Dirac Operators and Spectral Geometry

Joseph C. Várilly

Notes taken by
Pawel Witkowski

January 2006
Contents

Introduction and Overview 4

1 Clifford algebras and spinor representations 5

1.1 Clifford algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 The universality property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 The trace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Periodicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Chirality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.6 Spinc and Spin groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.7 The Lie algebra of Spin(V ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.8 Orthogonal complex structures . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.9 Irreducible representations of Cl(V ) . . . . . . . . . . . . . . . . . . . . . . . 15
1.10 Representations of Spinc (V ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2 Spinor modules over compact Riemannian manifolds 18

2.1 Remarks on Riemannian geometry . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2 Clifford algebra bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3 The existence of Spinc structures . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4 Morita equivalence for (commutative) unital algebras . . . . . . . . . . . . . . 22
2.5 Classification of spinor modules . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.6 The spin connection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.7 Epilogue: counting the spin structures . . . . . . . . . . . . . . . . . . . . . . 31

3 Dirac operators 32
3.1 The metric distance property . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2 Symmetry of the Dirac operator . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3 Selfadjointness of the Dirac operator . . . . . . . . . . . . . . . . . . . . . . . 35
3.4 The Schrödinger–Lichnerowicz formula . . . . . . . . . . . . . . . . . . . . . . 36
3.5 The spectral growth of the Dirac operator . . . . . . . . . . . . . . . . . . . . 38

4 Spectral Growth and Dixmier Traces 41

4.1 Definition of spectral triples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.2 Logarithmic divergence of spectra . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3 Some eigenvalue inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.4 Dixmier traces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

2
5 Symbols and Traces 49
5.1 Classical pseudodifferential operators . . . . . . . . . . . . . . . . . . . . . . . 49
5.2 Homogeneity of distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.3 The Wodzicki residue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.4 Dixmier trace and Wodzicki residue . . . . . . . . . . . . . . . . . . . . . . . 60

6 Spectral Triples: General Theory 62

6.1 The Dixmier trace revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
6.2 Regularity of spectral triples . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
6.3 Pre-C*-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.4 Real spectral triples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.5 Summability of spectral triples . . . . . . . . . . . . . . . . . . . . . . . . . . 77

7 Spectral Triples: Examples 79

7.1 Geometric conditions on spectral triples . . . . . . . . . . . . . . . . . . . . . 79
7.2 Isospectral deformations of commutative spectral triples . . . . . . . . . . . . 82
7.3 The Moyal plane as a nonunital spectral triple . . . . . . . . . . . . . . . . . 87
7.4 A geometric spectral triple over SUq (2) . . . . . . . . . . . . . . . . . . . . . 93

A Exercises 94
A.1 Examples of Dirac operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
A.1.1 The circle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
A.1.2 The (flat) torus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
A.1.3 The Hodge–Dirac operator on S2 . . . . . . . . . . . . . . . . . . . . . 96
A.2 The Dirac operator on the sphere S2 . . . . . . . . . . . . . . . . . . . . . . . 98
A.2.1 The spinor bundle S on S2 . . . . . . . . . . . . . . . . . . . . . . . . 98
A.2.2 The spin connection ∇S over S2 . . . . . . . . . . . . . . . . . . . . . 99
A.2.3 Spinor harmonics and the Dirac operator spectrum . . . . . . . . . . . 101
A.3 Spinc Dirac operators on the 2-sphere . . . . . . . . . . . . . . . . . . . . . . 102
A.4 A spectral triple on the noncommutative torus . . . . . . . . . . . . . . . . . 104

References 108

3
Introduction and Overview

Noncommutative geometry asks: “What is the geometry of the Quantum World?”

Quantum field theory considers aggregates of “particles”, which are of two general species,
“bosons” and “fermions”. These are described by solutions of (relativistic) wave equations:
• Bosons: Klein–Gordon equation, ( + m2 )φ(x) = ρb (x) – “source term”;
• Fermions: Dirac equation, (i∂/ − m)ψ(x) = ρf (x) – “source term”;
P3
where x = (t, ~x) = (x0 , x1 , x2 , x3 );  = −∂ 2 /∂t2 + ∂ 2 /∂~x2 ; and ∂/ = µ µ
µ=0 γ ∂/∂x . In
order that ∂/ be a “square root of ”, we need (γ 0 )2 = −1, (γ j )2 = +1 for j = 1, 2, 3 and
γ µ γ ν = −γ ν γ µ for µ 6= ν. Thus, the γ µ must be matrices; in fact there are four (4 × 4)
matrices satisfying these relations.
Point-like measurements are often ruled out by quantum mechanics; thus we replace points
x ∈ M by coordinates f ∈ C(M ). The metric distance on a Riemannian manifold (M, g) can
be computed in two ways:
dg (p, q) := inf{ length(γ : [0, 1] → M ) : γ(0) = p, γ(1) = q }
/ f k ≤ 1 },
= sup{ |f (p) − f (q)| : f ∈ C(M ), kD,
where D/ is a Dirac operator with positive-definite signature (all (γ µ )2 = +1) if it exists, so
the Dirac operator specifies the metric. D/ is an (unbounded) operator on a Hilbert space
H = L2 (M, S) of “square-integrable spinors” and C ∞ (M ) also acts on H by multiplication
/ f ]k = k grad f k∞ .
operators with k[D,
Noncommutative geometry generalizes (C ∞ (M ), L2 (M, S), D) / to a spectral triple of
the form (A, H, D), where A is a “smooth” algebra acting on a Hilbert space H, D is an
(unbounded) selfadjoint operator on H, subject to certain conditions: in particular that [D, a]
be a bounded operator for each a ∈ A. The tasks of the geometer are then:
1. To describe (metric) differential geometry in an operator language.
2. To reconstruct (ordinary) geometry in the operator framework.
3. To develop new geometries with noncommutative coordinate algebras.
The long-term goal is to geometrize quantum physics at very high energy scales, but we
are still a long way from there.
The general program of these lectures is as follows.
(A) The classical theory of spinors and Dirac operators in the Riemannian case.
(B) The operational toolkit for noncommutative generalization.
(C) Reconstruction: how to recover differential geometry from the operator framework.
(D) Examples of spectral triples with noncommutative coordinate algebras.

4
Chapter 1

Clifford algebras and spinor

representations

Here are a few general references on Clifford algebras, in reverse chronological order: [GVF,
2001], [Fri, 1997/2000], [BGV, 1992], [LM, 1989] and [ABS, 1964]. (See the bibliography for
details.)

1.1 Clifford algebras

We start with (V, g), where V ≃ Rn and g is a nondegenerate symmetric bilinear form.
If q(v) = g(v, v), then 2g(u, v) = q(u + v) − q(u) − q(v). Thus g is determined by the
corresponding “quadratic form” q.
Definition 1.1. The Clifford algebra Cl(V, g) is an algebra (over R) generated by the
vectors v ∈ V subject to the relations uv + vu = 2g(u, v)1 for u, v ∈ V .
The existence of this algebraLcan be seen in two ways. First of all, let T (V ) be the tensor
algebra on V , that is, T (V ) := ∞k=0 V
⊗n . Then

Cl(V, g) := T (V )/ Idealhu ⊗ v + v ⊗ u − 2g(u, v) 1 : u, v ∈ V i. (1.1)

Since the relations are not homogeneous, the Z-grading of T (V ) is lost, only a Z2 -grading
remains:
Cl(V, g) = Cl0 (V, g) ⊕ Cl1 (V, g).
The second option is to define Cl(V, g) as a subalgebra of EndR (Λ• V ) generated by all
expressions c(v) = ε(v) + ι(v) for v ∈ V , where

ε(v) : u1 ∧ · · · ∧ uk 7→ v ∧ u1 ∧ · · · ∧ uk
k
X
ι(v) : u1 ∧ · · · ∧ uk 7→ (−1)j−1 g(v, uj )u1 ∧ · · · ∧ ubj ∧ · · · ∧ uk .
j=1

Note that ε(v)2 = 0, ι(v)2 = 0, and ε(v)ι(u) + ι(u)ε(v) = g(v, u) 1. Thus

c(v)2 = g(v, v) 1 for all v ∈ V,

c(u)c(v) + c(v)c(u) = 2g(u, v) 1 for all u, v ∈ V.

Thus these operators on Λ• V do provide a representation of the algebra (1.1).

5
 Dimension count: suppose {e1 , . . . , en } is an orthonormal basis for (V, g), i.e., g(ek , ek ) =
±1 and g(ej , ek ) = 0 for j 6= k. Then the c(ej ) anticommute and thus a basis for Cl(V, g)
is {c(ek1 ) . . . c(ekr ) : 1 ≤ k1 < · · · < kr ≤ n}, labelled by K = {k1 , . . . , kr } ⊆ {1, . . . , n}.
Indeed,
c(ek1 ) . . . c(ekr ) : 1 7→ ek1 ∧ · · · ∧ ekr ≡ eK ∈ Λ• V
and these are linearly independent. Thus the dimension of the subalgebra of EndR (Λ• V )
generated by all c(v) is just dim Λ• V = 2n . Now, a moment’s thought shows that in the
abstract presentation (1.1), the algebra Cl(V, g) is generated as a vector space by the 2n
products ek1 ek2 . . . ekr , and these are linearly independent since the operators c(ek1 ) . . . c(ekr )
are linearly independent in EndR (Λ• V ). Therefore, this representation of Cl(V, g) is faithful,
and dim Cl(V, g) = 2n .
The so-called “symbol map”:

σ : a 7→ a(1) : Cl(V, g) → Λ• V

is inverted by a “quantization map”:

1 X
Q : u1 ∧ u2 ∧ · · · ∧ ur 7−→ (−1)τ c(uτ (1) )c(uτ (2) ) . . . c(uτ (r) ). (1.2)
r!
τ ∈Sr

To see that it is an inverse to σ, one only needs to check it on the products of elements of an
orthonormal basis of (V, g).
From now, we write uv instead of c(u)c(v), etc., in Cl(V, g).

1.2 The universality property

Chevalley [Che] has pointed out the usefulness of the following property of Clifford algebras,
which is an immediate consequence of their definition.
Lemma 1.2. Any R-linear map f : V → A (an R-algebra) that satisfies

f (v)2 = g(v, v) 1A for all v∈V

extends to an unique unital R-algebra homomorphism f˜: Cl(V, g) → A.

Proof. There is really nothing to prove: f˜(v1 v2 . . . vr ) := f (v1 )f (v2 ) . . . f (vr ) gives the unique-
ness, provided only that this recipe is well-defined. But observe that

f˜(uv + vu − 2g(u, v)1) = f˜((u + v)2 − u2 − v 2 ) − 2g(u, v) f˜(1)

= [q(u + v) − q(u) − q(v) − 2g(u, v)] 1A = 0.

Here are a few applications of universality that yield several useful operations on the
Clifford algebra.
1. Grading: take A = Cl(V, g) itself; the linear map v 7→ −v on V extends to an automor-
phism χ ∈ Aut(Cl(V, g)) satisfying χ2 = idA , given by

χ(v1 . . . vr ) := (−1)r v1 . . . vr .

This operator gives the Z2 -grading

Cl(V, g) =: Cl0 (V, g) ⊕ Cl1 (V, g).

6
 2. Reversal : take A = Cl(V, g)◦ , the opposite algebra. Then the map v 7→ v, considered
as the inclusion V ֒→ A, extends to an antiautomorphism a 7→ a! of Cl(V, g), given by
(v1 v2 . . . vr )! := vr . . . v2 v1 .

3. Complex conjugation: the complexification of Cl(V, g) is Cl(V, g) ⊗R C, which is iso-

morphic to Cl(V C , g C ) as a C-algebra. Now take A to be Cl(V, g) ⊗R C and define
f : v 7→ v̄ : V C → V C ֒→ A (a real-linear map). It extends to an antilinear automor-
phism of A. Note that Lemma 1.2 guarantees R-linearity, but not C-linearity, of the
extension even when A is a C-algebra.

4. Adjoint: Also, a∗ := (ā)! is an antilinear involution on Cl(V, g) ⊗R C.

5. Charge conjugation: κ(a) := χ(ā) : v1 . . . vr 7→ (−1)r v̄1 . . . v̄r is an antilinear automor-

phism of Cl(V, g) ⊗R C.

Notation. We write Cl(V ) := Cl(V, g)⊗R C to denote the complexified Clifford algebra. Up to
isomorphism, this is independent of the signature of the symmetric bilinear form g, because
all complex nondegenerate bilinear forms are congruent.

1.3 The trace

Proposition 1.3. There is an unique trace τ : Cl(V ) → C such that τ (1) = 1 and τ (a) = 0
for a odd.

Proof. If {e1 , . . . , en } is an orthonormal basis for (V, g), then

τ (ek1 . . . ek2r ) = τ (−ek2 . . . ek2r ek1 ) = −τ (ek1 . . . ek2r ) = 0.

(Here we have moved ek1 to thePright by anticommutation, and returned it to the left with the
trace property.) Thus, if a = K even ak1 ...k2r ek1 . . . ek2r lies in Cl0 (V ), then τ (a)
P = a∅ . We
will check that a∅ does not depend on the orthonormal basis used. Suppose e′j = nk=1 hkj ej ,
with H t H = 1n , is another orthonormal basis. Then
X
e′i e′j = (~hi · ~hj ) 1 + ckl
ij ek el ,
k<l


but ~hi · ~hj = [H t H]ij = 0 for i 6= j. Next, the matrix of ek el 7→ e′i e′j is H ∧ H, of size n2 ,
that is also orthogonal, so e′i e′j e′r e′s has zero scalar part in the ek el ep eq -expansion; and so on:
the same is true for expressions e′j1 . . . e′j2r by induction. Thus τ (a) = a∅ does not depend on
{e1 , . . . , en }.

Remark 1.4. At this point, it was remarked that for existence of the trace, one could use the
restriction of the (normalized) trace on EndR (Λ• V ) ⊗R C = EndC (Λ• V C ), in which Cl(V ) is
embedded. True enough: although one must see why odd elements must have trace zero. For
that, it is enough to note that if a ∈ Cl1 (V ), then c(a) takes even [respectively, odd] elements
of the Z-graded algebra EndC (V C ) to odd [respectively, even] elements; thus, in any basis,
the matrix of c(a) will have only zeroes on the diagonal, so that tr(c(a)) = 0. Nonetheless,
Proposition 1.3 is useful in that it establishes the uniqueness of the trace.
Now Cl(V ) is a Hilbert space with scalar product

ha | bi := τ (a∗ b).

7
1.4 Periodicity
Write Clpq := Cl(Rp+q , g), where g has signature (p, q), and the orthonormal basis is written
as {e1 , . . . , ep , ε1 , . . . , εq }, where e21 = · · · = e2p = 1 and ε21 = · · · = ε2q = −1. For example,
Cl10 = R ⊕ R;
Cl01 = C, with ε1 = i;
     
0 1 1 0 0 −1
Cl20 = M2 (R), with e1 = , e2 = , e1 e2 = ;
1 0 0 −1 1 0
Cl02 = H, with ε1 = i, ε2 = j, ε1 ε2 = k.
Lemma 1.5 (“(1,1)-periodicity”). Clp+1,q+1 ≃ Clpq ⊗M2 (R).
Proof. Take V = Rp+q+2 , A = Clpq ⊗M2 (R). Define f : V → A on basic vectors by
 
1 0
f (er ) := er ⊗ , r = 1, . . . , p,
0 −1
 
1 0
f (εs ) := εs ⊗ , s = 1, . . . , q,
0 −1
 
0 1
f (ep+1 ) := 1 ⊗ ,
1 0
 
0 −1
f (εq+1 ) := 1 ⊗ . (1.3)
1 0
Thus f (ek )2 = +1, f (εl )2 = −1 in all cases, and all f (ek ), f (εl ) anticommute. This entails
that f extends by linearity to a linear map satisfying f (v)2 = g(v, v) 1 for all v ∈ V . Hence
there exists a homomorphism f˜: Clp+1,q+1 → A, which is surjective since the right hand
sides of (1.3) generate A as an R-algebra. It is an isomorphism, because the dimensions over
R are equal.

Lemma 1.6. Cl0p+1,q ≃ Clqp .

Proof. Define f : Rq+p → Cl0p+1,q on basic vectors by
f (er ) := er ep+1 , r = 1, . . . , q,
f (εs ) := εs ep+1 , s = 1, . . . , p.
Then
f (er )2 = er ep+1 er ep+1 = −e2r e2p+1 = −e2r = −1,
f (εs )2 = εs ep+1 εs ep+1 = −ε2s e2p+1 = −ε2s = +1,
and all f (er ), f (εs ) anticommute. The rest of the proof is like that of the previous Lemma.

Lemma 1.7. Clp+4,q ≃ Clpq ⊗M2 (H) ≃ Clp,q+4 .

Proof. We will prove the first isomorphism. Take A = Clpq ⊗M2 (H); define f : Rp+4+q → A
by
 
0 −k
f (er ) := er ⊗ , r = 1, . . . , p,
k 0
 
0 −k
f (εs ) := εs ⊗ , s = 1, . . . , q,
k 0

8
and on the remaining four basic vectors, define
   
0 −i 0 −j
f (ep+1 ) := 1 ⊗ , f (ep+2 ) := 1 ⊗ ,
i 0 j 0
   
0 1 1 0
f (ep+3 ) := 1 ⊗ , f (ep+4 ) := 1 ⊗ .
1 0 0 −1

Corollary 1.8 (“(+8)-periodicity”). Clp+8,q ≃ Clpq ⊗M16 (R) ≃ Clp,q+8 .

Proof. This reduces to M2 (H)⊗R M2 (H) ≃ M16 (R), that in turn reduces to H⊗R H ≃ M4 (R),
which is left as an exercise.

All Clpq are given, up to MN (R) tensor factors, by Clp0 for p = 1, . . . , 8:

Cl10 = R ⊕ R
Cl20 = M2 (R)
Cl30 = M2 (C)
Cl40 = M2 (H)
Cl50 = M2 (H) ⊕ M2 (H)
Cl60 = M4 (H)
Cl70 = M8 (C)
Cl80 = M16 (R) (1.4)

Two algebras Cl10 and Cl50 are direct sums of simple algebras, and the others are simple. We
could also define Cl00 = R (the base field), so that Corollary 1.8 holds even when p = q = 0.
Those eight algebras Clp0 can be arranged on a “spinorial clock”, which is taken from
Budinich and Trautman’s book [BT].

C
✛
6

7
✛

H R
✛
5

0
✛

H⊕H R⊕R
✲
4

1
✲

H R
✲
3

2
✲

C
If p − q ≡ m mod 8, then Clpq is of the form A ⊗ MN (R), where A is the diagram entry at
the head of the arrow labelled m. Moreover, Lemma 1.6 says that the even subalgebra Cl0pq
is of the same kind, where A is now the diagram entry at the tail of the arrow labelled m.
The matrix size N is easily determined from the real dimension, in each case. In this way,
the spinorial clock displays the full classification of real Clifford algebras.

9
1.5 Chirality
From now on, n = 2m for n even, n = 2m + 1 for n odd. We take Cl(V ) ≃ Cl(V, g) ⊗R C
with g always positive definite. P
Suppose {e1 , . . . , en } is an oriented orthonormal basis for (V, g). If e′k = nj=1 hjk ej with
H t H = 1n , then e′1 . . . e′n = (det H) e1 . . . en , and det H = ±1. We restrict to the oriented
case det H = +1, so the expression e1 e2 . . . en is independent of {e1 , e2 , . . . , en }. Thus

γ := (−i)m e1 e2 . . . en

is well-defined in Cl(V ). Now

γ ∗ = im en . . . e2 e1 = (−i)m (−1)m (−1)n(n−1)/2 e1 e2 . . . en = (−1)m (−1)n(n−1)/2 γ,

and ( )
n(n − 1) m(2m − 1), n even
= ≡ m mod 2,
2 (2m + 1)m, n odd
so γ ∗ = γ. But also γ ∗ γ = (en . . . e2 e1 )(e1 e2 . . . en ) = (+1)n = 1, so γ is “unitary”. Hence
γ 2 = 1, so 1+γ 1−γ
2 , 2 are “orthogonal projectors” in Cl(V ).
Since γej = (−1)n−1 ej γ, we get that if n is odd, then γ is central in Cl(V ); and for n
even, γ anticommutes with V , but is central in the even subalgebra Cl0 (V ). Moreover, when
n is even and v ∈ V , then γvγ = −v, so that γ(·)γ = χ ∈ Aut(Cl(V )).

Proposition 1.9. The centre of Cl(V ) is C1 if n is even; and it is C1 ⊕ Cγ if n is odd.

Proof. Denote this centre by Z(Cl(V )).

2 0
Even case:
P a ∈ Z(Cl(V )) implies γaγ = aγ = a, soPa lies in Cl (V ).
If a = K even aK ek1 . . . ekr , then 0 = a − ej aej = K even, j∈K 2aK ek1 . . . ekr , so aK = 0
if j ∈ K. Since this holds for any j, we conclude that a = a∅ 1 = τ (a) 1. Therefore
Z(Cl(V )) ≃ C1 when n is even.
Odd case: If a = a0 + a1 (even + odd) lies in Z(Cl(V )), then

0 = [a, v] = [a1 , v] + [a0 , v] for all v ∈ V,

| {z } | {z }
even odd

so [a0 , v] = [a1 , v] = 0 for all v ∈ V . In particular, a0 ∈ Z(Cl0 (V )) ≃ C1, and thus a0 = τ (a) 1.
Also, a1 γ is even and central, so a1 γ = τ (aγ) 1 and a1 = τ (aγ) γ. Thus Z(Cl(V )) =
C1 ⊕ Cγ when n is odd.

1.6 Spinc and Spin groups

Let v be a unit vector, g(v, v) = 1. Then v 2 = 1 in Cl(V, g), so v = v ∗ and v ∗ v = vv ∗ = 1 in
Cl(V ). If w = λv ∈ V C with |λ| = 1, then ww∗ = w∗ w = 1 in Cl(V ) also. Now

hwa | wbi = τ (a∗ w∗ wb) = τ (a∗ b) = ha | bi,

haw | bwi = τ (w∗ a∗ bw) = τ (ww∗ a∗ b) = τ (a∗ b) = ha | bi,

so a 7→ wa, a 7→ aw are unitary operators in L(Cl(V )).

[[ Exercise: Conversely, if u ∈ Cl(V ) and a 7→ ua and a 7→ au are both unitary, then
u∗ u = uu∗ = 1. ]]

10
 If v, x ∈ V , with g(v, v) = 1, then

−vxv −1 = −vxv = (xv − 2g(v, x))v = x − 2g(v, x)v ∈ V .

This is a reflection of x in the hyperplane orthogonal to v. For w = λv, |λ| = 1 we also get
−wxw−1 = −λλ̄vxv −1 = −vxv −1 , which is the same as above. If a = w1 . . . wr is a product
of unit vectors in V C , then

χ(a)xa−1 = (−1)r w1 . . . wr xwr−1 . . . w1−1

is a product of r reflections of x ∈ V . If r = 2k is even, and a = w1 . . . w2k , then axa−1 ∈ V

after k rotations. Thus φ(a) : x 7→ axa−1 lies in SO(V ) = SO(V, g).
Definition 1.10. The set of all even products of unitary vectors,

Spinc (V ) := { u = w1 . . . w2k : wj ∈ V C , wj∗ wj = 1, k = 0, 1, . . . , m },

is a group included in Cl0 (V ), and φ : Spinc (V ) → SO(V ) is a group homomorphism.

The inverse of u = w1 . . . w2k is u−1 = u∗ = w̄2k . . . w̄1 .
Suppose u ∈ ker φ, which means that uxu−1 = x for all x ∈ V . Thus ker φ ⊂ Z(Cl(V ))
for n even, and ker φ ⊂ Z(Cl0 (V )) for n odd; in both cases, u lies in C1. It follows that
ker φ ≃ { λ ∈ C : |λ| = 1 } = T = U(1). Therefore, there is a short exact sequence (SES) of
groups:
φ
1 → T → Spinc (V ) −
→ SO(V ) → 1. (1.5)
If u = w1 . . . w2k ∈ Spinc (V ) with wj = λj vj where λj ∈ T and vj ∈ V , then u! =
w2k . . . w1 , and u! u = λ21 λ22 . . . λ2n ∈ T. Thus, u! u is central, so (u1 u2 )! u1 u2 = u!2 u!1 u1 u2 =
u!1 u1 u!2 u2 , so that u 7→ u! u is a homomorphism ν : Spinc (V ) → T, which restricts to T ⊂
Spinc (V ) as λ 7→ λ2 . The combined (φ, ν) : Spinc (V ) → SO(V ) × T is a homomorphism with
kernel {±1}.
Definition 1.11. Spin(V ) := ker ν ≤ Spinc (V ).
Indeed Spin(V ) is included in (the even part of) the real Clifford algebra Cl0 (V, g):

u∗ u = 1, u! u = 1 =⇒ u∗ = u! =⇒ ū = u =⇒ u ∈ Cl0 (V, g).

The SES (1.5) now becomes

φ
1 → {±1} → Spin(V ) −
→ SO(V ) → 1, (1.6)

so that φ is a double covering of SO(V ). Furthermore, Spinc (V ) ≃ Spin(V ) ×Z2 T.

Example 1.12. Case n = 2: We write Spin(n) ≡ Spin(Rn ). It is easy to check that

Spin(V ) = { α + βe1 e2 : α, β ∈ R, α2 + β 2 = 1 }
ψ ψ
= { u = cos + sin e1 e2 : −2π < ψ ≤ 2π } ≃ T.
2 2
We compute
 ψ ψ   ψ ψ 
ue1 u−1 = cos + sin e1 e2 e1 cos − sin e1 e2 = (cos ψ) e1 − (sin ψ) e2 ,
2 2 2 2
 ψ ψ   ψ ψ 
ue2 u−1 = cos + sin e1 e2 e2 cos − sin e1 e2 = (sin ψ) e1 + (cos ψ) e2 ,
2 2 2 2

11
so that  
cos ψ − sin ψ
φ(u) = ∈ SO(2),
sin ψ cos ψ

which is (nontrivial) double covering of the circle.

Example 1.13. Spin(3) ≃ SU(2) = {unit quaternions} in Cl030 ≃ Cl02 ≃ H, and φ : SU(2) →
SO(3) is the adjoint representation of SU(2).
Example 1.14. Spin(4) ≃ SU(2) × SU(2) in Cl040 ≃ Cl03 ≃ H ⊕ H. If u = (q, p) with
q, p ∈ SU(2), then φ(u) becomes x 7→ qxp−1 for x ∈ H ≃ R4 , and this map lies in SO(4). If
φ(u) = 1H , then 1 7→ qp−1 , so p = q, and x 7→ qxq −1 = x, so q is central; hence p = q = ±1
and φ is indeed a double covering of SO(4).

1.7 The Lie algebra of Spin(V )

Recall the linear isomorphism Q : Λ• V → Cl(V, g), inverse to σ : a 7→ c(a)1. Write

b = Q(u ∧ v) = 12 (uv − vu) = uv + g(u, v) 1 ∈ Cl0 (V, g).

Note in passing that b! = 21 (vu − uv) = −b.

Although the algebra Cl(V, g) is not Z-graded, it is Z-filtered : we may write Cl≤k (V, g)
to denote the vector subspace generated by products of at most k vectors from V . With
that notation, the subspace Q(Λ2 V ) may also be described as the set of all even elements
b ∈ Cl≤2 (V, g) with τ (b) = 0.
For x ∈ V , we compute

[b, x] = [uv, x] = uvx + uxv − uxv − xuv = 2g(v, x)u − 2g(u, x)v ∈ V ,

so ad b : V → V . Also

[b, b′ ] = 12 [b, u′ v ′ − v ′ u′ ] = 21 [b, u′ ]v ′ + 21 u′ [b, v ′ ] − 21 [b, v ′ ]u′ − 12 v ′ [b, u′ ]

so that [b, b′ ] ∈ Cl≤2 (V, g) with τ ([b, b′ ]) = 0. Hence [b, b′ ] ∈ Q(Λ2 V ), and this is a Lie algebra.
Next,
g(y, [b, x]) = 2g(v, x)g(y, u) − 2g(u, x)g(y, v) = −g([b, y], x),

so that ad b is skewsymmetric: thus ad b ∈ so(V ). By the Jacobi identity,

[[b, b′ ], x] = [b, [b′ , x]] − [b′ , [b, x]] for all x ∈ V,

and so ad([b, b′ ]) = [ad b, ad b′ ]. Thus, ad : Q(Λ2 V ) → so(V ) is a Lie algebra homomorphism.

If ad b = 0, so that [b, x] = 0 for all x ∈ V , then b ∈ Z(Cl0 (V )) ≃ C1. But τ (b) = 0
then implies b = 0, so ad is injective. Since dim Λ2 V = n(n − 1)/2 = dim so(V ), we see that
ad : Q(Λ2 V ) → so(V ) is a Lie algebra isomorphism.
There is an important formula for the inverse of ad. For A ∈ so(V ), define
n
1 X 1X
µ̇(A) = g(ej , Aek ) ej ek = g(ej , Aek ) ej ek . (1.7)
4 2
j,k=1 j<k

12
Since τ (µ̇(A)) = 0, we get µ̇(A) ∈ Q(Λ2 V ). Also
1X

[µ̇(A), er ] = g(ej , Aek ) ej {ek , er } − {ej , er } ek
4 | {z } | {z }
j,k
δkr δjr
1X 1X X
= g(ej , Aer ) ej − g(er , Aek ) ek = g(ej , Aer ) ej
2 2
j k j

= Aer ,
where we have used the anticommutator notation {X, Y } := XY + Y X. Hence ad(µ̇(A)) =
A ∈ so(V ). P
Now consider u = exp b := 1 + k≥1 k! b ∈ Cl0 (V, g) for b ∈ Q(Λ2 V ). Then u∗ u = u! u =
1 k

exp(−b) exp b = 1 since b! = −b. Also, u is unitary and even, and if x ∈ V then
X 1
uxu−1 = bk x(−b)l
k!l!
k,l≥0
X 1 X r  
r k
= b x(−b)r−k
r! k
r≥0 k=0
X 1
= (ad b)r (x) ∈ V ,
r!
r≥0

and thus u = exp(b) lies in Spin(V ). When b = µ̇(A), we get φ(exp(b)) = exp(ad b) = exp(A),
and it is known that exp : so(V ) → SO(V ) is surjective (a property of compact connected
matrix groups).
Now exp(Q(Λ2 V )) is a subset of Spin(V ) covering all of SO(V ). If we can show that −1 =
exp c for some c, then − exp b = (exp b)(exp c) = exp(b + c), provided that c, b commute. If
b = µ̇(A), we can express the skewsymmetric matrix A as a direct sum of 2×2 skewsymmetric
blocks in a suitable orthonormal basis:
 
0 ∗
∗ 0 
 
 0 ∗ 
 
 ∗ 0 
 
A=  .. .
. 
 
 0 ∗ 
 
 ∗ 0 
 
..
.

That is, we can choose the (oriented) orthonormal basis {e1 , . . . , en } so that
b = 21 g(e1 , Ae2 ) e1 e2 + 21 g(e3 , Ae4 ) e3 e4 + · · · + 12 g(e2r−1 , Ae2r ) e2r−1 e2r
with r ≤ m. Now this particular e1 e2 commutes with b: (e1 e2 )b = b(e1 e2 ); take c := πe1 e2 .
Then exp c = exp(πe1 e2 ) = cos π + sin πe1 e2 = −1. We have shown that exp : Q(Λ2 V ) →
Spin(V ) is surjective.
Note that t 7→ exp(te1 e2 ), for 0 ≤ t ≤ π, is a path in Spin(V ) from +1 to −1. Since
π1 (SO(V )) ≃ Z2 for n ≥ 3, the double covering Spin(V ) → SO(V ) is nontrivial. We get an
important consequence.
Corollary 1.15. Spin(n) is simply connected, for n ≥ 3.

13
1.8 Orthogonal complex structures
Suppose that n = 2m is even, V ≃ R2m . Then V can be identified with Cm , but not
canonically.
Definition 1.16. An operator J ∈ EndR V is called an orthogonal complex structure,
written J ∈ J (V, g), if
(a) J 2 = −1 in EndR V ;

(b) g(Ju, Jv) = g(u, v) for all u, v ∈ V .

Then also g(Ju, v) = −g(Ju, J 2 v) = −g(u, Jv), so that J is skewsymmetric with respect
to g: J t = −J. Note that (b) says that J t J = 1.
We can now make V a C-module by setting iv := Jv, that is,

(α + iβ)v := αv + βJv for all α, β ∈ R.

We define a hermitian scalar product on V by

hu | viJ := g(u, v) + i g(Ju, v)

Note that hJu | viJ = −ihu | viJ and hu | JviJ = +ihu | viJ (check it!). We denote the resulting
m-dimensional complex Hilbert space by VJ .
If {u1 , . . . , um } is an orthonormal basis for VJ := (V, h·|·iJ ), then {u1 , Ju1 , . . . , um , Jum } is
an orthonormal oriented basis for V (over R). The orientation may or may not be compatible
with the given one on V .
Exercise 1.17. If 2m = 4, show that all such J can be parametrized by two disjoint copies
of S2 , one for each orientation.
If J 2 = −1, J t J = 1 and if h ∈ O(n) = O(V, g) is an orthogonal linear transformation,
then K := hJh−1 is also an orthogonal complex structure. In that case,

hhu | hviK = g(hu, hv) + i g((hJh−1 )hu, hv)

= g(u, v) + i g(Ju, v) = hu | viJ ,

so that h : VJ → VK is unitary. Thus hJh−1 = J if and only if h ∈ U(VJ ) ≃ U(m). In short:

O(n) = O(2m) acts transitively on J (V, g) with isotropy subgroups isomorphic to U(m).
Hence, as a manifold,
J (V, g) ≈ O(2m)/ U(m).
Those J which are compatible with orientation on V form one component (of two), homeo-
morphic to SO(2m)/ U(m).
We may complexify V to get V C = V ⊗R C = V ⊕ iV . Take

WJ := { v − iJv ∈ V C : v ∈ V } = 12 (1 − iJ)V = PJ V.

This is an isotropic subspace for the symmetric billinear form g C on V C .

g(u − iJu, v − iJv) = g(u, v) − ig(u, Jv) − ig(Ju, v) − g(Ju, Jv) = 0.

The conjugate subspace

W J = { v + iJv : v ∈ V } = 12 (1 + iJ)V = PJ V

14
satisfies WJ ⊕ W J ≃ V C , an orthogonal direct sum for the hermitian scalar product

hhw | zii := 2 g(w̄, z) for w, z ∈ V C .

Note that PJ2 = PJ and PJ = PJ∗ with respect to this product. We say that WJ is a
polarization of V C . Also PJ : VJ → WJ is an unitary isomorphism.
Conversely: given a splitting V = W ⊕ W , orthogonal with respect to hh· | ·ii, write
w =: u − iv for w ∈ W , with u, v ∈ V ; then JW : u 7→ v lies in J (V, g), and WJW = W
(exercise). Thus the correspondence J ↔ WJ is bijective.

1.9 Irreducible representations of Cl(V )

We continue to suppose that n = 2m is even.
Definition 1.18. The (fermionic) Fock space corresponding to J ∈ J (V, g) is defined as

FJ (V ) := Λ• WJ ,

with hermitian scalar product given by

 
hhw1 ∧ · · · ∧ wk | z1 ∧ · · · ∧ zl ii := δkl det hhwi | zj ii . (1.8)

This is a complex Hilbert space of dimension 2m . Choose and fix a unit vector Ω ∈ Λ0 WJ :
it is unique up to a factor λ ∈ T. For w ∈ WJ (so that w̄ ∈ W J = WJ⊥ ), we write

ε(w) : z1 ∧ · · · ∧ zk 7→ w ∧ z1 ∧ · · · ∧ zk ,
k
X
ι(w̄) : z1 ∧ · · · ∧ zk 7→ (−1)k−1 hhw | zj ii z1 ∧ · · · ∧ zbj ∧ · · · ∧ zk .
j=1

For v ∈ V , write w = 12 (v − iJv) = PJ v ∈ WJ and define

cJ (v) := ε(w) + ι(w̄) = ε(PJ v) + ι(P−J v).

Then
c2J (v) := hhw | wii 1 = hv | viJ 1 = g(v, v) 1,
so that cJ : V → EndC (FJ V ) ≡ L(FJ V ). That is to say, cJ is a representation of Cl(V ) on
the Hilbert space FJ V .
Note that we complexify the representation of Cl(V, g), given by universality. One can
check that
cJ (w) = ε(w) if w ∈ Wj ; cJ (z̄) = ι(z̄) if z̄ ∈ W J .
From (1.8) and the properties of determinants, it is easy to check that the operators ε(w) and
ι(w̄) are adjoint to one another, that is, ε(w)† = ι(w̄) for w ∈ WJ ; in particular, cJ (v)† = cJ (v)
for v ∈ V . (This is a consequence of our choice of g to have positive definite signature: were
we to have taken g to be negative definite, as in done in many books, then the operators cJ (v)
would have been skewadjoint.) More generally, we get cJ (a)† = cJ (a∗ ) for a ∈ Cl(V ): we say
that cJ is a selfadjoint representation of the ∗-algebra Cl(V ) on the Fock space FJ (V ).
Now, if T ∈ L(FJ (V )) commutes with cJ (V C ), then in particular ι(z̄)T Ω = T ι(z̄)Ω =
T (0) = 0 for z̄ ∈ W J . Therefore T Ω ∈ Λ0 WJ , i.e., T Ω = tΩ for some t ∈ C. Now

T (w1 ∧ · · · ∧ wk ) = T ε(w1 ) . . . ε(wk )Ω = ε(w1 ) . . . ε(wk )T Ω = t w1 ∧ · · · ∧ wk

15
for w1 , . . . , wk ∈ WJ . Thus T = t 1 ∈ L(Λ• WJ ). By Schur’s lemma, the representation cJ is
irreducible.
Suppose K ∈ J (V, g) with K = hJh−1 for h ∈ O(2m). Then hJ = Kh, hP±J = P±K h,
and so cK (hv) = (Λ• h) cJ (v). By universality again, we get cK ◦ Λ• h = Λ• h ◦ cJ , so that the
irreducible representations cK and cJ are equivalent.
The Fock space is Z2 -graded as Λeven WJ ⊕ Λodd WJ . ¿What operator determines its Z2 -
grading? In fact, this operator is cJ (γ). To see that, write γ = (−1)m e1 e2 . . . e2m , where
e2j = Je2j−1 for j = 1, . . . , m. If z1 := PJ e1 = 21 (e1 − ie2 ) we get

z̄1 z1 − z1 z̄1 = 14 (e1 + ie2 )(e1 − ie2 ) − 41 (e1 − ie2 )(e1 + ie2 ) = −e1 e2 .

With zj := PJ e2j−1 = 12 (e2j−1 − ie2j ), this gives

γ = (z̄1 z1 − z1 z̄1 ) . . . (z̄m zm − zm z̄m ) in Cl0 (V ).

Now cJ (z̄j zj − zj z̄j ) = ι(z̄j )ε(zj ) − ε(zj )ι(z̄j ) is the operator

(
−zk1 ∧ · · · ∧ zkr if j ∈ K
z k 1 ∧ · · · ∧ zk r 7 →
−
+zk1 ∧ · · · ∧ zkr if j ∈
/ K.

Thus cJ (γ) acts as (−1)k on Λk WJ : this is indeed the Z2 -grading operator.

Finally, the odd case is treated as follows. Let U := R- span{e1 , . . . , e2m } ≤ V . Then
Cl(U ) ≃ Cl0 (V ) via u 7→ i ueem+1 , extended to Cl(U ). Now FJ (U ) is an irreducible Cl0 (V )-
module, while Z(Cl(V )) = C1 ⊕ Cγ. Since γ 2 = 1, we can extend the action of Cl0 (V ) on
FJ (U ) to the full Cl(V ) by setting either cJ (γ) := +1 or c′J (γ) := −1 on FJ (U ).
These representations cJ , c′J are inequivalent, since T (1) = (−1)T is not possible unless
T = 0, using Schur’s lemma again. Thus Cl(V ) has two irreducible Fock representations of
dimension 2m in the odd case.

Proposition 1.19. The Fock representations yield all irreducible representations of Cl(V ). If
dimR V = 2m, the irreducible representation is unique up to equivalence; if dimR V = 2m + 1,
there are exactly two such representations.

