Vous êtes sur la page 1sur 126

FOURIER

TRANSFORMS
AND CONVOLUTIONS

FOR THE EXPERIMENTALIST

by

R.C. JENNISON
Ph.D., B.Sc., F.R.A.S.
MANCHESTER UNIVERSITY DEPARTMENT OF PHYSICS
NUFFIELD RADIO ASTRONOMY LABORATORIES
JODRELL BANK EXPERIMENTAL STATION

PERGAMON PRESS
OXFORD LONDON NEW YORK PARIS

1961
PERGAMON PRESS LTD
Headington Hill Hall, Oxford
4 & 5 Fitzroy Square, London W.l

PERGAMON PRESS INC.


122 East 55th Street, New York 22, N. Y.
Statler Center 640, 900 Wilshire Boulevard
Los Angeles 17, California

PERGAMON PRESS S.A.R.L.


24 Rue des E:coles, Paris V.

PERGAMON PRES S G.m.b.H.


Kaiserstrasse 75, Frankfurt am Main

Copyright 1961
Pergamon Press Ltd.

Library of Congress Card Number 60-13824

Printed in the German Democratic Republic


VEB Offizin Andersen Nexo in Leipzig
CONTENTS

PREFACE

Page
I INTRODUCTION 1
What is a Fourier Transform? 2
Typical Applications of the Transform 3
Formal Statement of the Fourier Transform 5
Convolutions 6
Notations 6
The Meaning of Negative Frequencies 7
Basic Formulae 8

II THE ELEMENTARY TRANSFORM 10

The Spatial Relationships 10


Reciprocity 12
Time and Frequency 13
The Delta Function 15

III EXTENDED FUNCTIONS: THE SUPERPOSITION


OR SUMMATION PROPERTY 19
The Superposition of Elementary Functions 19
The Rectangular Function 21
Complex Distributions 24
A List of Common Fourier Pairs 26

IV THE DIRECT APPLICATION OF FOURIER TRANSFORMS 31


Frequency and Time Relationships in Simple Circuits 31
One Dimensional Aerial Systems 32
Optical Slits and Gratings 34
An Optical Example - The Rayleigh Refractometer 35
Examples in Acoustics 37
Two Dimensional Fourier Transforms 37
The Numerical Evaluation of Fourier Transforms 41
VI CONTENTS
Page
V LIMITATIONS, PRODUCTS AND CONVOLUTIONS 45
The Effect of Interposing Limits 45
Convolutions 47
Physical Interpretation of Convolutions 49
Some Examples of Convolutions 54
The Solution of Fourier Transforms by the Application
of the Convolution Theorem 56
The Isosceles Triangle 56
The Doublet Pulse 57
Convolutions involving the Sine Integral Si (x) 60
Two Dimensional Convolutions 65

VI THE DIFFERENTIATION OF FOURIER TRANSFORMS 68


Differentiation and Repeated Differentiation 69
The Fourier Transform of a Step Function or Straight Edge 70
The Convolution - Differentiation Relationship 74
The Differential Operator h(x) 75
The Integral Operator 79

VII THE AUTO-CORRELATION FUNCTION


AND THE TRANSFER FUNCTION OF A SYSTEM LINEAR
IN INTENSITY 82
The Auto-correlation Function 82
The Transfer Function of a System Linear in Intensity 86
Examples on Chapter VII: 90
1. The Telescope 90
2_ Stellar Interferometers 95
(i) The Michelson Stellar Interferometer 95
(ii) The Switched Interferometer 99
(iii) The Post-Detector Intensity Interferometer 102
Note on Fourier Synthesis of Apertures 105

ApPENDIX I ANALOGUE COMPUTERS OF FOURIER


TRANSFORMS 106
(i) The Diffraction Computer 106
(ii) A Mechahical Computer 107
(iii) A Coherent Electronic Computer 108
(iv) A Coherent Electronic Computer with an Incoherent Source 109
ApPENDIX II III
J1 X sin X
Tables of--, Si (x),
X X
INDEX 116
PREFACE

THE reasoning applied in the solution of a large number of physical


problems involves the use, often unconsciously, of the principles of
the Fourier transform. A pictorial solution in which the primary
dimensions may be simply computed, together with an appreciation of
how the finer details are likely to modify the result, is often all that
is required in the initial stages of an investigation. The rigorous
solution may follow and may often be expressed entirely in mathe-
matical form with complete precision but without necessarily in-
dicating the form of solution or the behaviour of the system to the
reader or student whose approach is not that of the mathematician.
It is the aim of this book to make available to the experimentalist
a guide to the principles and practical uses of the Fourier trans-
formation. It is hoped that it will help to bridge the gap between the
more abstract account of a purely mathematical approach and the
rule of thumb calculation and intuition of the practical worker.
The monograph springs from a lecture course which I have given
in recent years and for which I have drawn upon a number of
sources. In particular I was most fortunate to acquire in 1951 a set of
notes compiled by the late Dr. 1. C. Browne from a series of lectures
given by Mr. J.A.Ratcliffe of the Cavendish Laboratory. I most
gratefully acknowledge my debt to Mr. Ratcliffe for this valuable
introduction to the subject.
I am indebted to Dr. H.J.J.Braddick, at whose instigation the
book was compiled, and to many friends and colleagues for their
,helpful discussions and assistance.
R.C.JENNISON
CHAPTER I

INTRODUCTION

IN THE majority of books which deal with the Fourier Transform the
transform is derived by first setting up the Fourier series, and then
showing how we may extend the principle so that it goes over to the
integral form of the Fourier transform when the interval, or funda-
mental period, of the harmonic analysis is extended to infinity. In
this form it is possible to obtain the spectral function corresponding
to isolated transients of any description, subject only to .minor
restrictions which need not concern us at present.
This approach is excellent and rigorous but it is not necessarily
easy to assimilate if the reader's approach is that of the non-mathe-
matical experimentalist. Here it seems that an experimentalist's own
tactics must be employed. The mathematician provides the theory,
we may follow it and check it if we feel so inclined or we may accept
it as a theory to be put directly to the test and, if proven adequate
and accurate, to be used thereafter as a tool without necessarily
questioning the rigorous derivation of its formulation.
This is the approach that will be applied in this book. It is not to
teach the derivation of the Fourier transform nor to prove its ele-
gant versatility as a mathematical operation, but to help the reader to
grasp the simple application of the results of the theory in those
practical spheres where they are most frequently encountered. For
this reason we jump straight into the formal mathematical state-
ment of the transform in the early stages of this chapter and later
break it down, examine it piece by piece, apply it ~nd conjure with
it as a beautiful workshop tool.
2 FOURIER TRANSFORMS

What is a Fourier Transform1


To the question "what is a Fourier transform~" we may most
simply reply: "A method whereby we may obtain the variation of a
quantity as a spectral function (e.g. plotted against frequency) from
the variation of the quantity as a function of period (e.g. plotted
against time)."
The transform holds equally in the reverse direction so that we
may restate the previous paragraph, with equal validity, when we
interchange the phrases "spectral function" and "function of
period".
The frequency and period referred to above need not apply only
in the dimension of time (frequency in cycles per second and time in
seconds), they can also be used in less familiar circumstances. Thus
the period may be in spatial measure whilst the spectrum is angular.
It may cause some concern that the analysis of a simple function,
such as a single rectangular pulse, gives a result (the function sin xix
in the case of a rectangle) which appears very different from the line
spectrum that would be obtained by the application of Fourier
series (lines at zero, /0' 3/0 , 5/0 , etc. with amplitude decreasing line-
arlywith the order of the harmonic). The answer to this is to be
found in the fact that in the use of the series we assume implicitly
that the function continues to perform in the same manner, ad
infinitum, to either side of the period used in the analysis, i.e. we
assume an infinite series of identical patterns and hence a fundamental
component wave to which the harmonics are related. In the case of the
transform we extend the fundamental component to cover all time,
so that, in the case quoted, we see the behaviour of the square wave
as only one pulse in all time, never to be repeated; thus no assump-
tions are made as to its past or future behaviour. In this respect the
transform is more general than the Fourier series as in most practical
problems we are concerned with events which are contained within a
finite time scale though they may perform many cycles within this
period.
To illustrate the example quoted of a rectangular pulse which may
or may not form part of a longer train of similar waves we may refer
INTRODUCTION 3

to Fig. 7() in Chapter III. According to the width of the pulses


relative to the spaces between them, we obtain one or more spectral
lines within each of the maxima of the envelope shown dotted on the
right. When we have such a long time in the spaces to either side of
the pulse that no other pulses ever actually occur, the envelope be-
comes so packed with lines that it becomes just the transform of a
single pulse shown in Fig. 7 (d).
The reader may care to consider this problem in reverse and notice
that when we have a completely packed distribution of spectral
lines we have the white light condition which, by confining the
angular distribution, forms a very useful tool for setting up many
optical interferometers.

Typical Applications of the Transform


In many instances where we are concerned with the frequency or
spectral response of a system to a pulse or complex distribution of
amplitude with time, the Fourier transform relates the distribution
of the signal expressed in terms of time with its distribution in terms
of frequency. In another case, where we have a coherent spatial
distribution across an aperture, the relationship exists between the
reception or transmission pattern of the system in terms of the sine
of the angle of projection and the distribution of the field along the
aperture of the system.
In certain electronic apparatus, where the circuits are linear and
introduce no distortion, the Fourier transform is directly applicable
to the amplitude and phase of the signals everywhere within the
circuit when it is desired to represent a variable of frequency as a
variable of time, or vice versa. Examples of linear circuits in elec-
tronics are simple static combinations of resistance, inductance and
capacity, valves or transistors acting as distortionless class A ampli-
fiers, frequency changers and other devices in which the output
amplitude is linearly proportional to the input amplitude and the
phase of the input is linearly transferred to the output terminals.
The amplitude and phase distribution in the Fraunhofer diffrac-
tion pattern of an optical system is related by the Fourier transform
4 FOURIER TRANSFORMS

to the amplitude and phase patterns of the associated spatial distri-


bution across the diffracting object in linear, wavelength measure.
When an optical system is illuminated with incoherent light it is
again amenable to analysis by the Fourier transform, though in this
case the relationship holds between a different set of variables, the
power spectrum, or Fraunhofer intensity distribution, and the trans-
fer function of the optical system to incoherent illumination.
Many electronic circuits are not linear to amplitude, though they
may be linear to intensity. If, for example, a linear amplifier feeds a
thermocouple or square law detector, the ouptut is no longer linear
to the amplitude of the input signal though it may be directly pro-
portional to the power or intensity of that signal. The circuit then
becomes the analogue of an optical system, for in all optical systems
the ultimate detector, whether it be the human eye, the photographic
plate, photocell or thermocouple, responds to the intensity of the
incident radiation. If the circuit is fed with an incoherent noise
signal it possesses a Fourier transform relationship between the
output power spectrum and the autocorrelation function or transfer
function* relating the behaviour of its output to changes in the
input occurring as a function of time.
The Fourier transform frequently occurs in the course of analysis
when the variables may appear to have no direct physical signi-
ficance. For example, we may recognize the convolution integral in
the analysis and find it advantageous to change the variable by
means of the associated Fourier transform (See Ohapter V). Such
examples often occur when dealing with probability distributions
and must be treated on their individual merits. In this monograph
we shall confine the use of the transform to those aspects of optics,
electronics and acoustics where its material significance is readily
appreciated, bearing in mind that there are countless other problems
which may M tackled with this most versatile tool.
* The Transfer Function gives the relationship between the input and output
signals of a "black box" (electronic, optical, acoustical, etc.) without detailing its
contents. It is usually complex, e. g. for a passive filter for which the output varies
in amplitude and phase over the spectrum. It may include terms of higher power,
e. g. a non-linear system introducing di&tortion (by accident) or to measure r. m. s.
levels, etc. (by design).
INTRODUCTION 5

Formal Statement of the Fourier Transform


We must at this stage formally define the operation* which we
hope later to manipulate and understand. It is worthwhile commit-
ting to memory the first formulae, marked (1) and (2) as these will be
used frequently; the others need only be referred to if required. The
use of p and q below is only intended for generality and these
symbols will not be found in later chapters.
The Fourier transform f (p) of a function of q, F (q), may be stated as

f(p) = f F (q) e2nipq dq, (1)


q=-oo
and likewise
F(q) = f f(p) e-2nipq dp (2)
p=-oo

This almost symmetrical relationship is an outcome of the Fourier


integral**, which in itself is purely the extension of the summation of
Fourier series to cover terms of zero frequency and the analysis of
nonperiodic events.
The transform relationships

IF(q) = F.T. of f(p)


f(p) = F.T. of F(q)
l
are also frequently referred to, from their symmetry, as Fourier pairs
and oile is termed the mate of the other. Either may carry the
negative exponent provided that it is used consistently.
* The Fourier transform is applicable to linear systems and is the outcome
of the reciprocal relationship which can exist between a pair of conjugate variables.
** The Fourier integral
f(p) = f
q=-oo
[e 2ni pq f
p=-oo
1
f(p) e- 2ni pq dp dq

can be seen to invite the transformation relationship by the symmetry it receives


on substituting F (q). An alternative expression, for the variable, p, considered
about the epoch, g, is

f{p) = f f e2ni q(P-Y)g(g)dgdq


f/=-OO q=-oo
6 FOURIER TRANSFORMS

The transform may be quoted in a variety of equivalent forms,


either trigonometrically or in exponential form, as a complex func-
tion or as a real function. For example, a very useful real form of the
transform may be written*
~

F(q) = f f(p) cos (2n pq + IP)dp


Very often the transform is encountered with the frequency com-
ponent expressed in angular measure, OJ, but this leads to asymmetry
in the mate and the simplest form arises when frequency in terms
of a pure number per unit interval of the spectrum is used as the
mate of the epoch of the event.

Convolutions
The convolution will not be dealt with at such length as the
Fourier transform though its operation is so bound up with that of
the transform that it forces its presence upon us at an early stage.
A convolution may be simply considered as the operation whereby
a structure under observation is smeared or spread out by the response
or resolution of an instrument or mathematical operation.

Notations
Throughout this monograph we shall confine the transformation to
two systems. In one the time variable function f (t) will be mated with
the frequency variable denoted -by the corresponding capital letter
F (f). Where the quantities are specifically voltage, current or ad-
mittancethenotation will bee(t), i (t) andy(t) as the mates ofE (f),I(f)
and Y (f). In the other system of spatial coordinates the corresponding
quantities will be f(x), where xis measured in wavelengths,mated with
F(O) where 0 is measured in terms of sin O. The corresponding nota-
tions for intensity will be i(x) and J(O) whilst s(x), g(x), p(x), h(x),

* The phase angle iP indicates the relative position of the component waves
on the axil! of 'P. It is not an angle in the complex plane (c. f. page 17).
INTRODUCTION 7

g(t) andh(t) may occasionally be used as the respective mates of S (0),


G(O), P(O), H(O), G(f) and H(f) where it is necessary to introduce
additional parameters.
The Fourier transform relationship between one function and its
mate will frequently be shown diagrammatically by the double hal-
arrow
f(x) ~ F(O)

The Meaning of Negative Frequencies


Throughout this monograph we shall use the concept of a spectrum
containing both positive and negative frequencies. The existence of
negative frequencies may cause some confusion to those who have
not previously encountered their use. It may be agreed that a posi-
tive spectrum is all that is required and that one cannot have the
. existence of an event in negative time. The term frequency refers to
the rate of repetition of an event or the number of events per. unit
time. It.can therefore have a negative polarity if either the UI~it of
time is negative or the number of events is negative.
If the flywheel of an engine makes one hundred revolutions in a
second we may express the frequency of the passage of a fiduciary
mark on the wheel as lOO per second, and the mark will perform
100 cycles per second. This engine may be coupled to another. which
also rotates at 100 cycles per second and the pair may drive a larger
machine. But what if the second engine, though rotating at 100 cycles
per second, revolved in the opposite direction 1 The two would
differ by 200 cycles per second and sparks would fly when they were
coupled. The simplest way out of the anomoly is to refer to the fre-
quency of one or the other flywheel as negative -100 cycles per
second, which then correctly expresses their relative frequencies. It
will be apparent that the distinction of a negative spectrum is a
necessary outcome of the description of the system in terms of ro-
tating vectors. The vectors, in the example given, are of unit length
as we have not discussed the behaviour of any property (such as fuel
consumption) which may be represented as a function of the fre-
quency by adjusting the length of the vector.
8 FOURIER TRANSFORMS

A similar situation may be demonstrated in an interferometer


system. Interferometers used in radio astronomy usually depend
upon the transit of the radiating objects over the instrument to
generate the fringe pattern. It is possible to introduce into one of the
two limbs of an interferometer a phase rotating network which may
be driven to slow down or speed up the fringes*. When the interfero-
meter system is adjusted so that the phaser rotates at the same
speed as the natural fringe pattern, it produces a constant deflection
or a fringe of zero frequency. The connections may now be changed
so that the phase rotation is performed in the opposite order
(0 0-270-180-90-00 instead of 00-90-18~270-00). The final fringe
speed will be, not zero, but double the original speed. This will corre-
spond to the motion of stars from west to east rather than east to
west, and the angular frequency of the phaser, or of such hypothe-
tical retrograde stars, would be the same as in the stationary case
but of negative sign.
As an alternative to the use of rotating vectors with a negative
frequency spectrum we can express the vectors in terms of their pro-
jections on to two orthogonal planes and thus derive relationships
for the frequency in terms of cos () in one plane and sin () in the
other. In this form the notation is commensurate with the usual
derivation of Fourier series, but the use of the exponential to denote
the rotating vectors is a neater and more compact formulation. In
both cases two sets of data are required to define the spectral func-
tion, in one it is the combination of the spectra in positive. and nega-
tive frequencies, whilst in the other it is the combination of the odd
and even sine and cosine projections of the spectrum.

Basic Formulae
The basic formulae governing the relationships between the pairs
of functions participating in the Fourier transform are set out in
shorthand notation on page 9. Most of these operations will be dealt
with more fully in later chapters.
See for example: JENNISON, R.C., Monthly Notices Roy. Astron. Soc., 118,
p. 276, 1958.
INTRODUCTION

Simple transformation f(x) ~ F(O)


Negative spectra f (- x) ~ F( -0)
Complex conjugates f* (x) ~ F* ( - 0)
Superposition Ef(x) + g(x) +... ~ EF(O) + G(O) + ...
Elementary convolutionf(x - Xl) ~ F(O) e- 2 "iOx,
(A function off set by
constant Xl transforms
to product of mate with
cisoid generated by Xl.)
Elementary convolution f(x) e2"iO,x ~ F(O - 0l)
Convolution integral ft(x - Xl) g(x) dx ~ F(O) . G(O)

Convolution integral f (x) . g (x) f


~ F (0 - ( 1 ) G (()) dO
d
Differentiation dxf(x) ~2niOF(O)

d
Differentiation - 2ni xf(x) ~ de . F(O)
CHAPTER II

THE ELEMENTARY TRANSFORM

NEWCOMERS to the field of the Fourier integral are likely to be per-


fectly at home with the behaviour of optical systems illuminated by
point sources or by spectral lines. In this chapter, we shall show how
Sotoo

i
l
i
I
.... -
~-~~~~=jjj~~~~~~~~ ~:~l~~~~~-~~~L~~~~=~~~~~}
, I
Wavefronq

A 0 0' B

FIG.l

the basic patterns derived from these concepts can be treated to


yield their corresponding transforms. Towards the end of the chapter
we shall assign an integral nature to the elementary points, and in
the following chapter we shall see how the superposition of these
elementary integrals may be controlled to deal with any normal
distribution.
Consider a point source of light, S, situated at infinity, Fig. 1.
AB is a line in an infinite plane normal to the line joining it to S. In
a practical case S might be a star and AB a tangent to the surface of
the earth. If an observer is stationed at the point 0 anywhere in the
line AB he will record that the source emits radiation and that this
THE ELEMENTARY TRANSFORM 11

radiation is equally intense at all points along AB. If he could make


simultaneous observations at two points in the plane, 0 and 0' he
could also confirm that the phase of the radiation that he received
was identical at these points, this is readily apparent "from the fact
that the whole of AB lies in the plane of the wavefront from the
source, S, at infinity.
Let us now consider the situation if S subtends an angle which is
displaced from the normal to the line AB, Fig. 2. Once again the
intensity recorded by the observer, 0, will be the same at all points
)( Sol <Xl
'.\

\\

A o 0' B

FIG. 2

in the plane, but if he were able to make simultaneous observations


between the two points 0 and 0' he would establish that there was
a change of phase introduced by the additional path difference CO'.
If we change the separation b~tween 0 and 0' whilst keeping the
star fixed off the zenith, the relative phase observed between 0 and
0' will exhibit a cyclic variation and hence a periodicity which is a
function* of the angle () subtended between the star and the normal,
and the wavelength of the radiation. If we measure the separation
00' in terms of this wavelength, then we need only concern ourselves
with the angle () and the distance 00'. The greater the angle () the
more wavelengths will lie in the path difference CO, and the faster
will be the periodicity of the phase variations as 00' is increased.
* Strictly a function of sin 0 which is itself, of course, a function of O.
12 FOURIER TRANSFORMS

If we kept 00' fixed but varied the angle () over a small range, we
should also observe a-periodicity in the variation of phase of the two
signals received; the amplitude (or the intensity) received at either
o or 0' would, as before, be constant.

