EngineeringCybernetics (英文版) 钱学森着

KANSAS
CITY,
MO PUBLIC LIBRARY
3 1148-0105^5390
ENGINEERING CYBERNETICS
H.
S.
TSIEN
Daniel and Florence Guggenheim Jet Propulsion Center

California Institute of Technology
Pasadena, California
McGRAW-HILL BOOK COMPANY, INC.

New York
Toronto
1954
London
McGraw-Hill Book Company, Inc. Printed in the
Copyright, 1954, by the
United States of America. All rights reserved. This book, or parts thereof,
may not be reproduced in any form without permission of the publishers.
Library of Congress Catalog Card
Number 54-8098
THE MAPLE PEBSS COMPANY, YOKK, PA.
To
Tsiang Yin
PREFACE
The celebrated physicist and mathematician A. M. Ampere coined the
word cyberne~tique to mean the science of civil government (Part II of
"
Essai sur la philosophic des sciences/' 1845, Paris)

Ampere's grandiose
of political sciences has not, and perhaps never will, come to
.
scheme
In the meantime,
between governments with the use

development of another branch of science,
the science of control and guidance of mechanical and electrical systems.
It is thus perhaps ironical that Ampere's word should be borrowed by
N. Wiener to name this new science, so important to modern warfare.
The " cybernetics " of Wiener ("Cybernetics, or Control and Communication in the Animal and the Machine," John Wiley & Sons, Inc., New
York, 1948) is the science of organization of mechanical and electrical
components for stability and purposeful actions. A distinguishing feafruition.
conflict
of force greatly accelerated the
ture of this
and
new
science
is
the total absence of considerations of energy,
which are so important in other natural sciences.

In fact, the primary concern of cybernetics is on the qualitative aspects
of the interrelations among the various components of a system and the
synthetic behavior of the complete mechanism.
The purpose of " Engineering Cybernetics" is then to study those parts
of the broad science of cybernetics which have direct engineering appliheat,
efficiency,
cations in designing controlled or guided systems.

It certainly includes
such topics usually treated in books on servomechanisms. But a wider
range of topics is only one difference between engineering cybernetics
and servomechanisms engineering. A deeper and thus more important difference lies in the fact that engineering cybernetics is an engi-
neering science, while servomechanisms engineering is an engineering

An engineering science aims to organize the design principles
practice.
used in engineering practice into a discipline and thus to exhibit the

similarities between different areas of engineering practice and to emphasize the power of fundamental concepts.
In short, an engineering science
is predominated by theoretical analysis and very often uses the tool of
advanced mathematics.
this quite evident.
The
ponents of the system

almost never discussed.
glance at the contents of this book makes

and design of the com-
detailed construction
the actual implementation of the theory
No
gadget
is
vii
mentioned.
are
PREFACE
viii
What
is
practice?
the justification of this separation of the theory from the

With knowledge of the very existence of various engineering
and
sciences
seems
their recent rapid development, such justification
Fluid
cited:
be
a
example could
specific
Moreover,
mechanics exists as an engineering science separate from the practice
hardly necessary.
of
and many
aerodynamics engineers, hydraulic engineers, meteorologists,
others
the results of investigations in fluid mechanics in their

In fact, without fluid mechanists, the understanding and
who use
daily work.
to
the utilization of supersonic flows would certainly be greatly delayed,
of
establishing engineering
Therefore, the justification
lies in the possibility that looking
science
an
as
engineering
cybernetics
to fruitful
at things in broad outline and in an organized way often leads
say the least.
new avenues
vistas.
of
approach to old problems and gives new, unexpected

of multifarious developments in control
At the present stage
a very real advantage in trying to
and guidance engineering, there
is
grasp the full potentialities of this

of the whole field.
new science by a comprehensive survey
Therefore a discussion on engineering cybernetics should cover reasonhave engineering appliably well all aspects of the science expected to
a
should not avoid topic for the mere reason of
in
cations
and,
mathematical
particular,
that the mathematical

artificial.
This
difficulties.
With a
little
is
all
the more true
when one
realizes
are
difficulties
usually quite
any subject
be
înterpretation, the matter could generally
of
The mathematical
brought down to the level of a research engineer.
a course in elehad
has
who
student
level of this book is then that of a
ments
of
mathematical analysis.
variational calculus,
and
On
requisite for the study.
mathematical argument
Knowledge
of
complex integration,
forms the pre-
ordinary differential equations
is
the other hand, no rigorous and elegant

if a heuristic discussion suffices.
introduced
to the practicing electronics specialist, the treatment here must

"
a mathematician interested
appear to be excessively long-hair/' but to
well
in this field, the treatment here may
appear to be amateurish. If
Hence
indeed these are the only criticisms, then, with all due respect to them,
the author shall feel that he has not failed in what he aimed to do.
During the course
of writing these chapters, the
author had the benefit
of many conversations with his colleagues at the California Institute of

R. DePrima, which
Technology, Dr. Frank E. Marble and Dr. Charles
often led to sudden clarification of an obscure point.
The task
of pre-
rendered
paring the manuscript was greatly lightened by the efficient help
author
the
all
of
To
Winkel.
L,
Ruth
them,
by Sedat Serdengecti and
wishes to extend his sincere thanks.
H.
8,
TSIEN
CONTENTS
PREFACE
CHAPTER
vii
1.
INTRODUCTION
1.3.
Linear Systems of Constant Coefficients

Linear Systems of Variable Coefficients
Nonlinear Systems
1.4.
Engineering Approximation
1.1.
1.2.
CHAPTER
2.
METHOD
2.4.
2.5.
Response to Unit Impulse
2.2.
2.3.
CHAPTER
3.
INPUT, OUTPUT, AND TRANSFER FUNCTION

.
of First-order
Systems
3.3.
Examples
3.4.
Second-order Systems
Determination of Frequency Response
Composition of a System from Elements
Transcendental Transfer Functions
3.6.
3.7.
CHAPTER
4.1.
4.2.
4.3.
4.
FEEDBACK SERVOMECHANISM
Concept of Feedback
Design Criteria of Feedback Servomechanisms
Method of Nyquist
4.6.
Method of Evans
Hydrodynamic Analogy
Method of Bode
4.7.
Designing the Transfer Function
4.8.
Multiple-loop Servomechanisms
4.4.
4.5.
CHAPTER
5.
of
...
7
8
9
10
11
First-order Systems
3.2. Representations of the Transfer Function
3.1.
3.5.
OF LAPLACE TRANSFORM
Laplace Transform and Inversion Formula

Application to Linear Equations with Constant Coefficients
"
Dictionary" of Laplace Transforms
Sinusoidal Forcing Function
2.1.
Root Locus
NONINTERACTING CONTROLS
12
12
15
18
24
29
31
32
34
34
36
38
42
46
49
49
50
53
5.3.
Noninteraction Conditions
5.4.
Response Equations
53
54
58
62
5.5.
Turbopropeller Control
Turbojet Engine with Afterburning
63
66
5.1. Control of a Single-variable

5.2. Control of a Many-variable
5.6.
System
System
ix
CONTENTS
CHAPTEE 6. ALTERNATING-CURRENT SERVOMECHANISMS AND OSCILLATING CONTROL SEKVOMECHANISMS
................
............
...
.........
6.2.
Alternating-current Systems
Translation of the Transfer Function to a Higher Frequency
6.3.
Oscillating Control Servomechanisrns
6.4.
Frequency Response
6.1.
6.5.
7.
77
80
............
.............
......
...............
..............
Stibitz-Shannon Theory
7.3.
Nyquist Criterion
7.4.
Steady-state Error
Calculation of ft* ()
83
85
87
Sampling Servomechanisrns
for
88
89
Continuously Operating with Sampling Servomecha-
of
Comparison
7 6
74
Sampling Circuit
of a
7.2.
7.5.
73
SAMPLING SERVOMECHANISMS
Output
7.1.
70
72
......
.......
Relay
with Built-in Oscillation
Oscillating Control Servomechanisrns
General Oscillating Control Servomechanism
6.6.
CHAPTER
....
of a
01
^
nisms
Pole of ft() at Origin
7.7.
CHAPTER
70
8.
..........
.............
................
...
..........
........
LINEAR SYSTEMS WITH TIME LAG
8.3.
Time Lag in Combustion

Satche Diagram
Servo
System Dynamics of a Rocket Motor with Feedback
8.4.
Instability without
8.1.
8.2.
8.5.
8.6.
CHAPTER
Feedback Servo
Complete Stability with Feedback Servo

General Stability Criteria for Time-lag Systems
9.
LINEAR SYSTEMS WITH STATIONARY RANDOM INPUTS
Random Function
9.1.
Statistical Description of a
9.2.
Average Values
9.3.
Power Spectrum
Examples of the Power Spectrum
Direct Calculation of the Power Spectrum
9.4.
9.5.
9.6. Probability of
9.7.
......
Frequency
of
.......
................
................
...........
Large Deviations from the
Exceeding a Specified Value
of a Linear
9.8.
Response
9.9.
Second-order System
9.10. Lift
System
........
Mean ......
........
to Stationary
Random Input
..............
...................
...............
...........
Random
............
RELAY
.......
.............
......
on a Two-dimensional
Airfoil in
Flow
CHAPTER
10.
Input
Design for
Input
SERVOMECHANISMS
10.4.
Approximate Frequency Response of a Relay

Method of Kochenburger
Other Frequency-insensitive Nonlinear Devices
Optimum Performance of a Relay Servomechanism
10.5.
Phase Plane
10.6.
Linear Switching
10.7.
Optimum Switching Function
10.1.
10.2.
10.3.
94
97
100
103
104
108
Ill
Ill
H3
115
117
118
123
126
127
129
an Incompressible Turbulent
9.11. Intermittent
9.12. Servo
94
.....
.................
...............
.......
.....
131
132
133
136
136
138
140
141
142
145
150
CONTENTS
Switching Line for Linear Second-order Systems
xi
10.8.
Optimum
10.9.
Multiple-mode Operation
158
NONLINEAR SYSTEMS
160
CHAPTER
11.
154
11.2.
Nonlinear Feedback Relay Servomechanism

Systems with Small Nonlinearity
11.3.
Jump Phenomenon
163
11.4.
164
11.6.
Frequency Demultiplication
Entrainment of Frequency
Asynchronous Excitation and Quenching
11.7.
Parametric Excitation and
166
11.1.
11.5.
CHAPTER
12.
164
165
Damping
LINEAR SYSTEM WITH VARIABLE COEFFICIENTS
Rocket during Burning

Linearized Trajectory Equations
12.1. Artillery
12.2.
160
162
168
168
an Artillery Rocket
Stability and Control of Systems with Variable Coefficients
171
172
12.3. Stability of
12.4.
CHAPTER
13.1.
13.
CONTROL DESIGN BY PERTURBATION THEORY
178
Equations of Motion of a Rocket
178
Perturbation Equations
13.3. Adjoint Functions
183
13.2.
13.4.
13.6.
13.7.
13.8.
185
186
Range Correction
13.5. Cutoff
Condition
188
Guidance Condition
Guidance System
Control Computers
189
190
192
Appendix: Calculation of Perturbation Coefficients
CHAPTER
14.1.
14.
14.3.
195
CONTROL DESIGN WITH SPECIFIED CRITERIA
198
198
Control Criteria
Problem
14.2. Stability
200
201
General Theory for First-order Systems
Application to Turbojet Controls

14.5. Speed Control with Temperature-limiting Criteria
14.6. Second-order Systems with Two Degrees of Freedom
14.4.
14.7.
14.8.
CHAPTER
Control Problem with Differential Equation as Auxiliary Condition

of Concepts of Control Design
Comparison
15.
OPTIMALIZING CONTROL
15.1. Basic
204
205
209
212
213
214
214
Concept
Optimalizing Control
15.3. Considerations on Interference Effects
15.2. Principles of
15.4.
176
Peak-holding Optimalizing Control
216
220
221
222
15.5.
Dynamic
15.6.
Design for Stable Operation
228
FILTERING OF NOISE
231
CHAPTER
16.1.
16.
Mean-square Error
16.2. Phillips'
16.3.
Effects
Optimum
231
Filter
Design
Wiener-Kolmogoroff Theory
235
236
CONTENTS
xii
16.4.
Simple Examples
240
16.5.
Applications of Wiener-Kolmogoroff Theory
242
16.6.
16.7.
Optimum Detecting Filter

Other Optimum Filters
16.8.
General Filtering Problem
CHAPTER
17.
ULTRASTABILITY AND MULTISTABILITY
17.1. Ultrastable
17.2.
247
250
251
253
253
System
of an Ultrastable System
An Example
256
17.4.
Terminal Fields
259
261
17.5.
Multistage System
264
17.3. Probability of Stability
CHAPTER
18.
CONTROL OF ERROR
268
by Duplication
Basic Elements
268
18.2.
...
269
18.3.
Method
of Multiplexing
271
18.4.
18.5.
Error in Executive Component

Error of Multiplexed Systems
18.6.
Examples
18.1. Reliability
INDEX
274
^
280
283
285
CHAPTER
INTRODUCTION
of
Consider a system of one degree of freedom so that the physical state

the system can be specified by a single variable y. The behavior of
the system
is
then described by taking y as a function of time t. To

y(t), it is necessary to know the structure
determine this behavior or
of the system and the properties of the individual elements of the system.
This knowledge about the system, together with fundamental physical
laws, when translated into mathematical language gives an equation for
This equation could be an integral equaintegrodifferential equation, but very often it is a differential
It is also an ordinary differential equation, because there is
calculating the function

tion or
an
equation.
y(f).
only one independent variable, the time
t.
equation is called linear, and the system described by the

differential equation, a linear system, if each term of the equation conA. differential
tains at
most only
derivatives.
products
of
first
powers of the dependent variable y or
The terms should not contain higher powers

y and
called nonlinear,
its
its
time
of y or cross
Otherwise, the differential equation
derivatives.
and the system described by the
differential equation,
a nonlinear system. Linear systems can be further subdivided into
systems with constant coefficients and systems with variable coefficients.
is
Constant- coefficient systems have constants independent of time t as

terms in the differential equation describing the system.
coefficients of the
Variable- coefficient systems have coefficients that are functions of t.

This concern about the classification of the differential equation has
its justification in
that the character of the solution of the equation
and hence the behavior of the system depend closely on the type" of the
Even more than this, the type
differential equation which describes it.
of differential equation specifies the kind of questions that can logically
be asked about the system. In other words, the type of differential
equation determines the proper approach to the solution of the engineer-
ing problem of the system. We shall see this presently.

1.1
Linear Systems of Constant Coefficients. Let us consider the
simplest system a first-order system. That is, the differential equation
of the system is a first-order linear differential equation of constant
If the system is assumed to be free and is not subjected to
coefficients.
1
"forcing functions," then the differential equation can be written as
f+%
k
(1.1)
be called the spring constant and is real. When there is no variay with respect to time, dy/dt vanishes, and Eq. (1.1) requires
Therefore the stationary, or equilibrium, state of the system
0.
may
tion of
ij
corresponds to y
The
0,
solution of Eq. (1.1)
is
where
yo is the initial value of y
= y^-kt
(1.2)
or
(1.3)
2/o
yQ
is
thus the
state.
initial
The behavior
disturbance of the system from the equilibrium

of the
is illustrated in Fig. 1.1

system for t >
It is
for both positive and negative k.
>
seen that f or k
the magnitude of y
Then, as the
decreases with time.
time
increases
indefinitely,
Therefore, for k
of
>
0,
y*Q.
the disturbance
the system will eventually disapThe system can then be said to
pear.
FlG L1
'
be
When
stable.
<
0,
the motion of
the system increases with time, and eventually the disturbance will
become very large no matter how small the initial displacement is: the
system
will
never return to the equilibrium state once disturbed.
Such
systems are thus unstable.
For systems
derivatives.
of higher order, the differential
The nth-order system has the
a? +
equation will have higher
differential equation
+*-<>
^g3+
For a physical system, the coefficients a n -i,

solution of Eq. (1.4) can be written as
ao are real.
Then the
vV**
sin (A*
(l-fi)
a,-, ft are real and are related to the coefficients

and the p -'s are the phase angles. The motion of the system is thus
stable only if all a/s are negative.
If one of them is positive, the disturbance will eventually diverge, and the system is thus unstable.
where
INTRODUCTION
From the above examples it is seen that the crucial question to ask
about the behavior of a linear system of constant coefficients is the
Needless to say, the usual aim of an engineering
The question of stability can be answered, however,
design is stability.
once the coefficients of the differential equation are specified. In the
question of stability.
case of the simple first-order system specified by Eq.

information that matters is the sign of the coefficient k.
(1.1),
the only
Linear Systems of Variable Coefficients. If there is a variable

parameter in the system under study, the stationary, or equilibrium,
state of the system can be changed by changing this parameter.
It is
1.2
natural, then, to expect the coefficients of the linear differential equation
describing the system to be also functions of this parameter. For

instance, the aerodynamic forces acting on an aircraft are functions of
the speed of the aircraft is changing owing

deceleration, the aerodynamic forces will change
the speed of the aircraft.

to
acceleration
or
If
accordingly while the inertial properties of the aircraft remain practically

the same. As a result, if we wish to calculate the disturbed motion of the
fundamental
aircraft from, say, horizontal flight, the
differential equa-
tion will be an equation with variable coefficients.
Let us return to the simple example of a
by Eq.
(1.1).
aircraft
and
first-order system, described

a function of the speed of the
the aircraft has a constant acceleration a, then k is a
If the spring
if
function of the velocity

written as
u =
constant k
Thus the
at.
^ + k(at)y
The
solution
of
this
equation
is
differential
equation can be
(1.6)
is
r^
f
(1.7)
If k is always positive then log

increases
and
as
time
log (y/yo) will be increasingly
always negative,
than
less
is
Therefore
y Q and eventually will vanish.
always
y
negative.
Thus the system is stable. If k is always negative, log (y/yo) will be
where
y$ is the initial disturbance.
is
increasingly positive with time.

large even
unstable.
if
the
initial
Then y
disturbance y Q
is
become very
The system is thus
will eventually
very small.
These characteristics of the linear system with variable

remain positive or negative are very similar to those of
coefficients that
systems with constant
coefficients.
however, when k has both positive and negative

values.
Let us assume k(a) to be first positive, then negative, but
If the first zero of k is denoted by ui = ati
finally positive again.
The interesting
case
is,
and the seeond zero denoted by u 2
ThenEq.
at*,
then according to our previous
unstable in the velocity range from u\ to u<i (Fig.

minimum value of y and i/ raax be the maximum value
concepts, the system

Let 2/ m iu be the
1.2).
dy.
is
(1.7) gives
(1.8)
and
(1.9)
2/o
Of primary engineering interest is the question: How large is t/ max ?

system cannot function properly? We note that
Is it so large that the
'k(at)
."Unstable".
Zone
FIG. 1.2
to answer this question the knowledge of two things, in addition to the

functional dependence of k upon u, is required. These are: How large
What is the magnitude of the initial disturbance
is the acceleration a?
2/o?
for
For any
any
fixed a,
2/
max is
proportional to y Q
fixed initial disturbance, the
maximum
But more important,

value of the deviation
can be greatly reduced by increasing the acceleration a, as shown by

This means that by going through the "unstable" zone
(1.9).
quickly, the undesirable effects can be minimized.
t/max
Eq.
Therefore, for the more general linear systems with variable coefficients,
the simple question of stability has no definite meaning. The more
meaningful question is to ask specifically whether under a definite criterion the system will behave satisfactorily with specified disturbances
and circumstances.
In our simple example of a first-order system, the

the specified disturbance is
2/ max
definite criterion of proper behavior is
and the specified circumstance is the acceleration a. Thus by going

from systems with constant coefficients to systems with variable coeffi2/o,
cients, the character of the
problem
is
already considerably modified.
INTRODUCTION
1.3
Nonlinear Systems. If the spring constant k of the simple firstorder system described by Eq. (1.1) is a function of the disturbance y
itself, then the differential equation is
(1.10)
f+/(2/)=0
where f(y)
The system
It is seen that the differential equation is nonlinear,
k(y)y.
by Eq. (1.10) is thus the simplest example of a

nonlinear system. The solution y(f) can be computed from the following
relation obtained by integrating Eq. (1.10):
described
where y Q
On
is again the initial disturbance.

the other hand, repeated differentiation of Eq. (1.10) gives
dt*
dy
dt
(1.12)
=
z
dy dt
if yi is a zero of the function f(y) and if f(y) is regular at yi so that

the derivatives of f(y] with respect to y are finite at yi, then from
Thus
all
Eqs. (1.10) and (1.12)
This means that y approaches yi asymptotically. In fact, if y Q > yi and

> 0, then y will become yi eventually. If y Q < yi, then f(y Q } < 0,
This pattern of behavior of #
and y will again become yi at t > <*>
/(yo)
repeated with other zeros of the/fe) (Fig. 1.3).

If the initial disturbance yo coincides with one of the zeros of f(y),
Thus the zeros
this value of y will be maintained with increasing time.
is
>
If df/dy
at a zero such as yi,
of f(y) are equilibrium positions.
small deviations from this equilibrium position will eventually disappear
and the system will finally return to the initial state. Thus the system
may be said to have stability for small disturbances at yi. If, however,
at a zero such as t/ 2 the slightest disturbance from this equidf/dy <
librium position will cause the system to move to the next equilibrium
,
positions yi or y*.
y% is thus
an unstable equilibrium
state.
We have seen that even for the very simple nonlinear system described
by Eq.
(1.10),
the behavior of the system
is
very "complicated.
The
6
system
it is
may have
Thus
both stability and instability.
for
such systems,
about stability; rather,

entirely senseless to ask a general question
each specific problem must be considered individually.
FIG. 1.3
1.4
Engineering Approximation.
It is
almost certain that any physi-
analyzed in great detail, is always nonlinear. We speak

of the system as linear only with the understanding that we mean the
system can be approximated sufficiently accurately by a linear system.
cal system,
if
Furthermore; sufficient accuracy means that the deviation from linearity

so small as to be unimportant for the specific problem considered.
is
Thus whether a system
is
linear or not can be determined only
clearly defined circumstances.
The same can be
There
is
no general absolute
under
criterion.
said for the classification of linear systems into systems
of constant coefficients
and systems with variable
coefficients.
Take our
simple examples described by Eqs. (1.1) and (1.6). If the acceleration a

is very small, i.e., flight at almost constant speed, Eq, (L8) shows that
much smaller than the initial disturbance yo, and this
2/mia will be very
occur at a very large value of t. The behavior of the

finite time interval is thus very similar to a system
Therefore the system of variable
described by Eq. (1.1) with positive k.
coefficients can be sufficiently accurately approximated by a system with
value of
t/mia
will
system within a
constant coefficients under certain circumstances.
Needless to say, linear systems with constant coefficients are the

It is fortunate that a very large number of engineering
easiest to study.
systems
"
tion
is
falls
into this classification
made.
This
is
the reason
when the
why
"
engineering approxima-
this particular field of stability
and control theory is the most developed one. In fact, the present
theory of servomechanism deals almost exclusively with such systems.
We shall therefore begin with linear systems of constant coefficients.
CHAPTER
METHOD OF LAPLACE TRANSFORM

For
time
is
and with
method using Laplace transforms
linear differential equations with constant coefficients
as the independent variable, the
particularly useful in finding the solution.
Of course the problem can
be solved by a number of other methods; but the Laplace-transform

method appeals especially to the engineering scientist in that it reduces
all
problems to a uniform
The procedure
basis.
standardized, and a general approach
is
of solution is
possible.
practice of the Laplace transform are discussed in

not the purpose of the present chapter to do this.
then
The theory and
many
texts.
It is
The purpose here
is
rather to give for easy reference a summary of results which are useful
For details and proofs, the
to our discussion in the subsequent chapters.
reader should consult the texts cited.
of the
Laplace Transform and Inversion Formula. If y(t) is a function

time variable t defined for t > 0, the Laplace transform Y(s) of
y(t) is
defined as 2
2.1
<r* V
where
s is
dt
(2.1)
a complex variable having a positive real part, (Sis > 0. For

continuation.
s, the function Y(s) is defined by the analytic
other values of
The dimension
dimension of y multiplied by time.

the
Y(s)
known,
original function for which Y(s) is the Laplace
transform can be obtained in all cases by the inversion formula:
When
of Y(s) is the
is
y(t)
See for instance H.
S.
=,
&M Jy-i*
Carslaw and
e?Y(8)d8
J. C. Jaeger,
Mathematics," Oxford University Press,
New York,
(2.2)
"Operational Methods in Applied

1941; or K. V. Churchill, "Modern
Operational Methods in Engineering," McGraw-Hill Book Company, Inc., New York,

1944.
For a more complete theory, one should consult G. Doetsch, "Theorie und
der Laplace-Transformation," Verlag Julius Springer, Berlin, 1937; or
D. V. Widder, "The Laplace Transform/' Princeton University Press, Princeton, N. X,
Anwendung
1946.
2
Throughout
of quantities
this book, capital letters will
denoted by lower-case
letters.
be used to denote the Laplace transform
where 7
is
a constant greater than the real part of
The actual evaluation
the singularities of
all
can be done by properly deforming

7(s).
the path of integration according to the character of F(s).
2.2
Application to Linear Equations with Constant Coefficients.
of y(f)
Since the Laplace transform is defined as an operation on a function

t > 0, the method is particularly adapted to initial-value
the initial state of the system and the forcing function
Given
problems:
defined for
"
motion" of the system for t > 0. Let us consider

a f or the derivaan ftth-order system, with coefficients a nj a n -i,
Then the
or
and
a
forcing function, x(t).
tives,
nonhomogeneous term,
for
>
0, find
the
differential equation is
(2.3)
The
initial conditions are
usually specified as
(2.4)
The differential equation (2.3) together with the condition

mines uniquely the behavior of the system for t > 0.
To
solve the problem
sides of
Eq.
(2.3)
by
e~~
by the Laplace transformation, we multiply both

Then
and integrate from t =
to t =
s*
dt
And by
(2.4) deter-
Y(s)
partial integration,
dt
= -2
H<"
dt
-2/0
+ sF(s)
(2.5)
and
e
-3t
Therefore,
as X(s),
if
the Laplace transform of the forcing function
x(t) is
written
i.e.,
Z(s)
"
er*x(f) dt
(2.6)

then Eq.
together with the initial conditions (2.4) can be written as
(2.3)
Z()
Hence
if
we
define the polynomials D(s)
D(s)
a n sn
and Nn(s) as
dn^s"'
(2.7)
ais
(2.8)
and
ATo(s)
a ny Q s n
-1
+
+
(a ny^
then the solution of Eq.
"- 2
+
+ an_i2/r +
a n-i2/o)s n
- 1}
2)
=*
Ti1
that the
is
o2/o)
(2.9)
term
first
*
(210)
of the solution given
initial conditions
and
D(s)
D(s)
We note
'
(2.7) is
7(s)
v '
on the
'
'
through Eq. (2.9).

thus a lower order than D(s).
by Eq.
(2.10)
at
most
(s)
./Vo($)
is
will
vanish
depends
of order
if all
the
values specified by Eq. (2.4) vanish. In that case, Y(s) is given

the
second term alone. The second term depends upon the forcing
by
function.
Therefore the first term, No(s)/D(s), can be called the compleinitial
mentary function, and the second term X(s)/D(s), the particular inteThe actual solution y(t) can be obtained from Y(s) of Eq. (2.10)
gral.
by applying the inversion formula
of Eq. (2.2).
"Dictionary" of Laplace Transforms. The forcing function x(f)
is often of a character such that X(s) is the ratio of two polynomials in s.
Then the complete solution Y(s) given by Eq. (2.10) is also the ratio
2.3
of two polynomials of s.
Therefore the expressions for Y(s) can be
broken down into a number of simple fractions. Each of the fractions
can be inverted by the inversion formula,
or,
better yet, the original
TABLE 2.1
DICTIONARY OF LAPLACE TRANSFORMS
F
y(t)
I
l/s
l/s
l/s
a/(s
a/(s
s/(s
s/0
s/(s
t"- l /T(n)
l/(s
+a
+a
2
)
at
sin at
cos at
a2)
sinh at
a2 )
cosh at
+a
2 2
+a
2 2
)
sin at
(sin at
at cos at)
10
functions of
functions of
is
2
t
be found by the use
their Laplace transforms.
given in Table 2.1.

2.4 Sinusoidal Forcing Function.
down
can be broken
nomial D(s) are
where
a "dictionary," a
'can
and
JD'(s) is
of
say
simple
The ratio of polynomials N Q (s)/D(s}

If
into partial fractions.
all different,
list of
A much abridged dictionary
$1, s 2 ,
.
,
the roots of the poly-
*, then
the derivative of D(s) with respect to
s.
By
"interpret-
ing" the sum term by term according to our dictionary, the part y c (t)
of the solution due to the initial conditions, or the complementary
function
is
(2 - n)
In general, the roots s r of D(s) are complex. For physical systems, the
Eq. (2.8), are real; then the sr 's have complex conjugate pairs.
a's in Z)($),
But
s/s have negative real parts, then y e (t) will decrease exponenwith respect to time, and eventually y c (f) > 0. Then the system
if all
tially
is
stable.
If
the forcing function
sinusoidal
x(() is
x(t)
where x^ is the amplitude and
is
co
it
can be written as
aw*""
(2.12)
the circular frequency.
Then, accord-
ing to the dictionary,
la
Therefore the second term of Eq. (2.10)
is
now
(s
We can generalize this to include systems determined by a set of simultaneous equations by putting another polynomial N(s) of order lower
than n in the numerator. Then the Laplace transform Ft-(s) of the
particular integral
When
N(s)
Eq. (2.10).
1,
is
the problem is reduced to the simpler one specified by

the partial-fraction rule can be applied again. But the
Now
polynomial in the denominator

$2,
.
,
and
io).
Thus
is
now
(s
ico)JD(s),
and the roots are
Si,
__ __
^ _ ^ ]D M
>
11
-14
(Tir^J
Therefore the particular integral yt (t) due to a sinusoidal forcing function

of
form
(2.12) is
V
Z^
For stable systems,
all
part of 2/iOO vanishes as
and the
tion,
V
'
(2.15)
s/s have negative real parts.

t
>
<*>
The remaining part
Thus the second

is
the steady soluis thus
ratio of steady solution to the forcing function
simply given by
This equation gives a very direct way of computing the steady-state

solution under a sinusoidal forcing function.
When
the frequency co of the forcing function decreases to zero, the

is reduced to a constant, non varying with respect to
forcing function
time.
when #
at s
Equation (2.16) then indicates that F(0) is the ratio of y to x

This is the physical meaning of the value of F(s)
is a constant.
We shall use this interpretation frequently in
0.
2.5
to Unit
Response
be a continuous function.
instant
Impulse.
may
It
0, i.e.,
x(f)
x(t)
our discussions.
The
forcing function x(t) need not

be a unit impulse applied at the time
=0
for
5*
for
and
Then the Laplace transform X(s) of the forcing function is simply equal
The Laplace transform of the response to the unit impulse of the
to 1.
generalized system as described
by Eq.
Y<(s)
(2.13) is
1
simply
F(s)
(2.17)
solution y(t) due to this unit impulse is usually denoted as h(t).

According to our inversion formula, Eq. (2.2),
The
h(f)
When
the system
eF(s) ds
-.
is stable,
have negative real parts.

imaginary axis of
taken as the path
7 in Eq. (2.18) can be set equal to zero.
the roots
sr all
singularities of F(s) all lie to the left of the

the complex s plane. Then the imaginary axis can be
Then the
of integration for
the
h(t),
(2.18)
ZTTt J-y-ioo
that
is,
CHAPTER
AND TRANSFER FUNCTION
INPUT, OUTPUT,
We have seen in the previous chapter that by the application of the

Laplace transform, the behavior of a linear system with constant coefficients is made to depend essentially on the polynomial D(s) given by
Eq. (2.8), formed out of the coefficients of the differential equation.
}
Even in the generalized case, if the initial values of y(f) and the initial
derivatives necessary to specify the problem are all zero, the behavior of
the system is completely determined by the ratio N(s)/D(s) of two polyThis ratio is denoted as F(s). If the Laplace transform of the
nomials.
forcing function
integral
is
F -(s)
t
is
X(s) and the Laplace transform of the particular
Eq. (2.13) gives
7<()
F(s)X(s)
(3.1)
This equation can be considered as an operator equation: X(s) when

operated by F(s) gives F;(s), or F(s) transfers X(s) into F,-(s). There-
X(s) is the Laplace transform

the
Yi(s)
Laplace transform of the output yi(t).
In order to specify the fact that yi(t) implies the particular integral only,
without the complementary function introduced by the initial conditions,
fore F(s)
is
called the transfer function.
of the input x(t),
yi(f) is
y c (f)
is
and
is
Then the complementary function
called the output due to input.
called the output due to initial conditions.
The advantage of the Laplace transform method is thus to reduce a

problem in differential equations to one of algebraic operation. The step
of going from Y(s) to y(t) is seldom necessary, because the behavior of
y(t)
is
fully
Thus
determined by F(s).
it is
possible to translate the
engineering requirements on y(t) to a set of requirements on Y(s) or,

with the input characteristics specified, to a set of requirements on F(s)j
the transfer function. The study and design of a system by means of
the transfer function are the fundamental technique in servomechanism
engineering.
In this chapter,
we
shall
As a
first
expound
this technique
by a
series of examples.
3.1
First-order Systems.
example, consider a cantilever
spring which has a dashpot at one end and a sliding movement on the
other end (Fig. 3.1).
The position of the end with the dashpot is denoted
12
INPUT, OUTPUT,
by
y(t),
y(f) is
13
the position of the sliding end
by x(t). Because of the dashpot,

but
behind
The problem is to study
x(t)
lags
x(t).
when the sliding end is made to describe a specified
not equal to
the output y(t)

motion. x(f) is thus the input.
Let k be the spring constant and
.Slide
be the damping
coefficient of the
Spring
dashpot. Then if the motion is

slow enough for the inertial forces
to be neglected, the equilibrium of
forces requires
'/////////////////////////////777777s
rM
If
we
4- Mil
ri
FIG. 3.1
specify a characteristic time r\ as
_
Then the equation
initial
condition
(3.2)
motion can be written as
of
dy
The
is
(3.3)
simply
=
(3.4)
2/o
By multiplying
Eq.
(3.3)
by
e~ st
and integrating from
to
=
,
we
have the transformed equation
Therefore
(3.5)
Hence the output due
to input
is
given by
1
(3.6)
and the output due
to the initial condition

Tr
r.W /
The
transfer function F(s)
is
is
given by
7"l2/0
^FTT
(3.7)
thus
(3.8)
14
Equation
This simple
can be represented graphically as shown in Fig. 3.2.

visual aid is very helpful in picturing the situation and is
(3.6)
generally called the block diagram.
x(t)
*(*)
FIG. 3.3
FIG. 3.2
Let us investigate the output for a few special cases of the input
Consider
Fig. 3.3.
first
the case
when
x(f).
the unit step function 1(0 shown in
x(t) is
Then
e~8t dt
= s
and
1
Thus the output due to input
lfr(0
The output due
according to our dictionary, Table
is,
(1
to the initial condition
y e (t)
2/
<r"'0
Eq.
is,
2.1,
(3.9)
(3.7),
e-'/n
(3.10)
These output characteristics are shown in Fig. 3.4. Thus the output
due to the initial condition is a pure
subsidence with the characteristic time
The output due to input is an exponential approach to the asymptote,

again with the characteristic time T\.
TI.
In fact at t
r ly the output
yi(t)
reaches
63 per cent of the asymptotic final value.

The error signal e(t), defined as the
difference
output
between the input x(f) and the
yi(t) is,
e(t)
x(t)
for the case
under study,
(3.11)
Therefore the error signal vanishes as

>.
Consider
now another example
oidal, or
x(t)
00.
of the input.
Let the input be sinus-

where x m
is
the amplitude and w
is
15
Then
the frequency.
(3.12)
The output due to the initial condition

put due to input is given by
is
the same as before.
Therefore, according to our dictionary, the output
first
term
is
-f-
yi(t) is
1 -f- ictfTi
ZCOTi
a pure subsidence, and the second term
is
the steady-
state output.
Thus
This
agreement with our general result given in Eq.
in full
is
out-
The
The
(2.16).
Since
1
(3.13)
the steady-state output can be expressed as
1+
wVJ
Therefore the amplitude of 'the steady-state output is reduced by the

i*1
1/A/l
comparison with the input, and the phase of
+ ^M
factor
the output lags behind the input by the
amount
tan"" 1
cori.
For low-
frequency inputs,
r iw
That
is,
the amplitude
the characteristic time
is
(3.14)
not modified, but there is a time lag equal to

the transfer function. For high-frequency
TI of
inputs,
^*-^^l
dy
Then the amplitude

by
7T/2.
3.2
F(s)
is
reduced by the factor
These characteristics
of the
T!CO
1/cori,
and the phase lags
output are
Representations of the Transfer Function.
is
ratio of
a function of the complex variable
two polynomials
in
s,
s.
shown
in Fig. 3.5.
The transfer function
Since
the function F(s)
(3.15)
is
it is
generally the
determined up to a
16
constant by the zeros and the poles of F(s). This constant can be fixed
by knowing the value of F(s) at any particular s. The most convenient
In fact,
s is the origin.
\F(Q)\
has a physical meaning;
it is
(3-16)
the ratio of output to input with a constant,

is actually called the gain of the system.
non-time-varying input.
Therefore the transfer function F(s) is uniquely determined by the gain,
the zeros, and the poles. This is one possible representation of the
For example, the simple transfer function specified
transfer function.
by Eq.
(3.8)
has a gain of unity, a simple pole at -1/ri, and no zero.
x(tf
JJ
steady
Low Frequency
High Frequency
(a)
(b)
FIG. 3.5
imaginary parts of F(s) along the imaginary axis

plane are given, then by the principle of analytical continuation,
Therefore another possible
essentially determined for any s.
If loth the real
of the s
F(s)
is
and
the
is the complex function F(iu), where co is real.

For physical systems, the coefficients in the numerator polynomial N(s)
and the denominator polynomial D($) of F(s) are all real Then if we
representation of F(s)
F by
denote the complex conjugate of

E7/
F\
103)
==
F,
/O
~W7~"'
"
N
\1(JO)
Therefore for physical systems, knowledge of

sufficient for the determination of F for any s.
(o.it)
F(io>) f or
>
will
be
But we know from Eq,
steady output to the sinusoidal input

values of co is called the frequency response
(2.16) that F(ioi) is the ratio of

of
frequency
of the system.
F(iu) for
all
Therefore the frequency response
tion of the transfer function.

first-order
system
is
given
is
another representaof our simple
The frequency response
by Eq.
(3.13),
INPUT, OUTPUT,
One way
and
to present the frequency response
called the
is
the argument be
Bode diagram.
17
was devised by H. W. Bode

and
.F(ico) be
Let the magnitude of
6, i.e.,
F(iu)
= Me i8
(3.18)
The Bode diagram consists of the plots of log

and 6 against log
but not for 8 will be made meanThe choice of a logarithmic scale for
ingful
by
For our simple system
later discussions.
of
Eq.
(3.13),
U =
(3.19)
and
where u
system
is
is
The Bode diagram of this

The behavior of the frequency response
the dimensionless frequency.
shown
in Fig. 3.6.
Iog 10 u
FIG. 3.6
and at high frequencies is that already indicated by Eqs. (3.14)

As
against logio u has the
*>, the graph of logio
1.
For small values of u, the slope is nearly 0. Therefore the
slope
u diagram for a first-order system can be approximated by two
at low
and
(3.15).
u*
M~
straight lines.
In acoustical and electrical literature,
it is
customary to plot 20
logio
instead of logio
in order to convert the amplitude units into decibels.
A doubling in frequency is called an octave, and thus the region of the
plot in Fig. 3.6 where the logio
M curve has a slope of
1 is
described
We note also
6.02 db per octave.
20 logio 2 =
at
that on the same plot, the approximate logio
line goes through
u = 1, that is, at co = I/TI. Therefore we can measure the frequency
as a region of slope
response of a first-order system and plot the measurement as indicated.

The characteristic time n of the system can easily be estimated by
noting the frequency at which a straight-line approximation for large

values of co crosses the horizontal axis.
Another way, to present the frequency response was devised by H.
18
Nyquist and
F(zw) or
l/F(i)
parameter
1/(1
+i) =
for w-~
directly plotted in the
is
curve
of the
system, F(ico)
is
complex F or 1/F plane. The

For a simple first-order
the frequency w.
= 0, going through
1 for
un = u = 1, and ending at the origin
is much simpler: 1/^ = 1 + fwri, and
a semicircle, starting at
is
(l/-v/2)(l
f)
at
The 1/F diagram
oo.
Here the complex quantity
called the Nyquist diagram.
is
JL
FIG. 3.7
thus the diagram
is
simply a straight line parallel to the imaginary
These Nyquist diagrams for the

3.3
Examples
of First-order
first-order
Systems.
system are shown in Fig.

There are many elements
complex system that can be approximated by a

tion.
We
shall briefly discuss a
number
of
axis.
3.7.
of
first-order transfer func-
examples of such elements
in this section, together with the proper diagrams for their frequency
responses.
Integrator.
An
an input voltage
electric
v follows
motor whose speed
proportional to
d(j>/dt is
the equation
J = Kv
where
K is
a scale factor.
(3-20)
Thus the angular

the motor
position
is
4>
of the rotor of
proportional to
vdt
v(t)
This relation is represented by the

block diagram in Fig. 3.8 with the
Laplace transforms 7(s) and *(s) for v and
This transfer function
= K/s is the limiting case of the function
oo
F(s)
I/ (TIB
1) when T\
FIG. 3.8
<j>.
and
represented by a simple pole at the origin. In order to consider

the constant
as the gain of the transfer function
F(s), we have to
the
definition
first introduced in the
modify
previous section. The defiit is
nition there given
is
suitable for transfer functions

having
no zero or
INPUT, OUTPUT,
The gain
polecat the origin.
K of an
integrating system,
whose transfer function F(s) has a simple pole
i.e.,
at the origin s
19
a system
0, should
be defined as
K~lim\sF(s)\
(3.21)
s->0
is
Therefore, according to Eq. (3.18),

,, = K
M
7T
The Bode diagram is thus that shown

gram is that shown in Fig, 3.10.
(3,22)
CO
in Fig. 3.9,
and the Nyquist dia-
log
10^
FIG. 3.9
,,
/-
K
FIG. 3.10
A rate gyro gives a voltage output

Differentiator.
the angular velocity d$/dt of the precession axis, i.e.,
proportional to
-*
where
is
the factor of proportionality.
the preceding one.

origin.
The
Thus the gain
of
This case
= Ks
a differentiating system,
is
the inverse of
has a zero at the
i.e.,
transfer function F(s) as a simple zero at the origin 5
a system whose
= 0, should be
20
defined as
K
The block diagram is shown
(3.23)
Bode diagram in Fig. 3.12,

and the Nyquist diagram in Fig. 3.13.
in Fig. 3.11, the
Simple
V(s)
capacitance
C and
if
there
is
JR
Lag
Network.
Consider
the resistance] and capacitance network of Fig. 3.14. If j is the current
v (t]
flowing in the resistance R and the

in the capacitance at t = 0, then
no charge
dt
dt
By
st
multiplying these equations by e~ and integrating from
Iog 10
to
FIG. 3.12
oo
y
we have
Therefore
(3.24)
INPUT, OUTPUT,
Hence the
transfer function of this
lever spring with dashpot,
RC
circuit is the
This circuit
is
same as the
is TI = RC.
canti-
The
and the characteristic time
Bode diagram and the Nyquist diagram

3.7.
21
are thus given
by
and
Figs. 3.6
frequently used in order to introduce a phase lag
into a system.
wsA/Vv
vvww
-|
FIG. 3.14
Lead Network.
_ -*-;
j-
The
-T-
t'
A/VAV
1
ri_^^-T
jp
fr
FIG. 3.15
more complex
circuit is that
shown
in Fig. 3.15.
controlling equations are
and
The corresponding transformed equations
=:
J\ ~h
are
t/a
/2
-^r
and
K 1 ==
R\Jl
-f"
Therefore
w/
Fi(s)
Hence the gain
and
If
it is
(fix
R,}
is
necessarily less than unity
we introduce
and
is
generally between 0.1 and
1.
the symbol wi as
4)1
-RI
R%
0i \
/0
(3 26)
'
then the transfer function can be written as
Therefore the transfer function has a zero at
rwi
and a pole at
wi.
22
The frequency
response
is
then
x
rP/,
Utt)
= rm
COi
If
we introduce
+
+
:
fra
r-
(3.28)
CO
the nondimensional frequency w as
(3.29)
then
)
(3.30)
Hence
1
+
=
(u'/r)
logio
/r)
and
'l
The Bode diagram has thus a
log lo
certain
symmetry with
respect to
u =
1,
FIG. 3.16
as
shown
is
equal to
in Fig. 3.16.
tan- 1
The maximum value
of 6 occurs at
-p -
tan- 1
2 tan- 1
and
(3.31)
Therefore this circuit gives a considerable phase lead over a band of

= r.
For very large
If = 1. For very small co,
frequencies.
Restricted
Lag Network.
following transfer function:
The
RC
circuit
shown
in Fig. 3.17 has the

The gain
and
of the
is
system
thus unity.
If
23
we introduce the parameters
defined as
1
(3.32)
vi
then the transfer function

_i_
F(s]
U2
;*!
is
/w,.,^
FIG. 3.17
(3.33)
comparing this equation with Eq. (3.28) for the lead network, we
see that the two circuits have transfer functions that are reciprocal to
each other. In fact the frequency response in the present case can be
By
written as
where
the dimensionless frequency,
u,
is
(3.34)
Thus
if
The corresponding Bode diagram

is
is
thus that shown in Fig. 3.18.
(3 35)
-
There
The maximum phase
a phase lag for a range of the frequency.
log 10
"
tan
lag
Iog 10 w
l+-s
6/1
F(icJ)
-r^'^vF^
FIG. 3.18
occurs at
1,
or
co
\fr
on,
and
its
magnitude is given by Eq. (3.31).

Let I be the moment of
Simplified Rolling Motion of an Airplane.
inertia of the airplane about its longitudinal axis,
<t>
the
roll angle,
Lp
the aerodynamic damping due to roll, and kd the torque applied by

is thus
the aileron deflection 5. The equation for the roll angle
<t>
W+
I
"~di
24
Now
If
let
the
p =
roll
d<t>/dt
speed
is
be the
roll
zero at
then the transformed equation
0,
+ Lp)P(s) =
(Is
The
speed; then the above equation becomes
is
k A(s)
thus
is
(3 - 36)
The behavior of the system, as determined by the transfer function, is

thus similar to the cantilever spring with dashpot and the simple lag
If the damping is
network. Here the characteristic time n is I/L P
.
very small, then
TI
>
co
and the behavior
of the
system becomes that
of the simple integrator.
3.4
end.
Let us return to the cantilever spring
Second-order Systems.
with a dashpot (Fig.
The mass
tion of motion
with the
is
m to the dashpot
m d y/dt and the equa-
But now we attach a mass
3.1).
will introduce
an
inertial force
now
initial conditions
2/(0)
2/o
(3.37)
The
differential equation of
motion can be rewritten
in a
more convenient
form by introducing the following parameters:

k
coo is
thus the natural frequency of the mass-spring combination when

is absent,
f is the ratio of actual damping to critical damp-
the dashpot
ing.
The meaning of this nondimensional parameter

Then the differential equation becomes
will
be
made
clear
presently.
+
Equation
(3.39) together
2f
;|
with the
+ &
(3.39)
&5
initial conditions of
Eq.
(3.37)
can be
converted into the following relation in terms of Laplace transforms:

The output due
to initial conditions
is
then given by
transfer function
is
(3.40)
2f
The
then
"
X(s)
The
(s/coo)
(3.41)
+
=
transfer function thus has a gain
two simple poles
25
and no
zeros.
It has
at
(3.42)
When
the damping coefficient of the dashpot is smaller than the critical

damping, the value of f will be less than unity. In that case, the poles
and $2 will be complex conjugates, having

X and v respectively:
$1
real
and imaginary parts
where the
SI/COQ or
last expression is possible
s 2 /coo is
one.
Since X
it is
is
= ^-
Now
lt
because the absolute value of either
cos
sin vt
sinusoidal function.
damped
is
critical value,
shown
(3.43)
of the transfer function are
d* sin
vt
vt
(3.44)
a negative number, the output
the output
meaning
<l
For positive damping, X is a negative number.

due to initial conditions can be easily determined
The output y c (t)

from Eq. (3.40). Thus, for f 2 < 1, the poles
given by Eq. (3.43), and we have
y c (t)
a pure subsidence.
there
is
no
is damped, but nevertheless

on
the
other hand, f 2 > 1, then
If,
Thus for dampings greater than the
oscillation in the
This
is
the
the unit step function
l(t)
output y c (t).
of critical
let
damping.
us assume that the input
in Fig. 3.3.
Then X(s) =
"w
The output
yi(f)
due to input
yi(t)
x(t) is
1/s,
and for f 2
s[(s-\y
is
<
1,
then
cos
vt
(-^J
sin vt
xt
(3.45)
26
When
>
1,
the output
where
yi(t)
Si
are
and s 2
shown
is
Vr 2
are given
not oscillatory and
if
calculated as
is
(3.46)
by Eq.
(3.42).
Such behaviors
in Fig. 3.19 for various values of the
of the
damping
output
ratio
f.
seen that for a quick approach to the asymptotic value, f should

not be too large. On the other hand, if f is too small, there will be rather
It
is
1.6
1.2
0.8
0.4
10
FIG. 3.19
Here an engineering compromise

persistent oscillations with high peaks.
has to be made, and the usual practice is to select a damping ratio f
between 0.4 and
1.
a sinusoidal oscillation with amplitude x m and frequency

w as specified by Eq. (3.11), then
If -the
input
is
s)== ^~
_~VF(
s
tco
-
'
Therefore the output ^(Q due to input
ius*
is,
for f
2f coos
<
co
1,
coo
(3-47)
where X and
v are given by Eq, (3.43).

Since X is negative for positive
damping, the steady-state output is again simply given by the first
INPUT, OUTPUT,
-1
27
+1
lo
eioQ
FIG. 3.20
term of Eq.
Eq. (2.16).
(3.47).
This

Eq.
is
in agreement with our general result of
of
our second-order system
(3.41),
C\
rJF(*IW)
(
MpW
IVJLV
f:
r^T
is,
according to
28
Therefore
=
[2f(/
(3.48)
)]
and
tan*-The corresponding Bode diagram

of
M occur near
CO/OJQ
1,
where
is
shown
M ~ ^ and
in Fig. 3.20.
9
?r/2.
The maxima
As co/coo ^
,
INPUT, OUTPUT,
0-
TT
and
M~
engineer will
tical
l/(oj/co
M~
or log
2 log
(co/o>o).
then say that the slope for high frequency
29
The acous12.04 db
is
per octave.
The Nyquist diagram
for our second-order
system
is
shown
in Fig.
3.21.
Other physical systems can usually be approximated by a second-order

The hydraulic servo system is one example. A better
transfer function.
approximation to the behavior
of
the rate gyro discussed in Sec. 3.3
the transfer function
is
Ft*
f(S>
Ks
(/) +
2f
(/) +
This more accurate transfer function should be compared with that given
The transfer function for an accelerometer is
in Fig. 3.11.
+
The
electric
2r(*/o)
motor used as an integrator has a more accurate transfer
function
F( s)
This transfer function should be compared with the previous crude

approximation given in Fig. 3.8. All these transfer functions have a
second-order polynomial in the denominator. The constants in them
have meanings similar to those in the examples discussed previously.
3.5
Determination of Frequency Response.
In the discussions of the
previous sections, we have considered the problem of knowing the structural details of a system and of calculating the transfer function F(s)
and the frequency response F(ico) by elementary physical laws. This
procedure of determining the frequency response is thus theoretical, and

accuracy depends upon the accuracy of our knowledge of the system.
its
Very often in engineering
practice, our
knowledge
of the detailed struc-
ture of the system is incomplete, or if sufficiently complete, the system

is so complicated as to make the theoretical calculation of the frequency
response too lengthy to be practical. In such cases, it is often necessary

to determine the frequency response experimentally.
The simplest
method conceptually is to utilize the fact that the ratio of steady output
with a sinusoidal input of frequency
co
to the input is equal to the fre-
quency response F(i<a) as shown byEq. (2.16). The ratio of the amplitudes
of output and input is M, and the phase difference between output and
input
is 6.
Therefore this experimental method involves the determinaand phase differences for a number of frequencies
tion of amplitude ratios

co
in the desired range.
It has
been applied to problems varying from
30
such a relatively simple system as a fuel pump to as complicated a

2
The drawsystem as the longitudinal motion of a complete airplane.
of this method is the lengthy experimentation generally required
a
wide
for
range of frequencies. Sometimes it is also difficult to determine the phase difference between the output and the input.
back
more
instead
method is to excite all frequencies simultaneously

The best method for doing this is to use a
efficient
of individually.
x(t)
Or
2r
Triangular Pulse
Rectangular Pulse
FTG. 3.22
Then, according to Eq.
unit impulse as the input.
systems (7
(2.18), for stable
0),
*-s/-.
1
cos
where
(R
0F(ieo) sin ut]
and 6 denote the real and the imaginary parts,
last step, is possible
because
an input impulse excites
When
cot
of
all
Eq.
(3.17).
(3.49)
This
respectively.
Equation
(3.49)
shows that
the frequencies of the system uniformly.
the system to a unit impulse

can
be
computed as
frequency response
the response
dw
h(t) of
is
determined, the
"
f
iot
h(t)e'
dt
(3.50)
This integration can be carried out numerically for a number of frequencies.
Practically, however,
The more
it is difficult
to
make
the input be an impulse.
practical inputs are a single rectangular pulse
triangular pulse, as
shown
in Fig. 3.22.
Such inputs
and a
will
single
not excite
But if we make the length T of the pulse

the
ideal
uniform
excitation
can be approached. This method of
small,
excitation
has
been
to
determine the frequency response of
pulse
applied
all
frequencies uniformly.
an airplane by R, C. Seamans and his coworkers. 3

Himmel, D. Blivas, NACA TN 2109
Aeronaut Sci., 14, 493 (1947).
They have
H. Shames,
W.
E. C. Seamans, B. P. Blasingame, G. C. Clementson, /, Aeronaut,
(1950).
S.
C.
also
(1950),
F. MiUiken, J.
Sci., 17,
22
INPUT, OUTPUT,
31
from the
developed an approximate method for computing the JP(iw)
has
been
reduction
data
generalmeasured output y(f). The method of
1
by H. J. Curfman and R. A. Gardiner.
The
Elements.
from
3.6
systems studied
Composition of a System
much more
in
the
elements
are
in Sees. 3.1, 3.3, and 3.4
really only
in
be
to
found
stability and
necessary
complicated system generally
ized to arbitrary inputs
Take the example
control engineering.
The
plane.
is
of the rolling
usually in the
motion
form
of
of
an
an
air-
electric
"
"
This signal is the input to an amplifier and computer group,

a rather involved electric circuit and may contain vacuum tubes.
current.
which
signal
to move the aileron
is
The behavior
and computer
of the amplifier
is
determined by
its
transfer
and computer is then taken

function FI(S).
which
moves the aileron. The
servo
to
the
be
the
to
hydraulic
input
The output
of the amplifier
behavior of the hydraulic servo
is specified
by
the servo transfer function
7(5)
FIG. 3.23
is taken as the
Finally, the output of servo, the aileron motion,
The airplane
the
the
to
dynamics.
airplane
representing
system
input
of the airplane
The
F
function
transfer
8 ().
the
output
dynamics gives
F 2 (s).
connection
dynamics is the rolling motion. Here we have a series
between the various elements of the system from rolling signal to rolling
motion. If the rolling signal is denoted by x(t) and the roll angle by
then the Laplace transforms are related by
$({),
*<()
Fi(s)F*(s)Fi(8)X(s)
Therefore the over-all transfer function of the roll-control system is the

illustrates clearly the fact
product Fi(s)Ft(s)Fz(s). This example also
that a transfer function is generally dimensional: it is the ratio of two
The input, or roll signal, is an electric

quantities of different dimensions.
has the dimension of an angle.
current; and the output, or roll angle,
In general, then, if a system is composed of individual elements of
Fn (s) with gains Ki,
Fr (s)
transfer functions FI(S), F,(s),
.
K* and if tlie elements are in series (Fig. 3.23),

r
K*,
then the over-all transfer function F(*) is the product
.
F( 8 )
The gain
From
Fi(*)ft(a)
K of the system
K=
is
then
i# 2
Fr (s)
Fn (s)
&&
H.
(3.52)
transfer function F(s) has

Eq. (3.51), it is evident that the system
This fact
elements.
individual
the
of
of the zeros and poles
the totality
1
(3.51)
J.
Curfman and R. A. Gardiner,
NACA TR
984 (1950).
32
together with the gain K, computed by Eq. (3.52), completely determines

the transfer function F(s).

response of rth
quency
Me* 9 =
of the
element
system
is
re
ier
is
= Me
F(iw)
iQ
.
If
the fre-
then, according to Eq. (3.51),
(Jfi
Therefore
logic
logio
Mi
+ logio M +
'
'
logio
(3.53)
and
Equation
in the
0i
02
'
6r
6n
(3.53) gives the reason for the choice of a logarithmic scale
Bode diagram.
characteristics easier
This choice makes the work of finding the system

by requiring only simple addition of the ordinates
of individual diagrams.
Transfer Functions. The Laplace-transform

not
only to initial- value problems of linear ordinary
applicable
differential equations with constant coefficients, but also to linear partial
Transcendental
3.7
method
is
differential equations
variable
t,
with coefficients that are independent of the time
and with boundary conditions partially described as
initial-
By applying the Laplace transform to the original

differential
equation, the time variable t is removed, and in its place
partial
a parameter s appears. The resultant equation is a linear differential
value conditions in
t.
equation with respect to the remaining independent variables and can be

solved as such by using the remaining boundary conditions. The procedure involved here is evidently much more complicated than in the case of
ordinary differential equations discussed in Chap. 2. On the other hand,
any two specific quantities in the solution of the transformed equation
If one of them is conare compared, they still bear a linear relation.
if
sidered to be an input and the other an output, the ratio of output to
input is still a function of the parameter s and can still be considered as

the transfer function F(s). There is, however, one difference: the transIt is, in
fer function is no longer the quotient of two polynomials of s.
general, a transcendental function in
For instance, for two-dimensional
s.
flows,
W.
R. Sears 2 has calculated
the effects of a small vertical "gust" velocity v in the fluid on an airfoil
a stream moving horizontally with a uniform velocity U.

vary sinusoidally with respect to x, the horizontal coordinate, and
of chord c in
Let
1
See for instance H,
S.
Carslaw and
J.
Mathematics/' Oxford University Press,
C. Jaeger, "Operational Methods in Applied

1941; or R. V. Churchill, "Modern
New York,
Operational Methods in Engineering," McGraw-Hill

1944.
2
W, R.
Sears, /. Aeronaut. Sri., 8, 104 (1941).
Book Company,
Inc.,
New
York,
INPUT, OUTPUT,
i
33
(Fig. 3.24), so that
= aJJe<v-wn
v(x,t)
where am
is
the amplitude and
Sears has shown that the
lift
the
o>
"
(3.54)
frequency." For this gust velocity,

(7z, that is, the average lift force per
coefficient
unit area of the airfoil divided
by the "dynamic pressure" ?pU*
(p is
FIG. 3.24
the density of the
fluid), is
given by
Ci
2iram e i(at <p(k)
(3.55)
where
(3.56)
jfj
and
<p(k)
(3.57)
Ck] 4- /rVfel
The J's and K's in Eq. (3.57) are the Bessel functions of the first kind
and the modified Bessel functions of the second kind, respectively.
Therefore,
if
we take X(s)
as the Laplace transform of v(0,), the input,
and Y(s) as the Laplace transform

function F(s)
of Ci(t), the output,
then the transfer
is
and the frequency response
F(icc) is
(3.58)
The frequency response is thus a transcendental function.

