Aequationes Mathematicae 25 (t982) 257-268 0001-9054/82/002257-12 $1.50 + 0.

20/0
University of Waterloo © 1982Birkh~iuser Verlag, Basel

Constructing cospectrai graphs

C. D. GODSIL AND B. D. MCKAv

Abstract. Some new constructions for families of cospectral graphs are derived, and some old ones are
considerably generalized. One of our new constructions is sufficiently powerful to produce an estimated
72% of the 51039 graphs on 9 vertices which do not have unique spectrum. In fact, the number of graphs
of order n without unique spectrum is believed to be at least an~g,_l for some a >0, where g, is the
number of graphs of order n and n >- 7.

I. Introduction

1.1. W e use G to d e n o t e a simple g r a p h with vertex set V ( G ) = {v~, v~. . . . . v,}

and edge set E ( G ) . T h e adjacency matrix of G is the n x n matrix with (i,j)th
entry equal to 1 if vertices i and j are adjacent and equal to 0 otherwise. T h e
adjacency matrix o f G will also be d e n o t e d by the symbol G. The characteristic
polynomial of G is the polynomial ~ ( G ) = ~b(G, x ) = det(xI, - G), w h e r e / , is the
n × n identity matrix.
Two graphs G and H are cospectral if 4~(G) = ~b(H). W e say that G is
characterized by its spectrum if every g r a p h cospectral to G is isomorphic to G. It
was proved by S c h w e n k [14] that the p r o p o r t i o n of trees on n vertices which are
characterized by their spectra converges to zero as n increases. T h e c o r r e s p o n d i n g
asymptotic question for graphs in general remains one of the outstanding unsolved
problems in the t h e o r y of g r a p h spectra.
Schwenk's p r o o f d e p e n d s o n a construction which provides pairs of cospectral
trees. Thus, if we wish to settle the question for graphs in general, it is natural to

AMS (1980) subject classification: Primary 05C55. Secondary 15A18.

Manuscript receioed March 1, 1982, and in ]inal form, January 4, 1983.

257
258 C, D. GODSIL AND B, D. MCKAY AEO. MATH.

look for constructions for pairs of cospectral graphs. In this paper we present some
new and powerful constructions for pairs of cospectral graphs and considerably
generalize some old ones. One of our new methods is sufficiently powerful to
generate an estimated 72% of the 51039 graphs on 9 vertices which are not
characterized by their spectra.

1.2. T E R M I N O L O G Y . We will use Jm, to denote the m x n matrix with each

entry one and I, to denote the identity matrix of order n. In each case the subscripts
will be deleted if the order is clear from the context. The column vector Jm i will also
be denoted by jm.
The concept of switching was introduced by Seidel [13]. Let S be a subset of
V(G). T h e n the graph H formed from G by switching about S has

V(R)-- V(G),
and
E ( n ) = { x y E E ( G ) [ x , y E S or x, y f ~ S } U ( x y ~ E ( G ) [ x E S and y E S } .

We say that G and H are switching equivalent.

2.1. C O N S T R U C T I O N . Let G be a graph and let ,r = ((71, C: . . . . . Ck, D ) be a

partition of V(G). Suppose that, whenever 1 -< i, j -< k and o E D, we have
(a) any two vertices in C~ have the same number of neighbours in Ci, and
(b) v has either 0, n, [2 or n, neighbours in C~, where n, = [ C~ I-
The graph G °° formed by local switching in G with respect to ,r is obtained from G
as follows. For each o E D and 1 <- i -< k such that v has n~/2 neighbours in C,
delete these n~/2 edges and join v instead to the other n~[2 vertices in C .

For our purposes the most important property of our construction is provided
by the next theorem.

2.2. T H E O R E M . Let G be a graph and let lr be a partition of V(G) which

satisfies properties (a) and (b) above. Then G (~) and G are cospectral, with cospectral
complements.

