Bulk-Edge Correspondence For Two-Dimensional Topological Insulators

Bulk-edge correspondence for two-dimensional topological
insulators
G.M. Graf and M. Porta
Theoretische Physik, ETH Zurich, 8093 Zurich, Switzerland
arXiv:1207.5989v2 [math-ph] 19 Mar 2013
March 20, 2013
Abstract
Topological insulators can be characterized alternatively in terms of bulk or edge prop-
erties. We prove the equivalence between the two descriptions for two-dimensional solids in
the single-particle picture. We give a new formulation of the Z2 -invariant, which allows for a
bulk index not relying on a (two-dimensional) Brillouin zone. When available though, that
index is shown to agree with known formulations. The method also applies to integer quan-
tum Hall systems. We discuss a further variant of the correspondence, based on scattering
theory.
1 Introduction
Topological insulators are materials that behave as ordinary insulators in the bulk, in that
they exhibit an excitation gap, whereas the edge has robust, gapless modes. They have been
theoretically predicted as a class [16] and as a specific compound [2], in which the effect was
then observed [19, 15]. In analogy with quantum Hall (QH) systems, the presence of edge states
is a robust property of the system, in that it is stable under moderate changes of parameters.
That calls for an explanation in terms of topological invariants.
As pointed out in [16], a key feature of topological insulators is fermionic time-reversal
symmetry. It was shown that two-dimensional time-reversal symmetric insulators admit two
topologically distinct phases: The phases can not be deformed into one another as long as the
bulk gap and the symmetry are there. In one phase the insulator is an ordinary one; that is, it
does not carry currents in the bulk nor at the edges. In the other phase edge states are present.
They come in pairs with opposite velocity and spin. This phase is called the quantum spin Hall
(QSH) phase and its signature is a nonzero spin current, which in contrast to the QH current is
not quantized, as a rule. For integer quantum Hall systems the topological invariant is a Chern
number, and can be read off from the value of the Hall conductivity; for quantum spin Hall
systems the topological invariant is not expressing the value of the spin current, but the parity
of the number of edge states.
In this paper we prove the bulk-edge correspondence for two-dimensional topological insula-
tors and for independent particles. For short we introduce a Z2 bulk topological invariant (to be
shown equivalent to others), and we show that it is equal to the number of pairs of edge states
1
modulo 2. For sure, that duality has been discussed in the literature, see e.g. [9], but we main-
tain that there is room for a strict mathematical approach; just as for quantum Hall systems,
where Laughlin’s argument has gained in precision and detail by the subsequent mathematical
discussion. We should though mention [25, 1], whose results are further compared to ours below.
The paper is organized as follows. In Section 2.1 we consider a general class of two-
dimensional single-particle lattice Hamiltonians, with and without edge. They are symmetric
under fermionic time reversal and hence potentially describe topological insulators. One note-
worthy feature is that periodicity is postulated only in the direction parallel to the edge. In
Section 2.2 we formulate the duality in its most basic version, in a precise though still prelim-
inary form. The purpose of Section 3 is to show how graphene and related model materials,
which have been considered as candidates for topological insulators, fit into our scheme. As a
side remark, we give a simple new proof of the absence of edge states for armchair boundary
conditions. In Section 4 we discuss some abstract vector bundles on the 2-torus. By considering
their sections and transition matrices, we classify them in terms of a Z2 -invariant, defined simi-
larly but not identically to one found in [9]. So equipped, we define the bulk index in Section 5
and state in full the basic version of duality (Theorem 5.4). It is immediately followed by the
main steps of the proof, while technical details are postponed to main part of Section 7. The
part of the article so far described provides a full and self-contained account of basic bulk-edge
duality. It is the most general part, and yet makes up for less than half of its length.
In Section 5.1 we formulate a bulk index for the specific and more familiar case where the
lattice Hamiltonian is doubly periodic. As it is to be expected, the relevant 2-torus is now
the Brillouin zone. We discuss how that index arises from the general one, leaving details to
Section 8. All the results obtained up to that point have a counterpart in the case of QH
systems, which we present in Section 6 for illustration, because they are simpler in that context.
As a matter of fact, for such systems we include an independent, alternate version of the duality
based on scattering theory and more precisely on Levinson’s theorem (Section 6.2), as well as a
comparison between the two versions (Section 6.3). Related details are found in Section 9. We
conjecture an analogous alternate version for QSH systems. Finally, Section 7.1 contains some
results about indices, including a comparison of ours with some of those found in the literature.
Let us comment on the relation to other work. In [16] an edge and a bulk characterization of a
topological insulator is given, but without proof of a mathematical link. In [25] a correspondence
between bulk and edge description is given. There however the bulk is put in correspondence
with twisted boundary conditions allowing tunneling between a pair of edges; they include open
boundaries as a special case. Moreover, the bulk invariant used there is physically different from
a Z2 -classification and may not reflect itself in open boundaries. In [1] the bulk-edge duality for
topological insulators is derived in the same sense as ours, among other results. The setting,
however, is not quite as general as ours, as far as the side of the bulk is concerned. As the edge is
concerned, an index related to the Maslov index is defined and shown to be equivalent to that of
[16]. As for the bulk, it is assumed that the Hamiltonian is the perturbation of one commuting
with spin, and hence consisting to several copies of a quantum Hall system. It is assumed that
the gap remains open as the perturbation is switched on. Thus spin resolved Chern numbers
can be defined by homotopy [24, 29] and the duality inferred from that of the quantum Hall
case [28].
Further, assorted comments are: Some of our examples are discussed in [1] in a similar vein.
Related indices for topological insulators are discussed in [1, 9, 12, 16, 20, 27]. A bulk-edge
duality result for general symmetry classes is found in [4], differing from the results presented
2
here in various ways; for instance the index takes values in Z, with just hints at the Z2 case, at
least in two dimensions. A choice of a torus and bundle similar to ours in Sect. 2.2 is found in
[30]. The results of Section 5.1 depend on the analytic properties of band functions [18]. In [13]
bulk-edge duality for Hall systems is pinpointed at the birth of edge states of band edges. This
insight is the reason for using Levinson’s theorem, though the method is otherwise different and
the result more general. Last but not least, the existence of a complementary approach to gapped
systems, including interacting ones, should be mentioned. It is based on effective, topological
field theories as a tool to explore the response of a system in the limit of low frequencies and
long wavelengths; see [31, 5, 8, 32, 6, 7] for early examples.
2 Setting and results

We shall introduce a class of bulk, resp. edge, single-particle discrete Schrödinger operators,
by which we describe insulators, topological or otherwise, which extend over a plane, resp. a
half-plane. We conclude the section with a loose description of the results.
2.1 The Schrödinger operators

Consider, at first and in less than final generality, a tight binding Hamiltonian with nearest
neighbor hopping on the lattice Z ˆ Z, resp. N ˆ Z. The Hamiltonian is assumed to be periodic
in the direction along the edge, but not necessarily across it. The period may be taken to be
equal to one without loss: In fact sites within a period may be regarded as labels of internal
degrees of freedom, among others like e.g. spin. We may thus perform a Bloch decomposition
with respect to the longitudinal quasi-momentum k P S 1 :“ R{2πZ, which remains a good
quantum number even in presence of the boundary.
We so end up with a family of Hamiltonians Hpkq defined on the one-dimensional lattice
Z Q n, resp. N, and acting on wave-functions ψn P CN , where N is the number of internal
degrees of freedom.
More generally, these objects are stated as follows.
Definition 2.1. [Bulk Hamiltonian] The Hamiltonian, acting on ψ P `2 pZ; CN q and parametrized
by k P S 1 , is
pn P Z, ψn P CN q .
` ˘
Hpkqψ n “ Apkqψn´1 ` Apkq˚ ψn`1 ` Vn pkqψn , (1)
The potential Vn pkq and the hopping matrices Apkq are N ˆN matrices having a C 1 -dependence
on k, uniformly in n. We assume Vn pkq “ Vn pkq˚ , where ˚ denotes the matrix-adjoint, and that
Apkq P GLpN q. (Recall that GLpN q Ă MN pCq consists of invertible matrices of order N .)
We then consider the restriction of the Hamiltonian to N “ t1, 2, . . .u (we find it convenient
to omit zero), while allowing for changes within a finite distance n0 ě 0 from the edge.
Definition 2.2. [Edge Hamiltonian] The Hamiltonian, acting on ψ P `2 pN; CN q, is
pn P N, ψn P CN q ,
` 7 ˘
H pkqψ n “ Apkqψn´1 ` Apkq˚ ψn`1 ` Vn7 pkqψn , (2)
where V 7 satisfies the above properties of V , as well as
Vn7 pkq “ Vn pkq , pn ą n0 q . (3)
3
Moreover we assume the Dirichlet boundary condition, meaning that for n “ 1 Eq. (2) is to be
read with ψ0 “ 0.
Remark 2.3. The Dirichlet condition is by no means special. Since
pϕ, H 7 ψq ´ pH 7 ϕ, ψq “ ϕ˚1 Aψ0 ´ ϕ˚0 A˚ ψ1 ,
any boundary condition ψ0 “ Λψ1 , (Λpkq P MN pCq) defines a self-adjoint Hamiltonian if

pAΛq˚ “ AΛ. That amounts to the Dirichlet condition after adding δn1 AΛ to Vn7 .
Contrary to what the above motivation might suggest, the Hamiltonians (1, 2) are prompted
by more than just the square lattice. In the next section we will show that several models based
on the honeycomb lattice, which have been considered [16] in relation with the quantum spin
Hall effect, fit the scheme. Moreover, one-dimensional spin pumps [9] also match the description,
with k playing the role of time.
The topological classification applies to insulators that are invariant under odd (or fermionic)
time-reversal symmetry. In the sequel we specify these notions.
Definition 2.4. [Time-reversal symmetry] The symmetry is a map Θ : CN Ñ CN with the

following properties.
i) Θ is antilinear and Θ2 “ ´1;
ii) Θ˚ Θ “ 1;
iii) For all k P S 1 ,

Hp´kq “ ΘHpkqΘ´1 , (4)
where Θ also denotes the map induced on `2 pZ; CN q. Likewise for H 7 .
As a result, N is even. In the models described in the next section, properties (i, ii) arise
from the time-reversal of a spin- 21 particle, and (iii) from the symmetry of the Hamiltonian.
The Bulk Hamiltonian of an insulator is supposed to have a spectral gap at Fermi energy µ
and for all k:
µ R σpHpkqq , pk P S 1 q . (5)
Typically H 7 does not satisfy the gap condition. In fact, while for the essential spectra we have
σess pH 7 pkqq Ă σess pHpkqq, the edge Hamiltonian may have discrete eigenvalues crossing µ for
some values of k.
2.2 The main result in brief

We can informally introduce two indices, I, I 7 P t˘1u, defined in terms of H and H 7 , respectively.
On the circle the involution k ÞÑ ´k has two fixed points, k “ 0, π. The edge index I 7 “ p´1qn is
the parity of the number n of those k P r0, πs at which an eigenvalue of H 7 pkq equals µ, at least
if the eigenvalue crossings are simple. The index I requires more explanation. By definition
Hpkq does not have eigenvalues z R σpHpkqq; yet we may regard the Schrödinger equation
pHpkq ´ zqψ “ 0
4
as a second-order difference equation in n P Z. As such, it has 2N linearly independent solutions,
but we focus attention on those which decay as n Ñ `8. They form a linear space, Ez,k , of
dimension N . By (i, iii) we have
pHp´kq ´ z̄qΘψ “ ΘpHpkq ´ zqψ ,
and thus Ez̄,´k “ ΘEz,k . We choose a reflection symmetric complex contour, γ “ γ̄, encircling
the part of the spectrum of Hpkq lying below µ, and we set T “ γ ˆ S 1 . We so have: (a) an
involution pz, kq ÞÑ pz̄, ´kq on the torus T; (b) a vector bundle with base T and fibers Ez,k ;
which (c) are compatible with Θ in the stated sense. To any vector bundle with these features,
including (i), but irrespective of the concrete definition of its fibers, an index will be associated
in Sect. 4. The bulk index I is that index for the particular bundle arising from H as described.
The main result of this work is that I “ I 7 .
k P S1
Im z
π
γ
0 µ Re z
´π
Figure 1: The torus T is the product of the loop γ in the z-plane with the k-axis (both red); it
is to be glued along the dashed loops. Thick lines (in black) delimit parts of the bulk spectrum
σpHpkqq, as a function of k. Curves (in green) connecting them are discrete eigenvalues εpkq P
σpH 7 pkqq, representing edge states. They intersect the torus only at crossing points (dots) along
the Fermi line tz “ µu ˆ S 1 (in blue), one half of it (thick) being relevant for the index I 7 . The
spectra are symmetric in k ÞÑ ´k.
It should be noted that the result relies on just one translational symmetry.
3 Examples
We show how to obtain Hamiltonians (1, 2) from the Schrödinger operator on the honeycomb
lattice, which models graphene in the single-particle approximation. We will consider two types
of boundary conditions: zigzag and armchair. It is known that the spectrum of the Schrödinger
operator depends on which boundary condition is chosen, [21, 14]. In particular, a zigzag
boundary implies the presence of zero-energy edge states, while they are absent for the armchair
boundary. In Example 3.1 we give a new proof of the last statement.
5
3.1 Graphene
Let Λ “ ΛA Y ΛB be the (infinite) honeycomb lattice and its bipartite decomposition into two
triangular lattices ΛA , ΛB . Upon fixing an origin, they are
ΛA :“ ~n “ n1~a1 ` n2~a2 | pn1 , n2 q P Z2 , ΛB :“ ΛA ` ~δ ,

(
π ~δ “ 1 p~a1 ` ~a2 q .
|~a1 | “ |~a2 | , =p~a1 , ~a2 q “ ,
3 3
Any site in ΛA has three nearest neighbors in ΛB , shifted by ~δ or by another equivalent vector.
See Fig. 2. The model for graphene is simply the Schrödinger operator H0 for a particle hopping
between nearest neighbors (with hopping parameter ´t).
pn1 ` 1, n2 q
B A
pn1 , n2 q
pn1 , n2 ´ 1q
~a2
~a1
Figure 2: The honeycomb lattice. Coordinates of lattice sites are relative to the basis t~a1 , ~a2 u.
The (Bravais) lattice of translations of Λ is ΛA . We reduce Λ to ΛA by dimerizing neighbors

shifted by ~δ. We retain the position ~n of the A-site as that of the dimer, and the values of the
wave function at the two sites as pseudospin components:
ˆ A˙
ψ~n
ψ~n :“ P C2 . (6)
ψ~nB
With these notations, the Schrödinger operator H0 takes the form
ˆ B
ψn1 ,n2 ` ψnB1 ,n2 ´1 ` ψnB1 ´1,n2
˙
` ˘
H0 ψ n1 ,n2 “ ´t . (7)
ψnA1 ,n2 `1 ` ψnA1 `1,n2 ` ψnA1 ,n2
The expression for H0 could of course have ended up differently. Let us name the choices
underlying its construction: a sublattice shift vector (above: ~δ), defining the dimer, and two
1
primitive lattice vectors (~a1 and ~a2 ), defining adjacency between dimers. For instance for ~a1 “
1
~a1 , ~a2 “ ~a1 ` ~a2 , and hence n11 “ n1 ´ n2 , n12 “ n2 , Eq. (7) becomes
ˆ B
ψn1 ,n2 ` ψnB1 `1,n2 ´1 ` ψnB1 ´1,n2
˙
` ˘
H
r0ψ “ ´t ,
n1 ,n2 ψnA1 ´1,n2 `1 ` ψnA1 `1,n2 ` ψnA1 ,n2
after dropping primes. See Fig. 3.

