1. Introduction
Let E 1 and E 2 be two real Banach spaces with duals $E_{1}^{*}$ and $E_{2}^{*},$ respectively, and let C and Q be nonempty closed and convex subsets of E 1 and $E_2,$ respectively. Let $A:E_{1}\to E_{2}$ be a bounded linear operator. Censor and Elving [Reference Censor and Elfving7] introduced the concept of split feasibility problem (SFP), which is formulated as
The SFP has been found useful in solving numerous real-life problems, including medical image reconstruction, phase retrieval, signal processing, radiation therapy treatment planning, among others. See, for example [Reference Godwin, Izuchukwu and Mewomo17, Reference Godwin, Izuchukwu and Mewomo18, Reference Kassay, Reich and Sabach24] and the references therein.
For solving the SFP (1.1) using Bregman projection, Schöpfer et al. [Reference Schöpfer, Schuster and Louis40] proposed the following method in the framework of p-uniformly convex real Banach spaces: for $x_1 \in E_1$, set
where $\Pi_{C}$ denotes the Bregman projection from E 1 onto C and $J_{E}^{p}$ is the duality mapping.
Closely related to the SFP (1.1) is the following split variational inequality problem (SVIP) introduced by Censor et al. [Reference Censor, Gibali and Reich9] in the framework of real Hilbert spaces as follows: find $x^* \in C$, which satisfies the inequality
such that $y^*=Ax^* \in Q$ solves the inequality
where C and Q are nonempty, closed and convex subsets of real Hilbert spaces H 1 and $H_2,$ respectively, $F:H_1 \to H_1$ and $G:H_2 \to H_2$ are two given operators, $A:H_1 \to H_2$ is a bounded linear operator.
Recall that when problems (1.3) and (1.4) are viewed separately, then Equation (1.3) is the classical variational inequality problem (VIP) in H 1 with its solution set ${\rm VIP}(C, F)$ and Equation (1.4) is another VIP in H 2 with its solution set ${\rm VIP}(Q, G).$ To solve the SVIP (1.3) and (1.4), Censor et al. [Reference Censor, Gibali and Reich9] put forward the following algorithm. Let $x_1 \in H_1,$ the sequence $\{x_n\}$ is generated by
where $\gamma \in (0, {1}/{L})$ and L is the spectral radius of the operator $A^*A$. They proved that the above Algorithm (1.5) converges weakly to a solution of the SVIP under the assumption that $U, V$ are $\alpha_1,\alpha_2$-inverse strongly monotone operators and $\lambda\in(0,2\alpha)$ (where $\alpha:=\min\{\alpha_1,\alpha_2\}$).
Let C be a nonempty, closed and convex subset of a real Banach space E with dual $E^*$. Let $f:C\times C\to \mathbb{R}$ be a bifunction. The equilibrium problem (EP) studied by Blum and Oettli [Reference Blum and Oettli5] is to locate a point $x^*\in C$ such that
We denote by EP(f) the set of solutions of Equation (1.6). The EP was formerly introduced as the Ky Fan inequality [Reference Fan and Shisha15]. This class of problem has been extensively studied by numerous scholars because of its several applications. It is well known that many problems arising in economics, optimization and physics can be reduced to problem (1.6). Moreover, several iterative algorithms have been proposed to solve the EP (1.6) and related optimization problems in both Hilbert and Banach spaces (see [Reference Alakoya and Mewomo1, Reference Alakoya, Uzor and Mewomo3, Reference Blum and Oettli5, Reference Combettes and Hilstoaga13, Reference Godwin, Alakoya, Mewomo and Yao16, Reference Godwin, Izuchukwu and Mewomo17, Reference Ma, Wang and Cho28, Reference Moudafi31, Reference Owolabi, Alakoya, Taiwo and Mewomo36, Reference Ogwo, Izuchukwu and Mewomo34, Reference Reich and Sabach39, Reference Uzor, Alakoya and Mewomo45] and other references contained therein). However, most of the existing results on the EP are of the monotone type.
In 2011, Moudafi [Reference Moudafi30] extended the SVIP Equations (1.3) and (1.4) to split equilibrium problem (shortly, SEP), which is defined as
where C and Q are nonempty, closed and convex subsets of real Hilbert spaces H 1 and $H_2,$ respectively, $f_1:C \times C \rightarrow \mathbb{R}$ and $f_2:Q \times Q \rightarrow \mathbb{R}$ are bifunctions with a bounded linear operator $A: H_1 \rightarrow H_2$.
In addition, for solving the SEP (1.7), the author in [Reference He20] put forward a proximal technique without product space formulation as follows:
where $A^*$ is the adjoint of A. The author in [Reference He20] obtained a weak convergence result when the bifunctions f 1 and f 2 are monotone on C and Q, respectively. Since then, several iterative schemes have been proposed when the bifunctions are either monotone or pseudomonotone, see for example [Reference Anh4, Reference Eskandani, Raeisi and Rassias14, Reference Hieu21–Reference Hieu23, Reference Kim and Dinh25] and check also the references therein.
In Section 6.1 of Censor et al. [Reference Censor, Gibali and Reich9], the authors proposed an improvement of the SVIP (1.3) and (1.4), which they called multiple set SVIP (MSSVIP), which is formulated as follows:
where $A:H_{1}\to H_{2}$ is a bounded linear operator, $F_{i}:H_1\to H_1$, $i=1,2,\ldots,N$ and $G_j:H_2\to H_2$, $j=1,2,\ldots,M$ are given operators and $C_i, \ i=1,2,\ldots,N$ and $Q_j,\ j=1,2,\ldots,M$ are nonempty, closed and convex subsets of real Hilbert spaces H 1 and $H_2,$ respectively. They proposed an algorithm for solving the problem and proved that the sequence generated by the proposed iterative scheme converges weakly to the solution set of MSSVIP when $F_{i},\ i=1,2,\ldots,N$ and $G_j, \ j=1,2,\ldots,M$ are inverse strongly monotone operators.
Moudafi [Reference Moudafi31] proposed a new SFP, which he called split equality problem (SEqP). Let H 1, H 2 and H 3 be real Hilbert spaces. $A: H_1\rightarrow H_3$ and $B:H_2 \rightarrow H_3$ be two bounded linear operators, and $C \subset H_1$ and $Q \subset H_2$ be two nonempty, closed and convex sets. The SEqP is formulated as follows:
Let $T:E\to E$ be a mapping. We denote the fixed point set of T by $\mbox{F}(T)$; that is $\mbox{F}(T):=\{x\in E: Tx=x\}$. The Fixed Point Problem has application in various fields, such as optimization theory, economics, game theory, as well as in establishing the existence of solutions of several physical problems arising in differential and integral equations [Reference Godwin, Taiwo and Mewomo19, Reference Ogbuisi and Mewomo32, Reference Ogwo, Izuchukwu and Mewomo34, Reference Taiwo and Mewomo42].
If $C:=F(S)$ and $Q:=F(T)$ in Equation (1.8), where $S:H_1\rightarrow H_1$ and $T:H_2\rightarrow H_2$ are two nonlinear mappings, then the SEqP becomes the split equality fixed point problem.
Motivated and inspired by the above mentioned results, we introduce and study in the framework of p-uniformly convex Banach space an extension of the MSSVIP to multiple set split equality equilibrium and common fixed points problem of Bregman quasi-nonexpansive mappings. Using Bregman distance, we make use of the Halpern extragradient technique for solving the pseudomonotone EP, which guarantees strong convergence. We design our algorithm in such a way that it does not depend on the prior estimates of the Lipschitz-like constants.
