Sabri Youssef

RESEARCH ARTICLE | SEPTEMBER 01 2023
Flat bands of periodic graphs 

Special Collection: New Directions in Disordered Systems: In Honor of Abel Klein
Mostafa Sabri  ; Pierre Youssef
J. Math. Phys. 64, 092101 (2023)

https://doi.org/10.1063/5.0156336
 
View Export
Online Citation
Articles You May Be Interested In
Spectral gaps of frustration-free spin systems with boundary

J. Math. Phys. (May 2019)
More light on the 2ν5 Raman overtone of SF6: Can a weak anisotropic spectrum be due to a strong
transition anisotropy?
J. Chem. Phys. (January 2014)
Analytical techniques for linear response formula of equilibrium states

Chaos (January 2020)
15 October 2024 06:43:54

Journal of
ARTICLE pubs.aip.org/aip/jmp
Mathematical Physics
Flat bands of periodic graphs

Cite as: J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336
Submitted: 28 April 2023 • Accepted: 12 August 2023 •
Published Online: 1 September 2023
Mostafa Sabri1 , 2,a) and Pierre Youssef2 , 3,b)
AFFILIATIONS
1
Department of Mathematics, Faculty of Science, Cairo University, Giza 12613, Egypt
2
Science Division, New York University Abu Dhabi, Saadiyat Island, Abu Dhabi, United Arab Emirates
3
Courant Institute of Mathematical Sciences, New York University, 251 Mercer St., New York, New York 10012, USA
Note: Paper published as part of the Special Topic on New Directions in Disordered Systems: In Honor of Abel Klein.
a)
Author to whom correspondence should be addressed: mostafa.sabri@nyu.edu
b)
E-mail: yp27@nyu.edu
ABSTRACT
We study flat bands of periodic graphs in a Euclidean space. These are infinitely degenerate eigenvalues of the corresponding adjacency matrix,
with eigenvectors of compact support. We provide some optimal recipes to generate desired bands and some sufficient conditions for a graph
to have flat bands, we characterize the set of flat bands whose eigenvectors occupy a single cell, and we compute the list of such bands for small
15 October 2024 06:43:54

cells. We next discuss the stability and rarity of flat bands in special cases. Additional folklore results are proved, and many questions are still
open.
Published under an exclusive license by AIP Publishing. https://doi.org/10.1063/5.0156336
I. INTRODUCTION
Consider a connected, locally finite graph Γ, which is invariant under translation by some linearly independent vectors a1 , . . . , ad (we say
that Γ is Zd -periodic). Using Floquet theory (Sec. II), one sees that σ( AΓ ), the spectrum of the 0/1 adjacency matrix of Γ, consists of bands.
At least one of these bands will be non-degenerate (absolutely continuous), but curiously, degenerate bands do occur here, in contrast to the
classical periodic Schrödinger operators in Rd . These degenerate bands are infinitely degenerate eigenvalues for AΓ , and they always have
corresponding eigenvectors of compact support. See Fig. 1 for an example. They are called flat bands for Γ, the terminology coming from the
fact that the corresponding Floquet eigenvalue Ej (θ) is constant in the quasimomentum θ and thus traces a single point as θ varies in [ 0, 1 )d ,
instead of a nontrivial interval [a, b]. The graph of θ ↦ Ej (θ) is thus completely flat; see, e.g., Ref. 23, Fig. 6 for an illustration.
There has been intensive activity in the physics community in recent years regarding these flat bands, as they have found applications in
the contexts of superfluidity, topological phases of matter, and many-body physics; see Refs. 9, 20, 22, and 23 and references therein.
Mathematically, the topic has been much less explored. The basic problem that one would want to consider is the following: “Given a
connected periodic Γ, by looking at the graph, decide whether or not Γ has a flat band.”
More precisely, let V f be the vertices of the motif, which is copied by integer translations. The translated motifs should not intersect one
another and should span Γ. Let
ν = ∣Vf ∣,
e.g., ν = 2 in Fig. 1. Several choices of V f are possible; fix one. For vi ∈ V f ⊂ Γ, let N vi be the set of neighbors of vi in the full graph Γ.
Problem 1. For each fixed ν, find all possible choices of N vi and d such that Γ has a flat band. If possible, find the corresponding values
of the flat band λ.
This problem is completely open, and it may be too ambitious as stated. Note that vi can hop to large lattice distances, at least from a
mathematical standpoint. We reformulate this question algebraically in Sec. I B.
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-1

Published under an exclusive license by AIP Publishing
Journal of
FIG. 1. Flat bands λ = −1 (left) and λ = 0 (right). The values of eigenvectors are given (they are extended by zero to the infinite Γ).
A. Main results
In this paper, we take some small steps toward understanding the emergence of flat bands. In Sec. II, we gather and prove some folklore
results. In Sec. III, we provide recipes to generate graphs with or without flat bands. More precisely, we give a construction in Theorem 3.1
to generate flat bands from eigenvalues of finite graphs, and we identify corresponding eigenvectors. Our generator is shown to be optimal in
the size of V f . In the special case where the eigenvalue comes from a regular graph, we provide a more efficient generator. On the other hand,
we give simple operations to construct new graphs from old ones, which preserve the lack of flat bands in Sec. III B.
Next, we characterize the set of flat bands for which there exist eigenvectors localized within a fundamental cell in Theorem 4.6. These
types of flat bands have been known to be more tractable in the physics community.9 We apply our criterion to deduce further structural results
about these bands and also to compute explicitly the set F sν of such flat bands for ν ≤ 5 in Sec. IV (the superscript s stands for “single-cell”).
In doing so, we note a curious property about the growth of F sν with ν and formulate an open question.
Back to the main problem (problem 1) of characterizing flat bands, in general, we show in Sec. VI that if Γ is invariant under some
symmetries, then it has flat bands. This approach appeared originally in Ref. 23, and we give a more general and precise statement using a
different proof. These symmetry conditions are not necessary, however.
The converse, namely, finding sufficient conditions for the absence of flat bands, seems a lot more challenging. We have the following.
Fact. If ν = 1 and d is arbitrary, then the spectrum consists of a single band of an absolutely continuous spectrum.
Indeed, A is then equivalent to multiplication by E(θ) = ∑Dp=1 2 cos (2πθ ⋅ n(p) ) in L2 [ 0, 1 )d for some n(p) ∈ Nd /{0}. As E(θ) is analytic
and nonconstant, the claim follows.
15 October 2024 06:43:54

Even the simple case ν = 2 is not completely understood, however. We investigate it in more detail and give some partial results in Sec. V.
For ν > 1, the only general result we know concerning the absence of flat bands is the following one of Ref. 10.
Fact. Fix a finite multigraph G. If G has a 2-factor, then there is a recipe to create a Z-periodic graph Γ with quotient graph G such that Γ
has no flat bands. In particular, if G has a 2-factor, then the maximal Abelian cover of G has no flat bands.
Recall that a graph G is called Hamiltonian if there is a cycle in G, which covers all vertices of G. The condition that G has a 2-factor is a
generalization, which means that there is a collection of disjoint cycles, which cover all vertices of G. This condition is satisfied, in particular,
if G is even-regular or bipartite odd-regular, and the authors of Ref. 10 conjectured that the maximal Abelian cover of any regular multigraph
should have no flat bands.
In the preceding discussion, we were only interested in the adjacency matrix of Γ, so there are no weights on the edges and no potentials.
This is in the hope of getting a completely geometric characterization, and this restriction makes some aspects of the mathematics more
interesting. We do not expect the questions to be any different for the variant D − AΓ (the discrete Laplacian, where D is the degree matrix).
Physicists do endow their graphs with periodic weights. The Floquet theory is still applicable, and it makes sense to study flat bands of AΓ + Q,
for instance, for a periodic potential Q. Looking at the physics literature, it is apparent that flat bands are “rare” and that they are a “fine tuning”
phenomenon, but we only found examples rather than theorems in this respect. This motivates the following problem.
Problem 2. Show that for any periodic connected Γ, the set

ν
PΓ = {(Q1 , . . . , Qν ) ∈ R : AΓ + Q has a flat band} (1.1)
has dimension at most ν − 1.
We show in Sec. VII that PΓ is a closed semialgebraic set. An answer to problem 2 would imply that PΓ has Lebesgue measure zero,
reflecting the rarity of flat bands, and is nowhere dense, reflecting the fine tuning (it would contain no balls: given AΓ + Q, there would exist
Q′ arbitrarily close to Q such that AΓ + Q′ has no flat bands). Showing these weaker properties would already be interesting for physical
purposes.
We also show in Sec. VII that for any ν, there exists some Γ with dim ( PΓ ) ≥ ν − 1, so the conjectured bound on the dimension cannot
be improved, in general. Moreover, we answer problem 2 completely for ν = 2; it remains open for ν ≥ 3.
For the reader’s convenience, we summarize the results of this paper:
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-2

Journal of
(1) For any strictly positive periodic weight w on the edges and any periodic potential Q, the top band of H = Aw + Q is never flat. There
are examples of complex Hermitian weights w such that the spectrum of H is entirely flat.
H has a localized eigenvector iff the Floquet matrix H(θ) has an eigenvector, which is a trigonometric polynomial in each entry.
There is a constructive recipe to derive one from the other (Sec. II).
(2) If Γ is invariant under a permutation of V f consisting of r cycles, which exchanges stars around vertices, then Γ has at least ν − r flat
bands. In particular, if N vi /{v j } = N v j /{vi } for some vi , vj ∈ V f , then Γ has a flat band λ ∈ {0, −1} (Sec. VI).
(3) We have a recipe to make all eigenvalues of any finite graph G appear as flat bands such that the eigenvectors of AΓ are entirely
supported in V f , with ∣V f ∣ = 2∣G∣. We call these single-cell flat bands. The recipe is optimal in the size of V f . We also have recipes to
create graphs with arbitrary d, ν, which have no flat bands (Sec. III).
(4) We characterize the set F sν of single-cell flat bands and apply our criterion to compute F sν explicitly for ν ≤ 5. In general, F sν ⊆ F sν+1 ,
we have a description of F sν+1 / F sν , and ∪ν F sν is the set of all totally real algebraic integers (Sec. IV).
(5) Flat bands for ν = 2 are necessarily integers. They are not necessarily in F s2 . A sufficient condition is given for the absence of flat bands
(Sec. V).
(6) We answer problem 2 for ν = 2. For general ν, we show that the set PΓ in (1.1) is closed, semialgebraic and can satisfy dim PΓ ≥ ν − 1
for certain Γ. A stronger form of problem 2 was conjectured in Ref. 15; we show that version to be untrue (Sec. VII).
Point (1) is essentially folklore, while point (2) is inspired from Ref. 23; in both cases, we provide more precise and general statements
with simpler proofs. Flat band generators occupy a large part of physics literature; the one in (3), though very simple, is to our knowledge
new. Points (4)–(6) are new. We recall here that our main interest is in the adjacency matrix, so in all the previous results, except (1) and (6),
we assume that there is no nontrivial potential or edge weight. We have not investigated the effect of potential in points (2)–(5).
We refer to Refs. 13, 14, and 22 for further results, in particular concerning spectral gaps and singularities in band crossings. Flat bands
have also been investigated in the special framework of Archimedean tilings of the plane in Ref. 21. There is a total of 11 such tilings, with
explicit Floquet matrices. By analyzing each of them, the authors show that only two of them have flat bands. The question of perturbation by
edge weights was later studied in Ref. 12, and it was found that for flat bands to survive, some explicit relations between the weights must be
satisfied. This is in accord with problem 2. We also mention the recent paper6 where compactly supported eigenfunctions have been studied
in the Penrose and Amman–Beenker aperiodic tilings of the plane.
To conclude, let us mention that “flat bands” also arise in the context of universal covers of finite undirected graphs. The spectra of
15 October 2024 06:43:54

these trees T have a quite similar structure to those of the present Zd -periodic graphs, though the mathematical tools to analyze them are
different. (Floquet theory is not applicable.) It is shown in Ref. 1, Sec. 4.3 that the spectrum of T consists of bands, some of which may be
degenerate, but that the absolutely continuous spectrum always exists if min deg T ≥ 2. The authors of Refs. 3 and 25 characterized the flat
bands, completing earlier work of Ref. 2, and 3 answered a variant of problem 2 in that setting.
B. Algebraic formulation
The question of flat band existence can be reduced to the following. Let hi j (z) = ∑k ∈Ii j zk be Laurent polynomials on Cd /{0} for 1 ≤ i, j
≤ ν, where Ii j ⊂ Zd is finite and zk = z1k1 ⋅ ⋅ ⋅ zdkd . We assume hji (z) = hij (z−1 ), so I ji = −I ij . We assume 0 ∉ I ii for all i. Some I ij can be empty.
Now, consider the ν × ν matrix A(z) = (hij (z)). What are all the possible choices of I ij and λ ∈ R such that the characteristic polynomial
pλ (z) = det(A(z) − λ) vanishes identically as a Laurent polynomial? This is essentially problem 1.
For example, if ν = 2, this becomes a question of factorization. Namely, what are all the choices of I ij and λ such that (h11 (z) − λ)(h22 (z)
− λ) = h12 (z)h12 (z−1 )? In the special case h12 (z−1 ) = h12 (z), the question is when can we have Pλ (z)Qλ (z) = H(z)2 for Laurent polynomials
whose zk coefficients are 0 or 1 for k ≠ 0. An obvious choice is Pλ = Qλ = H, but there are other choices (see Sec. V B), yet we do not know all
of them.
Technically, not all choices of I ij and λ will answer the question of flat band existence: one also needs to assume that the periodic graph
generated by A(z) is connected (see Sec. II for the connection to Γ). For example, while (z2 + z−2 + 2)(z4 + z−4 + 2) = (z3 + z−3 + z + z−1 )2 is
a valid factorization, the corresponding Γ is disconnected, so this choice of I ij and λ has to be disregarded.
Another concept of algebraic flavor related to periodic operators is that of irreducibility. The question is whether the characteristic
polynomial pλ (z) is irreducible as a Laurent polynomial, either for fixed λ (Fermi irreducibility) or as a polynomial in both λ and z (Bloch
irreducibility). In the case of Bloch irreducibility, there are no flat bands [if λ = c was flat, then pλ (z) = (λ − c)qλ (z) would be a nontrivial
factorization], but irreducibility is strictly stronger. So far, irreducibility has been established for graphs with ν = 1, such as Zd endowed with
periodic potentials,7,18 and for planar graphs with ν = 2.17 The number of irreducible components for more general graphs with ν > 1 has
recently been investigated in Ref. 8.
C. Index of notation
Due to there being quite a lot of notations used in different places throughout this article, we summarize them in Table I and introduce
them again as needed.
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-3

