Brown University Physics 0030

Physics Department Analysis Of Experimental Uncertainties

131104 1


ANALYSIS OF EXPERIMENTAL UNCERTAINTIES


All experimental values are subject to uncertainties which arise from inherent limitations in the
instruments used. It is not possible for us to construct a perfect instrument or a perfect observer
even in the most carefully designed experiments random effects influence results, so that many
repetitions of any measurement will result in a distribution of final results that peaks at the most
probable value.

We discuss here mainly such random uncertainties.

The best way to obtain a good final number is to make a number of independent measurements,
as many as feasible. The best number we can then give is the arithmetical average of all the
measurements. For N measurements of a quantity L, this arithmetic
average, bar L , is obtained from all the individual measurements
i
L as follows:
) ... (
1
3 2 1 N
L L L L
N
L + + + + =

=
=
N
i
i
L
N
1
1
(1)

But no experimental measurement is complete without an estimate of the uncertainty, which is
reflected in the spread among the individual measurements. This spread is measured
quantitatively by the Standard Deviation of the mean, which is calculated (approximately) from
the root mean square deviation RMS by the following equation:

=
N
i
i
L L
N
RMS
1
2
) (
1
1
(2)

The RMS value is the uncertainty we associate with each of the repeated measurements. Using
the results of the following section, we can then see that the uncertainty in the average of N
repeated measurements is:
. / N RMS L = A (3)

The best value of repeated measurements of some quantity L is then L L A

Sometimes it is not feasible to do multiple measurements of a quantity; in such a case the
experimenter has to use his or her judgment about the uncertainty, based on the instruments used.
For example, if you are reading a meter stick, you should judge whether the uncertainty in your
reading is 1 . 0 or 2 . 0 of the smallest division on the meter scale. (Never assume it is better
than . 1 . 0 )

Every experiment measurement and result must be given in the form: . X X A
Brown University Physics 0030
Physics Department Analysis Of Experimental Uncertainties

131104 2

Propagation of Errors

Frequently, measured values are used to calculate further results. These results will of course be
affected by the uncertainties (often called errors) in the measured quantities that enter into their
derivation. In fact, we can calculate directly the uncertainty to associate with a derived result
from the uncertainties in the measurements involved in the calculation. We give here some basic
rules to use in propagating the uncertainties of measurements into the uncertainty of a result.
1


First, if k is a constant, then if kx z = the uncertainty in z is x k z A = A , where x A is the
uncertainty in x .

Now, consider two measured quantities X X A and Y Y A . The following rules hold for
calculating the uncertainty in , Y X + , Y X , XY and Y X / :


Z =X + Y or Z= X - Y


AZ=
2 2
) ( ) ( Y X A + A

Z =XY or Z = X/Y

Z
Z A
=
2 2
|
.
|

\
| A
+
|
.
|

\
| A
Y
Y
X
X



AX, AY, AZ are absolute uncertainties and ,
X
X A
,
Y
Y A
A
A A
are relative uncertainties.

These rules can be extended to any number of factors entering the calculation -- there can be
three or more, not just the two used in the examples above.

As an example, we can find the uncertainty L A in the average of N repeated measurements of the
same quantity L (see Eqn.1) each having uncertainty
L A
(perhaps found from Eqn. 2) as given
by:
) (
1 1
)
1
(
2
2 `
2
1 `
1 1
L
N
N
L L
N
L
N
L
N
L
N
i
i
N
i
i
A
|
|
.
|

\
|
= + A + A
|
.
|

\
|
= |
.
|

\
|
A
|
.
|

\
|
= A = A

= =

where L L L A = = A = A
2 1


1
These rules may be derived from the following more general relation:
Consider the variable z as a function f of the variables , ,
2 1
x x or ) , , (
2 1
x x f z = .
Then the uncertainty z A can be found form the uncertainties , ,
2 1
x x A A as
+ A
|
|
.
|

\
|
c
c
+ A
|
|
.
|

\
|
c
c
= A
2
2
2
2
2
1
2
1
) ( ) ( x
x
f
x
x
f
z

Brown University Physics 0030
Physics Department Analysis Of Experimental Uncertainties

131104 3

so, L
N
L A = A
1

In addition to the random uncertainties or errors discussed above, systematic errors can arise
from non-random sources of error. An example of a systematic error is the mis-calibration of
your measuring instrument. While random errors can be reduced by making additional
measurements, systematic errors are independent of the number of measurements. If you see any
evidence of such problems in your data, you should mention it in your lab report and try to
suggest specific causes.

