University of Groningen Convolution on homogeneous spaces Capelle, Johan IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite from it. Please check the document version below. Document Version Publisher's PDF, also known as Version of record Publication date: 1996 Link to publication in University of Groningen/UMCG research database Citation for published version (APA): Capelle, J. (1996). Convolution on homogeneous spaces. Groningen: s.n. Copyright Other than for strictly personal use, it is not permitted to download or to forward/distribute the text or part of it without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license (like Creative Commons). Take-down policy If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim. Downloaded from the University of Groningen/UMCG research database (Pure): http://www.rug.nl/research/portal. For technical reasons the number of authors shown on this cover page is limited to 10 maximum. Download date: 16-06-2020
41
Embed
Top 100 University | Rijksuniversiteit Groningen - University of … · 2016-03-08 · Chapter V The Homogeneous Space G/N V.1 Set-up In this chapter we consider the following example
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
University of Groningen
Convolution on homogeneous spacesCapelle, Johan
IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite fromit. Please check the document version below.
Document VersionPublisher's PDF, also known as Version of record
Publication date:1996
Link to publication in University of Groningen/UMCG research database
Citation for published version (APA):Capelle, J. (1996). Convolution on homogeneous spaces. Groningen: s.n.
CopyrightOther than for strictly personal use, it is not permitted to download or to forward/distribute the text or part of it without the consent of theauthor(s) and/or copyright holder(s), unless the work is under an open content license (like Creative Commons).
Take-down policyIf you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediatelyand investigate your claim.
Downloaded from the University of Groningen/UMCG research database (Pure): http://www.rug.nl/research/portal. For technical reasons thenumber of authors shown on this cover page is limited to 10 maximum.
In this chapter we consider the following example of a non-symmetric homogeneous space G/H, with
non-compact H. The group G will be a real non-compact connected semi-simple Lie group with finite
centre, with Cartan involution ∆. Let G=KAN be an Iwasawa decomposition for G. We study the
homogeneous space ≈=G/N.
First we describe the N–orbit structure (Theorem V.2.5). Then we determine explicitly the
algebra å of G–invariant operators in ∂æ(≈): it turns out that å is isomorphic to the convolution
algebra of compactly supported distributions on the subgroup MA (Theorem V.3.1 and Corollary).
Using some theory from Chapter IV we then reduce the existence problem of fundamental solutions
for the operators in å to the similar problem on MA, which is a well-studied problem (Theorem
V.4.1). Finally, in Section V.5 we look at decomposition and multiplicity problems for invariant
Hilbert subspaces of ∂æ(≈). The result is that (G,N) is not a Gelfand pair. We then try to obtain a
Gelfand pair by a natural group extension as in Section IV.14, but whether this always leads to a
Gelfand pair remains undecided: we can only reduce this question to another, well-known, open
problem (Proposition V.5.9).
We formally choose a point p∑≈, p=eN. Furthermore, we fix the following notations and recall a
number of well-known facts [82] .
Denote the Lie-algebras of G, K, A and N by g, k, a and n respectively. Let aæ denote the
dual of a. Let Í≤aæ denote the root system associated to the pair (g,a), and for å∑Í let gå denote the
corresponding root space. Let Í+ denote the subset of positive roots associated with the choice of N,
82 We refer to the first chapter of Garth Warner, Harmonic Analysis on Semi-Simple Lie Groups I
(Berlin:Springer-Verlag, 1972).
139
— Chapter V The Homogeneous Space G/N —
so n=å¶∑Í
@+gå. Let ® denote half of the sum over the positive roots, counted with multiplicity, that is
to say that ®=2å¶∑Í
@+(dimgå).å.
Let aæÇ be the complexified dual of a. Its elements are called weights, and they correspond
to the character group of a, through the definition
(V.1.0.a) a~:=e~(H) ~∑aæÇ, and a∑A, H∑a such that a=exp(H)
Choose Haar measures dk, da, dn on K, A, and N, with dk having total measure 1. Then with respect to
the Iwasawa decomposition dg=dk¤a™®da¤dn is a suitable choice of Haar measure. It follows that
G/N possesses an invariant measure. With respect to the coordinates KAë(k,a) éêâ kap∑G/N we
choose µ:=dk¤a™®da. Consequently, the theory developed in Chapter IV applies.
As to other coordinates, let Nä denote the nilpotent group opposite to N, so Nä=∆(N). It
follows directly from the general theory [83] that the set NäMAp is an open and dense subset of G/N
whose complement is a finite union of lower-dimensional submanifolds (so it has measure 0), and that
for a suitable choice of Haar measure dn_ on Nä one has
(V.1.0.b) ªG/N
Ï(x)µ(dx)=ªK*A
Ï(kap)dka™®da
=ªäN
*M*AÏ( n_map)d_ndma™®da Ï∑∂(≈).
Let W denote the Weyl group of the root system. Let M be the centralizer of A in K. Let M* denote
the normalizer of A in K. Then M is obviously normal in M*, and through the adjoint representation
M*/M acts on a. The group M*/M is finite, and through the adjoint representation M*/M is a
manifestation of the Weyl group. For each w∑W we choose an element m*w∑M* such that the
restriction of Adm*w to a corresponds to the action of w on a.
We mention the fact that MA and M*A are closed subgroups of G, equalling respectively the
centralizer and the normalizer of A. The Weyl group W also equals the quotient M*A/MA. Let B
denote the minimal parabolic subgroup B=MAN. Pick m*w∑M*. Then M being contained in B, the
double coset Bm*wB depends only on w. The important Bruhat Lemma [84] states that
(*) G is the disjoint union of the double cosets Bm*wB, for w ranging over the Weyl group.
A consequence of immediate importance to us is the fact that B=MAN is the whole normalizer of N83See, for example, Nolan R. Wallach, Harmonic Analysis on Homogeneous Spaces (New York: Marcel
Dekker, 1973), Section 7.6.84 Garth Warner, Harmonic Analysis on Semi-Simple Lie Groups I, Theorem 1.2.3.1.
140
— V.1 Set-up —
(that B normalizes N is obvious). Indeed, suppose g normalizes N, and write g=b¡m*wb™, b¡,b™∑B,
w∑W. Since B normalizes N, so does m*w. Equivalently, w maps Í+ into itself, so that w is the
identity, i.e., m*w∑M, yielding g∑B.
Since B=MAN is the semi-direct product of MA and N, where N is normal, it follows that
B/N equals MA. The results in Section IV.8 then yield that:
Proposition V.1.1 The G–invariant diffeomorphisms of ≈=G/N are of the form
equipped with this representation. Then Ó≈~ is a Ë(b)–module as well as a B–module.