Reciprocity
Let us now consider the situation reversed. If we take the line AB
emitting light at constant intensity and equal in phase over the
whole of its length, by application of the Reciprocity Theorem -
or simply by retracing the path lengths used above, we may say the
line AB would illuminate a point at infinity which lies on a line
normal to AB.
If the line AB emits light of constant intensity but with the phase
continuously varying along its length in a periodic manner, then we
may retrace the paths illustrated in Fig. 2 and derive a narrow beam
of radiation aimed at the point S situated at infinity at an angle
away from the normal which is such as to enable the point S to
receive in identical phase the radiation from each point in the line
AB. Only at S will the radiation from every point in the line AB add
in phase.* If the phase rotation along AB were more rapid, then S
wouid be further from the normal through 0; if it were slower,
S would be nearer the normal. We may invert Fig. 2 and it is then
apparent that an infinite aperture in which the phase varied contin-
uously, as it did along the original line AB, would illuminate the
new baseline at only one point. It is apparent that there is a recipro-
cal relati~nship between the distribution of the field along a line and
distribution of the far field in terms of the sine of the angle of the
radiation from the line.
Let us now summarise the position:
If a point S emitting radiation is located at an angle () from the
normal to a line AB then we may equally describe the situation by
plotting the variation of phase along the line AB of a vector whose
It is assumed that the system is confined to the plane of the paper and that
the line AB is opaque, otherwise there will be an ambiguity due to the image of S
traced through the origin below AB.
THE ELEMENTARY TRANSFORM 13

length is constant and equal to the amplitude of the signal received.


That is to say we may transform the distribution as a function of
angle into a corresponding amplitude and phase distribution along
a.line which is calibrated in terms of wavelengths. In the simple case
that we have chosen, that of a point source, the amplitude is con-
stant at all points in the line.
By the principle of reciprocity we have shown that we can trace
back the conditions above to prove that a simple distribution of
phase and amplitude (constant again in this instance) along a line of
infinite length will project a beam of radiation at a single discrete
angle to that line. We have transformed a linear distribution into an
elementary far field or Fraunhofer pattern and we have shown that
they possess a symmetrical relationship whereby we may interchange
the distributions.
Let us not be disturbed by the fact that we have referred to a
distribution of amplitude when in fact the distribution in these cases
is a constant. It is evident that doubling the amplitude of one
pattern doubles that of the other, i. e. the amplitude is linearly
transferred. We shall shortly see that in any general case when there
is not just one infinitely narrow beam, as in the above examples, but
a multiplicity or finite distribution of beams, the amplitude distri-
bution will also vary along the line AB.
Before we consider these more complex cases, let us see how the
same argument can be applied in a somewhat different frame of
reference, that which includes time and frequency.

Time and Frequency

In the previous section we were concerned with the relationship


between a function of angle and a corresponding distance measured
in wavelengths, and though we were dealing with waves which were
a function of time, we purposely chqse our system to exclude time
as a variable. Let us take the case where we exclude distance and
consider the behaviour of the time function. We shall see that we
have available a similar transformation between time and fre-
14 FOURIER TRANSFORMS

quency, * and, lest this seem too sudden a jump, we will first quote a
transformation of this kind in such terms that it may seem almost
obvious; a monochromatic spectral line may be depicted on a plot of
amplitude against frequency as a single line at a definite point on the
abscissa, it may equally well be depicted in a plot of amplitude
against time as a simple periodic wave extending, if the line is in-
finitely thin, for all time.
The case equivalent to Fig. 1 in the previous section is that of a
spectral line at zero frequency. This yields a constant amplitude
signal lasting for all time - that is a simple static field.
The case equivalent to Fig. 2 is that already quoted, a spectral
line of other than zero frequency yielding a wave function whose
period is reciprocally related to the difference in frequency of the
spectral line from zero.
The inverse cases are a little more difficult to visualize. By the
reciprocity theorem we should expect an infinitely fine pulse which
appears at time t = 0 to yield a spectrum of all frequencies stretch-
ing to infinity. As these frequencies are all generated only when
t = 0 it is apparent that they must all be in phase at that instant.
This condition is approxi~ated in practice by a sudden electrical
discharge. The waves from such a discharge may be detected over
a very broad band of frequencies.
The reciprocal case to that of the spectral line of finite frequency
is that of an infinitely fine pulse generated at a time t', other than
t = o. This again yields a spectrum of waves of every frequency but
in this case they will be in phase at time t' and hence will all have
reached varying degrees of phase at the origin where t = o. The
integration of the contributions from these waves at any point on

.* The similarity is apparent from the fact that in the previous cases the rate of
rotation of the phase vector was simply a function of angle of incidence, in the
present case the geometrical paths are considered equal and the rate of rotation
of the phase vector is simply a function of frequency. In both cases the amplitude
of the vector is determined by that of the monochromatic generator or point
source. The conditions are somewhat analogous to the amplitude and phase of
standing waves in the case of the spatial transform and travelling waves in the
case of the time transform. A fully rigorous treatment of coherence embraces both
the functions which are treated separately here.
THE ELEMENTARY TRANSFORM 15

the time axis will be found to yield a resultant of zero at all points
except that where t = t'.
We shall shortly proceed to show how the summation of elemen-
tary transforms can be employed to illustrate the phenomena asso-
ciated with more complex distributions, but first, in order to put
our formulations on a more rigorous footing and to retain the inte-
gral nature of the transforms as stated in the introduction, we shall
define our elementary points in terms of the delta function.

The Delta Function


The delta function is a very useful tool in the analysis of trans-
formation problems, for it describes a function in the form of a rudi-
mentary point or irresolvable pulse.
The delta function t5(O - 01} is a function whose amplitude is zero
at all 0 except at 0 = 01 , where its amplitude is such that the total
area of the function is unity, i. e.

(3)

We may obtain the Fourier transform of a delta function by sub-


stituting t5 (0 - 01) for F (O) in the basic transform formula
co

f(x} = ! F(O} e-2nix8dO

Thus
co

f (x) = ! t5 (0 - 0I) e- Znix 8d 0

and since
is oJ?1y very slowly varying in the immediate vicinity of 01 , we may
take it outside the integral as the constant e-Znix8..
Thus
co

f(x} = e-ZniX8,! t5(O - OI}dO


16 FOURIER TRANSFORMS

In which the integral is unity by definition. Hence


I (x) = e- 2n ;xo, (4)

Equation (4) is of the form

(5)

where Q(x) is an amplitude function = 1, and ;(x) = xO describes


the phase of the unit vector.
The phase function; (x) = X 01> depends upon the system of re-
ference, such as the angle of inclination of a collimation plane or the
zero on a time axis.
The function I (x) may be represented in three equivalent ways. It
may be directly represented as a unit vector rotating in phase space*
with the phase function; (x) = X 01 , whilst moving forward with x;
in this representation the locus of the tip of the unit vector describes
a helix whose pitch is constant and is determined by the constant 01"
If 01 is negative the helix has a left hand "thread" and if positive
the "thread" is right handed. In the helical representation the nota-
tioncis (xO), which is really a corruption of (cos + i sin) (xO) may
be used for I (x).
An equivalent representation of I (x) is the projection of the helix
on to the orthogonal y and z planes of the phase space, or complex
plane, and this yields a separate sine and cosine wave in the ortho-
gonal systems, these sine and cosine waves are the components of
the usual trigonometric expression for a complex exponential.
The third method of describing I(x) is simply two separate dia-
grams for the amplitude and phase functions. The amplitude func-
tion is a straight line of constant height and the phase function is
a straight line tilted at an angle proportional to 01, since the value
of 01 directly determines the rate of change of phase. See for
example, the spectral and time responses of the transmission line in
Fig. 20.
* The complex plane. This conception of phase space simply means that the
+.
y, z plane is used as a system of angular measure, 00 being in the direction of z
90 0 at y +,180 0 at z-and 2700 at y-.
TJI,E ELEMENTARY TRANSFORM 17

The transform of a point is a flat spectrl,lm extending to infinity in


either direction, but it will be seen from the preceding remarks that
this description is incomplete and assumes that the system of re-
ferencelieswithin the point; that is that Oliszero andhencee2ni .,9 , = 1
otherwise we should also stipulate that the phase of the spectral

Cis. (1 (Locus of Hp of vectorl -sin 1(/


Source function cos. 10 ~ inverse
- 516'1 + side elellOllon of CIS pIOnot CIS
Side eleVOTion PIon End elevatIOn

:F +--
Oblique elevation

-=t. t- +
{J 6'

I
t Y
/" "..,
i
"
",
L
',.---'"
/'
':
:
Y

-l -I
Y

~ A\tV C1 W Ai -I

A ~ ~ 'W o~
*WWt4A (9r 'W -I

-l '-1
~N-tJ
:W ,lrAf\
-

,
I
i ~ W1W c!1 -l Y

FIG. 3

components advances linearly with the order of the component and


the displacement of the point from the system of reference. In
general the spectrum is that of a vector of constant amplitude which
moves forward whilst rotating in the complex plane at a rate de-
pendent upon the departure of the point from the axis of reference.
In Fig. 3 the spectrum of a point is shown both in the form of
cis (x 0) and in the form of cos (x 0) and sin (x 0) for various values of O.
The extreme left hand column shows the generating function S (0)
with the position of the generating point displaced by the appro-
18 FOURIER TRANSFORMS

priate angle () from the chain-dotted collimating or reference line


corresponding to () = O. The next four columns depict the trans-
formation into f(x) in the form of the cisoid cis xO, whilst the last
two columns give the orthogonal components corresponding to the
even function, cos x(), and the odd function, sin x(), which together
may be used to portray f(x) in the form cos x() + i sin x(). It will
be seen that the trigonometrical form may be obtained from the
ciscoidal projections in side elevation and plan.
CHAPT'E R III

EXTENDED FUNCTIONS: THE SUPERPOSITION


OR SUMMATION PROPERTY

SO FAR we have only considered the transformation of a single point


or spectral line, though the point itself was given an integral nature
by the introduction of the delta function.
We are rarely fortunate enough to be concerned only with the
elementary transform of a point and we must now examine the
addition of other points into the system, eventually arriving at the
case where the points may be replaced by a continuous, though
possibly limited, function of any desired shape or size.
If we examine the transformation relationship

I(x) = j F(fJ)e2" ix8 dfJ


it is evident that we can substitute for F (fJ)

and we are at liberty to form separate integrals with each of these.


Each of these integrals may in turn be transformed independently
to give the mates

And therefore we may write the general relationship


~

Il(X) + 12 (X) +/a(x) + ... =jFl(fJ)e2"iX8dfJ +


~ ~

+ /F2(fJ)e2"iX8dfJ + /Fa(fJ)e 2" iX8 dfJ + ...


t-:)
CO

Saurce funCtion Locu s of tip of vector Projections


-(j- -(j Oblique elevation Side elevation Plan End elevation Odd Even
4-V
,'qV'\."x C-_v
F
I Components Cis ~xdFCx Son ~kr- COS~~
I t-z -I -I
~

-~~~ cl2__y
o

i
~ tos~
I I:j
z SIN o
i Resultant
CIS
t-z Y
q
!;d
4- Y t' .z .....
I Components CiS
~A-;x~ C)-y son~~ COS~~
l?l
!;d

i t-z ty -I -I 1-3
!;d
I :..

A i1X
0
0 4-Y
x ~-., zr:J:l
~
I : I
i Resultant CIS
:
0 SIN
tos \ I

V
I:j

I t-z ty o
y.
4
!;d
~
I
COI'f1')OI'IenfS Cis ~~x
~_: rY (3-y sm~~ cos~WI r:J:l

I t-z ty -I

I
~~
0 0
I
~"\7i
t-y
I
I
Resultant
z
CIS
ty t-z
@Y
- .., I
SIN

h looWi
FIG. 4
EXTENDED FUNCTIONS 21
and conversely,

00 00

+ !f2(x)e-z" ix8 dx +!f3(x)e- 2 iX8 dx + ...


1<

This facile property of summation is one of the most elegant


features of the Fourier transform. We may take as many points, or
distributions of points, as we wish, transform them independently
and add the transforms as we added their generating functions.
In this way any source distribution may be converted into its
spectrum.
Figure 4 shows a few simple examples of summation, * sufficient to
illustrate the effect of asymmetry, but the representation of the
spectrum of a limited continuous source becomes somewhat con-
fusing in the form of cis (x 0) unless an actual model is used, and the
form cos (xO) + i sin (xO) is more manageable in two dimensions.
An even simpler model is available from a plot of the modulus of the
cisoidal vector and a separate plot of its phase angle in the complex
plane at each point along the axis of O. When the phase plot is con-
fined to real values, either 0 or 1800 , we may omit the phase function
and show a single plot with corresponding polarity. We shall shortly
refer again to this method of portrayal as it will be adopted through-
out the remainder of the book.

The Rectangular Function


The cis representation shows how the elementary components are
actually present in the case, for example, of a rectangular source
situated astride the axis of reference, but that their effect in the
imaginary plane is destroyed by mutual cancellation. The resultant
cosine component, however, reappears after touching the first zero,
but its reappearance is in the direction of z- and hence the nett phase
of the transform undergoes it discontinuity at the zeros and reverses
instantaneously. This is shown in Fig. 5.
* See also Figure 15, page 43.
22 FOURIER TRANSFORMS

As an exercise and for reference we may transform this function


mathematically. We have the basic transformation formula

I(x) = f F(O) e2 ,,; x6 dO

Inve~ plan

Side elevation

Pion

Oblique elevatian

/\ ResuHant amplitude

Resultant phase

FIG. 5

and we wish to find the transform in terms of x of a function of 0


which is everywhere zero except between, say, 8/2 and - (8/2)
where it has an amplitude A which is constant.
It is apparent that although the integral in the right-hand side of
the equation extends from minus infinity to infinity, 0 will only
contribute in the range 8/2 to - (8/2) and we may therefore replace
the limit signs with these values if we give F (0) the value of the
constant A.
EXTENDED FUNCTIONS 23
Hence:
f)
2 f)

I(x) = fA e2n ; x6 dO = A [_1_. e2niX6]2


2n'~x f)
f) -2
2

= _~ (en,Xf) _ e-"ixf))
2n~x

Now from a standard trigonometric relationship


2 sinri
e'" - e- r = --~c:-.-

and we may therefore write:


A (2Sinn'ixe.
f( X) = --.- . i) = - -A SIn
. ( - nxo-
Q)
2n'~x ~ n'X

= A e [Sinn'xeJ
n'xe

The final expression is of the form f (x) = sin X/X where X =nx e
but it is multiplied by the constant A e. This informs us that the
maximum ordinate of the resultant amplitude curve shown in
Fig. 5 will have the value A e and is therefore proportional to the
total area of the original rectangle. The horizontal scale of the dia-
gram is determined by the rate of change with X of the sine term in
the numerator of the right-hand side; this, in turn is directly depen-
dent upon the only other variable within the argument of the sine
term, that is the width of the original rectangle, e. As e increases it
is evident that the width of the sin X/X function decreases whilst its
form remains the same. From a simple argument based upon the
conservation of energy it is now apparent why the height of the
patterns must undergo a corresponding increase with e as well as
increasing with the amplitude A.
The positions of the zeros on the sin X/X pattern may be obtained
from the values of X required to reduce the whole expression to zero.
When X is zero both the numerator and denominator are zero but we
24 FOURIER TRANSFORMS

may show that as X tends to zero the whole expression sin X/X tends
to unity. When X =:re, 2:re, 3:re the sin X term is zero whilst the
denominator is finite and in these circumstances the whole expression
is zero. Thus sin X/X = 0 when X = n:re except where n = 0, when
sin X/X = 1. Replacing X by its true value X = :rex 8 we see that the
zeros occur when :rex 8 = n:re, that is when
123
x=---
8'8'8'
When the source is a symmetrical function in its own frame of
reference, such as a rectangle, but is situated off the axis of reference
.of the system in which it is to be transformed, we have a similar
situation to that occurring when the source was a point, or pair of
points, off the axis. The sine component returns at the partial expense -
of the cosine component, and the final transform E cis (xO), corres-
ponding to the integrated components of cis (x 0), may be repre-
sented as the usual transform of a rectangle twisted in phase space
(the complex plane) about the x axis at a pitch dependent upon the
shift of its centre of radiation from the axis of reference. To illustrate
the meaning of this in practice we may consider the effect upon a
stellar interferometer. The rectangular source corresponds to a slit and
if this is moved across the field of view the effect is merely to cause
the output to exhibit a moving fringe pattern, whose period is depen-
dent on the spacing of the interferometer apertures and the rate of
motion of the source, and whose intensity is determined by the mo-
dulus of the transform of the source at that spacing. It also shows
that fringes may alternatively be produced by keeping the source
fixed off the collimation plane and extending the baseline of the
instrument, that is the separation of the interferometer apertures.
In the latter case the period of the fringes will be a variable dependent
upon the spacing of the apertures.

Complex Distributions
Suppose now that we have a complex distribution in which both
the amplitude and the phase vary over the pattern. Each point in
the pattern will transform into a corresponding frequency compo-
EXTENDED FUNCTIONS 25
nent bearing an amplitude proportional to that of the generating
point and a phase relative to the other frequency components which
is in the same relation as that of the generating point to the other
points in the original pattern. We may use closely packed delta
functions to represent a function of variable amplitude for we have
B

1lli
ill
+
.Eees ID
~

Amplitude
--
~
i
0
\
\,
O 00 (1' .Esin 01
phase

fM
Amplitude
phase
ill
+90 00 +90
--
~
.Eees

.Esin
f
ID
t
0
01
\

Amplitude
phase
ili
-90. O +900
--
....--
.Eees

.Esin
11

W
FIG. 6

already seen that though the amplitude of a delta function tends to


infinity, it is restricted by asigning a finite width to the function. We
may, if we wish, replace the delta function by a very similar tool, a
unit impulse, and multiply this by a factor chosen to fit the amplitude
of the required function at the point concerned.
Figure 6 illustrates these points. The transformation is shown in
terms of the separate sin and cos integrals (or sums in this case).
It could also be shown in the form of a single amplitude function and
a phase function, and in the first two cases above the amplitude
26 FOURIER TRANSFORMS

function will merely follow the}; cos curve whilst the phase function
is zero. The reader may satisfy himself that in the case of the third
asymmetrical transform of Fig. 6 the amplitude function will be
proportionate to the length or modulus of the vector in the cis
representation and the phase function will follow the angle between
that vector and the z axis. Alternatively both may be obtained from
the}; sin and}; cos forms by vector addition. In this instance the
behaviour of the resultant vector may be pictured by considering
the top of the vector free to run along a rail through the origin,
whilst the lower end traces a sine curve in the horizontal plane at
height B. The modulus of the resultant vector in this case appears to
exhibit a component of twice the frequency of the sine and cosine
projections but, as the phase function is also vibrating to either side
of the vertical axis, only the phase of each alternate peak is identical
and the true frequency remains the same.

