Network Models PDF
Network Models PDF
1, 237–269
This paper gives an overview of the formulation and solution of network equa-
tions, with emphasis on the historical development of this area. Networks are
mathematical models. The three ingredients of network descriptions are dis-
cussed. It is shown how the network equations of one-dimensional multi-port
networks can be formulated and solved symbolically. If necessary, the network
graph is modified so as to obtain an admittance representation for all kinds of
multi-ports. N -dimensional networks are defined as graphs with the algebraic
structure of N -dimensional vectors. In civil engineering, framed structures in two
and three spatial dimensions can be modeled as 3-dimensional or 6-dimensional
networks. The separation of geometry from topology is a characteristic feature
of such networks.
1. Introduction
Network modeling is applicable to any real-world system that fulfills the following
conditions. The signals occurring in the real-world system involve two types of vari-
ables:
(a) flow variables (FVs for short, also called through variables) obeying a cut law,
i.e., the flow quantities going through any closed cutting surface sum up to zero,
(b) difference variables (DVs for short, also called across variables) obeying a circuit
law, i.e., the difference quantities across adjacent points along any closed path
add up to zero.
Networks are interconnections of a finite number of network elements (modeled as
spatially lumped) which require interrelations between FVs and DVs defining the set
of network element relations (NERs).
For the special case of networks consisting of electrical wires, in 1845 Kirchhoff,
a 21-year-old student at that time, published what is now called the ‘node and mesh’
rules for electrical circuits (Kirchhoff, 1845). In the second half of the 19th century,
∗ Institut für Regelungs- und Steuerungstheorie, Technische Universität Dresden, Mommsen-
str. 13, D–01062 Dresden, Germany, e-mail: kr@erss11.et.tu-dresden.de
238 K.J. Reinschke
spring housing
coil housing
connections to
measurement coil
(transducer 2)
PSfrag replacements connection to
excitor coil
(transducer 1)
(a)
Transducer
Transducer
PSfrag replacements
1
PSfrag replacements
Transducer Transducer
(b)
2 1
PSfrag replacements
Transducer Transducer
(c)
1
9 6
16 20
3 19 7 4 17
8 14 3 5 0 8 1 15 5
6 11
2
10 12 21
1 18 2
7 9 13 4
2
(d)
Fig. 1. Electromechanical oscillating system: (a) cross-section of physical device,
(b) network model, (c) network graph consisting of five node-disjoint di-
graphs, (d) network graph modified to one connected digraph.
240 K.J. Reinschke
16 20
3 19 7 4 17
14 5 8 15
S1
6 11
10 12 21
1 18 2
9 13
ζ · · · 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
..
.
S1 0 1 0 1 0 0 -1 -1 0 0 1 1 1 0 1 0 1 0 0 0 0 = K̃
.
..
In this paper, the term circuit denotes an oriented closed path on the network
graph. Given any circuit, each branch is associated with an integer indicating how
many times the circuit passes through the branch. One additional passing in accor-
dance with the branch orientation increases the index by 1, one additional passing in
the opposite direction decreases the index by 1. Consequently, the given circuit can
be uniquely characterized by a circuit branch indicator vector. If we consider several
On network models and the symbolic solution of network equations 241
circuits, say t altogether, all the information can be gathered in a (t, z) indicator
matrix M̃ used by Weyl (1923). For the example system, one circuit C1 and the
corresponding row vector of M̃ are depicted in Fig. 3.
16 20 C1
3 19 7 4 17
5 8
14 15
6 11
10 12 21
1 18 2
9 13
ζ · · · 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
..
.
C1 0 0 0 -2 0 0 1 0 0 -1 -1 -1 3 0 0 0 2 0 0 0 0 = M̃
.
..
9 6
16 20
3 19 7 4 17
8 14 3 5 0 8 1 15 5
6 11
10 12 21
1 18 2
7 9 13 4
2
Fig. 4. Independent node cut sets and one possible set of tree branches.
incident with node κ. Any non-vanishing minor of order k of K, i.e., the determinant
242 K.J. Reinschke
of any (k, k) submatrix formed by the common entries of the row vectors 1, 2, . . . , k
and any pairwise different set of k column vectors ζ1 , ζ2 , . . . , ζk of K, corresponds
to a (spanning) tree of G. More specifically,
1 or −1 if the branches ζ1 , ζ2 , . . . , ζk form a tree,
K ζ11 ζ22···k
···ζk
= (2)
0 otherwise.
16 20
3 19 7 4 17
14 5 8 15
6 11
10 12 21
1 18 2
9 13
Fig. 5. A spanning tree in the network graph and the two basic meshes
that are defined by the links 7 and 14.
Rank M = z − k since each mesh contains one branch not contained in all the
other meshes. The orthogonality relations read
K M T = 0. (3)
Equation (3) is a special case of (1). Fortunately, the incidence matrices K and M are
sufficient to describe completely the cut and circuit laws. This is true since every cut-
set and every circuit-set, respectively, can be described as a linear combination of the
k independent branch sets incident with the nodes as well as the z − k independent
basic meshes associated with any chosen tree.