Define the module
148
— V.3 Convolution Operators in ∂æ(G/N) —
(V.3.2.c) √≈~:=Ë(g)
ˤ(b )Ó≈
~ ≈∑MÀ, ~∑aæÇ.
This type of module is called a generalized Verma module (when M is discrete, so m=(0), it is an
ordinary Verma module)[88] . It is a (Ë(g),B)–module in the usual way, that is, B acts by
b(¨Ë
¤(b)
v)=·Ad(b)(¨)‚Ë
¤(b)
(≈¤a~¤1)bv, b∑B, ¨∑Ë(g), v∑Ó ≈~, cf. Definition III.4.3 and Proposition
III.4.4.
Since the action of N on Ó≈~ is trivial, it is obvious that transversal order 0 elements, i.e.,
those of the form 1Ë
¤(b)
v, v∑Ó≈~, are fixed by N. (In the theory of Verma modules such elements are
usually called highest weight (see the proof of the following lemma), but in the present context of
distributions with transversal derivatives we prefer the expression 'transversal order 0').
Lemma V.3.3 The generalized Verma module √≈~ is irreducible if and only if all
n–annihilated elements are of transversal order 0.
In fact we need only the "only if" direction of this lemma. For the sake of completeness we prove the
lemma in both directions.
Proof Fix ≈ and ~, and set √=√≈~.
Let n_ denote the nilpotent subalgebra of g opposite to n, opposite with respect to the
Cartan involution, so g=n_@m@a@n=n_@b. Let √(+) denote the subspace n_√ of √, and let √(º)
denote the space of transversal order 0 elements in √. Then it is easy to see that √(º) and √( +) are
complementary:
(V.3.3.a) √= √(º) @ √(+) .
For example, it follows from the direct sum decomposition g=n_@b that the canonical map Ë(g)¤Ó≈~
ììâ Ë(g)Ë
¤(b )Ó≈
~ = √ is an isomorphism when restricted to Ë(_n)¤Ó≈
~ (to see this, set ¬=Ë(n_) in
Proposition I.7.5). Accordingly, the decomposition Ë(n_)=Ç.1@ n_Ë( n_) induces the decomposition
(V.3.3.a). The subspace √( +) can also be described as the sum over the weight spaces for the action
of a for weights other than ~.
n_ is an invariant subspace for the adjoint representation of M. And so, for m∑M, X∑_n, v∑√,88 In the theory of Verma modules usually a ®–shift is involved, that is, ~ is replaced by ~+®, where ® is half
the sum of the positive weights (counted with multiplicity). At this point the ®–shift is irrelevant.
149
— Chapter V The Homogeneous Space G/N —
mXv=·Ad(m)X‚(mv)∑ n_√=√( + ), so that √( + ) is M–invariant. Moreover, set b_=n_@m@a, then
[_b,nä]≤nä, so that b_√( +)=_bnä√≤nä√=√( +). In other words,
(V.3.3.b) √(+) is a (Ë(b_),M)– submodule of √.
Now assume that √ is irreducible. Assume that v∑√ is an n–annihilated element, or, equivalently,
N–fixed. Decompose v=vû+v+ in accordance with (V.3.3.a). Since vû is of transversal order 0, it is
n–annihilated. Therefore, so is v+. Since g=b_@n is an AdM–invariant decomposition, the
(Ë(g),M)-submodule of √ generated by an n–annihilated element equals the (Ë(b_),M)–submodule
generated by that element. In view of (V.3.3.b) this implies that the (Ë(g),M)–submodule of √generated by v+ is contained in √( +). And so this submodule cannot be all of √. Since √ is
irreducible it must be (0). This forces v+=0, in other words, v=vû is of transversal order 0. This
proves the "only if" direction of the lemma.
To show the reverse implication in the lemma we demonstrate that for every generalized
Verma module √,
(V.3.3.c) every non-zero (Ë(g),M)–submodule of √ contains a non-zero n–annihilated vector.
To see that this will do, assume, as in the lemma, that all n–annihilated vectors have transversal order
0. Then (V.3.3.c) implies that every non-zero submodule of √ contains a transversal–order–0 vector
1Ë
¤(b)
h, h∑Ó≈~, h≠0. But such a vector generates all of √, as (Ë(g),M)–module. This is because Ó≈
~ is
by definition an irreducible M–module, so the M–module generated by h≠0 is all of Ó≈~, and Ó≈
~
generates all of √=Ë(g)Ë
¤(b )Ó≈
~. In other words, every non-zero (Ë(g),M)–submodule of √contains a vector that generates all of √, so √ is indeed irreducible.
We prove (V.3.3.c) as follows. One way of expressing the triviality of the action of n on
Ó≈~ is through the equation X
ˤ(b)
h=0, for all X∑n, h∑Ó≈
~. This leads to
(V.3.3.d) X(¨Ë
¤(b)
h)=(adX(¨))Ë
¤(b)
h , X∑n, ¨∑Ë(g), h∑Ó≈
~.
The adjoint representation of n on g is nilpotent, that is to say, the derivations adX, X∑n, operating on
g, are nilpotent endomorphisms. This is easy to see, for example by considering the root space
decomposition of g. It follows that the adjoint representation of n on the universal enveloping algebra
of (the complexification of) g is nilpotent on every subspace Ë(m)(g) of elements of order not
exceeding m. Therefore (V.3.3.d) implies that the action of n on √ is nilpotent on every subspace
√(m) of elements of transversal order not exceeding m (cf. Definition I.7.8). Every √(m) being
150
— V.3 Convolution Operators in ∂æ(G/N) —
finite dimensional, and every v∑√ being of finite transversal order, it follows that for every v∑√,
v≠0, the submodule Ë(n)v≤√ is a finite dimensional vector space in which n operates by nilpotent
endomorphisms. Then apply the following theorem: for a Lie-algebra of nilpotent endomorphisms
operating on a finite-dimensional vector space there exists a non-zero vector annihilated by all these
endomorphisms. This theorem is one of the main ingredients in the proof of Engel's Theorem, see
[89] . This then proves (V.3.3.c) !
As to the irreducibility of generalized Verma modules we need only the following global property,
which is known in the literature on the subject [90] :
Lemma V.3.4 For every ≈∑MÀ the set of ~∑aÇæ such that √≈~ is irreducible is an open and
dense set in aæÇ.
