1950 1PROCEEDINGS OF THE I.R.E.

291

Frequency Analysis of Variable Networks*

LOTFI A. ZADEHt, ASSOCIATE, IRE
Summary-This paper describes an approach to the analysis of A basically different approach to the analysis of vari-
linear variable networks which is essentially an extension of the able networks is developed in this paper. The method
frequency analysis techniques commonly used in connection with
fixed networks. It is shown that a function H(jco; t), termed the to be described consists essentially of an extension of
system function of a variable network, possesses most of the fun- the frequency analysis techniques commonly used in
damental properties of the system function of a fixed network. Thus, connection with fixed linear networks. The advantages
once H(jw; t) has been determined, the response to any given input of frequency domain analysis, as compared to time do-
can be obtained by treating H(jw; t) as if it were the system functionmain analysis, are well known to electrical engineers. It
of a fixed network. It is further shown that H(jco; t) satisfies a linear
differential equation in t, which has complex coefficients and is for will be shown that similar advantages can be secured in
the same order as the differential equation relating the input and the case of variable linear networks in which the ele-
the output of the network. Two methods of solution of this equation ments are functions of time and possibly the input, but
covering most cases of practical interest are given. In addition to not the output.
H(jco; t), a network function introduced is the bi-frequency system It should be remarked that several attempts have
function r(jw; ju) which is shown to relate the Fourier transforms
of the input and the output through a superposition integral in the been made in the past, notably by Carson,6 and Neu-
frequency domain. On the basis of the results obtained in this paper feld,1 toward extension of Heaviside's operational cal-
it appears that the frequency domain approach using H(jw; t) culus to variable linear systems. Carson's method is
possesses significant advantages over the conventional approach similar in some respects to Schelkunoff's wave per-
using the impulsive response of the network. turbation method, while that of Neufeld is based on
regarding variability of an element as due to a continu-
ous series of switching operations. Both procedures are
I. INTRODUCTION essentially "time domain" methods, as contrasted to the
N RECENT years there has been an increasing "frequency domain" approach used in this paper.
recognition on the part of communication en-
gineers and investigators in allied fields of the po- II. SYSTEM FUNCTION OF A VARIABLE NETWORK
tentialities of so-called dynamic or variable networks. The conventional steady-state as well as transient
Generally speaking, a variable network is one in which analysis of fixed linear networks is based on the use of
one or more element-values are dependent in a specified a function H(jw) which is variously known as the trans-
manner upon a combination of the three variables- mission function, transfer factor, gain, system func-
time, input, and output. The simplest, though not tion, etc., of the network under consideration. The
necessarily the most useful type of variable network, is physical significance of fI(jw) needs no explanation, but
that in which the element-values are functions of time for the purposes of this analysis it will be useful to recall
only. that II(jw) may be regarded as the Fourier transform
The rather limited use of variable networks at the of the response of the network to a unit impulse applied
present time is due largely to lack of practical means at t=O. More generally, denoting a unit impulse oc-
for their analysis, synthesis, and mechanization. In par- curring at t= by the usual symbol 5(t- ), and repre-
ticular, the available tools for the analysis of linear senting the associated response by W(t-h), we can
variable systems consist essentially of Green's functioni write
approach and various perturbation methods of which
certain variants of Picard's method,2 and Schelkunoff's H(jw) = Wf(t - (1)
method,8'4 are probably the most powerful. These meth- -x

ods are quite useful in certain cases, but on the whole

would not be considered practical by an electrical en- The fundamental characteristic of fixed networks is
gineer. that their impulsive response is dependent solely upon
the so-called age variable, that is, the difference between
the instant of observation i and the instant t of applica-
*Decimal classification: R143. Original manuscript received by tion of the unit impulse. No such property is possessed
the Institute, April 11, 194Q. Presented, 1949 IRE West Coast Con- by variable networks, for in these the impulsive response
vention, San Francisco, Calif., Septenmber 1, 1949. is of the general form W(t, t). Nevertheless, by analogy
t Columbia University, New York, N. Y.
1 E. L. Ince, "Ordinary Differentiail Equations," Dover Publica- with (1) we define the system function of a variable
tions, New York, N. Y., p. 254, et seq.
2 See p. 63 of footnote reference 1. Additional references are given
in the text.
I S. A. Schelkunoff, "Solutiion of linear and slightly non-linear 'J. R. Carson, "Theory and calculation of variable systems,"
equations," Quart. App. Math., vol. 3, pp. 348-355; January, 1946. Phys. Rev., vol. 17, pp. 116-134; Februiary, 1921.
' M. C. Gray and S. A. Schelkunoff, "Approximate soluition of 6 J. Neuifeld, "Extension of Heaviside's calculus to networks
linear differential equations," Bell Sys. Tech. Jour., vol. 27, pp. 350- whose parameters vary with time," Phil. Mag., vol. 15, pp. 170-177;
364; April, 1948. Additional references will be found in the text. January, 1933.

