Experimentation An Introduction To Measurement Theory and Experiment Design DC Baird

Experimentation
Experimentation:
Prentice-Hall, Inc.
Englewood Cliffs, New Jersey
An
Introduetion
to
Measurement
Theory
and
Experiment
Design
D. C. Baird Associate Professor of Physics, Royal Military College of Canada
EXPERIMENTATION: AN INTRODUCTION TO MEASUREMENT

THEORY AND EXPERIMENT DESIGN .
BY D. C. BAIRD.
1962 by PRENTICE-HALL, INC.

All rights reserved. No part of this book
may be reproduced in any form, by
mimeograph or any other means, without
permission in writing from the publisher.
Current printing (last digit):

15
14
13
12
11
10
Library of Congress Catalog Card Number 62-11883

Printed in the United States of America
29534C
Prefaee
This text is intended for use in first and second year physics
laboratory courses for scientists and engineers. The function of such courses varies widely, and this function is
changing with time as new approaches to physics education
are tried. However, regardless of the actual aim which the
designer of a laboratory course has in mind, one point remains invariable. That is that experiments involve measurement. Unless the nature of measurement is clear to the
experimenter, the benefit which is sought from carrying
out the experiment cannot be fully realized. This text is
written, therefore, on the premise that whatever the purpose being served by the laboratory course, consideration
of measurement theory and elementary experiment design
should not be entirely neglected. In any case, the emphasis
on experiment design in engineering is growing to the point
where an introduction at elementary levels is becoming
essential, and this text is intended to contain the material
necessary for this work.
For such a purpose, it is not enough merely to insist that
physics laboratory students perform routine error calculations; the process as a whole has to make sense. and sufficient detail is included here to ensure that it does. The
v
vi
PREFACE
text is not intended to be a complete treatise on industrial

statistics or on numerical methods in observation processing. It is merely an introduction, because the writer feels
that it is necessary for the student to acquire, as a first
step, a personal feeling for the nature of measurement and
uncertainty, leaving mathematical sophistication to be acquired later. The amount of mathematical work has, therefore, been kept to the minimum which makes the treatment
understandable. Examples have been added to some of the
chapters because experience has shown that the average
laboratory program does not provide the necessary amount
or variety of exercise in the mechanics of observation processing.
The writer wishes to express his grateful thanks to all
those who have helped in the assembling of this material.
In particular he wishes to thank Dr. T. M. Brown and
Mr. C. D. Pearse, the Royal Military College, who have
provided invaluable assistance in discussing the principles
on which it is based and Mr. A. J. Filmer, also of Royal
Military College, who checked the answers to the problems.
D. C. B.
Contents
Introduction te Laboratory Work 1
The Nature of Measurement 6

2.1
2.2
2.3
2.4
2.5
Measurement and Confidence, 6

Types of Uncertainty, 10
Distribution Curves, 16
The Gaussian or Normal Distribution, 26
Correspondence between the Normal Distribution and
Actual Observations, 28
2.6 Significance of the Standard Deviation in Actual
Measurements, 30
2.7 Sampling, 31
2.8 Practical Considerl)tions, 38
2.9 Rejection of Readings, 40
2.10 Numerical Example, 41
Problems, 46
The Propagation of Uncertainties 48

3.1
3.2
3.3
Absolute and Relative Uncertainty, 48

Propagation of Uncertainty, 49
General Method for Uncertainty in Functions
of a Single Variable, 51
3.4 Uncertainty in Functions of Two or More Variables, 54
3.5 General Method for Uncertainty in Function
of Two or More Variables, 56
3.6 Compensating Errors, 60
3.7 Standard Deviation of Computed Values : General Methods, 61
3.8 Standard Deviation of Computed Values: Special Cases, 64
. 3.9 Combination of Different Types of Uncertainty, 67
3.10 Application of Results, 68
Problems, 69
The Nature of Experimenting 75

4.1
4.2
Experiment Planning 88
5.1
5.2
5.3
5.4
5.5
5.6
Precision of Measurement, 89
Experimenting with No Background, 91
Dimensional Analysis, 95
Experimenting with a Theoretical Background, 100
Graphical Analysis, 105
Experiment Analysis and Design, 111
Problems, 117
Experiment Evaluation 123

6.1
6.2
6.3
6.4
6.5
6.6
6.7
6.8
Nature of Scientific Theory, 75

Types of Experiment, 79
The Aims of Experiment Evaluation, 123

Computation of Elementary Quantities, 125
Graphs, 127
The Validity of the Experiment, 128
The Calculation of the Answer and Uncertainty, 130
The Principle of Least Squares, 133
Function Finding, 140
Over-all Precision of the Experiment, 142
Problems, 146
The Scientific Report 152

7.1
7.2
7.3
7.4
7.5
7.6
7.7
Aims in Report Writing, 152

General Principles, 153
Introductory Material, 155
Procedure, 158
Results, 161
Graphs, 162
Discussion, 164
Laboratory Practice 167

8.1
8.2
Appendix
The Laboratory Note Book, 167

Experiment Technique, 169
The Statistics of the Gaussian or Normal

Distribution 177
Al .l The Equation of the Normal Distribution Curve, 177
A1.2 Standard Deviation of the Normal Distribution, 183
A1.3 Areas Under the Normal Distribution Curve, 183
Least Squares Fit to Straight Lines 186

Bibliography 189
Problem Solutions 192
Index 196
Introduction to Laboratory Work
I t is the aim of this introduction to explain some of the

purposes in physics laboratory work for the student of
physics or engineering. So much time is spent either in the
laboratory or on the work of the laboratory that value for
the time spent will be gained only if the student understands clearly the purposes of the laboratory course and the
means by which these purposes are being achieved. This
understanding can best be found through a study of the
nature of progress in science and technology, and the realization that one of the main steps in scientific or technological advancement is the experimental one. The nature
of experiments will be discussed more completely in Chapter 3 and it will suffice for the moment merely to remind
the reader of the vast significance of, and consequences
from, experimental work. This gives observation a dominating role in almost every field of human activity. Consider, too, the enormous range of topic covered, from the
well-defined, strictly controlled experiment of the physicist
to the work of the biologist or physiologist, dominated to
a large extent by statistical fluctuation, or the observations
of the astronomer whose measurements may be precise but
whose degree of control over his subject matter is somewhat
INTRODUCTION TO LABORATORY WORK
limited. It is obvious, therefore, that it is useless to lay

down a set of rules to say in general how experiments
should be done: the range of activity is far too wide. However, there are habits of thought which will prove useful
whether the observer is studying rabbits, stars, or tar in a
road surface mix. It is the aim of this book to describe
some of these principles and to show how the student may
use introductory or secondary laboratory work to become
practiced in such arts of experimental investigation as will
be useful, no matter what line of work is ultimately to be
pursued.
The important part of the approach is that the range of
experimental technology is so vast that it is completely
futile to attempt to prepare the student for every contingency. The chances are overwhelming that the graduate
engineer or scientist will meet situations in his work which
are new to him. Thus, one of the most important single
factors in his education should be the development of the
capacity to cope with new situations, and this book is a
study in the methods whereby this can be accomplished.
Furthermore, professional research or technological development work is usually difficult through its very nature.
This is natural, since the necessity for experimental work
will always arise from the existence of a problem. Therefore, the future scientist or engineer must become accustomed to accepting practical work as a problem. He
must gain practice in working at such problems until a
way of solving them experimentally is found. He must, in
addition, become capable of solving these problems on his
own, since, unless he is capable of independent work, his
worth to his employer is lessened. It must be clear that the
necessary practice in solving problems experimentally i~

not gained by a laboratory system which provides the stu
dent with extensive and detailed instructions for the performance and calculation in the experiment. As will be
seen later, the major part of experimental investigation is
the preliminary analysis and the final evaluation of the experiment. If the experiment is properly planned, the actual performance can be reduced to a simple collecting of
observations. Most of the thought goes before and afterwards. To permit the student to bypass these, the most
important parts of the experiment, is to give him a totally
wrong impression of the importance of the performance of
the experiment and to deny him the opportunity of acquiring the real skill of experimenting. This capacity for
independent thought is one of the most important single
qualities of a research or development worker and he will
acquire the necessary judgement only through the exercise
of judgement.
This, then, suggests the attitude which the student should
have towards his laboratory work. He should regard each
experiment as a model of a problem which might be encountered in actual research or development. To be of any
use at all the experiment must be a new situation and it
must be a problem. Therefore, the student must expect to
have to work at his laboratory problems, otherwise the
time is wasted. He will be expected to make his own decisions about the method of making the measurements so
as to achieve maximum efficiency of information collection.
These decisions may often be wrong, but the student will
learn more effectively in retrospect, and he will be strengthened by having made his own decisions. He will have to
work within the framework of the apparatus provided,

since he must learn that the skill in experimenting lies
largely in achieving the maximum experimental yield with
the resources available. Restrictions of time, too, merely
simulate the conditions of most actual research or development. The experiments will never be ideal. However, this
should not be regarded as a defect, but as a challenge. The
real work of evaluating an experiment lies in sifting the
information desired from the yield of the experiment, which
is always clouded to a certain extent by uncertainty. The
experimenter must learn to identify sources of uncertainty
or error and, if possible, eliminate or allow for them. Whatever the degree of control he has over his experiment,
however, he must evaluate the reliability of his result. This
critical evaluation of the experiment is just as important as
obtaining a numerical answer. The ability to cope with
such conditions can be acquired only by experience and it
is a common injustice to the student to leave him with the
impression that the experiments are perfect.
It is necessary to keep an open mind towards experimenting
and not allow an objectively analytical attitude to be
hindered by a preconceived opinion of what "ought" to
happen in the experiment. This open-minded spirit of investigation can be cramped by the setting down of a toodetailed "requirement" for the student to follow. The
emphasis in the teaching laboratory should be on learning
rather than doing.
Report writing should be approached in the same constructive spirit. The necessity for engineers and SCIentists to
express themselves clearly and informatively through the
written or spoken word has been frequently maintained,
INTRODUCTlON TO LABORATORY WORK
and is a matter of widespread concern. It is the writer's

conviction that the acquiring of such fluency by scientistr
and engineers is as much the responsibility of the science
departments as of the arts departments and the medium of
such education is, to a very large extent, the laboratory report. Facility in clear and elegant exposition of scientific
material can only be acquired through much practice, and
this is the purpose in writing laboratory reports. This purpose will not be achieved unless the report writing is taken
seriously as an opportunity to improve one's powers of exposition. A report which degenerates into a mere indication that the experiment has been performed is a waste
of time to student and teacher alike. The student who takes
pains with his reports deserves, and will profit from, careful,
constructive discussion and criticism of his report; the
writer regards such discussion of report and experiment as
an essential feature of the system. The student will learn
from his mistakes only after careful clarification.
Thus, in conclusion, we may say that the laboratory offers
the student the opportunity to acquire many of the skills
connected with the performance of his professional work.
Hewill gain facility in the analysis of problems, the evalu<.tion of the solution achieved, and the ability to describe
his work clearly and informatively for the benefit of others.
These qualities are among the most important tools of his
trade and are well worth acquiring.
The Nature of Measurement
2.1 Measurement and Confidence
The observation of nature which constitutes an experiment

will almost invariably take the form of a measurement.
This is the case whether the experiment is of a precision
type in which the answer is the magnitude of a quantity, or
whether the measurements merely have the purpose of substantiating a qualitative conclusion. Because of this absolutely fundamental role of measurement it is necessary to
consider in some detail what a measurement actually is. A
true understanding of the nature of measurement would
prevent many errors of interpretation which may be perpetrated at the conclusion level.
To answer the question-"What is a measurement?," it is
easier to say what it is not. There are two classes of statement which can be made in human knowledge. The first
is the kind exemplified by Euclid theorems which deal with
humanly constructed definitions, and these are absolutely
incontrovertible. If Euclid defines lines, planes, angles,
etc., he can then say with absolute certainty that the sum of
the angles in his triangle is 180 0 This is exact truth because
it is little more that a restatement of his own definitions. It
6
THE NATURE OF MEASUREMENT
is the kind of statement which a measurement is not. The

other class of statement is concerned, not with precise,
humanly constructed definitions, but with experience. In
attempting to make a statement, not about our own ideas,
but about the external world, we immediately lose the advantage of unequivocal exactness and lay ourselves open to
all the frailties of human judgement. This, then, is the
nature of a measurement; it is a statement oj the result oj a
human operation of observation. It is not a matter of measuring
the length of a box, getting a value of 6 in., and then saying
"the length of this box is 6 in." This last statement can
never be made (unless, of course, the Government introduces legislation to define the inch in terms of our box). In
general the only possible statement that can be made is, "I
have measured the length of this box and get a result of 6
in." This last statement illustrates the extent and limitations of statements made about the external world.
We have stressed the human and fallible nature of measurement and its consequently limited validity. This leads to a
second concept. The days are past when people made
measurements for their own satisfaction. Today we are
almost invariably under the necessity of conveying our experimental results to someone else, either the man in charge
of our work, or else other workers in the field. This n.::cessity immediately raises the question of confidence.
Quite plainly-are people going to believe your measurement or not? It is clear that only your professional reputation can instill confidence in other people but some
measure of the reliability of your measurement is obviously
called for. Convinced of the subjective nature of measurement, we can ask, on being told that an experimenter has
measured the box and produced an answer of 6 in.: Is the
observer a metrologist who means that he has an answer of

6.000000 in., or was it measured by the local drunk who
had lost his glasses and really meant that he thought it was
somewhere between 5 and 7 in.? It should be noted here
that such concern tends to arise only when the measurement reaches beyond the bounds of the immediately familiar. If a painter whom I shall employ to paint my
house asks how long it is, I say "30 feet" and he is happy,
for he knows in general terms how much paint to order.
But if someone at a nuclear reactor research establishment
has measured the thermal expansion coefficient of a new
alloy for cladding reactor fuel elements, those who are going
to use his measurements in new reactor design will inquire very carefully into the reliability of his stated result.
If, therefore, we wish to convince other people of the usefulness of our experimental result, the statement of the
result must be amplified by the quotation of some range of
confidence. The intelligent way to quote the answer would
be, "I find the length of the box to be 5.95 in. with 95 per
cent confidence that it lies between 5.90 and 6.00 in."
Note that this still is merely an expression of opinion by the
observer and if real confidence is to be justified, sufficient
description of the mode of measurement must be given to
allow the reader to form his own judgement of the value of
the measurement. Our inebriated friend without his
glasses might use his imagination and claim a measurement
of 5.8279436 in., but no one would believe him. Moreover,
there have, in fact, been many instances in physics where
work, even by very distinguished physicists, has been shown
later to contain errors much larger than the limits of un-
certainty quoted by the original authors. The only thing

one can do is to give as complete a description of the experiment as is feasible and, relying on your own professional
reputation, give your suggestion as carefully as possible for
the limits of confidence. The remainder of this work will
be largely concerned with the methods available to do this
It can be said that the confidence one has in a measurement
and the closeness of the limits set for the uncertainty are
inversely related. For example, I have a right to be as
certain as any human being can be that the length of this
room is between 1 m and 10 m but my faith in the statement that it lies between 5.478339 m and 5.478340 m is
limited. In connection with this topic it is of interest at
this point to note another aspect of the statement that confidence and precision are inversely related. It is commonly
said that there is a distinction between observations permitting a unique, exact answer (normally of a counting
nature), such as the number of people in this room, and
those which are subject to uncertainty, such as the length
of the room. However, the distinction is one of degree only,
because most physical measurements have initially the
character of a counting process (I can lay down my meter
stick 5 complete times in succession along the wall of the
room) and the uncertainties (and declining confidence)
appear only when one tries to specify the limits more
closely (is the length between 5.4 and 5.5 m? Or, is it, even
more unlikely, between 5.4783 m and 5.4784 m?). This
problem does not arise in many counting type experiments
such as that mentioned above because they deal essentially
with units only: one does not subdivide people.
10
2.2 Types of Uncertainty
The methods by which the uncertainty can be estimated

do not depend on particular sources of discrepancy but it
will be instructive before proceeding to clarify our ideas by
reference to a few common situations. The sources of
perturbation of measurements cover such a wide range that
it is possible to do no more than list a few typical headings.
It is common to divide such perturbations into categories
such as random, systematic, personal, etc. However, these
terms are frequently too precise for the average experimental situation and their usefulness is consequently
limited. They are commonly defined as follows:
Random error is said to be shown when repeated measurements of the same quantity give rise to differing values.
Systematic error refers to a perturbation which influences all
measurements of a particular quantity equally.
However, these terms must be used with caution since a set
of readings will show truly random error only if there are a
large number of small perturbing influences. If the discrepancies arise from only a few types of experimental
defect, an analysis into a few competing systematic errors
may be possible and, even more important, the statistical
theory to be described later will not be applicable. On the
other hand, an error which is systematic under one system
of measurement (e.g., a set of ammeter readings all taken
going the same way when the meter bearings are sticky)
may become apparently random if the mode of measurement is changed (e.g., if the meter readings are taken with
arbitrary current changes). We shall use the terms systematic and random to indicate only clear-cut cases.
11
As has been stated above, any measurement is subject to an

enormous range of sources of perturbation. It is impossible
to consider all of these as systematic errors and even the
skilled experimenter will probably give up and leave a residuum of uncertainty to be treated under the heading of
"random" error.
A few factors of common occurrence which limit the precision of a measurement are the following:
(a) Instrument Calibration

Clearly this is of paramount importance, for the observer
is dependent on his scale for the accuracy of his measurement. His meter stick might have warped, his stop watch
changed its rate, his ammeter be sensitive to position. This
type of error is usually so regular as to justify the term
"systematic." Only comparison with a standard instrument will enable this source of uncertainty to be removed
or allowed for, and such comparison should be contemplated wherever a measurement plays a dominating role
in an experiment. Those instruments which have built-in
standards for comparison purposes (e.g., most direct reading potentiometers) are a great convenience and give corresponding confidence in the measurement.
(b) Instrument Reproducibility

Even when the over-all calibration of an instrument has
been checked under a certain set of circumstances, mechanical defects can still influence the readings and can
remove the value of the calibration unless the subsequent
readings are taken in exactly the same fashion as the calibration values. Such defects are slackness and friction in
meter bearings, backlash in micrometer measurements, and
12
many other cases where the final meter indication is de~

pendent on the path by which it is reached. Depending on
,the measurement schedule, this error may be systematic
under very regular circumstances (even zero under some
circumstances, such as a travelling microscope with all the
readings taken going one way) or otherwise the deviations
may be scattered sufficiently to justify the term "random."
Many instruments such as precision ammeters will have a
value marked on them for the magnitude of this uncertainty.
(c) Observer Skill

Even in this day of automation in large-scale research and
development, the operation of scientific measurement is, to
a very large extent, based on personal manipulation of the
apparatus and visual reading of a resultant indication.
Judgement will, therefore, enter at a number of points
depending on the number of variables on which the result
is dependent. This, consequently, magnifies the influence
of uncertainties in such settings and has the result that the
precision of an observer's measurement may be much less
than that suggested by his ability to read the final answer
on the scale. As an elementary example of this, consider
the observation of the position of a spectrum line on a spectrometer. Here the preliminary manipulation consists of
setting the spectrometer telescope cross wires on the
spectrum line, and the measurement is a simple reading on
a vernier scale. If the visual conditions are poor, the
difficulty in the preliminary setting may be so great as to
exercise a dominant influence over the over-all precision of
the measurement. It must be stressed that the influence of
the observer's judgement extends very far beyond the mere
scale reading, and it is this integrated uncertainty which
13
must be considered. Uncertainties of personal origin may

be systematic, as in the case of the astronomer watching a
star in a transit circle and consistently pushing the chronograph button late, or may be so varied in composition as
to justify the term "random."
(d) Miscellaneous Errors
In an experiment involving more than one or two variables or factors which influence the final measurement,
there are bound to be perturbations which influence the
final reading. These perturbations are usually random in
nature, such as line voltage fluctuations, vibration of instrument supports, variable cosmic ray background, etc.
I t is part of the task of the experimenter to reduce the
influence of these perturbations to the minimum, but there
will usually be a residual contribution.
(e) Fineness oj Scale Division
Even with an ideally calibrated instrument under ideal
conditions a fundamental limit is set to the precision of the
measurement by the instrument scale, which is necessarily
subdivided at finite intervals. Note that this is similar to
the circumstances in which a numerical value is rounded
off to some particular number of significant figures. This
rounding off is equivalent to a statement that, whatever
the actual value of the uncertainty, it cannot be less than 5
in the next decimal place. Thus, if we quote 7r as 3.1416
we are conveying no more information than that 7r lies
between 3.14155 and 3.14165 and so the range of uncertainty of the statement is .00005 about 3.1416. This
is not a real statistical uncertainty, but just another way of
saying, "the value of 7r is stated to 4 places of decimals."
Reading an instrument like a meter stick to millimeters is
l4
just a form of rounding off and consequently carries an

immediate implication of a minimum uncertainty of onehalf the finest scale division. The possibility exists, of course,
of visual interpolation between the finest scale divisions, but
the validity of such a procedure depends largely on the size
of the undivided interval. A claim to be able to interpolate
within millimeter divisions would be difficult to justify, but
if the scale were divided in centimeters a practised observer
could probably judge millimeters quite reliably. In all
cases of visual interpolation, however, the estimation of the
actual reliability is very difficult.
The foregoing are merely a few of the headings under which
experimental uncertainties can be grouped. Sometimes it is
possible to assign a cause to a particular observed uncertainty, and sometimes the cause can be eliminated.
Certainly all justifiable effort should be expended on such
diagnosis and cure. However, there is always a limit to the
resources of time or equipment available for such attempts,
and there comes a stage at which the observer wishes to
state the residual uncertainty in his reading. It is with this
residual uncertainty that the following treatment is concerned. If the observer's measurement is to be regarded as
significant, he must be realistic in his assessment of the
uncertainty. If he claims too high a precision, he may be
guilty of dishonesty and he will certainly cause grief to
those who use his results for continued work. If he claims a
pessimistically low precision, there may not be the serious
consequences of the other extreme, but he is certainly doing
himself an injustice. He is also degrading the worth of his
work since it will not be accorded the significance which
may be its due. This is unfortunate, not only for the
experimenter himself, but also for the workers in the field
15
who are denied confidence in results which could be of

value. The point is, therefore, that pains should be taken
to assess the uncertainty of the measurement, as carefully
as possible, striving to be realistic without undue optimism
or pessimism.
I t is now necessary to face the problem of how such an
estimate of precision can actually be made. We are, for the
moment, considering only the uncertainty of a single
measured value, and the over-all precision of an experiment
will be considered later. If an observer makes one, and only
one, measurement of a quantity, the problem of estimating
its precision is a difficult one. It is not an uncommon
suggestion to use the finest scale division as a measure of
the "maximum" range of uncertainty. However, this practice does not take into account the possibility of fluctuation
arising from other sources. It is foolish to claim limits of
uncertainty of ! mm on a meter stick if the influence of
parallax could give a range of 2 or 3 mm, or to claim that
voltages read on a meter divided in -h v are significant to
.05 v if random perturbations are causing the needle to
fluctuate over 2 or 3 tenths of a volt. On the other hand,
the manipulation of the apparatus could be sufficiently
precise that settings can be made well within the limits of
the scale subdivision. In this case interpolation between
the finest scale divisions may become significant and the
use of the scale division as a measure of the precision would
be unnecessarily pessimistic. I t is therefore of rather limited
value merely to quote the scale division as the limit of uncertainty. If such a criterion is used, the nature of the uncertainty quoted should be clearly stated so that the significance of the estimate can be evaluated. This method of
quoting uncertainty may possibly be justified by con-
16
venience, but only in the case of the most simple, direct

measurements. In the case of most experimental observations such an estimate of precision is hardly of any more
value than no estimate at all, for it can never be anything
more than an estimate and we shall always require an actual measurement of the uncertainty. Thus, for an experiment of any significance, a single measurement of any
quantity is almost useless. The only way of meeting this
difficulty is to repeat the measurement, not so much with a
mystical faith that the averaging results in a "better" value,
as to see how closely reproducible the measurement is.
2.3
Distribution Curves
The process of duplicating readings is at the heart of any

experimental process and much of the treatment to be given
later is concerned with the ways in which this duplication
can be carried out so that the yield of the experiment
can be maximized.
In the meantime we shall concern ourselves solely with the
consequences of straightforward duplication of the same
measurement. Since we are dealing with a system which is,
by hypothesis, subject to perturbation, the resulting readings will not be the same. Depending on the degree of
perturbation or of judgement required, the spread may be
a large or small fraction of the magnitude of the measurement. If the spread is large compared with the absolute
value of the measured quantity, we can say, in general
terms, that the reliability or precision of the measurements
is poor, and vice versa. However, the problem is not restricted merely to making the observations. If the measurement is to be of any use in further work, or to other people,
17
it must be capable of being described in simple terms. The

simplest thing is merely to reproduce your table of results
saying, "I have measured this quantity n times and here are
the n answers." This places the burden of evaluating the
significance of the results on the reader. However, such a
practice, although convenient to the experimenter, is
clumsy and obscure, since few people can see clearly the
essential characteristics of a set of numbers just by looking
at them. Thus, it is an enormous convenience and source
of clarification if the way in which the readings are distributed along the range of possible values is apparent, and
a visual representation is much the best way. Methods of
picturing ranges of numerical values can take many forms,
but the most profitable device is the "histogram."
This diagram is drawn by dividing the original set of observations into intervals of predetermined magnitude and
counting the numbers of observations found within each
interval. If one plots, on a suitable scale, this frequency
versus the readings themselves, a block diagram is obtained
which is the required picture of the distribution of the readings along the scale of values. It actually contains no more
information than the original set of readings but has, as
required, the enormous advantage of visual presentation of
the nature of the distribution of the readings.
Such a histogram and the set of readings from which it is
derived are shown in Fig. 2.1 . If the number of readings is
very high, so that a fine subdivision of the scale of values
can be made, the histogram approaches a continuous curve,
and this is called a distribution curve.
Clearly a histogram or a distribution curve contains the
information about the spread of the experimenter's read-
18
Table!
109
109
110
110
111
111
111
111
111
112
112
112
113
113
113
113
113
113
114
85
92
96
97
97
97
100
101
101
102
102
103
103
, 105
106
106
107
108
108
121
121
122
122
122
122
122
123
123
123
123
123
124
124
124
125
125
125
126
114
114
114
115
116
116
116
117
117
118
118
119
119
120
120
120
120
121
121
40
30
~
c
~
20
C"
....e
10
80
Value
Fl 2.1
A set of observations and its histogram.
100
127
127
127
127
128
128
128
128
128
128
130
130
130
130
130
130
131
131
131
131
132
133
134
134
134
134
135
136
137
137
137
144
148
149
19
ings which he wishes to communicate to other people, and

does so in a clearer and more informative manner than does
the original table of values. The histogram is an excellent
way of presenting experimental results and is commonly
used in reports on work which shows a high level of statistical fluctuation. It has the further advantage of presenting the observations themselves, free from manipulation
on the part of the experimenter, and thus permits the
reader to form his own judgement regarding the value of
the work.
However, such a full-scale representation of the results may
be undesirable, perhaps on grounds of difficulty of presentation, or more significantly, the results may be required
for further work. In either case, one wants to find numbers
which will represent the distribution of the values so as to
answer the questions, "Which single number shall I take
as the answer, and how reliable is it?" It is necessary,
therefore, to define quantities which will serve as a "best"
value and an "uncertainty."
It is difficult to say at this stage exactly what we mean by
"best" and "uncertainty" because we are, for the moment,
using a frankly intuitive approach. The uncertainty is obviously associated in some way with the spread of the resuI ts,
and the best value with any tendency of the results to
cluster in the middle of the distribution. As the development of the uncertainty theory progresses our concept of
the terms above will gradually become clearer and clearer.
If these numbers are to have general significance, they must
be defined in some standard way so that their significance
will be widely understood and accepted. Unfortunately,
the significance of most numbers one can suggest for the
20
purpose above depends on the actual shape of the distribution curve. We have not yet said anything about the
shape of the curve itself and, clearly, this depends on the
nature of the measurement. Consider an experiment which
consists of making a measurement on a scale marked in
inches and attempting to estimate tenths. If the most commonly found value were roughly halfway between divisions,
it might be found that the probability of getting a reading
declined to zero at the two boundaries and was approximately constant in the center third of the range. The
distribution curve for such a set of readings might appear
as in Fig. 2.2. On the other hand, the experiment might
consist of measuring the tensile strength of a set of supposedly identical wire samples. Here repetition might show
that, although there was a sharp upper limit to the tensile
strength obtained, many specimens showed lower strength
because of imperfections. This distribution curve might appear as in Fig. 2.3. The characteristics of these two curves
are completely different, and it is obviously impossible to
Scale reading
Fig 2.2
One type of distribution curve.
21
Scale reading
FIC. 2.3
Another type of distribution curve.
quote numbers which will represent them with any reliability.

