In this paper, we introduce an iterative method for approximating a common solution of monotone inclusion problem and fixed point of Bregman nonspreading mappings in a reflexive Banach space. Using the Bregman distance function, we study the composition of the resolvent of a maximal monotone operator and the antiresolvent of a Bregman inverse strongly monotone operator and introduce a Halpern-type iteration for approximating a common zero of a maximal monotone operator and a Bregman inverse strongly monotone operator which is also a fixed point of a Bregman nonspreading mapping. We further state and prove a strong convergence result using the iterative algorithm introduced. This result extends many works on finding a common solution of the monotone inclusion problem and fixed-point problem for nonlinear mappings in a real Hilbert space to a reflexive Banach space.
National Research Foundation111992Department of Science and Technology and National Research Foundation, Republic of South Africa Center of Excellence in Mathematical and Statistical Sciences (DST-NRF COE-MaSS)BA 2018/0121. Introduction
Let E be a real reflexive Banach space with a norm ⋅ and E∗ be the dual space of E. We denote the value of x∗∈E∗ at x∈E by x∗,x. A mapping A is called a monotone mapping if for any x,y∈domA, we have(1)μ∈Ax,ν∈Ay⟹μ−ν,x−y≥0.
A monotone mapping A:E⟶2E∗ is said to be maximal monotone if its graph, GA≔x,u∈E×E∗:u∈Ax, is not properly contained in the graph of any other monotone operator. A basic problem that arises in several branches of applied mathematics [1–7] is to find x∈E such that(2)0∈Ax.
One of the methods for solving this problem is the well-known proximal point algorithm (PPA) introduced by Martinet [8]. Let H be a Hilbert space and let I denote the identity operator on H. The PPA generates for any starting point x0=x∈H, a sequence xn in H by(3)xn+1=I+λnA−1xn,n=1,2,…,where A is a maximal monotone mapping and λn is a given sequence of positive real numbers. It has been observed that (3) is equivalent to(4)0∈Axn+1+1λnxn+1−xn,n=1,2,….
This algorithm was further developed by Rockafellar [5], who proved that the sequence generated by (3) converges weakly to an element of A−10 when A−10 is nonempty and liminfn⟶∞λn>0. Furthermore, Rockafellar [5] asked if the sequence generated by (3) converges strongly in general. This question was answered in the negative by Güler [9] who presented an example of a subdifferential for which the sequence generated by (3) converges weakly but not strongly. Also, the works of Bruck and Reich [10] and Bauschke et al. [11] are very important in this direction. For more recent results on PPA, see [12–14].
The problem of finding the zeros of the sum of two monotone mappings A and B, is to find a point x∗∈E such that(5)0∈A+Bx∗,has recently received attention due to its significant importance in many physical problems. One classical method for solving problem (5) is the forward-backward splitting method [15], which is as follows: for x1∈E,(6)xn+1=I+rB−1xn−rAxn,n≥1,where r>0. This method combines the proximal point algorithm and the gradient projection algorithm. In [16], Lions and Mercier introduced the following splitting iterative methods in a real Hilbert space H:(7)xn+1=2JrA−I2JrB−Ixn,n≥1,xn+1=JrA2JrB−Ixn+I−JrBxn,n≥1,where JrT=I+rT−1. The first one is called Peaceman–Rachford algorithm and the second one is called Douglas–Rachford algorithm [15]. It was noted that both algorithms converge weakly in general [16, 17].
Many authors have studied the approximation of zero of the sum of two monotone operators (in Hilbert space) and accretive operators (in Banach spaces), but the approximation of the sum of two monotone operators in more general Banach spaces other the Hilbert spaces has not enjoyed such popularity.
Throughout this paper, f:E⟶−∞,+∞ is a proper lower semicontinuous and convex function, and the Fenchel conjugate of f is the function f∗:E∗⟶−∞,+∞ defined by(8)f∗x∗=supx∗,x−fx:x∈E.
We denote by domf the domain of f, that is, the set x∈E:fx<+∞. For any x∈intdomf and y∈E, the right-hand derivative of f at x in the direction of t is defined by(9)fox,y≔limt⟶0+fx+ty−fxt.
The function f is said to be Gâteaux differentiable at x if the limit as t⟶0+ in (9) exists for any y. In this case, fox,y coincides with ∇fx, the value of the gradient ∇f at x. The function f is said to be Gâteaux differentiable if it is Gâteaux differentiable for any x∈intdomf. The function f is Fréchet differentiable at x if the limit is attained with y=1 and uniformly Fréchet differentiable on a subset C of E if the limit is attained uniformly for x∈C and y=1.
The function f is said to be Legendre if it satisfies the following two conditions:
(L1) intdomf≠∅ and the subdifferential ∂f is single-valued in its domain
(L2) tdomf∗≠∅ and ∂f∗ is single-valued on its domain
The class of Legendre functions in infinite dimensional Banach spaces was first introduced and studied by Bauschke et al. in [18]. Their definition is equivalent to conditions (L1) and (L2) because the space E is assumed to be reflexive (see [18], Theorems 5.4 and 5.6, p. 634). It is well known that in reflexive Banach spaces, ∇f=∇f∗−1 (see [19], p. 83). When this fact is combined with conditions (L1) and (L2), we obtain(10)ran∇f=dom∇f∗=intdomf∗,ran∇f∗=dom∇f=intdomf.