The Bloch decomposition w.r.t. n2 P Z
ż
dk ikn2
ψ~n “ e ψn1 pkq , pψn1 pkq P C2 q (8)
S 1 2π
6
aq B A bq
n2 n2 n1
n1
~a1 ` ~a2
~a2 ~a2 ~a1

~a1
Figure 3: The honeycomb lattice with zigzag aq and armchair bq boundary conditions. The
lattice has been rotated, so as to have n1 fixed along the vertical.
fibers the Hamiltonian, pH0 ψqn1 pkq “ pH0 pkqψpkqqn1 . In fact we obtain from Eq. (7)
` ˘
H0 pkqψ n “ A0 pkqψn´1 ` A0 pkq˚ ψn`1 ` V0 pkqψn ,
where we set ψ “ ψpkq, n “ n1 and

ˆ ˙ ˆ ˙
0 1 0 1 ` eík
A0 pkq “ ´t , V0 pkq “ ´t . (9)
0 0 1 ` eik 0
Likewise for H
r 0 pkq:
` ˘
H
r 0 pkqψ “A r0 pkq˚ ψn`1 ` Vr0 pkqψn ,
r0 pkqψn´1 ` A (10)
n
where now ˆ ˙ ˆ ˙
0 1 0 1
A
r0 pkq “ ´t , V0 pkq “ ´t
r .
eik 0 1 0
It should be noted that only the second Hamiltonian satisfies the condition Apkq P GLpN q in
Eq. (1). The two equivalent bulk Hamiltonians no longer are once they are turned into edge
Hamiltonians by means of the Dirichlet boundary condition. Indeed, they correspond to (a)
zigzag and (b) armchair boundary conditions, respectively. See again Fig. 3.
Proposition 3.1. (i) The Hamiltonian H07 has E “ 0 as an eigenvalue. Actually, H07 pkq has it
r 7 does not have any eigenvalue, i.e. its spectrum
for k P p2π{3, 4π{3q. (ii) The Hamiltonian H0
is purely continuous.
The result is known [10, 21], but perhaps not the argument below for (ii).
Proof. i) Related to A0 pkq being singular, H07 pkqψ “ 0 reduces to the first order equations
B
ψn´1 ` p1 ` eík qψnB “ 0 , A
ψn`1 ` p1 ` eik qψnA “ 0 , pn P Nq .
The boundary condition ψ0 “ 0 implies ψ B “ 0, but is ineffective for ψ A , for which there is a
non-trivial solution in `2 pNq, as long as |1 ` eik | ă 1, i.e. for k in the stated range.
7
ii) Every second dimer of the armchair boundary is exposed (see Fig. 3 right) and the axis
containing them is n1 “ 0. Let us consider, on the whole lattice Λ, the reflection ~n ÞÑ r~n of
lattice sites about the axis n1 “ 0 in Fig. 3, as well as parity, pP ψq~n “ ψr~n . Clearly, rP, H0 s “ 0.
Any odd eigenfunction ψ of H0 satisfies the Dirichlet boundary condition on the line n1 “ 0,
whence its restriction ψr to n1 ą 0 defines one for H r 7 . What matters more is that the converse
0
is true as well: The odd extension ψ of ψr satisfies the Schrödinger equation on n1 “ 0, and
hence at all ~n. The conclusion is by noting that H0 has no eigenvalues; actually it has purely
absolutely continuous spectrum. This is known and by the way has a short proof: It suffices to
establish the property for the fiber Hamiltonian H r 0 pkq. Since the latter has real analytic Bloch
eigenvalues λpqq, absolute continuity holds unless λpqq is constant in q. That, finally, is ruled
out by A r0 pkq P GLpN q in Eq. (10) and the remark below.
Remark 3.2. ([1], Prop. 7) Suppose the Hamiltonian on `2 pZ; CN q
pHψqn “ Aψn´1 ` A˚ ψn`1 ` V ψn
(note: V independent of n) has a Bloch eigenvalue λpqq independent of the quasi-momentum q.

Then A is singular.
Proof. Let λ be that eigenvalue. Then detpAz ´1 ` A˚ z ` V ´ λq “ 0 for z “ eiq , (q P R) and,

by analyticity, for z ‰ 0. Letting z Ñ 0 gives det A “ 0. .
3.2 The Kane-Mele model

In the next example we discuss a model with odd time-reversal symmetry. Wave functions are
now of the form ψ “ pψ~n q~nPZ2 , where ~n labels dimers, see Fig. 2, and ψ~n P C2 b C2 . The spin of
the particle is represented by the Pauli matrices σ1 , σ2 , σ3 acting on the second factor C2 . The
time-reversal operator is Θ “ íp1 b σ2 qC, where C denotes complex conjugation, which is the
usual expression for a spin- 21 particle.
Let H0 be the Schrödinger operator on the honeycomb lattice defined in Eq. (7). We consider
the operator
HKM “ H0 b 1 ` H1 b σ3 ,
where the last term implements a spin-orbit coupling. There H1 describes hopping between
next-to-nearest neighbors; it thus acts diagonally on the pseudospin components of Eq. (6), and
specifically as ˆ ˙
h1 ψ A
H1 ψ “ (11)
´h1 ψ B
with
´ ˘
ph1 ψq~n “ ´t1 ipψn1 ´1,n2 `1 ` ψn1 `1,n2 ` ψn1 ,n2 ´1 q ´ ipψn1 ,n2 `1 ` ψn1 `1,n2 ´1 ` ψn1 ´1,n2 q .
The grouping of terms reflects that, starting from a given A-site, its next-to-nearest neighbors
are reached by turning right or left at a B-site. The turns go along with phases ˘i “ e˘iπ{2
modeling a magnetic flux π{2 through any (positively oriented) triangle ABA; yet the total flux
through the hexagonal cell vanishes, since no phases are associated with the bonds forming its
boundary, see Eq. (7). Likewise for B and A interchanged, resulting in the sign in (11). Indeed,
8
for a given dimer, the next-to-nearest neighbors form again dimers, but the turns linking A,
resp. B-sites are opposite.
The model is the special case of the Kane-Mele model, [16], where two further terms (a
Rashba term and a staggered chemical potential) have been set to zero. Even so, it exhibits a
non-trivial topological phase, i.e. I “ ´1, for small t1 {t ą 0 and µ “ 0. In fact, it is the direct
sum of (a special case of) the Haldane model [11], HH “ H0 ` H1 , and of its time-reversed copy
H0 ´ H1 . In such a situation we have I “ p´1qN , where N is the integer associated with the
quantum Hall effect of HH . In the stated regime, a band gap of order Op|t1 |q opens, and N “ 1,
[11].
We conclude the example by giving the fibers of HKM w.r.t. the Bloch decomposition (8).
It will suffice to do so for HH : Instead of Eq. (9) we have
ˆ 1 ik ˙ ˆ 1 ik ˙
ít pe ´ 1q ´t it pe ´ eík q ´tp1 ` eík q
Apkq “ , V pkq “ .
0 it1 peik ´ 1q ´tp1 ` eik q ít1 peik ´ eík q
4 Time-reversal invariant bundles

The purpose of this section is to define the index of bundles of the kind mentioned in Sect. 2.2,
see Def. 4.8 below, as well as some auxiliary indices; and to formulate some of their properties.
We refer the reader to Sect. 7 for the proofs of the lemmas stated here.
4.1 The index of endpoint degenerate families

We shall define an index for certain families of points on the unit circle. We will give the most
general definition, in order to make evident the stability of the index under homotopy. At the
end of this section we give a procedure to compute it in a more restrictive setting, which is
suitable for our application.
Following ([17], Sect. II.5.2) we consider unordered N -tuples Z “ pz1 , . . . , zN q of complex
numbers zi P C. A distance is defined by
dpZ 1 , Zq “ min max |zn ´ zn1 | ,

n
where the minimum is taken over all possible relabellings of Z or Z 1 . Let us recall ([17],
Thm. II.5.2): Given a family Zpxq, px P ra, bsq, which is continuous w.r.t. d, there exists a
(non-unique) continuous labeling Zpxq “ pz1 pxq , . . . , zN pxqq.
In the following we will consider N -tuples with zi P S 1 . Any continuous labeling zi pxq
induces continuous arguments θi pxq, i.e. zi pxq “ eiθi pxq . Two simple observations are in order:
ˇb
(i) wpZq :“ p2πq´1 N
ř
i“1 θi pxq a is independent of the choice of arguments, as well as of labeling.
ˇ
śN
In fact i“1 zi pxq is independent of the latter. (ii) Any family Zpxq with Zpaq “ Zpbq has a
winding number given as
N pZq :“ wpZq P Z . (12)
We next consider endpoint degenerate families Zp¨q: for x “ a, b each z P Zpxq occurs with
even multiplicity. We may concatenate such a family with one, Z̃pxq, px P rb, csq, such that
Z̃pbq “ Zpbq and Z̃pcq “ Zpaq, while keeping z̃i pxq even degenerate. Clearly, Z̃ is not unique,
but by (12)
wpZ̃1 q ´ wpZ̃2 q P 2 ¨ Z .
9
By the same reason, N pZ#Z̃q is an integer; by N pZ#Z̃q “ wpZq ` wpZ̃q it is determined
mod 2 by Z.
Definition 4.1. [Index] We set

IpZq “ p´1qN pZ#Z̃q , (13)
as the index of endpoint degenerate families.
Consider now the special situation where there is z P S 1 such that z P Zpxq occurs only at
finitely many x, which moreover are simple crossings: z “ zj pxq for a single j, and zj1 pxq ‰ 0.
Then
IpZq “ p´1qn , (14)
where n is the number of crossings of z. Indeed, one can choose Z̃ without crossings of z. Then
N pZ#Z̃q is the number of signed crossings of z by Zp¨q but that qualification is irrelevant for
parity.
4.2 The index of Kramers families of matrices

Let ε be the matrix of even order N given by the block diagonal matrix with blocks
ˆ ˙
0 ´1
,
1 0
C : CN Ñ CN the complex conjugation, and Θ0 “ εC the standard time-reversal on CN .

Suppose T P GLpN q satisfies
Θ0 T “ T ´1 Θ0 . (15)
Then the eigenvalues of T come in pairs λ, λ̄´1 with equal algebraic multiplicity, which is
moreover even if λ “ λ̄´1 . In particular their phases z “ λ{|λ| are even degenerate regardless.
Indeed, Θ0 pT ´ λqn “ T ń p1 ´ λ̄T qn Θ0 , as seen inductively for n “ 0, 1, . . .; for λ “ λ̄´1 the
corresponding pairs of eigenvectors v, Θ0 v remain linearly independent by Θ20 “ ´1.
Definition 4.2. [Kramers property] We call Eq. (15) the Kramers property. We say a family
T pϕq P GLpN q, which is continuous in 0 ď ϕ ď π, has that property if the endpoints T p0q and
T pπq have it.
The repeated eigenvalues λi pϕq of T pϕq form a continuous family in the sense of the previous
section ([17], Thm. II.5.1), and so do the zi “ λi {|λi |. Moreover, Zpϕq “ pz1 pϕq, . . . zN pϕqq is an
endpoint degenerate family.

IpT q “ IpZq (16)
as the index of a Kramers family T . See Eq. (13).
All it in fact takes for the definition is the endpoint degeneracy of Z and not the stronger
Kramers property of T . However we shall not need such an extension.
10
Lemma 4.4. Suppose
T2 pϕq “ M´ pϕqT1 pϕqM` pϕq´1
with continuous M˘ pϕq P GLpN q (0 ď ϕ ď π), as well as
M´ p0qΘ0 “ Θ0 M` p0q , M´ pπqΘ0 “ Θ0 M` pπq . (17)
Then M pϕq :“ M´ pϕqM` pϕq´1 has the Kramers property, and T2 pϕq has it iff T1 pϕq does. If
so,
IpT2 q “ IpT1 qIpM q . (18)
The claims are of immediate verification, except for Eq. (18). However in the special case
that M` pϕq “ M´ pϕq the equality
IpT2 q “ IpT1 q (19)
is also immediate, because T1 pϕq, T2 pϕq then have the same eigenvalues. That case suffices for
the basic result on bulk-edge duality.
4.3 The index of time-reversal invariant bundles

Let S 1 “ R{2πZ be the circle and T “ S 1 ˆ S 1 Q ϕ “ pϕ1 , ϕ2 q the torus with involution
τ : ϕ ÞÑ ´ϕ. It has four fixed points: ϕ0 “ p0, 0q, p0, πq, pπ, 0q, pπ, πq. Let E be a complex
vector bundle with base T Q ϕ and fibers Eϕ of dimension N . We say that E is time-reversal
invariant if there is a map Θ : E Ñ E with Θ2 “ ´1 and Θ : Eϕ ÞÑ Eτ ϕ antilinear.
We also consider the associated frame bundle F pEq with the following operations induced
on frames v “ pv1 , . . . , vN q P F pEqϕ :
• right multiplication by M P GLpN q:
M : F pEqϕ Ñ F pEqϕ , v ÞÑ vM , (20)

řN
with pvM qj “ i“1 vi Mij . Any two frames are so related by a unique M .
• Θ : F pEqϕ Ñ F pEqτ ϕ , v ÞÑ Θv, with pΘvqi “ Θvi . Note that
ΘpvM q “ pΘvqM . (21)
In order to classify time-reversal invariant bundles we consider the torus cut along the circle
tϕ1 | ϕ1 “ π – ´πu ˆ S 1 , or more precisely T9 “ r´π, πs ˆ S 1 . Note that only ϕ “ p0, 0q, p0, πq
remain among the fixed points of τ , and that any bundle on T naturally defines one on T. 9
Lemma 4.5. [Existence of time-reversal invariant sections] On the cut torus T, 9 there are
9
(smooth) sections v : T Ñ F pEq of the frame bundle (whence vpϕq P F pEqϕ ) satisfying
vpτ ϕq “ Θvpϕqε . (22)
As a result, they are compatible with right multiplication by M pϕq P GLpN q iff
Θ0 M pτ ϕq “ M pϕqΘ0 . (23)
11
We remark that the condition (22) was shown [9] to be obstructed on the (uncut) torus by
the Z2 -invariant.
Let v˘ pϕ2 q :“ vp˘π, ϕ2 q be the boundary values of vpϕq along the two sides of the cut, and
T pϕ2 q P GLpN q the transition matrix,
v` pϕ2 q “ v´ pϕ2 qT pϕ2 q , pϕ2 P S 1 q . (24)
Lemma 4.6. [Time-reversal symmetry of the transition functions]
Θ´1
0 T p´ϕ2 qΘ0 T pϕ2 q “ 1 .
In particular, T pϕ2 q has the Kramers property on the interval 0 ď ϕ2 ď π, which hints at a
the possibility of assigning an index to the bundle.
Lemma 4.7. [Independence of the index from the section] Let v piq , (i “ 1, 2) be time-reversal
9 and let Ti pϕ2 q the corresponding transitions matrices
invariant sections on the cut torus T,
across the cut. Then
IpT1 q “ IpT2 q .
We may thus proceed to the following definition.
IpEq “ IpT q (25)
as the index of a time-reversal invariant vector bundle E over T. See Eq. (16).
Such bundles can hence be distinguished in two topologically distinct classes, according to
the value of the index.
We also retain the following remark, which appears as a byproduct of Lemma 4.5.
Remark 4.9. Let Θ : CN Ñ CN satisfy the time-reversal symmetry condition (i) of Def. 2.4.
Then there is a basis v “ pv1 , . . . , vN q of CN such that v “ Θvε. If condition (ii) applies too, the
basis can be chosen orthonormal. Letting U map v to the standard basis of CN , a restatement
is U Θ “ ´Θ0 U with U unitary.
5 The bulk-edge correspondence

It pays to look at first at wave-functions ψ “ pψn qnPZ as just sequences, i.e. ψ P C :“ Z ˆ CN . It
can be shown that for any z P ρpHpkqq in the resolvent set, the Schrödinger equation Hpkqψ “ zψ
has N linearly independent solutions which are square-integrable at n Ñ `8. Let Ez,k Ă C the
linear space they form. As explained in Sect. 2.2 we obtain a vector bundle
E “ tppz, kq, ψq P T ˆ C | ψ P Ez,k u (26)
over the torus T “ γ ˆ S 1 , which enjoys the property of being time-reversal invariant. The
following definition is thus natural.
Definition 5.1. We set
I “ IpEq (27)
as the bulk index. See Eq. (25).
12
The edge index has been loosely introduced in Sect. 2.2. In preparation for precise definition
let us take a closed interval I containing the Fermi energy µ in its interior and such that
σpHpkqq X I “ ∅ for k P S 1 . The spectrum of H 7 pkq is discrete and finite in I (uniformly in k),
and the eigenvalues are locally given by branches εi pkq which are C 1 in k ([17], Thm. II.6.8).
By possibly adjusting µ we can arrange that if εi pk˚ q “ µ for some k˚ then ε1i pk˚ q ‰ 0. We note
that the number of crossings at k˚ “ 0 or k˚ “ π is even by (4).
Definition 5.2. Let n be the number of crossings k˚ P r0, πs, with those at endpoints counted
half. Set
I 7 “ p´1qn
as the edge index.
The next remark about the edge index is inessential for the following main result.
Remark 5.3. The index I 7 could be also defined in terms of a Kramers family. To this end
let f pεq be a continuous real function with f pεq “ 0, (ε ă I) and f pεq “ 1, (ε ą I), where
I is the aforementioned interval. Then, in view of ΘgpHqΘ´1 “ ḡpΘHΘ´1 q, the operator
T pkq “ expp2πif pH 7 pkqqq has the Kramers property, albeit w.r.t. Θ, which is however irrelevant.
The definition I 7 “ IpT q would be legitimate, since that of wpZq in Sect. 4.1 extends to countable
families Z of points zi pxq P S 1 , as long as only finitely many (uniformly in x) are ‰ 1. It would
agree with Def. 5.2 by Eq. (14).
Theorem 5.4. [Bulk-edge correspondence for topological insulators] Let the Bulk and the Edge
Hamiltonian be as in Defs. 2.1 and 2.2. Assume the time-reversal symmetry conditions of
Def. 2.4 and the gap condition (5). Then
I “ I7 .
We can give the main steps of the proof right away. For given pz, kq P ρpHpkqq ˆ S 1 we
7
consider, besides of Ez,k , also the linear space Ez,k of solutions ψ 7 “ pψn7 qnPZ decaying at n Ñ 8
of the Schrödinger equation H 7 pkqψ 7 “ zψ 7 , without imposing boundary conditions at n “ 0.
Eq. (3) establishes a bijection
7
Ez,k Ñ Ez,k , ψ ÞÑ ψ 7 (28)
determined by ψn7 “ ψn for n ą n0 .