Proof. We have already described and classified the Fock representations. It remains to show
that this list is complete.
We have seen that up to tensoring with a matrix algebra MN (R), the real Clifford algebras
occur in eight species. The periodicity of complex Clifford algebras is much simpler, and may
be obtained from (1.4) by complexifying each algebra found there. Since C ⊗R C ≃ C ⊕ C
and H ⊗R C ≃ M2 (C), we obtain directly that

Cl(R2m ) ≃ M2m (C) and Cl(R2m+1 ) ≃ M2m (C) ⊕ M2m (C). (1.9)

From this it is clear that, when dim V is even, Cl(V ) is a simple matrix algebra and therefore
all irreducible representations are equivalent and arise from matrix multiplication on a mini-
mal left ideal, whose dimension is 2m . Similar arguments in the odd case show that there are
at most two inequivalent representations of Cl(V ). Thus the Fock representations we have
constructed account for all of them: there are no others.

16
1.10 Representations of Spinc (V )
We obtain representations of the group Spinc (V ) by restriction of the irreducible representa-
tions of Cl(V ).
Spinc (V ) = { w1 w2 . . . w2k : wi ∈ V C , wi∗ wi = 1 }.
We have to check whether these restrictions are irreducible or not.
Even case, n = 2m: γ belongs to Spinc (V ) and is central there, so cJ (γ) commutes with
cJ (Spinc (V )). Thus the group representation reduces over Λ• WJ = Λeven WJ ⊕Λodd WJ : there
are two subrepresentations. Since

w1 ∧ · · · ∧ w2k = ε(w1 ) . . . ε(w2k ) = cJ (w1 ) . . . cJ (w2k ) Ω,

we get at once that cJ (Spinc (V )) Ω = Λeven WJ : the “even” subrepresentation is irreducible.

If w1 , w2 ∈ WJ are unit vectors, then

cJ (w2 w̄1 ) w1 = ε(w2 )ι(w̄1 ) w1 = ε(w2 ) Ω = w2 .

From there we soon conclude that cJ (Spinc (V )) w1 = Λodd WJ : the “odd” subrepresentation
is also irreducible.
¿Are these subrepresentations equivalent? No: for suppose R : Λeven WJ → Λodd WJ inter-
twines both subrepresentations. Then in particular RcJ (γ) = cJ (γ)R means that R(+1) =
(−1)R : Λeven WJ → Λodd WJ , so that R = 0.
Conclusion: The algebra representation cJ of Cl(V ) restricts to a group representation
cJ of Spinc (V ) which is the direct sum of two inequivalent irreducible subrepresentations, if
dim V is even.
Odd case, n = 2m + 1: There are two irreducible representations cJ and c′J of Cl(V ) on
FJ (U ), but they coincide on Cl0 (V ): in this case, γ is odd. Declaring cJ (γ) to be, say, +1
on F(U ), we get for w1 , . . . , w2k+1 ∈ WJ :

w1 ∧ · · · ∧ w2k = cJ (w1 . . . w2k ) Ω,

w1 ∧ · · · ∧ w2k+1 = cJ (w1 . . . w2k+1 γ) Ω,

so that in this case, Λ• WJ is an irreducible representation if dim V is odd.

Conclusion: The two algebra representations cJ and c′J of Cl(V ) restrict to the same
group representation cJ of Spinc (V ) which is already irreducible, if dim V is odd.

17
Chapter 2

Spinor modules over compact

Riemannian manifolds

2.1 Remarks on Riemannian geometry

Let M be a compact C ∞ manifold without boundary, of dimension n. Compactness is not
crucial for some of our arguments (although it may be for others), but is very convenient, since
it means that the algebras C(M ) and C ∞ (M ) are unital : the unit is the constant function 1.
For convenience we use the function algebra A = C(M ) —a commutative C ∗ -algebra— at
the beginning. We will change to A = C ∞ (M ) later, when the differential structure becomes
important.
Any A-module (or more precisely, a “symmetric A-bimodule”) which is finitely generated
and projective is of the form E = Γ(M, E) for E → M a (complex) vector bundle. Two
important cases are

X(M ) = Γ(M, TC M ) = (continuous) vector fields on M ;

A1 (M ) = Γ(M, TC∗ M ) = (continuous) 1-forms on M.

These are dual to each other: A1 (M ) ≃ HomA (X(M ), A), where HomA means “A-module
maps” commuting with the action of A (by multiplication).

Definition 2.1. A Riemannian metric on M is a symmetric bilinear form

g : X(M ) × X(M ) → C(M )

such that:

1. g(X, Y ) is a real function if X, Y are real vector fields;

2. g is C(M )-bilinear: g(f X, Y ) = g(X, f Y ) = f g(X, Y ), if f ∈ C(M );

3. g(X, X) ≥ for X real, with g(X, X) = 0 =⇒ X = 0 in X(M ).

The second condition entails that g is given by a continuous family of symmetric bilinear
maps gx : TxC M × TxC M → C or gx : Tx M × Tx M → R; the latter version is positive definite.

Fact 2.2. Riemannian metrics always exist (in abundance).

18
 Since each gx is positive definite, there are “musical isomorphisms” between X(M ) and
A1 (M ),as A-modules
(
X → X✲♭ X ♭ (Y ) := g(X, Y )
X(M ) ✛ A1 (M ), given by
α♯ ← α α(Y ) =: g(α♯ , Y ).

They are mutually inverse, of course. In fact, they can be used to transfer the metric form
X(M ) to A1 (M ):
g(α, β) := g(α♯ , β ♯ ), for α, β ∈ A1 (M ).
One should perhaps write g −1 (α, β) —as is done in [GVF]— since in local coordinates gij :=
g(∂/∂xi , ∂/∂xj ) and g rs := g(dxr , dxs ) have inverse matrices: [g rs ] = [gij ]−1 .
If f ∈ C 1 (M ), the gradient of f is grad f := (df )♯ , so that

g(grad f, Y ) = df (Y ) := Y f.

2.2 Clifford algebra bundles

More generally a real vector bundle E → M is a Euclidean bundle if, with E = Γ(M, E C ),
there is a symmetric A-bilinear form g : E × E → A = C(M ) such that

1. g(s, t) ∈ C(M ; R) when s, t lie in Γ(M, E) —the real sections;

2. g(s, s) ≥ 0 for s ∈ Γ(M, E), with g(s, s) = 0 =⇒ s = 0.

By defining (s | t) := g(s∗ , t), we get a hermitian pairing with values in A:

• (s | t) is A-linear in t;

• (t | s) = (s | t) ∈ A;

• (s | s) ≥ 0, with (s | s) = 0 =⇒ s = 0 in E;

• (s | ta) = (s | t) a for all s, t ∈ E and a ∈ A.

These properties make E a (right) C ∗ -module over A, with C ∗ -norm given by

p
kskE := k(s | s)kA for s ∈ E.

For each x ∈ M , we can form Cl(Ex ) := Cl(Ex , gx ) ⊗R C. Using the linear isomorphisms
σx : Cl(Ex ) → (Λ• Ex )C , we see that these are fibres of a vector bundle Cl(E) → M , isomor-
phic to (Λ• E)C → M as C-vector bundles (but not as algebras!). Under (κλ)(x) := κ(x)λ(x),
the sections of Cl(E) also form an algebra Γ(M, Cl(E)). It has an A-valued pairing

(κ | λ) : x 7→ τ (κ(x)∗ λ(x)).

By defining kκk := supx∈M kκ(x)kCl(Ex ) , this becomes a C ∗ -algebra.

Lemma 2.3. If g, h are two different “metrics” on E = A1 (M ), the corresponding C ∗ -alge-

bras
Bg := Γ(M, Cl(T ∗ M, g) ⊗R C) and Bh := Γ(M, Cl(T ∗ M, h) ⊗R C)
are isomorphic.

19
Proof. We compose α 7→ α♯g : A1 (M ) → X(M ) and X 7→ X ♭h : X(M ) → A1 (M ) to get an
A-linear isomorphism ρ : A1 (M ) → A1 (M ). Now

h(ᾱ, ρ(α)) = ᾱ(α♯g ) = g(ᾱ, α) ≥ 0, for all α ∈ A1 (M ). (2.1)

At each x ∈ M , the C-vector space TxC (M ) may be regarded as a Hilbert space with scalar
product hαx | βx ih := hx (ᾱx , βx ), and now (2.1) says that each ρx ∈ EndC (TxC M ) is a pos-
itive operator with a positive square root σx : we thereby obtain an A-linear isomorphism
σ : A1 (M ) → A1 (M ) such that ρ = σ 2 . We may regard σ as an injective A-linear map from
A1 (M ) into the algebra Bh ; when α ∈ A1 (M ) is real, we get

σ(α)2 = h(σ(α), σ(α)) 1 = h(α, ρ(α)) 1 = g(α, α) 1.

By a now familiar argument, applied to each σx separately, we may extend σ to an A-

linear unital *-algebra homomorphism σ̃ : Bg → Bh . Exchanging g and h gives an inverse
homomorphism σ̃ −1 : Bh → Bg . Since any unital ∗-homomorphism between C ∗ -algebras is
automatically norm-decreasing and thus continuous, we conclude that σ̃ : Bg → Bh is an
isomorphism of C ∗ -algebras.

Definition 2.4. A Clifford module over (M, g) is a finitely generated projective A-module,
with A = C(M ), of the form E = Γ(M, E) for E a (complexified) Euclidean bundle, together
with an A-linear homomorphism c : B → Γ(M, End E), where B := Γ(M, Cl(T ∗ M )) is the
Clifford algebra bundle generated by A1 (M ), such that

(s | c(κ)t) = (c(κ∗ )s | t) for all s, t ∈ E, κ ∈ B.

Example 2.5. Take E = A• (M ) = Γ(M, (Λ• T ∗ M )C ) —all differential forms on M — with

c(α) : ω 7→ ε(α)ω + ι(α♯ )ω for ω ∈ A• (M ) and α ∈ A1 (M ) real. Then E is indeed a Clifford
module, but it is a rather large one: it may have nontrivial submodules. The goal of the next
subsection is top explore how some minimal submodules may be constructed.

2.3 The existence of Spinc structures

Suppose n = 2m + 1 = dim M is odd. Then the fibres of B are semisimple but not simple:
Cl(Tx∗ M ) ≃ M2m (C) ⊕ M2m (C). We shall restrict to the even subalgebras, Cl0 (Tx∗ M ) ≃
M2m (C), by demanding that c(γ) act as the identity in all cases. Then we may adopt the
convention that
c(κ) := c(κγ) when κ is odd.
Notice here that κγ is even; and c(γ) = c(γ 2 ) = +1 is required for consistency of this rule.
We take A = C(M ), but for B we now take
(
Γ(M, Cl(T ∗ M )), if dim M is even,
B := (2.2)
Γ(M, Cl0 (T ∗ M )), if dim M is odd.

The fibres of these bundles are central simple algebras of finite dimension 22m in all cases.
We classify the algebras B as follows. Taking
(
{ Bx = Cl(Tx∗ M ) : x ∈ M }, if dim M is even
B :=
{ Bx = Cl0 (Tx∗ M ) : x ∈ M }, if dim M is odd

20
to be the collection of fibres, we can say that B is a “continuous field of simple matrix
algebras”, which moreover is locally trivial. There is an invariant

δ(B) ∈ H3 (M ; Z)

for such fields, found by Karrer [Kar] and in more generality —allowing the compact operators
K as an infinite-dimensional simple matrix algebra— by Dixmier and Douady [Dix].
Here is a (rather pedestrian) sketch of how δ(B) is constructed:
If x ∈ M , take px ∈ Bx to be a projector of rank one, that is,

px = p∗x = p2x and tr px = 1.

On the left ideal Sx := Bx px , we introduce a hermitian scalar product

hax px | bx px i := tr(px a∗x bx px ). (2.3)

Notice that the recipe

|ax px ihbx px | : cx px 7→ (ax px )(bx px )∗ (cx px ) = (ax px b∗x )(cx px )

identifies L(Sx ) —or K(Sx ) in the infinite-dimensional case— with Bx , since the two-sided
ideal span{ ax px b∗x : ax , bx ∈ Bx } equals Bx by simplicity.
By local triviality, this can be done locally with varying x. If {Ui } is a “good” open
cover1 of M , we get local fields S i = { Si,x : x ∈ Ui } with isomorphisms θi : L(S i ) → B U
i
of fields of simple C ∗ -algebras. On nonempty intersections Uij := Ui ∩ Uj , we get ∗-algebra
isomorphisms θ−1 i θ j : L(S j ) → L(S i ), so there are fields of unitary maps uij : S j → S i such
that θi θj = uij (·)u−1
−1
ij .
On Uijk := Ui ∩ Uj ∩ Uk , we see that (Ad uij )(Ad ujk ) = Ad uik , and so

uij ujk = λijk uik ,

where λijk : Uijk → T are scalar maps. We may now check that λjkl λ−1 ikl λijl = λijk on Uijkl .
Thus λ is a Čech 2-cocycle, and its Čech cohomology class lies in Ȟ2 (M ; T) ≃ H3 (M ; Z). We
may go one more step in order to exhibit this isomorphism: if we write λijk = exp(2πif ijk )
—we can take logarithms since Uijk is simply connected— then

aijkl := f ijk − f ijl + f ikl − f jkl

takes values in Z (and since each Uijkl is connected, these will be constant functions); thus,
these aijkl form a Z-valued 3-cocycle, a. Finally, one may check that its class [a] ∈ H3 (M ; Z)
is independent of all choices made so far. We define δ(B) := [a], which is called the Dixmier–
Douady class of B.
m
Suppose now that the Hilbert spaces Sx ≃ C2 can be chosen globally for x ∈ M —not
just locally for x ∈ Ui — that is, they are fibres of a vector bundle S → M (that may b gifted
with a Hermitian metric) such that L(Sx ) ≃ Bx , for x ∈ M , via a single field of isomorphisms
θ : L(S) → B such that θi = θ U for each Ui . Then uij = θ−1 i θ j = id over Uij , and so λijk = 1
i
over Uijk , and aijkl = 0 over each Uijkl ; hence δ(B) = [a] = 0 in H3 (M ; Z).
1
The word good has a precise technical meaning: namely, that all nonempty finite intersections of open
sets of the cover are both connected and simply connected. On Riemannian manifolds, good open covers may
always be formed using geodesically convex balls.

21
 Conversely if δ(B) = 0, so that [λ] is trivial in Ȟ2 (M ; T), i.e., λ is a 2-coboundary, then
there are maps ν ij : Uij → T such that λijk = ν ij ν −1 −1
ik ν jk on Uijk . Setting v ij := ν ij uij , we get
local fields of unitaries such that v ij v jk = v ik on each Uijk . These v ij : S j → S i are therefore
transition functions for a (Hermitian) vector bundle S → M such that S U ≃ S i for each
i
Ui . Let S := Γ(M, S) denote the A-module of sections of this bundle. Now the pointwise
isomorphisms Bx ≃ End Sx , for each x ∈ M , imply that B ≃ EndA S as A-modules, and
indeed as C ∗ -algebras. We summarize all this in the following Proposition.
Proposition 2.6. Let (M, g) be a compact Riemannian manifold. With A = C(M ) and B the
algebra of Clifford sections given by (2.2), the Dixmier–Douady class vanishes, i.e., δ(B) = 0,
if and only if there is a finitely generated projective A-module S, carrying a selfadjoint action
of B by A-linear operators, such that EndA (S) ≃ B.

2.4 Morita equivalence for (commutative) unital algebras

Definition 2.7. If A, B are unital C-algebras, we say that they are Morita equivalent if
there is a B-A-bimodule E and an A-B-bimodule F, such that E ⊗A F ≃ B as B-B-bimodules
and F ⊗B E ≃ A as A-A-bimodules. We say that such an E is an “equivalence bimodule”.
In general, we may choose F ≃ E ♯ := HomA (E, A) to be the “dual” right A-module with
a specified action of B. We can then identify EndA E ≃ B.
Fact 2.8. EndA E ≃ B whenever E is an equivalence B-A-bimodule.
Fact 2.9. Since A, B are unital, each E is finitely generated and projective (and full).
Remark 2.10. There is a C*-version, due to Rieffel, whereby all bimodules are provided with
compatible A-valued and B-valued Hermitian pairings. This becomes nontrivial in the more
general context of nonunital algebras. The full story is told in [RW]. We shall not need this
machinery in the unital case: remember that M is taken to be compact.
Notation. For isomorphism classes [E] of bimodules, we form the set
Mrt(B, A) := { [E] : E is a B-A-equivalence bimodule }.
In the case B = A, we write Pic(A) := Mrt(A, A); this is called the “Picard group” of A.
We call an A-bimodule E symmetric if the left and right actions are the same: a ⊲x = x ⊳a
for x ∈ E and a ∈ A. When A is commutative, a symmetric A-A-bimodule can be called, more
simply, an “A-module” —as we have already been doing. Even when A is commutative, an
A-A-equivalence bimodule E need not be symmetric. Indeed, suppose φ, ψ ∈ Aut(A). Then
we define φ Eψ to be the same vector space E, but with the bi-action of A on E twisted as
follows:
a1 ⊲ a0 ⊳ a2 := φ(a1 ) a0 ψ(a2 )
For φ = ψ = id, this is the original A-A-bimodule (when either φ = id or ψ = id, we shall
not write that subscript). In particular, we can apply this twisting to E = A itself.
Lemma 2.11. If A is a unital algebra, there exists an A-A-bimodule isomorphism θ : A → φ A
if and only if φ is inner.
Proof. If θ : A → φ A is an A-bimodule isomorphism, then
φ(a)θ(1) = θ(a) = θ(1)a
so that φ(a) = uau−1 , where u = θ(1) is invertible.

22
 Thus the “outer automorphism group” Out(A) := Aut(A)/ Inn(A) classifies the asym-
metric A-bimodules. When A is commutative, so that Inn(A) is trivial, this is just Aut(A).
Recall that

Aut(C(M )) ≃ Homeo(M ), Aut(C ∞ (M )) ≃ Diff(M ),

where φ(f ) : x 7→ f (φ−1 x) for f ∈ C(M ). We shall write PicA (A), following [BW], to denote
the isomorphism classes of symmetric A-bimodules. (This repairs an oversight in [GVF,
Chap. 9], which did not distinguish between Pic(A) and PicA (A), as was pointed out to me
by Henrique Bursztyn.)

Fact 2.12. Pic(A) ≃ PicA (A) ⋊ Aut(A) as a semidirect product of groups, with product given
by ([E], φ) · ([F], ψ) = ([ψ Eψ ⊗A F], φ ◦ φ).

The proof is not difficult, but we refer to the paper [BW].

Lemma 2.13. For A = C(M ) or C ∞ (M ), PicA (A) ≃ H2 (M ; Z).

Proof. Since invertible A-modules L are given by L = Γ(M, L) —either continuous or smooth
sections, respectively— where L → M are C-line bundles; and these are classified by the first
1
Chern class c1 (L) ∈ H2 (M ; Z), obtained from [λ] = [λij ] ∈ Ȟ (M ; T) ≃ H2 (M ; Z). Indeed,
here L♯ = Γ(M, L∗ ), where L∗ → M is the dual bundle and L ⊗A L♯ = Γ(M, L ⊗ L∗ ) ≃
Γ(M, M × C) ≃ C(M ) or C ∞ (M ), respectively.
The group operation in PicA (A) is [L1 ]·[L2 ] = [L1 ⊗A L2 ]: since L1 ⊗A L2 ≃ Γ(M, L1 ⊗L2 ),
it is again a module of sections for a C-line bundle.

2.5 Classification of spinor modules

In this section, A = C(M ) and as before, B = Γ(M, Cl(T ∗ M )) or B = Γ(M, Cl0 (T ∗ M )),
according as the dimension of M is even or odd.
Consider now the set Mrt(B, A) of isomorphism classes of B-A-bimodules: we have seen
that δ(B) = 0 if and only if Mrt(B, A) is nonempty. We shall assume from now on that
indeed δ(B) = 0, so that there exists at least one B-A-bimodule S = Γ(M, S) —continuous
sections, for the moment— such that at each x ∈ M , Sx is an irreducible representation of
the simple algebra Bx . Therefore, any such S has a partner S ♯ = HomA (S, A) such that
S ⊗A S ♯ ≃ B and S ♯ ⊗B S ≃ A: in other words, S is an equivalence B-A-bimodule, and its
isomorphism class [S] is an element of Mrt(B, A).
Since S ♯ ≃ Γ(M, S ∗ ) where S ∗ → M is the dual vector bundle to S → M , we can write
this equivalence fibrewise: Sx ⊗C Sx∗ = EndC (Sx ) ≃ Bx and then Sx∗ ⊗Bx Sx ≃ C, for x ∈ M .

Lemma 2.14. Mrt(B, A) is a principal homogeneous space for the group PicA (A), when
δ(B) = 0.

Proof. There is a right action of PicA (A) on Mrt(B, A), given by [S] · [L] := [S ⊗A L]. We
say that the spinor module S is “twisted” by the invertible A-module L.
If S ⊗A S ♯ ≃ B and S ♯ ⊗B S ≃ A, then for S 1 := S ⊗A L we get

S1 ⊗A S1♯ = S ⊗A L ⊗A L♯ ⊗A S ♯ ≃ S ⊗A A ⊗A S ♯ ≃ S ⊗A S ♯ ≃ B,
S1♯ ⊗B S1 = L♯ ⊗A S ♯ ⊗B S ⊗A L ≃ L♯ ⊗A A ⊗A L ≃ L♯ ⊗A L ≃ A.

23
Thus S1 is again an equivalence B-A-bimodule. Moreover, under the natural isomorphism

HomB (S ′ , S) ⊗A L ≃ HomB (S ′ , S ⊗A L) : F ⊗ l 7→ [s′ 7→ F (s′ ) ⊗ l],

we see that HomB (S, S ⊗A L) ≃ S ♯ ⊗B S ⊗A L ≃ A ⊗A L = L, so that the right action of

PicA (A) is free. On the other hand, the identification

S ⊗ HomB (S, S ′ ) ≃ HomB (EndA (S), S ′ ) : s ⊗ F 7→ [G 7→ (F ◦ G)(s)]

yields, for L := HomB (S, S ′ ), the isomorphism

S ⊗A L ≃ HomB (B, S ′ ) ≃ S ′ ,

so that the action of PicA (A) is transitive.

To proceed, we explain how B acts on S ♯ = HomA (S, A). The spinor module S carries an
A-valued hermitian pairing (2.3) given by the local scalar products defined in the construction
of S, that may be written

(ψ | φ) : x 7→ hψx | φx i, for x ∈ M. (2.4)

We can identify elements of S ♯ with “bra-vectors” hψ| using this pairing, namely, we define
hψ| to be the map φ 7→ (ψ|φ) ∈ A. Since A is unital, there is a “Riesz theorem” for A-modules
showing that all elements of S ♯ are of this form. Now the left B-action is defined by

b hψ| := hψ| ◦ χ(b! ).

Recall that b 7→ χ(b!) is a linear antiautomorphism of B.

Remark 2.15. In these notes, there are many inner products. As a convention, angle brackets
h· | ·i take values in C —we shall call them scalar products to emphasize this— while round
brackets (· | ·) take values in an algebra A —we use the word pairing to signal that.

Lemma 2.16. Let LS := HomB (S ♯ , S) be the A-module for which S ♯ ⊗A LS ≃ S. Then

LS⊗A L ≃ LS ⊗A L ⊗A L, so that the twisting [S] 7→ [S ⊗A L] on Mrt(B, A) induces a
translation by [L ⊗A L] = 2[L] on PicA (A).

Proof. First observe that, since L is an invertible A-module, the dual of S ⊗A L is isomorphic
to S ♯ ⊗A L♯ . The commutativity of the group PicA (A) the shows that

(S ⊗A L)♯ ⊗A (LS ⊗A L ⊗A L) ≃ S ♯ ⊗A L♯ ⊗A (L ⊗A LS ⊗A L)
≃ S ♯ ⊗A LS ⊗A L ≃ S ⊗A L,

and the freeness of the action now implies the result.

Thus, the “mod 2 reduction” j∗ [LS ] ∈ H2 (M ; Z2 ), coming from the short exact sequence
×2 j
of abelian groups 0 → Z −−→ Z − → Z2 → 0, is independent of [S]. Indeed, it defines an
invariant κ[B] ∈ H2 (M ; Z). This is clear, when one takes into account the corresponding
long exact sequence in Čech cohomology and the governing assumption that δ(B) = 0:

∂ (×2)∗ j∗ ∂
· · · → H1 (M ; Z2 ) −
→ H2 (M ; Z) −−−→ H2 (M ; Z) −→ H2 (M ; Z2 ) −
→ H3 (M ; Z) → · · · (2.5)

24
Remark 2.17. It can be shown that κ(B) = w2 (T M ) = w2 (T ∗ M ), the familiar second Stiefel–
Whitney class of the tangent (or cotangent) bundle. See, for instance, the original papers of
Karrer [Kar] and Plymen [Ply], and the lecture notes by Schröder [Schd].
¿What is the meaning of the condition κ(B) = 0? It means that, by replacing any original
choice of S by a suitably twisted S ⊗A L, we can arrange that LS is trivial, i.e. LS ≃ A, or
better yet, that
S ♯ ≃ S as B-A-bimodules.
We now reformulate this condition in terms of a certain antilinear operator C; later on, in
the context of spectral triples, we shall rename it to J.

Proposition 2.18. There is a B-A-bimodule isomorphism S ♯ ≃ S if and only if there is an

antilinear endomorphism C of S such that

(a) C(ψ a) = C(ψ) ā for ψ ∈ S, a ∈ A;

(b) C(b ψ) = χ(b̄) C(ψ) for ψ ∈ S, b ∈ B;

(c) C is antiunitary in the sense that (Cφ | Cψ) = (ψ | φ) ∈ A, for φ, ψ ∈ S;

(d) C 2 = ±1 on S whenever M is connected.

Proof. Ad (a): We provisionally define C by C(ψ) := T hψ|, where T : S ♯ → S is the given

B-A-bimodule isomorphism. Now since hψa| = hψ| ā because (ψ a | φ) = ā (ψ | φ) = (ψ | φ) ā =
(ψ | φ ā) —since A is commutative— we get C(ψ a) = T (hψ| ā) = T hψ| ā = C(ψ) ā.
Ad (b): The formula for the B-action on S ♯ , and the relation (bψ | φ) = (ψ | b∗ φ) give

C(b ψ) = T hb ψ| = T (hψ| ◦ b∗ ) = T (χ(b! ) hψ|) = χ(b̄) T hψ|.

Ad (c): The pairing (φ | Cψ) is antilinear (and bounded) in both φ and ψ, and thus of
the form (ψ | χ) for some χ ∈ S, by the aforementioned “Riesz theorem”. Thus we get an
adjoint map to C, namely the antilinear map C † : φ 7→ χ —obeying the rule for transposing
antilinear operators, i.e., (ψ | C † φ) = (φ | Cψ). Next, notice that C † C is an A-linear bijective
endomorphism of S, that commutes with each b ∈ B:

(φ | C † Cbψ) = (Cbψ | Cφ) = (χ(b̄)Cψ | Cφ)

= (Cψ | χ(b! )Cφ) = (Cψ | Cb∗ φ)
= (b∗ φ | C † Cψ) = (φ | bC † Cψ).

Therefore C † C ∈ EndB (S) ≃ A, i.e. there is an invertible a ∈ A such that C † C = a1S .

1
If a 6= 1, we can now redefine C 7→ a− 2 C, keeping (a) and (b), so with the redefinition
1
C(ψ) := a− 2 T hψ|, we get C † C = 1, i.e., (Cφ | Cψ) = (ψ | C † Cφ) = (ψ | φ).
Ad (d): Finally, C 2 is A-linear, and C 2 b = C χ(b̄) C = b C 2 for b ∈ B, so C 2 = u1S with
u ∈ A. From the relations

uC = C 3 = Cu = ūC by antilinearity of C,
† 2 2 −2 2
ūu 1S = (C ) C = C C = 1S ,

we get u = ū and hence u2 = 1. Thus u ∈ A = C(M ) takes the values ±1 only, so u = ±1

when M is connected. (More generally, C 2 lies in H0 (M, Z2 ).)

25
 The antilinear operator C : S → S, which becomes an antiunitary operator on a suitable
Hilbert-space completion of S, is called the charge conjugation. It exists if and only if
κ(B) = 0.
¿What, then, are spinc and spin structures on M ? We choose on M a metric (without
losing generality), and also an orientation ε, which organizes the action of B, in that a change
ε 7→ −ε induces c(γ) 7→ −c(γ), which either

(i) reverses the Z2 -grading of S = S + ⊕ S − , in the even case; or

(ii) changes the action on S of each c(α) to −c(α), for α ∈ A1 (M ), in the odd case —recall
that c(α) := c(αγ) in the odd case.

Definition 2.19. Let (M, ε) be a compact boundaryless orientable manifold, together with a
chosen orientation ε. Let A = C(M ) and let B be specified as before (in terms of a fixed but
arbitrary Riemannian metric on M ). If δ(B) = 0 in H3 (M ; Z), a spinc structure on (M, ε)
is an isomorphism class [S] of equivalence B-A-bimodules.
If δ(B) = 0 and if κ(B) = 0 in H2 (M ; Z2 ), a pair (S, C) give data for a spin structure,
when S is an equivalence B-A-bimodule such that S ♯ ≃ S, and C is a charge conjugation
operator on S. A spin structure on (M, ε) is an isomorphism class of such pairs.

Remark 2.20. There is an alternative treatment, given in many books, that defines spinc or
spin structures using principal G-bundles for G = Spinc (Rn ) or G = Spin(Rn ) respectively.
The equivalence of the two approaches is treated in [Ply] and [Schd].
Atiyah, Bott and Shapiro [ABS] called a spinc structure a “K-orientation”, for reasons
which may be obvious to K-theorists. At any rate, it is a finer invariant than the orientation
class [ε], provided it exists.
In the long cohomology exact sequence there is a boundary homomorphism

∂
H2 (M ; Z2 ) −
→ H3 (M ; Z).

By examining the definitions of the various Čech cocyles that we have obtained so far, one
can show that δ(B) = ∂(κ(B)).
Remark 2.21. It is known that δ(B) = 0 for dim M ≤ 4: manifolds of dimensions 1, 2, 3, 4
always carry spinc structures. There are 5-dimensional manifolds for which δ(B) 6= 0; the
best-known is the homogeneous space SU(3)/ SO(3). A homotopy-theoretic proof of the
obstruction for this example in given in [Fri].
A complex manifold has a natural orientation and a natural spinc structure coming from
its complex structure. Thus CP m come with a spinc structure, for all m. However, it is known
that CP m admits spin structures if and only if m is odd: therefore, CP 2 is a 4-dimensional
manifold without spin structures.

2.6 The spin connection

We now leave the topological level and introduce differential structure. Thus we replace
A = C(M ) by A = C ∞ (M ), and continuous sections Γcont by smooth sections Γsmooth . Thus
S = Γ(M, S) will henceforth denote the A-module of smooth spinors.
Our treatment of Morita equivalence of unital algebras passes without change to the
smooth level. We can go back with the functor − ⊗C ∞ (M ) C(M ), if desired.

26
Definition 2.22. A connection on a (finitely generated projective) A-module E = Γ(M, E)
is a C-linear map ∇ : E → A1 (M )⊗A E = Γ(M, T ∗ M ⊗E) ≡ A1 (M, E), satisfying the Leibniz
rule
∇(f s) = df ⊗ s + f ∇s.
It extends to an odd derivation of degree +1 on A• (M )⊗A E = Γ(M, Λ• T ∗ M ⊗E) ≡ A• (M, E)
with grading inherited from that of A• (M ), leaving E trivially graded, so that ∇(ω ∧ σ) =
dω ∧ σ + (−1)|ω| ω ∧ ∇σ for ω ∈ A• (M ), σ ∈ A• (M, E).
Employing the usual contraction of vector fields with forms in A• (M ), namely,

ιX ω(Y1 , . . . , Yk ) := ω(X, Y1 , . . . , Yk ) for ω ∈ Ak+1 (M ),

extended to A• (M )⊗A E as ιX ⊗idE —but still written ιX — we get operators ∇X on A• (M, E)

of degree 0 by defining
∇X := ιX ◦ ∇ + ∇ ◦ ιX .
This is A-linear in X. Moreover, if ω ∈ A• (M ) and s ∈ E, one can check that ∇X (ω ⊗ s) =
LX ω ⊗ s + ω ∇X s, where LX = ιX d + d ιX is the Lie derivative of forms with respect to X.
Exercise 2.23. Verify that ∇X (ιY σ) = ιY (∇X σ) + ι[X,Y ] σ for σ ∈ A• (M, E).
Exercise 2.24. If E = X(M ) = Γ(M, T M ), then show that

∇X Y − ∇Y X − [X, Y ] = ιY ιX ∇θ,

where θ ∈ A1 (M, T M ) is the fundamental 1-form defined by ιX θ := X. We say that ∇ is

torsionfree if ∇θ = 0 in A2 (M, T M ).
Exercise 2.25. Show that ιY ιX ∇2 = ∇X ∇Y −∇Y ∇X −∇[X,Y ] , where the degree +2 operator
∇2 on A• (M, E) is the curvature of ∇.
We mention two natural constructions for connections, on tensor products of A-modules
and on dual A-modules. Firstly, if ∇′ : F → A1 (M ) ⊗A F is another connection in another
A-module, then
e ⊗ t) := ∇s ⊗ t + s ⊗ ∇′ t
∇(s
e a connection on E ⊗A F.
(extneded by linearity, as usual) makes ∇
Next, if E = HomA (E, A), then the dual connection ∇♯ on E ♯ is determined by
♯

d(ζ(s)) =: (∇♯ ζ)(s) + ζ(∇s) in A1 (M ); or equivalently

X(ζ(s)) =: (∇♯X ζ)(s) + ζ(∇X s) in A, for X ∈ X(M ),

whenever ζ ∈ E ♯ and s ∈ E.
Definition 2.26. If E an A-module equipped with an A-valued Hermitian pairing, we say
that a connection ∇ on E is Hermitian if

(∇s | t) + (s | ∇t) = d (s | t), or, in other words,

(∇X s | t) + (s | ∇X t) = X (s | t), for any real X ∈ X(M ).

If ∇, ∇′ are connections on E, then ∇′ −∇ is an A-module map: (∇′ −∇)(f s) = f (∇′ −∇)s,

so that locally, over U ⊂ M for which E U → U is trivial, we can write

∇ = d + α, where α ∈ A1 (U, End E).

27
Fact 2.27. On X(M ) = Γ(M, T M ) there is, for each Riemannian metric g, a unique torsion-
free connection that is compatible with g:

g(∇X, Y ) + g(X, ∇Y ) = d(g(X, Y )) for X, Y ∈ X(M ), or

g(∇Z X, Y ) + g(X, ∇Z Y ) = Z(g(X, Y )) for X, Y, Z ∈ X(M ).

The explicit formula for this connection is

2g(∇X Y, Z) = X(g(Y, Z)) + Y (g(X, Z)) − Z(g(X, Y ))

+ g(Y, [Z, X]) + g(Z, [X, Y ]) − g(X, [Y, Z]). (2.6)

It is called Levi-Civita connection associated to g. (The proof of existence consists in

showing that the right hand side of this expression is A-linear in Y and Z, and obeys a
Leibniz rule with respect to X, so it gives a connection; and uniqueness is obtained by
checking that metric compatibility and torsion freedom make the right hand side automatic.)
The dual connection on A1 (M ) will also be called the “Levi-Civita connection”. At
the risk of some confusion, we shall use the same symbol ∇ for both of these Levi-Civita
connections.

Local formulas From now on, we assume that U ⊂ M is an open chart domain over which
the tangent and cotangent bundles are trivial. Local coordinates are functions x1 , . . . , xn ∈
C ∞ (U ), and we denote ∂j := ∂/∂xj ∈ X(M ) U for the local basis of vector fields; by definition,
their Lie brackets vanish: [∂i , ∂j ] = 0. We define the Christoffel symbols Γkij ∈ C ∞ (U ) by

∇∂i ∂j =: Γkij ∂k , or ∇∂j =: Γkij dxi ⊗ ∂k .

The explicit expression (2.6) for the Levi-Civita connection reduces to a local formula over U ,
namely
Γkij := 21 g kl (∂i gjl + ∂j gil − ∂l gij ); here [g rs ] = [gij ]−1 . (2.7)
Notice that Γkji = Γkij ; this is beacuse of torsion freedom.
Dually, the coefficients of the Levi-Civita connection on 1-forms are −Γkij (note the change
of sign):
∇∂i (dxk ) = −Γkij dxj , or ∇(dxk ) = −Γkij dxi ⊗ dxj .
Since the Riemannian metric gives a concept of (fibrewise) orthogonality on the tangent
and cotangent bundles, we can select local orthonormal bases:

{E1 , . . . , En } for X(M ) U

= Γ(U, T M ) : g(Eα , Eβ ) = δαβ ;
{θ1 , . . . , θn } for A1 (M ) U
= Γ(U, T ∗ M ) : g(θα , θβ ) = δ αβ .

We rewrite the Christoffel symbols in these local bases:

e β dxi ⊗ Eβ ,
∇Eα =: Γ e β dxi ⊗ θα .
∇θβ = −Γ
iα iα

e • are skewsymmetric matrices:

Metric compatibility means that, for each fixed i, the Γ i•

eβ + Γ
Γ e α = −g(∇∂ θβ , θα ) − g(θβ , ∇∂ θα ) = −∂i (δ αβ ) = 0.
iα iβ i i

e lies in A1 (U, so(T ∗ M )) ≃ A1 (U ) ⊗R so(Rn ).

Thus Γ

28
Definition 2.28. On a spinor module S = Γ(M, S), a spinc -connection is any Hermitian
connection ∇S : S → A1 (M ) ⊗A S which is compatible with the action of B in the following
way:

∇S (c(α)ψ) = c(∇α)ψ + c(α)∇S ψ for α ∈ A1 (M ), ψ ∈ S; or

∇SX (c(α)ψ) = c(∇X α)ψ + c(α)∇SX ψ 1
for α ∈ A (M ), ψ ∈ S, X ∈ X(M ), (2.8)

where ∇α and ∇X α refer to the Levi-Civita connection on A1 (M ).

If (S, C) are data for a spin structure, we say ∇S is a spin connection if, moreover,
each ∇X : S → S commutes with C whenever X is real.
e
Before discussing existence, let us look first at local formulas. We thus write “∇ = d − Γ
locally” for the Levi-Civita connection, with an implicit choice of local orthonormal bases of
1-forms. We recall that there are isomorphisms of Lie algebras

µ̇ : so(Tx∗ M ) → Q(Λ2 Tx∗ M ) ≡ spin(Tx∗ M )

with the property that ad(µ̇(A)) = A for A ∈ so(Tx∗ M ); in other words, [µ̇(A), v] = Av for
v ∈ Tx∗ M —this is a commutator for the Clifford product in Cl(Tx∗ M, gx ). On the chart
e fibrewise; this means that
domain U , we can apply µ̇ to Γ
e c(α)] = c(Γ
[µ̇(Γ), e α)

for α ∈ A1 (M ) with support in U , Γe ∈ Γ(U, End T ∗ M ) is mapped to µ̇(Γ)

e ∈ Γ(U, End S),
and c(α) again denotes the Clifford action action of α on S U = Γ(U, S).
In this way we get the local expression of a connection,
e
∇S := d − µ̇(Γ), acting on S . (2.9)
U

Suppose we take α ∈ A1 (M ) with support in U , and ψ ∈ S U

. Then

e
∇S (c(α)ψ) = d(c(α)ψ) − µ̇(Γ)c(α)ψ
e
= c(dα) ψ + c(α) dψ − µ̇(Γ)c(α)ψ
e c(α)]ψ)
= c(α)(dψ − µ̇(Γ̃)ψ) + (c(dα) − [µ̇(Γ),
= c(α)∇S ψ + c(dα − Γ e α)ψ
= c(∇α)ψ + c(α)∇S ψ. (2.10)

Thus ∇S := d − µ̇(Γ) provides a local solution to the existence of ∇S : S → A1 (M ) ⊗A S

satisfying the Leibniz rule:

∇S (c(α)ψ) = c(∇α)ψ + c(α)∇S ψ.

Physicists like to write γ α := c(θα ) for a given local orthonormal basis of A1 (M ) —so
that the γ α are fixed matrices. For convenience, we also write γβ = δαβ γ α also (in the
Euclidean signature, which we are always using here); in other words, γβ = γ β but with its
index lowered for use with the Einstein summation convention. Thus the Clifford relations
are just
γ α γ β + γ β γ α = 2δ αβ , for α, β = 1, . . . , n.
The formula (1.7) for µ̇ can now be rewritten as
e = −1Γ
µ̇(Γ) eβ α
4 •α γ γβ .

