high geometrical quality that they can be regarded as free
from errors, at least in comparison with the device under
calibration, the concepts “true” values and “true” errors
can be applied, although always with care. Consequently,
the concept and term accuracy can be used in this con
nection.
2.2. The Concepts and Terms Deviation and
Error.
At this moment there is reason to discuss briefly the
terms deviation and error, which frequently are used
in statistics and theory of errors, individually and in the
combinations standard deviation and standard error (of
unit weight), mean deviation and mean error, probable
deviation and probable error, etc.
There seem to be good reasons to combine the term
deviation with the concept and term precision and the term
error with the concept and term accuracy.
In statistics deviation is usually applied to express the
difference between individual repeated or replicated
measurements and the average. It seems suitable to define
deviation = measured value — average = x — x
The variance is then defined as
n—1 n—1
where n is the number of measurements and n — 1 the de
grees of freedom 1
The standard deviation is then defined as the positive
square root of the variance or
Ordinarily, s is an expression for the precision (rather
imprecision) of one and each of the measurements x. The
precision of the average x is evidently higher than that of
one single measurement, i.e. the standard deviation of the
average is lower than that of a single measurement. This is
also expressed by the well known formula for the standard
deviation of the average
s
This formula can be derived by applying the special law of
propagation of errors and deviations to the expression for
the average
-_I]*_*l+*2 + + X n
n n
1 It should be noted that the latin character 5 is used instead of
the greek a (sigma). This is in agreement with ordinary statistical
practice where o represents the entire population and s a sample
of deviations. In measurements the population of deviations
and errors is infinitely large. Therefore, here s is always used
instead of o.
where each of the measured values x x ... x n is assumed to
be affected with the standard deviation s. Each value is
then assumed to be independent and free from correlation
with the others. If all measured values were affected with
errors of the same magnitude and direction (a constant
error) they would not appear in the standard deviation.
The measured values would in such a case be physically
correlated.
If the results of measurements, for instance the average
x, is compared with given (true) values, which is the case in
all calibration procedures, the discrepancy or error is
defined as
+ e = measured value—given value 2
Provided that the given value can be regarded as free
from errors, at least in comparison with the measured
value, the quantity e can be regarded as a true error or
discrepancy.
If there are several such determinations of errors of
similar character, for instance in photogrammetric model
coordinates, each of the errors represents the concept of
accuracy. Statistically, they can be represented by the
root mean square error (discrepancy) defined as
Frequently and particularly in connection with calibra
tions, the discrepancies e are regarded as indirect measure
ments of (gross errors), systematic (regular) errors and
irregular errors of the measured values.
The discrepancies e are interpreted as a linear differential
formula (mathematical model) where the parameters are
possible systematic errors according to physical and other
circumstances. In particular the manufacturers of instru
ments should be the natural source of information as to
possible systematic errors of the instruments. Because in
principle there always shall be redundant discrepancies in
the calibration procedure, the parameters shall ordinarily
be determined under the condition that the sum of the
squares of the residuals shall be a minimum. This leads to
the system of normal equations, the solution of which
gives the parameters, the minimized sum of squares, etc.
If this minimized sum of squares is divided by the number
of redundant discrepancies the variance of the residuals or
residual variance is obtained. The positive square root of
this variance is denoted standard error of unit weight (s 0 )
and is an expression for the accuracy of each of the mea
sured values after removing the regular errors. Because
each systematic error is a well defined linear function of the
measured values which are assumed to be independent,
2 A correction v, to be added to the measured value in order
to obtain the given value, is consequently defined as v = - e.
If, for some reason, other definitions of the signs of e and v are
used, as for instance in geodesy, this should always be clearly
stated.
»
the stan
from tl
standar
standar
weight
the n<
error
instanc<
propagi
be pai<
parame
weight
in a m
dents
With
and of
ficance
functio
other
with
of the
ment
with
of th<
II: 6
shall
actual
when
2.3
set
in
they
be
that
mat!
orth
influ
edge
assu
10