EIGENVECTORS OF NONLINEAR POSITIVE OPERATORS
AB~ THE LINEAR KREIN-RUTMAN THEOP~
BY
@
ROGER D, NUSSBAUM
Mathematics Department
Rutgers University
New Brunswick, New Jersey 08903
O, INTRODUCTION
In a recent pa~er [13] Massabo and Stuart prove an existence theorem for non-
zero eigenvectors of a nonlinear operator which maps a normal cone into itself. They
conjecture that normality of the cone is unnecessary; in Section I below we prove their
conjecture. Our proof is quite different from that of Massabo and Stuart and involves
some results from asymptotic fixed point theory. We hope that even the relatively
simple case considered here will illustrate the usefulness of these ideas.
In the second section of this paper, which is essentially independent of the
first, we prove a new fixed point theorem for nonlinear cone mappings. We then prove
that our nonlinear theorem implies as a corollary the most general versions of the
linear Krein-Rutman theorem. Finally, we discuss briefly an example of a linear ope-
rator which is best studied in non-normal cones.
Although the linear theorem we obtain is new, our central point is methodolo-
gical. The linear Krein-Rutman theorem has played an important role in the study of
nonlinear cone mappings, particularly in computing the so-called fixed point index of
such mappings. Our results cone full circle and show that the linear Krein-Rutman
theorem follows from a simple fixed point theorem. Partial results in this spirit
have been obtained before [3, 8, L3, 18, 20, 21], but here we avoid unnecessary hypo-
theses like normality of cones (see Section 5 of [18]).
The approach to the linear Krein-Rutman theorem given here is suita~ole for a
*Partially supported by a National Science Foundation Grant.
310
course on nonlinear functional analysis, and in fact that was our original motivation
for obtaining the results in Section 2. After development of the Leray-Schauder degree
theory, the most general versions of the Krein-Rutman theorem for linear compact maps
can be obtained in one lecture by our method.
1, EIGENVECTORS OF NONLINEAR CONE MAPPINGS
By a cone K in a Banach space X we mean a closed subset of X such that
(i) if x,y E K and k and ~ are nonnegative reals, then Xx+~y E K and (2) if
x E K-{O }, then -x ~ K. If K only satisfies (1), K is a "wedge". Notice that
induces a partial ordering on X by x < y if and only if y-x E K. A cone K is
"normal" if there exists a positive constant T such that for all elements x and
of K, Ilx+ylI ~ zlIxIl. The cone of nonnegative functions in C[O,I] or LP[0,1],
1 ~ p ~ ~, is normal; the same cone in ck[o,1], k ~ 1, or in a Sobolev space (other
than L p) is not.
We also need to recall Kuratowski's notion of measure of noneompactness [10].
If S is a bounded subset of a Banach space X (or, more generally, of a metric
space) define a(S), the measure of noncompactness of S, by
I n 1 ~(S) = inf d>O: S = U Si, n < ~ and diameter (Si) _~ d for l~i_~n . i=1
In general suppose that ~ is a map which assigns to each bounded subset S of X
a nonnegative real number ~(S). We will call ~ a generalized measure of noncompact-
ness if ~ satisfies the following properties:
(1 ~(S) = 0 if and only if the closure of S is compact.
(2 ~(c~(S)) : ~(S) for every bounded set S in X (c~(S) denotes the t
convex closure of S, i.e., the smallest closed, convex set which eontains
S).
(3 ~(S+T) ~ ~(S)+~(T) for all bounded sets S and T, where
S+T = {s+t: s E S, t E T}.
(4) ~(S U T) = max(~B(S), ~3(T)).
The measure of noncompactness ~(S) is well-known to satisfy properties 1-4. Only
311
property 2, first proved by Darbo [5], presents any difficulties.
If D is a subset of a Banach space X, ~ is a generalized measure of non-
compactness, and f:D ~ X a continuous map, f is called a k-set-contraction with
respect to ~ if
$( f (S ) ) .e k#(S) (1.1)
for every bounded set S in D. If ~ = a we shall simply say that f is a k-set-
contraction. Now suppose that C is a closed, convex subset of X and that W is
a bounded, relatively open subset of C (so W = 0 n C for some open subset 0 of X).
Assume that f:W ~ C is a k-set-contraction with respect to ~ and that k < 1.
If {x ( W: f(x) = x} is compact or empty or (less generally) if f is a k-set-
contraction on W and f(x) $ x for x ( W-W, it is proved in [16] that there is
defined an integer ic(f, W), the fixed point index of f:W ~ C, which is roughly an
algebraic count of the fixed points of f in W. We shall only need a few facts about
the fixed point index. If ic(f , W) r 0, then f has a fixed point in W. If
fs(X) = sf(x) for 0 ~ s ~ 1 and fs(X) $ x for x ( W-W, then iC(fs,W ) is constant
for 0 ~ s ~ 1 and ic(f,W ) = iC(fo,W) = i if 0 ( W. If C is a cone (or a wedge),
x ~ ( C, ft(x) = f(x)+tx ~ for t > 0 and ft(x) # x for x ( W-W, then ic(ft,W ) is
constant for 0 ~ t.
