David C. Baird Experimentation An Introduction To Measurement Theory and Experiment Design PDF

EXPERIMENTATION
An Introduction to Measurement Theory

and Experiment Design
Third Edition
D. C. Baird
Royal Military’ College
Kingston, Ontario
^ ♦
Prentice-Hall, Inc.
Englewood Cliffs, New Jersey 07632
Library o f Congress Cataloging-in-Publication Data
Baird, D. C. (David Carr)
(Experimentation : an introduction to measurement theory and
experiment design / D. C. Baird. — 3rd ed.
p. cm.
Includes bibliographical references and index.
ISBN 0-13-303298-1
1. Physical measurements. I. Title.
QC39.B\1 1995
507'.24—dc20 94-13384
CIP
Acquisition Editor: Ray Henderson

Assistant Acquisition Editor: Wendy Rivers
Editorial Assistant: Pam Holland-Moritz
Production Editors: Rose Keman and Fred Dahl
Copy Editor: Rose Keman
Production Coordinator: Tmdy Pisciotti
© 1995 by Prentice-Hall, Inc.

A Paramount Communications Company
Englewood Cliffs, New Jersey 07632
All rights reserved. No part of this book may be

reproduced, in any form or by any means,
without the permission in writing from the publisher.
Printed in the United States of America
10 9 8 7 6 5 4 3 2
ISBN D-lB-aOBaTfl-l
Prentice-Hall International (UK) Limited, London

Prentice-Hall of Australia Pty. Limited, Sydney
Prentice-Hall Canada Inc., Toronto
Prentice-Hall Hispanoamericana, S. A., Mexico
Prentice-Hall of India Private Limited, New Delhi
Prentice-Hall of Japan, Inc., Tokyo
Simon & Schuster Asia Pte. Ltd., Singapore
Editora Prentice-Hall do Brasil, Ltda., Rio De Janeiro
To M argaret
Contents
Preface, vii
Chapter 1: Approach to Laboratory Work, 1

1-1 Nature of Scientific Knowledge, 2
1- 2 Role of the Physics Laboratory, 4
Chapter 2: Measurement and Uncertainty, 10

2- 1 Basic Nature of Measuring Process, 10
2-2 Digital Display and Rounding Off, 12
2-3 Absolute and Relative Uncertainty, 13
2-4 Systematic Errors, 14
2-5 Uncertainty in Calculated Quantities, 15
2-6 Uncertainty in Functions of One Variable Only, 16
2-7 General Method for Uncertainty in Functions of a Single
Variable, 17
2-8 Uncertainty in Functions of Two or More Variables, 20
2-9 General Method for Uncertainty in Functions or Two or
More Variables, 22
2-10 Compensating Errors, 25
2- 11 Significant Figures, 26
Problems, 26
Chapter 3: Statistics of Observation, 29

3- 1 Statistical Uncertainty, 29
3-2 Histograms and Distributions, 30
3-3 Central Values of Distribution, 32
3-4 The Breadth of Distributions, 34
3-5 Significance of the Mean and Standard Deviation, 35
IV
CONTENTS
3-6 Gaussian Distributions and Sampling, 37

3-7 Relation Between Gaussian Distributions and Real
Observations, 40
3-8 Sample Means and Standard Deviation of the Mean, 41
3-9 Sample Standard Deviation, 43
3-10 Application of Sampling Theory to Real
Measurements, 43
3-11 Effect of Sample Size, 45
3-12 Standard Deviation of Computed Values, 47
3-13 Standard Deviation of Computed Values: Special
Cases, 50
3-14 Combination of Different Types of Uncertainty, 53
3- 15 Rejection of Readings, 54
Problems, 55
Chapter 4: Scientific Thinking and Experimenting, 57

4- 1 Observations and Models, 57
4-2 Construction of Models, 64
4-3 Testing Theoretical Models, 73
4-4 Use of Straight-Line Analysis, 77
4- 5 Case of Undetermined Constants, 79
Chapter 5: Experiment Design, 84

5- 1 To Text and Existing Model, 85
5-2 Straight-Line Form for Equations, 86
5-3 Planning and Experiments, 93
5-4 Experiment Design When There is No Existing Model, 99
5-5 Dimensional Analysis, 100
5-6 Difference-Type Measurements, 104
5-7 Experimenting with No Control Over Input Variables, 107
Problems, 109
Chapter 6: Experiment Evaluation, 113

6-1 General Approach, 113
6-2 The Stages of Experiment Evaluation, 115
6-3 Graphs, 118
6-4 Comparison Between Existing Models and Systems, 119
6-5 Calculation of Values from Straight-Line Analysis, 123
6-6 Cases of Imperfect Correspondence Between System and
Model, 128
6-7 The Principle of Least Squares, 129
6-8 Least-Squares Fit to Nonlinear Functions, 133
6-9 Precautions with Least-Squares Fitting, 134
VI CONTENTS
6-10 Function Finding, 135

6-11 Polynomial Representation, 137
6-12 Overall Precision of the Experiment, 138
6-13 The Concept of Correlation, 140
6- 14 Use of Computers in Experiment Evaluation, 145
Problems, 152
Chapter 7: Writing Scientific Reports, 157

7- 1 Good Writing Does Matter, 157
7-2 Title, 158
7-3 Format, 159
7-4 Introduction, 160
7-5 Procedure, 162
7-6 Results, 164
7-7 Graphs, 166
7-8 Discussion, 168
Appendix 1: Mathematical Properties of the Gaussian or

Normal Distribution, 171
A1-1 The Equation of the Gaussian Distribution Curve, 171
A1-2 Standard Deviation of the Gaussian Distribution, 176
A1-3 Areas Under the Gaussian Distribution Curve, 177
Appendix 2: The Principle of Least Squares, 179
A2-1 Least Squares and Sample Means, 179
A2-2 Least-Squares Fitted to Straight Lines, 180
A2-3 Weighting in Statistical Calculations, 183
Appendix 3: Difference Tables and the Calculus of Finite
Differences, 186
A3-1 Mathematical Foundations, 186
A3-2 Application of Difference Tables to Measured
Values, 193
Appendix 4: Specimen Experiment, 195
A4-1 Experiment Design, 195
A4-2 Report, 203
Bibliography, 208
Answers to Problems, 211
Index, 215
Preface
The first edition of this book was written with the conviction that, regardless
of the chosen objectives for an introductory physics laboratory, the basic
principles of experimenting should not be neglected and could in fact become
the principal topic. Introductory laboratories in physics are particularly suited
to this purpose since the systems and theories found there are usually simple
enough that the basic characteristics of measurement and experimenting can
easily be made visible and understandable Such an approach to physics
laboratory work can, therefore, be beneficial for a wide range of students, not
only those who will proceed to professional work in physics.
That purpose on which the 1962 edition was based seems still to be
valid. Many changes have taken place in the practice of experimenting, partly
through the introduction of new instrumentation, but mostly because of the
revolutionary impact of computing. Not only can we easily attain a level of
post-experiment data analysis that would have been completely impracticable
30 years ago, but the possibilities for the conduct of the experiment itself
have been enormously expanded by the availability of on-line data analysis
and computer-based control of the apparatus.
VII
VIII PREFACE
Revolutionary though such changes have been in the actual conduct of

experiments there has, nevertheless, been little or no change in the basic
principles underlying the experimenting, and training in these basic principles
is still required. Indeed, emphasis on these basic principles may be even more
necessary today than it was 30 years ago on account of the present-day pos
sibility than an experimenter can be completely insulated from the phenom
ena under study by an almost impenetrable barrier of data processing equip
ment and procedures. Under these circumstances, wholly invisible defects
can produce final results with little or no meaning. Unless we have complete
and clear understanding of all phases of our experiment and data analysis, we
turn over our experiment wholly to the computer at our peril.
The plan of the present edition is largely the same as in that first edition
but the text has been almost completely rewritten. Chapter 1 gives an outline
of an approach to introductory physics laboratory work that emphasizes the
basic nature of experimenting. Chapters 2, 3, and 4 provide the information
on measurement, statistics, and scientific procedures that is needed to
understand the basic principles of experiment design. Chapter 5 treats in a
step-by-step way the practical requirements in designing an experiment, and
Chapter 6 provides the corresponding procedures for evaluating the results of
the experiment after the measurements have been made. At the end of the
main text. Chapter 7 contains some suggestions for writing laboratory
reports. The main addition to the present edition is material on computer
processing of experimental observations.
The appendices contain material which, although desirable in itself, would

have interrupted the development within the main text. This includes
mathematical derivation of some of the equations quoted in the main text. In
addition, a sample experiment is described in extensive detail, starting at the
beginning of the experiment design, continuing through the conduct of the actual
experiment and the evaluation of results, and ending with the final report.
The material in the text has been based on many years of teaching in
our First Year Physics Laboratory, and I am grateful to the generations of
students whose sometimes painful experience with it provided the opportu
nity for continued refinement. I wish to express my appreciation, too, to Mr.
Peter Snell for valuable discussions and review of the text.
D.C. Baird
1
Approach
to Laboratory Work
This book is intended for use in introductory physics laboratories. It was

written, however, in the hope that it will serve a much wider purpose. It
provides an introduction to the study of experimenting in general, irrespec
tive of the area in which the experimenting is carried out. Some students in
an introductory physics laboratory may pursue careers in physics research,
and it is hoped that the book will serve as a suitable introduction to their
continued studies. Many others will pursue careers in completely different
areas, perhaps in other sciences, or perhaps in areas outside science alto
gether. Whatever the need, the introductory physics laboratory can, if suitably
oriented, provide a useful introduction to the fundamental principles that
underlie experimenting of any kind. For our purposes, the word ex
perimentation has a very broad definition. By it we mean the whole process
of identifying a portion of the world around us, obtaining information from it,
and interpreting that information. This definition covers a wide range of
activities— all the way from the traditional picture of a biologist in a white
coat splicing DNA molecules to a manufacturer taking a poll to determine
individual preferences in toothpaste. This book is intended to meet the needs
of all who are either engaged themselves in any kind of study of the world
around us, or who must form a judgment on scientific statements made by
others.
2 APPROACH TO LABORATORY WORK CHAP. 1
1 -1 NATURE OF SCIENTIFIC KNOWLEDGE
It is natural to ask why everyone, not just practicing scientists, should be

familiar with the processes by which we gain knowledge of the world around
us. The answer is found in the extensive part experimenting plays in our
lives, whether we are aware of it or not. Even if we are not practicing scien
tists ourselves, almost all of us in our daily lives are faced with the require
ment to use or to pass judgment on experimental information offered by
others. For example, our professional work may require us to make a choice
between competitive bids on equipment having certain specifications, or as
members of the general public, we may be called on to form opinions on such
issues as the possible health hazards of nuclear power plants, the safety of
food additives, the impact of acid rain on the environment, or the influence of
national monetary policy on unemployment. All these issues require us to be
familiar with the nature of the processes of scientific experiments and to
make decisions that are based on our own, appropriately skeptical assessment
of the reliability of the experimental information.
To do this we must first become knowledgeable about the nature of

measurement itself. In specific, we must be clearly aware that measured val
ues cannot be exact. What we call the uncertainty in the measurement can
arise either from limitations of the instruments or from statistical fluctuation
in the quantity being measured. Whatever the origin of the uncertainty, we
must be aware of its existence and know how to estimate it. Only then can we
know how much confidence to place in the measured value.
Once the nature of measurement has been settled, a second, equally

significant requirement remains. Despite all the best efforts of science edu
cators worldwide, much misunderstanding still remains concerning the basic
nature of scientific statement. The misunderstanding commonly centers on
the question of the reliability or authority of the statements. Attitudes vary all
the way from unquestioning faith that some point has been “scientifically
proved,” and so must be infallibly correct, to complete scepticism that all
science is “just theories,” and so can be confidently disregarded.
As can easily be expected, neither extreme position is valid, and public

life will be better served if we are all able to take the scientific or technical
statements we hear and place them appropriately along the scale of
credibility. To be able to do so, one vital point must be appreciated even be
fore we turn our attention to the mechanical details of the processes by which
the information is obtained. It is a point that is frequently neglected in public
statements about scientific or technical affairs, and yet it is indispensable for
their proper comprehension.
CHAP 1. APPROACH TO LABORATORY WORK 3
The point is this. There exists an all-important distinction between the

portion of the real world that is under discussion (we refer to it as the system
under study) and the ideas and concepts (generally referred to as the model)
that we invent in our heads as a consequence of observing the system. There
is usually little difficulty in comprehending the nature and status of the first
of these, the measurements. The second, however, needs a little elaboration.
Our purpose in inventing ideas is to represent the observed properties of the
system in a kind of shorthand way so that we can talk to each other about the
system conveniently, easily, and with a conunon basis of understanding. For
example, if we were the first people on earth exploring our environment for
the first time, we might notice each day a certain type of tree in our wander
ings. But instead of reporting today’s sighting and yesterday's sighting
separately, and so on for the preceding month, as if each observed event were
unrelated, it would be much more convenient to invent, using a certain set of
described properties, the abstract concept “banana.” Such a concept would
allow us to plan together for tomorrow's dinner much more expeditiously and
with better economy of communication than would be the case without it.
Beyond such simple examples, the use of models is widespread, significant,
and sophisticated.
As we pursue our daily lives, it is easy to forget that many of the items
that appear in routine communication concern concepts and ideas and are not
genuine statements about the real world. Frequently, the distinction is com
pletely unimportant, and we can get away with careless use of language. Oc
casionally, however, the distinction is vital, and serious error can result from
inattention to it.
The danger arises because the two aspects of our knowledge of the ex
ternal world have completely different character. On the one hand, observa
tions that we make on our system belong to the real world and can (subject to
the necessary presence of uncertainty that is discussed in Chapter 2) have the
status of genuine incontrovertibility. As an example, no thinking person
would question the assertion that we can measure the width of the Atlantic
Ocean to be greater than the length of our living room. With varying degrees
of confidence, therefore, all statements about observations of the real world
have the potential of being immune against refutation, and this can mislead
us by falsely reinforcing the notion that all scientific statements contain
incontrovertible truth about the universe.
Unlike statements about measurements, however, statements about our

ideas and concepts have no claim whatsoever as absolute knowledge about
the universe. They are nothing more than ideas invented in our heads, and
even if they are ideas that have been very carefully chosen to represent the
APPROACH TO LABORATORY WORK CHAP. 1
properties of the system as closely as possible, they remain nothing more

than just that: ideas in our heads. Not only must they remain provisional,
subject to improvement or replacement if someone comes along with a better
idea, but it is impossible for them to qualify, as can observations, as certain
knowledge of the external world.
Failure to comprehend the complementary roles of observation and

concept in science is very common in public debates about scientific matters
and is the source of much confusion. For example, a famous economist was
recently heard to marvel publicly that a drop in interest rates had failed to
provide the desired stimulus to the economy. He neglected to mention that
the hoped-for relationship between interest rates and economic activity be
longed to his model; the real economy, of course, has ideas of its own.
All statements that come from scientific study of a system fall into one
or other of the categories that have been mentioned. They may be statements
about observations of a system, they may be statements about models, or they
may be statements about the relationship between a system and a model. If
we keep these possibilities clearly in mind as we listen to scientific
statements and analyse them into these categories, we shall have gone far
toward forming accurate judgements.
As far as our own statements on scientific matters are concerned, we

should accept the responsibility to be precise about the language we use. We
still hear pronouncements from famous scientists that a “correct theory” for
something has been found. They may themselves have clear understanding
about what they mean, and they do not mislead those of us who know how to
decode such conventional language. But the potential for misunderstanding in
nonscientific circles is too serious, and all of us who make scientific
statements should watch our language very carefully.
1- 2 ROLE OF THE PHYSICS LABORATORY
But what does all this have to do with possible purposes for the introductory
physics laboratory? Physics teaching laboratories have played such a familiar
role that it is natural to wonder how the normal laboratory with its usual
experiments can be used to provide an introduction to experimenting in gen
eral. The answer lies not so much in the experiments themselves as in the
attitude with which we approach them. This will become clear as our studies
of experimental methods proceed, but it may be helpful at this point to illus
trate the suggestion by anticipating in outline the arguments that are dis
cussed in more detail later. We have already mentioned that our studies on
experimental methods will be clarified if we regard the piece of the universe

under study as a system. By a system we mean, in general, any isolated, de
fined entity that functions in a specific manner. We assume that we can in
fluence or control the system, and we refer to the methods we have available
to do this as inputs. We also assume that the system performs some identifi
able function or functions, and we refer to these as outputs. The various ex
amples that follow will make clear the use of the terminology. An economist,
for example, may view the economy of a country as a system with an exten
sive set of inputs and a correspondingly varied set of outputs. The system it
self includes the whole productive capacity for goods and services, transpor
tation facilities, supply of raw materials, inhabitants, opportunities for foreign
trade, weather, and many other things. The inputs are those things that can be
controlled by us—the money supply, tax rates, government spending, tariffs
on imports, and so on. The outputs are those things that we cannot control
directly; their magnitudes are determined by the system, not by us. Outputs
of an economic system include the gross national product, unemployment
rate, inflation rate, external trade balance, and the like. It would be very
comforting and convenient if we could secure the desired values of these
outputs by simple manipulation, but we cannot. No matter how desirable it
may be, we cannot instruct the country's gross national product or
unemployment rate to have a certain value; we are restricted to controlling
the inputs. Even there we have problems. In a system as complex as a na
tional economy, the linkages between the inputs and the outputs are tangled
and indirect. A change in one input variable will likely have an effect on a
number of output variables, instead of solely on the single output in which
we may be interested. For example, an attempt to increase the gross national
product of a country by reducing taxation rates will possibly be at least par
tially successful, but the simultaneous effects on other outputs may be
equally prominent and not nearly as desirable—for example, a possible in
crease in the rate of inflation. The methods available for handling such
situations are sophisticated, but given the complexity of the system, the level
of success achieved by the politicians and economists shows that substantial
room for improvement still remains.
There are other systems that, although still complex, are simple enough
for us to control them reasonably successfully. Consider, for example, an
electrical power generating and distribution system. It, too, is an enormous
and complex system. It has many inputs, such as the number of generators
that are started up, the routing of the power through the transmission lines,
the salaries paid to the staff, the hours of work on the shifts, the price charged
to the customers per kilowatt-hour, and so forth. It has obvious outputs, such
as the power delivered to various districts, and also other, less tangible
outputs, such as the overall efficiency of the system and the reliability of
APPROACH TO LABORATORY WORK CHAP. 1
service. These are quantities that cannot be controlled directly by the man
agement; the system decides their values, and so they must be counted as
outputs of the system.
How does all this refer to the introductory physics laboratory? If we are
to prepare people to enter a scientifically literate population, would it not be
better to tackle the important problems right away and start deciding whether
the mercury content of fish makes it safe to eat or whether our use of fossil
fuels is contributing to global warming? The trouble is that these are ex
tremely difficult problems. Evidence is hard to obtain and its interpretation is
usually uncertain; even the experts themselves disagree, often vigorously and
publicly. It is almost impossible to make a significant contribution to the
solution of such complex problems without first developing our skills
through using simpler situations.
To see how this can be done, let us think about some of these simpler
systems. An automobile engine is a system that is simple in comparison with
any of the earlier examples. The system includes only the engine, fuel supply,
mounting, surrounding atmosphere, and so on. The inputs may be the
obvious controls such as fuel supply, fuel-air ratio, and ignition timing (even
though some of our direct control may have been usurped by the computer
ized systems now found in automobiles, these quantities remain as inputs to
the engine itself). The output as always are the factors whose values are set
by the system—for example, the number of rpm, the amount of heat pro
duced, the efficiency of energy conversion, and the composition of the ex
haust gases. This is still a somewhat complex system, but we can begin to see
that relatively simple relationships between inputs and outputs can exist. For
example, the input-output relation between accelerator setting and rpm for a
gasoline engine is sufficiently direct and predictable for most of us to invoke
it daily. Notice, however, that the system is still sufficiently complex so that
the effect of that one input is still not restricted to the single output in which
we are interested—rpm. Other outputs such as heat produced, composition of
exhaust gas, and efficiency are also affected by the accelerator setting, even if
we are generally prepared to ignore the connection.
In the example of the automobile engine, we are beginning to reach the

stage at which the system is simple enough for us to start identifying basic
principles of experimenting. Let us go one stage further and consider the ex
ample of a simple pendulum. It, too, is a system but one that includes very
little other than the string, the mass at the end of the string, the supporting
stand, and the surrounding air. Furthermore, it has only two immediately
obvious inputs—the length of the string and the initial conditions according
to which the motion is started. The outputs, too, are few in number. Apart
CHAP 1. APPROACH TO LABORATORY WORK
from small, secondary effects, they include only the frequency of vibration
and the amplitude of oscillation. Last, the connection between the inputs and
outputs is relatively direct and reproducible. Altering the length of the pendu
lum's string offers few surprises when we measure the frequency of vibration.
Here is a system that is simple enough to allow the basic principles of
experimenting to be clearly visible. If we use it to develop expertise in
studying systems and evaluating their outputs, we shall acquire the compe
tence to tackle later the more important but more complex problems.
This gives us the key to at least some constructive uses of the intro
ductory physics laboratory. There is real point in working with a pendulum
but only if we view it properly. If we look at it as “just a pendulum,” our only
reaction will be total boredom. If, however, we view it as a system, just like a
supermarket, an electrical power system, or the national economy, but
differing from these only in that it is simple enough for us to understand it
relatively well, our battered old pendulum supplies excellent simulation of
the problems of the real world.
The introductory physics laboratory, therefore, can offer us the oppor

tunity to practice on simple systems and develop the expertise that we shall
require later in the real world with its important but complicated systems. We
must be careful, however, about the ways we practice on these simple
systems. We shall derive only very limited benefit if, for example, we restrict
ourselves to sets of instructions that tell us exactly how to do particular
experiments. The range of experimental situations in society is enormous. In
some areas, random fluctuation dominates, as in the biological sciences; in
others, measurement may be precise, as in astronomy, but control over the
subject matter is limited. If we are to learn to function independently within
this enormous range, it is necessary to identify general principles of ex
perimenting that can be applied later to any future subject matter or type of
experimenting. The remainder of this text is concerned with some of those
principles, and we assume henceforth that laboratory experiments will be
regarded as exercises to illustrate the principles.
It may now be obvious that many of the traditional procedures in in

troductory laboratories are inappropriate for our purpose. For example, we
must avoid thinking of an experiment as a procedure to reproduce some
“correct” answer, deviation from which makes us “wrong.” Instead, we
should simply assess the properties of our particular system dispassionately
and take the results as they come. Also, there is no point in seeking some
“procedure” to follow: that is nothing more than asking someone else to tell
us how to do the experiment. In real life we rarely find someone waiting to
tell us what to do or what our result should be; our usefulness depends on our
8 APPROACH TO LABORATORY WORK CHAP. 1
ability to make our own decisions about how to handle a situation. It takes a
great deal of practice and experience to develop confidence in our own de
cisions about experiment procedures, and the introductory laboratory is not
too early to start. We therefore place a great deal of emphasis on experiment
planning, for this is the stage at which much of the skill in experimenting is
needed. It is important to avoid the temptation to regard preliminary planning
as a waste of time or a distraction from the supposedly more important task
of making the measurements. Time must be explicitly set aside for adequate
analysis and planning of the experiment before a start is made on the actual
measuring process.
It is also necessary to learn to work within the framework of the appa

ratus available. All professional experimenting is subject to limits on re
sources, and much of the skill in experimenting lies in optimizing the yield
from these resources. Restrictions on time, too, merely simulate the circum
stances of most actual experimenting. The apparatus itself is not always good
enough, but this should not be regarded as a defect but as a challenge, for this
aspect as before accurately simulates real life. Good evaluation of
experimental results always requires us to separate the grain of useful meas
urement from the chaff of error, uncertainty, and mistaken interpretation.
Experimenters must learn to identify sources of error or uncertainty for
themselves and, if possible, eliminate them or correct for them. Even with the
greatest care there is always an irreducible residuum of uncertainty, and it is
experimenters' responsibility to evaluate it accurately. The ability to cope
with such requirements can be acquired only by actual contact with realistic
working conditions. It is a common injustice to students in introductory
physics laboratories to provide apparatus that is too carefully adjusted or
manipulated, which can give students the impression that experiments always
give the “right” answer. This is unfortunate, because the foundation of future
expertise lies in learning how to respond constructively to the limitations of
experiments.
The most fruitful use of laboratory time results when the experiments
are accepted as problems that we must solve on our own. Certainly, errors in
judgment will be made, but students can learn more effectively when they see
the consequences of their decisions through direct, personal experience than
when they rigidly follow some established “correct” procedure. What is
learned from an experiment is more important than the production of some
supposedly “good” result. This is not to say that they should be complacently
indifferent to the outcome of the experiment. Development of experimenting
skills comes about only if the challenge of obtaining the best possible result
in every experiment is taken seriously.
The writing of laboratory reports should be tackled in the same con

structive spirit. In professional life there is very little point in spending time
and trouble on an experiment unless we can adequately convey the outcome
to others. We have an obligation to our readers to express ourselves as lu
cidly, if not elegantly, as possible. It is wrong to regard this as solely the re
sponsibility of our local Departments of English. Report writing in the in
troductory science laboratory is an opportunity for exercise in descriptive
composition. Report writing that degenerates into a mere indication that the
experiment has been performed is little more than a waste of time and a loss
of opportunity for necessary practice. Report writing at the level suggested
here is almost pointless without adequate review and criticism. Opportunities
for improvement become much more obvious in hindsight, and such detailed
review should be regarded as an indispensable part of the work in a teaching
laboratory.
Measurement
and Uncertainty
2 -1 BASIC NATURE QF MEASURING PROCESS
Measurement is the process of quantifying experience of the external world.

The nineteenth-century Scottish scientist, Lord Kelvin, once said that “when
you can measure what you are speaking about and express it in numbers, you
know something about it; but, when you cannot measure it, when you cannot
express it in numbers, your knowledge is of a meager and unsatisfactory
kind; it may be the beginning of knowledge, but you have scarcely in your
thoughts advanced to the stage of science.” Although this may be a slight
overstatement, it remains true that measurements constitute one of the indis
pensable ingredients of experimenting. We cannot reach a satisfactory level
of competence in experimenting without knowledge of the nature of meas
urement and the significance of statements about measurements.
It is obvious that the quantifying process (almost invariably involves

comparison with some reference quantity (how many paces wide is my
backyard?). It is equally obvious that the good order of society requires ex
tensive agreement about the choice of these reference quantities. Such meas
urement standards, defined by legislation and subject to international agree
ment, are extensive and important. No one seriously interested in measure
ment can ignore defining and realizing such standards in his or her area of
work. A discussion of this important topic here would distract us from our
10
CHAP. 2 MEASUREMENT AND UNCERTAINTY 11
chief concern, the actual process of measuring, so we leave the topic of stan
dards without further mention (except reference to the texts listed in the Bib
liography), and take up the study of measuring processes.
We start at the most basic level with an apparently simple measurement

and use it to find out what kind of process is involved and what kind of
statement can be made about its outcome. If I give someone my three-ring
binder containing this text with the request to measure its length with a meter
stick, the answer is absolutely invariable: The length of the notebook is 29.5
cm. But that answer must make us wonder: Are we really being asked to be
lieve that the length of the book is exactly 29.50000000 cm? Surely not;
such a claim is clearly beyond the bounds of credibility. So how are we to
interpret the answer? A moment's thought in the presence of the notebook
and a meter stick makes us realize that, far from determining the “right” or
“exact” value, the only thing we can realistically do is to approach the edge
of the notebook along the scale, saying to ourselves as we go: Am I sure the
answer lies below 30 cm? Below 29.9 cm? Below 29.8 cm? The answer to
each of these questions will undoubtedly be: Yes. As we progress along the
scale, however, we eventually reach a point at which we can no longer give
the same confident reply. At that point we must stop and we identify one end
of an interval that will become our measured value. In a similar way we can
approach the edge of the notebook from below, asking ourselves at each
stage: Am I sure that the answer lies above 29.0 cm? Above 29.1 cm? And so
on. Once again, we reach a value at which we must stop, because we can no
longer say with confidence that the answer lies above it. By the combination
of these two processes we identify an interval along the scale. It is the
smallest interval that, as far as we can be certain, does contain the desired
value; within the interval, however, we do not know where the answer lies.
Such is the only realistic outcome of a measuring process. We cannot look
for exact answers; we must be content with measured values that take the
form of intervals. Not only does this example illustrate the essential nature of
a measuring process, it also provides guidance for actually making meas
urements. The process of approaching from each side separately the value we
seek is a reminder of the necessity of stating the result as an interval, and it
also makes it easier to identify the edges of that interval.
The final outcome of our discussion is most important. As we make

measurements and report the results, we must constantly keep in mind this
fundamental and essential point—measurements are not exact, single
numbers but consist of intervals, within which we are confident that our
desired value lies. The act of measuring requires us to determine both the
location and width of this interval, and we do it by the careful exercise of
visual judgment every time we make a measurement. There are no simple
12 MEASUREMENT AND UNCERTAINTY CHAP. 2
rules for determining the size of the interval; that depends on many factors in
the measuring process. The type of measurement, the fineness of the scale,
our visual acuity, the lighting conditions—all play a part in determining the
width of the measurement interval. The width, therefore, must be determined
explicitly each time a measurement is made. For example, it is a common
error to believe that, when a measurement is made using a divided scale, the
“reading error” is automatically one-half of the finest scale division. This is
an erroneous oversimplification. A finely divided scale used to measure an
object with ill-defined edges can give a measurement interval as large as
several of the finest scale divisions. A well-defined object and good viewing
conditions, on the other hand, may permit the identification of a measurement
interval well within the finest scale division. Every situation must be assessed
individually.
2 - 2 DIGITAL DISPLAY AND ROUNDING OFF
Other aspects may confuse the issue. Consider, for example, a piece of
equipment that gives a digital readout. If a digital voltmeter shows that a
certain potential difference is 15.4 V, does that readout imply that the value is
15.40000 ... exactly? Clearly not, but what does it mean? That depends on
circumstances. If the instrument is made in such a way that it reads 15.4 V
because the actual value is closer to 15.4 than it is to 15.3 or 15.5, then the
meaning is: This reading lies between 15.35 V and 15.45 V. On the other
hand, a digital clock may be made in such a way that it changes its indication
from 09.00 to 09.01 at the time of 09.01. When we see it reading 09.00, we
know that the time lies between 09.00 and 09.01, a slightly different interpre
tation from that appropriate to the digital voltmeter. Again, each situation
must be judged by itself
These two examples of digital display illustrate a more general con

cept—the inaccuracy inherent in the process of rounding off Even without
the uncertainty that arises from limited ability to make measurements, the
mere statement of a numerical quantity can contain uncertainty. Consider the
statement
71 = 3.14
We all know that this is not so because we can remember some of the num
bers that follow: 3.14159 ... and so on. So what can we mean by quoting n as
3.14? Presumably, we mean only that n has a value closer to 3.14 than it does
to 3.13 or 3.15. Our statement, therefore, can be translated to read—n lies
between 3.135 and 3.145. This range of possibility represents what is
sometimes known as a rounding-off error. The effect of such errors can be
small and unimportant, or they can become significant. In a long calculation,

there is a chance that rounding-off errors either can accumulate, or they can
be important in other ways. For example, the calculation may require us to
find the difference between two large calculated values. If these two calcu
lated values are close together, the result that we need may be greatly affected
by premature rounding off Because calculators make accurate calculation so
easy, it is always advisable to carry the calculation through more figures than
one might initially think would be necessary. We can always do appropriate
rounding off at the end of the calculation.
A similar rounding-off error can appear in statements about measure

ment. We sometimes hear that someone has made a measurement on a scale
that was “read to the nearest millimeter” or some such phrase. This is not a
very good way of reporting a measurement because it obscures the actual
value of the measurement interval. When we encounter such statements, we
can only assume that the quoted scale division represents some kind of
minimum value for the size of the measurement interval.
2 - 3 ABSOLUTE AND RELATIVE UNCERTAINTY
Whatever the means by which we make a measurement, the final outcome

should be the interval that best represents the range inside which the desired
value lies. In the example we used first, the experimenter might be able to
state with confidence nothing more than that, as a consequence of a certain
measuring process, the length of the notebook lay between 29.4 and 29.6 cm.
This is a perfectly satisfactory and accepted way to quote the result. Although
such a statement corresponds with the reality of the measuring process, it is
frequently desirable to rephrase the quoted value. Take the interval 29.4 to
29.6 cm and rename it 29.5 ± O.l cm. Although this expression is no more
than a renamed statement of the original interval, the new form does offer
certain advantages. It gives us a central value, 29.5 cm, which can be used in
further calculations. It also gives us a value, ±0.l cm, that we call the
uncertainty of the measurement. First, the magnitude of the uncertainty en
ables us to judge the quality of the measuring process. Second, we can use
this numerical measure of the uncertainty in continued calculations on uncer
tainties. One disadvantage of this mode of expression is the return to a central
value, 29.5 cm. Unless we remember clearly that only the complete quantity,
29.5 ± O.l cm, serves as an adequate statement of the answer, we may
become sloppy in making and reporting measurements and may forget the
essential presence of the uncertainty. It should be an invariable practice to
associate an uncertainty value with a reading, both at the time of the
measurement and subsequently, whenever the value is quoted or used in

further calculation.
Because the figure ±0.1 cm represents the actual amount, or range, by
which the reading of 29.5 cm is uncertain, it is often called the absolute un
certainty of the measurement, and we consistently use this terminology. For
the purpose of perceiving the significance of the uncertainty, however, it is
frequently convenient to extend the definition of uncertainty. How significant
is an uncertainty of ±0.1 cm? When we measure the length of a notebook, it
is significant to a certain extent. When we measure the distance between
two cities, an uncertainty of ±0.l cm surely is completely insignificant. At
the other end of the scale, however, if we measure the size of a microscopic
bacterium, an uncertainty of ±0 .l cm clearly makes the measurement
meaningless. Obviously, the significance of a particular uncertainty value
depends on the magnitude of the measurement itself For this reason, it is
frequently desirable to compare an uncertainty figure with the actual value of
the measurement. For this purpose, we define a quantity called the relative
uncertainty of the measurement. It is defined by
Absolute Uncertainty
Relative Uncertainty =
Measured Value
In the case of our example
± 0 .l
Relative Uncertainty = — ^ = ±0.003
29.5
This relative uncertainty is often quoted as a percentage, so that in the present
case the relative uncertainty is ±0.3%. Such a quantity gives us a much better
feeling for the quality of the measurement, and we often call it the precision
of the measurement. The absolute uncertainty has the same dimensions and
units as the basic measurement (29.5 cm is uncertain by O.l cm), whereas the
relative uncertainty, being a ratio, has neither dimensions nor units and is a
pure number.
2 - 4 SYSTEMATIC ERRORS
The kind of uncertainty that we have been considering arises from natu
rally occurring limitations in the measuring process. A different type of
error can appear when something affects all the measurements of a series
in an equal or a consistent way. For example, a voltmeter or a micrometer
caliper can have a zero error, a wooden meter stick may have shrunk, a
stopwatch may be running fast or slow, and so on. These errors are
termed systematic errors, a subclass of which are calibration errors.
Because such systematic errors may not be immediately visible as one
makes a measurement, it is necessary to be vigilant and remember at all

times the possibility of their presence. Instrument zeroes, for example,
should automatically be checked every time an instrument is used. Al
though it may be less easy to check calibration, the accuracy of electrical
meters, timing devices, thermometers, and other such instruments should
not be taken for granted and should be checked whenever possible. Also,
the presence on an instrument of a precise-looking, digital readout with
four or five supposedly significant figures should not be taken as proof of
precision and freedom from systematic error. Most of a batch of elec
tronic timers that our laboratory once acquired for laboratory teaching,
which could supposedly measure time intervals with millisecond accu
racy, turned out to have calibration errors as large as 14%. Do not be de
ceived; view all measuring instruments with suspicion and check instru
ment calibration whenever possible.
2 - 5 UNCERTAINTY IN CALCULATED QUANTITIES
The preceding sections have been concerned solely with the concept of
uncertainty in a single measurement. It is rare, however, that a single
measurement ends the process. Almost invariably the result we desire is a
combination of two or more measured quantities or is at least a calculated
function of a single measurement. We might wish, for example, to
calculate the cross-sectional area of a cylinder from a measurement of its
diameter, or to calculate its volume from measurements of both diameter
and length. The various measurements will sometimes be of different
types, as in a calculation of g from values of the length and period of a
pendulum. In all such cases the presence of uncertainty in the basic meas
urements obviously entails the presence of uncertainty in the final com
puted value. It is this final uncertainty that we now wish to calculate. For
the purposes of this section we assume that our uncertainties have the
character of ranges or intervals within which we are “almost certain” that
our answer lies. For the computed values we calculate intervals within
which we again wish to be “almost certain” that our answer lies. That
means that we must do our calculation for the “worst case” of combined
uncertainties in which the deviations in the various measured quantities
happened to occur in such directions as to reinforce each other. This is
perhaps a pessimistic assumption, and we see later in Chapter 3 how the
probabilities that are associated with various error combinations enable us
to make a more realistic and less pessimistic estimate. For the moment,
assume that we wish to calculate from the uncertainties in the primary
values the maximum range of possibility for the computed answer.
2 - 6 UNCERTAINTY IN FUNCTIONS OF ONE VARIABLE ONLY
Consider a measured quantity with an uncertainty ±3x (where we are

using finite differences such as 8x to represent absolute uncertainties in the
corresponding variable x, etc.), and consider a computed result z to be some
function of the variable x. Let
z = f{x)
The function /enables us to calculate the required value Zq from a measured
value Xq. Moreover, the possibility that x can range from Xq- 5x to Xq+ 6x
implies a range of possible values of z that range from Zq- 8z to Zq + 8z,
where 5z is the value of the absolute uncertainty in z. We now wish to calcu
late the value of 8z. The situation is illustrated graphically in Figure 2-1, in
which, for a given /(x), we can see how the measured value Xq gives rise to
the computed result Zq, and how the range ±5x about Xq produces a corre
sponding range ±5z about Zq.
Before considering general methods of evaluating 8z it is instructive to
see how finite perturbations are propagated in simple functions. Consider, for
example, the function
z = x*"
Figure 2-1 Propagation o f uncertainty from one variable to another.

