M.Sc. Physics Lab.

ANALYSIS OF ERRORS
To understand a physical phenomenon experiments are performed. The result
of each experiment depends on various measurements and each measurement has
certain uncertainty or error. We distinguish between a mistake and an error; a mistake
arises due to carelessness of an observer and cab be avoided whereas error is inherent
in the measurement resulting from the process of measurement and limitations of the
instrument. An experimentalist must know how the uncertainty of each measurement
affects the final result. The uncertainty must be carried along in all calculations in a
suitable way as indicated in the following to give an estimate of the error in the final
result. The routine calculation of evaluation of error should be carried out together
with regular numerical calculations and should never be dealt with separately
afterwards.

I. GENERAL SCHEME FOR PRESENTING A CALCULATION

1. Present the formula and define all symbols.

2. Show numerical values put into formula with an estimate of
uncertainty for every single value. Include the proper units for each
numerical value inserted in the formula and satisfy yourself that the
units are consistent and that the equation is dimensionally correct.
3. Show the important steps in the calculation of the results and the
combining of errors, and state the result together with the final estimate
of uncertainty.
4. List immediately the basis for every estimate of uncertainty for the
values used in the formula.
5. In a brief summary, point out what you consider to be the most
important errors discussed above.

II. MATHEMATICAL METHODS FOR COMBINING ERRORS

Let us note some short rules for the combination of errors in independent
quantities. Remember that errors are generally rough estimates – round them off
sensibly.

Separate discussions follow for estimated errors, statistical errors, and

combination of both.

A. Estimated errors

An estimated error is an estimate of the maximum extent to which a reading

might deviate from its true value. It includes random, and in some cases, systematic
errors. You will note that this same technique is used in computing the propagating
errors in dependent quantities where this dependence is not known analytically.

Convenient formulas can often be calculated for combining such error limits
by simply taking derivatives:
 Z = X log Y
Z = X log Y + ( X / Y )Y .

Such formulae usually work only for “small errors” (what does this mean?).

Derive for yourself the following rules:

1. In addition or subtraction, add the absolute errors:

(9.03  0.06) − (3.01  0.01) = 6.02  0.07
2. In multiplication or division, add the percentage errors:
2/3%
(9.03  .6)
= (3.00  0.03)
(3.01  .01)
1/3% 1%
3. When taking powers or roots, multiply the % error by the power to
which the number is raised:
2/3% 1.5x2/3=1%
(9.03  .06) 1.5
(27.2  0.3)
= (3.00  0.03)1.5 = = (9.04  0.12)
(3.01  .01) (3.01  .01)
1/3% 1/3% 1 1/3%
Example
A
Q= −C2
B
(6.024  0.006) dyne cm 2
= − (10.00  0.4) 2 cm 2
(0.050  0.001) dyne
(6.024  0.1%) dyne cm 2
= − (10.00  4%) 2 cm 2
(0.050  2%) dyne
= (120.5  2.1%) − (100  8%) cm2
= (20.5  10.55) cm2

B. Statistical errors

Statistical errors are used when the distribution of repeated readings of the
same variable can be expected to be random and requires the analysis of a “number”
of repeated readings. You should be acquainted with the definition and method of
calculation of quantities such as the mean, standard deviation of the mean, and
probable error. (You will come across these data in the G.M. Counter experiment.)
Recall that the variables are assumed to be completely independent in this
combination of errors.

Instead of combining absolute or percentage errors directly, one combines

them by taking the square root of the sum of the squares (srss). The examples will
make this clear.
1. In addition or subtraction, take the srss of the absolute errors:
.0036 .0001 .0037
(9.03 + 0.06) − (3.01 + 0.01) = (6.02 + 0.06)

2
 (Here, for example, 9.03 is the mean of a set of readings, and 0.06
might be the standard deviation of the mean.)
2. In multiplication or division, take the srss of the percentage errors.
4. When taking powers or roots, multiply the absolute or percentage error
by the power to which the number is raised before squaring it.

These rules are examples of the following more general formula which applies to
statistical errors. Let F be a function of two variables: F = F ( x, y ) . Then if
 F ,  x ,  y designate the probable standard deviations of individual measurements of
F , x and y, respectively, from their mean values, then
2
 F   F 
2

F =  x  +  y  .
 x   y 

C. Combination of standard deviations and estimated errors

Occasionally, it is necessary to combine several variables, some of which have

been measured once or twice so that their errors must be estimated and others
characterized by a mean value and a standard deviation from this value. The estimated
error of the final result is found by treating all errors like standard deviations from the
mean.