We remark that although the term “tree” was coined by Cayley (1857), Kirchhoff
had made use of the concept ten years earlier (Kirchhoff, 1847). Weyl was the first to
On network models and the symbolic solution of network equations 243
Figure 6 shows, in pictorial form, the branch representation in its most general form,
PSfrag replacements
including the sign conventions used in this paper. The branch is associated with an
independent FV source and an independent DV source. The three dots across which
ieζ
κ1 κ2 κ1 κ2
iζ ζ
uζ ueζ
(a) (b)
the DV uζ appears in Fig. 6 symbolize the NERs of branch ζ. The three dots may
stand for the symbol of a resistor (i.e., an ohmic resistor in electrical networks), a
FV storage (e.g., a capacitor) or a DV storage (e.g., an inductor). In all these cas-
es, the NERs represent a unique mapping between the FV iζ and the DV uζ . This
is typical of the so-called one-port branches. Figure 6, however, remains valid also
for multi-port networks, i.e., networks with interdependencies between FVs and DVs
of different branches caused by inductive couplings, controlled sources, or genuine
multiport elements such as transistors, transducers, etc. A general n-port network
element is depicted in Fig. 7. Examples of 2-ports are ideal transformers, gyrators,
ideal amplifiers, controlled DV sources, controlled FV sources, nullors (i.e., a nullator
branch combined with a norator branch). Figure 8(a) gives the general pictorial rep-
resentation of a two-port network element together with its NERs in implicit form.
Figure 8(b) explains the special case of a nullor.
The branch DVs, the branch FVs, the DVs across the independent DV sources,
the FVs through the independent FV sources, and the node DVs are considered to
be components of the vectors u, i, ue , ie , i0 and uφ , respectively.
The cut law for the FVs may be formulated as K (i + ie ) = 0, or, taking into
account (3) and expressed differently, i + ie ∈ image {M T }, i.e., M T i0 = i + ie ,
where the z − k componenets of i0 may be interpreted as mesh DVs. In the sequel,
we can make use of both the mathematically equivalent formulations of the cut law,
K(i + ie ) = 0 or M T i0 = i + ie . (4)
Analogously, there are two equivalent formulations of the circuit law:
M (u + ue ) = 0 or K T uφ = u + ue , (5)
244 K.J. Reinschke
port 1
port 2
PSfrag replacements
port n
(a) (b)
I1 I2
U1 U2 I1 U1 0
A +B =
I2 U2 0
(a)
I1 I2
1 0 I1 0 0 U1 0
U1 U2 ± =
0 0 I2 1 0 U2 0
(b)
Fig. 8. Two-port network elements: (a) general two-port and its NERs,
(b) nullor and its NERs.
(u + ue )T (i + ie ) = uφ KM T i0 = 0.
f (u, i) = 0. (6)
i0
i + ie
0
T K cut law
M e
−i
iφ
network
f (u, i) = 0
element relations
u0
−ue
circuit law
M KT uφ
0
u + ue
f (K T uφ − ue , i) = 0, Ki = −K ie = iφ (8)
with k + z unknowns,
246 K.J. Reinschke
f (u, M T i0 − ie ) = 0, M u = −M ue = u0 (9)
with 2z − k unknowns,
(iv) equations for node-DV mesh-FV analysis (NMA)
f (K T uφ − ue , M T i0 − ie ) = 0 (10)
with z unknowns.
The duality between (8) and (9) deserves to be noted. It appears that Maxwell was
the first to use the idea of duality in his study of frameworks (Maxwell, 1870). In this
paper, we mainly restrict our attention to NBA. In general, little can be said about
the unique solvability of the nonlinear equations (8). However, if the NERs are of the
form
f (u, i) = Y u − i = 0,
where uT Y u 6= 0 for all u 6= 0, the linear operator Y is termed “ohmic.” Then the
network problem has a unique solution (Roth, 1959).
Network models may be classified according to the algebraic structure of the
branch FVs and the branch DVs. If they are real or complex numbers, i.e., they
are elements of a one-dimensional vector space, then one refers to one-dimensional
networks. If the branch FVs and branch DVs are elements of an N -dimensional vector
space, then the networks are called N -dimensional networks.
3. One-Dimensional Networks
3.1. One-Port Networks
For one-port networks, the NERs are given by z FV-DV relations (for the z individual
branches)
fζ (uζ , iζ ) = 0 (ζ = 1, . . . , z). (11)
For linear time-invariant network elements, the NERs have the well-known time and
frequency domain equations:
In the frequency domain, the concept of branch admittances defined by yζζ (s) =
Uζ (s)/Iζ (s) has proved to be useful. Making use of the branch admittance matrix
Y (s) = hy11 (s), y22 (s), . . . , yzz (s)i ,
On network models and the symbolic solution of network equations 247
can be solved symbolically by inspecting the network graph (Kirchhoff, 1847; Maxwell,
1882). The Cauchy-Binet formula (Cauchy, 1815) gives a key to a thorough under-
standing of the facts announced in (Kirchhoff, 1847; Maxwell, 1882):
(kz )
X 1 2··· k
det(KY K T ) = det(K(Y K T )) = K i11 i2···
2 ···ik
k YK
T
i1 i1 ···ik
i=1
z
(kz )
X Xk
= K i1 i1 ···ik
1 2··· k Y ji 1 ji2 ···j1
T 1 2 ··· k
···i (K )j j ···j
k
2 k 1 2 k
i=1 j=1
(kz )
X
= K i11 i2···
2 ···ik
k Y ji11 ji22 ···j
···ik
k
K j11 j2···
1 ···jk
k , (15)
i,j
P
{i i · · · ik } = {j1 j2 · · · jk }. Then (Maxwell, 1882) det(KY K T ) = (i) Y i11 ii22 ···i
i
···ik =
k
P1 2
(i) (product of the admittances of all branches of tree i).