If X can range between X(, - 8x and Xq+ 8x , then z can range between
Zfl “ 8z and Zq+ 6z, where
z „ ± 8z = (x„±5x)^
= Xg ± 2 x„8x + ( 8x)^
We can ignore (Sx)^, since 5x is assumed to be small in comparison with Xq,
and equate Zq to x^, giving for the value of 5z
8z = 2 xq5x
This can more conveniently be expressed in terms of the relative uncertainty
8z/zo
8z _ 2xq8x _ 2
Zo ^0 ^0
Thus, the relative uncertainty of the computed result is twice that of the initial
measurement.
Although it is helpful to bear in mind the nature of propagated uncer

tainty, as illustrated by the use of finite differences, considerable simplifica
tion of the formulation can be achieved using differential calculus.
2 - 7 GENERAL METHOD FOR UNCERTAINTY

IN FUNCTIONS OF A SINGLE VARIABLE
The finite differences 8z and 8x that were used in the preceding section could
be regarded as components of the derivative dz! dx. We can therefore obtain
our value of 8z by first using standard techniques to obtain d z ! dx in the
form
dz ^ d { f{ x ))
dx dx
and then writing
dx (2- 1)
This is a relatively simple procedure, and it works well in cases for which the
elementary, finite-difference approach would lead to algebraic complexity. If,
for example we have to deal with the function
z=■
(x^ + l)
then
dz _x~ + \ - x x 2 x
l-jc '
and
\-x ^
5z = 8x
This calculation would have been very awkward if any other approach had
been used. Furthermore, it gives 8z generally as a function of x and 8x ; any
particular desired value can be obtained by setting x = x^. We now use this
technique to evaluate uncertainties for some common functions.
Powers
Consider the function
z= r'
Differentiating and replacing the derivative by finite differences gives us
dz
= nx
dx
6z = nx'’ ' 8jc
The significance of this result becomes a little more obvious when ex
pressed in terms of the relative uncertainty. Thus,
5z _ 5x
z X
Thus, when evaluating powers of a measured quantity, we must compute the
uncertainty in the final answer using the relative uncertainty of the measured
quantity. The relative uncertainty in the final answer is the relative uncer
tainty of the basic quantity multiplied by the power involved. This method is
valid for either powers or roots, so that precision diminishes as a quantity is
raised to powers and improves on taking roots. This situation must be care
fully watched in an experiment in which powers are involved. The higher the
power, the greater is the need for good initial precision.
Trigonometric Functions
We give only one example, because all the others can be treated in similar
ways. Consider
z = smx
Here
dz
■cosx
dx
and
8z = (cosx) 8x
This is one case where the elementary method of inserting jco± 8x in the
function shows the result more clearly. We obtain
Zo ±8z = sin(Xo ±8x)

= sin os 8
Xqccos
sinjCo jc± cos Xqsin 8.;c
8 jc
Since Zq = sinjCo and cos8jc « I, this becomes

8z : cosjCqsin 8x
This result makes it clear that the dx in the original expression was really an
approximate form of sin 8jc. Only in the case of very large uncertainty would
this difference be significant, but it is best to be aware of the situation. For
one thing, it allows us to understand that, when we use that original expres
sion, we must express the uncertainty in angle,8x, in radian measure, becau
se only then can we replace sin 8x by dx itself. Such uncertainty calculations
using trigonometric functions normally have straightforward application when
dealing with apparatus such as spectrometers.
Logarithmic and Exponential Functions
Consider the function

z = logx
Here
dz _ I
dx X
and
5z = —8x
JC
and the relative uncertainty can be calculated as usual.
If
z=e
dz
=e
dx
and
8z = e"8x
This is an important case, because exponential functions occur frequently in
science and engineering. These functions can become very sensitive to the
exponent when it takes values much over unity, and the uncertainty 8z may
become very large.
As stated earlier, the method can be easily applied to any function not
listed above by evaluating the appropriate derivative and using Equation (2-
1).
2 - 8 UNCERTAINTY IN FUNCTIONS OF TW O OR MORE VARIABLES
If the result is to be computed from two or more measured quantities, jc, y,

and so on, the uncertainty in the result can, as was mentioned in Section 2-5,
be regarded in two ways. We could be as pessimistic as possible and suppose
that the actual deviations of x and y happen to combine in such a way as to
drive the value of z as far as possible from the central value. In this way, we
would calculate the value for 8z that gives the extreme width of the range of
possible z values. On the other hand, we could argue that it is more probable
for the uncertainties in the basic measurements to combine in a less extreme
way, some making positive contributions to 8z and some negative, so that the
resulting 8z would be smaller than for the pessimistic assumption. This
argument is valid, and later we deal with the question of probable uncertainty
in computed quantities. For the moment, however, we calculate the value of
8z that represents the widest range of possibility for z. Such an approach, if
pessimistic, is certainly safe, because, if 8jc, 8y, and so forth, represent limits
within which we are “almost certain” the actual values lie, then the calculated
8z will give those limits within which we are equally certain that the actual
value of z lies.
The most instructive initial approach uses the elementary substitution

method, and we use this for the first two functions.
Sum of Two or More Variables
Consider the function of two variables

z = x+y
The uncertainty in z will be obtained from
Zo ± 8z = (Xo ± + (>'o ± 87)

and the maximum value of 5z is obtained by choosing similar signs through
out the right-hand side of the expression. Thus,
8z = 8x -I- 8y
As might be expected, the uncertainty in the sum is just the sum of the indi
vidual uncertainties. This can be expressed in terms of the relative uncer
tainty
8z _ 8x -f 8y
z x-\- y
but no increased clarification is achieved. If the quantity z contains the sum
of more than two variables, the expression for the uncertainty in z can obvi
ously be extended as necessary.
Difference of Two Variables
Consider a quantity that must be calculated as the difference between two

measured values. Let
z = X -y
As in the preceding case, 8z is obtained from
Zo ± 8z = (Xo ± dx) - (>^0 ± 8>^)

Here, however, we can obtain the maximum value of 8z by choosing the
negative sign for 5y, giving again,
8z = 8x -f- 8y
We can see from this equation that, when Xq and yo are close together and the
value of x - y is small, the relative uncertainty can rise to very large values. This
is at best an unsatisfactory situation, and the precision can be low enough to
destroy the value of the measurement. The condition is particularly hazardous
because it can arise unnoticed. It is perfectly obvious that, if it were possible to

avoid it, no one would attempt to measure the length of my notebook by measur
ing the distance of each edge from a point a mile away and then subtracting the
two lengths. However, a desired result can be obtained by subtraction of two
measurements made separately (two thermometers, clocks, etc.), and the char
acter of the measurement as a difference may not be strikingly obvious. Conse
quently, treat all measurements involving differences with the greatest caution.
Clearly, the way to avoid the difficulty is to measure the difference directly,
rather than obtaining it by subtraction between two measured quantities. For
example, if you have an apparatus within which two points are at potentials
above ground of V\ = 1500 V and V2 = 1510 V, respectively, and the required
quantity is V2 - Vi, only a voltmeter of very high quality would permit the values
of Vi and V2 to be measured with the exactness required to achieve even 10%
precision in V2 - V\. On the other hand, an ordinary 10 V table voltmeter, con
nected between the two points and measuring V2 - V\ directly, immediately gives
the desired result with 2% or 3% precision.
2 - 9 GENERAL METHOD FOR UNCERTAINTY IN FUNCTIONS

OF TW O OR MORE VARIABLES
The last two examples, treated by the elementary method, suggest that the
differential calculus may offer considerable simplification of the treatment. It
is clear that if z is a function of the two variables x and y,
z = f{x^y)
the appropriate quantity for calculating 5z is the total differential d z . This is
given by
^ 5/ ^ 5/ ,
dz = — dx-\-— dy
dx dy ( 2- 2)
We treat this differential as a finite difference 8z that can be calculated from

the uncertainties 8x and dy. Thus,
s: 5/ ^ df ^
5z = — 8jc -F — 5y
dx dy
and the derivatives df / dx and df / dy are normally evaluated for the values,
Xq and yo, at which 8z is required. We may find that, depending on the func
tion /th e sign of df / dx or df / dy turns out to be negative. In this case, using
our pessimistic requirement for the maximum value of 8z, we choose
negative values for the appropriate 8jc or 8y, obtaining thereby a wholly
positive contribution to the sum.
Product of Two or More Variables
Suppose
z = xy
To use Equation (2-2) we need the values of dz / dx and dz I dy. They are
dz dz
- y and dy = X
dx
Thus, the value of 5z is given by
6z = y8x + xby
The significance of this result is more clearly seen when it is converted to the
relative uncertainty
8z _ 5x ^ 3y
z X y
Thus, when the desired quantity is a product of two variables, its relative uncer
tainty is the sum of the relative uncertainties of the components. Notice the con
trast with the result for uncertainty in the case of two added variables, where we
must combine the uncertainties using the absolute uncertainties.
The most general case of a compound function, very commonly found

in physics, involves an algebraic product or quotient that has components
raised to powers. Consider the function
z=Xy
a h
where a and b may be positive or negative, integral or fractional. This formu

lation is greatly simplified by taking logarithms of both sides before differ
entiating. Thus,
logz = alogx-\-b log y
whence, differentiating implicitly,
dz dx , dy
— = a — + b—
z X y
As usual, we take the differentials to be finite differences and obtain
5z dx , dy
— =a — -hi —
z X y
MEASUREMENT AND UNCERTAINTY CHAP. 2
If the original expression for z contains more than two variables, we can
simply extend the result for 8z by adding terms as appropriate. This process
gives the relative uncertainty directly, which is frequently convenient. If the
absolute uncertainty 8z is required, it can be evaluated simply by multiplying
the relative uncertainty by the computed value Zq, which is normally avail
able. This form of implicit differentiation still offers the simplest procedure even
when z itself is raised to some power. For example, if the equation reads
z = xy
it is unnecessary to rewrite it
and work from there, because, by taking logs

21ogz = logjc-i-log>^
whence
_8x
z X y
giving 8z / z as required.
Quotients
Quotients can be treated as products in which some of the powers are nega
tive. As before, the maximum value of 8z is obtained by neglecting negative
signs in the differential and combining all the terms additively.
If a function other than those already listed is encountered, some kind

of differentiation usually works. It is frequently convenient to differentiate an
equation implicitly, thereby avoiding the requirement to calculate the un
known quantity explicitly as a function of the other variables. For example,
consider the thin-lens equation. If we had made measurements of the object
distance 5- and the image distance s' for a thin lens with the intention of cal
culating a value for its focal length, /, the equation we would use is
1-1 1
/ ^
To obtain the uncertainty in /, we can differentiate the equation imphcitly to
obtain
df _ ds ds’
■yy-
5^ 5'^
It is now possible to calculate d f ! f directly and more easily than by writing
/explicitly as a function of 5 and s' , and differentiating. In this way we can
prepare a formula for the uncertainty into which all the unknowns can be in
serted directly. Make sure that appropriate signs are used so that all contri
butions to the uncertainty add positively to give outer limits of possibility for
the answer.
If the function is so big and complicated that we cannot obtain a value

for 5z in general, we can always take the measured values, Xq, and so on,
and work out Zq. We can then work out two different answers, one using the
actual numerical values of Xq+ 8x , >ô + (or yo ~ ^y if appropriate), and so
forth to give one of the outer values of z, and the other using Xq- 5x, and so
on. These two values correspond to the outer limits on z, and we know the
value of 5z.
2 -1 0 COMPENSATING ERRORS
A special situation can appear when compound variables are involved. Con
sider, for example, the well-known relation for the angle of minimum devia
tion for a prism of refractive index n and vertical angle A:
sin + £)„,)
sin] A
If A and D,,, are measured variables with uncertainties 5A and the
quantity n will be the required answer, with an uncertainty 8 „,. It would be
fallacious, however, to calculate the uncertainty in A + D^^^, then in
sinY (^-}-D,„), and combine that with the uncertainty in s in i^ , as if the
function were a quotient of two independent variables. This can be seen by
thinking of the effect on n of an increase in A. Both siny(^ + D,„) and sin^^
increase, and the change in n is not correspondingly large. The error lies in
applying the methods of the preceding sections to variables, such as A and
A+ that are not independent. The cure is either to reduce the equation to
a form in which the variables are all independent or else to go back to first
principles and use Equation (2-2) directly. Cases that involve compensating
errors should be watched carefully, because if they are treated incorrectly,
they give rise to errors in uncertainty calculations that are hard to detect.
2 -1 1 SIGNIFICANT FIGURES
Because computations tend to produce answers consisting of long strings

of numbers, we must be careful to quote the final answer sensibly. If, for
example, we are given the voltage across a resistor as 15.4 ± 0.1 volts and
the current as 1.7 ± 0.1 amps, we can calculate a value for the resistance.
The ratio V/1 comes out on my calculator as 9.0588235 ohms. Is this the
answer? Clearly not. A brief calculation shows that the absolute
uncertainty in the resistance is close to 0.59 ohms. So, if the first two
places of decimals in the value for the resistance are uncertain, the rest
are clearly meaningless. A statement like /? = 9.0588235 ± 0.59 ohms is,
therefore, nonsense. We should quote our results in such a way that the
answer and its uncertainty are consistent, perhaps something like
i? = 9.06 ± 0.59 ohms. But is even this statement really valid? Remember
that the originally quoted uncertainties for V and I had the value ±0.1,
containing one significant figure. If we do not know these uncertainties
any more precisely than that, we have no right to claim two significant
figures for the uncertainty in R. Our final, valid, and self-consistent
statement is, therefore, /? = 9.1 ± 0.6 ohms. Only if we had a good reason
to believe that our original uncertainty was accurate to two significant
figures, could we lay claim to two significant figures in the final uncer
tainty and a correspondingly more precisely quoted value for R. In gen
eral terms, we must make sure that our quoted values for uncertainty are
consistent with the precision of the basic uncertainties, and that the num
ber of quoted figures in the final answer is consistent with the uncertainty
of that final answer. We must avoid statements like z = 1.234567±0.1 or
z = 1.2 ± 0 . 000001.
PROBLEMS
1. I use my meter stick to measure the length of my desk. I am sure that the
length is not less than 142.3 cm and not more than 142.6 cm. State this
measurement as a central value ± uncertainty. What is the relative uncer
tainty of the measurement?
2. I read a needle-and-scale voltmeter and ammeter and assess the range of
uncertainty visually. I am sure the ammeter reading lies between 1.24
and 1.25 A and the voltmeter reading between 3.2 and 3.4 V. Express
each reading as a central value ± uncertainty and evaluate the relative
uncertainty of each measurement.
3. My digital watch gives a time reading as 09:46. What is the absolute un
certainty of the measurement?
4. If I can read a meter stick with absolute uncertainty ±lmm, what is the
shortest distance that I can measure if the relative uncertainty is not to
exceed (a) 1%, (b) 5%?
5. I use a thermometer graduated in fifths of a degree Celsius to measure out
side air temperature. Measured to the nearest fifth degree, yesterday's tem
perature was 22.4° Celsius and today's is 24.8° Celsius. What is the relative
uncertainty in the temperature difference between yesterday and today?
6. The clock in the lab has a seconds hand that moves in one-second steps.
I use it to measure a certain time interval. At the beginning of the inter
val it reads 09:15:22 (hours:minutes:seconds), and at the end it reads
09:18:16. What is the relative uncertainty of the measured time interval?
7. For the desk mentioned in Problem 1 ,1 measure the width, and I am sure
the measurement lies between 78.2 cm and 78.4 cm. What is the
absolute uncertainty of the calculated area of the desktop?
8. In measuring the resistance of a resistor, the voltmeter reading was 15.2
± 0.2 V and the ammeter reading was 2.6 ±0.1 A. What is the absolute
uncertainty of the resistance calculated using the equation R = V / H
9. A simple pendulum is used to measure the acceleration of gravity using
[7
T = 2n -
The period T was measured to be 1.24 ± 0.02 s and the length to be

0.381 ± 0.002 m. What is the resulting value for g with its absolute and
relative uncertainty?
10 . An experiment to measure the density, J, of a cylindrical object uses the
equation
m
nr
where
m = mass = 0.029 ± 0.005 kg

r - radius = 8.2 ± 0 . 1 mm
^ - length = 15.4 ± 0.1 mm
What is the absolute uncertainty of the calculated value of the density?
MEASUREMENT AND UNCERTAINTY CHAP. 2
11. The focal length, /, of a thin lens is to be measured using the equation
1-1 1
f s s'
where
s = object distance = 0.154 ± 0.002 m
s' - image distance = 0.382 ± 0.002 m
What is the calculated value for focal length, its absolute uncertainty,
and its relative uncertainty?
12. A diffraction grating is used to measure the wavelength of light using
the equation
(isinO = X
The value of 0 is measured to be 13° 34' ± 2'. Assuming that the value
of d is 1420x 10"^ m and that its uncertainty can be ignored, what are
the absolute and relative uncertainties in the value of ^ ?
13. A value is quoted as 14.253 ±0.1. Rewrite it with the appropriate num
ber of significant figures. If the value is quoted as 14.253 ± 0.15, how
should it be written?
14. A value is quoted as 6.74914 ± 0.5%. State it as a value ± absolute un
certainty, both with the appropriate number of significant figures.
Statistics of Observation
3 -1 STATISTICAL UNCERTAINTY
In the preceding chapter we considered measurements in which the uncer

tainty could be estimated by personal judgment. In these, supposing that we
have judged the situation accurately, repeated measurements should give
consistent answers. Sometimes, however, systems behave in a different way
and repeated measuring gives clearly different answers. For example, if we
are using a particle detection and counting system to measure the activity of a
radioactive source and we decide, with given geometry, to obtain the number
of counts in a 10-second interval, we find that the results obtained by
counting in successive 10-second intervals are not the same. We can encoun
ter the same situation in measurements that involve visual judgment. If, for
example, we wish to find the image formed by a thin lens, we may be unable
to judge the position of the image accurately enough to obtain repeatedly the
same reading on a good, finely divided distance scale. In many other systems,
too, measurements show random fluctuation. Whether the fluctuation is
intrinsic to the system under investigation (as in the radioactive source, where
it arises from the basic nature of radioactive decay) or whether the variation
arises from difficulty we have in making the measurement, we must find out
how to make sensible statements about measured values that show such
fluctuation.
What kind of statement will it be possible to make? No longer is it

possible for us to make such statements as we made earlier. They had the
form “I am virtually certain that the answer lies within the interval __ ”
29
30 STATISTICS OF OBSERVATION CHAP. 3
Now, when measured values appear seemingly at random along a scale,

we are unable to identify the edges of an interval within which we can be
almost certain our answer lies. In fact, apart altogether from the impos
sibility of obtaining “right” answers, we find that the difficulty lies not so
much in constructing sensible answers as in knowing the sensible ques
tions to ask. We discover that the only sensible questions involve, as be
fore, intervals along the scale of values. This time, however, we interpret
these intervals in terms of probability rather than certainty. Our search for
a solution is fairly lengthy, but at the end the answer turns out to be sim
ple and elegant.
To start the search, we go back to the basic situation. Assume that

we have made a single measurement and that we have made the meas
urement a second time to check our work. This time we obtain a different
value. What are we supposed to do? We have no way of saying that one
value is “right” and the other “wrong.” Which one should we choose to be
“right”? In response to this ambiguity the natural reaction iws to try a
third time, hoping, perhaps, that the third reading will confirm one or
other of the first two. Very likely it will not be so obliging and will sim
ply add to the confusion by supplying a third possibility. Faced with
growing complexity, we can decide to keep on making measurements to
see what happens. Suppose that our curiosity has prompted us to make a
substantial number of repeated measurements, say 100, and we now ask:
What is the answer? As was mentioned earlier, it is more significant to
ask: What is the question? That depends very much on the use to which
we wish to put the measurements. If we are measuring the position of an
optical image, we may need a value that we can use in the design of some
piece of optical equipment. If we are measuring the activity of a radioac
tive source, we may wish to make a guess at the number of counts that
will be observed in a certain 10-second interval tomorrow. A sociologist
counting political opinions wishes to predict the outcome of the next
election, and so on. There is no single question and no unique answer.
The treatment we give our fluctuating numbers depends on circum
stances. We now consider some of the possibilities.
3 -2 HISTOGRAMS AND DISTRIBUTIONS
Assume that we have made 100 measurements of some quantity and that
we must now report our results. The first response to the question—What
did you obtain?—is the rather feeble reply: “I made the measurement 100
times and here are the 100 answers.” This statement may be clear and free
from the possibility of misinterpretation, but it is hardly helpful. Our
CHAP. 3 STATISTICS OF OBSERVATION 31
audience will find it difficult to make any sense out of a plain list of
numbers, and questions will naturally arise, such as: Is there anything
systematic about the numbers? Are there any regularities? Do any appear
more frequently than others? And so on. To show the characteristics of
the measurements more clearly, some kind of graphic display would
clearly be helpful.
One common mode of presentation is the histogram. To construct this

diagram, divide the scale along which the measurements are spread into in
tervals, and count the readings that fall within each interval. Then plot these
numbers on a vertical scale against the intervals themselves. It is conven
tional to use a bar diagram to indicate the number of readings; and the result
will be similar to Figure 3 - 1. At once we improve our comprehension of the
measurements enormously, because we can see at a glance how the values are
distributed along the scale. This distribution is the key to satisfactory in
terpretation of the measurements. Usually we find that the readings tend to
occur more frequently in the middle of the range If this is so and we are un
able to make any other sensible statement, we can always content ourselves
with the simple assertion that the observations have central tendency. This
may suffice, and when we have drawn the histogram we may be able to stop.
Many spreadsheet programs for computers such as Lotus I-2-3, Quat

troPro, and thge like, have built-in programs for calculating and displaying
frequency distributions and histograms. These allow convenient and rapid
statistical analysis of sets of observations, and it is good to use any available
opportunity to practice with them. Their very convenience, however, can be
deceptive, and it is very important to know by personal experience what kind
of calculation the machine is doing. For this purpose start by first doing a few
calculations by hand to see the way the numbers work. The examples at the
end of this chapter can be used for this purpose.
Many results from measuring processes are presented simply by offer

ing the histogram of the observations; readers can view the distribution and
draw their own conclusions.
Value
Figure 3-1 A set of obervations and its histogram.
85 109 114 121 127 131

92 109 114 121 127 132
96 no 114 122 127 133
97 no 115 122 127 134
97 111 116 122 128 134
97 111 116 122 128 134
100 111 116 122 128 134
101 111 117 123 128 135
101 111 117 123 128 136
102 112 118 123 128 137
102 112 118 123 130 137
103 112 119 123 130 137
103 113 119 124 130 144
105 113 120 124 130 148
106 113 120 124 130 149
106 113 120 125 130
107 113 120 125 131
108 113 121 125 131
108 114 121 126 131
3 -3 CENTRAL VALUES OF DISTRIBUTIONS
Frequently, however, we wish to go further. If, for example, we could find

some single number that could be used as a substitute for the whole distribu
tion, it might simplify the reporting of our results. As candidates for a single
number to represent the distribution as a whole, there are several possibilities,
and we choose one or another, depending on the future use of the infor
mation. The various possibilities are mode, median, and mean.
Mode
Many distributions have a peak near the center. If the peak is well defined,
the value on the horizontal scale at which it occurs is called the mode of the
distribution. Whenever we wish to draw attention to such central concentra
tion in our measured values, we quote the modal value. Sometimes a dis
tribution shows two peaks; we call it a bimodal distribution and quote the
two modal values.
Median
If we place all our readings in numerical order and divide the set into two
equal parts, each containing the same number of readings, the value at which
the dividing line comes is called the median. Because it is obvious that areas
under distribution graphs represent numbers of observations (the left-hand
bar in Figure 3-1 represents 5 observations, the second from the left rep
resents 9, so that the two together represent 14, and so on), the median is that
value at which a vertical line divides the distribution into two parts of equal
area. The median is frequently quoted in sociological work; people talk about
median salaries for certain groups of employees, for example.
Mean
The third of the commonly quoted numbers is the familiar arithmetic average,
or mean. For a group of N observations, x., the mean x is defined by
X= (3-1)
We shall discover that for our purposes the mean is the most useful of the
three quantities we have defined.
Notice that for a symmetrical distribution the mean, median, and mode
all coincide at the center of the distribution. On the other hand, if the distri
bution is not symmetrical, each has a separate value. For the histogram
shown in Figure 3-1, the values of the mean, median, and mode are shown in
Figure 3-2, which illustrates their relationship to the distribution as a whole.
If the distribution is markedly asymmetric, the difference between the

mode, median, and mean can be substantial. Consider, for example, the distribu
tion of family income in a country. The presence of the very wealthy, although
they are relatively few in number, has an effect on the mean that counterbalances
many members of the population at the low end of the salary scale. The mode
34 STATISTICS OF OBSERVATIDN CHAP. 3
30
20
10
20 40 60 80 100 120
Q
140
J
160
Mean 119 II
Median 120 I
Mode 125
Figure 3-2 The relationship between a histogram and its mean, median, and mode.
and the mean thus differ substantially. This example illustrates the care required
in interpreting quoted statistics; people who quote statistics frequently do so in
the way that best suits their particular purpose.
3 -4 THE BREADTH OF DISTRIBUTIONS
Let us now turn to the question: To what extent is our chosen number represen
tative of the distribution as a whole? That is, how rehable is it to use a single
number as a substitute for a whole distribution? In answering that question, we
have at the present stage no justification to offer for the procedures that will be
described. We rely, instead on an intuitive feehng that narrow distributions give
us more confidence in the results than do broad distributions.
Let us, therefore, construct a quantity that is a measure of the breadth of

the distribution. We could invent many such quantities, but, for reasons that
need not concern us at the moment, we define a quantity that is almost
universally used. Consider a set of N measurements x-, of which the mean is
X. We define the standard deviation of the set of values, 5, to be
^ (x-x,r
5' = (3-2)
N
The definition is to some extent arbitrary, for in defining a measure of
the breadth of the distribution we could have chosen other powers to which
the quantity (x-Xj) could be raised, and we could have chosen other de
nominators. There are, however, reasons for these choices; these reasons and
the significance of the standard deviation will become clear shortly.
To calculate the standard deviation for a set of numbers, it is important

to interpret the quantities in Equation 3-2 correctly. Each of the quantities
x-Xj is the difference between the mean and an individual number in the set.
That difference may be positive or negative. When we square these dif
ferences, we obtain a series of positive numbers whose sum is dependent on
the breadth of the distribution and provides the numerator in the expression
for 5. Almost all spreadsheet programs for computers and many calculators
have built-in facilities for calculating standard deviations, and these greatly
reduce the effort required. As was mentioned earlier with regard to distribu
tion curves, however, it is important to become personally familiar with the
way the numbers work. The examples at the end of the chapter, therefore,
should be worked out by hand so that we may later know what calculators or
computers are doing.
We can pause at this stage to summarize the progress so far. If we have

made repeated measurements of a quantity and wish to state the result in
numerical terms, we can do a number of things: ( 1) we can show the histo
gram, (2) we can quote the mode, median, or mean as a measure of the loca
tion of the distribution, and (3) we can quote the standard deviation as a
measure of the breadth of the distribution. We sometimes leave the outcome
of a measuring process in this form; the quantities involved are universally
understood, and the procedure is acceptable.
For our present purpose, however, we seek more detailed, numerical

interpretation of the quoted values.
3 -5 SIGNIFICANCE OF THE MEAN AND STANDARD DEVIATION
In this and the following sections, for reasons that will become clear, we ig
nore the mode and median and restrict ourselves to numerical interpretation
of the mean and the standard deviation. Because the presence of random
fluctuation has denied us the opportunity to identify a realistic interval within
which we can feel certain our answer lies, we must alter our expectations of
the measuring process. As said before, it is not so much a matter of obtaining
sensible answers to questions as of knowing the sensible questions to ask.
Specifically, it is not sensible to ask: What is the right answer? It is not even
sensible to ask: Having made 100 observations of a quantity, what shall I
obtain when I make the measurement the next time? The only sensible ques
tions involve not certainty but probability, and several different questions
about probabilities are possible.
For example, we could ask: What is the probability that the 101st
reading will fall within a certain range on our scale of values? That is a sen
sible question, and sensible answers can easily be imagined. If, for example,
of our 100 original readings, a certain fraction of the values fell within some
particular range, we might feel justified in choosing that fraction as the prob
ability that the 101st observation will fall within that interval. This would not
be an unrealistic guess, and we could attempt a standardized description of
our distribution by quoting the fraction of the total number of readings that
fall within various specified intervals. This would satisfactorily convey in
formation about our set of readings to other people, but a major problem ap
pears when we discover that our answers for these probabilities are specific
to our particular histogram. If we were to make another series of 100 read
ings, holding all the conditions the same as they were before in the hope of
obtaining the same histogram, we would be disappointed. The new histogram
would not duplicate the first exactly. It might have similar general character
istics with respect to location and breadth, but its detailed structure would not
be the same as before, and we would obtain different answers to questions
about probabilities.
How, then, are we going to find answers to our questions that have
some kind of widely understood numerical significance? One solution is to
abandon the attempt to describe our particular histogram and to start talking
about defined theoretical distributions. These may have the disadvantage of
uncertain relevance to our particular set of observations, but there is the
enormous advantage that, because they are defined theoretical constructs,
they have properties that are definite, constant, and widely known. Many
such theoretical distributions have been constructed for special purposes, but
we restrict ourselves to one only, the Gaussian, or normal, distribution.
We use the Gaussian distribution to interpret many kinds of physical

measurement, partly because the mechanical circumstances of many physical
measurements are in close correspondence with the theoretical foundations of
the Gaussian distribution, and partly because experience has shown that
Gaussian statistics do provide a reasonably accurate description of many real
events. For only one common type of physical measurement is another
distribution more appropriate. In counting events like radioactive decay we
must use a distribution called the Poisson distribution, but even for it the
difference from Gaussian statistics becomes significant only at low counting
rates. Further information about Poisson statistics can be found in books de
scribing experimental methods in nuclear or high-energy physics. Apart from
these special cases, we can feel relatively confident that Gaussian statistics
can be usefully applied to most real measurements. Always remember,
however, that unless we actually test our measurements for correspondence
with the Gaussian distribution, we are making an assumption that Gaussian
statistics are applicable, and we should remain alert to any evidence that the
assumption may be invalid.
3 -6 GAUSSIAN DISTRIBUTIQNS AND SAMPLING
Even if to use it successfully we need not know very much about the origins
of the Gaussian distribution, it is interesting to understand why its axiomatic
foundations make it particularly relevant to many physical measurements.
The equation for the Gaussian distribution can be derived from the assump
tion that the total deviation of a measured quantity, x, from the unperturbed
value X that would be obtained in the absence of perturbing fluctuations is
the consequence of a large number of small fluctuations that occur randomly
in positive and negative directions. To construct a simple model of such a
situation, suppose that there are m such contributions to the total deviation,
each of equal magnitude a and equally likely to be positive or negative. Any
individual measured value, x, therefore differs from X by an amount that
contains a random number of positive contributions to the deviation, and a
corresponding number of negative contributions. If we repeat the measuring
process many times, therefore, we obtain a set of values that range from
X-\- ma for a measurement in which all the fluctuations happened to be
positive simultaneously, to X - m a if the same happened in the negative di
rection. Either of these possibilities is most unlikely. If perturbations occur
randomly, it is much more likely that they will occur in a mixture of positive and
negative contributions, and so the total measured quantity x is more likely to be

found near the middle of the range of possibihties than at the ends.
Such a situation, in which we have a random summation of positive and
negative quantities, is similar to the so-called random walk. In this, we depart
from a starting point by deciding to take a step backward or forward,
depending on whether a coin that we toss comes up heads or tails. It can eas
ily be demonstrated that our most probable location after any number of steps
is the position from which we started. It is the same with the measurements
that have been perturbed by random perturbations. The most probable sum of
the perturbations is zero, meaning that for repeated measurements, .x, the
most common values are in the vicinity of X. The distribution curve,
therefore, has a peak in the middle, is symmetrical, and declines smoothly to
zero at X -\-ma and X - m a . If this concept is taken to the limiting case in
which an infinite number of infinitesimal deviations contribute to the total
deviation, the curve has the form shown in Figure 3-3. Treating the curve
solely from the mathematical point of view for the moment, we can easily
prove that its equation can be written
y = Ce - h ^ { x - X f (3-3)
We need not be concerned about the derivation of this equation; for our pur
poses, it is sufficient to know the principles on which it is based and the re
sulting properties of the function. If we wish to know more about the origins
of the function, the full derivation is in Appendix 1.
In the equation the constant C is a measure of the height of the curve,
since y = C fox x = X at the center of the distribution. The curve is symmet
rical about x = X and approaches zero asymptotically. The quantity h obvi
ously governs the width of the curve, because it is only a multiplier on the x
scale. If h is large, the curve is narrow and high in relation to its width; if
small, the curve is low and broad. The quantity h clearly must be connected
with the standard deviation of the distribution. Now that we are dealing
with a mathematically defined theoretical distribution instead of a finite set of

real observations, we modify our terminology a little. We used the latin letter
S to denote the standard deviation of a finite set of real observations. We use
the greek letter a to represent the standard deviation that can be calculated
for a mathematically defined distribution, such as the Gaussian function. For
the Gaussian distribution, the relationship between its standard deviationa
and the geometrical measure of the width h is
1
a = (3 ^ )
4lh
Now that we have a definite equation for the distribution, all the original
ambiguity about interpreting the standard deviation in terms of probability
disappears. We now have definite, unique, and permanent values. For
example, the area enclosed within the interval X ± a for a Gaussian
distribution is 68% of the total area under the curve, and within the interval
X ± 2 a it is 95%, and this is so for all Gaussian distributions. The relation
between the a values and areas on the distribution curve is shown in Figure
3-4 by the lines drawn vertically at intervals of l a and 2a from the central
value.
It is very comforting to have such definite numbers, because we can say

definitely that any particular value in a Gaussian set has a 68% chance of
falling within the interval X ±C5 and a 95% chance of falling within X ± 2 a ,
Figure 3-4 The relationship of 1a and 2 a limits to the Gaussian distribution.

and we shall have the occasion to use these probability values repeatedly. A
more extensive account of the mathematical properties of the Gaussian dis
tribution is in Appendix 1.
3 -7 RELATION BETWEEN GAUSSIAN DISTRIBUTIONS

A N D REAL OBSERVATIONS
The results given in the preceding section provide useful, precise methods for
interpreting means and standard deviations, but a problem arises when we
apply such thoughts to real measurements. Numbers like 68% and 95% refer
specifically to a theoretical construct, the Gaussian distribution. When we
engage in a real measuring process, all we have is one, or at most a few, ac
tual measurements of our desired quantity. We have at first no way of
knowing which Gaussian distribution, with attached values of X and a , is
appropriate to our observations. So what are we to do? The answer lies in a
concept that provides a bridge between the world of theoretical constructs
and the world of real measurements. For the particular circumstances of our
measurement, we invent the concept of the infinite set of measurements that
could be made. For rather obvious reasons, this infinite set of measurements
will never be made, but the concept enables us to interpret our real meas
urements. The construct is called the universe, or population, for that par
ticular measurement. Once we have made, say, 100 measurements with a
particular apparatus, we tend to feel that nothing exists but our 100 values.
We must now invert our thinking and view the measurements as a sample of
the infinitely large universe, or population, of measurements that could be
made. The universe, however, is permanently inaccessible to us; we shall
never know the universe distribution or its mean or its standard deviation.
Our task is to construct inferences about these quantities from the definitely
known properties of our sample.
We do this on the basis of some assumptions. First, we assume that the

universe distribution is Gaussian, and we call the universe mean X and the
universe standard deviation a . This assumption enables us to make state
ments such as: If we make just one measurement with our equipment, that
one measurement has a 68% chance of falling within the interval X ± a and a
95% chance of falling within X ± 2 g . This seems like an encouragingly exact
and explicit statement, but it has an overwhelming defect; we do not (and
never shall) know the values of X and a . In other words, having made only
one observation of a quantity that is subject to random fluctuation, we have
gained practically nothing. We can say only that our single measured value
has a 68% chance of falling within something of somewhere, which is not too
helpful. Our only hope lies in obtaining some information, even if uncertain.
about the universe distribution. As already mentioned, we are never going to

be able to determine the universe distribution exactly, because that would
require an infinite number of readings. We can only hope that, if we repeat
our measuring process to obtain a sample from the universe, that sample will
enable us to make some estimate of the universe parameters.
Because we make the basic assumption that the universe distribution is

a mathematical, defined function (whether Gaussian or some other equally
well-defined distribution), we can evaluate mathematically the properties of
samples taken from that universe and compare the sample properties with
those of the distribution of single observations. We simply state these prop
erties of samples without proof. The reader who is curious about the mathe
matical derivation of these results is encouraged to turn to the standard texts
on statistics, in which there are sections dealing with sampling theory.
The properties of samples become clear if we consider the concept of

repeated sampling. Consider that with a certain piece of apparatus we make
lOO observations. This is our first sample; let us calculate its mean and stan
dard deviation and record them. Now let us make another set of 100 obser
vations and record for it the mean and standard deviation. We continue such
repetition until we have an infinite number of samples, each with its own
mean and standard deviation, and we then plot the distribution curves of
these sample means and of the sample standard deviations. Of course, we
shall never carry out a process like this with actual observations, but knowing
the mathematical function for our original universe of single readings, we can
simulate such repeated sampling mathematically and so derive the properties
of the samples in comparison with those of the original universe of single
readings. The results of such calculations of the distribution of sample means
and sample standard deviations are shown in Figure 3-5 and Figure 3-6, and
they are described in the following sections.
3 -8 SAMPLE MEANS AND STANDARD DEVIATION OF THE MEAN
If the universe distribution of single readings is Gaussian, the theory of

sampling shows that the distribution of sample means is also Gaussian. The
distribution of sample means has two other very important properties. First, it
is centered on X, the center of the original distribution of single readings;
second, it is narrower than the original distribution. This narrowness is highly
significant because it demonstrates immediately the improvement in
precision that comes from samples as opposed to single readings; the means
of samples cluster more closely around the universe mean than do single
readings. The reduced scatter of sample means is represented by an important
42 STATISTICS QF OBSERVATION CHAP. 3
(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.)
quantity—the standard deviation of the distribution of sample means. This

quantity is called the standard deviation of the mean; its symbol is and
sampling theory gives its value as
a... = (3-5)
yfN
where N is the number of readings in the sample. This result gives us the op
portunity to make a very important statement. A particular sample mean has a
68% chance of falling within the interval X and a 95% chance for the
interval X ±2a„,. These intervals are smaller than the corresponding intervals
for single readings, and they supply a numerical measure of the improved
precision that is available from sampling.
Notice that the statement about sample means, although precise and im
portant, still does not help us much, because it still involves the unknown
quantities X and a The resolution of this difficulty and the significance of the
standard deviation of the mean will become clear soon. In the meantime we
turn our attention briefly to the other important property of samples—the
distribution of sample standard deviations.
Figure 3-6 The distribution of sample standard deviations.
3 -9 SAMPLE STANDARD DEVIATION
It can be proved by sampling theory that the sample standard deviations also fall
on a Gaussian distribution that is centered on the value of the universe standard
deviation a . The distribution is illustrated in Figure 3-6. As will become clear,
however, the variance of the sample standard deviations will not concern us as
much as the variance of sample means, and we postpone to Section 3-11 further
discussion of the variance of sample standard deviations.
3 -1 D APPLICATION OF SAMPLING THEORY TO

REAL MEASUREMENTS
The sample properties just presented are interesting, but how do they help us
when we do not have access to the actual distributions, either for sample
means or sample standard deviations? When making real measurements, we
obtain our lone sample with its mean and standard deviation, and we have no
idea how these values relate to the universe values. The problem, therefore, is
to find a connection between the theoretical results and the sample properties
that allows us to infer the universe properties from the sample values. We
cannot expect to obtain exact information. In addition, we must make one
basic, obviously imprecise assumption. Assume that our single number, the
standard deviation of our sample, provides us with a value for the universe
standard deviation. In fact, it can be proved that the “best estimate” (a

mathematically defined term) of the universe standard deviation is given by
the quantity
(3-6)
N -\
This quantity is only slightly different from the original value for the
standard deviation of a set of observations. The N in the denominator of the
original expression has been replaced by - 1, and the difference between
the two quantities is significant for only small values of N. In the future,
when we talk about a sample standard deviation, we shall assume that we are
using the equation in the new form and that we are really talking about the
“best estimate” of the universe value a .
Accepting the sample standard deviation as the best estimate of a , we

are now able to make a definite statement about the single sample. We can
rephrase Equation (3-5) and define
(3-7)
yfN
as our standard deviation of the mean, now a known quantity obtained from
our real sample. We can now say: The sample mean x has a 68% chance of
falling within the interval X ± S,„ and a 95% chance of falling within
X ± 2S„,. This statement is close to what we want, but it is not yet completely
satisfactory. It tells us something about a quantity that we know, x , in terms
of a quantity that we do not know, X. We really want the statement to be the
other way around; we want to be able to make an assertion about the un
known, X, in terms of a quantity, x , of which we do know the value. Fortu
nately, it is possible to prove that the above statement about probabilities can
be inverted to yield the desired result. We obtain thereby the statement to
ward which we have been working ever since we started the discussion of the
statistics of fluctuating quantities. The final statement is: There is a 68%
chance that the universe mean, X, falls within the interval x ± S„,. and a 95%
chance that it falls within the interval x ±25',,,. This is now, finally, a state
ment about the unknown quantity, X, in terms of wholly known quantities, x
and 5',,,. Along the scale of x values, we now have a real and known interval
between x - 5,„. and 3c -i-5',„, and we know that there is a 68% chance that the
desired quantity X lies within this interval.
This statement provides us with the answer we have been seeking and
brings us as close as we can come to exact information about the unperturbed
value of the measured quantity. It is worth becoming familiar with the

arguments that have been given in the preceding sections; there is more to
measurement than simply making a few measurements and “taking the aver
age” just because it seems to be the right thing to do. We should understand
fully the significance of what we are doing.
3 -1 1 EFFECT OF SAMPLE SIZE
Clearly, in any sampling process, the larger the sample, the more precise the
final statements. Even though the precision of a mean value increases only as
the square root of the number of observations in the sample [Equation (3-5)],
it does increase, and larger samples have more precise means. There may,
however, be limitations of time or opportunity, and we cannot always obtain
samples of the size we would like. Usually, a compromise must be sought
between the conflicting demands of precision and time, and good experiment
design incorporates this compromise into the preliminary planning. Never
theless, it may occasionally be necessary to be content with small samples. In
this undesirable eventuality, we should be aware of the magnitude of the
resulting loss of precision. There is, first, the influence on the value of the
standard deviation of the mean; the smaller N is, the larger the value of S’,,,,
and the longer the interval on the x scale that has the 68% chance of contain
ing the universe value X.
Second, for small samples, we must place declining faith in the use of
the sample standard deviation S as the best estimate of the universe value a .
To illustrate this, recall the distribution curve for sample standard deviations
shown in Figure 3-6. It is worth asking: Given the existence of this distribu
tion, how good is our best estimate of the universe standard deviation, and
how does it vary with sample size? The answer must be based on the width of
the distribution of sample standard deviations, and so we should calculate the
standard deviation of this distribution. It is called the standard deviation of
the standard deviation. (This process could obviously go on indefinitely,
but we stop at this stage.) The value of the standard deviation of the standard
deviation, calculated mathematically by sampling theory on the basis of the
equation of the Gaussian distribution, is
(3-8)
The breadth of the distribution of sample standard deviations is thus

related to its central value a by the numerical factor 1/ j 2 ( N - \ ) . As one
might expect, therefore, the accuracy of the sample standard deviation as

the best estimate of the universe value depends on the sample size. For
example, with a sample size of 10, Equation (3-8) shows that the S value
from the sample has a 68% chance of falling within an interval of
± a/^ /T 8 , approximately ± a / 4 , about the universe value a . Corre
spondingly, the interval that has a 95% chance of containing the sample
standard deviation is as wide as ± a / 2 about the universe value a . This
does not represent high precision of measurement. We have, therefore,
confirmation of the warning given earlier; statistical exercises with small
samples should be undertaken only when no alternative exists. To provide
an overall feeling for the reliability of a estimates from samples of differ
ing size. Table 3-1 contains some typical values of ^2(N - \ ) for various
values of N.
These values are illustrated in Figure 3-7 for N = 3, N = \0, and
= 100. The ±lcjs limits are marked on these curves, showing, for various
sample sizes, the intervals within which there is a 68% probability that the
single sample standard deviation lies. For values of N less than about 10, it is
clear that the intervals for 68% or 95% probability become so large in com
parison with the central value that it is almost pointless to attempt an estimate
of a . It is rarely worth attempting any kind of statistical analysis with samples
containing fewer than about 10 observations. When reporting the outcome of
statistical work, it is essential to quote the sample size. If we intend our values for
the mean and standard deviation of the mean to be interpreted in accordance with
the 68% and 95% prescription, we must give our readers the opportunity to judge
the accuracy of our estimates.
TABLE 3-1 Accuracy of G Estimates from Samples of Varying Size
6 8 % C o n fid e n c e 9 5 % C o n fid e n c e
N ■j2{N-\) N yl2(N-\)
2 1.4 2 0.7
3 2 .0 3 1.0
4 2.4 4 1.2
5 2.8 5 1.4
6 3.1 6 1.6
7 3.4 7 1.7
8 3.7 8 1.8
9 4 .0 9 2 .0
10 4 .2 10 2.1
15 5 .2 15 2 .6
20 6.1 20 3.2
50 9.8 50 4 .9
100 14.1 100 7 .0
(c)
Figure 3-7 Sample standard deviation distributions for samples of various sizes.
3 -1 2 STANDARD DEVIATION OF COMPUTED VALUES
In Chapter 2 we considered the uncertainty of computed values z, and we as

sumed that the uncertainty of the basic measurements constituted intervals
within which we were almost certain that the values lay. We calculated the
maximum range of variability of the computed answer on the pessimistic as
sumption that the errors in the various measured values had combined in a
worst-case fashion to drive the computed answer as far away from the central
value as it could go. We have already suggested that this represents an
unrealistically pessimistic approach and that a more useful quantity would be
a probable value for the uncertainty in z that is based on the various prob
abilities associated with deviation of the basic quantities x, y, and so on, from
their central values. The limits given by this quantity will naturally be smaller
than the ± 8z that we calculated before, but we can hope to find actual
numerical significance for them. Such statistical validity is available only if
the uncertainties in x and y have statistical significance, and we assume in the
following calculations that the measurements of .r and y have been
sufficiently numerous to justify a calculation of the standard deviations Sx
and Sy. We hope now to calculate a value for Sz that will have the same
significance for z values as Sx and Sy had for the values of .;c and y.
We must first ask what we mean by Sz. To construct an interpretation,

we assume that the measuring process has given us N pairs of observations jc,
y that were obtained by repetition of the observing process under identical
conditions (for example, the current through and the potential across a
resistor that had been measured for the purpose of calculating the resistance
R). Each pair of observations provides a value of z through some functional
relation z = / (jc,y). Because repetition of the basic measurements yielded N
pairs, we now have a set of N values of z. These are not identical because of
fluctuations in the basic measurements x and y. The z values therefore fall on
a distribution curve, and the quantity we require, is the standard deviation
of this set of z values. These individual values of z may, of course, never be
calculated separately because a simpler mode of calculating our final answer
exists. We can calculate the means x and y of the sets of x and y values and
obtain z directly by using the assumption (valid if Sx, Sy, and Sz are small
compared, respectively, with x , y , and z ) that
z=f(x,y)
Even if we never calculate them individually, that distribution of separate z
values provides the significance of the S, that we are about to calculate.
If we assume that the universes of the separate jc, y, and z values have
Gaussian distributions, the quantity a , (of which we are about to calculate
the best estimate in terms of the various 5 values) has the usual significance
(i.e., any particular z value has a 68% chance of falling within ± a. of the
central value). As before, let
z = f(x,y)
and consider perturbations and gy that lead to a perturbation Sz in the
computed value of z. The value of Sz will be given, as before, by
dz dz
8z : — OX+ — oy
dx dy
This perturbation in z can be used to calculate a standard deviation for the N
values of z. Each perturbation in z is equivalent to the difference between a mean
and a measured value that appears in the definition of a standard deviation.