Proo[. The most direct way of showing that two graphs are cospectral is to show
that their adjacency matrices are similar. We now proceed to do this.
For any positive integer m, define O m = 2J~,/m - Ira. T h e following claims can
be verified easily.
Vol. 25, 1982 Constructingcospectralgraphs 259

(a) OZ = / ~ .
(b) If X is an m x n matrix with constant row sums and constant column sums,
then QmXQ~=X.
(c) If x is a vector with 2m entries, m of which are zero and m of which are
one, then Q2wr = af~ - x.
If the vertices of G are labelled in an order consistent with ~, the adjacency
matrix of G has the form

• Dj
I CCI T12 Clz
C2 "'" '' Clk ~Dkl
C2k D2
G =
cT, cX ..- Ck
DT DT
2 "" • DI

The required properties of 7r ensure that each C, and each C,j has constant row
sums and constant column sums, and that each column of each 19,
has either 0, n,/2
or n, ones. Therefore QGO
is the adjacency matrix G ~ , where O is the
block-diagonal matrix diag(On,, O,~. . . . . O,k, Itol).
Since O 2 = L this proves G and
G ~*) to be cospectral. G and G °° have cospectral complements by the same
argument, since (t~) °° is the complement of G ~ . []

2.3. E X A M P L E S . We consider some cases of local switching that are of

particular interest.
(a) Form G by taking a regular graph H with an even number of vertices and
adjoining a new vertex v adjacent to exactly half the vertices of H. Then G ~) for
~'=(V(H),{v}) is formed by joining v instead to the other vertices of H. An
example with H = C8 is shown in Figure 1.

Figure 1

If H has 2m vertices and a trivial automorphism group, it is easy to show that all
of the (2~) possibilities for G are nonisomorphic. Hence, for example, if we choose
H to be a cubic graph on 12 vertices with no nontrivial automorphisms we can
COnstruct (~)/2 = 462 pairs of cospectral nonisomorphic graphs on 13 vertices.
(b) Let G be regular with degree k and let S be a subset of V(G) such that the
graph H obtained from G by switching about S is also regular with degree k. Then
260 C.D. GODSIL AND B. D. MCKAY AEQ. MATH.

H and G are cospectral because H = G t'°, where lr = ( V ( G ) \ S , S). This construc-

tion is well known.
In examples (a) and (b), G (') and G are switching equivalent. However this is
not always the case, as in the next example.
(c) This example is most easily explained by reference to Figure 2, where
¢r = ( { a , b , c , d } , H ) .

a b c d

. ) ( . )
Figure 2

A similar example is shown in Figure 3, where again ~r = ({a, b, c, d}, H). If H is

chosen to have vertices of degree three, except that vertices 1 through 6 have
degree one, then G and G t'° are both cubic. This construction provides 2 of the 3
pairs of cospectral cubic graphs on 14 vertices. There are no cospectral cubic graphs
on less than 14 vertices [2].

Figure 3

2.4. STATISTICS. We now demonstrate the efficiency of local switching as a

means of producing pairs of cospectral graphs. We will concentrate on the subcase
which appears to produce the most examples, namely that when 7r = (C~, D) and
Ic t =4.
Define I, to be the proportion of (unlabelled) graphs G on n vertices such that
there exists zr = (C1, D ) with I C~ I = 4 satisfying conditions (a) and (b) of Section 2.1
and G (') ~ G. Also define c, to be the proportion of all graphs on n vertices which
are not characterized by their spectra.
It is easy to show that l~ = 0 for n -<6. For .each value of n in the range
7 <-n-< 16 a large number 5/, of random labelled graphs with n vertices was
generated. For each graph the order of the automorphism group was found, and
then, for each ~ of the required type, G ('~) was tested for isomorphism with G. An
estimate of I, was then obtained by weighting each graph according to the order of
Vol. 25, 1982 Constructing cospectral graphs 261

its a u t o m o r p h i s m g r o u p , to get an unbiased estimator for unlabelled graphs. T h e

results are shown in Table 1. T h e values of cn are taken f r o m [10] for n -< 9 and are
unknown for n > 9. T h e estimates of I, are given with approximate 95% confidence
limits. T h e latter were c o m p u t e d u n d e r the untested assumption that the distribu-
tion of the estimates of I, o b t a i n e d f r o m samples of size N , / I 0 is normal, and
should only be taken as a r o u g h guide to the accuracy of l,. T h e r e are convincing
reasons to suspect that I, ~ (~)21-~, which would imply the claim m a d e in the last
sentence of the abstract.
Table 1

n N~ c. I. t,/c~

5 -- 0.059 0 0.00
6 -- 0.064 0 0.00
7 50000 0.105 0.037-+0.003 0,35
8 50000 0.139 0.084-+0.004 0.60
9 50000 0.186 0.135-+0.003 0.73
10 30000 0.164-+0.008
11 20000 0.165+0.008
12 10000 0.145-+0.011
13 100120 0.127_+0.010
14 6000 0.095 -+0.010
15 5000 0.074 -+0.008
16 3000 0.042 -+0.009

