Professional Documents
Culture Documents
Editors:
J.-M. Morel, Cachan
B. Teissier, Paris
Iterative Methods
for Fixed Point Problems
in Hilbert Spaces
123
Andrzej Cegielski
University of Zielona Góra
Faculty of Mathematics,
Computer Science and Econometrics
Zielona Góra, Poland
Mathematics Subject Classification (2010): 47-02, 49-02, 65-02, 90-02, 47H09, 47J25, 37C25,
65F10, 65K15, 90C25
In this monograph we deal with iteration methods for finding fixed points (if they
exist) of nonexpansive operators defined on a Hilbert space, i.e., operators T having
the property
d.T x; T y/ cd.x; y/ for all x; y 2 X (1)
and for some constant c 2 Œ0; 1. The origin of these methods dates back to 1920,
when Stefan Banach (1892–1945) formulated his famous contraction mapping
principle. Banach proved that if T W X ! X is a contraction (an operator satisfying
(1) with c < 1) defined on a complete metric space, then T has a unique fixed point
x , i.e., a point for which T x D x . Furthermore, for any x 2 X a sequence
fT k xg1
kD0 converges geometrically to x . Many practical problems can be reduced
to finding a fixed point of a nonexpansive operator or a common fixed point of a
family of nonexpansive operators. A simple example is a system of linear equations
Ax D b. Any solution of this system can be identified with a common fixed point of
(nonexpansive) operators of orthogonal projections onto hyperplanes corresponding
to particular equations of the system.
Under some additional conditions a nonexpansive operator has a fixed point (see,
e.g., the Browder–Göhde–Kirk fixed point theorem), but the Banach theorem does
not guarantee the convergence of the sequence fT k xg1 kD0 . Therefore, it is of great
interest to develop methods for finding fixed points of nonexpansive operators.
The first iterative method for solving a linear system was proposed in 1937 by
a Polish mathematician, Stefan Kaczmarz (1890–1945), in a very short paper (3
pages) “Angenäherte Auflösung von Systemen linearer Gleichungen” (“Approxi-
mate solution of systems of linear equations”) published in Bulletin International
de l’Académie Polonaise des Sciences et des Lettres. A year later, an Italian math-
ematician, Gianfranco Cimmino (1908–1989), proposed another iterative method
for a linear system. He published his result in a paper “Calcolo approssimato per
le soluzioni dei sistemi di equazioni lineari” (“Approximate computation of the
solutions of linear systems”) in a La Ricerca Scientifica. As opposed to earlier
methods for solving systems of linear equations, both methods are motivated rather
by geometrical operations than algebraic ones. The main operation in both methods
ix
x Preface
is a cyclic (in the Kaczmarz method) or simultaneous (in the Cimmino method)
orthogonal projection onto hyperplanes described by particular equations of the
linear system. The results of Kaczmarz and Cimmino were disregarded for several
decades except few references. A great interest in these results started in the 1970s,
when it suddenly turned out that the Kaczmarz and Cimmmino methods can be
efficiently applied in computed tomography (CT), because the mathematical model
of CT can be reduced to a solution of large systems of linear equations with sparse
and unstructured matrices, and the methods behave well for such systems. At the
end of the twentieth century the interest in the Kaczmarz and Cimmino results was
still increasing. Both results have become fundamental to modern iterative methods
for fixed point problems for nonexpansive operators.
The third result which influenced the development of this area concerns alterna-
tive projections onto two subspaces of a Hilbert space. The result was published
in 1950 by John von Neumann in a paper “Functional Operators—Vol. II. The
Geometry of Orthogonal Spaces” in Annals of Mathematical Studies. Von Neumann
proved that a sequence generated by his method converges to the projection of
the starting point onto the intersection of the subspaces. The results of Kaczmarz,
Cimmino, and von Neumann have been generalized several times in the last decades.
Today it is known that the convergence holds for an essentially wider class of
operators than orthogonal projections onto hyperplanes, e.g., for nonexpansive
operators or for quasi-nonexpansive operators, satisfying some additional condi-
tions. All three methods belong today to classical iterative methods for finding
fixed points of nonexpansive operators defined on a Hilbert space. These methods
served as the basis for several methods, e.g., for: the Landweber method, projected
Landweber method, Douglas–Rachford method, sequential projection methods,
methods of cyclic and simultaneous subgradient projections, Dos Santos method
of extrapolated simultaneous subgradient projections, reflection-projection method,
surrogate projection method, and many others. Irrespective of their theoretical value,
they have found application in many areas of mathematics, physics, and technology.
The most spectacular application of the methods is an intensity-modulated radiation
therapy (IMRT).
Iterative methods for finding fixed points of nonexpansive operators in Hilbert
spaces have been described in many publications. In this monograph we try to
present the methods in a consolidated way. We introduce some classes of operators,
give their properties, define iterative methods generated by operators from these
classes, and present general convergence theorems. On this basis we present the
conditions under which particular methods converge. A large part of the results
presented in this monograph can be found in various forms in the literature. We
tried, however, to show that the convergence of a big class of iteration methods
follows from general properties of some classes of operators and from some
general convergence theorems, in particular from Opial’s theorem or from its
modifications. This theorem was presented in 1967 by a Polish mathematician,
Zdzisław Opial (1930–1974), in a paper “Weak convergence of the sequence of
successive approximations for nonexpansive mappings” published in the Bulletin of
the American Mathematical Society.
Preface xi
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
1.1.1 Hilbert Space .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
1.1.2 Notations and Basic Facts . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5
1.2 Metric Projection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 16
1.2.1 Existence and Uniqueness of the Metric Projection .. . . . . . . 17
1.2.2 Characterization of the Metric Projection.. . . . . . . . . . . . . . . . . . 19
1.2.3 First Applications of the Characterization Theorem . . . . . . . 20
1.3 Convex Optimization Problems .. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 23
1.3.1 Convex Minimization Problems .. . . . . . . .. . . . . . . . . . . . . . . . . . . . 24
1.3.2 Variational Inequality . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 27
1.3.3 Common Fixed Point Problem . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 27
1.3.4 Convex Feasibility Problem . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 27
1.3.5 Linear Feasibility Problem . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 30
1.3.6 General Convex Feasibility Problem .. . .. . . . . . . . . . . . . . . . . . . . 33
1.3.7 Split Feasibility Problem . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 34
1.3.8 Linear Split Feasibility Problem . . . . . . . .. . . . . . . . . . . . . . . . . . . . 35
1.3.9 Multiple-Sets Split Feasibility Problem.. . . . . . . . . . . . . . . . . . . . 35
1.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 36
2 Algorithmic Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 39
2.1 Basic Definitions and Properties . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 40
2.1.1 Nonexpansive Operators .. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 41
2.1.2 Quasi-nonexpansive Operators . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 45
2.1.3 Cutters and Strongly Quasi-nonexpansive
Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 53
2.2 Firmly Nonexpansive Operators . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 65
2.2.1 Basic Properties of Firmly Nonexpansive Operators .. . . . . . 66
2.2.2 Relationships Between Firmly Nonexpansive
and Nonexpansive Operators . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 70
2.2.3 Further Properties of the Metric Projection .. . . . . . . . . . . . . . . . 76
xiii
xiv Contents
4.4.2
Properties of the Simultaneous Projection .. . . . . . . . . . . . . . . . . 165
4.4.3
Simultaneous Projection for a System of Linear
Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 168
4.4.4 Simultaneous Projection for the Linear
Feasibility Problem . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 169
4.5 Cyclic Projection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 171
4.5.1 Cyclic Relaxed Projection . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 172
4.5.2 Cyclic-Simultaneous Projection .. . . . . . . .. . . . . . . . . . . . . . . . . . . . 173
4.5.3 Projections with Reflection onto an Obtuse Cone .. . . . . . . . . 174
4.5.4 Cyclic Cutter . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 176
4.6 Landweber Operator .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 176
4.6.1 Main Properties . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 177
4.6.2 Landweber Operator for Linear Systems . . . . . . . . . . . . . . . . . . . 178
4.6.3 Extrapolated Landweber Operator for a System
of Linear Equations . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 181
4.7 Projected Landweber Operator.. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 184
4.8 Simultaneous Cutter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 185
4.9 Extrapolated Simultaneous Cutter . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 187
4.9.1 Properties of the Extrapolated Simultaneous
Cutter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 187
4.9.2 Extrapolated Simultaneous Projection . .. . . . . . . . . . . . . . . . . . . . 189
4.9.3 Extrapolated Simultaneous Projection for LFP .. . . . . . . . . . . . 190
4.9.4 Surrogate Projection . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 191
4.9.5 Surrogate Projection with Residual Selection . . . . . . . . . . . . . . 196
4.9.6 Extrapolated Simultaneous Subgradient Projection . . . . . . . . 198
4.10 Extrapolated Cyclic Cutter . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 199
4.10.1 Useful Inequalities . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 200
4.10.2 Properties of the Extrapolated Cyclic Cutter . . . . . . . . . . . . . . . 201
4.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 202
5 Projection Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 203
5.1 Alternating Projection Methods .. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 204
5.1.1 General Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 204
5.1.2 Alternating Projection Method for Closed
Linear Subspaces.. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 206
5.2 Extrapolated Alternating Projection Methods .. . .. . . . . . . . . . . . . . . . . . . . 208
5.2.1 Acceleration Techniques for Consistent Problems . . . . . . . . . 209
5.2.2 Acceleration Techniques for Inconsistent Problems . . . . . . . 210
5.2.3 Douglas–Rachford Algorithm .. . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 212
5.3 Projected Gradient Method.. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 213
5.4 Simultaneous Projection Method . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 215
5.4.1 Convergence of the SPM . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 215
5.4.2 Projected Simultaneous Projection Methods . . . . . . . . . . . . . . . 217
xvi Contents
References .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 275
Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 295
Chapter 1
Introduction
1.1 Background
In this section we present the notation, definitions and basic facts of convex subsets
and convex functions defined on a Hilbert space, convergence and differentiation
properties, properties of matrices, etc., which will be used in further parts of the
book.
pH with an inner product h; i, which is complete with respect to the
A linear space
norm kk WD h; i induced by this inner product is called a Hilbert space.
In what follows, we consider a real Hilbert space H, and denote by X a nonempty
closed convex subset of H and by I a finite subset f1; 2; : : : mg N, where m 2 N.
Furthermore,
(i) hx; yi D kxk kyk if and only if x and y are positive linearly dependent vectors,
i.e., ˛x D ˇy for some ˛; ˇ 0,
(ii) hx; yi D kxk kyk if and only if x and y are negative linearly dependent
vectors, i.e., ˛x D ˇy for some ˛; ˇ 0.
Moreover, one can take ˛ D kyk and ˇ D kxk in (i) and in (ii).
Triangle inequality. The following inequality holds for all x; y 2 H:
Furthermore,
(i) kx C yk D kxk C kyk if and only if x and y are positive linearly dependent,
(ii) kx C yk D j kxk kyk j if and only if x and y are negative linearly dependent.
Property (i) is known as the strict convexity of the norm and can also be
formulated equivalently in the form:
(iii) For any 2 .0; 1/ there holds
If, moreover, !i > 0, i D 1; 2; : : : ; m, then the equality holds in (1.7) if and only if
all xi , i D 1; 2; : : : ; m, are pairwise positive linearly dependent.
Let w D .!1 ; : : : ; !m / 2 Rm . The following equalities hold for all xi 2 H,
i D 1; 2; : : : ; m;
m 2
X X
m X
m X
m
!i xi D !i !j hxi ; xj i D !i2 kxi k2 C 2!i !j hxi ; xj i.
i D1 i;j D1 i D1 i;j D1;i ¤j
Theorem 1.1.1. A Banach space with a norm kk is a Hilbert space with an inner
product defined by the equality
1.1 Background 3
1
hx; yi D .kxk2 C kyk2 kx yk2 /,
2
if and only if the parallelogram law holds.
X
n
hx; yi WD x > y D j j ,
j D1
1
where x D .1 ; : : : ; n /, y D .1 ; : : : ; n /, and with the norm kxk WD .x > x/ 2 .
Example 1.1.3. The space Rm with the inner product induced by a positive definite
matrix G D Œgij mm , defined by
X
m X
m
hw; viG WD w> Gv D gij !i j , (1.9)
i D1 j D1
1
where w D .!1 ; : : : ; !m /, v D .1 ; : : : ; m /, and with the norm kukG WD .u> Gu/ 2 .
Example 1.1.4. The space l2 of real sequences x D .1 ; 2 ; : : :/ which are square-
summable, i.e.,
1
X
l2 WD fx W k 2 R, k D 1; 2; : : : and k2 < 1g,
kD1
P1 P1 2 12
with the inner product hx; yi WD kD1 k k , and with the norm kxk WD. kD1 k / .
Example 1.1.5. L2 .Œa; b/, where a; b 2 Œ1; C1, a < b—the space of
(equivalence classes of) Lebesgue measurable functions f W R ! R which are
square-integrable, i.e.,
Z b
L2 .a; b/ WD ff W f 2 .x/dx < 1g
a
4 1 Introduction
Example 1.1.6. Let Hi be a Hilbert space with an inner product h; iHi and with
the norm kkHi induced by this inner product, i 2 I WD f1; 2; : : : ; mg. Let H D
H1 : : : Hm . The function h; i W H H ! R defined by
X
m
hx; yiw WD !i hxi ; yi iHi , (1.10)
i D1
X
m
1
kxkw WD . !i kxi k2Hi / 2
i D1
is the norm in H induced by this inner product. The space H with the inner product
h; iw defined by (1.10) is a Hilbert space and is called a product Hilbert space.
B.x; / WD fy 2 H W ky xk g
denotes a ball with a centre x 2 H and radius > 0. C 0 denotes the complement of
a subset C H, i.e., C 0 D HnC . The subset
6 1 Introduction
C C D WD fz 2 H W z D x C y; x 2 C; y 2 Dg
C C a WD C C fag D fx C a W x 2 C g,
where C H and a 2 H. The subsets int C , cl C and bd C denote the interior, the
closure and the boundary of a subset C H, respectively. The subset
S.x; / WD bd B.x; / D fy 2 H W ky xk D g
denotes a sphere with a centre x 2 H and radius > 0. Lin S and Lin a denote
a linear subspace generated by the subset S H and by the vector a 2 H,
respectively. The subset
HC .a; ˇ/ WD fx 2 H W ha; xi ˇg
and
H .a; ˇ/ WD fx 2 H W ha; xi ˇg.
A subset K H is called an affine subspace if for all x; y 2 K and for any 2 R
we have .1 /x C y 2 K. The number
ha; bi
^.a; b/ WD arccos
kak kbk
1.1.2.3 Functions
S.f; ˛/ D fx 2 X W f .x/ ˛g
o.t/
where limt !0 D 0. The element y is called a derivative (or differential) of f
t
at x and is denoted by f 0 .x/.
8 1 Introduction
f 0 .x; s/ D hg; si
1.1.2.4 Operators
kAxkH2
kAk WD sup ,
xW0<kxk1 kxkH1
1.1 Background 9
where kkH1 and kkH2 denote the norms in H1 and H2 , respectively, is called a
norm of A in L.H1 ; H2 /.
Definition 1.1.17. Let A W H1 ! H2 be a bounded linear operator. An operator
A W H2 ! H1 with the property
for all x 2 H1 and y 2 H2 , where h; iH1 and h; iH2 are inner products in H1 and
H2 , respectively, is called an adjoint operator.
For every A an adjoint operator A exists and is uniquely determined. Futher-
more, A is a bounded linear operator.
Definition 1.1.18. We say that a bounded linear operator A W H ! H is unitary if
A A D Id holds.
Definition 1.1.19. We say that a bounded linear operator A W H ! H is self-
adjoint, if A D A.
Definition 1.1.20. We say that a self-adjoint operator A W H ! H is nonnegative
if hAx; xi m kxk2 for some constant m 0 and for all x 2 H. If m > 0, then we
say that A is positive (or elliptic).
Theorem 1.1.21. For any bounded linear operator A W H1 ! H2 it holds
hT x T y; x yi kx yk2
for a constant 0 and for all x; y 2 X . If > 0, then we say that T is strongly
monotone or -strongly monotone.
Definition 1.1.29. We say that an operator T W H ! H is Lipschitz continuous if
kT x T yk
kx yk .
T .x C h/ D T .x/ C Gh C o.khk/.
T .C / WD fT x W x 2 C g Y
1.1 Background 11
T 1 .D/ WD fx 2 X W T x 2 Dg
Im A WD fy 2 H2 W y D Ax for some x 2 H1 g
.ker A/? D Im A ,
consequently,
H1 D ker A ˚ Im A .
1.1.2.5 Matrices
2 3
1 0 ::: 0
6 0 2 ::: 0 7
diag v D 6
4:::
7.
::: ::: :::5
0 0 ::: m
Theorem 1.1.34. For a symmetric matrix A the following conditions are equiva-
lent:
(i) A is positive definite (positive semi-definite),
(ii) All eigenvalues of A are positive (nonnegative),
(iii) There exists an orthogonal matrix U and a diagonal matrix D with positive
(nonnegative) elements on the diagonal such that A D U > DU ,
(iv) There exists a positive definite (positive semi-definite) matrix B such that
A D B 2.
1
We denote by A 2 the matrix B with property (iv) above.
Definition 1.1.35. Let A be an m n matrix. The matrix A> A is called the Gram
matrix of the columns of A.
The Gram matrix A> A is positive semi-definite. Furthermore, it is positive
definite if and only if A has full column rank, i.e., the columns of A are linearly
independent.
Definition 1.1.36. The Moore–Penrose pseudoinverse AC of an m n matrix A is
a uniquely determined n m matrix satisfying the following conditions:
or
AC D A> .AA> /1 ,
respectively.
xk x
TC .x/ WD s 2 H W 9fx k g1
kD0 C;x !x
k 9ftk g1
kD0 ;tk #0
with s D lim
k tk
holds for all x; y 2 X and for all 2 Œ0; 1. If the inequality is strict for all x; y 2 X ,
x ¤ y and for all 2 .0; 1/, then we say that f is strictly convex. We say that f is
˛-strongly convex, where ˛ > 0 or, shortly, strongly convex if
1
f ..1 /x C y/ .1 /f .x/ C f .y/ ˛.1 / kx yk2
2
for all x; y 2 X and for all 2 Œ0; 1. A function f W X ! R is called concave if
f is convex.
f 0 .x; s/
ksk .
All the facts presented in this section can be found in the following references
[33, 38, 40, 61, 133, 158, 180, 181, 209, 210, 214, 234, 262–264, 267, 298, 300, 317,
325].
In this section we define the metric projection in a Hilbert space H and present its
basic properties. This operator plays an important role in further parts of the book.
Other properties of the metric projection will be presented in Chap. 2.
Definition 1.2.1. Let C H be a nonempty subset and x 2 H. If there exists a
point y 2 C such that
1.2 Metric Projection 17
C
z
y = PC x
ky xk kz xk
for any z 2 C , then y is called a metric projection of x onto C and is denoted
by PC x (Fig. 1.1). The vector s D PC x x is called a projection vector of x
onto C . If PC x exists and is uniquely determined for all x 2 H, then the operator
PC W H ! C is called the metric projection (onto C ).
Remark 1.2.2. Let C H. If x 2 C , then it follows from the definition of the
metric projection that PC x D x. If x … C and there exists a metric projection PC x,
then PC x 2 bd C . Indeed, if PC x 2 int C , then B.PC x; "/ C for a sufficiently
small " > 0. Then
"
z WD x C .1 /.PC x x/ 2 B.PC x; "/ C ,
kPC x xk
Definition 1.2.1 does not yield the existence and the uniqueness of a metric projec-
tion PC x for a point x 2 H and for a subset C H. If C H is compact, then the
existence of a metric projection follows from the Weierstrass theorem but it needs
not to be defined uniquely. For example, both points a; b 2 H are metric projections
of the midpoint x WD aCb 2 of the segment Œa; b onto the subset C D fa; bg.
Therefore, the convexity of C seems to be a natural assumption for the uniqueness of
the metric projection PC x for any x 2 H. Furthermore, it follows from the definition
of a metric projection and from the continuity of the norm that PC x (if it exists)
belongs to the closure of C . Actually, PC x 2 bd C for x … C (see Remark 1.2.2).
Therefore, the closedness of C is a natural assumption for the existence of the
metric projection PC x for any x 2 H. It turns out that in a Hilbert space these two
assumptions (the closedness and the convexity of C ) are sufficient for the existence
18 1 Introduction
and the uniqueness of the metric projection PC x for all x 2 C . The following
theorem can be found, e.g., in [61, Sect. 1.2.2, Theorem 1.7], [140, Theorem 3.4(2)],
[173, Theorem 7.43], [209, Chap. III, Sect. 3.1] or [267, Theorem 8.25].
Theorem 1.2.3. Let C H be a nonempty closed convex subset. Then for any
x 2 H there exists a metric projection PC x and it is uniquely determined.
Proof. First we prove the assertion for x D 0. Let d WD inffkyk W y 2 C g and
a sequence fy k g1 y k ! d . We split the proof onto three
kD0 C be such that
parts.
(a) We show that fy k g1kD0 is a Cauchy sequence. Let " > 0 and k0 0 be such
k 2
d C "=4 for k k0 . Let k; l k0 . Since C is convex, we have
2
that y
1 k l
1 k
2 y C 1 l
2 y 2 C . Therefore, 2 y C y d . The parallelogram law yields
now k
y y l 2 D 2 y k 2 C 2 y l 2 y k C y l 2
i.e., y D y 0 .
We have proved that PC 0 exists and is uniquely defined. Of course, C x is
closed convex, because C is closed convex. Therefore, PC x 0 also exists and is
uniquely defined. It follows easily from the definition of the metric projection that
PC x D x C PC x 0,
The theorem below gives a criterion for y 2 C to be the metric projection of a point
x 2 H onto a convex set C H and is often used in applications. The theorem
can be found, e.g., in [140, Theorem 4.1], [173, Theorem 7.45], [209, Chap. III,
Theorem 3.1.1] and in [185, Proposition 3.5] which contains several other equivalent
conditions.
Theorem 1.2.4 (Characterization Theorem). Let x 2 H, C H be a convex
subset and y 2 C . The following conditions are equivalent:
(i) y D PC x,
(ii) hx y; z yi 0 for all z 2 C .
Proof. (i) )(ii). Let y D PC .x/, z 2 C and
z D y C .z y/
kz xk2 D kz y C y xk2
D kz yk2 C ky xk2 C 2hz y; y xi
ky xk2 ,
x−u
PC (x − u)
−u
u C+u
PC+ u x
x
It follows from Lemma 1.2.5 that the inequality in condition (ii) of the Char-
acterization Theorem 1.2.4 can be replaced by one of the conditions (a)–(d) of
Lemma 1.2.5.
In this section we give some useful facts which follow from the Characterization
Theorem 1.2.4.
Lemma 1.2.6. Let x 2 H, C H be a convex subset and u 2 H. If a metric
projection PC .x u/ exists, then a metric projection PC Cu x also exists and
PC Cu x D PC .x u/ C u (1.12)
A∗ P C (A ∗ u) C
A A∗ (u)
0 hA u y; z yi
D hA u A Ay; z yi
D hu v; Az Ayi
D hu v; w vi.
PC x D .PC1 x1 ; : : : ; PCm xm /:
x
z
NC (y) C
y = PC x
Proof. By the definition of the normal cone, we can write the second part of the
right hand side of (ii) in the form hx y; z yi 0 for any z 2 C (Fig. 1.6).
The lemma follows now from the characterization of the metric projection (see
Theorem 1.2.4). t
u
For T W H ! H and 2 R denote T WD Id C.T Id/.
Corollary 1.2.10. Let T W H ! H, C H be a nonempty closed convex subset.
Then Fix.PC T / D Fix.PC T / for any > 0.
Proof. Let > 0. It follows from Lemma 1.2.9 that
minimize f .x/
(1.13)
with respect to x 2 C
for a sufficiently small > 0. The contradiction shows that f .x/ f .x / for all
x 2 C , i.e., x is a global minimizer of f on C . Suppose now that f is strictly
convex and that f .x 0 / D f .x / holds for some x 0 ¤ x . Then for x D 12 x C 12 x 0 ,
we obtain
1 1
f .x/ < f .x / C f .x 0 / D f .x / f .x/,
2 2
a contradiction which shows that the minimizer is uniquely determined. Now
suppose that f is continuous and strongly convex and that C is closed. Recall
that a strongly convex function is coercive. Therefore, Corollary 1.1.53 yields
Argminx2C f .x/ ¤ ;. t
u
The subset C in (1.13) is often given in the form
C WD fx 2 H W ci .x/ 0; i 2 I g,
minimize f .x/
subject to ci .x/ 0; i 2 I , (1.14)
x 2 H.
The functions ci are called the constraints of problem (1.14). If the functions
f; ci ; i 2 I , are convex, then (1.14) is called a convex minimization problem (CMP).
If f; ci ; i 2 I , are differentiable, then (1.14) is called a constrained differentiable
minimization problem. In order to solve problem (1.13) or (1.14) one can apply
suitable methods (algorithms) which generate sequences approximating a solution
x of the problem. These algorithms should be constructed in such a way that the
sequences converge (in some sense) to a solution of the problem. Since an algorithm
should terminate after a finite number of iterations, we cannot expect in general that
the algorithm gives an exact solution. Let " 0. We say that x" 2 C is an "-optimal
solution of (1.13) or (1.14) if f .x" / f .x/ C " for all x 2 C .
If f is a convex function and C is a closed convex subset (or ci ; i 2 I , are
convex functions), then problem (1.13) (or (1.14)) is called a convex (constrained)
minimization problem.
For a constrained differentiable minimization problem (1.14) with H D Rn
(equipped with the standard inner product) a point .x ; y / 2 Rn Rm satisfying
the following system of equalities and inequalities
for all y 2 C .
Proof. See [209, Chap. VII, Theorem 1.1.1] for finite dimensional case or [325,
Sect. 7.1] for an infinite dimensional one. t
u
Corollary 1.3.5. Let C H be a closed convex subset and f W H ! R be a
differentiable convex function. Then for any > 0 it holds
Proof. Let > 0. It follows from Theorem 1.3.4 and from Lemma 1.2.9 that
x 2 Argmin f .x/
x2C
hF x; y xi 0 (1.17)
holds for all y 2 C . Note that, by Theorem 1.3.4, the differentiable convex
minimization problem (1.13) is a special case of the variational inequality. The proof
of the following theorem is similar to the proof of Corollary 1.3.5 and is left to the
reader.
Theorem 1.3.8. Let > 0. A point x 2 C is a solution of variational inequality
(1.17) if and only if x 2 Fix PC .Id F /.
For a general discussion on the existence of solutions of variational inequality
problems we refer the reader to [165, 226, 358]. Below we only recall sufficient
conditions for the existence and uniqueness of a solution of VIP.F ; C /.
Theorem 1.3.9. If F W H ! H is a strongly monotone and Lipschitz continuous
operator, then the variational inequality (1.17) has a unique solution x 2 C .
Proof. See, e.g., [358, Theorem 46.C]. t
u
is to find a point x 2 C . We see that the CFP can be formulated as the CFPP with
Ui WD PCi , i 2 I . On the other hand, the CFPP is a CFP with Ci D Fix Ui . In
applications concerning the CFP the subsets Ci are often called the constraints sets
or, shortly, constraints and are given in the form
where ci W H ! R are convex functions, called the constraints functions or, shortly,
constraints, i 2 I .
The CFP can also be considered without the assumption C ¤ ;. In this case
we should define an appropriate convex proximity function f W H ! RC which
measures a “distance” to the constraints. The proximity function should have the
property
f .x/ D 0 if and only if x 2 C . (1.19)
The convex feasibility problem can be formulated as a minimization of the
proximity function f . The proximity function f W H ! RC for the CFP can be
defined by the following general form
1X
f .x/ D !i .ciC .x//2 ,
2 i 2I
P
where w D .!1 ; : : : ; !m / 2 RmCC . One can additionally suppose that i 2I !i D 1
which leads to the assumption that w 2 ri m . This assumption does not change
the minimization problem. Both proximity functions are convex and the other one
is even differentiable.
As mentioned above, the CFP is, in general, to minimize a convex proximity
function f . Denote f WD infx2H f .x/. It is clear that f 0. We emphasize that
the proximity function f needs not to attain its minimum. If C ¤ ;, we have C D
Argminx2H f .x/ for any proximity function with the above described properties.
In this case, the problem of determining x 2 C and the problem of minimization
of the proximity function f are equivalent. Furthermore, in this case, a significant
information about f is available, namely f WD minx2H f .x/ D 0. One can say
that in the general case, i.e., without the assumption C ¤ ;, the CFP is to find an
element x 2 H, for which “distances” to the constraints Ci , i 2 I , are minimal in
some sense, if such x exists. For x 2 H, the value f .x/ expresses this “distance”.
If C D ; and the proximity function attains its minimum, then, of course, f > 0.
Note, however, that the nonemptiness of C does not follow from the fact that f WD
infx2H f .x/ D 0.
In some applications, the CFP is to find an element x 2 C or to state that C D ;.
We can also say that the CFP is to find a zero of the proximity function or to state
that the proximity function is positive for all arguments.
30 1 Introduction
Ci WD H .ai ; ˇi / D fx 2 H W hai ; xi ˇi g
ai> x ˇi , i D 1; 2; : : : ; m, (1.26)
Ax b,
or by
.ai> x ˇi /C
fi .x/ WD hi (1.28)
kai k
for some convex increasing functions hi W RC ! RC with hi .0/ D 0, i 2 I .
In most cases the functions hi are given by hi .u/ WD !i u or hi .u/ WD 12 !i u2 for
!i > 0, i 2 I . If we apply these definitions of fi and hi to formulas (1.20)–(1.25)
we obtain the following collection of proximity functions:
(a)
f .x/ WD max.ai> x ˇi /C , (1.29)
i 2I
in order to determine a point for which the value of the most violated constraint
is minimal,
(b)
.ai> x ˇi /C
f .x/ WD max , (1.30)
i 2I kai k
1.3 Convex Optimization Problems 31
in order to determine a point for which the distance to the furthest constraint is
minimal,
(c) X
f .x/ WD !i .ai> x ˇi /C , (1.31)
i 2I
in order to determine a point for which the weighted sum of the values of the
constraints is minimal,
(d)
X .ai> x ˇi /C
f .x/ WD !i , (1.32)
i 2I
kai k
in order to determine a point for which the weighted sum of distances to the
constraints is minimal,
(e)
1X
f .x/ WD !i Œ.ai> x ˇi /C 2 , (1.33)
2 i 2I
in order to determine a point for which the weighted sum of squared values of
the constraints is minimal, or
(f)
> 2
1X .ai x ˇi /C
f .x/ WD !i , (1.34)
2 i 2I kai k
in order to determine a point for which the weighted sum of squared distances
to the constraints is minimal.
All of the proximity functions presented herein are convex. The last two functions
are even differentiable. If we denote the residuum of the i th constraint at a point
x by i .x/, i.e., i .x/ WD ai> x ˇi , i 2 I , and by r.x/ the residual vector,
i.e., r.x/ WD . 1 .x/; : : : ; m .x//, then the proximity function (1.31) has the form
f .x/ D w> rC .x/, where w D .!1 ; : : : ; !m / 2 Rm CC , and the proximity function
2
(1.33) has the form f .x/ D 2 krC .x/kW for W WD diag w. A system of linear
1
˛i ai> x ˛i ˇi ; i 2 I; (1.36)
is, in general, not equivalent to the original problem, even if we consider the
same proximity function as for the original problem. However, there are some
relationships between these two problems. The problems are equivalent if we use
the proximity functions (1.30), (1.32) or (1.34). If we apply the proximity functions
(1.31) or (1.33) with weights !i D ˛i1 i or !i D ˛i2 i , respectively, i 2 I , to
problem (1.36), then we obtain a new LFP which is equivalent to the original one
with the proximity function (1.31) or (1.33), respectively, and with weights !i D i ,
i 2 I . In particular, if we take ˛i D kai k1 , i 2 I , in (1.36), then the half-spaces in
the new LFP are defined by normalized vectors. Therefore, if we apply the proximity
function (1.33) with weights !i D kai k2 i , i 2 I , to the new LFP, then the problem
is equivalent to the original one with the proximity function (1.33) and with weights
!i D i , i 2 I . Note that we can suppose, without loss of generality, that the
vectors ai , i 2 I , are normalized. Otherwise, an appropriate change of the weights
leads to an equivalent problem, because the weights !i , i 2 I , can be multiplied
by a constant ˛ > 0 without changing the minimizers of the proximity function.
Therefore, we can suppose that w 2 ri m instead of w 2 Rm CC .
is called a linear least squares problem (LLSP). This problem is equivalent to the
following compatible system of linear equations
minimize 12 kxk2
subject to A> .Ax b/ D 0 (1.39)
x 2 Rn .
It follows from the definition of the metric projection that a solution x of (1.39) is
equal to PL 0, where L WD fx 2 Rn W A> Ax D A> bg. The uniqueness of the metric
projection yields the uniqueness of x . The Lagrange function L W Rn Rn ! R
has the form
1
L.x; y/ D kxk2 C y > A> .Ax b/
2
and the KKT-system has the form
rx L.x; y/ D x C A> Ay D 0
(1.40)
ry L.x; y/ D A> .Ax b/ D 0.
In many applications of the convex feasibility problems one seeks a point, for
which some constraints are satisfied and a proximity function defined for the rest
of constraints is minimal. The constraints which should be satisfied are called hard
constraints and the rest of constraints are called soft constraints. We call the problem
a general convex feasibility problem (GCFP). In other words, the GCFP is to find
a point satisfying the hard constraints for which “distances” to soft constraints are
minimal in some sense. Such model was described in details in [123].
Let fCi gi 2I be a family of closed convex subsets. Let Ci for i 2 Ih be the
hard constraints and Ci for i 2 Is beTthe soft constraints,Twhere I D Ih [ Is and
Ih \ Is D ;. Furthermore, let Ch WD i 2Ih Ci and Cs WD i 2Is Ci . If Ch \ Cs ¤ ;,
then GCFP is to find x 2 Ch \Cs . If we do not suppose Ch \Cs ¤ ;, then, similarly
as for the CFP, one defines a convex proximity function f W Ch ! RC , defined on
the subset Ch of hard constraints, which measures “distances” to the soft constraints.
Similarly as for the CFP, we require the proximity function f to have the property
34 1 Introduction
A−1 (Q)
Ax
x
C Q
In some applications the general convex feasibility problem has the form: Given
closed convex subsets C H1 , Q H2 of Hilbert spaces H1 ; H2 and a bounded
linear operator A W H1 ! H2 , find a point x 2 C such that Ax 2 Q (if such a point
exists) (see Fig. 1.7). If H1 ; H2 are Euclidean spaces, then the above problem has
the form: Given closed convex subsets C Rn , Q Rm and a matrix A of type
m n, find a point x 2 C such that Ax 2 Q (if such a point exists).
The subset C is a hard constraint and the subset D WD A1 .Q/ D fx 2 H1 W
Ax 2 Qg is a soft constraint. It is clear that D is closed convex, because for a
bounded linear operator A, the inverse image of a closed convex subset Q is closed
and convex. This problem was introduced by Censor and Elfving [88] and was called
a split feasibility problem (SFP). If we do not suppose that C \ D ¤ ;, then,
similarly as for the GCFP, we introduce a convex proximity function f W C ! RC
and we present the SFP as minimization of f . Byrne [55, Sect. 2] has proposed a
proximity function defined by
1
PQ .Ax/ Ax 2 .
f .x/ WD (1.41)
2
x 2 C \ D , f .x/ D 0.
Now consider the SFP in the finite dimensional case. If the subset Q Rm has the
form
Q WD fy 2 Rm W y bg, (1.42)
where b 2 Rm , then the problem is called a linear split feasibility problem (LSFP).
In other words, the problem can be written as follows: find x 2 C such that Ax b
(if such a point exists). In general, we do not suppose, however, that C \ fx 2 Rn W
Ax bg ¤ ;. Now we present a few examples of a proximity function for the LSFP
which are often used in applications. In Sect. 4 we will show that for the subset Q
defined by (1.42) we have
PQ .Ax/ Ax D b Ax.
1X >
m
1
f .x/ WD krC .x/k2 D Œ.a x ˇi /C 2 .
2 2 i D1 i
Note that f is differentiable. Alternatively, one can also apply a proximity function
given by
Xm
f .x/ WD w> rC .x/ D !i .ai> x ˇi /C
i D1
or by
1X
m
f .x/ WD krC .x/k2W D !i Œ.ai> x ˇi /C 2 , (1.43)
2 i D1
The following problem, called a multiple-sets split feasibility problem, has been
studied in the literature (see, e.g., [93, 103, 257, 344]): Given closed convex subsets
Ci H1 , i D 1; 2; : : : ; p, Qj H2 , j D 1; 2; : : : ; r, of Hilbert spaces H1 ; H2 ,
respectively, and a bounded linear
T p Tr operator A W H1 ! H2 , find a point x 2 C WD
C
i D1 i such that Ax 2 Q WD j D1 Qj .
This problem can be considered as a split feasibility problem with the proximity
function given by (1.41). The application of this proximity function requires,
however, a simple structure of C and Q, which allows an easy computation of PC x
36 1 Introduction
1X 1X
p r
2
f .x/ WD kPCi x xk2 C ˇj PQj Ax Ax .
2 i D1 2 j D1
Note that in this case the problem is a special case of the general convex feasibility
problem, where all subsets Ci , i D 1; 2; : : : ; p, and Qj , j D 1; 2; : : : ; r, are treated
as soft constraints. If Ch has a simple structure allowing an easy computation of
PCh x for any x 2 Rn , one can also use the proximity function f W Ch ! R
given by
1X 2
r
f .x/ WD ˇj PQj Ax Ax .
2 j D1
In this case the problem is a special case of the general convex feasibility problem,
where Ci , i D 1; 2; : : : ; p, are hard constraints and Qj , j D 1; 2; : : : ; r, are soft
constraints.
1.4 Exercises
Exercise 1.4.1. Prove that in a Hilbert space the Cauchy–Schwarz inequality and
the triangle inequality are equivalent.
Exercise 1.4.2. Let w D .!1 ; !2 ; : : : ; !m / 2 m . Prove that
m 2
X X m
1 X
m
2
!i xi D !i kxi k2 !i !j xi xj . (1.44)
2 i;j D1
i D1 i D1
NC x D conefai W i 2 I.x/g,
x kC1 D Tk x k (2.1)
Let H be a Hilbert space. In what follows, we consider operators which are defined
on a nonempty closed convex subset X H.
P
Remark 2.1.1. Let Ui W X ! X , i 2 I WD f1; 2; : : : ; mg. If (i) U D i 2I !i Ui ,
where w D .!1 ; !2 ; : : : ; !m / 2 m , or (ii) U D Um Um1 : : : U1 , then the following
obvious inclusion holds \
Fix Ui Fix U
i 2I
The converse inclusion needs not to be true even if all Ui , i 2 I , have a common
fixed point (see Example 2.1.27).
Definition 2.1.2. Let T W X ! H and 2 Œ0; 2. The operator T W X ! H
defined by
T WD .1 / Id CT
is called a -relaxation or, shortly, relaxation of the operator T . Obviously, T D
Id C.T Id/. We call a relaxation parameter. If 2 .0; 1/, then T is called an
under-relaxation of T . If 2 .1; 2/, then T is called an over-relaxation of T and
if D 2, then T is called the reflection of T . If 2 .0; 2/, then T is called a strict
relaxation of T .
A relaxation T of an operator T can be defined for any 2 R. However, if we
do not extend explicitly the range of , we assume that 2 Œ0; 2.