Hence, the standard deviation of the set of z values is given by
S ( 6z)'
S. =
Thus
■ dy
( 8y)^ + 2 — — 6x5y
Af ■ 11 ^\ ^ J dxdy
1 /s \2 2 dz dz /o 5 \
=[|] N Ndxoy
But and
N
Also, because hx and 53^ may be considered for the present purpose to be in
dependent perturbations,
^ ( 6x 8y) = 0
Thus, finally.
s = 5? + (3-9)
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
with some degree of reliability, a value can be found for the probable uncertainty
of the answer, where “probable” has real numerical significance.
The calculation has been performed in terms of the variance or standard

deviation of the x and y distributions. In actual practice, however, we do not
use the sample variance directly; we must calculate the best estimates of G v,
Gy, and so on, and in accordance with Equation (3-6), we use the modified
value for standard deviation with denominator N -1 instead of N. The final
result is then a best estimate for a . . The standard deviation of the mean for z
is then calculated by direct use of Equation (3-5) and gives the limits that
have a 68% chance of containing the unperturbed value of z.
Note that most experiments are not carried out in accordance with the restricted
assumptions of the foregoing development. If, for example, we are studying the flow
rate of water through a pipe, we would measure the flow rate, pipe radius, and pipe
length independently and would choose the number of readings in each sample on the
basis of the intrinsic precision of the measurement. We cannot, therefore, use
Equation (3-9) directly, because the various 5's are not compatible. The solution is to
calculate the standard deviation of the mean for each of the elementary quantities first.
If these are used in Equation (3-9), the result of the calculation is immediately a
standard deviation of the mean for z.
3 -1 3 STANDARD DEVIATION OF COMPUTED VALUES:

SPECIAL CASES
Let us now apply Equation (3-8) to a few common examples. In all the fol
lowing cases the various 5s are all assumed to be best estimates of the ap
propriate universe value a .
Sum of Two Variables
If
z-x-\-y
then
dz
1, ^=1
& dy
and
Note that this result provides justification for Eq. (3-5). The mean value for
the sample, Z (-^/) / N, is just such a function as z = x-\-y, where x and y
happen to be independent measurements of the same quantity. Thus, if
Z = — (X, -FX , + X 3 - F . . . )
dz dz
= l / N, = \ / N, and so on.
dx. dx.
and
= NS
=s j 4n
which is the result we had earlier for the standard deviation of the mean.
Difference of Two Variables
If
z =x-y
dz dz
= l, = - \
dx dy
but, again,
s. = ^sl+sl
Recalling Section 2-8, we note that the earlier discussion of measured differ
ences is still valid. The standard deviations in x and y combine additively,
even though the quantity x - y can have quite small values.
Product of Two Variables
If
z = xy
dz dz
=X
dx dy
and
5, = +x^Sl
The specific value of SI at any particular values of jc and y, say Xq and yo, can
be obtained by substituting Xq and yo in this expression. As was the case for
uncertainty in products, the equation is more clearly expressed in terms of

relative values of Sz- We obtain
Variables Raised to Powers
If
z=X
dz
= ax
dx
and
Again, this result is more instructive and more easily remembered when ex
pressed in terms of relative values:
S.
A
X
The General Case of Powers and Products
If
z^xy"
the results of the two preceding sections can obviously be extended to give
the result
z at ) y J
In contrast to the case of combined uncertainty, negative powers in the original

function need not be given special consideration; in the equation for Sz powers
occur in squared form and automatically make a positive contribution.
If a function other than those we have listed is encountered, the use of

Equation (3-9) yields the desired result. Incidentally, we may note that, for a
function of a single variable. Equation (3-9) reduces to the same form as for
uncertainties. Equation (2-1). This correspondence could easily have been
predicted for a situation in which we do not have the probability-based
interplay between two or more variables.
3 -1 4 COMBINATION OF DIFFERENT TYPES

OF UNCERTAINTY
Unfortunately for the mathematical elegance of the development, we

frequently require the uncertainty in a computed result to contain quantities
having different types of uncertainty. We may require the uncertainty in a
function
in which, for example, x is a quantity to which have been assigned outer lim
its, ±5jc, within which we are “almost certain” that the actual value lies, and y
is a quantity whose uncertainty is statistical in nature, a sample standard
deviation, Sy, perhaps, or a standard deviation of the mean, S^,/J~N. We
require an uncertainty for z. The initial difficulty is even to define the
uncertainty in z. We are trying to combine two quantities that in effect have
completely different distibution curves. One is the standard Gaussian
function; the other is a rectangle. This rectangle is bounded by the values
Xq + 8jc and Xq - dx and is flat on top because the actual value of x is equally
likely to be anywhere within the interval ± bx. Any general method of
solving this problem is likely to be far too complex for general use, but a
simple approximation is obtainable by using the following procedure.
In the calculation for z we could use the sample mean, y , for the y
value, implying that the universe mean has an approximately two-thirds
chance of falling within the interval, y ± s j . We, therefore, calculate
limits for x that also have a two-thirds probability of enclosing the actual
value. Because the probability distribution for x is rectangular, two-thirds of
the area under the distribution curve is enclosed by limits that are separated
by a distance equal to two-thirds of the total range of possibility (i.e., two-

thirds of 25jc. The total width of the region for two-thirds probability is
therefore, (4 / 3)6x and the uncertainty limits are ±(2 / 3)5x.
The quantity (2 / 3)dx is compatible with because both refer to

two-thirds probability. Equation (3-9) can now be used, inserting (2/3)8 jc
for the value of the standard deviation of the mean for x and / ^fN for the
y function. This yields a value for uncertainty in z which can be interpreted in
accordance with the two-thirds prescription. Note that the limits for 95%
probability are not simply twice as wide as those for two-thirds probability;
they must be calculated separately using the foregoing method.
3 -1 5 REJECTION OF READINGS
One last practical property of distribution curves concerns outlying values.

There is always the possibility of making an actual mistake, perhaps by
misreading a scale or in accidentally moving an instrument between setting
and reading. There is the temptation to assign some such cause to a single
reading that is well separated from an otherwise compact group of values.
This is a dangerous temptation because the Gaussian curve does permit
values remote from the central part of the curve. Furthermore, once we admit
the possibility of pruning the observations, it can become difficult to know
where to stop. We are dependent, on the judgment of the experimenter. This
is not unreasonable, because the experimenter knows more about the
measurement than anyone else, but criteria for making the choices can be
helpful. Many empirical “rules” for rejection of observations have been
formulated, but they must be used with discretion. It would be foolish to use
a rule to reject one reading that was just outside the limit set by the rule if
there are other readings just inside it. There is also the possibility that extra
information relating to the isolated reading was noted at the time it was made,
and this can help us decide in favor of retention or rejection.
The guidance for making such decisions can be found in the properties
of the Gaussian distribution. In a Gaussian distribution the probability of
obtaining readings outside the 2a limits is 5% (as we have seen before),
outside 3a limits it is approximately 0.3%, and outside 4 a limits the chance
is no more than 6 x 10"^ The decision to reject is still the responsibility of the
experimenter, but we can say in general terms that readings that fall outside
3a limits are likely to be mistakes and candidates for rejection. However, a
problem can arise because of our lack of information about the universe of
readings and its parameters X and a . The better our knowledge of a , the
more confident we can be that any far-out and isolated reading arises from a
genuinely extraneous cause such as personal error, malfunction of apparatus,
and the like. Thus, if we make 50 observations that cluster within 1% of the
central value and then obtain one reading that lies at a separation of 10%, we
can be fairly safe in suggesting that this last reading did not belong to the
same universe as the preceding 50. The basic requirement before any
rejection is justified is confidence in the main distribution of readings.
Clearly, there is no justification for taking two readings and then rejecting a
third measurement on the basis of a 3a criterion. Unless the case for rejection
is completely convincing, the best course is to retain all readings, whether we
like them or not.
It is wise also to remember that many of the greatest discoveries in

physics had their origin in outlying measurements.
PROBLEMS
The following observations of angles (in minutes of arc) were made while
measuring the thickness of a liquid helium film. Assume that the observa
tions show random uncertainty, that they are a sample from a Gaussian uni
verse, and use them in Problems 1 to 14.
34 35 45 40 46
38 47 36 38 34
33 36 43 43 37
38 32 38 40 33
38 40 48 39 32
36 40 40 36 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.
5. Calculate the standard deviation of the mean.
6. Calculate the standard deviation of the standard deviation.
7. (a) Within which limits does a single reading have a 68% chance of
falling?
(b) Which limits give a 95% chance?

8. Within which limits does the mean have (a) a 68% chance, and (b) a
95% chance of falling?
9. Within which limits does the sample standard deviation stand (a) a 68%
chance and (b) a 95% chance of falling?
10. Calculate a value for the constant h in the equation for the Gaussian dis
tribution.
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?
12. Take two randomly chosen samples of five observations each from the
main set of readings. Calculate their sample means and standard de
viations to see how they compare with each other and with the more
precise values obtained from the complete sample.
13. If the experiment requires that the standard deviation of the mean should
not exceed 1% of the mean value, how many readings will be required?
14. If the standard deviation of the universe distribution must be known
within 5%, how many readings will be required?
15. Repeated measurements of the diameter of a wire of circular cross sec
tion gave a mean of 0.62 mm with a sample standard deviation of 0.04
mm. What is the standard deviation for the calculated value of the cross-
sectional area?
16. The wavelength of the two yellow lines in the sodium spectrum are
measured to be 589.11x10"'^ m and 589.68xlO"^^ m, each with a
standard deviation of 0.15 x 10 ^^ m. What is the standard deviation for
the calculated difference in wavelength between the two lines?
17. A simple pendulum is used to measure g using
T = 2n -
Twenty measurements of T gave a mean of 1.82 s and a sample standard

deviation of 0.06 s. Ten measurements of £ gave a mean of 0.823 m and
a sample standard deviation of 0.014 m. What is the standard deviation
of the mean for the calculated value of g?
Scientific Thinking
and Experimenting
4 -1 OBSERVATIONS AND MODELS
In this chapter we briefly review the nature of scientific activity in the hope
that the procedures used in various types of experimenting will be seen to
arise naturally from the problems that are encountered. To understand the
nature of scientific thinking, it helps to go back to fundamentals and pretend
that we are inventing a new area of scientific study right from the beginning.
Identification of Significant Variables
As we encounter through observation a totally new phenomenon, our natural

first question is—What causes this? The question was asked with respect to
the diffraction of light, radioactivity, superconductivity, pulsars, and every
other physical phenomenon. It is still being asked with respect to the nature
of elementary particles, climatic change, cancer, and many other topics.
Asking questions about causes can lead to philosophic difficulties, and it is
better to recognize that our natural questions about causes and explanations
for phenomena are really questions about the relationships between observed
variables. The flow of electrical current through a conductor, for example,
can be observed by using ordinary lab equipment to depend strongly on the
potential difference across it and not at all on whether the conductor is ori
ented North-South or East-West, and this observation can be used to guide
future study. This may seem like a foolishly oversimplified example, but at
57
58 SCIENTIFIC THINKING AND EXPERIMENTING CHAP. 4
the frontiers of scientific work when we know nothing about a new phe
nomenon, we may have to consider a wide range of possibilities. Normally,
the first phase of research on a totally new phenomenon consists of a search
for the variables that seem to be related. By identifying these significant
variables, we narrow the field of investigation to practical levels and facilitate
continued work at both experimental and theoretical levels.
It is interesting that, at this primary stage of scientific development, we

can make relatively definite statements because we are talking about actual
observations. This accounts for the reputation of scientific activities that they
lead to “scientific truth” about the universe. The claim must be restricted to
the early, diagnostic stage at which we identify the significant variables.
Following this come later stages in which we deal with a totally different
type of activity that involves a much lower level of certainty.
Concept of a Model
After we have analyzed a new phenomenon and are aware of the significant
variables, we can proceed to the next level of sophistication. To illustrate this
stage, consider an elementary example. Suppose we were going to paint a
wall and wished to know the amount of paint to order. We would have to
know the area of the wall, so what would we do? The natural reaction would
be to measure the length of the wall and its height and then multiply the two
numbers together. But what would that give us? And why would we think
that the numerical product has anything to do with the wall? When we mul
tiply these two numbers together, we do obtain something, but it is the area
of the completely imaginary rectangle that is defined by the two lengths. This
imaginary rectangle may or may not have any relationship to the wall. The
important thing to notice is that we are dealing with two completely different
categories. First, there is the real wall whose area we need. Second, there is a
completely invented, conceptual rectangle that is constructed from defini
tions, exists in our imagination only, and in the present case is specified by
the two measured lengths. We are commonly insensitive to this important
distinction because we are all so familiar with the concept of rectangles that a
simple, almost subconscious, glance at the wall reassures us that a rectangle
is a satisfactory representation of the wall.
But suppose we were not able to make that judgment. Suppose we were
blind and had done nothing more than measure the base and one side of the wall
without thinking about angles or any other property of the wall. We could
multiply our two dimensions to obtain an area that had no relevance at all to the
wall if the wall happened to have the shape of a parallelogram. To avoid that kind
of error, as blind experimenters, we would have to recognize the necessity to
CHAP. 4 SCIENTIFIC THINKING AND EXPERIMENTING 59
check that the imaginary rectangle defined by the two dimensions compared
sufficiently closely with the actual wall. To do this, we would have to know the
various properties of rectangles, and we would have to compare with the actual
wall as many of these properties as possible. For example, we could test such
properties as straightness of sides, right-angle comers, equality of diagonals, and
so on. Only after a sufficient number of properties had been compared between
the rectangular constmct and the real wall, and found to correspond adequately,
could we have faith that the area of the imaginary rectangle was a good enough
approximation to the actual area of the real wall.
The distinction we have been discussing is most significant and must be

borne clearly in mind as we pursue scientific work. In all areas of scientific
study we shall find, on the one hand, the real world and our perceptions of it,
and, on the other hand, hypothetical, imaginary constructs fabricated out of
sets of definitions. Such a construct is often called a model of the situation,
and the use of models is almost universal in our thinking, whether scientific
or nonscientific. The painter contemplating the task of painting the wall has
in mind the imaginary rectangle. In addition to the real flower that is being
studied a botanist is aware of the concept of the particular species to which
the flower is assigned. In contrast to the real flower, the species is a construct
that has been defined by a standardized list of properties. Economists
studying the economy of a country constmct models that consist of a set of
definitions and equations and have properties, they hope, that are similar to
the actual properties of the real economy. As shorthand descriptions of
systems, models give us a framework for thought and communication, a basis
for calculation, a guide for future study, and many other advantages.
The use of models is universal in scientific work. Models come in

many different kinds and they serve many different purposes, but we must
remember their most important characteristic—they all are invented con
cepts. They are constmcted so that their properties correspond as closely as
possible to those of the real world, but no model can ever be an exact replica
of its real counterpart. Models belong to different categories; a wall cannot
actually be a rectangle, nor a wheel a circle. The properties of a model, how
ever, may be similar to the properties of the real world, and in general terms
the usefulness of a model depends on the extent that its properties do corre
spond with those of the real world.
Comparison Between Models and the Real World
At the beginning of an experimental study, we are usually unaware of the

extent to which the properties of our model and its real-world counterpart
correspond. It is necessary, as a basis for all later work, to start by testing the
model against the real system. Only if the properties of the model are shown
experimentally to be adequately in correspondence with those of the real
system are we justified, like the painter about to order paint, in proceeding to
the next step.
Notice that, to be useful scientifically, a model or concept must be ac

tually testable against observation. Thus, a proposition regarding the number
of angels who can dance on the head of a pin cannot qualify as science. This
is not to say that the only useful ideas are those that can be tested against ex
perience, only that other propositions do not come under the heading of sci
ence. Those other propositions may perhaps be perfectly valid as mathe
matical or philosophical statements, or as aesthetic or ethical judgments.
Refinement of Models
In general, an experimental situation contains, first, the system itself, and,

second, a model or models of the system. Whatever else is involved, it is an
essential part of the experimenter's task to test the properties of the model
against the properties of the real system. In principle, our model will inevi
tably be incomplete and inaccurate. For example, let us return to the problem
of ordering paint for the wall. If as blind painters we test the properties of
rectangles with increasing precision against those of the actual wall, we
inevitably reach a point at which we begin to find discrepancies. If at that
point a simple rectangle is not a good enough model of the wall at the new
level of precision, we must modify the model in an effort to improve the
match between the model and the wall. To do this, we could progressively
make small changes in angles or lengths and hope that the areas calculated
for the revised models will provide increasingly accurate estimates for the
actual area of the real wall. Even with these adjustments, the model remains
an invented concept, and the area calculated from the model belongs to the
model and not to the wall.
In scientific work generally, we should feel free to change our models

at any time as the need arises. The model is our construct to begin with, and it
is only an idea that exists in our heads. In contemplating change, our only
consideration is the basic usefulness of the idea and its improved utility if it
is altered in any way. Because it is presumably impossible to construct for a
piece of the natural world a verbal or mathematical description that is the ex
act and total equivalent of the real thing, a process of continued refinement
and eventual replacement of models must be accepted as the natural course of
events. It is the normal business of scientists, whether “pure,” technological,
or social, to use the process of comparing models and systems in a con
tinuous search for improvement in models. This is usually not an easy proc
ess. The models we have now are as good as generations of intelligent and
hard-working scientists in the past have been able to make them. We should
consider ourselves fortunate in our professional work if we are able to make a
few small improvements to existing models. Major revisions or the intro
duction of completely new models are rare and tend to be associated with
Nobel prizes.
On the other hand, we need not be totally preoccupied with improving

models. Even if no model can be the exact equivalent of the real thing, the
properties of our models and systems can frequently correspond sufficiently
well for our purposes. If so, we need not be excessively concerned with the
remaining defects. We can proceed confidently with our particular task,
provided that we remember periodically to recheck the situation and confirm
the continued suitability of the model. It is not appropriate to think about
“rightness” or “wrongness” of models. We cannot claim that a model is
“correct,” only that it is “adequate,” or “suitable,” or “appropriate” for the
purposes in hand.
Model Building in the History of Science
It is possible to gain the impression from the foregoing discussion that in sci
entific development there is some kind of unique sequence that starts with
observation and ends with a satisfactory model. Indeed, scientific thinking
has quite frequently progressed in this way, but the sequence is not invari
able. There are many examples of a basic, invented idea, the foundation of a
model, that was the fruit of pure speculation by the originator, without
awareness of the observations that could be directly associated with the con
jecture. We can recall, as examples, de Broglie's speculation on the wave
model of matter, which was published in 1924 before any of the relevant
phenomena were observed directly, and Fermi's invention of the concept of
the neutrino almost 40 years before the particle itself was directly observed.
There is no single process of scientific development, no single “scientific
method.” Ideas and observations tend to shuffle forward roughly together but
with no automatic leadership from one or the other. Regardless of the precise
order of development, one point remains invariable—the fundamental activ
ity in scientific experimenting is to compare the properties of models with the
corresponding properties of the real world.
We have not discussed at all the processes by which new ideas are in
troduced to serve as a basis for totally new theories. Sometimes an existing
idea can undergo a process of continued refinement and attain closer and
closer correspondence with observation without any alteration of the basic
concepts on which the theory is founded (for example, the Ptolemaic theory
of planetary epicycles). On the other hand, a theory such as Einstein's theory

of general relativity or Schroedinger's wave mechanics can be introduced
only after completely radical revision of basic concepts and ways of think
ing—not a simple process. The manner in which such major revolutions in
scientific thinking have occurred is described in the books by Kuhn, Cohen,
and Harre that are listed in the Bibliography.
One might think that, following such major revolutions, a superseded

model or theory would be immediately discarded to make way for its succes
sor. Indeed, many models or theories have found no lasting usefulness—one
does not hear too much these days about phlogiston, or about earth, fire, air,
and water—but this is not always the case. Superseded models have quite
often sufficiently close correspondence with the system that, usually on ac
count of simplicity, they continue to be very useful. If one wishes to deter
mine the depth of a well by dropping a stone into the water, one does not
need to use Einstein's general theory of relativity as a model for gravitational
acceleration. The more sophisticated model must be used when circum
stances demand it—for example, when we wish to predict the motion of the
planet Mercury.
Detailed Comparison Between Models and Systems
To summarize our development so far, we have four ingredients in the sci

entific recipe: (1) observation, (2) an idea constructed in our imagination, (3)
the process of comparing the properties of the idea with the those of the real
world, and (4) the possibility of modifying the idea progressively to improve
the fit between the model and the system. We now turn our attention to the
actual procedures by which we can compare the properties of models and
systems. It is not sufficient to have a vague pictorial concept of the situation;
to supply an adequate basis for comparison we must be as explicit as possi
ble. This normally requires quantitative observation of the system and
mathematical procedures for specifying the model. Let us consider some
specific examples and investigate the various levels of sophistication in the
methods for constructing models and comparing them with real systems.
Consider an elastic band, suspended from its upper end, from the lower
end of which we can hang weights. The most primitive form of construct
with respect to the properties of the system would be a verbal description of
its behavior. We could say something like: As I hang more weights at the
bottom of the elastic band, it stretches farther. This verbal description could
prompt us to invent the general concept “springiness” to serve as a model.
But if we wish to refine the model to make it more useful for detailed com
parison with observation, our purely verbal methods of description start to
fail us; we cannot refine such a vague concept as springiness without resort
ing to the precision of description that is available in numerical and mathe
matical modes of expression. We would then make a series of measurements
of the extension of the elastic band as a function of load, hoping that they will
suggest a more explicit concept. We would obtain a set of measurements
such as those shown in Table 4-1.
TABLE 4-1 Extension Versus Load

for a Rubber Sample
Load, Extension,
kg m
0 .0 5 0.03 ± 0 .0 1
0 .1 0 0 .0 4
0 .1 5 0 .08
0 .2 0 0.13
0 .2 5 0 .1 9
0 .3 0 0 .3 0
0 .3 5 0 .3 4
0 .4 0 0.38
0 .4 5 0 .3 9
(Notice that for simplification we are pretending to know the weights exactly
so that we can ignore the uncertainty in them. The values of extension for the
rubber band are measurements made by us, and so the uncertainty must be
included.)
Now that we have the measurements, do they give us a complete and

adequate description of the results? Not really. It is difficult to judge the be
havior of a system from a set of numbers in a table; some form of visual
presentation is much superior. A simple graph of the observations can com
prise all the information contained in the table and can in addition confer the
enormous benefit of facilitating visual Judgment of the results. Such a dia
gram is shown in Figure 4-1, in which we have plotted, in addition to to the
central values of the measured variables, the actual intervals over which the
measurements of extension are uncertain.
In Figure 4-1 we have done nothing more than plot the observations on
the graph. At this stage the set of observations is the only thing we have, and
there is no justification for putting anything else on the graph.
This completes the first stage of the process, that is, observation. We
must now undertake the next stage, in which we construct a model, or mod
els, of the system.
Figure 4-1 Graph of extension vs. load for rubber sample.
4 -2 CQNSTRUCTIQN OF MODELS
The type of process required at the various stages depends much on the par
ticular experiment. For example, we may be experimenting on a phenomenon
that is being observed for the first time and for which there are no existing
ideas. In such a case, our task would be to identify the significant variables
and possibly to generate some kind of model. Or we may be working on
some relatively familiar phenomenon, in which case we would probably have
some existing proposal or theory that could be applied to our system, thus
creating a model. Whatever the circumstances, we draw a distinction between
models that are empirical and models that are theoretical. The word
empirical means that models of this type are based solely on the observations
themselves, without any reference to the detailed, internal operation of the
system. The processes by which we can generate empirical models and the
usefulness of such models is described as we proceed. A theoretical model is
constructed more generally, not just for one particular set of observations,
and is based on some basic concept or principle about the actual mode of
operation of the system. The nature of theoretical models and their usefulness
is also described. We consider each type in turn.
Empirical Models
Assume that we have made a set of observations on a system for which there is
no existing model. All we have is a set of observations on some property of the
system. It could be the load versus extension measurements on our elastic band,
and the results probably take the form of a graph hke that in Figure 4-1. Our
problem is to construct a suitable model. What can we do? There are several
possibilities, and we consider them in order of increasing sophistication.
Verbal statement. The simplest model of all is a simple verbal description

of the variation. We could say something like: The extension increases
smoothly with load in an S-shaped curve. Notice that even this simple sen
tence is a construct. As soon as we stop talking about the individual observa
tions and start talking about the whole variation of extension with load, we
have made the transition from statements about particular observations to
constructed presumptions about the behavior of the system. Even such a
vague proposition as the foregoing statement could, on closer measurement,
turn out to be unsatisfactory. Perhaps, for example, the variation is really
stepwise rather than smooth. The constructed nature of even such simple
statements is stressed here as a reminder that we must always be aware of the
distinction between statements about the observations themselves and state
ments that sound as if they were about the observations but are actually
statements about our ideas concerning the observations.
Drawing a smooth curve through the points. The next stage of sophisti
cation in model construction is represented by a process that is so commonly
carried out (usually without due regard to its significance) that its name is
used as the heading for this section. As we view the graph of observations
initially (Figure 4-1), we must remember that it contains the observational
points and nothing else; we have no basis yet for putting anything else on the
diagram. There will inevitably be some scatter in the points because of their
inherent uncertainty, but it is possible to base the model construction on the
single basic assumption that, uncertainty and scatter notwithstanding, the
actual behavior of the system is smooth and continuous. This is our concept,
or idea, and as we draw a smooth curve (Figure 4-2) through the points, we
assume that it is valid to apply that concept to our system.
The assumption of smooth, continuous behavior can be valid to a high

degree of accuracy for many systems. An example is planetary motion, for
which many of the procedures for treating observations were first devised.
The responsibility for deciding to assume smooth and regular behavior lies
with the experimenter, who should make the assumption only if familiarity
with the system leads to the carefully considered conviction that it is valid.
The benefits of assuming regular behavior and drawing a smooth curve

through the points can be substantial. One of the most obvious benefits is as
sociated with interpolation and extrapolation. Consider that we have the set
of observations shown originally in Figure 4-1 and that we have drawn a
smooth curve through the points as shown in Figure 4-2. Our knowledge of
the system is good at the points at which measurements have actually been
made but if we want to find the value of the extension at a load intermediate
between two of the measured values, we have a problem. We could go back
to the apparatus and make the desired measurement, but for many reasons
this course of action could be either impossible or undesirable. We are then
left with the possibility of only inferring the desired value on the basis of the
existing measurements. The smooth curve provides one way of doing so, as
shown in Figure 4-3. We must remember that the answer obtained by inter
polation is an inferred value that depends on the decision to draw a particular
smooth curve.
Likewise, it is possible to use a smooth curve to extrapolate beyond the

existing range of values, as shown in Figure 4-4. Such a procedure enables
us to make a guess at values outside the measured range, but the validity of
the procedure is obviously much more limited than was the case for interpo
lation. Before proceeding with extrapolation, we must have very good rea
sons for believing that the behavior of the system remains regular beyond the
measured range. Smooth variation inside the measured range does not by it
self offer any guarantees about a wider range of behavior (Figure 4-5).
Mathematical methods for interpolation and extrapolation are given in

Appendix 3, and they can be used to obtain interpolated and extrapolated values
by calculation, without actually drawing the smooth curve. Such methods still
depend on the assumption of smooth, regular behavior of the system, and the
inferred values draw their vahdity from the rehability of that assumption.
Because the validity of interpolation and extrapolation is limited by the

assumption of smooth and regular behavior, opportunities for error abound.
If, for example, we were to offer someone the graph of temperature versus
time (Figure 4-6), without specifying the system, and ask that person to infer
the value of temperature for a time halfway between two measured values.
the usual answer would be to draw the smooth curve and obtain the interpo
lated value as shown on the graph. We could then reveal that the graph de
picts the noonday temperatures for the first few days of this month and that
we were asking for a temperature at midnight. Likewise, people who have
belief in the infallible validity of extrapolation can be asked why they have
not made a fortune on the stock market.
Before we leave the topic of drawing smooth curves through points,

one final procedure deserves mention. We commonly encounter graphs in
which the points have been connected by straight-line segments, as shown in
Figure 4-7. Spreadsheet or graphics programs for computers usually auto
matically draw line graphs in this form. How are we to interpret such a dia
gram? Surely we are not being asked to believe that these segments represent
the actual behavior of the system between the measured points. The only
possible benefit seems to be to supply some kind of emphasis. In a diagram
containing a number of possibly intersecting graphs, the segments do help
sometimes in identifying the various graphs. However, such segments repre
sent satisfactorily neither of the two basic ingredients of experimenting, ob
servations and models, so their use is rarely beneficial and they can be mis-
Figure 4-5 Extrapolation is not one of the exact sciences.
leading. For scientific work they are not recommended except in special
cases. The way in which the common spreadsheet programs can be adapted
for our purposes is described later.
Function finding. As a more sophisticated form of drawing smooth curves

through points, we can use a variety of mathematical methods to find various
analytical functions the graphs of which match, to a greater or lesser extent,
the variation of the measured values. Obviously, despite all the mathematical
sophistication that may be involved, such procedures still depend for their
validity on the basic assumption of regular behavior in the system; the curves
and functions are our concept of the behavior of the system. Nevertheless,
functions generated empirically to fit sets of observations can be useful. As
mathematical models of the system, they can, be used with varying precision,
to obtain inferred values for some characteristic of the system by inter
polation and extrapolation.
It is important to remember clearly that interpolation and extrapolation

using an empirical function depend on the validity of that particular function
as a model for the system. This does not usually pose too much of a problem
for interpolation, where good knowledge of the actual behavior of the system
Figure 4-6 Potentially fallacious use of interpolation.
is available on both sides of the interpolated value. Extrapolation is a differ

ent matter. We usually notice this when we are dealing with extrapolation in
time; forecasting is an uncertain business. We can forecast accurately the
time of sunset a week away because the relevant models are very good. We
have much less success in forecasting the weather a week ahead because our
models are much less satisfactory, and forecasting in other areas, such as the
stock market, proves to be almost impossible.
It suffices at present to note the possibility of constructing empirical

mathematical models of systems. The methods for doing this are described in
Chapter 6 .
Theoretical Models
Theoretical models are part of familiar theoretical physics. All analytical

theories in physics are constructed out of basic building blocks—definitions,
axioms, hypotheses, principles, and so on, followed by analytical derivation
from these basic starting points. Because all the elements of theories are
constructs of human imagination, the theories themselves and the results of
the theories are similarly imaginary constructs. Their relevance to actual

systems must be evaluated through experiment.
Let us illustrate the situation by using a particular example. Consider a

system in which we can release a steel ball bearing to fall freely under grav
ity, and we measure its time of fall from various heights. If we wished to
construct an empirical model of this system, we could simply measure the
time of fall over a number of different distances and graph the result, which
would look something like Figure 4-8. We could then use one of the tech
niques from the preceding section to obtain an empirical model for the sys
tem. If we wished to construct a theoretical model of the situation, however,
our approach would be completely different. We would have to choose a set
of basic axioms or hypotheses from which we would derive the required re
sults. For example, we might decide to use as a basic hypothesis a presumed
value for the acceleration of the ball bearing:
a = 9.8 ms'2
Notice that this hypothesis already contains several assumptions about the
system, thereby starting our process of constructing an invented model. By
choosing a constant value for the acceleration, we are implicitly neglecting

the presence of air resistance. We have every right to do so. The model is
ours; we are free to construct it in any way we please. Whether that assump
tion makes it a good model we may not yet be able to tell. As a second ex
ample, we are also neglecting effects associated with general relativity;
whether this is a serious defect also remains to be seen. We should try to es
timate in advance the validity of the assumptions that are built into the
model, but we are often limited in our ability to do this. There is always some
point at which we have to decide to start experimenting on the basis of the
model as it is and to rely on the experimental results to tell us if further
refinement of the model is necessary.
We are now ready to proceed with the development of our theoretical*

model. By integration we obtain
v = 9.8r assuming V= 0 at / = 0
and
assuimng X = 0 at ^ = 0
2
or
1
4.9
We have chosen to write the final equation in a form in which t is expressed
as a function of jc because this corresponds to the way in which the experi
ment was set up. We chose x as the input, or independent, variable and
measured t as the dependent, or output, variable. So we want our equation to
tell us r as a function of jc.
In the course of the derivation, all the assumptions that we insert con
stitute further components of the model. The final result for the measurable
variable, the time of fall, is thus a property of the model. Its applicability to
the system is the next topic of investigation.
4 -3 TESTING THEORETICAL MODELS
Consider actual measurements for the free-fall experiment. We have

treated the distance of fall as the independent, or input, variable for which
we chose the values; the time of fall is then the dependent, or output,
variable of which the system gives us the values. The results of the
experiment are shown in Table 4-2. In this experiment, the measurements
of the distance of fall could be made much more precisely than those of
the time of fall. For simplification, we consider the uncertainty in the x
values to be negligible. Normally, we would have to include the
uncertainty in all the measured quantities.
TABLE 4 -2 E x p e rim e n ta lly M e a su re d T im e o f F all V e rsu s D ista n c e for

a F re e ly F a llin g O b je c t
D ista n c e , jc, m T im e , /, s
0.1 0 .1 4 8 ± 0 .0 0 5
0 .2 0 .1 9 6
0.3 0 .2 4 4
0 .4 0 .2 9 0
0.5 0 .315
0 .6 0 .3 5 2
0.7 0 .385
0 .8 0 .403
The measurements given in Table 4-2 describe the behavior of the

system. We also have the behavior of the model, in the form of the function
that was the outcome of the analytical derivation:
t=
4.9
The task is somehow to compare these two, but it is not at all clear how that
should be done. One simple suggestion is to insert the various values of x in
the equation and calculate corresponding values of t. We could then compare
these with the measured values. If they agreed exactly, we could perhaps be
confident that the system and the model were in correspondence. The
probability of that happening, however, is minute; apart from anything else,
the presence of uncertainty in the measurements eliminates the possibility of
exact correspondence. The major point, though, is that the model is most
unlikely to be totally free of systematic defects and deficiencies. It is one
of the principal purposes in experimenting to detect these discrepancies
and deal with them constructively. The possibility of doing this
effectively by using simple arithmetic comparison is small. Much more
significant for our purpose is the overall behavior of the system; the best
way to view that behavior is on a graph.
The graph of our observations, shown in Figure 4-9(a), consists of a

series of points. The graph of the model's behavior is a curve, which is shown
in Figure 4-9(b). The two graphs together give us a visual impression of the
relationship between the properties of the system and those of the model. The
comparison would be more detailed yet if we could pick up one of the graphs
and lay it over the other. By doing so, we obtain the composite diagram
shown in Figure 4-9(c). Notice that this diagram has two different
components: ( 1) points representing the properties of the system, and (2) a
line corresponding to the analytical function that belongs to the model.
At last we can make a detailed comparison between the overall proper

ties of the system and those of the model. By straightforward visual inspec
tion, we can say that the model and the system are in correspondence, or are
divergent, or whatever. We list the various possibilities in more detail in
Chapter 6 . For the present, we must note carefully the kind of statement we
are able to make at the end of an experiment. We can say only that the be
havior of the model and of the system were in correspondence (or were not)
to such and such an extent. It is pointless to agonize over whether a theory is
“true,” “correct,” “wrong,” or whatever. As was mentioned when we first
discussed the nature of models, we should avoid using such terms, even if we
are sure we understand the situation. Others may be less clear about the use
of words than we are, and there are too many opportunities to be misun
derstood. It is far better to categorize a theory or model as “satisfactory” or
“good enough,” or some similar phrase, because all such decisions are rela
tive to the purposes we have in mind.
(j,) Distance, m
Figure 4 -9 The process of comparing the properties of a real system
with the properties of a model.
For example, our simple model of constant acceleration under gravity is

perfectly satisfactory for finding the depth of a well by dropping a stone
down in it, but it is not satisfactory for calculating the trajectory of a space
vehicle en route to the moon. If that were our purpose, we would have to
construct a more refined theory until we had one that is good enough for that
purpose. Even then, we would find that a theory adequate for moon rockets is
inadequate to describe the motion of the planet Mercury. For that, as was
mentioned earlier, the theory of Newtonian gravitation must be replaced by
Einstein's theory of general relativity. And the adequacy of Einstein's theory
for describing Mercury's orbit (at a particular level of precision) does not
“prove” that it is true or correct, simply that it is good enough for that pur
pose. Equally, the presence of Einstein's theory does not discredit either
Newton's theory of gravitation or our simple model of constant acceleration
under gravity. Most people do not measure the depth of wells by using Ein
stein's relativity theory. In general, we use a particular theory because it is
good enough for the purposes in hand. If increased precision is desired at any
time, the necessary refinements can be introduced as required (unless, of
course, we are working at the limits of knowledge in a particular area, and the
chief obstacle is the absence of an improved theory).
Because we are no longer going to use the misleading concept of the

