Documenta Math.

219

L2 -Invariants of Locally Symmetric Spaces

Martin Olbrich

Received: August 9, 2001

Communicated by Ursula Hamenstädt

Abstract. Let X = G/K be a Riemannian symmetric space of the

noncompact type, Γ ⊂ G a discrete, torsion-free, cocompact subgroup,
and let Y = Γ\X be the corresponding locally symmetric space. In
this paper we explain how the Harish-Chandra Plancherel Theorem
for L2 (G) and results on (g, K)-cohomology can be used in order to
compute the L2 -Betti numbers, the Novikov-Shubin invariants, and
the L2 -torsion of Y in a uniform way thus completing results previ-
ously obtained by Borel, Lott, Mathai, Hess and Schick, Lohoue and
Mehdi. It turns out that the behaviour of these invariants is essen-
tially determined by the fundamental rank m = rkC G−rkC K of G. In
particular, we show the nonvanishing of the L2 -torsion of Y whenever
m = 1.

2000 Mathematics Subject Classification: 58J35, 57R19, 22E46

Keywords and Phrases: locally symmetric spaces, L2 -cohomology,
Novikov-Shubin invariants, L2 -torsion, relative Lie algebra cohomol-
ogy

1 Introduction

During the last two decades L2 -invariants have proved to be a powerful tool in
the topology of compact manifolds (see [15] for an overview). Although they
can be defined in purely combinatorial terms we are interested here in their
equivalent analytic versions: They are spectral invariants of the p-form Lapla-
cians of the universal cover of the manifold. For particular nice manifolds these
might be computable. Indeed, the aim of the present paper is to extract from
the representation theoretic work of Harish-Chandra [9] and Borel-Wallach [2]
information on the spectral decomposition of the form Laplacians on Rieman-
nian symmetric spaces of the non-compact type which is sufficiently explicit
in order to compute the spectral invariants of interest. We try to do this in a

Documenta Mathematica 7 (2002) 219–237

220 Martin Olbrich

rather detailed way which, we hope, keeps the paper readable for nonspecialists
in harmonic analysis. 1
Let X → Y be the universal cover of a compact Riemannian manifold. Set
Γ := π1 (Y ). The form Laplacian ∆p = d∗ d + dd∗ defines a non-negative,
elliptic, self-adjoint operator acting on L2 (X, Λp T ∗ X), the square integrable
p-forms on X. By ∆0p and ∆cp = (d∗ d)0 we denote the restriction of ∆p to
the orthogonal complement of its kernel and the coclosed forms in this orthog-
onal ∗complement, respectively. We consider the corresponding heat kernels
e−t∆p (x, x0 ) := (P∗ e−t∆p )(x, x0 ), x, x0 ∈ X, for ∗ = ∅, 0 or c. Here P∗ denotes
the orthogonal
∗
projection to the corresponding subspace. The local traces
tr e−t∆p (x, x) are Γ-invariant functions on X. For a thorough discussion of the
following definitions we refer to [14], [17], [16], and [15].
We set Z
−t∆∗ ∗
TrΓ e p := tr e−t∆p (x, x) dx ,
F
where F ⊂ X is a fundamental domain of the action of Γ on X and dx is the
Riemannian volume element of X. Then the L2 -Betti numbers are given by
−t∆p
b(2)
p (Y ) := lim TrΓ e ∈ [0, ∞) .
t→∞

They are equal to the von Neumann dimension of ker ∆p viewed as Hilbert
N (Γ)-module, where N (Γ) is the group von Neumann algebra of Γ. If the
spectrum of ∆p has no gap around 0 the Novikov-Shubin invariants of Y are
defined by
0 β
α̃p (Y ) := sup{β | TrΓ e−t∆p = O(t− 2 ) as t → ∞} ∈ [0, ∞] .
It measures the asymptotic behaviour of the spectral density function of ∆ p at
0. In case of a gap around 0 we set α̃p (Y ) := ∞+ . Replacing ∆0p by ∆cp we
obtain the analogously defined Novikov-Shubin invariants αp (Y ) of d∗ d. Using
the action of the exterior differential d on the Hodge decomposition of L2 -forms
we obtain
α̃p (Y ) = min{αp (Y ), αp−1 (Y )} . (1)
Finally, if αp (Y ) > 0 for all p (or, more generally, if X is of determinant class
(see [15])), then the L2 -torsion of Y is defined by
1X 1X
ρ(2) (Y ) := (−1)p+1 p log detΓ (∆0p ) = (−1)p log detΓ (∆cp ) ,
2 2
where for ∗ = 0, c
µ Z ε ¶
∗ d 1 −t∆∗ s−1
− log detΓ (∆p ) := TrΓ e p t dt
ds |s=0 Γ(s) 0
Z ∞
∗
+ TrΓ e−t∆p t−1 dt
ε
1 Note added in proof: In the meanwhile the interested reader can find a discussion with-
out proofs of the results of the present paper in the recent monograph [16] which gives a
comprehensive treatment of the theory of L2 -invariants.

Documenta Mathematica 7 (2002) 219–237

 L2 -Invariants of Locally Symmetric Spaces 221

for any ε > 0, where the first integral is considered as a meromorphic function
in s. By Poincaré duality ρ(2) (Y ) = 0 for even dimensional manifolds Y .
From now on let X = G/K be a Riemannian symmetric space of the noncom-
pact type. Here G is a real, connected, linear, semisimple Lie group without
compact factors, and K ⊂ G is a maximal compact subgroup. It is the uni-
versal cover of compact locally symmetric spaces of the form Y = Γ\X, where
Γ∼ = π1 (Y ) can be identified with a discrete, torsion-free, cocompact subgroup
of G. It will be convenient to consider also the compact dual X d of X. X d
is defined as follows: Let g, k be the Lie algebras of G, K. Then we have the
Cartan decomposition g = k ⊕ p, and gd := k ⊕ ip is another subalgebra of the
complexification of g. Let Gd be the corresponding analytic subgroup of the
complexification GC of G. Then Gd is a compact group, and X d = Gd /K. We
normalize the Riemannian metric on X d such that multiplication by i becomes
an isometry TeK X ∼ = p → ip ∼ = TeK X d . In this paper we are going to prove
the following theorem:

Theorem 1.1 Let n = dim Y and m = m(X) := rkC G − rkC K be the funda-
mental rank of G. Let χ(Y ) be the Euler characteristic of Y . Then
(2) n
(a) bp (Y ) 6= 0 ⇔ m = 0 and p = 2.
(2) n vol(Y )
In particular, b n (Y ) = (−1) 2 χ(Y ) = χ(X d ).
2 vol(X d )
(b) αp (Y ) 6= ∞+ ⇔ m > 0 and p ∈ [ n−m n+m
2 , 2 − 1].
In this range αp (Y ) = m.
(c) ρ(2) (Y ) 6= 0 ⇔ m = 1.