Authorized licensed use limited to: Peter the Great St. Petersburg Polytechnic Univ. Downloaded on March 29,2022 at 13:25:58 UTC from IEEE Xplore. Restrictions apply.
 292 PROCEEDINGS OF THE I.R.E. March
linear network7 by the relation
el(O=-JEl(jiw)eiwtdw,
2 7r _00
(7)
HJ(j,w; t) = f W(t, ei()d
_00 (2)
-co and substituting (7) into (5), we obtain upon inversion
where the upper limit of the integral is actually I, since of the order of integration,
the impulsive response is zero prior to application of the
unit impulse, i.e., W(t, 0)_O for t<t. e2(t) = f dwE,(jfr) f W(t, )eiltd{. (8)
27r _x -x
It will be observed that in contrast to H(jw), H(jw; t)
as defined by (2) is a function of jw involving t as a But, by the definition of H(jco; t) (see equation (2))
parameter. In what follows it will be demonstrated that
H(jco; t) represents a natural extension of the notion of W(t, {)ei-4dt = H(jw; t)eiol, (9)
the system function of a fixed network. Thus, it will
appear that H(j&o; t) not only possesses many of the and hence
fundamental properties of H(jw) but, what is more
important, can also be used in a similar manner to ob- 1 f0
tain the steady-state as well as the transient response e2(t
()=2 7r _oc H(jo; t)Ej(jo)ejloldc.
- (10)
of a variable network to any prescribed input.
This result represents a generalization of a familiar rela-
Formulation tion used in the analysis of fixed networks, namely,
Analysis of a variable linear network in which the
parameters are functions of time reduces in general to
the solution of a linear differential equation of the form
e2(t)
= 217r fi 00
H(jco)Ej(jc)eioldcw.
(11)

[an(t)pn + * + al(t)p + ao(t)]e2(t) Heaviside's Expansion in Variable Networks

= [bm(t)pm + * * * + b1(t)p + bo(t)]ej(t), (3) It is not difficult to see that the conventional tech-
niques used for the evaluation of (11) can be used as
or, more compactly, well for the evaluation of (10). In applying these tech-
niques to (10), H(jw; t) should be treated as if it were the
L(p; t)e2(t) = K(p; t)ei(t) (4) system function of a fixed network; that is, in
H(jw; t),
where el(t) and e2(t) denote respectively the input and t should be regarded as a fixed parameter. For example,
output signals; the a's and b's are known functions of assuming that H(jw; t) El(jw) has only simple poles,
time; and p=d/dt. Throughout this paper it will be we can express e2(t) in a standard form of Heaviside's
assumed that the system is of the type described above; expansion
that the a's and b's are continuous functions of time, and
that the system is unexcited, though not necessarily e,2 (t) = E P(jco -; t) (12)
fixed, prior to application of e1(t). Equation (3) will be ______ Q(jcvil; t)
referred to as the fundamental equation of the system. la (jcol)
We shall now proceed with establishing a few of the where
more important properties of the system function of a
variable network. First, it will be recalled that by the P(jw; t) =numerator of H(jw; t) El(jo)
principle of superposition, Q(jw; t) =denominator of H(jw; t) Ei(jw)
rco jw,,'s=poles of H(jw; t) El(jw).
e2(t) = W(t, t)ei(t)dS (5) The main advantage of frequency domain analysis
over time domain analyses is due largely to the fact
where W(t, t), the impulsive response of the system, is that once H(jc; t) has been determined, the output
defined by the relation can be obtained from Laplace transform tables-or
L (p; t) W (t, t) = K(p; t)5 (t - (6) more generally-by means of contour integration.
Methods for the direct determination of H(jw; t) will be
Expressing ej(t) in terms of its Fourier transform described in part III. In the meantime, it will be in-
El(jw), structive to compare the steps involved in the calcula-
tion of e2(t) (on the assumption that W(t, t) is known
7The relation of H(jw; t) to Green's function of the system, as a priori), first through use of the superposition integral,
well as certain other aspects of the problem, will be considered in a and alternatively through use of Heaviside's expansion.
forthcoming paper. It recently came to the attention of the writer When W(t, t) is known, calculation of e2(t) reduces
that a similar, though somewhat less general, definition has been
given by C. Blanc in his paper "Sur les equations differentielles line- essentially to evaluation of (5), which in most cases
aires a coefficients lentement variables," Bull. Tech. Suisse Romande,
vol. 74, pp. 185-188; July, 1948. Blanc's paper does not contain the would be a difficult problem. On the other hand,
calcu-
results given in the present paper. lation of H(jco; t) requires evaluation of a generally

