Matthieu Kowalski, Charles Soussen. Bernoulli and Gauss Take a Look at the MAP. 2026. ⟨hal-04782425v2⟩
1478 Consultations
452 Téléchargements

Partager

  • More