TRANSLATION-INVARIANT CONES OF FUNCTIONS

ON SEMI-SIMPLE LIE GROUPS

HARRY FURSTENBERG 1

Introduction. Originally, the phrase harmonic analysis had a

function-theoretic meaning, referring to the decomposition of a func-
tion into exponentials. In the current interpretation, particularly in
connection with noncommutative groups, the term refers not to
functions but to representations, and harmonic analysis is regarded
as part of the theory of group representations. This shift in inter-
pretation was motivated by the L 2 -theory for compact groups. The
decomposition of the L 2 -space of a noncommutative compact group
analogous to the Fourier decomposition for the circle involves multi-
dimensional subspaces, and, as a result, there is no longer a canonical
choice of a basis for the L 2 -space analogous to the set {ein6} for the
circle. The subspaces, on the other hand, are canonically determined,
and correspond to the various irreducible representations of the
group. It therefore became natural to regard irreducible representa-
tions as the basic building blocks of the theory in the place of the
exponential functions.
Our purpose here is to call attention to some examples in the theory
of semi-simple noncompact Lie groups where the classical setup pre-
vails. T h a t is, we shall find a class of functions on these groups which
appear to play a role similar to that played by the exponentials for
the circle or the real line. In terms of these functions, a form of
spectral synthesis will be valid. Namely, for certain translation-
invariant classes of functions on the group, we shall find that each
function of such a class admits a unique representation as a general-
ized linear combination of the "exponentials" in t h a t class. Admit-
tedly, this result corresponds to a relatively easy case of spectral
synthesis for the line. However, it is hoped that by pursuing this
analogy further, other fruitful applications may be found.
The prototype of the theorem we shall prove is a theorem of
Choquet and Deny [3 ] for Rn (or any locally compact commutative
group). Let /z be a positive bounded Borel measure on Rn that does

An address delivered before the New York meeting of the Society on April 21,
1964 under the title Aspects of harmonic analysis on Lie groups, by invitation of the
Committee to Select Hour Speakers for Eastern Sectional Meetings; received by the
editors September 20, 1964.
1
Work supported in part by Army Research Office, Project in Probability, Prince-
ton University and the Air Force Office of Scientific Research.
271
272 HARRY FURSTENBERG [March

not have its support in a proper subgroup of Rn. Form the cone V
of (unbounded) positive measures v on Rn satisfying fi * v = \vt where
X > 0 is fixed. A general result of Choquet and others to which we
shall refer more explicitly later ensures that V is spanned by its ex-
tremal rays. Moreover, choosing appropriately a single element from
each extremal ray, there will be a unique integral representation of
each member of V in terms of these extremals. The question arises
as to the nature of the extremal measures of V. The answer given by
Choquet and Deny is that the extremal measures have the form
f(x)dx where ƒ(x) is proportional to an exponential: ƒ(x) =ceu'x. In
other words, the measures of V all correspond to functions (in fact,
convex functions), and these functions are linear combinations of the
exponentials in V: -F£ Vt=>F(x) — feu'xdo>(u) where co is a measure
concentrated in the set of u for which Jeu'x dfj,(x) =X.
It is worthwhile sketching a proof of the fact that the extremals of
V correspond to exponentials. The cone V is invariant under transla-
tion. Now the equation v = \~~1fi, * v implies that v is a linear combina-
tion of its translates. This means that v cannot be an extremal of V
unless all its translates by points in the support of /x are proportional
to v. If the support of fx generates Rn as a group, this implies that all
translates of v are proportional to v. One sees easily that v must then
be given by an exponential.
For a noncommutative group the convolution p * v may still be
defined and we can form the cone V of solutions to fi * v—\v as be-
fore. This time, however, V is invariant under translation on the
right : v—>>v * g (this will presently be defined more precisely) and the
equation v=\~1ix*v implies that v is a linear combination of left
translates. The above argument, therefore, breaks down, and one can
inquire what class of functions can occur as the extremals of V in
the general case. For a compact group G this question is easily
answered, because V will be empty unless X = /x(G) and v must be
proportional to the Haar measure of G. But for noncompact groups
the question is more interesting and for the most part unsolved. We
shall solve this problem only for semi-simple Lie groups, and even
here only under additional hypotheses regarding /z.

1. Preliminaries. If G is a topological group, by a G-space is meant

a space X and a continuous map (g, x)—>gx of GXX-+X satisfying
(&i&2)# = gi(g2#) and ex — x for e the identity of G. X is a homogeneous
G-space if G is transitive on X; that is for x, y (EX, there exists gÇzG
with gx = y. X may then be identified with the coset space G/H,
where, for some tfoG-X', H is the subgroup that leaves #0 fixed.
 1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 273

If X is a G-space, fx a Borel measure on G and v a Borel measure

on Xy then the convolution /x * v may be defined by

I f(y) dfx * v(y) = I I ƒ($*) dfx(g)di>(x),

J X J XJQ

providing this integral is finite for continuous functions ƒ of compact

support. In particular, the convolution of two bounded measures on
G is defined. We shall interchangeably denote the unit mass concen-
trated at a point x of a space X by 5* or by x itself. The convolution
b0 * ix for g G G and /x a measure on a G-space (or on G itself) will some-
times be denoted g/x; similarly ixx is the same as fx * öx where /x is a
measure on G and x belongs to the G-space X. Notice that 80 * 8X = ôgXl
so that our notation for convolution is consistent with the notation
for multiplication. It is frequently useful to regard convolution as a
linear extension of multiplication so that

jx*v= I gvdfi(g) = I iixdv(x) = I I gxdp(g)dv(x).

J Q J x J QJX

The space of unbounded regular measures on a locally compact

space may be identified with the dual of the space of continuous func-
tions with compact support on the space. When we speak of the weak
topology on measures, we have in mind the weak* topology of this
dual space. In addition, it will often be convenient to denote an
integral Jf(x) djx(x) simply as MCO-
On a locally compact group G a locally integrable function f(g)
determines a measure co on G by dœ(g) =f(g)dg, dg denoting the left
invariant Haar measure on G. We shall often find it convenient to
identify the function ƒ with the measure co. Thus the convolution of
two functions f\ */ 2 , or of a function with a measure, ƒ * \x or \x * ƒ,
is defined as the function corresponding to the measure which is the
convolution of the two measures involved.
On a group G we use the notation Ry to denote the operator Ryf(g)
—f(gy). It might be noticed that, as a measure, Ryf corresponds to
the convolution ƒ * ôy-i. In our discussion we shall be concerned pri-
marily with translates on the right : ƒ—>Ryf. As a result we shall al-
ways understand, by the translate of a function ƒ, a right translate
Ryf-
If X and Y are two G-spaces and j is a map of X into F, we shall
say that j is equivariant if j(gx) ~gj(x), for all xÇzX, gÇîG.
By a semi-simple group G we shall mean one which is noncompact
and has a finite center. Such a group admits an Iwasawa decompose
274 HARRY FURSTENBERG [March

lion G — K-A'N. Here K is any maximal compact subgroup of G,

A is a vector group and N is nilpotent, A • N is a solvable group whose
commutator subgroup is N. The decomposition is not direct but each
g £ G admits a unique representation g = kan. The letter K will al-
ways refer to a fixed but arbitrary maximal compact subgroup of G.
triK denotes the normalized Haar measure on K. In general, if ju is a
positive measure on a space X and jx(X) = 1, we refer to ju as a proba-
bility measure.

2. Multipliers and multiplier functions. To generalize the notion

of an exponential function to an arbitrary group we may proceed
as follows. The exponentials are characterized by the property t h a t
they have unit value at the identity and any translate is proportional
to the function itself. On a semi-simple group only the constant func-
tion has this property. (Such a function defines a homomorphism
from the group to the complex numbers and the homomorphic image
of a semi-simple group is semi-simple.) So we modify this condition
by supposing instead that we have a family of functions s(-, £),
where £ ranges over an index set X, with the property that s(e, £) = 1
and that a (right) translate of each of the s(-, £) is proportional to
some other of the s(-, £). We may take X to be the family of func-
tions occurring here with the identification of proportional functions.
M is then a G-space, and we may write Rys(-, Q~s(-, y£). More
precisely, we have

Rys(g, Q = s(gy, Ö = hi(y)s(g, yQ.

Setting g = e, we have s(y, £) =h(y) or

(2.1) s(gy9Q =s(g,yQs(y,Q.

A function satisfying (2.1) is usually referred to as a multiplier.

As a function from G to the module of functions on M it is referred to
as a crossed homomorphism or a 1-cocycle. Our usage will be more
restrictive and we take the following as our definition.
D E F I N I T I O N 2.1. Let X be a G-space. A continuous function 5 from
GXX to the positive reals is called a multiplier if it satisfies (2.1).
The group of all such functions is denoted M(X) (where the group G
is assumed fixed). If s£M(X) and £ £ X then the function s(-, £)
is called a multiplier function. The set of all multiplier functions cor-
responding to multipliers in M(X) is denoted E(X).
If X consists of a single point, then (2.1) defines a positive char-
acter (homomorphism into (0, 00)) on G. Thus the positive exponen-
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 275

tials are the multiplier functions on Rn corresponding to a trivial X.

As we have remarked, for G semi-simple, if X is trivial, then E(X)
consists of the function 1. On the other hand, if X is taken too large,
E(X) is again uninteresting because it contains all positive continu-
ous f unctions ƒ with f(e) = 1. For example, if we take X = G, and if ƒ
satisfies f(e) = 1, define s^M{G) by s(gu g2) =/(gig2)//(g2). It is easily
seen that s satisfies (2.1) and/(g) = s(g, e). What we shall find is that
in the case of a semi-simple G, there is a choice of X for which the
functions of E{X) play the role of the exponentials. In particular,
these functions will turn out to be the extremals in our version of the
Choquet-Deny theorem.
If £(£) is a positive continuous functions on the G-space X, we
may form a multiplier in M(X) by setting

(2.2) *(g,Q = P(g&/p®.

We call these multipliers trivial. (In the terminology of cohomology

theory, these would be coboundaries.) We shall call two multipliers
equivalent if their quotient is trivial. The (cohomology) group of
equivalence classes will be denoted H(X).
Similar considerations arise in the representation theory of G. If
X is a G-space we may define a representation of G on some linear
space of functions on X by setting Ugf(£) =s(g~ 1 , ö/ÜT"1©* The con-
dition that U0 be a representation is just the multiplier equation
(2.1) for s. Moreover, if Ug and Vg arise from equivalent multipliers,
then Ug and Vg are equivalent representations. It will not be surpris-
ing that the space X that we will study has also been studied in con-
nection with the irreducible representations for semi-simple Lie
groups [ l ] .
Let G be semi-simple and K a maximal compact subgroup, and let
X be a G-space.
D E F I N I T I O N 2.2. a is a K-multiplier if aGM(X) and a(k, £) = 1 for
all kÇzK, £ £ X . The set of ^-multipliers forms a group MK(X), and
the corresponding set of multiplier functions is denoted EK(X).
LEMMA 2.1. Let X be a homogeneous G-space and suppose that the
subgroup K is already transitive on X. Then the natural map of MK(X)
into H(X) is an isomorphism onto.
PROOF. T O show that this map is 1-1 suppose <TÇZMK{X) and is
trivial: a(g£), =p(gO/P(0- Then p{kÇ) =p(Ç) and since K is transitive,
p is constant so that <7 = 1. To show the map is onto we associate to
every multiplier an equivalent üT-multiplier. For sÇï.M(X), set
276 HARRY FURSTENBERG [March

ƒ s(kg, Ö dk ƒ s(k, gQ dk
(2.3) *(g, Q = — = — s(g, Ö,
r *(*, © ^ r s(k, & dk
JK JK

from which we see that a~s and cr(k, 0 = 1.

From this we can see that, in many cases, although the group M(X)
is infinite-dimensional, the group H(X) is finite-dimensional. For if
we let X = G/L, where L is a subgroup of G, then K is transitive on X
if and only if KL = G. If GCHMK and ^CLG/L corresponds to the coset
L, then (r(g, £) =c-(g, &£0) = <r(gk, £o) for some kÇ.K. So (T is determined
by the function cr(g, £0). Since g = kfl for some fe'£^, /G-^, <r is deter-
mined by the restriction of a-(*, £0) to L. But this is a positive char-
acter by (2.1), since /£o = £o- In particular, if L has finitely many con-
nected components then its (positive) character group is a subgroup
of a vector group. It follows that MK(X) and, by Lemma 2.1, also
H(X), is isomorphic to a subgroup of a vector group.
In certain cases we can assert that H(X) is trivial.
LEMMA 2.2. If L is compact then H(G/L) = 1.
PROOF. Let sÇzM(G/L) and let %o£:G/L correspond to the coset L.
Then s(lt £0) is a positive character on L. But, since L is compact,
s(h £o) = l. It follows that s(gl, %o)=s(g, /£oMZ, £o)=s(g, £0), so that
s(g> £o) = p(gèo) for some positive continuous function p on G/L. Hence
s(g, g%)=s(gg', | 0 ) A ( g ' , W - p C g g W ^ É o ) or s(g, p=P(gÇ)/p(ï).
As a final remark we point out that if j is an equivariant map of one
G-space X into another G-space F, then there is induced a map
M(Y)-*M(X). Namely we assign to each sÇ.M(Y) the multiplier s*
defined by s*(g, x)=s(g, j(x)). s* satisfies (2.1) if 5 does. If K is
transitive on X, then it is possible to show that the map of M(Y)
-*M(X) induces an isomorphism of H(Y) into H(X).
3. Examples. In this section we shall recall some familiar integral
representation formulae which we shall see express a function as a
linear combination of multiplier functions. In each case the function
being represented satisfies a convolution equation or a family of such
equations, and the formula will turn out to be an instance of the
generalized Choquet-Deny representation.
Let G be the connected group of rigid motions of the plane <R2. G is
the semidirect product K-R2 where K is the subgroup of notations
and R2 the subgroup of pure translations. R2 is a normal subgroup of
G and G/R2 may be identified with K. We shall compute MK(G/R2).
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 277

(Note that by Lemma 2.2, H(G/K) is trivial, so that all multipliers

of M(G/K) are trivial.) Identifying R2 with the complex plane, we
may denote an element g £ G by the pair (eid, w) if g(z) = ei6z+w. The
action of G on K = G/R2 is given by (eie, w)Ç = ei6Ç where |f | = 1 . We
then find that if (rGM^(G/i? 2 ),

a(e«, w; f) = a(l, e-t'r1™; 1) = ^ ^ r 1 » ) ,

where A denotes a homomorphism of C-^R. This means A(2) = (Rsr
for some complex r. The general form of a multiplier in MK(G/R2)
is therefore
(3.1) *r(eid, w\ r) » « « • - ' • r W .

