Fourier Transform For Signals On Dynamic 

Graphs 
Arash Golibagh Mahyari , Selin Aviyente 
mahyari@msu.edu , aviyente@egr.msu.edu Dept. of Electrical and Computer Engineering Michigan State 
University East Lansing, MI 48824, USA Tel: (517) 355-7649, Fax: (517) 353-1980 
Abstract—Signal  processing  on  graphs  offers  a new way of analyzing multivariate signals. The different relationships among 
the  sources generating the multivariate signals can be captured by weighted graphs where the nodes are the signal sources and the 
edges  correspond  to the relationships between these signals. Classical signal processing concepts need to be adapted to signals on 
graphs.  In  this  paper,  we  propose  a  graph  Fourier  transform  for  signals  on  dynamic  graphs,  where  the  relationships  vary  over 
time. The proposed transform is evaluated on both simulated and real dynamic social networks with signal defined on its nodes. 
I. INTRODUCTION 
Classical  multivariate  signal processing methods analyze signal properties over time. However, examining signal 
charac-  teristics  using  the information about their sources is of interest in many applications. For example, it is often 
desirable  to  analyze  the  spectrum  of  the  air  temperature  values  over  a  set  of  sensors  rather  than  evaluate  the 
spectrum  of  a  sensor’s values over time. Moreover, multivariate signal processing methods consider the dependency 
among the signals without making any assumption about their sources. These methods assume that the sources are in 
a  regular  domain,  and  their  relationships  are  uniform.  However,  in  most  applications,  the  relationships  among  the 
sources  are  not  the  same and needs to be taken into account while processing the signals generated by these sources. 
For  instance,  in  social  networks,  friendships  among  individuals  influence  their  behavior  or  opinion. Therefore, it is 
important  to  consider  their  friendships  while  analyzing  the  individuals’  behaviors  and  opinions.  One  way  to 
represent  the  relationship  among  the  sources  of  multivariate  signals  is to use graphs. Recently, this has led to a new 
research area, called signal processing on graphs [1]. 
Signal  processing  on  graphs  assume  that the signal sources are the vertices of a graph and each node has a value, 
and  edges  show  the  relationship  among  the  sources  [1].  Figure  1  shows  a  typical  signal  on  a  graph.  Signal 
processing  on  static  graphs  have  looked  at  the  signal  samples  over  sources  assuming  that  the  relationship  among 
sources  does  not  change.  The  irregular  topology  of  signals  on  graphs  makes  use  of  well-  known  signal  processing 
methods  such  as  filtering,  denoising,  and  compressing  difficult  [2],  which  is  why,  signal  processing  concepts  like 
graph filtering [3], downsampling [2], inpainting [4], sparse signal representation [5], and Fourier Transform 
This work was in part supported by the National Science Foundation under Grant No. CCF-1218377. 
have been defined based on the graph Laplacian matrix [1], [6], [7]. 
However,  the  relationships among the sources may change with time, and can be represented by dynamic graphs. 
For  example,  wireless  sensor  networks  are  redeployable  sources  whose  proximities  may  change  over  time.  Social 
networks  are  another  example  of  dynamic  graphs,  where  members’  friend-  ships  change  with  time.  Generally,  the 
extension  of  signal  processing  concepts for graphs relies on the graph Laplacian matrix. The graph Laplacian matrix 
of  dynamic  graphs  change  with  their  adjacency  matrices,  and  cannot  be  directly  used  for  signal  processing  on 
dynamic  graphs.  Therefore,  in  this  paper,  we  introduce  an  approach  to  find  a  common  subspace  representation  for 
multiple  graphs  across  time  [8],  [9]  and  define  a  common  Laplacian  matrix  for  them.  The  eigenvectors  of  this 
modified Laplacian matrix are employed to derive Fourier transform for signals on dynamic graphs. 
Fig. 1. An example of signal on a graph [1]. 
