Open Access

Damped projection method for split common fixed point problems

Journal of Inequalities and Applications20132013:123

https://doi.org/10.1186/1029-242X-2013-123

Received: 20 October 2012

Accepted: 4 March 2013

Published: 22 March 2013

Abstract

The paper deals with the split common fixed-point problem (SCFP) introduced by Censor and Segal. Motivated by Eicke’s damped projection method, we propose a cyclic iterative scheme and prove its strong convergence to a solution of SCFP under some mild assumptions. An application of the proposed method to multiple-set split feasibility problems is also included.

1 Introduction

The split feasibility problem (SFP) [1] consists of finding an element x ˆ H satisfying
x ˆ C , A x ˆ Q ,
(1)
where C and Q are closed convex subsets in Hilbert spaces and K , respectively. Moreover, if C and Q are the intersections of finitely many closed convex subsets, then the problem is known as the multiple-set split feasibility problem (MSFP) [2]. Note that SFP and MSFP model image retrieval [3] and intensity-modulated radiation therapy [4], and they have recently been investigated by many researchers (see, e.g., [511]). One method for solving SFP is Byrne’s CQ algorithm [5]: For any initial guess x 1 H , define { x n } recursively by
x n + 1 = P C ( x n λ A ( I P Q ) A x n ) ,
(2)
where P C stands for the metric projection onto C, I is the identity operator on K and λ is the step-size satisfying 0 < λ < 2 A 2 . By using Hundal’s counterexample, Xu [12] showed the CQ algorithm does not converge strongly in infinite-dimensional spaces. Motivated by Byrne’s CQ algorithm, Wang and Xu [13] proposed the following iterative method: For any initial guess x 1 H , define { x n } recursively by
x n + 1 = P C [ ( 1 α n ) ( x n λ A ( I P Q ) A x n ) ] ,
(3)
where { α n } ( 0 , 1 ) satisfies lim n α n = 0 ; n = 1 α n = ; either n = 1 | α n + 1 α n | < or lim n | α n + 1 α n | / α n = 0 . It is worth noting that this algorithm is in fact a generalization of Eicke’s damped projection method [14] for solving convexly constrained linear inverse problems (see [15]). Motivated by Krasnosel’skii-Mann’s iteration, Dang and Gao [16] proposed the following algorithm: For any initial guess x 1 H , define { x n } recursively by
x n + 1 = ( 1 β n ) x n + β n P C [ ( 1 α n ) ( x n λ A ( I P Q ) A x n ) ] ,
(4)

where { α n } ( 0 , 1 ) satisfies (i) lim n α n = 0 , n = 1 α n = ; (ii) lim n | α n + 1 α n | = 0 ; (iii)  0 < lim inf n β n lim sup n β n < 1 . It is clear that such an algorithm is an extension of (3). However, algorithm (4) fails to include the original one (3) because of condition (iii).

In the case where C and Q in (1) are the intersections of finitely many fixed-point sets of nonlinear operators, problem (1) is called by Censor and Segal [17] the split common fixed-point problem (SCFP). More precisely, SCFP requires to seek an element x ˆ H satisfying
x ˆ i = 1 p Fix ( U i ) , A x ˆ j = 1 s Fix ( T j ) ,
(5)
where p , s N , Fix ( U i ) and Fix ( T j ) denote the fixed point sets of two classes of nonlinear operators U i : H H , i = 1 , , p and T j : K K , j = 1 , , s . In this situation, Byrne’s CQ algorithm does not work because the metric projection onto fixed point sets is generally not easy to calculate. To solve the two-set SCFP, that is, p = s = 1 in (5), Censor and Segal [17] proposed the following iterative method: For any initial guess x 1 H , define { x n } recursively by
x n + 1 = U ( x n λ A ( I T ) A x n ) ,
(6)

where λ > 0 is known as the step-size. They proved that if U and T in (6) are directed operators, then λ should be chosen in ( 0 , 2 A 2 ) . Some further generations of this algorithm were studied by Moudafi [18] for demicontractive operators and by Wang-Xu [19] for finitely many directed operators.

