Skip to main content

Parallel computing subgradient method for nonsmooth convex optimization over the intersection of fixed point sets of nonexpansive mappings

Abstract

Nonsmooth convex optimization problems are solved over fixed point sets of nonexpansive mappings by using a distributed optimization technique. This is done for a networked system with an operator, who manages the system, and a finite number of users, by solving the problem of minimizing the sum of the operator’s and users’ nondifferentiable, convex objective functions over the intersection of the operator’s and users’ convex constraint sets in a real Hilbert space. We assume that each of their constraint sets can be expressed as the fixed point set of an implementable nonexpansive mapping. This setting allows us to discuss nonsmooth convex optimization problems in which the metric projection onto the constraint set cannot be calculated explicitly. We propose a parallel subgradient algorithm for solving the problem by using the operator’s attribution such that it can communicate with all users. The proposed algorithm does not use any proximity operators, in contrast to conventional parallel algorithms for nonsmooth convex optimization. We first study its convergence property for a constant step-size rule. The analysis indicates that the proposed algorithm with a small constant step size approximates a solution to the problem. We next consider the case of a diminishing step-size sequence and prove that there exists a subsequence of the sequence generated by the algorithm which weakly converges to a solution to the problem. We also give numerical examples to support the convergence analyses.

1 Introduction

Convex optimization theory has been widely used to solve practical convex minimization problems over complicated constraints, e.g., convex optimization problems with a fixed point constraint [18] and with a variational inequality constraint [913]. It enables us to consider constrained optimization problems in which the explicit form of the metric projection onto the constraint set is not always known; i.e., the constraint set is not simple in the sense that the projection cannot easily be computed (e.g., the constraint set is the set of all minimizers of a convex function over a closed convex set [7, 14], or the set of zeros of a set-valued, monotone operator ([15], Proposition 23.38)).

This paper focuses on a networked system consisting of an operator, who manages the system, and a finite number of participating users, and it considers the problem of minimizing the sum of the operator’s and all users’ nonsmooth convex functions over the intersection of the operator’s and all users’ fixed point constraint sets in a real Hilbert space.

The motivations behind studying the problem are to devise optimization algorithms which have a wider range of application compared with the previous algorithms for smooth convex optimization (see, e.g., [1, 3, 5, 7]) and to tackle outstanding nonsmooth convex problems over complicated constraint sets (e.g., the minimal antenna-subset selection problem ([16], Section 17.4)).

Many algorithms have been presented for solving nonsmooth convex optimization. The Douglas-Rachford algorithm ([15], Chapters 25 and 27), [1720], forward-backward algorithm ([15], Chapters 25 and 27), [18, 21, 22], and parallel proximal algorithm ([15], Proposition 27.8), ([18], Algorithm 10.27), [23] are useful to solve the sum of nonsmooth convex optimization problems over the whole space. They use the proximity operators ([15], Definition 12.23) of nonsmooth, convex functions. The incremental subgradient method ([24], Section 8.2) and the projected multi-agent algorithms [2528] can minimize the sum of nonsmooth, convex functions over a simple constraint set by using the subgradients ([29], Section 23) of the nonsmooth, convex functions instead of the proximity operators. To our knowledge, there are no references on parallel algorithms for nonsmooth convex optimization with fixed point constraints.

In this paper, we propose a parallel subgradient algorithm for nonsmooth convex optimization with fixed point constraints. Our algorithm is founded on the ideas behind the two useful algorithms. The first is the Krasnosel’skiĭ-Mann algorithm ([15], Subchapter 5.2), [30, 31] for finding a fixed point of a nonexpansive mapping. It ensures that our algorithm converges to a point in the intersection of the fixed point sets of nonexpansive mappings. The second algorithm is the parallel proximal algorithm ([15], Proposition 27.8), ([18], Algorithm 10.27), [23] for nonsmooth convex optimization. Since the operator can communicate with all users, our parallel algorithm enables the operator to find a solution to the main problem by using information transmitted from all users.

This paper has three contributions in relation to other work on convex optimization. The first is that our algorithm does not use any proximity operators, in contrast to the algorithms presented in [16, 18, 2123]. Our algorithm can use subgradients, which are well defined for any nonsmooth, convex functions.

The second contribution is that our parallel algorithm can be applied to nonsmooth convex optimization problems over the fixed point sets of nonexpansive mappings, while the previous algorithms work in nonsmooth convex optimization over simple constraint sets ([15], Subchapter 5.2), [18, 2123] or smooth convex optimization over fixed point sets [13, 5, 7].

The third contribution is to present convergence analyses for different step-size rules. We show that our algorithm with a small constant step size approximates a solution to the problem of minimizing the sum of nonsmooth, convex functions over the fixed point sets of nonexpansive mappings. We also show that there exists a subsequence of the sequence generated by our algorithm with a diminishing step size which weakly converges to a solution to the problem.

This paper is organized as follows. Section 2 gives the mathematical preliminaries and states the main problem. Section 3 presents the parallel subgradient algorithm for solving the main problem and studies its convergence properties for a constant step size and a diminishing step size. Section 4 provides numerical examples of the algorithm. Section 5 concludes the paper.

2 Mathematical preliminaries

2.1 Nonexpansivity and subdifferentiability

Let H be a real Hilbert space with inner product \(\langle\cdot, \cdot\rangle\) and its induced norm \(\| \cdot\|\). Let \(\mathbb{N}\) denote the set of all positive integers including zero.

A mapping, \(T \colon H \to H\), is said to be nonexpansive ([15], Definition 4.1(ii)) if \(\|T(x) - T(y)\| \leq\|x-y\|\) (\(x,y\in H\)). T is said to be firmly nonexpansive ([15], Definition 4.1(i)) if \(\| T(x) - T(y) \|^{2} + \|(\mathrm{Id} - T)(x) - (\mathrm{Id} - T)(y) \|^{2} \leq\|x-y\|^{2}\) (\(x,y\in H\)), where Id stands for the identity mapping on H. It is clear that firm nonexpansivity implies nonexpansivity. The fixed point set of T is denoted by \(\operatorname{Fix}(T) := \{ x\in H \colon T(x) = x \}\). The metric projection ([15], Subchapter 4.2, Chapter 28) onto a nonempty, closed convex set C (H) is denoted by \(P_{C}\). It is defined by \(P_{C}(x) \in C\) and \(\| x - P_{C} (x)\| = \inf_{y\in C} \|x-y\|\) (\(x\in H\)).

Proposition 2.1

Let \(T \colon H \to H\) be nonexpansive, and let C (H) be nonempty, closed, and convex. Then:

  1. (i)

    ([15], Corollary 4.15) \(\operatorname{Fix}(T)\) is closed and convex.

  2. (ii)

    ([15], Remark 4.24(iii)) \((1/2)(\mathrm {Id} + T)\) is firmly nonexpansive.

  3. (iii)

    ([15], Proposition 4.8, equation (4.8)) \(P_{C}\) is firmly nonexpansive with \(\operatorname{Fix}(P_{C}) = C\).

The subdifferential ([15], Definition 16.1), ([29], Section 23) of \(f \colon H \to\mathbb{R}\) is defined for all \(x\in H\) by

$$ \partial f (x) := \bigl\{ u\in H \colon f(y) \geq f(x) + \langle y-x,u \rangle\ ( y\in H ) \bigr\} . $$

We call u (\(\in\partial f (x)\)) the subgradient of f at \(x \in H\).

Proposition 2.2

([15], Proposition 16.14(ii) and (iii))

Let \(f \colon H \to\mathbb{R}\) be continuous and convex with \(\operatorname{dom}(f) := \{ x\in H \colon f(x) < \infty\}=H\). Then \(\partial f(x) \neq\emptyset\) (\(x\in H\)). Moreover, for all \(x\in H\), there exists \(\delta> 0\) such that \(\partial f(B(x;\delta))\) is bounded, where \(B(x;\delta)\) stands for a closed ball with center x and radius δ.

2.2 Notation, assumptions, and main problem

This paper deals with a networked system with an operator (denoted by user 0) and I users. Let

$$ \mathcal{I} := \{1,2,\ldots, I\} \quad \text{and} \quad \bar{\mathcal{I}} := \{0 \} \cup\mathcal{I}. $$

We assume that user i (\(i\in\bar{\mathcal{I}}\)) has its own private mappings, denoted by \(f^{(i)} \colon H \to\mathbb {R}\) and \(T^{(i)} \colon H \to H\), and its own private nonempty, closed convex constraint set, denoted by \(C^{(i)}\) (H). Moreover, we define

$$ X := \bigcap_{i\in\bar{\mathcal{I}}} \operatorname{Fix} \bigl( T^{(i)} \bigr),\qquad f := \sum_{i\in\bar{\mathcal{I}}} f^{(i)},\qquad X^{\star}:= \Bigl\{ x\in X \colon f(x) = f^{\star}:= \inf_{y\in X} f (y ) \Bigr\} . $$

The following problem is discussed.

Problem 2.1

Assume that:

  1. (A1)

    \(T^{(i)} \colon H \to H\) (\(i\in\bar{\mathcal{I}}\)) is firmly nonexpansive with \(\operatorname{Fix}(T^{(i)}) = C^{(i)}\).

  2. (A2)

    \(f^{(i)} \colon H \to\mathbb{R}\) (\(i\in\bar{\mathcal{I}}\)) is continuous and convex with \(\operatorname{dom}(f^{(i)}) = H\).

  3. (A3)

    User i (\(i\in\bar{\mathcal{I}}\)) can use its own private \(T^{(i)}\) and \(\partial f^{(i)}\).

  4. (A4)

    The operator can communicate with all users.

  5. (A5)

    \(X^{\star}\) is nonempty.

The main objective is to find \(x^{\star}\in X^{\star}\).

