Error Analysis in Experimental Physical Science

Mini-Version
by David Harrison and Jason Harlow
Last updated September 14, 2014 by Jason Harlow.
Original version written by David M. Harrison, Department of Physics, University of Toronto in January 2001. The regular version of this document is
available at: http://www.upscale.utoronto.ca/PVB/Harrison/ErrorAnalysis/index.html

1. Introduction
Almost every time you make a measurement, the result will not be an exact number, but it will be a range
of possible values. The range of values associated with a measurement is described by the uncertainty, or
error. Since all of science depends on measurements, it is important to understand errors and get used to
using them.
[An exception might be counting apples in a bowl. Three is an exact number of apples that does
not have an error. However, if you estimate the number of apples in the back of a large apple
truck and find there are 1600, this number does have an error!]
Note that in this course the word error does not mean mistake! It is better thought of as an
uncertainty, meaning it quantifies how certain (or uncertain) you are about the value you measured.
Basically, the error is the number that appears to the right of the symbol.
We should emphasize right now that a correct experiment is one that has been correctly performed.
You do not determine the error in an experimentally measured quantity by comparing it to some number
found in a book or web page! You determine each error as part of your experiment; the value and the
error are two numbers that are measured together, and you report them both, independently of what
anyone else may have reported in the past.
Also, although we will be exploring mathematical and statistical procedures that are used to determine the
error in an experimentally measured quantity, as you will see these are often just rules of thumb and
sometimes a good experimentalist uses his or her intuition and common sense to simply guess what the
error is. That is okay sometimes!
Example 1: Drawing a Histogram of repeated measurements
Imagine you have a cart on a track with a fan attached to it which causes it to
accelerate along the track. You release the cart from rest and then use a digital
stopwatch to measure the time it takes the cart to travel 1.50 m. You estimate
the error in the distance you measured to be 1 cm, so actually the distance is
1.50 0.01 m. You then repeat the time measurements for a total of 30 trials.
The time measurements are shown in the Table 1.
Trial #
1
2

Time (s)
5.55
5.50

Error Analysis Document for PHY131/132 Page 2 of 10

3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

5.56
5.51
5.49
5.46
5.49
5.51
5.32
5.50
5.49
5.45
5.30
5.67
5.64
5.61
5.66
5.69
5.38
5.37
5.74
5.54
5.56
5.47
5.49
5.54
5.45
5.59
5.52
5.45

Table 1. Time for a cart to start from rest, accelerate at a constant rate, and travel a distance of 1.50
0.01 m. The measurement was repeated 30 times.
a. Make a histogram of the results. On the horizontal axis, plot the time from 5.0 to 6.0 seconds with
tick-marks at intervals of 0.05 seconds. On the vertical axis, plot the number of trials that fall into
each range.
b. What is the height of the histogram? [number of trials in the most populated bin]
c. What is the width of the histogram? [size of the central part of the range which contains about 2/3
of the trials.]
d. Sketch a smooth bell-shaped curve through your histogram to approximate the tops of the
histogram. What is the centre of this curve? [this approximates the mean].

Error Analysis Document for PHY131/132 Page 3 of 10

Example 1: Answers
a. See the Excel spreadsheet, first sheet, available for download along with this document for work. Here
is a plot of the histogram:

b. The height of the histogram is 9.

c. 2/3 of the 30 trials is 20. In the three bins with centres at 5.45, 5.5 and 5.55, there are a total of 5 + 9 +
5 = 19 trials, which is almost 2/3 of the total. Therefore, I would say the width of the histgram is about 3
bins, or 30.05 = 0.15 seconds.
d. The smooth curve was not done in Excel, but if I was to do it by hand on a print-out, it would probably
have its centre slightly to the right of 5.5 s, because the tail of this histogram extends further toward the
right than the left. So perhaps the mean is about 5.51 s.

Error Analysis Document for PHY131/132 Page 4 of 10

2. Normal Distribution
Each time you make a measurement, there is some probability that you will get a certain answer. A
probability distribution is a curve which describes what the probability is for various measurements. The
most important and widely used probability distribution is called the Normal Distribution. It was first
popularized by the German mathematician Carl Friedrich Gauss in the early 1800s. It is also sometimes
called the Gaussian distribution, or the bell-curve. The formula is:

(x x )2

N(x) Ae

2 2

which looks like:

The symbol A is called the maximum amplitude.

