In this paper, we first introduce a new algorithm which involves projecting each iteration to solve a split feasibility problem with paramonotone equilibria and using unconstrained convex optimization. The strong convergence of the proposed algorithm is presented. Second, we also revisit this split feasibility problem and replace the unconstrained convex optimization by a constrained convex optimization. We introduce some algorithms for two different types of objective function of the constrained convex optimization and prove some strong convergence results of the proposed algorithms. Third, we apply our algorithms for finding an equilibrium point with minimal environmental cost for a model in electricity production. Finally, we give some numerical results to illustrate the effectiveness and advantages of the proposed algorithms.
Hinweise
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
1 Introduction and the problem statement
Let \(H_{1}\) and \(H_{2}\) be two real Hilbert spaces with inner product \(\langle\cdot,\cdot\rangle\) and reduced norm \(\|\cdot\|\), C and Q be nonempty closed convex subsets in \(H_{1}\) and \(H_{2}\), respectively.
In [7], Censor and Elving first introduced the split feasibility problem (shortly, SFP) in Euclidean space, which is formulated as follows:
$$ \text{Find } x^{*}\in C \text{ such that } Ax^{*}\in Q, $$
where \(A:H_{1}\rightarrow H_{2}\) is a bounded linear operator. The SFP can be a model for many inverse problems where constraints are imposed on the solutions in the domain of the linear operator as well as in its range. It has a variety of specific applications in real world such as medical care, image reconstruction and signal processing (see [5, 14‐17, 21, 38, 39] for more details).
Anzeige
Let \(f:C\times C \rightarrow\mathbb{R}\) be a bifunction such that \(f(x,x) = 0\) for all \(x\in C\). The equilibrium problem (shortly, EP)
$$ \text{Find } x^{*}\in C \text{ such that } f\bigl(x^{*},y\bigr)\geq0 \text{ for all } y\in C $$
was firstly introduced by Fan [19] and further studied by Blum and Oettli [2]. The solution set of the EP is denoted by \(\operatorname{Sol}(E P)\). The EP is a generalization of many mathematical models, including variational inequality, fixed point, optimization, complementarity problems (see, for instance, [2, 4, 13, 18, 20, 22, 23, 25, 31, 33‐36]).
Recently, Yen et al. [37] investigated the following split feasibility problem involving paramonotone equilibria and convex optimization (shortly, SEO):
Problem 1.1
Find \(x^{*}\in C\) such that \(f(x^{*},y)\geq0\) for all \(y\in C\) and \(g(Ax^{*})\leq g(z)\) for all \(z\in H_{2}\),
where g is a proper lower semi-continuous convex function on \(H_{2}\). Also, they introduced the following algorithm to solve Problem 1.1:
Anzeige
Algorithm 1.1
For any \(x^{k}\in C\), take \(\eta^{k}\in\partial_{2}^{\epsilon _{k}}f(x^{k},x^{k})\) and define
$$\alpha_{k}=\frac{\beta_{k}}{\gamma_{k}}, $$
where \(\gamma_{k}=\max\{\delta_{k},\|\eta^{k}\|\}\). Compute
In Algorithm 1.1, \(\operatorname{prox}_{\lambda g}\) denotes proximal mapping of the convex function g with \(\lambda> 0\), and the parameters \(\{a_{k}\}\), \(\{\delta_{k}\}\), \(\{\beta_{k}\}\), \(\{\epsilon_{k}\}\) and \(\{\rho_{k}\}\) are taken as in Algorithm 3.1 (see below Sect. 3).
Note that Algorithm 1.1 involves two exact projections onto the feasible set C, which limits the applicability of the method, especially when such projections are hard to compute. It is well known that only in a few specific instances the projection onto a convex set has an explicit formula. When the feasible set C is a general closed convex set, we must solve a nontrivial quadratic problem in order to compute the projection onto C.
