ERROR ANALYSIS

"To err is human; to evaluate and analyze the error is scientific".

I Introduction
Every measured physical quantity has an uncertainty or error associated with it.
An experiment, in general, involves
(i) Direct measurement of various quantities (primary measurements) and
(ii) Calculation of the physical quantity of interest which is a function of the measured
quantities. An uncertainty or error in the final result arises because of the errors in the
primary measurements (assuming that there is no approximation involved in the
calculation). For example, the result of a recent experiment to determine the velocity of
light (Phys. Rev. Lett. 29, 1346 (1972)) was given as

C = (299, 792, 456.2  1.1) m/s.

The error in the value of C arises from the errors in primary measurements, viz.,
frequency and wavelength. [g, σ]
-----------------------------------------------------------------------------------------------------------
Physics Today, July 2007
C = 299, 792, 458 m/sec.

Physical Quantity Relative standard uncertainty

2006 2002

Electronic charge 2.5 x 10-8 8.5 x 10-8

Planck’s constant 5 x 10-8 17 x 10-8

Avogadro number 5 x 10-8 17 x 10-8

Fine structure constant (e2/ħc) 6.8 x 10-10 33 x 10-10

Compton wavelength (ħ/m0c) 1.4 x 10-9 6.7 x 10-9

“G” 1.0 x 10-4 1.5 x 10-4

“C” < 3.3 x 10-9 ----

-----------------------------------------------------------------------------------------------------------
Error analysis, therefore, consists of
(i) Estimating the errors in all primary measurements, and
(ii) Propagating the error at each step of the calculation. This analysis serves two
purposes. First, the error in the final result ( 1.1 m/sec in the above example) is an
indication of the precision of the measurement and, therefore, an important part of
the result. Second, the analysis also tells us which primary measurement is causing

1
more error than others and thus indicates the direction for further improvement of the
experiment.

For example, in measuring 'g' with a simple pendulum, if the error analysis
reveals that the errors in 'g' caused by measurements of l (length of the pendulum)
and T (time period) are 0.5 cm/sec2 and 3.5 cm/sec2 respectively, then we know that
there is no point in trying to devise a more accurate measurement of l. Rather, we
should try to reduce the uncertainty in T by counting a larger number of periods or
using a better device to measure time. Thus, error analysis prior to the experiment is
an important aspect of planning the experiment.

Nomenclature

(i) 'Discrepancy' denotes the difference between two measured values of the same
quantity.
(ii) 'Systematic errors' are errors which occur in every measurement in the same
way-often in the same direction and of the same magnitude - for example, length
measurement with a faulty scale. These errors can, in principle, be eliminated or
corrected for.
(iii) 'Random errors' are errors which can cause the result of a measurement to deviate
in either direction from its true value. We shall confine our attention to these
errors, and discuss them under two heads: estimated and statistical errors.

II Estimated Errors
Estimating a primary error

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

quantity might deviate from its true value. For a primary measurement, the estimated
error is often taken to be the least count of the measuring instrument. For example, if
the length of a string is to be measured with a meter stick, the limiting factor is the
accuracy in the least count, i.e. 0.1 cm. [1/2 LC] A note of caution is needed here.

What matters really is the effective least count and not the nominal least count.
For example, in measuring electric current with an ammeter, if the smallest division
corresponds to 0.1 amp, but the marks are far enough apart so that you can easily make
out a quarter of a division, then the effective least count will be 0.025 amp. On the
other hand, if you are reading a vernier scale where five successive marks on the
vernier scale (say, 27th to 31st ) look equally well in coincidence with the main scale, the
effective least count is 3 times the nominal one. Therefore, make a judicious estimate
of the least count. The estimated error is, in general, related to the limiting factor in the
accuracy. This limiting factor need not always be the least count. Another example,
in a null-point electrical measurement, suppose the deflection in the galvanometer
remains zero for all values of resistance R from 351  to 360 . In that case, the
uncertainty in R is 10  (rather 5 ), even though the least count of the resistance box
may be 1  . [Stop-watch]

Propagation of estimated errors

How to calculate the error associated with f, which is a function of the measured
quantities a, b, and c ?