A Short List of Common Fourier Pairs


We may now list in Fig. 7 a few more complicated transforms and for
simplicity in illustration we shall depict them in the form of separate
plots of the amplitude and phase functions, except in such cases
where the function is wholly contained in the real plane, in which
case the phase can only be 00 or 1800 , and the exhibition of a positive
or negative polarity may be used in lieu of the phase function.
Let us now examine how the various transformations in Fig. 7
may be constructed from the elementary delta function, we shall
refer to each subheading in Fig. 7 in turn.
(a) This is the direct transform of the delta function. If we give
the line on the left a finite height then the continuous spec-
trum in the transform on the right will increase in height pro-
portionally.
(b) A cosine wave extending from + to -
00 is the transform
00

of a symmetrical pair of delta functions each spaced astride


the datum by a distance proportional to the reciprocal of the
length of the cosine wave.
EXTENDED FUNCTIONS 27

.._-------_....
-00 00
(0) .-------~--------+
o o

(d)
__---.EEL __
o

. '_
.,
J:LJlilJl.illULJ.
(I)
-(0 -0- 0 ~d.... co

(g)

___o._1_ :--_-_--L---:::
_ .. "'Phase
.<.:;....-:Amp,i'Ude
(h)
~/O------:hase
~~mplilude
---.~ ---?--O----L-
--
.'

"
~o--- ~~A~'_
...Phase

(k)
J:L_.o__R

FIG. 7
28 FOURIER TRANSFORMS

(c) In this case the function is a sine wave extending to in the00

real plane. As this function is odd, i. e. not symmetrical, it


transforms to two elementary lines spaced astride the datum
or collimating point and bearing opposite pha,ses. We have
already seen that a cosine wave, which is an even or symmetrical
function, transforms into two lines having the same phase. In
both cases the spacing of the lines to either side of the centre
is given by the reciprocal of the spacing between the crests of
the sine or cosine curve.
(d) This is one of the most frequently encountered transformations,
that corresponding to a uniformly illuminated slit or a rectan-
gular pulse. The transform follows a curve of the form* sin X/X.
This may be built up by the summation of the spectra of
adjacent delta functions spreading over the whole rectangular
pattern (see Fig. 5).
(e) A fine ruled grating or a "comb" filter transforms into a
similar series of beams or time signals. This is simply a case
of direct transformation of spaced delta functions multiplied
by an appropriate constant amplitude function.
(f) This case is similar to the previous example but the lines have
a finite width. The rectangular pattern of each line is equi-
valent to that of example (d) and hence we see a modulation
envelope of the form sin X/X. This may be seen clearly from the
product or convolution theorem (Chapter V).
(g) A finite wave train, contained in a rectangular modulation
envelope transforms into two functions of the form sin X/X
corresponding to the envelope, spaced apart by a distance
inversely proportional to the distance between the crests (i. e.
the wavelength) of the enclosed wave function. Looked at
from the viewpoint of delta functions the wave component is
a pure cosine vibration similar to that referred to in note (b)
and it transforms into two delta functions spaced astride the
datum at a distance determined by its wavelength. The square

* Often referred to as the "sine function", sin xix is tabled in Appendix II.
EXTENDED FUNCTIONS 29

(a)

Isosceles
(b) triongle

rm-2~~.
10
~ Phase .".
(e) "-------------2"
Allernative
o representation
in imaginary
plane
--~----~----~--

(d)

(e)
o-d~ ----
(I)
0/1 --
...--

FIG. 8
30 FOURIER TRANSFORMS

envelope produces a similar effect to that in (d) and hence,


by convolution, we obtain the complete transfurmation.
(h) This is the elementary case previously encountered of a line or
delta function offset from the datum transforming into a cisoid-
al vibration, that is a function of constant amplitude but
continuously varying phase.
(i) The delta function in (h) convolved with a sin x/x function trans-.
forms into a rectangle over which the phase continuously ro-
tates.
(j) This is the same as case (i) but the function is to the opposite
side of the datum, giving a phase function sloping in the
opposite direction. The addition of this example (j) with the
previous example (i) yields the function in (g). This is because
two contra-rotating cisoids, as seen in an earlier diagram,
cancel out in the imaginary plane to yield a cosine function in
real plane. (See Fig. 4, centre diagram.)
(k) By interchanging the rectangular functions with the sin X/X
functions in (g) we obtain the cosine transform modulated by
a sin X/X function.
Figure 8 on page 29 shows a few more transformations, the
derivations of which will be explained in later chapters.
(a) A Gaussian error function e-"'" transforms to another
Gaussian e-"'/' .
(b) An isosceles triangle transforms to a sin 2 X/X 2 function.
(c) The transform of a straight edge. This completely asymme-
trical function transforms to an amplitude function which
is wholly contained in the imaginary plane and which has a
discontinuity at the datum where the phase reverses whilst
still remaining imaginary.
(d) The harmonic spectrum from a full wave rectified cosine wave.
(e) The transform of a doublet pulse is a spectrum of the form
sin2X/X. (See Chapters V and VI.)
(f) The transform of the triangular doublet illustrated is dis-
cussed in Chapter VI.
CHAPTER IV

THE DIRECT APPLICATION


OF FOURIER TRANSFORMS

FOURIER transforms may be used to investigate, synthesise, break


down or predict variol!S systems and distributions in many branches
of science for which the 'present treatment up to this point is quite
sufficient for the required result to be obtained.
Before continuing in later chapters with more varied and power-
ful manipulations and applications of the Fourier transform we
shall quote a few simple examples of the use of the transform in
analysis.

Circuit under
examination

(0)

(b)

FIG. 9

Figure 9 (a) shows a simple block diagram of the apparatus on


an electronic test bench. The apparatus on the left produces a pulse,
this is passed through a circuit whose frequency response (ampli-
tude and phase) is known and the output is viewed on the oscillo-
scope. We wish to know how the pulse will appear on the cathode
ray tube, i. e. how its distribution with time is affected by a circuit
whose frequency response only is known. If we were engaged in the
32 FOURIER TRANSFORMS

design of a certain piece of apparatus we could dispense with the


test apparatus shown in Fig. 9 (a) provided that we could predict
the response on the oscilloscope. This prediction can be made by
using two transformations, one to transform the initial pulse into a
frequency distribution and a second transformation to transform
back the product of this distribution with the amplifier frequency
response into a time distribution corresponding to the picture on
the cathode ray tube. We shall see later that it is possible to per-
form the same analysis more simply by the use of one transformation
but this must await the later discussion of convolutions.
In the present case let us assume that the circuit under examina-
tion is just a simple tuned rejector circuit of the type shown in
Fig. 9(b). In Fig. lO(a) are shown the transformations giving rise to
the resulting picture when the applied pulse is just an impulse or
spike. The result is just that of the L.C. circuit ringing to the im-
pressed impulse.
In Fig. lO (b) we see the effect of the same tuned circuit on a
simple square wave input. The distortion of the pulse is again large-
ly due to the ringing effect of the tuned circuit. The principle of
this operation may obviously be applied to much more complicated
networks or pulse distributions. It may also be applied to the case
of a simple pendulum. If the pendulum is pushed aside by a gust of
air then we may consider it fed from a high impedance and we get
similar responses, and there is some response right down to zero
frequency; for a long gust we therefore get a small general displace-
ment. In the case of a short gust, it is easy to see how the inertia of
the pendulum can carry it on over the first part of the curve. If the
pendulum is pushed aside rigidly it corresponds to switching in and
then removing an impulse from an associated low impedance sop-roe,
whilst after the passage of the impulse the impedance is again high.
Thus in these circumstances the rise time is forced.
The Fourier transform may be directly applied to compute the
radiation pattern of an aerial system or the Fraunhofer diffraction
pattern of an optical system. As we have, up to the present, only
considered one-dimensional Fourier transforms, we shall confine our
attentions first to one-dimensional aerials.
DIRECT APPLICATION 33

Gil) I
o~------------~i
----........
(f)
Y(1l

o
~ 0 I

-90"

91(1)

(a)

g(~t
.
G(Ilf\
...... --- I

~
Y(I)
o 10 I

-90"

~ -- ......-
G(I)Y(I) 90"
L Amplitude
.,
Phase
0
(b)

- 270"
FIG. 10
FOURIER TRANSFORMS

Suppose that we have a long line of aerial elements such as simple


dipoles. If the line is very long we may considerit as a one dimension-
al aperture. We may connect together the various elements in a
large variety of ways and by introducing extra lengths of feeder we
may control their relative phase. By the use of lossy cable we may
also control the relative amplitude of the various elements. It will
be seen that we can build up various distributions of the field in the
aperture according to the way we make these connections. In
every case we may compute the far-field pattern of the aerial system
(and hence the polar diagram on conversion from sin 0 to 0) by
simply taking the Fourier transform of the apertlire distribution.
For example, if all the elements are connected in phase and fed
with equal strength, the aperture distribution will be rectangular,
and the radiation pattern, in amplitude and phase, will be given by
a sin xix function as in Fig. 7 (d). The power polar diagram will
correspond to the square of this function and will follow a curve of
the form sin 2 X/X2.
If the aerial ts connected so that all the elements on one side of
the centre are exactly 1800 out of phase with those on the other side
of the centre then the aperture distribution will be in the form of the
doublet in Fig. 8(e), this will give an amplitude radiation pattern
as shown in that diagram corresponding to a function of the form
sin 2 xix.
By keepIng all the elements in phase but by introducing progres-
sively greater loss to the elements away from the centre it is possible
to produce triangular or Gaussian aperture functions. The Gaussian
function, Fig. 8 (a), has the advantage that the resulting radiation
pattern is also Gaussian. It does not suffer from the presence of
confusing side lobes but this advantage is paid for by the fact that
the main lobe is wider than that of a uniformly illuminated array.
Precisely the same reasoning as that applied to a one-dimensional
aerial system also applies to a one-dimensional optical system, for
the only difference is one of wavelength. For a corI'esponding dis-
tribution of the radiation pattern the optical aperture may there-
fore be much smaller, for the critical dimension is the size of the
aperture in terms of wavelengths. The most frequently encountered
DIRECT APPLICATION 35
one-dimensional optical systems are the single slit, Young'l:! double
slits, and the diffraction grating. These will be recognized in Fig. 7 (d),
Fig. 7 (k) and Fig. 7 (f) respectively, in which the left-hand diagrams
give the aperture functions whilst the right-hand diagrams show
the Fraunhofer (amplitude) diffraction patterns given by the three
systems when they are coherently illuminated with parallel light.

An Optical Example: The Rayleigh Refractometer

As an example of an optical system we may. taKe the case of


the Rayleigh refractometer: This is rendered effectively one-
dimensional by the use of a cylindrical lens as the eyepiece for the
telescope.
The usual arrangement of the Rayleigh refractometer, starting
from the light source, consists of a filter to render the light mono-
chromatic and a lens to focus the light on to an entrance slit. On the
far side of the entrance slit a collimator renders the light from the
entrance slit parallel and projects it upon a pair of parallel slits in an
otherwise opaque screen. In line with each of these two slits there
is a chamber containing in one case the gas whose refractive index
is to be measured and in the other chamber there is a control gas
or a vacuum. Beyond the chambers the light passing through them
is again united in a telescope with a standard objective but a cylin-
drical eyepiece. Interference fringes are seen as a result of the path
difference of the light from the two slits occurring across the field of
view. If the refractive index of the gas in the two chambers is not
identical then the fringes will be displaced as a result of the addition-
al optical path difference introduced by the difference of velocity
of the light in the two media. It is a simple matter to compute the
refractive index of a gas if its pressure may be varied in one of the
chambers.*
Neglecting for a moment the effe,ct of the entrance slit and assu-
ming that it is narrow, the pair of parallel slits are illuminated with

* See, for example, JENKINS and WHITE, Fundamental8 01 PhY8ical Optic8,


McGraw-Hill (1937), or any standard book on optics.
36 FOURIER TRANSFORMS

plane monochromatic light and as they are viewed at infinity through


the telescope we see i~ the field of view an intensity distribution
which is simply the square of the Fourier transform of the slit
illumination, i. e. the amplitude distribution is given directly by the
transform depicted in Fig. 7 (k). The introduction of a difference in
the refractive index of the gas in one of the chambers will change
the phase of the light from one of the apertures and hence move the
fringe pattern (derived from Fig.7b) relative to the stationary
envelope (derived from Fig. 7d). If the two slits are of identical
width but further apart there will be a greater number of fringes
within the envelope, but the evelope will remain the same. If the
slits are made wider but of constant s~paration the envelope will be
narrower whilst the fringe width is constant so that the number of
fringes within the envelope will be fewer. The situation may be
described in terms of the superimposed individual diffraction patterns
of the two slits crossed by their mutual interference pattern but it
will be seen that from a simple one-dimensional Fourier transform
we may obtain the answer directly.
The size of the entrance slit of the Rayleigh refractometer deter-
mines the brightness and visibility of the image seen in the eyepiece.
If it is too narrow there may be insufficient illumination and if it
is too wide the fringes dissolve as a result of the incoherence of the
light at the double slit. The coherence may be retrieved by using a
closer spacing for the double slit but this will reduce the number of
fringes seen in the field of view and may otherwise be impracticable.
It will be seen from the discussions in Chapters V and VII that the
criterion to be adopted here is that the two apertures of the double
slit lie well within the primary maximum of the Fourier transform
of the brightness distribution across the first slit. A rhombic prism
may be used to divide the light beam in place of the double slit and
thereby preserve coherence for a given size of entrance slit over a
larger separation of the light paths. The most efficient but needlessly
impracticable arrangement wOllld be to allow the light from the
collimator to illuminate a cosinusoidal grating and a mirror which
could be adjusted to concentrate two coherent beams on to the
apertures of the double slit (Fig. 7 b).
DIRECT APPLICATION 37
Examples in Acoustics
The Fourier transform may be used to compute the radiation
pattern from an acoustical aperture in precisely the same manner
as that indicated for aerials and optical apertures. The acoustical
apertures that are most frequently encountered are two-dimensional
and an example in this category, that of the vibrating disc, will be
given in the next section.
The frequency spectrum of a transient sound whose amplitude-
time distribution is known, for example by recording on a cathode
ray tube, may be computed directly from the transform relationship.
The frequency spectrum from an echo-sounding pulsed transducer is
analogous to the case of the spectrum of an electrical pulse, for
example Fig. 7 g, and the radiation pattern of the transducer at
the various frequencies of the spectrum may'be obtained by apply-
ing the transform as a spatial relationship. These transformations
are of considerable help in determining how the details of the re-
flector may modify the distribution of the returned pulse and the
effects of a Doppler shift resulting from relative motion.

Two Dimensional Fourier Transforms


The majority of optical, acoustical and aerial systems have aper-
tures which extend in two dimensions and similarly their Fraunhofer
diffraction patterns, or radiation patterns, also extend in two dimen-
sions. We can apply the Fourier transform to these cases by intro-
ducing an extra dimension into each coordinate system. Thus, where
previously we had (), we now have the orthogonal systems () and (/J,
and where we had only x we now have orthogonal axes of x and y.
The simple transform relationship now becomes

F((), (/J) = f
X=-oo
f
11=-00
t(x, y) e2ni (z6+ Y (/)dxdy

and, correspondingly,

t(x, y) = I f F((), (/J) e- 2 ,.;(z6 +Y(/)d ()d(/J.


8:0--00 ,z,=-oo
38 FOURIER TRANSFORMS

In principle, and often in practice, we may take as many dimen-


sions as we wish, though these cases usually arise in more abstract
analysis where the dimensions may be less tangible than in the cases
that we shall consider. It is worth while to remember that the Fourier
transform may often be applied within a mathematical analysis
when the physical meaning of the variables may not be at all clear.
Such examples often occur in the theory of probability.
Frequently, we encounter a simple transitional case where a two
dimensional distribution is examined through a one-dimensional

~
=
(ED (~) C:)

O@)@)~@J@)O
C:) (=) C:)
~
=

FIG. 11

system. Examples of this type may be found in the Michelson stellar


interferometer in which the star is essentially two-dimensional
whilst the interferometer derives the one dimensional transform of
the brightness distribution across the star. Further examples occur
in the case of one dimensional aerial systems used to compute the
intensity distribution of cosmic radiation. We must exercise care
in these cases to decide whether to work in amplitude or intensity
and whether to use the aperture function of the instrument or its
overall (itensity) transfer function. This is dependent upon the quality
of the radiation from the source, whether it be coherent or incoherent,
and a discussion of this will be left to a later chapter.
If we consider a uniformly illuminated rectangular aperture we
may consider the illumination to exist in the third dimension to
give the form of a box. Examining the Fraunhofer pattern from the
DIRECT APPLICATION 39
face of the box we should expect to find sin X/X patterns for the
amplitude along axes parallel to the sides of the box. The scale of
these patterns will be inversely proportional to the sides of the
rectangular aperture. This is illustrated in Fig. 11 in which the left
hand diagram shows the uniformly illuminated aperture in x, y
coordinates, whilst the centre diagram shows the amplitude contours
of the Fourier transform. The lower and right-hand diagrams depict
the distributions along the axes of () and (/J.
When the aperture is a uniformly illuminated disc the aperture
function is cylindrical, whilst the projection of the illumination into

Aperture distribution Equivalent strip distribution Transform (l

FIG. 12

one dimension gives a radially symmetrical pattern which falls off


from the central axis as

where a o is the radius and a is the distance from the axis corres-
ponding to the particular value of I (a). This yields a spectrum which
is somewhat similar to that from a rectangle but it is also radially
symmetrical and the sine term in the function sin xix is replaced by
the Bessel function J 1 Thus the spectrum is specified by J 1 xlx in all
directions from the centre. The function (J1 x)1x has its first zero at
122 units of n along the axis compared to n for the sin X/X function,
and its maxima are less pronounced. This may be directly inferred
from the tapering of the generating function which is shown in
Fig. 12.
40 FOURIER TRANSFORMS

As an example of the application of this transform in acoustics we


may take the case of a rigid disc of radius a vibrating in an infinite
baffle. The distribution across the aperture is uniform and the trans-
form gives the spatial sound pressure distribution at any point
distance r from the source provided r ~ a. Thus, quantitatively, if
P, III be the pressure at the point whose coordinates are given by the
suffices rand W, if (! is the density of the air and if ~ is the axial
acceleration of the disc, then*
P = (!~a2 J 1 (K asinW)
r, III r K a sin W

Aperture dis1ribulion ( ) Eql-ivalent strip distribution (S;X) Transform

FIG. 13

In Fig. 13 is shown the reversal of the above procedure, where the


function (J 1 '1,)/'1, is transformed back to a cylinder. Note that in this
case the equivalent one dimensional distribution follows a sin X/X
curve.
The treatment for an elliptical uniformly filled aperture is similar
to that for a disc but the axes of the mate are extended reciprocally
to the axes of the generating ellipse.
The Gaussian function similarly transforms to a Gaussian function
in two dimensions, whilst a lattice such as a "bed of nails" also
transforms into itself.
* See, for example McLACHLAN, N. W., Bessel Functions for Engineers, Oxford
Univ. Press (1934), p. 50.
DIRECT APPLICATION 41

The Numerical Evaluation of Fourier Transforms


If as the result of some experiment we determine a certain func-
tion and desire to know the transform of that function in order to
find the variation of the transformed parameter in the systeIp, how
do we set about making the transformation~ A number of methods
are open to us.
(i) If the function can be given a mathematical form we may
refer to a list of tabulated transforms, such as those in
Fourier Tra'n<~form8 by Campbell and Foster (Van Nostrand
Company Inc., New York) or we may solve the integration in
the manner used for the Fourier transform of a rectangle
earlier in this chapter.
(ii) We may use a computer or wave analyser to carry out elec-
trically, mechanically or optically the equivalent operation to
a transformation. See Appendix I.
(iii) We may perform an approximate practical analysis which
effectively transmutes the transform into the integral at
selected points. The Fourier integral is the sum to infinity of
the products of each value of the ordinate of the function with
a cisoidal vibration which is itself the transform of a delta
function. We may perform an approximate practical ana-
lysis which transforms the function f(x) into its transform
F(O) at selected points. We rely on the fact that variations
between these points are small and select them accordingly.
Let us take as an example the simple transformation in
Fig. 14, and assume that we require an accuracy of only a few
per cent. The information is given in the form of the function
at the top on the left, we require the transform of this. For
simplicity in the diagram the illustration is confined to the
real plane. We know that a delta function at a certain point 01 ,
on the abscissa of the transform will transform back into a
cisoidal wave. Multiply f (x) by this wave and integrate, then
multiply a unit function at 01 by the resultant, this will yield
the point P on the transform. Repeat for points O2 , 03 , etc.,
42 FOURIER TRANSFORMS

as required, giving points P 2 , P a, etc., thus building up the


transform. If f (x) is wholly in the real plane then multiplica-
tion by a simple cosine wave is equivalent to multiplication
by a cisoid.
Note that the zero frequency term is just f f(x) dx and the
order of magnitude of this and of the most dominant harmo-

fIx)

-x

ez,r\,e'r ~ ~ ~
Realpart~X o

fIx) e
271 ,e, 1-+--1'-----\f--+--+->OO'.....-

F(O)!"\
, '
Integrate I \
, \, ~
Integral =jf (x)e 2""e, dx =P , - _ _~:.--_..;:."-~-~
b 01 "8
FIG. 14

nic may often be judged by inspection. Thus, in Fig. 14 it is


evident that the area contained in the second maximum of f(x)
is less than that in the first whilst their polarities are reversed,
thus the zero frequency integral exists but is small. Cursory in-
spection also reveals that multiplication by a wave between
four and five times longer than that generated by 01 would
rectify the polarity of the second maximum and result in a
large value for the corresponding point on F(O). This indi-
DIRECT APPLICATION 43

ca,tes that the most dominant part of the spectrum occurs at


between one quarter and one fifth of the way to 01 , It may also
be appreciated from Fig. 14 that if the function contains only
positive values in the real-plane the transform of that function
has its maximum value at the origin, thus in Fig. 14 if we
transform the wholly positive function F (0) we arrive back
at f(x) where the maximum value occurs at x = O. On the
other hand, f (x) itself contains both positive and negative

