Preprint typeset in JHEP style - HYPER VERSION

Quantum Symmetries and K-Theory

Gregory W. Moore

Abstract: Modified from Vienna Lectures August 2014 for PiTP, July 2015. VERY
ROUGH VERSION: UNDER CONSTRUCTION. Version: July 23, 2015
Contents

1. Lecture 1: Quantum Symmetries and the 3-fold way 3

1.1 Introduction 3
1.2 Quantum Automorphisms 4
1.2.1 States, operators and probabilities 4
1.2.2 Automorphisms of a quantum system 5
1.2.3 Overlap function and the Fubini-Study distance 6
1.2.4 From (anti-) linear maps to quantum automorphisms 8
1.2.5 Wigner’s theorem 9
1.3 A little bit about group extensions 9
1.3.1 Example 1: SU (2) and SO(3) 11
1.3.2 Example 2: The isometry group of affine Euclidean space Ed 12
1.3.3 Example 3: Crystallographic groups 13
1.4 Restatement of Wigner’s theorem 14
1.5 φ-twisted extensions 16
1.5.1 φ-twisted extensions 16
1.6 Real, complex, and quaternionic vector spaces 18
1.6.1 Complex structure on a real vector space 18
1.6.2 Real structure on a complex vector space 19
1.6.3 The quaternions and quaternionic vector spaces 21
1.6.4 Quaternionic Structure On Complex Vector Space 25
1.6.5 Complex Structure On Quaternionic Vector Space 25
1.6.6 Summary 26
1.7 φ-twisted representations 26
1.7.1 Some definitions 26
1.7.2 Schur’s Lemma for φ-reps 27
1.7.3 Complete Reducibility 29
1.7.4 Complete Reducibility in terms of algebras 30
1.8 Groups Compatible With Quantum Dynamics 32
1.9 Dyson’s 3-fold way 35
1.10 The Dyson problem 36

2. Lecture 2: Phases of gapped systems: The 10-fold way 38

2.1 Gapped systems and the notion of phases 38
2.2 Z2 -graded, or super-, linear algebra 41
2.2.1 Super vector spaces 41
2.2.2 Linear transformations between supervector spaces 43
2.2.3 Superalgebras 44
2.2.4 Modules over superalgebras 45
2.2.5 Star-structures and super-Hilbert spaces 46

–1–
 2.3 Clifford Algebras and Their Modules 48
2.3.1 The real and complex Clifford algebras 48
2.3.2 Clifford algebras and modules over κ = C 50
2.3.3 Morita equivalence of algebras 52
2.3.4 Morita equivalence of Clifford algebras and the complex K-theory
of a point 54
2.3.5 Real Clifford algebras and Clifford modules of low dimension 56
2.3.6 The periodicity theorem for Clifford algebras 57
2.3.7 KO-theory of a point 59
2.3.8 Digression: A hint of the relation to topology 61
2.4 The 10 Real Super-division Algebras 69
2.5 The 10-fold way for gapped quantum systems 70
2.6 Realizing the 10 classes using the CT groups 73

3. Lecture 3: Free Fermions, the Altland-Zirnbauer classification, and Bott

Periodicity 78
3.1 Fermions and the Spin Representation 78
3.1.1 Finite dimensional fermionic systems 79
3.1.2 Spin representations from complex isotropic subspaces 80
3.1.3 Fermionic Oscillators 84
3.1.4 Bogoliubov transformations 85
3.2 Free fermion dynamics and their symmetries 87
3.2.1 FDFS with symmetry 87
3.2.2 Free fermion dynamics 89
3.2.3 Symmetries of free fermion systems 91
3.2.4 The free fermion Dyson problem and the Altland-Zirnbauer classifi-
cation 92
3.2.5 Analog for bosons 94
3.3 Symmetric Spaces and Classifying Spaces 95
3.3.1 The Bott song and the 10 classical Cartan symmetric spaces 95
3.3.2 Cartan embedding of the symmetric spaces 98
3.3.3 Application: Uniform realization of the Altland-Zirnbauer classes 99
3.3.4 Relation to Morse theory and loop spaces 100
3.3.5 Relation to classifying spaces of K-theory 103

4. Lecture 4: K-theory classification and application to topological band

structure 105
4.1 Reduced topological phases of a FDFS and twisted equivariant K-theory of
a point 105
4.1.1 Definition of G-equivariant K-theory of a point 105
4.1.2 Definition of twisted G-equivariant K-theory of a point 106
4.1.3 Appliction to FDFS: Reduced topological phases 108
4.2 Motivation: Bloch theory 109

–2–
 4.2.1 Bands and insulators 112
4.3 Groupoids 114
4.3.1 Categories 114
4.3.2 Groups and groupoids 115
4.3.3 Vector bundles on groupoids 116
4.4 Twisted equivariant K-theory of groupoids 117
4.4.1 Central extensions as line bundles over a group 117
4.4.2 φ-twisted extensions and representations of a graded group 118
4.4.3 Twisting the K-theory of a groupoid 118
4.4.4 Definition of a twisted bundle and twisted K-theory of a groupoid 119
4.5 Applications to topological band structure 120
4.5.1 The canonical twisting 120
4.5.2 E as a ν can -twisted bundle over T ∨ //P (C) 123
4.5.3 Modifications for including (φ, χ) 124
4.5.4 K-theory classification of phases in topological band theory 125
4.6 Computation of some equivariant K-theory invariants via localization 126
4.6.1 Localization of rings and modules 126
4.6.2 The localization theorem 127
4.6.3 Example: Some twisted equivariant K-theory invariants for diamond
structure 128

A. Background material: Cartan’s symmetric spaces 130

B. Group cohomology and central extensions 133

B.1 Central extensions 133
B.2 Group cohomology in other degrees 140
B.2.1 Definition 140
B.3 The topology behind group cohomology 142

1. Lecture 1: Quantum Symmetries and the 3-fold way

1.1 Introduction
These lecture notes are a truncated version of an more extended set of notes which can be
found in [43].
The notes are meant to accompany a set of lectures given at the Erwin Schrödinger
Institute in Vienna, August 18-21, 2014. They constitute a review of material most of
which has been drawn from the paper [26]. That paper, in turn, was largely meant as an
exposition, in formal mathematical terms, of some beautiful results of C. Kane et. al., A.
Kitaev, A. Ludwig et. al., and M. Zirnbauer et. al.
There is very little here which has not been published previously. There are only
two novelties: One is a stress on a presentation of the 10-fold classification of gapped

–3–
phases which follows only from general properties of quantum mechanics. The presentation
deliberately avoids the mention of free fermions. It in principle could apply to systems of
fermions or bosons, interacting or not. The second, rather minor, novelty is the use of
localization techniques in equivariant K-theory to demonstrate some of the refined K-
theory invariants in the theory of topological band structure which appear when one works
equivariantly with respect to a crystallographic group.
For a more extensive introduction describing material to be covered see the first section
of [43]. The URL is
http://www.physics.rutgers.edu/ gmoore/QuantumSymmetryBook.pdf
The present notes can be found at
http://www.physics.rutgers.edu/ gmoore/ViennaLecturesF.pdf
The lectures are an extended version of a talk at the SCGP:
http://www.physics.rutgers.edu/ gmoore/SCGP-TWISTEDMATTER-2013D.pdf

Acknowledgements

I thank Martin Zirnbauer and Nick Read for the invitation to attend the ESI workshop
and for the invitation to give this series of lectures. I also thank the participants at the
workshop for asking many good questions. I particularly thank Johannes Kellendonk,
Andreas Ludwig and Guo Chuan Thiang for certain particularly good questions. I am
supported by DOE grant DOE-SC0010008 to Rutgers. I am also grateful for the hospitality
of the Aspen Center for Physics (under NSF Grant No. PHY-1066293) where I wrote the
lectures, and of course I acknowledge the hospitality of the Erwin Schrödinger Institute in
Vienna.

1.2 Quantum Automorphisms

1.2.1 States, operators and probabilities
Let us recall the basic tenets of quantum mechanics (these are the most important of the
Dirac-von Neumann axioms)

1. To a physical system we associate a complex Hilbert space H such that

2. Physical states are identified with traceclass positive operators ρ of trace one. They
are usually called density matrices. We denote the space of physical states by S.

3. Physical observables are identified with self-adjoint operators. We denote the set of
self-adjoint operators by O. 1

4. Born rule: When measuring the observable O in a state ρ the probability of measuring
value e ∈ E ⊂ R, where E is a Borel-measurable subset of R, is

Pρ,O (E) = TrPO (E)ρ. (1.1)

Here PO is the projection-valued-measure associated to the self-adjoint operator O

by the spectral theorem.
1
We are skating over many subtleties here.

–4–
 Now, the set of states S is a convex set and hence the extremal points, known as the
pure states are distinguished. These are the dimension one projection operators. They are
often identified with rays in Hilbert space for the following reason:
If ψ ∈ H is a nonzero vector then it determines a line

ℓψ := {zψ|z ∈ C} := ψC (1.2)

Note that the line does not depend on the normalization or phase of ψ, that is, ℓψ = ℓzψ
for any nonzero complex number z. Put differently, the space of such lines is projective
Hilbert space
PH := (H − {0})/C∗ (1.3)
Equivalently, this can be identified with the space of rank one projection operators. Indeed,
given any line ℓ ⊂ H we can write, in Dirac’s bra-ket notation: 2
|ψihψ|
Pℓ = (1.4)
hψ|ψi
where ψ is any nonzero vector in the line ℓ.

1.2.2 Automorphisms of a quantum system

Now we state the formal notion of a general “symmetry” in quantum mechanics:

Definition An automorphism of a quantum system is a pair of bijective maps s1 : O → O

and s2 : S → S where s1 is real linear on O such that (s1 , s2 ) preserves probability
measures:
Ps1 (O),s2 (ρ) = PO,ρ (1.5)
This set of mappings forms a group which we will call the group of quantum automorphisms. ♣Need to state
some appropriate
One can show that that s = s2 must take extreme states to extreme states, and hence continuity
properties. ♣
induces a single map
s : PH → PH. (1.6)
Moreover, the preservation of probabilities, restricted to the case of self-adjoint operators
given by rank one projectors and pure states (also given by rank one projectors) means
that the function
o : PH × PH → [0, 1] (1.7)
defined by
o(ℓ1 , ℓ2 ) := TrPℓ1 Pℓ2 (1.8)
must be invariant under s:
o(s(ℓ1 ), s(ℓ2 )) = o(ℓ1 , ℓ2 ) (1.9)
2
We generally denote inner products in Hilbert space by (x1 , x2 ) ∈ C where x1 , x2 ∈ H. Our convention
is that it is complex-linear in the second argument. However, we sometimes write equations in Dirac’s
bra-ket notation because it is very popular. In this case, identify x with |xi. Using the Hermitian structure
there is a unique anti-linear isomorphism of H with H∗ which we denote x 7→ hx|. Sometimes we denote
vectors by Greek letters ψ, χ, . . . , and scalars by Latin letters z, w, . . . . But sometimes we denote vectors
by Latin letters, x, w, . . . and scalars by Greek letters, α, β, . . . .

–5–
 Definition The function defined by (1.7) and (1.8) is known as the overlap function.

Remarks

1. The upshot of our arguments above is that the quantum automorphism group of a
system with Hilbert space H can be identified with the group of (suitably continuous)
maps (1.6) such that (1.9) holds for all lines ℓ1 , ℓ2 . We denote the group of such maps
by Autqtm (PH).

2. The reason for the name “overlap function” or “transition probability” which is also
used, is that if we choose representative vectors ψ1 ∈ ℓ1 and ψ2 ∈ ℓ2 we obtain the -
perhaps more familiar - expression:
|hψ1 |ψ2 i|2
TrPℓ1 Pℓ2 = (1.10)
hψ1 |ψ1 ihψ2 |ψ2 i

1.2.3 Overlap function and the Fubini-Study distance

If H is finite dimensional then we can identify it as H ∼ = CN with the standard hermitian
metric. Then PH = CP N −1 and there is a well-known metric on CPN −1 known as the
“Fubini-Study metric” from which one can define a minimal geodesic distance d(ℓ1 , ℓ2 )
between two lines (or projection operators). When the FS metric is suitably normalized
the overlap function o is nicely related to the Fubini-Study distance d by
 
d(ℓ1 , ℓ2 ) 2
o(ℓ1 , ℓ2 ) = cos (1.11)
2
Let us first check this for the case N = 2. Then we claim that for the case

PH2 = CP 1 ∼
= S2 (1.12)

d is just the usual round metric on the sphere and the proper normalization will be unit
radius. Let us first check this:
First we write the most general general density matrix in two dimensions. Any 2 × 2
Hermitian matrix is of the form a + ~b · ~σ where ~σ is the vector of “Pauli matrices”:
!
0 1
σ1 =
10
!
2 0 −i
σ = (1.13)
i 0
!
3 1 0
σ =
0 −1

a ∈ R and ~b ∈ R3 . Now a density matrix ρ must have trace one, and therefore a = 21 . Then
the eigenvalues are 21 ± |~b| so positivity means it must have the form
1
ρ = (1 + ~x · ~σ ) (1.14)
2

–6–
where ~x ∈ R3 with ~x2 ≤ 1.
The extremal states, corresponding to the rank one projection operators are therefore
of the form
1
P~n = (1 + ~n · ~σ ) (1.15)
2
where ~n is a unit vector. This gives the explicit identification of the pure states with
elements of S 2 . Moreover, we can easily compute:
1
TrP~n1 P~n2 = (1 + ~n1 · ~n2 ) (1.16)
2
and ~n1 · ~n2 = cos(θ1 − θ2 ) where |θ1 − θ2 | (with θ’s chosen so this is between 0 and π) is
the geodesic distance between the two points on the unit sphere. Thus we obtain (1.11).
There is another viewpoint which is useful. Nonzero vectors in C2 can be normalized
to be in the unit sphere S 3 . Then the association of projector to state given by
1
|ψi → |ψihψ| = (1 + ~n · ~σ ) (1.17)
2
defines a map π : S 3 → S 2 known as the Hopf fibration.
The unit sphere is a principal homogeneous space for SU (2) and we may coordinatize
SU (2) by the Euler angles:
φ 3 θ 2 ψ 3
u = e−i 2 σ e−i 2 σ e−i 2 σ (1.18)
with range 0 ≤ θ ≤ π and identifications:

(φ, ψ) ∼ (φ + 4π, ψ) ∼ (φ, ψ + 4π) ∼ (φ + 2π, ψ + 2π) (1.19)

We can make an identification with the unit sphere in C2 by viewing it as a homogeneous

space: ! !
ψ+φ
e−i 2 cos θ/2 1
ψ= ψ−φ =u· (1.20)
e−i 2 sin θ/2 0
The projector onto the line through this space is
1
Pℓψ = |ψihψ| = (1 + ~n · ~σ ) (1.21)
2
with ~n = (sin θ cos φ, sin θ sin φ, cos θ) as usual. Alternatively, we could map π : S 3 → S 2
by π(ψ) = [ψ1 : ψ2 ] ∼ = CP 1 , and this will correspond to the point in S 2 by the usual
stereographic projection. ♣from which pole?
♣
In any case, for the case N = 2 we see that Autqtm (PH) is just the group of isometries
2
of S with its round metric. This group is well known to be the orthogonal group O(3).
It is known that the FS metric on CP N has the property that the submanifolds CP k →
N
CP embedded by [z1 : · · · : zk+1 ] → [z1 : · · · : zN +1 ] are totally geodesic submanifolds.
Definition If (M, g) is a Riemannian manifold a submanifold M1 ⊂ M is said to be
totally geodesic if the geodesics between any two points in M1 with respect to the induced
metric (the pullback of g) are the same as the geodesics between those two points considered
as points of M .

–7–
 If M1 is the fixed point set of an isometry of (M, g) then it is totally geodesic. Now ♣simple proof or
ref? ♣
note that the submanifolds CP k are fixed points of the isometry

[z1 : · · · : zN +1 ] → [z1 : · · · : zk+1 : −zk+2 : · · · : −zN +1 ] (1.22)

Another way to see this from the viewpoint of homogeneous spaces is that if we exponentiate
a Lie algebra element in p to give a geodesic in U (N + 1) and project to the homogeneous
space we get all geodesics on the homogeneous space. But for any t ∈ p we can put it into
a U (2) subalgebra.
Now, any two lines ℓ1 , ℓ2 span a two-dimensional sub-Hilbert space of H, so, thanks to
the totally geodesic property of the FS metric, our discussion for H ∼= C2 suffices to check
(1.11) in general.

1.2.4 From (anti-) linear maps to quantum automorphisms

Now, there is one fairly obvious way to make elements of Autqtm (PH). Suppose u ∈ U (H)
is a unitary operator. Then it certainly takes lines to lines and hence can be used to define
a map (which we also denote by u) u : PH → PH. For example if we identify ℓ as ℓψ for
some nonzero vector ψ then we can define

u(ℓψ ) := ℓu(ψ) (1.23)

One checks that which vector ψ we use does not matter and hence the map is well-defined.
In terms of projection operators:
u : P 7→ uP u† (1.24)

and, since u is unitary, the overlaps Tr(P1 P2 ) are preserved.

Now - very importantly - this is not the only way to make elements of Autqtm (PH).
We call a map a : H → H anti-linear if

a(ψ1 + ψ2 ) = a(ψ1 ) + a(ψ2 ) (1.25)

but
a(zψ) = z ∗ a(ψ) (1.26)

where z is a complex scalar. It is in addition called anti-unitary if it is norm-preserving:

k a(ψ) k2 =k ψ k2 (1.27)

Now, anti-unitary maps also can be used to define quantum automorphisms. If we try
to define a(ℓ), ℓ ∈ CH by
a(ℓψ ) = ℓa(ψ) (1.28)

–8–
1.2.5 Wigner’s theorem
In the previous subsection we showed how unitary and antiunitary operators on Hilbert
space induce quantum automorphisms. Are there other ways of making quantum automor-
phisms? Wigner’s theorem says no:

Theorem: Every quantum automorphism Autqtm (PH) is induced by a unitary or

antiunitary operator on Hilbert space, as above.
In addition to Wigner’s own argument the paper [49] cites 26 references with alternative
proofs! (And there are others, for examples [52][24].) One proof, which follows Bargmann’s
paper, is in the notes [43]. Reference [24] gives two proofs. One is fairly elementary and
quick: It uses the identification of the tangent space to PH at a line ℓ as Hom(ℓ, ℓ⊥ ).
WLOG one can assume the isometry takes ℓ to ℓ and then such an isometry induces one on
the tangent space. This is a priori only a real linear operator, but then one can prove that
it must be complex linear or anti-linear. Then the complex linear or anti-linear isometry
of ℓ⊥ can be extended to H = ℓ ⊕ ℓ⊥ as a complex linear or anti-linear isometry.

1.3 A little bit about group extensions

If N , G, and Q are three groups and ι and π are homomorphisms such that
ι π
1→N → G → Q→1 (1.29)

is exact at N, G and Q then the sequence is called a short exact sequence and we say that
G is an extension of Q by N .
Note that since ι is injective we can identify N with its image in G. Then, N is a
kernel of a homomorphism (namely π) and is hence a normal or invariant subgroup (hence
the notation). Then it is well-known that G/N is a group and is in fact isomorphic to the
image of π. That group Q is thus a quotient of G (hence the notation).
There is a notion of homomorphism of two group extensions
ι
1 π
1→N → G1 →1 Q→1 (1.30)
ι
2 π
1→N → G2 →2 Q→1 (1.31)

This means that there is a group homomorphism ϕ : G1 → G2 so that the following diagram
commutes:

ι1 π1
1 // N // G1 // Q // 1 (1.32)
ϕ
ι2  π2
1 // N // G2 // Q // 1

When there is a homomorphism of group extensions based on ψ : G2 → G1 such that ϕ ◦ ψ

and ψ ◦ ϕ are the identity then the group extensions are said to be isomorphic extensions.
Given group N and Q it can certainly happen that there is more than one nonisomor-
phic extension of Q by N . Classifying all extensions of Q by N is a difficult problem. We

–9–
Figure 1: Illustration of a group extension 1 → N → G → Q → 1 as an N -bundle over Q. The
fiber over q ∈ Q is just the preimage under π.

will just mention one simple result here, since it introduces a concept that will be quite
useful later.
We would encourage the reader to think geometrically about this problem, even in
the case when Q and N are finite groups, as in Figure 1. In particular we will use the
important notion of a section, that is, a right-inverse to π: It is a map s : Q → G such that
π(s(q)) = q for all q ∈ Q. Such sections always exist.3 Note that in general s(π(g)) 6= g.
This is obvious from Figure 1: The map π projects the entire “fiber over q” to q. The
section s chooses just one point above q in that fiber.
Now, given an extension and a choice of section s we define a map

ω : Q → Aut(N ) (1.33)

q 7→ ωq (1.34)
The definition is given by
ι(ωq (n)) = s(q)ι(n)s(q)−1 (1.35)
Because ι(N ) is normal the RHS is again in ι(N ). Because ι is injective ωq (n) is well-
defined. Moreover, for each q the reader should check that indeed ωq (n1 n2 ) = ωq (n1 )ωq (n2 ),
therefore we really have homomorphism N → N . Moreover ωq is invertible (show this!)
and hence it is an automorphism.
3
By the axiom of choice. For continuous groups such as Lie groups there might or might not be continuous
sections.

– 10 –
 Remark: Clearly the ι is a bit of a nuisance and leads to clutter and it can be safely
dropped if we consider N simply to be a subgroup of G. The confident reader is encouraged
to do this. The formulae will be a little cleaner. However, we will be pedantic and retain
the ι in most of our formulae.

Let us stress that the map ω : Q → Aut(N) in general is not a homomorphism and in
general depends on the choice of section s. Let us see how close ω comes to being a group
homomorphism:

ι (ωq1 ◦ ωq2 (n)) = s(q1 )ι(ωq2 (n))s(q1 )−1

(1.36)
= s(q1 )s(q2 )ι(n)(s(q1 )s(q2 ))−1

In general the section is not a homomorphism, but clearly something nice happens when
it is:

Definition: We say an extension splits if there is a section s : Q → G which is also a

group homomorphism.

Theorem: An extension is isomorphic to a semidirect product iff there is a splitting. ♣If you discuss
central extensions
should explain
cocycle and H 2
1.3.1 Example 1: SU (2) and SO(3) here. Otherwise,
drop the Zp by Zp
example. ♣
Returning to (1.38) there is a standard homomorphism

π : SU (2) → SO(3) (1.37)

defined by π(u) = R where

u~x · ~σ u−1 = (R~x) · ~σ (1.38)

Note that:
1. Every proper rotation R comes from some u ∈ SU (2): This follows from the Euler
angle parametrization.
2. ker(π) = {±1}. To prove this we write the general SU (2) element as cos χ+sin χ~n·~σ .
This only commutes with all the σ i if sin χ = 0 so cos χ = ±1.
Thus we have the extremely important extension:

ι π
1 → Z2 → SU (2) → SO(3) → 1 (1.39)

The Z2 is embedded as the subgroup {±1} ⊂ SU (2), so this is a central extension.

Note that there is no continuous splitting. Such a splitting πs = Id would imply that
π∗ s∗ = 1 on the first homotopy group of SO(3). But that is impossible since it would have
factor through π1 (SU (2)) = 1.
Remarks

– 11 –
 1. The extension (1.39) generalizes to
ι π
1 → Z2 → Spin(d) → SO(d) → 1 (1.40)

as well as the two Pin groups which extend O(d):

ι π
1 → Z2 → Pin± (d) → O(d) → 1 (1.41)

we discuss these in Section *** below.

1.3.2 Example 2: The isometry group of affine Euclidean space Ed

Definition Let V be a vector space. Then an affine space modeled on V is a principal
homogeneous space for V . That is, a space with a transitive action of V (as an abelian
group) with trivial stabilizer.
The point of the notion of an affine space is that it has no natural origin. A good
example is the space of connections on a topologically nontrivial principal bundle.
Let Ed be the affine space modeled on Rd with Euclidean metric. The isometries are
the 1-1 transformations f : Ed → Ed such that

k f (p1 ) − f (p2 ) k=k p1 − p2 k (1.42)

for all p1 , p2 ∈ Ed . These transformations form a group Euc(d).

The translations act naturally on the affine space. Given v ∈ Rd we define the isome-
try:
Tv (p) := p + v (1.43)
so Tv1 +v2 = Tv1 + Tv2 and hence v 7→ Tv defines a subgroup of Euc(d) isomorphic to Rd .
One can show that there is a short exact sequence:

1 → Rd → Euc(d) → O(d) → 1 (1.44)

The rotation-reflections O(d) do not act naturally on affine space. In order to define such
an action one needs to choose an origin of the affine space.
If we do choose an origin then we can identify Ed ∼ = Rd and then to a pair R ∈ O(d)
and v ∈ Rd we can associate the isometry: 4

{R|v} : x 7→ Rx + v (1.45)

In this notation -known as the Seitz notation - the group multiplication law is

{R1 |v1 }{R2 |v2 } = {R1 R2 |v1 + R1 v2 } (1.46)

which makes clear that

1. There is a nontrivial automorphism used to construct the semidirect product: O(d):

{R|v}{1|w}{R|v}−1 = {1|Rw} (1.47)

4
Logically, since we operate with R first and then translate by v the notation should have been {v|R},
but unfortunately the notation used here is the standard one.

– 12 –
and π : {R|v} → R is a surjective homomorphism Euc(d) → O(d).
2. Thus, although Rd is abelian, the extension is not a central extension.
3. On the other hand, having chosen an origin, the sequence is split. We can choose a
splitting s : O(d) → Euc(d) by
s : R 7→ {R|0} (1.48)

Figure 2: A portion of a crystal in the two-dimensional plane.

1.3.3 Example 3: Crystallographic groups

A crystal should be distinguished from a lattice. The term “lattice” has several related
but slightly different meanings in the literature.

Definition An embedded lattice is a subgroup L ⊂ V where V is a vector space with a

nondegenerate symmetric bilinear quadratic form b such that L ∼= Zd as an abelian group.
Now there are several notions of equivalence of embedded lattices used in crystallog-
raphy. The most obvious one is that L1 is equivalent to L2 if there is an element of the
orthogonal group O(b) of V taking L1 to L2 .
Sometimes it is important not to choose an origin, so we can also have the definition:
Definitions Let L be an embedded lattice in Euclidean space Rd . Then:
a.) A crystal is a subset C ⊂ Ed invariant under translations by a rank d lattice
L(C) ⊂ Rn ⊂ Euc(n).
b.) The space group G(C) of a crystal C is the subgroup of Euc(d) taking C → C.

– 13 –
 c.) Recall that there is a projection π : G(C) → O(d). The point group P (C) of G(C)
is the image of the projection of G(C). Thus, G(C) sits in a group extension:

1 → L(C) → G(C) → P (C) → 1 (1.49)

and P (C) ∼ = G(C)/L(C).

d.) A crystallographic group is a discrete subgroup of Euc(d) which acts properly
discontinuously on Ed and has a subgroup isomorphic to an embedded d-dimensional lattice
in the translation subgroup. It therefore sits in a sequence of the form (1.49).
e.) If the group extension (1.49) splits the crystal is said to be symmorphic. Simi-
larly, for a crystallographic group G if the corresponding sequence splits it is said to be a
symmorphic group.

Remark: An important variation on this is the case where the crystal C is endowed with
some degrees of freedom, such as spins, which transform under time reversal. The magnetic
space group is the subgroup of Euc(d) × Z2 which preserves C with its degrees of freedom
and the magnetic point group is the subgroup of O(d) × Z2 which is the image under the
projection.

Example: An example of a two-dimensional crystal is shown in Figure 2 where the group

of lattice translations is Z2 . The point group is trivial. If we replace the starbursts and
smiley faces by points then the point group is a subgroup of O(2) isomorphic to Z2 × Z2
corresponding to reflection in the two axes. If the starbursts are replaced by the points Z2
and the smiley faces are replaced by points Z2 + (δ, 21 ) with 0 < δ < 12 then G(C) is not
split. Indeed the group is generated by
1
g1 : (x1 , x2 ) → (−x1 + δ, x2 + ) (1.50)
2
and
g2 : (x1 , x2 ) → (x1 , −x2 ) (1.51)
These project to reflections in O(2), but no lift of π(g1 ) will square to one because g12 is a
translation by (0, 1).

1.4 Restatement of Wigner’s theorem

Now that we have the language of group extensions it is instructive to give simple and
concise formulation of Wigner’s theorem.
Let us begin by introducing a new group AutR (H). This is the group whose elements
are unitary and anti-unitary transformations on H. The unitary operators U (H) form a
subgroup of AutR (H). If u is unitary and a is anti-unitary then ua and au are also anti-
unitary, but if a1 , a2 are antiunitary, then a1 a2 is unitary. Thus the set of all unitary and
anti-unitary operators on H form a group, which we will denote as AutR (H). Thus we
have the exact sequence

ι φ
1 → U (H) → AutR (H) → Z2 → 1 (1.52)

– 14 –
where φ is the homomorphism:
(
+1 S unitary
φ(S) := (1.53)
−1 S anti − unitary

In these notes the group Z2 will almost always be understood as the multiplicative group
{±1}.
Now, in Section §1.2.4 above we defined a homomorphism π : AutR (H) → Autqtm (PH)
by π(S)(ℓ) = ℓS(ψ) if ℓ = ℓψ . (Check it is indeed a homomorphism.) Now we recognize the
state of Wigner’s theorem as the simple statement that π is surjective. What is the kernel?
We also showed that ker(π) ∼ = U (1) where U (1) is the group of unitary transformations:

ψ 7→ zψ (1.54)

with |z| = 1. We will often denote this unitary transformation simply by z. Thus, we have
the exact sequence

ι π
1 → U (1) → AutR (H) → Autqtm (PH) → 1 (1.55)

Remarks:

1. For S ∈ AutR (H)) we have

(
φ(S) zS φ(S) = +1
Sz = z S= (1.56)
z̄S φ(S) = −1

So the sequence (1.55) is not central!

2. If we restrict the sequence (1.55) to ker(φ) then we get (taking dimH = N here, but
it also holds in infinite dimensions):
ι π
1 → U (1) → U (N ) → P U (N ) → 1 (1.57)

which is a central extension, but it is not split. This is in fact the source of interesting
things like anomalies in quantum mechanics.

3. The group AutR (H) has two connected components, measured by the homomor-
phism φ used in (1.52). This homomorphism “factors through” a homomorphism
φ′ : Autqtm (PH) → Z2 which likewise detects the connected component of this two-
component group. The phrase “factors through” means that φ and φ′ fit into the
diagram:
ι // π //
1 // U (1) AutR (H) Autqtm (PH) // 1 (1.58)
❖❖❖
❖❖❖φ
❖❖❖ φ′
❖❖❖
❖'' 
Z2

– 15 –
 Example: Again let us take H ∼
= C2 . As we saw,

Autqtm (PH) = O(3) = SO(3) ∐ P · SO(3), (1.59)

where P is any reflection. 5 Similarly, if we choose a basis for H then we can identify

AutR (H) ∼
= U (2) ∐ C · U (2) (1.60)

where C is complex conjugation with respect to that basis so that Cu = u∗ C. (Note that C
does not have a 2 × 2 matrix representation.) Now

P U (2) := U (2)/U (1) ∼

= SU (2)/Z2 ∼
= SO(3) (1.61)

1.5 φ-twisted extensions

So far we have discussed the group of all potential automorphisms of a quantum system
Autqtm (PH). However, when we include dynamics, and hence Hamiltonians, a given quan-
tum system will in general only have a subgroup of symmetries. If a physical system has
a symmetry group G then we should have a homomorphism ρ : G → Autqtm (PH).
In terms of diagrams we have

G (1.62)
ρ

ι π
1 // U (1) // AutR (H) // Autqtm (PH) // 1

The question we now want to address is:

How are G-symmetries represented on Hilbert space H?
Note that each operation ρ(g) in the group of quantum automorphisms has an entire
circle of possible lifts in AutR (H). These operators will form a group of operators which is
a certain extension of G. What extension to we get?

1.5.1 φ-twisted extensions

Now, let us return to the situation of (1.62) we define a group Gtw that fits in the diagram:

// Gtw π̃
1 // U (1) // G // 1 (1.63)
ρtw ρ
 
ι π
1 // U (1) // AutR (H) // Autqtm (PH) // 1

The definition is simple:

Gtw := {(S, g)|π(S) = ρ(g)} ⊂ AutR (H) × G (1.64)

This is known as the pullback construction.

That is, the group of operators representing the G-symmetries of a quantum system
form an extension, Gtw , of G by U (1).
5
Please do not confuse this with the notation P GL(n), P U (n) etc!

– 16 –
 What kind of extension is it?
This motivates two definitions. First

Definition: A Z2 -graded group is a pair (G, φ) where G is a group and φ : G → Z2 is a

homomorphism.
When we have such a group of course we have an extension of Z2 by G. Our examples
above show that in general it does not split. The group is a disjoint union G0 ∐ G1 of
elements which are even and odd under φ and we have the Z2 -graded multiplications:

G0 × G0 → G0
G0 × G1 → G1
(1.65)
G1 × G0 → G1
G1 × G1 → G0

This is just saying that φ is a homomorphism.

Next we have the

Definition Given a Z2 -graded group (G, φ) we define a φ-twisted extension of G to be an

extension of the form

// Gtw π
1 // U (1) // G // 1 (1.66)

where Gtw is a group such that

(
zg̃ φ(g) = 1
g̃z = z φ(g) g̃ = (1.67)
z̄g̃ φ(g) = −1

where g̃ is any lift of g ∈ G, and |z| = 1 is any phase. Put differently, if we define φtw := φ◦π
then
tw
g̃z = z φ (g̃) g̃ ∀g̃ ∈ Gtw (1.68)

Example
Take G = Z2 It will be convenient to denote M2 = {1, T̄ }, with T̄ 2 = 1. Of course,
M2 ∼
= Z2 . We take the Z2 grading to be φ(T̄ ) = −1, that is, φ : Z2 → Z2 is the identity
homomorphism. There are two inequivalent φ-twisted extensions:

// M tw π̃
1 // U (1)
2
// M2 // 1 (1.69)

Choose a lift T̃ of T̄ . Then π(T̃ 2 ) = 1, so T̃ 2 = z ∈ U (1). But, then

T̃ z = T̃ T̃ 2 = T̃ 2 T̃ = z T̃ (1.70)

on the other hand, φ(T̄ ) = −1 so

T̃ z = z −1 T̃ (1.71)

– 17 –
Therefore z 2 = 1, so z = ±1, and therefore T̃ 2 = ±1. Thus the two groups are

M2± = {z T̃ |z T̃ = T̃ z −1 & T̃ 2 = ±1} (1.72)

These possibilities are really distinct: If T̃ ′ is another lift of T̄ then T̃ ′ = µT̃ for some
µ ∈ U (1) and so
(T̃ ′ )2 = (µT̃ )2 = µµ̄T̃ 2 = T̃ 2 (1.73)

Remarks

1. It turns out that M2± is also a double cover of O(2) and in fact these turn out to be
isomorphic to the Pin-groups Pin± (2).
♣Simplify this to
one brief statement!
♣
1.6 Real, complex, and quaternionic vector spaces
1.6.1 Complex structure on a real vector space

Definition Let V be a real vector space. A complex structure on V is a linear map

I : V → V such that I 2 = −1.
Choose a squareroot of −1 and denote it i. If V is a real vector space with a complex
structure I, then we can define an associated complex vector space (V, I). We take (V, I)
to be identical with V , as sets, but define the scalar multiplication of a complex number
z ∈ C on a vector v by

z · v := x · v + I(y · v) = x · v + y · I(v) (1.74)

where z = x + iy with x, y ∈ R.
If V is finite dimensional and has a complex structure its dimension (as a real vector
space) is even. The dimension of (V, I) as a complex vector space is

1
dimC (V, I) = dimR V (1.75)
2
This is based on the simple ♣Note that in
several of the later
chapters our

Lemma If I is any 2n × 2n real matrix which squares to −12n then there is S ∈ GL(2n, R) basepoint complex
structure is −I0 .
Need to straighten
such that ! out this convention!

−1 0 −1n ♣
SIS = I0 := (1.76)
1n 0

It follows from our Lemma above that the space of all complex structures on R2n is a
homogeneous space for GL(2n, R). The stabilizer of I0 is the set of GL(2n, R) matrices of
the form !
A B
= A ⊗ 12 + iB ⊗ σ 2 (1.77)
−B A

– 18 –
and since σ 2 is conjugate to σ 3 , over the complex numbers this can be conjugated to
!
A + iB 0
(1.78)
0 A − iB

The determinant is clearly |det(A + iB)|2 and hence A + iB ∈ GL(n, C). Therefore, the
stabilizer of I0 is a group isomorphic to GL(n, C) and hence we have proven:

Proposition: The space of complex structures on R2n is:

CplxStr(R2n ) = GL(2n, R)/GL(n, C) (1.79)

If we introduce a metric g on V then we can say that a complex structure I is compatible

with g if
g(Iv, Iv ′ ) = g(v, v ′ ) (1.80)
So, when expressed relative to an ON basis for g the matrix I is orthogonal: I tr = I −1 .
But I −1 = −I, and hence I is anti-symmetric. Then it is well known that there is a matrix
S ∈ O(2n) so that
SIS −1 = I0 (1.81)
Now the stabilizer of I0 in O(2n) is of the form (1.77) and can therefore be conjugated to
(1.78). But now A + iB must be a unitary matrix so
The space of complex structures on R2n compatible with the Euclidean metric a ho-
mogeneous space isomorphic to

CmptCplxStr(R2n ) ∼
= O(2n)/U (n) (1.82)

where A + iB ∈ U (n) with A, B real is embedded into O(2n) as in (1.77). ♣Put here a
discussion of the
two components
1.6.2 Real structure on a complex vector space and the relation
between I and
orientation. ♣

Definition Suppose V is a complex vector space. Then a real structure on V is an

antilinear map C : V → V such that C 2 = +1.
If C is a real structure on a complex vector space V then we can define real vectors to
be those such that
C(v) = v (1.83)
Let us call the set of such real vectors V+ . This set is a real vector space, but it is not a
complex vector space, because C is antilinear. Indeed, if C(v) = +v then C(iv) = −iv. If
we let V− be the imaginary vectors, for which C(v) = −v then we claim

VR = V+ ⊕ V− (1.84)

The proof is simply the isomorphism

   
v + C(v) v − C(v)
v 7→ ⊕ (1.85)
2 2

– 19 –
Figure 3: The real structure C has fixed vectors given by the blue line. This is a real vector space
determined by the real structure C.

Moreover multiplication by i defines an isomorphism of real vector spaces: V+ ∼

= V− . Thus
we have
dimR V+ = dimC V (1.86)

Example V = C,
C : x + iy → eiϕ (x − iy) (1.87)
The fixed vectors under C consist of the real line at angle ϕ/2 to the x-axis as shown in
Figure 3.
In general, if V is a finite dimensional complex vector space, if we choose any basis
(over C) {vi } for V then we can define a real structure:
X X
C( zi vi ) = z̄i vi (1.88)
i i

and thus X
V+ = { ai vi |ai ∈ R} (1.89)
The space of real structures on Cn is

RealStr(Cn ) ∼
= GL(n, C)/GL(n, R) (1.90)

GL(n, C)/GL(n, R).

Remarks:

1. We introduced a group AutR (H). This is the automorphisms of H as a Hilbert space

which are real-linear. It should be distinguished from Aut(HR ) which would be a
much larger group of automorphisms of a real inner product space HR .

2. Suppose W is a real vector space with complex structure I giving us a complex vector
space (W, I). An antilinear map T : (W, I) → (W, I) is the same thing as a real linear
transformation T : W → W such that

T I + IT = 0 (1.91)

– 20 –
1.6.3 The quaternions and quaternionic vector spaces

Definition An algebra A over a field κ is a κ-vector space together with a κ-bilinear map
µ : A × A → A.
Usually we just write a · b for µ(a, b). The product satisfies:

1. a · (b + c) = a · b + a · c,

2. (b + c) · a = b · a + c · a,

3. α(a · b) = (αa) · b = a · (αb), for α ∈ κ.

