Splitting Techniques in the Face of Huge Problem Sizes: Block-Coordinate and Block-Iterative Approaches

Similar documents
In collaboration with J.-C. Pesquet A. Repetti EC (UPE) IFPEN 16 Dec / 29

A Parallel Block-Coordinate Approach for Primal-Dual Splitting with Arbitrary Random Block Selection

A Dykstra-like algorithm for two monotone operators

About Split Proximal Algorithms for the Q-Lasso

A NEW ITERATIVE METHOD FOR THE SPLIT COMMON FIXED POINT PROBLEM IN HILBERT SPACES. Fenghui Wang

Splitting methods for decomposing separable convex programs

FINDING BEST APPROXIMATION PAIRS RELATIVE TO TWO CLOSED CONVEX SETS IN HILBERT SPACES

arxiv: v1 [math.oc] 20 Jun 2014

Proximal splitting methods on convex problems with a quadratic term: Relax!

Operator Splitting for Parallel and Distributed Optimization

Convergence of Fixed-Point Iterations

Learning with stochastic proximal gradient

On the order of the operators in the Douglas Rachford algorithm

Iterative Convex Optimization Algorithms; Part One: Using the Baillon Haddad Theorem

A generalized forward-backward method for solving split equality quasi inclusion problems in Banach spaces

ADMM for monotone operators: convergence analysis and rates

Dual and primal-dual methods

Coordinate Update Algorithm Short Course Operator Splitting

Canadian Mathematical Society Société mathématique du Canada

Extensions of the CQ Algorithm for the Split Feasibility and Split Equality Problems

Visco-penalization of the sum of two monotone operators

Primal-dual coordinate descent

1 Introduction and preliminaries

Tight Rates and Equivalence Results of Operator Splitting Schemes

Inertial Douglas-Rachford splitting for monotone inclusion problems

PARALLEL SUBGRADIENT METHOD FOR NONSMOOTH CONVEX OPTIMIZATION WITH A SIMPLE CONSTRAINT

Convergence Rates for Projective Splitting

A Primal-dual Three-operator Splitting Scheme

Fonctions Perspectives et Statistique en Grande Dimension

Primal-dual coordinate descent A Coordinate Descent Primal-Dual Algorithm with Large Step Size and Possibly Non-Separable Functions

A Relaxed Explicit Extragradient-Like Method for Solving Generalized Mixed Equilibria, Variational Inequalities and Constrained Convex Minimization

A General Iterative Method for Constrained Convex Minimization Problems in Hilbert Spaces

Variational inequalities for fixed point problems of quasi-nonexpansive operators 1. Rafał Zalas 2

Approaching monotone inclusion problems via second order dynamical systems with linear and anisotropic damping

Primal-dual algorithms for the sum of two and three functions 1

Douglas-Rachford; a very versatile algorithm

Compressed Sensing: a Subgradient Descent Method for Missing Data Problems

arxiv: v1 [math.oc] 12 Mar 2013

Convex Feasibility Problems

Convergence analysis for a primal-dual monotone + skew splitting algorithm with applications to total variation minimization

Dual Proximal Gradient Method

Solving monotone inclusions involving parallel sums of linearly composed maximally monotone operators

I P IANO : I NERTIAL P ROXIMAL A LGORITHM FOR N ON -C ONVEX O PTIMIZATION

On the convergence rate of a forward-backward type primal-dual splitting algorithm for convex optimization problems

Iterative common solutions of fixed point and variational inequality problems

arxiv: v3 [math.oc] 18 Apr 2012

A splitting algorithm for coupled system of primal dual monotone inclusions

Nonconvex notions of regularity and convergence of fundamental algorithms for feasibility problems

EE 546, Univ of Washington, Spring Proximal mapping. introduction. review of conjugate functions. proximal mapping. Proximal mapping 6 1

Math 273a: Optimization Overview of First-Order Optimization Algorithms

Self-dual Smooth Approximations of Convex Functions via the Proximal Average

A New Primal Dual Algorithm for Minimizing the Sum of Three Functions with a Linear Operator

arxiv: v4 [math.oc] 29 Jan 2018

Analysis of the convergence rate for the cyclic projection algorithm applied to semi-algebraic convex sets

ON A HYBRID PROXIMAL POINT ALGORITHM IN BANACH SPACES

Convergence rate estimates for the gradient differential inclusion

Research Article On an Iterative Method for Finding a Zero to the Sum of Two Maximal Monotone Operators