29
 A more sensible notation arrives by introducing matrix-valued functions ω1 , . . . , ωn ∈
Γ(U, End T ∗ M ) as follows:
e β γ α γβ .
ωi := − 41 Γ iα

Let us look at the calculation (2.10) again, after contracting with a vectorfield X. We get

[∇SX , c(α)] ψ = [LX − X i µ̇(ωi ), c(α)] ψ

= c(LX α) − X i c(ωi α) ψ = c(∇X α) ψ.

e β γ α γβ , for X ∈ X(M ).
Thus the local coefficients of ∇SX are − 14 X i Γ iα

Now suppose S comes from a spin structure on M . Since C(ψa) = (Cψ)ā for a ∈ A =
C ∞ (M ), the operator C acts locally (as a field of antilinear conjugations Cx : Sx → Sx ); and
since C(b) = χ(b̄)C for b ∈ B, we get, for α, β = 1, . . . , n:

Cγ α γ β = Cc(θα )c(θβ ) = c(θα )c(θβ )C = γ α γ β C.

Thus ∇SX C − C∇SX vanishes over U , provided X U is real.

Suppose that ∇′ is another local connection defined on S U
and satisfying the Leibniz
rule (2.10) there. Then
∇′ − ∇S = β ∈ A1 (U, End S)
and c(κ)βψ = βc(κ)ψ for all κ ∈ B U . Thus βx is a scalar matrix in End(Sx ), for each x ∈ U .
To fix β, we ask that both ∇′ and ∇S be Hermitian connections; this entails that each βx is
skew-hermitian:
hβx φx | ψx i + hφx | βx ψx i = 0 for x ∈ U,
so this scalar is purely imaginary. On the other hand, if ∇′X , like ∇SX , commutes with C
whenever the coefficients X i of X are real functions, then this scalar must be purely real.
Therefore, β = 0.

Proposition 2.29. If (S, C) are data for a spin structure on M , then there is a unique
Hermitian spin connection ∇S : S → A1 (M ) ⊗A S, such that

∇S (c(α)ψ) = c(∇α)ψ + c(α)∇S ψ, for α ∈ A1 (M ), ψ ∈ S,

and such that ∇SX C = C∇SX for X ∈ X(M ) real.

Proof. We have shown that ∇S exists locally with the recipe (2.9) on any chart domain. This
e is skewadjoint —because the represen-
recipe gives a local Hermitian connection since µ̇(Γ)
tation c is selfadjoint— and it commutes with C. Any other local connection with these
properties must coincide with (2.9) over U .
Furthermore, on overlaps U1 ∩ U2 of chart domains, we have shown that β := ∇S |U1 −
∇S |U2 ∈ A1 (U1 ∩ U2 , End S) vanishes. Therefore, the local expressions can be assembled into
a globally defined spin connection.

Remark 2.30. If S is only a spinor module for a spinc structure, then the uniqueness argument
for the local spin connection fails. We can only conclude that ∇S |U1 − ∇S |U2 = i(α1 − α2 ) ⊗
1End S , where α1 ∈ A1 (U1 ) and α2 ∈ A1 (U2 ) are real 1-forms. We may be able to patch
these “gauge potentials” to get a connection ∇ of a line bundle L♯ = Γ(M, L∗ ). Then one
can show that on S ⊗ L, there is a connection ∇S,α that satisfies the Leibniz rule above, and
hermiticity. These are “spinc connections” for the twisted spinc structures.

30
 If ∇ is any connection on an A-module E = Γ(M, E), then



∇2 (f s) = ∇(df ⊗ s + f ∇s) = d(df ) ⊗ s − df ∇s + df ∇s + f ∇2 s = f ∇2 s,

for f ∈ A, so that ∇2 is tensorial: ∇s = R s for a certain 2-form R ∈ A2 (M, End E), the
curvature of ∇. For the Levi-Civita connection, a local calculation gives
e
∇2 α = (d − Γ)(dα e α) = −d(Γ
−Γ e α) − Γ
e dα + Γ(
e Γe α) = (−dΓ
e+Γ
e ∧ Γ)
e α,

which yields the local expression for the Riemannian curvature tensor:

R e+Γ
= −dΓ e ∈ A2 (U, so(T ∗ M )).
e∧Γ
U

Likewise, the curvature RS of the spin connection is locally given by

e + µ̇(Γ)
µ̇(R) = −dµ̇(Γ) e ∧ µ̇(Γ)
e ∈ A2 (U, End S).

One can check these formulas to get more familiar expressions by computing R(X, Y ) =
ιY ιX R and likewise RS (X, Y ), for X, Y ∈ X(M ).

2.7 Epilogue: counting the spin structures

A spin structure on (M, ε) is an equivalence class of pairs (S, C), but ¿what can be said about
the equivalence relation?
First, S has a class [S] ∈ Mrt(B, A): these are classified by H2 (M ; Z). If (S1 , C1 ) is another
spin structure, then C1 : S1♯ → S1 comes from a B-A-bimodule isomorphism T1 : S1 → S1 .
But now S1 ≃ S ⊗ L for some L, where [L] ∈ H2 (M ; Z) is well defined. Thus we get

− ⊗A L✲♯
S♯ S ♯ ⊗A L♯ = (S ⊗A L)♯
T T1
❄ −⊗ L ❄
A ✲
S S ⊗A L

and therefore (S1 , C1 ) ∼ (S, C) if this diagram commutes. Now

S1 ≃ S1♯ ⊗A HomB (S1♯ , S1 ) ≃ S1♯ ⊗A HomB (S ♯ ⊗A L♯ , S ⊗A L)

≃ S1♯ ⊗A L ⊗A HomB (S ♯ , S) ⊗A L ≃ S1♯ ⊗A L ⊗A L,

since HomB (S ♯ , S) is trivial: the existence of T shows that [S ♯ ] = [S] in Mrt(B, A). The
conclusion is that S1 ≃ S1♯ ⊗A L ⊗A L. Thus S1 is also selfdual if and only if L ⊗A L is
trivial: (×2)∗ [L] = 0 in H2 (M ; Z). But, using the long exact sequence (2.5), we find that
ker(×2)∗ = im{∂ : H1 (M ; Z2 ) → H2 (M ; Z)}.
Conclusion: Those [S ⊗A L] ∈ Mrt(B, A) for which L ⊗A L is trivial, but L is not, i.e.,
the distinct spin structures on (M, ε), are classified by H1 (M, Z2 ).
Remark 2.31. The group H1 (M, Z2 ) is known to classify real line bundles over M . If a twist by
L exchanges the spinor modules for two spin structures, there is an antilinear automorphism
of L which matches the two charge conjugation operators, and the part of L fixed by this
automorphism comprises the sections of the corresponding R-line bundle over M .

31
Chapter 3

Dirac operators

Suppose we are given a compact oriented (boundaryless) Riemannian manifold (M, ε) and
a spinor module with charge conjugation (S, C), together with a Riemannian metric g, so
that the Clifford action c : B → EndA (S) has been specified. We can also write it as ĉ ∈
HomA (B ⊗A S, S) by setting ĉ(κ ⊗ ψ) := c(κ) ψ.
Definition 3.1. Using the inclusion A1 (M ) ֒→ B —where in the odd dimensional case this
is given by c(α) := c(αγ), as before— we can form the composition
/ := −i ĉ ◦ ∇S
D (3.1)

where
∇S ĉ
S −−→ A1 (M ) ⊗A S −
→ S,
/ : S → S is C-linear. This is the Dirac operator associated to (S, C) and g.
so that D
The (−i) is included in the definition to make D / symmetric (instead of skewsymmetric)
as an operator on a Hilbert space, because we have chosen g to be positive definite, that is,
γ α γ β + γ β γ α = +2 δ αβ . Historically, D
/ was introduced as −iγ µ δµ = γ µ pµ where the pµ are
components of a 4-momentum, but in the Minkowskian signature.
Using local (coordinate or orthonormal) bases for X(M ) and A1 (M ), we get nicer formulas:
/ = −i ĉ(∇S ψ) = −i c(dxj )∇S∂j ψ = −i γ α ∇SEα ψ.
Dψ (3.2)

/ is the commutation relation:

The essential algebraic property of D
/ a] = −i c(da),
[D, for all a ∈ A = C ∞ (M ). (3.3)

Indeed,
/ a] ψ = −i ĉ(∇S (aψ)) + ia ĉ(∇S ψ)
[D,
= −i ĉ(∇S (aψ) − a ∇S ψ)
= −i ĉ(da ⊗ ψ) = −i c(da) ψ, for ψ ∈ S.

3.1 The metric distance property

As an operator, we can make sense of [D, / a] by conferring on S the structure of a Hilbert
space: if we write det g := det[gij ] for short, then
p
νg := det g dx1 ∧ dx2 ∧ · · · ∧ dxn ∈ An (M )

32
is the Riemannian volume form (for the given orientation ε and metric g). In the notation,
we assume that all local charts are consistent with the given orientation, which just means
that det[gij ] > 0 in any local chart. The scalar product on S is then given by
Z
hφ | ψi := (φ | ψ) νg for φ, ψ ∈ S.
M
p
On completion in the norm kψk := hψ | ψi, we get the Hilbert space H := L2 (M, S) of
L2 -spinors on M .
Using the gradient grad a := (da)♯ ∈ X(M ), we can compute

/ a]k2 = kc(da)k2 = sup kcx (da(x))k2

k[D,
x∈M
= sup gx (dā(x), da(x)) with gx on (Tx∗ M )C
x∈M


= sup gx grad ā x , grad a x using the dual gx on (Tx M )C
x∈M
2
=: sup grad a x
=: k grad ak2∞ .
x∈M

Classically, we compute distances on a (connected) Riemannian manifold by the formula

d(x, y) := inf{ length(γ) : γ : [0, 1] → M ; γ(0) = x, γ(1) = y },

with the infimum taken over all piecewise-smooth paths γ in M from x to y. For a ∈ C ∞ (M ),
we then get
1 Z
d
a(y) − a(x) = a(γ(1)) − a(γ(0)) = [a(γ(t))] dt
0 dt
Z 1 Z 1 Z 1
= γ̇(a) γ(t) dt = da(γ̇) γ(t) dt = daγ(t) (γ̇(t)) dt
0 0 0
Z 1


= gγ(t) grad a γ(t) , γ̇(t) dt,
0

and we can estimate this difference by

Z 1
|a(y) − a(x)| ≤ grad a γ(t)
|γ̇(t)| dt
0
Z 1
≤ k grad ak∞ |γ̇(t)| dt = k grad ak∞ length(γ)
0
/ a]k length(γ).
= k[D,

Thus
/ a]k ≤ 1 } ≤ inf length(γ) =: d(x, y).
sup{ |a(y) − a(x)| : a ∈ C(M ), k[D, (3.4)
γ

In this supremum, we can use a ∈ C(M ) not necessarily smooth; a need only be continuous
with grad a (ν-essentially) bounded. Since we have obtained |a(y)−a(x)| ≤ k grad ak∞ d(x, y),
we see that a need only be Lipschitz on M —with respect to the distance d— with Lipschitz
constant ≤ k grad ak∞ . In fact, this is the best general Lipschitz constant: fix x ∈ M , and
set ax (y) := d(x, y). This function lies in C(M ), and |ax (y) − ax (z)| ≤ d(y, z) by the triangle

33
inequality for d. Since k grad ax k∞ = 1 by a local geodesic calculation, we see that a = ax
makes the inequality in (3.4) sharp:
d(x, y) = sup{ |a(y) − a(x)| : k grad ak∞ ≤ 1 }
/ a]k ≤ 1 },
= sup{ |a(y) − a(x)| : a ∈ C(M ), k[D, (3.5)
so that D / determines the Riemannian distance d, which in turn determines the metric g.
(The Myers–Steenrod theorem of differential geometry says that g is uniquely determined by
its distaqnce function d.)
Example 3.2. Take M = S1 (n = 1, m = 0, 2m = 1). The trivial line bundle is a spinor
bundle, with S = C ∞ (S1 ) = A, and C is just the complex conjugation K of functions. With
the flat metric on S1 ≃ R/Z, we can identify S with the set of smooth 1-periodic functions
on R, so both ∇ and ∇S are trivial since Γ111 = 0. Therefore,
d
/ = −i
D
dθ
is the Dirac operator in this case. Thus [D,/ f ] = −if ′ for f ∈ A, and for α, β ∈ [0, 1], we get
Z β Z β
|f (β) − f (α)| = f ′ (θ) dθ ≤ |f ′ (θ)| dθ ≤ |β − α| whenever kf ′ k∞ ≤ 1.
α α

Using fα (β) := |β − α| for α − ≤ β ≤ α + 12 wrapped around R/Z, we get a Lipschitz

1
2
function making the inequality sharp. Thus d(α, β) = |β − α| provided |β − α| ≤ 21 : this is
just the arc length on the circle of circumference 1.
More generally, the formula for d(x, y) yields the length of the minimal geodesic from x
to y, provided y is closer to x, than the “cut-locus” of x.

3.2 Symmetry of the Dirac operator

We now regard D / as an operator on L2 (M, S), defined initially on the dense domain S =
Γsmooth (M, S).
/ is symmetric: that is, whenever φ, ψ ∈ S, the following equality holds:
Proposition 3.3. D
/ | ψi = hφ | Dψi.
hDφ /

Proof. We compute the pairings (Dφ/ | ψ) and (φ | Dψ),

/ which take values in A = C ∞ (M ).
We need a formula for the divergence of a vector field: LX νg =: (div X) νg , so that
Z Z Z Z
(div X) νg = LX (νg ) = ιX (dνg ) + d(ιX νg ) = d(ιX νg ) = 0
M M M M

by Stokes’ theorem (remember that M has no boundary). This formula is

div X = ∂j X j + Γjjk X k = dxj (∇∂j X),
as can easily be checked; on the right hand side we use the Levi-Civita connection on X(M ).
Now we abbreviate cj := c(dxj ) ∈ Γ(U, End S), for j = 1, . . . , n. Then we compute the
difference of A-valued pairings:
/ | ψ) = (φ | cj ∇S∂j ψ) + (cj ∇S∂j φ | ψ)
/ − i(Dφ
i(φ | Dψ)
= (φ | ∇S∂j cj ψ) − (φ | c(∇∂j dxj )ψ) + (∇S∂j φ | cj ψ)
= ∂j (φ | cj ψ) − (φ | c(∇∂j dxj )ψ).

34
Here we have used the Leibniz rule for ∇S , the selfadjointness of cj since dxj is a real local
1-form, and the hermiticity of ∇S .
By duality, the map α 7→ (φ | c(α)ψ), which takes 1-forms to functions, defines a vector
field Zφψ —because X(M ) = EndC ∞ (M ) (A1 (M ), C ∞ (M ))— so the right hand side becomes

∂j (dxj (Zφψ )) − (∇∂j dxj )(Zφψ ) = dxj (∇∂j Zφψ ) = div Zφψ ,

where we have used the Leibniz rule for the dual Levi-Civita connections on A1 (M ) and on
X(M ), respectively. Thus

/ − (Dφ
(φ | Dψ) / | ψ) = −i div Zφψ

which has integral zero.

3.3 Selfadjointness of the Dirac operator

If T is a densely defined operator on a Hilbert space H, its adjoint T ∗ has domain

Dom T ∗ := { φ ∈ H : ∃ χ ∈ H with hT ψ | φi = hψ | χi for all ψ ∈ Dom T }

and then T ∗ φ := χ, of course, so that the formula hT ψ|φi = hψ|T ∗ φi holds. If T is symmetric,
then clearly Dom T ⊆ Dom T ∗ with T ∗ = T on Dom T : that is, T ∗ is an extension of T to a
larger domain.
The second adjoint T ∗∗ =: T is called the closure of T (symmetric operators always have
this closure), where the domain of the closure is

Dom T := { ψ ∈ H : ∃ φ ∈ H and a sequence {ψn } ⊂ Dom T,

such that ψn → ψ and T ψn → φ in H }

In other words, the graph of T in H ⊕ H is the closure of the graph of T . And then, of course,
we put T ψ := φ. When T is symmetric, we get

Dom T ⊆ Dom T ⊆ Dom T ∗ .

Definition 3.4. We say that T is selfadjoint if T = T ∗ ; thus T is symmetric and closed.

Otherwise, we say that T is essentially selfadjoint if it is symmetric and its closure T is
selfadjoint.

Remark 3.5. Selfadjoint operators have real spectra: sp(T ) ⊆ R. This is crucial: an un-
bounded operator that is merely symmetric may have non-realR elements in its spectrum.
Moreover, selfadjoint operators obey the spectral theorem: T = R λ dET (λ), where ET is a
“projector-valued measure” on Borel subsets of R with support in sp(T ).
The main result of this chapter is that the Dirac operator on a compact Riemannian spin
manifold is essentially selfadjoint. This was proved by Wolf in 1973; he actually showed the
result also for noncompact manifolds which are complete with respect to the Riemannian
distance given by the metric [Wolf]. In his proof, completeness is needed to establish that
closed geodesic balls are compact; that proof is also given in the book by Friedrich [Fri]. For
simplicity, we deal here only with the compact case.

Theorem 3.6. Let (M, g) be a compact boundaryless Riemannian spin manifold. The Dirac
/ is essentially selfadjoint on its original domain S.
operator D

35
 ∗
/ , given by
Proof. There is a natural norm on Dom D

/ ∗ ψk2 .
|||ψ|||2 := kψk2 + kD
∗
We claim that S = Γsmooth (M, S) is dense in Dom D / for this norm. Using a finite partition of
unity f1 + · · · + fr = 1 with each fi ∈ A supported in a chart domain Ui over which S|Ui → Ui
is trivial, it is enough to show that any fi φ, with φ ∈ Dom D / ∗ , can be approximated in the
|||·|||-norm by elements of Γsmooth (Ui , S). Thus we can suppose that supp φ ⊂ Ui , and regard
φ ∈ L2 (Ui , S) as a 2m -tuple of functions φ = {φk } with each φk ∈ L2 (Ui , νg ).
Previous formulas now show that
Z Z
∗
/ φ | ψi = hφ | Dψi
hD / = (φ | c ∇∂j ψ) = (cj φ | ∇S∂j ψ)
j S

Z M
 
= ∂j (cj φ | ψ) − (∇S∂j cj φ | ψ) νg
Z
 
= −(cj φ | ψ) (div ∂j ) − (∇S∂j cj φ | ψ) νg
M

∗ ∗
after an integration by parts, so that D / is given by the formula D / = −(∇S∂ + div ∂j ) c(dxj ),
j
as a vector-valued distribution on Ui ; in particular, it is also a differential operator (the
difference D /∗ − D / will soon be seen to vanish).
Now, if {hr } is a smooth delta-sequence, then for large enough r we can convolve both φ
and D / ∗ φ with hr , while remaining supported in Ui —the convolution is defined after pulling
back functions on the chart domain Ui to an fixed open subset of Rn . Thus we find that
/ ∗ (φ ∗ hr ) → D
/ ∗ φ in L2 (Ui , νg )2 , so that |||φ ∗ hr − φ||| → 0. But the spinors
m
φ ∗ hr → φ and D
φ ∗ hr are smooth since the hr are smooth, so we conclude that S is |||·|||-dense in Dom D / ∗.
∗
But now D / (φ ∗ hr ) = D(φ / ∗ hr ) since S = Dom D, / so we have shown that φ lies in Dom D /
∗ ∗ ∗∗ ∗
and that Dφ / =D / φ. Thus Dom D / = Dom D / , and it follows that D / =D / =D / : which
establishes that D / is selfadjoint.

3.4 The Schrödinger–Lichnerowicz formula

If E → M is any smooth vector bundle with connection ∇E on E = Γ(M, E), we can consider
′
not only ∇E : E → A1 (M ) ⊗A E, but also the connection ∇E := ∇ ⊗ 1 + 1 ⊗ ∇E on the
tensor product bundle E ′ = A1 (M ) ⊗A E; here ∇ is once again the Levi-Civita connection
′
on A1 (M ). Their composition is an operator ∇E ◦ ∇E from E to A1 (M ) ⊗A A1 (M ) ⊗A E;
using the metric g on A1 (M ) we can take the trace over the first two factors, ending up with
a Laplacian:
′
∆E := − Trg (∇E ◦ ∇E ) : E → E. (3.6)
The minus sign is a convention to yield a positive operator (instead of a negative one) [BGV].
Locally, this means:
∆E = −g ij (∇E E k E
∂i ∇∂j − Γij ∇∂k ).

Definition 3.7. In particular, when E = M × C is the trivial line bundle, we get the “scalar
Laplacian”
∆ = −g ij (∂i ∂j − Γkij ∂k ), (3.7)
also known as the “Laplace–Beltrami operator” on A = C ∞ (M ). Likewise, when E = S, we
get the spinor Laplacian for a spin manifold.

36
 Before examining the relation between the Dirac operator and the spinor Laplacian, we
collect a few well-known formulas for the Riemann curvature tensor, R. These can be found
in many places, for instance [BGV]; perhaps the best reference is Milnor’s little book [Mil].
The square of the Levi-Civita connection on X(M ) is C ∞ (M )-linear, so it is given by
∇ X = R(X), where R ∈ A2 (M, End T M ). In local coordinates, its components are Rijkl :=
2

g(∂i , R(∂k , ∂l ) ∂j ).
Taking a trace over the first and third indices, we get the Ricci tensor, whose components
are Rjl := g ik Rijkl . The trace of the Ricci tensor is the scalar curvature (or “curvature
scalar”) s := g jl Rjl = g jl g ik Rijkl ∈ C ∞ (M ). Under exchange of indices, R has the following
skewsymmetry and symmetry relations:

Rijkl = −Rjikl = −Rijlk ; Rijkl = Rklij .

The (first) Bianchi identity says that the cyclic sum over three indices vanishes:

Rijkl + Riljk + Riklj = 0.

Moreover, the Ricci tensor is symmetric: Rjl = Rlj .

The formula in the next Proposition is generally attributed to Lichnerowicz [Lich, 1963],
but was anticipated by Schrödinger in a little-known paper [Sch1, 1932].

Proposition 3.8. Let (M, g) be a compact Riemannian spin manifold with spinor module S.
Then
D/ 2 = ∆S + 14 s (3.8)
as an operator on S,

Proof. It is enough to prove the equality when applied to spinors ψ supported in a chart
/ = −i cj ∇S∂ , we get
domain, so we may use local coordinate formulas. Since D j

/ 2 = −ci ∇S∂i cj ∇S∂j = −ci cj ∇S∂i ∇S∂j − ci c(∇∂i dxk ) ∇S∂k

D
= −ci cj (∇S∂i ∇S∂j − Γkij ∇S∂k ),

and from Γkij = Γkji (torsion freedom) and the Clifford relation ci cj + cj ci = 2g ij , we get

/ 2 = −g ij (∇S∂i ∇S∂j − Γkij ∇S∂k ) − 21 ci cj [∇S∂i , ∇S∂j ]

D
= ∆S − 12 ci cj [∇S∂i , ∇S∂j ].

Since [∂i , ∂j ] = 0, the commutator [∇S∂i , ∇S∂j ] is a spin-curvature term:

[∇S∂k , ∇S∂l ] = RS (∂k , ∂l ) = 41 Rijkl ci cj ,

because the curvature RS of ∇S is given by µ̇(R). Hence,

/ 2 − ∆S = − 18 Rijkl ck cl ci cj = 81 Rjikl ck cl ci cj .
D (3.9)

Since Rjikl has cyclic sum zero in the indices i, k, l, we can also skewsymetrize ck cl ci =
c(dxk ) c(dxl ) c(dxi ). It is a simple exercise to check that

ck cl ci = Q(dxk ∧ dxl ∧ dxi ) + g li ck − g ki cl + g kl ci ,

37
If we now skewsymmetrize the right hand side of (3.9) in the indices i, k, l —which does not
change its value— the Q-term contributes zero to the result. Also, the term g kl ci cj = g lk ci cj
contributes zero, while g li ck cj and −g ki cl cj contribute equally. Thus,

/ 2 − ∆S = 41 Rijkl g ik cl cj = 41 Rjl cl cj = 18 Rjl (cl cj + cj cl )

D
= 41 Rjl g jl = 41 s.

One consequence of the formula (3.8) is a famous “vanishing theorem” of Lichnerowicz.

Corollary 3.9. If s(x) ≥ 0 for all x ∈ M , and s(x0 ) > 0 at some point x0 ∈ M , then
/ = {0}.
ker D

/ = 0. Then
Proof. Suppose that ψ ∈ S satisfies Dψ
Z
0 = kDψk / 2 ψi = hψ | ∆S ψi +
/ 2 = hψ | D 1
4 s (ψ | ψ) νg . (3.10)
M

Now it is easy to check that, after an integration by parts over M and discarding a divergence
term,
hψ | ∆S ψi = g ij h∇S∂i ψ | ∇S∂j ψi. (3.11)

Since the matrix [g ij ] is positive definite, this (by the way) shows that ∆S is a positive
operator; and since s ≥ 0, both terms on the right hand side of (3.10) are nonnegative; so
they must both vanish, since their sum is zero.
Moreover, (3.11) shows that hψ | ∆S ψi = 0 implies ∇S ψ = 0. This in turn implies that
∂j (ψ | ψ) = (∇S∂j ψ | ψ) + (ψ | ∇S∂j ψ) vanishes for each j, so that k := (ψ | ψ) is a constant
R
function. But now (3.10) reduces to 0 = k M s νg , which entails k = 0 and then ψ = 0.

We saw by example (Appendix A.2) that on S2 , the Dirac operator for the round metric
/ = N \ {0}: here s ≡ 2 and ker D
has spectrum sp(D) / = {0}. Thus there are no “harmonic
2
spinors” on S .

3.5 The spectral growth of the Dirac operator

2
Since D/ = ∆S + 41 s, and ∆S is closely related to the Laplacian ∆ on the (compact, bound-
aryless) Riemannian manifold (M, g), the general features of sp(D)/ may be deduced from
those of sp(∆).
We require two main properties of Laplacians on compact Riemannian manifolds without
boundary. Recall that
∗ M ⊗T ∗ M ∗M
∆ = − Trg (∇T ◦ ∇T ) = −g ij (∂i ∂j − Γkij ∂k )

is the local expression for the Laplacian (which depends on g through the Levi-Civita con-
nection and g ij ). Thus ∆ is a second order differential operator on C ∞ (M ).

Fact 3.10. The Laplacian ∆ extends to a positive selfadjoint operator on L2 (M, νg ) —also
denoted by ∆— and (1 + ∆) has a compact inverse.

To make ∆ selfadjoint, we must complete C ∞ (M ) to a larger domain, by defining

|||f |||2 := hf | (1 + ∆)f i = kf k2 + g ij h∂i f | ∂j f i,

38
 R
where hf | f i := M |f |2 νg . Taking Dom ∆ := { f ∈ L2 (M, νg ) : |||f ||| < ∞ }, ∆ becomes
selfadjoint and (1 + ∆)−1 : L2 (M, νg ) → (Dom ∆, |||·|||) is bounded. Then one shows that
the inclusion (Dom ∆, |||·|||) ֒→ L2 (M, νg ) is a compact operator (by Rellich’s theorem); and
(1 + ∆)−1 , as a bounded operator on L2 (M, νg ), is then the composition of these two, so it
is also compact.

Corollary 3.11. ∆ has discrete (point) spectrum of finite multiplicity.

Proof. Since (1 + ∆)−1 is compact, its spectrum —except for 0— consists only of eigenvalues
of finite multiplicity. Therefore, the same is true of 1 + ∆, and of ∆ itself. Indeed,
 
−1 1 1 1
sp((1 + ∆) ) = , , ,...
1 + λ 0 1 + λ1 1 + λ2

with λ0 ≤ λ1 ≤ λ2 ≤ · · · being the list of eigenvalues of ∆ in increasing order. These are

counted with multiplicity: an eigenvalue of multiplicity r appears exactly r times on the list.
This λk → ∞, since (1 + λk )−1 → 0, as k → ∞.

As a convention, when A is a compact positive operator, we write λk (A) to denote the

k-th eigenvalue of A in decreasing order (with multiplicity): λ0 (A) ≥ λ1 (A) ≥ · · · ; on the
other hand, if A is an unbounded positive selfadjoint operator with compact inverse, we write
the eigenvalues in increasing order, as we did for ∆.

Fact 3.12 (Weyl’s theorem). The counting function for sp(∆) is

N∆ (λ) := #{ λk (∆) : λk (∆) ≤ λ }.

For large λ, the following asymptotic estimate holds:

N∆ (λ) ∼ Cn Vol(M ) λn/2 as λ → ∞,

R
where n = dim M , and Vol(M ) = M νg is the total volume of the manifold M . The constant
Cn , that depends only on the dimension n, is

Ωn 1
Cn = = ,
n(2π)n (4π)n/2 Γ( n2 + 1)

where Ωn = Vol(Sn−1 ) = 2 π n/2 /Γ( n2 ).

We shall not prove Weyl’s theorem, in particular why the number of eigenvalues (up to λ)
is proportional to Vol(M ), but we shall compute the constant by considering an example.
For a simple and clear exposition of the proof, we recommend Higson’s ICTP lectures [Hig].
Example 3.13. Take M = Tn = Rn /bZ n to be the n-torus with unit volume. Identify C ∞ (Tn )
with the smooth periodic functions on the unit cube [0, 1]n . For the flat metric on Tn , and
local coordinates t = (t1 , . . . , tn ), we get
     
∂ 2 ∂ 2 ∂ 2
∆=− − − ··· − .
∂t1 ∂t2 ∂tn

Thus we find eigenfunctions, labelled by r = (r1 , . . . , rn ) ∈ Zn :

φr := e2πir·t , for t ∈ [0, 1]n .

39
Since { φr : r ∈ Zn } is an orthonormal basis for L2 (Tn ), these are a complete set of eigen-
functions, and therefore
sp(∆) = { 4π 2 |r|2 : r ∈ Zn }.
If B(0; R) is the ball of radius R, centered at 0 ∈ Rn , then

N∆ (λ) = #{ r ∈ Zn : 4π 2 |r|2 ≤ λ }
 
p
2
λ n/2
∼ Vol(B(0; λ/4π )) = Vol(B(0; 1))
4π 2
λn/2 Ωn Ωn
= n
= λn/2 , as λ → ∞,
(2π) n n(2π)n
using
Z Z 1 Z  Z 1
1 n n−1 Ωn
Vol(B(0; 1)) = dx ∧ · · · ∧ dx = ν r dr = Ωn rn−1 dr = .
B(0;1) 0 Sn−1 0 n

For the spinor Laplacian ∆S , a similar estimate holds, but with Cn replaced by 2m Cn
m
(recall that in the flat torus case with untwisted spin structure, S ≃ C ∞ (Tn ) ⊗ C2 ). Now
by Lichnerowicz’ formula, D / 2 differs from ∆S by a bounded multiplication operator 41 s, thus
ND/ 2 (λ) ∼ N∆S (λ) as λ → ∞, hence

2m Ω n
ND/ 2 (λ) ∼ Vol(M ) λn/2 , as λ → ∞.
n(2π)n

Consider the positive operator |D| / 2 )1/2 ; remember that µ is an eigenvalue for |D|
/ := (D / if
2 2
/
and only if µ is an eigenvalue for D (with the same multiplicity). We arrive at the following
estimate.

Corollary 3.14.
2m Ω n
N|D|
/ (λ) ∼ Vol(M ) λn , as λ → ∞.
n(2π)n
Example 3.15. For M = S2 , with n = 2, we have seen (in Appendix A.2) that

/ = { ±(l + 12 ) : l +
sp(D) 1
2 ∈ N + 12 }, with multiplicities 2l + 1
= { ±k : k = 1, 2, 3, . . . }, with multiplicities 2k.

Therefore
X
N|D|
/ (λ) = 4k = 2⌊λ⌋(⌊λ⌋ + 1) ∼ 2λ(λ + 1) ∼ 2λ2 , as λ → ∞.
1≤k≤λ

Ω2 2π 1 1
Now C2 = = 2 = and 2C2 = for spinors. Therefore 2C2 Area(S2 ) λ2 = 2λ2 ,
2(2π)2 8π 4π 2π
so
1
Area(S2 ) = = 4π.
C2
In other words, Weyl’s theorem allows us to deduce the area of the 2-sphere S2 from (the
knowledge of the circumference of the circle Ω2 = 2π and) the growth of the spectrum of the
Dirac operator on S2 .

40
Chapter 4

Spectral Growth and Dixmier

Traces

4.1 Definition of spectral triples

We start with the definition of the main concept in noncommutative geometry.

Definition 4.1. A (unital) spectral triple (A, H, D) consists of:

• an algebra A with an involution a 7→ a∗ , equipped with a faithful representation on:

• a Hilbert space H; and also

• a selfadjoint operator D on H, with dense domain Dom D ⊂ H, such that a(Dom D) ⊆

Dom D for all a ∈ A,

satisfying the following two conditions:

• the operator [D, a], defined initially on Dom D, extends to a bounded operator on H,
for each a ∈ A;

• D has compact resolvent: (D − λ)−1 is compact, when λ ∈

/ sp(D).

For now, and until further notice, all spectral triples will be defined over unital algebras.
The compact-resolvent condition must be modified if A is nonunital: as well as enlarging A
to a unital algebra, we require only that the products a(D − λ)−1 , for a ∈ A and λ ∈ / sp(D),
be compact operators.
Example 4.2. Let (M, ε) be an oriented compact boundaryless manifold which is spin, i.e.
admits spin structures, and (S, C) be data for a specific spin structure. Choose a Rie-
mannian metric g on M (which allows us to define ∇ and ∇S ) and let D / = −i ĉ ◦ ∇S
be the corresponding Dirac operator, extended to be a selfadjoint operator on L2 (M, S).
Then (C ∞ (M ), L2 (M, S), D)/ is a spectral triple. Here [D,
/ a] = −i c(da) is a bounded op-
erator on spinors, with k[D, / a]k = k grad ak∞ , for a ∈ C ∞ (M ). We know by now that
2
/ + 1)−1 = (D
(D / − i)−1 (D
/ + i)−1 is compact, so (D/ ± i)−1 is compact. We refer to these
spectral triples as “standard commutative examples”.
/ sp(D), then (D − λ)−1 − (D − µ)−1 = (λ − µ)(D − λ)−1 (D − µ)−1
Note that, if λ, µ ∈
—this is the famous “resolvent equation”— since


(D − λ) (D − λ)−1 − (D − µ)−1 (D − µ) = (D − µ) − (D − λ) = λ − µ.

41
Thus (D − λ)−1 is compact if and only if (D − µ)−1 is compact, so we need only to check
this condition for one value of λ. In the same way, we get the following useful result.
Lemma 4.3. D has compact resolvent if and only if (D2 + 1)−1 is compact.
Proof. We may take λ = −i, since the selfadjointness of D implies that ±i ∈
/ sp(D). Thus,
D has compact resolvent if and only if (D + i) is compact. Let T = (D + i)−1 ; then the
−1

proof reduces to the well-known result that a bounded operator T is compact if and only if
T ∗ T is compact.

By the spectral√theorem (D2 + 1)1/2 − |D| = f (D), where f : R → R is the continuous

1
function f (λ) := λ2 + 1 − |λ| = √λ2 +1+|λ| ; and 0 < f (λ) ≤ 1 for all λ ∈ R, so that
kf (D)k ≤ 1. Or more precisely: the operator f (D) := (D2 + 1)1/2 − |D|, defined initially on
Dom D, extends to a bounded operator on H, of norm at most 1.
In many arguments to come, we shall employ |D| and |D|−1 as if we knew that ker D = {0}.
However, even if ker D 6= {0}, we can always replace |D| by (D2 + 1)1/2 and |D|−1 by
(D2 + 1)−1/2 , at the cost of some extra calculation.

4.2 Logarithmic divergence of spectra

If A is a positive selfadjoint operator with compact resolvent, let { λk (A) : k ∈ N } be
its eigenvalues listed in increasing order, λ0 (A) ≤ λ1 (A) ≤ λ2 (A) ≤ · · · (an eigenvalue of
multiplicity r occurs exactly r times in the list). The counting function NA (λ), defined for
λ > 0, is the number of eigenvalues not exceeding λ:

NA (λ) := #{ k ∈ N : λk (A) ≤ λ }.

If A is invertible (i.e., if λ0 (A) > 0), we can define the “zeta function”
X
ζA (s) := Tr A−s = λk (A)−s , for s > 0,
k≥0

where we understand that ζA (s) = +∞ when A−s is not traceless. For real s, ζA (s) is a
nonnegative decreasing function.
It is actually more useful to consider finite partial sums.
Notation. If T ∈ K(H) is any compact operator, and if k ∈ N, let sk (T ), called the k-th
singular value of T , be the k-th eigenvalue of the compact positive operator |T | := (T ∗ T )1/2 ,
where these are listed in decreasing order, with multiplicity. Thus s0 (T ) ≥ s1 (T ) ≥ s2 (T ) ≥
· · · and each singular value occurs only finitely many times in the list, namely, the finite
multiplicity of the that eigenvalue of |T |; therefore, sk (T ) → 0 as k → ∞. Note that
s0 (T ) = kT k since s0 (T )2 is the largest eigenvalue of T ∗ T , so that s0 (T )2 = kT ∗ T k = kT k2 .
For each N ∈ N, write
N
X −1
σN (T ) := sk (T ).
k=0
We shall see later that for many spectral triples, the counting function of the positive
(unbounded) operator |D| has polynomial growth: for some n, one can verify an asymptotic
relation N|D| (λ) ∼ Cn′ λn . In that case we can take A := |D|−n , which is compact. Then
the number of eigenvalues of A that are ≥ ε equals N|D| (λ) for λ = 1/ε. This suggests
heuristically that for N close to N|D| (1/ε), the N -th eigenvalue is roughly C/ε for some
constant C, so that σN (|D|−n ) = O(log N ). We now check this condition in a few examples.

42
Example 4.4. We estimate σN (|D| / −s ) for s > 0, where D/ is the Dirac operator on the sphere
2
S with its spin structure and its rotation-invariant metric. We know that the eigenvalues of
/ are k = 1, 2, 3, . . . with respective multiplicities 2(2k) = 4, 8, 12, . . . . For r = 1, 2, 3, . . . ,
|D|
let
Xr
Nr := 4k = 2r(r + 1) ∼ 2r2 as r → ∞,
k=1

so that log Nr ∼ 2 log r as r → ∞. Next,

r
X r
X
/ −s ) =
σNr (|D| 4k(k −s
)=4 k 1−s ,
k=1 k=1

and thus Z r
r
/ −s )
σNr (|D| 4 X 1−s 2
∼ k ∼ t1−s dt as r → ∞,
log Nr 2 log r log r 1
k=1

by the “integral test” of elementary calculus. There are three cases to consider:
Z r
2 r2−s − 1
• If s < 2, then t1−s dt = diverges as r → ∞;
log r 1 2−s
Z r
2
• if s > 2, then t1−s dt → 0 as r → ∞; while
log r 1
/ −s )
σNr (|D| 2 log r
• if s = 2, then ∼ → 2.
log Nr logr
Finally, note that if Nr−1 ≤ N ≤ Nr , then

/ −s )
σNr−1 (|D| / −s )
σN (|D| / −s )
σNr (|D|
≤ ≤ ,
log Nr log N log Nr−1
while log N ∼ log Nr−1 ∼ log Nr ∼ 2 log r as r → ∞. Thus

+∞ if s < 2,
/ −s )
σN (|D| / −s ) 
σNr (|D|
lim = lim = 2 if s = 2,
N →∞ log N r→∞ log Nr 

0 if s > 2.

/ −2 diverges logarithmi-
We express this result by saying that for s = 2, “the spectrum of |D|
cally”. There is precisely one exponent, namely s = 2, for which this limit is neither zero nor
infinite.

Exercise 4.5. Do the same calculation for D / on the torus Tn , whose spectrum we know:
/ −s diverges logarithmically if and only if s = n = dim Tn .
show that the spectrum of |D|

4.3 Some eigenvalue inequalities

Let H be a separable (infinite-dimensional) Hilbert space, and denote by L(H) the algebra
of bounded operators on H. Let K = K(H) be the ideal of compact operators on H. Each
T ∈ K has a polar decomposition T = U |T |, where |T | = (T ∗ T )1/2 ∈ K, and U ∈ L(H) is
a partial isometry. This factorization is unique if we require that U = 0 on ker |T |, since U
must map the range of |T | isometrically onto the range of T .