Note that n − m is always even and positive. Part (a) of the theorem was
known for a long time, at least since Borel’s paper [1]. For special cases see
also [4], [5]. For the convenience of the reader we include a proof here. In fact,
we prove a stronger statement which should have been known to the experts
although we were not able to find it in the literature:

Proposition 1.2 The discrete spectrum of ∆p on L2 (X, Λp T ∗ X) is empty

unless m(X) = 0 and p = n2 . In this case 0 is the only eigenvalue of ∆p .

The strategy of the proof of (b) can already be found in Lott’s paper [14],
Section VII. To be more precise, Equation (1) implies the slightly weaker result
½
∞+ p 6∈ [ n−m n+m
2 , 2 ] or m = 0
α̃p (Y ) = n−m n+m . (2)
m p ∈ [ 2 , 2 ] and m 6= 0

Lott proved the first line of (2) and that α̃p (Y ) is finite and independent of p
for the remaining values of p. He indicated how one should be able to compute
the precise value of α̃p (Y ). But he finished the computation in the real rank
one case, only. In addition, already Borel [1] showed that the range of the

Documenta Mathematica 7 (2002) 219–237

222 Martin Olbrich

differential dp of the L2 -de-Rham-complex is not closed for p ∈ [ n−m n+m

2 , 2 −1].
Theorem 1.1 (b) can be considered as a quantitative refinement of this result.
After the present paper was written I was informed by S. Mehdi that there is
a joint paper of him with N. Lohoue [13] which has recently appeared in print
and which contains a proof of (2). The interested reader will also find there
more information concerning the material presented here in Sections 2 and 3.
But he should be aware that in that paper the range of finiteness of α̃p (Y ) is
constantly misprinted and that the proof of the second line of (2) as written
down there is not quite complete (it is not mentioned that it is important to
know that pξ (0) > 0, see Equation (14) below).
The main motivation to do the present work was to obtain part (c) of the
theorem. In our locally homogeneous situation we have for any x ∈ X
∗ ∗
TrΓ e−t∆p = vol(Y ) · tr e−t∆p (x, x)

and thus
ρ(2) (Y ) = vol(Y ) · T (2) (X)
for a certain real number T (2) (X). Note that in contrast to ρ(2) (Y ) the num-
ber T (2) (X) depends on the normalization of the invariant metric on X (of
course only via the volume form). A well-known symmetry argument ([18],
Proposition 2.1) yields that T (2) (X) = 0 whenever m 6= 1. Lott [14] (see also
[17]) showed that T (2) (H n ) 6= 0 for n = 3, 5, 7, where H n is the real hyperbolic
space (his values for T (2) (H n ) for n = 5, 7 were not correct). This led to the
conjecture that T (2) (H n ) 6= 0 for all odd n which was open until the work of
Hess-Schick [11] who found a trick in order to control the sign of log detΓ (∆cp )
in terms of p and n. So they were able to show that there is a positive rational
number qn such that
µ ¶ n−1
(2) n 1 2
T (H ) = − qn . (3)
π
Here the metric on H n is normalized to have sectional curvature −1. Along
the same lines Hess [10] obtained

T (2) (SL(3, R)/SO(3)) 6= 0 . (4)

Let us introduce
2qn
Qn = n−1
( 2 )!
and rewrite (3) as
n−1 πQn
T (2) (H n ) = (−1) 2 . (5)
vol(S n )
Recall that S n is the compact dual of H n . The rational number Qn which
does not depend on the normalization of the metric has a nice interpretation
in terms of Weyl’s dimension polynomial for finite-dimensional representations,
see Proposition 5.3. But its significance remains to be clarified further, and it

Documenta Mathematica 7 (2002) 219–237

 L2 -Invariants of Locally Symmetric Spaces 223

seems to be difficult to write down a practical formula valid for all odd n. One
has Q3 = 13 , Q5 = 31 221
45 , Q7 = 210 . For further information see [11].
We will reduce Theorem 1.1 (c) to (4) and the positivity of Qn . Let X be an
arbitrary symmetric space satisfying m(X) = 1. By the classification of simple
Lie groups X = X1 × X0 , where m(X0 ) = 0 and X1 = SL(3, R)/SO(3) or
X1 = Xp,q := SO(p, q)0 /SO(p) × SO(q) for p, q odd. (Here as throughout the
paper a superscript 0 denotes the connected component of the identity.) Note
that a corresponding decomposition of Y does not necessarily exist. We show
Proposition 1.3

pq−1
d πQp+q−1
(a) T (2) (Xp,q ) = (−1) 2 χ(Xp−1,q−1 ) d )
.
vol(Xp,q
n0

(2) (−1) 2 χ(X0d ) (2)

(b) If m(X) = 1, then T (X) = T (X1 ).
vol(X0d )
Here n0 = dim X0 .
d
Note that Xp−1,q−1 = SO(p + q − 2)/SO(p − 1) × SO(q − 1) and for p, q > 1
µ p+q−2 ¶
d 2
χ(Xp−1,q−1 ) = 2 p−1 .
2
d
In fact, it is a classical result that χ(X ) > 0 whenever m(X) = 0 (compare
Theorem 1.1 (a)). It is equal to the quotient of the orders of certain Weyl
groups (see Section 5). Now Theorem 1.1 (c) follows from Proposition 1.3, (4)
and the positivity of Qn .
We are also able to identify the missing constant in (4).
Proposition 1.4 If X = SL(3, R)/SO(3), then
2π
T (2) (X) = .
3vol(X d )
If the invariant metric on X is induced from twice the trace form of the standard
representation of sl(3, R), then vol(X d ) = 4π 3 , and we have
1
T (2) (X) = .
6π 2
n−1
In particular, we see that (−1) 2 T (2) (X) is positive for all X with m(X) = 1.
Proposition 5.3 provides a uniform formula for the L2 -torsion of all these spaces.
Acknowledgements: I am grateful to Wolfgang Lück and Thomas Schick for
inspiring discussions which have provided me with a sufficient amount of moti-
vation and of knowledge on L2 -invariants in order to perform the computations
which led to the results of the present paper. I am also indebted to Wolfgang
Lück for giving me the opportunity to report on them at the Oberwolfach con-
ference “L2 -methods and K-theory”, September 1999. In addition, I benefited
from discussions with J. Lott, P. Pansu, E. Hess and U. Bunke.

