6th Mediterranean Conference on Embedded Computing
Resistive Circuits Analysis by Using Graph Spectral
Decomposition
Milos Dakovic, Ljubisa Stankovic,
Budimir Lutovac
University of Montenegro
Podgorica, Montenegro
Email: {milos, ljubisa, budo}@ac.me
AbstractA method for resistive circuit analysis based on the
spectral decomposition of the corresponding graph is proposed.
It is shown that the Laplacian matrix can be used in order to
calculate node potentials. Based on the Laplacian eigenvalues and
eigenvectors it is possible to decompose complex resistive circuit
into smaller, weakly connected sub-circuits.
KeywordsGraph spectra; Electrical networks; Resistive cir-
cuits; Spectral decomposition; Laplacian
I. I NTRODUCTION
Graphs theory have been used in wide variety of problems
including electrical networks, social networks, machine learn-
ing, communication networks, signal and image processing.
An electrical network is frequently modeled as a collection
of interconnected two-pole components. This model corre-
sponds to the weighted graph where edges correspond to the
individual components and edge weights correspond to the
component parameters. There exist many techniques for circuit
analysis based on graph incidence matrix, node voltages, graph
trees, independent contours, etc. [1].
Spectral graph theory [2] is emerging field of graph theory,
theoretically developed over past 60 years with many recently
developed applications [3][6]. In [3], graph Laplacian is used
to perform Kron reduction of the electrical network. The effec-
tive graph resistance and closed-form solution for equivalent
circuit resistance, are developed in [4], [5]. Signal processing
on graphs is reviewed in [6]. Graph spectra is obtained by
eigen-decomposition of the corresponding Laplacian matrix
[2], [6], [7].
In this paper, we will consider resistive electrical network. It
is shown that network voltages and currents can be calculated
by using the spectral decomposition of the Graph Laplacian
matrix.
This research is supported by the Montenegrin Ministry of Science, project
grant CS-ICT New ICT Compressive sensing based trends applied to:
multimedia, biomedicine and communications.
MECO2017, Bar, Montenegro
Ervin Sejdic Tomislav B. Sekara
University of Pittsburgh University of Belgrade
Pittsburgh, Pennsylvania, USA Belgrade, Serbia
Email: esejdic@ieee.org Email: tomi@etf.rs
The proposed method is demonstrated on two example cir-
cuits. It will be shown that spectral decomposition can be used
to identify (almost) independent part of the circuit (sub-circuit
with weak connections to the remaining part of the network).
In this case, the Laplacian eigenvectors are concentrated on
a subset of nodes that belongs to the considered weakly
connected sub-circuit.
II. P ROPOSED APPROACH
Let us consider passive resistive electric circuit with N
nodes and Nb branches. For branch connecting node n with
node m denote branch resistance with Rnm . The circuit can
be represented as a weighted graph form where edge weights
are conductances wnm = 1/Rnm .
The graph Laplacian can be obtained as L = D W
where W is edge weight matrix with elements wnm = 1/Rnm
if there is a branch between node n and node m and
wnm =P0 otherwise. Matrix D is a diagonal matrix with
N
dnn = m=1 wnm . Laplacian matrix eigenvalues are denoted
with k and corresponding eigenvectors with uk .
If the corresponding graph is connected, there is only one
zero eigenvalue 1 = 0 of the Laplacian matrix and all
other eigenvalues are positive. Here we will assume that
eigenvalues are sorted in non-decreasing order k k+1
for k = 1, 2, . . . , N 1. The eigenvector u1 corresponding to
the 1 = 0 is constant
1
u1 = [1, 1, . . . , 1]T .
N
If we arrange eigenvalues into diagonal matrix and
eigenvectors into square matrix U = [u1 , u2 , . . . , uN ], then
we can write
L = UUT (1)
Note that matrix U is unitary, UUT = E where E is unity
matrix with ones on the main diagonal and zeros elsewhere.
 6th Mediterranean Conference on Embedded Computing MECO2017, Bar, Montenegro