The circumventing of this problem takes us into the domain
of mathematical statistical theory. The principle is that, if
any arbitrarily found distribution curve is too difficult to
work with, we must define one which will prove tractable
and, we hope, not too unrealistic. It must be stressed that
most of the material in the rest of this chapter refers to
this defined distribution curve only. If the statistical theory
is to be of any significance in a set of measurements, the
correspondence between the theoretical form of the curve
and the distribution curve of the actual observations must
be established, or else only limited validity of the statistical
theory must be accepted and admitted.
Before working with the formal curve, however, there are
some definitions relating to the distribution curves which
are made independently of the shape of the curve. It must
be remembered, however, that the significance of the
quantities so defined does depend on the particular distri-
22
bution, and so caution must be used in their interpretation.

The mean, x, of a set of n numbers Xl, X2, etc., is defined as*
_
~Xi
= ._---
(2.1)
This is nothing more than the familiar average. The mean

does have, for reasons which will become clear later (Sec.
6.6), the characteristics of a "best" value and is commonly
quoted as such.
However, if the distribution is not symmetrical, the mean
value is not the most commonly found value, and the value
at which the peak of the distribution occurs is called the
mode. For many purposes it is the most frequently found
value which is the significant feature (for example, to a
clothing shop ordering a new supply of hats) and so, for
many purposes, the mode is quoted as the single number
which best characterizes the distribution. Note that the
mode and the mean can have differing values and this
shows again the need for caution, since asymmetry of the
distribution curve would probably be revealed only by
actual plotting. In such a case, careful thought should be
given to the question of which number best typifies the set
of values.
Yet another quantity is commonly defined at the center of
the set of readings so that as many readings fall below it in
value as above it. This is called the median and a line drawn
vertically from the median on the distribution curve will
divide it into two equal areas. For a symmetric distribution
it coincides with the mean and the mode. These three
* The symbols ~ Xi
refer to the sum of all the

~ Xi =
Xl
values in the set, i.e.,
+ + xa + ... + :r"
X2
23
quantities are illustrated in Fig. 2.4. Short of drawing the

complete distribution curve for a set of observations, the
quotation of one or more of these quantities will help the
reader in his interpretation of the observations, but such an
attempt to describe the curve is necessarily of limited value
"I,. 2..
The constants of a skew distribution.
In any case, for a symmetric distribution these three values

coincide and so it is now necessary to define a quantity
connected with the actual width of the curve.
This quantity is obviously associated with the deviations of
the readings from some central value and one quantity
which could be defined is the mean deviation. If a set of n
readings XI, X2, etc. have a mean x, the mean deviation is
defined as
mean deviation =
~ IXi - xl
n
(2.2)
However, a more commonly used term is the variance and

is defined as
.
~ [(Xi - X)2]
varIance = - .- .- - Yo
24
The square root of the variance

deviation, denoted by s, so that*
1S
called the standard
(2.3)
Thus it may suffice, when stating the result of a repeated
observation, to quote the mean value as the best value, and
this standard deviation for the set of readings as a measure
of the uncertainty. (For a slight correction to this statement
see page 35.) It must be stressed again, however, that the
significance of these quantities depends on the actual shape
of the distribution curve. If for any reason the curve of the
actual observations is markedly asymmetric, considerable
care must be exercised in the interpretation of these calculated quantities.
It has been common in the past to define a quantity known
as the probable error. This is a value which divides the area
under the distribution curve into two equal parts denoted
I and II in Fig. 2.5. It has the significance that any reading
of the set has an equal probability of being inside and outside the limits set by the probable error. This is a useful and
reasonable definition, but the size of the probable error, and
its relation to the standard deviation depend on the partiClllar shape of the distribution curve. This limits the usefulness of the quantity and it has become more common to
disregard it in favor of the standard deviation.
The problem is thus a matter of reducing a set of observations to such a condensed form as will permit further
work or calculation. The nature of this condensed form
depends on the nature of the results. If the distribution
* Note that this formula involves the sum of the squares of the deviations,
not the square of the sum of the deviations.
25
2.5
The probable error in a frequency distribution.
curve is symmetric, the obvious quantities to quote are the

central value (i.e., the mean) and some measure of the
width of the curve, such as the standard deviation of the set.
If the distribution curve is markedly asymmetric, it may be
necessary, in order to provide an accurate impression of the
results, to quote the mode and/or median, in addition to
the mean, and, in extreme cases, there is no alternative to
quoting the whole distribution curve.
Two difficulties remain for the physicist. First, the distribution curve for a particular measurement is not always
available. If one is measuring a wire diameter with a
micrometer, one rarely takes the hundreds of measurements
required to provide a significant distribution curve. Second,
it would be useful if the standard deviation could have a
definite numerical significance as a measure of the uncertainty for, as we have stated above, the relationship of
a quantity such as the standard deviation to the distribution
curve as a whole depends on the shape of the curve. For
this purpose it is common in physical measurement not to
26
consider the actual distribution curve relating to a particular measurement, but to discuss the situation in relation to
a defined curve. We do not pretend that all physical observations actually follow this curve but many carefully made
observations may be adequately close. Furthermore the
procedure will assign a definite numerical significance to
quantities such as the standard deviation and also permit
deductive work like the theory of sampling to have a
definite numerical significance.
2.4 The Gaussian or Normal Distribution
Many theoretical distribution curves have been defined and

their properties evaluated, but the one which has most
significance in the theory of measurement is the Gaussian
distribution.
The Gaussian error curve can be defined on the assumption
that the total deviation of a measured quantity, x, from a
central value, X, is the consequence of a large number of
small fluctuations. If there are rn such contributions to the
total deviation, each of equal magnitude, a, and either
positive or negative, the total set of observations may range
from X
rna, if all fluctuations happen to be positive
simultaneously, to X - rna if the same happens in the negative direction. It can be shown, in such a random summation of positive and negative quantities (as in the
"random walk"), that the most probable sum is zero,
meaning that the most common values of x are in the
vicinity of X. The distribution curve is thus peaked in the
middle, is symmetric, and declines smoothly to zero at
x = X
ma and x = X - mao If this concept is taken to
the limiting case of an infinite number of infinitesimal con-
27
tributions to the total deviation, the curve has the form

shown in Fig. 2.6. Treating the curve solely from the mathematical point of view for the moment, its equation can be
written
y = C e-h'(x-X)'
(2.4)
Here the constant C
y = C for x = X at
curve is symmetric
asymptotically. The
is a measure of the height scale, since

the center of the distribution. The
about x = X and approaches zero
quantity h obviously governs the
Small h, large
Fle.2.6
The Gaussian distribution curve.
width of the curve, since it is only a multiplier on the x

scale. If h is large, the curve is narrow and high; if small,
the curve is low and broad. The quantity h must dearly be
connected with the standard deviation, (1', of the distribution and it can be shown that the relationship is
(1'
=--
v2h
(2.5)
28
x
FIC. 2.7
The relationship of 10" and 20" limits to the Gaussian distribution.
(We shall use latin letters, e.g., s for standard deviation, for
quantities related to finite sets of actual observations, and
greek letters, e.g., (1, when referring to the normal distribution itself, or a "universe" of readings as described on
page 30). The relationship of the standard deviation to the
scale of the curve is indicated in Fig. 2.7 by the lines
drawn at intervals of 1(1 and 2(1 from the central value. The
probable error can also be evaluated and it is given by
P=
0.48
h
0 67 (1
rov
rov.
(2.6)
For a more complete account of the mathematical properties of the Gaussian error curve see Appendix 1.
2.5 Correspondence between the Normal Distribution and
Actual Observations
The normal distribution is a smooth and continuous curve.

It can, therefore, correspond only with a limiting case as
the number of observations tends to infinity, and this rather
29
stringent requirement is the first point of difference between

the Gaussian curve and actual sets of observations. Even
when a numerous set of observations is made, however,
there is no guarantee that the distribution will be Gaussian.
One of the assumptions in the definition of the Gaussian
distribution is that, in each reading, there is an infinite
number of infinitesimal perturbations, each of which is
equally likely to be positive or negative. This is an idealized
condition, and the coarser perturbing influences in actual
observation make it rather difficult to establish correspondence between actual observations and the ideal form
of the curve. For example, one result of this is that the
Gaussian curve approaches the axis only asymptotically,
thus permitting a finite possibility of extremely far-out
readings. This probability is certainly small but practical
considerations make such possibilities completely unacceptable.
Some forms of observation may not even accord in principle
with the normal distribution. One common example is
particle counting in nuclear physics. Here the repetition
of counts, over equal time intervals, of randomly occurring
events follows a skew distribution known as the Poisson
distribution. If the number of counts in each interval is small
the distribution is markedly skew, but the curve approximates the normal distribution more closely as the number
of counts increases. The statistics of this distribution play
a very important role in nuclear physics but it is not discussed here further because its application is limited to
counting experiments. For a description of the distribution
and the ways of using it the reader is referred to Reference
18 in the Bibliography.
30
2.6 Significance of the Standard Deviation in Actual Measurements
From Fig. 2.7 it can be seen that the values 10", 20", etc.
divide the area under the curve into various regions. Since
areas under a frequency distribution curve represent
numbers of readings, areas, expressed as a fraction of the
total area, are a measure of the probability of obtaining
readings wi thin the range defining the area. I t is now possible to see the significance of the curve and the standard
deviation to the experimenter.
Imagine that the night before an experimenter is going to
make a measurement of, say, the period of a pendulum, the
ghost of Galileo comes to the laboratory and, with the same
apparatus and under the same conditions to be used in the
experiment, makes a very large number of observations (sa~T
a million) of the pendulum's period. He can then draw the
distribution curve of these and this distribution curve can,
for the present purpose, be assumed to be very closely Gaussian. It will be centered on a mean value X which is very
close indeed to the value obtained if the number of observations actually went to infinity. This last limit would define (ignoring systematic error) what we could term the
"true" value. There will also be a certain value for the
standard deviation which can be calculated from the readings using Equation (2.3). This picture is merely a way of
visualizing what is called the "universe" or "population" of
a particular reading. This term refers to the infinite set of
readings which could be made with the apparatus, and thus
provides a link between actual observations and the statistical theory. This population and its distribution curve exist
for any measurement we care to consider and it is the con-
31
stants of this curve, particularly the mean value X, which if::

the object of the experimenter's work. The only trouble i~.
its obvious inaccessibility.
Next mornin~ our experimenter arrives and makes his
single measurement. It will fall somewhere along the range
of the previously obtained million readings. Its exact positiop is a matter of chance, and all that can be said is that
th probability of its being within a range of 10" from the
value of X is the fraction that the area enclosed by the 10'
limits is of the total area under the curve. This can be
shown (see Appendix 1) to be about 68 per cent for a
Gaussian distribution. The probability that it falls within
20" is, correspondingly, about 95 per cent. Thus, the only
thing that can be stated about our experimenter's measurement is that he stands a 68 per cent chance of being within
10" of Galileo's almost "true" value of X and a 95 per cent
chance of being within 20". The only trouble is that, in
the absence of psychic contact with Galileo, he does not
know what values of X and 0" are appropriate to the experiment. All he can say is that his single reading stands
a 68 per cent chance of being within something of something-a somewhat restricted statement. This brings us
back to the earlier point of the futility of isolated readings.
The obvious solution, which has already been propounded,
is the duplication of readings in an attempt to increase the
amount of information available, and this requires an
extension of the statistical theory.
2.7 Sampling
The two unknowns which constitute our experimenter's

problem are X, effectively the "true" value, and 0". Obvi-
32
ously, if the experimenter could make another million readings of his own, he would be able to get sufficiently close to
X from the peak of his distribution curve or the mean of his
observations. In this case he need not necessarily be interested in the standard deviation since he would have obtained his answer sufficiently precisely. The trouble is that
obvious practical considerations preclude the taking of
another million readings. He will attempt to duplicate his
readings, of course, but the best he will achieve will be a
sample of the next million. Say he takes 10 readings. He
hopes that this will do two things. First, it will give him
some measure of fT, and second, he trusts that he will be
further rewarded for his pains by improved reliability of
his mean value, although he cannot hope that the mean of
his sample should actually coincide with Galileo's mean of
a million readings. The improvement that does result is
given by the statistical theory of sampling.
Observations taken singly
FIC. 2.8
Distribution curve of single observations and sample means. (Note that

the vertical scale for the two curves is not the same. They have been
plotted with a common peak value solely for purposes of illustration.)
'B
Suppose that Galileo returns again the next night and takes
another million readings, Let us suppose that he then takes
these million readings and divides them into groups of ten.
He will have a hundred thousand of these. Suppose that
he works out the mean of each group and plots the distribution curve of these means. A derivation based on the
statistics of sampling shows that this new distribution curve
of means will also be Gaussian (even if, in fact, the distribution of single observations was not Gaussian) and will be
centered on the same value X as was the first curve. However, its most striking feature is that, as illustrated in Fig.
2.8, it is narrower than the distribution curve of the readings taken singly and it can be shown that the standard
deviation of this set of means, which we shall call CTm , is
given by
CTrn- =
Vn
(2.7)
where n is the number of readings in each group of the complete set. (For an indication of how this can be proved see
page 64.) In the case of our illustration n equals 10, and so
Galileo will find the distribution curve of his collection of
means to be about one-third as wide as that of the readings
taken singly.
Hence, the improvement which the experimenter hoped to
gain from duplicating his readings comes from this smaller
standard deviation of the means of the groups. The point
is that his action of taking a set of 10 readings is just that
of selecting, at random, one of Galileo's groups. If, therefore, he evaluates the mean of his sample, the likely position
of this quantity along the scale of values is governed by the
34
distribution curve of means, and not that of the values taken

singly. Thus, if he takes n readings and evaluates their
mean, he stands a 68 per cent chance that this mean will
come within, not O", but O"/Vn of Galileo's central
value X and, correspondingly a 95 per cent chance of coming within 20"/Vn.
This does justify his faith that duplicating his readings will
yield him a better answer since, with 10 readings he is
likely to be 3 times better off than before. However, even
with this improvement, notice that, in the absence of the
value of 0", he is still in the position of being able to say
nothing more than that he stands a 68 per cent chance of
being within something of something. He can improve his
situation only by making an estimate of 0". Towards this,
all he can do is work out the s for his sample. To see what
luck he will have, let us suppose that Galileo goes back to
his hundred thousand groups of readings and, for each
group, works out its s value. They will, of course, not all
be the same, and, again, an extension of the mathematical
treatment of the original Gaussian equation shows that this
set of s values will form a distribution curve which is, as
before, bell-shaped and which is centered on 0" as shown in
Fig. 2.9. The breadth of this curve is related to the standard
deviation of the set of s values. We shall call this 0". and it
can be shown that it is given by
0"8
0"
-Vi=2=(n=_=1)
(2.8)
where n is the number of readings in the sample. Thus, if

our observer with his sample of 10 observations evaluates
the s value of his sample, he will get a value which will lie
somewhere along the scale of possible s values, and this is
35
s
Fig. 2.9
The distribution of sample standard deviations.
as far as will be able to go towards a value of 0'. It can be

shown, actually, that the best estimate of 0' that can be
made from a sample is not exactly the sample standard
deviation itself. Instead, the best estimate of 0' is given by
best estimate of
0'
V~[(Xi - :x)2l!(n
-=--1)-
(2.9)
with n - 1 in the denominator instead of n. We shall accept

this value and when we quote the standard deviation for a
set of readings in the future we shall do so, not as a measure
of the variance of the set itself, but rather as the best
estimate of the universe standard deviation 0'. The difference is, in any case, small if we have enough readings to
justify a calculation of s.
This estimate of 0' from the sample is the best that the
experimenter can do but he can now state limits to his
ignorance, for Equation (2.8) shows that, with 10 readings,
he stands a 68 per cent chance that his s value will lie
within a range of
O'/Vls
or, approximately
0'/4,
36
about the unattainable and elusive o. He has correspondinglya 95 per cent chance of coming within u/2 approximately.
Let us list some typical values of V2(n - 1):
Table 2
Confidence 68%
n
2
3
4
5
6
7
8
9
10
15
20
50
100
Confidence 95%
V2(n - 1)
1.4
2.0
2.4
2.8
3.1
3.4
3.7
4.0
4.2
5.2
6.1
9.8
14.1
2
3
4
5
6
7
8
9
10
15
20
50
100
V2(n - 1)
0.7
1.0
1.2
1.4
1.6
1.7
1.8
2.0
2.1
2.6
3.2
4.9
7.0
These values are illustrated In Fig. 2.10 for n = 10 and

n = 3, giving an impression of the range of possible s
values. On these curves are marked the lu s limits, showing the range within which the observer stands a 68 per
cent chance of having his sample standard deviation fall .
The situation looks correspondingly worse for the 2u. limit~
. and it can be seen that for n = 10, 95 per cent of the possible s values lie within limits of u/ 2, and for n = 3, 95 per
cent lie within a range which is large enough to cover 0 up
to 2u. In actual practice the value of u is, of course, un
37
n= 10
n=3
u
.'C.2.10
Sample standard deviation distributions for small samples.
known. One would therefore use the value for the best
estimate of (J' in Equation (2.8) to give the actual numerical
value for (J' s . This provides the actual range within which
we have 68 per cent confidence that the standard deviation
estimate lies.
The important point about this result is that the observer
may be guilty of misrepresentation in quoting his results if
he does not state his number of readings. For, if he quotes
his mean value and the standard deviation of the mean
with the intention of expecting an interpretation that his
38
mean stands a 68 per cent chance of being within this

quantity of the "true value," a glance at Fig. 2.10 will
convince the reader that, even with as many as 10 readings,
the particular value for s can easily be so far from CT as to
destroy the worth of the numerically exact statement that
the observer was trying to make.
2.8 Practical Considerations
An understanding has now been reached of the nature of

the result of repeating the measurement of a single quantity.
To put this into practice in achieving optimum results from
an experiment involves the balancing of the effort required
to improve the intrinsic precision of the measurement itself
against that of taking sufficient readings to improve the
precision by a sort of brute force method. The use of the
Expression (2.7) will enable a decision to be made regarding the number of duplicated readings required to achieve
a certain specified precision. Here it must be noted that,
since the standard deviation of the mean involves V~, the
precision of the experiment improves rather slowly. It is
rather tedious to obtain an additional significant figure by
taking a hundred times as many readings. Usually it is
easier, and it is certainly more satisfactory, to get a travelling microscope from the store room than it is to measure
lengths with a meter stick to tenths of a millimeter. In
any case, the theoretical reasoning cannot be pushed too
far, since no amount of repetition of a meter stick reading
will give a precision of a micron (i.e., 10-6 m). The
finest scale division always provides a fundamental lower
limit to the uncertainty.
39
The principal benefit, therefore, of repeated readings is not

so much the improvement of the accuracy of the measurement as the provision of an estimate of the precision.
Caution must be exercised in interpreting such estimates,
however, since, as we have seen, the reliability with small
numbers of readings is poor. It is useful to keep in mind
the figures for the limits giving 68 per cent and 95 per cent
confidence for 10 measurements. These limits of 0"/4 and
0"/2 will serve as a rough mental guide but, if a definite precision in 0" value is required by the experimental conditions,
the figures in Table 2 should be used to determine the numbers of observations required in the sample. It must be
stressed again, however, that regardless of what is actually
done in the experiment, the number of readings must be
stated, in quoting the answer, to permit assessment of the
significance of the resul t.
Of course, it may happen that duplication of the measurement does not give rise to differing answers. This is the case
where the statistical fluctuations are sufficiently small to
permit the distribution curve to shrink into the space between the finest scale divisions. In this case the result of the
repetition will serve as evidence that the precision of
measurement is limited only by the scale divisions. In all
cases, however, the observer has to try repetition to discover
which type of measurement he is dealing with, perturbation-limited or scale-limited.
In actual practice, then, the number of readings will be

determined by a required precision of the mean or of the
standard deviation. This number is normally so small that
no conclusions about the actual distribution can be drawn.
One therefore assumes that the set of observations is a
40
THE NATURY. OF MEASUREMENT
sample from a Gaussian population. One quotes the mean

as the best value and the standard deviation of the mean
as the pr~cision, and assumes that this last has the numerical
significance appropriate to a Gaussian distribution.
2.9 Rejection of Readings
One last practical consideration of the distribution curve is

concerned with outlying values. There is always the possibility of an observer making an actual mistake, perhaps in
misreading a scale or in moving an instrument, accidentally and unnoticed, between setting and reading.
There is always the temptation, therefore, to assign some
such cause to a single reading which is well separated from
an otherwise compact group of values. However, this is a
dangerous temptation since the Gaussian curve does permit
values remote from the central part of the curve, and also
since the admission of the possibility of pruning of results
makes it very difficult to know where to stop. The only
answer can be in the judgement of the experimenter. This
is not unreasonable since the measurement is his creation,
and if his readers are going to give his conclusions any
weight they are, in effect, expressing faith in his judgement.
Many empirical rules for rejection of observations have
been set up but they cannot replace personal judgement. It
would be foolish to use a rule to reject one reading which
was just outside the limit set by the rule, if there are other
readings just inside it. There is also the possibility of extra
information relating to the isolated reading which was
noted at the time of making the reading, and which can
help decide in favour of retention or rejection.
For a normal distribution curve the probability of obtaining
a deviation greater than 2(1 is 5 per cent (as we have seen
41
before), greater than 3/T is -l per cent and greater than

4/T is no more than 6 X 10-5 . The criterion for rejection is
still the responsibility of the observer, of course, but one can
say, in general terms, that readings falling outside 3/T limits
are more likely to be mistakes and so candidates for rejection. However, a problem arises because of our lack of
information about the universe of readings and its constants
X and /T. The better our knowledge of /T the more confident
we can be that any far-out and isolated reading arises from
a genuinely extraneous cause such as personal error in reading, malfunction of apparatus, etc. Thus, if we make 50
observations which cluster within 1 per cent of the central
value and then obtain a reading which lies at a separation
of 10 per cent, we would be fairly safe in suggesting that the
circumstances had changed and that this last reading did
not belong to the same universe as the previous 50. The
necessity is to build up confidence in the main set of
measurements before feeling justified in doing any rejecting.
Thus, there is no justification for taking two readings and
then rejecting a third on the basis of a 3/T criterion. Unless
the situation is absolutely clear-cut it is by far the best to
retain all the readings whether one likes them or not.
It is wise to remember also that many of the greatest discoveries in physics have taken the form of outlying observations.
2.10 Numerical Example
The considerations above of distribution curves and

sampling are the keys to understanding the nature of
measurement and it is of value to clarify our ideas by actual
numerical examples. Let us pretend that the set of figures
given in Table 1 on page 18 is a complete population for a
42
particular experiment. Thus, by playing the role of Galileo,

we shall be able to see just what luck the experimenter is
going to have with his single sample. Since we are only
pretending to have a complete set we cannot, of course,
expect to get smooth distribution curves like Galileo's, but,
even although rough, they will help to make the situation
clear. We assume, therefore, that the distribution curve in
Fig. 2.1 is that of the population, that the mean is the "true"
value and that s for the set is (1. The mean can be found to
be X = 119 and, using Equation (2 .3) , the standard deviation is found to be (1 = 12. The distribution curve is replotted in Fig. 2.11 and the limits (1 = 12 from the mean
are marked on it. Notice, first, how these 1(1 limits divide
x
o
Observations
Means
3
30
'Vi'
c
:;:;
'"
1:
'Vi'
c
~
o
.0
Mean
~20
c
:e
1)'
10'm
- 2
'"
:::J
rr
~
u..
OJ
:0
U.
10
\
150
Fl 2.11
Distribution of single readings and sample means.
>.
limits-==J==t~=:;:~-
'"
5'"
43
the area under the curve to give approximately the 68 per

cent fraction that we have been using, and so give an impression of the chances our experimenter would have with
a single reading.
We now wish to perform Galileo's grouping process to investigate the sample fluctuations. Since the numbers in
Table 1 were arranged in ascending order to aid in interpretation of the histogram, the numbers in the set were shuffled
to give the impression of samples of random readings.
Random grouping gives the following samples:
1
130
114
122
121
92
109
96
124
103
137
7
10
4
5
6
8
9
11
2
3
- - - - - - - - - -- - - -- -- -- 128
114 125 130 128 116 122 137 137 101
117 113 134 133 125 148 130 134 127 101
110 125 121 130 108 111 121 124 120 115
120 132 109 124 111 113 112 128 114 130
85 116 131
97 114 127 116 113 106 126
149 108 120 130 113 100 144 111 128 127
97
131 122 121 123 134 127 110 137 117
128 123 123 119 111 193 122 120 102 102
97 106
131 122 135 113 136 119 112 118
111
112
105
123
134
118
113 107 131 128
We first evaluate the means of the groups. They are:

1
10
11
114.8 121.0 119.1 125.1 124.4 117.8 116.1 120.5 121.6 115.6 115.5
We cannot plot much on a distribution curve with only

11 values, but by selecting a range of size unity, we can,
44
as shown in Fig. 2.11, get an impression of how the means

of the groups are clustered around the mean (119) of the
population. This clustering shows us the degree of improvement in precision obtained by taking the mean of 10
cr
(= 12)
2 cr.
(= crf2) limits
10",
(= crf4) limits\
\
/
2-
e-
I
4
'o"lg. 2.12
10
12
I
14
I
16
18
Distribution of sample standard deviations.
values, instead of one single reading. We now evaluate the

standard deviation of the mean from Equation (2.7). It is
12/VlO = 3.8. These limits of 4 are plotted about the
mean on Fig. 2.11. Note how this value has the same kind
of relationship to the distribution of means as (J' had for the
distribution curve of single values, and how it therefore
serves as a measure of the improvement in precision which
45
has resulted from the duplication of readings to form a

sample.
To see what chance our observer, with his single sample,
will have of getting a value of s reasonably close to u, we
shall evaluate s for each of the 11 samples using Equation
(2.9). These standard deviations are:
14.7
10
11
14.4
8.2
7.9
6.6
11.3
14.0
11.4
15.4
12.4
13.2
Again, 11 values does not give a very good distribution

curve, but by selecting a range of size unity, we can get an
impression, as shown in Fig. 2.12, of how these sample
standard deviations are distributed along the scale of s
values and centered on 12, the value of the standard deviation of the population. Again we cannot expect too close
numerical correspondence but we can believe, on looking
at Fig. 2.12, that with more samples, 68 per cent of them
would be contained within a range o/4 = 3 about
U = 12, as suggested by Equation (2.8) and illustrated in
Fig. 2.12, and that only 5 per cent would lie outside limits
of u/2 = 6.
As he looks at these illustrations, the reader should consider
the situation of the observer with his single sample, and thus
gain an impression of the chances involved in the result of
the experiment. The most important lesson is the obviously
high degree of sample fluctuation, even with as many as 10
readings in each sample. This should serve as the warning
against the drawing of unjustifiable conclusions from sets of
observations.
46
PROBLEMS
The following observations of angles (in minutes of arc) were

made while measuring the thickness of a liquid helium film.
Assume that the observations show random uncertainty and that
they are a sample from a Gaussian universe.
34
38
33
38
38
36
35
47
36
32
40
40
45
36
43
38
48
40
40
38
43
40
39
36
46
34
37
33
32
34
1. Draw the histogram of the observations.
2. Identify the mode and the median.

