the centre) within which the true value is confined has a specified probability. The true
value of ox is unknown and an estimate of its value is to be obtained from the set of
observations consisting of a limited number of observations. The quantity called variance
given by 3,» y <1) is the best estimateloh squaneeh siandard deviation for a set
Tea
of n readings. The best estimate of standard deviation of the mean is then given by Oxm =
Sel, The terms AX AY... AZ are to be replaced by Gin, Gem. Gam Fespectively in the formula
for propagation of errors.
Mathematical results from the theory show that the probability of the true value of X lying
within an interval [xn- Om, Xm- Gin] Where Xm (represented also by + in statistical formulas)
represents the mean value of the n readings of X is about 68%. The quantity Gm is called the
standard uncertainty in the value of xi
The combined standard uncertainty in R due to random errors in various quantities X, ¥,...2
etc., is given by the square root of the sum of the squares of standard uncertainties of the
components,
Errors in different quantities could vary in magnitude as well as sign. Hence it becomes
necessary to consider probability of their distribution before combining them. It is accepted
that the combination is to be done in accordance with the rules for Normal (Gaussian)
Probability Distribution Function (PDF). We have considered AX as the error in X without
thinking how it is evaluated in measurement. If in the measurement we get different values
for X occurring in a random fashion, then the errors are of random type and the measure of
uncertainty is the standard deviation of the distribution evaluated using statistical formula
for normal distribution POF. (This is what is called type A uncertainty). When we know some
error is present but it does not change the reading as in the case of random errors we may
call it type B error (systematic error). To estimate its contribution to the uncertainty of the
result we have to estimate its maximum magnitude and assign an appropriate PDF to the
distribution of this error using our information about the measurement. When we know the
worst case limits of the measurement we can specify the interval within which the true
value of the measurand lies, If there is equal probability of obtaining a reading anywhere
within the interval then the distribution is said to be rectangular and the standard deviation
for this PDF when evaluated turns out to be equal to magnitude of the interval divided by
2V3 (or the worst case error divided by V3).
‘With the assumption that the standard uncertaint for all types of distributions combine
according to the rules of Gaussian PDF the procedure to obtain the combined uncertainty u.
involves adding the variances u’, uj? etc., to get the variance u.’. The combined uncertainty
giving the uncertainty in R is represented as u(A) and so we have2 a \ a
Pie geen) ae (eg
tax") “Vor
)
This is called the formula for propagation of uncertainties.
‘The above rule termed RSS (Root of Sum of Squares) is used to evaluate standard combined
uncertainty.
In this note, | have used Ax to represent the worst case error in X(which forms the upper
limit of uncertainty for type B errors) and u(x) u(Y) u(Z) etc. to represent standard
uncertainty in X, Y, Zetc. The quantity Aris estimated on the basis of our information and
u(X) is calculated for the PDF we assign to the distribution. Though it is possible to have
many types of distribution, in most cases the rectangular distribution is applicable
ESTIMATION OF UNCERTAINTIES IN SLOPE AND INTERCEPTS OF GRAPHS
(The method suggested below does not claim any authenticity. The standard method to
estimate uncertainty in the slope of a straight line graph is linear regression or least square
fit method which is applicable for type A uncertainties. More often the uncertainties which
we come across in the variables we plot to get linear graphs are of type B and that method
cannot be applied. So this method is suggested. )
Graphical methods are very useful for data analysis in physics experiments. Especially, when
there are constant systematic errors in either the independent or dependent variable, such
errors get clubbed in the intercept. They do not affect the slope. So, it is always advisable to
use slopes to calculate the results whenever possible. Hence we must see how to estimate
the uncertainty in the value of the slope from the uncertainties in the data.
After drawing the best fit line which is nearest to maximum number of points of the graph,
mark a point on it with coordinates (x1, y:) near one end of the graph. Choose another point
at the other end with coordinates (x2, y2) such that (x; ~ %4} is some round number. The
calculation of slope becomes easy with such choice. The slope is then equal to (ys- ¥2)/ (x: ~
x). Though data is represented on the graph by points, we know that the data have
uncertainties. This means in reality there are regions of uncertainties around the points on
the graph. One way to estimate the uncertainty in the slope of the graph is to determine the
uncertainties in (x: —%2) and (ys~ y2) from the uncertainties of the data. (This is equivalent to
calculation of error in the slope using the error bars on the graphs.) Though the points
chosen for determination of slope are selected arbitrarily we can assume that the
uncertainties of the points are of the same value as that of points nearby. The same idea is
8