Assume that node 1 is reference node and that there exist for k = 1, 2, . . . , N .
external current generators connected between reference node
Relation (5) can be solved for Vk for each k except for
1 and nodes 2, 3, . . . , N . Now we can form external currents
k=1
vector i as 1
i = [i1 , i2 , . . . , iN ]T Vk = Ik
k
where i1 = i2 i3 . . . iN , according to the first For k = 1 we have 1 = 0 resulting in I1 = 0 and arbitrary
Kirchhoffs law. V1 . Note that I1 = uT1 i = 0 according to the first Kirchhoffs
law.
The potential of node n will be denoted with vn . The
potential vector v is Value of V1 can be defined if we state that the reference
node potential v1 is zero. Form (3) we can write
v = [v1 , v2 , . . . , vN ]T
v = UV
According to the first Kirchhoffs law for each node n sum
of all currents should be equal to the external current in . and
Current at branch that connects nodes n and m is equal to v1 = u1 (1)V1 + u2 (1)V2 + + uN (1)VN

inm = (vn vm )/Rnm = (vn vm )wnm . from this equation we can find V1 having in mind that v1 = 0
as
We can write first Kirchhoffs law for node n as u2 (1)V2 + u3 (1)V3 + + uN (1)VN
V1 =
N
X N
X u1 (1)
in = inm = (vn vm )wnm .
m=1 m=1 III. E XAMPLES
Note that wnm = 0 if there is no branch between node n and
We will illustrate proposed approach on two examples. A
node m. This relation can be rewritten as
simple circuit is analyzed in Example 1. More complex circuit,
N N
X X composed of two weakly connected subcircuits, is analyzed in
i n = vn wnm vm wnm . Example 2.
m=1 m=1

or in a matrix form as A. Example 1

i = Dv Wv = (D W)v
Let us consider circuit presented if Fig. 1. The correspond-
i = Lv. (2)
ing weighted graph is presented in Fig. 2.
This relation can be considered as Ohms law on a whole Graph Laplacian is
circuit, where conductances are included in the Laplacian
matrix L. In the classical circuit analysis system of equations 12 6 0 6 0 0 0
(2) is known as vertex potentials equations. 6 19 4 6 3 0 0

0 4 16 0 12 0 0
Vectors i and v can be considered as signals defined on 1 6 6

L= 0 24 6 6 0
a given graph. Spectral representation of these signals is 24000
obtained as their projection on the Laplacian eigenvectors 0 3 12 6 27 6 0
0 0 0 6 6 36 24
I = UT i 0 0 0 0 0 24 24
V = UT v. (3) Eigenvalues of the Laplacian are
1 = 0
According to (1), (2) and (3) we can write
2 = 0.251 103
i = UUT v
3 = 0.406 103
UT i = UT v
4 = 0.956 103
I = V (4)
5 = 1.091 103
The relation (4) can be considered as Ohms law in the spec- 6 = 1.542 103
tral domain. The eigenvalue matrix is diagonal, resulting in 7 = 2.337 103
Ik = k Vk (5) and corresponding eigenvectors are presented in Fig. 3
 6th Mediterranean Conference on Embedded Computing MECO2017, Bar, Montenegro

Vector of external currents is 1k

7 6
T
ig = 103 7 0 0

0 0 0 7
4 k 4 k
Vector of node potentials is
T
7 mA

v = 0 12 18 16 20 32 39 4 k
4 5
and in the spectral domain
 T 4 k 4 k 2 k
V = 51.78 30.28 7.35 3.41 1.29 0.74 1.79
8 k