3. Calculate the mean.
4. Calculate the best estimate of the universe standard deviation.
S. Calculate the standard deviation of the mean.
6. Calculate the standard deviation of the standard deviation.
7. Within which limits does a single reading have (a) a 68 per

cent chance of falling, and (b) which limits give a 95 per cent
chance?
8. Within which limits does the mean have (a) a 68 per cent
chance, and (b) a 95 per cent chance of falling?
9. Within which limits does the sample standard deviation stand
(a) a 68 per cent chance, and (b) a 95 per cent chance offalling?
10. Calculate a value for the constant h in the equation for the
Gaussian error curve and for the probable error of the distribution.
11. If a single reading of 55 had been obtained in the set would
you have decided in favor of accepting it or rejecting it?
47
12. Take two randomly chosen samples of five observations each

from the main set of readings. Calculate their sample means and
standard deviations to see how they compare with each other and
with the more precise values obtained from the big sample.
13. If the experiment requires that the standard deviation of the
mean should not exceed 1 per cent of the mean value, how many
readings are required?
14. If the standard deviation of the universe distribution must

be known within 5 per cent, how many readings are required?
The Propagation of Uncertainties
3.1 Absolute and Relative Uncertainty
The previous chapter has provided a means whereby an

estimate of the uncertainty of measurement of a single
quantity can be made. However, it is normally the case,
in even the simplest experiments, that the final answer is to
be computed in some way from values of several different
quantities, each independently measured and subject, individually, to uncertainty. Thus, in the simple pendulum
experiment, g is obtained as a function of T% and I, and
the uncertainties in T and I will bo~h contribute to an uncertainty in g. We are concerned with this resultant uncertainty in this chapter.
The word "uncertainty" will be used to signify outer limits
of confidence within which we are "almost certain" (i.e.,
perhaps 99 per cent certain) that the measurement lies.
This will commonly be valid only in simple measurements
rounded off to the nearest scale division. When a measurement has been repeated often enough to give statistical
significance to the result, one would obviously quote the
standard deviation of the sample or the standard deviation
of the mean, but we shall restrict the word "uncertainty" to
the case of outer limits of confidence.
48
49
THE PROPAGATION OF UNCERTAINTIES
Let us suppose, therefore, that we have measured values of

x and y (which we shall call Xo and yo) with either outer
limits of uncertainty ox and oy DC else standard deviations
S", and SI/. These uncertainties can be expressed in two
forms, either of which can be the more useful, depending
on circumstances. The uncertainty itself ox or oy will be
referred to as an "absolute uncertainty." However, injudging the significance of this uncertainty in comparison with
the actual value Xo, the ratio ox/xo (or s",/xo) is the more
useful quantity. Obviously a ! mm uncertainty is far more
significant when the measurement is 1 cm than when it is
1 m. This ratio will be referred to as the "relative uncertainty" or "precision" of the measurement. It will normally be expressed as a percentage.
3.2 Propagation of Uncertainty
The simplest case in which a result is computed from a

measurement occurs when the result is a function of one
variable only as, for example, the computation of the area
of a circle from a measurement of a diameter.
Consider a computed result z to be a function of a variable x
z = f(x)
Here a measured value Xo permits the required value Zo to
be calculated. However, the possibility that x can range
from Xo
ox to Xo - ox means that there is a range about
Zo of possible values of z from Zo
oz to Zo - oz. It is this
value of oz which it is now desired to calculate. The situation is illustrated graphically in Fig. 3.1 in which it can
be seen, for a given f(x) , how the measured value Xo gives
50
8z
Zo~-L--------------~
Xo
1'1,. 3.1
Propagation of uncertainty from one variable to another.
rise to the computed result Zo, and how the range ox

about Xo gives a corresponding range oz about zoo
Before proceeding to any general methods of evaluating oz,
it is instructive to see how finite perturbations are propagated in simple functions. For example, consider the function
If x can range between Xo

ox and Xo - ox then z can
range between Zo
oz and Zo - oz where
Zo oz
(xo OX)2
= x~
2xo ox
+ (OX)2
we can ignore (OX)2, since ox is assumed to be small compared with Xo, and equate Zo to x~, giving us the value
of oz as
oz = 2xo ox
This can more conveniently be expressed in terms of the
relative uncertainty oz/ Zo as
51
ox/x~
oz/Zo = 2xo
= 2ox/xo
Thus, the relative uncertainty of the computed result

twice that of the initial measurement.
IS
Although it is essential to bear in mind the nature of the

propagation of uncertainty, as illustrated by this example
with finite differences, a considerable simplification of the
formulation can result from the use of the techniques of
the differential calculus.
3.3 General Method for Uncertainty in Functions of a
Single Variable
It will be noticed that these finite differences oz and Ox are

merely an expression of the derivative dz / dx . We can therefore obtain our value of oz by using standard techniques
to obtain
t~
dx
J'(x)
and then writing
oz = J'(x) Ox
(3.1)
This is a relatively simple procedure and will work in cases

where the elementary finite difference approach would lead
to algebraic complexity.
Thus, if
z = x2
dz
x2
-- =
dx
+1
+ 1 - x . 2x
(x + 1)2
2
52
This would have been very awkward by any other approach. It gives oz generally as a function of x and ox,
and the particular value desired would be obtained by
setting x = Xo. Let us now use this technique to evaluate
the uncertainty for some common functions.
(a) Powers
Consider
dz
nxn
oz
nxn-1ox
dx
The significance of this result becomes a little more obvious

when expressed in terms of the relative uncertainty. Thus,
OZ
-
ox
nx
This will hold for either powers or roots, so that the precision diminishes as a quantity is raised to powers or improves
on taking roots. This is a situation which must be carefully
watched in an experiment in which powers are involved.
The higher the power, the greater is the initial precision
that is needed.
(b) Trigonometric Functions
We shall do only one example since all the others can be
treated in a similar fashion.
z =
Consider
dz
-
dx
SIn
cos x
OZ = cos x ox
This is one case where the elementary method of inserting
53
xo Ox shows the nature of the result more clearly. This

substitution can be easily verified to give
oz =
cos x sin Ox
showing that the Ox in the previous result is really sin Ox in

the limit. Only in the case of a very large uncertainty
would this difference be significant, but it is best to understand the nature of the result. Clearly Ox should be expressed in radian measure. The result will normally have
straightforward application when dealing with apparatus
such as the spectrometer.
(c) Logarithmic and Exponential Functions
Consider
z = log x
dz
dx
oz =
Ox
and the relative uncertainty can be calculated as usual.

If
z = eX
dz = eX
dx
oz
eX
Ox
This is a rather more important case since the exponential

function is one of common occurrence in physics and engineering. These functions can become very sensitive to the
exponent when it takes values much over unity, and the
uncertainty oz can be seen to have potentially large values.
This will be familiar to anyone who has watched the cur-
54
rent fluctuations in a thermionic diode which can result

from quite small filament temperature variations.
As stated above, the method can be easily applied to any
function not listed above by evaluating the appropriate
derivative and using Equation (3.1).
3.4 Uncertainty in Functions of Two or More Variables
If the resul t is to be computed from two or more measured

quantities, x and y, the uncertainty in the result can be
regarded in two different ways. We can, first, be as pessimistic as possible and suppose that the actual deviations
of x and y happen to combine additively in such a way that
the value of z is driven as far as possible from the central
value. We shall, in this way, calculate a oz which gives
the extreme width of the range of possible z values. It is
possible to argue against this that the probability is small
of a number of uncertainties combining in magnitude and
direction to give the worst possible result for z. This is true,
and we shall deal later with the matter of the probable uncertainty in z. For the moment, however, let us calculate
the oz which represents the widest range of possibility of z.
This is certainly a safe, though pessimistic, approach since
if
etc. represent limits within which we are "almost
certain" the actual value lies, then this OZ will give limits
within which we are equally certain that the actual value
of z lies.
ox, oy
The most instructive approach initially is to use the elementary substitution method, and we shall use this for the
first two functions
55
(a) Sum oj Two or More Variables

Consider
z
+y
The uncertainty in z will be obtained from
oz
= Xo
ox
~ =
+ yo
oy
and the maximum value of oz is given by choosing similar
signs throughout. As might be expected, the uncertainty
in the sum is just the sum of the individual uncertainties.
This can be expressed in terms of relative uncertainties
Zo
+ 0)'
z
x+y
but no increased clarification is achieved.
(b) Difference oj Two Variables
Consider
z = x - Y
As in the case above, oz will be obtained from

Zo
oz
(xo ox) - (yo oy)
Thus, we can obtain the maximum value of oz by choosing

the negative sign for oy giving, once again,
OZ = ox
+ oy
The significance of this is more clearly apparent if we consider the relative uncertainty given by
~ = ox
02'
Z
x-y
This shows that, if Xo and yo are close together, x - y IS

small, and this relative uncertainty can rise to very large
values. This is, at best, an unsatisfactory situation and it
can become sufficiently bad to destroy the value of the
56
measurement. It is a particularly dangerous condition since

it can arise unnoticed. It is perfectly obvious that no one
would attempt to measure the distance between two points
a millimeter apart by measuring the distance of each from
a third point a meter away, and then subtracting the two
lengths. However, it can happen that a desired result is to
be obtained by subtraction of two measurements made
separately (two thermometers, clocks, etc.) and the character of the measurement as a difference may not be
strikingly obvious. All measurements involving differences
should be treated with the greatest caution. Clearly the
way to avoid this difficulty is to measure the difference
directly, rather than obtain it by subtraction between two
measured quantities. For example if one has an apparatus
within which two points are at potentials above ground of
VI = 1500 v and V2 = 1510 v respectively, and the required
quantity is V2 - VI, only a very high quality voltmeter
would permit the values of VI and V2 to be measured to
give V2 - VI with even say 10 per cent. But an ordinary
10 v table voltmeter connected between the two points and
measuring V2 - VI directly will immediately give the
answer with 2-3 per cent precision.
1.5 General Method for Uncertainty in Functions of Two

or More Variables
These last two examples, treated by the elementary method,

suggest that, once more, the differential calculus may offer
a considerable simplification of the treatment. It is clear
that if we have
z = j(x,y)
57
the appropriate quantity required in order to calculate

is the total differential dz, given by
dz =
oz
(*) dx + (~) dy
We shall take this differential and treat it as a finite difference, oz, given, in terms of the uncertainties ox and oy, by
oz =
(*) ox + (~) By
(3.2)
where the derivatives aj/ax and aj/ay will normally be

evaluated for the values Xo, yo at which oz is required. We
shall find that the sign of aj/ax or aj/ay may be negative,
in which case, using our pessimistic requirement for a maximum value of oz, we shaH choose negative values for the
appropriate ox or oy giving a wholly positive contribution
to the sum.
(a) Product oj Two or More Variables

Suppose
z = xy
Using Equation (3.2) we need
az
ax = y
Thus, the value of
and
az =x
ay
oz is given by
oz = y ox + x oy
The significance of this is more clearly seen in the relative

uncertainty
oz = ox
z
+~
y
i.e., when the result is a product of two variables, the rela-
58
tive uncertainty of the result is the sum of the relative uncertainties of the components.
The most general case of a compound function, and one
very commonly found in physics, is the one in which an
algebraic product has components raised to powers in the
form
z = x'JIb
where a and b may be positive or negative, integral or

fractional powers. In this case the formulation is greatly
simplified by taking logs of both sides before doing the differentiating.
Thus,
log z = a log x
b logy
Therefore, differentiating implicity,
~~
= a
~~
x
+ b dy
Y
As usual, we take the differentials to be finite differences,

giving
oz = a ox + b.~
z
Note that this process gives the relative uncertainty directly.

This is frequently convenient but, if the absolute uncertainty
is required, it is simply evaluated by multiplying
by the computed value zo, which is normally available.
This form of implicit differentiation is still the simplest even
when z is itself raised to some power. For if the equation
reads
Z2 = xy
oz
it is unnecessary to rewrite it
z =
Xl/yl/2
~9
and work from there because, by taking logs,

2 log z = log x
.
l.e.,
giving
oz
z
+ logy
~ + .~
V
ozl z as required.
(b) Quotients
These come under the heading of the previous section,
which permits negative values, and we repeat that the maximum value of oz will be obtained by neglecting the negative sign in the differential.
If a function other than those already listed is encountered,
some kind of a differentiation will usually be found to work.
It is frequently a convenience to differentiate an equation
implicitly, thus simplifying the working by avoiding the
necessity for calculating the unknown explicitly as a function of the other variables. For example, consider the lens
equation
!=!+l
s
s'
where] is a function of the measured quantities sand s'.

We can differentiate the equation implicitly to obtain
df _
-]2 -
ds
-~
ds'
S'2
It is now possible to calculate d] or d]I] directly and more

easily than would have been the case by writing] explicitly
as a function of sand s'. Thus, a formula may be prepared
for the uncertainty into which all the unknowns can be
inserted directly. Make sure that the appropriate signs are
used so that the contributions to the resultant uncertainty
60
THE PROPAGATION OF tTNCERTAINTmS
all add positively to give the outer limits of possibility for

the answer.
If the function is too big and complicated to work out a
value of oz in general, one can always take the measured
values Xo, yo and work out zoo Then if one evaluates the
result by substituting the actual numerical values of Xo
ox,
yo
oy (or yo - oy if appropriate) to give one of the
outer values of z and then repeating the other way, the
limits on z have been determined and oz obtained.
3.6 Compensating Errors
A special situation can arise when compound variables are

involved. Consider, for example, the well-known relation
for the angle of minimum deviation D in a prism of refractive index p, and vertical angle A
p,=
sin! (A
D)
sin! A
If A and D are measured variables with uncertainties oA

and oD, the quantity p, will be the required answer, with
an uncertainty op,. It would be fallacious, however, to cal.
culate the uncertainty in A
D, then in sin! (A
D), and
combine it with the uncertainty in sin! A, treating the function as a quotient of two variables. This can be seen by
thinking of the effect on p, of an increase in A. Both
sin! (A
D) and sin! A increase, and the change in p, is
not correspondingly large. The fallacy is in the application
of the particular methods of the previous sections to variD and A). The
ables which are not independent (e.g., A
cure is either to reduce the equation to a form in which the
61
variables are all independent, or else to go back to first

principles and use the equation of Sec. 3.5 directly.
Cases which involve compensation of errors should be
watched carefully since they can, if treated incorrectly,
give rise to large errors in uncertainty calculations.
3.7 Standard Deviation of Computed Values: General Methods
As has been frequently stressed, this last section has been

concerned with outer limits of possibility for the computed
value z. We have already suggested that this represents an
unrealistically pessimistic approach and that the more useful quantity would be a probable value for oz, provided we
can attach a numerical meaning to "probable." The limits
given by this quantity will be smaller than oz, but we
have the hope of an actual numerical significance for them.
Such statistical validity will be possible only if the uncertainties in x and y have such validity, and we shall, therefore, assume that the measurements have been sufficiently
numerous to justify a calculation of the standard deviation
of the x values Sx, and correspondingly, of SII' We then
hope to be able to calculate an SZ'
However, we must first inquire what we mean by Sz. We
assume that the measurement has taken the form of pairs
of observations x, y (for example, the current through and
the potential across a resistor, which have been measured
with the aim of obtaining the resistance) obtained by repetition under the same conditions. Each pair will define a
value of z and, if the repetition had yielded n pairs, we
shall have a set of n values of z showing statistical fluctuations. The quantity we require, Sz, is the standard devia-
62
tion of this set of z values. Now these individual z values

may never be calculated, because one would calculate the
means x and y and obtain z directly using the assumption
(valid if s"" SIl and Sz are small compared, respectively,
with x, y, and z) that
z = j(x,y)
Nevertheless, that is the significance of the
to calculate.
Sz
we are about
If we assume that the universes of the x, y, and z values

have a Gaussian distribution, the quantity U z (of which we
are about to calculate the best estimate in terms of sz) wi"1
have the usual significance that any z value will stand a
68 per cent chance of falling within Uz of the true value.
As before, let
z = j(x,y)
and consider perturbations
tion oz given by
oz =
OX, oy which lead to a perturba-
(::)
ox + (~;) oy
This perturbation can be used to calculate a standard deviation for the n different z values since
Thus
s; =
=
~}; [(::) Ox + (~;) oy
!}; [(OZ)2(OX)2 + (OZ)2 (0

n
ox
oy
+ 2 (::) (~;) ox oyJ
ry
)2
63
(aZ)2 ~~)2 + (~~)2 ~ (oy)2

aX
n
ay
n
But
and, since ox, oy may be considered for the present purpose

to be independent perturbations,
~
ox oy
Thus, finally
s.
v(az/ax)2s; +
(az/ay)2s~
(3.3)
If z is a function of more than two variables the equation

is extended by adding similar terms.
Thus, if the components of a calculation have standard
deviations of some degree of reliability, a value can be found
for the probable uncertainty of the answer where "probable" has a real numerical significance.
The calculation has been carried out in terms of the variance or standard deviation of the x and y distributions.
However, in actual practice the quantities we want are the
best estimates of U x , U y , etc., and so we would use the
modified value with denominator n - 1 in accordance with
Equation (2.9). The result would then be a best estimate
for U z The standard deviation of the mean for z can then
be calculated by direct use of Equation (2.7) and this will
give the limits within which the mean value of z, 1, stands
a 68 per cent chance of falling.
Note that most actual experiments do not accord with the
assumptions of the development just given. If we are me as-
64
uring the flow rate of water through a pipe, we shall

measure the flow rate, pipe radius and pipe length independently and each one with a number of readings dictated
by the intrinsic precision of the measurement. We cannot,
therefore, use Equation (3.3) directly, since the various s's
are different types of quantity. The solution is to calculate
the standard deviation of the mean for each of the elementary quantities first. If these are used in Equation (3.3),
the result of the calculation will be immediately a standard
deviation of the mean for z.
3.8 Standard Deviation of Computed Values: Special Cases
Let us now apply Equation (3.3) to a few common examples. In all the following cases the various s's are all assumed to be best estimates of the appropriate universe
value (T.
(a) Sum
of
Two Variables
z = x
oz
hence
ox
= 1
+y
oz
' oy
= 1
and
Note that this result affords ajustification for Equation (2.7)
on page 33. The mean value for the sample, ~ x;jn, is just
a function such as z = x y, where x and y happen to be
independent measurements of the same quantity. Thus if
z = -1n
oz
OXI
+ + Xa + ... )
oz = -,1 etc.
1
-,
(Xl
X2
OX2
65
and
s.
~GY s; + (~Y s; +
_/--
v ns2/n2
x
_x
V~
(b) Difference oj Two Variables
Here
az
-1
'ay
sz =
but again
az
= 1
ax
x -
- /
+ sv
sx
As dealt with in Sec. 3.4 on page 56, the previous considerations regarding measurements of differences are still valid.
(c) Product oj Two Variables
z= xy
az
hence
thus
ax = y,
Sz
az
- =
~y
+ x s;
= Vy 2s;
and the specific value for Sz at the particular values Xo, yo

of x and y would be obtained by substituting Xo and yo in
the equations.
Just as in the previously treated case of products, the equation is more clearly expressed in terms of relative values of
s i.e. s./z. We obtain
sz
z
(d) Variables Raised to Powers

z = x tJ
az -_ a..
ax
.A-I
66
Again this is more instructive when expressed in terms of

the relative value
sz
2 s!
- _
- ~a
-2
z
sz
=ax
(e) The General Case of Powers and Products
z=
x~l
Obviously the results of (c) and (d) can be extended to

give the result
In this result note that the presence of negative indices in

the original function is unimportant, since they occur only
squared in the expression for Sz.
If a function other than those listed above is encountered,
the use of Equation (3.3) will yield the required result.

It can be seen that, for the case of a function of a single
variable, Z = I(x), Equation (3.3) reduces to the same form
as that for uncertainties, Equation (3.1). The result is,
therefore, the same for standard deviations as it was for
uncertainties in the case of the trigonometric, exponential
and logarithmic functions treated in Sec. 3.3.
Note that, although we listed in Secs. 3.2 to 3.5 a number
of different approaches to the problem of outside limits to
uncertainty, the standard deviation of z is a uniquely defined quantity and there is no alternative to the use of
E0.luation (3.3).
67
3.9 Combination of Different Types of Uncertainty
Unfortunately for the mathematical elegance of the development, it very frequently occurs that the uncertainty
in a computed result is required when the component quantities have different types of uncertainty. Thus we may
require the uncertainty in
z = j(x,y)
where x is a quantity to which have been assigned outer
limits ox within which we are "almost certain" that the
actual value lies and y is a quantity whose uncertainty is
statistical in nature, and for which a sample standard deviation Sy and a standard deviation of the mean Sy/v'; have
been calculated. We require the uncertainty in z. The
problem is that the uncertainty in z is a difficult thing even
to define. We are trying to combine two quantities which
have, in effect, completely different distribution curves.
One is the standard Gaussian function but the other is a
rectangle, bounded by the outer limits of uncertainty, and
flat on top because the actual value of the unknown x is
equally likely to be anywhere between the outer limits
xo ox. Any general method of solving this problem is
likely to be far too complex for general use, but particular
solutions can be found following a method suggested by
Dr. T. M. Brown.
In the calculation for z one uses the sample mean y for the
y value. This has the significance that it stands approximately a i chance of coming within Sy/v'; of the true
value. Let us therefore calculate limits for x which, similarly, give a i probability of enclosing the true value. Since
the probability distribution for x is rectangular, i of the
68
area under the distribution curve is enclosed by limits

which are separated by a distance equal to j of the total
range of possibility, i.e., j of 2 ox. The limits for j probabili ty are therefore t ox or j ox.
This quantity, j ox, is therefore one which can be compared
with sill V;, since both quantities correspond to j probability. Equation (3.3) can now be used, inserting j ox for the
value of the standard deviation of the mean for x and
sjVn for the y function. This will yield a quantity for the
uncertainty in z which has a ~ probability attached to it,
and will serve instead of an SZ. Note, however, that it would
not be true to say that 95 per cent probability would be
represented by limits twice as widely spaced as those calculated for j probability. The limits for 95 per cent probability would have to be calculated separately using the method
above.
3.10 Application of Results
Two approaches to the problem of calculating the uncertainty of a computed value have been used in this
chapter. In the first the pessimistic calculation was made
of the outer limits of possibility for the answer. The use of
such a calculation is restricted to cases in which there does
not exist sufficient precision of measurement to justify the
calculation of a standard deviation. This would be the
case where the scale is not divided with sufficient fineness
to permit the statistical fluctuations to be observed. The
use of the outer limits would also be ,appropriate in the
preliminary analysis of an experiment,as. described later,
to serve as a guide to the conduct of the experiment. The
use of the second approach is limited, as stated above, to
69
cases of genuine statistical significance. These will usually

be encountered in the evaluation of an experiment on completion of the collection of observations. Here the emphasis
will be on the precision actually achieved, and a quantity
with numerical significance will be sought.
PROBLEMS
1. A meter stick (read to the nearest mm) is used to measure a

length of 12 cm. What is the absolute uncertainty? What is the
relative uncertainty?
2. A travelling microscope can be read to 0.1 mm. What is the
precision of the measurement of a distance of 1 em?
3. What is the smallest distance which can be measured using a
meter stick (read to mm) so that the uncertainty shall not exceed
(a) 1 per cent, (b) 5 per cent?
4. A distance of 2 cm must be measured to 1 per cent. (a) Would
a meter stick be suitable? (b) Would a travelling microscope
(read to -fo mm)?
5. A barometer reading normal atmospheric pressure can be
read to 0.1 mm. What is the precision?
6. An ammeter reading 0-5 amp is graduated m 0.1 amp.
Assuming that it is read to the nearest scale division, what is the
precision of measurement (a) at full scale? (b) at 1 amp?
7. A stop watch is graduated in ! sec. What is the minimum
time interval which can be measured with a precision of (a) 5 per
cent, (b) 0.1 percent?
8. A wrist watch gains 1 min/ day. What is the precision with
which it can be used to time an interval of 1 hr?
9. It is stated that today is 5.4 warmer than yesterday. Both
10
measurements were made on the same thermometer read to 0.2.