II. GRAPH THEORY 
Let  G  =  {V,E,A}  be  an  undirected,  weighted  graph  with  a  set  of  vertices  V,  |V|  =  N,  a  set  of  edges  E,  and  the 
adjacency  matrix  A.  A  signal  on  the vertices of the graph f : V −→ R is defined as a vector f ∈ RN×1, where its ith 
component is the signal value on the ith vertex. 
The normalized graph Laplacian of the graph G is defined as L = D− 1 2 
(D − A)D− 1 2 
, where D is the degree matrix. The complete set of orthonormal eigenvectors of the 
Laplacian matrix λ purposes 0 
≤ {u 
λ 
1 such k } 
≤ k=0,1,...,N−1 
as ... graph ≤ and their corresponding eigenvalues 
λ 
N−1 spectral have clustering been used [10]. for The numerous spectral clustering algorithms project the graph 
Laplacian matrix onto a new subspace such that the clusters within the graph become distinguishable. through 
solving the The optimization basis of this problem 
new subspace U∈R 
min 
N×K 
is obtained tr(U LU) 
978a1a4799a8297a4/14/$31.00 ©2014 IEEE 2001 Asilomar 2014 
 
subject to U U = I, where I is a unitary matrix, and K is the number of clusters. 
III. DYNAMIC GRAPH FOURIER TRANSFORM 
Dynamic graphs is a time series of static graphs 
G(t) = {V,E(t),A(t)} ∀t ∈ {1,2, ..., T}, 
where  it  is  assumed  that  the  number  of  vertices  remains  constant  while  the  edges  and  the  adjacency  matrices  vary 
with time. 
The  graph  Fourier  transform  theory  relies  on  the  spectral  content  of graphs. However, since the spectral content 
of  the  time  series  of  graphs  change  with  time  in  dynamic  graphs,  a  common  subspace  has  to  be  found  in  order  to 
obtain a unique spectral representation for all time steps. 
A. Common Subspace Across Time-Varying Graphs 
In  [9],  the  vertex  connectivity  of  a  graph  G  =  {V,E,A}  with  |V| = N is represented in a N-dimensional subspace 
defined  by  N  eigenvectors  of  the  graph  Laplacian  matrix.  The  information  carried  by  the  graph  connectivity  is 
summarized  in  this  N-dimensional  subspace.  In  this  paper,  we  propose  an  approach  based  on  the  Grassmann 
manifold  to  obtain  a  unique  subspace  representing  a  time  series  of  graphs  [9].  A  Grassmann  manifold  is  a  set  of 
N-dimensional subspaces in Rn. Each graph G(t) = {V,E(t),A(t)} is a unique point on the Grassmann manifold. 
Let U(t) ∈ RN×N be the matrix of the eigenvectors of the Laplacian U time, matrix ˆ 
be where the L ˆ 
basis is matrix the located. vectors modified L(t) The corresponding that graph common span G ˆ 
a common with subspace to the the modified tth subspace span( 
graph U) Laplacian ˆ across G(t). repre- Let all 
sents this subspace a point on span( 
the U) Grassmann ˆ 
manifold. The goal is to find such that it is close to all subspaces span(U(t)) ∀∈{1,2, ..., T} by minimizing the 
projection distance defined as follows [9]: 
d2 
proj 
( 
U,U(t) ˆ 
) 
= 
( 
U ˆ 
) 
= NT − 
∑ 
T t=1 
( N U ˆ 
− tr 
U ˆ 
U(t)U ) 
(t) 
. 