Documenta Mathematica 7 (2002) 219–237

224 Martin Olbrich

2 The Harish-Chandra Plancherel Theorem

We want to understand the action of the Laplacian and of the corresponding

heat kernels on L2 (X, Λp T ∗ X). Since the Laplacian coincides (up to the sign)
with the action of the Casimir operator Ω of G (Kuga’s Lemma [2], Thm. 2.5.)
it is certainly enough to understand the “decomposition” of L2 (X, Λp T ∗ X) into
irreducible unitary representations of G. There is an isomorphism of homoge-
neous vector bundles Λp T ∗ X ∼= G ×K Λp p∗ , and, hence, of G-representations

L2 (X, Λp T ∗ X) ∼
= [L2 (G) ⊗ Λp p∗ ]K .

Thus our task consists of two steps: First to understand L2 (G) as a repre-
sentation of G × G which is accomplished by the Harish-Chandra Plancherel
Theorem recalled in the present section and, second, to understand spaces of
the form [Vπ ⊗ Λp p∗ ]K , where (π, Vπ ) is an irreducible unitary representation of
G occurring in the Plancherel decomposition. For general G, the second step
will resist a naive approach. However, if π(Ω) = 0, then the space [Vπ ⊗ Λp p∗ ]K
has cohomological meaning, and the theory of relative (g, K)-cohomology as
recalled in the next section will provide a sufficient amount of information.
An irreducible unitary representation (π, Vπ ) of G is called a representation
of the discrete series if there is a G-invariant embedding Vπ ,→ L2 (G). Let
Ĝd denote the set of equivalence classes of discrete series representations of G.
Then we have

Theorem 2.1 (Harish-Chandra [7]) Ĝd is non-empty if and only if

m(X) = 0.

Note that m(X) = 0 means that G has a compact Cartan subgroup. The
Plancherel Theorem provides a decomposition of L2 (G) which is indexed by
discrete series representations of certain subgroups M ⊂ G which we are going
to define now.
Let a0 ⊂ p be a maximal abelian subspace. It induces a root space decompo-
sition M
g = g0 ⊕ gα .
α∈∆(g,a0 )

·
Choose a decomposition ∆(g, a0 ) = ∆+ ∪ −∆+ into positive and negative
roots, and let Π ⊂ ∆+ be the subset of simple roots. For any subset F ⊂ Π we
define

aF = {H ∈ a0 | α(H) = 0 for all α ∈ Π} , AF = exp(aF ) ,

M
nF = gα , NF = exp(nF ) .
{α∈∆+ |α|aF 6=0}

Furthermore, there is a unique (possibly disconnected) subgroup MF ⊂ G

with Lie algebra mF such that MF AF is the centralizer of aF in G and mF

Documenta Mathematica 7 (2002) 219–237

 L2 -Invariants of Locally Symmetric Spaces 225

is orthogonal to aF with respect to any invariant bilinear form on g. MF

is a reductive subgroup with compact center. The corresponding parabolic
subgroup PF := MF AF NF is called a standard parabolic. PF is called cuspidal
if MF has a compact Cartan subgroup. If for two subsets F, I ⊂ Π the spaces
aF and aI are conjugated by an element of K (thus by an element of the Weyl
group W (g, a0 )) we call PF and PI associate. (In many cases this already
implies F = I.) The assignment PF 7→ AF T , where T is a compact Cartan
subgroup of MF , gives a one-to-one correspondence between association classes
of cuspidal parabolic subgroups and conjugacy classes of Cartan subgroups of
G.
For two subsets F ⊂ I ⊂ Π we have PF ⊂ PI , MF ⊂ MI , AF ⊃ AI , NF ⊃ NI .
If F is understood we will often suppress the subscript F .
For illustration let us consider the two extreme cases. The minimal parabolic
arises for F = ∅. Since M∅ is compact P∅ is always cuspidal. For F = Π we
have P = M = G, and G is cuspidal iff m(X) = 0. For any cuspidal parabolic
subgroup we have dim A ≥ m(X), and there is exactly one association class of
cuspidal parabolic subgroups, called fundamental, with dim A = m(X).
Theorem 2.1 also holds in the context of such reductive groups like M . Thus
a parabolic P = M AN is cuspidal iff M has a non-empty discrete series M̂d .
Let a∗C be the complexified dual of the Lie algebra a of A. For a discrete series
representation (ξ, Wξ ) of M and ν ∈ a∗C we form the induced representation
(πξ,ν , H ξ,ν ) by
½ ¾
f (gman) = a−(ν+ρa ) ξ(m)−1 f (g) for all
H ξ,ν = f : G → Wξ | ,
g ∈ G, man ∈ M AN, f|K ∈ L2 (K, Wξ )
(πξ,ν (g)f )(x) = f (g −1 x) .
1 X
Here ρa = α|a . If ν ∈ ia∗ , then πξ,ν is unitary. An invariant bilinear
2 +
α∈∆
form on g induces corresponding forms on m and a∗C and determines Casimir
operators Ω and ΩM of G and M , respectively. Then we have
πξ,ν (Ω) = hν, νi − hρa , ρa i + ξ(ΩM ) . (6)
Note that ξ(ΩM ) is a non-negative real scalar.
Let C(G) ⊂ L2 (G) be the Harish-Chandra Schwartz space (for a definition see
e.g. [19], 7.1.2). It is stable under the left and right regular representations l
and r of G. Let C(G)K×K = {f ∈ C(G) | dim span{lk1 rk2 f | k1 , k2 ∈ K} < ∞}
be the subspace of Schwartz functions which are K-finite from the left and the
right. Note that C(G)K×K is dense in L2 (G). Suppose that ν ∈ ia∗ . Then for
f ∈ C(G) Z
πξ,ν (f ) := f (g)πξ,ν (g) dg
G
is a well-defined trace class operator on H ξ,ν which has finite rank if f ∈
C(G)K×K . Note that the map f 7→ πξ,ν (f ) intertwines the G-actions in the
following way: πξ,ν (lx ry f ) = πξ,ν (x)πξ,ν (f )πξ,ν (y −1 ) for x, y ∈ G.

