Download Errors and Error Estimation and more Lecture notes Physics in PDF only on Docsity!
First Year PhYsics LaboratorY ManuaL xvii
Errors and Error Estimation
Errors, precision and accuracy: why study them?
People in scientific and technological professions are regularly required to give quantitative
answers. How long? How heavy? How loud? What force? What field? Their (and your)
answers to such questions should include the value and an error. Measurements, values
plotted on graphs, values determined from calculations: all should tell us how confident
you are in the value. In the Physics Laboratories, you will acquire skills in analysing and
determining errors. These skills will become automatic and will be valuable to you in
almost any career related to science and technology. Error analysis is quite a sophisticated
science. In the First Year Laboratory, we shall introduce only relatively simple techniques,
but we expect you to use them in virtually all measurements and analysis.
What are errors?
Errors are a Measure oF the Lack oF certaintY in a vaLue.
Example: The width of a piece of A4 paper is 210.0 ± 0.5 mm. I measured it with a
ruler
1 divided in units of 1 mm and, taking care with measurements, I estimate that I
can determine lengths to about half a division, including the alignments at both ends.
Here the error reflects the limited resolution of the measuring device.
Example: An electronic balance is used to measure the weight of drops falling from an
outlet. The balance measures accurately to 0.1 mg, but different drops have weights
varying by much more than this. Most of the drops weigh between 132 and 139 mg. In
this case we could write that the mass of a drop is (136 ± 4) mg. Here the error reflects
the variation in the population or fluctuation in the value being measured.
Error has a technical meaning, which is not the same as the common use. If I say that
the width of a sheet of A4 is 210 cm, that is a mistake or blunder, not an error in the
scientific sense. Mistakes, such as reading the wrong value, pressing the wrong buttons on
a calculator, or using the wrong formula, will give an answer that is wrong. Error estimates
cannot account for blunders.
xviii First Year PhYsics LaboratorY ManuaL
Learning about errors in the lab
The School of Physics First Year Teaching Laboratories are intended to be places of learning
through supervised, self-directed experimentation. The demonstrators are there to help you
learn. Assessment is secondary to learning. Therefore, do not be afraid of making a poor
decision—it’s a good way to learn. If you do, then your demonstrator will assist you in
making a better decision.
Please avoid asking your demonstrator open ended questions like “How should I estimate
the error”. That is a question you are being asked. Instead, try to ask questions such as
“Would you agree that this data point is an outlier, and that I should reject it?”, to which
the demonstrator can begin to answer by saying “yes” or “no”. However, do not hesitate
in letting your demonstrator know if you are confused, or if you have not understood
something.
Rounding values
During calculations, rounding of numbers during calculations should be avoided, as
rounding approximations will accumulate. Carry one or two extra significant figures in all
values through the calculations. Present rounded values for intermediate results, but use
only non-rounded data for further processing. Present a rounded vaLue For Your FinaL answer.
Your final quoted errors should not have more than two significant figures.
xx First Year PhYsics LaboratorY ManuaL
Systematic and
random errors
A systematic error is one that is reproduced on every simple
repeat of the measurement. The error may be due to a calibration
error, a zero error, a technique error due to the experimenter, or
due to some other cause. A random error changes on every repeat
of the measurement. Random errors are due to some fluctuation
or instability in the observed phenomenon, the apparatus, the
measuring instrument or the experimenter.
Independent and
dependent errors
The diameter of a solid spherical object is 18.0 ± 0.2 mm. The
volume, calculated from the usual formula, is 3.1 ± 0.1 cm
3
(check this, including the error). These errors are dependent: each
depends on the other. If I overestimate the diameter, I shall cal-
culate a large value of the volume. If I measured a small volume,
I would calculate a small diameter. Any measurements made with
the same piece of equipment are dependent.
Suppose I measure the mass and find 13.0 ± 0.1 g. This is
an independent error , because it comes from a different
measurement, made with a different piece of equipment.
There is a subtle point to make here: if the error is largely due to
resolution error in the measurement technique, the variables mass
measurement and diameter measurement will be uncorrelated :
a plot of mass vs diameter will have no overall trend. If, on the
other hand, the errors are due to population variation, then
we expect them to be correlated: larger spheres will probably
be more massive and a plot will have positive slope and thus
positive correlation. Finally, if I found the mass by measuring the
diameter, calculating the volume and multiplying by a value for
the density, then the mass and size have inter-dependent errors.
Standard deviation
(σ n- 1
The standard deviation is a common measure of the random
error of a large number of observations. For a very large number
of observations, 68% lie within one standard deviation (σ) of
the mean. Alternatively, one might prefer to define their use of
the word “error” to mean two or three standard deviations. The
sample standard deviation (σ n- 1
) should be used. This quantity is
calculated automatically on most scientific calculators when you
use the ‘σ+’ key (see your calculator manual).
Absolute error The error expressed in the same dimensions as the value. e.g.
43 ± 5 cm
Percentage error The error expressed as a fraction of the value. The fraction is
usually presented as a percentage. e.g. 43 cm ± 12%
First Year PhYsics LaboratorY ManuaL xxi
Error Estimation
We would like you to think about the measurements and to form some opinion as to how to
estimate the error. There will possibly be several acceptable methods. There may be no “best”
method. Sometimes “best” is a matter of opinion.
When attempting to estimate the error of a measurement, it is often important to determine
whether the sources of error are systematic or random. A single measurement may have
multiple error sources, and these may be mixed systematic and random errors.
To identify a random error, the measurement must be repeated a small number of times. If the
observed value changes apparently randomly with each repeated measurement, then there
is probably a random error. The random error is often quantified by the standard deviation
of the measurements. Note that more measurements produce a more precise measure of the
random error.