Assumption (A2) and Proposition 2.2 ensure that \(\partial f^{(i)}(x) \neq\emptyset\) (\(i\in\bar{\mathcal{I}}\), \(x\in H\)). Suppose that the operator sets \(\hat{x} \in H\). Accordingly, (A4) guarantees that the operator can transmit \(\hat{x}\) to all users. Assumption (A3) implies that user i (\(i\in\bar{\mathcal{I}}\)) can compute in parallel \(\hat{x}^{(i)} := \hat{x}^{(i)} (\hat{x}, T^{(i)}, \partial f^{(i)})\) by using the information \(\hat{x}\) transmitted from the operator and its own private information. Moreover, (A4) ensures that the operator has access to all \(\hat {x}^{(i)}\) and can compute \(\bar{x} := \bar{x}(\hat{x}^{(0)}, \hat{x}^{(1)},\ldots, \hat{x}^{(I)})\). The next section describes a sufficient condition for satisfying (A5).

3 Parallel subgradient algorithm for nonsmooth convex optimization over fixed point sets

This section presents a parallel subgradient algorithm for solving Problem 2.1.

Algorithm 3.1

Step 0.:

The operator (user 0) and all users set α (\(\in(0,1)\)) and \((\lambda_{n})_{n\in\mathbb{N}}\) (\(\subset(0,\infty)\)). The operator chooses \(x_{0} \in H\) arbitrarily and transmits it to all users.

Step 1.:

Given \(x_{n} \in H\), user i (\(i\in\bar{\mathcal{I}}\)) computes \(x_{n}^{(i)} \in H\) by

$$ \begin{cases} g_{n}^{ (i )} \in\partial f^{ (i )} ( x_{n} ), \\ x_{n}^{ ( i )} := \alpha x_{n} + ( 1 - \alpha ) T^{ ( i )} ( x_{n} - \lambda_{n} g_{n}^{ (i )} ). \end{cases} $$

User i (\(i\in\mathcal{I}\)) transmits \(x_{n}^{(i)}\) to the operator.

Step 2.:

The operator computes \(x_{n+1} \in H\) as

$$ x_{n+1} := \frac{1}{I+1} \sum_{i\in\bar{\mathcal{I}}} x_{n}^{(i)} $$

and transmits it to all users. Put \(n := n+1\), and go to Step 1.

Our convergence results depend on the following assumption.

Assumption 3.1

The sequence, \((x_{n}^{(i)})_{n\in\mathbb{N}}\) (\(i\in\bar{\mathcal{I}}\)), generated by Algorithm 3.1 is bounded.

We shall provide examples satisfying Assumption 3.1. User i (\(i\in\bar{\mathcal{I}}\)) in an actual network [2, 3234] has a bounded \(C^{(i)}\) defined by the intersection of simple, closed convex sets \(C_{k}^{(i)}\) (\(k \in\mathcal{K}^{(i)} := \{1,2,\ldots,K^{(i)}\}\)) (e.g., \(C_{k}^{(i)}\) is an affine subspace, a half-space, or a hyperslab) and \(P_{k}^{(i)} := P_{C_{k}^{(i)}}\) can easily be computed within a finite number of arithmetic operations [35], ([15], Chapter 28). Then user i can choose a bounded \(X^{(i)}\) (\(\supset C^{(i)}\)) such that \(P^{(i)} := P_{X^{(i)}}\) is easily computed (e.g., \(X^{(i)} = \operatorname{Fix}(P^{(i)})\) is a closed ball with a large enough radius). Since \(X^{(i)}\) is bounded and \(X \subset C^{(i)} \subset X^{(i)}\) (\(i\in\bar{\mathcal{I}}\)), X is also bounded. Hence, the continuity and convexity of f ensure that \(X^{\star}\neq\emptyset\), i.e., (A5) holds ([15], Proposition 11.14). In this case, user i can use

$$ T^{ (i )} := \frac{1}{2} \biggl[ \mathrm{Id} + \prod _{k\in\mathcal{K}^{(i)}} P_{k}^{ (i )} \biggr]\quad \text{with } \operatorname{Fix} \bigl( T^{ (i )} \bigr) = C^{ (i )} \subset X^{ (i )}. $$
(1)

Proposition 2.1(ii) and (iii) guarantee that \(T^{(i)}\) defined by (1) satisfies the firm nonexpansivity condition. Moreover, user i can compute

$$ x_{n}^{ ( i )} := P^{(i)} \bigl( \alpha x_{n} + ( 1 - \alpha ) T^{ ( i )} \bigl( x_{n} - \lambda_{n} g_{n}^{ (i )} \bigr) \bigr) $$
(2)

instead of \(x_{n}^{(i)}\) in Algorithm 3.1. Since \(X^{(i)}\) is bounded and \((x_{n}^{(i)})_{n\in\mathbb{N}} \subset X^{(i)}\), \((x_{n}^{(i)})_{n\in\mathbb{N}}\) is bounded. We can prove that Algorithm 3.1 with (2) satisfies the properties in the main theorems (Theorems 3.1 and 3.2) by referring to the proofs of the theorems.

The following lemma yields some properties of Algorithm 3.1 that will be used to prove the main theorems.

Lemma 3.1

Suppose that Assumptions (A1)-(A5) and 3.1 are satisfied, \(\limsup_{n\to\infty} \lambda_{n} < \infty\), and \(y_{n}^{(i)} := T^{(i)}( x_{n} - \lambda_{n} g_{n}^{(i)})\) (\(n\in\mathbb{N}\), \(i\in\bar {\mathcal{I}}\)). Then the following properties hold:

  1. (i)

    \((g_{n}^{(i)})_{n\in\mathbb{N}}\), \((y_{n}^{(i)})_{n\in\mathbb{N}}\) (\(i\in\bar{\mathcal{I}}\)), and \((x_{n})_{n\in\mathbb{N}}\) are bounded.

  2. (ii)

    For all \(x\in X\) and for all \(n\in\mathbb{N}\),

    $$ \Vert x_{n+1} - x \Vert ^{2} \leq \Vert x_{n} -x \Vert ^{2} + M_{1} \lambda_{n} - \frac{1-\alpha}{I+1} \sum_{i\in\bar{\mathcal{I}}} \bigl\Vert x_{n} - y_{n}^{ (i )} \bigr\Vert ^{2}, $$

    where \(M_{1} := \max_{i\in\bar{\mathcal{I}}} (\sup\{ 2 |\langle y_{n}^{(i)} - x, g_{n}^{(i)} \rangle| \colon n\in\mathbb{N} \}) < \infty\).

  3. (iii)

    For all \(x\in X\) and for all \(n\in\mathbb{N}\),

    $$ \Vert x_{n+1} - x \Vert ^{2} \leq \Vert x_{n} -x \Vert ^{2} + \frac{2(1-\alpha)\lambda_{n}}{I+1} \bigl( f(x) - f(x_{n}) \bigr) + M_{2} ( 1- \alpha ) \lambda_{n}^{2}, $$

    where \(M_{2} := \max_{i\in\bar{\mathcal{I}}} (\sup\{ \| g_{n}^{(i)} \| ^{2} \colon n\in\mathbb{N} \}) < \infty\).

Proof

(i) Assumption 3.1 and the definition of \(x_{n}\) (\(n\in\mathbb {N}\)) ensure the boundedness of \((x_{n})_{n\in\mathbb{N}}\). Hence, from (A2) and Proposition 2.2, we find that \((g_{n}^{(i)})_{n\in\mathbb{N}}\) (\(i\in\bar{\mathcal{I}}\)) is also bounded. Assumption (A1) implies that, for all \(x\in X\), for all \(n\in\mathbb {N}\), and for all \(i\in\bar{\mathcal{I}}\),

$$ \bigl\Vert y_{n}^{ (i )} - x \bigr\Vert = \bigl\Vert T^{(i)} \bigl( x_{n} - \lambda_{n} g_{n}^{(i)} \bigr) - T^{(i)} (x) \bigr\Vert \leq \bigl\Vert \bigl( x_{n} - \lambda_{n} g_{n}^{(i)} \bigr) - x \bigr\Vert . $$

Accordingly, the boundedness of \((x_{n})_{n\in\mathbb{N}}\) and \((g_{n}^{(i)})_{n\in\mathbb{N}}\) (\(i\in\bar{\mathcal{I}}\)) and \(\limsup_{n\to\infty} \lambda_{n} < \infty\) imply that \((y_{n}^{(i)})_{n\in\mathbb{N}}\) (\(i\in\bar {\mathcal{I}}\)) is also bounded.

(ii) Choose \(x\in X\) arbitrarily and put \(M_{1} := \max_{i\in\bar{\mathcal{I}}} ( \sup\{ 2 | \langle y_{n}^{(i)} - x, g_{n}^{(i)} \rangle| \colon n\in\mathbb{N} \} )\). Lemma 3.1(i) guarantees that \(M_{1} < \infty\). Assumption (A1) ensures that, for all \(n\in\mathbb{N}\) and for all \(i\in\bar{\mathcal{I}}\),

$$\begin{aligned} \begin{aligned} \bigl\Vert y_{n}^{ (i )} - x \bigr\Vert ^{2} &= \bigl\Vert T^{ ( i )} \bigl( x_{n} - \lambda_{n} g_{n}^{ (i )} \bigr) - T^{(i)} (x) \bigr\Vert ^{2} \\ &\leq\bigl\Vert \bigl( x_{n} - \lambda_{n} g_{n}^{ (i )} \bigr) - x \bigr\Vert ^{2} - \bigl\Vert \bigl( x_{n} - \lambda_{n} g_{n}^{ (i )} \bigr)- y_{n}^{ (i )} \bigr\Vert ^{2}, \end{aligned} \end{aligned}$$

which, together with \(\|x-y\|^{2} = \|x\|^{2} - 2 \langle x,y\rangle+ \|y\| ^{2}\) (\(x,y \in H\)), means that

$$\begin{aligned} \bigl\Vert y_{n}^{ (i )} - x \bigr\Vert ^{2} \leq&\Vert x_{n} - x \Vert ^{2} -2 \lambda_{n} \bigl\langle x_{n} - x, g_{n}^{ (i )} \bigr\rangle + \lambda_{n}^{2} \bigl\Vert g_{n}^{ (i )} \bigr\Vert ^{2} \\ &{} - \bigl\Vert x_{n} - y_{n}^{ (i )} \bigr\Vert ^{2} +2 \lambda_{n} \bigl\langle x_{n} - y_{n}^{ (i )}, g_{n}^{ (i )} \bigr\rangle - \lambda_{n}^{2} \bigl\Vert g_{n}^{ (i )} \bigr\Vert ^{2} \\ \leq&\Vert x_{n} - x \Vert ^{2} - \bigl\Vert x_{n} - y_{n}^{ (i )} \bigr\Vert ^{2} + M_{1} \lambda_{n}. \end{aligned}$$
(3)