Authorized licensed use limited to: Peter the Great St. Petersburg Polytechnic Univ. Downloaded on March 29,2022 at 13:25:58 UTC from IEEE Xplore. Restrictions apply.
 1950 Zadeh: Frequency Analysis of Variable Networks 293

simpler integral, namely, response of the network to eiwt

H(j ; t) =
eJwt
H(jc,; 1) = e-iwtf W(t, t)eiutd(. (13)
-00 which is identical with the usual definition of the sys-
Once H(jw; t) has been determined, the second step tem function of a fixed network. In other words,
in the calculation of e2(t), i.e., use of Heaviside's H(jw; t) is the complex envelope of the response of the
expansion, does not entail any difficulties, since it in- system to e '"t.
volves only algebraic operations, Hence, on the whole it The Case of Noisy Input
would be simpler, in genieral, to use frequency analysis
even when W(t, t) is known a priori. The case where the input signal contains some noise
When W(t, t) is not known a priori-as would usually can be treated very conveniently by means of (16). It
be the case in practice-the simplest procedure to fol- will be necessary, however, to restrict the analysis to
low in most cases would be to determine H(jco; t) di- systems in which the variable parameters are largely
rectly by using one of the methods described later; and independent of noise. In other words, the parameters of
then use Heaviside's expansion or some other standard the system must either be completely independent of
technique for the evaluation of (10). the input signal, or else the signal-to-noise ratio at the
In passing, it may be noted that the question of input must be reasonably high. The method of analysis
stability of a variable network may be resolved in much outlined below applies only to such cases.
the same way as in the case of a fixed network. That is, In view of the assumption made above, the output
a variable network is stable at all times if the poles noise e2N (t) is related to the input noise elN(t) through
of H(jw; t) (whose location in general varies with time) the fundamental equation of the system, i.e.,
do not pass into the right-half of the complex frequency L(p; t)e2N (t) = K(p; t)elN(t)c (17)
plane.
Assuming further that ei2(t) is stationary and Gaussian
Steady-State Response to Periodic and Almost Periodic with a power spectrum S(f), we can express elN(t) in the
Inputs form of a Fourier series8
In many cases of practical interest the input signal is 27r 2r 47r
periodic or almost periodic in nature; in other words it elN(t) A1 cos - t + B1 sin - t + A2 cos- t
T T T
is of the form
4w
el(t) = Co + A1 cos wit + B1 sin cwit + B2sin-t + (18)
T
+ A2 cos C02t + B2 sin W2t + * * (14) -

where T is a long period of time, and the A's and B's

where C0, the A's, the B's, and the c's are constants. are normally distributed random variables with the
Steady-state response to signals of this type can be following properties:
expressed in a convenient form through use of (10).
Thus, it can easily be verified that for (14) A = B,, = 0, (19)
El(jw) = 7r[2Co(co) + (A1 - jBi)b(ci - co,) A 2 = B,,2 =S(fA)Af, (Af=-T f,= Af) (20)
+ (A1 + jBi)>(Q. + co,) + * * ] (15)
and
where b(co) represents a unit impulse in the frequency
domain. Substituting (15) into (10), we obtain e2(t); AABl3 = 0. (21)
the result can be expressed in several different forms of In (19), (20), and (21), the bars indicate, as usual, the
which the one most useful for the purposes of this mean (expected) values.
analysis is A similar representation of the output noise may be
e2 (t) = CoH(0; t) + A1 Re { H(joil; t)eiwlt } + * * obtained immediately through use of (16). Thus, we
-

(16) find
+ B1 Im {H(jwi; t)eiwlt} + * * -
where Re and Im have their conventional meaning. It e2N(t) = >
IL
[A, Re { H(jw,; t)ei i'ytI
will be noticed that (16) is identical in form to the ex- + B,,, Im {H(jw,.; t)eidt} ]. (22)
pression used in the case of fixed networks, the only
difference being, as usual, that in (16) H(jw; t) involves It will be observed that e2N(t) is a linear combination
t, whereas in the case of fixed networks H(jw) is inde- of normally distributed random variables having zero
pendent of t. It will also be noted that from (16) it fol- mean, and hence at any fixed instant of time e2N(t) is
lows that the response of a variable network to a com- likewise a normally distributed random variable with
plex exponential eiOt is H(jw; t)eiOt. It is thus evident 8 S. 0. Rice, "Mathematical analysis of random noise," Bell Sys.
that H(jw; t) may alternatively be defined as Tech. Jour., vol. 23, p. 328; July, 1944.