B. Example 2
1 4 k 2 6 k 3
Consider now, more complex circuit shown on Fig. 4.
Fig. 1. Circuit for Example 1
Weighted circuit graph is presented in Fig. 5. The Laplacian
eigenvalues are:
1 = 0
2 = 0.001 103
3 = 0.251 103
4 = 0.398 103
5 = 0.406 103
6 = 0.956 103
7 = 0.984 103
8 = 1.091 103
9 = 1.341 103
Fig. 2. Weighted graph for circuit shown in Fig. 1
10 = 1.543 103
11 = 1.778 103
12 = 2.338 103
We can see that eigenvalue 2 is very close to zero.
The Laplacian eigenvectors are shown in Fig. 6.
Form Fig. 6 we can see that eigenvectors 4, 7, 9 and 11
have zero values at nodes 17 while eigenvectors 3, 5, 6, 8,
10 and 12 nave zero values for nodes 812.
Eigenvectors 1 and 2 span over all nodes. Corresponding
eigenvalues are almost same, 2 1 = 0, meaning that these
two eigenvectors belongs to the same sub-space determined
with zero eigenvalue. Instead of them we can use their linear
combinations
unew
1 = a1 u1 + a2 u2
unew
2 = b1 u1 + b2 u2
with
a21 + a22 = 1
b21 + b22 = 1
a1 b1 + a2 b2 = 0. Fig. 3. Laplacian eigenvectors for Example 1
 6th Mediterranean Conference on Embedded Computing MECO2017, Bar, Montenegro

7 1k 6 1M 8 2 k 9

4 k 4 k 4 k
2 k
4 4 k 5 1M
7 mA 10 4 k

4 k 4 k 2 k 4 k
6 k
8k

1 4 k 2 6 k 3 12 3k 11

Fig. 4. Circuit for Example 2

Fig. 5. Weighted graph for circuit shown in Fig. 4, edge weights are in mS

We can select parameters a1 , a2 , b1 and b2 such that we obtain

eigenvectors unew
1 and unew
2 spanned over nodes 17 and
812. Orthonormality of the q matrix U is preserved.q For the
5 7
considered case a1 = b2 12 and a2 = b1 12 . Fig. 6. Laplacian eigenvectors for Example 2

In this way, from spectral decomposition we can conclude

that the analyzed circuit can be split into two, almost indepen-
dent, parts.
IV. C ONCLUSION
We studied possible application of the graph spectral theory
to the resistive electrical network analysis and decomposition.
It is shown that graph Laplacian relate node potentials and
external node circuits for resistive network. Similar relation is
derived in the graph spectral domain. These two relations can
be considered as Ohms law for a whole graph in physical
and spectral domain.
By analyzing graph spectra we can check is the considered
graph (circuit) strongly connected or there exists sub-graphs
(sub-circuits) that are weakly connected with the rest of the
graph.
R EFERENCES
[1] C. A. Desoer, Basic circuit theory, Tata McGraw-Hill Education, 2009.
[2] D. M. Cvetkovic, M. Doob, H. Sachs, Spectra of GraphsTheory and
Application, Johann Ambrosius Barth Verlag, HeidelbergLeipzig, 1995.
[3] F. Dorfler, F. Bullo, Kron reduction of graphs with applications to
electrical networks, IEEE Transactions on Circuits and Systems I:
Regular Papers Vol.60, No.1, pp. 150163, 2013
[4] W. Ellens, F. M. Spieksma, P. Van Mieghem, A. Jamakovic, R. E. Kooij,
Effective graph resistance, Linear algebra and its applications, Vol.
435, No.10, pp. 24912506, 2011
[5] M. Kagan, On equivalent resistance of electrical circuits, American
Journal of Physics, Vol. 83, No. 1, pp. 5363, 2015
[6] D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, P. Vandergheynst,
The emerging field of signal processing on graphs: Extending high-
dimensional data analysis to networks and other irregular domains,
IEEE Signal Processing Magazine, Vol. 30, No. 3, pp. 8398, 2013
[7] Lj. Stankovic, M. Dakovic, E. Sejdic Vertex-Frequency Analysis: A
Way to Localize Graph Spectral Components, submitted to IEEE Signal
Processing Magazine, 2017