It is seen that for n <-9 the ratio l~/c, is steadily increasing and p r o b a b l y
exceeds 70% for n = 9. It is interesting to note that lim i n f ~ In/cn > 0 would imply
c~ ~ 0, since 1, --* 0. H o w e v e r , we feel that our data provide only a small a m o u n t of
evidence that c, ~ 0, and that the behaviour of cn for small n may not be typical.

2.5. In the special case where zr = (C1, D ) and there are no edges within Cl, the
construction a b o v e can be considerably generalized. W e will call two rn x n
matrices A and B congruent if A T A = B T B . If we view the columns of A and B as
points in ~ " then it is clear that A and B are congruent if and only if the
corresponding sets of points in ~R" are c o n g r u e n t in the geometric sense, i.e., there
is an m × m o r t h o g o n a l matrix which m a p s the columns of A o n t o the correspond-
ing columns of B.

2.6. C O N S T R U C T I O N . Let H be a graph on n vertices and let A be an

m × n 0-1 matrix. T h e n H ( A ) is the g r a p h with adjacency matrix

(°T HA)
262 C.D. GODSIL AND B. D, MCKAY AEQ. MATH

2.7. T H E O R E M . Let H be a graph on n vertices and let A and B be two

congruent m x n 0-1 matrices. Then the graphs H ( A ) and H ( B ) are cospectral.

Proof. Let Q, = diag(O,/~), where Q A = B and Q is orthogonal. Then Q~ is

orthogonal and Q~H(A )QT = H ( B ). []

One source of congruent matrices is Construction 2.1 with k = 1. Another

example is provided when A and B are the transposed incidence matrices of two
BIBDs with the same parameters. If the graph H has no edges in this case then
Construction 2.7 reduces to a known result. If H is not empty, then H ( A ) may be
not isomorphic to H ( B ) even if the two corresponding designs are isomorphic. As
an example, let A be the transposed incidence matrix of a Steiner Triple System
with v = 15 and trivial automorphism group, and let H be any 15 vertex graph with
trivial automorphism group. Then H ( A ) ~ H ( B ) for any column permutation B of
A. Thus we get a family of 15! = 130767438000 nonisomorphic graphs on 50
vertices, all of which are cospectral and have cospectral complements. More
generally, there are v t~)/3+°t°) labelled Steiner triple systems with v points (see [1]),
and 2(P/v !(1 + O(1/v)) graphs of order v with trivial automorphism groups. Thus
we can construct 2"~)/v !(1 + O(1/v)) families of graphs of order v(v + 5)/6, each
containing v tp/3~°tv) cospectral graphs. Although high, the number of graphs
involved here is miniscule compared to the number produced by Construction 2.1.
Construction 2.6 has also been investigated (without proof) by Davidson [6],
who gives many examples of congruent matrices.

3. Tensor products

3.1. In this section we describe a very general procedure which uses the matrix
tensor product to construct families of cospectral graphs. Although many special
cases of this construction have appeared before, the general case is new.
All the matrices in this section are real, but otherwise not restricted. The
necessary conditions for the matrices so constructed to be adjacency matrices of
graphs will be obvious in every case.
The tensor (direct) product of matrices A and B will be denoted by A ~) B. For
the most elementary properties of this operation we refer the reader to [11] or [7].

3.2. Consider a sequence of matrices

.4 = (A ~l),A (21), ~ o). a ~2) A (2) A (k2)),

Vol. 25, 1982 Constructingcospectra!graphs 263

where A II) has order n~ x n~ and A I2) has order n2 x n2, for 1-< i -< k. The next
Iemma is just an elementary property of the tensor product.