In this paper, by expanding the domain of function f, we introduce a new algorithm which just involves a projection onto C. Also, we revisit Problem 1.1 and replace the unconstrained convex optimization by a constrained convex optimization. Further, we introduce two iterative algorithms to solve the new model and prove some strong convergence results of the proposed algorithms.
The paper is organized as follows: Sect. 2 deals with some definitions and lemmas for the main results in this paper. In Sect. 3, we introduce a new algorithm, which involves a projection in each iteration. In Sect. 4, we introduce two algorithms and study their convergence. In Sect. 5, we provide a practical model for an electricity market and some computational results for the model.
2 Preliminaries
The following definitions and lemmas are useful for the validity and convergence of the algorithms.
Definition 2.1
Let H be a Hilbert space, \(T:H\rightarrow H\) be a mapping and let \(K\subseteq H\).
The composition of finitely many averaged mappings is averaged. In particular, if\(T_{1}\)is\(\alpha_{1}\)-averaged and\(T_{2}\)is\(\alpha_{2}\)-averaged, where\(\alpha_{1},\alpha_{2}\in(0,1)\), then the composition\(T_{1}T_{2}\)isα-averaged, where\(\alpha=\alpha _{1}+\alpha_{2}-\alpha_{1}\alpha_{2}\).
It is easy to show that firmly nonexpansive mappings are \(\frac{1}{2}\)-averaged, and averaged mappings are nonexpansive.
For a mapping \(T:H\rightarrow H\), \(\operatorname{Fix}(T)\) denotes the set of fixed points of T, i.e.,
$$ \operatorname{Fix}(T):=\{x\in H: Tx=x\}. $$
It is well known that every nonexpansive operator \(T:H\rightarrow H\) satisfies the following inequality:
for all \(x\in H\) and \(y\in\operatorname{Fix}(T)\) (see, for example, [11, Theorem 3], [12, Theorem 1]).
Let H be a real Hilbert space and K be a nonempty convex closed subset of H. For each point \(x\in{H}\), there exists a unique nearest point in K, denoted by \(P_{K}(x)\), such that
for all \(y\in K\). The mapping \(P_{K}:H\rightarrow K\) is called the metric projection of H onto K. It is well known that \(P_{K}\) is a nonexpansive mapping of H onto K and even a firmly nonexpansive mapping. So, \(P_{K}\) is also \(\frac{1}{2}\)-averaged, which is captured in the following lemma:
Lemma 2.3
For any\(x,y\in H\)and\(z\in K\), the following hold:
Let\(\{v^{k}\}\)and\(\{\delta_{k}\}\)be the nonnegative sequences of real numbers satisfying\(v^{k+1}\leq v^{k}+\delta_{k}\)with\(\sum_{k=1}^{\infty}\delta_{k}<+\infty\). Then the sequence\(\{v^{k}\}\)is convergent.
Lemma 2.7
LetHbe a real Hilbert space, \(\{a_{k}\}\)be a sequence of real numbers such that\(0< a< a_{k}< b<1\)for all\(k\geq1\)and\(\{v^{k}\}\), \(\{w^{k}\}\)be the sequences inHsuch that
It is well known that \(\partial\iota_{K}(x)=N_{K}(x)\) and \((I+\lambda N_{K})^{-1}=P_{K}\) for any \(\lambda>0\).
Let \(f:H\times H\rightarrow\mathbb{R}\) be a bifunction. We need the following assumptions on \(f(x,y)\) for our algorithms and convergence:
(A1) For each \(x\in C\), \(f(x,x)=0\) and \(f(x,\cdot)\) is lower semi-continuous and convex on C;
(A2) \(\partial_{2}^{\epsilon}f(x,x)\) is nonempty for any \(\epsilon> 0\) and \(x\in C\) and is bounded on any bounded subset of C, where \(\partial_{2}^{\epsilon}f(x,x)\) denotes ϵ-subdifferential of the convex function \(f(x,\cdot)\) at x, that is,
(A3) f is pseudo-monotone on C with respect to every solution of the EP, that is, \(f(x,x^{*})\leq0\) for any \(x\in C\), \(x^{*}\in\operatorname{Sol}(EP)\) and f satisfies the following condition, which is called the para-monotonicity property:
$$ x^{*}\in\operatorname{Sol}(EP),\qquad y\in C, \qquad f \bigl(x^{*},y\bigr)=f\bigl(y,x^{*}\bigr)=0 \quad\Longrightarrow\quad y\in\operatorname{Sol}(EP); $$
(A4) For all \(x \in K\), \(f(\cdot,x)\) is weakly upper semi-continuous on C.
