Grant: Discovery Grants Program - Individual
The University of British Columbia
$37,000.00
New horizons in nonconvex and nonsmooth optimization: regularity theory and splitting algorithms
EF 2023-24
Résumé
Optimization plays a critical role in Canada's economy and society. Over the past decade, nonconvex and nonsmooth optimization problems have become more and more popular in statistics, deep learning or neural networks, and communication theory in engineering, etc. To solve these challenging optimization problems, theories and algorithms from convex analysis do not apply because the functions involved are not convex. Techniques from differentiable optimization do not apply neither because the functions are not differentiable. New research is needed pressingly to advance our understanding of nonconvex and nonsmooth optimization. From the point of view of iteration methods, almost all splitting algorithms need both operators to be monotone. It is not clear what will happens if one operator is not monotone. From the point of view of theory, proximal mappings of nonconvex functions are set-valued. This is a huge challenge for convergence analysis of iteration methods. Moving from convexity to nonconvexity, from monotone mappings to nonmonotone mappings, is intrinsically hard and often requires new methodologies and techniques. Although some progress has been made, they are ad hoc and not satisfactory. A comprehensive investigation is needed. The goal of this proposal is to study variational monotone operators and variational convex functions, and to develop convergence theory and algorithms for solving inclusions in which one operator is monotone and the other is not necessarily monotone, or for solving optimization problems in which one function is lower semicontinuous and the other is only differentiable with or without a Lipschitz gradient. The long-term goal is to establish a systematic theory on local convergence analysis of many splitting algorithms, such as the forward-backward, Douglas-Rachford, and Spingarn partial inverse methods involving possibly nonconvex and nonsmooth functions or mappings without monotonicity. Recently, with his deep insights, Rockafellar made a groundbreaking advance in studying proximal point algorithms for nonconvex functions, namely variational convexity. This discovery can beautifully extend some key results from convex optimization to nonconvex optimization. I will carry out a systematic study of variational convex functions and explore how to use them in splitting algorithms effectively. Variational convexity, variational monotonicity, metric regularity, etc., will be used to study convergence rates. Developing generalized and local nonexpansive mappings will be at the heart of the convergence theory. Seeing the success of each of my HQP brings me joy, and this program will provide opportunities to train a diverse group of HQP. While postdoctoral fellows and graduate students can concentrate on theory developments, undergraduates can do some numerical experiments. This research will profoundly advance our knowledge and produce efficient algorithms for solving nonconvex and nonsmooth optimization problems.
Programme
Discovery Grants Program - Individual
Bénéficiaire
Emplacement
British Columbia, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
NSERC's Awards Data
Tel que publié
Application id
RGPIN-2023-03656
Name nom
Wang, Shawn
Department département
Department of Mathematics
Organization id
2
Institution établissement
The University of British Columbia
Province en
British Columbia
Province fr
Colombie-Britannique
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2023
Competition year année de concours
2023
Award amount
37000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1508
Committee name en
Mathematics and Statistics
Committee name fr
Mathématiques et statistique
Area of application code
1200
Area of application group en
Advancement of knowledge
Area of application group fr
Avancement des connaissances
Area of application en
Advancement of knowledge
Area of application fr
Avancement des connaissances
Research subject code
2956
Research subject group en
Applied mathematics
Research subject group fr
Optimisation and optimal control theory
Research subject en
Mathématiques appliquées
Research subject fr
Optimisation et théories de commande optimale
Application title
New horizons in nonconvex and nonsmooth optimization: regularity theory and splitting algorithms
Keyword
convergence rates; iteration methods; local convergence; metric regularity; nonconvex functions; nonsmooth functions; optimization; problems from communication or machine learning; Variational analysis; variational convexity
Application summary
Optimization plays a critical role in Canada's economy and society. Over the past decade, nonconvex and nonsmooth optimization problems have become more and more popular in statistics, deep learning or neural networks, and communication theory in engineering, etc. To solve these challenging optimization problems, theories and algorithms from convex analysis do not apply because the functions involved are not convex. Techniques from differentiable optimization do not apply neither because the functions are not differentiable. New research is needed pressingly to advance our understanding of nonconvex and nonsmooth optimization. From the point of view of iteration methods, almost all splitting algorithms need both operators to be monotone. It is not clear what will happens if one operator is not monotone. From the point of view of theory, proximal mappings of nonconvex functions are set-valued. This is a huge challenge for convergence analysis of iteration methods. Moving from convexity to nonconvexity, from monotone mappings to nonmonotone mappings, is intrinsically hard and often requires new methodologies and techniques. Although some progress has been made, they are ad hoc and not satisfactory. A comprehensive investigation is needed. The goal of this proposal is to study variational monotone operators and variational convex functions, and to develop convergence theory and algorithms for solving inclusions in which one operator is monotone and the other is not necessarily monotone, or for solving optimization problems in which one function is lower semicontinuous and the other is only differentiable with or without a Lipschitz gradient. The long-term goal is to establish a systematic theory on local convergence analysis of many splitting algorithms, such as the forward-backward, Douglas-Rachford, and Spingarn partial inverse methods involving possibly nonconvex and nonsmooth functions or mappings without monotonicity. Recently, with his deep insights, Rockafellar made a groundbreaking advance in studying proximal point algorithms for nonconvex functions, namely variational convexity. This discovery can beautifully extend some key results from convex optimization to nonconvex optimization. I will carry out a systematic study of variational convex functions and explore how to use them in splitting algorithms effectively. Variational convexity, variational monotonicity, metric regularity, etc., will be used to study convergence rates. Developing generalized and local nonexpansive mappings will be at the heart of the convergence theory. Seeing the success of each of my HQP brings me joy, and this program will provide opportunities to train a diverse group of HQP. While postdoctoral fellows and graduate students can concentrate on theory developments, undergraduates can do some numerical experiments. This research will profoundly advance our knowledge and produce efficient algorithms for solving nonconvex and nonsmooth optimization problems.