Principles of Error Analysis

It does not suffice to just give a reader the average value obtained from a series of measurements. We must also convey to the reader our methods abilities, personally and/or experimentally, to make the measurements from which the reported average is derived. The reader must have a feel for our precision and accuracy before they can give proper merit to a reported value. In this document, we discuss how to evaluate this metric of experimental merit. We want to append every reported value with an error and to write reported values and their errors using the following form: Value Error in Value. We will now discuss the various methods of evaluating the term Error in Value. This document assumes the reader is familiar with normal (Gaussian) distributions and can apply qtests and student t-tests to appropriately reduce a data set. Section I will focus on evaluating errors from a series of measurement while section II will deal mainly with propagation of errors. I. Probable Errors The reliability of a method of measurement of a quantity X is indicated by specifying the probable error, qx, of a single measurement. The meaning of qx is that if a single measurement of X is made at random, there is a 50-50 chance that the magnitude of the error associated in measuring X will be less than qx or greater than qx. Thus, for example, if the measurement were repeated many times and these errors listed in order of increasing magnitude, there would be just as many errors below qx as above it. For a hypothetical Gaussian distribution, the probable error can be expressed in terms of the variance, sx, as qx = 0.6745 sx where sx
2

A. An Average from a Series of Measurements

(1)

" (x =

! x)2 N
i

(2)

for a series of measurements (x1, x2, . . . . xN) with the mean value x and with N approaching infinity. Equation 1 follows form the fact that the area under the Gaussian probability curve lying between x 0.6745sx and x + 0.6745sx is exactly 1/2.

For an actual series of measurements with finite N, sx2 is replaced by x2 where x2 is called the standard deviation, (Note: the divisor N is replaced by the degrees of freedom N-1) x
2

" (x =

! x) , N -1
2 i

(3)

and 0.6745 is replaced by the student t-Test for 50% probability, t0.5(N). Values for t0.5(N) versus N can be found in the CRC Handbook.
Table 1: t0.5(N) for values of N commonly encountered.

N t0.5(N) Thus we write qx = t0.5(N) x

2 1.000

3 0.816

4 0.765

5 0.741

6 0.727

0.6745

(4)

To vary the resolution of an error, the confidence level for the student t-test may be changed from 50%. Student-t values are listed for 80%, 90%, 95% and 99% confidence levels. Actually, however, we are really interested in the probable error of the mean, called x , and not the probable error of a single measurement (qx). Increasing the number of measurements will reduce the probable error of the mean: it can be shown that x is proportional to that finally we have the result we want to use in practice: x =

1 so N

qx t ! = 0.5 x N N

(5)

Example: What is the probable error of the mean of the following series of observations? Measurements: length of a table (cm): 34.22, 34.28, 34.16, 34.40, 34.30

x = 34.27
x =

" (x

! 34.27) 2 = 0.0901 4

t0.5(5) = 0.741, therefore x =

(0.741)(0.0901) = 0.02986 5

and we write for the value of the length of the table, X, X = x x = 34.27 0.03 cm What this means is that if we make a series of five (5) measurement 100 times (basically repeating the experiment of making the above five measurements 100 times) giving 100 means, 50 of the means will differ in magnitude from the true mean by less than 0.03 and 50 of them will differ by more that 0.03. Some of course might differ by exactly 0.03, so what is meant is that an equal number will be found with a difference of less than 0.03 and with a difference of more than 0.03. B. Curve Fitting Pairs of Data. For many experiments, one graphs pairs of data and obtains a desired value from either the slope and/or the intercept of the best-fit curve through the data. The most common fit is a Least-Squares fit to a straight line. If (x1, y1), . . . , (xN, yN) are measured pairs of data, then the best straight line, y= b + mx, to fit these N points has

b =
and

(! x )(! y ) - (! x )(! x y )
2 i i i i i

"

(6)

m =
where

( ! x y ) - (! x )(! y )
i i i i

"