3 A new algorithm for Problem 1.1 and its convergence analysis
In this section we give a new algorithm for Problem 1.1 and analyze its convergence.
Recall that the proximal mapping of the convex function g with \(\lambda> 0\), denoted by \(\operatorname{prox}_{\lambda g}\), is defined as the unique solution of the strongly convex programming problem:
The proximal mapping has some good properties, namely, it is firmly nonexpansive and \(\operatorname{prox}_{\lambda g}=P_{Q}\) when \(g=\delta _{Q}\) (see, e.g., [26]).
By using the necessary and sufficient optimality condition for convex programming, we can see that \(h(x) = 0\) if and only if Ax solves \(P(u)\) with \(u = Ax\). Note that, even though g may not be differentiable, h is always differentiable and \(\nabla h(x) = A^{*}(I-\operatorname{prox} _{\lambda g})Ax\) (see, for example, [28]).
Algorithm 3.1
Take positive parameters δ, ξ and the real sequences \(\{a_{k}\}\), \(\{\delta_{k}\}\), \(\{\beta_{k}\}\), \(\{\epsilon_{k}\}\), \(\{\rho_{k}\}\) satisfying the following conditions: for each \(k\in\mathbb{N}\),
where \(A_{k}=2(1-a_{k})(\alpha_{k}\epsilon_{k}+\beta_{k}^{2})\). This completes the proof. □
Theorem 3.1
Suppose that Problem1.1admits a solution. Then, under Assumptions (A1)–(A4), the sequence\(\{x^{k}\}\)generated by Algorithm3.1strongly converges to a solution of Problem1.1.
Proof
Claim 1. The sequence \(\{\|x^{k}-y\|^{2}\}\) is convergent for all \(y\in S\). Indeed, let \(y\in S\). Since \(y\in \operatorname{Sol}(E P)\) and f is pseudomonotone on C with respect to every solution of \((E P)\), we have
where \(A_{k}=2(1-a_{k})(\alpha_{k}\epsilon_{k}+\beta_{k}^{2})\). Since \(\alpha_{k}=\frac{\beta_{k}}{\gamma_{k}}\) with \(\gamma_{k}=\max\{ \delta_{k},\|\eta_{k}\|\}\),
Now, using Lemma 2.6, we see that \(\{\|x^{k}-y\|^{2}\}\) is convergent for all \(y\in S\). Hence the sequence \(\{x^{k}\}\) is bounded. Then, by Lemma 3.2, we can see that \(\{y^{k}\}\) is bounded too.
Claim 2. \(\limsup_{k\rightarrow\infty}f(y^{k},y)=0\) for all \(y\in S\). By Lemma 3.3, for each \(k\geq1\), we have
On the other hand, using Assumption (A2) and the fact that \(\{x^{k}\}\) is bounded, we see that \(\{\|\eta_{k}\|\}\) is bounded. Then there exists \(L>\delta\) such that \(\|\eta_{k}\|\leq L\) for each \(k\geq1\). Therefore, we have
and \(y^{*}\) is a weakly cluster point of \(\{y^{k_{j}} \}\). Then \(y^{*}\) belongs to \(\operatorname{Sol}(EP)\).