2
Let f = f (a, b, c). (1)

From differential calculus (Taylor's series in the 1st order)

f f f
df  da  db  dc . (2)
a b c

Eq. (2) relates the differential increment in f resulting from differential increments in a,
b, c. Thus, if our errors in a, b, c (denoted as a, b, c) are small compared to a, b, c,
respectively, then we may say

f f f
f  a  b  c . (3)
a b c

where the modulus signs have been put because errors in a, b, and c are independent of
each other and may be in the positive or negative direction. Therefore, the
maximum possible error will be obtained only by adding absolute values of all the
independent contributions. (All the ‟s are considered positive by definition). Special
care has to be taken when all errors are not independent of each other. This will become
clear in special case (V) below.

Some simple cases

(i) For addition or subtraction, the absolute errors are added, e.g.,
if f  a  b  c , then
f  a  b  c . (4)

(ii) For multiplication and division, the fractional (or percent) errors are added,
e.g.,
ab
if f  , then
c
1 1 1 1
f  a  b  c . (5)
f a b c

(iii) For raising to constant powers, including fractional powers, the fractional error
is multiplied by the power, e.g.,

if f  a 3.6 , then
1 1
f  3.6  a . (6)
f a

(iv) In mixed calculations, break up the calculation into simple parts, e.g.,

3
 3
a
if f   c 2 , then
b
a  3
f        c 2  .
b  
b a 1 1
As    a  b
a b a b

1  3
 3
and 3
  c 2   c
2   2c
c
1
1 a 3
So, f  a  2 b  c 2 c . (7)
b b 2

Note that the same result could have been derived directly by differentiation.
ab
(v) Consider f   a2 .
c
The relation for error, before putting the modulus signs, is

b a  ab 
f   a   b   2 
c  2aa .
c c c 

Note that the a factors in the first and fourth terms are not independent errors.
Therefore, we must not add the absolute values of these two terms indiscriminately.
The correct way to handle it is to collect the coefficients of each independent errors
before putting modulus signs, i.e.,

b a ab
f   2a a  b  2 c . (8)
c c c

III Statistical Errors

Statistical distribution and standard deviation

Statistical errors arise when making measurements on random processes, e.g.,

counting particles emitted by a radioactive source. Suppose we have a source giving
off 10 particles/sec. on the average. In order to evaluate this count rate
experimentally, we count the number of particles for 20 seconds. Shall we get 200
counts? Not necessarily. In fact, we may get any number between zero and infinity.
Therefore, in a measurement on a random process, one cannot specify a maximum
possible error. A good measure of uncertainty in such a case is the standard
deviation (s.d.) which specifies the range within which the result of any measurement is
"most likely" to be.

4
The exact definition of "most likely" depends on the distribution governing the
random events. For all random processes whose probability of occurrence is small
and constant, Poisson distribution is applicable, i.e.,

mn  m
Pn  e , (9)
n!
where Pn is the probability that you will observe a particular count n, when the
expectation value is m.

It can be shown that if an infinite number of measurements are made, (i) their
average would be m and (ii) their standard deviation (s.d.) would be √m, for this
distribution. Also, if m is not too small, then 68% or nearly two-thirds of the
measurements would yield numbers within one s.d. in the range m √m.

In radioactive decay and other nuclear processes, the Poisson distribution is

generally valid. This means that we have a way of making certain conclusions without
making an infinite number of measurements. Thus, if we measure the number of
counts only once, for 100 sec , and the number is, say 1608, then (i) our result for
average count rate is 16.08/sec , and (ii) the standard deviation is √1608 = 40.1 counts
which corresponds to 0.401/sec. So our result for the count rate is (16.1  0.4) sec-1.
The meaning of this statement must be remembered. The actual count rate need not
necessarily lie within this range, but there is 68% probability that it lies in that
range.

The experimental definition of s.d. for k measurements of a quantity x is

k
 x n2 
x     ,
n 1  k  1  (10)

where xn is the deviation of measurement xn from the mean. However, since we
know the distribution, we can ascribe the s.d. even to a single measurement.

Propagation of statistical errors

For a function f of independent measurements a, b, c, the statistical error f is

 f   f   f
2 2 2

f   a    b    c  . (11)
 a   b   c 

A few simple cases are discussed below.