We organize the rest of this article as follows: Section 2 presents preliminaries and some existing results, $\S$ 3 is the design of our iterative method, whereas $\S$ 4 focuses on the convergence analysis of the proposed algorithm. In $\S$ 5, we apply our result to solve a certain class of variational inequality problems. We present some numerical experiments in $\S$ 6 and conclude with some final remarks in $\S$ 7.
2. Preliminaries
In this section, we call up some important definitions and existing results, which will be needed in the proof of our main result. We denote strong and weak convergence of the sequence $\{x_{n}\}$ to a point x by ‘$\rightarrow$’ and ‘$\rightharpoonup$’, respectively.
Let E be a real Banach space and $1 \lt q\leq 2 \lt p \lt \infty$ with $\frac{1}{p}+\frac{1}{q}=1.$ The modulus of smoothness of E is the function $\rho_E: \mathbb{R}^{+}:=[0, \infty)\rightarrow \mathbb{R}^{+}$ defined by
The space E is called uniformly smooth if and only if $\frac{\rho_E(\tau)}{\tau}\rightarrow 0$ as $\tau \rightarrow 0$. Let q > 1, E is said to be q-uniformly smooth if there exists $\kappa_q \gt 0$ such that $\rho_E(\tau) \leq \kappa_q \tau^{q}$ for all τ > 0. The modulus of convexity of E is defined as
The Banach space E is called uniformly convex if and only if $\beta_E(\epsilon) \gt 0$ for every $\epsilon \in (0,2]$. Now, suppose p > 1, then E is called p-uniformly convex if there exists a constant $C_p \gt 0$ such that $\beta_E(\epsilon) \geq C_p\epsilon^{p}$ for all $\epsilon \in (0,2]$.
Remark 2.1. It is well known that every p-uniformly convex space is also strictly convex and reflexive. In addition, if a Banach space E is p-uniformly convex and uniformly smooth, then its dual space $E^*$ is q-uniformly smooth and uniformly convex (see [Reference Cholamjiak and Sunthrayuth11]).
Definition 2.2. see [Reference Cioranescu12]
Let p > 1 be a real number, the generalized duality mapping $J_{E}^{p}:E \rightarrow 2^{E^*}$ is defined by
where $\langle.,.\rangle$ denotes the duality pairing between elements of E and $E^*$. In particular, $J_E^{p}=J_E^{2}$ is called the normalized duality mapping. If E is p-uniformly convex and uniformly smooth, then $E^*$ is q-uniformly smooth and uniformly convex. In this case, the generalized duality mapping $J_{E}^{p}$ is one-to-one, single-valued and satisfies $J_E^{p}=(J_{E^*}^{q})^{-1}$, where $J_{E^*}^{q}$ is the generalized duality mapping of $E^*$. Furthermore, if E is uniformly smooth, then the duality mapping $J_{E}^{p}$ is norm-to-norm uniformly continuous on bounded subsets of E, and E is smooth if and only if $J_{E}^{p}$ is single valued.
Let $f: E \rightarrow (-\infty, + \infty]$ be a proper, lower semicontinuous and convex function, then the Fenchel conjugate of f denoted as $f^*: E^* \rightarrow (-\infty,+ \infty]$ is defined as
See [Reference Simons and Reich41] for more information about Fenchel conjugate.
Let the domain of f be denoted by $({\rm dom}\, f)=\{x \in E: f(x) \lt +\infty\}$; hence, for any $x \in {\rm int}({\rm dom}\, f)$ and $y \in E$, we define the right-hand derivative of f at x in the direction y by
The function f is said to be Gâteaux differentiable at x if $\lim_{t \rightarrow 0^+}\frac{f(x+ty)-f(x)}{t}$ exists for any $y.$ In this case, $f^0(x,y)$ coincides with $\bigtriangledown f(x)$ (the value of the gradient $\bigtriangledown f$ of f at x). The function f is said to be Gâteaux differentiable if it is Gâteaux differentiable for any $x \in {\rm int}({\rm dom}\, f)$. The function f is said to be Fréchet differentiable at x if its limit is attained uniformly in $||y||=1$. Moreover, f is said to be uniformly Fréchet differentiable on a subset C of E if the above limit is attained uniformly for $x \in C$ and $||y||=1$.
A function f is said to be Legendre if it satisfies the following conditions:
(1) The interior of the domain of f, ${\rm int}({\rm dom}\, f)$ is nonempty, f is Gâteaux differentiable on ${\rm int}({\rm dom}\, f)$ and ${\rm dom}\bigtriangledown f={\rm int}({\rm dom}\, f)$.
(2) The interior of the domain of $f^*$, ${\rm int}({\rm dom}\, f^*)$ is nonempty, $f^*$ is Gâteaux differentiable on ${\rm int}({\rm dom}\, f^*)$ and ${\rm dom} \bigtriangledown f^*={\rm int} ({\rm dom}\, f)$.
Definition 2.3 [Reference Bregman6] Let $f:E \rightarrow (-\infty,+ \infty]$ be a convex and Gâteaux differentiable function. The function $\Delta_f: E \times E \rightarrow [0, + \infty)$ defined by
is called the Bregman distance with respect to f.
