Вы находитесь на странице: 1из 20

DATA REDUCTION AND ERROR ANALYSIS FOR THE ION PROBE

Class notes for GG710, given by Gary R. Huss, Fall, 2015


This document summarizes the statistical tools that are used to evaluate data collected by
the ion microprobe. The presentation here is intended as a practical guide and does not include
all of the background and derivations that you will find in a good statistics book. I will try to
convey the concepts that are directly relevant to measurements using the ion microprobe. If you
are interested in studying statistics in more depth, I recommend Data Reduction and Error
Analysis for the Physical Sciences, by Bevington and Robinson (McGraw Hill).
Random and Systematic Errors
It is not possible to exactly measure the composition of a material by the ion probe. We
do not count all of the atoms, so we must infer the composition from the sample of atoms that we
do count. There are many reasons for this. First, only about 1 in 1000 to 10,000 of the atoms
sputtered from the sample make it into the detector. This is because only a small fraction of the
sputtered material comes off the sample as an atomic ion in the right charge state, and only a
fraction of those ions make it through the mass spectrometer to the detector. Many atoms come
off the sample as molecules, neutral atoms, or ions of charge other than the +1 or -1 that we
generally measures, and many of the ions of correct charge are blocked in the mass spectrometer
by the slits and apertures we use to tune the mass spectrometer. We assume that the fraction of
appropriately charged ions that reach the detector is representative of the atoms in the sample, or
that any differences between them, such as that caused by mass-dependent instrumental
fractionation, can be understood and corrected for.
Measurement uncertainties are typically called errors. In this usage, the error is the
difference between an observed value and the true value. But because we do not know the true
value, it is necessary to determine in a systematic way the likely magnitude of the difference
between our measurement and the true value. If reported correctly, the smaller the uncertainty,
or error, the better we know the true value. I will not discuss errors that are caused by
mistakes in the experiment or in calculation of the result. These clearly should be avoided, and it
is up to the experimenter to devise tappropriate safeguards against these mistakes. Here I discuss
the possible and unknown differences between our measured value and the true value that are
inherent in the measurement process. Errors that are reported with ion probe data (or any other
measurement data) give a quantitative assessment of how large this difference is likely to be.
Random errors can be thought of as fluctuations in the observed results from
measurement to measurement that require the analyst to make repeated measurements to yield a
precise result. Counting errors are random errors that derive solely from the fact that we have
not measured every atom in the sample, and the subset of atoms that we do measure may not
have exactly the composition of the whole. Counting errors describe how well the composition
of the ions arriving in the detector represents the true composition. If we have done everything
correctly, counting errors should dominate the uncertainty in our measurements. There are other
types of random errors that can be treated in much the same way as counting errors. For
example, if several points on a sample or standard are measured, and the degree of sample
charging is different for each spot, the variations in instrumental mass fractionation tend to be
random around the typical mass fractionation value.

Systematic errors can be of several types. There are systematic errors associated with
calibration of detectors and in determining the level of mass-dependent fractionation produced in
the mass spectrometer, with variations in tuning of the instrument, and with differences in the
fraction of each element that is in the correct ionization state after sputtering, among other
things. It is necessary for the analyst to consider all of these potential sources of error and
evaluate them appropriately. There is little that is more embarrassing and that can affect
someones scientific career more than to publish an exciting experimental result and then have to
retract it later when you discover that the result was wrong, particularly when the mistake is
something that should have been identified during the data analysis. The only thing worse is that
someone else discovers the error and publishes the correct result.
You have heard the terms precision and accuracy with regard to experimental data.
Counting statistical errors describe the precision of a measurement, which should be thought of
as the reproducibility of the measurement. This is the degree to which further measurements will
show the same result. On the other hand, the accuracy of a measurement is the degree to which
the measurement accurately represents the true state of the measured system. A measurement
can be highly precise and can be repeated to a very high degree of precision, and yet be
inaccurate because of a systematic error introduced by the experiment itself. A result is
considered to be valid if it is both accurate and precise. In general, the reported error or
uncertainty on a result should reflect the degree to which the experimenter believes the result to
be valid. In other words, the error should describe random uncertainties, usually dominated by
statistical errors, and an estimate of any systematic uncertainties introduced by the measurement.
In this document, I attempt to provide a comprehensive discussion of statistical and other
random errors. Systematic errors will be discussed, but in a much less comprehensive way.
Poisson and Gaussian statistics
The primary statistical tools for analyzing random errors in counting experiments are
probability distributions. These distributions allow you to estimate the most probable value for a
parameter. There are several different probability distributions. The Binomial distribution gives
you the probability of getting exactly k successes in a sequence of n independent yes/no
experiments, each having a probability p of success.