What is the precision of the statement?
10. In measuring a resistance the voltage is read as 5.4 v and the
current as 1.3 amp (both read to the first decimal place). What
is the absolute uncertainty of the resistance value?
11. A density measurement gives the following figures: mass,
24.32 g 0.005; volume, 10.2 .05 cc. What is the absolute
uncertainty in the density?
12. A simple pendulum experiment to measure g using
T = 211' Vl/g gave T to 2 per cent and I to 1.5 per cent. What
is the precision of the g value?
13. Young's modulus Y for a material can be found from the
deflection of a loaded beam using the equation
=
= deflection =
d
4Wl3
Yab 3
14.2 0.1 cm
W = load = 500 g (exact)
a = beam width = 2.1 cm
0.05
b = beam thickness = 0.3 cm 0.05

I = beam length = 45.1 cm 0.1
What is the absolute uncertainty in the measurement of Y?
14. The focal length of a thin lens is measured using the equation
1+1=!
s
s'
s is found to be 24.3 0.05 cm, s' 17.4

precision of the f measurement?
0.05 m. What is the
15. Using a diffraction grating for which
d sin 8 = >.
angles 8 are measured to 1 min of arc. A wavelength >. is determined from a 8 measurement of 15 35'. What is the relative
71
uncertainty of the A measurement? (d can be considered to be

precise).
16. A-c measurements at angular frequency ware made on a
series circuit of a resistor and inductor. The impedance Z is
given by
The resistance R is known to be 50 ohms with a precision of :;

per cent, L is known to be 2 henry with a precision of 10 per cent
and w is exactly 211" X 60. What is the absolute uncertainty in
the measurement of Z?
17. Ice was added to water in a measurement of the heat of
fusion, H, giving the equation
miH
+ mi T
= mw(Tl
T 2)
mi = mass of ice = 14.2 0.1 g

Tl = initial temperature of water = 25.40 0.1
T2 = final temperature of water = 7.8 0 0.1
mw = mass of water = 72.3 0.1 g
What is the absolute uncertainty in the measured heat of fusion?
18. The coefficient of linear expansion a of a solid is to be
measured using the equation
I
= 10(1
+ all T)
The length 10 is about 50 cm and the expansion I - 10 can be

measured to 5/100 mm. Knowing that a is about 2 X 10-5 and
neglecting the uncertainties in measuring 10 and Il T, calculate
the minimum temperature range Il T which will permit a to be
measured to 10 per cent.
19. A 1 m slide wire is used as two arms of a Wheatstone bridge
so that an unknown resistance R is calculated from
R =
~R.
12
72
R" the standard resistor is 10.0 0.05 ohms, 11 is 72.3 0.05

and 11 + 1% = 100 exactly. What is the absolute uncertainty in
R?
20. The heat capacity, S, of a liquid is measured by a continuous
flow calorimeter for which
VI = JSQ(T2
T 1)
V and I can be read to 2 per cent each, J is known exactly, Q

can be measured to t per cent and the thermometers for T1 and
T! can be read to 0.1. What is the minimum value of T2 - T1
which will allow S to be measured to 10 per cent?
21. The decay constant of a ballistic galvanometer can be obtained from the equation
where 81 and 82 are successive deflections (on the same side) in
the damped oscillation and T is the period of vibration. Twas
measured to be 5.4 sec 0.1 and 81 and ()2 were observed to be
24.1 and 16.5, each measurement being uncertain to 0.2.
What is the uncertainty in the k measurement?
22. The resistance R of a parallel network of two resistors R1 and
R2 is given by
R
where
R1
= 5.4 0.1,
R1R2
R1
R2
R2
= 1.4 0.05
In the form of the equation given here this is a compensating

error case. Work it out the wrong way (i.e., finding the uncertainty in R1R2 and in R1
R2 and combining) and the correct
way so as to see the difference.
23. An experiment to measure the refractive index J.' of a prism

uses the relation
sin teA
D)
J.'=
sin tA
73
The angle A is measured to be 60 2'; the angle D is

23 14' 2'. What is the uncertainty in p.?
24. Repeated measurements of the diameter of a wire of circular
cross section gave a mean of 0.41 mm with a sample standard
deviation of 0.07. What is the sample standard deviation of the
resulting calculation of the cross-sectional area?
25. The wavelengths of the two yellow lines 10 the sodium
spectrum are measured to be 5891.1 A with a standard deviation
of 1.5 A and 5896.8 with a standard deviation of 1.5 A. What is
the standard deviation for the wavelength difference between
the two lines?
26. A potentiometer with a 1 m wire is being used to measure
the emf of a cell in terms of that of a standard which is known
exactly as 1.0183 v. Only a rather insensitive null point galvanometer was available and 20 determinations of the balance
point gave a mean of 68.3 cm with a sample standard deviation
of 1.4 cm. Using
V unknown
[1
t; Vstandard
calculate the standard deviation of the mean for
VUnknown.
27. A simple pendulum is used to measure g using
T = 211" Vl/g
20 measurements of T give a mean of 1.82 sec and a sample
standard deviation 0[0.06 . Ten measurements of I give a mean of
82.3 cm and a sample standard deviation of 1.4. What is the
standard deviation of the mean for g?
28. The coefficient of viscosity of water
PoiseuiIlc's equation
IS
being measured by
P has been measured to be definitely between 17 and 18 cm of
74
water pressure, Q to be within 1 of 204 cc/min, I was measured

with a meter stick (read to mm) to be 32.1 cm, a was difficult to
measure and was observed 10 times yielding a mean of 1.2 mm
and a sample standard deviation of 0.2 mm. Within which
limits are we 68 per cent certain that the value of 1] lies?
The Nature of ExperImenting
4.1 Nature of Sc:ientiflc: Theory
An experiment can be defined as any planned observational

process by which man increases his experience of the external world. This definition is really so broad as to be
useless, but it does serve to emphasize the enormous range
of human activity which comes under the heading of experimenting. Because of this range, the advice which could
be given to help a student is likely to be so generalized and
vague as to be of very little practical assistance. Even so,
the nature of the experimental process in principle, and its
role in establishing the nature of human knowledge should
be part of the education of everyone who is likely to be
faced with the problem of gathering observational information of any type. For the present purpose we shall restrict ourselves to a few general remarks and then consider
especially the principles of experimenting in the physics
laboratory.
The general course by which a science grows is normally an
alternation of experimenting and thinking. The sequence
is usually observation-hypothesis-experiment. The first
step in the sequence is often the chance observation of a new
phenomt>.llon. Since this is automatically outside the range
75
76
THE NATURE OF EXPERIMENTING
of previous human experience, the observation is naturally

followed by speculation regarding the nature of the phenomenon. This speculation will probably result in some
genius introducing a new idea which has been stimulated by
the observation. For example, the observation might be
the mode of scattering of a particles from a gold foil, and
the resultant proposal is that of Rutherford's nuclear atom.
Note that our theoretician could, in principle, have thought
of this without knowledge of the experimental discovery.
The experiment thus acts only as a trigger on the imagination of the scientist but, because of the vast range of
speculative ideas which the theoretical scientist could have
produced, the observation is an absolutely essential guide.
It may be that the axiom which is accepted is not a completely new idea but is nothing more than an idealization
of observed behavior e.g., Ohm's law as a basis for circuit
theory, or Hooke's law in the theory of elastic media. This
axiomatic foundation for the development is frequently
called a "model."
Such exercise of the imagination by our theoretical scientist
is perhaps amusing but not really profitable until he can
develop the idea to the point when he can make a prediction about some aspect of the phenomenon which has
not yet been treated experimentally. This development
gives the scientist's idea the ;;tatus of a theory, and the
predictions made will immediately suggest experiments.
The requirement to subject theoretical predictions to experimental test can be used to define the "scientific method."
The results of these experiments decide whether the original
speculation was fruitful or not. If it was a good idea, the
predictions of the theory will conform with the experimental observations. This does not mean, of course, that
77
the theory is "right," since the only evidence we have in its

favor is correspondence with experiment within a certain
degree of instrumental precision. It would be a bold scientist who ever claimed that a theory represented absolute
truth about nature. If the so-called laws of physics of the
present day seem unshakable to the reader, he should reflect a little on the strength of the convictions held by the
propounders of the Ptolemaic system of planetary orbits or
of phlogiston. One can, however, prove ideas to be wrong
more or less easily. The postulate that nights are dark because the sun is swallowed by a dragon at the end of each
day is a very satisfactory one, for which every evening provides fresh confirmation. It is satisfactory, however, only so
long as the observer stays in one place. An expedition to the
opposite side of the earth could easily relay radio signals
back home which would transmit evidence that the sun
had not been swallowed but had merely gone around the
corner. In general, then, one can prove ideas about nature
to be wrong but one cannot prove them to be right.
Note that this concept of rightness or wrongness applies
only when we wish to make a statement about the natural
world. Once the basis for a theory has been postulated
(such as Euclidean geometry or Newtonian mechanics)
there is no question about the "truth" of the deductions,
where one defines "truth" as "correctly derived from the
axioms." These deductions, however, are statements about
the theory, not about the natural world. This philosophical
question arises only at the frontiers of knowledge where
ques~ions of the nature of the universe are considered. In
everyday work in the laboratory one is not so much concerned with "rightness" or "wrongness." It is better to
talk of "appropriateness," because even supplanted theories
78
are used extensively, normally because they tend to be

simpler than their replacements. The introduction of
general relativity by Einstein does not destroy the usefulness
of the Newtonian inverse square force "law," and the introduction of wave mechanics does not prevent one from
sometimes thinking of electrons as little charged billiard
balls. One uses theories more on a basis of convenience
than anything else, provided they give satisfactory correspondence with experiment at the level of precision being
considered.
It is frequently found, when subjecting a theory to experimental investigation, that one does not find complete correspondence, or lack of it, between prediction and observation. More commonly it is found that the original
speculation is partly satisfactory. If the idea is too elementary (as it is almost bound to be when first propounded)
then the experiment will probably show correspondence
with the predictions within a limited range only. This
partial discrepancy then constitutes another observation
which starts the cycle all over again.
The structure of scientific thought is therefore a complex of

deductions from a set of observation-stimulated axioms.
Sometimes the axiomatic nature of the science is clear, as in
the case of Euclidean geometry, but, even in the most
empirical and observational sciences it is still there. One
discusses the resistance of a metal wire in terms of Ohm's
law or the motion of electrons in a cathode ray tube using
the charged billiard ball model. It is important that this
axiomatic structure be recognized, and rem"embered, so
that the significance of any scientific or technological statement or conclusion may be judged. The axiomatic found a-
79
tion provides the background of thought to which all experimental processes are referred.
4.2 Types of Experiment
As has been stated, the range of activity covered by the term

experimenting is so broad that a detailed description is
almost useless. There are, however, some considerations
which enter into every type of work, and which help to
influence the conduct of the work. These considerations,
listed below, will enter an experimental situation in varying
degree and any practical problem will be constructed out
of them.
The three main factors involved in determining the type of
experiment are:-completeness of background material,
degree of control over the subject material, and degree of
statistical fluctuation. The influences of these will be considered in turn.
(a) Degree oj Reference to Background

In simple terms this is nothing more than the degree of
familiarity with the subject of the experiment. It obviously
can vary widely between extreme limits of complete familiarity and complete strangeness. It can be quite close, as in
the case of a measurement of a quantity using a theory,
concept or principle which is well established. An example
of such a situation would be the measurement of the electrical resistivity of a copper wire using a Wheatstone bridge.
In this C<.t:;c every aspect of the measurement is well-tried,
familiar , t!' d, wi thin certain limits of precision, the answer
is unequivoca1. No difficulties in interpretation arise be-
~o
cause everyone knows what the experimenter

about.
IS
talking
However, at the other end of the scale the degree of reference to background material may be very limited since, in
the case of a newly discovered phenomenon there is no
background material which is directly relevant. The study
of this type of situation constitutes exploratory type research, and this is very difficult work just because of the
lack of guidance. The aim of the work is, in this case, to
gather as much information as possible (usually as quickly
as possible because of human impatience) covering as wide
a range of approach as possible. This will limit the range
of speculation about the phenomenon and increase the
probability of someone making a lucky guess. Prior to this
lucky guess the interpretation of the measurements is very
difficult in view of the lack of guidance towards such an
interpretation. The only course open to the experimenter
is to try to find some function which fits the observation
and which he hopes will act as a guide to theoretical
thought. In this he may be optimistic since many empirically established functions have turned out later to have no
relation whatsoever to the theoretical functions later
developed.
The point to be noted is that, although the familiarity of
the concepts of an experiment range along the whole scale
as described above, some degree of reference to background
ideas is always found. This may be obvious half way along
the familiarity scale where one of the points of an experiment may be the establishment of the validity of the concepts on which the experiment is based. One would not
81
quote a measured value for the half life of a radioactive

isotope without verifying an exponential law for the decay
of its activity. However, the degree of reference to background material may be forgotten at the two ends of the
familiarity scale, and for a different reason each time. At
the familiar end one tends, just because of the familiarity,
to forget the axiomatic foundation of the measurement.
The observer with his Wheatstone bridge may forget that
he is automatically invoking Ohm's law. Even such a
seemingly clear-cut measurement as the distance between
two buildings on the University campus involves an interpretation of the theodolite readings in terms of the axioms
of Euclidean geometry. At the other end of the familiarity
scale the experimenter has to be especially careful with the
interpretation of his newly discovered and still obscure
phenomenon. He will automatically use his already existing ideas in considering it, just because, from the very
nature of the situation, he has nothing else available. There
is, however, no guarantee that the earlier ideas have any
validity for the new phenomenon. Even such a startling
discovery as that of superconductivity in 1911 was, for long,
interpreted in terms of the established concepts of conductivity. The inappropriateness of such an interpretation
will be revealed only by the development of a theory based
on it and subsequent comparison with experiment.
To summarize, we repeat that, even though the degree of
reference to background in an experiment may vary, it is
always there. The concept of the impersonal, objective
measurement of the "hard facts of science" is not appropriate. All measurements and experiments are conceived and
interpreted in terms of the current modes of thought.
82
(b) Degree of Control

The degree of control which the experimenter has over his
experimental material varies enormously. The production
staff of a large-scale industrial process which is not functioning correctly may be able to do little more than watch
it in the hope of making a guess at the fault. On the other
hand, a metal specimen in a solid state physics laboratory
may have every known physical property wrung out of it.
In between, there is every gradation of control, and
normally the degree of control is fixed by the circumstances
of the experiment. The student must learn to accept the
available extent of his influence, or lack of it, over the experiment and surmount this difficulty in obtaining his
result.
As one might expect, the difficulties arise under conditions
where control over the actual conduct of the experiment is
poor. Here the whole secret of success is in preliminary
planning because, once the experiment is under way, the
experimenter may be able to do little more than watch its
progress helplessly. The planning of the experiment will
determine if it is going to be successful, and, with poor
planning, the results may not be merely uninformative but
actually may be misleading.
The physics laboratory is misleading to the student in this
respect because, if the student makes a mess of an experiment, the opportunity is commonly available to retrieve
any missing readings or even repeat the whole experiment.
Under actual working conditions, however, this may not
be so easy. A scientist who has been given the job of
measuring the expansion of nuclear reactor fuel rods under
neutron irradiation would be most unpopular if, after
83
wasting many thousands of dollars worth of neutrons in the

reactor, he found he had forgotten to measure the initial
size of his rods. Experiment planning will be considered in
more detail in the next chapter and it will suffice for the
moment to say that the student will find it well worth while
acquiring, at the earliest possible stage, the habit of detailed
planning of an experiment.
The degree of control available during the actual conduct
of the experiment is low in large-scale industrial processes,
as has already been suggested. It is also low in principle in
those experiments of an essentially observational type, such
as are common in sociological work (e.g., What is the most
popular brand of toothpaste in North America? For the
average observer this is not a controllable variable.) In
cases like this the problem reduces, almost invariably, to
one of sampling, and once again the essence of work is
planning, because it is the only activity open to the experimenter. The fallibility of such procedures consequent on
poor planning is well known. Such fallibility can arise
either as a result of sampling errors or because of poorly
defined measurement methods. Consider the failure of the
Gallup Poll forecast in the Landon-Roosevelt presidential
contest in 1936. It used a telephone survey and therefore,
in a day when telephones were less common than they are
now, biased its sample towards a higher income group
favoring the Republican candidate. Consider also failure
of the automobile purchasing surveys in the early 1950's
which showed that people wanted smaller, less brightly
finished cars when in fact they continued to buy longer,
more exotic vehicles for many more years. The defect in
the first instance is a sampling error, and, in the second,
inadequate exactnetii'> in the defining of the quantity to be
84
measured. Methods are available for the programming of

such observational type experiments by statistical methods
to minimize the chance of errors, and many texts are devoted exclusively to this topic. (See, for example, Reference
3 in the BibliogTaphy.)
(c) Statistical Fluctuation

I t is probably true to say that the influence of statistical
fluctuation does not vary a great deal. It is always too
much. This seems to be true whether an experimenter is
studyiBg the influence of fluorine on tooth decay where the
spread of results may be comparable with the size of the
effect, or a metrologist is losing sleep over the ninth significant figure in a wavelength measurement.
The influence of statistical fluctuation has been considered
in the last chapter, and it must be apparent that continued
duplication of readings is unrewarding in a measurement of
low intrinsic precision. One way of attacking this problem
is to measure differences rather than absolute values. The
dangers of obtaining a measurement as the difference between two absolute values has been pointed out earlier and,
if it is possible to measure the difference itself, much higher
precision can be achieved.
The advantages of this are obvious in mere numerical considerations, before any experimental problems are discussed. If one has a resistor of around 1000 ohm resistance
(suppose 1124 ohms) and wishes to know its resistance to
within 1 ohm it is necessary to measure it with a precision
better than -h per cent. If however we have a standard
resistor of value 1000 ohms, and arrange that we measure
thediiference between the two resistances, we must measure
85
this quantity (124 ohms) to a precision of about 1 per cent

only to achieve our objective. The use of "beats" for frequency measurement is another example of this technique.
This use of a standard for comparison purposes is especially
valuable when some perturbing influence is at work on both
the sample and the reference simultaneously. Thus the
effect of temperature changes on the leads to a platinum
resistance thermometer is eliminated by subtracting the
resistance changes of a dummy pair of leads subject to the
same temperature changes, and in instruments measuring
the optical transmittance of liquids, changes in source intensity and detector sensitivity are eliminated by alternating
the light path between the sample and a vacant cell, and
measuring the difference between the two intensities. These
instrumental methods are really a process of continued
calibration, but the same approach is found in other
sciences where it is desired to subtract out the effect of
perturbing influences. The botanist will study the effect of
fertilizer on wheat not by taking a wheat patch, adding
fertilizer and measuring the resulting crop. The perturbing
effects of rainfall, temperature, sunshine, etc. are much too
great for this approach to be effective, and so the fertilized
product is compared with another plot of wheat called a
"control," grown under exactly the same conditions except
for the fertilizer. The difference between the two plots of
wheat is the effect of the fertilizer, which can therefore be
measured in isolation from unwanted influences. Another
example of common occurrence is particle counting in
nuclear physics. Here the perturbation is the addition of
unwanted counts from cosmic rays and a control experiment is necessary in which the "background" counts are
measured by themselves for future subtraction from tht'
86
THE NATURE OF EXPERIMENTNG
main measured value. Note that all these experimental

procedures have one thing in common ; the difference
which is measured is caused by the factor under investigation (the difference between the two plots of wheat is the
fertilizer, the difference between the two counting experiments is the presence of the radioactive source, etc.) and
everything else is held constant. Make sure that the experiment is designed to do that and not something else: it
is rather easier than one might think to try the physics
laboratory equivalent of investigating the effect of fertilizer
by varying the rainfall. (See also on page 114 another example of this technique in the "sample in-sample out"
method.) Such "control" type experiments are especially
valuable when the control and the specimen are subject to
the same fluctuations. Even when they are not, however,
a control or standard is still useful, as mentioned above, for
transforming a -10 per cent measurement into a 1 per cent
measurement.
Many measurements in science are of the type in which the
quantity desired is a function of some variable e.g., electrical resistance as a function of temperature. In many cases
the important quantity is this variation itself and the absolute value is less important. This situation is, once more,
a difference type problem. Suppose a resistor varies between 1000 and 1100 ohms over a certain range and the
variation is required to 1 per cent. Clearly one should
measure the change in resistance directly, either with respect
to the initial resistance itself or in comparison with some
fixed standard of suitable value.
Difference methods are very powerful and should be considered whenever possible in designing experiments. The
87
only requirement in realizing the potentialities of the

method is the reliability and stability of the standard
quantity. There is no point in attempting to measure a
frequency to 1 cps using beats, if the reference oscillater is
unstable to 5 cps.
A special case arises when, in such comparison-type measurements, the reference value can be controlled so that the
measuring instrument reads zero at the time the measurement is being made. This is the case, for example in the
Wheatstone bridge where the reference resistor is adjusted
to give zero current in the galvanometer, in the potentiometer when the unknown emf is balanced against a value
obtainable from the slide wire, and in the optical example
mentioned on p. 85 in which the reference intensity can
be controlled so that the difference being measured is zero.
In all these cases one very important advantage is secured
and that is that the final reading is independent of the
characteristic of the detecting instrument and possibly also
of other components in the system. This constitutes the
so-called "nul" measurement. It is a very valuable method
because it eliminates problems of calibration of instruments
like meters and amplifiers, and replaces them by the preferable problem of calibration of reference resistors, potentiometer slide wires or optical shutters. It may also eliminate
problems of source stability as, for example, in the Wheatstone bridge for which, at the balance point, non-constancy
of the supply battery is unimportant.
:; Experiment Planning
The conclusion from the last chapter is that an experiment

is a much more sophisticated procedure than an isolated
measurement, because it is the investigation of the properties of a physical system. The results of such an investigation must necessarily be expressed in terms of the commonly accepted ideas concerning the nature of the physical
system. The conduct of the experiment will therefore depend very largely on how much of this background material
is available, but in all cases the essence of good experimenting is planning. It is courting disaster to have the (not
uncommon) attitude of rushing through the measurements
and worrying later about what to do with them. The
experimenter should remember this and take time, before
starting an experiment in the laboratory, to analyze his
experiment in the ways to be described, and to layout his
analysis, and the consequent measurement and calculation
program, formally and neatly in his laboratory note book.
We have been stressing the wide range of experimental
activity, but from now on we shall be restricted more to the
practices of the physics laboratory.
88
89
EXPERIMENT PLANNING
5.1 Precision of Measurement
Whatever the nature of the experiment, it will be CCl.nstructed out of measurements, and we must consider these
first. The conduct of the experiment will be determined
very largely by the precision which is required. The instruments used to measure individual quantities, and the
whole method of measurement will depend on it. Thus
every experiment should be considered in the light of some
desired precision for the answer. This quantity should be
chosen realistically, since too optimistic a value will very
quickly lead to too great complexity. A desire to measure
elm for electrons to 10 per cent in an elementary laboratory
would almost certainly lead to disappointment. Once the
experimenter knows, in general terms the precision for
which he is aiming, he must then turn his attention in turn
to each of the quantities involved in the measurement.
Remember at this stage the important distinction between
readings whose precision is limited by statistical fluctuation
and those whose precision is limited by the measuring scale.
The only way to distinguish between those two types is to
try the measurement to see. Does repetition give the same
scale reading or not? Once this point has been settled is
the apparent value of the uncertainty acceptable or not?
If not, then some improvement is required. If the precision
is scale-limited, the acquisition of a more precise instrument
is indicated. If such is not available, then a lower limit to the
uncertainty of measurement is already set. If the uncertainty
of the reading proves to be statistical in nature, obtain an
estimate of the standard deviation using, say, 10 readings.
Does the precision thus calculated appear adequate or not?
90
EXPERIMENT PLANNING
If not, can the precision be improved ei ther by taking more

readings or must the measurement procedure itself be improved? Remember from page 33 how the standard deviation of the mean involves V~. Consequently, if 10 readings suggest that the precision must be improved by a factor
of 10 (the fluctuations may be 10 per cent, when the experimenter wants 1 per cent), the number of readings must be
increased by a factor of 100. This is an undesirable method
of improving precision and, in this case, some improved
measurement procedure is called for. [Refer to Sec. 4.2
(c).] If this is not possible then, once again, a lower limit
to the uncertainty has been set by the time and resources
available for duplication of the readings. The whole skill
in experiment design lies in the optimum choice of the above
procedures to give maximum precision, taking into account
the available resources of equipment, time and money.
At this stage use the methods of Chapter 3 to find out how
each elementary uncertainty is propagated through to the
final answer. This will give an estimate of the over-all
uncertainty to be found in the experiment, and enable the
experimenter to identify those quantities which dominate
the construction of the final uncertainty either because of
poor intrinsic precision or because they are raised to a high
power. The experimenter is now in the position of having
considered fully the measurement aspect of his experiment.
He has either secured the apparatus necessary to achieve
the required precision, or else he has obtained an estimate
of the precision to which he is limited by the resources available. The subsequent planning of the experiment will depend on how much theoretical material he has to suggest
the behavior of his system.
91
EXPERIMENT PLANNING
5.2 Experimenting with No Background
The circumstances here are that we are presented with a

physical situation about whose characteristics we know
virtually nothing. The situation might involve the strength
of a concrete mix which is dependent on the proportions of
its constituents, the angular dependence of protons scattered from an atomic nucleus, or the efficiency of a gasoline
engine which is known to depend on speed and fuel mixture
richness. The task is to investigate the system. This investigation can have two aims. First, the experimental results will describe the system to anyone else. Once the
curve of concrete strength vs. proportion of sand has been
established, other people can then make up concrete to
have a given strength by using the prescribed quantity of
sand. Secondly, the results will serve as a guide to a
theoretician who wishes to construct a model of the situation. In engineering the first aim is more common, in
physics the second.
In both cases, the design of the experiment starts with a
selection of the variables. There is usually one obvious
variable which we can select to be the main measured
quantity (the strength of the concrete, the intensity of the
scattered proton beam or the efficiency of the engine).
This is called the dependent variable. This quantity is probably influenced by a whole host of other factors and the next
task is to select those whose influence we wish to measure
(proportions of sand, angle of proton scattering, engine
speed, and mixture). These are called independent variables
since we are able to choose their values at will, thus determining tr..:ough the properties of the system the value
92
EXPERIMENT PLANNING
of the dependent variable. Thus, it is important to ensure

that, once the independent variables are chosen, all other
factors which may influence the dependent variable be
held closely constant (e.g., setting time of the concrete,
energy of the incident proton beam, engine intake air
temperature). If everything starts varying simultaneously
we shall find it very difficult to understand the results.
Once the variables are selected and the measurement
techniques established according to the precision desired,
as described in Sec. 5.1, the measurement program can be
constructed. The form of the measurement program will
obviously be the measurement of the dependent variable
for several values of the independent variable, the aim
being to learn as much as possible about the behavior of
the system. The values of this independent variable should
cover as wide a range as possible, and obviously the number
of values should be as large as possible. The actual number
will be predetermined by the time available for the whole
experiment and the time required for each measurement.
If there is more than one independent variable, the technique is to choose one of the independent variables and to
hold it at a constant value while one measures the variation
of the dependent variable on the other. Thereupon another
constant value is chosen for the variable and the process
repeated. Thus, we would have for the gasoline engine a
series of curves of efficiency vs. engine speed, each curve
referring to a certain value of the fuel mixture richness as
shown in Fig. 5.1. If anyone wanted to have the variation
of efficiency with mixture richness for a given speed, he
would have to read the values off the graph along such a
line as ABeD and replot the values accordingly. A variable
which is held constant while an experiment is in progress
93
EXPERIMENT PLANNING
l!..._--__
Mixture 4
L _ - -__ Mixture 3
I __--_-"'Mlixture 2
~_-_.:.:M::i,xture 1
rpm
(a)
Mixture
(b)
Fig. 5.1
Engine efficiency as a function of mixture and rpm.
and given a few discrete values, as described above, could

be termed a "subsidiary variable" as opposed to the main
or principal variables.
The end point of such an investigation is a curve or a set of
curves. If this is an empirical study it may suffice merely
94
EXPERIMENT PLANNING
to present the curves saying, in effect, "Regardless of what

is really going on, this is what the system actually does."
If the work is merely a prelude to theoretical thought,
further analysis is needed and this will be described in
Chapter 6.
Although it is usually taken as an axiom of good experimenting in the scientific laboratory to hold all independent
variables constant except the one under investigation, this
commonly turns out to be a counsel of perfection in more
complex systems, e.g., in industry. Here the variables may
be interdependent so that the effect of varying one may be
to alter all the others which the experimenter had hoped to
keep constant. As an elementary example, consider an experiment to evaluate the static constants of a pentode tube.
The plate and the screen grid are connected through suitable potentiometers to a power supply, and a separate grid
supply is provided. The aim of the experiment might be to
measure a defined parameter of the tube such as the variation of plate current with grid potential for constant screen
potential. One would do this by setting the grid and screen
potentials to the required initial values and reading the
plate current. The next step would be to change the grid
potential to the next required value and read the plate
current again. However, the experimenter is very likely to
find that the altered plate current has changed the source
voltage so that the screen potential is altered from its original value. In other words the supposedly independent
variables are not independent. This is a trivial example, of
course, because it is a simple matter to reset the required
screen potential. However, the experiment may be on a
rather larger scale such as the rate of production in industrial continuous flow processes in which the flow rate
95
EXPERIMENT PLANNING
depends on the reaction rate which depends on the temperature which itself depends on the flow rate which .... ....... ,
or a problem in meteorology in which the variables, wind
velocity, pressure, temperature, humidity, height, etc. may
be all interdependent to a high degree.
All one can say about systems such as these is that the
dependent variable is a function of the independent variables and various constants
If one wants information about such a system, one wants

the values of the constants al, a2, etc. and, to determine
them, all one can do is make observations of the required
quantity under various sets of circumstances with the values
of the various variables falling where they may. With
many variables such a procedure is sure to lead to chaos
unless some system of selecting the observations is used.
The methods which have been developed for choosing
a measurement scheme, which will yield the maximum
amount of information for a given experimental effort, form
a very important part of experiment design, they are of
widespread use in science and technology, and much of
the literature is devoted to them. They are not considered
here further because the physics laboratory normally enjoys
the luxury of controllable variables. For further information the reader is referred to References 3,11, 18, and 19 in
the Bibliography.
5.3 Dimensional Analysis
Even if no complete theory of a physical phenomenon

exists, it is still possible to obtain very useful guidance to
the performance of an experiment by the method of di-
96
EXPERIMENT PLANNING
mensional analysis. The "dimensions" of a physical (mechanical) quantity are its expression in terms of the elementary quantities of mass, length, and time, denoted by
M, L, and T. Thus, velocity has dimensions Lr-1, acceleration Lr-2, density ML-3, force (equal to mass X acceleration) MLr-2, work (equal to force X distance) MUr- 2,
etc.
The principle used in dimensional analysis is that the dimensions on each side of an equation must match. Thus,
if g is known to be dependent on the length and period of a
pendulum, it is obvious that the only way in which the
LT-2 of the acceleration can be balanced on the other side
is by having the length to the first power (to give the L)
and the period squared (to provide 1'-2). We can thus say
immediately that, whatever the final theoretical form for
the equation it must have the structure
'
. Iess constant) X length
g = (d lmenSlOn
~ dO)
peno Note that the treatment can give no information about dimensionless quantities (pure numbers, 7r, etc.) and so we
must always add in such a possibility to the form of an
equation obtained by dimensional analysis.
The general method is as follows: Consider a quantity z
which is assumed to be a function of variables x, y, etc.
Write the relation in the form
za:x~l
where a and b represent the numerical powers to which x

and y have to be raised. The values of a and b will then be
found by writing down the dimensions of the right-hand
side in terms of the dimensions of x and y and the powers
97
EXPERIMENT PLANNING
a and b, and writing down the condition that the total
power of M on the right-hand side must be the same as that

known for z, and similarly for Land T. Three simultaneous
equations result which enable values for a, b, etc. to be
calculated.
For example, consider the velocity v of transverse waves
in a string. We might guess that this velocity is governed
by the tension T in the string and the mass per unit
length m.
Let us write
Lit
(force), MLi 2
(mass/ unit length), ML-l
dimensions of v,
dimensions of T,
dimensions of m,
Therefore,
Lit
(MLT-2)a(ML-t)b
Ma+bLa-bT-2a
Therefore, by comparing powers of