Since

o And

FIG. 15

states and thus transforms into F(O) which does not have its
maximum at the origin.
Note that the function chosen for illustration in Fig. 14 was
symmetrical and it was therefore only necessary to use posi-
tive values of x.
(iv) Where the components oscillate so rapidly that the labour
involved in the previous method renders it prohibitive we
may use a more elegant method of numerical approximation
such as that due to Filon*.
* FlLON, L.N.G., Proc. Roy. Soc. Edinburgh 49,38 (1928-29); See also'SNED'
DON, LN., Fourier Transforms, McGrawHill (1951), p. 519.
!4 FOURIER TRANSFORMS

(v) We may in many cases compute the transform by observing


that the function is the sum of two or more better known
functions whose transforms are easily recognised or listed.
We then make use of the superposition or summation pro-
perty to obtain the complete transform. Even when the function
cannot .be expressed as the sum of well known functions such
as the rectangles in Fig. 15, the general behaviour of the trans-
form may still be implied by similar reasoning.
Perhaps the most powerful treatment by this method is
the superposition of Gaussian functions. We have already
seen in Fig. 8(a) that a Gaussian transforms into a Gaussian
whose scale is reciprocally related to its mate. It is possible to
build up an approximation to any function by the superpos-
ition of appropriately chosen Gaussian functions and the
mates of these may be superposed on the transform, just as with
the rectangles of Fig. 15. In this connection it is worth noting
that an off-set Gaussian (or rectangle, etc.,) is simply trans-
posed to its mate by the convolution property (see vi, below)
whereby the mated Gaussian (or sin xix function, etc.) is
multiplied by the transform of a simple off-set delta function.
(vi) We may make use of the product or convolution property to
transpose a number of better known functions, see for example
the many examples in Fig. 7. This will be discussed in Chap-
ter V. We may use in a similar way, the differentiation pro-
perty discussed in Chapter VI.
CHAPTER V

LIMITATIONS,
PRODUCTS, AND CONVOLUTIONS

FIGURE 16 illustrates an optical system in which two slits are facing


each other at a considerable distance relative to their apertures. We
have chosen the slits so that the right hand slit is of such a size as to

I ....==:'------
_ _ _ _ _ _ _J
---------------=----"'---------------
-------------h----h----_=_-it

-------------------,
FIG. 16

allow through only the first maximum of the Fraunhofer diffraction


pattern of the left hand slit (shown also as a polar diagram). Now as
the beam pattern is reciprocally related to the aperture, the larger
slit on the right will have an associated Fraunhofer pattern at the
left hand slit which is also of such a size as to allow that slit to
accept only the first maximum. We can see that a similar relation-
ship will also hold whatever the width of the slit that we may choose;
the slit will always cut off the pattern of its mate at exactly the
same point as in the reverse process.
We have assumed in this illustration that both slits were uniformly
illuminated, however, we see that if we uniformly illuminate one of
the slits only, the distribution of illumination over the other will not
be uniform but will be given by the product of the first diffraction
46 FOURIER TRANSFORMS

pattern with the second aperture. On viewing the second slit from a
great distance on the furthermost side we shall not now see the
diffraction pattern of the first slit but that of the second slit illu-
minated with an amplitude and phase distribution depending upon
the transform of the first slit.
There are a number of useful principles embodied in this example.
Firstly it shows how, in the study of planetary or stellar bodies, we
may work from either end of the system, i. e. we may consider the
aperture of the telescope limiting the diffraction image of the star at
the surface of the earth, or we may consider the diffraction or beam
pattern of the telescope only partially filled by the star which is
therefore incompletely resolved. As the diffraction pattern of the
star is never actually projected on the earth, the use of this principle
is obvious, though we shall see in a later chapter that, as most
celestial objects are incoherent radiators, we must work with the
intensity transfer function of the telescope rather than its aperture
function, or, alternatively, use the opposite treatment with the
power polar diagram. .
Secondly, we see that a slit may behave as an angular filter, that
it may therefore restrict some of the information transmitted to it by
the source and that it can superimpose its own transform upon the
radiation passing through it. To make this point doubly clear we may
consider an infinite plane wavefront falling on a slit. On the far side
of the slit the diffraction pattern will be that of the slit alone and we
can consider this either as being due to the action of the slit as a fil-
ter upon the original radiation or as the pattern due to the slit
itself when filled with uniform illumination. (Huygen's principle).
Thirdly, if the radiation is incoherent across the plane of a slit-
i. e., if there is no correlation between the radiation from any two
points, putting the slit in front of the source introduces a spatial
correlation which is dependent only on the characteristics of the slit.
This is directly analogous to the way in which an electrical filter can
select a narrow band of "white" noise and give a quasi monochro-
matic wave output. The white noise can be considered as a random
succession in time of delta functions of arbitrary phase. Each contrib-
utes its own spectrum over an infinite frequency band and each of
LIMITATIONS: PRODUCTS: CONVOLUTIONS 47

these spectra is limited by the filter so that the resulting output bears
the indelible stamp of the latter and the filter "rings" at its charac-
teristic frequency, superimposing its own time function on that
of the random input and giving an output that is still random within
the narrow band.

Convolutions
We have seen how we may solve a number of optical or electrical
problems by the Fourier transform by breaking up the initial func-
tion into its transform, passing this through the selective device
under consideration and then transforming back into the original
system, thus finding the effect upon the output of having the inter-
mediate component or selective device in the system. This method
involves two transformations, one into the intermediate system
and one out again.
It is clear that in many cases we shall save ourselves much labour
if we can perform the same operation with only one transformation.
This can be done by taking the transformation of the intermediate
system, applying to it the original function and determining the
resultant output as the integrated product of each point in the ori-
ginal function operating on the transformed response.
We owe this choice to the convolution theorem. We may state it
thus:
The Fourier transform of the product of two variables is equal to the
convolution of the Fourier transform of one of the variables with the
Fourier transform of the other.
The convolution of the two functions g(t) and h(t) may be ex-
pressed mathematically as
oc

f g(1') h(t - 1') d1'

and if g (t) transforms to G (f) whilst h (t) transforms to H (f) we may


state the convolution theorem:
oc oc

f G(f) H(f) eh i l l df = f g(1') h(t - 1')d l'


48 FOURIER TRANSFORMS

Let us derive this relationship by the use of the delta function. We


have already seen that a delta function in time applied to a system
having a frequency response Y (f) will give an output, i, varying with
time, t, as i (t) where
00

i(t) = f Y(f) e ";ttdl


2

[see, for example, the voltage spike injected into the electrical filter
of Fig. 10 (a)].
We wish to find the current output for any voltage input e(t). We
may arrive at the result in two ways.
(i) Fourier analyse e (t) into its components E (I), multiply this
function by the frequency response of the system, Y (I), and
transform the resultant back into time, thus
00

i (t) = f Y (I) E (f) e 2 "ift dl

This method was used in deriving Fig.10(b).


(ii) Consider a series of unit impulses representing voltage spikes
following each other in rapid succession, each impulse bearing
an amplitude modulation factor which is equal to the ordinate
of the original time function e (t) at that particular time or, i. e. the
spike at a time or has an amplitude e(or). Now if y(t) is the time
response of the filter, that is the transform of its spectral
response or frequency admittance function Y (f), it is clear
that a voltage spike occurring at time or will trace out the
time response y (t) delayed to start at time t = or, and the
amplitude of the output will be proportional to that of the
spike, this is shown in Fig. 17. If we now consider the time res-
ponse i (t) to a time function e (t) which is formed .by packing
together a continuous distribution of impulses whose ampli-
tude is adjusted to follow e (t) then it is clear that
00

i(t) = Jy(t - or) e(or) dor


LIMITATIONS: PRODUCTS: CONVOLUTIONS 49

Since the two methods (i) and (ii) are equivalent ways of
reaching the same result, we have again the relationship

.r E(f) Y(f) e2niftdf =.r e(T) y(t - T) dT

'=0 '=0 ,-

1=0 I=r ,- '=0 t=r t-

FIG. 17

Physical Interpretation of Convolutions


If we examine the function

.r e(T) y(t - T) dT

we see that the convolution of the two functions within it is the


result that we would obtain by scanning one function with each
element of the other (from - 00 to + 00) and summing the pro-
ducts at each position of the scan. Thus the convolution can extend
beyond the bounds of one function taken singly, due to the product
of the overlap within the second function. A convolution is not a
Fourier transform but an entirely separate operation in its own
right. It may be used independently of transformations or in con-
junction with them by the application of the theorem that we have
referred to as the "convolution theorem".
In Fig. 18 (a), a "white" frequency spectrum* is injected into two
filters having different pass bands E (f) and Y (F - f) where F is
the tunable centre frequency of the passband, Y (f), of the second
* A white frequency spectrum is one in which the intensity is constant at all
frequencies.
50 FOURIER TRANSFORMS

filter. As F is tuned through all frequencies an output will appear over


that range of frequencies where the spectra overlap. This output,
as a function of F, will be given by the convolution formula

i(F) = f E(I) Y(F - f) df

In Fig. 18(b), the filters, instead of being connected in series, are


fed into a balanced linear multiplier, such as a ring demodulator
White spectrum

White spectrum

Output spectrum I(F)


=convolution of the fwo responses

(a)

(b)
FIG. IS

or balanced hexode mixer, the d. c. output selected from the filter in


the output of the linear multiplier will now vary as the square of
the Jommon input but will follow the convolution of E (I) and Y (I)
as F is tuned throughout the entire frequency range. Figure 19 (a) is a
direct extension of the case in Fig. 18(b). It invokes the principle of
the linear mixer that any linear change may equally well be per-
formed before or after mixing.* Thus, instead of sweeping Y(f)
through all frequencies by varying its centre frequency F, we per-
form the same operation by looking at the output of the mixer -
* Apart from an ambiguity of sidebands giving rise to two output spectra in
the case of Fig. 19, one centred on the sum frequency and one on the difference
frequency. Where the difference is centred on zero the latter spectrum is folded on
itself.
LIMITATIONS: PRODUCTS: CONVOLUTIONS 51
which is now not filtered except by its self balancing against direct
break through of Y (f) and E (I) - and recording the output at fre-
quency F as F is swept through the whole spectrum, i. e. we observe
the spectrum of the output given by the cross beats formed between
E (f) and Y (f) around the difference frequency and a similar spectrum
around the sum. These have the form of identical convolutions but
may be confused by overlap.
White spectrum

Broadband output contolning crossbeats


Spectrum of crOSSbaatsa/:E(f)Y{F-ndf
(a)
White spectrum

White spectrum

(t
Broadband output with spectrum i(F)
where iCF)=1:Y(flY(f -Fldf Output spect<um J detector '1~ Y(fl Y(f-Fldf

(b) (e)

FIG. 19

The lafot case is of importance in the analysis of mixing or detecting


systems. It leads to the case of the convolution of a spectrum with
itself (Fig. 19b) and as the multiplier is accepting two identical fre-
quency bands, it may be replaced by a square law detector as in
Fig. 19 (c). If the passband is rectangular, the resulting form of each
convolution will be the triangular spectrum derived in Fig. 23(a).
If Y (f) extends to zero, the two spectra add to give a right angle
triangle on a base equal to twice the passband.
52 FOURIER TRANSFORMS

We have been concerned above with convolutions of the frequency


parameter, Fig.20 shows a simple example of the convolution of
time parameters. The simple delay line has linear dispersion, that is
the output is constant in amplitude but the phase increases in angle
Phose

,.0 .L. } I~
Pulse Delay line Convolution
Frequency response TIme resoonse

FIG. 20

with increasing frequency. Transforming this elementary character-


istic it will be recognised as an off-set delta function occurring at
time t = T. The convolution of this with the applied rectangular
pulse is a similar rectangular pulse delayed by time T.

(0)

=
(b)

FIG. 21

In the previous chapter we referred to the possibility of simpli-


fying the analysis of the electronic apparatus set out as a block dia-
gram again in Fig. 21 (a). We wish to find the form of the output
time response when an arbitrary time function e(t) is impressed upon
a circuit having a frequency response Y (f) and the output is viewed
LIMITATIONS: PRODUCTS: CONVOLUTIONS 53

on an oscilloscope. This problem was previously tackled by trans-


forming e (t) into the frequency function E (f), multiplying this by Y (f)
and transforming the product back to give the resulting time re-
sponse. Obviously, the problem may be simplified by using the con-
volution treatment shown in Fig. 21 (b). Y (f) is transformed into
its impulse response y(-c) and it then only remains to form the con-
00

volution integral! y (-cr e (t - -c) d -c in order to obtain the output.


The curves drawn to demonstrate the responses in Fig. 21 (b) are
arbitrary and are intended to portray only the method of operation.

- -g(x)

-hex)

G(6') HC,s-6'}

FIG. 22

As a first example of the use of the convolution in optics we may


consider a light chopper consisting of an opaque rotating disc in
which is cut a single radial slit. Parallel light falls on the disc which
periodically illuminates a second fixed slit as the two become aligned.
If the two slits are indentical rectangular apertures the light output
of the system will not be a rectangular pulse but an isosceles triangle
determined by the convolution of the two identical rectangles. If the
two slits are rectangular but of different width, the light output as a
function of time will be spread out to the form of a symmetrical tra-
pezium or a truncated isosceles triangle.
Secondly, we may consider an example in diffraction. A diffract-
ing screen f (x) is formed by superimposing two gratings g (x) and
h(x) as in Fig. 22.
If we consider the ray path in Fig. :~2 we see that in a direction r;p
with the normal there will be a ray which was diffracted at an angle
54 FOURIER TRANSFORMS

oby the first grating g(x) and a further angle (C/J -0) by the second
grating h(x). To find the total amplitude of the ray diffracted in a
direction C/J we must scan over all values of 0 and so obtain
00

F(C/J) = f G(O) H(C/J - 0) dO

Thus the diffraction pattern of the composite screen is given by the


convolution of their separate patterns.
Note that we could otherwise have obtained the same result by
taking the product of the aperture distributions to give a composite
aperture distribution I(x) = h(x) g(x) then upon transforming I(x)
we obtain
00

F(C/J) = f I(x) e "i(/;"'dx


2

Upon application of the convolution integral theorem this will be


seen to be identical with the previous expression for F (C/J).

Some Examples of Convolutions


We have shown how a convolution may be built up from a con-
sideration of the response of a system to adjacent impulses, similar
to delta functions but of controllable amplitude. In Fig. 23 three
simple convolutions are shown as being formed in this way. For
simplicity in the diagram only a small number of impulses have
been used in the pictorial analysis, but the integral form of the final
curve is shown for comparison. The first example, Fig. 23 (a) shows
a rectangle in convolution with itself and this will be seen to yield
a triangular spectrum. It will be seen that in the integral form the
convolution commences smoothly from zero rather than rising with
a discrete step, as in the simple sum. This is due to the fact that though
the first contrib~tion is admittedly that from one impulse, if these
are taken infinitely closely packed in the integral the central part
of the convolution will be the sum of an infinity of contributions
whilst the ends only correspond to one. If the centre is to be finite
the ends must therefore be infinitesimal. Fig. 23(b) shows a triangle
LIMITATIONS: PRODUCTS: CONVOLUTIONS 55

in convolution with a rectangle and Fig. 23 (c) shows the convolution


of two rectangles of different width. These cases provide examples of
First spectrum Second spectrum

--'---I -Q- !~
""':0--->< -0---0--..

Sumo! components~
Integ'OI.conVOlutionQ

(0) 1+----2 - -

I
I

Sum of components~

Inlegrc:i=convolutIO~
(b)

={~
Sum of components!\

In.eg,OI:: convoutlOA
FIG. 23
(C)

the benefit to be derived from reversing the process. If we wish to


transform the convolved curves F (f) we can equally well operate
with the more manageable component curves G(f) and H(f). The
Fourier transforms of the curves G (f) and H (f) are elementary and
56 FOURIER TRANSFORMS

they may be simply solved to reconstruct the Fourier transforms


of F(f).
Somewhat similar to the case of Fig. 23 (b) is the convolution of a
rectangle with an exponential. This gives a function of the form
shown in Fig. 24 which is precisely the same as that appearing when
a square pulse is passed into an integrating network formed by a
series resistance and parallel condenser.
~
T
FIG. 24
In this case the exponential time function of the integrating net-
work may be considered as fixed in time whilst the pulse performs
the scanning operation.

The Solution of Fourier Transforms by the Application


of the Convolution Theorem
We referred in Chapter III to the possibility of solving certain
transformation problems by the application of the Convolution
Theorem. In using this technique it is only necessary to judge
whether or not the function concerned can be generated by either
the product or the convolution of two simpler functions and then
apply the theorem that we have just discussed.
To illustrate the power of this method we shall take t:wo simple
examples.
(i) The Fourier Transform of an Isosceles Triangle.
This transform may be obtained directly from the formula
00

F(O) = f f(x) e2" ix6 dx by substituting a function representative of


-00

the triangle for f(x). We may then break the integral into three parts
one of which is zero from the negative limit of the triangle to - 00,

another is zero from the positive limit to + and the third is an


00

integral of the form D

f m(D - x) e2nix6dx
-D
in which D represents half the base of the triangle.
LIMITATIONS: PRODUCTS: CONVOLUTIONS 57

An alternative and shorter method of solution is to obtain the


transform from a convolution. We have previously seen (Fig. 23a)
that the convolution of a rectangle with itself is an isosceles triangle
of base 2D where D is the width of the rectangle. Now the transform
of a rectangle is a function of the form sin X/X and from the convo-

o ........
/\ --~
Convolved IIoilh x

o o

= =

--zo- I+!'"
o
FIG. 25

lution theorem we may thus obtain directly the Fourier transform


of the triangle as a function of the form
sin I(
- - x sin
- -X
X X
This is illustrated in Fig. 25.
(ii) The Doublet Pulse
We may use the convolution theorem to solve in a very simple
manner the transform of an elementary function which has not pre-
viously been discussed. This function is the simple doublet pulse
shown with its transform in Fig. 26. To the left hand side of the ori-
gin its amplitude is - 1 for a distance d after which it remains at
zero. To the right hand side of the origin its amplitude.is + 1 for a
further distance d, after which it is again zero. A simple example of
58 FOURIER TRANSFORMS

the occurrence of the transform is in the polar diagram of a large


aerial, such as a collinear broadside array, in which the two halves
are connected in anti-phase.
The transformation from the aperture distribution of the doublet
to the resulting expression F (0) may be performed in a great variety
of ways. We shall first solve the problem mathematically.
00

F(O) = f f(x)e 2niz6 dx

o d

=f -e Z,,; z6 dx + f eZ,,; z6 dx
-d 0

= _ [_1_._
2:roO
e + [_1_. e2 "iZ6]d
2 "iZ6]O
2:n:~0
-d
= _ [__
1_ eO __1_ e-2"i6d] + [_1_ e2"i6d _ _1_ eo]
2:n:iO 2:n:iO 2:n:iO 2:n:iO

I I -2"j6d 1 2"j6d 1
= 2:n:iO + 2:n:iO e + 2:n:iO e - 2:n:iO

i i 2 Od
= :n:0 - :n:O cos :n: ~ ~

~
= :n:O (1 - cos 2:n: 0 d)

2ill
= - (- - - cos 2:n; 0 d)
:n:O 2 2

= 2 i d. sin2 (:n: 0 d)
:n:Od

This function is plotted on the right of Fig. 26. It is an asymmetrical


function in which the maxima to either side of the centre line bear
opposite sign. It passes through zero at 0 = 0, lid, 21d, 3Id, . ..
IfF(O) was the pattern generated by an aerial in which f(x) repre-
LIMITATIONS: PRODUCTS: CONVOLUTIONS 59.

sented the field distribution it is apparent that we should obtain two


main lobes with zero response along the axis. The difference of phase
between the two lobes could be detected by an inferometer technique
and would not be apparent in the power polar diagram that is nor-
mally measured.

F(fJ) o~.,
=~~-
-ij-

FIG. 26

Let us now consider how we may derive the expression for F(O)
in a less arduous manner. The doublet pulse may be obtained from the
convolution of a simple rectangle with the elementary odd function

o
Convolved wifh

I
o-.-d~

= =

FIG. 27

shown in Fig. 27. We have encountered this function before, in


Fig. 7 (c), and we know it to be the transform of a continuous sine
wave. If the width of the rectangle is d we require the two delta
functions to be separated also by d and the period of the sine wave
will be 21d. Now the problem is solved, from the convolution theor-
em the transform of the doublet is given by the product of a func-
tion of the form sin xix with another function of the form sin x, the
scale being such as to cause the minima to occur at spacings of lid.
60 FOURIER TRANSFORMS

Hence

A further simple method of solving the transform of the doublet


pulse is by the use of the differentiation theorem. It will be demons-
trated in Chapter V.