On the split equality common fixed point problem for quasi-nonexpansive multi-valued mappings in Banach spaces

1. Standing assumptions, problem statement, and motivation. We assume throughout this paper that

Proximal methods. S. Villa. October 7, 2014

Convergence Theorems for Bregman Strongly Nonexpansive Mappings in Reflexive Banach Spaces

GENERAL NONCONVEX SPLIT VARIATIONAL INEQUALITY PROBLEMS. Jong Kyu Kim, Salahuddin, and Won Hee Lim

Convergence rate analysis for averaged fixed point iterations in the presence of Hölder regularity

A New Modified Gradient-Projection Algorithm for Solution of Constrained Convex Minimization Problem in Hilbert Spaces

A General Framework for a Class of Primal-Dual Algorithms for TV Minimization

THE CYCLIC DOUGLAS RACHFORD METHOD FOR INCONSISTENT FEASIBILITY PROBLEMS

Signal Processing and Networks Optimization Part VI: Duality

On a result of Pazy concerning the asymptotic behaviour of nonexpansive mappings

Hybrid Steepest Descent Method for Variational Inequality Problem over Fixed Point Sets of Certain Quasi-Nonexpansive Mappings

ARock: an algorithmic framework for asynchronous parallel coordinate updates

Victoria Martín-Márquez

A GENERALIZATION OF THE REGULARIZATION PROXIMAL POINT METHOD

M. Marques Alves Marina Geremia. November 30, 2017

The Split Hierarchical Monotone Variational Inclusions Problems and Fixed Point Problems for Nonexpansive Semigroup

CONSTRUCTION OF BEST BREGMAN APPROXIMATIONS IN REFLEXIVE BANACH SPACES

Convex Optimization. (EE227A: UC Berkeley) Lecture 15. Suvrit Sra. (Gradient methods III) 12 March, 2013

On the equivalence of the primal-dual hybrid gradient method and Douglas Rachford splitting

On Slater s condition and finite convergence of the Douglas Rachford algorithm for solving convex feasibility problems in Euclidean spaces

Proximal Methods for Optimization with Spasity-inducing Norms

INERTIAL ACCELERATED ALGORITHMS FOR SOLVING SPLIT FEASIBILITY PROBLEMS. Yazheng Dang. Jie Sun. Honglei Xu

An Overview of Recent and Brand New Primal-Dual Methods for Solving Convex Optimization Problems

WEAK CONVERGENCE THEOREMS FOR EQUILIBRIUM PROBLEMS WITH NONLINEAR OPERATORS IN HILBERT SPACES

A primal dual Splitting Method for Convex. Optimization Involving Lipschitzian, Proximable and Linear Composite Terms,

ARock: an Algorithmic Framework for Asynchronous Parallel Coordinate Updates

A splitting minimization method on geodesic spaces

Monotone variational inequalities, generalized equilibrium problems and fixed point methods

ADMM and Fast Gradient Methods for Distributed Optimization

Variable Metric Forward-Backward Algorithm

Convex Optimization Notes

You should be able to...

Second order forward-backward dynamical systems for monotone inclusion problems

Heinz H. Bauschke and Walaa M. Moursi. December 1, Abstract

Extrapolation algorithm for affine-convex feasibility problems

Abstract In this article, we consider monotone inclusions in real Hilbert spaces

Master 2 MathBigData. 3 novembre CMAP - Ecole Polytechnique

Distributed Optimization via Alternating Direction Method of Multipliers

Neighboring feasible trajectories in infinite dimension

Conductivity imaging from one interior measurement

Weak and strong convergence theorems of modified SP-iterations for generalized asymptotically quasi-nonexpansive mappings

Transcription:

Splitting Techniques in the Face of Huge Problem Sizes: Block-Coordinate and Block-Iterative Approaches Patrick L. Combettes joint work with J.-C. Pesquet) Laboratoire Jacques-Louis Lions Faculté de Mathématiques Université Pierre et Marie Curie Paris 6 75005 Paris, France Edinburgh, May 7, 2015 Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 1/25

2/25 Framework A wide range of problems in applied nonlinear analysis can be reduced to finding a point in a closed convex subset F of a Hilbert space H. The solution set F is often constructed from prior information and the observation of data. Mathematical model: find x F n N Fix T n, T n : H H quasi-nonexpansive Algorithmic model: x n+1 = T n x n Asymptotic analysis: under suitable assumptions, x n ) n N converges weakly/strongly/linearly) to a point in F Application areas: variational inequalities, game theory, optimization, statistics, partial differential equations, inverse problems, mechanics, signal and image processing, machine learning, computer vision, transport theory, optics,... Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 2/25