Remark 2.1.3. Note that the equality .T / D T holds for all ; 2 R, con-
sequently .T /1 D T for ¤ 0.
Remark 2.1.4. It is clear that Fix T D Fix T whenever ¤ 0.
Let Ui W X ! X , i 2 I WD f1; 2; : : : ; mg, U WD Um Um1 : : : U1 and Qi WD
Ui Ui 1 : : : U1 Um : : : Ui C1 , i D 1; 2; : : : ; m. Denote Q0 WD Qm D U and U0 WD
Um . There exists a relationship among the subsets of fixed points of operators Qi ,
which is expressed by the following theorem.
Theorem 2.1.5. For i D 1; 2; : : : ; m there holds
zi D Ui zi 1 D Ui Qi 1 zi 1 D Ui Ui 1 : : : U1 Um : : : Ui C1 Ui zi 1
D Ui Ui 1 : : : U1 Um : : : Ui C1 zi D Qi zi ,
2.1 Basic Definitions and Properties 41
kT x T yk kx yk
for all x; y 2 X ,
(ii) Strictly nonexpansive if
kT x T yk < kx yk or x y D T x T y
for all x; y 2 X ,
(iii) An ˛-contraction, where ˛ 2 .0; 1/ or, shortly, a contraction if
kT x T yk ˛ kx yk
for all x; y 2 X .
The theorem below, called the Banach fixed point theorem or the Banach theorem
on contractions, is widely applied in various areas of mathematics. The theorem
holds for any complete metric space, and hence, in particular, for every closed subset
of a Hilbert space.
Theorem 2.1.7 (Banach, 1922). Let X be a complete metric space and T W X !
X be a contraction. Then T has exactly one fixed point x 2 X . Furthermore,
for any x 2 X , the orbit fT k xg1
kD0 converges to x with a rate of geometric
progression.
Proof. See, e.g., original paper of Banach [15], [185, Theorem 1.1], [267, Theorem
24.2], [184, Theorem 2.1], [183, Theorem 2.1] or [36, Theorem 2.1]. t
u
42 2 Algorithmic Operators
The Banach fixed point theorem is a widely applied tool for an iterative approxi-
mation of fixed points. Unfortunately, its application is restricted to contractions. We
will need, however, appropriate tools for an iterative approximation of fixed points
of nonexpansive operators T with Fix T ¤ ;.
Below, we present several classical fixed points theorems.
Theorem 2.1.8 (Brouwer, 1912). Let X Rn be nonempty compact and convex
and T W X ! X be continuous. Then T has a fixed point.
Proof. See, e.g., original paper of Brouwer [43] or [191, Chap. II, 5, Theorem 7.2]
or [183, Theorem 7.6]. t
u
The Brouwer fixed point theorem was generalized by Juliusz Schauder.
Theorem 2.1.9 (Schauder, 1930). Let X be a nonempty compact and convex
subset of a Banach space and T W X ! X be continuous. Then T has a fixed point.
Proof. See, e.g., original paper of Schauder [302] or [191, Chap. II, 6, Theorem
3.2] or [183, Theorem 8.1]. t
u
For nonexpansive operators in a Hilbert space H the compactness of X H in the
Schauder theorem can be replaced by the boundedness of X . The following theorem
was proved independently by Browder [45, Theorem 1], Göhde [188] and by Kirk
[227]. The proof can also be found, e.g., in [185, Theorem 5.1], [191, Chap. I, 4,
Theorem 1.3], [183, Theorem 4.1] or [36, Theorem 3.1].
Theorem 2.1.10 (Browder–Göhde–Kirk, 1965). Let X be a nonempty closed,
convex and bounded subset of a uniformly convex Banach space (e.g., of a Hilbert
space H) and U W X ! X be nonexpansive. Then U has a fixed point.
Contrary to the Banach fixed point theorem, the theorems of Brouwer, Schauder
and of Browder–Göhde–Kirk are only of existential nature. In Chap. 3 we present
theorems which can be applied to iterative methods for determining fixed points of
nonexpansive operators.
Below, we present some properties of the subset of fixed points of a nonexpansive
operator. The following result can be found in [185, Proposition 5.3].
Proposition 2.1.11. The subset of fixed points of a nonexpansive operator T W
X ! H is closed and convex.
Proof. (cf. [185, Proposition 5.3]) Let x k 2 Fix T and x k ! x. We have x 2 X
because X is closed. By the continuity of T ,
x D lim x k D lim T x k D T x,
k
i.e., Fix T is a closed subset. Now we show the convexity of Fix T . Let x; y 2 Fix T ,
x ¤ y and z D .1 /x C y for 2 .0; 1/. By the nonexpansivity of T and by
the positive homogeneity of the norm we have
kx T zk D kT x T zk kx zk D kx yk (2.4)
2.1 Basic Definitions and Properties 43
and
kT z yk D kT z T yk kz yk D .1 / kx yk . (2.5)
Now, the triangle inequality yields
kx yk kx T zk C kT z yk
kx yk C .1 / kx yk
D kx yk .
Consequently,
kx yk D kx T zk C kT z yk .
By the strict convexity of the norm, the vectors x T z and T z y are positive
linearly dependent. Therefore, ˛.x T z/ C ˇ.y T z/ D 0 for some ˛; ˇ 0.
ˇ
Since x ¤ y, it follows that ˛ C ˇ > 0, and hence, T z D ˛Cˇ
˛
x C ˛Cˇ y. Now, the
nonexpansivity of T and inequalities (2.4) and (2.5) yield
ˇ
kx yk D kx T zk D kT x T zk kx zk D kx yk (2.6)
˛Cˇ
and
˛
kx yk D kT z yk D kT z T yk kz yk D .1 / kx yk . (2.7)
˛Cˇ
If at least one inequality in (2.6) and (2.7) is strict, then by summing up (2.6)
ˇ
and (2.7) we would obtain a contradiction. Therefore, ˛Cˇ D and ˛Cˇ
˛
D .1 /,
consequently T z D .1 /x C y D z. t
u
The closedness and convexity of the subset of fixed points of a nonexpansive
operator follows also from a property presented in Sect. 2.2 (see Corollary 2.2.48).
Lemma 2.1.12. Let Si W X ! X , i 2 I WD f1; 2; : : : ; mg, be nonexpansive.
Then:
P
(i) A convex combination S WD i 2I !i Si , where w D .!1 ; : : : ; !m / 2 m , is
nonexpansive. If, furthermore, at least one operator Si is a contraction and the
corresponding weight !i > 0, then S is a contraction;
(ii) A composition S WD Sm Sm1 : : : S1 is nonexpansive. If, furthermore, at least
one operator Si is a contraction, then S is a contraction.
Proof. Let x; y 2 X and Si be nonexpansive, i.e., kSi x Si yk ˛i kx yk,
where ˛i 2 .0; 1, i 2 I .
P P
(i) Let w 2 m , S WD i 2I !i Si and ˛ D j 2I !j ˛j . It is clear that ˛ 2 .0; 1.
By the convexity of the norm and the nonexpansivity of Si , i 2 I , we have
44 2 Algorithmic Operators
X
kS x Syk D !i .Si x Si y/
i 2I
X
!i kSi x Si yk
i 2I
X
!i ˛i kx yk
i 2I
X !i ˛i
D P ˛ kx yk
i 2I j 2I !j ˛j
D ˛ kx yk ,
T
Proof. The inclusion i 2I Fix Ui Fix U is always true (see RemarkT 2.1.1). Now
we show that the converse inclusion
T also holds. Let z 2 Fix U and u 2 i 2I Fix Ui .
If z D u, then, of course, z 2 i 2I Fix Ui . Otherwise, for z ¤ u, by the convexity
of the norm and by the nonexpansivity of Ui , i 2 I , we have
2.1 Basic Definitions and Properties 45
kz uk D kU z uk
X X
D !i Ui z u D !i .Ui z u/
i 2I i 2I
X X
!i kUi z uk D !i kUi z Ui uk
i 2I i 2I
X
!i kz uk D kz uk .
i 2I
Consequently,
X X X
!i .Ui z u/ D !i kUi z uk D !i kz uk . (2.8)
i 2I i 2I i 2I
Since !i > 0 for all i 2 I , the first equality in (2.8) yields a positive linear
dependence of all pairs of vectors Ui z u and Uj z u, i; j 2 I , i ¤ j , i.e.,
kUi z uk .Uj z u/ D Uj z u .Ui z u/. (2.9)
The second equality in (2.8), together with the inequality kUi z uk kz uk,
i 2 I , and the assumption !i > 0, i 2 I , yield
kUi z uk D kz uk (2.10)
T
for all i 2 I , i.e., z 2 i 2I Fix Ui . t
u
kT x zk kx zk
x z
x+ y
2
kT x zk < kx zk
is nonempty, then Fej T is the largest subset, with respect to which T is Fejér
monotone.
2.1 Basic Definitions and Properties 47
kT x zk kx zk
kT x zk < kx zk
T is sQNE H) T is C -sQNE
The metric projection onto a closed convex subset is a typical example of a strictly
quasi-nonexpansive operator.
A nonexpansive and strictly Fejér monotone operator is also called attracting
(see [22, Definition 2.1]). Yamada and Ogura use the name an attracting quasi-
nonexpansive operator for a strictly quasi-nonexpansive one (see [346, page
623]). Vasin and Ageev call these operators strongly Q-quasi-nonexpansive
48 2 Algorithmic Operators
(see [333, Definition 2.2]). Reich and Zaslavski define a more general operator
than the strictly quasi-nonexpansive one and call it an F -attracting mapping,
where F D Fix T (see [297, Sect. 1]). A continuous strictly quasi-nonexpansive
operator is also called a paracontraction (see, [164, Definition 1]). The class of
quasi-nonexpansive operators is denoted in [126, page 161] by F 0 . Properties
of quasi-nonexpansive operators in metric spaces have been intensively studied
since 1969 (see, e.g., [145, 148, 283], [50, Sect. 1], [113]), but the name quasi-
nonexpansive was introduced by Dotson [147].
Lemma 2.1.20. A nonexpansive operator U W X ! H with a fixed point is quasi-
nonexpansive.
Proof. Let U be nonexpansive and z 2 Fix U . Then
kUx zk D kUx U zk kx zk ,
i.e., U is quasi-nonexpansive. t
u
It is clear that the class of nonexpansive operators having a fixed point is an essential
subclass of quasi-nonexpansive operators, because a quasi-nonexpansive operator
needs not to be continuous. Moreover, a quasi-nonexpansive operator needs not to
be nonexpansive even if it is continuous (see Exercise 2.5.2). In this section we
present properties of the family of quasi-nonexpansive operators. In further parts of
the book we show that these operators play an important role in iterative methods
for fixed point problems.
The following lemma gives a relation between the subset Fej T and the subset
Fix T for an operator T W X ! H (cf. [24, Proposition 2.6 (ii)]).
Lemma 2.1.21. For any operator T W X ! H the inclusion Fej T Fix T holds.
If Fix T ¤ ; and T is quasi-nonexpansive, then the converse inclusion also holds.
Consequently, the subset of fixed points of a quasi-nonexpansive operator is closed
and convex.
Proof. If Fej T D ;, then the first part of the assertion is obvious. Now let
Fej T ¤ ; and w 2 Fej T . Then, for z D x D w in (2.13), we obtain
TwCw
0 hw ; T w wi
2
1
D kT w wk2 0,
2
i.e., T w D w. Therefore, Fej T Fix T . Now suppose that Fix T ¤ ; and that
T is quasi-nonexpansive, i.e., T is Fejér monotone with respect to Fix T . Then,
Lemma 2.1.17 yields the inclusion Fix T Fej T , which together with the first
part of the lemma gives Fix T D Fej T . The convexity and the closedness of Fix T
follows now from Lemma 2.1.17 and from the fact that the intersection of closed
half-spaces is closed and convex. t
u
2.1 Basic Definitions and Properties 49
Fix T
−h
Remark 2.1.22. It follows from Remark 2.1.18 (ii), Lemmas 2.1.17 and 2.1.21 that
a quasi-nonexpansive operator T W X ! X is Fejér monotone with respect to any
nonempty subset of Fix T . Therefore, we will restrict our further consideration of
Fejér monotone operators to quasi-nonexpansive ones. Note, however, that without
the quasi nonexpansivity of T the equality Fix T D Fej T needs not to be true. In
this case, Fix T needs not to be convex, even if T is Fejér monotone.
Example 2.1.23. Let H D R2 , X WD Œa; b R for 1 a b C1
and h W X ! RC be a function with infx2Œa;b h.x/ D 0. Define the operator T W
X ! R2 by
x if j2 j h.1 /
T x WD
.1 ; 0/ if j2 j > h.1 /,
where x D .1 ; 2 / (see Fig. 2.2).
The reader may check that Fej T D Œa; b f0g and that Fix T D fx 2 X W j2 j
h.1 /g. If h is positive in at least one point, then Fej T ¤ Fix T . If, moreover, h is
not concave, then Fix T is not convex.
Let Ui W X ! X , i 2 I WD f1; 2; : : : ; mg, and:
P
(i) U WD i 2I !i Ui , where w D .!1 ; !2 ; : : : ; !m / 2 m or
(ii) U WD Um Um1 : : : U1 .
As we observed before, the following inclusion holds
\
Fix Ui Fix U (2.14)
i 2I
(see Remark 2.1.1) and the converse inclusion holds in case (i) when all Ui ,
i 2 I , are nonexpansive operators with a common fixed point and w 2 ri m (see
Theorem 2.1.14). It turns out that, in both cases (i) and (ii), the inclusion converse
to (2.14)) is true for strictly quasi-nonexpansive operators (see [22, Proposition
2.12], where the property was formulated for attracting operators). P In case (i) this
property is also true for a more general form of the operator U D i 2I !i U , where
the weights !i , i 2 I , may depend on x.
50 2 Algorithmic Operators
!j .x/ k Uj x x k ¤ 0. (2.15)
Denote
I.x/ WD fi 2 I W x … Fix Ui g (2.16)
for a family of operators Ui W X ! H, i 2 I . The subset I.x/ is called a subset of
violated constraints. Note that w is appropriate if and only if
P
(i) Let Ux D i 2I !i .x/Ui x, where the weight function w W X ! m is
appropriate. Then the convexity of the norm, (2.19) and (2.15) yield
X
kUx zk D !i .x/.Ui x z/
i 2I
X X
!i .x/ kUi x zk !i .x/ kx zk D kx zk ,
i 2I i 2I
kUx zk D kUm : : : U1 x zk
D Um : : : Uj x z
Um1 : : : Uj x z
: : : Uj x z kx zk ,
kx zk D kUx zk < kx zk ,
T
a contradiction. We have proved that Fix U i 2I Fix Ui . Hence, (2.18) holds
and, in both cases (i) and (ii), U is C -strictly quasi-nonexpansive. t
u
Note that equality (2.18) needs not to be true for nonexpansive operators, even if
they have a common fixed point.
Example 2.1.27. (cf. [22, Remark 2.11]) Let X H be a subspace with dim X > 0.
Let Ui W X ! X , Ui WD Id, i D 1; 2. We have U2 U1 D Id, consequently,
Fix.U2 U1 / D X , but Fix U1 \ Fix U2 D f0g.
The assumption on the C -strict quasi nonexpansivity in Theorem 2.1.26 (i) can
be weakened. In this case it suffices to suppose that all Ui are quasi-nonexpansive,
i 2 I , and at least one of them is C -strictly quasi-nonexpansive. The assumption
that the weight function w is appropriate should be replaced in this case by a stronger
one, namely: wj .x/ > 0 for all x such that I.x/ ¤ ; and for all j 2 I.x/. We leave
the proof of this fact to the reader.
A stronger version of the first part of Theorem 2.1.26 (ii) for two operators is
stated below (cf. [346, Proposition 1(d) (i)]).
52 2 Algorithmic Operators
kx zk D kS T x zk kT x zk < kx zk .
kx zk D kT S x zk < kS x zk kx zk .
kS T x zk kT x zk kx zk ,
kT S x zk kS x zk kx zk ,
where the second inequality is strict if x … Fix S and the first one is strict if x 2
Fix S and x … Fix T . Hence, T S is strictly quasi-nonexpansive. u
t
2.1 Basic Definitions and Properties 53
x y
cl conv C
Proof. The corollary follows from Theorem 2.1.28. We leave to the reader an easy
proof by induction with respect to m. t
u
The assumption of Theorem 2.1.28 that T is strictly quasi-nonexpansive is essen-
tial. Note that the composition of quasi-nonexpansive operators needs not to
be quasi-nonexpansive (see Example 2.1.52). Furthermore, the assumptions of
Theorem 2.1.28 do not yield the strict quasi nonexpansivity of the operator S T
(see Example 2.1.54).
Tα x z
Tx
x − T x; u − T x 0
cutters a T -class (see [24, Definition 2.2] and [121, Definition 2.1]). Yamada and
Ogura (see [346, Sect. B]) and Măruşter (see [254]) call the operators firmly quasi-
nonexpansive. Zaknoon, Segal and Censor denoted cutters as directed operators (see
[104–106, 307, 356]). In [69] these operators were called separating operators. The
name cutter was proposed by Cegielski and Censor in [70].
Note that a separator and, in particular, a cutter need not to be continuous
operators.
Remark 2.1.31. Let T W X ! H have a fixed point. Then, by Lemma 1.2.5, the
operator T is a cutter if and only if
hT x x; z xi kT x xk2 (2.21)
holds for all x 2 X and for all z 2 Fix T (cf. [121, Proposition 2.3 (ii)]), and T is
an ˛-relaxed cutter, where ˛ 2 Œ0; 2, if and only if
holds for all x 2 X and for all z 2 Fix T . Furthermore, if T is a cutter, then
T jFix T D Id. Therefore, T is a cutter (respectively, an ˛-relaxed cutter) if and only
if inequality (2.21) (respectively, (2.22)) is satisfied for all x … Fix T and for all
z 2 Fix T . Relaxed cutters were also studied in [253, 255, 346] and in [249], where
they were called averaged quasi-nonexpansive mappings.
is nonempty, then Sep T is the largest subset for which T is a separator. Further-
more, Sep T is a closed convex subset.
Proof. The first part of the lemma follows directly from Definition 2.1.30. The
second part follows from the fact that an intersection of closed convex subsets is
closed and convex. t
u
If T is nonexpansive, then Fix T is a closed convex subset (see Proposition 2.1.11).
It turns out that cutters have the same property. The second part of the following
lemma was proved in [24, Proposition 2.6 (i)–(ii)].
Lemma 2.1.36. Let T W X ! H. The following inclusion holds
If T is a cutter, then a converse inclusion is also true. Hence, the subset of fixed
points of a cutter is closed and convex.
Proof. Let y 2 Sep T , i.e., hx T x; y T xi 0 for all x 2 X . If we take x D y,
we get ky T yk 0, and hence, T y D y, i.e., y 2 Fix T . Now suppose that T is a
cutter and that y 2 Fix T . Then for any x 2 X we have hy T x; x T xi 0g, i.e.,
56 2 Algorithmic Operators
y 2 Sep T . Therefore, we have Sep T D Fix T . The subset Fix T is closed convex
as an intersection of closed convex subsets. t
u
It follows from Remark 2.1.32 (ii) and from Lemmas 2.1.35 and 2.1.36 that a cutter
T W X ! X is a separator of any nonempty subset of Fix T . Therefore, we will
restrict our further considerations of separators to cutters. Note, however, that the
converse inclusion of (2.23) is not true in general (see Example 2.2.7). Hence, there
is a separator with a fixed point which is not a cutter.
It is an immediate consequence of the characterization of the metric projection
(see Theorem 1.2.4) that an operator T W H ! H is a metric projection onto a closed
convex subset if and only if T 2 D T and T is a cutter (a more general fact will be
presented in Theorem 2.2.5). In this case, we have T D PFix T . Even if a cutter
T is not idempotent, T is closely related to the metric projection. The following
corollary was proved in [121, Proposition 2.3 (iii)].
Corollary 2.1.37. Let T W X ! H be a cutter. Then, for any x 2 X , it holds
kT x xk kPFix T x xk . (2.24)
Proof. If x 2 Fix T , then inequality (2.24) is obvious. Now let x … Fix T . Then it
follows from inequality (2.21) for z WD PFix T x together with the Cauchy–Schwarz
inequality that
hT x x; PFix T x xi
kT x xk kPFix T x xk
kT x xk
for all x 2 X and z 2 Fix T . If T satisfies (2.25) with ˛ > 0, then T is called
strongly quasi-nonexpansive (SQNE).
A property which is more general than the strong nonexpansivity was introduced
by Halperin [198, Sect. 2] and was called '-property, where ' W Œ0; 1/ ! Œ0; 1/ is
a nondecreasing function. If '.t/ D t 2 for all t 2 Œ0; 1/, then '-property is equiva-
lent to the strong quasi-nonexpansivity. The notion strong quasi nonexpansivity was
introduced by Bruck [50, Sect. 1] for operators defined on a metric space. Strongly
quasi-nonexpansive operators are widely studied in the literature. Bauschke and
Borwein use the name strongly attracting operators for operators which are NE
and SQNE (see [22, Definition 2.1]). Reich and Zaslavski define a more general
operator and call it a uniformly F -attracting mapping, where F D Fix T (see
[297, Sect. 1]). Vasin and Ageev call the ˛-SQNE operators, where ˛ 2 .0; 1/,
Q-pseudocontractive operators (see [333, Definition 2.3]). Yamada and Ogura
2.1 Basic Definitions and Properties 57
use the notation ˛-attracting quasi-nonexpansive for the ˛-SQNE operators [346].
Crombez denotes the class of ˛-SQNE operators by F ˛ (see [126, pages 160–161])
and gives several equivalent conditions for T 2 F ˛ (see [126, Theorem 2.1]).
It follows easily from the equivalence (a),(c) of Lemma 1.2.5 that an operator
T which has a fixed point is a cutter if and only if it is 1-strongly quasi-
nonexpansive. The following theorem extends this property to relaxations of T (cf.
[121, Proposition 2.3 (ii)]).
Theorem 2.1.39. Assume that T W X ! H has a fixed point and let 2 .0; 2.
Then T is a cutter if and only if its relaxation T is 2
-strongly quasi-nonexpansive,
i.e.,
2
kT x zk2 kx zk2 kT x xk2 (2.26)
for all x 2 X and for all z 2 Fix T .
Proof. Since
T x x D .T x x/,
the properties of the inner product yield
2
kT x zk2 kx zk2 C kT x xk2
D kx z C .T x x/k2 kx zk2 C .2 / kT x xk2
D 2.kT x xk2 hz x; T x xi/
D 2hz T x; x T xi
for all x 2 X and for all z 2 C . The assertion follows directly from the equalities
above. t
u
The following corollary is an equivalent formulation of Theorem 2.1.39.
Corollary 2.1.40. Assume that U W X ! H has a fixed point and let ˛ 2 .0; 2.
Then U is an ˛-relaxed cutter if and only if U is 2˛
˛ -strongly quasi-nonexpansive.
2ı
kT x zk2 kx zk2 .2 C / kT x xk2 , (2.27)
kT x xk
where ı > 0 is such that B.z; ı/ Fix T . If X is bounded, then T is int Fix T -
strictly quasi-nonexpansive for any 2 .0; 2.
58 2 Algorithmic Operators
γ = γ(λ, μ)
1
0
2
2
1
1
μ 0 λ
is
2 4. C /
D D . (2.29)
. 2 C 1
2
/ C1 4
Moreover,
4 minf; g 4 maxf; g
0 < minf; g < < 2.
minf; g C 2 maxf; g C 2
Proof. Suppose that T is a -relaxed cutter and that U is a -relaxed cutter. Take
a WD T x x and b WD U T x T x. Then it follows from inequality (2.22) that
hz x; ai 1 kak2 and hz T x; bi 1 kbk2 for any z 2 Fix U \ Fix T .
Let ; 2 .0; 2/ and be defined by (2.29). Then Lemma 2.1.45 yields
1
hz x; UTx xi kU T x xk2
1
D hz x; a C bi ka C bk2
1
D hz x; ai C hz x; bi ka C bk2
1
D hz x; ai C hz T x; bi C ha; bi ka C bk2
1 1 1
kak2 C kbk2 C ha; bi ka C bk2
1 1 1 1 2
D. / kak2 C . / kbk2 C .1 /ha; bi
s s 2
1 1 1
1
D a b 0.
1
ıD . (2.30)
1
C 1
2
m D ˛m 1 . (2.31)
˛1
. 2˛ 1
C ˛2
2˛2
C ::: C 2˛m
/ C1
1
ım D . (2.32)
1
ˇ1
C 1
ˇ2
C ::: C 1
ˇm
Moreover,
2m mini 2I ˛i 2m maxi 2I ˛i
0 < min ˛i < m <2
i 2I .m 1/ mini 2I ˛i C 2 .m 1/ maxi 2I ˛i C 2
(2.33)
and
mini 2I ˇi maxi 2I ˇi
0< ım . (2.34)
m m
Proof. The assertion is obvious for m D 1. Note that m D ım2C1 and that
Corollary 2.1.43 yields the equivalence of conditions (i) and (ii). We prove by
induction with respect to m that these conditions hold for any m 2.
10 If m D 2, then conditions (i) and (ii) follow directly from Theorem 2.1.46 and
from Corollary 2.1.47.
20 Suppose that (ii) is true for some m D k. Consequently, Uk is ık -SQNE. It
follows now from Corollary 2.1.47 that the operator UkC1 D TkC1 Uk is ı-SQNE,
where
1 1
ıD D D ıkC1 .
1
ık
C ˇkC1
1 1
ˇ1
C 1
ˇ2
C ::: C 1
ˇk
C 1
ˇkC1
Now, for m D k C 1, equality (2.31) follows from the above mentioned equivalence
of (i) and (ii).
Hence, we have proved that conditions (i) and (ii) hold for all m 1. Both
inequalities in (2.34) follow immediately from equality (2.32). Now we have
2 2 2 2m maxi 2I ˛i
m D D D < 2.
ım C 1 mini 2I ˇi
C1 mini 2I
2˛i
.m 1/ maxi 2I ˛i C 2
m
m
˛i
C1
62 2 Algorithmic Operators
2m mini 2I ˛i
m > min ˛i > 0
.m 1/ mini 2I ˛i C 2 i 2I
D kUx xk2
Moreover,
and
X X i 1X 1
U D i Ui D Id C .Ti Id/ D Id C !i .Ti Id/ D Id C .T Id/,
i 2I i 2I
˛i ˛ i 2I ˛
i.e., T D Id C˛.U Id/ and T is an ˛-relaxed cutter. The second part of the theorem
follows now immediately from Corollaries 2.1.40 and 2.1.43. Inequalities in (2.37)
are obvious and inequalities in (2.38) follow easily from (2.36). t
u
Bauschke and Borwein proved that a convex combination of ˇi -SQNE operators,
i 2 I , with a common fixed point is ˇ-SQNE, where ˇ WD mini 2I ˇi (see [22,
Proposition 2.12]). By inequality (2.38) this result is weaker than Theorem 2.1.50.
Note that this inequality is strict if ˇi ¤ ˇj for at least one pair i; j 2 I for which
!i and !j are nonzero. The second part of Theorem 2.1.50 for m D 2 was proved
by Yamada and Ogura (see [346, Proposition 1(c)]).
The following important result extends Theorem 2.1.39.
Theorem 2.1.51. Let S W H ! X be nonexpansive, T W X ! H be a cutter and
2 .0; 2/. If Fix S \ Fix T ¤ ;, then, for any x 2 Fix S and z 2 Fix S \ Fix T , the
following estimations hold
and
2
kS T x zk2 kx zk2 kS T x xk2 . (2.40)
Consequently, the operator S T jFix S is 2
-strongly quasi-nonexpansive.
Proof. Let x 2 Fix S and z 2 Fix S \ Fix T . Then the assumptions that S is
nonexpansive and T is a cutter yield
64 2 Algorithmic Operators
One can easily check that S; T are quasi-nonexpansive, Fix S D Fix T D f0g and
Fix S T D f0; 1g. The operator S T is not quasi-nonexpansive, because a subset of
fixed points of a quasi-nonexpansive operator is convex (see Lemma 2.1.21).
Example 2.1.53. Let X D H WD R2 , A WD fx 2 R2 W he; xi 1g, B WD fx 2 R2 W
2 0g, U1 WD PB , U2 WD PA and U WD U2 U1 . Then U1 and U2 are cutters and it
follows from Theorem 2.1.26 that Fix U D Fix U1 \ Fix U2 D A \ B ¤ ;. For x D
.0; 1/ and z D .1; 0/ 2 A \ B, we have Ux D . 12 ; 12 / and hx Ux; z Uxi D 12
(see Fig. 2.6). Therefore, U is not a cutter.
2.2 Firmly Nonexpansive Operators 65
kS T x zk D kS x zk D kx zk .
1
hy x; S x xi C hx y; Sy yi k.S x x/ .Sy y/k2 . (2.42)
66 2 Algorithmic Operators
' x
1
hT y x; S x xi C hT x y; Sy yi .kS x xk2 C kSy yk2 /.
2 1
hyx; S xxiChxy; SyyiC hS xx; Syyi .kS x xk2 CkSy yk2 /.
The latter inequality is equivalent to (2.42). t
u
kT x T yk kx yk hT x T y; x yi kT x T yk2 0,
is nonexpansive and monotone for ' 2 .0; =2/, but T is not firmly nonexpansive
(see Fig. 2.7).
Now we prove a property of firmly nonexpansive operators, which also appears
in Theorem 1.2.4. In particular, the characterization of the metric projection is,
2.2 Firmly Nonexpansive Operators 67
actually, a corollary of the following theorem which is due to Goebel and Reich
(see [185, pp. 43–44]).
Theorem 2.2.5. Let T W X ! H be an operator with a fixed point.
(i) If T is firmly nonexpansive, then T is a cutter, i.e.,
hz T x; x T xi 0 (2.43)
hT y T x; x T xi C hT x T y; y T yi 0,
hT v T u; u T ui 0, (2.44)
hT u T v; v T vi 0. (2.45)
hT u T v; .T u T v/ .u v/i 0;
Example 2.2.7. (cf. [204] and [78, Sect. 4.10]) Let a 2 H, kak D 1 and ˛ > 0.
Furthermore, let A WD fx 2 H W ha; xi D 0g, B1 WD fx 2 H W ha; xi D ˛g,
B2 WD fx 2 H W ha; xi D ˛g and B WD fx 2 H W jha; xij ˛g. The subset
B is a band with a width of 2˛ and is bounded by two hyperplanes B1 and B2 .
The hyperplane A cuts the band B into two bands bounded by A and B1 and by A
and B2 . Define the operator T W H ! H as follows
8
ˆ
ˆ PA x if jha; xij 2˛
<
2PB1 x x if ˛ < ha; xi < 2˛
Tx D (2.46)
ˆ 2PB2 x x if 2˛ < ha; xi < ˛
:̂
x if jha; xij ˛
Note that T projects onto A all points with the distance to A equal at least 2˛, T
reflects (with respect to the closest hyperplane B1 or B2 ) the points which do not
belong to the band B with the distance to A less than 2˛ and T does not move
the elements of the band B (see Fig. 2.8). The reader can easily check that T is
nonexpansive and that T is a separator of A but T is not firmly nonexpansive. Note
that Fix T D B and that T is not a cutter, i.e., it does not separate Fix T , but T
separates A (see Fig. 2.8).
T x WD Œ1 .x/PA x C .x/PB x,
12
where .x/ D 12 C22
for x D .1 ; 2 / 2 R2 (see Fig. 2.9). We have PA x D .1 ; 0/,
PB x D .0; 2 /. Consequently,
1 22 2 2
Tx D . ; 2 1 2/
12C 2 1 C 2
2
for x ¤ .0; 0/. Note that z WD .0; 0/ is the unique fixed point of T .
The operator T is a cutter, because
12 22
hz T x; x T xi D 0
12 C 22
2.2 Firmly Nonexpansive Operators 69
x
PB x Tx
FixT PA x
22
22
kT x T yk2 .
12
21 1
1 /2 C .2
2 /2
D 1.
kx yk2 22
22
.1 C 12
/.1 C
21
/Œ.1
1 /2 C .2
2 /2
1 2
hT x T y; x yi D < D kT x T yk2 .
5 25
Therefore, T is not firmly nonexpansive.
The following property of firmly nonexpansive operators (cf. [22, Lemma 2.4
(iv)]) is often used in applications.
70 2 Algorithmic Operators
One can find in the literature several equivalent definitions of firmly nonexpansive
operators. The properties of these operators were studied by Zarantonello [357,
Sect. 1], Bruck [49, Sects. 2 and 3], Rockafellar [299], Bruck and Reich [51, Sect. 1],
Goebel and Reich [185, Chap. 1, Sect. 11], Reich and Shafrir [296], Goebel and Kirk
[184, Chap. 12], Bauschke and Borwein [22, Sects. 2 and 3], Byrne [56, Sect. 2], and
by Crombez [127, Sect. 2].
The class of firmly nonexpansive operators is included in the class of nonexpan-
sive ones (see Theorem 2.2.4). Further important relationships between these two
classes are also useful for the investigation of firmly nonexpansive operators. These
relationships are given in the following theorem.
Theorem 2.2.10. Let T W X ! H. Then the following conditions are equivalent:
(i) T is firmly nonexpansive.
(ii) T is nonexpansive for any 2 Œ0; 2.
(iii) T has the form T D 12 .S C Id/; where S W X ! H is a nonexpansive operator.
(iv) Id T is firmly nonexpansive.
(v) For all x; y 2 X it holds
i.e., T is nonexpansive.
(ii))(iii) This implication is obvious. It suffices to take S D T for D 2.
(iii))(iv) Let S be nonexpansive, T WD 12 .S C Id/ and G WD Id T . Then we
have G D 12 .Id S / and
for all x; y 2 X .
(iv))(v) Let G WD Id T be firmly nonexpansive. Then, for all x; y 2 X we
have
(v))(i) Let x; y 2 X . If (2.48) holds, then, by the properties of the inner product,
we have
1
h.˛/ D k˛.x y/ C .1 ˛/.T x T y/k2
2
h0 .0/ D hT x T y; .x y/ .T x T y/i.
1
T D .1 / Id C .Id CS / D ŒId C.1 / Id CS
2 2
U D Id C.T= Id/:
The corollary follows now from the fact that T= is firmly nonexpansive (see
Corollary 2.2.11). t
u
Corollary 2.2.13. Let X H be a closed convex subset and S W X ! H. The
following conditions are equivalent:
(i) S is nonexpansive,
(ii) S D 2T Id, where T W X ! H is a firmly nonexpansive operator.
Proof. (ii))(i) Let S WD 2T Id for a firmly nonexpansive operator T . It follows
from the implication (i))(ii) in Theorem 2.2.10 that S is nonexpansive.
(i))(ii) Let S be nonexpansive and T WD 12 .S C Id/. By the implication
(iii))(i) in Theorem 2.2.10 the operator F is firmly nonexpansive. Furthermore,
S D 2T Id. t
u
Definition 2.2.14. (cf. [127, Definition 2.1]) We say that an operator U W X ! H
is -firmly nonexpansive (-FNE), where > 0, if
Vasin and Ageev call a -firmly nonexpansive operator for 2 .0; 1/, a pseudo-
contractive operator (see [333, Definition 2.5]). In [127, Theorem 2.3] several
equivalent conditions for U to be -FNE are presented.
By the equivalence (i),(v) of Theorem 2.2.10, an operator is firmly nonex-
pansive if and only if it is 1-firmly nonexpansive. Note, however, that there is a
difference between a -RFNE operator and a -FNE operator. Below, we present
the relationship between these two notions.
Corollary 2.2.15. Let 2 .0; 2/. An operator U W X ! H is -relaxed firmly
nonexpansive if and only if U is 2
-firmly nonexpansive, i.e.,
74 2 Algorithmic Operators
2
kUx Uyk2 kx yk2 k.x Ux/ .y Uy/k2
2
kUx zk2 kx zk2 kUx xk2
2
kUx Uyk2 kx yk2 C k.x Ux/ .y Uy/k2
D kUx Uyk2 kx yk2 C .2 / kGx Gyk2
D 2.hx y; Gx Gyi kGx Gyk2 /.
The first part of the corollary follows now from the equivalence (i),(iv) in
Theorem 2.2.10, and now the other part follows directly from the definition of an
˛-strongly quasi-nonexpansive operator. t
u
Definition 2.2.16. Let ˛ 2 .0; 1/. We say that an operator T W X ! H is
˛-averaged or, shortly, averaged (AV) if
T D .1 ˛/ Id C ˛S
U D .1 ˛/ Id C ˛.2T Id/
D .1 2˛/ Id C2˛T
2
Fix Fix
=(2- )/
=(2- )/
=2
P
that T D 12 .S C Id/ for S WD i 2I !i Si . By Lemma 2.1.12 (i), the operator
S is nonexpansive. The corollary follows now from the implication (iii))(i) in
Theorem 2.2.10. t
u
In Fig. 2.10 we shortly present important relationships among the FNE operators,
cutters, QNE operators SQNE operators and AV operators, which are proved in
Sects. 2.1.3, 2.2.1 and 2.2.2. In Fig. 2.10, T W X ! H and U WD I D Id
C.T Id/ is its -relaxation, where 2 .0; 2/. We will extend this figure in
Sect. 3.9.
The basic facts concerning firmly nonexpansive operators presented in the previous
section yield further properties of the metric projection.
Theorem 2.2.21. Let C H be a nonempty closed convex subset and PC W H ! H
be the metric projection onto C . Then the operator PC is:
(i) Idempotent, consequently Fix PC D C ,
(ii) A cutter,
(iii) Firmly nonexpansive,
(iv) Monotone and nonexpansive,
(v) Averaged.
Proof. (i) The property follows directly from the definition of the metric projec-
tion.
(ii) It follows from (i) and from the characterization of the metric projection (see
Theorem 1.2.4) that hz PC x; x PC xi 0 for all x 2 H and for all z 2
C D Fix PC , which means that PC is a cutter.
(iii) The property follows directly from (i), (ii) and from Theorem 2.2.5 (ii).
2.2 Firmly Nonexpansive Operators 77
(iv) By Theorem 2.2.4, any firmly nonexpansive operator is monotone and nonex-
pansive. Therefore, the property follows from (iii).
(v) By the firm nonexpansivity of PC and by the implication (i))(iii) in The-
orem 2.2.10, we can write PC D 12 .S C Id/ for a nonexpansive operator
S W X ! H. Hence, PC is averaged.
t
u
Definition 2.2.22. Let C H be a nonempty closed convex subset. We call a
relaxation of the metric projection PC W H ! C a relaxed metric projection onto
the subset C and we denote it by PC; or, shortly, by P . If < 1, then P is called
an under-projection. If > 1, then P is called an over-projection. If D 2, then
P is called the reflection.
We have
PC; D P D Id C.PC Id/.
Corollary 2.2.23. Let C H be a nonempty closed convex subset, 0 and
P W H ! H be a relaxed metric projection. Then
(i) P is a nonexpansive operator for all 2 Œ0; 2,
(ii) Fix P D C for all > 0,
(iii) For all x 2 H, z 2 C and 2 .0; 2 the following inequality holds
2
kP x zk2 kx zk2 kP x xk2 . (2.49)
Consequently, P is 2
-strongly quasi-nonexpansive for all 2 .0; 2.
Proof. Part (i) follows from the equivalence (i),(ii) in Theorem 2.2.10, because
PC is firmly nonexpansive (see Theorem 2.2.21 (iii)). Part (ii) is obvious, because
Fix PC D C . Part (iii) follows now from Corollary 2.2.9. t
u
Corollary 2.2.24. Let C H be a nonempty closed convex subset and x; y 2 H.