Without loss of generality, we can assume that \(\{y^{k_{j}}\}\) weakly converges to \(y^{*}\) as \(j\rightarrow\infty\). Since \(f(\cdot,y)\) is upper semi-continuous, by Claim 2, we have
Since \(y\in S\) and f is pseudomonotone, we have \(f(y^{*},y)\leq0\) and so \(f(y^{*},y)=0\). Again, by pseudomonotonicity of f, \(f(y,y^{*}) \leq0\) and hence \(f(y^{*},y)=f(y,y^{*})=0\). Then, by paramonotonicity (Assumption (A3)), we can conclude that \(y^{*}\) is also a solution of \((EP)\).
Claim 4. Every weakly cluster point x̄ of the sequence \(\{x^{k}\}\) satisfies \(\bar{x}\in K\) and \(A\bar{x}\in \mathop {\operatorname {argmin}}g\). Let x̄ be a weakly cluster point of \(\{x^{k}\}\) and \(\{x^{k_{j}}\}\) be a subsequence of \(\{x^{k}\}\) weakly converging to x̄. Then \(\bar{x}\in K\). From Lemma 3.3, if \(\nabla h(x^{k})\neq0\), then we have
Moreover, since ∇h is Lipschitz continuous with constant \(\|A\|^{2}\), we see that \(\|\nabla h(x^{k})\|^{2}\) is bounded. So, \(h(x^{k})\rightarrow0\) as \(k\in N_{1}\) and \(k\rightarrow\infty\). Note that \(h(x^{k})=0\) for \(k\notin N_{1}\). Consequently, we have
which implies that Ax̄ is a fixed point of the proximal mapping of g. Thus Ax̄ is a minimizer of g. From (8) and the fact that \(\|\nabla h(x^{k})\|^{2}\) is bounded, it follows that
Claim 5. \(\lim_{k\rightarrow+\infty}x^{k}= \lim_{k\rightarrow+\infty}y^{k}=\lim_{k\rightarrow+\infty}P(x^{k})=x ^{*}\), where \(x^{*}\) is a weakly cluster point of the sequence satisfying (7). From Claims 3 and 4, we can deduce that \(x^{*}\) belongs to S. By Claim 1, we can assume that
Employing arguments, similar to those used in the proof of Theorem 1 in [37], we have
$$ \lim_{k\rightarrow+\infty}x^{k}=x^{*}. $$
This completes the proof. □
4 Algorithms and convergence analysis
In [37], Yen et al. presented an application of Problem 1.1 to a model of electricity production, in which z denotes the quantity of the materials and \(g(z)\) is the total environmental fee that companies have to pay for environmental pollution while using materials z for production. So, from \(x\in C\), it follows that \(z=Ax\in\{z: z=Ax, x\in C\}\).
However, in actual production, since the resources are limited, there are usually stricter constraints on the quantity of the materials such as \(z\in Q\), where Q is a nonempty closed convex set of \(H_{2}\). Therefore, it is necessary to replace the unconstrained convex optimization problem \(\min_{x\in H_{2}}g(x)\) with the constrained convex optimization as follows:
$$ \min_{x\in Q}g(x), $$
(9)
whose solution is denoted by \(\operatorname{Sol}(Q,g)\). By using (9), Problem 1.1 becomes the following problem:
Problem 4.1
Find \(x^{*}\in C\) such that \(f(x^{*},y)\geq0\) for all \({y}\in C\) such that \(Ax^{*}\in Q\) and \(g(Ax^{*})\leq g(z)\) for all \(z\in Q\),
Throughout this paper, we assume \(\varGamma\neq\emptyset\).
In this section, we discuss two cases that the function g is differentiable or non-differentiable. The corresponding algorithms and their convergence are provided next.
