Discrete Mathematics 312 (2012) 129–147

Contents lists available at SciVerse ScienceDirect

Discrete Mathematics
journal homepage: www.elsevier.com/locate/disc

State transfer on graphs

Chris Godsil
Combinatorics & Optimization, University of Waterloo, Canada

article info abstract

Article history: If X is a graph with adjacency matrix A, then we define H (t ) to be the operator exp(itA). We
Available online 31 July 2011 say that we have perfect state transfer in X from the vertex u to the vertex v at time τ if the
uv -entry of |H (τ )u,v | = 1. State transfer has been applied to key distribution in commercial
Keywords:
cryptosystems, and it seems likely that other applications will be found. We offer a survey
Graph spectra
of some of the work on perfect state transfer and related questions. The emphasis is almost
Perfect state transfer
Continuous quantum walk entirely on the mathematics.
© 2011 Elsevier B.V. All rights reserved.

1. Perfect state transfer

Let X be a graph on n vertices with adjacency matrix A and let H (t ) denote the matrix-valued function exp(iAt ). If u and
v are distinct vertices in X , we say that perfect state transfer from u to v occurs if there is a time τ such that |H (τ )u,v | = 1.
(We will occasionally use ‘‘pst’’ as an abbreviation for ‘‘perfect state transfer’’.) We say that X is periodic relative to a vertex
u if there is a time τ such that |H (τ )u,u | = 1, and we say X itself is periodic if there is a time τ such that |H (τ )u,u | = 1 for all
vertices u.
We can use the complete graph K2 as an illustration. Here
 
0 1
A= ,
1 0

so An equals I if n is even and equals A if n is odd. Consequently

cos(t ) i sin(t )
 
H (t ) = cos(t )I + i sin(t )A =
i sin(t ) cos(t )

and hence
 
0 i
H (π /2) = .
i 0

This shows that we have perfect state transfer from u to v at time π /2. We also see that X is periodic with period π (because
H (π) = −I). Of course at t = π /2 we also have perfect state transfer from v to u; we will see that this is not an accident.
We note two properties of H (t ):
(a) Since A is symmetric, H (t ) is symmetric.
(b) Since exp(itA) = exp(−itA), we find that H (t ) is unitary.

Lemma 1.1. If we have perfect state transfer on X from u to v at time τ , then we have perfect state transfer from v to u at the
same time, and X is periodic at u and v with period dividing 2τ .

E-mail address: [email protected].

0012-365X/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.disc.2011.06.032
130 C. Godsil / Discrete Mathematics 312 (2012) 129–147

Proof. If u ∈ V (X ), let |u⟩ denote the vector that is one on u and zero elsewhere. (So |u⟩ is the characteristic vector of u,
viewed as a subset of V (X ).) If we have pst from u to v at time τ then there is a complex number γ of norm 1 such that
H (τ )|u⟩ = γ |v⟩.
But this says that H (τ )u,v = γ and, since H (τ ) is symmetric, we have H (τ )v,u = γ and therefore there is pst from v to u at
time τ .
Now we see that
H (τ )2 |u⟩ = γ 2 |u⟩, H (τ )2 |v⟩ = γ 2 |v⟩
and since H (τ )2 = H (2τ ), we have proved our second claim. 
If |H (t )u,v | = 1 then since H (t ) is unitary, the uv -entry of H (t ) is the only non-zero entry in the u-column and the only
non-zero entry in the v row. Hence if we have perfect state transfer from 1 to 2 at time τ , then H (τ ) has the form
 
T 0
H (τ ) =
0 H1
where H1 is unitary and
 
0 1
T =γ
1 0
with |γ | = 1.
The theory of perfect state transfer starts with the papers Bose [11] and Christandl et al. [18]; we will refer to the latter
frequently.
We note that our matrix H (t ) determines what is known as a continuous quantum walk, for background on this we refer
the reader to [31,32]. Physicists use exp(−itA) where we have used exp(itA); this makes absolutely no difference to the
theory, which is what we care about here. For recent surveys on state transfer see [33,38,29].

2. Spectral decomposition

The main tool we use is spectral decomposition of symmetric matrices. Suppose A is symmetric with distinct eigenvalues
θ1 , . . . , θm
and let Er denote orthogonal projection on the eigenspace belonging to θr . Then
Er2 = Er = ErĎ
and if f is a complex-valued function defined on the eigenvalues of A,
m
−
f (A) = f (θr )Er .
r =1

Taking f to be the exponential matrix we obtain the basic identity

m
−
H (t ) = exp(iθr t )Er .
r =1

One important consequence of this is that, for each t, the matrix H (t ) is a polynomial in A. Thus it commutes with A and,
more generally, with any matrix that commutes with A. Further
m
−
H (t )|u⟩ = exp(iθr t )Er |u⟩
r =1

where the non-zero vectors Er |u⟩ are eigenvectors for A (and H). The set of eigenvalues θr such that Er |u⟩ ̸= 0 is the eigenvalue
support of the vector |u⟩.

Lemma 2.1. Let E1 , . . . , Em be the idempotents in the spectral decomposition of A(X ) and let θ1 , . . . , θm be the corresponding
eigenvalues. Then there is perfect state transfer from u to v at time τ if and only if there is a constant γ such that
Er |u⟩ = γ exp(−iτ θr )Er |v⟩ (r = 1, . . . , m).

Proof. We have
H (τ )|u⟩ = γ |v⟩
if and only if, for all r,
γ Er |v⟩ = Er H (τ )|u⟩.
 C. Godsil / Discrete Mathematics 312 (2012) 129–147 131

Since
Er H (τ ) = H (τ )Er = exp(iθr τ )Er
the lemma follows. 

Corollary 2.2. If there is perfect state transfer from u to v , then Er |u⟩ = ±Er |v⟩, and accordingly u and v have the same eigenvalue
support.
Proof. If we have state transfer from u to v , then Er |u⟩ = β Er |v⟩ where |β| = 1. As Er |u⟩ and Er |v⟩ are both real, β = ±1. 

For later use we note some properties of the eigenvalue support of a vertex, but to prove these we will need to provide
an expression for the idempotents Er . If θ1 , . . . , θm are the distinct eigenvalues of A, define the polynomial pk (t ) by
∏ t − θr
pk (t ) = .
θ − θr
r ̸=k k

Then it is not hard to verify that Er = pr (A).

Lemma 2.3. Suppose u ∈ V (X ) and S is its eigenvalue support. If θ ∈ S then all algebraic conjugates of θ are in S. If X is bipartite
and θ ∈ S then −θ ∈ S. The spectral radius of the connected component of X that contains u belongs to S.
Proof. If X is not connected, then the elements of S are eigenvalues of the connected component of X that contains u, and
the associated eigenvectors are zero on vertices not in this component. So we may assume X is connected. If θr and θs are
algebraic conjugates then Er = pr (A) and Es = ps (A) are algebraic conjugates and so Er |u⟩ ̸= 0 if and only if Es |u⟩ ̸= 0. The
eigenvalue belonging to the spectral radius of a connected graph is simple and the corresponding eigenvector has no entry
zero. It follows that no entry of the associated idempotent is zero, which implies the third claim.
Suppose X is bipartite on n vertices. Let D be the n × n diagonal matrix such that Dv,v = 1 if v is at even distance from u,
and Dv,v = −1 otherwise. If DAD = −A and if Az = θ z then ADz = −θ Dz. 

3. Period

If we have perfect state transfer from u to v in X at time τ , then X is periodic at u with period 2τ . Our next result shows
that minimum time at which perfect state transfer involving u occurs is determined by the minimum period at u.
Some preliminaries. Assume X is connected and let T denote the set of times τ such that H (τ )eu is a scalar multiple of eu .
Then T is an additive subgroup of R. Since for small t,
H (t ) ≈ I + itA
and since Aeu ̸= 0, it follows that T is a discrete subgroup of R. Hence it is cyclic, generated by the minimum period of X at
u. If we have perfect state transfer from u to v at time τ then X is periodic at u with period 2τ , and hence τ ≥ σ /2.

Lemma 3.1. Suppose X is a connected graph and X is periodic at u with minimum period σ . Then if there is perfect state transfer
from u to v , there is perfect state transfer from u to v at time σ /2.
Proof. Suppose we have uv -pst with minimum time τ . Then X is periodic at u, with minimum period σ (say).
If σ < τ , then H (τ − σ )eu = γ ev for some γ and so τ is not minimal. Hence τ < σ . Since X is periodic at τ with period
2τ , we see that σ ≤ 2τ . If σ < 2τ then u is periodic with period dividing 2τ − σ and so σ ≤ 2τ − σ , which implies that
σ ≤ τ . We conclude that σ = 2τ .
Thus if the minimum period of X at u is σ and there is perfect state transfer from u to v , then there is perfect state transfer
from u to v at time σ /2 (and not at any shorter time). 
We have the following corollary, due to Kay [30, Section IIID]:

Corollary 3.2. If we have perfect state transfer in X from u to v and also from u to w , then v = w . 