In the situation described above, the fixed point index can be described in
= -- = co f(W R and let terms of Leray-Schauder degree. Define C 1 co f(W), C n Cn_ 1)
D denote a compact, convex set such that n C c D c C and f(W n D) c D (such n~l n --
a D exists). Let 0 be any bounded open set in X such that 0 rl D = W n D and
let g:0 ~ D be a continuous map such that glW N D = flW n D (g exists by virtue
of a theorem of Dugundji [6]). One can define ic(f, W) = deg(I-g, 0, 0) (observe
that the fixed point set of g in 0 is compact, so the Leray-Schauder degree can
be defined) and prove that the definition is independent of the particular D, g and 0
as above. Properties of the fixed point index now follow from properties of the Leray-
Schauder degree, and this is especially easy to see if f is compact.
We want to generalize now following theorem of Massabo and Stuart [18].
THEOREM (Massabo-Stuart [ L3 ] ) .
Let C
open subset of
that
be a normal cone in a Banach space X, let ~ be a bounded, relatively
C containing 0, and let f:C -~ C be a k-set-contraction. Suppose
312
8 > kd/y ( ] . 2 )
where 6 = inf{l[f(z)[I: z E ~-~}, d = max{l[z[J: z E ~-~} and y is the constant appearing
in the definition of a normal cone. Then there exist t > 0 and z E 8K(~) (the
boundary of ~ as an open subset of K) such that f(z) = tz.
The chief tool we shall use is an "asymptotic fixed point theorem". Theorem l.l
below is a special case of Theorem 3 in [15] or Propositions 2.4 and 3.1 in [17].
(i.i) THEOREM
Let C be a closed, convex subset of a Banach space X, U a bounded, relati-
vely open subset of C and f:U~ U a k-set-contraction with k < i. Assume that
fn(u) is contained in U for some integer n. Then ic(f, U) is defined, Lgen(f),
the generalized Lefschetz number of f:U ~ U is defined and Lgen(f ) = ic(f , U). In
particular, if Lgen(f ) # 0, f has a fixed point in U.
We have not defined here Leray's generalized Lefschetz number [ii], but it
suffices to know a few facts. The generalized Lefschetz number agrees with the ordi-
nary Lefschetz number for f:U ~ U if the ordinary Lefschetz number is defined. If
fm(u) c y c U, f(y) c y and g denotes f:Y ~ Y, then Lgen(f) = Lgen(g ). In par-
ticular, if Y is homotopic in itself to a point, Lgen(f ) = i.
We can now show that the assumption of normality in Theorem i is unnecessary,
at least if the set ~ is "radial".
( 1 . 2 ) THEOREM
Let C be a cone in a Banach space X and let s be a bounded, relatively
open neighborhood of 0 in C. Assume that for each x E S d~f {x E C: Jlxll = I}
there is a unique real number t = t x > 0 such that tx E 8C(~), where ~c(~) denotes
the boundary of ~ as a subset of C. Let f:Sc(~) ~ C be a k-set-contraction and
suppose that
6 > kd ( 1 . 3 )
where 6 = inf{ Iif(z) ll: z E 8C(2)} and d = sup{ [Izll: z E ~C(~)}. Then there exists
t >_ 6d -I and z E 8C(~ ) such that f(z) = tz.
313
Proof
D e f i n e s = {x E C: Ilxlt = d}. For each x E 2, t h e r e e x i s t s a u n i q u e s = s x
~C(a ) , < 1 The assump- such t h a t sx ~ and by t h e d e f i n i t i o n o f d one has 0 < s x _ .
t i o n t h a t s x i s u n i q u e i m p l i e s e a s i l y t h a t R(x) d~f Sx x i s a c o n t i n u o u s map.
S ince t h e image o f any s e t A in ~ l i e s i n c-~(A U {0}) , we f i n d t h a t R i s a 1- d e f
s e t - c o n t r a c t i o n . I t f o l l o w s t h a t f l ( x ) = f (Rx) i s a k - s e t - c o n t r a c t i o n and
i n f IIf~ (x) ll = 5 x~3
I f we can p r o v e t h a t Theorem 1 .2 i s v a l i d f o r t h e s p e c i a l c a s e ~ = Ix ~ C: Ilxl] < d},
t h e d i s c u s s i o n above shows t h a t t h e r e e x i s t s x ~ 3 and s ~ 5d -1 such t h a t f l ( x ) = s x .
I f we w r i t e z = R(x) ( @ C ( a ) , i t i s c l e a r t h a t z i s an e i g e n v e c t o r o f f w i t h
e i g e n v a l u e t ~ 6d -1 .
The above discussions shows that it suffices to prove Theorem 1.2 in the case
that ~ = {x ~ C: llxll < d}. In this case define g:E ~ E by g(x) = dr(x) . If A
IIf(x)l l
i s any s u b s e t o f 3, i t i s c l e a r t h a t
g(A) c d S - i { t f ( x ) : 0 ~ t ~ 1, x ~ A}
c d6-ic-6-(f(A) U {o})
(1.4)
where for a set T, XT d~f {kx: x E T}. Using (1.4) and the basic properties of the
measure of noncompactness, one finds that
~ ( g ( A ) ) ~ kd6-1~(A) ( 1 . 5 )
so g is a c-set-contraction for c = kd5 -1 < i.