We note that the existing algorithms for SCFP have only weak convergence in the framework of infinite-dimensional spaces (see [18, 19]). However, as pointed by Bauschke and Combettes [20], norm convergence of the algorithm is much more desirable than weak convergence in some applied sciences. It is therefore of interest to seek modifications of these algorithms so that strong convergence is guaranteed. Following the damped projection method, we propose in this paper a new iterative scheme and prove its strong convergence to a solution of SCFP. An application of our method to multiple-set split feasibility problems is also included. This enables us to cover some recent results on split feasibility problems.

2 Preliminary and notation

Throughout this paper, I denotes the identity operator on , Fix ( T ) the set of fixed points of an operator T, ‘→’ strong convergence, and ‘’ weak convergence. Given a positive integer p, denote by [ n ] : = ( n mod p ) the mod function taking values in { 1 , 2 , , p } .

Definition 1 An operator T : H H is called nonexpansive if T x T y x y , x , y H ; firmly nonexpansive if T x T y 2 x y 2 ( I T ) x ( I T ) y 2 , x , y H .

Definition 2 Assume that T : H H is a nonlinear operator. Then I T is said to be demiclosed at zero, if, for any { x n } in , the following implication holds:
x n x , ( I T ) x n 0 ] x Fix ( T ) .

Clearly, firm nonexpansiveness implies nonexpansiveness. It is well known that nonexpansive operators are demiclosed at zero (cf. [21]).

Definition 3 Let T : H H be an operator with Fix ( T ) . Then T is called directed if z T x , x T x 0 , z Fix ( T ) , x H ; ν-demicontractive with ν ( , 1 ) if T x z 2 x z 2 + ν ( I T ) x 2 , z Fix ( T ) , x H .

Lemma 1 (Bauschke-Combettes [20])

An operator T : H H is directed if and only if one of following inequalities holds for all z Fix ( T ) and x H :
(7)
(8)
It is clear that demicontractive operators include directed operators, while the latter include firmly nonexpansive operators with nonempty fixed-point sets. The concept of directed operators was introduced by Bauschke and Combettes [20]. Such a class of operators is important because they include many types of nonlinear operators arising in applied mathematics. For instance, the metric projections onto a closed convex subset. Recall that the metric projection, denoted by P C : H C , is defined by
P C x = arg min y C x y , x H .
It is well known that P C x is characterized by the variational inequality
x P C x , P C x z 0 , z C .
(9)

Lemma 2 (Wang-Xu [19])

Assume that A : H K is a bounded linear operator and T : K K is a directed operator. Let V λ = I λ A ( I T ) A with λ > 0 . Then
Fix ( V λ ) = A 1 ( Fix ( T ) ) ,

whenever A 1 ( Fix ( T ) ) : = { x H : A x Fix ( T ) } is nonempty.

Lemma 3 Assume that A : H K is a bounded linear operator and T : K K is a directed operator. Let V λ = I λ A ( I T ) A with 0 < λ < 2 A 2 . If A 1 ( Fix ( T ) ) is nonempty, then
V λ x z 2 x z 2 2 λ A 2 λ A 2 V λ x x 2 ,
(10)

for all z A 1 ( Fix ( T ) ) and x H .

Proof Since A z Fix ( T ) , it follows from (8) that
( I V λ ) x , x z = λ ( I T ) A x , A x A z λ ( I T ) A x 2 1 λ A 2 ( I V λ ) x 2 .
Consequently,
V λ x z 2 = ( x z ) + ( V λ x x ) 2 = x z 2 + V λ x x 2 + 2 x z , V λ x x x z 2 2 λ A 2 λ A 2 V λ x x 2 .

Hence the proof is complete. □

We end this section by a useful lemma.

Lemma 4 (Xu [22])

Let { a n } be a nonnegative real sequence satisfying
a n + 1 ( 1 α n ) a n + α n b n ,
where { α n } ( 0 , 1 ) and { b n } are real sequences. Then a n 0 provided that
  1. (i)
    n α n =
    , lim n α n = 0 ,
     
  2. (ii)
    lim n b n 0
    or α n | b n | < .
     

3 Algorithm and its convergence analysis

In this section, we consider the following problem.

