Error Analysis

Introduction
Physics is an experimental science. The physicist uses mathematics but the
models he or she constructs arent abstract fantasiesthey must describe the real
world. All physical theories are inspired by experimental observations of nature
and must ultimately agree with these observations to survive. The interplay
between theory and experiment is the essence of modern science.
The task of constructing a theory, inherently difficult, is compounded by the fact
that the observations are never perfect. Because instruments and experimenters
depart from the ideal, measurements are always slightly uncertain. This
uncertainty appears as variations between successive measurements of the same
quantity.
With better instrumentation and greater care, these fluctuations can be reduced
but they can never be completely eliminated1. One can, however, estimate how
large the uncertainty is likely to be and to what extent one can trust the
measurements. Over the years powerful statistical methods have been devised
to do this. In an imperfect world we can expect no more; and once we
understand the limitations of an experiment we can, if we feel the need, try to
improve on it.
In the laboratory you, like any scientist, will have to decide just how far you can
trust your observations. Here is a brief description of experimental uncertainties
and their analysis. Your TA will indicate how far you need to carry the analysis
in each experiment.

In certain instances, there is a fundamental limit imposed on this process of improvement. Quantum
mechanics, which deals with very small systems such as atoms and nuclei, suggests that there is an inherent
uncertainty in nature that even perfect instrumentation cannot overcome. Luckily, most experiments do not
confront this ultimate barrier.

6.2 Definitions
The difference between the observed value of a quantity and the true value is
called the error of the measurement. This term is misleading; it does not
necessarily imply that the experimenter has made a mistake. If this were the case,
and the experimenter would correct it! Uncertainty is a better term but it is not
as commonly used; well use them interchangeably here.
Errors may be conveniently classified into three types:
A. Illegitimate Errors: These are the true mistakes or blunders either in
measurement or in computation. Reading the wrong scale or misplacing a
decimal point in multiplying are examples. These errors usually stand out if
the data is examined critically. You can correct such errors when you find
them by eliminating their cause and possibly by repeating the measurement.
If its too late for that, you can at least guess where the mistake is likely to lie.
B. Systematic Errors: These errors arise from faulty calibration of equipment,
biased observers, or other undetected disturbances that cause the measured
values to deviate from the true valuealways in the same direction. The
bathroom scale that reads 3 lbs before anyone steps on it exhibits a
systematic error. These errors cannot be adequately treated by statistical
methods. They must be estimated and, if possible, corrected from an
understanding of the experimental techniques used.
Systematic errors affect the accuracy of the experiment; that is, how closely the
measurements agree with the true value.
C. Random Errors: These are the unpredictable fluctuations about the average
or true value that cannot be reduced except by redesign of the experiment.
These errors must be tolerated although we can estimate their size. Random
errors affect the precision of an experiment; that is, how closely the results of
successive measurements are grouped.
Note that accuracy and precision are different. See if you can think of an example
of something that is precise but not accurate, and something that is accurate but
not precise.

The concepts of error analysis that are introduced below are strictly applicable
only to random errors.

Practices and Concepts

The result of a measurement is usually recorded as A A where A is the best
estimate for the value and A means the uncertainty in A . The interpretation
is that the true value of A lies between A A and A A . The length of a
wood block recorded as 5..2 0 3 cm would be expected to lie between 4.9 and
5.5 cm. (Technical note: A is one standard deviation. The true value lies
within one standard deviation of the mean 68% of the time and within two
standard deviations of the mean 95% of the time, if the errors are random. See
section B).
A. Significant Figures: The number of figures in a result A is, by itself, often
indicative of the uncertainty. In the example above, the value of 15.2 cm
implies that the error will be at most several tenths of a cm since it isnt
physically realistic to compute an error to more than one (occasionally two)
significant figures. Accordingly, it wouldnt make sense to write the result as
15 0.3 cm since writing 15 implies that we know nothing about tenths of
centimeters. Likewise, writing 15.2 1 cm is inconsistent. The result and the
error estimate must always be in agreement concerning the least uncertain
decimal place.
B. Interpretation, Best Value: When a result is quoted with an error estimate,
the error value has, ideally, a rather precise meaning. The interpretation that
the true value must be within the error limits and that any value within these
limits (as shown on the left of Fig. 1) is unphysical. A more accurate picture
is given by the familiar bell shaped curve (on the right of Figure 1), which is
the Gaussian distribution. This is the most frequently occurring distribution
of sampling theory.