Consider now the zero-order Bessel function of an imaginary argu-

ment: ƒ(r) = Jo(*V). The function on i? 2 given by F(z) =f(\z\) is an
eigenfunction of the Laplacian A = Dl+D% on the plane. This implies
that it is an eigenfunction of the mean-value operators:
1 r2r
TpF(z) = — I F(z + peie) dB.

We now lift the function F(z) to the group G by setting <&(e**, w)

= F(w). If we write the equations TPF=\PF in terms of $ we find
that $ satisfies a family of convolution equations on the group. Ac-
cording to the theorem to be proven it follows from this that <£ can
be expressed as an integral over a set of multiplier functions in
E(G/R2). Moreover, since $(kg)=$(g) for fe£i£, these multiplier
functions will be in EK(G/R2)1 and so are given by (3.1). This is in-
deed the case, for by the Bessel formula

Sir) = — I e~' °°' « du = I trto'^dmijt),

2T J0 J

where dm(£) refers to normalized Lebesgue measure on the unit cir*

cle. It follows that

F(«)- J V f l U r dm®,

"3?(e",w)= f g-^f1 dmQ) = f e-®-™-*"1'1 dm(Ç)

= ƒ <ri(e« »;*•)<&»(*•).
278 HARRY FURSTENBERG [March

Bessel's formula is a special case of a formula for spherical functions

on any symmetric space of noncompact type. (See [8, Chapter X ]
for details.) Let us consider the case of a space of constant negative
curvature. The space may then be taken as G/K where G is semi-
simple. A spherical function F(x) on G/K is defined by the conditions:
F(kx) = F(x) for all feG-K, F(x0) = 1 where x0 corresponds to the coset
K, and F is an eigenfunction of all differential operators on G/K
which commute with the action of G on G/K. If F(x) is lifted to G
we obtain a function (f>(g) = F(gK) which, as may be shown, is char-
acterized by the equation:

(3.2) f <Kgikg2) dk = <t>(giMg2), gh g2 E G.

JK

We notice that 4> satisfies a family of convolution equations. In

fact, (3.2) may be rewritten (see §1) as

(3.3) m K * g r 1 * <t> = <t>(gi)4>-

The integral representation formula for these functions is due to
Harish-Chandra [8, Chapter X, Theorem 6.16]. We define a function
H from G to A, where A is the vector part of the Iwasawa decom-
position of G, by H(kan)=a. Then, for every spherical function </>
on G, there is a homomorphism A: A—>C with

(3.4) <t>(g) = fe^MVdk.

JK
When 0 is real and positive, À is real-valued. We shall show that in
this case, the integrand of (3.4) represents a multiplier in MK{G/AN)1
where KA N is the Iwasawa decomposition of G. Namely, set <r(g,kA N)
= eHH(0k))m since H(k') is the identity of A if fe'G-K, it follows t h a t
<r(k', kAN) = l. We next check the multiplier equation. Suppose
£ = kAN for kGK, and gu guCzG with
g2k = &2ÖW2, gl&2 == klCilti.

Then gig2k = gik2a2n2 = kiaifiia2n2 — kiaia2n[n2 since N is normal in AN.

Then
a(g!g2, £) = e A(H(fl, i'2*)) = eM^HM**),

<r(gh £20 = a(gi, k2AN) = eAore**,)) = eA(al)>

<r(g2, Ö = <r(g2) kAN) = e <*to*)) = e A W,

A

so that <r(gig2, Ç)=<r(gu g2^)(r(g2, £). (3.4) may therefore be rewritten

1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 279

(3.5) <t>(g) = f a(g, h AN) dk = f <r(g, £) d m ® .

•J X J Q IAN

Thus the Harish-Chandra formula in the case of positive spherical

functions is a special case of the generalized Choquet-Deny repre-
sentation. I t should be pointed out that our methods fail in the case
of complex-valued spherical functions although the formula (3.4) is
still valid.
Another example of a spectral decomposition for functions on a
semi-simple group is given b y the Poisson formula for harmonic
functions in the unit disc. Let G denote the connected group of con-
formal maps of the unit disc D= {z: \z\ < l } onto itself. A function
h(z) on the disc determines a function/(g) =A(g -1 (0)) on G. Let K
again denote the group of rotations about the origin. The mean-value
property for harmonic functions says that h(0) =fKh(kz) dk for any
zÇzD. If A is harmonic, so is the function obtained by composing h
with the transformation g"""1, g (E G. This gives

(3.6) Krx(o)) = f Kr^dk.

If we write 2 = 7""1(0), (3.6) becomes

(3.7) f(g) = ff(ykg)dk, g,yEG.

J K

Thus the functions/(g) which correspond to harmonic functions on D

satisfy a family of convolution equations:

(3.8) tnK*V~l *ƒ = ƒ.

Assume that h ^ 0. The Poisson formula gives

ft(*) = f P(«, 0) dv(0),

where P denotes the Poisson kernel, and v is a positive measure. Now
one may verify t h a t if z = g~1(0) then

P(*,0) = W(e«)\
or

(3.9) /(g)- J
f\fW\iv(ff).
0
280 HARRY FURSTENBERG [March

We recognize by the formula for differentiating a composite function

that the integrand in (3.9) is a multiplier in MK(9D) where 3D is
the unit circle.
The Poisson formula for harmonic functions also generalizes to
arbitrary noncompact symmetric spaces. Moreover, a similar repre-
sentation is valid for any positive eigenfunction of the Laplace-Bel-
trami operator on a symmetric space. This was shown by Dynkin [5]
for the symmetric space SL(w, C)/SXJ(n). Once again it may be shown
that the extremal solutions occurring in Dynkin's representation are
multiplier functions for a certain compact G-space which we will call
B(G) and regarding which we shall presently be more specific.
Another generalization of the Poisson formula for harmonic func-
tions was given in [6]. For our present purposes it will be convenient
to deviate somewhat from the notation used in [ó]. Let /* be a proba-
bility measure on G which is absolutely continuous and whose sup-
port contains a neighborhood of the identity of G. Consider the
bounded solutions to

(3.10) f(g) = f' f{g'g)dn(g').

This is a generalization of (3.7) which was satisfied by ordinary har-

monic functions. It is shown that, given /x, there is determined a
probability measure v on the G-space B{G) referred to in the preced-
ing paragraph, with the property that every bounded solution of
(3.10) is given by

(3.11) f(g) = f f(r^)dv(x),

where ƒ is a bounded measurable function on B(G). The measure v

satisfies fx*v = v and from this one can infer that the transforms gv
are absolutely continuous with respect to z>, and that (dgv/dv)(x) is
positive and continuous on GXB(G). (3.11) may therefore be written

(3.12) f(g) = f ^(X)f(x)dv(x).

We claim that s(g, x) — (dg"~1v/dv)(x) is a multiplier in M(B(G)). By

the definition of the Radon-Nikodym derivative it is easily seen that
dsrhi dv\
(3.13) JL_(*) = _ (gx).
ag lvt dvi
So
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 281

dgïlgïlv dg2lgïlv dgtlv

{x) = (x) (x)
dv dg2 lv dv
= —:— (gtx) —— W,
dv dv
which is the multiplier equation for s. Thus the measure /x on G
determines a multiplier sÇzM(B(G)) and the bounded solutions of
(3.10) are integrals over multiplier functions corresponding to s:

Kg) = I s(g, x)f(x) dv(x).

JBÇQ)

We note that this is the first example in which a multiplier is used

which is not a i£-multiplier. Naturally, a function cannot be repre-
sented in terms of X-multiplier functions unless it satisfies f (kg)
=/(g) for kEK.
It should be mentioned that in case the support of ix does not con-
tain a neighborhood of the identity a similar representation is valid,
but the relevant space may not be B(G) but rather one of finitely
many covering spaces of B(G).
The space B(G) which occurs in these formulas may be defined as
follows. If G = KAN is an Iwasawa decomposition of G we let T de-
note the normalizer of AN, that is the set of all g £ G with gANg~~l
CAN. T is also obtained by adjoining to AN the centralizer K0 of A
in K: T = K0AN. B(G) is then defined as G/T. T is not uniquely de-
termined, but it is unique up to conjugacy. As a result, the various
versions of G/T are all isomorphic as G-spaces. An alternative defini-
tion of B(G) occurs in [6]. We define a boundary of a Lie group G
to be a compact G-space with the property that for any probability
measure w on the space, some sequence gn7T, gndzG, converges to a
point measure. All boundaries of G are equivariant images of one of
them, and this one is B(G). The equivalence of the two definitions
was proven in [l0].
The space B(G) is an equivariant image of the G-space G/AN that
occurs in (3.5). It may be shown that, in fact, the multiplier cr in
formula (3.5) which belongs to MK(G/AN) actually is the image of a
multiplier in MR(B(G)) (see end of §2) so that the integral in (3.5)
may also be taken over B(G). Thus in each of the cases studied, the
integral representation involves multipliers in M(B(G)).
4. Tychonoff groups. The significance of the subgroup T may be
attributed to the fact that it possesses a certain "fixed-point" prop-
erty which we will elucidate here. We consider a subcone F of a linear
282 HARRY FURSTENBERG [March

topological space S. We say that V has a compact base if there exists

a continuous linear functional L on 5 such that L(x)>0 for x&V
except for x = 0, and such that Vx= { x £ V: L(x) = 1} is compact. If
r is a continuous linear transformation of S taking V into itself then
r has a fixed (invariant) ray in V. For, by the Schauder-Tychonoff
fixed-point theorem [4], the transformation
TX

L(TX)

which is continuous and takes Vx-^Vi must have a fixed point in V\.
The ray through this fixed point is clearly left fixed by r. This result
may be extended to a one-parameter group of transformations r{t)
of 5 which take V into itself, and there will exist a ray in V invariant
under the entire group.
D E F I N I T I O N 4.1. A Lie group G is a Tychonoff group if, whenever G
acts continuously by linear transformations on a locally convex,
linear, topological space 5, taking a cone V with compact base into
itself, then G has a fixed ray in V.
A Tychonoff group also has the following fixed-point property: If
it acts by affine transformations on a compact convex set, it has a
fixed point in the set. (An affine transformation is one that preserves
the convex structure: r(ax + (1 —a)y) =arx+(l —a)Tyt 0 ^a S1.) The
converse, however, is not true; a group may have the latter fixed-
point property without being a Tychonoff group. For example, every
solvable group has the fixed-point property relative to compact convex
sets. (See [6].) On the other hand, the group of matrices

{(ô y c ;)-•*«.•><>}
is solvable, maps the first quadrant of R2 into itself, but has no fixed
ray. An example of a connected solvable group which is not a Tych-
onoff group is the connected group of rigid motions of the plane. It
will develop t h a t if this group were a Tychonoff group, there would
be no nonconstant positive eigenfunctions of the Laplacian in the
plane.
The following lemmas will be useful in establishing that certain
groups are Tychonoff groups.
LEMMA 4.1. Let A be a linear transformation on Rr with strictly posi-
tive eigenvalues. If f or some vector xt the set of vectors {Anx: — & <n < °o }
is boundedy then Ax = x.
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 283

PROOF. Since A satisfies its characteristic equation, it follows that

all the vectors Anx are contained in the subspace H spanned by
{x, Ax f • • - , Ar~1x\. Clearly each vector yÇzH has the property
that the set \Any} is bounded. So the restriction of A to H is an oper-
ator with all its powers, positive and negative, bounded. Therefore
all its eigenvalues are on the unit circle. But they must be positive
and so the restriction of A to H has only the eigenvalue 1. We may
therefore express this restriction as I+N, where I is the identity on
H and N is nilpotent on H. Let i be the least integer with N{ = 0. If
i ^ 2 , form (I+N)nNi-2 = Ni~2+nNi-1. This must be bounded as
i 1
n—>oo which implies N ~ = 0 and this is a contradiction. It follows
that N = 0, or the restriction of A to H is the identity. Hence Ax — x.
LEMMA 4.2. Let G, G\ be connected Lie groups with G\ a normal sub-
group of G and let ®, ®i denote their respective Lie algebras. Suppose
(a) G\ is a Tychonoff group,
(b) G/Gi is 1-ditnensional,
(c) the adjoint representation of ® restricted to ®i has matrices with
only real eigenvalues. Then G is a Tychonoff group.
PROOF. We suppose that G operates on a linear space S, that F is a
cone in 5 with compact base, and that L is a continuous linear func-
tional on 5, positive on V, and such that {x: L{x) = 1} intersects V
in a compact set. Since Gi is a Tychonoff group, G\ leaves fixed at
least one ray of V. If x is a vector on such a ray there is determined
a positive character X on G\ defined by gix=\(gi)x. Also, given a char-
acter X, there will exist a cone (possibly empty) V\ C V of vectors
y G V for which giy = \(gi)y. For gGG, g~lGigC.Gi and we may define
the transform \° of the character X by \°(gi) =XQf 1 gig). If x G V\,
(4.1) X'(«i)g* = Mr^igïg* = gg~lgig% = gig*,
from which it follows that gV\C V\f. Another consequence of (4.1) is
that if V\ is nonempty, then \°(gi) is bounded as a function of g. For
we have
*9(gùL(gx) = L(gig*)>
X*(gi) = L(glgx)/L(gx)}
and L(giy)/L(y) ^ m a x L(giz), where z ranges over the compact set of
vectors in V for which L(z) = 1.
If G{ denotes the commutator subgroup of Gi, then the group of
positive characters on G\ may be identified with the group of positive
characters on Gi/Gl. Since the latter group is abelian, it is the image
under the exponential map of its Lie algebra ®i/[®i, ®i]. Thus the
284 HARRY FURSTENBERG [March

positive character group of G\ may be identified with a subspace W

of the dual space (®i/[®i, ®i])*. Now it may be checked that the
automorphism X—»X* of the positive character group of Gi corresponds
to the operation X—>Ad g(X) on ®x. Since ad ® restricted to ®i has
all real eigenvalues, its image under the exponential map, namely
Ad G, has all its eigenvalues positive. This remains true when we pass
to the quotient space ®i/[®i, ®i] of ®i, and to its dual, and to the
subspace W of the dual.
We see then that the set of positive characters on G\ forms a linear
space on which we have a representation of G given by X—>Xa, and the
eigenvalues of the transformations of this representation are all posi-
tive.
We know that for some Xo, V\0 is nonempty. This means by our
previous observation that X{J(gi) is bounded as a function of gÇzG.
But since the map X—^X° has positive eigenvalues, we see by Lemma
4.1 that XQ =Xo. If we also recall that g V\0 C V^, we find that G takes
V\0 into itself. We shall now show that some ray of V\0 is left fixed
by all of G.
Let Fxo denote the set of vectors x £ V\0 with L(x) = 1. If for each
g G G we define g* by
g*(x) = gx/L(gx)
then each g* takes V^ into itself, g—>g* is a homomorphism of G into
a group of continuous transformations of V^. However, G* leaves
V{0 pointwise fixed since each g\X is proportional to x for # G K 0 . Thus
the action of G* on V^ is that of a one-parameter group. As a result
G* has a fixed point and this means that G has a fixed ray in Fx0. This
completes the proof of the lemma.
COROLLARY. A connected subgroup of the group of upper triangular
real rXr matrices is a Tychonoff group.
PROOF. Such a group is solvable and we can find a chain G = GoDGi
DG2 • • • "DGn = e with Gi/Gi+i 1-dimensional and G,-+i normal in
Gi. Let Q denote the group of upper triangular matrices with positive
entries on the diagonal. Since G is connected, GQQ. If O is the Lie
algebra of Ç, then O consists of real triangular matrices. For g G O ,
the eigenvalues of ad q are differences of diagonal entries of q, and so
they are real. This will be true for any subalgebra of O and so by
Lemma 4.2 we may proceed inductively to prove that G is a Tych-
onoff group.
COROLLARY. If G is a semi-simple Lie group with an Iwasawa de-
composition G = KAN, then AN is a Tychonoff group.
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 285

