Measurements and Uncertainties

The establishment or verification of a physical law or the experimental determination of a

physical quantity usually involves measurement. A reading taken from the scale on a stopwatch,
a meter stick or a voltmeter, for example, may be directly related by a chain of analysis to the
quantity or law under study; any uncertainty in these readings would thus be directly reflected in
an uncertainty in the final result. A measurement by itself, without at least a rough quantitative
statement as to the uncertainty involved, produces a result of limited usefulness. It is therefore
essential that any introductory physics laboratory course include a discussion of the nature of
experimental uncertainties and the manner in which they should be assigned to experimental
results.
1. UNCERTAINTY vs. DISCREPANCY
When you report the result of a measurement of a quantity x, you should also give the
uncertainty x, e.g.,
5.0 m 0.1 m

x
x
The uncertainty tells you how precise you think your measurement is. It is also often useful to
compare your result with a "true" or accepted value; the difference between these is the discrepancy and is a reflection of the overall accuracy of the measurement.
An estimate of the uncertainty of a measurement should always be made; a calculation of the
discrepancy can be made only if an accepted value or result happens to be available for
comparison. The conclusion for an experiment should, whenever possible, address the question:
Do the uncertainties involved in my measurements account for the discrepancies between my
result and the accepted one?

gexp

g acc
9.81

gexp
9.75

As an example, suppose you do an experiment to measure the acceleration of gravity and you
m
report the experimental value ( g exp g exp ) to be 9.75 0.08 2 where the accepted value is
s
m
g acc = 9.81 2 . As you can see from the graphic representation below, the uncertainty g exp in
s
the measurement accounts nicely for the discrepancy between gexp and gacc.

Uncertainties

2. ORIGINS OF UNCERTAINTIES
Problems which lead to discrepancies between experimental values and "true" values can be
placed in two general categories:
I. Systematic Errors are inaccuracies due to identifiable causes and can, at least in principle, be
eliminated. Errors of this kind result in values for the measured quantity which are
consistently either too high or too low. Such errors can be
a) Theoretical - due to simplifications of the model system or approximations in the
equations describing it.
b) Instrumental - e.g., a poorly calibrated instrument.
c) Environmental - e.g., factors such as inadequately controlled temperature and pressure.
d) Observational - e.g., parallax in reading a meter scale.
II. Random Uncertainties are the result of small fluctuating disturbances which cause about half
the measurements of any quantity to be too high and half to be too low. It is often not
possible in practice to identify all the sources of such errors, to evaluate their effects
individually or to completely eliminate them. They can, however, often be well
characterized mathematically.
To illustrate the difference between systematic errors and random uncertainties, we
consider the measurement of the length of time Tf taken for a ball to fall some fixed distance, say
10 m. Suppose we drop the ball about 1000 times and obtain as many values of Tf, rounding
each time to the nearest .01 s. If N(Tf) is a function of Tf that represents the number of times a
particular measured value of Tf occurs, then a histogram of N versus Tf might look like

The "best" value for Tf is just the average (mean) value Tf :

2

Uncertainties
n

N T

i fi

Tf =

i=1

(1)

where Ni is the number of times the value Tfi appears as a measurement and n is the total number
of times the ball is dropped. In this example we see that Tf = 1.43 s; the discrepancy between
this number and the "true" value of Tf, if known, would provide a measure of the accuracy of the
measurement. Systematic errors (e.g. the reaction time involved in starting and stopping the
clock) will affect Tf and hence the accuracy. The spread of Tf values, indicated by the width of
the curve, is a reflection of the precision with which Tf is being measured. Random uncertainties
contribute to this width and may be attributable, to e.g., small fluctuations of the height from
which the ball was dropped or to the difficulty in determining the exact moment of impact with
the ground. In what follows we discuss the mathematical treatment of such random uncertainties
and the role they play in reporting laboratory measurements.
3. CHARACTERIZING A SET OF DATA: THE NORMAL DISTRIBUTION
It is most often the case that repeated measurements of the same quantity will, as in the
timing experiment example described above, exhibit a spread about the average value related to
the random errors associated with the measurement process. If we make "many" (say 106)
measurements of a quantity x and plot the frequency of occurrence N(x), we quite often obtain a
curve that approximates a Gaussian, or normal distribution, as pictured below.