43
 The spectral theorem yields an orthonormal family {ψk } in H, such that
X X
|T | = sk (T ) |ψk ihψk |, T = sk (T ) |U ψk ihψk |.
k≥0 k≥0

(If |T | is invertible, this is an orthonormal basis for H. Otherwise, we can adjoin an orthonor-
mal basis for ker |T | to the family {ψk }.) Since φk := U ψk gives another orthonormal family,
any T ∈ K has an expansion of the form
X
T = sk (T ) |φk ihψk |, (4.1)
k≥0

for some pair of orthonormal families {φk }, {ψk }.

If V1 , V2 are unitary operators on H, we can then write
X
V1 T V 2 = sk (T ) |V1 φk ihV2∗ ψk |,
k≥0

and conclude that sk (V1 T V2 ) = sk (T ) for each k, and hence that

σN (V1 T V2 ) = σN (T ).

Therefore, any norm |||T ||| that is built from the sequence { sk (T ) : k ∈ N } is unitarily
invariant, that is, |||V1 T V2 ||| = |||T ||| for V1 , V2 unitary.
Example 4.6. If kT k is the usual operator norm on K, then

kT k = kT ∗ T k1/2 = k |T | k = sup sk (T ) = s0 (T ).
k≥0

A compact operator T is called trace-class, and we write T ∈ L1 = L1 (H), if the following

series converges: X
kT k1 := Tr |T | = sk (T ) = lim σN (T ).
N →∞
k≥0

For 1 < p < ∞, there are Schatten classes Lp = Lp (H) consisting of operators for which the
following norm is finite:
X 1/p
p
kT kp = sk (T ) .
k≥0

There are strict inclusions L1 ⊂ Lr ⊂ Lp ⊂ K for 1 < r < p < ∞.

Soon, we shall introduce a “Dixmier trace class” L1+ (H), with yet another norm built
from singular values, such that L1 ⊂ L1+ ⊂ Lp for p > 1.
Much is known about the singular values of compact operators. For instance, the following
relation holds, for T ∈ K:

sk (T ) = inf{ kT (1 − P )k : P = P 2 = P ∗ , dim P (H) ≤ k }. (4.2)

This comes from a well-known minimax principle: see [RS], for instance. The infimum is
indeed attained at the projector Q of rank k whose P range is Q(H) := span{ψ0 , . . . , ψk−1 },
when T is given by (4.1), since T (1 − Q) = T − T Q = j≥k sj (T )|φj ihψj | is an operator with
norm kT − T Qk = sk (T ).

44
Lemma 4.7. If T ∈ K, then

σN (T ) = sup{ kT P k1 : P = P 2 = P ∗ , rank P = N }. (4.3a)

If A is a positive compact operator, then it is also true that

σN (A) = sup{ Tr(P AP ) : P = P 2 = P ∗ , rank P = N }. (4.3b)

Proof. If P is a projector of finitePN −1rank N , then (T P )∗ (T P ) = P T ∗ T P and thus |T P | has

finite rank ≤ N , so kT P k1 = k=0 sk (T P ) = σN (T P ). From the formula (4.2), it follows
that 0 ≤ A ≤ B in K implies sk (A) ≤ sk (B) for each k ∈ N, and in particular, since
0 ≤ P T ∗ T P ≤ T ∗ T , we get sk (T P ) ≤ sk (T ). Thus σN (T P ) ≤ σN (T ) also. We conclude that
the right hand side of (4.3a) is ≤ σN (T ).
If we write T in the form (4.1) and P then choose Q, as before, to be the projector with
k−1
range span{ψ0 , . . . , ψk−1 }, then |T Q| = j=0 sj (T ) |ψj ihψj | and thus kT Qk1 = σN (T ).
When A ∈ K is positive, and P is a projector of rank n, then Tr(P AP ) = Tr(AP ) ≤
kAP
P k1 ≤ σN (A). To see that the supremum in (4.3b) is attained, we can write A =
k≥0 sk (A) |ψk ihψk | and note that QA = AQ = QAQ, so that AQ is also a positive operator.
It then follows that Tr(QAQ) = Tr(AQ) = kAQk1 = σN (A).

Corollary 4.8. Each σN is a norm on K: σN (S + T ) ≤ σN (S) + σN (T ) for S, T ∈ K.

Proof. This follows from kSP + T P k1 ≤ kSP k1 + kT P k1 for P = P 2 = P ∗ , rank P = N .

Lemma 4.9. If T ∈ K, then

σN (T ) = inf{ kRk1 + N kSk : R, S ∈ K with R + S = T }.

Proof. If T = U |T |, then |T | = U ∗ T (by the details of polar decomposition, this is true even
though U might not be unitary), so T = R + S implies U ∗ T = U ∗ R + U ∗ S; thus, we can
suppose that T ≥ 0.
If we now split T =: R + S, then σN (T ) ≤ σN (R) + σN (S) ≤ kRk1 + σN (S), while
X X
σN (S) = sk (S) ≤ s0 (S) = N kSk.
0≤k<N 0≤k<N
P
For T = k≥0 sk (T ) |ψk ihψk |, we consider the special splitting into positive operators,
X
e :=
R (sk (T ) − sN (T )) |ψk ihψk |, Se := T − R.
e
0≤k<N

e 1 = σN (T ) − N sN (T ), while kSk
Then kRk e = sN (T ) by inspection.

The triangle inequality in Corollary 4.8 is not good enough for our needs: our goal is
get an additive functional, rather than just a subadditive one. The next step is to extract
from (4.3b) a sort of “wrong-way triangle inequality”, at least for positive compact operators.

Lemma 4.10. If A ≥ 0, B ≥ 0 are positive compact operators, and if M, N ∈ N, then

σM +N (A + B) ≥ σM (A) + σN (B).

45
Proof. From (4.3b) we obtain σM (A) = sup{ Tr(P AP ) : P = P 2 = P ∗ , rank P = M }
and σN (B) = sup{ Tr(P ′ BP ′ ) : P ′ = P ′2 = P ′∗ , rank P ′ = N }. Now rank(P + P ′ ) =
dim(P H + P ′ H) ≤ M + N , so if P ′′ is any projector of rank M + N whose range includes
the subspace P H + P ′ H, then P ≤ P ′′ and P ′ ≤ P ′′ as operators. Therefore,

Tr(P AP ) + Tr(P ′ BP ′ ) ≤ Tr(P ′′ AP ′′ ) + Tr(P ′′ BP ′′ ) = Tr(P ′′ (A + B)P ′′ ),

so that σM (A)+σN (B) ≤ supP ′′ Tr(P ′′ (A+B)P ′′ ) ≤ σ2N (A+B). (Notice how this argument
requires additivity of the trace: it would not have worked with k · k1 instead of Tr, hence the
restriction to the case of positive operators.)

Corollary 4.11. If A, B ∈ K with A ≥ 0, B ≥ 0, then

σN (A + B) ≤ σN (A) + σN (B) ≤ σ2N (A + B).

We see that the functional A 7→ σN (A)/ log N is not far from being additive functional
on the positive cone K+ . But to get a truly additive functional, we must try to take the limit
N → ∞, and here things become more interesting.

4.4 Dixmier traces

It is a bit awkward to deal with the index N of σN (A) as a discrete variable, but we can fix
this by a simple linear interpolation.
If N ≤ λ ≤ N + 1, so that λ = N + t with 0 ≤ t ≤ 1, we put

σλ (A) := (1 − t)σN (A) + tσN +1 (A).

Note that σλ (A + B) ≤ σλ (A) + σλ (B) now holds for all λ ≥ 0: every σλ is a norm on K.

Exercise 4.12. Check that σλ (A + B) ≤ σλ (A) + σλ (B) ≤ σ2λ (A + B), for A, B ≥ 0 in K,

also holds for all λ ≥ 0.

Definition 4.13. The Dixmier ideal L1+ = L1+ (H) = L1,∞ (H) is defined to be
 
1+ σλ (T )
L := T ∈ K : sup <∞ .
λ≥e log λ

(The e here is by convention: any constant > 1 would do. Also, the notation L1+ is
not universally accepted: some authors prefer the clumsier notation L1,∞ , or even L(1,∞) ,
which comes from the historical origin of these operator ideals in real interpolation theory:
see [Con, IV.C] for that.)
Since each σλ is a norm on K, so also is this supremum whenever it is finite. Thus L1+
has a natural (¡unitarily invariant!) norm

σλ (T )
kT k1+ := sup for T ∈ L1+ .
λ≥e log λ

As stated, the norm depends on the chosen constant e, but the ideal L1+ (H) does not.
Note that T ∈ K is traceclass if and only if σλ (T ) is bounded (by kT k1 , for instance)
without need for the factor (1/ log λ). Thus L1 (H) ⊂ L1+ (H).

46
Remark 4.14. If the bounded function σλ (T )/ log λ is actually convergent as λ → ∞, or
equivalently, if σN (T )/ log N converges as N → ∞, then clearly
σN (T ) σλ (T )
lim = lim ≤ kT k1+ .
N →∞ log N λ→∞ log λ

We get an additive functional defined on L1+ in three more steps. First, we dampen the
oscillations in σλ (T )/ log λ by taking a Cesàro mean with respect to the logarithmic measure
on an interval [λ0 , ∞) for some λ0 > e. For definiteness, we choose λ0 = 3. Our treatment
closely follows the appendix of the local-index paper of Connes and Moscovici [CM].
Definition 4.15. For λ ≥ 3, we set
Z λ
1 σu (T ) du
τλ (T ) := , for T ∈ L1+ (H). (4.4)
log λ 3 log u u
Exercise 4.16. Check the triangle inequality τλ (S + T ) ≤ τλ (S) + τλ (T ) for λ ≥ 3.
Lemma 4.17 (Connes–Moscovici). If A ≥ 0, B ≥ 0 in L1+ (H), then
 
log log λ
τλ (A) + τλ (B) − τλ (A + B) = O as λ → ∞.
log λ
σu (A + B)
Proof. First of all, it is clear that ≤ kAk1+ + kBk1+ for λ ≥ e. Next,
log u
Z λ 
1 σ2u (A + B) σu (A + B) du
τλ (A) + τλ (B) − τλ (A + B) ≤ −
log λ 3 log u log u u
Z 2λ   Z λ Z 2λ 
1 σu (A + B) σu (A + B) du 1 σu (A + B) du
= − − − .
log λ 6 log(u/2) log u u log λ 3 6 log u u
The second term can be rewritten as
Z 6 Z 2λ 
1 σu (A + B) du
− .
log λ 3 6 log u u
R6 R 2λ du 2 log 2
Since 3 duu = λ u = log 2, we get an estimate of log λ kA + Bk1 . For the first term, we
compute
Z 2λ   Z  
1 σu (A + B) log u du kA + Bk1+ λ log 2u du
−1 ≤ −1
log λ 6 log u log(u/2) u log λ 3 log u u
Z λ
kA + Bk1+ du kA + Bk1+
= log 2 < log 2 (log log λ).
log λ 3 u log u log λ
Since the failure of additivity of τλ vanishes as λ → ∞, the second step is to quotient out
by functions vanishing at infinity. For that we consider the “corona” C ∗ -algebra
Cb ([3, ∞))
B∞ := .
C0 ([3, ∞)
The function λ 7→ τλ (A), for A ≥ 0 in L1+ , lies in Cb ([3, ∞)), and its image τ (A) in B∞
defines an additive map, that is,

τ (A + B) = τ (A) + τ (B) for A ≥ 0, B ≥ 0 in L1+ .

The final step is to compose this map with a state on B∞ .

47
Definition 4.18. For A ≥ 0 in L1+ , let τ (A) ∈ (B∞ )+ denote the image, under the quotient
map Cb ([3, ∞)) → B∞ , of the bounded function λ 7→ τλ (A). This yields an additive map
between positive cones, τ : (L1+ )+ → (B∞ )+ . Since the “four positive parts” of any opera-
tor in L1+ also lie in L1+ , as is easily checked, this map extends in the obvious way to a
positive linear map τ : L1+ → B∞ . Moreover, τ is invariant under unitary conjugation, i.e.,
τ (U AU ∗ ) = τ (A) for each unitary U ∈ L(H).
For each state ω : B∞ → C, we can now define a Dixmier trace Trω on L1+ (H) by

Trω T := ω(τ (T )).

Since Trω (U AU ∗ ) = Trω (A) for positive A ∈ L1+ (H) and unitary U ∈ L(H), each such
positive linear functional on L1+ (H) is indeed a trace.

This definition has a drawback : since B∞ is a non-separable C ∗ -algebra, there is no way

to exhibit even one such state. However, Dixmier traces are still computable in a special case:
if limλ→∞ τλ (T ) exists, then τ (T ) coincides with the image of a constant function in B∞ , and
since the state ω is normalized, ω(1) = 1, the value ω(τ (T )) equals this limit:

Trω T = lim τλ (T )
λ→∞

is independent of ω, provided that the limit exists. Such operators are called measurable.
When this happens, we shall suppress the label ω and write Tr+ T for the common value of
all Dixmier traces.
The use of the Cesàro mean (4.4) simplifies the original definition that Dixmier [Dix1]
gave of these traces. A detailed analysis of these (and other related) functionals was made
recently by Lord, Sedaev and Sukochev [LSS], who called them “Connes–Dixmier traces”.
As an unexpected consequence of their work, they have shown that a positive operator A ∈
L1+ (H) is measurable if and only if the original sequence { σN (A)/ log N : N ∈ N } is already
convergent. Thus it is not necessary to compute τλ (A), since

σN (A)
Tr+ A = lim for positive, measurable A ∈ L1+ .
N →∞ log N

48
Chapter 5

Symbols and Traces

5.1 Classical pseudodifferential operators

In order to develop a symbol calculus for Dirac operators and their powers, we shall tem-
porarily restrict our attention to a single chart domain U ⊂ M , over which the cotangent
bundle is trivial: T ∗ M U ≃ U × Rn . If P is an operator on C ∞ (M ), or more generally, on a
space of sections Γ(M, E) of a vectorPbundle E → M , and if {φ1 , . . . , φm } is a finite partition
m
of unity in C ∞ (M ), then P (f ) = i,j=1 φi P (φj f ), so we may as well consider operators
which are defined on a single chart domain of M . At some later stage, we must ensure that
the important properties of such operators are globally defined, independently of the choice
of local coordinates.
We will work, then, in local coordinates (x1 , . . . , xn ) over a chart domain U ; the local
coordinates of the cotangent bundle T ∗ M U are

(x, ξ) = (x1 , . . . , xn , ξ1 , . . . , ξn ), where ξ ∈ Tx∗ M.

Let E → M be a vector bundle of rank r. We assume (without loss of generality) that the
vector bundle E is also trivial over U , so we can identify Γ(U, End E) with U × Mr (C).
A differential operator acting on (smooth) local sections f ∈ Γ(U, E) is an operator P
of the form X
P = aα (x) Dα , with aα ∈ Γ(U, End E),
|α|≤d

where we use the notation Dα := D1α1 . . . Dnαn , and Dj := −i ∂/∂xj , the positive integer d is
the order of P .
The local coordinates allow us to identify U with an open subset of Rn . The coefficients
aα are matrix-valued functions U → Mr (C).
By a Fourier transformation, we can write, for f ∈ Cc∞ (U, Rr ),
Z
P f (x) = (2π)−n
eixξ p(x, ξ) fˆ(ξ) dn ξ
n
ZRZ
−n
= (2π) ei(x−y)ξ p(x, ξ) f (y) dn y dn ξ, (5.1)
R2n

where p(x, ξ) is a polynomial of order d in the ξ-variable, called the (complete) symbol of P .
(Clearly, this symbol depends on the choice of local coordinates.) Here
Z
−n
Kp (x, y) := (2π) ei(x−y)ξ p(x, ξ) f (y) dn ξ (5.2)
Rn

49
is the kernel of P , as an integral operator: the inverse Fourier transform of p(x, ξ).
For the Dirac operator D,/ we can use the local expression of the spin connection to write
D/ = −i c(dxj ) ∇S∂ = −i c(dxj )(∂j + ωj (x)), so the corresponding symbol is
j

p(x, ξ) = c(dxj )(ξj − i ωj (x)). (5.3)

This is a first-order polynomial in the ξj variables, so that D / is a first order differential

operator. The leading term in p(x, ξ) —the part that is homogeneous in ξj of degree one—
is c(dxj )ξj = c(ξj dxj ) = c(ξ), where ξ = ξj dxj can be regarded as an element of A1 (U ).
More generally, a pseudodifferential operator P is given locally by an integral of the form
(5.1), where the symbol p(x, ξ) need no longer be a polynomial. In that case, we must specify
certain classes of symbols for which these integrals make sense.
Definition 5.1. The vector space S d (U ) of (scalar) symbols of order ≤ d, consists of func-
tions p ∈ C ∞ (U × Rn ) such that, for any compact K ⊂ U , and any multiindices α, β ∈ Nn ,
there exists a constant CKαβ such that
1
|Dxβ Dξα p(x, ξ)| ≤ CKαβ (1 + |ξ|2 ) 2 (d−|α|) for all x ∈ K, ξ ∈ R. (5.4)

Here Dxβ and Dξα denote derivatives in the xi variables and in the ξj variables, respectively.
1
We use (1 + |ξ|2 ) 2 instead of |ξ| to avoid problems at ξ = 0.
In the same way, we define matrix-valued symbols of order ≤ d as smooth functions
p : U × Rn → Mr (C) satisfying the same norm estimates, but with the absolute value | · | on
the left hand side of (5.4) replaced by a matrix norm in Mr (C). By a small abuse of notation,
we shall write p ∈ S d (U ) also in the matrix-valued case.
When p(x, ξ) is a polynomial in ξ, of order at most d, we can isolate its homogeneous
parts:
d
X
p(x, ξ) = pd−j (x, ξ), where pd−j (x, tξ) = td−j pd−j (x, ξ) for t > 0.
j=0

Definition 5.2. More generally, an element p ∈ S d (U ) is called a classical symbol if we can

find a sequence of terms pd (x, ξ), pd−1 (x, ξ), pd−2 (x, ξ),. . . , with pd−j (x, tξ) = td−j pd−j (x, ξ)
for t > 0, such that for each k = 0, 1, 2, . . . ,
k−1
X
p− pd−j ∈ S d−k (U ).
j=0

When this is possible, we write

∞
X
p(x, ξ) ∼ pd−j (x, ξ), (5.5)
j=0

and regard this series as an asymptotic development

T of the symbol p. This expansion does not
determine p(x, ξ) uniquely: a symbol in k∈N S d−k (U ) is called “smoothing”, and smoothing
symbols are exactly those symbols whose asymptotic expansion is zero.
Definition 5.3. A classical pseudodifferential operator of order d, over U ⊂ Rn , is an
operator P defined by (5.1), for which p(x, ξ) is a classical symbol in S d (U ) whose leading
term pd (x, ξ) does not vanish. This leading term is called the principal symbol of P , and
we also denote it by σ P (x, ξ) := pd (x, ξ).

50
 We need a formula for the symbol of the composition of two classical pseudodifferential
operators (“classical ΨDOs”, for short). It is not clear a priori when and if two such operators
are composable: we remit to [Tay], for instance, for the full story on compositions (and
adjoints) of classical pseudodifferential operators, and for the justification of the following
formula.
If P is a classical ΨDOs of order d1 with symbol p ∈ S d1 (U ), and if Q is a classical ΨDO
of orders d2 with symbol p ∈ S d2 (U ), then the symbol p ◦ q of the composition P Q lies in
S d1 +d2 (U ) and its asymptotic development is given by
X i|α|
(p ◦ q)(x, ξ) ∼ Dξα p(x, ξ) Dxα q(x, ξ). (5.6)
n
α!
α∈N

To find the terms (p ◦ q)d1 +d2 −j (x, ξ) of the symbol expansion, one must substitute (5.5) for
both p and q into the right hand side of (5.6) and rearrange a finite number of terms. For the
case j = 0, one need only use α = 0 —since Dξα lowers the order by |α|— and in particular,
the principal symbols compose easily:

(p ◦ q)d1 +d2 (x, ξ) = pd1 (x, ξ) qd2 (x, ξ).

The composition formula is valid for both scalar-valued and matrix-valued symbols, provided
the matrix size r is the same for both operators.
Exercise 5.4. If P and Q are classical ΨDOs with scalar-valued symbols, show that the
principal symbol of [P, Q] = P Q − QP is −i {σ P , σ Q }, where {·, ·} is the Poisson bracket of
functions:
n
X ∂σ P ∂σ Q ∂σ Q ∂σ P
−i {σ P (x, ξ), σ Q (x, ξ)} = −i − .
∂ξj ∂xj ∂ξj ∂xj
j=1

Conclude that the order of [P, Q] is ≤ d1 + d2 − 1. ¿What can be said about the order of [P, Q]
if P and Q have matrix-valued symbols of size r > 1?
Suppose U and V are open subsets of Rn and that φ : U → V is a diffeomorphism. If P
is a ΨDO over U , then φ∗ P : f 7→ P (φ∗ f ) ◦ φ−1 is a ΨDO over V , as can be verified by an
explicit change-of-variable calculation. If P is classical, then so also is φ∗ P . If pφ denotes the
symbol of φ∗ P , we find that the principal symbols are related by

pd (x, ξ) = pφd (φ(x), φ′ (x)−t ξ),

where φ′ (x)−t is the contragredient matrix (inverse transpose) to φ′ (x).

This is the change-of-variable rule for the cotangent bundle. The conclusion is that, for
any scalar ΨDO P that we may be able to define over a compact manifold M , the complete
symbol p(x, ξ) will depend on the local coordinates for a given chart of M , but the leading
term pd = σ P will make sense as an element of C ∞ (T ∗ M ) —i.e., a function on the total space
of the cotangent bundle. (The subleading terms pd−j (x, ξ), for j ≥ 1, will not be invariant
under local coordinate changes.)
When P is defined on sections of a vector bundle E → M of rank r, the principal symbol
σ becomes a section of the bundle π ∗ (End E) → T ∗ M , i.e., the pullback of End E → M via
P

the cotangent projection π : T ∗ M → M .

For the Dirac operator D, / which is a first-order differential operator on Γ(M, S), we get
σ D/ ∈ Γ(T ∗ M, π ∗ (End S)). From (5.3), we get at once

σ D/ (x, ξ) = c(ξj dxj ) = c(ξ).

51
Since taking the principal symbol is a multiplicative procedure, we also obtain
2
σ D/ (x, ξ) = (σD/ (x, ξ))2 = c(ξ)2 = g(ξ, ξ) 12m .
(Here we use the handy notation 1r for the r × r identity matrix.) Notice that the principal
/ 2 − ∆S = 14 s is a term of order zero (it is independent
symbol of ∆S is also g(ξ, ξ) 12m , since D
of the ξj variables), thus D/ 2 and ∆S have the same principal symbol.
2
Note that σ D/ (x, ξ) only vanishes when ξ = 0, that is, on the zero section of T ∗ M .
Definition 5.5. A ΨDO P is called elliptic if σP (x, ξ) is invertible when ξ 6= 0, i.e., off the
zero section of T ∗ M .
/ D
In particular, D, / 2 , ∆, ∆S are all elliptic differential operators.

5.2 Homogeneity of distributions

We now wish to pass from the symbol p of a classical ΨDO, with a given symbol expansion
N
X −1
p(x, ξ) = pd−j (x, ξ) + rN (x, ξ), rN ∈ S d (U ).
j=0

to the operator kernel (5.2), by taking an inverse Fourier transform. However, the terms in
this expansion may give divergent integrals when y = x. Therefore, we first need to look
more closely at the inverse Fourier transforms of negative powers of |ξ|.
Assume that n ≥ 2, for the rest of this section.
Definition 5.6. Let λ ∈ R. A function φ : Rn \ {0} → C is homogeneous of degree λ, or
“λ-homogeneous”, if
φ(t ξ) = tλ φ(ξ) for all t > 0, ξ 6= 0.
Thus if ξ = rω with r = |ξ| > 0 and ω = ξ/|ξ| ∈ Sn−1 , we can write φ(ξ) = rλ ψ(ω) for some
ψ : Sn−1 → C.
We can extend this definition to (tempered) distributions on Rn . Write φt for the dilation
of φ by the scale factor t, that is, φt (ξ) := φ(tξ), so that the λ-homogeneity condition can be
written as φt = tλ φ for t > 0.
The change-of variables formula for functions,
Z Z
n
u(tξ) φ(ξ) d ξ = t−n u(η) φ(η/t) dn η,
Rn Rn
suggests the following definition of homogeneity.
Definition 5.7. Let u ∈ S ′ (Rn ) be a temepered distribution on Rn . For t > 0, the dilation
ut of u by the scale factor t is defined by
hut , φi := t−n hu, φ1/t i, for φ ∈ S(Rn ).
We say that u is homogeneous of degree λ if ut = tλ u for all t > 0.
Example 5.8. The Dirac δ is homogeneous of degree −n, since for all φ ∈ S(Rn ),
hδt , φi = t−n hδ, φ1/t i = t−n φ1/t (0) = t−n φ(0) = t−n hδ, φi.
Suppose now that u is a smooth function on Rn \ {0}, such that
u(ξ) = rλ v(ω), for ξ = rω, r = |ξ| > 0, ω ∈ Sn−1 .
We would like to extend it to a (tempered) distribution on the whole Rn . There are several
cases to consider.

52
Case 1 If λ > 0, then just put u(0) := 0. In this case, u extends to Rn as a homogeneous
function.

Case 2 If −n < λ ≤ 0, then u(0) may not exist, but u(ξ) is locally integrable near 0, so
hu, φi is defined. Indeed, if B = B(0; 1) and 1B is its indicator function, and if σ denotes the
usual volume form on Sn−1 , then
Z Z Z 1
n
hu, 1B i := u(ξ) d ξ = v(ω) σ rλ (rn−1 dr)
B Sn−1 0
Z 1
=C rλ+n−1 dr < ∞, since λ + n − 1 > −1.
0

R
Case 3 Suppose λ = −n, and that Sn−1 v(ω) σ = 0.
We define a distribution Pu by the following trick. Let f : [0, ∞) → R be a cutoff function,
such that: (
1 if 0 ≤ t ≤ 12 ,
f (t) :=
0 if t ≥ 1,

and f decreases smoothly from 1 to 0 on [ 12 , 1]. Replace the test function φ by φ(ξ)−φ(0)f (|ξ|),
and put Z


hPu, φi := u(ξ) φ(ξ) − φ(0) f (|ξ|) dn ξ. (5.7)
Rn

If g(t) is another cutoff function with the same properties, the right hand side of this formula
changes by
Z Z Z 1


dr
u(ξ)φ(0) f (r) − g(r) dn ξ = φ(0) v(ω) σ f (r) − g(r) = 0,
Rn Sn−1 1/2 r

since u(ξ) dn ξ = r−n v(ω) σ rn−1 dr = v(ω) σ dr/r by homogeneity. Thus hPu, φi is indepen-
dent of the cutoff chosen. Indeed, since
Z Z 1 Z
n dr
u(ξ)f (|ξ|) d ξ = f (r) v(ω) σ = 0,
|ξ|>ε ε r Sn−1

for any ε > 0, we get another formula for Pu:

Z
hPu, φi = lim u(ξ)φ(ξ) dξ.
ε↓0 |ξ|>ε

Therefore, Pu is just the “Cauchy principal part” of u at ξ = 0.

Lemma 5.9. When u is a (−n)-homogeneous function on Rn \ {0} whose integral over Sn−1
vanishes, its principal-part extension Pu is a homogeneous distribution of degree (−n).

Proof. For each t > 0, we observe that

Z
h(Pu)t , φi = t−n hPu, φ1/t i = t−n lim u(ξ) φ(ξ/t) dξ
ε↓0 |ξ|>ε
Z
−n
= t lim u(η) φ(η) dη = t−n hPu, φi.
ε↓0 |η|>ε/t

53
Case 4 Consider the function u(ξ) := |ξ|−n for ξ 6= 0. (By averaging v(ω) over Sn−1 , one
can see that any smooth (−n)-homogeneous function on Rn \ {0} is a linear combination of
|ξ|−n and a function in Case 3.
We can try the cutoff regularization, anyway. Let Rf u be given by the recipe of (5.7):
Z


hRf u, φi := u(ξ) φ(ξ) − φ(0) f (|ξ|) dn ξ. (5.8)
Rn

However, in the present case, Rf u is not homogeneous!

Lemma 5.10. If δ : φ 7→ φ(0) is the Dirac delta, and if u(ξ) := |ξ|−n for ξ 6= 0, then
(Rf u)t − t−n Rf u = (Ωn t−n log t) δ. (5.9)
Proof. We compute h(Rf u)t − t−n Rf u, φi for φ ∈ S(Rn ). Since u(ξ) := |ξ|−n and f (|ξ|) are
both rotation-invariant, we can first integrate over Sn−1 , so we may suppose that φ is radial:
φ(ξ) = ψ(|ξ|) for some ψ : [0, ∞) → C. Then
Z
hRf u, φi = r−n (ψ(r) − ψ(0) f (r)) σ rn−1 dr
R n
Z ∞

dr
= Ωn ψ(r) − ψ(0) f (r)
r
Z0 ∞      
r r dr
= Ωn ψ − ψ(0) f , for any t > 0.
0 t t r
Therefore,
h(Rf u)t − t−n Rf u, φi = t−n hRf u, φ1/t − φi
Z ∞   
−n r dr
= Ωn φ(0) t f − f (r)
0 t r
Z ∞ Z r/t
dr
= Ωn φ(0) t−n f ′ (s) ds
r
Z0 ∞ Zr s
dr
= Ωn φ(0) t−n f ′ (s) ds
0 st r
Z 1
−n
= Ωn φ(0) t (− log t) f ′ (s) ds = Ωn φ(0) t−n log t.
1/2

The extra log t-term measures the failure of homogeneity of the regularization Rf u.

Case 5 u(ξ) = |ξ|−n−j for j = 1, 2, 3, . . . .

Any cutoff function f gives a regularization by “Taylor subtraction”, as follows:
Z  X iα 
ef u, φi :=
hR |ξ|−n−j φ(ξ) − Dα φ(0)ξ α f (|ξ|) dn ξ.
Rn α!
|α|≤j

Again one finds that Ref u is not homogeneous, by a straightforward calculation along the lines
of the previous Lemma. This can be simplified a little by the following observation [GVF].
One ef u −
P can find α
constants cα for |α| ≤ j, such that the modified regularization Rf u := R
|α|<j cα D δ has a “failure of homogeneity” of the form
X 
−n−j −n−j α
(Rf u)t − t Rf u = t log t cα D δ .
|α|=j

54
 That completes our study of the extensions of homogeneous functions to distributions on
Rn . We need a remark about their Fourier transforms. Recall that the Fourier transformation
F preserves the Schwartz space S(Rn ), and by duality it also preserves SR′ (Rn ). If u is a λ-
homogeneous function on Rn \ {0}, its Fourier transform is Fu(ξ) := Rn e−ixξ u(x) dn x,
thus Z Z
−n
(Fu)t (ξ) = eitxξ n
u(x) d x = t e−iyξ u(y/t) dn y = t−n−λ Fu(ξ).
Rn Rn

It follows that F, and also the inverse transformation F −1 , take homogeneous functions (or
distributions) of degree λ to homogeneous functions (or distributions) of degree (−n − λ).

5.3 The Wodzicki residue

Now we return to the symbol expansion of a classical ΨDO P , of integral order d ∈ Z, with
N
X −1
p(x, ξ) = pd−j (x, ξ) + rN (x, ξ),
j=0

where rN ∈ S d−N (U ), and pd−j (x, tξ) = td−j pd−j (x, ξ). Now apply F2−1 , the inverse Fourier
transform in the second variable, to this sum, to get the integral kernel
N
X −1
kP (x, y) = hj−d−n (x, x − y) + (F2−1 rN )(x, x − y).
j=0

If N > n + d, then rN ∈ S d−N (U ) is integrable in ξ, so the term F2−1 rN (x, z) is bounded as

z → 0. For the terms hj−d−n (x, z), there are 3 cases, which may give singularities. So before
applying F2−1 to pd−j (x, ξ), we must regularize pd−j (x, ξ) to Rf pd−j (x, ξ) by using a suitable
cutoff:
N
X −1
p(x, ξ) = Rf pd−j (x, ξ) + sN (x, ξ),
j=0

with sN integrable. Now take hj−d−n := F2−1 (Rf pd−j ).

Case 1 Suppose d − j > −n. Then k := j − d − n < 0, and Rf pd−j (x, ξ) is homogeneous of
degree greater than −n, so hk (x, z) is homogeneous of degree k. These terms have no failure
of homogeneity.
Before examining the other two cases, we return to the context of functions on Rn \ {0},
and look first at w0 (z) := (2π)n F −1 (Rf |ξ|−n ). Since (5.9) holds with u(ξ) = |ξ|−n for ξ 6= 0,
and since (2π)n F −1 (δ) = 1, we get

t−n w0 (z/t) − t−n w0 (z) = Ωn t−n log t for t > 0,

or more simply,
w0 (z/t) − w0 (z) = Ωn log t. (5.10)
Notice that C = w0 (z/|z|) is a constant, because w0 is rotation-invariant. Substituting t := |z|
in (5.10) gives
w0 (z) = C − Ωn log |z|, (5.11)

55
so that w0 “diverges logarithmically”. We can suppress the constant term if we replace
Rf |ξ|−n by Rf |ξ|−n − Cδ, since we must then subtracting the constant C from the inverse
Fourier transform.
For j = 1, 2, . . . , we define wj (z) := (2π)n F −1 (Rf |ξ|−n−j ). A similar analysis shows that
wj (z) = qj (z) − rj (z) log |z|, where both qj and rj are homogeneous of degree j > 0. In this
case, wj (z) remains bounded as z → 0.
We now return to the examination of the terms hj−d−n in the integral kernel k(x, y).

Case 2 Suppose d − j < −n. Then k = j − d − n > 0, and we find that hj−d−n (x, z) remains
bounded as z → 0.

Case 3 Consider the case d − j = −n. Then we get h0 (x, z) = −u0 (x) log |z|, after possibly
subtracting a term depending only on x. We have proved the following result.

Proposition 5.11. If P is a classical pseudodifferential operator of integral order d, then its

kernel has the following form near the diagonal:
X
kP (x, y) = hk (x, x − y) − u0 (x) log |x − y| + O(1), (5.12)
−d−n≤k<0

where each hk (x, ·) is homogeneous of negative degree k, u0 (x) is independent of x − y, and

O(1) stands for a term which remains bounded as y → x.

To compute u0 (x), the coefficient of logarithmic divergence, we change coordinates by a

local diffeomorphism ψ(x). Note that

log |ψ(x) − ψ(y)| ∼ log |ψ ′ (x) · (x − y)| ∼ log |x − y| as y → x,

while kP (x, y) 7→ kP (ψ(x), ψ(y)) L(x, y), where L(x, y) → | det ψ ′ (x)| as y → x, by the
change of variables formula for |dn y|. (We use a 1-density, not an oriented volume form,
to do integration; however, if we agree to fix an orientation on M and use only coordinate
changes that preserve the orientation, for which det ψ ′ (x) > 0 at each x, then we need not
make this distinction). Thus the log-divergent term transforms as follows:

−u0 (x) log |x − y| 7→ −u0 (ψ(x))| det ψ ′ (x)| log |x − y|.

For the case of scalar pseudodifferential operators, this is all we need. In the general case of
operators acting on sections of a vector bundle E → M , we replace u0 (x) ∈ End Ex by its
matrix trace tr u0 (x) ∈ C. The previous formula then says that the 1-density tr u0 (x) |dn x|
is invariant under local coordinate changes.
Now, when we regularize p−n (x, ξ) to obtain this 1-density after applying F2−1 , we can
first subtract the homogeneous “principal part”, at each x ∈ U , since this will not change
the coefficient of logarithmic divergence. This subtraction is done by replacing p−n (x, ξ) by
its average over the sphere |ξ| = 1 in the cotangent
R space Tx∗ M . That is to say, we get the
same u0 (x) if we replace p−n (x, ξ) by Ω−1n |ξ|
−n
|ω|=1 p−n (x, ω) σ. On applying (5.11) (with
C = 0) at each x, we conclude that
Z
tr u0 (x) = tr p−n (x, ω) σ.
|ω|=1

56
Definition 5.12. The Wodzicki residue density of a classical ΨDO P , acting on sections
of a vector bundle E → M , is well defined by the local formula
Z 
wresx P := tr p−n (x, ω) σ |dn x|, at x ∈ M.
|ω|=1

The Wodzicki residue of P is the integral of this 1-density:

Z Z Z 
Wres P := wresx P = tr p−n (x, ω) σ |dn x|.
M M |ω|=1

We shall now show that Wres is a trace on the algebra of classical pseudodifferential
operators on M acting on a given vector bundle.
We begin with another important property of homogneous functions on Rn \ {0}. We
shall make use of the Euler vector field on this space:
n
X ∂ ∂
R= ξj =r .
∂ξj ∂r
j=1

∂
Notice that h is λ-homogeneous if and only if Rh = λh, since Rh(rω) = r ∂r (rλ h(ω)) =
λrλ h(ω) = λh(rω).

Lemma 5.13. If λ 6= −n, any λ-homogeneous function h on Rn \ {0} is a finite sum of

derivatives.

Proof. It is enough to notice that

n
X ∂
(ξj h(ξ)) = nh(ξ) + Rh(ξ) = (n + λ)h(ξ),
∂ξj
j=1

1 Pn ∂
which implies h = n+λ j=1 ∂ξj (ξj h).

Lemma 5.14. If h : Rn \ {0} R → C is (−n)-homogeneous, with n > 1, then h is a finite sum

of derivatives if and only if Sn−1 h σ = 0.
R
Proof. Suppose first that Sn−1 h σ = 0. Since h(ξ) = r−n h(ω), h is determined by its
restriction to Sn−1 , and the hypothesis says that h1 | hi = 0 in L2 (Sn−1 , σ). Thus h ∈
(C1)⊥ = (ker ∆)⊥ = im ∆, where ∆ is the Laplacian on the sphere Sn−1 (which is a Fredholm
operator on the Hilbert space L2 (Sn−1 , σ), with closed range). Thus the equation h = ∆g
has a unique (and C ∞ , since ∆ is elliptic) solution g on Sn−1 . Extend g to Rn \ {0} by setting
g(rω) := r−n+2 g(ω) for 0 < r < +∞. Since the Laplacian on Rn is
 
1 ∂ ∂ 1
∆Rn = n−1 rn−1 + 2 ∆,
r ∂r ∂r r
P  
we get ∆Rn g = h, and thus h = nj=1 ∂ξ∂ j − ∂ξ ∂g
j
is a finite sum of derivatives.
∂f
Suppose instead that h(ξ) = ∂ξ 1
for ξ 6= 0, where f is (−n + 1)-homogeneous. Let σ ′ be
the volume form on Sn−2 , and notice that
Z Z Z
∂f ′ ′
(ξ1 , ω ) dξ1 ∧ σ = [f (+∞, ω ′ ) − f (−∞, ω ′ )] σ ′ = 0
Sn−2 R ∂ξ1 Sn−2

57
since f (ξ1 , ω ′ ) → 0 as ξ1 → ±∞, by homogenity. Thus we must show that
Z Z
hσ = h dξ1 ∧ σ ′ .
Sn−1 R×Sn−2

By Stokes’ theorem, we must show that the difference is the integral of the zero n-form on the
tube T , whose oriented boundary is (R × Sn−2 ) − Sn−1 . (Picture a ball stuck in a cylinder of
radius 1; T is the region inside the cylinder but outside the ball.) Consider the (n − 1)-form
n
X
σ̃ = cj ∧ · · · ∧ dξn ∈ An−1 (Rn \ {0}).
(−1)j−1 ξj dξ1 ∧ · · · ∧ dξ
j=1

If i : Sn−1 → Rn \ {0} is the inclusion, then σ = i∗ σ̃. Now σ̃ = ιR ν, where ν = dξ1 ∧ · · · ∧ dξn ,
and R is the Euler vector field on Rn \ {0}. Since Rh = −nh by homogenity, we find that

d(hσ̃) = dh ∧ σ̃ + h dσ̃ = dh ∧ σ̃ + nh ν
= dh ∧ ιR ν − (Rh) ν = dh ∧ ιR ν − ιR (dh)ν
= −ιR (dh ∧ ν) = ιR (0) = 0.

Thus hσ̃ is a closed form on Rn \ {0}, that restricts to hσ on Sn−1 and to h dξ1 ∧ σ ′ on
R × Sn−2 , therefore
Z Z Z Z
′ ′
hσ − h dξ1 ∧ σ = hσ = d(hσ̃) = 0,
Sn−1 R×Sn−2 ∂T T
R
by Stokes’ theorem. Thus Sn−1 h σ = 0, as required.

