a.

error analysis

INTRODUCTION
Laboratory investigations involve taking measurements of physical quantities, and the-process of
taking any measurements, no matter how careful one is, always involves some uncertainty or
experimental error. There are cases that two persons taking independent measurement on the same
quantity may get different results. In such a case, the following questions arise:
1. Whose data are better? How does one know which set of data is more accurate?
2. How do we determine the degree of deviation of the data points from the accepted value of
the physical quantity?
The objective of this artide is to guide and teach us on how to handle errors and analyze them.

I. RANDOM AND SYSTEMATIC ERRORS }

The precision of a result is an indication of its repeatability, it does not imply anything about its
correctness. Accuracy, on the other hand, tells us how close our result is to the accepted value of the
physical quantity. '
Random errors are those errors that are beyond the control of the person performing the
experiment. These result from unknown and unpredictable variations in experimental situations. The
following are examples of random errors:
1. Unpredictable fluctuations in temperature.
2. Mechanical vibrations of experimental set-ups. "
3. Unbiased estimates of measurement readings by the observer.
The effect of random errors can be reduced and minimized by improving experimental
techniques and taking more trials in the measurements so that erroneous readings become
statistically insignificant.

A systematic error .is one that is inherent in the method of measurement of the apparatus, it
makes the result incorrect. Detection and elimination of non-random errors usually require an
understanding of the particular instruments and methods involved. There are therefore, no universal
rules for dealing with all possible sources of error. On the other hand, you can learn to recognize a
few common categories of error, some of which are summarized in the examples below:
1. Alteration of the measured quantity by the measurement itself
2. Asymmetries in the apparatus
3. Parallax error
4. Genuine variations in the measured quantity

National Institute of Physics, University of the Philippines- Diliman 71

 Physics 71.1 Appendix: A. Error Analysis

5. Zero or index errors

6. Drift
A common problem in many measurements is drift of the zero reading. This often occurs with very
sensitive instruments that are subject to environmental influences. The usual technique for dealing with
the problem is to take sufficient repeated zero readings to construct a graph of the zero reading as a
function of time.
7. Calibration
Corrected reading = (calibration factor) * (actual reading-zero reading)
Example. Measurement of a voltage
Zero reading -0.3 V
Calibration factor 0.98
Instrument reading 25.1 V
Corrected reading [25.1-(-0.3)]*0.98 V = 24.9 V

11. SI.GNIFICANT FIGURES

In the measurement of physical quantities, the numbers obtained are never exact. The accuracy is
always limited to the degree of refinement of the measuring device and the skill of the observer. There
will always be a point at which the reading becomes uncertain even though the best measuring device is
used and the observer has the highest skill.
Most physical measurements involve the reading of some scale. The fineness of graduation of the
scale is limited and the width of the lines marking the boundaries is by no means zero. In every case
therefore, the final digit of the reading must be estimated. If, for example, we measure the length of the
block of wood above, we would read the length as 1.24 cm where the digits 1 and 2 are exact while 4 is
an estimate. Digits beyond this first estimated digit can no ranger be obtained. Thus, we consider that
for this reading, we have three significant figures (the last digit is significant even though it is doubtful).
Significant figures (SF) are therefore those figures in a number which are reasonably known with
certainty plus the first digit that is an estimated/uncertain figure.

:111~:::II:~:§~II~:IJ,§'III::IJg~§
::::::::::::~:.:::::~::\::::
:::::::::::::::::::::::::::::::::::~::::::;;::~/::::::::::~:::::::::::::::::r::\:
»: : :r:::::::::::.:::::;:~) .:::.:::::::::::::::::::::::::::::::::::::::
'i. One doubtful figure. Retain only one doubtful figure in your reading.
All non-zero digits are significant
Example. 69.56 (4 SF)
2. In the presence of zero
a. All zeros between two non-zero digits are significant.
Example. 101.003 (6 SF)

72 National Institute of Physics, University of the Philippines- Diliman

Physics 71.1 Appendix: A. Error Analysis

b. If a decimal point exists, all zeros to the right of the last non-zero decimal digit are significant. All
left-most zeros are not significant.
Example. 235.00 (5 SF)
0.050 (2 SF)
c. If no decimal point is indicated but a bar sign is placed on one of the zeroes, then all the right-most
zeros up to the bar sign are significant.

Example. 235000 -+ (6 SF)

3. Reading digital devices and constants. If a digital voltmeter shows that a potential difference is 15.4 V,
it does not mean that the value is 15.400 exactly. It means that the reading lies between 15.35 and 15.45
V so that the value has a magnitude of 0.05 V.

