Nonhnear Anolysrs. Theory, Merhods & Applicarrom, Vol. IS. No. 6, pp. 537-558. 1990. 0362-546X/W 53.001 .

CNl
Printed in Great Britain. c 19% Pergamon Press pit

NONEXPANSIVE ITERATIONS IN HYPERBOLIC SPACES

SIMEON REICH*
Department of Mathematics, University of Southern California, Los Angeles, CA 90089, U.S.A.

and
ITAI SHAFRIR
Department of Mathematics, Technion-Israel Institute of Technology, 32000 Haifa, Israel

(Received 23 July 1989; received

for publication
1 March 1990)

Key words and phrases: Co-accretive operator, hyperbolic space, iteration, nonexpansive mapping,
nonlinear semigroup, resolvent, uniform convexity.

1. INTRODUCTION

ONE OF THE most active research areas in nonlinear functional analysis is the asymptotics of
nonexpansive mappings. Most of the results, however, have been obtained in normed linear
spaces. It is natural, therefore, to try to develop a theory of nonexpansive iterations in more
general infinite-dimensional manifolds. This is the purpose of the present paper. More specifi-
cally, we propose the class of hyperbolic spaces as an appropriate background for the study of
operator theory in general, and of iterative processes for nonexpansive mappings in particular.
This class of metric spaces, which is defined in Section 2, includes all normed linear spaces
and Hadamard manifolds, as well as the Hilbert ball and the Cartesian product of Hilbert balls.
In Section 3 we introduce co-accretive operators and their resolvents, and present some of their
properties. In the fourth section we discuss the concept of uniform convexity for hyperbolic
spaces. Section 5 is devoted to two new geometric properties of (infinite-dimensional) Banach
spaces. Theorem 5.6 provides a characterization of Banach spaces having these properties in
terms of nonlinear accretive operators. In Sections 6, 7 and 8 we study explicit, implict and
continuous iterations, repectively, using the same approach in all three sections. We illustrate
this common approach with the following special case.
Let C be a closed convex subset of a hyperbolic space (X, p), let T: C --f C be a nonexpansive
mapping, and let x be a point in C. In order to study the iteration (T”x: n = 0, 1,2, . . .), we set
z,, = (1 - (l/n))x 0 (l/n)T”x, K = clco(zj;j I l), and d = inf(p(y, Ty): y E C).
The first step is to show that p(x, K) = lim p(x, T”x)/n = d. This leads to the convergence
“+m
of lz,) when X is uniformly convex and to the weak convergence of (z,,] when X is a Banach
space which is reflexive and strictly convex. When T is an averaged mapping we are also able
to establish the following triple equality. For all k 2 1,
lim p(T”x, TnC1x) = lim p(T”x, T”+kx)/k = lim p(x, Tnx)/n.
n-m n-m n-o
This result, in turn, yields weak and strong convergence of the sequence (T”x - T”+‘x) in
those Banach spaces which possess the geometric properties discussed in Section 5. It also

* Present address: Department of Mathematics, Technion-Israel Institute of Technology, 32000 Haifa, Israel.

537
538 S. REICH and I. SHAFRIR

leads to weak and strong convergence of the sequence (T”x) itself in the Hilbert ball. The
general case of all of these results is described in Section 6.
The first results in this direction, outside Banach spaces, were established in [19]. As a
matter of fact, theorem 6.6(a) provides an affirmative answer to the question raised at the end
of [19]. Special cases of some of the theorems of the present paper, as well as some related
results, were announced in (20, 21, 22, 231. Although some of our results are new even in the
Banach space setting, our main interest is in hyperbolic spaces. We hope, in fact, that the
concepts and techniques of our paper will soon lead to further developments in analysis on
hyperbolic spaces.

2. HYPERBOLIC SPACES

Let (X, p) be a metric space, and let R denote the real line. We say that a mapping c: R -, X
is a metric embedding of R into X if

Ms), c(0) = Is - tl
for all real s and 1. The image of R under a metric embedding will be called a metric line. The
image of a real interval [a, b] = (t E R : a I t I b) under such a mapping will be called a metric
segment.
Assume that (X, p) contains a family M of metric lines, such that for each pair of distinct
points x and y in X there is a unique metric line in M which passes through x and y. This metric
line determines a unique metric segment joining x and y. We denote this segment by [x, _Y].For
each 0 4 t I 1 there is a unique point z in [x,y] such that

P(X, z) = @(XYu) and P(Z,Y) = (1 - 0P(%Y).

This point will be denoted by (1 - t)x @ ty.
We shall say that X, or more precisely (X, p, M), is a hyperbolic space if

&x 0 +Y, tx 0 +z> 5 MY> z) (2.1)

for all x, y and z in X.
An equivalent requirement is that

P($ 0 +_K+w @z) 5 +(P(x, w) + P(Y, z)) (2.2)

for all x, y, z and w in X.
It is clear that all normed linear spaces are hyperbolic. So are all Hadamard manifolds, that
is, all finite-dimensional connected, simply connected, complete Riemannian manifolds of
nonpositive curvature (cf. [5, p. 3051). An infinite-dimensional example is provided by the
Hilbert ball equipped with the hyperbolic metric [lo, p. 1041. Other examples can be
constructed in the following manner. For each 1 I i 5 n, let (Xi, pi, Mi) be a hyperbolic space,
and let II-j/ be a norm on R” with the following property: l/r/l 5 llsll for all r = (r,, r2, . . . . r,J
and s = (s, , s2, . . ., s,) in R” such that 0 5 ri I si for all 1 I i 5 n. Define a metric p on the
Cartesian product X = n I= 1 Xi by

P(X,Y) = IkPI(X,9Yl),P,(X,, Yz), . . ..P”(X”9 VnNll,

wherex=(x,,x, ,..., x,Jandy=(y,,y,, . . ..u.) are any two points in X. Now suppose
that for each 1 5 i I n, the family Mi of metric lines consists of the images of the metric
 Nonexpansive iterations in hyperbolic spaces 539

embeddings (~;,~(f): CYE A;). Let M be the set of metric lines determined by all the metric
embeddings c: R + X of the form

c(t) = (Cl,&? + 410, c2.&2 + 42 09 . * * 9 Cn.&n + 4n Oh

whereaiEA;,piER, and (1(q1,q2 ,..., q.)[[ = 1. Then (X, p, M) is also a hyperbolic space.
When each Xi = B, the Hilbert ball, and ]I@,, r2, . . ., r,Jlj = max(Jri(: 1 s i 5 n), then the
space X is equal to B”, the Cartesian product of n Hilbert balls, equipped with its hyperbolic
metric.
A similar construction can be carried out for a countable product of hyperbolic spaces.
Choosing a “monotone” norm I( - (1 on a subspace S of R” and a fixed point a =
(al 9a 2, .**, a “,...)inX= nT=D=Xi,wedefinethemetricpon
Y= (XEX:(Pi(Xi,UJ:i= 1,2,...] ES]
by
P(X, Y) = ll(PAXl9Yl), /32(x2 9 Y2), * * * 9P”(X”, y,), *. .MI
*
As a matter of fact, this construction can be extended to an arbitrary product of hyperbolic
spaces. More information on spaces of hyperbolic type and on B” may be found, for example,
in [12] and in [14], respectively.

3. CO-ACCRETIVE OPERATORS

Given any two distinct points x and y in a hyperbolic space (X, p, M), there is a unique metric
line in M which passes through x and y. For r 2 0 we let (1 + r)x 0 ry stand for the unique
point z on this metric line satisfying

P(Z, x) = V(& Y) and P(Z, Y) = (1 + MX,Y).