Remarks

1. If there is a multiplicative unit A is called unital.

2. If a · (b · c) = (a · b) · c then A is called associative.

3. A good example of a unital associative algebra over κ is End(V ) where V is a vector

space over κ. Choosing a basis we can identify this with the set of n × n matrices
over κ. We will see many more examples.

4. Given any algebra A we can define the opposite algebra Aopp by

µopp (a, b) := µ(b, a) (1.92)

Definition The quaternion algebra H is the algebra over R with generators i, j, k satisfying
the relations
i2 = −1 j2 = −1 k2 = −1 (1.93)

ij + ji = ik + ki = jk + kj = 0 (1.94)
The quaternions form a four-dimensional algebra over R, as a vector space we can
write
H = Ri ⊕ Rj ⊕ Rk ⊕ R ∼
= R4 (1.95)
The algebra is associative, but noncommutative. It has a rich and colorful history, which
we will not recount here. Note that if we denote a generic quaternion by

q = x1 i + x2 j + x3 k + x4 (1.96)

then we can define the conjugate quaternion by the equation

q̄ := −x1 i − x2 j − x3 k + x4 (1.97)

and
q q̄ = q̄q = xµ xµ (1.98)

– 21 –
 One fact about the quaternions that will be quite useful is the following. There is a
left- and right-action of the quaternions on itself. If q is a quaternion then we can define
L(q) : H → H by
L(q) : q′ 7→ qq′ (1.99)
and similarly there is a right-action

R(q) : q′ 7→ q′ q (1.100)

The algebra of operators L(q) is isomorphic to H and the algebra of operators R(q) is
isomorphic to Hopp , which in turn is isomorphic to H itself. Now H is a four-dimensional
real vector space and L(q) and R(q) are commuting real-linear operators. Therefore there
is an inclusion
H ⊗R Hopp ֒→ EndR (H) ∼ = End(R4 ) (1.101)
Since H ⊗R Hopp has real dimension 16 this is isomorphism of algebras over R.

Definition: A quaternionic vector space is a vector space V over κ = R together with

three real linear operators I, J, K ∈ End(V ) satisfying the quaternion relations. In other
words, it is a real vector space which is a module for the quaternion algebra.

Example: The canonical example is given by taking a complex vector space V and forming

W = V ⊕ V̄ (1.102)

The underlying real vector space WR has quaternion actions:

I : (v1 , v2 ) 7→ (iv1 , iv2 ) = (iv1 , −iv2 ) (1.103)

J : (v1 , v2 ) 7→ (−v2 , v1 ) (1.104)

K : (v1 , v2 ) 7→ (−iv2 , −iv1 ) (1.105)

Just as we can have a complex structure on a real vector space, so we can have a
quaternionic structure on a complex vector space V . This is a C-anti-linear operator K on
V which squares to −1. Once we have K 2 = −1 we can combine with the operator I which
√
is just multiplication by −1, to produce J = KI and then we can check the quaternion
relations. The underlying real space VR is then a quaternionic vector space.
It is possible to put a quaternionic Hermitian structure on a quaternionic vector space
and thereby define the quaternionic unitary group. Alternatively, we can define U (n, H)
as the group of n × n matrices over H such that uu† = u† u = 1. In order to define the
conjugate-transpose matrix we use the quaternionic conjugation q → q̄ defined above.

Exercise

– 22 –
 Show that Hopp ∼
= H.

Exercise
a.) Show that
√ √ √
i→ −1σ 1 j → − −1σ 2 k→ −1σ 3 (1.106)

defines a set of 2 × 2 complex matrices satisfying the quaternion algebra. Under this
mapping a quaternion q is identified with a 2 × 2 complex matrix
!
z −w̄
q → ρ(q) = (1.107)
w z̄

with z = x4 + ix3 and w = x2 + ix1 .

b.)
Show that U (1, H) ∼= SU (2)
c.) Show that det(ρ(q)) = q q̄ = xµ xµ and use this to define a homomorphism SU (2) ×
SU (2) → SO(4).
d.) Show that the algebra Matn (H) of n × n matrices with quaternionic entries can be
identified as the subalgebra of Mat2n (C) of matrices A such that
!
0 −1
A∗ = JAJ −1 J= (1.108)
1 0

e.) Show that the unitary group over H:

U (n, H) := {u ∈ Matn (H)|u† u = 1} (1.109)

is isomorphic to
U Sp(2n) := {u ∈ U (2n, C)|u∗ = JuJ −1 } (1.110)

Exercise Complex structures on R4

a.) Show that the complex structures on R4 compatible with the Euclidean metric can
be identified as the maps
q 7→ nq n2 = −1 (1.111)
OR
q 7→ qn n2 = −1 (1.112)
b.) Use this to show that the space of such complex structures is S 2 ∐ S 2 .

– 23 –
 c.) Explain the relation to O(4)/U (2).

Exercise A natural sphere of complex structures

Show that if V is a quaternionic vector space with complex structures I, J, K then
there is a natural sphere of complex structures give by

I = x1 I + x2 J + x3 K x21 + x22 + x23 = 1 (1.113)

Exercise Regular representation ♣This is too

important to be an
Compute the left and right regular representations of H on itself Choose a real basis exercise, and is used
heavily later. ♣
for H with v1 = i, v2 = j, v3 = k, v4 = 1. Let L(q) denote left-multiplication by a quaternion
q and R(q) right-multiplciation by q. Then the representation matrices are:

L(q)va := q · va := L(q)ba vb (1.114)

R(q)va := va · q := R(q)ba vb (1.115)

a.) Show that:  
0 0 0 1
 0 0 −1 0
 
L(i) =   (1.116)
0 1 0 0
−1 0 0 0
 
0 0 1 0
0 0 0 1
 
L(j) =   (1.117)
−1 0 0 0
0 −1 0 0
 
0 −1 0 0
1 0 0 0
 
L(k) =   (1.118)
0 0 0 1
0 0 −1 0
 
0 0 0 1
0 0 1 0
 
R(i) =   (1.119)
 0 −1 0 0
−1 0 0 0
 
0 0 −1 0
0 0 0 1
 
R(j) =   (1.120)
1 0 0 0
0 −1 0 0

– 24 –
  
0 1 0 0
−1 0 0 0
 
R(k) =   (1.121)
0 0 0 1
0 0 −1 0
b.) Show that these matrices generate the full 16-dimensional algebra M4 (R). This is
the content of the statement that

H ⊗R Hopp ∼
= End(R4 ) (1.122)

Exercise ’t Hooft symbols and the regular representation of H

The famous ’t Hooft symbols, introduced by ’t Hooft in his work on instantons in
gauge theory are defined by
1
α±,i
µν := (±δiµ δν4 ∓ δiν δµ4 + ǫiµν ) (1.123)
2
where 1 ≤ µ, ν ≤ 4
a.) Show that
1 1 1
α+,1 = R(i) α+,2 = R(j) α+,3 = R(k) (1.124)
2 2 2
1 1 1
α−,1 = − L(i) α−,2 = − L(j) α−,3 = − L(k) (1.125)
2 2 2
b.) Verify the relations

[α±,i , α±,j ] = −ǫijk α±,k

[α±,i , α∓,j ] = 0 (1.126)
1
{α±,i , α±,j } = − δij
2
So
1 1
α+,i α+,j = − δij − ǫijk α+,k
4 2 (1.127)
−,i −,j 1 ij 1 ijk −,k
α α =− δ − ǫ α
4 2

1.6.4 Quaternionic Structure On Complex Vector Space

QuatStr(C2n ) ∼
= U (2n)/U Sp(2n) (1.128)

1.6.5 Complex Structure On Quaternionic Vector Space

CplxStr(Hn) ∼
= U Sp(2n)/U (n) (1.129)

– 25 –
1.6.6 Summary
To summarize we have described three basic structures we can put on vector spaces: ♣Actually, there are
four. We can have a
complex structure
1. A complex structure on a real vector space W is a real linear map I : W → W with on a quaternionic
space. We should
I 2 = −1. also derive the the
moduli spaces of all
four cases as
2. A real structure on a complex vector space V is a C-anti-linear map K : V → V with recorded in (A.14)
to (A.17). These
K 2 = +1. are used later. ♣

3. A quaternionic structure on a complex vector space V is a C-anti-linear map K :

V → V with K 2 = −1.

1.7 φ-twisted representations

Wigner’s theorem is the source of the importance of group representation theory in physics.
In these notes we are emphasizing the extra details coming from the fact that in general
some symmetry operators are represented as C-antilinear operators. In this section we
summarize a few of the differences from standard representation theory.

1.7.1 Some definitions

There are some fairly straightforward definitions generalizing the usual definitions of group
representation theory.

Definitions:

1. A φ-representation (or φ-rep for short) of a Z2 -graded group (G, φ) is a complex

vector space V together with a homomorphism

ρ : G → End(VR ) (1.130)

such that (
C − linear φ(g) = +1
ρ(g) = (1.131)
C − anti − linear φ(g) = −1

2. An intertwiner or morphism between two φ-reps (ρ1 , V1 ) and (ρ2 , V2 ) is a C-linear

map O : V1 → V2 , i.e., O ∈ HomC (V1 , V2 ), which commutes with the G-action:

Oρ1 (g) = ρ2 (g)O ∀g ∈ G (1.132)

We write HomG
C (V1 , V2 ) for the set of all intertwiners.

3. An isomorphism of φ-reps is an intertwiner O which is an isomorphism of complex

vector spaces.

4. A φ-rep is said to be φ-unitary if V has a nondegenerate sesquilinear pairing such

that ρ(g) is an isometry for all g. That is, it is unitary or anti-unitary according to
whether φ(g) = +1 or φ(g) = −1, respectively. If G is compact we can always choose
the representation to by φ-unitary.

– 26 –
 5. A φ-rep (ρ, V ) is said to be reducible if there is a proper (i.e. nontrivial) φ-sub-
representation. That is, if there is a complex vector subspace W ⊂ V , with W not
{0} or V which is G-invariant. If it is not reducible it is said to be irreducible.

Remarks:

1. In our language, then, what we learn from Wigner’s theorem is that if we have a
quantum symmetry group ρ : G → Autqtm (PH) then there is a Z2 -graded exten-
sion (Gtw , φ) and the Hilbert space is a φ-representation of (Gtw , φ). In general we
will refer to a φ-representation of some extension (Gtw , φ) of (G, φ) as a φ-twisted
representation of G.

2. In the older literature of Wigner and Dyson the term “corepresentation” for a φ-
unitary representation is used, but in modern parlance the name “corepresentation”
has several inappropriate connotations, so we avoid it. The term “φ-representation”
is not standard, but it should be.

3. An important point below will be that HomG C (V1 , V2 ) is, a priori only a real vector
space. If O is an intertwiner the iO certainly makes sense as a linear map from V1
to V2 but if any of the ρ(g) are anti-linear then iO will not be an intertwiner. Of
course, if the Z2 -grading φ of G is trivial and φ(g) = 1 for all g then HomG C (V1 , V2 )
admits a natural complex structure, namely O → iO.

Example: Let us consider the φ-twisted representations of M2 = {1, T̄ } where φ(T̄ ) = −1.
We showed above that there are precisely two φ-twisted extensions M2± . First, let us
suppose H is a φ-rep of M2+ . Then set

K = ρ(T̃ ). (1.133)

This operator is anti-linear and squares to +1. Therefore K is a real structure on H. On

the other hand, if the φ-twisted extension of M2 is M2− then K 2 = −1. Therefore we
have a quaternionic structure on H. Thus we conclude: The φ-twisted representations
of (M2 , φ), with φ(T̄ ) = −1 are the complex vector spaces with a real structure (for M2+ )
union the complex vector spaces with a quaternionic structure (for M2− ).

1.7.2 Schur’s Lemma for φ-reps

While many of the standard notions and constructions of representation theory carry over
straightforwardly to the theory of φ-reps, sometimes they come with very interesting new
twists. A good example of this is Schur’s lemma.
One very important fact for us below will be the analog of Schur’s lemma. To state it
correctly we recall a basic definition:
Definition An associative division algebra over a field κ is an associative unital algebra
A over κ such that for every nonzero a ∈ A there is a multiplicative inverse a−1 ∈ A, i.e.
aa−1 = a−1 a = 1.

– 27 –
 Then we have ♣Use V1 and V2
rather than V and
Theorem [Schur’s Lemma]. V ′ for consistency
of notation. ♣
a.) If A is an intertwiner between two irreducible φ-reps (ρ, V ) and (ρ′ , V ′ ) then either
A = 0 or A is an isomorphism.
b.) Suppose (ρ, V ) is an irreducible φ-representation of (G, φ). Then the commutant,
that is, the set of all intertwiners A of (ρ, V ) with itself:

Z(ρ, V ) := HomG
C (V, V ) = {A ∈ EndC (V )|∀g ∈ G Aρ(g) = ρ(g)A} (1.134)

is a real associative division algebra.

Schur’s lemma for φ-representations naturally raises the question of finding examples
of real division algebras. In fact, there are only three. This is the very beautiful theorem
of Frobenius:

Theorem: If A is a finite dimensional 6 real associative division algebra then one of three
possibilities holds:

• A∼
=R

• A∼
=C

• A∼
=H

Examples

1. Let G = M2+ with φ(T̄ ) = −1. Take V = C, ρ(T̃ ) = C ∈ EndR (C) given by complex
conjugation C(z) = z̄. Then Z(ρ, V ) = R.

2. Let G = U (1) with φ = 1, so the grading is trivial (all even). Let V = C and
ρ(z)v = zv. Then Z(ρ, V ) = C. Notice we could replace G with any subgroup of
multiplicative nth roots of 1 in this example, so long as n > 2.

3. Let G = M2− , with φ(T̃ ) = −1. Take V = C2 and represent

! !
iθ z1 eiθ z1
ρ(e ) = (1.135)
z2 eiθ z2
! !
z1 −z̄2
ρ(T̃ ) = (1.136)
z2 z̄1
One checks that these indeed define a φ-representation of M2− . We claim that in this
case Z(ρ, V ) ∼
= H.
To prove the claim let us map C2 → H by
!
z1
7→ z1 + z2 j = (x1 + iy1 ) + (x2 + iy2 )j (1.137)
z2
6
The Gelfand-Mazur theorem asserts that any unital Banach algebra over R is R, C or H.

– 28 –
 Thus, if we think of the φ-twisted representation as acting on the quaternions then
we have:
ρ(eiθ ) = cos θ + sin θL(i) (1.138)
and T is represented by
ρ(T̃ ) = L(j) (1.139)
Of course, the commutant algebra Z(ρ, V ) must commute with the real algebra gen-
erated by all the elements ρ(g). Therefore, it must commute with left-multiplication ♣Perhaps put this
general remark
by arbitrary quaternions. From this it easily follows that Z(ρ, V ) is the algebra of earlier. ♣

right-multiplication by arbitrary quaternions. In fact this identifies Z(ρ, V ) ∼

= Hopp
but as in an exercise above Hopp ∼= H as a real algebra.
so the algebra of operators generated by ρ(g) is the quaternion algebra acting on
V = R4 as the left regular representation. The commutant of these operations is
therefore right-multiplication by any quaterion R(q), and hence Z ∼
= H.

1.7.3 Complete Reducibility

A very important theorem in ordinary representation theory is the complete reducibility
of representations of compact groups. This extends more or less directly to φ-reps.
If (G, φ) is a Z2 -graded group then (G, φ)∨ , known as the “dual,” is the set of in-
equivalent irreducible φ-representations of G. For each element of λ ∈ (G, φ)∨ we select a
representative irrep Vλ . Thanks to Schur’s lemma it is unique up to isomorphism.
Theorem: If (ρ, V ) is a finite-dimensional φ-unitary rep of (G, φ) then V is isomorphic
to a representation of the form
⊕λ∈(G,φ)∨ Wλ (1.140)
where, for each λ, Wλ is itself (noncanonically) isomorphic to a direct sum of representations
Vλ :
Wλ ∼ = HomG ∼ λ ⊕ · · · ⊕ Vλ
C (Vλ , V ) ⊗Dλ Vλ = V (1.141)
| {z }
sλ times

(If sλ = 0 this is the zero vector space. The second isomorphism is noncanonical.)

Proof : The proof proceeds by choosing a sub-rep and observing that the orthogonal com-
plement is again a subrep. ♦

Remarks

1. The isomorphism of a representation with (1.140), (1.141) is known as an isotypical

decomposition. The nonnegative integers sλ are known as degeneracies. ♣Drop this? ♣

2. Concretely the theorem means that we can choose a “block-diagonal” basis for V so
that relative to this basis the matrix representation of ρ(g) has the form
 
..
.
 
ρ(g) ∼ 
 1sλ ⊗ ρλ (g) 
 (1.142)
..
.

– 29 –
 We need to be careful about how to interpret ρλ (g) because anti-linear operators
don’t have a matrix representation over the complex numbers. If we are working
with ordinary representations over C and dimC Vλ = tλ then 1sλ ⊗ ρλ (g) means a
matrix of the form
 
ρλ (g) 0 0 · · · 0
 
 0 ρλ (g) 0 · · · 0 

1sλ ⊗ ρλ (g) =  . .. .. ..  (1.143)

 .. . . ··· . 
0 0 0 · · · ρλ (g)

where ρλ (g) and each of the 0’s above is a tλ × tλ matrix and there is an sλ × sλ
matrix of such blocks. On the other hand, if ρ(g) is anti-linear then it does not have
a matrix representation over the complex numbers. If we wish to work with matrix
representations what we must do is work with (VR , I) where I is a complex structure
on VR , and similarly for the irreps (Vλ,R , Iλ ). Then ρλ (g) means a real representation
matrix which is 2tλ × 2tλ and anticommutes with Iλ . See the beginning of §1.7.4 for
a specific way to do this.

3. Now if we combine the isotypical decomposition with the second part of Schur’s
lemma to compute the real algebra of self-endomorphisms EndG C (V ). To lighten the
G
notation let Sλ := HomC (Vλ , V ) and let Dλ be the real division algebra over R of
self-intertwiners of Vλ . (That is Dλ := Z(ρλ , Vλ ).) Then we compute:

HomC (V, V ) ∼
= V ∗ ⊗C V
∼
= ⊕λ,λ′ (S ∗ ⊗R Sλ′ ) ⊗R (V ∗ ⊗C Vλ′ ) (1.144)
λ λ
∼
= ⊕λ,λ′ Hom(Sλ , Sλ′ ) ⊗R HomC (Vλ , Vλ′ )

Now G acts trivially on the Hom(Sλ , Sλ′ ) factors and in the natural way on HomC (Vλ , Vλ′ ).
Therefore, taking the G-invariant part to get the intertwiners we invoke Schur’s
lemma
HomGC (Vλ , Vλ′ ) = δλ,λ′ Dλ (1.145)
and hence
HomG ∼
C (V, V ) = ⊕λ End(Sλ ) ⊗R Dλ (1.146)
Of course, End(Sλ ) is isomorphic to the algebra of real matrices M atsλ (R) upon
choosing a basis and therefore

EndG ∼
C (V ) = ⊕λ Matsλ (Dλ ) (1.147)

is a direct sum of matrix algebras over real division algebras.

♣Drop this
subsection on
algebras? ♣
1.7.4 Complete Reducibility in terms of algebras
The complete reducibility and commutant subalgebra can also be expressed nicely in terms
of the group algebra R[G]. We work with VR with complex structure I with operators

– 30 –
ρR (g) commuting or anticommuting with I according to φ(g). This defines a subalgebra of
End(VR ). If G is compact this algebra can be shown to be semisimple and therefore, by a
theorem of Wedderburn all representations are matrix representations by matrices over a
division algebra over R. See Appendix ?? for background on semisimple algebras.
It is useful to be explicit and make a choice of basis. Therefore, we choose a basis to
identify V ∼= CN . Then we identify VR ∼ = R2N by mapping each coordinate
!
x
z→ (1.148)
y

The complex structure on R2N is therefore

! !
0 −1 0 −1
I0 = ⊕ ··· ⊕ (1.149)
1 0 1 0

While the real structure of conjugation with respect to this basis is the operation
! !
1 0 1 0
C= ⊕ ··· ⊕ (1.150)
0 −1 0 −1

Having chosen a basis V ∼

= CN the C-linear operators ρ(g) with φ(g) = 1 can identified
with N × N complex matrices and then they are promoted to 2N × 2N real matrices by
replacing each complex matrix element by
!
xij −yij
zij → (1.151)
yij xij

The operators with φ(g) = −1 must be represented by C times a matrix of the above type.
Now we want to describe the algebra ρ(G) over R generated by the real 2N × 2N
matrices ρ(g) together with I0 . To do this let us introduce some notation: If K is any
algebra then we let K[m] or K(m) denote the algebra of all m×m matrices whose elements
are in K. The notation mK will denote the subalgebra of m × m matrices over K of the
specific form Diag{k, k, . . . , k}. Thus mK and K are isomorphic as abstract algebras. Note
that m(K[n]) and (mK)[n] are canonically isomorphic so we just write mK[n] when we
combine the two constructions. Finally, with this notation we can state the:

Theorem The algebra A(ρ(G), I) ⊂ End(VR ) generated over R by the operators ρ(g) and
I is equivalent to
A(ρ(G), I) ∼
= ⊕λ sλ Dλ [τλ ] (1.152)
and the commutant Z(ρ, V ) is equivalent to
opp
Z(ρ, V ) ∼
= ⊕λ τλ Dλ [sλ ] (1.153)

Note that the dimensions τλ are slightly different from the complex dimensions tλ
of Vλ in general. Let us denote the real dimension of Dλ by dλ = 1, 2, 4 according to
Dλ = R, C, H. Then
2
τλ = tλ (1.154)
dλ

– 31 –
Recall that when Dλ ∼ = H there must be an action of H on V and hence tλ must be even,
so τλ is always an integer, as it must be.

Definition: A real semisimple algebra has the form (1.152). We say that it has Wedderburn
type {Dλ }, that is, we have an unordered list of the real division algebras determining the
simple summands.

1.8 Groups Compatible With Quantum Dynamics

With the possible exception of exotic situations in which quantum gravity is important,
physics takes place in space and time. Except in unusual situations associated with nontriv-
ial gravitational fields we can assume our spacetime is time-orientable. Then, any physical
symmetry group G must be equipped with a homomorphism

τ : G → Z2 (1.155)

telling us whether the symmetry operations preserve or reverse the orientation of time.
That is τ (g) = +1 are symmetries which preserve the orientation of time while τ (g) = −1
are symmetries which reverse it.
Now, suppose that G is a symmetry of a quantum system. Then Wigner’s theorem
gives G another grading φ : G → Z2 . Thus, on very general grounds, a symmetry of a
quantum system should be bigraded by a pair of homomorphisms (φ, τ ), or what is the
same, a homomorphism to Z2 × Z2 .
It is natural to ask whether φ and τ are related. A natural way to try to relate them
is to study the dynamical evolution.
In quantum mechanics time evolution is described by unitary evolution of states. That
is, there should be a family of unitary operators U (t1 , t2 ), strongly continuous in both
variables and satisfying composition laws U (t1 , t3 ) = U (t1 , t2 )U (t2 , t3 ) so that

ρ(t1 ) = U (t1 , t2 )ρ(t2 )U (t2 , t1 ) (1.156)

Let us - for simplicity - make the assumption that our physical system has time-translation
invariance so that U (t1 , t2 ) = U (t1 − t2 ) is a strongly continuous group of unitary transfor-
mations. 7
By Stone’s theorem, U (t) has a self-adjoint generator H, the Hamiltonian, so that we ♣There is an
obvious
may write   generalization of
this statement for
it
U (t) = exp − H (1.157) U (t1 , t2 ). Is it
~ proved rigorously
somewhere? ♣

Now, we say a quantum symmetry ρ : G → Autqtm (PH) lifting to ρtw : Gtw → AutR (H) is
a symmetry of the dynamics if for all g ∈ Gtw :

ρtw (g)U (t)ρtw (g)−1 = U (τ (g)t) (1.158)

7
In the more general case we would need an analog of Stone’s theorem to assert that there is a family
Rt
of self-adjoint operators with U (t1 , t2 ) = Pexp[− ~i t12 H(t′ )dt′ ]. Then, the argument we give below would
lead to ρtw (g)H(t)ρtw (g)−1 = φ(g)τ (g)H(t) for all t.

– 32 –
where τ : Gtw → Z2 is inherited from the analogous homomorphism on G.
Now, substituting (1.157) and paying proper attention to φ we learn that the condition
for a symmetry of the dynamics (1.158) is equivalent to

φ(g)ρtw (g)Hρtw (g)−1 = τ (g)H (1.159)

in other words,
ρtw (g)Hρtw (g)−1 = φ(g)τ (g)H (1.160)
Thus, the answer to our question is that φ and τ are unrelated in general. We should
therefore define a third homomorphism χ : G → Z2

χ(g) := φ(g)τ (g) ∈ {±1} (1.161)

Note that
φ·τ ·χ=1 (1.162)

Figure 4: If a symmetry operation has χ(g) = −1 then the spectrum of the Hamiltonian must be
symmetric around zero.

Remarks

1. We should stress that in general a system can have time-orientation reversing sym-
metries but the simple transformation t → −t is not a symmetry. Rather, it must be
accompanied by other transformations. Put differently, the exact sequence

1 → ker(τ ) → G → Z2 → 1 (1.163)

– 33 –
Figure 5: In this figure the blue crosses represent an atom with a local magnetic moment pointing
up while the red crosses represent an atom with a local magnetic moment pointing down. The
magnetic point group is isomorphic to D4 but the homomorphism τ to Z2 has a kernel Z2 × Z2
(generated by π rotation around a lattice point together with a reflection in a diagonal). Since D4
τ
is nonabelian the sequence 1 → Pb0 → Pb→Z2 → 1 plainly does not split.

in general does not split. Many authors assume it does, and that we can always
write G = G0 × Z2 where G0 is a group of time-orientation-preserving symmetries.
However, when considering, for example, the magnetic space groups the sequence
typically does not split. As a simple example consider a crystal





C = Z2 + (δ1 , δ2 ) ∐ Z2 + (−δ2 , δ1 ) ∐ Z2 + (−δ1 , −δ2 ) ∐ Z2 + (δ2 , −δ1 ) (1.164)

and suppose there is a dipole moment, or spin S on points in the sub-crystal

C+ = (Z + (δ1 , δ2 )) ∐ (Z + (−δ1 , −δ2 )) (1.165)

but a spin −S at the complementary sub-crystal

C− = (Z + (−δ2 , δ1 )) ∐ (Z + (δ2 , −δ1 )) (1.166)

such that reversal of time orientation exchanges S with −S. Then the time-orientation-
reversing symmetries must be accompanied by a π/2 or 3π/2 rotation around some
integer point or a reflection in some diagonal. See Figure 5. Therefore, the extension
of the point group is our friend:

1 → Z2 → Z4 → Z2 → 1 (1.167)

– 34 –
 which does not split.

2. It is very unusual to have a nontrivial homomorphism χ. Note that

ρtw (g)Hρtw (g)−1 = χ(g)H (1.168)

implies that if any group element has χ(g) = −1 then the spectrum of H must be
symmetric around zero as shown in Figure 4. In many problems, e.g. in the standard
Schrödinger problem with potentials which are bounded below, or in relativistic QFT
with H bounded below we must have χ(g) = 1 for all g and hence φ(g) = τ (g), which
is what one reads in virtually every physics textbook: A symmetry is anti-unitary iff
it reverses the orientation of time.

3. However, there are physical examples where χ(g) can be non-trivial, that is, there
can be symmetries which are both anti-unitary and time-orientation preserving. An
example are the so-called “particle-hole” symmetries in free fermion systems. We will
discuss those later.

4. It is often stated that the only physical systems with χ 6= 1 are free fermionic systems.
While free fermionic systems are a very natural set of examples, they are probably
not the only ones. Some other examples are 8

• Models on bipartite lattices: Ising and interacting fermions. ♣MORE DETAILS!

♣
• Photonic crystals [G. De Nittis and M. Lein] 9

• One can engineer phonon bandstructures with the desired E → −E symmetry

• The Hilbert-Polya Hamiltonian has the E → −E symmetry.

5. The transformations with χ(g) = −1 are sometimes called “charge-conjugation sym-

metries” and are sometimes called “particle-hole symmetries.” The CMT literature
is inconsistent about whether we should allow “symmetry groups” with χ 6= 1 and
about whether “particle-hole symmetry” should be a C-linear or a C-anti-linear op-
eration. So we have deliberately avoided using the term “particle-hole symmetry”
and “charge conjugation” associated with χ(g).

1.9 Dyson’s 3-fold way

Often in physics we begin with a Hamiltonian (or action) and then find the symmetries
of the physical system in question. However there are cases when the dynamics are very
complicated. A good example is in the theory of nuclear interactions. The basic idea
has been applied to many physical systems in which one can identify a set of quantum
states corresponding to a large but finite-dimensional Hilbert space. Wigner had the beau-
tiful idea that one could understand much about such a physical system by assuming the
Hamiltonian of the system is randomly selected from an ensemble of Hamiltonians with a
8
I thank Matt Hastings, Karin Rabe, and Shankar for useful discussions about this question
9
This claim has been strongly criticized and is controversial.

– 35 –
probability distribution on the ensemble. In particular one could still make useful predic-
tions of expected results based on averages over the ensemble.
So, suppose E is an ensemble of Hamiltonians with a probability measure dµ. Then
if O is some attribute of the Hamiltonians (such as the lowest eigenvalue, or the typical
eigenvalue spacing) then we might expect our complicated system to have the attribute O
close to the expectation value: Z
hOi := dµO. (1.169)
E
Of course, for this approach to be sensible there should be some natural or canonical
measure on the ensemble E, justified by some a priori physically reasonable principles. For
example, if we take the space of all Hermitian operators on some (say, finite-dimensional)
Hilbert space CN then any probability distribution which is

• Invariant under unitary transformation.

• Statistically independent for Hii and Re(Hij ) and Im(Hij ) for i < j

can be shown [34] to be of the form

N
Y Y 2 )+bTr(H)+c
dµ = dHii d2 Hij e−aTr(H (1.170)
i=1 i<j

The specific choice

N
1 Y Y N 2
dµ = dHii d2 Hij e− 2 TrH (1.171)
Z
i=1 i<j
R
where Z is a constant chosen so that dµ = 1 defines what is known as the Gaussian
unitary ensemble.
Now sometimes we know a priori that the system under study has a certain kind of
symmetry. Dyson pointed out in [20] that such symmetries can constrain the ensemble in
ways that affect the probability distribution dµ in important ways.

1.10 The Dyson problem

Now we can formulate the main problem which was addressed in [20]:

Given a Z2 -graded group (G, φ) and a φ-unitary rep (ρ, H), what is the ensemble of
commuting Hamiltonians? That is: What is the set of self-adjoint operators commuting
with ρ(g) for all g?

Note that the statement of the problem presumes that χ(g) = 1. In Section §2.5 below
we generalize the problem to allow for χ 6= 1.
The solution to Dyson’s problem follows readily from the machinery we have developed.
We assume that we can write the isotypical decomposition of H as

H∼
= ⊕ λ S λ ⊗ R Vλ (1.172)

– 36 –
This will always be correct if G is compact. Moreover, H is a Hilbert space and there are
Hermitian structures on Sλ and Vλ so that Vλ a φ-unitary rep and we have an isomorphism
of φ-unitary reps.
Now, if χ(g) = 1 then any Hamiltonian H on H must commute with the symmetry
operators ρ(g) and hence must be in EndG C (H). But we have computed this commutant
above. Choosing an ON basis for Sλ we have

Z(ρ, H) ∼
= ⊕λ Matsλ (Dλ ) (1.173)

The subset of matrices Matsλ (Dλ ) which are Hermitian is


 Dλ = R
Real
 symmetric
Hermsλ (Dλ ) = Complex Hermitian Dλ = C (1.174)


Quaternion Hermitian Dλ = H

where quaternion Hermitian means that the matrix elements Hij of H are quaternions and
Hij = Hji . (In particular, the diagonal elements are real.)
In conclusion, the answer to the Dyson problem is the ensemble:
Y
E= Hermsλ (Dλ ) (1.175)
λ

Each ensemble HermN (D) has a natural probability measure invariant under the uni-
tary groups 

O(N ; R)
 D=R
U (N, D) := U (N ) D=C (1.176)


Sp(N ) ∼ U Sp(2N ; C) D = H
=
such that the matrix elements (not related by symmetry) are statistically independent.
These are:

N
Y Y
1 N 2
dµGOE = dHii dHij e− 2σ2 TrH (1.177)
ZGOE
i=1 i<j

where H ∈ HermN (R) is real symmetric.

N
Y Y
1 N 2
dµGU E = dHii d2 Hij e− 2σ2 TrH (1.178)
ZGU E
i=1 i<j

where H ∈ HermN (C) is complex Hermitian.

N
Y Y
1 N 2
dµGSE = dHii d4 Hij e− 2σ2 TrH (1.179)
ZGSE
i=1 i<j

where H ∈ HermN (H) is quaternionic Hermitian.

– 37 –
 They lead to very different eigenvalue distributions
1 Y N X 2
dµ(λ1 , . . . , λN ) = |λi − λj |β exp[− 2 λi ] (1.180)
Z 2σ
i<j i

with β = 1, 2, 4. ♣CHECK!! ♣

2. Lecture 2: Phases of gapped systems: The 10-fold way

2.1 Gapped systems and the notion of phases

An active area of current 10 research in condensed matter theory is the “classification of
phases of matter.” There are physical systems, such as the quantum Hall states, “topolog-
ical insulators” and “topological superconductors” which are thought to be “topologically
distinct” from “ordinary phases of matter.” We put quotation marks around all these
phrases because they are never defined with any great precision, although it is quite clear
that precise definitions in principle must exist.
One way to define a “phase of matter” is to consider gapped systems.

Definition By a gapped system we mean a pair of a Hilbert space H with a self-adjoint

Hamiltonian H where 1/H exists as a bounded operator. (In particular, 0 is not in the
spectrum of the Hamiltonian.) ♣There was a lot of
push-back on this
Now suppose we have a continuous family of quantum systems. Defining this notion definition in the
talk. Use another
precisely is not completely trivial. See Appendix D of [26] for details. Roughly speaking, name? Discuss
relation to what
we have a family of Hilbert spaces Hs and Hamiltonians Hs varying continuously with other people mean?
♣
parameters s in some topological space S. 11
Suppose we are given a continuous family of quantum systems (Hs , Hs )s∈S which
is generically, but not always gapped. In general there will be a subspace D ⊂ S of
Hamiltonians for which 0 ∈ Spec(H) is a generically real codimension one subset of S. It
could be very complicated and very singular in places. Nevertheless, if it is real codimension
one then S − D will have well-defined connected components.

Definition Given a continuous family of quantum systems (Hs , Hs )s∈S we define a phase
of the system to be a connected component of S − D.
Another way to define the same thing is to say that two quantum systems (H0 , H0 ) and
(H1 , H1 ) are homotopic if there is a continuous family of systems (Hs , Hs ) interpolating
between them. 12 Phases are then homotopy classes of quantum systems in the set of all
gapped systems.
10
2007-2013
11
To be slightly more precise: We use the compact-open topology to define a bundle of Hilbert spaces
over S and we use this topology for the representations of topological groups. The map s → Hs should
be such that (t, s) → exp[−itHs ] is continuous from R × S → U (H)c.o. where we use the compact-open
topology on the unitary group.
12
Strictly speaking, we should allow for an isomorphism between the endpoint systems and the given
(H0 , H0 ) and (H1 , H1 ) so that homotopy is an equivalence relation on isomorphism classes of quantum
systems.

– 38 –
 PHASE1 PHASE2

Figure 6: A domain wall between two phases. The wavy line is meant to suggest a localized low
energy mode trapped on the domain wall.

Remark: A common construction in this subject is to consider a domain wall between

two phases as shown in Figure 6. The domain wall has a thickness and the Hamiltonian
is presumed to be sufficiently local that we can choose a transverse coordinate x to the
domain wall and the Hamiltonian for the local degrees of freedom is a family Hx . (Thus,
x serves both as a coordinate in space and as a parameter for a family of Hamiltonians.)
Then if the domain wall separates two phases by definition the Hamiltonian must fail to
be gapped for at least one value x = x0 within the domain wall. This suggests that there
will be massless degrees of freedom confined to the wall. That indeed happens in some nice
examples of domain walls between phases of gapped systems.
The focus of these notes is on the generalization of this classification idea to continuous
families of quantum systems with a symmetry. Thus we assume now that there is a group
G acting as a symmetry group of the quantum system: ρ : G → Autqtm (PH). As we have
seen that G is naturally Z2 -graded by a homomorphism φ, there is a φ-twisted extension
Gtw and a φ-representation of Gtw on H. Now, as we have also seen, if we have a symmetry
of the dynamics then there are also homomorphisms τ : Gtw → Z2 and χ : Gtw → Z2 with
φ(g)τ (g)χ(g) = 1. When we combine this with the assumption that H is gapped we see
that we can define a Z2 -grading on the Hilbert space given by the sign of the Hamiltonian.
That is, we can decompose:
H = H0 ⊕ H1 (2.1)
where H0 is the subspace on which H > 0 and H1 is the subspace on which H < 0. Put
differently, since H is gapped we can define Π = sign(H). Then Π2 = 1 and Π serves as the
grading operator defining the Z2 grading (2.1). From this viewpoint the equation (1.158),
written as
ρtw (g)H = χ(g)Hρtw (g) (2.2)
means that the operators ρtw (g) have a definite Z2 -grading: They are even if χ(g) = +1.
That means they preserve the sign of the energy and hence take H0 → H0 and H1 → H1
while they are odd if χ(g) = −1 and exchange H0 with H1 . See §2.2 below for a summary
of Z2 -graded linear algebra.

– 39 –
 This motivates the following definition:

Definition Suppose G is a bigraded group, that is, it has a homomorphism G → Z2 × Z2

or, what is the same thing, a pair of homomorphisms (φ, χ) from G to Z2 . Then we define
a (φ, χ)-representation of G to be a complex Z2 -graded vector space V = V 0 ⊕ V 1 and a
homomorphism ρ : G → End(VR ) such that
( (
C − linear φ(g) = +1 even χ(g) = +1
ρ(g) = and ρ(g) = (2.3)
C − anti − linear φ(g) = −1 odd χ(g) = −1

Figure 7: The blue regions in the top row represent different phases of a family of gapped Hamil-
tonians. The red regions in the bottom row represent different phases with a specified symmetry.
Some of the original phases might not have the symmetry at all. Some of the connected components
of the original phases might break up into several components with a fixed symmetry.

In terms of this concept we see that if G is a symmetry of a gapped quantum system

then there is a (φ, χ)-representation of Gtw . We can again speak of continuous families
of quantum systems with G-symmetry. This means that we have (Hs , Hs , ρs ) where the
representation ρs is a symmetry of the dynamics of Hs which also varies continuously with
s ∈ S. If we have a continuous family of gapped systems then we have a continuous family
of (φ, χ)-representations. Again we can define phases with G-symmetry to be the connected
components of S − D. This can lead to an interesting refinement of the classification of
phases without symmetry, as explained in Figure 7. We will denote the set of phases by

T P(Gtw , φ, χ, S) (2.4)

We have been led rather naturally to the notion of Z2 -graded linear algebra. Therefore
in the next section §2.2 we very briefly recall a few relevant facts and definitions.

Remarks

1. Except in quantum theories of gravity one is always free to add a constant to the
Hamiltonian of any closed quantum system. Typically, though not always, the con-
stant is chosen so that E = 0 lies between the ground state and the first excited state.