Proposition 5.15. Wres is a trace on the algebra of classical pseudodifferential operators

acting on a fixed vector bundle E → M .

Proof. We must show that Wres([P, Q]) = 0, for all classical pseudodifferential operators P ,
Q on M . First we consider the scalar case, where E = M . Let p(x, ξ), q(x, ξ) be the complete
symbols of P , Q respectively, in some coordinate chart of M . Then if r(x, ξ) is the complete
symbol of [P, Q], we know that
X i|α|
r(x, ξ) ∼ (Dξα p Dxα q − Dξα q Dxα p). (5.13)
n
α!
α∈N

In particular, the principal symbol of R = [P, Q] comes from the terms with |α| = 1 in this
expansion:
n 
X  n
X    
∂p ∂q ∂q ∂p ∂ ∂q ∂ ∂q
σR (x, ξ) = −i − = −i p j − j p .
∂ξj ∂xj ∂ξj ∂xj ∂ξj ∂x ∂x ∂ξj
j=1 j=1

In like manner, when αj = 2 and the other αi = 0, we get the terms

 
1 ∂2p ∂2q ∂2q ∂2p
− −
2 ∂ξj2 ∂(xj )2 ∂ξj2 ∂(xj )2
   
1 ∂ ∂p ∂ 2 q ∂3q 1 ∂ ∂p ∂ 2 q ∂3q
=− −p − −p 2 .
2 ∂ξj ∂ξj ∂(xj )2 ∂ξj ∂(xj )2 2 ∂xj ∂xj ∂ξj2 ∂ξj ∂xj

58
By induction, all terms in the expansion (5.13) that contribute to r−n (x, ξ) are finite sums
of derivatives.
In the general case, if p(x, ξ) = [pkl (x, ξ)] and q(x, ξ) = [qkl (x, ξ)] are square matrices, the
same argument applies to the sums
X
(Dξα pkl Dxα qlk − Dξα qlk Dxα pkl ), for each α ∈ Nn ,
k,l

that contribute to the expansion of tr r−n (x, ξ). Thus, tr r−n (x, ξ) is a finite sum of derivatives
in the variables xj and ξj . Write
n
X ∂fj ∂gj
tr r−n (x, ξ) = j
+ ,
∂x ∂ξj
j=1

where fj (x, ξ), gj (x, ξ) vanish outside K ×Rn for some compact P subset K ⊂ U ofPa coordinate
chart of M . (This can be guaranteed by first writing P = r ψr P and Q = r ψr Q for a
suitable partition of unity {ψr } on M .) Then
Z
Fj (x) := fj (x, ξ) σξ
|ξ|=1

has supp Fj ⊂ K, so that

Z Z
∂fj ∂Fj ∂Fj n
j
σ= , and |d x| = 0.
|ξ|=1 ∂x ∂xj U ∂xj

∂g
By construction, tr r−n (x, ξ), and each ∂ξjj (x, ξ) also, are (−n)-homogeneous in ξ. Lemma 5.14
now implies that
Z Z Xn  !
∂gj
Wres([P, Q]) = σ dn x = 0.
M |ξ|=1 ∂ξj j=1

To show that the trace is unique (up to constants) when n > 1, let T be any trace on
the algebra of classical pseudodifferential operators. Again we suppose that all symbols are
supported in a coordinate chart U ⊂ M , and we note that the formulas for composition of
symbols give the commutation relations

∂f ∂f
[xj , f ] = i , [ξj , f ] = −i .
∂ξj ∂xj

By Lemmas 5.13 and 5.14, T (P ) thus depends

R only on the homogeneous term tr p−n (x, ξ),
of degree −n, and moreover T (P ) = 0 if |ξ|=1 tr p−n (x, ξ) σ = 0. We can replace tr p−n (x, ξ)
R
with |ξ|−n |ξ|=1 tr p−n (x, ξ) σ, without changing T (P ). Now f 7→ T (f (x) |ξ|−n ) is a linear
functional on Cc∞ (U ) that kills derivatives with respect to each xj , so it is a multiple of the
Lebesgue integral: Z
T (f ) = C f (x) |dn x| for some C ∈ C.
U
Therefore, Z Z
T (P ) = C tr p−n (x, ξ) σ |dn x| = C Wres(P ).
U |ξ|=1

59
 / then
Example 5.16. If (M, g) is a compact Riemann spin manifold with Dirac operator D,

/ −n = 2m Ωn Vol(M ).
Wres |D|

Proof. Recall that the principal symbol of |D| / 2 (or of ∆S )

/ is σ D/ (x, ξ) = c(ξ), and that of D
2
is σ D/ (x, ξ) = c(ξ)2 = g(ξ, ξ) 12m . (Recall that 12m means the identity matrix of size 2m ,
/ −n
/ −n is σ |D|
which is the rank of the spinor bundle.) Thus the principal symbol of |D| (x, ξ) =
g(ξ, ξ)−n/2 12m . This is homogeneous of degree −n, so that p−n (x, ξ) is actually the principal
symbol when P = |D| / −n . Therefore, tr p−n (x, ξ) = 2m g(ξ, ξ)−n/2 .
Now g(ξ, ξ) = g ij ξi ξj in local coordinates on T ∗ M . To compute its integral over the
Euclidean sphere |ξ| = 1 [rather than over the ellipsoid g(ξ, ξ) = 1], we make a change
of coordinates x 7→ y = ψ(x), and we note that (x, ξ) 7→ (y, η) where ξ = ψ ′ (x)t η. We
can choose ψ such that ψ ′ (x) = [g ij (x)]1/2 , a positive-definite
R n × n matrix, in which case
ij kl 2 m −n m
g (x) ξi ξj = δ ηk ηl = |η| . Now tr p−n (y, η) = 2 |η| , so |η|=1 tr σ−n (y, η) = 2 Ωn , and
the Wodzicki residue density is

/ −n = 2m Ωn |dn y| = 2m Ωn det ψ ′ (x) |dn x|.

wresx |D|
p
But det ψ ′ (x)) = det g(x) by construction, so we arrive at
p
/ −n = 2m Ωn det(gx ) |dn x| = 2m Ωn νg .
wresx |D|

/ −n = 2m Ωn Vol(M ), as claimed.
Integrating this over M gives Wres |D|

What we have gained? We no longer need the full spectrum of the Dirac operator: its
principal symbol is enough to give the Wodzicki residue.

5.4 Dixmier trace and Wodzicki residue

There is a third method of computing the logarithmic divergence of the spectrum of D, / by
means of residue calculus applied to powers of pseudodifferential operators. We shall give
(only) a brief outline of what is involved.
Suppose that H is an elliptic pseudodifferential operator on Γ(M, E) that extends to a
positive selfadjoint operator (also denoted here 2
R by H) on the Hilbert space L (M, E),pwhich is
defined as the completion of { s ∈ Γ(M, E) : M (s | s) νg < ∞ } in the norm kψk := hψ | ψi,
where Z
hφ | ψi := (φ | ψ) νg
M
/ =
is the scalar product introduced in Section 3.1. We have in mind the example H = |D|
2 1/2 2 1/2
(D/ ) / + 1) , in case ker D
or else H = (D / 6= {0}.
Since M is compact, the operator H on L2 (M, E) is known to be Fredholm [Tay], thus
ker H is finite dimensional. We can define its powers H −s , for s ∈ C, by holomorphic
functional calculus: I
1
H −s := λ−s (λ − H)−1 dλ
2πi Γ
where Γ is a contour that winds once anticlockwise around the spectrum of H, excluding 0
to avoid the branch point of λ−s . (We define H −s ψ := 0 for ψ ∈ ker H.)
By applying the same Cauchy integrals to the complete symbol of H, one can show that
−s
H is pseudodifferential, and obtain much information about its integral kernel. This was
first done by Seeley [See]. He found that the following properties hold.

60
 • If H has order d > 0, then for ℜs > n/d, H −s is traceless and ζH (s) := Tr H −s is
holomorphic on this open half-plane.

• For x 6= y, the function s 7→ KH −s (x, y) extends from the half-plane ℜs > n/d to all of
C, as an entire function.

• For x = y, the function s 7→ KH −s (x, x) can be continued to a meromorphic function

on C, with possible poles only at { s = (n − k)/d : k = 0, 1, 2, . . . }.

• The residues at these poles are computed by integrating certain symbol terms over the
sphere |ξ| = 1 in Tx∗ M .
Later on, Wodzicki [Wodz] made a deep study of the spectral asymptotics of these oper-
ators, and in particular found that at s = n/d, the operator H −n/d is of order (−n), and the
residue at this pole depends only on its principal symbol; in fact,
1
Res KH −s (x, x) |dn x| = wresx H −n/d .
s=n/d d(2π)n

Corollary 5.17. If A is a positive elliptic ΨDO of order (−n) = − dim M on L2 (M, E),
then s 7→ Tr As is convergent and holomorphic on { s ∈ C : ℜs > −1 }, it continues mero-
morphically to C with a (simple) pole at s = 1, and
1
Res(Tr As ) = Wres A.
s=1 n(2π)n

(For the proof, one applies Seeley’s theory to H = A−1 .)

A basic result in noncommutative geometry is Connes’ trace theorem of 1988 [Con1],
which shows that this residue is actually a Dixmier trace.
Theorem 5.18 (Connes). If A is a positive elliptic ΨDO of order (−n) = − dim M on
H = L2 (M, E), the operator A lies in the Dixmier trace class L1+ (H), it is “measurable”,
i.e., Trω A =: Tr+ A is independent of ω, and the following equalities hold:
1
Tr+ A = Res(Tr As ) = Wres A.
s=1 n(2π)n
We omit the proof, but a few comments can be made. In view of what was already said,
it is enough to establish the first equality. The elliptic operator H = A−1 , of order n, has
compact resolvent [Tay], so that A itself is compact (we ignore any finite-dimensional kernel).
If the eigenvalues of A are λk = sk (A) (listed in decreasing order), the first equality reduces
to the following known theorem on divergent series:
P
Proposition 5.19 (Hardy). Suppose that λk ↓ 0 as k → ∞, that ∞ s
k=1 λk < ∞ for s > 1,
P∞ s 1 P N
and that lims↓1 (s − 1) k=1 λk = C exists. Then log N k=1 λk → C as N → ∞.
For a proof of the Proposition, see [GVF, pp. 294–295].
Next note that both Tr+ A and Wres A are bilinear in A, so we can weaken the positivity
hypothesis when comparing them. (There are other zeta-residue formulas available which are
bilinear in A, but we do not go into that here.)
Corollary 5.20. If A is a linear combination of positive elliptic pseudodifferential operators
of order (−n), then A ∈ L+ , A is measurable, and Tr+ A = n(2π)1
n Wres A.

61
Chapter 6

Spectral Triples: General Theory

6.1 The Dixmier trace revisited

Let (A, H, D) be a spectral triple, whose algebra A is unital. We continue to assume, for
convenience, that ker D = {0}, so that D−1 is a compact operator on H. Suppose now that
|D|−p ∈ L1+ , for some p ≥ 1. Then the functional on A given by a 7→ Trω (a |D|−p ), for some
particular ω, is our candidate for a “noncommutative integral”. To see why that should be
so, we first examine the commutative case.
Proposition 6.1. If M is a compact boundaryless n-dimensional spin manifold, with Rie-
/ then for each a ∈ C ∞ (M ) the operator a |D|
mannian metric g and Dirac operator D, / −n is
measurable, and Z
+ / −n
Tr (a |D| ) = Cn a νg ,
M
where Cn is a constant depending only on n, namely,
 1
m

C2m = if n = 2m,
2 Ωn m!(2π)m
Cn = n
, that is, 1
n(2π) 
C2m+1 = if n = 2m + 1.
(2m + 1)!! π m+1
/ −n
Proof. We know that |D| / −n is a ΨDO with principal symbol σ |D| (x, ξ) = gx (ξ, ξ)−n/2 12m ,
a scalar matrix of size 2m × 2m . As a multiplication operator on L2 (M, S), a is a ΨDO
of order 0, with principal symbol σ a (x, ξ) = a(x) 12m . Thus a|D| / −n is of order −n, with
a| / −n
D| m −n/2
tr σ (x, ξ) = 2 a(x) gx (ξ, ξ) .
Now by the trace theorem, we find, using the calculation in Example 5.16,
1
/ −n ) =
Tr+ (a|D| / −n )
Wres(a|D|
n(2π)n
Z Z 
2m
= a(x) gx (ξ, ξ)−n/2 σ |dn x|
n(2π)n M |ξ|=1
Z p
2m Ωn
= a(x) det gx |dn x|
n(2π)n
ZM
2m Ωn
= a(x) νg .
n(2π)n M
+ / −n ) is just the usual integral with respect to
Therefore, the functional a 7→ Tr (a |D|
the Riemannian volume form, expect for the normalization constant. Therefore, it can be
adapted to more general spectral triples as a “noncommutative integral”.

62
 However, in the noncommutative case, it is not obvious that a 7→ Tr+ (a |D| / −n ) will be
itself a trace. ¿Why should Tr+ (ab |D|
/ −n ) be equal to Tr+ (ba |D|
/ −n ) = Tr+ (a |D|
/ −n b)? To
check this tracial property of the noncommutative integral, we need the Hölder inequality for
Dixmier traces.
Fact 6.2 (Horn’s inequality). If T, S ∈ K and n ∈ N, then
n−1
X
σn (T S) ≤ sk (T ) sk (s). (6.1)
k=0

Proposition 6.3. (a) If T ∈ L1+ and S is a bounded operator on H, then for any Dixmier
trace Trω , the following inequality holds:

Trω |T S| ≤ (Trω |T |) kSk. (6.2a)

1 1
(b) Let 1 < p < ∞ and q = p/(p − 1), so that p + q = 1, and let T, S ∈ K be such that
|T |p , |S|q ∈ L1+ . Then for any Trω , we get

Trω |T S| ≤ (Trω |T |p )1/p (Trω |T |q )1/q . (6.2b)

Proof. Ad (a): By the minimax formula (4.2) for singular values, we find, for each k ∈ N,

sk (T S) = inf{ k(1 − P )T Sk : P = P 2 = P ∗ , rank P ≤ k }

≤ inf{ k(1 − P )T k kSk : P = P 2 = P ∗ , rank P ≤ k } = sk (T ) kSk,

Summing over k = 0, 1, . . . , n − 1, we get σn (T S) ≤ σn (T ) kSk. Thus

σn (T S) σn (T )
≤ kSk for all n ≥ 2,
log n log n
and linear interpolation gives the same relation with N replaced by and real λ ≥ 2. Using
the definition (4.4) of τλ and integrating over λ ≥ 3, we get

τλ (T S) ≤ τλ (T ) kSk for all λ ≥ 3,

and therefore Trω |T S| ≤ (Trω |T |)kSk for all ω.

Ad (b): From (6.1) and the ordinary Hölder inequality in Rn , we get
 X 1/p  X 1/q
p q
σn (T S) ≤ sk (T ) sk (S) = σn (|T |p )1/p σn (|S|q )1/q .
0≤k<n 0≤k<n

If n ≤ λ < n + 1 with λ = n + t, then, with an := σn (|T |p )1/p and bn := σn (|S|q )1/q ,

σλ (T S) = (1 − t)σn (T S) + tσn+1 (T S)
≤ (1 − t)an bn + tan+1 bn+1

1/p
1/q
≤ (1 − t)apn + tapn+1 (1 − t)aqn + taqn+1
= σλ (|T |p )1/p σλ (|S|q )1/q for all λ ≥ 2,

where we have used the Hölder inequality in R2 . Again we employ (4.4) and use the Hölder
Rλ
inequality for the integral log1 λ 3 (·) du
u . This gives

τλ (T S) ≤ τλ (|T |p )1/p τλ (|T |q )1/q , for λ ≥ 3.

63
Thus τ (|T S|) ≤ τ (|T |p )1/p τ (|T |q )1/q as positive elements of the corona C ∗ -algebra B∞ . Fi-
nally, we use the Hölder inequality for the state ω of this commutative C ∗ -algebra, namely


ω τ (|T |)p )1/p τ (|S|q )1/q ≤ ω(τ (|T |)p )1/p ω(τ (|S|q )1/q ,

and the result (6.2b) follows at once.

Proposition 6.4. Let (A, H, D) be any spectral triple whose operator D is invertible, and let
a ∈ A. Then the commutator [|D|r , a] is a bounded operator for each r such that 0 < r < 1.
We postpone the proof of this Proposition until later. It is a crucial property of spectral
triple that this bounded commutator property is not automatic for the case r = 1, that is,
the commutators [|D|, a] need not be bounded in general.
Theorem 6.5. If (A, H, D) is a spectral triple such that |D|−p ∈ L1+ (H) for some p ≥ 1,
then for each a ∈ A and any T ∈ L(H), the following tracial property holds:

Trω (aT |D|−p ) = Trω (T a |D|−p ) for all ω. (6.3)

Proof. Note that aT |D|−p and T a |D|−p lie in L1+ since L1+ is an ideal in L(H). Also the
Hölder inequality (6.2a) gives

Trω ([a, T ] |D|−p ) = Trω (T [|D|−p , a]) ≤ kT k Trω [|D|−p , a] ,

so we must show that Trω [|D|−p , a] = 0 for all a ∈ A. We have not supposed that p ∈ N, so
write p = kr with k ∈ N, 0 < r < 1, and let R := |D|−r , a positive compact operator. Then
k
X k
X
−p k j−1 k−j
[|D| , a] = [R , a] = R [R, a] R =− Rj [|D|r , a] Rk−j+1 .
j=1 j=1

Applying Hölder’s inequality to each term, we get

Trω Rj [|D|r , a] Rk−j+1 ≤ k[|D|r , a]k (Trω Rjpj )1/pj (Trω R(k−j+1)qj )1/qj ,

where qj = pj /(pj − 1) and the number pj > 1 must be chosen so that all Rjpj and all
R(k−j+1)qj are trace-class: for that, we need rjpj > p and r(k − j + 1)qj > p. This will
happen if we take
p p
pj := 1 , qj := ,
r(j − 2 ) r(k − j + 21 )
1
and then pj + q1j = 1, since rk = p. Since Trω vanishes on L1 (H), we need only to check that

|D|−p ∈ L1+ (H) =⇒ |D|−s ∈ L1 (H) for all s > p. (6.4)

This is a consequence of the next lemma.

Lemma 6.6. If A ∈ L1+ (H) and A ≥ 0, then As ∈ L1 (H) for s > 1.

Proof. We need the following result on sequence spaces. If E is a Banach space, we denote
by E ∗ the dual Banach space of continuous linear forms on E.
Fact 6.7. If s := { (s0 , s1 , . . . ) ∈ CN : (s0 + · · · + sn−1 )/ log n isPbounded }, if s0 is the
closure of the finite sequences in s, and if t := { (t0 , t1 , . . . ) ∈ CN : k≥0 |tk |/(k + 1) < ∞ },
then s0 , P
s and t are complete in the obvious norms, and under the standard duality pairing
hs, ti := k≥0 sk tk , there are isometric isomorphisms s∗0 ≃ t and t∗ ≃ s.

64
 Now let K− := { T ∈ K : {sk (T )}k≥0 ∈ t }, and let L1+
0 be the closure of the finite-rank
1+ 1+ ∗ − − ∗ 1+
operators in L . Then (L0 ) ≃ K and (K ) ≃ L as Banach spaces.
For T ∈ Lq with 1 < q < ∞, the Hölder inequality for sequences gives

X sk (T ) X 1/q X 1/p
1
≤ sk (T )q = kT kq ζ(p)1/p < ∞,
k+1 (k + 1)p
k≥0 k≥0 k≥0

so that Lq ⊂ K− for all 1 < q < ∞. Since (Lq )∗ ≃ Lp with p = q/(q − 1), we conclude that
L1+ ⊂ Lp for all p > 1. (This is why we employ the notation L1+ , of course.)
Now if A ∈ L1+ with A ≥ 0, then As ∈ Lp/s (H) whenever 1 < s ≤ p. In particular, when
p = s, we see that
X X X
kAs k1 = sk (As ) = λk (As ) = λk (A)s = (kAks )s < +∞,
k≥0 k≥0 k≥0

since A ∈ L1+ implies A ∈ Ls .

This establishes (6.4) and concludes the proof of Theorem 6.5.

Corollary 6.8. If A ≥ 0 is in L1+ , and Tr+ A > 0, then Tr+ As = 0 for s > 1.

To establish Proposition 6.4, we use the following commutator estimate, due to Helton
and Howe [HH].

Lemma 6.9. Let D be a selfadjoint operator on H, and let a ∈ L(H) with a(Dom D) ⊆
Dom D be such that [D, a] extends to a bounded operator on H. Suppose also that g : R → R
is smooth, with Fourier transform is a function ĝ such that t 7→ t ĝ(t) is integrable on R.
Then [g(D), a] extends to a bounded operator on H, such that
Z
1
k[g(D), a]k ≤ k[D, a]k |t ĝ(t)| dt.
2π R

Proof. We may define Z

1
g(D) := ĝ(t)eitD dt
2π R
for any smooth function g, since ĝ has compact support. Now
Z 1
itD d
hφ | [e , a]ψi = hφ | eistD a ei(1−s)tD ψi ds
0 ds
Z 1
d

= it hDe−istD φ | aei(1−s)tD ψi − hφ | eistD aDei(1−s)tD ψi ds
0 ds

for φ, ψ ∈ Dom D, and thus

hg(D)φ | aψi − hφ | ag(D)ψi

Z Z 1
1 d

= tĝ(t) hDe−istD φ | aei(1−s)tD ψi − hφ | eistD aDei(1−s)tD ψi ds dt
2π R 0 ds
Z
k[D, a]k
≤ kφk kψk |t ĝ(t)| dt,
2π R

so that [g(D), a] extends to a bounded operator, and the required estimate holds.

65
Proof of Proposition 6.4. We want to apply Lemma 6.9, using |x|r instead of g(x), x ∈ R.
But x 7→ |x|r is not smooth at x = 0 (although it is homogeneous of degree r), so we modify
it near x = 0 to get a smooth function g(x) such that g(x) = |x|r for |x| ≥ δ, for some δ > 0.
Thus g(x) = |x|r + h(x), where supp h ⊂ [−δ, δ]. We can write its derivative as a sum of two
terms, g ′ (x) = u(x) + h′ (x), where supp h′ ⊂ [−δ, δ] and u is homogeneous of negative degree
r − 1. Taking Fourier transforms on R, we get it ĝ(t) = û(t) + ĥ′ (t), where ĥ′ (t) is analytic
and û is homogeneous of degree −1 − (r − 1) = −r, with −1 < −r < 0. Thus tĝ(t) is locally
integrable near t = 0, and tĝ(t) → 0 rapidly for large t, since g ′ is smooth. We end up with
an estimate
k[|D|r , a]k ≤ Cr k[D, a]k + k[h(D), a]k,
R
where Cr := (2π)−1 R |tĝ(t)| dt is finite, and k[h(D), a]k is finite since h(D) is a bounded
operator.

6.2 Regularity of spectral triples

The arguments of the previous section are not applicable to determine whether [|D|, a] is
bounded, in the case r = 1. This must be formulated as an assumption. In fact, we shall ask
for much more: we want each element a ∈ A, and each bounded operator [D, a] too, to lie in
the smooth domain of the following derivation.
Notation. We denote by δ the derivation on L(H) given by taking the commutator with |D|.
It is an unbounded derivation, whose domain is
Dom δ := { T ∈ L(H) : T (Dom |D|) ⊆ Dom |D|, [|D|, T ] is bounded }.
We write δ(T ) := [|D|, T ] for T ∈ Dom δ.
Definition 6.10. A Tspectral triple (A, H, D) is called regular, if for each a ∈ A, the operators
a and [D, a] lie in k∈N Dom δ k .
The regularity condition does not depend on the invertibility of D (that is, the condition
ker D = {0}) which we have been assuming, to simplify certain calculations. One can always
replace |D| by hDi := (D2 + 1)1/2 in the definition, since f (D) := hDi − |D| is bounded. If
δ ′ denotes the derivation δ ′ (T ) := [hDi, T ] = δ(T ) + [f (D), T ], then clearly Dom δ ′ = Dom δ,
and it is easy to show by induction that Dom δ ′k = Dom δ k for each k ∈ N, so one may instead
define regularity using δ ′ . This is the approach taken in the work of Carey et al [CPRS], who
use the term “QC ∞ ” instead of “regular” for this class of spectral triples.
Definition 6.11. Suppose that (A, H, D) is a regular spectral triple, with D invertible. For
each s ≥ 0, consider the operator |D|s defined by functional
p calculus. Define also Hs :=
s
Dom |D|T for s ≥ 0, with T the Hilbert norm kξks := kξk2 + k|D|s ξk2 . Their intersection
H = s≥0 Dom |D| = ∞
∞ s k
k=0 Dom |D| is the smooth domain of the positive selfadjoint
operator |D|. Its topology is defined by the seminorms k · kk , for k ∈ N. Each Hs (and thus
also H∞ ) is complete, since the operators |D|s are closed, thus H∞ is a Fréchet space.
Since a ∈ A implies a ∈ Dom δ, we see that a(H1 ) ⊆ H1 , and then we can write a(|D|ξ) =
|D|(aξ) − [|D|, a] ξ for ξ ∈ H1 . Also,
kaξk21 = kaξk2 + k|D|aξk2
= kaξk2 + ka|D|ξ + δ(a)ξk2
≤ kaξk2 + 2kδ(a)ξk2 + 2ka|D|ξk2
≤ max{kak2 + 2kδ(a)k2 , 2kak2 } kξk21 ,

66
where we have used the parallelogram law kξ + ηk2 + kξ − ηk2 = 2kξk2 + 2kηk2 . Therefore,
a extends to a bounded operator on H1 . If (A, H, D) is regular, then by induction we find
that a(Hk ) ⊂ Hk continuously for each k, so that a(H∞ ) ⊂ H∞ continuously, too.

Definition 6.12. If r ∈ Z, let OprD be the vector space of linear maps T : H∞ → H∞ for
which there are constants Ck , for k ∈ N, k ≥ r, such that

kT ξkk−r ≤ Ck kξkk for all ξ ∈ H∞ .

Every such T extends to a bounded operator from Hk to Hk−r , for each k ∈ N. Note that
r+s
|D|r ∈ OprD for each r ∈ Z. If T ∈ OprD and S ∈ OpsD , then ST ∈ OpD .

Suppose (A, H, D) is regular. Then A ⊂ Op0D and [D, A] := { [D, a] : a ∈ A } ⊂ Op0D ,

too. Moreover, if a ∈ A, then

[D2 , a] = [|D|2 , a] = |D| [|D|, a] + [|D|, a] |D|

= |D|δ(a) + δ(a)|D| = 2|D|δ(a) − [|D|, δ(a)]
= 2|D|δ(a) − δ 2 (a),

so that [D2 , a] ∈ Op1D . Also [D2 , [D, a]] ∈ Op1D in the same way.
If b lies the subalgebra of L(H) generated by A and [D, A], we introduce

L(b) := |D|−1 [D2 , b] = 2δ(b) − |D|−1 δ 2 (b),

R(b) := [D2 , b] |D|−1 = 2δ(b) + δ 2 (b) |D|−1 . (6.5)

T
If b ∈ k≥0 Dom δ
k, then L(b) and R(b) lie in Op0D . The operations L and R commute:
indeed,

L(R(b)) = |D|−1 [D2 , [D2 , b] |D|−1 ] = |D|−1 [D2 , [D2 , b] |D|−1 = R(L(b)).

Note also that L2 (b) = |D|−2 [D2 , [D2 , b]].

T k l
T m
Proposition 6.13. If D is invertible, then k,l≥0 Dom(L R ) = m≥0 Dom δ ⊂ L(H).

Proof. We use the following identity for |D|−1 , obtained from the spectral theorem:

Z ∞
−1 2
|D| = (D2 + µ2 )−1 dµ, (6.6)
π 0

in order to compute the commutators. We shall show that Dom L2 ∩ Dom R ⊂ Dom δ.
Indeed, if b ∈ Dom L2 ∩ Dom R implies b ∈ Dom δ, then b ∈TDom L4 ∩ Dom L2 R ∩
T Dom R2
implies δb ∈ Dom L2 ∩ Dom R, so b ∈ Dom δ 2 . By induction, k,l≥0 Dom(Lk Rl ) ⊂ m≥0 δ m .
The converse inclusion is clear, from (6.5).

67
 Take b ∈ Dom L2 ∩ Dom R, and compute [|D|, b] as follows:

[|D|, b] = [D2 |D|−1 , b] = [D2 , b] |D|−1 + D2 [|D|−1 , b]

Z
2 ∞

= [D2 , b](D2 + µ2 )−1 + D2 [(D2 + µ2 )−1 , b] dµ
π 0
Z
2 ∞

= [D2 , b](D2 + µ2 )−1 − D2 (D2 + µ2 )−1 [D2 + µ2 , b](D2 + µ2 )−1 dµ
π 0
Z
2 ∞
= (1 − D2 (D2 + µ2 )−1 )[D2 , b](D2 + µ2 )−1 dµ
π 0
Z
2 ∞ 2 2
= µ (D + µ2 )−1 [D2 , b](D2 + µ2 )−1 dµ
π 0
Z
2 ∞

= [D2 , b](D2 + µ2 )−2 + [(D2 + µ2 )−1 , [D2 , b]](D2 + µ2 )−1 µ2 dµ
π 0
Z
2 ∞

= R(b)|D|(D2 + µ2 )−2 − (D2 + µ2 )−1 D2 L2 (b)(D2 + µ2 )−2 µ2 dµ
π 0
D2
Now R(b) and D2 +µ2
L2 (b) are bounded, by hypothesis. Also
Z ∞ Z ∞ Z π/2
2 xµ2 dµ 2 t2 dt 2 1
2 2 2
= 2
= sin2 θ dθ = ,
π 0 (x + µ ) π 0 1+t π 0 2
while Z Z Z
∞ 1 ∞
(D2 + µ2 )−2 µ2 dµ = (D2 + µ2 )−2 µ2 dµ + (D2 + µ2 )−2 µ2 dµ
0 0 1
is bounded by Z Z
1 ∞
1
kD−4 kµ2 dµ + µ−2 dµ = kDk−4 + 1.
0 0 3
Thus [|D|, b] is bounded with the estimate
 
1 1 −4 2
k[|D|, b]k ≤ kR(b)k + kDk + 1 kL2 (b)k.
2 3 π
Hence b ∈ Dom δ, as desired.

Corollary 6.14. The standard commutative example (C ∞ (M ), L2 (M, S), D)

/ is a regular
spectral triple.
Proof. We need one more fact from the theory of ΨDOs (see [Tay], for example): over a
compact manifold M , with a hermitian vector bundle E, a ΨDO of order zero is bounded as
an operator on L2 (M, E). Thus we need only show that, if b = a or b = [D,/ a] = −i c(da),
k l
then L R is a ΨDO of order ≤ 0, for each k, l ∈ N.
/ a]ψ = −i c(da)ψ are bounded multiplication
For k = l = 0, note that ψ 7→ aψ and ψ 7→ [D,
and Clifford-action operators. Their (principal) symbols are

σ a (x, ξ) = a(x) 12m ,

Xn
/
[D,a]
σ (x, ξ) = −i ∂j a(x)cj = −i {c(ξ), a(x)} = −i c(da).
j=1

2 2
For k + l > 0, we use Lk Rl = |D|/ −k (ad D / −l . Now D
/ )k+l (·)|D| / is a second-order ΨDO, with
2
D/ 2
σ2 (x, ξ) = g(ξ, ξ) 12m , so that when P is of order d then [D / , P ] is of order ≤ d + 1. Hence,

68
if a ∈ C ∞ (M ), then (ad D / 2 )k+l (a) is of order ≤ k + l, and thus Lk Rl is of order ≤ 0. The
same is true if a is replaced by −i c(da). Thus Lk Rl (b) is bounded, if b ∈ A or b ∈ [D,/ A].

This example also shows why regularity is defined using the derivation δ = [|D|, · ] instead
of the apparently simpler derivation [D, · ]. Indeed, we have just seen that for a ∈ C ∞ (M ),
the operator [|D|, / a]] has order zero (and therefore, it lies in Op0D/ . On the other hand,
/ [D,
/ [D,
[D, / a]] is in general a ΨDO of order 1 (and so it lies in Op1D/ ). Indeed, the first-order terms
in its symbol are

[σ D/ , σ [D,a]
/
](x, ξ) = [cj ξj , −i ck ∂k a(x)] = −i [cj , ck ] ξj ∂k a(x)

which need not vanish since cj , ck do not commute. In contrast, the principal symbol of |D|
/
/
is a scalar matrix, which commutes with that of [D, a], and the order of the commutator
drops to zero.

6.3 Pre-C*-algebras
If any spectral triple (A, H, D), the algebra A is a (unital) ∗-algebra of bounded operators
acting on a Hilbert space H [or, if one wishes to regard A abstractly, a faithful representation
π : A → L(H) is given]. Let A be the norm closure of A [or of π(A)] in L(H): it is a
C ∗ -algebra in which A is a dense ∗-subalgebra.
A priori, the only functional calculus available for A is the holomorphic one:
I
1
f (a) := f (λ)(λ1 − a)−1 dλ, (6.7)
2πi Γ

where Γ is a contour in C winding (once positively) around sp(a), and sp(a) means the
spectrum of a in the C ∗ -algebra A. To ensure that a ∈ A implies f (a) ∈ A, we need the
following property:
If a ∈ A has an inverse a−1 ∈ A, then in fact a−1 lies in A (briefly: A∩A × ×
H = A , where A×
1 −1
is the group of invertible elements of A). If this condition holds, then 2πi Γ f (λ)(λ1−a) dλ
is a limit of Riemann sums lying in A. To ensure convergence in A (they do converge in A), we
need only ask that A be complete in some topology that is finer than the C ∗ -norm topology.

Definition 6.15. A pre-C ∗-algebra is a subalgebra of A of a C ∗ -algebra A, which is stable

under the holomorphic functional calculus of A.

Remark 6.16. This condition appears in Blackadar’s book [Bla] under the name “local C ∗ -
algebra”. However, one can wonder how such a property could be checked in practice. Con-
sider the two conditions on a ∗-subalgebra A of a unital C ∗ -algebra A:

(a) A is stable under holomorphic functional calculus; that is, a ∈ A implies f (a) ∈ A,
according to (6.7).

(b) A is spectrally invariant in A [Schw], that is,

a ∈ A and a−1 ∈ A =⇒ a−1 ∈ A. (6.8)

In particular, (6.8) implies spA (a) = spA (a), for all a ∈ A.

69
Question: If A is known to have a (locally convex) vector space topology under which A is
complete (needed for convergence of the Riemann sums defining the contour integral) and
such that the inclusion A ֒→ A is continuous, ¿are (a) and (b) equivalent?
Ad (a) =⇒ (b): This is clear: if a ∈ A, a−1 ∈ A, use f (λ) := 1/λ outside spA (a).
Ad (b) =⇒ (a): To prove that the integral converges in A, because A is complete, we
need to show that the integrand is continuous. Note that since the inclusion i : A ֒→ A is
continuous, then A× = { a ∈ A : a−1 ∈ A } = A ∩ A× = i−1 (A× ) is open in A. But we
still need to show that a 7→ a−1 : A× → A× is continuous. This will follow if A is a Fréchet
algebra [Schw].

Corollary 6.17 (Schweitzer). If A is a unital Fréchet algebra, and if k · kA is continuous in

the topology of A, then Conditions (a) and (b) are equivalent.

If A is a nonunital algebra, we can always adjoin a unit in the usual way, and work
with Ae := C ⊕ A whose unit is (1, 0), and with its C ∗ -completion A e := C ⊕ A. Since the
e
multiplication rule in A is (λ, a)(µ, b) := (λµ, µa + λb + ab), we see that 1 + a := (1, a) is
e with inverse (1, b), if and only if a + b + ab = 0.
invertible in A,

Lemma 6.18. If A is a unital, Fréchet pre-C ∗ -algebra, then so also is Mn (A) = Mn (C) ⊗ A.

Sketch proof. It is enough to show that a ∈ Mn (A) is invertible for a close to the identity 1n
in the norm of Mn (A). But for a close to 1n , the procedure of Gaussian elimination gives
matrix factorization a =: ldu, where
     
1 0 ... 0 d1 0 . . . 0 1 ∗ ... ∗
∗ 1 . . . 0  0 d2 . . . 0 0 1 . . . ∗
     
l = . . . , d=. .. . . , u = . . . ,
 .. .. .. 0  .. . . 0  .. .. .. ∗
∗ ∗ ... 1 0 0 ... dn 0 0 ... 1

with dj ∈ A such that kdj − 1kA < 1, for j = 1, . . . , n. Thus d−1 exists, and a−1 =
u−1 d−1 l−1 ∈ Mn (A).
For n = 2, we get explicitly
   
1 0 a1 1 0 1 a−1
11 a12 ,
a=
a21 a−1
11 1 0 a22 − a21 a−1
11 a12 0 1

provided k1 − a11 kA < 1. For larger n, if k1n − akMn (A) < δ for δ small enough, we can
perform (n − 1) steps of Gaussian elimination (without any exchanges of rows or columns)
and get the factorization a = ldu in Mn (A) with d invertible.

Lemma 6.19. The Schwartz algebra S(Rn ) is a nonunital pre-C ∗ -algebra.

Proof. We represent S(Rn ) by multiplication operators on L2 (Rn ). Its C ∗ -completion is

C0 (Rn ). Note that C1 ⊕ C0 (Rn ) ≃ C(Sn ). Suppose f ∈ S(Rn ), and that there exists
g ∈ C0 (Rn ) such that (1 + f )(1 + g) = 1. Then f + g + f g = 0, and 1 + g = 1/(1 + f ) in
C(Sn ). Now, since f is C ∞ , then in particular g is smooth on Rn and all derivatives ∂ α g are
bounded. This entails that f g ∈ S(Rn ) also.
Finally, g = −f − f g lies in S(Rn ), so that (1 + g) = (1 + f )−1 lies in C1 ⊕ S(Rn ), as
required.

70
Example 6.20. If M is compact boundaryless smooth manifold, then C ∞ (M ) is a unital
Fréchet pre-C ∗ -algebra. The topology on C ∞ (M ) is that of “uniform convergence of all
derivatives”:
fk → f in C ∞ (M ) if and only if kX1 . . . Xr fk − X1 . . . Xr f k∞ → 0 as k → ∞,
for each finite set of vector fields {X1 , . . . , Xr } ∈ X(M ). This makes C ∞ (M ) a Fréchet space.
If f ∈ C ∞ (M ) is invertible in C(M ), then f (x) 6= 0 for any x ∈ X, and so 1/f is also smooth.
Thus C ∞ (M )× = C ∞ (M ) ∩ C(M )× .
We state, without proof, two important facts about Fréchet pre-C ∗ -algebras.
Fact 6.21. If A is a Fréchet pre-C ∗ -algebra and A is its C ∗ -completion, then Kj (A) =
Kj (A) for j = 0, 1. More precisely, if i : A → A is the (continuous, dense) inclusion, then
i∗ : Kj (A) → Kj (A) is an surjective isomorphism, for j = 0 or 1.
This invariance of K-theory was proved by Bost [Bost]. For K0 , the spectral invariance
plays the main role. For K1 , one must first formulate a topological K1 -theory is a category
of “good” locally convex algebras (thus whose invertible elements form an open subset and
for which inversion is continuous), and it is known that Fréchet pre-C ∗ -algebras are “good”
in this sense.
Fact 6.22. If (A, H, D) is a regular spectral triple, we can confer on A the topology given by
the seminorms
qk (a) := kδ k (a)k, qk′ (a) := kδ k ([D, a])k, for each k ∈ N. (6.9)
The completion Aδ of A is then a Fréchet pre-C ∗ -algebra, and (Aδ , H, D) is again a regular
spectral triple.
These properties of the completed spectral triple are due to Rennie [Ren]. We now discuss
another result of Rennie, namely that such completed algebras of regular spectral triples are
endowed with a C ∞ functional calculus.
Proposition 6.23. If (A, H, D) is a regular spectral triple, for which A is complete in the
Fréchet topology determined by the seminorms (6.9), then A admits a C ∞ -functional calculus.
Namely, if a = a∗ ∈ A, and if f : R → C is a compactly supported smooth function whose
support includes a neighbourhood of sp(a), then the following element f (a) lies in A:
Z
1
f (a) := fˆ(s) exp(isa) ds. (6.10)
2π R
Remark 6.24. One may use the continuous functional calculus in the C ∗ -algebra A to define
the one-parameter unitary group s 7→ exp(isa), for s ∈ R. Then the right hand side of (6.10)
coincides with the element f (a) ∈ A defined by the continuous functional calculus in A.