Authorized licensed use limited to: Peter the Great St. Petersburg Polytechnic Univ. Downloaded on March 29,2022 at 13:25:58 UTC from IEEE Xplore. Restrictions apply.
 294 PROCEEDINGS OF THE I.R.E. March
zero mean. The ensemble variance a2(t) of e2N(t) may be
determined as follows. H (jw;t) =
By the definition of o-2(t), 0
a2(t) = e2N'(t), (23)
or

o2(t) = ;
E [A,2Re2 { H(jwo; I)eiw t I

+ B 1lm2 { H(jw,; t)eiwtII J. (24)

Making use of (20), (24) reduces to
a,.2(t) = j S(f,)Lf
[Re2 I H(jw,; t)eis I}
I
1-I
+ Im2 { H(jw,o; t)eiwt I I (25)
or, L -~~~~~~~~-
2(t) = E S(fm)Af H(j]m; t) 12, (26)
Fig. 1-Expansion of a periodically varying network.
and in the limit as T-> oo we obtain The type of expansion indicated above suggests a
natural extension to networks in which the variation
^2(t) 10 TH(jc; t) S(f)df.
(27) of parameters is not periodical. Thus, we are led to in-
troducing the notion of what might be called the bi-
This result is a generalization of the well-known relation frequency system function of a variable network. This
function rFJc; ju) is simply the Fourier transform of
t)eiwc (the response to eiOw) with respect to time,
f2 H(j&o) 12S(f)df (28) HCjw;
i.e.,
which holds in the case of fixed networks. P(j; ju) H H(ji; t)eiwte-iutdt. (30)
Expansion of Variable Networks
The bi-frequency system function has a number of
It is -evident that the system function of a network interesting properties. For example, it can easily be veri-
having periodically varying parameters, is itself a peri- fied that the Fourier transforms of the input and out-
odic function of time. Letting cwo denote the funda- put are related to each other by rF(&; ju) in the follow-
mental frequency of variation of parameters, it follows ing manner:
that the system function of a periodically varying net-
work may be expressed in the form of a Fourier series, rco
such as E2(jU) =

-wc
r (jw; ju) El (jc) dco. (31)

H(jw; t) Ho(jcj) + Hl'(jce) cos coot

= This relation is a frequencydomain analogue of the super-
+ Hi"(jco) sin coot + H2'(jco) cos 2coot position integral (see (5)).
+@* - -
(29) Differential Equation Satisfied by H(Uw; t)
where HoUco), H,'jco), Hh"(jw), etc., assume the place An important property of H(jco; t) is that it satisfies
a linear differential equation, which, in many practical
of usual constants. It will be noticed that Ho(&c),
cases, is easier to solve than the fundamental equation
Ho'(jc), Hi"(jco), etc., do not involve time, and hence of the system. Furthermore, it should be remembered
represent system functions of fixed networks; in practice
they would be automatically provided in the process of that possession of H(jw; t) enables one to find the re-
determination of H(jw; t). (See the illustrative example sponse to any input signal provided, of course, the
at the end of part III.) parameters of the system are functions of time only; in
Physically, (29) means that a periodically varying contrast, solution of the fundamental equation yields
network may be "expanded" as a parallel-series com- only the response to a particular input signal.
bination of a number of fixed networks having sinusoid- A convenient starting point for the derivation of the
ally varying gains (see Fig. 1). The advantage of this differential equation in question is furnished by the
type of representation is that it places in evidence the definition of the impulsive response of the system, i.e.,
mechanism of formation of the sidebands in the output. L(p; t)W(t, t) = K(p; t)c(t t). (6) -

Authorized licensed use limited to: Peter the Great St. Petersburg Polytechnic Univ. Downloaded on March 29,2022 at 13:25:58 UTC from IEEE Xplore. Restrictions apply.
 1950 Zadeh: Frequency Analysis of Variable Networks 295

Thus, transforming both sides of (6) with respect to t bm(t) (jc,) m + * * * + bi (t)jcw + bo(t)
we obtain
.
(38)
+ al(t)jw + a.(Q)
an(i)(jco)n + * * *

L(p; t) = K(p; t) L(t - t)eiwtdt, (32)