It is actually possible to derive a lower bound on the minimum period in terms of the eigenvalues of X .

Lemma 3.3. If X is a graph with eigenvalues θ1 , . . . , θm and transition matrix H (t ). If x is a non-zero vector, then the minimum
π
time τ such that xT H (τ )x = 0 is at least θ −θ .
1 m

Proof. Assume ‖x‖ = 1. We want

eit θs xT Es x,
−
0 = xT H (t )x =

where the sum is over the eigenvalues θs such that Er x ̸== 0, i.e., over the eigenvalue support of x. Since
−
1 = xT x = xT Es x,
132 C. Godsil / Discrete Mathematics 312 (2012) 129–147

the right side is a convex combination of complex numbers of norm 1. When t = 0 these numbers are all equal to 1, and as
t increases they spread out on the unit circle of radius. If they are contained in an arc of length less than π , their convex hull
cannot contain 0, and for small(ish) values of t, they lie in the interval bounded by t θ1 and t θm . So for xT HX (t )x to be zero,
we need t (θ1 − θm ) ≥ π , and thus we have the constraint
π
t ≥ . 
θ1 − θm

If u ∈ V (X ) and x = |u⟩, then this bound is tight for P2 but not for P3 .

π
Lemma 3.4. If X is a graph with eigenvalues θ1 , . . . , θm , the minimum period of X at a vertex is at least θ 2−θ .
1 m

Proof. We want
eiθr t (Er )u,u ,
−
γ =
r

where ‖γ ‖ = 1, and for this to hold there must be integers mr ,s such that
t (θr − θs ) = 2mr ,s π .
This yields the stated bound. 

In the previous lemma, θ1 is the spectral radius of A(X ). However θm can be replaced by the least eigenvalue in the
eigenvalue support. If the entries of x are non-negative, these comments apply to Lemma 3.3 too. For more bounds along
the lines of the last two lemmas, go to [30, Section IIIC].

4. More examples

Our theory is developing nicely, but as yet we have just one example of perfect state transfer. We describe a second, also
from [18].
√
Lemma 4.1. There is perfect state transfer between the end-vertices of the path on three vertices at time π / 2.
√ √
Proof. The eigenvalues of P3 are 2, 0, − 2 with respective eigenvectors
   
√1 1 1
 
1 1 1 √
 2 , √ 0 , − 2 .
2 2 −1 2
1 1

If we denote these vectors by z1 , z2 , z3 respectively, then

Er = zr zrT .
Then
√ √
H (t ) = exp(it 2)E1 + E2 + exp(−it 2)E3
and consequently
0 0 −1
 
√
H (π / 2) = −E1 + E2 − E3 = 0 −1 0 . 
−1 0 0.

If X and Y are graphs then their Cartesian product X Y is defined as follows. Its vertex set is V (X ) × V (Y ), and (x1 , y1 ) is
adjacent to (x2 , y2 ) if either
(a) x1 = x2 and y1 is adjacent to y2 , or
(b) x1 is adjacent to x2 and y1 = y2 .
Thus the Cartesian product of the paths Pm and Pn is the m × n grid. We use X d to denote the dth Cartesian power of
X —the Cartesian product of d copies of X . The dth Cartesian power of P2 is the d-cube Qd .
The theory of the Cartesian product is very well developed, for details for [28]. We could have defined the Cartesian
product using adjacency matrices:
A(X Y ) = A(X ) ⊗ I + I ⊗ A(Y ).
This expresses A(X Y ) as the sum of two commuting matrices, and hence allows one to prove the following (due once again
to Christandl et al.).
 C. Godsil / Discrete Mathematics 312 (2012) 129–147 133

Lemma 4.2. For any graphs X and Y we have

HA(X Y ) (t ) = HA(X ) (t ) ⊗ HA(Y ) (t ).

This lemma is particularly important in physical terms, because it implies that a physical system modeled by X Y is a
composite of the systems modeled by X and Y . It also means that we now have infinitely many examples where perfect state
transfer occurs. The vertices of the dth Cartesian power of X are the d-tuples in V (X )d , and if u and v are two such d-tuples,
then the distance between them is
d
−
distX (ur , vr ).
r =1

Theorem 4.3. If we have perfect state transfer from u to v in X at time τ , then at time τ we have perfect state transfer in the dth
Cartesian power of X between the d-tuples
(u, . . . , u), (v, . . . , v). 

Since we have perfect state transfer on P2 and P3 , we might naturally expect that perfect state transfer is possible on all
paths. We will see that this is false.

5. Periodicity

We have seen that the existence of perfect state transfer implies periodicity. Studying periodicity is an effective stepping
stone to the study of state transfer, and so we take this step.
The first thing to note is that
−
Er = I
and so X itself will be periodic if there is a time τ and a scalar γ with |γ | = 1 such that
ei τ θ r = γ , r = 1, . . . , m.
Certainly taking τ to be zero works. What is much more interesting is that if the eigenvalues of X are integers then τ = 2π
works: if the eigenvalues of X are integers then X is periodic with period dividing 2π . The path P2 = K2 is an example.
With only a little more thought we see that if there is a number δ such that θr /δ ∈ Q for all r, then X is periodic with
period dividing 2π /δ . The basic question is to what extent this rationality condition is necessary.
The ratio condition is a necessary condition for a graph to be periodic at a vertex. The version we offer here is stated as
Theorem 2.2 in [23]; it is an extension of result from [37], which in turn extends an idea used in [18].

Theorem 5.1. Let X be a graph and let u be a vertex in X at which X is periodic. If θk , θℓ , θr , θs are eigenvalues in the support of
|u⟩ and θr ̸= θs , then
θk − θℓ
∈ Q. 
θr − θs
Using this one can prove that a graph is periodic if and only if the ratio of any two eigenvalues is rational, and this leads
to the following result from [23].

Theorem 5.2. A graph X is periodic if and only if either:

(a) The eigenvalues of X are integers, or √
(b) The eigenvalues of X are rational multiples of ∆, for some square-free integer ∆.
If the second alternative holds, X is bipartite. 
If X is regular then its spectral radius is an integer and so (b) cannot hold. Thus a regular graph is periodic if and only if
its eigenvalues are integers.
Since our actual concern is perfect state transfer, not periodicity, it will only be useful if there are interesting cases where
perfect state transfer implies periodicity (and just periodicity at a vertex). We take this up in the next section. Note that
there are graphs with perfect state transfer that are not periodic. Stevanović observes that the bipartite complements of an
even number of copies of P3 provide a family of examples, and another class is presented in [4]. (A bipartite graph Y is a
bipartite complement of a bipartite graph X with bipartition (A, B) if the edge set E (Y ) is the complement of E (X ) in the edge
set of the complete bipartite graph with bipartition (A, B).)

6. Vertex-transitive graphs

An automorphism of the graph X is a permutation α of V (X ) such that the vertices uα and v α are adjacent if and only if u
and v are. Any permutation can be represented by a permutation matrix, and a permutation matrix P is an automorphism
134 C. Godsil / Discrete Mathematics 312 (2012) 129–147

of X if and only if it commutes with A(X ). The set of all automorphisms of X forms its automorphism group Aut(X ). Our
graph X is vertex transitive if Aut(X ) is transitive as a permutation group, that is, for each pair of vertices u and v there is an
automorphism α such that uα = v .
Cayley graphs form an important class of vertex-transitive graphs. To construct a Cayley graph for a group G we first
choose a subset C of G. The vertex set of the Cayley graph X (G, C ) is G, and elements g and h of G are adjacent if hg −1 ∈ C .
We call C the connection set, and we do not assume that C generates G (so X might not be connected). To avoid loops and
multiple edges we do assume that 1 ̸∈ C and the C is inverse-closed; if g ∈ C then g −1 ∈ C . If a ∈ G then the map that send
g to ga is an automorphism of G. In fact G acts regularly on V (X ) by right multiplication, and so any Cayley graph is vertex
transitive.
There are two classes of Cayley graphs which are important to us. If G is the cyclic group Zn , then a Cayley graph for
G is a circulant. If G is the elementary abelian 2-group Zd2 , then X is a so-called cubelike graph. The cycle on n vertices is a
circulant with connection set {1, −1}, and the d-cube is a cubelike graph. Note that if G is abelian we are using + as our
group operation.
For vertex-transitive graph, the existence of perfect state transfer has very strong consequences, as shown by the
following result. (This is a consequence of [23, Theorem 4.1].)