Documenta Mathematica 7 (2002) 219–237

226 Martin Olbrich

The Harish-Chandra Plancherel Theorem can now be formulated as follows:

Theorem 2.2 (Harish-Chandra [9]) For each cuspidal parabolic subgroup

as constructed above and any discrete series representation ξ of the cor-
responding group M there exists an explicitly computable analytic function
pξ : ia∗ → [0, ∞) of polynomial growth (the Plancherel density) such that for
any f ∈ C(G)K×K and g ∈ G
X X Z
f (g) = Tr(πξ,iν (f )πξ,iν (g −1 )) pξ (iν)dν .
P ξ∈M̂d a∗

Here the first sum runs over a set of representatives P = PF of association

classes of cuspidal parabolic subgroups of G.

For more details on the Plancherel Theorem and the structure theory behind
it the interested reader may consult the textbooks [12], [19], [20].
Note that the Plancherel measures pξ (iν)dν depend on the normalization of
the Haar measure dg. In the remainder of the paper we use the following one.
Let dx be the Riemannian volume form of ZX = G/K and Z dk be the Haar
measure of K with total mass one. Then f (g) dg = f˜(x) dx, where
Z G X

f˜(gK) = f (gk) dk. We normalize the invariant bilinear form on g such that
K
its restriction to p ∼
= TeK X coincides with the Riemannian metric of X. Let dν
be the Lebesgue measure corresponding to the induced form on a∗ . By these
choices pξ is uniquely determined.
∗
We are now able to give a kind of spectral expansion of TrΓ e−t∆p .

Corollary 2.3
X X Z 2
+kρa k2 −ξ(ΩM ))
TrΓ e−t∆p = vol(Y ) e−t(kνk
P ξ∈M̂d a∗

dim[H ξ,iν ⊗ Λp p∗ ]K pξ (iν)dν (7)

X X Z 2 2
= vol(Y ) e−t(kνk +kρa k −ξ(ΩM ))

P ξ∈M̂d a∗

dim[Wξ ⊗ Λp p∗ ]KM pξ (iν)dν . (8)

Here KM := K ∩ M denotes the maximal compact subgroup of M . There are

only finitely many pairs (P, ξ) with [H ξ,iν ⊗ Λp p∗ ]K ∼
= [Wξ ⊗ Λp p∗ ]KM 6= {0}.

Proof. We define kt ∈ [C(G) ⊗ End(Λp p∗ )]K×K by kt (g) := e−t∆p (eK, gK) ◦

2 p ∗ ∼ 2 p ∗ K −t∆p
Z For f ∈ L (X, Λ T X) = [L (G) ⊗ Λ p ] we have e
g. f (g0 ) =
kt (g)f (g0 g) dg. In addition, tr e−t∆p (x, x) = tr kt (e) for any x ∈ X.
G
We consider πξ,iν (kt ) as an operator acting on H ξ,iν ⊗ Λp p∗ . Using Kuga’s

Documenta Mathematica 7 (2002) 219–237

 L2 -Invariants of Locally Symmetric Spaces 227

Lemma, Equation (6), and the K × K-invariance of kt one derives that

2 2
πξ,iν (kt ) = e−t(kvk +kρa k −ξ(ΩM )) P , where P is the orthogonal projection onto
the subspace of K-invariants in H ξ,iν ⊗ Λp p∗ . The Plancherel formula now
yields
X X Z 2 2
tr kt (e) = e−t(kvk +kρa k −ξ(ΩM )) dim[H ξ,iν ⊗ Λp p∗ ]K pξ (iν)dν .
P ξ∈M̂d a∗

This proves (7). Since H ξ,iν ∼

= L2 (K ×KM Wξ ) as a representation of K Equa-
tion (8) follows by Frobenius reciprocity. The last assertion is a consequence
of the Blattner formula (see e.g. [19], 6.5.4) for the KM -types of discrete series
representations of M . 2

3 (g, K)-cohomology

If (π, Vπ ) is a representation of G on a complete locally convex Hausdorff topo-

logical vector space, then we can form its subspace Vπ,K consisting of all K-
finite smooth vectors of Vπ . Vπ,K becomes a simultaneous module under g and
K, where both actions satisfy the obvious compatibility conditions. Such a
module is called a (g, K)-module (see [2], 0.2).
We are interested in the functor of (g, K)-cohomology V 7→ H ∗ (g, K, V ) which
goes from the category of (g, K)-modules to the category of vector spaces. It
is the right derived functor of the left exact functor taking (g, K)-invariants.
H ∗ (g, K, V ) can be computed using the standard relative Lie algebra cohomol-
ogy complex ([V ⊗ Λ∗ p∗ ]K , d), where
p
X
dω(X0 , . . . , Xp ) = (−1)i π(Xi )ω(X0 , . . . , X̂i , . . . , Xp ) ,
i=0
ω ∈ [V ⊗ Λp p∗ ]K , Xi ∈ p .

Note that for V = C ∞ (G)K this complex is isomorphic to the de Rham complex
of the symmetric space X.
Let Z(g) be the center of the universal enveloping algebra of g. If (π, V ) is
an irreducible (g, K)-module, then any z ∈ Z(g) acts by a scalar χπ (z) on V .
The homomorphism χπ : Z(g) → C is called the infinitesimal character of V .
The following basic result can be considered as an algebraic version of Hodge
theory.

Proposition 3.1 ([2], II.3.1. and I.5.3.) Let (π, Vπ ) be an irreducible uni-
tary representation of G, and let (τ, F ) be an irreducible finite-dimensional
representation of G. Then
½
[Vπ ⊗ F ⊗ Λp p∗ ]K π(Ω) = τ (Ω)
H p (g, K, Vπ,K ⊗ F ) = .
{0} π(Ω) 6= τ (Ω)

Documenta Mathematica 7 (2002) 219–237

228 Martin Olbrich

If H p (g, K, Vπ,K ⊗ F ) 6= {0}, then χπ = χτ̃ , where τ̃ is the dual representation

of τ .