Problem 1 Find an element x ˆ H satisfying
x ˆ i = 1 p Fix ( U i ) , A x ˆ i = 1 p Fix ( T i ) ,
(11)

where p is a positive integer and ( U i ) i = 1 p , ( T i ) i = 1 p are two classes of directed operators such that U i I and T i I are demiclosed at zero for every i = 1 , 2 , , p .

We remark here that problem (11) is a special case of (5). However, this is not restrictive. Indeed, following an idea in [19], one can easily extend the results to the general case. We now present our algorithm for SCFP: Take x 1 H and define a sequence { x n } by the iterative procedure:
x n + 1 = ( 1 β n ) x n + β n U n [ ( 1 α n ) ( x n λ n A ( I T n ) A x n ) ] ,
(12)

where U n : = U [ n ] , T n : = T [ n ] and { α n } ( 0 , 1 ) , { β n } [ 0 , 1 ] , { λ n } R + are properly chosen real sequences.

Theorem 1 Assume that the following conditions hold:
  1. (i)
    lim inf n β n > 0
    ,
     
  2. (ii)
    lim n α n = 0
    , n = 1 α n = ,
     
  3. (iii)
    0 < λ ̲ λ n λ ¯ < 2 A 2
    .
     

If the solution set of problem (11) denoted by Ω is nonempty, then the sequence { x n } generated by (12) converges strongly to P Ω ( 0 ) .

Proof We first show the boundedness of { x n } . To see this, let z = P Ω ( 0 ) and set V n = I λ n A ( I T n ) A , y n = ( 1 α n ) V n x n . Hence
y n z = ( 1 α n ) ( V n x n z ) α n z ( 1 α n ) x n z + α n z .
Since U n is directed, it follows that
x n + 1 z = ( 1 β n ) ( x n z ) + β n ( U n y n z ) ( 1 β n ) x n z + β n y n z .
Adding up these inequalities, we have
x n + 1 z ( 1 α n β n ) x n z + α n β n z .

By induction, the sequence { x n } is bounded, and so is { y n } .

Next we show the following key inequality:
s n + 1 ( 1 α n β n ) s n + 2 α n β n z y n , z c n ,
(13)
where s n = x n z 2 and
c n = β n [ ( 1 α n ) ( 2 λ ¯ A 2 ) λ ¯ A 2 ( I V n ) x n 2 + ( I U n ) y n 2 ] .
Indeed, in view of Lemma 3, we arrive at
U n y n z 2 y n z 2 ( I U n ) y n 2 , V n x n z 2 x n z 2 2 λ n A 2 λ n A 2 ( I V n ) x n 2 x n z 2 2 λ ¯ A 2 λ ¯ A 2 ( I V n ) x n 2 .
(14)
On the other hand, we deduce that
y n z 2 = ( 1 α n ) ( V n x n z ) α n z 2 ( 1 α n ) V n x n z 2 + 2 α n z y n , z ,
(15)
where we use the subdifferential inequality, and also that
x n + 1 z 2 = ( 1 β n ) ( x n z ) + β n ( U n y n z ) 2 ( 1 β n ) x n z 2 + β n U n y n z 2 .
(16)

Adding up (14)-(16), we thus get inequality (13).

Finally, we prove s n 0 . To see this, let { s n k } be a subsequence such that it includes all elements in { s n } with the property: each of them is less than or equal to the term after it. Following an idea developed by Maingé [23], we consider two possible cases on such a sequence.