For e > 0 let U= {x ~ C: d-s < llxll < d+E} and define a retraction dx
r: U~ 3 by r(x) - . By reasoning like that above one can see that r is a
-1 Ilxrl -1 d(d-e) -set-contraction, so h(x) = g(r(x)) is a cd(d-e) -set-contraction.
lect ~ so small that c I = cd(d-s) -I < i and observe that fixed points of h:U ~ U
are the same as the fixed points of g:Z ~ E. Since fixed points of g correspond
to the desired eigenvectors, it suffices to find a fixed point of h. By Theorem i.i
and the remark immediately following it, h will have a fixed point in U and Lgen(h)=l
if Z is homotopic in itself to a point. To show 3 is homotopic to a point, let
p:K-{O} ~ E be radial retraction, fix x E s and define a deformation ~: 3x[0,1]~3 o
by
314
~(x,t) : p[(l-t)x+tx ]. (1.6) o
The fact that K is a cone insures that (1-t)x+tx ~ # 0
is well-defined and continuous. �9
for 0 _< t _< 1 and x ( Z, so
(1 .1 ) REMARK
If X is an infinite dimensional Banach space, S = {x E x:llxll = d > 0} and
f: S + S is a k-set-contraction with k < 1, then it follows from a remark on p. 373
of [15] that f has a fixed point in S. (Note that S is a continuous retract
of B = {x ( K: llxll ! d} when X is infinite demensional, so S can be deformed in
itself to a point). The proof of Theorem 1.2 follows by essentially the same trick.
(1.2) Remark
We have not proved Theorem 1.2 for general neighborhoods ~ of the origin in
K, although we conjecture it is true for such neighborhoods. In fact, Massabo and
Stuart use their theoren for general ~ in proving Theorem 1.2 oF [13]. However,
as is observed in [13], Theorem 1.2 of [13] is a global bifurcation result which can
be proved (for general cones) by using the fixed point index in the same way degree
theory is used to prove the classical Banach space version of Theorem 1.2 in [13].
(1.3) Remark
The problem of proving Theorem 1.2 for general ~ may be related to an exten-
sion problem for certain functions. I am indebted to Heinrich Steinlein for a conver-
sation which led to the following observation.
(1.3) THEOREM
and
Let ~ be a bounded, relatively open neighborhood of the origin in a cone
f: 8K(~) ~ K-{0} a continuous map. Define
s 1= {x~ K: llxlt = 1}, d 1: inf{HxJl: x~ ~K(~)},
315
d 2 = sup{Hxll: x ~ ~K(~)} an d A = { ( x , X ) E S l X [ d l , d 2 ] :
D e f i n e a map cp:A -+ S 1 b y
f(Xx) ~p(x,X) = ~ .
Xx ff a K ( ~ ) } .
Assume that there exists a continuous map ~: SlX[dl,d2J ~ S 1 such that (i) ~IA =
and (ii) ~ is a c-set-contraction for some c < 1. Then there exists x ~ aK(~)
and t > 0 such that f(x) = tx.
Outline of proof
Define ~x(X) = ~(x,X). One can associate a fixed point index to ~x:S1 + S 1
and prove that this fixed point index equals one. Using this fact one can prove that
there exists a connected set C c SlX[dl,d2] , C c {(x,X): ~(x,~) = x}, such that C
has nonempty intersection with SlX{dl} and SlX{d2}. It follows that C must
intersect aK(~) , which is the desired result. W
Under the hypotheses of Theorem 1.2, one can prove that there exists an exten-
sion ~ as in Theorem 1.3, so Theorem 1.2 is a consequence of Theorem 1.3. Further-
more one can derive the original Massabo-Stuart theorem from Theorem 1.3. Thus the
known results reduce to extending ~ to �9 in such a way that �9 is a c-set-con-
traction, c < 1.
2i LINEAR AND NONLINEAR KREIN-RUTMAN THEOREMS
The linear Krein-Rutman theorem has been used to calculate the fixed point in-
dex of certain nonlinear cone mappings and to obtain fixed point theorems for such
mappings. We shall show here that the most general linear Krein-Rutman actually follows
by elementary arguments using the fixed point index for cone mappings. We wish to
emphasize the elementary nature of our proof. We do not need the apparatus of asymp-
totic fixed point theory, and (at least for compact linear operators) our proof is
suitable for a course in nonlinear functional analysis as an application of the Leray-
Schauder degree.
316
We begin by recalling a trivial but useful observation of Bonsall [1].
(2 .1 ) LEMMA (Bonsall [ 1 ] )
Let {am: m ~ i} be an unbounded sequence of nonnegative reals.
exists a subsequence {am. : i > i} such that i
Then there
(1) a ~ i m. i
(2) a >_ a. for 1 _< j _< m.. m. j i 1
If C is a cone in a Banach space and f:C ~ C is a map, we shall say that
f is "order-preserving" if 0 } u } v implies f(u) ~ f(v) for all u and v in
C. We shall say that f is "positively homogeneous of degree 1" if f(tu) = tf(u)
for all real numbers t ~ 0 and all u ~ C.
(2 .1 ) THEOREM
Let C be a cone in a Banach space X and f:C ~ C an order-preserving map
which is positively homogeneous of degree 1. Assume that there exists a generalized
measure of noncompactness ~ such that f is a k-set-contraction with respect to
and k < 1. Assume that there exists u E C such that {llfm(u) ll: m ~ 1} is unboun-
ded. Then there exists x E C with I[xll = 1 and t ~ 1 such that f(x) = tx, and
i f f ( y ) r y f o r Ilyll = 1 and U = {y ~ c: Ilylt < 1 } , i c ( f , U ) = O.
(2.1) REMARK
If C is a normal cone, Theorem 2.1 is a very special case of Proposition 6
on p. 252 of [18], so the whole point of the following argument is that it applies
to nonnormal cones.