(i) For addition or subtraction, the squares of errors are added, e.g.
if f  a bc
then,  2f   2a   b2   2c . (12)

(ii) For multiplication or division, the squares of fractional errors are added, e.g.

5
 ab
if f  ,
c
 f
2
      
2 2 2

then     a   b   c  . (13)
 f   a   b   c 
(iii) If a measurement is repeated n times, the error in the mean is a factor √n less than
the error in a single measurement, i.e.,
f
f  . (14)
n

Note that Eqs. (11-14) apply to any statistical quantities a, b, etc, i.e., primary
measurements as well as computed quantities whereas

m  m (15)

applies only to a directly measured number, say, number of -particle counts

but not to computed quantities like count rate.

IV Miscellaneous
Repeated measurements

Suppose a quantity f, whether statistical in nature or otherwise, is measured n

times. The best estimate for the actual value of f is the f average of all
measurements. It can be shown that this is the value with respect to which the sum of
squares of all deviations is a minimum. Further, if errors are assumed to be randomly
distributed, the error in the mean value is given by

f
f , (16)
n
where f is the error in one measurement. Hence one way of minimising random
errors is to repeat the measurement many times.

Combination of statistical and estimated errors

In cases where some of the primary measurements have statistical errors

and others have estimated errors, the error in the final result is indicated as a s.d. and
is calculated by treating all errors as statistical.

Errors in graphical analysis

The usual way of indicating errors in quantities plotted on graph paper is to

draw error bars. The curve should then be drawn so as to pass through all or most of the
bars.

6
 Here is a simple method of obtaining the best fit for a straight line on a graph.
Having plotted all the points ( x1 , y1 )........( xn , yn ), plot also the centroid ( x , y ).
Then consider all straight lines through the centroid (use a transparent ruler) and visually
judge which one will represent the best mean.

Having drawn the best line, estimate the error in slope as follows. Rotate
the ruler about the centroid until its edge passes through the cluster of points at the
'top right' and the 'bottom left'. This new line gives one extreme possibility; let the
difference between the slopes of this and the best line be m1 .Similarly determine m2
corresponding to the other extreme. The error in the slope may be taken as
m1  m2 1 (17)
m  . ,
2 n

where n is the number of points. The factor √n comes because evaluating the slope
from the graph is essentially an averaging process.
It should be noted that if the scale of the graph is not large enough, the least
count of the graph may itself become a limiting factor in the accuracy of the result.
Therefore, it is desirable to select the scale so that the least count of the graph paper is
much smaller than the experimental error.

Significant figures (Hide)

A result statement such as f = 123.4678  1.2331 cm contains many superfluous

digits. Firstly, the digits 678 in quantity f do not mean anything because they represent
something much smaller than the uncertainty f. Secondly f is an approximate estimate
for error and should not need more than two significant figures. The correct expression
would be f = 123.5  1.2 cm.

7
V Instructions
1. Calculate the estimated/statistical error for the final result. In any graph you plot,
show error bars. (If the errors are too small to show up on the graph, then write them
somewhere on the graph).

2. If the same quantity has been measured/calculated many times, you need not
determine the errors each time. Similarly, one typical error bar on the graph will be
enough.

3. In propagating errors, the contributions to the final error from various

independent measurements must be shown. For example

f a b
if f  ab;   ,
f a b
a  10.0  0.1, b  5.1  0.2
 0.1 0.2 
then , f  51.0 
10.0 5.1 
 0.51  2.0
 2.5
Therefore, f  51.0  2.5 .

Here the penultimate step must not be skipped because it shows that the contribution to
the error from b is large.

4. Where the final result is a known quantity (for example, e/m), show the discrepancy
of your result from the standard value. If this is much greater or even less than the
estimated error, this is abnormal and requires explanation. (g = 9.3  0.1 or 10.0 
0.5) OR (g = 9.8  0.5 or 9.8  0.1).

5. Where a quantity is determined many times, the standard deviation should be

calculated from Eq.(10). Normally, the s.d. should not be more than the estimated
error. Also the individual measurements should be distributed only on both sides of
the standard value.
EXERCISE: L = (1.1  0.1) cm. What is A  A ???