Journal of
TABLE I. Index of symbols.
Γ An infinite Zd -periodic graph for some d

Vf A fixed fundamental cell for Γ
ν Cardinality of V f
vi A generic vertex of V f , i = 1, . . . , ν
vi + ka A generic vertex of Γ, i = 1, . . . , ν, k ∈ Zd
Qi The values of the periodic potential, i = 1, . . . , ν
I ij Subsets of Zd encoding the hopping: k ∈ Ii j ⇔ v j + ka ∼ vi
hi Maximal hopping in ith coordinate
H(θ) The Floquet matrix H(θ) = (hij (θ))
hij (θ) The Floquet matrix element hi j (θ) = δi j Qi + ∑k ∈Ii j wi j (k)e2πiθ⋅k
A(θ) H(θ) with Qi ≡ 0 and wij ≡ 1, so hi j (θ) = ∑k ∈Ii j e2πiθ⋅k
p(θ; λ) The characteristic polynomial det(H(θ) − λI)
AG The adjacency matrix of the graph G = Γ, Vf , . . .
ϵij The matrix elements of AVf
bij (θ) bij (θ) = hij (θ) − ϵij
B(θ) The Floquet matrix of “external bridges,” B(θ) = (bij (θ))
zk zk = z1k1 ⋅ ⋅ ⋅ zdkd with z j = e2πiθ j
F sν s
F ν = {λ ∈ R : λ single−cell flat band for some (Γ, Vf ), ∣Vf ∣ = ν}
PΓ PΓ = {Q ∈ Rν : AΓ + Q has a flat band}
II. GENERALITIES
Let Γ be an infinite graph in some Euclidean space, which is locally finite and has no self-loops. We assume that Γ is invariant under
translation by linearly independent vectors a1 , . . . , ad . The vertex set V = V(Γ) satisfies
15 October 2024 06:43:54

V = Vf + Zda (2.1)
for some finite vertex set V f = {v1 , . . . , vν } of a motif, which is copied periodically under translations by ka ∶= ∑di=1 ki ai ∈ Zda , where
Zda = {ka : k ∈ Zd }.
Endow Γ with periodic potential Q(vp + ka ) = Q(vp ) and edge weights w(vp , vq + ka ) = w(vp − ka , vq ) for vq + ka ∼ vp . We consider the
Schrödinger operator
H = Aw + Q
on Γ, where ( Aw ψ)(vi + ra ) = ∑u∼vi +ra w(vi + ra , u)ψ(u)for vi + ra ∈ Vf + Zda . Let
I i j = {v j + ka : v j + ka ∼ vi }, Ii j = {k : v j + ka ∈ I i j }. (2.2)
I ij dictate the hopping (if k ∈ I ij , then there is an edge from V f to its kth translate, k ∈ Zd ). Note that 0 ∉ I ii since Γ has no self-loops. We
have
ν
N vi = ∪ j=1 I i j
and ( Aw ψ)(vi + ra ) = ∑νj=1 ∑k ∈Ii j w(vi + ra , v j + ra + ka )ψ(v j + ra + ka ).
We henceforth identify ℓ2 (Γ) ≡ ℓ2 (Zd )ν via ψ(vp + ka ) ↦ ψp (k) for p = 1, . . . , ν and k ∈ Zd . We similarly map Q(vp ) ↦ Qp and
w(vp , v j + ra ) ↦ wp j (r). Then, H is now regarded as operating on vector functions ψ ∈ ℓ2 (Zd )ν by
ν
( Hψ)i (r) = ∑ ∑ wi j (k)ψ j (r + k) + Qi ψi (r). (2.3)
j=1 k∈Ii j
1
For example, in Fig. 1 (left), we have I 11 = I 22 = {±1}, I 12 = {0, ±1} = I 21 , w ≡ 1, and Q ≡ 0, and the depicted eigenvector is ψ(0) = ( −1 ),
0
ψ(k) = ( 0 ) for k ∈ Z/{0}.
Much like the Fourier transform is used to check that AZd is unitarily equivalent to M f , the multiplication operator by
f (θ) = ∑dj=1 2 cos 2πθ j on L2 (Td∗ ), where Td∗ ∶= [ 0, 1 )d , the Floquet transform U : ℓ2 (Zd )ν → L2 (Td∗ )ν defined by
(Uψ)p (θ) = ∑ e−2πiθ⋅k ψp (k) (2.4)

k∈Zd
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-4

Journal of
for p = 1, . . . , ν satisfies the following.

Lemma 2.1. The operator U is unitary and
U HU −1 = MH , (2.5)
where (M H f )(θ) = H(θ) f (θ) is the multiplication operator on L2 (Td∗ )ν by the matrix
H(θ) = (hi j (θ))νi, j=1 , (2.6)

2πiθ⋅k 2πiθ⋅k
hii (θ) = Qi + ∑ wii (k)e , hi j (θ) ∶= ∑ wi j (k)e , i ≠ j.
k∈Iii k∈Ii j
Here, 0 ∉ I ii as Γ has no loops. If Q ≡ 0 and w ≡ 1, we denote A(θ) = H(θ).
This is quite standard,4,15 though our language is simpler and more general.
Proof. Being the Fourier transform in each coordinate, U is clearly unitary. The pre-image of f (θ) is ϕp (k) = ∫Td fp (θ)e2πik⋅θ dθ. Next
∗
given ψ ∈ ℓ2 (Zd )ν ,
⎛ ν ⎞
(U Hψ)p (θ) = ∑ e−2πiθ⋅k ∑ ∑ wp j (r)ψ j (k + r) + Qp ψp (k)
k∈Zd
⎝ j=1 r∈Ip j ⎠
ν
= ∑ ∑ wp j (r)e2πiθ⋅r ∑ e−2πiθ⋅(k+r) ψ j (k + r) + Qp (Uψ)p (θ)
j=1 r∈Ip j k∈Zd
ν
= ∑ hp j (θ)(Uψ) j (θ) = [H(θ)(Uψ)(θ)]p.
j=1
15 October 2024 06:43:54

◻
We next note that the map τ : I i j → I ji given by v j + ka ↦ vi − ka is a bijection. In fact, by translation invariance, v j + ka ∼ vi ⇐⇒ v j
∼ vi − ka . This implies the symmetries
Iii = −Iii , I ji = −Iij (2.7)
as index sets. Note, however, that I ij can be distinct from −I ij for i ≠ j. Assuming that w satisfies w ji (−k) = wij (k), we get by (2.7) that the
matrix H(θ) is Hermitian,
h ji (θ) = hij (θ).
Lemma 2.2. There exist ν continuous functions E1 (θ) ≤ ⋅ ⋅ ⋅ ≤ Eν (θ) on Td∗ consisting of the eigenvalues of H(θ).
There exists a real analytic variety X ⊂ Td∗ of dimension ≤ d − 1 (so X is a closed nullset) such that the eigenvalues Ej (θ) of H(θ) are
analytic on Td∗ /X. Each eigenvalue has constant multiplicity on each of the finitely many connected components of Td∗ /X.
If d = 1, the eigenvalues and eigenvectors can be chosen to be analytic over all T∗ . This is Rellich’s theorem; it holds, more generally,
when H(θ) has a compact resolvent, see Ref. 11, Theorem 3.9, p. 392, as long as H(θ) is self-adjoint and θ ∈ T∗ is a real parameter.
Proof. The continuity of Ej is shown in Ref. 11, pp. 106–109. The argument there is for d = 1, but holds without change for any d. In
steps, one shows that the resolvent is continuous in θ and then deduces the continuity of the spectral projection near the distinct Ej (θ0 ), from
which one gets that H(θ) has m eigenvalues Ej,k (θ) close to Ej (θ0 ) counting multiplicity, if Ej (θ0 ) has multiplicity m and θ is close to θ0 .
Analyticity cannot be directly deduced from the 1d arguments in Ref. 11. Still, the proof of our statement is the same as Ref. 27,
Lemmas 4.5–4.7, which consider periodic Schrödinger operators in R3 . In our case, the argument is simpler as we can define directly the
corresponding set E = {(θ, λ) ∈ Rd+1 : p(θ, λ) = 0}, where p(θ, λ) = det(H(θ) − λI) is the characteristic polynomial, which is analytic, being
a polynomial of the analytic entries of H(θ) − λI. We also fix n ∶= ν in Lemmas 4.6–4.7 of Ref. 27. ◻
It follows from (2.5) and the continuity of Ej (θ) that
σ( H) = ∪νj=1 σ j ,
where σ j = Ran Ej (θ) are bands, as M H is equivalent to M D , with D(θ) = Diag(Ej (θ)). The operator H has no singular continuous spectrum
(Ref. 10, Proposition 4.5). One can also see that from the analyticity of Ej (θ) on Td∗ /X, because all partial derivatives ∂θk E j (θ) are analytic as
well, hence either Ej (θ) is constant and σ j is reduced to a point, yielding a point spectrum, or ∇θ Ej (θ) ≠ 0 a.e. and σ j consists of an absolutely
continuous spectrum.
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-5

Journal of
We say that λ ∈ R is a flat band of H if λ is an eigenvalue of H.
Lemma 2.3. λ is a flat band of H iff λ is an eigenvalue of H(θ) for all θ ∈ Td∗ .
Proof. By (2.5), λ is an eigenvalue of H iff λ is an eigenvalue of H(θ) for all θ in a set Ω ⊂ Td∗ of positive measure. Indeed, if M H has
an eigenvalue λ, there is a corresponding eigenvector f ∈ L2 (Td∗ )ν such that M H f = λf , i.e., H(θ) f (θ) = λf (θ) for a.e. θ ∈ Td∗ . Conversely,
if there is some f (θ) such that H(θ) f (θ) = λf (θ) for all θ ∈ Ω of positive measure, then M H f 1Ω = λf 1Ω , so λ is an eigenvalue of M H with
eigenvector f 1Ω .
Thus, λ is an eigenvalue of H iff p(θ; λ) ∶= det(H(θ) − λI) = 0 for all θ ∈ Ω. As p(⋅; λ) is analytic on Td∗ , then p(⋅; λ) ≡ 0 on Ω iff it
vanishes on all Td∗ ; see, e.g., Ref. 10, Proposition 4.3. ◻
Define the maximal hopping range in the ith coordinate by
hi = max (ki : k ∈ ∪p, j≤ν Ip j ).
If hi = 1 for all 1 ≤ i ≤ d, we speak of nearest-neighbor hopping.
Theorem 2.4. If H has a flat band, then it has a corresponding eigenvector ψ on Γ of compact support. We may choose ψ p to be supported
inside ×di=1 [[−(ν − 1)hi , (ν − 1)hi ]] for each p = 1, . . . , ν.
This result should be contrasted with the one of Ref. 19, which shows that the wavefunctions corresponding to the absolutely continuous
bands are very delocalized. See also Ref. 4 and references therein for connections with transport. Theorem 2.4 appeared in Ref. 10, Theorem 3.2.
We give a different proof, which is constructive and is of practical interest, as we illustrate on some graphs. This also yields the bound on the
support, which seems new but is probably not sharp. The proof is based on two results.
Lemma 2.5. If H has a flat band, then H(θ) has a corresponding eigenvector f (θ) whose entries f p (θ) are trigonometric polynomials of
degree at most (ν − 1)hi in each θi .
15 October 2024 06:43:54

A statement similar to the first part, with a different derivation, appears in Ref. 22.
Proof. We may assume the flat band is λ0 = 0 by translating H by λ0 I. Say, 0 has multiplicity m for H(θ). Then, pθ (λ) ∶= det(H(θ)
− λI) = λm qθ (λ) for some polynomial qθ (λ) of degree ν − m in λ. Note that qθ (H(θ)) ≠ 0 since 0 ≠ qθ (0) ∈ σ(qθ (H(θ))).
By the Cayley–Hamilton theorem, H(θ)m qθ (H(θ)) = 0, so there exists 0 ≤ n ≤ m − 1 such that H(θ)n qθ (H(θ)) ≠ 0 and
H(θ)n+1 qθ (H(θ)) = 0. Hence, there is a nonzero column f (θ) in H(θ)n qθ (H(θ)) lying in the kernel of H(θ).
Finally, if pθ (λ) = ∑νr=0 cr (θ)λr , then cr (θ) = 0 for r < m and qθ (λ) = ∑νr=m cr (θ)λr−m . Here, cr (θ) consists of minors of H(θ) of size
ν − r. Since each entry of H(θ) is a trigonometric polynomial of degree at most hi in θi , then each entry of H(θ)n qθ (H(θ)) is of degree at
most (n + ν − m)hi in each θi . ◻
Proposition 2.6. If H(θ) has an eigenvector f (θ) of entries fp (θ) = ∑m ∈Λp αp (m)e−2πim⋅θ for some finite Λp ⊂ Zd , p = 1, . . . , ν, then
defining ψ p (k) = αp (k) if k ∈ Λp and ψ p (k) = 0 otherwise, ψ is a localized eigenvector for H.
Conversely, if H has an eigenvector ψ of finite support Λp in each entry p = 1, . . . , ν and we let fp (θ) = ∑k ∈Λp e−2πiθ⋅k ψp (k), then f (θ) is
an eigenvector for H(θ).
In particular, H has an eigenvector localized in V f iff H(θ) has an eigenvector independent of θ.
See the graphs in Figs. 10 and 11 for an illustration of this construction.
Proof. Let ψp (k) = ∫Td fp (θ)e2πik⋅θ dθ, the pre-image of f (θ) under U. Then, (U Hψ)(θ) = H(θ)(Uψ)(θ) = H(θ) f (θ) = λ f (θ)
∗
= λ(Uψ)(θ), so ψ is an eigenvector of H since U is unitary. The first claim follows since ∫Td e2πi(k−m)⋅θ dθ = δk,m .
∗
Next, for ψ in the converse statement, (2.4) becomes (Uψ)p (θ) = ∑k ∈Zd e−2πiθ⋅k ψp (k) = fp (θ). Moreover, (Uψ)(θ) is an eigenvector of
H(θ) since H(θ)(Uψ)(θ) = (U Hψ)(θ) = λ(Uψ)(θ), where we used Hψ = λψ.
The special case follows by taking Λp = {0} for each p. ◻
Proof of Theorem 2.4. This now follows from Lemma 2.5 and Proposition 2.6. ◻
Note that if H has an eigenvector of compact support, then it can be translated under the action of Zd to produce infinitely many such
vectors, which are mutually orthogonal. Hence, a flat band has infinite multiplicity for H.
Theorem 2.7. If the weight w is symmetric, wji (−k) = wij (k), and strictly positive, then the top band of H = Aw + Q is never flat; it consists
of an absolutely continuous spectrum.
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-6