The convexity of \(\| \cdot\|^{2}\) implies that, for all \(n\in\mathbb {N}\) and for all \(i\in\bar{\mathcal{I}}\),

$$\begin{aligned} \bigl\Vert x_{n}^{ (i )} - x \bigr\Vert ^{2} &= \bigl\Vert \alpha ( x_{n} -x ) + (1-\alpha) \bigl( y_{n}^{ (i )} - x \bigr) \bigr\Vert ^{2} \\ &\leq \alpha \Vert x_{n} -x \Vert ^{2} + (1-\alpha) \bigl\Vert y_{n}^{ (i )} - x \bigr\Vert ^{2}, \end{aligned}$$
(4)

which, together with (3), means that, for all \(n\in\mathbb {N}\) and for all \(i\in\bar{\mathcal{I}}\),

$$ \bigl\Vert x_{n}^{ (i )} - x \bigr\Vert ^{2} \leq \Vert x_{n} - x \Vert ^{2} - (1-\alpha) \bigl\Vert x_{n} - y_{n}^{ (i )} \bigr\Vert ^{2} + M_{1} \lambda_{n}. $$

Summing up this inequality over all i guarantees that, for all \(n\in \mathbb{N}\),

$$ \frac{1}{I+1}\sum_{i\in\bar{\mathcal{I}}} \bigl\Vert x_{n}^{ (i )} - x \bigr\Vert ^{2} \leq \Vert x_{n} - x \Vert ^{2} - \frac{1-\alpha}{I+1} \sum _{i\in\bar{\mathcal{I}}} \bigl\Vert x_{n} - y_{n}^{ (i )} \bigr\Vert ^{2} + M_{1} \lambda_{n}. $$

Accordingly, from the definition of \(x_{n}\) (\(n\in\mathbb{N}\)) and the convexity of \(\| \cdot\|^{2}\), we find that, for all \(n\in\mathbb{N}\),

$$\begin{aligned} \Vert x_{n+1} - x \Vert ^{2} &\leq\frac{1}{I+1} \sum _{i\in\bar{\mathcal{I}}} \bigl\Vert x_{n}^{ (i )} - x \bigr\Vert ^{2} \\ &\leq \Vert x_{n} - x \Vert ^{2} - \frac{1-\alpha}{I+1} \sum_{i\in\bar{\mathcal{I}}} \bigl\Vert x_{n} - y_{n}^{ (i )} \bigr\Vert ^{2} + M_{1} \lambda_{n}. \end{aligned}$$

(iii) Choose \(x\in X\) arbitrarily. Then (3) and the definition of \(g_{n}^{(i)}\) (\(n\in\mathbb {N}\), \(i \in\bar{\mathcal{I}}\)) imply that, for all \(n\in\mathbb{N}\) and for all \(i\in\bar{\mathcal{I}}\),

$$\begin{aligned} \bigl\Vert y_{n}^{ (i )} - x \bigr\Vert ^{2} & \leq \Vert x_{n} - x \Vert ^{2} + 2 \lambda_{n} \bigl\langle x - x_{n}, g_{n}^{ (i )} \bigr\rangle + \lambda_{n}^{2} \bigl\Vert g_{n}^{ (i )} \bigr\Vert ^{2} \\ &\leq \Vert x_{n} - x \Vert ^{2} + 2 \lambda_{n} \bigl( f^{ (i )} (x ) - f^{ (i )} (x_{n} ) \bigr) + M_{2} \lambda_{n}^{2}, \end{aligned}$$

where \(M_{2} := \max_{i\in\bar{\mathcal{I}}} (\sup\{ \| g_{n}^{(i)} \| ^{2} \colon n\in\mathbb{N} \}) < \infty\) (\(M_{2} < \infty\) is guaranteed by Lemma 3.1(i)). Accordingly, (4) guarantees that, for all \(n\in\mathbb {N}\) and for all \(i\in\bar{\mathcal{I}}\),

$$ \bigl\Vert x_{n}^{ (i )} - x \bigr\Vert ^{2} \leq \Vert x_{n} - x \Vert ^{2} + 2 (1 - \alpha) \lambda_{n} \bigl( f^{ (i )} (x ) - f^{ (i )} (x_{n} ) \bigr) + M_{2} ( 1 - \alpha ) \lambda_{n}^{2}, $$

which, together with the convexity of \(\| \cdot\|^{2}\) and \(f:= \sum_{i\in\bar{\mathcal{I}}} f^{(i)}\), implies that, for all \(n\in \mathbb{N}\),

$$\begin{aligned} \| x_{n+1} - x \|^{2} \leq&\frac{1}{I+1} \sum _{i\in\bar{\mathcal{I}}} \bigl\Vert x_{n}^{ (i )} - x \bigr\Vert ^{2} \\ \leq& \Vert x_{n} - x \Vert ^{2} + \frac{2 (1 - \alpha) \lambda_{n}}{I+1} \sum_{i\in\bar{\mathcal{I}}} \bigl( f^{ (i )} (x ) - f^{ (i )} (x_{n} ) \bigr) \\ &{} + M_{2} ( 1 -\alpha ) \lambda_{n}^{2} \\ =&\Vert x_{n} - x \Vert ^{2} + \frac{2 (1 - \alpha)\lambda_{n}}{I+1} \bigl( f (x ) - f (x_{n} ) \bigr) + M_{2} ( 1 -\alpha ) \lambda_{n}^{2}. \end{aligned}$$

This completes the proof. □

3.1 Constant step-size rule

The discussion in this subsection makes the following assumption.

Assumption 3.2

User i (\(i\in\bar{\mathcal{I}}\)) has \((\lambda_{n})_{n\in\mathbb{N}}\) satisfying

$$ (\mathrm{C}1) \quad \lambda_{n} := \lambda \in(0, \infty) \quad (n\in \mathbb{N}). $$

Let us perform a convergence analysis on Algorithm 3.1 under Assumption 3.2.

Theorem 3.1

Suppose that Assumptions (A1)-(A5), 3.1, and 3.2 hold. Then the sequence, \((x_{n})_{n\in\mathbb{N}}\), generated by Algorithm 3.1 satisfies, for all \(i\in\bar{\mathcal{I}}\),

$$ \liminf_{n\to\infty} \bigl\Vert x_{n} - T^{(i)} (x_{n} ) \bigr\Vert ^{2} \leq M \lambda \quad \textit{and} \quad \liminf_{n\to\infty} f (x_{n} ) \leq f^{\star}+ \frac { (I+1 ) M_{2} \lambda}{2}, $$

where \(M_{1}\) and \(M_{2}\) are constants defined as in Lemma  3.1, \(M_{3} := \max_{i\in\bar{\mathcal{I}}} (\sup\{ \| x_{n} - y_{n}^{(i)} \| \colon n\in\mathbb{N} \})\), and \(M := (I+1) M_{1}/(1-\alpha) + 2 M_{3} \sqrt{M_{2}} + M_{2} \lambda\).

Let us compare Algorithm 3.1 under the assumptions in Theorem 3.1 with previous algorithms ([24], Section 8.2), ([15], Chapters 25 and 27), [18, 2123]. The following sequence \(({x}_{n})_{n\in\mathbb{N}}\) is generated by a parallel proximal algorithm ([15], Chapters 25 and 27), [18, 21, 22] that can be applied to signal and image processing: given \((\lambda_{n})_{n\in\mathbb{N}} \subset(0,2)\), \({y}_{n}^{(i)} \in H\), \(({a}_{n}^{(i)})_{n\in\mathbb{N}} \subset H\) (\(i = 0,1,\ldots,m\)), and \({x}_{n} \in H\),

$$ \textstyle\begin{cases} {p}_{n}^{(i)} := \operatorname{prox}_{\gamma f^{(i)}/\omega^{(i)}} {y}_{n}^{(i)} + {a}_{n}^{(i)} \quad (i=0,1,\ldots,m), \\ {p}_{n} := \sum_{i=0}^{m} \omega^{(i)} {p}_{n}^{(i)}, \\ {y}_{n+1}^{(i)} := {y}_{n}^{(i)} + \lambda_{n} ( 2 {p}_{n} - {x}_{n} - {p}_{n}^{(i)} ) \quad (i=0,1,\ldots,m), \\ {x}_{n+1} := {x}_{n} + \lambda_{n} ( {p}_{n} - {x}_{n} ), \end{cases} $$
(5)

where \(\gamma\in(0,1)\), \((\omega^{(i)})_{i=0}^{m}\) (\(\subset(0,1)\)) satisfies \(\sum_{i=0}^{m} \omega^{(i)} = 1\), and \(\operatorname {prox}_{f^{(i)}}\) stands for the proximity operator of \(f^{(i)}\) which maps every \({x} \in H\) to the unique minimizer of \(f^{(i)} + (1/2)\| {x} - \cdot\|^{2}\). (See ([18], Tables 10.1 and 10.2) for examples of convex functions for which proximity operators can be explicitly computed.) When \((\lambda_{n})_{n\in\mathbb{N}}\) satisfies \(\sum_{n=0}^{\infty} \lambda_{n} (2 - \lambda_{n}) = \infty\) (e.g., \(\lambda_{n} := \lambda\in(0,2)\) (\(n\in\mathbb{N}\)) satisfies this condition) and \(\sum_{n=0}^{\infty} \lambda_{n} \| {a}_{n}^{(i)} \| < \infty\) (\(i=0,1,\ldots,m\)), \(({x}_{n})_{n\in\mathbb{N}}\) in algorithm (5) converges to a minimizer of \(\sum_{i=0}^{m} f^{(i)}\) over H ([22], Theorem 3.4).