It also follows that f is Legendre if and only if f∗ is Legendre (see [18], Corollary 5.5, p. 634) and that the functions f and f∗ are Gâteaux differentiable and strictly convex in the interior of their respective domains.
Several interesting examples of the Legendre functions are presented in [18, 20, 21]. A very important example of Legendre function is the function 1/s⋅s with s∈1,∞, where the Banach space E is smooth and strictly convex, and in particular, a Hilbert space. Throughout this article, we assume that the convex function f:E⟶−∞,+∞ is Legendre.
Definition 1.
Let f:E⟶−∞,+∞ be a convex and Gâteaux differentiable function, the function Df: domf× intdomf⟶0,∞ which is defined by(11)Dfy,x≔fy−fx−∇fx,y−x,is called the Bregman distance [22–24].
The Bregman distance does not satisfy the well-known metric properties, but it does have the following important property, which is called the three-point identity: for any x∈domf and y,z∈intdomf,(12)Dfx,y+Dfy,z−Dfx,z=∇fz−∇fy,x−y.
Let C be a nonempty subset of a Banach space E and T:C⟶C be a mapping, then a point x is called fixed point of T if Tx=x. The set of fixed point of T is denoted by FT. Also, a point x∗∈C is said to be an asymptotic fixed point of T if C contains a sequence xnn=1∞ which converges weakly to x∗ and limn⟶∞xn−Txn=0 [25]. The set of asymptotic fixed points of T is denoted by F^T.
Definition 2 [26, 27].
Let C be a nonempty, closed, and convex subset of E. A mapping T:C⟶intdomf is called
Bregman firmly nonexpansive (BFNE for short) if
(13)∇fTx−∇fTy,Tx−Ty≤∇fx−∇fy,Tx−Ty,∀x,y∈C.
Bregman strongly nonexpansive (BSNE) with respect to a nonempty F^T if
(14)Dfp,Tx≤Dfp,x,
for all p∈F^T and x∈C and if whenever xnn=1∞⊂C is bounded, p∈F^T and
(15)limn⟶∞Dfp,xn−Dfp,Txn=0,
it follows that
(16)limn⟶∞DfTxn,xn=0.
Bregman quasi-nonexpansive if FT≠∅ and
(17)Dfp,Tx≤Dfp,x,∀x∈C and p∈FT.
Bregman skew quasi-nonexpansive if FT≠∅ and
(18)DfTx,p≤Dfx,p,∀x∈C and p∈FT.
Bregman nonspreading if
(19)DfTx,Ty+DfTy,Tx≤DfTx,y+DfTy,x,∀x,y∈C.
It is easy to see that every Bregman nonspreading mapping T with FT≠∅ is Bregman quasi-nonexpansive. Also Bregman nonspreading mappings include, in particular, the class of nonspreading functions studied by Takahashi et al. in [28, 29]. For more information on Bregman nonspreading mappings, see [30].
In a real Hilbert space H, the nonlinear mapping T:C⟶C is said to be
Nonexpansive if
(20)Tx−Ty≤x−y,∀x,y∈C.
Quasi-nonexpansive if FT≠∅ and
(21)Tx−p≤x−p,∀x∈C and p∈FT.
Nonspreading if
(22)2Tx−Ty2≤Tx−y2+Ty−x2,∀x,y∈C.
Clearly, every nonspreading mapping T with FT≠∅ is also quasi-nonexpansive mapping. The class of nonspreading mappings is very important due to its relation with maximal monotone operators (see, e.g., [28]).
Let B:E⟶2E∗ be a maximal monotone operator. The resolvent of B,ResBf:E⟶2E, is defined by (see [26])(23)ResBf≔∇f+B−1∘∇f.
It is known that ResBf is a BFNE operator, single-valued, and FResBf=B−10∗ (see [26]). If f:E⟶ℝ is a Legendre function which is bounded, uniformly Fréchet differentiable on bounded subsets of E, then ResBf is BSNE and F^ResBf=FResBf (see [31]).
Assume that the Legendre function f satisfies the following range condition:(24)ran∇f−A⊆ran∇f.
An operator A:E⟶2E∗ is called Bregman inverse strongly monotone (BISM) if (domA) (domf), and for any x,y∈intdomf and each u∈Ax and v∈Ay, we have(25)u−v,∇f∗∇fx−u−∇f∗∇fy−v≥0.
The class of BISM mappings is a generalization of the class of firmly nonexpansive mappings in Hilbert spaces. Indeed, if f=1/2⋅2, then ∇f=∇f∗=I, where I is the identity operator and (25) becomes(26)u−v,x−u−y−v≥0,which means(27)u−v2≤x−y,u−v.
Observe that(28)domAf=domA∩intdomf,ranAf⊂intdomf.
In other words, T is a (single-valued) firmly nonexpansive operator.
For any operator A:E⟶2E∗, the antiresolvent operator Af:E⟶2E of A is defined by(29)Af≔∇f∗∘∇f−A.
It is known that the operator A is BISM if and only if the antiresolvent Af is a single-valued BFNE (see [32], Lemma 3.2(c) and (d), p. 2109) and FAf=A−10∗. For examples and further information on BISM, see [32].