We shall say that a set-valued operator T c X x X with domain D(T) and range R(T) is
co-accretive if
P(X, , x2) 5 P((l + ex1 0 rY1, (1 + 4X2 0 rY2) (3.1)
forallyiETxi,i= 1,2,andr>O.
When X is a normed linear space, the operator A = I - T, the complement of T with respect
to the identity operator I, satisfies

1x1- $1 5 1x1- x2 f r(z1 - z2)l (3.2)

for all zj E AXi, i = 1,2, and r > 0, and is therefore accretive. In this case, the operator T is
also said to be pseudo-contractive [4, p. 8761. A study of accretive operators on Finsler
manifolds can be found in [ 111.
Let D be a subset of X. Recall that a mapping T: D + X is said to be nonexpansive if

P(Tx, TY) 5 P(X, Y) (3.3)

for all x and y in D. Since the inequality (2.2) implies that

P((l - 0x 0 t.2, (1 - f)Y 0 W) 5 (1 - Op(x,y) + @(z, w) (3 -4)

for all x, y, z and w in X and all 0 I t 5 1, it is not difficult to see that all nonexpansive
mappings are co-accretive.
540 S. REICH and I.SHAFRIR

When T is co-accretive, we can define, for each positive r, a single-valued nonexpansive

mapping J,: R(( 1 + r)Z 0 rT) + D(T) by
J,((l + r)x 0 ry) = x, (3.5)
where x E D(T) and y E TX. These mappings (which in normed linear spaces are indeed the
resolvents of the accretive operator A = I - T) satisfy the following resolvent identity for all
t > s > 0 and x E D(J,):
J,x = J,((s/t)x @ (1 - s/t)J,x). (3.6)
Following [16, p. 3131 we shall say that the co-accretive operator T satisfies the range
condition at infinity if
lim inf p(x, R((l + t)Z @ tT))/t = 0, (3.7)
1-m
where x is an arbitrary point in X and p(x, D) = inf(p(x, y): y E D). Equivalently, T satisfies
(3.7) if and only if there is a sequence ((ti, xii): 1 I i c 00)such that ti + 00, x,~ E D(J,,), and
lim p(X, X,i)/ti = 0. (3.8)
i-m

The following result shows how this range condition provides a connection between the
asymptotic behavior of J, and the range of T. It extends (with a different proof) proposition 3
of [16].

THEOREM 3.1. Let X be a hyperbolic space, T c X x X a co-accretive operator that satisfies

(3.7), and ((ti,xfi)) a sequence satisfying (3.8). Then for each y in X, lim p(y, Jt,Xti)/ti =
i-r00
inf(p(u, v): u E D(T), v E TuJ.

Proof. We first note that for each positive t and x, E &I,), x, E (1 + t)J,x, 0 tTJ,x,. Hence
p(x,, J,x,) = tp(J,x,, v,) for some u, E TJ,x,, p(x,, J,x,)/t 2 d = inf(p(u, v): u E D(T), u E Tu],
and lim__%fp(y, Jl,Xr,)/ti 2 d.
Conversely, we let E be positive and find u E D(T) and v E Tu such that d s p(u, u) c d + E.
For t > s > 0 and x, E D(J,), we have u = J,((l + s)u 0 su) and, by the resolvent identity
(3.6) and (2.1),
P(U, .4x,) = p(J,((l + s)u 0 4, J,(Wt)x, 0 (1 - s/t)Jtx,))
I ~((1 + S)U 0 su, (s/t)x, @ (1 - s/t)J,x,)
s (s/t)p((l + s)u 0 su, x,) + (1 - s/t)p((l + S)U 0 su, J,x,).
Now,
P((1 + S)U 0 su, x,) 5 A(1 + S)U 0 su, u) + P(U, 4)
= s&u, u) + P(U, 4)
and
~((1 + s)u 0 su, 4-G 5 SP(U, u) + Au, .4x,),
so that
P(U, Ax,) 5 s&u, u) + (s/Mu, x,) + (1 - s/Mu, 4x,)
and
Au, 4x,)/t 5 P(U, 4 + P(U, x,)/t.
 Nonexpansive iterations in hyperbolic spaces 541

Letting t tend to infinity through (t;) we conclude that

lim supp(u, Jzix,,)/ti I p(u, v) < d + E,
;-CC
and the result follows.

4. UNIFORM CONVEXITY
We define the modulus of convexity 6: X x (0, co) x (0,2] + [0, l] of a hyperbolic space
X by
&a, r, e) = infll - p(a, +x @ +y)/r), (4.1)
where the infimum is taken over all points x and y in X satisfying p(a, x) 5 r, p(a, y) s r and
p(x, y) 2 ET. We say that the space X is uniformly convex if 6 is always positive.
When X is an Hadamard manifold, then the law of cosines yields the following
parallelogram inequality:

2(P(G x)2 + &a, y)2) L 4p(u, +x 0 iy,’ + p(x, y)2.

This inequality, in turn, shows that for these spaces
&a, r, E) L 1 - (1 - &2/4)“2 = f&f(&), (4.3)
the modulus of convexity of Hilbert spaces (with dimension 22).
Both (4.2) and (4.3) are also true in the (infinite-dimensional) Hilbert ball. As a matter of
fact, the modulus of convexity has been computed in this case and the formula can be found
on p. 107 of [lo]. Note, however, an error on that very page-all the balls in the Hilbert ball
are more rotund than the balls in Hilbert space.
Other examples of uniformly convex hyperbolic spaces are obtained when a strictly convex
norm on R” is used to define the metric on a product of n uniformly convex hyperbolic
spaces.
Recall now that the norm of a Banach space X is said to be Frechet differentiable if for each
x in its unit sphere S(X) = lx E X: Ix/ = 11, the limit

)i_mo(lx+ [Yl - l&/f (4.4)

is attained uniformly for y in S(X). We shall then write that X is (F).

We quote now a known result [7] which provides two equivalent properties for a dual space.

LEMMA4.1. The following properties are equivalent for a Banach space X:

(a) The dual space X* is (F).
(b) Every sequence lx,1 in the unit sphere S(X) such that lim f(x,) = 1 for some f in
S(X*) converges strongly. n-m
(c) Every sequence lx,,) in X such that
lim lx,/ = lim (inf((yl:y E clcolx,: m 2 n)))
n-02 PI-o)
converges strongly.

Here clco(D) denotes the closed convex hull of a subset D of X. We also quote an analog of
lemma 4.1 [7] for weak convergence.
542 S. REICH and I. SHAFRIR

LEMMA4.2. The following properties are equivalent for a Banach space X:

(a) X is reflexive and strictly convex.
(b) Every sequence (x,] in the unit sphere S(X) such that lim f(x,J = 1 for some f in
n+m
S(X*) converges weakly.
(c) Every sequence lx,] in X such that
lim Ix,/ = lim (inf((y(:y E clco(x,: m 2 n)])
converges weakly. “** “*=J

Returning to hyperbolic spaces, we recall that a subset C of a hyperbolic space X is said

to be convex if for each pair of points x and y in C the metric segment [x, y] is also contained
in C. The closed convex hull of a subset D of X is the intersection of all closed convex
subsets of X which contain D. Since the norm of the dual of a uniformly convex Banach space
X is uniformly Frechet differentiable, X* is certainly (F) and therefore X has property (c) of
lemma 4.1.
We observe now that complete hyperbolic spaces which are uniformly convex possess an
analogous property.

LEMMA4.3. Let (X, p) be a complete hyperbolic space. If X is uniformly convex, then every
sequence lx,] in X such that
lim p(x, x,) = lim (inf(p(x, y): y E clco(x, : m 2 n)))
n-ro;r n-C=
for some x in X, converges.