The application of such concepts as the transcendental
and frequency response to the problem
has been demonstrated by J. Dugundji. 1
tion
J.
Dugundji,
J".
Aeronaut.
Sci., 19,
422 (1952).
transfer func-
of flutter of airplane
wings
CHAPTER
In this chapter we shall introduce the central concept of modern stabiland control engineering: the concept of feedback. We shall intro-
ity
duce this concept by discussing the simplest systems linear systems

with constant coefficients. We shall show how the feedback can greatly
increase the degree of accuracy in control and the rapidity of response to a
signal
We
shall
then explain the principles of designing such feedback
servomechanisms for
and
for optimum performance.

us consider the problem of controlling
Let
Concept
the rotational speed of a turboalternator. The primary purpose here
The most eleis to keep the speed very close to the normal value.
4.1
of
stability
Feedback.
mentary approach to this problem would be the so-called open-cycle

control, where we try to balance the torque generated by the steam
turbine and the torque absorbed by the alternator, the load torque.
This could be done by measuring the load and by opening the steam
However, it is to be expected that such balancing
cannot be perfect; there is always an error torque, x(f). This error
torque tends to accelerate the machine. If we denote by y(f) the speed
throttle accordingly.
deviation from the normal,
A (S)
by / the moment
of inertia of the rotating

parts of the machine, and by c the
damping due to the windage loss,
'
then the differential equation
(S)
e. 4.1
The block diagram
of this
system
is
shown
cy=x(f)
in Fig. 4.1.
is
(4.1)
It is seen that
the system is the familiar first-order system studied in the last chapter,
that the characteristic time of the system is 7/c, and that the ratio of the
Now
steady-state value of speed deviation to the error torque is 1/c.
/ is a very large quantity because of the heavy weight of the rotor of a
turboalternator, but c
windage
loss.
is
Hence the
a very small quantity because of the small

characteristic time is extremely long.
This
means that any speed deviation will persist and be difficult to remove.
Furthermore, for small speed deviation, the error torque has to be
extremely small, because of the large magnification factor 1/c. Need34
FEEDBACK SERVOMEGHANISM
less to say, this
is
system
of
35
keeping the turboalternator at constant speed
quite useless in practice.
Now consider the change in performance caused by changing the system

In the closed-cycle control, we
make the control torque depend upon the controlled variable. That is,
we cause the steam-throttle opening to depend not only on the load but
to the so-called closed-cycle control.
also
on the speed deviation
of proportionality
throttle
ky.
is
When
closed,
Let the second component have a factor

is too high, or y > 0, then the
the speed
and the accelerating torque
ky.
is
reduced by the amount
too low, the accelerating torque is increased

Thus the differential equation for y is now
the speed
by the amount
y.
When
k.
is
The only difference between Eqs. (4.1) and (4.2) is the replacement of c
by the sum c + k. Hence the characteristic time is now //(c + fc), and
the ratio of steady-state speed deviation to the error torque
,
is
l/(c
k).
Mixer
Output
Feedback
FIG. 4.2
Therefore, in comparison with the open-cycle control, we can greatly

reduce both the characteristic time and the speed error by making k
very much larger than c. But this can be accomplished quite easily,
because c is so small. Therefore the closed-cycle control can be designed
for quick response and for accurate control and thus achieve a great
improvement in performance.
The block diagram of the closed-cycle control can be drawn as Fig. 4.2,
retaining the intrinsic transfer function of the turboalternator shown in
In Fig. 4.2, we have introduced a convention in servomechaFig. 4.1.
nism engineering: addition or subtraction has to be specifically indicated
If at the junction of two links only a dot is
at the symbol for the mixer.
The quantity is only "measused, no addition or subtraction occurs.
ured." Thus the speed deviation y is measured at the output side and is
used to generate the control torque. It is seen from Fig. 4.2 that the
closed-cycle control involves a feedback link.
The whole system
is
thus
appropriately called a feedback servomechanism.

Although for the simple example analyzed above the advantage of a
36
feedback servomechanism can be shown by comparing the differential

equations (4.1) and (4.2), for more complicated systems the analysis can
be conveniently carried out only by the concept of transfer functions.

This method is expounded in the following sections.
4.2
Design Criteria of Feedback Servomechanisms. Let us consider
a general feedback servomechanism with arbitrary transfer functions FI(S)
and
2 ($),
similar to that represented
by
Fig. 4.2.
FI(S)
is
called the
and
Fs(s) the transfer function

Then the input X(s) and the output Y(s) are
transfer function of the forward circuit

of the feedback circuit.
related as follows:
Y(8)
By
we have
solving for F(s),
*( S '
l( S )
Jf(8)"l+Fi(8)F,(a)
where
is
(s)
(~\
ftW
(A
<n
( *' 6>
thus the system transfer function, or the output-input
ratio of the complete system.
be convenient for later discussions to indicate explicitly the

Thus we write
gains Ki and K% of the transfer functions FI(S) and F 2 (s)
It will
nondimensional; the dimension of the transfer

absorbed into the gain K. All information about the
It is evident that G(s) is
function
"
all
is
7
of the transfer function is contained in (?(s), that is, G(s)

and the poles of the transfer function. In the subsequent
zeros
the
gives
discussions, we shall usually think of the effect of the transfer function
structure'
on the performance of the system as the result of two separate influences:

the influence of the locations of the zeros and the poles of the transfer
function, i.e., the influence of G(s) and the influence of the magnitude
;
K.
This separation of effects is further justified by the fact

that the structure of the transfer function G(s) is controlled by the computer element of the amplifier- computer group in the composite system
of the gain
of F(s).
The
gain
is
controlled
by the
amplifier element of the amplifier-
Moreover, these two controls in the design can be

affected almost independently of each other.
Therefore (?($) and
can
computer group.
indeed be modified separately and can be considered as separate.

By using Eq. (4.4), Eq. (4.3) can be written as
y ()X(s)
F
*
Mw
-frW
The Laplace transform

by
E(s), is then
a 5)
(
of the error e(t) defined
by Eq.
(3.11),
if
'
denoted
FEEDBACK SEB.VOMECHANISM
EM
X(s)
Y(s)
J_
37
_1
For simple feedback servomechanisms, as shown in Fig. 4.3, the transfer

function of the feedback link F$(s) is simply unity; i.e., the output is only
X($)
'
FIG. 4.3
measured, not modified for the feedback control.
Then Eqs.
(4.6)
and
(4.7) simplify to
(4.7)
l+KG(s)
X(s)
and
(4.8)
KG(s)
Y(s)
The
first
requirement of a servomechanism
is stability.
This means
that the output y(t) should be damped, except possibly for steady sinusoidal motion.
Our analysis in Sec. 2.4 shows, however, that the condition
of stability is mathematically equivalent to the statement that F 8 (s)
should have no poles in the right-half s plane, where the real part of s
For the general feedback servomechanism, as shown by
is positive.
Eq.
the poles of
(4.6),
(s)
are zeros of
(4.9)
For the simple feedback servomechanism, as shown by Eq.

poles of
(s)
Therefore the
(a)
(4.8),
the
are zeros of
first
The function
design criterion of feedback servomechanisms

l/f*(s), given by Eqs. (4.9)
and
is
(4.10), should not have
zeros in the right-half s plane.
The second requirement
of a servomechanism is quick response.

If
a pole of F,(s), then the analysis in Sec. 2.4 shows that the output
has the component e Srt
The quickness of response is thus determined
by the magnitude of sr The larger the magnitude of srj the shorter the
s r is
time scale and thus the quicker the response.

design criterion of a feedback servomechanism is
Therefore the second
38
The
(6)
zeros of
large
l/F
(s),
given by Eqs. (4.9)
magnitude and
and
lie sufficiently to the left
(4,10), should all be of
of the imaginary axis
of the s plane.
If the
feedback servomechanism
is
designed for controlling the output
to follow the input signal, the steady-state output after the removal
of transients should
be made as close as possible to the input.
Therefore,
the third requirement that the
ratio S(0)/F(0) between the steady-state error and the steady-state
output should be as small as possible. This condition can be translated
for such "positional" controls, there
is
into a condition on the gains of the transfer functions
and
(4.8).
For positional
,(c)
by using Eqs.
accuracy of control requires for general servo-
control,
mechanisms, Eq.
(4.7),
]~0
^-[l-tf
Xti
2
and for simple servomechanisms, Eq.
(4.11)
(4.9),
J5T1
The
conditions
servomechanisms.
and
(6)
We
(c)
(4.6)
Thus
(a),
and
(6),
In practice,
as fully as desired,
(4.12)
(c)
are the design criteria of feedback
it is
usually difficult to satisfy conditions
and a compromise has generally been made.
shall see this in the following sections.
4.3
Method
of Nyquist.
tions are usually ratios of
section
is
Since, as stated before, the transfer func-
two polynomials
in
s,
criterion (a) of the last
generally equivalent to specifying the nonexistence of roots
with positive real parts for a polynomial. This is a classic question and
answered by E. J. Routh using the so-called Routh inequalities,- involv-
is
ing the coefficients of the polynomial under investigation. This method,

however, is not favored by control engineers, because of the obscure manner
Routh inequalities with changes in the coefficients.

Engineers prefer a method of analysis which uses the transfer functions
written in Eqs, (4.9) and (4.10) directly without further modification;
of the variation of the
because these transfer functions are the immediate information possessed
and are understood "physically" by the engineer.

Such a method was devised by H. Nyquist. The Nyquist method is
based upon a theorem due to Cauchy for an analytical function /($) where
,
s is a
complex variable:
If f(s) has n zeros and
poles within a closed path C, then as s travels
See for instance Whittaker and Watson, "Modern Analysis," Sec. 6.31, p. 119,
Cambridge-Macmillan, 1943.
39
along C once in a clockwise direction, the vector f(s) carries out

wise revolutions about the origin.
To apply
C
this
m clock-
very powerful theorem to our problem, we have chosen
to enclose the whole right-half s plane, where zeros with

real
Such a path is shown in Fig. 4.4 and conparts would lie.
positive
sists of the imaginary axis and a semicircle to the right with the radius
the path
R-^
Take
co.
the simpler case, the case of a simple feedback

note from Eq. (4.10) that the poles of l/F,(s)
Let the number of zeros of
first
We
servomechanism.
are zeros of {?($).
G(s)
the right-half
in
plane be m.
Then
l/F s (s) has m poles in C. Therefore, in order

for l/F 8 (s) to have no zeros in the right-half s
plane, l/F 8 (s) has to carry out m counterclockwise revolutions around the origin when s describes the contour
But from Eq.
of Fig. 4.4
(4.10), it is easily
with
oo
seen that this
is
plane
equivalent to requiring l/KG(s) to carry out

counterclockwise revolutions around the point
1
FIG. 4.4
But since K is a constant, the above criterion is the same as requiring
l/(?(s) to
carry out
m counterclockwise revolutions around the point K.

s plane, or m = 0,
Needless to say, when G(s) has no zero in the right-half

then the Nyquist stability criterion requires the vector
K.
revolution around the point
Let us illustrate the application of the
1/ff (s) to
make no
method by taking the simple
transfer function
(4.13)
Then
1/0(3)
First consider the part of the
path
of Fig. 4.4 along the
imaginary
axis,
where
At
As
+00
i.
Therefore as
l/G(iu) >
the vector l/(r(iw) increases in magnitude, and
co increases from
to
its phase angle increases from ?r/2 to 3x/2.
For negative co, the curve
co
0,
l/(7(i)
iO.
co
by the end point of the vector l/G(iw) is simply the reflection

the curve for positive co about the real axis, as required by Eq. (3.17).
Thus as s traces the imaginary axis, l/(?(zco) traces the curve oboe shown
traced
of
in Fig. 4.5.
As
5 traces the large semicircle
s rotates
from io to
but three times as
fast.
shown in Fig.
4.4, l/G(s)
~s
3
.
Then
as
clockwise, l/G(s) will also rotate clockwise,
This part of !/(?()
is
thus represented by the
40
curve c to a in Fig,
4,5.
From
this figure, it is seen that
if
K-
Ki
as
indicated in the figure, then the vector l/G(s) will make no net revolutions
Since the function ff(s), given by Eq. (4.13),
around the point
/.
has no zero, this means that the feedback system will be stable. If
= Ka as indicated, then the vector l/G(s) will make two net clock-
wise revolutions around the point

Ku. Therefore, with this larger
In fact, there
value of K, the feedback servomechanism will be unstable.
be two poles of Fs (s) with positive real parts. The transition point
from stability to instability is the point 6. Stable values of the gain
must lie between the origin and this point.
will
FIG. 4.5
is
s
For a general feedback servomechanism, the question

any zero of the expression l/F (s) given by Eq. (4.9)
8
is
whether there
in the right-half
To
use this expression directly for the Cauchy theorem is inconwe have to add two vectors l/Kiffi($) and ^26*2(5).
let Gi(s) and (? 2 (s) have mi and
2 zeros, respectively, in the right-
plane.
venient, because then
Now
The respective numbers of poles in the right-half

and n 2
Then it is evident that the number of poles of
half s plane.
are n\
in the right-half s plane
KzGz(s).
This operation
is
m\
+ n%. Now
let
us divide l/F8 (s) by
introduce mz poles and n% zeros into the

a possibility that some of the zeros may be the
will
But there is
expression.
same as the poles, and thus both are removed.
and poles thus removed be a. Now
Let the number
+
The number of
number of poles
plane
l/F,(s)
of zeros
(4.14)
poles of l/F8 (s)KzG2($) in the right-half 5 plane

of l/KiK%Gi(s)Gz(s)
and
is
thus equal to mi
m*.
is
the
Now
let
us assume that there
no zero
is
41
l/F 8 (s) in the right-half
of
plane,
Then the numbers of zeros and poles

i.e., the feedback system is stable.
It can then be easily
of various expressions are as listed in Table 4.1.
a =
deduced that n%
0,
and \/F 8 (s)K$%(s)
TABLE
Hence
right-half 5 plane.
also has
4.1
path C specified in Fig.
as s traces the
+m
the vector l/F s (s)K2Gz(s) should make

(mi
around the origin. By referring to Eq. (4.14),
dition of stability
2)
4.4,
clockwise revolutions
it is
seen that this con-
equivalent to requiring the vector l/KiK$i(s)Gt(s)

1.
counterclockwise revolutions around the point
Or
is
make mi + m^
we may require the
to
no zeros in the
make mi
vector l/Gi(s)Gz(s) to
+w
counterclock-
This is the Nyquist

wise revolutions around the point
(KiKz)
terion for stability of a general feedback servomechanism.
.
The essential part of the locus

method is the part where s =
cri-
the Nyquist
demonstrated by our
example in Fig. 4.5. Therefore the stability problem can be solved by
using directly the data on the frequency response of the forward link
of the
iw,
path
as
is
of integration in
clearly
and the feedback

the system
is
link.
Since the frequency response of the elements in
often determined experimentally, a method allowing the
This
direct application of experimental information has advantages.
the merit of the Nyquist method.

about the degree of stability. That
Its
is, if
is
drawback is the uncertainty

stable, what is the magnitude
damping? To answer this question, we may modify the criterion to

require no zeros of l/F8 (s) to the right of a line parallel to the imaginary
X between
The distance
axis in the s plane but displaced to the left.
of
this line
and the imaginary
ing.
The Nyquist
tions
on the path C, for
axis specifies the
minimum amount
of
damp-
criterion can again be used, with the proper modificas
and the numbers
ever, to carry out this test,
we have
to
m, mi, and m%. Howthe value of the transfer
of zeros
know
X + iw. Hence information on the

i, but at s =
frequency response can no longer be used directly. Then the method
In fact, a different approach to
of Nyquist loses its main advantage.
the question devised by W. R. Evans is much better. We shall discuss
functions, not at s
this
method
in the following section.
42
Method
4.4
Let us consider
of Evans,
Then
back servomechanisms.
first
the case of simple feed-
the basic question
is
to find the roots of
the equation
(tl5)
with 0($) given. The Evans method determines such roots as functions
and is thus called the root-locus method. When this is done,
of the gain
on the roots gives a proper choice of the magniset of
any
specifications
tude
of
much beyond the mere satisfaction

and actually solves the design problem for all
This method thus goes
K.
of criterion (a) of Sec. 4.2
three criteria stated in that section.
Now
ffi,
02,
(3.21),
pm and its poles

be specified by its zeros pi, p 2
definition of gain given by Eqs. (3.16),
the
from
Then
q.
let (?($)
and
>
(3.23),
0(8)
. A
(JL^PI^^
- ?l)(* -
'
'
?s)
($
(4.16)
?n)
where
'
'
'
(-gi)(-gg)
^-P^-PO
JriÎ
(~p)
For physical systems, the polynomials in the numerator and the denomThen the p s are either real or
inator of G(s) have real coefficients.
the g's are either real or form
Similarly,
form
conjugate pairs.
;
complex
complex conjugate
pairs.
Therefore
is
always
real.
as to make
systems, usually things are so arranged
also
Hereafter, then, we shall consider
positive.
positive.
For engineering
not only real but

A to be real and
order
<?($) is of equal or higher
Let us express each of the factors
Generally, the. denominator of
than the numerator, that is, n

in Eq. (4.16) in vector form:
>
m.
(4.17)
(4.18)
s is
P re
goes from p r to s.
the variable point in the complex
The vector
i<pr
as
(4.18), G(s) can be written
= AA (P
The vector Qre iSr goes from

s
plane.
By using Eqs.
g r to s.
(4.17)
and
A
Since
is
real
and
positive,
we can
43
write Eq. (4.19) as
= Re iQ
<?(*)
(4.20)
where
_
ff ~
"
PP
1
'
'
'
Pm)
(x<3i
,,
91
(4>21)
*)
and
(pi
<P2
'
<pm)
(6l
9,
(4.22)
ffn)
Since the P's and Q's are magnitudes of vectors defined by Eqs. (4.17)
Therefore R is positive. The basic equa(4.18), they are positive.
tion for the roots of the inverse system transfer function, Eq. (4.15), is
and
thus
<r*
KR
Therefore, to satisfy this equation,
KR
~
we must have
(4.23)
and
6 =
(4.24)
consists of two steps The first step is to determine

that satisfy the appropriate angle condition of Eq. (4.24).
Then,
knowing such a root locus, we can compute R and hence K, by Eq. (4.23),
The Evans method
all s's
each point on the root locus.
for
Evans has developed a number
We
useful rules for plotting the root locus.
shall
now
of
explain these
rules.
Rule
0,
1.
For
K=
0,
Eq.
(4.15)
shows that G(s) ->
the roots of l/Fs (s) are poles of
(?($),
oo.
Thus
for
or the root locus starts at the
These poles of 0(s) will be denoted by a dot in the s plane.

the root locus
For K-+ <*, (?(s) -> 0. Thus for K ->
poles of 0(s).
Rule
2.
We
denote the zeros of 0(s) by a small

But if n > m, the number of zeros of 0(s) is less
circle in the s plane.
than the number of zeros of l/F8 (s). However, in that case, (?($) >
could be the zeros of
as
* oo
(?($).
shall
Therefore the missing roots are supplied by
thermore, for very large
oo
Fur-
s,
Therefore Eq. (4.15) can be approximated by

&n~m y^ __
Thus the asymptotes
of
_jr_
-m
"If
the root locus have the phase angles

Jfc7r_
-m
fc
2,' 3,
(4.25)
44
Rule 3. The root locus along the real axis is along alternate segments connecting zeros and poles of G(s) located on the real axis, starting
with the one farthest to the right.
This rule may be easily verified by considering any point s on the real
axis.
The angles to this point from a pair of complex conjugate zeros
+d and
+<p and -<p or
or poles are
sum
The angle
is zero.
for all poles or zeros to the left
is
to the right of
and zeros
Rule
s.
Thus, the sum
of G(s) to the right of
4.
If
there
is
respectively.
0,
Thus
their
from a pole or zero on the axis

of s, and it is TT for all poles or zeros
TT if there is an odd number of poles
to this point
is
s.
of the root locus
breakaway
from the
real axis,
may be estimated from the condition that, for a

small displacement Aw from the axis, the increase in angle due to the
poles and zeros of G(s) on the axis to the left must be just balanced by the
the point of breakaway
effect of those to the right.
Example: Consider the transfer function
_
~
G(s]
(r(s)
(0.001)(2)(6)
0.001)(*
At
0,
the locus starts from
Sections of the locus

oo
Here
m =
0,
between
lie
n =
2)(*
and
2,
0.001,
0.001 and
'
6)
6 on the real axis.
and between
2,
and
Therefore the phase angles for the asymp-
3.
according to Eq. (4.25), are +7T/3,

x/3, and?r.
Breakaway from
the real axis occurs at AI, between
0.001 and
2.
By applying Rule 4,
totes,
we have
Aw
Aw
Ax
0.001
A7+^
Aw
A7T6
or
(A:
2)(Xi
6)
(A!
0.001) (Xi
3A*
16.002A!
6)
(A:
0.001) (Xi
+ 2) =
Hence
12.008
Therefore
/16.002V
~
X,--^V\
12.008
Al
Rule
5.
The point
_
~
U yU4
'
at which the root locus crosses the imaginary axis
into the right-half s plane can often be estimated
by taking advantage
the properties of the right angle.
Example: Let us consider the same transfer function, of Eq. (4.26).
Away from the origin, it can be very closely approximated by
of
(0.001)(2)(6)
Then, as shown in Fig.
4.6, 0i
45
Therefore Eq. (4.24) gives
v/2.
or
But,
by
referring to the figure,

7T
and
hence
This
is
the geometrical condition for determining the crossover point U.
FIG. 4.6
The direction of locus departure from
a pole (or locus approach

be easily estimated by computing the angle at the pole
(or zero) under consideration from all the other poles and zeros in the
to a zero)
may
field.
Example: Figure 4.7 shows the root locus for a transfer function 6(s)
having two zeros and two poles on the real axis, and a pair of complex
conjugate poles. For small displacements away from the pole q*, the
angles <pi <pz, 0i,
Thus the angle
}
6%,
4
is
and
#3
from other zeros and poles remain constant.
given, according to Eq. (4.24),
This equation determines 04.

These rules give the essential characteristics
intermediate locations, the root locus
is
by
of the root locus.
For
found by taking a number of
46
The gain
trial points.
root locus.
When
can then be calculated along the path of the
the desired location of the roots of l/F B (s)
can thus be fixed.

chosen, the corresponding value of
then
is
of the feedback system
completed.
is finally
The design
FIG. 4.7
4.5
and
Hydrodynamic Analogy of Root Locus.

we have
By combining Eqs.
(fi
(s
~
-
gQ(s
'
'
(a
gg)
- g)
- pm
__
j-r
we take the logarithm of the above equation and then

we have
If
(4.15)
(4.16),
divide the result-
ant equation by 2r,
(4.27)
Equation
(4.27)
illuminating one
has
is
many
A very
complex potential function of a
possible physical interpretations.
to consider W(s) as the
two-dimensional irrotational flow of a perfectly incompressible fluid.

If </>(X,&>) and ^(X,w) are the potential function and the stream function,
respectively, then
W(s)
<KX,co)
+ iiKX,)
(4.28)
s = X
iw.
Therefore Eq. (4.27) for the root locus of l/F t (s)
can be interpreted as lines on which the stream function \p assumes the
constant value j. In the terminology of fluid mechanics, the root locus
with
See for instance V. L. Streeter, "Fluid Dynamics," McGraw-Hill Book Company,
Inc.,
New
York, 1948.
is
47
The
potential func-
thus composed of branches of the | streamline.
tion along the streamline changes from point to point
Equation
(4.27) also
shows that the flow
unit strength located at the points #1, ft,

of unit strength, located at the points p\ p 2
is
.
and
is
equal to
composed of n sources of
q n and of m sinks, also
,
In our graphical
Pm*
are
a
and
sources
indicated
sinks
by
dot,
by a small
representation,
circle.
With this interpretation, the pattern of root loci in Figs. 4.6 and
3
4.7 can be immediately "understood."
The hydrodynamic analogy
is
also very useful in suggesting modifica-
tions of the system to achieve a better feedback performance.

instance, a system characterized
For
by the transfer function
may have
the disadvantage of being unstable in closed-cycle performance at too low values of the gain K, and thus not being able to satisfy
The root locus is similar to that shown in Fig.
criterion (c) of Sec. 4.2.
The hydrodynamic analogy immediately suggests that the crossover point U can be moved up by pulling that part of the streamline near
U to the left, with a sink p c close to ft, and a source qc near to q%. Thus
4.6.
the modified transfer function
PC
is
q c ) s(s
ft) (a
ft)
The corresponding root locus is shown in Fig. 4.8. Since \p e < \qe \,
the additional transfer function put in series with the original transfer
\
function must be that of a lead network, as shown in Sec. 3.3
by Eq.
(3.27).
The hydrodynamic analogy also permits us to understand the possispeeding up the response of a slow mechanism by using the
bility of
feedback
link.
According to criterion
requires roots with large magnitudes.
(b)
of Sec.
Now,
4.2,
fast response
for simplicity, suppose
we
have a linear mechanical system of first order characterized by a small

If we put this system in series
qi on the negative part of the real axis.
with a fast-damped electric network characterized by a large q$ on the
negative real axis, the response will not be improved, because we still
But if we have closed the feedback cycle, then
qi root.
have the small
the streamline pattern, or the root locus, indicates that the smaller root
toward the larger root g 2
Thereqi will increase with increasing gain
fore, with proper choice of the gain K, we can make the roots much larger
than
qi
and thus greatly increase the rapidity
of response of the
system.
48
The technique of plotting the root locus can also be applied to general
feedback servomechanisms. There the problem is to plot the root locus
Thus the condition for the root locus is
of l/F 9 (s) given by Eq. (4.19).
1
Since the roots of l/F s (s) are different from the zeros of
(72(5),
we can
Modified
Original
FIG. 4.8
divide the above equation
by Gz(s)/Ki, and thus

(4.29)
Therefore,
if
we put
K.
and then compare Eq.
(4.29)
(4.30)
K.i
with Eq. (4.15), we see that the problem

feedback servomechanism is
of finding the root locus of the general
reduced to that
simple feedback servomechanism discussed previour careful analysis of the application of the Nyquist
of the
ously.
In
method
method
of reduction
fact,
to general feedback servomechanisms in Sec. 4.3
shows that the
given by Eq. (4.30) can also be used there. Therefore, as far as finding the qualitative performance specified by criteria (a),
(6), and (c) of Sec. 4.2 is concerned, there is no difference between the
simple feedback servomechanism and the general feedback servomechanism, if the relations of Eq. (4.30) are borne in mind. Only when the
quantitative performance of the system
is
required must the differences
in the
system transfer functions
as
(s)
shown by Eqs.
49
and
(4.3)
(4.7)
be properly recognized.
Method of Bode. At the point U where the root locus crosses

4.6
over to the right-half s plane, the root is by definition purely imaginary,
= i*, or
say iw*. In other words, Eq. (4.15) is satisfied by s
Therefore, the critical condition of transition from stability to instability

of the frequency response is equal to unity
if the amplitude
the
phase angle 6 of the frequency response is
and, simultaneously,
occurs
equal to
This
it.
critical
which
condition can also be deduced from the
the critical point as

1 in the l/F(iw)
a
such
as Fig. 4.5, it
typical example,
fact, by studying
diagram.
1
will be seen that for stability the l/F(io>) curve must encircle the
Nyquist
criterion,
specifies
In
Since the magnitude of l/F(ia>) generally increases with increascan be ensured by requiring that the magnitude of
encirclement
ing co,
1 when the phase angle of l/F(ia>) is equal to IT.
than
be
larger
l/F(ia>)
should be less than one when 8 is
This is equivalent to saying that
point.
equal to
T.
Or,
we say
that
This condition for stability is

frequency at which the amplitude
to
1 is
= 1.
should be larger than
TT when
the basis of the method of Bode: the
of the
The Bode
called the phase margin.
frequency response
The
called the point of gain crossover.
difference of 8
is
and
equal
IT
is
criterion for stability is thus stated
margin of 30 to 50 degrees at gain crossover. In a Bode dia= 0, and

the
point of gain crossover is the frequency for logio
gram,
the Bode criterion can be easily tested.
as a phase
The Bode method is similar to the Nyquist method in that the information on the frequency response can be used directly.
This advantage of
counterbalanced by the disadvantage, in comparison with

simplicity
the Evans root-locus method, of not being able to know the degree of
is
R.
stability.
M. Osborn
tried to
semiempirical rule to calculate the
His formula
critical root.
where a
is
the same as that for
l/(2
4.7
is
the phase margin in degrees at gain crossover, and m is the

The unit of time for f
against co at the gain crossover.
slope of Iog 10
is
remedy this situation by giving a

for the most
damping coefficient
1.7)
Thus,
if
30 degrees and
1.7,
then
0.3.
Designing the Transfer Function.
The
various methods dis-
cussed in the previous sections for determining the stability of feedback

1
R.
1949.
M. Osborn, paper presented
at
Summer
Meeting, IRE, San Francisco, August.
50
servomechanisms are mainly methods of analysis. They are partly

methods of synthesis, i.e., methods of designing the transfer function, but
Of
as they fix the range of possible values of the gain K,
in so
far
only
both methods can suggest changes in the structure of the transfer

This is particularly so in the case
function to improve performance.
of the root-locus method.
However, how to realize these desired changes
course,
by modifying the physical elements of the system

servomechanism
is mainly an
engineering practice.
Only in one aspect of this synthesis problem is a general solution
known. This is the problem of designing an electric circuit composed of
in the transfer function
art in
'
2J(s)
FIG. 4.9
resistances
and capacitances, an
RC
circuit,
such that the transfer func-
and poles. Since such a

"
often
to
used
flexibility
very
compensate" the
transfer-function characteristics of other elements in the system, and
tion of this electric circuit has the specified zeros
circuit
has great
and
is
since the desired modifications of the transfer function can indeed often
be put in terms of additional zeros and poles, the general solution of such
a problem is very important. Important contributions to this problem
were made by E. A. Guillemin 1 and L. Weinberg. 2 We shall not pursue
the subject here, but only emphasize the possibility of synthesizing an
RC
circuit of
very complex specified properties.
Multiple-loop Servomechanisms. The servomechanisms discussed thus far are single-loop servomechanisms. Engineering practice
often calls for much more complicated systems.
For instance, Fig. 4.9
4.8
is
the block diagram of a typical control system 3 for an airplane rotating

1
2
3
E. A. Guillemin, J. Math, and Phys., 28, 22-44 (1949).

L. Weinberg, J. Appl. Phys., 24, 207-216 (1953),
L. Becker, Aeronaut. Eng. Rev.,
1951, p. 17.
September,
51
about a single axis. The inner loop is the so-called control surfaceIf the inner loop is not closed, we
position feedback or "follow-up."
have the usual feedback control, and
X(s)
both loops are
If
(4.32)
.
then
closed,
A() =
fi(8)[X(*)
F,(8) 7(8)]
and
7(8)
F,(8)A(8)
Therefore
Fi(8)
(4.33)
and
7QQ _
1
and response
of
X(s)
The
stability
zeros of the expression
(4.34)
+F
the control system then depend upon the
+ 0Fj.(s) + Fi(s)F
(s)F^(s).
FIG. 4.10
One of the difficulties of designing a good control system is to have

accurate control and hence large gain
together with fast response and
to
the
led
idea of combining the closedThis
satisfactory damping.
Conby J. R. Moore.
where
the
control
system represented by Fig. 4.10,
closed-cycle
and the open-cycle control are put in parallel. We have thus
cycle control with open-cycle control, proposed
sider the
(4.35)
and
1
J.
R. Moore, Proe. IRE, 39, 1421-1432 (1951).
52
Solving for the output 7(5),
X(s)
we have
+F
(s)
The stability and the speed of response of the system are thus established
by the zeros of 1 + ^1(5)^2(5)^3(5). Since ^2(5) is fixed, the design
problem is to find the proper transfer functions FI(S) and F$(s). The
actual response, in particular the steady-state error,
the additional transfer function
F 4 (s)
of
is
dependent upon
the open-cycle
controller.
Therefore the feedback loop is designed primarily for stability and

dynamic response, while the steady-state or "synchronizing" operations
are largely taken care of
When
there are
when these
many
by the open-cycle portion
of the system.
variables to be controlled simultaneously
controlled variables are also coupled, as in a steam
and
power
system diagram has many loops with a complicated

feedback scheme. 1 An extreme example of such complex systems is,
2
The
perhaps, the automatic control and guidance system for airplanes.
plant, then the
analysis of such a system, although following the
same
principles as
explained in this chapter for simple servomechanisms, can hardly be
done without recourse to analog computers. But this is only engineering development work: the process of going from principles to practice.
1
See for instance
J.
Hanny, "Regelung Theorie," A. G. Gebr. Leemann Co.,
Ziirich, 1946.
2 J.
B. Rea, Aeronaut Eng. Rev., November, 1951, p. 39.
CHAPTER
For complex systems with several controlled quantities and with interaction between these controlled quantities, a new design criterion genThis is the criterion of noninteraction.
erally has to be introduced.
For instance, the variables of a turbojet engine with afterburning are
the engine speed, the fuel injection rate to the combustion chambers, the
fuel injection rate to the afterburner, and the cross-sectional area of
the tail-pipe opening. However, the operation of this engine may be
based upon specific settings of the speed, the fuel rate to the combustion
chambers, and the fuel rate to the afterburner. If this is the case, it is
obvious that one of the design criteria for the servocontrol of the system
is the independence of the three different control settings: a change in
change the engine speed, and a

change in fuel rate to the
fuel rate to the afterburner should not
in engine speed should not require a
change
combustion chambers.
The key to this particular design problem is

then the proper manipulation of the tail-pipe opening with respect to the
The purpose
a general method for designing such noninteracting controls for systems of arbitrary complexity. This general method
was first given by A. S. Boksenbom and R. Hood. 1
5.1
Control of a Single -variable System. Let us consider first a simple system with one controlled output y(t) and one control setting, or
other variables and the proper design of the control servos.
of this chapter is to give
Their Laplace transforms are Y(s) and X(s). Consider the

control designed according to Fig. 5.L E(s) is the "engine" transfer
function, L(s) is the instrument transfer function, S(s) is the servo
transfer function, and C(s) is the "control" transfer function.
Only
input, x(t).
C(s)
can be changed easily by the designer. The system is slightly

from the simple servomechanism of Fig. 4.2, in that an arbitrary
different
disturbance V(s)
is
introduced between the servo and the engine to
account for accidental outside influences.
The
relation
between input W(s) to the engine and the output Y(s)

7(a)
U(s)
is
E(*)W(*)
E(8)[8(8)U(s)
7(8)]
the output of the control transfer function and

U(8)
C(8)[X(8)
A. S. Boksenbom and IL Hood,
Z(S)]
C(8)[X(8)
NACA TR 980
53
(1950),
is
(5.1)
in turn given
- L(8)7(8)]
is
by
(5.2)
54
By
eliminating U(s) from Eqs. (5.1)
and
we have
(5.2),
E(s)
E(s}S(s)C(s)
V(s >
(5 3)
'
E(s)S(s)C(s)L(s)
E(s)S(s)C(,)L(s)
the equation for the Laplace transform of the output under

for y(f) and x(t).
Except for the second
appropriate initial conditions
the same as the previis
term, involving the disturbance F(s), Eq. (5.3)
This
is
simple servomechanisms. The analysis of

the performance of the system can also be carried out in a similar way.
be generalHowever, for more complicated systems, this scheme has to
(4.3), for
ous relation, Eq.
ized.
We shall
do
this presently.
*() -
FIG. 5.1
5.2
Fi(s),
Control of a
(s),
Many -variable
Y,($),
ber of inputs Wi(s), TT 2 (s),

generalization of Eq. (5.1)
Let the number of outputs

be i and the num-
System.
Fi(s) of the engine
k (*),
Fn(s) be
n.
Then the
is
(5.4)
Each Ejk (s)
is
the transfer function which,
when operated on the input
Ejk(s) is then generally

a ratio of two polynomials of s, either obtained theoretically from analyzing the engine characteristics, or determined experimentally through the
frequency response. Equation (5.4) can be compressed into
Wk($), gives a component of the output Yj(s).
=
The array
of the quantities
transfer-function matrix E.
E, k (s)Wk (s)
E vk (s) can be conveniently

We may then consider
(5.5)
called the engine
that the inputs
55
"
Wk(s) enter' the matrix as columns and that the outputs Y v (s) "leave'
the matrix as rows, as indicated in Fig. 5.2. We shall be concerned
7
with the cases where the number of inputs is greater than the number of
Therefore the matrix E is rectangular with more
outputs, that is, n > i.
For later use, a square matrix obtained by using

first
i
columns
is denoted by E*.
the
only
Since the number of inputs to the engine is greater than the number of
outputs, the system behavior is not determined by merely giving the
columns than rows.
settings -Xy(s),
where j
tion the settings
must
(s),
HP (V)
The
also be specified.
for v
1,
TH
1,
i,
for the outputs
for the variables
i,
W^s}, where
p.
Y (s)j but
= i + 1,
in addi-
n,
controlled quantities are then the outputs
and the n
11
If the
engine inputs Wn(s).
Wt
FIG. 5.2
measured values
of
W^s)
after the instrument are denoted
TM (s).
then the errors are S^s)

defined as the differences
(s)
by
T^(s),
The
errors of the engine output are
where Z,(s) for
(s),
1,
i,
are the measured values of the output after the instrument, as shown in
The function of the control is to take these errors as inputs and
Fig. 5.1.
This is the feedback

to generate correction signals 17* (s) for the servos.
In the present generalized control system, the correction signals
link.
17* (s)
are
made
to depend linearly
error signals, there are
upon
all errors.
Since there are
correction signals; k thus ranges
from
to n.
Thus
Ui(s)
= Cn(Zi -
tfi(a)
= Cn(Xi -
- Tn
X, - Z<)
- T)
2n
3n
+
V.(a)
C(X* -
(5.6)
56
where we have separated the control matrix into C and C to indicate

that two kinds of error signals are involved. Equation (5.6) can be
compressed into
(5.7)
=
=!
Each
of
M-i+1
Ckv and
tions (5.6)
and
Cj^
is,
(5.7)
of course, a ratio of two polynomials in s.

Equacan also be represented graphically as shown in
Fig. 5.3.
~T"
I
I
*i
Cni\ Cn,j
t
(5 n -Tn )
Y
C
FIG. 5.3
The measured values of Z v (s) and

by the transfer functions !/,,($) and
are related to
of
(s)
and
the instruments:
(5.8)
(5.9)
The
correction signals will act individually on the servos,
and the
outputs of the servo when combined with the accidental outside disturbances Ffc(s) give the inputs
If /$**() are the
k (s) to the engine.
transfer functions of the servos, then
h (*)
Equations
many
= Skk (s)Uk (s)
k (s)
1, 2,
(5.10)
(5.4) to (5.10)
variables.
completely describe the control system with

Figure 5.4 is a block diagram for a system with three
engine outputs Fi(s),
and Fs(s).
+V
The
2 (s),
and
() and two controlled inputs W^s)
entire system is enclosed except for the

settings
the outside disturbances, which can be imposed on the

system,
and
By
eliminating Uk(s),
equations,
Z v (s), and TM (s) from
57
the previous system of
we have
(5-11)
and
Vk
Equations
(5.11)
and
more compact block diagram for

shown in Fig. 5.5, involves a single
(5.12) suggest a
the system than Fig. 5.4.

Vi
(5.12)
This, as
V2 V3
V5
Control
Settings
FIG. 5.4
V,
58
error signals of the controlled

systems matrix whose inputs are the
The ESC
variables.
controlled
the
variables and whose outputs are
the
in
element
the
which
in
jth row and
matrix in Fig. 5.5 is a matrix
ESC'
the
for
matrix, the element
rth column is E S kk C k ,.
Similarly,
jk
in the jth
of the
row and
SC matrix
/*th
column
S kk C kv
are
is
Ejk S kk C
kfl
Similarly, the elements
and the elements of the SC' matrix are
Skk C
f
.
k(i
The outside disturbances V k are introduced through another matrix composed mainly of the engine matrix E.
5.3
Noninteraction Conditions.
controls can
criterion of noninteraction of
The
in concrete terms.
now be formulated
The problem
to
is
FIG. 5.5
determine conditions on the elements of the control matrix
Ckv (s)
and HM (s) will affect only their respeci

tive corresponding variables Y3-(s) and TFM (s), where j = 1, 2,
the
and
for
and \i = i + I,
else.
example,
Thus,
n,
nothing
C'klt (s) such that the settings Xj(s)
setting
(s) will
modify only
F 2 (s),
while the setting 3*+i(s) will modify
The mathematical problem is thus one
of "diagonalizing"
only Wi+i(s).
the system matrix of Fig. 5.5. We put the design condition on the control matrix, because this is the part of the whole system most easily
modified by the designer. The characteristics of the "engine," the

servos, and the instruments are considered to be fixed and not at the
disposal of the control engineer.
Let us study
first
of the possible values
written then as
1, 2,
i.
with g assuming any one

Equations (5.11) and (5.12) can be
a specific output
(s),
59
/
i
and
vi
2,
Now
F,, the last
for
not influence any Y} or

terms of the above two equations must be zero for
in order that a setting
>
fc
Therefore for any g
i.
will
ff
among
and
the set
C* f
for
1, 2,
>
fc
except
j^g and
i,
(5.13)
(5.14)
of our control
Equation (5.14) gives an immediate simplification
= 3,
5.4
of
the
corresponds to i
matrix. For instance,
Fig.
example
n =
Then Eq.
5.
(5.14) specifies that
(7 41
Equation
(5.14)
= C 42 = Cn = C
5i
= C 52 = C S3 -
can also be used to simplify Eq. (5.13);
it is
in fact
equivalent to
EfiSuC*
Al
where
delta,
is
5
J /A&A,
(5.15)
*=1
any among the
set
1,
2,
and 5^
is
the Kronecker
i.e.,
S,,
For any
specific
g,
Eq. (5.15)
algebraic equations for
jV
is
jf
(B16)
essentially a
unknowns
SkkCkg,
system of
where k = 1,
1 linear
.
2,
i.
unknowns but not

Therefore we can determine
is
what
This is exactly
desired, as we do not wish
their absolute values.
to fix the control transfer function absolutely and thus lose freedom of
only the ratios of these
design.
To
find these ratios of the control transfer functions,
a property of determinants: Let \E
be the cof actor
of
we
shall utilize
the
fl
element
60
in the
determinant \E*\ formed out of the square matrix E*, then