III. GRAPHS

A. General comments
A graph cannot be analyzed convincingly without an estimate of the random
and systematic errors associated with each plotted point. Suppose one has the
following collection of points:

3
What does the experimental evidence tell us about the relationship between Y and Y ?
We might, in fact, on the basis of the above alone, be tempted to try various
possibilities like curves A, B and C. Any one or more of these possibilities may be
correct – to decide between these ones must indicate the uncertainty estimated for
each measurement. One standard way of doing this is as follows:

This means that the estimated error or standard deviation for that point has the
indicated magnitudes in both the abscissa and ordinate. Often the error is estimated on
the basis of a variation in one coordinate alone:

If the estimated errors used to describe the uncertainty (uncertainty: range of values
within which the true value is estimated to lie) are too small to be indicated
conveniently by drawing the limits, one should state what they are somewhere on the
graph.
The indicated uncertainties on a graph make it possible to
(a) decide which features of a curve are only spurious variations, and
which are “real”,
(b) estimate the errors in the constants of the equation of any curve drawn
through the experimental points.

B. Calculation of errors

Suppose that there is an expected relationship between two variables such that
one variable is linear, parabolic or some other function of the second variable. It is
always possible by means of a “least squares” treatment to find the constants of the
“best” appropriate curve fitting the data by a purely mathematical computation. This
is a very precise method, but is often very tedious. An alternate approach is to plot the
data if possible, in such a way that a linear plot is expected. Then estimated error can
be computed for the slope and the intercept.

4
 EXPERIMENT NO.1: GRAPHICAL ANALYSIS OF DATA

N.B. Assigned as homework; report to be submitted on the first laboratory day of

the following week.

Material needed:
Qty. Graph Paper

3 mm X mm
1 Semi-log 3 scale = log X mm
1 log-log 3 scale = log X 2 scale

(in addition to pencils, eraser, instrument box, transparent ruler, etc.)

I. Experimental observations yield sets of numbers. In the simplest of cases the

value of one physical quantity (x) is varied and the effect of this variation on another
physical quantity (y) is studied. The experiment yields pairs of values (x1, y1), (x2, y2),
etc. The goal of the experiment may be, for example, to find a relation y(x) describing
the variation, or, if the form of the relation is known from theory, determine the
unknown quantities in the relation. The simplest and the most direct method of doing
this is to plot a graph between x and y.

Although a graph gives the same information as the table of experimental

values, there are several advantages in representing data by a graph. A graph is a
visual representation, and as such is more useful to enable one to see what the relation
y(x) is like. Unexpected features or irregularities are more easily detected in a graph
than in the corresponding table of numbers. From the ‘scatter’ of data points, one can
estimate the accuracy of the measurements. Sharpness of maxima or minima,
symmetry or skewness of maxima or minima etc. are more easily seen in a graph.

One further advantage is ‘interpolation’. By virtue of limitations on time,

experimental observations are generally discrete. Thus values of variables within the
range defined by two successive observations are not available, but they may be
needed. A graph enables any values to be read directly, i.e., without cumbersome
arithmetical procedures. The same can be said about ‘extrapolating’ from a graph.

As mentioned earlier, from a set of values of x and y (and a known relation

y (x) ), the unknown quantities in y (x) can be determined. This involves a great
amount of cumbersome arithmetic which can be avoided by using graphs. For
example, if y = mx + c , the graph will be a straight line and the unknowns m and c
can be determined. That is, graphs are a computational method too. The accuracy is
not as good as that obtained by using the cumbersome arithmetical method but is
usually sufficient for many purposes.

Modern instrumentation gives at the conclusion of an experiment an

automatically recorded curve representing the dependence of one quantity on another.
These graphs are then analyzed to extract information. The use of recording systems
is so widespread that it is essential to know all about graphs. The analysis is never as
simple as that you are supposed to do in this assignment. In this assignment you will
be introduced to some simple techniques.

5
 (It is assumed that you are familiar with the basic ideas about graphs. If you
have any difficulty in the following account, take the help of your instructor.)

II. Experimental observations are always uncertain to some extent. When they are
plotted on a squared sheet of paper they seem to be falling on a curve more or less
depending upon the inaccuracies and uncertainties of the experiment. There is some,
as it is called ‘scatter’. The object or drawing a graph is to circumvent the scatter of
data points. In other words, drawing a graph is in fact an average process just as
taking a mean of a number of readings.