Proof. Denote lim p(x, x,) by d. We may assume, of course, that d is positive. Since
n-+o;r
lim p(x, +x~ @ 9x”)
r?l.ll-+m
is also equal to d, the uniform convexity of X implies that limp(x,,,, x,,) = 0. Hence the result.

Our next lemma is a consequence of the resolvent identity (3.6).

LEMMA4.4. Let J,, r > 0, be the resolvents of a co-accretive operator in a hyperbolic space X.
If 0 < s < I, x, E D(J,), x, E D(J) and x E X, then
P((1 - (1/.9)x 0 (l/s)J,x,, Ax,)
I (1 - l/r)p(x, J,x,) + p(x, x,)/s + p(x, x,)/t.

Proof. We have
P(J,X,, 4x0 = NJ,, J,((s/Gx, 0 (1 - s/0.&))
5 (s/t)p(x,, x,) + (1 - s/G&, 0,)
5 (1 - s/M-& Ax,) + Ax, x,) + (S/l)P(X, &).
The result follows.

We are now ready to present a result which sharpens theorem 3.1.

 Nonexpansive iterations in hyperbolic spaces 543

THEOREM4.5. Let T be a co-accretive operator in a complete hyperbolic space X. Assume that

T satisfies the range condition at infinity and that the sequence ((1;) x,,): 1 I i c a) satisfies
(3.8). Fix a point x in X and set (for sufficiently large n) y, = (1 - 1/t,)x @ (l/t,)J,,xt, ,
K,, = clco(yj: j 2 n), n 2 1, and d = inf(p(u, u): u E D(T), u E Tu). Then:
(a) lim p(x, K,,) = lim p(x, y,) = d.
n-a n-m
(b) If X is uniformly convex, then lim y, exists.
n-+m
(c) If X is a Banach space which is reflexive and strictly convex, then the weak lim y, exists.
n-m
(d) If X is a Banach space such that X* is (F), then v = the strong lim y, exists, and
n-m
x - u is the unique point of least norm in cl(R(Z - T)).

Proof. If z belongs to the convex hull of (yj: n I j I pJ, r,, = ~up(p(x,x,~)/t~: j L n)

and 4 > p, then, by lemma 4.4, p(z, J,,x,,) I (1 - l/t&(x, J,,xJ + 2r,, . Hence p(x, z) 2
P(x, J,,xt,) - P(z, J+,) 2 p(x, J1,xr,)/tq - 2r,.
Letting 4 tend to infinity and using theorem 3.1, we obtain p(x, z) 2 d - 2r,,
p(x, K,) L d - 2r,, and (a). Assertions (b) and (c) now follow from (a) combined with lemmata
4.3 and 4.2, respectively. The first part of (d) follows from lemma 4.1. For the second part, see
the argument on [16, p. 3141.

We also take this opportunity to improve upon theorem 2 of [ 161. This result was the original
motivation for the introduction of the range condition at infinity (3.7).

THEOREM4.6. Let D be a closed subset of a Banach space X and T: D + D a nonexpansive

mapping. If X* is (F), then the following are equivalent:
(a) lim inf d(0, R(Z + t(Z - T)))/t = 0.
(b) -‘i& T”x/n is the unique point of least norm in cl(R(Z - T)).
n-m
(c) lim 1T”xl/n = d(0, R(Z - T)).
ll-+m

Proof. We already know by the proof of [16, theorem 21 that (a) implies (b), and it is clear
that (b) implies (c). To show that (c) implies (a), we first note that by [16, theorem 11,
w = - lim T”x/n
n-m
exists for each x in D and is independent of x. Its norm equals d(0, R(Z - T)) = d by (c).
Choose any f in J(w). Let a natural number m be given. Since X* is (F), there is, by lemma 4.1,
a positive 6 such that lyl < d + 6, and f(y) > d2 - 6 imply that ]y - WI < l/m. Choose a
point x in D such that Ix - TX] < d + 6. Since
k-l

d2 = ,“rn_f((x - Tkx)/k) = jl~(l/k) c f(T’x - T’+‘x),

i=O

we must have lim sup f( T’x - T’+‘x) L d2. We can therefore find a strictly increasing sequence
i-cc

{n,) of natural numbers such that for each n = n, and x = x, we have f(T”x - T”+‘x) >
dZ - 6 and IT”x/n + WI < l/m. Hence,
](l/n)(Z + n(Z - T))T”xl = IT”x/n + T”x - T”+‘x(
zs I T”x/n + WI + IT”x - TnC1x - WI < 2/m,
544 S. REICH and 1. SHAFRIR

and
lim d(0, R(I + n,(Z - T)))/n, = 0,
m-cc
as claimed.

5. DIFFERENTIABLE NORMS

In this section we study two new geometric properties of (infinite-dimensional, real) Banach
spaces.
Recall that the duality map from a Banach space X into the family of nonempty closed
convex subsets of its dual X* is defined by
J(x) = 1x* E x* : (x,x*) = /x1* = Ix*121.
If X is reflexive, the range of J is all of X*, and to each bounded linear functional f in X*
there corresponds the nonempty face
J-‘(f) = (x E X: f(x) = 1x1’ = ff]‘J
of the sphere lx E X: 1x1 = IfI).
We say that the norm of a Banach space X is locally uniformly Gateaux differentiable if for
each y in the unit sphere S(X) and each face F of S(X) the limit (4.4) is attained uniformly for
all x in F. We shall then write that X is (LUG). We say that the norm of X is locally uniformly
Frtchet differentiable if for each face F of S(X) the limit (4.4) is attained uniformly for all y
in S(X) and x in F. We shall then write that X is (LUF). It is clear that X is (LUG) whenever
the norm of X is uniformly Gheaux .differentiable or X is smooth with compact faces.
Similarly, X is (LUF) if X* is uniformly convex or X is (F) with compact faces.

LEMMA 5.1. Let X be a Banach space and let S(X*) be the unit sphere of its dual.
(a) If X is (LUG), and (f,,: n = 1,2, . . . ) c S(X*), f E S(X*) and (Y,,: n = 1,2, . ..) c
J-‘(f) are such that lim f,(y,) = 1, then the weak-star lim ,f, = f.
n-05 n+oo
(b) If X is (LUF), and (f,: n = 1, 2, . . . ) c SW*), f e S(X*) and ly,: n = 1,2, . ..) c
J-‘(f) are such that lim f,(y,) = 1, then the strong lim f, = f.
n-+m n-m

Proof. Since the proofs of (a) and (b) are similar, we give only the proof of (a). To this end,
suppose If,], f, and (y,) satisfy the hypotheses of (a). If (f,] does not converge weak-star to f,
then there are a positive r, z E S(X) and a subsequence of (f,), which we shall continue to
denote by [f,,), such that (f, - f)(z) 2 2r > 0. We then have
0 5 (f - f,)(Y,) 5 (f - fn)(Y,)Kfn - f )(z)/r - 11
= (f” - f )(Y”) + (f - fn)(Y,) * (fn - f )(z)/r
= (fn - f )(y, + (f - fn)(Y,W~)
5 IY, + ((f - f,)(Yn)/r)zl - IYA - (f - fn)(y,) *f(z)/r.
Denote (f - f,)(y,)/r by t, 2 0. If t, = 0, then f = f,. Hence we may assume that t, > 0. We
then have
0 < r = (f -fn)(Y,)/t, 5 (Iv, + tnzl - IYnl -f(t,z))/t, + 0 asn + 00,
because t, + 0 as n -, a~ and X is (LUG). We have reached a contradiction and (a) is proved.
 Nonexpansive iterations in hyperbolic spaces 545

This lemma leads to the following characterizations of the properties (LUG) and (LUF)
(cf. [18, lemma 2.21 and lemmata 4.1 and 4.2).