Proof. The map δ = ad |D| : A → L(H) is a closed derivation [BR] since |D| is a selfadjoint
operator. To show that f (a) ∈ Dom δ and that
Z
1
δ(f (a)) = fˆ(t) δ(exp(ita)) dt, (6.11)
2π R
we need to show that the integral on the right hand side converges. Indeed, by the same
token, the formula
Z 1
δ(exp(ita)) = it exp(ista) δ(a) exp(i(1 − s)ta) ds
0

71
shows that exp(ita) ∈ Dom δ because
Z 1 Z 1
it exp(ista) δ(a) exp(i(1 − s)ta) ds ≤ |t| kδ(a)k ds = |t| kδ(a)k,
0 0

and dominated convergence of the integral follows. Plugging this estimate into (6.11), we get
Z Z
1 ˆ 1
|f (t)| kδ(exp(ita))k dt ≤ kδ(a)k |tfˆ(t)| dt < +∞,
2π R 2π R

since f ∈ Cc∞ (R) implies fˆ ∈ S(R). Thus f (a) ∈ Dom δ, and (6.11) holds.
Now let Am , for m ∈ N, be the completion of A in the norm
m
X m
X
a 7→ qk (a) + qk′ (a) = kδ k (a)k + kδ k ([D, a])k.
k=0 k=0

For m = 0, we get
Z
1
kf (a)k + k[D, f (a)]k ≤ (|fˆ(t)| + k[D, a]k |tfˆ(t)|) dt
2π R

by replacing |D| by D in the

R previous argument.
Therefore, f (a) = 2π 1 ˆ
R f (s) exp(isa) ds is a limit of Riemann sums converging in the
′
norm q0 + q0 , so that f (a) ∈ A0 . Next, since δ and (ad D) are commuting derivations (on
A), we get [D, f (a)] ∈ Dom δ, with
Z
1
δ([D, f (a)]) ≤ |fˆ(t)| kδ([D, exp(ita)])k dt
2π R
Z
1

≤ |tfˆ(t)| kδ([D, a])k + |t2 fˆ(t)| kδ(a)k k[D, a]k dt,
2π R

since tfˆ(t) and t2 fˆ(t) lie in S(R). We conclude that δ extends to a closed derivation from A0
to L(H).
By an (ugly) induction on m, we find that for k = 0, 1, . . . , m f (a) and [D, f (a)] lie in
Dom δ k , and that δ T extends to a closed derivation from Am to L(H), and that f (a) ∈ Am .
By hypothesis, A = m∈N Am , and thus f (a) ∈ A.

Before showing how this smooth functional calculus can yield useful results, we pause for
a couple of technical lemmas on approximation of idempotents and projectors, in Fréchet
pre-C ∗ -algebras. The first is an adaptation of a proposition of [Bost].

Lemma 6.25. Let A be an unital Fréchet pre-C ∗ -algebra, with C ∗ -norm k · k. Then for
each ε with 0 < ε < 18 , we can find δ ≤ ε such that, for each v ∈ A with kv − v 2 k < δ and
k1 − 2vk < 1 + δ, there is an idempotent e = e2 ∈ A such that ke − vk < ε.

Proof. Consider the holomorphic function

1
√

f : { λ ∈ C : |λ| < 4 defined by f (λ) := 12 1 − 1 + 4λ ,
}→C
√
where we choose the branch of the square root for which 1 = +1. Note that f (0) = 0, and
that (1 − 2f (λ))2 = 1 + 4λ, so that

f (λ)2 − f (λ) = λ for |λ| < 14 .

72
If x ∈ A with kxk < 81 , then (1 + 4x)−1 exists since k1 − (1 + 4x)k < 12 , and
∞
X ∞
X kxk 1
kx(1 + 4x)−1 k ≤ kxk k(1 + 4x)−1 k ≤ kxk k(−4x)k k ≤ kxk k4xkk = < ,
1 − 4kxk 4
k=0 k=0

t
since increases from 0 to 14 for 0 ≤ t ≤ 18 .
1 − 4t
Now let x := v 2 − v, (thus kxk < 18 ), and let y := −x(1 + 4x)−1 = (v − v 2 )(1 − 2v)−2 ,
for which kyk < 14 . Note that kxk → 0 implies kyk → 0, which in turn implies kf (y)k → 0,
so that for each ε ∈ (0, 18 ), we can choose δ ≤ ε such that k1 − 2vkkf (y)k < ε whenever
k1 − 2vk < 1 + δ and kv − v 2 k = kxk < δ.
Finally, let vt := v + (1 − 2v)f (ty) for 0 ≤ t ≤ 1, and take e := v1 . Since f (0) = 0, we get
v0 = v. Our estimates show that ke − vk = k(1 − 2v)f (y)k < ε. By holomorphic functional
calculus, v ∈ A implies that x, y, vt , e all lie in A, too. We compute

vt2 − vt = (v + (1 − 2v)f (ty))2 − (v + (1 − 2v)f (ty))

= v 2 − v − (1 − 2v)2 f (ty) + (1 − 2v)2 f (ty)2


= v 2 − v + (1 − 2v)2 f (ty)2 − f (ty)
= v 2 − v + (1 − 2v)2 ty = (v 2 − v)(1 − t),

and in particular e2 − e = 0, as required.

Lemma 6.25 says that in a unital Fréchet pre-C ∗ -algebra A, an “almost idempotent”
v ∈ A that is not far from being a projector (since k1 − 2vk is close to 1) can be retracted
to a genuine idempotent in A. The next Lemma says that projectors in the C ∗ -completion
of A can be approximated by projectors lying in A.
Lemma 6.26. Let A be an unital Fréchet pre-C ∗ -algebra, whose C ∗ -completion is A. If
q̃ = q̃ 2 = q̃ ∗ is a projector in A, then for any ε > 0, we can find a projector q = q 2 = q ∗ ∈ A
such that kq − q̃k < ε.
Proof. For a suitable δ ∈ (0, 1), to be chosen later, we can find v ∈ A such that v ∗ = v and
kv − q̃k < δ, because A is dense in A. Now

kv 2 − vk ≤ kv 2 − q̃ 2 + q̃ − vk ≤ (kv + q̃k + 1)kv − q̃k < (3 + δ)δ < 4δ,

and
k1 − 2vk ≤ k1 − 2q̃k + 2kq̃ − vk < 1 + 2δ.
Lemma 6.25 now provides an idempotent e = e2 ∈ A such that ke − vk < ε/4, for δ small
enough (in particular, we must take δ < ε/4). To replace e by a projector q, we may use
Kaplansky’s formula (in the C ∗ -algebra A: see [GVF, p. 88], for example) to define

q := ee∗ (ee∗ + (1 − e∗ )(1 − e))−1 .

Indeed, ee∗ + (1 − e∗ )(1 − e) = 1 + (e − e∗ )(e∗ − e) ≥ 1 in A, so it is invertible in A, and thus

also in A because A is a pre-C ∗ -algebra. Thus q ∈ A. One checks that q ∗ = q and q 2 = q.
Note also that eq = q.
If A is represented faithfully on a Hilbert space H, we can decompose H as qH ⊕ (1 − q)H.
With respect to this decomposition, we can write
     
1 0 1 T R V
q= , e= , v= ,
0 0 0 0 V∗ S

73
where R = R∗ ∈ L(qH), S = S ∗ ∈ L((1 − q)H), and V, T : (1 − q)H → qH are bounded.
Now ke − vk < ε/4, so k(v − e)∗ (v − e)k < ε2 /16; it follows that

ε2 ε2
k(R − 1)2 + V V ∗ k < , k(V − T )∗ (V − T ) + S 2 k < .
16 16

Thus kV V ∗ k < ε2 /16, i.e., kV k < ε/4, and likewise kV − T k < ε/4. Therefore, kq − ek =
kT k < ε/2. Finally,
ε ε
kq − q̃k ≤ kq − ek + ke − vk + kv − q̃k < + + δ ≤ ε.
2 4
Theorem 6.27. Suppose (A, H, D) is a regular spectral triple, in which A is a unital Fréchet
pre-C ∗ -algebra; and assume that A is commutative. Let X = M (A) be the character space
of A, a compact Hausdorff space such that A ≃ C(X). Then, for each finite open cover
{ U1 , . . . , Um } of X, we can choose a subordinate partition of unity { φ1 , . . . , φm }:

φk ∈ C(X), 0 ≤ φk ≤ 1, supp φk ⊂ Uk , φ1 + · · · + φm = 1,

in such a way that φk ∈ A for k = 1, 2, . . . , m.

Remark 6.28. By definition, X = M (A) is the set of all nonzero ∗-homomorphisms φ : A → C.

Note that if φ ∈ M (A), then φ(1)2 = φ(1) implies φ(1) = 1. Gelfand’s theorem provided a
∗-isomorphism of unital C ∗ -algebras from A onto C(X), so that we can regard elements of A
as continuous functions on X.
Now let φ ∈ M (A). Recall that spA (a) = spA (a) for all a ∈ A, since A is a pre-C ∗ -algebra.
Because (a − λ1)b = 1 in A implies (φ(a) − λ)φ(b) = 1, we see that λ ∈ / spA (a) =⇒ λ 6= φ(a);
therefore, φ(a) ∈ spA (a) for all φ ∈ M (A). Thus |φ(a)| ≤ r(a) ≤ kakA , where r(a) is the
spectral radius of a, so that φ extends to A by continuity. This means that M (A) = M (A) =
X, so that we can regard X as the character space of the pre-C ∗ -algebra A.

Proof of Theorem 6.27. We first choose a partition of unity {φ̃1 , . . . , φ̃m } in C(X) =q A sub-
ordinate to {U1 , . . . , Um }. Let q̃ ∈ Mm (A) be the matrix whose (j, k)-entry is φ̃j φ̃k .
2 ∗
Then q̃ = q̃ = q̃ . Choose ε ∈ (0, 1/m). We apply Lemma 6.26 to the pre-C -algebra ∗

Mm (A) ⊂ Mm (A), to get q = q 2 = q ∗ ∈ Mm (A), with kq − q̃k < ε.

PmLet ψi := qij ∈ A for j = 1, . . . , m. Then ψ1 + · · · + ψm = tr q, and k tr q − tr q̃k ≤
j=1 kqjj − q̃jj k ≤ mε < 1, so that tr q = tr q̃, because q̃ 7→ tr q̃ is an integer-valued function
on X (and tr q̃ is the rank of the vector bundle corresponding to the projector q̃). Thus,

ψ1 + · · · + ψm = tr q = tr q̃ = φ1 + · · · + φm = 1.

Furthermore, 0 ≤ ψk ≤ 1 since ψk is the (k, k)-element of q = q ∗ q ∈ Mm (A), and thus

{ψ1 , . . . , ψm } is a partition of unity with elements in A. We can modify {ψk } to get {φk }
that will be subordinated to {Uk }, as follows.
Let g : R → [0, 1] be smooth with supp g ⊆ [ε, 1 + ε], and g(t) > 0 for ε < t ≤ 1. Now
Vk := { x ∈ X : ψk (x) > ε } ⊂ Uk , since kψk − φk k < ε. Let χ Pk m:= g ◦ ψk . By the smooth
functional calculus, P we find that χk ∈ A, for k = 1, . . . , m, and j=1 χj (x) > 0 for all x ∈ X,
since otherwise m j=1 ψj (x) ≤ mε < 1, impossible. Therefore, χ1 + · · · + χm is invertible
in C(X) = A, and thus also in A, so we can take φk := χk (χ1 + · · · + χm )−1 ∈ A, having
supp φk ⊂ Uk . Now {φ1 , . . . , φm } is the desired partition of unity.

74
6.4 Real spectral triples
Recall that a spin structure on an oriented compact manifold (M, ε) is represented by a
pair (S, C), where S is a B-A-bimodule and, according to Proposition 2.18, C : S → S is
an antilinear map such that C(ψ a) = C(ψ) ā for a ∈ A; C(b ψ) = χ(b̄) C(ψ) for b ∈ B;
and, by choosing a metric g on M , which determines a Hermitian pairing on S, we can also
require that (Cφ | Cψ) = (ψ | φ) ∈ A for φ, ψ ∈ R S. S may be completed to a Hilbert space
2
H = L (M, S), with scalar product hφ | ψi = M (φ | ψ) νg . It is clear that C extends to a
bounded antilinear operator on H such that hCφ | Cψi = hψ | φi by integration with respect
to νg , so that (the extended version of) C is antiunitary on H. Moreover, the Dirac operator
/ = −iĉ ◦ ∇S , where by construction the spin connection ∇S commutes with C: that is,
is D
S
∇X commutes with C, for each X ∈ X(M ).
The property C(ψa) = C(ψ) ā shows that, for each x ∈ X, ψ(x) 7→ C(ψ)(x) is an
antilinear operator Cx on the fibre Sx of the spinor bundle, which is a Fock space with
dimC Sx = 2m . Thus, to determine whether C commutes with D / or not, we can work with
the local representation D α S α α
/ = −iγ ∇Eα . Here γ = c(θ ), for α = 1, . . . , n, is a local section
of the Clifford algebra bundle Cl(T ∗ M ) → M , and the property C(bψ) = χ(b̄) C(ψ) says that
C(γ α ψ) = −γ α C(ψ) whenever ψ is supported on a local chart domain.
However, replacing b by γ α ∈ Γ(U, Cl1 (T ∗ M )) is only allowed when the dimension n is
even. In the odd case, B consists of sections of the bundle Cl0 (T ∗ M ), and we can only write
relations like C(γ α γ β ψ) = γ α γ β C(ψ) for ψ ∈ Γ(U, S). But since C is antilinear, in the even
case we get

/ = C(−iγ α ∇SEα ψ) = iC(γ α ∇SEα ψ) = −iγ α C(∇SEα ψ) = DCψ.

C Dψ /

/ C] = 0 on H, when n = 2m is even.
Thus [D,
¿What happens in the odd-dimensional case? Consider what happens on a single fibre
Sx , which carries a selfadjoint representation of Bx = Cl0 (Tx∗ M ). Recall that we use the
convention that c(ω) := c(ωγ) to extend the action of B to all of Γ(M, Cl(T ∗ M )), where
γ = (−i)m θ1 . . . θ2m+1 is the chirality element. For ω = θα , then gives C c(ω) C −1 =
C c(ωγ) C −1 = c(χ(ωγ)) = c(ωγ) since ωγ is even for ω odd, and ωγ = im θα θ1 . . . θ2m+1 =
(−1)m ωγ. We conclude that C c(ω) C −1 = (−1)m c(ω) for ω ∈ A1 (M ) real, and therefore
CD/ = (−1)m+1 DC / by antilinearity of C. We sum up:
(
/
+DC, if n 6≡ 1 mod 4,
/=
CD
/
−DC, if n ≡ 1 mod 4.

In the even case, B = Γ(M, Cl(T ∗ M )) contains the operator Γ = c(γ) which extends to a
selfadjoint unitary operator on H. Recall from Definition 1.18 that cJ (γ) is the Z2 -grading
operator on the Fock space Λ• WJ , the model for Sx . If H± = L2 (M, S ± ) denotes the
completion of S ± in the norm of H, then H = H+ ⊕ H− , with Γ being the Z2 -grading
operator. Now γ is even and γ̄ = (−1)m γ as before, so that CΓ = (−1)m ΓC whenever
n = 2m.
When M is a connected manifold, there is a third sign associated with C, since we know
that C 2 = ±1. Once more, the sign can be found by examining the case of a single fibre Sx ,
so we ask whether an irreducible representation S of Cl(V ) admits an antiunitary conjugation
C : S → S such that C cJ (v) C −1 = ±cJ (v) for v ∈ V (plus sign if dim V = 1 mod 4) and
either C 2 = +1 or C 2 = −1. By periodicity of the Clifford algebras, the sign depends only
on n mod 8, where n = dim V .

75
 Note that if {γ 1 , . . . , γ n } generate Cln,0 , then {−iγ 1 , . . . , −iγ n } generate Cl0,n = Cl(Rn , g)
with g negative-definite. Thus one can equally well work with Cl0,q , for q = 0, 1, . . . , 7. Since
Clp,0 ⊗R MN (R) ≃ Cl0,8−p ⊗R MN ′ (R) for p = 0, 1, . . . , 7 and suitable matrix sizes N, N ′ , we
get, from our classification (1.4) of the Clifford algebras Clp,0 :
• for q ≡ 0, 6, 7 mod 8, Cl0,q is an algebra over R,

• for q ≡ 1, 5 mod 8, Cl0,q is an algebra over C,

• for q ≡ 2, 3, 4 mod 8, Cl0,q is an algebra over H.

On a case-by-case basis, using this classification, one finds that C 2 = −1 if and only if
n = 2, 3, 4, 5 mod 8.
Exercise 6.29. Find five matrices ε1 , . . . , ε5 ∈ M4 (C), generating a representation of Cl05 ,
and an antiunitary operator C on C4 such that C εj C −1 = −εj for j = 1, . . . , 5. Show that
C is unique up to multiples C 7→ λC with λ ∈ C and |λ| = 1; and that C 2 = −14 .
Summary: There are two tables of signs

n mod 8 0 2 4 6 n mod 8 1 3 5 7
C 2 = ±1 + − − + C 2 = ±1 + − − +
/ = ±DC
CD / + + + + / = ±DC
CD / − + − +
CΓ = ±ΓC + − + −

There is a deeper reason why only these signs can occur, and why they depend on n mod 8:
the data set (A, H, D, / C, Γ) determines a class in the “Real” KR-homology KR• (A), and
j+8 j
KR (A) ≃ KR (A) by Bott periodicity. We leave this story for Prof. Brodzki’s course.
(But see [GVF, Sec. 9.5] for a pedestrian approach.)
“Real” KR-homology is a theory for algebras with involution: in the commutative case,
we may just take a 7→ a∗ , and we ask that C a C −1 = a∗ i.e., that C implement the involu-
tion. This is trivial for the manifold case, since C(ψa) = C(ψ)ā =: a∗ C(ψ), the a∗ here being
multiplication by ā.
In the noncommutative case, the operator Ca∗ C −1 would generate a second representation
of A, in fact an antirepresentation (that is, a representation of the opposite algebra A◦ ) and
we should require that this commute with the original representation of A.
Definition 6.30. A real spectral triple is a spectral triple (A, H, D), together with an
antiunitary operator J : H → H such that J(Dom D) ⊂ Dom D, and [a, Jb∗ J −1 ] = 0 for all
a, b ∈ A.
Definition 6.31. A spectral triple (A, H, D) is even if there is a selfadjoint unitary operator
Γ on H such that aΓ = Γa for all a ∈ A, Γ(Dom D) = Dom D, and DΓ = −ΓD. If no such
Z2 -grading operator Γ is given, we say that the spectral triple is odd.
We have seen that in the standard commutative example, the even case arises when
the auxiliary algebra B contains a natural Z2 -grading operator, and this happens exactly
when the manifold dimension is even. Now, the manifold dimension is determined by the
spectral growth of the Dirac operator, and this spectral version of dimension may be used for
noncommutative spectral triples, too. To make this more precise, we must look more closely
at spectral growth.

76
6.5 Summability of spectral triples
Definition 6.32. For 1 < p < ∞, there is an operator ideal Lp+ (H) = Lp,∞ (H), defined as
follows:
Lp+ (H) := { T ∈ K(H) : σN (T ) = O(N (p−1)/p ) as N → ∞ },
with norm kT kp+ := supN ≥1 σN (T )/N (p−1)/p .

1
For instance, if A ≥ 0 with sk (A) := , then A ∈ Lp+ by the integral test:
(k + 1)1/p
Z N
p
σN (A) ∼ t−1/p dt ∼ N (p−1)/p , as N → ∞.
1 p−1

Indeed, since p > 1, T ∈ Lp+ implies sk (T ) = O((k + 1)−1/p ). To see that, recall that
s0 (T )+· · ·+sk (t) = σk+1 (T ); since {sk (T )} is decreasing, this implies (k+1)sk (T ) ≤ σk+1 (T ),
1
and thus sk (T ) ≤ k+1 σk+1 (T ) ≤ C(k + 1)−1/p for some constant C.
1
Therefore, T ∈ Lp+ implies sk (T p ) = O( k+1 ) and then σN (T p ) = O(log N ), so that
T p ∈ L1+ , which serves to justify the notation Lp+ . It turns out, however, that there
are, for any p > 1, positive operators B ∈ L1+ such that B 1/p ∈ / Lp+ , so the implication
“T ∈ Lp+ =⇒ T p ∈ L1+ ” is a one-way street. For an example, see [GVF, Sec. 7.C].

Definition 6.33. A spectral triple (A, H, D) is p+ -summable for some p with 1 ≤ p < ∞
if (D2 + 1)−1/2 ∈ Lp+ (H). If D is invertible, this is equivalent to requiring |D|−1 ∈ Lp+ (H).

Definition 6.34. Let p ∈ [1, ∞). A spectral triple (A, H, D) has spectral dimension p if
it is p+ -summable and moreover

0 < Trω ((D2 + 1)−p/2 ) < ∞ for any Dixmier trace Trω .

If D is invertible, this is equivalent to 0 < Trω (|D|−p ) < ∞ for any Trω .

For positivity of all Dixmier traces, it suffices that lim inf N →∞ log1N σN ((D2 + 1)−p/2 ) > 0.
Note that, in view of Corollary 6.8, this can happen for at most one value of p.

Proposition 6.35. If (A, H, D) is a p+ -summable spectral triple, with D invertible, let

F := D |D|−1 (6.12)

be the phase of the selfadjoint operator D. Then, for each a ∈ A, the commutator [F, a] lies
in Lp+ (H).

Proof. First we show that [F, a] ∈ K(H), using the spectral formula (6.6) for |D|−1 . Indeed,

[F, a] = [D |D|−1 , a] = [D, a] |D|−1 + D [|D|−1 , a]

Z
2 ∞

= [D, a] (D2 + µ)−1 + D [(D2 + µ)−1 , a] dµ
π 0
Z
2 ∞

= [D, a] (D2 + µ)−1 − D(D2 + µ)−1 [D2 , a] (D2 + µ)−1 dµ (6.13)
π 0
Z
2 ∞ 2 2

= µ (D + µ)−1 [D, a] (D2 + µ)−1 − D(D2 + µ)−1 [D, a] D(D2 + µ)−1 dµ.
π 0

77
 In the integrand, [D, a] is bounded by the hypothesis that (A, H, D) is a spectral triple.
Next, (D2 + µ)−1 = (D − iµ)−1 (D + iµ)−1 ∈ K(H) and
1 1
D(D2 + µ)−1 = D(D2 + µ)− 2 (D2 + µ)− 2
| {z }| {z }
∈ L(H) ∈ K(H)

is also compact. Thus the integrand lies in K(H) for each µ, hence [F, a] ∈ K(H), that is, the
integral converges in the norm of this C ∗ -algebra.
Next to show that [F, a] ∈ Lp+ (H), we may assume that a∗ = −a, since

i [F, a] = [F, 2i (a∗ + a)] − i [F, 12 (a∗ − a)].

Note that this assumption implies that the bounded operators [F, a] and [D, a] are selfadjoint.
If we replace the term [D, a] by its norm k[D, a]k on the right hand side of (6.13), this
integral changes into
Z
2 ∞ 2 2

µ (D + µ)−1 k[D, a]k (D2 + µ)−1 − D(D2 + µ)−1 k[D, a]k D(D2 + µ)−1 dµ
π 0
Z ∞
2

= k[D, a]k (µ2 (D2 + µ)−2 + D2 (D2 + µ)−2 ) dµ
π
Z0 ∞
2

≤ k[D, a]k (µ2 (D2 + µ)−2 + D2 (D2 + µ)−2 ) dµ
π
Z 0∞
2
= k[D, a]k (D2 + µ)−1 dµ = k[D, a]k |D|−1 ,
π 0

where these are inequalities among selfadjoint elements of the C ∗ -algebra K(H). Therefore,
if we plug in the order relation

−k[D, a]k ≤ [D, a] ≤ k[D, a]k

among selfadjoint elements of L(H) into the right hand side of (6.13), we obtain the operator
inequalities
−k[D, a]k |D|−1 ≤ [F, a] ≤ k[D, a]k |D|−1 .
Thus the singular values of [F, a] are dominated by those of |D|−1 . We now conclude that
|D|−1 ∈ Lp+ implies [F, a] ∈ Lp+ , for all a ∈ A.

The assumption that D is invertible in the statement of Proposition 6.35 is not essential
(though the proof does depend on it, of course). With some extra work, we can modify the
proof to show that (D2 + 1)−1/2 ∈ Lp+ implies that all [F, a] ∈ Lp+ , where F is redefined to
mean F := D (D2 + 1)−1/2 , in contrast to (6.12). This is proved in [CPRS], in full generality.

78
Chapter 7

Spectral Triples: Examples

7.1 Geometric conditions on spectral triples

We begin by listing a set of requirements on a spectral triple (A, H, D), whose algebra A
is unital but not necessarily commutative, such that (A, H, D) provides a “spin geometry”
generalization of our “standard commutative example” (C ∞ (M ), L2 (M, S), D).
/ Again we
shall assume, for convenience, that D is invertible.

Condition 1 (Spectral dimension). There is an integer n ∈ {1, 2, . . . }, called the spectral

dimension of (A, H, D), such that |D|−1 ∈ Ln+ (H), and 0 < Trω (|D|−n ) < ∞ for any
Dixmier trace Trω .
When n is even, the spectral triple (A, H, D) is also even: that is, there exists a selfadjoint
unitary operator Γ ∈ L(H) such that Γ(Dom D) = Dom D, satisfying aΓ = Γa for all a ∈ A,
and DΓ = −ΓD.

Remark 7.1. It is useful to allow the case n = 0 as a possible spectral dimension. There are
two cases to consider:

• H is an infinite-dimensional Hilbert space, but the spectrum of the operator D has

exponential growth, so that N|D| (λ) = O(λε ) as λ → ∞, for any exponent ε > 0.
This is what happens in the example by Da̧browski and Sitarz [DS] of a spectral triple
on the standard Podleś sphere S2q with 0 < q < 1, where the operator D has the
same eigenvalue multiplicities as the Dirac operator on S2 (see Section A.2), but the
1 1
eigenvalue ±(l + 12 ) is replaced by ±(q −l− 2 − q l+ 2 )/(q −1 − q), for l = 21 , 23 , . . . .

• H is finite-dimensional, A is a finite-dimensional matrix algebra, and D is a hermitian

matrix. In this case, we assign to (A, H, D) the spectral dimension n = 0, and replace
the Dixmier traces Trω by the ordinary matrix trace tr.

Condition 2 (Regularity).
T For each a ∈ A, the bounded operators a and [D, a] lie in the
smooth domain k≥1 Dom δ k of the derivation δ : T 7→ [|D|, T ].
Moreover, A is complete in the topology given by the seminorms qk : a 7→ kδ k (a)k and
qk : a 7→ kδ k ([D, a])k. This ensures that A is a Fréchet pre-C ∗ -algebra.
′

T
Condition 3 (Finiteness). The subspace of smooth vectors H∞ := k∈N Dom Dk is a finitely
generated projective left A-module.
This is equivalent to saying that, for some N ∈ N, there is a projector p = p2 = p∗
in MN (A) such that H∞ ≃ AN p as left A-modules.

79
Condition 4 (Real structure). There is an antiunitary operator J : H → H satisfying
J 2 = ±1, JDJ −1 = ±D, and JΓ = ±ΓJ in the even case, where the signs depend only on
n mod 8 (and thus are given by the table of signs for the standard commutative examples).
Moreover, b 7→ Jb∗ J −1 is an antirepresentation of A on H (that is, a representation of the
opposite algebra A◦ ), which commutes with the given representation of A:
[a, Jb∗ J −1 ] = 0, for all a, b ∈ A.
Condition 5 (First order). For each a, b ∈ A, the following relation holds:
[[D, a], Jb∗ J −1 ] = 0, for all a, b ∈ A.
This generalizes, to the noncommutative context, the condition that D be a first-order differ-
ential operator.
Since
[[D, a], Jb∗ J −1 ] = [[D, Jb∗ J −1 ], a] + [D, [a, Jb∗ J −1 ]],
| {z }
=0
this is equivalent to the condition that [a, [D, Jb∗ J −1 ]] = 0.
Condition 6 (Orientation). There is a Hochschild n-cycle
P
c = j (a0j ⊗ b0j ) ⊗ a1j ⊗ · · · ⊗ anj ∈ Zn (A, A ⊗ A◦ ),
such that (
P 0 0∗ −1 Γ, if n is even,
πD (c) ≡ j aj (Jbj J ) [D, a1j ] . . . [D, anj ] = (7.1)
1, if n is odd.
In many examples, including the noncommutative examples we shall meet in the next two
sections, one can often take b0j = 1, so that c may be replaced, for convenience, by the cycle
P 0 1 n ◦
j aj ⊗ aj ⊗ · · · ⊗ aj ∈ Zn (A, A). In the commutative case, where A = A, this identification
may be justified: the product map m : A ⊗ A → A is a homomorphism.
The data set (A, H, D; Γ or 1, J, c) satisfying these six conditions constitute a “noncom-
mutative spin geometry”. In the fundamental paper where these conditions were first laid
out [Con2], Connes added one more nondegeneracy condition (Poincaré duality in K-theory)
as a requirement. We shall not go into this matter here.
To understand the orientation condition in the standard commutative example, we show
that c arises from a volume form on the oriented compact manifold M . Choose a metric g
on M and let νg be the corresponding Riemannian volume form. Furthermore, let {(Uj , aj )}
be a finite atlas of charts on M , where aj : Uj → Rn , and let {fj } be a partition of unity
subordinate to the open cover {Uj }; then for r = 1, . . . , n, each fj arj lies in C ∞ (M ) with
supp(fj arj ) ⊂ Uj . Over each Uj , let {θj1 , . . . , θjn } be a local orthonormal basis of 1-forms
(with respect to the metric g). Then
νg Uj
= θj1 ∧ · · · ∧ θjn = hj da1j ∧ · · · ∧ danj ,

for some smooth functions hj : Uj → C. We write a0j := (−i)m fj hj ∈ C ∞ (M ), where as

usual, n = 2m or n = 2m + 1. With that notation, we get
P
P
(−i)m νg = (−i)m j fj νg U = j a0j da1j ∧ · · · ∧ danj .
j

Now we define
1 X X σ(1) σ(n)
c := (−1)σ a0j ⊗ aj ⊗ · · · ⊗ aj . (7.2)
n!
σ∈Sn j

80
Exercise 7.2. Show that the Hochschild boundary bc of the chain (7.2) is zero because A is
commutative.
Therefore, c is a Hochschild n-cycle in Zn (A, A), for A = C ∞ (M ). Its representative as
a bounded operator on H is
1 X X 1 X X
(−1)σ / aσ(1)
a0j [D, j
/ aσ(n)
] . . . [D, j ]= (−1)σ
σ(1) σ(n)
a0j c(daj ) . . . c(daj )
n! n!
σ∈Sn j σ∈Sn j
(−i)m X X σ(1) σ(n)
= fj (−1)σ c(θj ) . . . c(θj )
n!
j σ∈Sn
X 
= fj (−i)m c(θj1 ) . . . c(θjn )
j

= c(γ) = Γ or 1,
since c(γ) = Γ for n = 2m, and c(γ) = 1 for n = 2m + 1.
This calculation shows that the elements a1j , . . . , anj occurring in the cycle c are local
coordinate functions for M . An alternative approach would be to embed M in some RN and
take the arj to be some of the cartesian coordinates of RN , regarded as functions on M . This
is illustrated in the following example.
Example 7.3. By regarding the sphere S2 as embedded in R3 ,
S2 = { (x, y, z) ∈ R3 : x2 + y 2 + z 2 = 1 },
we can write down its volume form for the rotation-invariant metric g as
ν = x dy ∧ dz + y dz ∧ dx + z dx ∧ dy.
The corresponding Hochschild 2-cycle is
i X
c := − (x ⊗ y ⊗ z − x ⊗ z ⊗ y),
2
cyclic

summing over cyclic permutations of the letters x, y, z.

/ is the Dirac operator on S2 for this “round” metric and the unique spin structure
If D
2
on S compatible with its usual orientation (see Section A.2), then
 
i X
1 0
− / y] [D,
x [D, / z] − x [D,
/ z] [D,
/ y] = Γ =
2 0 −1
cyclic

on H = H+ ⊕ H− , which is the completion of the spinor module S = S + ⊕ S − .

Consider the following element of M2 (A), with A = C ∞ (S2 ):
 
1 1 + z x + iy
p := . (7.3)
2 x − iy 1 − z
Note that tr(p − 21 ) = 0, where tr(a) := a11 + a22 means the matrix trace tr : M2 (A) → A.
Exercise 7.4. Show that, if A is any ∗-algebra and p ∈ M2 (A) is given by (7.3), then the
projector relations p = p∗ = p2 are equivalent to the following relations among x, y, z ∈ A:
x∗ = x, y ∗ = y, z ∗ = z,
[x, y] = [x, z] = [y, z] = 0,
x2 + y 2 + z 2 = 1.

81
Exercise 7.5. Check that tr(p dp ∧ dp) = − 2i ν.
If we replace − 2i ν by the Hochschild 2-cycle c, the same calculation that solves the
previous exercise also shows that πD (c) = Γ.
This computation has a deeper significance. One can show that the left A-module M2 (A) p
is isomorphic to E1 in our classification of A-modules of sections of line bundles over S2 ; and
we have seen in Section A.3 that E1 ≃ Γ(S2 , L) where L → S2 is the tautological line bundle.
The first Chern class c1 (L) equals (a standard multiple of) [ν] ∈ HdR 2 (S2 ). One can trace

a parallel relation between spinc structures on S2 defined, via the principal U (1)-bundle
SU (2) → S2 , on associated line bundles, and the Chern classes of each such line bundle. For
that, we refer to [BHMS].

7.2 Isospectral deformations of commutative spectral triples

To some extent, one can recover the sphere S2 from spectral triple data alone. Thus, if A is
a ∗-subalgebra of some C ∗ -algebra containing elements x, y, z, and if the matrix
 
1 1 + z x + iy
p= ∈ M2 (A)
2 x − iy 1 − z
is a projector, i.e., p = p∗ = p2 , then by Exercise 7.4, the elements x, y, z commute, they
are selfadjoint, and they satisfy x2 + y 2 + z 2 = 1. Thus, the commutative C ∗ -algebra A
generated by x, y, z is of the form C(X), where X ⊆ S2 is a closed subset. If A is now a
pre-C ∗ -subalgebra of A containing x, y, z, and is the algebra of some spectral triple (A, H, D),
then the extra condition πD (c) = Γ can only hold if X is the support of the measure ν. This
means that X = S2 .
A similar argument can be tried, to obtain an “algebraic” description of S4 . What follows
is a heuristic motivation, following [CL]. One looks for a projector p ∈ M4 (A), of the form
 
1+z 0 a b    
 0 1 + z −b∗ a  (1 + z)12 q a b

p= ∗  = , where q = .
a −b 1 − z 0  q∗ (1 − z)12 −b∗ a
b∗ a∗ 0 1−z

Then p = p∗ only if z = z ∗ , and then p2 = p implies that −1≤ z ≤ 1 in the C ∗ -completion

 z 0
 ∗ = q∗q = 1 − z2 0
A of A, , q = 0, and qq . From that, one finds that
0z 0 1 − z2
a, a∗ , b, b∗ , z commute, and aa∗ + bb∗ = 1 − z 2 . Thus A = C(X) with X ⊆ S4 . Again, it can
be shown that the equality X = S4 is reached by some extra conditions, namely,
tr(p − 12 ) = 0,
tr((p − 12 ) dp dp) = 0 in Ω2 (A),
πD ((p − 21 ) dp dp dp dp) = Γ in L(H).
 
a b
However, if one takes instead q := with λ = e2πiθ , then there is another,
−λ̄b∗ a∗
noncommutative, solution [CL]: now A is the C ∗ -algebra generated by a, b and z = z ∗ , where
z is central, and the other relations are
ab = e−2πiθ ba, a∗ b = e2πiθ ba∗ ,
aa∗ = a∗ a, bb∗ = b∗ b, aa∗ + bb∗ = 1 − z 2 . (7.4)

82
To find a solution to these relations, where the central element z is taken to be a scalar
multiple of 1, we substitute

a = u sin ψ cos φ
b = v sin ψ cos φ
z = (cos ψ) 1

with −π ≤ ψ ≤ π and −π < φ ≤ π, say. In this way, the commutation relations (7.4)
reduce to
uu∗ = u∗ u = 1, vv ∗ = v ∗ v = 1, vu = e2πiθ uv.
These are the relations for the unitary generators of a noncommutative 2-torus: see Sec-
tion A.4. Thus, by fixing values of φ, ψ with ψ 6= ±π and φ ∈ / π2 Z, we get a homomor-
phism 2 ∗
 from 
A to C(Tθ ), the C -algebra of the noncommutative 2-torus with parameters
0 θ
Θ= ∈ M2 (A).
−θ 0
We look for a suitable algebra A, generated by elements satisfying the above relations, by
examining a Moyal deformation of C ∞ (S4 ). One should first note that S4 ⊂ R5 = C × C × R
carries an obvious action of T2 , namely,

(t1 , t2 ) · (α, β, z) := (t1 α, t2 β, z), for |t1 | = |t2 | = 1,

which preserves the defining relation αᾱ + β β̄ + z 2 = 1 of S4 . The action is not free: there
are two fixed points (0, 0, ±1), and for each t with −1 < t < 1 there are two circular orbits,
namely { (α, 0, t) : αᾱ = 1 − t2 } and { (β, 0, t) : β β̄ = 1 − t2 }. The remaining orbits are copies
of T2 . The construction which follows will produce a “noncommutative space” S4θ that can
be thought of as the sphere S4 with each principal orbit T2 replaced by a noncommutative
torus T2θ , while the S1 -orbits and the two fixed points remain unchanged.
In quantum mechanics, the Moyal product of two functions f, h ∈ S(Rn ) is defined as
an (oscillatory) integral of the form
Z Z
−n −1
(f ⋆ h)(x) := (πθ) f (x + s) h(x + t) e−2is(Θ t) ds dt. (7.5)
Rn Rn

Here n = 2m is even, Θ = −Θt ∈ Mn (R) is an invertible skewsymmetric matrix, and

det Θ = θn with θ > 0. In the next section, we shall interpret this formula in a precise
manner (see Definition 7.17 below), and show that f ⋆ h lies in S(Rn ) also. Formally, at any
rate, one can rewrite it as an ordinary Fourier integral:
Z Z
(f ⋆ h)(x) := (2π)−n f (x − 12 Θu) h(x + t) e−iut du dt,
Rn Rn

with the advantage that now Θ need not be invertible (so that n need no longer be even). It
was noticed by Rieffel [Rie] that one can replace the translation action of Rn on f, h by any
(strongly continuous) action α of some Rl on a C ∗ -algebra A. Then, given Θ = −Θt ∈ Ml (R),
one can define Z Z
a ⋆ b := α 1 Θu (a) α−t (b) e2πiut du dt,
2
Rl Rl

provided that the integral makes sense. In particular, if α is periodic action of Rl , i.e.,
αt+r = αt for r ∈ Zn , so that α is effectively an action of Tn = Rn /Zn , then one can describe
the Moyal deformation as follows.

83
Definition 7.6. Let A be a unital C ∗ -algebra, and suppose that there is an action α of Tl
on A by ∗-automorphisms, which is strongly continuous. For each r ∈ Zl , let A(r) be the
spectral subspace

A(r) := { a ∈ A : αt (a) = tr a for all t ∈ Tl }, where tr := tr11 . . . trnn ∈ T.

Let A := { a ∈ A : t 7→ αt (a) is smooth } be the “smooth subalgebra” for the action of Tl .