“truth” or “correctness” of theories and models, we shall be dependent on our
own decision that a chosen model is good enough, or not, for our purpose.
One of the primary aims of experiment design is to test the models we use
and check their suitability. If it is properly planned, the experiment itself will
tell us whether the model or theory is good enough.
In passing, we can note one interesting point of philosophy. Even when

our system and model seem to be in complete correspondence, we have to be
careful about stating the outcome. All we can say is that, at a particular level
of precision, we have failed to observe any discrepancy between the system
and the model. It is possible to be more assertive if we are sure that the prop
erties of the model and system are in disagreement by an amount clearly in
excess of the measurement uncertainty; we can say definitely that the model
is not in correspondence with the system. We can say that we have “proved”
the theory to be “wrong”—although, even in this case, it would be better to
call it “unsuitable” or “inadequate.”
Before proceeding, we must note that this process of comparing sys

tems and models depends on our ability to draw the graphs of the functions
that appear in the models. At one time this presented substantial difficulties
for even simple functions, such as parabolas, and often insuperable difficul
ties for more complicated functions. Now, the use of computers allows us to
make the comparison directly. We can display the graph of our experimental
values along with the graph of the function that appears in the model, and
judge immediately how well the two correspond. The ways in which common
spreadsheet programs can be made to do this are described in Chapter 6 .
Even when taking advantage of the marvelous opportunites offered by

computers, we must not neglect the development of our own personal expertise in
experimenting. First, when no computer is available, we are forced to rely on our
own resources. Second, even when we are engaged in computer-based evaluation
of experiments, we stand the chance of obtaining completely meaningless results
unless we are clearly aware of every detail of the computations that are being
carried out invisibly by the computer. To develop our personal expertise, we must
turn our attention (as was recommended earlier for calculations on frequency
distributions and standard deviations) to the old-fashioned procedures that still
constitute the basis of almost all experimenting.
If we are restricted to pencils, graph paper, and rulers, we are virtually

compelled to compare the model with the system by using the only function
whose graph is easy to draw—a straight line. (For fairly obvious reasons we
exclude from consideration that other simple graph, the circle.) A large
amount of experimental analysis is still carried out in linear form, and even if
the graphical techniques are cumbersome and tedious in comparison with the
ease of computer-assisted evaluation, the methods are sufficiently powerful,
important, and commonly used that we must become familiar with them.
4 -4 USE OF STRAIGHT-LINE ANALYSIS
The objective is to arrange the plotting process so that the behavior of the
system and the model are represented on a graph in linear form.
Consider the function for the time of free fall under gravity
1/2
t=
4.9
This function, when plotted on an x, t graph has the shape of a parabola.
Consequently, if we were to plot the measurements of x and t, intending to
compare them with the graph of the function in the model, it would be almost
impossible to judge visually whether our results were compatible with a
parabola. Suppose, however, we were to plot as variables, not t versus x but t
versus . Tthe equation that appears in the model
r = 0.4515(x"')
would then take the form of a straight line, whose equation we write
vertical variable = slope x horizontal variable

in which the vertical and horizontal variables will be given by
vertical variable = t
horizontal variable = x
and
slope = 0.4515
The experimental values of and t are given in Table 4-3 and are plotted
in Figure 4-10. This graph also contains the line representing the function;
the resulting simplification is immediately obvious. The whole process of
comparison is facilitated, and we can identify immediately the degree of cor
respondence between the model and the system.
T A B L E 4 - 3 E x p e rim e n ta lly M e a su re d T im e o f F all V e rsu s S q u a re R o o t o f D ista n c e fo r a

F re e ly F a llin g O b je c t
Distance, x, m (Distance) ,x ,m Time, t, s

0.1 0 .3 1 6 0 .1 4 8 ± 0 .0 0 5
0.2 0 .4 4 7 0 .1 9 6
0.3 0 .548 0 .2 4 4
0 .4 0 .6 3 2 0 .2 9 0
0.5 0 .707 0 .3 1 5
0 .6 0.775 0 .3 5 2
0.7 0 .837 0 .3 8 5
0.8 0 .8 9 4 0 .4 0 3
In this example we chose the time of fall as the output variable and the
distance of fall as the input variable. This choice caused us to plot the vari
ables as t and . The process would have been equally effective and prob
ably more convenient if we had plotted versus x instead of t versus x* ^
The slope would have been different, but the opportunity to compare the
model and the system would have been equally good. In an experiment, there
may be several equivalent ways to plot the variables in the form of a straight
line. Sometimes one is more convenient than another; sometimes one gives
us a better basis for comparing the system and the model. We have to make a
decision each time on the basis of the particular circumstances.
Finally, in this example the properties of the model were completely

specified, and the line on Figure 4-10 representing the model's behavior is
Figure 4-10 C o m p a riso n b e tw e e n th e p ro p e rtie s o f a m o d el a n d o f the

sy ste m w h e n e x p re ss e d in s tra ig h t-lin e form .
unique. The situation is slightly different if the model contains quantities of

which we do not know the value; this is the topic of the following section.
4 -5 CASE QF UNDETERMINED CONSTANTS
Suppose we are doing an experiment on a spring to determine the extension

under various loads (again assuming that only the extension has an uncer
tainty that needs to be taken into account). Suppose we are aware of a pro
posal (due to Hooke) that extension x can be considered to be proportional to
load W. This proposal, expressed in mathematical form as
X = constant x W
constitutes an invented model of the system. Assume that we wish to test this
model against the system. The only trouble is that (unlike the former example
of the falling ball in which the model contained the known value of the
gravitational acceleration) we may not know the value of the constant (the
“springiness”) that appears in the equation defining the model. Suppose we

have made measurements of extension versus load and plotted them in Figure
4 - 1 1(a). What are we to do to represent the behavior of the model? The
equation
X = constant x W
really represents the infinite set of lines on the W - x plane that have all the
values of slope, from zero to infinity, that correspond to the infinite range of
possibilities for the value of the constant. Some of these lines are represented
in Figure 4 -1 1(b). What, then, constitutes the outcome of our comparison?
Laying one graph on top of the other produces the diagram shown in Figure
4 -1 1(c) and provides us with the opportunity to choose a line that is compat
ible with the experimental points.
But which line or lines are we to choose? Clearly the lines OA and OB
have no obvious relevance to the observations and can be disregarded. We
can, on the other hand, identify a certain bundle of lines that fall within the
region of uncertainty of the measured points. We can estimate visually the
edges of this bundle; they are represented by the lines OC and OD. Within
these limits all the lines have some degree of consistency with the observa
tions, but no single line stands out as uniquely suitable. All we can say for the
moment is that the observations are compatible with the model over a certain
range of slopes. This means that there is a certain range of values of
“springiness” (within the model) that are consistent with the system. The
conclusion is, then, that if we have an initially undetermined constant in the
model, the experimenting process can be used to determine, within a certain
interval, the value that is appropriate for the system. This is the normal way
of determining experimental values of physical quantities.
The process is so commonly used because, in addition to the almost

necessary use of the graph in comparing models and systems, graphical methods
of obtaining values of experimental constants offer so much additional advantage
in increased precision that their use is compellingly attractive. The opportunities
for error when using algebraic computation alone without graphical checking are
great. For example, suppose we are trying to obtain a measured value such as the
electrical resistance of a resistor from the variation of the voltage across it with
current through it. We make pairs of measurements of I and K and we use the
relationship V = R1 directly to obtain a value of R from each pair of /, V values
by purely algebraic means. We then hope to obtain an accurate value for R by
calculating the average of all the resulting R values.
This approach is deficient in many ways. Basically it fails to satisfy the

primary requirement—to compare the properties of the system and the
(Distance)’^^,
Figure 4-11 T h e u se o f s tra ig h t-lin e a n a ly sis to o b ta in th e v alu e o f an u n d e te rm in e d c o n s ta n t in a m odel.
model—and the consequences for the accuracy of the R value can be serious.
If all our pairs of /, V values gave the same, or nearly the same, value for /?,
we might feel confident in our measurement of R, even without drawing the /, V
graph. In the much more likely event that the R values do not all turn out to be the
same, we have no way of interpreting the variations without the graph.
We might, for example, encounter a case in which, as plotting the graph

would have revealed, the points show more scatter than we expected [see
Figure 4 - 12(a)]. Using a graphical approach, we could still choose a suitable
straight line (passing through the origin, if we are sure of the origin as a
measured point) and feel reasonably confident about the R value obtained
from the slope. Our confidence is justified because the appearance of the
graph convinces us that we are dealing with simple scatter about a basically
L oad
(a)
Figure 4-12 U se o f g ra p h s to av o id e rro rs in slo p e m e a s u re m e n ts .
linear variation. Our nongraphical, algebraic calculation would, on the other

hand, give us values that correspond to the slopes of the lines OA, OB, OC,
and so on. In a simple table of values, the resulting variability would make no
sense at all, and we would gain no insight into what is happening.
As a more significant illustration of the inadequacy of an algebraic,

nongraphical approach, consider a case in which some failure of correspon
dence between the model and the system gives rise either to an unexpected
intercept or to deviation from linearity beyond a certain range. Using a
graphical approach, we can easily detect and compensate for these discrep
ancies between model and system. In the first case the graph enables us to
obtain a reliable value of R from the slope, which is in such cases unaffected
by the presence of the intercept. In the second case we obtain the R value
from the slope of the linear portion of the /, V variation, rightly dismissing
the nonlinear points as lying outside the scope of the model. The opportunity
to make these judgments can come only from visual inspection of the graph,
which makes the situation clear at a glance. As before, nongraphical, alge
braic calculation from the pairs of /, V values alone yields values of R corre
sponding [see Figures 4 - 12(b) and (c)] to the slopes of OA, OB, OC, and so
on. These slopes have nothing to do with the slope we want; if we were to
include them in an average of algebraically calculated quantities, we would
succeed only in introducing error into our answers.
As we ensure in these ways that the final answer is free from such
sources of systematic error, it does not matter whether or not we know the
origin of the discrepancy. For the purposes of obtaining the value of the
quantity under study, it is sufficient at this stage merely to identify the existence
of the discrepancy and to ensure that it is not permitted to introduce errors into
the answer. We can later consider possible sources of the discrepancy.
We have discussed the process of obtaining values for initially unde

termined constants in terms of slopes only. In principle, since a line has two
degrees of freedom on a plane, it is possible to obtain from it two pieces of
information independently, such as a slope and an intercept. Because of this,
an experiment can be made to yield values for two separate quantities that
appear in a model. Specific methods of doing this in actual practice are dis
cussed in Chapter 6 .
Experiment Design
In Chapter 4 we described the various circumstances in which we

compare the properties of models and of systems. We encountered such
variety that it will come as no surprise to learn that there is no single way
to plan experiments. The techniques and procedures we use depend on
circumstances, and we describe procedures that are appropriate to a
number of cases. The list is not exhaustive but it identifies general
principles that are valid in a wide variety of experimental circumstances.
Foremost among these is the following. Whatever the circumstances we
encounter, we make sense in our experiment procedures only if we keep
clearly in mind this central point—the fundamental requirement in
experimenting, whatever else is going on, is to compare the properties of
a system with the properties of a model or models.
Assume at the outset that, as the outcome of preliminary

investigation, we already know the significant variables. Some of these
are under our control and can serve as input variables. Others take values
determined by the system they are the output variables. In the following
sections, we assume that the input variables can be separated and
individually controlled; otherwise, with everything varying
simultaneously, the interpretation of the results is much more difficult.
This unfortunate circumstance is frequently encountered in professional
experimenting, but we restrict ourselves here to the case of a fully
controlled experiment.
84
CHAP. 5 EXPERIMENT DESIGN 85
5 -1 TO TEST AN EXISTING MODEL
In this section we are concerned with situations in which a model of some

type is already available. This model can be the simplest of suggestions
(whether theoretically derived or empirical), such as F = kx or K = /?/, or it
can be derived from some grand, sophisticated theory such as Einstein's the
ory of general relativity. Whatever the nature of the model, its properties al
most invariably will take the form of a functional relation between two or
more variables. The primary objective is, as always, to compare the proper
ties of the model with those of the system. Only after we have by experiment,
satisfied ourselves that, over some range at least, the properties of the system
and of the model overlap, are we entitled to go ahead with the evaluation of
the quantity we wish to measure.
Notice that any decision about the appropriateness of the model for the
system must be based on the experiment itself. We are, of course, not going
to attempt to decide on such meaningless questions as whether the model or
theory is “true” or “false,” “correct” or “incorrect,” or whatever. As we have
said so often, all models are imperfect in principle, and we simply need to
know if the model is good enough for our purposes at our level of precision.
Only our own experiment can provide the basis for making that decision, and
it is one of our major objectives in designing the experiment to ensure that
this will be possible. Once we have checked that our model is good enough,
we can proceed to evaluate our unknown quantity, not forgetting that, if our
situation changes and increased precision is called for, we must reopen the
question of the model's adequacy for our purpose. As has been described in
Section 4-3, almost invariably the best ways of testing models of physical
systems involve a graphical approach. In principle we wish to draw a graph
of the model's behavior and to superimpose on it our observations of the be
havior of the system. To do this in simple form, however, there are some
requirements.
First, because a graph (as we are considering it) is a two-dimensional

diagram, we must limit ourselves initially to two variables. In many cases,
this requirement is automatically satisfied, as it has been in all the preceding
examples. In others, however, the output variable is a function of two (or
more) independent input variables. We cannot plot three variables as coordi
nates on a two-dimensional piece of graph paper (although three-dimensional
diagrams can easily be generated by computers and are frequently seen in the
scientific literature). Consequently, for our purposes, it is necessary to sim
plify the experiment by holding one of the input variables constant while
studying the dependence of the output variable on the other. We can then al
ter the first variable to a second fixed value and repeat the process. By a sue-
86 EXPERIMENT DESIGN CHAP. 5
cession of such measurements, we can build up a relatively complete picture

of the behavior of the system. Notice that the success of the process depends
on the primary assumption that it is possible to hold one of the input vari
ables constant, independently of variation in the other. If such isolation of the
input variables is not possible, we have problems; some of the necessary
techniques are mentioned in Section 5-2.
For our present purpose suppose that we have only one input variable,
either because only one exists or because we can isolate one by holding the
others constant. The procedure is clear; we must measure the variation of the
output variable with the input variable and plot the resulting values for com
parison with the corresponding graph for the model. As was suggested in
Section 4-3, however, it would take a computer to draw even simple nonlin
ear functions, and the advantages of drawing graphs in straight-line form are
so overwhelming that we consider only this approach.
5 -2 STRAIGHT-LINE FORM FOR EQUATIONS
Simple Cases
If the model we are considering contains only linear functions (such as dis
tance traveled at constant velocity as a function of time, or the potential dif
ference across a constant resistor as a function of current), we have no prob
lem; the equation is already in straight-line form. This is rarely the case,
however, and we are almost invariably faced with the necessity of converting
the functions found in the model into linear form. We have already encoun
tered this requirement in Section 4-3. There the function was
/ = 0 . 4 5 1 (in units of meters and seconds)

and, clearly, if we wish to represent this equation in the linear form,
vertical variable = slope x horizontal variable + constant
we must choose
vertical variable = t
horizontal variable =
slope = 0.4515
and
constant = 0
This is a simple case, and it is often less easy to see how an equation
can be converted into linear form. There are no definite rules for doing it. The
best way is to keep clearly in mind the form toward which we wish to work,
vertical variable = slope x horizontal variable + constant
and juggle the quantities in the original equation around until we have the
required form. Opportunities for practice are found in the problems at the end
of this chapter.
Notice that there is no unique answer in this process. A given equation

can sometimes be put in linear form in several different ways. For example,
the equation
/ = 0.4515x''^
can be used equally effectively in any one of the equivalent forms
1
-t r =0.2039x. x = 4.905r'
0.4515
with appropriate choices for vertical variable, horizontal variable, and slope.
There is a conventional tendency to plot graphs with the input variable hori
zontally and the output variable vertically, but there is no real requirement to
do so. We should choose the form of graph that most effectively serves our
purpose.
Our purpose should include not only the basic experimental require
ments, but also the comfort and convenience of the experimenter. For this,
one should plot variables as simply as possible. For example, consider an ex
periment to determine a coefficient of viscosity by studying the flow of liquid
through a pipe. The appropriate equation (Poiseuille's equation) is
Pna^
Q=
8 ti^
where Q = rate of flow

P = pressure difference across the pipe
a = pipe radius
£ = pipe length
r\ = coefficient of viscosity
In this case the measured variables are P and Q; a and £ have constant
measured values. Our intention to plot the quantities to obtain a value for r\.
One possible choice is to plot Q versus {na^ I ii)P . This seems to suit our
purposes because the resulting graph has a slope equal to l/r|, but it is an
unwise choice for a number of reasons. First, it greatly increases the amount
of arithmetic required to do the plotting, because each value of P must be
multiplied by / M . Second, each of the quantities a and i has an attached
uncertainty; if this were to be combined each time with the uncertainty in F,
we would have a falsely enhanced uncertainty for the compound quantity (for
example, in the actual experiment a would be measured only once, and its
uncertainty should not be combined with that of P as if every time we
measured P another measurement of a were made to combine with it).
Clearly, in this case the path of wisdom is to plot Q versus P and to use
na^ /8r|^ as the slope, thus avoiding all the difficulties just mentioned. We
are then able to calculate r\ from the value for the slope and the measured
values of a and i using
na
S ix slope
In general, it is best to plot variables that are as simple as possible and to
leave most of the arithmetic to be done just once in calculating the answer
from the slope.
Use of Compound Variables
In many cases it may suit our convenience (or else it may be absolutely nec
essary) to plot the graph using variables that are not single measured quanti
ties but are constructed out of the primary measurements. Consider, for ex
ample, the so-called compound pendulum, a rigid lamina of a certain shape
that is allowed to oscillate under gravity about an axis perpendicular to the
plane of the lamina, as shown in Figure 5-1 (a). The normal model of its os
cillation (for small angles of oscillation) gives the period of oscillation, T, as
Ih^+k^
T = 2 n i ---------
i gh
where T = period of oscillation (output variable)
h = distance from center of mass to point of support (input variable)
g = gravitational acceleration (constant and unknown)
k = radius of gyration about center of mass (constant and unknown)
Figure 5-1 T h e c o m p o u n d p e n d u lu m and its v a ria tio n o f o s c illa tio n

p e rio d w ith h.
Straight-line forms of this equation are not immediately obvious, but it is

clearly impossible to place it in the required linear form if we choose the
horizontal and vertical variables to be functions of h and T singly. Conver

sion into linear form by using compound variables is possible. Squaring both
sides of the equation, we obtain
^2 + k^)
gh
Therefore
A n\h^+k^)
T^h =
g
and
= ^ p h -k ^
47t'
which is now in linear form with
vertical variable =
horizontal variable = T^h
Q
slope =
4 tc'
and
intercept = -k^
Notice how in this case plotting in straight-line form is possible only by us
ing compound variables.
This example is worth studying because it illustrates very clearly the

superiority of linear analysis over other methods. A commonly encountered
approach to this experiment uses the graph of T versus h, which is shown in
Figure 5-1 (b). It turns out that the graph directly supplies only the value of k,
which can be obtained from the lengths of the intercepts AB and CD. If g is
required, it must be obtained by calculation from the value of k. The advan
tages of the linear form of analysis are clear. First, the T versus h graph gives
no basis for comparing the system with the model, unless one uses a com
puter to draw the graph of the function T(h). Second, no reliable estimate of
the uncertainty of the final answer can be obtained from this graph, whereas
the overall uncertainty can readily be obtained from the linear graph. Third,
the use of an intercept at such a low angle, as illustrated in Figure 5-1 (b), is
unreliable, because small changes in placing the lines can cause large
changes in the length of the intercepted portion; a linear form, on the other
hand, enables us to determine the slope of the graph reliably. Fourth, by using
the intercept method the answer is determined solely by a few points in the
vicinity of the intercepts, and we obtain no benefit from all the other points.
When drawing a straight line, however, all the points can contribute to the
choice of the line. Last, the linear graph gives g and k from almost
independent measurements on the graph, whereas with the other method any
inaccuracy in the value of k is propagated automatically into the value of g.
The use of compound variables can also be convenient when there are
two or more separate input variables. In such cases, even if the use of com
pound variables is not absolutely necessary for linear plotting (as it was for
the compound pendulum), they often provide the neatest and most effective
way of plotting graphs. It was mentioned earlier that, if a system involves
two independent input variables, we can study the variation of the output
variable with either input variable in isolation, while holding the other input
variable at a number of discrete levels. For example, if we wish to measure
the specific heat of a fluid, C, by continuous-flow calorimetry, we can allow
it to flow at a certain mass flow rate, m, through an electrically heated tube in
which the rate of heat generation is Q per unit time. The equation for the re
sulting heat balance (neglecting losses, etc.) is
Q = mC^T
where AT is the difference in temperature between the input and output ends
of the tube. Clearly, both Q and m are separately controllable, and we can
perform our experiment by studying the variation of AT with m, holding Q at
various fixed levels, or we can study the variation of AT with Q, holding m at
various fixed levels. We would then be able to plot either AT versus l/m, in
which case the various slopes would have values Q/C, or else AT versus 2,
which would give the various slopes as Cm. Neither possibility by itself
provides a complete description of the behavior of the system. Another
possibility exists. If we treat the product mAT as one variable and plot it
against Q, we obtain a single graph that summarizes all the information about
the system by incorporating both input variables simultaneously, whether we
did or did not control the values of m. The slope would simply have the value
C, and we would have a neat way of testing the model and obtaining our
unknown in one simple step.
Such use of compound variables is common, and, as before, the choice

of combination for the variables and the mode of plotting can be made to suit
both the convenience of the experimenter and the requirements of the ex
periment. If any difficulty in interpreting observations appears when they are
plotted by using compound variables (unexpected scatter, perhaps, or some

systematic deviation from linearity), we can always obtain extra information
about the system by reverting to a plot of the pairs of variables individually
rather than in combination. By isolating the effects of the separate input vari
ables we can usually identify the cause of any difficulty.
Logarithmic Plotting
It is frequently desirable and sometimes absolutely necessary to plot variables

in logarithmic form. For example, many physical processes involve ex
ponential functions of the form
where y and x are measured variables and a and b are constants whose values
are to be obtained from the experiment. The equation can be put in linear
form by taking logs of both sides to the base e. We obtain
log^ y = log^ a^-hx

Thus, if we plot log^ y vertically and x horizontally (known as a “semi-log
plot”), the model gives us a straight line. The slope gives a value for b, and
the intercept gives the value of log^ a. Notice that, if logs are taken to the base
10 instead of e, only the intercept is affected, and this can be convenient if we
are interested in the slope only.
The use of such logarithmic plotting is common because of the frequent

occurrence of exponential functions in the models of physical and chemical
processes. In addition, logarithmic plotting is used even for simple algebraic
functions. Consider, for example, the function
Taking logs of both sides, either to base 10 or base e, we obtain

log y = 2 log X
This equation is linear with
vertical variable = log y
horizontal variable = log x

and
slope = 2
Thus, functional dependence like a simple square can be tested by using such
logarithmic plotting (known as a “log-log plot”).
But what is the advantage of this type of graph over a plot, as we have
been recommending all along, of y versus One obvious answer is that it
allows us to plot on one piece of paper of reasonable dimensions variations
that are too extensive for traditional plotting. A range of one power of 10 in
our observations can be conveniently plotted on simple graph paper, a range
extending over a factor of 100 is difficult, and a factor of 1,000 makes satis
factory plotting impossible. For these very large excursions of the variables,
only logarithmic plotting allows realistic representation of the results.
A second advantage of logarithmic plotting concerns the power of the

function. If the system is behaving in such a way that the function
would be a better model than y = x^, that fact would probably elude us if we
plotted y versus x^. We would simply obtain a set of points that deviate from
a straight line, and the source of the discrepancy would not be immediately
obvious. The log-log plot, on the other hand, would still give us a straight
line, and this would tell us that some function involving a power was still a
good model. The slope of the line would not be 2, of course, and the
improved value of the exponent, 1.8, would be available from the slope of the
log-log graph. In Chapter 6 , we consider further the uses of log-log plotting
for the construction of empirical models involving such powers. It suffices
for the moment to note that, at the stage of designing an experiment, the
possibility of semi-log or log-log plotting should be kept in mind if either the
type of function or the range of variables suggests that either is appropriate.
5 -3 PLANNING AND EXPERIMENTS
We now list the actual, practical steps by which we prepare to do the experi
ment. These may seem tedious to those whose ambition is to get on with the
experiment as quickly as possible and worry later about what to do with the
results. Indeed, for many of the simple experiments commonly encountered
in teaching laboratories, the painstaking care that we are about to recommend
may seem pointless and pedantic. But remember that the simple experiments
in teaching laboratories are merely simulation in suitably simplified form of
the much more complicated and important situations that will be encountered
later in real systems. If in an introductory physics laboratory we forget to
measure the wire diameter while doing an elasticity experiment, it is probably
not going to matter too much. We can return to the laboratory later and
recover the offending measurement; even if we do not, the world will not
come to an end. But if ten years later we plan some space-based
astronomy with a very expensive telescope in a spacecraft and we notice
only when our experiment is in orbit that we have omitted to test the
optics properly, the consequences can be very serious. Acquire as early as
possible the habit of meticulous and painstaking planning of experiments
even if for the moment such planning may sometimes seem overmetic-
ulous and superfluous.
The planning steps are as follows:
1. Identify the system and the model.

2. Choose the variables.
3. Rectify the equation.
4. Choose the ranges for the variables.
5. Consider the overall precision of the experiment.
Identify the System and the Model
This step may seem somewhat trivial, but it is sometimes surprisingly

difficult to identify what, exactly, is the system we are studying.
Obviously, we cannot do an experiment if we are not clear, right at the
beginning, about the topic of the experiment. The actual phenomenon
under study is often surrounded by so much measuring apparatus that we
can lose sight of the fundamentals. If we encounter difficulty in
answering the question—What, exactly, is the system under study?— we
can try looking for the answer to the equivalent question— what is it
whose properties are described by the model?
Similarly, identify clearly the model and the limitations contained

within it. In the falling-ball experiment, for example, are we going to
worry about air resistance or not? If we decide to ignore the presence of
the surrounding air in the system, we are not being irresponsible— we are
merely defining one aspect of the model that will be used. Whether this is
a good decision will be made clear later by the experiment itself. If the
behavior of the system turns out to be in correspondence with the
behavior of the model at the level of precision we use, we can be satisfied
that there would have been no point in wasting time on small effects. If
we make a poor choice, the results of the experiment will very quickly
inform us of the necessity of reconsidering the matter. So at the outset we
decide on the limits of the system and the model, and we proceed to test
the situation.
Choose the Variables
Usually one quantity in the experiment is an obvious choice for an output

variable. If there is only one input variable, there is no problem. If there are
several input variables, try to identify one as the chief independent variable
and vary the others in discrete steps.
Rectify the Equation
The equation representing the behavior of the model must now be put into
straight-line form, as described in Section 5-2. As we have already men
tioned, there is no unique, correct choice for the straight-line form. Choose a
form that suits the purposes conveniently and effectively. For example, if the
equation contains some unknown quantity the value of which is to be deter
mined by the experiment, it is probably best to use a form for the straight line
that puts the unknown into the slope. It is possible to determine unknown
quantities from intercepts, but because intercepts can frequently be subject to
errors arising from instrument defects or other systematic errors, it is usually
preferable to obtain unknowns from slopes. If the equation contains two un
knowns, it is probably best to find a form that enables us to obtain one un
known from the slope and the other from an intercept.
Choose the Ranges for the Variables
Before starting the actual measurements, we should make decisions about the
ranges over which we hope to make them. It is usually best to plan on a range
for the input variable of at least a factor of 10. More is better, and less can
often give an unsatisfactory basis for comparing the behavior of systems and
models. Obviously, we cannot choose directly the range of the output
variables; the system itself will tell us these values, but we must still be
careful. There may be instrument limits beyond which damage can occur—
elastic limits, overheating of precision resistors, overloading of meters and
other instruments, and the like. Carefully made trial measurements will allow
us to determine the range of the input variables that will avoid overloading
any part of the system. This is the time to consider carefully all aspects of in
strument ratings, which can be particularly significant for electrical appara
tus. For example, does the resistance box have marked on it the maximum
current for each range? If so, we incorporate that limit into the choice of
range for the variables. If the limits for some piece of equipment are not
marked on the equipment itself, we must find the values in the manufacturer's
catalogue. In all cases we must ensure that limits are identified and observed.
It is too late when the smell of overheated insulation or a vertical column of
blue smoke above a meter alerts us to the frailties of physical apparatus and
the expense of replacement.
Consider the Overall Precision of the Experiment
We should not start an experiment until we have a general idea of the preci
sion we hope to attain in the overall result. This is not to say that we can
guarantee a final level of precision, but we should have a target figure to
serve as a guide for our choice of measurement methods. For example, the
request—Measure the acceleration of gravity using a pendulum—is by itself
virtually meaningless. In response to this request, we could spend ten minutes
with crude apparatus and obtain a result with a precision of 10%, or we could
spend weeks with refined, expensive equipment and attain 0.01%. We can
obtain a realistic impression of the expectation only from some request such
as: Measure g using a simple pendulum with a final uncertainty around 2%,
and try not to spend more than two hours on it. The figure 2% and the
specified time give a general idea of the kind of measurement we are being
asked to make and enable us to make sensible choices for experimental
method and care in measuring.
Whether or not it is specified in the requirement we have been given, we

should have such a target in mind for every experiment we do, for only then do
we have the basis for realistic design of the experiment. It will give us the
opportunity to ensure that all our measurements are of sufficient precision to
contribute usefully to the final result and that we do not waste time and effort
making some measurement with precision far in excess of that required.
To see how such a design could be carried out, we return to the exam
ple of the pendulum and the hoped-for value of 2 % for the final uncertainty
in the g value. We know that the result for g, although it will be obtained
graphically, in essence involves measurements of i and T (in the form T^). If
the uncertainty in any measurement of either £ or is in excess of 2%,
therefore, there is little chance that it will contribute usefully to a final de
termination of g within 2%. Suppose, as a first guess, we elect to restrict the
uncertainties in each of I and to fall below 1%. What are the implications
for the measurements of i and T\ The first step must be to make trial meas
urements to assess the absolute uncertainty with which we can make meas
urements of £ and T. Once we have determined these uncertainties, we can
find the limits on the ranges of the £ and T measurements that allow the
precision to be acceptable. Assume that with the apparatus available we are
fairly sure that we can measure lengths with an absolute uncertainty of ±1 mm.
What is the measured length at which this uncertainty corresponds to
precision of 1%? If the requirement is
^ = 0.01 in cm)
then the value of (, is given by
= 10 cm
Thus, so long as the measured lengths are greater than 10 cm, the contribu
tion to the overall uncertainty from ^ is within the acceptable range, and we
have identified one limit on the acceptable range of I .
What are the corresponding implications for the measurements of T1

If we are going to ascribe a precision of 1% to we need a relative
uncertainty of 0.5% in T. The period of oscillation is determined by
timing a specified number of oscillations with some kind of timer or
stopwatch, and the choice of measuring procedure is determined by the
basic uncertainty in that timing device. Suppose we are using a stopwatch
that (as we can determine by actually trying the measurement) allows us
to measure the time for a number of oscillations to within ±0.2 s. Notice
that this figure of 0.2 s must be the overall uncertainty in the whole
timing process, not just the uncertainty with which we can read the
stopwatch once it has stopped. We must include the complete sequence of
judging the pendulum’s position, pressing the button, and so on. The
resulting uncertainty in the overall timing process may substantially
exceed the simple uncertainty in reading the scale, and we shall probably
have to determine it for our particular circumstances by trying the
measurement several times. In any case, if we have an overall uncertainty
in the timing measurement of ± 0.2 s, we can calculate the relative uncer
tainty of any timed interval, t, as 0.2/r. This is the quantity that we wish
to restrict to values below 0.5%, and so the limiting condition is
0.2
= 0.005
The limiting value of t is therefore given by

0.2
t =- = 40s
0.005
Thus, provided we choose the number of pendulum oscillations so that we
are never measuring times less than 40 sec, we have a good basis for hoping
that our measurements of oscillation time will all contribute effectively to an
overall determination of g within 2%. We cannot guarantee that such an ex
perimental plan will result in a value for g with an uncertainty no greater than
2%; there is always the possibility of unexpected contributions to meas
urement uncertainty or of unsuspected systematic error. But at least we can

avoid making measurements that stand no chance at all of contributing use
fully to the overall result.
At this stage we must consider whether each measurement is going

to be considered in terms of uncertainty derived from personal estimate,
or whether random fluctuation is large enough to require the use of
statistical methods. If the latter, some trial measurements will allow us to
make a preliminary estimate of the variance and so enable us to choose
the sample size that will be required if we wish to attain a certain level of
precision. At this point we must recall the warnings about inaccuracy in
small samples that were given in Section 3-11. In addition, however, we
should remember that attempts to improve precision by increasing the
sample size can be unrewarding. The expression for the standard
deviation of the mean involves a/ A , s o that if a trial sample of 10
measurements suggests the desirability of, say, a tenfold increase in
precision, the sample size would have to be increased by a factor of 100.
A sample size of 1,000 may not be practicable, and we would have to
seek some other route to improved precision.
Whether the measurements are statistical in nature or whether they

have simple, estimated uncertainty, it should be possible at this stage to
decide if each measurement in the experiment can be satisfactorily made.
If it appears that some measurement is restricted to an uncertainty in
excess of the design aspirations, we must either obtain a more precise
method of measuring that particular quantity, or if that is not possible, we
must acknowledge that our former target for the overall uncertainty of the
experiment was unrealizable with the apparatus available, and that
revision of the target value is necessary. Also, by assessing the
contribution of each quantity in the experiment to the overall uncertainty,
we can identify any measurement that makes a dominating contribution to
the final result, either because of low intrinsic precision or because of the
way in which it enters into the calculation (e.g., some quantity raised to a
high power, or a quantity that has to be obtained as the difference
between two measured values). Once identified, these measurements can
be given special attention so that their uncertainty can be kept under
control as much as possible.
All the detail described in this section may seem to constitute an un
necessarily exacting approach to a small, simple experiment, but it is wise to
recall again that we are practicing for much bigger, more important experi
ments in which the consequences of failure to plan properly can be serious

and expensive.
Construction of Measurennent Program
After choosing the variables, ranges, and precision of measurement, it is best

to conclude the design of the experiment by constructing a complete and
explicit measurement program. This will normally take the form of a table
that includes all the quantities to be measured in the experiment and that also
provides space for any computations required for drawing the graphs. A
completed measurement program allows the experimenter to concentrate
during the course of the experiment on the actual conduct of the experiment.
While one is manipulating apparatus and making measurements, there is
usually enough to be done without the continuous necessity of deciding what
to do next. The measurement program also helps to guard against the acci
dental omission of some significant measurement that could be overlooked as
a consequence of the pressure of actual experimenting.
The complete process of experiment design is illustrated in the de

scription of a sample experiment in Appendix 4.
As we have remarked frequently, all this planning may seem like an

unnecessary amount of fuss for a simple experiment. These recommenda
tions, however, represent nothing more than the basic minimum of
preparation for any serious experimenting, and no opportunity should be
lost for the early formation of careful habits of experiment design and
planning. It is important to avoid the temptation to rush ahead with the
experiment, leaving until later the task of deciding what to do with the
results; it is much more beneficial to acquire the habit of setting aside the
time to design and plan an experiment properly before starting the actual
measurements.
5 -4 EXPERIMENT DESIGN W H E N THERE IS NO EXISTING MODEL
The problem of designing an experiment when there is no model appears

when, for example, we are making observations on some phenomenon that is
so new that a theoretical model has not yet been constructed, or else on some
system that is so complicated (e.g., a complex engineering system or some
aspect of national economics) that it will probably never be possible to con
struct a satisfactory theoretical model for it. If we do not have an existing
model to test, our objective in doing experiments on the system can take sev
eral forms. We may be motivated by simple curiosity or by a practical need
for information about the system. At a higher level, we may be interested in

the possibilities of model building. We may be seeking guidance for the
construction of a theoretical model, or if this is too difficult, we may wish to
obtain measurements to serve as a basis for a purely empirical model of the
system. As was mentioned in Chapter 4, even in the absence of detailed,
theoretical understanding, empirical models are extremely useful. They can
be helpful in systematizing our thoughts about a complex system, and they
are usually essential for such mathematical calculations on the system as in
terpolation, extrapolation, forecasting, and so on.
Whatever the motivation, we need a function or graph that provides

a good enough fit to the observations. The methods of finding suitable
functions are described in Chapter 6 . We restrict ourselves for the
moment to the question of designing the experiment. In the absence of an
existing model, experiment design can be relatively straightforward,
especially if we can isolate the input variables so that we can vary one
while holding the others at fixed values. Experiment design can consist
simply of measuring the output variable over suitable ranges of the input
variables to build up as complete a picture of the behavior of the system
as possible. If we cannot isolate the input variables, we have problems,
and this case is considered in Section 5-7.
Even if there is no existing theory for a phenomenon, it is wise to ac

cept any available hints about functions that might be appropriate to our
system and to test these possibilities against the system's behavior. One way
of obtaining such suggestions is discussed in the next section.
5 -5 DIMENSIONAL ANALYSIS
Even without a complete theory of a physical phenomenon, it is still possible

to obtain useful guidance for the performance of an experiment by dimen
sional 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 LT~\ acceleration
LT~^, density ML~^, force (equals mass x acceleration) MLT~^, work
(equals force x distance) ML^T~^ ând so on.
The principle used in dimensional analysis is based on the requirement

that the overall dimensions on the two sides of an equation must match. Thus,
if g is known to be related to the length and period of a pendulum, it is
obvious that the only way by which the LT~^ of the acceleration on the left
side can be balanced on the other side is to incorporate the length to the first
power (to give the L) and the period squared (to provide T~^). We can thus
say immediately that, whatever the final, theoretical form for the equation, it
must have the structure
^ length ^
g = (dimensionless constant) x
period^
The process can give no information about dimensionless quantities
(pure numbers such as 7i, etc.), and so we must always include their possible
presence in equations obtained by dimensional analysis.
The general method is as follows. Consider a quantity z that is assumed

to be a function of variables jc, y, and so on. Write the relation in the form
z OCXa y h
where a and b represent numerical powers, initially unknown, to which x and
y may have to be raised. Then write down the dimensions of the right-hand
side in terms of the dimensions of x and y and of the powers a and b. Second,
set down the condition that the total power of the dimension M on the right-
hand side must be the same as that known for z. Do this also for L and 7,
obtaining thereby three simultaneous equations that enable us to calculate
values for a, b, and so on.
For example, consider the velocity v of transverse waves on a string.