Since the monotone inclusion problems have very close connections with both the fixed-point problems and the equilibrium problems, finding the common solutions of these problems has drawn many people’s attention and has become one of the hot topics in the related fields in the past few years [33, 34]. Furthermore, interest in finding the common solution of these problems has also grown because of the possible application of these problems to mathematical models whose constraints can be present as fixed points of mappings and/or monotone inclusion problems and/or equilibrium problems. Such a problem occurs, in particular, in the practical problems as signal processing, network resource allocation, and image recovery (see [35, 36]).
In this paper, we introduce an iterative method for approximating a common solution of monotone inclusion problem and fixed point of Bregman nonspreading mapping in a reflexive Banach space and prove a strong convergence of the sequence generated by our iterative algorithm. This result extends many works on finding common solution of monotone inclusion problem and fixed problem of nonlinear mapping in a real Hilbert space to a reflexive Banach space.
2. Preliminaries
The Bregman projection [22] of x∈intdomf onto the nonempty, closed, and convex subset C ⊂int (domf is defined as the necessarily unique vector ProjCfx∈C satisfying(30)DfProjCfx,x=infDfy,x:y∈C.
It is known from [37] that z=ProjCfx if and only if(31)∇fx−∇fz,y−z≤0,for ally∈C.
We also have(32)Dfy,ProjCfx+DfProjCfx,x≤Dfy,x,for allx∈E,y∈C.
Note that if E is a Hilbert space and fx=1/2x2, then the Bregman projection of x onto C, i.e., argminy−x:y∈C, is the metric projection PC.
Lemma 1 [37].
Let f be totally convex on int (domf). Let C be a nonempty, closed, and convex subset of int (domf) and x∈intdomf; if z∈C, then the following conditions are equivalent:
z=ProjCfx
∇fx−∇fz,z−y≥0forally∈C
Dfy,z+Dfz,x≤Dfy,xforally∈C
Let f:E⟶ℝ∪+∞ be a convex and Gâteaux differentiable function. The function f is said to be totally convex at x∈intdomf if its modulus of totally convexity at x, that is, the function vf:intdomf×0,+∞ defined by(33)vfx,t≔infDfy,x:y∈domf,y−x=t,is positive for any t>0. The function f is said to be totally convex when it is totally convex at every point x∈intdom f. In addition, the function f is said to be totally convex on bounded set if vfB,t is positive for any nonempty bounded subset B, where the modulus of total convexity of the function f on the set B is the function vf:intdomf×0,+∞ defined by(34)vfB,t≔infvfx,t:x∈B∩ domf.
For further details and examples on totally convex functions, see [37–39].
Let f:E⟶ℝ be a convex, Legendre, and Gâteaux differentiable function and let the function Vf:E×E∗⟶0,∞ associated with f (see [23, 40]) be defined by(35)Vfx,x∗=fx−x∗,x+f∗x∗,∀x∈E,x∗∈E∗.
Then Vf is nonnegative and Vfx,x∗=Dfx,∇f∗x∗,∀x∈E,x∗∈E∗. Furthermore, by the subdifferential inequality, we have (see [41])(36)Vfx,x∗+y∗,∇f∗x∗−x≤Vfx,x∗+y∗,∀x∈E,x∗,y∗∈E∗.
In addition, if f:E⟶−∞,+∞ is a proper lower semicontinuous function, then f∗:E∗⟶−∞,+∞ is a proper weak∗ lower semicontinuous and convex function (see [42]). Hence, Vf is convex in the second variable. Thus, for all z∈E,(37)Dfz,∇f∗∑i=1Nti∇fxi≤∑i=1NtiDfz,xi,where xii=1N⊂E and ti⊂0,1 with ∑i=1Nti=1.
Lemma 2 (see [43]).
Let r>0 be a constant and let f:E⟶ℝ be a continuous uniformly convex function on bounded subsets of E. Then(38)f∑k=0∞αkxk≤∑k=0∞αkfxk−αiαjρrxi−xj,for all i,j∈ℕ∪0, xk∈Br, αk∈0,1, and k∈ℕ∪0 with ∑k=0∞αk=1, where ρr is the gauge of uniform convexity of f.
Recall that a function f is said to be sequentially consisted (see [37]) if for any two sequences xn and yn in E such that the first one is bounded,(39)limn⟶∞Dfyn,xn=0⟹limn⟶∞yn−xn=0.
The following lemma follows from [44].
Lemma 3.
If domf contains at least two points, then the function f is totally convex on bounded sets if and only if the function f is sequentially consistent.
Lemma 4 (see [45]).
Let f:E⟶−∞,+∞ be a Legendre function and let A:E⟶2E∗ be a BISM operator such that A−10∗≠∅. Then the following statements hold:
A−10∗=FAf
For any w∈A−10∗ and x∈domAf, we have
(40)Dfw,Afx+DfAfx,x≤Dfw,x.
Remark 1.
If the Legendre function f is uniformly Fréchet differentiable and bounded on bounded subsets of E, then the antiresolvent Af is a single-valued BSNE operator which satisfies FAf=F^Af (cf. [31]).
Lemma 5 (see [46]).
If f:E⟶ℝ is uniformly Fréchet differentiable and bounded on bounded subsets of E, then ∇f is uniformly continuous on bounded subsets of E from the strong topology of E to the strong topology of E∗.