T
P Analogously, it can be derived that (Kirchhoff, 1847) det(M ZM ) =
(i) (product of the impedances of all branches of co-tree i).
Furthermore,
det(M ZM T )
= det Z = (det Y )−1 .
det(KY K T )
Formally replacing every branch admittance by the real number 1, the total number
of trees can be calculated as
nT = det(KK T ) = det(M M T ).
In the case of RLC-networks we can state the following: The network deter-
minants det(KY K T ) or det(M ZM T ) can be calculated symbolically by means of
248 K.J. Reinschke
enumeration of all the trees of the network graph. Every tree corresponds to one
term in a sum, cf. the right-hand side of (15). All these terms have the same sign,
and no cancellations of terms can thus occur. The tree enumeration method was
published in the thirties (Ting, 1935; Tsai, 1939; Wang, 1934), later on algebraically
substantiated (Bellert, 1962; Bott and Duffin, 1953; Duffin, 1959; Seshu and Reed,
1961; Trent, 1955), and has been implemented for CAD purposes since the sixties
(Chen, 1967; Chua and Lin, 1975; Dmitrischin, 1969; Mayeda and Seshu, 1965; Trochi-
menko, 1972). For related papers published in the seventies and eighties, see the mono-
graphs (Gieben and Sansen, 1991; Lin, 1991). The applicability of this approach has a
limitation: the number of trees may increase exponentially with the number of nodes.
Indeed, for a complete network graph (i.e., a graph in which every pair of nodes is
connected by exactly one branch) the number of trees equals k (k−2) . Fortunately,
complete graph structures are not typical of practical network models. But even for
ladder networks we observe a growth in the number of trees depending exponentially
on the number of ladder sections (see Lin, 1991, p.47). As for the actual usefulness of
this method, much depends on the skills of the investigator. Frequently, it does not
make sense to explicitly print out thousands of symbolic expressions corresponding
to thousands of trees. The trees need not actually be determined. All information
is contained in the main diagonal elements of (KY K T )κκ , where the admittances
of all theQ
branches connected to the node κ are summed-up symbolically. Then the
k
product κ=1 (KY K T )κκ , if evaluated according to the rules of the Wang algebra,
W W
( x + x = 0, x · x = 0 ), yields the desired network determinant in symbolic form. Of
course, a complete resolution of all brackets will often appear as an inefficient way of
evaluating the determinant.
As an example, consider the small electrical network depicted in Fig. 10 and find
a symbolic expression for the network determinant det(KY K T ). Obviously, k = 3.
1 3
As for the number of branches, let us discuss two possibilities. If each passive element
corresponds to one branch, then z = 9 and the number of trees is
4 −1 0
nT = det −1 3 −2 = 39.
0 −2 5
On network models and the symbolic solution of network equations 249
The nomenclature has been slightly changed from the one previously used so as to
illustrate the role played by the multi-port network elements. To avoid confusion, an
example with 8 one-port branches and three two-ports is depicted in Fig. 11. The
node-DV equations appear now in the augmented form
!
T Y Ue − Ie
(K KM )hY YM i(K KM ) Uφ = (K KM ) e e
.
YM U M − IM
The network determinant can again be evaluated by applying the Cauchy-Binet for-
mula twice,
T
det[(K KM ) hY YM i (K KM )T ] = det(K Y K )
X i1 i2 ···ik j1 j2 ···jk j1 j2 ···jk
= K 1 2 ···k Y i1 i2 ···ik K1 2 ···k .
(i,j)
j1 j2 ···jk
In contrast to (15), there now exist non-vanishing minors Y i1 i2 ···ik taken from the row
sets i1 , . . . , ik and the column sets j1 , . . . , jk with {i1 , . . . , ik } 6= {j1 , . . . , jk }. The
associated ‘i’-tree and ‘j’-tree may differ from each other. Nevertheless, it is possible
to formulate topological rules for a symbolic evaluation of the network determinant,
see, e.g., (Reinschke and Schwarz, 1976). Due to a lack of space, the details are
omitted here. In Section 3.3, related questions will be discussed in a more general
250 K.J. Reinschke
(a)
(b)
I1 y11 y12 U1
=
I2 y21 y22 U2
(c)
6 3 7 4 8
1 2 3 4 5 6
9 10 11 12 13 14
1 2
y1
y2
..