PROPOSITION 5.2. Let J be the duality map of the Banach space X.

(a) X is (LUG) if and only if J is continuous from the strong topology of X to the weak-star
topology of X*, uniformly with respect to faces.
(b) X is (LUF) if and only if J is continuous from the strong topology of X to the strong
topology of X*, uniformly with respect to faces.
In other words, {Jx,] converges (weak-star or strongly, respectively) to f whenever
lim d(x,, , J-‘(f )) = 0.
n*m

PROPOSITION 5.3. Let J* be the duality map of the dual space X*.
(a) X is reflexive and X* is (LUG) if and only if, whenever 4 E S(X**),
If,,: n = 1,2, . ..I c (J*)-‘(4) and (x,: n = 1,2, . ..) c S(X)
are such that lim f,(x,) = 1, then {x,] converges weakly.
n-m
(b) X* is (LUF) if and only if, whenever 4 E S(X**),
(f,: n = 1,2, . ..I c (J*)-‘(9) and (x,: n = 1,2, . ..) c S(X)
are such that lim f,,(x,) = 1, then (x,) converges strongly.
“*m
Recall that a Banach space X is said to be locally uniformly convex (LUC) if for each x in
S(X) and E > 0 there is a positive 6 = &E, x) such that Ix - yI I E for all y in S(X) with
1(x + y)/2/ 2 1 - 6. It is said to be weakly locally uniformly convex (WLUC) if for each x in
S(X), x* in X* and E > 0 there is a positive 6 = &(E,x, x*) such that 1(x - y, x*)1 I E for all
y in S(X) with 1(x + y)/2/ 2 1 - 6.

PROPOSITION 5.4. Let X be a reflexive Banach space.

(a) If X is (WLUC), then X* is (LUG).
(b) If X is (LUC), then X* is (LUF).

Proof. Since the proofs of (a) and (b) are similar, we include only the proof of (b). Assume
then that x E S(X), (f,: n = 1,2, . . .) C J(x) and (x,: n = 1,2, . . .) C S(X) are such that
lim f,(x,) = 1. Since 1(x + x,)/21 2 fn((x + x,)/2) = (1 + f,(x,))/2 + 1, (x,) converges
n-m
strongly to x. The result now follows by proposition 5.3.
A similar argument shows that if X* is (LUC), then X is (LUF) and, of course, (F). We do
not know if the converse implications hold. There is [l, 251, however, a reflexive Banach space
which is (F), although its dual X* is not (LUC).
We turn now to the connection between these geometric properties and accretive operators.

LEMMA 5.5. Let X be a Banach space, A c X x X an accretive operator such that T = Z - A

satisfies the range condition at infinity (3.7), and d = d(0, R(A)). Let ((ri, xJ: 1 5 i < 00) be
a sequence that satisfies (3.8) with an increasing (ti), and let b belong to Aa. Then there is a
functional f in the weak-star closure of U (J((a - J,,X,,)/ti): 1 I i < 00) such that 1f 1 = d and
f(b) r d’.
546 S. REICH and I.SHAFRIR

Proof. Since A is accretive, there is, for each 1 I i c 00, a functionalfi in .Z((a - JliXti)/ti)
such that fi(b - (x,,- J,iX~i)/ti)2 0. Let f be a weak-star limit of a subnet of (file Since
theorem 3.1 implies that lim fi((xti - J,iX,i)/ti) = d2, it is clear that f(b) 2 d2. On the other
i-m
hand, the proof of lemma 4.4 shows that for i < j,

fj((Q - 4ixtJ/fi) = Lfjta - 4jXtj) - fj(J,ixti - 4jxtj)l/ti

1 l.tjl<lfjl - la - xt,l/ti - la - xtil/tj)*
Hence
f((a - JIiXti)/ti)2 d(d - 1~ - Xr;l/ti) for all i,
and therefore If) = d.

THEOREM 5.6. A Banach space X is reflexive and its dual X* is (LUG) (respectively, (LUF)) if
and only if every accretive operator A c X x X such that T = Z - A satisfies (3.7) has the
following property: every sequence lb,] in Z?(A) for which lim l&l = d(O,Z?(A)) converges
n+m
weakly (respectively, strongly).

Proof. We prove only the characterization of (LUF) duals. The assertion concerning (LUG)
can be established in a similar manner. Assume then that X* is (LUF), A c X x X is accretive,
T = Z - A satisfies (3.7), the sequence ((ti, x,,)) satisfies (3.8), the sequence (6,] is contained in
R(A), and that lim (b,,l = d(0, Z?(A))= d. By theorem 4.5(d), u = the strong lim(-.Z,,x,,/tJ
n-m i-m

exists and IuI = d(0, R(A)) = d. By lemma 5.5 there is, for each n, a functional f, in J(u) such
that If,1 = d and f,(b,) 2 d2. Hence (b,] converges strongly (to u) by proposition 5.3(b).
Conversely, suppose first that X is not reflexive. Then certainly X* is not (F) and there
is [ 131a nonexpansive T: X -+ X such that the sequence [ T”x/n) does not converge strongly for
any x in X. Define the averaged mapping U: X + X by U = (I + T)/2, and let ST and Su be
the nonlinear semigroups generated by -(I - T) and -(I - V), respectively. Since
S,(t) = S,(t/2) for all t 2 0 and IS,(n) - T”xl 5 fi Ix - Txl for all n, the sequences (U”x/nJ
and, Q fortiori, (U”x - U”+‘xj also diverge. But it is known [2] that lim 1U”x - u”+‘x\ =
d(0, R(Z - U)). We conclude, therefore, that X must be reflexive. If X* is not (LUF), then,
by proposition 5.3(b), there are a point x in S(X) and sequences (x, : n = 1,2, ...)c S(X) and
If,; n = 1,2, ..) c J(x) such that lim f,,(x,) = 1, but (x,] diverges. Define an operator
n-m
AcXxXby

A = (g(-fxvx)) ‘J (iil(Ovxi/f,(xi)))*

In order to see that A is accretive, it is enough to check that for each i L 1 and t > 0
there is a functional g E J(tx) such that g(Xi/fi(Xi) - X) 2 0. Indeed, g = ifi satisfies
g(Xi/fi(Xi) - X) = 0. In addition, d(O,Z?(A)) = 1 because IXi/f(Xi)l > 1 and 1x1 = 1. The
operator T = Z - A satisfies (3.7) since the even stronger condition
D(A) C R(Z + rA) for all r > 0
holds. But the sequence (6,] c R(A) defined by b, = x,/f,(x,,) does not converge strongly
although lim lb,,1 = 1 = d(0, R(A)). This completes the proof.
n-m
 Nonexpansive iterations in hyperbolic spaces 547

We remark in passing that in the necessity part of this proof, the operator A may be replaced
by an operator of the form Z - T where T is nonexpansive. We may, in fact, let T be the
resolvent (I + A)-‘. More explicitly, T is given by
T(Xi /f (Xi)) = 0 i= 1,2,...
T((l - t)x) = -tx, t > 0.

6. EXPLICIT ITERATIONS

In this section we return to general hyperbolic spaces in order to study certain aspects of the
asymptotic behavior of the explicit iteration (6.1) defined below.
Let D be a closed subset of a hyperbolic space (X, p), T: D -+ X a nonexpansive mapping,
and (c,: n = 0, 1,2, . . . ) a real sequence such that 0 < c, I 1 and Cy=,,c, = 43. Denote the
partial sum 1;;: Cj by a,, , n L 1, and consider the sequence of nonexpansive mappings
defined by T, = (1 - c,)Z @ c, T. Assuming that the range of each T, is contained in D, we
intend to study the explicit iteration process

Yn+l = TIY,* n = 0, 1) 2, . . .) (6.1)

where the initial point y, belongs to D. We denote the resolvents corresponding to T by J,,
r > 0, and begin with two lemmata.