This curve N(x) represents the relative probability with which values of x are obtained as the
result of any single measurement of this quantity, which may be, for example, the reading of a
meter or stopwatch. The analytical expression for such a curve is
2
N 0 - (x 2)
N(x) =
(2)
e 2
2
where the parameters and determine the position and width of the peak, respectively. The
"normalization" parameter NO would correspond, in our timing example, to the total number of
readings taken from the stopwatch.
3

Uncertainties
The curve representing Eq. (2) is of importance for the analysis of experimental data
because in many cases this data is distributed normally and thus has a frequency distribution that
can be "fit" to this curve. The more data taken, the better the fit will be. But for any set of data,
regardless of the number of data points or the nature of the distribution, we can define quantities
which characterize the "best value" and "spread" of the data.
For a set of data points xi, the quantity x is the mean of all values obtained for x, and is
defined by
n

x=

i =1

x1 + x 2 +L+ x n
n

(3)

where n is the number of measurements. For data in which each value of xi generally occurs
more than once, Eq. (1) may provide a more convenient way to calculate the mean value of x.
The "best value" that one finally reports for the quantity x is generally its mean, as defined
above. If a large number of normally distributed data points are taken, then _x should be close to
the value of for which the curve representing Eq. (2) best fits the data.
We can characterize the spread of any finite set of n data points by its standard deviation,
symbolized by s.d., which we will define as
n

s. d. =

(x

x )2

i =1

(4)

The s.d. for a set of data thus represents the square root of the average square of the deviation
(i.e., the "rms" deviation) that the data points exhibit with respect to their average value. For a
large number of normally distributed data points, s.d. , where is the value of the parameter
of Eq. (2) which produces the best description of the data.
4. REPORTING THE VALUE OF A MEASURED QUANTITY
Because repeated measurements of the same quantity give us, as discussed above, a range
of values reflecting the random fluctuations inherent in the measurement process, we need to
report the result of these measurements in a way which reveals not only the best value of the
quantity, but also the precision to which we are stating it: we report both the best value and the
uncertainty. The distributions discussed above refer to a collection of single measurements. So
if we make a single measurement of a quantity x and had some knowledge of the single
measurement distribution curve, then we could report this value of x as
(5)
x s. d.
where x is the single measured value and s.d. is the standard deviation of the single measurement
distribution. To determine an appropriate value for s.d. without actually making several
4

Uncertainties
measurements, we would have to know something, a priori, about the single measurement curve;
we would either have to be given a value of s.d. or we would have to guess.
Suppose, as is more common, we take n measurements instead of just one. Intuitively, it
seems reasonable that the mean value should be reported with a smaller uncertainty than the
single measurement discussed above. In fact, if one took many sets of n measurements and then
calculated the mean for each set, the distribution of these means would have a smaller standard
deviation (denoted S.D.) than that for a single measurement (s.d.). It is not hard to show that
s.d.
S.D.
(6)
n
What this means is that one could report the result of a set of n measurements as
s. d.
x S. D. x
(7)
n
so that the uncertainty associated with the average of 5 measurements, e.g., could be reduced by
about a factor of 2 from the single measurement s.d. NOTE: The expression above for reporting
an average is only meant to serve as a guideline in reporting results; it assumes that all
uncertainties involved are purely random in nature. Systematic errors must always, at some
level, be considered, so that in most cases we cannot create an arbitrarily small uncertainty
simply by taking more and more measurements!
Relative vs. Absolute Uncertainty. One often reports the relative (or percentage) uncertainty associated with the best value in lieu of the absolute uncertainty x = s.d. or S.D. No units
are associated with these relative uncertainties, so that these may be used to directly compare the
precision of two different measurements. On the other hand, the absolute uncertainty is more
convenient for comparing two measured values.
Example. Suppose we make 10 measurements of the period T of a pendulum, and that the
results are tabulated to the nearest millisecond (ms) as
T (s)
TRIAL #
T (s)
TRIAL #
1
2
3
4
5

1.411
1.452
1.403
1.414
1.459

6
7
8
9
10

1.468
1.437
1.446
1.425
1.434

We need to report the best value for T along with the uncertainty, which we assume is due to
random fluctuations. For the best value, we have the mean,
10

T=

= 1435
.
s.
10
To estimate the uncertainty, we calculate s.d. from Eq. (4):
i =1

Uncertainties
10

s. d. =

(T T )

