Abstract
The purpose of this article is to introduce two iterative algorithms for finding the least norm fixed point of nonexpansive mappings. We provide two algorithms, one implicit and another explicit, from which strong convergence theorems are obtained in Hilbert spaces. Then we apply these algorithms to solve some convex optimization problems. Furthermore, we use them to solve some split feasibility problems. The results of this article extend and improve several results presented in the literature in the recent past.
Mathematics Subject Classification (2000): 47H09; 47H10; 90C25.
Keywords:
nonexpansive mapping; minimumnorm; fixed point; convex optimization1 Introduction
Throughout, H is a real Hilbert space with inner product 〈·,·〉 and induced norm ║·║. Let C be a nonempty closed convex subset of H. Then, a mapping T, from C into itself is said to be a nonexpansive mapping if
for any x, y ∈ C. Fix(T) denotes the fixed point set of T, that is Fix(T) = {x ∈ C : Tx = x}. Iterative methods for finding fixed points of nonexpansive mappings are an important topic in the theory of nonexpansive mappings and have wide applications in a number of applied areas, such as, image reconstruction in computerized tomography [1], optics and neural networks [2], collective sensing [3], and image denoising and deblurring [4] etc. However, the Picard sequence often fails to converge even in the weak topology. To overcome the difficulties, the KrasnoselskiiMann iteration algorithm become prevail. This algorithm generates from an arbitrary initial guess x_{0 }∈ C and a sequence {x_{n}} by the recursive formula
where {α_{n}} is a sequence in (0, 1). Reich [5] proved that if X is a uniformly convex Banach space with a Fréchet differential norm and if {α_{n}} is chosen such that , then the sequence {x_{n}} defined by (1.1) converges weakly to a fixed point of T. On the other hand, Maingé [6] proposed the socalled inertial KrasnoselskiiManntype algorithm as follows
where I : H → H is the identity operator, x_{0}, x_{1 }∈ H, {θ_{n}} ⊂ [0, 1], {α_{n}} ⊂ (0, 1) are relaxation factor. The proposed algorithm unifies KrasnoselskiiMann iteration and inertial type extrapolation. He established some weak convergence theorems of the sequence {x_{n}} generated by (1.2). It is clear that if θ_{n}= 0 for all n, then the algorithm (1.2) reduces to the KrasnoselskiiMann iteration (1.1). The sequence {υ_{n}} is intended to speed up the convergence of algorithms. As a matter of fact, the above algorithms (1.1) and (1.2) have only weak convergence except in a finite dimensional space. To obtain strong convergence in the setting of an infinite dimensional Hilbert or Banach spaces, there exist several iterative algorithms to nonexpansive mappings (e.g., Viscosity iteration algorithm [7], Hybrid projection algorithm [8], Hybrid steepest descent algorithm [9], Halperntype iteration algorithm [10,11], Shrinking projection algorithm [12], etc.). In general, the nonexpansive mapping may have more than one fixed point. Without loss of generality, we may assume that (otherwise, C is additionally bounded), then Fix(T) is closed and convex (It is worth mentioning that Ferreira [13] proved that Fix(T) is closed and convex even in a strictly convex Banach space which includes Hilbert spaces as a special case). So there exists a unique x* ∈ Fix(T) satisfies:
That is, x* is the minimumnorm fixed point of T. In other words, x* is the metric projection of the origin into Fix(T), i.e., x* = P_{Fix(T)}0. It is an interesting thing to construct iterative sequence to find the minimumnorm fixed point of a nonexpansive mapping T, i.e., the minimumnorm solutions of x = Tx. Recently, Yao and Xu [14] and Cui and Liu [15] independently introduced two iterative methods (one implicit and one explicit) for finding the minimumnorm fixed point of nonexpansive mapping which is defined on a closed convex subset C of H. The proposed algorithms are based on the wellknown Browder's iterative method [16] and Halpern's iterative method [17]. We next briefly recall the Browder's iterative method and the Halpern's iterative method. Browder [16] introduced an implicit scheme as follows. Let u ∈ C and t ∈ (0, 1), x_{t}be the unique fixed point in C of the contraction T_{t}from C into C:
Browder proved that the strong limit of {x_{t}} as t → 0^{+ }is the fixed point of T which is nearest from Fix(T) to u, i.e., lim_{t→0+ }x_{t}= P_{Fix(T)}u. Besides, Halpern [17] introduced an explicit scheme. Let x_{0 }∈ C, define a sequence {x_{n}} by the following:
where {α_{n}} ⊂ (0, 1). It is known that the sequence {x_{n}} generated by (1.4) converges in norm to the same limit P_{Fix(T)}u as Browder's implicit scheme (1.3) if the sequence {α_{n}} satisfies the following conditions:
(C1) lim_{n→∞ }α_{n}= 0;
It is noticed that the Browder's and the Halpern's iterative methods do find the minimumnorm fixed point x* of T if 0 ∈ C. However, if 0 ∉ C, then neither Browder's nor Halpern's methods works to find the minimumnorm element x*. The reason is simple: if 0 ∉ C, we cannot take u = 0 either in (1.3) or (1.4) since the contraction T_{t}x = (1  t)Tx is no longer a selfmapping of C or the point (1  α_{n})Tx_{n}may not belong to C and consequently, {x_{n}_{+1}} may be undefined. In order to overcome this difficulties caused by possible exclusion of the origin from C, Yao and Xu [14] and Cui and Liu [15] put forward the improvement strategy to impose the metric projection P_{C}on the right side of the (1.3) and (1.4) when u = 0. The role of the metric projection P_{C}is to pull the substituted sequence back to C, then the iterative sequences are welldefined.
Motivated and inspired by the above studies, the purpose of this article is to consider another way to ensure the well defined of the iterative sequence. That is, we replace the closed convex subset C by a closed convex cone C (C is said to be a closed convex cone if (i) C is closed and convex; (ii) αx ∈ C, for all α ≥ 0 and x ∈ C; (iii) C ≠ {0}). We present new strongly convergent methods for approximating minimumnorm fixed point of nonexpansive mappings. The proposed algorithms consist of two types and generated by the following. For each λ ∈ (0, 1), (i) The implicit method
(ii) The explicit method
where {α_{n}} ⊂ (0, 1).
We prove that the sequence {x_{n}} generated by (1.5) and (1.6) converge strongly to the element of minimal norm fixed point of nonexpansive mappings. As applications, we provide iterative processes for solving the constrained convex optimization problem. And we use them to solve some split feasibility problems which attracted great attention in recent years. Our results improve and generalize the corresponding results of Cui and Liu [15], Yao and Xu [14], and Wang and Xu [18] et al.
2 Preliminaries
Let H be a Hilbert space with inner product 〈·,·〉 and norm ║·║, and let C be a nonempty closed convex subset of H.
We use the following notions in the sequel:
(i) ⇀ for weak convergence and → for strong convergence;
(ii) denotes the weak ωlimit set of {x_{n}}.
Recall that the orthogonal projection P_{C}x of x onto C is defined by the following
The orthogonal projection has the following wellknown properties. For a given x ∈ H,
(i) 〈x  P_{C}x, z  P_{C}x〉 ≤ 0, for all z ∈ C;
(ii) ║P_{C}x  P_{C}y║^{2 }≤ 〈P_{C}x  P_{C}y, x  y〉, for all x, y ∈ H.
We shall make use of the following results.
Lemma 2.1. (Demiclosedness principle of nonexpansive mapping) Let T : C → C a nonexpansive mapping with . If x_{n }⇀ x and (I  T)x_{n}→ 0, then x = Tx.
Lemma 2.2. (see, [19]) Let {x_{n}} and {y_{n}} be bounded sequences in a Banach space E and let {β_{n}} be a sequence in [0, 1] with 0 < lim inf β_{n }≤ lim sup β_{n }< 1. Suppose x_{n}_{+1 }= β_{n}y_{n }+ (1  β_{n})x_{n }for all n ≥ 0 and
Then lim_{n→∞ }║y_{n} x_{n}║ = 0.
Lemma 2.3. (see, [20]) Let {a_{n}} be an nonnegative real sequences satisfying the following inequality:
where {γ_{n}} ⊂ (0, 1) such that , and lim sup_{n→∞ }δ_{n}≤ 0. Then lim_{n→∞ }a_{n}= 0.
3 Main results
First, we prove the following strong convergence theorem by using the implicit method (1.5) for finding the minimumnorm fixed point of a nonexpansive mapping T.