"
%
n!
nk
'' p k (1 p)
Pr ( X = k ) = $$
# k!( n k )! &

(1)

for k = 0, 1, 2, , n
The Binomial distribution describes coin-toss experiments, for example.
The Poisson distribution is more relevant to our problem. It is used for counting
experiments where the data represent the number of items or events observed per unit time. It is
important for studying random processes. The Poisson distribution gives the probability that a
given number of events (k) occur in a fixed time interval if the events occur at a known average
rate () and are independent of the time since the last event. The Poisson distribution is only
defined for integer values of k.
Pr ( X = k ) =

x
e
k!

(2)

As the number of trials, n, approaches infinity while np remains fixed (n is large, p is small), the
Binomial distribution approaches the Poisson distribution with = np. Figure 1 shows some
examples of Poisson distributions for cases where is small. As gets larger, the Poisson
distribution approaches the Gaussian, or Normal distribution.

Figure 1: Poisson distributions for l = 1, 4, and 10.


The Central Limit Theorem is one of the most profound and important theorems in
mathematics. The Theorem states: Given certain conditions, the arithmetic mean of a
sufficiently large number of iterates of an independent random variable, each with a well-defined
expectation value and well-defined variance, will be approximately normally distributed
regardless of the underlying distribution (wording from Wikipedia). This is important because
each of our measurement cycles constitutes an iterate of an independent random variable, each
of which samples the parent population in a random way. The distribution of the means of these
iterates forms a normal distribution, the mean of which approximates the value of the parent
population. Mathematically, this Normal distribution can be described as: The average of n
mutually independent, identically distributed random variables (with finite measured values of
the mean 0 and variance 02) approaches the distribution:

f ( x, , ) =

( x )

1
2

(3)

with = 0 and 2 = 02/n


Figure 2 shows some examples of Normal or Gaussian distributions. This distribution is
the limiting form of the Poisson distribution as the event rate becomes large. It is the limiting
form of the Binomial distribution as the number of events n becomes large. It is the convergent
distribution of the Central Limit Theorem. And it is the most likely distribution for our
experiments.

Figure 2: Examples of Normal or Gaussian distributions.


Properties of Probability Distributions
The shape of a probability distribution can be described by four statistical moments. The
first moment is the Expectation Value ( = E[f]), or mean. The second moment is the
Variance, given by 2 = E[(f-)2]. The variance describes the width of the distribution. The
third moment is Skewness, given by 1 = E[((f-)/3]. Skewness describes the degree of
asymmetry of the distribution. The fourth moment is Kurtosis, given by 2 = E[((f-)/4].
Kurtosis describes the pointiness of the distribution. A Gaussian distribution has zero
skewness and zero excess kurtosis. It is described entirely by the mean () and the variance (2).
Figure 3 shows graphically the relationship of three parameters used to describe a Normal
or Gaussian distribution. The mean, , is the peak of the distribution. The width of the curve is
determined by the value of , such that for x = + , the height of the curve is reduced to e-1/2
of its value at the peak (Figure 1). The full-width at half maximium (), also called the halfwidth, is given by:
= 2.354

(5)

Figure 1: Gaussian probability distribution illustrating the relationship of , ,


and to the probability envelope. The probability that the measurement falls
within one standard deviation (1) of the mean is ~68%. A measurement will fall
with two standard deviations (2) of the mean ~95% of the time.
The standard form of the Gaussian equation is obtained by defining the dimensionless
variable z = (x- )/ . This frees the distribution from any particular value of to give:
PG ( z) dz =

$ z2 '
1
exp& ) dz
2
% 2(

(6)

This equation permits a single computer routine or table of values to give the Gaussian
probability function PG(x; , ) for all values of the parameters and by changing the variable

and scaling the function by 1/ to preserve the normalization.