Theorem 6.1. Suppose X is a connected vertex-transitive graph with vertices u and v , and perfect state transfer from u to v occurs
at time τ . Then H (τ ) is a scalar multiple of a permutation matrix with order two and no fixed points, and it lies in the center of
the automorphism group of X . 
An immediate consequence is that if perfect state transfer takes place on a vertex-transitive graph X , then |V (X )| is even.
We can weaken the assumption that X is vertex transitive in this theorem it is enough that A(X ) should belong to a
homogeneous coherent algebra. A coherent algebra is a vector space of matrices that is closed under both the usual matrix
multiplication and under Schur multiplication and contains I and J. Such an algebra has a unique basis of 01-matrices and it
is homogeneous if I is an element of this basis (rather than a sum of elements). The adjacency matrix of a vertex-transitive
graph belongs to a homogeneous coherent algebra, and so does the adjacency matrix of a distance-regular graph. For details
see [23, Theorem 4.1].
For vertex-transitive graphs we can specify the true period: if the eigenvalues are integers and 2e is the largest power of
2 that divides the greatest common divisor of the eigenvalues then the period is π /2d−1 .

7. Cayley graphs of abelian groups

In investigations of state transfer, Cayley graphs of abelian groups provide a useful test bed, because it is easy to compute
their eigenvalues and eigenvectors.
Another advantage of this class is that we can decide which vertices might be involved in state transfer. Each finite group
G gives rise to two regular permutation groups: the group of permutations given by right multiplication on G and the group
of permutations given by left multiplication. If n = |G|, these give two regular subgroups of the symmetric group Sym(n)
and each element in one group commutes with each element in the other. The intersection of these two groups consists of
the elements in the center of G. (So if G is abelian, the two groups are equal.) If T denotes the permutation matrix arising in
Theorem 6.1, then we have the following extension of this theorem:

Lemma 7.1. If X is a Cayley graph for a group G and perfect state transfer occurs at time τ , then T lies in the center of G. 

Even when G is abelian this is useful, because it tells us that T is an element of G. So if we have perfect state transfer on
a Cayley graph for an abelian group G, then it maps the vertex g to g + c for some element of order two in Z (G). If G is the
cyclic group of order n = 2ν , then perfect state transfer must send 0 to ν (and a to a + ν ). If G is cyclic then this element of
order two is unique (it if it exists).

7.1. Cubelike graphs

A cubelike graph is a Cayley graph for Zd2 . The adjacency matrix of a cubelike graph can written as sum of commuting
permutation matrices P such that P 2 = I and tr(P ) = 0. If
A = P1 + · · · + Pd
then
d
∏
H (t ) = exp(it (P1 + · · · + Pd )) = exp(itPr ).
r =1

But if P 2 = I then
exp(itP ) = cos(t )I + i sin(t )P
 C. Godsil / Discrete Mathematics 312 (2012) 129–147 135

and therefore
d
∏
H (t ) = (cos(t )I + i sin(t )Pr ).
r =1

Consequently H (π ) = (−1)d I and

d
∏
H (π /2) = id Pr .
r =1

Using these ideas we arrive at the following result from [9].

Lemma 7.2. Suppose C ⊆ Zd2 \ 0 and X = X (Zd2 , C ). Then X is periodic with period dividing π . Its period is equal to π if and
only the sum σ of the elements of C is not zero, and in this case we have perfect state transfer from 0 to σ at time π /2. 
The connection set of a cubelike graph on 2d vertices with valency m can be presented as a d × m matrix over Z2 with
distinct columns; the columns of the matrix are the connection set. If M is such a matrix then its row space is a binary code.
A binary code is even if the Hamming weight of any code word is even and it is easy to show that the code is even if and only
if M1 = 0, that, is, the columns of M sum to zero. Hence we have perfect state transfer on a cubelike graph at π /2 if and
only if its code is not even.
We may have perfect state transfer when the code is even. A binary code is doubly even if all code words have weight
divisible by four. The code is self-orthogonal if MM T = 0 and again it is not hard to show that a self-orthogonal code is
doubly even if and only if the weight of each row of M is doubly even. In [17] it is proved that if the code of a cubelike graph
is self-orthogonal and even, but not doubly even, then we have perfect state transfer at time π /4.

7.2. Circulants

A circulant is a Cayley graph for the cyclic group Zn . As we saw above there is no perfect state transfer if n is odd and, if
n = 2ν then any state transfer must be from a to a + ν .
However we have got ahead of ourselves—a circulant is periodic if and only its eigenvalues are integers. So we need to
determine when this holds. Computing the eigenvalues of a circulant is easy because we know its eigenvectors. Choose a
primitive nth root of unity in C, say ζ , and an integer s. Then the function that maps x in Zn to ζ sx is an eigenvector. If the
connection set is C , then the eigenvalue is
−
ζ sx .
x∈C

What is more surprising is that it is easy to characterize the connection sets C such that X (C ) has only integer eigenvalues.
Define two elements of an abelian group G to be equivalent if they generate the same subgroup of G. Bridges and Mena [13]
showed that the eigenvalues of X (G, C ) are integers if and only if C is a union of equivalence classes. For cyclic groups, two
elements are equivalent if and only if they have the same order.
In [7,6,8,36] Bašić, Petković and (in one case) Stevanović have investigated perfect state transfer on circulants. Among
their many results, they proved that if n is square-free or is congruent to 2 modulo 4, there is no perfect state transfer. More
recently Bašić [5] has completely characterized the circulants on which perfect state transfer occurs.
A bicirculant is a graph on 2n vertices admitting an automorphism of order n with two orbits of length n. (So the Petersen
graph is one example.) Any circulant of even order is a bicirculant but in general a bicirculant graph need not be vertex
transitive. If X is a bicirculant relative to an automorphism g, then the subgraphs induced by the orbits of g are circulants.
Angeles-Canul et al. [4] define a circulant join to be a bicirculant where the two induced circulants are isomorphic, and give
a condition for such a graph to admit perfect state transfer.

8. Equitable partitions

A partition π of the vertices of a graph X is equitable if, for each ordered pair of cells Ci and Cj from π , all vertices in Ci
have the same number of neighbors in Cj . (So the subgraph of X induced by a cell is a regular graph, and the subgraph formed
by the vertices of two cells and the edges which join them is bipartite and semiregular, that is, all vertices in the same color
class have the same valency.) We note that the orbits of any group of automorphisms of X form an equitable partition. The
discrete partition, with each cell a singleton, is always equitable; the trivial partition with exactly one cell is equitable if an
only if X is regular. For the basics concerning equitable partitions see for example [25, Section 9.3]. In particular the join of
two equitable partitions is equitable, and consequently given any partition of X , there is a unique coarsest refinement of
it—the join of all equitable partitions which refine it. (Note: here ‘‘join’’ refers to join in the lattices of partitions.)
If π is a partition of V (X ), its characteristic matrix is the 01-matrix whose columns are the characteristic vectors of the
cells of π , viewed as subsets of V (X ). If we scale the columns of the characteristic matrix so they are unit vectors, we obtain
the normalized characteristic matrix of π . If P and Q are respectively the characteristic and normalized characteristic matrix
136 C. Godsil / Discrete Mathematics 312 (2012) 129–147

of π , then P and Q have the same column space and Q T Q = I. The matrix QQ T is block diagonal, and its diagonal blocks are
all of the form
1
Jr
r
where Jr is the all-ones matrix of order r × r, and the size of the ith block is the size of the ith cell of π . The vertex u forms a
singleton cell of π if and only if Qeu = eu .
We use |π | to denote the number of cells of π .

Lemma 8.1. Suppose π is a partition of the vertices of the graph X , and that Q is its normalized characteristic matrix. Then the
following are equivalent:
(a) π is equitable.
(b) The column space of Q is A-invariant.
(c) There is a matrix B of order |π | × |π | such that AQ = QB.
(d) A and QQ T commute. 

If u ∈ V (G), we use ∆u to denote the partition of the vertices by distance from u. The following result is proved in [27].

Theorem 8.2. A graph X is distance regular if it is regular and for each vertex u in X , the distance partition ∆u is equitable.

(If X is not regular but all distance partitions are equitable, then it is a distance-biregular graph. For details see [27].)
Ge et al. [19] provide an application of the theory of equitable partitions unrelated to what we consider here.