Case 1. Assume that { s n k } is finite. Then there exists N N such that s n > s n + 1 for all n N , and therefore { s n } must be convergent. It follows from (13) that
c n M α n β n + ( s n s n + 1 ) ,
where M > 0 is a sufficiently large real number. Consequently, both ( I V n ) x n and ( I U n ) y n converge to zero. We have
y n x n y n V n x n + V n x n x n = α n V n x n + V n x n x n 0 ,
which implies
x n + 1 x n U n y n y n + y n x n 0 .
Take a subsequence { y n k } of { y n } so that
lim sup n z , z y n = lim k z , z y n k .
Without loss of generality, we assume that { y n k } weakly converges to an element y . Let an index i { 1 , 2 , , p } be fixed. Noticing that the pool of indexes is finite, we can find a subsequence { y m k } of { y n } such that y m k y and [ m k ] = i for all k. Since ( I U i ) y m k = ( I U m k ) y m k 0 , we thus use the demiclosedness of I U i at zero to conclude that y Fix ( U i ) . On the other hand, we deduce from (8) that
( I T i ) A x m k 2 ( I T i ) A x m k , A x m k A z = A ( I T i ) A x m k , x m k z 1 λ ̲ ( I V i ) x m k x m k z M ( I V i ) x m k 0 .
As x m k y m k 0 , the weak continuity of A yields that A x m k A y , which together with the demiclosedness of I T i at zero enables us to deduce A y Fix ( T i ) . Since the index i is arbitrary, we therefore conclude y Ω . Consequently,
lim sup n z , z y n = z , z y 0 ,
where the inequality uses (9). It then follows from (13) that
s n + 1 ( 1 α n β n ) s n + 2 α n β n z y n , z .

We therefore apply Lemma 4 to conclude s n 0 .

Case 2. Assume now that { s n k } is infinite. Let n N be fixed. Then there exists k N such that n k n n k + 1 . By the choice of { s n k } , we see that s n k + 1 is the largest one among { s n k , s n k + 1 , , s n k + 1 } ; in particular,
s n k s n k + 1 and s n s n k + 1 .
(17)
Then we deduce from (13) that c n k M α n k so that
( I V n k ) x n k + ( I U n k ) y n k 0 .
(18)
In a similar way to case 1, we deduce x n k + 1 x n k 0 and
lim sup n u z , y n k z 0 .
Since by (17) s n k s n k + 1 , it follows from (13) that
s n k 2 z , z y n k .
(19)
Hence lim k s n k 0 so that s n k 0 . Moreover,
| s n k + 1 s n k | = | x n k + 1 z 2 x n k z 2 | x n k + 1 x n k ( x n k + 1 z + x n k z ) 0 ,

which immediately implies s n k + 1 0 . Consequently, s n 0 follows from (17) and the proof is complete. □

We next use our algorithm to approximate a solution to the two-set SCFP: Find an element x ˆ H such that
x ˆ Fix ( U ) , A x ˆ Fix ( T ) ,
(20)

where U : H H and T : K K are directed operators so that U I and T I are demiclosed at zero.

Corollary 2 Suppose that the following conditions hold:
  1. (i)
    lim n α n = 0
    , n = 1 α n = ,
     
  2. (ii)
    0 < λ ̲ λ n λ ¯ < 2 A 2
    .
     
Then the sequence { x n } , generated by
x n + 1 = U [ ( 1 α n ) ( x n λ n A ( I T ) A x n ) ] ,
(21)

converges strongly to P Ω ( 0 ) , whenever such point exists.

4 Some applications

In this section, we extend our result to SCFP for demicontractive operators recently considered by Moudafi [18].

Problem 2 Find an element x ˆ H satisfying
x ˆ i = 1 p Fix ( U i ) , A x ˆ i = 1 p Fix ( T i ) ,
(22)

where p is a positive integer and ( U i ) i = 1 p , ( T i ) i = 1 p are respectively ν i -demicontractive and κ i -demicontractive operator so that U i I and T i I are demiclosed at zero for every i = 1 , 2 , , p .

The following lemma states a relation between directed and demicontractive operators.

Lemma 5 Let ν ( , 1 ) and τ ( 0 , 1 ν 2 ] . If T is ν-demicontractive, then T τ : = ( 1 τ ) I + τ T is directed.

Proof For z Fix ( T ) , we deduce that
T τ x z 2 = ( 1 τ ) ( x z ) + τ ( T x z ) 2 = ( 1 τ ) x z 2 + τ T x z 2 τ ( 1 τ ) ( I T ) x 2 x z 2 τ ( 1 ν τ ) ( I T ) x 2 = x z 2 1 τ / ( 1 ν ) τ / ( 1 ν ) ( I T τ ) x 2 x z 2 ( I T τ ) x 2 .