Journal of
FIG. 2. Each horizontal edge carries symmetric weight ±i, with the sign as shown (in red), while the diagonal edges carry weight 1. Two localized eigenfunctions are shown
in black and blue, corresponding to λ = 2 and λ = −2, respectively. Here, σ( H) = {±2}.
1
To our knowledge, this result was only proved rigorously in Ref. 16, Theorem 2.1 for w(u, v) = √
deg (u) deg (v)
, and also the authors of
Ref. 16 provided extra spectral information as part of their proof. We give a direct argument, which is a lot simpler and works for arbitrary
w > 0.
Proof. Suppose on the contrary that the top band is an eigenvalue λ.
Step 1. By Theorem 2.4, we may find a corresponding eigenvector ψ of finite support. Let S = supp ψ, S1 = {w ∉ S : w ∼ v, v ∈ S}, and
S2 = {w ∉ (S ∪ S1 ) : w ∼ v, v ∈ S1 } be the vertices at distances 1 and 2 from S. Let B = S ∪ S1 ∪ S2 . Since Γ is connected, for any v, v′ ∈ B,
we may find a path Pv,v′ connecting v to v′ . We take Pv,v′ to be the shortest one and fix one if there are many. Let C = B ∪ {Pv,v′ }v,v′ ∈B .
Then, the finite graph C is connected. Let ̃ ψ be the restriction of ψ to C and H C be the restriction of H to C (with Dirichlet conditions).
Then, 0 = ( H − λ)ψ(v) = (HC − λ)̃ ψ (v) if v ∈ S ∪ S1 and 0 = (HC − λ)ψ̃(v) for v ∈ C/(S ∪ S1 ) by definition of S, S1 .
Step 2. We may assume that Q has positive entries up to shifting H by cI. Then, the matrix H C is irreducible since C is connected, so it has a
Perron–Frobenius eigenvector ϕ with ϕ(v) > 0 on C. Here, λ is the top eigenvalue of H C since λ ≤ ∥HC ∥ ≤ ∥ H∥ = λ, so λ = ∥H C ∥. Since λ
is simple for H C and (HC − λ)ψ ̃ = 0, we get ̃ ψ = ϕ.
We showed that ψ(v) has only non-negative entries. Finally, take v ∈ S1 . Then, 0 = λψ(v) = Hψ(v), implying that ψ(w) = 0 for all w ∼ v.
Progressively, as Γ is connected, we deduce that ψ ≡ 0, a contradiction. ◻
Theorem 2.7 is no longer true if the edges of Γ are allowed to carry Hermitian weights: the spectrum can be entirely flat in that case, as
15 October 2024 06:43:54

the example taken from Ref. 5 in Fig. 2 shows. The complex weights are interpreted as magnetic fields.
Let p(θ; λ) = det(H(θ) − λI). We see from (2.6) that if z j ∶= e2πiθ j , then p(z; λ) is a Laurent polynomial in z and polynomial in λ. If λ is
a flat band, then p(⋅; λ) vanishes identically on the sphere Sd = {(z1 , . . . , zd ) ∈ Cd : ∣z1 ∣ = ∣z2 ∣ = ⋅ ⋅ ⋅ = ∣zd ∣ = 1} by Lemma 2.3. This implies the
following fact that we frequently use.
Lemma 2.8. If λ0 is a flat band for H, then the characteristic polynomial p(z; λ0 ) is the zero Laurent polynomial in z.
In particular, each coefficient of zk must vanish in this case. See also Remark 7.3.
Proof. It suffices to show that if p(z) is a Laurent polynomial vanishing identically on Sd , then p(z) ≡ 0 on Cd /{0}. Take Q ∈ Nd large
so that q(z) = zQ p(z) is a polynomial in z. We show that q(z) vanishes for all z ∈ Cd .
If d = 1, then q(z) is identically zero on C, as q is holomorphic and q ≡ 0 on S.
Suppose that the claim holds in dimension d − 1 and fix zd ∈ S. Then, (z1 , . . . , zd−1 ) ↦ q(z1 , . . . , zd ) is a polynomial in
d − 1 variables vanishing on Sd−1 , so it vanishes for all (z1 , . . . , zd−1 ) ∈ Cd−1 . Say, q(z) = ∑k∈Λ αk zk for some Λ ⊂ Nd . Since q(z)
kd−1
= ∑ β(k1 ,...,kd−1 ) (zd )z1k1 ⋅ ⋅ ⋅ zd−1 is trivial, where β(k1 ,...,kd−1 ) (zd ) = ∑ αk zdkd , then β(k1 ,...,kd−1 ) (zd ) = 0 for each (k1 , . . . , kd−1 ). As
k1 ,...,kd−1 kd :(k1 ,...,kd ) ∈Λ
zd ∈ S is arbitrary, this holds for all zd ∈ S. However, each such β is a polynomial in zd . Since β vanishes on S, then it vanishes on C, so each
coefficient αk vanishes. Thus, q(z) ≡ 0 on Cd . ◻
Remark 2.9. If λ is an eigenvalue of H(θ), then the characteristic polynomial p(z; λ) must be symmetric in z±k because p(z; λ) = 0
= p(z; λ) on Sd .
Lemma 2.8 is the set of “necessary and sufficient conditions” given in Ref. 26 for H to have a flat band. The authors compute the
coefficients of zk in p(z; λ0 ) explicitly when (d = h1 = 1, ν ≤ 4), or (d ∈ {2, 3}, hi = 1, ν ≤ 3), or (d = 1, h1 ∈ {2, 3}, ν ≤ 3)and give examples
for a range of parameters that make these coefficients vanish so that a flat band appears. An example is also given for d = h1 = 1, ν = 5.
III. GENERATORS
The aim of Secs. III A and III B is to generate graphs with and without flat bands, respectively.
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-7

Journal of
A. Flat band generators

We are interested in Q ≡ 0 and w ≡ 1 here, so H = AΓ .
Theorem 3.1. Any eigenvalue of a finite graph GF can be made to appear as a flat band of a periodic graph Γ with ν = 2∣GF ∣, with eigenvector
ψ supported in V f . This choice is optimal in the sense that there exists GF and λ ∈ σ(GF ) such that λ does not arise in any periodic Γ having both
ν < 2∣GF ∣ and ψ supported in V f .
The eigenvector is explicit in the proof. Of course, this theorem is non-trivial because we consider AΓ and do not allow potentials;
otherwise, any λ ∈ R could trivially be made flat by shifting a flat band 0 with Q = λI. Before we give the proof, let us mention a consequence.
Corollary 3.2. A flat band is necessarily a totally real algebraic integer. Conversely, any totally real algebraic integer can be made to appear
as a flat band of some periodic Γ.
Recall that a totally real algebraic integer is a root of a monic polynomial with integer coefficients and real roots.
Proof. The first part follows from Theorem 2.4, as an eigenvector of compact support is also an eigenvector of a finite subgraph. The
more interesting converse follows from Theorem 3.1 and Ref. 24. ◻
Proof of Theorem 3.1. We construct Γ as shown in Fig. 3. Each vertex of the graph of Fig. 1 (right) is replaced by a copy of GF . The edges
of this new Γ consist of the old edges of each GF , plus an edge from each vertex vp ∈ GF to its copy vp′ in each neighboring GF . Hence, degΓ (vp )
= degGF (vp ) + 4. More precisely, Γ is the Cartesian product Γ = Γ0 ◻GF , where Γ0 is the infinite graph of Fig. 1, so AΓ = AΓ0 ⊗ I + I ⊗ AGF .
1
If λ ∈ σ( AGF ) with eigenvector ϕ and if f (0) = ( −1 ), f (k) = ( 00 ), k ≠ 0, is the eigenvector of AΓ0 corresponding to 0; let ψ = f ⊗ ϕ, i.e.,
ψ i (0) = ϕi on the upper GF , ψ i (0) = −ϕi on the lower GF , and ψ(k) = 0 for k ≠ 0. Then, AΓ ψ = ( AΓ0 f ) ⊗ ϕ + f ⊗ ( AGF ϕ) = 0 + λ f ⊗ ϕ = λψ.
√
Optimality follows from the examples in Sec. IV (which shows much more). In fact, the eigenvalues ± 2 of the path on three vertices do
not appear as single-cell flat bands for any Γ with ν < 6, as we see from the explicit lists. ◻
Although this generator is optimal for general GF , it can be improved for regular GF .
15 October 2024 06:43:54

Proposition 3.3. All eigenvalues of a regular connected graph GF except the top one can be generated in some Γ with ν = ∣GF ∣ + 1, with
eigenvectors localized on GF ⊂ V f .
Proof. We add a new vertex o and attach it to each of the vertices of GF and then Z-periodize along the vertices o. See Fig. 4. Then,
⎛2 cos 2πθ 1 ⋅⋅⋅ 1
⎞
1
A(θ) = ⎜ ⎟. As GF is regular, its top eigenvector is f 1 = (1, . . . , 1); the remaining ones f i , 1 < i ≤ ∣GF ∣ corresponding to λi ,
⎜ ⋮ AGF ⎟
⎝ 1 ⎠
∣G ∣
are orthogonal to it, i.e., satisfy ∑ j=1F fij = 0. Hence, the vector ̃
f i = (0, fi )⊺ satisfies A(θ)̃
f i = λĩ
f i for each i = 2, . . . , ∣GF ∣ and all θ. ◻
Remark 3.4. In the Proof of Theorem 3.1, we can, of course, generalize the construction by replacing Γ0 by any periodic graph having
the flat band zero.
In particular, we can take the Lieb lattice or the graph in Fig. 5 (right) with the depicted eigenvector f , and if ϕ is an eigenvector of GF
for λ0 , then ψ = f ⊗ ϕ will correspond to the flat band λ0 but ψ is not supported on a single-cell. Generating such ψ seems to be of interest to
physicists. Similarly, long-range hopping (with some hi > 1) can be included by changing Γ0 .
FIG. 3. The Cartesian product of the graph in Fig. 7 and GF .
FIG. 4. General procedure (left), example GF = P2 , λ = −1 (right).
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-8

Journal of
FIG. 5. Nearest hopping, ν = 3. Flat bands λ = −2 (left) and λ = 0 (right).
Remark 3.5. To embed the eigenvalues of a finite graph G into an infinite periodic graph, one could consider using the more classical
procedure of graph decoration, which simply means attaching a copy of G to each vertex of a given periodic graph Γ0 , for example, Γ0 = Z.
The example in Fig. 4 (right) can be viewed as an instance of this, with G = C3 , the 3-cycle (rather than P2 if we follow our construction). It
seems, however, that this idea only works for graphs G, which have eigenfunctions of compact support, such as C3 . For example, decorating
Z with G = P2 would give an infinite comb without flat bands.
B. No flat bands
Starting from a finite GF and any Γ0 with ν = 1, such as Zd or the triangular lattice, we can construct infinitely many new graphs Γ, which
have no flat bands by considering the Cartesian product Γ = Γ0 ◻GF , as was shown in Ref. 19, Sec. 3.2 (see Fig. 6).
Another operation is to take the tensor product. Here, starting again from a graph Γ0 with ν = 1, we consider Γ = Γ0 × GF for finite GF .
However, GF must now be chosen more carefully. It suffices to have GF non-bipartite with 0 ∉ σ( AGF ). In that case, Γ will be connected with
no flat bands. See Ref. 19, Sec. 3.4.2 for more details.
IV. SINGLE-CELL FLAT BANDS

The simplest kinds of flat bands one can try to characterize are those where the corresponding eigenvector is entirely supported on a
15 October 2024 06:43:54

single cell; cf. Ref. 9. We may assume this cell to be V f by shifting V f if necessary. We will use a quite abusive terminology and describe
these as single-cell flat bands. The set of all possible single-cell flat bands of AΓ that can appear in connected periodic graphs Γ with ∣V f ∣ = ν is
denoted by F sν .
Remark 4.1. The notion of a single-cell flat band depends on the choice of V f . It is a property of (Γ, V f ), rather than Γ. There is an
obvious non-uniqueness in choosing V f by shifting to Vf + ka . However, it can be more significant, as in Fig. 7. Note that changing V f also
changes I ij , as one can check from Fig. 7. Despite this “non-canonicity,” this notion is very useful to obtain some results, as we will see in this
section.
We note that one can always choose V f to be a tree (in particular, connected). The argument can be found in Ref. 16, Lemma 3.1. For
example, in this convention, one chooses V f on the left of Fig. 7. In general, the “natural” V f will not be a tree; one obtains a tree by removing
edges, passing to a spanning tree. We will not follow this convention here: our V f is arbitrary and can even be disconnected as in Fig. 7 (right).
Motivated by Fig. 1, we study the following neighborhood condition.