.
y8
(1) (1)
hY YM i = y11 y12
(1) (1)
y21 y22
(2) (2)
y11 y12
(2) (2)
y21 y22
(3) (3)
y y
11 12
(3) (3)
y21 y22
(d)
Fig. 11. Active RC circuit: (a) RC circuit with three transistors, (b) small
signal network model, (c) admittance representation of a transistor,
(d) network graph and branch admittance matrix.
On network models and the symbolic solution of network equations 251
κ1 IL κ1 IL κnew
UL UL Unew = IL
L
κ2 κ2
I1 I2 κnew −I2
κ1
− g1 U2 1
g
U1 Caux = L/g 2
U1 U2
κ2
Fig. 12. Two possibilities to get modified nodal equations whose coefficients
are affine functions of the complex frequency s.
The price to be paid is obvious: By including one further inductor in the network, the
number of graph nodes increases by one. The total number of network trees, however,
remains unchanged.
A few remarks about the symbolic solution of linear network equations are in
order. If the inner structure of the network equations is neglected, any method of
symbolic evaluation of determinants can be applied. In particular, graph theory pro-
vides useful tools to tackle this problem. There are several possibilities of constructing
digraphs that have a one-to-one correspondence with a given square matrix A and
obtaining the determinant det A by inspection of the digraph. (For example, see the
Appendix in (Reinschke, 1988).) The first graph-theoretic interpretation of determi-
nants was published by Cauchy (1815), reformulated by Jacobi (1841), and re-invented
252 K.J. Reinschke
by Coates (1959). Each term of det A corresponds to a spanning cycle family in the
digraph. This method is cancellation-free, i.e., if the matrix entries are mutually inde-
pendent, then no terms which cancel each other are generated. In König’s monograph
(1936) square matrices are represented by bipartite graphs. Each term of the determi-
nant corresponds to a matching, and this method is also cancellation-free. Seen from
the mathematical point of view, both the approaches are equivalent. As for computer
implementations, the matching algorithms have advantages. The symbolic solution of
linear algebraic equations can be traced back to the evaluation
of determinants as
T T −1 A a
follows: Let Ax = a, y = c x. Then y = c A a = det −cT 0 /det A = D/N and
A a
det −c T
p = p N + D. The augmented determinant is a linear polynomial in the
parameter p. The coefficients are the denominator D and the numerator N . The so-
called signal-flow graphs published by Mason became popular in the sixties (Mason,
1953). These graphs have the disadvantage that the algebraic equations must be of
the form x = Ax + b.
k
X
K Y KT = yij K•i (K•j )T ,
i,j=1
where K•j denotes the j-th column of the node branch incidence matrix K. The
dyadic product K•i (K•j )T gives a structural “stamp” defined by the node branch
incidence relations of the branches j and i. Stated in another way, the network co-
efficient matrix is a weighted sum of stamps of the same size. Each term of the sum
reflects the influence of one network parameter appearing as an entry of the branch
admittance matrix Y of the network. Graph-based methods for the symbolic solution
of the network equations are discussed in the next section.
A IM + B U M = 0 (16)
On network models and the symbolic solution of network equations 253
with (n, n) matrices A and B. For any linear time-invariant network, a complete
set of network equations can be written as
E −Y 0 0 0 I 0
0 E 0 0 −K T U −U e
T e
UM = −UM ,
0 0 E 0 −KM
T e
0 0 0 A BKM IM BUM
0 0 0 KM KY K T Uφ −Iφe
1
L8
Ue
3
2
R5
R3 R4 C6 C7
1 2
(a)
8
1
5
3 2 4 5
3
6
1 4 7
2
(b)
new nodes
Given
Augmented network graph Uadd = IM
network graph:
Iadd
E 0 0 0
cut law: IM =
0 KM K −KI e −KM IM
e
I
Uadd E 0 0
T Uadd e
circuit law: UM = 0 K M + −UM
T Uφ
U 0 K −U e
Iadd A B 0 Uadd 0
NERs: IM − E 0 0 UM = 0
I 0 0 Y U 0
| {z }
=:Y
T
A BKM T
E 0 0 A BKM
= E 0 =
0 KM K KM KY K T
0 Y KT
and z be the total number of multi-port and one-port branches, respectively. Then
!