Lemma 6 (see [44]).
Let f:E⟶ℝ be a Gâteaux differentiable and totally convex function. If x1∈E and the sequence Dfxn,x1 is bounded, then the sequence xn is also bounded.
Lemma 7 (see [45]).
Assume that f:E⟶ℝ is a Legendre function which is uniformly Fréchet differentiable and bounded on bounded subset of E. Let C be a nonempty, closed, and convex subset of E. Let Ti:1≤i≤N be BSNE operators which satisfy F^Ti=FTi for each 1≤i≤N and let T≔wnTN−1…T1. If(41)∩FTi:1≤i≤N,and FT are nonempty, then T is also BSNE with FT=F^T.
Lemma 8 (Demiclosedness principle [30]).
Let C be a nonempty subset of a reflexive Banach space. Let g:E⟶ℝ be a strict convex, Gâteaux differentiable, and locally bounded function. Let T:C⟶E be a Bregman nonspreading mapping. If xn⇀p in C and limn⟶∞Txn−xn=0, then p∈FT.
Lemma 9 (see [47]).
Assume an is a sequence of nonnegative real numbers satisfying(42)an+1≤1−tnan+tnδn∀n≥0,where tn is a sequence in 0,1 and δn is a sequence in ℝ such that
∑n=o∞tn=∞
limsupn⟶∞δn≤0
Then, limn⟶∞an=0.
Lemma 10 [48].
Let an be a sequence of real numbers such that there exists a nondecreasing subsequence ni of n, that is, ani≤ani+1 for all i∈ℕ. Then there exists a nondecreasing sequence mk⊂ℕ such that mk⟶∞, and the following properties are satisfied for all (sufficiently large number k∈ℕ): amk≤amk+1 and ak≤amk+1, mk=maxj≤k:aj≤aj+1.
3. Main ResultsTheorem 1.
Let C be a nonempty, closed, and convex subset of a real reflexive Banach space E and f:E⟶ℝ a Legendre function which is bounded, uniformly Fréchet differentiable, and totally convex on bounded subsets of E. Let A:E⟶2E∗ be a Bregman inverse strongly monotone operator, B:E⟶2E∗ be a maximal monotone operator, and T:C⟶C be a Bregman nonspreading mapping. Suppose Γ≔FResAf∘Af∩FT≠∅. Let γn⊂0,1 and αn,βn, and δn be sequences in 0,1 such that αn+βn+δn=1. Given u∈E and x1∈C arbitrarily, let xn and yn be sequences in E generated by(43)yn=∇f∗γn∇fxn+1−γn∇fTxn,xn+1=ProjCf∇f∗αn∇fu+βn∇fyn+δn∇fResBf∘Afyn,n≥1.
Suppose the following conditions are satisfied:
limn⟶∞αn=0 and ∑n=1∞αn=∞
1−αna<δn,αn≤b<1,a∈0,1/2
0≤c<liminfn⟶∞γn≤limsupn⟶∞γn<1
Then xn converges strongly to ProjΓfu, where ProjΓf is the Bregman projection of E onto Γ.
Proof.
First we observe that FResBf∘Af=A+B−10 and FResBf∘Af=FResBf∩FAf. Thus, since ResBf and Af are BSNE operators and FResBf∩FAf=A+B−10≠∅, it then follows from Lemma 7 that ResBf∘Af is BSNE and FResBf∘Af=F^ResBf∘Af.
We next show that xn and yn are bounded.
Let p∈Γ, then from (43), we have(44)Dfp,yn=Dfp,∇f∗γn∇fxn+1−γn∇fTxn≤γnDfp,xn+1−γnDfp,Txn≤γnDfp,xn+1−γnDfp,xn=Dfp,xn.
Hence Dfp,xn is bounded. Therefore, by Lemma 6, xn is also bounded, and consequently, yn is also bounded.
We now show that xn converges strongly to x¯=ProjΓfu. To do this, we first show that if there exists a subsequence xni of xn such that xni⇀q∈C, then q∈Γ.
Let s=sup∇fxn,∇fTxn and ρs∗:E∗⟶ℝ be the gauge of uniform convexity of the conjugate function f∗. From Lemma 2 and (9), we have(46)Dfp,yn≤Dfp,∇f∗γn∇fxn+1−γn∇fTxn=Vfp,γn∇fxn+1−γn∇fTxn=fp−p,γn∇fxn+1−γn∇fTxn+f∗γn∇fxn+1−γn∇fTxn≤γnfp−γnp,∇fxn+γnf∗∇fxn+1−γnfp−1−γnp,∇fTxn+1−γnf∗∇fTxn−γn1−γnρs∗∇fxn−∇fTxn=γnDfp,xn+1−γnDfp,Txn−γn1−γnρs∗∇fxn−∇fTxn≤Dfp,xn−γn1−γnρs∗∇fxn−∇fTxn.
Thus, from (45), we have(47)Dfp,xn+1≤αnDfp,u+1−αnDfp,xn−γn1−γnρs∗∇fxn−∇fTxn.
We consider the following two cases for the rest of the proof.
Case A.