(2 ,2 ) RE~IRK
Suppose that g:V = {x E C: Nxll < ~ ~ c is a k-set-contraction w.r.t. ~,
k < 1, and that f is as in Theorem 2.1. If tf(x)+(1-t)g(x) ~ x for Ilxll = R, and
0 ~ t ~ 1, then there exists k ~ 1 and x ~ C with Ilxll = R such that f(x) = kx.
If not, the homotopy gs(X) = sg(x), 0 ~ s ~ 1, shows that ic(g ,~ = 1, while the
homotopy tf(x)+(1-t)g(x) shows that ic(g,~ = ic(f,~ = 0, a contradiction.
317
Proof of theorem (2.1)
Let ~ = Ix ~ C= IIxll < i} and assume that there does not exist t ~ 1 and a
point x ~ C with llxl] = 1 such that f(x) = tx. If we define fs(X) = sf(x) for
0 ~ s ~ 1, the above assumption shows that fs(X) r x for llxl] = 1 and 0 5 s ~ 1.
The homotopy property of the fixed point index implies that
ic(fl, 6) = ic(f, ~) = ic(fo, ~) = 1. (2.1)
If we can prove that ic(f,~ ) = O, we will have a contradiction of our original
assumption and the theorem will be proved. It is well-known that I-fl~ is a proper
map (the inverse image of any compact set is compact), so there exists 6 > 0 such
that llx-f(x) ll ~ 6 for x ~ C, Hxll = 1 (since we are assuming x-f(x) r 0 for
IIx]l = 1). Let u be as in the statement of the theorem. Because f is homogeneous
of degree 1, we can, by multiplying u by a small positive constant, assume that
0 < ]lull < 6 and {[]fm(u) ll: m ~ 1} is unbounded. Define a function g(x) by
g(x) = f(x)+u . (2.2)
The homotopy f(x)+su for 0 5 s 5 1 has no fixed points x with ]]xl] = 1 so
i c ( f , ~ ) = i c (g ,~ ) . (2.3)
To complete the proof it suffices to show ic(g,~ ) = O; and to prove
ic(g,~ ) = O, it suffices to prove that g has no fixed points in ~. Thus we assume
that g(x) = x for x ~ ~ and try obtain a contradiction.
If g(x) = x = f(x)+u we have
x ~ u. (2.4)
In general:, for purposes of induction, assume that
x ~_ ~ ( u ) . (2.5)
It follows from the order-preserving property of f that
x = g ( x ) = f ( x ) + u > f ( x ) >_ f ( f m u )
= fm+l (u)
so e q u a t i o n ( 2 . 5 ) h o l d s f o r a l l m > 0 by m a t h e m a t i c a l i n d u c t i o n .
(2.6)
318
Now we apply Lemma 2.1. Define a = llfm(u) ll, an unbounded sequence of nonne- m
gative reals, and let {ami : i > 1} satisfy the conclusions of Lemma 2.1. I]efine
fm(u) and let s = _ {Vmi - : i ~ 1}. We claim that ~(Z) = 0, so that Z has V m
II fm (u) lJ compact closure. To see this observe that for j > ~ we can write (using homogeneity
of f)
m ,
S i n c e Hfmi-J(u) ll < Irf l ( u J I P
J Z = ~_~ {Vmi } U fJ(Tj)
i = 1
Z. J
def =
I l l m i ( u ) ] l > j}
for i t j, we have
2.7)
T . �9 B = {x e K: ilxl] ~ i} J
2.8)
Equations (2.7) and (2.8) imply that
J Z c U {Vm. } U fJ (B) (2 ,9
i=1 i
Since fJ is a kJ-set-contraction with respect to ~, equation (2.9) implies
~(Z) _~ kJ~(B) (2.10)
The right hand side of (2.10) approches zero as j + ~, so ~(Z) = 0, Z is compact,
def and for some s u b s e q u e n c e Vm. = wj we c a n a s s u m e t h a t wj c o n v e r g e s s t r o n g l y
J t o w. Of c o u r s e Nwll = 1 a n d w ~ K.
If we now return to equation (2.5) and divide both sides by
Taking the limit as j +
theorem. R
am. we obtain i. J
(am.)-i x-wj ~ K (2..ii lj
yields that -w E K, and this contradiction proves the
Our first corollary generalizes Theorem 4.2 in [13] by removing the assumption
of normality of the cone. If the cone C is normal, however, Corollary 2.1 is an
easy corollary of an earlier result, Proposition 6 on page 252 of [18]. If the ope-
rator g below is compact and p = 1, Corollary 2.1 below is a result of Krein and
Rutman [9, Theorem 9.1].
319
( 2 . 1 ) COROLLARY
(Compare [13] and [ 1 8 ] ) . Let C be a cone in a Banach space X, ~ a g e n e r a l i -
zed measu re o f n o n c o m p a c t n e s s such t h a t ~ ( t S ) = t ~ ( S ) f o r e v e r y bounded s e t S in
C and e v e r y r e a l t T 0, and g:C + C a k - s e t - c o n t r a c t i o n w . r . t . 9- Assume t h a t
g i s o r d e r - p r e s e r v i n g and p o s i t i v e l y homogeneous o f d e g r e e 1 and t h a t t h e r e e x i s t s
u E C-{0} and e > k p such t h a t gP(u) ~ cu. Then t h e r e e x i s t s x ~ E C-{0} and
and X ~ c ( p - l ) such t h a t g(Xo) = kx o.