Convolutions Involving the Sine Integral Si(x)


We have seen that a rectangular spectrum transforms into a
distribution of the form sin X/X. As both the spectral function and
its transform occur very frequently in analysis it will be profitable to
study their convolutions.
Let us take two specific examples.
(1) An optical instrument similar to a telescope with a rectan-
gular aperture is used to study the light from a distant coherent slit
source. Close examination of the image will show that it is not a
true picture of the distribution across the slit but has less sharply
bounded edges and faint diffraction fringes.
Each point in the distant slit gives rise to an image from the
second rectangular aperture which is in the form of the diffraction
pattern of the latter, sin 2 x1x 2 These patterns are extended side by
side and we see the integrated result of all the patterns due to the
adjacent points in the generating rectangular brightness distribution.
As the light is coherent we may integrate the amplitude patterns
giving an integral of the form

!Si: x
o
'"

and we may square these ordinates to obtain the intensity distribu-


tion in the image.
(2) A square pulse is passed through a low frequency amplifier with
a sharp cut off at a certain frequency which limits the frequency
content of the impressed pulse. Here we have a rectangular pulse
convolved with a system having a time response to an impulse of the
form sin X/X.
LIMITATIONS: PRODUCTS: CONVOLUTIONS 61

In each case it is clear that the result that we shall obtain for the
amplitude is that formed by adding together adjacent functions of
the form sin xix over a range equal to the width of the rectangular
function which limits the response.
Si(x)

t
20
?I"
-----2-- I
15 ---

-20 -18 -16 -14 -12 -10 -8 -6 -4 -2 2 ?I" 4 6 8 10 12 14 16 18 20


-05
I
+0
I
---t-~----
- 20

I
FIG. 28

It is evident, therefore, that the behaviour of the function, Si (x),


obtained by integrating the sin X/X function is relevant to a great
many physical problems. The integral

f Sin~ dX
X
can be evaluated in a power series. Tables of the definite integral

Si(x) = !Sinr.
'"
. X
dX
o
were first prepared by Lord Rayleigh. *A five figure table listing values
of x up to x = 20 is given in the appendix at the end of this book.
A plot of the values in the table follows the form of the curve in
Fig. 28. It will be seen that the value of X does not actually contri-
* RAYLEIGH, Proc. Roy. Soc. A 90, 318, 1914.
62 FOURIER TRANSFORMS

bute to the result of the integration and Si (x) is a function of the


limit, x, alone.
-The function is odd and passes through the origin with unity slope.
As x increases it oscillates about the value 11:/2 with a wavelength of
211: and with decaying oscillatory amplitude. For negative values of
x the function oscillates about-11:/2.
The sine integral is clearly an approximation to a step function
and remains indefinitely at the level -11:/2 or 11:/2 except in the region
of x = o. An extended picture of the function would look like that
shown in Fig. 29 where we see that it appears as a step but with an
anticipatory ring and an overshoot. These effects have a peak am-
plitude of nine per cent and are inherent to the function, they are
an example of Gibbs phenomenon in Fourier integrals and have the
same amplitude as Gibbs phenomenon in Fourier series. The pheno-
menon was first observed by Michelson in his wave analyser and
was initially thought to be due to a fault in the machine.
Returning now to the case of a rectangular time function con-
volved with a rectangular passband, we see that the integration of the
transform of the latter function corresponding to each point in the
former gives rise to a Si (x) function which, for a rectangular pulse
of long duration, remains up at the higher level 11:/2 until the rectan-
gular time function ceases, where upon we have the identical situa-
tion in reverse, i. e. another step function descending this time to give
a response of the form shown in Fig. 30(e).
The other curves in Fig.30 depict the convolution of a function
with a rectangle of progressively increasing duration, T, from T =11:/2
in Fig. 30 (a) to T = 811: in Fig. 30(e). In all cases T commences
from the dotted line at t = o. The scale of 11: is that determined by
the sin X/X function, i. e. 11: is the distance from the peak to the first
minimum of that function. Note that the flat top only becomes
clearly established for rectangles which are wider than 211:. All the
curves in Fig. 30 may be obtained quite quickly by tracing the Si
-function in Fig. 28, reversing the paper, moving the zeros of the two
curves apart by the desired length, T, of the rectangle, and then
adding their ordinates to give the resultant convolution. Note that
for T = 0 the resultant is zero whilst the convolution is diminutive
LIMITATIONS: PRODUCTS: CONVOLUTIONS

Fw.29

1:0
Fw.30
64 FOURIER TRANSFORMS

and quite undeveloped for values of T much less than 2 :n. It should
be noted that the rectangular pass band is highly idealized and a
practical pass band in an electronic circuit would possess an asso-
ciated phase characteristic which would delay the response and pre-
vent the occurrence of an output at an epoch prior to the applica-
tion of the input. An optical or aerial system with an angular res-
ponse or beam pattern in amplitude of the form sin X/X will distort

fIt) I
ConllOlved with

Ylf)

= =

-I
dr------ '
-1_~
FIIl,Y!f)

,' _________ _..1 f


2
FIG. 31

the appearance of a rectangular uniformly illuminated source in just


the same way, so that, as we might expect, our knowledge of the
brightness distribution across the source is limited by the angular
response pattern of the instrument and hence by the aperture of the
latter. Though the image is made broader by the source, the fea-
tures of the rectangle only appear for T > 2:n and even then the dis-
tortion is considerable.
Figure 31 shows the convolution of a step function with a sin xIx
function. The result is equal to the transform of the product of their
frequency distributions. This example will be discussed in Chapter VI.
LIMITATIONS: PRODUCTS: CONVOLUTIONS 65

Two Dimensional Convolutions


In a great many physical problems it is frequently useful to employ
a convolution in two or more dimensions. The scanning of the object
in television transmission and the simple projection of the object
into the image space in any linear optical system are two obvious
examples.
Figure 32 shows a simple linear optical system in which a lens
focuses the detail in the object space, where the coordinates are lX
y

FIG. 32

and p, into the image space specified by x and' y. In this case the
image is specified by
00 00

f(x,y) = f
a=-oo
f h(x -
P=-oo
lX, Y - p) Y(lX, P) dlX dP
where f (x, y) expresses the two dimensional convolution of the true
object with the point source response of the lens.
The point source response of a circular aperture, such as a perfect
lens or mirror free from aberration, has an intensity distribution
given by (J~X)/(X2). If the lens forms the aperture of a telescope and
is used to view an object which may just be resolved, careful ex-
amination of the image will show the presence of faint diffraction rings
or contours close to the edge of the image. These contours are fami-
liar to the users of large aerial systems as the side lobe response of
the aerial. They remain even when the resolution is much greater
than the angular size of the object, but they then closely approximate
66 FOURIER TRANSFORMS

the true edge of the image, whilst the edge itself likewise becomes
sharper as the resolution increases. This is shown for a cylindrical
(uniformly illuminated disc) object in Fig. 51, page 94.

Cylinder convolved with itself =


FIG. 33

We have previously discussed the case of two one-dimensional


optical systems in cascade when, in Fig. 22, we dealt with a diffrac-
ting screen formed by superimposing two gratings. A much more
frequently encountered situation is that where the cascaded systems
are two-dimensional.
F.T

Convolved with

F.T.

FIG. 34

A uniformly illuminated disc may be represented as a cylindrical


distribution (see, for example, Fig. 12). If two such discs are cas-
LIMITATIONS: PRODUCTS: CONVOLUTIONS 67

caded or if one uniformly illuminated disc is convolved with itself, we


obtain the almost conical convolution shown in Fig. 33, in which the
left-hand diagram shows the two cylinders at one stage of the
scanning operation whilst the right hand diagram shows the final
convolution. By utilising the relationship that the convolution inte-
gral is equal to the Fourier transform of the product, we may obtain
the Fourier transform of the two cascaded cylinders as a radially
symmetrical function of the form (J~ X)/(X 2 ) as shown in Fig. 34.
This important relationship is frequently encountered in problems
concerning the transfer of intensity from a circular aperture.
CHAPTER VI

THE DIFFERENTIATION
OF FOURIER TRANSFORMS

ARMED with the superposition property discussed in Chapter III


and the product or convolutiorr theorem dealt with in Chapter V, it
is possible to handle most of the problems that we encounter in
practice. There are, however, other simple tricks that we can play
with Fourier transforms, which will often lead us more directly to
our goal or give us a more immediate appreciation of the behaviour
of the problem under investigation.
The differentiation property is equally as facile as those already
discussed. In common with the convolution property and unlike the
superposition property it utilizes an asymmetrical operation on the
two sides of the equation, but both operations are equally simple
and once committed to memory they prove invaluable in practice.
Let us once more write down the basic relationship.
00

f(x) = f F(O) e2nixodO


differentiating this expression we obtain
00

f'(x) =(Idxf(X) =/F(O).2nioe2n ;xO dO

If now we put G(O) = 2n i 0 F(O) we obtain


00

f'(x) = f G(O) e2nixodO


DIFFERENTIATION 69

and evidently f'(x) and G(O) are a Fourier pair; and whilst we may
obtain f'(x) from f(x) by differentiation, we obtain G (0) from F (0) by
multiplying it with the function 2 n i 0.
We may repeat this process indefinitely and in general if

Then
g(x) = (ddx r f(x)

00

g(x) = f (2n i 0)' F(O) e 2 ", ix odO

00

= f G(O) e2 ",; x odO

where G(O) = (2 n i Or F(O)


In this case g(x) and G(O) are Fourier pairs.
Verbally we may write:
The r th differential of a Fourier transform f (x) is another Fourier
transform which may be obtained by multiplying the original mate F (0)
by the coefficient (2 n i Or
To illustrate the practical application of this theorem, let us first
consider a simple function which has previously been discussed, the
transformation of a rectangle.
The theorem tells us that if we differentiate the rectangle f(x) we
shall obtain the transform G(O) of the resulting function by mul-
tiplying the transform of the rectangle (see Chapter III) by the
function 2 n i O.
Performing this operation we obtain

G(O) = i (2A sinn 0 D)

which is a simple sinusoidal wave in the imaginary plane having an


amplitude proportional to that of the original rectangle and a period
inversely proportional to its width. The sequence of operations is
shown in Fig. 35.
The result shown above was previously derived in Chapter III by
means of simple delta functions of opposite phase, I:\>nd the trans-
70 FOURIER TRANSFORMS

form was shown in Fig. 7 (c). The present method therefore seems to
be a tortuous method of gaining a simple result. The beauty of the
treatment lies in a reversal of the process. Differentiation tends to
simplify the system but if we work upwards in Fig. 35 and divide by

f(X~
-
-x !
,
:
I -
x
F(~
....f~
1\M:
-.
9
,, ,, :: !I I: l I::
I
:
:
:
I

,t---D -----j, : : : :01 : !


!, ,,I
! : ! : : : i
I I I " I

. g(xl=d~f(xl : G(t9l=21Tit9F(t9) iii i i


-x
i
- x
--....
~
-00_
-19
',I I I I I I I

\ _00
19

FIG. 35

2 :n; i () we see that we may derive the transform of a rectangle in a


very simple manner.

The Fourier Transform of a Step Function or Straight Edge


We may apply the differentiation theorem to solve, very simply,
a tra!lsformation which may not be directly tackled in a simple
manner. In Fig. 8 (c) we showed the transform of a step function but
we have deferred the derivation of this transform until now. We
might expect that we could substitute in the general formula

F(()) = f f(x) ez,. ix8 dx

two integrals to represent the constant behaviour of the function


before and after the single discontinuity at x = O. It is apparent,
however, that the functions will not converge. The perfect step
function remains at the higher level all the way to infinity. In order
to solve the function in the conventional manner we must, therefore,
cause it to diminish in the far distance, so that it reaches zero at
infinity. Though this treatment is quite in order and yields the
correct result, one tends to feel that it is slightly dishonest and sa-
vours a little p sleight-of-hand.
DIFFERENTIATION 71

Figure 36 shows how the transform may be derived from the


differentiation property. We reverse the method of operation com-
mencing with the step function in the bottom left hand corner of
the diagram. If we differentiate this function we obtain an infinitely
narrow line at x = O. This is a delta function and we know that the
transform of a delta function is the constant and infinite spectrum in
~f(x)

-x -00-0

fIx)
-+ ,
x I
.., _...... _-- - '2
1T"

FIG. 36

the top right hand corner of Fig. 36. All that remains is to find a func-
tion which, when mUltiplied by 2 ni () gives this constant spectrum.
The solution is obviously F(()) = 1/(2n()) contained everywhere in
the imaginary plane and this may be depicted, as in the lower right
hand corner of Fig. 36, as a function whose amplitude varies symme-
trically but whose phase is n/2 from () = - 00to () = 0 whereupon
it smartly jumps to the constant value -n/2 for the range () = 0 to
() = + The values n/2 and -n/2 define the imaginary plane
00.

whilst the change from n/2 to -n/2 is equivalent to the change of


sign introduced when () goes from negative to positive.
In the previous chapter we discussed the behaviour of a step
function when fed into a low pass filter having a sharp cut off; this
problem was treated as a convolution, and it was shown that the
output was in the form of a Sine Integral. The same result may be
obtained from the transform that we have just derived. If the trans-
fonn represents the frequency content of a step function in time,
then the low pass filter will limit this suddenly at a finite distance to
either side of the centre line. From the summation property we
72 FOURIER TRANSFORMS

know that the transform back into time of the resulting truncated
pattern added to the transform into time of the missing components
from the imposed frequency limit to infinity, must equal the original
step function. The missing high frequency components, therefore,
give rise to the undulations preceeding and following the rising edge
of the Sine Integral function in Fig. 28.
It is interesting to note that the spectrum of a step function is
equivalent to the impulse response of a perfect inductance.

The Differentiation of a Triangle


In the discussion of the convolution theorem in Chapter V, we
showed how it could be used to obtain the Fourier transform of an
isosceles triangle. The transform was found to be of the form sin 2 x/x 2
Let us now start from the isoscles triangle and apply repeated
differentiation. The first differentiation produces the doublet pulse
and we obtain directly the transform of this pulse as the function
sin2 mx 2' . sin 2 mx
(mx)2 X n ~x = constant X ~ X mx

which, upon substitution for m, is the form previously derived.


Repeating the differentiation we obtain the "carrier plus side-
bands" spectrum in the bottom left hand corner of Fig. 37. This,
it is seen, must transform into a cosinusoidal wave positioned so that
its peaks touch the baseline, since the product of two succesive
operations with i gives a negative sign this curve may be ~xpressed
from the previous function in the form
2
mx 2
~
. SIn
X n ~. x = constant X (cos 2x - 1)
mx
If we have a se;ries of triangular waves we may obtain the trans-
form in the manner shown in Fig. 38, where we commence with a
regular grating function and each successive operation reduces the
components of G (f) in proportion to their displacement.
Figure 39 shows the transform of a triangular doublet obtained by
differentiating the sin X/X function and multiplying the rectangle by
2n i (). The resulting function is similar to the transform of a doublet
but has more pronounced side lobes of alternate phase.
DIFFERENTIATION 73

x2n .. ~

-oo2.\IVVV\I\fi:"'oo
-x X

_!+1
o
FIG. 37

r
I I I I I I I I
-'-2'1Tif! I-

I II' Il I, ,!

g(1) FIG.3S G(fl

---"-
.....---
8-

.2'11'i 8
J
--"-
.....--- /1
FIG. 39
V 8-
74 FOURIER TRANSFORMS

The Convolution - Differentiation Relationship


There is an interesting connection between the convolution and the
differentiation theorems.
Consider f (x) convolved with (ddx ) g (x), by the convolution theo-
rem this is equivalent to
00

f F(O) . 2:n; i 0 G(O) e2n ;x6 dO


where F(O) is the Fourier transform of f(x) and G(O) is the Fourier

5l...:.0nvolved with~ ~nvolved with


nn
~D~ ~
~
...0 ...

= =

C1 C1 ... O.r! C1
LJ 0
FIG.40

transform of g(x). The 2:n; i 0 is brought down by the differentiation


of g(x). We may re-write the integral in the form
00

f G(O) . 2:n;i 0 F(O) e2nix6dO


and, again by the convolution theorem, this may be written as
ddx f (x) convolved with g (x).
This shows that the solution of the convolution integral is not
unique unless one of the participant functions is defined. It also pro-
vides a further method of building up functions. Figure 40 shows
the application of the relationship to the construction of a repetitive
square wave from the convolution of a single square wave with a
grating of alternately phased delta functions spaced apart by the
same width as the square wave. It will be seen that a similar function
DIFFERENTIATION 75
may be produced by convolving a pair of oppositely phased delta
functions, the derivative of the square pulse, with a pair of square .
waves. The latter function, when differentiated, gives the grating of
alternately phased delta functions.
Figure 41 shows a further example in which it is shown that an
isosceles triangle convolved with an appropriately spaced pair of

~OlvedWitl1 n
~ ~
= .=

]j1:G.41

oppositely phased delta functions, is equivalent to a doublet pulse


(the derivative of the triangle) convolved with a rectangle.

The Differential Operator h(s)


We have seen that in the application of the differentiation theo-
rem we obtain the transform of the derivative by multiplying the
transform of the original function by 2 n i O. Now it is apparent that
this operation represents the procedure that we follow on one pair
of variables in the application of the convolution theorem. 2n i 0 is
obviously a function of 0 which we may call H (0) and we may there-
fore write:
Fourier transform of [G(O) X H(O)] = Convolution of g(x) with h(x)
which is a statement of the convolution theorem. The left hand
side of this expression is the Fourier transform of [G(O) X 2n i 0]
and by the differentiation theorem this is equal to

(ddX ) [g(x)]
76 FOURIER TRANSFORMS

where g(x) is the Fourier transform of G(O). Thus we may write:

ddx g(x) = Convolution of g(x) with k(x)


and we see that the act of differentiation is equivalent to convolving
the function that is to be differentiated with another function k (x)
which is the Fourier transform of 2::n; i 0.
It remains to find the Fourier transform of 2::n; i 0. The function
2::n; i 0 is most disinclined to converge to either side of zero and is a

FIG. 42

perfectly asymmetrical or odd function; we would expect its trans-


form to be also odd, but, as with the transform of a step function, it
cannot be obtained directly and it could never exist, in its true
form, as a Fraunhofer diffraction pattern.
If we plot the function H (0) = 2::n; i 0 we see that it is a straight
inclined line passing through the origin and extending to infinity in
both directions (Fig. 42). We may readily obtain its transform pro-
vided that we are permitted to limit the function before it reaches
infinity. What is the effect of such a limitation and is it permissible in
the circumstances of its application to impose a limit on the func-
tion?
The simplest way to limit the function H (0) = 2::n; i 0 is to stop
it abruptly at a given value of 0. Performing this operation we obtain
the simple triangular doublet whose transform was studied earlier
in this chapter (see Fig. 39). We saw that the transform could be
obtained by differentiating a function of the form sin X/X.
DIFlfERENTIATION 77
Returning to the problem before us, it is apparent that we may
obtain the derivative of any function by convolving it with a differ-
ential operator
h(x) = ~ [sin x]
dx x
provided that the limitation imposed by restricting H (0), the trans-
form of h (x), over a finite range of values is permissible in its effect

"A
"""~~
A<>_
osfl-oo

Fw.43

upon the operation. If we study the effect of the restriction we find


that the greater the range of values used and the nearer that 0
approaches infinity in both directions, the narrower is the function
h(x) until, in the limit, it tends to the form shown in Fig. 43, in
which there is an infinitesimal displacement from left to right be-
I
~

o d 2d

I
-~

FIG. 44

tween the vertical lines above and below the axis. For a finite
restriction to be valid, it is necessary that the oscillations of h (x)
should be much more rapid than the variations in the function upon
which it is to operate to form a derivative, this implies that H (0)
must extend further than the limits of the transform of that function.
It is apparent that h (x) may take on a variety of forms, all that is
required is that H (0) should pass linearly through the origin and re-
78 FOURIER TRANSFORMS

main linear over the entire range of the function upon which it is to
operate. It is evident that this approximation is most nearly correct
in the case of the extended triangular doublet, as the limit to the
linearity then occurs discretely at a distance which may be simply
controlled, but we may also obtain a function which closely approxi-
mates to 2 n i () by transforming the doublet impulse shown in Fig. 44.
Let us consider this doublet impulse. It commences at x = 0
whereupon it rises instantaneously to the level h(x) = lld 2 remain-
ing there until x = d when it reverses to the level h(x) = -1/d2 at
which it remains until x = 2d when it instantaneously falls to and
remains at zero.
The Fourier transform, H(()), of the doublet impulse may be
obtained as follows:

We may now find the limit as d -+ 0 by expanding the exponentials


and neglecting the terms of higher power, thus

i - {[ 1
H(()) = - - -
2n()d2
+ 2n~. ()d + ----
(2ni()d)2]
2!
-1

-[1+4ni()d+ (4n;~d)2J

+ [1 + 2ni()d + (2n;t d )21}

= 2n~d2 {(2ni ()d)2 _ (4n;d)2}

= i {_ 4n2i2()2d2} = 2n i ()
2n()d2
DIF]t'ERENTIATION 79
which is the result required. Comparing this with the Fourier trans-
form of the doublet pulse in Fig. 26, we see that the approximation
implies that, in the limit, the part of that function running through
the origin is very nearly linear.
The doublet impulse is the most elementary and usual differential
operator but the foregoing discussion shows that in some circum-
stances it may be advantageous to use, by convolution, the oscilla-
tory operator h (x) obtained either by transforming a triangular
doublet or by differentiating sin X/X.