Suppose that \(C^{(i)}\) (\(i\in\bar{\mathcal{I}}\)) is simple in the sense that \(P_{C^{(i)}}\) can easily be computed (e.g., \(C^{(i)}\) is an affine subspace, a half-space, or a hyperslab). Algorithm 3.1 with \(\lambda_{n} := \lambda\in(0,\infty)\) (\(n\in\mathbb{N}\)) and \(T^{(i)} = P_{C^{(i)}}\) (\(i\in\bar{\mathcal{I}}\)) is as follows: given \(g_{n}^{(i )} \in\partial f^{(i)} ( x_{n} )\) (\(i\in\bar{\mathcal{I}}\)),

$$ \textstyle\begin{cases} x_{n}^{ ( i )} := \alpha x_{n} + ( 1 - \alpha ) P_{C^{ (i )}} ( x_{n} - \lambda g_{n}^{ (i )} ) \quad (i=0,1,\ldots,I), \\ x_{n+1} := {\frac{1}{I+1} \sum_{i\in\bar{\mathcal {I}}} x_{n}^{(i)}}. \end{cases} $$
(6)

We can see that algorithm (6) uses the subgradient \(g_{n}^{(i)} \in\partial f^{(i)}(x_{n})\), while algorithm (5) uses the proximity operator of \(f^{(i)}\). Theorem 3.1 says that under the assumptions in Theorem 3.1 algorithm (6) satisfies, for all \(i\in\bar{\mathcal{I}}\),

$$ \liminf_{n\to\infty} \bigl\Vert x_{n} - P_{C^{ (i )}} (x_{n} ) \bigr\Vert ^{2} \leq M \lambda \quad \text{and} \quad \liminf_{n\to\infty} f (x_{n} ) \leq f^{\star}+ \frac { (I+1 ) M_{2} \lambda}{2}. $$

Therefore, we can expect that algorithm (6) with a small enough λ approximates a minimizer of f over \(\bigcap_{i\in\bar{\mathcal{I}}} C^{(i)}\).

Let us also assume \(C := C^{(i)}\) (\(i\in\bar{\mathcal{I}}\)). The following incremental subgradient method ([24], Section 8.2) can solve the problem of minimizing f over C: given \(\lambda>0\) and \(x_{n} = x_{n}^{(0)} = x_{n-1}^{(I)} \in\mathbb{R}^{N}\),

$$ \begin{cases} x_{n}^{(i)} := P_{C} ( x_{n}^{ (i-1 )} - \lambda g_{n}^{ (i )} ),\quad g_{n}^{(i)} \in\partial f^{(i)} (x_{n}^{(i-1)} )\ (i=1,2,\ldots,I), \\ x_{n+1} := x_{n}^{ (I )}. \end{cases} $$
(7)

Algorithm (7) satisfies

$$ \liminf_{n\to\infty} f(x_{n}) \leq f^{*} + \frac{D^{2} \lambda}{2}, $$

where \(\{ x\in C \colon f(x) = f^{*} := \inf_{y\in C} f(y) \} \neq \emptyset\), \(D := \sum_{i\in\mathcal{I}} D_{(i)}\), \(D_{(i)} := \sup \{ \|g\| \colon g\in\partial f^{(i)} (x_{n}) \cup\partial f^{(i)} (x_{n}^{(i-1)}), n\in\mathbb{N} \}\) (\(i\in\mathcal{I}\)), and one assumes that \(D_{(i)} < \infty\) (\(i\in\mathcal{I}\)) ([24], Proposition 8.2.2). In contrast to the above convergence analysis of the incremental subgradient method (7), Theorem 3.1 guarantees that, if \(x_{0}\in C\), the parallel algorithm (6) with \(P_{C} = P_{C^{(i)}}\) (\(i\in\bar{\mathcal{I}}\)) satisfies

$$ x_{n} \in C \quad (n\in\mathbb{N})\quad \text{and}\quad \liminf _{n\to\infty} f (x_{n} ) \leq f^{*} + \frac{ (I+1 ) M_{2} \lambda}{2}. $$

We can see that the previous algorithms (5) and (7) can be applied to the case where the projections onto constraint sets can easily be computed, whereas Algorithm 3.1 can be applied even when \(C^{(i)}\) (\(i\in\bar{\mathcal{I}}\)) has a more complicated form (see, e.g., (1)).

Now, we shall prove Theorem 3.1.

Proof

First, let us show that

$$ \liminf_{n\to\infty} \sum_{i\in\bar{\mathcal{I}}} \bigl\Vert x_{n} - y_{n}^{ (i )} \bigr\Vert ^{2} \leq\frac{ (I+1 ) M_{1} \lambda}{1- \alpha}. $$
(8)

Assume that (8) does not hold. Accordingly, we can choose \(\delta> 0\) such that

$$ \liminf_{n\to\infty} \sum_{i\in\bar{\mathcal{I}}} \bigl\Vert x_{n} - y_{n}^{ (i )} \bigr\Vert ^{2} > \frac{ (I+1 ) M_{1} \lambda}{1- \alpha} + 2 \delta. $$

The property of the limit inferior of \((\sum_{i\in\bar{\mathcal{I}}} \| x_{n} - y_{n}^{(i)} \|^{2})_{n\in \mathbb{N}}\) guarantees that there exists \(n_{0} \in\mathbb{N}\) such that \(\liminf_{n\to\infty} \sum_{i\in\bar{\mathcal{I}}} \| x_{n} - y_{n}^{(i)} \|^{2} - \delta \leq \sum_{i\in\bar{\mathcal{I}}} \| x_{n} - y_{n}^{(i)} \|^{2}\) for all \(n \geq n_{0}\). Accordingly, for all \(n \geq n_{0}\),

$$ \sum_{i\in\bar{\mathcal{I}}} \bigl\Vert x_{n} - y_{n}^{ (i )} \bigr\Vert ^{2} > \frac{ (I+1 ) M_{1} \lambda}{1- \alpha} + \delta. $$

Hence, Lemma 3.1(ii) leads us to that, for all \(n \geq n_{0}\) and for all \(x\in X\),

$$\begin{aligned} \Vert x_{n+1} - x \Vert ^{2} &< \Vert x_{n} -x \Vert ^{2} + M_{1} \lambda - \frac{1 -\alpha}{I+1} \biggl\{ \frac{ (I+1 ) M_{1} \lambda}{1 - \alpha} + \delta \biggr\} \\ &= \Vert x_{n} -x \Vert ^{2} - \frac{1- \alpha}{I+1} \delta. \end{aligned}$$

Therefore, induction ensures that, for all \(n \geq n_{0}\) and for all \(x\in X\),

$$ 0 \leq \Vert x_{n+1} - x \Vert ^{2} < \Vert x_{n_{0}} -x \Vert ^{2} - \frac{1-\alpha}{I+1} \delta (n + 1 -n_{0} ). $$

Since the right side of the above inequality approaches minus infinity when n diverges, we have a contradiction. Therefore, (8) holds. Since \(\liminf_{n\to\infty} \| x_{n} - y_{n}^{(i)}\|^{2} \leq \liminf_{n\to\infty} \sum_{i\in\bar{\mathcal{I}}} \| x_{n} - y_{n}^{(i)}\|^{2}\) (\(i\in\bar{\mathcal{I}}\)), we also find that

$$ \liminf_{n\to\infty} \bigl\Vert x_{n} - y_{n}^{ (i )} \bigr\Vert ^{2} \leq \frac{ (I+1 ) M_{1} \lambda}{1- \alpha}\quad ( i\in\bar{\mathcal{I}} ). $$
(9)

From the triangle inequality we see that, for all \(n\in\mathbb{N}\) and for all \(i\in\bar{\mathcal{I}}\), \(\| x_{n} - T^{(i)} (x_{n}) \| \leq\| x_{n} - y_{n}^{(i)}\| + \|y_{n}^{(i)} - T^{(i)} (x_{n}) \|\), which, together with \(M_{3} := \max_{i\in\bar{\mathcal{I}}} (\sup\{ \|x_{n} - y_{n}^{(i)} \| \colon n\in\mathbb{N} \}) < \infty\) and \(\| y_{n}^{(i)} - T^{(i)} (x_{n}) \| \leq\| ( x_{n} - \lambda g_{n}^{(i)} ) - x_{n} \| \leq\sqrt{M_{2}} \lambda\) (\(n\in\mathbb{N}\), \(i\in\bar{\mathcal{I}}\)), means that, for all \(n\in\mathbb{N}\) and for all \(i\in \bar{\mathcal{I}}\),

$$ \bigl\Vert x_{n} - T^{ (i )} ( x_{n} ) \bigr\Vert ^{2} \leq \bigl\Vert x_{n} - y_{n}^{ (i )} \bigr\Vert ^{2} + 2 \sqrt{M_{2}} M_{3} \lambda+ M_{2} \lambda^{2}. $$

Thus, (9) guarantees that

$$\begin{aligned} \liminf_{n\to\infty} \bigl\Vert x_{n} - T^{ (i )} ( x_{n} ) \bigr\Vert ^{2} &\leq \liminf _{n\to\infty} \bigl[ \bigl\Vert x_{n} - y_{n}^{ (i )} \bigr\Vert ^{2} + (2 \sqrt{M_{2}} M_{3} + M_{2} \lambda ) \lambda \bigr] \\ &= \liminf_{n\to\infty} \bigl\Vert x_{n} - y_{n}^{ (i )} \bigr\Vert ^{2} + (2 \sqrt{M_{2}} M_{3} + M_{2} \lambda ) \lambda \\ &\leq \biggl( \frac{ (I+1 ) M_{1}}{1 - \alpha} + 2 \sqrt {M_{2}} M_{3} + M_{2} \lambda \biggr) \lambda. \end{aligned}$$