Proof
Let b be any real such that
k P < b < c .
-1 If q = q , define f(x) = b-qg(x)~ Using the homogeneity of f, it is clear that
fP (u ) ~ r u , r = cb -1 > 1 ( 2 . 1 2 )
and that f is a kl-Set-contraction w.r.t. 9, where k I = kb -q < 1.
By Theorem 2 .1 , f w i l l have an e i g e n v e c t o r in C o f norm 1 i f {llfm(u) ll: m~l}
i s unbounded . But e q u a t i o n ( 2 . 1 2 ) i m p l i e s t h a t
f J P ( u ) u > 0.
r j
I f I l fJP(u) ll were bounded , we would o b t a i n by l e t t i n g
a contradiction.
j ~ ~ in ( 2 . 1 3 ) t h a t
(2.13)
-u E C,
Theorem 2.1 t h u s i m p l i e s t h a t t h e r e e x i s t s x ( C, Ilxll : 1, and t ~ 1 such
t h a t f ( x ) = t x . W r i t i n g i n t e r m s o f g , t h e r e e x i s t s k ~ h q such t h a t
g ( x ) = Xx. ( 2 . 1 4 )
S e l e c t an i n t e g e r N such t h a t c_N-1 > kr .n -1 c - n = b , t h e above r e m a r k s show t h a t t h e r e e x i s t
such t h a t
If n ~ N and if we define
K, IlXnll = 1 and X > ( c - n - l ) q Xn n -
g(Xn) = XnXn. (2.15)
320
we define Z = ~ IXn: n ~ N}, �9 B = c-N -1 If and
that f kl-set-contraction with respect to
shows that
f(x) = B-qg(x), previous remarks show
for some k I < i. Equation (2.15)
E c ~%-(f(Z) U { 0 } ) , ( 2 . 1 6 )
Equation (2.16) implies that ~(s ~ klP(Z), so ~ is compact.
X > c q and a subsequence we can assume Xn. + x and kni 1
Therefore, by taking
f(x) = Xx, X ~ c q (2.17)
which is the desired result. �9
We now wish to show how Theorem 2.1 can be used to obtain a general version of
the linear Krein-Rutman theorem. First we need to recall some definitions. If X is
L:X ~X is a bounded linear operator, define r(L), the spectral a Banach space and
radius of L, by
If X is real and
I 1
r ( L ) d=ef lira HLni] n = inf I[Lntl n �9 (2.18)
n~ n_ > i
denotes the complexifieation of X, X = (x+iy: x,y E X1 with
IJx+iyll = sup I I(cos e) x+(sin e)yil o~ee2~
then L has an obvious linear extension L and Ilerl = I I L I I .
spectrum of L, then of course
If ~(g) denotes the
r(L) = sup{lzl: z ~ ~(~), z complex}.
If ~(L) is defined by
a(L) = inf{c ~ O: L is a c-set-contraction}
it is clear that ~(L) ~ I I L I [ .
and we can define p(L) by
Because a(LiL2) 2 ~(Li) ~ (L2) , lim
1 1
~(L) = lira (a(Ln))n = inf (a(Ln)) n. n-~o n~l
If L is as above, one can prove that ~(L) = ~(L).
o(L), the essential spectrum of L, and it is proved in [14] that
There is a subset
(2.19")
1
(~(Ln)) n exists,
( 2 . 2 0 )
ess(L) of
p(L) = sup{ I~1: ~ ~ e s s ( L ) } . ( 2 . 2 ~ )
321
Actually, there are several inequivalent definitions of the essential spectrum, but
(2.21) is valid for all of them. It is proved in C14] that if s > O, there is at
most a finite number of complex numbers z such that z E d(L) and Izl ~ p(L)+e.
If z is such a complex number and F a simple closed curve in ~ which contains
no other point of ~(L) in its interior or on F, then the spectral projection P
corresponding to z,
if F
is finite dimensional.
If K is a cone in X
L(K) c K, define numbers IILII K
and L:X ~X is a bounded linear operator such that
and aK(L) by
IILIIK def= sup{tlgull: u E K, rlull ~ 1}
aK(L ) = inf{c ~ O: LIK is a c-set-contraction}.
( 2 . 2 2 )
In analogy r(L) and p(L), define rK(L), the cone spectral radius of
PK(L), the cone e s s e n t i a l s p e c t r a l r ad ius of L, by
1 def
rK(L ) = lira (IILnN K )n
n ~ ( 2 . 2 3 )
1
PK(L ) def lim (~K(Ln)) n . (2.23) n~
As was remarked in [18] it is easy to see that
PK(L) ~ p(L); if the cone K
prove that PK(L) 2 rK(L ).
if LIK is compact.
L, and
p(L) A r(L), rK(L ) 2 r(L) and
is "reproducing" (so X = {u-v: u,v E K}) one can also
Note also that p(L) = 0 if L is compact and PK(L) = 0
We need to recall one more definition. If K is a cone in a Banach space X
with norm If'If, define Y = {u-v: u,v ~ K}. Define a norm I-l on Y by
lyl : i n f { I/ull+llvll: u, v ~ K, y = u-v} . (2 .24)
It is remarked in [2, 22] (and is not hard to prove) that Y is a Banach space in
this norm, IIYH ~ IYl for y ~ Y, and IJYll = IYl for y E K. If L:X~X is a
bounded, linear operator such that L(K) c K, then L(Y) c y. Furthermore, if ILIy
denotes the norm of L considered as a map from the Banach space Y to itself, it
is easy to see that
322
so that
[LIy = [ILil K ( 2 . 2 5 )
1 1 -- def
lim ILnJ n : ry(L) : rK(L) : lira I[LnIIK . n~oo Y n~o~
( 2 . 2 6 )
In (2.26), ry(L) denotes the spectral radius of L as a map from Y to Y.