We highlight the following interesting properties of Bregman distance (see [Reference Bregman6, Reference Reem, Reich and De Pierro37]):
(i) $\Delta_{f}(x,x)=0,$ but $\Delta_{f}(x,y)=0$ does not necessarily imply that x = y,
(ii) for $x\in {\rm dom}\, f$ and $y, z\in {\rm int}({\rm dom}\, f),$ we have
(2.1)\begin{align} \Delta_{f}(x,y)+\Delta_{f}(y,z)-\Delta_{f}(x,z)= \langle \bigtriangledown g(z)-\bigtriangledown g(y), x-y\rangle, \end{align}(iii) for each $z\in E,$ $\{x_i\}_{i=1}^{N}\subset E$ and $\{\alpha_n\}_{i=1}^{N} \subset (0,1)$ with $\sum\limits_{i=1}^{N}\alpha_i=1,$ we have
\begin{equation*}\Delta_f\left(z,\nabla g^{*}\left(\sum\limits_{i=1}^{N}\alpha_i\nabla g(x_i) \right)\right) \leq \sum\limits_{i=1}^{N}\alpha_i\Delta_f(z,x_i). \end{equation*}
It is well-known that in general the Bregman distance $\Delta_f$ is not a metric because it fails to satisfy the symmetric and triangle inequality properties. Moreover, it is well known that the duality mapping $J_{E}^{p}$ is the sub-differential of the functional $f_p(.)= \frac{1}{p}||.||^{p}$ for p > 1, see [Reference Chidume10]. Then, the Bregman distance $\Delta_p$ is defined with respect to fp as follows:
Definition 2.4. Let C be a nonempty, closed and convex subset of a real Banach space E and let $T:C\to C$ be a nonlinear map. The mapping I ‒ T is said to be demiclosed at zero if for any sequence $\{x_n\}\subset C,$ the following implication holds: $x_{n}\rightharpoonup x$ and $(I-T)x_n\to 0\implies x\in F(T).$
Definition 2.5. [Reference Censor and Reich8, Reference Martin-Marquez, Reich and Sabach29]
Let $T: C \rightarrow {\rm int}({\rm dom}f)$ be a mapping. Then,
(i) a point $p \in C$ is called an asymptotic fixed point of T if C contains a sequence $\{x_n\}$, which converges weakly to p such that $\lim_{n \rightarrow \infty}\|Tx_n-x_n\|=0$. We denote by $\hat{F}(T)$ the set of asymptotic fixed points of T;
(ii) T is called Bregman firmly nonexpansive if
\begin{align*} \langle \Delta_f (Tx)-\Delta_f (Ty), Tx-Ty \rangle\leq \langle\Delta_{f}(x)-\Delta_{f}(y), Tx-Ty\rangle, \quad \forall x,y\in C. \end{align*}(iii) T is called Bregman strongly nonexpansive if $\hat{F}(T)\ne\emptyset$ and
\begin{align*} \Delta_{f}(p, Tx)\leq \Delta_{f} (p, x), \qquad \forall p\in \hat{F}(T) \quad \mbox{and} \quad x\in C. \end{align*}(iv) T is said to be Bregman quasi-nonexpansive if
\begin{align*} F(T) \neq \emptyset\quad \text {and} \quad \Delta_f(p, Tx) \leq \Delta_f(p, x),\quad \forall x \in C,\ p \in F(T). \end{align*}
Recall that the metric projection PC from E onto C satisfies the following property:
It is well known that PC is the unique minimizer of the norm distance. Moreover, PC is characterized by the following property:
The Bregman projection from E onto C denoted by $\Pi_{C}$ also satisfies the property
Also, if C is a nonempty, closed and convex subset of a p-uniformly convex and uniformly smooth Banach space E and $x \in E$, then the following assertions hold (see [Reference Cholamjiak and Sunthrayuth11]):
(i) $z=\Pi_{C}x$ if and only if
(2.5)\begin{align} \langle J_{E}^{p}(x)-J_{E}^{p}(z), y-z\rangle \leq 0, \quad \forall y \in C; \end{align}(ii)
(2.6)\begin{align} \Delta_p(\Pi_{C}x, y)+ \Delta_p(x, \Pi_Cx)\leq \Delta_p(x,y),\quad \forall y \in C. \end{align}
Lemma 2.6. [ Reference Tiel43] Let C be a nonempty convex subset of a Banach space E. Let $g:C\to \mathbb{R}$ be a convex, subdifferentiable function on C. Then g attains its minimum at $x\in C$ if and only if $0\in \partial g(x)+ N_{C}(x),$ where $N_{C}(x)$ is the normal cone of C at x, that is
Lemma 2.7. [ Reference Chidume10] Let E be a Banach space and $x, y \in E$. If E is q-uniformly smooth, then there exists $C_q \gt 0$ such that
Lemma 2.8. [ Reference Cioranescu12] Let f and g be two convex functions on E, such that $x_{0}\in {\rm dom}\, f\cap {\rm dom} g$, where f is continuous, then
Lemma 2.9. [ Reference Kuo and Sahu27] Let E be a real p-uniformly convex and uniformly smooth Banach space. Let $z, x_k \in E (k=1,2,\ldots,N)$ and $\alpha_k \in (0,1)$ with $\sum_{k=1}^{N}\alpha_k=1$. Then, we have
for all $i, j \in \{1,2,\dots,N\}$ and $g^*_r:\mathbb{R}^+ \rightarrow \mathbb{R}^+$ being a strictly increasing function such that $g^*_r(0)=0$.
Lemma 2.10. [ Reference Schöpfer, Schuster and Louis40] Let E be a real p-uniformly convex and uniformly smooth Banach space. Let $V_p: E^* \times E \rightarrow [0, +\infty)$ be defined by
Then the following assertions hold:
(i) Vp is nonnegative and convex in the first variable.
(ii) $\Delta_p \left(x, J_{q}^{E^*}(x^*)\right)= V_p( x, x^*),\quad \forall x \in E,\ x^* \in E^*$.
(iii) $V_p( x, x^*) + \langle J_{q}^{E^*}(x^*)-x, y^*\rangle \leq V_p(x, x^*+y^*),\quad \forall x \in E,\ x^*, y^* \in E^*$.
Lemma 2.11. [ Reference Cholamjiak and Sunthrayuth11] Let E be a real p-uniformly convex and uniformly smooth Banach space. Suppose that $\{x_n\}$ and $\{y_n\}$ are bounded sequences in E. Then the following assertions are equivalent:
(i) $\lim_{n\rightarrow \infty} \Delta_p(x_n, y_n)=0;$
(ii) $\lim_{n\rightarrow \infty}||x_n-y_n||=0$.
Lemma 2.12. [ Reference Xu and Roach46] Let $q\geq 1$ and r > 0 be two fixed real numbers. Then, a Banach space E is uniformly convex if and only if there exists a continuous, strictly increasing and convex function $g:\mathbb{R}^{+}\to \mathbb{R}^{*}$, $g(0)=0$ such that for all $x, y \in B_r$ and $0\leq\alpha \lt 1,$
where $W_q(\alpha):=\alpha^q(1-\alpha)+\alpha(1-\alpha)^q$ and $B_r:=\{x\in E:\|x\|\leq r\}.$
Lemma 2.13. [ Reference Reich and Sabach38] Let E be a real Banach space and let $f:E\to \mathbb{R}$ be a Gâteaux differentiable and totally convex function. If $x_0\in E$ and the sequence $\{\Delta_f(x_n, x_0)\}$ is bounded, then the sequence $\{x_n\}$ is also bounded.
Lemma 2.14. [ Reference Kimura and Saejung26] Let $\{a_n\}\subset \mathbb{R}_{+}$, $\{\varepsilon_{n}\}\subset (0,1)$ be a sequence such that $\sum_{n=1}^{\infty}\varepsilon_{n}=\infty$ and $\{b_n\}\subset\mathbb{R}$. Assume that
If $\limsup\limits_{k\to\infty}b_{n_k}\le 0$ for every subsequence $\{a_{n_k}\}$ of $\{a_n\}$ satisfying the condition $\liminf\limits_{k\to \infty}(a_{n_k+1}-a_{n_k})\ge0,$ then $\lim\limits_{k\to \infty}a_n=0.$
To solve the EP (1.6), the following assumptions are needed:
Assumption A:
(C1) f is pseudomonotone, that is, for all $x, y \in C,$ $f(x, y)\geq 0\Rightarrow f(y, x)\leq 0$ and $f(x, x)=0,$ for all $x\in C$.
(C2) f satisfies the Bregman–Lipschitz type condition on C, that is, there exists two positive constants c 1 and c 2 such that
\begin{align*} f(x,y)+ f(y,z)\geq f(x,z)-c_1\Delta_p(y, x)-c_2\Delta_p(y, z),\quad \forall x, y,z \in C, \end{align*}where $p:E\to (-\infty, +\infty]$ is a Legendre function. The constants c 1 and c 2 are called Bregman–Lipschitz coefficients with respect to p.(C3) $f(x, .)$ is convex, lower semicontinuous and subdifferentiable on C for all $x\in C$.
(C4) f is jointly weakly continuous on C × C in the sense that if $x, y \in C$ and $\{x_n\}$ and $\{y_n\}$ converges weakly to x and y, respectively, then $f(x_n, y_n)\to f(x, y)$ as $n\to \infty$.