4.1 The case when g is differentiable
We need to make the following assumption on the mapping g:
(B) g is L-Lipschitz differentiable with \(L>0\), i.e.,
It is easy to verify that the constrained convex optimization problem (9) is equivalent to the following variational inequality problem:
$$ \bigl\langle \nabla g\bigl(z^{*}\bigr),z-z^{*} \bigr\rangle \geq0 \quad\text{for all } z\in Q, $$
(10)
and the variational inequality problem (10) is equivalent to the following fixed point problem:
$$ z^{*}=P_{Q}(I-\nu\nabla g)z^{*}, $$
(11)
where \(\nu>0\). So, the constrained optimization problem (9) and the fixed point problem (11) are equivalent. From the optimality condition of (9), we can also deduce the equivalence of problems (9) and (11) (see [24]). Next we construct an iterative algorithm based on this equivalence and Algorithm 3.1.
Assume that a mapping\(g:H_{2}\rightarrow H_{2}\)satisfies Assumption (B) and\(\nu\in(0,\frac{2}{L})\). Let\(y\in\varGamma\). If\(\|l(x^{k})\|\neq0\), then it follows that
Let \(T=P_{Q}(I-\nu\nabla g)\). Since \(y\in\varGamma\), it follows from (11) that Ay is a fixed point of T. From the proof of [32, Theorem 4.1], it follows that T is \(\frac{2+\nu L}{4}\)-averaged and so it is nonexpansive. By (13) and Lemma 2.3(i), we have
Next we establish the convergence of Algorithm 4.1.
Theorem 4.1
Under Assumptions (A1)–(A4) and (B), the sequence\(\{x^{k}\}\)generated by Algorithm4.1strongly converges to a solution of Problem4.1.
The proof of Theorem 4.1 is similar with that of Theorem 3.1, so here we omit it.
The only thing to note about the proof of Theorem 4.1 is that from \(h(\bar{x})=0 \) it follows that Ax̄ is a fixed point of \(P_{Q}(I-\nu\nabla g)\). Thus Ax̄ is a solution of (9).
4.2 The case when g is non-differentiable
Let \(g:Q\rightarrow\mathbb{R}\cup\{+\infty\}\) be a proper convex lower semi-continuous function. Denote by
the Moreau–Yosida approximate of the function g with the parameter λ. It is easy to see that the solution of (16) converges to that of \(\min_{x\in Q}g(x)\) as \(\lambda\rightarrow\infty\).
For the mapping defined in (16), we have following result.
Note that the differentiability of the Yosida-approximate \(g_{\lambda }\) (see, for instance, [28]) secures the additivity of the subdifferentials, and so we can write
Similar to Algorithm 3.1, using Lemma 4.3, we introduce the following algorithm:
Algorithm 4.2
Take the real sequences \(\{a_{k}\}\), \(\{\delta_{k}\}\), \(\{\beta_{k}\}\), \(\{\epsilon_{k}\}\) and \(\{\rho_{k}\}\) as in Algorithm 3.1. Take a positive parameter ν.
when \(Q=H_{2}\). So, Algorithm 3.1 is a special case of Algorithm 4.2.
We need the following lemmas for the proof of the convergence of Algorithm 4.2.
Lemma 4.4
Let\(\nu\in(0,1]\). Then operator\(P_{Q}(I-\nu(I-\operatorname{prox}_{\lambda g}))\)is nonexpansive.
Proof
By the fact that \(\operatorname{prox}_{\lambda g}\) is firmly nonexpansive and Lemma 2.1, \(I-\operatorname{prox}_{\lambda g}\) and \(\nu(I-\operatorname{prox}_{\lambda g})\) are also firmly nonexpansive. So, using Lemma 2.1 again, \(I-\nu(I-\operatorname{prox}_{\lambda g})\) is firmly nonexpansive. Thus, from Lemma 2.2, it follows that \(P_{Q}(I-\nu (I-\operatorname{prox} _{\lambda g}))\) is \(\frac{3}{4}\)-averaged and hence nonexpansive. This completes the proof. □
Using Lemma 4.4 and following the proof of Theorem 4.1, we obtain the convergence result of Algorithm 4.2.