LEMMA 6.1. If (y,) is defined by (6.1) and x, E D(J,), then

n-1

(1 + l/tYA4x,, Y,) 5 n (1 + (1 - c;)/t)AJ,x,t YO)

i=o
n-l n-l
+ (l/t) C c;(l + l/t)‘p(X,, Tyi) n (1 + (1 - Ck)/t)
i=O k=i+l
forallnz 1.

Proof. Since J,x, = (l/(1 + t))x, 0 (t/(1 + t))TJ,x,, we have for each n 2 0,

(1 + !W+‘AJ,x,, Y,,,) 5 (1 + l/0(1 - c,)(l + l/VN,x,, Y,)

+ c,(l + l/t)“(l/t)(l + t)p(J,x,, Ty,)
I (1 + l/t)(l - c,)(l + l/t)“p(J,x,, y,)

+ c,(l + l/t)“(l/t)(p(x,, TY,,) + @(4x,, Y,))

= (1 + (1 - c,)/t)(l + l/t)“p(J,x,, y,)
+ c,(l + l/t)“(l/t)p(x,, Ty,).
The result now follows by induction.

LEMMA 6.2. If (y,) and lz,,) are defined by (6.1), then

P(YO9 Y,) 5 ~P(Y, 9 zo) + a, Hzo 3 Tzo) for alln 2 1.
548 S.REICH and I. SHAFRIR

Proof. The sequence (p(z,, Tz,,) is decreasing because

P(Gl+19 RI+,) 5 PC%+19 n,) + Pm,, Tz,,,)

5 (1 - CnlAZ,, Tz,) + P(Zn,Zn+l)

= AZ,, Tz,).
Hence
P(Z09G+ I) 5 P(Zo
, Z”) + c, AZ, , Tz,)
5 P(Z09z,) + cn AZ0 9 Tzcl).
Since
P(YOl Y,,,) 5 P(Yo9 z,) + P(Z0, z,+J + P(Zn+i, Yn+i)
5 2P(Y,, z,) + P(Z0,Zn+,),
the result again follows by induction.

We now state and prove the first theorem of this section.

THEOREM 6.3.Let T be a nonexpansive mapping in a complete hyperbolic space X. Assume that

T satisfies the range condition at infinity and let the sequence [y,) be defined by (6.1). Assume
that ano 2 1 and set (for n L no)

z, = (1 - 1/4JY, 0 (1/4J_vn, K = clco(zj: j L n,),

and
d= inf(p(x, TX): x E D(T)).
Then
(a) KY,, K) =T,“-“, P(Y~,z,) = d.
(b) If X is uniformly convex, then lim z, exists.
n-m
(c) If X is a Banach space which is reflexive and strictly convex, then the weak lim z,
n-m
exists.
(d) If X is a Banach space and X* is (F), then u = the strong lim z, exists and y, - u is
“-0D
the unique point of least norm in cl(R(Z - T)).

Proof. Let the sequence lr;, x,,) satisfy (3.Q and for m L n, let K,,, = clco(zj: j I mj. Fix a
point z in the convex hull of lZj: m I j 5 n). On the one hand, lemma 6.2 shows that
lim sup p(y,, K,,) I lim sup p(yO, z,) I d.
n-02 n-oo

On the other hand, for each m 5 i 5 n and p > 1, we have, by lemma 6.1,

P(Zit 4, xrp) 5 (1 - 1/Q,)P(YO J,pXt,)+ (l/Ui)P(Yi Jt,Xt,)

9 9

(1 - cj/(l + tp))
>I P(YO, Jtpxtp)

+ [max(P(x,p, Tyj): 1 ~j I ill/t,.

 Nonexpansive iterations in hyperbolic spaces 549

Consequently, we have

Pk Jr&&) 5 (1 - l/(1 + g)P(Y,9 Jfp,J

+ [max(p(x+, TYJ: 1 22j 5 n)J/f,
+ b,P(Y,, &&)/(I + Q2
for some b, > 0. Hence

P(Y, 9z) 2 PCYCI

9 Jtpxtp) - Hz, Jfpxtp)

2 P(Ycl, Jtpxtp)41 + &)

- [max( P(x,, , rYj>: 1 ~ j ~ n)]/t,

- b,p(y,, J,+,)/(l + $)*.

Letting p --f 00 and using theorem 3.1, we conclude that p(yO, z) 2 d and p(y, , K,,,) 2 d. Since
the sequence l&y,, K,)) is increasing, it must be constant. This yields (a). Assertions (b), (c)
and (d) now follow from (a) as in the proof of theorem 4.5.
In order to establish our next result we need two more lemmata.

LEMMA 6.4. If (y,J is defined by (6.1) and (c,) is bounded away from 0 and 1, then for all k L 1,

Proof. Fix k I 1 and denote the limit of the decreasing sequence (p(y,, Ty,)] by L. Since
n+k-1 n+k-1 n+k-1

P(Yn 9 Yn+k) 5 c P(Yjv Yj+ 1) = c Cjh’jv TYj) 5 C cj P(YO9 SO),

j=n j=n ( j=n >

it is clear that lim sup p(y,, yn+k)/( Cj”L,“-’ Cj) 5 L. By (9, proposition I] we also have
n-m
P(y,, Ynik) 2 PtTY,, TYn+k) z &PI, TYr,+k) - d_h, TY,)

nik-1

z fl t1 - cj)-l(P( Yn+k 3 TYn+k) - dyn 3 TYn))

,=n

Since the sequence (c,] is assumed to be bounded away from 0 and 1, we deduce that

lit;“_itf P(Y,, Yn+k)

The result follows.

LEMMA 6.5. If (y,) is defined by (6.1) and (c,) is bounded away from 0 and 1, then the limit of
the decreasing sequence l&y,, , Ty,)] is independent of the initial point y,.

Proof. This lemma is a consequence of lemma 6.4.

550 S. REICH and I. SHAFRIR

THEOREM 6.6. Let D be a closed subset of a hyperbolic space X and T: D -, X a nonexpansive

mapping which satisfies (3.7). Assume that lc,,) is bounded away from 0 and 1 and that the
sequence (y,,] is defined by (6.1). Then:
(a) Forallkz 1,
n+k-1

tz P(Ynv Yn+l)/Cn = fz P(Ynt Yn+k) c cj

I( j=n >
n-1

= ic AYO 9 YJ jC, cj
I( >

= inf(p(x, TX): x E D,).

(b) If X is a Banach space which is reflexive and X* is (LUG), then the weak jl~(y,, - Ty,)
exists.
(c) If X is a Banach space and X* is (LUF), then the strong lim (y, - Ty,) exists and it
n-a,
is the unique point of least norm in cl(Z?(Z - T)).

Proof. Since lim p(y,, Y”+~)/c, is independent of the initial point y. by lemma 6.5, it must
n-oo
equal inf (p(x, TX): x E D(T)]. Assertion (a) now follows by combining theorem 6.3(a) with
lemma 6.4. Assertions (b) and (c) follow from (a) and theorem 5.6.
Theorem 6.6(a) provides an affirmative answer to the question raised at the end of [19].
We now mention several consequences of theorem 6.6 for the special case when X is the
Hilbert ball.

COROLLARY 6.7, Let C be a closed convex subset of the Hilbert ball (B, p) and T: C -+ C a
nonexpansive mapping. Assume that (c,) is bounded away from 0 and 1 and that the sequence
(y,) is defined by (6.1). Then T is fixed point free if and only if iit p(O,y,) = co.