3. Proposed method
In this section, we present our method and discuss some of its features. We begin with the following assumptions under which our strong convergence result is obtained.
Assumption 3.1. We assume that the following conditions hold:
(1) (a) E 1, E 2 and E 3 are three p-uniformly convex and uniformly smooth real Banach spaces.
(b) Ci and Qj are nonempty closed and convex subsets of E 1 and $E_2,$ respectively, for $i=1,2,\ldots,N$ and $j=1,2,\ldots,M.$
(c) $A:E_{1}\to E_{3}$ and $B:E_2\to E_3$ are bounded linear operators.
(d) $f_{i}:C_{i}\times C_{i}\to \mathbb{R}$ and $g_{j}:Q_{j}\times Q_{j}\to \mathbb{R}$ are bifunctions satisfying conditions $C_{1}-C_{4}$ of Assumption A.
(e) $D_{s}:E_{1}\to E_{1}$ and $G_{t}:E_{2}\to E_{2}$ are Bregman quasi-nonexpansive mappings such that $I-D_s$ and $I-G_t$ are demiclosed at zero for each $s=1,2,\ldots,l$ and $t=1,2,\ldots,m.$
(f) Assume that the solution set
$\Upsilon:=\{\bar{x}\in \bigcap_{s=1}^l F(D_{s})\cap\bigcap_{i=1}^N \mbox{EP}(C_{i}, f_{i}),\quad \bar{y}\in \bigcap_{t=1}^mF(G_{t})\cap \bigcap_{j=1}^M\mbox{EP}(Q_{j}, g_{j}):A\bar{x}=B\bar{y}\}\neq 0$.
(2) $\{\beta_n\}^{\infty}_{n=1}$, $\{\alpha_{n, s}\}_{s=0}^{l}$, $\{\eta_{n, t}\}_{t=0}^{m}$ are positive sequences satisfying the following conditions:
(a)$\{\beta_n\}\subset (0,1),\ \lim\limits_{n\to\infty} \beta_n=0,\ \sum_{n=1}^{\infty}\beta_n= \infty,\ \tau_0 \gt 0,\ \lambda_0 \gt 0,\ \kappa\in(0,1),\ \epsilon\in(0,1).$
(b) $\{\alpha_{n,s}\}\subset (0,1),\ \sum_{s=0}^{l}\alpha_{n,s}=1 \ \mbox{and}\ \liminf\limits_{n\to\infty}\alpha_{n,0}\alpha_{n,s} \gt 0.$
(c) $\{\eta_{n,t}\}\subset (0,1),\ \sum_{t=0}^{m}\eta_{n,t}=1 \ \mbox{and}\ \liminf\limits_{n\to\infty}\eta_{n,0}\eta_{n,t} \gt 0.$
We now present the proposed method of this paper.
Algorithm 3.2.
For fixed $\mu\in E_1$ and $\vartheta\in E_2$, choose an initial guess $(x_0, y_0)\in E_1\times E_2$. Suppose that the nth iterate $(x_n, y_n)\subset E_1\times E_2$ has been constructed; then we compute the $(n+1)$th iterate $(x_{n+1}, y_{n+1})$ via the iteration
where
for small enough ζ; Cq and Qq are constants of smoothness of E1 and E2, respectively. Otherwise, $\rho_{n}=\rho$ (ρ being any nonnegative value), where the set of indexes $\Omega=\{n:Ax_n-By_n\neq 0\}$.
and
(a) Algorithm 3.2 solves split equality EP consisting of two strongly convex optimization problems in parallel for $i=1,2,\ldots,N$, as well as another two strongly convex optimization problems in parallel for $j=1,2,\ldots,M$ under bounded linear operators.
(b) The step size $\{\rho_n\}$ given by Equation (3.2) is generated at each iteration by some simple computations. Thus, $\{\rho_n\}$ is easily implemented without the prior knowledge of the operator norms $\|A\|$ and $\|B\|$. Similarly, the step size $\{\tau_n\}$ given by Equation (3.3) and step size $\{\lambda_n\}$ given by Equation (3.4) do not depend on the prior estimates of the Lipschitz-like constants of the pseudomonotone bifunctions $f_{i},~i=1,2,\ldots,N$, and $g_{j},~j=1,2,\ldots,M$, unlike the step sizes used in [Reference Eskandani, Raeisi and Rassias14, Reference Hieu, Muu and Anh22], which require finding the prior estimates of the Lipschitz-like constants of the pseudomonotone bifunctions, which is known to be computationally expensive.
(c) Moreover, our result in this paper extends the results in [Reference Hieu, Muu and Anh22, Reference Kim and Dinh25] from the framework of Hilbert spaces to Banach spaces.
4. Convergence analysis
Lemma 4.1. The sequences $\{\tau_{n}\}$ and $\{\lambda_{n}\}$ of step sizes generated by Algorithm 3.2 are well defined and bounded.
Proof. Clearly, from Equations (3.3) and (3.4), we have $\tau_{n+1}\leq \tau_n\ \forall n\in \mathbb{N}$ and $\lambda_{n+1}\leq \lambda_n\ \forall n\in \mathbb{N}$. This implies that $\{\tau_n\}$ and $\{\lambda_n\}$ are monotonically decreasing sequences. Moreover, it follows from condition C 2 of Assumption A that
Hence, we obtain for all $i=1,2,\dots,N$
Similarly, we obtain
Hence, we conclude that $\{\tau_n\}$ has lower bound $\min\left\{\tau_0, \frac{\kappa}{\max\limits_{1\le i\le N}\{k_{1,i},k_{2,i}\}}\right\} \gt 0$ and $\{\lambda_n\}$ has lower bound
$\min\left\{\lambda_0, \frac{\epsilon}{\max\limits_{1\le j\le M}\{c_{1,j},c_{2,j}\}}\right\} \gt 0.$ It then follows that $\lim\limits_{n\to \infty}\tau_n=\tau \gt 0$ and $\lim\limits_{n\to \infty}\lambda_n=\lambda \gt 0.$
Lemma 4.2. Let $C_i,\ i=1,2,\ldots,N$ and $Q_j, j=1,2,\ldots,M$ be nonempty, closed and convex subsets of E1 and E2, respectively. Suppose that $f_{i}:C_{i}\times C_{i}\to \mathbb{R},~i=1,2,\ldots, N$ and $g_j:Q_j\times Q_j\to \mathbb{R},~j=1,2,\ldots, M$ are bifunctions satisfying conditions $C_{1}-C_{4}$. Then, for all $(\bar{x},\bar{y})\in \Upsilon$, we have
and
Proof. Since $z_{n}^{i}=\arg\min\left\{f_{i}(a_{n}^{i}, \sigma)+ \frac{1}{\tau_n}\Delta_p(\sigma, s_n):\sigma\in C_{i}\right\},$ then from Lemma 2.6, we get
Then, there exists $\xi\in \partial_{2}f_{i}(a_{n}^{i}, z_{n}^{i}),$ $\bar{\xi}\in N_{C_i}(z_{n}^{i}),$ such that
Also, by the definition of $\partial_{2}f_{i}(a_{n}^{i}, z_{n}^{i})$, we obtain
If we replace σ with $\bar{x}$ in the inequality above, we have
Using the definition of $N_{C_i}(z_{n}^{i})$ together with Equation (4.3), we have
Again, if we let $\sigma=\bar{x}$ in Equation (4.5), we get
The combination of Equations (4.4) and (4.6) gives
because $f_{i}(\bar{x}, a_{n}^{i})\geq 0$ and fi is pseudomonotone on $C_{i},\ \forall i=1,2,\ldots,N.$ Similarly, since $a_{n}^{i}=\arg\min\{f_{i}(s_{n}, \sigma)+ \frac{1}{\tau_n}\Delta_p(\sigma, s_n): \sigma\in C_{i}\},$ we obtain
Using Equations (4.7) and (4.8) together, we get
Applying Bregman three-point identity Equation (2.1), we obtain
Furthermore, by the definition of τn, we obtain
Following similar procedure, we obtain
Observe that since $\lim\limits_{n\to \infty}\left(1-\frac{\tau_{n}}{\tau_{n+1}}\kappa\right)=1-\kappa \gt 0,$ then there exists $K\in\mathbb{N}$ such that
Hence, from Equation (4.9), we get
Similarly, from Equation (4.10), we obtain
Lemma 4.3. Suppose $\{x_n\}$ and $\{y_n\}$ are iterative sequences generated by Algorithm 3.2 under Assumption 3.1. Then, the sequences $\{x_n\}$ and $\{y_n\}$ are bounded.
