On 1-eigenvectors of graphs

Dragan Stevanovic

University of Primorska, Slovenia and

Mathematical Institute, Serbian Academy of Science and Arts

June 22, 2015

Dragan Stevanovic On 1-eigenvectors of graphs

Motivation

Jerry Ray Dias was discussing in 1995structuralfactors associated

to the presence of eigenvalues 1, 2, . . . , 7 and 3 in
graphs with 3.

He could not find any graph with eigenvalues 8 and 3,
and made an unfortunate conjecture that no such graph exists :)

Together with Patrick Fowler and Marko Milosevic, we found a

number of constructions of families of such graphs back in 2010.

Dragan Stevanovic On 1-eigenvectors of graphs

A particular construction

The construction relies on the property of C4k that

its vertices can be divided in two parts such that
each vertex has a half of its neighbors in each of the parts.

In other words: C4k has a 1-eigenvector of eigenvalue 0.

Dragan Stevanovic On 1-eigenvectors of graphs

Wilfs overlooked question

While meticulously reading through references for a research

monograph on

it turned out that Herbert Wilf asked back in 1986: What kind of
a graph admits an eigenvector consisting solely of 1 entries?

Dragan Stevanovic On 1-eigenvectors of graphs

Wilfs question as implicitly stated

This question arose in the discussion of the independence number

bound
n
n +1 2 , M2


d + 1 + n max M+

for a d-regular graph with the smallest eigenvalue n and the

corresponding eigenvector xn , with
1 1
M+ = max , M = min .
(xn )i >0 (xn )i (xn )i <0 |(xn )i |

Hence it appears that Wilf assumed his question to be:

What kind of a regular graph admits an eigenvector of the smallest
eigenvalue consisting solely of 1 entries?

Dragan Stevanovic On 1-eigenvectors of graphs

Wilfs question as explicitly stated

Problem PMEIG. Given a simple graph G with adjacency

matrix A, find a 1 eigenvector of A, if it exists.

Slightly easier and more specific variant is

Problem PMEIG0. Given a simple graph G with adjacency

matrix A, find a 1 eigenvector of A corresponding to the
eigenvalue 0, if it exists.

Dragan Stevanovic On 1-eigenvectors of graphs

Eigenpartite graphs

We call a graph eigenpartite

if its vertex set V can be partitioned as V = V + V so that
for each vertex u the number of its neighbors in V + is
equal to the number of its neighbors in V .

An eigenpartite graph trivially has

a 1 eigenvector x of the eigenvalue 0:

u V +,

1,
xu =
1, u V .

Hence PMEIG0 is solvable exactly for eigenpartite graphs.

Dragan Stevanovic On 1-eigenvectors of graphs

PMEIG PMEIG0

Theorem
PMEIG is polynomially reducible to PMEIG0.

Proof.
Ax = x

Entries of A are integers, so is an integer with {, . . . , }.

Hence the Cartesian product G + K||,|| has eigenvalue 0,

with a 1 eigenvector x j or x j ,
where j is the all-one eigenvector of K||,|| corresponding to ||,
and j is the 1 eigenvector of K||,|| corresponding to -||.

Thus, to solve PMEIG it is enough to solve PMEIG0 for the graphs

G , G + K1,1 , G + K2,2 , . . . , G + K, .

Dragan Stevanovic On 1-eigenvectors of graphs

PMEIG0 is NP-complete

Theorem
PMEIG0 is NP-complete.

Proof communicated by Brendan McKay and Laszlo Lovasz.

A k-uniform hypergraph is a set E of k-element subsets of V .

A hypergraph is 2-colorable if V can be partitioned in two color

classes such that no edge in E is monochromatic.

Lovasz (1973) showed that it is NP-complete to decide

whether a 4-uniform hypergraph is 2-colorable
so that each edge has two vertices of each color.

Dragan Stevanovic On 1-eigenvectors of graphs

Let H = (V , E ) be a 4-uniform hypergraph.
Construct G on the vertex set
V E 0 E 00 ,
where E 0 and E 00 are two copies of E :
E 0 = {e 0 |e E }, E 00 = {e 00 |e E }.
For each edge e = {x, y , z, w } of H, add to G the edges
e 0 x, e 0 y , e 0 z, e 0 w , e 00 x, e 00 y , e 00 z, e 00 w .
Dragan Stevanovic On 1-eigenvectors of graphs
If H is 2-colorable with each edge having two vertices of each
color, then G is eigenpartite: partition V = V 0 V 00 according to
the 2-coloring, put E 0 together with V 0 and E 00 together with V 00 .

If G is eigenpartite, then H is 2-colorable: just color the vertices

according to the eigenpartition.

Dragan Stevanovic On 1-eigenvectors of graphs

Solving PMEIG0

Problem PMEIG0 can be solved by an exhaustive search algorithm.

Let E0 be the eigenspace of the adjacency matrix A corresponding

to the eigenvalue 0.

E0 is the kernel of A, so it consists of vectors representing

coefficients of linear combinations of columns of A with value 0.

Compute the column echelon form

   
A B
.
I C

The basis D of E0 is given by the columns of C

corresponding to the zero columns of B.

Dragan Stevanovic On 1-eigenvectors of graphs

Solving PMEIG0

Let E be the full rank submatrix of D, and

U be the set of row indices of E within D.

For each z {1, +1}U

Solve Ey = z (there is unique such y )
Let x = Dy (a E0 vector whose U-restriction is z)
If x is a 1 eigenvector of A
Then output x

Dragan Stevanovic On 1-eigenvectors of graphs

Solving PMEIG0

Let E be the full rank submatrix of D, and

U be the set of row indices of E within D.