Proof. Assuming that a subsequence (y,,) of ( y,) is bounded, we let R = liTsup p(yo , y,,)
and D = (x E C: lim supp(x, y,,) I R 1. It is clear that D is a nonempty, bounded, closed
k-m
convex subset of C. Since lim P(_Y,,. Ty,,) = 0 by theorem 6.6(a), it is also invariant under T.
Therefore it contains a fixid;oint of T (cf. [IO, p. 1201).

COROLLARY 6.8.If, in the setting of corollary 6.7, T has a fixed point, then the sequence ly,,)
converges weakly to a fixed point of T.

Proof. If the subsequence (y,,] converges weakly to z, then z is the asymptotic center of ly,,]
with respect to both C and B [lo, pp. 116-1171. Since lim p(y,, Ty,) = 0 by theorem 6.6(a),
n-03
we see that
lim sup p(Tz, yn,) = IiT s~~pp(Tz,TY,,,) 5 li~_supP(z,~~,).
k-m

Therefore z is a fixed point of T and Limr p(y,, , z) = lim p(y,, z). Hence z is the asymptotic
?I-+=
center and the weak limit of the whole sequence (y,).
 Nonexpansiveiterationsin hyperbolic spaces 551

As a matter of fact, this result remains true even if the sequence (c,) is not bounded away
from 0.
For a point (I in the Hilbert ball B or on its boundary, let 4,: B + (0, 00) be defined by

%7(X)= I1 - (XV@I201 - lx12)

and consider the set
E(a, k) = (x E B : 4,(x) < k],
where k > 1 - /al’. Recall [lo, p. 1001 that this ellipsoid is either a p-ball (if Ial < 1) or a
half-space (if la/ = 1).

COROLLARY 6.9. Let B be the Hilbert ball and T: B --* B a nonexpansive mapping. Assume that
[c,) is bounded away from 0 and 1 and that the sequence (Y,) is defined by (6.1). If T is fixed
point free, then (YJ converges strongly to e(T), a point on the boundary of B.

Proof. Since lim ~(0, Y,,) = co by corollary 6.7, we know that lim lY,,l = 1. Let e = e(T)
n-+m n-m
be the “sink point” obtained on p. 126 of [lo]. Then &(Y,+,) I &(y,) for all n. Hence
&(y,) I &(yo) for all n, lim(y,, e) = 1, and the strong jii_ y, = e.

7. IMPLICIT ITERATIONS
In this section we continue our study of products of nonexpansive mappings by replacing the
explicit iteration (6.1) with an implicit one.
Let X be a hyperbolic space and T C X x X a co-accretive operator with domain D(T) = D.
Let (rn: n = 0, 1,2, . . . ) be a positive sequence such that CzGo r, = 00. We denote the partial
sum C r=-, Ti by a,, and assume that for each i, D is contained in R((1 + ri)l 0 ri T). We can,
therefore, define the iterative process

YII+1 = J,~Y,, n = 0,1,2, ... (7.1)

where y, belongs to D and the operators {J,] are the resolvents corresponding to T. Such
infinite products of resolvents have already been studied in Banach spaces. See, for example,
[24, 3, 17, 81.

LEMMA7.1. If (y,] is defined by (7.1) and x, E D(J,) with t > maxlri: 0 5 i I n - l), then
n-1 n-l PI-1

P(Y,, 4x,) 5 n (1 -
i=o
ri/l)P(Yo, J,xt) + C
i=o
(ri/t) JJ (1 - ~k/t)
k=i+l I
p(_Y;,Xr).

Proof. For t > r,, we have, by the resolvent identity (3.6),

P(Yrl+ 17.4x,) = P(J,~Y,, J&n/Ox, 0 (1 - ~n/W,xt))

5 P(Y, , (m/Ox, 0 (1 - rn /O.~X,)

5 (~/OP(Y,, x,) + (1 - ~,/OP(Y,, 4-G.

The result now follows by induction.

552 S. REICH and I. SHAFRIR

LEMMA7.2. If (y,J is defined by (7.1), then the sequence (p(y,,, y,+i)/r,,) is decreasing.

Proof. Denote (1 + l/r,_,)y, O(l/r,_,)y,_, by x,,. Since y,,_i belongs to (1 + rn_l)y, 0

r,,-,Ty,, we have x,, E Ty, and x,,+i E Ty,+, . Hence

P(Y”, Y,+i) 5 P((1 + 0Y” 0 Kz, (1 + 0Yn+, 0 f&+1)

for all positive t. Substituting t = r,,, we obtain

P(Y,, y,+i) 5 p((l + r,)y, 0 r,x,, YJ

= r, P(Y, 9XJ
= (rn~rn-Mb-l~ _hh
as claimed.
In view of lemma 7.2, the proof of the next lemma is similar to that of lemma 6.2.

LEMMA7.3. If (y,) and (z,) are defined by (7.1), then

P(Y,, Y,) 5 ~P(Y,, zd + a,Azo, zJro, forallnr: 1.

The next theorem is an analog of theorem 6.3. In view of the preceding lemmata, it can be
proved in the same way.

THEOREM7.4. Let T be a co-accretive operator in a complete hyperbolic space X. Assume that

T satisfies the range condition at infinity and let the sequence (y,) be defined by (7.1). Assume
that ano 2 1 and set (for n L n,) z, = (1 - l/a,)y, 0 (l/a,)y,, K = clco(zj: j 2 no), and
d = inf(p(x, y): x E D(T) and y E TX]. Then
(a) P(Y~, K) = lim P(Y,, z,) = d.
“-0D
(b) If X is uniformly convex, then lim z,, exists.
n-+=0
(c) If X is a Banach space which is reflexive and strictly convex, then the weak lim 2,
n-m
exists.
(d) If X is a Banach space and X* is (F), then u = the strong lim z,, exists and y, - u is
n-m
the unique point of least norm in cl(R(1 - T)).

In order to establish an analog of theorem 6.6 we require two more lemmata.

LEMMA7.5. If (y,) is defined by (7.1) and M = sup(ri/ri+i: i L 0) < 00, then for all k L 1,

lim p(y, 9 yn+Nm = lim dy,, yn+k

n-m n-m

Proof. Denote the left-hand side limit, which exists by lemma 7.2, by L.. We may assume that
L is positive. We proceed by induction on k. Since the case k = 1 is obvious, we assume the
result is true for all 1 I j 5 k and prove it for k + 1.
Given a positive E, there is an integer N(E) such that
 Nonexpansive iterations in hyperbolic spaces 553

for all n r N and 1 5 j I k. Since x, = (1 + l/r,_,)Y, 0 (l/r,_,)Y,_, belongs to Ty,, and

x,+k = (1 + l/r,,+k-l)Yn+k @(l/rn+k-lbn+k-l to TYn+k, we have

&,I, Y,,+k) 5 p((l + t)y,, 0 6, (1 + f)Yn+k 0 %+k)

= d(l - f/r,-,)y, @ (t/r,-,)y,-l, (1 - t/r,+k-,)Y,+k @ (f/m+k-l)Yn+k-l)

for all 0 c t 5 min(r,_, , rn+k_l). Substituting t = (r,_rr,,+k_r)/(r,_r + f,,ck-r) and using (3.4)
we obtain
P(Y,, Yn+k) s (rn+k-l/(rn+k-, + rn-,))hy”-, 9 Yn+k)

+ (rn-l/(r,r+k-, + rn-,))P(Y,v Yn+k-1).

It follows that for n 2 N + 1

1 L - (2Mk + l)&.

LEMMA7.6. If (y,) is defined by (7.1) and lr,) satisfies either

(i) suplr,: n = 0,1,2, . ..] = cQ, or
(ii) inflr,: n = 0, 1,2, . . .] > 0,
then the limit of the decreasing sequence l&y,, y,+r)/r,) is independent of the initial pointy,.