The Integral Operator


In a similar manner to that used in deriving the differential opera-
tor, we may derive an integral operator by using the fact that inte-
gration on one side of the transform relationship is equivalent to
division by 2 :n; i 0 on the other side. The operator which we so derive
is, as we might expect, the simple step function and though there
is, accordingly, nothing novel in the result, it provides an interesting
exercise and 'check on the validity of the previous analysis.
Division by 2:n; i 0 is equivalent to multiplication by 1/(2:n; i 0)
and we have previously seen that S(O) = (1/2:n; i 0) is the Fourier
transform of a step function. We may now embark on the sequence
of operations set out below
F.T.
(ddX ) g(x) G (0)

! integrate !X [S(O) = 2:n;\O]


g(x) G(O) S(O)
[= F.T. of {G(O) X S(O)}]

!
= convolution of [F. T. of G (0)] with [F. T. of S (0)]

[= Step Function]
80 FOURIER TRANSFORMS

Thus the convolution of

with a step function is the original function g (x). See, for example,
Fig. 45., in which a step is convolved with the derivative of a doublet

___ ~-.- Convolved with

Step
~
t. g(x) I
= ILr-
g(x)

FIG. 45

to reproduce the original doublet. The constant part of the step


integrates to zero everywhere except within the doublet from the
contributions of the two positive delta functions balancing the one
negative but doubly powerful delta function.

~ Convolved wilh~ = ~conVOlved wilh~-,-_


Step. f (x) f.- g(x)1 g(x) i!-f(x)

FIG.46

From this we see that the derivative of a function is that function


which, when convolved with a step, yields the original function.
Alternatively, we may state that the process of integration may be
treated as the convolution of the function with a simple step func-
tion, though it should be noted that this operation is restricted to
functions which do not extend to infinity. Similarly, if the range of
the function is finite we may use a convolution with a limited step
function, provided that this step is a correct representation over a
range of
d'
x ~( g(x) dX)
and provided that we confine our analysis to the behaviour of the
convolution in the vicinity of the commencement of the step. This
property is often used in electronic integration where the step may
be the leading edge of a repetitive square wave.
DIFFERENTIATION 81

It will be appreciated that this last example is a simplification of


the general convolution differentiation relationship. The situation is
correctly represented in Fig. 46 in which it will be seen that the
doublet pulse is convolved with a delta function derived from differ-
entiating the step function. As ~his leaves the doublet unchanged,
the previous diagram, Fig. 45, represents the final form.
CHAPTER VII

THE AUTO-CORRELATION FUNCTION


AND THE TRANSFER FUNCTION
OF A SYSTEM LINEAR IN INTENSITY

IN Chapter V we encountered the self-convolution of a function. Let


us suppose that g (t) is a function of time, then we may write for its
self-convolution
~

j g(t) g(t - T) dt

If we normalize this function by dividing it by the value it has at


~

T = 0, i.e. j [g(t)]2dt we may write


~

j g(t) g(t - T) dt
e (T) = --~-----

j[g(t)]2dt

The function e(T) is known as the auto-correlation function for the


time function g (t). It expresses the relationship between the values
of g(t) at any two instants of time separated by the interval T. It is
a very powerful operator for use in systems involving partially ran-
dom processes.
The auto-correlation function contains no information about the
phase of the original function and is therefore not unique insofar as
the same autocorrelation function may be obtained from a variety
of generating functions. This result is apparent from consideration
of an elementary generating function such as
g(t) = A sin (2n It + (]J)
THE AUTO-CORRELATION FUNCTION 83
where A is the amplitude of the simple sinusoid and (]J expresses its
position or phase. .
If now we form the auto-correlation function

! A sin (2:>t I t + (]J) A sin (2:>t I t - 2:>t /,r + (]J) dt


e('r) = -~ 00

![Asin(2:>tlt + (]J)]2dt
The expansion of the numerator gives
00

A2 ! rt cos (2:>t I -r) - ~- cos (4:>t It - 2:>t I -r + 2 (]J)] dt


in which the integral from the second term is zero over the range
from - to
00 Similary, the expansion of the denominator gives
00.

00

A2! [t - f cos 2 (2:>tlt + (]J)] dt.


Where, again, the contribution from the second term is zero. Hence
e(-r) = cos2:>t/-r
This expression shows that, not only has all phase information
been lost, but the amplitude term has also cancelled out. Thus the
auto-correlation function of a sinusoidal wave is a unit cosine.
It is apparent that all sinusoids of identical period but arbitrary
phase and amplitude, will generate the same unit cosine auto-corre-
lation function. Thus if the generating wave is a cosine of amplitude
B we may refer to it as a sine wave B sin (2 n It + 90) and we see
that the same result is obtained. Alternatively, if we work through
this example as we did for the sine wave, we have
g(t) = B cos (2n I t + (]J)
00

! B cos (2:>t I t + (]J) B cos (2:>t It - -2:>t I -r + (]J) dt


e(-r) = -00 00

! [B cos (2:>t I t + (]J)]2 dt


84 FOURIER TRANSFORMS

The expansion of the numerator gives

B2 f H cos 2 n 1 + i
l' cos (4 nit - 2 n 1l' + 24] dt
and the expansion of the denominator gives

B2 f [i + i cos 2 (2 nit + 4] dt
In both expressions the integrals of the second terms are zero over
the range - 00 to 00, that is to say they produce no d. c. component,
and hence we obtain the previous result
(1(1') =cos2n/1'
If we consider a slightly more complex case, in which the generat-
ing function is the beating pattern produced by the sum of two
sinusoidal waves of different periodicity A sin (2 nit + 4 and
B sin (2n p t + VI), then a similar analysis shows that the auto-
correlation function is given by
1
(1(1') = A2+ B2(A2cos2n/1' +B2 cos2np1')

Thus the contributions due to the two components are proportional


to their respective powers in the original generating function. An
extension of this analysis into any number of simple generating
sinusoids shows that the power contributions from each appear in the
auto-correlation function without regard to their phase. As the ele-
mentary sinusoids may be looked upon as the Fourier components
of a more complicated distribution, it is apparent that the auto-corre-
lation function is related to the power spectrum of the generating
function. For a formal proof of this relationship we turn to the
Wiener-Khinchine theorem.
Writing the auto-correlation function in its usual form
00

fg(t)g(t+T)dt
(1(1') = --00-00----

f [g. (t)]2 d t
THE AUTO-CORRELATION FUNCTION 85
it is evident that we may substitute for the numerator the expression
. given by the convolution theorem:
00 00

j g(t) g(t + or) dt = j G(f) G(f) e2niJldi

It is also apparent that the denominator


00

j[g(t)]2dt

must be equal to

j [G(f)]2 dl

in order that the energy in the alternative expression may be con-


served. This relationship is attributed to Lord Rayleigh and may
be obtained from Parseval's Theorem.
Substituting the new expressions in the relationship for (!(or) we
have 00

j[G(f)]2e bi/1 dl
(!(or) = -00 -00----
j[G (1)1 2 d 1

This is the Wiener-Khinchine Theorem, which states that the auto-


correlation function of a function is the Fourier transform of the
power spectrum [G (/)]2 of that function.
Thus, for example, if the original function is a rectangular pulse.
This has a Fourier transform which is of the form G (f) = (sin xix) in
amplitude and phase space. The power spectrum is given by [G (/)]2
= (sin2 X/X 2 ), and the auto-correlation function is the Fourier trans-
form of this function. We have previously seen that the Fourier
transform of sin 2 X/X 2 is a triangular distribution having a base twice
the length of the original rectangular pulse (Fig. 25).
If we take a series of randomly spaced rectangles it is apparent
from the previous discussion that the auto-correlation function does
not register the phase components corresponding to the positions of
the pulses, whereas the powers from the pulses are additive. The same
86 FOURIER TRANSFORMS

reasoning applies to any random distribution of similar functions


and has practical importance in the powder photograph used in
crystallography and the reinforcement of the diffraction pattern
from a single particle of lycopodium powder when many particles
are sprinkled at random over a surface.
The autocorrelation function of a perfectly "white" noise spectrum
is a single line existing only at the origin. This is a statement of the
fact that in a pure random process there is no correlation between the
various events. The events in the white noise spectrum are infinitely
narrow elementary transients which would independently yield an
auto-correlation function of a simple vertical line.
In general the auto-correlation function is a symmetrical function
whereas the generating function may be symmetrical or asymmetri-
cal. From a consideration of the equivalence of the auto-correlation
function to the self convolution of the generating function it is also
apparent that the auto-correlation function is in general broader
than the generating function.

The Transfer Function of a System Linear in Intensity

It is well known that an optical system illuminated by incoherent


light responds linearly to the intensity of the illumination. Whilst in
the case of coherent illumination it is possible, for example, to com-
pletely cancel the contributions from a pair of emitting points, in the
case of incoherent illumination the intensity of each additional point
source at the object is linearly transferred to the output, for the
contributions from the points are completely uncorrelated. There can
be no interference other than that of one point interfering with
itself and the intensities due to independent points are additive.
The fact that the intensities may be superposed immediately
suggests that the Fourier transform may be applied to incoherent
distributions. Performing this operation, we obtain the transform
as a virtual pattern which may be combined with the intensity transfer
function* of the apparatus used to detect it.
* See footnote on page 4.
THE AUTO-CORRELATION FUNCTION 87

Let us consider a point source illuminating an optical system which


is free from defects other than the limitations of a finite aperture. * If
we examine the image of the system we shall find that it is not a
point but a diffraction pattern determined by the aperture of the
system. As the image is an intensity pattern it corresponds to the
power spectrum which is obtained by multiplying the Fourier
transform of the aperture by its complex conjugate. If we Fourier
transform the actual distribution of intensity in the image we shall
expect to arrive back at the equivalent aperture function for inten-
sity. Thus if we transform the power spectrum rather than the
amplitude spectrum we obtain the transfer function of the optical
system insofar as it determines the behaviour of the system to the
illumination of an incoherent distribution of light. In a similar
manner to that in which the aperture function determines the beha-
viour of the system to the amplitude and phase pattern where the
existence of coherence permits the source illumination to be allotted
an amplitude and phase di!;!tribution we may use the transfer func-
tion when there is no coupling between the individual contributions
to the illumination and it therefore fulfils the requirements of a
system linear in intensity. Where coupling is present** between the
individual contributions the cross beats formed upon detection of
the radiation can also contribute to give a real output, whereas in
* It is often helpful to consider systems from the point of view of their elec-
trical analogue. Thus the circuit in Fig. 19 (c) (Chapter V) represents the simplest
electronic analogue of an incoherently illuminated optical system; more complex
systems may be represented by the incorporation of delay lines, additional con-
nections to the detector, and extra filter units having suitable responses to simulate
their optical counterparts. The transfer function in Fig.53 may be considered as
the spectrum (including negative frequencie3) of sum and difference beats obtained
in the output of a square law detector (Fig. 19) when fed with noise via a band-
pass filter. The conjugate image of the pass-band corresponds to the other aperture.
** The Intensity radiation pattern of a coherently illuminated aperture is given
by multiplying the amplitude radiation pattern by its complex conjugate. Thus
the intensity radiation pattern, or power polar diagram, of a uniformly illuminated
rectangle will be given by a sin2 X/X2 function, and the system behaves as though
we had an "intensity aperture" with a triangular distribution on a base twice as
long as the true aperture. The radiation pattern from the coherent aperture will
be directed in a true beam, so that apparatus used to observe the radiation will
be sensitive to the absolute, as well as the relative, position of the detecting aper-
tures.
88 FOURIER TRANSFORMS

the incoherent case they are high frequency terms of random phase
and amplitude which integrate together to give zero real contribu-
tions.
The relationships between the intensity transfer function and the
other parameters of the optical system are set out below.
Amplitude and Phase
Aperture F.T. distribution of point source
response (Fraunhofer am
plitude pattern.)
x
convolved with itself complex conjugate

Transfer function (auto. F.T. Intensity distribution


correlation function or self of point source response
convolution of aperture.) (Power spectrum or Fraun.
hofer intensity pattern.)

We shall derive the relationships formally for a two dimensional


system.
In an incoherent system which is linear in intensity, if the aperture
function is given by f(x,y) then the amplitude distribution in the
Fraunhofer pattern is F(O,fP) where F(O,fP) is the two dimensional
Fourier transform:
F(O,fP) = f f f(x,y) e2ni(x8+y(I)dx dy
The corresponding intensity distribution in the Fraunhofer pattern
is I(O,<P) = F(O,<P). F*(O,<P) where the* denotes complex conju-
gate. If T (x, y) is the transfer function of the apparatus, then

and by the convolution theorem this may be written

T(X,y) = f f f(x,y) f*(x - x, y - y) dx dy


THE AUTO-CORRELATION FUNCTION 89
Thus the transfer function is the self convolution of the aperture
and is the function that we would obtain by moving the aperture
over itself, and integrating the overlap, see, for example, Fig. 33
which shows a circular aperture convolved with itself, or Fig. 34
which shows the whole sequence of operations set out in the table
on the opposite page.
As with the amplitude pattern we see that the highest Fourier
component that can enter the system is limited by the aperture but
the weighting of the components between this limit and the com-
ponents of zero frequency is now determined by the convolution
integral (the transfer function) and not the direct distribution of the
aperture. We may therefore proceed in the manner used for previous
treatments of the amplitude function of coherent systems if, in the
case of incoherent radiation, we replace the aperture function by the
transfer function and work in intensity rather than amplitude.
In simple optics we may consider that there is only one term
involving the Fourier component between the limits of the aperture,
and, although this has a counterpart in the negative frequency
spectrum, there is only one real contribution. The number of Fourier
components present which are smaller than the aperture will increase
as the components become narrower, for we may locate a corres-
pondingly larger number of pairs of points within the aperture which
define the limits of these components. The difference is here appa-
rent between the intensity response to incoherent illumination and
the situation existing under coherent illumination. The number
distribution of the components follows the convolution integral, giv-
ing usually a peaked transfer function such as that in Fig. 33.
A single aperture illuminated with incoherent illumination will '
be seen to behave in a manner precisely analogous to that of a low
pass filter fed with random noise, where the output of the filter is
passed to a square law detector. The analogy may be extended to
show that the high frequency response (resolution) is determined by
the limb or periphery of the aperture whilst the low frequency res-
ponse is more dependent upon the centre (see footnote on page 87).
It may seem that the convolution involved in the transfer func-
tion doubles the resolving power of the instrument by moving the
90 FOURIER TRANSFORMS

limits of the aperture twice as far apart. This is only an effect of the
symmetry of the complete system, including negative frequen-
cies. Only one-half of the symmetrical convolution may be con-
sidered to yield a real part. The function only indicates the distribu-
tion of the numbers of the Fourier components and not their spatial
position. As in the case of the autocorrelation function, the phase
information determining the position of the individual components
contributing to the distribution has been lost.
We may use the transfer function to operate upon another func-
tion which, for similar reasons, has no material existence. This func-
tion is the Fourier transform of the distribution of intensity across an
incoherent source. The value of combining these functions is de-
monstrated by the analysis of an interferometer system included in
the following examples.

EXAMPLES ON CHAPTER VII


The Telescope*
The objective lens of the telescope has a linear aperture distri-
bution f (x) which is obtained by projecting a circle into a line;
we can find from a list of known transformations that this gives rise
to the Fourier transform shown in Fig. 47, where F(O) = J] (0)/0
and J 1 (0) is a Bessel Function. ** F (0) represents the amplitude
reception pattern of the telescope and is the amplitude pattern
which a point source moving relative to the telescope would produce
if observations were made only in the centre of the field of view. The
intensity distribution which would be seen would correspond to the
same curve with the ordinates squared.
If we fixed our gaze on another part of the field of view the path
lengths to different parts of the aperture would vary linearly and
* For a detailed treatment of the transform applied to radio telescopes and
large aerial systems, reference may be made to R. N. BRAOEWELL, AU8tralian
Journal 01 PhY8ic8 9, 297, 1956.
** A plot of J 1 (0) is somewhat similar in appearance to that of a decaying sine
wave. The function J 1 (0)/0 is tabled in the appendix. The linear aperture distri-
bution refers to the relative sensitivity in the x direction. This is a maximum at
the centre where all points on a diameter in the y direction contribute to the
sensitivity in x.
THE AUTO CORRELATION FUNCTION 91

the aperture distribution would have the associated phase function


shown in Fig. 48. We have seen that a phase function of this type
causes a shift in the reception pattern so that the transit of the
source now appears to be advanced or delayed according to the
direction of tilt Gf the phase function and according therefore to the
direction of our gaze.

FIG. 47

If the telescope remains fixed on an axis near to the point source,


i. e. if we keep it pointing straight along the axis at () = 0 and then
sweep our gaze over the field of view we shall introduce a variable
phase shift and rotate the angle of the phase vector in Fig. 48, so
that the reception pattern again appears.

f(x)
---
.......-

FIG. 48

We have already mentioned that the human eye, or any photo-


metric device which we use to study this pattern, will respond to the
intensity rather than the amplitude pattern to which we have been
referring, and hence we shall actually see a pattern IF(()) 12 which
will appear as in Fig. 49 (a).
If another point source of the same intensity is radiating at an
angle ~ () from the first source at () = 0, then it will independently
contribute the pattern 1F (() + ~()) 12 shown in Fig. 49 (b).
92 FOURIER TRANSFORMS

If both sources are viewed simultaneously and if they are inco-


herent, the eye will see a superimposition of both the patterns, as
in Fig. 49(c). If the two points radiate in fixed relative phase, i.e. "
if they are coherent, then the pattern may be stronger than that
resulting from incoherent points due to the additional real contri-
bution of the cross product terms, and the shape of the pattern in
this case will depend upon the actual phase
(0) IF(e1l 2 as well as the disposition of the points. They
will combine to form an amplitude pattern
r; at the image and from this we may then
obtain the actual intensity pattern that is
seen. Clearly, if we add more radiating points
(b) IF(e+8eIl 2
to the source we may build up the pattern
-19
which is seen from the sum of the separate
amplitude and phase patterns due to each
point. If the points are radiating inco-
herently, then we simply add the intensity
patterns. If now, for the sake of simplicity,
(e) Sum we consider only the incoherent system and
we refer to the angular intensity distribution

FIG.49
-
19 of a source consisting of any number of radia-
tingpoints as I(fktP ),we have the relationship :
~ ~

Pattern in field of view = J JI((),tP)P(() -1X,tP -f3)d()dtP

where P((),tP) is the power polar diagram or Fraunhofer intensity


pattern of the telescope objective, and IX, f3 is the angle of view
(transferred from the eyepiece to the corresponding angle made by
the line of sight at the aperture).
Now we have seen that we may use the convolution theorem to
re-write the expression for the pattern in the field of view
~

J J I((), tP) P(() - IX, tP - f3) d(), dtP


~ ~

= J f i(x, y) 'Z'(x, y) e 2ni (",8+yd)"dx dy


THE AUTO-CORRELATION FUNCTION 93

This equivalent expression is the Fourier transform of the pro-


duct of the telescope intensity transfer function with a function
. i (x, y) which may be considered as that virtual diffraction pattern
or distribution of intensity along the line of the telescope aperture
which would project back into space an angular distribution I(O,<P)
and hence illuminate a portion of the celestial sphere with a star of
this pattern. Note that although the virtual pattern i(x, y) is mea-
sured in intensity it is permitted to have an associated phase and
can exhibit negative values.
If T (x, y) is constant or only very slowly varying over the whole
range of i (x, y), then we may remove T (x, y) from within the integral
sign and place it outside as a constant. Hence the pattern that is seen
in the eyepiece corresponds to a constant times the Fourier trans-
form of the virtual diffraction pattern of the source, and this ob-
viously is equal to a constant times the angular intensity distribution
of the source. Hence a telescope of very large aperture will present
an image of the source which is a true picture of the angular inten-
sity distribution of that source.
Now consider the case where the telescope transfer function T(X,y)
is smaller than i(x, y), the transformed angular pattern of the star.
From the summation property of transforms we may separate the
product transform into two separate transforms, one of which con-
tains those parts which multiply to give zero and the other trans-
form embraces the contribution where T (x, y) exists.
The first transform is obviously zero, whilst the second cannot be
distinguished from that which would exist if T (x, y) were large and
i(x, y) were reduced; such a distribution transforms back into an
angular distribution which is not a true likeness of the source: This
is shown in Fig. 50 in which the diagrams are confined to x and 0
for the sake of clarity. The sequence of operations referred to starts
in the top left hand corner and returns via the right hand column
back to the observed pattern in the bottom left hand corner. It will
be seen that we may arrive at the same result by working straight
down th~ left hand column, i. e. by taking the convolution of I (0, <P)
with P(O, <P). The longer sequence of operation employing the trans-
forms on the right of Fig. 50 is therefore somewhat redundant but
94 FOURIER TRANSFORMS

it is worth close study for in certain instruments, notably interfero-


meters, the function i (x, y) is that which is actually derived.
/'
,
f'V\1(J) ,,
,,,
-
(J -.................. -,~

x
-
x

Convolved with

-
(J -
x

= =
,
,,
,,:' i{x)r(x)
,,
-
(J
FIG. 50
-
x

The alternative approach, using the convolution of the intensity


reception pattern with the intensity distribution across the source,
is illustrated as a two dimensional convolution in Fig. 51 where the
left hand diagram shows the source distribution for a uniformly illu-
minated disc, the centre diagram shows the intensity reception pattern
of the telescope, and the right hand diagram shows their convolution.

