Experimental Uncertainty

Abstract
This is intended as a brief summary of the basic elements of uncertainty analysis, and a handy
reference for laboratory use. It provides some elementary "rules-of-thumb" which are satisfactory for use in an
introductory physics laboratory.

References
For thorough derivations and justifications of the material presented in this summary, the student
should consult one of the following:

Taylor, Introduction to Error Analysis, Second Edition, University Science Books (1997)
Young, Statistical Treatment of Experimental Data, McGraw-Hill (1960)
Beers, Introduction to the Theory of Error Analysis, Addison-Wesley (1958)

Measurement Uncertainties and Uncertainty Propagation

In most experiments, certain quantities are measured and then other quantities are determined from the
measured data. For example, we might measure the length and width of a room, and then determine its area by
multiplication. Associated with each measured quantity is a "measurement uncertainty". When a final result is
determined from measured quantities, these measurement uncertainties lead to an uncertainty in the derived
result. The procedure by which an experimenter determines the uncertainty in a final result from the
measurement uncertainties is called "error analysis", or "propagation of uncertainty".

Types of Uncertainty (Absolute, Fractional, Percent, Relative)

The absolute uncertainty in a quantity is the actual amount by which the quantity is uncertain, e.g.if
L = 6.0 ± 0.1 cm, the absolute uncertainty in L is 0.1 cm. Note that the absolute uncertainty of a quantity has
the same units as the quantity itself.

The fractional uncertainty is the absolute uncertainty divided by the quantity itself, e.g.if L = 6.0 ±
0.1 cm, the fractional uncertainty in L is 0.1/6.0 = 1/60. Note that the units cancel in this division, so that
fractional uncertainty is a pure number.

Percent uncertainty is fractional uncertainty expressed as a percent, i.e. fractional uncertainty

multiplied by 100. If L = 6.0 ± 0.1 cm, its percent uncertainty is 1.7%.

The word "uncertainty", by itself, normally means "absolute" uncertainty. Fractional or percent
uncertainties are both called "relative" uncertainties because they relate the size of the uncertainty to the size of
the result itself. While not normally used in the presentation of data or experimental results, relative
uncertainties are often useful in intermediate stages of "error analysis" calculations.

Rule #1: Express the uncertainty in each of your measurements or derived results as
an absolute uncertainty, using the same units for the quantity and its
associated uncertainty.
 Measurement Uncertainties
The first step in determining the reliability of an experiment is to evaluate the measurement
uncertainties associated with each measured quantity. While in many cases, measurement uncertainty is due to
the particular instruments used, it may also depend on human limitations of the experimenter or on some
randomness in the effect being measured.

As an example of instrumental measurement uncertainty, consider a measurement of the length of this

page using a meter-stick. When the zero of the meter stick is placed at the top of the page, the bottom appears
to lie between 27.9 and 28.0 cm, although somewhat closer to 27.9 cm. Since the marks on the meter-stick are
0.1 or 0.2 centimeters wide, and placing the zero at the top edge of the paper might be "off" by 0.1 centimeter,
the experimenter might feel confident saying that the page's length to the nearest tenth of a centimeter is 27.9
cm, but would certainly feel that specifying the length to the nearest hundredth centimeter is risky. Hence, it
would be prudent to state the length as 27.9 ± 0.1 centimeters. This situation illustrates two basic rules-of-
thumb:

Rule #2: Make uncertainty estimates large enough to give yourself a margin of safety.
You should feel confident that the "real value" of the measured quantity lies
somewhere within the uncertainty range you specify.

Rule #3: The measurement uncertainty due to an instrument itself may be assumed
equal to the smallest scale division of the instrument.