T
A BKM T
det = det K Y K
KM KY K T
z+zM
k+z
XM i1 ···izM +k j1 ···jzM ···jzM +k j1 ···jzM +k
= K 1···zM +k Y i1 ···iz ···izM +k K 1···zM +k
M
i,j=1
The last equality results from the particular structure of the augmented incidence
matrix K. The first zM columns of K must be used to get a regular (zM + k)-minor
On network models and the symbolic solution of network equations 255
of K that corresponds to a tree in the augmented network graph. The (zM + b)-
minors of the square matrix ( E A B ), where 0 < b < z , are most important for the
0 M
symbolic evaluation of the network determinant. The minor
!1···zM j1 ···jb
A B
=: (YM T )ji11···i
···jb
E 0 b
1···zM i1 ···ib
is associated with that part of the network graph which consists of the multi-port
branches i1 , . . . , ib within the ‘i’-tree and the multi-port branches j1 , . . . , jb within
the ‘j’-tree. The small example introduced in Fig. 13 may help us to explain the basic
ideas for the symbolic evaluation of the network determinant (see Fig. 15). The node
¯¯¯ ¯ ¯
®æ1 0 0 0 öæ1 ö¬
ç ÷ç ÷
®ç0 0 1 0 ÷ç 1 ÷¬
ç ÷ç ÷
æ1 ö ç1 0 ÷ç 1 0 0÷¬
ç ÷ ç ÷ç
ç 1 ÷ ®ç0 ÷
1 ÷ç 0 0 1÷
ç ÷
ç 1 0 0 0 0 0 -1 1÷ ç G4 ÷ç
0 0 1÷
ç ÷ç ÷
ç ÷
ç 0 0 0 1 0 1 1 0÷ ® ç G5 ÷ç 0 1 -1÷ ¬ j -tree
ç ÷ ç ÷ç ÷
è 0 1 1 -1 1 0 0 0ø ç sC6 ÷ç 0 0 1÷
ç ÷ç ÷
ç sC7 ÷ç 0 1 0÷
ç 1 ÷ç ÷
i -tree ®ç ÷ç -1 1 0 ÷ ¬
ç sL8 ÷ç ÷
è G3 ø è 1 0 0ø
branch incidence matrix of the augmented graph has 3+2 = 5 rows and 2+2+5 = 9
columns. The non-vanishing terms of the network determinant are 5-minors of the
augmented branch admittance matrix, i.e., determinants of (5, 5)-matrices whose rows
are defined by an i-tree whereas the columns are defined by a j-tree. It is evident that
each j-tree and each i-tree contain the two newly added branches. Consequently, the
(2, 2)-submatrix in the left upper corner must be contained in any (5, 5)-submatrix.
Continuing this discussion, we could conclude: There is exactly one non-vanishing
5-minor, and—apart from the sign—the network determinant is equal to the product
of admittances of branch 5 and of branch 8. On the basis of these preliminary remarks,
we are able to formulate a general evaluation rule:
zM
zM (X
X b )
j ···j
NW-Det = (YM T )i11···ibb · Remainder-NW-Det.
b=0 i,j=1
Comments on the individual terms in the double sum (see Fig. 16):
1. Let b = 0 : (YM T ) = det A. Removal of the zM multi-port branches creates the
remainder-NW-graph. Remainder-NW-Det = det(KY K T ).
256 K.J. Reinschke
zM b k−b
z }| { z }| { z }| {
↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ j-tree
→
→
zM → A B
→
1 ↓
→ 1 →
b → +
1
1 ◦
• =Y
•
•
→
→
•
k−b Y
•
→ •
•
→ •
i-tree
Fig. 16. Evaluation of the network determinant. Notation:
‘◦’ a multi-port branch to be removed,
‘+’ a multi-port branch to be contracted,
‘↓, →’ a pair of multi-port branches to be completed to branch-disjoint meshes.
(a) Both the terminal nodes of each multi-port branch within the branch set
{i1 , . . . , ib } ∩ {j1 , . . . , jb } coalesce to one node. The multi-port branches
not contained in the branch set {i1 , . . . , ib } ∪ {j1 , . . . , jb } are removed.
(b) If there remain d(< b) branches in {i1 , . . . , ib } and d branches in
{j1 , . . . , jb }—neither short-circuited in step (a) nor open-circuited in step
(a)—then the d branches of the j-tree correspond to d columns of B
whereas the d branches of the i-tree correspond to d differently indexed
rows of E. Within the network graph, the d j-tree branches and the d
i-tree branches may be pairwise completed (by means of topologically suit-
able existing branches) to d branch-disjoint meshes. Assume that there are
m different possibilities to construct d disjoint meshes. Then we denote by
yM, µ the µ-th mesh admittance defined as the product of the admittances
of all 1-port branches occurring within the d meshes of the µ-th mesh set
(µ = 1, . . . , m). We set yM,µ = 1 if the d meshes of the µ-th mesh set
do not contain a 1-port branch. Afterwards, the µ-remainder-NW-graph
On network models and the symbolic solution of network equations 257
!
a11 b12
(d) − det · Remainder-NW-Det,
a21 b22
provided the remainder network resulting from short-circuiting of the second
two-port branch and removal of the first two-port branch is connected;
" ! !#
a12 b12 a11 b11
(e) det − det
a22 b22 a21 b21
hP i
m
× µ=1 (µ-th mesh admittance) · (µ-remainder-NW-Det) ,
provided there exist m meshes containing both two-port branches such that the
µ-th remainder network resulting from coalescing the µ-th mesh to one node is
connected.
A two-port of special interest is the “nullor” defined in Fig. 2. In this case, the
terms mentioned above (a)–(d) vanish. In the term (e) we have
! !
a12 b12 a11 b11
det − det = 1.
a22 b22 a21 b21
This implies that the symbolic evaluation of network determinants for networks with
nullors should start with a search for nullor meshes, i.e., meshes containing one nul-
lator branch and one norator branch. Two conclusions can be drawn from this:
(i) A simple criterion for the solvability of the network problem: The determinant
of a network whose multi-port elements are modeled by means of n nullors may
be non-zero only if the n nullator branches and the n norator branches can be
completed to n branch-disjoint nullor meshes.
