014 - Network Synthesis, Vol - XIV (1957) (Prof. E. A. Guillemin, Prof. P. M. Lewis II, Dr. M. V. Cerrillo)
014 - Network Synthesis, Vol - XIV (1957) (Prof. E. A. Guillemin, Prof. P. M. Lewis II, Dr. M. V. Cerrillo)
NETWORK SYNTHESIS
Procedures for the analysis of linear networks are ordinarily confined to a consider-
ation of configurations in which the branches or elements are two-terminal devices,
such as the familiar resistances, inductances, and capacitances that characterize
passive circuits. When active elements,
such as vacuum tubes or transistors, are
5 - TERMINAL
ELEMENT included, this simple topological picture
no longer applies, since these elements
<_
jare characterized schematically as
rboxes from which three or more termi-
Snals emanate. A network containing
such elements, as well as the ordinary
two-terminal elements, is shown in
Fig. XIV-1. We shall show that only a
Fig. XIV-1. Network containing multi- minor addition to the existing analysis
terminal elements, as well
procedure is needed to accommodate
as
as the usual two-terminal
the usual two-terminal
elements. this more general type of network.
It is well known that a box with p + 1
terminals is uniquely characterized by p voltages and p currents. The volt-ampere
relations for the multiterminal device may be written
(1)
v = r + ... + r
p plul + ppjp
or in inverse form,
(2)
p gplv 1 + ... pp p
in which jl...jp are currents that enter terminals 1, 2,... p, and return through the
common terminal 0 (see Fig. XIV-2a), and vl... vp are voltage drops measured from
103
(XIV. NETWORK SYNTHESIS)
104
(XIV. NETWORK SYNTHESIS)
105
(XIV. NETWORK SYNTHESIS)
The complete set of relations expressing loop and source currents in terms of
branch currents reads
j l
all + alj +... + abb
2 i
=
at,+, 1 Jl + a+l, , 2 + ".' + a+ 1, b b is 1
11
j = . and i=
isl
_Jb
sn
aj = i
-1.
a i= j (6)
The first I equations in the set 3 define the loop currents; the remaining n equa-
tions are Kirchhoff current-law equations. The quantities i sl'..sn appearing on the
right-hand sides of these equations are current sources that feed the pertinent
node pairs. The last n rows of the matrix a yield the cut-set schedule, as usually
defined.
If a tree is chosen and the link currents, numbered from I to 1, are identified
with loop currents, then the a matrix assumes the special form
106
(XIV. NETWORK SYNTHESIS)
0 0 ... 1 0 ...... O
a1 aLl,
C 1 aI l, 2 " a1+l, 1 0 " (7)
ab2 . .. ab 0 0 ... 1
ab
u1 0
ca= (8)
in which the submatrices u and un are unit matrices of order I and n, respectively,
the zero in the upper right-hand corner is an I X n null matrix, and ani is a submatrix
of n rows and I columns. Its elements are ±1 or zero, according to whether the branch
in question is or is not contained in the pertinent cut set. Since the tree branches,
numbered I + 1 to b, are contained singly in the respective cut sets, the last n columns
and the last n rows of a form a unit matrix of order n (the submatrix un in Eq. 8).
Analogously, on a voltage basis, we define node-pair voltage variables as linear
combinations of the branch voltages. Algebraically, we obtain n equations which may
be arbitrarily written, subject only to the condition that they be independent of each
other and of the equations in a set of I Kirchhoff voltage-law relations that express
source voltages as linear combinations of branch voltages. The complete set of b
equations thus obtained may be inverted to yield the branch voltages in terms of the
node-pair and source voltages.
The situation, again, is clearly unambiguous and unique, although no a priori dis-
tinction need be made between source voltages and node-pair voltages, which are alike
in kind and differ only in that the former are presumably known and the latter unknown.
Again, the division of the integer b into its additive components I and n may be
revised according to what we wish to regard as known and unknown. The same kinship
exists between the source voltages that act around loops and the node-pair voltages
that exist, on a current basis, between the loop and the source currents.