Theorem 3.1. Let C be a closed convex cone of a real Hilbert space H. Let T : C → C be a nonexpansive with . For each t ∈ (0, 1), let x_{t }be the unique fixed point in C of the contraction T_{λ }: = (1  t)(λT + (1  λ)I), where λ ∈ (0, 1) is a constant. Then x_{t }converges strongly to the minimumnorm fixed point of T as t → 0^{+}.
Proof. Take p ∈ Fix(T), from (1.5), we have
that is,
Hence, {x_{t}} is bounded and so is {Tx_{t}}. Next, we prove that ║x_{t} Tx_{t}║→ 0 as t → 0^{+}. In fact, from (1.5), we have
that is,
Next we show that {x_{t}} is relatively normcompact as t → 0^{+}. Since {x_{t}} is bounded, there exists a null sequence {t_{n}} ⊂ (0, 1) such that . By Lemma 2.1 and (3.1), then .
It turns out that
Since , we may substitute for and t_{n}for t in (3.2) to obtain that .
Hence, {x_{t}} is indeed relatively compact (as t → 0^{+}) in the norm topology.
Observe that (3.2) is equivalent to
Hence,
This implies that
Therefore, , where x* is the minimumnorm fixed point of T, and we conclude that x_{t}→ x* as t → 0^{+}. This completes the proof. □
Now, we are in the position to prove the strong convergence of the explicit method (1.6). Our proofs of this theorem closely follows proofs given in [11] for some related results.
Theorem 3.2. Let C be a closed convex cone of a real Hilbert space H. Let T : C → C be a nonexpansive mapping and Fix(T) is nonempty. Assume that the sequence {α_{n}} ⊂ (0, 1) satisfies the following conditions:
(i) lim_{n→∞ }α_{n}= 0;
Then the sequence {x_{n}} generated by the algorithm (1.6) strongly converges to a fixed point of T which is of minimal norm.
Proof. First we prove that the sequence {x_{n}} is bounded. Let p ∈ Fix(T). By (1.6), we have
By induction,
for all n ≥ 0. Then {x_{n}} is bounded. Therefore, {Tx_{n}} is also bounded.
Let , then the iterative sequence (1.6) is equivalent to
Observe that lim_{n→∞}(α_{n}+ (1  α_{n})λ) = λ, then
Thus, {y_{n}} is bounded. Consequently, we have
From the fact that {x_{n}} and {Tx_{n}} are bounded sequences and lim_{n→∞ }α_{n}= 0, then
With the help of Lemma 2.2, we obtain that lim_{n→∞ }║y_{n} x_{n}║ = 0. Therefore,
On the other hand,
From the above inequality, and (3.4), we obtain
Next, we prove that lim sup_{n→∞ }〈x*  x_{n}, x*〉 ≤ 0. To achieve this, we take a subsequence of {x_{n}} such that
Since {x_{n}} is bounded, without loss of generality, we may assume that . Consequently,
Notice that lim_{n→∞ }║x_{n} Tx_{n}║ = 0. By the demiclosedness principle of nonexpansive mapping T, we have x' ∈ Fix(T). Since x* = P_{Fix(T)}0. It follows from the properties of Projection operator that
By (1.6), we have
Observe that
Therefore, by (3.6) and (3.7), we get
By the condition of (ii) and the inequality (3.5), we can apply Lemma 2.3 to (3.8) and conclude that {x_{n}} converges strongly to x* as n → ∞, that is, the minimumnorm fixed point of T. This completes the proof. □
Remark 3.1. (i) If the closed convex cone C in Theorems 3.1 and 3.2 are replaced by closed convex C with 0 ∈ C, then Theorems 3.1 and 3.2 are still true because of the iterative sequence (1.5) and (1.6) are welldefined now.
(ii) Theorem 3.2 also improve the [[14], Theorem 3.2] and [[15], Theorem 3.3], in which the restrictions are removed.
4 Some applications
From now on, we apply the proposed methods for approximating the minimumnorm solution of convex function and to split feasibility problems. Let's recall that the standard constrained convex optimization problem as follows:
where f : C → R is a convex, Fréchet differentiable function, C is closed convex subset of H.
It is known that the above optimization problem is equivalent to the following variational inequality:
where ∇f : H → H is the gradient of f.
It is wellknown that the optimality condition (4.2) is equivalent to the following fixed point problem:
where P_{C}is the metric projection onto C and μ > 0 is a positive constant. Based on the fixed point problem, we deduce the projected gradient method.
Using Theorems 3.1 and 3.2, we immediately obtain the following result.
Theorem 4.1. Suppose that the solution set of (4.1) is nonempty. Let the objective function f be convex, fréchet differentiable and its gradient ∇f is Lipschitz continuous with Lipschitz constant L. In addition, if 0 ∈ C or C is closed convex cone. Let μ ∈ (0, 2/L),