(ii) An efficient way to symbolically evaluate the numerators of transfer functions:
Augment a given network graph with an additional nullor (whose nullator branch
and norator branch, respectively, connect the nodes κ1 and κ2 with the reference
node), and then the network determinant of the augmented network is just the
(κ2 , κ1 )-cofactor of the determinant of the original network. The proof is sketched
in Fig. 17.
For illustrative purposes, an example system is depicted in Fig.18. The aim is to
symbolically determine the transfer function Uφ5 (s)/U e (s). The denominator is equal
to the network determinant. The network graph (see Fig. 18(b)) contains one nullor
mesh. The remainder network coalesces into the reference node, i.e., the remainder
network determinant equals 1. Hence, network determinant = nullor mesh admittance
= G1 G2 G3 G4 . The numerator is given by the (1,5)-cofactor. The associated network
graph results from augmenting the given network graph by a nullator branch and
a norator branch as depicted in Fig. 18(c). The augmented network contains two
branch-disjoint nullor meshes, each with a mesh admittance equal to 1. The remainder
network graph generated through coalescing both nullor meshes contains 21 trees.
Thus the numerator consists of 21 terms to be summed. The term with the highest
degree in s results from the branch set {8, 7, 6} and is equal to s3 C8 C7 C6 .
On network models and the symbolic solution of network equations 259
κ1 κ2
PSfrag replacements
T
B × KM
1 0 æç 0 0 ö÷ × K T
A 0 0 è 1 0ø M 0 1
1
T
PSfrag replacements det KM KYK = det KYKT = det KYKT
1 1
κ1
κ2
0
Fig. 17. The determinant of a network augmented by an added nullor as indicated
is equal to the (κ1 , κ2 )-cofactor of the original network determinant.
R5 1
PSfrag replacements 5 R4 4 R3 3 R2 2 R1
e
κ1 U
C8 C7 C6
κ2
0
(a)
1
PSfrag replacements 1 5
4
4
3
3
2
2
κ1
8 7 6
κ2 5
0
(b)
1
κ1 κ1 8
8 7 6 4 7 6 1
κ2 5 κ2
0 0
(c)
i.e., some components of the FV-vector i are explicitly known as functions of the DV-
vector u and its time derivative u̇, whereas the other components symbolized by i b
are implicitly represented. On the basis of augmenting the network graph as depicted
in Fig. 14, we are able to represent the entire FV-vector i = (iTb , iTa )T in the explicit
form
iadd f2 (u, u̇, ib , i̇b )
ib = uadd . (17)
ia f1 (u, u̇)
If the aim is to carry out an MNA, the cut and circuit laws should be respectively
formulated as
! iadd !
E 0 0 0
ib + ieb = (18)
0 K b Ka e
0
ia + i a
and
uadd E 0 !
u add
ub + ueb = 0 KbT . (19)
e T uφ
ua + u a 0 Ka
Combining (17)–(19), the network equations for the MNA are obtained as
Next, let us consider a class of nonlinear networks which are of particular impor-
tance for computer aided design of large-scale electronic networks (see, e.g., (Günther
and Feldmann, 1999) and many references cited therein). During simulation, the con-
servation of electric charges and magnetic fluxes should be ensured. Stated in general
terms, integrated FVs, in the sequel denoted by q, and integrated DVs, in the se-
quel denoted by ψ, must be taken into account. For this purpose, the linear NERs
uL = L i̇L and ic = C u̇c are replaced by the NERs of the form
d d
uL = ψ and ic = q,
dt dt
where ψ and q may be regarded as given nonlinear, possibly time-varying vector-
valued functions
and
! ! !
uadd E 0 uadd
= ,
u + ue 0 K uφ
ψ̇ − KLT uφ + ueL = 0,
Kc q̇ + Kb ib + Ka f1 (K T uφ − ue ) = −Kie ,
f2 (ib , K T uφ − ue ) = 0,
q = g1 (uc , t),
ψ = g2 (iL , t).
4. Multi-Dimensional Networks
By an N -dimensional network we understand a network model for which the FVs
and DVs associated with the individual nodes and branches of the network graph are
N -dimensional vectors.
As a first example, consider the Newtonian n-body problem (Reibiger and Elst,
1983). The question to be answered is how n spatially lumped bodies with mass-
es M1 , M2 , . . . , Mn , driven by their mutual gravitational forces, move in the three-
dimensional Euclidean space. The equations of motion are known from elementary
mechanics:
n
X Mj Mk
Mj üφj = γ (uφk − uφj ) ,
k=1
k uφk − uφj k3
k6=j
262 K.J. Reinschke
where γ represents the gravitational constant and uφj denotes the position of body j
in a global Cartesian coordinate system. The Newtonian n-body problem can be
formulated as a network problem as follows: The network graph consists of n + 1
nodes, where the position of body j is uφj and the origin of the coordinate system
serves as the reference position uφ0 = 0, and of n(n + 1)/2 branches connecting each
node with all the others. The branch direction can be chosen such that the orientation
arrow points to the higher indexed node. The case of n = 4 is depicted in Fig. 19.