It can be shown that P A is a Fréchet pre-C ∗ -algebra, and each a ∈ A can be written as a
convergent series a = r∈Zl ar , where ar ∈ A(r) and kar k → 0 rapidly as |r| → ∞.
Definition = −Θt ∈ Ml (R). The Moyal
P 7.7. Fix Θ P P product of two elements a, b ∈ A,
with a = r ar and b = s bs , is defined as a ⋆ b := r,s ar ⋆ bs , where
 P
ar ⋆ bs := σ(r, s) ar bs , with σ(r, s) := exp −πi lj,k=1 rj θjk sk . (7.6)

For actions of Tl , Rieffel [Rie] showed that the integral formula and the series formula for
a ⋆ b are equivalent, when a, b belong to the smooth subalgebra A.
Definition 7.8. Let M be a compact Riemannian manifold, carrying a continuous action
of Tl by isometries {σt }t∈Tl . Then αt (f ) := f ◦ σt is a strongly continuous action of Tl .
Given Θ = −Θt ∈ Ml (R), Rieffel’s construction provides a Moyal product on C ∞ (MΘ ) :=
(C ∞ (M ), ⋆), whose C ∗ -completion in a suitable norm is C(MΘ ) := (C(M ), ⋆). In particular,
0 θ
for M = S4 with the round metric and Θ = , these are the algebras C ∞ (S4θ ) and
−θ 0
C(S4θ ) introduced by Connes and Landi [CL].
To deform the spectral triple (C ∞ (M ), L2 (M, S), D),
/ we need a further step. Since
each σt is an isometry of M , it defines an automorphism of the tangent bundle T M (with
Tx M → Tσt (x) M ), and of the cotangent bundle T ∗ M (with Tσ∗t (x) M → Tx∗ M ), preserving
the orientation and the metric on each bundle. But the group SO(Tx∗ M, gx ) does not act
directly on the fibre Sx of the spinor bundle. Instead, the action of the Clifford algebra B on
H = L2 (M, S) yields a homomorphism Spin(Tx∗ M, gx ) → End(Sx ) for each x ∈ M , and we
know that there is a double covering Adx : Spin(Tx∗ M, gx ) → SO(Tx∗ M, gx ) by conjugation.
It turns out [CDV] that one can lift the isometric action α : Tl → SO(T ∗ M ) to an action
of another torus τ : Te l → Aut(S), where there is a covering map π : T e l → Tl such that
π(±1) = 1, making the following diagram commute:

el τt✲
T Aut(S)
π Ad
❄ ❄
α✲t
Tl SO(T ∗ M )

Fact 7.9. One can find a covering of Tl by a torus T e l , and a representation t̃ 7→ τ of T

e l on
t̃
l ∞
Aut(S) such that Ad(τt̃ ) = αt if π(t̃) = t ∈ T . For f ∈ A = C (M ) and φ, ψ ∈ S = Γ(M, S),
this implies that

τt̃ (f ψ) = αt (f ) τt̃ ψ, (7.7)

(τt̃ φ | τt̃ ψ) = αt (φ | ψ). (7.8)

Integrating over M , and recalling that σt is an isometry, we get hτt̃ φ | τt̃ ψi = hφ | ψi, so that
τ extends to a unitary representation of Te l on H = L2 (M, S).

84
 We can regard Tl as Rl /(Zl + Ẑl ), where Ẑl = Zl + ( 21 , 12 , . . . , 12 ). With this convention,
one can show that the set of commuting selfadjoint operators P1 , . . . , Pl on H which generate
e l , i.e.,
the unitary representation of T

τt̃ =: exp(it1 P1 + · · · + itn Pn ),

have half-integer spectra: sp(Pj ) ⊆ 21 Z.

Now define a family of unitary operators { σ(r, P ) : r ∈ Zl } by
 P 
σ(r, P ) := exp −πi j,k rj θjk Pk ,

that is, we substitute sk by Pk in the cocycle formula σ(r, s) of (7.6).

Exercise 7.10. Show that, since the action α is isometric, the unitary operator τt̃ commutes
el.
/ and with the charge conjugation operator C, for each t̃ ∈ T
with D
It follows that each σ(r, P ) commutes with D, / too. However, the operators σ(r, P ) need
not commute with the multiplication operators ψ 7→ f ψ, for f ∈ C ∞ (M ). Indeed, (7.7)
el.
implies that τt̃ f τ−t̃ = αt (f ) for each t̃ ∈ T
Exercise 7.11. If hs ∈ A(s) and r ∈ Zl , show that σ(r, P ) hs = hs σ(r, P + s).
We are now ready to exhibit the isospectral deformation of the standard commutative
example for a spin manifold M carrying an isometric action of Tl , with respect to a fixed
matrix Θ of deformation parameters. The deformation is called isospectral for the simple
reason that the operator D of the deformed spectral triple is the same Dirac operator of the
undeformed case, so it is no surprise that its spectrum does not change. What does change is
the algebra: in fact, the underlying vector space of A is unchanged, but the product operation
is deformed, and consequently its representation on H changes, too.
Theorem 7.12. If AΘ = (C ∞ (M ), ⋆), H = L2 (M, S) and D = D, / then there is a represen-
/
tation of AΘ by bounded operators on H, such that (AΘ , H, D) is a spectral triple with the
same Dirac operator as the standard commutative example (C ∞ (M ), H, D). / Moreover, the
charge conjugation operator C is a real structure on (AΘ , H, D), / and the first order property
holds.
P
Proof. If f ∈ A, write f = r∈Zl fr as a decomposition into spectral subspaces, where
r l l
αt (fr ) = t fr for t ∈ T , r ∈ Z . Define
X
L(f ) := fr σ(r, P ) ∈ L(H).
r

Then f 7→ L(f ) is a representation of the algebra AΘ :

X
L(f )L(h) = fr σ(r, P ) hs σ(s, P )
r,s
X
= fr hs σ(r, P + s) σ(s, P )
r,s
X
= fr hs σ(r, s) σ(r + s, P )
r,s
X
= fr ⋆ hs σ(r + s, P ) = L(f ⋆ h).
r,s

85
The last equality follows because fr ∈ A(r) , hs ∈ A(s) imply that both fr hs and fr ⋆ hs lie
in A(r+s) —these products differ only by the phase factor σ(r, s)— and therefore (f ⋆ h)p =
P
r+s=p fr ⋆ hs .

P Since αt (fr∗ ) =Pαt (fr )∗ = t−r fr∗ , we see that (f ∗ )s = (f−s )∗ for s ∈ Zl . Thus L(f )∗ =
∗ ∗ ∗
r fr σ(−r, P ) = r (f )−r σ(−r, P ) = L(f ), so that L is actually a ∗-representation.
/ commutes with each σ(r, P ), we get
Since D
X
/ L(f )] =
[D, / fr ] σ(r, P ) =: L([D,
[D, / f ]), (7.9)
r

where we remark that τt̃ [D,/ fr ] τ−t̃ = [D, / τt̃ fr τ−t̃ ] = tr [D,
/ fr ], so that the operators [D,/ f ],
for f ∈ A, decompose into spectral subspaces under the action t 7→ Ad(τt̃ ) by automorphisms
of L(H), which extends t 7→ αt by automorphisms of A.
Next, since the antilinear operator C commutes with all unitaries σ(r, P ), we deduce that
CPj C −1 = −Pj for j = 1, 2, . . . , l. Therefore, we can define an antirepresentation of AΘ on H
by
X X X
R(f ) := C L(f )∗ C −1 = σ(r, P )∗ Cfr C −1 = σ(−r, P ) fr = fr σ(−r, P ).
r r r

Notice that σ(r, P )∗ = σ(−r, P ) commutes with fr in view of Exercise 7.11 and the relation
σ(−r, r) = 1.
The left and right multiplication operators commute, since
X
[L(f ), R(h)] := [σ(r, P ) fr , hs σ(−s, P )]
r,s
X
= [fr , hs ] σ(r, s) σ(r − s, P ) = 0,
r,s

where [fr , hs ] = 0 because, with its original product. A is commutative. The same calculation
shows also that
  X
[[D,/ L(f )], R(h)] = L([D,/ f ]), R(h) = / fr ], hs ] σ(r, s) σ(r − s, P ) = 0,
[[D,
r,s

/ f ])r = [D,
since ([D, / fr ] for each r and [[D,
/ fr ], hs ] = 0 by the first-order property of the
∞ /
undeformed spectral triple (C (M ), H, D).

When dim M is even, and Γ is the Z2 -grading operator Γ on the spinor space H, we should
note that the orientation condition πD/ (c) = Γ says, among other things, that Γ appears in
the algebra generated by the operators f and [D, / f ], for f ∈ A. The representation L of AΘ
/ f ]). In the formula
extends to this algebra of operators by using (7.9) as a definition of L([D,
r r
(7.1) for πD/ (c), if we replace all terms aj by L(aj ), then we obtain L(πD/ (c)) = L(Γ) = Γ.
Thus c may also be regarded as a Hochschild n-cycle over AΘ , and the orientation condition
πD/ (c) = Γ is unchanged by the deformation. In odd dimensions, the same is true, with Γ
replaced by 1.
In conclusion: the isospectral deformation procedure of Connes and Landi yields a family
of noncommutative spectral triples that satisfy all of our stated conditions for a noncommuta-
tive spin geometry. (Moreover [CL], Poincaré duality in K-theory is stable under deformation,
too.)

86
7.3 The Moyal plane as a nonunital spectral triple
In order to extend the notion of spectral triple (A, H, D) to include the case where the algebra
A may be nonunital, we modify Definition 4.1 as follows.
Definition 7.13. A nonunital spectral triple (A, H, D) consists of a nonunital ∗-algebra
A, equipped with a faithful representation on a Hilbert space H, and a selfadjoint operator D
on H with a(Dom D) ⊆ Dom D for all a ∈ A, such that
• [D, a] extends to a bounded operator on H, for each a ∈ A;
• a (D2 + 1)−1/2 is a compact operator, for each a ∈ A.
In general, D may have continuous spectrum, so that the operator (D2 + 1)−1/2 will
usually not be compact. But it is enough to ask that it become compact when mollified by
any multiplication operator in A. An equivalent condition is that a(D − λ)−1 be compact, for
all λ ∈
/ sp(D). In the nonunital case, there is no advantage in supposing that D be invertible,
so it is better to work directly with (D2 + 1)1/2 instead of |D|.
Remark 7.14. The simplest commutative example of a nonunital spectral triple is given by
m ∂
A = C0∞ (Rn ), H = L2 (Rn ) ⊗ C2 , / = −i γ j
D ,
∂xj
m
describing the noncompact manifold Rn with trivial spinor bundle Rn × C2 → Rn and flat
metric: as always, n = 2m or n = 2m+1. Here C0∞ (Rn ) is the space of smooth functions that
vanish at infinity together with all derivatives: it is a ∗-algebra under pointwise multiplication
and complex conjugation of functions. Here sp(D) / = R and (D / 2 + 1)−1/2 is not compact.
However, it is known [Sim] that if f ∈ Lp (Rn ) with p > n, then f (D / 2 + 1)−1/2 ∈ Lp (H).
The simplest noncommutative, nonunital example is an isospectral deformation of this
commutative case, where we use the same Dirac operator D / = −i γ j ∂/∂xj on the same spinor
m
space H = L2 (Rn ) ⊗ C2 , but we change the algebra by replacing the ordinary product of
functions by a Moyal product.
Before giving the details, we summarize the effect of this nonunital isospectral deformation
on the conditions given in Section 7.1 to define a “noncommutative spin geometry”.
• The reality and first-order conditions are unchanged: we use the same charge conjuga-
tion operator C as in the undeformed case.
• The regularity condition is essentially unchanged: all that is needed is to replace the
/ 2 +1)1/2 , T ], because Dom δ1k =
derivation δ : T 7→ [|D|, T ] by the derivation δ1 : T 7→ [(D
Dom δ k for each k ∈ N since (D / 2 + 1)1/2 − |D| is a bounded operator.

• For the orientation condition, the Hochschild n-cycle will not lie in Zp (A, A ⊗ A◦ ) but
e Ae ⊗ Ae◦ ), where Ae is a unitization of A, that is, a unital ∗-algebra in
rather in Zp (A,
which A is included as an essential ideal.
• For the finiteness condition, we ask that H∞ = AN p, for some projector p = p2 = p∗
e Thus H∞ can be regarded as the pullback, via the inclusion A ֒→ A,
lying in MN (A). e
e
of the finitely generated projective left A-module AeN p.
• To define the integer n as the spectral dimension, we would like to be able to assert that
a (D2 + 1)−1/2 lies in Ln+ (H) for each a ∈ A, and that 0 < Trω (a (D2 + 1)−n/2 ) < ∞
whenever a is positive and nonzero. It turns out that we can only verify this for a
belonging to a certain dense subalgebra of A, in the Moyal-plane example: see below.

87
Exercise 7.15. Check the assertion on regularity: show that Dom δ1 = Dom δ and that
Dom δ1k = Dom δ k for each k ∈ N, by induction on k.
Exercise 7.16. Show that Proposition 6.13 holds without the assumption that D is invert-
ible. Namely, if L1 (b)T:= (D2 + 1)−1/2 [D2 , b] and R1 (b) := [D2 , b] (D2 + 1)−1/2 , show that
T k l m
k,l≥0 Dom(L1 R1 ) = m≥0 Dom δ1 by adapting the proof of Proposition 6.13.

In what follows, we will sketch the main features of the Moyal-plane spectral triple. A
complete treatment can be found in Gayral et al [GGISV], on which this outline is based.
Our main concern here is to identify the “correct” algebra A and its unitization Ae so that
the modified spin-geometry conditions will hold.
We now recall the Moyal product over Rn , discussed in the previous Section. It depends
on a real skewsymmetricmatrix  Θ ∈ Mn (R) of “deformation parameters”. For n = 2, such
0 θ
a matrix is of the form for some θ ∈ R; and for n = 2m or n = 2m + 1, Θ is
−θ 0
similar to a direct sum of m such matrices with possibly different values of θ (so Θ cannot
be invertible if n is odd). For convenience, we now take n to be even, and we shall suppose
that all values of θ are the same. (In applications to quantum mechanics, where the Moyal
product originated [Moy], θ = ~ is the Planck constant.) Thus, we choose
 
0 1m
Θ := θ S ∈ M2m (R), with S := , θ > 0. (7.10)
−1m 0

Note that det Θ = θ2m > 0.

Definition 7.17. Let n = 2m be even, let θ > 0, and let f, h ∈ S(Rn ). Their Moyal
product f ⋆θ h ∈ S(Rn ) is defined as follows:
Z Z
−n
(f ⋆θ h)(x) := (πθ) f (x + s) h(x + t) e2is(St)/θ ds dt
n n
ZR ZR
= (2π)−n f (x − 21 θSu) h(x + t) e−iut du dt
R n Rn
Z
−n
= (2π) f (x − 21 θSu) ĥ(u) eiux du. (7.11)
Rn
R
Here ĥ(u) := Rn h(t) e−iut dt is the Fourier transform. Since h 7→ ĥ preserves the Schwartz
space S(Rn ), the third integral is a twisted version of the usual convolution of f and ĥ, and
one can check that this integral converges to an element of S(Rn ).
The first or second integral in (7.11) can also be regarded as defining the Moyal product
f ⋆θ h, where f and h need not be Schwartz functions, provided that the integrals are under-
stood in some generalized sense. Thus Rieffel [Rie], for instance, considers them as oscillatory
integrals. Here we shall extend the Moyal product by duality, as follows. It is easy to see that
kf ⋆θ hk∞ ≤ (πθ)−n kf k1 khk1 , from the first integral in (7.11). By applying similar estimates
to the functions xα ∂ β (f ⋆θ h), for α, β ∈ Nn , one can verify that the product (f, h) 7→ f ⋆θ h
is a jointly continuous bilinear map from S(Rn ) × S(Rn ) to S(Rn ).
Here are some elementary properties of the Moyal product that are easy to check formally;
they can be verified rigorously by some work with oscillatory integrals: see [Rie].
1. The Moyal product is associative: (f ⋆θ g) ⋆θ h = f ⋆θ (g ⋆θ h).

2. The Leibniz rule holds: ∂j (f ⋆θ h) = ∂j f ⋆θ h + f ⋆θ ∂j h for j = 1, . . . , n.

88
 3. Complex conjugation is an involution: f ⋆θ h = h̄ ⋆θ f¯.

4. Integration over Rn is a trace for the Moyal product:

Z Z Z
(f ⋆θ h)(x) dx = (h ⋆θ f )(x) dx = f (x) h(x) dx. (7.12)
Rn Rn Rn

We denote Sθ := (S(Rn ), ⋆θ ). It is a Fréchet ∗-algebra, with the usual topology of S(Rn ).

The trace property gives us a (suitably normalized) bilinear pairing:
Z
−m
hf, hi := (πθ) (f ⋆θ h)(x) dx.
Rn

Together with associativity, this gives the relation

Z
hf ⋆θ g, hi = hf, g ⋆θ hi = (πθ)−m (f ⋆θ g ⋆θ h)(x) dx,
Rn

valid for f, g, h ∈ S(Rn ). Now, if T ∈ S ′ (Rn ) is a tempered distribution, and if f ∈ S(Rn ),

we can define T ⋆θ f, f ⋆θ T ∈ S ′ (Rn ) by the continuity of the Moyal product:

hT ⋆θ f, hi := hT, f ⋆θ hi, hf ⋆θ T, hi := hT, h ⋆θ f i.

In this way, S ′ (Rn ) becomes a bimodule over Sθ . Inside this bimodule, we can identify a
multiplier algebra in the obvious way.

Definition 7.18. The Moyal algebra Mθ = Mθ (Rn ) is defined as the set of (left and right)
multipliers for S(Rn ) within S ′ (Rn ):

Mθ := { R ∈ S ′ (Rn ) : R ⋆θ f ∈ S(Rn ), f ⋆θ R ∈ S(Rn ) for all f ∈ S(Rn ) }.

This is a ∗-algebra, and S ′ (Rn ) is an Mθ -bimodule, under the operations

hT ⋆θ R, f i := hT, R ⋆θ f i, hR ⋆θ T, f i := hT, f ⋆θ Ri.

This Moyal algebra is very large: for instance, it contains all polynomials on Rn . How-
ever, because it contains many unbounded elements, it cannot serve as a coordinate alge-
bra for a spectral triple. Even so, it is a starting point for a second approach, developed
in [GV]. Consider the quadratic polynomials Hr := 21 (x2r + x2m+r ) for r = 1, . . . , m. In the
quantum-mechanical interpretation, these are Hamiltonians for a set of m independent har-
monic oscillators; but for now, it is enough to know that they belong to Mθ . It turns out
that the left and right Moyal multiplications by these Hr have a set of joint eigenfunctions
{ fkl : k, l ∈ Nm } belonging to the Schwartz space S(Rn ), with the following properties:

• The eigenvalues are half-integer multiples of θ, namely,

Hr ⋆θ fkl = θ(kr + 12 )fkl , fkl ⋆θ Hr = θ(lr + 21 )fkl .

• The eigenfunctions form a set of matrix units for the Moyal product: fkl ⋆θ frs = δlr fks
and f¯kl = flk for all k, l, r, s ∈ Nm .
P
• Any f ∈ S(Rn ) is given by a series f = (2πθ)−m/2 kl αkl fkl , converging in the topo-
logy of S(Rn ), such that αkl → 0 rapidly.

89
 • The subset { (2πθ)−m/2 fkl : k, l ∈ Nm } of S(Rn ) is an orthonormal basis for L2 (Rn ).

For example, when n = 2 and k = l ∈ N, fkk is given by

2 2
fkk (x1 , x2 ) := 2 (−1)k e−(x1 +x2 )/θ L0k ( 2θ (x21 + x22 )),

where L0k is the Laguerre polynomial of order k.

Because of these properties,
P we can extend the Moyal Pproduct to pairs of functions in
2 n
L (R ). If f = (2πθ) −m/2 −m/2
k,l αkl fkl and h = (2πθ) k,l βkl fkl , we define
X
f ⋆θ h := (2πθ)−m αkr βrl fkl . (7.13)
k,r,l

The Schwarz inequality for sequences shows that

X 2 XX 2
kf ⋆θ hk22 = (2πθ)−2m αkr βrl fkl = (2πθ)−m αkr βrl
k,r,l 2 k,l r
X X
≤ (2πθ)−m |αkr |2 |βrl |2 = (2πθ)−m kf k22 khk22 . (7.14)
k,r r,l

This calculation guarantees that the series (7.13) converges whenever f, h ∈ L2 (Rn ); and
that the operator L(f ) : h 7→ f ⋆θ h extends to a bounded operator in L(L2 (Rn )) whenever
f ∈ L2 (Rn ). Moreover, it gives a bound on the operator norm:

kL(f )k ≤ (2πθ)−m/2 kf k2 .

Now the Schwartz-multiplier algebra Mθ can be replaced by an L2 -multiplier algebra. By

duality in sequence spaces, any T ∈ S ′ (Rn ) can be given an expansion in terms of the {fkl }
basis, and in this way one can define an algebra

Aθ := { R ∈ S ′ (Rn ) : R ⋆θ f ∈ L2 (Rn ) for all f ∈ L2 (Rn ) }.

This is actually a C ∗ -algebra, with operator norm kL(R)k := sup{ kR ⋆θ f k2 /kf k2 : f 6= 0 }.

There is a unitary isomorphism W : L2 (Rn ) → L2 (Rm ) ⊗ L2 (Rm ) (tensor product of
Hilbert spaces), such that W L(f ) W −1 = σ(f ) ⊗ 1, where σ is the (irreducible) Schrödinger
representation; that is to say, f 7→ L(f ) is equivalent to the Schrödinger representation with
infinite multiplicity. One can show that Aθ = W −1 L(L2 (Rm )) W , whereas the norm closure
of the ∗-algebra (S(Rn ), ⋆θ ) is W −1 K(L2 (Rm )) W . For the details, consult [VG] and [GGISV].
The analogue of Lemma 6.19 holds, too: Sθ is a nonunital pre-C ∗ -algebra. As in the proof
of Lemma 6.19, if f ∈ Sθ , suppose the equation (1 + f ) ⋆θ (1 + g) = 1 has a solution g in the
unital C ∗ -algebra Aθ . We may also write

f + g + f ⋆θ g = 0 and f + g + g ⋆θ f = 0, (7.15)

and we wish to show that g ∈ Sθ . Since g = −f − f ⋆θ g, it is enough to show that f ⋆θ g ∈ Sθ .

Now, left-multiplying the second equation in (7.15) by f gives f ⋆θ f + f ⋆θ g + f ⋆θ g ⋆θ f = 0,
so it is enough to check that f ⋆θ g ⋆θ f ∈ S(Rn ) whenever f ∈ S(Rn ) and g ∈ Aθ . This turns
out to be true: the necessary norm estimates are given in [VG].
However, the algebra Sθ is not the best candidate for the coordinate algebra of the Moyal
spectral triple. We now introduce a better algebra.

90
Definition 7.19. Consider the following space of smooth functions on Rn :

DL2 (Rn ) := { f ∈ C ∞ (Rn ) : ∂ α f ∈ L2 (Rn ) for all α ∈ Nn },

introduced by Laurent Schwartz

P in his book on distributions [Schz]. It is a Fréchet space,
under the norms pr (f ) := |α|≤r k∂ α f k2 , for r ∈ N. The Leibniz rule for the Moyal product
and the inequality (7.14) show that if f, h ∈ DL2 (Rn ), then
X
k∂ α (f ⋆θ h)k2 ≤ k∂ β f ⋆θ ∂ α−β h)k2
0≤β≤α
X α 
−m/2
≤ (2πθ) k∂ β f k2 k∂ α−β hk22 ,
β
0≤β≤α

so that DL2 (Rn ) is actually an algebra under the Moyal product; and that this product is
continuous for the given Fréchet topology. Moreover, since complex conjugation is an isometry
for each norm pr , it is a ∗-algebra with a continuous involution. We write Aθ := (DL2 (Rn ), ⋆θ )
to denote this Fréchet ∗-algebra.

It does not matter whether these derivatives ∂ α f are taken to be distributional derivatives
only, since arguments based on Sobolev’s Lemma show that if f and all its distributional
derivatives are square-integrable, then f is actually a smooth function.
The algebra Aθ is nonunital. Next, we introduce the preferred unitization of Aθ .

Definition 7.20. Another space of smooth functions on Rn is found also in [Schz]:

B(Rn ) := { f ∈ C ∞ (Rn ) : ∂ α f is bounded on Rn , for all α ∈ Nn }.

It is also a Fréchet space, under the norms qr (f ) := max|α|≤r k∂ α f k∞ , for r ∈ N.

We shall soon prove that B(Rn ) is also a ∗-algebra under the Moyal product; we denote
it by Aeθ := (DL2 (Rn ), ⋆θ ).

It is proved in Schwartz’ book that DL2 (Rn ) ⊂ B(Rn ), and that the inclusion is continuous
for the given topologies. (This is not as obvious as it seems, because in general square-
integrable functions on Rn need not be bounded.) Combining this with knowledge of the
Moyal multiplier algebras, we end up with the following inclusions [GGISV]:

Sθ ⊂ Aθ ⊂ Aeθ ⊂ Aθ ∩ Mθ .

The inclusion Aeθ ⊂ Aθ is a consequence of the Calderón–Vaillancourt theorem, which says

that a pseudodifferential operator of order zero on Rn , whose symbol is differentiable to a high
enough order, gives a bounded operator on L2 (Rn ); we may notice that the second integral
in (7.11) says that L(f ) is pseudodifferential, with symbol p(x, ξ) = f (x − 21 θSξ).

Proposition 7.21. B(Rn ) is a Fréchet ∗-algebra under the Moyal product.

Proof. If f, h ∈ B(Rn ) and if s ∈ N, we shall find estimates of the form

qs (f ⋆θ h) ≤ Crs qr (f ) qr (h) whenever r ≥ s + n + 2. (7.16)

This shows that f ⋆θ h lies in B(Rn ) whenever f, h ∈ B(Rn ), and that (f, h) 7→ f ⋆θ h is a
jointly continuous bilinear operation on B(Rn ). Since complex conjugation is clearly isometric
for each qr , the involution is continuous, too.

91
 To justify the estimates (7.16), we first notice that, for any k ∈ N,
ZZ
β γ −n ∂ β f (x + y) ∂ γ h(x + z)
(∂ f ⋆θ ∂ h)(x) = (πθ) (1 + |y|2 )k (1 + |z|2 )k e2iy(Sz)/θ dy dz
(1 + |y|2 )k (1 + |z|2 )k
ZZ β
∂ f (x + y) ∂ γ h(x + z)  2iy(Sz)/θ 
= (πθ)−n P k (∂y , ∂ z ) e dy dz
(1 + |y|2 )k (1 + |z|2 )k
ZZ  β 
−n 2iy(Sz)/θ ∂ f (x + y) ∂ γ h(x + z)
= (πθ) e Pk (−∂y , −∂z ) dy dz,
(1 + |y|2 )k (1 + |z|2 )k

where Pk is a certain polynomial of degree 2k in both yj and zj variables, and for the third
line we integrate by parts. It is not hard to find constants such that |∂ α ((1 + |x|2 )−k )| ≤
′ (1 + |x|2 )−k for each k ∈ N, α ∈ Nn . Thus, we get estimates of the form
Cαk

X ZZ
′′ |∂ β+µ f (x + y)| |∂ γ+ν h(x + z)|
|(∂ β f ⋆θ ∂ γ h)(x)| ≤ Cµν dy dz
(1 + |y|2 )k (1 + |z|2 )k
|µ|,|ν|≤2k
Z Z
′′′ dy dz
≤ Ckr qr (f ) qr (h) 2 k 2 k
,
Rn (1 + |y| ) Rn (1 + |z| )

provided r ≥ |β| + |γ| + 2k; and we need k > n/2 so that the right hand side is finite. For
|β| + |γ| ≤ s, we only need to choose k so that n < 2k ≤ r − s, and this is always possible for
r ≥ s + n + 2.

Rieffel, in [Rie], showed that Aeθ is the space of smooth vectors for the action of Rn (by
translations) on its C ∗ -completion; this entails that Aeθ is a pre-C ∗ -algebra.
Now, the inclusion Aeθ ⊂ Aθ means that k∂ α f ⋆θ ∂ β hk2 is finite, whenever f ∈ Aeθ and
h ∈ Aθ ; therefore, f ⋆θ h lies in Aθ also. A similar argument shows that h ⋆θ f lies in Aθ .
Thus, Aθ is an ideal in Aeθ . (In fact, it is an essential ideal; that is to say, if f ⋆θ h = 0 for all
h ∈ Aθ , then f = 0; this can be seen by taking h = fkl for any k, l ∈ Nn and checking that f
must vanish.)

Lemma 7.22. Aθ is a nonunital pre-C ∗ -algebra.

Proof. Since Aθ is Fréchet, we only need to show that it is spectrally invariant. In the
nonunital case, this means that if f ∈ Aθ , and the equations f + g + f ⋆θ g = f + g + g ⋆θ f = 0
have a solution g in the C ∗ -completion of Aθ , then g lies in Aθ . Now since f ∈ Aeθ and Aeθ is
already a pre-C ∗ -algebra, we see that g ∈ Aeθ . But Aθ is an ideal in Aeθ , and thus f ⋆θ g ∈ Aθ .
This implies that g = −f − f ⋆θ g lies in Aθ , too.

An important family of elements in Aeθ that do not belong to Aθ are the plane waves:

uk (x) := e2πikx , for each k ∈ Rn .

It is immediate from the formulas (7.11) that

uk ⋆θ ul = e−πiθk(Sl) uk+l , for all k, l ∈ Rn .

In particular, by taking k, l ∈ Zn to be integral vectors, we get an inclusion C ∞ (TnθS ) ֒→ Aeθ :

the smooth algebra of the noncommutative n-torus, for Θ = θS, can be identified with a
subalgebra of periodic functions in Aeθ .

92
 In particular, the Hochschild n-cycle c representing the orientation of this noncommuta-
tive torus can also be regarded as an n-cycle over Aeθ . We can write uk = v1k1 ⋆θ · · · ⋆θ vnkn
where vj = uej for the standard orthonormal basis {e1 , . . . , en } of Rn . The expression for c is

1 X
c= (−1)σ (vσ(1) vσ(2) . . . vσ(n) )−1 ⊗ vσ(1) ⊗ vσ(2) ⊗ · · · ⊗ vσ(n) .
n! (2πi)n
σ∈Sn

(When θ = 0, we can write vj = e2πitj , and the right hand side reduces to dt1 ∧ · · · ∧ dtn , the
usual volume form for either Rn or the flat torus Tn = Rn /Zn .)
We refer to [GGISV] for the discussion of the spectral dimension properties of the triple
m
(Aθ , L2 (Rn ) ⊗ C2 , D).
/ Briefly, the facts are these. If π(f ) := L(f ) ⊗ 12m denotes the
representation of Aθ on the spinor space H by componentwise left Moyal multiplication, then
one can show that, for any f ∈ Aθ , we get
2
/ + 1)−1/2 ∈ Lp (H),
π(f ) (D for all p > n.

In particular, these operators are compact, so this triple is indeed a nonunital spectral triple.
However, this is not quite enough to guarantee that
2
/ + 1)−1/2 ∈ Ln+ (H),
π(f ) (D (7.17)

for every f ∈ Aθ . Instead, what is found in [GGISV] is that (7.17) holds for f lying in the
(dense) subalgebra Sθ . The key lemma which makes the proof work is a “strong factorization”
property of Sθ , proved in [GV]: namely, that any f ∈ Sθ can be expressed (without taking
finite sums) as a product f = g ⋆θ h, with g, h ∈ Sθ . This factorization property fails for the
full algebra Aθ .
Once (7.17) has been established, one can proceed to compute its Dixmier trace. It
turns out that Tr / 2 + 1)−1/2 ) is unchanged from its value when θ = 0, namely
ω (π(f ) (D
R
(2m Ωn /n (2π)n ) Rn f (x) dx. The end result is that the spectral dimension condition for
nonunital spectral triples is the expected one, but that Dixmier-traceability as in (7.17)
should only be required for a dense subalgebra of the original algebra.

7.4 A geometric spectral triple over SUq (2)

(The notes for this part will appear later. In the meantime, have a look at [DLSSV].)

93
Appendix A

Exercises

A.1 Examples of Dirac operators

A.1.1 The circle
Let M := S1 , regarded as S1 ≃ R/Z; that is to say, we parametrize the circle by the half-open
interval [0, 1) rather than [0, 2π), say. Then A = C ∞ (S1 ) can be identified with periodic
smooth functions on R with period 1:

A ≃ { f ∈ C ∞ (R) : f (t + 1) ≡ f (t) }.

Since Cl(R) = C1 ⊕ Ce1 as a Z2 -graded algebra, we see that B = A in this case; and since
n = 1, m = 0 and 2m = 1, there is a “trivial” spin structure given by S := A itself. The
charge conjugation is just C = K, where K means complex conjugation of functions. With
the flat metric on the circle, the Dirac operator is just
d
/ := −i
D .
dt
Exercise A.1. Show that its spectrum is

/ = 2πZ = { 2πk : k ∈ Z },
sp(D)

by first checking that the eigenfunctions ψk (t) := e2πikt form an orthonormal basis for the
Hilbert-space completion H of S – using Fourier series theory.
/ contains
The point is that the closed span of these eigenvectors is all of H, so that sp(D)
no more than the corresponding eigenvalues.
Next, consider
S ′ := { φ ∈ C ∞ (R) : φ(t + 1) ≡ −φ(t) },
which can be thought of as the space of smooth functions on the interval [0, 1] “with antiperi-
odic boundary conditions”.
Exercise A.2. Explain in detail how S ′ can be regarded as a B-A-bimodule, and how C = K
acts on it as a charge-conjugation operator. Taking D / := −i d/dt again, but now as an
operator with domain S on the Hilbert-space completion of S ′ , show that its spectrum is now
′

/ = 2π(Z + 21 ) = { π(2k + 1) : k ∈ Z },
sp(D)

by checking that φk (t) := eπi(2k+1)t are a complete set of eigenfunctions.

94
 The circle S1 thus carries two inequivalent spin structures: their inequivalence is most
clearly manifest in the different spectra of the Dirac operators. Notice that 0 ∈ sp(D) /
for the “untwisted” spin structure where S = A, while 0 ∈ / for the “twisted” spin
/ sp(D)
structure whose spinor module is S ′ . There are no more spin structures to be found, since
H 1 (S1 , Z2 ) = Z2 .

A.1.2 The (flat) torus

On the 2-torus T2 := R2 /Z2 , we use the Riemannian metric coming from the usual flat
metric on R2 . Thus, if we regard A = C ∞ (T2 ) as the smooth periodic functions on R2 with
f (t1 , t2 ) ≡ f (t1 +1, t2 ) ≡ f (t1 , t2 +1), then (t1 , t2 ) define local coordinates on T2 , with respect
to which all Christoffel symbols are zero, namely Γkij = 0, and thus ∇ = d represents the
Levi-Civita connection on 1-forms.
In this case, n = 2, m = 1 and 2m = 2, so we use “two-component” spinors; that is, the
spinor bundles S → T2 are of rank two. There is the “untwisted” one, where S is the trivial
rank-two C-vector bundle, and S ≃ A2 . The Clifford algebra in this case is just B = M2 (A).
Using the standard Pauli matrices:
     
1 0 1 2 0 −i 3 1 0
σ := , σ := , σ := ,
1 0 i 0 0 −1
we can write the charge conjugation operator as
C = −i σ 2 K
where K again denotes (componentwise) complex conjugation.
Exercise A.3. Find three more spinor structures on T2 , exhibiting each spinor module as a
B-A-bimodule, with the appropriate action of C. (Use T2 = S1 × S1 .)
Exercise A.4. Check that
 
1 2 0 −∂2 − i∂1
/ = −i (σ ∂1 + σ ∂2 ) =
D
∂2 − i∂1 0
where ∂1 = ∂/∂t1 and ∂2 = ∂/∂t2 , is indeed the Dirac operator on the untwisted spinor
/ 2 ) by finding a complete set of eigenvectors. Then show that
module S = A2 . Compute sp(D
q
/ = { ±2π r12 + r22 : (r1 , r2 ) ∈ Z }
sp(D)
by finding the eigenspinors for each of these eigenvalues. ¿What can be said of the multiplic-
/
ities of these eigenvalues? and what is the dimension of ker D?
Notice that σ 3 does not appear in the formula for D;
/ its role here is to give the Z2 -grading
3
operator: c(γ) = σ —regarded as a constant function with values in M2 (C)— in view of the
relation σ 3 = −i σ 1 σ 2 among Pauli matrices.
On the 3-torus T3 := R3 /Z3 , where now n = 3, m = 1 and again 2m = 2, we get two-
component spinors. Again we may use a flat metric and an untwisted spin structure with
S = A2 . The charge conjugation is still C = −i σ 2 K on S, so that C 2 = −1 also in this
3-dimensional case. The Dirac operator is now
 
1 2 3 −i ∂3 −∂2 − i∂1
/
D = −i (σ ∂1 + σ ∂2 + σ ∂3 ) = .
∂2 − i∂1 i ∂3
2
/ for this Dirac operator on T3 .
/ ) and sp(D)
Exercise A.5. Compute sp(D

95
A.1.3 The Hodge–Dirac operator on S2
If M is a compact, oriented Riemannian manifold that has no spinc structures, ¿can one
define Dirac-like operators on an B-A-bimodule E that is not pointwise irreducible under the
action of B? It turns out that one can do so, if E carries a “Clifford connection”, that is, a
connection ∇E such that
∇E (c(α)s) = c(∇α) s + c(α) ∇E s,
for α ∈ A1 (M ), s ∈ E, and which is Hermitian with respect to a suitable A-valued sesquilinear
pairing on E. For instance, we may take E = A• (M ), the full algebra of differential forms
on M , which we know to be a left B-module under the action generated by c(α) = ε(α)+ι(α♯ ).
The Clifford connection is just the Levi-Civita connection on all forms, obtaining by extending
the one on A1 (M ) with the Leibniz rule (and setting ∇f := df on functions). The pairing
(α | β) := g(ᾱ, β) extends to a pairing on A• (M ); by integrating the result over M with
respect to the volume form νg , we get a scalar product on forms, and we can then complete
A• (M ) to a Hilbert space.
If {E1 , . . . , En } and {θ1 , . . . , θn } are local orthonormal sections for X (M ) and A1 (M )
respectively, compatible with the given orientation, so that c(θj ) = ε(θj ) + ι(Ej ) locally, then
⋆ := c(γ) = (−i)m c(θ1 ) c(θ2 ) . . . c(θn )
is globally well-defined as an A-linear operator taking A• (M ) onto itself, such that ⋆2 = 1.
This is the Hodge star operator, and it exchanges forms of high and low degree.
Exercise A.6. If {1, . . . , n} = {i1 , . . . , ik } ⊎ {j1 , . . . , jn−k }, show that locally,
⋆(θi1 ∧ · · · ∧ θik ) = ±im θj1 ∧ · · · ∧ θjn−k ,
where the sign depends on i1 , . . . , ik . Conclude that ⋆ maps Ak (M ) onto An−k (M ), for each
k = 0, 1, . . . , n.
(Actually, our sign conventions differ from the usual ones in differential geometry books,
that do not include the factor (−i)m . With the standard conventions, ⋆2 = ±1 on each
Ak (M ), with a sign depending on the degree k.)
The codifferential δ on A• (M ) is defined by
δ := −⋆d⋆.
This operation lowers the form degree by 1. The Hodge–Dirac operator is defined to be
−i(d + δ) on A• (M ). One can show that, on the Hilbert-space completion, the operators d
and −δ are adjoint to one another, so that −i(d + δ) extends to a selfadjoint operator. (With
the more usual sign conventions, d and +δ are adjoint, so that the Hodge–Dirac operator is
written simply d + δ.)
Now we take M = S2 , the 2-sphere of radius 1. The round (i.e., rotation-invariant) metric
on S2 is written g = dθ2 + sin2 θ dφ2 in the usual spherical coordinates, which means that
{dθ, sin θ dφ} is a local orthonormal basis of 1-forms on S2 . The area form is ν = sin θ dθ ∧ dφ.
The Hodge star is specified by defining it on 1 and on dθ:
⋆(1) := −i ν, ⋆(dθ) := i sin θ dφ.
To find the eigenforms of the Hodge–Dirac operator, it is convenient to use another set
of coordinates, obtained form the Cartesian relation (x1 )2 + (x2 )2 + (x3 )2 = 1 by setting
ζ := x1 + ix2 = eiφ cos θ, along with x3 = cos θ; the pair (ζ, x3 ) can serve as coordinates for
S2 , subject to the relation ζ ζ̄ + (x3 )2 = 1. (The extra variable ζ̄ gives a third coordinate,
extending S2 to R3 .)

96
Exercise A.7. Check that in the (ζ, x3 ) coordinates, the Hodge star is given by

⋆(ζ) = −i dζ ∧ dx3 , ⋆(dζ) = x3 dζ − ζ dx3 .

Exercise A.8. Consider the (complex) vectorfields on R3 given by

∂ ∂ ∂ ∂ ∂ ∂
L+ := 2ix3 + iζ 3
, L− := 2ix3 − iζ̄ , L3 := iζ̄ − iζ .
∂ ζ̄ ∂x ∂ζ ∂x3 ∂ ζ̄ ∂ζ

Verify the commutation relations [L+ , L− ] = −2i L3 , [L3 , L− ] = i L− and [L3 , L+ ] = −i L+ .