With these preliminaries we can prove our linear Krein-Rutman theorem.
( 2 . 2 ) THEOREM
Let X be a Banach space, K a cone in X and L:X ~ X a bounded linear
operator such that L(K) c K. If PK(L) = v and rK(L ) = K are defined by equations
(2.22) and (2.23), assume that ~ < R. Then there exists x E K-{0} such that
Lx = bx.
Proof
d e f i n e
Let s n be a sequence of real numbers such that v < s n
gn(X) = snlL(x). Our assumptions imply that
< b and s -~ ~ and n
-i -1 1 < s ~ = rK(gn). (2.27) pK(gn) = s n v < n
For notational convenience, fix n, write g = gn' s = s n and select N so large
that ~K(g m) < i for m > N. Just as in the proof of Proposition 7 in [18], define
a generalized measure of noncompactness by
N-1 1 j~0 ~(gj (A)) (2.28) ~3 (A) = ~
One can easily check that ~ is a generalized measure of noncompactness, that
~(XA) = IX[~(A) and that there exists a constant c < 1 such that
~(g(A) ) ~ c~(A) (2 .29)
for every bounded set A a K.
323
If Y is the Banach space defined immediately before the statement of Theorem
2.2, we have seen that rK(g ) = ry(g) = the spectral radius of g as a map of Y to Y.
Since rK(g ) > i, it follows that Ignly is unbounded. The uniform boundedness prin-
ciple on the Banach space Y implies that there exists y E Y such ~hat Ign(y) l
is unbounded. It follows that there exists u E K such that Ignul = llgnull is
unbounded. Theorem 2.1 now implies that there exists t ~ i and x E K, Iixll = 1,
such that
g(x) = t x . ( 2 . 3 0 )
= Of course t = t n and x = x n depend on n, and if we write ~n Snt n we have
L(Xn) = ~nXn . ( 2 . 3 1 )
< < = We must have Sn -< ~n ~, so lim ~n ~. Exactly as in the last paragraph of the n->oo
p r o o f o f P r o p o s i t i o n 7 in [18] o r as in t h e p r o o f o f C o r o l l a r y 2 . 1 , {Xn: n _> 1} has
compact c l o s u r e , so one can assume by t a k i n g a s u b s e q u e n c e t h a t x -~ x E K, ]lxH = 1, n
and t a k i n g t h e l i m i t as n ~ ~ o f ( 2 . 3 0 ) g i v e s
Lx = ~x ( 2 . 3 2 )
which is the desired result. �9
(2 .3) REMARK
If LIK is compact, so ~K(L) = O, Theorem 2.2 generalizes a result of Bonsall
[1,2]. The original Krein-Rutman theorem deals with the case that L is compact as
a map of X to X, K is total and r(L) > O, and one might believe that, at least
for total cones, Bonsall's result is equivalent to the Krein-Rutman theorem. However
Bonsall gives a simple example which shows this hope is false.
Let X = {x E C[0,1]: x(O) = O} and define L:X +X by (Lx) (t) = x(�89 Bonsall
constructs, for each y > O, a total cone Ky such that L(Ky) c KT, LIK T is compact, IT
and rKT(L) = (~) . Pe rhaps most s u r p r i s i n g i s t h e f a c t t h a t t h e cone s p e c t r a l r a d i u s
can v a r y f o r d i f f e r e n t t o t a l c o n e s . B o n s a l l p r o v e s , however , t h a t i f L i s compact
as a map o f • i n t o i t s e l f , t h e n rK(L ) = r ( L ) f o r e v e r y t o t a l cone K such t h a t
L(K) c K. Our n e x t p r o p o s i t i o n i s a g e n e r a l i z a t i o n o f t h i s f a c t , and t h e a rgument we
g i v e i s a g e n e r a l i z a t i o n o f B o n s a l l ' s a rgument f o r t h e c a s e L compact .
324
( 2 .2) COROLLARY
(Compare [1].) Let X be a Banach space and L:X ~ X a bounded linear map.
Assume that p(L) < r(L) (p(L) and r(L) defined by (2.18) and (2.20)). Then if
K is any total cone in X such that L(K) c K and rK(L ) is defined by (2.23),
one has
In particular, if
r z ( L ) : r ( L ) .
r : r ( L ) , t h e r e e x i s t s x ( K-{O}
If X is the dual space of X
exists f ( ~ such that
such that
( 2 . 3 3 )
Lx = rx. ( 2 . 3 4 )
K ~ - and - {f E X : f(x) ~ 0 for all x ~ K}, there
L*(f) = rf. (2.35)
( 2.4 ) REMARK
For reproducing cones, the latter half of Corollary 2.2 was proved in [7] by
a linear argument like that used by Krein and Rutman.