M,
0
1
-1
L,
T,
= a +b
= a- b
= -2a
of which the solutions are obviously

a --
.!
2,
so that we can write
v=
(dimensionless constant) X
VT/m
Such a treatment is very valuable, for it provides, even in

the absence of a detailed fundamental theory, a prediction
regarding the properties of the system. This is available for
experimental investigation and, if our original guess regarding the factors contributing to v was correct, the pre-
98
EXPERIMENT PLANNING
diction will be verified. If experiment shows a discrepancy,

then we must look again at our original guess. Notice that
we obtained three equations for only two unknowns. The
condition, therefore, was really overdetermined and we
were very fortunate that the equations were consistent.
Had they not been consistent we would have known immediately that our guess regarding the constituents of v
was wrong.
Powerful as this method is, difficulties will obviously arise
when the quantity under discussion is a function of more
than three variables. Thus, we shall have more than three
unknown powers but only three equations from which to
determine them. In this case a unique solution is not possible but a partial solution may be found in terms of combinations of some of the variables.
For example, consider the flow rate Q of fluid of viscosity
coefficient 7J through a tube of radius r and length I under
a pressure difference p. We may suggest a relation
ex:
palb7Jcrd
The dimensions of the quantities are as follows:

ViI
Q, volume per unit time,
MLi2 . L-2 = ML- 1i
p, force per unit area,
L
I,
7J, viscosity coefficient is defined as a force per unit area
per unit velocity gradient,
r,
(MLi2)(V)-I(Lii . L-l)-l = ML-1i I

L
Therefore,
99
EXPERIMENT PLANNING
Comparing powers of
M,
L,
T,
0 = a
c
3 = - a
b - c
-1=-2a-c
+d
Here we have four unknowns and only three equations so

that, in general, a complete solution is not possible. We
can find part of it, however, for obviously
= 1
c = -1
and
We must have, therefore,
Q 0;
11
The remaining part of the solution can be written only as
b+d=3
If we write this
d=3-b
we can see that Q must contain the product
contained lb so that we can write finally
Q 0;
r3/r b
It also
~ r3 GY
and this is as far as dimensional analysis can take us

towards the complete solution. However, even this partial
solution would be of enormous assistance as a guide to
experimenting in a situation in which no fundamental
theory existed.
Dimensional analysis can be extended to cover thermal and
electrical quantities, but ambiguities arise which require
special consideration. Treatments will be found in the
standard texts on heat and electricity.
100
EXPERIMENT PLANNING
5.4 Experimenting with a Theoretical Background
The situation is rather different when there is a theoretical background to the experiment. This theoretical background may range all the way from a mere suggestion about
how the system might behave to a well established and
highly developed theory. In many cases where the theory
is more highly developed the quantity which is required to
be measured may be more complicated than the simple
observational properties found in empirical work (the
strength of the concrete or the efficiency of the engine),
and may be defined only in terms of the theory. Such a
quantity would be the constant of gravitation G defined by
the equation for the gravitational force F between two
masses ml and m2 at distance r
This situation uses the inverse square law of gravitational

force as a model and the constant G is meaningless unless
interpreted according to this model. The point is that there
is no a priori guarantee that the conditions described in
the theory match those found in the apparatus. A discrepancy which exists between experiment and theory can range
from the simplest of systematic errors such as an unnoticed
instrumental zero error to definitive evidence on the status
of great theories.
The theory is a deduction from an idealized model (a pendulum is a point mass suspended from a weightless, inextensible string, etc.) and the experiment is conducted on
actual bits of apparatus in the laboratory. The extent to
101
EXPERIMENT PLANNING
which it is valid to interpret the results of the experiment

in terms of the theory is the responsibility of the experimenter to determine. In general, such proof of validity is
absolutely necessary before any experimental result can be
accepted. This is why, even in situations where the topic
is familiar (resistivity and a Wheatstone bridge), the experimenter should keep his theoretical background in mind.
(Some day our Wheatstone bridge operator might be given
a piece of non-ohmic material to measure.) Note that we
do not say the theory is "wrong" or the experiment is
"wrong." It is merely a matter of whether the idealized conditions of the model approach sufficiently closely the conditions of the experiment. If they do not, the behavior of the
actual system will deviate from that of the ideal system
and a systematic error may be introduced. The question
of the importance of this systematic error depends entirely
on the precision of the experiment, and only the experiment
itself will enable one to judge whether the correspondence
between theory and experiment is adequately close. It
should be unnecessary to mention that any superficially
obvious mismatch should be corrected before further work
is done. (If the pendulum string is noticed to be really
stretchy, the experimenter either obtains another piece of
string or else develops the theory of pendulums with elastic
strings.)
In general, then, the question of whether the theory and
experiment are adequately in correspondence is to be settled by observation of the behavior of the system itself,
i.e., is the experiment working out as expected or not? The
clue as to how this is to be determined lies in this expectation because any theoretical result can be regarded as a
102
EXPERIMENT PLANNING
prediction. Thus the expression for the flow rate Q of fluid

of viscosity 11 along a pipe of radius a and length I under
pressure p
_ 7rpa 4
Qis a prediction that
87J1
Q r:r.p
Q r:r.1- 1
Q
r:r. a 4
Clearly it is impossible to judge from one spot reading

whether these predictions are fulfilled or not; the only way
is to take a series of readings with different values of the
variables. Once again we have confirmation of the uselessness of isolated spot readings. Only if these ranges of
variables have been covered as widely as possible, and the
behavior found to be as expected, can the experimenter
claim that his theory and experiment are compatible within
the precision of measurement and that, consequently, his
final answer is a valid one. It is true that not all experimental defects can be detected in this way since a discrepancy may be such as to affect all readings equally, and this
does make things difficult. However, the covering of a
range of values of variables goes very far towards elimination of this source of uncertainty. The only way in which
one can improve the chances of eliminating all systematic
errors is by comparison of measurements of the required
quantity using completely different measurement methods.
Once the range of readings has been taken, the question
arises of the comparison between the observed variation
and that predicted. A column of figures on paper is obviously almost useless since no one can say by looking at
them that the variation is, say, a fourth power. The only
103
EXPERIMENT PLANNING
available way of judging is by graphical methods, because

if the results are laid out in pictorial form, it can become
obvious whether the predictions are fulfilled or not. However, it would not suffice merely to plot Q vs. a if one had
done an experiment with a lot of pipes of differing a while
holding p and l constant. The result of such a procedure
would be a curve, and no one can judge visually whether
such a curve is a fourth power or not. The only curve
which has the convenient property of permitting visual
judgment is a straight line, and so the task is to plot the
observations in such a way that they would form a straight
line if they obey the theoretical predictions. Clearly in this
case one would calculate the quantity a 4 for all the values
of a and plot Q vs. a 4 If the theory and apparatus are in
correspondence, the points will form a straight line and any
lack of correspondence will be revealed by a departure of
the points from linearity along all or part of the series. The
experimenter then knows the range of validity of his experiment. He will therefore calculate his required answer from
the points falling on the straight line, rejecting those which
deviate. This is not dishonest selection of observations but
merely a restriction of the calculations to the range of
validity within which his quantity is defined. For example,
a plot of fluid flow Q vs. p in the previous example might
give a result such as shown in Fig. 5.2. The experimenter
could then restrict his calculation of his required quantity
(probably 7]) to the linear region, and reject the discrepant
points. He could be sure that these last correspond to conditions outside the framework of his theory, whether or not
he knew anything about the onset of turbulence. Another
common form of discrepancy is a shifted origin. A set of
points which should lie on a line passing through the origin
104
Fig. 5.2
EXPERIMENT PLANNING
The identification of the range of validity of an experiment.
might actually give an intercept on one axis but here,

again, a mode of computation using the slope can be
selected so that no systematic error in the answer is permitted. It may not at this stage be possible to say in any
particular case what the discrepancy between theory
and apparatus is, but it does not matter. It is sufficient for
the moment to have it revealed, so that it is not permitted
to introduce errors into the answer.
So far we have described the graph merely as an indicator
of the validity of the experiment. There are many more
105
EXPERIMENT PLANNING
advantages to graphical analysis. These advantages and

the methods of working with graphs must now be described.
5.5 Graphical Analysis
It should be noted here that graphs are of two general

types-pictorial and computational. The first type is used
to illustrate the behavior of a physical system (period vs.
point of suspension for a physical pendulum, thermionic
current vs. plate potential for a thermionic diode), being
merely a pictorial description of the properties of the system, and these are commonly found in texts. They are very
pretty but are no help in judging an experiment or (in
general) in computing an answer. The second type, which
is almost always a straight line, has the purpose of assisting
the evaluation of the experiment and the computation of
an answer, and this is the type with which we are almost
exclusively concerned. In all cases the advantages of graphical presentations of experimental results are overwhelming.
They include: verification of validity of the experiment, as
has been described above; ease of calculation of final answers, as will be described below; a check on the over-all
precision of the experiment since the uncertainties in each
observation will be revealed by scatter of the points.
In order to extract all the information from the observations

it is necessary to plot not only the measured values but their
range of uncertainty (ways of doing this will be suggested
later on page 128). Only if the range of uncertainty is
plotted on the graph will the significance of any deviation
be apparent and will the over-all uncertainty of the experiment be evident.
106
EXPERIMENT PLANNIN("
In considering the linear graph as an aid to computation

we must consider what information is available from a
graph once it is drawn. Two pieces of information are
available, which can be taken as a slope and intercept or as
two intercepts. Clearly the object of the analysis of the
equation into linear form is to cast the unknown (or unknowns if there are two) into the role of constants, i.e.,
slope or intercept, while retaining only measured quantities in the variables. There is no standard method for doing
this and a unique solution does not necessarily exist. The
best way of attacking the problem is to keep clearly in mind
the straight line equation
y=mx+b
while considering the relation relevant to the experiment.
The best way of illustrating the process (known as the
"rectification of the curve") is by examples.
(a) Ohm's law
v=
IR
R = resistance of resistor (constant and unknown)

V = potential (dependent variable)
I = current (independent variable)
This is already linear in the form
y
= slope X
where y is V, x is I and the slope is R . Consequently a set

of measured V and I will enable a straight line to be drawn
and the slope measured, thus yielding R.
(b) Gas law
pv
RT
p = pressure} of one mole of gas

II
= voI ume
107
EXPERIMENT PLANNING
= temperature
R = gas constant per mole
If T has been held constant, the equation reads
pv
constant
This can be rewritten

1
v
P = constant -and, once again, we have straight line form with p and
l/v as y and x respectively. A measurement of the slope
will yield a value for the gas constant R, provided the
temperature is known.
(c) Fluid Flow
If the radius a and length I are kept constant, the equation

has the form
7ra 4
8~P
and is already in straight line form with y as Q, x as p and

the slope as 7ra 4 /8'YJ1.
If a and I are known then 'YJ can be calculated. If a and I
are also variable, a number of methods could be used.
Q vs. p could be plotted as a series of lines, each one referring to discrete values of a and 1 (see page 92). This
describes the system but does not check the a 4 and 1 dependence. One could plot Q vs. a 4 for constant p and I, or
Q vs. I for constant p and a 4 but a rather neater way of summarizing the results would be to use a compound variable.
The equation can be written
108
EXPERIMENT PLANNING
7r pa 4
Q = 8171
where y is Q, x is pa 4 /1 and the slope is 7r /817 (for a precaution regarding this case see page 128). The use of such a
compound variable is perfectly valid and, while optional in
this case, may facilitate the assessment of complicated sets
of observations.
(d) The Compound Pendulum
T = 27rV (h 2
+k
)/
gh
T = period (dependent variable)

h = distance from CG to point of support (independent
variable)
g = gravitational acceleration (constant and unknown)
k = radius of gyration about CG (constant and unknown)
The treatment of this is not obvious, but it is dearly impossible to place it in the required linear form where y
and x are functions of hand T singly. An analysis into
compound variables is, however, possible
= 47r2 h
+k
gh
T2h
Compare with
= 47r 2 (h 2 + P)
y=mx+b
to give a linear form with y as h2 , x as T 2h, the slope as

g/47r 2 and the intercept as P.
109
EXPERIMENT PLA~NING
This example has been chosen specifically since it illustrates

very clearly the superiority of linear analysis over other
methods. The graph of T vs. h which can be obtained is
shown in Fig. 5.3. It turns out that k can be obtained from
T
x
h
Fig. 5.3
The behavior of the compound pendulum.
the lengths of the intercepts AB and CD. If g is required.

it has to be obtained as a calculation from this value of k.
The advantages of the linear analysis are clear. The T
vs. h graph gives no check on the performance of the experi.
ment except in the most general terms. No reliable estimat(;
of the uncertainty of the final answer can be obtained from
this graph, but the over-all uncertainty can be obtaineu
110
EXPERIMENT PLANNING
readily from the linear graph (see page 131). The use of an
intercept at such a low angle, as illustrated in Fig. 5.3, is
very unreliable, since small changes in orientation of the
lines can make large changes in the length of the intercepted
portion. On the other hand, the slope of a linear graph can
be determined very reliably. The answer using the intercept method is determined solely by a few points in the
vicinity of the intercepts, and the value of all the other
points is not realized. When drawing a straight line, however, all the points contribute towards the choice of the line.
Lastly, the linear graph gives g and k from independent
measurements on the graph while, in the other method, any
inaccuracy in the value of k is propagated automatically
into the value of g.
(e) Logarithmic Functions

Many physical processes are of the form
where Q is a measured variable, T is the temperature and

Qo and a are constants. This can be linearized by taking
logs to the base e
log. Q = log. Qo -
-1'a
Thus, log Q plotted vs. 1/ T will yield a straight line of slope

a and intercept log. Qo. Note that if logs are taken to the
base 10, only the intercept is affected, and this is a convenience if only the slope is to be measured.
Other more complicated functions are commonly found
but, in all cases, some kind of linear analysis can be
achieved, provided one is prepared to accept compound
variables. Such a linear analy..,is should always be at-
111
EXPERIMENT PLANNING
tempted. If there is any suggestion at all that a particular

function can be expected in an experiment, try it. Any clue
is better than none and even if the suggested function is
wildly wrong, the nature of the discrepancy will almost
certainly be useful as a guide to further theoretical thought.
At the other end of the scale all expected behavior should
be checked, even when the experimenter feels sure that
everything is working as it should. If the current should be
proportional to the voltage, make sure it is. When the
experimenter is describing his experiment after it is all
over, someone is sure to ask, "Did you check?"
5.6 Experiment Analysis and Design
The task of approaching an experiment is therefore a little

more complicated than throwing together the pieces of apparatus provided, reading some scale just because it had
been left beside the apparatus, and deciding to postpone
worrying about what to do with the measurements until
after it is all over. The experimenter should remember that
an experiment is something one does because it is the only
way of obtaining the information one needs. Therefore,
the act of observation should be the consequence of planned
necessity. The point about an experiment is not so much
the observations themselves as why they were taken and
what the observer does with them.
There is no justification, therefore, in starting to take observations unless the experiment has been completely analyzed, the mode of computation selected, and a program
of measurement drawn up. The steps in this process have
already been described but we wish to summarize the
procedure.
112
EXPERIMENT PLANNING
(a) Selection oj Variables

This will usually be obvious. The experimenter will have
been given the task of measuring a specific quantity as a
function of other specified variables (e.g., triode plate current as a function of grid potential and plate potential), or
else the required answer will be known from theoretical
grounds to involve certain other measurements. Decide
which quantities are to be the principal variables and which
are to be the subsidiary variables, and try to make sure
that you do not inadvertently change two variables at once.
(b) Graphical Analysis

If there is any clue at all to graphical analysis of the problem in straight line form, use it. The process of putting
the equation in straight line form will determine the mode
of computation since the unknown will appear in a slope
or intercept. Remember that the aim of the experiment is
twofold, first, a check on the validity of the experiment and
second, a calculation of the answer. Make sure that both
requirements are met.
(c) Experimental Precision

Make trial measurements of the various quantities to consider the precision as outlined in Sec. 5.1. When a suitable
compromise has been worked out between the precision
desired and that attainable using the available apparatus,
make an estimate of the over-all precision of the experiment
which is likely to be achieved. This will serve as a very
useful guide in the conduct of the experiment.
(d) Measurement Program

Make up a complete schedule of measurement for all the
quantities shown by the experiment analysis to require
t 13
EXPERIMENT PLANNING
measurement. The mode of measurement (e.g., micrometer or meter stick) will already have been selected and so
the measurement program must allow for adequate duplication of those readings which require statistical treatment.
In general, take care to ensure that adequate attention is
given to the quantity in the experiment which has the
dominant influence on the uncertainty, but do not waste
time over others. If one quantity in an experiment cannot
be measured with an uncertainty less than 5 per cent, it is a
waste of effort to spend time in an attempt to reduce the
uncertainty of another, equally significant, quantity from
! per cent to i per cent.
The range of values to be taken will be determined by
instrumental limitations, and the number of different values
by the time available. This is the time to consider instrument and equipment ratings. The rather expensive smell of
charring insulation coming from a 1 ohm standard resistor
can easily be avoided by noting beforehand that a maximum current value of 1 amp is clearly stamped on the
casing. This matter of ratings is obvious in instruments with
scales (ammeters, pressure gauges, etc.) but is not less
important in other components like resistors where no scale
is involved. In such a case, always look carefully for the
rating value marked on the component, and adjust the
measurement program accordingly.
The measurement program should allow for all factors
which may act as a check on the progress of the experiment.
For example, all reversible quantities should be read both
ways, e.g., measure elastic deflections both when loading
and unloading, use both direct and reversed currents if
possible, take measurements both when heating and cool-
114
EXPERIMENT PLANNING
ing, read both ends of circular scales, etc. Each of these

procedures will detect some kind of systematic defect. In
elastic deflections measurements, unloading in addition to
loading will check that an elastic limit has not been exceeded. In an experiment using the magnetic field of coils,
the use of both direct and reversed currents will check the
uniformity of the magnetic fields. In temperature variation experiments, the problem is always to ensure that
temperature equilibrium has been achieved. Only a cooling as well as a heating experiment will be convincing. In
general, one can be sure that one is dealing with equilibrium conditions only if the experiment works as well backwards as forwards.
Another very useful check is the "sample in-sample out"
procedure as described on page 85. This means that the
characteristics of the system are studied both with and without the element whose properties are desired. The difference between the two observations must be due to the
object under test. This procedure thus makes the experiment act as its own control, and is especially useful when
there are a lot of perturbing influences. For example, if
we are measuring the thermionic current in a diode it is
wise to check the current when the filament is not heated.
Any current which is observed must be due to leakage, and
thus constitutes a systematic error in all the actual observations. Anyone who doubts the value of such nul-measurement checks should follow the advice of Wilson (Reference
18 in the Bibliography) and contemplate the statement,
"It has been conclusively demonstrated by hundreds of
experiments that the beating of tom-toms will restore the
sun after an eclipse."
In constructing a measurement program it is usuallv worth
115
EXPERIMENT PLANNING
considering the topic known as randomization of the readings. Consider that a series of different alloys are to be
given a plastic deformation test using a drop hammer. The
idea is to study the variation of deformation with alloying
concentration, but it may happen that, as the drop hammer
works during the course of the tests, its lubrication becomes
less sticky and the actual impulse delivered to the specimen
is increased. If this effect were not suspected, a serious
contribution to the observed variation is provided. Now,
if the alloy specimens are used in the expected sequence of
increasing hardness, this discrepancy could probably go
unnoticed since it is a smooth variation added to another
smooth variation. If, however, the samples are tested in
random order, the discrepancy will be detectable. The
error has not been removed but it has been changed from
a systematic error, difficult to detect, to an easily visible
scatter of the points. This discovery of an unsuspected
source of uncertainty can then be used to eliminate it.
Note that this method is equivalent to the method, already
mentioned, of taking readings "going both ways": they both
detect progressive systematic defects. If complete control
in the experiment is possible the first method is preferable,
but if this is not possible (as in the case of destructive testing
where only one sample of each type is available) randomization is recommended.
It will enormously simplify computation later on if the
measurement program is laid out in the form of a table,
which incorporates all the stages of the future calculation,
both of the best values and their uncertainties. Consider,
for example, the problem of a ~ass m oscillating at the end
of a spring. The period T of vibration can be written
27rv-;;;fi.
116
EXPERIMENT PLANNING
where k is a constant of the spring which it is desired to

measure.
Straight line analysis gives
T2
471"2
so that we plot T2 vertically and m horizontally to obtain I..

from the slope. The measurement program would then be.
as follows:
AUm
AUT
RUT
Tl
RUTl
AUTl
where AU and RU mean absolute uncertainty and relative

uncertainty respectively. The reasons for the format above
are as follows: m is obvious and we shall need AUm for
plotting on the graph; T is the measured dependent variable but we need P for the graph; we need AUP for
plotting and it must be calculated from the AUT (which
we assume to be known) through the medium of the relative uncertainties of T and ]'2. In this simple case it would
have been possible to calculate AUP as 2 T(AUT) instead
of using the relative uncertainties. This would have saved
a column of calculation but it is commonly very useful to
have the relative uncertainties available, especially when
dealing with compound variables (e.g., quotients). Each
case should be treated on its own merits.
In actual professional work where money, labor and equipment may all be involved to a large degree, the onus is on
the experimenter to perform his experiment so as to secure
117
EXPERIMENT PLANNING
the maximum yield of results with the resources available.

The highest efficiency can only be achieved by adequate
preparatory thought. If the analysis and measurement program construction above is completed, the taking of the
observations will be a profitable consequence of reasoned
necessity rather than optimistic guesswork, and the possibility of omitting necessary and probably irretrievable
measurements is eliminated. The personal strain of observation is much reduced and this is conducive to good
observing, so that the over-all chances of success are enormously improved . Without such planning an observational
process cannot be granted the title of an experiment and
reduces itself to an undignified scrabbling for data in the
hope that something will prove useful.
PROBLEMS
1. A scientist claims that the terminal velocity of fall of a parachutist is dependent only on the mass of the parachutist and the
acceleration due to gravity. Is it worth while setting up an
experiment to check this?
2. The range of a projectile fired with velocity v at angle a to
the horizontal may depend on its mass, the velocity, the angle
and the gravitational acceleration. Find the form of the function.
3. The pressure inside a soap bubble is known to depend on the
surface tension of the material and the radius of the bubble.
What is the nature of the dependence?
4. The period of a torsion pendulum is a function of the rigidity
constant (torque/ unit angular deflection) of the support and of
the moment of inertia of the oscillating body. What is the form
of the function?
118
EXPERIMENT PLANNING
5. The deflection of a beam of circular cross section supported

at the ends and loaded in the middle is dependent on the loading
force, the length between the supports, the radius of the beam
and Young's modulus of the material. Deduce the nature of the
dependence.
In all the following problems state the variables or combination
of variables which should be plotted to check the suggested variation and state how the unknown may be found (slope, intercept,
etc.).
6. The position of a body starting from rest and subject to a
uniform acceleration is described by
s =
at 2
sand t are measured variables. Determine a.

7. The fundamental frequency of vibration of a string is given by
1.~
n=-vT/m
2i
n, i, and T are measured variables. Determine m.
8. The velocity of outflow of an ideal fluid from a hole in the

side of a tank is given by
= V2P/p
v and P are measured variables. Determine
p.
9. A conical pendulum has a period given by
T = 2'11' Vi cos ot/g
T and ot are measured variables, i is fixed and known. Determine g.
10. The deflection of a cantilever beam follows

4Wi3
d = - -3
Yab
119
EXPERIMENT PLANNING
d, W, and 1 are measured variables, a and b are fixed and known.

Determine Y.
11. The capillary rise of a fluid in a tube is given by
h = RfT
pgR
hand R are measured variables, p and g are fixed and known.

Determine fT.
12. The gas law for an ideal gas is
pv = RT
P and
R.
T are measured variables, v is fixed and known. Determine
13. The Doppler shift of frequency for a moving source is given

by
v
/=/0-v - vo
/ and Vo are measured variables, /0 is fixed and known. Determine
v.
14. The linear expansion of a solid is described by
1 = 100
+a
. ~t)
l and ~t are measured variables, 10 is constant but unknown.

Determine a.
15. The refraction equation is

fJ.1
sin 01
01, O2 are measured variables,
fJ.1
fJ.2
sin O2
fixed and known. Determine
fJ.2.
16. The thin lens (or mirror) equation can be written
;+s'=j
s, s' are measured variables. Determine /. There are two ways of

plotting this function. Which is the better?
120
EXPERIMENT PLANNING
17. The resonant frequency of a parallel L-C circuit is given by

1
= VLC
wand C are measured variables. Determine L.

18. The force between electrostatic charges is described by
F=~2
411"for
F and r are measured variables for fixed and known ql, q2. How
do you check the inverse square law?
19. The force between currents is described by

F
J.l.o
i1i2L
211"
F, ii, i2, and r are measured variables, J.l.o and L are constant.
How do you check the form of the dependence?
20. The discharge of a capacitor is described by
Q = Qoe- t / RC
Q and t are measured variables. R is fixed and known. Determine C.
21. The impedance of a series R-C circuit is
Z = V R2
(1/W2C2)
Z and ware measured variables. Determine Rand C.