3.3. LEMMA. For any monomial f in k noncommuting variables

tr f ( A ~1)@ A~2), A ~21)@A ~2),..., A~l) @ A~2))

= tr f ( A ~'), A(2~). . . . . A~))trf(A~ 2), A~2),..., A~2)). []

Let B be a sequence of matrices with the same orders as those of A. Define

T(A ) = ~ (A ~')@ A 12)) and T(B) = Y~=z(B ~')@ B12)). The general construction
we are considering is based on the following theorem, which follows immediately
from the multinomiai theorem and Lemma 3.3.

3.4. T H E O R E M . Suppose that

tr f ( A •), A ~),..., A ~)) = tr f(B? ), B~),..., B ~)),

for any monomial f and j E {1,2}. Then T(A ) and T( B ) are cospectral. []

The cases where A ? ) E { L G i , Gj} for 1-<]---2 and l<-i<-k have been
investigated in depth by Cvetkovi6 and others ([3], [5]). In this connection, we note
that, although we are using tensor products of only two factors, the class of graphs
constructed is not thereby reduced, since the case of more than two factors can be
obtained by repeated application.
A simple family of applications of Theorem 3.4 can be obtained with the help of
the following lemma.

3.5. L E M M A . If G~ and G2 are cospectal, then tr G~ = tr Gi for any r >- O. If

also GI and G2 are cospectral, then trf(G~, t ~ , J ) = trf(G2, t~2,J) for any mono-
rnial f.

Proof. The first claim is obvious, while the second follows from Lemma 2.1 of
McKay [12]. []

3.6. CONSTRUCTION. Let GI and G2 be cospectral, and let X and H be

square matrices of the same order. Then H @ I + X @ G 1 is cospectral to
H@I+X@G2.
The choices X = I, X = J and X = I + H give the cartesian product H x Gi, the
lexicographic product G~ [H] and the strong product H * G~, respectively.
264 C+ D. GODSIL AND B. D+ MCKAY AEQ+ MATH,

3.7. C O N S T R U C T I O N . Let G~ and G2 be cospectral, with (~ and ( ~ also

cospectral. Let C, D, E and F be square matrices of the same order. Then
C~I+D@J+E@G~+F@t~ is cospectral to C @ I + D @ J + E @ G 2 +
F @ G2.

If C = F = 0 and E = I we have the lexicographic product D[G,]. Another

interesting subcase comes from taking C = D = 0, E = (~ ~) and F = (0 ~), which
yields the switching graphs S w ( G 0 and Sw(G2). These are regular even if Gt and
G2 are not, and are known [8] to be isomorphic if and only if G~ and G2 are
switching equivalent. Thus, for example, if G; and G2 are cospectral nonisomorphic
trees then S w ( G 0 and Sw(G2) are cospectral nonisomorphic regular graphs. This
follows from the easily proved fact that switching equivalent trees are isomorphic.
Constructions 3.6 and 3.7 have the property that the graphs constructed will be
isomorphic if G~ and G2 are isomorphic. This is not a necessary characteristic of
T(A) and T(B), as our next example demonstrates.
Let G~ and G2 be bipartite graphs with parts of cardinality m and n. For i = 1, 2,
label G+ so that
0 Bi)
G+ = BT 0 '

where the partition is m + n. Define

E= (/" 0), E*=(0 0 I0, ) and C~= (0 B,).

3.8. L E M M A . If G~ and G2 are cospectral then, for any monomial f in four

noncommuting variables, trf(E, E*, C;, C T) = try(E, E *, G , CT2). If also m = n
then trf(E,E*, C~, C~) = trf(E*,E, C T, (72).

Proof. The only monomials f for which trf(E,E*, C, C T) is possibly nonzero

are those for which f(E, E*, C, C T) equals (CCT) r or (CTC) ~ for some r ~ 0. The
first claim now follows from Lemma 3.5, together with the observation that
tr(CiCT)' = tr(CTC+)r _- ~+~
2,- ~2,
,-, + if r -> 1. The second claim can be proved by noting,
in addition, that tr(CCT) ° = m and tr(C~C)° = n.
Lemma 3.8 can be used to produce many diverse pairs of cospectral graphs. For
example, if G~ and G2 are cospectral as above, and H~ and /-/2 are arbitrary
cospectral graphs, then the four "half-cartesian-products" <3+ @ I + E* @/4/
(i, j = 1, 2) are cospectral. However, the most interesting application is perhaps the
partitioned tensor product, first defined in [9].
Vol. 25, 1 9 8 2 Constructingcospectral graphs 265