7TT
LI
and
I?*
J %{**
~~~
I1 =5^
'
jfl
'"'
-ill
(5.17)
''?
\
/
IT?*
IT?*!
T?
jiij]c\juji\
i*
IV
\Jii
y-i
Multiplying Eq. (5.15) by \E* and summing over

t
V~\
V~\
Lt
L<
j,
we obtain
Ip
V\
Li
L<
'
3d
31
Therefore, because of Eq. (5.17),
we have
SU C,S =
\E*
EitSuflt,
/ \E*\
1,2,
... ,i
(5.18)
In particular,
Then by taking the ratio
of
Eq. (5.18) to the above equation,
write
M!
This equation gives the off-diagonal elements of the matrix

of the diagonal elements.
The
we can
conditions of Eqs. (5.14)
and
(5.19) are
SC
in terms
then the necessary con-
ditions for noninteraction of the controlled variables
g.
They were
given by Boksenbom and Hood. The same authors proved that these
conditions are also the sufficient conditions for noninteraction.
Therefore the problem of finding the appropriate control matrix C is
completely
solved.
To solve the problem for the other part of the control matrix C', we
have to consider the noninteraction conditions for the controlled variables
Wn where p = i +
and
1,
.
,
n.
For
this purpose,
we rewrite Eqs.
(5.11)
(5.12) as
'+fa
- L^WJ + E
AC' &
kr
- L rWr
r
(5.20)
61
and
Skk C'kr (Sr -
+
where r is any among the sett
1,
,nandj =
l,2,
LrrWr)
.
(5.21)
Forthe
,i.
1
present purpose, the k index in Eq. (5.21) is any among the set i
these
are
the
because
controlled
variables.
It
is
evident
from
Wk's
only
n,
.
Eqs. (5.20) and (5.21) that in order for the setting S r to influence only
the variable
That is,
r the last terms in these equation must be zero.
'*
1,
(5.22)
ife=i
and
C'fr
for
fc,
1,
n and k
(5.23)
Again, Eq. (5.23) gives an immediate simplification of the control

matrix: For the example represented by Fig. 5.4, i = 3, n - 5, and so
Equation (5.23) can be used also to simplify Eq.

reduced to
(5.22).
That equation
is
Ejk SkkC'kr
Multiply both sides

i
By
above equation by
of the
|J57;*j
and sum over j.
using the properties of determinants as given by Eq. (5.17),
\E*\SuC'lr
we can
Therefore
replacing
= -I
by
Z,
~
1
we have
write the above equation in the following form,
by j, and j by
S C
Then
\E*\
LJ
^ ^^
l
ir
i 4. l
(5.24)
#,
This equation then gives the off-diagonal elements of the control matrix
SC f in terms of the diagonal elements. Equations (5.23) and (5.24) are
62
the necessary and sufficient conditions of noninteraction for the con-
trolled variables
W^s), for M
For complete noninteraction
specified
by Eqs.
n.
and
(5.23),
(5.19),
(5.14),
1,
of all controlled variables, the conditions
must be
(5.24)
satisfied.
off-diagonal elements of the complete control matrix are either zero

When the engine charor expressed in terms of the diagonal elements.
The
by the engine matrix are known, the diagonal

elements of the control matrix determine completely the whole control
acteristics as expressed
matrix.
Response Equations. With the noninteraction conditions all

For instance,
satisfied, Eqs. (5.11) and (5.12) can be made much simpler.
5.4
by interchanging the two summations,
Y,(s)
=
v=l
+ Y
/i
But according
-Lm (s)W,(s)]
[g,()
ElkSkk C'k,+
= t+l
to Eqs. (5.13)
term vanishes except when

term vanishes. Thus
and
j.
(5.14), the
sum over
k of the first
(5.22), the
According to Eq.
second
Ejk Skk C +
Ejk Vk
kj
Now
=l
SkkCkj can be expressed in terms of the diagonal element
according to Eq. (5.19).
Thus, using Eq. (5.17),
.=wy
E7 *
Ejk
E*i
'
S#C#
we have
=s
..
Cji
'
*Tl
\ii.
^
l
Jjl
Therefore, finally,
i
vH
w.*\
Eik Vk
By
(5.25)
using the noninteraction conditions, a similar calculation will reduce
Eq. (5.12) to
By
1,
(5.26)
writing
,,_
(5 ' 27)
63
and
a
JKnn
~ci
rtt
/-y/
o w L MJUjL'
At M
-7-
(O.ôJ
the solutions of Eqs. (5.25) and (5.26) can be written as
,t
()7()
(5-29)
Jfc-1
and
BW
= ^GOS,(a) - [#W (*)M<0 - l]^W
(5.30)
and
(5.30) give the relations for calculating the conEquations (5.29)

trolled variables from the settings and the disturbances.
They are quite
similar to Eq. (5.3) for the simple system of one controlled variable.
The function 5#(s) is the over-all transfer function from input X/(s)
to output F/($). The function J? M (s) is the over-all transfer function
from the input EM (s) to output TT^(s). These over-all transfer functions
are calculated according to Eqs. (5.27) and (5.28) using the characterthe engine, the servos, the instruments, and the control. In
the
procedure of design will be to determine for each j and ju the
fact,
istics of
proper control transfer function C#(s) or C^(s) for satisfactory performance by methods explained in Chap. 4. The nondiagonal elements of
the control matrix are then determined by Eqs. (5.14), (5.19), (5.23), and
When
(5.24).
this
is
done,
we have
a noninteracting control of good
performance for a complex, many-variable system.

5.5
theory
Turbopropeller Control.
As a simple example
of
of noninteracting controls, let us consider the control
the general
problem of a
The variables of the operation of such

turbopropeller engine (Fig. 5.6).
an engine are the speed of rotation, the turbine-inlet temperature, the
The control system has to be
propeller blade angle, and the fuel rate.
designed for various possible steady-state normal operating conditions.
For each steady-state operating condition, we have to investigate the
control performance in nonsteady states near that particular operating
Let Wi(s) be the Laplace transform of the deviation of propeller

blade angle from the normal point, and W^s) be the Laplace transform
of the deviation of fuel rate from the normal value.
Then since we are
point.
interested in nonsteady states near the normal point, the relation between
the excess of turbine torque over the torque absorbed by the compressor
and the propeller blade angle and fuel rate can be

Therefore the excess torque is represented by a linear combination of TFi(s) and Tf 2(5). Let the Laplace transform of the deviation
and the
propeller,
linearized.
of the rotating speed from its normal value be FI($).

Then the excess
rs) Fi(s), where r is the characteristic time
torque is represented by (1
constant due to inertia of the rotating components of the power plant,
64
Eq.
cf.
The value
(4.1).
of r
depends upon the normal operating point.
Thus
(1
where a and
+ T8)Yi(s) =
2 (s)
0.
Equation
The physical meanings

held at the normal value,
normal operating point.
a and b are as follows:
(5.31)
2 (s)
deduced from the engine
& are positive real constants,
characteristics near the

of
+ &W
-aTFi(s)
If
(5.31)
the fuel rate
is
then gives a as
- Ki(0)/Fi(0).
But
corresponds to the steady state, and therefore a is the ratio of decrease of

the steady-state engine speed to increase in propeller-blade angle with the
If the steady-state engine performance is given
fuel rate held constant.
as graphs of engine speed versus propeller blade angle at various constant fuel rates, then a
is
normal operating point.
the slope of this graph evaluated at the chosen

Similarly, the meaning of 6 is the slope of the
Combustion Chamber
Turbine
Compressor
Propeller
FIG. 5.6
steady-state engine-speed versus fuel-rate curve with constant propeller

blade angle, evaluated at the chosen normal operating point.
Thus the
constants a and 6 are specified by the steady-state performance curves

of the engine.
For an
mass air flow through the compressor

almost constant at a given compressor
speed.
Therefore, with given inlet conditions, the ratio of heat added to
the gas to the mass of the gas is a function of engine speed and fuel rate.
axial compressor, the
for a certain inlet condition
is
Hence the engine speed and
fuel rate determine the turbine-inlet temperbe the Laplace transform of the deviation of turbinetemperature from the normal value. Then an equation between
and Wi(s), with
a (a) similar to that in Eq.
(5.31), can be estab-
ature.
inlet
2 (s)
lished.
Let
2 (s)
However, since the characteristic time for reaching thermal
equilibrium of the gas
is
practically zero, the relation is simpler:
Yt(*)
cW*(8)
- eY
1 (s)
(5.32)
c and e are again

In fact, c is the slope
positive real constants.
of the turbine-inlet temperature versus fuel-rate curve at constant
engine
where
speed, while e
is
the slope of the turbine-inlet temperature versus

engineall evaluated at the chosen
steady-state
speed curve at constant fuel rate;

operating point.
By
solving for 7i($) and
7 2 ($)
and
in Eqs. (5.31)
(5.32),
65
we have
These equations specify the engine matrix JB in our theory. It is interis only one time constant r in the engine matrix.
esting to note that there
this time characteristic is intrinsic to the engine.

The complete
control system has, of course, other time constants.
But the other time
Only
constants are introduced
by the
control functions, the servos,
and the
instruments, and are not in the engine matrix.

Let us consider first the case of controlling the engine speed and the
fuel rate.
Thus the controlled variables are 7i(s) and Wz(s). We need,
equation of Eq. (5.33), and

the engine matrix has only two elements:
then, only the
first
and n
*-
*-
2.
Hence
(5 - 34)
and
\E*\
The
control system
is
= lEÊu = En
represented
|tf*
(5.35)
by
(5.36)
The noninteraction
conditions require
C() =
(5.37)
and, using Eq. (5.35),

&i(*)CJ,(*)
Since
is
..
- -
JH
#11
J
a
(5.38)
the partial derivative of engine speed with respect to the

and & is the partial derivative of engine speed
propeller blade angle,
with respect to the fuel rate, the ratio b/a is the rate of change of the
propeller blade angle with change in fuel rate at constant engine speed.
Clearly this ratio is a function of flight conditions of the turbopropeller
For instance, this ratio b/a increases as the altitude increases.

Therefore a properly designed control requires means of compensating
for changes in flight and operating conditions of the engine.
engine.
66
The response
and
function fin(s) for the engine speed
is
then
for the fuel rate
characterisThese equations determine the noninteracting response

is to design
The
rate.
fuel
the
problem
tics of the engine speed and
such a way that the
the control transfer functions C n (s) and C"22 (s) in
performance
is
of expected
satisfactory for the full range
operating
conditions,
control.
consider the second possibility of turbopropeller
turbine-inlet
the
and
temperature.
shall control the engine speed
Now
controlled variables are then Yi(s)
need both equations
of
Eq.
(5.33),
and F 2 (s).
and
= n =
We
The
case, we
Thus in this
Then the noninter-
2.
action conditions specify
S n (s)C u (s)
The response function
__
(c-6e) +CTS
s)
S)
for the engine
speed
is
then
**
ac
and
for the turbine-inlet temperature
5.6
lem
Turbojet Engine with Afterburning. We shall now treat the probwith afterburning, mentioned at the
of controlling a turbojet engine
beginning of this chapter.
We
The physical components
are sketched in
study the problem of control for nonsteady

Therefore
states near a chosen normal steady-state operating point.
Fig. 5.7.
shall again
linearization of the relations
between the different variables
is
allowed.
Let 7i(s) again denote 'the Laplace transform of the deviation of

engine speed from the normal value, Wi(s) the Laplace transform of the
the
deviation of the tail-pipe opening from the normal value,
2 (s)
Laplace transform of the deviation of combustion-chamber fuel rate from

the normal, and, finally, Ws(s) the Laplace transform of the deviation of
Then
the tail-pipe fuel rate from the normal.
for
the
to
similar
turbopropeller.
Eq. (5.31)
(1
where
a2
ai,
rs)Yi(s)
and a 3 are
diWi(s)
we can
dzW^s)
As
real constants.
67
write, in a
dzW$(s)
form
(5.42)
in the case of the turbo-
propeller, these constants are slopes of the steady-state performance

curves of the engine. Thus a x is the rate of change of engine speed
with respect to tail-pipe opening at constant fuel rates to the engine

combustion chamber and to the tail pipe. a 2 is the rate of change of
engine speed with respect to the engine fuel rate. a 3 is the rate of
change of engine speed with respect to the tail-pipe fuel rate, r in
Combustion Chamber
Tail
Turbine
Compressor-
Pipe Combustion
Variable Exit Nozzle
FIG. 5.7
again the only characteristic time of the engine system and

This
the
effects of the inertia of the rotating components.
represents
linearized relation between the engine speed and other, engine-input
Eq. (5.42)
variables
is
was derived by M.
S.
Feder and R. Hood. 1
is an axial compressor, Eq. (5.32) of

F 2 (s) represents the
again applicable here.
turbine-inlet temperature, and thus
If the
of
compressor
the previous section
By
the engine
is
solving for Yi(s) and
(s)
in the
above equation and in Eq.
(5.42),
we have
ai
+
+
(5.43)
+rs
F
The elements
M.
S.
matrix are thus
l+rs
l+rs
+TS
1
of the engine
TS
=
~
""
/22
Feder and E. Hood,
(5.44)
CTS
(c
l+rs
l+rs
NAGA TN 2183
(1950).
68
Let us consider the problem of controlling the engine speed, the turbineinlet temperature,
and the
and
F 2 (s),
are then Fi(s),
tail-pipe fuel rate.
The
3 (s).
The
controlled variables
control equations are then
(5.45)
where Xi(s),
2 (s),
and
the
g,(s) are the settings for the engine speed,
turbine-inlet temperature,
The
and the
tail-pipe fuel rate, respectively.
noninteraction condition of Eq. (5.14) requires immediately that
CaiOO
The condition
of
Eq.
= C 32 (s) =
(5.46)
(5.19) gives
(5.47)
Sn(s)Cu(s)
The noninteraction
(c
a 2 e)
or
condition of Eq. (5.24) gives
Su(s)C'n (s)
_ _
03
and
=
The
ical
ratios
a^/a\
meanings:
and
a 2/ai
(5.48)
a 3 /ai in the above equations have simple physthe rate of change of tail-pipe opening with
is
respect to engine fuel rate at constant engine speed and constant tail~a 3 /ai is the rate of change of tail-pipe opening with
pipe fuel rate.
respect to tail-pipe fuel rate at constant engine speed
and constant engine
fuel rate.
When
trol,
Eqs. (5.46) to (5.48) are satisfied, we have noninteracting confor the engine speed is
and the response function
+
for the turbine-inlet temperature
(5.49)
is
(5.50)
69
for the tail-pipe fuel rate is
These equations then give the starting point
of
control transfer functions Cn(s),
and hence
C^s),
C'33 (s),
proper design of the

Ci2($), ^21(5),
CHAPTER
ALTERNATING-CURRENT SERVOMECHANISMS
AND OSCILLATING CONTROL SERVOMECHANISMS
In this chapter and the two following chapters, we shall extend the
servomechanisms in Chaps. 2
concepts and methods developed for simple
and 3 to linear systems which are more complicated but nevertheless
Therefore they
servomechanism design.
and the next chapter follow closely the treatment of
can be treated by approximately the same technique.
demonstrate the power
The contents
of this
L. A. MacColl.
6.1
of the basic principles of
So
Alternating-current Systems.
far,
whenever we have been
considering a servomechanism containing an electric motor, we have

assumed implicitly that the motor is a d-c motor. In practice, however,
It is clear that the use of such
it may well be desirable to use a-c motors.
motors necessitates the reconsideration of some parts
of
our previous
discussion.
Consider a servomechanism as sketched in Fig. 6.1. The purpose

system is to turn the motor to angle <, according to the input
of the
measured by a potentiometer. The voltthe feedback signal. In this system all

age across the potentiometer
of the currents and voltages appearing in the amplifier, motor, and
The output angle $
signal.
is
is
potentiometer are modulated sinusoids, i.e., sinusoidal functions of fixed

frequency, say wo, but with time-varying amplitude. The basic alternat-
When a certain condition,

is generated by the oscillator.
be discussed presently, is satisfied, much of the earlier theory
ing current
which
will
applicable to this system.

Let us consider for a moment the general steady-state theory of
linear systems of constant coefficients subjected to signals which are
is
modulated sinusoids.
Here the expression "steady state" refers to the

assumed to be purely sinusoidal
Let the unmodulated "carrier" be cos CO Q
The
fact that the modulating signals are
functions of time.
phase angle
is
carrier is expressed in real form,

1
L. A, MacColl,
Company,
Inc.,
here neglected without loss of generality.

it is
"Fundamental Theory
New
York, 1945.
70
of
Since the
obviously legitimate to take the

Servomechanisms/' D. Van Nostrand
CONTROL SERVOS
A-C OSCILLATING
modulating signal in complex form,
x (f)
and
fche
e***
cos
co i
iat
.
Then
M
>*
$[e''<+
71
the modulated carrier
i(w ~ wo)i
]
is
(6.1)
steady-state response y sie&dy (t) of a system having the transfer

is, according to Eq. (2.16),
function F(s)
to
(6.2)
For a physical system, the function F(s) is generally a ratio of polys with real coefficients.
Then, as has already been indicated
nomials in
Input
FIG. 6.1
by Eq.
(3.17),
F(-t) = F(m)
(6.3)
where the bar over the symbol indicates the complex conjugate value.
we can
Therefore
write Eq. (6.2) as
^F*(iê^
+ F*(=
(6.4)
where
Now we
suppose that the system

F(ko
Then the
+ iw)
is
such that we have the relation

F(?wo
iw)
(6-6)
expression of (6.4) can be written in the form

cos
This result shows that when the condition of (6.5) is satisfied, the amplitude of the response of the system to the modulated carrier of Eq. (6.1)
is the same as the response of a system having the frequency response
72
F*(iti>)
This statement can be immediately
at the input frequency w.
generalized to apply to
more general input functions by the principle
of superposition for linear systems.
If
Eq.
(6.6) is satisfied,
at least
approximately, throughout a range of values of u> which includes the

more important parts of the Fourier spectrum of a modulating input
signal x(t), the amplitude of the resulting
modulated output signal
is
at
approximately equal to the response of a system having the frehave shown in

quency response F*(iu) to the input signal x(f).
least
We
Chap. 4 that the performance
of the
feedback servomechanism
is
deter-
mined completely by the frequency response. The approximate frequency response now is F*(iu>). Then all the methods for determining
the performance of a system developed in Chap. 4 can be applied to the
The only difference is the use of F*(ico) instead of F(iu)
a-c systems.
in the analysis.
6.2
Translation of the Transfer Function to a Higher Frequency.
we
leave out of account certain trivial systems,

follows from Eq. (6.3) that
e.g.,
pure resistances,
If
it
is different from the condition of Eq. (6.6).

Therefore Eq. (6.6)
cannot be satisfied exactly for all real values of
Or, if we alter our
view
we
of
can
that
the
point
slightly,
say
frequency responses F(iu)
and F*(iw) of two physical systems cannot rigorously satisfy the relation
This
of
Eq.
(6.5) for all real values of
a?.
Nevertheless,
it is
entirely possible,
and indeed quite common, for the frequency responses F*(iu) and F(iu)
of two physical systems to satisfy the relation of Eq. (6.5)
approximately
over a range of values of co which is large enough to include the more
important parts of the Fourier spectra of the particular input signals
we are concerned with. We can see this briefly as follows:
Consider the impedance
of
an inductance
in series, at a frequency a/
If
we make L and C
of such
magnitude that
w
.
then
Z=
Liu
L and
a capacitance
A-C OSCILLATING
For
a/
coo
small, or
co
CONTROL SERVOS
near
coo
That
is,
at the frequency
bination of
impedance
L and C
of
co'
co
coo,
co )
co,
the impedance of the series com-
satisfying Eq. (6.7) is approximately equal to the
an inductance 2L at the frequency co.

impedance Z of an inductance
Similarly, consider the
tance
73
in parallel, at the frequency

1
L and
a capaci-
co'
.
,
condition (6.7)
if
Therefore at the frequency
is satisfied.
co
co
coo,
the impedance of the parallel combination of L and C satisfying Eq. (6.7)

is approximately equal to the impedance of a capacitance 2C at the
frequency
co.
The impedance
of a pure resistance
is,
of course,
independent
of the
the same at the frequency co

coo as at the frequency co.
frequency
with
from
a
a
transfer
function F*(s), we
Thus, starting
physical system
L
a
series
inductance
combination
of inductance
can, by replacing any
by
it is
%L and capacity C
and by replacing any capacity C by a parallel

$0 and inductance L = 2/Cco^, obtain a physical
2/Lco
combination of capacity
system having a transfer function F(s) such that the relationship of Eq.
This procedure
(6.6) is satisfied approximately for small values of co.
of going from F*(s) to F(s)
the frequency scale by co
is
called translating the transfer function
on
Let
co
denote the frequency of the current supplied by the oscillator.

and voltages in the system are modula-
It is clear that all of the currents
tions of the carrier
wave
cos
co
Hence
it
immediately follows from
the above that, in order to design the amplifier for the system with
alternating current, we need only design a suitable amplifier for a system
with direct current by methods described in Chap. 4 and then translate

the characteristics of the amplifier upward on the co scale by the amount
coo,
in accordance with the procedure sketched above.
As we have indicated, the foregoing arguments involve a considerable

number of approximations of one kind or another. An entirely complete discussion of servomechanisms with alternating current
would
necessitate an examination of the effects of all these approximations.
We
not go into this investigation, because it would be involved and

tedious, and because it does not appear to be a very urgent matter as
shall
far as
6.3
servomechanism art
another class of
is
concerned.
Servomechanisms. We shall now consider

which
we call oscillating control servomechasystems,
Oscillating Control
74
These resemble servomechanisms with a-c motors, in that

modulate a periodic oscillation.
nisms.
in
cases the signals are caused to
servomechanisms the modulation
ever, in the case of oscillating control

is
not the ordinary amplitude modulation.
employed
duce the concept
an
of
both
How-
In order to intro-
oscillating control servomechanism intelligibly,
we must first give a little preparatory discussion.

One very primitive but very common kind of servomechanism can
be described as follows. Suppose that to the system we were to add
a circuit containing a relay, designed to function so that no voltage would
be applied to the output terminals unless the absolute value of the input
x(t)
exceeded a certain threshold, and so that when x did exceed the
of a source
threshold the output would be the full electromotive force
the
absolute
as
to
tend
to
reduce
value
with
such
a
polarity
applied
of the error.
We
would then have an example
of
what we
call
an
on-off
servomechanism.
On-off servomechanisms have the great advantage that comparatively
simple systems of this kind can be made to handle large amounts of
This
power.
often difficult to achieve with servomechanisms of other
is
On
the other hand, on-off servomechanisms are definitely nontypes.

linear systems, and, as will be shown in Chap. 10, their performances
tend to be inferior to those of the systems we have considered previously.

Briefly, an oscillating control servomechanism is a modification of an
on-off
servomechanism, which enables us to secure the advantage of

without sacrificing the advantage of large power-carrying
linearity
capacity.
Before proceeding to the treatment of oscillating control servomechanisms proper, we shall present a general theoretical result, upon which
the theory of all such systems is based.
Let us consider a device having
the following property: According as the input signal x(f) is positive or
is
or
A, where A is a fixed constant.
negative, the output signal y(t)
+A
We may
think of such a device as an ideal relay, an on-off system with

zero threshold.
Suppose that the input signal to the relay is
x(f)
where
Qj
A,
co
and
EQ sin u
are constants.
kE Q
sin
cat
(6.8)
In connection with oscillating
control servomechanisms, the term

will be a persistent oscillaQ sin co
tion in the system, and kE Q sin wt will be an applied signal or
modulating
shall calculate the corresponding output y(f)
signal.
presently.
We
of a Relay.
The output of the relay in
response to the input of Eq. (6.8) can be written in the form
6.4
Frequency Response
a nm sin
[(wcoo
na)t]
(6.9)
A-C OSCILLATING
CONTROL SERVOS
75
where the <z s are independent of t. When m = 0, the inner summation

is to be extended only over positive values of n.
For our purposes the
only coefficients that are of any immediate interest are am and a ij
for in the case of an oscillating control servomechanism operating under
?
normal conditions, the other
coefficients either are negligibly small, or
they correspond to oscillations which are suppressed by suitable filtering.

When fc = 0, i.e., when the input to the relay is a purely sinusoidal
function with frequency w 0j the output from the relay is obviously a
series of alternate positive and negative square waves of height A and
Output and Input
FIG. 6.2
It is known that the amplitude a i of the leading term

co /4T.
a Fourier expansion of such a square wave is equal to
duration
of
(6.10)
Tt
When
The
0,
the output from the relay
difference of the outputs with k
is
presented in Fig. 6.2.

= is a series of
and with k
When
rectangles of height 2 A, indicated as shaded areas in Fig. 6.2.
\k\ <<C 1, the change of switching points of the output from the evenly
spaced points of
tn
mr/wo
is
very small.
Thus the correction
rec-
The width of these

tangles are very narrow, as shown in the figure.
can
be
calculated
as
value
of
the
the
rectangles
modulating signal at t n
divided by the slope of the persistent oscillation at
sn
COS
tn .
Thus the width is
sn
0)0
These rectangular areas are to be added (+) or to be subtracted ( )

from the unmodulated output according to whether sin ut n is positive or
76
negative.
areas of the rectangles can be considered as
Thus the
2Ak
sm
wt n
0)0
The
coefficient a 10 in
Eq.
the coefficient of the leading sin ut
is
(6.9)
term in the Fourier expansion of this series of narrow rectangular waves.

Since the value of sin at at the rectangular areas is sin ut n we have, by
,
taking N such
correction rectangles,
N
aio
sin
ut dt
= 2A
ô
Jo
But
sin
ut dt
(1
cos
2ori) dt
l^r_ Igin^T^
4co
coo/
\
2 WQ
and
N
sin
wt n
sin
LI
Li
= o / 1 (^
2Z^L
\
coo/
]
cos [2nir
ô
cos
The sum remains finite as we

N large, we have
increase
indefinitely.
o
2n7r
Therefore,
by
making
a 10
lc
(6.11)
7T
(6.11) give the two important coefficients aoi

For general values of A, these coefficients were
1
When < k < 1,
computed by K, M. Kalb and W. R. Bennett.
Equations
and
(6.10)
and
aio for small k.
where K(k) and jB(fc) denote the complete elliptic integrals of the first
and the second kinds, respectively. For k small, the elliptic integrals
can be expanded; then
1
Bell System Tech. J., 14, 322-359 (1935).
A-C OSCILLATING
CONTROL SERVOS
77
4A
Equation
(6.13)
(6 13)
'
2A*,'
shows that our simple computation is correct within

However, it also shows that the simple results
the accuracy of analysis.

of Eqs. (6.10)
and
(6.11) are accurate
enough
for
even moderate values
Therefore the ratio of the component of frequency w in the output

to the component of the same frequency in the input, i.e., the frequency
of &.
response Fi(iu),
is
approximately equal to
Fi(iu)
94
(6.14)
~jr
As shown by Eqs. (6.10) and (6.13), when k is small, the amplitude of

the component of frequency o>o in the output is approximately a constant
determined entirely by the properties of the relay. Also, the ratio of
the component of frequency co in the output to the component of the
same frequency in the input is 4zA/(jrE }. Thus the amplification of
the relay for the component of frequency ao is 6 db greater than the
amplification for the component of frequency a?.
Now it is easily seen that the preceding considerations can be extended
to the case in
which we have, instead
of the signal
kE
sin ut, a signal
x() of arbitrary form whose magnitude is much smaller than E 0} the

amplitude of the persistent oscillation. The essential result can be stated
as follows: If we ignore higher-order modulation terms for the reason
that they are negligibly small or that they are to be suppressed ultimately by suitable filtering, the relay with the input
3
EQ sin
woZ
+ x(f)
where
x(f) is small compared with EQ, behaves, as far as the transmission

the signal x(f) is concerned, substantially as a linear system having a
constant frequency response given by Eq, (6.14).
6.5
Oscillating Control Servomechanisms with Built-in Oscillation.
of
We are now ready to discuss the oscillating control servomeehamsm.

We have already seen that the only effect of the persistent oscillation, as
far as the transmission of signals is concerned, is to
into
an
make
the relay
effectively linear element, with a positive real frequency response.
Hence we might have considered the
relay to be such an element
from
the beginning, avoided all explicit mention of the oscillation EQ sin co

and dealt with the system entirely by means of the concepts and methods
,
given in the earlier chapters. This

J. C. Lozier.
procedure proposed by
is,
in fact, the novel
and excellent
78
For the sake
of simplicity,
we have assumed
so far that the inherent
is necessary to supterms
introduced
unwanted
modulation
the
by the relay. It is at
press
least conceivable that, in practice, it may sometimes be necessary to
properties of the servo afford all of the filtering that
supply additional filtering by means

rally, whatever effective filters there
supplementary filters. Natumay be in the system must be
of
such that they pass the wanted signals.
This, in combination with our
other considerations, implies that the frequency w must be above the

important parts of the Fourier spectra of the signals.
No matter what filtering we may introduce into the system, the out-
always contain, as one component, an oscillation having the

frequency coo. It is worth remarking that it may not even be desirable
put
will
to reduce the amplitude of this
component below a certain
by
level,
FIG. 6.3
filtering.
In fact, such an oscillation furnishes "dynamical lubrica-
effects of static friction, backlash, and

other parasitic nonlinearities tending to degrade the performance of
tion,"
which diminishes the
servomechanisms.
We
have not said anything very
the oscillation
incidentally that
in
which the
in the
way
sin
it
co
is
may
specific
supplied to the relay;
about the
way
in
which
we have merely remarked
be supplied by a subsidiary oscillator. Systems

way have certain advantages
oscillation is supplied in that
However, they have the disadvantage
of flexibility.
of
We
shall now give

involving a certain amount of extra equipment.
a brief description of a variety of oscillating control servomechanisms in
itself is made to supply the oscillation.
Consider the system sketched in Fig. 6.3. Suppose that the system
so designed that in the absence of the input signal it oscillates at a
which the servomechanism
is
frequency o? determined by the phase shifts of the linear elements in the

feedback loop. As we have seen, the relay behaves, as regards the
oscillation, as an effectively linear element, having a frequency response
which decreases as the amplitude
is
increased.
The amplitude
so that the amplification
of
the
around the loop,

determined by the amplifications through the relay and through the
oscillation
adjusts
itself
linear elements, is unity.
A-C OSCILLATING
CONTROL SERVOS
79
Now
suppose that the system is subjected to an input signal. If the

corresponding error signal at the input of the relay is sufficiently small,
the amplification of the relay for the persistent oscillation is substantially
and the system continues to oscillate at substantially the

As we have seen, the relay behaves,
original frequency and amplitude.
as regards the signals, as an effectively linear element, having an amplification which is 6 db less than the amplification for the persistent oscillaIt is clear that under these conditions we have an oscillatingtion.
The only novelty
control mechanism such as we have discussed above.
in the present situation is the fact that the frequency and amplitude
unaffected,
of the persistent
oscillation
sin u$t are
determined by the system
instead of being determined independently, as
itself,
we have
tacitly
assumed heretofore.
In
all
our considerations of the system as a servomechanism,
we need
only ascribe the proper effective frequency response to the relay,
and
then proceed in the ways described in the preceding chapters. We do

not need to take account explicitly of the persistent oscillation. However, the requirement that the system shall also function as an oscillator
imposes certain restrictions on what we can do toward improving

performance as a servomechanism. This can be seen as follows.
its
Let F(s) denote the transfer function of the feedback loop for the
computed by using Eq. (6.14). Then the transfer function for

the persistent oscillation is 2F(s) and, by the very fact that the system
does oscillate, there is a purely imaginary root s = icoo for the system
transfer function 1
Therefore
[l/2F(s)].
signals
Hence, the curve l/F(s) in the Nyquist diagram, as
s traces the imagi2.

constrained to pass through the point
On the other
hand, in order that the performance of the system as a servomechanism
shall be satisfactory, the curve must meet the conditions we have dis-
nary
axis, is
cussed in Chap.
of the point
jected
makes
4,
1.
it
including the condition of avoiding the neighborhood

Obviously, the constraint to which the curve is sub-
more
difficult
to meet these conditions than
it is
in the
other systems, where no such constraint exists. In this sense, these selfoscillating servomechanisms are less flexible than are oscillating control
servomechanisms in which the
oscillation is supplied
by an independent
generator.
An
which
elementary precaution to be observed, in order that the curve,

is constrained to pass through the point
2, shall avoid the
the
neighborhood of the point
1, is that the curve should intersect
real axis at the point
2 perpendicularly. This implies that the vector
80
should be varying slowly in magnitude and rapidly in angle at the

frequency at which the system oscillates.
General Oscillating Control Servomechanism. A relay is a nonBut by mixing the signal with a sinusoidal oscillation of
6.6
linear device.
high frequency and large amplitude, the output is made to be linear with
Thus the essential concept of oscillating control
respect to the signal.
the linearization of a nonlinear system. J. M. Loeb 1
has shown that this concept is applicable to any nonlinear system, and
he calls this method the general linearizing process for nonlinear control
servomechanisms
is
We shall
systems.
call
the resulting servomeehanism the general
oscillat-
ing control servomechanism.
is
is
Let us consider a general function y(x), where y is the output and x

the input. If for the variable x we substitute the sum x
e, where e
much smaller than x, then, if the function y(x) is regular, we can expand
y(x
e)
into a Taylor series as
We now
specify the input # as a periodic function of time t with the

period T, and e as a constant. Then it is clear that y(x) is also a periodic
function of time with the same period T. The same is true for dy/dx and
2
Periodic functions can be expanded into Fourier series; thus
d*y/dx
.
we
if
y(x
neglect powers of
e)
higher than the
n cos
n^
first,
we have
^ Qn s^
CO
aio
+ y
(ain cos
nut
& in sin nut)
(6.16)
n-l
where
co
If e is
that
its
2ir/T, the frequency of the input x.

not exactly a constant but a slowly varying function of t such
fundamental frequency is very much lower than co, then Eq.
approximately correct. Now consider y(x) as the outputinput relation of the nonlinear device, e(0 as the signal, and x(t) as the
(6.16) is still
superimposed high-frequency, large-amplitude oscillation, not necessarily

The signal information in the output of the nonlinear device
sinusoidal.
is
represented
is
much
by the second term
higher than that of
(<),
of Eq. (6.16).
Since the frequency co
the periodic function represented by the
Fourier series
00
}
1
J.
M. Loeb, Ann.
ain cos
de Telecommunications,
.6,
65-71 (1950).
A~C OSCILLATING
CONTROL SERVOS
81
can be considered as the carrier, while e(t) is the modulating function.

While the above discussion is based upon a direct relation y(x) between
the input and output of the nonlinear device, Loeb has shown that Eq.
more general functional relation between y and x,

dependent not only upon the instantaneous x at the same
but also upon all past values of x. This extended concept of input-
(6.16) is also true for the
that
t,
is,
y at
t is
output relation includes hysteresis effects such as gear blacklash, and

for almost all nonlinear devices in practice.
Therefore, for
is sufficient
a general oscillating-control servomechanism, the signal in the output

of the nonlinear device appears as a modulated carrier wave.
Moreover,
the input-output relation
is
linear as far as the signal
is
concerned.
*(*)
FIG. 6.4
Now let us assume that the superimposed

wave form such as the sine wave
Then if y(x) is even, or
oscillation
or the saw-tooth
has a symmetrical
wave shown
= y(-x)
y(x)
(6.17)
and
we have
in Fig, 6.4.
the following relations for the periodic function y(x) and dy/dx
'(x) t
y(x)
and
These requirements then specify

ttoi
Therefore,
if
601
0,
for y
even
(6.19)
the higher harmonics are neglected, the carrier is a sinusoidal

This is the case of the a-c servomechanism
oscillation of frequency w.
discussed in the previous sections. The design method described there

can then be applied to this class of general oscillating control servomechanisms. If y(x) is odd, or y(x) =
y(x), then a set of conditions
similar to Eqs. (6.17)
can be written down, and
and
(6.18)
an = &n =
0,
for y
odd
(6.20)
82
By neglecting higher harmonics, the case is thus identical to the oscillating control servomechanism discussed in Sec. 6.4.
The preceding discussion shows that the characteristics of a nonlinear
component
in a
servomechanism can be linearized by the technique

and thus converting the
of adding a persistent oscillation to the input
system into an
formance.
oscillating control
servomechanism
of
improved per-
Furthermore, such servomechanisms can be designed with
methods already explained
in this chapter.
CHAPTER
All the servomechanisms that
we have
considered so far are designed
to deal with signals which are given as functions of the continuous

There are situations, however, in which the signals which a
variable L
servomechanism has to deal with are given as functions of a discrete

Such a situation arises, for example, when we have an input
which
has been obtained by determining the values of a function
signal
variable.
x(t)
at equally spaced instants 0,
to,
signal is not defined at all in the
2t 0}
....
In such a ease, the input
open intervals between the successive
sampling instants.
Naturally, when we have a situation of the kind just described, we
are interested in the values of the output signal at the sampling instants.
Consequently, the servomechanism should function so that the correcwhich it applies to the output signal is governed only by
those values and not by the values which the output signal may have
tive effect
during the intervening intervals. A servomechanism which is designed

to function in this way may be called a sampling servomechanism.
In
this chapter,
we
give a brief account of a theory of linear sampling servo-
mechanisms which
is
very similar in
point of view and procedure

servomechanisms that we have
its
to the theory of continuously operating
been discussing in the preceding chapters.

7.1
Output of a Sampling Circuit. The prototype sampling servomechanism which we shall consider is shown in Pig. 7.1. The system
contains the usual forward and feedback circuits.
of the
system
lies
The
essential novelty
in the fact that the feedback path contains a switch,
which
is operated periodically so that the feedback loop is closed only

during short time intervals located at the equally spaced instants 0, t Q
2
.... The location of the energy-storing, or frequency-selective,
,
elements in the system affects the theory in matters of detail only.

Hence we take the opportunity to simplify the exposition somewhat by
assuming that the transfer function

of the frequency.
Then
it is
of
the forward circuit
essential that the switch
is independent
be placed in the
position shown.
The
following analysis
is
during which the switch
is
based upon the assumption that the intervals

closed are so short that the feedback circuit
83
84
can be considered to be subjected to a sequence of impulses. It is also

based upon the assumption that the response hi(t) of the feedback circuit
This means that for
to an impulse is a continuous function of time.
n
Then hs(t) has no
small values of t, A 2 (0 behaves like t where n>l.
,
jump
in value at
circuit,
0.
If Fz(s) is
the transfer function of the feedback
then according to the general formula, Eq.
(2.18),
(7.1)
where 7
of
is
(s).
a real constant which

If for large
values of
s,
is
greater than the real part of
any pole
m
Fz(s) behaves like l/s , then according
x(t)
Amplifier
Periodically
Operated Switch
FIG. 7.1
to the "dictionary" of Laplace transforms, Table 2.1, hi(f) will behave

m~ l for small t. Our condition for the
like t
continuity of hz(f) at t =
thus requires
m to be at least 2.
at least as rapidly as 1/s 2
Now
Thus
for large
s,
Fz(s) approaches zero
that if the input signal x(t) vanishes identically for

of
values
value of the output signal y(t) at the typical
negative
t, the
instant
is
ufa
sampling
computed as the sum of the effects of all previous
and
is
given by the formula
impulses
it is clear
y(nt Q )
(7.2)
where F\ denotes the transfer function of the forward circuit, a constant,

and 8 is the fraction of the switching cycle during which the switch is
closed.
6tQy(kt Q ) is thus the "impulse" input to the feedback circuit at
t
= kto.
When
x(f)
of y(0), y(t
in
and
are known
... can be
hz(f)
at the sampling instants, the values
calculated successively by Eq. (7.2)

However, instead of proceeding in that way, we
follow a more illuminating course, which will bring the theory of
),
y(2to),
an elementary way.
shall
the sampling servomechanism into a form similar to that of the theory

of the ordinary servomechanism discussed in Chap. 4.
This approach
is
due to G. R. Stibitz and C. E. Shannon.
7.2
Let us write
Stibitz-Shannon Theory.
X*(s)
85
x(nto)e-
nt *
(7.3)
K=0
00
y*(s)
y(nt )e-
nt s
(7,4)
and
(7.5)
These functions are thus periodic functions
The
period 2iri/U.
The
of
s,
with the imaginary
functions of nt Q are thus the Fourier coefficients.
step of going from the functions x(t), y(t), and h z (f) to X*(s), F*(s),
is very similar to the formation of the corresponding Laplace
and FJ(s)
transforms X(s) Y(s), and F z (s), as indicated by Eq. (2.1). Here the
continuous time variable is replaced by the discrete time instants nU,
and thus the integral sign is replaced by the summation sign. Therefore
}
Eqs. (7.3) to (7.5) represent the natural adaptation of the Laplacetransform technique to the problem of the sampling servomechanism.
For the time being we
shall confine our attention to the case in
which
the poles of Fz(s) lie to the left of the imaginary axis. Then the
function hz(t) ultimately decays exponentially as t tends toward infinity,
and the series in Eq. (7.5) converges for all values of s with real parts
all of
which are greater than a certain negative constant. Of course, the convergence of the series in Eqs. (7.3) and (7.4) depends upon the nature
of
the input signal.
We
restrict
our attention to input signals for which
the series converge in the same manner as the series in Eq. (7.5).
This
amounts only to the mild sort of restriction on x(f) that we are accustomed
to
assume in transient theory.
nhs and then

Multiplying Eq. (7.2) through by e~
summing over
values of n, we obtain
7*(s)
= Fi
[**()
fto
e-"'
n=0
kt Q )
&=0
But
o
k=
y(kto)hi(nt,
A*,)]
all
86
The
and
over k
step of changing the summation over n and k to a summation
= n k is indicated in Fig. 7.2, where the shaded region is the
Therefore
region of summation.
or
(7.6)
Equation
FIG. 7.2
What difference there is between the cases
mechanisms, discussed before.

lies in
the analogue of the
(7.6) is
basic equation (4.3) for feedback servo-
the analytical natures of the functions involved.
We shall discuss
this point later.
Now
let
of Y*(s) is
the series
us assume that the character of y(nt Q ) is such that the series

Then
$ with nonnegative real parts.
convergent for values of
is
convergent for purely imaginary
s.
Let
iw.
Then
Therefore
or
By
putting iw
and io =
So,
we
have, finally,
By
the Cauchy theorem of complex integration,
<"'
where, as shown in Fig. 7.3, T
separated by the distance 27r/
passing to the right of
tion of
is
now
all
a path of integration joining two points

on the imaginary axis in the s plane, and
This descripsingular points of the integrand.
is
in such a general
tive real parts, Eq. (7.7) is
still
form that
if
Y*(s) has poles with posi-
true.
Because of the periodicity of X*(s) and FJfa), we can add to r the

dotted lines parallel to the real axis and leave the value of the
integral
unchanged.
The combined path then
encloses all the poles of the
integrand.
But
87
can be shown that for reasonable input X*(s) has

real parts.
Then the only possible source of
it
no poles with positive
unstable output is a zero of the denominator of the integrand of Eq. (7.7)

with a positive real part. Thus, in a way very similar to the requirement on conventional servomechanisms, the necessary and sufficient
condition for stability is that the
plane
equation
1
(7.8)
shall
have no roots in the right half of
the
plane.
We shall now show how
we can implement
this condition
by
an appropriate adaptation of the
Nyquist
FIG. 7.3
criterion of Sec. 4.3.
7.3
Nyquist Criterion for Sampling Servomechanisms. Because of
the periodicity of F%(s), it suffices to determine whether or not Eq. (7.8)
has any roots in a horizontal half strip, of width Sir/to, extending to the
from the imaginary axis. We are assuming that F$(s) has no

singular points on or to the right of the imaginary axis, and we also assume
that 1 + 6FiF*(s) has no zero on the imaginary axis. We can, and do,
right
assume that the half strip is adjusted vertically so that 1

$FiFf($)
has no zeros on the horizontal sides. Now let the point $ describe the
2tri
$0+
i- plane
D
FIG. 7.4
ABCDA
Then the corresponding tip of the

of Fig. 7.4.
closed curve
vector 6FiF*(s) describes a certain closed curve, such as A'B'C'D'A'
shown in Fig. 7.5. We do not try to show the curve described by the
vector realistically.
When s describes BC3
of Ff(s)j
B'C'
is
When
AB, we have the
describes
we have an
a closed curve.
arc
arc
A W.
f
B'C and, because of the periodicity

s describes CD, we have an arc
]
When
C'D'; and, because of the periodicity of FJ($), C'D' coincides, except for
sense, with A'B'.
Finally, when s describes DA, we have the arc D'A',
which
is
a closed curve.
By Cauchy's
rectangle
1
net
theorem, Eq.
ABCD
(7.8) does, or
does not, have roots in the

from the point
according to whether the radius vector
to the running point 6FiF*(s) does, or does not,
number
A'B'C'D'A'.
of revolutions as the
Now
consider
what happens when the
rectangle in Fig. 7.4 recedes to infinity.
that the closed curve formed
by the
-
make
a nonzero
running point describes the curve

side
It is easy to see
BC
of the
from Eq.
(7.5)
arc B'C' in Fig. 7.5 shrinks to a single

Therefore the effective curve
point.
plane
is
the arc D'A'.
Obviously, Eq.
does not, have roots in
the half strip according to whether
(7.8) does, or
the radius vector from the point

to the curve does, or does not,
make
a nonzero net number of revolutions

as the limiting curve
is
This
criterion
is
the Nyquist
described.
for
simple sampling servomechanisms.

We have now given the fundamen-
FIG. 7.5
tals of the
Stibitz-Shannon theory
sampling servomechanisms, which is very similar, both in its point

of view and in its form, to the theory of servomechanisms with continuous
of
operation.
7.4
If the
Steady-state Error.
input
is
a unit step function,
71
71-0
then, according to Eq. (7.7),
y(nto)
As n
>
oo
,
tor
i
-
TT
the only pole of importance
is
at the origin,
(7.9)
"
This equation gives the steady-state " output for a constant
input of
unit magnitude.
Therefore the condition for small
error is
steady-state
or
1
8Ff(0)
(7.10)
89
This gives the approximate magnitude of the gain for the forward circuit
the output is to follow the input accurately. Equation (7.10) for
sampling servomechanisms is the analogue of Eq. (4.11) for continuous
if
servomechanisms.
have one more step to take before our
sampling servomechanisms can be regarded as being
of
theory
Both
practical value.
ize
We
Calculation of F*(s).
7.5
of the functions F$(s)
the feedback circuit:
is
z (s)
and F*(s) serve
of
much
to character-
the significant characteristic
when the
used as part of a continuously operating servomechanism; and

the significant characteristic when the circuit is used as part of a
circuit is
is
F*(s)
matically much the
which
The
servomechanism.
sampling
common
used directly in our
is
It is important, therefore, that
we
function
familiar
simpler of the two,
Fz(s)
and furthermore
it is
techniques for designing
relate F*(s) to Fz(s) }
and
mathe-
is
the function
circuits.
in as direct
a manner as possible. We note again that F*(s) is periodic in s with

the imaginary period i2ir/U. Furthermore, the analysis of the perform"
ance of the system by the Nyquist method requires the
frequency
TT/O < w < T/ O
response' F*(iu>) only for
Assuming that the real part of s is greater than 7, which according
to our assumptions is a negative number, we have, by Eqs. (7.1) and
7
(7.5),
oo
/*7
+ i*
rcA-i_
V\
\
FM do W
/
re
(7.11)
We
proceed to evaluate the right-hand
member
of (7.11)
by the method
of residues.
The integrand has

path
of integration,
~ e)
certain poles: the poles of Fz(s)
lie
and the poles which are the roots
to the left of the

of the equation
to the right of the path of integration. It

00 to
seen that the integration upward along the line 7
7
1
6~ Jo(fi
lie
is
easily
is
equivalent to integration in a clockwise direction along the closed curve

formed by that line and the infinite semicircle in the right-half plane.
Hence the right-hand member
of
Eq.
U times the sum
(7.11) is
of the
residues of the integrand with respect to the several roots of the equation
1
_ e-*o(a~fi) = 0.
Now the typical
root of the equation
is
an integer and the residue
is
-(1/*
)^2[
(2irim/to)].
is
of the integrand
(27rim/
),
where
with respect to that pole
Therefore, finally,
90
L4
Tiiis
formula gives considerable insight into the properties of F*(s) and

may be useful in making approximate calculations. However,
at times
we can easily obtain an exact representation of F$(s) in finite form.

The function ^2(5) can be represented as the sum of a finite number
of
partial fractions, thus:

n
Ll
where the a^s and the
s^'s
(7.13)
K
Sk
n is the degree of the denomConsequently, we can write, by using Eq.
are constants and
inator polynomial of F*(s).

(7.12),
Y
Z/
i(2irim/U)
sk )
(s
,y
LI
Now it
is
known
sum over
m in
F* (s) =
By means
of
of this
= -
s k )\l
(s
that coth z has the following expansion
coth z
Therefore the
i|
(2Trim/to)
+ z*
Eq. (7.14) can be carried out, and we have
i 2/
ak coth
<7 -
r~~^J
formula we can compute
F * (s)
15 )
any value
exactly for
s.
When
is
very small and
interval -Tr/t Q
<
<
TT/ZQ,
F*(i<a) is negligibly
small outside of the
the qualitative nature of FJ(iw)
is
immedi-
In fact, Ff (iw) is approximately equal,

ately apparent from Eq. (7.12).
in the interval -x/$ <
< 7r/s5 0j to the function f 2
shall now
(iw).
see that
when
is
large
We
we can obtain an equally simple approximation
Let us write the roots
to F*(ico).
where the
and
X's
the X's are
co's
Now we
fc
iwjb
(7.16)
In accordance with our assumption,
are all real
all positive.
as
sk
= -X
sk
91
have, according to Eq. (7.15),
coth
Therefore, for large values of U,

n
|
When
s is large,
circuit to
as a condition for continuous response of the feed-
F 2 (s)
an impulse that
a*
A
Then Eq.
(7.17)
(7.17)
*in}
Eq. (7.13) can be written as
But we have assumed

back
~ 1/s
when
s is large.
Therefore
(7.18)
=l
becomes
n
(7.19)
A-l
For physical systems, the s/s are

Therefore the
of
Ff (iu>) as
a;
finite
sum
goes from
real or
form pairs
in Eq. (7.19)
to ir/U
ir/t Q
is
is
of
complex conjugates.
and the graph
actually real,
circle
with the radius
n
t<>
^ <^
osi
,
(7.20)
7.6
Comparison of Continuously Operating with Sampling Servomechanisms. For small fo, we have seen that F^(ica) is approximately
F 2 (iw). The stability criterion for the continuously operating servomechanism is that the curve FiF2 (tw) should avoid the point
For
1.
92
sampling servomechanism, the curve 6FiF*(it)) should avoid the

Therefore,
1, or the curve F\Fi(iw} should avoid the point 1/6.
point
the
stability is the only consideration, the sampling servomechanism can

have much larger gain than the conventional servomechanism.
Tor large values of U, because of Eqs. (7.19) and (7.20), the Nyquist
criterion becomes simply
if
V ak e
tosi
<
has a negative real part, the radius of the F 2*(tw) curve is very
This fact together with the smallness of 6, the fraction of time
when the switch is closed, allows very large gain for the forward circuit
Since
sk
small.
Thus for any value of the switching period t Q the

condition for stable operation of a sampling servomechanism is very
without instability.
much
less stringent
than that for a continuously operating feedback
Perhaps this is to be expected, because the time

interval when feedback of the output actually occurs is very brief, and
servomechanism.
no
imposed on the output other than that at the switching
restriction is
instant.
Pole of
7.7
F 2 (s)
at Origin.
have a pole at
likely to
In practice, the function
So far we have excluded
0.
shall now consider the

we observe that when s =
However, we
(s) is
quite
from
minor com-
consideration, in order to avoid having to deal with certain

plications.
this case
case briefly.
is a pole of F 2 (s) the

constant 7 must be positive and that our representation of F *(s) in
terms of infinite series is valid only for values of s with positive real
In the
first place,
In the second place, the Nyquist diagram for the system also
Specifically, instead of getting an actually closed
parts.
undergoes changes.
curve, we get an open curve, the ends of which are to be regarded as

being joined by an infinite semicircle in the clockwise sense. Our
representation of F*(s) in finite form, however, remains valid.

set $i
0,
[.
,
IQ,\
For U
If
then Eq. (7.15) gives
large,
we
coto
COt
-.
.j
t
\/
,r JO[A& ~r
&
djc
COtD.
HW
~
co/j]
obtain an expression similar to Eq. (7.17),
i.e.,
n
\
rr*/
Ffuw) =
But according
^0
pr
2 L
-~
jWtfo
cot
to Eq. (7.18),
+
,
L^
ak \l
+
i
n
r\ _LV ~
MI!
2e~ io[X *+z(w Wft) M
*
'J
we
93
so that
nit /
\
F% (to) =
The constant
to
+7^0,
the
ai
is,
first
wiÛ
1^
-<
+*
i
of course, real
cot
and
wv
-~
+
i
positive.
io^
*
*"
As
term gives a vertical straight
/
co
&e
iosfc
varies
from
/
nt N
(7.21)
w/U
line parallel to the
The other part of F*(i&) is a sinusoidal function.

axis.
Hence the Nyquist diagram is a sinuous variation of a straight line.
imaginary
CHAPTER

In this chapter we shall introduce another new element into our linear
systems with constant coefficients the time lag. By time lag r we mean
that the relation between the different variables of the system cannot
:
be expressed as a relation of these variables all taken at some time instant

involves some variables taken at the
t] but on the contrary, the relation
time instant
and some taken
t,
at the instant
at the instant
t.
istic
at
an
earlier instant
Those taken
r.
by the interval r behind the variables taken

This time lag is thus quite different from the characterr then lag
time constant of the first-order linear system introduced in Sec.
Time-lag systems are represented
by
3.1.
differential difference equations of
constant coefficients and are more complex than the linear systems
studied previously, which are represented by differential equations,
Systems with time lag were studied by
many
investigators: for instance,
A. Callander, D. Hartree, and A. Porter, 1 and N. Minorsky. 2 Our interWe wish to know:
est here is, however, somewhat more restricted.
How
if
can we analyze the performance of a feedback servomechanism

is a characteristic time lag T in the system?
We wish, specifically,
there
to modify the Nyquist
method
of Sec. 4.3 to
apply to time-lag systems.

a
by treating particular example of such
the
of
systems, namely,
example
stabilizing the combustion in a rocket
We
shall develop the theory
motor by feedback
control.
The problem
rocket motors has been treated by
many
analysis of combustion lag time originates
For simplicity
of calculation, 4
we
combustion instability in
authors, but the following
of
from the work
of L. Crocco. 3
shall consider only the case of so-called
low-frequency oscillation in a rocket motor using a single liquid propellant.

8.1 Time Lag in Combustion. Let
b (t) be the mass rate of genera-
tion of hot gas

injection at
1
by combustion
at the time instant
can be denoted by w*().
Let
r(t)
t.
The mass
rate of
be the time lag for that
A. Callander, D. Hartree, and A. Porter, Trans. Roy. Soc. London (A), 235, 415-444
(1935).
2
N. Minorsky, /. Appl Mechanics (ASME), 9, 67-71 (1942).

Am. Rocket Soc. 21, 163-178 (1951).
L. Crocco, J.
The
following discussion
is
based upon a paper in J. Am. Rocket
(1952).
94
Soc., 22,
256-262
95
parcel of propellant which is burned at the instant L Then the mass

dt must be equal to the mass
burned during the interval from t to t
injected during the time
tiii,(t)
The mass
chamber by
dt
= m
raising its
(t
r)d(t
T).
is
Therefore
r)
(8.1)
either used to
is
pressure p(() or
fill
the combustion
discharged through the rocket
chamber
chamber can be considered uniform,
frequency of the possible oscillations within the
low, then the pressure in the
is
+
+ d(t
hot gas generated
of
If the
nozzle.
T to
1
and, as a first approximation, the flow through the nozzle can be conTherefore the mass rate of discharge through
sidered quasi-stationary.
the nozzle is proportional to the density of hot gas in the rocket motor.
But
for a
is
"monopropellant" rocket motor, the temperature of the hot

nearly independent of the combustion pressure, and the density
gas
Thus if m is the steady
of the hot gas is proportional to pressure only.
is
mass rate of flow through the system,
the
g
average mass of hot gas
in the motor, p is the steady-state pressure in the combustion chamber,
and
if
the volume occupied
by the unburned liquid propellant is neglected,
we have
b
We now
dt
fa
dt
+ d [M
(8.2)
introduce the nondimensional variables
chamber pressure and the rate
<p
and p
for the
of injection, respectively, defined as
m
<p
and
ju
are then the fractional deviations of the pressure and injection

With Eq. (8.3), m& can be eliminated from
from the average.

Eqs. (8.1) and (8.2), and
rates
-'>
To
(8.4)
calculate the quantity dr/dt, Crocco's concept of the pressure

of time lag has to be introduced.
If the rate at which the
dependence
liquid propellant is prepared for the final rapid transformation into hot
gas
is
a function f(p), then the lag r
is
f(p) dt
determined by
const.
(8.5)
This constant can be thought of as the amount of heat that has to be

added to a unit mass of the cold injected propellant before "ignition"
1
H.
S. Tsien, /.
Am.
Rocket Soc. } 22, 139-143 (1952).
96
occurs.
Then f(p) has the physical meaning of the rate of heat transfer
from the hot combustion gas to the injected liquid propellant. By differentiating Eq. (8.5) with respect to t,
[/(P)].
=
)
We can now introduce explicitly the concept of a small perturbation from

the uniform steady state. Assume that the deviation of the pressure p
from the steady-state value p is small. Then f(p) at the instant t and
f(p) at the instant
By
can be expanded as Taylor's
series
around
p.
taking only the first-order terms,
Here
r is the lag at the average pressure p, a constant
By combining
Eqs. (8.4) and
J+9
M (*
now.
(8.6),
the following equation
+ n(v(z) -
5)
v(z
is
Then
obtained:
5)]
(8.7)
where
to/)
\d log pJ ,-
(8 8)
.
P f
and
-5 -i
'-I
6g is
thus the average gaseous mass in the motor divided by the average
mass flow through the motor, and is thus the average time between
rate of
the instant of production by combustion of the hot gas to the instant

6 g is therefore called the gas transit
of discharge through the nozzle.
time.
shall measure time by this fundamental time constant in
We
the following calculations,
z is
the nondimensional time variable, and
the nondimensional constant time lag of combustion.

If n is a constant independent of p, then f(p) is proportional to pn
This is the form of f(p) assumed by Crocco. The present formulation
5 is
of
the problem
value of n
If f(p) is
is
is
slightly
to be
more
and the
and is a function of p.
heat transfer from the hot combustion
general, in that f(p) is arbitrary
computed by using Eq.
considered as the rate of
(8.8)
97
gas to the propellant droplet, then the physical laws of heat transfer
indicate a value for n between | and 1.
8.2
Satche Diagram. Crocco called the instability
with a constant rate of injection the intrinsic instability.
combustion
of
the injection
rate is a constant not influenced by the chamber pressure p, then p, = 0.
Therefore the stability problem is controlled by the following simple
equation, obtained from Eq. (8.7)
^+
Equation
the same
chapters.
(1
way
In fact this
is
n<p(z
5)
(8.10)
by the Laplace-transform technique
the method used
method
by H.
I.
in
Ansoff. 1
of stability of combustion, the
problem
However,
fundamental
Thus a more direct approach is that
equation has no forcing term.
is,
as equations without time lag discussed in the previous
for the present
classical
n)<p(z)
(8.10) could be treated
If
of solving linear differential difference equations.
of the
That
let
<p(z)
sz
then
This
is
s
(l-ri) +ne-*
the equation for the exponent
=0
(8.11)
Stability of combustion then
s.
requires that the real part of s be negative.
Equation (8.11) can also be obtained by applying the Laplace-transform method to Eq. (8.10). By multiplying Eq. (8.10) by e~ss and then
integrating with respect to z from z

$(s) is the Laplace transform of ^(2),
to z
&, we have, noting that
- 8)<rdz -
<p(z
But
f
Therefore,
if
<p(z
sz
d)e~
dz
e~sS f
<p(z
S)er*c*-) fa
the initial conditions are such that
<p
for z
<
for the
so-called null initial conditions, then

[s
Hence we have Eq.
H.
I.
n)
+ ne-* ]3>(s) =
5
has the same "meaning" as the variable s

difference between the two is the
The only
made nondimensional by
also interesting to note that

1
(8.11).
in the previous chapters.
fact that here s is
(1
if
Ansoff, J. Appl. Mechanics
Eq.
(8.10)
(ASME),
the transit time Og
It is
were a nonhomogeneous equa-
16,
158-164 (1949).
98
term at the right side, then after applying the Laplace

transform to the equation, the resultant equation would also be nonhomoThen it will be seen that the function
geneous.
tion with a forcing
"
s
(1
n)
+ ne~
s8
the transfer function of the system, with <p(z) considered as an output.

F(s) is thus another example of a transcendental transfer function.
is
Crocco determined the value

of
two equations
for the real
of the complex root s by solving the set

and the imaginary parts of Eq. (8.11).
FIG. 8.1
However, if the point of interest is whether the system is stable or not,

one can again use Cauchy's theorem of Sec. 4.3 with advantage. Let
1
n
(8.12)
Then the question
of stability is
in the right half of the complex
determined by whether G(s) has zeros

s plane.
This question itself can be
answered in turn by watching the argument
of G(s)
when
s traces
curve enclosing the right-half s plane, as shown in Fig. 4.4. If the

vector G(s) makes a number of complete clockwise revolutions, then,
according to Cauchy's theorem, that number is the difference between
the
number
of zeros
and the number
of poles of G(s) in the right-half
Since G(s) evidently has no poles in the s plane, the number

of revolutions of G(s) is the number of zeros.
Hence, for stability, the
s plane.
make any complete revolutions as s traces the

Therefore the stability question can be answered by
vector G(s) must not

specified curve.
plotting the Nyquist diagram.
method to G(s) as given by Eq. (8.12),

inconvenient because of the complication caused by lag
direct application of this
however,
is

term e~ 3s
M. Satche proposed
1
a very elegant and ingenious method to

time
with
system
lag: Instead of treating G(s) break it into
treat such a
two
99
parts,
G(s)
gi(s)
()
(8.13)
where
gi(s)
e~* s
The vector G(s) is thus a vector with its vertex in gi(s) and its tail on
The graph of gi($) for s on the imaginary axis is the unit circle.
2 (s).
For
on the large half
within the unit
circle, gi(s) is
The graph
circle.
of gz(s) is the straight line (Fig. 8.1)
imaginary axis when

on the imaginary axis. When s
on the large half circle, g^(s) is a
parallel to the
s is
is
half of a great circle closing the curve

left.
A moment's reflection
show that in order for the vector
on the
will
(?($)
not to
make complete
revolu-
any value of the time lag 5,

^
g2
the g%(s) curve must lie completely
That is, for unconditional intrinsic stability,
outside the gi(s) curve.
tions for
It is easily seen
now
-n
>
>
or
>
(8.15)
that the separation of G(s) into two parts gi(s) and
The diagram
be called the Satche diagram.
g z (s) allows a great simplification of the respective curves.

of the
combination
When n >
gi(s)
the
and
gi(s)
gz(s) will
and
gz(s)
curves intersect.
Stability is
still
possible, however, if for g%(s) within the unit circle of Fig. 8.2 gi(s) is
to the right of gz(s).
This condition
is satisfied if
or
if
<
where
/2
1
M.
Satche, /.
Appl Mechanics (ASME),
3.16)
16,
419-420 (1949).
100
When
5*,
then with
w*
= V2n -
(8.17)
= 0. Therefore, when 5 = 5*, <p has an oscillatory solution with

5* and co* are thus the nondimensional critical time
the frequency co*.
critical frequency, respectively.
and
the
nondimensional
lag
ff(tw*)
8.3
System Dynamics
of
a Rocket Motor with Feedback Servo.
a system including the propellant feed and a feedback

In order to approximate the elasticity of
servo, represented by Fig. 8.3.
the feed line, a spring-load capacitance is put at the point midway
Consider
now
between the propellant
pump and
the injector.
Near the
injector there
Line Capacitance
Pump
FIG. 8.3
another capacitance controlled by the servo. The servo receives its

If the
signal from the chamber-pressure pickup through an amplifier.
feed system and the motor design are fixed by the designer, the question
is
is
whether
it is
possible to design
an appropriate amplifier so that the
whole system will be stable. Because there is no accurate information

on the time lag of combustion, a practical design should specify unconditional stability, i.e., stability for any value of time lag 5.
Let mo be the instantaneous mass flow rate out of the propellant pump
and PQ be the instantaneous pressure at the outlet of the pump. The
The pump
average flow rate must be m. The average pressure is p
.
characteristics
can be represented by the following equation:
m
If
of
(8.18)
the time rate of change of mass flow is small in comparison with speed
propagation of elastic waves in the liquid, but large in comparison
with the slow rate of change of the rotating speed of the pump, a is
simply the slope of the head-volume curve of the pump at constant speed
101
For conventional centrifugal
near the steady-state operating point.
pumps, a is approximately 1. For displacement pumps, a is very large.

For the constant-pressure pump, or the simple pressure feed, a. is zero.
Let mi be the instantaneous mass rate of flow after the spring-loaded
capacitance, % be the spring constant of the capacitance, and pi the
instantaneous pressure at the capacitance.
Then
m -mi =
PX^
where
(8.19)
p is the density of the propellant, a constant.
In the following calculation, the pressure drop in the line caused by

Then the pressure difference p
frictional forces will be neglected.
Pi
That is,
is due to the acceleration of the flow only.
""
where A, a constant,
is
the cross-sectional area of the feed
the total length of the feed line. Similarly,

pressure at the control capacitance,
is
If
the mass capacity of the control capacitance
mi
mi
if
is
2 is
line,
and
the instantaneous
C, then
=
-g-
(8.22)
is very close to the injector, the inertia

between the control capacitance and the
Since the control capacitance

of the
injector
mass
is
of propellant
negligible.
Then
where Ai is the effective orifice area of the injector. At can be eliminated

from the calculation by noting that at steady state the difference of
pressures $0 and p, or Ap, is
(8 - 24)
Equations (8.18) to (8.24) describe the dynamics of the feed system.

By a straightforward process of elimination of variables, a relation
between rii^ p, and C is obtained. To express this relation in nondimensional form, the following quantities are introduced:
102
and
*
4-
s 26 )
-
mdg
where
ff ,
the gas transit time,
dimensional equation relating
given by Eq.
is
<p,
p,
and
[1
[aE (P
|)
+ J] ^ +
ft
aJE (P
Then the non-
(8.9).
K is
(P
I)]
+ | J2
|)
TO
777
Trs+H" ^f
I
where
A
/c==0
/O 0*7\
(
8 2/ )
-
nondimensional time variable defined by Eq. (8.9).

of the servo control is specified by the composite of the
instrument characteristics of the pressure pickup, the response of the
2 is the
The dynamics
amplifier,
and the properties
of the servo.
Since
we do not propose
to
discuss the detailed design of the feedback servo, the over-all

of the servo control is represented
dynamics
by the following operator equation:
-,
F
where
is
(8.28)
the ratio of two polynomials with the denominator of higher
order than the numerator.
and
(8.28) are the three equations for the

Since they are equations with constant
coefficients, the appropriate forms for the variables are
Equations
(8,7),
three variables
(8.27),
^,
<p }
<p
By
and
ae sz
ft
be sz
substituting Eq. (8.29) into Eqs.
homogeneous equations
a[s
?{l
K.
+ aE(P + I)
[aE (P
for
a, 6,
and
(8.7),
ce 8Z
(8.27),
are obtained.
(8.29)
and (8.28), three

Thus we have
+ (1 - + ne~ ~ be~* =
+ \ JEs*} a + {[1 + a (P + |)]
Ss
ri)
+ i) + J] s +
[J
aJE (P
+ i JE] s + J J WJ
2
|)
+ S {a (P + |) + /* + | aJE (P + |) S + i JWJ c
2
F(s)a
a, 6, and c to be nonzero, the determinant formed by their

must vanish. This condition can be written as follows:
In order for
coefficients
+ \ JE [1 +
[aE (P
+ /] S +
|)
[1
nJE
(P
[n
+ J] +
I)
a (P
*)]
[I
a (P
a (P
aEP
})] s
103
J)]
a (P
(P
P}
|) *
+ Js +
a (P
=
J)|l
(8.30)
This is the equation for determining the exponent s. F(s) is

nized as the over-all transfer function of the feedback link.
plete system stability depends upon whether Eq.
have positive real parts.
now recogThe com-
(8.30) gives roots that
8.4 Instability without Feedback Servo.

The system characteristics
without the feedback servo can be simply obtained from the basic equaLet it be assumed that the polynomial
tion (8.30) by setting F(s) = 0.
5s
multiplied into e~ has no zero in the positive-half s plane, as is usually
the case. Then Eq. (8.30) can be divided by that polynomial without
introducing poles in the positive-half s plane into the resultant function.

That is, for the Satche diagram, one has again
0(s)
<7i(s) is
gi(s)
g*(s)
thus again the "unit circled
Ti
g z (s) is
e~**
now much more complicated:
+
*
gi(s)
/ 2Es 3 -
{aE (P
(aE (P
JE
{l
+ |)[1 +
I)
+ J]s +
+ a (P + i) +
(P/n)]
[1-f a (P
s2
(P/n)}
+ /} s + {l + a (P + i) +
(P/n)}l
(8.31)
The
s
intercept of gs(s)
in Eq. (8.31),
when
s is purely
imaginary
is
given by setting
i.e.,
ffi
(8-32)
104
Since
all
02 (0) is
the parameters n, a, and P are positive, the magnitude of

smaller than the magnitude of #2(0) given by Eq. (8.14)
now
Thus the
for the intrinsic stability problem.
effect of the feed
system
is to move the gz(s) curve towards the unit circle of 0i(s) in the Satche
= |, gz(s) is just tangent to the unit
diagram. For instance, for n
the intrinsic system without considering the propellant feed.

But with the propellant feed system, the 2 (s) curve will intersect the
circle for
unit circle, and the system will
a certain finite value.

destabilizing.
of
(s)
This
for large
is
The
become unstable
for time lag d exceeding
influence of the feed system
is
thus always
further confirmed by considering the asymptote
imaginary
s,
That
obtained from Eq. (8.31).