Addition / Subtraction
Add first the digits then round off the final answer corresponding to the digit with the least decimal
place.
Example. 27.36
+ 106.2
4.785
138.345 13803 final answer
The digit with the least decimal place is 106.2 (one decimal place). Therefore the final answer is
138.3.
Multiplication / Division
Final answer should follow the number of significant figures of the least accurate factor (LAF) -
the number with the least number of significant figures.
Example. (11.11) * (4.43) * (7.7) = 378.97321 -+ 380 final answer
..1,

LAF(2 SF)

~Bmil§::I:II:BQmll~I§:::III:::IJJMI:§!§:::H:~:::::::::n::.::::~H::~:::~:::::\r.::::::::.::~::::~::·~:):::::~~:.:.:::;.::::.::
...:::"....;:;::;.;:;;:;:::::~:
...
::.:;=:::: ... ::~.::.

If the digit to be dropped" off is greater than 5, increase the last digit (if less than 5, then retain) .
Example. Round off to 2 decimal places.
3.286 -+ 3.29
4.431 -+ 4.43

National Institute of Physics, University of the Philippines" Diliman 73

Physics 71.1 Appendix: A. Error Analysis

If the digit to be dropped off is exactly 5 (with no non-zero numbers following it), then retain the
last digit if it is less than 5 (if it is greater than or equal to 5, increase the last digit by 1).
Example. Round off to 2 decimal places.
3.455 ~ 3.46 (increase by 1)
3.435 ~ 3.43 (retain)

3.485 ~ 3.49 (increase by 1)

3.4451 ~ 3.45 (increase by 1)

-,
Ill. ERROR PROPAGATION
A. SINGLE MEASUREMENT
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 uncertainty in the measured diameter is not necessarily the
uncertainty in the calculated are or volume. We have to find out how the uncertainty in the diameter
affects the uncertainty of the computed area -the very reason why we need error propagation. .

Consider a measured quantity xo, with an uncertainty Ox and consider a computed result z to be
some function of the variable x. Let z = f(x).
The function f enables us to calculate the required value Zofrom a measured value Xo' Moreover,
the possibility that x can range from (xo - Ox) to (xo + Ox) implies a range of possible values of z that
range from (z, - Oz) to (z, + Oz), where Oz is the value of the absolute uncertainty in z. We now wish
to calculate the value if Oz. The situation is illustrated graphically in the figure below, in which, for a
given f(x), we can see how the measured value Xogives rise to the computed zo, and how the range
+OX about Xoproduces the corresponding range ±Oz about zoo

z=f(x)

8z
Zo~--~------------------------~

Figure 1. Propagation of uncertainty from one variable to another.

74 National Institute of Physics, University of the Philippines- Diliman

Physics 71.1 Appendix: A. Error Analysis

Example. Find the uncertainty Oz propagated by a measured uncertainty in Ox given the function
z = X2.

If x can range between (xo - Ox) to (Xo + Ox), then z can range between (z, - oz) to (z, + oz),
where .

Zo + Oz = (xo + OX)2= X02± 2 XoOx+ (OX)2

We can ignore (OX)2,since Ox is assumed to be small in comparison with Xo,and equate 20 to .X02,
giving for the value of Oz

Oz = 2xoOx where Zo = Xo2

This can be expressed more conveniently in terms of the relative uncertainty oz/zo
,
8z 2xo8x 8x
-=--2-=2-
Zo xo Xo

Thus, the relative uncertainty of the computed result is twice that of the intitial measurement.
Although it is helpful to bear in mind the nature of propagated uncertainty, as illustrated by the use
of finite differences, considerable simplification of the formulation can be achieved using differential
calculus.

The finite differences oz and Ox that were used in the preceding section could be regarded as
components of the derivative dz/ dx. We can therefore obtain our value of Oz by first using standard
techniques to obtain dz/ dx in the form

: = d(~~X>,)and then writing 8z = Id(~~x))18X

Examples.
Powers
Consider the function z = x", Differentiating and replacing the derivative by finite differences
.
gtves us -dz = nx n -I . Th e a b so Iute uncertainty
. '.1S gtven
. b y oz,
~ W
hil1 et hi'ere atrve 1S
. -8z = n-.Sx
dx z x

Trigonometric functions

z .
= smx I I
: 8z = cos x 8x and -8z = Icos
-.- xl8x
z sm x

z = cosx: I· I
8z = sin xlox an d -8z = --8x
sin x I I
z cosx

Consider a function z which is dependent on two variables x and y, z = f(x,y). The appropriate
quantity for calculating Oz is the total differential dz. This is given by

Of Of
dz=-dx+-dy
Ox Oy

National Institute of Physics, University of the Philippines- Diliman 75

 I "

Physics 71.1 Appendix: A. Error Analysis

We treat this as a finite difference oz that can be calculated from the uncertainties Ox and Oy.
Thus

Examples.
Addition / Subtraction

z=x+y: oz = IOx I + Ioy I and 8z I~II~

~=~+Y(~I
z=x-y: Oz = I Ox I + Ioy I and

For subtraction we choose the maximum possible uncertainty so we have oz = Ox + oy.