(i) For each t ∈ (0, 1), let x_{t }be the unique solution of the fixed point equation
Then {x_{t}} converges in norm as t → 0^{+ }to the minimumnorm solution of the minimization (4.1)
(ii) Define a sequence {x_{n}} by the following
where λ ∈ (0, 1) and the sequence {α_{n}} ⊂ (0, 1) satisfies conditions in Theorem 3.2. Then the sequence {x_{n}} converges strongly to the minimumnorm solution of the minimization (4.1).
Proof. Since ∇f is Lipschitz continuous with Lipschitz constant L, then the P_{C}(I  μ∇f) is nonexpansive mapping (see [[21], Sect. 4). Replace the mapping T in (1.5) and (1.6) with P_{C}(I  μ∇f). Therefore, the conclusion of Theorem 4.1 follows from Theorems 3.1 and 3.2 immediately. □
Next, we give an application of Theorem 4.1 to the split feasibility problem (say SFP, for short) which was introduced by Censor and Elfving [22].
where C and Q are nonempty closed convex subset of Hilbert space H_{1 }and H_{2}, respectively. A : H_{1 }→ H_{2 }is a bounded linear operator.
It is clear that x* is a solution to the split feasibility problem (4.5) if and only if x* ∈ C and Ax*  P_{Q}Ax* = 0. We define the proximity function f by
and consider the convex optimization problem
Then, x* solves the split feasibility problem (4.5) if and only if x* solves the minimization (4.6) with the minimize equal to 0. Byrne [21] introduced the socalled CQ algorithm to solve the (SFP).
where 0 < μ < 2/ρ(A*A) and where P_{C}denotes the projection onto C and ρ(A*A) is the spectral radius of the selfadjoint operator A*A. He obtained that the sequence {x_{n}} generated by (4.7) converges weakly to a solution of the (SFP).
In order to obtain strong convergence iterative sequence to solve the (SFP), Xu [23] investigated the following algorithm:
where 0 < μ < 2/ρ(A*A). He showed that when the sequence {α_{n}} satisfies the conditions (C1)(C3), then {x_{n}} converges strongly to the projection of u onto the solution set of the (SFP). In particular, if u = 0 in the algorithm (4.8), then the corresponding algorithms converges strongly to the minimal norm solution of the (SFP). Lately, Wang and Xu [18] introduced a modification of CQ algorithm (4.7) with strong convergence by introducing an approximating curve for the (SFP) in infinite dimensional Hilbert space, and obtained the minimumnorm solution of the (SFP) as the strong limit of the approximating curve. The sequence {x_{n}} is generated by the iterative algorithm
where {α_{n}} ⊂ (0, 1) such that (C1)(C3).
Applying Theorem 4.1, we obtain the following result which improve the corresponding results of Xu [23] and Wang and Xu [18].
Theorem 4.2. Assume that the split feasibility problem (4.5) is consistent. In addition, if 0 ∈ C or C is closed convex cone. Let the sequence {x_{n}} be generated by
where the sequence {α_{n}} ⊂ (0, 1) satisfies the conditions: (i) lim _{n→∞}, α_{n}= 0, (ii) , λ ∈ (0, 1) and μ ∈ (0, 2/ρ(A*A)), where ρ(A*A) denotes the spectral radius of the selfadjoint operator A*A. Then the sequence {x_{n}} converges strongly to the minimumnorm solution of the split feasibility problem (4.5).
Proof. By the definition of the proximity function f, we have
and ∇f is Lipschitz continuous (Lemma 8.1 of [21]), i.e.,
where L = ρ(A*A). Then the iterative scheme (4.10) is equivalent to
Due to Theorem 4.1, we have the conclusion immediately. □
Remark 4.1. Theorem 4.2 extends the corresponding results of Wang and Xu [18] and Xu [23] by discarding the assumption " or lim_{n → ∞ }(α_{n}/α_{n+1}) = 1".
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
All authors read and approved the final manuscript.
Acknowledgements
The authors would like to thank the anonymous referees for their constructive comments and helpful suggestions, which greatly improved the original manuscript of this article. The authors are also deeply grateful to Prof. Hongkun Xu (Editor) for managing the review process. This study was supported partly by the National Natural Science Foundations of China (11101204, 11102078), the Natural Science Foundations of Jiangxi Province (2009GZS0021, CA201107114) and the Youth Science Funds of The Education Department of Jiangxi Province (GJJ12141).
References