Drawing good graphs is essential if one hopes to get the best out of one’s
results. The following hints will help you draw good and accurate graphs.

Choose your scales well. The most pitiable use of a graph paper is to make use
of only a small part of it by (ignorantly) choosing a compressed scale. Let your scales
be fairly spread out. If a quantity varies between, say, 120 gm and 150 gm choosing a
scale such as 1 cm = 10 gm is absurd and inaccurate. Make full use of the sheet.

It is not at all necessary to start numbering divisions from zero (unless

occasion demands it). In the case cited above, you could very well start at 110 and end
at 160, say. At times it may happen that under these conditions the graph, say a
straight line, cuts the y-axis outside the sheet. This is no problem. By choosing two
suitable points on the graph you can deduce the intercept.

The scales chosen should be simple so that it does not take too long to plot or
locate points. A scale such as 1 cm = 1/3 gm is not a very sensible one (unless the
steps in the variation of the quantity are always 1/3 gm, 2/3 gm, etc. – which is
unlikely).

After you have plotted a data point, draw a small circle around it: . This
prevents the point from being lost should your curve go right through it, and also
makes it easier to locate it. Further, you can distinguish various groups or sets of
points by using different symbols around them: , , , etc.

The writing on a graph sheet (for example, the numbers specifying divisions
on the axes, the data points and their symbols, labeling of different curves) should be
in ink. The trials for the actual curve may be in pencil but if the final choice is in ink,
it is easier to study it. Needless to say that a sharp pencil or a fine pen should be used
to draw the curves. Otherwise the accuracy suffers.

No curve drawn on the basis of data points which show scatter (and scatter is
always present) is unique. If the number of points is too small, any kind of curve may
be made to represent them. On the other hand, it is usually not possible to take a very
large number of readings because of constraints on the available time. Use your
judgment in the matter. If you know that the graph is going to be a straight line, six or
seven points are usually sufficient. (This does not mean that every time you should
take only six or seven readings!) For graphs other than a straight line you need more.
Also if the curve shows maxima or minima more points are needed around the
possible maxima or minima. It is a good practice roughly to draw a graph as you are

6
doing the experiment to enable you to detect such features while you are still in a
position to take more readings. Another example of such features is curves where you
need to find the slope at a particular point.

Drawing a graph is an averaging process, which should give you an average of

all the observations. So do not leave out any data point unless you have justifiable
reason for it and you have explained it in your report. The line, straight or otherwise,
must represent all points. For this it is not necessary that it passes through any or all
points. Sometimes it may not pass through even a single point but may be the best
representative line. Some age-old golden rules may be reaffirmed. The number of
points on one side of the line in any limit region should be roughly the same as the
number of points on the other side. The average deviation of points on one side should
be roughly the same as that of points on the other side. Points, which characteristically
deviate a good deal from the possible line, may be given less weightage (if there is a
good reason to do so) but never ignored completely. When you want to find the slope
of a straight line or the tangent to a curve, choose two points as far away from each
other as possible. This gives better accuracy.

As the following examples will show, most equations can be re-cast to obtain
a straight-line graph. It is useful therefore to see how the best ‘fit’ can be obtained for
a straight line. One simple method of surprising accuracy is the following. Having
plotted all points ( x1 , y1 ), ( x 2 , y 2 ),..., ( x r , y r ),..., x n , y n ) , plot the ‘centroid’, the co-
ordinates of which are given by

x=
x r
, r
. y=
y
n n
Then place a transparent ruler on the graph paper so that the edge passes through this
centroid. Rotate the edge to find the best line, always letting the edge to pass through
the centroid. A line through the centroid drawn with the paragraph above in mind is
the best fit. This method has firm theoretical justification but we shall not go into it at
this stage. Having drawn the best line, rotate the ruler about the centroid so that it
passes through the cluster of points at ‘top right’ and that at ‘bottom left’. This new
line gives one limit (m)1 of the accuracy of the slope m of the best fit. A similar line
drawn on the other side (again, through the centroid) gives the other limit (m)2. You
will always find that the range defined by these two limiting lines is absurdly large. A
more realistic figure for the uncertainty in the slope of the best fit is obtained by
dividing m by n, where n is the total number of points plotted. Express the slope as

m
(m1 ) + (m2 ) .
2 n
As the angles of the limiting lines with the best fit are likely to be small, you can
approximate by taking ()1 and ()2, in radians, for (m)1 and (m)2, respectively.

It is well to remember that this method is justified only if the scatter of points
is due to random errors alone.