Frames Ψ P F pEqz,k consists of N -tuples Ψ “ pψ1 , . . . , ψN q of linearly independent solutions
ψi P Ez,k . Note that the index i does not denote the lattice site n P Z. Since ψin P CN , we
may equivalently say: Ψ “ pΨn qnPZ with Ψn P MN pCq belongs to F pEqz,k iff Ψ is a solution of
HpkqΨ “ zΨ decaying at n Ñ `8 , which is moreover fundamental in the sense that for any n
Ψn a “ 0 , Ψn`1 a “ 0 ñ a “ 0 , pa P CN q. (29)
The bijection (28) induces one between frame bundles, F pEqz,k Ñ F pE 7 qz,k , Ψ Ñ
Þ Ψ7 . It is
manifestly compatible with the right action (23) of GLpN q. The next lemma rests on the
bijection.
Lemma 5.5. i) A point pz˚ , k˚ q P T has det Ψ70 “ 0 for some (and hence all) Ψ P F pEqz˚ ,k˚
iff z˚ P σpH 7 pk˚ qq. If so, then z˚ “ µ. For a dense set of Hamiltonians H 7 near the given
13
one, the points k˚ are isolated in S 1 and for each of them there is a simple eigenvalue
branch εpkq with εpk˚ q “ µ, ε1 pk˚ q ‰ 0; moreover,
det Ψ71 ‰ 0 . (30)
Density is meant with respect to the topology of the class of Hamiltonians specified at the
beginning of Sect. 2.1.
ii) Let Ψpz, kq P F pEqz,k be a section defined in a neighborhood in C ˆ S 1 Ą T of any of the

crossing points pz˚ “ µ, k˚ q. The family of matrices
˚
Lpz, kq “ ´Ψ71 pz̄, kqApkqΨ70 pz, kq
has the reflection property Lpz, kq “ Lpz̄, kq˚ . Its eigenvalues are thus real for real z.
There generically is a single eigenvalue branch lpz, kq of Lpz, kq vanishing to first order at
pµ, k˚ q: There the derivatives Bl{Bz and Bl{Bk are real and non-zero.
iii) At any of the points pµ, k˚ q we have

Bl
ă0. (31)
Bz
iv) As k increases past k˚ the eigenvalue εpkq crosses µ as an increasing function if

Bl ˇˇ
ˇ ą0,
Bk pz“µ,k“k˚ q
and as a decreasing one in the opposite case.
The lemma, which is proven in Sect. 7, allows to complete the proof of the main result.
As a matter of fact, only items (i, ii) matter to that end. For later use we mention that no
use of time-reversal symmetry has been made in the lemma, in the sense that the statement
about density of the Hamiltonians in (i) holds in either class, with and without that specification.
Proof of Thm. 5.4. As a preliminary we recall Remark 4.9: At the price of conjugating the
Hamiltonians by U in the internal space CN , which preserves the assumptions, we may assume
Θ “ ´Θ0 . (32)
By density it will suffice to prove the theorem for Hamiltonians as specified in part (i) of the
lemma. By the first sentence there, we can define a section pz, kq ÞÑ F pEqz,k away from crossing
points by requiring
Ψ70 pz, kq “ 1 . (33)
Clearly, the requirement can not be imposed at such points. In a small reflection symmetric
disk D “ D̄ Ă T containing the generic crossing point pµ, k˚ q, Eq. (30) allows us to make an
alternate choice Ψpz,
p kq by requiring
p 7 pz, kq “ 1 .
Ψ (34)
1
14
pBDq`
pz, kq
J pµ, k˚ q k
D
pBDq´
Figure 4: The neighborhood in T (see Fig. 1) of a crossing point pµ, k˚ q. Two different local
sections are defined inside (Ψ)
p and outside (Ψ) of a small disk D containing pµ, k˚ q. They are
glued across the two half-circles pBDq˘ by means of matrices M˘ pkq.
9 where the cut is

Based on the latter section we shall extend the former to the cut torus T,
1
the Fermi line tz “ µu ˆ S . On the boundary BD both sections are defined, and we have
Ψpz,
p kq “ Ψpz, kqM˘ pkq , (35)
for pz, kq P pBDq˘ “ BD X t˘ Im z ą 0u and some matrices M˘ pkq P GLpN q, see Eq. (20),
which are parametrized by k in the interval J resulting from the intersection of the Fermi line
with D. The same matrices relate Ψ p 7 and Ψ7 . We may then obtain the announced extension
by (re)defining Ψpz, kq through Eq. (35) for pz, kq P D˘ “ D X t˘ Im z ą 0u. We observe that
the section on T9 so constructed satisfies Eq. (22); in fact its ingredients (33, 34) do by (32) and
p´Θ0 1qε “ 1.
The transition matrix T pkq across the cut differs from the identity only within the intervals
J. Along such an interval the boundary values of the section are Ψ˘ pkq “ Ψpµ, p kqM˘ pkq´1 ,
resulting in the transition matrix
T pkq “ M´ pkqM` pkq´1 , pk P Iq . (36)
We can compute the bulk index IpEq “ IpT q using Eq. (14). The theorem then reduces to
the claim that, for each crossing point, the eigenvalues of T pkq change with k P J from 1 to 1
without winding, except for a single one with winding number ˘1.
It is with the proof of this claim that our choice of sections comes to fruition. In fact, by
Eq. (33) and the sentence after Eq. (35) we have
p 7 pz, kq “ M˘ pkq ,
Ψ ppz, kq P pBDq˘ q .
0
p 7 , which is as required there. Then Lpz, kq “

We apply part (ii) of the lemma to the section Ψ
7
ÁpkqΨ0 pz, kq by Eq. (34) and hence
p
T pkq “ Apkq´1 L´ pkqL` pkq´1 Apkq , pk P Iq
with L˘ pkq “ Lpz, kq for pz, kq P pBDq˘ . Clearly, Apkq can be dropped, as it does not affect the
eigenvalues under investigation. We have
Lpz, kq “ lpz, kqΠpz, kq ‘ Lpz,

r kq ,
where Πpz, kq is the rank 1 projection on the eigenspace corresponding to the eigenvalue lpz, kq,
and Lpz,
r kq is regular as a map on the range of the complementary projection. Clearly, lpz, kq “
15
Op∆q as ∆ :“ pz ´ µ, k ´ k˚ q Ñ 0, but by (ii) above we also have lpz, kq´1 “ Op∆´1 q on T. We
use the notation f0 and f1 for the value of a function f and its gradient pBz f, Bk f q at the critical
point. Then
Lpz, kq “ lpz, kqΠ0 ` L r 1 ¨ ∆ ` Op∆2 q ,

r0 ` L
Lpz, kq´1 “ lpz, kq´1 Πpz, kq ‘ Lpz, r ´1 ` Op∆q .
r kq´1 “ lpz, kq´1 pΠ0 ` Π1 ¨ ∆q ` L
0
In view of L´ pkq “ Lpz̄, kq for pz, kq P pBDq` we compute
lpz̄, kq
Lpz̄, kqLpz, kq´1 “ Π0 ` R ` p1 ´ Π0 q ` Op∆q , (37)
lpz, kq
R “ lpz, kq´1 pL
r 0 Π1 ¨ ∆ ` L ¯ 0 q “ lpz, kq´1 L
r 1 ¨ ∆Π ¯ “ Op1q .
r 0 Π1 ¨ p∆ ´ ∆q
In the second line we used L

r 0 Π1 ` L
r 1 Π0 “ 0, as seen from expanding LΠ
r “ 0. Using p1´Π0 qΠ1 “
2
Π1 Π0 from Π “ Π , we find that R “ p1 ´ Π0 qRΠ0 is strictly block triangular. For small D and
pz, kq P pBDq` the eigenvalues of (37) wind as if the error Op∆q is omitted, provided those of the
explicit part do not vanish. Those in turn remain the same when R is omitted, and are in fact
equal to 1 with multiplicity N ´ 1, and to l´ pkq{l` pkq with multiplicity 1, where l˘ pkq “ lpz, kq,
(pz, kq P pBDq˘ ). The winding number of l´ pkq{l` pkq along J equals that of lpz, kq along BD.
Since the vanishing at the crossing point is of first order, lpz, kq “ l1 ¨ ∆ ` Op∆2 q, that number
is
sgnppBz lqpBk lqq “ ˘1 ,
as claimed. Though the sign is irrelevant (so far), we observe that by (iv, v) it is `1 if the
eigenvalue crossing is decreasing in k.
5.1 The bulk index as an index of Bloch bundles

A feature of definition of the Bulk index is that it applies to Hamiltonians which are periodic
just along the edge, see Eqs. (1, 27). If the Hamiltonian is periodic in both directions and hence
the Brillouin zone two-dimensional, the Bulk index allows for an alternate formulation in terms
of the bundle of Bloch solutions, as we are about to explain.
We temporarily suppress the longitudinal quasi-momentum k in Eq. (1) and consider Hamil-
tonians of `2 pZ, CN q of the form
pHψqn “ Aψn´1 ` A˚ ψn`1 ` Vn ψn , pn P Z, ψn P CN q (38)
where A P GLpN q and Vn is periodic in n, i.e. Vn`M “ Vn . By once again considering sites n
within a period as labels of internal degrees of freedom, we may assume that the period is 1.
This amounts to the replacement of ψn , A, Vn by
¨ ˚
˛
¨ ˛ V 0 A 0 ¨ ¨ ¨ 0
¨ ˛ 0 ¨¨¨ 0 A ˚A ... ... ...
˚ .. ‹
ψ0 ˚0 ¨ ¨ ¨ . ‹
˚ .. ‹ MN
0 ‹ ˚
.
‹
Ψ “ ˝ . ‚P C , A “ ˚.
˚
.. ‹ ,
‹
V “˚0
˚ . . 0 ‹
‹
.
˝ .. . ˚ ‹
.
‚
ψM ´1 . . .
˝ .. .. .. ..
˚ ‹
0 ¨¨¨ 0 A˚ ‚
0 ¨¨¨ 0 A VM ´1
(39)
16
In particular A is singular, unlike A. A Bloch solution pψn qnPZ of quasi-periodicity ξ ‰ 0,
ψn`pM “ ξ p ψn , (40)
and of energy z is then represented as a solution Ψ of
HpξqΨ ” pAξ ´1 ` A˚ ξ ` VqΨ “ zΨ , pΨ P CM N q . (41)
For κ P S 1 “ R{2πZ and ξ “ eiκ , the matrix Hpξq is hermitian, since now ξ¯ “ ξ ´1 . It thus has
eigenvalues z “ λl pκq (real and increasingly ordered) and eigenvectors Ψl pκq, pl “ 1, . . . , M N q.
The ranges of the energy curves λj pκq are known as energy bands. Let
∆l pkq “ inf λl`1 pκq ´ sup λl pκq (42)

κ κ
be the gap between successive bands of Hpkq, where we temporarily reinstated the dependence
on k P S 1 , implicit in (38). The gap is open if ∆l pkq ą 0. Let then ∆l “ inf k ∆l pkq. In
topological insulators the bands are degenerate for k “ 0, π, in fact at κ “ 0, π. A band thus
can not be separated from the rest of the spectrum, but a pair of them can. We will assume so
for the pairs p2j ´ 1, 2jq, pj “ 1, . . . , N0 {2q, i.e.
∆2 , ∆4 , . . . ∆N0 ą 0 , (43)
where N0 (even) is the uppermost band below the Fermi energy µ. For that band we actually
retain the stronger assumption (5) of a spectral gap. (It amounts to a positive gap in Eq. (42)
for l “ N0 even when extremizing jointly in κ, k.)
Definition 5.6. The Bloch bundle Ej of the j-th pair of bands has the Brillouin zone B “
S 1 ˆ S 1 Q pκ, kq as base and the span rΨ2j´1 pκ, kq, Ψ2j pκ, kqs Ă CM N as fibers.
It should be noted that while the eigenvectors are not smooth in κ, k, their span is ([17],
Sect. II.1.4), since degeneracies occur within the pair.
The result of this section is that the Bulk index can be expressed by means of the indices of
the Bloch bundles of the filled pairs of bands.
Theorem 5.7. Under the above assumption (43) and the gap condition (5) we have
N
ź 0 {2
I“ IpEj q , (44)
j“1
where I “ IpEq is the bulk index (27) and IpEj q is defined in Eq. (25) for the torus B.
Note the bundles seen on the two sides of Eq. (44) have different base spaces and fibers; in
fact the dimensions of the latter are N and 2, respectively.
The proof we will give makes the simplifying assumption that the energy curves λl pκq do not
have more critical points than required by the time-reversal symmetry of the Hamiltonian.
At first, we reduce the theorem to a lemma. The base space of the bundle E in Eq. (26)
can be extended from T to all of tpz, kq | z P ρpHpkqqu. In this notation, I “ IpE æ Tq. The
torus T “ γ ˆ S 1 may then be deformed and split into N0 {2 tori Tj , pj “ 1, . . . , N0 {2q each
surrounding a pair of bands p2j ´ 1, 2jq. Unlike T, the Tj do not need to be a Cartesian product
17
form, since (43) does not imply a spectral gap uniformly in k. We then consider the bundles
E pjq :“ E æ Tj . By homotopy and by the multiplicative property under splitting (Lemma 7.5),
we have:
`Nď0 {2
pjq
˘ Nź
0 {2
IpE æ Tq “ I E “ IpE pjq q ,

j“1 j“1
and Eq. (44) reduces to the following.

Lemma 5.8.
IpE pjq q “ IpEj q . (45)
The proof of this main lemma is deferred to Sect. 8.
6 Quantum Hall systems

Much of what has been said in the previous sections has a counterpart for Hall systems. In that
case the results are not completely new; nevertheless they generalize the bulk-edge correspon-
dence of [13].
The setting is the same as given by the Hamiltonians (1, 2) with gap condition (5), but
without postulating a time-reversal symmetry. For the sake of brevity, definitions and statements
given in the sequel rely on notations and contexts used for topological insulators.
Definition 6.1. [Index] Let T pϕq P GLpN q, pϕ P S 1 q be a continuous family. We set
N pT q “ N pZq , (46)
where T pϕq determines Zpϕq as in Eq. (16) and N pZq is defined in Eq. (12). Alternatively,
N pT q is the winding number of det T pϕq.
Let E be the vector bundle with base T “ S 1 ˆ S 1 Q pϕ1 , ϕ2 q “ ϕ and fibers of dimension
N . On the frame bundle F pEq, the right multiplication by GLpN q is defined as in (20).
The classification of such bundles may again proceed by considering the cut torus T9 “
r´π, πs ˆ S 1 . In fact on T9 there are smooth sections v : T9 Ñ F pEq; this is in analogy to
Lemma 4.5, but with simpler proof, as observed in Remark 7.1.
N pEq “ N pT q , (47)
where T pϕ2 q is the transition matrix introduced in Eq. (24). See Eq. (46). The definition is
again independent of the choice of the section v, as seen from the simplification of Lemma 4.7.
It should be noted that the sign of the index would flip upon interchanging ˘ in Eq. (24).
The index N pEq is just the Chern number of E, but that will not be needed.
The definitions of bulk and edge indices parallel Defs. 5.1 and 5.2.
Definition 6.3. The bulk index is
N “ N pEq , (48)
where E is the bundle (26). See Eq. (47). The edge index N 7 is the number of signed crossings
k˚ P S 1 of the Fermi energy µ by eigenvalues of H 7 pkq. They are counted positively for decreasing
eigenvalue branches.
18
Theorem 6.4. Let bulk and edge index be defined as above. Then
N “ N7 .
Proof. The proof is contained in that of Thm. 5.4. Now its last sentence matters.
6.1 The bulk index as an index of Bloch bundles

We consider the case of doubly periodic Hamiltonians in close analogy to Sect. 5.1. As in
Thm. 5.7 we will express the bulk index in terms of the Bloch bundles of the filled bands.
In the context of quantum Hall systems it is legitimate to assume that, for fixed longitudinal
momentum k, the bands do not overlap:
∆1 , ∆2 , . . . ∆N 0 ą 0 , (49)
where N0 is again the uppermost band below the Fermi surface, cf. (44).
Definition 6.5. The Bloch bundle E` of the `-th band has the Brillouin zone B “ S 1 ˆS 1 Q pκ, kq
as base and the lines rΨ` pκ, kqs Ă CM N as fibers.
Theorem 6.6. Under the above assumption (49) and the gap condition (5) we have
N0
ÿ
N “ N pE` q ,
`“1
where N “ N pEq is the bulk index (48) and N pE` q is defined in (47) for the torus B.
Proof. Consider the simplifying assumption that the energy curves do not have more critical
points than necessary in absence of time-reversal symmetry, as explained after Eq. (50) below.
The proof then parallels that of Thm. 5.7, but is much simpler. In fact the complex loop is of
the type seen in the first case of Fig. 7, but run through just once.
6.2 Bulk-edge correspondence through scattering theory