Therefore, for large imaginary
approaches asymptotically a line
gz(s)
s,
is,
parallel to the imaginary axis at a distance
The
to the left of the imaginary axis.
move
02 (s)
towards the unit
2P
effect of feed
is
system
again to
circle.
parameter n near | or larger than |, it

would be impossible to design the system for unconditional stability.
In the Satche diagram, the 0i(s) and 02 (s) curves will always intersect
It is thus evident that for the
without a feedback servo.

8.5
H(s)
Complete
Stability with
W+
{aE (P
i)
{i
JE [1 +
Feedback Servo.
a (P
|)
(aEP/n}(P
which multiplies into e" Ss in Eq.

half s plane, then the presence
s
}
J </W + J o/5 (P +
(8.30)
I)]
|) s
the polynomial
(JEP/2n}
{l
If
/*
a (P
J)
a (P
(P/n)}
(8.34)
has no poles or zeros in the right-
in the right-half s plane can be determined
|)
of zeros of the expression in
0i(s)
T(s),
Eq. (8.30)
from the Satche diagram with
e~ 5s
and
+
(P
+ |) + /] s +
[1
+a
(P
ce
(P
-s-
I)])
J)]'
#()
(8.35)

As
if
105
the path of Fig. 4.4, gi(s) is again a unit circle.

Therefore,
simultaneously the g%(s) curve is completely outside the unit circle,
s traces
there can be no root of Eq. (8.30) in the right-half s plane.

In other
if
the
of
the
transfer
function
link
servo-control
is
so
F(s)
words,
designed
as to place the gt(s) curve completely outside the unit circle (Fig. 8.4),
then the system
stabilized for all time lags.
is
FIG. 8.4
As an example, take
P=|
=4
This value of a corresponds to a centrifugal
Then without the servo

t
nupaber iw, for
g*(i<}
co
1 (2s
g * (S)
Of primary interest
control,
is
real.
6)
when
s is
a purely imaginary
Thus
(6
+ 3s + 9s +
+ 3s + 6s + 6
3
the behavior of gz(s)
^
1
for the propellant.
l)(2s
s
= -
g$(s) is
pump
(21
8co
3co
)(6
2
)
+
3co
2
o>
(6
(6
co )
21cu
+ 4o>
)(6
2
co )
> is plotted in Fig. 8.5. It is evident that for suffiOn the

values
of time lag, the system will be unstable.
ciently large
other hand, if the # 2 ($) curve can be changed by the servo control to, say,
This curve f or
2 (s)
= -2 ^
*
,
,
106
new
then, as plotted in Fig. 8.5, the
straightforward calculation
is
completely outside the
now
unconditionally stable.
(8.35) shows that the
curve
g%
Therefore the system
unit circle of gi(s).
from Eqs.
is
and
(8.31)
required transfer function F(s) for the servo link
4*
_
The servo
(s
LQ528) (s
is
Q.7164s
2.6304)
link has thus the character of an integrating circuit, first
discussed in Sec. 3.3.
pickup and
with given response of the chamber-pressure

an amplifier could be
If,
of the servo for the control capacitance,
designed to give an over-all transfer function close to that specified above,

the combustion could be stabilized by such a servo control.
As the second example, take
p=
^=1
=
E =
j = 4
a==0
the feed pressure p Q is thus constant and even when the flow
The case then corresponds to that of a simple presof propellant varies.
Without the feedback servo,
sure feed.
Since a
0,
g * (S)
When
s is
+ 4s + 4
2s 2
purely imaginary,
Q*W - -
1 (4
2co )(2
(4
1
_
This curve of
17co
(4
2co
2a>
4co
2
2
)
4
)
co
co
(4
(4
a, )
2
a;
(12
- 4co - (4 (2
2co
+ (4
(4
2
co )
)(12
2^
It
g% is plotted in Fig. 8.6.
is
4co
<o
co )
17co
4co
4
)
evident that without servo-
control the combustion will be unstable for sufficiently long time lag.
In fact, the system is even less stable than the system considered in the
example: it will become unstable at shorter time lag. The part of

the # 2 curve near co = 2 is of special interest.
Near co = 2, the cucve
comes so close to the unit circle of gi that if the value of time lag 5 is
first
such as to
almost
6 is
make
and
gi
undamped
f or
co
oscillation at
co
^ 2 very close to each

^ 2 can occur. This
other, then an
critical
value of
evidently smaller than the critical 5 determined from the true inter-
section of 02 with the unit circle at
co
~ 0.65.
For unconditional
stability, g% should be displaced out of the unit

the
same
"stable" curve as in the first example. The
circle, to, say,
transfer
function
required
F(s) is calculated to be
_
- -4.8o
,
0.8126)
(g
0.04337s
+ 2.6506)

The required servo
link
must then have the character
107
of a double inte-
Furthermore, the transfer function has two purely

grating
at
+ 2z. This unrealistic requirement on the amplifier
poles
imaginary
comes from the original feed-system dynamics and is due to the negleccircuit.
damping in the feed line. In any actual system, the

damping in the feed line will remove these purely imaginary
tion of frictional
frictional
-2
-1
FIG. 8.6
FIG. 8.5
poles of the required transfer function F(s)
and replace them by two
complex conjugate poles.

It should be emphasized that the advantage of using a feedback servo
to stabilize the combustion lies in its great flexibility in being able to
obtain unconditional stability with any value of time lag 5 or r. In as
much
as there are no accurate data on the time lag, the possibility of
unconditional stability is of very real, practical importance. Even more

than this, servo stabilization also makes it possible to design the system
From
for stability against any expected variation in the parameter n.
physical reasoning,
n probably takes
a value between
| and
1.
Take
108
the worst possibility, and design the system for unconditional stability
1.
Then the system will be stable for all expected values of n.
with n
Therefore stabilization by feedback servo can be assured without having

know the exact values of the parameters of the system.
8.6
General Stability Criteria for Time-lag Systems. In the preced-
to
ing discussion of servo stabilization, it is assumed that the polynomial

This,
H(s), of Eq. (8.34), has no pole or zero in the right-half s plane.
however, is not necessarily the case. In general then, one should first
investigate the
To do
number
of zeros
and poles
of
H(s) in the right-half
plane.
should be recognized that the polynomial in Eq. (8.34)

before the factor F(s) usually does not have zeros in the right-half s plane.
this, it
Therefore, instead of studying ff(s), one can study the ratio of H(s) to
That is, the number of zeros and poles of H(s) in the
that polynomial.
is
the same as the number of zeros and poles of the
s
right-half
plane
following function
H(s)
[iJW + \$JE(1 + a(P + 1)] + (JEP/2n)} +

+ (aEP/n)(P + |)} s + jl + a(P + i)
2
+
(P/n)}]
{aE(P
= l+L()
(8.36)
where
sF(s)
$J*Es*
JE[1
(aEP/n}(P
According to the Nyquist
+ L(s)
a(P
)]
+ I)} s +
criterion,
(l
the
(JEP/M)}
a(P
+ |) +
{aE(P
(P/n)
+ i)
(8.37)
j]
number
of poles
and zeros
for
be found by plotting the Nyquist

diagram of 1 + L(s) with s tracing the curve of Fig. 4.4. In fact,
if 1
L(s) or H(s) has r zeros and q poles in the right-half s plane, then
1
in the right-half s plane can
L(s) will carry out r
q clockwise revolutions around the point
traces the semicircle.
Hence the necessary information on H(s) can be
as s
obtained by plotting the Nyquist diagram of L(s).

When one divides Eq. (8.30) by H(&) in order to get gi(s) and # 2 (s)
as given by Eq. (8.35), q zeros and r poles are introduced in the
righthalf s plane.
The
q poles of L(s)
must come from
F(s), since the poly-
nomial in the denominator of Eq. (8.37) has no zero in the right-half

s plane.
Therefore the original expression in Eq. (8.30) also has q poles
in the right-half s plane.
Hence
in order for the original expression in
109
Eq. (8.30) to have no zero in the right-half s plane, g z (s) must make
= r clockwise revolutions around the unit circle. In
r)
q-{-(q
order for the stability to be unconditional, i.e., stable for all time lag, the
Therefore the general
gz($) curve should never intersect the unit circle.
unconditional stability criteria are,
first, the gz(s)
around
revolutions
curve must
must make
outside the unit circle; and, second, gz(s)
lie
completely
r counterclockwise
the unit circle as s traces the conventional
the right-half s plane.
These are the
criteria for stability

Curve for Positive
( Full
path enclosing
with the Satche
u>\
Dotted Curve for Negative u)
L(s)
(a)
Nyquist Diagram forL(s)
With Two Zeros for l+L(s)
in
Right Half $
plane
Stable Satche Diagram
FIG. 8.7
diagram. To determine r, one has to use the Nyquist diagram of L(s},

Eq. (8.37). Thus the stability problem for the general case requires
both the Satche diagram and the Nyquist diagram (Fig. 8.7).
It is evident that the stability criteria developed here using a combina-
and the Nyquist diagram are applicable to

any system with a time lag r. The stability of such systems always
reduces to the question of ascertaining whether there is root of
tion of the Satche diagram
M(s)
with a positive real part, where
e~TS
is
The
=
contains terms with the factor
(s)
from the previous discussions,
principle of the method, as seen
to divide
(s)
by the
coefficient of e~
M(s)
T
=
= nf
G(s)
in M(s), such that

.
,,
gi(s)
gz(s)
and
gi(s)
The curves
of g\(s)
and
e-7*
g$(s) as s traces
the right-half semicircle shown
in Fig. 4.4 then constitute the Satche diagram, with gi(s) represented
110
by the unit circle. To find out whether dividing M(s) by 1

L(s) has
introduced roots with positive real parts to the Satche diagram, we have
to plot the Nyquist diagram of L(s). The number of roots of M(s) =
with positive real parts can then be determined by Cauchy's theorem.
The 02 (s) function contains the transfer function of the feedback link
with an amplifier which is under the designer's control. g%(s) may also
contain transcendental functions of s coming from other parts of the
Since the transfer function of the amplifier in the feedback
two polynomials in s, it is difficult to compen-
system.
link is generally a ratio of
sate completely for the destabilizing effects of a transcendental function.

However, in the Satche diagram, the critical part of the g%(s) path is
But
that part close to the unit circle of 0i(s).
circle is generally obtained by small values of s.
part of
gr 2
series in
(s),
s.
(s)
close to the unit
Hence
for the critical
the transcendental function can be expanded into a power

The amplifier of the feedback link can then be designed
power series as an approximation. Thus the

the
destabilizing effects of the system in the critical
amplifier compensates
Of course, the performance of the system should be checked
region.
using a few terms
finally
by the
of the
stability criteria developed, using the amplifier design
This is the procedure suggested by F. E. Marble.

the original work 1 should be consulted,
characteristics.
details,
IF. E. Marble and D.
W.
Cox, J.
Am.
Rocket Soc. 23, 75-81 (1953).

}
For
CHAPTER
LINEAR SYSTEMS WITH STATIONARY

RANDOM INPUTS
In the previous chapters, the inputs to a system are considered to be
time L However, there are many engi-
definitely specified functions of
neering problems for linear systems with constant coefficients where the
inputs cannot be so definitely described. An example of such an engineering problem is the problem of the motion and the stresses induced
Here the
in the structure of an airplane wing in a turbulent air stream.
be
considered
to
the
air-flow
The
can
be
time-varying
pattern.
input
air-flow pattern cannot be described as a definite function of time but
has to be recognized as a random function of time, specified by certain
It is then evident that the output of the
statistical characteristics.
system, the stresses in this case, must also be a

can also be described only in statistical terms.
this chapter is
random function and

The first objective of
then to find a convenient method of calculating the statisthe output from the specified statistical properties of
tical properties of
the input. This forms an easy extension of the early investigations by

P. Langevin of Brownian motion.
Another example of random input is the so-called noise in control
The noise is introduced by disturbances and fluctuations beyond
signals.
the control of the designer. The problem of noise is a subject of much

research in connection with communications engineering.
There, the
central question is how to design the system so that the effects of the
unavoidable noise can be minimized and the useful information of the
signal not destroyed.
filtering in
different.
We
shall discuss this particular
problem
of noise
Chap. 16. The problem of this chapter is, however, somewhat

In our present problem, the random output is the only output
Our purpose in the design of the system, particularly the

the feedback servomechanism, is to obtain with a given input
an output of the desired statistical characteristics. We shall see that
of the system.
design of
the transfer-function method developed in the previous chapters remains

useful in the present task.
9.1
Statistical Description of
Random
Function.
system which generates a random function y\ (f).

ill
Let us consider a
Now
to formulate
112
the concept of a statistical description of such a random function, we

have to consider a great number of systems identical to the first. Such
The random functions gena group of systems is called an assembly.

The
erated by the members of the assembly are y\(t\ yz(f), y&(f), ....
random character of the function is exhibited by the fact that, although
the systems are identical, the value of the function generated by any
member of the assembly at any specified instant of time t is generally
different
instant.
from the value generated by another member at the same time

But we can ask for what fraction of the total number of the
This fraction will

dy.
systems y occurs in a given range y to y
when
to
t and will be
and
on
dy is small. This
dy
y
proportional
depend
fraction
is
the probability that y will lie between y and y

dy at time t.
is
the first
called
The
function
W\(yf)
Wi(y,t) dy.
It is written as
Next we can consider all the pairs of values of y

two given instants ti and tz. The fraction of the total numin which y occurs in the range y\ to y\ + dy\ at ti and in the
probability distribution.
occurring at
ber of pairs
range y% to y z
dy^ at
tion Wz(yi,ti ,yz,tz)
is
tz is
written as Wzfyifayzjiz) dy\ dy^.
The func-
called the second probability distribution.
Higher
probability distributions can be similarly constructed.
The fact that the above formulation of the statistical description of

random function depends upon observations carried out simultaneously on a very large number of identical systems may be objectionthe
able on the ground of practical difficulty in observation.
random function
However,
if
the
a stationary random function in the sense that all

statistical characteristics of the function are time independent, then the
is
all necessary obseris not necessary

a single system, for a very long period of time.
The record of observation can then be cut into pieces of length 0, 9 being
large assembly of identical systems
vations can be
made on
large in comparison with the characteristic time of the function.
Then
each piece contains the same statistical information about the behavior
of the system, since statistically the origin of the time scale is of no conse-
The different pieces can then be considered as an assembly of

quence.
observations on identical systems, and the various probability distribuFurthermore, these distributions now become
somewhat simpler: Wi will be independent of time t, and Wz will be
Hence for stationary
h.
dependent only upon the time interval r = tz
random functions, W\(y) dy is the probability of finding y between y and
tions can be determined.
+ dy,
Wz(yi,yz'jT) dy\ dy% is the probability of finding a pair of values
yi and yi + dy\ and between yz and yz

dy 2 at an interval of
time equal to r. Since the random functions of engineering problems
can very often be considered as stationary random functions, we shall
between
limit the following discussion to such

It
random
functions.
should be emphasized that these probability distributions
embody
RANDOM-INPUT LINEAR SYSTEMS
113
the information about the statistical properties of a random function.

"
"
may say that the probability distributions define the random
all
Or we
Of course, these distributions
function.
n are
not arbitrary but must
satisfy the following conditions:
Wn > because there no such thing as negative probability;

W symmetric with respect to variables y
(a)
is
0,
its
n is
(fe)
W*(yi,y*ri
This
is
(c)
obvious from the meaning of
it
Wt(y>,yi',T)
(9.1)
W$ as a joint probability distribution,
Higher distributions imply lower distributions.
Wi(yi,y*ri dy,
where the integration
is
that integration over
?/ 2
i.e.,
(y,)
That
is,
W,(y}
(9.2)
to be carried over all possible values of

also eliminates
y-
r.
Wi(y] dy
oo
2/2-
Note
Furthermore,
This simply means that the probability of
(9.3)
all
occurrences must be a
certainty.
9.2
From
Average Values.
the
first
the average value y of y can be found
y
Since
we have
{*
probability distribution Wi(y} }
yWi(y) dy
limited ourselves to stationary
(9.4)
random
functions, the
average value can be also obtained by taking the time average
That
The equality
(9.5), is
Eq.
of y(t).
is,
re/2
lim i /
y(f) dt
QÔO O J-0/2
(9.5)
of the assembly average, Eq. (9.4), and the time average,

a characteristic of stationary random functions. We shall
property repeatedly in the following calculations.

Equation (9.4) can be generalized to arbitrary powers of y.
utilize this
Thus
y Wi(y) dy
(9.6)
m n is
the
moment of the first probability distribution. From

and second moments, we can compute what is called the fluctuavariance, or mean deviation a
called the nth
first
tion,
(9.7)
114
<r
is
thus a measure of the "width" of the probability distribution Wi(y)

Similarly, the third moment gives a measure
y.
about the average value
More and more informa-
of the skewness of the probability distribution.
more moments
tion about W\(y) can be deduced as
some
tion.
cases, the
knowledge
For instance,
for &
m, k =
first
moments uniquely determines the
0, 1, 2,
(2fc
probability distribution W\(y]
distribution, or
distribu-
if
mab+i
then the
normal
.1
- iy*
is
the well-known Gaussian
distribution,
o"
Sometimes
For
are known.
of
(9 9)
_
v 2ir
convenient to choose the origin of the y coordinate

that
y vanishes, i.e., the origin is the average value of y.
way
this is done, we say the probability distribution is normalized.
it is
in such a
When
In this case, the square of the mean deviation

moment y 2 as seen from Eq. (9.7).
The most important average value derived from
is
simply the second
distribution W^(y\,y^]r}
the second probability

It is defined as
the correlation function R(r).
is
=
r
00
+ r)
y(t)y(t
oo
J/
00
?/ 12/2^2(2/1,2/2
For a stationary random function,
;r)
dyidys
(9.10)
from
this is clearly also obtainable
time averaging:
R(r)
lim
9-
The function R(r) thus

measured at two
eo
e/2
1
i
tf
gives a
/"
~ 0/2
y(t}y(t
measure
different time instants.
+ r) dt
(9.11)
of the interrelation of the y's

It is to
be expected that as
the time interval r increases, the interrelationship or

be weakened and, in the end, when r is very large,
"memory" must
y(t)
and
y(t
+ T)
be independent of each other. Then according to the rule of probability calculus, the second probability distribution is equal to the
product of Wi(yi) and Wi(y 2 ). Thus for large r,
will
i)Wi(y*) d yi dy z
For r
0, it is
(gy
(9.12)
obvious from Eq. (9.11) that

(0)
(9.13)

Now
of the
we
shift of the origin
scale,
r)
If
an arbitrary
since R(r) can be calculated with
time
115
y(t
r)y(t)
differentiate the equation with respect to T
R'(fy
and then put r
0,
= -y(W(t)
y(t)y'(t)
Therefore,
'(0)
= y(W(t) =
(9.14)
In these equations, a prime indicates differentiation with respect to time.

Thus the correlation of a random function with its derivative at the
is zero.
This means that the slope of the record of y
an
has
equal probability of being negative or positive.
any y
If we differentiate Eq. (9.14) twice with respect to r and then set
same time instant

at
0,
we have
fi"(0)
= -T*
y(f)y"(t)
(9.15)
This equation allows us to calculate the mean square of the derivative

of y from the correlation function.
Similarly, the mean square of the
second derivative of y can be calculated as

fl""(0)
77 *
(9.16)
Power Spectrum. Of special significance for our applications of

the theory of random functions is the notion of the spectrum of a random
9.3
Let us suppose that a function y(t) is observed for a long time

assumed to vanish outside the interval, then y(t) can be
function.
0.
If y(t) is
1
developed into a Fourier integral
where A(co)
from
y(f)
is
|_
Afa^'da
the amplitude of the frequency
(9.17)
co.
It
can be computed
the inversion formula:
by
A()
we denote
If
y(f)
A*(co) as the
y(t)e~^dt
=f
ZT J-e/2
complex conjugate
of
(9.18)
A(w), then, since
y(t) is
real, Eq. (9.18) indicates that
l*(o,)
Now we
=
lim
e-*
ex,
(9.19)
can calculate the average y in terms of
j - 9/2
A ()
as
y\t) dt
lim
e->*
= A(-w)
r-
re/2
5
/
o j-e/2
See for instance WMttaker and Watson,
/-
J~ * 7-
"Modern
Analysis," Sec. 9.7, p. 188,
116
By
the substitution
i
lim
o>"
-co',
we have
/-/-
re/2
'
r
lim
Af
we now
introduce the
new
A*f
03
variable
f,
dt
.A()A*(")
-r
e-oot77~oo7-M
If
e^~"
dudu"A(u)A*(u")
du
//
do)
co
defined as
then
Consequently,
lim
=
Therefore,
if
4-7T
lim
we put
$(o))
lim
(9.20)
|A(co)|
then
y*
$()
do)
(9.21)
The function $(0?) is thus a real function and is called the power spectrum
of the random function.
Equations (9.20) and (9.21) enable us to com2
pute the average value z/ from the Fourier coefficient A(co).
tion is the Parseval theorem.
Let us consider next the correlation function R(r).

we have
By
This rela-
combining
Eqs. (9.11) and (9.17),
R(r)
lim
e-
5
W
y(t)y(t
r) dt
J -0/2
i
lim
re/2
fr+vVJ-vJ-*
A()A(c/)e
dcoda/
7-6/2
Then, by an argument similar to the above, we have

R(r)
$()
COS wrdco
(9.22)

By
Eqs. (9.21) and (9.22) reduce to the relation of Eq.

differentiating Eq. (9.22) with respect to r and then setting
setting r
By
(9.13).
Eq.
0,
117
0,
According to the Fourier inversion
obtained.
is
(9.15)
theorem,
o /*
$() = *- /
JQ
Equations
(9.22)
and
tion function or the

called the
R(r) cos
(9.23) allow the
cor
dr
(9.23)
computation of either the correla-
power spectrum when one
known and
of these is
are
Wiener-Khintchine relations.
The power spectrum
contain peaks of the Dirac 5-function

is not zero, or, in the terminology
may
$(o))
This certainly is the case when y
type.
of electrical engineering,
when
there
is
a d-c term.
Then
+ $i()
where
(9.24)
d(x) is defined as
d(x) =
d(x)
such that
8(x) dx
8(x)
->
for x
for x
8(x) dx
oo
(9>25)
^
*
and
JQ
<x>
|-
For pure "noise," the peak at

0, corresponding to the d-c term,
be the only peak, so that $1(00) will be a regular function,
will usually
representing the really continuous spectrum. But it is also possible to

have several sinusoidal oscillations superimposed upon the noise. In that
case,
the power spectrum will have additional peaks at the discrete
frequencies of these oscillations.

9.4 Examples of the Power Spectrum.
We
spectra computed from correlation functions.

is
shall
show a few power
If the correlation function
given by a Gaussian curve,

R(r)
R(Q)<r**
(9.26)
then, according to Eq. (9.23), the corresponding
cos (wrje-*** dr
power spectrum
is
=
(9.27)
where
$(0)
=
note that as a
the correlation function becomes

and R(Q)
in such a manner that R(r) becomes
a 8 function. This means that subsequent y's are not correlated at all
and that the random function is the "most chaotic" of all. As a ~>
It is interesting to
zero for
all finite r,
<*>
<*>
the power spectrum
is
a constant, independent of the frequency.
This
118
most chaotic random function
Another example
is
random
the naturally generated
called the white noise
and often describes
variations.
the small isotropic turbulence in a fluid flow of
is
Krm&n
and
otherwise uniform velocity. It has been shown by von
Howarth 1 that the fundamental second-order correlation functions are
RI(T) and Rt(r): Ri(r) is the correlation function of the fluctuating velocity parallel to the
.A
tion at the
{'
T( f
F IG
mean
flow direction.
point, but
at time interval r apart (Fig. 9.1).
9.1
mean flow direc-
same space
U is
If
the
Z (T)
is
the correlation function of the
fluctuating velocity normal to the

a scale of turbu-
mean velocity and L is
lence, these correlation functions can
be approximately expressed as
(9.28)
(9.29)
By
using Eq. (9.23), the power spectra $i(o>) and $2(w) for velocity
and normal to the mean flow direction, respectively,
fluctuations parallel
are
*'
w - *'
(9
(0)
30 >
<"
where $i(0) and $ 2 (0) are the values of the power spectra at
They are related to #i(0) and S 2 (0) by
o>
0.
(9.32)
9.5
Power Spectrum. It is not necessary,

power spectrum from the correlation function.
possible to determine the spectrum directly from the
Direct Calculation of the
of course, to calculate the
Sometimes
it
is
random function
specified character of the
for example, the case
identical shape
where
y(f) itself.
y(t) consists of
Let us consider,
a series of pulses that have
and a constant repetition frequency but whose heights

some probability distribution. The heights of succes-
vary according to
sive pulses are, however, uncorrelated.
rectangular pulse.
If
ti(t)
y(t)
Krrnn
and Howarth,
This
is
shown
in Fig. 9.2 for a
represents a single pulse of unit height, then
aw(t
kT)
Proc. Roy. Soc. (A), 164, 192 (1938).
(9.33)

where
The
first
119
the spacing of the pulses and a k the height of the kth pulse.
step in computing the power spectrum is to determine the
is
Let 9
Fourier spectrum A(co) by Eq. (9.18).
A (to) =
= 2NT,
then
NT
NT
-I
dt
2ir
ak e- iukT
>
2L/
where a (ai)
is
a(co)
Z^
-N
27T./-
-N
the Fourier spectrum of the single pulse,
(9.34)
For a rectangular pulse

xv f,.A
width 2e and unit height,
of
(9.35)
7T
According to Eqs.
(9.19)
and
(9.20), the
"
$() =
To
to
()|* lim
CO
power spectrum
is
N N
(9.36)
carry out the limiting process in Eq. (9.36), it will be convenient first
average of the whole equation. Since the power
make an assembly
FIG. 9.2
the same for every
member
of the assembly, the left

the assembly. The
over
changed by averaging
side
will
such
an
of
be
right
averaging process.
simplified by
Eq. (9.36)
Let a be the average of a& and 0,1, and o^ the average of the square of
spectrum
$(aj) is
side of Eq. (9.36) is not
ak and
ai.
aifli
We
Then
(a*
a)(ai
5)
a[(ak
5)
(ai
a)]
(a)
substitute this expression into the right side of Eq. (9.36)
and then
average over the assembly. Then since the successive heights of the
pulses are not correlated, the assembly average of (a*
d)(ai
a) is
zero, unless
When k =
I,
the assembly average of
(a&
a) (ai
a)
120
is
a2
(a)
Thus the
li
>
>
term under the limit sign
vV
first
0)( fl
( ajfc
"^^
5) e
-M>(*-Z)T
is
a2
__
()2
-j^ _jv
The assembly averages
of
and
a)
(a&
a)
(o?
are obviously zero.
Therefore, finally,
N
/
4r
[a
The sum
(a)
2N
(a)
in Eq. (9.37) will be
(9.37)
Jim
1
for
co
= 2mr/T when n
is
an
second term of Eq. (9.37) will be

>
the sum will be finite, and for
in the limit, the
integer.
Hence,
infinite.
For other values
of
the limiting value will be zero. Clearly then, in the limit, the second
term has the character of a series of peaks, or 8 functions, at the fre-
To determine
quency 2mr/T.
we have
TT
<
the
coefficient
of
these
functions,
to calculate the area under the curve for the typical interval
o)T
<
Thus by integrating the sum, the required area
TT.
j
1
-x/r
is
cos co?
Since the area under the
d function, according to Eq. (9.25), is unity,

the required coefficient is 2ir/T.
Therefore, finally, the power spectrum
for the specified stationary random function is
[a
where w
(a)
(9.38)
is
the frequency corresponding to the fundamental period T Le.,

,
Hence the power spectrum contains a continuous part that has the same
shape as the power spectrum of a single pulse. The intensity of this
continuous spectrum is determined by the square of the mean deviation <r
of
the pulse heights.
frequencies ftw
for
by the spectrum
There
n an
is,
integer,
in addition, a discrete spectrum at the

where the intensities are also determined
of the single pulse.
Let us consider next a
series of pulses that have an identical shape and

height but a repetition period varying around an average value T. The
spacing between pulses will be
e,
with
distributed according to a

The average value
specified probability function P(c).
The
cation zero.
successive
e's
121
of
c is
by
impli-
be assumed to be uncorrelated.
will again
Figure 9.3 shows such a random function with rectangular pulses.

Therefore the random function is represented by
where
-kT-
y(t)
represents the single pulse.
??()
(9.40)
6*)
According to Eq.
(9.18),
with
= 2AT,
N
a(w)
and
is
the Fourier spectrum of the single pulse given by Eqs. (9.34)
(9.35).
The power spectrum
according to Eqs. (9,19) and
of y(f)j
nnnn
nn
FIG. 9.3
(9.20), is
then
$() =
lim
|<x(co)|
JVW
J-
ee
yyr
oi|
^^V
L ^-/
N
Now
let
us introduce the function
*.i^-ft,
(9.41)
<<"-/
defined
x()
by
(9.42)
x(co) is
sometimes called the
transform of P(e).
where x*(w)
We now
is
We
characteristic function of
shall write
and
the complex conjugate of
x(co)
and
is
assembly average
e's
first.
The
limiting process
are not correlated.
the Fourier
equal to x(
substitute this expanded form into Eq. (9.41)
fact that the
is
then
The
is
and make an
greatly simplified
by the
final result is
{[1
a> is the frequency defined by Eq. (9.39).

Here the shape of the
continuous spectrum and the intensities of the discrete spectrum are no
where
122
longer determined solely by the spectrum of the single pulse but depend
also on the characteristic function of the distribution of e.
As the third example
of direct calculation of the
power spectrum,
consider the stationary random function y(f) indicated in Fig. 9.4.

function takes the value of either +1 or
1 with the interval T.
The
T,
distributed according to a specified

It is further specified that
probability distribution P(T), where T > 0.
the successive time intervals T are not correlated. Let us denote the
however,
is
not a constant but
successive intervals
is
by Tk) where k
2,
1,
and take the time
FIG. 9.4
of integration in Eq. (9.18) to be
period
T is the average time interval defined
f =
is
to
= NT, where
TP(T) dT
(9.44)
Nf
tk
by
Then
where
from
the time instant at the end of the
/cth interval.
The above
expression can be rewritten as
VM
IT
[ii
t-co
<
11 (-i)*^* +
k=l
Hence, by Eq.
we have
(9.20),
the power spectrum as
*() =
Now
let
us consider
VV
^!>->^c
>
&',
say
AJ
- 1)fc
+ m,
.
(9.45)
then
.
-^^m
(9<46)
Since the successive intervals are uncorrelated, the probability of occurrence of the product to the right in Eq. (9.46) is the product of the probability of occurrence of each factor.
function
x()
of the distribution
we
If
introduce the characteristic
P(T),
dT
(9.47)
123
Therefore the assembly averthe average value of e~ iaT

m
is
In the double sum
of
the
simply [x(^)]
Eq. (9.46)
product
age of
such
are
there
of Eq. (9.45)
products, all with the
approximately
then
x(co) is
w
sign
(
l)
m
l)
terms
Hence
[x(w)]is
the limit these products contribute a term

The sum of all such
1 to <*> in the limit.
in
ranges from
then
The contribution from terms with k
the contribution from terms with k

there
sum
to the double
is
of
> k is just the complex conjugate of

> k These contributions are all
f
Eq. (9.45) with one exception, the case
and V give a contribution corresponding to

Hence Eq. (9.45) is finally
within the square bracket.
These values
fc'.
where
61
means the "real part
imaginary parts
Eq.
of k
(9.47).
of %(u)
be
of
77
4>(w)
Let the real and the
the expression.
and ^(),
respectively, as
shown
in
Then
.
If the distribution
41-
P(T)
is
2
<}!>
(co)
iA (co)
Poisson's distribution,
P(T)=e-"*
where f
(9.49)
the average time interval defined by Eq. (9.44), then the

power spectrum of such a randomly switched function of unit amplitude
is
is
The complete lack of any regular periodicity in the random function

considered makes the power spectrum continuous and smooth without
any
peaks of the previous examples.

Probability of Large Deviations from the
of the sharp
9.6
Mean.
If
the random
the stress in a structure, then it is not sufficient to know the

average value of the stress. For safety, we shall want to know the probability of occurrence of stresses exceeding the specified working stress of
function
is
the structural material,
i.e.,
the probability P[\y\
>
k] of
the magnitude
124
y exceeding the value
random function
of the
distribution function W\(y)
p[\y\
But
tion
is
the
first
probability
very simple:
many
However, even under these restricted circumstances, it

estimates of the probability
possible to give general but broad
deviation
still
fc
If
k.
known, the answer
distribution funcengineering problems, the probability

value
mean
is
the
known
What
is
y and the mean
not known.
in
is
>
is
<r.
is
of
(*)
FIG. 9.5
For instance, if g(y) is a

occurrence of large deviations from the mean.
is by definition nonnegative,
since
then
of
function
W\(y)
y,
nonnegative
Q(uWi(y)
The
last integration is to
g(y)
>
K.
But the
is
be carried over
Then according
is
the
dy
(9.52)
satisfying the condition
>
K\.
Thus
>K]<=
(y
(9.53)
Now let
-vY
to Eq. (9.7),
deviation from the mean.
mean
(9.53) gives the
TFifo)
all y'B
the so-called Chebyshev inequality.

g(y)
where a
>
last integral is just P[g(y)
P[g(y]
This
Let
fcV 2 then Eq.

;
Bienayme-Chebyshev inequality
P[\y
-y\>
M<
(9.54)
The Bienayme-Chebyshev inequality is known to be too broad for

most practical applications, and the upper limit given is, in general,
much
too high.
only a single
modal"
ity, let
sharper estimation can be given for W\(y) that has

This estimate for "uniso-called mode.
maximum, the
distribution
is
that due to Gauss.
To prove
us consider a function w(x) (Fig. 9.5) which
Gauss's inequalis
monotonically
125
can be considered
decreasing in the range x > 0. It is seen that w(x)
to a; = X Q)
from
x =
constant
as a superposition of functions which are
>
Then
XQ.
>
for x
and zero
for
Let v(x)
any
K>
<
for
< K <
if
w
2
0(3)
f
JK
The maximum of this quantity

Thus it is generally true that
superposition,
Now
XQ
and
is
for
&=#
for
- K)
(zo
K within
v(x)
dx
<
w(a)
<to
<
the range specified
is
we have
"
4 f
jO
x*w(x) dx
any unimodal distribution with the abscissa x

the mode. Then
consider
where y
v(x)
so,
By
<
x$,
v(x) dx
But
#o,
and
By adding
where
^ is
these expressions,
mean
the
deviation from the mode, defined by
fo
Let
K=
fc^;
then
we
the distribution
dy
(9.55)
obtain the Gauss inequality
P(\y
If
_ ytfWM
yo|
is symmetrical,
>
M<
then
&r
()
ffi
=&
G"
3
and Eq.
(9.56)
reduces to
Equation
Often
a sharper estimate of the probability than Eq. (9.54).

be possible to assume, at least approximately, a Gaussian
(9.57) is
it will
distribution.
the error
Then, by using the asymptotic expansion of
126
function,
it is easily
shown that
-f\>
P[\V
^=
for k
(9.58)
for k
3, the probability
the
that
is only 0.002.
probability is less
say only
Equation
the
that
will
than 0.1111, while Eq. (9.57)
probability is less than
say
is
of
course caused by the
in
estimates
The difference
these
0.0493.
This
is
a very small probability.
For instance
(9.54) will
different degree of information available to the estimation.
general the assumption, the broader the estimate.

9.7
Frequency of Exceeding a Specified Value.
If
the
The more
random func-
the design is to be based upon the

"
"
of stress, i.e., the
a
certain
value
of
occurrences
fatigue stress
repeated
of the material, then it is necessary to know the probable number of
tion
is
the stress in a structure and
if
times per unit time the random function will exceed the value y =
The quantity is evidently one half of the number of times per unit time
Let ATo() denote
the random function will pass through the value
.
the
S.
number
of times of passing.
We
0. Rice. 1
shall follow his
This number was
first
computed by
method.
Let W(y,y ) dy dy be the joint probability of having the random function y between y and y
dy and the time derivative y' between y
and y' + dy' at the same time instant. This probability can be also
interpreted as the fraction of time per unit time that the

y and
random function
its
ranges.
derivative y will simultaneously have values within the specified

But crossing the interval dy takes the time dy/\y f \. Hence
and y' per unit time

the expected or probable number of crossings at
f
f
is equal to the quotient of W(%,y') dy dy and dy/\y'
or \y'\W(%,y } dy'.
,
The number
JVo(f) is obtained
by integrating over
all
Thus
\y'\W(S,y') dy>
But Eq,
(9.59)
shows that for any differentiate random function, y and
(9.15)
y' are not correlated.
Then according
to the general laws of probability

the
calculation,
simply
product of the first probability distributions Wi(y) and W(y'). Hence Eq. (9.59) can be written as
(y,y') is
TO = W (&
l
When W(y
is
S.
\y
\W(y') dy'
(9.60)
symmetrical, Eq. (9.60) can be further reduced to
TO = 2Wi($ /
1
y'W(y'} dy'
0. Rice, Bell System Tech.
/., 23,
for
W(y
symmetrical
282 (1944); 26, 46 (1945).
(9.61)

If
is
W(y')
mean
a Gaussian distribution, with the
deviation
127
then,
<r',
according to Eq. (9.9),
(9.62)
The mean deviation

by using Eqs.
(9.15)
er'
can be computed from the power spectrum
and
*$() da
If
W\(y)
deviation
is
<r,
(9.63)
Gaussian distribution with the mean y and the mean
also a
then with Eqs.
and
(9.7)
(9.21)
we
obtain
r
'
W 2 $(a>)
-T
$()
(9.22):
;o
(9.64)
<T
the formula given by Rice.

Response of a Linear System to Stationary Random Input. We
shall finally give the answer to the question we posed in the introduction
This
is
9.8
to this chapter:
with constant
Given the stationary random input to a linear system

what is the output? From the exposition of
coefficients,
the elements of the theory of random functions in the previous sections,

evident that the key to this question is the calculation of the power
it is
spectrum
of the
output from the power spectrum
When
of the input.
the power spectrum of the output is known, we can easily compute the
correlation function by Eq. (9.22), and the mean-square value by Eq.
Then the probability of large deviations from the mean and the
(9.21).
frequency of exceeding a specified value can be estimated by methods

given in Sees. 9.6 and 9,7. For many engineering problems, knowledge
of these characteristics of the output is generally sufficient.
*
Let the input be x(t) with the power spectrum $(w) and the correlation
function Ri(r). Then by Eqs. (9.21) and (9.22), we have
<)
Jv
da
Ri(Q)
(9.65)
and
Rib)
jo
where the relation $(
$()
o>)
the output be
Similarly,
correlation function Ro(r).
let
cos
cor
dr
MT
/
J-
$(w)e*
$(w), as seen from Eq.
y(f)
dco
(9.23),
(9.66)
is
used.
with the power spectrum gr() and the
Then
128
and
=|
As
at
g(o>}edu
00
(9.68)
before, let h(f) be the response of the linear system to a unit impulse
For a process started at
0.
the output can be written
as
y(t]
where
for
= r. Since
impulse applied at the instant t
the upper limit of the integral can be extended to
X(T) dr is the
<
0,
y (t)
The
I
00
x (r)h(t
correlation function
r)
dr
"
00
rc(*
h(t)
<
w)A(w) <fe
Then
(9.69)
thus
(r) is
aft
- wM* +
u'}h(u}h(u')
dudu'
But
^0
ttX*
+T-
Hence by using Eqs.
w')
(9.66)
x(t)x(t
and
+ T + U-U') =
(9.68)
Ri(r
+ U-U
Now
if
the transfer function of the linear system,
is
^(s)
Thus
of h().
F(iu]
Hence Eq.
(9.70)
1^
g(u)e
du=
where the fact that
Therefore the power spectra

g(a)
[cf.
Eq.
e^ A(w) d^
#(co)
oo
and
it is
the
(3.50)],
as
$(u)F(iu)F(-ia)
iar
du
and *(w) are related by the equation
= *()F(i)F(-iw) =
F(iu>)
(9.70)
tt
now .considered
can be
ic*
oo
we have
a
Laplace transform
F(ia)
(9.71)
|F(zco)| $(co)
are complex conjugates has been
used.
Equation
spectrum
Even when
table, the
power spectrum of the output from the power

and the frequency response of the linear system.
(9.71) gives the
of the input
the frequency response
power spectrum
of the concept of transfer function
demonstrated.
It is
is
given in a graph or in a numerical
g(w) can be easily calculated.
The
and frequency response
usefulness
thus again
of interest to note that since F(iw) generally vanishes
is

for
a?
faster
129
oo, the output power spectrum 0() goes to zero for co

than the input power spectrum $(&). This has the effect of
"smoothing" the output function.

9.9
Second-order System. As a simple example, consider the
system to be of second order. Then the equation of motion is
m
The
w+
+ky =
jt
ms*
where
coo is
cs
(9 - 72)
x(t}
transfer function F(s) of the system
thus
is
linear
($ /co )
the natural frequency of the undamped system and f is the

damping to the critical damping, defined by Eq. (3.38),
ratio of actual
Therefore
F(i(j})F(
v
The power spectrum
If
we
i(ji)
77777
\9
& 2 {[(<VW)
of the
output
is
--
i"T9
1]
+
i
TT?7
2
4f
vJT
(oVW
thus
are interested in the mean-square output of the linear system,
Eq. (9.21) gives
*&**__
+
^^
,
4fWa,o)
I]
Now
is very small, the denominator of the integrand in Eq. (9.74)

if
Therefore, if $() is a slowly varying
very nearly zero at cc = o>
function, then
is
-s
o;oo
~--
=
'
This equation shows that
mean-square output
will
if
"*
M
/
--
"
the damping coefficient
become
infinitely large.
n n^
(9 - 75)
c vanishes,
When
the
c is zero, the
zW
This phenomtransfer function F(s) has the purely imaginary pole

enon of infinite output occurs in general whenever the transfer function
a linear system has a purely imaginary pole. Therefore, for satisrandom input, the condition on the system transfer
/unction is that all poles of F(s) should have negative real parts. This
of
factory operation under
fundamental requirement on the system property
is
then identical to
that for conventional input functions.

Other improvements on the output behavior can generally be accom-
130
plished
by further modification
instance,
it is
For
system transfer function.
of the
quite conceivable that the function <>(coo)/Uo i*1 Eq. (9.75)

has a minimum at a reasonable fre-
quency co* as shown in Fig. 9.6. Then

be possible to reduce the out-
it will
put random fluctuation by making the

*
This
system operate effectively at
FIG. 9.7
FIG. 9.6
can be accomplished by a simple proportional servo feedback, indicated

in Fig. 9.7.
Then
we have
instead of Eq. (9.72)
or
(9.76)
The natural frequency
system can then be made to be u* by
of the
requiring
cof
k
(9.77)
Therefore, by proper choice of a, the output can be reduced.

9.10 Lift on a Two-dimensional Airfoil in an Incompressible
As a second example, consider a
lent Flow.
moving with a constant velocity
flat,
Turbu-
thin airfoil of chord c
through turbulent
air.
Let x
lie
along the chord, z along the span of the wing, and y normal to the span
and cord. The components of the fluctuating turbulent velocity u, v,
and
are assumed to be small in comparison to U.
Because
of these
turbulent fluctuations, a time-dependent apparent angle of attack a exists

at the airfoil, and, hence, fluctuating lift forces are produced.
The fluctuating angle of attack a
is
given by
v
as long as the fluctuations are small.

a(f) now plays the role of the
"
The
is
function.
the
forcing
response"
fluctuating lift of the airfoil, or,
better, the
Liepmanu.
1
lift
coefficient Ci(t).
This
is
a problem studied
H. W. Liepmann, J. Aeronaut.
Sci., 19,
793-801 (1952).
by H. W.

To
find the
mean square
Cf(t) of the lift coefficient, it is first necessary
This has already been done
to define a transfer function for the airfoil.
In fact the frequency response F(iw)j with v as the input

coefficient Ci as the output, is specified by Eqs. (3.54) to
in Sec. 3.7.
and the
131
lift
(3.58).
Turbulent fluctuations
that
is,
u,
seems
it
v,
and
are,
however, essentially three-dimensional

For the first analysis,
y, z, and t.
w will be functions of x,
component v and its dependence

Thus, in turbulent flow we consider a fluctuating velocity
attack of the form
sufficient to consider only the
upon x and
or angle of
t.
v(x,t)
a(x,t)
If it is now assumed that the turbulence pattern does not change appreciably during a time of the order c/U, then the turbulent angle of attack
depend upon t
(x/U) only, and Sear's result, given in Sec.
can be applied. This assumption is frequently made in turbulence
will also
3.7,
analysis
and requires
fluid velocity
change of the
of
essentially the condition that the rate of
by following a
fluid velocity at
Cf
where
$(co) is
fluid particle is small
change
of
compared with the rate
a fixed position.
With this assumption,
*
47T
$(co)|?(&)|
&>
(9.78)
the power spectrum of v/U.
According to Eqs. (9.31)
and
(9.32),
Furthermore, Liepmann has discovered that
J9?(&)l
can be approximated
by
lî "
1
(9 80)
-
Thus
\
1
47T
l+3u*
9
2
jft2
o
(1
+
;
2 2
v> )
du
r
1 + W
~.
where
This dependence of the mean-square

is
shown
in Fig. 9.8.
lift
coefficient
upon the parameter
132
Evidently,
if
c/L
0,
we have an
airfoil of small
chord in a large-scale
turbulence, and
airfoil behaves in a quasi-stationary manner with a lift slope of 2?r.

the other hand, c/L becomes large, we deal with an airfoil of long
on
If,
chord in a small-scale turbulence. It follows from Eq. (9.80) that Of -> 0.
That is to say the "gusts" cancel each other out completely, and the net
The
lift is
zero, as
might be expected.
1.0
0.8
0.6
0.4
0.2
10
FIG. 9.8
9.11
Intermittent Input,
aerodynamic buffeting
That
is,
is
the edge of any
of
phenomenon
major importance for
the so-called "intermittency" in
wake
fluctuates with a large-scale
wake
flow.
motion so
that a point situated near the edge will sometimes be within the wake
and sometimes outside. For a tail surface situated near the edge of the
wake of a stalled or partially stalled wing, this "intermittency" may

thus be extremely important in determining the lift forces on the tail
wing. For a crude idea of the effect, consider the flow at the tail as a
region of uniform
down- wash switched on and
off at irregular intervals.
probably a good model for the conditions in the wake of an

intermittently stalling wing. If the probability of switching over from
Such a flow
is
one region to the other is assumed to be governed by Poisson's distribuone can then apply the power spectrum of Eq. (9,50) with some
The mean deviation is not unity but the mean angle
modifications.
tion,
\/v*/U; and the average time during which the forcing function
switched on is the mean interval f. Thus the power spectrum is
is
(cof/2)
(9.83)

Then the mean-square
lift
coefficient is
approximately
f or
This relation
y
>
co
is
plotted in Fig. 9.9.
are, of course, the
133
The
f\
2irc
-r-=
limiting values at
same for the case studied
f\
(9.84)
and
in the previous section.
10
9.12
Servo Design for Random Input. We have already indicated, in
connection with the response of a second-order system to random input,
the possibility of improving the behavior of the system by servo control.
In that instance, however, the feedback mechanism is rather primitive

is of the same order of magni-
in that the feedback control force required
tude as the input forcing function. In a more practical design, this feedback mechanism can be made much more subtle, so as to reduce the
For instance, a wing in a turbulent flow can be

by a feedback servo which moves the hinged flap. The force
control force required.
controlled
necessary to
aerodynamic
thought
move
the flap could be quite small in comparison with the

such movement produces. The servo link can be
effects
of as that indicated in Fig. 9.10.
The
The input random function
aerodynamic transfer function

FI(S) is the relation between the turbulent air stream and the aerodynamic
This function is approximated by
lift due to the turbulent air stream.
is
the turbulent air stream.
first
that of Eq. (3.56). As a result of the changing

wing is subject to vertical and torsional motion.
and moment, the

The relation between
lift
the aerodynamic forces and the wing motion is given by the structural
The wing motion will have two effects. There
transfer function F 2 (s).
134
are the aerodynamic forces due to
dynamic
transfer function
().
wing motion through the second aeroThis
is
the
first
This
feedback loop.
not under designer's control. The designer's control

loop, however,
The wing motion can be used to generate flap
is on the second loop.
is
motion through the transfer function
F 4 (s).
The
flap
motion
will again
FIG. 9.10
Thus
generate aerodynamic forces through the transfer function F&(s).

the input and output relation is as follows
:
Y =
or
J=F
The
over-all
( S)
system transfer function
(9.85)
(s)
can thus be modified by
changing the servo transfer function F(s).

If $(w) is the power spectrum of the input x the power spectrum g(u)
of the output, or the wing displacement y(f), is then, according to Eq.
}
(9.71)
-iw)
(9.86)
F8 (s) is given by Eq. (9.85). Now it is quite conceivable that in

order to maximize passenger comfort in the airplane, we would wish to
make the acceleration y"(t) of the wing as small as possible. This means
where
we have to minimize y 77*.

f
y'
Equation (9.16) shows that the mean square of

can be calculated from the correlation function. But the correlation
function can be calculated from the power spectrum by Eq. (9.22).

combining these equations with Eqs. (9.85) and (9.86), we have
y"*(t)
To
By
(9.87)
Jo
The minimization is to be carried out by modifying the servo transfer

F 4 (s). To do so we can construct the servo transfer function
function

with unspecified parameters.
and F 5 (s)
Then with
all
135
other transfer functions
and the input power spectrum given,

Fi(s), Fz(s), Fs(s),
2
be
can
as
Eq. (9.79), y" (t)
computed as a function of these unspecisay,
The problem of minimization is then an ordinary
fied parameters.
minimum problem with respect to these parameters, and the condition of
fixed
minimization determines the parameters. The resultant servo transfer

is then the best transfer function for the purpose of maximizing
function
passenger comfort.
The above
discussion
for a specified purpose
is
but one example
and
specified input.
optimum servo design

As another instance, the
of
design condition could be the minimization of the mean square of the

elastic stress induced in the wing structure by the turbulent air stream.
Then the system
transfer function will be a different one, but the general
formulation of the problem remains the same. Such a problem of quantitative optimum design can be considered as one step beyond the mere
requirement of stability and other qualitative criteria of servomechanisms

introduced in the previous chapters. This general concept was perhaps
Boksenbom and D. Novik. 1
first
formulated by A.
this
problem again in Chap.
A.
S.
S.
Boksenbom, D. Novik,
16.
NACA TN
2939 (1953).
We shall touch upon
CHAPTER
10
RELAY SERVOMECHANISMS
If there is an on-off relay in the servomechanism, the system is called
a relay servomechanism. As pointed out in Sec. 6.3, the one great
advantage of a relay servomechanism is its low cost. However, since
the output of a relay is not proportional to the input, i.e., the input-output
relation is not linear, the behavior of a relay servomechanism cannot be
we
In this chapter,
analyzed by a linear theory.
approximate
an
and other
shall first present
theory for investigating the stability of relay
similar servomechanisms, based again on a modification of the Nyquist
In the latter part of this chapter, the more advanced and

problem of how to utilize the inherent nonlinear character-
criterion.
more
difficult
istics of
a relay to achieve superior performance from the servomechanism
Unfortunately, this particular subject is still far from

being completely investigated; the general solution is yet to come.
10.1 Approximate Frequency Response of a Relay.
Let us consider
will
be discussed.
a sinusoidal input x(t) of frequency

x(f)
The
and amplitude
a,
a sin ut
(10.1)
characteristics of the relay will be idealized in that
and the action
considered,
is
co
no time
On
lag.
no time delay
is
considered to be instantaneous, i.e., there

the other hand, the hysteresis of the relay action is
is
When
the input is positive and increasing, the relay output

changes from zero to the full value of A at x = b. When the input is
= c.
positive but decreasing, the relay output changes from A to zero at
included:
b is greater
current.
than
Pull-in
c.
b is called
and drop-out
the pull-in current and
The input-output
c, respectively.
It
as that of Fig. 10.1.
is
the input and the output.
the drop-out
evident that there
&
gja-1
The output square waves have
and
is
a phase shift between
6 is
sin- 1 -1
2[
relation can then be presented
This phase lag
for negative input occur at x
(10.2)
^
a]
a time duration of 2a/w, where a
is
given by
~
2
+6-
sin- 1
1
2
TT
1
|_
136
sin
sin- 1 -1
aj
(10.3)
The period
of the
output
is
137
the same as that of the input and
is
equal to
27T/CO.
Now
the output
y(t)
can be analyzed into a Fourier
y(t)
The
= Y
coefficient ai of the first
a n sin [n(ut
harmonic
is
series,
g)]
(10.4)
simply
sm
where a
is
given
the relay output
by Eq.
is
(10.3).
In a relay servomechanism, Fig.
the correction signal which actuates the servo.
10.2,
The
FIG. 10.1
y(t)
FIG. 10.2
servo has generally the property of a filter and greatly reduces the
importance of higher harmonics. Therefore, as an approximation, we
may neglect all, the higher harmonics and consider the output to be
ai sin
(co
6).
Then
the ratio of output to input
4A
sin
ira
is
a
(10.5)
Fr (iu)
can then be considered the frequency response 6f the relay. Of

True frequency response,
it is not the true frequency response.
as defined in the previous chapters, is a function of frequency only, not
course
a function of the input amplitude. The frequency response of Eq.

On the other
(10.5) is, however, a function of the input amplitude a.
hand, Fr (iu) is not a function of the frequency co. Therefore the func-
138
tional notation
shall
keep
When
adopted
is
not appropriate; but for easy identification we
the notation in spite of this.
the amplitude a of the input
(10.2), (10.3),
and
the amplitude a
The
response.
very
large,
we
have, from Eqs.
(10.5),
W^
When
is
)==
is
MI
7T
f or
a ->
oo
(10.6)
(I
very very small, the relay will not give any

a = b. Then
cutoff point occurs at
for a
(10.7)
These limiting values of the relay frequency response are thus completely
determined by the relay characteristics.
Let us assume for the moment that
10.2 Method of Kochenburger.
to the relay
the amplitudes a of the harmonic components of the input
F
the
of
relay is a comThen the frequency response r (iu}
are all equal.
If Fi(iw) is the frequency response
plex constant, given by Eq. (10.5).
the over-all frequency response
then
of the rest of the circuit in Fig. 10.2,
of
If we apply the Nyquist criterion

traced on the complex plane by
curve
the
then for stability
*> must "enclose" the
to
from
point -1.
as co varies
the forward circuit
of Sec. 4.3,
is
Fr (i<*)Fi(iu).
l/[F r (io))Fi(ia)]
In other words the curve 1/[F r (ia)Fi(ia>)] must cross the real axis to the
But for constant input amplitude a to the relay,
left of the point -1.
a constant, and the above-stated condition for stability is

the frequency-response curve l/Fi(ia) to enclose
equivalent to requiring
This is the basis of the
to *>.
w
varies from
as
-~F
the point
T (iu),
1
of Kochenburger for determining the stabilmethod
frequency-response
2
servomechanism. Dutilh developed a similar method
of a
Fr (iu)
is
relay
ity
independently.
a
Kochenburger argues that when the amplitudes
of
the harmonic
that is necessary
input to the relay are not equal, all
components
in the previous paragraph to
is to apply the stability condition deduced
The locus
to <*
from
all values of F r (iw) as the amplitude a varies
of the
Of
jpT (tco)
for various a is a curve, starting at the cutoff point specified
This is
of the complex plane.
by Eq, (10.7) and ending at the origin
shown in Fig. 10.3. Kochenburger's sufficient condition for stability
-Fr (io>)j
is then: The curve l/Fi(ia) must enclose the complete locus of
as shown in Fig. 10.3.
Figure 10.4 shows the case of complete instability.
The arrows on the
- JV(ta>)
curve show the direction of increasing ampliThe arrows on the l/Fi(iw) curve show
tude of the input to the relay.
the direction of increasing frequency, starting at the origin with

*
R.
J.
J. Kochenburger, Trans. AIEE, 69, 270-284 (1950).

R. Dutiih, L'Onde factrique, 30, 438-445 (1950).
co
0.
139
Besides these cases of complete stability and complete instability, there

are various cases of partial stability or instability, with the possibility
of persistent oscillations of a certain
For
instance, Fig. 10.5
frequency at constant amplitude.
shows a case having a convergent point.
For
Fr (iu) points are "outside" the l/Fi(ico) curve, and

small amplitudes, the
Then the amplitude of the oscillations will
is
unstable.
the
thus
system
FIG. 10.3
FIG. 10.5
FIG. 10.4
FIG. 10.6
As the amplitude increases, the point Fr (iu>} moves towards

the curve l/Fi(iw), and finally the point PI is reached. Then the system
will oscillate with a frequency and an amplitude corresponding to that
increase.
point.
The
oscillation is thus self-starting,
and the behavior
is
called
"soft" self-excitation. There is no tendency to move away from PI,

because any increase in amplitude will meet damping effects by entering
PI is thus a convergent point, and the

Figure 10.6 shows a different case where
Fr (iw) curve and l/Fi(io>) curve
the point Pa of the intersection of the
the stable region of the diagram,
system will always
oscillate.
140
is
The system always tends
a divergent point.
point.
It
is,
to
move away from
that
however, stable for small disturbances.
Figures 10.7 and 10.8 show yet more complicated cases, having both
The system of Fig. 10.7

convergent points PI and divergent points P 2
This
of
with
disturbances
oscillate
will
sufficiently large amplitude.
.
behavior
is
called
"hard"
self-excitation.
The system
of Fig. 10.8 will
always oscillate unless the amplitude of the disturbance is too large. For
very large disturbances, the system will diverge. All the cases depicted
in Figs. 10.5 to 10.8
show the
system on the amplitude
peculiar dependence of the behavior of the
of the disturbances,
persistent oscillation at fixed frequency
and the
and amplitude.
possibility of
These are
all
FIG. 10.8
FIG. 10.7
and are not present in the linear

Such behaviors are of course
in
the
studied
previous chapters.
systems
to be expected from our introductory discussions of nonlinear systems in
characteristics of nonlinear systems
Sec. 1.3.
Other Frequency-insensitive Nonlinear Devices. The Kochenburger method is a very effective solution for the problem of stability
It can be applied to systems with a conof a relay servomechanism.
10.3
siderable degree of complexity;
and
it
allows the direct inclusion of
experimentally measured frequency-response data. For most applications where the servo after the relay gives sufficient filtering action, the
The
neglection of higher harmonics in the relay output is fully justified.
prediction of theory
observation.
is
the Kochenburger
full agreement with experimental

the only design criterion is stability, then
solves the problem completely for relay
found to be in
Therefore,
if
method
servomechanisms.
In fact, the Kochenburger method is not only applicable to relay servomechanisms, but can be applied equally well to many other nonlinear
devices.
The
essential point of this
method
of analysis is the discovery
141
that the behavior of the relay is frequency invariant but amplitude

variant, while the behavior of a linear -system is frequency variant but
amplitude invariant. Now there are many nonlinear devices which have
same behavior
For instance, a gear train with backlash

can see this in the following way: Let 0i be the
angular position of the shaft of the driving motor, which is rigidly connected to the first gear of the train,
the
such a device.
is
and
let
62
as the relay.
We
be the angular position
of the shaft of the last gear of the
Then the
train.
and
10.9,
relation
between
61
can be represented as in Fig.

where 25 is the total backlash
62
If 0i, the input to the

gear train, is a sinusoidal oscillation,
8 2 the output, is a sort of clipped
sinusoidal wave with a lag in phase
of the train.
(Fig. 10.10).
It is easy to see that,
since the relation
not
between
0i
and
FIG. 10.9
dependent upon
time, the wave form of 02 will not be
is
explicitly
modified by a change in frequency of
#1.
Thus the response of the gear

If we denote the
train is amplitude variant but frequency invariant.
ratio of the amplitude of the fundamental of 02 to that of the input 0i

and the phase lag by the response Fg (iu>), then Fg (iu>) is a function of a,
but not of co. This response function can then be used to study servo-
mechanisms containing such gear trains with backlash in exactly the same
manner as the relay response function Fr (io>) in the preceding section.
FIG. 10,10
10.4 Optimum Performance of a Relay Servomechanism.

It is to
be borne in mind that, according to the Kochenburger diagram for stability, the curve l/Fi(zw) has to avoid the whole locus of the relay response
1 as in the case of a conventional
not just the single point
This is a more stringent condition on other parts of
the system, and is also the reason why the performance of a relay servo-
Fr(iw),
servomechanism.
142
mechanism tends to be inferior to the performance of a conventional servoHowever, this is not an intrinsic shortcoming of the relay
servomechanism. In fact, instead of merely asking for stability, we can
mechanism.
consider the relay as a switching device, capable of producing a positive

constant correction signal, a negative constant correction signal, or no
correction signal,
and then ask the question:
How
should
we
switch the
accordance with the output so as to obtain the optimum performance of the over-all system? For instance, the requirement on performance may be the quickest possible return to the normal state after a
relay in
disturbance.
This requirement not only guarantees the return to normal
state (stability) but also specifies the speediest return.

this
problem
of
optimum performance
is
The
solution to
to specify the proper switching
action of the relay as a function of the output, and this switching function
is the basis for the actual design of the servo system.
relay servo-
mechanism designed with
more general point
this
have a performance superior to that
of
of
view
will certainly
a conventional linear servo-
mechanism, because the nonlinear characteristics
of the relay are utilized
to the fullest extent.

10.5
Phase Plane.
If
is
the output and x the input, the differential
equation of a general second-order system, linear or nonlinear,
can be
written as
f(y,y,W)
(10.8)
where the dot denotes differentiation with respect to time.
We
can
rewrite Eq. (10.8) as the system
(10.9)
"
eft
we
consider y and y as the dependent variables, Eq. (10,9) is a system

two simultaneous first-order differential equations for the unknowns
= 0, and if / is not a function of t,
y and y. If the input is absent, x
If
of
the system
autonomous, as is frequently the case, then the first

equation
Eq. (10.9) can be solved for dy/dt and gives dy/dt as a function of y and y.
Then the system can be written as
that
is,
is
of
(10.10)
"''
Tt
This system of equations does not contain t explicitly.
equation of Eq. (10.10) by the second, we have
first
By
dividing the
|
This
is
now
143
k(V,fi
(10.11)
a first-order equation, with y as the independent variable

After this equation is solved, Eq.
and y as the dependent variable.
(10.10) can be used to calculate the relation
between
and
y.-
Physically, the procedure outlined in the previous paragraph
is
based
upon the concept of characterizing the state of the system by y and

instead of the more conventional method of specifying it by y and t.
y,
If
the variable describing the position of a point mass, then y is the

"velocity." y can thus be considered as representative of the momentum
is
of the point mass,
y and y then rep-
resent the position
and the momen-
tum, respectively, of the point mass.

Physicists call such a representation
of state the representation in phase
In the particular case dis-
space.
cussed, the phase space has only
two
dimensions; it is thus a phase plane.

The behavior of the second-order
'
then specified by a curve in

the phase plane. Every point on the
system
is
FIQ
curve represents the state of the system at a certain time L
It is
custom-
ary to indicate the sense of increasing time along the curve by an arrow,
If the order n of the system is higher than 2, the phase
as Fig. 10.11.
space will be of n dimensions, and the behavior of the system is represented
by a curve in this n-dimensional space.
The
practical advantage of phase-plane representation is that a very
number
of nonlinear systems of second order are autonomous

systems and can be expressed as Eq. (10.11), and this equation can be
In fact, the charsolved, at least graphically, by the isocline method.
large
acter of the system

curves, as specified
is
at once clear
by Eq,
when the
field of directions of
(10.11), is indicated in the
phase plane.
the
The
use of such geometrical properties of the phase plane is at the heart

of the theory of nonlinear oscillations and is called the topological meth-
ods of nonlinear mechanics.
To
translate our previous concepts to the language of the phase plane,

us consider the simple problem of a linear second-order system without a forcing function,
let
which
way
is
obtained from Eq. (3.39) by choosing the time unit in such a

make the natural frequency &> unity, f is of course the ratio
as to
144
of the
|f
<
system to critical damping.

Equation (10.12) can be rewritten as
damping
1.
of the
For
oscillations,
(10,13)
Therefore, corresponding to Eq. (10.11),
%L
dy
= _ ^y
+y=
we have
_2r - ^
(10.14)
then that the lines of constant slope dy/dy are radial lines from
the origin of the phase plane. Figures 10.12 to 10.16 are the five cases
It is clear
-KUO
FIG. 10.13
FIG. 10.12
FIG. 10.15
FIG. 10.14
with {
<
-1,
-1< f <
0,
0,
<
<
1,
and,
finally,
<
f.
10.13 and
Figures 10.12 and 10.16 are the nonoscillatory cases, Figs.
10.15 are the oscillatory cases, and Fig. 10.14 is the case of purely har-
monic
oscillation.
In the above
figures,
the origin of the phase plane corresponds to the
145
equilibrium state, where both dy/dt and dy/dt vanish. Mathematically,

the origin is the singular point of the system of equations of Eq. (10.13).
The character of the equilibrium state is, however, quite different for the
<
three different cases f
< f Figures 10.12 and 10.13

behavior of the system all diverge
Therefore the origin is an unstable equi-
0,
show that when f < 0, the

from the equilibrium state.
converge to the equilibrium

Then the origin is a stable equilib-
state.
rium
all
point.
in Figs. 10.12
which
Mathematically, the origin
and 10.16 is a point through
all lines of
behavior pass and
The origin in Figs.
called the node.
and 10.15 is the center of the

and is called the focus. In the
case of Fig. 10.14,
where f
0,
then called the
is
10. 13
spirals
special
the lines
The
of behavior circle the origin.

is
Figures 10.15 and 10.16

< f the lines of
librium point.
show that when

behavior
and
0,
lines of
origin
center.
FIG. 10.16
the basic equation of the secondorder system has a constant forcing term,
If
i.e.,
(10.15)
where
c is
a constant, then
2 ' d(y
c)
,
"*"
c)
dt
c)
Therefore the lines of behavior in the phase plane are entirely similar to
those indicated in Figs. 10.12 to 10.16, with only the modification of
= c, on the y axis.
translating the equilibrium point to y
10,6
In the subsequent discussion,
Linear Switching.
plify the
problem
we
shall sim-
by assuming only two

output and unit negative output. The
states
evidently not a restriction to our problem.
But
of switching the relay
for the relay: unit positive

fication of unit output
is
speci-
before attacking the optimum switching problem proper, let us consider

the simpler case of linear switching, i.e., the case where the forcing function c generated
ay
by.
by the
The purpose
relay
here
is
is
equal to unity but has the same sign as

some of the characteristics
to demonstrate
of the problem.
The performance of a linearly switched relay servomechanism is

2
1
The following
analyzed by Flugge-Lotz, and Fliigge-Lotz and Klotter.
l
l.
Fliigge-Lotz, Z. angew. Math. Mech., 26-27, 97-113 (1947).
1.
Fliigge-Lotz
and K.
Klotter, Z. angew. Math. Mech., 28, 317-337 (1948),
146
discussion
The
is
of the qualitative results of their investigation.
summary
differential equation of the

<
LK-
-]_
2f
system studied by the above authors
+ by)
sgn (ay
<
f or
<
is
(10.16)
curve in the phase plane with unit positive forcing function is called
arc.
a P arc. A curve with unit negative forcing function is called an
all
of
the
and
P
the
P
is
called
all
arcs
The system of
system
system,
arcs
is
called the N system.
with Eq. (10.15)
from our discussion in connection

that for the particular equation under study, the P
It is clear
a system of converging
its focus at the point
is
system
spirals with
= +i
and
0;
that
the
a system of converging
system
its focus at the point
with
spirals
is
1,
The
0.
state of the system
is,
desired end
of course, the
= y = 0.
According to the signs of a and
origin y
in the switching function of
Eq.
(10.16), four cases can be defined.

Let us call Case 1 the case with
Switching Line
ay+by-Q
>
line
FIG. 10.17
b,
ay
>
Then the switching
0.
by
is
a straight line
passing through the origin of the

phase plane and lies in the second and the fourth quadrants. To the right
of it, the sign of ay
by is positive, and the region is that of a P system.
To the left
by is negative, and the region is that of an

system.
line, the two systems join, and there the
corners of the curves of behavior occur (Fig. 10.17).
The condition for
the sign of ay
At the switching
of
it,
the existence of a periodic solution
is
that there should exist a
intersections with the switching line are equidistant
P arc whose
from the
origin; for
by symmetry, there exists an N arc on the other side of the switching

line joining the same two points, and these two arcs together form a closed
curve in the phase plane. Such periodic solutions are called limit cycles in
then,
We
the terminology of nonlinear mechanics.
shall see presently that
periodic solution can occur in our case.
Let
SP and Sx be
the points on the switching line where a P and an

R P and R N be the last intersections
curve, respectively, are tangent; let
curves
preceding Sp and SN with the switching line of the P and
through these points. SP and S# are symmetrical with respect to the
origin, as are
RP
and
outside the segment
Ry
Suppose that f
SPSN
as
shown
a,
and
in Fig. 10.18.
b are
such that
RN
is
solution starting
sufficiently near the
segment
SP S N
will
147
move away from
the line in one
direction or the other, according to the side of the line on which its
initial
point
lies,
and never return to the switching
line at
all.
It can
shown that every P arc which lies to the right of the switching
and has its end on this line also has the property that its terminal
also be
line
nearer the origin than its initial point; hence the condition for a
The lines of behavior of the
periodic solution can never be satisfied.
point
is
system then always spiral to one
of the foci,
and the end state
is
not the
origin of the
phase plane.
however, R N and R P are on the segment SpSx (Fig. 10,19), then
The reason is as follows. The P arc
there exists a periodic solution.
If,
Closed Cycle
FIG. 10.19
FIG. 10.18
RpSp starts at a point which is nearer to the origin than its end point,
according to our assumption. However, at large distances from the
origin, where the effects of shifting the foci of the converging spirals
1 for the P system and the
from the origin to +1 and
system, respec-
tively, are negligible, a
arc
must
on the switching line at a point

end point on the switching line.
start
away from the origin than its

These arcs then have the reverse property
farther
of R PSp.
Therefore, by
some intermediate P arc of this type must begin and end
This is the condition for the
at the same distance from the origin.
existence of a periodic solution, which is shown in Fig. 10.19 as a closed
continuity,
Extended analysis bears this out and shows that the periodic
is unique and, more important, orbitally stable: All solutions
beginning outside the periodic solution spiral onto it, and those solutions
cycle.
solution
which begin inside the periodic solution but outside the area enclosed by
RpSP R NSN, the shaded area in Fig. 10.19, also spiral onto it. Lines of
148
behavior beginning within the shaded area will spiral to one of the foci.
Here again we have no possibility of reaching the origin.