Let V be a finite dimensional B–module. Set √=Ë(g)Ë
¤(b)
V, and let W be a finite dimensional
vector space. Then V¤W becomes a B–module by trivial extension of the action of B, that is, by
setting b(v¤w):=(bv)¤w, b∑B, v∑V, w∑W. Moreover, (Ë(g),B)–induction commutes with this
operation, that is,
(V.3.4.a) Ë(g)Ë
¤(b)
“V¤W‘= “Ë(g)Ë
¤(b)
V‘¤W.
As to the set of n–annihilated elements, it is easy to see that
(V.3.4.b) ·Ë(g)Ë
¤(b)
“V¤W‘‚n= “Ë(g)Ë
¤(b)
V‘n¤W
(where the subscript n has the meaning ‘subspace of elements annihilated by n‘). We use the
following consequence:
Lemma V.3.5 When the module √ =Ë(g)Ë
¤(b)
V has the property that the n–annihilated
elements are those of transversal order 0, then so has every module √¤W, for W any finite
dimensional complex vector space W.
89 James E. Humphreys, Introduction to Lie Algebras and Representation Theory (New York: Springer
Verlag, 1972), Sections 3.2 and 3.3.90 See Nolan R. Wallach, Real Reductive Groups II (New York: Academic Press, 1992), Pure and Applied
Mathematics Series Vol. 132-II, appendix to Chapter 11, p. 127, Lemma 11.A.3.2. This result (rather than the
better-known result for the complex case, or for m=(0)) was kindly pointed out to me by Dr. E.P. van den Ban.
151
— Chapter V The Homogeneous Space G/N —
Finally we recall a few well-known facts about the representation theory of compact and abelian
groups.
Consider again the representation ≈∑MÀ, on the space Ó≈. Let (.|.) denote an invariant inner
product on Ó≈, with respect to which ≈ is a unitary representation. Let ˜≈ be the complex span of the
matrix coefficients Mëméêâ·≈mx»y‚, x,y∑Ó≈. Then ˜≈ is a finite dimensional subspace of the
ç°–functions on M, and a closed Hilbert subspace of L™(M), depending only on the equivalence
class of ≈.
Let Ó… …≈ denote the Hilbert space conjugate to Ó≈, so Ó… …≈ is the linear space conjugate to
Ó≈, with inner product (x_|y_)Ó…… ≈
=(y|x)Ó≈, and equipped with the representation ≈_ defined by
≈_mx_=≈î …mx_, m∑M, x∑Ó≈. The span of matrix coefficients ˜≈ can be identified with the tensor product
Ó… –≈¤Ó≈, by setting (y_¤x)(m)=·≈mx»y‚Ó≈, x, y∑Ó≈, m∑M. This isomorphism intertwines the
representation ≈_¤≈ of M*M on Ó……≈¤Ó≈ with L¤R, where L and R are the left and right regular
representations of M in the functions on M. The space Ó… …≈¤Ó≈ has its own tensor product Hilbert
structure, and the identification ˜≈=Ó… …≈¤Ó≈ is isometric up to a factor only. More precisely,
˜≈=d≈“ Ó… …≈¤Ó≈‘, where d≈ equals the dimension of ≈. In these terms the Peter-Weyl Theorem and
the Schur orthogonality relations for the compact group M can be formulated as
(V.3.5.a) L™(M)=≈¶∑MÀ
@d≈“ Ó… …≈¤Ó≈‘
Let Ô≈ be the character of ≈, so
(V.3.5.b) Ô≈(m)= i¶d
=
≈
1 ·≈mei»ei‚ m∑M,
for ·ei‚i=1,Ú,d≈ an orthonormal basis of Ó≈. Then the orthogonal projection L™(M) ììâ ˜≈ is
convolution by d≈Ô≈. So, when seen as a Hilbert subspace of ∂æ(M)=´æ(M), ˜≈ has reproducing
distribution d≈Ô≈. So (V.3.5.a) can be reformulated as a decomposition of the Dirac delta distribution
at the origin of M:
(V.3.5.c) ∂eM=≈¶∑MÀ
d≈Ô≈ in ´æ(M).
All this is well-known [91] .
Since ˜≈ is contained in ´(M), its reproducing operator (i.e., convolution by d≈Ô≈) extends
91 See e.g. Charles F. Dunkl & Donald E. Ramirez, in their Topics in Harmonic Analysis, Appleton-Century
Mathematics Series (New York: Meredith Corporation, 1971), Chapters 7 and 8.
152
— V.3 Convolution Operators in ∂æ(G/N) —
continuously to ´æ(M), with unchanged (because finite dimensional) image, so
(V.3.5.d) ´æ(M) * d≈Ô≈=˜≈ ≈∑MÀ.
Take a weight ~∑aÇæ . Let ˜≈~ denote the space of matrix coefficients associated to the
representation ≈¤a~¤1, defined at (V.3.2.b). Let Óä≈–~ denote Óä≈ equipped with the representation
≈_¤a–~¤1. Then ˜≈~ equals Ó…≈
–~¤Ó≈~, when y_¤x is identified with the map MANëman éêâ
(≈mx»y)Ó≈a~, and under this identification the representation L¤R of B*B on ˜≈
~ corresponds to the
representation “≈_¤a–~¤1‘¤“≈¤a~¤1‘ on Ó…≈–~¤Ó≈
~. Furthermore, since the matrix coefficients that
belong to ˜≈~ are derived from a representation of B=MAN that is trivial on N, they are themselves
right N–invariant, so ˜≈~ can be identified with a left B–invariant submodule of ∂æ(MA). More
precisely, this identification is ˜≈~=˜≈¤“a~‘≤∂æ(MA).
Let Ô≈~ denote the character
(V.3.5.e) Ô≈~(ma)=Ô≈(m).a~
corresponding to i¶d
=
≈
1e_i¤ei, where ·ei‚i=1,Ú,d≈ is an orthonormal basis of Ó≈
~. When ~ is purely
imaginary, ˜≈~ is a bi-invariant Hilbert subspace of ∂æ(MA), with reproducing distribution d≈Ô≈
~.
Moreover, since its image is contained in ´(MA), and is finite dimensional, convolution by d≈Ô≈~
extends continuously from ∂(MA) to ´æ(MA), with identical image, so
(V.3.5.f) ´æ(MA) * d≈Ô≈~=˜≈
~ = Ó…≈–~¤Ó≈
~.This is in fact true for all ≈∑Mfl, ~∑aÇæ
. The difference when ~ is not purely imaginary is merely that in
that case the bi-invariant space ˜≈~ has no invariant Hilbert structure for the action of A.