Proof. Let $(\bar{x}, \bar{y})\in \Upsilon.$ Since Ds is Bregman quasi-nonexpansive for each $s=1,2,\ldots,l,$ we obtain from Equation (3.1) that
Similarly, we obtain
Furthermore, from Equation (3.1), Lemma 2.7 and Lemma 2.10, we obtain
Similarly, we have
Combining Equations (4.15) and (4.16) and noting that $A\bar{x}=B\bar{y}$, we have
Hence,
Also, from Equation (3.1) and applying Equation (4.11), we obtain
In like manner, we have
It follows from Equations (4.18), (4.19) and (4.20) that
Therefore, $\{\Delta_p(\bar{x}, x_{n})+ \Delta_p(\bar{y}, y_{n})\}$ is bounded, and consequently $\{\Delta_p(\bar{x}, x_{n})\}$ and $\{\Delta_p(\bar{y}, y_{n})\}$ are bounded. Hence, by Lemma 2.13, the sequences $\{x_n\}$ and $\{y_n\}$ are bounded. Therefore, $\{s_n\}$, $\{a_{n}^{i}\}$, $\{z_{n}^{i}\}$, $\{u_n\}$, $\{t_n\}$, $\{b_{n}^{j}\}$, $\{h_{n}^{j}\}$ and $\{v_n\}$ are all bounded.
Lemma 4.4. Assume that $r=\sup\{\|J_{E_1}^{p}(\bar{z}_{n})\|,~\|J_{E_1}^{p}(D_s\bar{z}_{n})\|\}$ and let $(\bar{x}, \bar{y})\in \Upsilon.$ Then, the following inequality holds:
where $W_q(\alpha_{n,s})= (\alpha_{n,0})^q\sum_{s=1}^{l}\alpha_{n,s} + \alpha_{n,0}(\sum_{s=1}^{l}\alpha_{n,s})^q$ and $W_q(\eta_{n,t})= (\eta_{n,0})^q\sum_{t=1}^{m}\eta_{n,t} + \eta_{n,0}(\sum_{t=1}^{m}\eta_{n,t})^q.$
Proof. Let $(\bar{x}, \bar{y})\in \Upsilon.$ Then, from Equation (3.1), Lemma 2.10 and Lemma 2.12, we obtain
By the Bregman quasi-nonexpansivity of Ds for $s=1,2,\ldots,l$, we get
By the definition of $x_{n+1}$ and applying Equations (4.11) and (4.13) from Equation (4.23), we obtain
Following similar argument, we have
By adding Equations (4.24) and (4.25) and applying Equation (4.18), we get
which is the required inequality.
We now present the main theorem for our proposed algorithm as follows.
Theorem 4.5 Suppose $\{(x_n, y_n)\}$ is a sequence generated by Algorithm 3.2 under Assumption 3.1. Then $\{(x_n, y_n)\}$ converges strongly to $(\bar{x}, \bar{y})\in \Upsilon,$ where $\bar{x}=\Pi_\Upsilon(\mu)$ and $\bar{y}=\Pi_\Upsilon(\vartheta).$
Proof. Let $(\bar{x}, \bar{y})=(\Pi_\Upsilon(\mu),\Pi_\Upsilon(\vartheta)).$ It follows from Algorithm 3.2 and by applying Lemma 2.10 (iii) that
In the same vein, we have
Hence, by adding Equations (4.26) and (4.27) and applying Equation (4.18), we get
where $\chi_{n}:=\left(\langle J_{E_1}^{p}(\mu)-J_{E_1}^{p}(\bar{x}), x_{n+1}-\bar{x}\rangle + \langle J_{E_2}^{p}(\vartheta)-J_{E_2}^{p}(\bar{y}), y_{n+1}-\bar{y}\rangle\right).$
In order to show that $\{(x_n, y_n)\}$ converges strongly to $(\bar{x}, \bar{y})$ by Lemma 2.14, we only need to show that $\limsup_{k\to \infty}\chi_{n_k}\leq 0$ for every subsequence $\{\Delta_{p}(x_{n_k}, \bar{x})\}$ of $\{\Delta_{p}(x_{n}, \bar{x})\}$ and $\{\Delta_{p}(y_{n_k}, \bar{y})\}$ of $\{\Delta_{p}(y_{n}, \bar{y})\}$ satisfy the inequality
Now, from Algorithm 3.2 and Lemma 4.2, we obtain
In the same vein, we obtain
Adding Equations (4.30) and (4.31) together, we obtain
Applying Equation (4.17) in Equation (4.32), we obtain
By Equation (4.29), Assumption 3.1(2)(a) and (4.33), we obtain
In the same vein as in Equation (4.34), we get using Equation (4.29), Assumption 3.1(2)(a) and Equation (4.33) that
Now, suppose we let $\varrho_{n_k}=C_q\|A^*J_{E_3}^{p}(Ax_{n_k}-By_{n_k})\|^q+Q_q\|B^*J_{E_3}^{p}(Ax_{n_k}-By_{n_k})\|^q.$ Using the condition we placed on our step size $\rho_{n_k},$ we have that
it follows that
Hence, by Equations (4.34) and (4.36), we have
Thus, $C_q\|A^*J_{E_3}^{p}(Ax_{n_k}-By_{n_k})\|^q+Q_q\|B^*J_{E_3}^{p}(Ax_{n_k}-By_{n_k})\|^q\to 0\quad \mbox{as}~ k\to \infty,$
which implies that
and
Additionally, we obtain from Equation (4.34) that
Thus, we conclude from Equations (4.34), (4.35) and (4.39) that
Therefore, by Lemma 2.11, we obtain
Observe that by Equation (4.42) and Lemma 2.11, we have
In like manner, we have
By the definitions of in and $j_n,$ it follows that
Consequently, we have
From Equation (4.1) and by applying the three-point identity (2.1) and (4.46), we have
Hence, we have
Consequently, we obtain
Following similar procedure, we have
Furthermore, using Equations (4.22) and (4.29), we have
Thus,
Hence, we have
By the property of $g,$ and sine $J_{E_1^*}^{q}$ and $J_{E_2^*}^{q}$ are norm-to-norm uniformly continuous on bounded subsets of E 1 and E 2, respectively, then we obtain
and
Observe that from Equation (3.1) and by Equation (4.40), we obtain
Also, because E 1 is uniformly smooth, $J_{E_1^*}^{q}$ is norm-to-norm uniformly continuous on bounded subsets of E 1, then we have
In the same vein, we get
Moreover, it is easy to see from Equations (4.46) and (4.53)
In the same way, we obtain from Equations (4.46) and (4.54) that
Moreover, we obtain from Equations (3.1) and (4.50) that
which implies that