= 0.02 s.
10
But since we made 10 measurements, we can report the uncertainty as
s.d.
T S.D.
0.006 s.
10
Our result for T thus appears as
i =1

T = 1.435 s 0.006 s (with absolute uncertainty)

or
T = 1.435 s 0.5 % (with relative uncertainty)
Significant Figures. A decision must always be made, consciously or otherwise, as to how
many significant figures should be used to report a quantity x and its uncertainty x. Since
uncertainties are merely estimates, it is seldom necessary to report x to more than 1 significant
figure! The number of significant figures with which you report x is often taken to imply its
uncertainty, therefore this number must be consistent with the explicitly stated uncertainty x.
Rule of thumb for significant figures: The measurement and the uncertainty should have
their last digits in the same location relative to the decimal point. Example:
(1.06 0.01) 103 m.
Thus, a length stated as 1.0 m 0.002 m would present a problem because the stated
uncertainty would imply that the observer is withholding information by not reporting more
decimal places in the length. A result stated as 1.06 m 0.2 m reports a length with too much
accuracy, given the uncertainty. If x and x imply different uncertainties, one of them should be
adjusted so that they are both consistent with the largest of the two uncertainties.
Guesstimating Uncertainties. The uncertainty to assign a given single measurement of x
depends on the technique employed in the measurement, the quality of the equipment used, and
the care with which the measurement is made. The number you come up with as x for a
particular measurement is often quite subjective, because only you know what happened during
your measurement. There are a couple of options for estimating x:
1. Take several measurements and calculate s.d., as in the example above for T;
2. Estimate x from the nature of the measurement. For example, when reading a scale,
the uncertainty is sometimes, but not always, half of the smallest scale division, but you
will encounter situations when the scale divisions have nothing to do with x. In other
words, you will always have to use your judgment.
You will find yourself using option #2 quite often; it's a real timesaver.
6

Uncertainties
5. PROPAGATION OF UNCERTAINTIES
Quite often the quantity of interest is measured indirectly, i.e., derived from one or more
quantities which are each measured directly and reported with uncertainties, as discussed above.
In these cases, the estimate of the uncertainty in the indirectly measured quantity must reflect the
uncertainties associated with all of the relevant directly measured ones.
For example, suppose we wish to infer the acceleration of gravity (g) from direct
measurement of both Tf (the time taken for an object to fall) and of L (the height from which it is
dropped). The equations of kinematics tell us that g can be calculated from Tf and L:
2L
g =
.
2
Tf
How, then, do the uncertainties in L and Tf contribute to g, the uncertainty in g?
You will need to know the answer to this question, and to others just like it. We first give
the general expression for propagating the uncertainties in directly measured quantities to find
the uncertainty in the derived result. We then treat two special cases which occur quite often.
General Case: Let z be a quantity to be derived from the direct measurement of two independent
quantities x and y by using the relationship z = f(x,y). Suppose we were to make small definite
(known) errors x and y in the measurements of x and y respectively. These would show up as
an error z in z that would be given by
f
f
z
x + y .
x
y
But because x and y are actually random uncertainties, a more realistic estimate of the
uncertainty z is
2

f
f
z x + y ,
x y
2

(8)

as can be demonstrated from considerations of probability theory. Given uncertainties in the two
quantities x and y, Eq. (8) can be used to estimate the propagated uncertainty z. The
generalization to more than two directly measured quantities is straightforward.
Origin of Eq. (8) (Optional): Consider two independent random variables P and Q, each with an
average of zero and with s.d.'s of P and Q, respectively. The s.d. of the sum P + Q (denoted
(P+Q)) is, according to the definition of Eq. (4),
(P+ Q) =

(P+ Q )2

If we compute the square of this quantity in terms of P and Q, we get

7

Uncertainties
from which the s.d. for P + Q is
(P+ Q) =

( P )2 + ( Q )2 .