Suppose Dfp,xn is monotonically nonincreasing. Then, Dfp,xn converges and Dfp,xn−Dfp,xn+1⟶0 as n⟶∞. Thus, from (47), we have(48)1−αn1−γnγnρs∗∇fxn−∇fTxn≤αnDfp,u−Dfp,xn+Dfp,xn−Dfp,xn+1.
Since αn⟶0,n⟶∞, then we have(49)limn⟶∞γn1−γnρs∗∇fxn−∇fTxn=0,and hence, by condition (iii) and the property of ρs∗, we have(50)limn⟶∞∇fxn−∇fTxn=0.
Since ∇f∗ is uniformly norm-to-norm continuous on bounded subset of E∗, we have(51)limn⟶∞xn−Txn=0.
Since ∇f∗ is uniformly norm-to-norm continuous on bounded subsets of E∗, we have that(53)limn⟶∞xn−yn=0.
Now, let wn=∇f∗βn/1−αn∇fyn+δn/1−αn∇fResBf∘Afyn, then(54)Dfp,wn=Dfp,∇f∗βn1−αn∇fyn+δn1−αn∇fResBf∘Afyn≤βn1−αnDfp,yn+δn1−αnDfp,ResBf∘Afyn≤βn+δn1−αnDfp,yn=Dfp,yn.
Therefore, we have(55)0≤Dfp,xn−Dfp,wn=Dfp,xn−Dfp,xn+1+Dfp,xn+1−Dfp,wn≤Dfp,xn−Dfp,xn+1+αnDfp,u+1−αnDfp,wn−Dfp,wn=Dfp,xn−Dfp,xn+1+αnDfp,u−Dfp,wn⟶0,as n⟶∞.
More so(56)Dfp,wn≤βn1−αnDfp,yn+δn1−αnDfp,ResBf∘Afyn=Dfp,yn−1−βn1−αnDfp,yn+δn1−αnDfp,ResBf∘Afyn≤Dfp,xn+δn1−αnDfp,ResBf∘Afyn−Dfp,yn.
Since 1−αna<δn and αn≤b<1, we have(57)aDfp,yn−Dfp,ResBf∘Afyn<δn1−αnDfp,yn−Dfp,ResBf∘Afyn≤Dfp,xn−Dfp,wn⟶0,as n⟶∞.
Thus,(58)Dfp,yn−Dfp,ResBf∘Afyn⟶0,as n⟶∞.
Therefore, since ResBf∘Af is BSNE, we have that limn⟶∞Dfyn,ResBf∘Afyn=0, which implies that(59)limn⟶∞yn−ResBf∘Afyn=0.
Setting un=∇f∗αn∇fu+βn∇fyn+δn∇fResBf∘Afyn, for each n≥1, we have(60)Dfyn,un=Dfyn,∇f∗αn∇fu+βn∇fyn+δn∇fResBf∘Afyn≤αnDfyn,u+βnDfyn,yn+δnDfyn,ResBf∘Afyn⟶0.
Thus,(61)limn⟶∞yn−un=0.
Therefore, from (47), we have(62)un−xn≤un−yn+yn−xn⟶0,n⟶∞.
Moreover, since xn+1=ProjCfun, then(63)Dfp,xn+1+Dfxn+1,un≤Dfp,un,and therefore, we have that(64)Dfxn+1,un≤Dfp,un−Dfp,xn+1≤αnDfp,u+βnDfp,yn+δnDfp,ResBf∘Afyn−Dfp,xn+1=αnDfp,u+1−αnDfp,yn−Dfp,xn+1≤αnDfp,u−Dfp,xn+Dfp,xn−Dfp,xn+1⟶0,n⟶∞,which implies(65)xn+1−un⟶0,n⟶∞.
Hence,(66)xn+1−xn≤xn+1−un+un−xn⟶0,n⟶∞.
Since xn is bounded, there exists a subsequence xni of xn such that xni converges weakly to q∈C as n⟶∞. Since limn⟶∞xni−Txni=0, it follows from Lemma 8 that q∈FT. Also, since xni−yni⟶0, it implies that yni also converges weakly to q∈E. Therefore, from (59), we have that q∈FResBf∘Af, and hence, q∈Γ=FT∩FResBf∘Af.
Next, we show that xn converges strongly to x¯=ProjΓfu.
Now from (43), we have(67)Dfx¯,xn+1≤Dfx¯,∇f∗αn∇fu+βn∇fyn+δn∇fResBf∘Afyn=Vfx¯,αn∇fu+βn∇fyn+δn∇fResBf∘Afyn≤Vfx¯,αn∇fu+βn∇fyn+δn∇fResBf∘Afyn−αn∇fu−∇fx¯−−αn∇fu−∇fx¯,∇f∗αn∇fu+βn∇fyn+δn∇fResBf∘Afyn−x¯=Vfx¯,αn∇fx¯+βn∇fyn+δn∇fResBf∘Afyn+αn∇fu−∇fw,un−x¯=Dfx¯,∇f∗αn∇fx¯+βn∇fyn+δn∇fResBf∘Afyn+αn∇fu−∇fx¯,un−x¯=αnDfx¯,x¯+βnDfx¯,yn+δnDfx¯,ResBf∘Afyn+αn∇fu−∇fx¯,un−x¯≤βnDfx¯,yn+δnDfx¯,yn+αn∇fu−∇fx¯,un−x¯=1−αnDfx¯,yn+αn∇fu−∇fx¯,un−x¯≤1−αnDfx¯,xn+αn∇fu−∇fx¯,un−x¯.