VI Mean and Standard Deviation

If we make a measurement x1 of a quantity x, we expect our observation to be
close to the quantity but not exact. If we make another measurement we expect a
difference in the observed value due to random errors. As we make more and more
measurements we expect them to be distributed around the correct value, assuming that
we can neglect or correct for systematic errors and avoid blunders. If we make a
very large number of measurements we can determine how the data points are
distributed in the so-called parent distribution. In any practical case, one makes a
finite number of measurements and one tries to describe the parent distribution as best as
possible.

8
Consider N measurements of quantity x, yielding values x1, x2 …..xN. One defines

 1 N 
Mean : x  lim   xi  , (18)
N 
 N i 1 

which is equivalent to the centroid or average value of the quantity x.

Deviations: The deviation di of any measurement xi from the mean x of the parent
distribution is defined as

d i  xi  x . (19)

Note that if the x is the true value of the quantity being measured, di is also the true
error in xi .
The arithmetic average of the deviations for an infinite number of observations
must vanish, by definition of (Eq.(18)

 1 N   1 N 
lim   xi  x   lim   xi   x  0 . (20)
N 
 N i 1  N 
 N i 1 

There are several indices one can use to indicate the spread (dispersion) of the
measurements about the central value, i.e., the mean value. The dispersion is a
measure of precision. One can define average deviation d as the average of the
magnitudes of the deviations (absolute values of the deviations)
d  lim   xi  x  .
1 N
N  N
i 1

This can be used as a measure of the dispersion of the expected observation about the
mean. However, a more appropriate measure of the dispersion is found in the
parameter called standard deviation , defined as

1 N 2
 1 
  lim   xi  x    lim  x   x  .
2 2 2
i (21)
 i 1
N  N
 
N  N


2 is known as VARIANCE and the STANDARD DEVIATION  is the square root

of the variance. In other words it is the root mean square (rms) of the deviations. That
is
N

d i
2

  i 1
. (22)
N
The expression derived from a statistical analysis is
N

d i
2

(23)
  i 1
( N  1)

9
where the denominator is N-1 instead of N. In practice, the distinction between these
formulae is unimportant. According to the general theory of statistics the reliability
of a result depends upon the number of measurements and, in general, improves with
the square root of the number.

Significance: x , the mean, is a parameter which characterizes the information

we are seeking when we perform an experiment. The mean is, of course, not the most
probable value if the parent distribution is not symmetrical but nevertheless it is
a reasonable parameter to characterize the distribution. In place of mean, one can
characterize the distribution in terms of median or most probable value (mode). It can
be proved that if we use the average (mean) of the measured values for
calculating the deviations, the sum of the square of the deviations is a minimum. The
standard deviation is simply related to this minimum value of the square of the
deviations and is used for specifying error quantitatively.
The standard deviation characterizes the uncertainties associated with our
experimental attempts to determine the "true" value, namely the mean value (defined
by Eq.(18)) for all practical purposes. , for a given finite number of observations, is
the uncertainty in determining the mean of the parent distribution. Thus it is an
appropriate measure of the uncertainty in the observations.

VII Method of Least Squares

Our data consist of pairs of measurements (xi, yi) of an independent variable x

and a dependent variable y. We wish to fit the data to an equation of the form

y  a  bx (24)

by determining the values of the coefficients a and b such that the discrepancy is
minimized between the values of our measurements yi and the corresponding
values y = f(xi) given by Eq. (24). We cannot determine the coefficients exactly with
only a finite number of observations, but we do want to extract from these data the
most probable estimates for the coefficients.
The problem is to establish criteria for minimizing the discrepancy and
optimizing the estimates of the coefficients. For any arbitrary values of a and b, we can
calculate the deviations yi between each of the observed values yi and the
corresponding calculated values

yi  yi  a  bxi . (25)

If the coefficients are well chosen, these deviations should be relatively small. The
sum of these deviations is not a good measure of how well we have approximated
the data with our calculated straight line because large positive deviations can be
balanced by large negative ones to yield a small sum even when the fit is bad. We might
however consider summing up the absolute values of the deviations, but this leads
to difficulties in obtaining an analytical solution. We consider instead the sum of the
squares of deviations.
There is no unique correct method for optimizing the coefficients which is
valid for all cases. There exists, however, a method which can be fairly well justified,
which is simple and straightforward, which is well established experimentally (???)
as being appropriate, and which is accepted by convention. This is the method of
least squares which we will explain using the method of maximum likelihood.