(7)

! = N

(" x ) - ( " x )
2
i

(8)

While these equations allow us to evaluate the slope and intercept of our fitted line, we still require the associated errors with the slope and the intercept. These values are tabulated as follows,
2 !y " x 2i #

!b =

(9) (10)

!m =
where
2 !y =

N! 2 y "
1 N ( yi " b " mx i ) 2 # N - 2 i =1

(11)

Students rarely need to perform these rigorous calculations, most graphing programs and calculators perform Least-Square fits and provide values for the slope, intercept, and their respective errors. Similar formulas likewise exist for polynomial and exponential fits. The reader is referred to An Introduction to Error Analysis by John R. Taylor for information on error analysis associated with higher-order fitting protocols. C. Counting Experiments Numerous instruments, such as a mass-spectrometer, function by counting over a fixed period of time the number of events detected at a certain setting. A resultant experimental spectra can then be created by plotting the number of counts detected at a single setting, the peak, as a function of setting value. Experimental values derived from a spectra based on counting experiments are governed by the statistics of Poisson Distributions, and the reader is again directed to An Introduction to Error Analysis by John R. Taylor for detailed information on Poisson Distributions. However, the error associated with a peak is remarkably simple and can be shown to vary as the square root of the total number of counts. Thus, if our mass spectrometer records over a period of time counts for a certain mass setting, then the error in , , is equal to ! . Hence, the longer the counting time, the larger is and the smaller the error in becomes. D. A Single Measurement and/or a Tabulated Value When the measurement is made only once, we must make an estimate of the probable error based on the features of the measuring device. For example, if a buret has each division equal to 0.1ml and the eye can subdivide the interval between two division into fifths, or 0.2, we get a possible total error equal to 0.02 (0.20.1) or a error equal to half this value, or 0.01; so we record the volume as equal to the reading 0.01 and set V equal to 0.01. Finally, if the number does not come from your measurement but is obtained from a reference book such as the CRC Handbook, then one (1) in the next to last digit is the maximum error you could expect and a reasonable assumption would be to take 1/2 of this as the probable error. Thus, if the CRC Handbook gives the Density of NaCl a value equal to 2.165, we would take Density equal to 0.005 and use 2.165 0.005 to report the value for the density of NaCl. As another example, you may need to get the density of water from the CRC Handbook but you are not sure of the temperature you had during your experiment. Now, the uncertainty in

the temperature together with the way in which the density of water varies with temperature determined the probable error in the density rather than the accuracy of the tabulated values, which have a much smaller probable error. In the subsequent section, we will discuss how to handle or propagate such compound errors. II. Propagation of Errors Assume now that we have values x , y , z , . . . for all the quantities x, y, z, . . . which enter into the calculation of a desired quantity f, f = f(x, y, z, . . .), together with their probable errors, x , y , z , . . .. Then the best value of f is given by
f = f( x , y , z , . . .)

A. Background Theory

and not by the average of all the different fs calculated from the different xs, ys, zs etc.. The question now is: What is the probable error associated with f due to the probable errors of x , y ,

z , . . .? In other words, knowing that x can be vary between x x , y can be vary between y
y , z can be vary between z z , by what extent does f change as x , y , and z change? The answer is

!f =

2 2 # # "f % "f % # "f % 2 2 ' !x + !y + ! z2 + . . . $ "x & x , y , z... $ "z & x , y , z ... $ "y & x , y , z...

(12)

So we can report the value of f as f = f f Equation 12 requires many variables. Thus, it is helpful to present the calculated values that enter Equation 12 in tabular form so that we can see the individual contributions to the total probable error, thus singling out the chief source(s) of error for the given experimental method. For our benefit, in the subsequent sections we will develop practical solutions to Equation 12.
Table 2: A tabular view of the variables and their associated error.