Theorem 4.2
Let\(\nu\in(0,1]\). Then, under Assumptions (A1)–(A4), the sequence\(\{x^{k}\}\)generated by Algorithm4.2strongly converges to a solution of Problem4.1.
The proof of Theorem 4.2 is similar to that of Theorem 3.1, so here we omit it.
One thing to note about proof of Theorem 4.2 is that from \(h( \bar{x})=0 \) it follows that Ax̄ is a fixed point of \(P_{Q}(I-\nu(I-\operatorname{prox}_{\lambda g}))\). Thus, by Lemma 4.3, Ax̄ is a solution of (17).
5 Numerical examples
In this section, we provide two numerical examples to compare different algorithms. All programs are written in Matlab version 7.0 and performed on a desktop PC with Intel(R) Core(TM) i5-4200U CPU @ 2.30 GHz, RAM 4.00 GB.
Example 5.1
First, we consider an equilibrium-optimization model which was investigated by Yen et al. [37]. This model can be regarded as an extension of a Nash–Cournot oligopolistic equilibrium model in electricity markets. The latter model has been investigated in some research papers (see, for example, [10, 27]).
In this equilibrium model, it is assumed that there are n companies. Let x denote the vector whose entry \(x_{i}\) stands for the power generated by company i. Following Contreras et al. [10], we suppose that the price \(p_{i}(s)\) is a decreasing affine function of s with \(s = \sum_{i=1}^{n} x_{i}\), that is,
$$ p_{i}(s) = \alpha-\beta_{i} s. $$
Then the profit made by company i is given by
$$ f_{i}(x) = p_{i}(s)x_{i}-c_{i}(x_{i}), $$
where \(c_{i}(x_{i})\) is the cost for generating \(x_{i}\) by the company i.
Suppose that \(C_{i}\) is the strategy set of company i, that is, condition \(x_{i}\in C_{i}\) must be satisfied for each i. Then the strategy set of the model is \(C:=C_{1}\times C_{2}\times\cdots \times C_{n}\).
Actually, each company seeks to maximize its profit by choosing the corresponding production level under the presumption that the production of the other companies are parametric input. A commonly used approach to this model is based upon the famous Nash equilibrium concept.
Now, we recall that a point \(x^{*}\in C = C_{1}\times C_{2}\times \cdots\times C_{n}\) is an equilibrium point of the model if
for all \(x_{i}\in C_{i}\) and \(i=1,2,\dots,n\), where \(x^{*}[x_{i}]\) stands for the vector obtained from \(x^{*}\) by replacing \(x_{i}^{*}\) with \(x_{i}\). By taking
the problem of finding a Nash equilibrium point of the model can be formulated as follows:
$$ \text{Find } x^{*}\in C \text{ such that } f\bigl(x^{*},x\bigr)\geq0 \text{ for all } x\in C. $$
(EP)
In [37], Yen et al. extended this equilibrium model by additionally assuming that the companies use some materials to produce electricity.
Let \(a_{l,i}\) denote the quantity of material l\((l = 1,\dots, m)\) for producing one unit of electricity by company i\((i=1, \dots,n)\). Let A be the matrix whose entries are \(a_{l,i}\). Then entry l of the vector Ax is the quantity of material l for producing x. Using materials for production may cause environmental pollution, for which companies have to pay a fee. Suppose that \(g(Ax)\) is the total environmental fee for producing x.
The task now is to find a production \(x^{*}\) such that it is a Nash equilibrium point with a minimum environmental fee, while the quantity of the materials satisfies constraint Q. This problem can be formulated as the split feasibility problem of the following form:
$$\begin{aligned} &\text{Find } x^{*}\in C \text{ such that } f\bigl(x^{*},x\bigr)\geq0 \text{ for all } x\in C \\ &\quad\text{and } g \bigl(Ax^{*}\bigr)\leq g(Ax) \text{ for all } Ax\in Q. \end{aligned}$$
(SEP)
Suppose that, for every i, cost \(c_{i}\) for production and environmental fee g are increasing convex functions. The convexity assumption here means that both the cost and fee for producing a unit of product increase as the quantity of the product gets larger.