Case 2 is the case where a > 0, but & < 0. Now the switching line
The P systhird quadrants.

passes through the first and
No
case.
the
in
as
periodic
tem and the
previous
system are the same
Let R P RN, S P and Sy be defined as before;
solution occurs in this case.
ay
by
R P SP
then the points

order, as
shown
0,
S N and R N
lie
But on the
in Fig. 10.20.
on the switching
interval
SP S N
line in this
new phenome-
Consider any solution which reaches this interval, say at

the point E. What does the solution do at this point? It should proon this line. However, the
ceed along an
arc, because switching occurs
half
the
same
into
E
back
arc from
plane from which the solution
non
occurs.
goes
ay+by^Q
FIG. 10.20
entered E, and on this side a solution can contain only P arcs; on the
other hand, the solution certainly cannot follow the P arc through E
beyond
this point.
beyond E,
RpRx
it
spirals in
- 1.
"ending" the
In
reality,
The paradox
has a time
say then that the solution
Any
RPSPj
If it
it will
reaches
is
not defined
solution starting outside the region
toward the origin until
reaches the interval
will spiral to
of
We may
"ends" at E.
it
it
reaches the interval RNRp.
If
+1. If it reaches RySy it

we have the curious phenomenon
spiral to
S P S N}
solution.
the behavior of the system cannot "end" but must go on.

is resolved by the observation that switching action always
lag,
and when a solution meets the switching
line, it
actually
proceeds for some distance beyond it before it has the change of sign
In Case 1 such a time lag, provided that it is
of the forcing function.
not too large, does not affect the essential behavior of the system. But
"end" of the solu-
in the present case, the time lag avoids the necessary

tion.
lag, it
Consider a solution entering an end point. Because of the time

no longer ends there, but proceeds for a certain distance beyond;
149
then switching occurs, and the solution in the phase plane makes a
"
corner/' where the solution is still defined. From this corner, it crosses
the switching line in the reverse direction, moves for a short distance
beyond, has another corner, and so on, as shown in Fig. 10.21. From
that figure, it is also seen that such zigzag action of the system results
in its creeping out of the region SpSx, and eventually the solution will
FIG. 10.21
one of the
spiral into
solution, but
the solution
Case 3
is
Time
foci.
lag will thus avoid the ending of the
the behavior of the system

not able to reach the origin.
still
is
the case where a
line lies again in the first
<
0,
>
0.
is
unsatisfactory because
In this case the switching

of the phase plane.
and the third quadrants
P system now lies to the left of the switching

system now lies to the right of the switchIn this case, a stable periodic
ing line.
But the
solution,
10.22),
exists,
and
dominates
it
and the
ay-\~by=Q
or a stable limit cycle (Fig.
always
line,
P-arc
the whole situation, for all other solutions

Here again the origin of
spiral into it.
the phase plane cannot be reached.

Case 4 is the case where both a and b
The switching line is the

are negative.
same as in Case 1, but the arcs are the
It may be shown that
same as in Case 3
no periodic solution can exist in this
FIG. 10.22
case.
Without time lag in the switching
action, the segment S P S N (Fig. 10.23) consists of end points, and by tracing
.
the solutions which end on
it
backwards one can see that these cover the

"end" on the segment S P Sx of
entire plane; thus in this case all solutions
the switching line.

But here again, as in Case
ence.
The time
lag
makes no
2,
the presence of time lag makes a differ-
difference of importance until the solution
150
in question reaches
for a small distance
&>&;
then, instead of ending, the solution proceeds
beyond the switching
line,
has a corner, crosses the
has another corner, and so on. It may be seen

from Fig. 10.23 that this motion forces the system to the origin. The
system will finally oscillate at high frequency and small amplitude
switching line again,
around the origin

This
is
what
is
the smaller the time lag, the higher the frequency.

"
"
chattering of the servo system.
called
It is thus seen that out of the four cases discussed, only Case 4 gives
a system seeking the desired equilibrium state.
But even then, the
FIG. 10.23
system
will chatter
near the equilibrium state.
Our
discussion of th<
analysis of Fltigge-Lotz and Klotter then demonstrates the shortcomings

of linear switching.
Optimum switching for best performance of the
servomechanism
sections,
10.7
we
is
definitely not
show
shall
linear switching.
In the following
this.
Optimum Switching
system with forcing function
Function.
of unit
An autonomous
second-ordei
magnitude can be written as
dy
Tt
(10.17)
g(y,y)
where
<p(y,y)
values
+1
or
is
a discontinuous function with only the two possible

Then the optimum switching problem can be defined
1.
as follows: find the function <p(y,y)
such that, beginning at any point p

phase plane, the solution will pass through the origin 0, and the
length of time necessary to move from p to
along the solution from y
is minimal with
respect to <p no other <p could make this time shorter.
of the
Then
151
the optimum switching function. This particular switching

was
studied
by Bushaw, and the special case of linear g(y,y),
problem
that is, g(y,y) = 2f# + y, was completely solved by him for all real
<p(y,y) is
values of f
are,
The mathematical arguments which
however, complicated and are
difficult to
led to
Bushaw's
results
extend to other cases.
We
shall thus limit ourselves to indicating his solution.
A general result, good for any

A
continuous g(y,y),
is
Bushaw's canonical
a line of behavior in the phase plane.

Since
path.
take only the value +1 or
1, a path consists of P arcs and
junction of these arcs is called a
path
is
<p(y,y)
can
N arcs. A
PN
corner
with increasing time the
if
switching
is
from
-f 1 to
-1
for the
forcing function. Similarly, a junction is called an

corner if the
NP
1 to +1.
A
path is called canonical if it contains
no NP corners above the y axis and
is
switching
from
no PN corners below the y axis.
The
importance of the canonical path

that a minimal path, i.e., a path
minimum
is
of
FIG. 10.24
must be canonical.
That is, given any path A from a point p which is not canonical, one can
find a canonical path from p which is shorter than A in terms of time.
This can be shown quite easily. Given, say a path with the NP corner
p above the y axis, as shown in Fig. 10.24, one denotes by p' either the
time,
path preceding p or the last intersection preceding p

the path with the y axis, whichever is nearer p, and one denotes by p"
the corresponding point following p. We then draw the P curve forward
last corner of the

of
from
p'
and the
N curve backward from p".
tions of Eq. (10.17),
Now*, from the basic equa-
we have
(10.18)
dy
Therefore, at any point in the phase plane, the slope of the P curve is
always greater, algebraically, than the slope of the AT" curve. Thus the
If
configuration of the paths must be like that indicated in Fig. 10.24.
corner
f
t(p
is
ff
we now modify
NP
fl
the
the given path by replacing p'pp" by p p "p
removed, and the path is made canonical. If we denote by
pp") the time interval in going from
>
p',
through p to p"] and by
t(p'p'"p"), the time interval along the canonical path, then

1
D. W. Bushaw, Experimental Towing Tank, Stevens Institute
Report 469, Hoboken, N. J. 1953.
of
Technology
152
is greater than the

y, the value of y on the canonical path
and
therefore
on
the
value of y
t(p'p'"p"} < t(p'pp"}.
original path,
canonical path.
non
than
the
is
"shorter"
the
Thus
canonical path
But
at
any
switching
theory of the optimum

Then the system of Eq.
of the application of the
As a simple example
function, let
us take g(y,y)
fy.
becomes
(10.17)
*y
v.
(10.19)
J
N
The
of f ;
system of arcs depend, of course, on the value

system and the
but since Eq. (10.19) does not explicitly contain y, these systems of
=0
0<C
FIG. 10.25
arcs consist of parallel curves shifted along the y axis.
typical
arc
and an Ar arc with one branch passing through the origin, for each of the
three possible values of f, are shown in Fig. 10.25.
The case of f <
is different from the other cases, in that in order to reach the
origin the
initial
value of y must be within the range
case then, the problem of

initial
We
lies
within this specified range.

shall denote by T that part of the
1/f to +l/f.
For
optimum switching has meaning only
if
this
the
is
below the y
axis,
and by T~
curve through the origin which

about the origin. r~ is
its reflection
therefore that part of the

curve through the origin which lies above
the y axis.
T and T~ together gives a curve C. Bushaw showed that
C is the optimum switching line in the sense that above C the
optimum
switching function <p(y y) should be
ing function <p(y,y) should be +1.
}
1,
153
and below C the optimum switchThis
is
indicated in Fig. 10.26.
From any point p above (7,

Physically, the switching is done as follows.
should
function
be
arc
and
the
the forcing
system follows the
1,
There the forcing function changes to +1, and

the system follows C to the origin. If the initial point p is below C, the
forcing function should be +1, and the system follows the P arc to the
to the switching line C.
0<C
c-o
FIG. 10.26
(a)
(c\
switching line C.
There the forcing function changes to
and the
system follows C to the origin.

That Bushaw's solution for the optimum switching line is correct
can be seen as follows. First of all, we note that in order to reach the
because C is the only
origin, we must use C for the last part of the path,
curve through the origin. Now suppose that the initial point is above C
and that the optimum path, according to Bushaw, is that indicated in
arc from p to C and then the path along C
Fig. 10.27 (a) by the single
to the origin.
Then
if
the switching
is
done too
early,
we have an
NP
154
corner before reaching C.
and make a
PN
In order to reach
C we have
to switch again
f
corner.
done at p when y
is still
switching
this
violates
the
the
then
below
we
have
corner
a
y axis;
negative,
rule for a canonical path.
The time along the modified path is definitely
If the PN corner is made at p" when y
longer than the optimum path.
If this
is
PN
is positive,
the time required will be even larger, because the path then
has a closed loop. Thus switching too early is disadvantageous. Figure

Since the paths p'Q and
10.27(6) shows the case of switching too late.
ff
time
the
same
are
and
thus
interval, it is easily
p"p'
equivalent
require
seen from the figure that switching too late is also detrimental.
Figure
10.27(c) shows yet another variation, where the first part of the path is
a P arc instead of an
arc.
But it is apparent from the figure that this
variation
is
mum path.
cate the
also
worse than the opti-
These considerations indi-
correctness of choosing the
canonical path as the
optimum switch-
line.
ing
10.8
Optimum Switching Line for

Linear Second-order Systems.
Bushaw has determined the optimum
switching line for the linear secondorder systems with g(y,y) = 2f#
y,
for all real values of f
We
shall only
state his result here without proof; but

FIG. 10.28
in
view
of
our discussion of the simple
case in the preceding section, the general character of the result can be
The P system and the
system for this g(y,y) are
easily understood.
simply the family of curves in Figs. 10.12 to 10.16 with the origin shifted
to (+1,0)
and (1,0),
respectively.
example is the case of f > 1. Then

the switching line C consists of the P arc from infinity to the origin of the
arc from infinity to the origin.
phase plane, and the
Again, above C,
1 and below C, <p is
the switching function ^ takes the value
equal to
The
case nearest to our simple
+1.
The optimum path from an
initial
point above
is
thus as indi-
cated in Fig. 10.28.
When f < 1, then, as in the simple example, only from points within
a limited region of the phase plane can the system reach the origin;
because without the forcing function the system is unstable. Bushaw
specifies the boundaries of this region as the P arc /row the point (+1,0)
the point (1,0) and the

arc from the point (1,0) to the point
(+1,0), as shown in Fig. 10.29. The switching problem has meaning
to
only for
initial
consists of the
points within this region.

arc to the origin and the
The optimum switching line C

Above C,
N arc to the origin.
The
1, and below (7, <p is +1.
C
above
is
shown
in
point p
Fig. 10.29.
system and the N system are circles with the
the switching function
<p
optimum path from an
initial
When
the
0,
centers (+1,0) and
is
equal to
line
center on (+1,0), until the path intersects
1,0), respectively.
The optimum switching
a series of semicircular
is
(Fig. 10.30)
155
arcs of radius unity, starting at the

origin,
and stretching along the y
both
in
directions.
The
axis
are
arcs
below the y axis for positive y and

are above the y axis for negative y.
Above the switching line (7, the switching function

<p
it,
is
<p
is
equal to
below
Thus, starting from a
+1.
point p as shown in Fig. 10.30, the path

follows an
arc, or a circular arc with
its
center on
intersects
1,0).
arc, or a circular arc
again at
When the path
at a, the path
with
its
At 6, the path changes again into an N arc,
6.
section with the switching line C, etc.
and, from
FIG. 10.29
becomes a P
there, the path follows
The
until the next inter-
last intersection
C into the origin.
more complex switching performance than that
with
C is d,
This is a considerably
for f
>
1.
C-0
FIG. 10.30
The next more
<
1.
For
the case of converging spirals, or

<f
showed that the optimum switching line
difficult case is
this case,
Bushaw
should be constructed as follows:
We
first
draw a
spiral
backward
in
The first arc of C is the first arc of P, from

time, starting at the origin.
the origin to the first intersection with the y axis. Then, we draw the
156
above the y axis about their right point of intersection

with the y axis. Then we assemble all these arcs below the y axis into a
continuous curve by moving the consecutive arcs parallel to the y axis
until they join end to end, starting from the origin and extending to
reflection of all arcs
This
the right.
is
the positive half of the
optimum switching line. The

by reflection about
negative half of the switching line is then obtained
Again, above the switching line
the origin.
function
<p
is
1;
below
is
<p
it,
the same as for the case f
is
(Fig. 10.31), the switching
+1. The situation is then very much

shown in Fig. 10.30; the only difference
the replacement of circular arcs by spiral ares.
0X1
FIG. 10.31
The last case is the case of diverging spirals, or

1 < f < 0.
The
method of construction of the optimum switching line is exactly the
same as in the preceding case, except that now the consecutive spiral arcs
diminish in size instead of increasing.
The length of the switching line
is thus finite, spanning the y axis from (
a,0) to (+a,0), as shown in
This is as it should be; because here we have negative
Fig. 10.32.
damping, and, as in the case
of Fig. 10.29, the
path can reach the origin
the initial point is within a certain region near the origin. In

fact, the boundary of this region consists of the P arc from the point
arc from the point (-a,0) to (a,0).
The
(a,0) to (-a,0) and the
1 for points above C in
optimum switching function <p takes the value
only
if
the phase plane, and the value
The boundary curves
+1
for points below the switching line.
of the regions of possible
optimum switching
as indicated in Figs. 10.26 and 10.29 are evidently the limit cycles with
switching points at y
of the relay
0.
They each then
servomechanism considered.
It
represent periodic solutions
is, however, equally apparent that these periodic solutions are unstable: the slightest disturbance
will cause the lines of behavior of the system to spiral to the
origin or to
diverge to infinity.
157
Therefore such periodic solutions cannot occur in
reality.
One property
out: for
of our solution for the
all cases,
the
optimum switching problem stands
optimum switching function
in the first quadrant of the phase plane
<p
takes the value
and takes the value
+1
in the
FIG. 10.32
third quadrant.
By
writing Eq. (10.19) in the following form,
.dy
dt
we can understand
this feature of the solution quite easily.
of the design is to return to the state y
0,
or the
axis, in
The purpose
the shortest
When
y and dy/dt are both positive, this purpose can be

accomplished by making d^y/dt* or the curvature of the y(t) curve, as
1.
When y and dy/dt
negatively large as possible, that is, <p should be
possible time.
are both negative, d 2 y/dt* should be as positively large as possible, that

This intuitive reasoning agrees with our result for
is, <p should be +1.
the optimum switching function. When y and y have different signs,

the optimum value of <p cannot be so simply determined, because then
t axis depends upon the complicated interaction
Bushaw's contribution is then to specify the optimum
the rate of return to the
between y and
y.
158
<p
just in these regions, the second
But it
plane.
line C must lie
For systems
is
clear
from
and the fourth quadrants
this discussion that the
in the second
of the
and the fourth quadrants.

and for systems of more than one degree
of higher order
of freedom, the phase-plane representation of the state of the
no longer
phase
optimum switching
possible.
We
have to use a phase space
of
many
system
is
dimensions.
analogy with the problem already discussed, we may expect the

solution of optimum switching of such complicated systems to be the
determination of optimum switching surfaces in the phase space. HOW -
By
ever, such
has been
10.9
problems have not yet been solved; only an
made by Kang and
Fett.
initial
attempt
Multiple-mode Operation.
What happens when
the system
guided by our switching action to the origin of the phase plane?
is
Clearly,
Ôutput
FIG. 10.33
the forcing function
if
the origin
again.
is
But
is continued at the value obtaining just before

reached, the system will move away from the desired state
as soon as it moves away from the origin, the switching action
designed into the system will function and drive the system back towards
the origin.
The net result will be a rapid approach to the origin from
any disturbed position, and then a high-frequency oscillation or chattering around the origin.
If small deviations from the rest state or the
origin are not objectionable, the chattering near the origin can be eliminated by cutting off the
forcing function whenever the system is near the rest state, i.e., whenever
y and y are small enough to be negligible. The system then has two
modes of operation: for large deviations, the switching action together
with the output of the relay functions; for small
deviations, such input
to the system is cut off.
We shall see the necessity of such multiple-
mode operation
of
servomechanisms again in the later chapters.
The optimum switching line, being nonlinear, cannot be implemented

by a simple linear circuit. In fact, the measured value of the output y(t)
1
C. L.
Kang and
G. H. Fett, J.
Appl
Phys., 24, 38-41 (1953).
has to be
is
"
77
digested
by a nonlinear
159
device, or a computer.
The com-
designed to generate the switching signal for the relay according
puter
to the optimum switching line.
Furthermore, there should be another
to
disconnect
the
relay output from the system when the
cutoff computer
and
are
small
y
enough. Therefore the block diagram of such
output y
a relay servomechanism
is
that of Fig. 10.33.
The
incorporation of
computers into control systems will be generally necessary for the more
complicated systems discussed in the following chapters. However,
conceptually, this really involves nothing new the compensating circuit
used in a conventional servomechanism for modifying the system transfer
But in these simpler systems, the computing

is also a computer.
function can be performed by a linear circuit, such as an RC circuit. We
shall give a more extensive discussion on computers in Chap. 13.
function
CHAPTER
11
NONLINEAR SYSTEMS
A
nonlinear system is a system for which the output is not linearly

proportional to the input. The relay servomechanism is a simple examIn Chap. 6, we have given a general
ple of such nonlinear systems.
method
of linearizing
system can be
made
any nonlinear servomechanism,

to
behave
like a linear
i.e.,
any nonlinear
system by modifying the

In the preceding
system into an oscillating- control servomechanism.
we have
presented a method for analyzing a servomechanism

nonlinear
device whose behavior is insensitive to the frequency
a
including
These methods for designing nonlinear servomechanisms
of the input.
chapter,
encompass a wide variety of nonlinear problems in engineering practice

and are quite sufficient for dealing with the usual systems synthesis
problems.
On the other hand, as indicated in the later sections of the last chapter,
the problem of optimum utilization of the nonlinear characteristics of a
system to improve the performance of the system is generally a much

more difficult problem than the problem of design for stability. In fact,
only a modest beginning has been
made
in this direction.
It
is
thus not
possible at the present time to give a satisfactory treatment of the sub-
servomechanisms. Furthermore, the problem

should perhaps be formulated more directly and in a different way.
Instead of rinding the performance of an assumed system, we should
specify the performance of the system, and then determine the required
ject of general nonlinear
This approach
nonlinearity.
of this
chapter
is
will
quite limited:
be discussed in Chap.
we
shall indicate only a
14.
The scope
few
possibilities
of purposefully utilizing the characteristics of a nonlinear system.
11.1
Nonlinear Feedback Relay Servomechanism. If we confine ourwhere the deviations from the equilibrium state are small,
selves to cases
the results of
of a relay
Bushaw
stated in Sec. 10.8 for the
servomechanism can be greatly
optimum switching
simplified.
It is seen
discussions in that section that near the origin the switching line
approximated by
y\y\
= -2y
line
from the
is
(n.i)
This indicates that an improvement in the performance of the system

over that of a system with linear switching (Sec. 10.6) can be achieved
160
NONLINEAR SYSTEMS
by using nonlinear switching defined by Eq.
(11.1).
161
If
is
the input and
y the output, Eq. (11.1) should be modified into

a*y\y\
(re
y)
(11.2)
or
sgn
where a
is
(:r
y) -\/
y\
ay
(11.3)
a constant.
The block diagram

condition of Eq.
servomechanism
of a relay
(11.2)
is
shown
in Fig.
11.1.
utilizing the switching
This method of
y\y\
feedback was proposed by Uttley and Hammond 1 for improving the

performance of the simple relay servomechanism. The computer here
FIG. 11.1
SERME
FIG. 11.2
only required to generate the signal a?y\y\ and thus can be relatively
The equivalent switching condition of Eq. (11.3) can also be
simple.
feedback. Since
y) \/\x
implemented by the y and sgn (x
y\
is
(a;
error
be called sign
y) is the error, the
system
(SERME)
The block diagram is shown in Fig.
may
system.
again, the computer is relatively simple.
2
J. C. West.
error root-modulus
11.2.
Here
This system was proposed by
These nonlinear feedback relay servomechanisms, although comparaIn fact, our argument in
tively simple, cannot be analyzed rationally.
their favor is only
a plausible one, not a complete one.
The
final design
1
A. M. Uttley, P. H. Hammond, "Automatic and Manual Control," p. 285, edited
by A. Tustin, Butterworth & Co. (Publishers) Ltd., London, 1952.
2
J. C. West, "Automatic and Manual Control/' p. 300, edited by A. Tustin,
Butterworth
&
Co. (Publishers) Ltd., London, 1952.
162
of
any particular system using these principles has to be determined by
actual experimentation.
11.2 Systems with Small Nonlinearity.
freedom
Eq.
is
(2.3),
If a system of n degrees of
not linear, then instead of a linear differential equation, such as
we have
"- 1
'
dt- 1
dt"
'
'
ua
'
dt
|---.^)-*w
(1L4 >
the output, and

on the left of
part
/
in
All the
as
differential
is
thus
the
linear
(2.3).
Eq.
operator,
Eq. (11.4)
nonlinearity of the system is represented by the last term on the left
Small nonlinearity means that /x is small in comparison
of the equation.
where the
is
and
a's
ju
are constants, #()
a nonlinear function of
with the
its
solution in a
of
the input,
The
power
series of
/*
try a formal expansion of the
may
substituting Eq. (11.5) into Eq. (11.4)
and by equating equal powers
we have
/*,
and other equations for higher-order terms.
The
"
zeroth-order
5J
approxthus the linear equation, as in Eq. (2.3). But more important

the fact that the first-order correction term due to nonlinearity is
imation
is
*/() is
first
a's.
For small nonlinearity then, we
By
is
variables.
is
determined by Eq.
(11.7),
an equation
as the zeroth-order approximation.

approximation shows that the system is
istics
of exactly the
same character-
In other words, if the linear

damped and has other desired
properties of a servo system, then the first-order correction y^(t) will

also
have such properties.
Moreover, because
of
the occurrence of the
small parameter p before y^(t) in the expansion of Eq. (11.5), the corrections for nonlinear effects are small.
Consequently, small nonlinearity
in a system of "satisfactory" performance will not alter essentially the
system behavior from its linear approximation. Therefore, as far as an
engineering approximation is concerned, we can treat such systems as
linear systems.
This
is
the reason that the linear theory of servomech-
NONLINEAR SYSTEMS
163
anisms has had such good success in spite of the ever present small noneven a "linear" system.
linearity in
On
the other hand,
small, then
even
if
we know
if
the input x(f)
approximate system
the
damping
that there
is
of
of the linear
approximation
(V
Eq. (11.6) can be very
much
than unity.
or the non-
magnitude as some
of the linear
uit
linear effects, can be of the
cedure of the last
very
larger
yW
terms, even with small
is
the possibility of resonance. That is,

w
of order unity, the output
y (t) of the linear
is
same order
of
V.1
In other words, our formal expansion proparagraph is not justified, and we must expect strong
p.
A Am P Iltude
1 Amplitude
(a)
(b)
Soft Self -excitation
Hard Self -excitation
FIG. 11.3
from even small nonlinearity if the system has only very weak
damping. In the following sections, we shall give a brief description
effects
of the kaleidoscopic
behavior of a nonlinear system. Detailed treatment

such phenomena in nonlinear mechanics can be found in the excellent
books by Minorsky 1 and Stoker. 2
of
11.3
Jump Phenomenon.
As already stated
oscillation in a system,
if it is
the oscillation
"hard"
in Sec. 10.2, self-excited
automatically built up from very small

deviations from the equilibrium state, is called "soft" self-excited oscillation; if it has to be started by large deviations from the equilibrium state,
is
called
self-excited oscillation.
In some
cases,
the coefficients of the differential equation of the system may

depend
upon a parameter X of the system. If, for some particular value of X, the
critical
the character of the equilibrium state changes from that of a
stable state to an unstable state, then the limit cycle or steady oscillation
will appear.
According to whether the system is one of soft or hard selfexcitation, the
1
phenomenon occurs
as in either Fig. 11. 3a or
6.
N. Minorsky, "Introduction to Nonlinear Mechanics," Edwards Bros.,
In the
Inc.,
Ann
Arbor, Mich., 1947.

2
J. J.
1950.
Stoker, "Nonlinear Vibrations," Interscience Publishers, Inc.,
New
York,
'ENGINEERING CYBERNETICS
164
the parameter X increases, nothing happens until X reaches

Xo, at which point the equilibrium state changes from stability
first case, if
the value
to instability with a simultaneous appearance of a stable limit cycle
amplitude begins to increase with
made
is
whose
to decrease, the phe-
retraces its path exactly, and the limit cycle disappears when
For a system with hard self-excitation, the picture is differ-
nomenon
X
If
X.
oscillation appears suddenly at X = X with a

with
increasing X, the amplitude increases. If X
amplitude, and,
or
limit cycle does not disappear when X = Xo
the
oscillation
decreases,
The
ent (Fig. 11.36).

finite
but disappears farther along the curve where the amplitude again jumps
Thus the jump phenomenon is associated
finite value to zero.
from a
with the hysteresis
system behavior.
Frequency Demultiplication. If a nonlinear system
11.4
of the
is
acted on
by a periodic input containing two frequencies <oi and co 2 the output of

the system occurs not only with these frequencies and their harmonics
,
of the so-called combination tones
but also with an additional spectrum
mî
x
nco 2 ,
m and n
where
#o(cos cut
cos
co2<0
is
current y
Thus, for example, if a voltage

impressed on a nonlinear conductor whose
2
a 3 3 then the output y(f) will
&2
aix
are integers.
is given by y
contain the following frequencies: MI,
wi
~~
&2>
2i +
o> 2
2coi
o?i
o>2,
2co 2 ,
co 2 ,
and
2coi,
coi
2co 2 ,
2co 2
3wi, 3o> 2
.
The
wi
o> 2 ,
first six fre-
quencies are regular harmonics, but the last six are the combination
tones resulting from the nonlinearity of the conductor. Some of these
are higher, and others are lower than the original frequencies a?i and co 2
These lower frequencies are called subharmonics, and the process by
.
which they are obtained is called frequency demultiplication.

It is seen that if wi and co 2 are fairly close together,
i
much
smaller than either of the original frequencies.
If,
o? 2
can be
in addition,
the system can be stabilized at such a low frequency, subharrnonics of

the order of
of the input frequency, and even lower, can be obtained.
TW
such systems are connected in series, with the subharmonic

output of one serving as input to the other, then still lower frequencies
can be reached.
If several
11.5
Entrapment
If a nonlinear system has a selfthen when the system is subjected

to an input of a slightly different frequency co 2 we expect the simultaneous occurrence of both coi and o> 2 and, through nonlinear interaction, the
of Frequency.
excited oscillation of frequency
&>i,
beat frequency
in Fig. 11.4.
co 2
coi.
In reality the phenomenon develops as shown

o> 2
coi disappears suddenly as soon as
The frequency
w 2 reaches a certain zone of synchronization AB. In this interval there

one frequency w 2 and everything happens as if the original
exists only
frequency coi were entrained by the variable frequency co 2

This phenomenon of frequency entrainment was first explained by van
.
NONLINEAR SYSTEMS
der Pol and was extended
by
others.
165
Let the system be one of second
order and determined by the differential equation
where
a, 7,
and
are positive constants.
If
B =
the system has
0,
negative damping for small amplitudes of oscillation, but positive damping for large amplitudes of oscillation. Thus there is an amplitude at
which the system can maintain a

steady oscillation. Furthermore, if
both a and 7 are small, then steady
self -ex cited oscillation
must occur
a frequency close to wi.

Pol showed that when w 2
Van
is
at
der
close to
the solution of Eq. (11.8) can be

written as
coi,
y(f)
bi(f) sin u%t
62(2)
cos <*$
FIG. 11.4
(11.9)
and that
bi(t)
and
bz(t)
are slowly varying functions of
t,
such that, for
example,
db]
Hi
and retaining only the terms
write the equations for bi and 62 as
By
substituting Eq. (11.9) into Eq. (11.8)
up
to
first order,
we can
/i(6i,6 2 ;w 2 )
(11.10)
These are autonomous equations

be solved by the method of
of the first order.
isoclines in the
Therefore they can
same manner as a second-
order system in the phase plane. Analysis shows that for a certain range
of co 2 near
the system of Eq. (11.10) has a stable node point in the
i,
&i&2 plane.
Thus, no matter what the initial values of 61 and 62 are,
the system tends to a fixed set of values of 61 and & 2 corresponding to
this point.
Thus no oscillations of the frequency o?2
coi appear at all.
For
co 2
outside this range, there
which accounts
is
a stable limit cycle in the &i& 2 plane
for the observed effect
shown
in Fig 11.4.
Asynchronous Excitation and Quenching. In some nonlinear

systems it is possible either to start or to stop an oscillation of frequency
co by means of another oscillation of an
entirely different frequency &>i.
11.6
166
In the
first case,
the second,
in
the
phenomenon
is
called asynchronous excitation,
An
of
understanding
and
these
asynchronous quenching.
existence
recalling the fact that the mere
phenomena can be obtained by
a limit cycle in the phase space of a system does not necessarily mean
the occurrence of steady oscillation. For steady oscillations to take
must be stable in the sense that the system has a
the limit
of
cycle
place,
if the system is disturbed

tendency to return to the limit cycle, even
limit
the
cycle in the phase space.
and displaced to a point away from
unstable limit cycle cannot be realized in a physical system. În

view of this, one can readily see that the appearance of a new oscillation
An
create or destroy
may, under certain circumstances, either
i
mg+a
cos
tte concjitions O f stability of another.
In the
chronous
between
11.7
first
asynchronous excitation
have the phenomenon

second that of asynchronous quenching.
of
case
we
and in the
The term asyn-
relation
merely emphasizes the fact that the
to and wi is purely arbitrary.

Parametric Excitation and Damping.
It
has been
a parameter of an oscillating
known
a frequency w, the system
with
system is varied periodically
the
frequency /2. Lord Rayleigh
begins to oscillate at
effect
demonstrated this
by attaching one end of a stretched
for a long time that
if
wire to a prong of a tuning fork. If the fork oscillates with frequency

A similar
of the wire appear with frequency to/2.
co, lateral oscillations
a
mass
a
weightless
supported by
problem is that of a simple pendulum,
sinusoidal force applied at the end of the bar
bar, under the influence of a
the small angular displacement of the pendulum from

the vertical, and, without loss of generality, if the frequency of the sinusis then
oidal force is unity, the differential equation for
(Fig. 11.5).
If 6 is
ml
-jTg
(m<Q
o,
cos
f)Q
the mass, g the gravitational attraction, I the length of the

This equation
force.
pendulum, and a the amplitude of the periodic
can be written as
where
is
thus equal to g/l and $ is equal to a/ml The case of an inverted

also be reprependulum, with the mass above the point of support, can
is
sented by Eq. (11.11) by taking g as negative. Therefore, for a normal

pendulum a is positive, and for an inverted pendulum a is negative.
is actually a linear equation but with periodic variation
Equation (11.11)
end force applied to the pendulum. Thus the system can be
considered as one with periodic variation of its parameter.
in the
NONLINEAR SYSTEMS
Equation (11.11)
167
the well-known Mathieu differential equation.

determined by the constants a and 0. In
is
The
stability of the solution is
fact,
plane can be divided into a region of stability and a region

shown in Fig. 11.6 (where the stable region is shaded).
thus seen that for positive a, the case of the normal pendulum, the
the
a/3
of instability as
It is
system
is
stable
of course,
well
when
the periodic end force
known.
it
is
absent, or
0.
This
is,
interesting to note that with

unstable.
Then the pendulum
However,
appropriate /3, the system can be made
will swing with increasing amplitude until,
is
finally,
nonlinear effects limit
FIG, 11.6
This is the phenomenon of parathe amplitude at a fixed large value.

For negative values of a, the case of inverted pendumetric excitation.
lum, the system without a periodic end force is naturally unstable. But
narrow range of 0, the system can be stabilized. This phe-
for a certain
nomenon
is called parametric damping.

Parametric excitation or parametric damping can occur in any system
with periodic variation in the parameter of the system. This phenome-
non or any
of the previously
shown nonlinear phenomena can be
utilized
In fact, many
components of a servo con-
to achieve the desired performance of a control system.

of
them
trol
are already incorporated into the many

system. However, these nonlinear components are
and they are "designed

retical
analysis.
The
mere
"
gadgets,"
rather by experience and testing than by theoapplication of the characteristically nonlinear
77
to the over-all design of a control system is as yet an unexOur discussion in the preceding sections serves only as an
plored field.
indication of its rich possibilities.
phenomena
CHAPTER
12
LINEAR SYSTEM WITH VARIABLE COEFFICIENTS

The only system with time-varying
in the previous chapters is the
coefficients considered in detail
pendulum with a
periodic force at the
supporting end, discussed in connection with the phenomena of parametric excitation and damping. All other systems considered do not
have
coefficients of their differential equations that are explicitly func-
We have, however, shown in Chap. 1 that linear systems

with time-varying coefficients can have behavior entirely different from
In this chapter, we shall again
that of systems with constant coefficients.
tions of time.
take up this question and discuss in some detail such a typical but simple
system: the short-range artillery rocket. We shall demonstrate that the
question of stability of such a system with variable coefficients cannot be
solved in the same manner as for the linear system with constant coeffi-
Not only
cients.
is
the method of Laplace transform and transfer funcwe are forced to change our entire
tion useless for the purpose, but
approach to the problem.

We shall study the motion of a fin-stabilized
artillery rocket
during
We shall be particularly
the period of action of the rocket thrust.
concerned about the angular deviations of the rocket axis from the launching angle caused by disturbances during the launching and the subsequent damping action of the fins. The general problem of the dynamics
was studied in great detail by various authors in

World War II. The American work is summarized by Rosser, Newton, and Gross. 1 The work done in England
2
Carriere's paper 3 represents a French investiis reported by Rankin.
Our discussion here will be greatly simplified
gation of the same subject.
of artillery rockets
different countries during
and has the purpose of bringing out only the salient points of interest
to our study of linear systems with variable coefficients.
12.1
Artillery
Rocket during Burning. For a fin-stabilized artillery

motion in the vertical plane and the horizontal
rocket, the interaction of
plane
1
J.
is negligible,
i.e.,
there are negligible aerodynamic forces in the
B. Rosser, R. R. Newton, and G. L. Gross, "Mathematical Theory of Rocket

McGraw-Hill Book Company, Inc., New York, 1947.
Flight,"
2
R. A. Rankin, Trans. Roy. Soc. London (A), 241, 457-585, (1949).

Mkm. artillerie frang., 25, 253-360 (1951).
P. Carriere,
168
VARIABLE-COEFFICIENT LINEAR SYSTEMS

produced by the motion
vertical plane
in the horizontal plane.
169
There-
fore the characteristics of the rocket are not lost
by confining our conplane and studying the motion in that plane
siderations to the vertical

only.
We
shall consider the earth to
Let v be the magnitude
rockets.
and
inclination of the velocity vector,
rocket (Fig. 12.1).
Then the angle
a
Let
be the mass
of the rocket,
gravitational attraction
be
flat for
short-range artillery
of the velocity of the rocket, 6 the

<j>
of
<fr
and
the inclination of the axis of the
attack
of the rocket is
(12.1)
g the gravitational constant.
thus a vertically downward force mg.
is
The
The
FIG. 12.1
aerodynamic forces are the
lift
and parallel
and the moment M. All these
L, the drag D, perpendicular
to the direction of motion, respectively,
forces act on the center of gravity of the rocket.
8 of the rocket motor may have an
The constant
angular misalignment
ft
arm 8 with respect to the center of gravity, as indicated

Then the equation for acceleration along the trajectory is
m jr =
cos (a
ft)
- mg
sin 6
-D
thrust
and a moment
in Fig. 12.1.
(12.2)
Cit
The equation
for acceleration
mv -77 = $
at
normal to the trajectory
sin (a
ft)
mg
cos
is
+L
(12.3)
the radius of gyration of the rocket about a lateral axis

through the center of gravity, the equation of angular acceleration is
Finally,
if
is
mk^ =S5-M
(12.4)
In Eq. (12.4), we have neglected the so-called jet damping moment since
small in comparison to the restoring moment of the tail fins.
it is
170
The aerodynamic
and the moment are dependent upon the
forces
But
the fins on the rocket are not properly aligned,
angle of attack
the lift and moment will not vanish even
<*.
ment
that
air
if
can be accounted for
L and
M vanish not at a
and d the diameter
coefficient
by
Ly
if
The
is zero.
misalign-
introducing a misalignment angle 7, such
but at a
the density of
If p is
7.
body, we can introduce the
of the rocket
and the moment
the drag coefficient Kr>,
coefficient
lift
KM
as follows:
K L pv*d sin (a -~ 7)
K DpvW
K upv*d* sin (a ~ 7)
L =
D =
M=
For short-range
artillery rockets, the
summit
(12.5)
(12.6)
(12.7)
of the trajectory is
therefore the density p can be taken as a constant.
maximum
velocity
is
small enough so that
all
not high;
Furthermore, the
the coefficients
KL KD
,
can be considered as constants, -not influenced by the Mach

number on the trajectory. Moreover, for short-range rockets, the pro-
KM
and
pellant fraction of the total
mass
is
small; then the
mass
can be taken to be a constant without serious error.
of the rocket
Equations (12.1)
to (12.7), together with these simplifying assumptions, then determine
the trajectory of the rocket.
The burning time
of this type of rockets is very short, say

of a
This necessarily makes the acceleration S/m very large. In
the thrust S is so large that the gravitational and drag forces are
-J-
second.
fact,
negligibly small in comparison.

line
from the
flight direction,
Furthermore, the deviation of thrust

Therefore an immep, is never large.
a.
diate zeroth-order approximation of Eq. (12.2)
at
is
v
(12-8)
'
Therefore the zeroth-order approximation to the trajectory
is given by a
The motion along
this straight line is one of constant acceleration S/m.
Thus if z is the
distance measured along this line, then the motion is represented by
straight line with the initial inclination
^. m
If
the rocket
is
(Fig. 12.1).
*
v
\m
launched without any
distance from the launching point.
initial velocity,
then z
of
it.
From Eq.
dt
(12.9)
the true
If there is initial velocity, z is
the distance measured from the launching point, but from
ahead
is
we have
'
not
some point

With
we can
this zeroth-order solution,
171
calculate the first-order solution
This will be done presently.

Linearized Trajectory Equations. Since the deviation of the
of the rocket trajectory.
12.2
trajectory from the zeroth-order solution during burning is never large,

we can replace the velocity v and the time derivative whenever they
occur in Eqs. (12.3) and (12.4) by those given by Eqs. (12.9) and (12.10).
ft is small, sin (a
Furthermore, since a
ft) can be replaced by a
ft.
We shall also neglect the lift L, since
cos 6 can be replaced by cos ft.
it is
small in comparison to the lateral thrust component and the weight

With these simplifications, Eqs. (12.3) and (12.4) become
of the rocket.
2z~ =
ft)
cos
(12.11)
ft
^j-
and
where
is
cr
defined
by
and evidently has the dimension
of a length,
characteristic length of the disturbed
wave length
considered as the
(12.11),
(12.1),
unknowns
a, 6,
and
a can be taken as the
of the rocket,
and
of the disturbed trajectory.
(12.12) are
and
motion
may
be
Equations
then linearized equations for the three

is made under the assumption of
Linearization
<j>.
small deviations from the straight-line trajectory at the ideal launching

angle
We
this,
ft.
can eliminate
we
and
to obtain a single equation for a.
<
by 2 \/z and
Then we have
divide Eq. (12.11)
tion with respect to
Vrz dz*
d*0
-j-;
Now we
equation.
--
d6
z.
2V* &
-p.
-j-
- &---VA
da
/
F{a -
Q
ft
am
- ~r
cos
S
4s
2\/^
divide Eq. (12.12) by 2 \/z and subtract from
Then, using Eq.
(12.1),
To do
differentiate the resultant equa-
it
the above
we have
ffa
da
4?r
ff
4z
Vz \
This equation clearly demonstrates the fact that the controlling differan artillery rocket is not an equation
ential equation of the stability of

of constant coefficients.
In
fact, this
equation can be put into the
172
standard form of Bessel's equation by introducing the nondimensional

distance
{,
defined
by
$
(12.14)
0"
where
<r
is
the
"wave length"
specified
by Eq.
(12.13).
Then we have
(12 15)
'
When a
can be calculated by integrating the following

determined,
from
obtained
Eq. (12.11):
equation,
is
The independent variable z or | in the differential equations is not a

time variable but a distance variable. However, since, as shown by
Eq. (12.9), z and thus are monotonic functions of t, the stability of the
is not modified by changing the independent variable from t to
the system is stable in terms of ,'it is stable in terms of t, in the sense
that the deviations from the ideal straight-line trajectory will decrease
system
if
with increase in | or t. Therefore the problem of stability can be ade{ increases from the initial value
quately discussed with the variable
.
at
velocity
12,3
bility,
The
as time increases.
is
initial
value
is
zero
if
the launching
zero.
Stability of
we must
an
Artillery Rocket.
solve Eqs. (12.15)
and
To
discuss the question of sta-
(12.16) with the specified initial
conditions and then determine whether the angle of attack a or,

appropriate, the deviation of the inclination of the trajectory, 6
tends to zero as
of
elementary functions.
Equation (12.15)
The
and
8^
actually BessePs equacomplementary functions are thus Bessel
increases.
tion for the order %.
functions of order
more
is
They are, however, expressible in terms

In fact, Eq. (12.15) can be rewritten as
.
where
f
= Vf
(12.18)
and
+ eo M ,
Therefore the complementary functions for f are simply sin
(12.19)
and cos

The
conditions of the rocket as
it
173
leaves the launcher, or the initial
conditions, are
=
=
=
V
6
(X.
and
The
d$/dt
"
(12.20)
ao
(d<t>/d)Q
thus indicates quantities at the instant
subscript
initial
VQ
0.
The
values of f and f are thus
and
fo
By
using Eq. (12.16),
V&ao
we have
at
(12.22)
cos
2V*
But
<
a,
thus
=a -
+T
,
TT
\Vo
0<
17
_^/r
V fo
-IT
cos
(gm/S)
-
1
,
or explicitly, then
<r(d/(tt)o
(gm/S) cos
ft
f
With the
initial conditions so translated,
(f,fo)
-7=
cos
we can
down
the solu-
)GW ^1
(12.24)
write
or f or a immediately,
tion of Eq. (12.17)
Vf
+ -L sin
[
fo
fo)
sn
17
L
(f
fo)
[^
+ T cos
fe
where Q
is the forcing function specified by Eq. (12,19).

Since Q(ij)
contains half powers of ??, the integrals of Eq. (12.24) are actually Fresnel
integrals.
When a is
calculated,
Eq. (12.16) then gives
by quadrature,
We can further separate the effects of different disturbances at the

launcher by writing Eq. (12.25) in a series of terms, each representing
174
Thus
one type of disturbance.
8,
(0o
fc)
ft
cos
<?!,&)
(12.26)
The first term represents the effect of the initial deviation of the trajectory
The second term gives the effect of thrust misalignment and
angle.
On
-0,4
10
12
14
FIG. 12.2
gravitational pull.
moment
ment.
The
The
last
rocket.
The
of the thrust.
fifth
third
term shows the influence
The fourth term
term represents the
term gives the influence
Each
of the turning
gives the effect of fin misalign-
effect of the initial angle of attack.
of the initial angular velocity of the
of the (7s is a function of the
two variables
and
fo,
and
a combination of FresnePs integrals. Rosser and his collaborators call

them the "rocket functions " and tabulate them in their book. Figures
is
12.2 to 12.5, representing these functions graphically, are also taken
from
this book.
An
inspection of Figs. 12.2 to 12.5 immediately brings out the fact

the G functions have persistent values for large
Thus the
disturbances do not damp out. The first and the last two terms of
that
all

Eq. (12.26) represent the
effects of
the
initial
175
disturbances at the launch-
The other
ing point.
nonvanishing values at large
in
are
the
to
due
the
or
terms
Eq. (12.26)
forcing
"output"
"input/
functions.
They also have values for large f The behavior of the <?i
They have
finite
even worse: at large

without
limit.
increases
Therefore,
function
is
it is
if
approximately log
we use the previously
and thus
established
0,75
criterion for stability of a system,
i.e.,
the vanishing of initial disturbances
the output under "reasonable" forcing funcOn the other hand, the
tions, then the artillery rocket is unstable.
complementary functions of the basic equation, Eq. (12.15), are Bessel
and the boundedness
of
functions which vanish for large values of the variable and thus seem
If we try to apply our experience
to indicate stability of the system.
with systems which are described by a linear differential equation with

constant coefficients, the behavior of a system with variable coefficients
176
However, this only indicates the inapplicability

confusing.
concepts developed for systems with constant coefficients to systems
with time-varying coefficients. A new approach to the problem of stabilapparently
is
of
ity
and control
is
required.
A discussion
on
this point will
be presented
in the following section.

0.6
0.5
0.3
0.6
03
0.2
0.1
10
12
14
FIG. 12.5
12.4
Stability
and Control of Systems with Variable
For a linear system with constant
Coefficients,
our previous study has

shown that the satisfactory performance of the system with respect to
stability and other control criteria can be guaranteed if the solutions
of
coefficients,
the homogeneous equation without forcing function or input are all

damped. Therefore, although the forcing function or input
sufficiently
considerably from one case to another, the performance criand thus the design of the system are based upon the study of the
solutions of the homogeneous equation or the complementary functions
of the complete equation.
This is the fundamental principle in the theory
of servomechanisms.
The actual method of analysis by the technique of
may vary
teria
transfer functions based
upon the Laplace transform is merely a useful

In principle, for instance, the classical methods for finding the
complementary functions are just as good as the root-locus method of
trick.
Evans.
Our preceding discussions in this chapter definitely showed that for

a linear system with time-varying coefficients, the fact that all the
complementary functions of the equation are damped is no guarantee
for the satisfactory
With some input,

in spite of the
performance of the system with forcing functions.

may even be unbounded,
or forcing function, the output
damped complementary
functions.
Then the question
of
satisfactory performance of the system cannot be answered without
177
This requirement of specifying the input,

together with the difficulty of actually determining the solution of a nonhomogeneous differential equation with variable coefficients, seems to
knowing the input function.
make the task of developing a general theory of stability and control for
such systems a hopeless one. However, we must distinguish the computational difficulties from the real difficulty in organizing a general theory.
Computational difficulties can be removed by fast electromechanical
computers and should not be considered real difficulties. When this is
realized, we see that specifying the input function for performance
analysis is in fact designing for a specified purpose: we must know first
what we want from the system under what circumstances, before we can
When this approach is adopted, the general problem
of stability is suppressed, because, if the system is designed to have a
For linear
specified satisfactory performance, that in itself is sufficient.
design the system.
systems with time-varying

can be formulated on this
coefficients,
basis.
a general theory of control design

is an application of the classical
This
perturbation theory and will be discussed in the following chapIn retrospect, we may say that the theory of conventional servomechanisms is a theory for general design of a specific type of systems.
ballistic
ter.
The perturbation theory of the following chapter

more general type of systems.
design of a
is
a theory of
specific
CHAPTER
13

The
object of the ballistic perturbation theory is to calculate the behavior

near the so-called normal trajectory. The normal tra-
of a projectile
jectory
is
a certain trajectory with specified initial conditions, propul-
and programed lift and drag. If

the actual conditions are slightly different from these specified conditions
or if the vehicle is disturbed from its normal trajectory by accidental
wind gusts, the trajectory of the projectile will be different from the
sion program, atmospheric conditions,
normal trajectory.
But
if
such disturbance influences are small, the
perturbed trajectory will lie in the neighborhood of the normal trajectory,

and the difference of the perturbed trajectory and the normal trajectory
will
remain small.
This fact of nearness to a calculated,
known
tra-
the basis for the linearization of the differential equations of

jectory
motion for the perturbed trajectory. The perturbed system is then
is
represented as a linear system with time- varying coefficients, time varying because of the varying conditions of the projectile with respect to
time.
The original purpose of ballistic perturbation theory was to calculate

the small modification of the trajectory of a projectile due to deviations
of the weight of the shell from standard value, to changes of atmospheric
However, with the advent of modern
computing machines, the tendency was to calculate all
conditions, to effects of wind, etc.

large
and
fast
neighboring trajectories separately. The usefulness of ballistic perturbation theory then vanishes. However, the problem of designing the
control for linear systems with variable coefficients is just the problem
to which the ballistic perturbation theory can be applied.
shall
show this in this chapter by studying the control problem of a long-range
We
This particular problem was studied by Drenick. 1 Our

however, more complete and includes the problem of the
rocket vehicle.
treatment
is,
automatic guidance of such vehicles. 2

13.1 Equations of Motion of a Rocket.
matters, the vehicle
1
is
assumed to move
In order not to complicate
in the equatorial plane of the
R. Drenick, /. Franklin InsL, 25, 423-436 (1951).

H. S. Tsien, T. C. Adamson, E. L. Knuth, J.
Cf.
(1952).
178
Am.
Rocket Soc., 22, 192-199
179
The planar motion is possible

rotating earth, as sketched in Fig. 13.1.
due to the absence of the cross Coriolis force in the equatorial plane. The
coordinate system is fixed with respect to the rotating earth, i.e., it
actually rotates with the angular velocity 0, the speed of earth rotation.
In the equatorial plane, the position of the vehicle at any time instant t
is
specified
by the radius
and the angle
from the starting point
of the
Elevator
Flight Path
FIG. 13.1
vehicle,
ro is
the radius of the earth,
is
the gravitational constant
at the surface of the earth without the centrifugal force due to rotation.
Let
mass due to thrust and aerodynamic

and the circumferential direcThen the equations of motion of the center of grav-
and 6 be the
force per unit
forces acting on the vehicle in the radial

tions, respectively.
ity of the vehicle are
dr
(13.1)
Q)
where the plus sign in the second terms on the right will be valid for
west.
flights toward the east, and the minus sign for flights toward the
180
The
forces
drag D.
Let
and 9 are composed
of the thrust S, the lift L,
and the
W be the instantaneous weight of the vehicle with respect
the gravitational constant, and 7 the magnitude of air velocity

Then it is convenient to introduce the parameters
relative to the vehicle.
to
g,
and
S, A,
A, defined as
- so
- Lg
be assumed that the natural wind velocity w is in the horizontal

and in the equatorial plane, with positive sign if it is a head
wind, w is considered to be a function of altitude r. If v r is the radial
It will
direction
velocity
and
v$
the circumferential velocity,
i.e.,
(13.3)
then the relative
air velocity
is
computed as
If j8 is the angle between the thrust line and the horizontal direction, then
the radial and the circumferential components of the forces R and 9 per
mass are
unit
R =
9 = S
If
is
the
moment
by the moment
acceleration
To
sin
ft
cos
+
-
of forces
+ to)A - v
- (vo + w)
Vrk
(v s
(13.5)
about the center of gravity, divided
of inertia of the vehicle, the
equation for the angular
is
specify completely the motion of the vehicle, the
and the moment
lift L, the drag D,

about the center of gravity have to be given as
According to aerodynamic convention, L and D will

be expressed in terms of the lift coefficient C L and the drag coefficient CD
functions of time.
as follows:
L =
where
/>
is
the air density, a function of the altitude r, and A is a fixed

In the present problem,
reference area, say the wing area of the vehicle.
181
motion of the vehicle is restricted to the equatorial plane, the

attitude of the vehicle, essential for aerodynamic calculations, is determined by the angle of attack a, that is, the angle between the thrust
since the
line
and the
motion
The
relative air velocity vector (Fig. 13.1).
control on the
of the vehicle is effected,
The parameters which
will
however, through the elevator angle 7.

affect C L and C D are thus a and 7.
In addi-
tion, the aerodynamic coefficients are functions of the Reynolds number

Re and the Mach number M. If a is the velocity of sound in the atmos-
phere, a function of the altitude
r,
the
Mach number
is
M=~
IfZ
is
(13.8)
a typical length of the vehicle and $
a function of the altitude
r,
is
the viscosity of
the Reynolds number
is
air,
again
Re =
(13.9)
/*
Therefore
CL =
?L(a,T,Af,Re)
CD = 0D(a,7,Af,Re)
(13.10)
We shall assume that the thrust line passes through the center of
gravity of the vehicle; thus the thrust gives no moment. Since the
angular motion of the vehicle during the powered flight is expected to be
The only
slow, the jet damping moment of the rocket is negligible.
moment acting on the vehicle is then the aerodynamic moment m. m can
also
be expressed as a
CM
coefficient
m=
as follows:
%pV*AlC M
The moment coefficient CM is again a function

My and Re, or
CM = <7*(a,7,Jlf,Be)
If
the
is
the instantaneous lateral
moment
center of gravity of the vehicle,
(13.11)
of the four variables a, 7,
(13.12)
of inertia of the vehicle
then the magnitude
of
about
in Eq.
(13.6) is
With the notations defined above, the system

is
as follows:
(13.13)
j
of equations of
motion
182
*
dS
=
^vj
r
dt
dt
^~ = S
sin
(v 9
(13.14)
=F
tlVe
^
= S
cos
/5
A 4
yr
(0*
-^-
+
i
\ \
u;)A
o
2y r
yv
_i_
^-
+
I
^"r
I
=
cos
- Â -
This system of equations
(?
(v,
is
a set of first-order equations for the six
to)A
2v r
AT
unknowns r, 6, /3, tv, #0, and /3. To solve it, the six initial values at the
= 0, for the unknowns must be specified. In addition, the
start, when t
thrust 5, the weight TF, and the moment of inertia I must be given for
every time instant. To determine the aerodynamic forces the elevator
The properties of the
angle 7 must be specified as a function of time.
atmosphere must be known; i.e., the wind velocity w, air density p, air
and sound velocity a must be given as functions of the alti/z,
viscosity
The angle of attack a of the vehicle cannot be specified; it is

r.
a quantity to be computed from the angle p and the relative air velocity
vector V.
tude
Let the properties
of the
atmosphere be standardized, and the average

and its power plant be taken to be repre-
characteristics of the vehicle
Then if the elevator angle 7 is specified as a function of time,

the flight path of the vehicle is determined and can be calculated by
The actual execution of this
integrating the system of Eq. (13.14),
sentative.
computation will probably be done on an electromechanical computer.

This flight path of a standardized vehicle in standard atmosphere can be
called the
normal
flight path.
The dominating
feature of the normal flight path is its range.

This
the distance between the take-off point and the landing point.
problem of navigation is then to calculate the proper time for cutoff
range
The
is
motor and the proper variation of the elevator angle during

that the resultant range is that desired. This problem of navigation for the standardized vehicle in the standard atmosphere can be
solved mathematically before the actual take-off of the vehicle, since all
of the rocket
flight so
information for the normal flight path
is
known
or specified beforehand,

Perturbation Equations.
13.2
183
Natural atmospheric characteristics
do not necessarily coincide with those specified for the standard atmosThe wind velocity at each altitude changes according to the
phere.
weather conditions; the temperature T is also a varying quantity.
Therefore one should expect variations from the normal flight path due
The
to changes in atmospheric conditions.
actual vehicle also
may
be
somewhat different from the standardized vehicle in weight, in rocket

motor performance, etc. Therefore the actual flight path will be different from the normal flight path if the elevator angle program of the
normal path is used. The problem of navigation of an actual vehicle
is
that of correcting the elevator angle program so that the range of the
same as the normal flight path and the destination
actual flight will be the

is
reached without error.
Because of the rapidity of
flight, this
navi-
gational problem cannot be solved by the conventional method, which

neglects completely the dynamic effects and is based upon only kine-
But instead, the problem should be solved by

an automatic computing system, which responds to every deviation from
the normal conditions with a speed approaching instant action.
The
matical considerations.
problem
is
thus more appropriately called the guidance problem, and the
control system, the guidance system.
The
general problem of guidance
is
very
difficult
indeed.
However,
the deviations from the normal conditions are expected to be small, since
the normal flight path is, after all, a good representation of the average
situation.
This fact immediately suggests that only first-order quantities

"
'
need be considered. This linearization is the basis of the
7
in deviations
ballistic perturbation theory.

The resultant system will have coefficients
that are evaluated on the normal flight path and are generally functions
of time.
Therefore the fundamental equations in the ballistic perturba-
tion theory are linear equations with time-varying coefficients.
Our
discussion of the guidance problem of a long-range rocket vehicle
is
thus an example of control design for such systems. The particular

design specification here is the vanishing of range error. The controlled
"input" here is the elevator angle corrections. We shall develop these
concepts in the following discussion.
Let the quantities of the normal flight path be denoted by a bar over
them, and deviations, by the 5 sign. Thus for the actual flight path,
r
5r
50
,.
_",_,_,
(13.15)
The deviations
of
the actual atmosphere from the standard atmosphere

5p, the deviation of temperature
are expressed as the deviation of density

5T,
and the deviation

p
of
dp
wind velocity 5w, thus
T = f
ST
w = w
dw
(13.16)
184
we assume no change
If
in the composition of the atmosphere at any

standard atmosphere, the knowledge of dp and
The
sufficient to calculate the deviation of pressure, if necessary.
from that
altitude
5T
is
of the
deviation of the actual vehicle from the normal vehicle

limited only to the deviation of weight
That
of inertia 81.
dW
is
assumed to be
and the deviation
of
moment
is.
W + dW
= 1+81
(13.17)
S is assumed to be fixed at the standard value. The wing

and the aerodynamic characteristics of the vehicle, as expressed
by Eqs. (13.10) and (13.12), are also assumed to be invariant.
Substituting Eqs. (13.15) to (13.17) into Eq. (13.14) and then subtracting the corresponding equation for the normal flight path, we have
The
thrust
area
the following equations, according to our linearizing principle:
ddr
(13.18)
dv r
~HT
0,1
dr
+a
2 5/3
a 3 dv r
a4
dve
+ a$ dp
dT
61 dr
dt
ddft
&2
64
a 8 dw
$7
+ a* dW
&6 8p
(13.19)
+6
101
coefficient a's b's,

;
defined
by Eq.
+c
The
65
and
c's
dT
dT
68
Bw
69
dW
+ CB dw + c dW + C
9
1Q
dl
are partial derivatives of the F, G,
and
(13.14), evaluated on the normal flight path.
For
example,
(13.20)
The
details of this calculation of coefficients are given in the
this chapter.
appendix to

Equations (13.18) and (13.19) are linear equations with variable
cients for the six deviation quantities.
pheric properties
dp,
5T,
and
The problem
5/3.
if
dW, and 81 are
8y,
equations determines
of differential
coeffi-
the deviations of the atmos-
and 5w are known, and
specified, then this system

Sv r , $V0,
If
185
5r, BO, 5/3,
guidance is, however, different from this.

the function 5% or the correction to the elevator angle,
What is required is
of
such that the range error is zero. As suggested by Drenick, this guidance
1
problem can best be solved by the method of adjoint functions of Bliss.
13.3
tions
is
system
Adjoint Functions. The principle of the method of adj oint funcLet y l (t), where i = 1,
n, be determined by a
as follows.
of
linear equations
for i
where o# are given
coefficients
which
1,
,n
(13.21)
be functions of the time
may
t.
the "forcing" functions or inputs. Now introduce a new

set of functions X$), where i = 1,
n, called the adjoint functions
to yi(t), which satisfy the following system of homogeneous equations
The
Yi(t) are
*
ai>
=
) ofa
t4
1,
(13.22)
y=i
By
multiplying Eq. (13.21) by

equations over i, we have
n
) *# -
ti
\ and Eq.
(13.22)
by
X A ^'
~~
Xi2/y
^'^
LJ
t*iyi
The two parts in the double sum
^ and summing the
^ iYi
t-i
evidently cancel each other, and
\Yi
-i
(13.23)
i-i
This equation can be integrated from the instant

t
2,
ti
to the instant
and
V
%
'*''
ffl
Bliss calls this equation the
1
we have
;4
t=tl
Jti
V
l
\t(
fundamental formula.
G. A. Bliss, "Mathematics for Exterior Ballistics," John Wiley
York, 1944.
&
Sons, Inc.,
New
186
For the problem

quantities, that
n =
is,
the
of long-range rocket,
=
-
y<
are the perturbation
and
5r
2/2
ov r
2/5
=
=
(13.25)
Then, according to Eq. (13.19), the adjoint functions X$) satisfy the
following system of equations
61X5
Ht
^2X5
+
(13.26)
fr
X3
The inputs Ft are

-
T = 7 = F =
2
F =
F =
4
5
F =
6
13.4
a 5 ^7
5p
ST
5p
65
Range
dW
dw
5w
(13.28)
c9
c 5 57
Correction.
functions completely.
at a certain instant.
(13.27)
+ a 8T + a
+ &T T + 6
8W
Equation (13.26) does not determine the X
To do that, a set of values for X must be specified

What values to pick for X at what instant depends
For our guidance

specific purpose of the control design.
problem, we require zero range error. Therefore the quantity of interest
is 86 at the instant of landing, or #2 if the subscript 2 denotes quantities at
the instant of landing. This is sufficient to determine the X s completely.
upon the
We
If
shall see this presently.

t-2
is
the time instant of landing of the actual vehicle and
1%
the time
instant of landing of the normal flight path, then

T
j.
fr2
T"
jj
^2
(13.29)
Similarly
f2
$2
=
=
fa
<5r 2
^2
"4"
^02
(13.30)
187
but
502
dr 2 is
However,
the surface
of
by
= -
(13.31)
(ve
definition zero, because landing
or r 2
earth,
f2
By
means contact with

from Eq.
eliminating
(13.31),
-MOO
W
L
if
Therefore,
the magnitudes of X
(13.32)
at the landing instant
fa
are
specified as
X2
=
(13.33)
then the error in range
=
When
[Xi 5r
+X
the normal
is
given by
X6
flight
path
is
50]^
(13.34)
determined, the coefficients in Eq.
(13.27) are specified as functions of time.
These equations together
-Actual Path
Normal
Flight
Path
FIG. 13.2
with the end conditions of Eq. (13.33) then determine the adjoint funcThe integration has to be performed " backwards" for t < 1%,
tions \i.
perhaps by an electromechanical computer. With the adjoint functions

so determined, one can use the fundamental formula of Eq. (13.25) to
for the range error given by Eq. (13.34): Let ti
denote the time instant for the power cutoff for the normal flight path.
modify the equation
Then the
condition for the error in range 50 2 to be zero can be expressed as
188
50 2
[\i tr
+X
+X
50
dp
+X
dv r
4
2
[X 4
+ X 8v + X 50]^
r4 + X F5 + X F ]^
6
(13.35)
This
is
It will be exploited in the
the basic equation for guidance.
following sections.
13.5
Cutoff Condition.
disturbances
The condition of Eq. (13.35) for arbitrary

into two parts; the sum and the integral
down
can be broken
Therefore the condition to be satisfied
are set equal to zero separately.

at the normal cutoff instant ti is
[Xi 8r
X 2 50
+X
+X
3 <30
Since the normal cutoff instant
t\
+X
Sv r
ti
dv 9
X 6 dft]^
(13.36)
a standard time instant, but not
is
necessarily the actual cutoff instant h,

ti
i.e.,
(13.37)
Bti
Eq. (13.36) should be converted into a more useful form involving the
This is easily done, because up
quantities at the actual cutoff instant.
to the first-order quantities, according to Eq. (13.35),
or
Similarly,
where F,
are the values of these quantities, given by Eq. (13.14),

(5, and
evaluated on the normal flight path. In fact they should be evaluated
at an instant just before the normal cutoff time ti so that the
accelerating
force of the rocket is included
those of a powered
flight.
and the rates
Now
define
of
/ and J
change
of velocities are
as follows:
[X*r
+ X*0 + X*/3 + Xftv + Xft>, + Xf/3],^
J =
[X*f
(13.38)
and
\*d
where X* are the values
Then the condition to be
of
+ \*j +
X,-
\*v r
+ \*v + X*/3U
9
evaluated at the normal cutoff time
satisfied at the actual cutoff instant
ti is
ti.

j =J
\*v r
L
+ X* 4r +
+ XfF + Xf<? +
X*/3
X*# 1
(h
fx)
189
(13.39)
J*-*i
the equation for determining the proper instant of power cutoff.

the normal flight path is known, J and the quantity within
the bracket to the right in Eq. (13.39) are fixed. Then the whole right-
This
is
When
hand
side of Eq. (13.39) can be considered as a linearly increasing func-
tion of time
if
is
substituted for
Simultaneously
ti.
can be com-
by using the predetermined X* and

the values of position and velocity of the actual vehicle obtained by
The magnitudes of the quantities on the two sides of
tracking stations.
puted at every instant before cutoff
Eq. (13.39) can then be continuously compared. When they are equal
is satisfied.
Then the power cutoff signal is
the
rocket
is
off.
and
shut
power
given,
to each other, Eq. (13.39)
When
Guidance Condition.
the rocket power is shut off earlier

the propellant left in the tank,
?i,
and the moment of inertia /
if not dumped, will alter the weight
It is also possible that the pay-load of the vehicle is
of the vehicle.
13.6
normal
or later than the
cutoff instant
not that specified for the standard vehicle. Then, after power cutoff,
there is a fixed dW and 1, fixed in the sense that they do not change
with time and are
known once the power
character are the deviations
from the standard atmosphere.

In the following,
measured.
it
Of a different
cutoff is effected.
8T and dw
of the actual
atmosphere
These are not known unless they are
is proposed to use the vehicle itself as a
8p,
measuring instrument, and we proceed as follows.

After the cutoff condition
the condition for zero range

Now since the
is satisfied,
error is that the integral in Eq. (13.35) should vanish.
Yi in that integrand involve arbitrary disturbances
known
beforehand, this condition
itself
vanishes.
(X 4a 5
+X
That
&5
(X 4 a 7
is,
X 6 c 5 ) dy
X 56 7
if
and dw not
the integrand
according to Eqs. (13.28),
+ X & + X c 5p
+ X c dT + (X a + X & + X c
+ (X a + X & + X c 8W
(X 4 a 6
7)
or,
8T,
dp,
can be satisfied only
6)
8)
8w
9)
with the following notation
(13.40)
J)
this conditiorj.
(X 4 &9
-f-
XsOg
-f-
XeCg)
5TF
XeCio
81
can be written as
4 8y + d* 8p + d
8T
+d
8w
= D
(13.41)
190
Equation (13.19) can be rewritten as
+
h+
a 5 8y
fes
c6
a,5T
+
c8
<
<
= A
= B
= C
(13.42)
where
-
at
^
&i or
(13.43)
02
at
~~~
Cz dV r
at
the tracking stations for the vehicle will measure the quantities A,
Bj and C, then the atmospheric disturbances dp, 8T, and 5w can be
determined by solving for these variations using Eq. (13.42). This is
If
measuring instrument for 5p, 5T,