W(t, t)eiwEd- The physical significance of (38) is rather obvious.
In plain words it means that the system function of a
very slowly varying system is the system function of the
which in view of (9) reduces to fixed network resulting from freezing the variable net-
L(p; t)H(jco; t)eiwS = K(p; I)eiwt. (33) work at the instant of consideration. Because of this fact,
the ratio Hf(jw; t) = K(jw; t)/LC(jw; t) shall be referred
This result may also be obtained by noting that to as the frozen system function of the network (the sub-
HCjw; t)eiwt is the response to ei"'. script f standing for "frozen").
Now, it can be verified easily that the result of opera- The notion of the "frozen system function" provides
tion of any polynomial operator F(p; t) upon the prod- the basis for a useful interpretation of (37). Thus, re-
uct of two time functions u(t) and v(t) is given by the ferring to (37) it will be seen that H(jco; t) may be
expression formally regarded as the response of a virtual system-
du dF(p; t) + of which (37) is the fundamental equation-to an input
F(p, t)uv uF(p; t)v + d
= v+* consisting of the frozen system function of the network.
This point of view will be found particularly useful in
1 d%u anF(p; t) connection with the solution of (37).
+ - - v. (34) It will be seen later that in the case of a slowly vary-
n! dtn dpn ing system Hj(jo; t) furnishes the first approximation
Identifying u with H(co; t), v with eOw, and noting that to H(jw; t). The question of goodness of this approxima-
tion will be discussed in part III. Tentatively, a con-
(35) venient though rather rough criterion may be stated as
dIPAl(;)e<e
apLOP ) a(io;)
p(jw)IA)
is X
follows:
and The frozen system function may be regarded as a first
approximation the actual system function of a variable
K(p; t)ejwt K(jw; t)ei"', (36) network wheneverto the coefficients of the fundamental equa-
we obtain after minor simplifications the following dif- tion do not vary appreciably over the width9 of the impulsive
ferential equation: response of the system.
1 d-nL dnH ~ 1 dL dH
It can be shown easily that for large values of co the
[nAL a(jc)n d + L.. + [7 (jw) actual
dt system function tends asymptotically to the
frozen system function, regardless of the rate of varia-
K
+H= (37) tion of the system. The physical significance of this
statement becomes quite obvious when one considers
the fact that in the case of input signals of very-high
where K, L, and H are abbreviations for K(jc; t), frequency the coefficients of the fundamental equation
L(jco; t), and H(Uw; t), respectively. It should be under- do not change appreciably over a relatively great num-
stood that in (37) jw is regarded as a fixed parameter. ber of cycles of the applied signal. In other words, from
Equation (37) could be obtained alternatively by noting the point of view of the signal, the system behaves as
that the left-hand member in (33) is equivalent to if it were varying at a very slow rate.
eico' L(p+jwo; t) H(jw; t).
Equation (37) shows that H(jco; t) satisfies a non- III. DETERMINATION OF SYSTEM FUNCTION
homogeneous linear differential equation with complex The problem of finding the response of a variable net-
coefficients, the order of which is the same as that of work through analysis in the frequency domain reduces
the fundamental equation of the system. The problem essentially to the determination of the system function
of solution of this equation will be discussed in part III. of the network. There are two distinct approaches to
In the meantime it will be expedient to examine the the solution of the latter problem. The first, and most
limiting form of (37) in the case of a slowly varying practical, is to obtain H(jw; t) through the solution of
system. (37). The alternative approach is to determine the
As the rate of variation of a system decreases the impulsive response of the system, and then obtain
time derivatives in (37) decrease accordingly until in H(jw; t) through use of (9). This approach is not, in
the limit H(Jw; t) becomes equal to

L(jw;t1)t)
H(jwt)1) = K(jw; ' Roughly speaking, the width of W(t, t) is the length of the time
H!(jco; interval outside of which W(t, E) is small by comparison with its
maximum value.

Authorized licensed use limited to: Peter the Great St. Petersburg Polytechnic Univ. Downloaded on March 29,2022 at 13:25:58 UTC from IEEE Xplore. Restrictions apply.
 296 296 PROCEEDINGS OF THE I.R.E. March
general, very convenient, even though it is preferable to should not have any effect upon the response, and hence
the use of the superposition integral. It will be found the initial values of H(jw; t) within the period of ob-
that the two direct methods described in the sequel are servation can be chosen at will. The apparent arbitrari-
generally adequate for the determination of H(jco; t) ness that is involved in the choice of a particular solu-
in most practical cases. As a prerequisite to the solution tion is eliminated in the process of evaluation of (10).
of (37), it is necessary to establish the boundary condi- It should be remarked that in most transitionless cases
tions to which H(jw; t) is subjected. This matter is evaluation of (10) is simplest when H(jco; t) is set equal
examined in the following article. to the steady-state solution of (37). Actually, this is
true in most practical cases so that ordinarily the prob-
Boundary Conditions lem of determination of H(jc; t) reduces to finding the
The question of boundary values of H(jco; t) in the steady-state solution of (37).
time domain arises whenever there is a transition in the
state of variability of the system. Physically, a transi- General Solution
tion can assume a variety of forms. For instance, the For the purpose of solution of (37) it is somewhat
process of initiation of variability represents essentially more convenient to rewrite (37) in the following form:
a transition from a fixed state to a variable state. A
further example is furnished by the case of a dynamic dnH dH
network controlled by the input. Here, at the instant (40)
dtR + + no dt + co(t)H =f
of application of the signal there occurs a transition
from the state of variability existing prior to applica- where
tion of the signal, to a different state which is dependent 1 &"L
upon the signal. From a more general point of view, a a, (t) =
p 01 ly ... n. .