Lemma 4.2. For any ν and d, we have
N vi /{v j } = N v j /{vi } ⇔ Iii = I j j = I ji /{0} = Ii j /{0} and Iir = I jr ∀r ≠ i, j.
This condition also implies that deg(vi ) = deg(vj ).
Proof. Since N vi /{v j } = ( I i j /{v j }) ∪ (∪k≠ j I ik ), with I rs of the form {vs + na }, we see that equality holds iff I i j /{v j } = I j j ,
I ji /{vi } = I ii , and I ik = I jk for k ≠ i, j. This holds iff the statements holds for I rs , and we use (2.7) to conclude. ◻
FIG. 6. The graphs Z ◻ P2 (left) and Z ◻ C4 (right).
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-9

Journal of
FIG. 7. Two choices for V f are shown in red. The “obvious” choice (right) has a single cell flat band. The one on the left does not: any eigenvector of Γ will live on more than
one copy of V f .
This simple geometric condition always generates single-cell flat band 0 or −1.
Proposition 4.3. For any ν and d, if N vi /{v j } = N v j /{vi } for some i ≠ j, then AΓ has a single-cell flat band and I ij is symmetric
(−I ij = I ij ).
This flat band is λ = 0 if vi ≁ vj and −1 if vi ∼ vj . The corresponding eigenvector localized on V f can be chosen to have values +1 on vi , −1
on vj , and zero elsewhere.
⎛ h11 (θ) h11 (θ) + ϵ gθ

⎞
Proof. Up to rearranging the indices, we may assume i = 1 and j = 2. Then, (2.6) takes the form A(θ) = h11 (θ) + ϵ h11 (θ) gθ by
⎝ g∗ ∗
gθ
′
H (θ)⎠
θ
Lemma 4.2 for g θ = (h13 (θ), . . . , h1ν (θ)) and some (ν − 2) × (ν − 2) matrix H ′ (θ). Here, ϵ = 0 if vi ≁ vj and ϵ = 1 if vi ∼ vj .
Clearly, if f = (1, −1, 0, . . . , 0)⊺ , then H(θ) f = −ϵf . As this holds for all θ, with f independent of θ, we have shown that Γ has a single-cell
flat band −ϵ.
The set I ij is symmetric since I ij /{0} = I ii is symmetric. ◻
Figures 1 and 8 illustrate Proposition 4.3.
Proposition 4.4. For ν = 2, AΓ has a single-cell flat band iff N v1 /{v2 } = N v2 /{v1 }. Moreover, F s2 = {0, −1}.
15 October 2024 06:43:54

Proof. We already showed the converse in Proposition 4.3. Hence, suppose that Γ has an eigenvector supported in a single cell. By
h11 (θ) h12 (θ) a a
Proposition 2.6, this means there are some a, b independent of θ such that ( )( ) = λ( ). Let z j = e2πiθ j . Then, as Laurent
h12 (θ) h22 (θ) b b
polynomials,
′
a ∑ zk + b ∑ zp = aλ and a ∑ z−p + b ∑ zk = bλ. (4.1)
k∈I11 p∈I12 p∈I12 k′ ∈I22
Recall that 0 ∉ I ii . Equating the coefficients of z0 on both sides of each equation, we deduce that bϵ = aλ, aϵ = bλ, where ϵ = 0 if v1 ≁ v2
and ϵ = 1 if v1 ∼ v2 .
0 ϵ
If I 12 ⊆ {0}, then (4.1) implies I 11 = I 22 = ∅. Hence, A(θ) = ( ) induces a disconnected Γ. Hence, I 12 must contain nonzero terms.
ϵ 0
Then, (4.1) implies (I 12 /{0} = I 11 and b = −a) and (I 12 /{0} = −I 22 ), respectively, by looking at the coefficients of zp . This implies λ = −ϵ
and I 11 = I 22 = I 12 /{0}, where we used (2.7). Hence, N v1 /{v2 } = N v2 /{v1 } by Lemma 4.2. ◻
We next give a simple but useful observation.
Lemma 4.5. Fix (Γ, V f ). If λ is a single-cell flat band for AΓ , then λ ∈ σ( AVf ). If V f is connected, then its top eigenvalue cannot be a
single-cell flat band for Γ.
Proof. If ψ is an eigenvector entirely supported in V f , then

ν ν
λψi (0) = ( AΓ ψ)i (0) = ∑ ∑ ψ j (k) = ∑ ϵi j ψ j (0) (4.2)
j=1 k∈Ii j j=1
because ψ j (k) = 0 for k ≠ 0, and we denoted ϵij = 1 if vi ∼ vj and 0 otherwise. This means ψ(0) is an eigenvector of the matrix (ϵi j )νi, j=1 = AVf .
The second part is a Perron–Frobenius argument as in Theorem 2.7, step 2. ◻
Before we move on to ν > 2, it will be helpful to have more criteria to eliminate eigenvalues of finite graphs instead of studying case by
case. We have the following general result, which strengthens Lemma 4.5.
Theorem 4.6. Let ∣V f ∣ = ν. If λ is a single-cell flat band for AΓ , with eigenvector ψ, then λ ∈ σ( AVf ), and there exist δi ∈ {0, 1},
i = 1, . . . , ν, not all zero, such that ∑νi=1 δi ψi = 0.
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-10

Journal of
FIG. 8. Graphs with flat band λ = 0, with all Irs symmetric (left) and I13 , I23 non-symmetric (right).
Conversely, if λ ∈ σ( AVf ) and there exist δi ∈ {0, 1}, i = 1, . . . , ν, not all zero such that ∑νi=1 δi ψi = 0, then λ ∈ F sν , as long as Vf is
connected.
The connectedness of Vf is not necessary, a weaker assumption suffices, and we will need this in Corollary 4.7. The connectedness
condition cannot be completely dropped, however: there are λ ∈ σ( AVf ) that satisfy ∑νi=1 δi ψi = 0 without being flat [take, e.g., Vf = G4 in
Example 4.8 below, λ = 1, δ = (0, 0, 1)].
The condition ∑νi=1 δi ψi = 0 greatly generalizes some old procedures of constructing flat bands by starting from a finite connected graph
GF and an eigenfunction, which has a zero in GF . It is clear in that case that one can periodize GF along this vanishing vertex and obtain a flat
band by extending ψ by zero. This corresponds to the very special case where δi = 0 for all i except the vanishing vertex i0 , where δi0 = 1.
Proof. Suppose that λ is a single-cell flat band with eigenvector ψ. We saw in Lemma 4.5 that AVf ψ = λψ. It follows from Proposition 2.6
that ψ is an eigenvector of A(θ) for all θ. Let z j = e2πiθ j . We write A(z) = AVf + B(z), where B(z) is the matrix of bridges from Vf to its
neighboring copies [i.e., in (2.6), hi j (z) = ϵi j + ∑k ∈Iij /{0} zk , the first term is the entry in AVf and the second term is the entry in B(z)].
Now, A(z)ψ = λψ for all z, so ( AVf + B(z))ψ = λψ for all z. Since AVf ψ = λψ, it follows that B(z)ψ = 0 for all z.
Denote the entries of B(z) by bij (z). Then, for all i, ∑νj=1 bi j (z)ψ j = 0. If B(z) = 0, then A(z) = AVf and Γ is a disconnected direct sum
of copies of Vf . Hence, there exist i, r such that bir (z) ≠ 0. Say, 0 ≠ k ∈ I ir . Fixing i, k, let δj be the coefficient of zk in bij (z). Then, δr = 1. On
15 October 2024 06:43:54

the other hand, since ∑νj=1 bi j (z)ψ j = 0, then as a Laurent polynomial, the coefficient of zk must vanish; hence, ∑νj=1 δ j ψ j = 0. This proves the
first part.
Conversely, suppose that we found δj ∈ {0, 1} not all zero such that ∑νj=1 δ j ψ j = 0. To prove that λ ∈ F sν , it suffices to construct some Γ
with ∣Vf ∣ = ν such that λ is a single-cell flat band for Γ. We take Γ to be the Z-periodic graph defined by the matrix A(θ) = AVf + B(θ), where
B(z) ∶= (z + z−1 )B and B ∶= (δi δj ). Then, B(z)ψ = 0 and Γ is connected because V f is connected, and B(θ) gives bridges to (Vf ± 1a ) by our
choice of taking the weight z + z−1 = 2 cos 2πθ at each nonzero entry of B, which imply connections to the subsequent copies of V f . ◻
Corollary 4.7. The sets F sν are increasing: F sν ⊆ F sν+1 for all ν.
Proof. Let λ ∈ F sν , with Γ-eigenvector ψ on V f . By Theorem 4.6, we know that there are δj , j = 1, . . . , ν, with ∑νj=1 δj ψ j = 0. Consider
the disjoint union V ′f = Vf ⊔ {vν+1 }. Let ψ̃ j = ψ j on Vf and ψ̃ ν+1 = 0. Then, AV ′f ψ̃ = λψ̃. Define δ′ by δ′j = δ j for vj ∈ Vf and δν+1
′
= 1. Then,
ν+1 ′ ′ ′
∑ j=1 δ j ψ̃ j = 0. Here, V f is disconnected, so we cannot directly apply the converse in Theorem 4.6. However, this is why we chose δν+1 = 1.
′
In fact, we know we started from a connected Γ. To show that the graph Γ induced by A(z) = AV ′f + B(z) is connected, where B(z) is
constructed as in Theorem 4.6 using δ′ , we only need to show that vν+1 is connected to V f and vν+1 + 1a . By construction of B(z) (and the
′
choice δν+1 = 1), we know that there is a bridge from vν+1 to some vertex in Vf ± 1a and hence to any vertex in Γ. This also means that there
is a bridge from vν+1 ± 1a to Γ and hence from vν+1 to vν+1 + 1a . Thus, Γ′ is connected and λ is flat for Γ′ , which has ν + 1 vertices in its Vf′ . ◻
Proposition 4.4 characterizes F s2 as those satisfying the neighborhood condition. Interestingly, this is no longer true for ν = 3, although
we get the same flat bands.
Example 4.8. Let ν = 3. We apply our results so far to show that the following holds.
(1) F s3 = {0, −1}.
(2) Each of these single-cell bands can appear with multiplicity one or two, but they cannot appear together.
(3) Single-cell flat bands may exist even if N vi /{v j } ≠ N v j /{vi } for all i, j.
To see that 0 can appear with multiplicity 2, consider A(θ) = (aij (θ)) with aij (θ) = 2 cos 2πθ for all i, j. Similarly, −1 appears
with multiplicity 2 if A(θ) = (aij (θ)) with aii (θ) = 2 cos 2πθ and aij (θ) = 1 + 2 cos 2πθ for all i ≠ j. Both cases have the eigenvectors
(−1, 1, 0)⊺ , (−1, 0, 1)⊺ independent of θ and yield connected Γ.
Examples where 0 and −1 appear as flat bands of multiplicity one for A(θ) are given in Fig. 8 (left) and Fig. 4 (right), respectively.
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-11

Journal of
We now show that there are no more bands in F s3 . By Lemma 4.5, if λ ∈ F s3 , then it is an eigenvalue of one of the graphs in Fig. 9.
The corresponding eigenvalues and eigenvectors are
G1 2, −1, −1 (1, 1, 1), (−1, 0, 1), (−1, 1, 0),

G2 0, 0, 0 (1, 0, 0), (0, 1, 0), (0, 0, 1),
√ √ √ √
G3 2, 0, − 2 (1, 2, 1), (−1, 0, 1), (1, − 2, 1),
G4 1, 0, −1 (1, 1, 0), (0, 0, 1), (−1, 1, 0).
√
Eigenvalues 2 and 2√only arise as top eigenvalues of the connected G1 , G3 , respectively, so by Lemma 4.5, they cannot be flat.
Theorem 4.6 implies that − 2 ∉ F s3 .
The eigenvalue 1 ∉ F s3 . If it was, we would have Vf = G4 . Since 1 is simple for AVf , if 1 ∈ F s3 , then (1, 1, 0)⊺ would have to be an
eigenvector of A(θ) for all θ with eigenvalue 1. This would imply h13 + h23 = 0 as polynomials, so h13 = h23 = 0. However, this would make Γ
disconnected, with v3 not connected to any vi + na for i = 1, 2. Thus, 1 ∉ F s3 .
Next, the only possibility to have both 0 and −1 appear as single cell flat bands is if Vf = G4 . However, 0 cannot be flat in this case. In
fact, 0 is simple for AVf , so the corresponding eigenvector would be a multiple of (0, 0, 1)⊺ . This would imply h13 = 0, h23 = 0, and h33 = 0,
yielding again a disconnected Γ. This shows that 0 and −1 cannot appear together.
⎛2 cos 2πθ 0 2 cos 2πθ
⎞
Finally, A(θ) = 0 2 cos 4πθ 2 cos 4πθ induces a connected Γ, with eigenvector (1, 1, −1)⊺ localized on a single cell, but
⎝2 cos 2πθ 2 cos 4πθ 2 cos 2πθ + 2 cos 4πθ⎠
N vi /{v j } ≠ N v j /{vi }, ∀i, j.
We can refine Corollary 4.7 by describing more precisely how F sν+1 differs from F sν .
Theorem 4.9. We may write F sν+1 as a disjoint union,
15 October 2024 06:43:54

s s
F ν+1 = F ν ⊔ Cν+1 ,
where λ ∈ Cν+1 implies that λ ∈ σ( AVf ) for a connected graph on ν + 1 vertices or a disconnected graph Vf = V 1f ⊔ V 2f such that V if are
connected, ∣V 1f ∣ = ∣V 2f ∣ = ν+1
2
, and λ ∈ σ( AV 1 ) ∩ σ( AV 2 ).
f f
The set Cν+1 can be empty.
In particular, if ν + 1 is odd, only connected graphs on ν + 1 vertices may contribute new flat bands in F sν+1 .
Besides its theoretical input, the purpose of Theorem 4.9 is to eliminate the disconnected graphs from consideration. For example, for
ν = 5, this reduces the analysis from 34 graphs to 21 graphs. Let us clarify that disconnected graphs on ν + 1 vertices do contribute flat bands.
We are just saying that the flat bands they produce are old ones in F sν , with the exception of having V f precisely partitioned into two equal
parts. This last scenario is in harmony with our recipe, Theorem 3.1.
Proof of Theorem 4.9. Let λ ∈ F sν+1 / F sν . Then, λ ∈ σ( AVf ) for some graph V f on ν + 1 vertices. Suppose that V f is disconnected,
Vf = ⊔ri=1 V i , for some connected V i , so ∣V i ∣ ≤ ν. Then, σ( AVf ) = ∪ri=1 σ( AV i ). We claim that if ∣V i ∣ < ν+1
2
, then λ ∉ σ( AV i ). Indeed, if
λ ∈ σ( AV i ), then Theorem 3.1 and Corollary 4.7 would imply λ ∈ F s2∣V i ∣ ⊆ F sν , a contradiction.
If ∣V i ∣ < ν+1
2
for all i, we get a contradiction. If not, we may assume ∣V 1 ∣ ≥ ν+1
2
. Then, either ∣V i ∣ < ν+1
2
for all other i or V f = V 1 ⊔ V 2 with
1 2 ν+1
∣V ∣ = ∣V ∣ = 2 . In the latter case, if λ ∈ σ( AV 1 ) ∩ σ( AV 2 ), then we are done. Therefore, suppose that λ ∉ σ( AV 2 ). Then, in all cases, we are
in a situation where λ ∈ σ( AV 1 ) and λ ∉ σ( AV i ) for i > 1.
Now, since λ ∈ F sν+1 , we know by Theorem 4.6 that for the corresponding ψ, we may find δi ∈ {0, 1} not all zero such that ∑ν+1 i=1 δi ψi = 0.
Denote ψ = (ψ (1) , . . . , ψ (r) ) for the components of ψ over V i . Then, for v ∈ V k , we have Aψ (k) (v) = Aψ(v) = λψ(v) = λψ (k) (v). Since
λ ∉ σ( AV k ) for k > 1, we must have ψ (k) = 0 for k > 1.
FIG. 9. All graphs on three vertices.
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-12