It is equally useful to remember that if the scatter is quite small the error in
reading the divisions of the graph and plotting and locating points is what will
determine the uncertainty in m (or in other quantities read from the graph).

7
III. You are supposed to work out the following problems.

1. The following table gives the values of two related quantities, x and y, as
measured in an actual experiment.

x 0.5 2.3 3.8 4.5 8.5 9.4 11.0

y 2.0 2.6 3.8 4.7 7.4 8.8 10.7

x 12.0 14.0 15.0 15.5 16.0 17.0 18.0

y 12.5 14.7 15.8 15.8 16.1 15.7 15.8

x 19.1 20.6 21.2 22.0 22.7 24.0

y 13.8 11.0 8.5 6.0 3.9 0.8

Plot y against x. You will see that there is some scatter as shown in Fig.1a. (All
figures are collected together at the end on a separate sheet.) One way of drawing a
mean line through the data points is to mark the centers (x in Fig.1b) of the short lines
joining neighboring points. You will find it easier to draw a mean line through these
rather than the data points ( in Fig.1a). Again, it is not at all necessary that your
curve passes through each center x. Marking the centers is only an aid to drawing the
mean line.

Question: (1) Why is it easier to draw a curve through the centers than it is to draw a
curve through the original points?

2. You will find that the curve shows a maximum of y. So the next problem is to
locate the maximum. This is done as follows. Draw lines parallel to the x-axis at
suitable intervals of y, starting at the foot of the hump. Mark the centers of the chords
cut by the curve. Then draw a smooth curve through the centers. The intersection of
the line of centers with the main curve defines the maximum. Record the location and
the value of the maximum (see Fig.2). You can use this intersection to improve the
main curve.

Question: (2) What is the slope of the line of centers at the point where it intersects
the main curve?

(3) Will this procedure be better for a broad or a sharp maximum? (See
Fig.3a and 3b)

3. The specific heat C of a metal at temperatures near the absolute zero is given
by
C = aT + bT 3 .

The first term gives the contribution of the electrons while the second gives that of the
crystal. a and b are constants which involve the properties of the metal concerned.
The experiment consists of measuring C at different temperatures (T in K). The table
below gives the experimental values obtained for potassium.

8
 T (K) 0.131 0.186 0.227 0.262 0.293
C (mJ/mole-K) 0.279 0.408 0.502 0.596 0.684

T 0.321 0.371 0.393 0.415 0.454 0.473 0.491

C 0.750 0.918 0.982 1.050 1.192 1.254 1.328

T 0.525 0.541
C 1.480 1.528

Question: (4) If a graph is plotted between T and C a complicated curve will be

obtained. So plot a suitable graph and obtain the values of a and b. Find the
uncertainties in them.

4. Experiment shows that the total energy E radiated by a body at temperature T

depends on some power of T. In other words, it is suspected that the relation is given
by
E = KT n ,

where n and K are not known. The following table gives the experimental values.

T (K) 462 595 680 725 875 920

-1 -2
E (joule s m ) 2583 7106 1213x101 1566x101 2500x101 4060x101

T 1015 1200 1312 1408 1500

E 6015x101 1176x102 1680x102 2228x102 2871x102

Question: (5) Plot a suitable graph so as to obtain values of n and K.

5(A). The next class of functions to study is the ‘exponential function’. Exponential
functions occur so often in the study of natural (and, sometimes, even economic and
social) phenomena that it is essential to know the methods by means of which a
variation can be established to be exponential and the constants involved in the
functional relation determined with some reliability.

In any exponential relation there are two constants involved. One, the
amplitude coefficient (‘a’ below) that gives, so to say, the initial value of a quantity
(‘Y’ below); and two, the rate factor (‘b’ below) that gives the rate of the variation of
Y. Thus

Y = a exp( −bx). (1)

There could be a positive sign in the exponent, depending on the situation, but we
take here the negative sign for the sake of definiteness and also because it occurs
more often.

Consider, for example, the rate of fall of temperature, T, of a body heated to an

initial temperature T0. This rate is proportional to the actual temperature at any
instant:

9
 dT dT
= −bT , or = −bdt , or T = T0 exp (−bt ). (2)
dt T

Another example is that of the decay of a radioactive material. The number, n, of

nuclei, which disintegrate per unit time, is, quite plausibly, a certain fraction of the
actual number of nuclei of the same kind. Thus we could write

dn dn
= − Kn, or = − Kdt , or n = n0 exp (− Kt ). (3)
dt n

(B). Methods of determining the constants a and b

In (C) below are given the experimentally determined values of two variable
quantities which are suspected to be related by Eq.(1). Determine a and b by the
following methods.