PROOF. According to [ i l , Exposé 11, p. 14], AN may be repre-

sented (faithfully) by upper triangular matrices.
THEOREM 4.1. If G is a semi-simple Lie group, the subgroup T de-
fined at the end of §3 is a Tychonoff group.
PROOF. We recall that T = K0AN where K0 is the centralizer of A
in K. Now since N is the commutator subgroup of AN, every positive
character X on A N is identically 1 on N, and is determined by its values
on A. If feoE-Ko, then \k°(a)l=X(fe^"1afe0) =X(a) for aE.A, so that
X*° = X. Hence k0V\C.V^o= V\, where Vx denotes the subset of x in
a cone V (with compact base) with gx = \(g)x for gÇiAN. Since AN
is a Tychonoff group, V\ will be nonempty for some X. Let x 0 G V\
and form

x = I &o#o dko.

Since i 0 ^ x C ^ , # E V\, and so AN leaves fixed the ray to which x

belongs. But clearly K0 leaves x fixed so that the ray to which x be-
longs is left fixed by all of T.
By [6, Theorem 1.10] it follows that the subgroup T is maximal
with respect to the property of being a Tychonoff subgroup of G.
5. Spherical and semi-spherical functions. Before obtaining our
generalization of the Choquet-Deny theorem, we shall discuss a spe-
cial case connected with a class of functions which we shall call semi-
spherical. Throughout this section, G is a semi-simple group, K a
fixed maximal compact subgroup, T the Tychonoff subgroup defined
in §3 and B(G) = G/T. If Y £ G , let Ay denote the operator

(5.1) Ayf(g) = f f(ykg) dk = tnK* 7" 1 */(g).

J
K

D E F I N I T I O N 5.1. A positive function ƒ on G is semi-spherical if it

is an eigenfunction of each Ay, 7 G G .
The reason for our terminology is that a spherical function fulfills
this condition, as we see from equation (3.2). But whereas a spherical
function is bi-invariant under K:<f>(kigk2)=<j>(g), a semi-spherical
function need only satisfy ƒ (kg) =/(g). (In fact, it is not difficult to
show that a semi-spherical function which is bi-invariant under K is
actually spherical.)
If ƒ is semi-spherical, there is defined a function $(7) on G by
Ayf=<j>(y)f. We notice that ^(7) is spherical:
286 HARRY FURSTENBERG [March

f f f(gihg#gù dkdk' = f ABlf(g2k'g3)dk'

(5.2)
= 4>(gùA.J(gi) = *(ft)*(fc)/(fi).
But the first integral of (5.2) is also

ƒ ^W(ft) dk = ( j <Kgi^) ^ ) /(g3)

and so

J <Kgi£g2) di = <Kgi)<Kg2),

which is the defining equation of a spherical function.

DEFINITION 5.2. We say a semi-spherical function ƒ belongs to the
spherical function <p if

(5.3) ( AyH) dk = 4>(i)f{g);

JK

the cone of semi-spherical functions belonging to <f> will be denoted V^.

V^ is never empty since <££ V^. Notice also that V<j, is closed under
right translation, RyV^QV^y since the translates Ry commute with
Ay*.
LEMMA 5.1. If<T(ELMK(X) where X is a G-space such that K is transi-
tive on X, then the multiplier functions <r( •, x) are semi-spherical and
belong to the spherical function

(5.4) 4>(g) = I trig, x)dm(x).

Jx
Here m denotes the unique K-invariant probability measure on X. (m is
unique because K is transitive on X.)
PROOF.

J <r(ykg, x)dk = (j <r(yk, gx) dk) <r(g, x)

= ( J <K7, H*) dkj <r(g, x)

= ( J <r(Y, y) dmK * gx(y) J <r(g, x).

1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 287

Now the measure MR * gx is a it-invariant probability measure on X

and so WIK * gx = m. Thus
Ay<r(',x) = <Ky)a(-, x)
which proves simultaneously that a is semi-spherical, that it belongs
to <£, and that the function <j> defined by (5.4) is a spherical function.
Notice that this yields the easier half of the Harish-Chandra repre-
sentation formula; namely, that the function defined by (3.4) is a
spherical function. For we have shown that the integral in (3.4) is of
the form (5.4) for the space X = G/AN.
We will show that the multiplier functions of Lemma 5.1 generate
the set of all semi-spherical functions. More precisely, we will show
that we may take X = B(G) and each V^ is spanned by the functions
in F*rYE*(£(G)).
To begin with we examine how the multiplier functions of EK(B(G))
are expressed as measures. Suppose x is a positive character on the
group T, and form the measure m\ on T defined by

(5.5) dmxT(t) = xif^dt,

dt denoting left invariant Haar measure on T. From (5.5) we obtain

(5.6) tniT = x(OmT

for / G T . Inasmuch as right-invariant Haar measure on T is related
to the left-invariant Haar measure by multiplication by a positive
character, it follows that the measure m\ is also related to right-
invariant Haar measure in a manner similar to (5.5). As a result we
may also write

(5.7) mrt = x'W^r»

ls
where %' ' some other positive character in T.
Now form the measure niK * in% = v. If g = kt, with &£i£, J £ T, then
triK * g * mj> = x(l)mK * m% = xQ)V' Since KT=G, every g has such a
decomposition. Viewing an arbitrary measure on G as a linear com-
bination of point measures we find that for any bounded measure w,

(5.8) % * œ*niT = cv
for some constant c. From this it follows that v is absolutely continu-
ous, since o> could be chosen absolutely continuous. Let ƒ be the den-
sity of v: dv{g) =f(g)dg. Then
(5.9) f(kg) =f(g), f(gt) = x ' C r W
288 HARRY FURSTENBERG [March

by (5.7). We may therefore define a multiplier by

*(g,yT) =f(gy)/f(y)
since if 7 is replaced by yt, both numerator and denominator are
multiplied by x'^"" 1 )- It is, moreover, clear that <r is a multiplier
and is in MK{B(G)). Thus the measure WK * m%, as a function, is
proportional to the multiplier function <r(g, T).
We shall need the following lemma whose proof we leave to the
Appendix.
LEMMA 5.2. If ir is a measure on G satisfying tir = x(t)nfor / £ ! , and
X a positive character on T, then ir = m? * p for some positive measure /*
on G.
We now turn to the proof of the fact that the i£-multiplier func-
tions for B(G) span the cones V<t>. Suppose /&£ V^. We then have
(5.10) tnK * 7""1 * h = <t>(y)h.
Form the cone V' of all positive measures ir with the property that
WK * 7"1 * TT is proportional to h for all 7 E G . Then hÇzV' and also
V' is closed under translation from the left ir—>gir. We claim that V'
is a cone with compact base in the weak topology on measures. If if/
is some positive function with compact support on G with

ĥt(g)h(g)dg > 0,

define a linear functional L by

L(v) = mK * irWO / J Hg)Kg) dg.

L is continuous and the intersection V{ of V7 with the set of ir such

that L(T) = 1 consists of those ir in V7 for which mK * ir = h. We claim
this set is compact. Since it is closed, it suffices to show that for each
compact set JC.G, ir(J) is bounded as w ranges over V{. But

tnK*T(KJ)= f f CKj(kg) dkdir(g),

with CKJ the indicator function of KJ. But CRj(kg) ^Cx(fe)0(g) so

f Hg) dg è mK(K)w(J) - TT(7),

which gives the desired result. So V' is a cone with compact base.
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 289

Since G and therefore T operates on V' on the left, and since T is a

Tychonoff group (Theorem 4.1), there is a measure 7r£ F ' with
tir = x(t)7r f ° r a ll t£zT and some positive character x- By Lemma 5.2,
7T = m% * /i for some measure fi on G. Thus for some %> the measure
m% * /x is in V7. This means that mR * 7""1 * w | * ^ is proportional to h
for every 7 and, in particular, mR * ^ r * M is proportional to A. But
m*:* ra$<-xr(', x0) for CÇZMR(B(G)) and for x0 the point of 23(G)
corresponding to the coset T. mR * m? * JU is a linear combination of
right translates of mR * m | and so it follows that

Kg) = I <r(gy,Xo)d[A(y) = I <r(g, 7^0)^(7, *0) ^ ( 7 )

(5.11)

- ƒ '«Kg, x) d/j*(x)
J
B(G)
for some positive measure JU* on B(G). Now by Lemma 5.1, (j(-, x)
belongs to some spherical function <f>' defined by

4>'(g) = I <r(g, *) dm(x),

J
B(G)
with m, as usual, denoting the unique 2£-invariant probability meas-
ure on B(G). But then the function h(g) satisfying (5.11) must belong
to the same <j>'. Our assumption, however, was that A £ V4,. We must,
therefore, have <£'=</>. We have thereby proven
T H E O R E M 5.1. Let <f> be a positive spherical function and V^ the cone
of semi-spherical functions belonging to <f>. For each hÇzV$ there is a
multiplier <TÇZMR(B(G)) with o-(-, x)ÇzV4)for each xÇ:B(G) and such
that

(5.12) h(g) = f v(g,x)d»*(x)

J B(Q)

for some positive measure yi* on B{G). The multiplier a which occurs
in (5.12) must satisfy

(5.13) <j>{g) = I a(g,x)dm(x),

J B{G)

where m is the unique K-invariant probability measure on B(G).

It should be noticed that we have incidentally proven that every
positive spherical function <j> admits a representation (5.13), which is
essentially the Harish-Chandra formula.
290 HARRY FURSTENBERG [March

If the i£-multiplier functions o-(«, X) belong to V<f, we shall simply

say that a belongs to <£. Thus every spherical function has a X-multi-
plier belonging to it. In fact there is always a finite set of i£-multi-
pliers belonging to each spherical function, the cardinality of this set
being, in general, equal to the order of the Weyl group of G. (See
[8, Chapter X].) It will develop that even though a number of dis-
tinct if-multipliers (over B{G)) belong to <f>, only one of them is
needed to span V$.
6. The Choquet-Deny representation. The considerations of this
section are based on [3]. We are interested in determining when the
points of a convex cone can be expressed as resultants of measures
placed on the extremal points of the cone. We call a point x of a cone
F a n extremal point if it lies on an extremal ray, namely, if y G V and
x—y(EV imply y=\x for X ^ l . We shall be concerned with cones of
positive (unbounded) measures on a separable locally compact space
X. We regard the space of such measures as the dual to the space C°
of continuous functions of compact support on X and we endow it
with the corresponding weak topology. Suppose F is a closed cone of
positive measures on X and suppose /* is a measure on V. We form the
resultant /Z by setting

(6.1) fl(*)= JT(*)**«,

where \p is a function of compact support, and the integral on the right

is assumed to exist. When it does, the resultant /z is again a measure
in V. First of all, it is a positive measure because (6.1) defines a
positive linear functional on C°. Also /zG V for if / z $ V there would
exist, by the Hahn-Banach theorem, a function i/'GC 0 with jz(\[/) < 0
and 71-0/0 è O for all 7r£ F. This is impossible by (6.1).
In general a cone F is a lattice if whenever Xi,x 2 GF, there exists
an infimum y — 'mî(xu x2) with respect to the ordering of F That
is, there exists y G F w i t h Xi — y, x2 — y&V, and if, for some other
y'GV, xi-y, X2-y'EV, then y-y'EV.
The result we shall need is the following:
T H E O R E M 6.1. Let Vbea weakly closed cone of positive measures on a
separable, locally compact space X and let S denote the extremal rays of
V. Suppose there is a positive function of compact support \[/ with
7 T ( ^ ) > 0 for all 7 T G F , WT^O, and let Vi= {T:T(\//) = 1}. Then f or each
7TG V there exists a measure /z on a Borel subset of ZC\Vi such that T
is the resultant /z. /ƒ, moreover, V is a lattice, then the measure JJL is
unique.
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 291

Clearly if we wish to have a unique representation in a certain set,

then this set can intersect each ray in at most one point. T h a t is the
reason for restricting the representation to V\.
The proof of Theorem 6.1 is based on the corresponding result for
compact, convex sets. If W is a compact, convex set it is possible to
define the resultant /z of a probability measure /z on W, and /z will
be a point of W. We then have [2], [9]
THEOREM A. If W is metrizable then the set of extremals of W is a
Borel set and every point of W is the resultant of some measure on the
set of extremals of W.
A convex set W may be taken to be the base of a convex cone. For
example, if W is in the linear space 5 we may form the cone W in
RXS by setting W = {(/, tw) : t> 0, wG W). We say that W is a sim-
plex if W is a lattice.
THEOREM B [2], [9]. If W is metrizable and is a simplex, then each
point of W is the resultant of a unique probability measure on the set of
extremals of W.
We apply these theorems to the cone F of Theorem 6.1 by showing
that each point of V lies in some compact convex subset WQ V. Let
7rG V. By the separability of X there will be some continuous func-
tion h everywhere positive in X and such that ic{h) < <*>. Suppose
ic(h) = 1. Let W be the set of x ' G V with ic'Qi) S1. This set is closed
in V and moreover, since h(x) > 0 everywhere, so that on every com-
pact set h(x) is bounded from below, the measures irf in W are
bounded in every compact subset of X. Thus W is a compact convex
metrizable set. Let us show that each extremal of W with the excep-
tion of the point 0 is an extremal of V. If ir' is an extremal of W and
7r' = xi' +7r2' with irl G V, then ici {h) > 0 unless ici = 0 . So

iri (h) TT2' (h)

which expresses ic' as a convex combination of measures in W since

ic{ (h)+T2 (h) = l. It follows that ic{ and ic{ are proportional to ic'.
Now by Theorem A, ic is the resultant of a measure on the extremal
points of W:

x = ƒ w' d»W) = ƒ - ^ r r *'(*) dnW)

and this expresses ic as the resultant of a measure on SP\ V\.