Figure 1: Uniform (left) and Gaussian (right) distributions.

The interpretation of this distribution is that the most likely value is the best
guess A ; nearby values are only slightly less likely; but there remains a small
chance that even a faraway value might be the correct one. For a careful
experiment in which the errors have been properly analyzed, the value of A
gives an estimate of the precision expected of the best guess A . There is a
68% chance that the true value lies within the limits A A and A A , and
one can predict from the Gaussian, or normal, distribution the probability for
larger or smaller limits.
C. Estimating Errors: How can one estimate the size of experimental
uncertainties? The first and simplest way is to use the least count of the
measuring instrument. For instance, a ruler graduated in millimeters would
certainly allow you to determine lengths to the nearest mm. Probably by
interpolating between gradations, you could estimate the length to 0.3 or
even 0.2 mm. You couldnt convincingly estimate the length to within 0.1
mm. Thus the precision of the instrument introduces an uncertainty whose
magnitude you know.
As a minimal effort whenever you are recording data, you should always
include such an error estimate for every bit of data recorded. You might not
subsequently use this information and its magnitude may turn out to be
unrealistic. But the practice you receive in considering such uncertainties is
very important.
D. Repeated Measurements: When there is time or when the piece of data is
crucial, you can repeat the measurement several times. The scattering of
values about the average shows how large the random error must be. For
instance, in the set of measurements 1.0, 1.1, 1.3, 0.9, 1.3 cm, a reasonable

guess for the best value is the average, 1.1 cm, and the uncertainty in each
measurement seems to be about 0.2 cm.
1. Best Value: Statistics does show that the average, or mean, of a set of
measurements provides the best estimate of the true value. This is simply
the sum of the measurements divided by the number taken:

xave x

1
N

x .
i

i 1

Here x is the mean of the N measurements of the quantity x , labeled xi

where the index i runs from 1 to N .
2. Standard Deviation: The best estimate of the error in this mean is shown
by statistics to be derived from the standard deviation. The sample
standard deviation is found by taking the differences of each xi value
from the mean, squaring these differences and added them, dividing by
N 1 , and then taking the square root:

1
N 1

i 1

The factor N 1 is there because for N 1 , S 0 / 0 ; the standard

deviation is indeterminate. This makes sense; from only one
measurement it is impossible to say anything about the experimental
error.
3. Standard Error of the Mean: The standard deviation is not quite the error
estimate that is needed. The sample standard deviation, S , is the
estimated uncertainty of each individual measurement xi . The value S
fluctuates as more samples are taken, but it doesnt systematically get
smaller. Haven taken N repeated measurements, our estimate of the true
value is the mean, x . The standard error (or standard deviation) of the mean
is the estimate of the error in this value. The standard error of the mean is
related to the standard deviation by a simple expression:

This important result will be derived (and understood) when we examine

another source of error, propagation errors, in the next computer lab.
However, notice that the uncertainty of the mean is reduced by a factor of
N from the uncertainty in an individual measurement. This is the
reason for repeatedly measuring the same quantity. Note, however, that
the reduction is only by N . Making 100 measurements instead of 10
only reduces the uncertainty by a factor of 3.
Calculating the standard deviation can be tedious. However there are
programs for computers and programmable calculators which give the
mean and standard deviation of a series of entries. Some scientific
calculators have this program built in.
Example: Suppose we measured the length of a line four times and obtained
values of 5.5, 5.3, 4.9, 4.7 cm. These values sum to 20.4 cm, and our best estimate
1
of the length is the mean value x (20.4) 5.1 cm. The error estimate
4
associated with each measurement is the standard deviation:

2
2
2
2
1
5.5 5.1 5.3 5.1 4.9 5.1 4.7 5.1 0.37 cm

The set of four measurements taken together gives us an estimated error m of

the mean:

0.37
0.18 cm
4

The final result for the length of the line is then 5..1 0 2 cm.

Significant Figures
There is a natural shorthand for the estimate of errors and their propagation in
the use of a definite number of figures to represent in a measurement. These
figures are called significant figures. The measurement 15.23 cm, for example, has
four significant figures; it is uncertain to a few one-hundredths of a centimeter
(the exact uncertainty is deliberately left vague). It is different from 15 cm, 15.2
cm, 15.230 cm, and 15.2300 cm; those numbers might all represent the same
measurement, but the uncertainty is different in each case.