(1) 
However,  in  order  to  preserve  the  information  about  the  vertex  connectivity  of  each  individual  graph,  the 
Laplacian quadratic form 
∑ 
T t=1 
( 
U ˆ 
) 
(3) 
where 
L ˆ 
= 
∈R 
ˆU∈
min 
N×K 
tr 
L ˆ 
U ˆ 
T∑ 
t=1 
L(t) − α 
T∑ 
U(t)U (t). t=1 
B. Dynamic Graph Fourier Transform 
The  classical  Fourier  transform  is  the  inner  product  of  the  signal  with  the  eigenfunction  of  the  1D  Laplace 
operator  ,  e−j2πωt.  Similarly,  the  graph  Fourier  transform is defined as the inner product of the signal on the graph 
f  and  the  eigenvectors  of  the  corresponding  Laplacian  matrix [1]. However, dynamic graphs consist of a time series 
of  graphs  with multiple Laplacian matrix. In order to obtain a unique frequency we propose representation to use the 
Laplacian for the signals matrix 
L. on ˆ 
the set of graphs, 
 Let  f  (t)  ∈  RN×1  be  a  signal  on  the  vertices  of  G(t)  at  time  step  t,  where  f  (t)(n)  is  the  signal  value  on the nth 
vertex of the tth graph. The dynamic graph Fourier transform of the signal f (t) is defined as: 
f ˆ 
(t)(λ 
k k 
= ) = 0,1, F(f ..., ) = 
N ∑ 
− N−1 n=0 1,t = f (t)(n)û∗ 1,2, ..., k 
T. (n) 
(4) 
whereû {1,2, ..., k 
T} is shows the kth eigenvector the variation of of 
the L. ˆ 
kth f ˆ 
(t)(λ spectral k 
) ; ∀t content ∈ 
of  the  signal  on  the  set  of  graphs.  Similar  to  classical  Fourier  transform,  smaller  eigenvalues  correspond  to  low 
frequencies. The smallest eigenvalue component of the signal. λ 
0 Consequently, , which is zero, represents the DC associated with the smallest eigenvalue the eigenvector, is 
constant and equal 
u 
0 
, 
to is, √ 
the 1 N 
lower at each vertex. Generally, the smaller the eigenvalue the frequency it represents and the eigenvector associated 
with it varies slowly over time. 
Similarly, the inverse Fourier transform is defined as: 
f (t)(n) = F−1( 
f ˆ (t)(λ)) n = 0,1, ..., N − = 
1,t ∑ 
= N−1 k=0 
1,2, f ˆ 
(t)(λ 
..., T. k 
)û 
k 
(n) 
(5) 
IV. EXPERIMENTAL RESULTS AND DISCUSSION 

In  order  to  evaluate  the  performance  of  the  proposed  dynamic  graph  Fourier  transform,  a  simulated  dynamic 
graph  and  the  corresponding  signals  on  its  vertices  as  well  as  a  real dataset is used in this section. Some discussion 
regarding the interpretation of the outcomes and the application of this new transform are provided as well. 
A. Simulated Dataset 
A  path graph is generated with N = 128 vertices and T = 128 time steps. To change the graph structure over time, 
two  edges  are  selected  at  each  time  step  randomly  from  an  uniform  distribution  and  set  to  one  while  other  edges 
remain unchanged. The signal on vertices is also defined as f (t) = u 
(1) 5 
+ u 
(10) 15 fifth eigenvector of the ;t = 1,2, ...,32, first graph, and where u 
(10) 15 
u 
(1) 5 
is the 
ˆ U U(t)U (t) 
tr 
( ∑ 
T t=1 
U ˆ 
L(t) U ˆ 
) 
must  be  minimized  as  well  [9].  Therefore, the problem of finding the common 
subspace can be addressed through solving the following minimization problem: 
min U∈R ˆ 
tr 
( 
U ˆ 
L(t) U ˆ 
) 
+α 
[ 
KT N×K − 
∑ 
sumT 
T t=1 
t=1 
( tr 
U ˆ 
)] 
(2) 
subject to 
Rearranging the terms in Eq. 2 results in the following equation: 
tr 
U ˆ 
U(t)U (t) U ˆ 
U ˆ 
= I. 
is the 15th eigenvector of the Laplacian matrix of the graph at t = 
10. 
2002 
 
For the time interval t = 33,34, ...,64, the signal is f (t) = u of (1) 5 
the + 40th u 
(10) 15 
+ graph. u 
(40) 40 
Similarly, , where the u 
(40) 40 
signal is the for 40th other eigenvector time steps are defined as: f (t) = u 
(1) 100 
(40) 
f (t) = the spectral u 
(30) 100 
content + u 
(100) 45 
of + u 
30 
;t = 65,66, ...,96, and ;t the = 97,98, signal on ...,128. Figure 2 shows the dynamic graphs for the whole time interval, 
t = 1,2, ...,128. The signal f (t) is also corrupted with white Gaussian noise with −20dB, and its graph Fourier 
transform is shown in Fig. 3. 