For S∑´æ(MA) define its Fourier transform SÀ as the map
(V.3.5.g) SÀ: MÀ*aÇæ ììâ ∂æ(MA)
SÀ(≈,~)=S*Ô ≈~ ∑ ˜≈
~ [92]
What we need is the fact that the map S éêâ SÀ is one-to-one. This follows from the Plancherel
formula for the group MA, seen as a decomposition of the Dirac delta distribution at the origin of MA
by means of characters Ô≈~:
92 These are ad hoc definitions. The Fourier transform on M alone, for example, would normally be defined
as a field over Mfl with values in L(Ó≈) [the space of linear operators on Ó≈,equipped with Hilbert-Schmidt
norm]. There is no essential difference, however. The map L(Ó≈)ëA éêâ ·g éêâ tr(A≈g)‚∑˜≈ leads from
one situation to the other. For the Fourier transform on MA, see Section V.4.
153
— Chapter V The Homogeneous Space G/N —
(V.3.5.h) ∂eMA= ≈¶∑MÀ
d≈ ªiaæ
d~ Ô ≈~
for a suitable choice of Lebesgue measure d~ on iaæ.
With these facts available, we can give the
Proof of Theorem V.3.1 Take ≈∑MÀ,~∑aæÇ. Consider the distribution ^*Ô≈~, where Ô≈
~ is defined as
above, at (V.3.5.e). Since this is a zonal distribution, it propagates a convolution operator ´æ(≈) ììâ
∂æ(≈), according to Theorem IV.2.2. Moreover, when U belongs to ´æMAp(≈), then according to
Proposition IV.4.5 supp (U*^*Ô≈~) is contained in supp(U).supp (^*Ô≈
~), which, being the product of
two subsets of MAp, is itself contained in MAp. So,
´æMAp(≈) * ^*Ô≈~ ≤ ∂æMAp(≈).
More explicitly, convolution being G–equivariant, it is Ë(g)–equivariant. So, for ¨∑Ë(g), U∑´æ(Ì),
V∑∂æ(Ì), one obviously has (¨^*U)*(^*Ô≈~ )=¨^*(U*Ô≈
~ ) (compare Proposition IV.9.4.(b)).
Therefore, in terms of the algebraic-topological form for ´æMAp(≈) and ∂æMAp(≈) as given in
Theorem III.4.1, the operator Ù of convolution by ^*Ô≈~ allows the following simple description:
(V.3.5.i) Ù : Ë(g)Ë
¤(b)´æ(MA) ììâ Ë(g)
ˤ(Àb)∂æ(MA)
(¨Ë
¤(b)
S) *(^*Ô≈~ ) = ¨
ˤ(b)
(S*Ô≈~).
For U∑´æMAp(≈) define its transform UÀ, as follows:
(V.3.5.j) UÀ: MÀ*aæÇ ììâ ∂æMAp(≈)
UÀ(≈,~)=U* ^*Ô≈~ .
It is easy to see that UÀ is continuous. The main thing needed for this is the separate continuity of the
ZW 153/81 (Amsterdam: Mathematical Centre, 1981), Theorem 3.2.94 S. Helgason, “Duality and Radon Transform for Symmetric Spaces,” Amer. J. Math.85 (1963), pp. 667-
692. See also his “Invariant Differential Operators and Eigenspace Representations,” Representation Theory
of Lie Groups; London Mathematical Society Lecture Note Series 34 (Cambridge: Cambridge University
Press, 1979), pp. 236-286, see Lemma 3.3 on page 267.95 Jens Carsten Jantzen, Moduln Mit Einem Höchsten Gewicht Lecture Notes in Mathematics Vol. 750
(Berlin:Springer Verlag,, 1979).
156
— V.3 Convolution Operators in ∂æ(G/N) —
Proposition V.3.7 A distribution concentrated on MAp is strongly zonal if and only if it is of
transversal order 0, and derives from a central distribution on MA.
The proof is straightforward, as follows.
Proof The only real problem is to show that when a distribution is concentrated on MAp is strongly
zonal, then it must be of transversal order 0, so of the form ^*S. It is easily seen that a distribution of
the form ^*S, which is already zonal, is strongly zonal if and only if S is central on MA. For this it is
sufficient to consider the commutation relations with respect to other distributions of the form ^*V,
and to use the fact that ^* is an injective homomorphism for the convolution product (Proposition
IV.9.4).
Let N<> be the stability group of p for the extended group G<>:=G*MA, as defined in Section
IV.14. From (IV.14.1.a) one sees that N<>=”(man,ma)»man∑MAN’. So we can and will view the action
of N<> as a diagonal action of B=MAN, which we denote by ÿ. From (IV.14.0.b) it follows that the
explicit form for this diagonal representation of B is
ÿmanU=†manU*∂pm¡a¡ m∑M, a∑A, n∑N, U∑∂æ(≈).
MAp being left MAN– and right MA–invariant, N<> operates on ∂æMAp(≈). In view of (III.4.4.b) its
Theorem V.4.1 Fundamental Solutions on G/N; Reduction to MA.
Let u be a convolution operator ∂æ(G/N) êêâ ∂æ(G/N).
Then u is induced from MA, and
i) u has a fundamental solution in ∂æ(G/N) if and only if its inducing operator uº has a
fundamental solution in ∂æ(MA).
ii) If u has a fundamental solution in ∂æ(G/N), it also has one that is zonal.
iii) If u is a bilaterally invariant convolution operator, then if it has a fundamental solution, it
also has one that is strongly zonal.
Reminders A bilaterally invariant convolution operator on G/N is one that is invariant for the group
G <>=G*MA acting bilaterally. Strongly zonal is invariant for the stability group N<>:=G<>p, so the
minimal parabolic subgroup B=MAN acting diagonally through (man,x) éêâ manxm¡a¡.
Corollary V.4.2 When M is discrete (i.e.: finite), every non-0 invariant differential operator on
G/N has a strongly zonal fundamental solution.
Proof of the Theorem i) and ii) are a rendering of Theorem IV.10.2 in this specific situation, with
Theorem V.3.1 in mind. To prove iii), note that when u is bilaterally invariant invariant, uº will be
bi–invariant (compare Proposition V.3.7). When u has a fundamental solution, uº has one too
(according to (i)). But uº being bi-invariant, and MA being what it is (the direct product of a compact
group and, essentially, a Euclidean space), a fundamental solution of a bi-invariant operator can be
integrated over M to yield a central fundamental solution E (central in ∂æ(MA), that is). But then ^*E is
a strongly zonal fundamental solution for u in ∂æ(G/N) !