Then
In particular,
kPC x zk2 kx zk2 kPC x xk2 (2.52)
for all x 2 H and z 2 C . Consequently, the metric projection PC W H ! C is
strongly quasi-nonexpansive.
Proof. By Theorem 2.2.21 (iii), the metric projection is firmly nonexpansive.
Therefore, inequalities (2.50) and (2.51) follow directly from the implication
(i))(v) in Theorem 2.2.10 and from the Cauchy–Schwarz inequality. The second
part follows directly from Theorem 2.2.21 (i). t
u
78 2 Algorithmic Operators
PX (x + α2u)
PX (x + α1u) X
u
Corollary 2.2.25. Let T W X ! H and 2 .0; 2/. If T is a cutter, then for any
x 2 X and z 2 Fix T the following estimations hold
and
2
kPX T x zk2 kx zk2 kPX T x xk2 : (2.53)
Consequently, the operator PXT W X ! X is 2
-strongly quasi-nonexpansive.
kPX .x C ˛2 u/ xk2
kPX .x C ˛1 u/ xk2 C kPX .x C ˛2 u/ PX .x C ˛1 u/k2 . (2.54)
The characterization of the metric projection (see Theorem 1.2.4) and its mono-
tonicity (see Theorem 2.2.21 (iv)) yield
hx PX z; PX y PX zi D hx z; PX y PX zi C hz PX z; PX y PX zi
hy z; PX y PX zi 0,
1
Therefore, we suppose without loss of generality that C is closed. It turns out that
the functions d.; C / and d 2 .; C / are convex and differentiable.
Lemma 2.2.27. Let C H be a closed convex subset. Then the function
f W H ! R, f .x/ WD 12 d 2 .x; C / is differentiable and Df .x/ D x PC x for
all x 2 H.
Proof. (cf. [167, Proposition 2.2] and [209, Chap. IV, Example 4.1.6]) Let x; h 2 H.
It follows from the definition of the metric projection and from the properties of the
inner product that
f .x C h/ f .x/ hx PC x; hi
1 1
D kx C h PC .x C h/k2 kx PC xk2 hx PC x; hi
2 2
1 1
kx C h PC xk2 kx PC xk2 hx PC x; hi
2 2
1
D khk2 .
2
Similarly, by the definition of the metric projection, the Cauchy–Schwarz inequality
and the nonexpansivity of the metric projection, we obtain
f .x C h/ f .x/ hx PC x; hi
1 1
D kx C h PC .x C h/k2 kx PC xk2 hx PC x; hi
2 2
1 1
kx C h PC .x C h/k2 kx PC .x C h/k2 hx PC x; hi
2 2
1
D khk2 C hPC x PC .x C h/; hi
2
80 2 Algorithmic Operators
1
khk2 kPC x PC .x C h/k khk
2
1 1
khk2 khk2 D khk2 .
2 2
Now we see that
1 1
khk2 .f .x C h/ f .x/ hx PC x; hi/ khk2 .
2 2
Consequently,
f .x C h/ D f .x/ C hx PC x; hi C o.khk/.
x PC x
Dh.x/ D . (2.56)
kx PC xk
Proof. Since h.x/ D infy2C kx yk, the convexity of h follows from the fact that
the function p W H H ! H, p.x; y/ WD kx yk is convex (as a composition
of a linear function .x; y/ ! x y and a convex function z ! kzk) and from
pthat for a convex function p, the function infy2C p.; y/ is convex. Since
the fact
h D d 2 .; C /, the differentiability of h as well as equality (2.56) for x … C
follow from Lemma 2.2.27 and from the formula D.kzk/ D kzk z
for z ¤ 0. t
u
Corollary 2.2.29. Let C H be a closed convex subset. Then the function f W
H ! R, f .x/ WD 12 d 2 .x; C / is convex.
Proof. The function f is convex as a composition f D g ı h of a convex function
h WD d.; C / and of a convex and increasing function g W RC ! R, g.t/ WD 12 t 2 .
t
u
(iii) PV is self-adjoint,
(iv) Id D PV C PV ? .
Proof. (i) Necessity. Let y WD PV x. By the characterization of the metric
projection (see Theorem 1.2.4), hx y; z yi 0 for all z 2 V . Suppose
that hx y; w yi < 0 for some w 2 V . Let u WD 2y w. Then u 2 V because
V is a linear subspace and we have
hx y; u yi D hx y; y wi > 0.
hx; PV ui D hPV x; PV ui
and
hu; PV xi D hPV u; PV xi.
By the symmetry of the inner product
i.e., PV is self-adjoint.
(iv) Let x 2 H. By (i), we have x PV x 2 V ? and x PV x D PV ? x. Since
x D PV x C .x PV x/, it holds x D PV x C PV ? x. t
u
82 2 Algorithmic Operators
PA x D PAa .x a/ C a (2.57)
y D PA x , y a D PAa .x a/ , hy a .x a/; z yi D 0
for any z 2 A.
2.2 Firmly Nonexpansive Operators 83
PA u PA v D PAa .u a/ C a .PAa .v a/ C a/
D PAa .u v/
D PAa .u v a/ PAa .a/
D PA .u v/ PA 0.
Therefore,
h.PA u u/ .PA v v/; PA u PA vi D 0,
i.e.,
hPA u PA v; u vi D kPA u PA vk2 .
(v) It follows from the properties of the inner product and from property (iv) that
(vi) Let 2 R. Since A a is a closed subspace, (2.57) and Theorem 2.2.30 yield
1C ˛ 1˛
T D UC Id ,
2 2
1C ˛
i.e.. T is 2 -averaged. By the convexity of the norm and the nonexpansivity of T ,
2 1˛
kUx Uyk D .T x T y/ .x y/
1C ˛ 1C ˛
2 1˛
kT x T yk C kx yk
1C ˛ 1C ˛
2˛ 1˛
kx yk C kx yk
1C ˛ 1C ˛
D kx yk ,
X
m
!i i
Pm D 1,
i D1 j D1 !j j
2.2 Firmly Nonexpansive Operators 85
the operator
X
m
!i i
T WD Pm Ti
i D1 j D1 !j j
X
m
U D !i ŒId Ci .Ti Id/
i D1
X
m
D Id C !i i .Ti Id/
i D1
0 1 !
X
m X
m
!i i X m
!i i
D Id C @ !j j A Pm Ti Pm Id
j D1 i D1 j D1 !j j i D1 j D1 !j j
D Id C.T Id/
y x D T y T x C a1 a2 . (2.59)
86 2 Algorithmic Operators
and from the fact that the family of nonexpansive operators is closed under
compositions and convex combinations (see Lemma 2.1.12). Note, however, that
[273, Theorem 3 (b)] is stronger than the result mentioned above, because
ı < ˛ C ˇ ˛ˇ
for ˛; ˇ 2 .0; 1/ and ı given by (2.60). It follows from Corollary 2.2.17 that the
result of Ogura and Yamada is equivalent to Theorem 2.2.37 with ; 2 .0; 2/.
Moreover, the proof of this theorem differs from the proof of [273, Theorem 3 (b)].
Note that the property of composition of relaxed cutters with a common fixed point,
expressed in Theorem 2.1.46 and the property of compositions of relaxed firmly
nonexpansive operators presented in Theorem 2.2.37 are similar. Therefore, it is
quite natural that the proofs of both theorems are similar. But Theorem 2.1.46 is no
special case of Theorem 2.2.37 because a cutter needs not to be firmly nonexpansive,
even if it is nonexpansive (see Example 2.2.8).
An equivalent formulation of the following result can be found in [349,
Lemma 1].
Corollary 2.2.39. Let T; U W H ! H be firmly nonexpansive. Then the com-
position V WD U T is 43 -relaxed firmly nonexpansive. Consequently, V is firmly
nonexpansive for all 2 Œ0; 34 and nonexpansive for all 2 Œ0; 32 . If, furthermore,
V has a fixed point, then V is strongly quasi-nonexpansive for all 2 .0; 32 /.
Proof. If we take D D 1 in Theorem 2.2.37, we obtain that V is 43 -relaxed
firmly nonexpansive. Recall that .V / D V (see Remark 2.1.3). Corollary 2.2.19
yields the firm nonexpansivity of V for all 2 Œ0; 34 . By the implication (i))(ii)
in Theorem 2.2.10, V is nonexpansive for all 2 Œ0; 32 . Now let Fix V ¤ ; and
2 .0; 32 /. Then V is strongly quasi-nonexpansive, by Corollary 2.2.9. t
u
Yamada et al. also proved that, for any > 32 , there exist firmly nonexpansive
operators T; U such that V is not nonexpansive, where V WD U T (see [349,
Remark 1 (b)]). This means that the constant 32 is optimal in Corollary 2.2.39.
Remark 2.2.40. Let T; U W H ! H be firmly nonexpansive having a common fixed
point. Then it follows from Corollaries 2.2.39 and 2.2.15 that U T is 12 -strongly quasi
nonexpansive. A special case of this property was proved in [152, Proposition 1] for
T; U being orthogonal projections onto subspaces of H.
Corollary 2.2.41. Let T W H ! H be firmly nonexpansive and 2 Œ0; 2. If V is a
closed affine subspace, then the operator U WD .1 /PV C PV T is 4
4
-relaxed
firmly nonexpansive.
Proof. Let V be a closed affine subspace. By Theorem 2.2.33 (vi), the operator PV
is affine, consequently,
.1 /PV C PV T D PV T .
4
Now it follows from Theorem 2.2.37 that U is 4 -relaxed firmly nonexpansive,
because the metric projection is firmly nonexpansive. t
u
88 2 Algorithmic Operators
A weaker formulation of Corollary 2.2.41 can be found in [349, Lemma 2], where
T WD PC for a closed convex subset C .
Theorem 2.2.42. Let Ti W X ! X be i -relaxed firmly nonexpansive, where
˛i 2 Œ0; 2, i 2 I . Then the composition Sm WD Tm Tm1 : : : T1 is m -relaxed firmly
nonexpansive, where m D 0 if i D 0 for all i 2 I , m D 2 if i D 2 for at least
one i 2 I and
2
m D
1 , (2.61)
1
21 C 2
22 C ::: C m
2m C1
otherwise. Moreover,
2m mini 2I i 2m maxi 2I i
m , (2.62)
.m 1/ mini 2I i C 2 .m 1/ maxi 2I i C 2
m 1 2 m
D C C ::: C . (2.63)
2 m 2 1 2 2 2 m
k kC1
D C
2 2 k 2 kC1
1 2 k kC1
D C C ::: C C ,
2 1 2 2 2 k 2 kC1
2.2 Firmly Nonexpansive Operators 89
mini 2I i m maxi 2I i
m m ,
2 mini 2I i 2 m 2 maxi 2I i
X
m
2i
D !i . (2.64)
i D1
i C1
2
kR x zk2 kx zk2 kR x xk2 (2.65)
2
2
kR x zk2 kx zk2 kR x xk2
2
D kx zk2 kR x xk2 .
2
(iv) The claim follows from Corollary 2.2.25.
t
u
If X D H, then R D T , nevertheless, estimation (2.65) is weaker than
estimation (2.47). Furthermore, estimation (2.65) is weaker than estimation (2.53).
Note, however, that we have supposed in Corollary 2.2.25 that the operator T W
X ! H is a cutter, consequently Fix T ¤ ;, while in Theorem 2.2.46 (iii) we have
supposed that Fix.PX T / ¤ ;, which is weaker than the assumption Fix T ¤ ;.
Proof. Since a firmly nonexpansive operator with a fixed point is a cutter (see
Theorem 2.2.5), the theorem follows from Lemmas 2.1.36 and 2.1.35. t
u
Corollary 2.2.48. Let X H be closed and convex. The subset of fixed points of a
nonexpansive operator S W X ! H has the form
\
Fix S D fz 2 X W 2hz x; S x xi kS x xk2 g, (2.66)
x2X
kT x T yk < kx yk or x y D T x T y
k.x nk y nk / .T x nk T y nk /k ".
x k y k 2 T x k T y k 2
2
D x k y k T x k T y k x k y k C T x k T y k ! 0,
2
k
i.e., .x y k / .T x k T y k / ! 0 and T is strongly nonexpansive. t
u
2.3 Strongly Nonexpansive Operators 93
In the previous sections we have proved that the following classes of operators are
closed under composition and under convex combination:
(a) The class of strictly relaxed cutters with a common fixed point (see Theo-
rems 2.1.46 and 2.1.50),
(b) The class of strongly quasi-nonexpansive operators with a common fixed point
(see Corollary 2.1.47 and Theorem 2.1.50),
(c) The class of strictly relaxed firmly nonexpansive operators (see Theo-
rems 2.2.37 and 2.2.35)
(d) The class of averaged operators (see Remark 2.2.38 and Corollary 2.2.36).
It turns out that the class of strongly nonexpansive operators has the same properties.
The first part of the theorem below was proved by Bruck and Reich in [51,
Proposition 1.1] and the other one by Reich in [295, Lemma 1.3].
Theorem 2.3.5. Let T1 ; T2 W X ! X be strongly nonexpansive and T have one of
the following forms:
(i) T WD T2 T1 ,
(ii) T WD .1 /T1 C T2 , where 2 Œ0; 1.
Then T is strongly nonexpansive.
Proof. By Lemma 2.1.12, the operator T is nonexpansive. Let the ksequences
fx k 1
g ; fy k 1
g X be such that .x k
y k
/ is bounded and x y k
kD0 kD0
T x k T y k ! 0.
k 0, consequently,
k
x y k T1 x k T1 y k ! 0
and
T1 x k T1 y k T2 .T1 x k / T2 .T1 y k / ! 0.
x
B(0, 1)
Ty
Uy
k
T x T y k D .1 /T1 x k C T2 x k .1 /T1 y k T2 y k
.1 / T1 x k T1 y k C T2 x k T2 y k
.1 / x k y k C x k y k D x k y k ,
consequently,
k
x y k T x k T y k
.1 /.x k y k T1 x k T1 y k / C .x k y k T2 x k T2 y k /.
Therefore, k
x y k T1 x k T1 y k ! 0
and k
x y k T2 x k T2 y k ! 0.
.x k y k / .T x k Ty k /
D .1 /..x k y k / .T1 x k T1 y k // C ..x k y k / .T2 x k T2 y k // ! 0,
T x WD ˛.x/Ux.
kT x T yk D k˛.x/Ux ˛.y/Uyk
D kU.˛.x/x/ U.˛.y/y/k
D k˛.x/x ˛.y/yk .
kx yk2 kT x T yk2
D kx yk2 k˛.x/x ˛.y/yk2
1 1
D kxk4 kyk4 C kxk3 C kyk3
4 4
1
hx; yi.kxk C kyk kxk kyk/
2
1
D .kxk kyk/2 .kxk C kyk/.1 .kxk C kyk//
4
1
C.kxk kyk hx; yi/.kxk C kyk kxk kyk/.
2
We have kxkCkyk kxkkyk, since kxk ; kyk 2 Œ0; 1. This fact and the Cauchy–
Schwarz inequality yield
1 1 1 1
kxk C kyk kxk kyk kxk kyk kxk kyk hx; yi,
2 2 4 4
consequently,
kx yk2 kT x T yk2
1 1
.kxk kyk/2 .kxk C kyk/.1 .kxk C kyk// C .kxk kyk hx; yi/2
4 4
(note that 1 14 .x k C y k / 12 ) and
k k
x y hx k ; y k i ! 0.
Now we have
k
x y k 2 D x k y k 2 C 2 x k y k hx k ; y k i ! 0,
hx; x T xi C 2 kx T xk2
D hx C .T x x/; x T xi
D hT x; x T xi
D h0 V x; x V xi 0.
1 hx; x T xi kxk2 1
< D D .
2 kx T xk 2 2
kxk C ˛ .x/ kxk
2 2
1 C .1 12 kxk/2
1 1 1
< lim D ,
2 k 1 C .1 1 x k /2 2
2
T
; x x D
.x/.T x x/ D .T
x x/, (2.68)
i.e., T
; is a -relaxation of an operator T
.
(d) For any ¤ 0 it holds Fix T
; D Fix T (cf. Remark 2.1.4).
The corollary below is a version of Theorem 2.1.39.
98 2 Algorithmic Operators
kT
; x zk2 kx zk2 .2 /
2 .x/ kT x xk2 (2.69)
2
kT
; x zk2 kx zk2 kT
; x xk2 .
Applying now (2.68) to the inequality above we obtain (2.69). t
u
Let T W X ! H be an operator with a fixed point. Our aim is to give sufficient
conditions for the step size function
W X ! .0; C1/, at which T
is a cutter. The
following definition was proposed in [70, Definition 9.17].
Definition 2.4.4. We say that an operator T W X ! H with a fixed point is oriented
if for all x … Fix T
hz x; T x xi
ı.x/ WD inf > 0. (2.70)
z2Fix T kT x xk2
If ı.x/ ı > 0 for all x … Fix T , then we say that T is strongly oriented.
It follows from Remark 2.1.31 that T W X ! H is strongly oriented if and only
if T is an ˛-relaxed cutter for some ˛ > 0.
Corollary 2.4.5. Let T W X ! H be an oriented operator with Fix T ¤ ;. If a
step size function
W X ! .0; C1/ satisfies the inequality
hz x; T x xi
.x/ (2.71)
kT x xk2
hz x; T
x xi D hz x;
.x/.T x x/i
k
.x/.T x x/k2
D kT
x xk2 .
2.4 Generalized Relaxations of Algorithmic Operators 99
hz T
x; x T
xi 0,
2
i.e., T
is a cutter. The
-strong quasi nonexpansivity of T
; follows now from
Corollary 2.4.3. t
u
The convergence of sequences generated by generalized relaxations of an algorith-
mic operator U , which we present in the next chapter, requires a stronger condition
than (2.71). As we will see, the convergence holds if we additionally suppose that
U is strongly oriented, or, equivalently, that the step size
.x/ ˛ for all x 2 X
and for a constant ˛ > 0. This leads to ˛-relaxed cutters (see Remark 2.1.31).
It is clear that if an operator T W X ! H with a fixed point is an ˛-relaxed cutter
for some ˛ > 0, then there exists a step size function
W X ! .0; C1/ satisfying
inequality (2.71), e.g.,
.x/ D ˛ 1 for all x 2 X (cf. (2.22)). In practice, however, it
is important to determine a step size
.x/ for which the difference between the right-
and the left-hand side of inequality (2.71) is as small as possible for all z 2 Fix T .
Theoretically, the best possibility would be
.x/ D ı.x/ for x … Fix U , where
ı.x/ is defined by (2.70), but the computation of ı.x/ is, in most cases, impossible,
because we usually do not know Fix T explicitly.
Having an ˛-relaxed cutter T we can construct its generalized relaxation T
;
with the range of the step size function
contained in Œ˛; C1/ and satisfying
assumptions of Corollary 2.4.5. The corollary below gives a collection of operators
which are ˛-relaxed cutters.
Corollary 2.4.6. Let U W X ! H have a fixed point. Then U is an ˛-relaxed cutter
with:
(a) ˛ D 1 if U is firmly nonexpansive,
(b) ˛ D if U is -relaxed firmly nonexpansive, where 2 .0; 2,
(c) ˛ D 2 if U is nonexpansive,
(d) ˛ D 2 if U is -averaged, where 2 .0; 1/,
(e) ˛ D 1Cˇ
2
if U is ˇ-strongly quasi-nonexpansive, where ˇ > 0.
Proof. (a) Let U be firmly nonexpansive. Then it follows from the first part of
Theorem 2.2.5 that T is a cutter, i.e., T is a 1-relaxed cutter.
(b) Let 2 .0; 2 and U WD Id C.T Id/ for a firmly nonexpansive operator T .
Then, by (a), we have
hz x; Ux xi D hz x; T x xi
1
kT x xk2 D kUx xk2 .
(c) Let U be nonexpansive. Then U D 2T Id for a firmly nonexpansive operator
T (see Corollary 2.2.13) and this case is covered by (b) for D 2.
(d) Let 2 .0; 1/ and U WD .1 / Id CS for a nonexpansive operator S . Then
U is 2-relaxed firmly nonexpansive (see Corollary 2.2.17). The claim follows
now from (b) with D 2.
100 2 Algorithmic Operators
.x/ D ı.x/, where ı.x/ is given by (2.70) for x … Fix T . Consequently, the
generalized relaxation T
; is strongly quasi-nonexpansive for any 2 .0; 2/ (see
Theorem 2.4.5). Note that
.x/ 1, by Remark 2.1.31. The following example
shows, however, that there is a cutter T such that the generalized relaxation T
;
is strongly quasi-nonexpansive for all 2 .0; 2/ if and only if
.x/ 1 for all
x … Fix T .
Example 2.4.8. Let T
; be a generalized relaxation of the metric projection PC W
H ! H, where C H is a nonempty closed convex subset, i.e., T
; .x/ D
x C
.x/.PC x x/ for a relaxation parameter 2 .0; 2/ and for some step
size function
W H ! .0; C1/. For any x 2 H we have
kT
; x PC xk2 D kx C
.x/.PC x x/ PC xk2
D kx PC xk2 C 2
2 .x/ kPC x xk2 2
.x/ kPC x xk2
D kx PC xk2
.x/.2
.x// kPC x xk2 ,
consequently,
2
.x/
kT
; x PC xk2 D kx PC xk2 kT
; x xk2 . (2.72)
.x/
kT
; x zk2 kx zk2 ˛ kT
; .x/ xk2 (2.73)
for some ˛ > 0, for all x 2 H and z 2 C WD Fix PC . Note that ˛ can depend on .
Let x … C and z D PC x. Then (2.72) and (2.73) yield
2
.x/
0<˛ .
.x/
There exists a constant ˛ satisfying the above inequalities for all 2 .0; 2/ if and
only if
.x/ 1.
2.4 Generalized Relaxations of Algorithmic Operators 101
T¾x
√
z 1 2 2
Tx Ty
C
y
D
It is clear that T is a projection with Fix T D C . For kxk > 2, let Ux be the
unique common point of the segment Œx; T x and the circle D. Define the function
W R2 ! R by (
1 if kxk 2
.x/ WD kUxxk
kT xxk if kxk > 2.
FNE operators
cutters with
a common fixed point
is closed under
relaxed cutters with convex combination
a common fixed point
strictly RFNE
operators is closed under
composition
SQNE operators with
a common fixed point
SNE operators
SUMMARY
2.5 Exercises
Exercise 2.5.4. Prove that the assumption on the C -strict quasi nonexpansivity
in Theorem 2.1.26 (i) can be weakened. In this case it suffices to suppose that all
Ui are quasi-nonexpansive, i 2 I , and at least one of them is C -strictly quasi-
nonexpansive. The assumption that the weight function w is appropriate should be
replaced in this case by a stronger one, namely: wj .x/ > 0 for all x such that
I.x/ ¤ ; and for all j 2 I.x/.
Exercise 2.5.5. Prove Corollary 2.1.29.
Exercise 2.5.6. Prove Lemma 2.1.45.
Exercise 2.5.7. Show that the operator T W R2 ! R2 ,
is nonexpansive and monotone for ' 2 .0; =2/, but T is not firmly nonexpansive.
Exercise 2.5.8. Prove Lemma 2.2.2.
Exercise 2.5.9. Show that the operator T presented in Example 2.2.7 is nonexpan-
sive and that T is a separator of A, but T is not firmly nonexpansive.
Exercise 2.5.10. Let H D R2 , A WD fx 2 R2 W 2 D 0g and B WD fx 2 R2 W
1 D 2 g. By Theorem 2.2.21 (iii) PA and PB are firmly nonexpansive. Check that
T WD PB PA is not firmly nonexpansive.
•
Chapter 3
Convergence of Iterative Methods
A. Cegielski, Iterative Methods for Fixed Point Problems in Hilbert Spaces, 105
Lecture Notes in Mathematics 2057, DOI 10.1007/978-3-642-30901-4 3,
© Springer-Verlag Berlin Heidelberg 2012
106 3 Convergence of Iterative Methods
In this section we present properties of the weak convergence which will be applied
in the sequel.
Lemma 3.2.1. If x k * x 2 H, then lim infk x k kxk.
Proof. Let x k * x. The lemma is clear for x D 0. Let now x ¤ 0. By the Cauchy–
Schwarz inequality, we have
lim inf kxk x k lim infhx; x k i D kxk2 ,
k k
i.e., lim infk x k kxk. t
u
Note that a sequence which is weakly convergent needs not to converge strongly.
Example 3.2.2. (cf. [267, Example 2.13]). Let H WD l 2 and x k D ek D
.ek1 ; ek2 ; : : :/, where
1 for i D k
eki WD
0 for i ¤ k.
k
Then x D 1, although fx k g1 kD0 does not contain a convergent subsequence,
p
because x k x l D 2 for all k; l, k ¤ l, i.e., fx k g1
kD0 is not a Cauchy sequence.
k
Note, however, that x * 0.
Under some additional assumptions the weak convergence yields the strong one.
Lemma 3.2.3. If x k * x 2 H and x k ! kxk, then x k ! x.
Proof. Let x k * x. It follows from the properties of the inner product that
x x k 2 D kxk2 C x k 2 2hx; x k i ! 0
as k ! 1. t
u
A Hilbert space has an important property which is expressed in the following
lemma.
Lemma 3.2.4 ([279]). If x k * y 2 H, then, for any y 0 2 H, y 0 ¤ y, the following
inequality holds
lim inf x k y 0 > lim inf x k y . (3.1)
k k
sequence is bounded, both lower limits in (3.1) are finite. The properties of the inner
product yield
3.2 Properties of the Weak Convergence 107
k
x y 0 2 D x k y C y y 0 2
2 2
D x k y C y y 0 C 2hx k y; y y 0 i
2
D x k y C ı C 2hx k y; y y 0 i.
By assumption, hx k y; y y 0 i ! 0, consequently,
2 2 2
lim inf x k y 0 D lim inf x k y C ı > lim inf x k y ,
k k k
i.e., lim infk x k y 0 > lim infk x k y . t
u
Lemma 3.2.4 was proved by Zdzisław Opial in [279, Lemma 1]. The property
described in this lemma is known under the name Opial property. A Banach space
for which the property holds is called an Opial space. Therefore, Lemma 3.2.4 can
be formulated as follows: any Hilbert space is an Opial space.
The following lemma, also proved by Opial in [279, Lemma 2], is a consequence
of the Opial property.
Lemma 3.2.5. Let T W X ! Hk be nonexpansive
and y 2 X be a weak cluster point
of a sequence fx k g1 . If T x x k ! 0, then y 2 Fix T .
kD0
k
Proof. Let x *y for a subsequence fx nk g1 k 1 k
kD0 of fx gkD0 and T x x ! 0.
nk
Suppose that T y ¤ y. Then, by the triangle inequality and by Lemma 3.2.4, we have
D lim inf kT x nk x nk C x nk T yk
k!1
lim inf.kx nk T yk kT x nk x nk k/
k!1
D lim inf kx nk T yk
k!1
to Browder [44] (see [278, Proposition 2.5]). The property says, actually, that for a
nonexpansive operator T in a Hilbert space, the operator T Id is demi-closed at 0.
Proof. (i) By the definition of the metric projection and the Fejér monotonicity of
fx k g1
kD0 , we have
d.x kC1 ; C / D x kC1 PC x kC1 x kC1 PC x k
x k PC x k D d.x k ; C /.
(ii) (cf. [22, Theorem 2.16 (iv)]) Let l k 0. By the parallelogram law with
x D PC x l x l and y D PC x k x l , the definition of the metric projection,
the convexity of C and the Fejér monotonicity of fx k g1
kD0 , we have
PC x l PC x k 2 D 2 PC x l x l 2 C 2 PC x k x l 2
2
1
4 .P
2 C x l
C PC x k
/ x l
2 2 2
2 PC x l x l C 2 PC x k x l 4 PC x l x l
2 2
D 2.PC x k x l PC x l x l /
2 2
2.PC x k x k PC x l x l /
D 2.d.x k ; C / d.x l ; C //.
By (i), the right hand side of the latter equality goes to zero as k; l ! 1.
Therefore, fPC x k g1
kD0 is a Cauchy sequence, consequently, it converges in
norm to a point z 2 C .
(iii) Because any subsequence of a Fejér monotone sequence is Fejér monotone,
we can suppose, without loss of generality, that the whole sequence converges
weakly to xN 2 H. Let x 2 C and " > 0. Denote zk D PC x k . Let k0 0 be
such that
1
maxfzk x k zk z ; kz xk zk z ; hx k x;
N z xig < "
3
for all k k0 . Then, by the characterization of the metric projection and the
Cauchy–Schwarz inequality, we have
110 3 Convergence of Iterative Methods
We have
2 2
lim x k 2hx k ; zi D lim x k z kzk2 D ˛.z/.
k k
3.4 Asymptotically Regular Operators 111
Let fx mk g1 nk 1 k 1
kD0 and fx gkD0 be two subsequences of fx gkD0 such that x
mk
* z0 2
00
C and x * z 2 C as k ! 1. Then
nk
h 2 2 i
2 limhx k ; z0 z00 i D lim x k 2hxk ; z00 i x k 2hxk ; z0 i
k k
D ˛.z00 / ˛.z0 /.
But
limhx mk ; z0 z00 i D hz0 ; z0 z00 i
k
and
limhx nk ; z0 z00 i D hz00 ; z0 z00 i.
k
x 0 ¤ x , is a cluster point of fx k g1 mk 1
kD0 and that a subsequence fx gkD0 converges to
x . Let " WD 2 kx x k > 0, k0 be such that kx x k < " and kx nk x k < ",
0 1 0 mk 0
for all k k0 and let mk > nk0 . By the triangle inequality and by the Fejér
monotonicity of fx k g1 kD0 with respect to C , we obtain
2" D x 0 x x 0 x mk C kx mk x k < 2",
a contradiction. Therefore, x k ! x . t
u
for all x 2 X .
112 3 Convergence of Iterative Methods
X
m
R WD PX T D PX .Id C !i .Ti Id//
i D1
is asymptotically regular.
Pm
Proof. The operator T WD i D1 !i Ti is firmly nonexpansive (see Corol-
lary 2.2.20). Therefore, the asymptotic regularity of R follows from Corol-
lary 3.4.5. t
u
Corollary 3.4.8. Let Ti W X ! X , i 2 I , be strictly relaxed firmly nonexpansive
and the composition T WD T1 T2 : : : Tm have a fixed point. Then T is asymptotically
regular.
Proof. By Theorem 2.2.42, the operator T is strictly relaxed firmly nonexpansive.
Therefore, the asymptotic regularity of T follows from Corollary 3.4.6. t
u
The following result is due to Bruck and Reich (see [51, Corollary 1.1]).
Theorem 3.4.9. Let T W X ! X be a strongly nonexpansive operator with a fixed
point. Then T is asymptotically regular.
Proof. The operator T is quasi-nonexpansive, because a strongly nonexpansive
operator is nonexpansive and a nonexpansive operator with a fixed point is quasi-
nonexpansive (see Lemma 2.1.20). Let x 0 D x 2 X , x k D T k x and z 2 Fix T be
such that
114 3 Convergence of Iterative Methods
lim x k z D inf lim x k z .
k z2Fix T k
The
existence
and uniqueness of z follows from Theorem 3.3.2. Since the sequence
fx k z g1
kD0 is decreasing, it converges and the sequence fx k 1
g kD0 is bounded.
We have
k
x z T x k z D x k z x kC1 z ! 0.
The equality above, the strong nonexpansivity of T and the equality T z D z yield
now k
T x x k D .T x k T z / .x k z / ! 0.
Now, it follows from the demi-closedness principle (Lemma 3.2.5) that y 2 Fix U .
Theorem 3.3.2 yields that y D y ; where y is the unique fixed point of U
satisfying equality (3.3). t
u
If X H is a subset of a finite dimensional Hilbert space, then theorem holds by
weaker assumptions. In this case it suffices to suppose the closedness of U Id at 0
and quasi nonexpansivity of U instead of its nonexpansivity. The following theorem
holds.
3.5 Opial’s Theorem and Its Consequences 115
consequently, fx k g1
kD0 is bounded. Let z
2 X be an arbitrary cluster point
of fx gkD0 and fx gkD0 be a subsequence which converges to z . Since U is
k 1 nk 1
asymptotically regular,
kUx nk x nk k D U nk C1 x U nk x ! 0.
x kC1 D Uk x k , (3.5)
or, equivalently,
lim Uk x k x k D 0,
k
then fx k g1
kD0 converges weakly to a point z 2 Fix S .
(ii) If H is finite dimensional and the sequence of operators fUk g1 kD0 has the
property
lim Uk x k x k D 0 H) lim inf S x k x k D 0, (3.7)
k k
then fx k g1
kD0 converges to a point z 2 FixS .
3.6 Generalization of Opial’s Theorem 117
Opial’s Theorem is, actually, a corollary of Theorem 3.6.2 (i). Indeed, suppose
that the assumptions of Theorem 3.5.1 are satisfied. Then U is quasi-nonexpansive
(see Lemma 2.1.20) and U Id is demi-closed at 0 (see Lemma 3.2.5). We see that
all assumptions of Theorem 3.6.2 are satisfied for a constant sequence of operators
Uk D U , k 0. Therefore, for an arbitrary x 2 C , the sequence fU k xg1 kD0
converges weakly to a point x 2 Fix U .
Proof of Theorem 3.6.2. (cf. [70, Theorem 9.9]) Let x 2 X , z 2 Fix S and the
sequence fx k g1
kD0 be generated by recurrence (3.5). Since Uk is quasi-nonexpansive
and Fix Uk Fix S , we have
kC1
x z D Uk x k z x k z ,
kUk x xk ˛ kS x xk (3.8)
holds for all x 2 X , for all k and for some ˛ > 0. Condition (3.7) is satisfied if,
e.g., inequality (3.8) holds for all x 2 X , for infinitely many k and for some ˛ > 0.
118 3 Convergence of Iterative Methods
Remark 3.6.4. It follows from the proof that Theorem 3.6.2 remains true if
we replace the assumption that fUk g1 kD0 is asymptotically regular and the
assumption
(3.6) in
case (i) or (3.7) in case (ii)
by a weaker
assumption
limk!1 S x k x k D 0 in case (i) or lim infk!1 S x k x k D 0 in case (ii),
respectively. The formulation presented in Theorem 3.6.2 is preferred, because in
applications, the operators Uk are often relaxed cutters with relaxation parameters
guaranteeing the asymptotic regularity of fUk g1kD0 . Furthermore, various practical
algorithms which apply relaxed cutters have properties which yield (3.6), (3.7) or
some related conditions.
T1
Proof. Let C D kD0 Fix Tk and z 2 C . Denote Uk WD PX .Id C k .Tk
Id//. Inequality (3.10) and the strong
k quasi nonexpansivity of Uk follow from
Corollary 2.2.25. Consequently, x z converges as a bounded and monotone
sequence and lim .2 / Tk x k x k D 0. Since lim infk k .2 k / > 0, we
k k k
have limk Tk x k x k D 0. The rest part of the theorem can be proved by similar
arguments as in the proof of Theorem 3.6.2. t
u
A special case of Corollary 3.7.1 (ii) with X D H and Tk D S for all k 0
can be found in [253, Theorem 1]. Note that condition (3.11) holds in this case
automatically.
Bauschke and Combettes proved a theorem [24, Theorem 2.9 (i)] which is related
to Corollary 3.7.1 (ii). They supposed that X D H, k D 2 Œ1; 2/ and, instead
T1
of (3.11), supposed that all cluster points of fx k g1
kD0 belong to kD0 Fix Tk (see
also [124, Theorem 2.6 (iii)] and [352, Theorems 2.1–2.3] for related results).
Remark 3.7.2. If int Fix S ¤ ;, then the range of the relaxation parameters in
Corollary 3.7.1 (ii) and (iii) can be extended. By Proposition 2.1.41, for any
z 2 int Fix S there is " > 0 such that
kC1 2 2 2
x z x k z k .2 C " k / Tk x k x k . (3.13)
If lim infk k > 0 and lim sup k 2, then, similarly, as in in the proof of
Corollary 3.7.1, we obtain limk Tk x k x k D 0, because
Therefore, if int Fix T ¤ ;, then the results of Corollary 3.7.1 (ii) and (iii)
remain true if we suppose that lim infk k > 0 and lim sup k 2 instead of
lim infk k .2 k / > 0. This explains the convergence of sequences generated by
several projection methods which apply the extended range of relaxation parameters
(see, e.g., [54, 266]).
If we iterate inequality (3.10) k-times, we obtain for an arbitrary z 2 Fix S the
following inequality
kC1 2 2 Xk
2
x z x 0 z l .2 l / Tl x l x l . (3.14)
lD0
X
k
2
d 2 .x kC1 ; Fix S / R2 l .2 l / Tl x l x l ,
lD0
lim inf k .2 k / > 0 ” .lim inf k > 0 and lim sup k < 2/.
k k
x kC1 D PX .x k C k .T x k x k //,
where x 0 2 X and k 2 Œ0; 2. If lim inf k .2 k / > 0, then x k converges weakly
to a fixed point of T .
Proof. Denote Tk D T , for all k 0, and S D T . Then implication (3.11) is
obvious. Therefore, the corollary follows from Corollary 3.7.1. u
t
A related result to Corollary 3.7.3 was obtained by Reich [294, Theorem 2]
for a uniformly convex Banach space with a Frechét differentiable norm, where
it
Pis supposed (in an equivalent form) that T is firmly nonexpansive and that
1
kD0 k .2 k / D C1 instead of lim inf k .2 k / > 0.
If we take k D 2 .0; 2/ and X D H in Corollary 3.7.3, then we obtain
Corollary 3.5.3.
Denote Uk WD PX .Id C k .Sk Id// for an operator Sk W X ! H and for
k 2 .0; 2, k 0.
Corollary 3.7.4. Let U W X ! H be an operator with a fixed point and such that
U Id is demi-closed at 0, x 0 2 X and a sequence fx k g1 kD0 be generated by the
recurrence
x kC1 D PX .x k C k .Sk x k x k //, (3.15)
where k 2 .0; 2, lim supk k < 2, and T fSk g1
kD0 is a sequence of firmly
nonexpansive operators Sk W X ! H such that 1
kD0 Fix .PX Sk / Fix U , k 0.
Uk x
Tk x
X
x
4
Uk x k x k D Tk x k x k ,
4 k
where x 0 2 X , lim infk k .2 k / > 0 and the step size function W X ! .0; C1/
satisfies the following condition
hz x; Ux xi
˛ .x/ (3.19)
kUx xk2
for all x 2 X , for all z 2 Fix U and for some ˛ > 0. Then
kC1 2 2 2
x z x k z k .2 k / 2 .x k / Ux k x k (3.20)
2 2 k
x k z U; x k x k 2
k
k
k 2 2
D x z k .2 k / 2 .x k / Ux k x k
2 2
x k z "2 ˛ 2 Ux k x k
kC1 2 2 Xk
2
x z x 0 z l .2 l / 2 .x l / Ux l x l (3.21)
lD0
X
k
2
d 2 .x kC1 ; Fix U / R2 l .2 l / 2 .x l / Ux l x l , (3.22)
lD0
Theorem 3.6.2 as well as other results presented in the last two sections, yield
only the weak convergence of a sequence fx k g1 kD0 generated by a sequence
of asymptotically regular and quasi-nonexpansive operators to a fixed point of
a nonexpansive operator. If H is finite dimensional, the weak convergence is
equivalent to the strong convergence. Under some assumptions on the structure of
the convex minimization problem we can obtain, however, the strong convergence
in any Hilbert space.
The following theorem is due to Bauschke and Borwein (see [22, Theorem 2.16
(iii)]).
Theorem 3.8.1. Let fx k g1
kD0 H be Fejér monotone with respect to a subset
C H. If int C ¤ ;, then fx k g1
kD0 converges strongly.