The cohomology groups H p (g, K, Vπ,K ⊗ F ) for the representations π = πξ,iν

occurring in the Plancherel Theorem have been computed in [2]. We shall need
the following information.

Proposition 3.2 ([2],II.5.3., III.5.1., and III.3.3.)

(a) Let (τ, F ) be an irreducible finite-dimensional representation of G and

π ∈ Ĝd with χπ = χτ̃ . Then
½
1 p = n2
dim H p (g, K, Vπ,K ⊗ F ) = .
0 otherwise

(b) Let (πξ,ν , H ξ,ν ) be a representation occurring in the Plancherel Theorem.

Then
ξ,iν
H ∗ (g, K, HK ) = {0}
unless P is fundamental, ν = 0 and ξ belongs to a certain non-empty
finite subset Ξ ⊂ M̂d . If P is fundamental and ξ ∈ Ξ, then
½ ¡ m ¢
ξ,0 p− n−m p ∈ [ n−m n+m
2 , 2 ]
dim H p (g, K, HK )= 2 .
0 otherwise

Choose a Cartan subalgebra h ⊂ g and a system of positive roots. Via the

Harish-Chandra isomorphism any infinitesimal character χπ : Z(g) → C is
given by an element Λπ ∈ h∗C , which is uniquely determined up to the action
of the Weyl group W (g, h) ([19], 3.2.4.). The set of infinitesimal characters
of discrete series representations coincides (in case m = 0) with the set of
infinitesimal characters of finite-dimensional representations, which are of the
form µτ + ρg , µτ and ρg being the highest weight of τ and the half-sum of
positive roots, respectively. In the following we will represent infinitesimal
characters of discrete series representations by elements of this form. We in-
troduce a partial order on h∗C by saying that µ > ν, if µ − ν is a sum of (not
necessarily distinct) positive roots. Then a careful examination of the proof of
Proposition II.5.3. in [2], which only rests on some basic knowledge of the pos-
sible K-types occurring in discrete series representations, shows that slightly
more than Proposition 3.2 (a) is true.

Proposition 3.3 Let (τ, F ) be an irreducible finite-dimensional representa-

tion of G and π ∈ Ĝd with Λπ 6< Λτ̃ . Then
½
1 p = n2 , χπ = χτ̃
dim[Vπ,K ⊗ F ⊗ Λp p∗ ]K = .
0 otherwise

Documenta Mathematica 7 (2002) 219–237

 L2 -Invariants of Locally Symmetric Spaces 229

4 L2 -Betti numbers and Novikov-Shubin invariants

In this section we shall prove parts (a) and (b) of Theorem 1.1 as well as
Proposition 1.2.
Let L2 (X, Λp T ∗ X)d be the discrete subspace of L2 (X, Λp T ∗ X), i.e., the di-
rect sum of the L2 -eigenspaces of the Laplacian. The Plancherel Theorem in
particular says that as a representation of G
M
L2 (X, Λp T ∗ X)d ∼
= Vπ̃ ⊗ [Vπ ⊗ Λp p∗ ]K .
π∈Ĝd

Let (τ0 , C) be the trivial representation of G. Let π ∈ Ĝd . Then Λπ 6< Λτ0 = ρg .
Proposition 3.3 now yields
½
1 p = n2 , χπ = χτ0
dim[Vπ ⊗ Λp p∗ ]K = . (9)
0 otherwise

Since in case m = 0 discrete series representations with infinitesimal character

χτ0 always exist (see e.g. [12], Thm. 9.20 or [19], Thm. 6.8.2) and τ0 (Ω) = 0
(2)
this implies Proposition 1.2. In particular, bp (Y ) is non-zero exactly when
n 2
m = 0 and p = 2 . Using that the L -Euler characteristic coincides with the
usual Euler characteristic and applying Hirzebruch proportionality we obtain

(2) n vol(Y )
b n (Y ) = (−1) 2 χ(Y ) = χ(X d ) . (10)
2 vol(X d )

We will give an alternative, purely analytic proof of that formula in the sequel
of Corollary 5.2. This finishes the proof of part (a) of Theorem 1.1.
We now turn to part (b). In order to compute αp (Y ) we need an expression
c
for TrΓ e−t∆p .

Proposition 4.1 For any triple (P, ξ, ν) appearing in (7) let B p (ξ, ν) =
d([H ξ,iν ⊗ Λp−1 p∗ ]K ) be the space of coboundaries in the relative Lie algebra
cohomology complex and bp (ξ, ν) be its dimension. Then

c X X Z 2 2
TrΓ e−t∆p = vol(Y ) e−t(kνk +kρa k −ξ(ΩM ))
P 6=G ξ∈M̂d a∗

bp+1 (ξ, ν) pξ (iν)dν . (11)

Here the first sum runs over a set of representatives P = PF of association

classes of proper cuspidal parabolic subgroups of G. If m > 0, P is fundamental,
ξ ∈ Ξ (see Proposition 3.2 (b)), and ν 6= 0, then
( ¡
m−1 ¢
p+1 p− n−m
p ∈ [ n−m n+m
2 , 2 − 1]
b (ξ, ν) = 2 . (12)
0 otherwise

Documenta Mathematica 7 (2002) 219–237

230 Martin Olbrich
c
Proof. We proceed exactly as in the proof of Corollary 2.3. The kernel e−t∆p
determines a function ktc ∈ [C(G) ⊗ End(Λp p∗ )]K×K . Then one computes that
2 2
πξ,iν (ktc ) = e−t(kvk +kρa k −ξ(ΩM )) P c , where P c is the projection to the orthog-
onal complement in [H ξ,iν ⊗ Λp p∗ ]K of the space of cocycles in the relative Lie
algebra cohomology complex. The dimension of that complement is equal to
bp+1 (ξ, ν). If m = 0, P = G, and ξ ∈ Ĝd , then bp+1 (ξ) = 0 for all p by (9).
This proves (11).
Let now m > 0, P be fundamental, ξ ∈ Ξ, and ν ∈ a∗ . In this case
ξ,0
H p (g, K, HK ) 6= {0} for some p. Hence Proposition 3.1 implies that πξ,0 (Ω) =
0 and
ξ,0
hp (ξ) := dim[H ξ,iν ⊗ Λp p∗ ]K = dim[H ξ,0 ⊗ Λp p∗ ]K = dim H p (g, K, HK ).
¡ ¢
By Proposition 3.2 (b) we have hp (ξ) = p− m n−m . On the other hand,
2
ξ,iν
dim H p (g, K, HK ) = 0 for ν 6= 0 implies that hp (ξ) = bp (ξ, ν) + bp+1 (ξ, ν).
(12) now follows inductively. 2
ξ,iν
If H p (g, K, HK ) = {0} for all ν ∈ a∗ , then by Proposition 3.1 dim[H ξ,iν ⊗
Λ p ] = 0 (which is independent of ν) or inf∗ (kνk2 + kρa k2 − ξ(ΩM )) > 0.
p ∗ K
ν∈a
Now Proposition 4.1 implies that the spectrum of ∆cp has a gap around zero,
which means αp (Y ) = ∞+ , unless m > 0 and p ∈ [ n−m n+m
2 , 2 − 1]. In the latter
case we obtain for some c > 0
µ ¶XZ
c m−1 2
TrΓ e−t∆p = vol(Y ) n−m e−tkνk pξ (iν)dν + O(e−ct )
p− 2 a∗
ξ∈Ξ