Proof. Assume first that condition (i) holds, and let (z,] be another sequence defined by (7.1).
Denote the limits of (p(y,, y,+r)/r,,) and (p(z,, .z,+,)/r,) by L(y,) and L(zo), respectively.
By the triangle inequality we have

I&Y”? Yn+J - P(Z,, I,,lW” 5 WY”, z,) + P(Y,+, 9zn+,))/m

Hence
IL(Y,) - L(zo)l 5 lim inf 2p(y,, zo)/rn = 0.
"-CC

If condition (ii) holds, then by the first part of the proof we may assume that 0 c u I r, 5 b
for all n. Therefore we can apply lemma 7.5 and the result follows.

Now we can present an analog of theorem 6.6. It is established by the same method
of proof.

THEOREM7.7. Let X be a hyperbolic space and T C X x X a co-accretive operator which

satisfies (3.7). Assume that (r,) is a positive sequence which is either unbounded or bounded
554 S. REICHand I. SHAFRIR

away from 0, and that the sequence ly,) is defined by (7.1). Then:
(a) Forallkr 1,

lim p(y,, y,,+&/r,, = j’_” P(Y, FYn+k)

n+m

= h P(Y0, YJ
n-m
= inf(p(x, y): x E D(T) and y E TX).
(b) If X is a Banach space which is reflexive and X* is (LUG), then the weak
lim (y, - yn+&/rn exists.
n-m
(c) If X is a Banach space and X* is (LUF), then the strong lim (y, - y,+,)/r, exists and
n-m
it is the unique point of least norm in cl(R(I - T)).
We turn now to several consequences of theorem 7.7 for the special case of the Hilbert ball.
We begin with a lemma and with a definition.

LEMMA7.8. If X is a hyperbolic space, T C X x X a co-accretive operator, s > 0, and

Y E Wi) n D(T), then P(Y, J,Y)/s I inf(p(y, z): z E Ty).

Proof. If z belongs to Ty, then y = J,((l + s)y 0 sz) and therefore

P(Y, 4Y) 5 P((1 + S)Y 0 .% Y) = wty, z).

We shall say that a co-accretive operator is m-co-accretive if
R((1 + r)l @ rT) = X (7.2)
for all positive r.

COROLLARY 7.9. Let (B, p) be the Hilbert ball and T C B x B an m-co-accretive operator.
Assume that (r,,) is a positive sequence which is either unbounded or bounded away from 0,
and that the sequence (y,] is defined by (7.1). Then T is fixed point free if and only if
lim ~(0, y,) = 00.
n-m

Proof. If x E TX, then x = J,x for all positive r and p(y,+, , x) I p(y,, x) I . -. I p(yo, x).
Assuming now that a subsequence ly,,) of {y,l is bounded, we let R = liTs_upp(y,, y,,)
and D = lx E B: lim sup p(x, y,,) % R ). It is clear that D is a nonempty, bounded, closed
k-m

convex subset of B. Since (1 + l/r,_,)y, 0 ( l/r,,_l)y,_, belongs to Ty,,

inffAynf z): z E TY,,~5 P(Y,-, , yn)/rn-l.
Therefore lemma 7.8 shows that

P(Y,, 9J,Yn,) 5 Wdy,, , 4: z E Tu,,l 5 P(Y,,-, , y,k)/rnk-l.

Thus $_nmAY,, , J,ynk) = 0 by theorem 7.7(a) and D is invariant under J,. Therefore it
contains a fixed point of J, which is also a fixed point of T.
 Nonexpansive iterations in hyperbolic spaces 555

Remark. As a matter of fact, corollary 7.9, as well as corollary 6.7, remain true if the Hilbert
ball B is replaced by any complete hyperbolic space X the modulus of convexity of which
satisfies the following property: for each T, E and a bounded A C X, inf(d(a, r, E): a E A) > 0.

COROLLARY 7.10. If, in the setting of corollary 7.9, T has a fixed point, then the sequence {Y,)
converges weakly to a fixed point of T.

Proof. If a subsequence (Y,J converges weakly to z, then z is the asymptotic center of {Y,,).
Since lim p(Y,, JrY,) = 0 by theorem 7.7(a) and lemma 7.8, it follows, as in the proof of
n-m
corollary 6.8, that z is a fixed point of Jr, hence of T, and that z is the asymptotic center and
the weak limit of the whole sequence {y,l.

LEMMA 7.11. Let (B, p) be the Hilbert ball and T C B x B an m-co-accretive operator. If T is
fixed point free, then there exists a unique point e = e(T) on the boundary of B such that all
the ellipsoids E(e, k), k > 0, are invariant under all the resolvents J,., r > 0, of T.

Proof. We already know [ 10, p. 1261that for each positive r there is a unique point e, of norm
one such that all the ellipsoids E(e,, k), k > 0, are invariant under J,. We have to show that
e, = e, for all 0 < s < t. For a positive k, denote the p-closure of E(e,, k) by C. This subset of
B is p-convex, p-closed and invariant under J,. Let x belong to C and consider the mapping
S: C -, C defined by Sy = J,((s/t)x @ (1 - s/fly), y E C. This mapping is Lipschitzian with a
Lipschitz constant 1 - s/t. Its unique fixed point coincides with J,x by the resolvent identity
(3.6). Hence J,x belongs to C. It follows that all the ellipsoids E(e,, k), k > 0, are invariant
under J, and therefore e, must coincide with e,.

COROLLARY 7.12. If, in the setting of corollary 7.9, T is fixed point free, then the sequence (y,)
converges strongly to e(T), a point on the boundary of B.

Proof. We already know that lim lynl = 1 by corollary 7.9. Let e = e(T) be the “sink point”
n-m
obtained in lemma 7.11. Then &(y,+,) I &(y,) for all n. Hence &(y,) I 4,&J for all n,
lim (y,, e) = 1, and lim y, = e.
n-m n-03

8. NONLINEAR SEMIGROUPS

In this section we briefly consider continuous analogs of some of the discrete results obtained
in the previous sections.
Recall that a (continuous) semigroup (of nonlinear nonexpansive mappings) on a (closed)
subset D of a hyperbolic space (X, p) is a family of operators S(t): D -+ D, 0 5 t 5 co,
satisfying the following conditions:
s(t + S)X = S(&S(s)x for all s, t 2 0 and x E D; (8.1)
S(O)x = x for all x E D; (8.2)
S(t)x is continuous in f L 0 for each x E D; (8.3)
P(S(0% S(0Y) 5 P(X, Y) for all t 2 0 and x, y E D. (8.4)
556 S. REICHand I. SHAFRIR

Using the ideas of [6], we can show that such semigroups are generated by co-accretive
operators satisfying the following range condition:
cl@(T)) c R(( 1 + r)Z 0 rT) (8.5)
for all positive r.

THEOREM 8.1. Let X be a complete hyperbolic space and T C X x X a co-accretive operator

that satisfies the range condition (8.5). Then T generates a nonlinear semigroup on cl@(T)) via
the exponential formula S(t)x = lim J:/,,x, where t 1 0 and x E cl@(T)).
n-oo

LEMMA 8.2. Let X be a complete hyperbolic space, T C X x X a co-accretive operator that

satisfies the range condition (8.5), and S the nonlinear semigroup generated by T. Then for all
x in D(T) and t > s > 0,
(a) p(S(.s)x, J,x) I e-“‘p(x, J,x) + (l/t) j8 e-(s-ti’fp(S(r)x, x) dr;
(b) p(x, S(t)x) I t inf(p(x, y): y E TX].
Proof. To establish (a), let n L 1 and 0 I j I n - 1. Then we have, by the resolvent
identity (3.6),
p(J:A’x, J,x) = p(J$‘x, J,,,((s/nt)x @ (1 - s/nt)J,x))

I (.s/nt)p(x, J,,,x) + (1 - s/nt)p(Jj,,,x, J,x).