124 3 Convergence of Iterative Methods
Proof. Let z 2 int C and " > 0 be such that B.z; "/ C . Denote d k D
"
.x k x kC1 /. Then, obviously, we have zk D z C d k 2 C . By the Fejér
kx k x kC1 k
monotonicity of fx k g1 kD0 with respect to C ,
kC1 2 2
x z C d k 2hx kC1 z; d k i
2
D x kC1 z d k
2 2 2
D x kC1 zk x k zk D x k z d k
2 2
D x k z C d k 2hx k z; d k i,
i.e.,
kC1 2 2
x z x k z 2hx k x kC1 ; d k i
2
D x k z 2" x k x kC1
and
k
x x kC1 1 x k z2 x kC1 z2 .
2"
Consequently,
k m1 X
x x kCm x kCl x kClC1 1 x k z2 x kCm z2
2"
lD0
k(3.23)
for all m 1. By the Fejér monotonicity of fx k g1 with respect to C , x z
kD0
converges and inequalities (3.23) yield that fx k g1
kD0 is a Cauchy sequence, therefore
it converges strongly. t
u
Before we formulate our next result, we prove some auxiliary lemmas.
T
Lemma 3.8.2. Let C WD i 2J H .ai ; ˇi /, where ai 2 H and ˇi 2 R, i 2 J , and
fx k g1kD0 be Fejér monotone with respect to C . Then x 2 x C Linfai ; i 2 J g for
k 0
all k 0.
P
Proof. Denote V WD Linfai ; i 2 J g and fix k 0. Since x kC1 D x 0 C klD0 .x lC1
x l /, it suffices to prove that x lC1 x l 2 V for all l 2 f0; 1; : : : ; kg. We leave it to
the reader to check that the Fejér monotonicity of fx k g1 kD0 with respect to C yields
that
1 lC1
w x x l ; x lC1 x l 0 (3.24)
2
for any w 2 C (see equivalence (2.11)). Let l 2 f0; 1; : : : ; kg, z 2 C , and ul 2 V
and vl 2 V ? be such that x lC1 x l D ul C vl . Suppose that x lC1 x l … V . Then
vl ¤ 0. For any ˛ 2 R and i 2 J , we have
3.8 Strong Convergence of Fejér Monotone Sequences 125
we obtain
1
hz C ˛l vl .x lC1 x l /; x lC1 x l i
2
1
D hz; x lC1 x l i C h˛l vl .x lC1 x l /; x lC1 x l i
2
1
D hz; x lC1 x l i C h˛l vl .ul C vl /; ul C vl i
2
1 2 1 2
D hz; x lC1 x l i C .˛l / vl ul < 0,
2 2
to z.
Proof. Let x k * z 2 X . Denote y k WD x k x 0 . It is clear that y k 2 V WD X x 0 ,
y k * z x 0 and V is a finite dimensional subspace of H. Therefore, y k ! z x 0
and limk x k D limk y k C x 0 D z. t
u
A special case of the following Theorem was proved by Gurin et al. (see [196,
Theorem 1 (d)]).
Theorem 3.8.4. Let fx k g1 kD0 H be Fejér monotone with respect to a subset C
being the intersection of finitely many half-spaces. If fx k g1 kD0 converges weakly,
then it converges strongly.
T
Proof. Let C WD i 2I H .ai ; ˇi /, where ai 2 H and ˇi 2 R, i 2 I WD
f1; 2; : : : ; mg. Denote X D x 0 C Linfai ; i 2 I g. Then X is a finite dimensional
affine subspace. By Lemma 3.8.2, fx k g1 k 1
kD0 X . Suppose that fx gkD0 converges
k 1
weakly. Then the strong convergence of fx gkD0 to a point of X follows from
Lemma 3.8.3. t
u
Corollary 3.8.5. Suppose that a sequence fx k g1
kD0 generated by recurrence (3.5)
satisfies the assumptions of Theorem 3.6.2. Furthermore, suppose that one of the
following conditions is satisfied:
(i) X is a finite dimensional
T affine subspace of H,
(ii) The subset F WD T1 kD0 Fix Uk has a nonempty interior,
(iii) The subset F WD 1 kD0 Fix Uk is an intersection of finitely many half-spaces.
126 3 Convergence of Iterative Methods
U NE U contraction
U /2-AV
T - FNE U SNE
U -RFNE
FixT FixU
FixU
T 2=T =(2- )/
T cutter =(2- )/ U - AR
U -SQNE
U NE
FixU
=2 Ukx
U QNE x* FixU
3.10 Exercises
As we have mentioned in Chap. 2, the algorithms (or methods) for solving convex
optimization problems are defined by algorithmic operators. Usually, particular
iterations of these algorithms have the form
x C D Ux,
where x is the current approximation of a solution and x C is a next approximation
(also called an actualization or an update). If an operator U describing this actualiza-
tion belongs to the class of strongly quasi-nonexpansive operators (or, equivalently,
to the class of relaxed cutters), then we call them algorithmic projection operators.
The name projection operator can be explained by the following property of a cutter
T W H ! H: For any x … Fix T , T x is the metric projection of x onto a hyperplane
H.x T x; hT x; x T xi/ D fy 2 H W hy T x; x T xi D 0g
A. Cegielski, Iterative Methods for Fixed Point Problems in Hilbert Spaces, 129
Lecture Notes in Mathematics 2057, DOI 10.1007/978-3-642-30901-4 4,
© Springer-Verlag Berlin Heidelberg 2012
130 4 Algorithmic Projection Operators
PH(a,β ) x
ha; xi ˇ
hx y; z yi D .ha; zi ha; yi/ D 0.
kak2
Now, the characterization of the metric projection (see Theorem 1.2.4) yields y D
PH.a;ˇ/ x (Fig. 4.1). We can also write
.x/
PH.a;ˇ/ x D x a;
kak2
where .x/ D ha; xi ˇ is the residuum of the equality ha; xi D ˇ at the point
x 2 H.
Example 4.1.1. In a Euclidean space (Rn with the standard inner product) we have
where a 2 Rn and ˇ 2 R. In this case equality (4.1) can be written in the form
a> x ˇ
PH.a;ˇ/ x D x a
kak2
or
.x/
PH.a;ˇ/ x D x a,
kak2
where .x/ D a> x ˇ.
Example 4.1.2. Denote by PCG x the metric projection of x 2 Rn onto a closed
convex subset C of Rn equipped with the inner product h; iG induced by a positive
definite matrix G of type n n, defined by hx; yiG WD x > Gy. We call PCG x an
G
oblique projection. We calculate the oblique projection PH.a;ˇ/ x for x 2 Rn , where
4.1 Examples of Metric Projections 131
H a,β
x
G
PH a,β x
(see Fig. 4.2). By the properties of the metric projection, this problem has a unique
solution. Since the minimization problem (4.2) is convex, the solution z D PH.a;ˇ/
G
x
can be derived from the Karush–Kuhn–Tucker conditions (see Theorem 1.3.7). The
Lagrange function L W Rn R ! R has the form
1
L.z; / D kz xk2G C .ha; zi ˇ/
2
G.z x/ C a D 0
ha; zi D ˇ.
Consequently,
hG 1 a; G.z x/ C ai D 0
and
ha; z xi ha; xi ˇ
D 1
D .
ha; G ai kak2G 1
If we put the computed Lagrange multiplier into the first equation of the KKT-
system, we obtain
ha; xi ˇ 1
P G .x/ D x G a. (4.3)
kak2G 1
Example 4.1.3. In H D L2 .Œ˛; ˇ/ with the inner product defined by
Z ˇ
hf; gi WD f .x/g.x/dx
˛
132 4 Algorithmic Projection Operators
H(a2; ¯2 )
PH x
H (a1; ¯1 )
H
we have Z ˇ
H.g; / D fh 2 L2 .Œ˛; ˇ/ W g.x/h.x/dx D g,
˛
H x D fy 2 Rn W A.y C x/ D bg.
H−(a,β)
PH − (a,β) x
ha; xi ˇ
hx y; z yi D .ha; zi ha; yi/ 0.
kak2
Now, it follows from the characterization of the metric projection (see Theorem
1.2.4) that y D PH .a;ˇ/ x. We can also write
.ha; xi ˇ/C
PH .a;ˇ/ x D x a (4.7)
kak2
(Fig. 4.4) or
C .x/
PH .a;ˇ/ x D x a,
kak2
where C .x/ D .ha; xi ˇ/C .
C WD fx 2 H W ˇ1 ha; xi ˇ2 g,
134 4 Algorithmic Projection Operators
z = PC z
PC x PC y
C
y
Let RnC be the nonnegative orthant in the Euclidean space Rn . Obviously, RnC is
closed and convex. We show that PRnC .x/ D xC . Let y WD xC . It is clear that y 2
RnC . Let z 2 RnC be arbitrary. Since x D xC x , hx ; zi 0 and hxC ; x i D 0,
we have
hx y; z yi D hx xC ; z xC i
D hx ; z xC i
D hx ; zi C hx ; xC i 0.
PRnC x D xC
w = P R +n w
PR+n y P R +n z
y z
PLC f D fC
2
and
PL
2
f D f .
Example 4.1.4. Let Q WD fu 2 Rm W u bg, where b 2 Rm . We have Q D Rm Cb
or, equivalently, Rm
D Q b. Applying equalities (1.12) and (4.8), we obtain
PQ .y/ D PRm .y b/ C b
D .y b/ C b
D y .y b/C .
®1 ¯1
g
PC f
a b
It follows from the above equality and and from Lemma 1.2.8 that
PC x D maxfminfx; bg; ag
(see Fig. 4.7). This formula can also be used when ˛j D 1 or ˇj D C1 for
some j , j D 1; 2; : : : ; n.
We obtain a similar formula in H D L2 WD L2 .Œa; b/ for the metric projection
of f 2 L2 onto the subset C WD ff 2 L2 W g f hg, where g; h 2 L2 and
g h. Obviously, C is closed and convex. It follows from the characterization of
the metric projection (see Theorem 1.2.4) that
i.e., 8
< g.x/ if f .x/ < g.x/
.PC f /.x/ D f .x/ if g.x/ f .x/ h.x/
:
h.x/ if f .x/ > h.x/
(see Fig. 4.8).
The formula above can be extended to the case C D ff 2 L2 W g f h
on S g, where S Œa; b is a measurable subset and g h on S . We leave it to the
reader to check that
8
< g.x/ if f .x/ < g.x/ and x 2 S
.PC f /.x/ D f .x/ if g.x/ f .x/ h.x/ or x … S
:
h.x/ if f .x/ > h.x/ and x 2 S .
4.1 Examples of Metric Projections 137
PB(z,ρ) x
y = PB(z,ρ) y
z
B(z,ρ)
ρ
PC x
y = PC y
z
J (D; z ; ½)
1
L.y; / D ky xk2 C Œ.y z/> D.y z/ .
2
y x C 2D.y z/ D0
.y z/> D.y z/
(4.12)
0
Œ.y z/> D.y z/ D 0.
by the complementary condition (fourth equality in (4.12)). The first equality in the
KKT-system gives
y D .Id C 2D/1 .x C 2Dz/
By the existence and uniqueness of the metric projection PC x, equation (4.14) has
a unique solution . Summarizing, we obtain
x if .x z/> D.x z/
PC x D
.Id C 2D/ .x C 2Dz/ if .x z/> D.x z/ > ;
1
Since
2 3
.1 C 2!1 /1 0 ::: 0
6 0 .1 C 2! 2/
1
::: 0 7
.Id C 2W /1 D6
4
7,
5
::: ::: ::: :::
1
0 0 : : : .1 C 2!n /
X
n
!j j2
D1 (4.16)
j D1
.1 C 2!j /2
(note that this equation is considered only in the case x … J.W; 0; 1/, i.e.,
P n 2
j D1 !j j > 1). As mentioned before, this equation has a unique solution > 0,
which can be computed, e.g., by the Newton method with the starting point 0 D 0.
P !j 2
Note that the function h W RC ! R, h./ D nj D1 .1C2!j /2 is convex (h00 ./ > 0
j
for all 0), h.0/ > 1 and lim!C1 h./ D 0. Therefore, the Newton method
generates an increasing sequence k which converges to a unique solution
of (4.16). The details can be found in [316, Theorem 3.4-2].
140 4 Algorithmic Projection Operators
is called an ice cream cone with axis a and angle ˛. Without loss of generality we
suppose that kak D 1. Then the ice cream cone has the form
where D cos ˛ 2 Œ0; 1. This subset is closed and convex as a sublevel set S.f; 0/
of a continuous convex function f W H ! R, f .x/ WD kxk ha; xi. It is clear
that C.a; ˛/ is a cone. The polar cone to an ice cream cone is again an ice cream
cone and has the form
.C.a; ˛// D C.a; ˛/ (4.17)
2
ha; u xi
D sin ˛ (4.18)
ku xk
4.1 Examples of Metric Projections 141
π
a 2 ® u a
®
xy C (a; ®)
y
or, equivalently,
q
kxk2 ha; xi2
D cos ˛. (4.19)
ku xk
and
hx; x ui kxk2 ha; xi
D 2
D . (4.21)
ku xk ka xk2
Consequently,
and
q
kxk2 ha; xi2
kyk D kx C .u x/k D . (4.23)
ka xk
By (4.19), (4.22) and (4.23), we obtain hy; ai D kyk cos ˛, i.e., y 2 C.a; ˛/.
Furthermore, (4.17) and (4.18) yield x y D .x u/ 2 C .a; ˛/, and we have
hx y; yi D hx u; yi D 0,
consequently,
hx y; z yi D hx y; zi 0
142 4 Algorithmic Projection Operators
for all z 2 C.a; ˛/. Now, the characterization of the metric projection yields that
y D PC.a;˛/ x. Subsuming all cases, we obtain
8
<x if ha; xi kxk cos ˛
PC.a;˛/ x D 0 if ha; xi kxk sin ˛
:
x C .a x/ if kxk sin ˛ < ha; xi < kxk cos ˛,
where and are given by (4.20) and (4.21). Equivalent formulas can also be found
in [17, Theorem 3.3.6] or in [316, Sect. 3.5].
4.2 Cutters
hT x x; T x zi 0
It follows from Lemma 2.1.36 that, for T being a cutter, Fix T is closed and
convex,
T consequently, Fix T is an intersection of half-spaces, i.e., Fix T D
i 2J H .ai ; ˇi /, where ai 2 H, ai ¤ 0, ˇi 2 R, i 2 J . This fact is applied
in the theorem below which gives a necessary and sufficient condition for T to be a
cutter. Denote V WD Linfai ; i 2 J g.
Theorem 4.2.1. Let T W H ! H. For any y 2 H the following statements are
equivalent:
(i) T is a cutter.
(ii) For all x 2 H and for all w 2 .V C y/ \ Fix T , it holds
hT x x; T x wi 0 (4.24)
and
Tx x 2 V. (4.25)
Proof. Let z 2 Fix T , v? 2 V ? and ˛ 2 R. We prove that z C ˛v? 2 Fix T . Note
that hai ; v? i D 0 and hai ; zi ˇi , i 2 J , Therefore, we have
Tx Fix T
x
(ii))(i) Let x; y 2 H and z 2 Fix T . Suppose that conditions (4.24) and (4.25)
are satisfied for all w 2 .V C y/ \ Fix T . Let y; N zN 2 V and y ? ; z? 2 V ? be such
? ?
that y D yN C y and z D zN C z . It follows from the first part of the proof that
zN 2 Fix T and that zN C y ? 2 Fix T . It is clear that V C y D V C y ? , consequently,
zN C y ? 2 V C y. By (4.24) and (4.25), we have
hT x x; T x zi D hT x x; T x .Nz C y ? /i hT x x; z? y ? i 0
and T is a cutter.
(i))(ii) Suppose that T is a cutter. Let x 2 H and w 2 .V C y/ \ Fix T .
Inequality (4.24) is obvious. Suppose that (4.25) does not hold. Then T x x D
vN C v? for some vN 2 V and v? 2 V ? , v? ¤ 0. It follows from the first part of the
2
proof that w C ˛v? 2 Fix T for any ˛ 2 R. For ˛ < hT x x; T x wi= v? we
obtain
0 hT x x; T x .w C ˛v? /i D hT x x; T x wi ˛hNv C v? ; v? i
2
D hT x x; T x wi ˛ v? > 0,
hT x x; T x ui 0 (4.26)
144 4 Algorithmic Projection Operators
hT x x; v ui D 0 (4.27)
2nd Proof of Corollary 4.2.3. The sufficiency of the conditions and the necessity
of (4.26) is obvious. Suppose that T is a cutter and that
hT x x; v ui ¤ 0 (4.28)
for some x 2 H and for some v; u 2 Fix T . By the symmetry of (4.28) with respect
to u and v, it suffices to consider only the case hT x x; v ui > 0. Let w D
u C t.v u/ for t 2 R. Of course, w 2 Fix T , because Fix T is an affine subspace.
For
hT x x; T x ui
t<
hT x x; v ui
it holds
hT x x; T x wi > 0.
We have obtained a contradiction with the assumption that T is a cutter, which yields
the necessity of (4.27). t
u
x
Pf x
S (f,α)
x
Pf ,αx
S(f x, α)
and, by (4.7),
.f .x/ ˛/C
PS.fNx ;˛/ x D x gf .x/ D Pf;˛ x.
gf .x/2
because for x … S.f; ˛/ we have .f .x/ ˛/C D f .x/ ˛ and gf .x/ ¤ 0 (see
Theorem 1.3.2), and for x 2 S.f; ˛/ we have .f .x/ ˛/C D 0. Obviously, Pf;˛ x
depends on the choice of gf .x/ 2 @f .x/, x 2 H. Nevertheless, the properties of
the subgradient projection presented below hold for any choice of gf .x/ 2 @f .x/,
x 2 H. Note that Pf;˛ is well defined even if S.f; ˛/ D ;, i.e., if ˛ < infx2H f .x/
or if ˛ D infx2H f .x/ and the function f does not attain its minimum.
Lemma 4.2.5. Let Pf;˛ W H ! H be a subgradient projection relative to a convex
subdifferentiable function f by a level ˛ 2 R. If S.f; ˛/ ¤ ;, then
146 4 Algorithmic Projection Operators
Proof. The inclusion S.f; ˛/ Fix Pf;˛ is obvious. Suppose that x … S.f; ˛/ ¤ ;.
Then f .x/ > ˛ infx2H f .x/, consequently, gf .x/ ¤ 0 (see Theorem 1.3.2) and
f .x/ ˛
gf .x/ ¤ 0,
gf .x/2
.f .x k //C
D Pf x k x k ! 0. (4.32)
gf .x k /
Note that fx k g1
kD0 is bounded as a weakly convergent sequence. By the global
Lipschitz continuity of f on bounded subsets, there exists a constant > 0 such
that
f 0 .x k ; gf .x k // gf .x k /
4.3 Alternating Projection 147
and Pf z D z. t
u
hz T x; x T xi hPB x T x; x PB xi kT x PB xk2 ,
hz T x; x T xi D hz T x; x PB xi C hz T x; PB x T xi
hz T x; x PB xi
148 4 Algorithmic Projection Operators
PB x B
D hz PB x; x PB xi C hPB x T x; x PB xi
hPB x T x; x PB xi
kT x PB xk2 0,
x 2 Argmin f .x/
x2A
” PB x x D Df .x / 2 NA .x /
” x D PA PB x (4.33)
” x 2 Fix PA PB .
Dh.x; y/ D .x y; y x/
for all x 2 A and y 2 B. We leave it to the reader to check that NAB .x; y/ D
NA .x/ NB .y/. Now, using similar arguments as in the first part of the proof, we
150 4 Algorithmic Projection Operators
Fix( PA PB )
y∗
B
obtain
.x ; y / 2 Argmin h.x; y/
x2A;y2B
” .y x ; x y / D Dh.x ; y / 2 NAB .x ; y /
” y x 2 NA .x / and x y 2 NB .y /
” x D PA y and y D PB x
” x D PA PB x and y D PB x ,
The alternating projection PA PB does not need not to be firmly nonexpansive even
if A an B are closed subspaces of H (see Exercise 2.5.9). Nevertheless, if we
suppose that A is a closed affine subspace and if we restrict the operator PA PB
to the subset A, then PA PB is firmly nonexpansive. This fact is formulated in the
following theorem which is an extension of an important result of Combettes [117,
Proposition 3].
Theorem 4.3.7. Let A H be a closed affine subspace and B H be a closed
convex subset. Then, for all x; y 2 A,
and
T WD PA PB W A ! A,
is firmly nonexpansive.
Proof. Let a 2 A. Applying the property
hPA u PA v; wi D hu v; PAa wi
hT x T y; x yi
D hPA PB x PA PB y; x yi
D hPB x PB y; PAa .x y/i
D hPB x PB y; x yi
kPB x PB yk2
kPA PB x PA PB yk2 C k.PA PB x PB x/ .PA PB y PB y/k2
D kT x T yk2 C k.T x PB x/ .T y PB y/k2
kT x T yk2 C .kT x PB xk kT y PB yk/2
kT x T yk2
2
kT x T yk2 kx yk2 k.T x x/ .T y y/k2 (4.36)
2
kT x zk2 kx zk2 kT x xk2 ,
2
i.e., the operator T is -strongly quasi-nonexpansive and asymptotically regular.
Proof. By Theorem 4.3.7, the operator T WD PA PB is firmly nonexpansive.
Therefore, the corollary follows from Corollary 2.2.15 and Theorem 3.4.3. t
u
T
; .x/ D x C
.x/.T x x/, (4.37)
x kC1 D PA T
;k .x k /.
Therefore, the ability to construct such step sizes is of big interest. In the remaining
part of this section we will present various step sizes for which T
is a cutter. Recall
that Fix T
; D Fix T for all > 0 (see Remark 2.4.2 (d)). Denote
ı WD d.A; B/ D inf kx yk
x2A;y2B
and
N
ı.x/ WD kT x PB xk ,
4.3 Alternating Projection 153
T x PB x D .x PB x/ (4.38)
for some . By the characterization of the metric projection PA .PB x/, we have
.x/ WD kT x xk2
:̂ 1 if x 2 Fix T;
(4.41)
Q
where ı.x/ is a fixed element of the segment Œı; kT x PB xk (see Fig. 4.18). This
step size was proposed in [76, equality (13)].
Lemma 4.3.10. Let x 2 A and the step size
.x/ be defined by (4.41). Then
1
.x/ (4.42)
2
and the inequality is strict if T x … Fix T .
Proof. If x 2 Fix T , then
.x/ D 1. Let now x … Fix T . Denote a WD PB x x,
b WD T x x, c WD PB x T x (see Fig. 4.18) and ıQ WD ı.x/.
Q
154 4 Algorithmic Projection Operators
δ( x )
B
PB x
˛Cˇ
Let ˇ 2 R. Note that the function f W .2ˇ; C1/ ! R, f .˛/ WD ˛C2ˇ is
increasing. Therefore, if we take ˛ WD .kak kck/ and ˇ WD kak kck ha; ci in
2
1
.x/ f .˛/ f .0/ D ,
2
i.e.,
.x/ 12 . If T x … Fix T , then a and c are linearly independent by
Lemma 4.3.9. In this case, ˇ D kak kck ha; ci > 0 and
.x/ > 12 . t
u
Inequality (4.42) can be strengthened. One can prove that
1 1
.x/ ;
1 C cos ˛.x/ 2
where ˛.x/ denotes the angle between the nonzero vectors x PB x and T x PB x
(for details see [76, Lemmas 3 and 4]).
4.3 Alternating Projection 155
A
x∗
Tx
y δ
Tσ x y∗
δ B
PB x
Let x 2 A be such that T x … Fix T . One can prove that the step size
.x/ given
by (4.41) is characterized by the equality
hx C
.x/.T x x/ y; T x xi D 0, (4.43)
Q
hy PB x; x PB xi D ı.x/ kPB x xk (4.44)
hy PB x; T x PB xi D kT x PB xk2 (4.45)
Q
and ı.x/ 2 Œı; kT x PB xk (see Fig. 4.19). By Lemma 4.3.9, such a solution is
defined uniquely. For details, see [76, Lemma 6].
Theorem 4.3.11. Let A; B H be nonempty closed convex subsets, T WD PA PB ,
Fix T ¤ ; and the step size function
W A ! .0; C1/ be defined by (4.41). Then,
for any x … Fix T and z 2 Fix T ,
hz x; T x xi
.x/ (4.46)
kT x xk2
Fix T
; D Fix T
D Fix T .
hz PB x; PB x xi D hz w; PB x xi C hw PB x; PB x xi
hz w; PB x xi
kz wk kPB x xk D ı kPB x xk .
Therefore, if we apply again the characterization of the metric projection PA .PB x/,
we obtain
hz x; T x xi D hz PB x; T x xi C hPB x x; T x xi
D hz PB x; T x PB xi C hz PB x; PB x xi
ChPB x x; T x xi
kT x PB xk2 ı kPB x xk C hPB x x; T x xi.
Now, (4.46) follows from the inequality ı ı.x/Q and from the definition of the
step size given by (4.41). We see that the step size
.x/ satisfies the assumptions
of Corollary 2.4.5. Therefore, the operator T
is a cutter and the operator T
; is
2
-strongly quasi-nonexpansive. The asymptotic regularity of T
; follows from
Theorem 3.4.3. t
u
Q
If A \ B ¤ ;, then, of course, ı D 0. If we take ı.x/ WD ı D 0 in (4.41), then we
obtain
8 2
< kT x PB xk C hPB x x; T x xi
if x … A \ B
.x/ D kT x xk2 (4.47)
:
1 if x 2 A \ B.
i.e., both sides of inequality (4.48) are well defined. Inequality (4.48) can be written
in the form
kb ak2 C ha; bi kak2
kbk2 ha; bi
4.3 Alternating Projection 157
Tx
A
Tσ1 x
y
PB x Tσ 2 x
B
which is equivalent to
consequently, (4.50) is true. Suppose now that A is a closed affine subspace. Then
hc; bi D 0 (see Theorem 2.2.33 (i)), consequently, ha; bi D kbk2 and the equality
in (4.50) holds. t
u
Suppose that A \ B ¤ ;. Define the step size function
W A ! .0; C1/ by the
formula 8
< kPB x xk2
if x 2 AnB ;
.x/ D hPB x x; T x xi (4.52)
:
1 if x 2 A \ B,
where T WD PA PB is the alternating projection. The step size defined by (4.52)
was proposed by Gurin et al. in [196, Sect. 3] and was applied in an accelerated
alternating projection method of the form x kC1 D PA T
x k . Note that Lemma 4.3.12
says that in the case A \ B ¤ ;, the step size given by (4.47) is equal to at least the
step size given by (4.52). Therefore, the estimation (2.69) for T WD PA PB with the
step size
.x/ given by (4.47) is better than the one with the step size
.x/ given
by (4.52) (see Fig. 4.20). The following lemma shows that both step sizes lead to
generalized relaxations T
which are extrapolations of T .
Lemma 4.3.13. Let A \ B ¤ ; and the step size function
W A ! .0; C1/ be
defined by (4.52). Then
.x/ 1 for all x 2 A and the equality holds if and only if
PB x 2 A.
158 4 Algorithmic Projection Operators
Therefore,
.x/ 1 and the equality holds if and only if PB x 2 A. t
u
Corollary 4.3.14. Let A; B H be nonempty closed convex subsets, A \ B ¤ ;
and the step size function
W A ! .0; C1/ be defined by (4.52). Then the operator
T
W A ! H is a cutter. Consequently, for any 2 .0; 2/, the operator T
; is
2
-strongly quasi-nonexpansive.
Proof. Let x 2 A and 2 .0; 2/. It follows from Lemma 4.3.12 that
.x/ is not
greater than the step size given by (4.47). Theorem 4.3.11 yields now that
.x/
satisfies conditions of Corollary 2.4.5. Therefore, T
is a cutter and T
; is 2
-
strongly quasi-nonexpansive. t
u
Remark 4.3.15. Let x 2 AnB. Note that the step size
.x/ defined by (4.52) is the
unique solution of the equality
hx C
.x/.T x x/ PB x; PB x xi D 0
(see Fig. 4.21). This step size was proposed by Bauschke et al. in [25, Corollary
4.11] and was applied in an accelerated alternating projection method of the form
x kC1 D T
x k .
Now suppose that A and B are closed subspaces of H. In this case, the metric pro-
jections PA , PB are, actually, orthogonal projections (see Theorem 2.2.30 (i)) and
4.3 Alternating Projection 159
PB x B
PB x
B
(see Fig. 4.22). This step size was applied in [30, equality 3.1.2] in order to
accelerate globally the alternating projection method (see [30, Theorem 3.23]).
Now we consider the case in which A is a closed affine subspace and B is a closed
convex subset. We will propose a step size function
W A ! .0; C1/ which leads
to an extrapolation T
of T and we will show that the generalized relaxation T
;
which employs this step size function is strongly quasi-nonexpansive. Define the
step size function
W A ! .0; C1/ by the formula
8
ˆ Q
< 1 C .kT x PB xk ı.x//
2
if x … Fix T
.x/ WD kT x xk2 (4.55)
:̂ 1 if x 2 Fix T ,
Q
where ı.x/ 2 Œı; kT x PB xk is an upper approximation of ı WD d.A; B/ D
infx2A;y2B kx yk. A geometrical interpretation of the steps size
.x/ is presented
Q
in Fig. 4.23, where
.x/ ı.x/ < kT x PB xk, tan ˛ D kT xxk
Q . We have
kT xPB xkı.x/
kT xxk
.x/ D 1 C cot2 ˛ D 1
sin2 ˛
and kT
x xk D
.x/ kT x xk D sin2 ˛
.
160 4 Algorithmic Projection Operators
δ(x)
PB x
hz x; T x xi
.x/ kT x xk2
hz x; T x xi D kT x xk2 C hz T x; T x xi
D kT x xk2 C hT z T x; z xi kT z T xk2
kT x xk2 C .kT x PB xk kT z PB zk/2
.kT x PB xk ı/2
D .1 C / kT x xk2
kT x xk2
Q
.kT x PB xk ı.x// 2
.1 C / kT x xk2
kT x xk2
D
.x/ kT x xk2 .
Tx
RA R B x
A
T D .1 /Id C RA RB
2 2
More general results than the following one can be found in [28, Sect. 3].
Corollary 4.3.17. Let A; B H be closed convex, U WD RA RB W H ! H and
T WD 12 .U C Id/ be the averaged alternating reflection. Then
(i) A \ B Fix T ,
(ii) If int.A \ B/ ¤ ;, then A \ B D Fix T ,
(iii) If A\B ¤ ; and x 2 Fix T , then PB x D PA RB x, consequently, PB .Fix T / D
A \ B,
(iv) T is firmly nonexpansive, consequently, T is nonexpansive and 2
-strongly
quasi-nonexpansive for any 2 .0; 2/.
Proof. (i) The inclusion is obvious (see Remark 2.1.1).
(ii) Let C WD int.A \ B/ ¤ ;. It is clear that
Tλ x
PC 2 x
PC 3 x
C2 C3
X !i i
D Id C .PCi Id/
i 2I
!
X
D Id C i PCi Id
i 2I
X
D i .Id C .PCi Id//
i 2I
X
D i PCi ; .
i 2I
We see that in both cases ( D 0 and > 0), the simultaneous relaxed projection
P
i 2I !P
i PCi ;i can be presented as a relaxation U of a simultaneous
P projection
U WD i 2I i PCi or as a simultaneous relaxed projection i 2I i PCi ; with the
same relaxation parameter 2 Œ0; 2 for all projections PCi . Therefore,
P we restrict
our further analysis of simultaneous relaxed projection operators i 2I !i PCi ;i to
the case i D 2 Œ0; 2 for all P i 2 I , or, equivalently, to the relaxation T of a
simultaneous projection T WD i 2I !i PCi with a relaxation parameter 2 Œ0; 2.
Let Ci H be nonempty
P closed convex subsets, i 2 I , w D .!1 ; !2 ; : : : ; !m / 2
ri m and T WD i 2I !i PCi be a simultaneous projection. In this subsection we
present the operator T as an alternating projection in a product Hilbert space H D
Hm with the inner product h; iH W H ! R defined by
X
m
hu; viH WD !i hui ; vi i,
i D1
164 4 Algorithmic Projection Operators
C WD C1 : : : Cm
and
D WD fx D .x1 ; x2 ; : : : ; xm / 2 H W x1 D x2 D : : : D xm g.
By Lemma 1.2.8, we have
1
PD y D argmin kx yk2H
x2D 2
1X
m
D argmin !i kx yi k2 .
x2H 2 i D1
P
Since the function f W H ! R defined by f .x/ WD 12 m 2
i D1 !i kx yi k is
differentiable and convex, x is a minimizer of this function if and only if x is its
stationary point. After differentiation we obtain the following condition
X
m
!i .x yi / D 0.
i D1
Pm Pm
By the equality i D1 !i D 1, we obtain x D i D1 !i yi , i.e.,
X
m X
m
PD y D . !i yi ; : : : ; !i yi /.
i D1 i D1
X
m X
m
Tx D . !i PCi x; : : : ; !i PCi x/ D .T x; : : : ; T x/.
i D1 i D1
T x D x C .PD PC x x/
X
m X
m
D. !i PCi ; x; : : : ; !i PCi ; x/
i D1 i D1
D .T x; T x; : : : ; T x/ D PD PC; x,
2
kT x zk2 kx zk2 kT x xk2 .
166 4 Algorithmic Projection Operators
Now inequality (4.58) follows from the obvious equality T x x D .T x x/.
The asymptotic regularity of T follows from Theorem 3.4.3. t
u
T
If i 2I Ci ¤ ;, then inequality (4.58) can be strengthened. In the following
theorem we allow that the weights !i , i 2 I , can depend on x 2 H.
Theorem 4.4.5. Let Ci H be nonempty closed convex subsets, i 2 I , C WD
T
Pi 2I Ci ¤ ;, w W H ! m be an appropriate weight function and T WD
i 2I !i PCi . Then
X
kT x zk2 kx zk2 .2 / !i .x/ kPCi x xk2 , (4.59)
i 2I
(see Lemma 2.2.27) Therefore, the sufficient and necessary optimality condition
(see Corollary 1.3.3) yields that z 2 Argminx2H f .x/ if and only if Df .z/ D
z T z D 0, i.e., z 2 Fix T . Now suppose that a subset Ci is bounded for some
i 2 I and that !i > 0. Then the function f is coercive. By Corollary 1.1.53,
Argminx2H f .x/ ¤ ; and Fix T ¤ ;. t
u
The proximity function f W H ! R defined by (4.57) has the following nice
property (cf. [215, Lemma 2]).
P
Lemma 4.4.7. Let T WD i 2I !i PCi , where w 2 m . Then, for any x 2 H, it
holds
1
f .T x/ f .x/ kT x xk2 .
2
Proof. Let x 2 H. It follows from the definition of the metric projection that
kPCi T x T xk kPCi x T xk. Therefore, the properties of the inner product
yield
1X 1X
f .T x/ D !i kPCi T x T xk2 !i kPCi x T xk2
2 i 2I 2 i 2I
1X 1X X
D !i kPCi x xk2 C !i kT x xk2 !i hPCi x x; T x xi
2 i 2I 2 i 2I i 2I
1X 1X
D !i kPCi x xk2 C !i kT x xk2 kT x xk2
2 i 2I 2 i 2I
1
D f .x/ kT x xk2
2
which completes the proof. t
u
Let x 2 H. By the equality Df .x/ D x T x, Lemma 4.4.7 can be written in the
form
f .x Df .x// f .x/ kDf .x/k2 .
The equality above is related to the efficiency of a minimization method defined by
the recurrence x kC1 D x k Df .x k / (see [178, Definition 4.5]). It is known that in
the case H D Rn the sequence fx k g1 kD0 generated by this recurrence converges to a
minimizer of f (see [178, Theorem 4.6]). P
It is clear that for aTsimultaneous projection T WD i 2I !i PCi the following
inclusion is true C WD i 2I Ci Fix T . If C ¤ ; and w 2 ri m , then C D Fix T
168 4 Algorithmic Projection Operators
(see Theorem 2.1.14) and, for the proximity function f W H ! R defined by (4.57)
it holds that minx2H f .x/ D 0.
1 X .hai ; xi ˇi /2
f .x/ D !i . (4.61)
2 i 2I kai k2
and
1
f .x/ D .Ax b/> D.Ax b/, (4.64)
2
where D WD diag. !1 2 ; : : : ; !m 2 / We can suppose without loss of generality that
ka1 k kam k
w 2 ri m . In this case, the proximity function f has the form
1
f .x/ D kAx bk2D , (4.65)
2
1
where kukD D .u> Du/ 2 denotes the norm of the vector u 2 Rm , induced by
the positive definite matrix D. The operator T is firmly nonexpansive and its
4.4 Simultaneous Projection 169
relaxation T is 2
-strongly quasi-nonexpansive and asymptotically regular (see
Corollary 4.4.4), and Fix T D Argminx2H f .x/ (see Theorem 4.4.6). Furthermore,
both subsets are nonempty.
Theorem 4.4.8. Let a function f W H ! R be defined by (4.61). Then
Argminx2H f .x/ ¤ ;:
Proof. The minimization of the proximity function f W H ! R given by (4.61) is
equivalent to the following minimization problem
p .hai ; xi ˇi /
X WD f.x; z/ 2 H Rm W i D !i ; i D 1; 2; : : : ; mg
kai k
1 X Œ.hai ; xi ˇi /C 2
f .x/ D !i (4.68)
2 i 2I kai k2
and
1
f .x/ D .Ax b/>
C D.Ax b/C , (4.71)
2
where D WD diag. !1 2 ; : : : ; !m 2 /. We can suppose without loss of generality that
ka1 k kam k
w 2 ri m . In this case, the proximity function f has the form
1
f .x/ D k.Ax b/C k2D . (4.72)
2
The remaining part of the proof is similar to the proof of Theorem 4.4.8 with X
replaced by
p .hai ; xi ˇi /
X WD f.x; z/ 2 H Rm W i !i ; i 0; i D 1; 2; : : : ; mg.
kai k
De Pierro and Iusem proved a property which yields Theorem 4.4.9 in case H D Rn
(see [137, Proposition 13]).
x C2
Tx
C3
C2
x = Tx
C3
Corollary 4.5.4. If at least one of the subsets Ci , i 2 I , is bounded, then the cyclic
projection T WD PCm PCm1 : : : PC1 W Cm ! Cm has a fixed point z 2 Cm .
Proof. Since the metric projection is nonexpansive and PCi .X / D Ci , i 2 I , the
corollary follows from Theorem 2.1.13. t
u
A fixed point of a cyclic projection is illustrated in Fig. 4.27.
Let Ci H be closed half-spaces, i.e., Ci WD H .ai ; ˇi / D fx 2 H W hai ; xi
ˇi g, where ai 2 H, ai ¤ 0 and ˇi 2 R, i 2 I . The following theorem was proved
in [137, Lemma 1 and Proposition 13].
Theorem 4.5.5. If Ci Rn are half-spaces, i 2 I , then the cyclic projection
T WD PCm : : : PCm has a fixed point.
C3
Tx
Corollary 4.5.7. Let i 2 .0; 2/, i 2 I , and T WDT PCm ;m PCm1 ;m1 : : : PC1 ;1 be
the operator of cyclic relaxed projection. If C WD i 2I Ci ¤ ;, then Fix T D C .