Proof of Corollary V.4.2 When u is a differential operator, the propagator of the inducing operator uºis a distribution Uº concentrated in eMA (the unit of MA). M being discrete, this can be seen as a
distribution Uºº concentrated in eA. Let uºº be the operator in ∂æ(A) of convolution by Uºº. Then uººcan be seen as a differential operator with constant coefficients on Ân, and so, according to the
Theorem of Malgrange-Ehrenpreis, it has a fundamental solution in ∂æ(A) [96] . Since M is discrete,
96 Lars Hörmander, The Amalysis of Partial Differential Operators; Part I: Distribution Theory and Fourier
Analysis (Berlin: Springer Verlag, 1983).
160
— V.4 Fundamental Solutions —
A is the connected component of MA, so Eºº can be extended trivially (i.e., by 0) to a fundamental
solution for uº in ∂æ(MA) !
As to other interesting convolution operators, consider difference-differential operators, so those
whose propagators have finite supports. Every (non-0) invariant difference-differential operator uû on
A is known to have a fundamental solution (since A=Ân, see the first paragraph of Section IV.10).
When M is discrete one would expect the same conclusion, but this is not in general true when the
support of the propagator of uû is not contained in A. The following simple argument gives a precise
criterion.
When M is discrete consider a distribution U on MA as a function m éêâ Um on M with
values in the distributions on A, so U=m
¶∑M
∂m¤Um. With these conventions the convolution product
on MA can be expressed as a two-step convolution product:
(E * U)k=m
¶∑M
Em*Um¡k E, U ∑ ´æ(MA), k∑M.
This implies
(V.4.2.a) (E * U)n¡k=m
¶∑M
En¡m*Um¡k k, n ∑M.
For E a distribution on MA let Eõ be the convolution kernel Eõ(n,k)=En¡k on M*M. Then (V.4.2.a) reads
as
(V.4.2.b) (E*U)õ= Eõ*Uõ,
where the right-hand side should be interpreted as a matrix product. By convolution U propagates a
left MA–invariant operator uû on MA, and E will be a fundamental solution for uû if and only if Eõ*Uõ
equals the identity, that is, the element I¤∂eA. This will happen if and only if and only if the operator
propagated by Det(Uõ) has a fundamental solution. Indeed, this condition is necessary, because
Eõ* Uõ=I¤∂eA implies that Det(Eõ)*Det( Uõ)=Det(I¤∂eA)=∂eA. Moreover, let co(Uõ) denotes the
transposed cofactor matrix of Uõ, so that
Uõ*co(Uõ)=co(Uõ) * Uõ =Det(Uõ)*·I¤∂eA‚.
Then if F∑∂æ(A) satisfies F*Det(Uõ)=∂eA, a solution for Eõ * Uõ =I¤∂eA is given by Eõ=F*co(Uõ).
For a map Ï defined on M*M with values in an abelian algebra å define DetÏ in the
obvious way, that is, as the element in å with formula ß∑P¶erm(M) (–1)|ß|
m•∑M
Ï(m,ß(m)). When uûis a convolution operator with propagator U=m
¶∑M∂m¤Um, Um∑´æ(A), let Uõ:M*M ììâ´æ(A) be the
161
— Chapter V The Homogeneous Space G/N —
convolution kernel Uõ(n,m)=Un¡m, n,m∑M. Let Det(uû) denote the operator of convolution by
Det( Uõ).
With these notations one we have the following.
Proposition V.4.3 M discrete; reduction to A
Let M be discrete. Let uû be a left invariant operator in the distributions on MA.Then uû has a
fundamental solution in ∂æ(MA) if and only if Det(uû) has a fundamental solution in ∂æ(A).
In particular, a difference-differential operator on uû on MA has a fundamental solution
if and only if Det(uû)≠0.
Later on we prove a similar result in a concrete context, using another approach (Corollary VI.3.3).
As to the existence of fundamental solutions for differential operators on MA for general M (no longer
discrete), precise criteria (in the form of growth estimates on Fourier coefficients) have been obtained
by Cerezo and Rouvière [97] . Their results being so directly relevant, we give a brief summary in a
way that fits in with the notation as used in the introduction to the proof of Theorem V.3.1.
For the estimates, one first constructs a series of bi-invariant differential operators on M of
increasing order. Being reductive, the Lie algebra m is the sum of its centre and its semi-simple
derived algebra [m,m]. Choose an orthonormal basis of [m,m] (orthonormal with respect its Killing
form), and extend this to a basis X¡,ÚXm of m. Set D:=i¶=
m
1 Xi™, a bi-invariant differential operator on M.
For k∑ˆ set
Dk:=0¯
¶q¯k
(-D)q
Being bi-invariant, Dk operates by scalars on the traces Ô≈, ≈∑Mfl, say DkÔ≈=dk(≈)Ô≈. The dk(≈) are
strictly positive numbers, and increasing in k (for ≈ fixed). They function as a kind of basic
polynomials in ≈, of degree k. For S a continuous function on M its Fourier transform is defined as a
field of Fourier coefficients Sfl(≈) with Sfl(≈)∑¬(Ó≈), the space of linear operators on the representation
space Ó≈, equipped with the Hilbert-Schmidt inner product (A»B)HS=trace(B*A) [98] . The
97 André Cerezo and François Rouvière, “Solution Élémentaire d’un Opérateur Différentiel Linéaire Invariant
a Gauche Sur un Groupe de Lie Réel Compact et Sur un Espace Homogène Réductif Compact,” Ann. Scient.
Éc. Norm. Sup. 4th series, part 2 (1969), pp. 561-581. In the paper the compact group is assumed to be
connected. The results that we need are just as much valid when M is non-connected. See note to page 562 in
the paper.
162
— V.4 Fundamental Solutions —
coefficient Sfl(≈) is obtained by integrating ªMdm. S(m)≈m¡. As remarked at (V.3.5.d) Fourier
transformation extends to the distributions. One shows that a field (a≈)≈∑MÀ is the Fourier transform of
a (uniquely determined) distribution if and only if it is “of no more than polynomial growth”, in the
sense that
‰k∑ˆ ‰C>0 Å≈∑ Mfl »»a≈»»¯Cdk(≈).
It follows from this that a left invariant differential operator u¡ on M has a fundamental solution in
∂æ(M) if and only if the Fourier coefficients of its propagator u¡∂ are invertible, and satisfy
(V.4.3.a) ‰k∑ˆ ‰C>0 Å≈∑ Mfl »»“ u fl¡∂(≈)‘¡»» ¯ Cdk(≈).
That all the Fourier coefficients should be invertible (equivalently, that the differential operator
should be injective) is an obvious condition. In the simplest case of a one-torus (i.e., the circle)
invertibility of the Fourier coefficients implies (V.4.3.a), but already on the two-torus there are
examples of first order differential operators where all the coefficients are invertible (non-zero, that is,
the group being abelian), but where this growth condition is not fulfilled. This is not obvious, and the
clue is an arithmetic subtlety. A first order polynomial like (p–åq+2), p,q∑Û™, with å irrational,
though nowhere vanishing, will still approach 0 on certain sequences in Û™. Moreover, depending on
å, the rate of this approach may on certain sequences be faster than any negative power of 1+p™+q™.