We propose a further perspective on the bulk-edge correspondence in the doubly periodic case. It
does not rely on decaying bulk solutions, as Thm. 6.4 did. In contrast to that result, where edge
states are intercepted at Fermi energy, here they are right at inception, i.e. as they are born at
band edges. That will be done by means of a result from scattering theory known as Levinson’s
theorem. In its usual form ([26], Thm. XI.59) it computes the phase of the scattering matrix at
thresholds. The version below computes the phase difference when a parameter is changed.
We focus on a single band ` which is separated from its neighbors,
∆`´1 , ∆` ą 0 , (50)
and on its Bloch bundle E` , cf. Def. 6.5. We also assume that, for fixed k P S 1 , the energy curve
λ “ λ` pκ, kq has as a function of κ just two critical points, both non-degenerate, namely a max-
imum κ` pkq and a minimum κ´ pkq (That assumption would not be consistent with topological
insulators, cf. Fig. 7.) The curves κ˘ pkq cut the Brillouin zone B into two open domains B`
(resp. B´ ) where λpκ, kq is increasing (resp. decreasing) in κ w.r.t. the orientation of S 1 .
19
1111
0000
k
0000
11110000
1111
0000
1111
0000
1111
0000
11110000
1111
0000
11110000
1111
0000
1111
0000
1111
00001111
11110000
00001111
11110000
0000
1111
κ´ κ` κ
Figure 5: The Brillouin zone B with the domain B

r ´ (shaded) and the curves κ˘ pkq.
Lemma 6.7. i) There is a map r : B Ñ B defined by λpκ, kq “ λprpκ, kq, kq and the property
that is has the two extrema κ “ κ˘ pkq as its only fixed points. It interchanges B˘ and is
real analytic in κ.
ii) There is a domain Br ´ Ą B´ and on there a section Ψ´ pκ, kq ‰ 0 of the Bloch bundle E`
which is smooth in k and analytic in κ.
We will occasionally omit k from the notation in the sequel. In order to ensure that edge
states are indeed “born at band edges”, we make the assumption that there are none embedded
in the band,
σpp pH 7 q X rλpκ´ q, λpκ` qs “ H . (51)
The assumption is generically satisfied, but counterexamples can be constructed by taking the
direct sum of two Hamiltonians, such that the pure point edge spectrum of one overlaps the
band of the other.
Lemma 6.8. At energies λ “ λpκq, pκ P B´ q the edge Hamiltonian H 7 has a bounded eigenso-
lution ψ 7 “ pψn7 pκqqnPN which is unique up to multiples. It satisfies
ψn7 pκq “ ψn´ pκq ` ψn` prpκqq ` op1q , pn Ñ `8q (52)
where the Bloch solution ψ ` prpκqq ‰ 0 is uniquely determined by ψ ´ pκq. (We recall the relation
(39) between Ψ and ψ.) Hence Ψ` is a section of E` on B` “ rpB´ q.
The bounded solution ψ 7 ought to be interpreted as a scattering solution for the reflection
at the boundary n “ 0. In fact, since λ1 pκq ă 0 on B´ , ψ ´ represents an incoming wave, and
ψ ` an outgoing one.
Since B` and Br ´ overlap near κ˘ we may introduce scattering amplitudes S˘ pκ, kq for κ
near κ˘ pkq and κ ą κ` pkq, resp. κ ă κ´ pkq:
Ψ` prpκqq “ S˘ pκqΨ´ prpκqq . (53)
The scattering amplitudes S˘ pκ˘ ˘ δq, which by the lemma do not vanish, play the role of
transition matrices (24) for the line bundle E` . Therefore, by (47),
N pE` q “ N pS` q ´ N pS´ q ,
where N pS˘ q is defined in (46).
20
Definition 6.9. We say that H 7 has a semi-bound state at the upper band edge λpκ` q if it
admits a bounded solution ψ 7 “ pψn7 qnPN of H 7 ψ 7 “ λpκ` qψ 7 .
Lemma 6.10. Suppose that a branch εpkq of discrete eigenvalues of H 7 pkq touches the `-th band
from above at k˚ , i.e.
εpkq ´ λpκ` pkq, kq Ñ 0 , pk Ñ k˚ q . (54)
Then H 7 pk˚ q has a semi-bound state.
Theorem 6.11. [Relative Levinson Theorem] Let k “ ki P S 1 , pi “ 1, 2q not correspond to
semi-bound states of H 7 pkq. Then
lim arg S` pκ` pkq ` δ, kq |kk21 “ 2πN` , (55)
δÑ0
where arg denotes a continuous argument and N` is the signed number of discrete eigenvalue
branches of H 7 pkq emerging p´q or disappearing p`q at the upper band edge, as k runs from k1
to k2 in the orientation of S 1 . Likewise for the lower band edge and S´ , except for a reversed
count of signs in N´ .
In particular the theorem may be applied to k1 “ k2 , i.e. to a full circle S 1 . Then it states
N pS˘ q “ N˘ , pδ ą 0q. If ` “ N0 is the uppermost band below the Fermi energy, then N` “ N 7 .
If the same assumptions hold true for all bands below it, then the bulk-edge correspondence
N0
ÿ
N pE` q “ N 7
`“1
p`q p`´1q
is recovered in view of N´ “ N` .
The proofs of the results of this section are found in Sect. 9.
6.3 Comparison between two approaches

In this section we compare the independent approaches to bulk-edge correspondence underlying
Thms. 6.4 and 6.11.
Let H 7 pkq have a semi-bound state at the isolated point k˚ . Suppose that a branch εpkq of
discrete eigenvalues disappears there into the N0 -th band from above, cf. Eq. (54) with k Ò k˚ ,
whereas none emerges; see Fig. 6. By a suitable, k-dependent energy shift we may assume that
the upper band edge λpκ` pkqq ” λ` is constant, and pick the Fermi energy µ ą λ` arbitrarily
close to it. The branch εpkq is then decreasing and crosses the (dashed) Fermi line. For k1 , k2
near k˚ with k1 ă k˚ ă k2 we have
lim arg S` pκ` pkq ` δ, kq|kk21 “ 2π ,
δÑ0
arg det T pkq|kk21 “ 2π
for µ close enough to λ` , where T is the transition matrix (36) across the Fermi line, as used in
the proof of Thms. 5.4 and 6.4. In fact, the first equation is the specialization of Eq. (55) and
the second one follows from the claim made after Eq. (36). In case of an emerging branch the
r.h.s. of both equations would change sign.
Both expressions on the l.h.s. arose as winding numbers. The point we wish to make here is
that their equality can also be seen by homotopy, rather than by separate evaluation. We thus
emphasize:
21
1111111
0000000
0000000
1111111
εpkq
εpkq
0000000
1111111 1111111
0000000 µ
0000000
1111111 0000000
1111111 λ`
k˚ k k1 k˚ k2 k
Figure 6: Left: A discrete edge eigenvalue εpkq disappearing into the uppermost band below
the Fermi energy µ. Right: Same, after energy shift.
Proposition 6.12. Under the above assumptions we have
lim arg S` pκ` pkq ` δ, kq|kk21 “ arg det T pkq|kk21 . (56)

δÑ0
The proof is given in Sect. 9.

We close this section with a remark. The duality of Sect. 6.2 relied on the assumption that,
for fixed k, the energy curve λl pκ, kq had a single local maximum and minimum in κ. That
naturally assigns a unique outgoing wave to a given incoming one, with transversal momenta
κ ÞÑ rpκq. Should the assumption fail, multiple reflected waves could arise, preventing that
essential assignment; at least without further ado, like dealing with scattering matrices instead
of amplitudes. That failure is in fact unavoidable for time-reversal symmetric Hamiltonians (see
Fig. 7 left). However, Prop. 6.12 still indicates that the index of the Bloch bundle Ej (associated
to the pair of bands l “ 2j´1, 2j) can be determined by the glitches of arg S` pκ, kq along the lines
κ “ κ˘ pkq of global maxima (or minima). In fact points where edge state emerge or disappear
generically occur only on those lines, see Eq. (51), whence at nearby energies scattering remains
described by a single amplitude S` .
7 Basic bulk-edge correspondence: Proofs

The present section contains the proofs of our basic results for topological insulators (Thm. 5.4)
and for integer quantum Hall systems (Thm. 6.4). We begin by supplying those relating to
Sect. 4 (except for Lemma 4.4) and concerning the index of an abstract time-reversal invariant
bundle.
Proof of Lemma 4.5. The main claim (22) states that there are linearly independent vectors
v1 pϕq, . . . , vN pϕq P Eϕ such that
ˆ ˙
` ˘ ` ˘ 0 ´1
vi´1 pτ ϕq, vi pτ ϕq “ Θvi´1 pϕq, Θvi pϕq ,
1 0
pi “ 2, 4, . . . , N q, i.e.
vi´1 pτ ϕq “ Θvi pϕq , vi pτ ϕq “ ´Θvi´1 pϕq . (57)

9 We will prove by induction in n “ 0, 2, . . . , N
We first consider either fixed point ϕ0 “ τ ϕ0 P T.
that there are linearly independent vectors vk “ vk pϕ0 q P Eϕ0 , pk “ 1, 2, . . . , n ´ 1, nq such
22
that (57) holds true for i “ 2, . . . , n. Indeed, pick vn`1 such that v1 , . . . , vn , vn`1 are linearly
independent and set vn`2 “ ´Θvn`1 . Then Eqs. (57) also hold for i “ n ` 2 by Θ2 “ ´1.
řn`2
vn`2 are still linearly independent: k“1 λk vk “ 0 implies, by v “ Θvε, the
Moreover v1 , . . . ,ř
same with λ̃k “ n`2 `“1 εk` λ̄` ; or for short for λ̃ “ Θ0 λ instead of λ “ pλ1 , . . . , λn`2 q. Since the
rank of those vectors is at least n ` 1 we have
aλ ` ãλ̃ “ 0 ,
with pa, ãq P C2 , pa, ãq ‰ 0. Applying Θ0 yields á̃λ¯ ` āλ̃ “ 0 and together λ “ λ̃ “ 0 by
2 2
|a| ` |ã| ‰ 0.
The frames vpϕ0 q at the two fixed points may be interpolated over the line tϕ1 “ 0u ˆ r0, πs.
In fact, there the bundle E is trivial ([22], Cor. 9.5), since r0, πs is contractible to a point; and
GLpN q is connected. Eq. (22) is then used as a definition to extend vpϕq to a section over the
circle tϕ1 “ 0u ˆ S 1 .
The section vpϕq may then be extended further to the half torus T9 1{2 :“ r0, πs ˆ S 1 by
homotopy. (We still denote by E the bundle restricted to it.) More precisely, two maps T9 1{2 Ñ
T9 1{2 are clearly homotopic, namely the identity map and f : pϕ1 , ϕ2 q ÞÑ p0, ϕ2 q. Hence F pEq
and f ˚ F pEq are equivalent bundles ([22], Thm. 9.4). Since the latter has fibers pf ˚ F pEqqϕ “
F pEqf pϕq , the above section over the circle extends trivially to T9 1{2 ; by equivalence the same
holds true for F pEq. Finally, Eq. (22) extends the section v to all of T. 9 An inspection of the
procedure shows that the necessary smoothness of the section can be ensured. Eq. (23) then
follows in the equivalent form εM pτ ϕq “ M pϕqε using (21).
Remark 7.1. In absence of time-reversal symmetry, see Sect. 6, the existence of a global section
on T9 is quite obvious. In view of the above we may just note that it exists on the circle tϕ1 “
0u ˆ S 1 ” S 1 . In fact we may pick any frame vp0q “ vp2πq and interpolate along r0, 2πs.
Proof of Lemma 4.6. Note that Eq. (22) states v˘ pϕ2 q “ Θv¯ p´ϕ2 qε for the boundary values.
We thus have
v` pϕ2 q “ v´ pϕ2 qT pϕ2 q “ pΘv` p´ϕ2 qqεT pϕ2 q

“ pΘv´ p´ϕ2 qqT p´ϕ2 qεT pϕ2 q “ pΘ2 v` pϕ2 qqεT p´ϕ2 qεT pϕ2 q .
Since Θ2 “ ´1 and the right action is transitive, the claim follows.
Proof of Lemma 4.7. The two sections obey the relation v p2q pϕq “ v p1q pϕqM pϕq for some matrix
M pϕq P GLpN q continuous on T. 9 Let M˘ pϕ2 q :“ M p˘π, ϕ2 q be its boundary values. The two
transition matrices are then related by
T1 pϕ2 q “ M´ pϕ2 qT2 pϕ2 qM` pϕ2 q´1 . (58)
More generally, we consider
T pϕ1 , ϕ2 q :“ M p´ϕ1 , ϕ2 qT2 pϕ2 qM pϕ1 , ϕ2 q´1
as a homotopy in 0 ď ϕ1 ď π of families in ϕ2 . Eq. (23) implies that, at fixed ϕ1 , the matrices

M p˘ϕ1 , ¨q satisfy (17). Hence the family T pϕ1 , ¨q has the Kramers property. Clearly the index
of T pϕ1 , ¨q is constant in ϕ1 by continuity. Note that one end of the homotopy is
T p0, ϕ2 q “ M p0, ϕ2 qT2 pϕ2 qM p0, ϕ2 q´1 ,
23
which by (19) has the same index as T2 , while the other is (58). We conclude IpT2 q “ IpT1 q.
We next come to the the proof of the main technical lemma in relation with the bulk-edge
correspondence.
Proof of Lemma 5.5 i). The proof of the lemma follows quite closely that of ([3], Lemma 3).
Consider the finite difference equation pH 7 ´ zqψ 7 “ 0 with z P C and without imposing the
boundary condition ψ07 “ 0. Solutions ψ 7 are square-summable at n Ñ `8 iff ψ 7 “ Ψ7 a for
some a P CN . Hence 0 is eigenvalue of Ψ70 iff z is an eigenvalue of the operator H 7 , which now
includes the boundary condition. In particular z is then real, because H 7 is self-adjoint. For γ
as in Fig. 1 we have γ X σpH 7 q Ă tµu.
We may arrange for the absence of “flat” crossings, εpk˚ q “ µ, ε1 pk˚ q “ 0, by adding to
Vn pkq and Vn7 pkq in Eqs. (1, 2) an arbitrarily small constant. In particular, the points k˚ are
isolated, as claimed. Moreover, they are generically simple. To show this, we perturb V , V 7
by tWn pkq and determine the splitting µ ` tµ̃ ` opt2 q, pt Ñ 0q of a degenerate eigenvalue µ of
Hpk˚ q. On general grounds µ̃ is found by diagonalizing W after orthogonally projecting it onto
the unperturbed eigenspace of µ. Here, that eigenspace is the image of ker Ψ70 Ă CN under Ψ7 .
Hence the eigenvalue problem reads
8
´ÿ ¯ 8
´ÿ ¯
P0 pΨ7n q˚ Wn pk˚ qΨ7n P0 a “ µ̃P0 pΨ7n q˚ Ψ7n P0 a , pa P CN q , (59)
n“0 n“0
where P0 is the orthogonal projection onto ker Ψ70 . By Eq. (29) the matrix in brackets on the
r.h.s. is positive definite on CN , while that on the l.h.s. may take arbitrary Hermitian values,
along with Wn pk˚ q. As a result, the eigenvalues µ̃ are generically distinct and, since ε1 pk˚ q ‰ 0,
the points k˚ split into non-degenerate ones. Similarly, points k˚ with det Ψ71 “ 0 correspond to
µ being a Dirichlet eigenvalue for n “ 1. Its perturbative splitting is determined by (59) with
the replacement of 0 by 1. By (29), ker Ψ70 X ker Ψ71 “ t0u, so that the operator on the l.h.s. can
be chosen with independent projections under P0 and P1 . Any coincidence between eigenvalues
of the two Dirichlet problems is thus generically lifted.
The above argument did not pay attention to the time-reversal symmetry of the Hamiltonian
and hence of W pkq. We do so now: A crossing at k˚ “ 0, π may be displaced by perturbing
by an arbitrarily small constant; one at k˚ ‰ 0, π by letting W pkq satisfying Eq. (4). It still
remains arbitrary at k˚ .
In preparation for the proof of part (ii) let us introduce the Casoratian, which is to finite
difference equations what the Wronskian is to ordinary differential equations (both linear and
of second order). Let H be as in Eq. (38). Given ψ “ pψn qnPZ , ϕ “ pϕn qnPZ with ψn , ϕn P CN
viewed as column, resp. row vectors, let
Cn pϕ, ψq “ ϕn A˚ ψn`1 ´ ϕn`1 Aψn . (60)
Suppose that ψ, ϕ satisfy the Schrödinger equation in the form
pH ´ zqψ “ 0 , ϕpH ´ zq “ 0 , (61)
where pϕHqn “ ϕn´1 A˚ ` ϕn`1 A ` ϕn Vn . Then
24
C1) Cn pϕ, ψq is independent of n P Z and denoted Cpϕ, ψq.
C2) Let ψpzq solve the first Eq. (61). Then ψpz̄q˚ solves the second. In particular Cpψpz̄q˚ , ψpzqq
is well-defined.
C3) Items (C1–2) apply if H, ψ, Z are replaced by H 7 , ψ 7 , N.
Property (C1) follows from the identity
Cn pϕ, ψq ´ Cn´1 pϕ, ψq “ ϕn pAψn´1 ` A˚ ψn`1 q ´ pϕn´1 A˚ ` ϕn`1 Aqψn (62)
and Eq. (61); the others are straightforward. The Casoratian may be extended literally to ma-
trix solutions Ψ, Φ of Eqs. (61), in which case Cn pΦ, Ψq is itself a matrix. Its entries are the
Casoratians of the rows and columns of Φ, resp. Ψ. Properties (C1–3) hold correspondingly.
Proof of Lemma 5.5, continued. ii) We drop k and notice that
C0 pΨ7 pz̄q˚ , Ψ7 pzqq “ Ψ70 pz̄q˚ A˚ Ψ71 pzq ´ Ψ71 pz̄q˚ AΨ70 pzq “ Lpzq ´ Lpz̄q˚ .
By (C2) the l.h.s. equals limnÑ8 Cn pΨ7 pz̄q˚ , Ψ7 pzqq “ 0; whence the reflection property. The
statement about the eigenvalue branch follows from part (i) and the definition of L.
iii) We drop k “ k˚ . Let u P CN be the normalized eigenvector of Lpµq with eigenvalue
lpµq “ 0, whence Ψ70 pµqu “ 0 by (30). Thus,
Bl ˇˇ ` BL ˇˇ ˘ ` 7 7 ˘ ` ˘
ˇ “ u, ˇ u “ ´ u, Ψ1 pµq˚ ApBz Ψ0 pµqqu “ u, C0 pΨ7 pµq˚ , Bz Ψ7 pµqqu .
Bz µ Bz µ
Next we observe that
Cn pΨ7 pµq˚ , Bz Ψ7 pµqq ´ Cn´1 pΨ7 pµq˚ , Bz Ψ7 pµqq “ Ψ7n pµq˚ Ψ7n pµq ě 0 ,
because of (62) and of pH 7 ´ zqBz Ψ7 “ Ψ7 . Since Cn pΨ7 pµq˚ , ř Bz Ψ7 pµqq Ñ 0, pn Ñ 8q, we