These commutation relations show that if L± =: L1 ± i L2 , then L1 , L2 , L3 generate a

representation of the Lie algebra of the rotation group SO(3). One obtains representation
spaces of SO(3) by finding functions f0 (“highest weight vectors”) such that L3 f0 is a multiple
of f0 , L+ f0 = 0, and { (L− )r f0 : r ∈ N } spans a space of finite dimension. To get spaces of
differential forms with these properties, one extends each vector field Lj to an operator on
A• (S2 ), namely its Lie derivative Lj , just by requiring that Lj d = d Lj . Since the Hodge
star operator is unchanged by applying a rotation to an orthonormal basis of 1-forms, one
can also show that Lj ⋆ = ⋆Lj , so that the Hodge–Dirac operator −i(d + δ) commutes with
each Lj . This gives a method of finding subspaces of joint eigenforms for each eigenvalue of
the Hodge–Dirac operator.
We introduce the following families of forms:

φ+ l
l := iζ (1 − iν), l = 0, 1, 2, 3, . . . ; ψl+ := ζ l−1 (dζ + ⋆(dζ)), l = 1, 2, 3, . . . ;
φ−
l
l
:= iζ (1 + iν), l = 0, 1, 2, 3, . . . ; ψl− := ζ l−1
(dζ − ⋆(dζ)), l = 1, 2, 3, . . . .

Clearly, ⋆(φ± ± ± ± + + − −
l ) = ±φl and ⋆(ψl ) = ±ψl . Thus φl and ψl are even, while φl and ψl are
• 2 + 2 − 2
odd, with respecting to the Z2 -grading on forms given by A (S ) = A (S ) ⊕ A (S ), where
A± (S2 ) := 12 (1 ± ⋆) A• (S2 ).

Exercise A.9. Show that

−i(d + δ)φ± ∓
l = lψ , for l = 0, 1, 2, . . .
−i(d + δ)ψl± = (l + 1)φ∓ , for l = 1, 2, 3, . . .
− −
and conclude that each of φ+ + 2
l , φl , ψl and ψl is an eigenvector for (−i(d + δ)) = −(dδ + δd)
with
p eigenvalue l(l + 1). Find corresponding eigenspinors for −i(d + δ) with eigenvalues
± l(l + 1).

Exercise A.10. Show that L3 (ζ l ) = −il ζ l , L+ ζ l = 0, and that (L− )k (ζ l ) is a linear

combination of terms (x3 )k−2r ζ̄ r ζ l−k+r that does not vanish for k = 0, 1, . . . , 2l, and that
(L− )2l+1 (ζ l ) = 0. Check that L+ (L− )k (ζ l ) is a multiple of (L− )k−1 (ζ l ), for k = 1, . . . , 2l.

Exercise A.11. Show that

L3 φ± ±
l = −il φl , L+ φ±
l = 0; L3 ψl± = −il ψl± , L+ ψl± = 0;

for each possible value of l. Conclude that the forms Lk− (φ± k ±
l ) and L− (ψl ) vanish if and only
if k ≥ 2l + 1. ¿What can now be said about the multiplicities of the eigenvalues of −i(d + δ)?

With some more works, it can be shown that all these eigenforms span a dense subspace
of the Hilbert-space completion of A• (S2 ), so that these eigenvalues in fact give the full
spectrum of the Hodge–Dirac operator.

97
A.2 The Dirac operator on the sphere S2
A.2.1 The spinor bundle S on S2
Consider the 2-dimensional sphere S2 , with its usual orientation, S2 = C ∪ {∞} ≃ CP 1 . The
usual spherical coordinates on S2 are

p = (sin θ cos φ, sin θ sin φ, cos θ) ∈ S2 .

The poles are N = (0, 0, 1) and S = (0, 0, −1). Let UN = S2 \ {N }, US = S2 \ {S} be the two
charts on S2 . Consider the stereographic projections p 7→ z : UN → C, p 7→ ζ : US → C given
by
θ θ
z := e−iφ cot , ζ := e+iφ tan ,
2 2
so that ζ = 1/z on UN ∩ US . Write
2 q
q := 1 + z z̄ = , and q ′ := 1 + ζ ζ̄ = .
1 − cos θ z z̄
The sphere S2 has only the “trivial” spin structure S = Γ(S2 , S), where S → S2 has
rank two. Now S = S + ⊕ S − , where S ± → S2 are complex line bundles, and these may be
(and are) nontrivial. We argue that S + → S2 is the “tautological” line bundle coming from
S2 ≃ CP 1 . We know already that

S ♯ ≃ S ⇐⇒ S ∗ ≃ S ⇐= (S + )∗ ≃ S −

and the converse S ∗ ≃ S =⇒ (S + )∗ ≃ S − will hold provided we can show that S ± → S2 are
nontrivial line bundles. (Otherwise, S + and S − would each be selfdual, but we know that
the only selfdual line bundle on S2 is the trivial one, since H 2 (S2 , Z) ≃ Z.)
Consider now the (tautological) line bundle L → S2 , where
z1
Lz := { (λz0 , λz1 ) ∈ C2 : λ ∈ bC }, if z = , L∞ := { (0, λ) ∈ C2 : λ ∈ C }.
z0
In other words, Lz is the complex line through the point (1, z), for z ∈ C. A particular
1 1
local section of L, defined over UN , is σN (z) := (q − 2 , zq − 2 ), which is normalized so that
(σN | σN ) = q −1 (1 + z̄z) = 1 on UN : this hermitian pairing on Γ(S2 , L) comes from the
standard scalar product on C2 —each Lz is a line in C2 .
1 1
Let also σS (ζ) := (ζq ′− 2 , q ′− 2 ), normalized so that (σS | σS ) = 1 on US . Now if z 6= 0,
then    
−1 1 1 1/2 1 z
σS (z ) = √ , √ ′ = (z̄/z) √ ,√ = (z̄/z)1/2 σN (z).
z q′ q q q
To avoid ambiguity, we state that (z̄/z)1/2 means e−iφ , and also (z/z̄)1/2 will mean e+iφ .
+
A smooth section of L is given by two functions ψN (z, z̄) and ψS+ (ζ, ζ̄) satisfying the
+
relation ψN (z, z̄)σN (z) = ψS+ (ζ, ζ̄)σS (ζ) on UN ∩ US . Thus we argue that
+
ψN (z, z̄) = (z̄/z)1/2 ψS+ (z −1 , z̄ −1 ) for z 6= 0,
+
and ψN , ψS+ are regular at z = 0 or ζ = 0 respectively. Likewise, a pair of smooth functions
− −
ψN , ψS on C is a section of the dual line bundle L∗ → S2 if and only if
−
ψN (z, z̄) = (z/z̄)1/2 ψS− (z −1 , z̄ −1 ) for z 6= 0.

98
 We claim now that we can identify S + ≃ L and S − ≃ L∗ = L−1 —here the notation
L−1 means that [L−1 ] is the inverse of [L] in the Picard group H 2 (S2 , Z) that classifies C-line
bundles— so that a spinor in S = Γ(S2 , S) is given precisely by two pairs of smooth functions
! !
+
ψN (z, z̄) ψS+ (ζ, ζ̄)
− on UN , on US ,
ψN (z, z̄) ψS− (ζ, ζ̄)

satisfying the above transformation rules. (The nontrivial thing is that the spinor components
must both be regular at the south pole z = 0 and the north pole ζ = 0, respectively.)
Since S ⊗A S ∗ ≃ EndA (S) ≃ B ≃ A• (S2 ) as A-module isomorphisms (we know that
B ≃ A• (S2 ) as sections of vector bundles), it is enough to show that, as vector bundles,

A• (S2 ) ≃ L0 ⊕ L2 ⊕ L−2 ⊕ L0 ,

where L2 = L ⊗ L, L−2 = L∗ ⊗ L∗ , and L0 = S2 × C is the trivial line bundle. It is clear that

A0 (S2 ) = C ∞ (S2 ) = A = Γ(S2 , L0 ); and furthermore, A2 (S2 ) ≃ A = Γ(S2 , L0 ) since Λ2 T ∗ S2
has a nonvanishing global section, namely the volume form ν = sin θdθ ∧ dφ.
With respect to the “round” metric on S2 , namely,

4
g := dθ2 + sin2 θ dφ2 = (dx1 ⊗ dx1 + dx2 ⊗ dx2 ),
q2
   
dz dz̄ dζ dζ̄
the pairs of 1-forms , and − ′ , − ′ are local bases for A1 (S2 ), over UN and US
q q q q
respectively.

Exercise A.12. Write, for α ∈ A1 (S2 ),

dz dz̄
α =: fN (z, z̄) + gN (z, z̄) on UN ,
q q
dζ dζ̄
=: −fS (ζ, ζ̄) ′ − gS (ζ, ζ̄) ′ on US .
q q

Show that

fN (z, z̄) = (z̄/z) fS (z −1 , z̄ −1 )

gN (z, z̄) = (z/z̄) gS (z −1 , z̄ −1 )

on UN ∩ US , and conclude that A1 (S2 ) ≃ Γ(S2 , L2 ⊕ L−2 ).

Note that the last exercise now justifies the claim that the half-spin bundles were indeed
S + ⊕ S − ≃ L ⊕ L∗ .

A.2.2 The spin connection ∇S over S2

Given any local orthonormal basis of 1-forms {E1 , . . . , En }, we can compute Christoffel sym-
bols with all three indices taken from this basis, by setting Γ b βµα := (Eµ )i Γ
e β , or equivalently,
iα
by requiring that
∇Eµ Eα =: Γ b β Eβ
µα

for µ, α, β = 1, 2, . . . , n. (This works because the first index is tensorial).

99
Exercise A.13. On UN , take z =: x1 + ix2 . Compute the ordinary Christoffel symbols Γkij
in the (x1 , x2 ) coordinates for the round metric g = (4/q 2 )(dx1 ⊗ dx1 + dx2 ⊗ dx2 ), and then
show that
b β = δµα xβ − δµβ xα for µ, α, β = 1, 2.
Γ µα

This yields the local orthonormal bases E1 := 21 q ∂/∂x1 , E2 := 12 q ∂/∂x2 for vector fields,
and dually θ1 = (2/q) dx1 , θ2 = (2/q) dx2 for 1-forms. However, since S2 = CP 1 is a complex
manifold, it is convenient to pass to “isotropic” bases, as follows. We introduce

∂ dz
E+ := E1 − iE2 = q , θ+ := 12 (θ1 + iθ2 ) = ,
∂z q
∂ dz̄
E− := E1 + iE2 = q , θ− := 12 (θ1 − iθ2 ) = .
∂ z̄ q

Exercise A.14. Verify that the Levi-Civita connection on A1 (S2 ) is given, in these isotropic
local bases, by
   
dz dz dz dz
∇E+ = z̄ , ∇E− = −z ,
q q q q
   
dz̄ dz̄ dz̄ dz̄
∇E+ = −z̄ , ∇E− =z .
q q q q
 
0 1
The Clifford action on spinors is given (over UN , say) by γ1 := σ1 = and γ 2 :=
1 0
 
2 0 −i
σ = . The Z2 -grading operator is given by
i 0
 
1 2 3 1 0
χ := (−i) σ σ = σ = .
0 −1

The spin connection is now specified by

b β γ α γβ .
∇SE± := E± − 14 Γ ±α

Exercise A.15. Verify that, over UN , ∇S is determined by

∂ ∂
∇SE+ = q + 1 z̄ χ, ∇SE− = q − 1 z χ.
∂z 2 ∂ z̄ 2
/ = −iσ 1 ∇SE1 − iσ 2 ∇SE2 is given, over UN , by
Conclude that the Dirac operator D
!
∂
0 q ∂z − 21 z̄
/ = −i
D .
q ∂∂z̄ − 21 z 0

A similar expression is valid over US , by replacing z, z̄, q by ζ, ζ̄, q ′ respectively, and by

changing the overall (−i) factor to (+i). This formal change of sign is brought about by
the local coordinate transformation formulas induced by ζ = 1/z. (Here is an instance of
the “unique continuation property” of D:/ the local expression for the Dirac operator on any
one chart determines its expressions on any overlapping chart, and then by induction, on the
whole manifold.)

100
Exercise A.16. By integrating spinor pairings with the volume form ν = sin θdθ ∧ dφ =
2iq −2 dz ∧ dz̄, check that D
/ is indeed symmetric as an operator on L2 (S2 , S) with domain S.

Exercise A.17. Show that the spinor Laplacian ∆S is given in the isotropic basis by


∆s = − 21 ∇SE+ ∇SE− + ∇SE− ∇SE+ − z∇SE+ − z̄∇SE− ,

and compute directly that D / 2 = ∆S + 12 . This is consistent with the value s ≡ 2 of the scalar
curvature of S2 , taking into account how the metric g is normalized.

A.2.3 Spinor harmonics and the Dirac operator spectrum

Newman and Penrose (1966) introduced a family of special functions on S2 that yield an
orthonormal basis of spinors, in the same way that the conventional spherical harmonics
Ylm yield an orthonormal basis of L2 -functions. For functions, l and m are integers, but the
spinors are labelled by “half-odd-integers” in Z+ 12 . When expressed in our coordinates (z, z̄),
they are given as follows.
For l ∈ { 21 , 23 , 52 , . . . } = N + 12 , and m ∈ {−l, −l + 1, . . . , l − 1, l}, write
X l − 1 l + 1 
+ −l
Ylm (z, z̄) := Clm q 2 2
z r (−z̄)s ,
r s
r−s=m− 21
X l + 1 l − 1 
− −l
Ylm (z, z̄) := Clm q 2 2
z r (−z̄)s ,
1
r s
r−s=m+ 2

1
where r, s are integers with 0 ≤ r ≤ l ∓ and 0 ≤ s ≤ l ± 12 respectively; and
2
r s
l−m 2l + 1 (l + m)!(l − m)!
Clm = (−1) .
4π (l + 21 )!(l − 12 )!
±
Exercise A.18. Show that Ylm are half-spinors in S ± , by applying the transformation laws
−1
under z 7→ z and checking the regularity at the poles.
Then define pairs of full spinors by
! !
+ +
′ 1 Ylm ′′ 1 −Ylm
Ylm := √ − , Ylm := √ − .
2 iYlm 2 iYlm

These turn out to be eigenspinors for the Dirac operator.

Exercise A.19. Verify the following eigenvalue relations:
′ ′ ′′ ′′
/ lm
DY = (l + 12 ) Ylm , / lm
DY = −(l + 12 ) Ylm .

Goldberg et al (1967) showed that these half-spinors are special cases of matrix elements
l
Dnm of the irreducible group representations for SU (2), namely,
r
± 2l + 1 l
Ylm (z, z̄) = D∓ 1 ,m (−φ, θ, −φ),
4π 2

1
By setting h±lm (θ, φ, ψ)
R
±
:= e∓ 2 (φ+ψ) Ylm (z, z̄), we Rget an orthonormal set of elements of
L2 (SU(2)), such that SU(2) |h±lm (g)|2 dg = (1/4π) ± 2
S2 |Ylm | ν. The Plancherel formula for

101
 / Thus we
SU(2) can then be used to show that these are a complete set of eigenvalues for D.
have obtained the spectrum:

/ = { ±(l + 12 ) : l ∈ N + 21 } = {±1, ±2, ±3, . . . } = N \ {0},

sp(D)
±
with respectively multiplicities (2l + 1) in each case, since the index m in Ylm takes (2l + 1)
distinct values.

Postscript: / 2 = ∆S + 12 , we also get

Since s ≡ 2 and D

sp(∆S ) = { (l + 21 )2 − 1
2 = l2 + l − 1
4 : l ∈ N + 21 }

with multiplicities 2(2l + 1) in each case. Note that

2
/ ) = { (l + 12 )2 = l2 + l +
sp(D 1
4 : l ∈ N + 21 }.
1 1
The operator C given by C := ∆S + 4
/−
=D 4 has spectrum

sp(C) = { l(l + 1) : l ∈ N + 21 },

with multiplicities 2(2l + 1) again. This C comes from the Casimir element in the centre of
U(su(2)), represented on H = L2 (S2 , S) via the rotation action of SU(2) on the sphere S2 .
There is a general result for compact symmetric spaces M = G/K with a G-invariant spin
structure, namely that D / = C G + 18 s, or ∆S = C G − 81 s. This is a nice companion result,
albeit only for homogeneous spaces, to the Schrödinger–Lichnerowicz formula. Details are
given in Section 3.5 of Friedrich’s book.

A.3 Spinc Dirac operators on the 2-sphere

We know that finitely generated projective modules over the C ∗ -algebra A = C(S2 ) are of
the form p Ak , where p = [pij ] is an k × k matrix with elements in A, such that p (= p2 = p∗ )
is an orthogonal projector, whose rank is tr p = p11 + · · · + pkk . To get modules of sections
of line bundles, we impose the condition that tr p = 1, so that p Ak is an A-module “of rank
one”. It turns out that it is enough to consider the case k = 2 of 2 × 2 matrices.

Exercise A.20. Check that any projector p ∈ M2 (C(S2 )) is of the form

 
1 1 + n3 n1 − in2
p= ,
2 n1 + in2 1 − n3

where n21 + n22 + n23 = 1, so that ~n = (n1 , n2 , n3 ) is a continuous function from S2 to S2 .

n1 − in2
After stereographic projection, we can replace ~n by f (z) := . where z = e−iφ cot 2θ
1 − n3
is allowed to take the value z = ∞ at the north pole. Then f is a continuous map from the
Riemann sphere C ∪ {∞} = CP 1 into itself. If two projectors p and q are homotopic —there
is a continuous path of projectors { pt : 0 ≤ t ≤ 1 } with p0 = p and p1 = q— then they give
the same class [p] = [q] in K 0 (S2 ); and this happens if and only if the corresponding maps ~n,
or functions f (z), are homotopic.

102
Exercise A.21. Consider, for each m = 1, 2, 3, . . . , the maps

z 7→ fm (z) := z m and z 7→ f−m (z) := z̄ m

of the Riemann sphere into itself. ¿Can you describe the corresponding maps ~n of S2 into
itself ? Can you show that any two of these maps are not homotopic?
Let E(m) = pm A2 and E(−m) = p−m A2 , where
 m m m  m m m
1 z z̄ z 1 z z̄ z̄
pm (z) = m m m , p−m (z) = m m m ,
1 + z z̄ z̄ 1 1 + z z̄ z 1

with the obvious definition (¿what is it?) for z = ∞.

Exercise A.22. Show that E(1) is isomorphic to the space of sections of the tautological line
bundle L → CP 1 [hint: apply p1 to any element of A2 and examine the result]. Show also
that E(−1) gives the space of sections of the dual line bundle L∗ → CP 1 .
Exercise A.23. For m = 2, 3, . . . , show that E(m) ≃ E(1) ⊗A · · · ⊗A E(1) (m times) by
examining the components of elements of pm A2 . ¿What is the analogous result for E(−m) ?

For m ∈ Z, m 6= 0, we redefine E(m) := pm A2 with A = C ∞ (S2 ); so that E(m) now denotes

smooth sections over a nontrivial line bundle on S2 . We can identify each element of E(m) with
a smooth function fN : UN → C for which there is another smooth function fS : US → C,
such that
fN (z) = (z̄/z)m/2 fS (z −1 ) for all z 6= 0. (m)
Here, as before, (z̄/z) means eiφ in polar coordinates.
Exercise A.24. Writing E+ := q ∂/∂z and E− := q ∂/∂ z̄ as before, where q = 1 + z z̄, show
that when the operators

(m) ∂ (m) ∂
∇E+ = q + 1 mz̄, ∇E− = q − 1 mz,
∂z 2 ∂ z̄ 2
are applied to functions fN that satisfy (m), the image also satisfies (m). Thus they are
components of a connection ∇(m) on E(m) .
To get all the spinc structures on S2 , we twist the spinor module S for the spin structure,
namely S = E(1) ⊕ E(−1) , by the rank-one module E(m) . On the tensor product S ⊗A E(m) we
use the connection
∇S,m := ∇S ⊗ 1E(m) + 1S ⊗ ∇(m) .

Exercise A.25. Show that the Dirac operator D / m := −i ĉ ◦ ∇S,m , that acts on S ⊗A E(m) ,
is given by
! !
0 D /−m
0 ∂
q ∂z + 21 (m − 1) z̄
/m ≡
D = −i .
/+
D m 0 q ∂∂z̄ − 21 (m + 1) z 0

Check also that

∂ −(m+1)/2 ∂ (m−1)/2
/+
D m = −i q
(m+3)/2
q and /−
D m = −i q
−(m−3)/2
q ,
∂ z̄ ∂z
where these powers of q are multiplication operators on suitable spaces on functions on UN .

103
 +
/ m is of the form a(z) q (m+1)/2 where
Exercise A.26. If m < 0, show that any element of ker D
a(z) is a holomorphic polynomial of degree < |m|. Also, if m ≥ 0, show that ker D /+m = 0.

Exercise A.27. If m > 0, show that any element of ker D /−m is of the form b(z̄) q
−(m−1)/2

where b(z̄) is an antiholomorphic polynomial of degree < m. Also, if m ≤ 0, show that

/−
ker D / m equals −m in all cases.
m = 0. Conclude that the index of D

The sign of a selfadjoint operator D on a Hilbert space is given by the relation D =:

F |D| = F (D2 )1/2 , where we put F := 0 on ker D. Thus F is a bounded selfadjoint operator
such that 1 − F 2 is the orthogonal projector whose range is ker D. When ker D is finite-
dimensional, 1 − F 2 has finite rank, so it is a compact operator.
An even Fredholm module over an algebra A is given by:
1. a Z2 -graded Hilbert space H = H0 ⊕ H1 ;
 0 
π (a) 0
2. a representation a 7→ π(a) = of A on H by bounded operators that
0 π 1 (a)
commute with the Z2 -grading;
 
0 F−
3. a selfadjoint operator F = on H that anticommutes with the Z2 -grading,
F+ 0
such that F 2 − 1 and [F, π(a)] are compact operators on H, for each a ∈ A.
We can extend the twisted Dirac operator D / m to a selfadjoint operator on H = H0 ⊕ H1 ,
where H and H are two copies of the Hilbert space L2 (S2 , ν) where ν = 2i q −2 dz dz̄. We
0 1

define π 0 (a) = π 1 (a) to be the usual multiplication operator of a function a ∈ C ∞ (S2 ) on

this L2 -space.
/ m , given by the above formulas on its original domain, is a
Exercise A.28. Show that D
symmetric operator on H.
/ m determines a
Exercise A.29. Check that the sign Fm of the twisted Dirac operator D
∞ 2
Fredholm module over C (S ).

A.4 A spectral triple on the noncommutative torus

To define a spectral triple over a noncommutative algebra, we introduce the so-called non-
commutative torus. In fact, there are many such tori, labelled by a dimension n and by a
family of parameters θij forming a real skewsymmetric matrix Θ = −Θt ∈ Mn (R).
Fix an integer n ∈ {2, 3, 4, . . . }. In the algebra A0 := C(Tn ), one can write down Fourier-
series expansions:
X Z
f (φ1 , . . . , φn ) ←→ cr e2πir·φ
, cr := e−2πir·φ f (φ) dn φ ∈ C,
r∈Zn [0,1]n

where r · φ := r1 φ1 + · · · + rn φn , as usual. To ensure that this series converges uniformly and

represents f (φ), we retreat to the dense subalgebra A0 := C ∞ (Tn ), in which the coefficients
cr decrease rapidly to zero as |r| → ∞. On the space of multisequences c := {cr }r∈Zn , we
introduce the seminorms
X 1/2
k 2
pk (c) := (1 + r · r) |cr | , for all k ∈ N.
r∈Zn

104
We say that “cr → 0 rapidly” if pk (c) < ∞ for every k. Notice that pk+1 (c) ≥ pk (c) for
each k; these seminorms induce, on rapidly decreasing sequences, the topology of a Fréchet
space, which indeed coincides with the usual Fréchet topology on C ∞ (Tn ), i.e., the topology
of uniform convergence of the functions and of all their derivatives.
We can think of A0 as the C ∗ -algebra generated by n commuting unitary elements, namely
the functions uj defined by uj (φ1 , . . . , φn ) := e2πiφj , for j = 1, . . . , n.
Noncommutativity appears when we choose a real skewsymmetric matrix Θ ∈ Mn (R),
and introduce the (universal) C ∗ -algebra AΘ generated by unitary elements u1 , . . . , un which
no longer commute: instead, they satisfy the commutation relations

uk uj = e2πiθjk uj uk , for j, k = 1, . . . , n.

(In quantum mechanics, these are called “Weyl’s form of the canonical commutation rela-
tions”.) To form polynomials with these generators, we introduce a Weyl system of unitary
elements { ur : r ∈ Zn } in AΘ , by defining
 P
ur := exp πi j<k rj θjk rk ur11 ur22 . . . urnn .

Exercise A.30. Show that (ur )∗ = u−r for r ∈ Zn , and that

 P
ur us = σ(r, s) ur+s , where σ(r, s) := exp −πi j,k rj θjk sk .

Verify directly that

σ(r, s + t) σ(s, t) = σ(r, s) σ(r + s, t), for r, s, t ∈ Zn .

Notice that σ(r, ±r) = 1 by skewsymmetry of Θ.

We now define AΘ =: C ∞ (TnΘ ) to be the dense ∗-subalgebra of AΘ consisting of elements
of the form X
a= ar ur
r∈Zn

where ar ∈ C for each r, and ar → 0 rapidly.

Exercise A.31. Check that this series converges in the norm of AΘ , by considering the series
P −k for large enough k.
r (1 + r · r)

There is an action of the abelian Lie group Tn by ∗-automorphisms on the C ∗ -algebra

AΘ , given by
z · ur := z1r1 z2r2 . . . znrn ur for r ∈ Zn ,
or, more simply, z · uj = zj uj , where z = (z1 , . . . , zn ) ∈ Tn . This action is generated by a set
of n commuting derivations δ1 , . . . , δn , namely,

d
δj (a) := e2πitφj · a,
dt t=0

whose domain is the set of all a ∈ A for which the map t 7→ e2πitφj · a is differentiable.

Exercise A.32. Show that ur ∈ Dom δj , and that δj (u r r r ∈ Zn and

T ) = 2πirj u m1for all
j = 1, . . . , n. Conclude that the common smooth domain m∈Zn Dom(δ1 . . . δnmn ) is equal to
the subalgebra AΘ .

105
 The result of the previous exercise shows that AΘ is just the “smooth subalgebra” of
the C ∗ -algebra AΘ with respect to the action of Tn . It is known that any such smooth
subalgebra, under a continuous action of a compact Lie group on a C ∗ -algebra, is actually a
pre-C ∗ -algebra.
Exercise A.33. Define a linear operator E : AΘ → AΘ by averaging over the orbits of this
Tn -action: Z
E(a) := (e−2πiφ1 , . . . , e−2πiφn ) · a dφ1 . . . dφn .
[0,1]n

Check that E(1) = 1, that E(a∗ ) = E(a)∗ , that E(a∗ a) ≥ 0 and kE(a)k ≤ kak for all a ∈ AΘ ;
where “x ≥ 0” means that x is a positive element of AΘ . Then show the “conditional
expectation” property:


E E(a) b E(c) = E(a) E(b) E(c) for all a, b, c ∈ AΘ .

By considering b = a − E(a), show also that E(a∗ a) ≥ E(a)∗ E(a) for a ∈ AΘ .

P
Exercise A.34. If a = r ar ur ∈ AΘ , check that E(a) = a0 1. Conclude that the range
of E is the ∗-subalgebra C 1, and that

τ (a) 1 := E(a)

defines a trace on AΘ ; by continuity, it is enough to check the trace property on the dense
subalgebra AΘ .
Exercise A.35. If instead we only consider the action of a subgroup Tk of Tn , we can define
a conditional expectation
Z
Ek (a) := (e−2πiφ1 , . . . , e−2πiφk , 1, . . . , 1) · a dφ1 . . . dφk .
[0,1]k

In this case the range of Ek will be isomorphic to a C ∗ -algebra AΦ where Φ is a certain real
skewsymmetric matrix in Mn−k (R). Compute the matrix Φ in terms of the matrix Θ. In
particular, ¿what is the range of Ek for the case k = n − 1?
We now define Hτ to be the completion of AΘ in the norm
p
kak2 := τ (a∗ a).

We remark that kak2 ≤ kak for all a, so that the inclusion map ητ : AΘ → Hτ is continuous.
It is convenient to write a := ητ (a) to denote the element a ∈ AΘ regarded as a vector in Hτ .
It turns out that the trace τ is faithful, so that Hτ is just the Hilbert space of the “GNS
representation” πτ of AΘ . This representation is defined —first on ητ (AΘ ), then extended by
continuity— by
πτ (a) : b 7→ ab : Hτ → Hτ , for each a ∈ AΘ .
Exercise A.36. Define an antilinear operator J0 : Hτ → Hτ by setting

J0 (a) := a∗ , for a ∈ ητ (AΘ ).

Show that J0 is an isometry on this domain, so that it extends to all of Hτ ; and show that
the extended J0 is an antiunitary operator on Hτ . For b ∈ AΘ , consider the operator

πτ′ (b) := J0 πτ (b∗ )J0 .

Check that πτ′ (b) : c 7→ cb for c ∈ AΘ . Conclude that [πτ (a), πτ′ (b)] = 0 for all a, b ∈ AΘ .

106
 The analogue of the L2 -spinor space for the noncommutative torus is just the tensor
m
product H := Hτ ⊗ C2 , where as usual, n = 2m or n = 2m + 1 according as n is even
or odd. (In the commutative case Θ = 0, this means that we are using the spinor module
m
for the untwisted spin structure on Tn .) Recall that we can regard C2 as a Fock space
Λ• Cm , carrying an irreducible represenation of the matrix algebra B = Cl(Rn ) if n is even,
or B = Cl0 (Rn ) if n is odd. In the even case, there is a Z2 -grading operator Γ := 1Hτ ⊗ c(γ),
satisfying Γ2 = 1 and Γ∗ = Γ.
The charge conjugation on B, that we have written b 7→ χ(b̄), is implemented by an
m
antiunitary operator on C2 of the form C0 K, where K is complex conjugation and C0 is a
certain 2m × 2m matrix: this means that (C0 K)  b (C0 K)−1 = χ(b̄) as operators on C2m .

0 −1
For instance, if n = 2 or 3, then C0 = i σ 2 = .
1 0
Now let J := J0 ⊗ C0 . This is an antiunitary operator on H, such that J 2 = ±1 according
as C02 = ±1.
Exercise A.37. Show that δj (a∗ ) = (δj (a))∗ and that τ (δj (a)) = 0 for all a ∈ AΘ . Conclude
that the densely defined operator δ j : a 7→ δj (a), with domain ητ (AΘ ), is skewsymmetric in
the sense that
hδ j (a) | bi = −ha | δ j (b)i, for all a, b ∈ Dom δ j .
The closure of this operator, still denoted by δ j , is then an unbounded skewadjoint oper-
ator on H.
m
Let γ 1 , . . . , γ n be the generators of the action of the Clifford algebra Cl(Rn ) on C2 : they
are a set of 2m × 2m matrices such that γ j γ k + γ k γ j = 2δ jk for j, k = 1, . . . , n. The operator
C0 K is determined by the relations
(C0 K) γ j (C0 K)−1 = −γ j for j = 1, . . . , n.
We can now define the Dirac operator on H by
n
X
D := −i δj ⊗ γ j .
j=1

Exercise A.38. Show that JDJ −1 = ±D on the domain AΘ .

m
Exercise A.39. If { sα : α = 1, . . . , 2m } is an orthonormal basis of C2 , define ψrα :=
ur ⊗ sα ∈ H. Show that { ψrα : r ∈ Zn , α = 1, . . . , 2m } is an orthonormal basis of H that
diagonalizes D2 , by checking that
D2 ψrα = 4π 2 (r · r) ψrα for each r, α.
¿What is the spectrum (with its multiplicities) of |D|? What is the spectrum of D itself ?
Exercise A.40. We can invert D on the orthogonal complement of the finite-dimensional
space ker D = span{ ψ0α : α = 1, . . . , 2m }. Show that, for each s > 0, the expression
σN (|D|−s )
Tr+ |D|−s := lim
N →∞ log N
either exists as a finite limit, or diverges to +∞. (Show that we may use a subsequence where
N = NR := #{ r ∈ Zn : r · r ≤ R2 } for some R > 0.) Verify that the 0 < Tr+ |D|−s < +∞
if and only if s = n; and compute the value of Tr+ |D|−n .
Exercise A.41. If a ∈ AΘ , show that both a and [D, a], considered as bounded operators
on H, lie in the smooth domain of the operator T 7→ [|D|, T ].

107
Bibliography

[ABS] M. F. Atiyah, R. Bott and A. Shapiro, “Clifford modules”, Topology 3 (1964),

3–38.

[BHMS] P. Baum, P. M. Hajac, R. Matthes and W. Szymański, “Noncommutative geometry

approach to principal and associated bundles”, Warszawa, 2006, forthcoming.

[BGV] N. Berline, E. Getzler and M. Vergne, Heat Kernels and Dirac Operators, Springer,
Berlin, 1992.

[Bla] B. Blackadar, K-theory for Operator Algebras, 2nd edition, Cambridge Univ. Press,
Cambridge, 1998.

[Bost] J.-B. Bost, “Principe d’Oka, K-théorie et systèmes dynamiques non commutatifs”,
Invent. Math. 101 (1990), 261–333.

[BR] O. Bratteli and D. W. Robinson, Operator Algebras and Quantum Statistical Me-
chanics 1, Springer, New York, 1987.

[BT] P. Budinich and A. Trautman, The Spinorial Chessboard, Trieste Notes in Physics,
Springer, Berlin, 1988.

[BW] H. Bursztyn and S. Waldmann, “Bimodule deformations, Picard groups and con-
travariant connections”, K-Theory 31 (2004), 1–37.

[CPRS] A. L. Carey, J. Phillips, A. Rennie and F. A. Sukochev, “The Hochschild class of

the Chern character for semifinite spectral triples”, J. Funct. Anal. 213 (2004),
111–153.

[Che] C. Chevalley, The Algebraic Theory of Spinors, Columbia Univ. Press, New York,
1954.

[Con1] A. Connes, “The action functional in noncommutative geometry”, Commun. Math.

Phys. 117 (1988), 673–683.

[Con] A. Connes, Noncommutative Geometry, Academic Press, London and San Diego,
1994.

[Con2] A. Connes, “Gravity coupled with matter and foundation of noncommutative ge-
ometry”, Commun. Math. Phys. 182 (1996), 155–176.

[CDV] A. Connes and M. Dubois-Violette, “Noncommutative finite-dimensional manifolds.

I. Spherical manifolds and related examples”, Commun. Math. Phys. 230 (2002),
539–579.

108
[CL] A. Connes and G. Landi, “Noncommutative manifolds, the instanton algebra and
isospectral deformations”, Commun. Math. Phys. 221 (2001), 141–159.

[CM] A. Connes and H. Moscovici, “The local index formula in noncommutative geome-
try”, Geom. Func. Anal. 5 (1995), 174–243.

[DLSSV] L. Da̧browski, G. Landi, A. Sitarz, W. van Suijlekom and J. C. Várilly, “The Dirac
operator on SUq (2)”, Commun. Math. Phys. 259 (2005), 729–759.

[DS] L. Da̧browski and A. Sitarz, “Dirac operator on the standard Podleś quantum
sphere”, in Noncommutative Geometry and Quantum Groups, P. M. Hajac and
W. Pusz, eds. (Instytut Matematyczny PAN, Warszawa, 2003), pp. 49–58.

[Dix] J. Dixmier, Les C ∗ -algèbres et leurs Représentations, Gauthier-Villars, Paris, 1964;

2nd edition, 1969.

[Dix1] J. Dixmier, “Existence de traces non normales”, C. R. Acad. Sci. Paris 262A
(1966), 1107–1108.

[Fri] T. Friedrich, Dirac Operators in Riemannian Geometry, Graduate Studies in Math-

ematics 25, American Mathematical Society, Providence, RI, 2000.

[GGISV] V. Gayral, J. M. Gracia-Bondı́a, B. Iochum, T. Schücker and J. C. Várilly, “Moyal

planes are spectral triples”, Commun. Math. Phys. 246 (2004), 569–623.

[GV] J. M. Gracia-Bondı́a and J. C. Várilly, “Algebras of distributions suitable for phase-

space quantum mechanics. I”, J. Math. Phys. 29 (1988), 869–879.

[GVF] J. M. Gracia-Bondı́a, J. C. Várilly and H. Figueroa, Elements of Noncommutative

Geometry, Birkhäuser, Boston, 2001.

[HH] J. W. Helton and R. E. Howe, “Integral operators: traces, index, and homology”,
in Proceedings of a Conference on Operator Theory, P. A. Fillmore, ed., Lecture
Notes in Mathematics 345, Springer, Berlin, 1973; pp. 141–209.

[Hig] N. Higson, “The local index formula in noncommutative geometry”, in Con-

temporary Developments in Algebraic K-Theory, M. Karoubi, A. O. Kuku and
C. Pedrini, eds. (ICTP, Trieste, 2004), pp. 443–536. Also available at the URL
<http://www.math.psu.edu/higson/Papers/trieste.pdf>

[Kar] G. Karrer, “Einführung von Spinoren auf Riemannschen Mannigfaltigkeiten”, Ann.

Acad. Sci. Fennicae Ser. A I Math. 336/5 (1963), 3–16.

[LM] H. B. Lawson and M.-L. Michelsohn, Spin Geometry, Princeton Univ. Press, Prince-
ton, NJ, 1989.

[Lich] A. Lichnerowicz, “Spineurs harmoniques”, C. R. Acad. Sci. Paris 257A (1963),

7–9.

[LSS] S. Lord, A. A. Sedaev and F. A. Sukochev, “Dixmier traces as singular symmetric

functionals and applications to measurable operators”, J. Funct. Anal. 224 (2005),
72–106.

[Mil] J. W. Milnor, Morse Theory, Princeton University Press, Princeton, NJ, 1963.

109
[Moy] J. E. Moyal, “Quantum mechanics as a statistical theory”, Proc. Cambridge Philos.
Soc. 45 (1949), 99–124.

[Ply] R. J. Plymen, “Strong Morita equivalence, spinors and symplectic spinors”, J. Oper.
Theory 16 (1986), 305–324.

[RW] I. Raeburn and D. P. Williams, Morita Equivalence and Continuous-Trace C ∗ -

algebras, Amer. Math. Soc., Providence, RI, 1998.

[RS] M. Reed and B. Simon, Methods of Modern Mathematical Physics, I: Functional

Analysis, Academic Press, New York, 1972.

[Ren] A. Rennie, “Smoothness and locality for nonunital spectral triples”, K-Theory 28
(2003), 127–165.

[Rie] M. A. Rieffel, Deformation Quantization for Actions of Rd , Memoirs of the Amer-

ican Mathematical Society 506, Providence, RI, 1993.

[Schd] H. Schröder, “On the definition of geometric Dirac operators”, Dortmund, 2000;
math.dg/0005239.

[Sch1] E. Schrödinger, “Diracsches Elektron in Schwerefeld I”, Sitzungsber. Preuss. Akad.

Wissen. Phys.-Math. 11 (1932), 105–128.

[Schz] L. Schwartz, Théorie des Distributions, Hermann, Paris, 1966.

[Schw] L. B. Schweitzer, “A short proof that Mn (A) is local if A is local and Fréchet”, Int.
J. Math. 3 (1992), 581–589.

[See] R. T. Seeley, “Complex powers of an elliptic operator”, Proc. Symp. Pure Math.
10 (1967), 288–307.

[Sim] B. Simon, Trace Ideals and their Applications, Cambridge Univ. Press, Cambridge,
1979.

[SDLSV] W. van Suijlekom, L. Da̧browski, G. Landi, A. Sitarz and J. C. Várilly, “The local
index formula for SUq (2)”, K-Theory (2006), in press.

[Tay] M. E. Taylor, Partial Differential Equations II, Springer, Berlin, 1996.

[VG] J. C. Várilly and J. M. Gracia-Bondı́a, “Algebras of distributions suitable for phase-

space quantum mechanics. II. Topologies on the Moyal algebra”, J. Math. Phys.
29 (1988), 880–887.

[Wodz] M. Wodzicki, “Local invariants of spectral asymmetry”, Invent. Math. 75 (1984),

143–178.

[Wolf] J. A. Wolf, “Essential selfadjointness for the Dirac operator and its square”, Indiana
Univ. Math. J. 22 (1973), 611–640.

110