9. Stabilizers

If u ∈ V (X ), then Aut(X )u denotes the group of automorphisms of X that fix u. Our next result says that if perfect state
transfer from u to v occurs, then any automorphism of X that fixes u must fix v (and vice versa).

Lemma 9.1. Suppose perfect state transfer from u to v occurs on X at time τ . If MA = AM and M |u⟩ = |u⟩ then M |v⟩ = |v⟩.

Proof. Since M must commute with Er ,

MEr |u⟩ = Er M |u⟩ = Er |u⟩.

By Lemma 2.1 it follows that MEr |v⟩ = Er |v⟩, and therefore M |v⟩ = |v⟩. 

Corollary 9.2. If X admits perfect state transfer from u to v , then Aut(X )u = Aut(X )v . 

By way of example, suppose X is a Cayley graph for an abelian group G. The map that sends g in G to g −1 is an
automorphism of G and also an automorphism of X . The fixed points of this automorphism are the elements of G with order
one or two. So if perfect state transfer from 1 to a second vertex a occurs, then a has order two. Consequently perfect state
transfer cannot occur on a Cayley graph for an abelian group of odd order. (Another proof of this is presented as Corollary
4.2 in [23].)
If there is perfect state transfer from u to v , then it follows from the previous corollary that the orbit partitions of Aut(X )u
and Aut(X )v are equal. Since equitable partitions can be viewed as a generalization of orbit partitions, it is not entirely
unreasonable to view the following result as an extension of this fact.

Corollary 9.3. Let u and v be vertices in X . If there is perfect state transfer from u to v , then ∆u = ∆v .

Proof. Let Q be the normalized characteristic matrix of the partition ∆u , let R = QQ T and assume H (t ) = exp(iAt ). Then
H (t ) is a polynomial and so RH (t ) = H (t )R. If we have perfect state transfer from u to v at time τ , then

H (τ )eu = γ ev

and accordingly

γ Rev = RH (τ )eu = H (τ )Reu = H (τ )eu = γ ev .

So Rev = ev and this implies that {v} is a cell of ∆u . 

By way of example, consider a distance-regular graph X with diameter d. If u ∈ V (X ), then ∆u is the distance partition
with respect to u with exactly d + 1 cells, where the ith cell is the set of vertices at distance i from u. We conclude that if
perfect state transfer occurs on X , then for each vertex u there is exactly one vertex at distance d from it. In particular perfect
state transfer does not occur on strongly regular graphs. (This also follows from some observations in [23, Section 4].)
 C. Godsil / Discrete Mathematics 312 (2012) 129–147 137

Lemma 9.4. Suppose C and D are cells of an equitable partition π of X and that we have perfect state transfer from u in C to v in
D at time τ . If y is the characteristic vector of the cell of π that contains u, then H (τ )y is the scalar multiple of the characteristic
vector of the cell that contains v . Further these two cells have the same size.

Proof. Let Q be the normalized characteristic matrix of the partition. Then

HQ |u⟩ = QH |u⟩ = γ Q |v⟩.

If w ∈ V (X ) then Q |w⟩ is the normalized characteristic vector of the cell containing w and so our first claim holds. Since H
is unitary, z and Hz have the same length. 

10. Finiteness

From [23] we know that if a graph is periodic, then the squares of its eigenvalues are integers and, if the graph is not
bipartite, the eigenvalues themselves are integers. A variant of this fact was derived in [21], it implies that if perfect state
transfer occurs on X , then the spectral radius of X is an integer or a quadratic irrational.

Theorem 10.1. Suppose X is a connected graph with at least three vertices where perfect state transfer from u to v occurs at time
τ . Let δ be the dimension of the A-invariant subspace generated √ by eu . Then δ ≥ 3 and either all eigenvalues in the eigenvalue
support of u are integers, or they are all of the form 12 (a + br ∆) where ∆ is a square-free integer and a and br are integers. 

Corollary 10.2. There are only finitely many connected graphs with maximum valency at most k where perfect state transfer
occurs.

Proof. Suppose X is a connected graph where perfect state transfer from u to v occurs at time τ and let S be the eigenvalue
√
support of u. If the eigenvalues in S are integers then |S | ≤ 2k + 1, and if they are not integers
√ then |S | < (2k + 1) 2. So
the dimension of the A-invariant subspace of RV (X ) generated by |u⟩ is at most ⌈(2k + 1) 2⌉, and this is also a bound on the
maximum distance from u of a vertex in X . If s ≥ 1, the number of vertices at distance s from u is at most k(k − 1)s−1 , and
the result follows. 

Corollary 10.3. Suppose X is a bipartite graph with spectral radius θ1 . If perfect state transfer takes place on X , then θ12 ∈ Z.

Proof. Suppose X is bipartite are we have perfect state transfer from u to v . Let S be the eigenvalue support of u. By
Theorem 10.1 there is a square-free integer ∆ and integers a and br such that each eigenvalue in S has the form
1 √
(a + br ∆).
2
√
We assume by way of contradiction that a ̸= 0. The spectral radius θ1 is then (a + br ∆)/2 and from Lemma 2.3 we see
that θ1 and −θ1 belong to S. This implies that a = 0. 

11. Cospectral vertices

We use φ(X , x) to denote the characteristic polynomial of A(X ). Vertices u and v in the graph X are cospectral if

φ(X \ u, t ) = φ(X \ v, t ).
Of course two vertices that lie in the same orbit of Aut(X ) are cospectral, but there are many examples of cospectral pairs of
vertices where this does not hold. A graph is walk regular if any two of its vertices are cospectral. Any strongly regular graph
is walk regular.
We note the following identities:
φ(X \ u, t ) −
= ((tI − A)−1 )u,u = (t − θr )−1 (Er )u,u .
φ(X , t ) r

(Here the first inequality is a consequence of Cramer’s rule and the second is spectral decomposition.) Since

(Er )u,u = ‖Er |u⟩‖2

we see that u and v are cospectral if and only the projections Er |u⟩ and Er |v⟩ have the same length for each r. So Lemma 2.1
yields immediately:

Lemma 11.1. If X admits perfect state transfer from u to v , then Er u = ±Er v for all r, and u and v are cospectral. 
138 C. Godsil / Discrete Mathematics 312 (2012) 129–147

We note that if the eigenvalues of A are simple and

‖Er |u⟩‖2 = ‖Er |v⟩‖2
then necessarily Er |u⟩ = ±Er |v⟩.
Let A be the adjacency matrix of the graph X on n vertices and suppose S is a subset of V (X ) with characteristic vector
|S ⟩. The walk matrix of S is the matrix with columns
|S ⟩, A|S ⟩, . . . , An−1 |S ⟩.
We say that the pair (X , S ) is controllable if this walk matrix is invertible. If u ∈ V (X ) then the walk matrix of u is just the
walk matrix relative to the subset {u} of V (X ). From our discussion at the start of the proof of Theorem 10.1, we have that
the rank of the walk matrix relative to u is equal to the size of the eigenvalue support of u. Thus if (X , u) is controllable, the
eigenvalues of A must be distinct.

Lemma 11.2. Let u and v be vertices of X with respective walk matrices Wu and Wv . Then u and v are cospectral if and only if
WuT Wu = WvT Wv . 
One consequence of this is that if u and v are cospectral and (X , u) is controllable, then so is (X , v). We also have the
following (from [26]):

Corollary 11.3. If u is a vertex in X with walk matrix Wu , then rk(Wu ) is equal to the number of poles of the rational function
φ(X \ u, x)/φ(X , x). Hence (X , u) is controllable if and only if φ(X \ u, x) and φ(X , x) are coprime.
Proof. By the spectral decomposition,
− 1
((xI − A)−1 )u,u = (Er )u,u . 
r
x − θr

The characteristic polynomial of the path Pn satisfies the recurrence

φ(Pn+1 , x) = xφ(Pn , x) − φ(Pn−1 , x)
and using this it easy to show that either end-vertex of a path is controllable. On the other hand, if X has an eigenvalue of
multiplicity greater than one then no vertex in X is controllable. The following result comes from [21].

Theorem 11.4. Let X be a connected graph on at least four vertices. If we have perfect state transfer between distinct vertices u
and v in X , then neither u nor v is controllable. 
For information on controllability, go to [20].

12. Transfer without exponentials

If u and v are controllable and cospectral we cannot get perfect state transfer between them. Our next result shows that
if these conditions hold, there is nonetheless a symmetric orthogonal matrix Q which commutes with A and maps |u⟩ to |v⟩.