7 ˚ 7
conclude pBz lqpµq ď 0. Actually, equality is excluded, because 8 n“1 Ψn pµq Ψn pµq is positive
definite, as remarked earlier.
iv) Let εpkq the eigenvalue branch crossing µ at k˚ , whence lpεpkq, kq “ 0 near k˚ . The claim
follows from
d ˇ Bl ˇˇ Bl ˇˇ
lpεpkq, kqˇ ε1 pkq `
ˇ
“ ˇ ˇ
dk k“k˚ Bz µ,k˚ Bk µ,k˚
together with (31).
This concludes the proof of Thms. 5.4 and 6.4.
7.1 Supplementary results

This section contains a few details related to Sect. 4, some of independent interest and some
needed in connection with Thm. 5.7, but none with the proof of the basic results, which is by
now complete. Also addressed is the relation with other indices found in the literature.
Polar decomposition. Given a matrix T P GLpN q, let T “ P U be its (unique, left) polar
decomposition, i.e. P “ P ˚ ą 0, U ˚ U “ 1.
25
Lemma 7.2. Let T P GLpN q satisfy Eq. (15). Then so does U and the following deformations
of T retaining that property are possible: (i) T to U , while keeping the polar part fixed; (ii) U
to 1, while preserving unitarity; (iii) and hence T to 1.
Proof. The right polar decomposition is T “ U P̃ with P̃ “ U ´1 P U . Eq. (15) states T “

Θ´1
0 T
´1 Θ “ pΘ´1 U ´1 Θ qpΘ´1 P ´1 Θ q and is hence equivalent to
0 0 0 0 0
U “ Θ´1
0 U
´1
Θ0 , U ´1 P U “ Θ´1
0 P
´1
Θ0 . (63)
In particular the preliminary

ř claim holds true.
i) Denoting by P “ i λi Πi , (λi ą 0) the spectral decomposition of P , the second Eq. (63) is
equivalent to the existence of an involution i ÞÑ ı̄ such that λı̄ “ λ´1
i and U ´1 Πı̄ U “ Θ´1 0 Πi Θ0 .
By interpolating the eigenvalues but not the eigenprojections, it becomes clear that T can be
deformed as stated. ř
ii) Denoting now by U “ i zi Πi , (|zi | “ 1) the spectral decomposition of U , the first Eq. (63)
is equivalent to Πi “ Θ´1
0 Πi Θ0 . By again interpolating eigenvalues only we arrange for U going
to 1.
Let us also mention the following consequence:
Remark 7.3. Let T pϕq “ P pϕqU pϕq be the polar decomposition of T pϕq P GLpN q. If the
family T has the Kramers property, then so does U . Even though the zi pϕq are generally not the
eigenvalues of U pϕq, it holds true that IpT q “ IpU q.
Proof. By (i) of the previous lemma, T pϕq can be continuously extended to the right on an
interval ϕ P pπ, bs in such a way that there Eq. (15) holds throughout, U pϕq ” U pπq, and
P pbq “ 1. Likewise on an interval ra, 0q. By continuity neither index, IpU q or IpT q, changes in
the process. At this point they are manifestly equal. In fact tP | P ą 0u is a convex set; hence
any continuous map P : ra, bs Q x ÞÑ P pxq ą 0 with P paq “ P pbq “ 1 is homotopic to P ” 1.
Proof of Lemma 4.4. As mentioned, only Eq. (18) requires proof. Consider the families
T21 pϕq “ M´ p0qT1 pϕqM` p0q´1 , T22 pϕq “ M´ pϕqT1 pπqM` pϕq´1 .
They enjoy the Kramers property along with T1 ; they can be concatenated, since T21 pπq “ T22 p0q;
and T2 is homotopic to T21 #T22 , as seen by postponing the change of M˘ till after that of T1 .
Hence
IpT2 q “ IpT21 #T22 q “ IpT21 qIpT22 q ,
where IpT21 q “ IpT1 q by deforming M˘ p0q to 1 while preserving (17); and IpT22 q “ IpM q by
Lemma 7.2 (iii).
Alternate definition. There is an alternative way to Def. 4.8 of computing the index IpEq
of a time-reversal invariant bundle E. Let upϕq : T : Ñ F pEq be a section satisfying (22) on the
: 1 : The
twice cut torus T “ pr´π, 0s \ r0, πsq ˆ S , where τ exchanges the left and right halves of T.
section gives rise to four boundary values and two transition matrices parametrized by ϕ2 P S 1 :
up0`, ϕ2 q “ up0´, ϕ2 qT0 pϕ2 q , up`π, ϕ2 q “ up´π, ϕ2 qTπ pϕ2 q . (64)
26
Lemma 7.4. In this situation, the families T0 , Tπ enjoy the Kramers property and
IpEq “ IpT0 qIpTπ q . (65)
Def. 4.8 corresponds to the special case of a single cut at ϕ1 “ ˘π, whence T0 ” 1; a single
cut at ϕ1 “ 0 could have been used there instead.
Proof. The two families are Kramers because Lemma 4.6 still applies. Let v : T9 Ñ F pEq be a
time-reversal invariant section as in Lemma 4.5 and let M pϕq, (ϕ P T) : be the change of frame
vpϕq “ upϕqM pϕq. It satisfies (23). Since vpϕ1 , ϕ2 q is continuous along ϕ1 “ 0˘ and satisfies
(24) along ϕ1 “ ˘π we have
1 “ M´0 pϕ2 qT0 pϕ2 qM`0 pϕ2 q´1 , T pϕ2 q “ M´π pϕ2 qTπ pϕ2 qM`π pϕ2 q´1 ,
where M˘0 pϕ2 q “ M p0˘, ϕ2 q, M˘π pϕ2 q “ M p˘π, ϕ2 q. By Eq. (18),

´1 ´1
1 “ IpT0 qIpM´0 M`0 q, IpT q “ IpTπ qIpM´π M`π q.
As remarked in the proof of Lemma 4.7, the family M p´ϕ1 , ¨qM pϕ1 , ¨q´1 is Kramers for fixed
ϕ1 P r0, πs. It is continuous in ϕ1 in that interval, and its index constant. Hence the claim.
Splitting the torus. We consider a time-reversal invariant bundle E on the torus T, as

described in Sect. 4.3. We suppose moreover that ϕ1 “ ˘π{2 is a distinguished pair of lines in
the sense that
Ep´ π2 ,ϕ2 q “ Ep π2 ,ϕ2 q , pϕ2 P S 1 q (66)
(In applications, that identification of fibers may occur because the bundle E is effectively the
pull-back of another one under a map f with f p´π{2, ϕ2 q “ f pπ{2, ϕ2 q.) Let us define the torus
T1 “ tpϕ1 , ϕ2 q | |ϕ1 | ď π{2, ϕ2 P S 1 u with identified edges ϕ1 “ ˘π{2, whence it still has four
time-reversal invariant points. By (66) E1 “ E æ T1 is a well-defined time-reversal invariant
bundle. Similarly for T2 “ tpϕ1 , ϕ2 q | |ϕ1 ´ π| ď π{2, ϕ2 P S 1 u and E2 “ E æ T2 .
Lemma 7.5. [Splitting lemma] In the situation set by Eq. (66) we have
IpEq “ IpE1 qIpE2 q . (67)
Proof. By Lemma 4.5 there exists a section v1 pϕq of F pE1 q satisfying (22) on T9 1 , the torus T1 cut
along ϕ1 “ ˘π{2; likewise, v2 pϕq on T9 2 with cut ϕ1 “ ˘π. Let T1 pϕq, T2 pϕq be the corresponding
transition matrices (24). By (66) we may also introduce transition matrices M˘ pϕ2 q, (ϕ2 P S 1 )
by
π π
v2 p˘ , ϕ2 q “ v1 p˘ , ϕ2 qM˘ pϕ2 q . (68)
2 2
Multiplying by Θ, ε from the left, resp. right we obtain
π π
v2 p¯ , ´ϕ2 q “ v1 p¯ , ´ϕ2 qΘ´1
0 M˘ pϕ2 qΘ0
2 2
and thus
Θ0 M˘ p´ϕ2 q “ M¯ pϕ2 qΘ0 . (69)
27
Moreover, the l.h.s. of (68) is independent of ˘, whence
T1 pϕ2 q “ M´ pϕ2 qM` pϕ2 q´1 . (70)
On T9 we define the section

#
v1 pϕq , p|ϕ1 | ď π{2q
vpϕq “
v2 pϕqM˘ pϕ2 q , p π2 ď ˘ϕ1 ď πq .
´1
It is continuous by (68) and satisfies (22): by hypothesis for |ϕ1 | ď π{2, but also for π{2 ď
˘ϕ1 ď π by (69). Its transition matrix is read off as
T pϕ2 q “ M´ pϕ2 qT2 pϕ2 qM` pϕ2 q´1 .
Now Eq. (67) follows in the form IpT q “ IpT1 qIpT2 q from (18, 70).
Indices in the literature. The Z2 -index has been introduced in various other forms before
[16, 9], but to our knowledge not in the form (25). We shall first recall a formulation [9] (there
connected to [16]), which rests on an additional, metric structure, and later establish equivalence
with ours when there is overlap. Like our index, it first deals with continuous families of matrices
W pϕq, (0 ď ϕ ď π) which, in lieu of the Kramers property, enjoy antisymmetry W T “ ´W at
endpoints W “ W p0q, W pπq. There,
det W “ ppf W q2 ,
a
where pf W is the Pfaffian of W ; in between, consider continuous branches ˘ det W pϕq. One of
them will connect pf W p0q to IpW p qpf W pπq, where IpW p q “ ˘1 defines the index of the family,
cf. ([9], Eq. (3.24)). In the special case that pf W p0q “ pf W pπq, and hence det W p0q “ det W pπq,
the index reduces to
p q “ p´1qn ,
IpW (71)
where n is the winding number of det W pϕq, (0 ď ϕ ď π).
Let us move on to time-reversal invariant bundles E. They are assumed equipped with a
compatible hermitian metric, i.e. with an inner product x¨, ÿ on any fiber Eϕ such that Θ˚ Θ “ 1.
Use is made of the fact that E is trivial if Θ is disregarded. There thus is a section vpϕq of F pEq
9 but at the price of forgoing
on the torus T Q ϕ “ pϕ1 , ϕ2 q. We stress: not just on the cut torus T,
time-reversal symmetry. It can be taken to consist of orthonormal frames v “ pv1 , . . . vN q. Let
Wij pϕq “ xvi pϕq, Θvj pτ ϕqy ;
see [12] and ([9], Eq. (3.16) with ϕ, τ ϕ switched). Then W pϕq˚ W pϕq “ 1 and W pϕqT “ ´W pτ ϕq.
By this last property, one may define the index as
IpEq
p “ IpW
p 0 qIpW
p πq ,
where W0 pϕ2 q “ W p0, ϕ2 q, Wπ pϕ2 q “ W pπ, ϕ2 q, p0 ď ϕ2 ď πq; see ([12], Eq. (10)). The relation
to the indices of the present work is as follows.
Proposition 7.6. [Relation between indices]
28
i) Suppose T pϕq, p0 ď ϕ ď πq has the Kramers property and W pϕq :“ T pϕqε is antisymmetric
at endpoints. Then
IpT q “ IpW
p q. (72)
ii) For a time-reversal invariant vector bundle E as above we have
IpEq “ IpEq
p .
Proof. We begin with a preliminary remark. Consider the three properties of a matrix T P
GLpN q:
T εT ε “ ´1 , T ˚T “ 1 , T T “ ´εT ε ,
called Kramers, unitarity and antisymmetry, the first one being indeed a restatement of Eq. (15).
In terms of W “ T ε they respectively read
W W “ ´1 , W ˚W “ 1 , W T “ ´W .
Clearly any two of them imply the third.

i) The assumptions imply that T p0q, T pπq are unitary. By Lemma 7.2 (ii) we may deform the
family at its endpoints in such a way that T p0q “ T pπq, while retaining the Kramers property
and unitarity, and hence antisymmetry. In the process both indices (72) remain defined and
constant. At this point, see Eqs. (13, 16), IpT q “ p´1qn , where n is the winding number of
det T pϕq. The claim follows by (71) and det T pϕq “ det W pϕq.
: as
ii) We define a time-reversal invariant section upϕq, see Eq. (22), on the twice cut torus T
#
vpϕq , p´π ď ϕ1 ď 0q
upϕq “
Θvpτ ϕqε , p0 ď ϕ1 ď πq.
It has ϕ1 “ 0˘ and ϕ1 “ ˘π as lines of discontinuity, where Eq. (64) becomes
Θvp0, ´ϕ2 qε “ vp0, ϕ2 qT0 pϕ2 q , Θvpπ, ´ϕ2 qε “ vpπ, ϕ2 qTπ pϕ2 q ,
by the continuity of v. Taking the inner product with the vectors of the orthonormal frames
vp0, ϕ2 q resp. vpπ, ϕ2 q yields W0 pϕ2 qε “ T0 pϕ2 q and Wπ pϕ2 qε “ Tπ pϕ2 q. The claim follows by
(65, 72).
8 The bulk index as an index of Bloch bundles: Proofs