We might guess that this velocity is determined by the tension T in the string
(not to be confused with the T that appears as the dimensional symbol for
time) and the mass per unit length m. Let us write
r <uamI
The appropriate dimensions are;
of v: LT~'
of 7 (force): MLT'^
of m (mass per unit length): ML-'

Therefore,
L r ' ={M LT-^y{M L-')'’
= M “*'’ X ! “■ * X
Thus, by comparing in turn the powers of M, L, and T on the two sides of the
equation, we obtain
for M 0 = a-\-b
for L \= a -b
forT - \ = -2 a
of which the solutions are obviously

a = M2, and h = -{M 2 )
We obtain finally
-(dimensionless constant) x
Such a procedure is very valuable, for even in the absence of a de

tailed, fundamental theory, it provides a prediction regarding the behavior
of the system. This can be a starting point for experimental investigation.
If the experiment shows consistency between the system's behavior and
the model produced by dimensional analysis, we have confirmation of the
validity of our original guess regarding variables. If the experiment shows
a discrepancy, we must look again at our primary suppositions about the
quantities involved in the experiment. Notice that in the foregoing
example we obtained three equations for only two unknowns. The
situation, therefore, was really overdetermined, and we were fortunate
that the equations containing a and b were consistent. Had that not been
the case, we would have known immediately that our guess regarding the
constituents of v was wrong.
Powerful as this method is, difficulties obviously arise when the quan
tity under discussion is a function of more than three variables. We then have
more than three unknown powers but only three equations from which to
determine them. A unique solution is not possible, but a partial solution may
be found in terms of combinations of variables.
For example, consider the flow rate Q of fluid of viscosity coefficient r\

through a tube of radius r and length £ under a pressure difference P. All
these quantities are clearly significant in determining the flow rate, and so we
may suggest a relation
Q oc V '
The dimensions of the quantities are as follows:

Q (volume per unit time) L^r'
P (force per unit area) MLT'^ X = ML-'r^
1 (tube length) L
r| (viscosity coefficient,
defined as force per
unit area per unit
velocity gradient) (A/Lr-2)(L2)-'(Lr> X L-')-' =ML-'7’-'
r (tube radius) L
Therefore
L^r' = {M V ' r ^ y L \M L -'T -'y Lf
Comparing powers of
M 0= a +c
L 'i = -a -\-b -c -\-d
T -1 = - 2 a - c
Here we have four unknowns and only three equations, so that in general a
complete solution is not possible. We can obtain part of it, however, for it is
obvious that the M and T equations give us
a=I
c = -\
The equation for Q must therefore contain the term P ! r\. 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 It also contains so that
we can write
n
Q o z
P X r 3X
— - ^
r| \r
Because it is inconceivable that Q should increase with f , if all other quanti
ties are kept constant, it is obvious that b must have a negative value, and we
can invert the ^ / r term to obtain, finally.
p
Q cc — X X
The quantity b remains unknown, and this is as far as dimensional analysis

can take us toward the complete solution. Even this partial solution, however,
could serve as a guide to experimenting in a situation in which no fun
damental theory existed. Dimensional analysis can be extended to cover
thermal and electrical quantities, but in those cases ambiguities arise and they
require special consideration. The appropriate discussion can be found in the
standard texts on heat and electricity or in the specialized texts on dimen
sional analysis.
5 -6 DIFFERENCE-TYPE MEASUREMENTS
In all the preceding sections we have assumed that there was a clear and
definite relationship between the input and output variables, and that the in
put variables themselves were readily identifiable and relatively well con
trolled. We do, however, encounter circumstances in which we are not so
fortunate. Perhaps our input variables cannot be clearly isolated, so that, with
everything varying at once, it is difficult to identify the effect of each on the
output of the system. Or perhaps the system is so complex and subject to so
many variable factors that we find it hard to judge whether the effect in
which we are interested even exists. Many experimental techniques, mostly
of a statistical nature, have been devised for use in such circumstances. De
scriptions of these can be found in the texts on statistics listed in the Bibliog
raphy. For our present purpose, we restrict ourselves to a brief description of
the problems appearing at the various levels of complexity and uncertainty.
Difference-Type Experimenting in the Physical Sciences
Suppose we wish to study some relatively small effect, such as the extension
of a hard steel wire under load. Not only is the effect small, but it also is
subject to a number of perturbing factors—for example, temperature. If we
simply measure, therefore, the extension of a particular wire under a certain

load without ensuring temperature stability, we cannot be certain that the
extension we measure can be ascribed uniquely to the influence in which we
are interested, namely, load. Not only that but if in addition we are
actually unable to control the temperature, we shall never be able to be
sure about the effect of load on the wire. The solution is a null-effect
measurement. We study two identical specimens simultaneously, one
loaded, the other not, and we measure the difference in length between
the two specimens. The wire under load shows the behavior of the system
as it is tested, while the unloaded wire provides the null effect—that is,
the behavior of the system in the absence of the load. We can then hope
to ascribe the measured difference in length to the influence in which we
are interested, the load, and the perturbing influences that affect the two
wires equally are prevented from introducing errors into the
measurements. We must obviously try to ensure that, as far as possible,
the two specimens be identical, be subject to exactly the same influences,
such as temperature, and differ in only the one respect— load.
Fortunately, such correspondence is not too hard to achieve if we are

talking about steel wires. We can come close to making the situation of the
two wires identical by mounting them close together (to minimize tempera
ture differences between them) and by taking other similar precautions. And
because we wish the basic properties of the two specimens to be as close to
identical as possible, we can simply take one length of wire and cut it in two,
making one piece the specimen to be loaded and the other the comparison
specimen that will indicate the null effect. Our ability to cut the specimen in
half allows us conveniently to perform a great variety of difference types of
measurements and to obtain high precision in the detection of small effects
that would otherwise be hopelessly obscured by perturbing factors. Such ex
perimenting is common and is encountered over the whole range of physical
phenomena.
It is always good experimental practice to check the performance of an

experimental system in the absence of the influence we are studying as well
as in its presence. Sometimes the results are surprising, and we do well to
take the advice of Wilson (see the Bibliography) and reflect on the statement:
It has been conclusively proved by numerous tests that the beating of drums
and gongs during a solar eclipse causes the sun's brightness to return.
Difference-Type Experimenting in the Biological Sciences
In illustrating null-effect measurements using the extension of a loaded steel

wire, we have encountered one very convenient aspect. To guarantee the
similarity of the experimental specimen and the comparison specimen, we cut

the basic specimen in half. In the case of steel wires and other similar
materials, that presents no problem, but other systems are not so cooperative.
Suppose we wished to measure the effectiveness of a new drug for a
particular type of illness. It would clearly be of little value if we did
nothing other than simply administer the drug to a patient suffering from
the disease and watch for improvement. There are far too many variable
and perturbing factors for us to ascribe confidently any change in the
patient’s condition to the drug. If we wish to isolate the effect of the drug
alone, we should clearly try to design some sort of difference-type
experiment in which we observe the null effect as well as the influence of
the drug. Such a requirement raises obvious difficulties not encountered
when experimenting on steel wires. The reluctance of most human
specimens to be cut in half makes it impossible to create a genuine null-
effect specimen. We could use a second person as a null-effect specimen,
but we would immediately encounter all the variability of response that
we had sought to evade by using identical specimens.
Faced with the inevitability of biological variability, our only

recourse is to compensate with increased numbers. We abandon attempts
to experiment on single specimens and construct an experimental group,
which we expose to the influence under study, and a control group. The
control group is constructed to be as closely comparable as possible to the
experimental group, differing only in that it does not receive the
treatment that forms the topic of the research. It will, we hope, be
exposed to all the perturbing influences that affect the experimental
group, will respond to them in the same way as that group, and will
therefore provide the null-effect measurement.
Many refinements may have to be built into this kind of

experimenting, because the effects we seek to measure can often be quite
small in comparison with all the perturbing influences. For example, to
diminish subconscious distortion of the results in medical experimenting
on human subjects, it is common to offer the members of a control group
a simulation of the real material given to the experimental group (a
placebo), and to keep both the experimenters and the subjects in
ignorance of the allocation of real and simulated material (the so-called
double-blind experiment).
Experiment designs involving an experimental group and a carefully

matched control group are virtually universal in biological studies,
whether we are trying to measure the possible carcenogenicity of some
food dye in large numbers of unfortunate mice or the beneficial effects of
musical activities on the academic achievement of elementary school

students.
5 -7 EXPERIMENTING W ITH NO CONTROL

OVER INPUT VARIABLES
Sometimes we have to design a process to study some system over which we

have no control at all. If this is the case, we have no alternative to simple
unmanipulative observation of the system, and our task is to design the ob
servational procedure (perhaps we are not justified in calling it an experi
ment) to optimize our chances of effective comparison between the properties
of the system and those of any model we have in mind. In cases of clear-cut
behavior of the system and well-defined models, we may not have too much
of a problem. For example, astronomers may suffer the frustration of inability
to influence their subject matter, but their system usually functions in a well-
defined manner, often permitting extremely accurate measurement. In this
way it is not too hard to decide that Einstein's theory of general relativity fits
the observations on the orbit of the planet Mercury better than does Newton's
theory of gravitation.
In other cases, however, the questions we ask may be harder to answer.

For example, has the introduction of a new detail of manufacture altered the
quality of a manufactured product? Even when everything in the manufactur
ing process is kept as nearly constant as possible, observation shows that the
product varies from specimen to specimen. Does this variance mask the ef
fect in which we are interested? Without control over our input variables, the
study becomes an exercise in sampling procedures, and a whole field of in
dustrial study exists under the title of “quality control.” The literature on sta
tistics and statistical experiment design is extensive; some of the texts listed
in the Bibliography provide a starting point.
Even industrial processes, with their inherent fluctuations and their lack
of input control, pose problems that are simple in comparison with some of
the questions to which we seek answers today. Does the addition of fluorides
to municipal water supplies improve the condition of peoples' teeth, and does
it have other, possibly harmful effects? Do nuclear power stations cause a
higher incidence of leukemia in their vicinity? In seeking the answers, we
have almost every problem that can face an experimenter. There is little or no
control over the input variables, there is wide variation in individual re
sponse, the response may be of only a probabilistic nature, there may be long
delays in observing a response, there is rarely an opportunity to observe a
genuine null effect (we do not normally carry out surveys of sufficient sensi
tivity before the municipality starts to add fluorides or before the nuclear
power station is built), and there is commonly a multitude of confusing ex
traneous factors. The only thing we can do is to carry out the sampling pro
cedure as carefully as possible. We must obtain an artificial null-effect
measurement by constructing an experimental group as large as possible that
is under the influence we are studying and a control group that is exempt
from that influence but that in every other respect matches the experimental
group as closely as possible.
The whole point in this kind of experimenting or survey work lies in

the skill and care with which the sampling is done. The effects under
study are usually so subtle that, as a consequence of no more than
changes in sampling procedure, it is not uncommon for different surveys
to provide completely contradictory conclusions. It is not completely
unknown that people with special interests in mind can supply results of
surveys to “prove” their point, obtaining the result they want by careful
control over their sampling procedures. Many of the issues in which
scientific matters have a bearing on public policy have this characteristic
of uncontrollable input variables, and we should all become as familiar as
possible with the procedures used for sampling and significance testing.
In this way we may be able to judge as accurately as possible the usually
conflicting claims of the protagonists.
When faced with problems of such complexity, we must frequently

abandon familiar patterns of thought that have been successfully used in
other areas. For example, the word proof is legitimately used in many con
texts. We can, for example, prove mathematical results as consequences of
mathematical principles. The word is also used (perhaps less legitimately)
with reference to measurements when the uncertainty level permits. Most
reasonable people would accept it as “proved” that the sun is more distant
from the earth than the moon (although it would be better to say simply that
the distance is measured to be greater). But there are other areas in which we
cannot use the word at all. We have all heard claims from some of the inter
ested parties that the evidence linking cigarette smoking to lung cancer is
“only statistical” and that harm has not been “proved.” This is a common
form of argument in such matters of public policy. Situations in which such
controversy exists are usually difficult to deal with, partly because the ob
servable effects may appear only in terms of probabilities, and also because
of long delays in the appearance of the effects. In such cases the concept of
proof must be modified. It is actually replaced by the concept of correlation.
Correlation studies give results, phrased in terms of probabilities, that differ
in character from the clear-cut cause-and-effect relationships with which we
are familiar in other experiments. Nevertheless, they can be equally valid for
identifying the factors that influence systems. The concept of correlation re
ceives further consideration in Section 6-14.
PROBLEMS
1. A scientist claims that the terminal velocity of fall of a parachutist is depen

dent on only the mass of the parachutist and the acceleration due to gravity.
Is it reasonable setting up an experiment to check this suggestion?
2. The range of a projectile fired with velocity v at angle a to the hori
zontal may depend on its mass, the velocity, the angle, and the gravita
tional 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 the moment of inertia
of the oscillating body. What is the form of the function?
5. The deflection of a beam of circular cross section supported at the ends
and loaded in the middle depends on the loading force, the length be
tween 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
that should be plotted to check the suggested variation and state how the un
known (slope, intercept, etc.) may he found.
6. The position of a body starting from rest and subject to a uniform accel
eration is described by the function
s = O.Saf^
s and t are measured variables. Determine a.
7. The fundamental frequency of vibration of a string is given by
1
/=
2 e \m
/, 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
V=
Vand P are measured variables. Determine p .

9. A conical pendulum has a period given by
^cosa
T = 2n
g
T and a are measured variables, i is fixed and known. Determine g.
10. The deflection of a cantilever beam is expressed by
" = ------ r
Yab^
d, W, and (. are measured variables, a and b are fixed and known. De
termine Y.
11. The capillary rise of a fluid in a tube is given by
PS®
h and R are measured variables, p and g are fixed and known. Deter
mine a .
12. The gas law for an ideal gas is
pv=R T
p and T are measured variables, v is fixed and known. Determine R.
13. The Doppler shift of frequency for a moving source is given by
/ = / o -------
v-Vo
/an d Vq are measured variables, is fixed and known. Determine v.
14. The linear expansion of a solid is described by
^ = ô(l + a-A T)
i and AT are measured variables, is constant but unknown. Deter
mine a .
15. The refraction equation is
sin 0 , = rij sin 02

0, and 02 are measured variables, is constant and known. Determine
16. The thin-lens (or mirror) equation can be written
1 1-1
5^ and s' are measured variables. Determine /. There are two ways of
plotting this function. Which is better?
17. The resonant frequency of a parallel L-C circuit is given by
1
CO =
JT c
CO and C are measured variables. Determine L.
18. The force between electrostatic charges is described by
1 ^ 1^2
F=
47T8o
F and r are measured variables. and ^2 fixed and known. How do

you check the form of the function?
19. The force between adjacent current-carrying conductors is described by
F=
An
F, /,, /2, and r are measured variables, pg and i are constant. How do
you check the form of the function?
20. The discharge of a capacitor is described by
Q and t are measured variables. R is fixed and known. Determine C.

21. The impedance of a series R-C circuit is given by
1
Z = AR^+-
co'C'
Z and CO are measured variables. Determine R and C.
22. The relativistic variation of mass with velocity is
m=
m and v are measured variables. Determine and c.

23. The wavelengths of the lines in the Balmer series of the hydrogen
spectrum are given by
X U n-
X and n are measured variables. Determine R.
Experiment Evaluation
6 -1 GENERAL APPROACH
Even when we have finished making the measurements in an experiment, an

equally significant part of the process still remains—we must evaluate the
significance in what has been done. Our objective in doing an experiment is
to be able to make some statement, usually about the relationship between
some system and a model. It is important to identify clearly the statement we
wish to make and to ensure that the statement is as accurate and complete as
possible and fully justified by our measurements. The precise way in which
we evaluate the experiment as a whole depends on the type of experimenting
we have been doing. As described in Chapters 4 and 5, we may have been
operating with or without a theoretical model, and our measurements may or
may not have been dominated by statistical variance. The procedures we
must now follow will vary accordingly.
Before we proceed, we must note two general points. First, we should

always remember that experimental results are precious. They have often
been obtained from an extensive experimental program involving many
people and large amounts of money. At any level of cost, the results may be
unique and irretrievable. We should accept the obligation to extract every
available bit of information from the observations and to ensure that our final
statement is as complete as possible. The second general remark concerns
objectivity. It is almost impossible to avoid approaching an experiment with
out some preconception of what “ought” to happen. We must discipline our
selves to be as objective as we can, and if the outcome of the experiment is
113
114 EXPERIMENT EVALUATION CHAP. 6
different from what we expected or hoped for or is disappointing in some

way, we must be prepared to state the result honestly and realistically and to
obtain from it the guidance required for future work.
In the teaching laboratory, where it is sometimes difficult to keep our

ultimate objectives clearly in mind and easy to forget that the experiments
serve to simulate real tasks in the working world, we commonly encounter
the mistaken belief that the objective is to reproduce the known values of ex
perimental quantities. If we measure the acceleration of gravity and obtain a
value of 9.60 m S'^, our answer is different from the “right” answer and so we
are “wrong.” The “error” can then be conveniently blamed on the apparatus.
Because there are no “right” answers for experimental quantities, the
situation really involves the comparison of two measured values for a quan
tity. Each measured value has its own characteristics, and each has its own
range of uncertainty. To assess the significance of a discrepancy between two
independently measured values of a quantity is actually a complex and
difficult task. It is far better first to develop the ability to make our own
measurements as reliably as possible and to assess their range of uncertainty
as accurately as possible; we can worry later about comparing our measure
ments with those of other people.
So when we make measurements on quantities for which we are sure

we already know a more precise value, it is best to discipline ourselves to
avoid thinking about the more precise, or “standard,” value; it is better to ac
quire experience and build up confidence in our own work. This confidence
will be necessary later as we undertake professional experimenting, in which
we must take responsibility for the experiments and measure things that have
never been measured before.
So if we obtain 9.60 m s"^ for g, we must be equally aware that the

measured uncertainty is ±0.3 m s"^ and that the result is not as bad as we
might think at first. If we are going to grumble about anything, let it be the
±0.3 m s"-, but we must not feel guilty about it if the experimental apparatus,
with normal effort, is not capable of precision better than 3%. We must not
be misled by the way in which accepted values for physical quantities are
quoted in textbooks. The values are often mentioned rather casually, and the
texts rarely make it clear that these numbers represent the outcome of so
phisticated work by generations of expert scientists. It is instructive to read
the detailed history of such measurements. Excellent accounts of some of
them are in the book by Shamos listed in the Bibliography. We should not be
too casual about numbers such as these and should not hope to reproduce
them exactly in two hours of work in an elementary laboratory.
CHAP. 6 EXPERIMENT EVALUATION 115
The main point is to state the result of the experiment honestly and ob
jectively. The experimenter should strive earnestly to maximize the yield of
the experiment by making the final answer as reliable as possible and the
limits of uncertainty as close as the experiment will permit, but in all cases it
is important to be realistic.
6 -2 THE STAGES OF EXPERIMENT EVALUATION
The process of evaluating the result of an experiment has several parts. First,
we must obtain the values of the basic measurements and their uncertainties.
Second, we must assess the degree of correspondence between the properties
of the system and of the model. Third, we must calculate the values of what
ever property of the system we set out to measure. Last, we must make an
estimate of the overall precision of the experiment. Let us consider each of
these steps in turn.
Computation of Elementary Quantities
The first step in working out the result of an experiment consists of calculat
ing the elementary quantities of which the experiment is composed. For ex
ample, an experiment on a simple pendulum that has the purpose of obtaining
a value for g will probably yield, as its input variable, a set of measurements
of length i . The output variable will be presented by a set of measurements
of the times required for a certain number of oscillations, and from them
values of the period T can be calculated. Our present purpose is to compute
the values of i and T and their uncertainties; these will be the basis of the
subsequent graphical analysis. The choice of procedure here depends on
whether we have elected to make a subjective assessment of the uncertainty
range of each measurement or have decided that random fluctuation is
sufficiently prominent that statistical treatment is desirable.
Estimated Uncertainty
In the case of the simple pendulum, the first variable to consider is I . Here
we may have found that measuring the length of the pendulum with a meter
stick has enabled us to identify intervals, as described in Section 2-3, within
which we are almost certain our values lie. Our experimental results will
therefore take the form of a set of values for ( in the form: value ± uncer
tainty. It is conceivable, too, if we have been counting swings and measuring
the times with a stopwatch, that we are similarly able to identify intervals on
the time scale within which we are almost certain that our values for time lie.
These, too, would be expressed as time value ± uncertainty. This, however, is
not yet our variable T. We might have counted 15 oscillations of the pendulum,
obtaining a time value of 18.4 ± 0.2 s, and the value for the period, the time
required for one oscillation, must be obtained by division as 1.227 ±0.013 sec.
Notice that not only the central value must be calculated in this way but also the
value of the uncertainty. In simple algebraic terms
(18.4 + 0.2) = — ± — = 1.227±0.013

15 15
Do not ignore this kind of significant modification of the uncertainty value; it
is necessary whenever any arithmetic process is carried out on the basic
measurements.
The end result of this experiment will be a set of f and T values, com
plete with uncertainties, and we shall then be ready to start drawing our
graph.
Statistical Uncertainty
If repetition of our measuring process has shown random fluctuation in

one or more of the variables, we may have decided, as described in Sec
tion 5-3, to take a sample of readings, the number of readings being cho
sen on the basis of the apparent magnitude of the scatter to give the pre
cision we require. Because we must reduce the resulting set of readings to
a form suitable for plotting, we must express the sample in the form: cen
tral value ± uncertainty. As described in Section 3-10, the most suitable
form to choose is usually the sample mean and the standard deviation of
the mean because of the readily identifiable significance of these quanti
ties. Provided we make it clear in our report that we are quoting sample
means and standard deviations of means, everyone will understand that
we are specifying intervals that have a 68 % chance of containing the uni
verse mean.
While we are making these claims about the numerical significance

of our measurements, we must remember the warnings given in Section
3-5. The measurement samples encountered in the work of the physics
laboratory are frequently too small to permit any assessment of the actual
frequency distribution of the universe from which the measurements were
taken. We are therefore making an assumption when we ascribe the nu
merical properties of the Gaussian distribution to our sample. It is usually
a good enough assumption, but we should remember that it is an only
assumption.
At this point, remember also the warnings about a estimates from

small samples that were given in Section 3 - 11 and check that the computa
tions are significant. In general, it is not worth using a statistical approach
with fewer than lO observations; for some particular purposes many more
may be required.
It is useful to think in advance about the interpretation of the uncer

tainty regions on the graph. If both variables in the experiment have similar
statistical character, the mean and standard deviation of the mean for each
point will enable us to draw, for each point on the graph, a little rectangle
whose interpretation will be clear. We may have a little more of a problem if
the experiment has yielded variables of two different kinds. It is quite con
ceivable in, say, the experiment on free fall under gravity, which was used as
an example in Section 4-2, that one variable, the distance of fall, will have an
estimated uncertainty and the other will require a statistical treatment
yielding standard deviations of the mean. If we were to plot values derived
from these two different types of treatment, the uncertainty ranges along the
two axes would be different. The uncertainty interval in one dimension would
give almost 100% probability of containing the desired value, whereas the
probability in the other dimension would be only 68%. It would be difficult
to know how to interpret the graph, and it would be better to bring the two
variables into better correspondence. Remembering that a range of twice the
standard deviation of the mean gives us a 95% chance of including the
universe value, we can use 2S„, as our uncertainty for the statistically treated
variable, thus giving a range of uncertainty for each point on the graph with
roughly the same significance in both dimensions.
At this stage, by one process or another, the measurement of every

quantity in the experiment will have been reduced to a central value and its
uncertainty, but we are not yet quite ready to start drawing the actual graph.
If the graph is to be drawn with one variable on one axis and the other vari
able on the second axis (like load vs. extension for a spring or current vs.
potential difference for a resistor), then we can proceed directly. If, however,
the process of rectifying the equation for the model has led to a choice of
more complicated variables for plotting (such as T~ vs. (. for the simple pen
dulum, or h~ vs. T~h for the compound pendulum, etc.), we must construct
these variables by some process of arithmetic computation. We obviously have
no problem in performing such simple arithmetic calculations, but we must not
forget that the uncertainty values also must be recalculated. If we are going to
plot values of T~ on the graph, the uncertainty bars or rectangles must give the
actual interval over which T~ itself is uncertain. All such computed quantities
must be provided with their own uncertainty intervals, and only then are we ready
to start drawing the graph.
6 -3 GRAPHS
Whether the graph is to be merely an illustration of the behavior of a physical

system or whether it is to be the key to assessing the experiment and cal
culating the answer, the aim is to set out the results in such a way that their
characteristics are displayed as clearly as possible. This will involve appro
priate choices of scale, proportions, and so on. First, ensure that the graph
paper is large enough. It is a waste of time to plot observations having a pre
cision of 0.1% on a piece of graph paper 18 cm x 25 cm, where a typical
plotting uncertainty is perhaps 2%. As we shall see later, valuable informa
tion will be lost unless the uncertainties on the points are clearly visible, aad
so it is necessary to make sure that the graph paper is big enough. Second,
make the graph fill the available area. This can be done by choosing the
scales so that the general course of the graph runs diagonally across the paper
and by suppressing the zero if necessary. When 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 little information.
There are times when it may be important to preserve the origin as part
of the graph. It may be desirable or even necessary to examine the behavior
of the graph at the zero of one or both axes. At other times, for purposes of
illustration, it may be useful to show clearly the scale of some variation in
relation to its zero value. However, for the purposes of the graphical analysis
with which we are here concerned, it is generally best to make the graph fill
the graph paper.
The method of marking each measurement on the graph paper depends

to some extent on preference. One essential feature is to make sure that the
range of uncertainty is clearly indicated. Only if this is done can the process
of comparing the behavior of the system and the model have any meaning
and the uncertainty of any future calculations of slopes, be assessed. To each
point on the graph we can attach a cross with horizontal and vertical bars to
indicate the range of uncertainty, or we can make each point a little rectangle
surrounding the measured value and indicating by its horizontal and vertical
dimensions the range of uncertainty in each coordinate. So long as the ranges
of uncertainty are clearly indicated, it may not matter which method we
choose; the important thing is to acquire the habit of marking uncertainties on
every graph. It is also important to note on the graph itself, or in its caption,
the nature of the uncertainties, whether estimated outer limits of uncertainty,
statistical uncertainties of 15 or 25, or other such information. It can be very
frustrating when trying to judge the significance of a graph if we have to
search through the text to find out what the uncertainty marks mean. If
several graphs are to be plotted on one piece of paper, make sure they are
clearly distinguished by the use of some different symbol or by color or by
some other means.
6 -4 COMPARISON BETWEEN EXISTING MODELS A N D SYSTEMS
Once all our observations have been plotted on the graph paper, we are ready
to proceed with the next stage—the comparison between the properties of the
system now displayed before us and the properties of any models we have
available. The procedure depends on circumstances, and we describe the
various situations in turn. In all the following we assume that, on account of
the difficulty in representing nonlinear properties of models on hand-drawn
graphs, we have chosen or rearranged our variables so that the graphs take on
linear form.
Let us suppose, first, that we have a model that is fully specified and
that has no undetermined quantities. The purpose of the investigation would
then be only to see how well the properties of the model match the properties
of the system. To do this we would simply draw on the graph, using the same
scale, the graph of the function that represents the properties of the model. A
typical case was illustrated in Figure 4-10, in which observations of the time
of fall of an object as a function of distance are compared with the behavior
of the analytical expression
t = QA5\5x''^
which represents the theoretical model of the situation.
But how are we to judge the degree of correspondence? This is

where the presence of the uncertainty intervals becomes of dominating
importance. If we simply plot points without uncertainty bars, the inevi
table scatter in the points would mean that the probability of the line that
represents the model's properties actually passing through even one (not
to say more than one) of the points would be vanishingly small. So how
can we say anything sensible about the outcome of the comparison? If,
however, the points on the graph represent intervals of possibility for the
location of the plotted values, it becomes possible to make logically satis
factory statements. If, as was the case in Figure 4-10, the line represent
ing the model passed through the region of uncertainty of each point, we
could say just that. Notice again that this does not mean that we have
“proved” that the equation is “true,” or “correct,” or whatever. All we can
say is that the model and the system are “consistent,” or “in agreement,”
or “compatible,” or some such phrase. As we have said before, we must

make sure that we use the right language, for otherwise we may misrepre
sent the situation and have a good chance of misleading people. Notice
also that we must be careful to say that we have found “correspondence,”
“consistency,” “agreement,” or whatever between the model and the sys
tem only at the level of precision of our experiment. Nothing in our proc
ess entitles us to ignore the fact that, at a higher level of precision of
measurement, discrepancies might appear that were undetectable at the
level of precision in our experiment.
Now that we have considered the case in which a model and a system
turned out to have properties that are undistinguishable at the level of preci
sion involved, we must consider the other possibilities in which the proper
ties of the model and the system do not overlap completely.
No Detectable Discrepancy
This is the case we have already considered in detail. It is illustrated in Figure

6-l(a).
Correspondence over Part of the Range
Sometimes a model provides a satisfactory description of a system, provided

the value of some variable does not exceed or fall below some limit. In this
case the graphical comparison would appear as in Figure 6-l(b ) or 6-l(c).
An example of case (b) would be the flow of a fluid through a pipe, in which
the proportionality between flow rate and pressure head is satisfactory only
below the onset of turbulence. Figure 6-l(c ) could be a representation of the
variation with temperature of the resistivity of a metal, for which the linear
model breaks down at low temperatures.
In any case that comes within this category, we would state the re
sult of the comparison using some phrasing such as: We observed agree
ment (compatibility, consistency, etc.) between the model and the obser
vations only over the range so-and-so. Or: The properties of the model
and the system are observed to diverge significantly after the value such
and such. Notice again that we must resist the temptation to think that
something is “wrong” because we do not encounter complete correspon
dence between models and systems over the whole range. Both models and
systems exist in their own right, and we cannot prejudge the extent to which
their properties overlap. In fact, the detection of the limits on the validity of a
particular model can furnish important clues for its improvement.
Figure 6-1 T h e c o m p a riso n b e tw e e n sy s te m s and m o d els.

Intercepts
A frequently encountered circumstance involves intercepts. The graph of the

model's behavior may pass through the origin, but the observed behavior of
the system may not, as illustrated in Figures 6-l(d ) and 6-l(e ). Such a dis
crepancy can arise from many different types of mismatch between the model
and the system, and information about such intercepts can be helpful in
analyzing experimental situations. It is usually advisable, when drawing
graphs, to check the behavior of the model and system at the origin; reference
was made to this point in Section 6-3 on graph drawing. As we saw in
Section 4-5, the graphical analysis of an experiment is invaluable for ena
bling us to obtain answers that are free from the systematic errors associated
with unexpected intercepts. Even with this protection, however, it is usually
best to know whether an unexpected intercept exists so that we can check the
overall degree of correspondence between the model and the system.
Unexpected Scatter of Points
As was described in the chapter on experiment planning, we should have

carefully judged the uncertainty of our measurements before starting the ex
periment, and in the light of our target value for final precision, we should
have made appropriate choices for our measurement methods. If we have
done this satisfactorily, we shall find, on plotting the graph, that there is
consistency between the scatter of the points and the uncertainties of the
measurements, as illustrated in Figure 6-l(a ). However, things do not always
work out as we wish, and we not uncommonly find ourselves in the situation
illustrated in Figure 6-l(f). It results, simply, from the presence of factors in
our measurement methods that we failed to identify as we made the initial
assessment of the uncertainty of the measurements.
We should not be content to leave the situation like this. It is worth

checking the apparatus in an attempt to discover the cause of the fluctuation.
It can be something as simple as a loose electrical connection or failure to stir
a heating bath, and it is always satisfying to see such a discrepancy disappear.
If for any reason it is not possible to keep the experiment going and take
steps to reduce the scatter, it may be necessary to work with the results as
they are and make the best statement we can about the degree of corre
spondence between the model and the system. We might be able to say
something like: The observations are distributed uniformly about the line
representing the model. For cases in which we have to obtain numerical in
formation from lines drawn using such experimental measurements, see Sec
tion 6-7.
No Correspondence Between System and Model
We rarely encounter circumstances in which the behavior of the system bears

no resemblance at all to the behavior of the model [Figure 6-l(g)]. If every
thing in the experiment is working as it should, this is a most unlikely out
come. Models may be in principle inadequate representations of the behavior
of the physical world, but they would not be models if they were as bad as we
are suggesting. Such complete failure of correspondence points clearly to an
actual error in the experiment. It can be an error of interpretation of the
variables, a mistake in the rectification of the equation, an error in setting up
the apparatus, or a mistake in making the observations, in calculating the re
sults, or in plotting the graph. If possible, go back to the beginning, check
everything, and start again. If it is not possible to check the instruments used
in the experiment, check for errors in all the analytical and arithmetic proc
esses. If every attempt to discover an error fails, state the outcome of the ex
periment honestly and objectively. There is always the chance that we have
discovered something new. In any case, if we are truly baffled by some fail
ure of correspondence between a well-checked piece of apparatus and a reli
able model, an honest statement of the situation is bound to be of interest to
other people.
In all the foregoing we have stressed one important point: We must not
think that an experiment is giving us a “right” or a “wrong” result. We just
carry out the experimental process as carefully as possible and then state the
outcome as honestly and objectively as we can. It is not a bad thing to be
reminded from time to time that models may provide only partially satisfactory
representation of the behavior of systems. It is most important for us to know the
hmits for the vahdity of models, and the manner in which models fail can furnish
invaluable evidence to those who seek to improve them.
6 -5 CALCULATION OF VALUES FROM STRAIGHT-LINE ANALYSIS
In all the preceding sections, we have dealt with models that were completely
specified, including the numerical values of all quantities. The purpose of
experiments was simply to compare the behavior of the system and of the
model. As was considered in Section 4-5, however, it is also possible, indeed
very common, to use straight-line analysis in determining for some quantity
in the model the numerical value that is appropriate for our system. In such
cases the model is not fully specified, because it contains a quantity or
quantities of initially unknown value. It is not possible, therefore, to draw a
graph for the model to compare with the points, for the graph initially con
tains nothing but the points alone, as shown in Figure 6-2(a).
(a)
Suppose that we have measured the values of current through, and the
potential difference across, a resistor and we wish to test the observations
against the model
V = IR
In the absence of a specified value for R, the behavior of the model is repre
sented by all the lines on the I-V plane that have the equation
V = constant x /
where the constant can take all values from zero to infinity. In principle, we
can simply draw all these lines on the same graph as the measurements, and
determine first the extent to which the behavior of the system and of the
model overlap. Second, from the bundle of lines that fall within the regions
of uncertainty of all the points, we can determine the range of R values that
are appropriate for our system (as illustrated in Figure 4-11). For our present
purpose, it is not quite as simple as that because, on the basis of the measured
values shown in Figure 6-2(a), we have no right to prejudge the behavior of
the system at the origin. It is best to leave the question of intercepts to a later
stage and simply decide on the range of straight lines that is consistent with
the observations.
There are several ways of doing this. The most satisfactory, a statis
tical method, is described later. In the meantime we content ourselves
with simpler, mechanical procedures and carry out the time-honored
practice of drawing the best straight line through the points. To do this by
eye requires some mechanical aid that does not obscure half the points.
An opaque ruler cannot be used, but a transparent straight edge is accept
able. Probably the most satisfactory aid is a length of dark thread, which
can be stretched over the points and easily moved until the most satisfac
tory position is found. If difficulty is encountered in judging visually the
trend of a set of points, it is often helpful to hold the graph paper at eye
level and sight along the points. This makes the clustering of the points
around a straight line or a systematic deviation from a straight line much
clearer than in the direct view.
We can profitably identify several straight lines. The “best” straight

line (whatever we mean by “best ”) is one obvious candidate. In addition we
can make a guess at how far we can twist that “best” line in either direction
until it can no longer be regarded as an acceptable fit to the points. These two
extreme positions supply us with a value for the uncertainty in the slope. If,
on account of wide scatter in the points we find it difficult to identify a “best”
line and its uncertainty limits, it is sometimes helpful to remember that the
points and their uncertainties are really a sample from a whole band of values
on the plane. The occupation of this band by the measurements may be spotty
on account of the limited number of observations, and this can make it
difficult to choose the lines. If this is the case, it is often helpful to imagine
the band to be populated by the million or so readings that we could have
made with the apparatus. We can then try to guess from the graph where the
center and the edges of that band might be, and that will enable us to make
our choice of lines. In Figure 6-2(b) we could have chosen AB as our “best”
line, and we could have decided that the lines CD and EF would contain
almost all the infinite universe of points. The lines CF and ED (not drawn)
would then represent, respectively, the steepest and least steep slope of the
set of lines that are consistent with the observational points.
Once we have chosen the lines, we can set about determining the nu
merical value of their slopes so that we can obtain the answer we want, such
as, in the case of the V = IR example, the value of R. For our purpose, the
question of slope has nothing to do with the angle made by the lines on the
graph paper; we are talking about intervals of the measured variables I and V,
and so the slopes must be calculated analytically. For a line such as AB on
Figure 6-3 look carefully near the ends, and identify as exactly as possible
places at the top and bottom of the line where it crosses an intersection of
lines on the graph paper. Identify the coordinates (/i,kl) and (I2 ,V2 ) of these
intersections and evaluate the slope as
V -V
slope = —-----^
A -/,
We then immediately have
R = slope
which gives us the answer we want. In more complicated expressions, the
value for the slope may give the desired answer only after computation with
other measured quantities.
Figure 6 - 3 T h e “b e s t” slo p e an d o u te r lim its.