– 40 –
 For example, if we were studying the Schrodinger Hamiltonian for a single electron
p2 2
in the Hydrogen atom instead of the usual operator Ha = 2m − Zer we might choose
Ha + 12eV so that the groundstate would be at −1.6eV and the continuum would
begin at Ec = 12eV .

2. Again we stress that situations with χ 6= 1 are somewhat rare. However, they do not
just include the case of noninteracting fermions. Any time a Hamiltonian has some
set of energy levels in a range [E1 , E2 ] with a symmetry of the spectrum about some
value E0 ∈ [E1 , E2 ] (and E0 is not in the spectrum) there is an effective system with
χ 6= 1. ♣Explore the
possibilities of BEC.
Phonons ♣
3. In general T P(Gtw , φ, χ, S) is just a set. In some nice examples that set turns out
to be related to an abelian group which in turn ends up being a twisted equivariant
K-theory group. An example of how this refinement is relevant to condensed matter
physics is that in topological band structure we can consider families of one-electron
Hamiltonians which respect a given (magnetic) space-group. Then there is an inter-
esting refinement of the usual K-theoretic classification of band structures [26] which
will be discussed in Chapter 4.5.

2.2 Z2 -graded, or super-, linear algebra

In this section “super” is merely a synonym for “Z2 -graded.” Super linear algebra is
extremely useful in studying supersymmetry and supersymmetric quantum theories, but
its applications are much broader than that and the name is thus a little unfortunate.
Superlinear algebra is very similar to linear algebra, but there are some crucial differ-
ences: It’s all about signs.

2.2.1 Super vector spaces

It is often useful to add the structure of a Z2 -grading to a vector space. A Z2 -graded vector
space over a field κ is a vector space over κ which, moreover, is written as a direct sum

V = V 0 ⊕ V 1. (2.5)

The vector spaces V 0 , V 1 are called the even and the odd subspaces, respectively. We may
think of these as eigenspaces of a “parity operator” PV which satisfies PV2 = 1 and is +1
on V 0 and −1 on V 1 . If V 0 and V 1 are finite dimensional, of dimensions m, n respectively
we say the super-vector space has graded-dimension or superdimension (m|n).
A vector v ∈ V is called homogeneous if it is an eigenvector of PV . If v ∈ V 0 it is
called even and if v ∈ V 1 it is called odd. We may define a degree or parity of homogeneous
vectors by setting deg(v) = 0̄ if v is even and deg(v) = 1̄ if v is odd. Here we regard 0̄, 1̄
in the additive abelian group Z/2Z = {0̄, 1̄}. Note that if v, v ′ are homogeneous vectors of
the same degree then
deg(αv + βv ′ ) = deg(v) = deg(v ′ ) (2.6)
for all α, β ∈ κ. We can also say that PV v = (−1)deg(v) v acting on homogeneous vectors.
For brevity we will also use the notation |v| := deg(v). Note that deg(v) is not defined for
general vectors in V .

– 41 –
 Mathematicians define the category of super vector spaces so that a morphism from
V → W is a linear transformation which preserves grading. We will denote the space of
morphisms from V to W by Hom(V, W ). The underline is there to distinguish from the
space of linear transformations from V to W discussed below. The space of morphisms
Hom(V, W ) is just the set of ungraded linear transformations of ungraded vector spaces,
T : V → W , which commute with the parity operator T PV = PW T . One reason for this
definition is that only then can we say the super-dimension

sdimV := (dimV 0 |dimV 1 ) (2.7)

is an isomorphism invariant.
If κ is any field we let κp|q denote the supervector space over κ of superdimension (p|q).
That is:
κp|q = |{z}
κp ⊕ |{z}
κq (2.8)
even odd
So far, there is no big difference from, say, a Z-graded vector space. However, important
differences arise when we consider tensor products.
Put differently: we defined a category of supervector spaces, and now we will make it
into a tensor category. (See definition below.)
The tensor product of two Z2 graded spaces V and W is V ⊗ W as vector spaces over
κ, but the Z2 -grading is defined by the rule:

0
b
V ⊗W := V 0 ⊗ W 0 ⊕ V 1 ⊗ W 1

1 (2.9)
b
V ⊗W := V 1 ⊗ W 0 ⊕ V 0 ⊗ W 1

Thus, under tensor product the degree is additive on homogeneous vectors:

deg(v ⊗ w) = deg(v) + deg(w) (2.10)

Thus, for example:
′ ′ ′ ′ ′ ′
b ne |no ∼
Rne |no ⊗R = Rne ne +no no |ne no +no ne (2.11)

For example:
R8|8 ⊗R
b 8|8 = R128|128 (2.12)
Now, in fact we have a braided tensor category:
In ordinary linear algebra there is an isomorphism of tensor products

b → W ⊗V
cV,W : V ⊗W b (2.13)

b 7→ w⊗v.
given by cV,W : v ⊗w b In the category of super vector spaces there is also an
isomorphism (2.13) defined by taking

b → (−1)|v|·|w| w⊗v
cV,W : v ⊗w b (2.14)

on homogeneous objects, and extending by linearity.

Let us pause to make two remarks:

– 42 –
 1. Note that in (2.14) we are now viewing Z/2Z as a ring, not just as an abelian
group. Do not confuse degv + degw with degvdegw! In computer science language
degv + degw corresponds to XOR, while degvdegw corresponds to AND.

2. It is useful to make a general rule: In equations where the degree appears it is

understood that all quantities are homogeneous. Then we extend the formula to
general elements by linearity. Equation (2.14) is our first example of another general
rule: In the super world, commuting any object of homogeneous degree A with any
object of homogeneous degree B results in an “extra” sign (−1)AB . This is sometimes
called the “Koszul sign rule.”

Remark: (Reversal of parity)

Introduce an operation which switches the parity of a supervector space: (ΠV )0 = V 1
and (ΠV )1 = V 0 . In the category of finite-dimensional supervector spaces V and ΠV
isomorphic iff n = m.

2.2.2 Linear transformations between supervector spaces

If the ground field κ is taken to have degree 0 then the dual space V ∨ in the category of
supervector spaces is the supervector space:

(V ∨ )0 := (V 0 )∨
(2.15)
(V ∨ )1 := (V 1 )∨

Thus, we can say that (V ∨ )ǫ are the linear functionals V → κ which vanish on V 1+ǫ .
Taking our cue from the natural isomorphism in the ungraded theory:

Hom(V, W ) ∼
= V ∨ ⊗W
b (2.16)

we use the same definition so that the space of linear transformations between two Z2 -
graded spaces becomes Z2 graded. We also write End(V ) = Hom(V, V ).
In particular, a linear transformation is an even linear transformation between two
Z2 -graded spaces iff T : V 0 → W 0 and V 1 → W 1 , and it is odd iff T : V 0 → W 1 and
V 1 → W 0 . Put differently:

Hom(V, W )0 ∼
= Hom(V 0 , W 0 ) ⊕ Hom(V 1 , W 1 )
(2.17)
Hom(V, W )1 ∼
= Hom(V 0 , W 1 ) ⊕ Hom(V 1 , W 0 )

The general linear transformation is neither even nor odd.

If we choose a basis for V made of vectors of homogeneous degree then with respect
to such a basis even transformations have block diagonal form
!
A 0
T = (2.18)
0 D
while odd transformations have block diagonal form

– 43 –
 !
0 B
T = (2.19)
C 0

As in the ungraded case, End(V ) is a ring, but now it is a Z2 -graded ring un-
der composition: T1 T2 := T1 ◦ T2 . That is if T1 , T2 ∈ End(V ) are homogeneous then
deg(T1 T2 ) = deg(T1 ) + deg(T2 ), as one can easily check using the above block matrices.
These operators are said to graded-commute, or supercommute if

T1 T2 = (−1)degT1 degT2 T2 T1 (2.20)

2.2.3 Superalgebras
The set of linear transformations End(V ) of a supervector space is an example of a super-
algebra. In general we have:
Definition
a.) A superalgebra A is a supervector space over a field κ together with a morphism

b →A
A⊗A (2.21)

of supervector spaces. 13 b ′ 7→ aa′ . Note this implies that

We denote the product as a⊗a

deg(aa′ ) = deg(a) + deg(a′ ). (2.22)

We assume our superalgebras to be unital so there is a 1A with 1A a = a1A = a. Henceforth

we simply write 1 for 1A .
b.) Two elements a, a′ in a superalgebra are said to graded-commute, or super-commute
provided
′
aa′ = (−1)|a||a | a′ a (2.23)

If every pair of elements a, a′ in a superalgebra graded-commmute then the superalgebra

is called graded-commutative or supercommutative.
c.) The supercenter, or Z2 -graded center of an algebra, denoted Zs (A), is the subsu-
peralgebra of A such that all homogeneous elements a ∈ Zs (A) satisfy

ab = (−1)|a||b| ba (2.24)

for all homogeneous b ∈ A.

Example 1: Matrix superalgebras. If V is a supervector space then End(V ) as described

above is a matrix superalgebra. As an exercise, show that the supercenter is isomorphic
to κ, consisting of the transformations v → αv, for α ∈ κ. So in this case the center and
super-center coincide.
13 b as a tensor product of supervector spaces, not yet a tensor product of
Warning!! We are here using ⊗
superalgebras!

– 44 –
 Definition Let A and B be two superalgebras. The graded tensor product A⊗Bb is the
superalgebra which is the graded tensor product as a vector space and the multiplication
of homogeneous elements satisfies

(a1 ⊗b b 2 ) = (−1)|b1 ||a2 | (a1 a2 )⊗(b

b 1 ) · (a2 ⊗b b 1 b2 ) (2.25)

b
Example For matrix superalgebras we have End(V )⊗End(V ′ ) ∼ End(V ⊗ V ′ ), and in
=
particular:
′
ne |no ∼′ ′ ′ ′ ′ ′ ′
End(Cne |no )⊗End(C
b b ne |no ) ∼
) = End(Cne |no ⊗C = End(Cne ne +no no |ne no +no ne ) (2.26)

Remarks

1. Every Z2 -graded algebra is also an ungraded algebra: We just forget the grading.
However this can lead to some confusions:

2. An algebra can be Z2 -graded-commutative and not ungraded-commutative: The

Grassmann algebras are an example of that. We can also have algebras which are
ungraded commutative but not Z2 -graded commutative. The Clifford algebras Cℓ±1
described below provide examples of that.

3. The Z2 -graded-center of an algebra can be different from the center of an algebra

as an ungraded algebra. Again, the Clifford algebras Cℓ±1 described below provide
examples.

4. One implication of (2.25) is that when writing matrix representations of graded al-
gebras we do not get a matrix representation of the graded tensor product just by
taking the tensor product of the matrix representations. This is important when
discussing reps of Clifford algebras, as we will stress below.

5. As for ungraded algebras, there is a notion of simple, semi-simple, and central super-
algebras. These are discussed in Appendix A of [43]. A key fact is that the Clifford
algebras over κ = R, C are all central simple superalgebras over κ = R, C respectively.

2.2.4 Modules over superalgebras

Definition A super-module M over a super-algebra A (where A is itself a superalgebra

over a field κ) is a supervector space M over κ together with a κ-linear map A × M → M
defining a left-action or a right-action. That is, it is a left-module if, denoting the map by
L : A × M → M we have
L(a, L(b, m)) = L(ab, m) (2.27)
and it is a right-module if, denoting the map by R : A × M → M we have

R(a, R(b, m)) = R(ba, m) (2.28)

– 45 –
In either case:
deg(R(a, m)) = deg(L(a, m)) = deg(a) + deg(m) (2.29)
The notations L(a, m) and R(a, m) are somewhat cumbersome and instead we write
L(a, m) = am and R(a, m) = ma so that (ab)m = a(bm) and m(ab) = (ma)b. We also
sometimes refer to a super-module over a super-algebra A just as a representation of A.

Definition A linear transformation between two super-modules M, N over A is a κ-linear

transformation of supervector spaces such that if T is homogeneous and M is a left A-
module then T (am) = (−1)|T ||a| aT (m) while if M is a right A-module then T (ma) =
T (m)a. We denote the space of such linear transformations by HomA (M, N ). If N is a left
A-module then HomA (M, N ) is a left A-module with (a · T )(m) := a · (T (m)). If N is a
right A-module then HomA (M, N ) is a right A-module with (T ·a)(m) := (−1)|a||m| T (m)a.
When M = N we denote the module of linear transformations by EndA (M ).
Just as in the case of supervector spaces, we must be careful about the definition of a
morphism:

Definition A morphism in the category of A-modules is a morphism T of supervector

spaces which commutes with the A-action.

2.2.5 Star-structures and super-Hilbert spaces

There are at least three notions of a real structure on a complex superalgebra which one
will encounter in the literature: ♣Use a symbol for
structure 2 which is
more easily
1. It is a C-antilinear involutive automorphism a 7→ a⋆ . Hence deg(a⋆ ) = deg(a) and distinguished from
structure 3. Star
(ab)⋆ = a⋆ b⋆ . with circle? ♣

2. It is a C-antilinear involutive anti-automorphism. Thus deg(a∗ ) = deg(a) but

(ab)∗ = (−1)|a||b| b∗ a∗ (2.30)

3. It is a C-antilinear involutive anti-automorphism. Thus deg(a⋆ ) = deg(a) but

(ab)⋆ = b⋆ a⋆ (2.31)

If A is a supercommutative complex superalgebra then structures 1 and 2 coincide:

a → a⋆ is the same as a → a∗ . See remarks below for the relation of 2 and 3.

Definition A sesquilinear form h on a complex supervector space H is a map h : H×H → C

such that
1. It is even, so that h(v, w) = 0 if v and w have opposite parity
2. It is C-linear in the second variable and C-antilinear in the first variable
3. An Hermitian form on a supervector space is a sesquilinear form which moreover
satisfies the symmetry property:

(h(v, w))∗ = (−1)|v||w| h(w, v) (2.32)

– 46 –
 4. If in addition for all nonzero v ∈ H0

h(v, v) > 0 (2.33)

while for all nonzero v ∈ H1

i−1 h(v, v) > 0, (2.34)

then h is super-positive definite. If every Cauchy sequence in H converges to an element

of H it is a super-Hilbert space.

For bounded operators we define the adjoint of a homogeneous linear operator T :

H → H by
h(T ∗ v, w) = (−1)|T ||v| h(v, T w) (2.35)

The spectral theorem is essentially the same as in the ungraded case with one strange
modification. For even Hermitian operators the spectrum is real. However, for odd Hermi-
tian operators the point spectrum sits in a real subspace of the complex plane which is not
√
the real line! It lies on the line passing through eiπ/4 = (1 + i)/ 2 in the complex plane,
as shown in Figure 8

Figure 8: When the Koszul rule is consistently implemented odd super-Hermitian operators have
a spectrum which lies along the line through the origin which runs through 1 + i.

Remark:
In general star-structures 2 and 3 above are actually closely related. Indeed, given a
structure a → a∗ of type 2 we can define a structure of type 3 by defining either
(
⋆ a∗ |a| = 0
a = (2.36)
ia∗ |a| = 1

or (
⋆ a∗ |a| = 0
a = (2.37)
−ia∗ |a| = 1

– 47 –
It is very unfortunate that in most of the physics literature the definition of a star structure
is that used in item 3 above. For example a typical formula used in manipulations in
superspace is
θ1 θ2 = θ̄2 θ̄1 (2.38)
and the fermion kinetic energy Z
d
dtiψ̄ ψ (2.39)
dt
is only “real” with the third convention.

2.3 Clifford Algebras and Their Modules

2.3.1 The real and complex Clifford algebras
Clifford algebras are defined for a general nondegenerate symmetric quadratic form Q on a
vector space V over κ. They are officially defined as a quotient of the tensor algebra of V by
the ideal generated by the set of elements of T V of the form v1 ⊗ v2 + v2 ⊗ v1 − 2Q(v1 , v2 ) · 1
for any v1 , v2 ∈ V . A more intuitive definition is that Cℓ(Q) is the Z2 graded algebra
over κ which has a set of odd generators {ei } in one-one correspondence with a basis, also
denoted {ei }, for the vector space V . The only relations on the generators are given by ♣Poor choice of
notation since eij is
an element of the
{ei , ej } = 2Qij · 1 (2.40) Clifford algebra but
also our notation for
matrix units... ♣
where Qij ∈ κ is the matrix of Q with respect to a basis {ei } of V , and 1 ∈ Cℓ(Q) is
the multiplicative identity. Henceforth we will usually identify κ with κ · 1 and drop the
explicit 1.
Because ei are odd and 1 is even, the algebra Cℓ(Q) does not admit a Z-grading.
However, every expression in the relations on the generators is even so the algebra admits
a Z2 grading:
Cℓ(Q) = Cℓ(Q)0 ⊕ Cℓ(Q)1 (2.41)
Of course, one is always free to regard Cℓ(Q) as an ordinary ungraded algebra, and this is
what is done in much of the physics literature. However, as we will show below, comparing
the graded and ungraded algebras leads to a lot of insight.
Suppose we can choose a basis {ei } for V so that Qij is diagonal. Then e2i = qi 6= 0. It
follows that Cℓ(Q) is not supercommutative, because an odd element must square to zero
in a supercommutative algebra. Henceforth we assume Qij has been diagonalized, so that
ei anticommutes with ej for i 6= j. Thus, we have the basic Clifford relations:

ei ej + ej ei = 2qi δij (2.42)

When {i1 , . . . , ip } are all distinct is useful to define the notation

ei1 ···ip := ei1 · · · eip (2.43)

Of course, this expression is totally antisymmetric in the indices, and a moment’s thought
shows that it forms a basis for Cℓ(Q) as a vector space and so we have
Cℓ(Q) ∼
= Λ∗ V
(2.44)
= Λev V ⊕ Λodd V

– 48 –
We stress that (2.44) is only an isomorphism of super-vector spaces. If V is finite-dimensional
with d = dimκ V then the graded dimension is easily seen to be

s − dimκ Cℓ(Q) = (2d−1 |2d−1 ) (2.45)

We must also stress that while the left and right hand sides (2.44) are both algebras over
κ the equation is completely false as an isomorphism of algebras. The right hand side of
(2.44) is a Grassmann algebra, which is supercommutative and as we have noted Cℓ(Q) is
not supercommutative.
If we take the case of a real vector space Rd then WLOG we can diagonalize Q to the
form !
+1r 0
Q= (2.46)
0 −1s
For such a quadratic form on a real vector space we denote the real Clifford algebra Cℓ(Q)
by Cℓr+,s−. We will see that Cℓ−s,r and Cℓ−r,s are in general inequivalent. ♣It is probably
better to use the
In general, one can show that Cℓ(Q)opp ∼= Cℓ(−Q). So they are opposite superalgebras. notation Cℓr,−s
where r, s are
We can similarly discuss the complex Clifford algebras Cℓn . Note that over the complex always understood
to be nonnegative
numbers if e2 = +1 then (ie)2 = −1 so we do not need to account for the signature, and integers. ♣

WLOG we can just consider Cℓn for n ≥ 0.

The even subalgebra

The even subalgebra is an ungraded algebra and is isomorphic, as an ungraded algebra,

to another Clifford algebra. For example, if d ≥ 1 then

Cℓ0d ∼
= Cℓd−1 ungraded algebras. (2.47)

Relations by tensor products

One important advantage of regarding Cℓ(Q) as a superalgebra, rather than just an

algebra is that if Q1 ⊕ Q2 is a quadratic form on V1 ⊕ V2 then

Cℓ(Q1 ⊕ Q2 ) ∼ b
= Cℓ(Q1 )⊗Cℓ(Q 2) (2.48)

As we will see below, this is completely false if we regard the Clifford algebras as ungraded
algebras.
From (2.48) we have some useful identities: First, note that for n > 0:

Cℓn ∼ b · · · ⊗Cℓ
= Cℓ1 ⊗ b 1 (2.49)
| {z }
n times

Cℓ−n ∼ b · · · ⊗Cℓ
= Cℓ−1 ⊗ b −1 (2.50)
| {z }
n times
More generally we have
b · · · ⊗Cℓ
Cℓr+,s− = Cℓ1 ⊗ b 1⊗b Cℓ ⊗
b · · · ⊗Cℓ
b −1 (2.51)
| {z } | −1 {z }
r times s times

– 49 –
 Moreover
Cℓn ∼ b · · · ⊗Cℓ
= |Cℓ1 ⊗ b 1 (2.52)
{z }
n times

If we view the Clifford algebras as ungraded algebras then the tensor product relations
are a bit more complicated.
It is important to stress that when producing matrix representations we use the un-
graded tensor product, so that the structure of the algebras is more intricate than the
above simple identities would suggest.

The Clifford volume element

A key object in discussing the structure of Clifford algebras is the Clifford volume
element. When V is provided with an orientation this is the canonical element in Cℓ(Q)
defined by
ω := e1 · · · ed (2.53)

where d = dimκ V and e1 ∧ · · · ∧ ed is the orientation of V . Since there are two orientations
there are really two volume elements.
Note that:
Remarks

1. For d even, ω is even and anti-commutes with the generators ei ω = −ωei . Therefore
it is neither in the center nor in the ungraded center of Cℓ(Q). It is in the ungraded
center of the ungraded algebra Cℓ(Q)0 .

2. For d odd, ω is odd and ei ω = +ωei . Therefore it is in the ungraded center Z(Cℓ(Q))
but, because it is odd, it is not in the graded center Zs (Cℓ(Q)).

3. Thus, ω is never in the supercenter of Cℓ(Q). In fact, the super-center of Cℓr,s is R

and the super-center of Cℓd is C.

4. ω 2 is always ±1 (independent of the orientation). Here is the way to remember

the result: The sign only depends on the value of r+ − s− modulo 4. Therefore
we can reduce the question to Cℓn and the result only depends on n modulo four.
For n = 0mod4 the sign is clearly +1. For n = 2mod4 it is clearly −1, because
(e1 e2 )2 = −e21 e22 = −1 as long as e21 and e22 have the same sign. For Cℓ+1 and Cℓ−1
it is obviously +1 and −1, respectively.

2.3.2 Clifford algebras and modules over κ = C

Let us begin by considering the low-dimensional examples. We will contrast both the
graded and ungraded structures, to highlight the differences.
Of course Cℓ0 ∼ = C is purely even. Nevertheless, as a superalgebra it has two inequiv-
alent irreducible graded modules M0+ ∼ = C1|0 and M0− ∼ = C0|1 . As an ungraded algebra it
has one irreducible module - the regular representation N0 ∼ = C.

– 50 –
 On the other hand, it is easy to show that Cℓ1 has a unique graded irreducible module.
We can take it to be M1 ∼ = C1|1 with ρ(e) = σ 1 . Note that Cℓ1 cannot be a matrix
superalgebra.
By contrast, for Cℓ2 we can take two generators ρ(e1 ) = σ 1 and ρ(e2 ) = ±σ 2 . Now
we find that for either of these two irreps M2± the Clifford algebra is the full supermatrix
algebra End(C1|1 ).
What about dimensons n > 2? Now we can use tensor products to get the general
structure.
First, in the superalgebra case we can invoke (2.52). We know how to multiply matrix
superalgebras by (2.26). As we have stressed, Cℓ1 is not a matrix superalgebra, but Cℓ2
is. Therefore, since Cℓn+2 ∼ b 2 we have the key fact
= Cℓn ⊗Cℓ

Cℓn+2 ∼
= End(C1|1 )⊗Cℓ
b n (2.54)

Therefore, one can show inductively that ♣BLACKBOARD

TECHNIQUE!!
Don’t write twice.
k−1 k−1
Cℓ2k ∼
= End(C2 |2 ) (2.55) Build the table. ♣

and hence
k−1 k−1
Cℓ2k+1 ∼
= End(C2 |2 )⊗Cℓb 1 (2.56)

In both cases they are central simple superalgebras.

If we define (
ω n = 0, 1mod4
ωc := (2.57)
iω n = 2, 3mod4
then ωc2 = 1
The irreducible modules are:

Clifford Algebra Cℓ2k Cℓ2k+1

k−1 k−1 k−1 |2k−1
Graded algebra End(C2 |2 ) End(C 2 b
)⊗Cℓ 1
Ungraded algebra C(2k ) C(2k ) ⊕ C(2k )
k−1 k−1 k k
Graded irreps ±
M2k ∼
= C2 |2 , ρ(ωc )|M ±,0 = ±1 M2k+1 ∼
= C2 |2
2k
k k
Ungraded irreps N2k ∼
= C2 ±
N2k+1 ∼
= C2 , ρ(ωc ) = ±1

– 51 –
2.3.3 Morita equivalence of algebras
There is a very useful equivalence relation on (super)-algebras known as Morita equivalence.
The basic idea of Morita equivalence is that, to algebras A1 and A2 are Morita equiv-
alent if their “representation theory is the same.” More technically, if we consider the
categories ModL (Ai ) of left Ai -modules then the categories are equivalent.

Example 1: A1 = C and A2 = Mn (C) = C(n) are Morita equivalent ungraded algebras.

The general representation of A1 is a sum of n copies of its irrep C. So the general left
A1 -module M is isomorphic to
M∼ = |C ⊕ ·{z
· · ⊕ C} (2.58)
m times

for some positive integer m. On the other hand, the general representation N of A2 can
similarly be written
N∼ =C n
· · ⊕ Cn}
| ⊕ ·{z (2.59)
m times

again, for some positive integer m. Now, Cn is a left A2 -module, but is also a right A1 -
module. Then, if M is a general left A1 -module we can form Cn ⊗A1 M which is now a left
A2 -module. Conversely, given a left A2 -module N we can recover a left A1 -module from

M = HomA2 (Cn , N ) (2.60)

Example 2: More generally, if A is a unital algebra then A and Mn (A) are Morita
equivalent, by considerations similar to those above.

In more general terms, a criterion for Morita equivalence is based on the notion of
bimodules. An A1 − A2 bimodule E is a vector space which is simultaneously a left A1
module and a right A2 module (so that the actions of A1 and A2 therefore commute).
Now, given an A1 − A2 bimodule E we can define two functors:

F : ModL (A1 ) → ModL (A2 ) (2.61)

G : ModL (A2 ) → ModL (A1 ) (2.62)

as follows: For M ∈ ModL (A1 ) we define

F (M ) := HomA1 (E, M ) (2.63)

Note that this is in fact a left A2 module. To see that suppose that a ∈ A2 and T : E → M
commutes with the left A1 -action. Then we define (a · T )(p) := T (pa) for p ∈ E. Then we
compute

(a1 · (a2 · T ))(p) = (a2 · T )(pa1 )

= T (pa1 a2 ) (2.64)
= ((a1 a2 ) · T )(p)

– 52 –
On the other hand, given a left A2 -module N we can define a left A1 -module by

G(N ) = E ⊗A2 N (2.65)

For Morita equivalence we would like F, G to define equivalences of categories so there

must be natural identifications of


M∼
= E ⊗A2 HomA1 (E, M ) ∼
= E ⊗A2 E ∨ ⊗A1 M (2.66)


= HomA1 (E, E ⊗A2 N ) ∼
N∼ = E ∨ ⊗A1 E ⊗A2 N (2.67)
Therefore, for Morita equivalence E must be invertible in the sense that there is an A2 − A1
bimodule E ∨ with
E ⊗A2 E ∨ ∼
= A1 (2.68)
as A1 − A1 bimodules together with

E ∨ ⊗A1 E ∼
= A2 (2.69)

as A2 − A2 bimodules. In fact we can recover one algebra from the other

A2 ∼
= EndA1 (E) (2.70)

A1 ∼
= EndA2 (E) (2.71)
and within the algebra of κ-linear transformations End(E) we have that A1 and A2 are
each others commutant: A′1 = A2 .
Moreover, E determines E ∨ by saying

E∨ ∼
= HomA1 (E, A1 ) as left A2 module, (2.72)

E∨ ∼
= HomA2 (E, A2 ) as right A1 module. (2.73)
Another useful characterization of Morita equivalent algebras is that there exists a full
idempotent 14 e ∈ A1 and a positive integer n so that

A2 ∼
= eMn (A1 )e. (2.74)

Example A1 = Mn (κ) is Morita equivalent to A2 = Mm (κ) by the bimodule E of all

n × m matrices over κ. Indeed, one easily checks that

E ⊗A2 E ∨ ∼
= A1 (2.75)

(Exercise: Explain why the dimensions match.) and

E ∨ ⊗A2 E ∼
= A2 (2.76)

Similarly, we can check the other identities above.

14
If R is a ring then an idempotent e ∈ R satisfies e2 = e. It is a full idempotent if ReR = R.

– 53 –
Remark: One reason Morita equivalence is important is that many aspects of representa-
tion theory are “the same.” In particular, one approach to K-theory emphasizes algebras.
Roughly speaking, K0 (A) is defined to be the Grothendieck group 15 or group completion of
the monoid of finite-dimensional projective left A-modules. The K-theories of two Morita
equivalent algebras are isomorphic abelian groups.

The above discussion generalizes to superalgebras. A sufficient condition (and the only
condition we will need) for two superalgebras A1 and A2 to be Morita equivalent is that
there is a matrix superalgebra End(V ) such that

A1 ∼ b
= A2 ⊗End(V ) (2.77)

or the other way around.

Tensor product induces a multiplication structure on Morita equivalence classes of
(super) algebras.
b κ B]
[A] · [B] := [A⊗ (2.78)
If we take the algebra consisting of the ground field κ itself then we have an identity element
[κ] · [A] = [A] for all algebras over κ.
If, moreover, A is central simple then there is an isomorphism

b opp ∼
A⊗A = Endκ (A) (2.79)

where on the RHS we mean the algebra of linear transformations of A as a κ vector space.
In fact, this property is equivalent to A being central simple. For the real Clifford algebras,

b −1 ∼
Cℓ+1 ⊗Cℓ = End(R1|1 ) (2.80)

so by this criterion all the Cℓn are central simple superalgebras. Since A is assumed finite
dimensional the RHS is isomorphic to a matrix (super-)algebra over κ and hence Morita
equivalent to κ itself. Therefore the above product defines a group operation and not just
a monoid. If we speak of ordinary algebras then this group is known as the Brauer group
of κ, and if we speak of superalgebras we get the graded Brauer group of κ.

2.3.4 Morita equivalence of Clifford algebras and the complex K-theory of a

point
Equation (2.54) shows that the Morita equivalence classes of complex Clifford algebras
have a mod two periodicity:
[Cℓn+2 ] = [Cℓn ] (2.81)
As explained above there is a group structure on Morita equivalence classes

b m ] = [Cℓn+m ] = [Cℓ(n+m)mod2 ]
[Cℓn ] · [Cℓm ] := [Cℓn ⊗Cℓ (2.82)

Therefore, the graded Brauer group of C is the group Z2 . ♣Need to

distinguish Morita
equivalence classes
15
See Remark 1 at the end of this section from notation for
K-theory classes
below. ♣

– 54 –
 At this point we are at the threshhold of the subject of K-theory. This is a generaliza-
tion of the cohomology groups of topological spaces. At this point we are only equipped to
discuss the “cohomology groups” of a point, but even this involves some interesting ideas.

Definition
a.) Mcn is the abelian monoid of finite-dimensional complex graded modules for Cℓn .
The monoid operation is direct sum and the identity is the 0 vector space.
b.) Mc,triv
n is the submonoid of modules M that admit an invertible odd operator
T ∈ End(M ) which graded-commutes with the Cℓn -action. 16
c.) We define
K n (pt) := Mcn /Mc,triv
n . (2.83)
Note that since Cℓn = Cℓ−n it follows that K n (pt) = K −n (pt) so this makes sense for
n ∈ Z.
There is a well-defined sum on equivalence classes:

[M1 ] ⊕ [M2 ] := [M1 ⊕ M2 ] (2.84)

and in the quotient monoid there are additive inverses. The reason is that

[M ] ⊕ [ΠM ] = [M ⊕ ΠM ] = 0 (2.85)

Therefore, K n (pt) is an abelian group.

Example 1: Mc0 is just the abelian monoid of finite-dimensional complex super-vector-

spaces. It is generated by C0|1 and C1|0 and is just Z+ ⊕ Z+ . The submonoid Mc,triv0
requires the existence of an isomorphism t : V 0 → V 1 . In this case we can take
!
0 t
T = −1 (2.86)
t 0

So the submonoid of trivial modules is generated by C1|1 . Therefore

K 0 (pt) := Mc0 /Mc,triv

0
∼
=Z (2.87)

The homomorphism
[V ] 7→ ne − no (2.88)
is in fact an isomorphism.

Example 2: K 1 (pt). Then there is a unique irreducible module M1 for Cℓ1 . We can
take M1 ∼ = C1|1 with, say, ρ(e) = σ 1 . Then we can introduce the odd invertible operator
T = iσ which graded commutes with ρ(e). Therefore M1 ∈ Mtriv
2
1 and since Cℓ1 is a
super-simple algebra all the modules are direct sums of M1 . Therefore Mtriv
1 = M1 and
−1 ∼
hence K (pt) = 0.
16
We can put an Hermitian structure on M and require that T be odd anti-Hermitian. This is useful for
generalizations from the K-theory of a point.

– 55 –
 It is easy to show, using the above table, that
(
Z n = 0(2)
K n (pt) ∼
= (2.89)
0 n = 1(2)

Remarks

1. In general, given an abelian monoid M there are two general methods to produce an
associated abelian group. One, the method adopted here, is to define a submonoid
Mtriv so that the quotient M/Mtriv admits inverses and hence is a group. A second
method, known as the Grothendieck group is to consider the produce M × M and
divide by an equivalence relation. We say that (a, b) is equivalent to (c, d) if there is
an e ∈ M with
a+d+e=c+b+e (2.90)

The idea is that if we could cancel then this would say a − b = c − d. Now it
is easy to see that the set of equivalence classes [(a, b)] is an abelian group, with
[(a, b)] = −[(b, a)]. A standard example is that the Grothendieck group of M = Z+
produces the integers. Note that if we took M = Z+ ∪ {∞} then the Grothendieck
group would be the trivial group. This idea actually generalizes to additive categories ♣Explain the
Grothendieck group
where we have a notion of sum of objects. In that case (2.90) should be understood approach to
K −n (pt)? ♣
to mean that there exists an isomorphism between a + d + e and c + b + e. Then one
takes the monoid of isomorphism classes of objects to the Grothendieck group of the
category.

2. There are very many ways to introduce and discuss K-theory. In the original approach
of Atiyah and Hirzebruch [8], K −n (pt) was defined in terms of stable isomorphism
classes of complex vector bundles on S n . One of the main points of [7] was the
reformulation in terms of Clifford modules, an approach which culminated in the
beautiful paper of Atiyah and Singer [11]. We have chosen this approach because it
is the one closest to the way K-theory appears in physics. For more about this see
the remark in Section §2.3.8 below.

3. A third method uses algebraic K-theory of C∗ -algebras. If we identify K −n (pt) ∼ =

K0 (Cℓn ) then the results on Morita equivalence of the Clifford algebras implies the
Bott periodicity of the K-theories.

2.3.5 Real Clifford algebras and Clifford modules of low dimension

In this section we consider the real Clifford algebras Cℓn for n ∈ Z. We also describe their
irreducible modules and hence the abelian monoid Mn of isomorphism classes of Z2 -graded
representations. ♣Should go back to
previous section and
Here is a summary of the graded and ungraded irreps of the low dimensional real change notation to
Mcn . etc. ♣
Clifford algebras. (Here ε± is odd and ε2± = ±1. ):

– 56 –
 Clifford Algebra Ungraded algebra Graded algebra Ungraded irreps Graded irreps
Cℓ+4 H(2) b
End(R1|1 )⊗H H2 µ̃±
Cℓ+3 C(2) b
R[ε− ]⊗H C2 η̃ 3
Cℓ+2 R(2) C[ε+ ], zε+ = ε+ z̄ R2 η̃ 2
Cℓ+1 R⊕R R[ε+ ] R± , ρ(e) = ±1 η̃
Cℓ0 R R R R , R0|1
1|0

Cℓ−1 C R[ε− ] C η
Cℓ−2 H C[ε− ], zε− = ε− z̄ H η2
Cℓ−3 H⊕H R[ε+ ]⊗Hb H± , ρ(e1 e2 e3 ) = ±1 η3
Cℓ−4 H(2) End(R )⊗H1|1 b H2 µ±

Explicit models for the modules η, µ± and so forth can be found explained in detail in
[43].

2.3.6 The periodicity theorem for Clifford algebras

The fact that Cℓ±4 are the same, and are matrix superalgebras over the even division
algebra H is very significant.
Note that since left and right multiplication of the quaternions acting on H we have

H ⊗ Hopp ∼
= R(4) (2.91)

is an ungraded matrix algebra. Thus,

Cℓ±8 ∼
= End(R8|8 ). (2.92)

It immediately follows that there is a mod-eight periodicity of the Morita equivalence class
of the graded real Clifford algebras Cℓn for all positive n and, separately, for all negative
n.
Moreover, one can also show that

Cℓ±3 ∼ b
= Cℓ∓1 ⊗H (2.93)

Using this identity we can

1. Build the remaining Cℓn for 5 ≤ |n| ≤ 8 by taking graded tensor products with the
Cℓn for |n| ≤ 4. The result is shown in Table **** .

2. Establish mod-eight periodicity of the Morita classes of Cℓn for all integers, both
positive and negative. For example, note that, at the level of Morita equivalence we
have
[Cℓ±1 ] = [Cℓ∓7 ] [Cℓ±2 ] = [Cℓ∓6 ] [Cℓ±3 ] = [Cℓ∓5 ] (2.94)

– 57 –
 Clifford Algebra Ungraded Algebra Mr|s ⊗ Dαs
Cℓ+8 R(16) EndR (R8|8 )⊗D
ˆ 0s
Cℓ+7 C(8) 4|4
EndR (R )⊗D s
ˆ −1
Cℓ+6 H(4) EndR (R2|2 )⊗D s
ˆ −2
Cℓ+5 H(2) ⊕ H(2) EndR (R1|1 )⊗D s
ˆ −3
Cℓ+4 H(2) ˆ 4s
EndR (R1|1 )⊗D
Cℓ+3 C(2) D3s
Cℓ+2 R(2) D2s
Cℓ+1 R⊕R D1s
Cℓ0 R D0s
Cℓ−1 C s
D−1
Cℓ−2 H s
D−2
Cℓ−3 H⊕H s
D−3
Cℓ−4 H(2) EndR (R1|1 )⊗D
ˆ 4s
Cℓ−5 C(4) 1|1
EndR (R )⊗D s
ˆ +3
Cℓ−6 R(8) EndR (R2|2 )⊗D s
ˆ +2
Cℓ−7 R(8) ⊕ R(8) EndR (R4|4 )⊗D s
ˆ +1
Cℓ−8 R(16) EndR (R8|8 )⊗D
ˆ 0s

The upshot is that if we define the following 8 basic superalgebras:

D0s := R
s
D±1 := Cℓ±1
s
D±2 := Cℓ±2 (2.95)
s
D±3 := Cℓ±3
D4s = s
D−4 := H
where D0s and D4s are purely even, then all the Clifford algebras are matrix superalgebras
over the Dαs : The graded Morita equivalence class of [Cℓn ] where n ∈ Z is positive or
negative is determined by the residue α = nmod8, and we have:

[Cℓn ] = [Dαs ] (2.96)

and moreover, the multiplication on Morita equivalence classes is just given by

[Dαs ] · [Dβs ] = [Dα+β

s
] (2.97)

Thus the real graded Brauer group over R is Z/8Z.

The Wedderburn type of the ungraded algebras is now easily determined from the
graded ones by using the explicit determination we gave above for the basic cases Cℓn with
|n| ≤ 4. Notice that there is a basic genetic code in this subject

R, C, H, H ⊕ H, H, C, R, R ⊕ R, R, . . . (2.98)

We will meet it again and again. One would do well to memorize this sequence. It is
illustrated in Figure 9.