U x y z

!f !U -

U -

" !f $ 2 & # !U % U 2

# "f % 2 ! = ' ! $ "U & U

2 2 f

B. Propagation Rule for Addition and Subtraction Suppose the functional relation of f to x, y, and z involves only addition and/or subtraction, for example, f = x + y z with coefficients of unit magnitude. Then either 1 since

!f will be !U

!f !f !f = 1, = 1, and = -1, and we can reduce Equation 12 to a result that !y !x !z

2 2 !2 x + !y + ! z

should be familiar to most,

!f =

(13)

So, for addition and subtraction, the probable error of f is the square root of the sum of the squares of the probable errors of x, y, and z, rather than simply the sum of the probable errors. The latter would give the worst possibility, corresponding to the situation where all the errors are in the same direction. The correct formula involving the Pythagorean type relationship allows for the possibility that some of the errors might be in opposite directions (in the case of addition) or in the same direction (in the case of subtraction) thereby canceling out. Example: What is the probable error in weighing a sample? If we have two weighings say of a flask, once empty, wE = 64.2793 0.0002 g, and then filled with sample, wF = 82.9624 0.0002 g, where the estimated probable errors are taken to be 0.0002 g based on experience acquired with balances which can be read to 0.1 mg, we would say the weight of sample, WS, obtained is given by WS = WF WE = 18.6831 0.0003 g since

! WS =

(0.0002) + (0.0002) = 2.810-4.

2 2

Note, however, that if the coefficients are not of unit magnitude, for example, if f = x + 3y z, we get instead

!f =

2 2 !2 x + 9! y + ! z

(14)

thus showing the greater importance on an error in y. In general, for f = ax + by cz, we can reduce Equation 12 to get

!f =

a !x + b !y + c !z
2 2 2 2 2

(15)

C. Propagation Rule for Multiplication and Division Suppose the functional relation of f to x, y, and z involves only multiplication and/or division with exponents of unit magnitude, for example, f =

xy . Instead of the above result for z

the addition and subtraction case where the errors are related, we find here that it is the relative errors which are related by the Pythagorean type formula since example for f, so

!f f = . Namely, for the above !U U

" !f $ y f ; = = # !x % x , y, z... z x

" " !f $ xy f x f $ !f % ; and = = = - 2 = - , # !z % x , y , z... # !y & x, y, z... z y z z

!f =
or

" " " $f% $f% $ f% !2 !2 ! z2 x + y + # x & x , y , z... # y & x , y , z... # z & x, y, z...

(16)

!f = f

" !x $ # x%

"! $ + & y # y%

"! + # z$ z%

(17)

So, for multiplication and division, the relative probable error of f is the square root of the sum of the squares of the relative probable errors of x , y , and z . Note, however, that if the exponents are not of unit magnitude, for example, if f = then

x 4y , z

" !f $ 4x 3 y 4f and we get = = # !x % x , y, z... z x

!f = f

"! 16 x $ # x %

"! $ + & y # y %

"!z $ # z%

(18)

thus showing the greater importance on an error in x. In general, for f = Equation 12 to get

x a yb , we can reduce zc

!f = f

"! a # x$ x %
2

"! $ + b & y # y%
2

"! + c # z$ z%
2

Both Equations 13 and 17 are generalizations derived from Equation 12. Equation 12 is the work horse for error analysis calculations and should be referred to when complicated functions, and hence derivatives, are encountered.

An Illustrative General Example of Propagation of Errors A. The Experiment: Determination of the volume of a pycnometer by weighing the water it holds and observing the temperature. To allow for the buoyancy effect of weighing in air, we have

" d d % mtrue = m app $1 + air ! air ' # d water dWTS. &

where mapp is the apparent mass read from the scale and d stands for density. Therefore,

Vpyc =

m true d water

" d d % m app$ 1 + air ! air ' dwater d WTS. & # = d water

B. The Data: We have the following data with the given probable errors.
Table 3: Data for Determination of the Volume of a Pycnometer.