Parameter Estimation
Ion probe data should be considered to be measurements that sample a large parent
population. We wish to use our data to estimate the characteristics of the parent population, and
to estimate how well we know those characteristics (the uncertainties). We will assume we are
sampling from a Gaussian parent population with mean 0 and standard deviation 0. We do not
know the value of 0 and we wish to estimate it from our data. Let us define our estimate of the
population mean, 0, to be m. We hypothesis that our data follow a normal distribution with
parameters and (equation 3). If we perform a set of n measurements, all with the same
standard deviation, how do we estimate the most accurate ?
We will use the Method of Maximum Likelihood. The probability of observing all n
measurements is the joint proability of x1, x2, . . ., xn. This is simply the product of all of Pis
(equation 3):
n
" 1 n ( x )2 %
" 1 %
i
'
Pall = Pi = $
' exp $$
2
'
2
2

# 2 &
i=1
# i=1
&
n

(7)

We seek the m that maximizes this probability, assuming constant . The method of maximum
likelihood says that this quantity is maximized at = 0. To maximize Pall, we just maximize
the argument of the exponential. To do this, we set the derivative of this argument, with respect
to , to zero.

d " 1 n " xi %
$ $
'
d $# 2 i=1 # &

d " 1 n " xi %
$ $
'
d $# 2 i=1 # &

%
'' = 0
&

(8a)

2
n
%
" x %
1 n d " x %
'' = $ i
=
' $ i
'=0
#
&
#
&
2
du

&
i=1
i=1

(8b)

since we assumed s is constant:

1 n
= xi = x
n i=1

(9)

The most likely value of the population mean is just the average of the data!
What is the uncertainty on our estimation of ? The variance of a normal distribution is
2
. How can we calculate the variance? Consider x, a function of u, v, . . . [x = f(u,v, . . .)]. The
variance of x can be expressed as a function of the variance of u, v, . . . and partial derivatives by
expanding f in a Taylor series:
2

# x &
# x &
# x &# x &
% ( + v2 % ( +... + 2 uv2 % (% ( +...
$ u '
$ v '
$ u '$ v '
2
x

2
u

(10a)

so
" " %2 %

= $ x2i $ ' '


$ # x & '
i
i=1 #
&

(10b)

#1 n & 1
=
% xi ( =
xi xi $ n i=1 ' n

(11)

With all xi = , this reduces to:

! ! 1 $2 $ 2
= ## 2 # & && =
"n% % n
i=1 "
n

(12)

This is the uncertainty on our estimation of the population mean, the parameter we computed
from our data. But what is ? We dont know this a priori. We can construct an estimator of
the population standard deviation from our sample: the sample standard deviation, s:

1 n
s=
xi x
n 1 i=1

(13)

Substituting into equation 12 and taking the square root gives the uncertainty in our
estimation of , the mean of the population:

s
n

(14)

Our estimate of the variance of the population is given the variance of our data:
1 n
s =
xi x
n 1 i=1
2

(15)

The uncertainty in our estimate of the variance (given here without derivation) is:

" 2
%
s2 = 4 $
+ '
# n 1 n &

(16)

where is the kurtosis, which is zero for a Gaussian distribution.


In summary, for a normal distribution consisting of measurements with a constant
standard deviation, we have calculated an estimator of the mean of the population (Equation 9)
and for the uncertainty in the estimation of the mean (Equation 14). We have calculated the
standard deviation of our data and from that estimated the variance of the population (Equation
15). One can also estimate the uncertainty in our estimate of the variance (Equation 16). Now I
will describe how to use these tools to understand ion probe data.

Figure 4: Example of a data set and the distribution inferred from it, giving
estimates of the mean and standard deviation of the parent population.
Propagation of errors
So far, we have discussed the probability distribution produced by several measurements
of a single quantity. However, we often must deal with several measured quantities, such as a
measured count rate, a background correction, and an instrumental fractionation correction, each
with its own error. To calculate an isotope ratio, we must combine the uncertainties in the
measured, background-corrected count rates for numerator and denominator into an error for the
ratio. These various errors can be combined using some standard formulae. Derivations of these
formulae can be found in Bevington and Robinson or another statistics book. They are based on
the error propagation equation, which, for the situation where x = f( , ,. . .) gives the variance,
2 for x as:
2

2
" dx %
" %
" dx %
2 " dx % dx
$ ' + 2 $ ' +... + 2
$ '$ ' +...
# d &
# d &# d &
# d &
2
x

(17)