292 HARRY FURSTENBERG [March

We observe that conversely, if T is given as the resultant of a

measure on 8>C\ Vi, then this measure corresponds to one on the ex-
tremals of W. Namely, if

then fwf(h) dfx(irf) = 1 and Tr'(h)<<x> for almost all T' with respect
to fi. We can then write

and TT'{h)dix{irf) is a probability measure, and the measures ir'/irr{h)

are extremals of W if the x' are extremals of V. As a result of this
observation it follows that to prove the uniqueness portion of Theo-
rem 6.1 it suffices, in view of Theorem B, to show that if F is a lattice
then W is a simplex. Equivalently we must show that W is a lattice.
This follows readily. If Xi — (ty /7Ti), X2 = (S, ST2) are two points of W
where 7Ti, TT^W, then if t<s, we set

y = inf #-(#!, x2) = ( t, t imV ( TTI, — T2

) ) •

Notice that y€: W^ inasmuch as inf^C^i, sir2/t) is in W (being bounded

by 7TiG W). If we rewrite ;y as
(/, inîv(tTi, sw2)),

it becomes clear that Xi — y, x2 — y(EW and if X\ — y', x2 — y'GW then

y-y'GW.
The uniqueness of the representation in Theorem 6.1 for the case
t h a t F is a lattice may also be formulated as follows:
COROLLARY. If Vis a cone of measures as in Theorem 6.1 and Vis a
lattice, then if v is a signed measure on &C\ V\ such that fir' dv{irf) £ V,
then v^O.
PROOF. Write v = v\ — v2 with P » ^ 0 and suppose that Jir' dv{jl)
= fir' dvz{irf) with v^O. Then v\ = v2+vz and, by uniqueness, v\
— v2-\-v%. Then J/ = J > 3 ^ 0 .
Theorem 6.1 is used to prove the following result which we refer
to as the Choquet-Deny representation.
T H E O R E M 6.2. Let G be a separable locally compact group and suppose
that {fjiat a £ r } is a commuting family of positive measures with com-
1965Î TRANSLATION-INVARIANT CONES OF FUNCTIONS 293

pact supports on G (/x« * jti/3=M/3 * /za, a, jöGT). Let V denote the set of
positive measures w on G satisfying jxa * IT = T for all a G T . Assume
moreover that there exists a continuous function of compact support
^ 0 on G with ir(\l/)>0 for all 7rG V, TT^O. If 8 denotes the set of ex-
tremals of V and Vi= {w: 7r(\f/) = 1}, then, for each TÇZV there exists a
unique Borel measure vv on 8>r\ V\ with ir = vv.
PROOF. Since the measures jua have compact support it follows that
V is closed. Thus the existence of a measure vT is guaranteed by the
first part of Theorem 6.1. To prove uniqueness we must show that
F is a lattice. Suppose then that 7Ti, 7T2G V. Denote by ir' the infimum
of TTI and 7T2 in the lattice of all positive measures, TT' may be defined by
d d dlr
*' / x . / *i , x * /v
dyiri + fl-2) \d(wi nr iry dyiri + 7r2; •

Let 5 denote the convolution semigroup of measures on G generated

by the jua, a g r . We shall consider the family S * ir' of measures
fx * 7r', /-tGS. We have
Ma * 7r' ^ lla * 7T» = 7Tt', i = 1, 2,

so that lia * 7r' ^7r'. Hence /x * 7r' ^7T' for any /*GS. By the commuta-
tivity of S we also have
(6 .2) /*' * (fi * 7r') ^ M * TT'

for jLt, /x'G-S.

Choose a dense subset {/xn * ir'} in »S * TT'. This may be done because
G is separable. Let
7T = Hm /Zi * • • • * fln * 7r'.
n-»°o

This limit exists because the sequence of measures is decreasing by

(6.2). Clearly 7r^jLt * T' for any J U G S . Also /*« * 7r^7r. But jiia * 7r is a
limit of measures in 5 * TT' and so irSixa * TT. Hence 7T = JU« * ?r and
T T G F Since TT^TT' we have TTI — 7rG F and 7r2 — 7TG V. Now suppose
7Ti —#, 7r2 —7rG V for some T G F Then T^TT' and ju * 7r^/x * ir' for
all J U G 5 . But if T G F then JA * T=T, SO X^JLI * 7r' for juGS, and so
7râ7r or 7T —7rGF. Therefore 7r = inf(7ri, 7r2) with respect to F. This
proves the theorem.
7. Application to semi-spherical functions. We apply the foregoing
to obtain more information regarding the representation of semi-
spherical functions in terms of multiplier functions. In particular we
shall see that for each spherical function <t> ( # > 0 ) there is a single
294 HARRY FURSTENBERG [March

üC-multiplier <T<I>E:MK{B{G)) such that V4, is spanned by o>(-, x). G

is again a semi-simple Lie group.
LEMMA 7.1 [7], [8]. The measures mK*g*MKfor g £ G commute
with one another.
PROOF. We use the fact that there is a map g—>g' of G into itself
with the following properties : (a) (g')' = g, (b) (gig2)' = gi g{ , (c) k' = k~l
for kGK, (d) if P = {g:g' = g) then KP = PK = G. (In case
G = SL(w, i?), g' denotes transposition.)
We extend the operation g—»g' to measures /x—»//. We then have
m'K — mK and (mK * p * ms:)'= mK * P * MR for pÇ^P. Now every
g £ G may be expressed as g = &£ for &£i£, pÇLP and m^ * g * w^
= mx * £ * Wis;. It follows that (mm * g * m*)1 = WIR * g * ^Ü:. From
this it follows that generally (w^ * ju, * m^)' = m^ * /x * m^. So
mK * gi * wx * mK * g2 * WK is invariant under ( )', but, on the other
hand, using (b) it must equal (w^ * g2 * wx)' * (wir * gi * m^)'
= wx * gi * Wx * m# * gi * m*:. This proves the lemma.
Now the functions in V^ are the solutions ^ 0 of
(7.1) trtK + y-1*/ = *(?)ƒ.
The measure solutions of (7.1) are actually functions because ƒ is
proportional to mK * M * ƒ for any /*, in particular, an absolutely con-
tinuous one. Since mK * ƒ =ƒ we may rewrite (7.1) as
(7.2) ^ ( T ) " 1 ^ ^ * 7-1 * m^ * ƒ = ƒ.

Since the measures w^ * 7 - 1 * m^ are mutually commutative and of

compact support, the cone V<j, of solutions to (7.2) is one to which
Theorem 6.2 applies. Moreover, a function ƒ G V^ must be everywhere
positive unless ƒ = 0 , since by

(7.3) f ftokg) dk - *(y)f(g);

J K

if Kg) == 0» since ƒ is continuous, ƒ must vanish everywhere. Hence

for any function \p ^ 0 of nonempty compact support, ff(g)$(g) dg > 0.
Choose \p so that ^(gk) =^(g) for fe£i£ and fy{g)4>(g) dg = l. Then

f f(g)Ug)dg= f f f(gk)Hg) dgdk

J Q J K J G

- f<t>W(g)dgf(e)=f(e).
J
G

The set Fi of Theorem 6.2 is therefore the set of functions with unit
1965J TRANSLATION-INVARIANT CONES OF FUNCTIONS 295

value at the identity. We refer to SP\ Vi as the normalized extremals.

Theorem 6.2 then asserts that each semi-spherical function in V<t> has
a unique representation in terms of normalized extremals of V^
Now by Theorem 5.1, every function in V^ is an integral over multi-
plier functions a(', X) in V^ (equation (5.12)). It follows that an ex-
tremal of V<j> must already be proportional to a multiplier function.
Since the multiplier functions satisfy <r(e, x) = l, we see that the
normalized extremals of V<t> are multiplier functions.
We next show that only one i£-multiplier o> G-Wx (£((•?)) can be
such that the functions o>(-, x) are extremals of V+. Suppose that
o"i(-, x) and cr2(-, x) are all extremals in V+. (Note that if o*i(', Xi) is
an extremal so is Ryai(•,Xi)= oi(•, 7x1)0*1(7, xi) so that all <7i(•, x) are
extremals if one of them is.) Now if CTI( •, x) G V* then JB(G)<Ti(g, x) dm(x)
= <t>(g). The same holds for cr2. But <£G ^ has a unique representation
in terms of its extremals (Theorem 6.2) and so the set of functions
(7i(-, x) must be identical with the set o-2(', x). Now both these sets
of functions B\ and £ 2 form G-spaces with the operations h-*Rgh/h(g).
Moreover it is easy to see that #—xr,-( •, x) then gives an equivariant
map of B(G) onto Bi} i = l, 2. But the spaces Bi are the same, so if
a"i7^cr2, we obtain two distinct equivariant maps of B(G) onto B\ = B%.
LEMMA 7.2. If ji: B(G)—>B, j 2 : B(G)-±B are two equivariant maps
of B(G) onto the same space B, thenji=j2.
PROOF. It suffices to show that for some point x 0 G ^ ( G ) , Ji(#o)
—J2(x0). For any x<E.B(G) form the measure ^r=i{oy1(aJ) + Syt(a.)} on
B. Since B is a boundary (end of §3), there exists a sequence {gn} C.G
with gnT converging to a point measure. But extracting a subsequence
we will have gn^—*#o
for some XoG^(G) and so gn]F-*\ {Sy1(afo) + Sy1(Xo)}.
If this is a point measure we must haveji(xo) =j 2 (xo). Hence J i = j 2 .
It follows by this lemma that the maps x—><7;(-, x) must be the
same for i = 1, 2, and so ai = a2. We have thereby proved
THEOREM 7.1. TO ea£& positive spherical function <f> on G there is a
multiplier (T4><E.MK{B{G)) such that every positive solution to

f f(ykg) dk = 4>(y)f(g)
J K

is given by

(7.4) /(g)- f <r*(g, x) dr(x),

J B(O)

where v is a positive measure on B(G).

296 HARRY FURSTENBERG [March

For, the function ƒ G V^ has a representation in terms of normalized

extremals of V<f>. We have seen that these extremals must be multi-
plier functions and, moreover, only one multiplier belonging to <f> can
give rise to extremals. We denote that multiplier by &$.
There is still one improvement that may be made on Theorem 7.1.
We know that a measure on the set of extremals of V<f> is uniquely
determined by its resultant, but we cannot as yet identify B(G) with
the set of normalized extremals. Namely the map x—»Ö>(', X) may
not be 1-1. We shall see later that this map is 1-1, so that the measure
v in (7.4) is uniquely determined by the function ƒ.
According to Lemma 5.1, every multiplier function in EK(B(G)) is
semi-spherical and belongs to some V<t>. We may then apply Theorem
7.1 to obtain an integral representation of any multiplier function
with respect to B{G) in terms of some o>. Even in the simplest cases
this leads to nontrivial relationships.
Let us illustrate this for the case of the group G of analytic auto-
morphisms of the unit disc. In the Iwasawa decomposition of this
group G, the vector group is 1-dimensional. It follows that MK(B(G))
is 1-dimensional (o-(g, yT) is determined by the positive character
(T(/, 7") which is determined by a(a, T)). For this group, B{G) is the
boundary of the unit disc, B{G)= {ei9}. We have seen in §3 t h a t
| g'(eid) | is a i£-multiplier with respect to B{G) and so it follows that
all the i£-multipliers in MK(B(G)) are given by
(7.5) *8(g, eiG) = | g'(eie) |', - oo < s < co.
If we denote by m normalized Lebesgue measure on the unit circle,
then (7.5) can also be written as

We recall from §3 that this expression always represents a multiplier.

The spherical functions on G and therefore given by

Mg) = f<r*(g,eie)dtn(ei(>).

Now, in general, if G is a semi-simple group and B(G) is its asso-

ciated G-space we can form a 1-parameter family of spherical func-
tions by taking
rdg~lm "I*
J — — (x) dm(x),
B(Q)L dm J
- co < s < oo.

We then have
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 297

\±r-(*)\ ~^—(x)dm(x)
B (G) L dm j dm

= Jamt
f r ^dm( * ) lJ °dg~hn(x)
(7.7) =f [^(g-i^V^^)
r dm 1~*
J —(*)
B(G)Ldgm J
**(*>

= f r^(x)]8^(o:) = ^(r1).
In our case it can also be verified that 4>s(g) = <t>s(g~~1)- (Every g £ G
can be written g — kihkz where k\ andfe2are rotations and

*(*) = > - 1 < r < 1.

1 — rz
But h = h~l, so <£*(g) = <j>s(h) =4ts(h~1) =<t>s(g~1)') For this group it fol-
lows that <TS and cri_8 belong to the same spherical function. As a re-
sult, for each 5 we must be able to express ov(-, eid) in terms of
(Ti_s(-, ei9), or vice versa. Now it can be seen that 0*1/24-* grows more
rapidly than cri/2-t and, as a result, the latter must be expressed in
terms of the former. Granting that such an expression exists, it is not
difficult to determine what it must be. If we write <ri_s(g, 1)
= f<Ts(g, eiB) dv(d), and use the multiplier property for (Ti_s and a3
and the uniqueness of the decomposition in terms of crs(-, x)—the
extremal of V<f,a—we conclude that

- ^ ( O = <n-.{t, l)*(r», e<°)

dv
for t<E:G and £1 = 1. This determines the measure v. If we write
/ 1-r2 V
*.(«, «") = ( — - — — = P(r, 0 - *)•,
\ 1 — 2r cos (0 — </>) + f V
the relationship in question is

J
» 2T

PM-^Hi-costfO*- 1 ^,
0
(7.8)
7. = I ƒ (1 - cos4)*-1 d*\-K
298 HARRY FURSTENBERG [March

(7.8) is proved by these means to be true for s> 1/2. Since both sides
of (7.8) are analytic for <3U>l/2, the same formula holds for these
complex values as well. Apparently a limiting form of (7.8) is still
valid for 6is = 1/2. In this case the kernel (1 —cos </))1/2+it is singular,
but with the proper interpretation it defines a unitary operator on
L 2 (0, 27r). Equation (7.8) then shows that this operator is the inter-
twining operator relating the equivalent unitary representations ob-
tained from the complex multipliers ai/2+it and o-i/2-u (see §2).
8. The general convolution equation. We now come to our main
result which deals with solutions to the general convolution equation

(8.1) ff(g'g)d»(g') = \f(g), X>0.