By the uniform continuity of $J_{E_1^*}^{q}$ on bounded subsets of $E_{1}^{*}$, we have
Hence, from Equations (4.55) and (4.57), we obtain
Similarly, we obtain
It is easy to see from Equations (4.55) and (4.58) that
Similarly, we obtain from Equations (4.56) and (4.59) that
Furthermore, from Equation (3.1) and the fact that $\lim\limits_{k\to \infty}\beta_{n_k}=0$, we obtain
In the same way, we get
Since $J_{E_1^*}^{p}$ is norm-to-norm uniformly continuous on bounded subsets of E 1, we obtain
Similarly, we get
Hence, from Equations (4.60) and (4.62), we obtain
In the same vein, from Equations (4.61) and (4.63), we get
Since $\{x_n\}$ and $\{y_n\}$ are bounded, then $w_\omega(x_n)$ and $w_{\omega}(y_n)$ are nonempty. Now, let $(x^*, y^*)\in w_\omega(x_n, y_n)$ be arbitrary elements. Then, there exists subsequences $\{x_{n_{k}}\}$ of $\{x_n\}$ and $\{y_{n_{k}}\}$ of $\{y_n\}$ that converge weakly to $x^*\in E_{1}$ and $y^*\in E_{2}$, respectively. Also, from Equations (4.55) and (4.56), $\{z_{n_{k}}^{i}\}$ converges weakly to $x^*\in C_{i}$ for each $i=1,2,\ldots,N$ and $\{h_{n_{k}}^{j}\}$ converges weakly to $y^*\in Q_{j}$ for each $j=1,2,\ldots,M.$ Using Equations (4.50) and (4.51) and by the demiclosedness of $I-D_s$ and $I-G_t$, we obtain
which implies that
Next, recall that
Using Lemma 2.6 and applying condition (C4), we get
Hence, there exists $\varsigma_{n_k}^{i}\in \partial_{2}f_i(s_{n_k}, a_{n_k}^{i})$ and $\bar{\varsigma}_{n_k}^{i}\in N_{C_i}(a_{n_k}^{i})$ such that
Since $\bar{\varsigma}_{n_k}^{i}\in N_{C_i}(a_{n_k}^{i})$, $\langle \omega-a_{n_k}^{i}, \bar{\varsigma}_{n_k}^{i} \rangle \leq 0$ for all $\omega\in C_{i},$ then this together with Equation (4.68) gives
Again, since $\varsigma_{n_k}^{i}\in \partial_{2}f_i(s_{n_k}, a_{n_k}^{i})$, we obtain
Combining Equations (4.69) and (4.70), we obtain
which implies that
Since $J_{E_1}^{p}$ is uniformly continuous, applying Equation (4.47) to Equation (4.71) and using Equation (4.53) together with the fact that $x_{n_k}\rightharpoonup x^*,$ we get
which implies that
Hence, we have
Similarly, we obtain
which implies that
Next, recall that $\{x_{n_k}\}$ and $\{y_{n_k}\}$ converges to $x^*$ and $y^*,$ respectively, where $A:E_{1}\to E_{3}$ and $B:E_{2}\to E_{3}$ are bounded linear operators. Then, by Equation (4.40) and the weakly lower semi-continuity of the norm, we have
which implies that
Since $(x^*, y^*)\in w_\omega(x_n, y_n)$ is an arbitrary element, then it follows that
Next, by the boundedness of $\{x_{n_k}\}$ and $\{y_{n_k}\},$ there exist subsequences $\{x_{n_{k_j}}\}$ of $\{x_{n_k}\}$ and $\{y_{n_{k_j}}\}$ of $\{y_{n_k}\}$ such that $x_{n_{k_j}}\rightharpoonup \hat{x}\in E_1$ and $y_{n_{k_j}}\rightharpoonup \hat{y}\in E_2$ and
Since $\bar{x}=\Pi_\Upsilon(\mu)$ and $\bar{y}=\Pi_\Upsilon(\vartheta),$ then by Equation (2.5), (4.64) and (4.65), we have
Hence, by Equation (4.72), we have $\limsup\limits_{k\to \infty}\chi_{n_k}\leq 0.$ Therefore, by applying Lemma 2.14 to Equation (4.28), it follows that $\{(x_n, y_n)\}$ converges strongly to $(\bar{x}, \bar{y})\in \Upsilon$ as required.
Some corollaries
The following consequent result can easily be obtained from Theorem 4.5 by setting $l=m=N=M=1$.
Corollary. Let E1, E2 and E3 be three p-uniformly convex Banach space and $C, Q$ be nonempty, closed and convex subsets of E1 and E2, respectively. Suppose $f:C \times C \to \mathbb{R}$ and $g:Q \times Q \to \mathbb{R}$ be bifunctions satisfying (C1)–(C4) of Assumption A. Let $A:E_1\to E_3$ and $B:E_2\to E_3$ be bounded linear operators and let $D:E_1\to E_1$ and $G:E_2\to E_2$ be Bregman quasi-nonexpansive mappings such that I ‒ D and I ‒ G are demiclosed at zero and $\Upsilon:=\{\bar{x}\in F(D)\cap \mbox{EP}(C, f)$, $\bar{y}\in F(G)\cap \mbox{EP}(Q, g):A\bar{x}=B\bar{y}\}\neq \emptyset$. Suppose other conditions of Theorem 4.5 hold. For fixed $\mu\in E_1$ and $\vartheta\in E_2$ and initial point $(x_0, y_0)\in E_1\times E_2,$ let $\{(x_n,y_n)\}$ be a sequence generated as follows:
where $\rho_{n}\in \left(\zeta, \left(\frac{q\|Ax_n-By_n\|^p}{C_q\|A^*J_{E_3}^{p}(Ax_n-By_n)\|^q+ Q_q\|B^*J_{E_3}^{p}(Ax_n-By_n)\|^q}-\zeta\right)^\frac{1}{q-1}\right),\ n\in \Omega,$ for small enough ζ; Cq and Qq are constants of smoothness of E 1 and E 2, respectively. Otherwise, $\rho_{n}=\rho$ (ρ being any nonnegative value), where the set of indexes $\Omega=\{n:Ax_n-By_n\neq 0\}$.
and
Then, the sequence $\{(x_n, y_n)\}$ generated by Equation (4.73) converges strongly to $(\bar{x}, \bar{y})\in \Upsilon$.
Let $E_r=H_r$, $r=1,2,3$ be real Hilbert spaces, then we obtain the following consequent result for approximating a common solution of multiple sets split equality pseudomonotone EP and common fixed point problems of quasi-nonexpansive mappings in real Hilbert spaces.