Proof. It follows from Theorem 2.2.21 (i) and from Remark 2.1.4 that Fix PCi ;i D
Ci , i 2 I . The relaxed metric projection PCi ;i is strictly quasi-nonexpansive for
i 2 .0; 2/, i 2 I , (see Corollary 2.2.23 (iii)). Therefore, the claim follows from
Theorem 2.1.26. t
u
Corollary 4.5.8. Let i 2 Œ0; 2, i 2 I . The relaxed cyclic projection T WD
PCm ;m PCm1;m1 : : : PC1 ;1 is -relaxed firmly nonexpansive, where
2mmax
D
.m 1/max C 2
X
m
S WD !i Si , (4.75)
i D1
where Si WD PCi : : : PC1 , is called the cyclic-simultaneous projection (see Fig. 4.29).
174 4 Algorithmic Projection Operators
S3x
C3
P WD Id C .P Id/, where 2 .0; 2/, and RK WD .PK /2 D 2PK Id, i.e., P is
a relaxation of the cyclic projection P and RK is the reflection operator onto K.
Definition 4.5.11. The operator RK P is called a projection-reflection and the
operator RK P , where 2 .0; 2/ is called a relaxed projection-reflection (see
Fig. 4.30). The operator PRK is called a reflection-projection and the operator
P RK , where 2 .0; 2/ is called a reflection-relaxed projection.
The following result can be found in [31, Lemma 2.1 (v)].
Lemma 4.5.12. If K H is a closed convex and obtuse cone, then RK x 2 K for
any x 2 H.
4.5 Cyclic Projection 175
RK PC 1 x
PC1 x
C
C1
(see Theorem 2.1.28). Now we see that all assumptions of Theorem 2.1.51 are
satisfied for S D RK . Therefore, the operator RK P jK is ˇ-strongly quasi-
nonexpansive, where
2 2 ˛ .1 /m C 1
ˇD D D
˛ m
which completes the proof. t
u
U WD Um Um1 : : : U1
is called a cyclic cutter.
Note that the cyclic cutter U WD Um Um1 : : : U1 doe not need to be a cutter even
if Ui , i 2 I , have a common fixed point (see Example 2.1.53).
Corollary 4.5.15. Let Ui W H ! H be cutters with a common T fixed point, i 2 I ,
and U WD Um Um1 : : : U1 be a cyclic cutter. Then Fix U D i 2I Fix Ui and U is
1
m -strongly quasi-nonexpansive. Consequently, U is asymptotically regular.
Ax
Tx
s= PQAx-Ax
x+A* s
Q
A*
The Landweber operator is closely related to a method for the problem: find x 2 H1
such that Ax 2 Q. The method was proposed by Landweber for approximating
least-squares solution of a first kind integral equation [240] (see also [37, Sect. 6.1]).
Note that kAk ¤ 0, because A is a nonzero operator, consequently, the operator
T is well defined. We can also take max .A A/ or max .AA / instead of kAk2 in
Definition 4.6.1 if A is a compact linear operator (see Theorem 1.1.27). Define a
proximity function f W H1 ! RC by
1
PQ .Ax/ Ax 2
f .x/ WD (4.78)
2
Proof. The function f is differentiable and Df .x/ D A .Ax PQ .Ax//. It follows
from the necessary and sufficient optimality condition (see Corollary 1.3.3) that
z 2 Argminx2H1 f .x/ if and only if A .Az PQ .Az// D 0, i.e., T z D z. t
u
Theorem 4.6.3. The Landweber operator is firmly nonexpansive.
Proof. Since PQ is firmly nonexpansive (see Theorem 2.2.21 (iii)), the implication
(i))(iv) in Theorem 2.2.10 yields that the operator Id PQ is firmly nonexpansive,
i.e.,
2
h.u PQ u/ .v PQ v/; u vi .u PQ u/ .v PQ v/
178 4 Algorithmic Projection Operators
hG.x/ G.y/; x yi
1
D 2
hA .Id PQ /Ax A .Id PQ /Ay; x yi
kAk
1
D h.Id PQ /Ax .Id PQ /Ay; Ax Ayi
kAk2
1
.Id PQ /Ax .Id PQ /Ay 2
2
kAk
kA k2
.Id PQ /Ax .Id PQ /Ay 2
D 4
kAk
1
A .Id PQ /Ax A .Id PQ /Ay 2
4
kAk
2
1 1
D
A .Id PQ /Ax A .Id PQ /Ay
kAk 2
kAk 2
D kG.x/ G.y/k2 ,
1
Tx D x A> .Ax b/C . (4.80)
max .A> A/
1
Tx D x A> D.Ax b/ (4.82)
max .A> DA/
and
1
Tx D x A> D.Ax b/C , (4.83)
max .A> DA/
respectively. We call the operators T given by (4.82) and (4.83), the Landweber
operators related to the matrix D. The corresponding proximity functions obtain
the form
1 1
2
1
f .x/ D D 2 .Ax b/ D kAx bk2D
2 2
and
1
f .x/ D k.Ax b/C k2D ,
2
respectively. Note that in both cases the Landweber operators and the proximity
functions differ from the for the original systems.
Remark 4.6.5. The Landweber operators defined by (4.82) and (4.83) related to the
matrix D given by (4.81), have a nice property. One can easily show that
X
m
!i
>
A D.Ax b/ D ai .ai > x ˇi /
i D1 kai k2
180 4 Algorithmic Projection Operators
and
X
m
!i
A> D.Ax b/C D ai .ai > x ˇi /C .
i D1 kai k2
Therefore, the terms A> D.Ax b/ and A> D.Ax b/C in operators T defined
by (4.82) and (4.83), respectively, do not change if we rescale equations or
inequalities of the system. Moreover, max .A> DA/ also does not depend on the
rescaling. Therefore, the Landweber operators defined by (4.82) and (4.83) do
not change after rescaling. Note that they depend only on the vector of weights
w 2 ri m . Furthermore, if A has normalized rows and w D . m1 ; m1 ; : : : ; m1 /, then
D D diag. m1 ; m1 ; : : : ; m1 / and the Landweber operators (4.82) and (4.83) reduce
to (4.79) and (4.80), respectively.
Now we state some relationships between the Landweber operators defined
by (4.82) or (4.83), where D is given by (4.81), and the simultaneous projection
applied to the system Ax D b or Ax b, respectively, where A is a matrix of type
m n with nonzero rows ai , i 2 I , x 2 Rn and b 2 Rm .
Lemma 4.6.6. Let A WD Œa1 ; : : : ; am > be an m n matrix with nonzero rows ai ,
i 2 I , and D WD diag. !1 2 ; : : : ; !m 2 /, where w D .!1 ; : : : ; !m / 2 m . Then
ka1 k kam k
Proof. Note that max .˛1 A1 C ˛2 A2 / ˛1 max .A1 / C ˛2 max .A2 /, where A1 ; A2
are positive semi-definite matrices and ˛1 ; ˛2 0. Therefore,
X
m
!i
max .A> DA/ D max . ai ai> /
i D1 kai k2
X
m
!i
max .ai ai> /
i D1 kai k2
X
m
!i
D max .ai> ai / D 1.
i D1 kai k2
Corollary 4.6.7. The Landweber operators T defined by (4.82) and (4.83), where
D is given by (4.81), are extrapolations of the simultaneous projection operators
U defined by Ux WD x A> D.Ax b/ and by Ux WD x A> D.Ax b/C ,
respectively, i.e., T x D x C
L .Ux x/, where the Landweber step size
L 1.
Remark 4.6.8. Suppose that A is a matrix with normalized rows. If we take w D
. m1 ; m1 ; : : : ; m1 / in Lemma 4.6.6, then we obtain
One can show even more: max .A> A/ s, where s denotes the maximal number
of nonzero elements in any column of A (see [55, Proposition 4.1] or [58, Corollary
2.8]), which is of course a stronger result than Lemma 4.6.6.
where D WD diag. !1
; : : : ; !m 2 / (see (4.64)), or, equivalently, in the form
ka1 k2 kam k
1 >
f .x/ D x Gx C g > x C c; (4.85)
2
ks.x/k2
EL .x/ WD (4.88)
s.x/> A> DAs.x/
i.e., x 2 Fix T , a contradiction. Therefore, s.x/> A> DAs.x/ > 0 and the step size
kGx C gk2
EL .x/ D . (4.89)
.Gx C g/> .G 2 x C Gg/
Lemma 4.6.9. Let x … Fix T and z 2 Fix T . Then the step size
EL .x/ defined by
(4.88) satisfies the following inequalities
1 hz x; T x xi
1 >
EL .x/ (4.90)
max .A DA/ kT x xk2
Proof. Let
EL .x/ be defined by (4.88). The first inequality in (4.90) follows from
Lemma 4.6.6. We have
ks.x/k2 ks.x/k2 1
EL .x/ D D ,
s.x/> A> DAs.x/ max .A> DA/ ks.x/k2 max .A> DA/
i.e., the second inequality in (4.90) is true. Applying (4.89), we write the third
inequality in (4.90) in the form
4.6 Landweber Operator 183
and
kGx C gk2 D v> 2 v > 0,
i.e., Gx C g ¤ 0. Note that x … Fix T since x T x D s.x/ D Gx C g ¤ 0.
Similarly, we obtain
and
.x z/> .Gx C g/ D v> v.
Therefore, (4.91) can be written in equivalent forms
v> 2 v v> v
> 3
> 2
v v v v
or
.v> 2 v/2 .v> v/.v> 3 v/ 0:
We prove that the latter inequality is true. Denote I 0 WD fi 2 I W ıi > 0g. We have
XX ıj 2 2
D ıi2 ıj2 .1 /
ıi i j
i 2I 0 j 2I 0
184 4 Algorithmic Projection Operators
X ıi ıj
D ıi2 ıj2 .2 /i2 j2
ıj ıi
i;j 2I 0 ;i >j
X ıj
D ıi ıj3 . 1/2 i2 j2 0
ıi
i;j 2I 0 ;i >j
i.e., the third inequality in (4.90) is true. Note that T is a cutter as a firmly nonex-
pansive operator having a fixed point (see Theorem 2.2.5 (i) and Corollary 4.4.4),
consequently, T is oriented (see Definition 2.4.4). By Corollary 2.4.5, the operator
T
EL is a cutter. Let 2 .0; 2/. The 2
-strong quasi nonexpansivity and asymptotic
regularity of T
EL ; follow from Theorems 2.1.39 and 3.4.3. t
u
Let H1 and H2 be two Hilbert spaces. Consider the following split feasibility
problem
find x 2 C with Ax 2 Q,
if such an x exists, where C H1 and Q H2 are closed convex subsets and
A W H1 ! H2 is a bounded linear operator.
Definition 4.7.1. An operator U W H1 ! H1 defined by
1
U WD PC .Id C 2
A .PQ Id/A/ (4.92)
kAk
is called the projected Landweber operator or an oblique projection (see Fig. 4.32).
The following result is due to Byrne [55, Proposition 2.1].
Proposition 4.7.2. Let R WD PC T , where > 0, be a projected relaxation of the
Landweber operator T defined by (4.77), i.e.,
R x WD PC .x C A .PQ Id/Ax/, (4.93)
kAk2
Ax
s =PQ Ax-Ax
1
x x+ A* 2 A*s
Ux
C Q
A*
regular.
Proof. By Theorem 4.6.3 the Landweber operator T is firmly nonexpansive.
4
The 4 -relaxed firm nonexpansivity of R follows from Theorem 2.2.46 (i).
Let Fix.PC T / ¤ ;. The 2 2 -strong quasi nonexpansivity of R follows
from Theorem 2.2.46 (iii) and the asymptotic regularity of R follows from
Corollary 3.4.5. t
u
U2x
FixU2
and
kU x zk2 kx zk2 .2 / kUx xk2 (4.96)
for all 2 Œ0; 2, x 2 H and z 2 Fix U . Consequently, for all 2 .0; 2/, the
operator U is 2
-strongly quasi-nonexpansive and asymptotically regular.
X
kx zk2 C 2 !i .x/ kUi x xk2
i 2I
X
2 !i .x/ kUi x xk2
i 2I
X
D kx zk2 .2 / !i .x/ kUi x xk2 ,
i 2I
i.e., inequality (4.95) holds. Inequality (4.96) follows now from the convexity of the
function kk2 . Let 2 .0; 2/. The 2
-strong quasi nonexpansivity of U follows
from (4.96) and from the obvious equality
2
.2 / kUx xk2 D kU x xk2 .
The asymptotic regularity of U follows now from Theorem 3.4.3. t
u
U
; .x/ WD x C
.x/.Ux x/ (4.97)
T
(cf. (2.67)). Denote Ci WD Fix Ui and C WD i 2I Ci . If we suppose that the weight
function w W H ! m is appropriate, then Theorem 2.1.26 yields Fix U D C .
.x/
w .x/, (4.98)
188 4 Algorithmic Projection Operators
U2x
FixU2
where P 2
i 2I !i .x/ kUi x xk
w .x/ WD P 2 (4.99)
i 2I !i .x/Ui x x
T
for all x … i 2I Fix Ui , and the weight function w W H ! T m is appropriate.
We suppose without loss of generality that
.x/ D 1 for all T x 2 i 2I Fix Ui . The
fact that w is appropriate ensures that Fix U
; D Fix U D i 2I Fix Ui ¤ ; (see
Theorem 2.1.26 (i)), consequently, the step size
w .x/ is well defined. If
.x/ 1
for all Px 2 H, then the operator U
is an extrapolation of the simultaneous cutter
U WD i 2I !i Ui . We call an operator U
with a step size function
1 satisfying
inequality (4.98) for all x 2 H an extrapolated simultaneous cutter (ESC) (see
Fig. 4.34). The existence of an ESC follows from the convexity of the function
kk2 . It turns out that an extrapolated simultaneous cutter is a cutter.
P
Theorem 4.9.1. Let U WD i 2I !i Ui be a simultaneous cutter with an appropriate
weight function w W H ! m , U
; be a generalized relaxation of U , defined
by (4.97) with the T step size function
W H ! .0; C1/ satisfying inequality
(4.98) for x … i 2I Fix Ui , and with 2 Œ0; 2. Then the operator U
is a
cutter. Consequently, for all 2 .0; 2/, the operator U
; is 2
-strongly quasi-
nonexpansive and asymptotically regular and
kU
; x zk2 kx zk2 .2 /
2 .x/ kUx xk2 (4.100)
T
for all x 2 H and for all z 2 i 2I Fix Ui .
Proof. Since a cutter is strictly
T quasi-nonexpansive (see Theorem 2.1.39), it holds
that Fix U
; D Fix U D i 2I Fix Ui for an appropriate weight functionT wWH!
m (see Remark 2.4.2 (d) and Theorem 2.1.26). Let x 2 H and z 2 i 2I Fix Ui . If
T
x 2 i 2I Fix Ui , the inequality hz x; U
x xi kU
x xk2 is clear. Now let
T
x … i 2I Fix Ui . It follows from (2.21) that hz x; Ui x xi kUi x xk2 , i 2 I ,
consequently,
X
hz x; Ux xi D hz x; !i .x/.Ui x x/i
i 2I
X
D !i .x/hz x; Ui x xi
i 2I
4.9 Extrapolated Simultaneous Cutter 189
X
!i .x/ kUi x xk2
i 2I
D
w .x/ kUx xk2 .
hz x; Ux xi
.x/ kUx xk2 .
.x/
w .x/, (4.101)
where
P 2
i 2I !i .x/ kPCi x xk
w .x/ WD P 2 (4.102)
i 2I !i .x/PCi x x
190 4 Algorithmic Projection Operators
for x … C and
w .x/ D 1 for x 2 C . We have C D Fix U D Fix U
; , because
w is an appropriate weight function (see Theorem 2.1.26 and Remark 2.4.2 (d)).
Therefore,
w is well defined. We call an operator U
with a step size function
Furthermore, let U
; be the generalized relaxation of U , defined by (4.97) with
2 Œ0; 2 and with the step size function
W H ! .0; C1/ satisfying inequality
(4.101). Then U
is a cutter. Consequently, for all 2 .0; 2/, U
; is 2
-strongly
quasi-nonexpansive and asymptotically regular.
Proof. Since the metric projection PC is a cutter with Fix PC D C (see Theo-
rem 2.2.21 (i) and (ii)), the corollary follows directly from Theorem 4.9.1. t
u
P h i2
.ai> xˇi /C
i 2I !i .x/ kai k
>
Tw x WD x A D.x/.Ax b/C (4.105)
A> D.x/.Ax b/C 2
.a> xˇ /
for x … C and Tw x D x for x 2 C . Note that the terms i kai k i C , i 2 I , and
A> D.x/.Ax b/C do not change if we rescale inequalities of the system (see
Remark 4.6.5), consequently, the operator Tw does not depend on the rescaling of the
inequalities. Therefore, we can suppose, without loss of generality, that kai k D 1,
i 2 I . Then we have
P
!i .x/Œ.ai> x ˇi /C 2 >
Tw x D x i 2I A W .x/.Ax b/C , (4.106)
A> W .x/.Ax b/C 2
Denote X
a.x/ WD i .x/ai D A> v.x/,
i 2I
192 4 Algorithmic Projection Operators
X
ˇ.x/ WD i .x/ˇi D v.x/> b
i 2I
and
Cv.x/ WD fy 2 Rn W v.x/> Ay v.x/> bg.
It is clear that C Cv.x/ for any weight function v. If a.x/ ¤ 0, then Cv.x/ is a
half-space, Cv.x/ D H .a.x/; ˇ.x//, which is called a surrogate constraint for the
system of linear inequalities Ax b.
Definition 4.9.4. We say that a weight function v W Rn ! m is essential (for the
system Ax b) if for any x … C it holds that
The contradiction proves that a.x/ WD A> v.x/ ¤ 0. Now we have C Cv.x/ D
H .a.x/; ˇ.x//. t
u
Let v W Rn ! m be an essential weight function and Sv x be the metric projection
of x 2 Rn onto the surrogate constraint H .a.x/; ˇ.x//. If x 2 C , then, of course,
Sv x D x. If x … C , then
.a.x/> x ˇ.x//C
Sv x D x a.x/
ka.x/k2
Œ.A> v.x//> x v.x/> bC >
Dx A v.x/
A> v.x/2
C
x Sv x
C2
Note that Sv x is well defined, because A> v.x/ ¤ 0 (see Lemma 4.9.5). We call
Sv W Rn ! Rn a surrogate projection (see Fig. 4.36). Note that an essential weight
function v can be arbitrarily defined on C .
A surrogate projection Sv is not a metric projection, because the weights vi
defining the half-space H .a.x/; ˇ.x// depend on x … C .
Theorem 4.9.6. Let C ¤ ;. If v W Rn ! m is an essential weight function, then
Fix Sv D C and Sv is a cutter. Consequently, for any 2 .0; 2/, its relaxation Sv;
is 2
-strongly quasi-nonexpansive and asymptotically regular.
hz Sv x; x Sv xi 0 (4.110)
1 1 1
1 .x/a1 C 2 .x/a2 C 3 .x/a3 D a C a a D 0,
4 4 2
x 2 Ci H) i .x/ D 0. (4.111)
194 4 Algorithmic Projection Operators
j .x/.aj> x ˇj / > 0,
i.e., v is essential. Now the second part of the Lemma follows from Lemma 4.9.5.
t
u
Note that S˛v D Sv for any weight function v W Rn ! m and for any function
˛ W Rn ! .0; C1/. Furthermore, Sv does not change if we rescale the inequalities.
Therefore, we can suppose without loss of generality that kai k D 1, i 2 I , and that
v W Rn ! Rm C nf0g instead of v W R ! m .
n
More general versions of the following two results were proved in [68, Theo-
rems 5 and 8]. The first result shows that any extrapolated simultaneous projection
is a surrogate projection.
Proposition 4.9.10. Let C ¤ ;, w W Rn ! m be an appropriate weight function
and v W Rn ! RmC nf0g be defined by
X
.Ax b/>
C W .x/.Ax b/ D .ai> x ˇi /C !i .x/.ai> x ˇi /
i 2I
X
D !i .x/Œ.ai> x ˇi /C 2
i 2I
D .Ax b/>
C W .x/.Ax b/C .
x 2 Rn , where
X i .x/
˛.x/ WD >
.
a x ˇi
i 2I.x/ i
!i .x/.ai> x ˇi /C D !i .x/.ai> x ˇi /
and
X X
!i .x/Œ.ai> x ˇi /C 2 D !i .x/.ai> x ˇi /.ai> x ˇi /C
i 2I i 2I
X
D ˛.x/1 i .x/.ai> x ˇi /C
i 2I
Therefore,
P >
i 2I !i .x/Œ.ai x ˇi /C
2
>
Tw x D x 2 A W .x/.Ax b/C
A> W .x/.Ax b/C
Proof. The equality in (4.113) follows from Propositions 4.9.10 and 4.9.11. The
inclusion in (4.113) follows from Lemma 4.9.9. t
u
One can easily prove that for a system Ax b containing at least two inequalities
defining two different half-spaces the inclusion in (4.113) is strict. We leave the
details to the reader.
Corollary 4.9.13. Let C ¤ ;. If a weight function v W Rn ! Rm C nf0g considers
only violated constraints, then a surrogate projection Sv W Rn ! Rn is a
cutter. Consequently, for any 2 .0; 2/, its relaxation Sv; is 2
-strongly quasi-
nonexpansive and asymptotically regular.
Proof. The corollary follows from Lemma 4.9.9 and from Theorem 4.9.6. Since
T D S, the corollary also follows directly from Corollary 4.9.3. t
u
is essential for the system AL x bL , consequently, the weight vector v.x/ with
vI nL .x/ D 0 2 Rmr is essential for the system Ax b. Therefore, the surrogate
projection Sv x given by (4.109) is well defined.
We say that a subset L.x/ satisfying the conditions (i) and (ii) above is a residual
selection (RS) of I . Note that such a subset exists, e.g., L.x/ WD fj.x/g, where
j.x/ 2 I.x/ WD fi 2 I W ai> x > ˇi g. Residual selections were studied in
[72, Sect. 3.2] and [73, Sect. 3.1], where several constructions of maximal subsets
L satisfying the conditions (i) and (ii) above were presented. If we replace the
condition (ii) above by the following stronger condition:
(iii) .AL A>
L/
1
0 and AL x bL 0,
then the residual selection L.x/ I is called an obtuse cone selection (OCS).
The notion OCS can be explained by the fact that for a full row matrix AL the
cone generated by the rows of AL is obtuse (in Linfai W i 2 Lg) if and only if
.AL A>L/
1
0 (see [232, Lemma 3.1] or [65, Lemma 1.6]). A special case of an
obtuse cone selection is a regular obtuse cone selection (ROCS), for which the
condition (iii) above is replaced by
(iv) ai> aj 0 for i ¤ j , i; j 2 L, and AL x bL 0
(see Fig. 4.37).
The first inequality in (iii) says that the inverse of the Gram matrix G WD
AL A>L of rows of AL is nonnegative. Nonsingular matrices with nonnegative
inverses, called inverse-positive matrices, play an important role in many areas
of mathematics and were studied in [35]. The properties of Gram matrices with
nonnegative inverse were also studied in [232, Sect. 3] and [65, Sect. 2], where their
198 4 Algorithmic Projection Operators
PC 3 x SÀ x
x
C
a2 PC 4 x
PC 2 x a4
a3 C6
C2 C4
a6
relations to obtuse cones are presented. We see that L is an obtuse cone selection if
and only if G is inverse-positive and AL x bL 0.
The first inequality in (iv) says that G has nonpositive off-diagonal elements.
Nonsingular matrices with a nonnegative inverse having this property are called M-
matrices or Minkowski matrices. The properties of such matrices were studied in
[166], [326, Sect. 4], [62, Chap. 5], [232, Sect. 3] and [65, Sect. 2]. In particular, it
follows from [166, Theorem 4.3] that a nonsingular Gram matrix with nonpositive
off-diagonals has a nonnegative inverse. This result also follows from [62, Theorem
5.4.A], from [232, Lemma 3.2], and from [65, Corollary 2.9]. Several constructions
of maximal subsets L.x/ being obtuse cone selection or regular obtuse cone
selection are presented in [326, Sect. 4], [62, Chap. 5], [65, Sect. 2]. In [62, Chap. 6],
[64, 230, 231] one can find applications of such selections to algorithms for convex
minimization problems.
Suppose that for any x … C the subset L.x/ I is a residual selection. If we
take vL .x/ D .AL A> 1
L / .AL x bL / and vI nL .x/ D 0, then equality (4.109) obtains
the form
Sv x D x A> > 1
L .AL AL / .AL x bL /. (4.114)
The operator Sv W Rn ! Rn defined by (4.114) is called a surrogate projection with
residual selection.
Further properties of the residual selection as well as its applications to the
surrogate projection methods are presented in Sect. 5.13.
P
We call the operator U WD i 2I !i Ui , where w W H ! m is a weight function,
a simultaneous subgradient projection. In this section we apply Theorem 4.9.1 to
the extrapolated simultaneous subgradient projection U
with the step size function
1
.x/
w .x/, (4.116)
where
2
Pm ciC .x/
i D1 wi .x/ kgi .x/k
w .x/ WD 2 (4.117)
Pm ciC .x/
i D1 wi .x/ kgi .x/k2 gi .x/
T
for all x … i 2I Ci ¤ ;, and the weight function w is appropriate (for simplicity,
.c .x//
we use the convention that kgi i .x/kC D 0 if .ci .x//C D 0). Note that the step
size
w .x/ defined by (4.117) is a special case of the step size (4.99), where Ui
is the subgradient projection defined by (4.115), i 2 I .
Corollary 4.9.14. Let U
; beP a generalized relaxation of the simultaneous
subgradient projection U WD i 2I !i Ui with an appropriate weight function
w W H ! m , with
T a step size function
W H ! .0; C1/ satisfying inequality
(4.116) for x … i 2I Ci ¤ ; and with 2 Œ0; 2. Then the operator U
is a cutter.
Consequently,
kU
; x zk2 kx zk2 .2 /
2 .x/ kUx xk2
2
D kx zk2 kU
; x xk2
T
for all x 2 H, z 2 i 2I Fix Ui and for all 2 .0; 2/, i.e., U
; is 2
-strongly
quasi-nonexpansive and asymptotically regular.
Proof. Since a subgradient projection Ui is a cutter and Fix Ui D Ci , i 2 I , the
corollary follows from Theorem 4.9.1. t
u
The properties of a special case of the extrapolated simultaneous subgradient projec-
tion were studied by Dos Santos in [146], where
D
w and an appropriate weight
function w is supposed to be constant, i.e., w 2 ri m . The study was continued
for various convex optimization problems by Cegielski [62, Sect. 4.3], Kiwiel [229,
Sects. 3 and 8], [230, 231], Kiwiel and Łopuch [233, Sect. 2], Combettes [118,
Sect. 5.6], [120, Sect. V] and by Cegielski and Censor [70].
U
; x WD I C
.x/.Ux x/.
Tm
where x 2 H and z 2 i D1 Fix Ui ¤ ;.
Proof. (cf. [71, Lemma 7]) Let x 2 H be arbitrary. We prove the first inequality in
(4.119) by induction with respect to m. For m D 1 this inequality follows directly
from the definition of a cutter and from inequality (2.21). Suppose that the first
inequality in (4.119) is true for some m D k. Let w 2 H. Define V1 WD Id, Vi WD
Ui Ui 1 : : : U2 for i D 2; 3; : : : ; kC1, v1 WD w, vi WD Ui vi 1 and zi WD vi vi 1 , i D
TkC1
2; 3; : : : ; k C 1. Take w WD U1 x and z 2 i D1 Fix Ui . Then, of course, Si x D Vi w,
ui D vi and y i D zi , i D 2; 3; : : : ; k C 1. It follows from the induction assumption
that
X
kC1
hVkC1 w w; z wi hzi C : : : C zkC1 ; zi i (4.120)
i D2
PkC1
Note that VkC1 w w D i D2 zi and w x D y 1 . Since U1 is a cutter, it follows
from (4.120) that
hSkC1 x x; z xi
D hVkC1 w x; z xi
D hVkC1 w w; z xi C hU1 x x; z xi
2
hVkC1 w w; z wi C hVkC1 w w; w xi C y 1
4.10 Extrapolated Cyclic Cutter 201
X
kC1 X
kC1
2
hzi C : : : C zkC1 ; zi i C h zi ; y 1 i C y 1
i D2 i D2
X
kC1 X
kC1
2
D hy i C : : : C y kC1 ; y i i C hy i ; y 1 i C y 1
i D2 i D2
X
kC1
D hy i C : : : C y kC1 ; y i i.
i D1
Therefore, the first inequality in (4.119) is true for all m 2 N. The second inequality
follows from the following equality
m
Xm X
m
X 2
1 y i D
2 1
hy i C : : : C y m ; y i i yi .
i D1
2 i D1
2
i D1
The third inequality in (4.119) follows from the convexity of the function kk2 . t
u
We use the notation from the previous subsection. Define the step size function
where x … Fix U .
Note that
X
m X
m
hy i C : : : C y m ; y i i D hy 1 C : : : C y i ; y i i.
i D1 i D1
hUx x; z xi
max .x/ kUx xk2 . (4.124)
Consequently,
hUx x; z xi
.x/ kUx xk2 .
Now Corollary 2.4.5 yields that U
is a cutter and U
; is 2 -strongly quasi-
nonexpansive. The asymptotic regularity of U
; follows from Theorem 3.4.3. u t
4.11 Exercises
Exercise 4.11.1. Let B.z; / H, be a ball with the centre z 2 H and a radius
> 0. Prove that B.z; / is a nonempty closed convex subset and that
(
x if kx zk
PB.z;/ .x/ D
zC kxzk .x z/ if kx zk > .
x0 2 X – arbitrary
x kC1 D Rk x k ,
A. Cegielski, Iterative Methods for Fixed Point Problems in Hilbert Spaces, 203
Lecture Notes in Mathematics 2057, DOI 10.1007/978-3-642-30901-4 5,
© Springer-Verlag Berlin Heidelberg 2012
204 5 Projection Methods
In autonomous methods the sequence of operators fTk g1 kD0 as well as the sequence
of relaxation parameters .k / are constant and these methods have the form
The alternating projection method (APM), also known as the von Neumann method
is an iterative method for finding a common point of two nonempty closed convex
subsets A; B H, if such a point exists. One iteration of the method has the form
x kC1 D PA PB x k (5.4)
Suppose that A and B are closed subspaces of a Hilbert space H. Then the strong
convergence holds in Corollary 5.1.3. The following theorem is due to John von
Neumann [271, Theorem 13.7].
Theorem 5.1.5 (von Neumann, 1933). Let A; B H be closed subspaces. Then,
for any x 0 2 H, the sequence fx k g1kD0 generated by the alternating projection
method (5.4) converges in norm to y WD PA\B x 0 .
Proof. Let x 0 2 H and x k D T k x 0 , k 0, where T WD PA PB . We divide the proof
into three parts.
(i) Since the metric projection is strictly quasi-nonexpansive (see Theorem 2.2.21
(iii) and Corollary 2.2.9) and A \ B ¤ ;, we have Fix T D A \ B (see
Theorem 2.1.26). The weak convergence of x k to a point x 2 A \ B follows
from Corollary 5.1.3.
(ii) We show that x D PA\B x 0P . Denote y 2l WD x l and y 2lC1 WD PB x l , l 0.
2k1 l
It is clear that y y D
i 2k
lDi y y
lC1
, for all i; k 0, i 2k. Let
z 2 A \ B be arbitrary. By Theorem 2.2.30 (i), we have
hy 2l y 2lC1 ; zi D hx l PB x l ; zi D 0
and
hy 2lC1 y 2lC2 ; zi D hPB x l PA PB x l ; zi D 0,
l 0. Consequently,
X
2k1
hy i x ; zi D limhy i y 2k ; zi D limh y l y lC1 ; zi
k k
lDi
X
2k1
D lim hy l y lC1 ; zi D 0,
k
lDi
5.1 Alternating Projection Methods 207
i 0. In particular,
hPB x 0 x ; x i D hy 1 x ; x i D 0
and
hx 0 x ; zi D hy 0 x ; zi D 0,
i.e., x D PA\B x 0 .
(iii) We show that limk x k x D 0. Since the metric projection onto a
closed subspace is self-adjoint (see Theorem 2.2.30 (iii)) and idempotent (see
Theorem 2.2.21 (i)), we have
hx k x ; x k i D hT x k1 T x ; x k i D hx k1 x ; T x k i.
If we iterate the above equalities k-times and apply the facts that x k * x and
PB x D x , we obtain
hx k x ; x k i D hx 0 x ; .T /k x k i D hx 0 x ; .T /k T k x 0 i
D hx 0 x ; PB T 2k1 x 0 i D hPB x 0 PB x ; T 2k1 x 0 i
D hPB x 0 x ; x 2k1 i ! hPB x 0 x ; x i D 0.
i.e., x k ! x D PA\B x 0 . t
u
Several proofs of the von Neumann theorem can be found, e.g., in [152], [140,
Theorem 9.3] [235, Theorem 1.1] and [236].
Theorem 5.1.5 remains true if A and B are closed affine subspaces with a
common point. The details are left to the reader.
One can also estimate the rate of convergence of the alternating projection
method x kC1 D PA PB x k . It turns out that when A and B are closed subspaces of
H the rate depends on the angle between the subspaces. The following definition is
due to Friedricks [170, Sect. 1.1] (see also [139, Sect. 6] and [140, Definition 9.4]).
Definition 5.1.6. Let A; B H be closed subspaces and
The value ˛.A; B/ WD arccos c.A; B/ 2 Œ0; 2 is called an angle between the
subspaces A and B.
It follows from the Cauchy–Schwarz inequality that c.A; B/ 2 Œ0; 1, conse-
quently, the the angle between the subspaces A and B is well defined. The proof of
the following theorem can be found in [11, page 379] or in [140, Sect. 9.8].
Theorem 5.1.7. Let A; B H be two closed subspaces, c WD c.A; B/, x 0 2 H,
and fx k g1
kD0 be generated by the alternating projection method (5.4). Then
k
x PA\B x 0 c 2k1 x 0 PA\B x 0 ,
consequently fx k g1
kD0 converges geometrically, if c < 1.
In papers [290, 291, 313] the angles between subspaces are also used to obtain
convergence rates for iterative projection methods. Further results on the alternating
projection method can be found, e.g., in [20, 21], [140, Chap. 9] and [174].
Applications of the alternating projection method in various areas of mathematics
were presented in [139].
The sequences generated by the alternating projection method often converge very
slowly. Such a slow convergence is observed, e.g., if A is a hyperplane and B is a
ball which is tangent to A (see, e.g., [21, Example 5.3]) or A and B are two closed
subspaces with c.A; B/ close to 1. In these cases the angle between x k PB x k and
PA PB x k PB x k is close to 0 which causes very short steps PA PB x k x k . Since
the alternating projection method has many practical applications, any acceleration
technique seems to be important. In this section we give several modifications of
the alternating projection method which lead in practice to an acceleration of the
convergence. All these methods have the form
for x 2 H, where fk g1 kD0 W H ! .0; C1/ is a sequence of step size functions
and k 2 Œ0; 2 is a relaxation parameter (cf. Sect. 2.4). Suppose, for simplicity, that
.x/ D 1 for x 2 A \ B. We can also write
Let
kPB x xk2
GPR .x/ WD , (5.8)
hPB x x; T x xi
for x 2 AnB. The step size (5.8) was proposed by Gurin et al. in [196, Sect. 3].
Recall that GPR .x/ 1 for all x 2 A and that GPR .x/ D 1 if and only if
PB x 2 A\B (see Lemma 4.3.13). Therefore, recurrence (5.7) with k D GPR .x k /
describes an extrapolated alternating projection method. The following theorem
extends the results of [196, Theorem 4].
Corollary 5.2.1. Let x 0 2 A and the sequence fx k g1
kD0 be generated by (5.7),
where
1 k .x/ GPR .x/
and lim inf k .2 k / > 0. Then x k converges weakly to a point x 2 A \ B.
Proof. Denote Tk WD Tk . It follows from Corollary 4.3.14 that the operator TGPR is
a cutter. Therefore, Tk are also cutters, k 0, (see Lemma 2.4.7 (ii)).The operator
T is nonexpansive as a composition of nonexpansive operators PA and PB . It is
clear that Fix Tk D Fix T D A \ B. We have
Tk x k x k D k .x k / T x k x k T x k x k
consequently, T x k x k ! 0 whenever Tk x k x k ! 0. Corollary 3.7.1 with
S D T and X D A yields now the weak convergence of x k to a point x 2 A\B. u t
Gurin et al. proved a special case of Corollary 5.2.1 with k D GPR and k D 1
for all k 0 (see [196, Theorem 4]).
210 5 Projection Methods
ı WD ı.A; B/ D inf kx yk
x2A;y2B
and
N
ı.x/ WD kT x PB xk
for x 2 A.
First we consider the general case, where A; B H are nonempty closed convex
subsets. Let
Q
kT x PB xk2 ı.x/ kPB x xk C hPB x x; T x xi
CS1 .x/ WD
kT x xk2
Tk x k x k D k .x k / T x k x k 1 T x k x k ,
2
consequently, T x k x k ! 0 whenever Tk x k x k ! 0. Corollary 3.7.1 with
S D T and X D A yields now the weak convergence of x k to a point
x 2 Fix PA PB . t
u
The convergence of sequences generated by the extrapolated alternating projection
method (5.7) with the step size function k WD CS1 , k 0, was proved in [76,
Theorem 15 (i)].
Now we consider the case, where A H is a closed affine subspace. Let
Q 2
.kT x PB xk ı/
CS 2 .x/ WD 1 C
kT x xk2
Q
for x … Fix T , where ı.x/ N
2 Œı; ı.x/. It is clear that CS 2 .x/ 1.
212 5 Projection Methods
k
x kC1 D x k C .RA RB x k x k /, (5.10)
2
where k 2 Œ0; 2. If k D 1, then we obtain (5.9). We can also write x kC1 D Tk x k ,
where T denotes a -relaxation of the AAR operator T . The following corollary
extends the result of [26, Fact 5.9].
Corollary 5.2.4. Let A; B H be closed convex. If Fix.RA RB / ¤ ; and
lim infk k .2 k / > 0, then for any x 0 2 H the sequence fx k g1
kD0 generated
by (5.10) converges weakly to a fixed point of RA RB and the sequence fPB x k g1
kD0
is bounded. Furthermore, if A \ B ¤ ;, then:
5.3 Projected Gradient Method 213
1
limŒPA .2PB x k x k / PB x k D lim .RA RB x k x k / D lim.T x k x k / D 0.
k k 2 k
(i) Let y be a weak cluster point of fPB x k g1 kD0 . Because B is weakly closed
(see Theorem 1.1.40) and fPB x k g1 kD0 B, we have y 2 B. We prove that
y 2 A. Suppose y … A and let a subsequence fPB x nk g1 k 1
kD0 fPB x gkD0
converge weakly to y. By the boundedness of fPB x gkD0 and of fx k g1
k 1
kD0
and by the nonexpansivity of PA , the sequence fPA .2PB x nk x nk /g1 kD0 is
bounded, consequently, it contains a subsequence fPA .2PB x mk x mk /g1 kD0
which converges weakly to a point z 2 H. Since A is weakly closed, z 2 A.
Because fPA .2PB x k x k / PB x k g1
kD0 converges to 0, it converges weakly to
0 and fPA .2PB x mk x mk / PB x mk g1kD0 also converges weakly to 0. On the
other hand, fPA .2PB x mk x mk / PB x mk g1
kD0 converges weakly to z y ¤ 0.
A contradiction shows that y 2 A.
(ii) Let dim H < 1. The first part of the corollary and the nonexpansivity of PB
yield
lim PB x k PB x lim x k x D 0,
k k
(iii), y 2 A \ B. t
u
closed convex subset. The method is called a projected gradient method and has
the form
x kC1 D PC .x k F x k /. (5.11)
The convergence of sequences generated by the method follows from the theorem
below, which shows an application of the Banach fixed point theorem and the
properties of the metric projection.