M2
M1
uφ2
uφ1
M3
x3
M4
uφ3
uφ4
x2
x1
(a)
1 2
4
3
0
(b)
iζ = Mκ üζ .
The NERs of the remaining n(n − 1)/2 branches ζ leading from κ1 to κ2 , where
1 ≤ κ1 < κ2 ≤ n, are given by
M κ1 M κ2
iζ = γ uζ .
k uζ k3
On network models and the symbolic solution of network equations 263
qe Pe
joint 1
1 2 3 4
1 3 4
2 5
5 6
(b)
2
2
x
N
2
M
2
Q
z2
αζ
ζ
x
ϕ
2
ϕ
zζ
τ1
τ2
lζ
1
ϕ
1
M
z1
x
1
N
1
Q
about the y-axis perpendicular to the plane. The six end loads can be arranged as a
T
column vector iφζ = N1 Q1 M1 N2 Q2 M2 . Since the member ζ is thought of as
ζ
being released from the framework, its end loads must obey a cut law appearing as
three scalar equations of equilibrium. Therefore, only three of the six end loads are
T
independent. Choosing iζ = N2 M1 M2 as the three independent end loads, the
ζ
cut law for the released member reads
−1 0 0
0 −1/l −1/l
0 1 0
iφζ =
iζ =: cζ iζ . (21)
1 0 0
0 1/l 1/l
0 0 1 ζ
correspond component-wise to the FVs iφζ . The member movement can be split into
two parts: the rigid body movement which is not comprised in the load-deformation
relations, and deformations which cause the member to alter its shape. Again, several
choices of the three independent scalar relative movements can be realized in the local
T
member coordinate representation. We choose uζ = ∆x τ1 τ2 in accordance
ζ
with the chosen iζ . Then the kinematic compatibility conditions (circuit law for the
On network models and the symbolic solution of network equations 265
member ζ) require
−1 0 0 1 0 0
uζ = 0 −1/l 1 0 1/l 0 uφζ = cTζ uφζ .
(22)
0 −1/l 0 0 1/l 1 ζ
The local member stiffness matrix yζ relates the member deformations uζ to the
member loads iζ ,
i ζ = y ζ uζ . (23)
Neglecting shear deformations, Hooke’s law yields for prismatic members the local
member stiffness matrix as
EA 0 0
1
yζ = 0 4E I 2E I ,
lζ
0 2E I 4E I ζ
where E, A and I denote Young’s modulus of the material, the member cross-
sectional area and the member cross-sectional moment with respect to the y-axis,
respectively.
Together, eqns. (21)–(23) provide a complete representation of the NERs in local
coordinates:
iφζ = cζ yζ cTζ uφζ =: yφζ uφζ . (24)
The symmetric (6, 6) admittance matrix yφζ is called the complete member stiffness
matrix in civil engineering. Note that yφζ is singular since rank yφζ = 3.
The enforced member displacements, distributed member loads, and the effects
of heating act as independent member DV and/or FV sources. They can be replaced
by an equivalent set of concentrated loads acting on the joints at the two ends of the
member (see modern texts on structural analysis, e.g., (Krätzig, 1998)). The resulting
FV source vector ieφζ acts “parallel” to iφζ (ζ = 1, 2, . . . , z).
Up to now, each member stiffness matrix has been expressed in terms of the local
member coordinates. Before combining the stiffness matrices to an overall stiffness
matrix of the entire framework, it is necessary to describe all the individual member
end loads and displacements, using one and the same coordinate system. As we are
considering only plane frames here, the local member coordinates simply arise from
the global member coordinates by rotation on the common y-axis (with angle α).
Using the abbreviations c := cos α and s := sin α, the member end loads defined by
iφζ in local coordinates have the global coordinates
c s 0
−s c 0 0
0 0 1
g iφζ =
iφζ =: gζ iφζ . (25)
c s 0
0 −s c 0
0 0 1 ζ
266 K.J. Reinschke
is called the global member stiffness matrix. Now we are able to derive the overall
nodal DV equations for the entire framework by means of a suitably defined frame-
node member-node incidence matrix K. To illustrate this, the (6, 2 × 5) incidence
matrix K for the example framework (see Fig. 20(a)) is written out here,
frame member
11 12 21 22 31 32 41 42 51 52
nodes nodes
1 E
2 E E E
3 E E E
= K. (28)
4 E
5 E
6 E
The symbols E represent (3, 3) unit matrices. Let us compare the displacements Uφκ
of the frame nodes (measured in global coordinates) with the member end displace-
ments g uφζ . The circuit law (kinematic compatibility conditions between the frame
and members) requires
g uφ = K T Uφ . (29)
The cut law (static equilibrium conditions for each node) becomes
K g iφ = −Iφe , (30)
where Iφe denotes the vector of independent joint loads (including support reactions)
measured in the global coordinates. The combination of (26), (29) and (30) yields the
desired node DV equations,
For the example system (Fig. 20), member 2 whose NERs are given by the (6, 6)
member stiffness matrix
!
y211 y212
g y2 =:
y221 y222
generates the “stamp”
0 0 0 0 0 0
0 y222 0 y221
0 0
T
0 0 0 0 0 0
K h0 g y2 0 0 0 0i K =
0 0
.