,u!!49(jw)A
-,
,

transition taking place at, say, t=to, corresponds to a

nonanalyticity at t =to in one or more of the coefficients Now H(jf; t) is, so to say, the response of (40) to
of the fundamental equation of the system. K = K(jcw; t) and hence it can be expressed in the form
It is obvious that, regardless of the nature of vari- of a superposition integral. Thus denoting the impul-
ability of a given system, it is always permissible to sive response of (40) by'0 UH(t, t) it follows that
assume that the system has been fixed prior to t =-o. -
Adopting this point of view we can restate a statement 00
made earlier, in a more expressive form which is as fol- H(jw; t)-= UH(t, t)K(jco; t)d(. (41)
_00
lows:
The system function of a variable network may be The general solution provided by (41) is largely of
formally regarded as the response of an initially unex- academic value since the determination of UH(t, t)
cited system, of which (37) is the fundamental equation, would usually constitute a difficult problem. This means
to the frozen system function of the network. that ordinarily it would be necessary to use approximate
The above statement is perfectly general and applies, methods for the solution of (40). Two such methods
in particular, to cases where the period of observation, covering many, if not most, cases of practical interest
i.e., the time interval during which the response is ob- are described in the sequel.
served, includes one or more transitions in the state of
variability of the network. Thus, considering the fact First Method
that H(jcw; t) satisfies a nonhomogeneous linear differ- This method is essentially an adaptation of Schel-
ential equation of nth order having continuous co- kunoff's wave perturbation method ;3'4 it yields H(jw; t)
efficients, it follows immediately that: in the form of a series
H(jw; t) and its first n-I derivatives with respect to t
are continuous at a transition. That is, H(jco;t) = IHI(jw;t) + H2(jf;t) + H3(jCC;t) + * (42)
which is rapidly convergent when the coefficients of
- H(jCA; t)(t=to+) = dt H(jo; t)(1 to-)I (40) oscillate in the neighborhood of their mean values.
This condition is present in many practical cases, par-
=u-O,1, ,n-1. (39) ticularly in systems having periodically varying parame-
Equation (39) provides the necessary information re- ters. the first
garding the boundary values of H(jco; t). As step in application of the method, (40) is
In many practical cases the period of observation is rewritten as
"transitionless." In such cases, it is not difficult to see dnH dH
that any particular solution of (37) within the interval an + *I + -+oH K + P{HI} (43) =

of observation may be chosen as the system function of dt'n dt

the network. This follows from the fact that the be- 1 In mathematical terminology UH(t, t) is a Green's function of
havior of H(jco; t) outside the period of observation (40) satisfying homogeneous boundary conditions at t= - *.

Authorized licensed use limited to: Peter the Great St. Petersburg Polytechnic Univ. Downloaded on March 29,2022 at 13:25:58 UTC from IEEE Xplore. Restrictions apply.
 1950 Zadeh: Frequency Analysis of Variable Networks 297
where &; is the mean value of ax,(t) or, more generally, As the first-order approximation to H, (47) gives
a constant approximating a,(t); and P{H} represents
the aggregate of perturbation terms, i.e., H1 = Hf = frozen system function. (49)
dnH dH Successive corrections to Hi are given by the iterative
P{IH I = [~-&,-an(t) ] ~
dtn
Y it
dt relation
HU= PIHy 1} ,u= 2,3, ... (50)
+ [-di- ao(t)IH. (44)
It will be observed that (43) is a nonhomogeneous It will be noticed that the solution obtained by this
linear equation with constant coefficients. The first ap- procedure is in reality a particular solution of (37);
proximation to H(jc; t) is obtained by solving the hence, strictly speaking, the method is valid only in
equation transitionless cases. This limitation, however, would
generally not be important in practice.
dnH, dHj
an~~+*' +
dtn dt
+ aoH = K, (45) Illustrative Example
The main points of the two methods described above
while the successive corrections appearing in (42), i.e., may be conveniently illustrated by applying these meth-
H2, H3, etc. are obtained iteratively from the solution of ods to a simple variable network such as, for example, a
d nH,A, d H, low-pass bandwidth-modulated RC half-section. The
an dtn
dtn + -a-
***+ dtdt + Pt,
= PH, }.(46o) fundamental equation of this network, on the assump-
tion that the bandwidth is varied sinusoidally, may be
In this manner the solution of (45) and (46) yields the expressed in the following normalized form:
successive terms of (42). The boundary conditions that
should be used in conjunction with (45) and (46) are (p + I + p cos cot)e2(t) = (1 + p cos wot)ei(t) (51)
the same as those used with (40). It should be remem- where
bered, of course, that in transitionless cases it will be
sufficient to determine the steady-state (or particular) amplitude of bandwidth variation
solutions of (45) and (46). mean bandwidth
In practice, the first two terms of (42) will generally frequency of bandwidth variation
provide an adequate approximation. Usually it is found coo = .."