Then the result follows from Lemma 1. □

We now propose an algorithm to solve problem (22). Take x 1 H and define a sequence { x n } by the iterative procedure
x n + 1 = U τ n [ ( 1 α n ) ( x n λ A ( I T γ n ) A x n ) ] ,
(23)

where { α n } ( 0 , 1 ) , U τ n = ( 1 τ [ n ] ) I + τ [ n ] U [ n ] and T γ n = ( 1 γ [ n ] ) I + γ [ n ] T [ n ] . By using the previous lemma, we can easily extend our result to demicontractive operators.

Theorem 3 Let 0 < τ i 1 ν i 2 and 0 < γ i 1 κ i 2 for every i = 1 , 2 , , p . Assume that the following conditions hold:
  1. (i)
    lim n α n = 0
    , n = 1 α n = ,
     
  2. (ii)
    0 < λ < min { 1 κ i 2 γ i A 2 : 1 i p }
    .
     

If the solution set of problem (22) denoted by Ω is nonempty, then the sequence { x n } generated by (23) converges strongly to P Ω ( 0 ) .

Remark 1 Theorem 3 also holds true if we relax hypothesis (ii) above as 0 < λ < min { 1 κ i γ i A 2 : 1 i p } .

We next consider the multiple-set split feasibility problem (MSFP): Find an element x ˆ H satisfying
x ˆ i = 1 p C i , A x ˆ i = 1 p Q i ,
(24)
where { C i } i = 1 p and { Q i } i = 1 p are closed convex subsets in and K , respectively. Take x 1 H and define a sequence { x n } by the iterative procedure
x n + 1 = ( 1 β n ) x n + β n P C n [ ( 1 α n ) ( x n λ n A ( I P Q n ) A x n ) ] ,
(25)

where C n : = C [ n ] , Q n : = Q [ n ] , and { α n } ( 0 , 1 ) , { β n } [ 0 , 1 ] , { λ n } R + are properly chosen real sequences.

Theorem 4 Assume that the following conditions hold:
  1. (i)
    lim inf n β n > 0
    ,
     
  2. (ii)
    lim n α n = 0
    , n = 1 α n = ,
     
  3. (iii)
    0 < λ ̲ λ n λ ¯ < 2 A 2
    .
     

If the solution set of MSFP denoted by Ω is nonempty, then the sequence { x n } generated by (25) converges strongly to P Ω ( 0 ) .

Proof We note that the metric projection P C is firmly nonexpansive, which implies P C is directed and I P C is demiclosed at zero. Hence, by using Theorem 1, one can immediately get the desired result. □

Remark 2 Theorem 4 covers [[16], Theorem 3.1], and we relax the condition on { β n } as lim inf n β n > 0 . Moreover, the choice of variable { λ n } is more flexible than the fixed one. Also, we cover the result of [19] and remove one condition posed on { α n } : either n = 1 | α n + 1 α n | < or lim n | α n + 1 α n | / α n = 0 .

Declarations

Acknowledgements

We would like to express our sincere thanks to the referees for their valuable suggestions. This work is supported by the National Natural Science Foundation of China, Tianyuan Foundation (11226227), the Basic Science and Technological Frontier Project of Henan (122300410268) and the Foundation of Henan Educational Committee (12A110016).