Now considering the function z = f(x,y): the uncertainty in z (z), will be due to the combined
effects of the uncertainty in x ( x) and the uncertainty in y (y). If P and Q represent the effect
on z of changes in x and y, respectively, then P and Q, the s.d.'s for P and Q, will be given by
f
f
P =
x; Q =
y,
x
y
where x and y are the s.d.'s, or uncertainties, in x and y. The spread, or uncertainty z in z
will then be (P+Q), given above. Substitution for P and Q results directly in Eq. (8) for z.
It should be noted that this same argument can also be applied to derive Eq. (6) above for the
uncertainty (S.D.) in the average.
Many times, it is not necessary to use Eq. (8) directly because the form of the expression
for z will fit one of two special cases.
Special Case I: The quantity of interest (z) is given in terms of two directly measured quantities
x and y, each with their respective uncertainties x and y, as z = Ax + By. In this case, the
absolute uncertainty in z is given by
z = (A x)2 + (B y)2 + L

(8a)

where the " + L " indicates that this expression is easily generalized to the case where z is given
as the sum of three or more quantities.
Example: We wish to determine the mass of water, mw, contained in a beaker from independent
measurements of the mass of the beaker alone, mb, and the combined mass of the beaker + water,
mt. Then mw = mt - mb and the uncertainty in the derived quantity mw is
m w = ( m t ) 2 + ( m b ) 2 ,

where mt and mb must either be estimated or determined from repeated measurements of these
quantities.
Special Case II: The derived quantity (z) is the product of powers of measured quantities x and
y: z = K xa yb, where K is a constant. In this case we can give the fractional (relative)
uncertainty in z as
z
x y
= a
+ b

z
x y
2

(8b)

Uncertainties
where we can again generalize to any number of quantities, as indicated for Eq. (8a).
Example: If the acceleration of gravity g is given, as discussed above, in terms of L and Tf as
g = 2L/Tf2, what is g in terms of the uncertainties L and Tf? If g is rewritten as
g = 2(L1)(Tf-2), then we have something that fits the form of special case II with x = L, a = 1 and
y = Tf, b = -2. So for the relative uncertainty g/g we get
2

g
L
Tf
=
+
4
T .
g
L

Note that because g depends on the second power of Tf but on only the first power of L, the
uncertainty in g is much more "sensitive" to the uncertainty in Tf than to the uncertainty in L.
If neither of the special cases applies to the quantity being investigated, it is sometimes
possible to derive the uncertainty by applying Eq. (8a) and (8b) in combination; otherwise, use
the general expression of Eq. (8) directly.
Graphs and Uncertainties. Physical laws often predict that two quantities (x and y) that can be
measured (directly or indirectly) are proportional to each other. For example, Newton's second
law in one dimension (F = Ma) predicts that the acceleration (a) of an object is proportional to
the force F (the mass M being the constant of proportionality); Ohm's law (V = IR) predicts, for
certain materials, that the voltage V and current I are proportional. It is often the constant of
proportionality between these measured quantities (M or R in the above examples) that is to be
determined by the experiment, rather than just individual pairs of values for x and y. The linear
relationship between x and y can be expressed as
slope

y = mx + b

y-intercept

If y is plotted vs. x, as in graph (a) below, then the result should be a line with a slope m,
calculated as y/x. (Note that here y and x refer to the rise and run, respectively, not to the
uncertainties in y and x!).
y

m (best fit) = y
x

m+

m
__
(x,y)

x
x
(a)

(b)

Uncertainties

Of course, experiments being what they are, the data points (as indicated in graph (a)) will not all
be on a single line; it is up to you to draw the "best" line that "fits" the data so that the slope m
(or y-intercept b) may be determined. There is a well-known method for determining the best
line for a given set of points (x,y) -- the least squares fit (see below). For the purposes of this
lab, however, it will often be sufficient to use a straightedge and your eyes to draw the line which
represents the best compromise, i.e., "comes as close to as many points as possible."
NOTE: If the slope or the intercept is to be measured directly from the graph, it is most
often advantageous to use as much of the page as possible to draw the graph!
If the slope m is the quantity to be reported as the result of the experiment, then the
uncertainty m must be reported along with it: we need m m. The uncertainty m can be
thought of as arising from the uncertainties in the position of each plotted point on the graph,
since each y value may be characterized by an uncertainty y. One way of representing these
uncertainties is by means of the "error bars" drawn for each point on graph (b) above. As with
the slope m, there is a well known (and somewhat complex) expression that can be used to
determine m, but for our purposes it is sufficient to "eye ball" the range of slopes. As indicated
in graph (b), this can be done by using the error bars to estimate m+, the highest possible slope
for a line that passes through or near most of the error bars, and m-, the lowest slope. The
uncertainty m is then very roughly given by
m+ m
m
.
2
To estimate the slope (m) and its uncertainty (m) by eye:
1.