Induction yields
n-1

P(JL x, 4 xl 5 (1 - s/nt)“p(x, J,x) + c (.s/nt)(l - s/nt)“-j-$(x, J;,,x).

i=o
The result now follows by letting n tend to infinity.
To prove (b), we note that by lemma 7.8,
n-1
p(x, J:/,, x) I c p(J&, x, J:A’x) I np(x, Jl,,, x) 5 t inf(p(x, y): y E TX).
i=o

Remark. In the Banach space case, the inequality in part (a) of this lemma is due (with a
different proof) to Plant [ 151.

Combining lemma 8.2 with the methods of the previous sections we obtain the following
analog of theorems 6.3 and 7.4.
THEOREM 8.3. Let T be a co-accretive operator in a complete hyperbolic space X. Assume that
T satisfies the range condition (8.5) and let S be the nonlinear semigroup generated by T. Fix
a point x in cl@(T)) and set (for t 2 I), zl = (1 - l/t)x @ (l/t)S(t)x, K = clcolz,: s L l],
and d = inflp(u, u): u E D(T) and u E Tu). Then
(a) P(X, K) = pi; P(X, z,) = d.
(b) If X is uniformly convex, then lim zt exists.
(c) If X is a Banach space which is iezexive and strictly convex, then the weak lim zI exists.
(d) If X is a Banach space and X* is (F), then v = the strong lim z, exists ini x - u is
I-m
the unique point of least norm in cl(R(Z - T)).
 Nonexpansive iterations in hyperbolic spaces 557

As a matter of fact, theorem 8.3 remains true, without the identification of the limit, for any
nonlinear semigroup defined on a closed subset of X, even if it is not assumed to be generated.
This can be established by modifying the methods of [16].
Although we do not have at present an analog of theorems 6.6 and 7.7 in the spirit of (21, we
use theorem 8.3 to prove a result which is related to corollaries 6.9 and 7.12.

THEOREM 8.4. Let (B, p) be the Hilbert ball, T c B x B an m-co-accretive operator, and S the
nonlinear semigroup generated by T on the p-closure of its domain. If inf(p(u, u): u E D(T)
and u E Tu) is positive, then for each x in the domain S the strong lim S(t)x = e(T), a point
f-r0
on the boundary of B.

Proof. Since T is fixed point free, there exists, by lemma 7.11, a unique point e = e(T) on
the boundary of B such that all the ellipsoids E(e, k), k > 0, are invariant under all the
resolvents J,, r > 0, of T. Let C be the domain of S. Given a point x in C, there is a positive
k such that x belongs to E(e, k). Since S is defined by the exponential formula, $,(S(t)x) I k
for all t 1 0. But lim p(x, S(t)x)/t is positive by part (a) of theorem 8.3. Hence
f+m
lim (S(t)x[ = 1, lim (S(t)x, e) = 1,
r-m I-rrn
and the result follows.

Acknowledgements-The first author was partially supported by the Fund for the Promotion of Research at the Tech-
nion, and by the Technion VPR Fund.

REFERENCES

1. ANDERSON K. W., Midpoint local uniform convexity and other geometric properties of Banach spaces, Ph.D.
thesis, University of Illinois, Urbana (1960).
2. BAILLON J.B., BRUCK R.E. & REICH S., On the asymptotic behavior of nonexpansive mappings and semigroups in
Banach spaces, Houston J. Math. 4, l-9 (1978).
3. BREZIS H. & LIONS P. L., Produits infinis de r&.olvantes, Israel J. Math 29, 329-345 (1978).
4. BROWDER F. E., Nonlinear mappings of nonexpansive and accretive type in Banach spaces, Bull. Am. math. Sot.
73, 875-882 (1967).
5. BUSEMANN H., Spaces with non-positive curvature, Acta Mafh. 80, 259-310 (1948).
6. CRANDALL M. G. & LIGCETT T. M., Generation of semigroups of nonlinear transformations on general Banach
spaces, Am. J. Math. 93, 265-298 (1971).
1. FAN K. & GLICKSBERG I., Some geometric properties of the spheres in a normed linear space, Duke math. J. 25,
553-568 (1958).
8. FUJIHARA T., Asymptotic behavior of nonexpansive mappings in Banach spaces, Tokyo J. Math. 7, 119-128
(1984).
9. GOEBEL K. & KIRK W. A., Iteration processes for nonexpansive mappings, Contemp. Math. 21, 115-123 (1983).
10. GOEBEL K. & REICH S., Uniform Convexity, Hyperbolic Geometry and Nonexpansive Mappings. Marcel Dekker,
New York (1984).
11. HOYOS GUERRERO .I. J., Differential equations of evolution and accretive operators on Finsler manifolds, Thesis,
University of Chicago (1978).
12. KIRK W. A.,Krasnoselskii’s iteration process in hyperbolic space, Num. Funcf. Analysis Opfim. 4, 371-381 (1982).
13. KOHLBERG E. & NEYMAN A., Asymptotic behavior of nonexpansive mappings in normed linear spaces, Israel J.
Math. 38, 269-275 (1981).
14. KUCZUMOW & STACHURA A., Fixed points of holomorphic mappings in the Cartesian product of n unit Hilbert
balls, Can. marh. Bull 29, 281-286 (1986).
15. PLANT A. T., The differentiability of nonlinear semigroups in uniformly convex spaces, Israel J. Mafh. 38.257-268
(1981).
16. PLANT A. T. & REICH S., The asymptotics of nonexpansive iterations, J. Funcf. Analysis 54, 308-319 (1983).
17. REICH S., On infinite products of resolvents, Atti. Acad. naz. Lincei Rc. 63, 338-340 (1977).
558 S. REICHand I. SHAFR~

18. REICH S., On the asymptotic behavior of nonlinear semigroups and the range of accretive operators, J. math.
Analysis Applic. 79, 113-126 (1981).
19. REICHS., Averaged mapping; m the Hilbert ball, J. math. Analysis Applic. 109, 199-206 (1985).
20. REICH S., Integral equations, hyperconvex spaces and the Hilbert ball, in Nonlinear Analysis and Applications
517-525. Marcel Dekker, New York (1987).
21. REICHS., Fixed point theory in the Hilbert ball, Contemp. Math. 72, 225-232 (1988).
22. REICHS. & SHAFRIRI., On the method of successive approximations for nonexpansive mappings, in Nonlinear and
Convex Analysis 193-201. Marcel Dekker, New York (1987).
23. REICH S. & SHAFRIRI., The asymptotic behavior of firmly nonexpansive mappings, Proc. Am. math. Sot. 101,
246-250 (1987).
24. ROCKAFELLAR R. T., Monotone operators and the proximal point algorithm, SIAM J. Control Optim. 14.877-898
(1976).
25. St.nr~ M. A., Some examples concerning rotundity in Banach spaces, Math. Ann. 233, 155-161 (1978).

Note added in proof-(l) Using an idea of J. M. Borwein, we are now able to show that theorems 6.6 and corollaries
6.7, 6.8 and 6.9 continue to hold even if the sequence lc,) is not bounded away from 0. (2) There is a continuous analog
of theorems 6.6 and 7.7. See the paper by the first author entitled “The asymptotic behavior of a class of nonlinear
semigroups in the Hilbert ball” which is to appear in J. Math. Analysis Applic.