It makes no difference, however, whether we write 15.23 cm, 0.l523 m, or 152.3

mm; each number has four significant figures. A number with m significant
figures has a fractional uncertainty of 1 part in 10m .
In adding or subtracting numbers, the largest uncertainty will dominate. This
belongs to the number whose least significant figure is farthest to the left (that is,
the last significant digit of the most imprecise number); the sum (or differences)
will have no significant figures beyond this point. For instance, adding the
numbers below gives:
15.2304
2.13
489.5
62
569

For multiplication and division, the largest fractional error will dominate. It occurs in
the number with the fewest significant figures. Hence the result can have no
more significant digits than the least accurate of the factors. As an example:

15.23 471
380
19
The answer is not 377 or 377.5437. We could be even more definite by writing
3.8 102 .

Propagation of Errors
In the laboratory we seldom measure directly the quantities or results of interest.
Instead we must measure others from which the results are derived. For
example, to measure the volume of a rectangular solid, we measure the three
sides and multiply these values. In the course of such a calculation, the errors in
the measured quantities propagate through the computation to affect the
result.
The basic equation which describes the propagation of errors is best expressed in
terms of partial derivatives. Suppose that we require a quantity P that is a

that we actually can measure. P may be

function of a set of variables a , b , c ,

expressed as

P f a,b, c,

For the rectangular solid, V is a function of the length

and we could write:
V f , w, h w h

, width w , and height h ,

The mean values of a , b , c ,

are substituted into f to give a value for P . The
relation between the error in P ( P ) and the errors in a , b , c ,
( a , b , c , )
is given by the following equation:

P2

f
f
f
a2 b2 c2
a
b
c

(1)

Consult the following references if you want to see how this result is obtained:
P. R. Bevington, Data Reduction and Error Analysis for the Physics Sciences
H. Young, Statistical Treatment of Experimental Data
Y. Beers, Introduction to the Theory of Error

f
. It means: take the
a
derivative of f with respect to a keeping all the other variables b , c ,
constant.
The partial derivative of f with respect to a is written as:

We will show how this general formula is applied to some specific cases of error
propagation.

y a b

A. Addition:

The estimated errors on the measured quantities a and b are a and b . What
is the uncertainty in y ?
Evaluate:

y
1 ;
a

y
1
b

and use Eqn. (1) to find:

 y2 a2 b2
B. Subtraction:

y a b

You can check that:

y2 a2 b2

C. Multiplication:

y a b

Evaluate:

The errors are added in quadrature.

The same result as in addition.

y
y
b (when b is kept constant);
a (when a is kept constant)
a
b

y2 b 2a2 a 2b2

So that

This result is better expressed in terms of fractional or relative errors, y y , a a ,

b b :
2

y
a 2 b 2

y a b

Thus the fractional errors are added in quadrature.

y

D. Division:

We have:

y 1
;
a b

a
b

y
a
2
b
b

y2

so that:

1
b

2
2 a

a2
b

b2

a 2
By dividing each term by y 2 , we see that, again, the fractional errors are
b
added in quadrature:
2

2
2
y
a b

y
a b

E. Power Law:

y an

We find:

y
na n 1
a

y na n 1 a

so that:

or

n a
y
a

Thus if y a 3 , the fractional error in y is three times the fractional error in a .

Example: Calculate the acceleration g due to gravity by dropping a ball a
distance x in time t . You know that x 21 gt 2 . How do the uncertainties in x
and t affect the uncertainty in g ?
Answer: Since x 21 gt 2 , then g

2x
t

. Then

g
2
g
4x
2 and
3 , and
x t
t
t

therefore:
2

2 2 4x 2
x 3 t
t2
t

g2

What Do We Expect from You in the Lab?

By now you may feel overwhelmed by the depth to which error analysis can be
carried. We do not want to obscure the physics involved, however. There would
be no point in exhaustively discussing errors for an experiment that you do not
fully understand. What you should attempt to do is:
1. understand what sources of error are present in your
experiment,
2. estimate how significant they are and how they influence your
result, and
3. determine the deviation of the result that you find from the one
you expect can be accounted for by these errors.
It is better to find an inaccurate result with large sources of error that you can
explain than to fudge a right answer that cannot be justified by the method.