22. The relativistic variation of mass with velocity is
m = ---;.==m::::o==
V1 - (V 2/C 2)
m and v are measured variables. Determine mo and c.
23. The wavelengths of the lines in the Balmer series of tM

hydrogen spectrum are given by
~ = R (~_l)
2
>. and n are measured variables. Determine R.
121
EXPERIMENT PLANNING
24. The thermionic current emitted from a heated filament

described by
J = AT2e -'P/kT
IS
J and T are measured variables, k constant and known. Determine A and <p.
25. The specific heat S of a solid is to be measured by the method

of mixtures giving
mass of solid-measured variable
TIl,
Tl
T2
T3
mw
=
=
=
=
initial temperature of solid-fixed and known

final temperature of mixture-measured variable
initial temperature of water-fixed and known
water equivalent of calorimeter and contents-measured
variable
Find S.
23. The thermal conductivity K of a solid is measured by measuring the temperature gradient along a bar while heat is flowing,
according to the equation
Q = KA T2 - Tl
d
The quantity of heat Q is determined by measuring the temperature rise of a stream of water circulating round the cooled end.
Thus,
where m is the quantity of water flowing per second and T4 - Ts
is the temperature rise.
m, T4 - T3 and T2 - Tl are measured variables, d and A are
fixed and known. Determine K.
27. A continuous flow calorimeter experiment is to measure the

heat capacity of a liquid S. An amount of electrical energy VI is
supplied and the resulting temperature change T2 - Tl of the
122
EXPERIMENT PLANNING
fluid is measured. m is the mass of fluid flowing in unit time.

Heat is lost from the sides of the tube constituting a systematic
error. We can, however, presume that such heat loss is proportional to the difference between the mean temperature of the
fluid (Tl + T 2 )/ 2 and that of the surroundings. To. The final
heat balance equation could therefore be written
V, I, m, T 2 , and Tl are all measured variables, J and To are

constant and known. Determine S.
Experiment Evaluation
6.1 The Aims of Experiment Evaluation
Once the measurements have been completed, the aim of

the work is to determine the range of validity of the
measurements, compute the answer and estimate its uncertainty. At this stage it is well to remember that the
experimental results are precious. They may have been obtained from a massive experimental program involving
many people and a lot of money. At any level they may
well be irretrievable. It is necessary, therefore, to extract
the maximum value from the results by squeezing every
bit of information out of them. One must be objective too,
because, if the outcome proves disappointing, it is necessary
to state the result honestly and realistically, and obtain
from it the necessary guidance for future work. It is a
familiar misunderstanding amongst students that they are
in the laboratory to reproduce the known values of their
experimental qualities. If they then get 960 cm/sec 2 for g,
this is different from the "right" answer, and so they are
"wrong." The "error" is then probably blamed on the
apparatus. This deplorably ungrateful attitude is described
merely to point out its naivete and wrongness. If the
student had taken the trouble to estimate his uncertainty
123
124
EXPERIMENT EVALUATION
he would have been able to quote his answer as 960

cm/sec 2 30 cm/sec 2 and so would have realized that he
is actually right. If he is going to grumble about anything,
let it be the 30 cm/sec 2, but he should not feel guilty
about it, if the experiment with normal effort is not capable
of a precision greater than 3 per cent. Part of this mistaken
attitude arises from the fact that one normally meets the
accepted values of the constants of physics in a textbook,
and not in the laboratory. In the texts one usually finds a
casual mention of the accepted value of a particular constant. The fact may not be made clear that this number
is the outcome of many years of hard work by dedicated
and brilliant men of science. To gain such insight the future
experimenter should read the history of the measurements
of such quantities as the mechanical equivalent of heat or
the velocity oflight. (See Reference 12 in the Bibliography.)
One should not be too casual about numbers such as these,
and should not hope to reproduce them in two hours of
work in the elementary laboratory.
The point which we wish to emphasize is that the result of
the experiment should be stated honestly and objectively as
a best value, its uncertainty, and its range of validity.
Certainly the experimenter should strive earnestly to maximize the yield of the experiment by making his best value
as reliable as possible and his limits of uncertainty as narrow
as the experiment will permit, but it is in all cases important
to be realistic.
The evaluation of the experiment is in four parts, computation of the individual quantities, consideration of the
validity of the results, computation of the answer, and
125
estimation of the over-all precision of the experiment.

These will be considered in turn.
6.2 Computation of Elementary Quantities
The first step in experiment evaluation is the calculation of

the elementary quantities of which the final result is composed. For example the simple pendulum experiment will
yield a set of values of T as dependent variable and I as
independent variable. The present purpose is to compute
the values of I and T, and their uncertainties, which will
enable the subsequent graphical analysis to be accomplished.
The treatment of the observations depends on whether the
precision of observations is scale-limited or statistical.
Consider first the case of the scale-limited type. In the
example above I may have been measured with a meter
stick. Repetition may have shown that the reading, within
1 mm, did not show fluctuations. We therefore have a
single value for I quoted to the nearest millimeter. This
will consequently (see page 13) carry an automatic uncertainty of ! mm meaning that we are "almost certain"
that the real value lies within this range. The set of I values
which constitutes the range covered by the experiment is
therefore a set of lengths, quoted to the nearest millimeter
and each with an absolute uncertainty of ! mm leading
to the appropriate value for the relative uncertainty. This
will hold for all readings with scale-limited precision.
If the readings have an uncertainty which is statistical in

nature, the necessity is to compute from the duplicated
readings a best value and its uncertainty. For example, the
126
simple pendulum experiment might reveal that the uncertainty in the timing measurement is sufficiently large
that duplication of the measurement does give differing
values. The experiment will then have been arranged to
allow for sufficient duplication of T values for each value of
I to give a precision which is estimated to be acceptable.
The present task is to compute the precision which was
actually achieved.
As was pointed out on page 39, the numbers of readings in
the samples normally found in the work of the physics
laboratory are too low to permit one to draw any conclusions regarding the actual frequency distribution of the
universe from which they were drawn. Therefore, one
tacitly assumes that the set of readings is a sample from a
Gaussian universe and applies the results based on a
Gaussian distribution given in Chapter 2. For a Gaussian
distribution the best estimate of the true value X and the
universe standard deviation is the sample mean x and the
sample standard deviation s [using Equation (2.9)]. For a
discussion of what we mean by the "best" value see page 133.
If the mean of a set of quantities is required when the
members of the set are, themselves, of unequal precision,
a procedure known as "weighting" must be followed. This
will control the contribution of each quantity to the final
average in proportion to its precision. The weighting of
observations will be described in Sec. 6.6. Given a set of
repeated measurements, one therefore reduces it to the
mean and the standard deviation or the standard deviation
of the mean. At this point bear in mind the warnings about
q estimates from small samples and so make sure the computations are significant. In general, it is not worth while
127
using a statistical approach with fewer than 10 observation ~

and, for particular purposes, many more may be required.
At this stage the measurement of every quantity in the ex
periment has been reduced to a best value and its un
certainty, and the way is clear to the graphical analysis of
the experiment.
6.3 Graphs
Regardless of whether the graph is to be merely an illus~

tration of the behavior of a physical system or whether it is
to be the key to assessing the experiment and calculating
the answer, the aim is to set out the results so as to display
their characteristics as clearly as possible. This will involve
appropriate choice of scale and other physical arrangements. Ensure that the graph paper is large enough. It is
a waste of time to plot observations having a precision of
! per cent on a piece of graph paper 12 cm X 18 cm where
a typical uncertainty is perhaps 2 per cent. As we shall see
later, we want the uncertainty on the points to be clearly
visible or else valuable information is lost, and so it is necessary to make sure the graph paper is big enough. Make the
graph fill the available area. This can be done by choosing
the scales so that the general course of the graph runs at
about 45 to the axes anu by suppressing the zero if necessary. If one is plotting the resistance of a copper wire as a
function of temperature, and the values run from 57 to 62
ohms, start the resistance scale at 55 ohms and run it to 65.
If the scale is started at zero, the graph will look like a flat
roof over a sheet of empty graph paper and convey no information at all. The only exception to this suggestir.m
arises sometimes when, for pictorial reasons, one wants to
128
keep the scale of some effect clearly visible and we may want
to show it in relation to some zero. However, for computational purposes, make the graph fill the paper.
It is necessary to indicate the uncertainties of the variables
on the graph. This will enable the experimenter to judge
how far the results show straight line behavior. Unless the
range of possible variation is marked on each point, it is
impossible to say whether any particular departure from
the general trend is significant or not. Furthermore, the
range of possibility for each point will determine the overall precision of the experiment as described on page 131.
When the variables used are compound variables, the uncertainty in the compound quantity will have to be calculated for each point. The best way of marking the points
on the graph is by a very fine dot at the center of bars
indicating the possible range of variation. Make sure that
the nature of this possible range of variation is clear, outer
limits of possibility, 1s or 2s limits etc. When using compound variables it is very frequently useful to choose different symbols to distinguish the points arising from different values of a subsidiary variable. In this way a clue to
the source of any discrepancy is provided (see Fig. 7.2)
where the influence of differing pipe diameters on the discrepancy, the onset of turbulence, is thus clarified).
6.4 The Validity of the Experiment
If the experiment has been analyzed in terms of a straight

line graph, the problem is to decide to what extent the
measured points can be said to fall on a straight line. This
decision will almost certainly be complicated by the in-
129
evitable scatter of the points and it is for this reason that

the range of uncertainty of each point is plotted with it.
The best way of judging a set of plotted points is to hold the
graph paper at eye level and sight along the points. This
makes the trend much more obvious than does the direct
view. The clustering of the points about any linear region
and any departures from linearity become quite clear. In
this way a linear region of the curve can be identified and
any discrepancies isolated. If no linear region is observed,
contemplate the possibility that it lies below the lowest
point, above the highest or in between two of the measured
points, thus seeking guidance to future experimenting.
Even if theory suggests that the graph should pass through
the origin, do not take this into account when judging
linearity unless the origin is actually a measured point. If
the graph determined by the measured points fails to pass
through the origin, this is precisely one of the discrepancies
we wish to uncover.
Once a portion of the graph has been selected as a straight
line, it is useful directly to draw in what is considered to be
the best straight line. A method will be given later for the
mathematical computation of this, but it is surprising how
good visual judgment can be. A black thread stretched
along the points, or a transparent ruler is good for drawing
the best straight line. An opaque ruler is unsuitable. The
final result will be as illustrated in Fig. 5.2, and the region of
validity, in which the theory and experiment can be assumed to be compatible, has been determined. If the curve
is completely linear as far as can be detected, the theory
can be assumed to be adequate for calculation of the answer
from all points. Note that this might not be the case if the
130
experiment were to be repeated with increased precision.

In this case the validity of the experiment would have to
be re-checked.
6.5 The Calculation of the Answer and Its Uncertainty
If the preliminary analysis of the experiment has been

satisfactorily accomplished, the final evaluation of the
answer will present no problems. The answer, or some
simple function of it, will be directly obtainable from the
slope or intercept of the graph. The slope is, of course, not
the geometrical slope of the line on the graph paper, for
this depends on the scales chosen. The slope must be obtained by selecting two points on the graph, reading off the
values of the variables for these points, and calculating the
ratio of the difference in the value of the variables in the
vertical direction to that in the horizontal direction. The
dimensions of this slope will emerge automatically as the
dimensions of the ratio of the two measured variables. In
the case of the simple pendulum illustration, the dimensions
will be that of liP or Lr-2. Check that this is appropriate
to the answer required and make sure that the units are
correct. The final step will probably be a simple calculation from the slope to the required answer. For the
simple pendulum the slope was g/47r 2 The slope and 7r
are known, hence g. In all cases of numerical computation,
enormous simplification and reduced possibility of error
will result from the use of a logarithmic mode of expressing
very large or very small numbers. Thus one writes 176,000
as 1.76 X 105 and 0.00023 as 2.3 X 10-4 In extended
computation it is necessary to guard against accidental
accumulation of rounding off errors. For this reason it is
131
safer to work through a long calculation using more figures

than are indicated to be significant by the precision of the
original observations. Note that the normal slide rule does
not give more than three significant figures, and that for
only part of the scale. In extended calculations such
rounding-off errors can easily accumulate to constitute an
uncertainty greater than the experimental uncertainty.
The use of a slide rule in such a case would be a very
foolish waste of valuable information, and for all but the
very simplest of calculations the use of a set of log tables is
indicated.
If some deviation from linear behavior is found, it may
suffice merely to have its presence detected and its influence
on the answer eliminated. If, however, any speculation on
its origin is required, the deviation from linearity must be
treated as an observation of a new phenomenon. Perhaps
someone can make a lucky guess about its origin, and one
can always use the function-finding techniques to be described in Sec. 6.7.
The uncertainty in the answer is determined by the uncertainty in the slope from which it was determined. This
uncertainty arises from the finite range of possible positions
for each point given by its plotted uncertainty. This means
that the points define a band of values, delineated by AIAz
and BIB2 in Fig. 6.1, instead of a unique line. The result of
this is that a whole set of lines are contained within the
limits of possibility. The central one of these will already
have been chosen as the best line, CD, but limits exist,
shown as AIB2 and A 2B 1, within which no slope is precluded
by the observations. It must be admitted, therefore, that all
slopes within these limits are possible, and the slopes of
132
c
Fie. 6.1
The uncertainty in slope of the best straight line.
these limiting lines therefore represent limits of uncertainty

for the slope of the best line. This uncertainty for the slope
is thus obtained by calculating, in the same way as for the
best line, the slopes of the limiting lines.
The uncertainty in the answer is then calculated from the
theoretical expression for the slope, using the uncertainty
in the slope and the uncertainties of any other measured
quantities. It is necessary at this stage to recall clearly what
133
kind of uncertainty has been marked on the graph. If the

limits are outer limits of possibility (scale-limited or perhaps
2s) then the limits on the slope will similarly be of this
nature. If the points have been marked with Is limits, th<".
limiting slopes will be not quite outer limits of uncertainty,
but will probably be safer than 68 per cent confidence
since the limiting lines are drawn with a pessimistic bias.
There are really two cases to be considered. If the scatter
encountered in the actual results is within the predicted
range of uncertainty, then the use of the limiting lines gives
rise to a fairly well-defined value for the uncertainty in
slope. If, however, the scatter found is well outside the expected range of uncertainty (due to an unsuspected source
of perturbation) then there is no unique setting for lines
within which we are "almost certain" the answer lies. In
such a case and in all precise work there is no substitute for
the method of least squares to be described in the next
section. In general, remember that experimental results are
very precious quantities. If the experimenter is to justify his
existence he must extract the maximum amount of information from them, either for his own answer or as a guide
to future work.
6.6 The Principle of Least Squares
What is meant by the "best" value from a measurement

process? It is merely begging the question to answer, "The
one that stands the highest chance of coming close to the
true value." Indeed there is no unique answer to our question but we can rephrase it as follows: If we have a set of
observations XI, X2, Xa, and we try to find best estimates
134
of the values X and (T appropriate to the universe of read

ings we are, in effect taking a whole set of Gaussian curves
with differing X and (T and trying them on our Xl, X2, Xa,
Xn for fit. If we can find some process or criterion for
our choice of the "best" X and (T whereby, on repeated
sampling, i.e., by repeatedly taking batches Xl Xn of the
readings, the values of our estimates of X and (T cluster most
closely around some central value, and in an unbiassed
fashion, then we can use this process to define a "best"
value.
A similar consideration holds if we have a set of observations
which we have plotted in the hope of obtaining a straight
line. It will essentially show scatter as shown in Fig. 6.2.
The question really is-does this set of observations provide
support for any particular hypothetical variation and, if so,
how much support? It is possible to draw a straight line
through these observations but it is also possible, as shown
in Fig. 6.2, to draw any number of other curves. What is
to be the choice? The mathematical answer is, as stated
above for repeated readings of a single quantity, that curve
whose parameters show least variance on repeated sampling. By that we mean the following: If we consider the
unperturbed (i.e., the ideal) variation to have the form
y = ax
we could use some criterion (as yet undefined) to compute

a best estimate of a. If we take another sample of observations we shall obtain another best estimate. Repeated
sampling gives a whole series of best estimates of a with a
certain spread.
Suppose now we consider the unperturbed variation to be
of the form
135
y = ax
+ bx
We could use the same criterion (still undefined) to obtain,

on repeated sampling, a whole series of best estimates of a
and b. The question now is-which postulated ideal variation gave the least spread in estimates of a? Whichever
postulate it was, we can accept it as the better guess. However, who is to say that there is not a yet better guess, as yet
unguessed? What then, is the justification for accepting any
postulated ideal variation to describe a set of observations?
The answer is almost invariably that someone has made a
prior suggestion regarding the form of the function. If
someone has taken a lot of trouble to make up a theory,
FIC. 6.2
The possibilities for a functional relation in observations showing statisti.

cal uncertainty.
136
the cause of science is best advanced by interpreting the

results in terms of it. \Vhether it fits well or not the comparison will be the most useful way of progressing from the
present state of knowledge to an improved state.
Granted, then, that a decision is made by the experimenter
to interpret his observations in terms of some particular
function, we must ask-what is the best way of determining
the parameters of the function? We shall assume that the
experimenter has been able to reduce his postulated function to straight line form
.Y
mx
+b
and his problem will be solved if he can find some criterion

for choosing the best estimate of the parameters m and b so
that, on repeated sampling, he would find the least variance
of his estimates.
This is really all one can say about the nature of a best
value and, because of the vagueness, there has been much
discussion over the choice for the criteria mentioned above.
One criterion stands out through its simplicity and widespread use. This is the principle of least squares. It must
be remembered however, that it is just a way of satisfying
the requirements listed above for the choice of a best value,
and is not a magical way of attaining absolute truth. It
must be clearly borne in mind that, when a straight line fit
is being adjusted by least squares, the answer for a slope or
intercept has no greater validity than is possessed by the
original decision to interpret the results in terms of a
straight line.
The writer makes no apology for this wordy preamble to a
very simple and commonly used technique. It is quite easy
137
(and widely practiced) to use the least squares method with~

out any idea of the significance of the procedure. It is the
conviction of the writer that users should be aware not only
of its powers but also of its limitations.
For the present purpose we shall accept the principle as
an axiom. It is stated as follows:
"The most probable value of an observed quantity is
such that the sum of the squares of the deviations of the
observations from this value is a minimum."
Let us consider the applications of the principle to two cases.
(a) Repeated Observations oj the Same Quantity

Consider n observations Xl, X2, X3, and call the most
probable value X. Application of the least squares principle
gives
2; (Xi -
X)2 = minimum
Let x be the mean of the Xi. Then

2; (Xi -
X)2 =
2;
= 2; [(Xi -
+ 2(Xi or since
2;
+ (x - X)P
X)2 + (x - Xp
[(Xi - x)
x)(x - X)]
(Xi - x) = 0,
2; (Xi - X)2 = 2; (Xi - x)2
+ (x -
X)2
This last expression is clearly a minumum when x = X,

thus confirming that the use of the mean as the most probable value of a sample is consistent with the principle of
least squares.
If the individual Xi do not all have the same precision, but
have each a standard deviation, it can be shown that the

principle of least squares gives for the best value
138
~ x;fs1.
~
(6.1)
l/sl
This equation clearly biases the value X towards those Xi

which have the smallest Si, thus favoring the more precise
measurements. This process is called "weighting" and X
is a "weighted mean." It can be shown that the standard
deviation of such a weighted mean is given by
X)2j (n __ 1) ~ .(
~~ (Xi -
s;
s;
(6.2)
(b) Best Fit to Straight Lines

Suppose that an experiment has yielded a set of n values ofy
as a function of x. Assume that, for the sake of simplicity,
we can regard the values of X as exact, and that the values
ofy are subject to uncertainty.
When plotted, these readings appear as in Fig. 6.3. The
task is to determine the constants of the straight line AB
which best fits the observations. Let the equation of this
b so that our objective is the best values
line be y = mx
of m and b.
The scatter of the points about the line is characterized by

differences of the form
0Yi = Yi - (mxi
+ b)
The criterion of least squares gives the constants m and b

from the condition
~
Write
+ b)J2 = minimum
(mxi + b)J2 = M
[Yi - (mxi
~
[Yi -
Then the condition for the minimum is
aM
am
= 0 and
aM
ab
= 0
139
y
FIg. 6.3
The least squares fit to a straight line.
Elementary algebra then gives for the constants m
m=
d110
n ~ (X iYi) - ~ Xi ~ Yi
n~Xi2
(~Xi)2
(6.3)
b = ~xl ~ Yi"":- ~ Xi ~ (XQi2

n~xl- (~Xi)2
This then defines the best straight line by a rather more

acceptable criterion than visual judgment. Since it is statistically significant we can assign standard deviations to
the estimates of m and b, so that their reliability has a
definite meaning. They are expressed in terms of a quantity SIl which is the standard deviation of the
values
about the best line. This is given by
oy
(6.4)
140
and
Sm
and
Sb
are given by
S
Sb
I
n
Y'Jn~xi2 - (~Xi)2
I
Sy
Xi
(6.5)
'J n ~ Xi 2 - (~Xi)2
For a more complete, mathematically detailed treatment

of the method of least squares the reader is referred to
Appendix 2.
Functions other than straight lines can be fitted to observations by the method of least squares but Equations (6.3),
(6.4), and (6.5) are, of course, not applicable. Recourse to
first principles as described in Appendix 2 is necessary and
the methods used will be found in Reference 2 in the
Bibliography. If the uncertainties of the points are known
to be different and values for the uncertainties are available, a weighting procedure can be followed to bias the
m and b values in favor of the more precise points. A
description of this weighted least squares method will be
found in Reference 18 of the Bibliography.
1.7 Function Finding
It has been said before that, if the experimenter has any

clue at all to the expected behavior of a system, he should
use it in analyzing the results. Failure to do so would mean
that there is some information in the results which is not
being extracted. It may happen, however, that there is
absolutely no information available to aid in the interpretation of a set of results. In such a case algebraic analysis of
the results has a twofold aim. First, the establishment of
some function which fits the observations adequately will
141
be very useful to others who wish to use the results empirically. Although values can be read directly off a graph
and this is very convenient, the representation of the results
by an equation is much shorter and neater. Second, the
discovery of some function which describes the results may
(but only "may") be of assistance in the establishment of a
model for the system. In any case it is worth trying.
The information available to the experimenter is a set of
readings y = J(x) which, when plotted simply in terms of
the measured variables y vs. x (for lack of any better
proposal), give a smooth curve of no readily identifiable
shape. It is desired to obtain the form of the functionJ(x)
and this is easy in two particular cases, fortunately both of
reasonably common occurrence in physical systems.
(a) Power Law
Suppose
where a is a constant.
Then
logy = a log x
Consequently a plot of log y vs. log x is a !>traight line of
slope a. If a power law is suspected, therefore, plot the
variables logarithmically, either on log - log graph paper,
or on linear graph paper by obtaining the logs from tables_
If a straight line is obtained, a power law dependence for
the results can be claimed, with an index equal to the slope.
(b) Exponential Law
Many physical phenomena follow an exponential relation,
= ae bx
where a and b are constants. In this case

logy = log a
+ bx
142
and a plot of log y vs. x gives a straight line. If, therefore,

this type of variation is suspected, either plot logy vs. x on
linear graph paper or y vs. x on semi-log paper. Obtain
the constant b from the slope and a from the intercept. It
is necessary to be careful about intercept measurements, for
if logs to the base 10 instead of to the base e are used, the
origin is shifted.
If neither of these processes works, it is possible to express

any set of readings as a power series of one variable as a
function of the other using a difference table (see Reference
17 in the Bibliography). This polynomial representation
may be useful for empirical work but it is still difficult to
guess from it at the basic function which would be revealed
by a theoretical treatment.
'.8 Over-aU Precision of the Experiment
At the beginning of the experiment, the experimenter

looked forward with a guess at the limits of uncertainty
likely to be found. This was only an estimate for his own
information. At the end of the experiment he should look
back and, by a critical assessment of the results, evaluate
the precision actually achieved. To be useful at all to other
people, this figure must be realistic and honest, and it is
worthwhile taking pains to make it reliable. On the other
hand, there is no intention of persuading the student to
immerse himself in a set of formal calculations, blindly following the mathematical requirements and finally quoting
an uncertainty to three supposedly significant figures. Due
to obscurity of the main issues it is quite likely to be significant to none. It is far better to make a realistic estimate
143
guided by reason, than it is to lose sight of the point by

overemphasis on the formalism.
The kind of uncertainty to be quoted does not matter
provided the quantity used is made clear. Thus the experimenter
may quote outer limits for the range of his answer (we
presume he means something like 95-99 per cent confidence), or he can quote the standard deviation of his set of
results or the standard deviation of the mean. It does not
matter, but only so long as it is stated quite clearly what
kind of quantity is being quoted. It is essential, of course, to
ensure that all known sources of uncertainty are included
in this evaluation of the precision. There is no point in
claiming that potentials read off a 1 m slide wire potentiometer are precise to ! per cent just because the balance point
could be read to within a millimeter. It could well be that
lack of uniformity in the slide wire introduces uncertainties
much greater than! per cent. Similarly a thermometer
could perhaps be read to 10 deg but errors in calibration
could be even! deg. The lower uncertainty can, of course,
be claimed as the uncertainty of reading the scale, but we
are concerned at present with the over-all precision of the
experimen t.
I t is assumed, then, that the readings at this stage have been
pruned of all systematic deviations, and the remainder of
the uncertainty calculation depends on circumstances.
(a) Result Is the Mean oj a Set oj Readings

The best quantity to quote is the standard deviation of the
mean, since this has an immediate numerical significance.
Sometimes the standard deviation itself is quoted. In all
144
cases it is essential to quote the number of readings so that

the reliability of the (J' estimate can be judged.
(b) Result Is the Consequence oj a Single Calculation
In the undesirable event that no graphical analysis has

been possible and the result is obtained algebraically from
a number of measured quantities, use the methods of Chapter 3 to calculate outer limits for the uncertainty, or the
standard deviation.
(c) Result Is Obtained Graphically
If the straight line has been established by a least squares
method, the uncertainties in the constants m and bare
obtained directly. Note that these uncertainties have the
advantage that they have been obtained from the actual
scatter of the points, regardless of their estimated uncertainties. (This does not mean, however, that if one intends to
make a least squares fit to a straight line, one should not
bother to plot in the uncertainties, or even to draw a graph
at all. The graph, with the uncertainties on the points, is
still needed to judge the range of validity of the experiment,
and thus to detect possible systematic errors. A least squares
fit should be attempted only after one has, by visual inspection of the graph, selected the range which is to be considered linear). If the straight line has been drawn by eye,
the lines at the limits of possibility will give the possible
range of slope and intercept. This uncertainty in slope will
then probably have to be combined with the uncertainties
of some other quantities before the final uncertainty of the
answer can be stated.
In all cases, it probably does not matter too much what kind
145
of uncertainty is quoted, so long as one is quoted, and the

nature of the quoted value is made clear.
When working through lengthy uncertainty calculations the
arithmetic may be simplified by dropping insignificant uncertainty contributions. There is no point in adding a 0.01
per cent contribution to one of 5 per cent. In the final
statement of uncertainty it is not commonly valid to quote
uncertainties to any more than two significant figures. Only
work of real statistical significance would justify more.
Once the over-all uncertainty of the final answer has been
obtained, the question of the number of significant figures
to be retained in the answer can be considered. There is
no unique answer to this question but, in general, one
should not keep figures after the first uncertain figure. For
example, 5.4387 0.1 is 5.4 because, if the 4 is uncertain,
the 387 are much more so. However, if the uncertainty is
known more precisely, it might be justifiable to keep one
more figure. Thus, if the uncertainty above were known
to be 0.15, it might be possible to quote 5.44 0.15. When
quoting observations with scale-limited uncertainty one
cannot use the suggestion of keeping one uncertain figure.
A meter stick reading is quoted as 43.2 0.05, not as
43 .20 0.05. This is because the statement of the measurement is that the value lies somewhere between 43.15 and
43.25, and 43.2 is just the name which is given to this
interval.
If a measurement is quoted with a percentage precision, the
number of significant figures is automatically implied. For
example, a measurement might be quoted as 527.64182 1
per cent. This means that the absolute uncertainty is
146
around 5.276. However the precision itself is quoted to only

one significant figure (1 per cent, not 1.00 per cent) so that
we are not justified in using more than one significant
figure in the absolute uncertainty. We shall call the absolute uncertainty 5 and this implies that, if the 7 in the
original number is uncertain by 5, the .64182 has no meaning. The measurement could then be quoted as 528 5.
If a set of readings has yielded a mean as the answer, the
number of significant figures in the mean will be governed
by the standard deviation of the mean, and the number of
significant figures in the standard deviation will be governed, in turn, by the standard deviation of the standard
deviation.
The experimenter should make sure that he quotes his answer and his uncertainty in such a way that they are consistent, i.e., neither as 16.2485 0.5 nor as 4.3 0.0002.
PROBLEMS
1. A value is quoted as 14.248 0.1. State it with the appropriate number of significant figures.
2. If the value is quoted as 14.248 0.15, how is it quoted with
the appropriate number of significant figures?
3. A value is quoted as 6.74914 ! per cent. State it with an
absolute uncertainty, both with the appropriate number of
significant figures.
4. State the results of the calculations in the problem at the end
of Chapter 2 as a best value and an uncertainty, both with the
appropriate number of significant figures.
5. An experiment was done to measure the impedance of a series
147
R-L circuit. The impedance Z is given as a function of the resistance R, the frequency of the source f and the inductance L by
Z2
J(l
+ 4~fL2
The experiment was done by measuring Z as a function off with

the intention of plotting Z2 vertically and f horizontally to obtain
L from the slope and R from the intercept. The observations
obtained are given in the table.
f
c. sec-1
ohms
123 4
158
194
200
229
245
269
292
296
7.4 0.2
8.4
9.1
9.6
10.3
10.5
11.4
11.9
12.2
AUZ2
AUr
r
f(AUf)
= 2f(AUf)
Z2 Z.(AUZ)
= 2Z(AUZ)
The uncertainties given in the first line refer to all the readings
in each column.
(a) Plot these readings in the appropriate fashion and mark
the uncertainties on the points. Suggested table headings
to expedite the calculations are given above.
(b) Check to see if the observations can be interpreted in
terms of a straight line for any part of the range or all of
it.
(c) Obtain the slope of the best line.
(d) Calculate the best value for L.
(e) Obtain the slopes of the lines at the outer limits of possibility and so state the range of uncertainty for the slope.
(f) Calculate the absolute uncertainty in the measurement
of L.
148
(g) Obtain the best value of R from the intercept.