Unfortunately, we shall not be able to handle the case of an arbitrary

positive measure /*. We shall have to assume, to begin with, that fi is
absolutely continuous and we shall also make a special assumption re-
garding the support of ju. Specifically, we shall take dfi(g) =^(g)dg
where \p(g) is a bounded measurable function of compact support A,
and assume that A has the property that some power Aw contains a
neighborhood of the identity. These hypotheses regarding if/ will be
assumed in force throughout our discussion. Our equation now takes
the form

(8.2) ƒ f(g'g)Hg') dg' = \f(g)

and we denote by V\(\[/) the cone of non-negative solutions to (8.2).

We shall show that V\(\p) is spanned by its extremals and that the
extremals are multiplier functions. Note that unlike equation (5.3),
the solutions to (8.2) need not satisfy ƒ (kg) = ƒ (g). As a result, the
multiplier functions that we deal with now are no longer if-multiplier
functions. As usual, V\(\p) is translation invariant, and so, for a
multiplier sEM(B(G)), if s(-, ^ i ) G W for some XiGB(G), then
*(•, * ) G W ) for all x<EB(G).
LEMMA 8.1. In each equivalence (cohomology) class of H(B(G)) there
is at most one multiplier s with s(-, x)£:V\(\p).
PROOF. Let us denote by \[/(n) the n-îold convolution of \[/ with it-
self. Notice that under our hypotheses, \p(n) will be continuous of
compact support containing a neighborhood of the identity for some
n ^ 2 . (8.2) implies that
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 299

(8.3) ƒ fig'gWKg') dg' = X»/(g).

If s ( ' , x) satisfies (8.2), then

Js(g,x)^(g)dg = \\

If s ' ~ s then s'(g, x) = (p(gx)/p(x))s(g, x), so that if s'(-, x ) G F ^ ) ,

J P(gx)s(g, %Wn)(g) dg = An/>(z).

Then

J P(gx)s(g, *Wn)(g) dg
(8.4) ƒ>(*) =
J *(g, x)Vn)(g) dg

If r C ^ ( G ) denotes the set where p(x) attains its maximum, and

*A(n)(&)>0, then g r c r , according to (8.4). Since this set of g com-
prises a neighborhood of the identity, GT C T so that T = B(G). Hence
p is constant and s' = s. This proves the lemma.
If the function ƒ on G corresponds to the measure ir, then equation
(8.1) may be rewritten p * 7r = X7r. Here p is the measure correspond-
ing to the functions $(g~l). Because p is absolutely continuous, T
must also be, so the set of positive solutions to p * w = \w is F\(^).
We also note that the non-negative solutions to (8.2) are all con-
tinuous and everywhere positive. This follows from (8.3) since ^ ( n )
is continuous, and if f(g) = 0 then ƒ{g'g) = 0 for g' in a neighborhood
of the identity of G.
We may now apply Theorem 6.2 to the cone V\(\[/). For, to begin
with, this cone coincides with the cone of measures satisfying p * T
=XTT. Secondly, ff(gMg) dg=\f(e)>0 for every ƒ G Fx(</>). So if S
denotes the set of extremals of V\(\p) and V{ 0/0 denotes the set of
functions in V\(yp) with f(e) = 1 we have
LEMMA 8.2. For every function ƒ'£ VxWO there is a unique measure v
on Vii$)C\Z with

(8.5) f(g) = f ƒ«(*)*(«).

The deeper part of our analysis concerns the identification of the

 300 HARRY FURSTENBERG [March

extremal functions in V\ (jp) with multiplier functions. We shall first

prove
LEMMA 8.3. For each extremal ƒ £ V\ ( ^ ) n 8 , there is a G-space X on
which K is transitive, and a multiplier sÇzM(X)f such that f (g) = s(g, £)
for some £ £ X .
PROOF. T O begin with, let us notice that the space V\ (^)H6 is a
G-space. For if f(g) is an extremal of V\ ty), so is the function ƒ (gy) /ƒ (7).
Thus G operates on V\(ip)r\& by right translation (and normaliza-
tion). Let us form the subset X of extremals f(gy)/f(y), 7 G G . If
£ = {y'f(gy) ~c(y)f(g) }» where c(y) is some arbitrary constant, then
X = G/L. We can define a function s on G XX by
(8.6) s(g,yL) =f(gy)/f(y).
s is well defined since if 7 is replaced by 7Z, Z£L, both numerator and
denominator in (8.6) are multiplied by c(l). Moreover, 5 is a multiplier
in M(X). To prove the lemma, it is merely necessary to show that K
is transitive on X. (This will, incidentally, imply that X is compact,
a fact which is not otherwise obvious.)
We shall use the following notation. If T\ and 7r2 are two measures
we shall write 7Ti>^7r2 if, for some € > 0 , iri^tT^ We call a measure x
radial if it satisfies k\ * T * k2 — 7r for ki, kzÇzK.
If p is any continuous function on G with compact support, there
is an n with \//(n)y-p- \j/ is defined by vKg) —*Kg_1)- Choose p radial and
positive in a neighborhood of the identity. Then
^<»> * ƒ = \ n /,
and so p *ƒ-<ƒ. If we now choose p > l on the support of ^, we will
also have ^ * ƒ^Cp * ƒ. Hence
(8.7) f<P*f<f-
Since p is radial, mR * p = p * mR = p. By Lemma 7.1,
(8.8) p * f*niK = p *ntK*f*wiK = mR *f*MR *p = w& *ƒ * p.
If p' is another radial function whose support contains a neighbor-
hood of the identity, and p ' ^ 0 , we will have p' * p>-p. So
(8.9) P*f*p' = P *niR *ƒ * p = ntR *ƒ * p' *p>- % *ƒ * p,
or p * ƒ * p'>-p * ƒ * w*. By (8.7),
(8.10) f*p'>-f*niR.
Returning to the functional notation we have
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 301

f/w-won *? >- ff(gk)dk.

J G JR.
Setting/(g) = s(g, £0) (see (8.6)), we can write

f s(g, g'-%)s(g'-\ £o)p'(g') dg' >- f s(g, k£0)s(k, f o) dk

J G JK
or

r s(g, g'-^)P'{g') dg' >- r .(g, «o) <z£

which may be written

(8.11) f s(g, Ö iwift) >- f j(g, Ö <M£).

Jx J x
Now the functions s(g, £) are extremals and V\(y//) is a lattice. By the
corollary to Theorem 6.1, (8.11) implies that coi>-co2. Now X is a
manifold as the quotient of a Lie group by a closed subgroup, and
the measure «i, as image under the natural map of G-+G/L of an
absolutely continuous measure, is itself absolutely continuous on X.
(This means that the restriction to a coordinate neighborhood is ab-
solutely continuous with respect to Lebesgue measure induced on
that coordinate neighborhood.) On the other hand, co2 = W2?:*£o is
concentrated on the submanifold KÇoQX. A measure on a submani-
fold of X can be absolutely continuous on X only if the submanifold
is open. But K$j0 is closed and X is connected. It follows that K^ — X
which is what was needed to prove the lemma.
To prove t h a t the space X occurring in Lemma S.3 is B(G) or an
equivariant image of B{G), we use the next lemma. We shall continue
to use the notation ir{p-T2 introduced in the proof of Lemma S.3.
LEMMA 8.4. Let Y and Z be two homogeneous K-spaces and suppose
there is a map from Y to the space of probability measures on Z, y-+dVl
such that Oky — kdy for kÇiK, and whenever /zi, /z2 are two positive meas-
ures on K such that JJLI * 0„>-/ji2 * 0y, we have /zi * yy-fx2 * y- Then there
is a map j : Z—+Y, equivariant with respect to K, and such that j~"l(y)
coincides with the support of 6yfor each y(£Y.
PROOF. We set Y=K/Ki, Z = K/K2. If mKl denotes the Haar
measure on Kh then Y may be identified with the space of measures
{kntKxi kÇ:K\ on K. Similarly, the space of measures on Z can be
identified with the space of measures {v * m#,: v a measure o n l } .
Namely, if co is a measure on Z, set
302 HARRY FURSTENBERG [March

JKIK%L JK* J

so that w is a measure on K. Then o)k = co for &£i£ 2 , and so co = co * m^,.

Since cô—-KO under the natural map of K—>Zt the correspondence
co<->co is 1-1.
Let yoG Y correspond to the coset K\ or to the measure niKv If
dV0 corresponds to v * TUKV then, since K\ leaves y0 fixed, it must leave
6V0 fixed, so that Oy^m^ * 6VQ1 and 0j,o corresponds to W,KX *V* m,KY
The map y—>9U is then given by
(8.12) kniKi ~» ^ i T j * *> * wir r
The hypothesis of the lemma may then be rewritten : /Ji * raj^ * p * m#,
>-jLt2 * WÜTJ * v * ra^jj implies jui * m ^ ^ ^ * mirlt We can make this
more precise. Suppose, in fact, that /xi * rrtKx * P * w^ 2 ^ jtz2 * wx t * v * m ^ .
Then for some O 0 , / i i * w x ^ C / i 2 * m ^ . Take the largest possible C
for which this is true and suppose C<\. Then
(Ml * W>K ~ CjLt2 * MKI) * WjST! * J' * ^ Z 2 è (1 — C)H2 * m>Kx * V * MKV

Now /xi * wx — C/z2 * w^j is a positive measure, so we can apply the

condition in the lemma again. We get, for some e > 0 ,
0*1 * WKI ~ CjJL2 * tJtKi) * W>KX è €/i2 * WK»
or
/*i * WJTJ 2> (C + e )A*2 * mKl.

This contradicts the choice of C. It follows that C ^ l . What we have

shown is that, for a signed measure ju on K,
(8.13) fi * ntRx * v * mKt ^ 0 implies /z * /w^ ^ 0.
Suppose, more generally, that we have two positive measures o?i
and co2 on K with the property that JJL * coi^O implies jit * co 2 ^0. We
claim that we must have co2=coi * co for some positive measure co on
K. The proof is based on the Hahn-Banach theorem. Namely, form
the cone of all coi * co, co^O. This cone is weakly closed. If co2 did not
belong to this cone, it could be separated from it by a continuous
linear functional. T h a t is, there would exist a continuous function h
with coi * œ(h) § 0 for all co^O on K, but co2(&) < 0 . Now the first in-
equality implies that % * co3 ^ 0 (h(g) =h(g~1)). Hence h * co 2 ^0 which
in turn implies co2(A) ^ 0. So co2 must belong to the cone in question, or
C0 2 = C0i * CO.
By (8.13) we therefore have ntK1 = fnK1 * v * w^, * co for a positive
measure co on K. Let P be the support of v and Q the support of co.
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 303

The support of a convolution of positive measures is the product of

the supports, so PK2Q CKX. This gives PK2P~l <ZPK2QQrxK2P-x CKi-
To define the map j it is merely necessary to verify that every point
of Z belongs to the support of some 0y and that the supports of 0VI
and 0V2 do not overlap unless yi = y2. As to the first question, we notice
t h a t since some point belongs to the support of some dy, and Z is
homogeneous, every point belongs to the support of some 0y. For the
second question we observe that the support of OUKI is the image in
ZoikK1PK2 by (8.12). Soif
k'KxPKifW'KxPKi ?* 0,
then kf"1kffeK1PK2P-lKT1CK1 and k'K1 = W'Ki. This completes
the proof of the lemma.
We return to the extremal ƒ in V{ (\p) and the space X which it
generates. We have ƒ = $ ( • , £0), 5 6 M ( I ) , and we have shown that K
is transitive on X. At this point we invoke the theory of semi-spheri-
cal functions. By Lemma 2.1, every multiplier in M(X) is equivalent
to a X-multiplier:

(8.14) s(g, Ö = " ^ « r f c , 0 , «r S MK(X),

where p(£) is continuous and positive in X. By Lemma 5.1, the func-

tions cr(», £) are semi-spherical and belong to the spherical function

Jx
where m is the unique X-invariant measure on X. By Theorem 7.1,
the semi-spherical functions belonging to <j> are all generated by some
v*eMK(B(G)):

J B(G)

for some measure 6' on B{G). Let B' be the equivariant image of B(G)
obtained by identifying Xi,x2Ç:B(G) if o^(g, Xi)=a<i>(gt x2). (One
verifies readily that this is an equivalence relationship.) o> defines a
multiplier G I in MK{B') and we may write

(8.15) »(*,*) = f °i(g,y)dei{y).

J B'

Because the <r/(-, y), yÇzB', are all distinct, the points of B' cor-
respond to the normalized extremals of V4. In particular, the meas-
304 HARRY FURSTENBERG [March

ures 0$ in (8.15) are uniquely determined. Since a(g, kÇ)=cr(gk, £),we

conclude from (8.15) that kd^ — 6k^ We now show that the map
£—»0£ from X to the probability measures on B' satisfies the condition
of Lemma 8.4. Suppose then that jui, ju2 are positive measures on K
with /xi * 0£>-/i2 * 0£ for some £ G Z . Note that by (8.14), s>-cr>-s. As
functions of g we then have

f s(g, kQ dn(k) >- f *(g, kQ dn(k)

J K J K

= f *(gk9Q dpi(k) = f f ^(gk9y)dm(k)dek(y)

J K J Br J K

= f f */ (g, *y) dnWdOtiy) = f er/ (g, 2) dMl * ^(*)

J B' J K J B'

>- I o-/ (g, 2) <*/i2 * 0*(«) = I I or/ (g, fry) dfi2(k)dd^y)

J B, J B, J K

= f f «I (gk, y) d,x2(k)d0^y) = f <r(s£, Q itx^k)

J B' J K J K

= J <Kg> H) <W*) >- J s(g, kQ dfx2(k).