Under this convexity assumption, it is not hard to see (see also Quoc et al. [27]) that problem (EP) with f given by (22) can be formulated as follows:
$$ \text{Find } x^{*}\in C \text{ such that } f\bigl(x^{*},x\bigr):=\bigl\langle \tilde{B}_{1}x^{*}-\bar{a}, x-x^{*}\bigr\rangle + \varphi(x)-\varphi\bigl(x^{*}\bigr)\geq0 \text{ for all } x\in C, $$
(23)
where
(24)
Note that, when \(c_{i}\) is differentiable and convex for each i, problem (23) is equivalent to the following variational inequality problem:
$$ \text{Find } x^{*}\in C \text{ such that } \bigl\langle \tilde{B}_{1}x^{*}-\bar{a}+ \nabla\varphi\bigl(x^{*}\bigr), x-x^{*} \bigr\rangle \geq0 \text{ for all } x\in C. $$
We tested the proposed algorithm with the cost function given by
where \(p_{i}\geq0\). In [37], Yen et al. showed that function \(f(x,y)\) defined by (23), (24) and (25) satisfies Assumptions (A1), (A2) and (A4).
In [37], the author denoted by \(g(z)\) the total environmental fee. It is unreasonable. Firstly, the total environmental fee should be included in the cost, that is, it is a part of \(c_{i}(x_{i})\). Secondly, it is supposed that the companies behave as players in an oligopolistic market, but at the same time they are subordinated to the centralized planning decision in order to minimize the total environmental fee for the whole system. That is, the model is not concordant with the real system behavior.
It may be reasonable to denote by \(g(z)\) the restriction for the emission of contaminants. To protect the environment, governments generally adopt policies to restrict emissions of contaminants.
Assume that the production of electricity brings p contaminants and governments require that the quantity of contaminants brought by the production of one unit of electricity is in a given region. We use a set \(K\subset\mathbb{R}^{p}\) to denote this region.
Let \(b_{k,l}\) denote the quantity of the contaminant k\((k = 1, \dots, p)\) for consuming one unit of material l\((l = 1,\dots, m)\). Let B be the matrix whose entries are \(b_{k,l}\). Then entry k of vector Bz is the quantity of contaminant k for consuming one unit of material \(z_{l}\)\((l = 1,\dots, m)\). So, the quantity of contaminant k\((k = 1,\dots, p)\) for producing one unit of electricity is entry k of \(BAx\), and \(BAx\) should be in the set K, i.e., \(BAx\in K\). We get \(Bz\in K\) when letting \(z=Ax\).
for each \(k\geq1\) and take \(\nu=\frac{1.99}{\|B\|}\). The entries of matrix A were randomly generated in the interval [0, 5]. In the bifunction \(f(x,y)\) defined by (23), (24) and (25), the parameters \(\alpha=0.5\) and \(b_{i}\), \(p_{i}\) and \(q_{i}\) for each \(i = 1,\dots, n\) were generated randomly in the interval \((0, 1]\), \([1, 3]\), and \([1, 3]\), respectively. In the function \(g(z)\), we take \(B\in\mathbb{R}^{p}\times\mathbb{R}^{m}\), and its elements are generated randomly in \((0,1)\).
Since function \(g(z)\) is differentiable, we use Algorithm 4.1 to solve Problem 4.1 and compare it with Algorithms 1.1 and 3.1. In Algorithms 1.1 and 3.1 we substitute \(\operatorname{prox}_{\lambda g}\) with \(I-\nu\nabla g\) and do not consider the constraint set Q.
The computational results are shown in Figs. 1 and 2. The horizontal and vertical axes show iteration k, as well as error1\((k):= \|x^{k}-x ^{k-1}\|\) and error2\((k):= \|Ax^{k}-P_{Q}(Ay^{k})\|\), respectively. We solve the model with \(m=15\) and take \(n=10\) as the number of companies.