2/25 Basic convergence principle T n is quasi-nonexpansive, i.e., with F Fix T n. Then: x H) y Fix T n ) T n x y x y, Fejér monotonicity: n N) y F) x n+1 y x n y Suppose that the set Wx n ) n N of weak cluster points of x n ) n N is in F. Then x n x F Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 3/25

2/25 Basic convergence principle T n is quasi-nonexpansive, i.e., with F Fix T n. Then: x H) y Fix T n ) T n x y x y, Fejér monotonicity: n N) y F) x n+1 y x n y Suppose that the set Wx n ) n N of weak cluster points of x n ) n N is in F. Then x n x F Elementary example: alternating projection method for finding a point in the intersection of two closed convex sets C 1 and C 2. Set T 2n = P 2 and T 2n+1 = P 1. Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 3/25

2/25 Example: Special cases Fixed point methods: Krasnosel skĭı Mann, string averaging, extrapolated barycentric methods, Martinet s cyclic firmly nonexpansive iteration, etc. Projections methods for convex feasibility problems Projections methods for split feasibility problems Subgradient projections methods for systems of convex inequalities Splitting methods for monotone inclusions: forward-backward, Douglas-Rachford, forward-backward-forward, Spingarn, etc. Convex optimization methods: projected gradient method, augmented Lagrangian method, ADMM, various proximal splitting methods, etc. Iterative methods for variational inequalites in mechanics, traffic theory, and finance etc. Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 4/25

2/25 Very large scale problems I: Block-coordinate approach The basic iteration x n+1 = T n x n in the Hilbert space H may be too involved computations, memory) to be operational We assume that H can be decomposed in m factors H = H 1 H m in which each T n has an explicit decomposition T n : x T 1,n x,...,t m,n x) The strategy is to update only arbitrarily chosen coordinates of x n+1 = x 1,n+1,...,x m,n+1 ) up to some tolerance: x i,n+1 = x i,n +ε i,n Ti,n x n + a i,n x i,n ), where ε i,n {0, 1} activation variable) Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 5/25

2/25 Very large scale problems I: Block-coordinate approach The basic iteration x n+1 = T n x n in the Hilbert space H may be too involved computations, memory) to be operational We assume that H can be decomposed in m factors H = H 1 H m in which each T n has an explicit decomposition T n : x T 1,n x,...,t m,n x) The strategy is to update only arbitrarily chosen coordinates of x n+1 = x 1,n+1,...,x m,n+1 ) up to some tolerance: x i,n+1 = x i,n +ε i,n Ti,n x n + a i,n x i,n ), where ε i,n {0, 1} activation variable) Our goal is to extend available fixed points methods to this block-coordinate setting while preserving their convergence properties Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 5/25

2/25 A roadblock The nice properties of an operator T: H H are destroyed by coordinate sampling For instance, consider the nonexpansive 1-Lipschitz) operator T: x 1, x 2 ) x 2, x 1 ) Then Q 1 : x 1, x 2 ) x 1, x 1 ) Q 2 : x 1, x 2 ) x 2, x 2 ) Q 1 Q 2 Q 2 Q 1 are no longer nonexpansive Fejér monotonicity is destroyed Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 6/25

2/25 A roadblock The nice properties of an operator T: H H are destroyed by coordinate sampling For instance, consider the nonexpansive 1-Lipschitz) operator T: x 1, x 2 ) x 2, x 1 ) Then Q 1 : x 1, x 2 ) x 1, x 1 ) Q 2 : x 1, x 2 ) x 2, x 2 ) Q 1 Q 2 Q 2 Q 1 are no longer nonexpansive Fejér monotonicity is destroyed.. and even for jointly convex functions with a unique minimizer, alternating minimizations fail, etc. Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 6/25

2/25 A roadblock The nice properties of an operator T: H H are destroyed by coordinate sampling For instance, consider the nonexpansive 1-Lipschitz) operator T: x 1, x 2 ) x 2, x 1 ) Then Q 1 : x 1, x 2 ) x 1, x 1 ) Q 2 : x 1, x 2 ) x 2, x 2 ) Q 1 Q 2 Q 2 Q 1 are no longer nonexpansive Fejér monotonicity is destroyed.. and even for jointly convex functions with a unique minimizer, alternating minimizations fail, etc. introduce stochasticity and renorming Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 6/25