In writing the complete set of b equations in this case, we write the Kirchhoff equa-
tions first and the defining equations for the node-pair voltages last. Thus we have
107
(XIV. NETWORK SYNTHESIS)
S1pV
1 +1 2v 2 + ... bb = esl
+ + + =
+1, 1 +1, 2 V2 -* P+1, b Vb el
esl
v1
v = and e = (10)
Lvb
3v = e (11)
-le = v
3 e= v (12)
The quantities es1. . . esk are voltage sources acting upon the contours of loops for
which the Kirchhoff voltage-law equations are written. The first I rows of the matrix
3 yield the tie-set schedule, as usually defined.
If a tree is chosen and the tree-branch voltages, numbered from I + 1 to b, are
identified with node-pair voltages, then the P matrix assumes the special form
0 0 ... O 1, (13)
1, I+2
0 1 0
0 0 1 ... PO
ii o0 0 0 . 1
108
(XIV. NETWORK SYNTHESIS)
= | (14)
n0
n
Its interpretation is analogous to that discussed for the abbreviated matrix 8. Elements
in the submatrix BIn are ± 1 or zero, according to whether the branch in question is or
is not part of the pertinent tie set. Since the links, numbered 1 to 1, are contained
singly in the respective tie sets, the first I columns and the first I rows of P form a
unit matrix of order I (the submatrix u in Eq. 14).
We recall that the columns of a cut-set schedule yield the coefficients in a set of
equations that expresses the branch voltages in terms of the node-pair voltages; and
that the columns of a tie-set schedule yield the coefficients in a set of equations that
expresses the branch currents in terms of the loop currents. The latter may be
regarded as expressed by Eq. 6 and the former by Eq. 12, since the matrices a and
3 may be interpreted as representing cut-set and tie-set schedules in which the normal
loop currents are augmented by including the source currents, and the normal node-
pair voltages are augmented by including the source voltages. If the closed paths for
which Kirchhoff voltage-law equations are written are the same as those used to define
loop currents, and if the cut sets for which the Kirchhoff current-law equations are
written pertain to the same node pairs that are used to define node-pair voltages, then
we observe that the columns in the matrix P yield coefficients in the relations expressed
by.Eq. 6, while columns in the matrix a yield coefficients in the relations expressed by
Eq. 12. Under these conditions, which we refer to as "consistency conditions" because
the Kirchhoff equations are consistent with the defining equations for the variables, it
follows that P is the inverse transpose (or the reciprocal) of a and vice versa. That is,
at-1 = a
= -1 *
and. = 3 = (15)
It follows from Eqs. 8 and 14 that in this case Pp n is the negative of the transpose of
which may be readily verified, since the matrix 8 multiplied by the transpose of matrix
14 (or vice versa) must yield a unit matrix of order b.
Matrix Eqs. 5 and 11 (or their inverses, Eqs. 6 and 12) contain the information about
Kirchhoff relations, as well as the definition of voltage or current variables. All that
need be added in order to construct the equilibrium equations, on either a voltage or
current basis, is the set of volt-ampere relations linking the branch currents with the
109
(XIV. NETWORK SYNTHESIS)
in which the operator matrix D contains the branch inductance, resistance, and
elastance parameter matrices. If an active and/or nonbilateral multiterminal device
is involved, the branch resistance matrix (and in D-1 the branch conductance matrix)
is no longer diagonal but contains some nondiagonal elements. However, the formation
of D or of D-l follows precisely the same pattern as for the ordinary linear passive
two-terminal elements and needs no further elaboration.
If we want the equilibrium equations on a loop basis, we begin with Eq. 11 whose
first I rows are the Kirchhoff voltage-law equations. Since we want these expressed
in terms of the loop currents, we express v in Eq. 11 in terms of j by means of Eq. 17,
and subsequently express j in terms of i by means of Eq. 6. This double substitution
yields
-l
XDX a Xi= e (18)
If we want the equilibrium equations on a node basis, we begin with Eq. 5 whose
last n rows are the Kirchhoff current-law equations. Since we want these expressed
in terms of the node-pair voltages, we express j in Eq. 5 in terms of v by means of
Eq. 17, and subsequently express v in terms of e by means of Eq. 12. This double
substitution yields
aX D - 1 X -1 e = i (19)
which we recognize as the inverse of Eq. 18. In this approach the equilibrium equa-
tions on the loop and node bases are contained in matrix equations that are mutually
inverse.