The symbol x is called the mean or average.
The symbol is called the standard deviation of the distribution. Statisticians often call the square of the
standard deviation, 2, the variance; we will not use that name. Note that is a measure of the width of
the curve: a larger means a wider curve. ( is the lower case Greek letter sigma.)
Note that 68% of the area under the curve of a Gaussian lies between the mean minus the standard
deviation and the mean plus the standard deviation. Similarly, 95% of the curve is between the mean
minus twice the standard deviation and the mean plus twice the standard deviation.

Error Analysis Document for PHY131/132 Page 5 of 10

3. Using the Gaussian

Suppose you make N measurements of a quantity x, and you expect these measurements to be normally
distributed. Each measurement, or trial, you label with a number i, where i = 1, 2, 3, etc. You do not
know what the true mean of the distribution is, and you cannot know this. However, you can estimate the
mean by adding up all the individual measurements and dividing by N:
Estimate of Mean:

1 N
xest xi
N i 1

Similarly, it is impossible to know the true standard deviation of the distribution. However, we can
estimate the standard deviation using our N measurements. The best estimate of the standard deviation is:

Estimate of Standard Deviation:

est

1 N
2
(
x

x
)
i est
N 1 i 1

The quantity N 1 is called the number of degrees of freedom. In the case of the standard deviation
estimate, it is the number of measurements minus one because you used one number from a previous
calculation (mean) in order to find the standard deviation.
Example 2: Consider the 30 measurements of time in the Table 1.
a. What is the estimated mean?
b. What is the estimated standard deviation?
c. How many of these 30 measurements fall in the range xest est ? What percentage is that?
Example 2: Answers
a. See the Excel spreadsheet, second sheet, available for download along with this document for
work. The estimated mean is found to be 5.517 s.
b. The estimated standard deviation is found to be 0.103 s.
c. 21 of the 30 measurements fall within the range 5.517 0.103 s, or 5.414 < t < 5.620 s. This is
70% of the trials.
There is roughly a 68% chance that any measurement of a sample taken at random will be within one
standard deviation of the mean. Usually the mean is what we wish to know and each individual
measurement almost certainly differs from the true value of the mean by some error. But there is a 68%
chance that any single measurement lies with one standard deviation of this true value of the mean. Thus
it is reasonable to say that:
The standard deviation is the error in each individual measurement of the sample.
The error in a quantity is usually indicated by a (delta), so the above statement may be written as xi
= . This error is often called statistical. We shall see another type of error in the next section.

Error Analysis Document for PHY131/132 Page 6 of 10

4. Reading Error
In the previous section we saw that when we repeat measurements of some quantity in which random
statistical factors lead to a spread in the values from one trial to the next, it is reasonable to set the error in
each individual measurement equal to the standard deviation of the sample. Here we discuss another error
that arises when we do a direct measurement of some quantity: the reading error.
For example, imagine you use a metric ruler to measure the length of a pencil. You line up the tip of the
eraser with 0, and the image below shows what you see over near 8 cm. The pencil appears to be about
8.25 cm long. What is the reading error?

Image source: http://www.amazingrust.com/experiments/background_knowledge/Measurement.html