as t → ∞, where a corresponds to a fundamental parabolic subgroup and thus

has dimension m.
Let P = M AN be fundamental and ξ ∈ M̂d . Then dim n =: 2u is even.
Choose a compact Cartan subgroup T ⊂ M with Lie algebra t and a system
∆+ (m, t) of positive roots. Considering the pair (m, t) instead of (g, h) we can
define Λξ ∈ t∗C in the same way as at the end of Section 3. Then h := t ⊕ a
is a Cartan subalgebra of g. Let Φ+ be a system of positive roots for (g, h)
containing ∆+ (m, t). Then there exists a positive constant cX depending only
on the normalization of the volume form dx such that
Y hα, Λξ + νi
pξ (ν) = cX (−1)u (13)
+
hα, ρg i
α∈Φ

(see [9], Thm. 24.1, [20], Thm. 13.5.1 or [12], Thm. 13.11). In particular, p ξ
is an even polynomial of degree dim n. The factor (−1)u makes it nonnegative
on ia∗ .
The element Λξ + ν gives the infinitesimal character of πξ,ν : Λπξ,ν = Λξ + ν
([12], Prop. 8.22). Let now ξ ∈ Ξ. Then Propositions 3.2 (b) and 3.1 imply
that πξ,0 has the same infinitesimal character as the trivial representation. It

Documenta Mathematica 7 (2002) 219–237

 L2 -Invariants of Locally Symmetric Spaces 231

follows that Λξ is conjugated in h∗C by an element of the Weyl group W (g, h) to

ρg . By (13) we obtain pξ (0) = ±cX 6= 0. On the other hand pξ (0) ≥ 0, hence

pξ (0) > 0 . (14)

X u
We decompose pξ (ν) = pξ,2k (ν) into homogeneous polynomials. Set qξ,k =
Z k=0

pξ,2k (ν) dν. Then qξ,0 > 0 and

kνk=1
Z X Z ∞
−tkνk2 2
e pξ (iν)dν = e−tr rm−1+2k dr
qξ,k
a∗ 0
X Z ∞
2
−( m
= t 2 +k)
qξ,k e−y y m−1+2k dy .
0
c
−t∆p
Thus for p ∈ [ n−m n+m
2 , 2m − 1] the leading term of TrΓ e as t → ∞ is a
−2
non-zero multiple of t . This completes the proof of Theorem 1.1 (b).

5 L2 -torsion

For even dimensional manifolds the L2 -torsion vanishes. Thus we may assume
that m is odd, in particular m ≥ 1. Then ∆0p = ∆p . We first want to compute

X n
1
kX (t) := (−1)p p TrΓ e−t∆p .
2vol(Y ) p=0

Then
µ Z ε ¶ Z ∞
d 1
T (2)
(X) = kX (t)t s−1
dt + kX (t)t−1 dt .
ds |s=0 Γ(s) 0 ε

Let P = M AN be a parabolic subgroup appearing in (8). Set KM = K ∩ M

and pm = p ∩ m. Then an elementary calculation in the representation ring
R(KM ) of KM yields
n
X
(−1)p p Λp p∗ = 0 , if dim a ≥ 2 ,
p=0

while for dim a = 1 we have

n
X n−1
X
(−1)p p Λp p∗ = (−1)p+1 Λp (p∗m ⊕ n∗ )
p=0 p=0
dim
Xn
= (−1)l+1 (Λev p∗m − Λodd p∗m ) ⊗ Λl n
l=0

Documenta Mathematica 7 (2002) 219–237

232 Martin Olbrich

(see [18], Prop. 2.1 and Lemma 2.3). It follows from (8) that kX (t) ≡ 0 for
m > 1, hence ρ(2) (Y ) = 0.
From now on let m = 1. Let P = M AN be a fundamental parabolic subgroup
of G. Then (8) gives

dim n
1 X X
kX (t) = (−1)l+1 dim [Wξ ⊗ (Λev p∗m − Λodd p∗m ) ⊗ Λl n∗ ]KM
2
l=0 ξ∈M̂d
Z
2 2
e−t(kνk +kρa k −ξ(ΩM )) pξ (iν)dν . (15)
a∗

Now X = X1 × X0 , X1 = G1 /K1 , X0 = G0 /K0 , m(X0 ) = 0 as explained in

the introduction. Although (15) can be evaluated directly for general X with
m(X) = 1 we prefer to reduce the computation to the irreducible case X = X1 .
In order to compute T (2) (X) it is sufficient to compare ρ(2) (Y ) with vol(Y ) for
one particular Y = Γ\X. If we choose Γ of the form Γ1 × Γ0 , where Γ0 ⊂ G0
and Γ1 ⊂ G1 , then
ρ(2) (Y ) = χ(Y0 )ρ(2) (Y1 ) ,

where Y1 = Γ1 \X1 , Y0 = Γ0 \X0 . Applying Hirzebruch proportionality we

obtain the assertion of Proposition 1.3 (b)

ρ(2) (Y ) χ(Y0 ) (2)

T (2) (X) = = T (X1 )
vol(Y0 )vol(Y1 ) vol(Y0 )
n0
(−1) 2 χ(X0d ) (2)
= T (X1 ) . (16)
vol(X0d )