As a preliminary to the proof of Lemma 5.8, we consider Eq. (41) not just for |ξ| “ 1, but for
ξ ‰ 0, together with its characteristic polynomial
P pξ, zq “ detpHpξq ´ zq “ detpAξ ´1 ` A˚ ξ ` V ´ zq . (73)
Its basic properties, to be proven later, are as follows.
29
Lemma 8.1. P is a polynomial of degree M N in z and a Laurent polynomial in ξ of degrees N
and Ń . Moreover
P pξ, zq “ P pξ¯´1 , z̄q (74)
¯ also
and, in the time-reversal invariant case ΘHpξqΘ´1 “ Hpξq,
¯ z̄q .
P pξ, zq “ P pξ, (75)
(As a function of k that case holds true for k “ 0, π.)
We then consider the Riemann surface (Bloch variety) defined by
B “ tpξ, zq P C˚ ˆ C | P pξ, zq “ 0u , (76)
where C˚ “ Czt0u. We make some assumptions, which are typically true: B is non-singular as
a Riemann surface, meaning that pBP {Bξ, BP {Bzq ‰ 0 at all points pξ0 , z0 q P B; and, if either
partial derivative vanishes, the corresponding second derivative does not. As a result, near a
point pξ0 , z0 q where BP {Bz ‰ 0 we can solve P pξ, zq “ 0 as z “ zpξq with z analytic; and if
BP {Bξ “ 0 at a critical point ξ “ ξ0 , then z 1 pξ0 q “ 0, but z 2 pξ0 q ‰ 0. Similarly for interchanged
roles of variables and at branch points. We conclude that eigenvalue branches z “ zpξq have
only non-degenerate critical points, and only branch points of order 2.
For k “ 0, π we will have to exceptionally allow singular points pξ0 , z0 q P B where both
derivatives of P vanish. However they shall be ordinary double points, i.e. the Hessian of P is
non-degenerate.
Of importance is also the real Bloch variety
B0 “ tpξ, zq P B | |ξ| “ 1u . (77)
In fact, for any z P C let m ě 0 be the number of ξ with |ξ| “ 1 and P pξ, zq “ 0, i.e. of points
pξ, zq P B0 . Then m ą 0 implies z P R, since now ξ¯ “ ξ ´1 and P pξ, ¨q becomes the characteristic
polynomial of a hermitian matrix. Moreover m is the multiplicity of z as a point of the spectrum
σpHq. By (74) the remaining 2N ´ m “ 2pN ´ m{2q points pξ, zq with |ξ| ‰ 1 come in pairs
pξ, zq, pξ¯´1 , zq. In particular m is even.
For each pξ, zq P B we consider the (geometric) eigenspace E rξ,z of the eigenvalue z of the
matrix seen in Eq. (73).
Lemma 8.2. For k ‰ 0, π

E
r :“ tpξ, z, Ψq | Eq. (41) holdsu
is a line bundle with base B and fibers E

rξ,z .
In order to properly state the relation between this bundle and E we make the following
definition.
Definition 8.3. A generalized Bloch solution pψrn qnPZ of energy z and quasi-periodicity ξ ‰ 0
satisfies pH ´ zqψr “ 0 and
ψrn`pM “ ξ p pψrn ` pξ ´1 ψn q , (78)
where pψn qnPZ is a Bloch solution for the same ξ, z.
30
Eq. (78) characterizes a generalized eigenvector (of order 2) of the translation operator. In
r P CM N with
terms of Eq. (39) a generalized Bloch solution ψr corresponds to Ψ
r ` pA˚ ξ ´ Aξ ´1 qξ ´1 Ψ “ z Ψ
HpξqΨ r.
Lemma 8.4. Let the eigenvalue branch z “ zpξq have a critical point at ξ “ ξ0 and let Ψpξq be
r at pξ0 , z0 “ zpξ0 qq. Then Ψ1 pξ0 q corresponds to a generalized Bloch solution
a local section of E
for ξ0 , z0 .
Proof. Follows by differentiating in ξ Eqs. (40, 41). In fact,

dH dz
pHpξq ´ zqΨ1 ` Ψ“ Ψ (79)
dξ dξ
with z 1 pξ0 q “ 0 and ξdH{dξ “ A˚ ξ ´ Aξ ´1 .
Finally, we consider
B´ “ tpξ, zq P B | |ξ| ă 1u
pBB´ “ B0 q equipped with the projection
π : B´ Ñ C, pξ, zq ÞÑ z . (80)
The preimage π ´1 pzq of a point z P C consists of N points iff z P ρpHq. In that case we have
π˚ pE
rπ´1 pzq q “ Ez ,
where the l.h.s. is to be understood as a direct image: The fiber at z P ρpHq is ‘E rξ,z with
´1
sum over pξ, zq P π pzq; the equation itself on rests on the relation (39) between Ψ P E rξ,z
and ψ P Ez . Some care applies at points pξ, zq where BP {Bξ “ 0, i.e. to ξ a multiple (in fact,
double) preimage of z, because this seemingly results in a missing solution on the l.h.s.. It should
however be read so as to include the there existing generalized Bloch solution. Equivalently, for
z 1 near z the fiber π˚ pE
rπ´1 pz 1 q q is of dimension N and has a limit (of the same dimension) as
1
z Ñ z. With this reading both sides are continuous in z P ρpHq.
Likewise, the preimage π ´1 pγq of a loop γ Ă ρpHq is a cycle in B´ , possibly consisting of
several loops, and we have
r æ π ´1 pγqq “ E æ γ .
π˚ pE (81)
Besides of π we consider the projection
σ : B´ Ñ tξ P C | |ξ| ă 1u, pξ, zq ÞÑ ξ
and the bundle σ˚ pE

rσ´1 p¨q q. Its fibers are of dimension M N , even at branch points by continuous
interpretation.
Proof of Lemma 5.8. We discuss a typical case only. Let us first recall the base spaces of the
two bundles, E pjq “ E æ Tj and Ej , seen in Eq. (45): On the l.h.s. it is Tj Q pz, kq and on the
r.h.s. B Q pκ, kq. In a nutshell, we shall show that each side of the equation is associated with
a cycle in B´ Q pξ “ eiκ , kq or actually with a family of cycles parametrized by k P S 1 ; and,
more precisely, with the bundle E r on B´ of Lemma 8.2 restricted to that family. The proof
31
(a)
(b)
(c)
Figure 7: Left: Real energy curves λl peiκ q, pκ P S 1 , l “ 2j ´ 1, 2jq for various values of
k. The first case corresponds to time-reversal invariant points k “ 0, π. Right: Loops λl pξq,
p|ξ| “ q ă 1q in the complex plane. In the first case the loop is run through twice. The second
case is illustrated thrice (a–c) for different values of q (or k). The dashed lines represent energy
bands of multiplicity 2, resp. 4 if thick.
then proceeds by an interpolation deforming one family into the other, up to contractible cycles.
Most of the discussion occurs at fixed k, which is hence once more omitted.
For E pjq “ E æ Tj on the l.h.s. of Eq. (45) the association is by (81) with γ “ tz | pz, kq P Tj u.
For Ej on the r.h.s. a longer discussion is needed. The Bloch bundle Ej (see Def. 5.6) can be
analytically continued from real κ to the path I “ tκ P C | Im κ “ εu, as long as ε is small
enough. By ξ “ eiκ there correspond two paths Iplq “ tpξ, λl pξqq | |ξ| “ e´ε u, in the sense that
Ip2j´1q Y Ip2jq Ă σ ´1 pIq is a subcycle lying in the sheets l “ 2j ´ 1, 2j of B´ . Then
Ej “ σ˚ pE
r æ pIp2j´1q Y Ip2jq q . (82)
For the sake of illustration Fig. 7 shows the real energy curves λl pκq ” λl peiκ q, pκ P S 1 , l “
2j ´ 1, 2jq for various values of k (left) next to the complex loops λl pξq, p|ξ| “ qq with q ă 1
close to 1 (right). The loops are understood on the observation that a portion of the energy curve
λl pκq increasing in κ is flanked by a nearby branch running in the upper half-plane. The loops
can be mutually and self-intersecting, and intersect the spectrum. However the outer boundary
γ1 of both loops does not. In the last case of Fig. 7 γ1 consists of two disconnected loops, each
surrounding one of the bands l “ 2j ´ 1, 2j, the second one being shown in Fig. 8, right.
32
The contour γ Ă ρpHq underlying E pjq may be picked as γ1 , at least in the cases where the
latter is connected.
κ γ1 λ
Γλ
γλ γ2
I1 I2 I1
ε
A1 A2 A1
´π π
Figure 8: Basic elements. Left: The sheet l “ 2j of B´ near B0 and the path Ip2jq “
I1 Y I2 : |ξ| “ q parametrized as ξ “ eipκìεq , (κ P S 1 ) for q “ e´ε ă 1. Right: Its image
under ξ ÞÑ λ2j pξq is a cycle γ1 Y γ2 with loops γi “ λ2j pIi q. Further elements. Annulus
A “ A1 Y A2 “ tκ | 0 ď | Im κ| ă εu – tξ | q ă |ξ| ď 1u and its subsets Ai , pi “ 1, 2q, where λ2j
is 1 to 1, resp. 2 to 1. The two sets are shown separated by a line (dotted). The set A2 is mapped
onto the inside of γ2 ; A1 onto the complementary subset of the inside of γ1 . Interpolating curve
γλ and one of its preimages, Γλ (both dashed); another preimage (dashed-dotted).
We claim:
Lemma 8.5. The bundle E æ γ contains a subbundle Ep which is homotopic to the Bloch bundle
Ej æ I – Ej . Moreover
E
p “ π˚ pEr æ Γq
p , (83)
p Ă π ´1 pγq is a subcycle of the cycle π ´1 pγq Ă B´ . It covers γ twice.
where Γ
q :“ π ´1 pγqzΓ
As a result, Γ p is a cycle too, and
E
q :“ π˚ pE
r æ Γq
q
provides a complementary subbundle,
Eæγ“E
p‘E
q.
Lemma 8.6. Γ
q is contractible within B´ .
As a result Er æ Γ,
q and hence E,q are trivial.
This is the core of the proof of Lemma 5.8. In summary,
E pjq – E æ γ – E
p – Ej
where – is by homotopy or by dropping trivial subbundles. Some further comments in relation

with Lemma 8.5 are appropriate.
First, one ought to worry about whether these equivalences persist once the dependence on
k P S 1 is included into the picture. This will be understood by inspection of the proofs of the
two lemmas. Indeed, the construction of Γ,p and hence that of E,
p E,
q will not involve any choices,
except among homotopically equivalent constructions. In particular, as k runs from 0 to 2π,
33
the cycles Γ
p match at the endpoints, whence the family glues up to tori; likewise for deformed
cycles used in the interpolation.
Second, no deformation is needed at k “ 0, π, see Fig. 7 (first case), where
Γ
p “ Ip2j´1q Y Ip2jq . (84)
Third, a cycle C Ă B´ (parametrized by k) may give rise to a (i) torus (still called C) and to
a (ii) bundle E r æ C, which are both time-reversal invariant. For (i) it suffices that the map
¯ z̄, ´kq on B´ ˆ S 1 leaves C invariant. For (ii) the bundle, being of dimension
τ : pξ, z, kq ÞÑ pξ,
1, does not qualify as such. An additional structure is needed, namely an involution α on C
preserving k, which moreover commutes with τ . Then τ descends to the quotient C{α, which
consists of pairs of points related by the involution as well as by fixed points. Now E r æ C shall
actually stand for E æ pC{αq, whose fibers Eξ,z ‘ Eαpξ,zq are, by continuous interpretation, of
r r r
dimension 2 even at fixed points. Both bundles E r æ Γ
p and E r æ pIp2j´1q Y Ip2jq q induce such
involutions, α1 on Γ,p resp. α2 on Ip2j´1q Y Ip2jq : Points pξ, zq and pξ 1 , z 1 q are related by α1 if
z “ z ; and by α2 if ξ “ ξ 1 . We remark that α1 , α2 are consistent with Eqs. (83, 82), respectively.
1
At k “ 0, π, where the cycles agree by (84), the involutions α1 , α2 nevertheless differ. We will
prove the homotopy in Lemma 8.5 by constructing cycles Cλ , p1 ď λ ď 2q interpolating between
C1 “ Γp and C2 “ Ip2j´1q Y Ip2jq . The bundle E r æ Cλ goes with the ride because of the following
fact.
Lemma 8.7. Let α1 , α2 be as above. Then there is an interpolating involution αλ on Cλ ,

p1 ď λ ď 2q commuting with τ æ Cλ .
8.1 Further details

Proof of Lemma 8.1. We expand the determinant (73) along the first row of blocks. Then, by
the Laplace rule,
ˇ ˇ
ˇ A V1 ´ z
ˇ A˚ ˇ
ˇ
ˇ .. .. .. ˇ
ˇ 0
ˇ . . . ˇ
ˇ
P pξ, zq “ p´1qM pN `1q
pdet Aqξ Ń ˇ
¨ˇ . .. .. .. ˚ ˇ ` P` pξ, zq
ˇ
ˇ . . A ˇ
ˇ .. .. ˇ
ˇ . . VM ´2 ´ z ˇ
ˇ ˚ ˇ
ˇ A ξ 0 A ˇ
where P` contains only powers of ξ of degree ´pN ´ 1q or higher. When looking for the leading
term in ξ ´1 we may replace A˚ ξ by 0, which leaves an upper block triangular matrix. That term
thus is p´1qM pN `1q pdet AqM ξ Ń . Similarly, the leading term in ξ is p´1qM pN `1q pdet A˚ qM ξ N .
Eq. (74) follows from det M “ det M˚ . In the time-reversal invariant case we use det M “
detpΘMΘ´1 q, where Θ “ diagpΘ, . . . , Θq in the notation of Eq. (39). Hence the last claim.
Proof of Lemma 8.2. To be shown is that, for each pξ, zq P B, the eigenvalue z of Hpξq,
cf. Eq. (41), is geometrically simple.
34
If BP {Bz ‰ 0 the eigenvalue z “ zpξq is even algebraically simple. In that case the eigenspace
depends analytically on ξ. If BP {Bz “ 0 at pξ0 , z0 q P B, meaning a branch point, then we have
by our general assumptions,
B2P BP
2
‰0, ‰0;
Bz Bξ
in particular z0 is an eigenvalue of Hpξ0 q of algebraic multiplicity 2 and ξ 1 pz0 q “ 0, ξ 2 pz0 q ‰ 0.
It follows that nearby eigenvalues of Hpξq are given as
d
2
z ´ z0 “ ˘ 2
pξ ´ ξ0 q ` Opξ ´ ξ0 q , pξ Ñ ξ0 q . (85)
ξ pz0 q
In order to show that the geometric multiplicity of z0 is nevertheless 1, we denote by Πpξq the
projection onto the group (85) of eigenvalues z (it is analytic in ξ; [17], Sect. II.1.4), as well as
by Π0 “ Πpξ0 q and by
N0 “ pHpξ0 q ´ z0 qΠ0
the eigenprojection and eigennilpotent of z0 . The claim amounts to N0 ‰ 0, and we prove it by
contradiction: N0 “ 0 implies that
Hpξq ´ z0
Πpξq
ξ ´ ξ0
has a removable singularity at ξ “ ξ0 . Hence z ´ z0 “ Opξ ´ ξ0 q, in violation of (85).
The proofs of Lemmas 8.5 and 8.6 will mostly deal with a single k at a time and will be
given them in two parts. The first part is, so to speak, a test run which avoids the complications
due to the overlapping bands. It is thus restricted to the last case of Fig. 7 and pretends that
γ “ γ1 even there. The limitations will be overcome in a second part.
Proof of Lemma 8.5, first part. We consider a homotopy of curves γλ , p2 ě λ ě 1q which lie
between γ2 and γ1 . In particular, we keep the endpoints of γλ fixed, as indicated by the dashed
line in Fig. 8 (right). Among the many curves in π ´1 pγλ q let us track the curve, Γλ Ă B´ , arising
from I2 at λ “ 2 by continuity, i.e. Γ2 “ I2 . The curve Γλ has to stay outside of A. Indeed, it
can cross neither I2 , as πpΓλ q would end up inside γ2 , nor I1 , as πpΓλ q would approach γ1 from
the outside. As Γλ does not penetrate into A, it does not approach the real κ axis (i.e. |ξ| “ 1)
on the branch of λ2j . Moreover it does not either on any other branch of B0 , since those map
to other, disjoint bands. That ensures that Cλ “ I1 Y Γλ is a cycle in B´ and that the bundle
Er æ pI1 YΓλ q is defined and continuous in λ. Finally, the homotopy of bundles π˚ pE r æ pI1 YΓλ qq,
p2 ě λ ě 1q does the job: For λ “ 2 we have
π˚ pE
r æ pI1 Y I2 qq “ Ej æ I
by construction; for λ “ 1, πpI1 Y Γ1 q “ γ1 Y γ1 “ γ1 and
π˚ pE r æ π ´1 pγ1 qq “ E æ γ1
r æ pI1 Y Γ1 qq Ă π˚ pE
by (81). Thus Lemma 8.5 holds with Γ