Next, let us show that

$$ \liminf_{n\to\infty} f (x_{n} ) \leq f^{\star}+ \frac { (I+1 ) M_{2} \lambda}{2}. $$
(10)

Assume that (10) does not hold. Since (A5) guarantees that \(x^{\star}\in X\) exists such that \(f(x^{\star}) = f^{\star}\), we can choose \(\epsilon> 0\) such that

$$ \liminf_{n\to\infty} f (x_{n} ) > f \bigl(x^{\star}\bigr) + \frac { (I+1 ) M_{2} \lambda}{2} + 2 \epsilon. $$

From the property of the limit inferior of \((f (x_{n}))_{n\in\mathbb{N}}\), there exists \(n_{1} \in\mathbb{N}\) such that \(\liminf_{n\to\infty} f (x_{n}) - {\epsilon} \leq f (x_{n})\) for all \(n \geq n_{1}\). Accordingly, for all \(n \geq n_{1}\),

$$ f (x_{n} ) - f \bigl( x^{\star}\bigr) > \frac{ (I+1 ) M_{2} \lambda}{2} + {\epsilon}. $$
(11)

Therefore, from Lemma 3.1(iii) and (11) we see that, for all \(n \geq n_{1}\),

$$\begin{aligned} \bigl\Vert x_{n +1} - x^{\star}\bigr\Vert ^{2} &< \bigl\Vert x_{n} -x^{\star}\bigr\Vert ^{2} + M_{2} (1-\alpha ) \lambda^{2}+ \frac{2(1-\alpha)\lambda }{I+1} \biggl\{ - \frac{ (I+1 ) M_{2} \lambda}{2} - {\epsilon} \biggr\} \\ &= \bigl\Vert x_{n} -x^{\star}\bigr\Vert ^{2} - \frac{2(1-\alpha)\lambda }{I+1} {\epsilon}, \end{aligned}$$

which implies that, for all \(n \geq n_{1}\),

$$ \bigl\Vert x_{n +1} - x^{\star}\bigr\Vert ^{2} < \bigl\Vert x_{n_{1}} -x^{\star}\bigr\Vert ^{2} - \frac{2(1-\alpha)\lambda}{I+1} {\epsilon} (n+1 - n_{1} ). $$

Since the above inequality does not hold for large enough n, we have arrived at a contradiction. Therefore, (10) holds. This completes the proof. □

3.2 Diminishing step-size rule

The discussion in this subsection makes the following assumption.

Assumption 3.3

User i (\(i\in\bar{\mathcal{I}}\)) has \((\lambda_{n})_{n\in\mathbb{N}}\) satisfying

$$ (\mathrm{C}2)\quad \lim_{n\to\infty}\lambda_{n} = 0 \quad \text{and}\quad \sum_{n=0}^{\infty} \lambda_{n} = \infty. $$

An example of \((\lambda_{n})_{n\in\mathbb{N}}\) is \(\lambda_{n} := 1/(n+1)^{a}\) (\(n\in\mathbb{N}\)), where \(a\in(0,1]\).

Let us perform a convergence analysis on Algorithm 3.1 under Assumption 3.3.

Theorem 3.2

Suppose that Assumptions (A1)-(A5), 3.1, and 3.3 hold. Then there exists a subsequence of \((x_{n})_{n\in\mathbb{N}}\) generated by Algorithm 3.1 which weakly converges to a point in \(X^{\star}\).

Let us compare Algorithm 3.1 under the assumptions in Theorem 3.2 with the previous gradient algorithms with diminishing step sizes ([24], Section 8.2), [2]. Suppose that \(C := C^{(i)}\) (\(i\in\bar{\mathcal{I}}\)). The sequence \((x_{n})_{n\in\mathbb{N}}\) is generated by the incremental subgradient method ([24], Section 8.2) as follows (see also (7)): given \((\lambda_{n})_{n\in\mathbb{N}}\) with (C2), and \(x_{n} = x_{n}^{(0)} = x_{n-1}^{(I)} \in\mathbb{R}^{N}\),

$$ \begin{cases} x_{n}^{(i)} := P_{C} ( x_{n}^{ (i-1 )} - \lambda_{n} g_{n}^{ (i )} ),\quad g_{n}^{(i)} \in\partial f^{(i)} (x_{n}^{ (i-1 )} )\ (i=1,2,\ldots,I), \\ x_{n+1} := x_{n}^{ (I )}. \end{cases} $$

The incremental subgradient method satisfies

$$ \liminf_{n\to\infty} f(x_{n}) = f^{*}, $$

where \(\{ x\in C \colon f(x) = f^{*} := \inf_{y\in C} f(y) \} \neq \emptyset\), \(D_{(i)} := \sup \{ \|g\| \colon g\in\partial f^{(i)} (x_{n}) \cup\partial f^{(i)} (x_{n}^{(i-1)}), n\in\mathbb{N} \}\) (\(i\in\mathcal{I}\)), and one assumes that \(D_{(i)} < \infty\) (\(i\in\mathcal{I}\)) ([24], Proposition 8.2.4).

The following broadcast gradient method ([2], Algorithm 4.1) can minimize the sum of convex, smooth functionals over the intersection of fixed point sets: given \(x_{0}^{(i)} \in H\) (\(i\in\bar{\mathcal{I}}\)),

$$ \textstyle\begin{cases} x_{n+1}^{ (i )} := \alpha_{n} x_{0}^{ (i )} + (1-\alpha_{n} ) T^{(i)} ( x_{n} -\lambda_{n} \nabla f^{(i)} (x_{n} ) ) \quad (i=0,1,\ldots,I), \\ x_{n+1} := {\frac{1}{I+1} \sum_{i\in\bar{\mathcal {I}}} x_{n+1}^{ (i )}}, \end{cases} $$

where \(\nabla f^{(i)}\) (\(i\in\mathcal{I}\)) is the Lipschitz continuous gradient of \(f^{(i)}\), and \((\alpha_{n})_{n\in\mathbb{N}}\) and \((\lambda_{n})_{n\in\mathbb {N}}\) are slowly diminishing sequences such as \(\lambda_{n} := 1/(n+1)^{a}\) and \(\alpha_{n} := 1/(n+1)^{b}\) (\(n\in\mathbb{N}\)), where \(a \in(0,1/2)\), \(b\in(a,1-a)\). The sequence \((x_{n})_{n\in\mathbb{N}}\) weakly converges to a minimizer of f over X ([2], Theorem 4.1).

Meanwhile, Algorithm 3.1 works even when \(f^{(i)}\) (\(i\in \mathcal{I}\)) is convex and nondifferentiable and \(T^{(i)}\) (\(i\in\bar{\mathcal{I}}\)) is firmly nonexpansive. Theorem 3.2 guarantees that there exists a subsequence of \((x_{n})_{n\in\mathbb{N}}\) in Algorithm 3.1 with (C2) such that it weakly converges to a point in \(X^{\star}\).

The rest of this subsection gives the proof of Theorem 3.2.

Proof

Fix \(x\in X\) arbitrarily. We will distinguish two cases.

Case 1: Suppose that \(m_{0} \in\mathbb{N}\) exists such that \(\| x_{n+1} -x \| \leq\| x_{n} -x\|\) (\(n\geq m_{0}\)). Lemma 3.1(ii) means that, for all \(n\in\mathbb{N}\),

$$ \frac{1-\alpha}{I+1} \sum_{i\in\bar{\mathcal{I}}} \bigl\Vert x_{n} - y_{n}^{ (i )} \bigr\Vert ^{2} \leq \Vert x_{n} -x \Vert ^{2} - \Vert x_{n+1} - x \Vert ^{2} + M_{1} \lambda_{n}, $$

which, together with the existence of \(\lim_{n\to\infty} \Vert x_{n} - x \Vert \) and \(\lim_{n\to\infty} \lambda_{n} =0\), implies that \(\lim_{n\to\infty} (1-\alpha)/(I+1) \sum_{i\in\bar{\mathcal {I}}} \| x_{n} - y_{n}^{(i)} \|^{2} = 0\), i.e.,

$$ \lim_{n\to\infty} \bigl\Vert x_{n} - y_{n}^{ (i )} \bigr\Vert = 0\quad (i\in\bar{\mathcal{I}} ). $$
(12)

Moreover, (A1) (the nonexpansivity of \(T^{(i)}\) (\(i\in\bar{\mathcal{I}}\))) guarantees that, for all \(n\in\mathbb{N}\) and \(i\in\bar{\mathcal{I}}\), \(\| y_{n}^{(i)} - T^{(i)} (x_{n})\| \leq \| ( x_{n} - \lambda_{n} g_{n}^{(i )} )- x_{n} \| \leq \sqrt{M_{2}} \lambda_{n}\), which, together with \(\lim_{n\to\infty} \lambda_{n} = 0\), means that

$$ \lim_{n\to\infty} \bigl\Vert y_{n}^{ (i )} - T^{ (i )} (x_{n} )\bigr\Vert = 0 \quad (i\in\bar {\mathcal{I}} ). $$
(13)

Since the triangle inequality implies \(\| x_{n} - T^{(i)} (x_{n}) \| \leq \| x_{n} - y_{n}^{(i)} \| + \| y_{n}^{(i)} - T^{(i)} (x_{n}) \|\) (\(n\in\mathbb {N}\), \(i\in\bar{\mathcal{I}}\)), (12) and (13) guarantee that

$$ \lim_{n\to\infty} \bigl\Vert x_{n} - T^{ (i )} (x_{n} )\bigr\Vert = 0 \quad (i\in\bar{\mathcal{I}} ). $$
(14)

Here, we define, for all \(n\in\mathbb{N}\),

$$ M_{n} := (1-\alpha) \biggl\{ \frac{2}{I+1} \bigl( f(x_{n}) - f(x) \bigr) - M_{2} \lambda_{n} \biggr\} . $$