Journal of
Thus, ψ = (ψ (1) , 0, . . . , 0). Hence, λ ∈ σ( AV 1 ), 0 = ∑ν+1 1 (1)

i=1 δi ψi = ∑v ∈V 1 δv ψv , and V is connected. By Theorem 4.6 and Corollary 4.7,
s s
this implies λ ∈ F ∣V 1 ∣ ⊆ F ν , again a contradiction. This completes the proof of the statement. ◻
So far, all flat bands have been {0, −1}, so integers are nonpositive. Interestingly, if ν = 4, the flat band can be positive and irrational.
Example 4.10. The set of single-cell flat bands for ν = 4 and ν = 5 is the same, with
√ √
s −1 + 5 −1 − 5
F 4 = {0, 1, −1, −2, , } = F s5. (4.3)
2 2
The fact that F s5 does not bring any new flat bands is quite remarkable, as there is a total of 34 graphs on five vertices, which, in principle,
could provide some.
To prove (4.3), we simply apply Theorem 4.6 to the subset of graphs that Theorem 4.9 tells us to consider. This, however, requires the
tables of eigenvalues and eigenvectors of graphs on four, five vertices; we provide this in the Appendix.
We showed that
∅ = F s1 ⊂ F s2 = F s3 ⊂ F s4 = F s5 (4.4)
and ∪ν F sν is the set of all totally real algebraic integers. It would be interesting to know if (4.4) is a general fact or a coincidence, i.e., do we
always have F s2k+1 = F s2k ? This would avoid the whole discussion in Subsection 2 of the Appendix. In view of Theorem 4.9, we may state the
question equivalently as follows: is it true that the connected graphs on 2k + 1 vertices only offer flat bands in F s2k ? In other words, do the
eigenvectors of such graphs corresponding to “new” eigenvalues never satisfy ∑ν+1 j=1 δ j ψ j = 0, δj ∈ {0, 1}, except when δj = 0 for all j?
To conclude, let us mention that there exist flat bands beyond single-cells even in the nearest-neighbor hopping case, as shown in Fig. 5.
For the graph in Fig. 5 (left), −2 is not even an eigenvalue of a graph on three vertices. This is because I 12 is not symmetric. When
all hopping are symmetric, such as in Fig. 5 (right), then the flat band is necessarily an eigenvalue of V f if d = h1 = 1. This is because the
15 October 2024 06:43:54

Floquet matrix takes the form A(θ) = AVf + 2 cos 2πθB for some matrix B encoding the first-hopping, so if λ is an eigenvalue for all θ, it is,
in particular, for θ = 41 .
V. CASE OF TWO VERTICES

In the following, d is arbitrary, ν = 2, Q ≡ 0, w ≡ 1. Proposition 4.4 characterizes single-cell flat bands. Here, we study existence beyond
single-cells. We also show that flat bands are always integers if ν = 2, and we give two formulas for them in terms of I ij .
f1 (z) g(z)
Let z j = e2πiθ j . We denote (2.6) as A(z) = ( −1 ), with fi (z) = ∑k ∈Iii zk and g(z) = ∑p ∈I12 zp . Then, λ is flat iff ( f 1 (z) − λ)( f 2 (z)
g(z ) f2 (z)
− λ) = g(z)g(z−1 ) as Laurent polynomials, by Lemma 2.8.

f1 (z) 0
We first note that we cannot have ∣I 12 ∣ = 0; otherwise, A(z) = ( ) induces a disconnected graph Γ in which v1 is not connected
0 f2 (z)
to any v2 + na . Hence,
∣I12 ∣ ≥ 1 (5.1)
for any periodic connected graph with ν = 2.
A. Absence of flat bands

Most arguments here are based on degree analysis of Laurent polynomials. Let Ii j (r) = {kr ∈ Z : k ∈ Ii j } for 1 ≤ r ≤ d and
⎧
⎪ kmax (r) = max I11 (r),
⎪ k′max (r) = max I22 (r),
⎨
⎩ pmax (r) = max I12 (r), pmin (r) = min I12 (r).
⎪
⎪
If Γ has a flat band, then for all r,

kmax (r) + k′max (r) = pmax (r) − pmin (r). (5.2)
′
In fact, we have (∑I11 zk − λ)(∑I22 zk − λ) = (∑p zp )(∑p z−p ). Evaluate at z = (zr , 1), so zi = 1 for i ≠ r. As Laurent polynomials in zr , the
largest degrees on both sides are kmax (r) + k′max (r) and pmax (r) − pmin (r), respectively, so they coincide.
It follows, in particular, that if Γ has a flat band, then
∣I11 ∣ + ∣I22 ∣ ≥ 2. (5.3)
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-13

Journal of
p
0 ϵz
In fact, if I 11 = I 22 = ∅, then kmax (r) + k′max (r) = 0 = pmax (r) − pmin (r) for each r, so I 12 = {p} or I 12 = ∅, i.e., A(z) = ( −p ) for ϵ ∈ {0, 1}
ϵz 0
and Γ is disconnected. Hence, I 11 or I 22 is nonempty, and since each is symmetric without 0, each has even cardinality.
Expand the characteristic polynomial,
⎛ ′⎞ ′ ′
p(z; λ) = λ2 − ∑ zk + ∑ zk λ + ∑ zk+k − ∑ zp−p . (5.4)
⎝k∈I11 k′ ∈I22 ⎠ k∈I11 p,p′ ∈I12
′
k ∈I22
√
Lemma 5.1. If AΓ has a flat band λ, then λ = ± ∣I12 ∣ − ∣I11 ∩ I22 ∣ ∈ Z.
Proof. Let λ be a flat band. By Lemma 2.8, the coefficient of z0 in (5.4) vanishes identically. Since 0 ∉ I ii , this coefficient is precisely
2
λ + ∣I 11 ∩ I 22 ∣ − ∣I 12 ∣. Indeed, it comes from taking, for each k ∈ I 11 , a corresponding k′ ∈ I 22 with k′ = −k and taking, for each p ∈ I 12 , a cor-
responding p′ = p. The latter has cardinality ∣I 12 ∣. For the former, if k′ = −k, then since I 11 is symmetric, we get k′ ∈ I 11 , so k, k′ ∈ I 11 ∩ I 22 .
Conversely, for each k ∈ I 11 √ ∩ I 22 , there is a corresponding k′ = −k contributing to the coefficient.
We showed that λ = ± ∣I12 ∣ − ∣I11 ∩ I22 ∣. Finally, since p(z; λ) ≡ 0 and the double sums in (5.4) have coefficients ±1, then λ must be an
integer. ◻
√
∣I11 ∣+∣I22 ∣− (∣I11 ∣−∣I22 ∣)2 +4∣I12 ∣2
Lemma 5.2. If λ is a flat band, we also have λ = 2
.
√
2 2 ∣I11 ∣+∣I22 ∣± (∣I11 ∣+∣I22 ∣)2 −4∣I11 ∥I22 ∣+4∣I12 ∣2
Proof. Consider z = 1 in (5.4). This yields λ − λ(∣I 11 ∣ + ∣I 22 ∣) + ∣I 11 ∥I 22 ∣ − ∣I 12 ∣ = 0. Hence, λ = 2
√
∣I11 ∣+∣I22 ∣± (∣I11 ∣−∣I22 ∣)2 +4∣I12 ∣2
= 2
.
√
(∣I11 ∣−∣I22 ∣)2 +4∣I12 ∣2
Using (5.3), ∣I 11 ∣ + ∣I 22 ∣ ≠ 0. If the + sign occurs in λ, then λ > 2
≥ ∣I12 ∣. This contradicts Lemma 5.1, so the sign must be
negative.
15 October 2024 06:43:54

◻
Lemma 5.3. If ν = 2, then AΓ has no flat bands if any of the following holds:
(i) kmax (r) + k′max (r) ≠ pmax (r) − pmin (r) for some r.
(ii) ∣I 12 ∣ = 1.
(iii) I 11 = ∅ or I 22 = ∅.
For example, the honeycomb lattice has I ii = ∅ and no flat bands. Note that this is not true for ν > 2. The graph in Fig. 4 (right) has
I 11 = I 22 = ∅ but has a flat band. The graph in Fig. 12 has I ii = ∅ for all i, but has flat bands.
Proof. We proved (i) in (5.2). For (ii), if Γ has a flat band and ∣I 12 ∣ = 1, then by (5.2), kmax (r) + kmin (r) = 0, so I 11 = I 22 = ∅, contradicting
(5.3). Thus, Γ has no flat bands.
For (iii), suppose on the contrary that λ is flat. By (5.3), I 11 and I 22 cannot both be empty, so we may assume, by symmetry, that I 11 ≠ ∅
and I 22 = ∅. ′
Then, λ2 − λ∑k ∈I11 zk = ∑p,p′ ∈I1,2 zp−p for all z. On the RHS, if d = 1, then the highest power pmax − pmin appears only once. In general, the
maximum power in lexicographic order appears only once. More precisely, we choose the power n with highest pmax (1) − pmin (1). This will,
in general, be a set of powers. Among these elements, we choose the one with highest pmax (2) − pmin (2) and so on. Then, the coefficient of zn
is 1, and we must have λ = −1 on the LHS.
However, by Lemma 5.1, λ2 = ∣I 12 ∣, so ∣I 12 ∣ = 1. This contradicts (ii). ◻
B. Existence beyond single-cells

It is easy to generate flat bands, which do not live on a single cell by taking Γ such that I 11 = I 22 = (I 12 /{ℓ}) − ℓ, for some integer ℓ. For
4πiθ
2 cos 2πθ 1+e
example, consider I 11 = I 22 = {±1} and I 12 = {0, 2}. This gives the graph in Fig. 10. Here, ℓ = 1, A(θ) = ( −4πiθ ), with eigenvector
1+e 2 cos 2πθ
1
( −2πiθ ) for λ = 0. We may construct an eigenvector for Γ according to Proposition 2.6; here, α1 (0) = 1, α2 (1) = −1, and αk (m) = 0 otherwise.
−e
We note, however, that such a graph is isomorphic to a graph Γ′ on which the flat band lives on a single cell; simply take ϕ : (v1 + ka )
↦ (v1 + ℓa + ka ), ϕ : (v2 + ka ) ↦ (v2 + ka ), and for edges, ϕ : {u, u′ } ↦ {ϕ(u), ϕ(u′ )}. For the graph in Fig. 10, this corresponds to a shearing
of the top layer; the isomorphic graph Γ′ is the one of Fig. 1 (right).
We may wonder if these are the only non-single-cell flat bands. Note that such I 12 are not symmetric if ℓ ≠ 0. Therefore, we take the
special case where I 12 is symmetric and ask whether all flat bands are in F s2 .
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-14

Journal of
FIG. 10. Flat band λ = 0 with I12 non-symmetric and localized ψ.
FIG. 11. Flat band λ = −2 with localized eigenvector on five cells (in red). The edge colors are only there to slightly improve visibility.
3 −3 −1 2 −2
z +z z+z +z +z
The answer is no. In fact, take I 11 = {±3}, I 22 = {±1}, and I 12 = {±1, ±2}, so A(z) = ( −1 2 −2 −1 ). Then, we have a flat
z+z +z +z z+z
−1
−2 − z − z
band λ = −2 corresponding to the eigenvector ( −1 2 −2 ). Using the recipe in Proposition 2.6, we construct a corresponding eigenvector
z+z +z +z
for Γ in Fig. 11.
In contrast to the single-cell case, here, −2 is not an eigenvalue of a graph on two vertices.
15 October 2024 06:43:54

It is then natural to ask the following: what are the flat bands that can be generated by ν = 2? We know that they are always integers, so
not all flat bands can appear in this fashion. Are there other values of λ besides 0, −1, −2? If not, this would induce some constraints on the
possible choices of I ij .
In this example, ∣I 12 ∣ = 4. We know that if ∣I 12 ∣ = 1, then there are no flat bands by Lemma 5.3. If ∣I 12 ∣ = 2 or ∣I 12 ∣ = 3, then the only flat
bands that can appear are those living on single-cells, up to the isomorphism explained before. This fact is not entirely obvious, but since
we do not need it, let us simply mention that the main idea is to combine Lemmas 5.1 and 5.2 to deduce constraints on the sets I ii and then
analyze the degrees and coefficients of the possible characteristic polynomials.
VI. FLAT BANDS FROM SYMMETRY

The basic idea in Ref. 23 is that the appearance of flat bands can sometimes be attributed to the presence of symmetries in Γ, such as in
Fig. 1. We should be a bit careful by what we mean by symmetry since the infinite ladder in Fig. 6 is very symmetric yet features no flat bands.
We explain this in some detail in this section and compare our results with Ref. 23 at the end.
Given an infinite Γ and its V f , the symmetries that interest us are those permuting some vertices of V f while keeping those outside of
V f fixed. Importantly, when we permute vertices, we exchange the whole star attached to them. See Figs. 12 and 13 for illustration. We are
interested in the case when Γ remains the same after such permutation.
Figures 12 and 13 motivate the following definition. Let ϵi,j = 1 if vi ∼ vj and ϵi,j = 0 otherwise. We say that Γ has a local symmetry if there
exists a nontrivial permutation ϕ of V f such that for any i, j = 1, . . . , ν,
Ii, j /{0} = Iϕ(i), j /{0} and ϵi, j = ϵϕ(i),ϕ( j). (6.1)
FIG. 12. The 1d pyrochlore is preserved by the symmetry v 1 ↔ v 2 , v 3 ↔ v 4 . Vertices outside V f are kept fixed. We colored the edges to show their position after performing
the symmetry, but they carry no weight.
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-15

Journal of
FIG. 13. The ladder is not preserved by exchanging v 1 and v 2 .
If Γ has a symmetry preserving it as in Fig. 12, then (6.1) must be satisfied because if k ∈ I ij /{0}, then vi ∼ v j + ka , so after the symmetry,
since v j + ka stays fixed, we must have vϕ(i) ∼ v j + ka , i.e., k ∈ I ϕ(i),j /{0}. Similarly, I ϕ(i),j /{0} ⊆ I ij /{0} since ϕm = id for some m. On the other
hand, if ϵi,j = 1, then vi ∼ vj , so after performing the symmetry, we should have vϕ(i) ∼ vϕ( j) and vice versa.
In particular, the ladder does not satisfy this condition under v1 ↔ v2 since I 11 = I 22 = {±1}, while I 12 = {0} = I 21 . However, the
pyrochlore satisfies it under ϕ(v1 ) = v2 , ϕ(v2 ) = v1 , ϕ(v3 ) = v4 , and ϕ(v4 ) = v3 . Here, I ii = ∅ for all i, I 12 = I 34 = {0}, I 13 = I 24 = {−1},
I 14 = I 23 = {0, −1}, and I ji = −I ij for the rest as usual. In addition, ϵ12 = ϵ14 = ϵ21 = ϵ23 = ϵ32 = ϵ34 = ϵ43 = 1; the rest are zero. One checks that,
indeed, (6.1) is satisfied.
We have shown that if Γ is preserved by a symmetry, then it is necessary for (6.1) to be satisfied. We now show that (6.1) is enough to get
some flat bands. Actually, a weaker condition on ϵij suffices; see Remark 6.6.
We first note that this symmetry concept is a generalization of the neighborhood condition that we discussed before.
Lemma 6.1. We have N vi /{v j } = N v j /{vi } iff (6.1) is satisfied for the permutation ϕ(vi ) = vj , ϕ(vj ) = vi , ϕ(vr ) = vr for r ≠ i, j.
Proof. This follows from Lemma 4.2. ◻

