Skip to Main content Skip to Navigation
Theses

First-order noneuclidean splitting methods for large-scale optimization : deterministic and stochastic algorithms

Abstract : In this work we develop and examine two novel first-order splitting algorithms for solving large-scale composite optimization problems in infinite-dimensional spaces. Such problems are ubiquitous in many areas of science and engineering, particularly in data science and imaging sciences. Our work is focused on relaxing the Lipschitz-smoothness assumptions generally required by first-order splitting algorithms by replacing the Euclidean energy with a Bregman divergence. These developments allow one to solve problems having more exotic geometry than that of the usual Euclidean setting. One algorithm is hybridization of the conditional gradient algorithm, making use of a linear minimization oracle at each iteration, with an augmented Lagrangian algorithm, allowing for affine constraints. The other algorithm is a primal-dual splitting algorithm incorporating Bregman divergences for computing the associated proximal operators. For both of these algorithms, our analysis shows convergence of the Lagrangian values, subsequential weak convergence of the iterates to solutions, and rates of convergence. In addition to these novel deterministic algorithms, we introduce and study also the stochastic extensions of these algorithms through a perturbation perspective. Our results in this part include almost sure convergence results for all the same quantities as in the deterministic setting, with rates as well. Finally, we tackle new problems that are only accessible through the relaxed assumptions our algorithms allow. We demonstrate numerical efficiency and verify our theoretical results on problems like low rank, sparse matrix completion, inverse problems on the simplex, and entropically regularized Wasserstein inverse problems.
Document type :
Theses
Complete list of metadata

https://tel.archives-ouvertes.fr/tel-03154499
Contributor : Abes Star :  Contact
Submitted on : Monday, March 1, 2021 - 8:51:28 AM
Last modification on : Tuesday, March 2, 2021 - 3:22:48 AM

File

sygal_fusion_28704-silveti_fal...
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-03154499, version 1

Citation

Antonio Silveti Falls. First-order noneuclidean splitting methods for large-scale optimization : deterministic and stochastic algorithms. Optimization and Control [math.OC]. Normandie Université, 2021. English. ⟨NNT : 2021NORMC204⟩. ⟨tel-03154499⟩

Share

Metrics

Record views

96

Files downloads

53