Then Lemma 3.1(iii) implies that, for all \(n\in\mathbb{N}\), \(\lambda_{n} M_{n} \leq\| x_{n+1} -x \|^{2} - \|x_{n} -x\|^{2}\), which means \(\sum_{n=0}^{m} \lambda_{n} M_{n} \leq\| x_{0} -x \|^{2} - \|x_{m+1} -x\|^{2} \leq\| x_{0} -x \| < \infty\) (\(m\in\mathbb{N}\)). Accordingly, we find that

$$ \sum_{n=0}^{\infty}\lambda_{n} M_{n} < \infty. $$

Therefore, from \(\sum_{n=0}^{\infty}\lambda_{n} = \infty\), we find that

$$ \liminf_{n\to\infty} M_{n} \leq0. $$
(15)

Indeed, let us assume that \(\liminf_{n\to\infty} M_{n} \leq0\) does not hold, i.e., \(\liminf_{n\to\infty} M_{n} > 0\). Then there exist \(m_{1} \in\mathbb{N}\) and \(\gamma> 0\) such that \(M_{n} \geq\gamma\) for all \(n \geq m_{1}\). From \(\sum_{n=0}^{\infty}\lambda_{n} = \infty\), we have \(\infty= \gamma\sum_{n=m_{1}}^{\infty}\lambda_{n} \leq\sum_{n=m_{1}}^{\infty}\lambda_{n} M_{n} < \infty\), which is a contradiction. Hence, (15) holds. Accordingly, from \(\lim_{n\to\infty} \lambda_{n} = 0\), we find that

$$\begin{aligned} 0 &\geq\liminf_{n\to\infty} \biggl\{ \frac{2}{I+1} \bigl( f(x_{n}) - f(x) \bigr) - M_{2} \lambda_{n} \biggr\} \\ &= \frac{2}{I+1} \liminf_{n\to\infty} \bigl( f(x_{n}) - f(x) \bigr) - M_{2} \lim_{n\to\infty} \lambda_{n} \\ &= \frac{2}{I+1} \liminf_{n\to\infty} \bigl( f(x_{n}) - f(x) \bigr). \end{aligned}$$

This means there is a subsequence \((x_{n_{l}})_{l\in\mathbb{N}}\) of \((x_{n})_{n\in\mathbb{N}}\) such that

$$ \lim_{l\to\infty} f ( x_{n_{l}} ) = \liminf _{n\to\infty }f(x_{n}) \leq f(x) \quad (x\in X ). $$
(16)

The boundedness of \((x_{n_{l}})_{l\in\mathbb{N}}\) guarantees that \((x_{n_{l_{m}}})_{m\in\mathbb{N}}\) (\(\subset(x_{n_{l}})_{l\in\mathbb {N}}\)) exists such that \((x_{n_{l_{m}}})_{m\in\mathbb{N}}\) weakly converges to \(x_{\star}\in H\). Here, fix \(i\in\bar{\mathcal{I}}\) arbitrarily and assume that \(x_{\star}\notin\operatorname{Fix}(T^{(i)})\). From Opial’s condition ([36], Lemma 1), (14), and the nonexpansivity of \(T^{(i)}\), we produce a contradiction:

$$\begin{aligned} \liminf_{m\to\infty} \Vert x_{n_{l_{m}}} - x_{\star} \Vert &< \liminf_{m\to\infty} \bigl\Vert x_{n_{l_{m}}} - T^{(i)} (x_{\star}) \bigr\Vert \\ &= \liminf_{m\to\infty} \bigl\Vert x_{n_{l_{m}}} - T^{(i)} ( x_{n_{l_{m}}} ) + T^{(i)} ( x_{n_{l_{m}}} ) - T^{(i)} (x_{\star}) \bigr\Vert \\ &= \liminf_{m\to\infty} \bigl\Vert T^{(i)} ( x_{n_{l_{m}}} ) - T^{(i)} (x_{\star}) \bigr\Vert \\ &\leq \liminf_{m\to\infty} \Vert x_{n_{l_{m}}} - x_{\star} \Vert . \end{aligned}$$

Hence, \(x_{\star}\in\operatorname{Fix}(T^{(i)})\) (\(i\in\bar{\mathcal {I}}\)), i.e., \(x_{\star}\in X\). Moreover, since f is weakly lower semicontinuous ([15], Theorem 9.1) and (16), we find that

$$ f ( x_{\star}) \leq \liminf_{m\to\infty} f ( x_{n_{l_{m}}} ) = \lim_{l\to \infty} f ( x_{n_{l}} ) \leq f(x) \quad (x\in X ). $$

Therefore, \(x_{\star}\in X^{\star}\).

Let us take another subsequence \((x_{n_{l_{k}}})_{k\in\mathbb{N}}\) (\(\subset(x_{n_{l}})_{l\in\mathbb{N}}\)) which weakly converges to \(x_{\star\star} \in H\). A similar discussion to the one for obtaining \(x_{\star}\in X^{\star}\) ensures that \(x_{\star\star} \in X^{\star}\). Assume that \(x_{\star}\neq x_{\star\star}\). The existence of \(\lim_{n\to\infty} \| x_{n} - x \|\) (\(x\in X\)) and Opial’s condition ([36], Lemma 1) imply that

$$\begin{aligned} \lim_{n\to\infty} \Vert x_{n} - x_{\star} \Vert &= \lim_{m\to\infty} \Vert x_{n_{l_{m}}} - x_{\star} \Vert < \lim_{m\to\infty} \Vert x_{n_{l_{m}}} - x_{\star\star} \Vert \\ &= \lim_{n\to\infty} \Vert x_{n} - x_{\star\star} \Vert = \lim_{k\to\infty} \Vert x_{n_{l_{k}}} - x_{\star\star} \Vert < \lim_{k\to\infty} \Vert x_{n_{l_{k}}} - x_{\star} \Vert \\ &= \lim_{n\to\infty} \Vert x_{n} - x_{\star} \Vert , \end{aligned}$$

which is a contradiction. Hence, \(x_{\star}= x_{\star\star}\). Accordingly, any subsequence of \((x_{n_{l}})_{l\in\mathbb{N}}\) converges weakly to \(x_{\star}\in X^{\star}\), i.e., \((x_{n_{l}})_{l\in\mathbb{N}}\) converges weakly to \(x_{\star}\in X^{\star}\). This means that \(x_{\star}\) is a weak cluster point of \((x_{n})_{n\in \mathbb{N}}\) and belongs to \(X^{\star}\). A similar discussion to the one for obtaining \(x_{\star}= x_{\star\star}\) guarantees that there is only one weak cluster point of \((x_{n})_{n\in \mathbb{N}}\), and hence, we can conclude that, in Case 1, \((x_{n})_{n\in\mathbb{N}}\) weakly converges to a point in \(X^{\star}\).

Case 2: Suppose that \((x_{n_{j}})\) (\(\subset(x_{n})_{n\in\mathbb{N}}\)) exists such that \(\| x_{n_{j}} - x \| < \| x_{n_{j} +1} - x \|\) for all \(j\in\mathbb{N}\). Lemma 3.1(ii) means that, for all \(j\in\mathbb{N}\),

$$ \frac{1-\alpha}{I+1} \sum_{i\in\bar{\mathcal{I}}} \bigl\Vert x_{n_{j}} - y_{n_{j}}^{ (i )} \bigr\Vert ^{2} \leq \Vert x_{n_{j}} - x \Vert ^{2} - \Vert x_{n_{j} + 1} - x \Vert ^{2} + M_{1} \lambda_{n_{j}} < M_{1} \lambda_{n_{j}}, $$

which, together with \(\lim_{n\to\infty} \lambda_{n} =0\), implies that

$$ \lim_{j\to\infty} \bigl\Vert x_{n_{j}} - y_{n_{j}}^{ (i )} \bigr\Vert = 0 \quad (i\in\bar{\mathcal{I}} ). $$
(17)

Therefore, a similar discussion to the one for obtaining (14) ensures that

$$ \lim_{j\to\infty} \bigl\Vert x_{n_{j}} - T^{ (i )} (x_{n_{j}} ) \bigr\Vert = 0 \quad (i\in\bar{\mathcal {I}} ). $$
(18)

Since Lemma 3.1(iii) implies that \(\lambda_{n_{j}} M_{n_{j}} \leq\| x_{n_{j}} -x \| - \| x_{n_{j} +1} -x\| < 0\) (\(j\in\mathbb{N}\)) and \(\lambda_{n_{j}} > 0\) (\(j\in\mathbb{N}\)), we find that \(M_{n_{j}} < 0\) (\(j\in\mathbb{N}\)), i.e., for all \(j\in\mathbb{N}\),

$$ \frac{2}{I+1} \bigl( f (x_{n_{j}} ) - f(x) \bigr) < M_{2} \lambda_{n_{j}}. $$

Since \(\lim_{n\to\infty} \lambda_{n} =0\) implies that

$$ \frac{2}{I+1} \limsup_{j\to\infty} \bigl( f (x_{n_{j}} ) - f(x) \bigr) \leq M_{2} \lim_{j\to\infty} \lambda_{n_{j}} = 0, $$

we find that

$$ \limsup_{j\to\infty} f (x_{n_{j}} ) \leq f(x) \quad (x\in X). $$
(19)

Inequality (19) ensures the existence of \((x_{n_{j_{k}}})_{k\in \mathbb{N}}\) of \((x_{n_{j}})_{j\in\mathbb{N}}\) such that

$$ \lim_{k\to\infty} f (x_{n_{j_{k}}} ) = \limsup _{j\to\infty} f (x_{n_{j}} ) \leq f(x)\quad (x\in X). $$
(20)