For each z {1, +1}U

Solve Ey = z (there is unique such y )
Let x = Dy (a E0 vector whose U-restriction is z)
If x is a 1 eigenvector of A
Then output x

The running time is controlled by the multiplicity of 0.

Although it is believed that almost all graphs have simple

eigenvalues only, the interesting ones have large multiplicity of 0:
the graph corresponding to the 2-coloring of a 4-uniform
hypergraph H = (V , E ) has multiplicity of 0 at least 2|E | |V |.

Dragan Stevanovic On 1-eigenvectors of graphs

Applying the algorithm

This simple algorithm can be used

to test the existence of a 1 eigenvector for any fixed eigenvalue
simply by applying it to A I instead of A.

Dragan Stevanovic On 1-eigenvectors of graphs

Applying the algorithm

This simple algorithm can be used

to test the existence of a 1 eigenvector for any fixed eigenvalue
simply by applying it to A I instead of A.

Its application shows that there is little hope to answer

the Wilfs implicit question of which regular graphs have
a 1 eigenvector corresponding to the smallest eigenvalue.

Dragan Stevanovic On 1-eigenvectors of graphs

Applying the algorithm

This simple algorithm can be used

to test the existence of a 1 eigenvector for any fixed eigenvalue
simply by applying it to A I instead of A.

Its application shows that there is little hope to answer

the Wilfs implicit question of which regular graphs have
a 1 eigenvector corresponding to the smallest eigenvalue.

The triangular graph L(K8 ) and the three Chang graphs are
all strongly regular graphs with the same parameters (28, 12, 6, 4)
and the same spectrum [12, 4(7) , 2(20) ].

Dragan Stevanovic On 1-eigenvectors of graphs

Applying the algorithm

This simple algorithm can be used

to test the existence of a 1 eigenvector for any fixed eigenvalue
simply by applying it to A I instead of A.

Its application shows that there is little hope to answer

the Wilfs implicit question of which regular graphs have
a 1 eigenvector corresponding to the smallest eigenvalue.

The triangular graph L(K8 ) and the three Chang graphs are
all strongly regular graphs with the same parameters (28, 12, 6, 4)
and the same spectrum [12, 4(7) , 2(20) ].

Yet, the algorithm responds that

each of the three Chang graphs has a 1 eigenvector of 2,
while L(K8 ) does not have such an eigenvector.

Dragan Stevanovic On 1-eigenvectors of graphs

Examples of eigenpartite graphs

Due to a simple defining condition,

constructions of eigenpartite graphs are abundant.

The simplest construction is to take two copies of a graph and add

edges between vertices in different copies whenever they are
adjacent in the original graph.

Dragan Stevanovic On 1-eigenvectors of graphs

Further examples

Let G have degree sequence S, and

let H have degree sequence T .

Let K be a bipartite graph with bipartition (U, V )

with (S, T ) as its degree sequence.

Superimposing G on U and H on V gives an eigenpartite graph.

Dragan Stevanovic On 1-eigenvectors of graphs

NEPS and eigenpartite graphs

Let
G = NEPS(G1 , . . . , Gn ; B)
If i is an eigenvalue of Gi with the eigenvector xi , i = 1, . . . , n,
then X
= 1 1 2 2 n n
B

is an eigenvalue of G with the eigenvector

x = x1 x2 xn .

Lemma
The set of eigenpartite graphs is closed with respect to the NEPS
of graphs.

Dragan Stevanovic On 1-eigenvectors of graphs

Partial Cartesian product of graphs
Let G = (V , E ) with S V and let H = (W , F ).
The partial Cartesian product G +S H has the vertex set V W ,
with two vertices (v1 , w1 ) and (v2 , w2 ) adjacent in G +S H
if either v1 = v2 S, (w1 , w2 ) F or (v1 , v2 ) E , w1 = w2 .

Dragan Stevanovic On 1-eigenvectors of graphs

Partial Cartesian product of eigenpartite graphs
The adjacency matrix of G +S H is

A = IS AH + AG I .

If x is an eigenvector of AG corresponding to ,
and y is an eigenvector of AH corresponding to 0, then

A(x y ) = (IS x) (AH y ) + (AG x) (Iy ) = x y .

Lemma
If H has eigenvalue 0, then the spectrum of G is contained in the
spectrum of G +S H for each S V (G ).

Corollary
If both G and H are eigenpartite, then G +S H is eigenpartite for
each S V (G ).

Dragan Stevanovic On 1-eigenvectors of graphs

Conclusions (1)

The Wilfs implicitly stated question of

Which regular graphs admit a 1 eigenvector of the smallest

eigenvalue?

does not appear to have a simple answer,

due to the difference in behavior between
the three Chang graphs and the triangular graph L(K8 ),
which all share the same local neighborhood structure,
yet each of the Chang graphs has a 1 eigenvector of 2,
while L(K8 ) does not have such an eigenvector.

Dragan Stevanovic On 1-eigenvectors of graphs

Conclusions (2)

The Wilfs explictly stated question of

Which graphs admit a 1 eigenvector?

has a simple combinatorial answer: those graphs which admit a

partition of their vertex set in two parts such that
each vertex has equally many neighbors in both parts (for
eigenvalue 0), or
there exists k so that each vertex has k more neighbors in its
part than in the other part (for eigenvalue k).

Dragan Stevanovic On 1-eigenvectors of graphs

Conclusions (3)

The bad news is that it is NP-complete to test this combinatorial

condition of eigenpartiteness.

Further, the set of eigenpartite graphs is surprisingly rich

and closed under many different compositions of graphs.

Thus, having in mind simplicity of the defining condition for

eigenpartiteness and the abundance of such graphs,
it appears unlikely that further characterizations of
graphs admitting 1 eigenvectors could be made.

Dragan Stevanovic On 1-eigenvectors of graphs