For such å the differential operator that has p–åq+2 for its Fourier polynomial has therefore no
fundamental solution, at least not in the distributions [99].
Next consider compactly supported distributions on the product M*A as compactly
supported distributions on A with values in the distributions on M. It follows that the Fourier
transform of a distribution U∑ æ(MA) can be seen as a function on the dual Afl=aæÇ with values in the
fields in ·¬(Ó≈)‚≈∑Mfl. More precisely, let Uû denote the element in ∂æ(A;∂æ(M)) corresponding to U,
defined by <<Uû,¥>,ƒ>=<U,ƒ¤¥>, ƒ∑∂(M), ¥∑´(A), then flU(≈,~)=< U
û,a–~>À(≈), ≈∑Mfl, ~∑aæÇ . So for
fixed ≈ one has a function Ufl≈ on aæÇ with values in ¬(Ó≈). This is implicit in (V.3.5.f).
When U is the propagator uº.∂eMA of a left invariant differential operator uº on MA, the
maps Ufl≈ are polynomial maps with values in ¬(Ó≈). They are described succinctly by
Ufl≈(~)=(uº(≈¤a~))(eMA), ≈∑ flM, ~∑aæÇ .
For a polynomial map P on aæÇ define the norm ||P||=”(9å¶∑ˆr» îîîÿÿ
åî~îå
P(0)»™0)»
™. Similarly, for a polynomial
98 ¬(Ó≈)= Ó… … ≈¤Ó≈, the form we used at (V.3.5.a).99 André Cerezo and François Rouvière, page 570.
163
— Chapter V The Homogeneous Space G/N —
map P≈ on aæÇ with values in Ó≈ define ||P≈||=”(9å¶∑ˆr»» îîîÿÿ
åî~îå
P≈(0)»»™Ó≈0) »
™. Let DetP≈ denote the
Determinant of P≈, a polynomial on aæÇ . Finally, let coP≈ denote the transposed cofactor matrix of
P≈, so that P≈.coP≈=coP≈.P≈=Det(P≈).Id≈.
Theorem V.4.4 Cerezo-Rouvière
Let uº be a left invariant differential operator on MA, with propagator U. Then uº has a
fundamental solution in ∂æ(MA) if and only if for every ≈∑Mfl the polynomial Det( flU≈) is not
identically 0, and
‰k∑ˆ ‰C>0 Å≈∑ Mfl
»î »î»»î
D
c î
e
oî
t
Uîfl
flUîî≈î
≈îî»»
»
» ¯Cdk(≈).
Finally a somewhat similar result concerning differential operators on the group G. By means of a
(generalized) Radon transform François Rouvière has succeeded in reducing bi–invariant differential
equations on the group G to bi-invariant differential equations on MA, under the assumption that G
have only one conjugacy class of Cartan subalgebras [100] . To compare that result, let z be a bi-
invariant differential operator on G. Being in the centre Z(g) of the universal enveloping algebra, it
acts as a bilaterally invariant differential operator õz on G/N. Proposition V.3.8 implies that õz is
induced from a bi-invariant operator zº in MA. This yields an algebra homomorphism from Z(g) to
Z(m@a). It is actually into: this is because Z(g)§Ë(g)n=(0), so that the map z éìâzº is one-to-one
[101] . The homomorphism thus arising coincides with the map † introduced by Rouvière, defined as
the restriction to Z(g) of the projection onto the first summand in the direct sum decomposition
Ë(g)=Ë(m@a) @ Ë(m@a@_n)_n @ Ë(g)n [102]
Rouvière’s result is that (again, supposing G has only one conjugacy class of Cartan subalgebras) the
existence of a fundamental solution for zº ensures the existence of a central fundamental solution for
z. This result suggest a relationship between the existence of fundamental solutions on G/N and the
existence of fundamental solutions on G, but it is not clear how the actual solutions themselves should
be related.
100 François Rouvière, “Invariant Differential Equations on Certain Semi-Simple Lie Groups” Transactions
of the American Mathematical Society Vol. 243, September 1978, pp. 97-114, Theorem IV.2101 Garth Warner, Harmonic Analysis on Semi-Simple Lie Groups, Lemma 2.3.3.5. The kernel of the map
éêâ ¨∂p is Ë(g)n, see our Example III.4.2.ii.102 François Rouvière, “Invariant Differential Equations on Certain Semi-Simple Lie Groups,” Lemma 2.3.
164
V.5 Plancherel Formulas;
(G,N) a Generalized Gelfand Pair ?
In the preceding sections some questions in Harmonic Analysis on G/N have been reduced to similar
questions on MA. In the present section we induce Hilbert subspaces of ∂æ(G/N) from Hilbert
subspaces of ∂æ(MA). This is based on Sections IV.11-13. We then deal with the question whether the
representation of G in G/N is multiplicity free.
The modular functions introduced in Section IV.11 take the following form. One calculates that the
modular function of the right action of MA on G/N is given by ®µMA(ma)=a™®, m∑M, a∑A (use
(IV.11.0.b), and µ=dk.a™®da). Since MA is unimodular it follows that ®MA=1, and çµ(ma)=a™®,
m∑M, a∑A. According to Section IV.12 a left MA–invariant Hilbert subspace Ó of ∂æ(MA) give rise to
a G–invariant Hilbert subspace Óind of ∂æ(G/N). To obtain an explicit expression for the latter space
one can choose a smooth strictly positive measure on ≈/MA and then calculate the associated
modular function ç~µ, determined by (IV.11.1.a). Since ≈/MA=G/B=K/M an obvious choice is the
uniquely determined K–invariant measure dk$ with total mass 1. It follows from the expression
(V.5.0.a) ª≈Ï(x)dx=ªK/M
dk$ªMA
Ï(kmap)a™®dmda Ï∑∂(≈),
that ç~µ(kap)=a™®, k∑K, a∑A. Denote the function “ç~
µ(x)‘™ by x®, so (kap)®=a®. For a smooth
function on a smooth manifold let µå denote the operator in the distributions of multiplication by å.