8w
are
known, Eq. (13.41) gives the proper
dT, and
essentially using the vehicle itself as a
and 8w.
When
5p
elevator angle correction Sy.
mathematically equivalent
directly for By using the

$6
system
a?
way
of
to calculate dy would be to solve
Eqs. (13.41) and (13.42).
A
B
67
C7
Thus
(13.44)
This equation specifies the necessary change in the elevator angle at

every instant, to be calculated from a's, 6 s, c's, and A, B, C, D at the
?
same
instant.
These quantities consist partly of predetermined inforflight path, and partly of measured information
mation from the normal
on the position and the velocities of vehicle obtained by tracking the

At high altitudes where the air density is very small, the aero-
vehicle.
dynamic forces will be almost negligible in comparison with the gravitational and inertial forces.
Then the quantities A, B, and C of Eq. (13.43)
be the small difference of large magnitudes. These are then the
If the actual elequantities most difficult to determine accurately.
will
vator angle is made to conform with the one calculated by Eq. (13.44),
then in conjunction with the proper power cutoff, as specified in the last
section, the vehicle will be navigated to the chosen landing point in spite
of the
atmospheric disturbances.
Guidance System. When the general character of the flight

path has been chosen from over-all engineering considerations, the first
step is the calculation of the normal flight path using the properties of
13.7
191
the standard atmosphere and the expected performance of the vehicle

with normal weight. The knowledge of the normal flight path then
?
Equation (13.27) together with the end

s, and c's.
conditions of Eq. (13.33) allows the calculation of the adjoint functions
AV All this information should be on hand before the actual flight of the
"
stored data."
vehicle and may be called the
determines the
a's,
fc
Before the power cutoff, the elevator angle may be programed according to that for the normal flight path, and the stability of the vehicle
by the auxiliary rockets. The tracking

into action and supply the vehicle with
immediately
however,
go
stations,
information on its positions and velocities. This information goes first
is
supplied by the jet vanes or
computer which, using the stored information, continuthe magnitudes of quantities on the two sides of Eq.
compares
ously
into the cutoff
(13.39), the cutoff condition.
When that
condition
is satisfied,
the power
cutoff is effected.
At the instant
of power cutoff, the tracking information is switched

computer for the guidance system. The instant of power cutoff
also fixes the amount of propellant in the tank and thus determines the
and the moment of inertia / from the normal.
variations of the weight
to the
This information together with the stored data on the normal flight path
then allows the computer to generate the elevator correction angle $7
according to Eqs. (13.40), (13.43), and (13.44). Theoretically the value
of ^7 must be obtained without time delay from the instant when the
information is received, because Eq. (13.44) is a condition of equality
of
two quantities evaluated at identical time instants. The computed

combined with the elevator angle f, determined for the
correction dy
normal flight path, then gives the actual elevator angle setting 7. The
design of the control mechanism for the elevator from here on can follow
the practice of the conventional feedback servomechanism, with the
usual criteria of quick action, stability, and accuracy.
The general
scheme of the guidance system can then be represented by Fig. 13.3.
The computers envisaged here are carried in the vehicle and receive
the information on positions and velocities of the vehicle from the fixed
ground tracking stations along the flight path. Then, as indicated in
Fig. 13.3, this is the feedback link.
Properly designed computers here
assure us of the specified performance of the system and thus function

as the compensating circuit or the amplifier of the conventional servo-
mechanism.
In over-all conception then, our guidance system is very

servomechanism studied in the preceding chapters.
similar to the simple
However, the guidance system is a system of great complexity. Its

design requires the theory of perturbation together with the concept of
Our example of the guidance of a long-range rocket,
adjoint functions.
although somewhat oversimplified, serves the purpose of showing how
192
the theory can be used to design the control system. In the example,
In
there is only one design criterion, i.e., the vanishing of range error.
more complicated systems, more than one design criterion can be speci-
Neverset of adjoint functions will be required.

in the
shown
that
as
the
same
will
still
be
the
design principle
theless,
fied,
and more than one
simple example.
Tracking Information
Tracking Information
PIG. 13.3
13.8
Control Computers.
Although it is not the purpose here to disand the engineering of any components
cuss the detailed construction

of
the systems considered, the role of the computer,
connection with
switching in
first
introduced in
so important in the
optimum
Chap.
more advanced control systems that a general discussion of its characteristics and requirements would be appropriate.
For details, the reader
10, is
should refer to books on this special subject. 1

There are in current use two different kinds of computers: the analog
computer and the digital computer, The analog computer is just what
its
name
implies: a physical analogy to the type of problem its designer
It is then a system which is described by the same

mathematical formalism as the computation to be carried out. The
wishes to solve.
Hill
Engineering Research Associates, "High-speed Computing Devices," McGrawBook Company, Inc., 1950. For d-c analog computers, see G. A. Korn and T. M.
Korn, "Electronic Analog Computers," McGraw-Hill Book Company,

York, 1952.
Inc.,
New
193
inputs to the machine are in terms of the value of some physical quantity
an electric voltage or current, the degree of angular rotation of a
The machine transshaft, or the amount of compression of a spring.
forms these inputs into other physical quantities, the outputs, according
to the rules of its construction, chosen by the designer to represent the
The inputs to an analog computer

specified mathematical procedure.
are then the instrument readings of the several physical quantities of
the system to be controlled, and the outputs of the computer are com-
mand
signals fed
directly to the individual
servomechanisms
of
the
controlled quantities.
In contrast to the analog computer, a digital computer works by

Data on the problem must be supplied in the form of num-
counting.
machine processes this information according to the rules of

arithmetic or other formal logic demanded by the computing problem,
bers; the
and expresses the final result in numerical form. There are two very
important consequences of this manner of operation. First, input and
output equipment, or the "transducers," must be designed to make an
appropriate translation between the logical world of the digital machine
and the physical world of the controlled system. Second, the problem
to be solved must be formulated explicitly for the digital computer.
In
the case of the analog computer, the problem is implicit in the construction of the machine itself; construction of a digital computer is determined not by any particular problem or class of problems but by the
logical rules
lar
which the machine must follow in the solution
of the particu-
computing problem.
In comparing digital and analog computers as components in a control

we observe, first, that for simple control applications the analog
system,
machine
is
almost always
less elaborate
Even the most elementary
digital
unit, a storage unit, a control unit,
than a digital machine would be.
computer requires an arithmetical

and input and output transducers.
For simple problems, this array of equipment is wastefully elaborate. In

contrast, an analog computer need be no more complicated than the
problem demands. In fact, as mentioned before, the compensating
circuit in a simple
servomechanism
is
such an analog computer.

as for the guidance
As the computing task becomes more complex,
problem discussed in this chapter, the analog machine loses its advantage,
and we see a second fundamental difference between the two types of
The analog computer, being a physical analogy to the problem,

must be more complicated for more complicated computing problems.
If it is mechanical, longer and ever longer trains of gears, ball-and-disk
integrators, and other devices must be connected together; if it is electric,
more and more amplifiers must be cascaded. In the mechanical case,
the inevitable looseness in the gears and linkages, though tolerable in
machine.
194
simple setups, will eventually add up to the point where the total backlash or "play" in the machine is bigger than the significant output
quantities,
and the device becomes
useless.
In the electric case, the
which always occur in electric

circuits, will similarly build up until they overwhelm the desired signals.
Since noise is far less obtrusive than backlash, electric analog computers
can be more complicated than their mechanical counterparts, but still
random
there
is
electric disturbances, the noise,
The
a limit.
digital computer,
the hazards of backlash and noise.
of
on the other hand, is entirely free

There is no intrinsic limit to the
complexity of the problem that a digital machine can handle.
The
third important difference between analog and digital computers
The precision of the analog computer

with
which physical quantities can be
the
accuracy
by
handled and measured and by the accuracy of representing a physical
is
in their potential accuracy.
is
restricted
system by idealized laws. In practice, the best such a machine can

is an accuracy of about 1 part in 10,000; many give results
achieve
accurate to only
or 2 parts in 100.
For some applications
this
range
On
the other hand, a

digital computer, which deals only with numbers, can be as precise
To increase accuracy we need only increase the
as we wish to make it.
adequate; for others
of precision is
it is not.
number
of significant figures carried by the machine to represent each

quantity being handled. Of course, the over-all accuracy is still limited
by the accuracy of the input and output transducers; but this does not
alter the fact that where high precision is required, a digital computer is
preferable to the analog computer.

There is a fourth aspect in which the
An
what is
two types
of
computers
differ.
That is, it continua

offers
solution
of
the
it
is
and
this solution is
ously
problem
solving,
appropriate at every instant to all the inputs which have so far entered
the machine. On the other hand, a digital machine works by formulating and solving an explicit logical model of the computing problem.
Therefore a digital machine gives only a sequence of values of the output
analog machine works
in
called real time.
at discrete time instants, and, furthermore, because of the finite, although
computing time, the output lags behind the input. Two problems
then arise: the problem of interpolation of the output between the discrete time instants, and the problem of prediction of the
output from
short,
fast values to
remove the time lag
of the output.
Obviously, if the
very much shorter than the characteristic time of the
controlled system, the prediction question can be ignored, and the computer considered to be working in real time. The present-day electronic
computing time
is
digital computer seems to be fast enough in this respect for the guidance
problems of the long-range rocket discussed previously; but for a high-
speed guided missile, this time-lag effect must be properly included in

the design of the control system.

Appendix
195
Chapter 13
to
CALCULATION OF PERTURBATION COEFFICIENTS

The
eters
quantities F, G,
S, A,
(13.13),
A,
and
are denned by Eq. (13.14).

They contain the paramAccording to their definition, as given by Eqs. (13.2) and
and N.
they can be written as follows:
^ ~
A-^gpAC L
'
(13.45)
= \~
N
where the aerodynamic
coefficients CL, CD.
the elevator angle 7, the
dj
and
CM are functions of the
Mach number M, and
the Reynol ds
angle of attack
number Re.
These
aerodynamic parameters are related to quantities immediately connected with the

flight path as follows:
jS
tan-i
= -|r
Re -
'a(r)
where
(13.46)
sound velocity in the atmosphere, and /*(r) is the coefficient of viscosboth functions of altitude r. In the following calculation, the thrust S will
a(r) is the
ity of air,
be considered to be a function of altitude only. It is also assumed that the composition of atmosphere at different altitudes remains the same as that of the standard
atmosphere only the density p and the temperature T changes.

of a and M at any altitude are variations due to temperature T.
For IB:
Thus the variations
=
~
dr
AH
W dr
(13.47)
dW
other partial derivatives are zero.
For A:
aA
_ AA
"
dr
fldp
IP
^V
CL aRej
(M
\CL
UAX
aA
ReaCA
1
dw[/M dCL + Re dCL +
+ 7Mr
C L ^Re
LVC'Lalf
Re dC L
aRe
dC L
dM ^ CL
*\
dC L lda
\
+ + C L a<7^
da ve + w)
1
^r
CL da
(13.48)
a7~"
3A
ReaC L \
'
a.A
__
aA
dw
(M
dCi,
TL^_
JL
If 2T
\C~L 'dM"2f
dT
_
=A
ve -
jM ac/L
~
.
"^
+
Re aCL 1 a/z
CLaReâT,
Jute ac/L
!<
-L.
]
-J-
_..
J_
J_
,.
iac/L
____
^r
\
I
aA
__
196
For
A, the partial derivatives are
and CD
For N:
for A,
ReaCjiA
jldp (
\plr V
dr
NF
dvr
72
CMaleV
\CM dM
obtained from the above equations by substituting
for CL.
CM aRe
\CM dM
far/M.agjf
V*dr [\C M
+2
CM
CM aRe
Re dC M
dM ^ CM aRe
da
vr
CM
da v&
-^
wj
CM
(13.49)
^L -
AT-
ap
I (}
\
dT
dJV
_ dN
dw
dve
Re aCaA
CM aRe/
\C M dM 2T
CM aRe M^^/
dN
az^-T
With these partial
derivatives, the coefficient a's,
op
dF
_
(13.50)
aA
dp
dp
aA
)_
and c's can be easily calculated:
^-
aA
ay
r-
6's,
aA
_
^_^_^5
1
~~
~~
a^
/?
i^
dr
dr
^
~~
__
aA
sm
__
dr
VT
a/3
a/3
dG
a/3
aA
aA
= dG
dy
dG
dr
dr
aA
__^r
dy
aA
as
a^
(13.51)
dy
A .
a^
a^
aG
__
__
= dG = -y aA
197
(ve 4-
w)
aA
a^
aA
^âFâF 008 ^"^^_,
ar
'
-
= ^J =
/3
(^0
aA
vr
_M-ir_ ^
Cfi
-s
dp
ajff
C7
cs
Cg
ClQ
=:
--
- r
r L
i r
i
</l
ajy
= -if
â^
-~
Vr
oJL
|_
vr
r L
dH
IdW
dl
A
/'
-f-
4.
.,
(^
\r
COS P
-Vr
aA
3A i
^TJ
aAl
-j
op
aAl
oJ.
(ve 4-
dN
-f-
w)
Ojt
aAl
aiyj
(13.52)
^T
a^
apj
.
rt l>r
1)
tf;)
+w
aA
aw
as
(VQ
(z;^
-^ _ -
(VQ
I
1A
ap
aA
dW
aff
aA
__
a^r
a7
|_
aA
_|- ^y)
aA
dH
__
#r
a^r
rtTTir
= aJ.1 = 1
T
dw
^-
aA
ay
5A
z; r
-Ssin^-z;r
.
C4
cos
a^v
+ -r~
aw
aA
dW
_ dN
~
a/
After the power cutoff, the thrust

tives are zero.
vanishes.
Thus
for
>
?i,
S and
its
deriva-
CHAPTER
14

In the earlier chapters, we discussed the design problem of control
systems mainly from the point of view of analysis. That is, assume
the construction of the system, and then find out whether the performance
In the last chapter, we introduced for the
of the system is satisfactory.
first time a different and more direct point of view: we specify the per-
formance
first,
and then
will give the desired
system which
In this chapter, we shall extend this
find out the necessary control
performance.
which the performance

terms of integrals of the controlled variables.
The resultant system behavior is represented by a very general equation,
which is usually nonlinear. Thus the control system so designed is
principle to arbitrarily controlled systems, for
criteria are expressed in
The nonlinearity here, however, is purgenerally a nonlinear system.

posefully chosen to give the optimum performance of the over-all system.
Mathematically,
we can
describe the principle of control design with

In the system to be controlled, we intro-
specified criteria as follows.
duce one or more extra variables. These extra variables, being artiare not determined by the intrinsic physical laws of
ficially created,
We
the controlled system.

extra variables
formance.
by
These
obtain the conditions for determining the
satisfying the specified criteria of the over-all perconditions are then enforced through the computer
This principle of control design was
built into the system.
The
by Boksenbom and Hood.
cited work of these authors.
1
14.1
first
suggested
following discussion follows partly the
Control Criteria.
If y is the output of the controlled

system,
reasonable to expect that the measure of over-all performance
of the system is expressed as the time
integral of some function / of y.
then
it is
Then the
mum
criterion of the
performance
or a constant; that
is
that this integral
is
to be mini-
is,
I
L f(y)
dt
const, or min.
dt
(14.1)
or, specifically,
l
JQ
1
A.
S.
(y
2/*)
Boksenbom and R. Hood,
const, or min.
NACA TR
198
1068 (1952).
(14.2)

where
199
the time at the end of transient and y, is the

setting or the
y.
Equation (14.2) weights the error in y as the square
and according to the time duration of that error, and is thus a measure
ti is
desired value of
mean-square error from the setting. Another type of criterion

be that which requires a criterion time duration to be a minimum
of the
may
or a constant; that
is,
l
The use
stant.
dt
const, or min.
(14.3)
such as Eq. (14.1), will yield f(y) = conis reasonable because f(y) can
usually be made to be a
no additional criteria are imposed on other variables in the
of a single criterion,
This result
constant
if
system. Usually, however, certain limiting conditions exist on other

variables in the system, and these conditions must be included in the
original criterion.
Thus, for example, a possible criterion could be writ-
ten as follows:
(y
dt
min.
f(z) dt
const.
~~
2/s)
for
If,
for instance, y is the engine speed
and
z is
the characteristic tempera-
ture of a gas turbine engine, the criterion of (14.4) states that
it is
desir-
able to design a control system such that, for a particular value of a

temperature integral, the integral of the speed-error squared is a mini-
This criterion may be used if, for instance, it is known that an

overtemperature condition can be tolerated for a certain period of time
and it is desired to keep the average speed error at a minimum during
mum.
Then
this transient.
the integral of
represents the total heat input to
the turbine blades.
The general theory will show that as many criteria as desired of the
type shown in Eqs. (14.1) to (14.4) can be included together, and a control system can be derived that automatically satisfies all these criteria
simultaneously.
Another aspect
of the control criteria is the
end conditions
of
the
The time interval for which these

a constant must be chosen. A reason-
integrals of Eqs. (1,4.1) to (14.4).

integrals are to
be a minimum or
is any duration during which essential external disturbances are constant and during which the system to be controlled
able time interval
The essential
essential level of operation to another.
external disturbances are those that cannot be immediately corrected
moves from one
by the control system. If an essential external disturbance occurs in the

time interval of the criteria, no physically realizable system could be
designed to anticipate this disturbance so a's to behave properly before
the disturbance takes place. An essential level of operation is any
200
specific
must be continuous.
condition of only those variables that
In the case of a turbojet engine, the transient behavior of which can be

described by a first-order differential equation, the engine speed determines the level of operation. If the fuel system must be considered
or if the temperature does not respond to speed immediately, then both
the engine and the acceleration are required to describe the essential
We shall see this presently.
level of the engine.
The control system resulting from any design method must be phys-
There are two aspects to
ically realizable.
possible to set
down
or are incompatible with each other.
unrealizability will
this
problem.
First, it is
criteria that are not realizable with any system

If
such
criteria
appear either as a requirement
are used, the
on the control to
look ahead into the future or as an inability to satisfy the boundary

In most cases, a clear underconditions of some differential equation.
standing of the criteria used and of the system to be controlled will

preclude incompatibilities of this sort.
The second aspect of physical readability
is
purely mathematical.
It is desired to derive a description (a differential equation) of the control
or the controlled system that satisfies the criteria of control
and
all
the
necessary boundary conditions that arise in the derivation of this equation.

Although the mathematical solution of the problem may be any
derivative or integral of this differential equation, the physical solution
problem requires the differential equation that itself satisfies the

boundary conditions and for which no undetermined constants of integraof the
tion exist.
Thus, such forms as

y
ex
ex
and
are not necessarily interchangeable as descriptions of

controlled system, because the forms differ
stant of integration.
some part of a
by an undetermined con-
For stable linear systems, the
effect of this
constant
becomes vanishingly small; for the general nonlinear systems presented

here, however, this constant must be considered.
The requirement of stability is a special

not enter into the design of the main control system
during the transient. This situation is the same as that of the last
chapter where the stability criterion is also suppressed, because the
14.2
Stability
Problem.
criterion that does
satisfactory performance of the over-all system
by the imposed performance specifications.

necessary to add to the controlled system a
not go into action until the
final instant of
fore, this stability device will
is
already
However,
made
it
certain
is
usually
stability device that does
the transient interval.
There-
not affect the behavior of the system as

far as satisfying the other criteria
then
This device can be
concerned.
is
described as follows: For a first-order system,
201
when
(14.5)
For second-order systems, when
then
and
(14.6)
When
such a stability device is added to the control system, the consystem has two modes of operation and is thus a multiple-mode
During the transient, the main control system
system (of. Sec. 10.9).
trol
At the end
in operation to produce the specified performance.
is
of
switched to a second system, represented by

or
Eq. (14.6), to ensure stability of the system at the end
Eq. (14.5)
avoid running away from the desired operating point of
thus
to
and
state
transient, the control
is
the system.
General Theory for First-order Systems.
14.3
control criteria
given previously by
late the control equation in the following

criteria,
one
of
the integrals
is
Xi
manner:
for such a list of
if,
minimum under
to be a
the other integrals are to be constant,

tional calculus, to
With the type of

we can formu-
Eq. (14.1) to (14.4),
it is sufficient,
the condition that
according to varia-
make
2
I
(y
ys )
dt
X2
fo(z) dt
+ Xa
dt
min.
or
'"'
t/(2/)
The X
+ Xifo ~ y? +
X/o(*)
XJ
dt
min.
(14.7)
are arbitrary constants that enter into the control system as the
adjustable parameters and are precisely determined by the choice of

values that the constant integrals are to have. The technique of the X
multipliers
is
to be a
is
widely used for problems of this type, where one condition
minimum under
other restrictive conditions.
Indeed, the con-
need not be in integral form, and any functional or differential

1
relation among variables can be handled in a similar manner.
Equaditions
made very
general when all possible restrictive condiIn the final equations, if any one criterion is not to
be used, then the corresponding X > 0. If any of the criteria are to be
> <*>
zero, then the corresponding X
tion (14.7) can be
tions are included.
See for instance C. Lanczos, "The Variational Principles of Mechanics/' UniverToronto Press, Toronto, 1946.
sity of
202
If the system to be controlled is of first order with constant

and with one essential output y, then the variables y and
related so that z
coefficients
z
must be
Equation (14.7) can then be written, in
z(y,y).
general, as
^ F(y,y)
where
is
dt
min.
(14.8)
a continuous function of y and y, and yisa, continuous funct

We note that F is not explicitly dependent on the time t.
tion of time
Let us consider
to be a solution, that
y(t)
is,
y(f) is
the output
among
To
admissible outputs which satisfies the condition of Eq. (14.8).
all
FIG. 14.1
test this,
we
construct a neighboring solution as
an arbitrary function, and

(14.8) is satisfied
Eq.
by
y(t),
y(t)
a small parameter.
then
e is
+
If
8y(t).
dy(t) is
the condition of
or
By
The
variation
$ti
(14.9)
dy
occurs because of the fact that the end point of the

is not fixed, but lies on the curve
y = /(*), as shown
integral of Eq.- (14.8)

in Fig. 14.1.
from one
By
This
is
to allow the proper
partial integration, Eq. (14.9)
The
boundary conditions
of
moving
essential level of operation to another, as

previously discussed.
relationship between $y(ti}
end condition.
That
is,
%+
becomes
and
can be easily computed from the

Then by eliminating
8y(ti),
and
since
dt is
arbitrary,
we have
}\dydt-0
~r[
203
(14.10)
and
interval during which the criterion of Eq. (14.8) is to hold

considered as that during which the system moves from one essential
operating level to another; in this case, from one definite value of y to
The time
is
another definite value of

straight line
with
f(t)
Thus the end curve y =

Hence
y.
f(t)
must be a
constant.
=
/'&) =
dy(0)
Thus Eqs.
(14.10)
and
(14.11)
become
2E.
A (^l.
and
if
(|is
finite
(14.14)
= because 5y(0) = 0, The only

Equation (14.13) need not hold at t
are that (dF/dy)o is finite and y is
conditions that need hold at t =
continuous.
At the
start of a
new
transient, y, F, (BF/dy),
may be discontinuous, whereas at other points

be continuous because of Eq. (14.13).
Equation (14.13)
is
(0
<
<
and (dF/dy)
ti),
dF/dy
will
the differential equation for the y(t) that satisfies

This equation is the so-called Euler-
the original criterion of Eq. (14.8).
Lagrange differential equation of our variational problem. For the problem considered here, F does not explicitly depend upon t. Then we can
immediately obtain a first integral of this equation. The first integral

of Eq. (14.13), which satisfies the boundary condition of Eq. (14.14), is
y-.
By
differentiating this equation with respect to
dF
Ty^
or
(14.15)
whenever
y,
..dF
^Ty
..
dF
Ty
+y
we have
d idF
dt(
dF/dy, and so forth are continuous
204
= or Eq. (14.13) is
y
vanish
generally
during the transient.
Thus
either
as specified
by Eqs.
(14.13)
satisfied.
However, y does not
Thus the two conditions on
and (14.14) are now replaced by the
y(t]
single
equation (14.15).
Thus Eq. (14.15) is the description of that physically realizable system

the behavior of which will automatically and simultaneously satisfy those
criteria included in the function F during that time interval for which
the external disturbances are constant and during which the system
goes from one operating level to any other operating level. At the end
of the transient
when the end
level of y is reached, a stability device
to the system; the description of such an ideal device
must be added
is
that of Eq. (14.5).

14.4 Application to Turbojet Controls.
In the usual case of designing

that sets the essential operthe
turbojet engine controls,
engine speed
or
As the result, other
set
controlled.
level
of
is
to
be
the
engine
ating
Limiting condipertinent characteristics, such as thrust, are also set.

tions of the engine are those of overspeed, overtemperature, compressor
Let
s be the engine speed setting, T
temperature to the turbine, and P be the discharge pressure
of the compressor; then the criteria on the behavior of this engine can be
surge,
and rich burner blowout.
be the
inlet
specified as the following integrals:

1
fi(N
s)
dt
for speed control
fi
fa(N) dt
for speed overshoot
fz(T) dt
for
temperature overshoot and
undershoot
l
fi[P
g(N)] dt
f*f*[P-h(N)]dt
(14.16)
for compressor surge

for
blowout
'
and
dt
The nature
g(N)
is
for rise
time
sketched in Fig. 14.2. The quantity

amount by which the compressor discharge pressure
of these functions is
the
exceeds the safe pressure for surge, and g(N)
is
the compressor discharge
pressure for each engine speed at a safe value below surge.

Rich burner
blowout can be handled in a similar manner. The rise time is the total
time for the system to move from one essential operating level to the other.
Similarly to the treatment on turbojet behavior in Sec. 5.6, the linearized engine characteristics
can be expressed as follows:
T = aN
+ arN

where r
is
the engine time constant.
we
into the integrals of Eq. (14.16),
where /
is
function of
By
substituting these relations
see that they all take the
a continuous function of
205
form
"
and N, and A
is
a continuous
t.
P-g(N)
k(N)
P-h(N)
FIG. 14.2
14.5
Speed Control with Temperature-limiting
error in speed control
Then
is
Criteria.
If
only the
considered important, the criterion becomes
the control condition of Eq. (14.15) simply gives
fr(N
Then because of the nature

means that, in the absence
this speed control should
s)
of the function
of other criteria
/i,
9.
This result
on the engine behavior,

zero, which is
keep the speed error identically
206
physically possible only in the sense of allowing infinite temperatures.

This result, however, is inconsistent with the previous development of
This
is not a discontinuous function of time.
Eq. (14.15), in that
instance
is
actually a trivial case of the general problem.
does indicate that a
But the
result
must be accompanied by an
this
criterion like
additional criterion to give a physically sensible system.

Now if the error in speed control is to be combined with the condition
on the overshoot and undershoot

1
[f^N
Therefore^
fr(N
N,)
N + X/
8)
temperature, then
of
\MT)]
By
(!T).
dt
min.
(14.18)
using Eq. (14.17), Eq. (14.15)
becomes
-N +
fr(N
s)
Xfa(T)
\arNfi(T)
(14.19)
the control equation during the transient. At the end of the

transient, the ideal stability device is switched on, so that when
This
is
N
N
then
=
=
(14.20)
Equations (14.19) and (14.20) describe the complete control system.

we can visualize a computer so designed that it takes informa-
Therefore
tion
from the measurements on
and
T,
and the
N and
T, the stored information on A, a,
between the fuel rate and
relation
N and
T,
and gener-
ates the signal for the proper fuel rate in accordance with Eq. (14.19).
the stability device takes over, so

Then shortly before
reaches
S}
that at the end of the transient, Eq. (14.20) is satisfied. In general, the
control equation (14.19) is nonlinear, and the computer cannot be a
linear device, such as a simple
As an example
and fz(T) = (Li
> 1,
because
T)
when n
RC
circuit.
<
for T > L 2
T < LI. In general, the power n should be
T may be infinite and of such nature as to make
convenient to let fz(T)
it is
(T
L 2) n
for
1,
N discontinuous and physically unreal,
even though the integral
In the example of this discussion, let n = 2, and, furthermore,

2
= (N
Therefore we again take the mean-square
s)
fi(N
)
deviation from the setting as a measure of the error.
Then Eq. (14.19)
is finite.
becomes
AMs} 2
(TV"
where, for acceleration, or
when
N>
(L
aAO
aVW
(14.21)
N<N
s,
and
I =L
(14.22)

For deceleration, or
N>N
8)
N<
The block diagram
is
207
L =
and
for the control
system
LI
is
(14.23)
shown
in Fig. 14.3.
the actual engine speed. We assume the case of deceleration with

s is positive, the switch between
> s During the transient, e
the computer and the engine servomechanism is closed, and the signal
from the computer is in action. The computer generates the signal
arN according to the control Eq. (14.21). In Fig. 14.3, the signal is
schematically indicated by a rectangular triangle. The engine servomechanism is so designed that the signal cnN from the computer is
actually obeyed closely
by the
engine.
This
is
done by using a high-
FIG. 14.3
When the speed error (N e gain circuit as indicated.

to a very small value, the computer signal is switched
stability of the engine
servomechanism
the system at the engine setting

condition of Eq. (14.20).
The
N,
will
JV.) is
off.
reduced
Then the
guarantee the stability of
and the system essentially satisfies the

For any value
control system has one adjustable parameter X.
of X, this system will, for the value of integral temperature overshoot

error squared.
obtained, give the minimum value of the integral speed
The value
overshoot.
of
X determines the actual value
of the integral
Let us consider the special case
where
temperature
aN =
s
L',
that
is,
acceleration or deceleration to the speed that corresponds to the limiting

to note
temperatures occurs, according to Eq. (14.17). It is interesting
= 8 according to the control

= when
that for this special case,
condition of Eq. (14.21). Therefore no separate stability device is necesand the switch in the control system as shown by Fig. 14.3 can be
sary,
208
eliminated.-
In this special case, Eq. (14.21) becomes linear and can be
written as
E(L
aN)
= arN
(14.24)
where
Now the integrals

integral
For instance, the temperature
can be easily computed.
is
A!
'*,
(T
- LY dt =
-L + arN)
(aN
dt
Jo
= (E-
I)
\L-aN)*dt
Jo
'h
(-
I)
a 2 (S-
I)
a2
rNs
2
//AT
(NS
JN*
-N)*^
N
Thus
r Jo
(L
aJV Q ) 2
dt
"^1
where NQ is the engine speed at the beginning of the transient.

the speed integral is
a p / AT _
1
(f -
(14.26)
Similarly,
r J
and
if
T^x
is
the
maximum
(14.24),
=
$E
(W 27)
'
temperature, then
fev^
From Eq.
'
\L
= % "
(14.28)
we have
Ea s
Ea(N
-N) = ar~
dt
so that the characteristic time r* for the controlled transient is
(14.29)
The left sides of these equations have been put in dimensionless form.
The maximum temperature rmax occurs at the beginning of the transient.
209
E = 1 (X = co), the temperature does not overshoot, in agreeour previous statement that the constant integral has the
with
ment
value zero when X-* <*>. The speed integral is 0.5, and T* = r. As E
increases (or X decreases), the temperature integral and the maximum
For
temperature increase, whereas the speed integral and the time constant
A compromise value for E may be \/2, or a 2 X = 1. Once
decrease.
the choice of
or X
is
settled, the specification for
The design can then
is fixed.
the control computer
proceed.
For the general case of Eq. (14.21), the calculation of the values of the
integrals is somewhat more cumbersome, but a similar procedure applies
Boksenbom and Hood
for the design of the control system.
give solutions of Eq. (14.26), that
is,
as a function of time
actually
But
t.
we should emphasize here that such explicit solutions are not necessary
The information for control design is fully set
for the control design.
forth by Eq. (14.24) itself, because that equation tells how the control
computer should be constructed. If the control computer is made
according to that condition, then the desired performance is ensured.
with respect to time is thus of no importance.
The actual variation of
the
to
Hence our approach
design problem is to "design the nonlinear
3 '
control equation itself rather than to design according to the solutions

of
an assumed equation.
Second-order Systems with
14.6
Two Degrees
of
Freedom.
For the
case of a second-order system with two degrees of freedom and with
constant coefficients, Eq. (14.8) becomes

'1
/
where y and
The
z are the
F(y,y y z
)
F(y
min.
(14.30)
outputs and are independent functions of time.
condition to satisfy Eq. (14.30)
rf
/
a*
Jo
z,z) dt
is
+
z
e 8z,z
6z,z
e 8z)
dt
at
The time
(t
=
=
(1431)
*
interval of the integral of Eq. (14.30) begins at a definite time
= 0) but does not end at any definite time, but rather along the curves
= fi(f) y = fa(t) % = 9i(f)) an-d ^ = ^(Q. The functions by and dz
y
}
are arbitrary and, naturally, independent functions of time.
Performing the operation indicated by Eq. (14.31) gives
}
210
After integration
by
__
parts,
dF
-|
we have
\y f
d idF\
Jt
~dz
VaJ J
d?_
dfi
OF
As
=0
&
^-Ûr
(14.32)
and the boundary-condition

the specified end condition, we
before, the integrands of the integrals
terms must vanish separately.

have
From
(14.33)
The
three conditions from Eq. (14.32) are the
two simultaneous Euler-
Lagrange equations
^1W +^!Yî
dy
aF__rf
a^
"-'
"
a -'
(14.34)
^^
*
\dz /
and
.
Equation (14.34)
criterion of
is
dF
dF\
a system, of two equations that
Eq. (14.30).
The physical
&F
satisfies
the original
solution to the problem
must
Eq. (14.34) and, in addition, satisfy the boundary-condition equations of Eq. (14.35).
However since F does not explicitly depend upon t,
satisfy
a modification of this set of conditions

(14.34) is multiplied
is,
possible: If the first of Eqs.
by y and the second by
added, an exact derivative
is
formed the integral
and the equations are

of which is as follows:

.dF
d (dF\
-y dy
.dF
..dF
y
~'
'"
SdF\
+,.d
..
BF
" c_
211
, oa
(14 36)
.
'
F is a function of # and y, it is unlikely that dF/dy and dF/d$ can

be zero, or that dF/dy can be a constant with respect to time. Then
Since
idF\
T^
dt\dy/
dF
---d
T: I
dy
zer
=
in
Eq. (14.35).
_,
,
inerelore
1S
50(0)
it
is
=0
argument applies to the variable
similar
reasonable set of boundary conditions
Sz(G)
reasonable to have
=
=
=
=
is
z.
Thus a
as follows:
f((ti)
ft(ti)
g((ti]
0'2 (*i)
=
=
=
=
(14.37)
These boundary conditions are also the ones corresponding to fixed

starting values and fixed end values of y, y, z, and 2, but with a variable
With the conditions of Eq. (14.37), Eq.

ti for the transient.
shows
that
the
constant
C to the right in Eq. (14.36) must vanish.
(14.35)
The final solution to the problem of Eq. (14.30) is as follows:
interval
and
dF
+
.
--'
*-'d
..
...
00
(14 38)
'
either one of the following equations:
dF
.
dt\di,
_
^dt*\dz'
JL.
dtVdzJ
(14.39)
Equations (14.38) and (14.39) constitute a system of two equations for

two unknowns y and z. They are the control equations and are the
equations for the computer design.
The boundary condition (14.37) defines the original criteria for that
duration during which the system moves from one essential operating
Thus, if all conditions of Eq. (14.37) must hold, the

system goes from one definite y, y, z, and z to any other definite y, y z
and z. Equation (14.38) is a third-order equation. Equation (14.39)
level to another.
a fourth-order equation. Thus besides the four initial values

we can still assign three values of y z z at the final y. That
is
=
2/5;'
0, z
z S)
and
0.
z} z
when
A stability device must still be added
to the system, so that
y,
is,
and
212
We
at the final point.
above
is
cussed in
see then that although the
system considered
considerably more involved than the first-order system disthe previous sections, yet the same general approach is entirely
applicable.
Control Problem with. Differential Equation as Auxiliary CondiLet us consider y to be the essential variable whose performance
has to be controlled, z is the variable which we put in the system to
The
ensure that y can be made to have the desired performance.
14.7
tion.
inherent dynamics of the system then gives one a relation between y

and 2. This relation is in general a nonlinear differential equation, say
of
second order,
The performance
of
y during a transient
dt
is specified,
say, as
min.
(14.41)
would be an example of this type of

is
to
derive a control equation satisfying
The
problem
specification.
both Eqs. (14.40) and (14.41).
The mathematical problem is then a problem of the calculus of varia-
The
error integral of Eq. (14.2),
tions with the differential Eq. (14.40) as an auxiliary condition.
This
can again be solved by using the method of the Euler-Lagrange multiplier

1
XOO.
That
is,
1
/J
F(yd,y ?,*,&$
^=
min.
(14.42)
with
F *
The only novel
f(y)
feature of the problem
is
the introduction of the time-
varying multiplier \(t). The problem of Eq. (14.42) is exactly the same
as that of Eq. (14.30). Therefore all the equations developed in the
last section can be used.
However, we now have three unknowns: y, z,
and
X.
The
three equations are Eqs. (14.34) and (14.40) with
by Eq. (14.32). Equation

and is thus automatically
has to be
artificially
F defined
(14.40) is intrinsic to the physical

satisfied
enforced
is
by the
Eq. (14.34).
controlled system.
That system
of
system
What
two equa-
tions then forms the basis of the operation of the control computer.
properly constructed control computer then takes information about
the essential output y digests it, and then generates a continuous signal
}
z, when fed into the controlled system, then forces

the system to behave according to the specification of Eq. (14.41).
for
z.
This signal of
See for instance 0. Bolza, "Vorlesungen tiber Variationsrechnung," Chap.

11,
Teubner, 1909.

14.8
and
Comparison
of
in this chapter,
213
In the last chapter

discussed methods of designing control
Concepts of Control Design.
we have
systems when the performance is specified quite rigidly. The method

of the last chapter, based upon the perturbation theoty, is applicable to
The method of this chapeven more general in that the system to be controlled may itself be
For such general systems, these newer methods are the only
nonlinear.
available tools for designing the control; and the resultant complicated
linear systems with time- varying coefficients.
ter is
control systems involving electromechanical computers seem to be the

only logical solution. However, the methods of the last two chapters are
equally applicable to the simpler physical systems treated in the earlier
For such simpler
chapters, i.e., linear systems with constant coefficients.
systems, then,
we have two
different general approaches to the control
compare these two design concepts.

engine-control problem was treated in Chap. 5, using the conventional principles of servomechanism. In this chapter, we have discussed almost the same problem with the new design method with
problem.
It is illuminating to
The
One point is immediately clear: the

the
older
method is linear, and the control
by
a
RC
can
be
simple
circuit; while the control system designed
component
is
new
method
nonlinear, and the essential control unit is a
by the
specified
performance
criteria.
control system designed
computer which even in its simplest form

than an RC circuit. But this complication
gain: while the control system based
tional
servomechanism
may
is
is
upon the
much more complicated

not introduced without a
principle of the conven-
be entirely satisfactory in performance, the
control system involving a nonlinear computer is guaranteed to give the
optimum performance no other control system can be better under

the same design specification. But this comparison has meaning only
if we know specifically the desired performance.
For instance, if we
do not know exactly what the temperature integral in Eq. (14.18) is, we
cannot apply the method discussed in this chapter at all. On the other
hand, to design a satisfactory control system according to the older concepts of servomechanisms requires no such sharply defined specification.
Of course, strict insistence on the best performance must come after

a clear understanding of what constitutes optimum control behavior.
Therefore, when we do want an optimum control system, we naturally
should have the information to define sharply the design criteria. From
this point of view, then, the newer principles of the last chapters for
control systems with specified performance certainly go one step beyond
the conventional theory of servomechanisms and are principles for more
advanced control design.
That the more advanced control system
should also be more complicated
is
to be expected.
CHAPTER
15
In the previous chapters we have discussed the design principles of
control systems with increasing degrees of generality and complexity.
However, one basic assumption was made throughout the treatment: the
properties and characteristics of the system to be controlled were always
assumed to be known. In the case of conventional linear servomecha-
nisms, the transfer functions of the servos and other components are
In the case of linear systems with timespecified before the design.
varying coefficients, we take the example of the guidance system for the
long-range rocket vehicle. There, the dynamic and the aerodynamic
properties of the rocket were determined previous to the design of the
In the case of the general system control according to
specified criteria treated in the last chapter, the response of the system to
The control
variations in the controlled input is again predetermined.
control system.
is
design
thus based upon this knowledge of the properties of the system.
The feedback merely conveys the information on the state of the output
The computer then uses its built-in knowledge of
to the "computer."
37
the system properties to generate the "intelligent control signal.

In this chapter, we wish to relax even this seemingly elementary
requirement for control design. We wish to introduce the principle of
continuously sensing and continuously measuring control systems where no
exact knowledge of the properties of the controlled system is necessary
for the design.
Instead, the properties of the controlled system are
measured during the control process.
In particular, we shall discuss a
simple case of such control systems, the optimalizing control

15.1 Basic Concept.
No matter what degree of accuracy we can
obtain from the control computer, the accuracy of the controlled behavior
of
a system
before
dependent upon the accuracy
is still
in its design.
we know
of
the information used
we determine
the properties of the controlled system

the over-all design of the control system, as is tacitly
If
assumed
in the previous chapters, then extreme accuracy of the controlled

behavior cannot be expected for two reasons: First, the manufacturing
process always introduces small differences into supposedly identical
objects.
wing
For instance, the wing of a rocket cannot be identical to the

model on which the wind-tunnel tests are made to
of the rocket
214
determine the aerodynamic properties.
erties of the rocket in reality
must
215
Therefore the aerodynamic propfrom the test results.
differ slightly
Secondly, any engineering system is subject to small variations with

This may be due to the normal deterioration of the
respect to time.
system caused by wear and fatigue, or it may be due to the drift of

In short,
conditions in the environment in which the system operates.
the properties of an engineering system can never be known exactly
Therefore, when
prior to the instant of actual operation of the system.
is
a
behavior
the
of
controlled
necessity, we must use the
great accuracy
principle of a continuously sensing control system.
Nor is the required accuracy of control the only reason for changing
our concept of control; very often the fact that large unpredictable
may occur forces us to use the con-
variations of the system properties
tinuously sensing system.
We
have already introduced
this principle
in connection with the disturbance effects of atmospheric changes in the
guidance problem of the long-range rocket. There we used the dynamic

behavior of the rocket itself as an instrument for continuous measurement
But a more illuminating example is the flight of an

The deposition of ice on the
airplane through icing weather conditions.
surfaces of the wing and the fuselage changes the shape of these airplane
components. Moreover, the exact manner of deposition is somewhat
of these effects.
variable and cannot be predicted with accuracy.
Therefore the aero-
the airplane can be profoundly altered, and altered

dynamic
Even worse, the changes all tend to
in an unpredictable way, by ice.
of
the
the
airplane,
i.e., decrease the number of
degrade
performance
properties of
We
are thus intermiles that can be flown with one gallon of gasoline.
ested in knowing the combination of engine throttle, engine rpm, and
airplane trim that will give the maximum miles per gallon of gasoline,
because
we should
fly
the airplane at the
But
the strained fuel load.
knowledge
of the airplane
deposition.
optimum condition
to conserve
just at this critical situation,
performance
is
our prior
rendered useless by the ice
Hence the only
solution to this problem of cruise control

an automatic sensing and measuring control
an optimalizing system which automatically holds the air-
in adverse circumstances
system, i.e.,
plane at the measured
Of course, a skilled
is
optimum operating
conditions.
human
operator controls the performance of a

machine on the optimalizing principle. He watches the instrument
readings of the inputs and outputs of the machine, and then uses his
knowledge and experience to decide in what directions the controls should
be adjusted. The adjusted inputs bring new output readings, which have
to be interpreted by the operator to determine whether the optimum
operating condition
controls will
is
reached or exceeded.
have to be made.
New
adjustments
The continuous adjustment
of
the
of inputs is
216
the sensing process, and the reading of the outputs
is
How-
the feedback.
slow in
ever, manually controlled optimalizing systems are necessarily
well
how
matter
no
human
for
skill,
complicated systems,
response, and,
Automatic optimalizing control was conis not sufficient.
developed,
1
Its application to
ceived by C. S. Draper, Y. T. Li, and H. Laning, Jr.
2
R.
Shull.
J.
discussed
was
of
an
control
cruise
by
airplane
15.2
The heart of an optimalizthe nonlinear component which characterizes the
Principles of Optimalizing Control.
ing control system
is
operating conditions. For simplicity of discussion, we shall

assume that this basic component has a single input and a single output.
optimum
For the time being, we
shall also neglect
any time
effects
and assume
that the output is determined only

by the instantaneous value of input.
is an optimum operating
the
output as a function of
point,
has
at y and x as
a
maximum
input
shown in Fig. 15.1. It is convenient
Since there
and the input to

optimum point, and thus put the
physical input as x + x and the
to refer the output
the
optimum
x
y*
point
optimalizing control is then to search out this
is
The
y
then the point
physical output as y*
The purpose of an
optimum point and to keep
0.
the system in the immediate neighborhood of this point. In this neighborhood, the relation between x and ?/*.can be represented as
y*
The
simplest
method
= -kx*
of obtaining
(15.1)
an optimalizing system,
in concept,
Let us assume that we start with a negative input, i.e., an

input smaller than the optimum input, and that we increase this input
at a constant time rate as shown in Fig. 15.20.
The corresponding output
is
as follows.
y* will first increase, then reach the optimum value, and start to decrease,
as shown in Fig. 15.26.
The time derivative of y*, dy*/dt, is then first
positive, decreases to zero at the point 1 (Fig. 15.2c), and becomes
negative after that point.
At the point
2,
the value of dy*/dt reaches the
magnitude designed into the control system so that the direction

variation of the input is reversed, and the input x now decreases with
critical
of
the same constant rate,

1
y*
now
increases again,
and dy*/dt jumps to
Y. T. Li, Instruments, 25, 72-77, 190-193, 228,

324-327, 350-352 (1952). C. S.
Draper and Y. T. Li, "Principles of Optimalizing Control Systems and an Application
to Internal Combustion Engine," ASME Publications
(1951).
2
J. R. Shull, Trans IRE (Electronic
Computers), December, 1952, pp. 47-51.
positive values.
At the point
3,
the output reaches
217
its
maximum value,
and dy*/dt becomes zero again. At the point 4, dy*/dt again reaches
the critical value, and the drive for the input variation is again reversed.
This process repeats itself, and the behavior of the system is periodic.
The system
is
is
optimum point. The period T*

The minimum value of the output is A* and is
said to hunt around the
the hunting period.
called the hunting zone of the output y*.
Because of the parabolic
(b)
Effective
Output
Input Drive Reversal Points
FIG. 15.2
relation of Eq. (15.1), the average value of the output with hunting is
fA* lower than the optimum output. This difference is a loss, the
hunting loss D*, and
is
the price to be paid for the control.
Thus
D* = A*
system can be calculated in terms of

using Eq. (15.1), the extreme values of the input are
Other characteristics
A* and T*: By
VA*A- The rate of

critical
if
we
(15.2)
of the
change of the input
is
thus 2 <\/A*/k/T*.
The
value of the time rate of output change

-4A*/r*. Therefore,
the hunting zone A* or the hunting loss D*, and the hunting
such an
T*, the system is specified. The essential elements of
is
fix
period
the input, the output

optimalizing system are the test variations of
the
and
detection and differentiating device,
switching of the input drive
The
at the
sensing and searching for the
magnitudes.
predetermined
218
optimum point
by the forced input
are accomplished
the constantly changing input also causes a small loss
make
It is desirable to
the hunting zone A* small.
But
variations.
of the output.
But small A*
also
This
of the critical dy*/dt for input drive reversal.
reduces the magnitude

then increases the danger of accidental input drive reversals caused by
It is apparent that
the unavoidable disturbances or noise in the system.
displaced from the optimum point, the recovery time is
short hunting period
directly proportional to the hunting period T*.
it
will be difficult to
made
too
T*
is
if
But
is thus desirable.
small,
if
the system
is
differentiate the output variation
random
We
disturbances.
from the hunting operation and other

next
shall discuss this point again in the
section.
The
testing input variation can also be a continuous function of time
For instance, we can make
instead of the saw-tooth curve of Fig. 15.2a.
the input # be a combination of a slowly varying part x a and a sinusoid

Thus
of constant amplitude a and frequency w.
Then, according to Eq.
x*
xa
(15.1), the
+ ~2
sin at
(15.3)
corresponding output y*
2kx a a sin ut
/f-
cos
is
(2orf)
(15.4)
This output signal can be fed to a band-pass filter to remove both the
first term and the double harmonic of the third term.
slowly varying
The
filtered
signal is thus
sinusoidal signal a sin ut are
~2kxaa
sin at.
Now
this signal
and the
combined in a rectifying multiplier which
multiplies these signals to give
-2kx aa*
sin
<d
= -kx aa,\l -
cos (2w$)]
(15.5)
and then we again remove the double harmonic. We have then, finally,
ko?x a
This signal can be used to vary the part x a of the
the signal
input, such that
a^=
fca
xa
UAr
Then xa tends
(15.6)
to zero with a decay time 2T*,
2T * -
&
<
15 7 )
-
Because of the parabolic relation between input and output, the decay
time constant for the output is T*. Therefore such a control system
will also search out the optimum
point and approach it asymptotically.
The operation
of this optimalizing control with continuous test

signal is
shown
in Fig. 15.3.
Fig.
shows the
15.3c
filtered
219
output
signal,
and
Fig. 15.3d indicates the effects of the rectifying multiplier.

When the system is operating near the optimum point, because of the
sinusoidal oscillation of the input, the output
is
ka 2 sin 2
ut.
Therefore
(a)
(b)
2kx a a
sin at,
Fundamental Component of y*
(c)
-2fcc a a sin tit

2
-~ka x a Output
,
of Rectifying Multiplier
FIG. 15.3
again an output loss D*

of the test input should be small.
there
is
foz /2.
But this
on incidental 'disturbances and noise

zone A* of the output is &a 2 and thus
tions
For low
is
losses,
again limited
of the system.
the amplitude
by consideraThe hunting
D* =
(15.8)
220
Equation (15.7) shows that the design constant a of the input drive
determined by the time constant f* and D* according to the relation
= 4D*f * = 2A*f*
The
rectified part of the signal given
from the
test input of
by Eq.
(15.3), is actually
Eq.
is
(15.9)
(15.5),
-ka 2 x a
obtained
a measure of the deviation
input from the optimum input. The continuous drive of the

input according to Eq. (15.6) is but one of the many possible uses of this
signal.
Evidently, this signal can also be used to give a saw-tooth variaof the
by using a constant rate variation of input with a superimposed

and by reversing the input drive at critical values
z
of the signal \ka x a
The hunting operation of this optimalizing control then consists of two separate frequencies, a low-frequency component
tion of x a
sinusoidal oscillation,
\.
in the variation of x a
and a high-frequency component produced by the
sinusoidal input oscillation.

15.3
Considerations on Interference Effects.
The previous
discus-
have shown the importance

the variation of test input and the time
sions on the idealized optimalizing controls

of reducing the
amplitude
of
Amplitude
of
Noise
Range
for
Optimalizing
Control
High Frequency Interference
Frequency
FIG. 15.4
constant T*. However the actual design is limited in this respect by the
In order to
ubiquitous noise and interference in the physical system.
measure effectively the output variation as a result of input variation
for testing the distance to the
optimum point, it is necessary for this

output variation with time to be made up of frequency components that
may with certainty be distinguished from output variations due to noise
and interference effects. The relative amplitudes of the interference
output frequency components can be plotted as a function of the frequency. This output interference spectrum generally has a low-fre-
quency part, the drift interferences, and a high-frequency part, as shown

Between these two parts, there is generally a range of
in Fig. 15.4.
frequency relatively free of the noise effects. If the optimalizing control

designed with test variations with frequencies in this range, the amplitude of these test variations can be made small without fear of
losing
is
them
in the
background
of interferences.
In general then, the test func-
tion
must be made up
tinguishable from
of
221
input variations that are fast enough to be disand at the same time slow enough
drift interference
to prevent confusion with high-frequency noise.