Notation H: separable real Hilbert space; : weak convergence Wx n ) n N : set of weak cluster points of x n ) n N H N Γ 0 H): proper lower semicontinuous convex functions from H to ],+ ] Ω, F, P): underlying probability space Given a sequence x n ) n N of H-valued random variables, X = X n ) n N, where n N) X n = σx 0,...,x n ) l + X ): set of sequences of [0,+ [-valued random variables ξ n ) n N such that, for every n N, ξ n is X n -measurable { l p } +X )= ξ n ) n N l + X ) ξn p < + P-a.s., p ]0,+ [ l + {ξ X ) = n ) n N l + X ) n N sup n N } ξ n < + P-a.s. Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 7/25

Stochastic quasi-fejér sequences Ø F H, φ: [0,+ [ [0,+ [, φt) + as t + Deterministic definition: A sequence x n ) n N in H is Fejér monotone with respect to F if for every z F, n N) φ x n+1 z ) φ x n z ) Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 8/25

Stochastic quasi-fejér sequences Ø F H, φ: [0,+ [ [0,+ [, φt) + as t + Stochastic definition 1: A sequence x n ) n N of H-valued random variables is stochastically Fejér monotone with respect to F if, for every z F, n N) Eφ x n+1 z X n ) φ x n z ) Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 8/25

Stochastic quasi-fejér sequences Ø F H, φ: [0,+ [ [0,+ [, φt) + as t + Stochastic definition 2: A sequence x n ) n N of H-valued random variables is stochastically quasi-fejér monotone with respect to F if, for every z F, there exist χ n z)) n N l 1 + X ), ϑ n z)) n N l + X ), and η n z)) n N l 1 +X ) such that n N) Eφ x n+1 z ) X n )+ϑ n z) 1+χ n z))φ x n z )+η n z) Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 8/25

Stochastic quasi-fejér sequences Ø F H, φ: [0,+ [ [0,+ [, φt) + as t + Stochastic definition 2: A sequence x n ) n N of H-valued random variables is stochastically quasi-fejér monotone with respect to F if, for every z F, there exist χ n z)) n N l 1 + X ), ϑ n z)) n N l + X ), and η n z)) n N l 1 +X ) such that Theorem n N) Eφ x n+1 z ) X n )+ϑ n z) 1+χ n z))φ x n z )+η n z) Suppose x n ) n N is stochastically quasi-fejér monotone. Then z F) [ n N ϑ nz) < + P-a.s. ] [Wx n ) n N F P-a.s.] [x n ) n N converges weakly P-a.s. to an F-valued random variable] Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 8/25

An abstract stochastic iterative scheme Theorem Let Ø F H. Suppose that x n ) n N, t n ) n N, and e n ) n N are sequences of H-valued random variables such that: n N) x n+1 = x n +λ n t n + e n x n ), λ n ]0, 1] n N λ n E en 2 X n ) < + P-a.s. For every z F, there exist θ n z)) n N l + X ), µ n z)) n N l + X ), and ν n z)) n N l + X ) such that λ n µ n z)) n N l 1 +X ), λ n ν n z)) n N l 1/2 + X ), and n N) E t n z 2 X n )+θ n z) 1+µ n z)) x n z 2 +ν n z) Then z F) [ n N λ nθ n z) < + P-a.s. ] and [Wx n ) n N F P-a.s.] x n ) n N converges weakly P-a.s. to an F-valued random variable. Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 9/25

Single-layer algorithm H = H 1 H m, H i ) 1 i m separable real Hilbert spaces T n : H H: x T i,n x) 1 i m quasinonexpansive F = n N Fix T n Ø x 0 and the errors a n ) n N are H-valued random variables ε n ) n N identically distributed D-valued random variables with D = {0, 1} m {0} Algorithm: for n = 0, 1,... for i = 1,...,m xi,n+1 = x i,n +ε i,n λ n Ti,n x 1,n,...,x m,n )+a i,n x i,n ) Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 10/25

Single-layer algorithm Theorem Set n N) X n = σx 0,...,x n ) and E n = σε n ). Assume that inf n N λ n > 0 and sup n N λ n < 1. n N E an 2 X n ) < +. Wx n ) n N F P-a.s. For every n N, E n and X n are independent. For every i {1,...,m}, p i = P[ε i,0 = 1] > 0. Then x n ) n N converges weakly P-a.s. to an F-valued r.v. Proof. We achieve stochastic quasi-fejér monotonicity w.r.t. the norm x 2 = m i=1 x i 2 /p i Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 11/25