Theorem 5.3.1. Let C H be closed convex, F W H ! H be -Lipschitz
continuous and -strongly monotone over C , where ; > 0, and 2 .0; 2 2
/. Then
the operator T WD PC .Id F / is a contraction. Consequently, for any x 0 2 H, the
sequence fx k g1kD0 defined by (5.11) converges to the unique solution of VIP.F ; C /
with a rate of the geometric progression.
Proof. (cf. [358, Theorem 46.C]) Define G D Id F . Let x; y 2 H. Then
i.e., PC .Id F / is a .1 /-contraction. The Banach fixed point theorem and
Theorem 1.3.8 yield that PC .Id F / has a unique fixed point xN 2 C and xN is
the unique solution of VIP.F ; C /. Furthermore, the Banach fixed point theorem
yields that for any x 0 2 H, the sequence fx k g1
kD0 generated by (5.11) converges
N
geometrically to x. t
u
In the literature one can find several method for solving VIP.F ; C / in particular
for solving VIP.F ; Fix T /, where T W H ! H is a quasi-nonexpansive operator, for
solving convex constrained minimization problems, for finding the metric projection
onto the intersection of a finite family of convex subsets, for solving elliptic control
problems, etc. For details we send the reader to [18,52,77,94–97,101,144,171,200,
203,211,216–218,237,244,250,251,269,273,274,276,318,320,338,342,343,345–
349, 353].
5.4 Simultaneous Projection Method 215
x0 2 H – arbitrary
P (5.13)
x kC1
D x C k . i 2I !i PCi x x /,
k k k
1X
f .x/ WD !i kPCi x xk2 (5.14)
2 i 2I
X
k
2
d 2 .x kC1 ; Fix T / d 2 .x 0 ; Fix T / l .2 l / T x l x l , (5.15)
lD0
T
k 0. If, furthermore, C WD i 2I Ci ¤ ;, then
216 5 Projection Methods
X
k X 2
d 2 .x kC1 ; C / d 2 .x 0 ; C / l .2 l / !i PCi x l x l , (5.16)
lD0 i 2I
k 0.
Proof. Let Fix T ¤ ;. The operator T is firmly nonexpansive as a convex combi-
nation of firmly nonexpansive operators PCi (see Corollary 2.2.20). Furthermore,
T is nonexpansive (see Theorem 2.2.4) and a cutter (see Theorem 2.2.5 (i)).
Consequently, T Id is demi-closed at 0 (see Lemma 3.2.5). If we take X D H,
Tk WD T and S WD T in Corollary 3.7.1, then we obtain that x k converges weakly
to a point x 2 Fix T and that
lC1 2 2 2 2
x z D T x l z x l z l .2 l / T x l x l ,
2 2 Xk
2
D T x k z x 0 z l .2 l / T x l x l
lD0
X
k
2
D d 2 .x 0 ; Fix T / l .2 l / T x l x l .
lD0
If C ¤ ;, then estimation (5.16) can be proved in the same way as above applying
Theorem 4.4.5. t
u
If Ci H are closed subspaces, then the convergence in Corollary 5.4.1 is strong
and x D PC x 0 . This result is due to Reich [295, Theorem 1.7], where the
convergence was proved for uniformly convex Banach spaces.
The strong convergence of sequences generated by the simultaneous projection
method does not hold in general. Bauschke, Matoušková and Reich, relying on a
result of Hundal [213, Theorem 1], gave an example of two closed convex subsets
C1 ; C2 H with a nonempty intersection, a simultaneous projection T WD 12 PC1 C
1 k
2 PC2 and a point x for which T x converges weakly but does not converge strongly
[32, Theorem 5.1].
If H D Rn and Ci are hyperplanes or half-spaces, i.e., Ci WD H.ai ; ˇi / D fx 2
R W ai> x D ˇi g or Ci WD H .ai ; ˇi / D fx 2 Rn W ai> x ˇi g, then it follows from
n
equalities (4.63) and (4.70) that the simultaneous projection method (5.13) can be
written in the matrix form
or
x kC1 D x k k A> D.Ax k b/C , (5.18)
respectively, where D WD diag. !1 2 ; !2 2 ; : : : ; !m 2 /.
ka1 k ka2 k kam k
The simultaneous projection method (5.13) was introduced by Cimmino in [116],
where H D Rn , Ci WD H.ai ; ˇi /, i 2 I , C ¤ ; and k D 2 for all k 0. It follows
from (5.17) that Cimmino’s method has the form
P
If Fix PX T ¤ ;, where T WD i 2I !i PCi , then x k converges weakly to a point
x 2 Fix PX T .
Proof. T is firmly nonexpansive (see Corollary 4.4.4). If we set Sk WD T , k 0,
in Corollary 3.7.4, then we obtain the weak convergence of x k to a point x 2
Fix PX T . t
u
Consider now the following recurrence
x kC1 D PX .x k C k .T x k x k //,
x0 2 H – arbitrary
(5.22)
x kC1 D PCm : : : PC1 x k .
If we replace the operators PCi in (5.22) by their relaxations PCi ;i , where i 2
.0; 2/, i 2 I , then we obtain the following, more general method
5.5 Cyclic Projection Methods 219
x0 2 H – arbitrary
(5.23)
x kC1
D PCm ;m : : : PC1 ;1 x ,
k
which we call the cyclic relaxed projection method. We can also write x kC1 D T x k ,
where T WD PCm ;m : : : PC1 ;1 is the cyclic relaxed projection. Denote S0 WD Id and
Si WD PCi ;i : : : PC1 ;1 for i D 1; 2; : : : ; m. We have Sm D T .
5.5.1 Convergence
X
k
2
d 2 .x kC1 ; Fix T / d 2 .x 0 ; Fix T / ˛ T x l x l , (5.24)
lD0
T
where ˛ D 2 max
mmax
and max WD maxfi W i 2 I g. If C WD i 2I Ci ¤ ; and i D 1
for all i 2 I , then
X
k X
d 2 .x kC1 ; C / d 2 .x 0 ; C / Si x l Si 1 x l 2 . (5.25)
lD0 i 2I
Proof. Let Fix T ¤ ;. It follows from Corollary 4.5.8 that T is -RFNE, where D
2mmax
.m1/max C2
and that T is asymptotically regular. Since T is nonexpansive, it follows
from Theorem 3.5.3 that x k converges weakly to x 2 Fix T . Corollary 2.2.9 yields
the ˛-strong quasi nonexpansivity of T , where ˛ WD 2
D 2 max
mmax
. Therefore, T is
˛-SQNE, i.e.,
l
T x z2 x l z2 ˛ T x l x l 2
for all z 2 Fix T , l D 0; 1; : : : If we take z D PFix T x 0 and apply the above inequality
for l D 0; 1; : : : ; k, we obtain
2
d 2 .x kC1 ; Fix T / x kC1 z
2 2 Xk
2
D T x k z x 0 z ˛ T x i x i
i D0
X
k
2
D d 2 .x 0 ; Fix T / ˛ T x i x i .
i D0
220 5 Projection Methods
P WD Id C.P Id/, where 2 .0; 2/, and RK WD .PK /2 D 2PK Id, i.e., P
is a relaxation of the cyclic projection P and RK is the reflection operator onto K.
Consider the iterative method defined by the following recurrence
5.5 Cyclic Projection Methods 221
x kC1 D RK P x k , (5.26)
y kC1 D P RK y k (5.27)
Furthermore, y kC1
D P x (see Fig. 5.1).
k
Therefore,
.P RK /kC1 y 0 .P RK /k y 0 D y kC1 y k ! 0,
where fk g1 1
kD0 Œ0; 2 is a sequence of relaxation parameters and fik gkD0 I
is a control sequence or, shortly, control. Note that ik can depend on the current
approximation x k 2 H. In this case we identify ik with ik .x k /. It is convenient
to define ik .x/ for any x 2 H, i.e., to define a sequence of functions fik g1 kD0 W
H ! I . We will, however, relate this sequence of functions to recurrence (5.28).
If the sequence of functions fik g1
kD0 is constant, i.e., ik D i for all k 0, then the
function i W H ! I is called a control function or, shortly, control. In this case we
write ik D i.x k /. A comprehensive overview of the sequential projection methods
can be found in [78] and in [108, Chap. 5].
5.6.1 Convergence
Remark 5.6.3. Since I is a finite subset, the equality on the right hand side of
(5.29) can be written in the following equivalent form: limk PCi x k x k D 0 for
all i 2 I . Therefore, implication (5.29) is equivalent to the following one
lim PCik x k x k D 0 H) lim PCi x k x k D 0 for all i 2 I .
k k
Note that a remotest set control is an almost remotest set control. Furthermore,
an almost remotest set control is approximately remotest set control.
Definition 5.6.7. Let Cj WD fx 2 H W haj ; xi ˇj g, j 2 I , be half-spaces, where
aj 2 H, aj ¤ 0, and ˇj 2 R , j 2 I . A control i W H ! I is called a maximal
residual control if
i.x/ D argmax.haj ; xi ˇj /.
j 2I
Note that if aj D 1, j 2 I , then a maximal residual control is a remotest set
control.
Lemma 5.6.8. Let Cj D fx 2 H W haj ; xi ˇj g, j 2 I , be half-spaces, where
aj 2 H, aj ¤ 0, and ˇj 2 R , j 2 I . The maximal residual control is an almost
remotest set control. Consequently, it is an approximately remotest set control.
Proof. Let ˛ WD minj 2I aj , ˇ WD maxj 2I aj and i W H ! I be a maximal
residual control. We have
haj ; xi ˇj
hai.x/ ; xi ˇi.x/ D max.haj ; xi ˇj / D max.aj /
j 2I j 2I aj
haj ; xi ˇj
˛ max D ˛ max PCj x x
j 2I aj j 2I
T
for any x … j 2I Cj . Therefore,
hai.x k / ; x k i ˇi.x k /
˛
PCi.xk / x k x k D max PCj x x
ai.x k / ˇ j 2I
k 1
kCr rX kClC1
x k x k x x kCl
lD0
k 1
rX
D kCl PCikCl x kCl x kCl
lD0
X
s2
kCl PCikCl x kCl x kCl ,
lD0
consequently,
lim x kCrk x k D 0. (5.32)
k
The definition of the metric projection and the triangle inequality yield,
PC x k x k PC x kCrk x k
i i
PCi x kCrk x kCrk C x kCrk x k
D PCikCr x kCrk x kCrk C x kCrk x k .
k
Therefore, (5.31) and (5.32) yield limk PCi x k x k D 0, i.e., fik g1
kD0 is
approximately regular.
(ii) Let H be finite dimensional, C ¤ ;, lim infk k .2 k / > 0 and fik g1 kD0 be
repetitive. Since (5.28) is a special case of (3.9) with X D H and Tk WD PCik ,
by Corollary 3.7.1, we obtain
kC1 2 2 2
x
z x k z k .2 k / PCik x k x k
i.e., fik g1
kD0 is an approximately semi-remotest set control sequence. t
u
Since the cyclic control is almost cyclic, it is an approximately remotest set
control. Therefore, Corollary 5.6.4 yields that the sequences generated by the cyclic
projection method with lim infk k .2 k / > 0 converge (at least weakly) to a point
x 2 C if C ¤ ;.
5.6 Successive Projection Methods 227
5.6.3 Examples
Example 5.6.13. Consider the successive projection method (5.28) with the cyclic
control fik g1
kD0 and with k 2 Œ"; 2 " for some " 2 .0; 1/. The method (also called
a cyclic projection method) was studied by:
(i) Kaczmarz in [223], where H D Rn , Ci WD fx 2 Rn W hai ; xi D ˇi g, i D
1; 2; : : : ; n, ai 2 Rn are linearly independent and k D 1 for all k 0.
Kaczmarz proved that for any starting point x 0 2 Rn the sequence fx k g1 kD0
generated by the method converges to a unique solution of the system hai ; xi D
ˇi , i D 1; 2; : : : ; n;
(ii) Tanabe in [321], where H D Rn , Ci WD fx 2 Rn W hai ; xi D ˇi g, i 2 I , (with
possibly an empty intersection) and k D 1 for all k 0. Tanabe proved the
convergence of any sequence generated by the method (see [321, Corollary 9]);
(iii) Bregman in [42], where Ci H are closed convex subsets with C ¤ ; and
k D 1 for all k 0. Bregman proved that any sequence generated by the
method converges weakly to a point x 2 C (see [42, Theorem 1]);
(iv) CensorTet al. in [86], where H D Rn , Ci D fx 2 Rn W hai ; xi D ˇi g, i 2 I ,
C D i 2I Ci D ; and k D 2 .0; 2/. Let x 0 2 Rn be arbitrary, x k be
generated by the successive projection method with cyclic control and x ./ D
limk x km . Censor et al. proved that if # 0, then x ./ converges to a least
squares solution (see [86, Theorem 1]). Small relaxation parameters for the
Kaczmarz method were also used in [201];
(v) Gurin et al. in [196, Theorem 1], where Ci H, i 2 I , are closed convex
subsets with C ¤ ; and k 2 Œ"; 2 " for some " 2 .0; 1/ and for all
k 0. Gurin et al. proved the strong convergence of sequences generated by
the method if the subsets Ci , i 2 I , have a special structure (e.g., they are
half-spaces or are uniformly convex or C has a nonempty interior).
Example 5.6.14. Let C ¤ ;. Consider the successive projection method (5.28)
with the approximately remotest set control and with k 2 Œ"; 2 " for some " 2
.0; 1/. Corollary 5.6.4 yields x k * x 2 C .
(i) The convergence properties of the successive projection method with the
remotest set control were studied by Agmon in [1] and by Motzkin and
Schoenberg in [266], where H D Rn , Ci D fx 2 Rn W hai ; xi ˇi g,
k D 2 .0; 2/. Agmon proved that the sequences generated by the method
converge linearly [1, Theorem 3]. Motzkin and Schoenberg also considered the
case D 2 if C has a nonempty interior and proved the finite convergence in
this case.
(ii) Bregman in [42] considered the convex feasibility problem with infinitely
many subsets Ci H, i 2 I , having a nonempty intersection, and studied
successive projection methods for this problem. In [42, Theorem 2] the
convergence x k * x 2 C was proved for the remotest set control (the
existence of such control was supposed). Bregman also considered a special
case of the problem with infinitely many half-spaces Ci H, Ci WD fx 2 H W
228 5 Projection Methods
k
x kC1 D x k C A .PQ .Ax k / Ax k /,
kAk2
where x 0 2 H1 is arbitrary and k 2 Œ0; 2. We can also write x kC1 D Tk x k .
where T is a -relaxation of the Landweber operator T WD 1 2 A .PQ Id/A.
kAk
Landweber proposed the method with k D 2 .0; 2/, k 0, for solving an
integral equation (see [240]). The projected Landweber method (PLM) is an iterative
method for solving the following split feasibility problem: find x 2 C such that
Ax 2 Q, where C H1 , Q H2 are nonempty closed convex subsets of Hilbert
spaces H1 ; H2 and A W H1 ! H2 is a nonzero linear bounded operator, if such a
point exists. The iterative step of the projected Landweber method has the form
k
x kC1 D PC .x k C A .PQ .Ax k / Ax k //, (5.33)
kAk2
where k 2 Œ0; 2. We can also write x kC1 D PC Tk x k , where T is the -relaxation
of the Landweber operator T or x kC1 D Rk x k , where R WD PC T is the projected
-relaxation of the Landweber operator. It is clear that if C D H1 , then the projected
Landweber method reduces to the Landweber method. The projected Landweber
method was studied by Byrne [55] for H1 and H2 being Euclidean spaces and
for k D 2 .0; 2/, k 0. Actually, Byrne called the method a CQ algorithm
5.7 Landweber Method and Projected Landweber Method 229
k
x kC1 D PC .x k C A> .b Ax k //
max .A> A/
k
x kC1 D PC .x k C A> D.b Ax k // (5.34)
max .A> DA/
1
(cf. (4.80)) and the projected Landweber method for the modified system D 2 Ax
1
D 2 b has the form
k
x kC1 D PC .x k A> D.Ax k b/C / (5.35)
max .A> DA/
(cf. (4.83). If we take a positive definite matrix D with nonnegative elements and
C D Rn in (5.35), then we obtain a method which was studied in [87, equality (12)
and Theorem 2] (see also [322], [132, Sect. 4], [98], [220, Theorem 2], [160, 222]
for related results for a system of linear equations). Note that method (5.34) for
1 1
solving the system D 2 Ax D D 2 b is an extrapolation of the simultaneous projection
method (5.17) for solving the system Ax D b (see Corollary 4.6.7). Similarly,
1 1
method (5.35) for solving the system D 2 Ax D 2 b is an extrapolation of the
simultaneous projection method (5.18) for solving the system Ax b, where the
extrapolation parameter (step size) L D .A1> DA/ 1.
max
In this section we consider the common fixed point problem for a finite family of
cutters U D fUi gi 2I , where the operators Ui W H ! H, i 2 I WD f1; 2; : : : ; mg,
have a common fixed point. We study the convergence ofPsequences generated by
relaxations of simultaneous cutters of the form V WD i 2J !i Vi , where V WD
fV
T i g i 2J is a finite
T family of cutters V i W H ! H, i 2 J , with the property
i 2J Fix V i i 2I Fix U i and w W H !
jJ j is an appropriate weight function.
One iteration of the recurrence is defined by
X
x C D V x D x C . !i .x/Vi x x/,
i 2J
x0 2 H – arbitrary
P (5.36)
x kC1
D x C k . i 2Jk !i .x /Vi x x /,
k k k k k k
x kC1 D x k C k .V k x k x k /,
P
where V k x WD i 2Jk !i .x/Vi x for x 2 H. It is clear that V is a cutter (see
k k k
Corollary 2.1.49). Iterations of type (5.36) were studied by several authors (see [70]
and the references therein).
T T
Suppose that i 2J Fix Vi i 2I Fix Ui . Note that a weight function which is
regular with respect to the family U WD fUi gi 2I is appropriate with respect to the
family V WD fVi gi 2J (cf. Definition 2.1.25). It is also clear that if V U, then a
constant weight function w with all positive weights !i , i 2 J , is regular.
Example 5.8.3. Let V WD U, I.x/ WD fi 2 I W x … Fix Ui g and m.x/ WD jI.x/j
be the number of elements of I.x/, where x 2 H. The following weight functions
w W H !
m , where w.x/ D .!1 .x/; !2 .x/; : : : ; !m .x//, x 2 H, are regular:
(i) Positive constant weights, i.e.,
w.x/ WD w 2 ri
m (5.38)
i 2 I , x 2 H, where v D .1 ; : : : ; m / 2 ri
m . To verify that w is regular
it suffices to choose j 2 Argmaxi 2I kUi x xk and take ˇ WD mini 2I !i in
Definition 5.8.2. In particular, a weight function ! with
(
1
for i 2 I.x/
!i .x/ D m.x/ (5.40)
0 for i … I.x/,
( kUi xxk T
P for x … Fix Uj
!i .x/ WD kUj xx k
j 2I
T
j 2I
(5.41)
0 for x 2 j 2I Fix Uj ,
fˇk g1
kD0 .0; C1/ with lim infk ˇk > 0 such that w applied to the family V is
k k
T
for any z 2 i 2I Fix Ui , consequently, Vik x k x k is bounded. Since !ik .x k /
is also bounded, we have the following equivalence
2
lim !ikk .x k / Vikk x k x k D 0 ” lim !ikk .x k / Vikk x k x k D 0 (5.48)
k k
lim !ikk .x k / Vikk x k x k D 0 H) lim Ui x k x k D 0 for all i 2 I ,
k k
(5.49)
and implication (5.47) is equivalent to the following one
lim !ikk .x k / Vikk x k x k D 0 H) lim inf Ui x k x k D 0 for all i 2 I .
k k
(5.50)
(ii) The following condition yields the approximate regularity of fwk g1 kD0 :
There is a sequence fik g1
kD0 with i k 2 Jk for all k 0, such that
and
lim Vikk x k x k D 0 H) lim Ui x k x k D 0 for all i 2 I . (5.51)
k k
for all k 0 and for a constant ˇ > 0. If Jk D I , k 0, then the following con-
dition also implies the approximate regularity of fwk g1 k k
kD0 : lim infk !i .x / > 0
and for all i 2 I with !i .x / > 0 it holds
k k
k k
V x x k ˇ Ui x k x k for all k 0. (5.53)
i
2
lim ˇk max Ui x k x k D 0
k i 2I
Consequently, limk maxi 2I Ui x k x k D 0, i.e., limk Ui x k x k D 0 for
all i 2 I .
(ii) Let a subsequence fwnk g1 k 1
kD0 fw gkD0 applied to the sequence of families
V , k 0, be regular, i.e., for an arbitrary k, there exists ink 2 Jnk such that
nk
2
!innk .x nk / Vinn k x nk x nk ˇnk max kUi x nk x nk k2
k k i 2I
then, of course, 2
lim !innk .x nk / Vinn k x nk x nk D 0,
k k k
consequently,
lim ˇnk max kUi x nk x nk k2 D 0.
k i 2I
i.e.,
2
lim inf max Ui x k x k D 0
k i 2I
or, in other words, lim infk Ui x k x k D 0 for all i 2 I . t
u
Now we present theorems which give sufficient conditions for important control
sequences to be approximately (semi-)regular. Consider a special case of recurrence
(5.36), where Jk D I , Fix Vik Fix Ui for all k 0 and for all i 2 I . We can write
recurrence (5.36) in the form
X
x kC1 D x k C k !ik .Vik x k x k /, (5.54)
i 2I
X
x kC1 D x k C k !ik .Vik x k x k /, (5.55)
i 2Ik
Note that the latter condition is weaker than the approximate regularity of fwk g1
kD0
with
respect to the family fPFix Ui gi 2I , because
Ik I , consequently maxi 2Ik
PFix U x k x k maxi 2I PFix U x k x k .
i i
Definition 5.8.8. (cf. [22, Definition 3.18]) We say that the sequence fIk g1kD0 I
is p-intermittent, where p 2 N, or intermittent if I D Ik [ IkC1 [ : : : [ IkCp1 for
any k 0.
A simple example of an intermittent sequence is an almost cyclic control
(cf. Definition 5.6.10).
Theorem 5.8.9. Let fx k g1 1
kD0 be generated by (5.54), fIk gkD0 I be intermittent.
Suppose that for any k 0 there is lk 2 Ik , k 0, satisfying condition (5.56). If
wki D 0 for all i … Ik , k 0, then fwk g1
kD0 is approximately regular with respect to
the family of cutters U WD fUi gi 2I .
Proof. Let wki D 0 for all i … Ik , k 0. Define ik WD argmaxi 2Ik Vik x k x k ,
2
k 0. Suppose that limk !lkk Vlkk x k x k D 0. Denote Ci WD Fix Ui , i 2 I . By
property (5.56), we have
lim max PCi x k x k D 0. (5.57)
k i 2Ik
Corollary 2.1.37, the assumption Fix Vik Fix Ui D Ci for all i 2 I and k 0,
and the definition of the metric projection yield
k
V y y
P k y y
kPCi y yk ,
i Fix V i
y 2 H, consequently,
k k
V x x k D max V k x k x k max PC x k x k . (5.58)
ik i i
i 2Ik i 2Ik
238 5 Projection Methods
k 1
kCr rX kClC1
x k x k x x kCl
lD0
k 1
rX X
D
kCl kCl kCl kCl
!j .Vj x kCl
x /
lD0 j 2IkCl
k 1
rX X
kCl !jkCl VjkCl x kCl x kCl
lD0 j 2IkCl
X
p1
X
kCl
kCl !jkCl VikCl
kCl
x kCl
x
lD0 j 2IkCl
X
p1
kCl
D kCl VikCl
kCl
x kCl
x !0
lD0
as k ! 1, i.e.,
lim x kCrk x k D 0. (5.60)
k
Further, by the definition of the metric projection and by the triangle inequality,
PC x k x k PC x kCrk x k PC x kCrk x kCrk C x kCrk x k .
i i i
(5.61)
Since i 2 IkCrk , equality (5.57) yields limk PCi x kCrk x kCr
k D 0. Now
2
k 0. Suppose that limk !ikk Vikk x k x k D 0. By equivalence (5.48), we have
lim !ikk Vikk x k x k D 0. (5.64)
k
k 1
kCr rX kClC1
x k x k x x kCl
lD0
k 1
rX X
D
kCl kCl kCl kCl
!j Vj x kCl
x
lD0 j 2IkCl
k 1
rX X
kCl !jkCl VjkCl x kCl x kCl
lD0 j 2IkCl
X
p1
kCl kCl
m kCl !ikCl
kCl
VikCl x x kCl ! 0
lD0
240 5 Projection Methods
as k ! 1, consequently,
lim x kCrk x k D 0. (5.65)
k
Further, by the definition of the metric projection and by the triangle inequality, we
have
PC x k x k PC x kCrk x k (5.66)
i i
PCi x kCrk x kCrk C x kCrk x k ,
i.e., limk VikCrk x kCrk x kCrk D 0. By condition (5.63),
lim PCi x kCrk x kCrk D 0:
k
Inequalities (5.66) and equality
(5.65) imply
now limk PCi x k x k D 0. Since
Ui are cutters, we have limk Ui x k x k D 0 (see Corollary 2.1.37), i 2 I , i.e.,
condition (5.46) is satisfied, which means that fwk g1
kD0 is approximately regular. ut
Remark 5.8.12. Since a sequence generated by (5.54) is bounded as a Fejér
monotone sequence, it has a weak cluster point y . Therefore, property (5.63) and
the demi-closedness of PFix Ui Id at 0 yields that y 2 Fix Ui . This leads to
the assumption that algorithm (5.54) is focusing (see [22, Definition 3.7], where
this property was applied to firmly nonexpansive operators). Furthermore, property
(5.63) applied to operators Vik being metric projections leads to the assumption that
algorithm (5.54) is strongly focusing (see [22, Definition 4.8]).
Consider now the following recurrence
X
x kC1 D x k C k !ik .Ui x k x k /, (5.67)
i 2Ik
kC1 2 2 Xk X 2
x z x 0 z l .2 l / !il Ui x l x l .
lD0 i 2Il
Consequently,
1
X X 2
k .2 k / !ik Ui x k x k < 1.
kD0 i 2Ik
Since the sum of an absolutely convergent series does not depend on the order of
summands, we have
X
m X
2
k .2 k /!ik Ui x k x k
i D1 k2Ki
1
X X 2
D k .2 k / !ik Ui x k x k < 1.
kD0 i 2Ik
Therefore, X 2
k .2 k /!ik Ui x k x k < 1,
k2Ki
i 2 I , and
2
lim k .2 k /!ik Ui x k x k D 0,
k!1;k2Ki
242 5 Projection Methods
The operators V k are cutters (see Corollary 2.1.49), consequently, Tk are strongly
quasi-nonexpansive (see Theorem 2.1.39). By Theorem 2.1.26, we have
\ \
Fix Tk D Fix V k D Fix Vik Fix Ui ,
i 2Jk i 2I
5.8 Simultaneous Cutter Methods 243
consequently,
1
\ \
Fix Tk Fix Ui .
kD0 i 2I
Let " > 0 be such that lim infk k > " and lim infk .2 k / > " and let z 2
T
i 2I Fix Ui . For a sufficiently large k we have 2 k 2 . Now, it follows from
"
as k ! 1.
(i) Suppose that fwk g1 kD0 is approximately regular with respect to the family U WD
fUi gi 2I . Letik 2 Jk , k
0, be such that implication (5.46) holds. Then (5.69)
yields limk Ui x k x k D 0 for all i 2 I . Therefore, condition (3.11) is
satisfied for Tk WD V k and for S WD Ui , i 2 I . We have proved that all
assumptions of Corollary 3.7.1 (ii) are satisfied for X D H, Tk WD V k and for
S WD Ui , i 2 I . Therefore, x k converges weakly to a common fixed point of
Ui , i 2 I .
(ii) Suppose that H is finite dimensional and fwk g1 kD0 is approximately semi-
regular with respect to the family U WD fUi gi 2I . Let ik 2 Jk , k 0,
be such
that implication (5.47) holds. Then (5.69) yields lim infk Ui x k x k D 0 for
all i 2 I . Therefore, condition (3.12) is satisfied for Tk WD V k and for S WD Ui ,
i 2 I . We have proved that all assumptions of Theorem 3.7.1 (iii) are satisfied
for X D H, Tk WD V k and for S WD Ui , i 2 I . Therefore, x k converges to a
common fixed point of Ui , i 2 I . t
u
244 5 Projection Methods
Remark 5.8.16. (i) We can also consider recurrence (5.36), where k depends on
x 2 H, i.e.,
x0 2 H – arbitrary
P (5.70)
x kC1 D x k C k .x k /. i 2Jk !ik .x k /Vik x k x k /,
x0 2 H – arbitrary
P (5.71)
x kC1 D x k C i 2I ki .ik .x k /Vik x k x k /,
Note, however, that recurrence (5.71) can be reduced to (5.70) (or to (5.36)
if we apply constant weight functions vk ) if we substitute the sequence of
weight functions fvk g1 k 1
kD0 and the sequence of relaxation vectors f gkD0 to
k 1 1
the following sequences fw gkD0 W H !
m and fk gkD0 W H ! .0; 2,
where X
k .x/ D ki ik .x/ (5.72)
i 2I
and
!ik .x/ D ki ik .x/=k .x/; (5.73)
i 2 I and x 2 H (cf. Remark 4.4.3). Note that this transformation maintains
the assumptions on weight functions and on relaxation parameters presented
before. In particular, if lim infk ki " and lim supk ki 2 " for all i 2 I
and for some " 2 .0; 1/, then lim infk k .x/ " and lim supk k .x/ 2" for
all x 2 H. Furthermore, if vk is an approximately regular sequence of weight
functions and lim infk k .2 k / > 0, then wk is also approximately regular.
Therefore, we apply recurrence (5.36) instead of (5.71) in Theorem 5.8.15 in
order to simplify the notation.
(iii) If we replace the assumption on the demi-closedness of Ui Id at 0, i 2 I and
the assumption on the approximate regularity of fwk g1 kD0 in Theorem 5.8.15
by the following one:
k k k
k 2
T k !ik .x / Vik x x
lim k
D 0 H) all weak cluster points of fx k g1
kD0 lie in
i 2I Fix U i ,
then the theorem remains true (cf. Remark 3.7.5).
5.8 Simultaneous Cutter Methods 245
5.8.3 Examples
Example 5.8.17. (Iusem and De Pierro [215]) Consider recurrence (5.36), where
Jk D I , k D 1, Vik D PCi , for all k 0, Ci H are closed convex subsets,
i 2 I , the sequence fwk g1
kD0 of weight functions w W H !
m is constant, i.e.,
k
(
P i for i 2 I.x/
!i .x/ WD j 2I.x/ j
0 for i … I.x/,
where v D T .1 ; 2 ; : : : ; m / 2 ri
m and I.x/ WD fi 2 I W x … Ci g. Suppose
that C WD i 2I Ci ¤ ;. Since wk is regular (see Example 5.8.3 (ii)), it follows
from Theorem 5.8.15 that x k * x 2 C . The convergence was proved by Iusem
and de Pierro [215, Corollary 4] for H D Rn . Iusem and de Pierro have also
considered the case C D ; and have k
P proved the local convergence of x to a
the operator T WD
fixed point of [ i 2I !i PCi with Fix T ¤ ; for a starting
point x 0 2 B.z; 12 mini 2I kPCi z zk/, see [215, Theorem 3]). Note that
z2Fix T
this convergence can be proved simply as follows. If x 0 2 B.z; r/ for some
z 2 Fix T , where r WD 12 mini 2I kPCi z zk, then !i .x 0 / D i , i 2 I . The quasi
nonexpansivity of T (see Corollary 4.4.4) yields that x 1 2 B.z; r/. One can prove
by induction with respect to k that !i .x k / D i for all i 2 I and for all k. The
convergence follows now from Corollary 5.4.1.
Example 5.8.18. (Aharoni and Censor [3]) Consider recurrence (5.36), where
, Jk D I , Vik D PCi for closed convex subsets Ci Rn , i 2 I , with
H D RnT
C WDP i 2I Ci ¤ ;, k 2 Œ"; 2 " for some " 2 .0; 1/, k 0, wk 2
m
with 1kD0 !i D C1, i 2 I . By Theorem 4.4.5, we have
k
kC1 2 2 X
m
2
x z x k z k .2 k / !ik PCi x k x k
i D1
kC1 2 2 Xk X
m
2
x z x 0 z l .2 l / !il PCi x l x l .
lD0 i D1
Consequently,
X 1
m X
2
k .2 k /!ik PCi x k x k
i D1 kD0
1
X X
m
2
D k .2 k / !ik PCi x k x k < C1
kD0 i D1
246 5 Projection Methods
and
1
X 2
k .2 k /!ik PCi x k x k < C1
kD0
(see Example 5.8.3 (i)). Therefore, fwk g1 kD0 is approximately semi-regular (see
Lemma 5.8.7 (ii)). Theorem 5.8.15 (ii) yields now x k ! x 2 C . If we suppose
that all cluster points w of fwk g1
kD0 have the property !i ı for all i 2 I and for
some ı 2 .0; m1 , then fwk g1kD0 is regular (see Example 5.8.3 (iv)) and the weak
convergence holds in general Hilbert spaces H.
Example 5.8.20. (Flåm and Zowe [168]) Consider recurrence (5.36), where Jk D
Ti are cutters with Fix Vi
k k
I for all k, lim infk k > 0 and lim supk k < 2 and V
Ci for closed convex subsets Ci H with C WD i 2I Ci ¤ ;, satisfying the
inequality k k
V x x k ˛ PC x k x k , (5.74)
i i
(a) If fwk g1kD0 is approximately regular, then it follows from Theorem 5.8.15 (i)
that x k converges weakly to a point x 2 C .
(b) If fwk g1kD0 is approximately semi-regular and H is finite dimensional, then it
follows from Theorem 5.8.15 (ii) that x k converges to a point x 2 C .
Example 5.8.23. (Censor and Elfving [87]) Consider a consistent system of linear
inequalities Ax b and a method defined by the recurrence
x kC1 D x k C k .Tk x k x k /,
where Tk W Rn ! Rn is given by
1
Tk x WD x A> Mk .Ax b/C ,
max .A> Mk A/
and 1 1
Mk2 D U > Gk2 U: (5.76)
Denote y.x/ D .1 .x/; 2 .x/; : : : ; m .x//> D UAx. By Theorem 1.1.27, we have
2
12 2 12 2 > 12
max .A Mk A/ D Mk A
>
D sup
M Ax D sup
k
U G UAx
k
kxkD1 kxkD1
2
12 X
D sup G
k y.x/ D sup
gik .i .x//2
kxkD1 kxkD1 i 2I
X 1 2
sup gi .i .x//2 D sup G 2 y.x/
kxkD1 i 2I kxkD1
1 2 2
1
D sup G 2 UAx D sup U > G 2 UAx
kxkD1 kxkD1
1 2 1 2
D sup M 2 Ax D M 2 A D max .A> MA/.
kxkD1
1
S x WD x A> M.Ax k b/C .
max .A> MA/
Note that Mk .Ax k b/C D M.Ax k b/C . Because the system Ax b is equivalent
1 1 1
to M 2 Ax M 2 b, Lemma 4.6.2 yields that Fix S D fx 2 Rn W M 2 Ax
1
M 2 bg D C . Therefore,
1 >
Tk x k x k D A Mk .Ax k b/C
max .A> Mk A/
1 >
A Mk .Ax k b/C
>
max .A MA/
1 >
D A M.Ax k b/C
max .A> MA/
D S x k x k ,
Consider the convex feasibility problem: find a common point of a finite family
of nonempty closed convex subsets Ci H, i 2 I . If m is a large number, it is
convenient to split the set I into smaller blocks and to perform the operations of the
projection method in two stages: first on blocks and then on the family of blocks.
Let I D L1 [ L2 [ : : : [ Lr , where Lj are nonempty subsets which are pairwise
disjoint, j 2 J D f1; 2; : : : ; rg. Then each i 2 I can be transformed in a unique
way to a pair .j; l/ such that j 2 J and l 2 Lj . The subset Ci canTnow beTpresented
as Cj l and the convex feasibility problem obtains the form: x 2 j 2J l2Lj Cj l .
Consider the following two stage recurrence
where jk 2 J is a control for the outer loop and Ujkk is a relaxed simultaneous
projection which “employs” only projections PCjk ;l for l 2 Ljk , i.e., Ujkk has the
form X
Ujkk WD Id Ck . !ik PCjk ;l Id/, (5.78)
l2Ljk
Consider the two stage recurrence (5.77), where the control sequence:
fjk g1
kD0 J applied in the outer stage is an almost cyclic control, i.e.,
for some p r and, for all k, Ujkk is defined by (5.78), where the weight function
wk W H !
jLj j is ˛-regular with respect to the family fPCjk ;l gl2Ljk , where ˛ 2
k
.0; 1, i.e., there exists lk 2 Ljk such that
2 2
!lkk PCjk ;lk x k x k ˛ max PCjk ;l x k x k , (5.82)
l2Ljk
k 0. Examples of such weight functions were presented in Sect. 5.8.1. We call the
sequence f.jk ; wk /g1
kD0 defined above a double layer control. A special case of this
control is the following: ˛ D 1, k D 1, wk D elk 2 ext
jLj j and lk 2 Ljk is a
k
remotest set control, i.e., lk D lk .x k / 2 Ljk is such that
PCjk ;lk x k x k D max PCjk ;l x k x k . (5.83)
l2Ljk
The double layer control f.jk ; wk /g1kD0 can be identified with a sequence of
weight functions fvk g1
kD0 defined by (5.80).
Corollary 5.8.24. The double layer control is approximately regular with respect
to the family U WD fPCi gi 2I .
Proof. Recurrence (5.77), where Ujkk is defined by (5.78), is a special case of (5.55),
where Ik D Ljk , V k D U WD fPCi gi 2I . Furthermore, (5.82) yields condition (5.56).
Since fjk g1 1
kD0 is almost cyclic, fIk gkD0 is intermittent. The claim follows now from
Theorem 5.8.9. t
u
Theorem 5.8.25. Let C ¤ ; and x k be generated by recurrence (5.77) with a
double layer control and with a sequence of relaxation parameters fk g1 kD0 such
that lim inf k .2 k / > 0. Then for an arbitrary x 0 2 H the sequence fx k g1
kD0
converges weakly to a point x 2 C .
Proof. It follows from Corollary 5.8.24 that the double layer control is approxi-
mately regular. The claim follows now from Theorem 5.8.15 (i). t
u
Several block iterative projection methods are presented in [3, 53, 81, 229], [118,
Chap. 5], [108, Sect. 5.6], [5, 59, 159].
In this section we consider the common fixed point problem for a finite family of
cutters Ui W H ! H, i 2 I WD f1; 2; : : : ; mg, having a common fixed point, and
sequences generated by the following recurrence
x0 D x 2 H – arbitrary
(5.84)
x kC1 D x k C k .Uik x k x k /
5.9 Sequential Cutter Methods 251
where fk g1 1
kD0 .0; 2 is a sequence of relaxation parameters and fik gkD0 I is a
control sequence.
Definition 5.9.1. Let fx k g1
kD0 be generated by recurrence (5.84). A control
sequence fik g1
kD0 W H ! I is called:
(i) Approximately regular if the following implication holds
lim Uik x k x k D 0 H) lim Ui x k x k D 0 for all i 2 I , (5.85)
k k
sequential cutter method. In this section we present some control sequences which
are approximately (semi-)regular. The results presented in this section are special
cases of Theorems 5.8.11 and 5.8.14, where V k D U D fUi gi 2I and wk D eik .