These considerations on the noise effects point out the difficulties
in
both types
The
of optimalizing controls discussed in the previous section.
type, with saw-tooth input drive, uses the time derivative

output as the control signal. If there is random interference in
first
of the
the output, then the relative amplitudes of the high-frequency components will be increased by taking the time derivative; thus the range
This is
of frequency available for optimalizing control will be reduced.
a serious disadvantage. The second type of optimalizing control, using

the continuous sinusoidal test function, requires a wide band of noise-free
frequencies, because in addition to the variation of the level of input, x a
,
there
the sinusoidal variation with
is
its
own
higher frequency
co.
Thus
the system to be controlled has only a narrow range of noise-free

frequencies, the two optimalizing controls so far discussed are not easily
A better system is the so-called peak-holding optimalizing
applicable.
if
control to
be discussed in the next section.
Peak-holding Optimalizing Control.
15.4
The input
variation for a
peak-holding optimalizing control is the same as for the first type of

optimalizing control studied here, a constant rate variation with periodic
The essential improvement here lies in the method of generatthe

drive-reversal
Reversal of the input drive should occur
signal.
ing
when the output has passed its maximum and has decreased to a value
reversals.
approaching the hunting zone
limit.
This fact
itself is
used as the con-
dition for producing the input drive reversal for the peak-holding opti-
malizing control.
The output y*
is
Essentially this could be accomplished as follows.

This voltage is applied to a
measured as a voltage.
condenser through a gate allowing only charging but not discharging.

of the condenser follows the output y* until the maxi-
Then the voltage
mum
mum
value
is
reached.
When
the input x
value, the output y* decreases.
is increased beyond the optiBut the condenser voltage will
its maximum value, and a voltage difference v exists between

This voltage
the condenser and the indicating voltage for the output.
At this point a
difference v builds up to the hunting zone limit A*.
remain at
relay
is
triggered to reverse the input drive, and at the
condenser
same time the
discharged to a voltage equal to the instantaneous y*.

Thus the operation of this optimalizing control can be represented by
is
Fig. 15.5.
The
relation
between the hunting zone A* and the hunting
loss
D*
the same as given by Eq. (15.2). The extreme values of the input
are still
-\A*A> and the input rate is again 2 \/A*/k/T*. It is
seen that the peak-holding optimalizing control has only one essential
is
222
frequency of output, determined by the hunting period T *, and no

It is thus particularly adapted' to a
differentiation of the output is used.
system with a narrow noise-free frequency range. In fact a further

improvement in this direction could be made by basing the input drive
reversal not directly
on
v,
the voltage difference between condenser and

Then
of v with respect to time.
output indicator, but the integral
Input Drive Reversal
D*, Hunting Loss
T*Hunting Period
(c)
Critical Voltage
t
Voltage Difference
Difference
Between Condenser
&
Output Indicator
FIG. 15.5
the high-frequency interference effects will be suppressed, and the hunting zone and hunting loss can be further reduced without accidental input
drive reversal.
15.5
Dynamic
Effects.
In the discussions of previous sections, we
have assumed that the input-output relation is specified by Eq. (15.1)

and is independent of the rate of change of the input or the higher time
This is true if the response of the output to input
instantaneous without the slightest time delay. In any
physical system, this is not possible; there are always the inertial and
other dynamic effects. We have then to consider the output y* given
derivatives of input.
variation
is
by Eq. (15.1) as the fictitious "potential output" but not the actual
output y measured by the output-indicating instrument, y* is equal to y
only when the time constant !F* of the optimalizing control approaches
223
The relation between y* and y is determined by the dynamic

But we have seen previously that such dynamic effects can be
If the optimalizing control is
closely approximated by a linear system.
to be applied to an internal-combustion engine, as was done by C. S.
Draper and Y. T. Li, the potential output is essentially the indicated
infinity.
effects.
mean
mean
effective pressure of the engine, while the actual
output
The dynamic
effective pressure of the engine.
is
the brake
effects here are
mainly due to the inertia of the piston, the crankshaft, and other moving
For small changes in the operating conditions of
parts of the engine.
the engine, such
dynamic
equation with constant
effects
can be represented as a linear differential

Since we have set the reference
coefficients.
input and output at the optimum input x and the optimum

y 0j and the physical
output y 0} the physical potential output is y*
the
relation
Thus
between
the physical
indicated output is y
y
level of
potential output
and the physical indicated output can be written
as an
operator equation
(V
+ Vc) =
F,
U] (y* + y
where F is generally a quotient of two polynomials in the operator d/dt.

In the language of Laplace transforms, F (s) is the transfer function.
Let us call the linear system which transforms the potential output to the
indicated output used for controlling the input variation, the output
Then F (s) is the transfer
linear group of the optimalizing control.
function of the output linear group. By implication, however, when the

= 0, the potential output is
dynamic effects are negligible, or when 5
indicated
Therefore
we
have the condition that
the
to
output.
equal
fo(0)
(15.10)
and the
That is,
Then the operator equation between the
potential output
cated output can be simplified because y
is
a constant.
indi-
In a similar manner, we can introduce a "potential input' x* which

is actually the forcing function generated by the optimalizing control
system but not the actual input x. The relation between x and x*
determined by the inertial and dynamic effects of the input drive

system. This input drive system we shall call the input linear group
And the operator equation between the
of the optimalizing control.
is
potential input x*
and the actual input x

*
is
M*
(15-12)
224
thus the transfer function of the input linear group.
Fi(s) is
toEq.
(15.10),
Similarly
we have
ft(0)
(15.13)
Thus the block diagram of the complete optimalizing control system

can be drawn as shown in Fig. 15.6. The nonlinear components of the
system are the optimalizing input drive and the controlled system itself.
The general relation between the input x and the output y is then
determined by the system of Eqs. (15.1), (15.11), and (15.12) and by the
For instance, if the
optimalizing input drive adopted.
of
the
drive
is
peak-holding type discussed in the last
optimalizing input
particular
Controlled
Input Linear Group
System
Output Linear Group
input
FIG. 15.6
section, then the potential input
and amplitude
a,
as
shown
a;* is
a saw-tooth curve with period
in Fig. 15.7a.
2T
Let
27T
_.
Then x* can be expanded
(15.14)
into a Fourier series,
8a
e
According to the general relation of Eq. (2.16), the actual input

by Eq. (15.12), can then be calculated as
(15.15)
a?,
given
I)
The
have
potential output y*
is
given by Eq. (15.1).
Using Eq. (15.16), we
*
VV
225
(-1)"+"
LI LI
(2n
2n
l)
(2m
_,
r~ to y
l)
/2m
T~
(n
m)i
(15.17)
By again applying Eq. (2.16),

VV
LI LI
+m+
have, finally, the indicated output
y,
(-1)"+"
(2n
l) (2wi
I)
IJtwoR.
FJ- (n - m)^
-
we
+m+
/2w+l too\
.
B,
-wo^^
jec*
]F
l)tcooK f
^^
ft
icoo)
(15.18)
Equations (15.17) and (15.18) clearly indicate that the hunting period
of the
output
is
only half of the period of the input variation.
course, to be expected
The average value

optimum output
from the basic parabolic relation
This
is,
T
of
of input to output.
y with respect to time, being here referred to

Equation (15.18)
y
gives the hunting loss D.
shows that this average value is the sum of second and third terms of that
the
equation with n
m.
of
Thus, noting Eq. (15.10),

CO
32fl^c
\^
~^~ 2y
=
(2n
(15.19)
I)
This equation can be easily checked by observing that when the dynamic
effects are absent, Fi = 1 then the series can be easily summed, and
;
226
a fc/3 = A*/3, as required by Eq. (15.2). Equation (15.19)

the average output and the hunting loss are independent
that
also shows
linear
of the output
group. This is, of course, to be expected, since the
is
determined
level of output
by the input #, and is not influenced by the
D = D* =
output linear system. Only detailed time variation of the output is modified by the dynamics of the output linear group.
In the case of an internal-combustion engine, the output level is the
dynamic
effects of the
(a)
(b)
r Critical Voltage Difference
(O
Input Drive Reversal
Voltage Difference
Between Condenser
&
Output Indicator
FIG. 15.7
power of the engine. The dynamics of the output linear group is determined by the inertia of the moving parts. The power of the engine is
certainly independent of the inertia of the
moving parts.
The numerical calculation of the output y from Eq. (15.18) for general
input and output transfer functions is rather difficult. However, any
practical design of an optimalizing control usually has a rather long
hunting period
to avoid the high-frequency interference.
dynamic effects, although not negligible, are not large.

we can assume that the ratios of the time constants
Then the
In other words,
of the input and
output linear groups to the hunting period are small, and carry out the
For instance, if the input linear group is approxi-
analysis accordingly.
mated by a
first-order
system with the time constant rt
-,
i.e.,
227
*w-nbs
and
if r* is
small in comparison with T, then the nondimensional quantity

Under this condition, the first few corresponding
also small.
is
Tt-ojo
(15 20)
harmonics of the series of Eqs. (15.15) and (15.16) will have practically
the same amplitudes, according to Eq. (3.14), The only difference
between these corresponding lower harmonics in x* and in x is a time
shift of
magnitude
Therefore, for regions of x* and x
r$.
drive reversal points where the curvatures of x*(t)
and
away from the

x(t)
small and where the values are determined mainly by the
monics, the
x(f)
in magnitude.
off,
curves are
first
few har-
curve merely lags the x*(t) curve by n without change

In going from a;* to #, the sharp corners will be rounded
but the general shape
of the curves
remains unchanged, as shown in
the output linear group is also approximated by a firstorder system with a characteristic time r then similar considerations
will show that in going from y* to y, the pattern of the curves remains
If
Fig. 15.7a.
This fact is shown in Fig. 15.76.

the same, but y lags behind y* by T O
With the input transfer function given by Eq. (15,20), the hunting
Thus
loss can be calculated by Eq. (15.19).
.
2
n _ 32a fc
JJ
2n
I)
(2n
00
32a%^
(2n
_ /TôY
V 2 / //
I)
1
(27i
I)
(2n
But we have
Z^
(2^
T4
4
I)
96
on
Z/
n=0
n=0
and,
V
(2n
_T
2
I)
by using the well-known expansion for the hyperbolic cotangent cited
p. 90,
V
/
Z/
/
Therefore,
finally,
'
1
1
4.
41
"T f9,
(,^^ T*
by
IW^M
79^
rÔ/^J
J-J
7T
~T
T{0)0
substituting the period
27T
C0th
L
T;
Ttt^
T from
"
T
* C0th
T'tWO
Eq. (15.14),
we
have,
228
D -
12
(?)'
48
coth
?)' (
~
I
* coth
When the time lag rt is much smaller than T, the hyperbolic cotangents
Then
are approximately equal to unity.
-
12
(j)
[l
24
(15.22)
(j)]
Since the input amplitude a can be expressed in terms of the input

drive speed and the period f, Eqs. (15.21) and (15.22) give the hunting
loss
in terms of input drive speed and the hunting period T, for a
peak-holding optimalizing control with a first-order input linear group

of lag r,-.
These equations apparently indicate that a decrease of hunting
loss is
caused by the lag of the input linear group.
However,
it is
decep-
tive: for a given critical voltage difference v for input drive reversal,
determined by considerations of noise and interference, the hunting

period T and hence the amplitude a will be much larger with the time
The net result is an increase
lags n and r than without the time lags.
instead of a decrease in hunting loss.
16.6 Design for Stable Operation.
Stability of any control system

means that the design performance of the system will be obtained even
with the presence of internal and external disturbances. We have seen
how
this
requirement
is
satisfied in the
case of conventional servo-
mechanisms and other more general control systems in the preceding

For optimalizing control systems, the essential part of the
chapters.
operation
is
the proper coordination of the input drive with the output
behavior, so that the output stays within a close neighborhood of the

optimum. This operation must not be influenced by internal and
external disturbances.
When
this is achieved
by a good design
of the
system, we have stable operation.

For a peak-holding optimalizing control, we have described the input
drive signal as the result of the charging and discharging of a condenser
by a voltage representing the magnitude

reversal signal
is
given
of the output.
when the voltage
difference v
The input
drive-
between the out-
put indicating voltage and the condenser builds up to a critical value

because of decreasing output. At the instant of reversal of the input
drive, the condenser voltage is reset by discharging to the output indicat-
When there are dynamic effects, the

ing voltage at the same instant.
time lags of the input and output linear groups will cause the output to
continue to decrease even after the signal for input drive reversal has
been given. Then the voltage v again builds up and will be removed
only after the output has risen to a value equal to that at the instant of
This is shown in Fig. 15.7c. This
signaling for input drive reversal.
229
positive spurious voltage v between the instants 1 and 2 (Fig. 15.7) is

undesirable because of the danger of tripping the input drive control
during this wrong interval of time. To greatly reduce the positive
spurious voltage, the condenser voltage
Condenser Voltage
is
reset at the instant of input
Output Indicating Voltage
(a)
(b)
FIG. 15.8
Output With Noise
^r Output Without Noise
Minimum
1 1
Critical
Voltage Difference
FIG. 15.9
drive reversal to a lower voltage than the instantaneous output indicatThus during an interval of time after the reset, the coning voltage.
of
charged by the output indicating voltage. The capacity
chosen
the condenser and the resistance of the electric circuit are so
as to make the voltage of the condenser approach that of the output
denser
is
230
when the output

shown in Fig. 15.8.
indicating voltage
of voltages is
is
increasing again.
The dangerous
The
variation
positive spurious
voltage difference is thus greatly reduced (Fig. 15.8&), and the stability
of the control system improved.
We have already stated that the reduction of the hunting zone and
hunting loss is limited by the interference and noise of the system.

This is again a problem of stable operation: we do not wish to have
Such
false signals for reversing the input drive caused by interference.
false signals will occur
reversal
is
too small.
the critical voltage difference for input drive

This can be shown by Fig. 15.9 where the
if
output y contains a high-frequency sinusoidal noise.

that
It is easily
seen
the critical voltage difference is too small, the noise will trip
the input drive in an erratic manner. For stable operation, the critical
if
voltage difference must be larger than the amplitude of the interference.

Thus the hunting loss of an optimalizing control cannot be less than the
interference
or noise of the system.
Of course,
if
the interference
actually were a pure high-frequency sine wave of constant amplitude, as

shown in the figure, it could be removed by, a filter, and a much smaller
hunting zone could be used. In fact, if the interference or noise has any
definite pattern at
limitation.
all,
we can
design a proper
filter
to ameliorate this
CHAPTER
16
FILTERING OF NOISE
In
all
the previous discussions with the exception of the last chapter,
we have tacitly assumed that the control system does not generate noise
and interference, so that theoretically there is no limit to the accuracy
In the last chapter, we have shown that noise and interof control
ference in fact obscure the output signal used for the optimalizing control
and are the fundamental design restrictions in such control systems. But
and interference are present in any engineering system, because
noise
77
even the "perfect

systems have thermodynamic fluctuations. Their
effects on the control system are negligible only if the signal is relatively
strong in comparison to the interference. For the optimalizing system,
to minimize hunting loss of the output, we design for "weak" signals, and
thus the problem of noise is of paramount importance. In general then,
whenever the control signal is weak in comparison to interference, the
and interference cannot be neglected.

disturbing influence of noise on the control system can be mini-
effects of noise
The
77
mized by introducing a proper device which will "filter out the noise
This
as much as possible without reducing the strength of the signal.
We shall first give

subject of noise filtering is the theme of this chapter.
a discussion of the theory of optimum linear filters developed by N.
Wiener 1 and A. Kolmogoroff. 2 Later parts of this chapter will treat
the various applications and extensions of this very powerful theory.
The concepts and mathematical tools introduced in Chap. 9 on random
inputs are very useful in onr present discussion.

16.1
Mean-square Error. Let f(f) be the control signal and n(t) be
the noise.
The input x(t) to tLe filter is thus
x(t)
The output from the

a linear
filter
a linear
is
and the
filter is y(t),
f(f)
as
+ n(t)
shown
(16.1)
in Fig. 16.1.
If
the
filter is
differential equation for the output-input relation
differential equation of constant coefficients,
properties are completely determined
by
its
then the
filter
transfer function F(s).
N. Wiener, "The Extrapolation, Interpolation, and Smoothing of Stationary Time

with Engineering Applications," John Wiley & Sons, Inc., New York, 1949.
2
A. Kolmogoroff, Bull acad. sci. U.R.S.S., Ser. Math., 5, 3-14 (1941).
Series
231
232
When
is
known, the response
is
F(s)
If F(s)
given by Eq. (2.18),

write
a unit impulse
of the filter to
h(t)
has poles only in the left-half
s plane,
we can
The output
r*
2iri
J - ita
2?r
due to the input
y(t)
y(t)
x(f) of
Eq. (16.1)
x(j])h(t
J-
J_
77)
= [~
Let the desired output be
x(t
which
z(i),
and the desired impulse response
then
dr\
assuming that the input extends far into the past.

y(f)
is
Let
r)h(r) dr
is
r;
then
(16.3)
determined by the signal
/(i)
h\(t}, i.e.,
/(<-r)Ai(r)dr
(16.4)
can be calculated from the desired transfer function
-Fi(s),
and we
/(.*)
(*)
FIG. 16.1
have, similar to Eq. (16.2),

r<-
I
= A-
Ai(0
ds
Since the actual output
is
not
z(f)
but
=
?/(),
(16.5)
the error
e(t) is
their differ-
ence, and, according to Eqs. (16.3) and (16.4),

e(t)
y(f)
~
=
The square
z(t)
//
/o" /o"
{[/(<
We
~r)+ n(t - r)]A(r) - /(* - r)*^)}
rfr
(16.6)
+ n(i " T)]A(r) - ^ - T) ^r)J

+ n(t - r')W) - f(t - r'W) dr dr'
(16.7)
of the error is
n(i) is
[[f(t
[/(i
rO
now make
then
T)
a very important assumption. Since the noise

a random function, only its statistical properties can be specified.
shall
FILTERING OF NOISE
In addition,
be,
but only
we do not
its
really
know
233
advance what the signal f(t) will

Therefore, even for the signal we can
broad character.
in
Then we can
specify only the statistical properties.
specify the sta-
measured by the assembly average of e*(). In

is
a function of the time instant L
this
But if we make
average
general
the assumption that the random functions f(t) and n(t) are stationary
tistical
error to be
2
as defined in Sec. 9.1, then e is independent of time.
shall
assume that both the
signal /(i)
and the noise
Furthermore, we
have zero mean
n(t)
evident from Eq. (16.6) that the mean value of the

Now we can introduce the following correlation
error e() also vanishes.
functions between the signal and the noise, entirely similar to the correvalue.
Then
it is
lation functions of Sec. 9.2:
(16.8)
Here we take the general case of nonvanishing correlation between the

Very often these cross-correlation functions Rf n and
signal and noise.
R nf
are zero,
and only the
auto-correlation functions Rf/
and
R nn
remain,
The
auto-correlation functions are symmetrical functions of the argument. The cross-correlation functions are not symmetrical functions,
but they have the following relations according to their definitions
using these definitions, the mean-square error

be written as
By
{Rff(r
r/)[h(T}
+ Rfn (r - T')[&(T) -
from Eq.
(16.7)
can
- ^
hi(T}]h(r')
Rnf(r
Rnn(r
r )h(r)[h(^
f
r%(r)h(r
dr dr
^{r
}}
(16.10)
This equation allows the calculation of the mean-square error from the
and the response to a unit impulse.
correlation functions
Control designers, however, prefer to

the transfer functions F(s) and Fi(s).
make the analysis directly with

To do this, we introduce the
Fourier transforms of the correlation functions.
transforms be
*//(o>),
$/*(), */(w), and
the following equations:
$ n n(o>),
Let these Fourier
respectively, defined
by
234
(16.11)
-;/-**
Because
of the fact that R//(T) is
$//(&>)
= -1
iwr
/
By comparing
dr
= -2
n n(co) is
is
cor
with Eq. (9.23), we see immediately that

actually the power spectrum of the signal /(Q.
the power spectrum of the noise n(t).
since the cross-correlation functions are related as

it is
cos
this equation
the function $//()

Similarly,
Rf/(T)(e
a symmetrical function of r, we can write
shown
Furthermore,
in Eq. (16.9),
easily seen that the Fourier transforms are connected in a similar
manner:
$/()
/(-&>)
= $/()
= $/()
(16.12)
1
According to the Fourier integral theorem, the inverse of Eq. (16.11)
is
T')
By
Fi(s).
==
)
we can obtain an equation

mean-square error in terms of the transfer functions F($) and
For instance, the first part of Eq. (16.10) becomes
substituting Eq. (16.13) into Eq. (16.10),
for the
But
(16.13)
= ^
F(s) and FI(S) are the Laplace transforms of h(t)
and
hi(t),
See for instance Whittaker and Watson, "Modern Analysis," Sec.
i.e.,.'.,
6.31, p. 119,
FILTERING OF NOISE
iat
h(f)e-
dt
235
(16.14)
hi(t)e^
Hence the
dt
part of Eq. (16.10) can be written as
first
at
J-
oo
$ff (uW(to>)
- Fi(iuW(-to) -Fi(-i)]d
Other terms in Eq. (16.10) can be converted in a similar manner.

then obtain,
We
finally,
+ ^ Bn (w)F(i)F(-ico)
do;
(16.15)
The integrand within the brace can be considered the power spectrum
The last term of the integrand is in fact the power
of the error e(f).
spectrum of the filtered noise, according to Eq. (9.71). Evidently the
and the last terms are real. The second and the third terms are
first
However, because of Eq.

conjugates, and therefore their sum
complex.
(16.12), these
two terms are complex
is real.
With the statistical properPhillips's Optimum Filter Design.

the signal and the noise given as the various correlation |unctions,
the Fourier transforms can be computed by using Eq. (16.11).
Then
16.2
ties of
if
the transfer function FI(S)
is
specified, the only function in the
square error integral of Eq. (16.15) yet to be fixed

of the filter.
F(s)
The optimum
filter is
transfer function F(s) such that the
is
then the
mean-square
mean-
the transfer function
which has a
filter
error
is
the
minimum
and
FI(S).
straightforward method for solving this
filter
of
optimum
problem
design is to assume a reasonable form for
with
but
undetermined
constants.
Then e 5 can be determined as a
F(s)j
function of these undetermined constants by substituting the assumed
for specified $'s
The
F(s) into Eq. (16.15).
constants are finally determined
that the mean-square error must be a

constants.
The optimum filter design
finding the
maximum
or
minimum
of
minimum with
is
by requiring
respect to these
thus reduced to a problem of

R. S. Phillips, 1
a known function.
worked out such a theory of the optimum filter by taking F(s)

be a ratio of two polynomials in s. This form of F(s) is indeed a
natural one, because our experience of linear systems indicates just such
in fact,
to
a choice.
For
1
The
actual calculation involved, however,
this reason, the
R.
S. Phillips,
Vol. 25,
Chap.
7,
more elegant theory
of
is
considerable.
Wiener and Kolmogoroff
"Theory of Servomechanisms," MIT Radiation Laboratory

McGraw-Hill Book Company, Inc., New York, 1947.
is
Series,
236
generally preferred, and
we
shall
not pursue the Phillips theory any
further here.
Wiener -Kolmogoroff Theory. The theory of the optimum filter

by Wiener and Kolmogoroff is based upon an application of the calculus
16,3
Eq. (16.15). If F(s) is indeed the optimum

filter with fixed <'s and Fi(s) then by forming the neighboring function
this
fj(s) where 17 (s) is the arbitrary variation and by substituting
F(s)
of variations to the integral of
neighboring function into Eq. (16.15),

of the mean-square error as
we
find the first-order variation
+
the
If F(s) is
optimum
filter
should vanish for arbitrary
*/()]}
eta
(16.16)
2
transfer function, then the variation 5e
rj(s).
This condition will yield an equation
However, before we can actually take this step, we have to

make some very important modifications of Eq. (16.16).
First, the power spectra $//() and <3?nn(co) are symmetrical functions
for F(s).
Thus by taking
of w.
$//,
$/, $ n/, and $ nn
expect that this
by
an even function of
sum $ can be "factored"
*() = $/,()
is
into account Eq. (16.12),
is
w.
ty(
s plane.
Now
see that the
sum
of
so that
+ $/() + *() + *() = ^M^(-ico)
definition a function with poles
5 plane.
we
It is thus reasonable to
and zeros in the
(16.17)
left-half
a function with poles and zeros in the right-half

the transfer function of the filter, for reasons of stability,
s)
then
is
can have poles only in the left-half 5 plane. Thus F(s) and 97 (s) are
functions with poles only in the left-half 5 plane. F(s) and i/( s)
are functions with poles only in the right-half s plane.
Thus the physical
requirements limit the class of functions for F(s) and
understanding, we can rewrite Eq. (16.16) as
77
(s).
With
this
__
However, not
important.
If
all of
the second terms in the braces of Eq. (16.18) are
H(s) and K(s) are functions with poles in the left-half
FILTERING OF NOISE
s
plane,
and
if
for large s they
behave
like l/s n ,
with n
237
>
1,
then
"
oo
H(i
where the closed path of integration of the second integral is the imaginary
and the semicircle enclosing the right-half 5 plane, as shown in
But the singularities of H(s)K(s) are outside this path, and
16.2a.
Fig.
axis
hence the value of the integral is zero. For the product H(~iu>)K(iu)
with singularities in the right-half s plane, the path of integration can be
made to enclose the left-half 5 plane, as shown in Fig. 16.2&. Then
s- plane
Poles
in this
Half Plane
(a)
FIG. 16.2
again no singularities will be enclosed, and the value of the integral will
be zero. For products H(iu)K(iu>) or H(iu)K(iu), the path will
always enclose some singularities, and the integral has value. Thus
to
H(-i
(16.19)
But
H(ia)K(-iu) dw
oo
(16.20)
H(-iu)K(i<*) du 7*Q
With these facts in mind, we divide Fi(s)^ff (s/i)/^f(s) into two parts:
one part with singularities in the left-half s plane, denoted by [ ]+, and
another part with singularities in the right-half s plane, denoted by [ ]_.
That
is,
tw)
(16.21)
238
Then because
of Eqs. (16.19)
and
Eq. (16.18) can be written as
(16.20),
J
(16 22)
.
Now if F(s)
is
indeed the transfer function of the optimum filter, then

Thus the quantities in the brackets
??(s).
Be*
should vanish for arbitrary
of
Eq. (16.22) must vanish.
This condition determines the optimum
transfer function as
the solution of the optimum filter problem given by Wiener and

Kolmogoroff. F(s) has poles only in the left-half s plane, since ^(s) has
This
is
zeros only in the left-half 5 plane.
The operation
with poles in the left-half
The
F(s)
is
thus a stable transfer function.
^ nf (s/f)}/^f(s}
done
be
also
In fact,
can
plane
analytically.
of picking the part of Fi(s){$ff(s/i)
validity of this equation can be easily verified
tion in the complex
co
Which
plane.
(16.24) is to be used for actual calculation

cases.
Usually, however, Eq. (16.23)

event, the properties of the optimum
by the
specified operation FI(S)
the signal and noise.
When
by contour integratwo equations (16.23) and
of the
is
depends upon the individual

easier to use.
In any
much
are completely determined
filter
$//, $/ n $ n/, and $, of

$ nn = $/n = n/ = 0, and
and the spectra
noise
is
absent,
<i>
*(ia>)(-ia>), according to Eq. (16.17). Then F(s) = Fi(s)

as expected, and the error e(t) is always zero.
When there is noise, F(s)
$/,()
not equal to FI(S), and the mean-square error cannot be eliminated

even with the best filter.
is
different interpretation of the operation of picking the part of a

function with poles in the left-half s plane can be given. As shown by
Eq. (16.2), we can consider F(iu) to be the Fourier transform of the
response
Eq.
to a unit impulse.
h(t)
calculate h(t)
from F(iw).
(16.2), the
for positive
That equation also shows how to

Because of the factor e iut in the integrand of
proper paths to take in the complex
and
for negative
t,
and are shown
has poles only in the left-half s plane, then
F(zco)
CD
plane are different
in Fig. 16.3.
If F(s)
has poles only in the
FILTERING OF NOISE
239
If F($) has poles only in the right-half s

upper-half co plane.
plane, then
has
in
the
lower-half
w
t >
Thus
for
poles
F(iu)
only
plane.
0, the
will
enclose
of
if
in
the
has
left-half
path
poles
F(ia) only
F(s)
poles
s
and
in that case the response h(t) will be different from zero.

the
0,
path will enclose no pole of F(iw) only if F(s) has all
in
the
left-half
s plane, and then h(f) = 0.
On the other hand,
poles
plane,
For
its
<
F(s) has poles in the right-half
s plane, the case of unstable F(s) accordthe

to
present interpretation, h(t) will be different from zero even
ing
for negative t.
Since the impulse is applied at t = 0, response at negative t means response before the impulse is applied.
Such behavior is
if
Contourfor*>0
Enclosing Upper
Half Plane
Complex
Contour for
<0
Enclosing Lower
Half Plane
FIG. 16.3
impossible for any physical system. Therefore we can consider the

operation of picking the part of function with poles only in the left-half
plane as the operation of making the transfer function physically

that h(t) =
for t < 0.
An explanation of the WienerKolmogoroff solution of Eq. (16.22) and (16.23) based upon this concept
realizable, so
of a physically realizable transfer function
was given by Bode and
Shannon. 1
There remains one point to be
cleared.
One
of the
assumptions in
the possibility of factoring the sum of spectra as

shown in Eq. (16.17). This is not always possible for an even positive
function $(o>) of co. To make factoring possible, $(w) must satisfy the
the present theory
Wiener-Paley
is
criterion,*
|log$(co)|
+0) 2
Actually
zero as w
$(co) is either
>
<*>
a consent, as
The Wiener-Paley
da
in-.case-.of
(16.25)
white jnois^r Approaches
criterion say&,th$n that the
approach
H. W. Bode and C. E. Shannon, Proc. IRE, 39, 417-425; (1950).

2
R. E. A. C. Paley and N. Wiener, 'Fourier Transforms IE the Complex Domain/'
Am. Math. Soc. Colloquium Publication, Vol. 19, p. 17, 1934.
1
'
240
to zero at large
a?
An approach
should not be too rapid.
w2
will
or 6~
of the
type
make
the integral
or is allowed, but an approach
Fortufactored.
be
cannot
latter
the
Thus a *(w) of
types
diverge.
noise
are
and
the
generally
actually occurring signal
nately the spectra of
71
like
ratios of polynomials of
co
2
.
usually possible.
16.4 Simple Examples.
e~W
Factoring according to Eq. (16.17)
is
thus
As a simple example, we take the power
spectrum of the signal to be
The
flat
noise
is
assumed to be white
$
The
Then
noise.
its
power spectrum
is
and a constant; thus
cross-correlation functions
= n4
(16.27)
and their Fourier transforms thus vanish,
i.e.,
$ /n = $ n/ =
(16.28)
Let the problem be to design the optimum filter for differentiating the
= s. First, we note that
signal, that is, FI(S)
+ n + nV
4
(1
Thus
+ n + nV
4
(1
Obviously to factor this function is quite straightforward.

ty(s) has zeros and poles only in the left-half s plane,
that
ately write
down
--
Remembering
we can immedi-
^(s) as
?iV
+ n V2
v7
7=
S + A/2 S +
2
Then
As
+B
-
where A, B,
C and D
}
are constants.
poles only in the left-half s plane
F (s^ff (s/i)l
l
is
the
V2
It is evident that the part

first
term.
As
+B
Hence
with
FILTERING OF NOISE
By
solving for
and $, we have,
241
finally,
(16.29)
This
the transfer function for the
is
optimum
filter.
As the noise
is
- 0, and F(s) approaches s as it should.

reduced, n
Another interesting example is the case of very high noise intensity
and a weak signal Let the noise be white noise and the power spectrum
of the noise $ W n be
Here again there
is
no
cross correlation
$//()
where k
is
is
between noise and
and
signal,
Let the power spectrum of the signal be
Eq. (16.28) remains true.
ft?()
(16.31)
a small quantity and p(w) is an even function of

<p(s/i) with poles only in the left-half s plane, i.e.,
.
If
K(s)
the part of
K( s =
)
then since
e-**
n =
J\vJ+ ^T ô
/
even in
<p(co) is
*(<*)<?*
da
(16.32)
co,
^(^ZW+JJ:(-)
\ /
(16.33)
and
[1
Therefore
optimum
if
(s)
filter is
fcX(t)][l
fc(-fo)]
(16.34)
represents the desired operation on the signal, the
given by
.
(16 35)
-
This
is
is
the second approximation for small
k.
The
first
approximation
even simpler,
F(s)
As a
specific
example,
let
fc^(*)?
(16.36)
the power spectrum of the signal be specified by
242
and
let FI(S)
Then when
s.
is
small,
1
This result checks with Eq. (16.29) by making n very large and k
Thus under strong noise

fer function is
Fi(s)
very
interference, the
much
distorted
optimum
1/n
4
.
differentiating trans-
and bears no resemblance at
all
to
s.
Besides the simApplications of Wiener -Kolmogoroff Theory.

are a number of
there
the
in
discussed
section,
preceding
ple examples
the
of
theory.
Wiener-Kolmogoroff
very important applications
16.5
We
shall consider several of
them
in this section.
These filters take the input x(f), the sum of signal

closest representaand
give an output y(f) which is the
n(t),
the signal not at t but at t + a, where a is positive. Thus
Predicting Filters.
f(t)
and noise
tion of
F,(s]
#*
(16.37)
Now let us assume that the signal is a random switching function. Then
can be written as
according to Eq. (9.50), the power spectrum
__
Ths
noise,
assumed to be white
noise, has a
(16.38)
power spectrum
$ nn = n*
The
(16.39)
Then
cross spectra vanish.
Therefore
M = Vl +T+-+
*Tr
n2
nig
Hence
In this case,
First of
all,
it will
for
>
be necessary to use Eq. (16.24) for calculating F($).

0,
+ Vl +
FILTERING OF NOISE
243
Hence, according to Eq. (16.24),
l+s
<r
._.,
(w
Therefore, finally, the optimum predicting

characterized by the transfer function F(s),
A/1
filter of
ft )
(Vl
+ n + ns)
2
predicting time
a.
is
F(s)
(16.40)
This type is similar to the predicting filters except

Filters.
the "predicting" time a is negative. A straightforward adaptation of the preceding calculation to this case will not lead to success.
Lagging
that
now
/(.*)
--.
t~
__M,..}
FIG. 16.4
In fact there
is
no
linear
system
the role of an
of finite order that will exactly fulfill
optimum lagging
obtained by using the approximation
filter.
A
y
y)1
_1
(as/2vv)\
more
direct solution can be
<
*>
(16.41)
integer
We
then obtain an approximation to the optimum lagging filter.

Servomechanisms with Noise. Let Ff(s) be the transfer function of
the forward circuit and F&(s) be the transfer function of the feedback
circuit of a feedback servomechanism, as shown in Fig. 16.4.
Let Fi(s)
represent the desired operation on the signal The problem is to find
the optimum JF&(s) with Ff (s), Fi($), and the signal and noise properties
specified.
As shown
in the figure, the equivalent F(s) is
But according to the theory, the optimum F(s)

or Eq.
(16.24).
Knowing
F(s),
is given by Eq. (16.23)

we can obtain the optimum transfer
244
function Fb(s) for the feedback circuit as

1
* b(s)
__
(16.42)
Noisy Servomechanisms. In the previous paragraph we assumed the

source of noise to be outside the servomechanism and the servo itself
However, very often there is noise generated in the servomechanism. For instance, the system shown in Fig. 16.5 has the external
noise n(t\ and in addition an internal noise m(f) from the output measuring instrument, or an output disturbance. Here again, let Ff(s) be the
quiet.
transfer function of the forward circuit,
Fb (s)
be the transfer function
(t)
FIG. 16.5
and
of the feedback circuit,
Let the Laplace transforms
be the desired operation on the signal.

m(), y(f), and z(t) be S(s), N(s),
FI(S)
of f(f), n(f),
M(s), Y(s), and Z(s), respectively.
Ff (s){S(s)
+ N(s}
Then
- F,(s)[Y(s) + M(*)]} =
Y(s)
and
Z(s)
Then
= F 1 (s}S(s)
the Laplace transform E(s) of the error

Y(s)
Z(s)
Ff (s)F
[S(s)
e(f) is
+ N(s)]
t (s)
l+Ff (s)F
M(s)
- Fl (s)S(s)
t (s)
Let us put
F(s)
(16.43)
Then we have
1
-.
l+Ff(8)Ft (8)
E(s)
[1
- F(s)][Ff(s)S(s) +Ff (s)N(s) - F
(s}8(s}}
F(s)[M(s)
+F
This equation indicates that our servo problem is equivalent to the filter
problem, with the filter transfer function F(s), and the equivalent signal
FILTERING OF NOISE
245
input S'(s) and equivalent noise input N'(s) given by

S'(s)
#'(*)
The
original
= {Ff (s) ~~F

= M(s) + F (s}S(s)
problem
+F
f (8)N(*)
(16.44)
of
optimum Fj>(s)
is
now reduced to that
of
optimum
F(s), with the equivalent signal and noise independent of the unknown
Let us assume that the original signal and noises are independent
and only autocorrelations exist. Then we have only the power spectra
$//, $ nn and $ TOW
By using Eq. (16.44), the $'s of the equivalent filter
.
problem are found.
-F&WA-ti-Ft-*
-)*.(?
=
Thus the equivalent
[Ff (-s)
filter
(16.45)
problem has cross spectra $/v and $ n 'f, in

and noises in the original problem. The
spite of the uncorrelated signal
function to be factored
is
then, using Eq. (16.45),
+ 3w(co)
The optimum
F($)
is,
P
When
F(s)
is
FI(S) for the
known, Eq.
feedback
(16.46)
(16.43) gives the
optimum
(16.47)
transfer function
circuit as
(16.48)
Saturation Constraint.
designed to
f(t)
x(t).
Consider the servomechanism of Kg. 16.6,

closely as possible the input
amplifier is F a (s), and the
make the output y(t) follow as

The transfer function of the
246
transfer function of the servo

is
to
make
mean square
the
motor
of
is
Fm (s).
the error
e(f)
the design condition

f(f) as small as
y(f)
If
by varying Fa (s), then the power of the control input to the

To avoid
servo motor may reach a very high level during operation.
the
of
we
must
that
the
of
the
input to the
power
specify
average
this,
in
this case is
error
a
must
be
value.
The
motor
mean-square
specified
possible
?=
-
$/,()
dco
(16.49)
The average power of the input to the servo motor is represented by the
mean square of the servo motor signal. This is to be kept at the fixed
FIG. 16.6
value
Then
<r .
(16.50)
j-*/
using the Lagrange-multiplier method, this problem of minimizing

with the constraint of Eq. (16.50) can be converted into minimizing
2
X being the constant amplifier. Therefore the integral to be
Xcr
By
e
e
2
2
minimized
is
-*/-'
d"
FS?/
( 16 - 51 ^
where
F(s)
Comparing Eq.
3>n/
0,
(16.52)
(16.51) -with the integral of
the present problem

*/n
is
equivalent to the
filter
Eq.
(16.15),
we
see that
problem with Fi(s)
and the equivalent noise power spectrum
1,
FILTERING OF NOISE
The function
$(co) to
be factored
is
247
thus
m(
According to Eq. (16.23), the optimum F(s)
is
_,J *()
(16-53)
by
given
Equations (16.52) and (16.54) determine the transfer function F a (s) of

optimum amplifier, with the exception of the constant X. This con-
the
stant
is
The
fixed
by the power
level
2
a,
according to Eq. (16.50).
discussion in the preceding paragraphs has perhaps demonstrated
the wide range of problems that can be solved by the Wiener-Kolmogoroff

In fact, the problem of servomechanism design for random
theory.
in Sec. 9.12, can also be conveniently solved by this
formulated
input,
theory.
Such an application to servomechanisms
is
another example
a subject treated in its generality

Whenever such criteria can be formulated, the optimum
of design according to specified criteria,
in
14.
Chap.
system characteristics are completely determined.
Furthermore, because
and the properties of the contype

trolled system, the resultant control system is linear and has constant
This concept of a more specific design of servomechanisms
coefficients.
of criteria chosen
of the particular
is
thus one step beyond the design principles of feedback servos discussed
Boksenbom and Novik 1 were perhaps the first
in the earlier chapters.
to suggest this particular application of
In
filter
theory.
control systems, the probthe detection of the signal f(t) under heavy noise interference.
In this problem, the pattern of the signal is generally known; what has
16.6
lem
Optimum Detecting
Filter.
many
is
to be detected
is
U when the
the time instant
signal has reached the
expected value /(Jo). For instance, in the case of radar, we know that
the signal is a pulse having a specified shape. The problem is to know
when the signal has reached its maximum strength. If the signal has its
maximum
at the instant
Q,
then the
filter
should give the least distortion
Therefore the optimum filter must be so designed

that the value of the signal after filter action, f (t), at the instant J is, in
at the instant
fact, /(Jo).
Jo.
Thus the
constraint
/o(Jo)
The
noise input
to,
the
filter is
is
/(ô)
a const.
We
wish to reduce the noise as much as possible by

we can write
=
i
A. S. Boksenbom, D. Novik,
(16.55)
n (t).
and thus
n(); the corresponding output
NACA TN
min.
2939 (1953).
filtering,
is
(16.56)
248
The problem
its
is
to determine the transfer function F(s) of the filter or
equivalent, the response h(f) of the
/() and
and the noise
filter
characteristics
where X
is
or
$w, such that
- mn.
The
the Lagrange multiplier.
given in this generalized
to a unit impulse, with given
R nn
(16.57)
solution of this problem
form by Zadeh and Eagazzini.
is
We shall follow
their analysis,
If
^(w)
can be factored, so that
$n
(w)
(16.58)
(i)tf(-e'w)
in the left-half s plane; then it is

^r(s) has zeros and poles only
convenient to consider two amplifiers in series, where one has the transfer
where
"
>TS)~
<*>
FIG. 16.7
function l/*(s) and the other the transfer function *($), both in series
again with the filter, as shown in Fig. 16.7. This system is equivalent
to the original system.
one unites),
We
consider the last two transfer functions as
i.e. t
f"(s)
V(s)F(s)
(16.59)
with the response k'(t) to an impulse. The inputs to F'($) are the signal
and the noise n'(f). If S(iu) is the Fourier transform of the signal
jf'(tf)
/(O, M.,
S(to)
dt
f *^f ()***
(16.60)
then
HA
fin
(ID.OIJ
&(*<*) ,***.,
*0
^77T6
The power spectrum
of
n x (^)
is
now, according to Eq.
=
Therefore the noise
is
now a white noise with the autocorrelation function
B*vM =
J
(9.71),
L. A. Zadeh, J. R. Ragazzini, Proc,
IRE,
(r)
40,
1223-1231 (1952).
(16.62)
FILTERING OF NOISE
The outputs from
now be
the system can
M0
The mean square

Or
of
A'(r)n'(*-T)dr. v
Jo
output noise
is
A'TA'r'flnvT
But the
dT'h'(T)h'(T )n'(t
correlation of noise n'(t)
r)n
(t
(16.64)
dr dr
T'}
Hence
given by Bq. (16.62).
is
(16.63)
then the assembly average of nj(0-
dr
written as
l*VW(t-r)to
and
(f)
249
(16.65)
Therefore the minimization condition of Eq. (16.57) can be converted
by Eqs.
(16.63)
and
(16.65) to
2
[h (f)]
dt
2X
A'(0/'(*o
-f)dt = min.
This can be rewritten as

2
*)]
*-
X2
[/'ft,
O]
min.
(16.66)
JT
But with/(0 and

calculated by Eq.
is
ntt (w)
(16.61).
a fixed constant.
is
The
by the problem, f'(t)
is
a fixed function
Therefore the second integral of Eq. (16.66)

For the
first integral is positive or zero.
the sum, the first integral of Eq. (16.66) should be zero.

so only if the quantity within the square bracket of that integral
minimum
This
specified
vanishes.
of
Thus
h'(t)
X/'ft,
t)
for
>
(16.67)
= for t < 0. In
Naturally, for physically realizable systems h'(()
r
is identical,
an
to
F
of
the
impulse
other words,
optimum response
This result
to
with
of
the
with
U/2.
respect
(0
for positive t,
image
derived by
first
was
and
time
some
for
known
been
has
noise
for white
D. 0. North.
With the
result of Eq. (16.67), the
problem can be immediately written
and
optimum filter F(s) of the original

down with the help of Eqs. (16.59)
(16.61),
'
(16 68)
-
250
where X
to be finally fixed
is
by the constant /(fo) in Eq. (16.55). This

optimum filter given by Eq.
result is very similar to that of Wiener's
In fact we can write
(16.24).
(16 69)
'
[
]+ again denotes picking the part of the function within the
bracket having poles only in the left-half s plane, or making the transfer
function physically realizable. Equations (16.68) and (16.69) are the
where
given by Zadeh and Ragazzini.

using the artifice of an equivalent
problem (Fig. 16.7), in which the noise component of the input is a
white noise. This device is generally very useful in simplifying the
formulae for optimum detecting
The derivation
is
filters
made elementary by
analysis of complex optimization problems.
Other Optimum
16.7
filter
One
Filters.
of the basic
theory discussed in the preceding sections
random
functions, either as signal or as noise.
assumptions in the
the stationarity of the

No true stationarity can
is
hold for very long time, because of the natural drift of the system or a
purposeful change in the state of operation. More likely, the random
inputs are stationary only for a limited time interval T. For time intervals longer than T, the random functions are not stationary.
Therefore,
if the filter is designed with the assumption of a stationary random
function and
if
filter is much larger than T,

and the performance of the
would be better to deviate from the
the characteristic time of the
reality will not correspond to the theory,

filter suffers.
theoretical
In such a case,
"optimum"
filter
it
and use one with a shorter characteristic
time.
more
filter
satisfactory solution
calculated from the input
is
is
to include the time
explicitly in
our
We
can do this by requiring the impulse response h(t) of the

< t < T, Then the output y(t)
to be zero outside the interval
theory.
x(t)'
as follows:
h(r)x(t
r)
dr
(16.70)
This equation demonstrates the fact that the output is dependent upon
the input only for a finite time T back from the present.
Therefore such
filters
may
be called
filters of finite
memory.
The
filters
discussed in
<*>
the previous sections, having T
are thus infinite-memory filters.
1
Finite memory filters are discussed by Zadeh and Ragazzini.
They
,
give a solution of the
optimum
filter for
to that of the preceding section.

1
Zadeh and Ragazzini, op. at.

L. A. Zadeh, J. R. Ragazzini, J.
Appl
the problem of detection similar
They
also give a solution of the
Phys., 21, 645-654 (1950).
FILTERING OF NOISE
optimum
filter
composed
more complicated problem with the signal input

parts, a stationary random function and a nonrandom
for a
two
of
251
function expressible in a polynomial of nth degree in

of a
finite-memory
mean
of the
error,
mean
For
t.
the performance criteria are,
filter,
this
first,
problem
vanishing
and second, minimum mean-square error. The vanishing

error is no longer automatically ensured because of the
nonrandom part
of the signal.
The
solutions given
by
their theory for
these problems are, however, generally difficult to realize
RC
circuits.
filters,
by simple
In fact, even for the simpler problem of infinite-memory
the solution specified by Eqs. (16.68) and (16.69) is sometimes

Practical filters can be only approximations to the
difficult to build.
theoretical
optimum.
Then the value
mainly in giving a guide for design

16.8
of
the theoretical solution
and a standard
General Filtering Problem.
Of course,
lies
of ideal performance.
it will
be
still
possible to
use the complicated theoretical optimum filter design if we abandon the

inadequate RC circuits and use an analog computer or even a digital
computer to serve as the filter. Then the theoretical optimum performance can actually be attained. However, the introduction of an electromechanical computer as a component of the filtering system greatly
increases the complexity of the over-all system
and can be
justified only
we have made the system very complicated at

we have actually obtained the very best
we
ask
whether
may
high cost,
The
optimum performance in the theory discussed in the
performance.
is
sections
only optimum within the limitations of the assumpprevious
For instance, two random signals with the same
tions of the theory.
in very critical cases.
If
correlation function or the
same power spectrum, according
developed, require the same optimum

looseness in design criteria.
Surely,
This
filter.
if
is,
we have more
to the theory
in a sense, a certain
statistical
informa-
power spectrum, we should be able

to distinguish these two signals and to improve our design by utilizing
such additional knowledge. Then we can obtain even better performtion about the signal than just the
ance than possible with the so-called "optimum" filter. It is evident

that this generalized approach to the filtering problem must require a
more advanced theory
of probability
than we have used.
The
recently
may also find important applicamade in this " probabilistic " approach
developed science of information theory

tions here.
A beginning
has been
to the problem of detecting a signal in noise.
But much remains
to be
done.
The Wiener-Kolmogoroff theory of the optimum filter

the mean-square-error criterion. By using this criterion,
1
See for instance P.
M. Woodward,
Phil Mag.,
is
based upon
we
essentially
1001-1017 (1950);
Proc. IRE, 39, 1521-1524 (1951); J. InsL Eke. Engrs. London, 99(3), 37-51 (1952);
and T. G. Slattery, Proc. IRE, 40, 1232-1236 (1952).
I.
L. Davies,
41,
252
without much considerput the emphasis on minimizing the large errors,

most
ation of the small errors.
However, on many occasions, we may be
while
not
error as small as possible,
the
interested in
making
frequent
It is also possible
error.
being particular about an infrequent large
mean far from
the
with
that the probability function is very lopsided,
is entirely
criterion
the
the mode. For such cases,
mean-square-error
of prethe
consider
we
problem
a
As
simple example,
inappropriate.
will be a clear day, quoted by Bode and
tomorrow
whether
dicting
Shannon. 1 Since clear days are in the majority, and there are no days
with negative precipitation to balance days when there is precipitation,
the probability function is very much lopsided. With this function, the
the
average point, which is the one given by a prediction minimizing
be represented by a day with a light drizzle.
mean-square error, might

To a man planning a picnic, however, such a prediction would have no
value at all. He is interested in the probability of having a really clear
would ruin a picnic.
day, since even a smaU amount of precipitation
filter of the present chapter also assumes
of the
The
optimum
theory
a linear filter in that the differential equation relating the input to
the output of the filter is a linear equation with constant coefficients.
This is clearly a self-imposed limitation, taken with the purpose of
the theory and with the knowledge that such filters are easily
simplifying
RC circuits. For controlled systems that have timeof Chap.

varying characteristics, such as the ballistic guidance problem
of
case
In
this
suitable.
time-varying
13 such filters are clearly not
have time-varying characteristics.
systems, the proper filter should also
If the filter is still linear, so that the principle of superposition holds, the
synthesized out of
But
input-output relationship is still controlled by the impulse response.

now this response is a function h of two time variables t and J*.
t is
when the response is taken, if* is the time instant when

Then the output y(f) can be calculated from the
applied.
the time instant

the impulse
input
is
x(t) as follows:
y(f)
The optimum
filter
then of finding a
problem
way
Bode and Shannon,
is
r}x(t
then that of
r)
first
to actually implement this
impulse.
1
h(t,t
op.
tit.
dr
determining
(16.71)
h(r,t)
optimum response
and
to an
CHAPTER
17

In the preceding chapters, we have indicated how control systems
complexity can be designed to give almost any specified per-
of great
formance.
Of course, the greater the complexity
greater is the chance of malfunction caused

the system or the failure of an individual
of the system, the
by mistakes in assembling
component of the system.
Therefore, for complex control systems, the problem of the reliability

In
of the design in actual operation becomes one of utmost importance.
this and the last chapter, we shall consider this problem from two differ-
ent points of view.

In this chapter we shall discuss the possibility of building into the
system a certain measure of flexibility and adaptability so that incidental
and unexpected mistakes in design will be automatically corrected by the

control system itself without external human aid.
Such a control system
is thus capable of learning how to behave properly and has almost the
homeostatic mechanisms of living organisms, which enable them to survive
under varying conditions of environment. Naturally this concept of
self-adjustments in a complicated system came from the study of the
behavior of living creatures, because there these characteristics are most
In fact, our discussions in this chapter are based upon a remarkevident.
able book by W. R. Ashby 1 on the origin of the nervous system's unique
There may be different opinions
ability to produce adaptive behavior.
about how the brain of an animal is actually constructed. But our task
is
simply to indicate that
by mechanical means.
it is
possible to achieve such adaptive behavior
Whether the suggested mechanism
is
the only
one possible does not concern us here.

17.1
For simplicity, let us consider an autonovariables

two
by
yi and t/ 2 such as those considered
Ultrastable System.
mous system
in Sec. 10.5.
specified
The phase plane
is
then the yiyz plane.
If
is
the time,
then the simultaneous differential equations determining the behavior

of the system can be written as
4=
1
W. R. Ashby, "Design for a
Brain," John Wiley

253
&
Sons, Inc.,
New
York, 1952.
254
where we have included
in the functions /i
and /2 an extra parameter f
indicating that the functional relationships between dyi/dt and dy*/dt

on one hand, and y\ and 2/2 on the other are fixed only if the parameter f
In particular, we shall allow f to take a series of discrete values.
is fixed.
The pattern of behavior of the system is determined by lines of the loci of the
point
(2/1,2/2)
as time increases, starting
from various
initial
points in the
phase plane. Clearly then, there are as many different patterns of

behavior of the system as there are different values of the discrete parameter f
For instance, the
linear
system discussed in Sec. 10.5 (with y\
and
y) has the parameter f

2/2
one
negative, less than
values,
to
one
0;
positive, between
equal
than
then the
1,
If f is capable of taking five different
1 and 0; one
one negative, between
1 and finally one positive, greater
1;
and
behavior of the system are indicated by

Another example would be
five patterns of
Figs. 10.12 to 10.16.
(17.2)
where the
Then
coefficients an, #12, #21,
there are as
different values of the
and a^ are monotonic functions
of f
different sets of these coefficients as there are
many
parameter f
Each
set of the coefficients gives a
definite pattern of behavior.
Some
of the patterns of behavior will
be stable in that
all lines of
behavior tend to a point in the phase plane, the stable equilibrium point.
Some of the patterns of behavior will be unstable in that the lines of
behavior tend to diverge from the equilibrium point.
of the system naturally requires stability.
formance
Satisfactory per-
We
shall achieve
the desired adaptive behavior of the system if we can cause the system
to reject automatically the unstable patterns of behavior and to retain
the stable patterns of behavior. Now see what happens if we surround
the desired equilibrium point of the system by a closed boundary and if
we build the system with a switching device such that the parameter f will
jump
to a different value whenever the line of behavior of the system

If we start with a pattern of behavior as shown
crosses the boundary.
in Fig. 17. la at the point
system
will cross the
the pattern of behavior is unstable, and the

at the point PI.
At the instant of cross-
boundary
ing, the switching device acts so that
the parameter f jumps to a different value, and the pattern of behavior of the system changes to that of
This pattern
is also unstable, and the

system moves from
point PI to P 2 where it crosses the boundary again. The switching
device changes f, and the pattern of behavior becomes that of Fig. 17. Ic.
Fig. 17.16.
But
this pattern of behavior, although it contains a stable

equilibrium
255
Thus
system to move out of the boundary at P 2
the
and
time
a
third
the switching device operates
pattern of
changes
stable
the
is
This
17.
Id.
behavior to that of Fig.
pattern, and the
This
the
P
to
2
equilibrium point P 3
system moves from the point
the
of
condition
in
a
system
stability,
pattern will be retained because,
will not cross the boundary and thus will not activate the switching
point, requires the
device to change f
FIG. 17.1
device and a preTherefore, with only the addition of the switching

scribed switching boundary in the phase plane, the system is made to
seek stability automatically, rejecting unstable patterns and retaining
Furthermore, the switching action in
Stable patterns of behavior.
stable pattern
changing the parameter f can be entirely random. If the
attained by the first switching, so much the better. But the end
result is the same, whether the system has to switch once or three times.'
is
Stability
is
always achieved.
Therefore
we are able to
obtain purposeful,
by purely mechanical means. Such a system is

thus automatically stable; it does not have to be designed to have stabilgoal-seeking behavior
256
ity, it will
become
stable
by
Therefore
learning,
an
it
has something more
ultrastable system, after
than just stability and will be called

Ashby.^
Our two-variable autonomous system can be generalized to an ^-vari= 1,
Then the
n.
able system with the variables y i} where i
.
differential equations
can be written as
(17.3)
the
the parameter, f jumps to a different value whenever
the
boundary.
crosses
switching
line of behavior in the phase space &- 1 dimensions
The switching boundary here is a closed hypersurface of n
where f
is
n dimensions. Such a system is also ultrastable.

In order to demonstrate
17,2 An Example of an Ultrastable System.
a relatively
constructed
the behavior of an ultrastable system, Ashby
1
The
homeostat.
the
called
he
which
of four
in the phase space of
variables,
simple system
four variables y 1} y%
whose motion
trolled
is
by four
y$,
and
z/ 4
are the angular deflections of four
heavily damped.
coils,
each
of
which
The
is
position of each
magnets
magnet
is
con-
fed with currents generated by the

Because of the heavy damping, the
angular position of the four magnets.

motion of the magnets is slow, and the inertial forces can be neglected.
The equations of motion can be obtained by equating the turning torques
Thus
produced by the coils to the damping torques.
-jr
at
(17.4)
=7-
CLnyi
"f*
#32^/2
#332/3
at
~dt
The magnitude of the coefficients a can be modified by the experimenter

by regulating the current in the coils with a variable potentiometer. The
the coil circuits.
sign of the a's can also be changed by the commutators in
is
the
coils
fed
one
of
each
In addition, for
by current passing
magnet,
The position
has
25
which
possible positions.
through a uniselector
of the uniselector will
jump randomly
to a different one whenever the
magnet reaches 45 degrees in either direction.

Thus four of the coefficients a#, for i,j = 1, 2, 3, 4, are each capable of
angular deflection of that
having one of the 25 values chosen randomly by the four uniselectors.

For each magnet, or each variable y^ we have a block diagram similar
to that
1
shown
in Fig. 17.2.
W. R. Ashby,
Electronic Eng., 20, 379 (1948).

The switching boundary
of the
homeostat
257
thus a "cube" in the
is
four-dimensional phase space centered around the origin with sides corresponding to 90-degree angular motion. For every setting of the a's by
the experimenter, the four uniselectors give 25 4 = 390,625 combinations
of the four coefficients controlled by them.
Therefore, for every hand
setting, there are 390,625 patterns of behavior of the homeostat, some
[Uniselector Action Whenever]
= 45
FIG. 17.2
stable,
some unstable.
The unstable
patterns
will,
however, be auto-
matically rejected.
The
ultrastability can
now be demonstrated.
First, for simplicity,
a single unit is shown arranged to feed back into itself through a single
The behavior of the single
uniselector; the other coils are disconnected.
magnet
is
shown
in Fig. 17.3,
where the upper curve
is
the deflection
-Time
Wl/lr
FIG. 17.3
magnet and the lower curve indicates the uniselector action. At

magnet is disturbed by hand; but the uniselector position happens
to give a stable pattern of behavior, and the deflection is promptly
corrected.
At jRi, the feedback is reversed by hand. The old uniselector
position now makes the system unstable, and the deflection of the magnet
The
reaches the switching boundary (the dotted line in the figure).
uniselector acts.
After one jump of the uniselector, the pattern becomes
At Rz, the feedback
stable, as shown by the test disturbance at D%.
is again reversed by hand.
This time it takes four random jumps of the
of the
DI, the
258
At
uniselector to obtain stability.
jD 8 ,
the system
is
again shown to be
stable.
As the next example, we have two magnets yi and 2/2 interacting. The
coefficient a 2 i is set by experimenter, and the coefficient a n by the uniFor each setting
All the other coefficients are set to be zero.
selector.
due to the 25
behavior
of
different
of tt2i, there are then 25
patterns
can
be shown as
of
results
The
experiment
settings of the uniselector.
of yi and
deflections
the
indicate
curves
Fig. 17.4, where the two upper
At
uniselector.
the
of
the
action
the lowest curve indicates
DI, the
2/2
;
setting of the uniselector

yi
and
2/2
are in the
happens to be a stable one, and the deflections
same
sense.
At
Ri, the sign of the coefficient a 2 i
FIG. 17.4
is
changed by reversing the polarity of the
coil for
the y% magnet fed
with current from the yi magnet. This induces instability and causes
the 7/1 deflection to reach the switching boundary. One jump in the
uniselector setting corrects the situation.
At D 2 a test disturbance
,
shows that the system
is
now
stable,
with yi and y%
of opposite signs as
expected. At jR 2 the sign of a 2 i is again changed to that at DI; and

the uniselector setting is again unstable. Now it takes three jumps
of the uniselector to reach stability.
At D%, the system is seen to be
,
and t/ 2 in the same sense.