(i) Plot the points on the usual squared paper, choosing good scales. You
will need a sheet of size 25 cm x 20 cm. Draw a smooth curve representing the points
(not necessarily through them!) using the smoothing procedure given in III earlier.
Remember to mark the smoothing points differently from the data points.

Extrapolate the curve to x=0 so as to intersect the y-axis at P (Fig.4). Putting

x=0 in Eq.(1), we get y0=a, so that the distance OP directly gives a, if the curve really
represents an exponential variation. This can be established in the following way.

Draw a horizontal line at y=a/2 (using a=OP) to cut the smooth curve at Q.
Then draw a similar line at Y=a/4 to cut the curve at R, and so on, as far as you can.
Let the coordinates of P, Q, R, S, … be (0, a), (x1, a/2), etc. We then have the
relations:

Y0 = a,
Y1 = Y0 / 2 a exp (−b x1 ),
Y2 = Y1 / 2 a exp (−b x2 ),
Y3 = Y2 / 2 a exp (−b x3 ), etc. (4)
Therefore
Y0 / Y1 = exp b ( x1 ),
Y1 / Y2 = exp b ( x2 − x1 ),
Y2 / Y3 = exp b ( x3 − x2 ), etc. (5)

The ratios at the left are all equal, each being equal to 2. Hence the right-hand sides of
these equations must all be equal if the relation between x and Y is exponential. That
is, if x1=(x2-x1)=(x3-x2), etc., we have established the relationship (1). Do this,
commenting on the differences in the values as measured. The property expressed by
Eq.(5) can be stated thus: equal increments of x produces equal fractional changes in
Y. The fraction chosen for the graphical analysis does not need to be 2. You could
choose 3/2 or something else if you wanted.

10
 The next step is to determine b. For doing this write Eq.(5) as
ln 2 = bx1 ,ln 2 = b( x2 − x1 ),ln 2 = b( x3 − x2 ), etc. (6)

Find the average of the multipliers of b. Call it  . Then

ln 2
b= . (7)


Notice that we have used all of our data to get this value of b. This value, however, is
subject to the accuracy with which the point P of Fig.4 can be located. In general, y0 is
very sensitive to the way the curve is drawn but this method suffices for many
purposes.

(ii) For the second method we put Eq.(1) in another form:

ln Y = ln a − bx, or log Y = log a − (log e) bx. (8)

A plot of log Y against x should give a straight line with a slope of –(log e) b, and a y-
intercept of log a, if the quantities x and Y are connected exponentially. Determine a
and b in this way.

Plotting this straight line is very easy if a ‘semi-log graph paper’ is used. In
this, one of the two axes is marked linearly as usual but the other is marked by
divisions proportional to the logarithms (to base 10) of numbers, much as in a slide
rule. Then there is no need to refer to log tables. Plot Eq.(8) on a semi-log paper so as
to get familiar with it. For the purpose of the present study, a 3-cycle semi-log graph
paper is adequate.

(C) At sufficiently high temperatures the electric current through ionic crystals
like sodium chloride is carried by positive or negative ions (or their vacancies).
Unlike the (electronic) current in metals, this current increases with the temperature
and does so rapidly. The ionic conductivity (C) of an ionic crystal is given as a
function of the temperature (T) by the following equation:

−E
C = C 0 exp  , (9)
 kT 

where k is the Boltzmann constant. E is a quantity of the nature of energy, which

specifies the ease with which current carriers are formed and move through the
crystal. An experiment gave the following values for potassium chloride.

T (K) 746 807 830 854 876

Cx107 (cgs units) 1.82 2.91 5.80 9.38 19.2

T 886 913 950 964 991 1014

C 32.8 66.2 119 243 418 800

Question: (6) Analyze these data by the methods described above, and determine C0
and E. Look up the value of k from the table of constants.

11
(D) Questions and problems

(a) Which of the two methods is more accurate? Why? Estimate the errors
numerically.

(b) Prove that the tangent to Y=a exp (-bx) at x=0 cuts the x-axis at a
distance 1/b. This indicates yet another method of determining b. What
objections, if any, would you advance to this method?

(c) When radioactive nuclei of one kind (n, say) disintegrate, they produce
nuclei of another kind (N, say). As n decreases from an initial value n0,
N increases from zero. Draw a schematic graph of n and N on the same
axes.

12
13