Choose a subsequence xnj of xn such that(68)limsupn⟶∞∇fu−∇fx¯,xn−x¯=limj⟶∞∇fu−∇fx¯,xnj−x¯.
Since xnj⇀q, it follows from Lemma 1(ii) that(69)limsupn⟶∞∇fu−∇fx¯,xn−x¯=limj⟶∞∇fu−∇fx¯,xnj−x¯=∇fu−∇fx¯,q−x¯≤0.
Since un−xn⟶0, n⟶∞, then,(70)limsupn⟶∞∇fu−∇fx¯,un−x¯≤0.
Hence, by Lemma 9 and (67), we conclude that Dfx¯,xn⟶0,n⟶∞. Therefore, xn converges strongly to x¯=ProjΓfu.
Case B.
Suppose that there exists a subsequence nj of n such that(71)Dfxnj,w<Dfxnj+1,w,for all j∈ℕ. Then, by Lemma 10, there exists a nondecreasing sequence mk⊂ℕ with mk⟶∞ as n⟶∞ such that(72)Dfp,xmk≤Dfp,xmk+1,Dfp,xk≤Dfp,xmk+1,for all k∈ℕ. Following the same line of arguments as in Case I, we have that(73)limk⟶∞Txmk−xmk=0,limk⟶∞ResBfAfymk−ymk=0,limk⟶∞wmk−xmk=0,limsupk⟶∞∇fu−∇fp,wmk−p≤0.
From (67), we have(74)Dfp,xmk+1≤1−αmkDfp,xmk+αmk∇fu−∇fp,wmk−p.
Since Dfp,xmk≤Dfp,xmk+1, it follows from (74) that(75)αmkDfp,xmk≤Dfp,xmk−Dfp,xmk+1+αmk∇fu−∇fx∗,wmk−p≤αmk∇fu−∇fp,wmk−p.
Since αmk>0, we obtain(76)Dfp,xmk≤∇fu−∇fp,wmk−p.
Then from (73), it follows that Dfp,xmk⟶0 as k⟶∞. Combining Dfp,xmk⟶0 with (74), we obtain Dfp,xmk+1⟶0 as k⟶∞. Since Dfp,xk≤Dfp,xmk+1 for all k∈ℕ, we have xk⟶p as k⟶∞, which implies that xn⟶p as n⟶∞.
Therefore, from the above two cases, we conclude that xn converges strongly to x¯=ProjΓfu.
This completes the proof.
Corollary 1.
Let C be a nonempty, closed, and convex subset of a real reflexive Banach space E and f:E⟶ℝ a Legendre function which is bounded, uniformly Fréchet differentiable, and totally convex on bounded subsets of E. Let A:E⟶2E∗ be a Bregman inverse strongly monotone operator, B:E⟶2E∗ be a maximal monotone operator, and T:C⟶C be a Bregman firmly nonexpansive mapping. Suppose Γ≔FResAf∘Af∩FT≠∅. Let γn⊂0,1 and αn,βn, and δn be sequences in 0,1 such that αn+βn+δn=1. Given u∈E and x1∈C arbitrarily, let xn and yn be sequences in E generated by(77)yn=∇f∗γn∇fxn+1−γn∇fTxn,xn+1=ProjCf∇f∗αn∇fu+βn∇fyn+δn∇fResBf∘Afyn,n≥1.
Suppose the following conditions are satisfied:
limn⟶∞αn=0 and ∑n=1∞αn=∞
1−αna<δn,αn≤b<1,a∈0,1/2
0≤c<liminfn⟶∞γn≤limsupn⟶∞γn<1
Then, xn converges strongly to ProjΓfu, where ProjΓf is the Bregman projection of E onto Γ.
Corollary 2.
Let C be a nonempty, closed, and convex subset of a real Hilbert space H. Let A:H⟶H be a single-valued 1-inverse strongly monotone operator, B:E⟶2E∗ be a maximal monotone operator, and T:C⟶C be a firmly nonexpansive mapping. Suppose Γ≔FI+B−1I−A∩FT≠∅. Let γn⊂0,1 and αn,βn, and δn be sequences in 0,1 such that αn+βn+δn=1. Given u∈E and x1∈C arbitrarily, let xn and yn be sequences in E generated by(78)yn=γnxn+1−γnTxn,xn+1=PCαnu+βnyn+δnI+B−1I−Ayn,n≥1.
Suppose the following conditions are satisfied:
limn⟶∞αn=0 and ∑n=1∞αn=∞
1−αna<δn,αn≤b<1,a∈0,1/2
0≤c<liminfn⟶∞γn≤limsupn⟶∞γn<1
Then, xn converges strongly to PΓu, where PΓ is the metric projection of H onto Γ.
4. Application
In this section, we apply our result to obtain a common solution of variational inequality problem (VIP) and equilibrium problem (EP) in real reflexive Banach spaces.