Now the extreme ends of this series of inequalities are integrals over
s(g> rç), rç£-X\ and these are extremals in V\ (\p). By the corollary to
Theorem 6.1, since Fx(t/0 is a lattice, we must have jUi * £>-ju2 * £.
The conditions of Lemma 8.4 are thereby verified and we can con-
clude that there is a map 7: B'-^X such that 0$ has its support equal
to/"KO.
Finally we must show that j is equivariant as a map of G-spaces:
j(gy)=ti(y). By (8.15),
<Kg7,Ö = <Kg> yQ<r(y9 Q
(8.16) f
= *(7,Ö *j(g,z)dOyi(z).
J gr

On the other hand,

<Kg7,Ö = I <r+(gy,y)d6t(y) = I <rj (g,yg)<rj (y,y) dBfa)

J B' J B'
(8.17)
= J <r# (g, « W (7, 7 _1 z) <fyöï(z).
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 305

By the uniqueness of the Choquet-Deny representation in V^ we

must have

This implies that

(8.18) y0s-<0yi
for any 7 G G . Suppose now that j(y) = £. Then y G support(0$) and
gy G support(g0$). By (8.18) this implies that gy G s u p p o r t ^ ) or
j(Û)—gy- This proves that j is an equivariant map of B' onto X.
Now B', in turn, is an equivariant image of B{G). But the map
B(G)-*X induces an injection of M(X)—*M(B(G)) and so the multi-
plier s£zM(X) corresponds to a multiplier s*£M(B(G)). We have
proved
THEOREM 8.1. The cone V\(\[/) of non-negative solutions to

ff(i'i)Hi')dg' = \f(g)
admits a Choquet-Deny representation in terms of its normalized ex-
tremals. Each normalized extremal has the form s(-, x), where
sÇ:M(B(G)) and xÇ.B(G). Moreover, if rj(EH(B(G)), there is at most
one multiplier snÇ:r} such that sv(', x ) G F \ ( ^ ) .
9. Basic multipliers. Theorem 8.1 tells us that the extremals of
V\(\f/) are proportional to multiplier functions in E(B(G)). However,
not all the multiplier functions in V\(\p) need be extremals. For
example, taking <j>^l on G, the multipliers (dg~1m/dm)(x)=ai(gJ x)
and (To(g, #) = 1 in MK(B(G)) both belong to V^. However,
dg~lm
J B(Q) dm
(x) dm(x),

so that (To cannot be an extremal of V^. In this section we shall intro-

duce a class of multipliers, the basic multipliers, such that the nor-
malized extremals of V\(\f/) coincide exactly with the set of basic
multiplier functions in Vx(i^). In the course of this discussion we shall
resolve another question. Namely we shall show that the space X of
the preceding section is not just an image of B(G) but coincides with
B(G). In addition we will find that the space J3' = J3(G) so that the
space B{G) coincides with the set of extremals in each V<j>.
DEFINITION 9.1. s(E:M(B(G)) is basic if, for a signed measure o> on
B(G),
306 HARRY FURSTENBERG [March

I s(g, x) dw(x) ^ 0 for all g implies w ^ 0.

We denote by Mb(B(G)) the set of all basic multipliers, and by

Eh(B(G)) the set of all multiplier functions coming from basic multi-
pliers.
In particular, Js(g, x) dco(x) = 0 implies co = 0 for 5 basic, and so the
functions s(-, x), x G ^ ( G ) , must all be distinct. It follows that a
multiplier on B(G) induced by a multiplier in an equivariant image
B' of B(G) cannot be basic unless B'= B(G).
LEMMA 9.1. s<EM(B(G)) is basic if and only if

(9.1) J s(g, x) dù)i(x) >- J s(g, x) do>2(x)

implies o>i>-a>2 for any two positive measures coi, œ2 on B(G).

The proof is similar to the proof of (8.13) in Lemma 8.4.
LEMMA 9.2. If s\ and s2 are equivalent multipliers then Si is basic if
and only if s% is basic.
PROOF. For, if si and s2 are equivalent, Si^-s^si, so that (9.1) for
Si is equivalent to the same assertion for s2.
Thus the condition of being basic depends only on the equivalence
class of 5 and we may introduce the set Hb{B{G)) of basic (cohomol-
ogy) classes.
LEMMA 9.3. s is basic if and only if the cone of functions on B(G)
spanned by the functions s(g, •) is dense in the cone of all positive con-
tinuous functions on B(G).
That s is basic under the stated condition is immediate. The con-
verse is a consequence of the Hahn-Banach theorem.
According to Lemma 9.2, 5 is basic if and only if the i£-multiplier
equivalent to s is basic. We will now show that the basic iC-multipliers
are just the a^ which we saw spanned the cones V$.
T H E O R E M 9.1. <T£.MK{B(G)) is basic if and only if <r = <r<t>for some
spherical function <t>.
PROOF. We shall first show that if a is basic it coincides with some
Ö> This is the easier of the two directions. Suppose then that a is
basic. In any case o-£ V^ for some spherical function <j> by Lemma 5.1.
Then, as in the proof of Theorem 8.1,
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 307

(9.2) »(&*)- f °i{g,y)dex(y).

J B,

Now suppose that JJLI * 0$>-jU2 * 0$ where /JI, ju2 are measures on if. Then

I <K& y) dm * «(y)
•J B

= I *(& is) dfii(k) = I or(g*, ou) dfjLi(k)

JK JK

n 3'
a{(gk,z)ddx(z)dixi(k)= f f *{ (g, kz) ddx(z)dm(k)
J E." B'

= I *j(g,u)dlH*0s(u)9
J B'

for i= 1, 2. Since MI * 0x>-/x2 * 0X it follows that

J <Kg, y) ^ i * *0y) >- J *{&> y)fo** *60*

By Lemma 9.1, since a is basic, this implies jUi * #>^JU2 * x. We see

that the map x—>6X from B{G) to B ' fulfills the conditions of Lemma
8.4. This means that there is an equivariant map (see the proof of
Theorem 8.1) j : Br—>B(G) such that dx has its support on j~l(x). But
B' is itself an equivariant image of B{G)—ƒ : B{G)-*B', so that j / is
an equivariant map of B{G) onto itself. Now, by Lemma 7.2, the
identity is the only equivariant map of B{G) onto itself. So jf is
the identity. Since ƒ is onto, j is an isomorphism. Hence B' &B(G)
and j is an equivariant map of B{G) onto itself; hence the identity.
Since dx has its support on j^ix), it follows that dx = 8x and cr(g, x)
= o>' (g, x) = <r<t>(g, x). This proves that if a is basic, then a = a<t).
Now consider the X-multiplier &$ for some spherical <£. We wish to
prove that o> is basic. We must first establish that o> separates points
in B(G). To do this introduce the space B' as before: B' is obtained
from B{G) by identifying Xi, x 2 £ £ ( G ) if o>(g, Xi)=(r^(g, x2) for all
g £ G . Let j denote the ensuing natural map of B(G)—>B'. We then
define cr/ as before by cr/ (g, j(x)) = o"4>(g, #).
For each £ £ - 5 ' , let K^C.K denote the subgroup of i£ that leaves
£ fixed. i£s takes j " 1 ^ ) into itself and is transitive on j - 1 (£) since K is
transitive on B(G). So there is a unique probability measure 0$ on
j~ x (£) invariant under i£$. We see readily that kdz = 9kz for fe£-K*. If
308 HARRY FURSTENBERG [March

7 G G and Y£ = £, then 7 determines a differentiate homeomorphism

of j _ 1 (£) with itself and so y0$ is absolutely continuous with respect to
0£. Since every g £ G is a product ky with &£i£, Y£ = £ (since K% = Bf),
it follows that gd^ — kyd^ is absolutely continuous with respect to
&0£ = 0/k£ = 0*7$ = 6QZ- Then g " - ^ is absolutely continuous with respect
to 0$ and the expression

(9.3) ^(g,^) = ^ i ( x W ( g , 0

is well defined for x in the support of 0$, that is, for j(x) = £. We claim
that with this definition of £, <r(g, x) defines a multiplier in MK(B(G)).
a(k, x) = 1 since k~ldki — 0^ and cr/ (fe, £) = 1. What remains to be veri-
fied is the multiplier identity. Letting £ = j(x) we have, making use of
(3.13):
. . dgrlgrlQoiQrt , . ,f ..
<r(gig2, *) = ( * W (gig*, Q

= 'lg* (g2*W (gi, g 2 fW («2, Ö

rfg20f
dgT^QiQzt ^W
= (g2«) — — ( g 2 * W (gl, g£)<T* (g2, Ö
^i€ ^g20{

= 0-(gl, g 2 *) — ( * W (g2, Ö
U0£

= <Kgi>g2*Mg2,*).
Here we have used the fact that j{gtx) =g2? if j(%) =£•
We next verify that this multiplier a also belongs to 0. It suffices
to show that

(£.4) I cr(g, x) dm{x) = 0(g).

B(O)
But

J B(Q)
cr(g, tf) d w ( x ) = I <r(g, kx0)
J
K
dk

o-(g, kk'xo) dkdk'.

Here x0 is any point of B(G) and X ' is any subgroup of K. If we let

1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 309

%o=j(xo) and set K' = K^ then k'x0 has the distribution 0£o for ¥
uniformly distributed in K'. So

(9.5)
J B(O)
a(g, x) dm(x) = I I
JRJB(G)
a(g, ky) d6h(y)dk

= \ [ *(«*» y) deh(y)dk.
J K J B(O)

Now

) <r(g, y) d0zo(y) = I cr(g, y) dd^(y)

(9.6) = f Il^^^^dB^iy)

= ** (g, So) J dg-%z0(y) = <r/ (g, fo).

By (9.5),

J B(O)
o-(g, a) dm(x) = I o>' (g&, £0) dk=
JK "K
I <r<i,(gk, x0) dk

= I o>(g, £*0) d i = I <r*(g, *) dm(#) = <f>(g).

J
K JB(G)

This proves that a belongs to <f>. But (9.6) also shows that o>(g, x0)
— <*l (g? £0) is expressed as a linear combination of functions in V$.
Since the <r<t> are extremals this can only take place if all the cr(g, y)
for yGj-^fêo) coincide with <r0(g, # 0 ). In other words, dg^dg^/dd^
must be identically 1, or, equivalently, the measure g""1^^ coincides
with 0£o. The same argument implies that for all g £ G , £G^>', g0£ = 0ff£.
But -B(G) is a boundary (see §3) so that some gn0$ tends to a point
measure. Setting gn — knyni where 7 n £ = £ and knÇ£K, we must have
7n0£ = 0<yns = 0$ tending to a point measure. This means that 0$ is a
point measure and so j is 1-1. We have thereby proved that ö>(g, #)
separates points in B(G).
It is now an easy matter to show that a^ is basic. For, since the
functions o>(«, x) are distinct for distinct xÇzB(G), it follows that
B{G) coincides with the normalized extremals of V4,. Now suppose
JB(Q)cr<t>(g, x) dœ(x) §^0 for some signed measure o> on B(G). From the
fact that V<t, is a lattice, by the corollary to Theorem 6.1, this inequal-
310 HARRY FURSTENBERG [March

ity implies that co^O. This, however, was just the condition for 0>
to be basic. This completes the proof of the theorem.
COROLLARY. A multiplier s is basic if and only if it is equivalent to
a> for some spherical function cf>.
Very little is known regarding the subset Hh(B(G)) of the linear
space H(B(G)). For example, is it a semigroup, and is it closed under
the operations a—><7r, r ^ 1?
10. Application to the Choquet-Deny representation. We shall
apply the results of §9 to obtain more precise information regarding
the extremals of V\(\f/).
THEOREM 10.1. If s(«, x) is an extremal of V\(\p), then the multiplier
s is basic.
PROOF. As in the proof of Theorem 9.1, we form the space B'
which is obtained from B{G) by identifying xu x2GB(G) if s(g, xî)
= s(g, x 2 ). We have an equivariant map j : B(G)—>B', and s(g, x)
==
s'(g, j(%)), where sfÇiM(Bf). Again, it will suffice to prove that
B'=B(G). For, V\(\p) is a lattice, and so by the corollary to Theorem
6.1, if jfe'fe, Ö dco(£) ^ 0 , then co^O.
Let cr' be the i£-multiplier equivalent to s' on B\ The functions
<r'(«, £) are semi-spherical and belong to V^ for some <£. So, by Theo-
rem 7.1,

(10.1) </(g,Q = f <r*(g,x)d0t(x),

where o> is the basic multiplier belonging to <£. Because a> is basic,
the measures 0$ in (10.1) are uniquely determined. Now use the
multiplier equation :

<r'(yg, Ö = J <r*(yg, x) d6t(x)

= J <r*(y, gx)*<t>(g, *) dOfc) = J 0^(7, x)<r+(g, f%x) dgO^x).

But <r'(yg, 0 is also equal to

Ay, giW(g, Q = ƒ*•(?, *V(g, 0 <»,«(*).

1965Î TRANSLATION-INVARIANT CONES OF FUNCTIONS 311

Comparing the two expressions as functions of 7, we have

(10.2) <rt(g, r 1 * ) dtffc) = </(g, Q Mgi{x).
Now we can write

PKZ)

where p' is a positive continuous function on J5'. Since s''(•»£) G V\{yp),

(10.3) ƒ p'(gZW(g, Ö#Cg) <fe = X#'(Q.

Form the convex set Q of all functions p(x), continuous and non-
negative on 5(G), and satisfying

(10.4) f P(x) d8t(x) - ?(Q.

If we recall the proof of Theorem 8.1 we see that the maps £—»0$ of
(8.15) and (10.1) are the same. As a result 0$ has its support on j -1 (£).
It follows that Q is nonempty since the function p = p' oj~l will be-
long to Q.
Define a linear transformation T on the space of continuous func-
tions on B(G) by
Tp{x) = X-1 ƒ p(gx)<n(g, x)t(g) dg.
Then

X-1 ƒ ƒP(g*)<r*(g> *)4>(d igdBfa) = X-1 ƒ ƒ * ( * ) * • & r**) WfcMi) dg.

By (10.2), we have

(10.5) ƒ 2>(*) <»«(*) = X-1 ƒ ƒ ƒ> (*)</(& Ö dBoi{x)Hg) dg.

If pGQi this equals

*-lf?(gQ*'(g,Q*(g)dg-t'(Q.
It follows that r takes Q into itself.
The functions in TQ form an equicontinuous family. For
312 HARRY FURSTENBERG [March

| Tp(kx) - Tp(x) |

=S x-1 ƒ p(gx)<r*(g, *)( I Hk-'g) - *(*) I) dg.