Multiplication / Division

Z = xy: Oz = yox + xoy and 8z ~18:1+1;1

Z

z=x/y: and 8: =18:1+1;1

For division we choose the maximum possible uncertainty so we have oz = Ox + oy.
Exercises.
Derive the uncertainty in z of z = X·ybassuming an initial uncertainty of Ox and oy.
. 2
Try it for z =xyan ,-=-+-.x y
dIll
Z

B. MULTIPLE MEASUREMENT
In the first part of this section, we discussed the error propagation of a single reading, the
uncertainties of which can be estimated using personal judgement. In these, supposing 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.
In this case, it is no longer possible to say that "I am certain that it lies within the interval" and so the
edges of the interval begin to blur and we can only interpret the intervals in terms of probability
rather than uncertainty.

1. Central location
There are many ways of describing a set of readings. Among the most commonly used are the
distribution table and the histogram.

76 National Institute of Physics, University of the Philippines- Diliman

Physics 71.1 Appendix: A. Error Analysis

In a distribution table, the values are tabulated along with the frequency of their occurrence. If, for
example, we have a box of resistors and we sorted according to resistance, the distribution table will
look something like
Value (Q) Number of readings
327 5
328 9
329 7
.:»

330 3
331- 1
TOTAL 25
The corresponding histogram is
a

No. of readings

r -I ~
•..
327 328 329 330 331
Resistance

It is difficult to quantitatively handle a histogram because you cannot plug it into computations. If a
simple representation can be devised that could be used as a substitute for the figure then it will simplify
reporting of results.
The most common measure of location is the arithmetic mean. If the values of the n readings are
denoted by Xi = 1,2,3 ... n, then their mean is
n

- i=1
L>i
X=--
n

2. Spread
In order to answer questions like "How reliable is it to use a single number as a substitute for the
whole distribution?" or "Up to what extent can the central value represent the whole?" we introduce the
concept of breadth or spread. One measure of the spread of a set of n values about their mean value x
is the quantity s, which is defined as the root mean square (rms) deviation of the readings from x.
Translating this definition into symbols gives the formula
n n
L(x i -x)2 L Xi
2

S=
i=1 i=1 X2
n n

National Institute of Physics, University of the Philippines- Diliman 77

Physics 71.1 Appendix: A. Error Analysis

3. Sampling and Precision

So far we have discussed only the description of a particular set of values. In experimental work we
often take a set of repeated readings which we can regard as a sample of all the readings which could
possibly be taken. We call "all possible readings" the population of readings.' We imagine that this
a
population has mean value. Let's call it X, which is what we really wish to know.
The question to be answered now is: "What can we learn about the population by taking a sample
from it?" Statistical theory indicates that if the sample is a random selection of readings then we can use
the sample mean (x) as an estimate of the population mean eX) and we can use s as an estimate of the
population standard deviation S.
The next problem is this: "How good are those estimates?" .That question can be recast as: "By how
much would the estimates vary if we took more random samples?" An obvious (but brutal) way of
answering that would be to actually take more samples. We could, for example, get 20 sets of readings,
calculate 20 mean values and calculate the standard deviation of those 20 means. We could then use
that standard deviation of the mean values as a measure of how good the sampling procedure is for
estimating S.
Statistical theory saves us the trouble of actually taking more samples! The formula for the standard
deviation of the population or the standard error of the mean (SEIvf)is

s
s=--= n(n-l)
In-l
where s is the sample standard deviation and n is the number of samples. So we can actually write our
reading as a representative of the population X =X ± s.
An interpretation of the SEM is that there is 68% chance that the population mean X actually lies
within the range x ± S. The meaning of such an uncertainty estimate is therefore quite different from
the meaning implied when we quote limits. If you wish to estimate limits based on S, note that there is
about 68% probability that X lies in the range x±S, 95% probability that X lies in the
range x±2S, and about 99.7% probability that it lies in the range x±3S. These estimates require that n
is large enough to give a representative sample.

Example. The resistance of a resistor is tested using a voltmeter. The derived readings are given below.
Find the resistor's best value.