p “ I1 Y Γ1 .
Let us recall the count of preimages π ´1 pzq done in relation with Eq. (80). We find it conve-
nient to pretend that all points z P C, i.e. including z P σpHq, have precisely N preimages under
35
π. This is enforced by including the m preimages on B0 , but by counting them with weight 1{2,
as explained in the sequel of Eq. (77). The preimages of a loop intersecting the spectrum then
is a pseudo-cycle, where it is tolerated that curves break up at pairs of points in B0 with same
λl pκq (see the joined dashed and dashed-dotted curves in Fig. 8, left).
Proof of Lemma 8.6. We first observe from the previous proof that any point z inside γ1 has
N ´ 2 preimages pξ, zq outside of and away from I1 Y Γ1 . We next extend γλ from 1 ď λ ď 2
to 2 ď λ ď 3 so that it contracts γ2 to a point γ3 “ tz0 u inside of it. We then consider the
deformation Γ p λ , p1 ď λ ď 3q of pseudo-cycles covering γλ twice and starting with Γ
p 1 “ I1 Y Γ 1 .
By continuity in λ, Γλ moves inside of I1 Y Γ1 , while N ´ 2 preimages of γλ remain outside. As
p
q λ :“ π ´1 pγλ qzΓ
a result Γ p λ stays away from B0 and is a cycle in B´ .
Proof of Lemma 8.5, second part. We extend the proof to general values of k. They are illustrated
by the three cases seen in Fig. 7 with the middle one having three variants (a–c). The argument
will rely on two sheets (l “ 2j ´ 1, 2j) of B´ , as opposed to just one before (see Fig. 8 left).
Each comes with an annulus in κ bounded above by Ip2j´1q , Ip2jq . The sheets share two branch
points: In the first case (k “ 0 or π) they rather are, to be precise, two double points at κ “ 0
and π. In the variant (a) each of them gives way to two branch points, out of which one is in
B´ , and actually in the annuli; in (b) the branch points have reached the paths Ip2j´1q , Ip2jq
(this might occur at different k for the two points), and in (c) they have left the annuli. This
is seen in Fig. 7 (right) as a rearrangement of loops through an intersection point at (b). It
should be noted that the Bloch bundle with fibers rΨ2j´1 pκ, kq, Ψ2j pκ, kqs remains continuous as
the branch point reaches the paths Ip2j´1q , Ip2jq , at least if properly interpreted: That point is
then common to the two paths, resulting in a just 1-dimensional fiber, cf. Lemma 8.2; however
the 2-dimensional fiber has a limit at the branch point, as remarked after Eq. (85), and unlike
the two spanning vectors individually. That limit provides the appropriate fiber at the branch
point.
In the first case of Fig. 7, where the curves do not intersect the spectrum, the conclusion is
reached as before, but without need for deformations. We next turn to the three variants (a–c)
of the middle case. Let us specify the deformations of the curves in C and describe their lifts in
B.
(a-b) We deform the inner parts γ2 of the curve till they reach the outermost parts γ1 , while
keeping the intersection points fixed. By the same arguments as before the preimage selected
by continuity remains outside the annuli of both sheets.
(c) Let us expand the inner loops. Quite soon the situation will look as in the undeformed
variant (b) with the inner loops touching essentially the same curves as they did there. By
continuity from (b) the touching also takes places between their respective lifts in B. From
there on the deformation proceeds as in variant (b). As a whole it includes and reverts the
rearrangement of loops which occurred between (a) and (c).
In the last case we first perform the deformation described in the first part of this proof. The
two components of γ1 , which at this point are run through twice, still require a deformation to
a countour γ encircling both bands at once. When the two components of γ1 first touch, so do
the two lifts of each, as inherited from (c) by continuity.
Likewise the proof of Lemma 8.6 extends to the other cases, too. In the first case (k “ 0, π)
the loops γλ (2 ď λ ď 3) are to be chosen so that γλ “ γλ .
36
Proof of Lemma 8.7. We begin with some preliminary observations about circle homeomorphisms
β : S 1 Ñ S 1 , p ÞÑ βppq. Their degree is dpβq “ ˘1 and any two of them are homotopic iff they
have the same degree (β1 „ β2 ô dpβ1 q “ dpβ2 q); if so, the homotopies βλ , p1 ď λ ď 2q between
them fall into connected components which, relatively to one another, are labeled by n P Z.
Indeed, for β1 “ β2 and fixed p the map λ ÞÑ βλ ppq has a winding number n (independent of p).
Continuous maps β : S 1 Ñ S 1 are called involutions if β ˝ β “ 1. They are homeomorphisms.
Involutions have either all or no points p P S 1 as fixed points, or else just two [23]. The two
cases correspond to dpβq “ ˘1, respectively. The above statements about homotopy remain
true within the class of involutions.
We remark that the properties to be proven,
α˝α“1, α˝τ “τ ˝α, (86)
are relations within or between the fibers at k and ´k, respectively.

We first consider the cases k “ 0, π. There the cycle Cλ “ S` 1 \ S 1 is independent of λ
´
1 1 1
and consists of two loops S˘ with the involution τ : S˘ Ñ S¯ , pξ, zq ÞÑ pξ, ¯ z̄q (see (75, 76))
1 1
interchanging them. Maps α : S˘ Ñ S¯ are expressible in terms of β “ α ˝ τ : S˘ 1 Ñ S 1 and
˘
properties (86) are (jointly) equivalent to β ˝ β “ 1, β ˝ τ “ τ ˝ β; actually the latter property
just determines β æ S´ 1 in terms of β æ S 1 . It is clear from the definitions of α , α : S 1 Ñ S 1
` 1 2 ˘ ¯
that they satisfy (86). Moreover it is impossible that α1 pξ, zq “ α2 pξ, zq ” pξ 1 , z 1 q for some
pξ, zq P S˘1 , since that would imply ξ “ ξ 1 , z “ z 1 in contradiction with pξ 1 , z 1 q P S 1 . Hence
¯
α1 ˝ α2 has no fixed points. But that map equals β1 ˝ β2 , whence dpβ1 qdpβ2 q “ dpβ1 ˝ β2 q “ 1
and dpβ1 q “ dpβ2 q. Therefore there is an interpolation βλ , p1 ď λ ď 2q with αλ satisfying (86).
In addition, there are more of them, differing from it by any given winding n P Z.
We now turn to 0 ă k ă π. The second property (86) will simply determine αλ at ´k
in terms of its value at k. Hence only the first one matters in the following construction of
αλ , p1 ă λ ă 2q. To begin, for (i) k near 0 the cycle Cλ can still be identified with S` 1 \ S1 ;
´
further on it may (ii) collapse into a ”figure 8” (for λ “ 2 this occurs when a single branch point
intersects C2 ; but does not for λ “ 1), and thereafter (iii) become a single loop. Eventually, the
cycle reverts to two separate loops before k reaches π . In range (i) αλ ought to interchange the
two loops and (86) just determines αλ æ S´ 1 in terms of α æ S 1 . That half of α can simply be
λ ` λ
chosen by continuity from k “ 0. At (ii) continuity requires that the intersection point be a fixed
point of αλ , which can be arranged for on the side of (i). In the range (iii) αλ is an involution
on a single loop with just two fixed points. At the end of that range, the two fixed points are
again to coalesce into an intersection point, but that does not obstruct the construction of αλ
(see the observations at the beginning of the proof).
The whole construction could have begun from k “ π instead. We have to ensure that the
two interpolations αλ constructed from the two ends match at some 0 ă k ă π. This can be
arranged, because their relative winding n can be chosen to vanish.
9 Bulk-edge correspondence through scattering theory: Proofs

Proof of Lemma 6.7. i) The map rpκq ” rpκ, kq is well-defined by the required properties,
and it remains to show that it is real analytic in κ. Away from critical points, κ ‰ κ˘ , let
F pκ, rq “ λpκq ´ λprq, so that F pκ, rpκqq “ 0. Since BF {Br “ ´λ1 prq ‰ 0 for r “ rpκq, the
37
claimed analyticity follows by the implicit function theorem. Near a critical point, say κ` , the
argument must be modified. Let there
#
λpκq´λprq
κ´r , pκ ‰ rq
F pκ, rq “ 1
λ pκq, pκ “ rq.
Note that F pκ, rq “ 0 still has the solution r “ rpκq, but no longer r “ κ, except for κ “ κ` .
Since
BF λpκq ´ λprq ´ λ1 prqpκ ´ rq
“
Br pκ ´ rq2
equals λ2 pκ` q{2 ‰ 0 at pκ, rq “ pκ` , rpκ` qq the solution r “ rpκq is analytic also near κ “ κ` .
ii) We recall the Def. 6.5 of the Bloch bundle E` . As shown in Fig. 5, B r ´ does not contain any
1
non-contractible loop winding around κ P S . Thus there is no obstruction for a smooth section
Ψpκ, kq ‰ 0 of E` on B r ´ , or even on small complex neighborhood in κ thereof. To be shown is
that there is one, Ψ´ pκ, kq, which at fixed k is analytic in κ for pκ, kq in that neighborhood. As
a preliminary, let P pκ, kq be the Riesz projection onto the fiber pE` qκ,k P CN M ,
¿
1
P pκq “ ´ pHpκq ´ zq´1 dz ,
2πi
where Hpκq ” Hpeiκ q is defined in Eq. (41), the variable k is suppressed, and the integration
contour surrounds once the eigenvalue λ` pκq of Hpκq, and no further ones. We observe that P
is analytic in κ.
We construct Ψ´ first by setting Ψ´ pκ, kq “ Ψpκ, kq for, say, κ “ κ` pkq; then by extending
it in κ (at fixed k) through parallel transport:
P Bκ Ψ´ “ 0 , Ψ´ “ P Ψ´ . (87)
This is feasible. In fact, the ansatz
Ψ´ pκq “ Ψpkqλpκq , pλpκq P C˚ q
reduces both equations to

P pBκ Ψq ` pP ΨqBκ log λ “ 0 .
Here P pBκ Ψq “ f pP Ψq for some function f pκq P C, since the fibers are lines. The equation
thus reads Bκ log λ “ ´f , which can be integrated starting from λpκ` q “ 1. It remains to
verify the Cauchy-Riemann condition Bκ Ψ´ “ 0. By the Eqs. (87) and Bκ P “ 0 we indeed have
Bκ Ψ´ “ P pBκ Ψ´ q “ 0.
In preparation for the proof of Lemma 6.8, or actually of an extension thereof, let us list two
further properties of the Casoratian, Eq. (60). This time, the energy z in Eq. (61) is real.
C4) Let ψ be a Bloch solution of quasi-periodicity ξ “ eiκ , pκ P S 1 q and energy z “ λpκq. Then
M
ÿ ´1
˚ 1
Cpψ , ψq “ íλ pκq ψn˚ ψn . (88)
n“0
38
C5) Let ψ “ ψpκq be a Bloch solution as in (C4) for κ near κ˘ . Then ψ 1 pκ˘ q is a generalized
Bloch solution by Lemma 8.4 and
M
ÿ ´1
Cpψ 1˚ , ψq ´ Cpψ 1˚ , ψq˚ “ íλ2 pκ` q ψn˚ ψn . (89)
n“0
The derivation is as follows. We recall that Ψ in Eq. (39) satisfies Eq. (41) with Hpξq “ Hpξq˚ .
Eq. (79) thus yields the Feynman–Hellmann equation
dH dλ
pΨ, ξ Ψq “ í pΨ, Ψq
dξ dκ
by ξpd{dξq “ íd{dκ, the inner product being the standard one on CM N . The l.h.s. equals
˚
ξψM ˚ ¯ ˚ ˚ ˚ ˚ ˚
´1 A ψ0 ´ ξψ0 AψM ´1 “ ψM ´1 A ψM ´ ψM AψM ´1 “ CM ´1 pψ , ψq ,
as claimed in (C4). Differentiation of (88) yields (89) because of Cpψ ˚ , ϕ˚ q “ Ćpϕ, ψq˚ and
λ1 pκ˘ q “ 0.
Let us consider a complex neighborhood of the band λ under consideration:
N pδ0 q “ tλpκq | | Im κ| ă δ0 u , pδ0 ą 0q
The following lemma manifestly implies Lemma 6.8, which therefore will not require a separate
proof.
Lemma 9.1. [Solutions of at most small exponential growth] For small enough ε0 ą 0 there is
δ0 ą 0 such that we have: For z “ λpκq P N pδ0 q there is (up to multiples) a unique solution
ψ 7 ‰ 0 of
pH 7 ´ zqψ 7 “ 0 , ψ07 “ 0
with ψn7 “ Opeε0 n q, pn Ñ `8q. It is of the form
ψn7 “ ψnp12q ` Ope´ε0 n q , pn Ñ `8q (90)
where
ψnp12q “ ψnp1q ` ψnp2q ‰ 0 (91)
and
p1q p2q
i) for z ‰ λpκ˘ q: ψn , ψn are Bloch solutions for H with ξ “ eiκ , resp. ξ “ eirpκq ;
p1q p2q
ii) for z “ λpκ˘ q: ψn , ψn are Bloch, resp. generalized Bloch solutions with ξ “ eiκ “ eirpκq .
p1q p2q
Moreover, if z “ λpκq, pκ P S 1 ), then ψn “ Op1q and ψn , ψn ‰ 0 in case (i); and ψn “ Opnq
in case (ii).
Proof. At first, let z belong to the band under consideration, i.e. z “ λpκq, pκ P S 1 ). As
explained in connection with the Bloch variety Eq. (76), we have P pξ, zq “ 0 for m “ 2 values
of ξ with |ξ| “ 1 (counting multiplicities) and for N ´ 1 values with |ξ| ď e´ε0 and small enough
ε0 ą 0. By possibly making it smaller the same remains true for z P N pδ0 q and small δ0 ą 0,
39
provided |ξ| “ 1 is replaced by e´ε0 ă |ξ| ă eε0 . Any eigensolution ψ of H, which remains
Opeε0 n q, pn Ñ `8q is thus of the form
Nÿ
`1
ψn “ ψnp1q ` ψnp2q ` ψnpjq , pn P Zq (92)
j“3
pjq
where pψn q is a (generalized) Bloch solution corresponding to the above ξ’s, in the stated order.
Disregarding the boundary condition, the same applies to eigensolutions ψ 7 of H 7 for n ą n0 ,
by means of the map (28).
We claim that up to multiples there is precisely one such solution ψ 7 , once the boundary
condition ψ07 “ 0 is imposed. Without any conditions the space of solutions ψ 7 has dimension
2N . The subspaces determined by ψ07 “ 0, resp. Opeε0 n q, have dimensions N , N ` 1. Their
intersection V thus has dimension at least 1. We make two claims: paq V Xtψ 7 | ψ p12q “ 0u “ t0u,
cf. (90), which implies Eq. (91) and dim V ď 2; and the stronger: pbq dim V ď 1, and hence
dim V “ 1, which is the statement of uniqueness of ψ 7 . Properties pa, bq are stable, because the
nullity dimpV1 X V2 q of a pair of subspaces is an upper semi-continuous function of them ([17],
Thm. IV.4.24); hence it suffices to prove them for z “ λpκq with κ real.
We begin with paq: The opposite would amount to an embedded eigenvalue, which is ruled
out by (51). To prove pbq, together with the remaining claims, we consider first the case κ ‰ κ˘ ,
where ξ “ eiκ , eirpκq are different. Suppose, indirectly, dim V “ 2. Then, by taking a suitable
linear combination of solutions we could arrange for ψ p1q ‰ 0, ψ p2q “ 0 (or viceversa). This
would imply the contradiction
0 “ C0 pψ 7˚ , ψ 7 q “ lim Cn pψ 7˚ , ψ 7 q “ Cpψ p1q˚ , ψ p1q q ‰ 0

nÑ8
by ψ07 “ 0 and the preliminary remark (C4). Finally we are left with κ “ κ` (or κ´ ). This time
p1q p2q
dim V “ 2 would imply that there exist two solutions ψa7 , ψb7 P V : one with ψa “ ψ, ψa “ 0
p1q p2q
in Eq. (92) and the other with ψb “ 0, ψb “ ψ 1 . The resulting contradiction is
0 “ C0 pψb7˚ , ψa7 q “ Cpψ 1˚ , ψq ‰ 0
by (C5) and λ2 pκ` q ‰ 0. Moreover, the bounds Op1q and Opnq follow from Eqs. (40, 78) with
|ξ| “ |eiκ | “ 1.
Proof of Lemma 6.10. Let again k be fixed till further notice. We consider z in a complex
neighborhood of the upper band edge λpκ` q. There the expanded part of the solution (90) may
alternatively be written as a linear combination of
# ´
ψ pκq´ψ ´ prpκqq
, pκ ‰ κ` q
p1q ´ ´
ψ pκq “ ψ pκq ` ψ prpκqq , ψ pκq “ Bψ´ κ´rpκq
p2q
Bκ , pκ “ κ` q
where ψ ´ pκq is the section of Bloch solutions of Lemma 6.7 (ii), rather than as linear combination
of Bloch solutions ψ ´ pκq, ψ ´ prpκqq. The advantage of the basis tψ p1q pκq, ψ p2q pκqu is that it does
not degenerate as κ Ñ κ` . The solution (90) is unique up to a multiple, which we shall fix by
means of a prescription independent of κ. For instance, since ψ17 ‰ 0, there is a linear functional
40
` on CN such that `pψ17 pκqq ‰ 0 for κ near κ` . We impose the normalization `pψ17 pκqq “ 1. We
then have
ψn7 pκq “ αpκqψnp1q pκq ` βpκqψnp2q pκq ` Ope´ε0 n q , pn Ñ `8q (93)
with α, β analytic near κ` . Moreover, α, β are even under r since ψ 7 , ψ p1q , ψ p2q are. A semi-
bound state is tantamount to βpκ` q “ 0. For κ ‰ κ` we may also write
ψn7 pκq “ f pκqψn´ pκq ` f prpκqqψn´ prpκqq ` Ope´ε0 n q
where
βpκq
f pκq “ αpκq ` (94)
κ ´ rpκq
is analytic in a punctured neighborhood of κ` . Comparison with Eqs. (52, 53) yields for later
use
f prpκqq
S` pκq “ , pκ ą κ` q . (95)
f pκq
For Im κ ă 0 the Bloch solution ψn´ pκq is exponentially diverging for n Ñ 8 due to |ξ| “ e´ Im κ .
Based on z “ λpκq, edge state energies ε ą λpκ` q close to λpκ` q occur iff f pκq “ 0 for some κ
close to κ` with Im κ ă 0. For that to happen βpκq has to be correspondingly small:
βpκq “ αpκqprpκq ´ κq . (96)
Let us now reintroduce the parameter k of the lemma and of its assumption Eq. (54). Then
βpk˚ , κ` pk˚ qq “ 0 follows by continuity. This concludes the proof.
Proof of Theorem 6.11. Let us first deal with the simple case that the interval rk1 , k2 s does not
contain any k˚ where H 7 pk˚ q has a semi-bound state. Then N` vanishes by Lemma 6.10 and
so does the l.h.s. of Eq. (55). In fact, by (94, 95) we have
lim Spκ` pkq ` δ, kq “ ´1