We next make the following observation.
Lemma 6.2. Suppose that (6.1) is satisfied for a permutation ϕ. Write ϕ as a product of r cycles of sizes kj (we allow kj = 1, i.e., fixed
elements, and allow ki = kj ). Then, up to renumbering the vertices of V f , the Floquet matrix A(θ) acquires an r × r block structure, with blocks
15 October 2024 06:43:54

Dst (θ) of sizes ks × kt . Moreover, Dst (θ) = bst (θ)J st + Cst , where J st is the all-one matrix of size ks × kt and Cst is circulant, independent of θ, for
s, t ≤ r.
(1) (1) (1) (2) (2) (r) (r)
Proof. Let ϕ = (i1 i2 ⋅ ⋅ ⋅ ik1 )(i1 ⋅ ⋅ ⋅ ik2 ) ⋅ ⋅ ⋅ (i1 ⋅ ⋅ ⋅ ikr ) and k0 ∶= 0.
By renumbering the vertices of V f , we can replace this cumbersome permutation by (1, 2, . . . , k1 )(k1 + 1, . . . , k2 ) ⋅ ⋅ ⋅ (kr−1 + 1, . . . , kr ).
′
The renumbering is simply vm = vi(1) and vk′ n +m = vi(n+1) for n ≥ 1. These cycles induce the block structure of A(θ).
m m
Let bi j (θ) = ∑Iij /{0} e2πik⋅θ . Then, hij (θ) = ϵij + bij (θ). Condition (6.1) entails that the column of each block Dst (θ) has a fixed dependence
on θ. Indeed, I ij /{0} = I ϕ(i),j /{0} = I i+1,j /{0} by the cycle structure, where i + 1 ∶= ks−1 + 1 if i = ks . Hence, bij (θ) = bi+1,j (θ) as asserted for
i = ks−1 + 1, . . . , ks and j = kt−1 + 1, . . . , kt . The same holds for Dts (θ), so we get bji (θ) = bj+1,i (θ) for j = kt−1 + 1, . . . , kt and i = ks−1 + 1, . . . , ks .
However, h ji = hij , so b ji = bij . Thus, bi j = bji = bj+1,i = bi, j+1 . This shows that the rows have the same θ dependence. Thus, all entries of Dst (θ)
take the form bi0 j0 (θ) + ϵi j for some fixed i0 , j0 .
It remains to show that Cst = (ϵij ) is circulant. Condition (6.1) says that ϵij = ϵi+1,j+1 in each block, that is, diagonals are constant and we
have a Toeplitz structure. However, Cst is, in fact, circulant using ϵi,kt = ϵi+1,kt−1 +1 and ϵks , j = ϵks−1 +1, j+1 . ◻
Theorem 6.3. Under the assumptions of Lemma 6.2, Γ has at least ν − r single-cell flat bands.
Proof. Let A(θ) = (Di j (θ))i, j≤r be in the block structure provided by Lemma 6.2. Consider the partition {X 1 , . . . , X r } of V f induced by
ϕ, where X j contains the elements of the jth cycle. It follows from Lemma 6.2 that
r
Colk (A(θ)) = ∑ bik (θ)1Xi + Ck , (6.2)
i=1
where Ck is a column vector independent of θ and bik = bik′ if k, k′ are in the same block.
Let Yi = √1 1Xi for i = 1, . . . , r. Clearly, Y i ’s are orthonormal. Complete this with Y i , i = r + 1, . . . , ν, as an orthonormal basis of ℓ2 (V f ).
∣Xi ∣
We now write A(θ) in the basis {Yi }νi=1 ; in other words, the matrix elements become aij = ⟨Y i , A(θ)Y j ⟩.
Let W be the subspace generated by Y 1 , . . . , Y r . Then, W is invariant under A(θ). In fact, if v ∈ X i , then (A(θ)1X j )(v) = ∑p ∈X j A(θ)v,p is
the vth row sum of the block Dij (θ), which is equal to some dij (θ) independent of v since Dij (θ) is circulant. Thus, A(θ)1X j = ∑ri=1 di j (θ)1Xi
∈ W as required.
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-16

Journal of
It follows that W is invariant as well. Hence, ⟨Y i , A(θ)Y j ⟩ = 0 for (i ≤ r ∧ j > r) or (i > r ∧ j ≤ r), i.e., A(θ) takes a block-diagonal form
of sizes r and ν − r, respectively, in the basis {Yk }k≤ν and σ(A(θ)) is the union of the respective spectra. The eigenvectors ϕp of the submatrices
are also eigenvectors ϕ̃p for A(θ) if we extend ϕp by zero.
Since M f = ∑m ν
k=1 f (k)Colk (M)for an m × m matrix M, then for i, j > r, we have by (6.2), ⟨Yi , A(θ)Y j ⟩ = ∑k=1 Y j (k)⟨Yi , Colk (A(θ))⟩
= ∑νk=1 Y j (k)⟨Yi , Ck ⟩ = ⟨Yi , CY j ⟩. As the submatrix (⟨Yi , CY j ⟩)i, j>r is independent of θ, so are its eigenvalues and eigenvectors {μp , ϕp }ν−r
p=1 .
Now, the matrix A(θ) in the basis {Y k } is Ã(θ) = Y ∗ A(θ)Y, where Y is the unitary matrix of columns Y k . We showed that Ã(θ)ϕ̃p = μp ϕ̃p ,
so A(θ)Y ϕ̃p = μp Y ϕ̃p for all θ, and Y ϕ̃p = ∑k>r ϕ̃p (k)Yk is independent of θ. Therefore, μp is of single-cell type. ◻
1 1 1
Example 6.4. The graph in Fig. 12 has ϕ = (12)(34), Y1 = √
2
(1, 1, 0, 0), Y2 = √
2
(0, 0, 1, 1), Y3 = √
2
(1, −1, 0, 0), and Y4 =
1
√
2
(0, 0, 1, −1), so
⎛ 0 1 e−2πiθ 1 + e−2πiθ ⎞ ⎛ 1 1 + 2e−2πiθ 0 0⎞

⎜ ⎟ ⎜ ⎟
⎜ 1 0 1+e −2πiθ
e−2πiθ ⎟ ⎜1 + 2e2πiθ 1 0 0⎟
⎜ ⎟ ⎜ ⎟
A(θ) = ⎜ 2πiθ 2πiθ
⎟, Ã(θ) = ⎜ ⎟,
⎜ e 1+e 0 1 ⎟ ⎜ 0 0 −1 −1⎟
⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟
⎝1 + e2πiθ e2πiθ
1 0 ⎠ ⎝ 0 0 −1 −1⎠
and we get the flat bands {−2, 0}. The eigenvectors are given by Y ϕ̃1 = Y3 + Y4 , Y ϕ̃2 = Y3 − Y4 for ϕ1 = ( 11 ), ϕ2 = ( −1
1
). The estimate
ν − r = 2 is exact here.
Example 6.5. The graph corresponding to the Floquet matrix at the end of Example 4.8 has a single-cell flat band λ = 0, which is not
induced by any symmetry discussed in this section. In fact, since we showed that it does not satisfy the neighborhood condition, this excludes
15 October 2024 06:43:54

the permutations of two vertices fixing the third. The remaining permutation would be (123), but this is also not valid either since A(θ) would
then have the form of a single block h(θ)J + C by Lemma 6.2, with J all one, which is not the case here.
Remark 6.6. Condition (6.1) on ϵij means that ϕ : V f → V f is a graph automorphism. We saw in Lemma 6.2 that this implies the blocks
are circulant. However, in Theorem 6.3, we only used the weaker property that each block has a constant row sum. In particular, we can
recover Proposition 3.3 by considering ϕ to be cyclic on GF and fixing o so that ν = ∣GF ∣ + 1 and r = 2. In general, keeping the condition on
I ij /{0}, for ϵij , we only need the cycles of ϕ to induce an equitable partition on V f . That is, each cell forms a regular subgraph (Css have
constant row sum) and all vertices in a given cell s have the same number of edges to the cell t (so Cst have constant row sum). There can be
no edges at all, whether within the cell or between different cells.
Compared to Ref. 23, our cells can be of arbitrary sizes; we do not need ϕ to be “basic of order k.” In addition, Ref. 23 needs some sites
in V f to be fixed by ϕ, as the corresponding eigenvectors vanish on them, and this allows them to connect the copies of V f using these points
[Ref. 23, Eq. (19) and Sec. III B]. Our argument does not need any vertex to be fixed by ϕ.
We saw in Lemma 6.1 that the permutation invariance in this section generalizes the concept of having two vertices sharing the same
neighbors, a criterion we had before. A different generalization is to continue to look at only two vertices, which can now be far apart, but
have the same number of walks to a given subset of V f . This idea is explored in Ref. 20 and is shown to yield flat bands as well.
VII. DESTROYING FLAT BANDS

The aim of this section is to take small steps toward answering problem 2, that is, showing that flat bands are rare and disappear under
generic perturbations.
Lemma 7.1. The set PΓ = {Q ∈ Rν : AΓ + Q has a flat band} is semialgebraic.
Proof. Let p(z; Q, λ) ∶= det(A(z) + Q − λI). We have by Lemma 2.8,

ν
PΓ = {Q ∈ R : ∃λ ∈ R, p(z; Q, λ) = 0 ∀z ≠ 0}
ν+1
= π1 {(Q, λ) ∈ R : p(z; Q, λ) = 0 ∀z ≠ 0},
where π1 is the projection π1 (Q, λ) = Q.
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-17

Journal of
Now, p(z; Q, λ) is a Laurent polynomial in z, and it is also polynomial in Q1 , . . . , Qν , λ. We expand p(z; Q, λ) = ∑r∈Λ pr (Q, λ)zr , where
Λ is a finite subset of Zd determined by ν, d, and I ij , i, j ≤ ν. Then, the Laurent polynomial p(z; Q, λ) vanishes for all z iff each coefficient of zr
vanishes. Each pr (Q, λ) is a polynomial in (Q1 , . . . , Qν , λ). Hence,
ν+1
PΓ = π1 {(Q, λ) ∈ R : pr (Q, λ) = 0 ∀r ∈ Λ}
is the projection of an algebraic variety and is thus semialgebraic. ◻

We now observe that the following holds.
(1) There exists some r ≠ 0 in Λ such that pr (Q, λ) is not the trivial polynomial. Indeed, if pr (Q, λ) was trivial for all r ≠ 0, we would have
p(z; Q, λ) = p0 (Q, λ) for all z, Q, λ. This would mean that the characteristic polynomial of A(z) + Q is independent of z, and so are its
roots, which would imply that all bands of AΓ + Q are flat, contradicting Theorem 2.7.
(2) The polynomial p0 (Q, λ) is nontrivial and distinct from the polynomials pr (Q, λ) for r ≠ 0. In fact, p0 (Q, λ) contains the term (−1)ν λν ,
which does not appear in any other pr (Q, λ), since p(z; Q, λ) = (−1)ν λν + q(z; Q, λ), with q(z; Q, λ) being a polynomial of lower order
in λ
It follows from (1) and (2) that {(Q, λ) ∈ Rν+1 : pr (Q, λ) = 0 ∀r} is given by the vanishing of at least two distinct nontrivial
polynomials. We would like to conclude that the algebraic variety has dimension at most ν − 1, and so does its projection PΓ . The
problem is that p0 and pr might share a common factor q, i.e., pr = q f and p0 = qg for some polynomials q, f , g. In that case, {(Q, λ) :
p0 (Q, λ) = pr (Q, λ) = 0} ⊇ {(Q, λ) : q(Q, λ) = 0} could have dimension ν. Let us show that this does not happen for ν = 2.
Lemma 7.2. If ν = 2, then PΓ has dimension at most one.
Proof. Here,
′ ′ ′
p(z; Q, λ) = ∑ ∑ zk+k + (Q1 − λ) ∑ zk + (Q2 − λ) ∑ zk + (Q1 − λ)(Q2 − λ) − ∑ zp−p ,
k∈I11 k′ ∈I22 k′ ∈I22 k∈I11 p,p′ ∈I12
15 October 2024 06:43:54

so p0 (Q, λ) = n + (Q1 − λ)(Q2 − λ) for n = ∣I 11 ∩ I 22 ∣ − ∣I 12 ∣. The nontrivial pr (Q, λ) must have the form pr (Q, λ) = m + δ1 (Q1 − λ)
+ δ2 (Q2 − λ) for some m ∈ Z and δi ∈ {0, 1}, the coefficient m coming from the double sums.
If δ1 = δ2 = 0, then either pr = m ≠ 0 never vanishes and so PΓ = ∅ or m = 0 and pr is trivial, contradicting the assumption. Therefore, let
δ1 + δ2 ≠ 0. Then, pr = 0 implies that λ = m+δδ1 1Q+δ
1 +δ2 Q2
2
. Substituting into p0 , we get p0 (Q, λ) = n + (δ2 (Q1 −Q2 )−m)(δ 1 (Q2 −Q1 )−m)
(δ +δ )2
. This reduces to
1 2
2 2
p0 (Q, λ) = n − (Q1 −Q42 ) −m or p0 (Q, λ) = n + m2 ± m(Q1 − Q2 ). Either way, for p0 to vanish, we must have Q2 = Q1 + c for some c ∈ R. This
proves the result. ◻
Remark 7.3. It is worthwhile to note that if there exists a coefficient pr (Q, λ) = m ≠ 0 independent of (Q, λ), then PΓ = ∅. This was
−1 −1
Q1 1 + z1 + z2
mentioned in the previous proof and holds for any ν. A simple example is the honeycomb lattice H(z) = ( ).
1 + z1 + z2 Q2
Let us turn back to general ν > 1 and prove a couple of results.