Plot the data points (x,y) on graph paper. Use as much of the page as possible.

2.

Place error bars through each point to indicate roughly the extent to which the location of
that point on the graph is uncertain. The size of these bars can be estimated from the
discussion under the "uncertainty" section in each experiment. An alternative to bars is
to enlarge your dot to indicate its "fuzziness." If the uncertainty is too small to show up
on your graph, state this in the report.

3.

Plot the point ( x, y ) on your graph (where x and y are the means of the x and y values,
respectively). It so happens that this point always lies on the "best line" defined below by
the least squares fit procedure.

10

Uncertainties
4.

Using a straight edge, draw the line passing through ( x, y ) that "comes closest" to the
rise
data points. Measure the slope m as
, using as much of the line as possible and
run
paying attention to the units on the axis scales.

5.

Estimate the uncertainty m by drawing 2 dotted lines through ( x, y ) and passing within
most of the error bars--one with the minimum slope, m, and one with the maximum
slope, m+. Estimate m from the expression in the paragraph above.

Least Squares Fit (Linear Regression): Though somewhat more involved and less intuitive than
the graphical method described above for finding the slope and intercept of a best fit line, the
analytic method of linear regression finds the best fit by minimizing the sum of squares of
deviations of y values from the fitted straight line. For this reason, it is called the method of least
squares. Given N data points x1, y1; x2, y2; ...; xN, yN, the slope m and y-intercept b of the best fit
line (given by y = mx + b) are calculated as:
N

(x x)(y y)
m=

i=1

N x 2

; b = y mx,

(9)

where x and y are the means of the x and y values respectively, given by Eq. (3), and x is the
s.d. of the x values as given by Eq. (4).
As an alternative to the graphical method described above for estimating the uncertainty
m for the slope, you may use the least squares methodology to estimate this quantity along with
the uncertainty in the intercept, b. The expressions for these are
N

1
m =
x

d
i =1

N( N 2)

b = m x 2 + x 2 ,

11

(10)

Uncertainties
in which di yi (mxi + b) is the y deviation of the point (xi,yi) from the best fit straight line
y = mx + b. If your calculator does linear regression calculations which result in the least
squares fit slope m and correlation coefficient r, then it may be simpler to calculate the
uncertainty in the slope (m) with the equation*
m = m

tan ( arccos ( r ) )
N2

m
r

1 r2
.
N2

(11)

References and further reading:

1.
Lichten, William, Data and Error Analysis in the Introductory Physics Laboratory (Allyn
and Bacon, Inc.)
2.

Taylor, John R., An Introduction to Error Analysis (University Science Books)

J. Higbie, American Journal of Physics, 59(2) 1991

12

Summary on Uncertainties and Least Squares Fitting

I.

Mean (average) and Standard deviation of means

a) Measurements : data
x1 x, x2 x, x3 x,., xN x
where x = uncertainty of measurement (random error)
b) Mean (average): the best value to represent this set of data
x + x2 + ... + x N
x= 1
N
c) Standard deviation of a single measurement x
s.d . = x =

(x x1 )2 + (x x2 )2 + ... + (x x N )2
N

d) Standard deviation of Means: S. D.

s.d .
x
S .D.
=
N
N
e) Report of the measurements;
x S.D.
S.D. in one significant figure
II.

Propagation of Uncertainties
Given: AA, BB and f = f(A,B)
a) In general
2

f
f

f = A +
B
A B

b) Addition/ subtraction
f = A + B

f =

( A)2 + ( B )2

c) Multiplication/ division/ powers

f = A B
A B
f = f
+

A B
2

2006 Eric R. Dietz

Uncertainties
III.

Physics 204B Laboratory

Least Squares Fit (Linear Regression):

a) Measurements: (x1, y1), (x2, y2),, (xN, yN), N pairs of data.
b) In theory

y = mx + b
m = slope,

b = y-intercept

c) Result of Least Squares Fit

m=

N xi y i xi y i
N xi2 ( xi )

y x x x y
b=
N x ( x )
2
i

m =

[y
i =1

2
i

(mxi + b)]

N ( N 1)

1
1
2
1 r 2
=m r
m
N
N

r = correlation coefficient

(x x )

b = m + x , x =
2
x

14

x=

x
N