Lemma 12.1. Let u and v be vertices in X with respective walk matrices Wu and Wv . If u and v are controllable and Q := Wv Wu−1
then Q is a polynomial in A. Further Q is orthogonal if and only if u and v are cospectral.
Proof. Let Cφ denote the companion matrix of the characteristic polynomial of A. Then
AWu = Wu Cφ
for any vertex u in X . Hence if u and v are controllable,
Wu−1 AWu = Wv−1 AWv
and from this we get that
AWv Wu−1 = Wv Wu−1 A.
Since X has a controllable vertex its eigenvalues are all simple, and so any matrix that commutes with A is a polynomial in
A. This proves the first claim.
From Lemma 11.2, the vertices u and v are cospectral if and only if WuT Wu = WvT Wv , which is equivalent to

Q −T = Wv−T WuT = Wv Wu−1 = Q . 

This lemma places us in an interesting position. If u and v are cospectral and controllable, there is an orthogonal matrix Q
that commutes with A such that Q |u⟩ = |v⟩ but, by the result of the previous section, Q cannot be equal to a scalar multiple
of H (t ) for any t.
 C. Godsil / Discrete Mathematics 312 (2012) 129–147 139

13. Pretty good state transfer

We say we have pretty good state transfer from u to v if there is a sequence {tk } of real numbers and a scalar γ such that

lim H (tk )|u⟩ = γ |v⟩.

k→∞

As an example consider P4 with eigenvalues

1 √ 1 √ 1 √ 1 √
θ1 = ( 5 + 1), θ2 = ( 5 − 1), θ3 = (− 5 + 1), θ4 = (− 5 − 1).
2 2 2 2
Then by straightforward computation
 
0 0 0 1
0 0 1 0
E1 − E2 + E3 − E4 =  ;
0 1 0 0
1 0 0 0

r (−1)
r −1
∑
denote the right side by T . (We could prove that T = Er by verifying that if zr is an eigenvector with eigenvalue
θr , then Tzr = (−1)r −1 zr .) Now choose integers a and b so that
√
a 1+ 5
≈ .
b 2
Then bθ1 ≈ a and

bθ2 = b(θ1 − 1) ≈ a − b, bθ3 ≈ b − a, bθ4 ≈ −a.

For example take a = 987 and b = 610 and set τ = 610π /2. Then the values of τ θr are (approximately)

987π /2, 377π /2, −377π /2, −987π /2

and these are congruent modulo 2π to

3π /2, π /2, 3π /2, π /2.

Hence

H (305π ) ≈ −iT .

(The approximation is accurate to five decimal places.)

In general we have to choose a and b so that a ∼
= 3 and b ∼
= 2 modulo 4. If fn is the nth Fibonacci number with f0 = f1 = 1
then modulo 4

= 2,
f4m+2 ∼ f4m+3 ∼
=3
√
and the ratios fn+1 /fn are the standard continued fraction approximation to ( 5 + 1)/2. We conclude that we have pretty
good state transfer between the end-vertices of P4 . We leave the reader the exercise of verifying
√ that there is also pretty
good state transfer between the end-vertices of P5 (for which the eigenvalues are 0, ±1, ± 3). In the next section we see
that we do not have perfect state transfer on Pn when n ≥ 4. Also, to get a good approximation to perfect state transfer on
P5 the numerical evidence is that t must be very large. This suggests that pretty good state transfer will not be a satisfactory
substitute for perfect state transfer in practice.
Dave Morris has noted the following in a private communication.

Lemma 13.1. If we have pretty good state transfer from u to v , then Er |u⟩ = ±Er |v⟩ for each r.

Proof. By assumption there is a sequence {tk } of real numbers such that

lim H (tk )|u⟩ = γ |v⟩.

k→∞

Since the unit circle is compact there is a subsequence {tℓ } and a complex number ζ such that exp(itℓ ) → ζ . Now

ζ Er |u⟩ = lim exp(itℓ θr )Er |u⟩ = lim Er H (tℓ )|u⟩ = γ Er |v⟩.

ℓ→∞ ℓ→∞

Since Er |u⟩ and Er |v⟩ are real vectors, the lemma follows. 
140 C. Godsil / Discrete Mathematics 312 (2012) 129–147

14. Paths

In [18] Christandl et al. proved that perfect state transfer between the end-vertices of a path on n vertices did not occur if
n ≥ 4. It is possible to extend their arguments to show that P2 and P3 are the only paths where perfect state transfer occurs
at all. Our approach has benefited from discussions with Dragan Stevanović.
To begin we note some simple properties of paths. First, the characteristic polynomials φ(Pn , x) satisfy the recurrence

φ(Pn+1 , x) = xφ(Pn , x) − φ(Pn−1 , x).

One consequence of this is φ(Pn+1 , x) and φ(Pn , x) are coprime. Since interlacing implies that any multiple eigenvalue of
Pn+1 must be an eigenvalue of Pn , we also see that the eigenvalues of a path are simple.
Given the Φ (Pn , x) and φ(Pn−1 , x) are coprime, it follows from Theorem 11.4 that we do not have perfect state transfer
between end-vertices in Pn when n ≥ 4. Using the identity

φ(Pm+n , x) = φ(Pm , x)φ(Pn , x) − φ(Pm−1 , x)φ(Pn−1 , x)

(which can be derived easily by induction from our recurrence above), it is not hard to show that φ(Pn , x) and φ(Pm , x) have
a non-trivial common factor if and only if m + 1 divides n + 1. This rules out many more possible cases of perfect state
transfer.
If n is odd then the stabilizer in Aut(Pn ) of the middle vertex of Pn has order two, while the stabilizer of any other vertex
is trivial. So the middle vertex cannot be involved in perfect state transfer.
But nothing we have mentioned will rule out perfect state transfer between (say) vertices 3 and 6 in P8 . We address this
problem now.
Let ζn (x) denote the vector
1
 
 φ(P1 , x) 
 .. .
.
 
φ(Pn−1 , x)
If the eigenvalues of Pn are

θ1 ≥ · · · ≥ θn
then ζn (θr ) is an eigenvector for Pn with eigenvalue θr . This is easy to verify using the recurrence, and using this we can also
see that two consecutive entries of ζr cannot be zero. Nor can the last entry be zero. We say there is a sign change at r in the
sequence (φ(Pr , x))rn= −1
0 if

φ(Pr −1 (x))φ(Pr (x)) < 0

or if φ(Pr (x)) = 0 and

φ(Pr −1 (x))φ(Pr +1 (x)) < 0.

From the recurrence we see that if φ(Pr (x)) = 0 then there is a sign change at r. It follows from Sturm’s theorem that there
are exactly m − 1 sign changes in the sequence (φ(Pn , θm ))nr = −1
0.
Let T be the permutation matrix representing the non-identity automorphism of Pn . Since the eigenvalues of the path
are simple, if z is an eigenvector for A = A(Pn ), then Tz is also an eigenvector and consequently Tz = ±z. Since the bottom
entry of ζn (θ2 ) must be negative, we have

T ζn (θ2 ) = −ζn (θ2 ).

We conclude that either no entry of ζn (θ2 ) is zero, or n is odd and the middle entry is zero. Hence if m is not the middle vertex
then neither E1 |m⟩ nor E2 |m⟩ is zero. If D is the n × n diagonal matrix with Dr ,r = (−1)r and z is an eigenvector of the path
with eigenvalue θ , then Dz is an eigenvector with eigenvalue −θ . Hence En−1 |m⟩ and En |m⟩ are not zero. By Theorem 5.1 we
have that
θ2 θn − θn−1
= ∈ Q.
θ1 θ1 − θn
The eigenvalues of Pn are the numbers
rπ
 
2 cos , r = 1, . . . , n
n+1
and it follows that

θ2 = θ12 − 2.
 C. Godsil / Discrete Mathematics 312 (2012) 129–147 141

Our rationality condition now implies that θ1 must be, at worst, a quadratic irrational. But by Corollary 10.3 we have that θ12
is an integer, and therefore θ2 must also be an integer. Since θ2 < 2 and θ1 > 0 we have
θ2 ∈ {−1, 0, 1}.
√ √
If θ2 = −1 then θ1 = 1 and X = K2 . If θ2 = 0, then θ1 = 2 and X = P3 . If θ2 = 1 then θ1 = 3 and according to the ratio
condition
√
θ1 − θ2 3−1 √
= √ =2− 3
θ1 + θ2 3+1
should be rational.
Depending on one’s mood, it is either instructive or depressing to see how much effort is needed to deal with perfect
state transfer on paths.