Since \((x_{n_{j_{k}}})_{k\in\mathbb{N}}\) is bounded, we have \((x_{n_{j_{k_{l}}}})_{l\in\mathbb{N}}\), which weakly converges to \(x_{*} \in H\). A similar discussion to the one for obtaining \(x_{\star}\in X\) and (18) leads us to \(x_{*} \in X\). Moreover, the weakly lower semicontinuity of f ([15], Theorem 9.1) and (20) guarantee that

$$ f(x_{*} ) \leq\liminf_{l\to\infty} f (x_{n_{j_{k_{l}}}} ) = \lim _{k\to\infty} f (x_{n_{j_{k}}} ) \leq f(x) \quad (x\in X), \textit{i.e.}, x_{*} \in X^{\star}. $$

Therefore, there exists a subsequence of \((x_{n})_{n\in\mathbb{N}}\) such that it weakly converges to a point in \(X^{\star}\). This completes the proof. □

4 Numerical examples

Let us look at some numerical examples to see how Algorithm 3.1 works depending on the choice of step size. Consider the following problem: given \(a^{(i)} > 0\), \(b^{(i)} \in\mathbb{R}\), \(d_{k}^{(i)} \in\mathbb {R}\), and \(c_{k}^{(i)} \in\mathbb{R}^{I+1}\) with \(c_{k}^{(i)} \neq0\) (\(i\in\bar{\mathcal{I}} := \{0,1,2,\ldots,I\}\), \(k \in\mathcal{K} := \{1,2,\ldots,K\}\)),

$$ \text{minimize} \quad \sum_{i\in\bar{\mathcal{I}}} \bigl\vert a^{ (i )} x_{ (i )} + b^{ (i )} \bigr\vert \quad \text{subject to}\quad (x_{ (i )} )_{i\in\bar {\mathcal{I}}} \in C \cap \bigcap _{i\in\bar{\mathcal{I}}} C^{ (i )}, $$
(21)

where \(f^{(i)}(x) := |a^{(i)} x + b^{(i)}|\) (\(i\in\bar{\mathcal{I}}\), \(x\in\mathbb{R}\)), \(C_{k}^{(i)}\) (\(\subset\mathbb{R}^{I+1}\)) (\(i\in \bar{\mathcal{I}}\), \(k\in\mathcal{K}\)) is a half-space defined by \(C_{k}^{(i)} := \{ x\in\mathbb{R}^{I+1} \colon\langle c_{k}^{(i)}, x \rangle\leq d_{k}^{(i)} \}\), \(C^{(i)} := \bigcap_{k\in\mathcal{K}} C_{k}^{(i)} \neq\emptyset\) (\(i\in\bar{\mathcal{I}}\)), C (\(\subset\mathbb{R}^{I+1}\)) is a closed ball, and \(C \cap\bigcap_{i\in\bar{\mathcal{I}}} C^{(i)} \neq\emptyset\).

We will assume that user i (\(i\in\bar{\mathcal{I}}\)) computes

$$ x_{n}^{(i)} := P_{C} \bigl( \alpha x_{n} + ( 1 - \alpha ) T^{(i)} \bigl( x_{n} - \lambda_{n} g_{n}^{(i)} \bigr) \bigr)\quad (n\in \mathbb{N} ), $$

where \(T^{(i)}\) is defined by

$$ T^{ (i )} := \frac{1}{2} \biggl[ \mathrm{Id} + P_{C} \prod_{k\in\mathcal{K}} P_{k}^{(i)} \biggr], $$

\(P_{k}^{(i)} := P_{C_{k}^{(i)}}\) (\(k\in\mathcal{K}\)), \(g_{n}^{(i)} = (0,0,\ldots,0, \bar{g}_{n}^{(i)}, 0,0, \ldots, 0)\), and

$$ \bar{g}_{n}^{(i)} \in\partial f^{(i)} ({x_{n}}_{ (i )} ) := \begin{cases} - a^{ (i )}& ( - \infty< {x_{n}}_{ (i )} < - \frac{b^{(i)}}{a^{(i)}} ), \\ {[- a^{ (i )}, a^{ (i )} ]}& ( {x_{n}}_{ (i )} = - \frac{b^{(i)}}{a^{(i)}} ), \\ a^{ (i )} & (- \frac{b^{(i)}}{a^{(i)}} < {x_{n}}_{ (i )} < \infty ). \end{cases} $$

Since \((x_{n}^{(i)})_{n\in\mathbb{N}} \subset C\) (\(i\in\bar{\mathcal{I}}\)), the boundedness of C means Assumption 3.1 holds (see also (1) and (2)). Moreover, the continuity and convexity of f ensures that \(X^{\star}\neq\emptyset\) ([15], Proposition 11.14). The projections \(P_{C}\) and \(P_{k}^{(i)}\) (\(i\in\bar{\mathcal{I}}\), \(k\in \mathcal{K}\)) can be computed within a finite number of arithmetic operations ([15], Chapter 28), and hence, \(T^{(i)}\) (\(i\in\bar{\mathcal{I}}\)) can also be computed easily. User i can randomly choose \(\bar{a}^{(i)} \in\partial f^{(i)} (- b^{(i)}/a^{(i)}) = [-a^{(i)}, a^{(i)}]\).

The experiment used a 15.4-inch MacBook Pro with a 2.6 GHz Intel Core i7 processor and 16 GB 1600 MHz DDR3 memory. Algorithm 3.1 was written in MATLAB 8.2. We set \(I:=3\) and \(K:=3\), and used \(a^{(i)}\), \(b^{(i)}\), \(c_{k}^{(i)}\), \(d_{k}^{(i)}\), and \(\bar {a}^{(i)}\) randomly generated by MATLAB. We used

$$ \alpha:= \frac{1}{2}, \qquad \lambda_{n} := \frac{1}{10}, \frac{1}{10^{3}}, \frac{1}{(n+1)^{a}} \quad (n\in\mathbb{N}), \text{ where } a = 0.5, 1. $$

We performed 100 samplings, each starting from different random initial points given by MATLAB, and averaged their results.

We used the following performance measures: for each \(n\in\mathbb{N}\),

$$\begin{aligned}& D_{n} := \frac{1}{100} \sum_{s=1}^{100} \sum_{i\in\bar{\mathcal{I}}} \bigl\Vert x_{n} (s ) - T^{ (i )} \bigl(x_{n} (s ) \bigr) \bigr\Vert ^{2} \quad \text{and} \\& F_{n} := \frac{1}{100} \sum_{s=1}^{100} \sum_{i\in\bar{\mathcal{I}}} \bigl\vert a^{ (i )} x_{n(i)} (s ) + b^{ (i )} \bigr\vert , \end{aligned}$$

where \((x_{n} (s))_{n\in\mathbb{N}}\) is the sequence generated by the initial point \(x(s)\) (\(s=1,2,\ldots,100\)) and Algorithm 3.1, and \(x_{n} (s) := (x_{n(i)}(s))_{i\in\bar {\mathcal{I}}}\) (\(n\in\mathbb{N}\), \(s=1,2,\ldots,100\)). \(D_{n}\) (\(n\in\mathbb{N}\)) stands for the mean value of the sums of the squared distances between \(x_{n}(s)\) and \(T^{(i)}(x_{n}(s))\) (\(i\in\bar{\mathcal{I}}\), \(s=1,2,\ldots,100\)). If \((D_{n})_{n\in\mathbb{N}}\) converges to 0, Algorithm 3.1 converges to a point in \(\bigcap_{i\in\bar{\mathcal{I}}} \operatorname{Fix}(T^{(i)}) = C \cap \bigcap_{i\in\bar{\mathcal{I}}} C^{(i)}\). \(F_{n}\) (\(n\in\mathbb{N}\)) is the mean value of the objective function \(\sum_{i\in\bar{\mathcal{I}}} f^{(i)} (x_{n(i)}(s))\) (\(s = 1,2,\ldots,100\)).

Figure 1 indicates the behavior of \(D_{n}\) for Algorithm 3.1. We can see that the sequences generated by Algorithm 3.1 with \(\lambda_{n} := 1/(n+1)^{a}\) (\(a=0.5,1\), \(n\in\mathbb{N}\)) converge to a point in \(\bigcap_{i\in\bar{\mathcal{I}}} \operatorname {Fix}(T^{(i)})\). Meanwhile, Figure 1 shows that Algorithm 3.1 with \(\lambda_{n} := 1/10\) (\(n\in\mathbb{N}\)) does not converge in \(\bigcap_{i\in\bar {\mathcal{I}}} \operatorname{Fix}(T^{(i)})\), and \((D_{n})_{n\in\mathbb{N}}\) in Algorithm 3.1 with \(\lambda _{n} := 1/10^{3}\) (\(n\in\mathbb{N}\)) initially decreases. This is because the use of \(\lambda:= 1/10^{3}\) satisfies \(\liminf_{n\to\infty} \| x_{n} - T^{(i)}(x_{n}) \| \leq M/10^{3} \approx0\) (\(i\in \bar{\mathcal{I}}\)) (see Theorem 3.1).

Figure 1
figure 1

Behavior of \(\pmb{D_{n}}\) for Algorithm 3.1 when \(\pmb{\lambda_{n} := 1/10, 1/10^{3}, 1/(n+1)^{a}}\) ( \(\pmb{a=0.5, 1}\) ).

Figure 2 plots the behavior of \(F_{n}\) for Algorithm 3.1 and shows that Algorithm 3.1 with \(\lambda_{n} := 1/(n+1)\) (\(n\in\mathbb{N}\)) is stable during the early iterations and converges to a solution to problem (21), as promised by Theorem 3.2. This figure indicates that the \((F_{n})_{n\in\mathbb{N}}\) generated by Algorithm 3.1 with \(\lambda:= 1/10^{3}\) (\(n\in\mathbb{N}\)) decreases slowly. Therefore, Figures 1 and 2, and Theorem 3.2 show that Algorithm 3.1 with \(\lambda_{n} := 1/(n+1)\) (\(n\in\mathbb{N}\)) converges to a solution to problem (21).

Figure 2
figure 2

Behavior of \(\pmb{F_{n}}\) for Algorithm 3.1 when \(\pmb{\lambda_{n} := 1/10, 1/10^{3}, 1/(n+1)^{a}}\) ( \(\pmb{a=0.5, 1}\) ).