0.9 
0.8 
0.7 
0.6 
0.5 
0.4 
0.3 
0.2 
0.1 
Fig. 2. Time versus spectral content representation of the graph signal. 
1.2 
1 
0.8 
0.6 
0.4 
0.2 
Fig. 3. Time versus spectral content representation of the noisy graph signal. 
B. Real Dataset: Mobility traces of taxi cabs in San Francisco 
 The  real dataset used in this experiment contains mobility traces of taxi cabs in San Francisco, USA in May 2008 
[11],  [12].  500  taxi  cabs  are  equipped  with  GPS  devices,  and  their  location  and  their  occupancies  are  collected 
approximately  every  30  seconds.  Here,  we  use  GPS  coordinates  of  64  cabs  to  calculate  their  proximity  and  build 
dynamic  graphs  with  128  time  steps.  Each  time  step  is  the  average  of  information  over  30  minutes.  The  cabs’ 
occupancies  are  used  as the signal on the graph at each time step. The dynamic graph Fourier transform is computed 
for analyzing the time-varying the occupancies of the taxi cabs. 
Figure  4(a)  shows  the  dynamic  graph  Fourier  transform  of  the  occupancies  of  the  cabs,  where  the  first  spectral 
component has the highest value. In order to better demonstrate other 
20 
40 l a r t c e p S 
60 
80 
100 
120 
20 40 60 80 100 120 
20 
40 l a r t c e p S 
60 
80 
100 
120 
20 40 60 80 100 120 
Time 
Time 
0.8 
(b) 
Fig. 4. (a) The dynamic graph Fourier transform of taxi cabs; (b) The dynamic graph Fourier transform of taxi cabs when DC 
component is suppressed. 
spectral  components  of  the  cabs  occupancies,  the  DC  com-  ponent  is  suppressed  and  the  modified  spectrum  is 
displayed  in  Fig.  4(b).  This  figure  shows  the  variation  of  the  different spectral components over time. For instance, 
the  3rd  and  5th  spectral  components  are  important  within  the  time  interval  (16,25),  14th  and  47th  are  significant 
within  the  time  interval  (1,10).  Eigenvectors  carry  information  about  the  structure of graphs, which is why the 14th 
eigenvector  of  the  common  Laplacian  matrix  is  shown  in  Fig.  5.  Figure  5  indicates  that  nodes  7,14,43  or 
equivalently  the  taxi  cabs  have  contributed  mostly  to  the  significant  value  of  the  14th  spectral  content  within  the 
time interval (1,10). The signals on 7th, 14th, and 43th are displayed over time in Fig. 6. This analysis shows that the 
correspondence  between  the  spectral  components and the individual nodes in the graph can be used to determine the 
time changing energy distribution of different nodes in the network. 
V. CONCLUSIONS 
Graphs  are  flexible  tools  to  model  the  signals and their sources which are in topologically complicated domains. 
Sig-  nal  processing  on  graphs  has  provided  a  framework  to  analyze  such  signals  by  modifying  the  classical  signal 
processing  concepts.  However,  many  applications  like  wireless  sensor  net-  works  are  represented  by  dynamic 
graphs, where the topology 
2003 
10 
0.6 
20 
30 
40 
50 
60 
0.4 
l a r t c 
0.2 
e p S 
20 40 60 80 100 120 
0 
−0.2 
−0.4 
−0.6 
−0.8 
10 
20 l a r t c e p S 
30 
40 
50 
60 
20 40 60 80 100 120 
Time 
4 
(a) 
3 
2 
1 
0 
Time 
 
0.4 
0.2 
0 
−0.2 
−0.4 
−0.6 
−0.8 
Node 
Fig. 5. The 14th eigenvector of the modified Laplacian matrix 
1 
0.8 
0.6 
0.4 
0.2 
Time 
Fig. 6. The signals on 7th, 14th, and 43th nodes. 
of  sources  changes  with  time.  In  this  paper,  we  introduced  a  new  approach  to  obtain  a  common  subspace 
representation  for  a  time  series  of  graphs.  Eigenvectors  of  the  common  subspace  obtained  from  a  modified 
Laplacian  matrix  is  used  as  the  bases  to  calculate  the  graph  Fourier  transform  of  the signals on the dynamic graph. 