3.9. T H E O R E M [9]. Let G, and G2 be cospectral bipartite graphs with parts of

cardinality m and n. Define E, E*, C, and C2 as above. Let G and H be arbitrary
graphs and let Q be an r x s matrix, where r and s are the orders of G and H,
respectively. Define

P = EGG +E*QH+C~@O+G@Q r,

and
P*= E*@G + E @H +C~@O + C T ( ~ Q ~.
Then
4~(P)@(H) ...... = ~ ( P * ) 4 , ( G ) .... •

Consequently, da(P) = da(P*) if either m = n or oh(H) = cb(G).

Proof. W i t h o u t toss of generality, assume that m >-n. A d d m - n isolated

vertices to the second parts of Gt and G2. T h e effect is to add m - n isolated copies
of H to P and m - n isolated copies of G to P*. T h e claim is now immediate from
3.4 and 3.8. []

Informally, P is obtained as follows: Replace each vertex in the first part of G,

by a copy of G, and each vertex in the second part by a copy of H. Then, for each
edge of G,, join the c o r r e s p o n d i n g copies of G and H according to the entries of Q.
An example is shown in Figure 4.

GI " G2 " o---o----o G - H -

0-
0 0
0 0

P p-

Figure 4
266 C.D. GODSIL AND B. D. MCKAY AEQ. MATH.

In [9] it is shown that the smallest pair of cospectral graphs, the smallest pair of
cospectral forests and the smallest pair of cospectral connected graphs can each be
obtained using the partitioned tensor product.
Of course, there is no guarantee that P and P* are nonisomorphic in general,
although some sutficient conditions are known (see [9]). A special case of Theorem
3.9 has been rediscovered by Schwenk, Herndon and Ellzey [16].

4. Cespectrai points

4.1. If G is a rooted graph with root v, then Go denotes the graph formed by
deleting v from G. Given two rooted graphs G and H with roots v and w,
respectively, we can define the following composite graphs:
(i) G " H is formed from G and H by identifying v and w.
(ii) G ~ H is formed from disjoint copies of G and H by adding one edge
joining v and w.
(iii) G --- H is formed from disjoint copies of Go and Hw by joining every vertex
in G, which is adjacent to v in G to every vertex in Hw which is adjacent to w in H,
Examples are given in Figure 5.
The spectrum of G . H was first determined by Schwenk [14]. That of G ~ H is
just a special case. The spectrum of G -- H has not been previously determined.

OoH-

O-H-

OBH°~ ~ 7

Figure 5
Vol. 25, 1 9 8 2 Constructingcospectral
graphs 267

4.2. THEOREM.
(i) 4~(G" H ) = ~b(G)ga(H+)+ c k ( G ~ ) 6 ( H ) - x 4 ~ ( G o ) ~ ( H + ) ;
(ii) 4,(G ~ H ) = ¢ b ( G ) ( a ( H ) - 6(Go)~b(H+);
(iii) 4,(G =- H ) = ¢b( Go )~b(H+ ) - (xcb(G~ ) - ¢b(G))(x~b(H+ ) - ok(H)).

Proof. For part (i), and thus (ii), see [14] or [4]. To prove part (iii) we need some
additional notation. Let M and N be square matrices of order m and n,
respectively. Let a and b be (column) vectors of length m and n, respectively. The
notation M I,, represents the matrix (o ,~). The two claims following can be proved
by applying elementary row and column operations. For any a,

b ( M + aaa x) = a~b(M [a) + (1 - ax )~b(M). (1)

4~ M 0 =d~(M)4~(glb)+cb(N)cb(Mla)-xck(M)cb(N). (2)
0 N

Now suppose G = Go [g and H = H~ [h. Then

( H w -O h h T ) + kkT]

where kT= (gThT). Application of (1) and (2) produces

qS(G -~ H ) = (~(G~)dp(Hw - hh I ) + qb(G++)~b(G, - ggT)

- - ~b(H+ - hh T)~b(G~ - g g T ) .