Corollary 4.7. Let H 1, H 2 and H 3 be three real Hilbert spaces, and let Ci and Qj be nonempty closed and convex subsets of H 1 and $H_2,$ respectively, for $i=1,2,\ldots,N$ and $j=1,2,\ldots,M.$ Suppose $f_{i}:C_{i} \times C_{i} \to \mathbb{R}$ and $g_{j}:Q_{j} \times Q_{j} \to \mathbb{R}$ are bifunctions satisfying (C1)–(C4) of Assumption A. Let $A:H_1\to H_3$ and $B:H_2\to H_3$ be bounded linear operators. Let $D_{s}:H_1\to H_1$ and $G_{t}:H_2\to H_2$ be quasi-nonexpansive mappings such that $\Upsilon:=\{\bar{x}\in \bigcap_{s=1}^l F(D_s)\cap \bigcap_{i=1}^{N}\mbox{EP}(C_{i}, f_{i})$, $\bar{y}\in \bigcap_{t=1}^mF(G_t)\cap \bigcap_{j=1}^{M}\mbox{EP}(Q_j, g_j):A\bar{x}=B\bar{y}\}\neq \emptyset$. Suppose other conditions of Theorem 4.5 hold. For fixed $\mu\in H_1$ and $\vartheta\in H_2$ and initial point $(x_0, y_0)\in H_1\times H_2,$ let $\{(x_n,y_n)\}$ be a sequence generated as follows:
where $\rho_{n}\in \left(\zeta, \left(\frac{2\|Ax_n-By_n\|^2}{\|A^*(Ax_n-By_n)\|^2+ \|B^*(Ax_n-By_n)\|^2}-\zeta\right)\right),\ n\in \Omega,$ for small enough $\zeta.$ Otherwise, $\rho_{n}=\rho$ (ρ being any nonnegative value), where the set of indexes $\Omega=\{n:Ax_n-By_n\neq 0\}$.
and
Then the sequence $\{(x_n, y_n)\}$ generated by Equation (4.76) converges strongly to $(\bar{x}, \bar{y})\in \Upsilon$.
5. Application
5.1. Multiple set split equality variational inequality problem
In this section, we apply our result to study the multiple set split equality variational inequality problem (MSSEVIP).
Let $U:C\to E^{*}$ be a nonlinear mapping. The classical VIP is formulated as locating a point
The solution set of VIP (5.1) is denoted by $\mbox{VI}(C, U)$. Variational inequalities have been found very applicable in several real-world problems such as optimization problems, minimax theorems, differential equations and in certain applications to economic theory and mechanics. For more details on variational inequalities, see [Reference Alakoya, Mewomo and Shehu2, Reference Ogwo, Izuchukwu, Shehu and Mewomo35, Reference Owolabi, Alakoya, Taiwo and Mewomo36, Reference Uzor, Alakoya and Mewomo44] and the references therein.
Now we consider the MSSEVIP defined as follows:
where $U_i:E_1\to E_1$ and $V_j:E_2\to E_2$ are two nonlinear mappings, and $A:E_1\to E_3$ and $B:E_2\to E_3$ are two bounded linear operators. When viewed separately, Equation (5.2) consists of two classical multiple sets variational inequality problem (MSVIP) whose solution sets are denoted by $\mbox{VI}(C_{i},U_i)$ and $\mbox{VI}(Q_{j},V_j),$ respectively.
Let $U:C\to E^{*}$ be a nonlinear mapping. Then, U is said to be
(D1) pseudomonotone; if for any $x, y \in C$, we have
\begin{align*} \langle Ux, y-x \rangle \geq 0 \Longrightarrow\langle Uy, y-x\rangle \geq 0, \end{align*}(D2) K-Lipschitz continuous, if there exists a constant K > 0 such that
\begin{align*} \|Ux-Uy\|\leq K\|x-y\|,\quad \forall x, y \in C, \end{align*}(D3) sequentially weakly continuous, if for any sequence $\{x_n\}\subset C$, we have $x_n\rightharpoonup x\in C$ implying that $Ux_n\rightharpoonup Ux\in E^*.$
We need the following lemma to establish our next result.
Lemma 5.1. [ Reference Eskandani, Raeisi and Rassias14] Let C be a nonempty, closed convex subset of a reflexive, smooth and strictly convex Banach space E, $U:C\to E^*$ be a nonlinear mapping. Then
for all $x\in E$, $y\in C$ and $\lambda \in (0, +\infty).$
Setting $f_{i}(x,y)=\langle U_ix, y-x\rangle,\ \forall x, y \in C_{i},\ i=1,2,\ldots,N$ and $g_{j}(x, y)=\langle V_jx, y-x\rangle,\ \forall x, y \in Q_{j},\ j=1,2,\ldots,M$ in Algorithm (3.2), then the bifunctions fi and gj satisfy conditions (C1)–(C4) of Assumption A (see [Reference Eskandani, Raeisi and Rassias14]).
Hence, by applying Theorem 4.5 and Lemma 5.1, we obtain the following consequent result for approximating a common solution of MSSEVIP and common fixed point problem for finite families of Bregman quasi-nonexpansive mappings in p-uniformly convex real Banach spaces, which are also uniformly smooth.
Theorem 5.2. Let E1, E2 and E3 be three p-uniformly convex and uniformly smooth real Banach spaces. Let $C_i,\ i=1,2,\ldots,N$ and $Q_j,\ j=1,2,\ldots,M$ be nonempty, closed and convex subsets of E1 and E2, respectively. Let $U_i:C_i\to E^{*}$ and $V_j:Q_j\to E^{*}$ be two nonlinear mappings satisfying conditions (D1)–(D3) above. Let $D_s:E_1\to E_1,\ s=1,2,\ldots,l$ and $G_t:E_2\to E_2,\ t=1,2,\ldots,m$ be two finite families of Bregman quasi-nonexpansive mappings such that $I-D_s$ and $I-G_t$ are demiclosed at zero for each s and t, respectively. Suppose that Assumption 3.12(a)–2(c) holds and the solution set $\Upsilon:=\{\bar{x}\in F(D_s)\cap\mbox{VI}(C_{i},U_i),\ \bar{y}\in F(G_t)\cap\mbox{VI}(Q_j,V_j): A\bar{x}=B\bar{y}\}\neq \emptyset$. Then, the sequence $\{x_n, y_n\}$ generated by Algorithm (5.3) below converges strongly to $(\bar{x}, \bar{y})\in \Upsilon,$ where $\bar{x}=\Pi_\Upsilon(\mu)$ and $\bar{y}=\Pi_\Upsilon(\vartheta).$
Algorithm 5.3.