0 0 0 0
0 y212 0 0 y211 0
0 0 0 0 0 0
The constituents of K g ieφ can also be generated member-wise and then summed up.
Returning to the example system, let us consider the effect of heating member 5.
T
Heating generates axial forces ieφ5 = EA γT ∆T e (1 0 0 − 1 0 0) , whence
!
e
e e e T g iφ5,1
g iφ5 = g5 iφ5 = EA γT ∆T (c, −s, 0, −c, s, 0)5 =: e
.
g iφ5,2
The part of K g ieφ that results from heating member 5 has the form
T
(0, 0, g iφ5,1 , 0, 0, g iφ5,2 ) .
In the foregoing discussion, we have modeled plane frameworks as three-dimensional
networks. In the case of spatial frameworks we would obtain six-dimensional networks,
and the entries of the incidence matrix K would be (6, 6) identity matrices or (6, 6)
zero matrices. Similarly, modeling plane and spatial trusses would result in two- and
three-dimensional networks, respectively.
5. Conclusion
This paper constitutes an attempt to overview the formulation and solution of network
equations. In this presentation, the historic development of the concept of network
modeling has been emphasized. Beyond the usual electrical network applications,
network models can describe a wide variety of other real-world systems. Topological
properties of the underlying network graphs provide the key to a thorough understand-
ing of one-dimensional networks. It was shown that the network element relations can
always be explicitly formulated by means of an augmented network graph. General
topological rules for symbolic solution of multi-port network equations were derived.
In the case of multidimensional networks, the analyst has to cope with the geome-
try of the systems. This has been exemplified for plane framed structures commonly
occurring in civil engineering.
268 K.J. Reinschke
References
Bellert S. (1962): Topological analysis and synthesis of linear systems. — J. Franklin Inst.,
Vol.274, No.6, pp.425–443.
Branin F.H.jr. (1966): The algebraic-topological basis for network analysis and the vector
calculus. — Proc. Symp. Generalized Networks, Brooklyn, pp.453–491.
Bott R. and Duffin R.J. (1953): On the algebra of networks. — Trans. AMS, Vol.74, pp.99–
109.
Cauchy A.L. (1815): Mémoire sur les fonctions qui ne peuvent obtenir que deux valeurs. —
J. l’Ecole Polytechnique, Vol.10, pp.29–97, reprinted in Oeuvres Complètes, Serie II,
Tome I, pp.91–169.
Cayley A. (1857): On the theory of the analytical forms called trees. — Philos. Mag., Vol.4,
No.13, pp.172–176.
Chen W.K. (1967): Topological network analysis by algebraic methods. — Proc. IEE, Vol.114,
No.1, pp.86–88.
Chua L.O. and Lin P.-M. (1975): Computer Aided Analysis of Electronic Circuits—
Algorithms and Computational Techniques. — New Jersey: Prentice-Hall.
Coates C.L. (1959): Flow-graph solutions of algebraic equations. — IEEE Trans. Circuit
Theory, Vol.6, No.2, pp.170–187.
Dmitrischin R.W. (1969): Rasčet linejnych cepej s metodom rasširennych strukturnych čisel.
— Isv. Vuzov SSSR, Radioelektreonnika, Vol.12, No.8, pp.806–813 (in Russian).
Duffin R.J. (1959): An analysis of the Wang algebra of networks. — Trans. AMS, Vol.93,
pp.114–131.
Firestone B.F. (1933): A new analogy between mechanical and electrical systems. — J. Acoust.
Soc. Am., Vol.4, pp.249–267.
Gieben G. and Sansen W. (1991): Symbolic Analysis for Automated Design of Analog Inte-
grated Circuits. — Boston: Kluwer.
Günther M. and Feldmann U. (1999): CAD-based electric-circuit modeling in industry, Part I
and II. — Surv. Math. Ind., Vol.8, pp.97–129 and 131–157.
Hähnle W. (1932): Die Darstellung elektromechanischer Gebilde durch rein elektrische Schalt-
bilder. — Wiss. Veröff. a. d. SIEMENS-Konzern, Vol.11, No.1, pp.1–23.
Jacobi C.G.J. (1841): De formatione et proprietatibus determinantium. — Crelle’s J. f. reine
und angew. Mathem., Vol.22, pp.285–318.
Kirchhoff G. (1845): Über den Durchgang eines elektrischen Stromes durch eine Ebene, ins-
besondere durch eine kreisförmige. — Poggendorfs Annalen der Physik und Chemie,
Vol.64, pp.497–514.
Kirchhoff G. (1847): Über die Auflösung der Gleichungen, auf welche man bei der Unter-
suchung der linearen Verteilung galvanischer Ströme geführt wird. — Poggendorfs An-
nalen der Physik und Chemie, Vol.72, pp.497–508.
König D. (1936): Theorie der Endlichen und Unendlichen Graphen. — Leipzig: Akademische
Verlagsgesellschaft.
Krätzig W.B. (1998): Tragwerke 2. 3. Aufl. — Berlin: Springer,
Kron G. (1939): Tensor Analysis of Networks. — New York: Wiley.
On network models and the symbolic solution of network equations 269