The uncertainties or experimental errors discussed above do not include mistakes in reading or
setting instruments. Also note that the experimental error is derived solely from the precision of
the measurements you make and the limitations of the experiment itself. It is not the same as the
discrepancies that might exist between your measured value and a known or accepted value.

Although we have presented the mathematical rules and formulas for error analysis it is
important to have an intuitive understanding of the concepts involved. In this section we explain
the difference between accuracy and precision and give some examples of error analysis. In an
experiment, there is a difference between accuracy and precision. Accuracy means how close a
measurement is to the known (true) value. Precision means how close your data points are to
each other. It is possible for data to be accurate and not precise and vice-versa. For example:

The average in (a) is accurate (all measurements average to the bulls eye) but not precise,
whereas in (b) the average is precise but not very accurate.

Usually, if an experiment is accurate but not precise, this is due to random errors such as human
error or imprecise equipment. If an experiment is precise but not very accurate, there is
systematic error. A systematic error is an error with the equipment or an error in technique that is
repeated consistently.



(a) (b)
Brown University Physics 0030
Physics Department Analysis Of Experimental Uncertainties

131104 4




Average and Standard Deviation

An average tells you what value your data tends to. The average over N trials is given by:

L

=
=
N
i
i
L
N
1
1


The standard deviation is a measure of the error in your data. Whenever you measure a quantity
that tends to an average, there will be some values greater than the average and some values less
than the average. If an infinite number of trials are done, then the following distribution is
obtained:

The x axis shows the values. The y axis marks the occurrences of a particular value.

The maximum of the curve is the average. The width of this curve is measured by the standard
deviation. A large standard deviation means a wide curve etc.

The standard deviation measures how precise your values are. Recall that the standard deviation
is given by:

Standard Deviation = . / N RMS L = A

If you were to take another measurement, then the probability
2
that the value would follow
within 1 standard deviation is 68% and 98% that it would wall within 2 standard deviations.

2
Random measurement errors generally follow a Gaussian probability distribution, and the integral of such a
distribution to +/- 1std. deviation contains 68% of the area and to +/- 2 std. deviations, 98% of the area.
L
Brown University Physics 0030
Physics Department Analysis Of Experimental Uncertainties

131104 5

Thus a small standard deviation means a very precise experiment. This does not mean that the
results are accurate. This means that there were mainly systematic errors.

When presenting the value of something that was measured, one often presents the average value
plus or minus the standard deviation ( L L A ) where the standard deviation has the same units
as the average. This means that you are 68% sure that the data point lies within the interval (
L L A + ) and ( L L A ).

Is it possible to have a small percent error in a value and rotten precision? YES! This could be
indicative of a large amount of random error ( and a lot of good luck ). Is it possible to have a
small standard deviation and low accuracy? YES! This could be indicative of a systematic error.

For example if one were to measure the acceleration of gravity and measured the value to be
2
/ 1 . 0 2 . 9 s m , this would be a very precise answer but it does not mean the measurement gave a
good result. Although the percent error is low, the true value of 9.8
2
/ s m is 6 standard
deviations away. The probability of using the experimental apparatus to measure the correct
value of 9.8
2
/ s m is near 0. This is indicative of large systematic error. Likewise if one were to
measure that the acceleration due to gravity is
2
/ 7 7 s m , the percent error could be large but
the true value for the acceleration due to gravity of 9.8
2
/ s m is within the standard deviation.
This would be indicative of a lot of random error to make the standard deviation so large.

When writing about your results in your lab report it is important to keep these distinctions clear
in your mind. Also, understanding these distinctions can make it easier to judge what sources of
error are reasonable.