(h) Obtain the uncertainty for the R value.
(i) State the complete result for the experiment with the
appropriate number of significant figures in each quantity.
6. Ten different observers report on the intensity of a lamp
measured with a comparison photometer. Their results (in arbitrary units) are as follows:
Observer
Mean
Standard deviation of mean
2
3
4
5
6
7
8
9
10
17.3
18.4
17. t
16.6
19.1
17.4
18.5
14.3
16.8
17.4
2.1
1.9
2.5
2.8
3.2
1.2
1.8
4.5
2.3
1.6
What is the result of the experiment and what is its standard

deviation?
7. An experiment has been carried out to investigate the temperature dependence of the resistance of a copper wire. The ideal
variation can be represented by
R
= Ro(1
+ aT)
where R is the resistance at temperature Toe, R is the resistance

at ooe and a is the temperature coefficient of resistance. The
following observations of Rand T were obtained:
149
Calculation of best values of

Ro and ex
T"C
Rohms
x2
xy
T2
TR
Calculation of standard deviations

using calculated best values of
Ro and ex
RoexT
Calculated
ideal
value
of R
(= Ro
RoexT)
M
(obs. R
- ideal R)
(oR)2
+
10
20
30
40
50
60
70
80
12.3
12.9
13.6
13.8
14.5
15.1
15.2
15.9
--
~x
and
2:y
2: (t"y) 2: (x2 )
(~ X)2
2: (oy)!
and so,
s~
(a) Using the method of least squares (suggested headings

under which the calculations for this can be carried out
are given in the first part of the table) obtain the best
value for the slope and for the intercept.
(b) Hence obtain the best value for ex.
(c) Evaluate the standard deviation for the slope and for the
intercept (suggested headings for this part of the calculation are given in the second part of the table).
150
(d) Hence evaluate the standard deviation of a.

(e) State the final result of the experiment with the appropriate number of significant figures.
8. The following measurements were made in the investigation
of new phenomena. In each case identify the function and evaluate its constants.
(a)
v
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.61
0.75
0.91
1.11
1.36
1.66
2.03
2.48
3.03
2
4
6
8
10
12
14
16
18
3.2
16.7
44.2
88.2
150.7
233.5
337.9
464.5
618.0
(b)
151
(C)
100
150
200
250
300
350
400
450
500
0.161
0.546
0.995
1.438
1.829
2.191
2.500
2.755
2.981
(This last one is a little less obvious)
The Scientifie Report.
7.1 Aims in Report Writing
The value of skill in descriptive writing to any scientist or

engineer can hardly be overestimated. The written report,
whether privately circulated in a small organization, or published in an international journal, is the standard mode of
communication in all forms of technological or scientific
work. It is very frequently the only contact between the
experimenter and those potentially interested in his work
and so it must be able to stand by itself as the record of the
experiment and its interpretation. The aim in the report,
therefore, has to be to convey to the reader as clearly,
concisely and convincingly as possible the description of the
experimenter's work. The development of a fluent, elegant
style of writing should not be regarded as a waste of time
even for the most technical-minded of engineers. Irritation
on the part of the reader of a report which is caused by
obscurity, repetition or failure to create a logical sequence
can depreciate even the best experimental work. So often,
on reading a poor and obscurely written report, the reader
finally puzzles out what the writer meant to say and says
to himself in exasperation, "Why didn't he just say so?"
152
153
THE SCIENTIFIC REPORT
Try to avoid inducing this feeling in the readers of your

reports.
The circumstances of report writing can vary so much that
it is, unfortunately, impossible to lay down specific rules
for good reports. The range of topic and purpose is far too
wide, and it is the student's responsibility to acquire his own
style for his own purpose. This can happen only through
practice and correction, and such practice is the main purpose in writing laboratory reports. They should therefore
be treated, not so much as mere assignments to be completed, as opportunities for the practice of good descriptive
wntmg. This practice is necessary since the technique of
clear explanation (either written or spoken) does not necessarily come naturally. For this purpose there is no substitute for personal application, but one very useful method
of absorbing some of the atmosphere of the literature in a
particular field is to look at the pages of a reputable journal.
This will show the student, at any stage, the standard at
which he is aiming.
7.2 General Principles
This section contains some hints which may help the student
to become fluent in scientific writing. It is intended to be
very general but is illustrated specifically with reference to
the physics laboratory.
The most important single point is clarity. The aim of the
report is to convey information and it is a failure unless it
does so. At all stages of the writing, pause to ask,
"What is the point I wish to make here?" Once this is
154
settled, put it down in a simple and straightforward fashion

so that it emerges clearly. Obscurity can arise as easily from
too many words as from too few, so examine every statement
for information density. In almost all cases the most concise
way of saying something is the best, and one should discard
everything which does not tell the reader something significant. The report should be a logical sequence, one section
leading into another without abrupt changes in the train of
thought. The reader should be led gently through the argument by a series of smooth, logical steps, all the way through
the report from the first statement of the topic of the work
to the final conclusions. One very useful method of keeping
the main line of argument clear is the use of appendices.
Any detail of theory, method or discussion or any lengthy
table ofresults which, although judged worthy of inclusion,
would obscure the main line of the argument should be
relegated to an appendix. To do this, divide material relating to the experiment into three categories; descriptive,
reference, and personal. Put the first in the text of the report, put the second in the appendices and do not put the
third in at all. People are not interested in the trials and
tribulations of your experiment unless you have some really
specific point which will help them to avoid difficulties.
The place for your record of trials and tribulations is your
own note book. Write the report in the past tense and
avoid first person constructions and colloquial expressions.
Pay attention to the format of the report, taking care to
give the various sections a clear heading which stands out.
The reader should not have to hunt around in a featureless
maze of text and numbers. In a long, complicated report
a table of contents at the beginning is very helpful.
155
The general rule, one which will be of great assistance in

writing good reports is-think of your reader all the time.
Ask yourself what he wants to know, and tell it to him.
Your reader constitutes the purpose in writing the report
and he is one of the most important features of it.
It is not intended to suggest a rigid format for the report
since circumstances vary so much, but some features of the
report are of fairly general importance.
7.3 Introductory Material
The writer of a report must start with the assumption that

the reader, although intelligent and well-informed, knows
nothing about the work described in the report. The function of the introductory part of the report is, therefore, to
bring the reader to a point at which a detailed description
of the procedure and results makes sense.
The normal components of introductory material are:
(a) Title
(b) Topic
(c) Background Material
(d) Application oj the Background Material to the Particular
Experiment
(e) Specific Statement oj Intention
These headings, (a) to (e), are not intended to be taken as

suggestions for actual headings in a report. They are merely
the items which constitute the analysis of a report and suggestions regarding headings for use in actual reports will be
given later. The items above will now be considered in
turn.
156
(a) Title
This should be short and to the point. Its purpose is only
to identify the work, not describe it.
(b) Topic
The first thing the reader wants to know is the subject to be
considered in the report. He has already had a hint of this
in the title, but he is not ready yet for detail. Think of the
most general statement about the experiment you can make
and start with that. For example, an experiment on el m
for electrons could start with the statement, "The ratio of
charge to mass for electrons can be obtained from the trajectory of an electron beam in a combination of electric and
magnetic fields." This introduces the reader to the topic of
the work generally, and this kind of topic statement will be
found at the beginning of almost every good report. Make
sure this topic statement is clear, because if the reader misses
the point right at the beginning, the rest of the report may
be of very little value.
(c) Background Material

The reader must be brought up-to-date in the field of the
experiment, and must be provided with enough information
to make the rest of the report intelligible. This background
material may be either historical or theoretical. It is always
a difficult point to know how much of this to include, since
one does not know how much the reader already knows. A
good guide is to put in the minimum that will render the
rest of the report understandable. Unless you are writing
a review paper for the purpose of expounding the work of
other people, quote the results of established theory only,
giving a reference where the whole treatment may be found.
157
Make sure, however, that any limitations on the result are

stated if they are important in the experiment. Thus one
would write, "It can be shown (Reference 1) that the
period T of a simple pendulum of length I can, if the
amplitude is sufficiently small, be written etc., etc." If it is
considered that some detailed piece of theory is necessary,
consider the possibility of relegating it to an appendix to
keep the logical thread of the development clear.
(d) Application to the Particular Experiment
The reader must be shown how the background material

or theory is relevant to the present work. In many cases
the standard theory must be revised or extended. Since
such application is not standard and is probably the work
of the experimenter himself, it should be given at adequate
length to enable its validity to be judged. This section will
probably contain the analysis of the experiment in linear
form, and will thus reveal how the validity of the experiment is to be checked and the result computed. This section
may also contain a reference to any special measuring
technique which may be necessary.
(e) Specific Statement of Intention
It is a great help to the reader if the position is summarized
at this stage to say fairly precisely what is intended in the
experiment, e.g., "Thus, if the variation of x and y for
various values of z is measured, a straight line graph of XZ2
vs. y should be obtained and this will give a from the slope
and b from the intercept on the horizontal axis."
The description above of introduction material may be
subdivided to suit the particular circumstances. In an elementary experiment it need occupy no more than a single
158
paragraph and be labelled simply, "Introduction." In a

more complicated experiment, more extended discussion of
some particular topic may be necessary, thus justifying
other headings such as "Theory" or "Method." The actual
mode of subdivision is probably unimportant (provided it
is clear).
It will be noticed that an important concept, the aim of the

experiment, has not yet been discussed. This is again because it is undesirable to warp the report to fit too rigid a
format, and the place where the aim of the experiment
may be stated varies very much according to circumstances.
If the subject of the experiment is well known and well established, the aim makes a convenient topic statement, e.g.,
"The aim of the experiment was to measure n by the
method of Jones." On the other hand, if a relatively
obscure quantity is being measured, it might not be possible to state the aim until after a considerable amount of
definition and theory had been given. It probably does not
matter too much, provided only that the aim is clearly
stated somewhere and probably as early as is convenient.
At this stage the reader has been told the topic of the experiment, given any necessary background, shown how this is
applied to the particular experiment and finally how the
whole scheme is used to obtain the required answer. This
leaves the reader ready to be told how the experiment was
actually done.
7.4
Procedure
A good plan to follow in the "Procedure" section is:
159
(a) A Statement of the General Outline of the Performance

of the Experiment:
This should be stated as concisely as possible so that the
reader can see immediately the general course of the experiment. If the experiment consisted of the measurement
of the variation of y with x over the range a to b, say just that.
(b) Detailed Description of SPecific Measurement Methods,'

It is essential to ensure that the reader is told the method
of measurement of every factor contributing to the final
result, but this description should not degenerate into a list
of experimental trivia. Insert only enough detail so that
the reader can judge validity of the procedure. An exception to this advice is encountered in papers describing some
new experimental procedure which is described so that
other people can duplicate the method. Here, clearly, the
description must be complete, but this is a special case
and commonly a much more general description of procedure can be used. Well known, standard methods of
measurement need be identified by name only, and only
in special cases should they be described in detail.
The description of measurement method will, of course,
involve the description of any necessary apparatus but
make sure that the principle of the method is first described
clearly and prominently. Do not ask your reader to follow
through a long, detailed, step-by-step description of some
complicated piece of apparatus only to have him ask himself at the end, "Yes, but what is it measuring and how?"
(c) Any Additional Details of Care or Precaution Which
Proved Necessary.
160
There should be a diagram, with a figure number, of even

the simplest apparatus, but it is rarely worth while making
diagrams into works of art. They exist merely to aid the
reader and should be as simple as possible. They should,
however, be neat: it is not too much to expect the use of a
ruler. Notations on diagrams form a very neat and useful
way of providing reference-type information without cluttering the actual text. This is particularly true of circuit
diagrams, on which the values of resistors, capacitors and
other components, the ranges of meters, ratings of power
supplies, etc. can be marked. If there is no room on the
diagram itself, mark the components with a simple symbol,
R1 , C2 , etc., and make a little table beside the diagram with
the actual values on it, as suggested on Fig. 7.1.
If the main topic of the experiment is a study of the properties of a sample of material, the procedure section must
c
G
(a) Flow apparatus
F - Constant flow apparatus

T 1 , T2 - Thermometers 0 - 100C
G - Glass tube
C - Rubber hose clamp
Heater wire
AI>A 2 - Electrical terminals leading to R
R-
Flg.7.1
A completed apparatus diagram.
(b) Power control circuit
s-
100 v DC supply
Ru - Rheostat 80 ohms
V - Voltmeter 0-30 v DC
A - Ammeter 0-10 amp DC
161
contain an account of the nature of the sample, its origin,

physical form, mode of preparation, etc.
If for any purpose, any particular section has to be treated
in detail, e.g., method of making a particular measurement,
preparation of samples, etc., make it into a section on its
own, with its own heading. If this obscures the main line
of the argument, put the material in the appendix.
7.5
Results
The main point here, again, is clarity. From the preceding

sections the reader has been led to expect a set of values for
the variation of y with x and measurements of a, b, and c.
When he turns to the results page he wants to see just that.
He should not have to work his way through a page,
solidly filled with numbers and devoid of description and
notations. If the principal result of the experiment is a set
of values of y and x, put them in a big clear table, standing
out by themselves and headed clearly "Variation of y with
x." All other tables should also be clearly labelled. Give
each table a number so that it is easier to refer to it in the
text. Any subsidiary measurements should similarly stand
out by themselves, clearly labelled and prominently displayed. Note that there is a difference between the table of
results constructed by the experimenter in his laboratory
note book and that required for the report. The former
was laid out to expedite calculation and included a lot of
information with which the reader is not concerned and
whose presence would merely irritate him. In a column of
results the reader wants results and only that. All intermediate arithmetic should be eliminated and values should
162
have their uncertainties marked right beside them and not

in a separate column.
Every measured value should have its uncertainty indicated. If the entries in a whole column of readings have the
same uncertainty, this can be quoted just once at the top.
Otherwise, every reading should be accompanied by an
uncertainty, and the nature of the uncertainty (outside
limits, standard deviation, etc.) clearly stated. Original
readings should always be included, or else the mode of
computation should be made so clear that the original
readings can be recovered. If this is going to make the
Results section too long and bulky, the lengthy portions
can be reserved for an appendix.
Do not clutter up the Results section with calculations.
Keep all the work on derived quantities (e.g., slopes) separate. Do not include trivial algebra and ensure only that
sufficient detail is displayed to make the mode of calculation
obvious. Any calculations which are needed but are
lengthy should, again, be put in an appendix. At the end
of the process make the final answer and its uncertainty
really prominent.
7.6 Graphs
Graphs should be as clear and informative as possible. It

is quite likely that the reader's judgment of the work will
be based more on the graphical presentation of the results
than anything else. The graph that was drawn during the
evaluation of the experiment was for purposes of computation, and the emphasis was on fineness of drawing and
choice of format to permit the maximum extraction of infor-
163
mation from the results. The graph in the report is a copy

of this one, but its role is more that of illustration, and
attention should therefore be directed to clarity and completeness, as suggested in Fig. 7.2.
Choose suitable scales so that the results fill the available
space to a reasonable extent. Label the axes clearly and
include the units of measurement. Put a clear heading on
10
Dependence of water flow

rate on pressure head and
channel diameter
T = 20C
8
--,
(.)
.'"
CI>
Qo
~
a= 1cm
u:
4
+ ++
a=.05cm
Pa 4 (dynes cm 2 )
FIC. 7.2
A completed graph.
164
the graph so that there is no doubt about its identification.

Indicate the range of uncertainty of each point. When a
slope has been obtained from a graph, do not indicate this
by a drawn-in triangle because this can obscure the position
of the point from which the slope is to be calculated. Instead, the chosen point should be indicated in some way
without interfering with the line, and the coordinates of the
point written nearby. No other writing on the graph should
be permitted since a tendency to fill up blank spaces with
calculations, etc. merely obscures the message of the graph.
Graphs should be drawn in ink. All illustrations in a report,
apparatus diagrams, circuits, graphs, etc. should be given
a figure number to facilitate reference in the text.
7.7
Discussion
A suggested list of topics is as follows:
(a) A Summary of the Experiment and Conclusion:

This should be inserted even if the experiment is a simple
measurement: "n has been measured by the method of
Jones and found to be 52.3 0.5"
(b) Comparison with Expt!ctation:

The experiment was very likely to have started with a
piece of theory. As was stated on page 102, this was a
prediction, on the validity of which the successful measurement of the unknown was based. In the successful performance of the experiment this validity was investigated.
The extent to which it was confirmed should now be stated
as a reminder to the reader and care should be taken to
point out how any possible systematic error from any loss
of validity (such as the onset of turbulence in fluid flow)
165
has not been allowed to influence the answer. Consider,

too, the correspondence between the uncertainty estimate
at the start of the experiment and that actually found. If
the latter is greater than the former, some justification
should be sought.
Try to make the comments in the discussion refer to the
readings themselves because they are the contribution to
the cause that the experimenter is making. In the region
of any discrepancies from expected behavior, however,
speculation is justifiable. A guess at the origin of any
such discrepancy, either in behavior or uncertainty, may
be of great assistance to another experimenter who wishes
to make similar measurements.
In the discussion the experimenter himself should be his
most severe critic. If he does not point out the limitations
on his work himself, someone else is bound to do it for
him, and that is much more distressing.
(c) Additional Comments:
Any other noteworthy features of the experiment can be
mentioned if they shed light on the reliability of the experiment, or are potentially useful in future work. Do not,
however, list in self-defense a series of "sources of error,"
since anything which is identifiable to this degree should
not have been allowed to influence the result, or else
should have already been included in the quoted range of
uncertainty.
At this stage it may be legitimate to quote other measurements of the same quantity, for the sake of interest in the
comparison. I t must be stressed that this is done, not in
the spirit of quoting the "right" answer so that the "error"
166
THE SCffiNTIFIC REPORT
of the experiment can be calculated, but just out of interest. The reader may be reminded that the result of any
experiment is a best value and uncertainty, and the only
superiority of more professional measurements lies in the
reduced possibility of systematic errors (through multiplicity of measurement techniques) and reduction of the
range of uncertainty. Even so, the measurements are far
from final. It is best not to think of the "right" answer
but to have confidence in one's own ability to check the
validity of an experiment and so to obtain a best value
with a realistic uncertainty.
Laboratory Practice
Experimental skills can be acquired only through practice.

The following remarks are therefore limited to a few hints
of fairly general usefulness.
8.1 The Laboratory Note Book
It is mandatory in all research and development experimental work to keep a running record of the work. No
matter how confident one feels at the time, human memory
is too unreliable and every piece of information or bright
idea should be put in writing in some permanent form.
This is true not only for lists of observations, for which it is
obvious, but to all other aspects of description of the work.
It is best to write down almost everything, since experimental work is, essentially, exploration and one never
knows what information may subsequently be needed.
Rutherford discovered radon by observing that the results
of the experiment depended on whether the laboratory
door was open or shut! It is a good thing for the student to
form this habit of making a permanent record of observations and thoughts as early as possible. Use the note
167
l68
LABORATORY PRACTICE
book (preferably hard-backed) for all work and thoughts

relevant to the experiment. Layout in it the preliminary
analysis of the experiment, neatly and formally, before the
actual measurements are undertaken. At this stage of preliminary work the note book is very useful, since the average
person can think much more effectively on paper, writing
down any given information, equations, circuits, etc., and
working with them systematically, than he can by staring
vacantly into space hoping for inspiration for the over-all
solution of the experimental problem. Keep the note book
neat and, when recording actual observations, enter them
in carefully drawn tables marked clearly for identification.
All calculations, too, should be kept neat and orderly. Do
not scribble on loose bits of paper: if anything is worth
writing down, it is worth preserving. It is, of course, necessary to prevent the note book degenerating into a meaningless mass of symbols, so use clear section headings as much
as possible. Note that the type of information needed in
such a personal log is different from that required in a report intended for other people. For one's own purposes,
one requires specific information, serial numbers of meters,
lengths of heater wire, filament currents, etc., which enables one e~ther to check calibrations, should the necessity
arise, or else duplicate some process which has worked
successfully in the past. For this purpose records of failures
are just as important as those of successful operations so as
to avoid duplicating mistakes. One point to consider here
is that circumstances can arise in which one's experimental
work has to be continued by someone else. Make sure,
therefore, that your note book is clear enough so that someone else can take over.
169
LABORATORY PRACTICE
8.2 Experiment Technique
The range of choice of technique in the average student

laboratory is rather limited. However, within the limits of
this restriction, the student should learn to take responsibility himself for his choice of measurement conditions. He
is going to have to accept this responsibility later on in
professional life and first year university is none too soon to
start training for it. He must therefore learn to decide to
use a micrometer rather than a meter stick to measure a
wire diameter, the range over which readings are desirable,
and how many readings are necessary, etc. Even safety
factors to a certain extent, have to be left to the student, so
consider matters of personal hazard or apparatus limitations, and do not assume merely that someone else has already taken care of these things. This has all been stressed
in the sections on experiment design, but the necessity for
the student to plan his own experiment is also found at the
level of setting up. The general advice is to think continuously about what one is doing and make every action the
consequence of a considered experimental necessity. For
example, in electrical work decide in advance the value and
power rating of resistor requirements, and use these values
instead of wiring in the nearest rheostat because it happens
to have been left beside the apparatus. A large fraction of
the sales of General Radio potentiometers must arise from
failure to use a little preliminary thought.
There are two phases in the actual performance of the
experiment; choice of apparatus and its actual manipulation.
170
LABORATORY PRACTICE
(a) Choice oj Apparatus

This choice is naturally limited in elementary laboratories,
and any such restrictions will already have been incorporated in the measurement program. However, if choice
is available, the particular item will be selected on the basis,
once more, of the measurement program. This is the point
at which to watch again for ratings. It would be foolish to
write out an experiment plan calling for currents up to 10
amp. when the only available resistor is clearly marked
with a maximum current of 5 amp. Always bear such ratings in mind, look for them on equipment, and plan the
experiment accordingly. If there is no rating marked on a
particular piece of equipment, find its instruction book or
its entry in a catalogue and look there.
It is almost always best to tackle experimental problems
(like all problems) in pieces. Consider each independent
section of the experiment separately and thus simplify the
work. Choose the experimental conditions to have as many
built-in checks as possible (see page 113).
(b) Performance oj Experiment