Proof
Suppose we can prove that rK(L) = r(L). Since we clearly have PK(L) ~ p(L), it
will then follow from Theorem 2.2 that there exists x ( K satisfying (2.34). Sin~e
K is total, it is easy to see that K* is a cone and L*:K* ~ K*. It is proved in
[14] that p(L*) = p(L), so that pK,(L*) ! p(L). If we can prove that rK,(L*)~r=r(L),
the existence of f satisfying (2.35) is a consequence of Theorem 2.2 again. However,
if x satisfies (2.34), a version of the Hahn-Banach theorem implies that there exists
g ( K* with <g,x> > 0 (where < , > denotes the bilinear pairing between X* and
X) and one obtains
< ( L * ) n g , x > = <g , Lnx> = r n < g , x > . ( 2 . 3 6 )
Equation (2.36) easily implies that rK,(L* ) ~ r.
325
exists
Thus to complete the proof it suffices to prove (2.33).
x ( K such that
llnxH i im sup > 0.
n-~ IILnH
We claim that there
( 2 . 3 7 )
If (2.37) holds for some x, it is easy to see (using the fact that K is total)
that there exists u ( K such that
II Lnull lim sup > O. (2.38)
n-+~ llLnll
Inequality (2.38) implies that rK(L ) > r(L), while the opposite inequality is imme-
diate.
Thus it suffices to prove (2.37).
H Lnx rl lim sup
n-~ IILn[I
We argue by contradiction and assume
= O for all x ( K.
Select numbers Pl and P2 such that
p : p(L) < Pt < P2 < r : r ( L ) .
If B = {x: ]Ix]] ~ i} and a denotes the measure of noncompactness, select
that
~(Ln(B)) < pic~(B) _< 2Pi, n >_ N O .
We can also assume that N O
is so large that
2 o P 2 < r -
By definition of the measure of noncompactness, there exist sets
such that
L N~ (B) m No : U Si, diameter (Si) -< 2p i .
i:i
N O N O Select r with 0 < r < 2(p 2 -Pl ) and for each i, i <- i -< m, select
and N. such that 1
IITnxil] < ~l]Tnll , n >_ N . . 1
N such O
Si, i _< i -< m,
x. (S. 1 1
326
Finally, select an integer K ~ such that
N d~f k N ~ max N.. o o l!i!m i
I f y E B, there exists an integer i, 1 _< i _< m, such that
N N liT ~ <- 2P10.
If n>-N we obtain that
n+N N liT ~ <- rlTnxilr + IlTnlrH T Oy_xil I,
_< r + 2pNOlrTnll (2.39)
< 2p~ o IITnll
Inequality (2.39) implies that for n ~ N
N IITn+Noil ~ 2P2OllTn[I . (2.40)
Using (2.40) one sees that there exists a constant B independent of the integer
k ~ I such that
kN N k liT ~ -~ B(2P2 ~ (2.41)
and (2.41) implies that
1 1 kN (k~--)
l i m sup liT ~ o _< 2 ~ 2 < r . ( 2 . 4 2 ) k ~
Inequality (2.42) contradicts the definition of r. �9
Many cones used in analysis are normal. We would like to close this paper by
discussing a linear operator which is best studied in nonnormal cones. The operator
L which we shall define below plays an important role in recent work by K Bumby on
the problem of finding the Hausdorff dimension of certain sets of real numbers defined
by properties of their continued fraction expansions [4].
Let M denote a finite union of closed, bounded intervals of reals, let
Xn, n ~ 0, denote the Banaeh space of real-valued maps x:M ~ ~ which are n times
327
continuously differentiable and let
X n. If x (X n define IlXNn by
K denote the cone of nonnegative functions in n
n
IIxlJ n : X maxlx (j) ( t ) I . j:0 te~
(2.43)
There is a natural map
S is a bounded set in
Jn:Xn ~Xo by Jn(X) : x (n), the n th- derivative of x.
Xn, we leave to the reader the straightforward proof that
If
(S) = a(Jn(S)). (2.44)
We use the same letter
on the left and in X O
in (2.44) to denote the measure of noncompactness in
on the right.
X n
We now define a linear map L by
( L x ) ( t ) : N
bj(t)xCzj(t)). j=i
(2.45)
We shall always assume about the given functions b. and ~. in (2.45) ] ]
HI. The functions b.:M ~ ~ are C and nonnegative for i ~ j < N and -- ]
N
j~l= b4(t)~ > 0 f o r a l l t ( M . The f u n c t i o n s ~ . :M4N] a r e C , ~ j (N) c M and
max (max ~j ( t ) l) = c < 1. L~j~N t(M
Under assumption HI L defines a bounded linear map of X to X for n ~ 0 n n
and L(Kn) c Kn. In our next theorem we discuss the spectrum of L:• n ~Xn; for reasons
of length we shall not prove Theorem 2.3 here.