(10.6) = x_x ƒ ƒ p(kigx)(T^g) x) ( i ^-i£lg) _ ^ ^ I) dgdkl

J o
-1
where A = X max o>(g, x) and the maximum is taken over all
x£J3(G) and all g for which some k'g is in the support of xp, fe'£K.
Now, by (10.4),

fp(y)dfn(y) = fp'(£)drn(i;)

so t h a t the right-hand side of (10.6) goes to 0 as kx—>x independently

oîpGQ.
As a result, TQ is a compact convex set and T maps TQ into itself.
It follows that T has a fixed point in Q; that is, there exists a function
p(x) satisfying (10.4) and

(1
°- 7 ) ƒ ƒ ^* )<r * ( S> *>*<«> ^ = X^-
If we set fx(g) =p(gx)(r<t>(g, x), we then find t h a t / z £ ^xOA). Now

J /x(g) <»*(*) = J Pittig, x) ddiix) = J pWatig, fH) dgd^x).

By (10.2), this gives

ƒ Mg) dek{x) = fp(xW(g,0 ddttt(x) = cr\g^)p\gk) = *W(&0,

since pÇzQ> This expresses $'(•, £) as a linear combination of the

functions fx which are in V\(\f/). However, the s'(-, £) were taken to
be extremals of V\(\}/). It must therefore be the case that all the fx
with x in the support of 0$ are proportional. If j were not 1-1, that
is, j(xi)=j(x2) for some pair xi, X2&B(G)y then, since fXl and fXi are
proportional,

(10.8) — — <T4>(g, xt) = — — o>(g, * 2 ) .

£(*i) ^(^2)
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 313

Replace g by kg and integrate both sides of (10.8) over K. This gives

g»(g, # i ) _ o>(g, x2)
p(%i) p(oc2)
Setting g = e we find p(xi)=p(x2) and <?$(•, Xi) =<r<t>(-, x2). Since <r<f,
is basic this cannot happen unless Xi = x2. This proves that j is 1-1 and
hence that B' = B(G). This completes the proof of the theorem.
As a result, the statement of Theorem 8.1 may be made more pre-
cise. T h a t is, the normalized extremals of V\(\f/)t which are already
known to be multiplier functions, are now known to be basic multi-
plier functions. We now give a more precise form to the Choquet-Deny
representation for V\(\f/). For each cohomology class rj£:H(B(G)), we
denote by <r„ the unique i£-multiplier in that class. <rn belongs to a
spherical function which we shall denote <£„.
LEMMA 10.1. For each cohomology class r]ÇzH(B(G)), there exists a
uniquely determined positive constant A(rj) and a positive continuous
function pv(x) on B(G), unique up to a constant multiple, such that

(10.9) fpM',(g, x)t(g) dg = A(v)pn(x).

PROOF. Let Q denote the convex set of non-negative functions p(x)

on B(G) satisfying fp(x) dm(x) = 1. Let T denote the transformation
of Q into itself defined by

Tp(x) = c{p) ƒ p(gx)av(g, x)t(g) dg,

where c{p) is a constant chosen so that TpÇzQ- One verifies that the
functions of TQ form an equicontinuous family so that TQ is a com-
pact convex set. If pv is the fixed point of T in TQ then there is a
constant A(rj) so that (10.9) is valid.
The uniqueness of A(rj) follows from

J P(gx)<rn(g, x)\P(g) dg
X p(x)
(10.10)
A' p'(x)
J P'(gx)<Tv(g, x)Hg) dg

If X/X'<1 we choose x so that p/pf attains its minimum at x; then

(10.10) becomes contradictory. So X^X' and similarly X'^X. Thus
there is a unique eigenvalue for which (10.9) can have a solution. Now
if p„ satisfies (10.9), then the multiplier functions
314 HARRY FURSTENBERG [March

—T-r- *i(fc *)
Pui*)
belong to FAOOO/O, and, by Lemma 8.1, pn is thereby uniquely deter-
mined. This completes the proof of the lemma.
If we compare the multiplier functions on G with the exponentials
e°' on the real line, then we see that A(rj) corresponds to the Laplace
transform of the functions $(g).
We can now formulate the Choquet-Deny representation as fol-
lows:
T H E O R E M 10.2. Restrict the function A(r}) to the subset Hb(B(G))
CH(B(G)), so that for each \ > 0 , A^Çk) CHb(B(G)). Then, f or each
function f <E V\(\(/)t there exists a measure œ on A""1^) XB(G) such that

(10.11) f(g) = f p,(gx)<r,(g, *) Mv, x).

In considering A" 1 (X) X-B(G) as a measure space, we have imbedded

it in Fx(^) and taken the induced Borel structure. We do not as yet
know enough about the function A(rj) to claim that A_1(X) is a Borel
subset of H(B{G)).
What has been gained by our analysis as regards the integral equa-
tion
ff(g'gMg')dg' = \f(g)
JG

is this. We have replaced a single equation for a function on a non-

compact space and which, moreover, does not have a unique solu-
tion, by a family of equations (10.9) for a function A (rj) and a family
of functions pv(x) on the compact space B(G) which now have a unique
solutions.
Theorem 10.2 still leaves open the uniqueness of the measure o>.
Although we know that the representation in terms of normalized
extremals of V\(\p) is unique, we have not shown that every multiplier
function corresponding to a point in A""1(X)X-B(G) is an extremal.
This we shall do in the next theorem.
T H E O R E M 10.3. The normalized extremals of V\(yp) consist exactly of
those multiplier functions s(-, x), xG-B(G), such that s is basic, and
$(•, x) belongs to V\(\f/).
PROOF. We have already shown in Theorem 10.1 that if s(*, x) is
extremal in V\(yp) then s is basic. We now show the converse. Suppose
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 315

then that s is basic and s(-t # ) £ V x ( ^ ) . By Theorem 10.2 there is

some measure co with

(10.12) s(g,x0)= I Pi(£x)<ri(g,x)d<o(ri9x).

Here we may assume t h a t the multiplier functions

(10.13) sv(g, x) = -1— <rn(g} x)

ƒ>,(*)
are extremals of V\(\f/) for those rj occurring in this representation. If
s~<r where aGMK(B(G)), then, by (2.3),

J s(kg, x) dk = ( J s(k, x) dk)<r(g, x) = S(x)a(g, x).

By (10.12)

S(xo)<r(g, x0) = I (I p,(y) dm(y) ) cr,(g, x) dœ(Vy %)

J A~ 1 (X)X5(G)\ J /

or

(10.14) <r(g, XQ) = I cr,(g, x) da>'(v, x).

• / A _ 1 (X)XB(G)

Setting g = e, we see that co' is a probability measure. Replacing g by

gk and integrating both sides of (10.14) over K we find

'A^CX)

Here 0 is the spherical function to which a belongs. Now replace g by

gkg and integrate over K. We have, by (3.2),

(10.15)

Now œ" is again a probability measure, so (10.15) shows that $„ is

essentially constant as a function of rj. This means that co" is concen-
trated on classes 77 with the property that 0 , = 0 . All we need, how-
ever, is the fact that for some 77, <£, = $. We have 5 basic and cr~s. It
follows that <7 is basic; so tr = oy for some spherical function <£', by
316 HARRY FURSTENBERG [March

Theorem 9.1. Since a^ belongs to 0' and a belongs to </>, we must have
<j>' =<t> and so a~a^. Since the rj occurring in (10.12) are basic classes,
av is basic and so <Tv = <r<},i} = <r4> = G. As a result, s ^ o v But ff^s,
((10.13)), so ss^s,,. NOW the multiplier functions s(-, x) and s„(-, x)
are in V\(\p) and, by Lemma 8.1, since s~sv, we must have 5 = 5,.
Since the functions 5,(-, x) are extremal, it follows that the functions
5(-, x) are extremal and this proves our theorem.
This theorem implies that the normalized extremals of V\(\p) cor-
respond exactly to the points of A_1(X) XB(G), the correspondence
being given by

Since the representation of elements in V\(\p) in terms of measures

on the normalized extremals is unique, we have:
COROLLARY. For each function ƒ'£ V\(\p), the measure œ in (10.11) is
unique.
The eigenvalue function A (77) was defined originally for all
rjÇzH(B(G)). It is not an arbitrary function; in particular, it is deter-
mined in all of H(B(G)) by its values in Hh{B(G)). This follows from
the fact that the Choquet-Deny representation depends only on the
restriction of A(77) to Hb(B(G)). However, we can make a more precise
statement.
D E F I N I T I O N 10.1. We shall say that two classes 771 and 772 in H(B(G))
are related if (j>vl = 0, 2 .
Thus every class is related to one and only one basic class.
THEOREM 10.4. If 771 and 772 are related, then Afai) =A(rj2).
PROOF. Suppose 5 is a multiplier in the class 771 such that 5(-, x)
GFxOA). Then X=Afoi). By Theorem 10.2,

(10.16) s(g,Xo)= I Pi(gx)*i(g,x)d<u(ii,x).

Replacing g by k\gk2 and integrating both sides of (10.16) over KXK

we obtain, as in the proof of Theorem 10.3,

(10.17) # fl (g) = f *,(g) *»"(*).

As in the proof of Theorem 10.3, it follows that 77 is constant in the

1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 317

integration (10.17), and 0 n = #,. Since 7?GA_1(X), we have A=A(?7i)

= A(rj). Now 77 is the uniquely determined basic class related to rji
since <t>r,l = 4>v and rj is basic. The same reasoning then implies A (772)
= A(rj)y so A(7/i) =A(T/ 2 ). This completes the proof of the theorem.
11. Irreducible multipliers. If sEM(B(G)), let V(s) denote the
closed cone of functions on G generated by s(-, x), xÇ:B(G). Here
closure will mean with respect to the topology of pointwise conver-
gence on G. V(s) is identical with the functions

Mg) = I s(g, x) dœ(x),

J B(O)

where co ranges over the positive measures on B(G). That this family
is closed follows from two facts. First, if fu>{e) is bounded, then the
measures co are bounded and lie in a compact set. Secondly, the map
co—^/c is continuous with respect to the weak topology on positive
measures.
The cones V(s) are obviously closed under (right) translation.
We may inquire for which 5 the cone V(s) will be irreducible, that is,
have no proper closed, translation-invariant subcone. For example, if
v is a smooth measure on B(G), the multiplier s(g, x) = (dg~lv/dv){x)
is not irreducible. For fs(g, x) dv(x) =v(B(G)) is constant, and if V(s)
contains a constant it cannot be irreducible unless it reduces to the
cone of positive constants.
D E F I N I T I O N 11.1. We say sÇ:M(B(G)) is irreducible if the cone V(s)
contains no proper closed translation-invariant subcone.
Let m denote the X-invariant probability measure on B(G). We
shall use <r0 to denote the 2£-multiplier defined by

(11.1) <T0(g,x) = - f — ( * ) •
dm
LEMMA 11.1. If s* is a basic multiplier, then s = ao/s* is irreducible.
PROOF. A typical function in V(s) isf(g) =fs(g, x) dco(x). To prove
that 5 is irreducible, we must show that given any ƒ G V(s), the func-
tions s( •, x) can be retrieved from ƒ by taking limits of positive linear
combinations of translates. We first form

(11.2) F(g) = ff(gk)dk= f f s(g, kx)s(k, x) dœ(x) dk.

Now, fixing tfoE-B(G), let x = k'x^ Then

318 HARRY FURSTENBERG [March

J K
s(g, kx)s(k, x) dk = I s(g, kk'xo)s(kk', Xo) dk
s{k,X^JK
1 C 1
(11.3) - — I s(g, kx0)s(k, xo) dk = — -F0(g).
S(k , Xo) J K S{k , XQ)
From (11.2) and (11.3) we see that F(g) is proportional to F0(g),
where the latter function is independent of co, and thus independent
of the function ƒ in V(s). It follows that the function F0(g) belongs to
every closed translation-invariant subcone of V(s). F0(g) may be re-
written as

(11.4) Fo(g)= f s(g,x)h(x)dm(x),

J B{Q)

where h(x) is a positive continuous function on B(G).

Recall Lemma 9.3 according to which s* is basic only if the func-
tions of the form fs*(g, x) djji(g), /x^O, are dense in the positive con-
tinuous functions on B(G). In particular, if U is any neighborhood of
a point x0ÇzB(G) and e > 0 , there exists a measure /x on G with

N(x) - ƒ s*(g, x) dn(g) <e for x Ç U,

(11.5)
f N(x) dmix) = 1.
'B(G)
J R(G\
Then

Foigy1) = I s(g, yl»)s(y"\ x)h(x) dm(x)

J B(G)

= s
I (gi x)s(y"19 yx)h(yx) dylm{x)
J B(Q)

= I s(g,x)s(y7x)-1h(yx)ao(y9x)dm(x)
J B(.G)

= s
I (g> x)s*(7} x)h(yx) dm(x).
J 'B(G)
B(G1

So

I Foigy-1) dfji(y) = I s(g, x)\ I s*(y, x)h(yx) dy.(y) dm(x)

J G J B(G) LJQ J
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 319

Since SGS*(J1 X) dii(y) approximates a ô-function in x with respect to

the measure m, and since h{x) is bounded from above and below, it
follows that
x)h(yx) dfi(y)
J Q

approximates a multiple of the same ô-function. More precisely, we

see from (11.5) that

ƒ Fo(gy~l) dfx(y)
s(g, x0) g max CI s(g, x) — s(g, x0) | ,
xeu

1
1
Foiy- ) dix{y)

where C is a constant depending upon h(x). Consequently, with an

appropriate choice of measures /*»,

1
ƒ Foigy- ) d»n(y)
>*(g,#o), as **-»«>,

ƒ
1
^OCT" ) d»n(y)

for each g. It follows that s(-, x0) belongs to every closed translation-
invariant cone containing F0(g) and so V(s) is irreducible. This com-
pletes the proof of the lemma.
We turn now to K-multipliers.
LEMMA 11.2. For each spherical f unction <j> there is one and only one
K-multiplier belonging to <t> which is irreducible.
PROOF. T h a t at least one i£-multiplier belonging to <j> is irreducible
is easily established. If <j>{g) is spherical, then 4>{g) = <t>(g~~l) is also
spherical. Suppose <r* is the basic multiplier belonging to $, and form
a = <To/a*. Then
dgm
J '
'B(G)
<r(g, x) dm(x) = ^B(G)
I
C r
U> g~x%)
a(g, g~xx) —— (x)dm(x)
BCG) J B(G) 1 dm

-X B(G) cr(r\X)

**(fTl> x) dm{x)
-ƒ. ' B(G)
320 HARRY FURSTENBERG [March

So cr belongs to <j> and is irreducible by the foregoing lemma.