Authors’ Affiliations

(1)
Department of Mathematics, Luoyang Normal University

References

  1. Censor Y, Elfving T: A multiprojection algorithm using Bregman projections in a product space. Numer. Algorithms 1994, 8(2):221–239. 10.1007/BF02142692MathSciNetView ArticleGoogle Scholar
  2. Censor Y, Elfving T, Kopf N, Bortfeld T: The multiple-sets split feasibility problem and its applications for inverse problems. Inverse Probl. 2005, 21(6):2071–2084. 10.1088/0266-5611/21/6/017MathSciNetView ArticleGoogle Scholar
  3. Byrne C: A unified treatment of some iterative algorithms in signal processing and image reconstruction. Inverse Probl. 2003, 20: 103.MathSciNetView ArticleGoogle Scholar
  4. Censor Y, Bortfeld T, Martin B, Trofimov A: A unified approach for inversion problems in intensity-modulated radiation therapy. Phys. Med. Biol. 2006, 51(10):2353–2365. 10.1088/0031-9155/51/10/001View ArticleGoogle Scholar
  5. Byrne C: Iterative oblique projection onto convex sets and the split feasibility problem. Inverse Probl. 2002, 18(2):441–453. 10.1088/0266-5611/18/2/310MathSciNetView ArticleGoogle Scholar
  6. Lopez G, Martin V, Xu HK: Perturbation techniques for nonexpansive mappings with applications. Nonlinear Anal., Real World Appl. 2009, 10(4):2369–2383. 10.1016/j.nonrwa.2008.04.020MathSciNetView ArticleGoogle Scholar
  7. Qu B, Xiu N: A note on the CQ algorithm for the split feasibility problem. Inverse Probl. 2005, 21(5):1655–1665. 10.1088/0266-5611/21/5/009MathSciNetView ArticleGoogle Scholar
  8. Schöpfer F, Schuster T, Louis A: An iterative regularization method for the solution of the split feasibility problem in Banach spaces. Inverse Probl. 2008., 24(5): Article ID 055008Google Scholar
  9. Wang F, Xu H: Choices of variable steps of the CQ algorithm for the split feasibility problem. Fixed Point Theory 2011, 12: 489–496.MathSciNetGoogle Scholar
  10. Xu HK: A variable Krasnosel’skii-Mann algorithm and the multiple-set split feasibility problem. Inverse Probl. 2006, 22(6):2021–2034. 10.1088/0266-5611/22/6/007View ArticleGoogle Scholar
  11. Yang Q: The relaxed CQ algorithm solving the split feasibility problem. Inverse Probl. 2004, 20(4):1261–1266. 10.1088/0266-5611/20/4/014View ArticleGoogle Scholar
  12. Xu HK: Iterative methods for the split feasibility problem in infinite-dimensional Hilbert spaces. Inverse Probl. 2010., 26(10): Article ID 105018Google Scholar
  13. Fenghui W, Hong-Kun X: Approximating curve and strong convergence of the CQ algorithm for the split feasibility problem. J. Inequal. Appl. 2010., 2010: Article ID 102085Google Scholar
  14. Eicke B: Iteration methods for convexly constrained ill-posed problems in Hilbert space. Numer. Funct. Anal. Optim. 1992, 13(5–6):413–429. 10.1080/01630569208816489MathSciNetView ArticleGoogle Scholar
  15. Engl HW, Hanke M, Neubauer A 375. In Regularization of Inverse Problems. Springer, Berlin; 1996.View ArticleGoogle Scholar
  16. Dang Y, Gao Y: The strong convergence of a KM-CQ-like algorithm for a split feasibility problem. Inverse Probl. 2010., 27: Article ID 015007Google Scholar
  17. Censor Y, Segal A: The split common fixed point problem for directed operators. J. Convex Anal. 2009, 16(2):587–600.MathSciNetGoogle Scholar
  18. Moudafi A: The split common fixed-point problem for demicontractive mappings. Inverse Probl. 2010., 26(5): Article ID 055007Google Scholar
  19. Wang F, Xu HK: Cyclic algorithms for split feasibility problems in Hilbert spaces. Nonlinear Anal. 2011, 74(12):4105–4111. 10.1016/j.na.2011.03.044MathSciNetView ArticleGoogle Scholar
  20. Bauschke HH, Combettes PL: A weak-to-strong convergence principle for Fejé-monotone methods in Hilbert spaces. Math. Oper. Res. 2001, 26(2):248–264. 10.1287/moor.26.2.248.10558MathSciNetView ArticleGoogle Scholar
  21. Goebel K, Kirk WA 28. In Topics in Metric Fixed Point Theory. Cambridge University Press, Cambridge; 1990.View ArticleGoogle Scholar
  22. Xu HK: Iterative algorithms for nonlinear operators. J. Lond. Math. Soc. 2002, 66(1):240–256. 10.1112/S0024610702003332View ArticleGoogle Scholar
  23. Maingé PE: Strong convergence of projected subgradient methods for nonsmooth and nonstrictly convex minimization. Set-Valued Anal. 2008, 16(7):899–912. 10.1007/s11228-008-0102-zMathSciNetView ArticleGoogle Scholar

Copyright

© Cui et al.; licensee Springer 2013

This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.