Temperature = 24.3 0.4 C mapp = 24.9526 0.0003 g dwater = 0.9972 0.0001 g ml-1

Barometer reading = 740.0 0.2 mmHg dair = 0.001156 0.000002 g ml-1 dWTS.(brass) = 8.5 0.5 g ml-1

C. The Errors: The above listed probable errors were arrived at as follows: Temp.: Based on scale on thermometer and observation of variation in readings during the course of the experiment. Bar.: Based on vernier scale and reliability of location of meniscus. mapp: Based on estimate of 0.0002 g probable error of single weighing and propagation of this error when taking the difference of two weighings to get the weight of water. dair: Based on tables together with errors in temperature and barometer reading. dwater: Based on tables together with error in temperature dWTS.: Based on table value only given to the tenth, so use 5 times this as the error. D. The Value: Calculating the best value for the volume of the pycnometer, we have
Vpyc = $1 + (24.9526)" # 0.001156 0.001156 % ' ! 0.9972 8.5 & = 25.048 ml 0.9972

E. The Error in Value: Now we need to calculate the probable error of this value, Vpyc, In accordance with Equation 12,
2 !V = pyc

# "Vpyc & 2 ( ! )% $ "U ' U

2

To determine Vpyc, requires we take the partial derivative of Vpyc with every variable used in determining Vpyc.
" !Vpyc % Vpyc 25.048 = 1.0038 ml g-1. = $ ' = ! m m 24.9526 # app & app

" !Vpyc % " m app % " 1 1 % " 24.9526 % " 1 1 % '$ ' = 22.149 ml2 g-1. ( ( $ ' = $ ' $ '= $ # !d air & # dwater & # dwater d WTS. & # 0.9972 & # 0.9972 8.5& " !Vpyc % " m app % " -d air % Vpyc " -0.001156 % 25.048 " 24.9526 % ' $ = -25.15 ml2 g-1. = $ $ ' = $ '$ 2 '( 2 ' ( # & 0.9972 # (0.9972) & 0.9972 # !d water & # d water & # d water & d water. " !Vpyc % " m app % " dair % " 0.001156 % " 24.9526 % 2 -1 ' $ $ ' = $ ' $ 2 ' = $ ' = 0.00040 ml g . 2 # & ( ) ! d d d 0.9972 8.5 # WTS & # water & # WTS & # &
In Tabular form, we have
Table 4: A tabular view of the variables and their associated error.

U mapp dair dwater dWTS

!
2 Vpyc

!Vpyc !U

U 310-4 g 210-6 g ml-1 110-4 g ml-1 510-1 g ml-1 -

1.0038 ml g-1 22.149 ml2 g-1 25.15 ml2 g-1 0.00040 ml2 g-1
2

" !Vpyc % 2 $ ' ( # !U & U 0.110-6 ml2 0.010-6 ml2 6.3 10-6 ml2 0.010-6 ml2
2

# "V & = ) % pyc ( ! 2 $ "U ' U

6.410-6 ml2

F. The Reported Value: Thus, V = 2.510-3 ml and therefore, Vpyc = 25.048 0.003 ml. The greatest source of error, by far, was due to the uncertainty in dwater which in turn came about because of the uncertainty of temperature (the density of water is known very accurately for a given temperature). Thus, the way to improve the experiment and reduce the probable error

in the result is to exercise better temperature control and measure temperature with a better thermometer. If the probable error in the temperature of the water were reduced from the above value of 0.4C to 0.1C, this would reduce the probable error in dwater from 0.0001 to 0.00003,
" !Vpyc % 2 ' ( giving $ = 0.610-6 ml2 and Vpyc = 810-4 ml, so that we would have # !d water & d water
2

Vpyc = 25.048 0.001 ml. The experimenter would have to decide whether this improvement was of sufficient importance to warrant the extra trouble and/or expense of obtaining more accurate temperature data.