As an example of the adaptability of an ultrastable system to even
an unforseen situation, a situation not thought of until the machine has
been built, we consider the course of events shown in Fig. 17.5. Here
three magnets 7/1, 3/2, and t/ 3 are interacting.
The behavior of the system
is at first stable, as shown at DI.
The deflections z/i and y z are in the
same sense, but y 2 is in the opposite sense to them. Now at /, we substable again, with deflections of yi
homeostat to a new, unexpected circumstance by joining the

magnets t/i and ?/ 2 so that they must move together. This additional
ject the
constraint
is
counter to the pattern of behavior of the prevailing setting
259
The system becomes unstable and the consequent
of the uniselector.
large deflection causes the uniselector to act.
Three unstable patterns
are successively rejected before the terminal stable pattern
is reached, as
the
connection
between
the
.R,
magnet y\ and y% is
becomes
the
unstable
and
system again
broken;
requires new action of
shown
At
at D%.
the uniselector.
-/c^--v-~ Switching
Boundary
-Time
-fft
FIG. 17.5
17.3 Probability of Stability.

In the previous section, we have demonstrated the characteristic adaptive behavior of the ultrastable system
The
in seeking stable patterns of behavior under any circumstances.
question then arises: Is this searching for stability always crowned with
What is the probability of success? If we taken an autono-
success?
mous system
by Eq. (17.3), each value of the

an ordinary dynamic system. The whole
the parameter f then gives an assembly of autonomous dynamic
of
variables as specified
discrete parameter f gives
range of
systems of n variables. We may define the probability of stability within

the switching boundary in the phase space as follows. Pick a point
P(yj) of the phase space, and construct an infinitesimal neighborhood
dV around
this point P.
We
can then find the differential probability
dp as the fraction of the assembly of dynamical systems that will have a

Now integrate dp over
stable equilibrium point within the volume dV.
This then gives
the phase space enclosed by the switching boundary.
the system's general probability of stability p corresponding to the specified
switching boundary.
Needless to say, the actual calculation of this general probability of
In order to gain
stability is a very difficult mathematical problem.
some ideas about
this probability,
assembly of linear systems
dt
of the
Ashby made a
trial calculation for
an
form
i,
I?
>
(17.5)
260
only one equilibrium point, the origin. The question of stability

of the system is solved by considering the determinantal equation
There
is
1 f or i
for
the roots X
If
In
all
the root X
fact,
ability of stability
(17.6)
i 7* j
real parts, then the- system is stable.

The probcalled the latent root of the matrix o#.
have negative
is
is
thus the probability that the latent roots of the
sample matrix a# belonging to the assembly will all have negative real
Ashby took the simplest possible distribution, the rectangular
parts.
That is, each element o# of the matrix has an equal chance
distribution.
an integer between -9 to +9, inclusive. The sampling of aiy

was done with the aid of the table of random numbers. When n = 1,
For other orders of the system,
the probability of stability is obviously
1
His results are
rule.
Hurwitz's
the
tested
stability by using
Ashby
of being
shown
Table 17.1. It is seen that the probability of stability p

with increase in the order of the system. As an
decreases
steadily
n
approximation, the probability is }4
in
TABLE
17.1
The probability of stability can be increased by limiting the o# to

favorable values. For instance, we can make all diagonal elements
of the matrix have zero or negative values.
Then if the variables are
not interacting, the system will be always stable. The probability of
= 1, is now clearly unity. When n = 2,
stability for one variable, orn
the probability is f. Ashby's test results are shown in Table 17.2.
TABLE
The probability of
number of
17.2
thus higher, but nevertheless it decreases

increased.
From these investigations it
seems reasonable to say that the probability of stability of a randomly
as the
stability
variables
is
is
A. Hurwitz, Mathematischen Annalen, 46, 273 (1875).
ULTRASTABILITY AND MULTISTABILIT Y
261
constructed system steadily decreases as the system becomes more comLarge systems are very much more likely to be unstable than
plicated.
to be stable.
17.4 Terminal Fields.

Following Ashby, we shall now give a definite
name to the pattern of behavior at any one setting of the parameter f
The pattern in the phase space will be called the field of lines of behavior.
The fields will be different for different values of the parameter. The
.
final stable field after
the terminal
field.
the switching action on the parameter will be called

action of an ultrastable system is thus mainly
The
the search for the terminal
field.
It is
thus of primary importance to
of switching actions necessary to reach the

the average number
is very simply related to the probability
terminal field. This number
The probability of reaching the
of stability p of the ultrastable system.
know
stable terminal field at the
the probability p
first
The
itself.
action of the switching action is evidently

probability of not reaching the terminal
1
If the switching action is perfectly random, then the
p.
that
the
second
field will be stable is still p and the probability
probability
that the second field will be unstable is q. The relative probability that
field is q
the second
be terminal
field will
that the second
field will
is
thus pq.
not be terminal
is
The relative probability

Hence we arrive at the
g
conclusion that the relative probability for reaching the terminal field
m~ l
of switchThe average number
at the mth switching action is pq
the
is
thus
terminal
field
actions
for
reaching
ing
.
00
mpq
m~
m=l
When p is very small, for very large systems, the number of switching
actions necessary for reaching the stable terminal field will be very large.
That is, the search for the terminal field will be long and will seem to
follow a devious road.
terminal field
may
be singular in the sense that only a very small
fraction of the lines of behavior tend to the equilibrium point, while
other lines of behavior will diverge from the equilibrium point and cross
the switching boundary. Such a field will be terminal only if the line
of behavior from the switching boundary happens to be one of the small
former group. It will be shown presently that such singular terminal

fields are not favored.
If, among all possible fields of an ultrastable
system, there
of
such
fields
is
a certain fraction of such singular
used as terminal
fields is
much
smaller.
fields,
the fraction
To show
this, let
262
k be the fraction of the surface of the switching boundary from which
the lines of behavior will reach the equilibrium point.
the
fields
shown
Fjg. 17. Ic, k
is
in Fig. 17. la
and
6,
0.
For the
approximately i
dk be the fraction
For example,
For the
field
field
shown
stable system having k between k and k
dk.
f(k)
is
for
shown
in
17. Id,
in Fig.
of all possible fields of
Now let f(k)
1.
fe
the ultra-
thus the distribu-
tion function of the possible fields of the ultrastable system,
and by
definition
fc
Then because only the
lines of
(17.8)
behavior that start from the part
fc
of
the
terminal field, the relative probability

switching boundary can produce a
dk is
the terminal field with its k in the range from k to k
of
having
kf(k) dk.
Thus the
distribution function g(k]
related to the distribution function
stable system
/(fc)
of the terminal fields is
of the possible fields of the ultra-
by
Vf(fc') dk
Evidently
g(k)
The terminal
of k, as
The
fields
shown by
Fig.
dk
(17.10)
are thus definitely crowded toward larger values

17.6.
Singular terminal fields are thus not favored.
actual terminal-field distribution
is,
however, not necessarily the
potential terminal-field distribution

The reacalculated by Eq. (17.9).
son
is
that any equilibrium state

field is subject to
within a terminal
random disturbances.
If the equilib-
rium point in the phase space
is
closed to the switching boundary,
then even a relatively small disturbance might put the instantaneous
system outside the

switching boundary and thus destroy
that field. Therefore, for a terminal field to have a high probability of
being retained under random disturbances, the equilibrium must have a
state
central location in the switching boundary.
17.7 the field

field
To
is definitely
more
of
the
For instance, as shown in Fig
stable than the fields
and B.
The
contains both an unstable equilibrium point and a limit cycle.

put this concept of stability under random disturbances into a
quantitative form
we have
to introduce the probability
<r
of retaining a

terminal
field after
a single disturbance.
If
the
field
263
contains a single
shown in Fig. 17.7, and if the distribution

function of the random disturbances from P is specified, say by a Gaussian
stable equilibrium point P, as
distribution, then
cr
is
simply the integral of this disturbance distribution
function over the part of the phase space enclosed by the switching
boundary. If the terminal field contains a limit cycle S, then cr is the
average of the probability of retaining the field, taking as the equilibrium point each point on the limit cycle, and weighting this probability according to the proportion of time the system will spend at that
Boundary
(a)
(c)
(b)
FIG. 17.7
point of the limit cycle. We can then assign a probability cr to each

terminal field. Let the distribution function of the terminal fields in a
be
<p(o-)j
the probability of finding a terminal
i.e.,
range from
o-
to
cr
da- is <p(d) d<r.
da
p(tr)
The
calculate
\l/(d)
in terms of
with
<r
actual terminal-field distribution function
da
To
field
in the
Thus
^(<r),
(17.11)
is,
however,
$(<?),
with
(17.12)
we observe
that the distribution
$(<r), being the actual final terminal-field distribution, will not be altered
by random disturbances. Secondly, after one disturbance the fields with
a value of
<r
between
<r
and
<r
d<r, \l/(<r) dcr
in relative
number,
the probability a of being retained and the probability

destroyed.
The
disturbance
is
total relative
(1
of fields destroyed
have
d) of being
by one random
thus
The new terminal

minal
number
will
fields, i.e.,
fields
are distributed according to the potential ter-
according to
#>(<r).
Therefore the total relative
number
264
of terminal fields in the range
<^(cr) dff
cr
to
<r
+ da- after one random disturbance is

(1 -cr OlK*')
dff
?(*)
&'
because random disproportional to $(a) dc,

distribution
Thus, if C is
$(<r).
turbances should not change the final
But
this relative
number
is
the constant of proportionality,
By
to
integrating this expression with respect
can be shown to be unity with Eqs. (17.11)
<r
and
from
or
= Otocr =
(17.12).
C
we
Therefore
have
In this equation the integral

lK<r) is
of
just a constant independent
is
proportional to ?(<r)/(l
Or by using Eq.
cr).
Thus
cr.
(17.12),
(17.13)
actual terminal-field
Equation (17.13) allows the determination of the
Since the
distribution.
distribution from the potential terminal-field
be calcu
can
potential distribution
lated from the distribution of
all
pos-
sible fields of the ultrastable system

by Eq. (17.9), we can, in theory at
least,
obtain the actual terminal-field
distribution
\l/(<r)
from the specified
properties of the ultrastable system.

(17.13) shows, furthermore,
that the actual terminal-field dis-
Equation
tribution
FlG 17g
toward
is,
as expected
by our previous
dicated in Fig. 17.8,
random disturbances
function
<p(d) ,
Eq. (17.13),
is
^(<r)
is
far
more crowded
larger values of
intuitive arguments.
cr
than
This fact
<f>(a)
is in-
noted that the particular type of

only the calculation of the distribution
It should be
will affect
while the relationship between $(a) and
independent
of it
^(er) ?
as specified
by
and holds for any form of the distribution
of disturbances.
17,5
number
is
Multistable System.
As shown
in the preceding section, the
of switching actions necessary for reaching a terminal field
expected to be 1/p, where p
is
the general probability of stability
265
of the fields of an ultrastable system.

Since we have concluded that p
decreases to very small values for large systems, the number
may be
For
if
there
are
100
variables
in
the
very large.
instance,
system,
1/2
100
,
and
2 100
10 30
Even
if
we
allow for 10 switching
actions per second, the average time for reaching a terminal field will
still be 3
10 19 centuries. Such a long settling time may well be con-
sidered to be infinite, and the ultrastable system will practically never

reach a stable terminal field. Therefore, for large systems, or just where
the principle of ultrastability for automatically reaching stable behavior

important, we find the concept to be impractical.
is
To remedy
this situation,
One way
of the system.
to
we must
do
this
increase the probability of stability
would be a compromise.
We require
the design of the system to be such that the fields of the ultrastable
system are limited to those which are stable under expected operating
and minor adjustments need be made by the

In other words, the system is designed according to
the conventional methods without using the principle of ultrastability.
Ultrastability and switching are introduced only when trouble is expected.
conditions.
local
Only
switching action.
For instance, we can design an autopilot for an airplane according to the

But we have an appreprinciples discussed in the previous chapters.
hension that the assembly mechanic might plug the control signal from
the autopilot to the aileron in a reverse manner, so that the aileron down
If the mechanic actually
signal actually produces aileron up motion.
makes
this mistake,
then the autopilot
rolling motion.
will
not stabilize the airplane;
be unstable, with diverging

However, the designer's misgiving can be quieted by
instead, the autopilot-airplane
system
will
introducing ultrastability just at this point. The signal connection is

made to switch automatically whenever the rolling motion exceeds the
Then the system becomes an
specified values, the switching boundary.
ultrastable system of
maximum
two possible
fields,
one
stable,
one unstable.
one switching action is necessary to reach stability, in

spite of the fact that there may be a very large number of variables in
the autopilot-airplane system. The point is, of course, that we do not
of only
need to leave everything to chance.
under almost
all
conditions
We
can design for stable operation
and greatly reduce the choice
of fields of
behavior by providing safeguards only for those contingencies expected

to occur.
Hence it is a compromise between the principles of conven-
and the principle of ultrastability.

no prior knowledge of the conditions of the surroundings can be assumed, and therefore it is not possible to increase
tional control design
For
living organisms,
the probability of stability by limiting the choice of fields of behavior

of the system.
Ashby discovered a different way of increasing the probHe observed that for a very complicated system
ability of stability.
266
containing a large number of variables, any single disturbance or change

of operating conditions directly affects only a relatively small number of
Thus if the variables directly disturbed can be isolated

and form an ultrastable system by themselves, then the
these variables.
from the
rest
probability of stability for that particular type of disturbances can be

For instance, if the number of variables directly
greatly increased.
influenced at
first
any one disturbance
paragraph
is five,
instead of 100 as assumed in the
of this section, the expected
time of reaching
its
terminal
only 3.2 seconds, with 10 switching actions per second. Thus

the 100 variables are actually divided into 20 groups of five variables
field is
if
each, forming 20 separate ultrastable systems, the total time required

new set of operating conditions will be
for adapting to a completely
X 3.2 = 64 seconds. This is a tremendous reduction from the

X 10 centuries required for a completely connected ultrastable system
20
3
of
19
100 variables.
Of course a system composed of 20 separate systems of five variables

each does not have the flexibility and the rich response of a system of
100 interacting, connected variables.
however, any one disturbance
If,
involves only five variables, then separate systems of five variables can
be made to be equivalent to the system of 100 variables by making the
any subsystem change according to the

y*, and y$, then
2/1, y^ y$
these five variables are associated to form an ultrastable system of five
variables.
If a later disturbance affects the variables y z y*>, y^ y^ and
2/99, then these five variables are associated to form an ultrastable system.
association of the variables in
disturbance.
If
the disturbed variables are
This phenomenon of ever-changing organization of variables into subsystems according to operating conditions is called by Ashby the dispersion of behavior.
Dispersion can be actually achieved by making
the functions fi(yi,y%
,2/;f) in Eq. (17.3) zero when the point
Then
(2/1,2/2,
,y n ) is within a certain region of the phase space.
.
those
yi will
be constants with respect to time and are tentatively only

In our example above, for
parameters with respect to other variables.

the first disturbance, the point (yi,y^
,2/n) is in a region of phase

space such that /< is zero except when i = 1, 2, 3 4, and 5. For the
second disturbance, /, vanishes except for i = 2, 5, 10, 98, and 99. It
is evident that this behavior of the functions
$ only means that various
.
thresholds exist for the derivatives dyi/dt. Such thresholds are naturally
be expected in any real system. Therefore dispersion is something
easy to obtain.
to
system
dispersion
is
of ultrastable subsystems organized with the

possibility of
called by Ashby a multistable system.
multistable
system
obviously has the adaptive behavior of an ultrastable system simply

because it is composed of ultrastable subsystems. However it differs
ULTRASTABILITY AND MULTISTABILIT Y
267
single ultrastable system of an equal number of variables in

the time required for reaching a terminal field. The multistable system
has a very much shorter settling time and makes the principle of ultra-
from the
stability
practically realizable.
Moreover, a multistable system, by
making successive tentative adaptations as response

ances, shows learning by steps, or serial adaptation.
to successive disturb-
This
is
a character-
universally observed in living things.

Furthermore, since the second
and subsequent adaptations to a disturbance will certainly alter the
istic
of the system, the recurrence of a disturbance identical to

disturbance will not generally reproduce the behavior of the
system at the first adaptation. This is indeed the dispersion of behavior.
"
In other words, as the system grows older" it becomes "wiser," and
its behavior is appropriate for a wider range of activities than just for
parameter
the
first
coping with the single prevailing disturbance.
CHAPTER
18
CONTROL OF ERROR
In the preceding chapter, we have shown how the principle of ultrastability can make the control system insensitive to accidental errors
and occasional failures of the components by the simple device of changing the characteristics of the system whenever instability occurs. Since
an ultrastable system will automatically seek stability, the control
system,
when
designed, actually embodies unstable fields of behavior as

In other words, during the design of
well as stable fields of behavior.
an ultrastable system, we make no attempt to distinguish stability from

instability, to separate the right fields of behavior from the wrong fields
of behavior.
Errors of behavior are merely treated as a probability, but

In this chapter, we shall approach the reliability
otherwise unspecified.
a complex control system from a different point of view: we shall

specifically introduce errors into the system and ask how the system
of
should be designed so that the system will give satisfactory performance

That is, we wish to know how to control the error.
in spite of the errors.
This subject of control of error is now in its early period of developThe control of error can be discussed for only the most elementary
ment.
operations and the present theory is wholly due to J. von Neumann.

Our discussion in this chapter is then an exposition of Neumann's work.
Its purpose is to serve as an introduction to this very important topic

and to indicate the need for much further investigation.
18.1
It is common knowledge that the
Reliability by Duplication.
reliability of a system can generally be increased by the simple expedient
of duplication.
For instance, if a simple system as shown in Fig. 18. la
when it fails to operate, it merely gives no

output, then to guard against probability of failure, we can duplicate
the system with identical systems and put all in parallel, as shown in
has the characteristic that
Fig. 18.16.
If
the probability of failure of the original system is p, a

and 1, then the probability of failure of each individual
number between
unit in the parallel system is also p.

If the units in the parallel system
are independent of each other, the failure of the parallel system will
1
J.
von Neumann,
"Probabilistic Logics
and the Synthesis
from Unreliable Components," Printed notes

Institute of Technolog}^ Pasadena, Calif., 1952.
268
of Reliable
Organisms
of lectures given at the California
CONTROL OF ERROR
when every unit
occur only
269
The
probability of failure of the

the
number n of duplications,
parallel system
By increasing
we can make this probability very small indeed.
is
thus p n
failure of a
However,
fails.
component
of a control
system generally does
not cause zero output. Instead, the effect is much more damaging:
the system will still give an output, but the wrong output. Then simple
duplication of the system, as described above, will not be effective,
because a wrong output
when mixed with
a correct output results in a
Hence, in such a malfunction of the system, as opposed
wrong output.
to failure, the probability of the malfunction of the parallel
system
X(s)
(a)
is
equal to the probability of the malfunction of the single system and no

of reliability results.
Therefore the problem of control
;
improvement
of error is a
deeper and more
difficult
problem than
it
seems to be at
Nevertheless, as will be seen later, the principle of duplicathe necessity of increasing the number of components, remains
first sight.
tion,
i.e.j
basic.
What
has to be discovered
is
new
organization of these com-
ponents for controlling the error, because the simple parallel organization
of Fig. 18.16 is not effective.
To simplify the analysis, we shall not discuss

18.2 Basic Elements.
the case where the input and the output are continuous functions, but
we choose an elementary component with input and output capable
two discrete values, or 1, {.., the input is either on (activated)
the output is either on (activated) or off. The characteristic
of the element is then determined by the relation between the state of
of only
or
off,
the input and the state of the output. We shall always assume a single
output, but there may be more than one input to the element. There
270
is
between the output and the inputs
also a time lag
in the sense that
the
time interval after the

output is activated only at a certain specified
the properties of a
has
then
inputs are activated. Such an element
relay circuit.
To
describe the characteristic of the element,
we introduce four kinds
of input: excitory input; inhibitory input; per-
>
Excitory input
manently excitory, or live, input; and a kind of

or a grounded
input which is never activated,
inhibitory input
inpui
Uve
)n
ut
I-
^
-i||'
number
by symbols accord-
The element is then represented
circle
meaning that the number of

activated excitory inputs must be greater than
activated inhibitory inputs by * or more, for the outThus Fig. 18.3a represents an element whose output is
there
the
are represented
with inputs on the left side and the

Within the circle
single output on the right.
by a
<
Grounded input
They
ing to Fig. 18.2.
of
put to be on.
is
number
k,
both inputs a and 6 are on. It can be called an ab element.

whose output is on if either a or 6
Figure 18.36 represents an element
Figure 18.3c repreIt can be called an a + b element.
or both are on.
the
if
on
input a is off. It can
only
sents an element whose output is
on only
if
b*\^)
AjL/
(c)
(6)
(a)
FIG. 18.3
1
be called an or element. Incidentally, if we consider the inputs to be the
conditions for a given statement to be true (on) or false (off), then the
three elements of Fig. 18.3 represent the three fundamental operations

In a digital computer using binary numbers, these
of Boolean algebra.
elements are the basic elements for its operation. If memory is neces-
such a device
sary for the computing operation,
can be made from the second element of Fig. 18.3
by feedback
as
shown
in Fig. 18.4.
a is activated, the output
will
Once the input
be on, even if a is later

.clG.
18.4
turned on.
For
having three basic elements to deal with is a

However, all three are really special cases of one funda-
later discussions,
disadvantage.
mental element.
Consider the element represented by Fig.
the Scheffer stroke.
can eliminate them
Since the two live inputs are always excited, we

from the diagram and represent the element as shown
in the right diagram.
neither a nor 6
is
18,5, called
on, or
The output
if
either a or
of the Scheffer stroke is thus
on
Ms
both
on; but the output
is off if
if
CONTROL OF ERROR
a and
6 are on.
The
271
three basic elements of the last paragraph can be
constructed out of the Scheffer stroke as shown in Fig. 18.6. Of course

both the ab element and the a
b element involve two Scheffer-stroke
thus the time lag will be twice as large as for the or 1

element having only one Scheffer stroke. But since our notion of time lag
merely indicates that inputs precede output, the exact magnitude of time
elements in
series;
no consequence.
Operations differing only in time lag will be

all three basic operations can be
Therefore
equivalent.
a
Scheffer-stroke
element.
represented by
single
is of
lag
considered
(a/6)
Scheffer Stroke
Fig. 18.5
ab element
FIG. 18.6
The choice of a single fundamental element is, however, not unique.

There can be other choices besides the Scheffer stroke. But our choice
convenient for the following discussions.
is
control of error in an operation represented
any complex operation
built
We
shall consider first the
by the Scheffer
stroke; then
up from Scheffer strokes can be
similarly
designed.
Method of Multiplexing. Applying the concept of duplication

improving the reliability, we substitute any single input by a bundle
inputs with n individual lines. Thus in the system of the single
18.3.
for
of
Scheffer stroke,
i
1, 2,
and n
<
we would have n lines for the a input, labelled a where

n }n lines for the 6 input, labelled for i = 1, 2,
,n;
i}
&,-
lines in the output bundle.
<
|,
such that
if (1
We
S)n lines
then specify the fraction 5, where

of the output bundle are on or off,
272
the output as a whole
considered to be on or
is
off.
If
dn lines of the
output bundle are on or off, then the output as a whole is considered

to be off or on respectively.
Any intermediate value is considered to be a
?
malfunction.
5 is
thus the fiduciary level
The problem
is
how
to con-
struct the system, using Scheffer-stroke elements, such that the probaof errors
bility of malfunction can be reduced with a specified probability
in the input bundles
and a
malfunction of the
specified probability of
individual Scheffer-stroke elements.
approach to the problem would be to take a line a* from the a

input bundle and a line hi from the b input bundle, and to use these as
first
a bundle,^
n lines
Output
bundle,
lines
62
b bundle,
-bi
lines
*ba
FIG. 18.7
the two inputs to a Scheffer-stroke element. The organization of the

system is thus as shown in Fig. 18.7. It is evident that if almost all
lines of
bundle
both input bundles are on, then almost all lines of the output
be off. If almost all lines of both input bundles are off, then
will
almost all lines of the output bundle will be on. This over-all behavior
seems to be satisfactory. However, a more careful consideration will
show that it is not so. Since, for output of Scheffer stroke to be off, both
inputs have to be on, an error either in the a bundle or in the b bundle
will
be
sufficient to cause
the output
is
an error in the output bundle. Therefore,

off, then the error in output is the sum
supposed to be
errors in the input bundles.
and only one input bundle
if
of
if the output is supposed to be on

then the error in output is the same as
the output is supposed to be on and both
Similarly,
is off,
that in the input bundle. If

input bundles are off, then an error in the output bundle requires a
simultaneous error in both input bundles; and therefore the error level
in the output bundle is less than the error level in the input bundles.
Therefore the error level in the inputs is not maintained at the output.
situations result in a magnification of error, while others result in a
Some
reduction of error.
There
is
then a dispersion of error.
This
is
undesir-
CONTROL OF ERROR
able, because dispersion of error causes the
273
number
of activated lines of
the output bundle to drift to the undetermined region between 5n lines

and (1
S)n lines, and thus increases the chance of malfunction.
To
suppress the dispersion of error, we can introduce a restoring component of the system as follows. We take each line of the output bundle
from the
executive component of Fig. 18.7, executive in the sense of carrythe

out
Scheffer-stroke operation for the over-all system, and split
ing
We then obtain 2n lines. Then we permutate these
it into two lines.
2n
lines so that the order of lines is
successive pairs of lines and using
we then again obtain
element,
now made random.
them
Taking the
as inputs to a Scheffer-stroke
a bundle of n output lines.
lines
/i
This
is
lines
FIG. 18.8
Let there be a$n activated
indicated diagrammatically in Fig. 18.8.

lines in the original bundle.
vated
is
is
clearly ao<*o
= a.
The fraction of
The fraction ai of
the output lines not actithe output lines activated
thus
<4
(18.1)
If oj is the probability of having the original lines activated, then,

provided n is large, the probability of having the transformed lines activated is ai. This is not a restoring component yet. But if we put two
such units in
series,
the probability a 2 of having the final lines activated
is
(18.2)
The
series
Fig. 18.9
system is now a restoring component for the following reason.

shows the relationship between a 2 and
0:2 is equal to a
.
when
ai
or
when
0, -g-Ov/S
1),
2a*
or
1.
Thus
i(-V/5
1)
a,
=
if <*o lies
0.618034
between
and
274
is
<x 2
smaller than
than a
<*<>;
if
<*
lies
between J(\/5
1)
and
1, <*2 is
greater
Therefore the action of the restoring component is to drive

and 1, and
of activating the output towards the limits
the probability
thus to reduce the dispersion of error caused by the executive component.
On the basis of the preceding discussion then, our system for error
7
control consists of an executive
of
component
n individual Scheffer-stroke
of two units
elements, followed by a restoring component composed
elements and a randomScheffer-stroke
n
of
made
each
in
18.8
series,
Fig.
of
element of perfect
Therefore, for each Scheffer-stroke
accuracy, we have to expand the
izing device.
system to a complex system of 3n

Scheffer-stroke elements.
fixed fiduciary level
probabilities
lines
and with given
error in the input
and the Scheffer-stroke
we
ments,
make
of
With any
shall
see
ele-
that we can
the reliability of the over-all
system as high as we please by

n,
creasing
principle is
V5-1
tion.
then,
Basically
still
reliability
in-
the
by duplica-
However, our analysis
in this
section gives us a specific plan of
FIG. 18.9
This
organizing these elements.
of
unreliable
a reliable system out
particular method of synthesizing
elements is called the method of multiplexing by von Neumann.
We shall now compute the
18.4 Error in Executive Component.
error of the multiplexed Scheffer-stroke
section
and show that the error
that the direct source of error

itself,
in either the executive
is
is
system described in the preceding
We
indeed controlled.
observe
first
the individual Scheffer-stroke element
component or
in the restoring component.
Let the probability for each individual element of making a mistake be e.

If we assume that there are r parallel Scheffer-stroke elements and that
the probability for the
they are independent in their operation, then
elements to fail when in the system is still e. Therefore, of the r parallel
on the average r will make mistakes. This number is also
'
elements,
number of mistakes in the n parallel elements. The

making some other number of mistakes is less. In fact,
the most likely

probability of
the problem of determining the probability po(p,c,r) of having p mistakes

in r parallel elements with an individual probability of failure equal to e is
the classical problem of

1
random sampling.
See for instance H. Margenau and G.
and Chemistry,"
p. 422,
For large
r, it is
known 1
M. Murphy, "The Mathematics
D. Van Nostrand Company,
Inc.,
New
that
of Physics
York, 1943.
CONTROL OF ERROR
275
Therefore the probability distribution po(p,e,r) is a normal distribution

mean at er and a mean deviation equal to \/e(l
e)r.
with a
The other source
of error in the executive component is the misfitting

from the input bundles into the individual Scheffer-stroke element. For instance, if a fraction
of the a input bundle of Fig. 18.7
is activated and if a fraction 77 of the b input bundle is activated, we
of lines
would expect a similar fraction
But
of the
output bundle to be inhibited.
the particular o for ith element is activated and bi happens to be

nonactivated, the output of the ith element is still activated, even with
if
no mistake on the part of the element itself. Let f be the fraction of the
output bundle of the executive component that is activated. Then the
number
of
output
lines effectively inhibited is (1
ber of activated lines in the a input bundle
The number
it is r\n.
however, only
(1
total
num-
of effective, or properly fitted, a-input lines
The
tf)n.
difference
The number of ineffective &-input lines

number of effective output lines is (1
is
(1
[77
)w, the
output lines due to misfitting activated a input

of ineffective
lines
ineffective.
f)]n.
Therefore
number
is
is,
is
f)]n
(1
the
number
The
f)n.
n, in the b input bundle
is
of ineffective
(1
the
f)]ft,
due to misfitting activated
output
input is
the
number
of
ineffective
lines
due to
f )]n, and, finally,
(1
output
[T?
nonactivated input lines is the remainder
{i
(i
rt
The number
output
[(1
On
is
[*
(i
f)]
of possible effective
fo
- a -
r)]}*
(2
r)n
combinations of such a classification of
thus 1
f)n]!{[{
(1
the input side, the

with n activated,
lines,
f)]n}!{fo
number
(1
nl
(1
f)]n}I[(2
f)n]l
of
n a-input
n 6-input lines, with yn
activated,
of possible
combinations
)n nonactivated,
is
The number
(1
77)71
of possible combinations of
nonactivated,
is
nl
Therefore the probability pi of having a fraction f of the output actiand T\ of the inputs activated and with perfect
vated with fractions
1
See for instance H. Margenau and G.
M. Murphy,
op.
cit.,
p. 415.
276
operation of the individual Scheffer-stroke elements,
is
;^)
n!
[(1
[(1
{)]!
__
fon] ![(!
i/W
(18.4)
have meaning, none of the four classes

output lines discussed in the preceding paragraph can have less than
zero number.
Whenever that happens, the probability drops to zero.
Clearly, for the enumeration to
of
That
pi
is,
is
zero whenever the following conditions are violated,
-?>0
(1
(i
>0
>o
f)
rt
(18.5)
2-J-i7-f>0
and
We shall now
tion that
by
simplify the expression of Eq. (18.4) under the assump-
When n is
is large.
large, the factorials
their asymptotic value, given
by
(18.6)
we can
can be approximated
formula
v^re-W*
nl
By using Eq.
Stirling's
(18.6)
write the approximate expression for pi as
"
T**
(18.7)
V2irn
where
#
and
^
+f(1
f)
1) log (5
log (1
+ff)
- {log{- (1- 01og(l-
1)
(2
)
+ri,
-lylogiy-
Differentiating $ with respect to f
and
(1?
we have
+ f - 1)
-, f) log (2
(1-^loga -i?)
1) log
(iy
f)
(18.9)
CONTROL OF ERROR
We
find
by using
277
these equations that at f
&,
Aft
-
0.
of
J12/Q
Furthermore, because of the conditions

Therefore the only zero
positive.
of
Eq. (18.5),
of $ is at f
7.
is
Then
always
if
is
very large, the negative exponential in Eq. (18.7) shows that we need
near its zero. But at the zero of 0, where f = 1 - &,
only consider
3f
1(1
Thus near
17(1
17)
fa,
*)
is
^fa
T
(1
f)(l
{(1
n)
approximated by
1
[r
(1
fa)]'
When n is large, a is a relatively slowly varying function of f in comparison to the exponential of Eq. (18.7). Thus we can take the value
of a at the point f = 1
Or
%q.
Therefore, finally, the approximate expression for
large
pi(%w]n) with very
is
^=
' C1
"'
(18.14)
fin
Hence pi
also a normal distribution with respect to f

The mean is at
and
the
mean
deviation
is
fri)n,
V(l Q^(l ?;)n.
(1
We shall make a final modification of the probability pi(,i?,f ;n) by
is
passing to the continuous distribution function ]^F(f;,7?;?i) for large

values of n. If IF(f ;f,ij:n) df is the probability of having the number of
= I/ft,
activated output lines from fn to fn
1 = w(f
1/w), then if
and
this probability is exactly Pi(% } i],r, n )'
Therefore
ir
r-d-fe)
(18,15)
If (f ;f,i?;n) df in general is then the probability of having the fraction of

activated output lines between f and f
df, with specified fractions
77
input bundles activated. The size of bundles is determined
of lines n.
the
number
is a Gaussian distribution with the mean
by
mean
and
the
derivation
1
%q
V?(l ~~ )*?(1 ~~ ^)A- A^ equivalent
and
of the
way
of expressing the result is to write
278
where y denotes a random variable distributed according to the normal

Gaussian distribution, with the mean zero and the mean deviation equal
Then Eq. (18.16) states that f

Gaussian distribution with the mean 1
to unity.
is
distributed according to the
and the mean deviation

Therefore Eqs. (18.15) and (18.16) express the
ri)/n.
Equation (18.16) is, however, often more convenient to use.
fa
same
fact.
We
can
now combine
the two sources of errors and add to the f distri-
bution of Eq. (18.16) the effects of imperfect individual Scheffer-stroke

elements.
We can rewrite Eq. (18.3) in a way similar to Eq. (18.16)
:
Now
in our executive
One
stroke elements.
output.
The
er
+ V(i - > y
e
(18.17)
component there are two

type,
f??
in number,
is
classes of the Scheffer-
supposed to have activated
mistake will decrease the number of activated lines by one.
q mistakes in this type of elements are given

q
The
other type, (1
= fr
+ V(l
)n in number,
is
by Eq.
(18.17) as
-e)fny
(18.18)
will increase
q'
f
is
(1
f)n
supposed to be nonactivated.
the number of activated lines by one.

mistakes in this type of elements are distributed as
mistake
+ VX1 -)(!- Dn y
The
(18.19)
then the additional number of activated output lines due to
errors of the elements themselves.
According to Eqs. (18.18) and (18.19),
we have
q'
2 (|
f)n
+ -0(1 -)(!- i> V -
V*(l
~ *Kn y
(18.20)
The last two terms of the above equation are the difference of two
normally distributed random variables. We shall see presently that the
difference is again a normally distributed random variable.
Consider two normally distributed random variables z\ and
mean zero, and the mean deviations <TI and c- 2 respectively.
the
with
Then
(18.21)
^
}
Or, with Wi(zi)
Fa (22)
denoting the probability distribution function of

the probability distribution function of 2
,
Jo
V2?T
zi t
CONTROL OF ERROR
279
Now if the two random variables

probability of
23 to 2 2
having
are independent of each other, the joint

the range z\ to z\
dzi and 22 in the range
z\ in
dzz is
We now introduce new variables xi
The
By
joint probability in the
#1
22
new
and # 2 defined by
21
22
(#2
l)
variables
is
then
integrating this joint probability with respect to x 2 from
we obtain the
probability W(xi) dxi,
distribution of #1
Therefore
we can
22
21
is
co
to
o
?
the probability
Thus
write
21
By integrating
where W(xi)
22
VT+Î
(18.22)
2/
the joint probability with respect to

21
22
= V*i
+ *i
Xi,
we can show
that
(18.23)
2/
Hence the difference or the sum of two normally distributed and independent random variables is again a normally distributed random
variable with the square of the mean deviation equal to the sum of the
squares of the mean deviations of the original random variables. This
ability to
zero
mean
add or subtract normally distributed random variables with

values is to be expected, since they have equal probabilities
for positive or negative values.
With the
relation of Eq. (18.22),
we can
write the result of Eq. (18.20)
as
f
2e(|
f)n
+ V(l ~
280
Let
= Af
q)/n
(g
be the correction of the fraction f for imperfect
Scheffer- stroke elements; then
we have
(18.24)
We
can
now combine
Eqs. (18.16) and (18.24), and the corrected
'
fraction f of activated output lines
is
r =
.W-V + V-W.W^W
<i8 25)
-
The
last
two terms
of
normally distributed
Eq. (18.25) represent the sum of two independent,

variables.
We can thus use Eq. (18.23).
random
Therefore, finally, writing f in place of
(1
-2e)
(l
',
we have from Eq.
- gmJ
--,
cyx
(18.24)
c;
n
where y
is
a normally distributed random variable with the
to zero and the
(
v
mean
equal
mean
deviation equal to unity. Equation (18.26) specifies the performance of the executive component of our multiplexed
Scheffer-stroke system with the fractions
and
of the inputs and
t?,
,
output activated and with
as the probability of failure of the individual
Scheffer-stroke elements.
18.5
Error of Multiplexed Systems. After having calculated the perof the executive component of our multiplexed Scheffer-stroke
formance
system, we find that the rest of the computation is very easy. Each
unit (Fig. 18.8) of the restoring component is really equivalent to the
executive component. For the first stage of the restoring
component, the
Thus
inputs are the split output lines from the executive component.
instead of two different fractions
Therefore,
if ju is
and
17,
we have the same
the fraction of activated output of the
fraction
first stage,
f.
then,
Similarly if v is the fraction of activated output of the second stage of

the restoring component, then
2. (,
1)
~
(18 28)
.
CONTROL OF ERROR
281
Equations (18.27) and (18.28) have a first term identical with Eq.
The additional terms come from the imperfect elements and
(18.1).
from the
statistical distribution of errors.
With any specified

and n, Eqs. (18.26) to
e,
the
distribution
function
of v, the fraction
compute
,
lines of the
what
??,
clearer
by
of activated
can
make
output
some-
this
reverting to the notation of probability distribution
Thus, for instance, Eq, (18.26)
functions.
We
complete Scheffer-stroke system.
(18.28) enable us to
[(1
is
equivalent to
fc)
I)]
5
-17)
The
probability distribution function of
with respect to f and
of integrating
/*
v
TF^jf^jn), is thus the result
the joint probability of f, /*, and ?
:
Thus
df
exp
tf
W
~
/(I
We shall now show that
2) V(l
M)
(1
6)
under proper conditions we can obtain almost
perfect performance of the multiplexed Scheffer-stroke system
ing n.
that v
i)
>
<
1
5
(18.29)
by increas-
Consider a given fiduciary level 5. Perfect performance requires

< for the nonactivation of output is implied by 5 > 1 6 and
5 for the activation of both inputs; and that v > 1
5 by either
and
large and
)?
>
or>l~5 and
Eqs. (18.26) to (18.28).

r
(i
Then
<
??
so small that terms of order
is
so
in
Let us assume that n
and l/\/n can be neglected
5.
282
Or
?
Now let f = 1 - a, =
Hence
and
77
>
also gives
0(5
z>
).
5;
so that
Similarly, Eq. (18.30) gives v
5,
or f
0(
>
4
<
)for
and
5
8
and
provided
When n
is
17
and
(18.30)
>
Sand
r?
>1-
d.
17
< 5.
< 5.
0(5
for
<
Furthermore, Eq. (18.30)

Therefore perfect reliability
system can be indeed obtained with
of the multiplexed Scheffer-stroke
when n
(18.30) gives
- V)
- ft and a, <
TI
Then Eq.
(2fr
are small.
large but not infinite, the calculation
is
somewhat
tedi-
ous because of the necessity of evaluating the integral of Eq. (18.29).

Although the asymptotic values of the integral can be determined by the
classical
we
methods, we shall not enter into this calculation here. Instead

an example from von Neumann where 5 = 0.07, i.e., activa-
shall cite
tion of at least 93 per cent of the lines of a bundle represents a positive

message; activation of at most 7 per cent of the lines represents a negative
message. He then found that the probability e of malfunction of the
individual Scheffer-stroke elements

trolling the error.
over-all
For
>
must be
less
than 0.0107 for con-
0.0107, the probability of malfunction of the
system cannot be made arbitrarily small by increasing n. For

Neumann gave the numer|- per cent chance of failure, von
0.005, or
ical results in
Table
18.1.
It is seen that for as
many
as 1000 lines in a
bundle, the reliability is rather poor. In fact it is inferior to the original

1 per cent of e.
But a 25-fold increase in n will give extreme reliability.
For systems organized originally in Scheffer-stroke elements, the technique of multiplexing discussed in the preceding section can also be
applied without change. We replace each Scheffer-stroke element in
TABLE
5
Number
of lines,
=
n
o.07
18.1
0.005
Probability of malfunction
1,000
2.7X10- 2
2,000
2.6
3,000
2.5X10- 4
5,000
X
X
10~ 3
10~ 6
10,000
1.6X10- 10
20,000
2.8
25,000
1.2
X
X
10" 19
10~ 23
the original system by the Scheffer-stroke system of 3n elements, each

its executive component and restoring component.
The error in
with
the over-all system can be computed from the error of the individual
CONTROL OF ERROR
283
Scheffer-stroke systems as shown in the previous discussions.
Practically,
a very tedious calculation. However, for the purpose of estimatthe

order of magnitude of n required for specified reliability, we can
ing
consider the entire system to be equivalent to a single Scheffer-stroke
This will
element, and use the result of this over-all reaction directly.
this is
be done in the following section.

18.6 Examples.
To obtain an idea of the magnitude of required bundle size, let us consider a computing machine of 2,500 vacuum tubes.
Each
of the tubes is
supposed to be actuated on the average of once every

specify that the machine should on the average
We
5 microseconds.
run 8 hours before making a single mistake.

number of actuations of a single tube is
3,600
10 6
5.76
During
10
Consider each tube as a Scheffer-stroke element.
this period, the
Then the
specified
probability of malfunction, considering each tube as an independent unit,

9
is l/(5.76 X 10 ).
However, there are 2,500 interconnected tubes in
the system.
mistake in any one among the 2,500 tubes will mean a

Therefore, considering each tube as one unit
mistake of the machine.
within the system, the specified probability of malfunction should be

9 =
7 X 10~ u
only 1/2,500 of the above value, or 1/(2,500 X 5.76 X 10)
.
We see then that
the final probability of malfunction is the same as that

obtained by considering the whole system of 2,500 tubes as a single
This possibility greatly simplifies the calculaScheffer-stroke element.
number of lines n in the multiplexed system.

that the fiduciary level d and the probability e of failure
of the tubes are the same as specified in Table 18.1, then the specified
tion of required
If
we assume
= 14,000,
probability of malfunction obtained above will require n
according to the table. Therefore, in order to make the machine as
reliable as specified, it will
times.
This would
be necessary to multiplex the system 14,000

of every single tube in the
mean the replacement
machine by a system of 3n = 3 X 14,000 = 42,000 tubes. The original

machine of 2,500 tubes now becomes a giant of 105,000,000 tubes. This
is
clearly not practicable.
Now take a second example, a plausible quantitative picture of the

The number of neurons
organization of the human nervous system.
But considering the presence of
involved is usually given as 10 10
.
synaptic end bulbs and other possible autonomous subunits, this number
It ought to be a few hundred times larger.
Let
is certainly too low.
us take the
number
of basic
elements to be 10 13
The neurons can be
actuated up to a maximum of about 200 times a second. But the average

rate of actuations must be a good deal less; say 10 actuations per second.
We shall further assume that a mistake in our nervous system is serious
284
and should not happen in the time interval of a human life span. Take
the error-free interval to be 10,000 years.
During this interval, the
13
elements is
total number of actuations in the system of 10
10 1S
10,000
Thus the probability
31,536,000
10
of malfunction should
l/(3.2
10
25
)
3.2
3.2
10 25
be
10- 26
Again, assume that the basic nervous elements have the properties specified in Table 18.1 then an extrapolation from that table gives n = 28,000.
;
However, our calculation needs a correction:
if
the
human nervous
indeed multiplexed 28,000 times, the number of basic elements

system
13
as assumed above; the number
in the nonmultiplexed system is not 10
is
Then the probability

28,000).
3
factor
the
X 28,000. The corby
21
Table 18.1
of malfunction is now 2.7 X 10~
by a
should be reduced
factor of l/(3
of malfunction should be increased
rected
probability
then gives n
22,000.
Further iteration
will
not change this value
appreciably.
that, while our method of controlling the error
quite conceivably applicable to the microcomponents
of nervous systems, it is nevertheless impractical for engineering control
These examples show
by multiplexing
is
systems with the present technologies. One obvious direction of future

development is the reduction in bulk and power requirement of the
The transistor is a great improvement over the vacuum tube
elements.
from this point of view. Hence the method of multiplexing may yet
become practical in the future. Another direction of investigation
would be a deeper analysis of the process of controlling error. The organization of the executive
component and the restoring component
basic system for the Scheffer stroke

zation.
We
of
our
only one possible organiare fortunate in that such a crude attempt already is
is,
after
all,
successful in demonstrating the possibility of increasing the reliability.

There are probably other plans of organization of the duplicated com-
ponents which will produce the same degree of reliability with a smaller
number of components. In other words, only a beginning has been made
in the technique of error control in automatic systems.
For control
engineers, there
is
as yet no applicable solution to the problem.
INDEX
a~ l element, 270
ab element, 270
a
&
Clementson, G.
element, 270
C., 178
Adamson, T.
Control computer, 159, 192

analog, 192
Addition, normally distributed independent random variables, 278
198
criteria,
Control design with specified

additional parameter, 212
Ansoff, H. L, 97
first-order systems, 201
R., 253
second-order systems, 209

stability problem, 200
Assembly average, 113

Asynchronous excitation, 165
Asynchronous quenching, 165
Auto correlation function, 233
Control system, airplane rotation, 50

computer, 159, 192
Autonomous system, 142

of random function, 113
Control
systems,
continuously
Correlation function, 114

random function with
Ballistic perturbation theory,
178
its
derivative,
115
Becker, L., 50
W.
turbulent flow
field,
118
Cox, D. W., 110
R., 76
damping, 24
Bienayme-Chebyshev inequality, 124
Critical
Blasingame, G. C., 30
Bliss, G. A., 185
Crocco, L., 94
Cross-correlation function, 233
Curfman, H. J., 31
Blivas, D., 30
Block diagram, 14
construction, 35
Bode, H. W.,
17, 49,
sensing
and measuring, 214
Average
Bennett,
198
criteria,
application to turbojet, 204
Artillery rocket (see Rocket)
W.
192
digital,
Control
Adjoint functions, 185

A-c servomeehanism, 70
Ashby,
30
C.,
Close-cycle control, 35
Computer in control system, 159, 192
Cutoff point of relay frequency response,
138
239
Bode diagram, 17, 49

Boksenbom, A. S., 53,
Bushaw, D. W., 151
135, 198
Damping
Davies,
ratio,
I.
L.,
24
251
Decibel, 17
criteria, linear
Design
Callander, A., 94
Canonical path, 151
system with
con-
stant coefficients, 37
linear
Carrier, 70
system with random inputs,
133
Carriere, P., 168
Detecting
Cauchy's theorem, 38
Center of trajectory, 145
Differentiator, 19, 29
Characteristic function of probability dis-
Dispersion of behavior, 266

Drag coefficient CD, 180
Dirac
tribution, 121
Chattering of relay servomeehanism, 150

Chebyshev inequality, 124
Drag
filter,
function, 117
coefficient
Draper, C.
285
247
S.,
KD, 170
216
129j.
286
Gardiner, R. A., 31
Gauss, C. F., 124
Drenick, R., 178

Dugundji, J., 33
Gauss inequality for unimodal distribu-
Dutilh, J. R., 138
Dynamic
effects,
optimalizing control,
tion, 125
Gaussian distribution, 114
222
Gear
train, backlash, 141
Gross, G. L., 168
Engineering approximation, 6
Grounded
Error, 14
Guillemin, E. A., 50
input, 270
mean-square, 231
multiplexed system, 274
steady-state, of sampling servomecha-
nism, 88
W.
R., 41, 42
Himmal,
Excitory input, 270

Executive component, multiplexed system, 271
M.
Feder,
S.,
J.,
52
Hartree, D., 94
Error control, 268
Evans,
P. H., 161
Hammond,
Hanny,
S. C.,
30
Homeostat, 256
Homeostatic mechanism, 253
Hood, R., 53, 67, 198

Howarth, L., 118
Hunting loss, optimalizing
control, 217,
219
67
Feedback circuit (link), 36

Feedback servomechanism, 34-69
combined open-cycle and close-cycle, 51
Hunting period, optimalizing control, 217

Hunting zone, optimalizing control, 217,
219
Hydrodynamic analogy, root
design criteria, 37
general, 36
locus, 46
multiple-loop, 50-69
noninteracting, 53
Infinite-memory
simple, 37
Inhibitory input, 270

Input, 12
Fett, G. H., 158
Fiduciary
Finite
272
level,
memory
filter,
Fluctuation of random function, 113

Fliigge-Lotz, L, 145
Focus of trajectory, 145

circuit (link), 36
Forward
Integrator, 18, 29
Jump phenomenon,
Frequency of large deviations, 126-127

Frequency demultiplication, 164
Frequency entrainment, 164
Frequency response, 16
von Karman,
pulse excitation, 30
first-order system, 15
Knuth, E.
Th., 118
Klotter, K., 145
by
relay, 136
L.,
178
Kochenburger, R.
cutoff point, 138
J.,
138
Kochenburger diagram, 138
relay in oscillating control servos, 74

Bliss,
Kolmogoroff, A., 231
185
Lag network,
Gain crossover, 49
163
Kalb, R. M., 76
Kang, C. L., 158
determination, 29
Gain, 16, 19, 20
250
Input linear group, optimalizing control,

223
250
Fundamental formula,
filter,
restricted,
simple, 20
Lagging
filter,
243
22
INDEX
Laning, H.,
N arc, 146
N system, 146
216
Jr.,
287
Laplace transform, 7
application, to linear equations, 8
to linear equations with time lag, 97
dictionary, 9
inversion formula, 7
von Neumann, J., 268

Newton, R. R., 168
Node of trajectory, 145
Noise, 111
white, 118
Large deviations, 123

frequency of, 126-127
Noise
filter,
231
Lead network, 21
finite-memory, 250
Li,
Y. T., 216
Liepmann, H. W., 130
heavy noise with weak
Lift coefficient CL, 180
infinite-memory, 250
linear system with variable coefficient,
Lift coefficient
KJ
J}
Wiener-Kolmogoroff theory, 236
orbital stability, 147
Linear switching of relay servomechanism, 145

1
first-order, 12
lag,
Noninteraction conditions, 58
Nonlinear device, frequency insensitive,
Nonlinear system,
linearization, 80
Normal
second-order, 24
time
Noninteracting control, 53
response equations, 62
140
composition from elements, 31

1, 12-135
constant-coefficient,
stationary
241
252
170
Limit cycle, 146
Linear system,
signal,
random
1, 5,
136-167, 198-230
182
flight path,
Novik, D., 135
inputs, 111
Nyquist, H.,
94
18,
38
Nyquist diagram,
variable-coefficient, 3, 168-197
18,
38
acceleration effect, 4
Live input, 270

Loeb, J, M., 80
Octave, 17
Open-cycle control, 34
Optimalizing control, 214
Lozier, J. C., 77
MacColl, L. A., 70
Mach number,
181
deviation of
Mean-square
MiUiken,
W.
error,
F.,
random
function, 113
231
Mode of probability distribution,

Moment coefficient CM, 181
Moment coefficient KM, 170
J. R.,
219
output linear group, 223

peak-holding, 221
rate sensing, 216
124
51
Multiple-mode operation, 158, 201

Multiplexed system, 271
executive component, 271
malfunction probability, 282

reliability, 274
restoring component, 273
Multistable system, 264
hunting zone limit, 221

input linear group, 223
interference effects, 220, 228
30
Minorsky, N., 94, 163

Mixer, 35
Moore,
effects,
loss, 217,
hunting period, 217

hunting zone, 217, 219
Marble, F. E., 110
Mean
222
dynamic
hunting
sinusoidal testing input, 218

stability criteria,
228
Optimum operating point, 216

Optimum switching function, relay servomechanism, 150
Optimum
switching
line,
relay
servo-
mechanism, 152
Osborn, R. M., 49
Oscillating control servomechanism, 73
with built-in
general, 80
oscillation,
77
"ENGINEERING CYBERNETICS
288
Output, 12
due to initial conditions, 12
due to input, 12
linear group, optimalizing con-
Output
trol,
223
Randon
function, variance, 113
Random
switching function, power spec-
trum, 122
Rankin, R. A., 168
Rea,
J. B.,
52
Relay servomechanisms, 136

chattering, 150
P
P
arc,
linear switching, 145
146
nonlinear feedback, 160
system, 146
function, 150
Paley, R. E. A. C., 239

Parametric damping, 166
optimum switching
optimum switching
Parametric excitation, 166

Pendulum with sinusoidal force, 166
stability criteria, 138
by
duplication, 268
multiplexed system, 274
Phase margin, 49
Phase plane, 143
Phase-plane representation, second-order
linear system, 143
Phase space, 143

Phillips, R. S., 235
optimum
Reliability,
152
line,
filter
Response, linear system, to stationary

random inputs, 127
steady-state, to sinusoidal forcing function, 11
unit impulse, 11
design, 235
Physical realizability, 239
Restoring component, multiplexed system, 273
Poisson's distribution, 123
Reynolds number, 181
Porter, A., 94
Rice, S. 0., 126
Power spectrum, 115
Rocket,
Phillips
direct calculation, 118
pulse sequence, 118
turbulent flow
filter,
field,
118
242
malfunction, multiplexed system, 282

of stability, 259
distribution,
function, 121
first,
112
Gaussian (normal), 114
mode, 124
moment, 113
normalization, 114
equations of motion, 178

guidance condition, 189
guidance system, 190
Probability, of large deviations, 123
Probability
168
stability, 172
random-switching function, 122

Prediction
artillery,
linearized trajectory equations, 171
characteristic
perturbation coefficients,
195
perturbation equations, 183

cutoff condition, 188
power
range deviation, 186
Rocket functions, 174

Rocket motor, combustion
with feed system, 100
intrinsic,
94
instability,
97
Poisson's, 123
combustion-lag index
gas transit time, 96
second, 112
servo-stabilization
skewness, 114
calculation,
of
n,
96
combustion
in,
100
Root-locus, hydrodynamic analogy, 46
Root-locus method, 42
Ragazzini, J. R., 248, 250

function, 111
Random
Rosser, J. B., 168
Routh, E.
J.,
38
average, 113
fluctuation, 113
mean
deviation, 113
power spectrum, 118

stationary, 112
Sampling servomechanism, 83
comparison with continuously operating servomechanism, 91
INDEX
Sampling
servomechanism,
Nyquist
criterion for, 87
Stibitz-Shannon theory, 85
transfer function, 89
M.
design, 49
rate gyro, 29
sampling servomechanism, 89
Scheffer stroke, 270
Seamans, R.
W.
C.,
transcendental, 32
translating to higher frequency, 72
in sinusoidal gust, 32
30
R., 32
wing
Transfer-function matrix, engine, 54
Self-excitation, hard, 140, 163

soft, 139,
system, 58
163
Servomechanism, with external noise, 243

feedback (see Feedback servomechanism)
with internal noise, 244
relay (see Relay servomechanisms)
sampling (see Sampling servomecha-
nism)
with saturation constraint, 245
Servo stabilization, airplane wing, 133
combustion in rocket motor, 100
Shames, H., 30
Shannon, C. E.,
Shull, J. R., 216
average, 113
lag,
accelerometer, 29
99
Satche diagram, 97-110

Saturation constraint, 245
Sears,
Time
Time
combustion, 94
Transducer, 193
Transfer function, 12
steady-state error, 88
Satche,
289
84, 85,
H.
S., 94, 95,
178
Turbojet control, with afterburning, 66

design with temperature criteria, 204
design criteria, 204
Turbopropeller control, 63
Ultrastable system, 253
switching boundary, 255

terminal field, 261
Unimodal
239
Sign error root-modulus error

system, 161
Tsien,
distribution, 124
Unit impulse, 11
(SERME)
Uttley, A. M., 161
Singular point, phase plane, 145

Slattery, T. G., 251
Variance of random function, 113
Small nonlinearity, 162

Spring dashpot system, 12
Stability criteria, linear system, 38
with time
lag,
with variable
108
coefficient, 176
optimalizing control, 228

relay servomechanism, 138
sampling servomechanisms, 87
G. R., 84, 85
Stibitz,
Stoker, J.
J.,
163
Subtraction, normally distributed inde-
pendent random variables, 278

Switching boundary, ultrastable system,
255
Weinberg,
L.,
J. C.,
White
noise, 118
Wiener, N., 231, 239

Wiener-Khintchine relations, 117
Wiener-Kolmogoroff theory, noise

ing, 236
Wiener-Paley
Wing,
field,
criterion,
132
in sinusoidal gust, transfer function, 32

in turbulent air, 130
P. M., 251
261
distribution functions, 262
filter-
239
m intermittent wake,
Woodward,
Terminal
50
161
West,
Zadeh, L. A., 248, 250
r|
16638
5!

EngineeringCybernetics (英文版) 钱学森着

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

EngineeringCybernetics (英文版) 钱学森着

Uploaded by

Copyright:

Available Formats

KANSAS

Daniel and Florence Guggenheim Jet Propulsion Center

McGRAW-HILL BOOK COMPANY, INC.

THE MAPLE PEBSS COMPANY, YOKK, PA.

Essai sur la philosophic des sciences/' 1845, Paris)

between governments with the use

of force greatly accelerated the

the total absence of considerations of energy,

which are so important in other natural sciences.

cations in designing controlled or guided systems.

neering science, while servomechanisms engineering is an engineering

used in engineering practice into a discipline and thus to exhibit the

ponents of the system

glance at the contents of this book makes

the actual implementation of the theory

the justification of this separation of the theory from the

the results of investigations in fluid mechanics in their

say the least.

approach to old problems and gives new, unexpected

At the present stage

a very real advantage in trying to

and guidance engineering, there

grasp the full potentialities of this

new science by a comprehensive survey

that the mathematical

the more true

requisite for the study.

ordinary differential equations

the other hand, no rigorous and elegant

to the practicing electronics specialist, the treatment here must

During the course

of writing these chapters, the

author had the benefit

of many conversations with his colleagues at the California Institute of

often led to sudden clarification of an obscure point.

Linear Systems of Constant Coefficients

Response to Unit Impulse

INPUT, OUTPUT, AND TRANSFER FUNCTION

Designing the Transfer Function

Laplace Transform and Inversion Formula

5.1. Control of a Single-variable

CHAPTEE 6. ALTERNATING-CURRENT SERVOMECHANISMS AND OSCILLATING CONTROL SEKVOMECHANISMS

Oscillating Control Servomechanisrns

Continuously Operating with Sampling Servomecha-

LINEAR SYSTEMS WITH TIME LAG

Time Lag in Combustion

Complete Stability with Feedback Servo

LINEAR SYSTEMS WITH STATIONARY RANDOM INPUTS

Large Deviations from the

Exceeding a Specified Value

Approximate Frequency Response of a Relay

Optimum Switching Function

Nonlinear Feedback Relay Servomechanism

Parametric Excitation and

LINEAR SYSTEM WITH VARIABLE COEFFICIENTS

Rocket during Burning

CONTROL DESIGN BY PERTURBATION THEORY

Equations of Motion of a Rocket

Appendix: Calculation of Perturbation Coefficients

CONTROL DESIGN WITH SPECIFIED CRITERIA

General Theory for First-order Systems

Application to Turbojet Controls

Control Problem with Differential Equation as Auxiliary Condition