– 58 –
Figure 9: An illustration of the “Bott clock”: For Cℓn with decreasing n read it clockwise (=
decreasing phase) and with increasing n read it counterclockwise (= increasing phase).

Finally, we can now easily determine the structure of Cℓr+ ,s− for all r, s. The Morita
class is determined by:
s
[Cℓr+ ,s− ] = [Dr−s ] (2.99)
and hence, lifting α = (r − s)mod8 to |α| ≤ 4
n n
Cℓr+ ,s− ∼ b αs
= End(R2 |2 )⊗D (2.100)

for an n which can be computed by matching dimensions.

2.3.7 KO-theory of a point

Now in this section we describe the real KO-theory ring of a point along the lines we
discussed in Section §2.3.4. In order to complete the story we need to name the irreducible
representations of Cℓ±8 . These are supermatrix algebras and so we have simply λ± ∼ = R8|8
for Cℓ−8 and λ̃± ∼= R8|8 for Cℓ+8 . The superscript ± refers to the sign of the volume form
on the even subspace.
Now let us consider KO−n (pt) along the above lines. A useful viewpoint is that we are
considering real algebras and modules as fixed points of a real structure on the complex
modules and algebras. Recall we described Mtriv,c
n (where the extra c in the superscript
reminds us that we are talking about complex modules of complex Clifford algebras) as
those modules which admit an odd invertible operator which graded commutes with the
Clifford action. In order to speak of real structures we can take our complex modules to

– 59 –
have an Hermitian structure. Then the conjugation will act as T → ±T † where the ± is
a choice of convention. We will choose the convention T → −T † . The other convention
leads to an equivalent ring, after switching signs on the degrees. ♣check ♣

Note that we have introduced an Hermitian structure into this discussion. If one
strictly applies the the Koszul rule to the definition of Hermitian structures and adjoints
√
in the Z2 -graded case then some unusual signs and factors of −1 appear. See Section
§2.2.5 above. We will use a standard Hermitian structure on Rn|m and Cn|m such that
the even and odd subspaces are orthogonal and the standard notion of adjoint. Since we
introduce the structure the question arises whether the groups we define below depend on
that choice. It can be shown that these groups do not depend on that choice, and the main
ingredient in the proof is the fact that the space of Hermitian structures is a contractible
space.
This motivates the following definitions:

Definition
a.) For n ∈ Z, Mn is the abelian monoid of modules for Cℓn under direct sum.
b.) For n ∈ Z, Mtriv
n is the submonoid of Mn consisting of those modules which admit
an odd invertible anti-hermitian operator T which graded-commutes with the Cℓn action.
c.)
KO n (pt) := Mn /Mtriv
n (2.101)

We now compute the KOn (pt) groups for low values of n:

1. Of course KO0 (pt) ∼

= Z, with the isomorphism given by the superdimension.

2. Now consider KO 1 (pt). In our model for η̃ we had ρ(e) = σ 1 . Therefore we could
introduce T = ǫ. Thus [η̃] = 0 in KO-theory and KO 1 (pt) = 0.

3. Next consider KO −1 (pt). In our model for η we had ρ(e) = ǫ. Now we cannot
introduce an antisymmetric operator which graded commutes with ǫ. Thus, η is a
nontrivial class. However, we encounter a new phenomenon relative to the complex
case. Consider 2η = η ⊕ η. As a vector space this is R2|2 and as usual taking an
ordered bases with even elements first we have
 
0 0 −1 0
0 0 0 −1
 
ρ(e) =  = ǫ⊗1 (2.102)
1 0 0 0 
01 0 0

We can therefore introduce T = σ 1 ⊗ ǫ which is odd, anticommutes with ρ(e), and

squares to −1. Therefore, KO −1 (pt) ∼
= Z2 with generator [η].

Now, using the periodicity of the Clifford algebras we conclude that:

– 60 –
Theorem
KOn (pt) is mod-eight periodic in n and the groups KO−n (pt) for 1 ≤ n ≤ 8 are given
by 17

Z2 , Z2 , 0, Z, 0, 0, 0, Z (2.103)

2.3.8 Digression: A hint of the relation to topology

The original Atiyah-Hirzebruch definition of K(X) used the Grothendieck construction
applied to the abelian monoid Vect(X) of complex vector bundles over a topological space
X Thus, we consider equivalence classes [(E1 , E2 )] where [(E1 , E2 )] = [(F1 , F2 )] if there
exists a G with
E1 ⊕ F2 ⊕ G ∼= F1 ⊕ E2 ⊕ G (2.104)
Intuitively, we think of [(E1 , E2 )] as a difference E1 − E2 .

Example: If we consider from this viewpoint the K-theory of a point K 0 (pt) then we
obtain the abelian group Z, the isomorphism being [(E1 , E2 )] → dimE1 − dimE2 .
For vector bundles the Grothendieck construction can be considerably simplified thanks
to the Serre-Swan theorem:

Theorem[Serre; Swan] Any vector bundle 18 has a complementary bundle so that E⊕E ⊥ ∼ =
θN is a trivial rank N bundle for some N . Equivalently, every bundle is a subbundle of a
trivial bundle defined by a continuous family of projection operators.
This leads to the notion of stable equivalence of vector bundles: Two bundles E1 , E2
are stably equivalent if there exist trivial bundles θs of rank s so that

E1 ⊕ θs1 ∼
= E2 ⊕ θs2 (2.105)

Example: A very nice example, in the category of real vector bundles, is the tangent
bundle of S 2 . The real rank two bundle T S 2 is topologically nontrivial. You can’t comb
the hair on a sphere. However, if we consider S 2 ⊂ R3 the normal bundle is a real rank
one bundle and is trivial. But that means T S 2 ⊕ θ1 ∼= θ3 . So T S 2 is stably trivial.
Returning to the general discussion. In the difference E1 − E2 we can add and subtract
the complementary bundle to get (E1 ⊕ E2⊥ ) − θN for some N . If we restrict the bundle to
any point we get an element of K 0 (pt). By continuity, it does not matter what point we
choose, provided X is connected.
In other words, there is a homomorphism

K 0 (X) → K 0 (pt) (2.106)

The kernel of this homomorphism is, by definition, K̃ 0 (X). We can represent it by formal
differences of the form E − θN where N = rank(E).
17
This is easily memorized using the “Bott song.” Sing the names of the groups to the tune of ”Ah! vous
dirai-je, Maman,” aka ”Twinkle, twinkle, little star.”
18
over a suitable nice topological space

– 61 –
 There is another construction of K(X) called the difference-bundle approach which is
more flexible and useful for generalizations.
As in our discussion using Clifford modules we consider an abelian monoid and divide
by a submonoid of “trivial” elements. As we mentioned, the latter viewpoint is closer to
the physics. The abelian monoid consists of isomoprhism classes of pairs (E, T ) where E is
a Z2 -graded complex (or real) vector bundle over X equipped with an Hermitian structure,
and T is an odd anti-Hermitian operator (that operator could well be zero). The trivial
submonoid will be those pairs (E, T ) where T is invertible.
Very roughly speaking the difference E 0 − E 1 in the Grothendieck construction is to
be compared with a Z2 -graded bundle E = E 0 ⊕ E 1 with an odd operator T = 0. Note
that in the Grothendieck construction we could also have used

E 0 − E 1 = (E 0 ⊕ F 0 ) − (E 1 ⊕ F 1 ) (2.107)

if there is an isomorphism of t : F 0 → F 1 . This would correspond to to the Z2 -graded

bundle with even part E 0 = E 0 ⊕ F 0 and odd part E 1 = E 1 ⊕ F 1 with odd anti-hermitian
operator (in an ordered basis for the fibers of E 0 , E 1 , F 0 , F 1 , in that order):
 
0 00 t
 0 0 0 0
 
T =  (2.108)
 0 0 0 0
−t∗ 0 0 0

= kerT † ∼
= kerT while E 1 ∼
Now T ∈ End(E) is odd, and E 0 ∼ = cokT . Therefore:

E0 − E1 ∼
= (kerT |E 0 ⊕ (kerT |E 0 )⊥ ) − (kerT † |E 1 ⊕ (kerT † |E 1 )⊥ ). (2.109)

Now T provides a bundle isomorphism between (kerT |E 0 )⊥ and (kerT † |E 1 )⊥ , so these can
be canceled.

Remark: In string theory, T turns out to be the classical value of a tachyon field [53]. The
invertibility of the tachyon field has an interpretation in terms of being able to undergo
renormalization group flow to its true vacuum value. See [39] for more about this viewpoint.
In the applications to topological phases of matter T is related to “topologically trivial
pairing of particles and holes”. See, e.g. discussion in [36, 48].

Now let us briefly consider the relation to generalized cohomology theory. We would
like to define K −n (X).
In some treatments of K-theory one will find K −n defined in terms of suspension: ♣Careful. Need to
define K̃ and be
careful about
K̃ −n (X) := K̃ 0 (S n X) (2.110) basepoints and
suspension vs.
reduced suspension.
♣
However, another way to introduce the degree is enhance our Z2 -graded bundles E
with odd antihermitian endomorphism T by including a left graded Cℓn action that T

– 62 –
graded commutes with T . This is actually an example of twisting of K-theory by a bundle
of (Clifford) algebras. (For more about why that is, see [25].)
The relation between the old definition using suspension (which makes clear the relation
to generalized cohomology theory) and the definition of K −n (X) using Clifford algebras is
provided by the Thom isomorphism of K theory together with a nice model for the Thom
class constructed using Clifford modules.
The Thom isomorphism in K-theory states that if π : E → X is a real rank m vector
bundle (with some orientation data) then

K n (X) ∼ n+m
= Kcpt−vert (E) (2.111)

with the isomorphism given by [V ] → π ∗ [V ] ⊗ Φ for a K-theory class Φ called the Thom
class. This is a very general statement in generalized cohomology theory. See Section 12
of [7] for a very nice discussion.
Now, in order to explain the model for the Thom class Φ in real and complex K-theory
we need to look at some interesting ways that Clifford algebras are related to the homotopy
groups of the unitary groups.
To explain this let us begin with the following standard result from the theory of fiber
bundles:

Theorem. If d > 1 and G is connected then principal G-bundles on S d are topologi-

cally classified by πd−1 (G), i.e. there is an isomorphism of sets:

PrinG (S d ) ∼
= πd−1 (G) (2.112)

The isomorphism is very easily understood from choosing north and southern hemi-
spheres as patches and thinking about the gluing function for the fibers at the equator.
For complex vector bundles of rank N the relevant structure group is GL(N, C). We
can always put an Hermitian structure on the vector bundle and take the structure group
to be U (N ). Now we recall Bott’s result for the stable homotopy groups:

π2p−1 (U (N )) = Z N ≥p (2.113)

π2p (U (N )) = 0 N >p (2.114)

Note that these equations say that for N sufficiently large, the homotopy groups do
not depend on N . These are called the stable homotopy groups of the unitary groups and
can be denoted πk (U). The mod two periodicity of πk (U) as a function of k is known as
Bott periodicity. We will prove these statements using Clifford algebras and Morse theory
in Lecture 3.

– 63 –
 It follows from this theorem that, for N > d/2 we have
(
d ∼ Z d = 0(2)
VectN
C (S ) = (2.115)
0 d = 1(2)

where VectN d
C (S ) is the set of isomorphism classes of rank N complex vector bundles over
Sd.

Remark

1. One way to measure the integer is via a characteristic class known as the Chern
character ch(E) ∈ H 2∗ (X; Q). If we put a connection on the bundle then we can write
an explicit representative for the image of ch(E) in DeRham cohomology. Locally the
connection is an anti-hermitian matrix-valued 1-form A. It transforms under gauge
transformations like
(d + A) → g−1 (d + A)g (2.116)
The fieldstrength is
F = dA + A2 (2.117)
and is locally an anti-hermitian matrix-valued 2-form transforming as F → g −1 F g.
Then, in DeRham cohomology
 
F
ch(E) = [Trexp ] (2.118)
2πi
and the topological invariant is measured by
Z
ch(E) (2.119)
Sd

Note that since ch(E) has even degree this only has a chance of being nonzero for
d even. On a bundle with transition function g on the equator we can take A =
rg−1 dg on the northern hemisphere, where g(x) is a function only of the “angular
coordinates” on the hemisphere and A = 0 on the southern hemisphere. Note that
thanks to the factor of r, which vanishes at the north pole this defines a first-order
differentiable connection. For this connection the fieldstrength is

F = drg −1 dg − r(1 − r)(g−1 dg)2 (2.120)

and hence if d = 2ℓ
Z Z 1 Z
1
ch(E) = (−1)ℓ−1 ℓ−1
(r(1 − r)) dr Tr(g−1 dg)2ℓ−1
S 2ℓ (2πi)ℓ (ℓ − 1)! 0 S 2ℓ−1
Z (2.121)
(ℓ − 1)!
= (−1)ℓ−1 Tr(g −1
dg)2ℓ−1
(2πi)ℓ (2ℓ − 1)! S 2ℓ−1
The integral of the Maurer-Cartan form over the equator measures the homotopy
class of the transition function g. It is not at all obvious that this integral will be
an integer, but for U (N ) and the trace in the N it is. This is a consequence of the
Atiyah-Singer index theorem.

– 64 –
 Note that in the equations (2.112) and (2.115) there is no obvious abelian group op-
eration despite the fact that in these isomorphism of sets the RHS has a structure of an
abelian group. We can of course take direct sum, but this operation changes the rank. If
we want to make an abelian group we might take a direct sum in the case of VectN C (X),
but this changes the rank. This is one place where K-theory is useful. We can form the
abelian monoid obtained by taking the direct sum

Vect(S d ) := ⊕N ≥0 VectN d
C (S ). (2.122)

then we can take, e.g., the Grothendieck construction.

For spheres, we will get, from this result:

(
0 d Z d = 0(2)
K̃ (S ) = (2.123)
0 d = 1(2)

and this is the abelian group which is to be compared with the group (2.89) defined above. ♣Explain a little
more that ch is a
Now, let us finally relate all this to the Clifford algebras. We begin with projected homomorphism to
deduce this
bundles which have fibers given by the spin representation. isomorphism of
groups. ♣
Therefore, let us begin with an ungraded irreducible representation of Cℓ2k+1 by Her-
mitian matrices Γµ , µ = 1, . . . , 2k + 1. We consider the unit sphere S 2k ⊂ R2k+1 defined
by
2k+1
X
XµXµ = 1 (2.124)
µ=1

Therefore, restricted to the sphere (X µ Γµ )2 = 1 and hence

1
P± (X) := (1 + X µ Γµ ) (2.125)
2

are a family of projection operators, with X ∈ S 2k , acting on the fixed vector space
k
V = C2 .
We then form nontrivial projected bundles by taking E = S 2k × V to be the trivial
bundle and letting
E± |X := P± (X)V (2.126)

These will be nontrivial bundles with nontrivial Chern characters. Indeed, to see this let
us choose a trivialization and compute the transition function.
We consider northern and southern hemispheres as patches and hence divide up the
coordinates as
X µ = (xa , x2k , y) (2.127)

– 65 –
where a = 1, . . . , 2k − 1. Moreover, we can choose our ungraded representation of Cℓ2k+1
so that
!
a a 1 0 γa
Γ =γ ⊗σ = a = 1, . . . , 2k − 1
γa 0
!
0 −i
Γ2k = 12k−1 ⊗ σ 2 = (2.128)
i 0
!
1 0
Γ2k+1 = 12k−1 ⊗ σ 3 =
0 −1

where γ a is an ungraded representation of Cℓ2k−1 by Hermitian matrices.

Choose a basis vα , α = 1, . . . , 2k−1 for the irrep of Cℓ2k−1 on which γ a is an ungraded
representation. Then !
vα
(2.129)
0

is a trivialization of the bundle V+ at the north pole y = 1. Indeed:

! !
v α 1 (1 + y)v α
u+
α := P+ = (2.130)
0 2 (γ a xa + ix2k )vα

Similarly, !
0
(2.131)
vα

provides a trivialization at the south pole y = −1 and

! !
− 0 1 (γ a xa − ix2k )vα
uα := P+ = (2.132)
vα 2 (1 − y)vα

The transition function at the equator S 2k−1 , y = 0 is

u− +
α = −iυ(x)uα (2.133)

where
2k−1
X
υ(x) = x2k + (iγ a )xa (2.134)
a=1

is a unitary matrix.
We claim that this unitary matrix actually generates the homotopy group π2k−1 of the
unitary group. This construction generalizes the standard constructions of the magnetic
monopole and instanton bundles on S 2 and S 4 , respectively. Indeed, choosing the trivial
connection on E we can form projected connections on E ± . The projected connections on
E ± define the basic (anti)monopole and (anti)instanton connections.
The above construction provides motivation for the following discussion:

– 66 –
 Consider an ungraded representation (not necessarily irreducible) of Cℓd by anti-
Hermitian gamma matrices on a vector space (with basis) V where Γµ are such that
{Γµ , Γν } = −2δµν , where µ = 1, . . . , d. 19 Let dimC V = L.
Suppose x0 , xµ , µ = 1, . . . , d are functions on the unit sphere S d embedded in Rd+1 ,
so
x20 + xµ xµ = 1 (2.135)
Consider the matrix-valued function

υ(x) := x0 1 + xµ Γµ (2.136)
Note that
υ(x)υ(x)† = 1 (2.137)
and therefore υ(x) is a unitary matrix for every (x0 , xµ ) ∈ S d . We can view υ(x) as
describing a continuous map υ : S d → U (L). Therefore it defines an element of the
homotopy group [υ] ∈ πd (U (L)). The following examples show that the homotopy class of
the map can be nontrivial:

Example 1: If d = 1 then we could take either of the ungraded irreducible representations

V = C and Γ = ±i. If x20 + x21 = 1 then

υ ± (x) = x0 ± ix1 (2.138)

and, for either choice of sign, [υ ± ] is a generator of π1 (U (1)) = Z.

Example 2: If d = 3 then we may choose either of the ungraded representations V = C2

√
and Γi = ± −1σ i and then
υ(x) = x0 + xi Γi (2.139)
is one way to parametrize SU (2). Thus the map υ : S 3 → SU (2) is the identity map (with
the appropriate orientation on S 3 ). If we fix a an orientation on S 3 we get winding number
±1 and hence [υ ± ] is a generator of π3 (SU (2)) ∼= π3 (S 3 ) ∼
= Z.

Here is one easy criterion for triviality of [υ]: Suppose we can introduce another ♣Make the
connection between
anti-Hermitian L × L gamma matrix on V , call it Γ, so that Γ2 = −1 and {Γ, Γµ } = 0. Γ and the odd
invertible element υ
Now consider the unit sphere that graded
commutes with Cℓn
d we discussed above?
X But we are working
S d+1 = {(x0 , xµ , y)|x20 + xµ xµ + y 2 = 1} ⊂ Rd+2 (2.140) with ungraded reps
so it is not clear
µ=1 what “odd” means.
♣
Then we can define
υ̃(x, y) = x0 + xµ Γµ + yΓ (2.141)
When restricted to S d+1 ⊂ Rd+2 , υ̃ is also unitary and maps S d+1 → U (L). Moreover
υ̃(x, 0) = υ(x) while υ̃(0, 1) = Γ. Thus υ̃(x, y) provides an explicit homotopy of υ(x) to
the constant map.
19
Note that the notation µ has changed from the previous discussion.

– 67 –
Figure 10: The map on the equator extends to the northern hemisphere, and is therefore homo-
topically trivial.

Thus, if the representation V of Cℓd is the restriction of a representation of Cℓd+1

then υ(x) is automatically homotopically trivial.
Let us see what this means if we combine it with what we learned above about the
irreducible ungraded representations of Cℓd .

p p
1. If d = 2p we have irrep N2p ∼ ±
= C2 . It is indeed the restriction of N2p+1 ∼
= C2 and
hence, υ(x) must define a trivial element of π2p (U (L)), with L = 2p .
p
±
2. On the other hand, if d = 2p + 1 then N2p+1 = C2 is not the restriction of N2p+2 ∼
∼ =
2p+1
C . All we can conclude from what we have said above is that υ(x) might define a
homotopically nontrivial element of π2p+1 (U (L)) with L = 2p . On the other hand, if
+ −
we had used V = N2p+1 ⊕ N2p+1 then since V is the restriction of the representation
N2p+2 and υ = υ ⊕ υ , it follows that the homotopy classes satisfy [υ − ] = −[υ + ].
+ −

Now, a nontrivial result of [7] is:20

Theorem[Atiyah, Bott, Shapiro]. If V is an irreducible ungraded representation of Cℓd

then [υ] generates πd (U (L)).

∼ Z, with generator [υ + ] or [υ − ].
It therefore follows that π2p (U (L)) = 0 and π2p+1 (U (L)) =
Now, finally, we can use the above constructions to provide an explicit model for the
Thom class.
The simplest case to state is when V → X is a real, rank 8, oriented bundle with a
spin structure. Then we need to construct a K-theory class on V compactly supported
in the fiber directions. To do this we give V a metric so that we can form a bundle of
Clifford algebras Cl(V ) → X. Let S ± be the associated spin bundles, pulled back to V
itself. Then we take the difference bundle T : S + → S − where T is just T (v) = v · γ given
by Clifford multiplication by the vector v in the fiber. This is invertible away from v = 0,
the zero section and represents the Thom class, as described in [7]. This construction can
be generalized to arbitrary twisted equivariant vector bundles, as described in the papers
of Freed-Hopkins-Teleman, and as used extensively in my work with Distler and Freed on
orientifolds. See [25] for some further discussion.
20
More precisely, they used the above υ(x) to define a K-theoretic Thom class. Then the result we have
stated follows from the relation of K-theory to homotopy theory.

– 68 –
2.4 The 10 Real Super-division Algebras

Definition An associative unital superalgebra over a field κ is an associative super-division

algebra if every nonzero homogeneous element is invertible.

Example 1: We claim that Cℓ1 is a superdivision algebra over κ = C (and hence a

superdivision algebra over R). Elements in this superalgebra are of the form x + ye with
x, y ∈ C. Homogeneous elements are therefore of the form x or ye, and are obviously
invertible, if nonzero. Note that it is not true that every nonzero element is invertible!
For example 1 + e is a nontrivial zero-divisor since (1 + e)(1 − e) = 0. Thus, Cℓ1 is not a
division algebra, as an ungraded algebra.

Example 2: We also claim that the 8 superalgebras Dαs , with α ∈ Z/8Z defined in (2.95)
are real super-division algebras. The argument of Example 1 show that Cℓ±1 are super-
division algebras. For Cℓ±2 the even subaglebra is isomorphic to the complex numbers,
which is a division algebra. It follows that Cℓ±2 are superdivision algebras. To spell this
out in more detail: For Cℓ+2 we note that for even elements we can write

(x + ye12 )(x − ye12 ) = x2 + y 2 (2.142)

and for odd elements we can write

(xe1 + ye2 )2 = x2 + y 2 (2.143)

where x, y ∈ R. Thus the nonzero homogeneous elements are invertible. For Cℓ−2 the
equation (2.142) holds and (2.143) simply has a sign change on the RHS. So this too is a
superdivision algebra. More conceptually, note that Cℓ0±2 is isomorphic to C, which is a
division algebra, and Cℓ1±2 is related to Cℓ0±2 by multiplying with an invertible element. We
can now apply this strategy to Cℓ±3 : The even subalgebra is isomorphic to the quaternion
algebra, which is a division algebra and the odd subspace is related to the even subspace
by multiplication with an invertible odd element. Hence Cℓ±3 is a superdivision algebra.
Note well that Cℓ+1,−1 being a matrix superalgebra is definitely not a superdivision
algebra! For example
!
10
(2.144)
00

is even and is a nontrivial zerodivisor. By the same token, Cℓ±4 ∼ b is also

= End(R1|1 )⊗H
not a superdivision algebra.
The key result we need is really a corollary of Wall’s theorem classifying central simple
superalgebras. For a summary of Wall’s result see Appendix ??.

Theorem There are 10 superdivision algebras over the real numbers: The three purely
even algebras R, C, H, together with the 7 superalgebras Cℓ1 , Cℓ±1 , Cℓ±2 , Cℓ±3 .

– 69 –
2.5 The 10-fold way for gapped quantum systems
We are now in a position to describe the generalization of Dyson’s 3-fold way to a 10-fold
way, valid for gapped quantum systems. ♣NEED TO
REWRITE:
Recall from our discussion of a general symmetry of dynamics (§1.8) that if G is a ASSUME AN a
priori (G, φ, χ) rep
symmetry of the dynamics of a gapped quantum system then there are two independent on a super-Hilbert
space H. ♣
homomorphisms (φ, χ) : G → Z2 . In the Dyson problem one explicitly assumes that χ = 1.
Nevertheless, as we saw when discussing phases of gapped systems in Section §2.1, there
is a natural Z2 -grading of the Hilbert space so that, if χ 6= 1 then the Hilbert space is a
(φ, χ)-representation of G. (See Definition (2.3).) Therefore we can state the

Generalized Dyson Problem: Let G be a bigraded compact group, bigraded by (φ, χ), and
H a Z2 -graded (φ, χ)-representation H of G. What is the ensemble of gapped Hamiltonians
H such that G is a symmetry of the dynamics?

Remark: One must use caution here when thinking about the Z2 -grading induced by
the Hamiltonians in the ensemble. If we interpret the Hamiltonian-reversing property of
group elements ρ(g) with χ(g) = −1 as a graded commutation relation, then H must be
considered odd. On the other hand, if sign(H) is a grading then H must be even! It is
best to consider a (φ, χ)-graded representation and ask what are the compatible gapped
Hamiltonians. Then the Hamiltonians in the ensemble each induce different gradings on
the Hilbert space.

We can proceed to answer this along lines closely analogous to those for Dyson’s 3-fold
way.
First, we imitate the definitions of Section §1.7.1 for φ-representations:

Definitions:

1. If G is a bigraded group by (φ, χ) then a (φ, χ)-representation is defined in (2.3).

2. An intertwiner or morphism between two (φ, χ)-reps (ρ1 , V1 ) and (ρ2 , V2 ) is a C-linear
map T : V1 → V2 , which is a morphism of super-vector spaces: T ∈ HomC (V1 , V2 ),
which commutes with the G-action:

T ρ1 (g) = ρ2 (g)T ∀g ∈ G (2.145)

We write HomG
C (V1 , V2 ) for the set of all intertwiners.

3. An isomorphism of (φ, χ)-reps is an intertwiner T which is an isomorphism of complex

supervector spaces.

– 70 –
 4. A (φ, χ)-rep is said to be φ-unitary if V has a nondegenerate even Hermitian structure
21 such that ρ(g) is an isometry for all g. That is, it is unitary or anti-unitary

according to whether φ(g) = +1 or φ(g) = −1, respectively.

5. A (φ, χ)-rep (ρ, V ) is said to be reducible if there is a nontrivial proper (φ, χ)-sub-
representation. That is, if there is a complex super-vector subspace W ⊂ V , (and
hence W 0 ⊂ V 0 and W 1 ⊂ V 1 ) with W not {0} or V which is G-invariant. If it is
not reducible it is said to be irreducible.

As before, if G is compact and (ρ, V ) is a (φ, χ)-rep then WLOG we can assume that
the rep is unitary, by averaging. Then if W is a sub-rep the orthogonal complement is
another (φ, χ)-rep, and hence we have complete reducibility.
Now we need to deal with a subtle point. In addition to intertwiners we needed to
consider the graded intertwiners HomGC (V1 , V2 ) between two (φ, χ)-representations. These ♣Be consistent in
notation about
are super-linear transformations T such that if we decompose T = T 0 + T 1 into even and V, V ′ vs. V1 , V2 . ♣

odd transformations then T 0 ∈ HomG 1

C (V1 , V2 ) but T instead satisfies

T 1 ρ1 (g) = χ(g)ρ2 (g)T 1 ∀g ∈ G (2.146)

Two irreducible representations can be distinct as (φ, χ)-representations but can be graded-
isomorphic. The simplest example is G = {1} which has graded irreps C1|0 and C0|1 .
Let {Vλ } be a set of representatives of the distinct graded-isomorphism classes of
irreducible (φ, χ)-representations. We then obtain the isotypical decomposition of (φ, χ)-
representations:
H∼ = ⊕λ HomG b
C (Vλ , H)⊗Vλ (2.147)

Note that HomG C (Vλ , H) is not an even vector space in general. This will be more conve- ♣Careful. Then
decomposition is
nient to us because of the nature of the super-Schur lemma: not unique. ♣

Lemma[Super-Schur] Let G be a Z2 × Z2 -graded group, graded by the pair of homo-

morphisms (φ, χ).
a.) If T is a graded intertwiner between two irreducible (φ, χ)-representations (ρ, V )
and (ρ′ , V ′ ) then either T = 0 or there is an isomorphism of (ρ, V ) and (ρ′ , V ′ ).
b.) If (ρ, V ) is an irreducible (φ, χ)-representation then the super-commutant Zs (ρ, V ),
namely, the set of graded intertwiners of (ρ, V ) with itself is a super-division algebra.

Now we can now proceed as before to derive the analog of Dyson’s ensembles. We
consider the isotypical decomposition (2.147) of H. Let Sλ := HomG C (Vλ , H). It is a real
super-vector space of degeneracies. Now we compute the set of superlinear transforma-
tions:
HomC (H, H) ∼= ⊕λ,µ (Sλ∗ ⊗S
b µ )⊗Hom
b C (Vλ , Vµ ) (2.148)
21
See Section §2.2.5 above.

– 71 –
Now we take the graded G-invariants and apply the super-Schur lemma to get

Zs (ρ, H) = HomG ∼ b s
C (V, V ) = ⊕λ End(Sλ )⊗Dλ (2.149)

where to each isomorphism class of graded-irreducible (φ, χ)-rep λ we associate a unique real
0 1
super-division algebra Dλs . Meanwhile End(Sλ ) ∼ = End(Rsλ |sλ ) is a matrix superalgebra.
Finally, let us apply this to the generalized Dyson problem. If G is a symmetry of the
dynamics determined by H then

Hρ(g) = χ(g)ρ(g)H (2.150)

and hence the C-linear operator H is in the graded-commutant of the given (φ, χ) repre-
sentation H. Therefore, H is in the space (2.149). For each irreducible representation λ
there is a corresponding super-division algebra Dλs and this gives the 10-fold classification.
To write the ensemble of Hamiltonians more explicitly we recall that H must be a self-
adjoint element of Zs (ρ, H). There is a natural ∗ structure on the superdivision algebras
since the Clifford generators can be represented as Hermitian or anti-Hermitian operators.
That is, we take e∗i = ±ei with the sign determined by e∗i = e3i . We then extend this to
be an anti-automorphism, and for End(Sλ )⊗D b λs we take the tensor product of the natural
∗-structures. H must be a self-adjoint element of this superalgebra.
Moreover, if χ(g) is nontrivial for any g then H must be in the odd element of the
superdivision algebra.
Thus, the 10-fold way is the following:

1. If the (φ, χ) representation has χ = 1 then the generalized Dyson problem is identical
to the original Dyson problem, and there are three possible ensembles.

2. But if χ is nontrivial then there are new ensembles not allowed in the Dyson classi-
fication. In these cases, Dλs is one of the 7 superalgebras which are not purely even
and H is an odd element of the superalgebra End(Sλ )⊗D b λs .

Remarks:

1. It was easy to give examples of the three classes in Dyson’s 3-fold way. Below we will
give examples using the 10 bigraded “CT-groups” discussed in Section §2.6 below.

2. The above is, strictly speaking, a new result, although it is really a simple corollary
of [26]. However, it should be stressed that the result is just a very general statement
about quantum mechanics. No mention has been made of bosons vs. fermions, or
interacting vs. noninteracting, or one-body vs. many-body Hamiltonians. ♣Go back to
Fidkowski-Kitaev
and interpret their
3. A key point we want to stress is that the 10-fold way is usually viewed as 10 = 2 + 8, interacting models
in this framework.
where 2 and 8 are the periodicities in complex and real K-theory. And then the They should be
getting some of the
K-theory classification of topological phases is criticized because it only applies to 7 non-Dyson
classes. ♣
free systems. However, we believe this viewpoint is slightly misguided. The unifying
concept is really that of a real super-division algebra, and there are 10 such. They
can be parceled into 10 = 8 + 2 but they can also equally naturally be parceled into
10 = 7 + 3 (with the 3 referring to the purely even superdivision algebras).

– 72 –
 4. The Altland-Zirnbauer classification discussed below makes explicit reference to free
fermions.

5. As a curious digression we note that in Dyson’s original paper on φ-representations

[20] he in fact had a 10-fold classification of irreducible φ-representations! (Assuming
φ 6= 1.) Conjecturally, it is related to the 10 real superdivision algebras, although
this has never been demonstrated in a totally satisfactory way. See [43] for more
discussion.

2.6 Realizing the 10 classes using the CT groups

To make contact with some of the literature on topological insulators we describe here the
10 “CT groups.” (This is a nonstandard term used in [26].) This is a set of 10 bigraded
groups which we now define.
To motivate the 10 CT groups note that in some disordered systems, (sometimes well-
described by free fermions), the only symmetries we might know about a priori are the
presence or absence of “time-reversal” and “particle-hole” symmetry. Thus it is interesting
to consider the various φ-twisted extensions of the group

M2,2 = hT̄ , C̄|T̄ 2 = C̄ 2 = T̄ C̄ T̄ C̄ = 1i ∼

= Z2 × Z2 (2.151)

or of its subgroups.
A second motivation is that if G is a bigraded group by (φ, χ) then there is a sequence

1 → G0 → G → M2,2 → 1 (2.152)

It is easy to show that, WLOG, if the bigrading (φ, χ) is surjective then we can choose
generators of M2,2 so that:

φ(T̄ ) = −1 φ(C̄) = −1
χ(T̄ ) = +1 χ(C̄) = −1 (2.153)
τ (T̄ ) = −1 τ (C̄) = +1

where we have defined τ from φ and χ so that τ · φ · χ = 1.

In general we should allow for the possibility that one or two of the gradings φ, τ, χ is
trivial and this motivates us to study subgroups of the group M2,2 .
Therefore, let us consider the φ-twisted extensions of M2,2 and its subgroups. This is
a simple generalization of the example we discussed in Section §1.5, equation (1.69). First,
let us note that there are 5 subgroups of M2,2 depending on whether T̄ , C̄ or T̄ C̄ is in the
group. See Figure 11.
As in the example (1.69) the isomorphism class of the extension is completely deter-
mined by whether the lift T̃ and/or C̃ of T̄ and/or C̄ squares to ±1. After a few simple
considerations discussed in the exercises below it follows that one has the table of 10 φ-
twisted extensions of the subgroups of M2,2 :

– 73 –
 Figure 11: The 5 subgroups of Z2 × Z2 .

Subgroup U ⊂ M2,2 T̃ 2 C̃ 2 [Clifford]

{1} [Cℓ0 ] = [C]
{1, S̄} [Cℓ1 ]
{1, T̄ } +1 [Cℓ0 ] = [R]
M2,2 +1 −1 [Cℓ−1 ]
{1, C̄} −1 [Cℓ−2 ]
M2,2 −1 −1 [Cℓ−3 ]
{1, T̄ } −1 [Cℓ4 ] = [H]
M2,2 −1 +1 [Cℓ+3 ]
{1, C̄} +1 [Cℓ+2 ]
M2,2 +1 +1 [Cℓ+1 ]

Now, we can generalize the remark near the example of Section §1.7.1. Recall that we
could identify φ-representations of φ-twisted extensions of M2 with real and quaternionic
vector spaces. If we consider subgroups of M2 then for the trivial subgroup we also get
complex vector spaces. This trichotomy is generalized to a decachotomy for the CT groups:

– 74 –
Theorem There is a one-one correspondence, given in the table above, between the ten
CT groups and the ten real super-division algebras (equivalently, the 10 Morita classes
of the real and complex Clifford algebras) such that there is an equivalence of categories
between the (φ, χ)-representations of the CT group and the graded representations of the
corresponding Clifford algebra.

Sketch of Proof:
We systematically consider the ten cases beginning with a (φ, χ)-representation of a
CT group and producing a corresponding representation of a Clifford algebra. Then we
show how the inverse functor is constructed.

1. First, consider the subgroup U = {1}. A (φ, χ) representation W is simply a Z2 -

graded complex vector space, so V = W is a graded Cℓ0 -module.

2. Now consider U = {1, S̄}. There is a unique central extension and S = C̃T acts on
W as an odd operator which, WLOG, we can take to square to one. Moreover, S is
C-linear. Therefore, we can take V = W and identify S with an odd generator of
Cℓ1 .

3. Now consider U = {1, C̄}. On the representation W of U tw we have two odd anti-
linear operators C̃ and iC̃. Note that

(iC̃)2 = C̃ 2 {iC̃, C̃} = 0 (2.154)

since C̃ is antilinear. Therefore, we can define a graded Clifford module V = W with

e1 = C̃ and e2 = iC̃. It is a Clifford module for a real Clifford algebra, again because
C̃ is anti-linear. The Clifford algebra is Cℓ+2 if C̃ 2 = +1 and Cℓ−2 if C̃ 2 = −1.

4. Next, consider U = {1, T̄ }. The lift T̃ to U tw acts on a (φ, χ) representation W

as an even, C-antilinear operator. It is therefore a real structure if T̃ 2 = +1 and a
quaternionic structure if T̃ 2 = −1. In the first case, the fixed points of T̃ define a real
Z2 -graded vector space V = W |T̃ =+1 which is thus a graded module for Cℓ0 . In the
second case, T̃ defines a quaternionic structure on V = WR . As we have seen, Cℓ4 is
Morita equivalent to H, and in fact the Cℓ4 module is V ⊕ V . (Recall equation (??)
above.)

5. Now consider U = M2,2 . This breaks up into 4 cases:

6. If T 2 = +1 then, as we have just discussed T defines a real structure. As shown in

the exercises, WLOG we can choose the lift of C so that CT = T C. Therefore, C
acts as an odd operator on the real vector space of T = +1 eigenstates: V = W |T =+1 .
Then V is the corresponding module for Cℓ±1 according to whether C 2 = ±1.

7. If T 2 = −1, then, as we just discussed, T defines a quaternionic structure on V = WR .

Then C, iC, and iCT are odd endomorphisms of WR and one checks they generate
a Cℓ+3 action if C 2 = +1 and a Cℓ−3 action if C 2 = −1.

– 75 –
♦

Now, in order to give our application to the generalized Dyson problem we note a key:

Proposition: Let U tw be one of the 10 bigraded CT groups and let D be the associated
real superdivision algebra. Let (ρ, W ) be an irreducible (φ, χ)-rep of U tw . Then, the graded
commutant Zs (ρ, W ) is a real superdivision algebra isomorphic to D opp .

This is proved in detail in [43]

We can now give examples of all 10 generalized Dyson classes. If U tw corresponds to

one of the even superdivision algebras R, C, H then there are two irreps W± . The general
⊕n ⊕n
rep of U tw is isomorphic to H = W+ + ⊕ W− − . Then the graded commutant is

Zs (ρ, H) = End(Rn+ |n− )⊗D

b opp (2.155)

In these cases the group U tw (which is isomorphic to Pin± (1), see below) is purely even so
the Hamiltonian can be even or odd or a sum of even and odd. We can therefore forget
about the grading and we recover precisely Dyson’s 3 cases. If U tw corresponds to one of
the remaining 7 superdivision algebras (those which are not even) then there is a unique
graded irrep W and up to isomorphism H = W ⊕n so again

Zs (ρ, H) = End(Rn )⊗D

b opp (2.156)

As discussed above we can impose Hermiticity conditions on the graded commutant

to get the relevant ensembles of Hamiltonians.