3270 3280 3270 3290 3270

328 327 328 330 329
329 329 329 329 328
327 328 329 329 330
331 328 330 328 328

78 National Institute of Physics, University of the Philippines- Diliman

 ----,-----~~~~~~~---~----~
Physics 71.1 Appendix: A. Error Analysis

Calculations lead to
Mean value x = 328.4 n
Stdev of sample s = 1.0 n
Mean of population X = 328.4 n
Stdev of population S = 0.2 n
The best value is then R = X ± S = 328.4 ± 0.2 n

::::::::':::::::::::::::::::::::::,:::.::,:,::::::::::::::,:::::::::::::::,:,:!:::::,::::::',:':::!:!:::::::!:::::::::::::'::::::!:::!,:::::,:,:::!:,::'::::::::!':':'::!::'!':':!:::':':'::!::::'::",::!:,!:::::::::!:!:!:!:!!!:!:!!:!!!':!:::!!!!::!!:!:!!!:!:'§i@.·::lmm'EJ11$:::II'IMg~~lflw.!:'IE!I.$,!
Analogous to the general method of propagating uncertainties, the population standard deviation can
also be propagated using the general form .Ueral formula

8 (:)28x2+(:) 8
z

If z is a function of more variables, the equation is extended by adding similar terms.

Examples.

Addition / Subtraction
z = x +y: s, = ~r-s-~
-+-S-;

z = x - y: Sz = ~S~ +S;

Multiplication / Division
z xy := S, = ~'y-2S-~-+-X-2-S;-

S2 X2S2
z=x/y:
•
S
Z
= -2..+ __ Y
y2 y4
The specific value of
Szat any particular values of x and y, say Xo and Yo, can be obtained by substituting
Xo and Yo in this expression.

Power
z X4: =
Examples.

Combining standard error estimates

Data: y = xl - x2, where xl = 2.63 + 0.02 mm and x2 = 1.61 + 0.02 mm
Estimate of SEM:
Sy =~SXI2 +Sx/ =J2 2
+22 X 10-2 mm=0.03mm
y = 1.02 + 0.03 mm

Constants in the equation

. .

Assume that the uncertainty is an SEM.

Calculate r = d/2 where d = 95.24 + 0.04 mm. The factor 1/2 has no uncertainty and no SEM.

National Institute of Physics, University of the Philippines- Diliman 79

I~--------------------------------------------------------------~~~~~~~
Physics 71.1 Appendix: A. Error Analysis

Estimate of SEM:

S =S~ =O.02mm

r = 47.62 + 0.02 mm

Standard error estimates

Data: v = (a-b) / t
a = 5.63 + 0.02 mm, b = 1.48 + 0.02 mm and t = 5.16 + 0.05 s
Using the general formula

s = v

s = (_1_)2(0.02)2 +( __ 1_)2 (0.02)2+(_5.63-l.49J2 (0.05)2 = 0.01rnIs

v 5.16 5.16 (5.16)2

And the best value of the velocity is v = 0.804 ± 0.01 m/ s

C. MIXED MEASUREMENTS
What if, in our calculations we encounter variables that have. single reading uncertainties and
multiple reading uncertainties? Clearly, the two uncertainties are different and cannot readily be
combined because one has an uncertainty Ozand the other S,..
One approach is to try to estimate, for each contributing uncertainty, something that might
correspond to a standard error. (Recall that the SEM S, gives an estimate of the range within which the
result is likely to Iie within a probability of 68% or about 2/3). To do this we can guess that for each
quantity (say z + oz) which contributes to the final uncertainty, that there are 2 chances in 2 that the
correct value lies in the range z ± t 8z. We then treat each contributing uncertainty estimate (t 8z) as
though it were a standard error Sz = t 8z. Having converted Oz into S2'we can then propagate errors by
the preceding method.
Example. The area of a rectangular plate is measured by taking 10 readings of one of its two sides, x,
giving a best value x = 15.0 + 0.1 mm. Due to time constraints, the second side is measured only once
and approximated with an uncertainty with a value y =·10.3 + 0.2 mm. Find the best value of the area.
In this problem, we have two component uncertainties: one SEM (Sx=O..tmm) and one uncertainty
(Oy=0.2 mm). First we convert oy into S, and then we combine all uncertainties as though they were
standard errors.
Sy = 2/30y = 0.1

80 National Institute of Physics, University of the Philippines-Dilimen

Physics 71.1 Appendix: A. Error Analysis

Because the area is a product of the lengths (A = xy), the estimated uncertainty of the area is given
by

SArea= ~y2S; + x2S; = ~(1O.3)2(0.l? + (15.0)2(0.1)2 = 2 mm

So the area is quoted as A = 154 + 2 mm.

V.SUMMARY
The propagation of uncertainties is summarized in the table below for single reading and multiple
readings.
o eration
Addition/Subtraction
z=x+y Oz = loxl + loyl
z = x-
Multiplication/Division
s, = ~y2S; + x2S;
z =x*y
z =x / y S2 X2S2
S = _x + __ y
z y2 s'
Power
z = Xli S z = ax (a-l)s x

Exponential
z = eX
Logarithmic Ox
z = lnx oz=g

If the quantity you are calculating contains variables with single and multiple reading uncertainties;
convert the single uncertainty to a standard error by 2/3, i.e. S, = 20z/3 and then propagate by method
of multiple readings.

ational Institute of Physics, University of the Philippines- Dilirnan 81