A common situation in which measurement uncertainty relates more to human limitation than to
instrumental error is the measurement of time with an electronic stopwatch. The watch can measure to .01
second. When a careful experimenter makes five separate measurements of the time for a ball to roll a fixed
distance down an inclined plane, the results are likely to look like this: 4.63, 4.69, 4.64, 4.56, and 4.70
seconds. While each time is measured to .01 second, the range of values is 4.70 minus 4.56, or 0.14 seconds.
In another common experiment, a spring gun fires a projectile horizontally from the edge of a lab table. The
horizontal distance from the table's edge to the spot where the projectile hits the floor is easily measured to
within ± 1 cm, but due to an inherent randomness in the operation of the spring gun, typical results from a
series of five trials using the same apparatus might be: 256, 239, 263, 252, and 253 cm. In both these
situations, the best an experimenter can do is take an average of the measurements, namely 4.644 seconds and
252.6 cm, respectively. But from the time range of 0.14 seconds and the distance range of 263 - 239 = 24 cm,
the experimenter might reasonably estimate the uncertainty in time to be 0.14/2 = 0.07 → 0.1 seconds and
distances uncertainty to be 24/2 = 12 → 10 cm. The value for each would then be written as 4.6 ± 0.1 and 250
± 10 cm. Notice that the uncertainties are related in only an approximate way to the ranges of values
obtained in a series of 5 measurements, and that the uncertainties in both cases make it unreasonable to state
the results to more than two "significant figures" (i.e 4.6 and 2.5 x 102).

Rule #4: If a measurement is not reproducible within instrumental uncertainty, the

measurement uncertainty should be estimated from the range of values
obtained in a series of five (or more) trials.

Rule #5: No quantity should be stated to more significant figures than are justified by
its associated uncertainty.
 Uncertainty Propagation: Worst Case Method
The most straightforward way to find the uncertainty in the final result of an experiment is worst case
error analysis, a method in which uncertainties are estimated from the difference between the largest and
smallest possible values that can be calculated from the data. As an illustration of the method, suppose we
measure the length and width of this page to be 27.9 ± 0.1 cm and 21.6 ± 0.1 cm, respectively. Then the most
probable value for its area is (27.9)(21.6) = 602.64 sq cm. But the largest possible value ("worst case")
consistent with the data is (28.0)(21.7) = 607.6 sq cm. Similarly, the smallest possible area is (27.8)(21.5) =
597.7 sq cm. Since the actual area might be anywhere between these extremes, it is reasonable to state the
result as 603 ± 5 sq cm.

In a second example, the mass and volume of a sample of oil are measured as 7.54 ± 0.01 g and 10.0 ±
0.1 cm3, respectively. The density of this oil is most probably (7.54 g)/(10.0 cm3) = .754 g/cm3, but might be
as large as (7.55 g)/(9.9 cm3) = .763 g/cm3 or as small as (7.53 g)/(10.1 cm3) = .746 g/cm3. Note that in this
case we must use the smallest possible volume when calculating the largest possible density. This range of
possible density values suggests that the final result should be expressed as 0.75 ± 0.01 g/cm3.

Rule #6: When a result is calculated from several measured quantities, its uncertainty
is approximately half the difference between the largest and smallest
possible results that could be obtained from the measurements and their
associated measurement uncertainties.

Differential Error Analysis

A second, and somewhat more formal, method of treating the propagation of uncertainties is related to
differential calculus. When an experimental result, f, depends on several independent measured quantities (x,
y, z, ...) in such a way that f can be written as a differentiable function of the measured quantities, the change
in f which results from small changes in the measured values can be determined from the calculus of partial
derivatives:

f = f(x, y, z, ...)

∂f ∂f ∂f
df = dx + dy + dz + ...
∂x ∂y ∂z

Since an uncertainty can be interpreted as a range within which a quantity is free to change, we can
relate the absolute uncertainty in f (call it ∆f) to the absolute uncertainties ∆x, ∆y, ∆z, etc, in a simple way:

Rule #7: If a final result, f, is a differentiable function of several measured quantities

(x, y, z, ...), the uncertainty in f relates to measurement uncertainties
according to:

∂f ∂f ∂f
∆f = ∆x + ∆y + ∆z + ...
∂x ∂y ∂z

The absolute value signs (along with the fact that x, y, z, etc are positive by definition) guarantee that
the contributions of different measurement uncertainties to the uncertainty of the final result are always
additive. This is in contrast to the calculus of analytic functions, in which variables can change in such a way
as to produce canceling effects on the function itself.
 Special Cases--Results of Differential Error Analysis
We now proceed to look at some special cases of Rule #7, in an attempt to find some simple rules-of-
thumb:

∂f ∂f
1) f = x + y =1 =1 ∆f = ∆x + ∆y
∂x ∂y

∂f ∂f
2) f = x - y =1 = −1 ∆f = ∆x + ∆y
∂x ∂y

∂f ∂f
3) f = xy =y =x ∆f = y∆x + x∆y
∂x ∂y

x ∂f 1 ∂f x 1 x
4) f = = =− 2 ∆f = ∆x + 2 ∆y
y ∂x y ∂y y y y

The ∆f in each case is the absolute uncertainty in f. For #3, the fractional uncertainty turns out to have
a particularly simple form:

∆f y x
= ∆x + ∆y
f xy xy

∆x ∆y
= +
x y

The fourth example also yields a simple result if we look at the relative error:

∆f y xy
= ∆x + 2 ∆y
f yx xy

∆x ∆y
= +
x y

These four examples yield two simple rules:

Rule #8: The absolute uncertainty of the sum or difference of two quantities is the
sum of their absolute uncertainties.