Equations 18 and 19 contain, besides the desired equilibrium equations, some rather
interesting by-products which we shall place in evidence by partitioning the rows and
columns of the resultant matrices into groups of I and n. We shall assume henceforth
that consistency conditions 15 are fulfilled, so that the resultant matrices in Eqs. 18
and 19 become symmetrical. We then have for the loop basis,
XD X ptx i = e (20)
-1
aXD x at X e = i (21)
p x DX Pt = Z (22)
110
(XIV. NETWORK SYNTHESIS)
and
a XD X t = Y = Z (23)
Z = I(24)
Zni Znn]
Y = n (25)
Y ni Y
nnj
e= .. and i = K] (26)
in which e and i represent source quantities, and e and i represent variables, then
Eqs. 20 and 21 become
Z i + Zi = e
f v fn s s
(27)
Znni i v +Z nn i s = e
v
and
Y e +Y e = i
Yi es + Yln v vSs n v (28)
Y e +Y e = i
ni s nn v s
Equilibrium, on the loop basis, is expressed by the first of the matrix Eqs. 27 in
which the second term effects the conversion of current sources into equivalent voltage
sources. The second of Eqs. 28, similarly, expresses equilibrium on the node basis,
and the conversion of voltage sources to equivalent current sources is given by the
first term. The second equation in Eqs. 27 or the first in Eqs. 28 relates the node-
pair voltages and the loop currents. These relations may be useful, for example, if
for some reason it is desirable to express equilibrium in terms of some loop currents
and of some node-pair voltages, on a sort of mixed basis. Since the relations between
loop currents and node-pair voltages are available, we can trade variables of one sort
for variables of another.
Now if an active and/or nonbilateral element is embedded in an otherwise passive
bilateral network, then the branch resistance or conductance matrix has embedded in
111
(XIV. NETWORK SYNTHESIS)
it a submatrix like that pertaining to either Eq. I or Eq. 2, which does not
define a positive quadratic form and may be dissymmetrical. The operator matrix
-1
D in Eq. 20 or D-1 in Eq. 21 exhibits corresponding deviations from the form
for passive bilateral networks. (We shall refer to the corresponding network as
being non-P and/or non-B, as distinguished from the designation PB in the passive
bilateral case.)
By well-known methods, it is always possible to find a real nonsingular matrix
-1
t of order b so that D- Xt has PB character. Equation 21 is rewritten in the
form
=
a(D-t) ata t t at e i (29)
Let
-1 -1
at t at = T (30)
and
e = e' (31)
-l
a(D t) a e' = i (32)
-i
Since D1 t is PB, and transformation 31 is real and nonsingular, Eq. 32 represents
the equilibrium, on a node basis, of a linear, passive, bilateral, reference network
in which the voltage variables are uniquely and reversibly related to those in the origi-
nal non-P, non-B network. The matrix T which yields this relationship is easily
obtainable from the transformation matrix t by the collinear transformation expressed
by Eq. 30.
Observe that these manipulations are not possible unless the cut-set matrix a
possesses an inversel That is why the modified approach to the analysis problem given
here is essential.
By analogy with Eq. 23, we can write
-1 A
a(D t) at = Y (33)
and regard Y as the node-admittance matrix of the PB reference network. Equations 21,
23, 31, and 32 then yield
A
Y= YT (34)
A
and if we partition Y and T in the manner shown for Y in Eq. 25, we find that
112
(XIV. NETWORK SYNTHESIS)
A A
Yi Y I In Tnf
A A
Yin ii Tin in nn
A A (35)
Yn = Yn Tp Y 7
A A
Ynn Yni n nn Tnn
nIn
-1 -1 -1
at a at D at e = i (36)
Then, we let
-i
ata T
7 (37)
and
i'= Ti (38)
-1 A
a(tD )a t = Y (40)
is then the node-admittance matrix of a PB reference network for which the equilibrium
equations are contained in
A
Ye= i' (41)
-1A
Y =T Y (42)
113
(XIV. NETWORK SYNTHESIS)
Conclusions drawn from this result are similar to those just given with regard to
Eqs. 34 and 35.
