International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 2

ISSN: 2321-8169
484 - 488

_______________________________________________________________________________________________

Distributed Graph Isomorphism using Quantum Walks

Jayaprabha Yadav

Prof. Mrs. Aparna Junnarkar

Department of Computer Engineering

PES Modern College of Engineering
Pune, India
[email protected]

Department of Computer Engineering

PES Modern College of Engineering
Pune, India
[email protected]

AbstractGraph isomorphism being an NP problem, most of the systems that solves the graph isomorphism are constrained with some classes
of the graph, and do not work for all types of graphs in polynomial time. We exploited the two particle quantum walks on different classes of
graphs including strongly regular graphs which are co-spectral in nature. We simulated two particle quantum walks on graph using distributed
algorithm.
To show the effectiveness of the technique, we applied it to the large graphs derived from images using Delauney triangulation. The results show
a remarkable speedup for large data. The two-particle quantum walks is implemented in map-reduce programming technique which scales the
computation as the cluster get scaled to account Big data. We checked the isomorphism of the graphs with upto 100 vertices in polynomial time.
The system is scalable to accept big inputs from any other domain in graph format.
Keywords- Graph isomorphism; Distributed computing; Map-Raduce; Quantum Walks.

__________________________________________________*****_________________________________________________
I.

INTRODUCTION

In every application always there is a need to search and

hence obviously we need to compare the objects. Nowadays
the objects and the relations among them are getting more
and more complex. This complexity can be represented by a
data structure called graph. Two graphs are isomorphic if
one can be obtained from another by relabeling of the
vertices. Graph isomorphism is a basic methodology used in
various applications for comparing objects.
The graph isomorphism is basically a hidden subgroup
problem of the permutation group and is considered to be
NP-complete. Checking weather two graphs are isomorphic
is an NP-intermediate. A number of polynomial time
methods have been suggested for finding isomorphic graphs.
Some of these methods uses spectral analysis of graphs [1]
and applies to specific classes of graph and doesnt work for
co-spectral graphs which are strongly regular graphs. The
best general classical algorithm to date runs in O(CN 1/2log N ),
C is a constant value and N is the number of vertices in the
graphs being compared [2].
The two-particle quantum walks with hardcore bosons [3]
can find isomorphism between two strongly regular graphs.
The two particle QW is applied to graphs up-to 60 vertices
and checked to be working correctly to check isomorphic
graphs. The problem is to apply this method to graphs with
more number of vertices.
We exploited a map-reduce algorithm for two particle
quantum walks to calculate the evolution factor which can
differentiate even the SRGs. This enables us to determine the
isomorphism with highest responses in as O(nlogn*s*(1/p))
where n is m*m and m is no. of vertices in graph, s is the

number of CPUs and p is the communication delay for one

heartbeat time independent of the number of vertices in the
graph. The technique is applicable to a variety of graphs that
can be prepared from structures in image processing as well
as other domains. Our approach is not restricted to any
particular method or dataset; rather it provides the basis for
scaling the number of traditionally compute-intensive graph
processing operators from the hundreds or at most thousands
applied in current practice to millions. We first applied the
Delaunay triangulation method on images using batik library
to derive graphs from images. Then we derived the
adjacency matrix of the image. We demonstrate the efficacy
of the approach by scaling graph isomorphism to multiple
graph
classes
employing
Delaunay
triangulation
preprocessing of the images representing graphs. The
isomorphic images can be detected even if the target image
graph undergoes multiple transformations.
We believe the present work will accelerate the state of
the art in object detection by increasing the number of visual
categories by an order of magnitude or more while
simultaneously reducing run times by a comparable factor.
We demonstrate our approach in an implementation that
achieves respectable performance on a standard benchmark
for graph isomorphism, and exhibits graceful progress in
performance with larger, automatically generated datasets
consisting of tens of thousands of graph data.
II.