If the errors on and are uncorrelated, the cross terms in this equation vanish in the limit of a
large random selection of observations and the error equation reduces to:

2
" dx %
" dx %
$ ' + 2 $ ' +...
# d &
# d &
2
x

(18)

with similar terms for additional variables. We will return to the case where errors are correlated
later in this discussion. Below I give the specific formulas for propagating uncorrelated errors in
several standard cases.
Sums and differences
If the result, x, is related to a measured value by the relationship

(19)

x = +b

where b is a constant, then partial derivative x/ = 1 and the uncertainty in x is just:

(20)

x =
The relative uncertainty, also known as the fractional error, is given by

=
=
x
x +b

(21)

If x is the weighted sum of and such that

(22)

x = b + c

then the partial derivatives are constants

$ x '
& )=b
% (

$ x '
& )=c
% (

(23)

The variance and standard deviation are:

2
x2 = b 2 2 + c 2 2 + 2bc

2
= b 2 2 + c 2 2 + 2bc

(24)

For uncorrelated errors, cross terms in equation (17) vanish and

= b 2 2 + c 2 2

(25)

Multiplication and division


Equation (17) also applies for multiplication and division. For multiplication, where
x = b

(26)

the partial derivatives are


$ x '
& ) = b
% (

$ x '
& ) = b
% (

(27)

and the variance of x is

2
2
x2 = (b ) + (b ) + 2b 2

(28)

A more useful expression is given by the equation for relative variance, which can be derived
from equation (18) by dividing both sides by x2
2
2

x2 2
=
+
+2
x2 2 2

(29)

The standard deviation for x:

x = x

2
2

2
+
+
2
2 2

(30)

As with equation 25, the cross term becomes zero in the case where the variances in and are
uncorrelated (covariance = 0). The standard deviation becomes:
2
2
x = x 2 + 2

(31)

In our application, we will be calculating isotope ratios and elemental abundance ratios.
Uncertainties in the numerator and denominator must be combined to give the uncertainty in the
ratio. Thus, for division, if r is the weighted ratio of and :
r=

(32)

The partial derivatives are

$ r ' b
& )=
% (

$ r '
b
& )= 2

% (

(33)

and the relative variance for r and the standard deviation are given by

2
2

r2 2
=
+
2
r2 2 2

2
2

2
r = r 2 + 2 2

(34)

where the last term becomes zero when the errors are uncorrelated.
In summary, when doing addition or subtraction, the errors are determined by calculating
the sum of the squares of the absolute errors and then taking the square root. When you are
doing multiplication or division, the errors are determined by calculating the sum of the squares
of the relative errors, taking the square root, and then multiplying the result by the value of the
product or quotient.
Application to ion probe measurements
How does the above formalism apply to counted ions, the output of the ion microprobe?
When the mean number of counted ions is N during a given time t, it can be said to be the result
of repeated N*M tests for which the probability of counting one ion is 1/M. In an ion counting
experiment, the uncertainty of a measurement is equal to the square root of the number of counts
measured ( = N ).
Suppose that we want to calculate the uncertainty in a ratio R defined as:

R=

S1
S2

(35)

where S1 and S2 are count rates expressed in counts per second and are derived from two
numbers of counts, N1 and N2, collected over time intervals T1 and T2 such that
S1 = N1/T1 and S2 = N2/T2

(36)

The counting-statistical uncertainty in the ratio can be expressed as the ratio times the quadratic
sum of the relative uncertainties in the numerator and denominator as determined by counting
statistics (equation 23):

R = R

N1

( N1)

) +(

N2

(N 2 )

= R

T1S1

(T1S1)

) +(

T2 S2

(T2 S2 )

(37)

The following examples illustrate the use of Gaussian statistics to determine uncertainties
in data from the ion microprobe.
Example 1: We measure carbon isotopes, 12C and 13C. The measured counts and
uncertainties are: 12C = 125390 354 and 13C = 1420 38. The ratio, its variance, and
its 1 standard deviation are:
13
12

C
1420
=
= 0.011325
C 125390

! 2 2 $
! 382
354 2 $
= r ## 2 + 2 && = 0.113252 #
+
= 0.00000929
2
2&
%
" 1420 125390 %
"
2
r

10

*
= 0.011325 ,
,+

$ 38 2
354 2 ' /
+
= 0.000305
&
)
2
2
% 1420 125390 ( /.