δÑ0
uniformly in k P rk1 , k2 s, since βpκ` pkq, kq ‰ 0. We consider next the case where such points
k˚ are present. Though they are generically isolated, we will treat the general case, where they
form intervals I Ă rk1 , k2 s. Those are closed, countably many and possibly consisting of single
points. By the first case and by compactness it suffices to consider such an interval I and k1 , k2
sufficiently close to its endpoints.
Let first k P I. Then αpκ` pkq, kq ‰ 0, pk P Iq since βpκ` pkq, kq “ 0; see Eq. (93). We claim
there is a punctured disk of fixed radius centered at κ` pkq, which remains free of zeros κ of
f pκ, kq. Indeed, if that radius is small, Eq. (96) can not hold true there, because its r.h.s. is
linearly large in κ ´ κ` pkq Ñ 0, while the l.h.s. is quadratically small, as β is even. The claim
extends by continuity to k P rk1 , k2 s, provided the punctured disk is replaced by an annulus
Apkq of fixed radii.
We recall that by (51) real κ ‰ κ` pkq do not occur as zeros either. Let N pkq be the number
of zeros inside the annulus and having Im κ ă 0. Thus
´ N` “ N pk2 q ´ N pk1 q .
41
Eq. (55) then follows from the claim that
` 1˘
lim arg S` pκ` pki q ` δ, ki q “ ´2π N pki q ` , pi “ 1, 2q . (97)
δÑ0 2
In proving it we drop ki from the notation, e.g. A “ Apki q. We note that by assumption
βpκ` q ‰ 0, whence there is a disk D centered at κ` which is free of zeros of f . In particular,
κ` ` δ P D for small δ ą 0, and κ` ` ∆ P A for suitable ∆ ą δ; likewise for the images under
r, see Fig. 9.
A
κ`
κ` ` δ
C D κ` ` ∆
Figure 9: The disk D, the annulus A and the contour C in the κ-plane.
We consider the contour Λ Y Γ Y rpΛq Y γ, where Λ “ rκ` ` δ, κ` ` ∆s, Γ Ă A X tIm κ ă 0u

joins κ` ` ∆ to rpκ` ` ∆q, and γ Ă D X tIm κ ă 0u joins rpκ` ` δq to κ` ` δ. Let us denote
by C that contour with opposite (positive) orientation. By the argument principle we have
ż
Bf
f pκq´1 dκ “ 2πiN .
C Bκ
The contour may also be split into two parts, and their contributions computed otherwise. First,
by (95)
ż
Bf
arg Spκ` ` δq “ arg f prpκ` ` δqq ´ arg f pκ` ` δq “ Im f pκq´1 dκ .
ΛYΓYrpΛq Bκ
(Note that the first equality is consistent with arg being a continuous argument in k P rk1 , k2 s;
so is the second, because the path remains free of zeros as k changes, unlike γ). Second, by
Eq. (94) ż
Bf
lim f pκq´1 dκ “ iπ .
δÑ0 γ Bκ
Together, this proves (97) and hence the theorem for the upper band edge. The case of the
lower band edge is similar, except that incoming states are found at κ ă κ´ pkq. This explains
the reversed count of signs in N´ .
Proof of Proposition 6.12. We write k only when necessary. Since κ` is a non-degenerate

maximum, the energy curve z “ λpκq bijectively maps a neighborhood of κ` in the half-plane
Im κ ă 0 to one of the band edge λ` , however slit by the band itself. It will be understood that
z and κ are so related. We first consider the case of a disappearing branch and reformulate the
42
statement using the notation from the proofs of Thms. 6.11 and 5.4. Since no eigenvalue branch
is present at k2 ą k˚ , the two sides of Eq. (56) are, after dividing by 2π, those of
ż ż
1 1
f pκ, k1 q´1 df “ lpz, kq´1 dl , (98)
2πi C 2πi BD
where
• C is the contour in the κ-plane described in Fig. 9;
• BD Ă T is the contour encircling the crossing point with the Fermi line and described in
Fig. 4;
• lpz, kq is the eigenvalue of Lpz, kq described in Lemma 5.5 (ii). (We recall that the use
made of Lpz, kq rested on Eq. (30)).
Lpzq has an eigenvalue lpzq “ 0 iff f pκq “ 0 (with both zeros being of first order), since both
conditions are equivalent to z “ ε. Hence Eq. (98) holds provided the contours are homotopic
under z “ λpκq.
To show this, let us visualize BD Ă T in the setting of Fig. 1. That contour can be rotated
without intersecting the discrete eigenvalue branch, and so as to lie in a z-plane at fixed k “
k1 . The contour remains positively oriented w.r.t. the orientation of that plane and, after
substitution z “ λpκq, it is homotopic to C at k1 ă k˚ .
In the case of an emerging eigenvalue branch, the l.h.s. of (98) is evaluated at k2 ą k˚
and acquires a minus sign, cf. (56). However, when BD is rotated as just prescribed it ends up
negatively oriented in the z-plane at k “ k2 . Hence the modified Eq. (98) still holds.
Acknowledgments. We thank Y. Avron, C. Cedzich, M. Fraas, J. Fröhlich, H. Schulz-Baldes,

and D. Würsch for discussions. The work of M.P. is supported by the Swiss National Science
Foundation.
References
[1] Avila, J.C., Schulz-Baldes, H. and Villegas-Blas, C.: Topological invariants of edge states
for periodic two-dimensional models. arXiv:1202.0537
[2] Bernevig, B.A., Hughes, T.L. and Zhang, S.-C.: Quantum spin Hall effect and topological
phase transition in HgTe quantum wells. Science 314, 1757-1761 (2006).
[3] Bräunlich, G., Graf, G.M. and Ortelli, G.: Equivalence of topological and scattering ap-
proaches to quantum pumping. Commun. Math. Phys. 295, 243-259 (2010).
[4] Essin, A.M., Gurarie, V.: Bulk-boundary correspondence of topological insulators from their
Green’s functions. Phys. Rev. B 84, 125132 (2011).
[5] Fröhlich, J. and Kerler, T.: Universality in quantum Hall systems. Nucl. Phys. B 354,
369-417 (1991).
[6] Fröhlich, J. and Studer U.M.: Gauge invariance and current algebra in nonrelativistic many-
body theory, Rev. Mod. Phys 65, 733 (1993).
43
[7] Fröhlich, J., Studer U.M. and Thiran, E.: Quantum theory of large systems of non-relativistic
matter, Les Houches Lectures 1994, Elsevier (1995); arXiv:cond-mat/9508062.
[8] Fröhlich, J. and Zee A.: Large scale physics of the quantum Hall fluid. Nucl. Phys. B 364,
517-540 (1991).
[9] Fu, L. and Kane, C.L.: Time reversal polarization and a Z2 adiabatic spin pump. Phys. Rev.
B 74, 195312 (2006).
[10] Fujita M., Wakabayashi K., Nakada K. and Kusakabe K.: Peculiar localized state at zigzag
graphite edge. J. Phys. Soc. Jpn. 65, 1920-1923 (1996).
[11] Haldane, F.D.M: Model for a quantum Hall effect without Landau levels: Condensed-matter
realization of the “parity anomaly”. Phys. Rev. Lett. 61, 2015-2018 (1988).
[12] Hasan, M.Z. and Kane C.L.: Topological insulators. Rev. Mod. Phys. 82, 3045-3067 (2010).
[13] Hatsugai, Y.: Chern number and edge states in the integer quantum Hall effect. Phys. Rev.
Lett. 71, 3697 (1993).
[14] Hatsugai, Y. and Ryu, S.: Topological origin of zero-energy edge states in particle-hole
symmetric systems. Phys. Rev. Lett. 89, 077002 (2002).
[15] Hsieh, D., Qian, D., Wray, L., Xia, Y., Hor, Y.S., Cava, R.J. and Hasan, M.Z.: A topological
Dirac insulator in a quantum spin Hall phase. Nature 452, 970 (2008).
[16] Kane, C.L. and Mele, E.J: Z2 Topological order and the quantum spin Hall effect. Phys.
Rev. Lett. 95, 146802 (2005).
[17] Kato, T.: Perturbation Theory for Linear Operators. Springer-Verlag (1980).
[18] Kohn, W.: Analytic properties of Bloch waves and Wannier functions. Phys. Rev. 115,
809-821 (1959).
[19] König, M., Wiedmann, S., Brüne, C., Roth, A., Buhmann, H., Molenkamp, L.W., Qi, X.-L.
and Zhang, S.-C.: Quantum spin Hall insulator state in HgTe quantum wells. Science 318,
766 (2007).
[20] Moore, J.E. and Balents, L.: Topological invariants of time-reversal-invariant band struc-
tures. Phys. Rev. B 75, 121306(R) (2007).
[21] Nakada, K., Fujita, M., Dresselhaus, G. and Dresselhaus, M.S.: Edge state in graphene
ribbons: Nanometer size effect and edge shape dependence. Phys. Rev. B. 54, 17954 (1996).
[22] Nakahara, M.: Geometry, Topology and Physics. Graduate Student Series in Physics, In-
stitute of Physics Publishing (1990).
[23] Pfeffer, W.F.: More on involutions of a circle. Amer. Math. Monthly 81, 613 (1974).
[24] Prodan, E.: Robustness of the spin-Chern number. Phys. Rev. B 80 , 125327 (2009).
[25] Qi, X.-L., Wu, Y.-S. and Zhang, S.-C.: Topological quantization of the spin Hall effect in
two-dimensional paramagnetic semiconductors. Phys. Rev. B 74, 085308 (2006).
[26] Reed, M. and Simon, B.: Methods of Modern Mathematical Physics, III. Scattering Theory.
Academic Press (1979).
44
[27] Roy, R.: Z2 classification of quantum spin Hall systems: An approach using time-reversal
invariance. Phys. Rev. B 79, 195321 (2009).
[28] Schulz-Baldes, H., Kellendonk, J. and Richter, T.: Simultaneous quantization of edge and
bulk Hall conductivity. J. Phys. A: Math. Gen. 33 L27 (2000).
[29] Sheng, D.N., Weng, Z.Y., Sheng, L. and Haldane, F.D.M.: Quantum spin-Hall effect and
topologically invariant Chern numbers. Phys. Rev. Lett. 97, 036808 (2006).
[30] Thouless, D.J.: Quantisation of particle transport. Phys. Rev. B 27, 6083-6087 (1983).
[31] Wen, X. G.: Chiral Luttinger liquid and the edge excitations in the fractional quantum
Hall states. Phys. Rev. B 41, 12838–12844 (1990).
[32] Zhang, S.-C.: The Chern-Simons-Landau-Ginzburg theory of the fractional quantum Hall
effect. Int. J. Mod. Phys. B 6, 25-58 (1992).
45

Bulk-Edge Correspondence For Two-Dimensional Topological Insulators

Uploaded by

Copyright:

Available Formats

Bulk-Edge Correspondence For Two-Dimensional Topological Insulators

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Bulk-Edge Correspondence For Two-Dimensional Topological Insulators

Uploaded by

Copyright:

Available Formats

Bulk-edge correspondence for two-dimensional topological

March 20, 2013

2 Setting and results

2.1 The Schrödinger operators

where V 7 satisfies the above properties of V , as well as

Vn7 pkq “ Vn pkq , pn ą n0 q . (3)

Remark 2.3. The Dirichlet condition is by no means special. Since

pϕ, H 7 ψq ´ pH 7 ϕ, ψq “ ϕ˚1 Aψ0 ´ ϕ˚0 A˚ ψ1 ,

any boundary condition ψ0 “ Λψ1 , (Λpkq P MN pCq) defines a self-adjoint Hamiltonian if

Definition 2.4. [Time-reversal symmetry] The symmetry is a map Θ : CN Ñ CN with the

i) Θ is antilinear and Θ2 “ ´1;

iii) For all k P S 1 ,

2.2 The main result in brief

pHp´kq ´ z̄qΘψ “ ΘpHpkq ´ zqψ ,

ΛA :“ ~n “ n1~a1 ` n2~a2 | pn1 , n2 q P Z2 , ΛB :“ ΛA ` ~δ ,

The (Bravais) lattice of translations of Λ is ΛA . We reduce Λ to ΛA by dimerizing neighbors

after dropping primes. See Fig. 3.

~a2 ~a2 ~a1

where we set ψ “ ψpkq, n “ n1 and

Remark 3.2. ([1], Prop. 7) Suppose the Hamiltonian on `2 pZ; CN q

pHψqn “ Aψn´1 ` A˚ ψn`1 ` V ψn

(note: V independent of n) has a Bloch eigenvalue λpqq independent of the quasi-momentum q.

Proof. Let λ be that eigenvalue. Then detpAz ´1 ` A˚ z ` V ´ λq “ 0 for z “ eiq , (q P R) and,

3.2 The Kane-Mele model

4 Time-reversal invariant bundles

4.1 The index of endpoint degenerate families

dpZ 1 , Zq “ min max |zn ´ zn1 | ,

Definition 4.1. [Index] We set

4.2 The index of Kramers families of matrices

C : CN Ñ CN the complex conjugation, and Θ0 “ εC the standard time-reversal on CN .

Definition 4.3. [Index] We set

M´ p0qΘ0 “ Θ0 M` p0q , M´ pπqΘ0 “ Θ0 M` pπq . (17)

4.3 The index of time-reversal invariant bundles

• right multiplication by M P GLpN q:

M : F pEqϕ Ñ F pEqϕ , v ÞÑ vM , (20)

• Θ : F pEqϕ Ñ F pEqτ ϕ , v ÞÑ Θv, with pΘvqi “ Θvi . Note that

ΘpvM q “ pΘvqM . (21)

vpτ ϕq “ Θvpϕqε . (22)

v` pϕ2 q “ v´ pϕ2 qT pϕ2 q , pϕ2 P S 1 q . (24)

Lemma 4.6. [Time-reversal symmetry of the transition functions]

5 The bulk-edge correspondence

E “ tppz, kq, ψq P T ˆ C | ψ P Ez,k u (26)

determined by ψn7 “ ψn for n ą n0 .

det Ψ71 ‰ 0 . (30)

ii) Let Ψpz, kq P F pEqz,k be a section defined in a neighborhood in C ˆ S 1 Ą T of any of the

iii) At any of the points pµ, k˚ q we have

iv) As k increases past k˚ the eigenvalue εpkq crosses µ as an increasing function if

9 where the cut is

T pkq “ M´ pkqM` pkq´1 , pk P Iq . (36)

p 7 , which is as required there. Then Lpz, kq “

T pkq “ Apkq´1 L´ pkqL` pkq´1 Apkq , pk P Iq

Lpz, kq “ lpz, kqΠpz, kq ‘ Lpz,

Lpz, kq “ lpz, kqΠ0 ` L r 1 ¨ ∆ ` Op∆2 q ,

In view of L´ pkq “ Lpz̄, kq for pz, kq P pBDq` we compute

In the second line we used L

5.1 The bulk index as an index of Bloch bundles

pHψqn “ Aψn´1 ` A˚ ψn`1 ` Vn ψn , pn P Z, ψn P CN q (38)

and of energy z is then represented as a solution Ψ of

HpξqΨ ” pAξ ´1 ` A˚ ξ ` VqΨ “ zΨ , pΨ P CM N q . (41)