It remains to deal with the case X = X1 . We can assume that G = SO(p, q)0 ,
p ≤ q odd, or G = SL(3, R). Then M ∼ = SO(p − 1, q − 1), n ∼ = Rp+q−2 or
∼ ∼
M = GL(2, R) := {A ∈ GL(2, R) | | det A| = 1}, n = R , respectively, and
0 2

M acts on n via the standard representation. Note that M is not connected

unless G = SO(1, q)0 . The M 0 -representations Λl n∗ ⊗ C are irreducible unless
G = SO(p, q) and l = u = 12 dim n. In the latter case Λu n∗ ⊗ C decomposes
into two irreducible components Λ+ n ⊕ Λ− n. Since compact Cartan subgroups
of M are connected the discrete series representations of M are induced from
discrete series representations of M 0 : Wξ = IndM 0
M 0 (Wξ 0 ), ξ0 ∈ (M̂ )d (see [19],
∼ KM
6.9 and 8.7.1). As representations of KM we have Wξ = IndK 0 (Wξ0 ). By
M
Frobenius reciprocity we obtain

dim[Wξ ⊗ (Λev p∗m − Λodd p∗m ) ⊗ Λl n∗ ]KM

0
= dim[Wξ0 ⊗ (Λev p∗m − Λodd p∗m ) ⊗ Λl n∗ ]KM .
0
Note that the infinitesimal characters χξ and χξ0 coincide. By χ(m, KM , .)
we denote the Euler characteristic of relative Lie algebra cohomology. Set

Documenta Mathematica 7 (2002) 219–237

 L2 -Invariants of Locally Symmetric Spaces 233

v = 21 dim pm . Applying Propositions 3.1 and 3.2 (a) to M 0 instead of G we

obtain
0
dim[Wξ0 ⊗ (Λev p∗m − Λodd p∗m ) ⊗ Λ∗ n∗ ]KM
0 ∗ ∗
= χ(m, KM 0 ⊗ Λ n )
, Wξ0 ,KM
½
(−1)v χξ = χΛ∗ n
= .
0 otherwise

Here Λ∗ n∗ , ∗ = l, +, −, denotes an irreducible component of Λl n∗ ⊗ C.

In all cases under consideration the set {α|a | α ∈ ∆+ , α|a 6= 0} consists of a
single element α0 ∈ a∗ . It follows that ρa = uα0 . Moreover, ΩM acts on Λl n
as l(2u − l)kα0 k2 id (compare [18], Lemma 2.5).
In order to evaluate (15) further we have to determine the constant cX in
formula (13). This can be done in complete generality. So for a moment we
drop the assumptions m = 1, X = X1 .

Lemma 5.1 Let P = M AN ⊂ G be fundamental. We retain the notation

d
introduced before (13). Set WA = {k ∈ K|Ad(k)a ⊂ a}/KM , SA = exp(ia)K ⊂
d +
X , and let Φk be a positive root system for (k, t) with corresponding half sum
ρk . Then
Q
1 + hα, ρg i
cX = n+m
Qα∈Φ (17)
|WA |(2π) 2 α∈Φ+ hα, ρk ik
d
1 vol(SA ) 1
= . (18)
|WA | (2π) vol(X d )
m

Proof. Formula (17) is a combination of [8], Thm. 37.1, with [9], Cor. 23.1,
Thm. 24.1 and Thm. 27.3. In order to apply these results correctly one
has to take into account that Harish-Chandra’s and our normalizations of the
measures dg and dν all of them starting from a fixed invariant bilinear form
n−dim a0
on g differ by the factors 2 2 ([8], Section 7 and Lemma 37.2) and (2π)m ,
respectively. On the other hand we have
Y dim K/T vol(T )
hα, ρk i = (2π) 2 (19)
vol(K)
α∈Φ+
k

(see e.g. [8], Lemma 37.4). Here the volumes are the Riemannian ones corre-
sponding to the invariant bilinear form h., .i. Formula (19) holds for any pair
(K, T ) of a connected compact Lie group and a maximal torus. Applying it also
to the pair (Gd , H d ), where H d is the maximal torus of Gd with Lie algebra
t ⊕ ia, we obtain
Q d d
+ hα, ρg i n−m vol(K)vol(H ) n−m vol(S )
Qα∈Φ = (2π) 2 d
= (2π) 2 A
. (20)
α∈Φ+ hα, ρk i vol(G )vol(T ) vol(X d )
k

Documenta Mathematica 7 (2002) 219–237

234 Martin Olbrich

The second equality follows from the fact that the map H d /T → X d ,
d
hT 7→ hK, is an isometric embedding with image SA . This proves (18). 2

In particular, specializing (13) and (18) to the case m = 0 we obtain as a

consequence of Weyl’s dimension formula
Corollary 5.2 Let π be a discrete series representation of G having the same
infinitesimal character as the finite-dimensional representation τ , then
dim τ
pπ = .
vol(X d )
Let us now give the promised analytic proof of (10). For a fixed finite-
dimensional representation τ of G (which still is assumed to be connected)
there are exactly |W (g, t)|/|W (k, t)| equivalence classes of discrete series repre-
sentations with infinitesimal character χτ (see [19], Thm. 8.7.1, or [12], Thm.
12.21). But this quotient of orders of Weyl groups is equal to χ(X d ) (see e.g.
[3]). By (7), (9) and Corollary 5.2 we obtain
(2)
X 1
b n = vol(Y ) pπ = vol(Y )χ(X d ) .
2 vol(X d )
π∈Ĝd ,χπ =χτ0

We return to the evaluation of kX (t) for m = 1, X = X1 . The polynomial pξ

only depends on the infinitesimal character of ξ. By Λ∗ ∈ it∗ , ∗ = l, +, −, we
denote the infinitesimal character of the irreducible M -representation Λ∗ n. We
set
 Q hα,Λl +νi

 α∈Φ+ hα,ρg i G = SL(3, R)

 or l 6= u
pl (ν) := Q hα,Λ+ +νi Q hα,Λ− +νi .