Example 2: We measure carbon isotopes, 12C and 13C. The measured count rates are:
12
C = 125390 cps and 13C = 1420 cps and the counting times are T12 = 10 and T13 = 100
seconds. The ratio is calculated using count rates, the same as in Example 1:
13
12

C
1420
=
= 0.011325
C 125390

But for the uncertainty, we must consider the total number of counts measured for each
isotope, not simply the count rate. Total counts for 12C = 125390 10 = 1253900 and for
13
C = 1420 100 = 142000. The variance and 1 standard deviation are:

! 2 2 $
! 3772
1120 2 $
2

= r ## 2 + 2 && = 0.11325 #
+
= 0.000000101
2
2&
%
" 142000 1253900 %
"
2
r

( "
+
3772
1120 2 % *
= 0.011325 $
+
' = 0.0000317
*) # 142000 2 1253900 2 & -,

Because the 13C counts provide the main contribution to the error, increasing the total
number of 13C counts by a factor of 100 relative to Example 1 results in a factor of 10
improvement in the uncertainty for the ratio.
Example 3: Suppose we are using an electron multiplier that has a background of 0.5
counts per second. We determined this background by measuring it for 10 seconds,
where we got 5 total counts. The background is then (5/10) (sqrt(5)/10) = 0.5 0.22
counts per second. We measure a signal for four seconds and find that the count rate is
23.0 counts per second. The uncertainty in our measurement is sqrt(23*4)/4, so the
measurement is 23.0 2.4 counts per second. What is the count rate and uncertainty
after correcting for the background?
Count rate = 23 0.5 = 22.50

(2.4

+ 0.22 2 ) = 2.41

Note that the uncertainty in the background correction, which has an absolute magnitude
less than 10% of the uncertainty in the measurement, plays almost no role in the final

uncertainty.
Example 4: Suppose that we have a measured count rate of 1.2 counts per second,
measured for 10 seconds (1.2 0.35), and a background of 0.5 counts per second
measured for 5 seconds (0.500.31).

11

Count rate = 1.20 0.50 = 0.70

(0.35

+ 0.312 ) = 0.47

In this case, the two uncertainties contribute roughly equally to the total uncertainty.
Problems with low count rates
Low count rates can cause problems in isotope ratio measurements if counting times are
not sufficiently long. Consider a case where the average count rate is 0.3 counts per second. If a
counting time of 1 second per cycle is used, only one out of ~3 cycles will record a count, on
average. With enough cycles, one can calculate the average count rate reasonably accurately.
But suppose I want to calculate ratios for every cycle using these data. In the 0.3 cps mass is in
the numerator, the ratio will be zero two-thirds of the time. If it is in the denominator, the ratio
becomes undefined two-thirds of the time. Neither of these cases is handled well by the
techniques described above. The undefined ratios result in a completely unreasonable ratio that
is easy to spot. However, the zeros are equally bad and more insidious when calculating a ratio,
because one is typically making ratios with one mass that has an asymmetrical distribution about
the mean (Fig. 5) and a another that has a symmetrical distribution (Fig. 6). The result is that the
ratio appears reasonable but is too high. Thus, it is necessary to choose counting times such that
every cycle has at least one count, so all ratios are defined and non-zero. Alternatively, one can
recompute the counting interval by combining cycles into blocks such that each block has at
least one count. We will return to the issue of low counts later in the course.

Figure 5: Histogram of counts in a cosmic ray detector. The experimenter recorded


the number of counts in their detector for a series of 100 2-second measurements. The
Poisson distribution, shown as a smooth curve, is an estimate of the parent
distribution based on the measured mean, 1.69 counts per 2-second interval. Note
that the distribution is not symmetrical because one cannot measure negative counts.