that (42) develops as a power series in a small parameter mean bandwidth

so that in many cases the question convergence
of of t= actual time X mean bandwidth.
(42) can be resolved by inspection. An example il-
lustrating the use of this method is given at the end of Using (40) it is found immediately that the differential
the section. equation satisfied by H(jc; t) is
Second Method dH
-+(j& + I + p cscot)H = 1 + pCoswo1. (52)
This method is far simpler in use than the first meth- dt
od; its usefulness, however, is limited to slowly varying
systems. The system function furnished by this method First, we shall make the assumption that p is small
has the form of a series such as (42). As in the case of by comparison with unity, and on this basis apply the
the first method, it is usually found that (42) develops first method. Next, it will be assumed that co is small by
as a power series in a small parameter, so that the mat- comparison with unity, thus making the second method
ter of convergence of (42) can be investigated in most applicable. And, finally, it will be shown that the same
cases by inspection. result is obtained from the use of the two methods when
As is usual with all other perturbation methods, the it is assumed that both p and coo are small by comparison
first step in the procedure consists of rewriting (37) in with unity.
a form that places in evidence the perturbation terms. Application of the First Method
Ordinarily it will be found simplest to regard all deriva-
tives1' of H as perturbations, and write Following the general procedure (52) is rewritten as
H=Hf = P{FH} (47) dH
-+(1 + j)H = 1 + p cos ceot-Hp cos wot, (53)
dt
where
1 d nH dH from which it follows that the first two terms in (42) are
P{H} = - L (t) d~dtnddi+ *+ J( (48) the steady-state solutions of
L1
11 Moregenerally, only those terms involving derivatives of H
dli ± (1 ±frw)H, 1+p coswcot (54)
beyond a certain order would be regarded as perturbations. dt

Authorized licensed use limited to: Peter the Great St. Petersburg Polytechnic Univ. Downloaded on March 29,2022 at 13:25:58 UTC from IEEE Xplore. Restrictions apply.
 298 PROCEEDINGS OF THE I.R.E. March

and from where it follows that the first two terms in (42)
are
t+ (I + jw)H-= - HIP Cos ot, (55) 1 + p cos cMOt
di 1l7 = Hf = (60)
respectively.
W.1 + 1 + p Cos ,,t
and
Treating the complex coefficient 1+jw as if it were 1 dH1
real, we find from (54) the first-order approximation H2 = - (61)
to H, jc + 1 + p c os wot dt
I 1 respectively.
HI ==-I jco+ pcoosincoot +s jCO)2n+ c2
+ (1
Substituting Hi as given by (60) into (61) and adding
H1 and H2 we find the second-order approximation to
H, which is
+ p COS otj (56)
( 1+ j+) 2 + 2 I + p cos toot

Hl(f; t) ico + 1 + Cos coot =

r

Substituting H1 as given by (56) into (55) we find next

i + pcoo sin c
H2 [Cos Wot (jw + I + p cos coot) I
1
+ 0(cwo2). (62)
+ pwo sin wot (I-+ MI1,)2o+ (57) It will be observed that in this case H(jc; I) develops
as a power series in coo. If in (62) an additional assump-
Adding Hi and H2 we obtain the second-order approxi- tion is made to the effect that p is small by comparison
mation to H, namely, with unity, H(jw; t) reduces to
1 jcW 1
1 j
H(jw; t) + p cos wot ( + j)2 + o H(jco; t)-= + pcos woot
1 +jw (I2j)2+U0

3W)
+ PWo [ (1 + jco)2 + C,oQ] ( + jo.)
+ pWo sin ceot (1 +-j - + O(p2, W02). (63)

+ O(pi) (58)
It can easily be verified that (58) reduces to the same
expression when it is assumed that coo<<l. Thus, as
where 0(p2) stands for the terms involving the second should be expected, the first and second methods lead
and higher powers of p. It will be noticed that if higher- to identical results when they are based on identical as-
order approximations were included in (58), H(jco; t) sumptions concerning the behavior of the system.
would appear as a power series in p and, at the same As was mentioned previously, possession of H(jco;t)
time, a Fourier series of wot. This, of course, is in ac- enables one to find the response of the system to any
cordance with the general remarks made earlier. prescribed input by using the conventional "fixed net-
The second-order approximalion given by (58) would work" techniques. Thus, for example, assuming that the
be considered adequate when p is reasonably small by input signal is el(t) =cos co.t, and using (58) and (16),
comparison with unity. Higher-order approximations, we find
if necessary, may be obtained in exactly the same way @ 1
as the second-order approximation. Once the desired 62(t) - 2 cos CW8t + =sin w,,t
approximation is obtained, the response to any pre- +@ 1 +@, 2

scribed input may be found by standard means treat- pW.(2w. + wo)

ing H(ji; t) as if it were the system function of a fixed
2(1 W.2) [I + (Wo. + WO)2] + COS (W. + Wo)

network.
pW, (2 w. - W
o)
Application of the Second Method 2(1 C,2)[1 (WO -Co )21 cos (c- o)t

In this case the assumption is that coo is small by

comparison with unity; no restrictions are placed on the pC+ , [co,inc, + o) -1 ]

magnitude of p. Following the general procedure (52) 2(1 + W.2)[1 + (c. + Co)2-(2
sin , + WO)t

is rewritten as ps [C6(CO -
Wo) -
1]
1 + p Cos wot 2(1 + c,.2)f1 + (COC Wo)21 (in -

jco + 1 + p Cos wot + O(P2). (64)

1
______________ . dH (59) It can easily be verified that the direct solution of (51)
jc + I + pcos wo di leads to the same result.