Let C be a nonempty, closed, and convex subset of a real reflexive Banach space E. Suppose g:C×C⟶ℝ is a bifunction that satisfies the following conditions:
A1 gx,x=0,∀x∈C
A2 gx,y+gy,x≤0,∀x,y∈C
A3 limsupt↓0gtz+1−tx,y≤gx,y,∀x,y,z∈C
A4 gx,. is convex and lower semicontinuous, for each x∈C.
The equilibrium problem with respect to g is to find x¯∈C such that(79)gx¯,y≥0,∀y∈C.
We denote the set of solutions of (79) by EP(g). The resolvent of a bifunction g:C×C⟶ℝ that satisfies A1−A4 (see [49]) is the operator Tgf:E⟶2C defined by(80)Tgfx≔z∈C:gz,y+∇fz−∇fx,y−z≥0,∀y∈C.
Lemma 11 ([27], Lemma 1, 2).
Let f:E⟶−∞,∞ be a coercive Legendre function and let C be a nonempty, closed, and convex subset of E. Suppose the bifunction g:C×C⟶ℝ satisfies A1−A4, then
domTgf=E.
Tgf is single valued
Tgf is Bregman firmly nonexpansive
FTgf=EPg
EPg is a closed and convex subset of C
Dfu,Tgfx+DfTgfx,x≤Dfu,x, for all x∈E and for all u∈FTgf
Let A:E⟶E∗ be a Bregman inverse strongly monotone mapping and let C be a nonempty, closed, and convex subset of domA. The variational inequality problem corresponding to A is to find x∈C, such that(81)Ax∗,y−x∗≥0,∀y∈C.
The set of solutions of (81) is denoted by VIC,A.
Lemma 12 (see [25, 46]).
Let A:E⟶E∗ be a Bregman inverse strongly monotone mapping and f:E⟶−∞,∞ be a Legendre and totally convex function that satisfies the range condition. If C is a nonempty, closed, and convex subset of domA∩intdomf, then
PCf∘Af is Bregman relatively nonexpansive mapping
FPCf∘Af=VIC,A
Now let iC be the indicator function of a closed convex subset C of E, defined by(82)iCx=0,x∈C,+∞,otherwise.
The subdifferential of the indicator function ∂iCx¯=NCx¯, where C is a closed subset of a Banach space E and NC⊂E∗ is the normal cone defined by(83)NCx¯=v∈E∗:v,x−x¯≤0,for allx∈C,x¯∈C,∅,x∉C.
The normal cone NC is maximal monotone and the resolvent of the normal cone corresponds to the Bregman projection (see [50], Example 4.4) that is ResNCf=ProjCf.
Therefore, if we let B=NC and T=Tgf, then the iterative algorithm (77) becomes(84)yn=∇f∗γn∇fxn+1−γn∇fTgfxn,xn+1=ProjCf∇f∗αn∇fu+βn∇fyn+δn∇fProjCf∘Afyn,n≥1.
Thus, from Corollary 1, we obtain a strong convergence result for approximating a point x∈VIC,A∩ EPg.
Data Availability
No data were used to support this study.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
Acknowledgments
The work of the first author is based on the research supported wholly by the National Research Foundation (NRF) of South Africa (Grant no. 111992). The third author acknowledges the financial support from the Department of Science and Technology and National Research Foundation, Republic of South Africa Center of Excellence in Mathematical and Statistical Sciences (DST-NRF CoE-MaSS) (postdoctoral fellowship) (Grant no. BA 2018/012). Opinions expressed and conclusions arrived are those of the authors and are not necessarily to be attributed to the NRF and CoE-MaSS.
CombettesP. L.FloudesC. A.PardolisP. M.Fejer monotonicity in convex optimization20092Boston, MA, USAKluwer106114KaplonA.TichatschkeR.A general view on proximal point methods to variational inequalities in Hilbert spaces-iterative regularization and approximation20012305332LemaireB.PenstJ. P.The proximal point algorithm198987Boston, MA, USABirkhauser7387RockafellarR. T.Argumented Lagrangians and application of the proximal point algorithm in convex programming1976129711610.1287/moor.1.2.972-s2.0-0016950796RockafellarR. T.Monotone operators and the proximal point algorithm197614877898TossingsP.The perturbed proximal point algorithm and some of its applications199429212515910.1007/bf012041802-s2.0-0000115211ZeidlerE.1995New York, NY, USASpinger-VerlagMartinetB.Brève communication. Régularisation d’inéquations variationnelles par approximations successives19704R315415810.1051/m2an/197004r301541GülerO.On the convergence of the proximal point algorithm for convex minimization199129240341910.1137/0329022BruckR. E.ReichS.Nonexpansive projections and resolvents of accretive operators in Banach spaces19773459470BauschkeH. H.MatouškováE.ReichS.Projection and proximal point methods: convergence results and counterexamples200456571573810.1016/j.na.2003.10.0102-s2.0-0347269125EslamianM.VahidiJ.General proximal-point algorithm for monotone operators201768111715172610.1007/s11253-017-1322-x2-s2.0-85018776965MorosanuG.A proximal point algorithm revisted and extended20141612478489ShehuY.Convergence theorems for maximal monotone operators and fixed point problems in Banach spaces201423928529810.1016/j.amc.2014.04.0832-s2.0-84901286460LópezG.Martín-MárquezV.WangF.XuH.