Example 1: Let U = {1, S̄} ∼ = M2 with φ(S̄) = +1 and χ(S̄) = −1. Since φ = 1 the
extension of M2 must be trivial and U tw = U (1) × M2 and there is a single graded irrep
W = C1|1 with ! !
z1 z2
S: → (2.157)
z2 z1

Every (φ, χ) representation is of the form W ⊕n ∼

= Cn|n . In a basis of (even,odd) the general
commutant consists of complex matrices of the form
!
α β
(2.158)
−β α

The Hermitian ensemble consists of such matrices with α† = α and β † = −β. Since the
Hamiltonian must be odd the ensemble is
!
0 β
E(Cℓ1 ) = { |β ∈ Mn (C) & β † = −β} (2.159)
−β 0

– 76 –
and we further require that 0 is not in the spectrum, so β is gappped. ♣just say
β ∈ GL(n, C) ? ♣
How does this differ from the Dyson ensemble? If we compare with M2 acting C-
linearly but evenly then there are two irreps and we have Cn1 ⊗ 1 ⊕ Cn2 ⊗ ǫ. The ensembles
are Hermn1 (C) × Hermn2 (C).

Example 2: If U = {1, C̄} with φ(C̄) = χ(C̄) = −1 then the extension U tw is characterized
by a sign C 2 = ξ ∈ {±1}. The bigraded group U tw has a single (φ, χ)-irrep, W ∼ = C1|1
with action ! !
z1 ξ z̄2
C: → (2.160)
z2 z̄1
The most general super-linear transformation which graded commutes with C is
!
α β
A= (2.161)
−ξ β̄ ᾱ

the algebra of these matrices is Cℓ∓2 for ξ = ±1. The general (φ, χ) representation can be
written as W ⊕n and the graded commutant consists of matrices of the form (2.161) where
now α, β are n × n complex matrices and ᾱ means the complex conjugate. To pass to the
ensemble of compatible Hamiltonians we require that H be Hermitian and odd. Therefore,
again α = 0 and β † = −ξ β̄, that is, β tr = −ξβ, and again β must be gapped, and so,
labeling the ensemble by the superdivision algebra of the graded commutant:
!
0 β
E(Cℓ∓2 ) = { |β ∈ Mn (C) & β tr = −ξβ} (2.162)
−ξ β̄ 0

Example 3: If U = M2,2 then we again have four cases to consider. Suppose first that
T 2 = +1. Then C 2 = ξ = ±1 corresponds to the case D = Cℓ±1 . There is a unique (φ, χ)
irrep W = C1|1 of U tw with actions:
! !
z1 ξ z̄2
C: → (2.163)
z2 z̄1
! !
z1 z̄1
T : → (2.164)
z2 z̄2
The supercommutant for the representation W ⊕n is given by matrices of the form (2.161)
but now with α, β real, so that it commutes with T . Therefore
!
0 β
E(Cℓ∓1 ) = { |β ∈ Mn (R) & β tr = −ξβ} (2.165)
−ξβ 0

Example 4: Finally, for T 2 = −1 and C 2 = ξ there is a unique (φ, χ)-irrep of U tw and we

can take W = H1|1 with complex structure I = L(i). Then
! !
q1 L(j)q1
T : → (2.166)
q2 L(j)q2

– 77 –
 ! !
q1 −ξL(j)q2
C: → (2.167)
q2 L(j)q1

so that C 2 = ξ = ±1, is odd, and commutes with T . The odd elements generating the
graded commutant are ! !
q1 ξq2 q
R(q) : → (2.168)
q2 q1 q
Taking R(i), R(j), R(k) we get 3 odd generators generating Cℓ±3 . It is useful to write the
commutant in terms of 2 × 2 complex matrices. To that end, we identify H ∼ = C2 via
q = z1 + z2 j where z1 = x1 + iy1 and z2 = x2 + iy2 with xi , yi real. Then
! !
z1 −z̄2
L(j) : → (2.169)
z2 z̄1

and right-multiplication by a quaternion is C-linear and hence represented by a 2×2 matrix.

The general (φ, χ)-representation of U tw is W ⊕n . It has real superdimension (4n|4n) and,
identifying H ∼
= C2 as above the general graded intertwiner takes the form
!
α β
(2.170)
ξβ α

where α, β are n × n matrices whose entries are right-multiplication by quaterions. In other

words, we can consider them to be 2n × 2n complex matrices which satisfy JβJ −1 = β ∗
and Jα∗ J −1 = α, where J = 1n ⊗ ǫ, as usual. Thus, for the case C 2 = ξ = ±1,
!
0 β
E(Cℓ∓3 ) = { |β ∈ Mn (H) & β † = ξβ} (2.171)
ξβ 0

Remark: We motivated the study of M2,2 and its subgroups using the example of dis-
ordered systems. Unfortunately, in the literature on this subject it is often assumed that
given a pair of homomorphisms
(φ, χ) : G → M2,2 (2.172)
such that τ · χ = φ, we will always have G ∼ = G0 × U , where U is a subgroup of M2,2 and
G0 is ker(φ) ∩ ker(χ). This is not true in general! There is a sequence

1 → G0 → G → U → 1 (2.173)

and in general it will not split, let alone be a direct product. ♣Should make the
point better by
taking an example
of such a bigraded
3. Lecture 3: Free Fermions, the Altland-Zirnbauer classification, and group G and a
(φ, χ)-rep and
Bott Periodicity working out the
ensemble of
Hamiltonians. ♣
3.1 Fermions and the Spin Representation
We now return to quantum mechanics.

– 78 –
 The central motivation for this chapter, in the context of these notes, is that important
examples of the 10-fold way described above are provided by free fermions. They also
appear in the Altland-Zirnbauer classification, and in applications to topological band
structure.
Of course, the basic mathematics of free fermion quantization is very broadly applica-
ble. In this chapter we give a summary of that quantization and comment on the relation
to the Spin group and spin representations.

3.1.1 Finite dimensional fermionic systems

A finite dimensional fermionic system (FDFS) is a quantum system based on a certain kind
of operator algebra and its representation:

Definition: A finite dimensional fermionic system is the following data:

1. A finite-dimensional real vector space M ∼

= RN , called the mode space with a positive
symmetric bilinear form Q.

2. An extension of the complex Clifford algebra

A = Cliff(M, Q) ⊗ C (3.1)

to a ∗-algebra.

3. A choice of Hilbert space HF together with a ∗ homomorphism of A into the algebra

of C-linear operators on HF . 22

Here are a number of remarks about this definition:

1. As an algebra A is the complex Clifford algebra of V := M ⊗ C with Q extended

C-linearly.

2. From (M, Q) we can make the real Clifford algebra Cliff(M, Q). In quantum me-
chanics we will want a ∗-algebra of operators and the observables will be the operators
fixed by the ∗-action. For us the ∗-algebra structure on

A := Cliff(M, Q) ⊗ C (3.2)

is β ⊗ C, where β is the canonical anti-automorphism of Cliff(M, Q) and C is complex

conjugation on C. Thus ∗ fixes M and is an anti-automorphism. (These conditions
uniquely determine ∗.) Axioms of quantum mechanics would simply give us some
∗-algebra without extra structure. The fermionic system gives us the extra data
(M, Q). ♣Use font for ∗
consistently with
discussion above. ♣
22
The subscript “F ” is for “Fock.”

– 79 –
 3. Since we have a ∗ structure on a Z2 -graded algebra we must deal with a convention
issue. Here we are taking the convention that (ab)∗ = b∗ a∗ for any a, b because this
is the convention almost universally adopted in the physics literature. However, a
systematic application of the Koszul sign rule in the definition of ∗ would require
(ab)∗ = (−1)|a|·|b| b∗ a∗ . One can freely pass between these two conventions and, if
used consistently, the final results are the same. See Section §2.2.5 above for more
discussion.

4. If Q is positive definite then we can diagonalize it to the unit matrix. If ei is a choice

of basis in which Q is δij then the usual Clifford relations

ei ej + ej ei = 2δij i, j = 1, . . . , N (3.3)

are known in this context as the fermionic canonical commutation relations. Be-
cause of our choice of ∗-structure we have e∗i = ei . Of course, the choice of basis
is far from unique. Different choices are related by O(N ) transformations. Those
transformations commute with the ∗ structure. The ei are known in the literature
as real fermions or Majorana fermions. In terms of the ei the most general quantum
observable is
Xd
O = O0 + Oi1 ...ik ei1 ...ik (3.4)
k=1

where the coefficients are totally antisymmetric tensors such that O0 ∈ R and
1
Oi∗1 ...ik = (−1) 2 k(k−1) Oi1 ...ik . (3.5)

5. In quantum mechanics we must also have a Hilbert space representation of the ∗-

algebra of operators so that ∗ corresponds to Hermitian conjugation in the Hilbert
space representation. That is, we have an algebra homomorphism

ρF : A → EndC (HF ) (3.6)

to the C-linear operators on the Hilbert space HF . The is a ∗-homomorphism in the

sense that ♣Here a is a
general element of
(ρF (a))† = ρF (a∗ ) (3.7) the algebra. Not a
fermion oscillator.
Bad notation. ♣
In the fermionic system we are assuming that HF is a choice of an irreducible module
for A. We will describe explicit models for HF in great detail below. (Of course, we
have already discussed them at great length - up to isomorphism.)

6. The notation N is meant to suggest some large integer, since this is a typical case in
the cond-matt applications. But we will not make specific use of that property.

3.1.2 Spin representations from complex isotropic subspaces

If v ∈ Rr,s then we can define the contraction operator on the exterior algebra of V by:

– 80 –
 k
X
i(v)(vi1 ∧ · · · ∧ vik ) := (−1)s−1 Q(v, vis )vi1 ∧ · · · ∧ vc
i s ∧ · · · ∧ vi k (3.8)
s=1

where the hat superscript vb means we omit that factor. Similarly, we can define the wedge
operator by
w(v)(vi1 ∧ · · · ∧ vik ) := v ∧ vi1 ∧ · · · ∧ vik (3.9)

These operators are easily shown to satisfy the algebra:

{i(v1 ), i(v2 )} = 0
{w(v1 ), w(v2 )} = 0 (3.10)
{i(v1 ), w(v2 )} = Q(v1 , v2 )

Let us assume dimR M is even so that N = 2n. The standard finite-dimensional ♣Put remark below
on the odd case. ♣
fermionic Fock space construction begins by choosing a complex structure I on M. As
shown in (??) above we automatically have

M⊗C=V ∼
=W ⊕W (3.11)

given by the projection operators P± = 12 (1 ± I ⊗ i). Here we take

W := P− M ⊗ C = SpanC {e − iIe|e ∈ M} (3.12)

W ∼
= P+ M ⊗ C = SpanC {e + iIe|e ∈ M} (3.13)

If a vector P− v = v then Iv = iv, so W is the (1, 0) space of V .

Now, we henceforth assume that I is compatible with the quadratic form Q so

Q(Iv1 , Iv2 ) = Q(v1 , v2 ) (3.14)

We then extend Q to be a symmetric C-linear form on V . Note that W is a maximal

dimension isotropic complex subspace of V . For if w1 , w2 ∈ W then

Q(w1 , w2 ) = Q(Iw1 , Iw2 ) = Q(iw1 , iw2 ) = i2 Q(w1 , w2 ) = −Q(w1 , w2 ) (3.15)

and hence Q(w1 , w2 ) = 0. Note we have crucially used the fact that the extension of Q is
C-linear.

Remark: Recall that the space of complex structures compatible with Q is a homogeneous
space CmptCplxStr(M, Q) isomorphic to O(2n)/U (n). (See (1.82) above.) Once we have
extended Q in this C-linear fashion we can also understand the space of complex struc-
tures as the Grassmannian of maximal dimension complex isotropic subspaces in V . This
interpretation is sometimes quite useful, especially in giving a geometrical interpretation
of the spin representation in Section ?? below. We denote this Grassmannian by G(V, Q).

– 81 –
 Now, given the decomposition V ∼= W ⊕ W̄ it is fairly evident how to take a “square-
root” of
Λ∗ V ∼
= Λ∗ W ⊗ Λ∗ W (3.16)
We could, for example, consider the vector space

Λ∗ W = ⊕nk=0 Λk W (3.17)

We can make this vector space into an irreducible Clifford module for Cliff(V, Q) by simi-
larly taking “half” of the representation (??):

1. For w ∈ W we define ρF,W (w) := w(w)

2. For w̄ ∈ W we define ρF,W (w̄) = i(w̄).

3. Now define ρF on V by extending the above equations C-linearly:

ρF,W (w1 ⊕ w2 ) := ρF,W (w1 ) + ρF,W (w2 ). (3.18)

Now one checks that indeed the Clifford relations are satisfied and ρF,W defines the
structure of a Clifford module on Λ∗ W . We will often denote this module as

HF,W := Λ∗ W (3.19)

and to lighten the notation we sometimes abbreviate ρF,W by ρF if W is understood or

drop it altogether if the context is clear.
In fact (ρF,W , HF,W ) is naturally a graded representation with
0
HF,W ∼
= Λeven W := ⊕k=0(2) Λk W (3.20)
1
HF,W ∼
= Λodd W := ⊕k=1(2) Λk W (3.21)
Now if we think of Spin(2n) as a group of invertible elements in

Spin(2n) ⊂ Cliff(M, Q) ⊂ Cliff(V, Q) (3.22)

then through ρF the group Spin(2n) acts on HF,W , but not irreducibly. The operations
of contraction and wedging with a vector change the parity of k, but Spin(2n) involves
the action of an even number of vectors so we see that, as a representation of Spin(2n),
HF,W ∼= Sc and this decomposes into:

Sc+ ∼
= Λeven W (3.23)

Sc− ∼
= Λodd W (3.24)
♣Need to check
conventions for
In the physical applications it is important to note that we can put an Hermitian chiral vs. antichiral!
♣
structure on V by defining the sesquilinear form

h(v1 , v2 ) := Q(v 1 , v2 ) (3.25)

– 82 –
where v is defined from the decomposition W ⊕ W . Note that V = W ⊕ W̄ is an orthogonal
Hilbert space decomposition: W and W̄ are separately Hilbert spaces and are orthogonal.
To prove this note that orthogonality follows since W and W̄ are isotropic with respect
to Q. Then since W is maximal isotropic and Q is nondegenerate the sesquilinear form
restricted to W must be nondegenerate. Moreover, since Q > 0, this defines a Hilbert
space inner product on V .
The Fock space HF,W now inherits a Hilbert space structure since we can define

h(w1 ∧ · · · ∧ wk , w1′ ∧ · · · ∧ wℓ′ ) := δk,ℓ deth(wi , wj′ ) (3.26)

for k, ℓ > 0. We extend this to Λ0 W by declaring it orthogonal to the subspaces Λk W with

k > 0 and normalizing:
h(1, 1) := 1 (3.27)
Note that ρF,W (ei ) are self-adjoint operators so that ρF,W is indeed a ∗ homomor-
phism, as desired. Moreover, ρF,W (eij ) are skew-adjoint. Therefore with this Hilbert space
structure HF,W is a unitary representation of Spin(2n). Indeed, the operators representing
the group Spin(2n) are of the form exp[ 12 ω ij ρF,W (eij )] with real ω ij .
The upshot of this discussion is the theorem:

Theorem: There is a bundle of Z2 -graded finite dimensional Hilbert spaces over

CmptCplxStr(M, Q) ∼
= G(V, Q) ∼
= O(2n)/U (n) (3.28)

whose fiber at a complex maximal isotropic subspace W ⊂ V is the fermionic Fock space

HF,W ∼
= Λeven W ⊕ Λodd W (3.29)

The homogeneous subspaces in the fibers are naturally unitary chiral representations of
the spin group Spin(2n).

Remarks:

1. The reader might well be wondering: “Why not choose Λ∗ W ?” Indeed that works
too. Exchanging I and −I is equivalent to exchanging W and W . So, with our
construction, ρF,W is simply the module we would get from complex structure −I.
The space of complex structures O(2n)/U (n) has two connected components. These
can be distinguished by the extra data of a choice of orientation of M. For example,
we could associate to any basis in which I is of the form
! !
0 1 0 1
I= ⊕ ··· ⊕ (3.30)
−1 0 −1 0

the orientation e1 ∧ e2 ∧ · · · ∧ e2n . Then, thanks to (1.78) above, this association is

well-defined. ♣Maybe this
discussion should be
in Section 1.6.1. ♣
2. More generally, what happens if we change the choice of W ? This is discussed in
Section 3.1.4 below.

– 83 –
3.1.3 Fermionic Oscillators
Now let us connnect the construction of the spin representation in 3.1.2 to the usual
discussion in the physics literature using fermionic harmonic oscillators. In particular, we
would like to justify the terminology “fermionic Fock space” for HF,W .
Given a complex structure I on M compatible with Q we can find an ON basis ei for
M such that

Ie2j−1 = −e2j
(3.31)
Ie2j = e2j−1 , j = 1, . . . , n

Put differently, the ordered basis:

{e}2n
α=1 := {e1 , e3 , . . . , e2n−1 , e2 , e4 , . . . , e2n }, (3.32)

is a basis in which !
0 1
I= (3.33)
−1 0

Once again: The choice of such a basis is far from unique. Different choices are related by ♣Opposite sign
from I0 in Section
a subgroup of O(2n) isomorphic to U (n), as described in Section §1.6.1. We will explore 1.6.1. ♣

this in detail below.

Then applying projection operators gives us a basis for W and W̄ , respectively:

1
āj = P− e2j−1 = (e2j−1 + ie2j )
2 (3.34)
1
aj = P+ e2j−1 = (e2j−1 − ie2j )
2

e2j−1 = aj + āj
(3.35)
e2j = i(aj − āj )
We easily compute the fermionic CCR’s in this basis to get the usual fermionic har-
monic oscillator algebra:

{aj , ak } = {āj , āk } = 0

(3.36)
{aj , āk } = δj,k

The space Λ∗ W has a natural basis 1, āj , . . . where the general basis element is given
by āj1 · · · ājℓ for j1 < · · · < jℓ . In particular, note that 23

ρF,W (ai ) · 1 = 0 (3.37)

where 1 ∈ Λ0 W ∼
= C. We build up the other basis vectors by acting with ρF,W (āj ) on 1.
The transcription to physics notation should now be clear. The vacuum line is the
complex vector space Λ0 W ∼
= C. Physicists usually choose an element of that line and
23
Note that, if we drop the ρF,W then the equation would be wrong!

– 84 –
denote it |0i. Moreover, they drop the heavy notation ρF,W , so, in an irreducible module
we have just
ai |0i = 0. (3.38)

The state |0i is variously called the Dirac vacuum, the Fermi sea, or the Clifford vacuum.
However, irrespective of whose name you wish to name the state after, it must be stressed
that these equations only determine a line, not an actual vector, and, when considering
families of representations this can be important. Indeed, some families of quantum field
theories are inconsistent because there is no way to assign an unambiguous vacuum vector
to every element in the family which varies with sufficient regularity.
In our case we have a canonical choice 1 ∈ Λ0 W ↔ |0i ∈ HF , where HF is our
notation for the fermionic Fock space. Then, Λk W is the same as the subspace spanned
by āj1 · · · ājk |0i.
In physical interpretations Λk W is a subspace of a Fock space describing states with
k-particle excitations above the vacuum |0i. It is very convenient to introduce the fermion
number operator
Xn
n iX
F := āi ai = − eα Iαβ eβ (3.39)
2 4
i=1 α,β

so that Λk W is the subspace of “fermion number k.”

The operator (−1)F commutes with the spin group and decomposes the Fock space
into even and odd subspaces. That is, the eigenspaces (−1)F = ±1 are isomorphic to the
chiral spin representations.
Finally, consider the Hilbert space structure. With respect to the Hilbert space struc-
ture (3.26) we find that indeed

ρF,W (āi ) = ρF,w (ai )† , (3.40)

so in physics we would just write āi → a†i . The normalization condition (3.27) is written
in physics notation as
h0|0i = 1. (3.41)

3.1.4 Bogoliubov transformations

We now return to the fact that we had to choose a complex structure to construct an
irreducible spin representation in Section §3.1.2. However, as we saw in equation (1.82)
above there is a whole family of complex structures I which we can use to effect the
construction. On the other hand, the irreducible spin representations Sc± are unique up
to isomorphism. Therefore there must be an isomorphism between the constructions: this
isomorphism is known in physics as a “Bogoliubov transformation.” It can have nontrivial
physical consequences.
To a mathematician, there is just one isomorphism class of chiral spin representation
Sc or Sc− (distinguished, invariantly, by the volume element). However, in physics, the
+

fermionic oscillators represent physical degrees of freedom: Nature chooses a vacuum, and
if, as a function of some control parameters a new vacuum becomes preferred when those

– 85 –
parameters are varied then the Bogoliubov transformation has very important physical
implications. A good example of this is superconductivity. ♣Explain
somewhere what a
Returning to mathematics, suppose we choose one complex structure I1 with a com- polarization is. ♣

patible basis {eα } satisfying (3.31) for I1 . Next, we consider a different complex structure ♣We need to
explain that
I2 with corresponding basis {fα } satisfying (3.31) for I2 . although Spin(2n)
acts on HF,W for
With the different basis {fα } we can form fermionic oscillators according to (3.34): fixed W we can also
make it (through
the map to O(2n))
1 act on the space of
b̄j = (f2j−1 + if2j ) complex structures.
2 (3.42)
So we could ask
how the actions on
1 different complex
bj = (f2j−1 − if2j ) structures is
2 defined. Thus, we
could also ask about
Now we must have a transformation of the form whether the bundle
HF → G(V, Q) can
be interpreted as an
b̄i = Aji āj + Cji aj equivariant bundle.
♣
(3.43)
bi = Bji āj + Dji aj 1 ≤ i, j ≤ n

Observation: For a general transformation of the form (3.43), with complex n×n matrices
A, B, C, D the fermionic CCR’s are preserved iff the matrix
!
A B
g= (3.44)
C D

satisfies: ! !
01 01
gtr g= (3.45)
10 10

That is

Atr D + C tr B = 1
Atr C = −(Atr C)tr (3.46)
tr tr tr
D B = −(D B)

Proof : The proof is a straightforward computation. ♦

The proposition characterizes the general matrices which preserve the CCR’s. We
recognize (3.45) as the definition of the complex orthogonal group for the quadratic form
!
01
q= (3.47)
10

So, Bogoliubov transformations can be effected by complex orthogonal transformations

O(q; C) := {g ∈ GL(2n; C)|g tr qg = q} (3.48)

The form q has signature (n, n) over the real numbers but is, of course, equivalent to the
standard Euclidean form over the complex numbers.

– 86 –
3.2 Free fermion dynamics and their symmetries
3.2.1 FDFS with symmetry
Finally, let us define formally what it means for a FDFS to have a symmetry.

Definition: Let (G, φ) be a Z2 -graded group with φ : G → Z2 . We will say that (G, φ)
acts as a group of symmetries of the FDFS if

1. There is a homomorphism ρ : G → AutR (HF ) of Z2 -graded groups. That is, HF is a

φ-rep of G. (See Section §1.7.1.)

2. There is a compatible automorphism α of the ∗-algebra A so that A is a φ-representation

of G. That is, α(g) is C-linear or anti-linear according to φ and ρ and α are compatible
in the sense that: ♣Again a for
generic element of
−1
ρ(g)ρF (a)ρ(g) = ρF (α(g) · a) (3.49) A is bad notation.
♣

3. The automorphism preserves the real subspace M ⊂ A, and hence we have a group
homomorphism: α : G → AutR (M, Q) = O(M, Q) ∼ = O(N ).

Remarks:

1. Assuming ρF is faithful and surjective (as happens for example if N is even and we
choose an irreducible Clifford module for HF ) the map a 7→ a′ defined by

ρ(g)ρF (a)ρ(g)−1 = ρF (a′ ) (3.50)

defines the automorphism of A. When A is a central simple algebra it must be inner.

The condition (3) above puts a further restriction on what elements we can conjugate
by.

2. We put condition (3) because we want the symmetry to preserve the notion of a
fermionic field. The mode space M is the space of real fermionic fields. It should
then preserve Q because we want it to preserve the canonical commutation relations.
In terms of operators on HF :
X
ρ(g)ρF (ej )ρ(g)−1 = Smj ρF (em ) (3.51)
m

where g 7→ S(g) ∈ O(N ) is a representation of G by orthogonal matrices.

3. When constructing examples it is natural to start with a homomorphism α : G →

O(N ). We then automatically have an extension to an automorphism of Cliff(M, Q).
There is no a priori extension to an automorphism of A. The data of the φ-
representation determines that extension because a 7→ ρF (a) is C-linear. It follows
that ρ(g) is conjugate linear iff α(g) is conjugate linear. This tells us how to extend
α to AutR (A).

– 87 –
Examples

1. By its very construction, the group G = Pin+ (N ) with φ = 1 is a symmetry group of

the FDFS generated by (M, Q) for M of dimension N . We can simply take ρ = ρF .
This forces us to take α = Ad. 24

2. What about G = Pin− (N )? In fact we can make G = Pinc (N ) (which contains both
Pin± (N ) as subgroups) act. We think of generators of Pinc (N ) as ζei where |ζ| = 1
is in U (1). Then ρ(ζei ) = ζρF (ei ) and α(ζei ) = Ad(ei ). Again we take φ = 1 in this
example.

3. Now we can ask what Z2 -gradings we can give, say, G = Pin+ (N ). Since we take
φ to be continuous φ = 1 on the connected component of the identity. Then if we
take φ(v) = −1 for some norm-one vector then if v ′ is any other norm-one vector
vv ′ ∈ Spin(N ) and hence φ(vv ′ ) = 1 so φ(v ′ ) = −1. Therefore the only nontrivial
Z2 -grading is given by the determinant representation described in (??) above. If we
use this then in general there is no consistent action of (G, φ) on the N -dimensional
FDFS. ♣Maybe when CℓN
has real reps it is
ok? ♣
4. To give a very simple example with φ 6= 1 consider N = 2, hence a single oscillator
a, ā and let G = Z4 = hT |T 4 = 1i. Then, in the explicit representation of ?? take

ρ(T )|+i = −|−i

(3.52)
ρ(T )|−i = |+i

and extend by linearity for φ(T ) = +1, and by anti-linearity for φ(T ) = −1, to define
ρ : G → AutR (HF ). In either case α(T ) · e1 = −e1 , but a small computation shows
that (
e2 φ(T ) = +1
α(T ) · e2 = (3.53)
−e2 φ(T ) = −1
Note that

α(T ) · a = −ā
(3.54)
α(T ) · ā = −a

in both cases φ(T ) = ±1.

5. Inside the real Clifford algebra generated by ei is a group EN generated by ei . This

group is discrete, has 2N +1 elements and is a nonabelian extension of ZN 2 with cocycle
−1 −1
determined from ei ej ei ej = −1 for i 6= j. EN is known as an extraspecial group.
Suppose Ti , i = 1, . . . , k with 2k ≤ N generate an extraspecial group Ek of order 2k+1 .
Thus, Ti2 = 1 and Ti Tj Ti−1 Tj−1 = −1 for i 6= j. Then there are many Z2 gradings
φ of Ek because we can choose the sign of φ(Ti ) independently for each generator.
For each such choice (Ek , φ) acts as a symmetry group of the N -dimensional FDFS.
24 f This leads to some important signs below.
Note that it is Ad and not Ad.

– 88 –
 Using the basis for the explicit representation of ?? we can take ρ(Ti ) = ρF (e2i−1 ).
Since the latter matrix is real the operators ρ(Ti ) can be consistently anti-linearly
extended in the basis of ??. A small computation shows that
(
−e2j φ(Ti ) = +1
α(Ti ) · e2j = (3.55)
e2j φ(Ti ) = −1

but (
ai j=i
α(Ti ) · āj = (3.56)
−aj j 6= i
(
āi j=i
α(Ti ) · aj = (3.57)
−āj j 6= i
independent of the choice of φ.
♣Now comment on
possibilities for φ
and what the Dyson
3.2.2 Free fermion dynamics classes would be.
♣
♣Put general
In general, the Hamiltonian is a self-adjoint element of the operator ∗-algebra and thus comments on
dynamics here? ♣
has the form (3.4). We will distinguish a ∗-invariant element h ∈ A from the Fock space
Hamiltonian H := ρF (h).
Usually, for reasons of rotational invariance, physicists restrict attention to Hamilto-
nians in the even part of the Clifford algebra, so then
X
h = h0 + hi1 ...ik ei1 ...ik (3.58)
k=0(2)

with h0 ∈ R and h∗i1 ...ik = (−1)k/2 hi1 ...ik . These elements generate a one-parameter group
of automorphisms Ad(u(t)) on A where u(t) = e−ith . Related to this is a one-parameter
group of unitary operators
U (t) = ρF (u(t)) = e−itH (3.59)

on HF representing time evolution in the Schrödinger picture.

In the Heisenberg picture Ad(u(t)) induces a one-parameter group of automorphisms
of the algebra of operators and in particular the fermions themselves evolve according to
√ X
u(t)−1 ei u(t) = ei + −1t hI [eI , ei ] + O(t2 ) (3.60)
I

where we have denoted a multi-index I = {i1 < · · · < ik }. Terms with k > 2 will clearly
not preserve the subspace M in A.
By definition, a free fermion dynamics is generated by a Hamiltonian h such that
Ad(u(t)) preserves the subspace M. (Note well, when expressed in terms of harmonic
oscillators relative to some complex structure it might or might not commute with F.)
The most general Hamiltonian defining free fermion dynamics is a self-adjoint element of

– 89 –
A = Cliff(M, Q) ⊗ C which can be written with at most two generators. Therefore, the
general free fermion Hamiltonian is
√
−1 X
h = h0 + Ajk ej ek (3.61)
4
i,j

where Aij = −Aji is a real antisymmetric matrix.

Remarks

1. Note well that Aij is an element of the real Lie algebra so(N ) and indeed
1X
Ajk ej ek (3.62)
4
j,k

is the corresponding element of spin(N ) ∼

= so(N ).

2. As we remarked, there are two Hilbert spaces associated to the fermionic system. In
the Fock space HF we have Hamiltonian
√
−1 X
H = h0 + Ajk ρF (ej ek ) (3.63)
4
i,j

and, up to a trivial evolution by e−ih0 t , the free fermion dynamics is the action of a
one-parameter subgroup U (t) of Spin(2N ) acting on the spin representation, in the
Schrödinger picture. In the Heisenberg picture the corresponding dynamical evolu-
tion preserves the real subspace M ⊂ A is given by the real vector representation:
f
Ad(u(t)).

3. Upon choosing a complex structure we have a second Hilbert space, the Dirac-Nambu
Hilbert space HDN := V ∼ = W ⊕ W̄ and, (only in the case of free fermion dynamics)
f
U (t) induces an action on V . This is simply Ad(u(t)) on M extended C-linearly to
V = M ⊗ C. The “Dirac-Nambu Hamiltonian” is therefore just ρDN (h) := Ad(h)
acting on V , thought of as a subspace of Cliff(V, Q).

4. Any real antisymmetric matrix can be skew-diagonalized by an orthogonal transfor-

mation. That is, given Aij there is an orthogonal transformation R so that
! !
0 λ1 0 λ n
RARtr = ⊕ ··· ⊕ (3.64)
−λ1 0 −λn 0

The Bogoliubov transformation corresponding to R can be implemented unitarily and

hence if h0 is zero then the spectrum of H b must be symmetric about zero. Therefore
this is a system in which it is possible to have symmetries with χ 6= 0. In this basis
we simply have (with h0 = 0)
X 1
h= λj āj aj − (λ1 + · · · + λn ) (3.65)
2
P
The spectrum of the Hamiltonian on HDN is {±λj } and on HF is { 21 i ǫi λi } where
ǫi ∈ {±1}.

– 90 –
 Exercise
Compute the time evolution on M of the one-parameter subgroup generated by the
self-adjoint operator ei . 25

3.2.3 Symmetries of free fermion systems

Now suppose we have a Z2 -graded group (G, φ) acting as a group of symmetries of a
finite dimensional fermion system. We therefore have the following data: (M, Q) together
with a ∗-representation of A = Cliff(V ; Q) on the Hilbert space HF together with the
homomorphisms α and ρ satisfying (3.49).
Suppose furthermore that we have a free fermionic system, hence a Hamiltonian of the
form (3.61).

Definition: We say that G is acting as a group of symmetries of the dynamics of the free
fermionic system if
ρ(g)U (s)ρ(g) = U (s)τ (g) (3.66)
for some homomorphism τ : G → Z2 . Here U (s) = exp[−isH/~] is the one-parameter time
evolution operator. If (3.66) holds then we declare g with τ (g) = −1 to be time-reversing
symmetries.

1. The above definition looks like a repeat of our previous definition of a symmetry
of the dynamics from Section §1.8. The data (M, Q, HF , G, φ, α, ρ, H) determine
ρ(g)Hρ(g)−1 . With our logical setup here, a symmetry of the fermionic system is a
symmetry of the dynamics if there is some homomorphism χ : G → Z2 so that

ρ(g)Hρ(g)−1 = χ(g)H (3.67)

Then because general quantum mechanics requires φτ χ = 1, we will declare g to

be time-orientation preserving or reversing according to τ (g) := φ(g)χ(g). This
logic is reversed from our standard approach where we consider φ determined by
an a priori given homomorphism G → Autqtm (PH) together with an a priori given
homomorphism τ determined by an a priori action on spacetime.

2. There will be physical situations, e.g. a single electron moving in a crystal where
there is an a priori notion of what time-reversing symmetries should be and how
they should act on fermion fields.

3. Let us see what the above definition implies for the transformation of the oscillators
f Choose an ON basis for (M, Q) satisfying (3.3). Then, in terms of
under Ad.
operators on HF : X ♣Maybe Smj
should be α(g)mj .
ρ(g)ρF (ej )ρ(g)−1 = Smj ρF (em ) (3.68) ♣

25
Answer : ej (t) = cos(2t)ej + i sin(2t)ei ej for j 6= i and ei (t) = ei .

– 91 –
 Or, equivalently: X
α(g) · ej = Smj em (3.69)
m
so
iX
ρ(g)Hρ(g)−1 = h0 + φ(g) (SAS tr )mn ρF (em en )
4 m,n (3.70)
= χ(g)H

This shows that

1. If χ(g) = −1 for any g ∈ G then h0 = 0.
2. The matrix A must satisfy

S(g)AS(g)tr = τ (g)A (3.71)

for all g ∈ G, where τ (g) is either prescribed, or deduced from τ = φ · χ, depending

on what logical viewpoint we are taking.
The condition (3.71) can be expressed more invariantly: Given α : G → O(M, Q)
there is an induced action Adα(g) on o(M, Q), and we are requiring that

Adα(g) A = τ (g)A. (3.72)

3.2.4 The free fermion Dyson problem and the Altland-Zirnbauer classification
There is a natural analog of the Dyson problem suggested by the symmetries of free
fermionic systems:
Given a finite dimensional fermionic system (M, Q, HF , ρF ) and a Z2 -graded group
(G, φ) acting as a symmetry on the FDFS via (α, ρ), what is the ensemble of free Hamil-
tonians for the FDFS such that (G, φ) is a symmetry of the dynamics?
Note well! We have changed the Dyson problem for the φ-rep HF of G in a crucial
way by restricting the ensemble to free fermion Hamiltonians.
Our analysis above which led to (3.71) above shows that the answer, at one level,
is immediate from (3.71): We have the subspace in o(Q; R) satisfying (3.71). Somewhat
surprisingly, this answer depends only on α and τ as is evident from (3.72). For a given τ
there can be more than one choice for φ and χ.
However, the answer can be organized in a very nice way as noticed by Altland and
Zirnbauer [4]: Such free fermion ensembles can be identified with the tangent space at the
origin of classical Cartan symmetric spaces. This result was proved more formally in a
subsequent paper of Heinzner, Huckleberry, and Zirnbauer [31]. We explain the main idea:
Let us consider two subspaces of o(2n; R):

k := {A|Adα(g) (A) = A} (3.73)

p := {A|Adα(g) (A) = τ (g)A} (3.74)

– 92 –
 p is of course the ensemble we want to understand. If τ = 1 it is identical to k but in
general, when τ 6= 1 it is not a Lie subalgebra of o(2n; R) because the Lie bracket of two
elements in p is in k. This motivates us to define a Lie algebra structure on
g=k⊕p (3.75)
by
[k1 ⊕ p1 , k2 ⊕ p2 ] := ([k1 , k2 ] + [p1 , p2 ]) ⊕ ([p1 , k2 ] + [k1 , p2 ]) (3.76)
One can check this satisfies the Jacobi relation.
Note that we have an automorphism of the Lie algebra which is +1 on k and −1 on p,
so this is a Cartan decomposition.
Both k and g are classical Lie algebras: This means that they are Lie subalgebras of
matrix Lie algebras over R, C, H preserving a bilinear or sesquilinear form.
To prove this for k: Note that we have a representation of G on o(M; Q) ∼ = o(2n; R).
If G is compact this representation must decompose into irreducible representations. The
group algebra is therefore a direct sum of algebras of the form nK(m) where K = R, C, H.
By the Weyl duality theorem (1.152),(1.153) the commutant is mK(n). Since k is, by
definition, the commutant, when restricted to each irreducible representation exp[k] must
generate a matrix algebra over R, C, H. Therefore, k is a classical Lie algebra.
A similar argument works to show that g is a classical Lie algebra. There is a Lie
algebra homomorphism
g → o(M, Q) ⊕ o(M, Q) (3.77)
given by
k ⊕ p → (k + p) ⊕ (k − p) (3.78)
Now, we can characterize g as the commutant of a representation of G on M ⊕ M given
by ! ♣Should α(g) be
denoted S(g)? ♣
α(g) 0
g 7→ τ (g) = 1 (3.79)
0 α(g)
!
0 α(g)
g 7→ τ (g) = −1 (3.80)
α(g) 0
We embed g into o(M) ⊕ o(M). The matrices in the commutant of the form x ⊗ 12 is
isomorphic to k and the matrices in the commutant of the image of G which are of the
form x ⊗ σ 3 is isomorphic to p. Hence k and g are both classical real Lie algebras. ♣explain why we
don’t need to worry
Next, note that the Killing form of o(M; Q) restricts to a Killing form on k and on g. about other kinds of
matrices in the
It is therefore negative definite. Hence the real Lie algebras k and g are of compact type. commutant. ♣

This proves the theorem of [31]:

Theorem: The ensemble p of free fermion Hamiltonians in Cliff(M, Q) ⊗ C compatible

with (α, τ ) is the tangent space at the identity of a classical compact symmetric space
G/K.

Remarks

– 93 –
 1. See Appendix A for background material on symmetric spaces.

2. Cartan classified the compact symmetric spaces. They are of the form G/K where G
and K are Lie groups. There are some exceptional cases and then there are several
infinite series analogous to the infinite series A, B, C, D of simple Lie algebras. These
can be naturally organized into a series of 10 distinct classical symmetric spaces.
Thus, the Altland-Zirnbauer argument provides a 10-fold classification of ensembles
of free fermionic Hamiltonians. This gives yet another 10-fold way! We will relate it
to the 10 Morita equivalence classes of Clifford algebras (and thereby implicitly to
the 10 real super-division algebras) below. That relation will involve K-theory.

3. It is possible to give many examples of this classification scheme. See [43] for some of
them. Using the description of the 10 classes given in (A.7) - (A.17) one can give a
description of the 10 AZ classes along the following lines. Recalling (??) we can, with
a suitable choice of complex structure as basepoint write the free fermion hamiltonian
as X 1X

h= Wij āi aj + Zij āi āj + Z̄ij aj ai (3.81)
2
i,j i,j
where Wij is hermitian and Zij is a complex antisymmetric matrix. Then the 10
cases correspond to various restrictions on Wij and Zij . See Table 1 of [55]. We will ♣Should probably
reproduce that table
give a uniform construction of 10 examples using some mathematics related to Bott here and explain
the entries in detail.
periodicity below. Actually, his table
is confusing because
in some cases Zij is
3.2.5 Analog for bosons not antisymmetric,
so they are not all
Free Hamiltonians FF Hamiltonians in
an oscillator basis.
The quantization of a bosonic system is very similar. The mode space M is now a real Clear this up. ♣

symplectic vector space.