RELATED WORK

These include probabilistic methods [4] and graphspectral methods which utilizes the Eigen values and
eigenvectors of the Laplacian matrix [5]. The auxiliary graph
484

IJRITCC | February 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 2

ISSN: 2321-8169
484 - 488

_______________________________________________________________________________________________
method by David Emms [6] has the complexity dominated
by the simulation of the walk and is O(|V|6 ).By comparison,
for the Umeyama algorithm [7] the complexity of computing
the spectra of the two adjacency matrices is O(|V|3)and the
complexity of using the Hungarian search method is again
O(|V|3). This is typical of most graph-spectral methods.
Other approximate algorithms have similar complexity. For
instance, Gold and Rangarajan's [8] algorithm has
complexity O(|EG ||EH |) where EG and EH are edges in G
and H graph respectively. There are of course more
sophisticated inexact graph matching algorithms available.
Douglas and Wang [9] have recently explored the use of
discrete quantum walks for graph isomorphism. Their idea is
to use the probability amplitudes associated with the states of
walks on separate graphs as node-attributes. The amplitudes
for corresponding steps of the walk on different graphs are
compared to establish isomorphism or similarity. The
algorithm has complexity O (|V|7).
Two-particle interacting boson walks distinguish all nonisomorphic pairs of SRGs, the SRGs with up to 64 vertices [3].
If the two-particle quantum walks can check for all types of
isomorphic graphs, then the Graph isomorphism is a
polynomial problem and not an NP can be proved.
III.

t u (t) u

(1 )

uV

Unlike the classical walk, the quantum walk is not a Markov

chain. Given an initial state for the walk,
solved to give

t e iLt 0

, Eq. (1) can be

(2)

,where L=DA is the Laplacian matrix, A is the adjacency

matrix and D is the diagonal degree matrix.
If we restrict ourselves to single-particle states, we find that
adjacency matrix elements give Hamiltonian,

i | H | j = Aij

(3)

Hence, we can easily identify the a single-particle Hamiltonian

H1P = A

( 4)

And Quantum Walk time evolution operator can be defined as

U = eitH

(5)

TECHNICAL DETAILS

The architecture described in this paper applies to a wide range

of graph types, e.g., spectral and Strongly Regular graphs
(SRG). The application and experiments presented here make
use of the map-reduce model (MR) of Hadoop. Many other
graph matching applications can be adapted to use our
approach, including hidden sub graph detection [10], hidden
group in financial transaction network [11], and
communication networks [12].
A. One particle continuous time quantum walks
The state space for the continuous-time quantum walk on a
graph, G = (V, E), is the set of vertices, V, as is the case for the
classical random walk. In addition, transitions only occur
between adjacent vertices. If the walk is at a vertex u, it moves
to adjacent vertices at a rate proportional to 1/d(u), where d(u)
is degree at u. The basis states for the continuous-time quantum
walk are vectors corresponding to particular vertices, as is the
case for the classical random walk and unlike the discrete-time
quantum walk where basis states correspond to edges. The
basis state corresponding to the walk being at u V is

written, in Dirac notation, as

. A general state of the walk is
a complexlinear combination of these basis states and so the
state of the walk at time t is given by a vector, which we write
component wise as

An SRG is a graph in which (a) all vertices have the same

degree, (b) each pair of neighboring vertices has the same
number of shared neighbors, and (c) each pair of nonneighboring vertices has the same number of shared neighbors.
This definition permits SRGs to be categorized into families by
four integers (N, k, , ), each of which might contain many
non-isomorphic members. Here, N is the number of vertices in
each graph, k is the degree of each vertex (k regularity), is the
number of common neighbors shared by each pair of adjacent
vertices, and is the number common neighbors shared by
each pair of nonadjacent vertices.
The adjacency matrix of all strongly regular graphs satisfies the
useful relation [13]

A2 = (k )I + J +( )A

(6)

where I is the identity matrix, A is adjacency matrix, and J is

the matrix of all 1s. We can write

An = n I + n J + n A

(7)

,where , , are functions only of the family parameters.

That is, all SRGs of the same family have the same coefficients
and same Hamiltonian and hence the algorithm based on
single-particle quantum evolution fails to distinguish any nonisomorphic SRGs that are in the same family.

485
IJRITCC | February 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 2

ISSN: 2321-8169
484 - 488

_______________________________________________________________________________________________
B. Two particle continuous time quantum walks
First, we note that we may write the Hamiltonian for any twoboson Quantum Walk as

1
H 2B = (I + S)(AA) + UR
2
Where

(8)

( ) ( ) is a Kronecker sum,

the matrix special case of a direct sum, and

S ij ji
i, j

R ii ii

(9)

Ai , j X i , j | i, j is an edgein the graph

The double occupancy matrix R is calculated for every

graph,

Ri , j X ii, jj | X i , j adjacency matrix

Using equation (5) we get,

U = e it [ 2(I +S) (AA) +UR]

very large data sets. It also includes an implementation of MapReduce, a programming model designed for processing big
data sets on large clusters.
Input for the map reduce is <key, value> pair i. e. two files
containing adjacency matrices derived from two different graph
data. It is actually edge list of the graph.
<vertex1, vertex2, 1>