12

Figure 6: The same experiment was performed as that to generate Fig. 5 except that
60 measurements of 15 seconds each were made. The Poisson distribution is again
shown as a smooth curve. The mean in this case was 11.48 counts per 15-second
interval and the distribution is symmetrical.
In the two cases shown in the Figures above, the mean count rate and standard deviation
are the same. Poisson statistics correctly handles the asymmetrical distribution of counts per
measurement interval and the final results of the two experiments are statistically equivalent ( =
0.85, = 0.65 for the first experiment; = 0.77, = 0.23 for the second). The problem only
becomes evident if ratios are calculated using a distribution like in Figure 5.
Combining several measurements into a single result
An ion microprobe is best described as an isotope-ratio mass spectrometer. The data are
collected as a series of ratios, with the most abundant isotope typically serving as the
denominator. This method of collecting data permits us to recognize and correct for changes in
signal strength with time and to identify noise spikes or other problems with the data. However,
the final result we are after is a single value for each isotope ratio and its uncertainty.
Suppose we make 100 measurements of an isotope ratio and calculate the statistical
uncertainty for each ratio as described in the previous section. How can we combine these data
into a single measurement? The distribution of these measurements will be Gaussian (assuming
that there are no non-statistical uncertainties), characterized by a mean and standard deviation
. If the 100 ratios, xi, are all based on a similar number of counts, and thus have the same
uncertainty, i, then the mean is given by

= x

1
xi
N

(38)

13

and the uncertainty is given by

* $ '2
= , i2 & ) /
,+ % x i ( /.

(39)

Because the uncertainties on the individual ratios are all equal in this case, the partial derivatives
in equation (39) are

' 1
$ 1
=
& xi ) =
( N
x i x i % N

(40)

Combining equations (39) and (40), we find

) # 1 &2,
= + i2 % ( . = i
N
* $N' -

(41)

When calculated this way, is referred to as the standard deviation of the mean or the standard
error. If the scatter of the individual ratio measurements is due entirely to statistical effects, the
uncertainty calculated from equation (41) will be the same as that calculated from the total
counts from all of the ratios treated as a single measurement. We use this equivalency as a test
of how good our data is. The output from both the instrument software and our off-line data
reduction program gives the ratio of the uncertainty calculated from the spread of the individual
measurements for each ratio divided by that expected uncertainty from the total number of
counts collected. For good ion probe measurements, this ratio is very close to unity.
Often we find it necessary to combine data from several measurements that were
collected with different counting rates or with different numbers of ratios. The uncertainties of
the measurements are thus different. I will not attempt to derive the equations here (you can see
the derivation in Bevington and Robinson). I will just present and discuss them. The equation
for the mean of several values with different uncertainties is:

( x )
(1 )
i

2
i

(42)

2
i

In this equation, each data point xi in the sum is weighted inversely by its own variance. So a
measurement with a large error will contribute less to the mean than a measurement with a small
error. The uncertainty in the mean can be determined from the by evaluating x from
i
equation (42) for the mean :


=
x i x i

( x ) = 1
(1 ) 1
i

2
i

2
i

2
i

2
i

(43)

Substituting this into equation (39) gives:

14

1 i2

[ (1 )]
2
i

1
(1 i2 )

(44)

The error-weighted mean and its associated standard deviation of the mean (standard error) are
used extensively in ion-probe mass spectrometry and we have and Excel macro programmed to
make them easy to calculate.
When to use the standard deviation of the mean and when to use the standard deviation
The last section showed that the standard deviation of the mean or standard error of a set
of measurements is given by the standard deviation of the measurements divided by the squareroot of the number of measurements. This is true if the errors of the measurements are a truly
random distribution from a single population. Increasing the number of measurements of the
isotope ratios in a sample will rapidly improve the precision to which we know the mean of the
measurements. However, the standard error is not always the correct statistic to use when
analyzing ion probe data. Remember that the standard error asks the statistical question: How
well do I know the mean of the group of measurements that I have made and how closely is it
likely to represent the true value for the system being measured? In contrast, the standard
deviation asks the question: What is the statistical scatter of measurements around the mean
value and how close is my next measurement likely to be to the mean value? This uncertainty
does not decrease with an increasing number of measurements.
As an example, consider a set of ten measurements xi of an isotope ratio. Our estimate of
the mean of these measurements has a value x, and the uncertainty associated with that mean
describing how close it is to the true value of the system is 10 . However, the likelihood
that the next measurement will fall within one sigma of the mean, x, is given by the standard
deviation, .
It is not sufficient to keep measuring a sample until the standard deviation of the mean

becomes arbitrarily small. At some point something other than statistics will control the error on
the measurement. Consider the case of a typical ion probe measurement. In order to correct for
instrumental effects, we measure standards. To determine the magnitude of the instrumental
effect precisely, we measure the standards many times. If the variation of the standards is truly
random, then we can determine the value of, for example, the mass-dependent instrumental
fractionation, f, very precisely. The standard error of our set of standard measurements (
f N f ) tells us how well we know the instrumental fractionation of the instrument, and the

standard deviation of the measurements ( f) tells us the likelihood that the next measurement
will be within one sigma of the measured instrumental fractionation.
We then measure a sample of unknown composition many times and obtain and value, r,
for an isotope ratio and an uncertainty ( r N r ). The more we measure, the better determined
our mean value, r, becomes. This ratio must be corrected for the mass-dependent instrumental
fractionation and the uncertainties of our measurement and of the correction must be combined
appropriately. The corrected value and its uncertainty are:

15

R=r f

# &2 #
R = %% r (( + %% f
$ Nr ' $ N f

&2
(
(
'

(45)

But suppose that we cannot measure the unknown sample many times and are dependent
on a single measurement to determine the true value. What is the uncertainty in our
determination? Theuncertainty in the measured ratio, r, must be combined with the uncertainty
for the mass fraction, as in the previous example. However, in this case, because the sample was
only measured once, the uncertainty of the mass fractionation associated with our measurement
is the uncertainty associated with an individual measurement of the standard. You dont know if
this measurement is close to the mean or far from the mean and you cannot evaluate it
independently because the true ratio is also unknown. So the measurement error must be
combined with the measurement error for a single standard measurement, the standard deviation,
f. We have to ask the statistical question: How likely is the fractionation for this measurement
to be within one sigma of the fractionation value, f, that we have determined for the standard? In
this case, it doesnt matter how well we may have determined f through independent
measurements (except that any systematic error is reduced). The statistical error is the quadratic
sum of the measurement error and the standard deviation of the measurements on the standard

R = r2 + 2f

(46)

If we must apply the uncertainty associated with a single measurement of the


instrumental mass fractionation as the uncertainty on the mass fractionation correction, why
collect a large number of measurements of the standards? The standard deviation does not
decrease with the number of measurements, but the correction does become better determined.
We are interested in more than just the precision of our result. We are also interested in the
accuracy. If we were to use a single measurement of the standard, it might be an outlier in a
distribution obtained by a large number of measurements. Using this value to correct for
instrumental fractionation would add a systematic error to all of the measurements equivalent to
the differences between the true value of the instrumental fractionation and the value that we
inferred from our single measurement. The better we know the value of the instrumental
fractionation, the smaller are any potential systematic errors that may be introduced by the
correction.
In summary, the more measurements of a sample or standard you make, the better
determined the mean value is. The standard error will improve with additional measurements
by a factor of N . The statistical uncertainty associated with a correction of a single
measurement for instrumental mass fractionation, detector background, detector sensitivity,
deadtime, etc is given by the standard deviation of the measurements that were made to
determine the value of the correction. The level of the potential systematic error associated
correction is given by the standard error calculated from the same set of measurements.
with the
The 2 statistic: How well do the data follow Gaussian statistics?
In the discussions above, we have assumed that the data can be fit by a Gaussian (or
Poisson) distribution and thus can be described by the statistics based on that distribution.
However, in the real world, our data often have variations caused by something other than

16

random statistical variation. Sometimes it is clear that this is the case and sometimes it is not. It
would be nice to have a statistic that could help us to recognize non-Gaussian distributions. That
statistic is called chi-squared ( 2).
Assume that we make N measurements xi of the quantity x. There are j possible values
for xi ranging from 1 to n. We define the frequency of observations for each possible value xj as
h(xj). If the probability for observing the value xj in any random measurements is denoted by
P(xj), then the expected number of such observations is y(xj) = NP(xj), where N is the total
number of measurements. Figures 7 and 8 show the same six-bin histogram, drawn from a
Gaussian parent distribution with a mean = 5.0 and standard deviation = 1, corresponding to
100 total measurements. The parent distribution is illustrated by the solid Gaussian curve on
each histogram.
For each measured value xj, there is a standard deviation j(h) associated with the
uncertainty in the observed frequency h(xj). This is not the same as the uncertainties associated
with the spread of the individual measurements xi about their mean , but rather describes the
spread of the measurement of each of the frequencies h(xj) about its mean j. If we were to
repeat the experiment many times to determine the distribution of frequency measurements at
each value of xj, we should find each parent distribution to be Poisson with mean j = y(xi) and
variance 2j (y) = y(x j ) . Thus, for each value of xj, there is a distribution curve, Pj(yk), that
describes the probability of obtaining the value of the frequency hk(xj) in the kth trial for each
value of j that is characterized by j(h). These distributions are illustrated in Figures 7 and 8 as
dotted Poisson curves at each value of xj. In Figure 7, the Poisson curves are centered at the

observed frequencies h(xj) with standard deviations j ( h ) = h ( x j ) . In principle, we should


center the Poisson curves at the frequencies j = y(xj) with standard deviation j ( h ) = j of the
parent population, as in Figure 8. However, in an actual experiment, we generally would not
know these parameters.