Further application of (2) gives the desired form. []

4.3. C O R O L L A R Y . ~b(G H H ) + ~b(G - H ) = x~b(G. H). []

4.4. C O R O L L A R Y . For i = 1,2, let G °> and H ¢'~be rooted graphs with roots v <'~
and w ~'~, respectively. Suppose that ~b(G~°), ~b(G~?,,), 4~(H ~°) and ~b(H~,>) are
independent of i. Then 4~(G °~. H°~), ~,(G <')H H ~'~) and ¢b(G ~'~=--H ~')) are indepen-
dent of i. []

The case ~b(G °~. H °~) was used by Schwenk [14] to prove that almost no tree is
uniquely identified by its spectrum. Stronger results of similar form appeared in [10]
and especially in [12]. A generalization (for the " . " operation) to graphs rooted at
more than one point has been given by Schwenk [15]. Construction 2.1 can also be
described in this manner.
268 C. D. GODSILAND B. D. MCKAY AEQ. MATH.

REFERENCES

[1 ] BABA1,L., Almost all Steiner triple systems are asymmetric. Ann. Discrete Math. 7 (1980), 37-39.
[2] BUSSEMAKER,F. C., COBELJIC, S., CVETKOVI~,D. M. and SEIDEL, J. J., Computer investigation o/
cubic graphs. Technishe Hogeschool Eindhoven, Report 76-WSK-01 (1976).
[3] CVETKOVI~,D. M., The generating function for variations with restrictions and paths of the graph and
self-complementary graphs. Univ. Beograd Publ. Elektrotehn. Fak., Ser. Mat. Fiz 320-328 (1970),
27-34.
[4] CVETKOV|C,D. M., DOOB, M. and SACHS,H., Spectra of graphs. Academic Press, New York, 1980.
[5] CVETKOVt(',D. M. and LuC~(:, R. P., A new generalization of the concept of the p-sum of graphs.
Univ. Beograd Publ. Elektotehn. Fak., Ser. Mat. Fiz. 302-319 (1970), 67-71.
[6] DAVIDSON,R. A., Genesis and synthesis of cospectral and paracospectral graphs revelation of latest
symmetries. Preprint, 1980.
[7] GANTMAKHER,F. R., Theory of matrices. Vols. 1, 2. Chelsea, New York, 1960.
[8] GODSIL,C. D., Neighbourhoods of transitive graphs and GRRs. J. Comb. Theory, Ser. B 29 (1980),
116-140.
[9] GODSlL,C. D. and McKAv, B. D., Products of graphs and their spectra. Combinatorial Mathematics
IV, Lecture Notes in Math. 560, Springer-Verlag (1976), 61-72.
[10] GODSIL, C. D. and MCKAY, B. D., Some computational results on the spectra of graphs.
Combinatorial Mathematics IV, Lecture Notes in Math. 560, Springer-Vedag (1976), 73-92.
Ill] LANCASTER,P., Theory of matrices. Academic Press, New York, 1969.
[12] McKAY, B. D., On the spectral characterisation of trees. Ars Combin. 3 (1977), 219-232.
[13] SEIDEL, J. J., Graphs and two-graphs. In Proc. Fifth Southeastern Conference on Combinatorics,
Graph Theory and Computing (Boca Raton, FL., t974), Congressus Numerantium X, Utilitas
Math., Winnipeg, Man., I974, pp. 125-143.
[14] SCHWENK,A. J., Almost all trees are cospectral. In New Directions in the Theory of Graphs (Proc.
Third Ann Arbor Conf., Univ. Michigan, Ann Arbor, Michigan, 1971), Academic Press, New
York, 1973, pp. 275-307.
[15] SCHWENK, A. J., Removal-cospectral sets of vertices in a graph. In Proc. Tenth Southeastern
Conference on Combinatorics, Graph Theory and Computing (Boca Raton, FL., 1979), Congressus
Numerantium XXIII-XXIV, Utilitas Math., Winnipeg, Man., 1979, pp. 849-860.
[16] SCHWENK, A. J., HERNDON, W. C. and ELLZEY, M. L., The construction of cospectral composite
graphs. In Proc. Second Intern. Conf. on Combinatorial Math. (New York, 1978), Ann. N. Y. Acad.
Science 319 (1979), 490-496.

Simon Fraser University, Vanderbilt University,

Bumaby, BC VSA IS6, Nashville, TN 37235,
Canada. U.S.A.