We carry out this process three times. The line AB gives us our chosen
“best” value for R, and the other two lines, CF and ED, will give us the upper
and lower limits, outside which we are “almost certain” the value for R does
not lie. It usually happens that these extreme values for the uncertainty range
are roughly equidistant from the central value, and is then finally possible to
state the value for R as
R = value ± uncertainty
It may sometimes appear that the “best” line and the two outer limiting
lines are not equally spaced. The reason is usually that the graph contains too
few points to allow good assessment of the positions of the lines. Although
circumstances occasionally appear in which experimenters feel obliged to
express a result as
f+uncertainty 1
valuer
[-uncertainty 2
visual judgment of a graph is rarely sufficiently precise to justify such a pro
cedure. If it is genuinely difficult to identify a “best” line, it is acceptable to
delineate the edges of the band of values (lines ED and CF in Figure 6-3)
and to calculate simply the maximum slope (of the line CD) and the mini
mum slope (of the line EF). We can then give the experimental answer as the
interval between these two slopes, or else we can calculate a central value for
the slope as the average between these two extreme values with an uncer
tainty equal to ± half that interval.
If in our experiment the desired answer is not equal to the slope di
rectly, the expression for the slope may contain a number of quantities, and
the unknown may have to be calculated from the slope by a separate arith
metic process. If these other quantities are themselves uncertain, the uncer
tainty in the answer will have to be found by combining the uncertainty of the
slope with the other uncertainties, using the techniques of Chapter 2.
It is natural at this stage to think about the significance of the uncer

tainty associated with quantities obtained from graphs. The significance de
pends on the type of uncertainty marked on the graph. If the bars indicate
outer limits of possible variation (subjectively assessed or in the case of
statistical fluctuation), then the limits on the slope will similarly be of this
nature. If the points have been marked with IS,,, limits, the limiting slopes CF
and ED will probably represent limits implying better than 68% probability,
because the limiting lines are drawn with a pessimistic bias.
We have assumed in the foregoing that the scatter encountered in

the actual results is within the predicted range of uncertainty. If this is a
valid assumption, the use of the limiting lines gives rise to a fairly well-
defined value for the uncertainty in slope. If, however, the scatter is well
outside the expected range of uncertainty (owing to an unsuspected
source of fluctuation), then there may be 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,
which is described in Section 6-7.
In choosing our three lines, we have deliberately excluded the origin as

a factor in making the choice, precisely because the behavior of the system at
the origin may be one of the things we wish to examine. If the graph of the
model's behavior does pass through the origin, we should inspect the three
lines in that region. It is most unlikely that our central line will pass exactly
through the origin, but if the area between the two limiting lines does include
the origin, we can say that we have consistency between the model and the
system, at least at the present level of precision. Only if both limiting lines
clearly intersect an axis on one side of the origin can we claim that we have
unambiguously identified an unexpected intercept.
If the behavior of the model does lead us to expect an intercept from

which we hope to obtain the value of some quantity, the intersection of the
three lines on the axis in question will give us that intercept directly in the
desired form: value ± uncertainty.
6 -6 CASES OF IMPERFECT CORRESPONDENCE

BETWEEN SYSTEM AND MODEL
When the correspondence between model and system is only partial, we must
be careful to obtain answers without introducing systematic error from the
discrepancies. Refer to Figures 6-l(b ) and 6-l(c ) and consider first the cases
in which the measured values correspond adequately with the straight line of
the model over only a limited range. Obviously, our evaluation of slopes
should be confined to those regions in which the system and model are
compatible. The points that deviate systematically from the line clearly arise
from physical circumstances that are not included in the model, and it is
obviously inappropriate to include them in any calculations that are based on
the model. We disregard, therefore, all points that deviate systematically
from straight-line behavior by an amount clearly in excess of the estimated
uncertainties and observed scatter of the points, and we restrict our calcula
tions of the slope and its uncertainty to the linear region.
A second point concerns intercepts. Even if the model's behavior passes

through the origin, it is not uncommon to find that the graph shows an
intercept. Such a deviation can arise from a variety of causes; fortunately,
many of these prove to be harmless. If the discrepancy causing the intercept
affects all the readings in the same way (like an undetected zero error in an
instrument or a spurious and constant emf in an electrical circuit), then the
graph will give a slope that is free of the systematic error that would other
wise be introduced. It is wise, therefore, to arrange the experiment so that the
answer will be obtainable from the slope of the graph, whereas quantities that
may be subject to undetermined systematic error should be relegated to the role
of intercepts. The capacity of graphical analysis to provide answers that are free
from many types of systematic error is one of its chief advantages.
6 -7 THE PRINCIPLE QF LEAST SQUARES
All the procedures described in the preceding sections have one characteristic
in common; they are all based on the use of visual judgment by the ex
perimenter. Thus, although the procedures are commonly used and are useful,
they are vulnerable to the criticism that, even when they are carefully carried
out, we cannot be sure of the numerical significance of the results. It would
be comforting if we could use some mathematical procedure to identify the
“best” line for a set of points, for then we would be released from the
insecurity of personal judgment. In addition, we could hope to find out what
we mean by “best” and to assess the precision of that choice.
The procedure in question is based on the statistical principle of least

squares. We discuss the procedure mostly in the restricted application to
straight-line fitting to measured values. It is possible, in addition, to use the
least-squares principle to fit other functions to sets of observations, and this is
briefly considered later. For the moment, however, we restrict our attention to
straight lines only, so that the discussion gives a clear and simple illustration
of the principle. Further detail on the principle in general is in Appendix 2.
Consider that we have a set of N values of a variable y measured as a

function of a variable x. We must restrict ourselves to the special case in
which all the uncertainty is confined to the y dimension; that is, we shall as
sume that the x values are known exactly or at least so much more precisely
than the y values that the uncertainty in the x dimension can be neglected.
Fortunately, many common experimental situations involve one variable
which, if not exactly zero in uncertainty, is at least so much more precise than
the other that the assumption we are making is good enough. If this condition
cannot be satisfied, the simple treatment following is not valid. The least-
squares method can be extended to cover the case of uncertainty in both
dimensions, but the procedure is not simple. Anyone who wishes to pursue
the subject can find an excellent treatment in the text by Wilson listed in the
Bibliography.
The questions now to be answered by our mathematical procedure

are: Which of all possible lines on the x-y plane do we choose as the best
line? What do we mean by “best”? The principle of least squares makes
this choice on the basis of the deviations of the points in a vertical
direction from a line. Let AB in Figure 6-4 be one candidate for the status
of “best” line. Consider all the vertical intervals between the points and
the line, of which /[Q are typical. We define the best line to be
the one that makes the sum of the squares of deviations such as P^Q\
andP 202 ^ minimum.
F ig u r e 6 - 4 F ittin g a stra ig h t lin e to a se t o f p o in ts by th e p rin c ip le o f le a s t sq u a re s

Notice that we have no right to consider an invented criterion like this

as providing any automatic path to “truth” or to “correct” answers. It is sim
ply one choice of a criterion for optimizing the path of the line through the
points. It does, however, offer some advantages over other possibilities, such
as minimizing the third power of the intervals, or the first power, and so on.
Although we need not in general be concerned with the rationale for the
principle of least squares as we use it, the basis for its claim to validity is of
interest. It can be proved that the procedure of minimizing the squares of the
deviations gives rise on repeated sampling to smaller variance of the resulting
parameters, such as slope, than does the use of any other criterion. We
consequently have greater confidence in results obtained by using the prin
ciple of least squares than is the case for any competitor. As a result, use of
the principle of least squares is almost universal.
We now put the least-squares principle into mathematical form. We

define the best line to be that which leads to the minimum value of the sum
and we desire the values of the parameters, slope m and intercept b, of that
best line.
Let the equation of the best line be

y = mx-\-b
The magnitude of the deviation P-Q- is the interval between a certain meas
ured value and the y value of the point, on the line, at that value of x. This
y value, Y-, can be calculated from the corresponding x value by using
Y. = mx. -\-b
so that, if we call each difference 8y-, we have
= y,-{mx,+b)
The criterion of least squares then enables us to obtain the desired values of
m and b from the condition
-(wjc, -\-b)f = minimum
Write
Y,[y,-(mx,+b)f = M
Then the condition for the minimum is
dM dM
0 and = 0 ( 6- 2)
dm db
A brief algebraic exercise (given in full in Appendix A2-2) then allows us to
obtain the values of slope and intercept for the best line as
m= (6-3)
and
b= (6 ^ )
We have now succeeded in replacing the sometimes questionable

use of personal judgment by a mathematical procedure that leads to
results of well-identified significance and universal acceptability. In
addition, because there is some statistical meaning in the new method, we
can expect a more precise form of uncertainty calculation. The least-
squares principle allows us immediately to obtain values for the standard
deviations of the slope and the intercept, giving us uncertainties of known
statistical significance.
The standard deviations of the slope and intercept are calculated in

terms of the standard deviation of the distribution of 8y values about the best
line, which we call Sy. It is given by
EM
' N -2
(6-5)
(Do not worry about a standard deviation being calculated with N - 2

in the denominator instead of the familiar or A^ - 1; it is a consequence
of applying the definition of the standard deviation to the positioning
of a line on a plane.) The values for the standard deviation of the slope.
and for the standard deviation of the intercept Sf^ can then be
calculated to be
N
S... = S ., ( 6- 6)
and
Sf. = S, X (6-7)
The full derivation of these equations can be found in Appendix 2.
These values for the standard deviations can be used in association with
the values of m and b to determine intervals with the normal meaning,
namely, that intervals of one standard deviation give a 68% chance of enclos
ing the universe value, two standard deviations 95%, and so on. One
important advantage of the least-squares method is, therefore, that it supplies
statistically significant values for the uncertainties in our slope and intercept.
Not only are they statistically significant; they are in addition derived objec
tively from the actual scatter in the points themselves, irrespective of any
optimistic claims for the uncertainties of the measured values.
Besides the complete mathematical description of the least-squares

method in Appendix 2, there is also an extension to the method that we have
excluded from the present discussion. If, in the experiment, the points used in
the least-squares calculation are not equally precise, we should use some
procedure that accords greater importance to the more precise measurements.
This procedure is called weighting. The use of weighting is not limited to
straight-line fitting. The procedures are applicable whenever we wish to
combine observations in some way, even in such a simple process as deter
mining the mean of a set of values of unequal precision. The equations for
finding a weighted mean and for doing a weighted least-squares calculation
are given in Appendix 2.
6 -8 LEAST-SQUARES FIT TO NONLINEAR FUNCTIONS
The procedures used in Section 6-7 to determine the slope and intercept of
the best straight line can, in principle at least, also be applied to nonlinear
functions. We can write an equation analogous to Equation (6-1) for any
function, and we can still use a requirement similar to Equation (6-2) to ex
press the minimizing of the quantity M with respect to the parameters in our
chosen model. If the resulting equations for the parameters are easy to solve,
we can obtain their values just as we did for straight lines.
Frequently, however, it is not easy to solve the equations. In such cases

we abandon the attempt to obtain an analytical solution to the problem and
rely on the computer to provide us with approximate solutions by using it
erative techniques. We construct a trial function, calculate the sum of the
squared differences, and then vary the chosen function until a minimum is
found for that sum. Descriptions of such computer-based methods are given
in the text by Draper and Smith that is listed in the Bibliography. If, however,
a method can be found to test a model in linear form, this is certainly simpler.
In all cases, it is the responsibility of the experimenter to choose the

type of function to be used. All that the least-squares method can do is to give
us, for whatever function we choose, those values of the parameters that
provide the best fit with the observations.
6 -9 PRECAUTIONS W ITH LEAST-SQUARES FITTING
The mathematical procedures for least-squares fitting are completely

objective and impartial. The use of Equations (6-3) and (6-4) for linear
fitting drives a straight line through any set of points with complete
disregard for the appropriateness of a straight-line function. If, for
example, an experiment has given us a set of observations (Figure 6-5)
that clearly show the breakdown of a linear model and we heedlessly use
the least-squares procedure on the whole set of observations, we shall
obtain the parameters of a line, AB, that has no significance at all, neither
for the model nor the system. Unthinking use of the least-squares
procedures must be studiously avoided.
This warning is all the more important because of easy access to

calculators and computers that can give, at the touch of a few buttons,
least-squares parameters for any set of numbers we care to insert. We
must remember that, if we are comparing straight lines with our set of
observations, it is because we have made the decision that this is a
reasonable thing to do. We must not, therefore, even contemplate using a
least-squares procedure until we have plotted the observations on a graph
and satisfied ourselves, by visual inspection and personal Judgment, that
linear fitting is appropriate. In addition, as has been mentioned, it may be
necessary to decide that some of the observations fall outside the scope of
the model and are not appropriate for inclusion in the choice of the best
straight line. Only after we have carefully considered the whole situation
graphically and visually and are sure that linear fitting is appropriate over
all or part of the range of the observations, are we justified in starting the
least-squares procedure. Failure to heed this warning can give rise to
serious error in experiment interpretation.
F ig u r e 6 - 5 Im p ro p e r u se o f le a s t-s q u a re s fittin g .
6 -1 0 FUNCTIQN FINDING
All the foregoing discussion involved the assumption that we were already in
possession of a model that we wished to compare with a system. Although
this is commonly the case, it does happen sometimes that we have a set of
observations for which no model is available— for example, in research on a
phenomenon that has never been observed before, or in work on a system
that is so complex that a theoretical model will never be available. The
observations, when plotted in elementary form, will probably show a
curve of no readily identifiable form. In the absence of a model, what are
we to do?
One thing we can do is try to find functions that have some degree of
correspondence with the observations. Such a procedure can be very useful.
For example, in very complex systems for which there is little hope of con-
structihg theoretical models, it may be the only thing we can do. An

empirical model, even if is only a mathematical function that is nothing more
than a restatement in mathematical form of the actual behavior of the system,
can facilitate computer processing of the observations and is indispensable
for procedures such as interpolation or extrapolation. Such models can be
used, for example, to predict the response of a country's GNP to a change in
taxation or to obtain measurements of temperature from the calibration curve
of a resistance thermometer.
In simpler systems for which some hope of constructing a theoretical

model from first principles exists, some functions, if shown to be appropriate
to the observations, may be able to give valuable guidance in model building
by suggesting the type of physical process involved in the phenomenon. Even
so we must be careful. The fact that we have identified a function that seems
to be consistent with our set of observations at a particular level of precision
does not “prove” that we have found the “right” function. Quite often,
functions of widely varying type can show closely similar variation,
especially over a short range of the variables, and “guidance” from an
inappropriately identified function can be misleading. It can retard genuine
theoretical progress for years. The history of physics contains many examples
of such failure to understand that any choice of an empirical function must be
provisional.
With due attention, therefore, to the possibly limited significance of our

procedures we describe some of the methods used. They can be quite simple
in a few cases, and two of these are important because they involve functions
that are of relatively common occurrence. Assume that we have made
measurements of two variables that we can call and y.
Power Law
In the discussion of experiment planning, we have already described the nature of

logarithmic plotting and the uses to which it can be put. For the sake of
completeness in the topic of experiment evaluation, we briefly recapitulate that
description in this and the following section. Consider the function
T=^
where a is a constant. We have
logj^ = alogx
(where the logarithms can be taken to any base we please) and a graph of log
y versus log jc is a straight line with slope a. Consequently, if we wish to test
whether a power law is a good function for our observations, we can plot
them in the form log y versus log .x. If the resulting points plotted in this way
correspond well with a straight line, we can say that a function involving a
simple power, positive or negative, integral or fractional as determined by the
graph, is a good fit with our observations. The value of the appropriate
power, a, is derived from the slope of the graph and is obtained within
uncertainty limits that depend on the uncertainties plotted on the points. A
graph like this can be plotted on ordinary graph paper by plotting the actual
values of log x and log y, or we can use logarithmic graph paper. This paper
has rulings that are spaced in proportion to the logarithms of the numbers, so
that we can plot our observations directly on the paper.
Exponential Functions
For many physical phenomena an exponential function is appropriate.

Consider
y = ae'"
where a and b are constants. In this case
log^ y = log^ a + bx
(where the logarithms must be natural logarithms, taken to the base e), and
the graph of the function is a straight line when we plot log^ y versus x. If
there is reason to suspect that an exponential function is appropriate to a
particular system, we should do a semi-log plot, either on ordinary graph
paper by obtaining the values of log^ y, or on semi-log graph paper, which
has one logarithmic and one linear scale. The appropriate values of a and b
are obtainable from the intercept and slope of the line, with uncertainties
determined by the plotted uncertainties of the measured values.
6 -1 1 POLYNOMIAL REPRESENTATION
If neither a simple power nor an exponential function has been found to

provide a good match to a set of observations, the probability of stumbling on
a more complicated function that would be appropriate is very small. In such
cases it is often useful to resort to a polynomial representation, which we can
write in the form
y = cIq-\-a^x-¥a2X~ + . . .
A recourse of this nature is really an admission that we have no idea what is
going on in the system. Although such a representation may not contribute much
insight regarding the fundamental theoretical basis for the operation of the
system, it at least offers some of the advantages of empirical models. If nothing
more, it allows computerized processing of our knowledge of the system's

behavior and provides a satisfactory basis for interpolation and extrapolation.
The coefficients in such a polynomial expansion that make it appro

priate for our particular system can be found by using the least-squares principle.
Recalling the remarks of Section 6- 8, it will be appreciated that the as
sociated difficulties escalate rapidly with the number of terms that are required in
the polynomial to give satisfactory correspondence with the observations. A
fuller discussion of such methods is in the text by Draper and Smith that is
listed in the Bibliography.
A similar method is available if the scatter in the observations is not too

severe and if the highest precision is not required. The techniques of the
calculus of finite differences can be applied to the observations, and a differ
ence table can be used for interpolation and extrapolation or for polynomial
fitting. A complete discussion of difference-table methods is in the texts by
Whittaker and Robinson and by Hornbeck that are listed in the Bibliography,
and an elementary description is in Appendix 3.
6 -1 2 OVERALL PRECISION OF THE EXPERIMENT
At the beginning of the experiment, we guessed at the uncertainties that we

were likely to encounter. This was only an estimate made for the purpose of
supplying guidance for the conduct of the experiment. At the end of the ex
periment, we should look back and, by critical assessment of the results,
evaluate the precision actually achieved. It does not matter very much what
type of uncertainty we choose—estimated range of possible value, standard
deviation, standard deviation of the mean, and so on—provided only that we
clearly state the kind of uncertainty we are quoting.
To be useful, the overall uncertainty figure must be realistic and honest,

even if the outcome of the experiment is less favorable than we had hoped. It
should also include all identifiable sources of uncertainty. If the balance point
cannot be identified within 2 to 3 mm or if errors are introduced by
nonuniformities of the slide wire, there is no point in claiming that potentials
read on a 1 m slide-wire potentiometer are precise to 0 .2% simply because
the scale is graduated in millimeters.
Known contributions from systematic errors should not be included at

this stage, because the appropriate corrections to the measurements should
already have been made. On the other hand, a source of systematic error,
whose presence we suspect but whose contribution cannot be evaluated accu
rately, should be described and appropriate allowance made in the overall

range of uncertainty. The final statement depends on the circumstances.
Result Is the Mean of a Set of Readings
The best quantity to quote is the standard deviation of the mean, because it
has recognizable numerical significance. Sometimes the standard deviation
itself is quoted. It is always essential to quote the number of readings so that
the reliability of the a estimate can be judged.
Result Is the Consequence of 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 either outer limits for the uncertainty or
else a standard deviation.
Result Is Obtained Graphically
If the straight line has been established by a least-squares method, the uncer
tainties in the constants m and b will have been obtained directly. These un
certainties have the advantage that they have been obtained from the actual
scatter of the points, regardless of their estimated uncertainties. (This does
not mean that, if we intend to make a least-squares fit to a straight line, we
should not bother to plot the uncertainties or even not draw a graph at all. As
was emphasized in Section 6-9, the graph, with the uncertainties on the
points, is still needed to judge the range of matching between the model and
the system before the least-squares calculation is done.) If the straight line
has been drawn by eye, the lines at the limits of possibility will give the pos
sible range of slope and intercept. This uncertainty in slope may have to be
combined with the uncertainties of some other quantities before the final un
certainty of the answer can be stated.
As mentioned earlier, it probably does not matter much what kind of

uncertainty is quoted, so long as one is quoted and the nature of the quoted
value is made clear. Also, when one is working through lengthy uncertainty
calculations, the arithmetic may be simplified by dropping insignificant con
tributions to the total uncertainty. There is no point in adding a 0.01% con
tribution to one of 5% because we do not know the 5% with three-figure ac
curacy. In the final statement of uncertainty, it is not commonly valid to
quote uncertainties to more than two significant figures; only work of high
statistical significance justifies more.
Once the overall 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. This matter has already been covered in Section 2-11; we re
peat the discussion here simply for the sake of completeness as we discuss
experiment evaluation.
There is no unique answer to the question of significant figures, but in

general, one should not keep figures after the first uncertain figure. For example,
5.4387 ± 0.2 should be quoted as 5.4 ± 0.2, 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 were known to be
0.15, it would be valid to quote the answer as 5.44 ± 0.15.
If a measurement is quoted with a percentage precision, the number of

significant figures is automatically imphed. For example, what could be meant if
a measurement were quoted as 527.64182 ± 1%? The 1% means that the absolute
uncertainty could be calculated to be 5.2764. The precision itself, however, is
quoted to only one significant figure ( 1%, not 1.000%), 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 0.64182 has no meaning. The measurement could
then be quoted as 528 ± 5 or, alternatively, 528 ± 1%. 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.
Finally, always be sure to quote an answer and its uncertainty in such a

way that the two are consistent—that is, neither as 16.2485 ± 0.5 nor as
4.3 ± 0.0002.
6 -1 3 THE CONCEPT OF CORRELATION
Until now we have been considering the interpretation of experimental results

in which relatively precise observations were available and the models were
relatively satisfactory. We are not always so lucky, and much of modem
experimenting is less simple and clear-cut than the preceding sections might
suggest. In many areas of science it is common to be concerned with subtle
phenomena in which the effects we seek to measure can be wholly or
partially masked by statistical fluctuation or other perturbations. In such
cases, far from being able to make detailed comparisons between the system
and a model, we may find it difficult to obtain clear-cut evidence that the ef-
feet we are considering even exists. This is a not uncommon situation in, for
example, biological, medical, and environmental studies. We are all familiar
with the discussions about the role of smoking in lung cancer, of low levels
of ionizing radiation in leukemia, or of diet in cardiovascular disease. In such
cases the concept of “proof’ is almost always brought into the discussion in
phrases such as: We have not proved that smoking causes lung cancer. Can
we prove that heart attacks are less likely if we eat margarine instead of but
ter? And so on. In cases like these we are in a very different area of operation
from our earlier kind of experimenting, and it is worthwhile spending a mo
ment to think about what we mean by words like “proof’ and “cause.”
Consider two experiments. One might be a measurement of the current

through a resistor as a function of potential difference across it, and the result
might be as shown in Figure 6- 6(a). In this experiment, have we “proved”
that the current is “caused” by the potential difference? Certainly the current
at the top end of the range is different from that at the low end by an amount
greatly in excess of the uncertainty of measurement, and that gives us confi
dence that the variation actually existed. Given that it existed, was it “caused”
by the change in potential difference? On that one occasion we certainly did
observe that the current did increase as the potential difference increased.
However, it could be that the current has nothing to do with potential
difference and that the increase in current was caused by some totally
separate factor, such as atmospheric pressure. The apparent relationship with
potential difference could have been totally accidental. Philosophers for hun
dreds of years have been warning us that events observed to take place si
multaneously are not necessarily causally related. In the present case, how
ever, accumulated experience with the experiment, using multiple repetition
and careful attention to the control of other variables, will gradually convince
us that potential difference and current are genuinely related. Only a philo
sophical purist would quarrel with the claim that the potential difference
caused the current to flow.
The situation is quite different in less clear-cut cases. Another experi

ment might yield the result shown in Figure 6- 6(b). This result would be
likely if we were dealing with, perhaps, the number of colds experienced by
the whole student body of a university as a function of the amount of ascorbic
acid ingested daily. Can we say that the number of colds is dependent on
ascorbic acid dose or not? We might conduct a well-designed experiment
using an experimental group of 100 students who were given ascorbic acid
every morning and [as described in Section 5-6(b)], a control group of 100
students who unknowingly swallowed sugar pills instead of ascorbic acid
every morning. If we find that the control group had 125 colds in a particular
lar period and that the experimental group who took ascorbic acid had a total of
(a)
• • •
(b)
F ig u r e 6 - 6 E x tre m e s o f b e h a v io r w ith re sp e c t to c a u se -a n d -e ffe c t re la tio n sh ip s.
106 colds, the questions we must ask are: Is this difference significant? What
do we mean by “significant”? If the difference is significant, can we attribute
it to the ascorbic acid? And so on. Even painstaking attention to the details of
experimenting, control over samples, elimination of extraneous variables,
repetition of the experiment, and the like, may not clear up the situation very
much. Biological systems are so complex that we can rarely attain the degree
of control over variables that characterized the electrical experiment. It there
fore becomes inappropriate to seek the kind of “proof’ that is available in
other systems. We cannot say that we have “proved” that smoking causes
lung cancer or that ascorbic acid reduces the incidence of colds in the
same way that we can “prove” that a potential difference “causes” a
current to flow. We have to be content with another class of statement,
which, although less exact, can still be adequately significant and
completely convincing.
This type of statement can be illustrated by reference to a diagram such

as Figure 6-7. These measurements were made to test the proposition: The
number of counts obtained from a weak radioactive source depends on the
length of time of counting. Here, statistical fluctation is almost as big as the
effect we seek to observe, but we can still see that there is an upward trend in
the observations and we say that there exists a correlation between one vari
able and the other. This means that we can observe a tendency for one vari
able to follow the other, although fluctuations arising from other factors prevent
the observation of a unique, one-to-one correspondence. The mathematical
study of such correlation is called regression analysis, and it supplies a
numerical measure of the degree of correlation between two variables that we
call a correlation coefficient.
We encounter the concept of correlation in two significant cases: (I) if,

of two measured variables, one can be regarded as the cause of the other but
its effect is partially masked by random fluctuation, and (2) if two variables
can be regarded as simultaneous consequences of a common cause whose ef
fect, as before, is partially obscured by random fluctuation. In either case we
may be able to say that we can observe a certain degree of correlation be
tween one variable and the other.
The mathematical properties of correlated variables are described in the

standard texts on statistics. We confine ourselves here to quoting the equation
by which one calculates correlation coefficients. For a pair of measured
variables x and y, the expression for the correlation coefficient, commonly
called r, is
r=
Values of r calculated in this way can vary all the way from I, for perfect,
fluctuation-free dependence of one variable on another, to 0 if there is no
connection whatsoever between the two variables. For intermediate values,
the correlation coefficient indicates the extent to which the observed variation
in one quantity can be ascribed to variation in another. In case (I) it is the
extent to which the variation in the output variable can be ascribed to
variation in the input variable; in case (2) it is the extent to which the vari-
22 -
20 -
18 -
• •
16 - • • • •
o
o 14 -
12
10 - • • • •
• •
8 - • •
• • • • •
6 • •
• • • • •
4 - • • •
• • • •
2 -
1 - • •
JL
2 4 6 8 10 12 14 16 18 20
Time (Sec)
F ig u r e 6 - 7 T h e d e p e n d e n c e o f n u m b e r o f co u n ts on c o u n tin g tim e fo r a
w e a k ra d io a c tiv e so u rce.
ation in both variables can be ascribed to variation in whatever is the common

source of influence. In the circumstances commonly found in the type of
experimenting we discussed in this text, values of r close to unity are usual.
Even when we do observe a correlation, we must still be careful

about inferring causal connection between the various variables. If we
observe that one variable seems to correlate well with another, we have
not “proved” that one variable “causes” the other in the same sense as we
used these words in the example about electric current. The literature has
many examples of false and misleading correlations. One conference
speaker illustrated this point with a tongue-in-cheek claim to have
discovered the cause of cancer. He showed a graph of a quantity that
correlated beautifully with the increase in some type of cancer and only
later revealed that the other variable was the consumption of fuel oil in
the British Navy. In another case, intended apparently to be taken
seriously, a 1920s newspaper report described the “discovery” of the
cause of polio, because the incidence of the disease correlated so well
with the number of motor cars on the roads.
However, such amusements do not discredit the study of correlations or

the search for causal relations; they merely serve as another reminder of the
need for caution and clear thinking. When treated with great care, and
especially when the correlation can be observed repeatedly, correlation
studies can and do supply convincing evidence of causal connection. Because
of the immense importance that many of such issues have in public affairs, it
is important to have a clear understanding of the nature of correlation and the
methods available for significance testing. Further discussion is beyond the
scope of this volume, but pursuit of the topic in the texts on statistics (listed
in the Bibliography) is earnestly recommended.
6 -1 4 USE OF COMPUTERS IN EXPERIMENT EVALUATION
Many software applications have facilities that aid in processing experimen

tal observations. Apart from word processing, which is obviously helpful in
the preparation of reports (considered in the next chapter), any application
that contains a spreadsheet facility eases the burden of arithmetic calculation
enormously. Advanced spreadsheet programs such as Lotus 1-2-3 and Quattro
Pro have built-in mathematical functions that can cope with all but the most
specialized requirements. Nevertheless, some limitations and precautions must be
kept in mind. We consider a number of different aspects in turn.
Graph Drawing
Almost all spreadsheet programs produce attractive graphs of experimental

observations, but their use in serious experimenting is limited. As has been
stressed repeatedly, the function of a graph in experimenting is to enable the
experimenter to form a judgment about the degree of correspondence between the
system and the model. A graph can do this only if it is large enough to show
clearly both the scatter of the points and the uncertainty bars or boxes that have
been plotted on them. This can require large sheets of graph paper for an
experiment containing precise measurements, but spreadsheet programs and
normal printers usually produce graphs on 8 1/2-by-l 1-inch paper. These may be
acceptable as illustrations in a report, but they are rarely satisfactory for accurate
analysis of experiments. Unless the computer can produce output on a large
plotter, there is no substitute for a good, big, hand-drawn graph.
(1) Uncertainties. Apart from size, a problem appears with the common
spreadsheet programs when we try to represent uncertainties on the points.
Both Lotus 1-2-3 and Quattro Pro have an alternative to the normal “jc-y”
choice on the graph menu. They call it “High-Low,” and it is designed to
represent stock market prices. It produces a vertical bar to represent the
plotted values and so is suitable for representing experimental observations

that have uncertainty in the vertical direction only. Figure 6-8 shows the ob
servations from Table 4 - 1 in a graph drawn using Quattro Pro 4.
On the work sheet the load values were entered in column A. In column
B were entered the values for the extension incorporating the minus value of
the uncertainty, and in column C were entered the extension values calculated
with the positive sign for the uncertainty. For the graph choices, the primary
selection of graph type was “High-Low.” Columns B and C were selected as
the “ 1st Series,” and column A was selected as the “X-Axis Series.” Such a
representation is satisfactory if all one wants is to show the observations and
if the uncertainty is restricted to the vertical direction. The “High-Low”
option for graph type, however, does not allow us to plot graphs of functions
in any easy way, and we are restricted to showing experimental values only.
To plot functions easily we must use the ordinary “jc-y” choice for
graph type, as opposed to the “High-Low.” Figure 6-9 shows the observa
tions from Table 4-3 as re-plotted using Quattro Pro 4.
Because we wish to plot the graph of the model as well as the observa
tions, we must use the “.x-y” choice for graph type rather than the “High-
EXTEN SIO N O F RUBBER W ITH LOAD
LOAD, kg
F ig u r e 6 - 8 T h e o b se rv a tio n s o f T a b le 4 -1 re -p lo tte d u sin g Q u a ttro P ro.
TIM E vs. SQ UARE R O O T O F DISTANCE
SQUARE ROOT OF DISTANCE

F ig u r e 6 - 9 T h e o b s e rv a tio n s in T a b le 4 - 3 re -p lo tte d u sin g Q u a ttro P ro
Low” form, which raises problems about representing the uncertainty on the
points. The diagram was prepared as follows. In the work sheet, column A
contains the values of distance. Columns B and C contain the measured val
ues of time of fall using, respectively, the negative and positive signs for the
uncertainty. Column D contains values of the square root of each distance. To
draw the graph of the experimental observations, columns B and C were used
together as the “ 1st Series” (i.e., the y-axis variable), and column D was used
as the “X-Axis Series.” Under the menu entry “Customize Series,” the
“Format” was chosen to be “Symbols” (so that no connecting lines would be
drawn between the points), and the “Marker” selected was “Horizontal Line.”
This produced an indication on the graph of the vertical range of each
measured point. To draw the graph of the model function, column E was used
to calculate values of time for each distance of fall, using the model function.
That column was then chosen as the “2nd Series” under the graph menu and
plotted by using “Lines” as the format. In this way both the experimental
observations and the model function can be shown simultaneously on the
same graph. Such a diagram, although too small to allow accurate appraisal
of the experiment, can serve as an acceptable illustration in a report.
(2) Least Squares Analysis. The larger spreadsheet programs such as Lotus
1-2-3 and Quattro Pro, have built-in facilities that use the least-squares prin
ciple for fitting straight lines to sets of observations. These facilities can
make the process fast, accurate, and simple—almost too simple, in fact—and
it is essential to bear in mind the warning given earlier about proceeding to

least-squares analysis only after adequate personal appraisal of the situation
has been made by using an appropriately large graph.
To do a least-squares analysis, we enter the observations in two col

umns and then choose the option marked “Regression.” This immediately
supplies values for the slope and intercept of the best line, as well as other in
formation that we discuss in a moment. We usually want to illustrate also the
best line on the graph of the actual observations, and we do this by calculat
ing values that lie on the best line. In Figures 6- 10(a) and 6- 10(b) we show
both a printout of the work sheet and the resulting graph for the observations
listed in Table 4-3.
In the work sheet. Column A contains the measurements on distance.

Columns B and C contain the measurements of time calculated by using the
negative and positive values of the uncertainty. Column D gives the calcu
lated values of the square root of the distance. The regression analysis was
done on the basis of these four columns. In Quattro Pro the regression facility
is found in the “Tools” menu under “Advanced Math.” Using the
“Regression” option, column D was selected as the “Independent Variable,”
and columns B and C together were selected as the “Dependent Variable.”
The output of the regression calculation is shown below the tabulated values.
The meaning of the various items is as follows:
“Constant” means the intercept on the y-axis

“X Coefficient” means the slope
“Std Err of Coef ’ means the standard deviation for the slope
and these three represent a large part of what we require from the regression
analysis. The one remaining item, the standard deviation for the intercept, is
not represented directly in the regression output, but it can be obtained by a
short calculation. The quantity “Std Err of Y Est” means the standard devia
tion of the y values about the best hne. This is the quantity that, in Section 6-7 we
called 5y, and the value we need, the standard deviation of the slope, was
given by
i
If we want the standard deviation of the intercept, therefore, we must obtain it
by this separate calculation from “Std Err of Y Est.”
DISTANCE, m TIME(min), s TIME(max), s sqrt(DISTANCE) TIME(calc)

0.10 0.143 0.153 0.316 0.137
0.20 0.191 0.201 0.447 0.197
0.30 0.239 0.249 0.548 0.243
0.40 0.285 0.295 0.632 0.281
0.50 0.310 0.320 0.707 0.315
0.60 0.347 0.357 0.775 0.346
0.70 0.380 0.390 0.837 0.374
0.80 0.398 0.408 0.894 0.400
Regression Output:
Constant -0.00656
StdErrofYEst 0.005217
R Squared 0.997152
No. of Observations 8
Degrees of Freedom 6
X Coefficient(s) 0.454867834598
Std Err of Coef. 0.00992360903
(a)
TIME vs. SQUARE ROOT OF DISTANCE
F ig u r e 6 - 1 0 T h e le a s t-s q u a re s a n a ly s is o f th e fallin g ball e x p e rim e n t u sin g Q u a ttro Pro.
The regression output also gives us (along with a few things that do not
concern us) the value for the regression coefficient r. It is called “R Squared,”
and we can see that, as we would expect for a well-controlled experiment, its
value is very close to 1.
After fitting a line to observations by using least squares, we usually

want to illustrate the situation by using a graph that shows the observations
and the least-squares line together, as shown in Figure 6- 10(b). The graph
was drawn in much the same way as was the one shown in Figure 6-9. The
points were produced by exactly the same procedure, and the only difference
lay in the way the line was produced. For Figure 6-9 we wished to plot an
explicit model function, and we had to calculate the y values for that func
tion. In Figure 6-10 we are pretending that we do not have an explicit model
function, and we have used the least-squares procedure to generate a func
tion. To plot the line, therefore, we must use the least-squares values of slope
and intercept to calculate y values on that line. These values appear in Figure
6- 10(b) in column E, and in the graphing process, that column was selected
for the second graph and the “Lines” option was chosen for it. As before,
such a diagram makes a suitable illustration for a report.
(3) Function Finding. We have already discussed in Sections 6-10 and 6- 11

the philosophy and logic of finding functions to fit observations for which no
preexisting model is available. There is no need to repeat that discussion here
except to emphasize again that, even if one does find an empirical function
that matches the observations to a certain extent, there is no necessary
significance in that function. It may be useful for interpolation or, if treated
with great caution, extrapolation, but it cannot be claimed to supply anything
more than that. It may turn out to be useful as a guide in building a model
based on basic principles, but it also may not.
We have mentioned here (see also Appendix 3) some procedures

that can be used to find suitable functions by means of hand-done
calculations, tbut he methods are slow and tedious. Computers make the
task easy, and one can quickly try various possible functions. Most of
these programs fit their various functions by using the least-squares
criterion, either directly and analytically, or else by some process of
successive approximations that use repeated iteration. One such program
is SlideWrite, which was used to create Figure 6-11 with its various
curves fitted to the observations from Table 4 - 1.
SlideWrite has one additional advantage in that it will plot directly the
uncertainty on the points, either in the vertical or horizontal direction (but not
both together). Figure 6- ll( a ) shows the observations plotted directly. In
Figure 6- 1l(b) we see the result of fitting a straight line to the observations,
(c) shows an exponential function, (d) shows a logarthmic function, and
(e) shows a power law (i.e., a function of the type ax^). Clearly, none of
these is a good fit to the observations, but if we had had some good
reason to use such a function, the program would give us the best choice
of parameters for that chosen function on the basis of the least-squares
criterion. Figure 6- l l ( f ) shows a better fit. It is for the sigmoidal
function, which has the form
LOAD,kg
LOAD, kg
0.50 (d ) 0.50
0.40 / f 0.40
f/ E
0.30 f / §
(0
z
111
0.20 T X 0-20 T
UJ
f
0.10 0.10
--------"T
■ ^
0.00 ____________1____________ 1____________ 1____________1____________ 0.00
0.20 0.30 0.20 0.30
LOAD, kg LOAD, kg
0 .5 0
(f)
0.40 0.40
f
E
0.30 z 0.30
g
(0 /
z /
0.20 5
HI
0.20
X
0.10 0.10
0.00 0.00 -----------------------1----------------------- 1-----------------------1-----------------------1-----------------------

0.20 0.30
LOAD, kg
0.20 0.30
LOAD,kg LOAD.kg
Figure 6-11 V a rio u s fu n c tio n s fitte d to th e o b se rv a tio n s o f T a b le 4 -1
T = ao+-
l + exp
^3 y
and the program has given us the values of the various a's optimized to meet
the least-squares criterion using an iterative process.
Last, Figures 6- ll( g ) and 6- ll( h ) provide graphic illustration of an

interesting and important point. Both represent polynomial fits to the obser
vations, but they are calculated for two different orders in the polynomial
function. In Figure 6- ll( g ) the polynomial chosen contained terms up to the
third order—that is, the function was of the form
y-a^ -\-a^x^
and the computer gave us the optimized values of the coefficients on the
basis of least squares. Figure 6- l l ( h ) shows the result of fitting a sixth
order polynomial to the measurements. Both of these functions would
obviously serve us well if we were interested in interpolation only
(although it is apparent that the sixth order polynomial provides a slightly
better fit than the other). Their behavior outside the measured range,
however, provides convincing evidence that, as was pointed out in
Section 4-2, extrapolation on the basis of a model that is wholly
empirical (i.e., based on nothing other than the observations themselves),
is a highly questionable exercise.
PROBLEMS
1. An experiment was done to measure the impedance of a series R-L cir

cuit. The impedance Z is given as a function of the resistance /?, the fre
quency of the source/and the inductance L by
The experiment was done by measuring Z as a function of f with the

intention of plotting Z" vertically and p horizontally to obtain L from
the slope and R from the intercept. The observations obtained are given
in the table.
A U (/^ ) A U (Z ^ )
/ , H e rtz Z, O h m s /X A U (/) = 2 / X A U (/) Z X A U (Z ) = 2 Z X A U (Z )
123±4 7 .4 ± 0 .2
158 8 .4
194 9 .1
2 00 9 .6
229 10.3
245 10.5
269 11.4
292 11.9
296 12.2
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) 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.
(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.
2. Ten observers report on the intensity of a lamp measured repeatedly by
using a comparison photometer. Their results (in arbitrary units) are as
follows:
O b s e rv e r M e a n o f In te n sity M e a su re m e n ts S ta n d a rd D e v ia tio n o f M e an
I 17.3 2 .1
2 18.4 1.9
3 17.1 2.5
4 16.6 2.8
5 I9.1 3.2
6 17 .4 1.2
7 18.5 1.8
8 14.3 4.5
9 16.8 2.3
lO 17.4 1.6
What is the overall mean value for the intensity, and what is its standard
deviation?
3. An experiment has been carried out to investigate the temperature de
pendence of the resistance of a copper wire. A common model is repre
sented by the equation
R = R^(\ + aT)
where R is the resistance at temperature T°C, Rq is the resistance at 0°C,
and a is the temperature coefficient of resistance. The observations of R
and T that were obtained follow:
rc R O hm s
lO 12.3
20 12.9
30 13.6
40 13.8
50 14.5
60 15.1
70 15.2
80 15.9
Assume that the uncertainty in the measurements of temperature can be

neglected.
(a) Using the method of least squares (i.e., using directly the
equations of Section 6-7), evaluate the slope and intercept of the
graph of R versus T.
(b) Hence, obtain the best value for a.
(c) Evaluate the standard deviation for the slope and for the intercept.
(d) Hence, evaluate the standard deviation of a.
(e) State the final result of the experiment with the appropriate
number of significant figures.
4. It is desired to fit a set of observations to the function y = a-\-bx^ by
using least squares. Use the same procedures that are used in Appendix
2, Section A2-2, for calculating the constants of a linear function to
obtain equations for a and b in the parabolic function. Hence, calculate
the values of a and b appropriate to the following set of observations:
0.5 1.5
l.O 6.3
1.5 12.4
2 .0 12 .6
2.5 18 .0
3 .0 32.8
3.5 4 0 .2
4 .0 4 7 .4
Assume that uncertainty is confined to the y variable.