( 10 )

Expanding as a power series in t, we have

Kronecker sum B is calculated using map-reduce

implementation.
m

B {xi , j xi , j i m x, i m y x m i, y m i }
( it)n
1
i, x, y
U=
[ (I + S)(A A) + UR] n ( 11 )
n!
2
n=0

Numerically, it is determined [3] that a sixth-order expansion

was necessary, and that the term B2RB3 can be used to
distinguish the graphs up to N=40 vertices, where
. The component of evolution factor that
differentiate the value of evolution factor can be used to
represent the graph so that it can be compared with other
graph's evolution factor. This solution is used to formulate the
MapReduce algorithm to differentiate the graphs with number
of vertices upto 100.
C. Map-reduce algorithm for Quantum Walks
MapReduce is a programming model invented for
distributed data processing on large clusters. As part of two
particle Quantum Walk on graph, two algorithms
MapReduceKron and MapReduceMatrix are created for
computing Kronecker sum B and multiplications of matrices.
An algorithm for calculations on matrices requires a large
amount of memory to store the input operands, intermediate
results and output. Traditionally calculations on large arrays
require a uniform address space on a single machine and these
operations are performed on supercomputers with software
customized to exploit its particular memory and interconnect
architecture. When memory is exhausted during a computation
number of packages may have a poor performance or may not
work at all, as memory available in most machines is
insufficient by orders of magnitude.
In this paper we present a distributed implementation of
two-particle quantum walks for large graphs on commodity
hardware using a shared compute layer. Our prototype
implementation runs on Hadoop clusters. Hadoop is an open
source distributed computing framework developed by Apache
Software Foundation. It includes fault tolerant, distributed file
system namely HDFS, designed for high-throughput access to

A
distributed
multiplication
using
map-reduce
programming method is implemented. Evolution factor can be
calculated using the differentiating component as B2RB3.
The evolution factors for two graphs are U1and U2 such
that they are elements of the matrices. These values are placed
in the separately and sorted. The difference of these lists if zero
then the two graphs are isomorphic else are non-isomorphic.
It is found that evolution factor can identify distinct graphs
upto 64 vertices for SRGs and upto 100 vertices for images. T
he results are discussed in next section.
IV.

EXPERIMENTAL RESULTS

In the following, we use the term one-particle algorithm to

refer to the graph-isomorphism algorithm described in [2]
utilizing graphs in SRG dataset. To execute map-reduce
system we prepare the cluster of 5 machines having dual core
and 2GB RAM, single rack based structure and Hadoop 2.6.0
distributed system setup. One of the machine is namenode and
resource-manager while others slaves to work as datanode and
nodemanager. Replication factor set to four. By performing
several experiments, we compare the performance of the oneparticle algorithm with the distributed two-particle algorithm
described in the previous section, utilizing map-reduce model
of programming as mentioned earlier. First, we demonstrate
that the one-particle algorithm compares favorably with the
two-particle algorithm on the SRG dataset. Second, we show
that two-particle distributed algorithm can scale graphisomorphism to hundreds of thousands of graph nodes and
provide insight into the trade-offs involving accuracy, memory
and computation time. We use our system for data derived
from real-world data. For this we take images from the COIL
database and construct Delaunay triangulations. We consider
12 different objects from the database. We used our system to
compare images.
486

IJRITCC | February 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 2

ISSN: 2321-8169
484 - 488

_______________________________________________________________________________________________
A. Dataset and implementation Details
We first employed the SRG dataset upto 64 vertices to test
our implementation, which work correctly to identify the nonisomorphic SRGs from the same class of SRGs. The system
required more time as we checked on a cluster of multiple
nodes. In the case of a distributed system the network overhead
is more than that of computations for small datasets. But the
task of complex computation can be distributed in easy steps
and same application can handle big data. We can see in the
fig. 1 that the same number of vertices requires more time if
the distributed file system block size reduced.

The continuous time quantum walks are better than discrete

time quantum walks as continuous time walks corresponds to
vertices [2]. But while implemented in map-reduce it is
observed that more dense the adjacency matrix, more time it
takes for distributed computation. That is the time required is
more as the connectivity of the graph is increased.
B.