Authorized licensed use limited to: Peter the Great St. Petersburg Polytechnic Univ. Downloaded on March 29,2022 at 13:25:58 UTC from IEEE Xplore. Restrictions apply.
 1950 PROCEEDINGS OF THE I.R.E. 299

CONCLUSIONS and is varying at a rapid rate. It should be recognized

The frequency domain technique described in this however, that cases such as these are difficult to handle
paper provides a powerful mathematical tool for the by any means available at present.
analysis of linear variable systems. An important ad-
ACKNOWLEDGMENT
vantage of this technique, at least from the point of view
of an electrical engineer, is its similarity with the The author wishes to express his appreciation of the
standard procedures used in the analysis of fixed net- continued help and encouragement given him by
works. In general, it may be said that frequency analy- John R. Ragazzini, under whose supervision this work
sis works best when the system under consideration is was carried out. Thanks are also due John B. Russell
varying at a relatively slow rate, and when, further- for many helpful suggestions, and to staff members of
more, the period of observation is transitionless. On the department of electrical engineering, Columbia
the other hand, the advantages of frequency analysis University, for their constructive criticism and co-
are least when the system is controlled by the input, operation.

Input Impedance of a Two-Wire Open-Line and

Cylindrical-Center Driven Antenna*
T. W. WINTERNITZt, ASSOCIATE, IRE

Summary-Using the potential theory, a method of analysis of ance. The analysis applies only to the configuration
the input impedance of a transmission line terminated by a cylindri- shown in Fig. 1. The method used is to set up two equa-
cal antenna is described. The results of application of this method to tions in vector and scalar potential differences-one
a particular configuration of line and antenna are presented
along the line in a manner analogous to that used in an
Tl VHE ANALYSIS of the cylindrical antenna by exact transmission line analysis,4 but including the
Hallen' and King2.' using an integral equation to effect of the vector potential along the antenna, and one
obtain the current distribution has proven itself analogous to that of the Hallen and King analyses, but
to be a very effective approach and has given good including the effect of the vector potential due to the
values for the input impedance of such antennas. How- antenna currents.
ever, in this approach, the assumed driving source of
potential has been taken as an infinitesimally thin
charge separating region at the center of the antenna.
Since such a source is physically unrealizable, and
since the values of impedance predicted by this anal-
ysis can only apply to antennas whose driving condi- - 2 8~~~~2
tions closely simulate such a "slice generator," a further
analysis applying to an antenna center driven by a
balanced two-wire line terminated in the center of the
antenna across a finite gap has been carried through.
This analysis leads to values of impedance for the an-
tenna and line taken at a point down the line and away
from the antenna a sufficient distance so that the mutual
coupling terms have died out. For convenience, this Fig. 1-Cylindrical antenna driven by two-wire open
distance may be taken as an integral number of half- transmssion line.
wave lengths, so that for a lossless line, as assumed, the
impedance values may be construed as antenna imped- These equations may both be converted to equations
in current and scalar potential difference, and from the
line equation an expression for the desired input im-
Decimal classification: R221 X R326.61 1. Original manuscript pedance at the generator may be obtained in terms of
*

received by the Institute, February 3, 1949. Revised manuscript re-

ceived, October 5, 1949. This paper is based on a doctoral dissertation the ratio of scalar potential difference to current at the
presented at Harvard University in 1948 under the direction of junction of line and antenna (Zs). From the antenna
R.W. P. King.
t Formerly, Cruft Laboratory, Harvard University, Cambridge, equation a second independent value for this ratio may
Mass.; now, Bell Telephone Laboratories, WVhippany. N. J. be obtained, and it may be then eliminated by substitu-
E. Hall6n, Nova Acta Royal Soc. Sci., Upsala 11, p. 1; 1938.
'R. King and D. Middleton, 'The cylindrical antenna; current tion into the line equation.
and impedance," Quart. Appl. Math., vol. 3, p. 302; January, 1946.
" R. King and T. W. Winternitz "The cylindrical antenna with 4R. King, "Electromagnetic Engineering," vol. I, McGraw-Hill
gap," Quart. A pp. Math., vol. 5, p. 403; January, 1948. Book Co., New York, N. Y.; 1945.

Authorized licensed use limited to: Peter the Great St. Petersburg Polytechnic Univ. Downloaded on March 29,2022 at 13:25:58 UTC from IEEE Xplore. Restrictions apply.