Future work will focus on the graph Fourier transform for dynamic graphs with large number of time steps. Once the 
number  of  graphs  increases,  the  common  subspace  across  graphs  has  to be close to a larger set of subspaces, which 
reduced  the  accuracy  of  finding  the  common  subspace.  Thus,  finding  the  modified  graph  should  be  limited  to  the 
graphs within a moving window. 
REFERENCES 
[1] D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Van- dergheynst, “The emerging field of signal processing on 
graphs: Ex- tending high-dimensional data analysis to networks and other irregular domains,” Signal Processing Magazine, IEEE, 
vol. 30, no. 3, pp. 83–98, 2013. [2] H. Q. Nguyen and M. N. Do, “Downsampling of signals on graphs via maximum spanning 
trees,” Signal Processing, IEEE Transactions on, 2014. [3] A. Sandryhaila and J. M. Moura, “Discrete signal processing on 
graphs: 
Graph filters,” submitted for publication, 2013. [4] S. Chen, A. Sandryhaila, G. Lederman, Z. Wang, J. MF Moura, P. Rizzo, 
J. Bielak, J. H. Garrett, and J. Kovacevic, “Signal inpainting on graphs via total variation minimization,” in Acoustics Speech and 
Signal 
−1 
ˆ L. 
0 
0 20 40 60 80 100 120 140 
0 10 20 30 40 50 60 70 
Processing (ICASSP), 2014 IEEE International Conference on. IEEE, 2014, pp. 8267–8271. [5] D. Thanou, D. Shuman, and P. 
Frossard, “Learning parametric dictio- naries for signals on graphs,” Signal Processing, IEEE Transactions on, vol. 62, no. 16, 
pp. 3849–3862, 2014. [6] D. I. Shuman, B. Ricaud, and P. Vandergheynst, “A windowed graph fourier transform,” in Statistical 
Signal Processing Workshop (SSP), 2012 IEEE. IEEE, 2012, pp. 133–136. [7] B. Ricaud, D. I. Shuman, and P. Vandergheynst, 
“On the sparsity of wavelet coefficients for signals on graphs,” in SPIE Optical Engineering + Applications. International Society 
for Optics and Photonics, 2013, pp. 88581L–88581L. [8] X. Dong, P. Frossard, P. Vandergheynst, and N. Nefedov, “Clustering 
on multi-layer graphs via subspace analysis on grassmann manifolds,” in Proceedings of the 1th IEEE Global Conference on 
Signal and Information Processing (Global SIP). IEEE, 2013, pp. 993–996. [9] X. Dong, P. Frossard, P. Vandergheynst, and N. 
Nefedov, “Clustering on multi-layer graphs via subspace analysis on grassmann manifolds,” Signal Processing, IEEE 
Transactions on, vol. 62, no. 4, pp. 905–918, 2014. [10] A. Y. Ng, M. I. Jordan, and Y. Weiss, “On spectral clustering: Analysis 
and an algorithm,” Proceedings of Advances in Neural Information Processing Systems. Cambridge, MA: MIT Press, vol. 14, pp. 
849–856, 2001. [11] Michał Pi ́orkowski, Natasa Sarafijanovic-Djukic, and Matthias Gross- glauser, “A parsimonious model of 
mobile partitioned networks with clustering,” in Communication Systems and Networks and Workshops, 2009. COMSNETS 
2009. First International. IEEE, 2009, pp. 1–10. [12] “Dataset of mobility traces of taxi cabs in san francisco, usa,” 
http://crawdad.cs.dartmouth.edu/ crawdad/epfl/mobility/. 
2004