Lemma 7.4. For any d, ν > 1, the set PΓ is closed.
This does not follow from Lemma 7.1 as the projection of an algebraic variety may not be closed, e.g., π1 {(x, y) ∈ R2 : xy − 1 = 0}
= R/{0}.
Proof. Let {Q(n) } ⊂ PΓ , Q(n) → Q ∈ Rν . Then, for each n, AΓ + Q(n) has a flat band. Hence, there exists λn = λ(Γ, Q(n) ) such that
det(A(θ) + Q(n) − λn ) = 0 for all θ ∈ Td∗ .
As Q(n) converges in Rν , it is bounded, ∥Q(n) ∥∞ ≤ c, so AΓ + Q(n) is uniformly bounded; in particular, {λn } is a bounded sequence. Let
(λnk ) be a subsequence converging to some λ∗ . Since Q(nk ) → Q and λnk → λ∗ , we have for any fixed θ: det (A(θ) + Q(nk ) − λnk ) → det (A(θ)
+ Q − λ∗ ) because the determinant is a polynomial in Q(nk ) and λnk . However, we know det (A(θ) + Q(nk ) − λnk ) = 0 for all nk and all θ.
Hence, for any θ ∈ Td∗ , we have det(A(θ) + Q − λ∗ ) = 0. This shows that AΓ + Q has a flat band, i.e., Q ∈ PΓ . ◻
Proposition 7.5. For any d, ν > 1, there exists Γ such that PΓ ⊇ {Q ∈ Rν : Q1 = Q2 }, and thus, dim PΓ ≥ ν − 1.
Proof. Take Γ with N v1 /{v2 } = N v2 /{v1 } and add Q = (Q2 , Q2 , Q3 , . . . , Qν ) with Qi arbitrary for i ≥ 2. Then, (2.6) takes the form H(θ) =
h11 (θ) + Q2 h11 (θ) + ϵ gθ
⎛ ⎞
h11 (θ) + ϵ h11 (θ) + Q2 gθ for g θ = (h13 (θ), . . . , h1ν (θ)). Clearly, (1, −1, 0, . . . , 0)⊺ is an eigenvector with eigenvalue λ = Q2 − ϵ for all θ.◻
⎝ g∗ ∗
gθ
′
H (θ)⎠
θ
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-18

Journal of
FIG. 14. Here, ν = 2, AΓ has no flat bands, but AΓ + Q has the flat band λ = −2 if Q = −1 on the upper row and Q = 0 on the lower row.
It was conjectured in Ref. 15, P. 590–591 that if we perturb AΓ by a potential Q such that Q1 < Q2 < ⋅ ⋅ ⋅ < Qν , then AΓ + Q has no flat
bands. This is not true because such potentials can create flat bands if AΓ has no flat bands, as the example in Fig. 14, taken from Ref. 9, shows.
−2πiθ
−1 1+e
Here, H(θ) = ( 2πiθ ).
1+e 2 cos 2πθ
Still, this conjecture is almost true for ν = 2: we showed in Lemma 7.2 that for any Γ, there exists c such that PΓ ⊆ {Q ∈ R2 : Q1 = Q2 + c}.
⎛c c c
⎞ Q1 +Q2
However, it is wrong for higher ν. For example, for A(θ) = c c c with c = 2 cos 2πθ, we have the flat band λ = 2
for all perturbations
⎝c c 0⎠
of the form (Q1 , Q2 , Q1 +Q2
2
). The same holds for the Laplacian D − A(θ), with D = Diag(6, 6, 4) being the degree matrix, where flat bands
occur for all (Q1 , Q2 , Q1 +Q
2
2
+ 2).
DEDICATION
Dedicated to Abel Klein on the occasion of his 75th birthday.
AUTHOR DECLARATIONS
Conflict of Interest
The authors have no conflicts to disclose.
15 October 2024 06:43:54

Author Contributions
Mostafa Sabri: Writing – original draft (equal). Pierre Youssef: Writing – original draft (equal).
DATA AVAILABILITY
The data that support the findings of this study are available within the article.
APPENDIX: EIGENVALUES AND EIGENVECTORS OF SMALL GRAPHS

We give here the details for (4.3). The tables were produced by Wolfram Alpha.
1. Case of four vertices

Given F s3 , by Theorem 4.9, the only V f that can produce more bands for F s4 are those in Fig. 15.
We compute √ the eigenvalues
√
and eigenvectors of these graphs in Table II.
Here, α = 1+ 2 17 , α′ = 1− 2 17 , βi are the roots of β3 − β2 − 3β + 1 = 0, with β1 ≈ 2.170 09, β2 ≈ 0.311 11, and β3 ≈ −1.481 19 and
√ √ √ √
κ1 ≈ 1.854 64, κ2 ≈ −0.451 606, and κ3 ≈ 0.596 968. Finally, σ1 = 1+2 5 , σ2 = −1+2 5 , σ3 = 1−2 5 , and σ4 = −1−2 5 .
We know F s3 = {0, −1} ⊆ F s4 . By looking at the eigenvectors, we may deduce from Theorem√
4.6 that
√
−2, σ2 , σ4 ∈ F s4 (because σ 3 = −σ 2
and σ 4 = −σ 1 ). We can also see this more explicitly from Fig. 16, which shows that σ2 = −1+2 5 , σ4 = −1−2 5 can appear together, and Example
6.4, which showed that −2 arises in F s4 from the 1d pyrochlore.
FIG. 15. All connected graphs and the relevant disconnected graph.
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-19

Journal of
TABLE II. Eigenvalues and eigenvectors of the corresponding four-graphs.
G1 3, −1, −1, −1, −1 (1, 1, 1, 1), (−1, 0, 0, 1), (−1, 0, 1, 0), (−1, 1, 0, 0)
′ ′
G2 α, 0, −1, α′ ( α2 , 1, α2 , 1), (0, −1, 0, 1), (−1, 0, 1, 0), ( α2 , 1, α2 , 1)
G3 2, 0, 0, −2 (1, 1, 1, 1), (0, −1, 0, 1), (−1, 0, 1, 0), (−1, 1, −1, 1)
G4 √1 , β2 , −1,√
β β3 (κ1 , κ1√, β1 , 1), (κ2 , κ2 , β2 , 1), (−1, 1, 0, 0), (κ3 , κ√ 3 , β3 , 1)
G5 3, 0, 0, − 3 (1, 1, 3, 1), (−1, 0, 0, 1), (−1, 1, 0, 0), (1, 1, − 3, 1)
G6 σ1, σ2, σ3, σ4 (1, σ 1 , σ 1 , 1), (−1, σ 3 , σ 2 , 1), (1, σ 3 , σ 3 , 1), (−1, σ 1 , σ 4 , 1)
G7 1, 1, −1, −1 (0, 0, 1, 1), (1, 1, 0, 0), (0, 0, −1, 1), (−1, 1, 0, 0)
√ √
−1− 5 5−1
FIG. 16. Flat bands σ4 = 2
(left) and σ2 = 2
(right).
If we remove the middle vertical links in Fig. 16 between v2 and v3 , we obtain the flat bands −1 and 1, corresponding to (−1, 1, −1, 1)⊺
and (−1, −1, 1, 1)⊺ . Here, V f = G7 .
15 October 2024 06:43:54

We now exclude the remaining eigenvalues. In view of Lemma 4.5, we know that we can remove the top eigenvalues of connected graphs
′
Gk . Next, we exclude α′ as there are no δi ∈ {0, 1} such that α2 (δ1 + δ3 ) = −(δ2 + δ4 ) except all δj = 0, the LHS being irrational. Similarly, we
√
see that − 3, σ3 ∉ F s4 . From the values of βi , κi , we see that there are no δk ∈ {0, 1} such that κj (δ1 + δ2 ) + δ3 βj + δ4 = 0 except all δj = 0, so
βi ∉ F s4 .
2. Case of five vertices

By Corollary 4.7 and the case ν = 4, all the listed values are in F s5 . We provide some explicit graphs in Figs. 17 and 18 for illustration.
We now show that there are no more flat bands. By Theorem 4.9, it suffices to investigate the 21 connected graphs on five vertices, listed
in Fig. 19. We compute the eigenvalues and eigenvectors in Table III.
FIG. 17. Flat bands λ = 1 (left) and λ = −2 (right).
√ √
−1− 5 −1+ 5
FIG. 18. Flat bands λ = 2
(left) and λ = 2
(right).
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-20

Journal of
FIG. 19. All connected graphs on five vertices.
15 October 2024 06:43:54

TABLE III. Eigenvalues and eigenvectors of connected five-graphs.
H1 4, −1, −1, −1, −1 (1, 1, 1, 1, 1), (−1, 0, 0, 0, 1), (−1, 0, 0, 1, 0), (−1, 0, 1, 0, 0), (−1, 1, 0, 0, 0)
H2 3, 0, 0, −1, −2 (3, 2, 3, 2, 2), (0, −1, 0, 0, 1), (0, −1, 0, 1, 0), (−1, 0, 1, 0, 0), (−1, 1, −1, 1, 1)
H3 2, 0, 0, 0, −2 (2, 1, 1, 1, 1), (0, −1, 0, 0, 1), (0, −1, 0, 1, 0), (0, −1, 1, 0, 0), (−2, 1, 1, 1, 1)
H4 2, σ 2 , σ 2 , σ 4 , σ 4 (1, 1, 1, 1, 1), (σ 2 , σ 3 , −1, 0, 1), (−1, σ 3 , σ 2 , 1, 0), (−1, σ 1 , σ 4 , 1, 0), (σ 4 , σ 1 , −1, 0, 1)
′ ′ ′ ′
H5 √a, 0, a , −1, −1 √ √ (ã, ã,√ã, 1, 1), (0, 0, 0, −1, 1), (ã , ã , ã , 1, 1), (−1, 0, 1, 0, 0), (−1, √ 1, 0, 0, 0)√
H6 3, 1, 0, −1, − 3 (1, 3, 2, 3, 1), (−1, −1, 0, 1, 1), (1, 0, −1, 0, 1), (−1, 1, 0, −1, 1), (1, − 3, 2, − 3, 1)
H7 b, 0, 0, b′ , −2 (−b′ , 1, 1, 1, 1), (0, 0, −1, 0, 1), (0, −1, 0, 1, 0), (−b, 1, 1, 1, 1), (0, −1, 1, −1, 1)
H8 α, 1, −1, −1, α′ √
(−α ′
, 1,√1, 1, 1), (0, −1, −1, 1, 1), (0, 0, 0, −1, 1), (0, −1, 1, 0, 0), (−α, √
1, 1, 1,√1)
√ √
H9 6, 0, 0, 0, − 6 ( 2 , 1, 1, 2 , 1), (0, −1, 0, 0, 1), (−1, 0, 0, 1, 0), (0, −1, 1, 0, 0), ( 2 , 1, 1, − 2 6 , 1)
6 6 − 6
′
H 10 c, 0, σ 2 , c , σ 4 (c, 2, c, 1, 1), (0, −1, 0, 1, 1), (σ 2 , 0, σ 3 , −1, 1), (c′ , 2, c′ , 1, 1), (σ 4 , 0, σ 1 , −1, 1)
H 11 d1 , d2 , −1, −1, d3 (d1 , d1′ , d1′ , d1′ , 1), (d2 , d2′ , d2′ , d2′ , 1), (0, −1, 0, 1, 0), (0, −1, 1, 0, 0), (d3 , d3′ , d3′ , d3′ , 1)
H 12 e1 , e2 , 0, −1, e3 (h1 , i1 , i1 , h1 , 1), (h2 , i2 , i2 , h2 , 1), (−1, 0, 0, 1, 0), (0, −1, 1, 0, 0), (h3 , i3 , i3 , h3 , 1)
H 13 j1 , j2 , 0, −1, j3 (k1 , ℓ1 , ℓ1 , k1 , 1), (k2 , ℓ2 , ℓ2 , k2 , 1), (−1, 0, 0, 1, 0), (0, −1, 1, 0, 0), (k3 , ℓ3 , ℓ3 , k3 , 1)
H 14 m1 , m2 , −1, −1, m3 (n1 , 1, p1 , p1 , 1), (n2 , 1, p2 , p2 , 1), (0, −1, 0, 0, 1), (0, 0, −1, 1, 0), (n3 , 1, p3 , p3 , 1)
H 15 q1 , q2 , 0, q3 , −2 (r1 , 1, s1 , s1 , 1), (r2 , 1, s2 , s2 , 1), (0, −1, −1, 1, 1), (r3 , 1, s3 , s3 , 1), (0, −1, 1, −1, 1)
H 16 t1 , σ 2 , t2 , t3 , σ 4 (u1 , 1, v1 , v1 , 1), (0, −1, σ 3 , σ 2 , 1), (u2 , 1, v2 , v2 , 1), (u3 , 1, v3 , v3 , 1), (0, −1, σ 1 , σ 4 , 1)
H 17 w1 , 1, w2 , −1, w3 (x1 , y1 , y1√ , z1 , 1), (0, −1, −1, 2, √2), (x2 , y2 , y2 , z2 , 1), (0, −1, 1, 0,√ 0), (x3 , y3 , y3 , z3 ,√1)
H 18 ζ 1 , ζ 2 , 0, ζ 3 , ζ 4 (1, ζ1 , 1, ζ2 , 2), (1, ζ2 , 1, ζ4 , − 2), (−1, 0, 1, 0, 0), (1, ζ3 , 1, ζ1 , − 2), (1, ζ4 , 1, ζ3 , 2)
H 19 η1 , η2 , 0, η3 , η4 (η1 , μ1 , ξ 1 , μ1 , 1), (η2 , μ2 , ξ 2 , μ2 , 1), (0, −1, 0, 1, 0), (η3 , μ2 , ξ 3 , μ2 , 1), (η4 , μ1 , ξ 4 , μ1 , 1)
H 20 Ϛ1 , Ϛ2 , 0, Ϛ3 , Ϛ4 (Ϛ1 , ϕ1 , τ 1 , ϕ1 , 1), (Ϛ2 , ϕ2 , τ 2 , ϕ2 , 1), (0, −1, 0, 1, 0), (Ϛ3 , ϕ3 , τ 1 , ϕ3 , 1), (Ϛ4 , ϕ4 , τ 2 , ϕ4 , 1)
H 21 χ 1 , χ 2 , χ 3 , −1, χ 4 (χ 1 , ψ 1 , ω1 , ψ 1 , 1), (χ 2 , ψ 2 , ω2 , ψ 2 , 1), (χ 3 , ψ 3 , ω3 , ψ 3 , 1), (0, −1, 0, 1, 0), (χ 4 , ψ 4 , ω4 , ψ 4 , 1)
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-21