Herman, GT: Fundamentals of Computerized Tomography: Image Reconstruction from Projections. Springer, New York (2009)

Stark, H, Yang, Y: Vector Space Projections: A Numerical Approach to Signal and Image Processing, Neural Nets, and Optics. WileyInterscience, New York (1998)

Li, X: Finegranularity and spatially adaptive regularization for projection based image deblurring. IEEE Trans Image Process. 20(4), 971–983 (2011)

Beck, A, Teboulle, M: Fast gradientbased algorithms for constrained total variation denoising and deblurring problems. IEEE Trans Image Process. 18(11), 2419–2434 (2009)

Reich, S: Weak convergence theorems for nonexpansive mappings in Banach spaces. J Math Anal Appl. 67, 274–276 (1979)

Maingé, PE: Convergence theorems for inertial KMtype algorithms. J Comput Appl Math. 219, 223–236 (2008)

Moudafi, A: Viscosity approximations methods for fixedpoints problems. J Math Anal Appl. 241, 46–55 (2000)

Nakajo, K, Takahashi, W: Strong convergence theorems for nonexpansive mappings and nonexpansive semigroups. J Math Anal Appl. 279, 372–379 (2003)

Yamada, I, Ogura, N: Hybrid steepest descent method for the variational inequality problem over the fixed point set of certain quasinonexpansive mappings. Number Funct Anal Optim. 25(78), 619–655 (2004)

Chidume, CE, Chidume, CO: Iterative approximation of fixed points of nonexpansive mappings. J Math Anal Appl. 318, 288–295 (2006)

Suzuki, T: A sufficient and necessary condition for Halperntype strong convergence to fixed point of nonexpansive mappings. Proc Am Math Soc. 135(1), 99–106 (2007)

Takahashi, W, Takeuchi, Y, Kubota, R: Strong convergence theorems by hybrid methods for families of nonexpansive mappings in Hilbert spaces. J Math Anal Appl. 341, 276–286 (2008)

Ferreira, PJSG: The existence and uniqueness of the minimumnorm solution to certain linear and nonlinear problems. Signal Process. 55(1), 137–139 (1996)

Yao, YH, Xu, HK: Iterative methods for finding minimumnorm fixed points of nonexpansive mappings with applications. Optim. 60(6), 645–658 (2011)

Cui, YL, Liu, X: Notes on Browder's and Halpern's methods for nonexpansive mappings. Fixed Point Theory. 10(1), 89–98 (2009)

Browder, FE: Convergence theorems for sequences of nonlinear operators in Banach spaces. Math Z. 100, 201–225 (1967)

Halpern, B: Fixed points of nonexpanding maps. Bull Am Math Soc. 73, 957–961 (1967)

Wang, F, Xu, HK: Approximating curve and strong convergence of the CQ Algorithm for the split feasibility problem. Journal of Inequalities and Applications. 2010, 102085 (2010)

Suzuki, T: Strong convergence of Krasnoselskii and Mann's type sequences for oneparameter nonexpansive semigroups without Bochner integrals. J Math Anal Appl. 305, 227–239 (2005)

Xu, HK: Another control condition in an iterative method for nonexpansive mappings. Bull Aust Math Soc. 65, 109–113 (2002)

Byrne, C: A unified treatment of some iterative algorithms in signal processing and image reconstruction. Inverse Probl. 20, 103–120 (2004)

Censor, Y, Elfving, T: A multiprojection algorithm using Bregman projections in a product space. Numer Alg. 8, 221–239 (1994)

Xu, HK: A variable Krasnosel'skiiMann algorithm and the multipleset split feasibility problem. Inverse probl. 22, 2021–2034 (2006)