15. Joins

If X and Y are graphs let X + Y denote their join, which we get by taking a copy of X and a copy of Y and joining each
vertex in X to each vertex in Y . Angeles-Canul et al. [4,3] and the Ge et al. [19] provide many interesting results on perfect
state transfer in joins, including cases with weighted edges. Here we will focus simply on the joins of two regular graphs.
Assume X is k-regular on m vertices and Y is ℓ-regular on n vertices. Set A = A(X ) and B = A(Y ). If Z := X + Y then
 
A J
A(Z ) = .
JT B

If Az = θ z and 1T z = 0, then
   
z z
A(Z ) =θ .
0 0

Similarly, if Bz = θ z then
   
0 0
A(Z ) =θ .
z z

We see that n + m − 2 of the eigenvalues of X + Y are eigenvalues of X and eigenvalues of Y . The remaining two
eigenvalues are associated with eigenvectors that are constant on V (X ) and V (Y ). Assume that A and B have respective
spectral decompositions:
− −
A= θr Er , B= νs Fs
r s

where E1 and F1 are multiples of J. Then we have a decomposition

− −
A(Z ) = µ1 N1 + µ2 N2 + θr Êr + F̂r ,
r >1 s >1

where a lot of explanation is needed. Here

   
Er 0 0 0
Êr = , F̂s =
0 0 0 Fs

while since rk(N1 ) = rk(N2 ) = 1 they are respectively of the form

 √   √ 
√aJm,m abJm,n
, √cJm,m cdJm,n
abJn,m bJn,n cdJn,m dJn,n
with a, b, c , d to be determined, along with the eigenvalues µ1 and µ2 .
To determine µ1 and µ2 , we note that the partition of V (X + Y ) with two cells V (X ) and V (Y ) is equitable, with quotient
 
k n
.
m ℓ
Hence µ1 and µ2 are the zeros of the characteristic polynomial of this matrix
x2 − (k + ℓ)x + kℓ − mn,
that is they are equal to
1 
(k + ℓ ± (k − ℓ)2 + 4mn).
2
142 C. Godsil / Discrete Mathematics 312 (2012) 129–147

Now let u and v be distinct vertices in X . We determine conditions for perfect state transfer from u to v in X + Y . Since
(Fˆs )u,v = 0 and (Êr ) = (Er )u,v ,
−
(HX +Y (t ))u,v = exp(iµ1 t )(N1 )u,v + exp(iµ2 t )(N2 )u,v + exp(iθr t )(Er )u,v
r >1
−
= a exp(iµ1 t ) + c exp(iµ2 t ) + exp(iθr t )(Er )u,v .
r >1

On the other hand

1 −
HX (t )u,v = exp(ikt ) + exp(iθr t )(Er )u,v
m r >1

and thus if we have perfect state transfer from u to v in X at time τ , we will have perfect state transfer between the same
vertices in X + Y at time τ if
1
exp(ikτ ) = a exp(iµ1 τ ) + c exp(iµ2 τ ). (15.1)
m
As
− −
I = N1 + N2 + Êr + F̂s
r >1 s>1

we see that a + c = 1/m. Since exp(ikt ), exp(iµ1 τ ) and exp(iµ2 τ ) are roots of unity, we see that (15.1) can hold if and only
if

exp(ikτ ) = exp(iµ1 τ ) = exp(iµ2 τ ).

For this we need both (k − mu1 )τ and (k − µ2 )τ to be integer multiples of 2π , and hence that
k − µ1
∈ Q.
k − µ2

This can only happen if (k − ℓ)2 + 4mn is a perfect square.

In our treatment here we have followed Angeles-Canul et al. [4]. Using these ideas they prove the following results.

Lemma 15.1. Suppose Y is a k-regular graph on n vertices. Then there is perfect state transfer in K2 + Y between the vertices of
K2 if
√
(a) ∆ = k2 + 8n is an integer.
(b) n is even and 4|k.
(c) The largest power of two that divides k is not equal to the largest power that divides ∆. 

Lemma 15.2. Suppose Y is a k-regular graph on n vertices. Then there is perfect state transfer in K2 + Y between the vertices of
K2 if

(a) ∆ = (k − 1)2 + 8n is an integer.



(b) Both k − 1 and n are divisible by 8. 

The join of X and Y is the complement of the disjoint union of X and Y . Hence for regular graphs we can derive results
about joins from information about complements.

Lemma 15.3. Suppose X is regular graph on n vertices with perfect state transfer from u to v at τ . If τ is an integer multiple of
2π/n, then there is perfect state transfer from u to v at time τ in X .

Proof. Since X is regular, A and J − I commute and so

HX (t ) = exp(it (J − I ))HX (−t ).

Using the spectral decomposition of J − I we find that

 
1 1
exp(it (J − I )) = exp(i(n − 1)t ) J + exp(−it ) I − J
n n
and this is a multiple of I if exp(int ) = 1, that is, if t is an integer multiple of 2π /n. 

Using this lemma, it is immediate that we have perfect state transfer on nK2 (when n ≥ 2) and nC4 .
 C. Godsil / Discrete Mathematics 312 (2012) 129–147 143

16. The direct product

If X and Y are graphs then their direct product X × Y is the graph with adjacency matrix

A(X ) ⊗ A(Y ).

Lemma 16.1. Suppose X and Y are graphs with respective adjacency matrices A and B and suppose A has spectral decomposition
−
A= θr Er .
r

Then
−
HX ×Y (t ) = Er ⊗ HY (θr t ).
r

Proof. First,
−
A⊗B= θr Er ⊗ B
r

and since the matrices Er ⊗ B commute,

∏
HX ×Y (t ) = exp(iθr Er ⊗ B).
r

If E 2 = E then
− 1
exp(E ⊗ M ) = I + E ⊗ M k = (I − E ) ⊗ I + E ⊗ exp(M )
k≥1
k !

and accordingly
∏
HX ×Y (t ) = (I − Er ) ⊗ I + Er ⊗ HY (θr t ) .

r

(I − Er ) = 0, the lemma follows.

∏
Since Er Es = 0 if r ̸= s and r 

Lemma 16.2. Suppose that HY (τ )|u⟩ = γ |v⟩ where γ = exp(iϕ) and HY (2τ ) = γ 2 I. If the eigenvalues θ1 , . . . , θm of X are
odd integers, then

HX ×Y (τ ) = HX (ϕ) ⊗ γ −1 HY (τ ).

Proof. Assume that HY (t ) = exp(it θr )Er . If θr is an odd integer then

∑
r

(θr −1)/2
HY (θr τ ) = HY (2τ ) HY (τ ) = γ θr γ −1 HY (τ )

and accordingly
 
θr
−
HX ×Y (t ) = γ Er ⊗ γ −1 HY (τ ).
r

If γ = exp(iϕ) it follows that

 
−
HX ×Y (t ) = exp(iϕθr )Er ⊗ γ −1 HY (τ ) = HX (ϕ) ⊗ γ −1 HY (τ )
r

as required. 

A closely related result appears as Proposition 2 in [19]. We present two examples provided there.
If the eigenvalues of X are odd integers then HX (0) = I and HX (π ) = −I. The eigenvalues of the d-cube are the integers
d − 2r for r = 0, . . . , d, and it has perfect state transfer at π /2, with γ = id . Hence if X is a graph with odd integer
eigenvalues, then we have perfect state transfer on the product X × Qd when √ d is even. For a dsecond example, if Rd denotes
√ of P3 then Rd has perfect state transfer at time π / 2 with γ = (−1) . Therefore X × Rd has perfect
the dth Cartesian power
state transfer at π / 2.
Ge et al. [19] also give results for the lexicographic product.
144 C. Godsil / Discrete Mathematics 312 (2012) 129–147

17. Mixing

Questions about perfect state transfer might be viewed as asking at what times t does the transition matrix satisfy certain
restrictions on its entries. There are a number of interesting questions of this form.