These methods make it possible to express the response functions of active and/or
nonbilateral networks as linear or bilinear combinations of the response functions of a
passive, bilateral, reference network. The application of this technique to synthesis
problems will be discussed in later reports.
E. A. Guillemin
References
1. Cf. E. A. Guillemin, Introductory Circuit Theory (John Wiley and Sons, Inc.,
New York, 1953), Chap. X.
2. E. A. Guillemin, Notes for Principles of Electrical Communications 6.30
(Massachusetts Institute of Technology, Cambridge, Massachusetts, 1941), Chap. 3,
pp. 9-18. (Out of print.)
3. E. A. Guillemin, Introductory Circuit Theory, op. cit., pp. 494-495.
The following discussion will show how these relations can be extracted directly
from the Cauchy integral formula. Let y(s) be analytic in the right half of the
s-plane, inclusive of the j-axis.
Physically, y(s) may be a propa-
JW gation function, in which case we
have y = a + jP with a and P equal
Ij y(k) dX
.y(s) = - (1)
114
(XIV. NETWORK SYNTHESIS)
and for the external point -s the Cauchy integral law yields
0 / (X) dX
0 2rj (+s
s y(X)
y dX
y(s) = s2 2 (3)
or
1 y(X) dX
==y(s) (4)
S2 2
In either of these integrals, the path of integration may be abridged to the j-axis if
the contribution from the arc is zero. If the asymptotic behavior of y(X) is described
by -y() - ka for X - oo, then in Eq. 3 the contribution from the arc is zero if a < 1, and
in Eq. 4 if a < 0. Under these conditions, we can write integrals 3 and 4 in the forms
(s) T 2 2
and
where the change in algebraic sign of the integrand is the result of traversing the j-axis
in the opposite direction from that shown in Fig. XIV-4. Since X is now restricted to
the j-axis, we shall write X = j and consistently have
where a() and P(a) are the loss and phase functions along the j-axis, as usually
defined. Since these functions are even and odd, respectively, integrals 5 and 6
become
115
(XIV. NETWORK SYNTHESIS)
s a( ) d
y(s) = s (8)
-1 3(J) 5 d 5
y(s) 0 S (9)
2 2
-00 s +
and
s s 2
zj s j = u o() -
uo(~) - 2 2 (11)
( I 2)Sj
s-J Co
in which u (c) is the unit impulse function occurring at 0 = 0. In other words, we must
not overlook the j-axis impulses that result when the left-hand functions in Eqs. 10 and
11 are evaluated along the j-axis of the s-plane.
For integral 8, we thus obtain
001
) a( ) d _ Z0o a( ) d
(13)
r2 2
00 - W - o
116
(XIV. NETWORK SYNTHESIS)
-1 p( ) d5 0P(0 ) d
) = -a( (15)
T2 2 2 2
Equations 13 and 15 are the familiar Hilbert transforms in the form in which the
even and odd character of a(w) and P(w) are taken into account. In Eq. 13 we observe
that if the asymptotic behavior of a(,) is indicated by writing a(S) - a for - oo, then
a finite value for this integral can result only if a < 1. Essentially, this condition is
the Paley-Wiener criterion. If we consider integral 8 whence Eq. 13 stems, we can
say that the asymptotic character of a( ) must be such that this integral has a finite
value for all finite values of s in the right half-plane. For this to be so, it is obviously
sufficient that the integral have a finite value for s = 1 or that
0 () d
(16)
be finite, which is the more common form in which the Paley-Wiener criterion is
stated.
A related question is: How must the behavior of p(,) in Eq. 15 be restricted for
S- oo if this integral is to have a finite value? Here we find that if we write P(,) -. a
for - oo, we must require a < 0. In fact, the conditions a < 1 for a(,) and a < 0 for p(g)
are just the conditions assumed initially for the asymptotic behavior of y(X) in Eqs. 3
and 4, in order to render the contributions from the semicircular arcs zero, so that the
subsequent derivation of the Hilbert transforms can be carried out. The resulting
restriction upon a(g) we accept without reservation because it coincides with that
imposed by the Paley-Wiener criterion; but the restriction just stated with respect to
P() is somewhat puzzling because we know that phase functions of minimum phase-shift
networks are not so restricted. The condition a < 0 states that p(g) must become zero
for - oo; and we know that in networks whose transfer functions have zeros at s = 0o,
the phase approaches a constant nonzero asymptote. In fact we can conceive of mini-
mum phase-shift networks with continuously increasing phase, for which, therefore,
P(g) - g for g - oo. What is the explanation of this seeming inconsistency?