Up to a constant the general free boson Hamiltonian is an element of A of the form

h = hij ei ej (3.82)

This should be ∗ invariant and hence hij must be a real symmetric matrix. Now, notice
that from (??) that we can therefore identify hJ with an element of the symplectic Lie
algebra. Thus,
The space of free boson Hamiltonians is naturally identified with sp(2n; R).
Analog of the AZ classification of free bosonic Hamiltonians
Now we define a symmetry of the bosonic dynamics to be a group G with ρ : G →
End(HF ) such that *****
An argument completely analogous to that for (3.71) applies. The symmetry operators
act by X
ρ(g)ρF (vj )ρ(g)−1 = Smj (g)ρF (vm ) (3.83)
m
where now S(g) ∈ Sp(2n; R). The result is that the symmetry condition is just that
A = hJ ∈ sp(2n; R) is in the space

p := {A ∈ sp(2n; R)|S(g)AS(g)−1 = χ(g)A} (3.84)

– 94 –
 For bosons the Hamiltonian will have an infinite spectrum. It is natural to assume
that the Hamiltonian is bounded below, in which case χ = 1. From a purely mathematical
viewpoint one could certainly consider quadratic forms with Hamiltonian unbounded from
above or below. Consider, e.g., the upside down harmonic oscillator. Thus, one could still
contemplate systems with χ 6= 1, although they are a bit unphysical.

3.3 Symmetric Spaces and Classifying Spaces

3.3.1 The Bott song and the 10 classical Cartan symmetric spaces
Now we will give an elegant description of how the 10 classical symmetric spaces arise
directly from the representations of Clifford algebras. This follows a treatment by Milnor
[38]. Then, thanks to a paper of Atiyah and Singer [11] we get a connection to the classifying
spaces of K-theory. Milnor’s construction was discussed in the context of topological
insulators by Stone et. al. in [48].
We begin by considering the complex Clifford algebra Cℓ2d . In this section we will
be considering the algebras and representations as ungraded. 26 Thus, there is a unique
d
irreducible representation with carrier space Sc = C2 . Give it the standard Hermitian
structure. We can then take the representation of the generators Ji = ρ(ei ) so that Ji2 = −1,
Ji† = −Ji and hence Ji are both in the unitary group and its Lie algebra.
Then we define a sequence of groups

G0 ⊃ G1 ⊃ G2 ⊃ · · · (3.85)

We take G0 = U (2r) where we have denoted 2d = 2r and we define

Gk = {g ∈ G0 |gJs = Js g s = 1, . . . , k} (3.86)

We claim that G1 ∼
= U (r) × U (r).
One very explicit way to see this is to note that we could represent
!
0 1r
ρ(e1 ) = J1 = i (3.87)
1r 0

and hence the matrices which commute with it are of the form
!
AB
(3.88)
B A

by a unitary transformation we can bring this to the form

!
A+B 0
(3.89)
0 A−B

hence such matrices are unitary iff (A ± B) are unitary. So the group of such unitary
matrices is isomorphic to U (r) × U (r), as claimed.
26
It would be interesting to give a parallel discussion from the Z2 -graded viewpoint.

– 95 –
 While G1 is isomorphic to U (r) × U (r) it is embedded in U (2r) slightly nontrivially.
Now, at the next step, we can take
!
0 −1r
ρ(e2 ) = J2 = (3.90)
1r 0

Then clearly G2 is the subgroup of matrices of the form

!
A 0
(3.91)
0 A

Now, further Clifford generators will act within the 11 and 22 blocks, so the process repeats:
This is the mod-two periodicity.
The above line of reasoning would be a little tedious in the case of the real Clifford
algebras, so we rederive the above with a new strategy (that will generalize nicely):
1. Use the representation theory of Clifford algebras to characterize the Wedderburn
type of the image ρ(Cℓk ).
2. Then use Weyl’s commutant theory to characterize the Wedderburn type of Z(ρ(Cℓk )) =
ρ(Cℓk )′ .
3. Then read off the group Gk = Aut(S) ∩ ρ(Cℓk )′ .
For example, let us rederive G1 this way: As an ungraded algebra Cℓ1 has two irreps
and so we can write Sc as a sum of ungraded irreps of Cℓ1 : S = r+ N1+ + r− N1− . The
projector to the isotypical components is P± = 12 (1 + ρ(e1 )) but the trace of ρ(e1 ) is zero,
so
Sc ∼
= rN1+ ⊕ rN1− (3.92)
and therefore the algebra ρ(Cℓ1 ) has Wedderburn type

rC ⊕ rC (3.93)

so the commutant must have Wedderburn type

C(r) ⊕ C(r) (3.94)

and the intersection with Aut(Sc ), which gives precisely G1 , must be

G1 ∼
= U (r) × U (r) (3.95)

Next for G2 , Cℓ2 ∼ = M2 (C), so ρ(Cℓ2 ) has Wedderburn type rC(2) and hence the
commutant is 2C(r) so the group G2 is isomorphic to U (r). As r is a power of 2 we clearly
have periodicity so that our sequence of groups is isomorphic to

U (2r) ⊃ U (r) × U (r) ⊃ U (r) ⊃ · · · (3.96)

The successive quotients give the two kinds of symmetric spaces U (2r)/(U (r) × U (r)) and
(U (r) × U (r))/U (r). ♣Don’t get
Grassmannians
Now let us move on to the real Clifford algebra Cℓ−8d . We choose a real graded Grn,m with n 6= m.
♣
irreducible representation, R2N , with 2N = 24d . It is convenient to define an integer r by

– 96 –
2N = 16r. Again, we will regard the Clifford algebras as ungraded and the representation
S∼= R2N . Denote the representations of the generators Ji = ρ(ei ), so of course

Js Jt + Jt Js = −2δs,t . (3.97)

We can give S a Euclidean metric so that the representation of Pin− (8d) is orthogonal.
Therefore, Ji† = −Ji , so Jitr = −Ji , and hence Ji ∈ o(2N ). However, since Ji2 = −1 we
have Jitr = Ji−1 and hence we also have Ji ∈ O(2N ).
Now we define a sequence of groups

O(2N ) := G0 ⊃ G1 ⊃ G2 ⊃ · · · (3.98)

These are defined for k > 0 again by taking the commutant with Cℓ−k :

Gk = {g ∈ G0 |gJs = Js g s = 1, . . . , k} = Aut(S) ∩ Z(ρ(Cℓ−k )). (3.99)

Now, we claim that the series of groups is isomorphic to

O(16r) ⊃ U (8r) ⊃ Sp(4r) ⊃ Sp(2r) × Sp(2r) ⊃ Sp(2r) ⊃

(3.100)
⊃ U (2r) ⊃ O(2r) ⊃ O(r) × O(r) ⊃ O(r) ⊃ · · ·

We will show that this follows easily from the basic Bott genetic code:

R, C, H, H ⊕ H, H, C, R, R ⊕ R, R, · · · (3.101)

The argument proceeds along the lines of the three steps outlined above for the complex
case:

1. Determine the Wedderburn type of ρ(Cℓ−k ): Decompose S in terms of ungraded

irreps of Cℓ−k . For k 6= 3mod4 there is a unique irrep Nk up to isomorphism,
and for k = 3mod4 there are two Nk± . Therefore, S ∼ ⊕s
= Nk k for k 6= 3mod4 and
S ∼= (Nk+ )⊕sk ⊕ (Nk− )⊕sk for k = 3mod4. The number of summands is the same
±
Nk for k = 3mod4 because the decomposition is effected by the projection operator
using the volume form P± = 12 (1 ± ωk ) and TrS (ωk ) = 0 for all k. Now, the image of
the Clifford algebra in End(S) (as an ungraded algebra) will have be isomorphic to
sk K(tk ) for k 6= 3mod4 and sk K(tk ) ⊕ sk K(tk ) for k = 3mod4.

2. Apply the Weyl commutant theorem: Z(ρ(Cℓ−k )) will be isomorphic to tk K(sk ) for
k 6= 3mod4 and tk K(sk ) ⊕ tk K(sk ) for k = 3mod4.

3. Determine the group which can generate this commutant: When we intersect Z(ρ(Cℓ−k ))
with Aut(S) ∼
= O(16r) we get the group Gk . In this way we determine the following
table

– 97 –
 k Bott clock ρ(Cℓ−k ) Z(ρ(Cℓ−k )) Gk
0 R 16rR R(16r) O(16r)
1 C 8rC C(8r) U (8r)
2 H 4rH Hopp (4r) Sp(4r)
3 H⊕H 2rH ⊕ 2rH Hopp (2r) ⊕ Hopp (2r) Sp(2r) × Sp(2r)
4 H 2rH(2) 2Hopp (2r) Sp(2r)
5 C 2rC(4) 4C(2r) U (2r)
6 R 2rR(8) 8R(2r) O(2r)
7 R⊕R rR(8) ⊕ rR(8) 8R(r) ⊕ 8R(r) O(r) × O(r)
8 R rR(16) 16R(r) O(r)

We should stress that the entries for ρ(Cℓ−k ), Z(ρ(Cℓ−k )), and Gk just give the iso-
morphism type. Of course, r is some power of 2 and for large d we can repeat the periodic
sequence down many steps. ♣NEED TO DO
COMPLEX CASE
The series of homogeneous spaces Gk /Gk+1 for k ≥ 0 provide examples of the Cartan OF CARTAN
EMBEDDING ♣
symmetric spaces (for ranks which are a power of two!). Note that the tangent space at
1 · Gk+1 has an elegant description. First define g0 := T1 G0 = so(2N ). Now for k > 0 ♣Use so(2N) for
the Lie algebra to
define: distinguish it
clearly from the
gk := T1 Gk = {a ∈ so(2N )|aJs = Js a s = 1, . . . , k}. (3.102) group O(2N). ♣

−1
Observe that, for k ≥ 0, the map θk (a) = Jk+1 aJk+1 acts as an involution on gk and that
the eigenspace with θk = +1 is just gk+1 . Therefore we can identify

pk := TGk+1 Gk /Gk+1 = {a ∈ so(2N )|aJs = Js a s = 1, . . . , k & aJk+1 = −Jk+1 a}

(3.103)
so that
gk = gk+1 ⊕ pk (3.104)

3.3.2 Cartan embedding of the symmetric spaces

The involution θk described above extends to a global involution θk : Gk → Gk defined by
conjugation with Jk+1 :
−1
θk (g) = Jk+1 gJk+1 (3.105)
Of course, the fixed subgroup of θk in Gk is Gk+1 so the Cartan symmetric space is
Gk /Gk+1 . Moreover, the Cartan embedding of this symmetric space is just

Ok = {g ∈ Gk |θk (g) = g −1 } ⊂ Gk ⊂ O(2N ) k ≥ 0. (3.106)

Let us unpack this definition: The condition θk (g) = g−1 is equivalent to the condition
(Jk+1 g)2 = −1. Therefore, writing g̃ = Jk+1 g we can also write

Õk := Jk+1 Ok = {g̃ ∈ O(2N )|g̃ 2 = −1 {g̃, Js } = 0 s = 1, . . . , k} (3.107)

– 98 –
The map g 7→ g̃ = Jk+1 g is a simple diffeomorphism so Õk ∼= Gk /Gk+1 , and Õk is also
embedded in O(2N ). This manifestation of the homogeneous space will be more convenient
to work with. Note that:
Õ0 = {g̃ ∈ O(2N )|g̃ 2 = −1} (3.108)
and
Õ0 ⊃ Õ1 ⊃ Õ2 ⊃ · · · (3.109)
When we wish to emphasize the dependence on N we will write Õk (N ). It will also be
convenient to define Õ−1 := O(2N ).
Since g = ±1 is in Ok we have ±Jk+1 ∈ Õk , as is immediately verified from the
definition. (Note that ±1 are not elements of Õk .) Let us compute the tangent space to
Õk at Jk+1 . A path through Jk+1 must be of the form Jk+1 eta where a ∈ T1 Ok = pk .
Therefore there is an isomorphism T1 Ok ↔ TJk+1 Õk given simply by left-multiplication
by Jk+1 . Now a ∈ T1 Ok iff atr = −a, [a, Js ] = 0 for s = 1, . . . , k and {a, Jk+1 } = 0 and
therefore

p̃k := TJk+1 Õk = {ã ∈ so(2N )|{ã, Js } = 0, s = 1, . . . , k + 1} (3.110)

(A slightly tricky point here is that Jk+1 a is itself an anti-symmetric matrix, and hence in
the Lie algebra so(2N ).)

3.3.3 Application: Uniform realization of the Altland-Zirnbauer classes

The characterization (3.107) of p̃k is nicely suited to a realization of 8 of the 10 AZ classes
of free fermion Hamiltonians. We take a FDFS based on M = R2N with Q the Euclidean
metric. We take as our symmetry group G = Pin− (k + 1) with Clifford generators Ti .
We choose the nontrivial option for τ on G, thus τ (Ti ) = −1 for i = 1, . . . , k + 1. For
α we choose the embedding of G into O(2N ) using α(Ti ) = Ad(ei ) (not Ad) f acting on
M ⊂ Cℓ−8d . Comparing the definitions (3.73) and (3.74) we find that we have precisely

k = gk
(3.111)
p = p̃k

thus neatly exhibiting examples of 8 of the AZ 10 classes.

The remaining two AZ classes follow from completely analogous manipulations for the
series U (2r) ⊃ U (r) × U (r) ⊃ U (r) ⊃ · · · .
Remarks:

1. Note that our fermionic oscillators are a basis for the spin representation of Spin(8d).
So their Hilbert space will be a representation of the much larger group Spin(2N ) of
8r 4d−1
dimension 2N = 22 = 22 .

2. This example can be extended to compute the 3- and 10-fold classes on HDN and HF .
Again there are two options (φ(Ti ) = +1, χ(Ti ) = −1) and (φ(Ti ) = −1, χ(Ti ) = +1).
Representing the Ji by real matrices on HF we can take ρ = ρF restricted to Cℓ−k−1 .
♣Finish by
analyzing the
classes with this
choice. ♣

– 99 –
3.3.4 Relation to Morse theory and loop spaces
The homogeneous spaces Õk have a further beautiful significance when we bring in some
ideas from Morse theory.
We consider the quantum mechanics of a particle moving on these manifolds using the
action Z
dq
S[q] = − dtTr(q −1 )2 (3.112)
dt
where q(t) is a path in the orthogonal group or one of the Õk .
We begin with quantum mechanics on SO(2N ). We choose boundary conditions and
define P−1 to be the space of (continuously differentiable) paths q : [0, 1] → SO(2N ) such
that q(0) = +12N and q(1) = −12N . We are particularly interested in the minimal action
paths. Such paths will be geodesics in the left-right-invariant metric. The geodesics are
well known to be of the form q(t) = exp[πtA] with A ∈ so(2N ). We can always conjugate
A to the form ! !
0 a1 0 aN
⊕ ··· ⊕ (3.113)
−a1 0 −aN 0
P
where ai ∈ R. This has action 2π 2 a2i and the boundary conditions imply that ai are
odd integers. Therefore the minimal action paths have ai = ±1 and hence the space of
minimal action paths is precisely given by the conjugacy class of A ∈ o(2N ) with A2 = −1.
Moreover, such paths have a very simple form:

q(t) = cos πt + A sin πt (3.114)

Now, notice a trivial but significant fact:

1. If g ∈ O(2N ) is an orthogonal matrix and g2 = −1 then g ∈ so(2N ) is also in the Lie

algebra. Proof: g2 = −1 and g ∈ O(2N ) implies g tr = −g.

2. If A ∈ so(2N ) is in the Lie algebra and A2 = −1 then A ∈ O(2N ) is also in the Lie
group. Proof: Atr = −A and A2 = −1 implies Atr = A−1 .

Therefore, the space of minimal action paths in P−1 is naturally identified with

Õ0 := {g ∈ O(2N )|g 2 = −1} ⊂ O(2N ). (3.115)

Of course Õ0 is J1 O0 where O0 is the Cartan embedding of G0 /G1 = O(2N )/U (N ).

Now let us consider the quantum mechanics of a particle on the orbit Õ0 , again with the
action (3.112). We choose boundary conditions so that P0 consists of maps q : [0, 1] → Õ0
such that q(0) = J1 and q(1) = −J1 . The solutions to the equations of motion are of the
form 27 q(t) = J1 exp[πtA] where now A ∈ p0 implies {A, J1 } = 0, which guarantees that
the path indeed remains in Õ0 . Again, the boundary conditions together with the minimal
action criterion implies that A2 = −1, so we can write:

q(t) = J1 exp[πtA] = J1 cos πt + (J1 A) sin πt = J1 cos πt + Ã sin πt (3.116)

27
We use the fact that Õ0 is totally geodesic.

– 100 –
Because A ∈ p0 both A and à = J1 A are both antisymmetric and square to −1: A2 =
−1 = Ã2 . We can therefore consider A and à to be in O(2N ) and hence the minimal
action paths on Õ0 are parametrized by A, or better by Ã, and hence the space of minimal
actions paths is naturally identified, via the mapping q(t) 7→ Ã with Õ1 ⊂ O(2N ).
This stunningly beautiful pattern continues: We take

Pk := {q : [0, 1] → Õk |q(0) = Jk+1 q(1) = −Jk+1 } (3.117)

Since Õk is totally geodesic the solutions to the equations of motion are of the form
Jk+1 exp[πAt] with A ∈ pk . The minimal action paths have A2 = −1 and hence they
are of the form

q(t) = Jk+1 exp[πtA] = Jk+1 cos πt + (Jk+1 A) sin πt = Jk+1 cos πt + Ã sin πt (3.118)

But now Ã2 = −1 and à ∈ Õk+1 , so we can identify the space of minimal action paths in
Pk with Õk+1 .

The space of minimal action paths in the set Pk of all smooth paths [0, 1] → Õk from Jk+1
to −Jk+1 is naturally identified with Õk+1 by equation (3.118).

Figure 12: The minimal length geodesics on S N from the north pole to the south pole are
parametrized by S N −1 . Similarly, the geodesics in Õk from Jk+1 to −Jk+1 are parametrized by
Õk+1 .

Remark: A good analogy to keep in mind is the length of a path on the N -dimensional
sphere. If we consider the paths on S N from the north pole N to the south pole S then

– 101 –
the minimal length paths are great circles and are hence parametrized by their intersection
with the equator S N −1 . See Figure 12.
The great significance of this comes about through Morse theory. The action (3.112)
for the paths is a (degenerate) Morse function on Pk and the critical manifolds allow us to
describe the homotopy type of Pk . One considers a series of “approximations” to Pk by
looking at paths with bounded action:

PkΛ := {q ∈ Pk |S[q] ≤ Λ} (3.119)

As we have seen, the minimal action space is PkΛmin ∼

= Õk (N ) ⊂ O(2N ). Now – it turns out
– that the solutions of the equations of motion which are non-minimal have many unstable
modes. The number of unstable modes is the “Morse index.” The number of unstable
modes is linear in N . The reason this is important is that in homotopy theory the way PkΛ
changes as Λ crosses a critical value Λ∗ is

PkΛ∗ +ǫ ∼ (PkΛ∗ −ǫ × D n− )/ ∼ (3.120)

where n− is the number of unstable modes at the critical value Λ∗ and D n− is a ball
of dimension n− . This operation does not change the homotopy groups πj for j < n− .
Therefore, in this topological sense, Õk (N ) gives a “good approximation” to Pk (N ). ♣Need a separate
subsubsection on
On the other hand, the spaces Pk have the same homotopy type as the based loop how Morse function
gives a cell
spaces Ω∗ Õk . Indeed, choosing any standard path from −Jk+1 to Jk+1 we can use it to decomposition. ♣

convert any path in Pk to a loop Ω∗ Õk based, say, at Jk+1 by composition. Conversely,
composing the (inverse of) the standard path with any loop gives a path in Pk .
Putting these remarks together we can say that

Ω∗ Õk (∞) ∼ Õk+1 (∞) (3.121)

In the case of the complex Clifford algebras we have an even simpler slogan:

A loop of projectors is a unitary; A loop of unitaries is a projector.

From this discussion we get two nice outcomes:

1. We get a nice proof of Bott periodicity from the periodicity of the Clifford algebras
[38].

2. We thereby make a connection to generalized cohomology theory through the notion

of a spectrum.
♣Both of these
points require a lot
more explanation.
Definition: A sequence of spaces with basepoint, {Eq }q∈Z , is called a loop spectrum if ♣

there are homotopy equivalences

Eq → Ω∗ Eq+1 . (3.122)

– 102 –
 This is an important concept in algebraic topology. Given a loop spectrum one can
define generalized cohomology groups E q (X) := [X, Eq ], and conversely, given a generalized
cohomology theory there is a corresponding loop spectrum.
We see that the sequence of spaces . . . , U, (U ×U )/U, U, (U ×U )/U, . . . form a spectrum ♣Do we want
BU × Z. Comment
known as the complex K-theory spectrum. Similarly, the sequence of spaces on the index in
U (2r)/(U (r + x) ×
U (r − x)). Then
what is the
. . . , O/U, U/Sp, Sp/Sp × Sp, Sp × Sp/Sp, Sp/U, U/O, O/O × O, O × O/O, . . . (3.123) basepoint? ♣

forms the real K-theory spectrum. In the next section we will discuss a very nice realization
of these spectra in terms of operators on Hilbert space.

3.3.5 Relation to classifying spaces of K-theory

The fact that the Morse index for the space of paths Pk (N ) (where the N -dependence
comes from the fact that the paths are in Õk (N ) ⊂ O(2N ) ) grows linearly in N suggests
that it will be interesting to take the N → ∞ limit. We can do this as follows:
We make a real Hilbert space by taking a countable direct sum of copies of simple
modules of the real Clifford algebra Cℓ−(k+1) . Specifically we define, for k ≥ 0, 28
(
k Nk+1 ⊗ ℓ2 (R) k 6= 2(4)
HR := + −
(3.124)
(Nk+1 ⊕ Nk+1 ) ⊗ ℓ2 (R) k = 2(4)

k (n) be the sum of the first n representations N +

and for an integer n let HR k+1 or (Nk+1 ⊕ ♣Atiyah-Singer just
− write HR with no k
Nk+1 ). Now define a subspace of the space of orthogonal operators Ωk (n) ⊂ O(HR k ). These
since all separable
Hilbert spaces are
are operators which satisfy the following three conditions: isomorphic, but the
isomorphism
between different k
k (n) and Hk (n)⊥ .
1. They preserve separately HR is slightly
R nontrivial... ♣

k (n)⊥
2. They are just given by A = Jk on HR
k (n) they satisfy:
3. On HR

A2 = −1
(3.125)
{A, Ji } = 0 i = 1, . . . , k − 1

We recognize that Ωk (n) ∼

= Õk−1 (N ) where N and n are linearly related. (We define
Õ−1 (N ) := O(2N ) so this holds for k ≥ 0.) Now, from this description it is easy to see
that there are embeddings
Ωk (n) ֒→ Ωk (n + 1) (3.126)

and we can take a suitable “n → ∞ limit” and norm closure to produce a set of operators
Ωk (∞) on HR k . In [11] Atiyah and Singer show that this set of operators is closely related

to a set of Fredholm operators on HR k.

28
Recall that Nk denote irreducible ungraded Clifford modules.

– 103 –
 Define F0 to be the set of all Fredholm operators on HR
k , and let F1 ⊂ F0 denote the

subspace of skew-adjoint Fredholm operators: Atr = −A. (Formally, this is the Lie algebra
k ). ) Now for k ≥ 2 define Fk ⊂ F1 to be the subspace such that 29
of O(HR

T Ji = −Ji T i = 1, . . . , k − 1 (3.127)

Now, the space of Fredholm operators has a standard topology using the operator
norm topology. Using this topology Atiyah and Singer prove ♣You are changing
k’s here. Need to
clarify. ♣

1. Fk ∼ Ωk−1 (∞) ∼
= Õk−2 (∞), k ≥ 1, where ∼ denotes homotopy equivalence.

2. Fk+1 ∼ ΩFk , and in fact, the homotopy equivalence is given by

A 7→ Jk+1 cos πt + A sin πt 0≤t≤1 (3.128)

which should of course be compared with (3.118).

The relation to Fredholm operators implies a relation to K-theory because one way of
defining the real KO-theory groups of a topological space X is via the set of homotopy
classes:
KO−k (X) := [X, Fk ] k≥0 (3.129)
We summarize with a table

k Fk ∼ Gk−2 /Gk−1 Cartan’s Label

0 (O/(O × O)) × Z BDI
1 O D
2 O/U DIII
3 U/Sp AII
4 (Sp/(Sp × Sp)) × Z CII
5 Sp C
6 Sp/U CI
7 U/O AI

and the complex case is

29
For k = 3mod4 the subspace of F1 satisfying (3.127) in fact has three connected components in the
norm topology. Two of these are contractible but one is topologically nontrivial and we take Fk to be that
component. In fact for T satisfying (3.127) one can show that ωk−1 T is self-adjoint, where ωk−1 = J1 · · · Jk−1
is the volume form. The contractible components are those for which ωk−1 T is positive or negative - up to
a compact operator.

– 104 –
 k Fkc ∼ Gk−2 /Gk−1 Cartan’s Label
0 (U/(U × U )) × Z AIII
1 U A

where Fc is the space of Fredholm operators on a complex separable Hilbert space, and is
the classifying space for K 0 (X) and F1c is the subspace of skew-adjoint Fredholm operators
and is the classifying space for K −1 (X).
Remark: We indicate how this discussion of KO(X) is related to what we discussed
in Section §2.3.7 above. We take X = pt. Then, KO0 (pt) ∼ = Z. In terms of Fredholm
operators T the isomorphism is given by T 7→ Index(T ) := dimkerT − dimcokT . Thus,
“invertible part of T cancels out.” The idea that if T is invertible then it defines a trivial
class was the essential idea in the definition in Section §2.3.7. It is also worth noting the ♣Improve this
discussion by
Fredholm interpretation of KO−1 (pt) ∼ = Z2 in this context. For a skew-adjoint Fredholm rephrasing the AS
† results in terms of
operator ker(T ) = ker(T ) so the usual notion of index is just zero. However we can form Z2 -graded Hilbert
spaces. ♣
the “mod-two index,” which is defined to be dimkerT mod2. This is indeed continuous in
the norm topology and provides the required isomorphism.

4. Lecture 4: K-theory classification and application to topological band

structure

4.1 Reduced topological phases of a FDFS and twisted equivariant K-theory of

a point
4.1.1 Definition of G-equivariant K-theory of a point
Now let G be a compact Lie group. (It could be a finite group.)
KG (pt) is the representation ring of G. It can be defined in two ways:
1. Group completion of the monoid of finite-dimensional complex representations: A
typical element is a formal difference R1 −R2 with R1 , R2 , finite-dimensional representations
on complex vector spaces.
2. Divide by trivial representations: Define Reps (G) to be Z2 -graded fin. dim. cplx
reps (with even G-action) Define the submonoid: Trivs (G) to be those with an odd invert-
ible superlinear transformation T with

T ρ(g) = ρ(g)T (4.1)

Then we define:
KG (pt) = Reps (G)/Trivs (G) (4.2)
An important point is that KG (pt) is not just a group but a ring, because we can take
the tensor product of G-representations, and if we take a product of a representation with

– 105 –
a representation in Trivs (G) then the resulting representation is in Trivs (G). We will need
this in our discussion of localization below.
As an example, let us return to the automorphism groups of two-dimensional quantum
systems. We have
1 → U (1) → U (2) ⋊ Z2 → SO(3) ⋊ Z2 → 1 (4.3)
where the Z2 actions on U (2) by complex conjugation and on SO(3) by **** ♣work out z2 action
on so3 ♣
It is interesting to compare the K theories of the groups, but to do that we must use
pullback by the obvious homomorphisms to compare

ι∗1 : KU (2) ֒→ KSU (2)×U (1)

(4.4)
ι∗2 : KSO(3)×U (1) ֒→ KSU (2)×U (1)

Now, the representation ring of SU (2) × U (1) is generated by f , the two-dimensional

representation of SU (2) and δ±1 , the charge ±1 representations of U (1), so

KSU (2)×U (1) ∼

= Z[f, δ±1 ] (4.5)

In these terms we can say that the images are:

ι∗2 KSO(3)×U (1) ∼

= Z[δ±1 , f 2 − 1] (4.6)

ι∗1 KU (2) ∼
= Z[δ±1 , f ]+ (4.7)
where f represents the class of the two-dimensional fundamental representation of SU (2),
δn is the charge n representation of U (1), and the superscript + means we project onto the
even elements under f → −f and δ → −δ (so that the subgroup generated by (−1, −1) in
SU (2) × U (1) acts trivially.

4.1.2 Definition of twisted G-equivariant K-theory of a point

There is a general notion of a “twisting” of a generalized cohomology theory. This can be
defined in terms of some sophisticated topology (like using nontrivial bundles of spectra)
but in practice it often amounts to introducing some extra signs of phases. This is not
always the case: Degree shift in K-theory can be viewed as an example of twisting.
A simple example of a twisting of ordinary cohomology theory arises when one has a
double cover π : X̃ → X. Then the “twisted cohomology” of X refers to using cocycles,
coboundaries, etc. on X̃ that are odd under the deck transformation.
In the case of equivariant K-theory of a point, a “twisting of KG (pt)” is an isomorphism
class of a central extension of G by C∗ (or by U (1), if we include an Hermitian structure).
See appendix §B.1 for a description of central extensions. It can be constructed using a
C∗ - or U (1)-valued function λ(g1 , g2 ) on G × G satisfying the 2-cocycle condition

λ(g2 , g1 )λ(g3 , g2 g1 ) = λ(g3 , g2 )λ(g3 g2 , g1 ) (4.8)

Isomorphic central extensions induce a suitable equivalence such that equivalence class of
λ are classified by H 2 (G, U (1)). In general, isomorphism classes of twistings of K-theory
are classified by certain cohomology groups.

– 106 –
 Let τ denote such a class of central extensions. A “twisted G-bundle over a point with
twisting τ ” is a representation of a corresponding central extension Gτ . The τ -twisted
G-equiviariant K-theory of a point is then just the Gτ -equivariant K-theory of a point:
τ
KG (pt) = KGτ (pt) (4.9)

Now, when we add the other ingredients that we encountered from a general realization
of a quantum symmetry with gapped Hamiltonians we have a pair of homomorphisms
(φ, χ) : G → M2,2 :
1. The data of φ indicates that we take not central extensions but φ-twisted extensions.
For these the cocycle relation is modified to

λ(g2 , g1 )λ(g3 , g2 g1 ) = φ(g1 ) λ(g3 , g2 )λ(g3 g2 , g1 ) (4.10)

where λ are valued in U (1) and φ(g) λ is defined in (1.56). (If we take λ ∈ C∗ then instead
of λ̄ use λ−1 .) Given an isomorphism class τ of such a φ-twisted extension we can define
the monoid of representations, Reps (Gτ , φ, c) as usual.
2. The data of χ enters in modifying the definition of the submonoid of trivial repre-
sentations. These are the ones with an odd automorphism P such that

P ρτ (g) = χ(g)ρτ (g)P (4.11)

We can now define the more general twisted-equivariant K-theories:

φ τ,χ
KG (pt) := Reps (Gτ , φ, c)/Triv s (Gτ , φ, c) (4.12)

In general this is not a ring. It will be a module for KG0 (pt) where G0 is an untwisted
quotient of Gτ .
Example Take G = M2 = {1, T̄ }.
As we have seen when φ(σ) = −1 there are two inequivalent φ-twisted extensions by
U (1):
1 → U (1) → Pin+ (2) → M2 → 1 (4.13)
1 → U (1) → Pin− (2) → M2 → 1 (4.14)
Now, the untwisted K theory of Z2 is

KZ2 (pt) ∼
= Z[ǫ]/(ǫ2 − 1) (4.15)

where ǫ is the sign representation. The two twisted versions are:

τ
φ
KZ+2 (pt) ∼
= Z[f ] (4.16)

where f ∼
= R2 is a real representation, and
τ
φ
KZ−2 (pt) ∼
= Z[q]− (4.17)

where q = H are the quaternions, the complex structure is L(i), we choose a lift T of T̄ to
act L(j) and the superscript means we must take odd powers. The reason is that T must
act anti-linearly.

– 107 –
4.1.3 Appliction to FDFS: Reduced topological phases
Now let us return to our finite dimensional fermionic systems.
We take G to be a compact group and we restrict to a finite dimensional Hilbert space.
If we have not put any locality conditions on the Hamiltonians Aij then we simply
have
T Pfinite (Gτ , φ, χ) = Reps (Gτ , φ, χ) (4.18)

The proof is simply that for gapped Hamiltonians we can always homotope to H 2 = 1
and (twisted) representations of compact groups are discrete.
Now comes an important point: We would like to put an abelian group structure on
the set of phases. To do that, we need a monoid structure.
For free fermions we can use the monoid structure provided by the “single-particle”
Dirac-Nambu Hilbert space:
If we combine the systems 1 and 2 then

DN
H1+2 = H1DN ⊕ H2DN (4.19)

Then, given this monoid structure we can group-complete or quotient by a suitable

notion of “topologically trivial systems.” In this was we can get an abelian group of
“reduced” topological phases RT P(Gτ , φ, χ).

Remark: This raises the obvious question of what happens more generally when the
Dirac-Nambu structure is not available. There is in fact a general monoid structure on
quantum systems given by (graded) tensor product of Hilbert spaces:

b 2
H1+2 = H1 ⊗H (4.20)

The corresponding K-theory hasn’t been much studied in the literature, and with good
reason. Part of the problem is that the group completion can be fairly trivial. Consider
for example the K-theory of a point with this monoid structure using group completion.
Then (V1 , V2 ) ∼ (V3 , V4 ) if V1 ⊗ V4 ∼
= V2 ⊗ V3 . Instead of producing the integers, we produce
the abelian group of nonzero rational numbers under multiplication. This does not have
torsion and the resulting K-theory groups will tend to be trivial.
♣Should do this for
super vector spaces
If we interpret the existence of an odd automorphism (4.11) as the possibility of a total and take into
account the super
pairing of “particles” and “holes” and hence of a trivial system then we can identify the dimension. ♣

abelian group of reduced topological phases with a twisted K-theory group:

τ,χ
RT P(Gτ , φ, χ) = φ KG (pt) (4.21)

We have not yet made contact with standard statements found in the works of Kitaev
et. al. To do that we must consider the very special case that

Gtw = U tw × G0 (4.22)

– 108 –
where G0 = ker(τ, χ). In this case we can invoke the theorem of 2.6 to identify twisted
G-equivariance with G0 equivariance together with having a Clifford module for the corre-
sponding Clifford algebra. ♣Explain this point
more fully! ♣
In this way we get the correspondence between the 10-possibilities for the CT group
with the 10 possible K-theories

Subgroup U ⊂ M2,2 T2 C2 φ K τ,χ

G
{1} KG0 (pt)
0
{1, S̄} KG1 (pt)
0
{1, T̄ } +1 KOG0 (pt)
0
−1
M2,2 +1 −1 KOG 0
(pt)
−2
{1, C̄} −1 KOG0 (pt)
−3
M2,2 −1 −1 KOG 0
(pt)
−4
{1, T̄ } −1 KOG 0
(pt) = KSp4G0 (pt)
−5
M2,2 −1 +1 KOG 0
(pt)
−6
{1, C̄} +1 KOG0 (pt)
−7
M2,2 +1 +1 KOG 0
(pt)

The above is the kind of table (usually without the G0 ) that one finds in the papers. ♣Aren’t we missing
a twisting ν0 of the
G0 -equivariant K
4.2 Motivation: Bloch theory theory here? ♣

Next we would like to generalize to certain noncompact groups - the crystallographic groups
- and we would also like to include some notion of locality so we can speak of fermions in
different dimensions. To do that we need to develop a bit more mathematical machinery.
We consider noninteracting particles in an affine space E modeled on Euclidean space
d
R subjected to a potential with the symmetry of a crystal C ⊂ E. ♣Change notation
E →E ? ♣
In the absence of electromagnetic fields the symmetry group of the physical system
will be a magnetic crystallographic group G(C) ⊂ Euc(d) × Z2 . The Z2 factor is for time-
reversal, and hence G(C) is a Z2 -graded group, by restricting τ which is the projection
onto the second factor of Euc(d)× Z2 . As we have stressed many times, this does not imply
that G(C) is a product with Z2 .
The Hilbert space of a single electron is

H = L2 (E; W ), (4.23)

where W is a finite-dimensional Hilbert space accounting for internal degrees of freedom.

Different choices of W will change the relevant group G(C). For example, if we want to
take into account the spin of the electron the modification is described below. We will take
W = C for simplicity, and indicate the spin-generalization in remarks.

– 109 –
 Now, according to the general principles, a φ-twisted extension G(C)tw acts on the
Hilbert space. In the absence of electromagnetic fields the subgroup of lattice translations
is not extended so we have

1 → L(C) → G(C)tw → P (C)tw → 1 (4.24)

where P (C)tw is a φ-twisted extension of the magnetic point group P (C) ⊂ O(d) × Z2 , and
we assume the grading φ of G(C) is pulled back from that of P (C).
A nice geometrical way to think about Bloch waves is to introduce a pair of dual tori
T and T ∨ . ♣The notation T is
heavily overused!
The torus T is the quotient of an affine space E modeled on Rd by the sublattice Find another font.
♣
L = L(C) or Rd :
T := E/L (4.25)

On the other hand, we also have the Brillouin torus

T ∨ := E ∨ /L∨ (4.26)

where E ∨ is the dual space to Euclidean space. In physics it is the momentum space or
reciprocal space and the dual torus (normalized so that L∨ := Hom(L, Z)) is the set of
reciprocal lattice vectors. We should think of T ∨ as the set of characters of L. The general
characters are of the form:
v 7→ e2πik·v v ∈ L(C) (4.27)

where k ∈ E ∨ is a lift of k̄. This only depends on the projection k̄ ∈ T ∨ and we can denote
this character λk̄ , or just λ.
Next, we introduce the Poincaré line bundle, a complex line bundle

L → T∨ × T (4.28)

One way (albeit not the most conceptual way) to define it is to define the restriction
Lk̄ → T for each fixed k̄. We regard T as the base of a principal L-bundle L → E → T ,
and we take the associated complex line bundle determined by the representation k̄ of L.
Then Lk̄ varies smoothly with k̄ to define a line bundle over T ∨ × T . The space of sections
of the line bundle Lk̄ → T is naturally identified with equivariant functions on the total
space E,
T∨ × E → C (4.29)

with equivariance determined by k̄. This is a mathematical way of stating the familiar
Bloch periodicity condition:

ψ(k̄, x + v) = χk̄ (v)ψ(k̄, x) = e2πik·v ψ(k̄, x) ∀v ∈ L, x ∈ E (4.30)

We can define a Hilbert space of sections

Ek̄ = L2 (T ; Lk̄ ⊗ W ) (4.31)

– 110 –
As k̄ varies over T ∨ the Hilbert spaces fit into a Hilbert bundle E → T ∨ , and we can
identify
H := L2 (E; W ) ∼
= L2 (T ∨ ; E) (4.32)
♣Maybe comment
on Fourier
We would like to understand the isomorphism (4.32) as an isomorphism of G(C)tw transform
D.19-D.22 ♣
representations. Taking W = C, there is a canonical right action of G(C) on L2 (E; W ):
(
ψ(Rx + v) φ(R) = +1
({R|v} · ψ) (x) := (4.33)
ψ ∗ (Rx + v) φ(R) = −1

(If we had used the inverse transformation acting on x we would get a left-action.) We
would like to understand this in terms of an action on L2 (T ∨ ; E). ♣This is treating ψ
as a scalar. Discuss
The action of L is clear: It is just a multiplication operator on the fibers. Moreover, if how it generalizes
for a spinor and for
ψk̄ (x) is quasiperiodic as in (4.30), then it is easy to check that {R|v} · ψk̄ is quasiperiodic the magnetic point
group. That is
for k̄′ with ( necessary to get
T 2 = −1. Also,
R−1 · k φ(R) = +1
{R|v} · k̄ = k̄′ = (4.34) with this definition
G(C) is realized on
−1
−R · k φ(R) = −1 the Hilbert space
without being
twisted. ♣
where k is a lift of k̄. The tricky part is in lifting this P (C) action to the Hilbert space
fibers to make an equivariant bundle. It turns out that E is a twisted equivariant P (C)
bundle over T ∨ . To explain this tricky point we will introduce some general mathematical
ideas which allow us to interpret E as twisted vector bundles over a groupoid T ∨ //P (C).