The ordinate scale corresponds to intensity or brightness. The dis-


tortion of the image due to the use of a finite aperture is clearly
visible in the diagram.
THE AUTO-CORRELATION FUNCTION 95

The resolving power of a telescope with a circular aperture at the


objective is usually quoted as a = 1'~2 A., where A. is the wavelength
of the radiation and d is the diameter of the aperture. It will be
appreciated that this formula is a statement ofthe transform relation-
ship between the angular seperation to the first zero of the reception
pattern J 1 (())/() and the distance in wavelengths across the aperture
of the instrument. Its efficacy as a criterion is dependent upon the
equality of the intensity of the radiation between the points to be
resolved.

Stellar Interferometers
(i) The Michelson Stellar Interferometer
This instrument* consists of two apertures of small width spaced
apart by a much larger distance which may be preset as required
(Fig. 52a). Incident light from a star falls upon the slits and is then
focused through a telescope system on to the eye or a photographic
plate. The image seen by the eye is the diffraction pattern due to the
separate apertures crossed by their mutual interference fringes. The
radio frequency analogue system consists of two small aerials spaced
apart by a large number of wavelengths and connected together via
equal lengths of cable at the input of a single receiver (Fig. 52b). In
this case, if the two small aerials are identical, the record traced by
a point source is that given by the polar diagram of the individual
aerials modulated 100 % by the cosinusoidal fringe pattern produced
between them. If the apertures or aerials are unequal the modulation
of the pattern is not complete and is limited in extent to the region
defined by the product of the amplitude diffraction patterns of the
apertures or aerial systems. The fringes disappear when the amplitude
pattern of either aerial reaches zero and reappear in opposite phase
where the second maximum due to the larger aerial lies upon the first
maximum due to the smaller aerial. In general the pattern will be
seen to be related to the aperture by a Fourier transform relation-
* MICHELSON. Phil. Mag., I), 30, 1, 1890.
96 FOURIER TRANSFORMS

ship. It is in fact the power spectrum of the aperture. In the classical


Michelson stellar interferometer the apertures are made precisely
similar and the fringe pattern produced by a point source therefore
gives rise to 100 % modulation of the diffraction envelope, the light

\, '\
; ;
\ ,
,-----, ;!'-----J,

<--\---=>,:,'1
\
\\, :'
/
,I
(a)
,
\ ,
;
\

oQ
( )2 Detector

Integratar

Recarder

(b)

FIG. 52

rises to double the mean intensity on the crests of the fringes and
falls to zero in the troughs. When two incoherent point sources are
radiating and these have a small angular separation, they each pro-
duce independent fringe patterns in the image. The relative phase
of these ,fringe patterns depends upon the path lengths from the
points to the two apertures of the system, it therefore depends upon
the spacing of the apertures and the spacing of the radiating points.
If either of these spacings is adjusted, the two fringe systems may
THE AUTO-CORRELATION FUNCTION 97

be made to occur ~n antiphase so that the total illumination is


approximately constant, varying only with the overlapping diffrac-
tion patterns of the separate apertures and no fringes are seen. At
this spacing of the apertures the combined source is said to be re-
solved into two components and Michelson's criterion for this resolu-
tion is given by his fringe "visibility" formula
V = Imax - I m1n
Imax +I m1n

Where V is the visibility of the fringes, Imax is the bdghtness of the


peaks of the fringes and I m1n is the brightness in the troughs. When
the modulation is 100 % the value of V is unity and it becomes zero
when the modulation disappears due to resolution of the source. The
spacing of the interferometer apertures corresponding to this reso-
lution is related to the separation of the two sources by the formula
IX = ).j2d where A is the wavelength of the light.

If the source is a continuous distribution of incoherent radiators


such as the surface of a star, the individual points distributed over
the surface of the star each give rise to a faint fringe system and
these integrate together in the image giving a brighter pattern which
will exhibit fringes except when the spacing of the apertures of the
interferometer, or the diameter of the source, is such as to cause the
individual fringe patterns from the elementary radiating points to
integrate out in the final image. In these circumstances the visibility
function goes to zero at a value of the spacing of the interferometer
apertures corresponding to the extent of the source, and the diameter
of the source may be determined from this spacing. Resolution occurs
for a rectangular or "slit" source at a spacing of the interferometer
apertures given by IX = A/d. When the source is a uniformly illumi-
nated disc the value is given by IX = 1.22 A/d.
These values will be recognised as those corresponding to the first
minima of the Fourier transforms of the intensity distributions across
the respective sources, and, in fact, the variation of the visibility of
the fringes as the apertures of the interferometer are separated by
increasing amounts follows directly the Fourier transform of the
intensity distribution across the source. Thus for a rectangular star
98 FOURIER TRANSFORMS

the visibility would fall off from unity (100 % modulation) when the
apertures or aerials were very close together to a value at any other
spacing given by the corresponding point on a sin xix function. In
the case of a uniformly illuminated disc the fringe visibility would
follow a (J1 X)/i law. In the secondary maxima of these functions
and in the general case where the source is asymmetrical, the varia-
tion of phase in the Fourier transform appears as a shift in the posi-
tion of the final fringe pattern in the image. Thus the fringes reverse
in the secondary maxima of the (J1 X)/x pattern from a disc source
and the centre of the whole pattern from the star is a dark trough in-

Aperture function convolvedwithitself=Transfer function

FIG. 53

stead of a bright fringe as it would be within the first maximum. By


using a large number of separations for the interferometer apertures
and plotting the complete visibility function including this phase
component, we may obtain the actual brightness distribution or
structure of the source by Fourier transforming the oberservations
back again into an angular distribution.
A formal analysis of the operation of the Michelson stellar inter-
ferometer may be made in a large variety of ways but the most com-
pact approach is provided by the transfer function. As the light from
the star is incoherent we may obtain the transfer function by taking
the self convolution of the aperture. The self convolution of two
equal slits is shown in Fig. 53. It consists of three triangular func-
tions, the central function has twice the amplitude of the other two
and the spacing of each of the outer triangles from the centre corres-
ponds to the spacing between the two apertures of the interferometer.
To obtain the output of the interferometer we simply multiply this
function by the virtual amplitude diffraction pattern of the star across
the plane of the aperture. The Fourier transform of the resulting
function is the pattern that is actually seen at the chosen spacing
THE AUTO-CORRELATION FUNCTION 99

of the apertures (Fig. 54). It is immediately seen that the fringe


modulation will disappear at a spacing of the apertures which corres-
ponds to the distance to the first minimum of the Fourier transform
of the intensity distribution across the star. At this spacing the pro-
ducts formed by the outer triangles of the transfer function are zero
and only the Fourier transform of the central triangle will appear
Virtual amplitude Intensity
diffraction pattern distribution
of star across star

o
x
Transfer function of
interferometer

./S.
Observed
intenSily
pattern
Product

A A ... e
FIG. 54

in the output. Thus when the source is resolved the observed image
pattern will be a sin 2 X/X 2 distribution of intensity unadulterated by
modulating fringes.

(ii) The Switched Interferometer


An interesting and useful variation of the Michelson Interfero-
meter is the switched interferometer introduced by Martin Ryle
and used extensively in radio astronomy*. In this system the two
aerials forming the apertures are rapidly switched through an extra
half wavelength of cable so that they are alternately connected in
phase and in antiphase, Fig. 55 (a). In synchronism with the aerial
RYLE, M., Proc. Roy. Soc. A 211, 351, 1952.
100 FOURIER TRANSFORMS

switching the output of the receiver is reversed and then integrated


so that only the difference in the reception of the system in the two
states is observed. Analysing this system by means of the transfer
function we see that in each successive half cycle of the switching
operation the output of the detector alternates between the two
transfer functions shown in Fig. 55 (b). As the output of the receiver
is reversed during the second half of the switching cycle, and the two
states are added by integration, the effective transfer function of the
complete equipment is that formed by the outer triangles only.
Thus the output of the recording meter at any spacing of the aerials
is given by the Fourier transform of the product of the Fourier trans-
form of the intensity distribution across the source with the transfer
function just derived. This is shown in Fig. 56. It will be observed
that with this system the background radiation is removed, the
Michelson visibility criterion may not be directly applied and the
solution for a particular source distribution requires the measure-
ment of the relative magnitude of the observed patterns at various
spacings to that obtained at 'zero spacing. When a source is com-
pletely resolved, no output, other than random noise, is obtained on
the recording meter. To reconstruct the actual distribution of inten-
sity across a star or cosmic region it is necessary to move the inter-
ferometer apertures to all spacings at which a fringe output is ob-
servable and then by working backwards in Fig. 54 or Fig. 56 it will
be seen that we can arrive back at the distribution of intensity across
the source. A unique solution to this problem can only be obtained if
the phase of the virtual amplitude diffraction pattern of the star is
also recorded. This information appears as the relative position of
the central maximum of the fringe pattern to that which it would
occupy at zero spacing of the apertures and implies a measurement
of the apparent position of the source. Where the experimental diffi-
culties are such that the positional accuracy and phase information
would normally be lost the information may be retrieved by using
three apertures, combining these to give three separate fringe systems
and suitably combining the arguments of the fringe patterns.*
This technique also enables the measurement of the amplitude of the
* JENNISON, R.C., Monthly Notices Roy. Astron. Soc. 118, 276.1958.
THE AUTO-CORRELATION FUNCTION 101

Delector
SvnChronlsed

Reversing
switch

Integrator

Recorder

(0)

Aperture
first half cycle. ~
fun onvolved with itself =

secondho~cycfe ~
fun~Convolved ."h Itsett.
Aperture

Effective transfer
function ofter
~
reversing and integrating

(b)

FIG. 55

Virtual amplitude Intens~y distribution


diffraction pattern across star
ofstor

n
Effe~tive transfer
/
_ \ function

Observed intensity
PractJct f\ pallern
A ....-rtr.."- O~

FIG. 56
102 FOURIER TRANSFORMS

transform of the source to be made without recourse to reducing the


aperture spacing to zero.

(iii) The Post-Detector, or Intensity, Stellar Interferometer


This instrument was originally conceived* and constructed**
as a radio interferometer but has been adapted to form an optical
interferometer*** with an attainable resolution considerably higher
than that of the Michelson type. It is free from the effects of scin-
tillation and vibration, the apertures may be of poor optical quality
provided that they have a large light gathering power, and the
connections between the apparatus at the two apertures may be
made with simple cables, radio links or even separate synchronized
recordings fed into the central analysing system at a later date.
An elementary analysis of this interferometer may be made by
considering the star or radiating source composed of pairs of points
which may beat together to produce low frequency signals in the out-
puts of the detectors in the radio frequency equipment (Fig. 57 a) and
in the outputs ofthe photomultipliers in the optical version, (Fig. 57b.)
The outputs of the two channels are brought together and their corre-
lation coefficient is determined by a low frequency interferometer in
which the independent powers in each channelt as well as the cross
product power is recorded. If the pairs of points are integrated over
the complete source distribution it will be seen that the correlation
coefficient measured at the output will correspond to the ordinate
on the power spectrum of the intensity distribution of the source
corresponding to the spacing of the interferometer apertures on the
abscissa. Complete resolution is obtained where the spacing is such
as to correspond to a zero of the power spectrum (virtual Fraunhofer
intensity pattern) and at this spacing no correlation is observed.
The post detector interferometer is ideal for the measurement of
* HANBURY BROWN, R., and TWISS R Q., Phil. Mag. 40, 663, 1954.
** JENNISON, RC., and DAS GUPTA, M.K., Phil. Mag. 1, 55, 1956.
*** HANBURY BROWN, R., TWISS, R Q., Proc. Roy. Soc. A24S, 291, 1957, also
Nature 178, 1046, 1956.
t In the optical version the r. m. s. fluctuation level is recorded, to yield a simi-
lar index, in lieu of the two intensity recorders.
THE AUTO-CORRELATION FUNCTION 103

the angular diameter of small intense stars but the signal to noise
ratio falls off rapidly for fainter stars. It is not suita~le for the

~
Aerials

Amplifiers

Detectors
(r

Recorders

Integrat ion

Recorder

(o)

Linear Reversing
Photo multiplier switch Photo
multiplier multiplier

Mirror
"--/Mirror
: Synchronisation

Integrating
motor

ROIl counter
(flue tuations)

(b)

FIG. 57

measurement of the distribution of intensity within a cosmic source


as the phase information is lost and a unique reconstruction of the
source structure cannot be made.
104 FOURIER TRANSFORMS

The transfer functions of the equipment may be used to illustrate


its behaviour as shown in Fig. 58. The star "projects" a virtual
Fraunhofer amplitude pattern upon the ground and this may be
multiplied by the simple transfer functions of the separate apertures

Independent transfer X
function of chamel A A....____ sin 2X Transit recorded
Xl on chaMei A

PrOdUC_' _ _ _ ..,1\....____

JL
Source distribution
Independen' 'mnsf., . X
function ofchonnel B A...._ _ __
Product
_ _- - J/ \' -_ __