17.1. Perfect mixing

For the first, we can ask if there is a time t such that all entries of H (t ) have the same absolute value. We say a unitary
matrix is flat if all its entries have the same absolute value and we say that perfect mixing occurs at time t is H (t ) is flat. We
have
 
1 1 i
HK2 (π /4) = √
2 i 1
which is flat. Since
HQd (t ) = HK2 (t )⊗d
we have perfect mixing at time π /4 on the d-cube. Also
   
1 1 1 1
HK4 (t ) = exp(3it ) J + exp(−it ) I − J = exp(−it ) exp(4it ) J + I − J
4 4 4 4
and thus HK4 (π /4) is flat. Consequently any Cartesian product of copies of K2 and K4 is perfect mixing at time π /4.
The graphs K2 and K4 are the first two members of a series of graphs: folded cubes. The folded (d + 1)-cube is the graph we
get from the d-cube by joining each vertex to the unique vertex at distance d from itself. It can also be viewed as the quotient
of the (d + 1)-cube over the equitable partition formed by the pairs of vertices at distance d + 1, which is the origin of the
term ‘folding’. The first interesting example is the folded 5-cube, often known as the Clebsch graph. In [10] Best et al. prove
(in our terms) that when d is odd, the folded d-cube has perfect mixing.
Ahmadi et al. [2] prove that K3 is perfect mixing and in [14] Carlson et al. show that C5 is not. Konno [34, Section 10.3]
shows that C6 is not perfect mixing. We can prove a little more. If m is odd,
= K2 × Cm .
C2m ∼
Then by Lemma 16.1
   
1 1 1 1 1 −1
HK2 ×X (t ) = ⊗ HX (t ) + ⊗ HX (−t ).
2 1 1 2 −1 1
If HK2 ×X (t ) is flat and
a := (HX (t ))u,v , b := (HX (−t ))u,v
then |a + b| = |a − b| and this can hold if and only if b = ia. Hence H (−t ) = iHX (t ) and
 
1 1+i 1−i
HK2 ×X (t ) = ⊗ HX (t ).
2 1−i 1+i
Hence K2 × X is uniform mixing if and only if HX (2t ) = −iI and X is uniform mixing at t. As K3 is perfect mixing when
t = 4π /9, it follows that C6 is not perfect mixing. Since C5 is not perfect mixing, neither is C10 .
If perfect mixing occurs at time τ , then H (τ ) is a flat unitary matrix. Such matrices form an important class of so-called
type-II matrices. For further information see [16].

17.2. Average uniform mixing

For all t, each row of the Schur product

H (t ) ◦ H (−t )
is a probability density; it is equal to
− − −
exp(it (θr − θs ))Er ◦ Es = Er ◦ Er + 2 cos(t (θr − θs ))Er ◦ Es
r ,s r r <s

and its average value over time is

−
Er ◦ Er .
r

Lemma 17.1. If the n × n matrices F1 , . . . , Fk are positive semidefinite and

∑
r Fr is a multiple of J, then Fr is a multiple of J for
each r.
 C. Godsil / Discrete Mathematics 312 (2012) 129–147 145

Proof. If F is positive semidefinite z = |u⟩ − |v⟩, then

0 ≤ z T Fz = Fu,u + Fv,v − 2Fu,v

and if equality holds

Fu,u = Fv,v = Fu,v .

Now
 
− −
T
(Fr )u,u + (Fr )v,v − 2(Fr )u,v .

z Fr z=
r r
∑
Since Fr is positive semidefinite each summand above is non-negative. If r Fr is a multiple of J then the left side is zero,
and so each summand on the right is zero. 

The continuous quantum walk on X is average uniform mixing if the average value of H (t ) ◦ H (−t ) is a multiple of J. Our
next result is new.

Lemma 17.2. If |V (X )| ≥ 3 then the continuous quantum walk on X is not average uniform mixing.

Proof. If X is not connected, it cannot be average uniform mixing, so we assume that X is connected. If the continuous walk
on X is average uniform mixing, then
−
Er ◦ Er
r

is a multiple of J. Since Er is positive semidefinite, Er ◦ Er is positive semidefinite for each r and therefore the previous lemma
implies that Er ◦ Er is a multiple of J.
Since θ1 is the spectral radius of A, then entries of E1 are non-negative and as E1 ◦ E1 is a multiple of J, it follows that E1 is
a multiple of J. Hence we may assume X is k-regular. If n = |V (X )| then E1 = 1n J. If r ̸= 1 then Er is flat and since E1 Er = 0
we see that n is even and exactly half the entries in any row or column of Er are negative. This implies that each eigenvalue
of X is an integer, with the same parity as the valency k. If |(Er )u,v | = e, then

e = (Er )u,u = (Er2 )u,u = ne2

and thus e = 1/n. Therefore tr(Er ) = 1, which tells us that each eigenvalue of X is simple. Since all diagonal entries of Er
must be positive, they are all equal and by [24, Theorem 4.1] it follows that X is walk regular. From [24, Theorem 4.8] we
know that the only connected walk-regular graph with simple integer eigenvalues is K2 . 

Adamczak et al. [1] prove the above theorem for Cayley graphs of abelian groups. Our proof follows theirs closely.
We do not seem to know very much about the average value of H (t ). The following indicates that there may be some
surprises.

Theorem 17.3. Suppose E1 , . . . , En are the idempotents for the path Pn and let T be the permutation matrix such that T |u⟩ =
|n + 1 − u⟩ for u = 1, . . . , n. Then
n
− 1
Er ◦ Er = (2J + I + T ). 
r =1
2n + 2

For a proof and further information, see [22].

18. Weighted adjacency matrices

Suppose |V (X )| = n. We say a symmetric matrix M is a weighted adjacency matrix for X if Mu,v = 0 for each pair of
distinct nonadjacent vertices u and v . So if M is a weighted adjacency matrix for X , it is a weighted adjacency matrix for any
graph Y such that X is a subgraph of Y . If the off-diagonal entries of M are 0 or ±1 and the diagonal entries are zero, we
will call M a signed adjacency matrix. Much of the theory of perfect state transfer extends to weighted adjacency matrices
with very little effort, since spectral decomposition still applies. If π is an equitable partition of X then, as we saw, perfect
state transfer on X implies perfect state transfer on the quotient X /π ; the adjacency matrix of this is weighted and thus we
see that information about state transfer on weighted graphs may have a bearing on the unweighted case. See [19, Section
5] for an illustration. Since the d-cube is distance regular, the distance partition relative to any vertex is equitable and the
corresponding quotient is a weighted path. As shown by Christandl et al. [18] it follows that for each d there is a weighting of
the edges of a path of length d that admits perfect state transfer between the end-vertices. If ∆ denotes the diagonal matrix
of valencies of X , then it seems reasonable to consider the Laplacian ∆ − A and perhaps the unsigned Laplacian ∆ + A. Bose
146 C. Godsil / Discrete Mathematics 312 (2012) 129–147

et al. consider perfect state transfer relative to the Laplacian in [12]. (Whether the adjacency matrix or the Laplacian is used
by a physicist depends on the type of spin interaction postulated.)
To decide which weightings are natural we need to know the physical situation being modelled, and currently this is
largely a matter of speculation. We note that the fundamental papers [18,37] focus on the unweighted case.

19. Some physics

The states of a quantum system are the 1-dimensional subspaces of a complex vector space—equivalently the points of
a complex projective space. There are two ways we may avoid admitting that projective geometry is involved.
The first way is to regard vectors non-zero x and y as equivalent if they span the same subspace; thus we represent a
projective point by an equivalence class of complex vectors. This is the traditional approach in introductions to quantum
physics. Here a reversible change of state is modelled by the application of a unitary operator: if our state is x then the
new state is the subspace spanned by Ux, where U is unitary. It is traditional to use unit vectors to represent states (which
reduces the size of our equivalence classes), and so then we might say that the new state y is equal to γ Ux. Here γ is a
complex number of norm 1; physicists call it a phase factor.
The second way is to represent the 1-dimensional space spanned by the non-zero vector x using the projection
1
xx∗ .
x∗ x
If y = ax (and a ̸= 0) then
1 1
yy∗ = ∗ xx∗ .
y∗ y x x
If |x| = 1 and U is unitary then |Ux| = 1, and the projection on y is
Uxx∗ U ∗ .
So our phase factors are gone. The map that sends xx∗ to Uxx∗ U ∗ is a linear map on the space of Hermitian matrices. A
quantum state corresponds to a Hermitian matrix with rank 1 and trace 1. Our unitary operator U is now
U ⊗ U ∗.
(Physicists refer to linear maps on spaces of operators as superoperators.)
Note that a positive semidefinite Hermitian matrix can be written as a sum of matrices of the form xx∗ and, if its trace
is 1, it can be written as a convex combination or Hermitian matrices with rank and trace 1. Physicists refer to the latter as
pure states and to a positive semidefinite matrix with trace 1 as a density matrix.
We turn to our continuous quantum walks, where our operators are the operators H (t ). In the density matrix approach
these become
H (t ) ⊗ H (t )∗ = H (t ) ⊗ H (−t ).
If H (t ) = exp(itA), then
H (t ) ⊗ H (−t ) = exp(it (A ⊗ I − I ⊗ A)).
If H (t )|u⟩ = γ |v⟩ then H (−t )|v⟩ = γ −1 |u⟩ and so
H (t ) ⊗ H (−t )|u⟩ ⊗ |v⟩ = |v⟩ ⊗ |u⟩.
We also have
H (t ) ⊗ H (−t )|u⟩ ⊗ |u⟩ = |v⟩ ⊗ |v⟩.
This shows that questions about perfect state transfer on graphs can be translated to questions about ‘‘phase-free’’ perfect
state transfer on signed graphs, because we can view A ⊗ I − I ⊗ A as the adjacency matrix of a signed version of the Cartesian
square X X .
There is a very interesting recent paper by Pemberton-Ross and Kay [35] using signed adjacency matrices to obtain perfect
state transfer between vertices at distance n in graphs with cn edges (for some constant c).