Make sure that your experimental arrangement is orderly
and, especially, convenient. Good, thoughtful experimenting cannot be carried out if some delicate adjustment
has to be made by stretching to the back of the bench over
exposed terminals at 500 v while watching a meter at the
front. It even helps merely to have it look nice. In general,
use a thoughtful approach throughout. Make sure that you
are in charge of the experiment, not the other way around.
When setting up apparatus, treat the experiment in sections. Set up each section and ensure that it works correctly before starting on the next bit. If one assembles a
171
LABORATORY PRACTICE
huge, complicated piece of apparatus without such preliminary piece-by-piece checking, the probability that it all
works the first time is close to zero, and the problems of
identifying the trouble, or troubles, may be insurmountable.
When constructing an electrical circuit always work from
a carefully drawn circuit diagram. When starting up apparatus always start with the least sensitive condition, i.e.
the maximum ranges on meters, maximum resistance 10
rheostats, minimum settings on potential dividers, etc.
Always check the zero on every instrument. Check, too, the
calibration if at all possible, because errors in calibration
constitute a type of systematic error which it is not normally
possible to detect using the internal consistency methods of
Chapter 5. In precise work it is not normally adequate to
check instruments at one point only, and so they should be
tested at all points on the range in use. When this is done
a graph can be drawn giving either the corrected reading
at each point on the scale, or else the error correction to be
applied to each scale reading. Such a curve is called a
calibration curve and is usually supplied with instruments
designed for precision measurements.
The instruments most prone to errors of calibration are
frequently the easiest to check, clocks and watches, ther,
mometers, etc., and there is rarely any excuse for ignoring
this. Electrical meters may be more of a problem if standards are not available. It is quite surprising how inaccurate cheap meters can be, and no measurement in which a
precision of better than perhaps 5 per cent is desired should
be based on them. In the case of a multi-range meter for
which a detailed calibration is not possible, try to restrict
the readings to one range of the instrument. This will not
172
LABORATORY PRACTICE
eliminate calibration errors but will at least ensure that

one has only one potential calibration error to consider.
Calibration errors arise, too, from backlash in mechanical
instruments. If the thread is loose on a micrometer eyepiece, a finite rotation of the screw (and consequently a
finite reading on the scale) is required before the eyepiece
cross wire starts to move. Errors due to backlash in such
mechanical devices can be eliminated by taking all the
readings going the same way, and care should be taken to
ensure that this be done.
One practical point associated with the reading of instrument scales is the phenomenon of parallax. If the position
of some index (e.g., the top of a barometer mercury column
or an electrical meter needle) must be read on some scale
which does not lie in the same plane as the index, the indicated value will vary with the position of the observer's eye.
This is illustrated in Fig. 8.1 from which it can be seen that
the error increases as the distance between the scale and the
index increases. This source of uncertainty can be minimized in many ways. If the observer has control over the
distance between the measured point and the scale, he
~ Point
whose position
is being measured
111:. 1.1
Parallax.
173
LABORATORY PRACTICE
should use it to bring them as close together as possible e.g.,

a meter stick should be used standing on its edge so that the
graduations are actually in contact with the object to be
measured. Frequently, a barometer will have the sliding
part of its scale double, one part in front of and the other
behind the mercury column. Therefore, one lines up the
two parts of the scale with the top of the mercury column
and thus eliminates angular effects. Good quality electric
meters have a mirror in the plane of the scale so that if one
lines up the pointer with its reflection a reading is ensured
with the line of sight perpendicular to the scale.
During the actual taking of readings it is usually very help,
ful to plot a rough graph of the variables. This will prO\ ide
a running check on the performance of the experiment and
may suggest changes in the measurement program to maximize the yield of the experiment.
Appendix
The Jollowing sections are intended to give a more detailed

mathematical treatment oj some oj the topics introduced in the
main body oj the text. Such a treatment was unnecessary at that
point and would merely have served to interrupt the progress oj
the development. The reader who wishes a justification Jor
many oj the equations used will find it in these appendices.
Any topic which is not mentioned cannot be treated adequately
here and reference must be made to the more advanced texts.
The Statistics of the Gaussian or

Normal Distribution
Al.l The Equation of the Normal Distribution Curve
Consider that a quantity whose true value is X is subject to

random uncertainty. Consider that the uncertainty arises
from a number, 2n, of errors each of magnitude E and
equally likely to be positive or negative. The measured
value x can then range all the way between X - 2nE if all
the errors should happen to have the same sign in the
negative direction to X
2nE if the same thing happened
positively. We require the distribution curve of a very
large number of readings each subject to this uncertainty
so that the quantity we must calculate is the probability
of encountering a particular error R within this range of
2nE. This probability is governed by the number of ways
in which a particular error can be generated. For example, an error of the total value 2nE can be generated in only
one way, all the elementary errors must have happened to
have the same sign simultaneously. An error of magnitude
(2n - 2)E, on the other hand, can occur in many ways.
If anyone of the elementary errors had been negative the
total error would have added up to (2n - 2)E and this
l77
178
THE STATISTICS OF THE GAUSSIAN OR NORMAL DISTRIBUTION
situation can arise in 2n different ways. An error of

(2n - 2)E is, therefore, 2n times as likely as one of 2nE.
A situation in which two of the elementary errors had
negative signs can correspondingly be generated in many
more ways than the previous one and so on. This argument
can be generalized, therefore, by using the number of ways
in which a specific error R can be generated as a measure
of the probability of the occurrence of the error R, and
consequently as a measure of its frequency in a umverse
of observations.
Consider a total error R of magnitude 2rE (where r < n).
This must be the result of some combination of errors of
which (n + r) are positive and (n - r) are negative. The
number of ways in which this can happen can be calculated as follows: The number of ways of selecting any
particular arrangement of 2n items is (2n)!. However, not
all of these arrangements are different for our purpose since
we do not care if there is any internal rearrangement
between the errors in the positive group. We must divide
the total number of arrangements, therefore, by the number of these insignificant rearrangements, i.e., by (n + r) !.
Similarly we must divide by the number of internal rearrangements possible in the negative group, i.e., by (n - r)!.
The total number of arrangements is, therefore,
(n
(2r) !
r)!(n - r)!
This is not yet strictly a probability although it is a measure

of the likelihood of finding such a total error. The probability itself will be obtained by multiplying the above
number by the probability of this combination of (n + r)
179
positive and (n - r) negative choices. Since the probability of each choice is t the required multiplier is
.!(n+r)
.!(n-r)
2
2
The final result for the probability of the error R is then
.
(n
(2n) !
+ r)!(n -
___(!)(n+r)
r)! 2
(!)(n-r)
2
(al)
Our problem is now to evaluate this as a function of the

variable r. This is done subject to the condition that m is
very large, in fact tending to infinity.
The evaluation requires two auxiliary results.
(i) The first auxiliary result:
n! = V27rn e-nn n
This is known as Stirling's theorem. Although its full derivation is beyond our scope its plausibility can be indicated
as follows:
h log x dx = [x log x n
= n log n -
xJ~
n
+1
The graph oflog x vs. x is shown in Fig. A1.1. Clearly the

integral above can be approximated by the sum
log 1
+ log 2 + log 3 + ... + log n
which is log (1 X 2 X 3 X ... X n) or log n!. We can,

therefore, write approximately if n is large
log n! = n log n - n
i.e.,
This is an approximation to the formula given above and
the reader is referred to Reference 8 in the Bibliography
for a full derivation.
180
FIe. AI.1
The evaluation of log n!.
(ii) The second auxiliary result:

lim
n~""
(1 + l)n
n
The expansion for [1
+ .!!.. ! + n(n l!n
2!
1)
+ (l/n)Jn is
(!)2
+ n(n n
1) (n - 2)
3!
(!)3
+
n
As n becomes larger all the terms in n clearly tend to unity

so that the series tends to
as required.
We are now in a position to evaluate the expression (al).
Apply Stirling's theorem to the terms (21/)!, (n + r)! and
(n - r)!
181
=
r)!
(n -
nn+r+(l/2)
= nn-r+(l/2)
+ ;r)n+r+(l/2)
_ ;;:;e-n-rV 2'1l"
r)n-r+(l/z)
_ 1e-n+rV 2'1l"
1 - ;
Therefore,
(n
+ r) !(n -
r)!
The variable part of (al) can now be written

r~)-n-(L'2)
( 1 - n-2
r)-r (1+-r)r
( 1-11
(1 r2)(1 + -r)nIT(
- -
(n'/ r')(r'/n) (
n'2
-r'ln) (
11
Thus expression (al) can now be written
1
~
(1 _r2)-<n'/T')(T'/n) (1 _r2)-1/2
n2
n2
1 _ r:)-1/2
n_
~) -n/r(-r'/n)
182
Now
as
r2)-1/2
( 1 - -n 2
(1
n
...
r
~oo
~1
+;r)nlr ~e
r)-nlr
( 1 - n-
~e
Thus finally, the probability of error R is

___
1_ e- r2ln
V;'n
The significant feature of this result is the form e- r'. It
specifies the probability of an error R and is thus equivalent
to Equation (2.4) in which the error is the difference between the true value X and the measured value x. The only
problem that remains in putting the equation into standard
form is to redefine the constants. Put
r
hx
= -.;,;"7t
for the exponent and in the constant replace

The equation then reads
P(x) dx
l/Vn by h dx.
=~; e- h' x' dx
where P(x) dx is the probability of finding an error between x and x + dx.
183
A1.2 Standard Deviation of the Normal Distribution
We must calculate the sum of the squares of the errors

divided by the total number of observations. Let there be N
observations where N can be assumed to be a very large
number.
The number of errors of magnitude between x and x
Nh
equals vi; e- h' x ' dx.
Therefore,
(7'2
1. fOO
N
-00
N h_
e-h'x'
vl7r
x 2 dx
The integral is a standard one and has a value

Therefore,
2
(7'
h
vi;
= vi:;;: . 2h 3 =
+ dx
vi;' /2h 3
2h2
This provides the justification for Equation (2.5) and enables us to rewrite the probability function
P(x) dx =
---c=-
vl27r (7'
e- x' /2rs'dx
Al.3 Areas Under the Normal Distribution Curve
The probability that an error falls between x and x
+ dx is
184
Therefore, the probability that an error lies between

o and x is
{x
Jo
V21r
(T
e-
'I? d
-"
Although this integral can be easily evaluated for infinite

limits it is not so simple for fixed limits as we now require.
Probability that an
error lies between
o and x
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
1.9
2.0
2.5
3.0
0
0.04
0.08
0.12
0.16
0.19
0.23
0.26
0.29
0.32
0.34
0.36
0.38
0.40
0.42
0.43
0.45
0.46
0.46
0.47
0.48
0.49
0.499
185
FII. Al.2
The area evaluated in calculating the probability of occurrence of an

error up to x.
Numerical methods of integration are used with results

given in the table on page 184 (see Fig. A1.2).
If we require the probability that an error lies between
x/u the value is, of course, doubled. For example, the
entry at x/u = 1 is 0.34 giving the 68 per cent figure that
we have been using for 1u limits. The table is intended
to give only an indication of the way the probabilities run
and for statistical work reference should be made to one
of the many statistical tables available (e.g., Reference 7 in
the Bibliography).
Least Squares Fit to Straight Lines
Consider a set of observations (Xi, Yi) to which it is desired

to fit a linear relation
y=mx+b
We assume that the X values are precise, that all the uncertainty is contained in the Y values and that the weights
of the y values are equal. The differences whose sum of
squares it is desired to minimize are of the form
OYi
Yi - (mxi
+ b)
Therefore,
+ b) P
= y~ + m2x~ + b2 + 2mx ib -
(0)'i)2 = [yi - (mxi
2mxiYi - 2y ib
If there are n pairs of observations the sum is

M = ~ (OYi)2 = ~y~
m2~x~
nb 2 2mb~xi
- 2m
XiYi - 2b
The condition for the best choice of m and b is that

should be a minimum
a~
am
and
aM =
ab
The first condition gives
+ 2b ~ x, 2nb + 2m ~ Xi -
2m ~ x~
The second
186
2 ~ (XiYi) = 0
2 ~ Yi = 0
~Yi
~ (OYi)~
187
LEAST SQUARES FIT TO STRAIGHT LINES
Solution of the simultaneous equations for m and b gives
m=
n ~ (XiYi) - ~ Xi ~ _Yi
n ~ X; - (~X i) 2
~ X~~ Yi - ~Xi~ (XiYi)

n ~ X~ - (~ X i) 2
Standard deviations for the m and b can be calculated as

follows: The calculated values of m and b are functions of
the quantities Yi . The standard deviations for m and bare,
therefore, to be calculated using Equation (3.3) for the
standard deviation of computed functions. They will be
calculated in terms of the standard deviation of the Y
values. This was written as Equation (6.3) using the quantities 0Yi
SII
~~ (oyS~
n-2
A justification of the value n - 2 will not be attempted but

it is associated with the fact that the 0Yi are not independent but are connected by the existence of the best line
given by m and b.
Equation (3.3) for the standard deviation of a computed
value says
This formula is applied to our case by noting that the x

and Y of the formula are the Y1, Y2, etc., which form part
of our set of observations. Thus the function for m is
m =
~------~~
n ~ Xi
(~xiF
[nxIYl - Yl ~ Xi
+ nX2Y2 -
Y2 ~ x.
+ ... ]
188
LEAST SQUARES FIT TO STRAIGHT LINES
Therefore,
and
(~;y
(n
~ Xi2 ~ (~ Xi)2) [n2x~ + (~ Xi)

2
2 -
2nXk
~ x;]
Since S1I is common to all the contributions we can sum the

2
directly to obtain
(am/ak)
~ (~;y = (n ~ x1 ~ (~Xi)2)2
[n 2 ~ X~
since ~ Xk =
Therefore,
~k (aa m)2
~k
+ n(~ Xi)2 -
2n(~
Xi, etc.
= (
~ X,2 - 1 (~ Xi )2)2 [n 2~ x1 - n(~ Xi)2]

n
n ~ X~
or
The value for
Sm
Sb
S1I
(~Xi) 2
'
__
n_ _.,--,
\j~ xl - (~Xi)2
can be found by the same procedure.
Xi)2]
BIBLIOGRAPHY
The following selection of books may prove useful in the furthel

study of the topics mentioned in the text. It is not intended to be
exhaustive.
1
Beers, Y., Introduction to the Theory of Error, Addison-Wesley,

1957. A short text dealing with the applications of error
theory to experimenting.
Braddick, H. J. J ., The Physics of the Experimental Method,

Chapman and Hall, London, 1956. A study of practical
experimenting techniques including observation treatment.
Cox, D. R., Planning of Experiments, Wiley, 1958. An account

of the statistical methods for choosing an experiment program for maximum information (see page 94).
Fretter, W. B., Introduction to Experimental Physics, PrenticeHall, 1954. An account of the experimental methods used
in physics research.
Freund, J. E., Modern Elementary Statistics, Prentice-Hall,

1961. An introduction to statistics.
6 Jeffreys, H., Scientific Inference, Cambridge University Press,

1957. A study of the applications of probability theory to
experimental information.
180
Lindley, D. V., and J. C. P. Miller, Cambridge Elementary

Statistical Tables, Cambridge University Press, 1958.
Margenau, H., and G. M. Murphy, The Mathematics of

Physics and Chemistry, Van Nostrand, 1947.
190
BIBLIOGRp,PHY
Menzel, D. H., H. M. Jones, and L. G. Boyd, Writing a

Technical Paper, McGraw-Hill, 1961. A survey of the principles and practice of composition in scientific writing.
10
Parratt, L. G., Probability and Experimental Errors in Science,

Wiley, 1961. A study of the applications of probability and
statistics in measurement.
11
Schenck, H., Theories of Engineering Experimentation, McGrawHill, 1961. An account of those experiment design techniques which refer more specifically to engineering systems.
12
Shamos, M. H. (editor), Great Experiments in Physics, HoltDryden, 1960. A compilation of the writings of great
physicists describing their own experiments.
13
Stanton, R. G ., Numerical Methods for Science and Engineering,

Prentice-Hall, 1961. This text is devoted mostly to numerical methods in mathematics but many of the techniques
have application to curve fitting.
14
Strong,]., Procedures in Experimental Physics, Prentice-Hall,

1938. An account of the experimental methods used in
physics research.
15
Topping, J., Errors of Observation and t~eir Treatment, The

Institute of Physics, London, 1955. A short text dealing
with the applications of error theory to experimenting.
16
Tuttle, L., and J. Satterley, The Theory of Measurements,

Longmans, Green, 1925. An account of the application of
statistics to measurement and of numerical methods in the
treatment of observations.
17 Whittaker, E. T., and G. Robinson, The Calculus of Observt7tions, Blackie, Glasgow, 1944. An extensive treatment of the
mathematical aspects of observation treatment.
191
BIBLIOGRAPHY
18
Wilson, E. B., An Introduction to Scientific Research, McGrawHill, 1952. An account of the nature of scientific experimenting and the methods used.
19
Wortham, A. W., and T. E. Smith, Practical Statistics in

Experimental Design, Charles E. Merrill, 1959. A short treatment of the experiment design methods used in engineering.
20 Worthing, A. G., and J. Geffner, Treatment of Experimental

Data, Wiley, 1946. An account of the application of statistics
to measurement and of numerical methods in the treatment
of observations.
PROBLEM SOLUTIONS
Chapter 2
1.
b
c:
CT
C1)
U:
Reading
2.
3.
4.
5.
6.
Between 38 and 39; 38

38.30
4.388
0.801
0.576
7. a) 4.388 about the universe
mean
b) 8.776 about the universe
mean
mean

mean
standard deviation
standard deviation
10. h = 0.161,
P = 2.940
11. rejection
13. about 130

14. about 207
Chapter 3
192
1. 0.05 cm; 0.4%
6. a) 1%; b) 5%
2.0.5%
7. a) 2 sec; b) 100 sec
3. 5 cm; 1 cm
4. a) No; b) Yes
8. 0.07% assuming the watch

gains uniformly
5.0.007%
9.3.7%
193
PROBLEM SOLUTIONS
Chapter 3 (contInued)
10. 0.2 ohms

11. 0.012 g
20. 3.60
ce
21. 0.005
12. 5.5%
13. 1.2 X 10 11 dynes cm- 2
22. 7.6% the wrong way; 3.2%

the right way
15. 0.05%
23. 0.00015
24. 0.045
16. 77 ohms
25. 2.1
14. 0.25%
26. 0.032
17. 1.9 cals g-l

18. approximately 50
27. 14.9 cm sec-2
19. 0.2 ohms
28. 0.0027 cgs units
ChapterS
1. No
10. d vertically, W[3 horizontally,
2. range ex: (
velo!!city
~
)2
surface tension
3. pressure ex:
radius
4. period ex:
moment of inertia
rigidity constant
5. deflection ex: (load force 2 ).

y X radius
slope is 4/ Yab 3
11. h vertically, 1/R horizontally,
slope is 20"/pg
12.
P vertically, T horizontally,
slope is R/o
13.
vertically, / tally, slope is 0
/00
/0
horizon-
(len~th)b radius where a
14. l vertically, 6.t horizontally,

slope is loa and intercept is 1o,
whence a
6. s vertically, t 2 horizontally,
slope is ~a
15. sin (h vertically, sin 82 horizontally, slope is 1l2/III
7. l' vertically, n2(Z horizontally,
16. l/s vertically, l/s' horizontally, each intercept is 1/!; or

ss' vertically, l '
s' horizontally, slope is /
radIUs
and b are arbitrary constants
slope is 4 m
8. P vertically,
slope is p/2
02
horizontally,
9. 1'2 vertically, cos a horizontally, slope is 47r2l/g
17. l/c vertically, w2 horizontally,

slope is L.
194
PROBLEM SOLUTIONS
Chapter 5 (continued)
18. F vertically, 1/72 horizontally,

check for linearity
19. F vertically, i 1i 2/T 2 horizontally, check for linearity and
check F vs. it, F vs. i2 and
F vs. l/T separately, holding
other variables constant.
20. log Q vertically, t horizontally,

slope is -1/ RC
21. Z2 vertically, 1/w2 horizon-
23. 1/:\ vertically, 1/n 2 horizontally, slope is - R, intercept is

R/4.
24. log J/T2 vertically, l/T horizontally, slope is -cp/k, intercept is log A
25. mw(T2 -
T 3)
m.(T1 - T 2)
slope is S.
vertically,
horizontally,
tally, slope is 1/c2 and intercept is R2
26. m( T4 - T 3) vertically, T2 - Tl
horizontally, slope is KA/d
22. m 2 vertically, m 2v2 horizontally,

slope is 1/c 2 and intercept is
27. VI/(HT 1+ T 2) - To) verticalT 2)

ly, m(T2 - T1)/(HT 1
- To) horizontally slope is JS.
m5
Chapter 6
1. 14.2 0.1
g) 6.13
2. 14.25 0.15
h) 5.4 and 6.7 are the limits

giving a maximum range
of uncertainty of 0.7
3. 6.75 0.03
4. 38.30 0.80
5. c) 12.5 X 10-4 ohm 2 sec 2

d) 5.63 X 10-3 henry
e) 11.2 X 10--4 and 14.2 X
10-4 ohm 2 sec 2; +1.7 X
10-4, -1.3 X 10-4 ohm 2
sec 2
f) 5.33 X 10-3 henry to
6.00 X 10-3 henry gives a
maximum range of uncertainty of 0.37 X 10-3
henry
i) L = (5.6 .4) X 10-3

henry
R = 6.1 0.7 ohm
6. Mean is 17.44 and standard
deviation of mean is 0.25
7. a) Slope = 4.988 X 10-20hms
deg- 1 ; intercept = 11.916
ohms
b) 4.186 X 10-3 deg-1

c) Ssiope = 2.430 X 10-3 ohms
deg-1
Sintercept = 0.1227 ohms
195
PROBLEM SOLUTIONS
Chapter 6 (continued)
d) Sa = 2.084 X 10-4 deg-1

e) a = 4.19 X 1O-30.21 X
10-3 deg-1
Ro = 11.92 0.12 ohms
8. a) i = 0.5 e'"
b) y = 0.6 X 2 4
c) f = 6.2 e-366I'I'
Index
Absolute uncertainty, 49
Backlash, errors arising from, 11,

172
Best value:
and distribution curves, 19
of slope and intercept by least
squares, 138
Calibration, errors arising from,

11,171
Calibration curves, 171
Compensating errors, 60
Controls, 85
196
Diagrams, in reports, 160

Difference type measurements, 8487
Differences:
standard deviation of, 65
uncertainty in, 55
Differentiation, use of in error
propagation, 57
Dimensional analysis, 95-99
Discussion, in reports, 164
Distribution curves, 16-25
Error (see Uncertainty)

Experiment:
role of in research, 76
types of, 79-87
Exponential functions:
standard deviation in, 66
uncertainty in, 53
Function finding, 140
Gaussian distribution:
areas under, 31, 183-185
correspondence with actual observations, 23
definition of, 26, 178
equation of, 27, 178-182
probable error in, 28
standard deviation of, 31, 183
Graphs:
calculation of slope of, 130
drawing, 127
in reports, 162
uncertainty of slope, 131
use in checking validity of experiments, 103
Histogram, 17
Introduction, in reports, 155-158
Least squares principle:

applied to sets of observation,
137
application to straight lines, 138,
186-188
basis of, 133-136
Logarithmic functions:
standard deviation in, 66
uncertainty in, 53
Logarithmic plotting, 140-142
Mean, definition of, 22

Mean deviation, definition of, 23
Measurement, nature of, 6
Measurement program, 112-116
Median, definition of, 22
Mode, definition of, 22
Models, 76
Normal distribution (see Gaussian

distribution)
Note book, 167
Nul measurement, 87
t97
Parallax, errors arising from, 172

Personal errors, 12
Poisson distribution, 29
Population, definition of, 30
Powers :
uncertainty in, 52
Precision, definition of, 49
Probable error:
definition of, 24
in Gaussian distribution, 28
Procedure, in reports, 158-161
Products:
uncertainty in, 57
Propagation of uncertainty :
general methods for, 51 , 56
use of logarithmic implicit differentiation, 58
Quotients, uncertainty in, 59
Random error, definition of, 10

Randomization of readings, 115
Ratings, instrument, 113, 170
Rectification of curves, 106-110

Rejection of readings, 40
Relative uncertainty, definition of,
49
Reversibility of experiments, 113
"Rounding off," errors arising
from, 13, 130
5
198
Sample in-sample out methods,

84-87, 114
Sampling, 31-38
Scale division, errors arising from,
13
Significant figures, 145
Slide rule, accuracy of, 131
Standard deviation :
best estimate of universe, 35
combined with uncertainty, 67
definition of, 24
of computed values, 61-66
of the mean, 33
of the standard deviation, 34
significance of, for Gaussian distribution, 30
Stirling'S theorem, 179
Sums:
uncertainty in, 55
Systematic error, definition of, 10
T
Trigonometric functions:
uncertainty in, 52
u Uncertainty:
and distribution curves, 16, 19
combined with standard deviation,67
propagation of, 49-61
sources of, 11
Universe, definition of, 30
v Variable:
dependent, 91
independent, 91
Variance, definition of, 23
w Weighted mean:
equation for, 138
,11\
Other books of interes' ...

MODERN PHYSICS, 2nd Edition
by F. W . VAN NAME, Jr. , University of Delaware
Planned to take the reader through the various stages of elementary atomic
physics, the book progresses from the study of individual particles in atoms
and molecules to the discussion of extra-nuclear particles, and finally, subnuclear particles. It is divided as follows:
SECTION I -Electron properties and the special theory of relativity-The
first three chapters constitute a discussion of the properties of individual
particles, particularly the electron. The units ureu ,throughout the book are
those of the MK~ system. The Planck Radiation Law is featured in Chapter 2.
SECTION II - A'tomic and molecular spectra, structure and properties-The
Bohr theory is introduced in connection with the explanation of spectral
seli es and its defects are discussed. New additions to Chapter 5 are the
Tunnel Effect and the Stern-Gerlach Experiment.
SECTION III-Nuclear physics and cosmic rays-This section covers nuclei
and nuclear particles without being concerned to any extent with extranuclear electrons of atoms and molecules.
Published 1961
319 pages
OPTICS OF THE ELECTROMAGNETIC SPECTRUM

by CHARLES L. ANDREWS, State University of New York
In this guide to general methods for solving problems of wave optics in
anticipation of new developments throughout the spectrum, Dr. Andrew's
treats the properties of wave motion common to the whole electromagnetic
spectrum.
Two features of his treatment which will appeal especially to those working
in this field are: the use of hand-sized wave lengths of microwaves
(in detail sufficient for the reader to perform them) to illustrate those
fundamentals of wave optics that are either impossible or very, difficult to
show with light waves the presentation of recent research in diffraction
of transverse waves in the near, as well as far, regions.
Published 1960
501 pages
PRENTICE-HALL, Inc.

Experimentation An Introduction To Measurement Theory and Experiment Design DC Baird

Uploaded by

Copyright:

Available Formats

Experimentation An Introduction To Measurement Theory and Experiment Design DC Baird

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Experimentation An Introduction To Measurement Theory and Experiment Design DC Baird

Uploaded by

Copyright:

Available Formats

What is the intended audience and purpose of the book?

What is the intended audience and purpose of the book?

What topics does the book cover?

What topics does the book cover?

Experimentation

Englewood Cliffs, New Jersey

D. C. Baird Associate Professor of Physics, Royal Military College of Canada

EXPERIMENTATION: AN INTRODUCTION TO MEASUREMENT

1962 by PRENTICE-HALL, INC.

Current printing (last digit):

Library of Congress Catalog Card Number 62-11883

text is not intended to be a complete treatise on industrial

Introduction te Laboratory Work 1

The Nature of Measurement 6

Measurement and Confidence, 6

The Propagation of Uncertainties 48

Absolute and Relative Uncertainty, 48

The Nature of Experimenting 75

Experiment Evaluation 123

Nature of Scientific Theory, 75

The Aims of Experiment Evaluation, 123

The Scientific Report 152

Aims in Report Writing, 152

Laboratory Practice 167

The Laboratory Note Book, 167

The Statistics of the Gaussian or Normal

Least Squares Fit to Straight Lines 186

Introduction to Laboratory Work

I t is the aim of this introduction to explain some of the

INTRODUCTION TO LABORATORY WORK

limited. It is obvious, therefore, that it is useless to lay

INTRODUCTION TO LABORATORY WORK

necessary practice in solving problems experimentally i~

INTRODUCTION TO LABORATORY WORK

work within the framework of the apparatus provided,

INTRODUCTlON TO LABORATORY WORK

and is a matter of widespread concern. It is the writer's

The Nature of Measurement

2.1 Measurement and Confidence

The observation of nature which constitutes an experiment

THE NATURE OF MEASUREMENT

is the kind of statement which a measurement is not. The

THE NATURE OF MEASUREMENT

observer a metrologist who means that he has an answer of

THE NATURE OF MEASUREMENT

certainty quoted by the original authors. The only thing

THE NATURE OF MEASUREMENT

2.2 Types of Uncertainty

The methods by which the uncertainty can be estimated

THE NATURE OF MEASUREMENT

As has been stated above, any measurement is subject to an

(a) Instrument Calibration

(b) Instrument Reproducibility

THE NATURE OF MEASUREMENT

many other cases where the final meter indication is de~

(c) Observer Skill

THE NATURE OF MEASUREMENT

must be considered. Uncertainties of personal origin may

(d) Miscellaneous Errors