Figure 7: Histogram drawn form a Gaussian distribution mean = 5.0 and


standard deviation = 1, corresponding to 100 total measurements. The parent
17

distribution y(xj) = NP(xj) is illustrated by the large Gaussian curve. The smaller
dotted curves represent the Poisson distribution of events in each bin, based on the
sample data.

Figure 8: The same histogram as shown in Figure 7 with dotted curves


representing the Poisson distribution of events in each bin, based on the parent
distribution.

With the preceding definitions for n, N, xj, H(xj), P(xj), and j(h), the definition for 2 is

[h( x ) NP( x )]

j=1

j (h )

(47)

In most experiments, we do not know the values of j(h) because we make only one set of
measurements f(xj). Fortunately, these uncertainties can be estimated from the data directly
without measuring them explicitly.
If we consider the data in Figure 8, we observe that for each value of xj, we have
extracted a proportionate random sample of the parent population for that value. The
fluctuations in the observed frequencies h(xj) come from the statistical probabilities of making
random selections of finite numbers of items and are distributed according to the Poisson
distribution with y(xj) as mean.
For the Poisson distribution, the variance j(h)2 is equal to the mean y(xj) of the
distribution, and thus we can estimate j(h) from the data to be j ( h ) = NP ( x j ) = h ( x j ) .
Equation 47 becomes:

18

[h( x ) NP( x )] [h( x ) NP( x )]

NP ( x )
h( x )
n

j=1

j=1

(48)

As defined in equations (47) and (48), 2 is a statistic that characterizes how much the
observed frequency distributions differ from the expected frequencies based on Poisson statistics.

If the frequencies were to agree exactly with the predicted frequencies, then we should find 2 =
0. This is not a likely outcome. The numerator of equation 47 is a measure of the spread of the
observations; the denominator is a measure of the expected spread. We might imagine that for
good agreement, the average spread of the data would correspond to the expected spread, and
thus we should get a contribution of about one from each frequency, or 2 n for the entire
distribution. This is almost correct. The true expectation for the value of 2 is

2 = = n nc

(49)

where is the number of degrees of freedom and is equal to the number n of sample frequencies
minus the number nc of constraints or parameters that have been calculated from the data to
describe the probability function NP(x ). For our example, even if NP(x ) is chosen completely
j
j
independently of the distribution h(xj), there is still the normalizing factor N corresponding to the
total number of events in the distribution, so that the expectation value of 2 must at best be

2 = n 1

(50)

In order to estimate the probability that our calculated values of 2 are consistent with our
expected distribution of the data, we must know how 2 is distributed. If our value of 2

corresponds to a reasonably high probability, then we can have confidence in our assumed
distribution.
It is convenient to define the reduced chi-squared as

(40)

which should give a value of 1. Values of 2 much larger than 1 result from large deviations
from the assumed distribution and may indicate poor measurements, incorrect assignment of

uncertainties, or an incorrect choice of probability function. Very small 2 are also


unacceptable and may imply misunderstanding of the experiment. To quantitatively evaluate the

probability of observing a value of 2 equal to or greater than the value we calculated, we must
turn to a table that gives values of reduced chi-squared corresponding to the probability Px( 2; )

exceeding 2 versus the number of degrees of freedom. Details of how to use such a table are
given in Bevington and Robinson. In practice, we typically do not go to the trouble to

quantitatively evaluate the goodness of fit described by chi-square. Instead, we use it as a


general indicator of whether or not our data can be considered to obey a Gaussian distribution or
whether other sources of scatter are affecting the data. Reduced chi-squared is calculated by
several of our routines and you should look at the result. If reduced chi-squared is significantly
19

different from one, you should look for the problems in your data that cause such a result and
think hard about how to reduce, interpret, and report your data.

20

Вам также может понравиться