Example: Krasnosel skĭı Mann iteration T: H H: x T i x) 1 i m nonexpansive operator F = Fix T Ø x 0 and the errors a n ) n N are H-valued random variables ε n ) n N identically distributed D-valued random variables with D = {0, 1} m {0} Algorithm: for n = 0, 1,... for i = 1,...,m xi,n+1 = x i,n +ε i,n λ n Ti x 1,n,...,x m,n )+a i,n x i,n ) Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 12/25

Example: Krasnosel skĭı Mann iteration Theorem Set n N) X n = σx 0,...,x n ) and E n = σε n ). Assume that inf n N λ n > 0 and sup n N λ n < 1 n N E an 2 X n ) < + For every n N, E n and X n are independent For every i {1,...,m}, P[ε i,0 = 1] > 0 Then x n ) n N converges weakly P-a.s. to an F-valued r.v. Proof. Apply the single-layer theorem with T n = T. Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 13/25

Example: Block-coordinate, primal-dual splitting of coupled composite monotone inclusions Let F be the set of solutions to the problem find x 1 H 1,...,x m H m such that i {1,...,m}) 0 A i x i + p L ki B m k L kj x j ), where each A i : H i 2 H i, Bk : G k 2 G k are maximally monotone, L ki : H i G k linear&bounded k=1 Let F be the set of solutions to the dual problem find v 1 G 1,...,v p G p such that m p ) k {1,...,p}) 0 L ki A 1 i L li v l i=1 l=1 j=1 + B 1 k v k Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 14/25

Example: Block-coordinate, primal-dual splitting of coupled composite monotone inclusions Let Q j 1 j m+p) be the jth component of the projector P V onto the subspace { m } V = x, y) H G k {1,...,p}) y k = L ki x i Algorithm: γ ]0,+ [ and for n = 0, 1,... µ n ]0, 2[ for i = 1,...,m zi,n+1 = z i,n +ε i,n Qi x 1,n,...,x m,n, y 1,n,...,y p,n )+c i,n z i,n ) x i,n+1 = x i,n +ε i,n µ n JγAi 2z i,n+1 x i,n )+a i,n z i,n+1 ) for k = 1,...,p wk,n+1 = w k,n +ε m+k,n Qm+k x 1,n,...,x m,n, y 1,n,...,y p,n )+d k,n w k,n y k,n+1 = y k,n +ε m+k,n µ n JγBk 2w k,n+1 y k,n )+b k,n w k,n+1 ). i=1 Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 15/25

Example: Block-coordinate, primal-dual splitting of coupled composite monotone inclusions Under the same conditions as before: z n ) n N converges weakly P-a.s. to an F-valued random variable γ 1 w n y n )) n N converges weakly P-a.s. to an F - valued random variable Proof: This relies on the single-layer theorem and a nonstandard implementation of the Douglas-Rachford algorithm in the product space H G = H 1 H m G 1 G p : solve 0, 0) Ax By+N V x, y), where A: H 2 H : x m i=1 A ix i B: G 2 G : x p { k=1b k y k V = x, y) H G k {1,...,p}) y k = } m i=1 L kix i Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 16/25

Example: Nonsmooth, block-coordinate, primal-dual multivariate minimization Let F be the set of solutions to the problem minimize x 1 H 1,...,x m H m m f i x i )+ i=1 p m ) g k L ki x i where f i Γ 0 H i ), g k Γ 0 G k ), L ki : H i G k linear&bounded k=1 Let F be the set of solutions to the dual problem minimize v 1 G 1,...,v p G p m i=1 f i k=1 i=1 p ) L ki v k + p g k v k) k=1 Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 17/25

Example: Nonsmooth, block-coordinate, primal-dual multivariate minimization Algorithm: γ ]0,+ [ and for n = 0, 1,... µ n ]0, 2[ for i = 1,...,m zi,n+1 = z i,n +ε i,n Qi x 1,n,...,x m,n, y 1,n,...,y p,n )+c i,n z i,n ) x i,n+1 = x i,n +ε i,n µ n proxγfi 2z i,n+1 x i,n )+a i,n z i,n+1 ) for k = 1,...,p wk,n+1 = w k,n +ε m+k,n Qm+k x 1,n,...,x m,n, y 1,n,...,y p,n )+d k,n w k,n y k,n+1 = y k,n +ε m+k,n µ n proxγgk 2w k,n+1 y k,n )+b k,n w k,n+1 ). Under the same conditions as before: z n ) n N converges weakly P-a.s. to an F-valued random variable γ 1 w n y n )) n N converges weakly P-a.s. to an F -valued random variable Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 18/25