5. The following measurements were made in the investigation of phe
nomena for which no existing model was available. In each case iden
tify a suitable function and evaluate its constants.
V i X >’ T /
O.l 0.61 2 3.2 lOO 0.161

0 .2 0 .7 5 4 16.7 150 0 .5 4 6
0.3 0.91 6 4 4 .2 200 0 .9 9 5
0 .4 I.II 8 8.2 250 1.438
0.5 1.36 lO 15 0.7 3 00 1.829
0 .6 1.6 6 12 2 3 3.5 3 50 2.191
0.7 2 .03 14 3 3 7 .9 400 2 .5 0 0
0.8 2 .48 16 4 6 4 .5 450 2.755
0 .9 3.03 18 6 1 8 .0 500 2 .9 8 1
(a) (b) (c)
6. The following results come from a study on the relationship between

secondary-school matriculation averages and the students' overall aver
ages at the end of first-year university. The first number of the pair is
the secondary-school average and the second is the university average.
7 8 ,6 5 ; 8 0 ,6 0 ; 8 5 ,6 4 ; 7 7 ,5 9 ; 7 6 ,6 3 ; 8 3 ,5 9 ; 8 5,73; 7 4 ,5 8 ; 8 6,65; 80 ,5 6 ; 8 2 ,6 7 ; 8 1 ,6 6 ; 8 9 ,78
8 8 ,6 8 ; 8 8 ,6 0 ; 9 3 ,8 4 ; 8 0 ,5 8 ; 7 7 ,6 1 ; 8 7 ,7 1 ; 8 0,66; 8 5 ,6 6 ; 87 ,7 6 ; 8 1 ,6 4 ; 7 7 ,6 5 ; 9 6 ,8 7 ; 7 6 ,5 9
8 1 ,5 7 ; 8 4 ,7 3 ; 8 7 ,6 3 ; 7 4 ,5 8 ; 9 1 ,7 8 ; 9 2 ,7 7 ; 8 5,72; 8 6 ,6 1 ; 84 ,6 8 ; 82 ,6 6 ; 8 1 ,7 2 ; 9 1 ,7 4 ; 8 6 ,6 6
9 0 ,6 8 ; 8 8 ,6 0 .
(a) Draw a scatter diagram of university averages plotted against

school averages.
(b) Evaluate the correlation coefficient.
7. Evaluate the correlation coefficient for the values of ^/jc and t in Table
4-3.
7
Writing Scientific Reports
7 -1 GOOD W RITING DOES MATTER
It is almost impossible to overestimate the importance of good scientific writing.

The best experimenting in the world can be of Uttle or no value if it is not
communicated to other people—and communicated well by clear and attractive
writing. Although communication may sometimes be verbal, in the
overwhelming majority of cases people learn about our work from the printed
page. Our obhgation to become as literate as we can is therefore not trivial, and it
should be regarded as an essential and integral part of our experimenting
activities. Our writing must be sufficiently good to attract and retain the interest
and attention of our readers. This chapter contains some hints on how this may be
achieved, and the sample report in Appendix 4 is an example of the ways in
which the suggestions can be implemented in practice.
It is almost impossible to tell a person how to write well. It would be very

convenient if we could lay out a short list of instructions and guarantee thereby
fluent, lucid, and literate prose, but the list is not available. Each of us has
different ways of expressing thoughts, and each must allow his or her writing
style to develop in its own way. This needs extensive practice, and we should
regard report writing in the introductory physics laboratory as an excellent
opportunity to obtain it. We may end up acquiring different writing styles, but
provided the message is clear, the diversity can be enriching rather than
damaging.
We now turn to some practical considerations concerning the actual

writing of reports. Although there is no list of explicit instructions for good
157
158 WRITING SCIENTIFIC REPORTS CHAP. 7
writing, there is one principle that will make the production of good, readable
prose more likely. Whether one is preparing a report for internal circulation in a
private organization or a paper for publication in the open literature, there is one
person whose interests must claim the writer's first attention—the person who
will actually read the report. As far as the report is concerned, that person is the
most important person in the world. We are well advised to concentrate our
attention on him or her. Our readers are very hkely people we do not know,
perhaps in some far-distant part of the world and very hkely knowing nothing
about us or our work except the report that they hold in their hands. We probably
have only one chance to influence them—as they read our report—and the report
must do it alone. We cannot stand beside our readers, adding explanation and
clarification if they encounter difficulty in understanding what we have written.
Not only must the report stand on its own, the outcome of the reading can be
highly significant. The public recognition of our work, the opportunity for others
to benefit from it, our own reputation, perhaps even our chances of employment
or promotion may depend on these few minutes spent by our readers as they
work their way through our report. Do we have to be persuaded further that we
should take writing seriously?
It has been common in the past to recommend a detached, depersonalized

mode of expression characterized by the use of the passive voice and impersonal
constructions. There seems to be little point in perpetuating such stilted language.
We can simply tell our readers what we did in our experiment—for example:
“We measured the time of fall using an electronic timer that was accurate to I
millisecond.” Since there is no single “right” way to write a report of an
experiment, we should feel free to use such language and modes of expression as
allow us to express our thoughts in the most clear, attractive, and persuasive way
possible. For invaluable advice regarding writing style, consult the little book by
Strunk and White that is listed in the Bibliography.
We now consider the various sections of the report in turn, all as seen
through the eyes of our all-important reader.
7 -2 TITLE
The title is probably the first part of the report to draw the attention of read
ers. Because they are almost always busy people with many items competing
for their attention, our report will capture their interest only if the title is in
formative, appropriate, and attractive. It should not be too long, yet it should
specify quite explicitly the topic of the work. For example, if the purpose of
the experiment is to measure the specific heat of a fluid by using continuous-
flow calorimetry, we can use this fact directly as a title: “Measurement of the
CHAP. 7 WRITING SCIENTIFIC REPORTS 159
Specific Heat of Water by Using Continuous-Flow Calorimetry.” Notice that

three questions are answered in this title:
1. Is the work experimental or theoretical? That is, are we reporting a mea

surement or a calculation?
2. What is the topic of the work?
3. What general method did we use?
Attention to these three items will almost invariably result in a good choice
of title.
7 -3 FORMAT
The sections that follow analyze the various parts of a report. The various
subsections that are described under each section heading should not them
selves be used as headings in actual reports. Although practice obviously
varies with circumstances, reporting on most normal work in the introductory
physics laboratory needs only the minimum of division. The sections of a
report that are essential are the following:
INTRODUCTION
PROCEDURE
RESULTS
DISCUSSION
These divisions can be used as a basic starting point. The headings should be
neat, clear, and written in block capitals. Subsections within each of these
main sections should be used only when the length or complexity of the re
port makes them indispensable for clarity. Other main sections may be intro
duced in accordance with the requirements of particular experiments. Sug
gested possibilities are the following:
THEORY
SAMPLE PREPARATION
UNCERTAINTY CALCULATIONS
To make the report as inviting to read and as easy to understand as

possible, it should contain a clear, logical thread of argument, and we should
not allow anything to disrupt that development of thought. If we feel that we
must include some particular piece of description that is so lengthy and de
tailed that it would interrupt the smooth development of the main argument,
we should consider making it an appendix to the report. In that way all the
detail is available to any reader who wants it, but the main continuity of
thought is not broken.
Let us turn now to the details of each section of the report.
7 -4 INTRODUCTION
The various components that make up an informative introduction are, in or

der of presentation, as follows.
Topic Statement
Review of Existing Information
Application of Information to Specific Experiment
Summary of Experimental Intention
Topic Statement
With a good title, we can assume that we now have our readers' attention and
that they have picked up the report. However, they are almost certainly
starting from zero, or close to it, as far as our particular experiment is con
cerned. As they start to read, our first task is to orient their thinking toward
the particular area of study. We are not going to succeed in this by diving
immediately into unorganized detail about the experiment. Think instead of
the most general statement that can be made about the experiment and state it
directly. For example: “It is possible to measure gravitational acceleration by
using the oscillation of a simple pendulum.” In this way readers are taken
from their initial state of ignorance to direct awareness of the specific topic of
the work.
Review of Existing Information
At this point readers need some reminder of the basic information relating to
this particular area. We can meet this need by giving them a brief summary of
the existing state of knowledge relevant to the experiment. The summary may
include, as necessary, some aspects of the history of the subject, a summary
of earlier experimental work, or both. Two items are not discretionary and
must be included in every report on an experiment. One is a clear statement
of the system and the experimental circumstances with which the report
deals; and the second is a description of the model or models used.
It is generally best to give this summary of existing information quite

briefly for fear of obscuring the main line of argument, but it should be suf
ficiently detailed so that readers can understand the rest of the report. In the
interests of brevity and clarity, the derivation of standard theoretical results
associated with the model should not be included. (The way in which these
standard results are manipulated to refer to our particular system is another
matter, however, because that is specific to the experiment. This is the topic
of later discussion.) The behavior of the model, as represented by important
equations, should be quoted, and it is important at this stage to mention any
assumptions contained in the model that may limit the validity of the equa
tions. For example: “It can be proved that, in the limit of vanishingly small
amplitude of oscillation, the period of oscillation of a simple pendulum,
considered to be a point mass at the end of a massless, inextensible string, is
given by ... .” To compensate for the omission of standard derivations it may
be desirable to include in the references a source in which the complete deri
vation can be found.
Application of Information to Specific Experiment
Readers are now equipped to understand all that follows in the report, and
their natural reaction at this point will be to wonder: “How does all this refer
to this particular experiment?” We therefore supply a paragraph or two to
show how the basic information, such as an equation representing the behav
ior of the model, can be converted to provide a foundation for our particular
experiment. Commonly, this involves some procedure such as putting the
basic equation into straight-hne form (or some suitable equivalent) and
identifying the ways in which the model can be tested against the system. We can
also point out at this stage the information that will become available from the
parameters of the graph (such as slope and intercept in the case of straight-line
plotting). Readers thereby become fully aware of how our final answer will be
obtained.
Summary of Experimental Intention
It is helpful to readers to conclude the introduction with a summary of our

specific intention in the experiment. For example: “Thus, by measuring the
variation of index of refraction with wavelength, we shall be able to test
Cauchy's model using a graph of n vs. the values of Cauchy's coeffi
cients A and B that are appropriate to our glass specimen will then be ob
tained, respectively, from the intercept and slope of the graph.” Such a
statement is satisfying to the readers because, particularly in a long and
complicated experiment report with a lengthy introduction, it offers them a
review in summary form of the whole course of the experiment, and it en
ables them to make sense of the subsequent description of the actual conduct
of the experiment.
Statement of Experimental Purpose
No mention has yet been made of the traditional statement of purpose for the
experiment. It has been omitted so far because, although it should appear
somewhere in the introduction, there is no universally suitable location. If the
topic of the experiment is familiar, the statement of purpose can form an ac
ceptable topic statement right at the beginning of the introduction. For ex
ample: “It is the purpose of this experiment to measure the acceleration of
gravity by timing the fall of a freely falling object.” Under suitable circum
stances such a statement of purpose can make an excellent topic statement.
On the other hand, the basic purpose of an experiment might involve matters
so complicated and unfamiliar that a statement of it would be completely in
comprehensible unless it followed a substantial amount of preparatory mate
rial. It is easy to imagine a complicated theoretical description that could
profitably conclude with the phrase: “... and it is the purpose of this experi
ment to determine a value for the coefficient k in equation lO.” It does not
matter a great deal where the statement of purpose comes, so long as it is in
cluded and comes at a point in the introduction where it fits well and makes
good sense to the readers.
The introduction has performed a number of services for our readers.

Right at the beginning the topic statement has directed their attention to our
particular area of work. They have then been reminded of the existing state of
knowledge in that area. Next they have been shown how that applies to our
particular experiment. Finally, they have been given a concluding summary
of our specific experimental intention. They are now ready to hear how we
actually did the experiment.
7 -5 PROCEDURE
The report's introductory section takes the form of a descriptive sequence that
proceeds from general to specific. We start with a topic statement that is the
most general remark about the experiment we can make, and we end with a
completely specific statement of intention. Such a sequence is designed to
suit readers' requirements in the introductory section, and a similar sequence
is equally suitable for the procedure section. If we were to start the descrip
tion of procedure by launching immediately into a mass of unorganized de
tail, we would succeed only in irritating our readers. How can they appreciate
the great care we took in some experimental detail if they are not aware even
of our choice of measured variables? We should be as considerate of our

readers' mental efforts when we write the procedure section as we were in the
introduction. A second sequence from general to specific is clearly called for.
Outline of Procedure
To set the scene for the subsequent description of the details of procedure and
measurement, we first offer readers a review of the whole course of the
experiment. If the experiment really consisted of the measurement of the
variation of electrical resistance of a copper wire with temperature over the
range 20°C to 100°C, we say just that to provide the readers with a frame
work into which they can fit all subsequent description of detail. If we start
the description of procedure by saying that we connected terminal A to ter
minal B, switched on power supply C, read voltmeter D, and the like, we
shall have lost their attention in two lines.
Specific Measurement Details
Now that the readers know the general course of the experiment, they are
ready to be told the specific methods by which we measured each of the re
quired quantities, carried out sample preparation, and performed other steps.
This can be done quite simply by stating each in turn until we have com
pleted the list. We must make sure that no significant method of measure
ment is omitted; in such a thing as a timing measurement it is almost cer
tainly important that we used an electronic timer with millisecond accuracy
rather than a stopwatch that could be read to 0.2 second, and our readers need
to know that we did so. If a quantity in the experiment could be measured by
using some standard and familiar technique, it may be sufficient to mention it
by name. For example: “The resistances were measured using a Wheatstone
bridge accurate to 0.01%.” If we feel it is unusually significant, we can
discuss at this stage the accuracy of any particular measuring process, while
remembering that the overall precision of the experiment is a different topic
that will appear in a subsequent section of the report.
Precautions
After readers have learned the methods by which we made each measure
ment, they may recall the difficulties or possibilities for error that are inher
ent in particular procedures. They therefore need reassurance that we, too,
had thought of these possibilities and had been sufficiently careful to take the
necessary precautions. As we offer that reassurance, however, we need
not go to extremes. Care should obviously be taken with all measurements;
there is no point in making superfluous claims to virtue in describing
routine and obvious precautions. There are times, however, when special
care to avoid some particular source of error is a genuinely important part
of the experiment, and it is reasonable to draw attention to this before we
close the procedure section.
Apparatus Diagrams
Good diagrams of experimental apparatus are an essential part of any

good report. Although a published paper requires drawings of profes
sional quality, such resources are not available in introductory work.
However, we should early acquire the habit of taking care with drawings
of apparatus. Even if sophisticated drawing aids are not available, it is not
too much to expect the use of a ruler. Neatness and clarity are important,
and good, legible labeling assists enormously in understanding the ex
periment. Good diagrams can also help experimenters to write reports.
Reference to a good, clear, well-labeled diagram can save paragraphs of
written description and provide detail that would be intolerably tedious to
read if it were included in the text.
Reference to diagrams can be inserted at any appropriate point in the

text, but reference to a general diagram of the apparatus as a whole can make
a neat and convenient beginning for a procedure section. For example:
“Using the apparatus shown in Figure I, we measured the variation of time of
fall of a ball bearing with height over the range 20 cm to 150 cm.” Figure 7-1
is an example of an acceptable apparatus diagram.
7 -6 RESULTS
Measured Values
At this point in the report, readers have all the information they need for un
derstanding the experiment, and they are ready to receive the results directly.
Because any good experiment almost inevitably involves the variation of
some quantity with another, the results are usually best presented in a table.
As always, neatness and clarity are of paramount importance. Lines for the
table should be drawn with a ruler, and ample space should be provided for
the headings and for the columns of figures. The headings should be explicit
and should include, if possible, the name of the variable, its symbol, and the
units of measurement. Attached to each numerical entry should be its uncer
tainty, unless some separate discussion of uncertainties makes the precision
of the measurements absolutely clear. Tables should be clearly identified
(b) Power control circuit
F — Constant flow apparatus S - 100 V DC supply

7"i, T2 — Thermometers 0 - 100°C Ry — Rheostat 80 Q
G — Glass tube V - Voltmeter 0.30 V DC
C — Rubber hose clamp A — Ammeter 0.10 A DC
R — Heater wire
A^, A2 — Electrical terminals leading to R
F ig u r e 7 - 1 A c o m p le te d a p p a ra tu s d ia g ra m .
with a table number and a title. At this point it may be appropriate to refer to
any graphs we have drawn of the basic variables. A simple statement will
suffice. For example: “A graph of the time of fall versus height is shown in
Figure 2. Any tables of values that are so extensive and detailed that their
inclusion in the main text of the report would disrupt the readers' train of
thought should be relegated to an appendix.
Following the main table or tables, list the measured values of all the
other quantities in the experiment. As always, each should have its uncer
tainty attached, and the units of measurement should be clearly stated.
Description of Measurement Uncertainties
The report should state explicitly the kind of uncertainties we are quoting.
These are likely to be either estimated outer limits or statistical quantities
such as a standard deviation or a standard deviation of the mean. In the
case of statistical quantities, we must not omit mention of the number of
readings in the sample from which the results were derived. If any
quantities in our list of measured values were obtained by computation
from some basic measurement or measurements, we must state clearly the
type of calculation used to obtain the final uncertainty in the computed
quantity. It is not necessary to give much, if any, of the arithmetic details,
provided that our readers can see clearly the kind of calculation that was
performed.
Computation of Final Answer
If the experiment has been well designed, we will probably obtain our final
answer by some graphical procedure. It is now time to tell the readers exactly
what that procedure is. In simple cases we may obtain the answer from the
basic graph of one measured variable against the other, but even then we
must be explicit about what we have done. For example: “The value of the
resistance was obtained from the slope of the graph (shown in Figure 3) of V
versus I between 0.5 A and 1.5 A.” If the answer was not the slope itself but
was obtained from calculation with other measured quantities, we must state
explicitly what we have done. For example: “Our value for the coefficient of
viscosity was obtained from the slope of the graph of Q versus P in combi
nation with the measured values of a and i using Equation (3).”
The readers will wonder what kind of calculation we performed to ob

tain the uncertainty in the final answer. We simply say what we have done. If
we assessed visually the possible range of slopes, we say just that. We can
add, if necessary, that the basic uncertainty in slope was combined with other
uncertainties, and state explicitly the method of calculation. If we obtained
the slope by a least-squares calculation and incorporated any other standard
deviations to obtain a final value for the uncertainty of the answer, we again
state simply what we have done.
Throughout the results section of the report, we do not trouble our busy
readers with unnecessarily detailed calculations. They trust us to do simple
arithmetic, but they want to know what kind of calculation we did. If we feel
compelled, for some particular reaspn, to offer an unusual amount of detail
regarding such calculations, we can always put it in an appendix where it will
be available if wanted but where it will not obscure the clarity of the main
report.
7 -7 GRAPHS
Graphs in the report differ from the graphs used in doing the experiment.
Those graphs were working documents designed as computational aids.
For a precise experiment, the graphs are possibly quite large and finely
drawn to permit precise extraction of information. On the other hand, it is
extremely unlikely that our readers will want to do any numerical work of
their own using the graphs in the report. These graphs serve mostly as
illustrations. They allow the readers to see the behavior of the system so
that they can judge for themselves the validity of our assertions about the
results.
The graphs in the report must be clear, neat, and uncluttered so that
readers do not have to work too hard to get the message. The points on the
graphs should have their uncertainties clearly marked on them (by a box or a
cross), and the axes should be clearly labeled. Both the type of uncertainty
and any symbols used in labeling the axes should be explicitly identified in
some obvious way in or beside the graph; we do not want to cost our readers
the irritation of hunting through the text to find out how to interpret the
graph. Do not, however, fill up empty spaces on the graph with arithmetic
calculations of slopes, and the like. Each graph should obviously have a clear
title or, as is common in printed publications, a more extended caption. In
addition to supplying identification, an extended caption has the added ad
vantage of supplying a good location for the important details mentioned
earlier. A sample of acceptable layout for such an illustrative graph is given
in Figure 7-2.
(n e w to n m^)
F i g u r e 7 - 2 A c o m p le te d grap h .
7 -8 DISCUSSION
Comparison Between Model and System
The discussion is an integral part of the report and not an afterthought. It has
this importance because we have yet to describe the aspect of the experiment
that we have, right from the beginning, regarded as the basic issue in ex
perimenting—the relationship between the system and the model. The out
come of that comparison is vital for the experiment, and our readers will be
eager to hear what we have to say about it.
We have already listed in Section 6-11 the various categories of ex

perimental outcome. Remembering that, in evaluating our results, we had to
disengage ourselves from our hopes and aspirations for the experiment and
accept objectively the actual outcome, so now, at the reporting stage, we
must make a candid and unbiased statement of that outcome. We should
make it a plain, simple statement of the actual situation. For example: “The
behavior of the model is represented by Equation (I), in which the variation
of Q with P is a straight line passing through the origin. In the experiment the
results do show a linear variation over part of the range, but instead of
passing through the origin, the line that best fits the observations has a finite
intercept on the Q axis. In addition, at the high end of the P scale, the obser
vations show systematic deviation from linearity by an amount clearly in ex
cess of the uncertainty of measurement.”
At this stage it is sufficient to make this plain, factual statement. Be

cause such comparison was the fundamental objective of the experiment, it is
necessary for its outcome to be clearly, factually, and prominently stated. In
the report we shall be proceeding quite soon to matters of interpretation and
opinion, and it is important that we start the discussion section with a plain
statement of the actual, indisputable outcome of the experiment.
That statement will raise some questions in the mind of the readers, and
we must turn our attention to them now.
Consequences of Discrepancies Between Model and System
One of these questions concerns the possibility of error in the final answer
that could be caused by failure of correspondence between the system and the
model. Some of these possibilities have already been mentioned in Section
4-5, and readers need reassurance that we have protected the final answer
from that kind of error. We should point out, for example, that an unexpected
intercept will not contribute to error in a quantity that has been obtained from
the slope alone, or that a systematic departure from linearity over part of a
graph did not invalidate an answer that was obtained from the linear segment
only. Much of the skill in experimenting lies in the protection of the final
answers from such sources of error, and we can be quite explicit about our
claims to have done so.
Speculation Concerning Discrepancies Between System and Model
In describing the report's earlier sections, we have stressed objective and

factual reporting of the actual situation. Matters of opinion or conjecture
should not have played a significant role in those parts of the report, and we
have probably limited ourselves to such statements as would have been made
by most impartial observers. Now, however, comes a stage at which we not
only can but should introduce our own ideas. Our readers have in turn been
informed about the actual degree of correspondence between the system and
the model, and they have been reassured that the final answer has not been
contaminated (as far as we were able to tell) by any failure of correspondence
between the system and the model. Because we have met our basic
obligations as experimenters, we could quite justifiably leave the report there.
However, the interest of the readers will doubtless have been aroused by the
description of any discrepancies between the model and the system. We
presumably started with a model that was chosen to suit the system as closely
as possible. If any failure of correspondence between the system and the
model had been anticipated, such breakdown would have been incorporated
into the experiment design. If we want to measure a coefficient of viscosity
by using theories based on streamline flow, for example, we do not design an
experiment to run under turbulent conditions (unless, for some separate
purpose, we wish to detect the onset of turbulence). Any observed failure of
correspondence, therefore, is bound to attract attention, and readers will want
to know what we think about it. We are more familiar with the experiment
than anyone else and are in a better position than others to guess at the origin
of discrepancies.
Sometimes a discrepancy has (at least superficially) an origin that is

easy to identify. If we measure, for example, the flow rate of fluid through a
pipe, a departure of the flow-rate measurements below linear behavior at high
values of pressure difference may be ascribed quite confidently to the onset
of turbulence. If the objective of the experiment included the detection of the
onset of turbulence, such a statement could end the matter. At other times,
however, more comment is needed. If, in the preceding example, our
intention had been simply to measure the coefficient of viscosity from the
linear part of the Q, P variation, readers might wonder why we had not been
more successful in avoiding a region in which the streamline theory was

clearly invalid. Perhaps we had been surprised by an unexpectedly early on
set of turbulence; if so, we should be candid enough to admit it and perhaps
speculate on the origin of that discrepancy. If the situation is genuinely
puzzling, we may not be able to offer much in the way of speculation, but it
is always worth trying. As has been said, as the experimenters, we have a
better chance of speculating fruitfully than most others, and our ideas are al
most certainly to be of interest and possible value to other workers.
Sometimes, however, despite our best efforts we fail and are unable to offer
any constructive ideas. We must be completely honest. If we are dealing with a
well-tested system and a well-known, reliable model and if we have tried and
failed to resolve a failure of correspondence between them, our situation cannot
but be of interest to other workers. We should tell them about it, and perhaps we
shall all learn something from the resulting discussion.
As we attempt to be creative regarding our experimental discrepancies,

we should remember that we are doing something important. All models and
theories go through processes of refinement, and these processes are based on
the various stages of observed failure of the models. We should try to be
responsible, therefore, as we speculate. Instead of having a fling at every wild
idea we can imagine, we should try to make our suggestions have some
logical connection with the evidence of the discrepancy. For example, if we
have done an experiment on the oscillation of a load suspended from a
spring, we could write: “Since the unexpected intercept in the plot of T~ vs.
m gives a finite value for T at w = 0 , we have a clear indication of the pres
ence of an extra mass that was not included in the measured values of load."
Whether we can guess at the identity of this extra mass is less important. We
have at least offered a logically acceptable inference from the observed na
ture of the discrepancy, and further research and experimenting in this area
will have been facilitated.
Appendix 1
Mathematical Properties
of the Gaussian or
Normal Distribution
A1 - 1 THE EQUATION OF THE GAUSSIAN DISTRIBUTION CURVE
To derive the equation of the Gaussian distribution curve, we consider a

quantity, whose unperturbed value is X, to be subject to random uncertainty.
We assume that the uncertainty arises from a number, 2n, of fluctuations
from the central value, each of magnitude E and equally likely to be positive
or negative. The measured value x can then range all the way from X - 2nE,
if all the fluctuations have the same sign in the negative direction, to
X -\-2nE if the same thing happened positively. Intermediate values arise
from various combinations of positive and negative contributions. We wish
to determine the form of the resulting distribution curve for a very large
number of such measurements. This form will be determined by the prob
ability of encountering a particular deviation R within the total interval
±2nE. The probability is governed by the number of ways in which a par
ticular deviation can be generated.
For example, a deviation of the total value 2nE can be generated in only
one way: All the elementary contributions to the deviation must have the
same sign simultaneously. An error of magnitude (2n - 2)E, on the other
hand, can occur in many ways. If any one of the elementary fluctuations had
been negative, the total deviation would have added up to {2n - 2)E, and this
171
172 MATHEMATICAL PROPERTIES-GAUSSIAN OR NORMAL DISTRIBUTION APP. 1
situation can arise in In different ways. A deviation of (2n - 2)E is, therefore,
2n times as likely to happen as one of 2nE. A situation in which two of the
elementary fluctuations have negative signs can, correspondingly, be
generated in many more ways than for one, and so on.
The argument can be generalized by using the number of ways in which

a specific deviation R can be generated as a measure of the probability of the
occurrence of that deviation and, consequently, as a measure of the frequency
of its occurrence in a universe of observations.
Consider a total deviation R of magnitude 2rE (where r <n). This must

be the result of some combination of fluctuations of which {n-\-r) are posi
tive 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 things is {2n)\. However, not all these arrangements are
different for our purpose, because we do not care if there is any internal rear
rangement within the fluctations in, say, the positive group. We must, there
fore, divide the total number of arrangements by the number of these insig
nificant rearrangements, that is, by ( at -f r)!. Similarly we must divide by the
number of internal rearrangements that are possible in the negative group,
[i.e., by {n - r)!]. The total number of significant combinations is, therefore,
(2 a7)!
{ n r ) \ { n - r)\
This quantity is not yet strictly a probability, although it is a measure of the
likelihood of finding such a total deviation. The probability itself is obtained
by multiplying the foregoing number by the probability of this combination
of {n + r) positive and { n - r ) negative choices. Because the probability of
each choice is 1/ 2, the required multiplier is
(/»+r) {n-r)
The final result for the probability of the deviation R is then

\(/» + r) / - \{n-r)
{2n)\
(A l-1)
(A7-i-r)!(A7-r)! v2
Our problem now is to evaluate this result as a function of the variable

r. This is done subject to the condition that n is very large, tending to infinity.
The evaluation requires two auxiliary results.
1. The first auxiliary result:

APP.1 MATHEMATICAL PROPERTIES-GAUSSIAN OR NORMAL DISTRIBUTION 173
n\^yl2me rT
This is known as Stirling's theorem. Although its full derivation is beyond
our scope, its plausibility can be indicated as follows.
Note first that
log xdx = [x log ;c - x] j'

= n\ogn-n-\-\
Now the graph of log x versus x is shown in Figure A l- l, and it shows clearly
rn
\ogxdx
that the value of the integral can be approximated by the sum
logl + log2 + log3+ . . . + l 0gA7

which is lo g (lx 2 x 3 x . . . n) ox \ogn\. We can, therefore, write approxi
mately, if n is large,
log A?! = n \o g n -n
or
«! = c - V
This is an approximation to the formula given above. A full derivation is in
the text by Margenau and Murphy listed in the Bibliography.
2. The second auxiliary result is:
//-►oo yi )
This can be proved as follows. The expansion for [l + (I / n)^' is
, n I n ( n - \) ( I \2 n {n -\){n -2 )
l-h---- -h '
I! n 2! 3!
As n becomes larger, all the terms involving n clearly tend to unity, so that
the series tends to
1 I I I
1-|-----1----- 1----- h . . .
1! 2! 3!
and the sum of this series to infinity has the value e, which gives us the re
quired result.
We are now in a position to evaluate the expression (A 1-1). We apply

Stirling's theorem to the terms (2n)!, (n + r)!, and { n- r ) \ . We obtain
{2n)\ = i 2nf e - ^"^2nx2n =
(n + r)< = (n + r)"*''e-*"'"''>y/2K(n +r)

/ y j+ r+ (l/2 )
' H
and
/j-r+ (l/2 )
n
Therefore,
/i+ (l/2 ) r
{n + r)\ {n-r)\ = n^ 1 - -
(. 1 + -
0
1 - - ^ X In
V n)
The variable part of (A 1-1) can now be written
> -nV 1----

V n V
-(t?! r~) {r^hx) / -n/r{-rin)
n j
1-^1
n J fl
^ +-nlj f\ l - -n
Thus expression ( A l-l) can now be written

/ 2 ( 2\ nlr{~r^ln) / \-n/r(-r~/n)
I
>-V i+ r \ 'I ^
Vim > -nV « j
Now, a s ---- > 00, we have

r
1+ - e
/ \ -nir
fV l - -n
and so, finally, the probability of deviation R is

I -rhn
nn
The significant feature of this result is the form e~''. It specifies the
probability of a deviation R and is thus equivalent to Equation (3-3) in
which the deviation is the difference between the unperturbed value X and .
measured value x. The only problem that remains in putting the equation ini
standard form is to redefine the constants. Put
hx = - ^
for the value of the exponent, and in the constant replace I / hy h dx. The
equation then reads
P{x) dx = " dx
V7l
where P(x) dx is the probability of finding a deviation between x and x + dx.

A1 - 2 STANDARD DEVIATION OF THE GAUSSIAN DISTRIBUTION
To calculate the standard deviation for the Gaussian distribution, we must

calculate the standard deviation of a very large set of values that are distrib
uted in a Gaussian distribution. We must, therefore, calculate the sum of the
squares of the deviations from the central value and divide it by the total
number of observations. Let there be N observations, where N can be as
sumed to be a very large number. The number of deviations from the central
value that occur between x and x + dx equals
yjn
The value of the standard deviation, therefore, is given by
1 Nh -h^x^
- - - f yfn
/- Xx^dx
V7T
The integral is a standard one and has a value /2 /z \ and so the value of
the standard deviation is
h _ 1
Vtt 2hi^ 2h^
This provides the justification for Equation (3-4) and also enables us to
rewrite the probability function P(x) in terms of the standard deviation as
1
P{x)dx ■ e-^^^^dx
The equation is frequently used in this form.
A1 - 3 AREAS UNDER THE GAUSSIAN DISTRIBUTION CURVE
It is frequently desirable to know what fraction of the area under the Gaussian
distribution is enclosed within certain limits on the horizontal scale because
this tells us the probability that observations will ocbur within that interval.
To calculate these areas we proceed as follows. The probability that a
deviation falls between jc and x-\-dx
1
-e-^^'^^'dx
Therefore, the probability that a deviation lies between 0 and jc is

r 1
^dx
Although this integral can be easily evaluated for infinite limits, it is not so
simple for fixed limits as we now require. Numerical methods of integration
are used, with results that are given in Table A l-1 (see also Figure A l-2).
T A B L E A l - 1 A re a s u n d e r th e G a u ssia n C u rv e
P ro b a b ility th a t a d e v ia tio n lies

X l<J b e tw e e n 0 and x
0 0 .0
0.1 0 .0 4
0 .2 0.08
0.3 0 .1 2
0.4 0 .1 6
0.5 0 .1 9
0.6 0.23
0 .7 0 .2 6
0.8 0 .2 9
0 .9 0 .3 2
1.0 0 .3 4
1.1 0 .3 6
1.2 0.38
1.3 0 .4 0
1.4 0 .4 2
1.5 0 .43
1.6 0 .45
1.7 0 .4 6
1.8 0 .4 6
1.9 0 .47
2 .0 0 .4 8
2.5 0 .4 9
3 .0 0 .4 9 9
If we require the probability that a deviation lies somewhere within the

complete range on both sides of the origin, that is, ±(x / a) , the value for the
area is doubled. For example, the entry at x / a = 1 is 0.34, which gives the
68% figure that we have been using for ± la limits. The table gives only an
indication of the way the probabilities run, and for serious statistical work
reference should be made to one of the many statistical tables available (see
the Bibliography under Lindley and Miller).
F i g u r e A l - 2 T h e are a e v a lu a te d in c a lc u la tin g th e p ro b a b ility o f

o c c u rre n c e o f an e rro r u p to x.
Appendix 2
The Principle of
Least Squares
A 2 - 1 LEAST SQUARES AND SAMPLE MEANS
Consider that we make N measurements, x., of a quantity that contains ran

dom fluctuation. Let us calculate that value, X whose deviations from the x,
are minimized in accordance with the principle of least squares. X will be
obtained from the condition
' ^ ( x - - X y = minimum
Let X be the mean of the x.. Then
= - x f + { x - X ) ^ +2(x, - X) ( x - Z)]
or, since - x ) = 0,
This last expression clearly has a minimum value when x = X , thus confirm
ing that the use of the mean as the most probable value for a sample is con
sistent with the principle of least squares.
179
180 THE PRINCIPLE QF LEAST SQUARES APP. 2
A 2 -2 LEAST-SQUARES FITTED TO STRAIGHT LINES
Consider a set of observations {x.,y.) to which we wish 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.
(For the definition of the concept of “weights,” see Appendix A2-3.) The
deviations of the observed points from the straight line y = mx-[-b are of
the form
^y i =y i - { mx , +b )
and we wish to have a minimum value for the sum of the squares of these
quantities. We have
=[>', -{mx^ + b ) f
= yf + rn^xf +b^ 2mXjb - 2mx-y- - 2y.b
If there are N pairs of observations, the sum of the squares, M, is given by
M = Y .^ b y y
= Z Z ^ + 2 wZ) ^ X, - x^y, - I b ^ ^ y ,
The condition for the best choice of m and b is that ^ ( 5 y ,) “ should be a

minimum. We need, therefore,
dM ^ , dM ^
dm db
The first condition gives
2m Z xf + 2bY,x^ - ) =0
and the second gives
2Nb -\-2m^x, - 2 ^ y . = 0
Solution of the two simultaneous equations for m and b gives
m=
« Z « ,’ - ( ! * , ) ’
and
APP. 2 THE PRINCIPLE OF LEAST SQUARES 181
^_Z Z -y- ~ Z Z )
Standard deviations for m and b can be calculated as follows. Consider

m first. Because m is a computed value that is calculated in terms of the un
certain quantities y 2 , and the like, we can apply the equation that we had
earlier (Equation 3-9) for the standard deviation of a computed value, z, that
is a function of variables, ;c, y, and so on. It was
dy
We apply this result to our case by noting that the x and y of the formula are
the y \ , y 2 , and so on that appear in the expression for m. We can write,
therefore.
^dm^ 2 ( dm
S]y\ + s: +
J V
Now, in making our set of measurements of the .x and y values, we would not
normally have measured explicitly the standard deviation for each y value. In
the absence of these we assume that the values of the various SyS can be re
placed by a quantity based on the scatter of the y values about the line whose
m and b values we have calculated. These intervals, 8y,, have a standard de
viation whose value (Equation 6-5) was
Z ( S t ,)^
‘S' =
N-2
and this is the value that we shall use in place of all the separate ‘Sy,» ‘^>'2’ and
so on. Justification of the term - 2 is not attempted here. It is associated
with the fact that the 8y, are not independent but are connected by the
existence of the best line that is specified by the values of m and b. The
equation for the standard deviation of the slope can therefore be written
dm^
y 5' +
l^ J l^ 2 j
or
8m
5! = 5.^
dy^ J
182 THE PRINCIPLE OF LEAST SQUARES APP. 2
Let us write
^ dm ^
\^k
where we have introduced a second index k to denote terms in the summation
for the standard deviation. We must now evaluate the partial derivatives---- .
The value of m is given by

I
m=
and we can see that differentiation with respect to y gives us, for as an ex
ample,
dm I
or in general, for the k'ih term,

dm I
dyu NY^x]-(Y,xi)
and
dm
A '* ;
\^ k
To obtain the value of this must now be summed over the index k. If
we write
.2
^ dm^ I
\^ k j NE « M Z * .r l
and remember that ^ is the same thing as ^ x . , we can easily perform
the summation over the index k to obtain

dm 1
\^ k
________ I_______
a' Z « M Z * ,F
so that, finally, the value of 5,„ is given by
N
5... = 5„ X
The value for Sf^ can be found by using the same procedure.
A 2 -3 W EIGHTING IN STATISTICAL CALCULATIONS
When we perform some statistical calculation, such as obtaining the mean of

a set of observations or fitting a function to observations by using the least-
squares principle, the equations in Sections 3-3 and 6-7 are valid only when
all the observations are equally precise. If the measurements are of unequal
precision, it is obviously fallacious to allow them to make equal contributions
toward the final answer. Clearly, the more precise measurements should play
a more important part in the calculation than the less precise values. To
accomplish this we assign to the observations weights that are inversely
proportional to the standard deviations of the observations. The derivations of
the resulting equations can be found in the standard texts on statistics. We
simply quote the results here.
Weighted Mean of a Set of Observations
Consider that we have a set of independently measured quantities, x-, each of

which has arisen from a statistical process that has yielded a value for its
standard deviation, 5, . The weighted mean of the set of x values is defined to
be
184 THE PRINCIPLE OF LEAST SQUARES APP. 2
and the standard deviation of the weighted mean by
Straight-Line Fitting by Weighted Least Squares
Consider that we have a set of values of a variable y measured as a function

of X. As in Section 6-7 we assume that the x values are precise and that all
the uncertainty is confined to the y values. We assume also that the y values
are of unequal precision and have been assigned weights w. The equations by
which we can calculate the slope m and the intercept b of the best line can be
written as follows:
m=
^_ Z ~Z
Because of the cumbersome nature of these equations, we have used abbre

viated notation in which we have dropped the suffix i that should be attached
to each of the quantities. Also, the term w. that is used for the weight of each
pair of measured values (jc, , y^) is calculated in terms of the standard devia
tions of the y values as
I
w.
The best estimates of the standard deviations for m and b can be written
(as they were in Section 6-7) in terms of the deviations of the measured
points from the best line. For a weighted least-squares fit, these deviations
must now be weighted. The weighted value of Sy is given by
The best estimates of the standard deviation for the slope and intercept can
now be written as
=sl/w
and
I
s,=sl +
J^w W
where
w = ' £ { H x - x Y)
and X is the weighted mean of the x values, given by
Y, wx
X=
w
The suffix i has, as before, been omitted. Further detail on weighted least-
squares fitting to straight lines is in the text by Squires that is listed in the
Bibliography.
Appendix 3
Difference Tables
and the Calculus
of Finite Differences
A 3 -1 MATHEMATICAL FOUNDATIONS
The calculus of finite differences supplies a powerful tool for the treatment of
measured variables. For the moment, however, let us consider the situation
wholly from the mathematical point of view. After we have established
mathematically the results we need, we can proceed to apply them to
measurements.
Consider a known function y = f(x) (see Figure A3-1) that can be ex
pressed in terms of a Taylor expansion about the value x = a :
fix) = f(a) + { x - a ) \ ^ ]
dx^^ Ja
Such a function is said to be analytic at the point x = a. Any good book on
calculus will provide more detail.
The function is defined along a continuous range of values on the scale

of jc, and, to make the theory applicable to measured variables, we must con
vert the formulation so that it refers to discrete values of x. Let these discrete
186
APP. 3 DIFFERENCE TABLES AND THE CALCULUS OF FINITE DIFFERENCES 187
F ig u r e A 3 - 1 G ra p h o f the fu n c tio n y = f{x).
values of x be spaced equidistantly upward from x - a dX intervals of h, so

that the values of x in which we are interested are
x = a, x = a-hh, x = a-h2h, x-h3h, . . .