-K.Forward-Backward splitting methods for accretive operators in Banach space201220122510923610.1155/2012/1092362-s2.0-84864914328LionsP. L.MercierB.Splitting algorithms for the sum of two nonlinear operators197816696497910.1137/0716071BauschkeH. H.CombettesP. L.2011New York, NY, USACMS Books in Mathematics. SpingerBauschkeH. H.BorweinJ. M.CombettesP. L.Essential smoothness, essential strict convexity, and Legendre functions in Banach spaces20013461564710.1142/s02191997010005242-s2.0-23044531274BonnansJ. F.ShapiroA.2000New York, NY, USASpinger-VerlagBauschkeH. H.BorweinJ. M.Legendre functions and the method of random Bregman projections199742767ReemD.ReichS.Solutions to inexact resolvent inclusion problems with applications to nonlinear analysis and optimization201867233737110.1007/s12215-017-0318-62-s2.0-85050461772BregmanL. M.The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming19677320021710.1016/0041-5553(67)90040-72-s2.0-49949144765CensorY.LentA.An iterative row-action method for interval convex programming198134332135310.1007/bf009346762-s2.0-0019587748ReemD.ReichS.De PierroA.Re-examination of Bregman functions and new properties of their divergences201968127934810.1080/02331934.2018.15432952-s2.0-85057332760ReichS.1996Boca Raton, FL, USACRC PressBauschkeH. H.BorweinJ. M.CombettesP. L.Bregman monotone optimization algorithms200342259663610.1137/s03630129024071202-s2.0-1942509685ReichS.SabachS.Two strong convergence theorems for Bregman strongly nonexpansive operator in reflexive Banach space201013112213510.1016/j.na.2010.03.0052-s2.0-77955430773KohsakaF.TakahashiW.Fixed point theorems for a class of nonlinear mappings related to maximal monotone operators in Banach spaces200891216617710.1007/s00013-008-2545-82-s2.0-50849119713TakahashiW.WangN.-C.YaoJ.-C.Fixed point theorems and Convergence theorems for generalized nonspreading mappings in Banach spaces201211115918310.1007/s11784-012-0074-32-s2.0-84864532151NaraghiradE.ChingN.-C.YaoJ.-C.Applications of bregman-opial property to bregman nonspreading mappings in Banach spaces201420141427286710.1155/2014/2728672-s2.0-84893798187ReichS.SabachS.BauschkeH. H.Existence and approximation of fixed points of Bregman firmly nonexpansive mappings in reflexive Banach spaces201149New York, NY, USASpinger301316ButnariuD.KassayG.A proximal-projection method for finding zeros of set-valued operators20084742096213610.1137/0706820712-s2.0-55549108063PlubtiengS.PunpaengR.A new iterative method for equilibrium problems and fixed point problems of nonexpansive mappings and monotone mappings2008197254855810.1016/j.amc.2007.07.0752-s2.0-39449089199QinX.ChoY. J.KangS. M.Convergence theorems of common elements for equilibrium problems and fixed point problems in Banach spaces20092252030IidukaH.A new iterative algorithm for the variational inequality problem over the fixed point set of a firmly nonexpansive mapping201059687388510.1080/023319309028841582-s2.0-77954840484MaingéP. E.A hybrid extragradient-viscosity method for monotone operators and fixed point problems20084731499151510.1137/0606753192-s2.0-55549132648ButnariuD.ResmeritaE.Bregman distances, totally convex function and a method for solving operator equations in Banach spaces20062006398491910.1155/AAA/2006/849192-s2.0-31844434904BorweinJ. M.ReichS.SabachS.A characterization of Bregman firmly nonexpansive operators using a new monotonicity concept201112161184ButnariuD.IusemA. N.2000Dordrecht, NetherlandsKluwer AcademicAlberY. I.KartsatosA. G.Metric and generalized projection operators in Banach spaces: properties and applications1996New York, NY, USAMarcel Dekker1550KohsakaF.TakahashiW.Proximal point algorithms with Bregman functions in Banach spaces20056505523PhelpsR. P.Convex functions, monotone operators, and differentiability199313642nd EditionBerlin, GermanySpringer-VerlagNaraghiradE.YaoJ.-C.Bregman weak relatively nonexpansive mappings in Banach spaces2013201314310.1186/1687-1812-2013-1412-s2.0-84890087506ReichS.SabachS.Two strong convergence theorems for a proximal method in reflexive Banach space2010311224410.1080/016305609034998522-s2.0-77949538440KassayG.ReichS.SabachS.Iterative methods for solving systems of variational inequalities in reflexive Banach spaces20112141319134410.1137/1108200022-s2.0-84856027098ReichS.SabachS.A strong convergence theorem for a proximal-type algorithm in reflexive Banach space200910491485XuH. K.An iterative approach to quadratic optimization2003116365967810.1023/a:10230736215892-s2.0-0037269525MaingéP. E.The viscosity approximation process for quasi-nonexpansive mappings in Hilbert spaces2010591747910.1016/j.camwa.2009.09.0032-s2.0-73449122599ReichS.SabachS.A projection method for solving nonlinear problems in reflexive Banach spaces20119110111610.1007/s11784-010-0037-52-s2.0-79651473591BauschkeH. H.WangX.YaoL.General resolvents for monotone operators: characterization and extension2010Madison, WI, USAMedical Physics Publishing5774