Double-layer random block-coordinate algorithms T n : H H: x T i,n x) 1 i m is α n -averaged Id+α 1 n T n Id) is nonexpansive) R n : H H is β n -averaged F = n N Fix T n R n Ø x 0, a n ) n N, and b n ) n N are H-valued random variables ε n ) n N identically distributed D-valued random variables with D = {0, 1} m {0} Algorithm: for n = 0, 1,... y n = R n x n + b n for i = 1,...,m ) xi,n+1 = x i,n +ε i,n Ti,n y n + a i,n x i,n Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 19/25

Double-layer random block-coordinate algorithms Theorem Set n N) X n = σx 0,...,x n ) and E n = σε n ). Assume that n N E an 2 X n ) < +, n N E bn 2 X n ) < + Wx n ) n N F P-a.s. For every n N, E n and X n are independent For every i {1,...,m}, p i = P[ε i,0 = 1] > 0 Then x n ) n N converges weakly P-a.s. to an F-valued r.v. Proof. We achieve stochastic quasi-fejér monotonicity w.r.t. the norm x 2 = m i=1 x i 2 /p i Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 20/25

Block-coordinate forward-backward splitting The forward-backward splitting algorithm is important because: It models many problems of interest and tolerates errors: PLC and Wajs, Signal recovery by proximal forward-backward splitting, Multiscale Model. Simul., vol. 4, 2005. Applied to the dual problem of a strongly monotone/convex composite problem, it provides a primal-dual algorithm: PLC, Dung, and Vũ, Dualization of signal recovery problems, Set-Valued Var. Anal., vol. 18, 2010. Applied in a renormed product space, it covers/extends various methods e.g., Chambolle-Pock): Vũ, A splitting algorithm for dual monotone inclusions involving cocoercive operators, Adv. Comput. Math., vol. 38, 2013. It can be implemented with variable metrics: PLC and Vũ, Variable metric forward-backward splitting with applications to monotone inclusions in duality, Optimization, vol. 63, 2014. In minimization problems it provides Fejér-monotonicity, convergent sequences, and monotone minimizing sequences. Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 21/25

Block-coordinate forward-backward splitting Let F be the set of solutions to the problem find x 1 H 1,...,x m H m such that i {1,...,m}) 0 A i x i + B i x 1,...,x m ) where A i : H i 2 H i is maximally monotone and, for some ϑ > 0, m m x H) y H) x i y i B i x B i y ϑ B i x B i y 2 Algorithm: i=1 for n = 0, 1,... ε γ n 2 ε)ϑ for i = 1,...,m xi,n+1 = x i,n +ε i,n JγnA i xi,n γ n B i x 1,n,...,x m,n )+c i,n ) ) ) +a i,n x i,n i=1 Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 22/25

Block-coordinate forward-backward splitting Under the same conditions as before almost sure weak convergence of x n ) n N to a point in F is achieved. Proof: In double-layer theorem, set A: H 2 H : x m i=1 A ix i B: H H: x B i x) 1 i m T n = J γna R n = Id γ n B, F = zera+b) b n = γ n c n α n = 1/2 β n = γ n /2ϑ) Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 23/25

Block-coordinate forward-backward splitting: convex minimization Let F be the set of solutions to the problem minimize x 1 H 1,...,x m H m m p m ) f i x i )+ g k L ki x i i=1 where f i Γ 0 H i ), g k : G k R differentiable, convex, g k τ k -Lipschitz, L ki : H i G k linear&bounded Let Algorithm: k=1 i=1 k=1 p m ϑ = τ k L ki L ki ) 1 for n = 0, 1,... for i = 1,...,m p ri,n = ε i,n xi,n γ n k=1 L ki g m k j=1 L ) )) kjx j,n + ci,n ) x i,n+1 = x i,n +ε i,n λ n proxγnf i r i,n + a i,n x i,n. i=1 Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 24/25

References PLC and J.-C. Pesquet, Stochastic quasi-fejér block-coordinate fixed point iterations with random sweeping, SIAM J. Optim., 2015 arxiv, April 2014) H. H. Bauschke and PLC, Convex Analysis and Monotone Operator Theory in Hilbert Spaces. Springer, New York, 2011. 2nd ed., Fall 2015) Patrick L. Combettes joint work with J.-C. Pesquet) Block-coordinate splitting algorithms 25/25