Accuracy, speed and memory

Our implementation is able to detect isomorphism between

graphs of all classes with 100% accuracy. As the number of
vertices is increased, the input is increased and thus number of
map tasks is increased so the time required to collect result
from multiple nodes is greater and hence total time required
for computation is increased. But the system withstands big
input data.
CONCLUSIONS

Figure 1. The time required for checking isomorphism to the

number of vertices of graph with sparse connectivity and dense
connectivity.
We employed the standard image dataset, COIL-20, to test
our system. The COIL-20 dataset contains images from
different categories with 20 images for training and validation.
The triangulated graphs can be used to check isomorphism. We
can consider the triangulated images close to SRG class of
graph because certain set of triangular graphs are also SRGs.

Our key contribution is a distributed approach to graph

isomorphism that replaces computational density with
distributed simple operations by using an efficient MapReduce design. This approach is applicable to a variety of
graphs of different classes and derived from different
application domain. Through extensive empirical tests on
distributed graph isomorphism system, we have shown that (a)
the system performs comparably better to the original single
machine algorithm for big data, (b) performance upgrades
gracefully as the number of vertices in graph is increased, and
(c) up to 100 vertices graph can be processed on a single
machine in polynomial time.

REFERENCES

Figure 2. The time requires for detecting isomorphism of

SRGs from E. Spence Dataset [14] with 64MB block size and
128MB block size.
Fig. 2 shows the results for sparse matrix that is the graph
with less number of edges requires less time as compared to
that for a dense graph, which has more number of edges. In
dense graph the computations are more as we are considering
the adjacency edges for the calculation of evolution factor.
The Figure 2 shows the results for dense graphs derived from
images on multiple node clusters.

[1]

B. D. McKay, Practical graph isomorphism, Congr. Numer. 30

(1981) 4587.

[2]

A. Robles-Kelly, E. R. Hancock, Graph edit distance from

spectral seriation, IEEE Trans. Pattern Anal. Mach. Intell. 27 (3)
(2005) 365378.

[3]

John King Gamble, Mark Friesen, Dong Zhou, Robert

Joynt, and S. N. Coppersmith, Two-particle quantum walks
applied to the graph isomorphism problem", Phys. Rev.
A 81, 052313 - Published 13 May 2010.

[4]

W. J. Christmas, J. Kittler, M. Petrou, Structural matching in

computer vision using probabilistic relaxation, IEEE Trans.
Pattern Anal. Mach. Intell. 17 (8) (1995) 749764.

[5]

A. Robles-Kelly, E. R. Hancock, Graph edit distance from

spectral seriation, IEEE Trans. Pattern Anal. Mach. Intell. 27 (3)
(2005) 365378.

[6]

D. Emms, S. Severini, R. C. Wilson, and E. R. Hancock, Graph

matching using the interference of continuous-time quantum
walks", Pattern Recog. 42, 1988 (2009).

[7]

S. Umeyama, An eigen decomposition approach to weighted

graph matching problems, Pattern Anal. Mach. Intell. 10 (5)
(1988) 695703.
487

IJRITCC | February 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 2

ISSN: 2321-8169
484 - 488

_______________________________________________________________________________________________
[8]

[9]

S. Gold, A. Rangarajan, A graduated assignment algorithm for

graph matching, IEEE Trans. Pattern Anal. Mach. Intell. 18 (4)
(1996) 377388.

[12] J. Baumes, M. Goldberg, M. Magdon-Ismail and A. Wallace,

B.
Douglas and J.
Wang, Journal of Physics A: Mathematical and Theoretical 41,
075303 (2008).

[13] C. Gosil and G. Royle, Algebraic Graph Theory (Springer, New

[10] R. Jozsa, Quantum factoring, discrete logarithms, and the hidden

subgroup problem, Comput. Sci. Eng. 3 (2) (2001) 3443.

[11] Yuhua

Li, Dongsheng Duan, Guanghao Hu, Zhengding

Lu,Discovering Hidden Group in Financial Transaction
Network Using Hidden Markov Model and Genetic Algorithm,
Fuzzy System and knowledge Discovery,2009,pp. 253-258.

Discovering Hidden groups in communication Networks,

Intelligence and security Information, 2004.
York, 2001)
[14] E. Spence, Strongly regular graphs on at most 64 vertices.

Available at: http://www. maths. gla. ac. uk/~es/srgraphs. html].

[15] Columbia Object Image Library (COIL-20)," S. A. Nene, S. K.

Nayar and H. Murase, Technical Report CUCS-005-96,

February 1996.

488
IJRITCC | February 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________