20. Questions

We list some questions which seem interesting. Unless explicitly stated otherwise, we consider only unweighted graphs.
(1) Is there a graph where we have perfect state transfer from u to v , but there is no automorphism of X which swaps u
and v ?
(2) Let Pn (a) be the path of length n with loops of weight a on each end-vertex. Is it true that for each n there is a weight a
such that we have perfect state transfer between the end-vertices? Casaccino et al. state in [15] that they have numerical
evidence that the answer is yes.
 C. Godsil / Discrete Mathematics 312 (2012) 129–147 147

(3) Is there some useful theory about the case where H (t )u,u = 0 for some time t and some vertex u? Or at least when
H (t ) ◦ I = 0?
(4) There are cubelike graphs with perfect state transfer at times π /2 and cubelike graphs with perfect state transfer at
times π /4. Are there cubelike graphs with perfect state transfer at time τ , where τ is arbitrarily small?
(5) Are there any trees, K2 and P3 aside, on which perfect state transfer occurs? (I do not see this as being useful, but it might
be fun.)

Acknowledgments

I would like to thank the following people, who either provided useful comments on early versions of this paper, or
helped me to understand the material presented in it: XiaoXia Fan, Alastair Kay, Dave Morris, Simone Severini, Murray
Smith, Christino Tamon, Dragan Stevanović.

References

[1] W. Adamczak, K. Andrew, L. Bergen, D. Ethier, P. Hernberg, J. Lin, C. Tamon, Non-uniform mixing of quantum walk on cycles, International Journal of
Quantum Information 5 (2007) 12. arXiv:0708.2096.
[2] A. Ahmadi, R. Belk, C. Tamon, C. Wendler, On mixing in continuous-time quantum walks on some circulant graphs, Quantum Computation and
Information 3 (2003) 611–618. arXiv:quant-ph/0209106.
[3] R.J. Angeles-Canul, R. Norton, M. Opperman, C. Paribello, M. Russell, C. Tamon, On quantum perfect state transfer in weighted join graphs, International
Journal of Quantum Information 7 (2009) 16. arXiv:0909.0431.
[4] R.J. Angeles-Canul, R. Norton, M. Opperman, C. Paribello, M. Russell, C. Tamon, Perfect state transfer, integral circulants and join of graphs, Quantum
Computation and Information 10 (2010) 325–342. arXiv:0907.2148.
[5] M. Bašić, Characterization of circulant graphs having perfect state transfer, 2011, p. 14. arXiv:1104.1825.
[6] M. Bašić, M.D. Petković, Some classes of integral circulant graphs either allowing or not allowing perfect state transfer, Applied Mathematics Letters
22 (2009) 1609–1615.
[7] M. Bašić, M.D. Petković, Perfect state transfer in integral circulant graphs of non-square-free order, Linear Algebra and its Applications 433 (2010)
149–163.
[8] M. Bašić, M.D. Petković, D. Stevanović, Perfect state transfer in integral circulant graphs, Applied Mathematics Letters 22 (2009) 1117–1121.
[9] A. Bernasconi, C. Godsil, S. Severini, Quantum networks on cubelike graphs, Physical Review A 78 (2008) 5. arXiv:0808.0510.
[10] A. Best, M. Kliegl, S. Mead-Gluchacki, C. Tamon, Mixing of quantum walks on generalized hypercubes, International Journal of Quantum Information
6 (2008) 1135–1148. arXiv:0808.2382.
[11] S. Bose, Quantum communication through an unmodulated spin chain, Physical Review Letters 91 (2003) arXiv:quant-ph/0212041.
[12] S. Bose, A. Casaccino, S. Mancini, S. Severini, Communication in XYZ all-to-all quantum networks with a missing link, International Journal of Quantum
Information 7 (2009) 713–723. arXiv:0808.0748.
[13] W.G. Bridges, R.A. Mena, Rational G-matrices with rational eigenvalues, Journal of Combinatorial Theory, Series A 280 (1982) 264–280.
[14] W. Carlson, A. Ford, E. Harris, J. Rosen, C. Tamon, K. Wrobel, Universal mixing of quantum walk on graphs, Quantum Computation and Information 7
(2007) 738–751. arXiv:quant-ph/0608044.
[15] A. Casaccino, S. Lloyd, S. Mancini, S. Severini, Quantum state transfer through a qubit network with energy shifts and fluctuations, International Journal
of Quantum Information 7 (2009) 1417–1427. arXiv:0904.4510.
[16] A. Chan, C. Godsil, Type-II matrices and combinatorial structures, Combinatorica 30 (2010) 1–24. arXiv:0707.1836.
[17] W.-C. Cheung, C. Godsil, Perfect state transfer in cubelike graphs, Linear Algebra and its Applications 435 (2011) 2468–2474. arXiv:1010.4721.
[18] M. Christandl, N. Datta, T. Dorlas, A. Ekert, A. Kay, A. Landahl, Perfect transfer of arbitrary states in quantum spin networks, Physical Review A 71
(2005) 12. arXiv:quant-ph/0411020.
[19] Y. Ge, B. Greenberg, O. Perez, C. Tamon, Perfect state transfer, graph products and equitable partitions, arXiv:1009.1340.
[20] C. Godsil, Controllable subsets in graphs, 2010, p. 14. arXiv:1010.3231.
[21] C. Godsil, When can perfect state transfer occur?, 2010, p. 15. arXiv:1011.0231.
[22] C. Godsil, Average mixing of continuous quantum walks, 2011, p. 20. arXiv:1103.2578.
[23] C. Godsil, Periodic graphs, The Electronic Journal of Combinatorics 18 (2011) #23. arXiv:0806.2074.
[24] C.D. Godsil, B.D. McKay, Feasibility conditions for the existence of walk-regular graphs, Linear Algebra and its Applications 30 (1980) 51–61.
[25] C. Godsil, G. Royle, Algebraic Graph Theory, Springer, New York, 2001.
[26] C. Godsil, S. Severini, Control by quantum dynamics on graphs, Physical Review A 81 (2010) 5. arXiv:0910.5397.
[27] C. Godsil, J. Shawe-Taylor, Distance-regularised graphs are distance-regular or distance-biregular, Journal of Combinatorial Theory, Series B 43 (1987)
14–24.
[28] W. Imrich, S. Klavzar, Product Graphs: Structure and Recognition, Wiley, 2000.
[29] A. Kay, Perfect, efficient, state transfer and its application as a constructive tool, International Journal of Quantum Information 08 (2010) 641.
arXiv:0903.4274.
[30] A. Kay, The basics of perfect communication through quantum networks, 2011, p. 8. arXiv:1102.2338.
[31] J. Kempe, Quantum random walks—an introductory overview, Contemporary Physics 44 (2003) 20. arXiv:quant-ph/0303081.
[32] V. Kendon, Quantum walks on general graphs, International Journal of Quantum Information 4 (5) (2006) 791–805. arXiv:quant-ph/0306140.
[33] V.M. Kendon, C. Tamon, Perfect state transfer in quantum walks on graphs, Journal of Computational and Theoretical Nanoscience 8 (2011) 422–433.
[34] N. Konno, Quantum walks, in: Quantum Potential Theory, in: Lecture Notes in Math., vol. 1954, Springer, Berlin, 2008, pp. 309–452.
[35] P.J. Pemberton-Ross, A. Kay, Perfect quantum routing in regular spin networks, Physical Review Letters 106 (2011) 020503. p. 4. arXiv:1007.2786.
[36] M. Petković, M. Bašić, Further results on the perfect state transfer in integral circulant graphs, Computers and Mathematics with Applications 61
(2011) 300–312.
[37] N. Saxena, S. Severini, I. Shparlinski, Parameters of integral circulant graphs and periodic quantum dynamics, International Journal of Quantum
Information 5 (3) (2007) 417–430. arXiv:quant-ph/0703236.
[38] D. Stevanović, Applications of graph spectra in quantum physics, in: D. Cvetković, I. Gutman (Eds.), Selected Topics on Applications of Graph Spectra,
Mathematical Institute SANU, Belgrade, 2011, pp. 85–111.