 α∈Φ+ hα,ρg i + α∈Φ+ hα,ρg i

 Q
= 2 α∈Φ+ hα,Λ + +νi
hα,ρg i otherwise

For fixed infinitesimal character there are |W (m, t)|/|WKM | equivalence classes
of discrete series representations, where WKM = {k ∈ KM | Ad(k)t ⊂ t}/T .
Note that there is an embedding W (km , t) ,→ WKM which becomes an isomor-
phism if KM is connected. Furthermore, |WA | = 1 except for G = SO(1, q)0 ,
d
where |WA | = 2. In any case |WKM ||WA | = 2|W (km , t)|. Let XM be the
0 0
compact dual of XM = M/KM = M /KM . Then |W (m, t)|/|WKM ||WA | =
1 d n−1
2 χ(XM ). In addition, u + v = 2 and
d
vol(SA ) 1
= .
2π kα0 k
Summarizing the above discussion we obtain
d X 2u
n−1 χ(XM )
kX (t) = (−1) 2
d
(−1)l+1 kl (t) ,
4kα0 kvol(X )
l=0

Documenta Mathematica 7 (2002) 219–237

 L2 -Invariants of Locally Symmetric Spaces 235

where Z
2
+(u−l)2 kα0 k2 )
kl (t) = e−t(kνk pl (iν)dν .
a∗
For an even polynomial P and c ≥ 0 set
Z ∞
2 2
kP,c (t) = e−t(y +c ) P (iy)dy .
−∞

Then (compare [6], Lemma 2 and Lemma 3, [17], Lemma 6.4, [11], p.332)
µ Z ε ¶ Z ∞
d 1 s−1
kP,c (t)t dt + kP,c (t)t−1 dt
ds |s=0 Γ(s) 0 ε
µZ ε ¶ Z ∞
= s−1
kP,c (t)t dt + kP,c (t)t−1 dt
0 |s=0 ε
Z c
= −2π P (y) dy .
0

Since pl = p2u−l we obtain

Proposition 5.3 Let X = G/K with m(X) = 1, P = M AN ⊂ G a funda-
mental parabolic subgroup, u = 21 dim n, and XM = M/KM . Then
n−1 πQX
T (2) (X) = (−1) 2 d
χ(XM ) ,
vol(X d )
where
u−1
X Z u−l
l
QX = (−1) pl (y · α0 ) dy .
l=0 0

In fact, we have proved the proposition for irreducible X, but (16) shows that
it holds in the general case, too.
Let us first discuss the case G = SO(p, q)0 , X = Xp,q , p ≤ q odd. Then
n = pq and XM = Xp−1,q−1 . Furthermore, the polynomials pl depend only on
the complexification of G, i.e., on p + q. Thus QXp,q = QH p+q−1 . This proves
Proposition 1.3 (a). We emphasize again that QH p+q−1 is a positive rational
number [11]. In fact, Hess-Schick showed that
Z u−l
l
(−1) pl (y · α0 ) dy > 0 , l = 0, . . . , u − 1 .
0

Let X = SL(3, R)/SO(3). Then n = 5 and u = 1. We find that

Z 1
1
QX = y 2 dy = , XMd
= S 2 , χ(XM
d
)=2.
0 3

Using for instance (20) also vol(X d ) can be easily computed. This proves
Proposition 1.4 and finishes the proof of Theorem 1.1.

Documenta Mathematica 7 (2002) 219–237

236 Martin Olbrich

References

[1] A. Borel. The L2 -cohomology of negatively curved Riemannian symmetric

spaces. Ann. Acad. Sci. Fenn. Ser. A I Math. 10 (1985), 95–105.
[2] A. Borel and N. Wallach. Continuous Cohomology, Discrete Subgroups,
and Representations of Reductive Groups. Princeton University Press,
1980.
[3] R. Bott. The index theorem for homogeneous differential operators. In Dif-
ferential and Combinatorial Topology. A Symposion in Honour of Marston
Morse, pages 167–186. Princeton University Press, 1965.
[4] J. Dodziuk. L2 -harmonic forms on rotationally symmetric Riemannian
manifolds. Proc. Amer. Math. Soc. 77 (1979), 395–400.
[5] H. Donnelly. The differential form spectrum of hyperbolic space.
manuscripta math. 33 (1981), 365–385.
[6] D. Fried. Analytic torsion and closed geodesics on hyperbolic manifolds.
Invent. Math. 84 (1986), 523–540.
[7] Harish-Chandra. Discrete series for semisimple Lie groups II. Acta Math.
116 (1966), 1–111.
[8] Harish-Chandra. Harmonic analysis on real reductive groups I. The theory
of the constant term. J. Funct. Anal. 19 (1975), 104–204.
[9] Harish-Chandra. Harmonic analysis on real reductive groups III. The
Maass-Selberg relations and the Plancherel formula. Ann. of Math. 104
(1976), 117–201.
[10] E. Hess. Simpliziales Volumen und L2 -Invarianten bei asphärischen
Mannigfaltigkeiten. Dissertation Universität Mainz, 1998. Available at
wwwmath.uni-muenster.de/u/lueck/publ.
[11] E. Hess and T. Schick. L2 -torsion of hyperbolic manifolds. manuscripta
math. 97 (1998), 329–334.
[12] A. W. Knapp. Representation Theory of Semisimple Lie Groups. An
Overview Based on Examples. Princeton University Press, 1986.
[13] N. Lohoue and S. Mehdi. The Novikov-Shubin invariants for locally sym-
metric spaces. J. Math. Pures Appl. 79, 2 (2000), 111–140.
[14] J. Lott. Heat kernels on covering spaces and topological invariants. J.
Differential Geom. 35 (1992), 471–510.
[15] W. Lück. L2 -invariants of regular coverings of compact manifolds and CW -
complexes. In Handbook of geometric topology, pages 735–817. Elsevier,
Amsterdam, 2002.

Documenta Mathematica 7 (2002) 219–237

 L2 -Invariants of Locally Symmetric Spaces 237

[16] W. Lück. L2 -Invariants: Theory and Applications to Geometry and K-

Theory. Springer Verlag, 2002.

[17] V. Mathai. L2 -analytic torsion. J. Funct. Anal. 107 (1992), 369–386.

[18] H. Moscovici and R. Stanton. R-torsion and zeta functions for locally
symmetric manifolds. Invent. Math. 105 (1991), 185–216.

[19] N. R. Wallach. Real Reductive Groups. Academic Press, 1988.

[20] N. R. Wallach. Real Reductive Groups II. Academic Press, 1992.

Martin Olbrich
Mathematisches Institut
Universität Göttingen
Bunsenstr. 3-5
37073 Göttingen
Germany
olbrich@uni-math.gwdg.de

Documenta Mathematica 7 (2002) 219–237

238

Documenta Mathematica 7 (2002)