10
Method of maximum likelihood

Our data consist of a “sample of observations” extracted from a parent

distribution which determines the probability of making any particular observation.
Let us define parent coefficients a0 and b0 such that the actual relationship between
y and x is given by

y( x )  a0  b0 x . (26)

For any given value of x = xi, we can calculate the probability Pi for making the
observed measurement yi, assuming a Gaussian distribution with a standard deviation
i for the observations about the actual value y(xi). It is given by

1  1  y  yx  2 
Pi  exp   i i
  . (27)
 i 2  2   i  
The probability for making the observed set of measurements of the N values of yi is the
product of these probabilities
 1 
 1  yi  yxi   
2

Pa0 , b0   Pi    exp     ,
 i 2 i (28)

   
2

where the product  is taken for i ranging from 1 to N.
Similarly, for any estimated values of the coefficients a and b, we can calculate the
probability that we should make the observed set of measurements

 1    y 
2

Pa, b     exp  1  i
   . (29)
  2   2   i 
 i  

The method of maximum likelihood consists of making the assumption that the
observed set of measurements is more likely to have come from the actual parent
distribution of Eq. (26) than from any other similar distribution with different
coefficients and, therefore, the probability in Eq. (28) is the maximum probability
attainable with Eq. (29). The best estimates for a and b are therefore those values
which maximise the probability in Eq. (29). The first term of Eq. (29) is a constant,
independent of the values of a or b. Thus, maximizing the probability P(a, b) is
equivalent to minimizing the sum in the exponential. We define the quantity 2 to
be this sum

 1 
2
 y 
    i
2
    2  yi  a  bxi 2  , (30)
 i   i 
N

where  always implies  and consider this to be the appropriate measure of the
i 1
goodness of fit.
Our method for finding the optimum fit to the data will be to minimize this
weighted sum of squares of deviations and, hence, to find the fit which produces the
smallest sum of squares or the least-squares fit.

11
 Minimizing 2: In order to find the values of the coefficients a and b which
yield the minimum value for 2, we use the method of differential calculus for
minimizing the function with respect to more than one coefficient. The minimum value
of the function 2 of Eq. (30) is one which yields a value of zero for both of the partial
derivatives with respect to each of the coefficients.

 2   
   2  y i  a  bxi 2 
1
a a   i 
(31)

2
2 
 yi  a  bxi   0

 2
 x  y  a  bxi   0 ,
2
  2
b 
i i

where we have for the present assumed all of the standard deviations equal, i = . In
other words, errors in y's are assumed to be the same for all values of x. These
equations can be rearranged to yield a pair of simultaneous equations

y i  aN  b x i
(32)
and x y i i  a  x i  b x i2 ,
N

where we have substituted Na for a

i 1
since the sum runs from i = 1 to N.

We wish to solve Eqs. (32) for the coefficients a and b. This will give us the
values of the coefficients for which 2, the sum of squares of the deviations of the
data points from the calculated fit, is a minimum. The solutions are:

a
1

 x  y   x  x y 
2
i i i i i

b
1
N  xi yi   xi  yi  (33)


  N  xi2   xi 
2
.

Errors in the coefficients a and b: Now we enquire what errors should be assigned
to a and b. In general, the errors in y's corresponding to different values of x will be
different. To find standard deviation in 'a', say Sa, we approach in the following way.
The deviations in 'a' will get contributions from variations in individual yi's. The
contribution of the deviation, yn of a typical measured value, yn to the standard
deviation Sa is found using the first equation of Eq.(33) reproduced below as

a
 x  y   x  x y 
2
n n n n n
. (34)
N  x   x  2 2
n n

12
By differentiating it partially with respect to yi we get

a
yi 
 xn   xn xi y .
2
(35)
yi N  xn2   xn 
2 i

Since yi is assumed statistically independent of xi we may replace yi by its average
value
yi 2
yi   y   N
. (36)

Thus this contribution becomes

a   xn   xn xi 
2

yi   y   .
 N  xn2   xn  
yi 2 (37)
 

The standard deviation Sa of the slope „a‟ is found by squaring this expression,
summing over all measured values of y (that is, summing the index i from 1 to N), and
taking the square root of this sum. Also, it should be realized that xi = xn, and x2i =
x2n. The result of this calculation is

x 2

Sa   y
N  x   x  
n
2
. (38)
2
n n

In a similar manner, the standard deviation Sb of the intercept „b‟ can be found as

N
Sb   y
N  x 2
n   xn 
2
. (39)

13
EXAMPLE I

(T) for some Ni-rich Ni-Fe-Cr permalloys

(T)/min vs. T taken with a home-made apparatus.