We can now calculate corresponding values of y for these discrete values of
X. They are
/ ( a ) , f(a-hh), f ( a + 3h), . . .
and we can illustrate these values on a graph as shown in Figure A3-2.
If we concentrate our attention on these discrete values of x and y, and

if we wish to find a form of the Taylor expansion applicable to the discrete
values, we can simulate the required derivatives by using finite differences.
We define the quantity Af(a) to be
/^f(.a) = f { a + h ) - f { a )
Correspondingly, we also have.
188 DIFFERENCE TABLES AND THE CALCULUS OF FINITE DIFFERENCES APR. 3
A /(a + A) = / ( a + 2h) - f { a + h)
and so on, and these quantities are related to the first derivatives of the
function at the various values of x. In a similar way we define the second
differences
A V (a) = A /(a + /i) - A /(a)

and so on for third and higher differences.
When we arrange all these differences beside the tabulated values of

/ (jc), we obtain a difference table for the values. A difference table for
the function
y = 2x-\-x^
is shown in Table A3-1.
T A B L E A 3 - 1 D iffe re n c e T a b le fo r th e F u n c tio n y = 2x = x^
\ 3
9
2 12 12
2l
3 33 18
39
4 72 24
63
5 135 30
93
6 228 36
129
7 357 42
171
8 52 8 48
219
9 747 54
273
10 1020
This table illustrates several important features of difference tables, includ

ing, for this example, the constancy of the third differences and the conse
quently zero value of the fourth differences.
Now let us consider the problem of obtaining values of y at values of x

intermediate between the discrete values of x. Furthermore, let us find a way
of calculating these intermediate values from the known values of y, instead
of calculating them directly from the function itself. The advantage of such a
procedure is that it will be available for later use on values for which we do
not know the relevant function. To calculate these intermediate values, we
must rewrite the Taylor expansion in a form that is compatible with the
quantities found in the difference table and that also is suitable for the calcu
lation of intermediate values. In Figure A3-3 the gradient of the function / a t
x = a can be approximated by the ratio A/ h. Corresponding values for the
second derivative can be calculated in terms of the second difference A^, and
so on for higher derivatives.
Consider also a value of x intermediate between x = a and x = a-\-h,

and let it be specified by a new variable u that is defined by
Figure A3-3 A p p ro x im a tio n fo r th e g ra d ie n t o\' fix) at jr = a.
X = a -\-u h
The Taylor expansion can now be rewritten in terms of the above-mentioned

quantities to yield the intermediate values of y as
y = /(^?) + w A -h^w (w -l)A ^+ ^w (w -l)(w -2)A " + . . .
This form of the Taylor expansion is known as the Gregory-Newton formula

for interpolation. It can be used to calculate intermediate values whenever we
have tabulated values of two variables. For example, such methods used to be
commonly employed for interpolating in tabulated values of logarithms,
trigonometric functions, and the like.
To use the Gregory-Newton formula, we construct a difference table as

far as those differences that become either zero or else small enough that the
error involved in the interpolation calculation becomes acceptably small. If
we construct the table so that the value we are seeking lies between x = a and
X = a + /z, the various differences we need for insertion in the formula will be
found along the upper edge of the table. If we are seeking a value that lies
between x = a + h and x = a + 2h, the differences we need will be found
along the next lower row, and so on.
Extrapolation can be carried out by a similar process. Suppose we have

a set of values of y for values of x ranging from x - a to x = a + { n- \ ) h. If
we wish to calculate a value for y appropriate to x = a-\-nh, we must start
with the supposition that the y values beyond x = a- ^{n- \ )h are determined
by the same function as for the lower values. On the basis of that assumption,
there is a simple method of finding y for x = a + 77/2; it is based on a process
of extending the basic difference table. Starting with the column of
differences that are constant, or sufficiently close to constancy for our pur
poses, we calculate successively the lower differences in terms of the higher
until we reach the required value of y. The process is illustrated in Table A3-
2, in which we have pretended that we did not know the function that y was
of X, that we knew the values of y only for values of x up to 6, and that we
desired the value of y for x = 7. In this way the table can be extended in
definitely to provide further values of the function as required.
T A B L E A 3 - 2 T h e U se o f a D iffe re n c e T a b le fo r E x tra p o la tio n
X y A A2
2 8
19
3 27 18
37 6
4 64 24
6l 6
5 125 30
9l
6 216 6
30 + 6
= 36
9l + 36
= 127
7 2 1 6 + 127
= 343
The difference table for actual observations does not work out as neatly
as did the two foregoing examples. For purposes of illustration these exam
ples were constructed by using explicit mathematical functions, and absolute
constancy was found at some level in the differences. Real observations dif
fer from these examples in two respects. First, there is no guarantee, for a
particular set of real observations, that a simple function that would lead to
constancy of some difference column even exists. Second, even if some ele
mentary function is appropriate, the presence of uncertainty in the observa
tions makes it impossible to obtain complete constancy in any of the differ

ence columns. We just use our best judgment in dealing with each situation
as it arises. The situation is considered further at the end of this Appendix.
We can also use a difference table to construct a polynomial that will

either reproduce the actual functional relationship between y and x (if such
exists) or else provide an adequate approximation to a set of experimental
observations. To do this we must rewrite the Gregory-Newton formula in a
form suitable for our purpose. We earlier wrote it in terms of the variable m;
we now wish to write it in terms of the variable x while still incorporating the
differences A rather than the derivatives df/dx. Remembering the definition
of u and using the equation
X = a -\-u h
we have
x-a
u=
h
and the original form of the Gregory-Newton equation becomes

. . 1 1
y = f {x ) = f {a ) + - { x - a ) ^ + —- J { x - a ) { x - a - \ ) / ^ -
h 21 n
+2^{x- a) {x- a- \) {x- a- 2) ^^ + . . .

3! h
The equation is now in the form we desire. If we insert in it the appropriate
values of A, A^ A\ and so on, for a particular value such ^sf(a), we generate
the required polynomial in x.
As an example consider the difference table in Table A3-1 and choose

the values in the top row. They are
f{a) = X 1, /z = l, A = 9, A ' =12, A ' =6, A'=0

Inserting these values and performing some elementary algebra we obtain
y = 2x + x^
Because this is the function we started with, we should not be surprised.
However, we have confirmed the suitability of the Gregory-Newton formula
for generating a polynomial that is consistent with a set of tabulated values. It
is therefore of immense potential value for constructing a suitable polynomial
when we are dealing with a set of measurements alone, with no idea of a
suitable function to act as a model.
A 3 -2 APPLICATION OF DIFFERENCE TABLES

TO MEASURED VALUES
In the preceding section we illustrated the calculus of finite differences and

difference tables by using mathematical functions. When we turn to measured
variables and seek to apply these techniques, we encounter two differences:
(I) We may not know a function that will provide an adequate fit to the
observations, and (2) the variables will show scatter arising from uncertainty
of measurement.
Consider first the case in which the measurements are precise, so that
the scatter is negligible in comparison with the measured values. In this case
the difference table will contain values that behave relatively regularly, and
we can use it to perform immediately such procedures as interpolation and
extrapolation. Furthermore, if a polynomial of a certain order will serve as a
good fit to the observations, the differences of the appropriate order will turn
out to be nearly constant, and the next differences will scatter around zero.
We can then use the procedures of the preceding section to construct the ap
propriate polynomial.
If, on the other hand, our observations show larger scatter, we are faced
with a somewhat different problem of interpretation. It is in principle possible
to fit a polynomial exactly to any set of values, no matter how much scatter
they show. In fact, to any set of values it is possible to fit an infinite number
of polynomials, two of that infinity being represented by the curves in Figure
A3-^.
So which polynomial do we want? Is it going to be one like that repre

sented by the solid line in Figure A 3 ^ 7 Under some circumstances this may
be appropriate. On many other occasions, however, we may have good rea
son to believe that, measurement uncertainty apart, the basic behavior of the
system is regular, and we really want the function corresponding to the dotted
line in Figure A 3 ^ . We have to consider, therefore, smoothing the ob
servations, by which we mean choosing a function or curve that follows the
observations in general terms but ignores deviations smaller than a selected
limit. Many of the standard texts on measurement theory supply detailed de
scriptions of smoothing procedures. See, for example, the text by Whittaker
and Robinson listed in the Bibliography.
It is not always clear how far we have to go in smoothing observations.

We trade off simplicity of representation against possible loss of genuine
detail in the behavior of the system. This takes good judgment on the part of
the experimenter, and our decisions are not always greeted with universal
F i g u r e A 3 —4 F ittin g a p o ly n o m ia l to a se t o f o b se rv a tio n s.
approval. In any case, if we do want to make a choice of a certain order of

polynomial to represent the observations, we can choose the corresponding
difference column in the table to be constant and, on the basis of some aver
aged value of these differences, construct the polynomial we want.
If such a procedure is not to our liking and we are restricted to unavoidably

noisy observations, our only alternative may be to use a least-squares
procedure and thereby minimize the discrepancy between the observations and a
function of a chosen type. Notice, however, the important distinction between the
use of the difference table and the least-squares procedure. The difference table
will tell us the coefficients of a polynomial that is implied by the
observations; the least-squares procedure will simply give us the optimized
parameters of a function whose general form we must choose for ourselves.
Appendix 4
Specimen Experiment
A 4 -1 EXPERIMENT DESIGN
System
We are given a spring suspended from a stand, a pan to hold weights attached
to the lower end of the spring, a set of weights, and a stopwatch with a scale
divided into fifths of a second.
Model
We are told that, on the assumption that the extension of a spring is propor
tional to the load on it, the period of oscillation, 7, of a suspended load, m,
given by
7 = 2 ti
where k is a. constant for a particular spring.
Requirement
We are asked to measure k for the spring with an uncertainty not greater than
10% .
196 SPECIMEN EXPERIMENT APP. 4
Experiment Design
Following the steps listed in Section 5-3, we have the following:
1. The operating system consists of the spring alone. We have been

given no information about the pan for the weights, or any way of weighing
it, so we must proceed without that knowledge.
2. The model contains only two variables, load m and period of oscil
lation r, so it is simple to decide that our input variable, the one we can con
trol, will be m and the output variable will be T.
3. To put the equation into straight-line form, our first idea might be to
remove the square root. Squaring both sides of the equation gives us
r' = 4 n ^ -
k
Comparison with the straight-line equation
vertical variable = slope x horizontal variable
suggests that we could choose
vertical variable =
horizontal variable = m
slope = -----
k
This is an acceptable choice, but the unknown, k, appears in the denominator
of the slope. To simplify later arithmetic, it is equally valid and more conven
ient to write the equation
k
m= — r
47t'
where
vertical variable = m
horizontal variable =
slope = — -
47t'
4. The range of the input variable, m, may be governed by the weights

we have been given. In addition, we should consider the possibility of over
APP. 4 SPECIMEN EXPERIMENT 197
loading the spring. Has anyone suggested, or is it written anywhere, that

loads should be restricted? We might try a few weights on the pan to see how
the spring behaves. One way or another, we can choose a range of m that we
feel comfortable with. The range of T values presents no problem because it
is determined by the system.
5. Let us suppose that our weights are sufficiently precise so that their
uncertainty need not be considered. They are not totally precise, of course,
and if we want to know what uncertainty they do have, we look in the
supplier's catalogue.
The only uncertainty, therefore, arises from the timing measurements,
and that uncertainty depends on the precision with which we are capable of
timing the oscillations. The only way to find that out is to try it. We choose a
typical load, start the oscillations, and measure the time interval for, say, lO
oscillations. We must now decide what determines the uncertainty of the
measurement. Is it the accuracy of reading the stopwatch, or is it our ability
to watch the oscillations and start or stop the watch appropriately? Obviously,
we must test this by trying the measurement again, and we must continue to
probe the measuring system in this way until we are sure we know our
capabilities. We may decide, as in the present case, that we are sure that we
can measure time intervals with an uncertainty of ±0.3 s.
This, however, does not complete our consideration of precision; we
must evaluate the effect on our k values of this uncertainty in T. It is difficult
to plan ahead exactly, because we shall obtain our final value of k from the
graph, but it is only prudent to check that our individual measurements have
adequate precision to contribute significantly to the final result.
For example, suppose we time a certain number of oscillations that give
a time interval, r, of 2 seconds. What would be the contribution of our ±0.3 s
to the uncertainty in /:? /: is a function of Therefore
RU(/fc) = 2 X^ = 30%
Clearly, such a measurement makes little significant contribution to the de

termination of k with 10% precision. If we want I0% precision in k, we need
5% precision in r, and that requirement imposes limits on our measuring
process for t. We can determine these limits as follows:
If
RU(0 must not be greater than 5%

that is.
0.3
<0.05
then
0.3 ^
t > ----- = 6 s
0.05
Thus, whatever the actual value of T, if we time a number of oscillations that
gives a total time interval of 6 seconds or more, the uncertainty in our timing
measurements is likely to be acceptable. For convenience we might choose a
constant number of oscillations for the various loads, but if we were short of
time in a long experiment, we could choose for each load the number of os
cillations that gives us a satisfactory value of t.
6. Let us decide, as a first guess, that we shall measure the time for lO
oscillations, knowing that even for the lowest load this gives us a measured
time interval in excess of 6 s, and that we shall time oscillations for loads of
O.l, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45, and 0.5 kg. Since we want to plot m
versus and incorporate the value of the absolute uncertainty in that is,
IT X AU(r), we should lay out the table that expresses our measurement
program to have the following headings:
P e rio d "
N um ber o f
L oad, m, kg O s c illa tio n s T im e , t, s P e rio d , T, s T\ s~ A U (r^ ), r
Experimental Results
The next step is to make the actual measurements and fill in the table with the
measured values of t and the values calculated for Period and for Period “
with its absolute uncertainty. The result of this process will then appear as in
the following table.
N um ber o f P eriod^,
L o a d , m , kg
O s c illa tio n s
T im e , t, s P e rio d , T, s A V ( T \ s^-
O.IO lO 8.2 ± 0 .3 0 .8 2 ± 0.03 0 .6 7 ± 0 .0 5
0 .1 5 lO 9.8 0 .9 8 0 .9 6 0 .0 6
0 .2 0 lO 10.7 1.07 1.14 0 .0 6
0 .2 5 lO 11.5 1.15 1.32 0.07
0 .3 0 lO 12.5 1.25 1.56 0.08
0 .3 5 lO 13.0 1.30 1.69 0.08
0 .4 0 lO 13.8 1.38 1.90 0.08
0 .4 5 lO 14.5 1.45 2 .1 0 0 .0 9
0 .5 0 lO 15.2 1.52 2 .3 1 0 .0 9
These values of m and must now be graphed. Each plotted value

must contain its range of uncertainty, m has no uncertainty and T~ is uncer
tain by the amount listed in the final column, so that each point on the graph
will be a little horizontal line. Once the values are plotted, the graph will look
as shown in Figure A4-1.
The next step is to interpret what we see in terms of the categories de
scribed in Section 6 ^ . We first observe the scatter of the points and consider
if it is consistent with our prior estimate of the uncertainty. In the present
case there seems to be reasonable consistency between the scatter and the
estimated uncertainty, and no further consideration of this point seems to be
required at this stage. The next point to, consider is the extent to which the
behavior of the system is consistent with the model. The model predicted a
straight line passing through the origin, and we must judge our graph against that.
We can see immediately that in our case the correspondence with straight-hne
behavior seems quite adequate over the whole range. We are justified, therefore,
in including all the points when we decide on our choice of lines.
With regard to intercept, however, the situation is different. A glance at

the graph makes it clear that we are going to have an intercept on the T~ axis
that cannot be ascribed to measurement uncertainty. We return to consider
this discrepancy later, but in the meantime we can note that the final value of
k will be obtained from the slope alone, and that the slope can be calculated
even in the presence of an unexpected intercept.
The next step is to draw lines so that we obtain values for the slopes.
One choice is to draw our “best” line by eye and also lines that represent the
maximum and minimum slopes permitted by the range of uncertainty in the
scatter. At this stage, the graph will look as shown in Figure A4-2.
We now have to read values off the graph that will enable us to calcu
late these slopes. For each line we look for convenient intersections with the
graph paper, illustrated in Figure A4-3, that give us the coordinates of points
at the top and bottom of the line.
On the present graph, the chosen intersections are indicated by arrows, and
the appropriate coordinates are marked. Given these coordinates, it is easy to
calculate the slopes.
Steepest line:
0.55-0.075
slope =
2.40-0.73
= 0.284 k g s '
Central line:
0.55-0.05
slope =
2.50-0.52
= 0.253 k g s '
Least steep line:
0.55-0.075
slope =
2.64-0.51
= 0.223 kg s '
The corresponding values of k can now be calculated by using
slope = — 7
which gives
k = 4 ti^ Xslope
Highest value:
yt = l l . 2 l l kgs"'
Middle value:
A = 9.988 kgs"'
•S 1-0 hZ
Figure A 4-3 T h e id e n tific a tio n o f in te rse c tio n s at th e to p an d at th e b o tto m o f th e g ra p h to p erm it
c a lc u la tio n o f th e slo p e s.
Lowest value:
k = 8.804 kg s
Now that we have a measure of the overall uncertainty of the k values, we

can round off the values to obtain our final statement about k and its uncertainty:
/: = 10.0 ± 1.2 kg s'^
The final figure for the uncertainty at 12% of the k value is slightly
greater than the target figure of 10%, but perhaps we can say that we have
come close enough to our objective. If any further reconsideration is forced
on us, we could go back to the beginning and reassess our basic measurement
uncertainty in timing. Certainly, the low scatter of the points in comparison
with the estimated uncertainty in the upper part of the graph suggests that we
were slightly pessimistic about our estimate of ±0.3 s in timing, and
reappraisal might allow us to refine that estimate.
Given the completion of our calculations for k, we now have to return

to the question of the unexpected intercept. We satisfied ourselves that it was
harmless because our k value was obtained from the slope, and that could be
determined accurately even in the presence of the intercept. Nevertheless, we
should not ignore it altogether, because it constitutes failure of correspon
dence between the model and the system, and it is not good practice in ex
perimenting to leave things like that unconsidered. In guessing at possible
sources of the discrepancy, we note that it seems to be associated with some
load not counted in the m values, for when our added load, m, is zero, the
graph tells us that we would still observe a finite frequency of oscillation.
What could give rise to such uncounted mass? One obvious guess would be
the pan on which the weights were placed. Another obvious guess would be
the mass of the spring itself. Without further investigation, we cannot be
certain that either of these guesses is good, but our explanation for the unex
pected intercept seems reasonable enough so that we are probably justified in
terminating our present experiment at this point and leaving confirmation of
our guesses to further experimenting.
A 4 -2 REPORT
In this section we give a version of a final report that could be written on the
basis of the experiment that we have just described. The report is written ac
cording to the suggestions offered in Chapter 7. Only the final version of the
report is given; the details of its construction and their correspondence with
the suggestions in Chapter 7 can be elucidated by comparing the report with

the various sections of Chapter 7.
MEASUREMENT OF A SPRING CONSTANT BY AN OSCILLATION METHOD
Introduction
The stiffness of a spring can be measured by timing the oscillation of a sus
pended load. For an elastic spring (extension oc load) it can be proved that the
period of oscillation, 7, of a suspended mass, m, is given by
7 = 271 ( 1)
where /: is a constant for a particular spring. The objective of the present ex
periment is to measure the value of k for a spring with an uncertainty not
greater than 10%.
Equation (1) can be rewritten to read
m = —^ - Tr2
47t'
which is linear in m and 7^ with slope ^/47r~. Thus, by measuring the
variation of oscillation period with load we shall be able to plot a graph of m
versus 7^ and obtain the value of k from the slope.
Procedure
Using the apparatus shown in Figure 1, we measured the variation of
oscillation period with load.
The loads consisted of a Cenco weight set ranging from 0.1 kg to 0.5 kg
with stated precision 0.04%.
The period of oscillation was measured by timing a number of oscilla

tions using a stopwatch graduated in fifths of a second.
Results
The measurements of load and oscillation period are shown in Table 1.
TABLE 1 V a ria tio n o f O s c illa tio n P e rio d w ith L o a d

N um ber o f
L o a d , m, kg O s c illa tio n s T im e , r, s P e rio d , T, s Period^, T^, s^
0 .1 0 10 8 .2 ± 0 .3 0 .8 2 0 .6 7 ± 0.05
0 .1 5 10 9.8 0 .9 8 0 .9 6 ± 0 .0 6
0 .2 0 10 10.7 1.07 1.14 ± 0 .0 6
0 .2 5 10 11.5 1.15 1.32 ± 0 .0 7
0 .3 0 10 12.5 1.25 1.56 ± 0 .0 8
0 .3 5 10 13.0 1.30 1.69 ± 0 .0 8
0 .4 0 10 13.8 1.38 1.90 ± 0 .0 8
0 .4 5 10 14.5 1.45 2 .1 0 ± 0 .0 9
0 .5 0 10 15.2 1.52 2.31 ± 0 .0 9
The uncertainty shown for the measured times was estimated by visual
judgment to be the outer limits for possible values of time, and so the calculated
values for the uncertainty of also represent outer limits of possibility.
The graph of the m and T~ values is shown in Figure 2.

Period^ sec^
F ig u r e 2 G ra p h o f m v e rsu s T
The value of k and its uncertainty was calculated from the slopes of the
three lines shown, giving
A:= 10.0±1.2kg s"'
Discussion
Using an oscillation method, we have measured the value of the stiffness constant
for a spring to be lO.O ± 1.2 kg s - 2 . The model represented by equation (1) predicted
for the m versus graph a straight line passing through the origin. In our
experiment, the variation of m with proved to be consistent with a straight hne
within the present limits of uncertainty. Instead of passing through the origin,
however, the lines shown in Figure 2 can be seen to have a finite intercept on the
axis that cannot be ascribed to measurement uncertainty. Our value of k.
however, was obtained from the slope alone and should be free from error
arising from factors that would give rise to an intercept.
Because the intercept gives a finite value for 7 at w = 0, it seems to be

associated with the presence of some load not included in the measured val
ues of m. Although we have not tested these possibilities, we can suggest that
such extra load could arise from the pan supporting the weights and also from
the mass of the spring itself.
Bibliography
The subject of experimenting is vast, and the literature is correspondingly

extensive and diverse. To meet all our varied needs there is no satisfactory
substitute for going to the library and becoming familiar with its range of of
ferings. The following list contains some relatively recent books and some
classics, but it is very far from exhaustive. It is intended only to provide a
few suggestions that may serve as a starting point for individual study.
B acon , R. H. “The ‘Best’ Straight Line Among the Points,” American

Journal of Physics, 21, 428 (1953).
B arford, N. C. Experimental Measurements: Precision, Error and Truth.
Addison-Wesley, 1967.
B arry , R. A. Errors in Practical Measurement in Science, Engineering and
Technology. Wiley, 1978.
B eers, Y. Introduction to the Theory of Error. Addison-Wesley, 1957.
B evington, P. R. Data Reduction and Error Analysis for the Physical Sci
ences. McGraw-Hill, 1969.
B raddick, H. J. J. The Physics of the Experimental Method. Chapman and
Hall, 1956.
B ragg, G. M. Principles of Experimentation and Measurement. Prentice-
Hall, 1974.
B rinkworth, B. j . An Introduction to Experimentation. The English Uni
versity Press, 1973.
Cohen, 1. B. Revolution in Science. Harvard University Press, 1985.
Cook, N. H., Rabinowicz, E. Physical Measurement and Analysis. Addison-
Wesley, 1963.
Cox, D. R. Planning of Experiments. Wiley, 1958.
D eming, W. E. Statistical Adjustment of Data. Wiley, 1944.
D raper, N. R., S mith, H. Applied Regression Analysis. Wiley, 1981.
Fretter, W. B. Introduction to Experimental Physics. Prentice-Hall, 1954.
Freund , J. E. Modern Elementary Statistics. Prentice-Hall, 1961.
208
BIBLIOGRAPHY 209
Hall , C. W. Errors in Experimentation. Matrix Publishers, 1977.

Harre, R. Great Scientific Experiments. Oxford University Press, 1983.
Hornbeck, R. W. Numerical Methods. Quantum Publishers, 1975.
Jeffreys, H. Scientific Inference. Cambridge University Press, 1957.
Kuhn , T. S. The Structure of Scientific Revolutions. University of Chicago
Press, 1970.
Leaver , R. H., Thomas, T. R. Analysis and Presentation of Experimental
Results. Macmillan, 1974.
Lindley , D. V., M iller, J. C. P. Cambridge Elementary Statistical Tables.
Cambridge University Press, 1958.
Lyons , L. A Practical Guide to Data Analysis for Physical Science Students.
Cambridge University Press, 1991
M argenau , H., Murphy, G. M. The Mathematics of Physics and Chemistry.
Van Nostrand, 1947.
M enzel, D. H., Jones, H. M., B oyd , L. G. Writing a Technical Paper.
McGraw-Hill, 1961.
Parratt, L. G. Probabiliiy and Experimental Errors in Science. Wiley,
1961.
Rabinovich , S. Measurement Errors, Theory and Practice. American Insti
tute of Physics, 1993.
Rabinowicz, E. An Introduction to Experimentation. Addison-Wesley, 1970.
Rossini, F. D. Fundamental Measures and Constants for Science and Tech
nology. CRC Press, 1974.
SCHENCK, H. Theories of Engineering Experimentation. McGraw-Hill, 1961.
S hamos , M. H. (editor). Great Experiments in Physics. Holt-Dryden, 1960.
S hchigolev, B. M. Mathematical Analysis of Observations. American
Elsevier, 1965.
S quires, G. L. Practical Physics. McGraw-Hill, 1968.
Stanton , R. G. Numerical Methods for Science and Engineering. Prentice-
Hall, 1961.
Strong, J. Procedures in Experimental Physics. Prentice-Hall, 1938.
Strunk , W., W hite, E. B. The Elements of Style. Macmillan, 1979.
Taylor , J. R. An Introduction to Error Analysis. University Science Books,
1982.
Topping, J. Errors of Observation and Their Treatment. The Institute of
Physics, London, 1955.
Tuttle, L., S atterley, J. The Theory of Measurements. Longmans Green,
1925.
210 BIBLIOGRAPHY
Whittaker, E. T., Robinson , G. The Calculus of Observations. Blackie,

1944.
W ilson, E. B. An Introduction to Scientific Research. McGraw-Hill, 1952.
W orthing, A. G., Geffner, J. Treatment of Experimental Data. Wiley,
1946.
Answers to Problems
CHAPTER 2:
1. 142.45 ±0.15 cm; 0.11% 8. 0.30 ohm

2. 1.245 ±0.005 A, 9. 9.77 ± 0.04 m s-', 0.4%
3.3±0.1V ;0.4% ,3.0% 10. 1800 kg m-'
3. 0.5 minute 11. 0.110 m, 0.0012 m, 1.08%
4. (a) 10 cm, (b) 2 cm 12. 0.8 nm, 0.24%
5. 7.7% 13. 14.3 ±0.1; 14.25 ±0.15
6. 0.6% 14. 6.75 ± 0.03
7. 26 cm'
CHAPTER 3:
33 35 37 39 41 43 45 47 49 51
2. Between 39 and 41, 38 10. 0.16

3. 38.3 11. Rejection
4. 4.39 12. Sample 34, 47, 43, 40, 32 has mean
39.2 and standard deviation 5.56
5. 0.80 Sample 36, 40, 38, 43, 34 has mean
6. 0.58 38.2 and standard deviation 3.12
7. (a) 33.9 to 42.7, 13. More than 130
(b) 29.5 to 47. 1 14. More than 200
8. (a) 37.5 to 39. 1, 15. 0.47 mm^
(b) 36.7 to 39.9 16. 0.21xl0-’ m
9. (a) 3.8 to 5.0, 17. 0.11 ms-'
(b) 3.2 to 5.5
211
212 ANSWERS TO PROBLEMS
CHAPTER 5:
1. No 13, Jvq vertically, / - / o horizontally,

velocity V slope is V
2. range oc
. g J 14. i vertically, AT horizontally, slope is
and intercept is whence a
^ surface tension
3. pressure o c ----------------------------- 15. sin0, vertically, sin 02 horizontally,
radius
slope is ri2 /
^ , Imoment of inertia 16, l/^'vertically, 1/5’' horizontally,
4. period oc -----------------------------------
\ rigidity constant each intercept is Mf; or ss' vertically,
s-{-s' horizontally, slope i s /
17, 1/C vertically, co^ horizontally, slope
load force Y is L
5. deflection oc
^yx radius^ 18. F vertically, 1/ r ^ horizontally, check
for linearity
Xradius
Vradius J 19. F vertically, Z,/ ! horizontally,
where a and b are arbitrary constants check for linearity; also check
6. 5 vertically, horizontally, slope is Fvs. , Fvs. / and F vs. Mr~
all separately, while holding other
7. T vertically, f^(,^ horizontally, slope variables constant
is Am 20. log^, Q vertically, t horizontally, slope
8. P vertically, horizontally, slope is -l/RC
isp/2 21. vertically, 1/ co^ horizontally, slope
9. vertically, cos a horizontally, is 1/ and intercept is F"
slope is 471^^/g 22. vertically, horizontally, slope
10. d vertically, W horizontally, slope is 1/ and intercept is m]
hAIYab" 23 1/ X vertically, 1/ horizontally,
11. /i vertically, MR horizontally, slope slope is -F, intercept is F/4
is 2 a / p g
12. p vertically, 7 horizontally, slope is
R/v
ANSWERS TO PROBLEMS 213
CHAPTER 6;
(c) 0.00129 ohm^s^ 3. (a) Slope is 0.0499 ohm deg-^

(d) 0.00572 henry and the intercept is 11.92 ohm
(e) 0.00145 ohm^s^ (b) a = 0.00419 deg">
0.00117 ohm 's' (c) 0.0024 ohm deg"*
(f) The measured value of L 0.16 ohm
can range from 0.00544 (d) 0.00021 deg-'
henry to 0.00606 henry (e) a =0.00419 ± 0.00021 deg"'
(g) 6.00 ohm Ro= 11.92 ±0.16 ohm
(h) The measured value of R 4. a = 2.12, ^ = 2.98
can range from 5.39 ohm 5. (a) i = 0.5el'
to 6.43 ohm. (b) y = 0.6a:'^
(i) L = 0.0057 + 0.0004 henry (c) / = 6.2^--’^-’'^
R = 6.0 ± 0.6 ohm 6. (b) 0.73
Mean is 17.54 and the standard 7. 0.99
deviation of the mean is 0.26
Index
Absolute uncertainty, 13 Gaussian distribution, 36ff
areas under, 39,176
equation, 38,171
Calibration errors, 14 standard deviation, 39,176
Comparison, between models and systems, I'ijf, \\9 ff Graphs:
Compensating errors, 25 drawing, 118
Compound variables, plotting, 88 linear, 77#
Computed quantities: logarithmic, 92,136
standard deviation of, A ljf in reports, 166
uncertainty in, \5jf use of computers, 145
Computers, l^Sff Gregory-Newton formula, 190
calculation of standard deviations, 35
drawing histograms, 32
function finding, 150 Histogram, 31
graph drawing, 145
least squares fitting, 147 Input, 5
Control group, 106 Input variable, 5, 84
Correlation, 143 Intercepts, 86,95,122
Correlation coefficient, 143 by least squares, 132,181
Curve fitting: standard deviation of, 133
by computer, 150 Interpolation:
by least squares, 133 graphical, 66
using Gregory-Newton formula, 190
Introduction, in reports, 160
Diagrams, in reports, 164
Differences:
in experiment design, 104 Least squares:
standard deviation of, 51 intercept, 132,181
uncertainty in, 21 non-linear functions, 133
Difference tables, 188 precautions, 134
Digital displays, uncertainty in, 12 principle, 130
Dimensional analysis, lOOj^ sample means, 179
Discussion, in reports, 168 slope, 132,180
Distribution, concept of, 31 standard deviations, 132,181
use of computers, 147
weighted, 184
Experiment design, 84j9^ Logarithmic functions:
Exponential functions: standard deviation of, 53
plotting, 137 uncertainty of, 19
standard deviation of, 53 Logarithmic plotting, 92,136
uncertainty in, 19
Extrapolation:
graphical, 67 Mean:
using difference table, 191 definition, 33
using polynomial representation, 152 weighted, 183
Measurement:
nature of, 11
Finite differences, calculus of, 186# standards, 10
Format, in reports, 159 Median, 33
Function finding: Mode, 32
by computer, 150 Models:
by difference table, 192 concept, 58
graphical, 135 empirical, 65
215
216 INDEX
Models (Continued) Sampling, concept of, 40

testing, 73, 85# Significant figures, 26,140
theoretical, 70 Significant variables, 57
Slopes:
determination of, 126,200
Normal distribution, see Gaussian
by least squares, 132,180
distribution
standard deviation of, 133
uncertainty in, 127
Output, 5 Smooth curve, uses, 65#
Output variable, 5, 84 Smoothing, of observations, 193
Standard deviation:
Planning, of experiments, 84# best estimate of universe, 44
Poisson distribution, 37 of computed values, 47#
Polynomial fitting: definition, 35
using computers, 152 of Gaussian distribution, 39
using difference table, 192 of intercept, 133
Polynomial representation, 137 of means, 50
Population, see Universe of slope, 132,181
Powers: Standard deviation of the mean, 42
standard deviation of, 52 Standard deviation of the standard
uncertainty in, 18 deviation, 45
Precautions, in reports, 163 Statistical analysis, of experimental
Products: quantities, 43#, 116
standard deviation of, 51 Statistical fluctuation, 29
uncertainty in, 23 Stirling’s theorem, 173
Purpose, in reports, 162 Straight-line form, 86#
Sums:
standard deviation of, 50
Quotients: uncertainty in, 21
standard deviation of, 52 System, definition, 5
uncertainty in, 24 Systematic error, 14
Rectification of equations, 77# 86#

Rejection of readings, 54 Title, in reports, 158
Relative uncertainty, 14 Trigonometric functions:
Reports, 157# 203 standard deviation of, 53
diagrams, 164 uncertainty in, 19
discussion, 168
format, 159
Uncertainty:
graphs, 166
absolute, 14
introduction, 160
in calculated quantities, 15#
precautions, 163
definition, 13
procedure, 162
in experimental quantities, 115
purpose, 162
on graphs, 118
results, 164
relative, 14
significance of, 158
Uncertainty calculations, general method, 22
title, 158
Universe:
Results, in reports, 164
definition, 40
Rounding-off, 12
mean, 40,44
standard deviation, 40,44
Sample means
distribution, 41
and least squares, 179 Weighting:
Sample standard deviations, distribution in least squares, 133,184
of, 43 in means, 183

David C. Baird Experimentation An Introduction To Measurement Theory and Experiment Design PDF

Uploaded by

Copyright:

Available Formats

David C. Baird Experimentation An Introduction To Measurement Theory and Experiment Design PDF

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

David C. Baird Experimentation An Introduction To Measurement Theory and Experiment Design PDF

Uploaded by

Copyright:

Available Formats

What are some of the main topics covered in the book?

What are some of the main topics covered in the book?

How does the book describe the process of measurement?

How does the book describe the process of measurement?

EXPERIMENTATION

An Introduction to Measurement Theory

Acquisition Editor: Ray Henderson

© 1995 by Prentice-Hall, Inc.

All rights reserved. No part of this book may be

Prentice-Hall International (UK) Limited, London

Chapter 1: Approach to Laboratory Work, 1

Chapter 2: Measurement and Uncertainty, 10

Chapter 3: Statistics of Observation, 29

3-6 Gaussian Distributions and Sampling, 37

Chapter 4: Scientific Thinking and Experimenting, 57

Chapter 5: Experiment Design, 84

Chapter 6: Experiment Evaluation, 113

6-10 Function Finding, 135

Chapter 7: Writing Scientific Reports, 157

Appendix 1: Mathematical Properties of the Gaussian or

Revolutionary though such changes have been in the actual conduct of

The appendices contain material which, although desirable in itself, would

This book is intended for use in introductory physics laboratories. It was

1 -1 NATURE OF SCIENTIFIC KNOWLEDGE

It is natural to ask why everyone, not just practicing scientists, should be

To do this we must first become knowledgeable about the nature of

Once the nature of measurement has been settled, a second, equally

As can easily be expected, neither extreme position is valid, and public

The point is this. There exists an all-important distinction between the

Unlike statements about measurements, however, statements about our

properties of the system as closely as possible, they remain nothing more

Failure to comprehend the complementary roles of observation and

As far as our own statements on scientific matters are concerned, we

1- 2 ROLE OF THE PHYSICS LABORATORY

experimental methods will be clarified if we regard the piece of the universe

In the example of the automobile engine, we are beginning to reach the

The introductory physics laboratory, therefore, can offer us the oppor­

It may now be obvious that many of the traditional procedures in in­

It is also necessary to learn to work within the framework of the appa­

The writing of laboratory reports should be tackled in the same con­

2 -1 BASIC NATURE QF MEASURING PROCESS

Measurement is the process of quantifying experience of the external world.

It is obvious that the quantifying process (almost invariably involves

We start at the most basic level with an apparently simple measurement

The final outcome of our discussion is most important. As we make

2 - 2 DIGITAL DISPLAY AND ROUNDING OFF

These two examples of digital display illustrate a more general con­

small and unimportant, or they can become significant. In a long calculation,

A similar rounding-off error can appear in statements about measure­

2 - 3 ABSOLUTE AND RELATIVE UNCERTAINTY

Whatever the means by which we make a measurement, the final outcome

measurement and subsequently, whenever the value is quoted or used in

makes a measurement, it is necessary to be vigilant and remember at all

2 - 5 UNCERTAINTY IN CALCULATED QUANTITIES

2 - 6 UNCERTAINTY IN FUNCTIONS OF ONE VARIABLE ONLY

Consider a measured quantity with an uncertainty ±3x (where we are

Figure 2-1 Propagation o f uncertainty from one variable to another.

Although it is helpful to bear in mind the nature of propagated uncer­

2 - 7 GENERAL METHOD FOR UNCERTAINTY

The introductory physics laboratory, therefore, can offer us the oppor

It may now be obvious that many of the traditional procedures in in

It is also necessary to learn to work within the framework of the appa

The writing of laboratory reports should be tackled in the same con

These two examples of digital display illustrate a more general con

A similar rounding-off error can appear in statements about measure

Although it is helpful to bear in mind the nature of propagated uncer

where a and b may be positive or negative, integral or fractional. This formu

In the preceding chapter we considered measurements in which the uncer

Many spreadsheet programs for computers such as Lotus I-2-3, Quat

Many results from measuring processes are presented simply by offer

In Chapter 2 we considered the uncertainty of computed values z, and we as