c=> '
~
PI ~
&
XI X
Relative complex
tronsfer functions
ofAandB
AIA
~ '-- ~ ' -
=I =
~~~~~c::
multiplier
In ---"-i ~
'-.,.-'
Transit recorded
on cross-product
meter

Transfer function after


multiplier

FIG. 58

and then transformed to find the form of the response on the two
meters or counters recording the independent channels. The cross
product channel gives an output which corresponds to the Fourier
transform of the real function formed by convolving together the
conjugate functions, corresponding to the products of the complex
transfer functions of the apertures, with the same virtual Fraunhofer
amplitude pattern of the star. Note that the baseline of the instru-
ment, d, is half the spacing between the two complex transfer func-
tions.
THE AUTO-CORRELATION FUNCTION 105

The abstract nature of the transfer function in this example does


not detract from its remarkable power as an analytical tool to pro-
vide a rapid solution of the problem. The simple analysis in Fig. 58
shows directly the actual variation of the readings on the three re-
corders whilst it is obvious that the source distribution can be deter-
mined by finding the value of e2 from a division of the readings on
the cross product chart with the product of the other two readings.
The lack of uniqueness of the solution will be recognised in the ambi-
guity of taking the square root to determine the visibility, or ordi-
nate of the Fourier transform of the source distribution, e. The
phase information is lost in the convolution of the complex transfer
functions.
***
Note on the Fourier Synthesis of Apertures
The instruments described on the previous pages may be used to
measure the Fourier components at various spacings of the element-
ary apertures. Provided that both the amplitude and phase are
recorded and the radiation does not vary over the period of the
measurement, we may synthesise the equivalent continuously distri-
buted aperture of a very large instrument. The additional information
is obtained at the expense of the increased time required to make
these observations.
It is important to note that, if no assumptions are made as to the
distribution of intensity in the source, the information obtained with
an interferometer is limited to those Fourier components which
correspond to the spacing of the apertures. No conjuring with the
optics or electronics after the signals have entered the apertures can
therefore replace the missing components. Thus, for example, the
ability to resolve a pair of points is not increased by introducing
non-linear devices into each limb in order to combine the harmonics
of the signals.
APPENDIX I

ANAI.OGUE COMPUTERS
OF FOURIER TRANSFORMS

THERE is a very large variety of computers which may be conceived


to solve the Fourier transforms of the functions encountered in the
course of experiment. The adoption of a particular technique will
usually depend upon the nature of the transform under investigation
and the accuracy required in the result. A brief reference is given
here offour representative techniques. It is worth noting that digital
computation may be preferred for most final solutions but the ana-
logue techniques give insight into the behaviour of the solutions
when variations are performed upon the generating function.

(i) The Diffraction Oomputer


This instrument was deveioped by Lipson for the solution of pro-
blems in crystallography. The optical arrangement is shown in Fig.59.
Parallel light is passed into a chamber into which a slide may be

Lamp Collimating Lens Slide Lens Microscope


hole

FIG. 59

introduced. The light is then focused on to a microscope and either


photographed or examined by eye. If a slide is introduced consisting
of a punched card with the holes arranged in a pattern which is
thought to be the likeness of the crystal structure, the resulting
diffraction pattern seen in the eyepiece may be compared with the
ANALOGUE COMPUTERS 107

X-ray diffraction photograph. A discrepancy in the diffraction


patterns indicates an error in the model and a variation may then be
tried. The use of polarized light and mica discs in the punch holes
permits the control of the phase of the elements in the reciprocal
process.
The technique is excellent for X-ray diffraction but its use is
limited in other fields. Shading of the brightness distribution of the
pattern and smooth variations of phase such as are frequently re-
quired in problems in other spheres are extremely difficult, if not im-
possible to obtain. The measurement of the intensity distribution
of the image which is required in most other problems would necessi-
tate the examination of the image by means of a microphotometer.
The system is coherent and yields the square* of the two-dimensional
transform of the punch card on the photographic plate.

(ii) A Mechanical Oomputer

Mechanical computers may be designed to reconstruct the distri-


bution of intensity across an object from the available (interfero-
Drive
---J

FIO.60

metric) data on the amplitude and phase plots of the transform of the
source. The computer sketched in Fig. 60 was modified from an air-
craft position indicator.
* The product of the complex conjugates.
108 FOURIER TRANSFORMS

A small electric motor rotates the shaft marked "drive" and moves
a cursor in front of a wire model bent to form a plot of the experi-
mentally derived complex function f(x, (/J). The distance of the wire
from the axle on which it is mounted determines the amplitude of
the function at any value of x measured along the axle, whilst the
phase is adjusted by bending the wire to the corresponding angle in
the perpendicular plane to the axis at each value of x.
An infinitely variable gear 01' consisting of a disc and ball bearing,
may be set to any desired value of 0 by adjusting the position of the
ball. This varies the relative rate of rotation of the wire function
f (x, (/J) to the motion of the cursor in the x direction and hence
corresponds to different angular inclinations of radiating points in
the transform F(O) of f(x, (/J).
The cursor moving in the x direction is directly coupled, via the
main driving shaft, to the disc of a second infinitely variable gear O2
which acts as an integrating device. The position of the ball bearing
on this disc is adjusted by hand so that the second cursor may be
tracked with the first cursor to keep theirjntersection lined up with
the rotating function f(x, (/J) as the value of x increases. A mirror is
mounted behind the wire function to enable the cursors to be correctly
aligned. The ball-bearing on the gear O2 is coupled to a revolution
counter which records the integral of the function for each value of
O. The gear 0 1 is adjusted to give 0 as many values as desired and.
for each value the tracking operation is performed and the integral
is recorded on the revolution counter. A plot of the readings on the
counter for various values of 0 constitutes the Fourier transform of
the function f (x, (/J). The accuracy of this technique is of the order
of a few per cent.

(iii) A Ooherent Electronic Oomputer


A method developed by Levy* employs a bridge circuit in which
a matched artifical delay line of 80 sections forms one arm of the
bridge (Fig. 61). The bridge is fed initially from an impulse genera-
tor; with the line perfectly matched and the bridge balanced this
should give no output. The 80 capacity elements in the line may now
* LEVY, J. Inst. Elec. Engrs. 00 (III), 153, 1943.
ANALOGUE COMPUTERS 109
be adjusted to cause partial reflections at their respective delay
times and the magnitude of these reflections is adjusted so that the
overall time response exhibited on the cathode ray tube follows the
shape of the required function I (t). The transform F (I) of this func-
tion is simply determined by replacing the impulse generator by a
continuous oscillator, varying the frequency of this oscillator and
plotting the corresponding deflection on the cathode ray tube. If
desired an automatic scan may be obtained from the oscillator and

Generator.
FIG. 61

synchronised with the horizontal scan of the cathode ray tube to


give directly the frequency spectrum.
A suitable delay line may be constructed from a continuous tapped
inductance with preset condensers at the tapping points. The delay
per section is given by T = 0.4/ Ie where Ie is the cut-off frequency of
the line. The inductance of each section of the line LB = Z/(2nle),
the capacity 0 = 04/leZ, the mutual inductance M = 01 LB'
A'suitable value for the line impedance, Z, is the order of a few
hundred or a'thousand ohms. If it is too high the effects of strays will
be troublesome.

(iv) An Electronic Oomputer with an Incoherent Source


Figure 62 illustrates a technique which is useful for obtaining the
transforms of smoothly varying functions, such as those which may
be analysed numerically by the superimposition of Gaussian func-
tions.
no FOURIER TRANSFORMS

Noise at video frequencies is generated in a photomultiplier tube,


thyratron, or neon stabilizer, amplified over a broad band and fed
through a choice of electrical filters. The circuit then divides into
two variable delay lines which feed a multiplying circuit. The direct
current output, of this circuit actuates a meter.
The method of operation is to introduce a filter circuit having the
desired frequency characteristic and then plot the readings on the
output meter as a function of the setting of the delay line.

MUltiPher~-D-Q
Filter mnnrrrrrrtnlmTrmrrnTrYm1-----l Amplifier
and
Integratar

FIG. 62

The output meter may be of the recording variety and the paper
chart can then be synchronised to the setting of the delay line by
a simple stepping device. In principle, the system will operate with
one delay line, but two lines arranged so that the relative delay
changes whilst the net delay is constant are preferred in order to
cancel the variation in the loss of the line.
The mUltiplying circuit consists of an aperiodic phase reversing
switch in one input feeding one grid of a mixer valve. The other in-
put feeds directly on to the second mixer grid. It is advisable to
arrange a subsidiary circuit to cancel the non-linearity of the valve
characteristic so that only the alternating component due to the
product of the two inputs appears at the output of the mixer. The
alternating output is amplified, synchronously rectified, and passed
to the final meter.
ApPENDIX II

J 1 X (X in radians)
X

X 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

0 05000 04990 04975 04943 04900 04846 04778 04700 04610 04510
1 04401- 04280 04152 04015 03870 03719 03561 03398 03230 03058
2 02883 02706 02527 02347 02167 01988 01810 01635 01463 01294
3 01130 00970 00816 00668 00527 00392 00265 00145 00033 -00069
:'"d
4 -00165 -00251 -00330 -00399 -00460 -00513 -00557 -00593 -00621 -00642 '"d
I:':l
5 -00655 -00660 -00660 -00652 -00639 -00620 -00596 -00568 -00536 -00500 l2i
t:i
6 -00461 -00419 -00375 -00330 -00283 -00236 -00189 -00142 -00095 -00050 H
~
7 -00006 00035 00075 00113 00148 00180 00209 00235 00258 00277 H
H

8 00293 00305 00314 00320 00322 00321 00317 00310 00300 00287
9 00272 0.0255 i 00236 00215 00193 00169 00145 00120 00094 00069
10 00043 00018 -00006 -00030 -00053 -00075 -00095 -00114 -00131 -00147
11 -0016(t- -00172 -;-0.0182 -00189 -00195 -00198 -00200 -00199 -00196 -00192
12 -00186 -00178 -00168 -00157 -00145 -00132 -00118 -00102 -00087 -00070
13 -00054 -00037 -00020 -00003 00012 00028 00043 00057 00071 00083
14 0.0095 00105 00114 00122 00128 00133 00136 00138 00139 00138
15 .....
I 00136 00133 00128 00122 00115 00107 00098 00089 00078 00067
-
.....
.....
112 FOURIER TRANSFORMS

TABLE OF THE INTEGRAL Si (X)


X Si(x) X Si(x) x Si(x) x Si(x)

00 00000 50 15499 100 16584 150 16182


01 00999 51 15313 101 16525 151 ] 6223
02 01996 52 15137 102 16460 152 16258
03 02985 53 14973 103 16388 153 16287
04 03965 54 14823 104 16311 154 16309

05 04931 5.5 14687 105 16229 155 16326


06 05881 56 14567 106 16144 156 16336
07 06812 57 14462 107 16056 157 16340
08 07721 58 14374 108 15965 158 16337
09 08605 59 14302 109 15874 159 16328
-
10 09461 60 14247 110 15783 160 16313
11 10287 61 14209 11-1 15693 161 16292
12 11081 62 1-4187 112 15604 162 16266
13 11840 63 14182 113 15518 163 16234
14 12562 64 14192 114 15436 164 16197
---

15 1.3247 65 14218 115 15357 165 16157


16 13892 66 14258 116 15284 166 16111
17 1-4496 67 14312 117 15216 167 16063
18 15058 68 14388 118 15154 168 16011
19 15578 I 69 14457 119 15098 169 15957
---

20 16054 70 14546 120 15050 no 15901


21 16487 71 14644 121 15009 171 15844
22 16876 72 14751 122 14976 172 15786
23 17222 73 14864 123 14950 173 15729
24 17525 74 l-4983 124 14933 174 15671
-----
25 17785 75 15107 125 1-4923 175 15615
26 18004 76 15233 126 14922 176 15560
27 18182 77 15361 127 14929 177 15507
28 18321 78 15489 128 14943 n8 15457
29 18422 79 15617 129 1-4965 179 15410
- --

30 18487 80 15742 130 14994 180 1.5366


31 18517 81 15864 131 15029 181 15326
32 18514 82 15981 132 15071 182 15291
33 18481 83 16093 133 15119 183 16260
34 18419 84 16198 134 15172 I 184 15234
APPENDIX II 113
TABLE OF THE INTEGRAL Si (X) (continued)

x Si(x) x Si(x) x Si(x) x Si(x)

35 18331 85 16296 135 15229 185 15213


36 18220 86 16386 136 15291 186 15197
37 18086 87 16467 137 15335 187 15186
38 17933 88 16538 138 15423 188 15181
39 17765 89 16599 139 15492 189 15181

40 17582 90 16650 140 15562 190 15186


41 17387 91 16691 141 15633 191 15197
42 1-7184 92 16721 142 15704 192 15212
43 16973 93 16739 143 15773 193 15232
44 16758 94 16747 144 15841 194 15257

45 16541 95 16745 145 15907 195 15286


46 16325 96 16732 146 15970 196 15319
47 16110 97 16708 147 16030 197 15336
48 15900 98 16676 148 16085 198 15395
49 15696 99 16634 , 149 16136 199 15438
siny'
X

y'degrees 0 1 2 3 4 5 6 7 8 9
.....
.....
0 14'-
100000 100000 099971 099961 099928 099873 099818 099754 099670 099586
10 099495 099385 099269 099144 099005 098861 098707 098537 098363 098178
20 097980 097776 097562 097334 097101 096857 096601 096339 096067 095783
30 095492 095192 094881 094562 094233 093895 093547 093192 092855 092454
40 092073 091682 091281 090873 090457 090031 089598 089156 088705 088248
50 087782 087308 086826 086337 085839 085334 084822 084301 083775 083241
60 082699 082150 081595 081033 080463 079889 079306 078717 078123 0-77521
70 076914 076302 075682 075057 074427 073791 073150 072503 071850 071194
80 070531 069864 069193 068516 067835 066924 066460 065766 065069 064367 "':I
90 063660
o
062953 062240 061523 0608041 060081 059355 058628 057896 057161 q
100 056425 055686 054944 054201 053455 052708 ~
051958 051207 050455 049701 ....
110 048945 048189 047431 046673 045914 045154 044393 043633 042872 042110 f<I;J
~
120 041349 040588 039827 039066 038306 037547 036788 036030 035273 034517 >-3
130 033762 033008 032256 031506 030757 030010 029265 028522 027781 027103 ~
140 026314 025572 024841 024112 023386 022655 021944 021228 020515 019805 ~
00
150 019098 018395 017696 017001 016317 015622 014938 014259 013584 012913 >.Jj
160 012247 011602 010929 010277 009629 008987 008350 007717 - 007088 006496
o
~
170 005852 005241 004635 004032 003442 002853 002270 001694 001123 000558 i:(
00
180 000000 -000552 -001098 -001638 -002172 -002699 -003219 -003734 -004241 -004742
190 -005236 -005723 -006204 -006678 -007144 -007604 -008057 -008503 -008943 -009373
200 -009781 -010215 -010625 -011025 -011423 -011811 -012192 -012566 -012932 -013290
210 -013641 -013985 -014321 -014650 -014991 -015285 -015591 -015889 -016181 -016464
220 -016740 -017008 -017269 -017522 -017768 -018006 -018236 -018475 -018674 -018882
230 -019083 -019275 -019471 -019639 -019809 -019972 -020127 -020275 -020415 -020549
240 -020674 -020793 -020904 -021008 -021105 -021195 -021277 -021352 -021420 -021482
250 -021536 -021583 -021623 -021656 -021683 -021703 -021716 -021722 -021722 -021715
260 -021702 -021682 -021655 -021618 -021583 -021538 -021487 -021417 -021365 -021298
270 -021220 -021139 -021051 -020958 -0-20859 -020755 -020645 -020530 -020409 -020286
- - I - ---- ----
sin)!
(continued)
x
X degrees 0 1 2 3 4 5 6 7
8 I 9
280 -020151 -020015 -0:19873 -019726 -019575 -019418 -019257 -019091 -018920 -018745
290 -018558 -018381 -018192 -018000 -017803 -017602 -017397 -017188 -016976 -016759
300 -016545 -016316 -016089 -015858 -015625 -015388 -015148 -014905 -014824 -014410
310 -014158 -013904 -013647 -013387 -013125 -012861 -012595 -012326 -012056 -011783
320 -011509 -011232 -010970 -010675 -010394 -010111 -009828 -009543 -009256 -008969
330 -008681 -008392 -008102 -007811 -007519 -007228 -006933 -006643 -006350 -006056
340 -005763 -005470 -005177 -004878 -004591 -004298 -004006 -003714 -003423 -003132
350 -002842 -002553 -002265 -001978 -001691 -001406 -001122 -000840 -000558 -000278
360 000000 000276 000552 000826- 001098 001368 001636 001902 002166 002428
370 002689 002946 003201 003454 003706 003954 004200 004443 004683 004921 I>-
380 005156 005389 005618 '"d
005845 006068 006289 006506 006721 006932 007140 '"d
390 007345 007547 007745 007940 008131 008319 008504 008685 008863 009036
t;:j

400 009207 Z
009373 009536 009696 009851 010003 010149 010295 010435 010572 t::t
410 010705 010833 010958 011079 011196 011309 011418 011523 011625 011721
>-I

420 ><
011814 011903 011988 012068 012145 012218 012287 012351 012412 012468 >-I
>-I
430 012521 012569 012613 012654 012690 012722 012750 012775 012795 012811
440 012823 012832 012836 012837 012833 012826 012815 012800 012781 012758
450 012730 012702 012668 012630 012600 012544 012495 012443 012388 012329
460 012266 012192 012130 012057 011981 011901; 011818 011732 011643 011551
470 011455 011356 011255 011150 011042 010932 010818 010702 010583 010462
480 010337 010210 010080 009948 009814 009677 009539 009388 009251 009105
490 008957 008806 008644 008499 008343 008183 008024 007862 007698 007532
500 007365 007197 007026 006855 006681 006509 006331 006154 005976 005797
510 005617 005435 005253 005070 004886 004701 004516 004330 004143 003956
520 003768 003580 003391 003202 003013 0.02824 002635 002445 002256 002068
1
530 001876 001687 O 0149 001310 001121 0.00933 000745 000558 .....
000371 000185 .....
540 000000 1 Ot
INDEX
Aberration 65 Campbell and Foster 41
Acceleration 40 Capacity 3
Acoustics 4, 37 Carrier 72
Admittance 6, 48 Cascaded systems 66
Aerial 34, 58, 64, 90, 95 Cathode Ray Tube 32,109
Amplifier 3, 60 Celestial sphere 93
Amplitude 3 et al. Chopper (light) 53
distribution 13 Cis; cisoid; cisoidaI16-18, 21, 30, 41, 42
function 25, 89 Coherence: coherent illumination 14,
response 31 35, 36, 38, 86, 92
Analogue computers 106 Collimating point 28
Analyser, wave 41 Collimation; collimator 24, 35
Angular frequency 6, 8 Comb filter 28
response 64 Complex conjugate 9, 88
Anticipatory ring 62 Complex distribution 24
Aperture 3 et al. Complex function 6, 108
acoustical 37 Complex plane 16, 21, 24
circular 65, 67 Complex transfer function 104
distribution 34, 46 Computer 41, Appendix I
function 34, 38, 86, 87 Condenser 56
synthesis 105 Conjugate variables 5
Argument (of fringes) 100 Convolution 6 et al.
Array, aerial 34 integral 4, 9, 53, 54, 74
Asymmetrical transform 26 Correlation 46, 102
Auto-correlation function 4, 82, 83, 84, coefficient 102
85 Cosine 16 et al.
Cosmic radiation 38
Baffle 40 Criterion (of resolution) 95, 97
Beam 12, 13, 28 Cross beats; cross product 88, 92, 104
Beats 88 Crystallography 86
Bed of nails 40 Crystal structure 106
Bessel Function 39, 90, 111 Current 6
Bracewell, R.N. 90
Braddick, H. J. J. Preface Das Gupta, M.K., Jennison, R.C., and
Brightness distribution 64, 98 102
Browne, I. c. Preface Delay line 52, 108, 110
INDEX 117
Delta function 15 et al. Epoch 6
DeplOdulator, ring 50 Error function 30
Detector 51 Even and odd functions 8
Differential operator 75, 79
Differentiation 9, 68 et al. Filon, L.N.G. 43
Diffraction pattern 3 et al. Filter
computer 106 angular 46
grating 35 comb 28
screen 53, 66 electrical 46-50, no
virtual 93, 98, 100 low pass 71, 89
X-ray 107 Foster, Campbell and 41
Disc, illuminated 39, 94, 98 Fourier
rotating 53 components 2, 84 et al.
vibrating 37 Integral 5, 41, 62
Discharge, electrical 14 Pairs 5
Dispersion 52 series 2, 5, 62
Distribution, amplitude 13 synthesis (of apertures) 105
angular 93 Frequency changer 3
aperture 34 Frequency response 31, 32, 48, 52
brightness 64, 98 Fraunhofer 3 et al.
complex 24 Fringe
cylindrical 66 diffraction 60
frequency 64 interferometer 8, 24, 35, 36, 95, 97
intensity 4, 36, 38, 90, 97, 99
number 89 Gaussian 30, 34, 40
of field 34 Generating function 17, 39, 82, 84, 86,
of illumination 45 106
phase 13 Gibbs phenomenon 62
probability 4 Grating 28, 35, 74
pulse 32 function 72
source 21, 100, 105
spatial 4 Hanbury Brown, R., and Twiss, R. Q.,
Dominant harmonic 42 102
Doppler 37, 40 Harmonic, dominant 42
Doublet 30 et al. Helix 16
triangular 30, 72, 76 Hexode mixer 50
Huygen's principle 46
Echo sounding 37
Electrical discharge 14 Image 66, 87, 95
Electronics 4 space 65
Electronic apparatus 31, 52 Imaginary plane 21, 69
Electronic integration 80 Impedance 32
Energy, conservation of 23 Impulse 25, 32, 54, 60
U8 I~DEX

Incoherent noise 4 Object Hpace 65


Incoherent radiation or light 4, 38, 46, Objective 90, 93
67, 86-88, 96 Odd and even functions 8, 28, 59, 62,
Inductance 3, 72 75
Integral operator 79 Optical interferometer 3
Integration, electronic 80 Optics; optical instruments; systems 4,
Intensity 11, 12, 86, 87 53, 60, 64, 65, 87, 88
distribution 4, 36, 38, 90, 97 Oscilloscope 31, 53
interferometer 102 Overshoot 62
transfer function 46, 93
Interferometer 8, 24, 38, 59, 90, 94, 95 Parse val's theorem 85
optical 3 Path difference 11
post-detector (intensity) 102 Pattern
switched 99 radiation 32
Isosceles triangle 30, 53, 56, 72, 75 reception 90, 94, 95
Pendulum 32
Jenkins and White 35 Period, periodicity 2, 11, 12, 14, 84
Jennison, R.C., 8, 100 Phase 3, et al.
and Das Gupta, M.K., 102 distribution 13, 46
function 16, 21, 25, 91
L. C. circuit 32
information 83, 100, 105
Lens 65, 86
response 31
Levy, J., 108
rotation 8, 12, 14
Light chopper 53
space 16
Limitation 45, 76, 87
Photocell 4
Linear Multiplier 50
Photographic plate 4
Lobe 58
Photomultiplier 102
Lord Rayleigh 61, 85
Polar diagram 34, 46, 59
Lycopodium 86
power 92
Mate 5, 19,40,45, 69 Polarity 26, 42
McLachlan, N. W., 40 Post detector interferometer 102
Michelson 38, 62, 95, 99, 102 Powder photograph 86
visibility criterion 97, 100 Power polar diagram 92
Microphotometer 107 Power spectrum 4, 84. 85, 87, 102
Mixer, hexode 50, 51 Pressure 40
Modulation 28 Prism 36
Modulus 26 Probability 38
Multiplier, linear 50 distribution 4
Product 45, 53
Nails, bed of 40 Pulse 31, 32, 37
Negative exponent 5 retangular 2, 28, 85
Negative frequencies 7, 8
Non-periodic events 5 Quasi-monochromatic 46
INDEX 119

Radiation pattern 32, 34 Space


Radio astronomy 8 phase 16
Radio telescope 90 Spatial distribution 4
Random noise 89 Spectral line 3, 10, 14
Random process 82, 86 Spectrum
Ratcliffe, J. A., Preface spectral function 4 et al.
Rayleigh, Lord 61, 85 power 4, 84, 85
Rayleigh refractometer 35 Square pulse (see also rectangle) 56
Real form (of transform) 6 Square wave, repetitive 74, 80
Real plane 26 Star; stellar 8, 38, 46, 95, 103
Real values 21 Step function 62, 70, 72, 76, 79, 80
Reception pattern 90, 94, 95 Straight edge, see step function
Reciprocity 12 Summation 19, 44, 71
Rectangle; rectangular pulse 2 et al. Superposition 19, 44, 68
Reference, axis of 17, 24 Switched interferometer 99
Refractive index 35 Symmetrical function 24
Refractometer (Rayleigh) 35
Resistance 3, 56 Telescope 36, 46, 60, 65, 90, 95
radio 90
Resolution; resolving power 65
Television 65
Response
Thermocouple 4
angular 64
Time function 52
frequency 31, 32, 52
time 48,60 Time response 48, 60
Transducer 37
Ring, anticipatory 62
Transfer function 4 et al.
Ring demodulator 50
complex 104
Rotating vectors 7, 8
Rotation, phase 8, 12, 14 Transistor 3
Transmission 65
Ryle, Martin 99
line 16
Trapezium 53
Self convolution 51, 82, 86, 89, 98
Triangle, isosceles 30, 53, 56, 72, 75
Sidebands 72
Triangular
sin X/X function 2 et al. doublet 30, 72, 76
sin X/X function, table of 114 function 34, 85
sinc function, see sin X/X function
spectrum 51
sine integral Si(x) 60, 62, 71, 72 Tuned circuit 32
table of 112
Twiss, R. Q., Hanbury Brown, R. and
sine wave 11, 12, 16, 23, 25, 28, 59, 83
102
Sneddon, LN., 43
Source distribution 21, 100, 105 Valve 3
Source function 17 Vector 7, 16, 26
Space Virtual pattern 60, 65, 93, 98, 100, 102,
image 65 104
object 65 Voltage 6
120 INDEX

Wave White
analyser 41 spectrum 49
front 11 Wiener-Khinchin Theorem 84, 85
function 28
train 28 X-ray diffraction lO7
Wavelength 11, 13, 28, 34
White Young's slits 35
light 3
noise 46, 86 Zero frequency 5, 42, 89

Vous aimerez peut-être aussi