Corollary 5.9.3. Let Ui W H ! H be cutters having the following property: for
any bounded sequence fy k g1
kD0 H
lim Ui y k y k D 0 H) lim PFix Ui y k y k D 0, (5.86)
k k
for a constant ı > 0, and for arbitrary i 2 I and y 2 H. Of course, the latter
property is satisfied for operators Ui being strict relaxations of metric projections
PCi onto nonempty closed convex subsets Ci , i 2 I . Therefore, Corollary 5.9.3
yields in particular that an almost cyclic control sequence applied to a successive
projection method (5.28), where lim infk k .2 k / > 0, is an approximately
remotest set control sequence (see Lemma 5.6.12 (i)).
Corollary 5.9.4. Let Ui W H ! H, i 2 I be cutters with a common fixed point and
fx k g1
kD0 be generated by recurrence (5.84), where lim infk k .2 k / > 0. If the
control sequence fik g1 1
kD0 is repetitive, then fik gkD0 is approximately semi-regular.
5.9.3 Examples
Example 5.9.5. (Aharoni, Berman and Censor [2]) Consider method (5.84), where
H is finite dimensional and the cutters Ui satisfy the demi-closedness principle,
i 2 I , (in [2] it was supposed that Ui have property (5.87), i 2 I ), k 2 Œ; 2 for
some 2 .0; 1/ and the control fik g1 kD0 is repetitive. By Corollary 5.9.4 the control
fik g1
kD0 is approximately semi-regular. Note that for a cutter Ui with property (5.87),
the operator Ui Id is closed at 0, because, by the nonexpansivity of the metric
projection, the operator PFix Ui Id is closed at 0, i 2 I (seeTLemma 3.2.5).
Therefore, Corollary 5.9.2 (ii) yields the convergence x k ! x 2 i 2I Fix Ui . The
convergence was proved by Aharoni et al. [2, Theorem 1] (see also [108, Sect. 5.5]),
where the method was called the .ı; /-algorithm. An interesting convergence result
for method (5.84) with k D 1 and Ui W Rn ! Rn being paracontractions
(continuous strictly quasi-nonexpansive operators) and with repetitive control was
obtained by Elsner et al. [164, Theorem 1] (see also [163, Theorem 1] for linear
case). This result is not covered by Corollaries 5.9.2 (ii) and 5.9.4. Conversely, the
5.10 Extrapolated Simultaneous Cutter Methods 253
also [331, Theorem 2] and [6], where a quasi-cyclic order is considered, which
is slightly more general than the almost cyclic control).
(ii) If H is finite dimensional and the control sequenceT fik g1
kD0 is repetitive, then
Corollaries 5.9.2 (ii) and 5.9.4 yield x ! x 2 i 2I Fix Ui (see also [22,
k
Example 3.14], [331, Theorem 1] and [6, Theorem 3.1] for related results).
Example 5.9.7. (Sequential subgradient projection method) Consider method
(5.84), where lim infk k .2 k / > 0 and Ui , i 2 I , are subgradient projections,
i.e., Ui D Pci , where ci W H ! R are continuous convex functions, i 2 I , which
T holds if, e.g., H D R ).
n
are globally Lipschitz continuous on bounded subsets (this
Denote Ci WD S.ci ; 0/, i 2 I , and suppose that C WD i 2I Ci ¤ ;. Since Ui are
cutters (see Corollary 4.2.6) and Ui Id are demi-closed at 0 (see Theorem 4.2.7),
it follows from Corollary 5.9.2 (i) that x k converges weakly to a point x 2 C if the
control sequence fik g1 1
kD0 is approximately regular (e.g., if fik gkD0 is almost cyclic).
Similarly, it follows from Corollary 5.9.2 (ii) that x converges to a point x 2 C ,
k
if H is finite dimensional and the control sequence fik g1 kD0 is approximately semi-
regular (e.g., if fik g1
kD0 is repetitive). A special case of the sequential subgradient
projection method was presented in [102], where H D Rn and the control is
almost cyclic. In [136] a finitely convergent modification of the method of [102]
was presented. In [80] the method of [102] was applied for solving interval linear
inequalities. For a survey on subgradient projection methods, see [79].
We consider the common fixed point problemT for a finite family of cutters Ui W
H ! H, i 2 I WD f1; 2; : : : ; mg, with i 2I Fix Ui ¤ ;. In this section we
study convergence of sequences which are generated by generalized relaxations of
simultaneous cutters of the form
where k 2 Œ0; 2, fk g1 kD0 W H ! .0; 1/ is a sequence of step size func-
tions, fvk g1 kD0 W H !
jJk j is a sequence of weight functions, i.e., vk .x/ D
.1 .x/; : : : ; jJk j .x// 2
jJk j and V k WD fVik gi 2Jk , k 0, is a sequence of families
k k
T T
of cutters with i 2Jk Fix Vik i 2I Fix Ui for all k 0. The iterative method
described by (5.89) is called an extrapolated simultaneous cutter method (ESCM).
If we take, in particular, V k WD U for all k 0, we obtain the following recurrence
X
k
x kC1 D x k C k k .x k /. ik .x k /Ui x k x k /, (5.90)
i 2I
Let V WD fVi gi 2J be a finite family of cutters with a common fixed point and v W
H !
jJ j be an appropriate weight function. Define a step size function v W H !
R by the equality
8 P
< i .x/kVi xxk2 T
P
i 2J
if x … i 2J Fix Vi ,
k i .x/Vi xx k
2
v .x/ WD i 2J
T
:1 if x 2 Fix Vi .
i 2J
well defined. Furthermore, it follows from the convexity of the function kk2 that
v .x/ 1 for all x 2 H.
Definition 5.10.1. Let Vi W H ! H, i 2 J , be cutters with a common fixed
point and v W H !
jJ j be a weight function which is appropriate with respect
to the family V WD fVi gi 2J . We say that a step size function W H ! .0; C1/
is ˛-admissible with respect to the family fVi gi 2J , where ˛ 2 .0; 1, or, shortly,
admissible, if
˛v .x/ .x/ v .x/ (5.91)
T
for all x … i 2J Fix Vi .
and
max Vik x x max kUi x xk (5.93)
i 2Jk i 2I
Then:
(i) If the sequence of weight functions fvk g1 kD0 applied to the sequence of families
V k is regular with respect to the family U WD fUi gi 2I , then x k converges weakly
to a common fixed point of Ui , i 2 I .
(ii) If the sequence of weight functions fvk g1 kD0 applied to the sequence of families
V k contains a subsequence of weight functions which is regular with respect to
the family U WD fUi gi 2I and H is finite dimensional, then x k converges to a
common fixed point of Ui , i 2 I .
Proof. (cf. [70, Theorem 9.35]) Let V k W H ! H be defined by
X
V k x WD ik .x/Vik x
i 2Jk
256 5 Projection Methods
kC1 2 2 2 k
x z x k z Tk x k x k 2
k
T
for all z 2 m Fix kUi and
i D1 all k k0 . Consequently, fx k g is bounded, x k z is
monotone and Tk x x k ! 0.
(i) Let ˇ > 0, k1 k0 and jk 2 f1; 2; : : : ; Jk g be such that
2
jkk .x/ Vjkk x x ˇ max kUi x xk2
i 2I
for any x 2 H and k k1 . Since k is ˛-admissible, the norm kk is a convex
function and Vjk x x maxi 2I kUi x xk for all j 2 Jk , we have
Tk x k x k D k k .x k / V k x k x k
P
k k k k
k 2
i 2Jk i .x / Vi x x
k ˛ P V k x k x k
k 2
i 2Jk i .x /Vi x x
k k k k
P k k k k
k 2
i 2Jk i .x / Vi x x
D k ˛ P
k .x k /V k x k x k
i 2Jk i i
2
jkk .x k / Vjkk x k x k
k ˛ P
k k k k
k
i 2Jk i .x / Vi x x
2
ˇ maxi 2I Ui x k x k
k ˛ P .
. i 2Jk ik .x k // maxi 2I Ui x k x k
"˛ˇ
D max Ui x k x k
i 2I
Tk x k x k "˛ˇ max Ui x k x k (5.95)
i 2I
and Ui x k x k ! 0 for all i 2 I . Therefore, condition (3.6) is satisfied
for Uk D Tk and S D Ui , i 2 I . We have proved that all assumptions of
Theorem 3.6.2 (i) are satisfied for S D Ui , i 2 I . Therefore, x k converges
weakly to a common fixed point of Ui , i 2 I .
(ii) Suppose that H is finite dimensional and fvk g1 kD0 contains a ˇ-regular subse-
quence fvnk g1
kD0 . Let ˇ > 0, k1 k0 and jnk 2 I be such that
2
jnnk .x/ Vjnnk x x ˇ max kUi x xk2 .
k k i 2I
"˛ˇ
kTnk x nk x nk k max kUi x nk x nk k .
i 2I
Consequently, lim infk Ui x k x k D 0 for all i 2 I . If we take Uk D Tk and
S D Ui , i 2 I , in Theorem 3.6.2 (ii), we obtain the convergence of x k to a
fixed point of Ui for all i 2 I . t
u
The operators Tk defined by (5.94) do not need to be continuous, because we have
not assumed that the operators Vik as well as the weight functions ik are continuous.
This allows application of operators from an essentially broader family than the
nonexpansive ones. We have supposed in Theorem 5.10.2 that fvk g1 kD0 is regular
(in (i)) or that fvk g1
kD0 contains a regular subsequence (in (ii)). These assumptions
seem not to be too restrictive. The assumption on the step sizes guarantees that Tk
is a relaxed cutter. All these assumptions allow using algorithms which can treat the
violated and the nonviolated constraints in a more flexible way and provide long
steps.
Remark 5.10.3. In [119, Sect. 1] and [120, Sects. IV and V] Combettes presented
three methods which are quite similar to ESCM: two extrapolated methods of
parallel projections (EMOPP and EPPM2) and an extrapolated method of parallel
subgradient projection (EMOPSP):
(i) In EMOPP (see [119, Definition 1.2]) it is supposed that Jk D I , Vjk D PC k ,
i
where PC k is the metric projection onto a closed convex superset Cik of Ci
i
with
d.x k ; Cik / d.x k ; Ci /, (5.96)
for all i 2 I , k 0, and for some constant > 0, and the weights vk 2
m
satisfy condition (5.42), k 0.
(ii) In EPPM2 (see [120, Sect. IV.C]) it is supposed that Jk D I , Vjk D PC k , where
i
PC k is the metric projection onto a closed and convex superset Cik of Ci , i 2 I ,
i
258 5 Projection Methods
k 0, and v 2
m . The supersets Cik of Ci are chosen in such a way that the
following condition is satisfied: for any subsequence fx nk g1 k 1
kD0 fx gkD0 and
for all i 2 I it holds
.x nk * x and lim PC nk x nk x nk D 0/ H) x 2 Ci . (5.97)
k i
(iii) In EMOPSP (see [120, Sect. V]) it is supposed that Jk I , Jk ¤ ;, Vik D Pfi
are subgradient projections relative to a (lower semi-)continuous convex func-
tion fi W H ! R,
ik ı > 0 for xk … Ci ; i 2 Jk : (5.98)
The above methods EMOPP and EMOPSP are covered by ESCM, where Vjk are
cutters, j 2 Jk , k 0, satisfying (5.92) and (5.93) and fvk g1 kD0 is a sequence of
appropriate weight functions. Note that (5.96) with ik satisfying ik ı > 0 for
xk … Ci , i 2 Jk , implies that fvk g1 kD0 is a regular sequence of weight functions
(see Example 5.8.3 (iv) and Definition 5.8.4). In EPPM2 condition (5.97) is slightly
weaker (only in an infinite dimensional case) than the approximate regularity of vk
satisfying (5.98), but Theorem 5.10.2 enables application of more general sequences
of weight functions than the ones satisfying (5.98). Furthermore, Theorem 5.10.2
T not krequire Ci T Ci , i 2 Jk , k 0. The theorem requires only the inclusion
k
does
i 2Jk Ci C WD i 2I Ci to be satisfied, k 0.
.ci .x//C
k 0 (for simplicity, we use the convention that kgi .x/k
D 0 if .ci .x//C D 0). The
ESSPM is defined by
5.11 Extrapolated Cyclic Cutter Method 259
X .ci .x k //C
x kC1 D x k k k .x k / ik .x k / k
gi .x /, (5.99)
i 2I
gi .x k /2
where
Ukk ;k WD Id Ck k .U k Id/
is a generalized relaxation of the simultaneous subgradient projection
X
U k WD ik Pci ,
i 2I
k 0.
Theorem 5.10.4. Let x 0 2 H be arbitrary and x k be generated by recurrence
(5.99), where lim infk k .2 k / > 0.
(i) If fvk g1 k
kD0 is regular, then x converges weakly to a common fixed point of Ui ,
i 2 I.
(ii) If fvk g1
kD0 contains a subsequence of weight functions which is regular and H
is finite dimensional, then x k converges to a common fixed point of Ui , i 2 I .
Proof. The step sizes k are 1-admissible. Therefore, the theorem follows from
Theorem 5.10.2. t
u
The ESSPM was studied by Dos Santos in [146], where positive constant weights
w 2 ri
m were considered and the convergence in the finite dimensional case was
proved [146, Sect. 5]. Furthermore, the method was studied by Combettes in [120,
Sect. V] (cf. Remark 5.10.3 (iii)).
where k 2 .0; 2, the step size functions k W H ! .0; C1/ satisfy the inequality
(see again (4.122)) we can use in particular step size functions k satisfying the
inequality P
kSi x Si 1 xk2
m˛ i 2I k .x/ max .x/ (5.102)
kUx xk2
for all x 2 H. We can also write recurrence (5.100) in the form x kC1 D Uk ;k x k ,
where Uk ;k WD Id Ck k .U Id/ is a generalized relaxation of the cyclic cutter
U with the step size function k satisfying inequalities (5.101).
5.11.1 Convergence
In [71] one can find a slightly weaker version of the following theorem.
Theorem 5.11.1. Let fx k g1 kD0 be generated by (5.100), where k .x/ satisfies
(5.101) and let z 2 Fix U . Then the following inequality holds
kC1 2 2 2
x z x k z k .2 k /˛ 2 Ux k x k . (5.103)
Moreover:
(i) If U Id is demi-closed at 0 and lim infk k .2 k / > 0, then x k * x 2
Fix U .
(ii) If the step size function k satisfies inequalities (5.102), then
Pm
kC1 2 2 k 2 2
i D1 Si x Si 1 x
k
. /
x z x k z k .2 k /m2 ˛ 2 2 .
Ux k x k
(5.104)
If, furthermore, Ui Id are demi-closed at 0, i D 1; 2; : : : ; m, and lim infk k
.2 k / > 0, then x k * x 2 Fix U .
Proof. Since a T
cutter is strictly quasi-nonexpansive (see Theorem 2.1.39), we
have Fix U D i 2I Fix Ui . Note that Uk ;k x k x k D k k .x k /.Ux k x k /.
5.11 Extrapolated Cyclic Cutter Method 261
kC1 2 2 2 2 k
x z D U; x k z x k z U ; x k x k 2
k k
k
k 2 2
D x z k .2 k /k2 .x k / Ux k x k
2 2
x k z k .2 k /˛ 2 Ux k x k .
Therefore, x k z is decreasing, x k is bounded and
k
Ux x k ! 0. (5.105)
kU1 x nk x nk k D kS1 x nk S0 x nk k ! 0.
k.U1 x nk U1 x / .x nk x /k D kU1 x nk x nk k ! 0
Ux D Um : : : U1 x D x .
Even if we take WD max and D 2, the generalized relaxation U; needs not
to be an extrapolation of U because max needs not to be greater or equal to 12 .
We only know that max 2m1
(see (4.122)). Note, however, that U is m1 -SQNE (see
2m
Corollary 4.5.15), consequently, it is a mC1 -relaxed cutter (see Corollary 2.1.43) and
U 1Cm is a cutter (see Remark 2.1.3). Since Umax is a cutter (see Theorem 4.10.2),
2m
Lemma 2.4.7 (i) yields that U is also a cutter, where D maxf 1Cm 2m
; max g.
Therefore, U; with 2 . 1Cm ; 2 is an extrapolation of U .
2m
hai ; xi D ˇi ; i 2 I , (5.107)
hSi x; ai i D hUi ui 1 ; ai i D ˇi ,
i 2 I , consequently,
X
m X
m
hSi x x; Si x Si 1 xi D hSi x x; Ui ui 1 ui 1 i
i D1 i D1
X
m
D hSi x x; .ˇi hai ; ui 1 i/ai i
i D1
5.12 Surrogate Constraints Methods 263
X
m
D .ˇi hai ; ui 1 i/.hSi x; ai i hx; ai i/
i D1
X
m
D .ˇi hai ; xi/.ˇi hai ; ui 1 i/
i D1
where x … Fix U . It follows from Theorem 5.11.1 that for any x 0 2 H, the
sequence fx k g1
kD0 generated by the recurrence x
kC1
D Uk x k with k D max is
Fejér monotone with respect to C and converges weakly to a solution of the system
(5.107). Moreover, by Theorem 3.8.4, the sequence converges strongly.
Remark 5.11.2. For any u 2 H and z 2 Ci , we have hUi u u; zui D kUi u uk2 .
Therefore, it follows from the proof of Lemma 4.10.1 that in this case the first
inequality in (4.119) is, actually, an equality. Consequently, we have an equality
in (4.124) and the operator U with the step size function given by (4.123) has the
property hU x x; z U xi D 0 for all z 2 C , or, equivalently, hU x x; z xi D
kU x xk2 . Consequently, the step size function WD max is optimal in the
following sense:
One can expect that this property leads in practice to an acceleration of the
convergence to a solution of the system (5.107), of sequences generated by the
recurrence x kC1 D Umax x k in comparison to the classical Kaczmarz method.
A different acceleration scheme for the Kaczmarz method was presented in [247,
Sect. 3.1].
defined by (4.109) and Svk ;k WD Id Ck .Svk Id/ is the k -relaxation of the oper-
ator Svk , k 0. Recall that Svk x k is the projection of x k onto a surrogate constraint
Hk WD H .ak .x k /; ˇk .x k //, where ak .x k / WD A> vk .x k / and ˇk .x k / WD vk .x k /> b
(see Sect. 4.9.4) and that Svk is a cutter with Fix Svk D C (see Theorem 4.9.6).
Various versions of the surrogate projection method were studied by Merzlyakov
[260], Oko [277], Yang and Murty [351], Kiwiel [229–231], Kiwiel and Łopuch
[233], Cegielski [67, 68] and by Dudek [150].
It follows from Remark 5.8.6 (i) and from the equivalence of all norms in Rn that
condition (5.111) is equivalent to
lim vkik .x k /> .ai>k x k ˇik /C D 0 H) lim inf .Ax k b/C D 0. (5.112)
k k
Therefore, the assumption that a sequence of weight functions fvk g1 kD0 which
consider only violated constraints is approximately semi-regular is stronger than
5.12 Surrogate Constraints Methods 265
(5.110). This fact follows from a simple observation that if a sum of positive
quantities goes to zero, then all summands go to zero.
If the weight functions vk consider only violated constraints, k 0, then it
follows from Proposition 4.9.11 that recurrence (5.109) defines an extrapolated
simultaneous projection method. In this case one can apply Theorem 5.10.2 in
order to prove the convergence of a sequence generated by (5.109). In the next
subsection we prove that the convergence follows without assumptions that vk
considers only violated constraints and that the corresponding sequence of weight
functions fwk g1
kD0 (see (4.112)) is approximately semi-regular. It suffices to suppose
instead that fvk g1
kD0 is proper.
for any z 2 C . Consequently, x k z converges as a decreasing sequence, x k is
bounded and
vk .x k /> .Ax k b/
lim D 0.
k A> vk .x k /
This implies
lim inf PCi x k x k D lim inf.ai> x k ˇi /C D 0
k k
for all i 2 I . Corollary 3.7.1 (iii) for X D Rn , Tk WD Svk and S WD PCi yields now
the convergence of x k to x 2 Fix PCi D Ci for any i 2 I , i.e., x 2 C . t
u
266 5 Projection Methods
X
D !ik .x k /Œ.ai> x k ˇi /C 2
i 2I
Since fvk g1 k
kD0 is approximately semi-regular, lim infk PCi x x D 0, i.e.,
k
lim infk .ai> x k ˇi /C D 0 for all i 2 I , or, equivalently, lim infk .Ax k b/C D 0.
In particular, a constant sequence vk WD v is proper, where v W Rn ! Rm C nf0g is
defined by
v.x/ WD W .x/.Ax b/C ,
for x … C , W .x/ WD diag w.x/ and w has one of the following forms:
(i) w is constant with positive weights, i.e., w.x/ DW w 2 ri
m for all x 2 Rn .
Note that w is regular (see Example 5.8.3 (i)), consequently, it is approximately
regular (see Lemma 5.8.7 (i)). If we take w D . m1 ; m1 ; : : : ; m1 /, then we obtain
i .x/ D m1 .ai> x ˇi /C , i.e., the weights vi .x/ are proportional to the residua of
violated constraints. Since the rescaling of weights does not change the operator
Sv defined by (4.109), we can apply, equivalently, i .x/ WD .ai> x ˇi /C or
.ai> xˇi /C
i .x/ WD P > . The latter weights were proposed by Yang and Murty
j 2I .aj xˇj /C
[351, page 167].
(ii) w considers only almost violated constraints, i.e., !j .x/ D 0 for all j 2 I such
that
.aj> x ˇj /C < max.ai> x ˇi /C ,
i 2I
where 2 .0; 1. Note that w is approximately regular (see Example 5.8.3 (v)
and Lemma 5.8.7 (i)). If we take D 1, then the surrogate constraints method
reduces to a successive projection method with the remotest set control.
5.12 Surrogate Constraints Methods 267
1
max.ai> x nk ˇi /C .
m i 2I
jection method with the remotest set control is a special case of the method with
D 1.
Example 5.12.5. Let fx k g1 kD0 be generated by the surrogate constraints method
(5.109), where fvk g1
kD0 is a sequence of essential weight functions containing a
subsequence fvnk g1
kD0 satisfying the following condition
ı for some i 2 Argmaxi 2I .ai> x nk ˇi /
ink .x nk / (5.113)
D 0 for i … I.x nk /,
Then we have
X
vnk .x nk /> .Ax nk b/ D ink .x nk /.ai> x nk ˇi /
i 2I
X
D ink .x nk /.ai> x nk ˇi /C
i 2I.x nk /
268 5 Projection Methods
ı max.ai> x nk ˇi /C .
i 2I
Remark 5.12.6. Yang and Murty proved the finite convergence of sequences
fx k g1
kD0 generated by the surrogate constraints method (5.109) with weight func-
tions vk satisfying the condition
ı for i 2 I" .x k /
ik .x k / (5.114)
D 0 for i … I" .x k /,
> k
T constantsn "; ı>> 0, where I" .x/ WD fi 2 I W ai x ˇi > "g
for predefined small
and x … C" WD i 2I fx 2 R W ai x ˇi > "g (see [351, Theorem 3.3]) Note that
k
if a sequence fvk g1 nk 1
kD0 satisfies condition (5.114), then any subsequence fv gkD0
satisfies condition (5.113) for all x … C" . We leave it to the reader to check that
nk
the result of [351, Theorem 3.3] can be derived from Example 5.12.5.
We start with some relationships between surrogate projections and metric projec-
tions onto solution sets of equality and inequality systems. Let x … C be a current
approximation of the solution, L WD L.x/ I be a nonempty subset of indices (a
constraints selection) and let and r WDj L j. Let AL denote the submatrix of A with
rows ai , i 2 L, and bL —the subvector of b with coordinates ˇi , i 2 L. Further,
denote \
CL WD Ci D fu 2 Rn W AL u bL g
i 2L
5.13 SCM with Residual Selection 269
and
HL WD fu 2 Rn W AL u D bL g.
Suppose, for simplicity, that L D f1; 2; : : : ; rg, i.e.,
AL bL vL .x/
AD ,bD and v.x/ D .
AI nL bI nL vI nL .x/
Denote
vL .x/> .AL x bL / >
SvL x WD x > AL vL .x/
A vL .x/2
L
(cf. (4.109)).
Lemma 5.13.1. Let x … C and L WD L.x/ I be such that AL has full row rank.
Then PCL x D PHL x if and only if
.AL A> 1
L / .AL x bL / 0.
minimize 12 ku xk2
subject to AL u D bL (5.115)
u 2 Rn
uN D x A> > 1
L .AL AL / .AL x bL /
yN D .AL A> 1
L / .AL x bL /
minimize 12 ku xk2
subject to AL u bL (5.116)
u 2 Rn
uN D x A> > 1
L .AL AL / .AL x bL /. (5.117)
270 5 Projection Methods
uN D x A>
L y D 0.
y D .AL A> 1
L / .AL x bL /
vL D ˛.AL A> 1
L / .AL x bL / (5.118)
for some ˛ > 0. Consequently, if any of both conditions is satisfied, then SvL x D
PCL x.
Proof. ((H) Let vL be defined by (5.118). Applying (4.109) and (4.5) with A WD
AL and v.x/ WD vL .x/, we obtain
v>
L .AL x bL / >
S vL x D x > 2 AL vL
A vL
L
v>
L .AL x bL / >
Dx AL vL
v> >
L AL AL vL
D x A> > 1
L .AL AL / .AL x bL / D PHL x.
(H)) Suppose that SvL x D PHL x. By (4.5) and (4.109) with A WD AL , we have
v>
L .AL x bL / > > > 1
> 2 AL vL D AL .AL AL / .AL x bL /.
A vL
L
vL D ˛.AL A> 1
L / .AL x bL /,
5.13 SCM with Residual Selection 271
where > 2
A vL
L
˛D
v>
L .AL x bL /
Now we describe one iteration of the surrogate constraints method with residual
selection. Suppose that x k is obtained in the kth iteration. The next iteration consists
of two steps:
Step 1 (Constraints selection). Determine a subset Lk WD L.x k / such that
(a) AL has full row rank and
(b) The vector
ai> x k ˇi
ik WD argmaxf W i 2 I g, (5.121)
kai k
vL WD .AL A> 1
L / .AL x b/
Recall that AC
L denotes the Moore–Penrose pseudoinverse of AL which for a full
row matrix AL has the form AC > > 1
L D AL .AL AL / . The proof of the following
Lemma can be found in [65, Corollary 2.7].
5.13 SCM with Residual Selection 273
Lemma 5.13.4. Let A> L0 w > 0 for some w 2 R . If AL has full row rank,
n
C
.AL A>
L/
1
0 and ai> AL 0, then AL0 has full row rank and .AL0 A>
L0 /
1
0.
Let x D x k … C . If L D flg, where l 2 K, then, of course, AL has full row rank
(as a nonzero vector al ) and .AL A>
L/
1
D kal k2 0. If there is i 2 K such that
>
ai al 0, then
2
ai> AC >
L D ai al kal k 0.
Applying Lemma 5.13.4 we see that L0 has the same properties as L: the matrix AL0
has full row rank and .AL0 A> L0 /
1
0. By repeated application of Lemma 5.13.4
with L WD L we obtain a new subset L0 K such that AL0 has full row rank and
0
.AL0 A>L0 /
1
0.
Let L WD L.x/ be constructed by an obtuse cone selection. Since AL x > bL ,
the vector vL WD .AL A> 1
L / .AL x b/ has nonnegative coordinates. Therefore, all
assumptions of Lemma 5.13.2 are satisfied, consequently, SvL x D PHL x D PCL x.
If in each iteration of (5.120) we start the above defined construction of the subset
a> x k ˇ
Lk WD L.x k / with lk WD argmaxf i kai k i W i 2 I g, then Theorem 5.13.3 yields the
convergence of sequences generated by (5.120) to a solution of the system Ax b.
Now we present a special case of the obtuse cone selection. The corollary below
follows from Lemma 5.13.4. We use the same notation as in Sect. 5.13.3.
Corollary 5.13.5. Let A>
L0 w > 0 for some w 2 R . If AL has full row rank,
n
.AL AL / 0 and AL ai 0, then AL0 has full row rank and .AL0 A>
> 1
L0 /
1
0.
Proof. Suppose that AL has full row rank, .AL A>
L/
1
0 and AL ai 0. Then, by
Lemma 5.13.4, we have
5.14 Exercises
Exercise 5.14.1. Show that Theorem 5.1.5 remains true if A and B are closed
affine subspaces with a common point.
Exercise 5.14.2. Show that the result of [351, Theorem 3.3] can be derived from
Example 5.12.5.
Exercise 5.14.3. Prove equalities (5.75) and (5.76).
References
1. S. Agmon, The relaxation method for linear inequalities. Can. J. Math. 6, 382–392 (1954)
2. R. Aharoni, A. Berman, Y. Censor, An interior point algorithm for the convex feasibility
problem. Adv. Appl. Math. 4, 479–489 (1983)
3. R. Aharoni, Y. Censor, Block-iterative projection methods for parallel computation of
solutions to convex feasibility problems. Lin. Algebra Appl. 120, 165–175 (1989)
4. D. Alevras, M.W. Padberg, Linear Optimization and Extensions. Problems and Solutions
(Springer, Berlin, 2001)
5. A. Aleyner, S. Reich, Block-iterative algorithms for solving convex feasibility problems in
Hilbert and in Banach spaces. J. Math. Anal. Appl. 343, 427–435 (2008)
6. A. Aleyner, S. Reich, Random products of quasi-nonexpansive mappings in Hilbert space.
J. Convex Anal. 16, 633–640 (2009)
7. M. Altman, On the approximate solution of linear algebraic equations. Bulletin de l’Académie
Polonaise des Sciences Cl. III 3 , 365–370 (1957)
8. I. Amemiya, T. Ando, Convergence of random products of contractions in Hilbert space. Acta
Sci. Math. (Szeged) 26, 239–244 (1965)
9. R. Ansorge, Connections between the Cimmino-method and the Kaczmarz-method for
solution of singular and regular systems of equations. Computing 33, 367–375 (1984)
10. G. Appleby, D.C. Smolarski, A linear acceleration row action method for projecting onto
subspaces. Electron. Trans. Numer. Anal. 20, 253–275 (2005)
11. N. Aronszajn, Theory of reproducing kernels. Trans. Am. Math. Soc. 68, 337–404 (1950)
12. A. Auslender, Optimisation, méthodes numériques (Masson, Paris, 1976)
13. V.N. Babenko, Convergence of the Kaczmarz projection algorithm. Zh. Vychisl. Mat. Mat.
Fiz. 24, 1571–1573 (1984) (in Russian)
14. J.B. Baillon, R.E. Bruck, S. Reich, On the asymptotic behavior of nonexpansive mappings
and semigroups in Banach spaces. Houston J. Math. 4, 1–9 (1978)
15. S. Banach, Sur les opérations dans les ensembles abstraits et leur application aux équations
intégrals. Fundamenta Mathematicae 3, 133–181 (1922)
16. H.H. Bauschke, A norm convergence result on random products of relaxed projections in
Hilbert space. Trans. Am. Math. Soc. 347, 1365–1373 (1995)
17. H.H. Bauschke, Projection Algorithms and Monotone Operators, Ph.D. Thesis, Department
of Mathematics, Simon Fraser, University, Burnaby, BC, Canada, 1996
18. H.H. Bauschke, The approximation of fixed points of compositions of nonexpansive mapping
in Hilbert space. J. Math. Anal. Appl. 202, 150–159 (1996)
19. H.H. Bauschke, The composition of the projections onto closed convex sets in Hilbert space
is asymptotically regular. Proc. Am. Math. Soc. 131, 141–146 (2002)
A. Cegielski, Iterative Methods for Fixed Point Problems in Hilbert Spaces, 275
Lecture Notes in Mathematics 2057, DOI 10.1007/978-3-642-30901-4,
© Springer-Verlag Berlin Heidelberg 2012
276 References
20. H.H. Bauschke, J. Borwein, On the convergence of von Neumann’s alternating projection
algorithm for two sets. Set-Valued Anal. 1, 185–212 (1993)
21. H.H. Bauschke, J. Borwein, Dykstra’s alternating projection algorithm for two sets. J. Approx.
Theor. 79, 418–443 (1994)
22. H.H. Bauschke, J. Borwein, On projection algorithms for solving convex feasibility problems.
SIAM Rev. 38, 367–426 (1996)
23. H.H. Bauschke, J.M. Borwein, A.S. Lewis, The method of cyclic projections for closed
convex sets in Hilbert space. Contemp. Math. 204, 1–38 (1997)
24. H.H. Bauschke, P.L. Combettes, A weak-to-strong convergence principle for Fejér-monotone
methods in Hilbert spaces. Math. Oper. Res. 26, 248–264 (2001)
25. H.H. Bauschke, P.L. Combettes, S.G. Kruk, Extrapolation algorithm for affine-convex
feasibility problems. Numer. Algorithms 41, 239–274 (2006)
26. H.H. Bauschke, P.L. Combettes, D.R. Luke, Phase retrieval, error reduction algorithm, and
Fienup variants: A view from convex optimization. J. Opt. Soc. Am. A 19, 1334–1345 (2002)
27. H.H. Bauschke, P.L. Combettes, D.R. Luke, Hybrid projection-reflection method for phase
retrieval. J. Opt. Soc. Am. A 20, 1025–1034 (2003)
28. H.H. Bauschke, P.L. Combettes, D.R. Luke, Finding best approximation pairs relative to two
closed convex sets in Hilbert spaces. J. Approx. Theor. 127, 178–192 (2004)
29. H.H. Bauschke, P.L. Combettes, D.R. Luke, A strongly convergent reflection method for
finding the projection onto the intersection of two closed convex sets in a Hilbert space.
J. Approx. Theor. 141, 63–69 (2006)
30. H.H. Bauschke, F. Deutsch, H. Hundal, S-H. Park, Accelerating the convergence of the
method of alternating projections. Trans. Am. Math. Soc. 355, 3433–3461 (2003)
31. H.H. Bauschke, S.G. Kruk, Reflection-projection method for convex feasibility problems with
an obtuse cone. J. Optim. Theor. Appl. 120, 503–531 (2004)
32. H.H. Bauschke, E. Matoušková, S. Reich, Projection and proximal point methods: conver-
gence results and counterexamples. Nonlinear Anal. 56, 715–738 (2004)
33. M.H. Bazaraa, H.D. Sherali, C.M. Shetty, Nonlinear Programming, Theory and Algorithms,
3rd edn. (Wiley, Hoboken, 2006)
34. M. Benzi, C.D. Meyer, A direct projection method for sparse linear systems. SIAM J. Sci.
Comput. 16, 1159–1176 (1995)
35. A. Berman, R.J. Plemmons, Nonnegative Matrices in the Mathematical Sciences (Academic,
New York, 1979)
36. V. Berinde, Iterative Approximation of Fixed Points (Springer, Berlin, 2007)
37. M. Bertero, P. Boccacci, Introduction to Inverse Problems in Imaging (Institute of Physics
Publishing, Bristol, 1998)
38. D.P. Bertsekas, Nonlinear Programming (Athena Scientific, Belmont, 1995)
39. D. Blatt, A.O. Hero, Energy-based sensor network source localization via projection onto
convex sets (POCS). IEEE Trans. Signal Process. 54, 3614–3619 (2006)
40. J.M. Borwein, A. Lewis, Convex Analysis and Nonlinear Optimization, Theory and Examples
(Springer, New York, 2000)
41. R. Bramley, A. Sameh, Row projection methods for large nonsymmetric linear systems.
SIAM J. Sci. Stat. Comput. 13, 168–193 (1992)
42. L.M. Bregman, Finding the common point of convex sets by the method of successive
projection (in Russian). Dokl. Akad. Nauk SSSR 162, 487–490 (1965); English translation
in: Soviet Math. Dokl. 6, 688–692 (1965)
43. L.E.J. Brouwer, Über Abbildung von Mannigfaltigkeiten. Math. Ann. 71, 97–115 (1912)
44. F.E. Browder, Fixed-point theorems for noncompact mappings in Hilbert space. Proc. Nat.
Acad. Sci. USA 53, 1272–1276 (1965)
45. F.E. Browder, Nonexpansive nonlinear operators in a Banach space. Proc. Nat. Acad. Sci.
USA 54, 1041–1044 (1965)
46. F.E. Browder, Convergence of approximants to fixed points of nonexpansive nonlinear
mappings in Banach spaces. Arch. Rational Mech. Anal. 24, 82–90 (1967)
References 277
47. F.E. Browder, Convergence theorems for sequences of nonlinear operators in Banach spaces.
Math. Zeitschr. 100, 201–225 (1967)
48. F.E. Browder, W.V. Petryshyn, The solution by iteration of nonlinear functional equations in
Banach spaces. Bull. Am. Math. Soc. 72, 571–575 (1966)
49. R.E. Bruck, Nonexpansive projections on subsets of Banach spaces. Pac. J. Math. 47, 341–355
(1973)
50. R.E. Bruck, Random products of contractions in metric and Banach spaces. J. Math. Anal.
Appl. 88, 319–332 (1982)
51. R.E. Bruck, S. Reich, Nonexpansive projections and resolvents of accretive operators in
Banach spaces. Houston J. Math. 3, 459–470 (1977)
52. R.S. Burachik, J.O. Lopes, B.F. Svaiter, An outer approximation method for the variational
inequality problem. SIAM J. Contr. Optim. 43, 2071–2088 (2005)
53. D. Butnariu, Y. Censor, On the behavior of a block-iterative projection method for solving
convex feasibility problems. Int. J. Comp. Math. 34, 79–94 (1990)
54. D. Butnariu, Y. Censor, P. Gurfil, E. Hadar, On the behavior of subgradient projections
methods for convex feasibility problems in Euclidean spaces. SIAM J. Opt. 19, 786–807
(2008)
55. C. Byrne, Iterative oblique projection onto convex sets and the split feasibility problem.
Inverse Probl. 18, 441–453 (2002)
56. C. Byrne, A unified treatment of some iterative algorithms in signal processing and image
reconstruction. Inverse Probl. 20, 103–120 (2004)
57. C.L. Byrne, Applied Iterative Methods (AK Peters, Wellesley, 2008)
58. C.L. Byrne, Bounds on the largest singular value of a matrix and the convergence of
simultaneous and block-iterative algorithms for sparse linear systems. Int. Trans. Oper. Res.