Rule #9: The relative uncertainty of the product or quotient of two quantities is the
sum of their relative uncertainties.
 Graphs and Error Flags - Uncertainty in the Slope of a Line
In a typical experiment to determine velocity, the position of an object is recorded every second.
Suppose the uncertainty in each position measurement is ± 5 cm, the uncertainty in each time measurement is
± .25 s, and the data for a five-second interval is given by the table to the left of the following figure.

best
max slope line
60
t(sec) x(cm)
0 10 all times
1 21 50
± 0.3 s min slope
2 33
3 42 40
4 55 x (cm)
5 65 30

20 all distances
± 5 cm
10
00

1 2 3 4 5
t (sec)

In the graph to the right, the uncertainties in each data point are indicated by "error flags", the vertical
flags are for uncertainty in distance, horizontal for time. The lengths of this flags correspond to the uncertainty
ranges. For example, since the point at t = (3.0 ± 0.3) seconds corresponds to x = (42 ± 5) cm, the error flag
extends from 37 cm to 47 cm in the vertical direction, and from 2.7 sec to 3.3 sec in the horizontal direction.

The solid line drawn through the data points was estimated by an experimenter to be the "best fit" to
this data. The slope of this line, V = dx/dt = 11 cm/sec, represents the most probable value of the velocity. But
there are many other straight lines which could have been drawn through the error flags, and hence there is a
range of possible slopes consistent with the data. To estimate this uncertainty in the slope, we draw the "worst
case" lines, i.e.the line of greatest slope and the line of least slope which still pass through most of the error
flags. If you think about the slope being given by rise/run, then to get a "max" slope, you will need the "max"
rise/"min" run. The worst case lines are shown as dotted lines in the figure, and have slopes of Vmax = 13
cm/sec and Vmin = 9 cm/sec. Given these extremes, it is reasonable to express the final result for the velocity as
V = (11 ± 2) cm/sec.

Rule #10: The uncertainty in the slope of a straight line fit to experimental data is
approximately half the range of slopes of the set of all straight lines which
can be drawn through the data error flags.
 Standard Deviation and Confidence Limits
We have seen that when a measurement is not reproducible because of random errors, its uncertainty
can be estimated from the results of several trials. A complete theory of random error predicts that for a very
large number of trials, N, the distribution of results will look like the bell-shaped curve shown in the following
figure.
# of results

x − 2σ x + 2σ x→

The peak in the curve corresponds to the average result, x_, and the curve's width is a measure of the
range of results obtained, i.e. the uncertainty in the measurement. If x(i) is the result of the ith trial, we can
define the "deviation" of the ith result as [x(i)-x_], and the "standard deviation", σ , as:

σ=
1 N
[
∑ x(i) - x
N - 1 i =1
] 2

The standard deviation is defined in such a way that 68.3% of all results lie between x - σ and x + σ,
95.5% of all results lie between x - 2σ and x + 2σ, and 99.7% of all results lie between x - 3σ and x + 3σ.
Thus, if you make a very large number of measurements, you have "68% confidence" that a future
measurement will fall within ± σ of the average. In this sense, the standard deviation can be regarded as the
uncertainty in one measurement, for example, the next measurement you will make.

A more useful quantity is the "uncertainty in the average" of all N measurements. What this means is,
if you were to take many "sets" of N measurements and find the average for each, how close together would the
averages be? The answer is that the averages would be distributed in a bell shaped curve of their own, but the
standard deviation of the averages would be smaller than the standard deviation of the data in any one "set".
This is why it's better to take a large number of measurements and find the average than to just take a single
measurement! It turns out that the uncertainty in the average is √N times smaller than the standard deviation.

Rule #11: The uncertainty in the average of a large number, N, of measurements may
be taken as the standard deviation of the data divided by √N.