Journal of
TABLE IV. Equations satisfied by the parameters and their approximate values.
Equations Approximate values

3 2
d − 2d − 4d + 2 = 0, d1 ≈ 3.086 13, d2 ≈ 0.428 007, d3 ≈ −1.514 14,
dd′ = d + 2d′ , d2 = 3d′ + 1 d1′ ≈ 2.841 4, d2′ ≈ −0.272 27, d3′ ≈ 0.430 87
e3 − e2 − 6e + 2 = 0, e1 ≈ 2.855 77, e2 ≈ 0.321 637, e3 ≈ −2.177 41,
h = 2e , eh = 2i + 1, h1 ≈ 1.427 89, h2 ≈ 0.160 819, h3 ≈ −1.088 7
ei = e + i i1 ≈ 1.538 86, i2 ≈ −0.474 137, i3 ≈ 0.685 278
j3 − j2 − 4j + 2 = 0 j1 ≈ 2.342 92, j2 ≈ 0.470 683, j3 ≈ −1.813 61,
k = 1j , ℓ + 1 = jℓ k1 ≈ 0.426 817, k2 ≈ 2.124 57, k3 ≈ −0.551 388
ℓ1 ≈ 0.744 644, ℓ2 ≈ −1.889 23, ℓ3 ≈ −0.355 416
m3 − 2m2 − 5m + 2 = 0 m1 ≈ 3.323 4, m2 ≈ 0.357 926, m3 ≈ −1.681 33
n = m2 , p + 2 = mp, n1 ≈ 0.601 793, n2 ≈ 5.587 74, n3 ≈ −1.189 53
n + 2p + 1 = m p1 ≈ 0.860 806, p2 ≈ −3.114 91, p3 ≈ −0.745 898
q3 − 2q2 − 2q + 2 = 0 q1 ≈ 2.481 19, q2 ≈ 0.688 892, q3 ≈ −1.170 09,
r = q2 , qs = s + 1 r1 ≈ 0.806 063, r2 ≈ 2.903 21, r3 ≈ −1.709 28,
q=r+s+1 s1 ≈ 0.675 131, s2 ≈ −3.214 32, s3 ≈ −0.460 811
t 3 − t 2 − 5t − 2 = 0 t 1 ≈ 2.935 43, t 2 ≈ −0.462 598, t 3 ≈ −1.472 83
t = u + v, tu = 2 + 2v, u1 ≈ 1.594 77, u2 ≈ 0.699 104, u3 ≈ −1.793 87
tv = u + v + 1 v1 ≈ 1.340 67, v2 ≈ −1.161 7, v3 ≈ 0.321 037
w3 − 4w − 2 = 0 w1 ≈ 2.214 32, w2 ≈ −0.539 189, w3 ≈ −1.675 13
z = w1 , wx = 2y + 1, x1 ≈ 1.762 71, x2 ≈ 1.315 45, x3 ≈ −1.078 16
wy = x + y, w = x + z y1 ≈ 1.451 61, y2 ≈ −0.854 638, y3 ≈ 0.403 032
z1 ≈ 0.451 606, z2 ≈ −1.854 64, z3 ≈ −0.596 968
χ 4 − χ 3 − 5χ 2 + χ + 2 = 0, χ 1 ≈ 2.641 19, χ 2 ≈ 0.723 742, χ 3 ≈ −0.589 216,
15 October 2024 06:43:54

χω = 2ψ, χ 2 = 2ψ + 1 χ 4 ≈ −1.775 71, ψ 1 ≈ 2.987 93, ψ 2 ≈ −0.238 099,
χψ = χ + ψ + ω ψ 3 ≈ −0.326 412, ψ 4 ≈ 1.076 58, ω1 ≈ 2.262 57,
ω2 ≈ −0.657 965, ω3 ≈ 1.107 96, ω4 ≈ −1.212 56
√ √ √ √ √ √
Notation in Table III, as mentioned before, σ1 = 1+2 5 , σ2 = −1+2 5 , σ3 = 1−2 5 , σ4 = −1−2 5 , α = 1+ 2 17 , and α′ = 1− 2 17 . In addition,
√ √ ′ √ √ √ √ √ √ √ √
a = 1 + 7, a′ = 1 − 7, ã = a3 , ã ′ = a3 , b = 1 + 5, b′ = 1 − 5, c = 1+ 2 13 , and c′ = 1− 2 13 . Next, ζ1 = 2 + 2, ζ2 = 2 − 2, ζ 3 = −ζ 2 , and
√ √ √ √ √ √ √ √ √ √
ζ 4 = −ζ 1 . Next, η1 = 5+ 2 17 , η2 = 5− 2 17 , η3 = −η2 , η4 = −η1 , μ1 = 3+ 4 17 , μ2 = 3− 4 17 , ξ1 = 7+2 17 , ξ2 = − 7−2
17
, ξ 3 = −ξ 2 , and ξ 4 = −ξ 1 .
√ √ √ √ √ √ √ √ √ √ √
1+ √5+2 2 −1+ √5−2 2 1− √ 5+2 2 −1− √5−2 2 Ϛ Ϛ3
Next, Ϛ1 = 2
, Ϛ2 = 2
, Ϛ3 = 2
, Ϛ4 = 2
, ϕ1 = √2 , ϕ2 = √2 , ϕ3 = √2 , ϕ4 = √2 , τ1 = 1 + 2, and τ2 = 1 − 2.
1 −Ϛ2 −Ϛ 4
For the remaining values, we give the equations they solve and numerical approximations in Table IV.
The only new eigenvalues that H 1 –H 4 can offer √ are′ top ones,
√ which cannot be in F s5 . Ignoring the top eigenvalue, each of H 5 –H 10
′ ′ ′
may only offer one possible flat band, namely, a , − 3, b , α , − 6, c , respectively. By looking at the eigenvectors, we see that there are no
δj ∈ {0, 1} such that ∑5j=1 δ j ψ j = 0, except all δj = 0. Hence, these values are not in F s5 .
The same argument shows, in fact, that none of the graphs H j , j > 5, offers any flat band. Again, one looks at the eigenvalues outside the
list in (4.3), considers the corresponding eigenvectors ψ, and checks by hand that ∑5j=1 δ j ψ j = 0 for δj ∈ {0, 1} implies δj = 0 for all j. For this,
one can use the numerical approximations of the different quantities in Table IV.
REFERENCES
1
Anantharaman, N. and Sabri, M., “Recent results of quantum ergodicity on graphs and further investigation,” Ann. Fac. Sci. Toulouse: Math. 28, 559–592 (2019).
2
Aomoto, K., “Point spectrum on a quasihomogeneous tree,” Pac. J. Math. 147, 231–242 (1991).
3
Banks, J., Garza-Vargas, J., and Mukherjee, S., “Point spectrum of periodic operators on universal covering trees,” Int. Math. Res. Not. 2022, 17713–17744.
4
Boutet de Monvel, A. and Sabri, M., “Ballistic transport in periodic and random media,” in From Complex Analysis to Operator Theory: A Panorama. In Memory of
Sergey Naboko, Operator Theory: Advances and Applications series (Springer/ Birkhäuser, 2023), Vol. 291.
5
Creutz, M., “Aspects of chiral symmetry and the lattice,” Rev. Mod. Phys. 73, 119–150 (2001).
6
Damanik, D., Embree, M., Fillman, J., and Mei, M., “Discontinuities of the integrated density of states for Laplacians associated with Penrose and Amman-Beenker
tilings,” Exp. Math. (unpublished).
7
Fillman, J., Liu, W., and Matos, R., “Irreducibility of the Bloch variety for finite-range Schrödinger operators,” J. Funct. Anal. 283, 109670 (2022).
8
Fillman, J., Liu, W., and Matos, R., “Algebraic properties of the Fermi variety for periodic graph operators,” arXiv:2305.06471.
9
Flach, S., Leykam, D., Bodyfelt, J. D., Matthies, P., and Desyatnikov, A. S., “Detangling flat bands into Fano lattices,” Europhys. Lett. 105, 30001 (2014).
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-22

Journal of
10
Higuchi, Y. and Nomura, Y., “Spectral structure of the Laplacian on a covering graph,” Eur. J. Combinatorics 30, 570–585 (2009).
11
Kato, T., Perturbation Theory for Linear Operators, reprint of the 1980 ed., Classics in Mathematics (Springer-Verlag, Berlin, 1995).
12
Kerner, J., Täufer, M., and Wintermayr, J., “Robustness of flat bands on the perturbed Kagome and the perturbed Super-Kagome lattice,” arXiv:2301.05076.
13
Kollár, A. J., Fitzpatrick, M., Sarnak, P., and Houck, A. A., “Line-graph lattices: Euclidean and non-Euclidean flat bands, and implementations in circuit quantum
electrodynamics,” Commun. Math. Phys. 376, 1909–1956 (2020).
14
Kollár, A. J. and Sarnak, P., “Gap sets for the spectra of cubic graphs,” Commun. Am. Math. Soc. 1, 1–38 (2021).
15
Korotyaev, E. and Saburova, N., “Schrödinger operators on periodic discrete graphs,” J. Math. Anal. Appl. 420, 576–611 (2014).
16
Korotyaev, E. and Saburova, N., “Spectral estimates for Schrödinger operators on periodic discrete graphs,” St. Petersburg Math. J. 30, 667–698 (2019).
17
Li, W. and Shipman, S. P., “Irreducibility of the Fermi surface for planar periodic graph operators,” Lett. Math. Phys. 110, 2543–2572 (2020).
18
Liu, W., “Irreducibility of the Fermi variety for discrete periodic Schrödinger operators and embedded eigenvalues,” Geom. Funct. Anal. 32, 1–30 (2022).
19
McKenzie, T. and Sabri, M., “Quantum ergodicity for periodic graphs,” Commun. Math. Phys. (to be published); arXiv:2208.12685.
20
Morfonios, C. V., Röntgen, M., Pyzh, M., and Schmelcher, P., “Flat bands by latent symmetry,” Phys. Rev. B 104, 035105 (2021).
21
Peyerimhoff, N. and Täufer, M., “Eigenfunctions and the integrated density of states on Archimedean tilings,” J. Spectral Theory 11, 461–488 (2021).
22
Rhim, J.-W. and Yang, B.-J., “Classification of flat bands according to the band-crossing singularity of Bloch wave functions,” Phys. Rev. B 99, 045107 (2019).
23
Röntgen, M., Morfonios, C. V., and Schmelcher, P., “Compact localized states and flat bands from local symmetry partitioning,” Phys. Rev. B 97, 035161 (2018).
24
Salez, J., “Every totally real algebraic integer is a tree eigenvalue,” J. Comb. Theory, Ser. B 111, 249–256 (2015).
25
Salez, J., “Spectral atoms of unimodular random trees,” J. Eur. Math. Soc. 22(2), 345–363 (2020).
26
Toikka, L. A. and Andreanov, A., “Necessary and sufficient conditions for flat bands in M-dimensional N-band lattices with complex-valued nearest-neighbour
hopping,” J. Phys. A: Math. Theor. 52, 02LT04 (2019).
27
Wilcox, C. H., “Theory of Bloch waves,” J. Anal. Math. 33, 146–167 (1978).
15 October 2024 06:43:54
J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-23


Sabri Youssef

Uploaded by

Copyright:

Available Formats

Sabri Youssef

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

Sabri Youssef

Uploaded by

Copyright:

Available Formats

RESEARCH ARTICLE | SEPTEMBER 01 2023

Flat bands of periodic graphs 

Mostafa Sabri  ; Pierre Youssef

J. Math. Phys. 64, 092101 (2023)

Articles You May Be Interested In

Spectral gaps of frustration-free spin systems with boundary

Analytical techniques for linear response formula of equilibrium states

15 October 2024 06:43:54

Flat bands of periodic graphs

Mostafa Sabri1 , 2,a) and Pierre Youssef2 , 3,b)

15 October 2024 06:43:54

J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-1

15 October 2024 06:43:54

Problem 2. Show that for any periodic connected Γ, the set

has dimension at most ν − 1.

J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-2

15 October 2024 06:43:54

J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-3

TABLE I. Index of symbols.

Γ An infinite Zd -periodic graph for some d

15 October 2024 06:43:54

(Uψ)p (θ) = ∑ e−2πiθ⋅k ψp (k) (2.4)

J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-4

for p = 1, . . . , ν satisfies the following.

H(θ) = (hi j (θ))νi, j=1 , (2.6)

Here, 0 ∉ I ii as Γ has no loops. If Q ≡ 0 and w ≡ 1, we denote A(θ) = H(θ).

15 October 2024 06:43:54

J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-5

We say that λ ∈ R is a flat band of H if λ is an eigenvalue of H.

hi = max (ki : k ∈ ∪p, j≤ν Ip j ).

If hi = 1 for all 1 ≤ i ≤ d, we speak of nearest-neighbor hopping.

15 October 2024 06:43:54

J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-6

15 October 2024 06:43:54

J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-7

A. Flat band generators

15 October 2024 06:43:54

FIG. 3. The Cartesian product of the graph in Fig. 7 and GF .

FIG. 4. General procedure (left), example GF = P2 , λ = −1 (right).

J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-8

FIG. 5. Nearest hopping, ν = 3. Flat bands λ = −2 (left) and λ = 0 (right).

IV. SINGLE-CELL FLAT BANDS

15 October 2024 06:43:54

Motivated by Fig. 1, we study the following neighborhood condition.

N vi /{v j } = N v j /{vi } ⇔ Iii = I j j = I ji /{0} = Ii j /{0} and Iir = I jr ∀r ≠ i, j.

This condition also implies that deg(vi ) = deg(vj ).

FIG. 6. The graphs Z ◻ P2 (left) and Z ◻ C4 (right).

J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-9

⎛ h11 (θ) h11 (θ) + ϵ gθ

15 October 2024 06:43:54

Proof. If ψ is an eigenvector entirely supported in V f , then

J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-10

15 October 2024 06:43:54

Corollary 4.7. The sets F sν are increasing: F sν ⊆ F sν+1 for all ν.

J. Math. Phys. 64, 092101 (2023); doi: 10.1063/5.0156336 64, 092101-11

G1 2, −1, −1 (1, 1, 1), (−1, 0, 1), (−1, 1, 0),

Theorem 4.9. We may write F sν+1 as a disjoint union,

15 October 2024 06:43:54

FIG. 9. All graphs on three vertices.