From Figs. 1 and 2, we have two conclusions as follows:
(a) The “error1” of Algorithm 4.1 is smaller than that of Algorithms 1.1 and 3.1 and the “error1” of Algorithm 3.1 is slightly smaller than that of Algorithm 1.1.
(b) The “error2” of Algorithm 4.1 decreases with the iteration number k, while the “error2” of Algorithms 1.1 and 3.1 increases with the iteration number k. The “error2” of Algorithm 4.1 is smaller than those of Algorithms 1.1 and 3.1.
Next we give a numerical procedure in an infinite-dimensional space and compare Algorithm 4.1 with a numerical algorithm which is based on the Halpern modification of [8, Algorithm 6.1] as follows:
where \(T:=P_{Q}(I-\lambda\nabla g)\), \(U:=P_{C}(I-\lambda f)\), and \(\gamma\in(0,1/L)\), L is the spectral radius of the operators \(A^{*}A\), denoted by \(\rho(A^{*}A)\). The parameter λ depends on the constants of the inverse strong monotonicity of ∇g and f.
According to the condition of the convergence of Halpern-type algorithm, we assume that \(\lim_{k\rightarrow\infty} \tau_{k}=0\) and \(\sum_{k=1}^{\infty}\tau_{k}=\infty\).
Example 5.2
Suppose that \(H = L^{2}([0,1])\) with norm \(\|x\|:= (\int_{0}^{1} |x(t)|^{2}\,dt )^{\frac{1}{2}}\) and inner product \(\langle x,y\rangle:= \int_{0}^{1} x(t)y(t)\,dt\), \(x,y \in H\). Let \(C:=\{x \in H:\|x\|\leq1\}\) be the unit ball, \(Q:=\{x\in H:\langle x(t),\sin(10x(t))\rangle\leq1\}\). Define an operator \(F:C\rightarrow H\) by
As shown in [30], F is monotone and L-Lipschitz-continuous with \(L=2\). Let \(f(x(t), y(t))=\langle Fx(t),y(t)-x(t)\rangle\), \(g(x)(t)=\frac{1}{2}\|x(t)\|^{2}\) and \((Ax)(t)=3x(t)\) for all \(x\in H\).
Let \(x^{1}(t)=1\). Take the sequences \(\{\alpha_{k}\}\), \(\{\beta_{k}\}\), \(\{\epsilon_{k}\}\), \(\{\delta_{k}\}\) of the parameters as follows:
for each \(k\geq1\) and take \(\nu=\frac{1.99}{L_{g}}\). We take \(\lambda=1\) according to the numerical tests since the constants of the inverse strong monotonicity of ∇g and f are unknown. Take \(\tau_{k}=\frac{1}{k+1}\) and \(\gamma=\frac{0.9}{\rho(A^{*}A)}\) for Algorithm 5.1. We use \(\operatorname{error}=\frac{1}{2}\| P_{C}(x^{k})-x^{k}\|^{2}+ \frac{1}{2}\|P_{Q}(Ax^{k})-Ax^{k}\|^{2}\) to measure the error of the kth iteration.
Numerical results are given in Fig. 3, which illustrate that Algorithm 4.1 behaves better than Algorithm 5.1.
We first introduce a new algorithm, which involves a projection of each iteration, and show its strong convergence. We also improve the model proposed in [37] by adding a constraint to the minimization problem of the total environmental fee. Two algorithms are introduced to approximate the solution and their strong convergence is analyzed.
Acknowledgements
We sincerely thank Prof. S. He for his helpful discussion and the reviewers for their valuable suggestions and useful comments that have led to the present improved version of the original manuscript.
Availability of data and materials
Data sharing not applicable to this article as no datasets were generated during the current study.
Competing interests
The authors declare that they have no competing interests.
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.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