(2.3) THEOREM
(Compare [4]). Let L be defined by (2.45) and assume that H1 holds. Consider
L as a map from X n = cn(M) into itself and let Pn denote the essential spectral
radius of L in X n and r n = rn(L) the spectral radius of L in X n. Then if
N N
A = max ~ Ibj(t) I , B = min ~ b~(t) and c = max ~(t) l < 1, one has t~M j=l t(M j=l j,t
Pn < Acn ( 2.46 )
328
r _> B. (2.47) n
Inequality (2.46) holds even if the functions b. are not nonnegative. If Pn < rn 3
(as will be true for large n), r n = ro; and if Pn < rn and D is a total cone
in X n such that L(D) c D, then rD(L ) = r n. There exists a nonnegative, C = function
u, not identically zero, such that
Lu = r u. (2.48) O
( 2 .5 ) REMARK
If b. is strictly positive for 1 ~ j ~ N, Bumby [4] proves the existence 3
of a continuous function which is strictly positive on M and satisfies (2.48). How-
ever, his proof does not apply under the weaker assumption H1, and indeed we do not
prove here that u is positive on M under H1, although we conjecture that this is
so. The fact that r n = r ~ = rD(L ) when Pn < rn and the estimates on Pn appear
to be new; if the b. are strictly positive, we can prove that r = r for all n. 3 n o
( 2 .6 ) REMARK
If the hypotheses of Theorem 2.3 are weakened slightly, then r may fail to o
be a positive eigenvalue of L with corresponding eigenvector in K o, For example,
if (Lx)(t) = tx(~t) for x ~ C[0,1], one can verify directly that the spectral ra-
dius r ~ of L is 0; any C function with support in [~, ] is a C eigenfunc-
tion.
If (Lx)(t) = b(t)x(t) for x E C[0,1], where b(t) is not constant on any
subinterval of [0,1], one can prove directly that the spectral radius of L equals
max ]b(t) I , but that L has no point spectrum. t
2.7 ) REMARX
The spectrum of L varies depending on what space L acts in. For example
if (Lx)(t) = x(ct), 0 < c < i, one can prove that the spectrum of L as a map of
cn[0,1] into itself is {cJ: 0 ~ j ~ n-l} U {z: z complex, Izl ~ cn}; cJ(0 ~ j ~ n-l)
t j Further- has algebraic multiplicity one and corresponds to the eigenvector xj(t) = . n
more if Izl < c , z is in the point spectrum of L:C n § C n and has infinite multi-
plicity; if 0 ~ z < c n, there are infinitely many nonnegative, linearly independent
eigenvectors corresponding to z.
829
REFERENCES
[1]
[2]
[3]
[4]
[5]
[s]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
BONSALL, F.F.: Linear operators in complete positive cones, Proc. London Math.
Soc. 8 (1958), 53-75.
BONSALL, F.F.: Positive operators compact in an auxiliary topology, Pacific J.
Math. 10 (1960), 1131-1138.
BONSALL, F.F.: Lectures on Some Fixed Point Theorems of Functional Analysis,
Tata Institute of Fundamental Research, Bombay, 1962.
BUMBY, R.T.: Hausdorff dimension of Cantor sets, to appear.
DARB0, G.: Punti uniti in transformazioni a condiminio non compatto, Rend.
Sem. Mat. Univ. Padova 24 (1955), 353-367.
DUGUNDJI, J.: An extension of Tietze's theorem, Pacific J. Math. 1 (1951),
353-367.
EDMUNDS, D.A. and POTTER, A.J.B.and STUART, C.A: Noncompact positive operators,
Proc. Royal Soc. London A328 (1972), 67-81.
v KRASNOSEL'SKII, M.A.: Positive Solutions of Operator Equations, p. Noordhoff
Ltd., Groningen, The Netherlands, 1964.
KREIN, M.G. and RUTMAN, M.A: Linear operators leaving invariant a cone in a
Banach space (Russian), Uspehi Mat. Nauk 3, No. 1, 23 (1948), 3-95, English
translation, A.M.S. translation 26.
KURATOWSKI, C.: Sur les espaces complets, Fund. Math. 15 (1930), 301-308.
LERAY, J.: Th~orie des points fixes: indice total et nombre de Lefschetz,
Bull. Soc. Math. France, 87 (1959), 221-233.
LERAY, J. and SCHAUDER, J.: Topologic et 6quations fonctionnelles, Ann. Sci.
[cole Norm. Sup. 51 (1934), 45-78.
MASSABO, I. and STUART, C.A: Positive eigenvectors of k-set-contractions,
Nonlinear Analysis, T.M.A, 3 (1979), 35-44.
330
[14] NUSSBAUM, R.D.: The radius of the essential spectrum, Duke Math. J. 37 (1970),
473-478.
[15] NUSSBAUM, R.D.: Some asymptotic fixed point theorems, Trans. Amer. Math. Soc.
171 (1972), 349-375.
[16] NUSSBAUM, R.D.: The fixed point index for local condensing maps, Ann. Mat.
Pura. Appl. 89 (1971), 217-258.
[17] NUSSBAUM, R.D.: Generalizing the fixed point index, Math, Ann. 228 (1977),
259-278.
[18] NUSSBAUM, R.D.: Integral equations from the theory of epidemics, p. 235-255
in Nonlinear systems and applications, edited by V. Laksmikantham, Academic
Press, New York, (1977).
[ 1 9 ] NUSSBAUM, R.D.: Periodic solutions of some nonlinear integral equations,
p. 221-249, in Dynamical systems, Proceedings of a University of Florida Inter-
national Symposium, edited by A.R. Bednarek and L. Cesari, Academic Press,
New York, (1977).
[ 2 0 ] RABINOWITZ, P.: Th6orie du degr6 topologique et applications ~ des probl~mes
aux limites non lin6aires, notes by H. Berestycki from a course at Universit6
Paris VI, Spring, (1973).
[21] SCHAEFER, H,H.: On nonlinear positive operators, Pacific J. Math. 9 (1959),
847-860.
[22] SCHAEFER, H.H.: Some spectral properties of positive linear operators,
Pacific J. Math. 10 (1960), 1009-1019.