For fixed $\mu\in E_1$ and $\vartheta\in E_2$, choose an initial guess $(x_0, y_0)\in E_1\times E_2$. Suppose that the nth iterate $(x_n, y_n)\subset E_1\times E_2$ has been constructed, then we compute the $(n+1)$th iterate $(x_{n+1}, y_{n+1})$ via the iteration
where $\rho_{n}\in \left(\zeta, \left(\frac{q\|Ax_n-By_n\|^p}{C_q\|A^*J_{p}^{E_3}(Ax_n-By_n)\|^q+ Q_q\|B^*J_{p}^{E_3}(Ax_n-By_n)\|^q}-\zeta\right)^\frac{1}{q-1}\right)\ n\in \Omega,$ for small enough ζ, Cq and Qq are constants of smoothness of E1 and E2, respectively. Otherwise, $\rho_{n}=\rho$ (ρ being any nonnegative value), where the set of indexes $\Omega=\{n:Ax_n-By_n\neq 0\}.$
and
6. Computational Experiments
In this section, we demonstrate the efficiency and applicability of our proposed method with two numerical examples. In all the experiments, we consider the case when $l=m=N=M=5.$
Example 6.1. Let $E_r=\mathbb{R}^{m},\ r=1,2,3$, equipped with induced norm $\|x\|=\sqrt{\sum_{i=1}^{m}|x_{i}|}$ and the inner product $\langle x,y\rangle = \sum_{i=1}^{m}x_{i}y_{i}$, for all $x=(x_{1}, x_{2},\ldots, x_{m})\in \mathbb{R}^{m}$ and $y=(y_{1}, y_{2},\ldots, y_{m})\in \mathbb{R}^{m}$. Let $C_{i}=Q_{j}=C,$ where the feasible set C has the form
Consider the following problem:
where $f_{i}:\mathbb{R}^{m}\times \mathbb{R}^{m}\to \mathbb{R}$ is given by
where $q_{ik}\in (0,1)$ is randomly selected $\forall i=1,2,\ldots,N,\ k=1,2,\ldots,m$ and $D_{s}:\mathbb{R}^{m}\to \mathbb{R}^{m}$ is defined by
In the same vein, let $g_{j}:\mathbb{R}^{m}\times \mathbb{R}^{m}\to \mathbb{R}$ is given by
where $q_{jk}\in (0,1)$ is randomly selected $\forall j=1,2,\ldots,M,\ k=1,2,\ldots,5$ and $G_{t}:\mathbb{R}^{m}\to \mathbb{R}^{m}$ is defined by
It is easy to see that conditions (C1)–(C4) of Assumption A are satisfied and Ds and Gt are Bregman quasi-nonexpansive mappings for $s=1,2,\ldots,l$ and $t=1,2,\ldots,m$, respectively, $I-D_s$ and $I-G_t$ are demiclosed at zero. Moreover, we define $A(x)= \frac{x}{2}$ and $B(x)= \frac{x}{3},$ then A and B are bounded linear operators. Furthermore, $\Upsilon=\{0\}$. In this example, we choose $\beta_{n}=\frac{3}{2n+3}$, κ = 0.36, $\tau_{0}=0.24$, ϵ = 0.5, $\lambda_{0}=0.4$, $\alpha_{n,0}=\frac{3n}{8n+11}$, $\alpha_{n,s}=\frac{1}{5}(1-\frac{3n}{8n+11})$, $s=1,2,\ldots,5$, $\eta_{n,0}=\frac{2n}{4n+7}$, $\eta_{n,t}=\frac{1}{5}(1-\frac{2n}{4n+7})$, $t=1,2,\ldots,5$. Using $\frac{\|x_{n+1}-x_{n}\|}{\|x_2-x_1\|} \lt 10^{-4}$ as our stopping criterion, we generate randomly different starting points $(\mu, \vartheta),(x_0, y_0)\in E_1\times E_2$ for different cases of $m=20,\ 50,\ 100\ {\rm and}\ 500.$
We plot the graphs $\|x_{n+1}-x_{n}\|$ against the number of iterations. The numerical results can be seen in Figure 1.
The next example is presented in an infinite dimensional space setting.
Example 6.2. Let $E_r=L^{2}\left([0,1]\right)$, $r=1,2,3$, with the induced norm given by $\|x\|_{L}=\int_{0}^{1}|x(s)|^{2}\,{\rm d}s$ and the corresponding inner product $\langle x,y\rangle=\int_{0}^{1}x(s)y(s)\,{\rm d}s.$ Let the feasible sets Ci and Qj be defined as follows:
Let $f_{i}(x,y)=\langle S_{i}x, y-x\rangle$ and $g_{j}(x,y)=\langle T_{j}x, y-x\rangle$ with the operators $(S_ix)(t)=\max\left\{0, \frac{x(t)}{i}\right\}$ for $i=1,2,\ldots,5$ and $(T_jx)(t)=\max\left\{0, \frac{x(t)}{j}\right\}$ for $j=1,2,\ldots,5.$ Then, it is easy to see that each fi is monotone (and by implication, pseudomonotone) on $C_{i}.$ Similarly, gj is pseudomonotone on Qj. Furthermore, let $D_{s}:H\to H$ and $G_{t}:H\to H$ be defined by $D_{s}(x)(t)=\frac{x(t)}{2s}$ and $G_{t}(x)(t)= \frac{x(t)}{2t},$ then the mappings Ds and Gt are quasi-nonexpansive $\forall s=1,2,\ldots,5$ and $t=1,2,\ldots,5.$ Moreover, we define $A(x)(t)= \frac{x(t)}{3}$ and $B(x)(t)= \frac{x(t)}{5},$ then A and B are bounded linear operators. The solution set $\Upsilon=\{0\}$. We choose $\beta_{n}=\frac{1}{n+2}$, $\alpha_{n,0}=\frac{n+1}{2n+3}$, $\alpha_{n,s}=\frac{1}{5}(1-\frac{n+1}{2n+3})$, $s=1,2,\ldots,5$, $\eta_{n,0}=\frac{n+2}{2n+5}$, $\eta_{n,t}=\frac{1}{5}(1-\frac{n+2}{2n+5})$, $t=1,2,\ldots,5$, κ = 0.54, $\tau_{0}=0.63$, ϵ = 0.75, $\lambda_{0}=0.83,$ and using $\frac{\|x_{n+1}-x_{n}\|}{\|x_2-x_1\|} \lt 10^{-4}$ as stopping criterion.
We choose fixed points $\mu=t^2+2$, $\vartheta=4t^3+3$ and different starting points as follows:
Case I: $x_0=t^2 + 4$, $y_0=t^3 + 2t + 1$
Case II: $x_0=t^4 + 5$, $y_0=t^2 + t + 3$
Case III: $ x_0= \sin(2t),\quad y_0=\cos(5t)$
Case IV: $x_0=\exp(t), \quad y_0=\exp(2t)$.
We plot the graph of errors against the number of iterations in each case. The numerical results can be found in Figure 2.
7. Conclusion
In this article, using Bregman distance, we proposed and studied a new algorithm for approximating the common solution of multiple set split equality pseudomonotone EP and fixed points of Bregman quasi-nonexpansive mappings in real p-uniformly convex Banach spaces, which are also uniformly smooth. The algorithm under discourse is designed in such a way that its convergence does not rely on prior estimates of the Lipschitz constants of the pseudomonotone bifunctions as well as the prior knowledge of the norm of the bounded linear operators. We proved a strongly convergent theorem under some mild conditions on our parameters. We gave a theoretical application of our result and finally present two numerical examples to show the efficiency and applicability of our method. The result presented in this article extends numerous results in the literature in this direction of research.
Acknowledgements
The authors sincerely thank the anonymous referee for his careful reading, constructive comments and useful suggestions that improved the manuscript. Opinions expressed and conclusions arrived are those of the authors and are not necessarily to be attributed to the National Research Foundation.
Funding Statement
The second author is supported by the National Research Foundation of South Africa Incentive Funding for Rated Researchers (grant number 119903). The third author is funded by University of KwaZulu-Natal, Durban, South Africa, Postdoctoral Fellowship. He is grateful for the funding and financial support.
Competing Interests
The authors declare that they have no conflict of interest.