16, 465–479 (2009)
59. T.D. Capricelli, P.L. Combettes, Parallel block-iterative reconstruction algorithms for binary
tomography. Electron. Notes Discr. Math. 20, 263–280 (2005)
60. G. Casssiani, G. Böhm, A. Vesnaver, R. Nicolich, A geostatistical framework for incorporat-
ing seismic tomography auxiliary data into hydraulic conductivity estimation. J. Hydrol. 206,
58–74 (1998)
61. J. Cea, Optimisation: théorie et algorithmes (Dunod, Paris, 1971); Polish translation:
Optymalizacja: Teoria i algorytmy (PWN, Warszawa, 1976)
62. A. Cegielski, Relaxation Methods in Convex Optimization Problems (in Polish). Monographs,
vol. 67, Institute of Mathematics, Higher College of Engineering, Zielona Góra, 1993
63. A. Cegielski, in Projection Onto an Acute Cone and Convex Feasibility Problems, ed. by
J. Henry i J.-P. Yvon. Lecture Notes in Control and Inform. Sci., vol. 197 (Springer, London,
1994), pp. 187–194
64. A. Cegielski, A method of projection onto an acute cone with level control in convex
minimization. Math. Program. 85, 469–490 (1999)
65. A. Cegielski, Obtuse cones and Gram matrices with nonnegative inverse. Lin. Algebra Appl.
335, 167–181 (2001)
66. A. Cegielski, A generalization of the Opial’s theorem. Contr. Cybern. 36, 601–610 (2007)
67. A. Cegielski, Convergence of the projected surrogate constraints method for the linear split
feasibility problems. J. Convex Anal. 14, 169–183 (2007)
68. A. Cegielski, Projection methods for the linear split feasibility problems. Optimization 57,
491–504 (2008)
69. A. Cegielski, Generalized relaxations of nonexpansive operators and convex feasibility
problems. Contemp. Math. 513, 111–123 (2010)
70. A. Cegielski, Y. Censor, in Opial-Type Theorems and the Common Fixed Point Problem, ed.
by H.H. Bauschke, R.S. Burachik, P.L. Combettes, V. Elser, D.R. Luke, H. Wolkowicz. Fixed-
Point Algorithms for Inverse Problems in Science and Engineering, Springer Optimization
and Its Applications, vol. 49 (Springer, New York, 2011), pp. 155–183
71. A. Cegielski, Y. Censor, Extrapolation and local acceleration of an iterative process for
common fixed point problems. J. Math. Anal. Appl. 394, 809–818 (2012)
278 References
72. A. Cegielski, R. Dylewski, Selection strategies in projection methods for convex minimization
problems. Discuss. Math. Differ. Incl. Contr. Optim. 22, 97–123 (2002)
73. A. Cegielski, R. Dylewski, Residual selection in a projection method for convex minimization
problems. Optimization 52, 211–220 (2003)
74. A. Cegielski, R. Dylewski, Variable target value relaxed alternating projection method.
Comput. Optim. Appl. 47, 455–476 (2010)
75. A. Cegielski, A. Suchocka, Incomplete alternating projection method for large inconsistent
linear systems. Lin. Algebra Appl. 428, 1313–1324 (2008)
76. A. Cegielski, A. Suchocka, Relaxed alternating projection methods. SIAM J. Optim. 19,
1093–1106 (2008)
77. A. Cegielski, R. Zalas, Methods for variational inequality problem over the intersection of
fixed point sets of quasi-nonexpansive operators (2012) Numer. Funct. Anal. Optim. (in print)
78. Y. Censor, Row-action methods for huge and sparse systems and their applications. SIAM
Rev. 23, 444–466 (1981)
79. Y. Censor, Iterative methods for convex feasibility problems. Ann. Discrete Math. 20, 83–91
(1984)
80. Y. Censor, An automatic relaxation method for solving interval linear inequalities. J. Math.
Anal. Appl. 106, 19–25 (1985)
81. Y. Censor, Parallel application of block-iterative methods in medical imaging and radiation
therapy. Math. Program. 42, 307–325 (1988)
82. Y. Censor, Binary steering in discrete tomography reconstruction with sequential and
simultaneous iterative algorithms. Lin. Algebra Appl. 339, 111–124 (2001)
83. Y. Censor, M.D. Altschuler, W.D. Powlis, On the use of Cimmino’s simultaneous projections
method for computing a solution of the inverse problem in radiation therapy treatment
planning. Inverse Probl. 4, 607–623 (1988)
84. Y. Censor, A. Ben-Israel, Y. Xiao, J.M. Galvin, On linear infeasibility arising in intensity-
modulated radiation therapy inverse planning. Lin. Algebra Appl. 428, 1406–1420 (2008)
85. Y. Censor, T. Bortfeld, B. Martin, A. Trofimov, A unified approach for inversion problems in
intensity-modulated radiation therapy. Phys. Med. Biol. 51, 2353–2365 (2006)
86. Y. Censor, P.P.B. Eggermont, D. Gordon, Strong underrelaxation in Kaczmarz’s method for
inconsistent systems. Numer. Math. 41, 83–92 (1983)
87. Y. Censor, T. Elfving, New methods for linear inequalities. Lin. Algebra Appl. 42, 199–211
(1982)
88. Y. Censor, T. Elfving, A multiprojection algorithm using Bregman projections in a product
space. Numer. Algorithm 8, 221–239 (1994)
89. Y. Censor, T. Elfving, Block-iterative algorithms with diagonal scaled oblique projections for
the linear feasibility problems. SIAM J. Matrix Anal. Appl. 24, 40–58 (2002)
90. Y. Censor, T. Elfving, Iterative algorithms with seminorm-induced oblique projections. Abstr.
Appl. Anal. 8, 387–406 (2003)
91. Y. Censor, T. Elfving, G.T. Herman, in Averaging Strings of Sequential Iterations for Convex
Feasibility Problems, ed. by D. Butnariu, Y. Censor, S. Reich. Inherently Parallel Algo-
rithms in Feasibility and Optimization and their Applications (Elsevier, Amsterdam, 2001),
pp. 101–113
92. Y. Censor, T. Elfving, G.T. Herman, T. Nikazad, On diagonally relaxed orthogonal projection
methods. SIAM J. Sci. Comput. 30, 473–504 (2008)
93. Y. Censor, T. Elfving, N. Kopf, T. Bortfeld, The multiple-sets split feasibility problem and its
applications for inverse problems. Inverse Probl. 21, 2071–2084 (2005)
94. Y. Censor, A. Gibali, Projections onto super-half-spaces for monotone variational inequality
problems in finite-dimensional spaces. J. Nonlinear Convex Anal. 9, 461–475 (2008)
95. Y. Censor, A. Gibali, S. Reich, The subgradient extragradient method for solving variational
inequalities in Hilbert space. J. Optim. Theory Appl. 148, 318–335 (2011)
96. Y. Censor, A. Gibali, S. Reich, Extensions of Korpelevich’s extragradient method for the
variational inequality problem in Euclidean space. Optimization 61, 1119–1132 (2012)
References 279
122. P.L. Combettes, Solving monotone inclusions via compositions of nonexpansive averaged
operators. Optimization 53, 475–504 (2004)
123. P.L. Combettes, P. Bondon, Hard-constrained inconsistent signal feasibility problems. IEEE
Trans. Signal Process. 47, 2460–2468 (1999)
124. P.L. Combettes, S.A. Hirstoaga, Equilibrium programming in Hilbert spaces. J. Nonlinear
Convex Anal. 6, 117–136 (2005)
125. P.L. Combettes, H. Puh, Iterations of parallel convex projections in Hilbert spaces. Numer.
Funct. Anal. Optim. 15, 225–243 (1994)
126. G. Crombez, A geometrical look at iterative methods for operators with fixed points. Numer.
Funct. Anal. Optim. 26, 157–175 (2005)
127. G. Crombez, A hierarchical presentation of operators with fixed points on Hilbert spaces.
Numer. Funct. Anal. Optim. 27, 259–277 (2006)
128. Y.-H. Dai, Fast algorithms for projection on an ellipsoid. SIAM J. Optim. 16, 986–1006
(2006)
129. A. Dax, A note of the convergence of linear stationary iterative process. Lin. Algebra Appl.
129, 131–142 (1990)
130. A. Dax, Linear search acceleration of iterative methods. Lin. Algebra Appl. 130, 43–63 (1990)
131. A. Dax, On hybrid acceleration of a linear stationary iterative process. Lin. Algebra Appl.
130, 99–110 (1990)
132. A. Dax, The convergence of linear stationary iterative processes for solving singular
unstructured systems of linear equations. SIAM Rev. 32, 611–635 (1990)
133. L. Debnath, P. Mikusiński, Hilbert Spaces with Applications, 2nd edn. (Academic, San Diego,
1999)
134. A.R. De Pierro, A.N. Iusem, A simultaneous projections method for linear inequalities. Lin.
Algebra Appl. 64, 243–253 (1985)
135. A.R. De Pierro, A.N. Iusem, A parallel projection method of finding a common point of a
family of convex sets. Pesquisa Operacional 5, 1–20 (1985)
136. A.R. De Pierro, A.N. Iusem, A finitely convergent “row-action” method for the convex
feasibility problem. Appl. Math. Optim. 17, 225–235 (1988)
137. A.R. De Pierro, A.N. Iusem, On the asymptotic behavior of some alternate smoothing series
expansion iterative methods. Lin. Algebra Appl. 130, 3–24 (1990)
138. F. Deutsch, in Applications of von Neumann’s Alternating Projections Algorithm, ed. by
P. Kenderov. Mathematical Methods in Operations Research (Sophia, Bulgaria, 1983), pp.
44–51
139. F. Deutsch, in The Method of Alternating Orthogonal Projections, ed. by S.P. Singh. Approx-
imation Theory, Spline Functions and Applications (Kluwer Academic, The Netherlands,
1992), pp. 105–121
140. F. Deutsch, Best Approximation in Inner Product Spaces (Springer, New York, 2001)
141. F. Deutsch, in Accelerating the Convergence of the Method of Alternating Projections via
a Line Search: A Brief Survey, ed. by D. Butnariu, Y. Censor, S. Reich. Inherently Parallel
Algorithms in Feasibility and Optimization and their Application, Studies in Computational
Mathematics, vol. 8 (Elsevier Science, Amsterdam, 2001), pp. 203–217
142. F. Deutsch, H. Hundal, The rate of convergence for the cyclic projections algorithm, I. Angles
between convex sets. J. Approx. Theor. 142, 36–55 (2006)
143. F. Deutsch, H. Hundal, The rate of convergence for the cyclic projections algorithm, II. Norms
of nonlinear operators. J. Approx. Theor. 142, 56–82 (2006)
144. F. Deutsch, I. Yamada, Minimizing certain convex functions over the intersection of the fixed
point sets of nonexpansive mappings. Numer. Funct. Anal. Optim. 19, 33–56 (1998)
145. J.B. Diaz, F.T. Metcalf, On the set of subsequential limit points of successive approximations.
Trans. Am. Math. Soc. 135, 459–485 (1969)
146. L.T. Dos Santos, A parallel subgradient projections method for the convex feasibility problem.
J. Comput. Appl. Math. 18, 307–320 (1987)
147. W.G. Dotson Jr., On the Mann iterative process. Trans. Am. Math. Soc. 149, 65–73 (1970)
References 281
148. W.G. Dotson, Fixed points of quasi-nonexpansive mappings. J. Austral. Math. Soc. 13, 167–
170 (1972)
149. J. Douglas, H.H. Rachford, On the numerical solution of heat conduction problems in two or
three space variables. Trans. Am. Math. Soc. 82, 421–439 (1956)
150. R. Dudek, Iterative method for solving the linear feasibility problem. J. Optim Theor. Appl.
132, 401–410 (2007)
151. J. Dye, M.A. Khamsi, S. Reich, Random products of contractions in Banach spaces. Trans.
Am. Math. Soc. 325, 87–99 (1991)
152. J.M. Dye, S. Reich, On the unrestricted iteration of projections in Hilbert space. J. Math.
Anal. Appl. 156, 101–119 (1991)
153. J. Dye, S. Reich, Unrestricted iterations of nonexpansive mappings in Hilbert space.
Nonlinear Anal. 18, 199–207 (1992)
154. R. Dylewski, Selection of Linearizations in Projection Methods for Convex Optimization
Problems (in Polish), Ph.D. thesis, University of Zielona Góra, Institute of Mathematics, 2003
155. R. Dylewski, Projection method with residual selection for linear feasibility problems.
Discuss. Math. Differ. Incl. Contr. Optim. 27, 43–50 (2007)
156. M.G. Eberle, M.C. Maciel, Finding the closest Toeplitz matrix. Computat. Appl. Math. 22,
1–18 (2003)
157. J. Eckstein, D.P. Bertsekas, On the Douglas–Rachford splitting method and the proximal point
algorithm for maximal monotone operators. Math. Program. 55, 293–318 (1992)
158. I. Ekeland, R. Témam, Convex Analysis and Variational Problems (North-Holland, Amster-
dam, 1976)
159. T. Elfving, A projection method for semidefinite linear systems and its applications. Lin.
Algebra Appl. 391, 57–73 (2004)
160. T. Elfving, T. Nikazad, Stopping rules for Landweber-type iteration. Inverse Probl. 23, 1417–
1432 (2007)
161. V. Elser, I. Rankenburg, P. Thibault, Searching with iterated maps. Proc. Natl. Acad. Sci. USA
104, 418–423 (2007)
162. L. Elsner, I. Koltracht, P. Lancaster, Convergence properties of ART and SOR algorithms.
Numer. Math. 59, 91–106 (1991)
163. L. Elsner, I. Koltracht, M. Neumann, On the convergence of asynchronous paracontractions
with application to tomographic reconstruction from incomplete data. Lin. Algebra Appl. 130,
65–82 (1990)
164. L. Elsner, I. Koltracht, M. Neumann, Convergence of sequential and asynchronous nonlinear
paracontractions. Numer. Math. 62, 305–319 (1992)
165. F. Facchinei, J.-S. Pang, Finite-Dimensional Variational Inequalities and Complementarity
Problems, Volume I, Volume II (Springer, New York, 2003)
166. M. Fiedler, V. Pták, On matrices with non-positive off-diagonal elements and positive
principal minors. Czech. Math. J. 12, 382–400 (1962)
167. S. Fitzpatrick, R.R. Phelps, Differentiability of the metric projection in Hilbert space. Trans.
Am. Math. Soc. 270, 483–501 (1982)
168. S.D. Flåm, J. Zowe, Relaxed outer projections, weighted averages and convex feasibility. BIT
30, 289–300 (1990)
169. R. Fletcher, Practical Methods of Optimization (Wiley, Chichester, 1987)
170. K. Friedricks, On certain inequalities and characteristic value problems for analytic functions
and for functions of two variables. Trans. Am. Math. Soc. 41, 321–364 (1937)
171. M. Fukushima, A relaxed projection method for variational inequalities. Math. Program. 35,
58–70 (1986)
172. E.M. Gafni, D.P. Bertsekas, Two metric projection methods for constrained optimization.
SIAM J. Contr. Optim. 22, 936–964 (1984)
173. A. Galántai, Projectors and Projection Methods (Kluwer Academic, Boston, 2004)
174. A. Galántai, On the rate of convergence of the alternating projection method in finite
dimensional spaces. J. Math. Anal. Appl. 310, 30–44 (2005)
282 References
175. U. Garcı́a-Palomares, Parallel projected aggregation methods for solving the convex feasibil-
ity problem. SIAM J. Optim. 3, 882–900 (1993)
176. U. Garcı́a-Palomares, A superlinearly convergent projection algorithm for solving the convex
inequality problem. Oper. Res. Lett. 22, 97–103 (1998)
177. W.B. Gearhart, M. Koshy, Acceleration schemes for the method of alternating projections.
J. Comput. Appl. Math. 26, 235–249 (1989)
178. C. Geiger, Ch. Kanzow, Numerische Verfahren zur Lösung unrestingierter Optimierungsauf-
gaben (Springer, Berlin, 1999)
179. C. Geiger, Ch. Kanzow, Theorie und Numerik restringierter Optimierungsaufgaben (Springer,
Berlin, 2002)
180. J.R. Giles, Convex Analysis with Application in Differentiation of Convex Functions (Pitman
Advanced Publishing Program, Boston, 1982)
181. P.E. Gill, W. Murray, M.H. Wright, Numerical Linear Algebra and Optimization (Addison-
Wesley, Redwood City, 1991)
182. W. Glunt, T.L. Hayden, R. Reams, The nearest ‘doubly stochastic’ matrix to a real matrix
with the same first moment. Numer. Lin. Algebra Appl. 5, 475–482 (1998)
183. K. Goebel, Concise Course on Fixed Points Theorems (Yokohama Publishing, Yokohama,
2002); Polish translation: Twierdzenia o punktach stałych (Wydawnictwo UMCS, Lublin,
2005)
184. K. Goebel, W.A. Kirk, Topics in Metric Fixed Point Theory (Cambridge University
Press, Cambridge, 1990); Polish translation: Zagadnienia metrycznej teorii punktów stałych
(Wydawnictwo UMCS, Lublin, 1999)
185. K. Goebel, S. Reich, Uniform Convexity, Hyperbolic Geometry, and Nonexpansive Mappings
(Marcel Dekker, New York, 1984)
186. J.L. Goffin, The relaxation method for solving systems of linear inequalities. Math. Oper. Res.
5, 388–414 (1980)
187. J.L. Goffin, On the finite convergence of the relaxation method for solving systems of
inequalities. Operations Research Center, Report ORC 71–36, University of California,
Berkeley, 1971
188. D. Göhde, Zum Prinzip der kontraktiven Abbildung. Math. Nachr. 30, 251–258 (1965)
189. R. Gordon, R. Bender, G.T. Herman, Algebraic reconstruction techniques (ART) for three-
dimensional electron microscopy and X-ray photography. J. Theoret. Biol. 29, 471–481
(1970)
190. D. Gordon, R. Gordon, Component-averaged row projections: a robust, block-parallel scheme
for sparse linear systems. SIAM J. Sci. Comput. 27, 1092–1117 (2005)
191. A. Granas, J. Dugundji, Fixed Point Theory (Springer, New York, 2003)
192. K.M. Grigoriadis, A.E. Frazho, R.E. Skelton, Application of alternating convex projection
methods for computation of positive Toeplitz matrices. IEEE Trans. Signal Process. 42, 1873–
1875 (1994)
193. K.M. Grigoriadis, R.E. Skelton, Low-order control design for LMI problems using alternating
projection methods. Automatica 32, 1117–1125 (1996)
194. K.M. Grigoriadis, R.E. Skelton, Alternating convex projection methods for discrete-time
covariance control design. J. Optim. Theor. Appl. 88, 399–432 (1996)
195. J. Gu, H. Stark, Y. Yang, Wide-band smart antenna design using vector space projection
methods. IEEE Trans. Antenn. Propag. 52, 3228–3236 (2004)
196. L.G. Gurin, B.T. Polyak, E.V. Raik, The method of projection for finding the common point in
convex sets. Zh. Vychisl. Mat. Mat. Fiz. 7, 1211–1228 (1967) (in Russian); English translation
in: USSR Comput. Math. Phys. 7, 1–24 (1967)
197. R. Haller, R. Szwarc, Kaczmarz algorithm in Hilbert space. Studia Math. 169, 123–132 (2005)
198. I. Halperin, The product of projection operators. Acta Sci. Math. (Szeged) 23, 96–99 (1962)
199. H.W. Hamacher, K.-H. Küfer, Inverse radiation therapy planning – a multiple objective
optimization approach. Discrete Appl. Math. 118, 145–161 (2002)
200. S.-P. Han, A successive projection method. Math. Program. (Ser. A) 40, 1–14 (1988)
201. M. Hanke, W. Niethammer, On the acceleration of Kaczmarz’s method for inconsistent linear
systems. Lin. Algebra Appl. 130, 83–98 (1990)
References 283
251. P.-E. Maingé, New approach to solving a system of variational inequalities and hierarchical
problems. J. Optim. Theor. Appl. 138, 459–477 (2008)
252. W.R. Mann, Mean value methods in iteration. Proc. Am. Math. Soc. 4, 506–510 (1953)
253. Şt. Măruşter, The solution by iteration of nonlinear equations in Hilbert spaces. Proc. Am.
Math. Soc. 63, 69–73 (1977)
254. Şt. Măruşter, Quasi-nonexpansivity and the convex feasibility problem. An. Ştiinţ. Univ. Al.
I. Cuza Iaşi Inform. (N.S.) 15, 47–56 (2005)
255. Şt. Măruşter, C. Popı̂rlan, On the Mann-type iteration and the convex feasibility problem.
J. Comput. Appl. Math. 212, 390–396 (2008)
256. Şt. Măruşter, C. Popı̂rlan, On the regularity condition in a convex feasibility problem.
Nonlinear Anal. 70, 1923–1928 (2009)
257. E. Masad, S. Reich, A note on the multiple-set split convex feasibility problem in Hilbert
space. J. Nonlinear Convex Anal. 8, 367–371 (2007)
258. E. Matoušková, S. Reich, The Hundal example revisited. J. Nonlinear Convex Anal. 4, 411–
427 (2003)
259. S.F. McCormick, The methods of Kaczmarz and row orthogonalization for solving linear
equations and least squares problems in Hilbert space. Indiana Univ. Math. J. 26, 1137–1150
(1977)
260. Yu.I. Merzlyakov, On a relaxation method of solving systems of linear inequalities (in
Russian). Zh. Vychisl. Mat. Mat. Fiz. 2, 482–487 (1962)
261. D. Michalski, Y. Xiao, Y. Censor, J.M. Galvin, The dose-volume constraint satisfaction
problem for inverse treatment planning with field segments. Phys. Med. Biol. 49, 601–616
(2004)
262. W. Mlak, Introduction to Hilbert Spaces (in Polish) (PWN, Warsaw, 1982)
263. W. Mlak, Hilbert Spaces and Operator Theory (Kluwer Academic, Boston, 1991)
264. J.-J. Moreau, Décomposition orthogonale d’un espace hilbertien selon deux cônes mutuelle-
ment polaires. C. R. Acad. Sci. Paris 255, 238–240 (1962)
265. J. Moreno, B. Datta, M. Raydan, A symmetry preserving alternating projection method for
matrix model updating. Mech. Syst. Signal Process. 23, 1784–1791 (2009)
266. T.S. Motzkin, I.J. Schoenberg, The relaxation method for linear inequalities. Can. J. Math. 6,
393–404 (1954)
267. J. Musielak, Introduction to Functional Analysis (in Polish) (PWN, Warszawa, 1989)
268. J. Mycielski, S. Świerczkowski, Uniform approximation with linear combinations of repro-
ducing kernels. Studia Math. 121, 105–114 (1996)
269. N. Nadezhkina, W. Takahashi, Strong convergence theorem by a hybrid method for nonexpan-
sive mappings and Lipschitz-continuous monotone mappings. SIAM J. Optim. 16, 1230–1241
(2006)
270. F. Natterer, The Mathematics of Computerized Tomography (Wiley, Chichester, 1986)
271. J. von Neumann, in Functional Operators – Vol. II. The Geometry of Orthogonal Spaces.
Annals of Mathematics Studies, vol. 22 (Princeton University Press, Princeton, 1950) (Reprint
of mimeographed lecture notes first distributed in 1933)
272. O. Nevanlinna, S. Reich, Strong convergence of contraction semigroups and of iterative
methods for accretive operators in Banach spaces. Israel J. Math. 32, 44–58 (1979)
273. N. Ogura, I. Yamada, Non-strictly convex minimization over the fixed point set of an
asymptotically shrinking nonexpansive mapping. Numer. Funct. Anal. Optim. 23, 113–137
(2002)
274. N. Ogura, I. Yamada, Nonstrictly convex minimization over the bounded fixed point set of a
nonexpansive mapping. Numer. Funct. Anal. Optim. 24, 129–135 (2003)
275. S. Oh, R.J. Marks, L.E. Atlas, Kernel synthesis for generalized time-frequency distributions
using the method of alternating projections onto convex sets. IEEE Trans. Signal Process. 42,
1653–1661 (1994)
276. J.G. O’Hara, P. Pillay, H.-K. Xu, Iterative approaches to finding nearest common fixed points
of nonexpansive mappings in Hilbert spaces. Nonlinear Anal. 54, 1417–1426 (2003)
277. S.O. Oko, Surrogate methods for linear inequalities. J. Optim. Theor. Appl. 72, 247–268
(1992)
286 References
278. Z. Opial, Nonexpansive and Monotone Mappings in Banach Spaces. Lecture Notes 67-1,
Center for Dynamical Systems, Brown University, Providence, RI, 1967
279. Z. Opial, Weak convergence of the sequence of successive approximations for nonexpansive
mappings. Bull. Am. Math. Soc. 73, 591–597 (1967)
280. S.C. Park, M.K. Park, M.G. Kang, Super-resolution image reconstruction: A technical
overview. IEEE Signal Process. Mag. 20, 21–36 (2003)
281. J. Park, D.C. Park, R.J. Marks, M. El-Sharkawi, Recovery of image blocks using the method
of alternating projections. IEEE Trans. Image Process. 14, 461–474 (2005)
282. S.N. Penfold, R.W. Schulte, Y. Censor, A.B. Rosenfeld, Total variation superiorization
schemes in proton computed tomography image reconstruction. Med. Phys. 37, 5887–5895
(2010)
283. W.V. Petryshyn, T.E. Williamson Jr., Strong and weak convergence of the sequence of
successive approximations for quasi-nonexpansive mappings. J. Math. Anal. Appl. 43, 459–
497 (1973)
284. G. Pierra, Decomposition through formalization in a product space. Math. Program. 28, 96–
115 (1984)
285. C. Popa, Least-squared solution of overdetermined inconsistent linear systems using
Kaczmarz’s relaxation. Int. J. Comp. Math. 55, 79–89 (1995)
286. C. Popa, Extensions of block-projections methods with relaxation parameters to inconsistent
and rank-deficient least-squares problems. BIT 38, 151–176 (1998)
287. C. Popa, R. Zdunek, Kaczmarz extended algorithm for tomographic image reconstruction
from limited data. Math. Comput. Simulat. 65, 579–598 (2004)
288. S. Prasad, Generalized array pattern synthesis by the method of alternating orthogonal
projections. IEEE Trans. Antenn. Propag. 28, 328–332 (1980)
289. J.L. Prince, A.S. Willsky, A geometric projection-space reconstruction algorithm. Lin.
Algebra Appl. 130, 151–191 (1990)
290. E. Pustylnik, S. Reich, A.J. Zaslavski, Convergence of infinite products of nonexpansive
operators in Hilbert space. J. Nonlinear Convex Anal. 11, 461–474 (2010)
291. E. Pustylnik, S. Reich, A.J. Zaslavski, Convergence of non-cyclic infinite products of
operators. J. Math. Anal. Appl. 380, 759–767 (2011)
292. B. Qu, N. Xiu, A note on the CQ algorithm for the split feasibility problem. Inverse Probl. 21,
1655–1665 (2005)
293. B. Qu, N. Xiu, A new halfspace-relaxation projection method for the split feasibility problem.
Lin. Algebra Appl. 428, 1218–1229 (2008)
294. S. Reich, Weak convergence theorems for nonexpansive mappings in Banach spaces. J. Math.
Anal. Appl. 67, 274–276 (1979)
295. S. Reich, A limit theorem for projections. Lin. Multilinear Algebra 13, 281–290 (1983)
296. S. Reich and I. Shafrir, The asymptotic behavior of firmly nonexpansive mappings. Proc. Am.
Math. Soc. 101, 246–250 (1987)
297. S. Reich, A.J. Zaslavski, Attracting mappings in Banach and hyperbolic spaces. J. Math. Anal.
Appl. 253, 250–268 (2001)
298. R.T. Rockafellar, Convex Analysis (Princeton University Press, Princeton, 1970)
299. R.T. Rockafellar, Monotone operators and the proximal point algorithm. SIAM J. Contr.
Optim. 14, 877–898 (1976)
300. W. Rudin, Functional Analysis, 2nd edn. (McGraw-Hill, New York, 1991); Polish translation:
Analiza funkcjonalna (PWN, Warszawa, 2002)
301. A.A. Samsonov, E.G. Kholmovski, D.L. Parker, C.R. Johnson, POCSENSE: POCS-based
reconstruction for sensitivity encoded magnetic resonance imaging. Magn. Reson. Med. 52,
1397–1406 (2004)
302. J. Schauder, Der Fixpunktsatz in Funktionalräumen. Studia Math. 2, 171–180 (1930)
303. D. Schott, A general iterative scheme with applications to convex optimization and related
fields. Optimization 22, 885–902 (1991)
304. H.D. Scolnik, N. Echebest, M.T. Guardarucci, M.C. Vacchino, A class of optimized row
projection methods for solving large nonsymmetric linear systems. Appl. Numer. Math. 41,
499–513 (2002)
References 287
305. H.D. Scolnik, N. Echebest, M.T. Guardarucci, M.C. Vacchino, Acceleration scheme for
parallel projected aggregation methods for solving large linear systems. Ann. Oper. Res. 117,
95–115 (2002)
306. H.D. Scolnik, N. Echebest, M.T. Guardarucci, M.C. Vacchino, Incomplete oblique projections
for solving large inconsistent linear systems. Math. Program. 111, 273–300 (2008)
307. A. Segal, Directed Operators for Common Fixed Point Problems and Convex Programming
Problems, Ph.D. Thesis, University of Haifa, Haifa, Israel, 2008
308. H.F. Senter, W.G. Dotson Jr., Approximating fixed points of nonexpansive mappings. Proc.
Am. Math. Soc. 44, 375–380 (1974)
309. A. Serbes, L. Durak, Optimum signal and image recovery by the method of alternating
projections in fractional Fourier domains. Comm. Nonlinear Sci. Numer. Simulat. 15, 675–
689 (2010)
310. N.T. Shaked, J. Rosen, Multiple-viewpoint projection holograms synthesized by spatially
incoherent correlation with broadband functions. J. Opt. Soc. Am. A 25, 2129–2138 (2008)
311. G. Sharma, Set theoretic estimation for problems in subtractive color. Color Res. Appl. 25,
333–348 (2000)
312. K.K. Sharma, S.D. Joshi, Extrapolation of signals using the method of alternating projections
in fractional Fourier domains. Signal Image Video Process. 2, 177–182 (2008)
313. K.T. Smith, D.C. Solman, S.L. Wagner, Practical and mathematical aspects of the problem of
reconstructing objects from radiographs. Bull. Am. Math. Soc. 83, 1227–1270 (1977)
314. R.A. Soni, K.A. Gallivan, W.K. Jenkins, Low-complexity data reusing methods in adaptive
filtering. IEEE Trans. Signal Process. 52, 394–405 (2004)
315. H. Stark, P. Oskoui, High resolution image recovery from image-plane arrays, using convex
projections. J. Opt. Soc. Am. A 6, 1715–1726 (1989)
316. H. Stark, Y. Yang, Vector Space Projections. A Numerical Approach to Signal and Image
Processing, Neural Nets and Optics (Wiley, New York, 1998)
317. J. Stoer, R. Bulirsch, Introduction to Numerical Analysis, 3rd edn. (Springer, New York, 2002)
318. C. Sudsukh, Strong convergence theorems for fixed point problems, equilibrium problems
and applications. Int. J. Math. Anal. (Ruse) 3, 1867–1880 (2009)
319. S. Świerczkowski, A model of following. J. Math. Anal. Appl. 222, 547–561 (1998)
320. W. Takahashi, Y. Takeuchi, R. Kubota, Strong convergence theorems by hybrid methods for
families of nonexpansive mappings in Hilbert spaces. J. Math. Anal. Appl. 341, 276–286
(2008)
321. K. Tanabe, Projection method for solving a singular system of linear equations and its
applications. Numer. Math. 17, 203–214 (1971)
322. K. Tanabe, Characterization of linear stationary iterative processes for solving a singular
system of linear equations. Numer. Math. 22, 349–359 (1974)
323. G. Tetzlaff, K. Arnold, A. Raabe, A. Ziemann, Observations of area averaged near-surface
wind- and temperature-fields in real terrain using acoustic travel time tomography. Meteorol-
ogische Zeitschrift 11, 273–283 (2002)
324. Ch. Thieke, T. Bortfeld, A. Niemierko, S. Nill, From physical dose constraints to equivalent
uniform dose constraints in inverse radiotherapy planning. Med. Phys. 30, 2332–2339 (2003)
325. J. van Tiel, Convex Analysis, An Introductory Text (Wiley, Chichester, 1984)
326. M.J. Todd, Some Remarks on the Relaxation Method for Linear Inequalities, Technical
Report, vol. 419, Cornell University, Cornell, Ithaca, 1979
327. Ph.L. Toint, Global convergence of a class of trust region methods for nonconvex minimiza-
tion in Hilbert space. IMA J. Numer. Anal. 8, 231–252 (1988)
328. W. Treimer, U. Feye-Treimer, Two dimensional reconstruction of small angle scattering
patterns from rocking curves. Physica B 241–243, 1228–1230 (1998)
329. J.A. Tropp, I.S. Dhillon, R.W. Heath, T. Strohmer, Designing structured tight frames via an
alternating projection method. IEEE Trans. Inform. Theor. 51, 188–209 (2005)
330. M.R. Trummer, SMART – an algorithm for reconstructing pictures from projections. J. Appl.
Math. Phys. 34, 746–753 (1983)
288 References
331. P. Tseng, On the convergence of the products of firmly nonexpansive mappings. SIAM
J. Optim. 2, 425–434 (1992)
332. A. Van der Sluis, H.A. Van der Vorst, in Numerical Solution of Large Sparse Linear Algebraic
Systems Arising from Tomographic Problems, ed. by G. Nolet. Seismic Tomography (Reidel,
Dordrecht, 1987)
333. V.V. Vasin, A.L. Ageev, Ill-Posed Problems with A Priori Information (VSP, Utrecht, 1995)
334. S. Webb, Intensity Modulated Radiation Therapy (Institute of Physics Publishing, Bristol,
2001)
335. S. Webb, The Physics of Conformal Radiotherapy (Institute of Physics Publishing, Bristol,
2001)
336. R. Webster, Convexity (Oxford University Press, Oxford, 1994)
337. R. Wegmann, Conformal mapping by the method of alternating projections. Numer. Math.
56, 291–307 (1989)
338. R. Wittmann, Approximation of fixed points of nonexpansive mappings. Arch. Math. 58,
486–491 (1992)
339. P. Wolfe, Finding the nearest point in a polytope. Math. Program. 11, 128–149 (1976)
340. B.J. van Wyk, M.A. van Wyk, A POCS-based graph matching algorithm. IEEE Trans. Pattern
Anal. Mach. Intell. 26, 1526–1530 (2004)
341. Y. Xiao, Y. Censor, D. Michalski, J.M. Galvin, The least-intensity feasible solution for
aperture-based inverse planning in radiation therapy. Ann. Oper. Res. 119, 183–203 (2003)
342. H.-K. Xu, Iterative algorithms for nonlinear operators. J. Lond. Math. Soc. 66, 240–256
(2002)
343. H.-K. Xu, An iterative approach to quadratic optimization. J. Optim. Theor. Appl. 116, 659–
678 (2003)
344. H.-K. Xu, A variable Krasnosel’skiı̆-Mann algorithm and the multiple-set split feasibility
problem. Inverse Probl. 22, 2021–2034 (2006)
345. I. Yamada, in The Hybrid Steepest Descent Method for the Variational Inequality Problem
Over the Intersection of Fixed Point Sets of Nonexpansive Mappings, ed. by D. Butnariu,
Y. Censor, S. Reich. Inherently Parallel Algorithms in Feasibility and Optimization and their
Applications, Studies in Computational Mathematics, vol. 8 (Elsevier Science, Amsterdam,
2001), pp. 473–504
346. I. Yamada, N. Ogura, Hybrid steepest descent method for variational inequality problem over
the fixed point set of certain quasi-nonexpansive mappings. Numer. Funct. Anal. Optim. 25,
619–655 (2004)
347. I. Yamada, N. Ogura, Adaptive projected subgradient method for asymptotic minimization of
sequence of nonnegative convex functions. Numer. Funct. Anal. Optim. 25, 593–617 (2004)
348. I. Yamada, N. Ogura, N. Shirakawa, in A Numerically Robust Hybrid Steepest Descent
Method for the Convexly Constrained Generalized Inverse Problems, ed. by Z. Nashed,
O. Scherzer. Inverse Problems, Image Analysis and Medical Imaging, American Mathemati-
cal Society, Contemp. Math., vol. 313 (2002), pp. 269–305
349. I. Yamada, N. Ogura, Y. Yamashita, K. Sakaniwa, Quadratic optimization of fixed points of
nonexpansive mappings in Hilbert space. Numer. Funct. Anal. Optim. 19, 165–190 (1998)
350. Q.Z. Yang, The relaxed CQ algorithm solving the split feasibility problem. Inverse Probl. 20,
1261–1266 (2004)
351. K. Yang, K.G. Murty, New iterative methods for linear inequalities. JOTA 72, 163–185 (1992)
352. Q. Yang, J. Zhao, Generalized KM theorems and their applications. Inverse Probl. 22, 833–
844 (2006)
353. Y. Yao, Y.-C. Liou, Weak and strong convergence of Krasnoselski–Mann iteration for
hierarchical fixed point problems. Inverse Probl. 24, 015015, 8 (2008)
354. D. Youla, Generalized image restoration by the method of alternating orthogonal projections.
IEEE Trans. Circ. Syst. 25, 694–702 (1978)
355. M. Yukawa, I. Yamada, Pairwise optimal weight realization – Acceleration technique for set-
theoretic adaptive parallel subgradient projection algorithm. IEEE Trans. Signal Process. 54,
4557–4571 (2006)
References 289
356. M. Zaknoon, Algorithmic Developments for the Convex Feasibility Problem, Ph.D. Thesis,
University of Haifa, Haifa, Israel, 2003
357. E.H. Zarantonello, in Projections on Convex Sets in Hilbert Space and Spectral Theory, ed.
by E.H. Zarantonello. Contributions to Nonlinear Functional Analysis (Academic, New York,
1971), pp. 237–424
358. E. Zeidler, Nonlinear Functional Analysis and Its Applications, III – Variational Methods and
Optimization (Springer, New York, 1985)
359. J. Zhang, A.K. Katsaggelos, in Image Recovery Using the EM Algorithm, ed. by V.K.
Madisetti, D.B. Williams. Digital Signal Processing Handbook (CRC Press LLC, Boca Raton,
1999)
360. D.F. Zhao, The principles and practice of iterative alternating projection algorithm: Solution
for non-LTE stellar atmospheric model with the method of linearized separation. Chin.
Astron. Astrophys. 25, 305–316 (2001)
361. J. Zhao, Q. Yang, Several solution methods for the split feasibility problem. Inverse Probl. 21,
1791–1799 (2005)
•
Glossary of Symbols
H Hilbert space
hx; yi Inner product of x; y 2 H
kxk Norm of x 2 H induced by h; i
hx; yiG Inner product of x; y 2 Rn induced by a positive definite
p matrix G
kxkG WD hx; xiG The norm of x 2 Rn induced by h; iG
^.x; y/ Angle between nonzero vectors x; y 2 H
X A nonempty closed convex subset of H
I WD f1; 2; :::; mg Finite subset of indices
xC ; x Positive and the negative part of x 2 Rn
RnC ; Rn Nonnegative and the nonpositive orthant
m Standard simplex in Rm
jJ j The number of elements of a finite subset J
V? Subspace orthogonal to a subspace V H
B.x; / Ball with a centre x and radius > 0
C0 Complement of a subset C H
bd C Boundary of a subset C H
int C Interior of a subset C H
cl C Closure of a subset C H
Lin S Linear subspace generated by S H
aff S Affine subspace generated by S H
H.a; ˇ/ Hyperplane fx 2 H W ha; xi D ˇg
H .a; ˇ/ Half-space fx 2 H W ha; xi ˇg
HC .a; ˇ/ Half-space fx 2 H W ha; xi ˇg
fC ; f Positive and the negative part of a function f
Argminx2X f .x/ Subset of minimizers of f W X ! R
argminx2X f .x/ Minimizer of f W X ! R
A. Cegielski, Iterative Methods for Fixed Point Problems in Hilbert Spaces, 291
Lecture Notes in Mathematics 2057, DOI 10.1007/978-3-642-30901-4,
© Springer-Verlag Berlin Heidelberg 2012
292 Glossary of Symbols
A. Cegielski, Iterative Methods for Fixed Point Problems in Hilbert Spaces, 293
Lecture Notes in Mathematics 2057, DOI 10.1007/978-3-642-30901-4,
© Springer-Verlag Berlin Heidelberg 2012
294 Glossary of Acronyms
A. Cegielski, Iterative Methods for Fixed Point Problems in Hilbert Spaces, 295
Lecture Notes in Mathematics 2057, DOI 10.1007/978-3-642-30901-4,
© Springer-Verlag Berlin Heidelberg 2012
296 Index