To determine the reading error in this measurement we have to answer the question: what is the minimum
and maximum values that the position could have for which we will not see any difference? There is no
fixed rule that will allow us to answer this question. Instead we must use our intuition and common sense!
Could the pencil actually be as long as 8.3 cm? No, I dont think so. Its clearly shorter than that. Could
it be 8.28 cm? Maybe. And it could be as short as 8.23 cm, but, in my opinion, no shorter. So the range
is about 8.23 to 8.28 cm. A reasonable estimate of the reading error of this measurement is half the range:
0.025 cm, which, to be cautious, we might round up to 0.03 cm. Then we would state that the length of
the pencil is 8.25 0.03 cm.
For your eyes you may wish to instead associate a reading error of 0.02 cm with the position; this is also a
reasonable number. A reading error of 0.04 cm, though, is probably too pessimistic. And a reading error
much less than 0.01 cm is probably too optimistic.
We assume that the reading error indicates a spread in repeated measurements, just like the standard
deviation discussed in the previous section. So if we get a collection of objective observers together to
look at the pencil above, we expect most (ie more than 68%) of all observers will report a value between
8.22 and 8.28 cm.
Note that there is often a trade-off when assigning a reading error such as above. On the one hand we
want the error to be as small as possible, indicating a precise measurement. However we also want to
insure that measured value probably lies within errors of the "true" value of the quantity, which means we
don't want the error to be too small.
For a measurement with an instrument with a digital readout, the reading error is " one-half of the last
digit."

Error Analysis Document for PHY131/132 Page 7 of 10

We illustrate with a digital thermometer shown to the right. The phrase " onehalf of the last digit" above is the language commonly used in manufacturer's
specification sheets of their instruments. It can be slightly misleading. It does
not mean one half of the value of the last digit, which for this thermometer
photo is 0.4C. It means one-half of the power of ten represented in the last
digit. Here, the last digit represents values of a tenth of a degree, so the reading
error is 1/2 x 0.1 = 0.05C. You should write the temperature as 12.80 0.05
C.
It may be confusing to notice that we have two different specifications for the
error in a directly measured quantity: the standard deviation and the reading
error. Both are indicators of a spread in the values of repeated measurements.
They both are describing the precision of the measurement. So you may ask:
What is the error in the quantity?
The answer is both. But fortunately it is almost always the case that one of the two is much larger than the
other, and in this case we choose the larger to be the error.
For example, if every time you measure something you always get the same numerical answer, this
indicates that the reading error is dominant.
However, if every time you measure something you get different answers which differ more than the
reading error you might estimate, then the standard deviation is dominant. In these cases you can usually
ignore the reading error. For example, consider the rolling fan-cart we looked at in Table 1. The reading
error of the digital stopwatch is 0.005 seconds, which is negligible compared to the standard deviation,
which we found to be 0.1 seconds. Thus here the error in each measurement is the standard deviation.
Often just thinking for a moment in advance about a measurement will tell you whether repeating it will
be likely to show a spread of values not accounted for by the reading error. This in turn will tell you
whether you need to bother repeating it at all. If you don't know in advance whether or not you need to
repeat a measurement you can usually find out by repeating it three or four times.

Error Analysis Document for PHY131/132 Page 8 of 10

5. Significant Figures
Most introductory Physics textbooks (including the one chosen for this course) include a little section on
significant figures. It provides some rules for how many significant figures to include if your number
does not include an error. However, in experimental science every number you measure or compute
should include an error! Therefore, the rules in your textbook are useless for any actual numbers
measured in the Practicals. In this section, we will teach you the rules for significant figures when there
is an error associated with every value.
Lets consider again the 30 data points in Table 1. The estimated standard deviation is numerically equal
to 0.102933971 seconds, which is larger than the reading error for these measurements. (By
"numerically" we mean that is what the calculator read when we computed the standard deviation, or
what Excel produced when we used the STDEV function.) Since the estimated standard deviation is
larger than the reading error, it will be the error in the value of each of the data points.
Consider one of these data points, say, the 5th trial, for which we measured 5.49 seconds. If you take the
estimated standard deviation to be 0.102933971, then the data point has a value of 5.49 0.102933971.
What this means is that there is about a 68% chance that the true value is somewhere between
5.387066029 and 5.592933971 seconds.
Keep in mind that all of these are estimates, as we have only 30 data points, not infinity. Also, 68% is a
somewhat arbitrary probability, which we have chosen to represent most of the time. Clearly we are
using WAY too many significant figures here! It would be just as instructive to say that there is about a
68% chance that the true value is somewhere between 5.4 and 5.6 seconds. Or, you could say the
measurement is 5.5 0.1 s. In fact it is not only more concise to report this, but it is more honest.
This example illustrates two general rules for significant figures used in experimental sciences:

1. Errors should be specified to one or two significant figures.

2. The most precise column in the number for the error should also be the most
precise column in the number for the value.
So if the error is specified to the 1/100th column, the quantity itself should also be specified to the
1/100th column.
Example 3. Express the following quantities to the correct number of significant figures:
a. 25.052 1.502
b. 92 3.14159
c. 0.0530854 0.012194
Example 3: Answers
a. 25.1 1.5 (or 25 2 is also acceptable)
b. 92 3
c. 0.053 0.012 (or 0.05 0.01 is also acceptable)

Error Analysis Document for PHY131/132 Page 9 of 10

6. Propagation of Errors of Precision

When you have two or more quantities with known errors you may sometimes want to combine them to
compute a derived number. If your errors come from standard deviation and are not reading errors, it is
usually best to compute the derived number several times and use compute its standard deviation directly.
However, if some reading errors are larger than the standard deviation, you can use the rules of Error
Propagation to infer the error in the derived quantity.
We assume that the two directly measured quantities are x and y, with errors x and y respectively. The
measurements x and y must be independent of each other.
The fractional error is the value of the error divided by the value of the quantity: x / x. The fractional
error multiplied by 100 is the percentage error. Everything is this section assumes that the error is small
compared to the value itself, i.e. that the fractional error is much less than one.
For many situations, we can find the error in the result z using three simple rules:
Rule # 1 (sum or difference rule):
If
z=x+y
or
z=xy
then

z x 2 y 2
Rule #2 (product or division rule):
If
z = xy
or
z = x/y
then

z
x y

z
x y
Note if x is an exact number, so that x = 0, then z = x y.
2

Rule #3 (exponent rule):

If
z = xn
then

z
x
n
z
x

Example 4. As with Example 1, you have a cart on a track with a fan attached to it, which causes it to
accelerate along the track. You release the cart from rest and then use a digital stopwatch to measure the
time it takes the cart to travel 1.50 0.01 m. You then repeat the time measurements for a total of 30

Error Analysis Document for PHY131/132 Page 10 of 10

trials. The time measurements are shown in the Table 1. You model the distance as a function of time
with the kinematic equation d = a t2. From your measurements of d and t you wish to derive a, which
is a = 2d/t2.
a. Compute individual estimates of acceleration for each time measurement, using d = 1.50 exactly.
Estimate the mean and standard deviation of all these values of a.
b. As discussed in Section 5, the best way to report the measurement for the 5th trial is: t = 5.5 0.1
s. Combine this with d = 1.50 0.01 m and propagate errors to compute acceleration, a.
Example 4: Answers
a. The mean of the acceleration values is 0.0987 m/s2, and the standard deviation is 0.0037 m/s2.
b. The equation is a = 2d/t2, where 2 is an exact number (with no error). To compute the error, first
use Rule#3 to find the error in z = t2: z = 2z(t/t) = 2(5.52)(0.1/5.5) = 1.1. Then use Rule#2 to
d 2 z 2
0.012 1.1 2
get the error in y = d/t2 = d/z: y y 0.0496

= 0.0018.
d z
1.5 30.25
Then use Rule#2 again with the exact number 2 as a multiplier, which simply increases the error
by 2: a = 0.0037 m/s2. It should not be surprising to you that the error in one measurement of
acceleration should be equal to the standard deviation we computed in part a.

7. The Error in the Mean

We have seen that when the data have errors of precision we may only estimate the value of the mean.
We are now ready to find the error in this estimate of the mean.
Recall that to calculate the estimated mean we use:

1 N
xest xi
N i 1
Each individual measurement xi has the same error, x, which is usually the estimated standard deviation.
To calculate the error in the sum of all the xi, we use Error Propagation Rule #1 to find the error in the
sum as x 2 x 2 x 2 Nx 2 N x . We then use Rule #2 to find the error in the sum
divided by the exact number N, which gives:

xest

x
N

Example 5.
a. What is the mean and the error in the mean for the 30 time measurements in Table 1? [Note that
the error in the mean is not the same as the standard deviation!]
b. What is the mean and the error in the mean for the acceleration of the cart?
Example 5: Answers
a. 5.517 0.019 s.
b. 9.867 0.067 cm/s2.