The expressions in Theorem IV.12.1 then turn out as follows:
(V.5.0.b) Óind= µx®ª@
K/Mdk$†k^*Ó
=ª@
K/M dk$†k^*µa®Ó
for Ó a left MA–invariant Hilbert subspace of ∂æ(MA). Equally, however, (V.1.0.b) can be used to
derive the expressions
(V.5.0.c) Óind= µ ∫ ª@
Nädn_† n_^*Ó
=ª@
Nädn_† n_^*µa®Ó
where ∫ denotes the function ∫(n_map)=a®.
165
— Chapter V The Homogeneous Space G/N —
Use the Plancherel formula for the group MA, as described in the proof of Theorem V.3.1. That is,
choose Lebesgue measure d~ on iaæ so that
(V.5.0.d) ∂eMA= ≈¶∑MÀ
ªiaæ
d~ d≈ Ô≈~ (compare (V.3.5.h)).
The corresponding direct integral decomposition of L™(MA) in Hilb˚MA*MA(∂æ(MA)) is then
(V.5.0.e) L™(MA)=≈¶∑MÀ
@ ª
@
iaæd~ ˜≈~ ,
where ˜≈~ equals the space ˜≈¤“a~‘ of matrix coefficients of the representation ≈¤a~.
Decomposition (V.5.0.e) is a direct integral decomposition by means of minimal bi-invariant Hilbert
subspaces of ∂æ(MA). According to Corollary IV.13.7 this implies that
(V.5.0.f) L™(≈;µ)=≈¶∑MÀ
@ d≈ ª
@
iaæd~ “˜≈
~‘ind
is a direct integral decomposition of L™(≈;µ) by means of bilaterally invariant Hilbert subspaces of
∂æ(G/N).
There are a number of ways of obtaining a more explicit description of “˜≈~‘ind. One way
is to work out the integral (IV.11.2.a). We use the following argument, which requires no
calculations, and has some use in Chapter VII.
The reproducing distribution of “˜≈~‘ind is ^*d≈Ô≈
~+®, and this is a strongly zonal
distribution (Definition IV.13.4). According to Proposition IV.14.2 a strongly zonal distribution is
one that is invariant for N<>, the stability group of p in the extended group G<>=G*MA. According to
Proposition IV.14.1 this means that the convolution operator that ^*d≈Ô≈~+® propagates can be
determined by considering G/N as the quotient G<>/N<>. But by the same token, the convolution
product can be determined by considering G/N as a homogeneous space for any closed subgroup
G¡<G<>, as long as this group operates transitively on ≈=G/N. A particularly suitable subgroup is
G¡:=K*A≤G*MA (so, with A operating from the right). Its action on G/N allows a very simple
expression in the KAp coordinates. That is,
(V.5.0.g) (k,a)(kºaºp)=k(kºaº)pa¡ = (kkº)(aºa¡)p.
The result is that the bilaterally invariant convolution operator propagated by a strongly zonal
distribution can be calculated as if G/N were the group K*A. Explicitly, on the space of strongly
zonal distributions the left translation †ka equals Lk¤La, k∑K, a∑A, with Lk and La denoting the left
pairings that the G–representation induced by ≈¤a~ can only be reducible (though it may well happen
that it is not) when there exists a w∑W, w≠1, such that the representation ≈¤a~ is equivalent to
w(≈¤a~) [104] . This can happen only for (≈,~) on a set of measure 0 in Mfl¤iaæ (measure 0 in terms
of the product of counting measure and Lebesgue measure). And so (V.5.0.f) and (V.5.1.f) provide a
way of giving a direct integral decomposition of L™(≈;µ) in HilbG (∂æ(≈)):
(V.5.1.g) L™(≈;µ)=≈¶∑MÀ
@ d≈ ª
@
iaæd~k
d¶=
≈
1@
“˜(≈k)‘ind¤“a~+®da‘
with almost all “(≈k)‘ind¤“a~+®da‘ irreducible. This result (in various shapes) has been derived by
N. Wallach, by G. van Dijk, and by Mannes Poel [ 105] . It can be seen as a refinement of the
bilateral decomposition (V.5.0.f), the refinement taking place after restriction of a representation to a
smaller group. Only decomposition (V.5.0.f) is canonical (see Theorem V.5.3).
One sees immediately that when M is not abelian, some of its representations are of
dimension more than one. But this implies that the representation of G is not multiplicity free, the
“˜(≈k)‘ind¤“a~+®da‘ being equivalent when k varies, and ≈ and ~ are fixed. This is in keeping with
Theorem V.3.1. That theorem (or rather its corollary) implies that when M is non-abelian, the algebra
å of G–invariant operators in the distributions is non-abelian as well. For (G,N) to be a Generalized
Gelfand Pair å would have had to be abelian, essentially because in a multiplicity free situation åoperates by scalars on every irreducible Hilbert subspace (Corollary IV.14.10)
However, even when there is no such obvious obstruction, so when M is abelian, there is
multiplicity still. This is because of a second result in Bruhat theory: when ≈¡¤a~¡ is equivalent to
104 F. Bruhat, “Sur les Représentations Induites des Groupes de Lie,” Bull. Soc. Math. France 84 (1956), pp.
97-205.105 G. van Dijk, “A Plancherel Formula for the Isotropic Cone” Proceedings of the Koninklijke Nederlandse
Akademie van Wetenschappen Series A, Vol . 91 (March 28, 1988) no. 1. This paper in fact concerns a
generalization of a Plancherel formula for the cone, which is G/MN rather than G/N, but the paper includes a
formula for the latter space.
Nolan R. Wallach, Real Reductive Groups II (New York: Academic Press, 1992), Pure and Applied
Mathematics Series Vol. 132-II, Theorem 15.1.3.
Mannes Poel, unpublished work..
169
— Chapter V The Homogeneous Space G/N —
w(≈™¤a~™) for some w∑W, the representations induced by ≈¡¤a~¡ and w(≈™¤a~™) respectively are
unitarily equivalent. As Wallach points out, this means that at the regular points the multiplicity in
(V.5.1.g) is described by (≈,~) éêâ d≈ »W», with »W» denoting the order of the Weyl group.
The result is that
Theorem V.5.2
(G,N) is not a Generalized Gelfand Pair.
Formula (V.5.1.g) is nevertheless a Plancherel formula, in the sense that the integrand is reducible
only on a set of measure 0.
As such there is nothing remarkable about this. As argued in Section IV.14 similar things happen
when one considers the left regular representation in the distributions of a non-abelian unimodular
group. The noticeable thing is rather that when M is abelian one has here an example of a pair (G,H),
H=N, that fails to be a Generalized Gelfand Pair in spite of the fact that the algebra of invariant
operators in the distributions is abelian. (Note that N is not compact. Otherwise this phenomenon
could not occur, because when H is compact the commutativeness of å guarantees that (G,H) is a
Generalized Gelfand Pair. This is shown in [106]).