It is simply that a(w) in Eq. 15 (unlike P(w) in Eq. 13) is determined only within an
arbitrary additive constant; and the value of this constant is sometimes infinite. It is
infinite, for example, in any situation in which the phase p(w) approaches a constant
asymptote. In such a case we should not try to compute a(w) from P(w) but instead find
a(w) - a(0)
- =2 (17)
(iv)
117
(XIV. NETWORK SYNTHESIS)
in which any additive constant in the function a(w) drops out. In other words, integral 17
yields the functional variation in a(w), which is really all that we are interested in. For
its determination according to integral 17, we see that the asymptotic character
described by P( ) - a for
- c allows a to be larger than unity as long as a < 2. This,
then, is the true restriction upon the phase function.
As an example of the use of Eq. 17 we readily compute for the specification
the solution
which checks with the solution to this situation obtained by other methods.
Again, for p(,) = , Eq. 17 gives a(w) - a(O) - 0, which agrees with the well-known
physical fact that a linear phase shift over the infinite spectrum is associated with a
constant attenuation, albeit an infinite attenuation if we conceive of obtaining this result
with a minimum phase-shift network.
It is generally thought that the usefulness of Hilbert transforms is restricted to
situations in which the given real or imaginary part is either graphically or analytically
specified over intervals of the j-axis, as, for example, in specification 18, or in the
computation of the phase associated with an attenuation function defined by confluent
straight-line segments or arcs. Computation of the imaginary part from Eq. 13, when
the real part is given as a rational function of the frequency variable s, is regarded as
leading to an integral that is difficult to evaluate; and so such problems are usually
solved by other methods (the Bode or Gewertz or Miyata methods). Through use of
Eq. 5 or Eq. 8 (equivalent to Eq. 13) together with methods of complex integration, we
can use the Hilbert transforms for problems of this sort with the same facility as the
algebraic methods. In fact, we shall show that use of the Hilbert transform (Eq. 13)
or of Eq. 8 leads directly to Bode's form of the algebraic solution.
Let us do this for an impedance Z(s), for which we write Z(jw) = R(c) + jX(c), and
have Eq. 13 in the form
X(- ) 2 (20)
The j-axis real part of the impedance Z(s) is expressible in the familiar form
118
(XIV. NETWORK SYNTHESIS)
in which R(-s2 ) is the even part of Z(s). Since we want to use the methods of complex
integration for the evaluation of Eq. 20, it must be converted into a contour integral.
This process is obviously just the reverse of the above derivation of Eq. 13 from
Eqs. 5 and 8. If we use Eq. 5, and close the path of integration by adding the semi-
circular arc, as shown in Fig. XIV-4, we obtain
2 2
where traversal of the path is again counterclockwise. Since the even part of any physi-
cal impedance must be regular at infinity, the conditions for zero contribution from the
semicircular arc are obviously fulfilled.
If the left half-plane poles of Z(s) are denoted by k i , and the residues of Z(s) in
these poles are k i , then a partial-fraction expansion of R(-k2) reads
( )1 k i
2 1 k. (23)
i i
2)
because the residues of R(-k in its left half-plane poles are (according to Eq. 21)
equal to ki/2, while those in the right half-plane poles are the negatives of these values,
as is clear from the quadrantal symmetry of the pole-zero pattern of R(-X ).
The partial-fraction expansion of the integrand in Eq. 22 is then seen to be
sk
s R(- 2) R(-s2) R(-s) 1 i i 1
+ i _ iI (24)
2 2 2(x - s) 2(X + s) - s i 1
By Cauchy's residue theorem the value of the integral in Eq. 22 is equal to 2Trj multi-
plied by the sum of the residues of the integrand in those poles enclosed by the contour
of Fig. XIV-4. These poles are at k = s and k = -k.. The residues are evident in
1
Eq. 24, and so, for the evaluation of Eq. 22, we have
sk.