5 Conclusion

This paper discussed the problem of minimizing the sum of nondifferentiable, convex functions over the intersection of the fixed point sets of firmly nonexpansive mappings in a real Hilbert space. It presented a parallel algorithm for solving the problem. The parallel algorithm does not use any proximity operators, in contrast to conventional parallel algorithms. Moreover, the parallel algorithm can work in nonsmooth convex optimization over constraint sets onto which projections cannot be always implemented, while the conventional incremental subgradient method can only be applied when the constraint set is simple in the sense that the projection onto it can easily be implemented. We studied its convergence properties for the two step-size rules, a constant step size and a diminishing step size. We showed that the algorithm with a small constant step size will approximate a solution to the problem, while there exists a subsequence of the sequence generated by the algorithm with a diminishing step size which weakly converges to a solution to the problem. We also gave numerical examples to support the convergence analyses.

References

  1. Combettes, PL: A block-iterative surrogate constraint splitting method for quadratic signal recovery. IEEE Trans. Signal Process. 51(7), 1771-1782 (2003)

    Article  MathSciNet  Google Scholar 

  2. Iiduka, H: Fixed point optimization algorithms for distributed optimization in networked systems. SIAM J. Optim. 23, 1-26 (2013)

    Article  MATH  MathSciNet  Google Scholar 

  3. Iiduka, H: Acceleration method for convex optimization over the fixed point set of a nonexpansive mapping. Math. Program. 149, 131-165 (2015)

    Article  MATH  MathSciNet  Google Scholar 

  4. Iiduka, H, Hishinuma, K: Acceleration method combining broadcast and incremental distributed optimization algorithms. SIAM J. Optim. 24, 1840-1863 (2014)

    Article  MATH  MathSciNet  Google Scholar 

  5. Iiduka, H, Yamada, I: A use of conjugate gradient direction for the convex optimization problem over the fixed point set of a nonexpansive mapping. SIAM J. Optim. 19(4), 1881-1893 (2009)

    Article  MATH  MathSciNet  Google Scholar 

  6. Maingé, PE: A viscosity method with no spectral radius requirements for the split common fixed point problem. Eur. J. Oper. Res. 235, 17-27 (2014)

    Article  MATH  Google Scholar 

  7. Yamada, I: The hybrid steepest descent method for the variational inequality problem over the intersection of fixed point sets of nonexpansive mappings. In: Butnariu, D, Censor, Y, Reich, S (eds.) Inherently Parallel Algorithms for Feasibility and Optimization and Their Applications, pp. 473-504. Elsevier, Amsterdam (2001)

    Chapter  Google Scholar 

  8. Yao, Y, Cho, YJ, Liou, YC: Algorithms of common solutions for variational inclusions, mixed equilibrium problems and fixed point problems. Eur. J. Oper. Res. 212, 242-250 (2011)

    Article  MATH  MathSciNet  Google Scholar 

  9. Facchinei, F, Pang, J, Scutari, G, Lampariello, L: VI-constrained hemivariational inequalities: distributed algorithms and power control in ad-hoc networks. Math. Program. 145, 59-96 (2014)

    Article  MATH  MathSciNet  Google Scholar 

  10. Iiduka, H: Iterative algorithm for solving triple-hierarchical constrained optimization problem. J. Optim. Theory Appl. 148, 580-592 (2011)

    Article  MATH  MathSciNet  Google Scholar 

  11. Iiduka, H: Iterative algorithm for triple-hierarchical constrained nonconvex optimization problem and its application to network bandwidth allocation. SIAM J. Optim. 22(3), 862-878 (2012)

    Article  MATH  MathSciNet  Google Scholar 

  12. Iiduka, H, Yamada, I: Computational method for solving a stochastic linear-quadratic control problem given an unsolvable stochastic algebraic Riccati equation. SIAM J. Control Optim. 50, 2173-2192 (2012)

    Article  MATH  MathSciNet  Google Scholar 

  13. Maingé, PE: Projected subgradient techniques and viscosity methods for optimization with variational inequality constraints. Eur. J. Oper. Res. 205, 501-506 (2010)

    Article  MATH  Google Scholar 

  14. Combettes, PL, Bondon, P: Hard-constrained inconsistent signal feasibility problems. IEEE Trans. Signal Process. 47(9), 2460-2468 (1999)

    Article  MATH  Google Scholar 

  15. Bauschke, HH, Combettes, PL: Convex Analysis and Monotone Operator Theory in Hilbert Spaces. Springer, Berlin (2011)

    Book  MATH  Google Scholar 

  16. Yamada, I, Yukawa, M, Yamagishi, M: Minimizing the Moreau envelope of nonsmooth convex functions over the fixed point set of certain quasi-nonexpansive mappings. In: Bauschke, HH, Burachik, RS, Combettes, PL, Elser, V, Luke, DR, Wolkowicz, H (eds.) Fixed-Point Algorithms for Inverse Problems in Science and Engineering, pp. 345-390. Springer, Berlin (2011)

    Chapter  Google Scholar 

  17. Combettes, PL, Pesquet, JC: A Douglas-Rachford splitting approach to nonsmooth convex variational signal recovery. IEEE J. Sel. Top. Signal Process. 1, 564-574 (2007)

    Article  Google Scholar 

  18. Combettes, PL, Pesquet, JC: Proximal splitting methods in signal processing. In: Bauschke, HH, Burachik, RS, Combettes, PL, Elser, V, Luke, DR, Wolkowicz, H (eds.) Fixed-Point Algorithms for Inverse Problems in Science and Engineering, pp. 185-212. Springer, Berlin (2011)

    Chapter  Google Scholar 

  19. Eckstein, J, Bertsekas, DP: On the Douglas-Rachfold splitting method and proximal point algorithm for maximal monotone operators. Math. Program. 55, 293-318 (1992)

    Article  MATH  MathSciNet  Google Scholar 

  20. Lions, PL, Mercier, B: Splitting algorithms for the sum of two nonlinear operators. SIAM J. Numer. Anal. 16, 964-979 (1979)

    Article  MATH  MathSciNet  Google Scholar 

  21. Combettes, PL: Iterative construction of the resolvent of a sum of maximal monotone operators. J. Convex Anal. 16, 727-748 (2009)

    MATH  MathSciNet  Google Scholar 

  22. Combettes, PL, Pesquet, JC: A proximal decomposition method for solving convex variational inverse problems. Inverse Probl. 24, 065014 (2008)

    Article  MathSciNet  Google Scholar 

  23. Pesquet, JC, Pustelnik, N: A parallel inertial proximal optimization method. Pac. J. Optim. 8, 273-306 (2012)

    MATH  MathSciNet  Google Scholar 

  24. Bertsekas, DP, Nedić, A, Ozdaglar, AE: Convex Analysis and Optimization. Athena Scientific, Belmont (2003)

    MATH  Google Scholar 

  25. Lobel, I, Ozdaglar, A, Feijer, D: Distributed multi-agent optimization with state- dependent communication. Math. Program. 129, 255-284 (2011)

    Article  MATH  MathSciNet  Google Scholar 

  26. Nedić, A, Olshevsky, A, Ozdaglar, A, Tsitsiklis, JN: On distributed averaging algorithms and quantization effects. IEEE Trans. Autom. Control 54, 2506-2517 (2009)

    Article  Google Scholar 

  27. Nedić, A, Ozdaglar, A: Distributed subgradient methods for multi agent optimization. IEEE Trans. Autom. Control 54, 48-61 (2009)

    Article  Google Scholar 

  28. Nedić, A, Ozdaglar, A: Cooperative distributed multi-agent optimization. In: Convex Optimization in Signal Processing and Communications, pp. 340-386 (2010)

    Google Scholar 

  29. Rockafellar, RT: Convex Analysis. Princeton University Press, Princeton (1970)

    Book  MATH  Google Scholar 

  30. Krasnosel’skiĭ, MA: Two remarks on the method of successive approximations. Usp. Mat. Nauk 10, 123-127 (1955)

    Google Scholar 

  31. Mann, WR: Mean value methods in iteration. Proc. Am. Math. Soc. 4, 506-510 (1953)

    Article  MATH  Google Scholar 

  32. Low, S, Lapsley, DE: Optimization flow control. I: basic algorithm and convergence. IEEE/ACM Trans. Netw. 7(6), 861-874 (1999)

    Article  Google Scholar 

  33. Maillé, P, Toka, L: Managing a peer-to-peer data storage system in a selfish society. IEEE J. Sel. Areas Commun. 26(7), 1295-1301 (2008)

    Article  Google Scholar 

  34. Sharma, S, Teneketzis, D: An externalities-based decentralized optimal power allocation algorithm for wireless networks. IEEE/ACM Trans. Netw. 17, 1819-1831 (2009)

    Article  Google Scholar 

  35. Bauschke, HH, Borwein, JM: On projection algorithms for solving convex feasibility problems. SIAM Rev. 38(3), 367-426 (1996)

    Article  MATH  MathSciNet  Google Scholar 

  36. Opial, Z: Weak convergence of the sequence of successive approximation for nonexpansive mappings. Bull. Am. Math. Soc. 73, 591-597 (1967)

    Article  MATH  MathSciNet  Google Scholar 

Download references

Acknowledgements

I am sincerely grateful to the associate editor Lai-Jiu Lin and the two anonymous reviewers for helping me improve the original manuscript. This work was supported by the Japan Society for the Promotion of Science through a Grant-in-Aid for Scientific Research (C) (15K04763).

Author information

Authors and Affiliations

Authors

Corresponding author

Correspondence to Hideaki Iiduka.

Additional information

Competing interests

The author declares that they have no competing interests.

Rights and permissions

Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Iiduka, H. Parallel computing subgradient method for nonsmooth convex optimization over the intersection of fixed point sets of nonexpansive mappings. Fixed Point Theory Appl 2015, 72 (2015). https://doi.org/10.1186/s13663-015-0319-0

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s13663-015-0319-0

MSC

Keywords