One way of dealing with the non-uniqueness of decomposition (V.5.1.g) is to group
equivalent representations together, and to integrate over a suitable set of representatives of the
W–orbits in MÀ*iaæ. This is done by Wallach [107]. We prefer the approach of M. Poel, that is, to
consider the bilateral action of G*Ì, with Ì the group of G–invariant diffeomorphisms, Ì=MA. That
this is a ‘natural’ approach has been argued in Section IV.14. In this particular case this leads to the
following:
Theorem V.5.3 The direct integral decomposition in Hilb˚G*MA(∂æ(G/N)):
(V.5.3.a) L™(G/N;dx)=≈¶∑MÀ
@ d≈ ª
@
iaæd~ “˜≈
~‘ind
is the unique, multiplicity free, bilateral Plancherel decomposition.
106 E.G.F. Thomas, “ An Infinitesimal Characterization of Gelfand Pairs,” Contemporary Mathematics Vol
26: Conference in Modern Analysis and Probability, (Providence, Rhode Island: American Mathematical
Society, 1984), pp. 379-385.107 Nolan R. Wallach, Real Reductive Groups II, Section 15.1.4.
with anti-duality given by the G–invariant sesquilinear pairing
<S¤a~+®da»ƒ¤a~+®da>=<S»ƒ>, S∑´æ(K)≈, ƒ∑´(K)≈.
The G–invariant pairing between ´(K)≈¤“a~+®da‘ and ´æ(K)≈¤“a~+®da‘ is a particular case of a
general sesquilinear pairing occurring in the context of induction of (in this case finite dimensional)
unitary representations (for more details, see [111] ). Furthermore, when Ó is a unitarily induced
representation, then Ó° is contained in the space of ç°–functions in Ó, so in ´(K)≈¤“a~+®da‘, and
it is actually equal to ´(K)≈¤“a~+®da‘ because K/M is compact (this is based on a general theorem
by Poulsen [112] ). One may note that ^*Ô≈~+®, by our definition the reproducing distribution of
“˜≈~‘ind, is zonal and belongs to ´æ(K)≈¤“a~+®da‘, so it is an N–fixed co-ç°–vector associated to
“˜≈~‘ind. We will return to this in Sections VII.6-8.
It follows from (V.5.5.b) that
Proposition V.5.6 For every minimal bilaterally invariant Hilbert subspace of ∂æ(G/N)
there exist ≈∑MÀ and ~∑ia such that is a Hilbert subspace of ·“˜ ≈~‘ind‚–°.
We then use the following lemma, in addition to some general theory.
Lemma V.5.7 Let Ó° ≤â Ó ≤â Ó–° be as at (V.5.5.c).
Let Ó have a finite decomposition:
Ó=i¶=
n
1@
Ói, n<°, Ói irreducible and inequivalent.
Then any G–invariant Hilbert subspace of Ó–° is of the form
˚=i¶=
n
1@
¬iÓi,
for certain positive numbers ¬i, i=1,Ú,n.
Corollary V.5.8 Assume that Ó is the sum
(V.5.8.a) Ó=i¶=
n
1@
Ói
111 Garth Warner, Harmonic Analysis on Semi-Simple Lie Groups I, Chapter 5.112 Niels Skovhus Poulsen, “On ç°–Vectors and Intertwining Bilinear Forms for Representations of Lie
Groups,” Journal of Functional Analysis 9 (1972), pp. 87-120, Theorem 5.1.
175
— Chapter V The Homogeneous Space G/N —
with the Ói irreducible.
Then the representation †–° in Ó–° is multiplicity free if and only if (V.5.8.a) is
multiplicity free.
The lemma is quite what one would expect, but does require a proof. To see that it is not enough that
Ó is dense in Ó–° consider a homogeneous space ≈ with invariant measure dx such that L™(≈;dx) is
irreducible. (Such spaces exist, an example is Ân acted on by the semi-direct product of SL(n;Â) and
Ân). Then although L™(≈;dx) is dense in ∂æ(≈), there still exists the one-dimensional Hilbert subspace
[dx] of ∂æ(≈), which is not a multiple of L™(≈;dx). This example does not contradict Lemma V.5.7,
since dx does not belong to Ó–°. Roughly speaking, Lemma V.5.7 says that Ó–° fits Ó like a glove,
leaving no room for any G–invariant Hilbert subspaces other than the obvious ones.
Lemma V.5.7 is easily proved, using some well-known facts. One fact we use twice is that
if Ó is an irreducible Hilbert representation space, then the positive multiples of Ó are the only
G–invariant Hilbert subspaces of Ó. The reason is that due to Schur’s Lemma reproducing operators
of Hilbert subspaces of Ó must be trivial.
Proof First take n=1, so Ó is irreducible. Assume that ˚≠(0) is a Hilbert subspace of Ó–°. For a
testfunction ƒ consider the image of ˚ under the operator πƒ–°:=ª
Gƒ(g)π g–°dg. Choose ƒ so that
πƒ–°˚ is not (0). (This is always possible, otherwise varying ƒ to approximate ∂eG would show that
˚=(0)). It is well-known that the Gårding vectors πƒ°(v):=ª
Gƒ(g)πgvdg, ƒ∑∂(G), v∑V, belong to
Ó°. Since πƒ–° is the transpose of πƒõ
° it follows that πƒ–° maps Ó–° into Ó. So in particular
πƒ–°˚ is contained in Ó. On the other hand, ˚ is assumed π–°–invariant, and the restriction of
π–° to ˚ is a continuous representation. Therefore, πƒ–°˚ is also contained in ˚. This shows that
Ó§˚ is not 0. Now Ó§˚ is in its own right a Hilbert subspace of Ó–°, when equipped with the
norm ||T||™Ó§˚=||T||™Ó+||T|| ™˚ [113] . So one has the continuous inclusion (Ó§˚)≤âÓ, and Ó being
irreducible it follows that Ó§˚=¬Ó for some positive constant ¬. But this implies that Ó is contained
in ˚, that is, Ó≤â˚≤âÓ–°, both inclusions being continuous in view of the closed graph theorem.
This in turn implies that ˚ must be irreducible: indeed, if ˚ allowed a G–invariant decomposition
˚=˚¡@˚™, with neither ˚¡ nor ˚™ equal to (0), the preceding argument involving ˚ would apply
113 A standard construction: see Laurent Schwartz, "Sous-espaces Hilbertiens d'Espaces Vectoriels
Topologiques et Noyaux Associés (Noyaux Reproduisants)," Jour. Anal. Math. 13 (1964), p. 138, Proof of