Remark: The bundle E → T ∨ carries a canonical family of flat connections labeled by

x̄0 ∈ T . In order to see this it suffices to describe the parallel transport along straight-line
paths in T ∨ . Suppose k̄ ∈ T ∨ and we choose a lift k ∈ E ∨ and a path k(t) = k + t∆k,
where 0 ≤ t ≤ 1 and ∆k ∈ L∨ , so that k(t) is a closed loop in T ∨ . Then, for any ψk̄ ∈ Ek̄
we have to say how it is parallel transported along k(t). We interpret the fibers of E as
spaces of quasiperiodic functions and define a family of quasiperiodic functions

(U (t) · ψk̄ )(x) := e2πit∆k·(x−x0 ) ψk̄ (x) (4.35)

This has quasiperiodicity determined by k(t). Computation of the parallel transport around
contractible loops shows that the connection is a flat connection: the curvature is zero.
However, the holonomy around the closed loop γ represented by the path k(t), 0 ≤ t ≤ 1
is clearly multiplication by the periodic function:

h(γ) = e2πi∆k·(x−x0 ) (4.36)

Note that it only depends on the projection x̄0 of x0 to T .

Remark: Let us indicate how to include electron spin. The spin-orbit terms in the
Schrodinger equation arise from the Foldy-Wouthuysen expansion of the Dirac equation.
In QED the electron wavefunction transforms as a τ -representation of the Z2 -graded group
(Pin(−1, +3), τ ) which restricts to the 4-complex-dimensional pin representation on the

– 111 –
components that do not reverse time orientation. In solid state physics we ignore the
boosts and hence should define a double cover Õm of O(3) × Z2 by the pullback construc-
tion:
// Z2 π̃ //
1 // Õm O(3) × Z2 // 1 (4.37)

 
// Z2 ι // π
1 Pin(−1, 3) // O(1, 3) // 1

When pulled back to a representation of Õm the representation decomposes as a sum of

(identical) two-dimensional τ -twisted representations of Õm . This two-dimensional repre-
sentation is defined by taking the lift of a reflection Rn̂ in the plane orthogonal to a unit
vector n̂ to be
(R̃n̂ · ψ)(x, t) = (n̂ · σ)ψ(Rn̂ x, t) (4.38)
and the lift T of a time-reversal T̄ ∈ O(3) × Z2 to be

(T̃ · ψ)(x, t) = iσ 2 ψ(x, −t) (4.39)

These operators generate Õm . Now, given a crystallographic group G(C), the relevant
group that is represented on H is the pullback \
G(C) constructed as follows. We first pull
back
1 // Z2 \
// P (C)
π̃ // P (C) // 1 (4.40)

 
// Z2 ι π // O(3) × Z2
1 // Õm // 1

to construct the proper spin cover of the magnetic translation group acting on electron
wavefunctions and then we pull back again

1 // L(C) \
// G(C) π′ // \
P (C) // 1 (4.41)
π̃
 
ι π
1 // L(C) // G(C) // P (C) // 1

Note that a corollary of this discussion is that, because it is experimentally impor-

tant that T 2 = −1 on the electron wavefunction, the relevant pin double cover of spatial
reflections for electron wavefunctions in solid state physics is Pin+ (3) and not Pin− (3).

4.2.1 Bands and insulators

The Hamiltonian of a noninteracting electron in a crystal is
1 2
H=− ∇ + V + spin − orbit (4.42)
2m
acting on L2 (E; W ). Under this isomorphism it acts on Ek̄ with discrete spectrum, since it
is an elliptic operator acting on wavefunctions on a compact space. The discrete spectrum
varies continuously with k̄ ∈ T ∨ , producing a set of bands.

– 112 –
Figure 13: An example of band structure. The horizontal axis is a line of characters starting
from the origin (labelled Γ) to the edge of the Brillouin zone. The material shown here is silicon.
CHECK!!!

In materials the groundstate configuration of electrons fills up the lowest energy levels
consistent with the Pauli principle. In insulators the ground state configuration of electrons
fills all the bands up to some level but there is a gap between the filled (valence) and unfilled
(conduction) bands, as shown in Figure 13. We choose the Fermi energy Ef to be within
the gap.
It follows that in insulators there is a natural splitting of the Hilbert bundle

E = E− ⊕ E+ (4.43)

given by the spectrum below and above Ef . In families of materials where the bands do
not cross Ef we have a gapped system.
It is clear that E − will be finite rank and E + is infinite rank. In general E − will be a
topologically nontrivial vector bundle over the Brillouin torus in two and three dimensions.
Sometimes it is useful to focus on just a few bands near the Fermi level. In this case
E± will be finite rank. We thus have two natural cases to consider: E + of infinite rank and
E+ of finite rank.
Remark: Explain the family of flat connections on E and the “Berry connection”

– 113 –
4.3 Groupoids
4.3.1 Categories

Definition A category C consists of

a.) A set X0 of objects
b.) A set X1 of morphisms with two maps p0 : X1 → X0 and p1 : X1 → X0 . The map
p0 (f ) = x1 is the range map and p1 (f ) = x0 is the domain map. Morphisms are often
denoted as arrows:
f
x0 → x1 (4.44)

The set of all morphisms from x0 to x1 is denoted hom(x0 , x1 ).

c.) There is a composition law on the set of composable morphisms

X2 = {(f1 , f2 ) ∈ X1 × X1 |p0 (f1 ) = p1 (f2 )} (4.45)

The composition law takes X2 → X1 and may be pictured as the composition of arrows. If
f1 : x0 → x1 and f2 : x1 → x2 then the composed arrow will be denoted f2 ◦ f1 : x0 → x2 .
The composition law satisfies the axioms

1. For all x ∈ X0 there is an identity morphism in X1 , denoted 1x , or Idx , such that

1x1 ◦ f = f and f ◦ 1x0 = f .

2. The composition law is associative: If we consider 3 composable morphisms

X3 = {(f1 , f2 , f3 )|p0 (fi ) = p1 (fi+1 )} (4.46)

then (f3 ◦ f2 ) ◦ f1 = f3 ◦ (f2 ◦ f1 ).

Remarks:

1. When defining composition of arrows one needs to make an important notational

decision. If f : x0 → x1 and g : x1 → x2 then the composed arrow is an arrow
x0 → x2 . We will write g ◦ f when we want to think of f, g as functions and f g when
we think of them as arrows.

2. It is possible to endow the data X0 , X1 and p0 , p1 with additional structures, such as

topologies, and demand that p0 , p1 have continuity or other properties.

3. A morphism f ∈ hom(x0 , x1 ) is said to be invertible if there is a morphism g ∈

hom(x1 , x0 ) such that g ◦ f = 1x0 and f ◦ g = 1x1 . If x0 and x1 are objects with
an invertible morphism between then then they are called isomorphic objects. One
key reason to use the language of categories is that objects can have nontrivial au-
tomorphisms. That is, hom(x0 , x0 ) can have more than just 1x0 in it. When this is
true then it is tricky to speak of “equality” of objects, and the language of categories
becomes very helpful.

– 114 –
 4. There is a geometrical picture associated with a category where objects are points,
morphisms are one-simplices etc. See Figure 14. Elements of X2 are associated with
2-simplices, elements of X3 are associated with 3-simplices. See Figure 15. We can
clearly continue this to Xn for n-tuples of composable morphisms and create what is
known as a simplicial space.

Figure 14: Elementary 0, 1, 2 simplices in the simplicial space |C| of a category

Figure 15: An elementary 3-simplex in the simplicial space |C| of a category

4.3.2 Groups and groupoids

Let us consider a category with only one object, but we allow the possibility that there are
several morphisms. For such a category let us look carefully at the structure on morphisms
f ∈ X1 . We know that there is a binary operation, with an identity 1 which is associative.
But this is just the definition of a monoid!
If we have in addition inverses then we get a group. Hence:
Definition A group is a category with one object, all of whose morphisms are invertible.

We should think of it geometrically in terms of Figure 14 and Figure 15. (In these
figures identify all the objects to one point.)

– 115 –
Definition A groupoid G is a category all of whose morphisms are invertible.

Note that for any object x in a groupoid, hom(x, x) is a group. It is called the auto-
morphism group of the object x. It is useful to define a homomorphism µ of groupoids to
be a functor so if f1 , f2 are composable then µ(f2 ◦ f1 ) = µ(f2 ) ◦ µ(f1 ). In particular, we
will use below the notion of a homomorphism of a groupoid G to the group Z2 (thought
of multiplicatively). Since a group has a single object this amounts to a map µ : G1 → Z2
so that µ(f2 ◦ f1 ) = µ(f2 )µ(f1 ).

Example 1: The key example we will need is provided by the action of a group G on a
space X. We can then form the a groupoid denoted G = X// G. This groupoid has as
objects and morphisms given by

G0 := X G1 := X × G (4.47)

and p1 (x, g) = x while p0 (x, g) = g · x for a left-action or x · g for a right-action. The

picture we associate with a morphism (x, g) is an arrow from x to gx (for a left action) or
xg (for a right action). The groupoid X// G contains important information not available
in the ordinary quotient X/G in that it contains the automorphism group of the action on
x. We should think of the points x as having the isotropy group attached to them. When
identifying a group with a category with one object we are really speaking of the groupoid
G = pt// G, where the group G acts on a single point. The main example of groupoids
used later in this chapter is the groupoid T ∨ // P (C).

Example 2: Another useful example of a groupoid is provided by a manifold M with an

open cover {Uα } labeled by an ordered set. Then

G0 = ∐α Uα (4.48)

G1 = ∐α<β Uα ∩ Uβ (4.49)

and then we take p0 , p1 to be the inclusion maps into β, α, respectively.

4.3.3 Vector bundles on groupoids

By a complex vector bundle on a groupoid G we mean:

1. V1: A complex vector bundle E → G0 . Fibers will be denoted Ex for x ∈ G0 .

f
2. V2: An isomorphism ψ : p∗1 E → p∗0 E. Concretely, if (x0 −→x1 ) ∈ G1 then we have a
C-linear isomorphism
ψf : Ex0 → Ex1 . (4.50)

3. V3: Cocycle condition on G2 , if (f, g) ∈ G2 then we must have

ψg ψf = ψgf : Ex0 → Ex2 (4.51)

– 116 –
Example 1: In the case G = X// G, a bundle on the groupoid is the same thing as a
G-equivariant bundle on X. The maps ψf provide a lifting of the group action on X to
the total space of the bundle E.

Example 2: In the case of the groupoid G provided by an open cover {Uα }, the ψf are the
clutching transformations, and the cocycle condition is the standard one of bundle theory.

4.4 Twisted equivariant K-theory of groupoids

4.4.1 Central extensions as line bundles over a group

Let us now return to the twisted equivariant K-theory of a point, but now incorporating
the viewpoint of a group as a groupoid G = pt// G.
We will identify a central extension of G with a complex line bundle over G1 , the space
of morphisms: For each group element g we have a line Lg and there is a multiplication

λg2 ,g1 : Lg2 ⊗ Lg1 → Lg2 g1 (4.52)

The familiar cocycle is obtained by choosing basis vectors ℓg for each line Lg and then
comparing basis vectors:
λg2 ,g1 (ℓg2 ⊗ ℓg1 ) = λ(g2 , g1 )ℓg2 g1 (4.53)

The line bundle is defined over G1 , the multiplication is defined over G2 , and there is a
consistency condition defined for each element of G3 . This is the requirement that the
different ways of using λ to multiply

Lg 3 ⊗ Lg 2 ⊗ Lg 1 → Lg 3 g 2 g 1 (4.54)

yield the same answer. Choosing a basis we recover the 2-cocycle condition (4.8). Making
a change of basis ℓg induces a change of the 2-cocycle by a coboundary.
Of course, a given group will in general admit different isomorphism classes of such
line bundles with multipliation. The isomorphism class is labeled by an an element of
H 2 (BG; U (1)) ∼
= HG2 (pt; U (1)).

If we denote by τ the isomorphism class of the central extension then in this language
a representation of Gτ can be thought of as a vector bundle over the groupoid G together
with isomorphisms defined on G1

ρg : Lg ⊗ V → V (4.55)

satisfying the consistency conditions defined on G2 . These are obtained by demanding that
the two ways of multiplying
Lg 2 ⊗ Lg 1 ⊗ V → V (4.56)

yield the same answer. ♣Write out the

commutative
diagram ♣

– 117 –
4.4.2 φ-twisted extensions and representations of a graded group
Now for a Z2 -graded group G with φ : G → Z2 we can similarly interpret φ-twisted
extensions in terms of complex line bundles.
Now, in the categorical picture of a group, a Z2 graded group has a sign φ(g) = ±1
attached to each morphism (one-simplex).
Let us introduce the useful notation
(
φ(g) V φ(g) = +1
V := (4.57)
V φ(g) = −1

where V is a complex vector space.

Then, a φ-twisted extension of G is equivalent to a multiplication law

λg2 ,g1 : φ(g1 ) Lg2 ⊗ Lg1 → Lg2 g1 (4.58)

Again we require associativity of the map ♣Check it is the

same condition as
that quoted above.
φ(g2 g1 )
Lg3 ⊗ φ(g1 ) Lg2 ⊗ Lg1 → Lg3 g2 g1 (4.59) ♣

If we choose basis vectors ℓg for each complex line Lg we obtain the φ-twisted co-cycle
condition given in (4.10) above.
If, now G is a bigraded group with bigrading (φ, χ) then we can incorporate the idea
of a (τ, φ, χ)-twisted representation of G:
First, each morphism in G1 is bigraded by (φ, χ).
Next, we choose a complex line bundle over G1 , the space of morphisms, satisfying
(4.58). This determines a φ-twisted extension τ . We now consider these lines to be Z2 -
graded with parity determined by χ(g).
Then, a (τ, φ, χ)-twisted representation of G is a vector bundle over the groupoid
G = pt// G with a compatible set of even and C-linear isomorphisms

ρg : φ(g) (Lg ⊗ V ) → V (4.60)

satisfying the consistency conditions defined on G2 . These are obtained by demanding that
the two ways of multiplying
 
φ(g2 g1 ) φ(g1 )
Lg 2 ⊗ Lg 1 ⊗ V → V (4.61)

yield the same answer. ♣Write out the

commutative
diagram ♣
Remark: Once again, φ-twisted extensions of G are classified by cohomology, but now the
cohomology is itself twisted cohomology. ***** EXPLAIN MORE ****

4.4.3 Twisting the K-theory of a groupoid

It is now straightforward to generalize to groupoids G more general than pt// G.

Definition: Now suppose we have a bigraded groupoid G, that is, a groupoid with two
homomorphisms φ, χ : G → Z2 . Then we define a (φ, χ)-twisting of the K-theory of G to
be

– 118 –
 1. A collection of Z2 -graded complex lines Lf , for all f ∈ G1 , graded by χ(f ).

2. A collection of C-linear even isomorphisms

λf2 ,f1 : φ(f1 ) Lf2 ⊗ Lf1 → Lf2 ◦f1 (4.62)

3. Such that the λ satisfies the cocycle condition which requires that for all 3-composable
morphisms (f1 , f2 , f3 ) ∈ G3 the two ways of multiplying

φ(f2 ◦f1 )
Lf3 ⊗ φ(f1 ) Lf2 ⊗ Lf1 → Lf3 ◦f2 ◦f1 (4.63)

give the same map.

Remarks:

1. We define a twisting of the K-theory of G before we define the K-theory of G!

2. These twistings have a nice generalization to a class of geometrical twistings deter-

mined by bundles of central simple superalgebras (over G0 ) and invertible bimodules
(over G1 ). The isomorphism classes of such twistings, for G = X//G are classified -
as a set - by
0 1 3
HG (X; Z2 ) × HG (X; Z2 ) × HG (X; Zφ ) (4.64)

3. **** EXPLAIN THAT THE TWISTINGS WE ENCOUNTER ARE NONEQUIV-

ARIANTLY TRIVIAL ****

4. There are yet more general twistings, but they have not yet had an application in
physics. Not even in string theory.

4.4.4 Definition of a twisted bundle and twisted K-theory of a groupoid

Definition Given a (φ, τ, χ)-twisting of the K-theory of a groupoid G, a (φ, τ, χ)-twisted
bundle over G is

1. A complex Z2 -graded bundle over G0 .

2. A collection of C-linear, even isomorphisms over morphsms f : x0 → x1 in G1 :

ρf : φ(f ) (Lf ⊗ Vx0 ) → Vx1 (4.65)

3. Such that the ρx,f satisfy the compatibility condition over G2 obtained by requiring
that the two ways of multiplying
 
φ(f2 ◦f1 ) φ(f1 )
L f 2 ⊗ L f 1 ⊗ Vx 0 → Vx 2 (4.66)

give the same answer.

– 119 –
 Finally, we can give a definition of the twisted equivariant K-theory over a groupoid
G:
Let ν = (φ, τ, χ) denote a twisting of the K-theory of G.
The isomorphisms of ν-twisted bundles form a monoid Vectνs (G) under direct sum.
(The subscript s reminds us that the bundles are Z2 -graded.)
We define a sub-monoid of “trivial” elements, denoted Trivνs (G) to be those with an
odd invertible automorphism T : V → V . Then we set

K ν (G) := Vectνs (G)/Trivνs (G) (4.67)

In the case that the groupoid is a global quotient G = X// G we also use the notation
τ,χ
K ν (G) := φ KG (X) (4.68)

Remarks:

1. If we take τ and χ to be trivial and G = Z2 × G0 with G0 = kerφ then (4.68) reduces

to G0 -equivariant KR theory.

2. For X = pt we recover the previous description.

4.5 Applications to topological band structure

Now we will apply some of the above general ideas to the case of topological band structure.

4.5.1 The canonical twisting

We now consider the groupoid T ∨ // P (C). We claim there is a canonical (φ, χ)-twisting of
the K-theory of this groupoid.
The canonical twisting is actually a special case of a more general result discussed in
[26], Section 9, and references therein. We will explain the the statement of the general
result, omit the proof, but explain the upshot for the special case of interest for band
structure.
The general result concerns extensions

1→N →G→Q→1 (4.69)

Let X be the space of irreps of N . Then Q acts on X as follows:

Suppose ρV : N → Aut(V ) is an irrep of N . Then we can twist it by an element q ∈ Q
by choosing a lift s(q) ∈ G and defining

ρqV (n) := ρV (s(q)ns(q)−1 ) (4.70)

Note that since N is normal in G this makes sense: s(q)ns(q)−1 ∈ N , and ρqV is a new
representation of N on the vector space V . It clearly depends on the choice of section s,
but the isomorphism class of of ρqV does not depend on the choice of X. In general, the

– 120 –
isomorphism class is distinct from ρV . Denoting the isomorphism class by [V ] for brevity,
there is a right-action of Q on X,
[V ] → [V ] · q. (4.71)

Example 1: Consider the permutation group S3 as an extension of Z2 by Z3 with the Z3

normal subgroup identified with the subgroup A3 of even permutations:

1 → A3 → S3 → Z2 → 1 (4.72)

There are three characters of A3 labeled by χ(1), a third root of unity. So, choosing a
nontrivial third root of unity ω we can identify X = {1, ω, ω 2 }. Now choose the lift of the
nontrivial element q of Z2 to be s(q) = (12). Then (12)(123)(12) = (132) and hence the
action of Q on X fixes 1 and exchanges ω and ω 2 .

Example 2: In our crystallographic example, N = L(C), so X = T ∨ is the Brillouin

torus. An isomorphism class is determined by a character χk̄ and R ∈ P (C) ⊂ O(d) acts
by
R : χk̄ → χk̄′ (4.73)

where, if k and k′ are lifts of k̄ and k̄′ then

k′ = R−1 k (4.74)

Next, the statement of the theorem is that there is a canonical central extension ν of
G = X//Q such that a representation E of G is in one-one correspondence with ν-twisted
bundles 30 over G. In fact, if G is compact this sets up an isomorphism

KG (pt) ∼ ν
= KQ (X) ∼
= K ν (G). (4.75)

The result can be extended to (φ, χ)-twisted bundles, but we will restrict to the simplest
case here.
In the application to Bloch theory the extension of the groupoid is simply the following:
Given R ∈ P (C) and k̄ ∈ T ∨ we need to produce a line Lk̄,R . We identify it with the vector
space of equivariant functions from the fiber in G(C) above R. That is π −1 (R) ⊂ G(C) is
the set of space group elements g projecting to R. A typical element will be {R|v0 } where
v0 might or might not be in the lattice L. (If the group is non-symmorphic then for some
R it will be impossible to take v0 in L.) The set π −1 (R) is a torsor for L. We consider the
set of complex-valued functions
f : π −1 (R) → C (4.76)

which satisfy the equivariance condition:

f (Tv g) = χk̄ (v)f (g) (4.77)

30
actually, sheaves

– 121 –
for all v ∈ L(C) and all g ∈ π −1 (R). Note that the set of such functions (4.76) satisfying
the equivariance constraint (4.77) form a vector space. In fact, it is a one-dimensional
vector space: If we choose any lift s(R) ∈ π −1 (R) then the value of the function f at s(R)
determines its values on all other points in the fiber.
The one-dimensional vector space of equivariant functions on the fiber π −1 (R) is iden-
tified with the line bundle of the twisting Lk̄,R .
It remains to show that we can define a product

λk̄,R1 ,R2 : LR−1 k̄,R2 ⊗ Lk̄,R1 → Lk̄,R1 R2 (4.78)

1

which satisfies the cocycle condition. To do this we consider the general element of Lk̄,R1 . ♣careful about left
vs. right actions ♣
It is a function satisfying

f1 ({R1 |v0 + v}) = χk̄ (v)f1 ({R1 |v0 }) ∀v ∈ L (4.79)

where v0 represents a choice of section so that {R1 |v0 } ∈ G(C). Similarly,

f2 ({R2 |v0′ + v}) = χR−1 k̄ (v)f2 ({R2 |v0′ }) ∀v ∈ L (4.80)

1

Then, note that {R1 |v0 } · {R2 |v0′ } = {R1 R2 |v0 + R1 v0′ } so that we can define the product:
λk̄,R1 ,R2 (f1 ⊗ f2 ) to be the function f12 defined by:

f12 ({R1 R2 |v0 + R1 v0′ + v} := χk̄ (v)f1 ({R1 |v0 })f2 ({R2 |v0′ }) (4.81)

where v ∈ L. Note that f12 is, by construction, an equivariant function on π −1 (R1 R2 ) and
moreover it is bilinear in f1 and f2 . Moreover, one can easily check that the definition of
the product is in fact independent of the choice of section determined by v0 , v0′ . We leave
the verification of the cocycle relation to the reader. This completes the construction of
the canonical twisting ν can for the K-theory of the groupoid T ∨ //P (C) relevant for band
structure.
It is interesting to consider the restriction of the canonical twisting ν can to the sub-
groupoid of the fixed subgroup of a particular point k̄ ∈ T ∨ . That is, let P (C, k̄) ⊂ P (C)
denote the subgroup of P (C) which fixes k̄ ∈ T ∨ . Then there is a natural inclusion homo-
morphism
k̄//P (C, k̄) ֒→ T ∨ //P (C) (4.82)

and the twisting ν can will pull back to a line bundle with multiplication over the group
P (C, k̄). This is just a central extension of P (C, k̄) and we will now describe it. Since
P (C, k̄) is a finite group we can enumerate its elements as Ri , i = 1, . . . , n for some integer
n. Then we choose a section of π : G(C) → P (C) over P (C, k̄), that is, for each i we
choose an element {Ri |v i } ∈ G(C). Now we choose a basis for Lk̄,Ri by declaring the basis
vector to be the equivariant function determined by f ({Ri |v i }) = 1. Now, for any pair of
elements Ri , Ri′ ∈ P (C, k̄) there is an i′′ so that Ri Ri′ = Ri′′ . Then we can define vectors
′
v ii ∈ L by
′ ′ ′ ′′
{Ri |v i }{Ri′ |v i } = {Ri Ri′ |v i + Ri v i } = {1|v ii }{Ri′′ |v i } (4.83)

– 122 –
That is:
′ ′ ′′
v ii := v i + Ri v i − v i (4.84)
It follows from the definition of the cocycle that, with respect to this basis it is
′
λ(Ri , Ri′ ) = χk̄ (v ii ) (4.85)

This determines the central extension fairly explicitly.

Remarks

1. If G(C) is symmorphic (a split extension of P (C) by L(C) ) then we can choose

v i = 0 for all elements Ri and hence the central extension will be trivial.

2. The above central extension has a much more invariant and conceptual description,
given in Proposition 9.31 of [26]. ♣Give it here? ♣

Example: Let us return to the two-dimensional crystallographic group G(C) described by

1 2
the generators (1.50) and (1.51). The character on L = Z2 given by λ(n1 , n2 ) = (−1)n +n
is clearly fixed by the full point group Z2 × Z2 generated by
! !
−1 0 1 0
R1 = R2 = (4.86)
0 1 0 −1

since it is unchanged if we change the sign of n1 or n2 . A lift of λ = χk̄ is k = ( 12 , 12 ).

The generators are g1 = {R1 |(δ, 12 )} and g2 = {R2 |0}. Then we lift R1 R2 to the element
g1 g2 = {R1 R2 |(δ, 21 )} but then g2 g1 = {R1 R2 |(δ, − 12 )}. It follows that the cocycle satisfies

λ(R1 , R2 ) χ (v 12 )
= k̄ 21 = −1 (4.87)
λ(R2 , R1 ) χk̄ (v )

In general, if the group P (C, k̄) is abelian then the ratio

λ(Ri , Ri′ )
′
χ (v ii )  ′

= k̄ i′ i = exp[2πik · (1 − Ri )v i − (1 − Ri′ )v i ] (4.88)
λ(Ri′ , Ri ) χk̄ (v )
is gauge invariant and if it is nontrivial the central extension is nontrivial. In our example
the central extension is D4 . ♣check! ♣

4.5.2 E as a ν can -twisted bundle over T ∨ //P (C)

Returning to the general theorem (4.75) a representation E of G defines a vector bundle
over X whose fiber at [V ] is
E[V ] = HomN (V, E) (4.89)
That is, the degeneracy space in the isotypical decomposition of E as a representation of
N . On the RHS we have used a specific representative V of an isomorphism class. In the
application to band structure, this will just be the space Ek̄ corresponding to Bloch waves
of momentum k̄.

– 123 –
 The required action
Lk̄,R ⊗ Ek̄ → ER−1 k̄ (4.90)

is defined by choosing a lift {R|v} of R in G(C) and setting

ρk̄,R : f ⊗ ψ → ψ ′ (4.91)

where
ψ ′ (x) := f ({R|v})ψ(Rx + v) (4.92)

♣CHECK! ♣

4.5.3 Modifications for including (φ, χ)

Now let us suppose that G(C) is bigraded by (φ, χ) and show how the canonical twisting
is modified.
Returning to the general situation, (4.69) we assume that the grading on G factors
through a grading on Q. That is, there is a homomorphism φ : Q → Z2 defining a Z2
grading on Q and that on G is just the pullback. Then the action of Q on the space of
irreps X is modified to
(
ρV (s(q)ns(q)−1 ) φ(q) = +1
ρqV (n) := (4.93)
ρ̄V̄ (s(q)ns(q)−1 ) φ(q) = −1

In the case of band structure, we are assuming that the grading on G(C) is pulled back
from a grading on the (magnetic) point group and then the modified action on T ∨ is
(
−1 R−1 k φ(R) = +1
R · k̄ = (4.94)
−R−1 k φ(R) = −1

where k is a lift of k̄ to L∨ .
Now we must generalize to the multiplication of the form (4.58), that is

λk̄,R1 ,R2 : φ(R1 ) LR−1 k̄,R2 ⊗ Lk̄,R1 → Lk̄,R1 R2 (4.95)

1

This will be given by an equation of the form (4.81) but with a complex conjugation on
f2 . ♣check ♣

Now, including both φ and χ the generalization of equations (4.90)-(4.92) for W = C

isf
(
′ f ({R|v})ψ(Rx + v) φ(R) = +1
ψ (x) := (4.96)
f ({R|v}) ψ(Rx + v) φ(R) = −1

GIVE GENERALIZATION WITH SPIN.

– 124 –
4.5.4 K-theory classification of phases in topological band theory
Having set up this formalism it is fairly straightforward to relate the classification of phases,
and the reduced phases for the set of Hamiltonians of noninteracting electrons with G(C)-
symmetry, where G(C) is a bigraded group by (φ, χ) as follows:

1. If E + is finite-rank, then the components T P are in 1-1 correspondence with the

monoid of Z2 -graded twisted equivariant bundles
can
TP ∼
= Vectνs (T ∨ // P (C)) (4.97)

2. If E + is finite-rank and we define trivial phases by those such that E admits an

invertible odd intertwiner then the abelian group RT P is a K-theory group
can
RT P ∼
= K ν (T ∨ // P (C)) (4.98)

3. If E + is infinite-rank, then there is a projection of the components T P onto the

monoid of Z2 -graded twisted equivariant bundles
can
T P → Vectνs (T ∨ // P (C)) (4.99)

♣Make the arrow

an onto symbol. ♣

4. If E + is infinite-rank, and we define trivial phases by those such that E admits an

invertible odd intertwiner then the abelian group RT P is a K-theory group
can
RT P ∼
= K ν (T ∨ // P (C)) (4.100)

EXPLAIN A LITTLE BIT

In the very special case that the group G(C)τ splits as U tw × G0 where G0 = ker(φ) ∩
ker(χ) and U tw is a CT-group then, using the main theorem of Section §2.6 and the
relation between degree-shift and twisting by a Clifford algebra we can relate the above
twisted equivariant K-theories to more standard K-theories.
For example, in the case that E + is infinite-rank we necessarily have χ = 1. Then we
just have three cases, as we have seen repeatedly above. In this case we have
can
Subgroup U ⊂ M2 T2 K ν (T ∨ // P )
can
{1} KPν (T ∨ )
{1, T̄ } +1 KRPν00 (T ∨ )
{1, T̄ } −1 KRPν00−4 (T ∨ )

where P0 is the projection of G0 into O(d).

Finally, we compare with the literature. Essentially three special cases have been
discussed:

– 125 –
 1. Take G(C) = hIi × L, where I is inversion, and φ(I) = χ(I) = 1. Then the relevant
K-theory group is KZ02 (T ∨ ). This was studied in [50].

2. Take G(C) = hT̄ i × L, where T̄ is time-reversal, and φ(T̄ ) = −1 and χ(T̄ ) = +1.
Because we are working with electrons the lift T of T̄ has square T 2 = −1 determining
τ . Then the relevant K-theory group is KR−4 (T ∨ ). This was studied in [33, 44].

¯ × L, where T̄ is time-reversal, and the bigrading is defined as

3. Take G(C) = hT̄ , Ii
above. Because we are working with electrons the lift T of T̄ has square T 2 = −1,
while I 2 = +1. This determines τ . Then the relevant K-theory group is KOZ−42
(T ∨ ).
This was studied in [27, 28].

COMMENT ON KANE-MELE and CHERN-SIMONS INVARIANTS.

4.6 Computation of some equivariant K-theory invariants via localization

These equivariant K-theory groups are not easy to compute. What are some new invariants?
One way to get a handle on some of them is via the use of “localization.” This allows us
to simplify the K-theory groups and still retain some nontrivial information.

4.6.1 Localization of rings and modules

If R is a ring, and S ⊂ R a subset closed under multiplication then we can form the localized
ring S −1 R consisting of elements r/s, r ∈ R, s ∈ S with equivalence relation r1 /s1 = r2 /s2
if ur1 s2 = ur2 s1 for some u ∈ S, and we define the ring structure:

r1 r2 r1 r2 r1 r2 r1 s2 + r2 s1
:= + := (4.101)
s1 s2 s1 s2 s1 s2 s1 s2

Similarly, if M is a module over a ring R then we can localize the module S −1 M . This
is a module over the ring S −1 R. S −1 M consists of elements m s with m ∈ M, s ∈ S with
m1 /s1 = m2 /s2 if um1 s2 = um2 s1 for some u ∈ S.
The module S −1 M is called the localization of M at S.
Remarks and Examples:

1. If there is an element s∗ ∈ S with s∗ M = 0 then S −1 M = 0. Proof: 1 · m =

s−1 −1
∗ · s∗ · m = s∗ · 0 = 0.

2. A common construction is the following: For any r ∈ R we can form the multiplicative
set Sr := {r n }n∈Z . Then we denote Sr−1 R := Rr and if M is an R-module, Mr :=
Sr−1 M . A key remark for what follows is the following. If M is a module over R and
r annihilates M then Mr = {0} is the 0-module, because

1
m =1·m= · (rm) = 0 (4.102)
r
for all m ∈ M .

– 126 –
 3. In particular: For G = Z2 we have

R(Z2 ) = Z[ε]/(ε2 − 1) = Z ⊕ Zε.

where ε is the sign representation ε(σ) = −1, for σ the nontrivial element of Z2 . Let
us choose r = 1 − ε. Then we claim:
1
R(Z2 )1−ε ∼
= Z[ ] (4.103)
2

To see this note that

1 1
n + mε = (1 − ε)(n + mε) = (n − m)(1 − ε) = n − m (4.104)
1−ε 1−ε
Now note that 2 = (1 + ε) + (1 − ε) so, multiplying by (1 − ε) we derive
1
2 =1 (4.105)
1−ε
(Alternatively, divide the equation (1 − ε)2 = 2(1 − ε) by (1 − ε)2 . ) In any case, we
have inverted 2 ∈ R(Z2 ), with (1 − ε)−1 = 12 . So now we have the general element of
the ring S −1 R is
n + mε ∼ n − m
= (4.106)
(1 − ε)k 2k
where n, m, k are arbitrary integers. In this sense, inverting 1 − ε in RO(Z2 ) is
essentially the same thing as setting ε = −1:

4. One important construction is the following: Suppose P ⊂ R is a prime ideal. Then

S := R − P is a multiplicative set: for if s1 s2 ∈ P then, by definition s1 or s2 is
in P . In this case the localization S −1 R is called the localization at P and denoted
RP . Similarly, modules localized at P are denoted MP . As an important example
consider the representation ring R(G), i.e. the character ring, of a compact group G.
Let γ be a conjugacy class in G. The set of representations whose character vanishes
on γ is a prime ideal. We can localize at this ideal. That is, we can take S to be
the multiplicative set of virtual representations ρ whose character χρ (γ) 6= 0. This is
used in Atiyah and Segal [10].

5. Good references are Eisenbud [21] ch.2 and Atiyah-MacDonald [9], ch. 3

4.6.2 The localization theorem

Segal’s localization theorem for K-theory, nicely reviewed in Atiyah-Segal *****
Suppose a compact group G acts on a space X. Then KG (X) is a KG (pt) = R(G)-
module. If we choose a multiplicative set S ⊂ R(G) we can localize R(G) and the module
KG (X) at S.
Segal, and Atiyah-Segal consider a conjugacy class γ ⊂ G and consider localizing at
the prime ideal of representations whose character vanishes on γ. Then let

X γ := ∪g∈γ X g (4.107)

– 127 –
where for g ∈ G, X g is the fixed point set of the action by g. There is a pullback map

ι∗ : KG (X) → KG (X γ ) (4.108)

and the key theorem states that this becomes an isomorphism when we invert the characters
which do not vanish at γ. That is, if S = {χ|χ(γ) 6= 0} then

ι∗ : S −1 KG (X) → S −1 KG (X γ ) (4.109)

is an isomorphism.
The main idea of the proof is to break X up into a union of orbits G/H. Whenever H
is a proper subgroup of G there an argument in [10] shows there is an element χ∗ ∈ S such
that χ∗ KH (pt) = 0, but KG (G/H) ∼ = KH (pt) (the fiber of a G-equivariant bundle on G/H
at the coset 1 · H is an H-representation, and that fiber determines the entire bundle).
Therefore, by remark 1 above, the “contribution” of any orbit is zero.
For our application we will not use exactly Segal’s theorem, since we wish to consider
twisted K-theory and since we will invert a slightly different set.
As a simple example suppose that φ = 1 so we do not have time-reversing symmetries
and moreover that the point group has the form Z2 × P ′ (C) where Z2 = {1, I} and I is the
inversion transformation I : x → −x. A nontrivial example of such a point group is that
of the diamond structure, discussed below. Let ǫ be the one-dimensional representation of
P (C) which is trivial on P ′ (C) and is the sign representation of Z2 . We will localize by
inverting the multiplicative set S = {(1 − ǫ)n }n≥0 .
If I takes k̄ to k̄′ with k̄ 6= k̄′ then the restriction of ι∗ E to the set of two points
{k̄, k̄′ } has an isomorphism ι∗ E ∼= ǫι∗ E and hence the localization gives zero for all but the
fixed-points.
Therefore, if φ = 1 and E + is infinite-dimensional we have
can can
S −1 KPν (C) (T ∨ ) ∼
= S −1 KPν (C) ((T ∨ )I ) (4.110)

Example: If P (C) = {1, I} ∼ = Z2 then the fixed point locus is the set of order two
elements of T . If we think of T ∼
∨ ∨
= Rd /Zd then these can be represented by the points
k̄ = (k̄1 , . . . , k̄d ) with k̄i ∈ {0, 12 }. (In the CM literature they are called the TRIM points.)
The fixed points are isolated and at each fixed point k̄i , i = 1, . . . , 2d , we have an element
of the representation ring of Z2 , E − |k̄i ∼ = ni ⊕ mi ǫ. The localization just identifies this with
1
an element of Z[ 2 ] as discussed above. If we know the total number of bands ni + mi = nB
then from ni − mi we can recover the number of parity-odd states at each TRIM point.
These are the invariants discussed in [50].