To detect a systematic error is more difficult. The method and apparatus should be carefully
analysed. Assumptions should be checked. If possible, a measurement of the same quantity,
but by a different method, may reveal the existence of a systematic error. A systematic error
may be specific to the experimenter. Having the measurement repeated by a variety of
experimenters would test this.
Error Processing
The processing of errors requires the use of some rules or formulae. The rules presented here
are based on sound statistical theory, but we are primarily concerned with the applications
rather than the statistical theory. It is more important that you learn to appreciate how, in
practice, errors tend to behave when combined together. One question, for example, that
we hope you will discover through practice, is this: How large does one error have to be
compared to other errors for that error to be considered a dominant error?
An important decision must be made when errors are to be combined. You must assess
whether different errors are dependent or independent. Dependent and independent errors
combine in different ways. When values with errors that are dependent are combined, the
errors accumulate in a simple linear way. If the errors are independent, then the randomness
of the errors tends, somewhat, to cancel out each other and so they accumulate in quadrature ,
which means that their squares add, as shown in the examples below.
First Year PhYsics LaboratorY ManuaL xxiii
Averages
When performing an experiment, a common way of obtaining a better estimate of
something which you are trying to measure, is to take repeated measurements and calculate
the average, or mean, of these measurements. Random errors in the measurements cause
them all to be slightly different, so the mean may be thought of as an estimate of the “true
value”.
For a set of n measurements
x 1
, x 2
, x 3
,, x n
( ) the mean, x^ , is given by
n
x =
∑ x i
i=
n
For example, supposing we measured the time it takes for a ball to fall through a height of 3
metres, and repeat it another six times. The results might look like this:
Time (s) 1.11 1.18 1.02 1.09 1.10 1.13 1.
The mean of this data is given by
= 1.107 s
Just as there is uncertainty in each of the measurements, there is also uncertainty in the
mean. This uncertainty may be calculated in a variety of ways, which depend on how many
measurements have been made.
The simplest estimate of the uncertainty in the mean comes from the range of the data. The
range is simply the difference between the most extreme values in the set. In the above
example the extreme (highest and lowest) values are 1.18 and 1.02, so the range is 0.16.
The uncertainty in the mean, then, is given by the half of the range, that is
where and are the highest and lowest values in the set, respectively. This method
provides a good estimate when you have less than 10 measurements. So in our example,
the uncertainty in the mean time is 0.16/2 = 0.08, so that we can write the mean as (1.11 ±
0.08) s. Note that we have dropped the last significant figure in the mean, as the uncertainty
is large enough to make it meaningless.
xxiv First Year PhYsics LaboratorY ManuaL
If you have taken a lot of measurements, then the uncertainty in the mean is given by the
standard deviation, σ (mentioned earlier on page 18 ). Most calculators can calculate the
standard deviation, but you can also use the formula:
σ = √
∑ ( x 1
− x )
2
n
and the error in the mean is given by
Δ x = σ n−
σ
√ n^ − 1
Generally, in First Year Lab, you wont be taking very large sets of measurements, and so will
not be using this method. Use half the range for a quick simple error estimate, taking care
to be sure that your extreme values are not outliers (unusually small or large values, usually
the result of some one-off mistake). Outliers should generally be discarded when taking
averages, and when graphing.
xxvi First Year PhYsics LaboratorY ManuaL
The error in the gradient is
∆m =
m
− m −
To find the y-intercept, other lines of worst fit may have to be drawn. The worst fit that
produces the greatest y-intercept, and the worst fit that produces the smallest y-intercept
may not necessarily be the same as the worst fits used to find the extremes in gradient. The
extremes in the y-intercept may be produced by a combination of rotating the fitted line
and moving it without rotation.
In the case shown above, the deviation of the measured values from the fitted line are
comparable in size to the error bars. This is a ‘normal’ case.
g
f
In the graph to the left, the error bars are large
compared with the departure of the measured
points from the fitted line. This suggests that
the error estimates are too large: they should
be re-examined.
In this graph, the error bars are small in
comparison with the departure of the
measured points from the fitted line. It is
impossible to fit a straight line without
rejecting a substantial fraction of the data as
outliers. Such a result suggests either:
(a) the error estimates are too small;
(b) that the measurements were made
carelessly;
(c) that numerical blunders have been
made in treating the data; or
(d) that the relation is better described
as non-linear, which means that the theory
which gives a straight line in this plot is
wrong or inappropriate here.
g
f
Further, the general shape of the points suggests that it would be a good idea to try a
different plot, such as g vs for ln g vs ln f.
First Year PhYsics LaboratorY ManuaL xxvii
Automatic graphing routines
Most common software packages that graph data and fit lines or curves do not take errors
into account. Many do not even plot the errors. Further, they give all points equal weight,
even if there is a big variation in the error bars. Graphing by hand, as described above,
you give the points with small error bars more importance (the statistical term is weight )
because the line is more tightly constrained to pass through the smaller error bars. It is
possible to include appropriate weighting factors (usually the reciprocal of the error) in
automatic routines.
In the first year lab, you can use the excel template “Linear plot with Errors”. This will plot
the line of best fit and also the maximum and minimum gradient lines that satisfy your data.
References
‘Experimental Methods. An Introduction to the Analysis and Presentation of Data.’ Les
Kirkup, Wiley, (1994).
‘Data Reduction and Error Analysis for the Physical Sciences.’ Philip R. Bevington, McGraw
Hill (1969).
‘Statistical Methods in Medical research.’ (3rd Edition) P. Armitage & G. Berry, Blackwell:
Oxford, (1994).
‘Handling Experimental Data.’ Mike Pentz, Milo Shott and Francis Aprahamian, Open
University Press (1988).