Z(s) = R(-s) 2 2 (25)
-
i
1 k. k. k k.
Z(s) +
or s+ s s
or
119
(XIV. NETWORK SYNTHESIS)
k.
Z(s) (26)
i
Darlington showed that any positive real rational function (realizable driving-point
impedance) can be constructed as the impedance on the input side of a lossless two
j LOSSLESS 2 I LOSSLESS 2i
NETWORK NETWORK
(1+z ) Y22
22
Fig. XIV-5. Network for relating open- Fig. XIV-6. Network for relating short-
circuit transients to imped- circuit transients to imped-
ance functions, ance functions.
m 1 + n1 P(s)
- m (1)
1m2 + n2 Q(s)
in which, as usual, m and n denote the even and odd parts of either polynomial P(s)
or Q(s). If the network in Fig. XIV-5 is excited somehow (by having a little boy throw
coulombs at the capacitors), its natural frequencies are the poles of Z (s) or the zeros
of (1 + z 2 2 ), and hence are zeros of Q(s) = m 2 + n 2. Here z 2 2 is the open-circuit
impedance of the lossless network at its terminal pair 2, and (1 + z 2 2 ), therefore, is
the impedance that would result from cutting into the output mesh, as indicated in
Fig. XIV-5. It follows that we must have
120
(XIV. NETWORK SYNTHESIS)
m 2 + n2 m
Case A: 1+ z 21 + - 2 (2)
22 n2 n2
or
m? + n 2 n
Case B: 1+ z - 1 + -- 2 (3)
22 m, m2
If the network in Fig. XIV-6 is excited, the natural frequencies are zeros of Zl(s),
and hence zeros of P(s) = m 1 + n 1 ; but they must also be zeros of 1 + (1/y 22 ), where
YZ2 is the short-circuit admittance of the lossless network at its terminal pair 2. We
have
1 m1 + nl n1
Case A: 1 + - 1 + (4)
Y22 ml ml
or
+ ml
1 ml nl
Case B: 1 + - 1 + (5)
Y22 n1 n1
Let z 11 denote the open-circuit impedance of the lossless network at its terminal
pair 1. We may then say that the poles of z 11 are poles of z 2 2, and zeros of z 11 are
poles of 1/y 22 . From Eqs. 2, 3, 4, and 5 it follows that
=
For case A: zll ml/n 2 (6)
or
=
For case B: z11 nl/m 2 (7)
Let yll denote the short-circuit admittance of the lossless network at its terminal
pair 1. The poles of yll are the poles of y2' and the zeros of yll are poles of 1/z 2 2.
These conclusions (as well as those in the previous paragraph) follow from the con-
sideration that the critical frequencies in question are natural frequencies of the lossless
network under identical terminal constraints. For example, for the poles of y 1 1, the
constraints are: terminal pair 1 shorted, terminal pair 2 shorted; the same holds for
the poles of YZ2 . For the zeros of yll or for the poles of 1/z2 2 , the constraints are:
terminal pair 1 open, terminal pair 2 shorted. Thus, we have
or
From the fact that reactance or susceptance functions must be ratios of two
121
(XIV. NETWORK SYNTHESIS)
polynomials, one of which is even and the other odd, it follows, incidentally, that
cases A and B are mutually exclusive.
From the separation property of the zeros and poles of reactance functions, and of
the even and odd parts of Hurwitz polynomials, it follows that P(s) and Q(s), as well
as P+(s) = m i + n 2 and Q (s) = m 2 + nl, are Hurwitz polynomials. We may conclude
that the so-called double-alternance (1) property holds for the polynomials m 1 m 2 , and
n 1n2 . As Reza has shown (1), this double-alternance property is a necessary (not
sufficient) condition to ensure the positive real character of Z (s). He also showed
that a rational function with this property can be made to fulfil positive real conditions
merely through assigning to either m 1 m 2 or n n 2 an appropriate constant multiplier,
which in our situation amounts to scaling the impedance level of one end of the lossless
network. It follows, therefore, that the physical representation in Fig. XIV-5 enables
the construction of any given positive real impedance function Z (s). Q.E.D.
E. A. Guillemin
References
122