4.6.3 Example: Some twisted equivariant K-theory invariants for diamond

structure
As a somewhat more nontrivial example let us consider the three-dimensional diamond
structure. We consider the case with φ = 1 so the point group is a 48-element group
isomorphic to Z2 × S4 , where S4 is realized as the proper cubic group and the Z2 factor

– 128 –
Figure 16: The Wigner-Seitz fundamental domain for the Brillouin torus of the diamond structure.

is generated by inversion. The Wigner-Seitz cell is shown in Figure 16. Identifying sides
appropriately produces a copy of T ∨ . ♣Explain some
more of the
We take again the multiplicative set S = {(1 − ǫ)n } as above. There are now 8 fixed background details.
♣
points of I. These consist of the origin, labeled Γ in 16 together with two orbits under the
point group of four points labeled L and three points labeled X, respectively.
At Γ the stabilizer is all of Oh = Z2 × S4 . The contribution to the localized K-theory
from this point is just a representation of this group - it is the representation provided by
the valence bands.
There are four L-points. One of them is k = ( 41 , 41 , 14 ) and has stabilizer ∼ = Z 2 × S3 .
The canonical twist does not lead to a central extension and hence, again the K theory
invariants correspond to a representation of this group.
There is an orbit of three X points ( 12 , 0, 0), (0, 12 , 0), and (0, 0, 12 ) each with isotropy
group Z2 × D4 . For example ( 21 , 0, 0) is fixed by a group of order 16 generated by ǫi and
(23). To identify this with Z2 × D4 we observe that I = ǫ1 ǫ2 ǫ3 is in the center. Then ǫ2 (23)
generates a π/2 rotation in the 23 plane and together with ǫ2 generates D4 acting in the
23 plane. The lifts to G(C) of I and ǫ2 are {I|s} and {ǫ2 |s} and the group commutator is
′
{1|(−1, 1, 1)}, and exp[2πik · v ii ] = exp[2πi( 12 , 0, 0) · (−1, 1, 1)] = −1 is nontrivial. All other
group commutators are trivial. Therefore, the canonical twisting describes a nontrivial ♣CHECK!!! ♣

extension
1 → Z2 → Z^ 2 × D4 → Z2 × D4 → 1 (4.111)

at the three X-points. Now the twisting makes the localization a bit more interesting. We

– 129 –
are localizing in the representation ring of Z^
2 × D4 . Since {1, I} is a quotient of this group
ǫ is in the representation ring and we are inverting powers of (1 − ǫ).
DISCUSS SOME MORE ABOUT THE LOCALIZATION IN THIS CASE
To summarize: By localization we have identified several nontrivial K-theory invari-
ants:
1. A representation of Oh
2. A representations of Z2 × S3 given by the representation of the valence bands at
the L-points.
3. A of Z^2 × D4 given by the representation of the valence bands at the X-points. ♣There be 2-torsion
information lost
It would be very interesting to explore the physical implications in terms of edge modes upon localization...
♣
and electromagnetic response of these “extra” K-theory invariants that go beyond those
normally described in the literature on topological insulators and superconductors.

A. Background material: Cartan’s symmetric spaces

Definition: A symmetric space is a (pseudo) Riemannian manifold (M, g) such that every
point p is an isolated fixed point of an involutive isometry τp .
Near any point p, the involutive isometry τp can be expressed as the inversion of the
geodesics through p. That is, if (x1 , . . . , xn ) are normal coordinates in a neighborhood of p
with ~x = 0 the coordinate of p then τp (~x) = −~x. Importantly, τp extends to an involutive
isometry of the full Riemannian space (M, g)
One can show that the Riemannian curvature is covariantly constant, and hence there
are three families of examples where the scalar curvature (which is must be constant) is
positive, zero, or negative.
Cartan classified the symmetric spaces and found that they are always homogeneous
spaces of Lie groups. The positive curvature examples are of the form G/K where G is a
compact Lie group and K is a Lie subgroup.
Let us first examine G/K at the Lie algebra level. The tangent space of G at 1 is the
Lie algebra g and the tangent space of K at 1 is the Lie subalgebra k. If we write

g=k⊕p (A.1)

then there is a natural identification of p with TK (G/K). The involutive isometry τp where
p = 1 · K has a differential θ = dτp : p → p which in fact can be shown to be the restriction
of an involutive automorphism θ : g → g. That is, θ is a Lie algebra homomorphism
θ([X, Y ]) = [θ(X), θ(Y )] which is an isomorphism of vector spaces and θ 2 = Id. The
+1 eigenspace is k and the −1 eigenspace is p. The property that it is a Lie algebra
automorphism implies that ♣subsets not
proper? ♣

[k, k] ⊂ k
[k, p] ⊂ p (A.2)
[p, p] ⊂ k

– 130 –
The decomposition (A.1) associated with θ is called a Cartan decomposition. 31
Now let us consider G/K at the global level. The reduction of the τp to a single
involutive automorphism θ of g has a global analog: There is an involutive automorphism
τ of the group. (That is, τ is a group automorphism and τ 2 = Id) such that dτ = θ at the
identity. Given such an involutive automorphism τ we can define a subgroup K to be the
fixed points of τ :
K = {g ∈ G|τ (g) = g} (A.3)
Given such an involution we have a Cartan embedding by the “anti-fixed points”:

G/K ֒→ O := {g ∈ G|τ (g) = g −1 } (A.4)

gK 7→ τ (g)g−1 . (A.5)
Note that this is well-defined and indeed τ (τ (g)g−1 ) = (τ (g)g−1 )−1 because τ is an invo-
lution. One checks it is an embedding by looking at the neighborhood of g = 1. Then
we identify dτ1 = θ. To see it is surjective note that O admits a left G-action by twisted
adjoint action: If g0 ∈ G and g ∈ O then τ (g0 )gg0−1 ∈ O, and the isotropy group of this
action at g = 1 is precisely K. The metric tensor is just the pullback of the usual left-
right-invariant metric −Tr(g−1 dg) ⊗ (g−1 dg). The inversion τg∗ through g∗ ∈ O required
by the definition is τg∗ : g 7→ g∗ g−1 g∗ . One easily checks that this takes O → O and is an
isometry of the metric. To see that g∗ is an isolated fixed point of τg∗ use the left G-action
to translate to g∗ = 1 and use the involution θ on g above. We see that infinitesimally it
is the exponential of elements of p which lie in O in the neighborhood of 1.
It is also worth noting that the Cartan embedding O of G/K is a totally geodesic
submanifold, as follows from the same reasoning used at the end of 1.2.3
Now that we have these definitions we give the 10 classes of compact classical symmetric
spaces:
Whenever G is a compact simple Lie group the homogeneous space (G × G)/Gdiag is
a symmetric space. Suppose the action of the diagonal subgroup is on the right, then we
have an isomorphism of manifolds:

(G × G)/Gdiag ∼
=G (A.6)

where we take (g1 , g2 ) 7→ g1 g2−1 . Warning! This is not a group homomorphism. The
involution τ is just τ : (g1 , g2 ) 7→ (g2 , g1 ). In particular, if we take G = U (n, R) = O(n),
G = U (n, C) = U (n), or G = U (n, H) = Sp(n) then we get a series of 3 classical symmetric
spaces:
(O(n) × O(n))/O(n) (A.7)
(U (n) × U (n))/U (n) (A.8)
(Sp(n) × Sp(n))/Sp(n) (A.9)
31
A Cartan involution of a Lie algebra is an involutive Lie algebra automorphism s such that B(X, sY ) is
positive definite. θ is related to a Cartan involution. ♣ Clarify some confusing terminology. See Helgason
III.7 for the straight story.♣

– 131 –
 c

Figure 17: K and G/K locally divide the group into a product.

Another natural series of classical symmetric spaces are the Grassmannians. These
arise from the involutive automorphism coming from conjugation
!
1k 0
τ (g) = g0 gg0−1 g0 = (A.10)
0 −1n−k

We can consider Grassmannians in real, complex, and quaternionic vector spaces to get

Grk (Rn ) ∼
= O(n)/(O(k) × O(n − k)) (A.11)

Grk (Cn ) ∼
= U (n)/(U (k) × U (n − k)) (A.12)
Grk (Hn ) ∼
= Sp(n)/(Sp(k) × Sp(n − k)) (A.13)
With a little charity (regarding cases with k 6= n − k as nonzero index analogs of the cases
with k = n − k) we can consider this to be three more series of classical symmetric spaces.
Finally, as discussed in Section §1.6, we can put real, complex, or quaternionic struc-
tures on real, complex, or quaternionic spaces (when this makes sense). When these struc-
tures are made compatible with standard Euclidean metrics we obtain moduli spaces of
structures. This gives us:
Real structures on complex vector spaces: Rn ֒→ Cn . Moduli space

U (n)/O(n) (A.14)

This comes from τ (u) = u∗ .

Complex structures on real vector spaces: R2n ∼
= Cn . Moduli space:

O(2n)/U (n) (A.15)

– 132 –
This comes from τ (g) = I0 gI0−1 where I0 is (1.76).
Complex structures on quaternionic vector spaces: Cn ֒→ Hn . Moduli space:

Sp(n)/U (n) (A.16)

Viewing Sp(n) as unitary n × n matrices over the quaternions the involution is τ (g) = −igi,
i.e. conjugation by the unit matrix times i.
Quaternionic structures on complex vector spaces: C2n ∼
= Hn . Moduli space:

U (2n)/Sp(n) (A.17)

Viewing Sp(n) as U Sp(2n) := U (2n) ∩ Sp(2n; C) we can use the involutive automorphism
τ (g) = I0−1 g∗ I0 on U (2n). The fixed points in U (2n) are the group elements with gI0 gtr =
I0 , but this is the defining equation of Sp(2n, C).
When Cartan classified compact symmetric spaces he found the 10 series above (A.7)
- (A.17) together with a finite set of exceptional cases. 32

B. Group cohomology and central extensions

B.1 Central extensions

Now we study an important class of extensions. We change the notation from the previous
section to emphasize this.
Let A be an abelian group and G any group.

Definition A central extension of G by A, 33 is a group G̃ such that

ι π
1→A → G̃ → G→1 (B.1)
such that ι(A) ⊂ Z(G̃).
We stress again that what we called G in the previous section is here called G̃, and
what we called Q in the previous section is here called G.

Example . An example familiar from the quantum mechanical theory of angular momen-
tum, and which we will discuss later is:

1 → Z2 → SU (2) → SO(3) → 1 (B.2)

Here the Z2 ∼
= {±1} is the center of SU (2).
Remarks:

1. Central extensions are important in the theory of projective representations and occur
quite frequently in quantum mechanics. A simple example are the spin representa-
tions of the rotation group. We will explain this relation in more detail later, but for
32
The exceptional cases are presumably all related to the nonassociative real division algebra O, but
explaining how this comes about appears to be nontrivial [22].
33
Some authors say an extension of A by G.

– 133 –
 the moment the reader might find it useful to think about G as a group of classical
symmetries of a physical system and G̃ as a group of corresponding operators in the
quantum mechanical description of that physical system. The fiber of the map π
can be thought of as possible c-number phases which can multiply the operator on
Hilbert space representing a symmetry operation g. For a more detailed account of
this see Chapter *** below.

2. Central extensions appear naturally in quantization of bosons and fermions. The

Heisenberg group is an extension of a translation group. The symplectic group of
linear canonical transformations gets quantum mechanically modified by a central
extension to define something called the metaplectic group.

3. Central extensions are important in the theory of anomalies in quantum field theory.

4. Central extensions are very important in conformal field theory. The Virasoro group,
and the Kac-Moody groups are both nontrivial central extensions of simpler objects.

Exercise Another form of splitting

Show that an equivalent definition of a split exact sequence for a central extension is
that there is a homomorphism t : G̃ → A which is a left-inverse to ι, t(ι(a)) = a.
(Hint: Define s(π(g̃)) = ιt(g̃ −1 ))g̃.)

There is an interesting way to classify central extensions of G by A.

As before let s : G → G̃ be a “section” of π. That is, a map such that

π(s(g)) = g ∀g ∈ G (B.3)

As we have stressed, in general s is not a homomorphism. As in the general discussion

above, when s is in fact a homomorphism we say that the short exact sequence (1.39) splits.
In this case G̃ is isomorphic to a direct product G̃ ∼
= A × G via

ι(a)s(g) → (a, g) (B.4)

When the sequence splits the semidirect product of the previous section is a direct
product because A is central, so ωg (a) = a.
Now, let us allow that (B.1) does not necessarily split. Let us choose any section
s and measure by how much s differs from being a homomorphism by considering the
combination:
s(g1 )s(g2 ) (s(g1 g2 ))−1 . (B.5)

– 134 –
Now the quantity (B.5) is in the kernel of π and hence in the image of ι. Since ι is injective
we can define a function fs : G × G → A by the equation

ι(fs (g1 , g2 )) := s(g1 )s(g2 ) (s(g1 g2 ))−1 . (B.6)

That is, we can write:

s(g1 )s(g2 ) = ι(fs (g1 , g2 ))s(g1 g2 ) (B.7)

The function fs satisfies the important cocycle identity

f (g1 , g2 g3 )f (g2 , g3 ) = f (g1 , g2 )f (g1 g2 , g3 ) (B.8)

Exercise
Verify (B.8) by using (B.6) to compute s(g1 g2 g3 ) in two different ways.
(Note that simply substituting (B.6) into (B.8) is not obviously going to work because
G̃ need not be abelian.)

Exercise Simple consequences of the cocycle identity

a.) By putting g1 = 1 and then g3 = 1 show that

f (g, 1) = f (1, g) = f (1, 1) ∀g ∈ G (B.9)

b.) Show that

f (g, g−1 ) = f (g −1 , g). (B.10)

Now we introduce some fancy terminology:

Definition: In general
1. A 2-cochain on G with values in A, C 2 (G, A) is a function

f : G×G →A (B.11)

Note that C 2 (G, A) is naturally an abelian group.

2. A 2-cocycle f ∈ Z 2 (G, A) is a 2-cochain f : G × G → A satisfying (B.8). Z 2 (G, A)
inherits an abelian group structure from C 2 (G, A).

So, in this language, given a central extension of G by A and a section s we naturally

obtain a two-cocycle fs ∈ Z 2 (G, A) via (B.6).

– 135 –
 Now, if we choose a different section ŝ then

ŝ(g) = s(g)ι(t(g)) (B.12)

for some function t : G → A. It is easy to check that

fŝ (g1 , g2 ) = fs (g1 , g2 )t(g1 )t(g2 )t(g1 g2 )−1 (B.13)

where we have used that ι(A) is central in G̃.

Definition: In general two 2-cochains f and fˆ are said to differ by a coboundary if they
satisfy

fˆ(g1 , g2 ) = f (g1 , g2 )t(g1 )t(g2 )t(g1 g2 )−1 (B.14)

for some function t : G → A.

One can readily check that if f is a cocycle then any other fˆ differing by a coboundary
is also a cocycle. Moreover, being related by a cocycle defines an equivalence relation on
the set of cocycles f ∼ fˆ. Thus, we may define:

Definition: The group cohomology H 2 (G, A) is the set of equivalence classes of 2-cocycles
modulo equivalence by coboundaries. Moreover, this set carries a natural structure of an
abelian group.
The group multiplication making H 2 (G, A) into an abelian group is simply defined by

(f1 · f2 )(g, g′ ) = f1 (g, g ′ ) · f2 (g, g′ ) (B.15)

This descends to a well-defined muiltiplication on cohomology classes: [f1 ] · [f2 ] := [f1 · f2 ].

Now, the beautiful theorem states that group cohomology classifies central extensions:

Theorem: Isomorphism classes of central extensions of G by an abelian group A are in

1-1 correspondence with the second cohomology set H 2 (G, A). Moreover, with its abelian
group structure identity corresponds to the split extension A × G.

Proof : From (B.6)(B.8)(B.13) we learn that given a central extension we can unambigu-
ously form a group cohomology class which is independent of the choice of section. More-
over, if G̃ ∼
= G̃′ are isomorphic central extensions and ψ : G̃ → G̃′ is an isomorphism, then
ψ can be used to map sections of G̃ → G to sections of G̃′ → G: s′ (g) = ψ(s(g)). Then

s′ (g1 )s′ (g2 ) = ψ(s(g1 ))ψ(s(g2 ))

= ψ(s(g1 )s(g2 ))
= ψ(ι(fs (g1 , g2 ))s(g1 g2 )) (B.16)
= ψ(ι(fs (g1 , g2 )))ψ(s(g1 g2 ))
= ι′ (fs (g1 , g2 ))s′ (g1 g2 )

– 136 –
and hence we assign precisely the same 2-cocycle f (g1 , g2 ) to the section s′ . Hence the
isomorphism class of a central extension maps unambiguously to a cohomology class [f ].
Conversely, given a cohomology class [f ] we may construct a corresponding central
extension as follows. Choose a representative 2-cocycle f . With such an f we may define
G̃ = A × G as a set with multiplication law:

(a1 , g1 )(a2 , g2 ) = (a1 a2 f (g1 , g2 ), g1 g2 ) (B.17)

Now suppose that we use two 2-cocycles f and f ′ which are related by a coboundary
as in (B.14) above. Then we claim that the map ψ : G̃ → G̃′ defined by

ψ : (a, g) → (at(g)−1 , g) (B.18)

is an isomorphism of groups. (Check this!) On the other hand, we just showed above that
if [f ] 6= [f ′ ] then G̃ cannot be isomorphic to G̃′ . ♠

Remark: Using a coboundary one can usefully simplify cocycles. For example, by setting
t(1) = f (1, 1)−1 we may assume f (1, 1) = 1. Then, by (B.9) we have f (g, 1) = f (1, g) = 1
for all g. Similarly, if g 6= g−1 we may put f (g, g−1 ) = f (g−1 , g) = 1. If g = g−1 then we
might not be able to set f (g, g) = 1. We can “preserve this gauge” with further coboundary
transformations that satisfy t(1) = 1 and t(g −1 ) = t(g)−1 .

Example 1 . Extensions of Z2 by Z2 . WLOG we can take f (1, 1) = f (1, σ) = f (σ, 1) = 1.

Then we have two choices: f (σ, σ) = 1 or f (σ, σ) = σ. Both of these choices satisfies
the cocycle identity. In other words H 2 (Z2 , Z2 ) = Z2 . For the choice f = 1 we obtain
G̃ = Z2 × Z2 . For the nontrivial choice f (σ, σ) = σ we obtain G̃ = Z4 . Let us see this in
detail. We’ll let σ1 ∈ A ∼ = Z2 and σ2 ∈ G ∼ = Z2 be the nontrivial elements so we should
write f (σ2 , σ2 ) = σ1 . Note that (σ1 , 1) has order 2, but then

(1, σ2 ) · (1, σ2 ) = (f (σ2 , σ2 ), 1) = (σ1 , 1) (B.19)

shows that (1, σ2 ) has order 4. Moreover (σ1 , σ2 ) = (σ1 , 1)(1, σ2 ) = (1, σ2 )(σ1 , 1).
Thus,

Ψ : (σ1 , 1) → ω 2 = −1
(B.20)
Ψ : (1, σ2 ) → ω

where ω is a primitive 4th root of 1 defines an isomorphism. In conclusion, the nontrivial

central extension of Z2 by Z2 is:

1 → Z2 → Z4 → Z2 → 1 (B.21)

Recall that Z4 is not isomorphic to Z2 × Z2 .

Example 2. The generalization of the previous example to an odd prime p is extremely

instructive. These are Zp × Zp and Zp2 . So, let us study in detail the extensions

1 → Zp → G → Zp → 1 (B.22)

– 137 –
where we will write our groups multiplicatively. Now, using methods of topology one can
show that 34
H 2 (Zp , Zp ) ∼
= Zp . (B.23)
On the other hand, we know from the class equation and Sylow’s theorems that there
are exactly two groups of order p2 , up to isomorphism! How is that compatible with (B.23)?
The answer is that there can be nonisomorphic extensions involving the same central group.
To see this, let us examine in detail the standard extension:

1 → Zp → Zp2 → Zp → 1 (B.24)

We write the first, second and third groups in this sequence as

Zp = hσ1 |σ1p = 1i
2
Zp2 = hα|αp = 1i (B.25)
Zp = hσ2 |σ2p = 1i

respectively. Then the first injection must take

ι(σ1 ) = αp (B.26)

since it must be an injection and it must take an element of order p to an element of order
p. The standard sequence then takes the second arrow to be reduction modulo p, so

π(α) = σ2 (B.27)

Now, we try to choose a section of π. Let us try to make it a homomorphism. Therefore

we must take s(1) = 1. What about s(σ2 )? Since π(s(σ2 )) = σ2 we have a choice: s(σ2 )
could be any of
α, αp+1 , α2p+1 , . . . , α(p−1)p+1 (B.28)
Here we will make the simplest choice s(σ2 ) = α. The reader can check that the discussion
is not essentially changed if we make one of the other choices. (After all, this will just
change our cocycle by a coboundary!)
Now that we have chosen s(σ2 ) = α, if s were a homomorphism then we would be
forced to take:
s(1) = 1
s(σ2 ) = α
s(σ22 ) = α2 (B.29)
.. ..
. .
s(σ2p−1 ) = αp−1

But now we are stuck! The property that s is a homomorphism requires two contradictory
things. On the one hand, we must have s(1) = 1 for any homomorphism. On the other
34
You can also show it by examining the cocycle equation directly.

– 138 –
hand, from the above equations we also must have s(σ2p ) = αp . But because σ2p = 1 these
requirements are incompatible. Therefore, with this choice of section we find a nontrivial
cocycle as follows:
(
k ℓ k+ℓ −1 1 k+ℓ<p
s(σ2 )s(σ2 )s(σ2 ) = (B.30)
αp p ≤ k + ℓ
and therefore,
(
1 k+ℓ<p
f (σ2k , σ2ℓ ) = (B.31)
σ1 p≤k+ℓ
In these equations we assume 1 ≤ k, ℓ ≤ p − 1. We know the cocycle is nontrivial because
Zp × Zp is not isomorphic to Zp2 .
But now let us use the group structure on the group cohomology. [f ]r is the cohomology
class represented by
(
r k ℓ 1 k+ℓ<p
f (σ2 , σ2 ) = (B.32)
σ1r p ≤ k + ℓ

This corresponds to replacing (B.27) by

πr (α) = (σ2 )r (B.33)

and the sequence will still be exact, i.e. ker(πr ) = Im (ι), if (r, p) = 1, that is, if we compose
the standard projection with an automorphism of Zp . Thus πr also defines an extension of
the form (B.24). But we claim that it is not isomorphic to the standard extension! To see
this let us try to construct ψ so that

ι ♣♣♣
88 hαi ◆◆◆πr (B.34)
◆◆◆
♣♣♣ &&
1 // hσ1 i
◆◆◆ι
ψ
88 hσ2 i
// 1
◆◆◆  ♣π♣♣♣♣
&& ♣
hαi

In order for the right triangle to commute we must have ψ(α) = αr , but then the left triangle
will not commute. Thus the extensions π1 , . . . , πp−1 , all related by outer automorphisms
of the quotient group Zp = hσ2 i, define inequivalent extensions with the same group Zp2
in the middle.
In conclusion, we describe the group of isomorphism classes of central extensions of Zp
by Zp as follows: The identity element is the trivial extension

1 → Zp → Zp × Zp → Zp → 1 (B.35)

and then there is an orbit of (p − 1) nontrivial extensions of the form

1 → Zp → Zp2 → Zp → 1 (B.36)

acted on by Aut(Zp ).

– 139 –
B.2 Group cohomology in other degrees
Motivations:
a.) The word “cohomology” suggests some underlying chain complexes, so we will
show that there is such a formulation.
b.) There has been some discussion of higher degree group cohomology in physics in

1. The theory of anomalies (Faddeev-Shatashvili; Segal; Carey et. al.; Mathai et. al.;
... )

2. Classification of rational conformal field theories (Moore-Seiberg; Dijkgraaf-Vafa-

Verlinde-Verlinde; Dijkgraaf-Witten; Kapustin-Saulina)

3. Condensed matter/topological phases of matter (Kitaev; Wen; ...)

4. Three-dimensional Chern-Simons theory and three dimensional supersymmetric gauge

theory.

So here we’ll just give a few definitions.

B.2.1 Definition
Suppose we are given any group G and an abelian group A (written additively in this
section) and a homomorphism
α : G → Aut(A) (B.37)
Definition: An n-cochain is a function φ : G×n → A. The space of n-cochains is
denoted C n (G, A).
Note that C n (G, A) is an abelian group using the abelian group structure of A on the
values of φ, that is: (φ1 + φ2 )(~g ) := φ1 (~g ) + φ2 (~g ).
Define a group homomorphism: d : C n (G, A) → C n+1 (G, A)

(dφ)(g1 , . . . , gn+1 ) := αg1 (φ(g2 , . . . , gn+1 ))

− φ(g1 g2 , g3 , . . . , gn+1 ) + φ(g1 , g2 g3 , . . . , gn+1 ) ± · · · + (−1)n φ(g1 , . . . , gn−1 , gn gn+1 )
+ (−1)n+1 φ(g1 , . . . , gn )
(B.38)

We interpret a 0-cochain φ0 to be some element φ0 = a ∈ A. Then we have, for n = 0:

(dφ0 )(g) = αg (a) − a (B.39)

For n = 1 φ1 : G → A and the differential acts as:

(dφ1 )(g1 , g2 ) = αg1 (φ1 (g2 )) − φ1 (g1 g2 ) + φ1 (g1 ) (B.40)

(dφ2 )(g1 , g2 , g3 ) = αg1 (φ2 (g2 , g3 )) − φ2 (g1 g2 , g3 ) + φ2 (g1 , g2 g3 ) − φ2 (g2 , g3 ) (B.41)

(dφ3 )(g1 , g2 , g3 , g4 ) = αg1 (φ3 (g2 , g3 , g4 ))−φ3 (g1 g2 , g3 , g4 )+φ3 (g1 , g2 g3 , g4 )−φ3 (g1 , g2 , g3 g4 )+φ3 (g1 , g2 , g3 )
(B.42)

– 140 –
 The set of n-cocycles is defined to be the subgroup Z n (G, A) ⊂ C n (G, A) that satisfy
dφn = 0.
Next, one can check that for any φ, d(dφ) = 0. (We give a simple proof below.)
Therefore, we can define a subgroup B n (G, A) ⊂ Z n (G, A) by
B n (G, A) := {φn |∃φn−1 s.t. dφn−1 = φn } (B.43)
Then the group cohomology is defined to be the quotient
H n (G, A) = Z n (G, A)/B n (G, A) (B.44)
Remarks:

1. Remembering that we are now writing our abelian group A additively, we see that
the equation (dφ2 ) = 0 is just the twisted 2-cocycle conditions, and φ′2 = φ2 + dφ1 are
two different twisted cocycles related by a coboundary. See equations **** above.

2. Homogeneous cocycles: A nice way to prove that d2 = 0 is the following. We define

homogeneous n-cochains to be maps ϕ : Gn+1 → A which satisfy
ϕ(hg0 , hg1 , . . . , hgn ) = αh (ϕ(g0 , g1 , . . . , gn )) (B.45)
Let C n (G, A) denote the abelian group of such homogeneous group cochains. Define
δ : C n (G, A) → C n+1 (G, A) (B.46)
by
n+1
X
δφ(g0 , . . . , gn+1 ) := (−1)i ϕ(g0 , . . . , gbi , . . . , gn+1 ) (B.47)
i=0
where gbi means the argument is omitted. It is then very straightforward to prove
that δ2 = 0. Indeed, if ϕ ∈ C n−1 (G, A) we compute:
n+1
X i−1
X
δ2 ϕ(g0 , . . . , gn+1 ) = (−1)i (−1)j ϕ(g0 , . . . , gbj , . . . , gbi , . . . , gn+1 )
i=0 j=0
n+1
X

− (−1)j ϕ(g0 , . . . , gbi , . . . , gbj , . . . , gn+1 )
j=i+1
X (B.48)
= (−1)i+j ϕ(g0 , . . . , gbj , . . . , gbi , . . . , gn+1 )
0≤j<i≤n+1
X
− (−1)i+j ϕ(g0 , . . . , gbi , . . . , gbj , . . . , gn+1 )
0≤i<j≤n+1

=0
Now, we can define an isomorphism ψ : C n (G, A) → C n (G, A) by defining
φn (g1 , . . . , gn ) := ϕn (1, g1 , g1 g2 , . . . , g1 · · · gn ) (B.49)
That is, when φn and ϕn are related this way we say φn = ψ(ϕn ). Now one can check
that the simple formula (B.47) becomes the more complicated formula (B.38). Put
more formally: there is a unique d so that dψ = ψδ.

– 141 –
 3. Where do all these crazy formulae come from? The answer is in topology. We will
indicate it briefly in our discussion of categories and groupoids below.

B.3 The topology behind group cohomology

Now, let us show that this point of view on the definition of a group can lead to a very
nontrivial and beautiful structure associated with a group.
An interesting construction that applies to any category is its associated simplicial
space |C|.
This is a simplicial space whose simplices are:
0. 0-simplices = objects
1. 1-simplices = ∆1 (f ) associated to each morphism f : x0 → x1 ∈ X1 .
2. 2-simplices: ∆(f1 , f2 ) associated composable morphisms (f1 , f2 ) ∈ X2 .
3. 3-simplices: ∆(f1 , f2 , f3 ) associated to 3 composable morphisms, i.e. elements of:

X3 = {(f1 , f2 , f3 ) ∈ X1 × X1 × X1 |p0 (fi ) = p1 (fi+1 )} (B.50)

And so on. See Figures 14 and 15. The figures make clear how these simplices are
glued together:
∂∆1 (f ) = x1 − x0 (B.51)
∂∆2 (f1 , f2 ) = ∆1 (f1 ) − ∆1 (f1 f2 ) + ∆1 (f2 ) (B.52)
and for Figure 15 view this as looking down on a tetrahedron. Give the 2-simplices of
Figure 14 the counterclockwise orientation and the boundary of the simplex the induced
orientation from the outwards normal. Then we have

∂∆(f1 , f2 , f3 ) = ∆(f2 , f3 ) − ∆(f1 f2 , f3 ) + ∆(f1 , f2 f3 ) − ∆(f1 , f2 ) (B.53)

Note that on the three upper faces of Figure 15 the induced orientation is the ccw orien-
tation for ∆(f1 , f2 f3 ) and ∆(f2 , f3 ), but with the cw orientation for ∆(f1 f2 , f3 ). On the
bottom fact the inward orientation is ccw and hence the outward orientation is −∆(f1 , f2 ).
Clearly, we can keep composing morphisms so the space |C| has simplices of arbitrarily
high dimension, that is, it is an infinite-dimensional space.
Let look more closely at this space for the case of a group, regarded as a category with
one object. Then in the above pictures we identify all the vertices with a single vertex.
For each group element g we have a one-simplex ∆1 (g) beginning and ending at this
vertex.
For each ordered pair (g1 , g2 ) we have an oriented 2-simplex ∆(g1 , g2 ), etc. We simply
replace fi → gi in the above formulae, with gi now interpreted as elements of G:

∂∆(g) = 0 (B.54)

∂∆(g1 , g2 ) = ∆1 (g1 ) + ∆1 (g2 ) − ∆1 (g1 g2 ) (B.55)

∂∆(g1 , g2 , g3 ) = ∆(g2 , g3 ) − ∆(g1 g2 , g3 ) + ∆(g1 , g2 g3 ) − ∆(g1 , g2 ) (B.56)
See Figure 15.

– 142 –
 And so on.
To put this more formally: We have n + 1 maps from Gn → Gn−1 for n ≥ 1 given by
d0 (g1 , . . . , gn ) = (g2 , . . . , gn )
d1 (g1 , . . . , gn ) = (g1 g2 , g3 , . . . , gn )
d2 (g1 , . . . , gn ) = (g1 , g2 g3 , g4 , . . . , gn )
······ (B.57)
······
n−1
d (g1 , . . . , gn ) = (g1 , . . . , gn−1 gn )
dn (g1 , . . . , gn ) = (g1 , . . . , gn−1 )
On the other hand, we can view an n-simplex ∆n as
n
X
∆n := {(t0 , t1 , . . . , tn )|ti ≥ 0 & ti = 1} (B.58)
i=0

Now, there are also (n + 1) face maps which map an (n − 1)-simplex ∆n−1 into one of the
(n + 1) faces of the n-simplex ∆n :

d0 (t0 , . . . , tn−1 ) = (0, t0 , . . . , tn−1 )

d1 (t0 , . . . , tn−1 ) = (t0 , 0, t1 , . . . , tn−1 )
······ (B.59)
······
dn (t0 , . . . , tn−1 ) = (t0 , . . . , tn−1 , 0)

di embeds the (n − 1) simplex into the face ti = 0 which is opposite the ith vertex ti = 1
of ∆n .
Now we identify
(∐∞ n
n=0 ∆n × G ) / ∼

via
(di (~t), ~g ) ∼ (~t, di (~g )). (B.60)
The space we have constructed this way has a homotopy type denoted BG. Even for
the simplest nontrivial group G = Z/2Z the construction is quite nontrivial and BG has
the homotopy type of RP ∞ .
Now, an n-cochain in C n (G, Z) (here we take A = Z for simplicity) is simply an
assignment of an integer for each n-simplex in BG. Then the coboundary and boundary
maps are related by
hdφn , ∆i = hφn , ∂∆i (B.61)
and from the above formulae we recover, rather beautifully, the formula for the coboundary
in group cohomology.
Remark: When we defined group cohomology we also used homogeneous cochains.
This is based on defining G as a groupoid from its left action and considering the mapping
of groupoids G//G → pt//G. ♣Explain more
here? ♣

– 143 –
References
[1] A.A. Albert, Structure of Algebras, AMS monograph
[2] A. Alekseev, L. D. Faddeev and S. L. Shatashvili, “Quantization of symplectic orbits of
compact Lie groups by means of the functional integral,” J. Geom. Phys. 5, 391 (1988).
[3] A. Altland and M. Zirnbauer, “Nonstandard symmetry classes in mesoscopic
normal-superconducting hybrid structures,” Phys. Rev. B55 (1997) pp. 1142-1161.
[4] A. Altland and M. R. Zirnbauer, “Nonstandard symmetry classes in mesoscopic
normal-superconducting hybrid structures,” arXiv:cond-mat/9602137 ; Phys. Rev. B 55,
1142 (1997).
[5] O. Alvarez, I. M. Singer and P. Windey, “Quantum Mechanics and the Geometry of the Weyl
Character Formula,” Nucl. Phys. B 337, 467 (1990).
[6] O. Alvarez, I. M. Singer and P. Windey, “The Supersymmetric sigma model and the geometry
of the Weyl-Kac character formula,” Nucl. Phys. B 373, 647 (1992) [hep-th/9109047].
[7] M.F. Atiyah, R. Bott, and A. Shapiro, ”Clifford Modules”, Topology 3 (1964), ((Suppl. 1)):
338
[8] M.F. Atiyah and F. Hirzebruch, “Vector Bundles and Homogeneous Spaces,” Proc. Symp.
Pure Math. 3 (1961)53
[9] M.F. Atiyah and I.G. MacDonald, Introduction to Commutative Algebra, Addison-Wesley
1969
[10] M.F. Atiyah and G.B. Segal, “The Index of Elliptic Operators: II,” Ann. Math. 87(1968)531
[11] M.F. Atiyah and I.M. Singer, “Index theory for skew-adjoint Fredholm operators,” Publ.
Math. IHES, 37 (1969) pp. 5-26
[12] J.C. Baez, “The Octonions.” arXiv:math/0105155 [math.RA];
Bull.Am.Math.Soc.39:145-205,2002
[13] V. Bargmann, “Note on Wigner’s Theorem on Symmetry Operations,” J. Math. Phys. 5
(1964) 862.
[14] Bourbaki, Algebra, ch. 8
[15] R.A. Brualdi and H. Schneider, Determinantal Identities: Gauss, Schur, Sylvester,
Kronecker, Jacobi, Binet, Laplace, Muir, and Cayley, Linear Algebra and its applications,
Vol. 52-53, (1983), pp. 769-791
[16] B. Conrey, http://arxiv.org/pdf/math/0005300v1.pdf
[17] P. Deligne, et. al. , Quantum Fields and Strings: A Course for Mathematcians, vol. 1, AMS,
1999
[18] P. Deligne, Notes on Spinors [hep-th/9710198].
[19] Yu. A. Drozd and V.V. Kirichenko, Finite Dimensional Algebras, Springer 1994
[20] F.J. Dyson, “The Threefold Way. Algebraic Structure of Symmetry Groups and Ensembles in
Quantum Mechanics,’ J. Math. Phys. 3, 1199 (1962)
[21] D. Eisenbud, Commutative Algebra with a view toward algebraic geometry, Springer GTM 150

– 144 –
[22] J.-H. Eschenburg, “Symmetric spaces and division algebras.”
[23] Lukasz Fidkowski, Alexei Kitaev, “ Topological phases of fermions in one dimension,”
arXiv:1008.4138
[24] Freed, arXiv:1211.2133.
[25] D. Freed, “Lectures on twisted K-theory and orientifolds,”
http://ncatlab.org/nlab/files/FreedESI2012.pdf
[26] D. S. Freed and G. W. Moore, “Twisted equivariant matter,” arXiv:1208.5055 [hep-th].
[27] L. Fu and C. L. Kane, “Topological insulators with inversion symmetry,” Phys. Rev. B 76
(2007), 045302, arXiv:cond-mat/0611341.
[28] L. Fu, C. L. Kane, and E. J. Mele, “Topological Insulators in Three Dimensions,” Phys. Rev.
Lett. 98 (2007), 106803, arXiv:cond-mat/0607699.
[29] P. Griffiths and J. Harris, Principles of Algebraic Geometry,
[30] J. Harris, Algebraic Geometry, Springer GTM 133
[31] P. Heinzner, A. Huckleberry, and M. Zirnbauer, “Symmetry classes of disordered fermions,”
arXiv:math-ph/0411040 ; Commun.Math.Phys. 257 (2005) 725-771
[32] Hitchin, Nigel J. (1974), ”Harmonic spinors”, Advances in Mathematics 14: 155, MR 358873
[33] C. L. Kane and E. J. Mele, “Quantum Spin Hall Effect in Graphene,” Phys. Rev. Lett. 95
(2005), 226801; “Z2 Topological Order and the Quantum Spin Hall Effect,” Phys. Rev. Lett.
95 (2005), 146802.
[34] M.L. Mehta, Random Matrices
[35] L. Michel and B.I. Zhilinskii, “Symmetry, Invariants, Topology: Basic Tools,” Physics
Reports 341(2001), pp. 11-84
[36] A. Kitaev, “Periodic table for topological insulators and superconductors,” AIP Conf. Proc.
1134 (2009) 22-30; arXiv:0901.2686
[37] L. Michel, “Fundamental Concepts for the Study of Crystal Symmetry,” Physics Reports
341(2001), pp. 265-336
[38] J. Milnor, Morse Theory, Ann. Math. Studies, 51, Princeton (1963)
[39] G. W. Moore, “K theory from a physical perspective,” hep-th/0304018.
[40] G. Moore, Lecture notes at
http://www.physics.rutgers.edu/ gmoore/QuantumSymmetryKTheory-Part1.pdf
[41] G. Moore, Lecture at
http://www.physics.rutgers.edu/ gmoore/SCGP-TWISTEDMATTER-2013D.pdf
[42] G. Moore, “Linear ALgebra User’s Manual,” Lecture notes from Applied Group Theory
Course, Spring 2013.
[43] G. Moore, “Quantum Symmetries and Compatible Hamiltonians.” Notes available at
http://www.physics.rutgers.edu/ gmoore/QuantumSymmetryBook.pdf
[44] J. E. Moore and L. Balents, “Topological invariants of time-reversal-invariant band
structures,” Phys. Rev. B 75 (2007), 121306, arXiv:cond-mat/0607314.

– 145 –
[45] A. Pressley and G. Segal, Loop Groups, OUP
[46] R. Carter, G. Segal, and I. Macdonald, Lectures on Lie Groups and Lie Algebras, London
Math Soc. Student Texts 32.
[47] B. Simon, “Quantum Dynamics: From Automorphism to Hamiltonian.”
[48] M. Stone, C-K. Chiu, and A. Roy, “Symmetries, dimensions and topological insulators: the
mechanism behind the face of the Bott clock,” arXiv:1005:3213.
[49] R. Simon, N. Mukinda, S. Chaturvedi, and V. Srinivasan, “Two elementary proofs of the
Wigner theorem on symmetry in quantum mechanics,” arXiv:0808.0779
[50] A.M. Turner, Y. Zhang, R.K. Mong, and A. Vishwanath, “Quantized Response and Topology
of Insulators with Inversion Symmetry,” arXiv:1010.4335
[51] C.T.C. Wall, “Graded Brauer Groups,” Journal für die reine und angewandte Mathematik,
213 (1963/1964), 187-199
[52] S. Weinberg, Quantum Field Theory, vol. 1, pp.
[53] E. Witten, “D-branes and K theory,” JHEP 9812, 019 (1998) [hep-th/9810188].
[54] M.R. Zirnbauer, “Symmetry classes in random matrix theory,” arXiv:math-ph/0404058
[55] M.R. Zirnbauer, “Symmetry Classes,” arXiv:1001.0722

– 146 –