• All show minima at Tmin.
• Data every 50 mK below Tmin.

For 2 K < T < Tmin/2

Very dilute alloys show Kondo minima where (T) is given by

(T) = 0 – m ln T. Eq. (1)

e - e interaction in the presence of weak localization gives

(T) = 0 – m‟ T. Eq. (2)

14
 Random residuals, order of magnitude smaller values of normalized 2 ( ~10-10)
for the √T fit as against systematic residuals & 2 ~10-9 for the lnT fit.
Exptal. accuracy of ours is ~ 5 ppm (how ???, repeated mmts., fluctuations of
nanovoltmeter, crimp connections).
Goodness of fit ~ 10 ppm [2 ( ~10-10)]  Good fit to that extent!!!
If our accuracy was better than 1 ppm, the fit would have been considered poor.

EXAMPLE II

Spin waves in bulk permalloys

M (T) = M (0)*(1+T3/2+T5/2), Fit from 2 K till TC/10 or 10.

3.335 Only 1 % change in M in 80 K

3.330 Equation y=P1*(1+P2*x^1.5) Red Colour

Adj. R-Squ 0.99843

Value Standard E
3.325 Long Mom P1 3.33323 5.09382E-
Long Moment (EMU)

ent (EMU) 5
Long Mom P2 -1.11889 4.56836E-
ent (EMU) E-5 8
3.320

Equation y=P1*(1+P2*x^1.5+P3*x^ Blue Colour

2.5)
3.315 Adj. R-Squar 0.99941
Value Standard Err
Long Momen P1 3.33286 4.29883E-5
t (EMU)
3.310 Long Momen P2 -9.43288E- 1.43395E-7
t (EMU) 6
Long Momen P3 -2.34184E- 1.87559E-9
t (EMU) 8

3.305

0 10 20 30 40 50 60 70 80 90
Temperature (K)

15
 Sample M(0)   C R2 D
Composition (emu) (10-4 K-3/2) (10-6 K- (K) Norm meVÅ2

5/2
(Sample ID) )

Ni81.0Fe16.7Mo2.3 3.33323± -0.1189± - 720 0.099 0.99843 219

(Mo2) 0.00005 0.0004
3.33286± -0.094± -0.023 0.036 0.99941
0.00004 0.001 ±0.001

Spin waves in permalloy thin films

2.185
Sample1_30min@1000 Oe
3/2
Best fit using Bloch's T
2.180
M (10 emu)

2.175
-4

2.170

2.165

2.160
15 20 25 30

T(K)
M (T) between 10 and 30 K (<< Tc ) fit quite well to Bloch‟s T3/2 law giving spin-wave
stiffness constant D = 157 and 195 meV Å2 for the thickest films of Mo ~13.5 % and
10.9 %.
Change is still 1 % but the absolute change is only 2x10-6 emu as against 2x10-2 emu for
the bulk. The sample is just push-fitted into the straw with no packing whatsoever.

16
 Superparamagnetism
Superparamagnetism
Samples are single layers of Ni nanoparticles with non-conducting Al2O3 on
both sides deposited on both Si and Sapphire substrates using PLD technique.
M(H,T) measured using Quantum Design MPMS (SQUID magnetometer).
Diamagnetic contribution of substrate subtracted (typically χ = - 2 x 10-4 emu/tesla).

-6
4.0x10

-6
3.5x10

-6
3.0x10
M (T) at H = 200 Oe for 6 nm Ni
-6
2.5x10
sample. Langevin/Brillouin function
M (emu)

-6
2.0x10 M = (cothx-1/x) fits well with μ =
-6
1.5x10 2700 μB where x = μH/kBT.
-6
1.0x10

-7
5.0x10

0.0
0 50 100 150 200 250 300
T (K) 82

17