Now suppose that <T\ and <r2 are related iC-multipliers (Definition
10.1) which are both irreducible. Since <j>(g) = fB(G)<Ti(g, x) dm(x)1 we
have <£G^(cri)fW(<72). If both cones are irreducible, we must have
V(<Ti)= V(a2). In particular, we have

^ ( g , »o) = J <n(g, x) dœi(x),

(11.6)
i(g, *Q) = I <r*(g, x) dœ2(x)

for two positive measures coi, co2 on B(G). One can find measures
Mi> M2^0 on X with CO4 = M» * x0, and (11.6) becomes

*(g>
<T2l, *o) = I <ri(gk, xo) dm(k),
JK

i(g,
<rik Xo) = I o-2(g*, Xo) dfx2(k).
JK

Combining these two:

(11.7) (7i(g, xo) = I <ri(gk, x0) dfi2 * m(k).

JK
(n)
Let M denote the w-fold convolution of ju2 * ^ with itself. It is not
hard to see that n~1[/jt(1)+ • • • +ju ( n ) ] converges, as n—>oo, to the
Haar measure of the least subgroup Ki(ZK which contains the sup-
port of M2 * Mi [6> pp. 343-344]. Since (11.7) would be valid with M(n)
replacing M2 * Mi» it follows that one may write

<ri(g, Xo) = I (7l(g&, S 0 ) <*&.

Hence, <J\(gk, Xo)=<ri(g, x0) for every kÇ^K\. If P 2 and P\ denote,

respectively, the supports of M2 and jm, then P2P1C.K1. In particular,
Pi is contained in some left coset kiKi. Then

0-2(2, * o) = J I (Tiigkik, xo) dkr^iik).

KX

But the support of fe^Vi is in K\ and the kÇ:Ki leave cri invariant. So
<T2(g, *o) = <Ti(gftl, Xo) = 0-i(g, * i * 0 ) = Ci(g, ffi).

But we have seen in §7, using Lemma 7.2, that this implies GI = <T2.
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 321

This completes the proof of the lemma.

We can now prove the main result of this section.
THEOREM 11.1. A multiplier s is irreducible if and only if <xo/s is
basic.
PROOF. The implication in one direction is given in Lemma 11.1.
What remains to be shown is that if 5 is irreducible, then <T0/S is
basic. We claim that if 5 is irreducible and a is the ^-multiplier equiv-
alent to 5, then a is irreducible. To see this, consider the operator
with domain V(s) defined by

Tf(g)= ff(kg)dk = mK*f(g).

J
K

Since T commutes with right translation, it follows that TV(s) is a

translation-invariant cone. Moreover,

Ts(g, x) = J s(kg, x) dk = a(g, x)\ j s(k, x) dk

so that cr(-, X)ÇZTV(S). It is also easy to see that TV(s) is closed, so

it follows that TV(s)~Z)V(o). Now, since V(s) is irreducible, it follows
that TV(s) is irreducible, which implies that TV(s) = V{o), and that
V(<r) is irreducible, whence <r is irreducible. Let a belong to c/>. By
the foregoing lemma, a is the only irreducible multiplier belonging to
<f>. On the other hand, we know that if o-* is the basic multiplier be-
longing to <t>(g~l)i then (To/cr* is irreducible and belongs to 4>(g). So we
must have ÖT = (TO/Ö'*. In other words, if a is irreducible, <T0/(T is basic.
But then <JQ/S is basic, for <j~s implies a0/<T~<To/s. This completes
the proof of the theorem.
COROLLARY. If si and s2 are equivalent multipliers in M{B(G)), then
S\ is irreducible if and only if s^ is irreducible.
If we consider the exponentials in the real line we notice that they
obviously have the irreducibility property—the translation-invariant
cones they generate are irreducible (in fact, 1-dimensional). Thus, for
the real line, the properties of irreducibility and forming the extrem-
als in the Choquet-Deny representation are shared by the same class
of functions. In the case of semi-simple groups these properties split:
one characterizes the irreducible multiplier functions and the other
the basic multiplier functions. (It is probably true that the only multi-
pliers which are simultaneously basic and irreducible are those equiv-
alent to <rl/2.)
322 HARRY FURSTENBERG [March

12. Irreducible cones. We have mentioned the fact that the posi-
tive exponentials on the group of reals generate irreducible transla-
tion-invariant cones. In fact this property characterizes the exponen-
tials. More precisely
THEOREM 12.1. If V is a cone of non-negative continuous f unctions on
the reals, closed under translation and closed in the topology of pointwise
convergence, and if V is irreducible (that is, no proper subcone of V has
the same properties), then V consists of the positive multiples of some
exponential f (t) =eat, — <*> <a < oo.
We will not give the proof of this inasmuch as it resembles that of
the analogous result for semi-simple groups which we will give. Let us,
however, sketch an alternative proof. If we know that V has a com-
pact base, then, since the reals form a Tychonofï group and act on V
by translation, V must have a fixed ray, i.e., a 1-dimensional transla-
tion-invariant subcone. This must be all of V and, as is easily seen,
this implies that V consists of multiples of an exponential. The prob-
lem is, therefore, to show that V has a compact base. This may be
done by showing that if V is irreducible, the functions in V are con-
vex, secondly, that, in fact, their logarithms are convex, and, finally,
that their logarithms satisfy a uniform Lipschitz condition.
For an arbitrary topological group one can attempt to characterize
the irreducible (right) translation-invariant cones. In the semi-simple
case, this is most easily done if we assume further that the functions
in the cone are invariant under left translation by elements in a
maximal compact subgroup. Note that if a cone is irreducible (with
respect to right translations) and if one function has this property,
then so do all functions in the cone.
T H E O R E M 12.2. Let V be a cone of continuous non-negative f unctions
on G closed under (right) translation and closed in the topology of point-
wise convergence. Assume V is irreducible, that is, that no proper sub-
cone of V has these properties. If, furthermore, there is a function ƒ £ V
with f(kg)—f(g) for all kÇ^K, then V coincides with V(o) for some
irreducible K-multiplier a.
We remark that since the functions of V may be thought of as func-
tions on K\G, the question is that of characterizing cones of non-
negative functions on a symmetric space, invariant under the isom-
etries of the space, and irreducible with respect to this property.
PROOF. We convert the problem into one involving cones of meas-
ures. Identify F with a subcone of the cone of positive (unbounded)
measures on G and let V denote the closure of V in this space. Sup-
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 323

pose co is a measure on G corresponding to a continuous function of

compact support. Then V * co C V by the translation invariance of
V, and, since irn—»7r weakly implies irn * co—>T * co pointwise as func-
tions, it follows that V * co C V. Suppose that V is not irreducible,
and suppose that WQV is also invariant and weakly closed. Then
W * co C WC\ V so, then, WC\ V is nonempty. By the irreducibility of
V, WC\ V— V or W"D V and, since W is weakly closed, W=V. Hence
V is irreducible.
Let VK denote the set of measures T G V such that T * k = T for each
&£i£. Under the hypotheses of the theorem, it is true, for every
TTÇZV, that k * T = 7T. So the measures in VK have the form mK * ir' * w*.
By Lemma 7.1, measures of this form are commutative when the
product is defined.
Notice that VK is invariant under the operations T—»7r * 7 * rnK,
7 G G . We claim t h a t VK is minimal with respect to this property. Let
W(ZVK be closed and invariant under T—*TT * 7 * mK for all 7 G G .
Form W'= I T T G F : 7T * g * % G W for all g G G } . Clearly I T is right-
invariant, weakly closed, and W'Z)W so it is not empty. So W'= V,
or V*mK=VKCW. Thus W= VK.
Let co be a measure of compact support on G and form VK * co * m^.
This cone is closed, for, if 7rn * co * mK converges, a subsequence of
{irn} must converge weakly (T * co * mx(AiA2) ^7r(Ai)co * mx(A 2 )). We
also have VK * co * % * 7 * % = VK * Y * % * co * % C VK * co * m^.
Also VK * co * niK C.V * mK = Fir, so that VK * co * % is a subcone of
VK invariant under T—>T * 7 * rag;. S O F# * co * THK = F^.
Now VK is a weakly closed cone of positive measures on G, so, by
Theorem 6.1, it is spanned by its extremals. Let To be an extremal of
VK> By what has just been shown TO = TW * co * WIK for some measure
TTUCLVK- But this expresses 7r0 as a linear combination of measures
7To, * g * WxG Fir. Since 7r0 is extremal, these must be proportional for
almost all g in the support of co. By continuity, this will be true for all
g in the support of co. Moreover, if the identity of G is in the support
of co, then T0 will itself be proportional to TU. We then find that
To * g * mK is proportional to T0 for all g in the support of co. Since co
had arbitrarily large support, we must have

(12.1) TTO * g~l * mK = <Kg)iro.

Since Î 0 = % * 7r0 * Wz, we have

TO* g* MK = mK * TO * w^ * g * Wz = w^ * g * Wjr * TTO * MK = Wis: * g * TTO,

so that
324 HARRY FURSTENBERG [March

(12 .2) mK * g~* * 7T0 = «KgW

This, however, is the defining equation (7.1) for semi-spherical func-
tions. That is, <j> must be spherical and TTO corresponds to a function
$ £ 7 * . So

Hg) = { <r*(g, x) &<*)

' B(Q)

for some P ^ O on B(G). But <£(gfe) = $(g) since 7r 0 £ F*, so t h a t

*(g) = f fa*(g,kx)dkdv(x)

J B(G) J K

= I *>(g, *) * » ( * ) = <*>(g).
We have proven that if V is irreducible, some spherical function
<££ F. Now V * co C V if co corresponds to a continuous function with
compact support, and <f> * œ is proportional to <t> if <£ is spherical and
co is radial. Hence, <££ V. Now let cr be the (unique) irreducible K-
multiplier belonging to c/> (Lemma 11.2). Then V{&) contains <j> and,
hence, V(<x)r\ Vis nonempty. It follows that V— V{o). This completes
the proof of the theorem.
REMARK 1. The result would not have been altered if the topology
on the cone V had been taken to be that of uniform convergence on
compact subsets.
REMARK 2. Whenever a group is represented by automorphisms of
some structure, we may speak of an irreducible representation when
the structure is minimal with respect to invariance under the action of
the group. Traditionally, the structure is assumed to be that of a
(topological) linear space; in this section we have taken it to be that
of a closed cone. Another possibility would be to let the structure
be that of a compact convex set. For example, every bounded meas-
urable function on G generates a translation-invariant compact con-
vex set in L°°(G) (with the weak* topology), and we may inquire
when this gives an irreducible representation of G. Using [6] it is
easy to show that, up to isomorphism, the irreducible representations
of a semi-simple G on compact convex sets are just its representations
on the space of probability measures on one of its boundary spaces
(the equivariant images of J5(G)). With this it may be shown that
the irreducible weak* compact convex sets in L°°(G) are just those
which are spanned by a function f(g) satisfying ƒ (gk) =f(g) for all &GX*
and which, as a function on the symmetric space G/K, is harmonic.
1965] TRANSLATION-INVARIANT CONES OF FUNCTIONS 325

APPENDIX. LEMMA 5.2. If w is a measure on G satisfying /7r = x(0 7 r

for t(ET> and % a positive character on T, then w = m% * fx for some posi-
tive measure /x on G.
PROOF. Form the cone V of measures ic on G satisfying tir = x(t)7ri
7T^0. Let h{g) be a non-negative continuous function on G with com-
pact support and such that h(k) > 0 for fe£2£. If, for irÇz V, w(h) = 0,
then fh(tg) dw(g) = 0 for all tÇzT. But the interiors of the supports of
the functions h{tg) cover G since TK = G. This would then imply that
7T = 0, so we see that ir{h) > 0 for all 7r£ V, 7 ^ 0 . We can now apply
Theorem 6.1 to the weakly closed cone V, and we find that each
7r£ V has an integral representation in terms of normalized extremals
of V. We see then that the lemma will be proved once we establish
that the extremals of V are the measures m\ * 7, 7 £ G, up to a multi-
plicative constant.
Let 7r£ V and form the measure 7rA given by

*A(/) - ƒ f(g)C^Tg) d*(g),

where A is a set in 7 \ G and CA denotes its characteristic function.

Now it is readily seen that TA again belongs to V. So if ir is extremal
in V it must be concentrated on a single coset Ty. Then TT * 7~~1 = 7r0
is concentrated on T and satisfies ^x0 = x(07I"o. Define the measure
1
TT\ on 7 by

Ari(0 = X(0 *ro(fl.

If T G T , then

ÜTTQ awo

But (dTTr0/dro)(t) = x( r )> so

dfTri ^TTTI rfrxo , dir\ s

— (0 = — (0 — (0 = X(r-><)x(r) = x(0 = — (0.
tt7To aT7To »7T0 »7To

Hence T7TI = 7TI for all r(ET. So 7Ti is proportional to Haar measure on
T and
droit) = cxO"1) <Ö,

and iTo = cm%. Since every extremal 7r of V has the form 7r0 * 7""1 for
some 7(EG, our lemma is proven.
326 HARRY FURSTENBERG

BIBLIOGRAPHY
1. F . Bruhat, Sur les représentations induites des groupes de Lie, Bull. Soc. M a t h .
France 84 (1956), 97-205.
2. G. Choquet, Le théorème de représentation intégrale dans les ensembles convexes
compacts, Ann. Inst. Fourier ( Grenoble) 10 (1960), 333-344.
3. G. Choquet et J. Deny, Sur l'équation de convolution /*=ju * <r, C. R. Acad. Sci.
Paris 250 (1960), 799-801.
4. N . Dunford and J. T. Schwartz, Linear operators, Interscience, New York,
1958.
5. E. B. Dynkin, Non-negative eigenfunctions of the Laplace-Beltrami operator and
Brownian motion on certain symmetric spaces, Dokl. Akad. Nauk SSSR 141 (1961),
288-291 (Russian); translated as Soviet Math. Dokl. 2 (1961), 1433-1436.
6. H. Furstenberg, A Poisson formula f or semi-simple Lie groups, Ann. of Math.
(2)77(1963), 335-386.
7. I. Gel'fand, Spherical functions on symmetric Riemann spaces, Dokl. Akad.
N a u k SSSR 70 (1950), 5-8. (Russian)
8. S. Helgason, Differential geometry and symmetric spaces, Academic Press, New
York, 1962.
9. L. Loomis, Unique direct integral decompositions on convex sets, Amer. J. M a t h .
84 (1962), 509-526.
10. C. C. Moore, Compactifications of symmetric spaces, Amer. J. M a t h . 86 (1964),
201-218.
11. Séminaire Sophus Lie, Théorie des algèbres de Lie, École Normale Supérieure
1954/1955.
UNIVERSITY OF M I N N E S O T A